diff --git a/.gitattributes b/.gitattributes index 72bde8aa946b91f3aa2aa1bdaae342cf431ff564..030cc378cb9705521bfea85c13815666909c84eb 100644 --- a/.gitattributes +++ b/.gitattributes @@ -3876,3 +3876,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2024/LAKE-RED_[[:space:]]Camouflaged[[:space:]]Images[[:space:]]Generation[[:space:]]by[[:space:]]Latent[[:space:]]Background[[:space:]]Knowledge[[:space:]]Retrieval-Augmented[[:space:]]Diffusion/12722ca1-ee01-4244-ac76-dd0f56c3ada2_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/LAMP_[[:space:]]Learn[[:space:]]A[[:space:]]Motion[[:space:]]Pattern[[:space:]]for[[:space:]]Few-Shot[[:space:]]Video[[:space:]]Generation/a527aaf0-4fd2-4061-87f8-d8fd7f439d09_origin.pdf filter=lfs diff=lfs merge=lfs -text 2024/LAN_[[:space:]]Learning[[:space:]]to[[:space:]]Adapt[[:space:]]Noise[[:space:]]for[[:space:]]Image[[:space:]]Denoising/74bb68f0-698b-4fd3-96c1-35383b7bb71f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LASA_[[:space:]]Instance[[:space:]]Reconstruction[[:space:]]from[[:space:]]Real[[:space:]]Scans[[:space:]]using[[:space:]]A[[:space:]]Large-scale[[:space:]]Aligned[[:space:]]Shape[[:space:]]Annotation[[:space:]]Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LASIL_[[:space:]]Learner-Aware[[:space:]]Supervised[[:space:]]Imitation[[:space:]]Learning[[:space:]]For[[:space:]]Long-term[[:space:]]Microscopic[[:space:]]Traffic[[:space:]]Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LASO_[[:space:]]Language-guided[[:space:]]Affordance[[:space:]]Segmentation[[:space:]]on[[:space:]]3D[[:space:]]Object/4a958b96-aed8-4e26-96c9-3b681d296992_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LDP_[[:space:]]Language-driven[[:space:]]Dual-Pixel[[:space:]]Image[[:space:]]Defocus[[:space:]]Deblurring[[:space:]]Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEAD_[[:space:]]Exploring[[:space:]]Logit[[:space:]]Space[[:space:]]Evolution[[:space:]]for[[:space:]]Model[[:space:]]Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEAD_[[:space:]]Learning[[:space:]]Decomposition[[:space:]]for[[:space:]]Source-free[[:space:]]Universal[[:space:]]Domain[[:space:]]Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEAP-VO_[[:space:]]Long-term[[:space:]]Effective[[:space:]]Any[[:space:]]Point[[:space:]]Tracking[[:space:]]for[[:space:]]Visual[[:space:]]Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEDITS++_[[:space:]]Limitless[[:space:]]Image[[:space:]]Editing[[:space:]]using[[:space:]]Text-to-Image[[:space:]]Models/f8bff481-3f32-424c-92f1-b66a29a1df02_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LED_[[:space:]]A[[:space:]]Large-scale[[:space:]]Real-world[[:space:]]Paired[[:space:]]Dataset[[:space:]]for[[:space:]]Event[[:space:]]Camera[[:space:]]Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEMON_[[:space:]]Learning[[:space:]]3D[[:space:]]Human-Object[[:space:]]Interaction[[:space:]]Relation[[:space:]]from[[:space:]]2D[[:space:]]Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LEOD_[[:space:]]Label-Efficient[[:space:]]Object[[:space:]]Detection[[:space:]]for[[:space:]]Event[[:space:]]Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LION_[[:space:]]Empowering[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]with[[:space:]]Dual-Level[[:space:]]Visual[[:space:]]Knowledge/ee456ca8-65df-48b8-a118-352557caef54_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LISA_[[:space:]]Reasoning[[:space:]]Segmentation[[:space:]]via[[:space:]]Large[[:space:]]Language[[:space:]]Model/92907404-c48e-45be-bfa1-4c27e2c673e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LL3DA_[[:space:]]Visual[[:space:]]Interactive[[:space:]]Instruction[[:space:]]Tuning[[:space:]]for[[:space:]]Omni-3D[[:space:]]Understanding[[:space:]]Reasoning[[:space:]]and[[:space:]]Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLM4SGG_[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Weakly[[:space:]]Supervised[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLMs[[:space:]]are[[:space:]]Good[[:space:]]Action[[:space:]]Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLMs[[:space:]]are[[:space:]]Good[[:space:]]Sign[[:space:]]Language[[:space:]]Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLaFS_[[:space:]]When[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]Meet[[:space:]]Few-Shot[[:space:]]Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LLaMA-Excitor_[[:space:]]General[[:space:]]Instruction[[:space:]]Tuning[[:space:]]via[[:space:]]Indirect[[:space:]]Feature[[:space:]]Interaction/54e482bc-9979-46be-8235-fd3729928e88_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LMDrive_[[:space:]]Closed-Loop[[:space:]]End-to-End[[:space:]]Driving[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LORS_[[:space:]]Low-rank[[:space:]]Residual[[:space:]]Structure[[:space:]]for[[:space:]]Parameter-Efficient[[:space:]]Network[[:space:]]Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LOTUS_[[:space:]]Evasive[[:space:]]and[[:space:]]Resilient[[:space:]]Backdoor[[:space:]]Attacks[[:space:]]through[[:space:]]Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LP++_[[:space:]]A[[:space:]]Surprisingly[[:space:]]Strong[[:space:]]Linear[[:space:]]Probe[[:space:]]for[[:space:]]Few-Shot[[:space:]]CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LPSNet_[[:space:]]End-to-End[[:space:]]Human[[:space:]]Pose[[:space:]]and[[:space:]]Shape[[:space:]]Estimation[[:space:]]with[[:space:]]Lensless[[:space:]]Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LQMFormer_[[:space:]]Language-aware[[:space:]]Query[[:space:]]Mask[[:space:]]Transformer[[:space:]]for[[:space:]]Referring[[:space:]]Image[[:space:]]Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LSK3DNet_[[:space:]]Towards[[:space:]]Effective[[:space:]]and[[:space:]]Efficient[[:space:]]3D[[:space:]]Perception[[:space:]]with[[:space:]]Large[[:space:]]Sparse[[:space:]]Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LTA-PCS_[[:space:]]Learnable[[:space:]]Task-Agnostic[[:space:]]Point[[:space:]]Cloud[[:space:]]Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LTGC_[[:space:]]Long-tail[[:space:]]Recognition[[:space:]]via[[:space:]]Leveraging[[:space:]]LLMs-driven[[:space:]]Generated[[:space:]]Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LTM_[[:space:]]Lightweight[[:space:]]Textured[[:space:]]Mesh[[:space:]]Extraction[[:space:]]and[[:space:]]Refinement[[:space:]]of[[:space:]]Large[[:space:]]Unbounded[[:space:]]Scenes[[:space:]]for[[:space:]]Efficient[[:space:]]Storage[[:space:]]and[[:space:]]Real-time[[:space:]]Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LUWA[[:space:]]Dataset_[[:space:]]Learning[[:space:]]Lithic[[:space:]]Use-Wear[[:space:]]Analysis[[:space:]]on[[:space:]]Microscopic[[:space:]]Images/962a883b-464b-4830-ab92-8bd978dfedca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LaMPilot_[[:space:]]An[[:space:]]Open[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving[[:space:]]with[[:space:]]Language[[:space:]]Model[[:space:]]Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LaRE^2_[[:space:]]Latent[[:space:]]Reconstruction[[:space:]]Error[[:space:]]Based[[:space:]]Method[[:space:]]for[[:space:]]Diffusion-Generated[[:space:]]Image[[:space:]]Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Label[[:space:]]Propagation[[:space:]]for[[:space:]]Zero-shot[[:space:]]Classification[[:space:]]with[[:space:]]Vision-Language[[:space:]]Models/f789bcea-03da-4382-8fd5-69048545fa3d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Label-Efficient[[:space:]]Group[[:space:]]Robustness[[:space:]]via[[:space:]]Out-of-Distribution[[:space:]]Concept[[:space:]]Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Lane2Seq_[[:space:]]Towards[[:space:]]Unified[[:space:]]Lane[[:space:]]Detection[[:space:]]via[[:space:]]Sequence[[:space:]]Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LaneCPP_[[:space:]]Continuous[[:space:]]3D[[:space:]]Lane[[:space:]]Detection[[:space:]]using[[:space:]]Physical[[:space:]]Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LangSplat_[[:space:]]3D[[:space:]]Language[[:space:]]Gaussian[[:space:]]Splatting/591ba264-2503-43cc-8f70-dc10b9821047_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language[[:space:]]Embedded[[:space:]]3D[[:space:]]Gaussians[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Scene[[:space:]]Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language[[:space:]]Model[[:space:]]Guided[[:space:]]Interpretable[[:space:]]Video[[:space:]]Action[[:space:]]Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language[[:space:]]Models[[:space:]]as[[:space:]]Black-Box[[:space:]]Optimizers[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/40b8d27b-a1d9-46c0-919c-6079601bc412_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-Driven[[:space:]]Anchors[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Adversarial[[:space:]]Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-aware[[:space:]]Visual[[:space:]]Semantic[[:space:]]Distillation[[:space:]]for[[:space:]]Video[[:space:]]Question[[:space:]]Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-conditioned[[:space:]]Detection[[:space:]]Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-driven[[:space:]]All-in-one[[:space:]]Adverse[[:space:]]Weather[[:space:]]Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-driven[[:space:]]Grasp[[:space:]]Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-driven[[:space:]]Object[[:space:]]Fusion[[:space:]]into[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]with[[:space:]]Pose-Conditioned[[:space:]]Dataset[[:space:]]Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-guided[[:space:]]Image[[:space:]]Reflection[[:space:]]Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Language-only[[:space:]]Training[[:space:]]of[[:space:]]Zero-shot[[:space:]]Composed[[:space:]]Image[[:space:]]Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Laplacian-guided[[:space:]]Entropy[[:space:]]Model[[:space:]]in[[:space:]]Neural[[:space:]]Codec[[:space:]]with[[:space:]]Blur-dissipated[[:space:]]Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Large[[:space:]]Language[[:space:]]Models[[:space:]]are[[:space:]]Good[[:space:]]Prompt[[:space:]]Learners[[:space:]]for[[:space:]]Low-Shot[[:space:]]Image[[:space:]]Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Latency[[:space:]]Correction[[:space:]]for[[:space:]]Event-guided[[:space:]]Deblurring[[:space:]]and[[:space:]]Frame[[:space:]]Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Latent[[:space:]]Modulated[[:space:]]Function[[:space:]]for[[:space:]]Computational[[:space:]]Optimal[[:space:]]Continuous[[:space:]]Image[[:space:]]Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Layout-Agnostic[[:space:]]Scene[[:space:]]Text[[:space:]]Image[[:space:]]Synthesis[[:space:]]with[[:space:]]Diffusion[[:space:]]Models/cffc0cc1-9751-4215-8102-3e240ce1394d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LayoutFormer_[[:space:]]Hierarchical[[:space:]]Text[[:space:]]Detection[[:space:]]Towards[[:space:]]Scene[[:space:]]Text[[:space:]]Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LayoutLLM_[[:space:]]Layout[[:space:]]Instruction[[:space:]]Tuning[[:space:]]with[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]for[[:space:]]Document[[:space:]]Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/LeGO_[[:space:]]Leveraging[[:space:]]a[[:space:]]Surface[[:space:]]Deformation[[:space:]]Network[[:space:]]for[[:space:]]Animatable[[:space:]]Stylized[[:space:]]Face[[:space:]]Generation[[:space:]]with[[:space:]]One[[:space:]]Example/6836350f-601e-42fa-b6e0-49050e114ece_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Leak[[:space:]]and[[:space:]]Learn_[[:space:]]An[[:space:]]Attacker's[[:space:]]Cookbook[[:space:]]to[[:space:]]Train[[:space:]]Using[[:space:]]Leaked[[:space:]]Data[[:space:]]from[[:space:]]Federated[[:space:]]Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learn[[:space:]]from[[:space:]]View[[:space:]]Correlation_[[:space:]]An[[:space:]]Anchor[[:space:]]Enhancement[[:space:]]Strategy[[:space:]]for[[:space:]]Multi-view[[:space:]]Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learn[[:space:]]to[[:space:]]Rectify[[:space:]]the[[:space:]]Bias[[:space:]]of[[:space:]]CLIP[[:space:]]for[[:space:]]Unsupervised[[:space:]]Semantic[[:space:]]Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learnable[[:space:]]Earth[[:space:]]Parser_[[:space:]]Discovering[[:space:]]3D[[:space:]]Prototypes[[:space:]]in[[:space:]]Aerial[[:space:]]Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learned[[:space:]]Lossless[[:space:]]Image[[:space:]]Compression[[:space:]]based[[:space:]]on[[:space:]]Bit[[:space:]]Plane[[:space:]]Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learned[[:space:]]Representation-Guided[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Large-Image[[:space:]]Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learned[[:space:]]Scanpaths[[:space:]]Aid[[:space:]]Blind[[:space:]]Panoramic[[:space:]]Video[[:space:]]Quality[[:space:]]Assessment/4097298f-874f-411d-9a63-b8f9db525955_origin.pdf filter=lfs diff=lfs merge=lfs -text +2024/Learned[[:space:]]Trajectory[[:space:]]Embedding[[:space:]]for[[:space:]]Subspace[[:space:]]Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_content_list.json b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d18717d8b5fe35948844633b0da61d7ef6f1589e --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset", + "text_level": 1, + "bbox": [ + 135, + 128, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haolin Liu $^{1,2*}$ , Chongjie Ye $^{1,2*}$ , Yinyu Nie $^{3}$ , Yingfan He $^{1,2}$ , Xiaoguang Han $^{2,1\\dagger}$", + "bbox": [ + 174, + 200, + 795, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\\*equal contribution \\*corresponding author", + "bbox": [ + 341, + 223, + 624, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FNii, CUHKSZ $^{2}$ SSE, CUHKSZ $^{3}$ Technical University of Munich", + "bbox": [ + 178, + 243, + 787, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gap-lab-cuhk-sz.github.io/LASA", + "bbox": [ + 383, + 272, + 580, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "920 Scenes + 10,412 Unique CADs", + "bbox": [ + 125, + 339, + 390, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/250f7aba2f0d0ed674bac9d0d817575f4c2c315eb2c28d6bb78d3d454e6867b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 367, + 158, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2db398d93e153f11ff0a35b8312cdbe2f25278979f027dd91e6f39accf7b466.jpg", + "image_caption": [ + "Figure 1. We introduce LASA, a Large-scale Aligned Shape Annotation Dataset containing 10,412 unique object CAD models aligned with 920 real-world scene scans. Supported by LASA, we achieve state-of-the-art in real-world object reconstruction and 3D object detection." + ], + "image_footnote": [], + "bbox": [ + 96, + 429, + 158, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e21ddd04138c3c785c1c5b5fdca31b0e0665726e661a24e944ea65317457f7e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 367, + 223, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ffc6a5a8d2e899ace848d26b2e134cb11cf1d3d92cb9a896053a9e6437abacfa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 429, + 222, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e76eb70f3920aa8b231a84824906074c69598e1bd7749a69e443a526c4161ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 367, + 285, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/96dabc861db0528e2d36664c51b2c24b6927d4178e72a52169ff8bb3e25f22e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 429, + 285, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ebe7ab88b8ddf99b334e59a60375e35afd494e0ea99a2c0685dbf8f7b4e2c1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 367, + 351, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f124df7c5e8989ebf26c41de3b0694980d77c6df5cf746c699c0d5293e407705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 429, + 351, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f53da1c19e3f1167d744266c5711a57b0ae598e6a6854823265e3c87e08908c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 367, + 413, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8eb88f5dd7c7870c5c60ed4b6c6a877265ef2787e72d788b71998e7086670644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 429, + 413, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale Aligned CAD Annotations", + "bbox": [ + 99, + 506, + 406, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4f314354ca6b825ac15f587ae20156c86c8b3844faedf64e6c6aeb09df164a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 415, + 470, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/63b8636a0b3f66e55b0a7d7e1a17a452212975dc20ec173641c99db1228e7ae5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 477, + 333, + 552, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/968cdb0a1c96d242eeec51a027813212ffe68957e2d75a0a56ef252ba10da2d6.jpg", + "image_caption": [ + "Monocular RGB-D video" + ], + "image_footnote": [], + "bbox": [ + 480, + 410, + 550, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1fa87864012d14ec262b6d4a2e33b8c3d9043193a3f874b7e92b464d303ff53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 330, + 633, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/052bc906431840c2762d7f94457aa1aed39906726dc16b186df3d9cf68b81f94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 410, + 632, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d28a971c530ebb099f0d0f03385485d35b6d0989b0b35bc0b5e499b74ee9d797.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 330, + 728, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ddd3e77910a700d0f50e7307bdd9865806a2d601d82447a9f699a99ba674d77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 409, + 728, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/292844475f887dafc5f80d7ee29acb8b9496cfa4e0967f67ce9e69f4a4e2f5c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 463, + 728, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c38814998b2bf490b7f9c3b15c0c32d08919b752ee535456ab8e1b9046b7fdcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 333, + 823, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac9ab39c2f2507753a599fd65c3cef306f00a0a1ad9a6b9d75f808d669123b5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 406, + 823, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86eca4bfb0549f5ba7db7f2add5682ef78d28f8cc2a712fa60ab639d1d80d3dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 463, + 823, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/733dcbe6abad5b4c0f85072a097c9273b4ea1b2ef29978444e5a7cd7cd283b05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 462, + 875, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstruction Results", + "bbox": [ + 571, + 507, + 781, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 575, + 313, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance shape reconstruction from a 3D scene involves recovering the full geometries of multiple objects at the semantic instance level. Many methods leverage data-driven learning due to the intricacies of scene complexity and significant indoor occlusions. Training these methods often requires a large-scale, high-quality dataset with aligned and paired shape annotations with real-world scans. Existing datasets are either synthetic or misaligned, restricting the performance of data-driven methods on real data. To this end, we introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising 10,412 high-quality CAD annotations aligned with 920 real-world scene scans from ArkitScenes, created manually by professional artists. On this top, we propose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo) method. It is empowered by a hybrid feature aggregation design to fuse multi-modal inputs and recover high-fidelity object geometries (see Fig. 1). Besides, we present an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate that our shape", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "annotations provide scene occupancy clues that can further improve 3D object detection. Supported by LASA, extensive experiments show that our methods achieve state-of-the-art performance in both instance-level scene reconstruction and 3D object detection tasks.", + "bbox": [ + 498, + 578, + 893, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 664, + 632, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The widespread use of hand-held RGB-D sensors has facilitated the effortless acquisition of indoor scene scans. However, these scans often suffer from noises and incompleteness due to limitations in sensor accuracy, the complexity of indoor environments, and occlusion among objects. This further limits its applications in scenarios such as VR/AR and 3D industry where a complete and high-quality reconstruction is desired. This shortage absorbs great attention in 3D vision and graphics community, particularly in the realm of indoor instance-level scene reconstruction. In this task, the objective is to reconstruct the shapes of observed objects based on 3D scans or images captured by sensors. Many advances [8, 36, 56, 58, 61] have been seen by leveraging the power of deep learning methods for this task. They", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20454", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are data-driven and demand a substantial number of paired scene scans and objects' CAD ground truths.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing data-driven methods rely on two types of datasets. [1, 7-9, 19, 27, 39, 50, 52, 55-58] utilized synthetic datasets [5, 10, 15] by synthesizing images and scans that mimic real-world distributions. Synthetic data provides CAD models perfectly aligned with input observations (images or scans) though. Models trained on it are vulnerable to domain gaps from the real world [53] that could impair the generalizability. Scan2CAD[2] attempts to bridge this gap by annotating objects' CAD in real-world scene scans. However, their CADs are retrieved from a synthetic database [5] and their poses are manually adjusted to roughly align the object scans. However, retrieved shapes are often unlike the real observed objects, introducing shape misalignment issues. Many works [13, 18, 28, 32, 47, 48] using it as instance shape supervision are potentially biased with inferior alignment. In summary, The absence of a well-aligned real-world dataset barriers the further advancement of the community.", + "bbox": [ + 75, + 125, + 470, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To surmount this challenge, we present a new dataset LASA, a Large-scale Aligned Shape Annotation dataset that contains 10,412 high-quality instance CAD annotations meticulously crafted by skilled artists. Each CAD shape is designed and placed to precisely align with objects' scans from 920 real-world scene scans in ArkitScene [3]. We deliberately annotate objects in ArkitScene instead of ScanNet[11] because it provides scans obtained from both accurate Laser sensors and hand-held RGB-D sensors. The accurate Laser scan benefits high-quality and aligned manual annotations. Meanwhile, the less accurate scans from hand-held devices support research on reconstruction and completion using consumer-level devices.", + "bbox": [ + 75, + 398, + 470, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A large-scale scan dataset with high-quality instance CAD annotations empowers many downstream applications. We first investigate how LASA benefits indoor instance-level scene reconstruction. Given an indoor scene, images and scene scans can be obtained through scanning. Typically, the initial step involves 3D object detection, after which the partial point clouds and multi-view images of each detected instance are acquired, serving as visual cues for subsequent shape reconstruction. Point cloud provides natural 3D information though, they are often noisy and incomplete. On the other hand, images present rich appearance clues but lack 3D constraints. Inspired by the complementary nature of these two modalities in describing object surfaces, we advocate utilizing both modalities as inputs to fuse their advantages. Recently, diffusion models have shown appealing performance in shape generation [7, 19, 44, 61, 64]. We advance it further for instance reconstruction and propose Diffusion-based Cross-modal Shape Reconstruction, namely DisCo. DisCo is a triplane diffusion model tailored to accommodate multi-modal in", + "bbox": [ + 75, + 598, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "puts, towards robust real-world object reconstruction, where a hybrid feature aggregation layer is proposed to aggregate and align two input modalities effectively. Supported by the LASA dataset, extensive experiments show that our method achieves state-of-the-art reconstruction performance with real-world inputs.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the LASA dataset has full annotations covering every instance within each of the 920 real-world scenes. This extensive coverage enables LASA to provide scene-level occupancy labels. We further explore how they can affect indoor scene 3D object detection. 3D Object detection [37, 42, 49] usually comprises a backbone and a detection head. We propose to integrate an occupancy prediction head for occupancy-guided 3D object detection, namely OccGOD. Our experiments demonstrate that incorporating occupancy prediction leads to substantial improvements in detection performance. In summary, our key contributions are four-fold:", + "bbox": [ + 496, + 181, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a large-scale dataset, LASA. It contains 10,412 manually crafted, high-quality instance CAD annotations geometrically aligned with 920 real-world scene scans.", + "- We propose DisCo, a novel diffusion-based method that leverages hybrid representation, effectively interacting with both input partial point cloud and multi-view images, achieving state-of-the-art reconstruction.", + "- With LASA's scene-level annotation, we introduce occupancy-guided 3D object detection (OccGOD) with decent improvements.", + "- We strongly believe the large-scale dataset of well-aligned shape annotations can break the bottleneck of current research on 3D indoor scene understanding and reconstruction." + ], + "bbox": [ + 500, + 366, + 890, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 609, + 650, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Indoor Instance Shape Dataset", + "text_level": 1, + "bbox": [ + 500, + 633, + 772, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, many advances have been seen in learning-based 3D object and scene reconstruction from images and videos. However, numerous challenges persist due to the limitations of existing datasets.", + "bbox": [ + 496, + 657, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These learning-based approaches are usually trained either on existing synthetic datasets or real-world datasets. While synthetic datasets are demonstrated valuable for training models, they lack realism. Large synthetic object collections like ShapeNet[5] and ABO[10] provide diverse 3D models but lack environment context. Synthetic scene datasets such as 3D-Front[14], Replica[45], and Structured3D[63] possess complete synthetic environments with CAD annotations for objects. However, models trained on these datasets often struggle to generalize with real inputs due to substantial domain gaps[53].", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some works [2, 29, 34, 46, 53, 54] annotate CAD", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20455", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/1494b02ad58fea5756601f0600a2e7bc6691b15a5a1c1927e6d5c051e16a2078.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAligned#Scenes#CADsSensor TypeAnnotation Method
Scan2CAD [2]-1,5063,049RGB-DRetrieval
CAD-Estate [29]-19,51212,024RGBRetrieval
IKEA [16]✓null90RGBRetrieval
Pix3D [46]✓null219RGBRetrieval
ScanSalon [53]✓413800RGB-DArtist
Aria's Digital Twin [34]✓2370RGB-DArtist
LASA (Ours)✓92010,412RGB-DArtist
", + "bbox": [ + 163, + 87, + 807, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons with existing 3D indoor datasets with instance shape annotations", + "bbox": [ + 223, + 224, + 743, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models on real-world data to bridge the domain gaps. Scan2CAD [2] and CAD-estate [29] have provided scanned object-CAD pairs by retrieving them from ShapeNet [5] and are further manually aligned to the real-world scenes. However, these retrieved CAD models lack alignment with real objects, potentially biasing data-driven reconstruction methods towards inferior reconstruction. This motivates us to build a real-world dataset with well-aligned CAD annotations. Aria's Digital Twin [34] and ScanSalon [53] provide such CADs though, the limited quantities restrict their application for data-hungry tasks. Datasets such as Pix3d [46] and PASCAL3D+ [54] supply single-view images with aligned shapes but are limited in scene modalities lacking point cloud and multi-view images.", + "bbox": [ + 75, + 251, + 472, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Shape Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 469, + 308, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object-level Reconstruction Existing approaches leverage images or partial point clouds as input. Images based reconstruction methods accept either single-view [6, 17, 19, 20, 25, 27, 30, 31, 33, 35, 50, 55, 57, 62, 64] or sparse multiview images [4, 9, 39, 56, 59] as inputs for shape reconstruction. They extract 2D image features and utilize them for shape reconstruction. Some [27, 43] leverage 2D pixel-aligned local features for reconstruction, demonstrating appealing performance for high-quality reconstruction. Other works [1, 8, 12, 26, 40, 52, 58, 61] conduct shape reconstruction from noisy, incomplete partial point clouds. Both paradigms have their advantages, with image-based methods perceiving rich 2D appearance features, while point-based methods process native 3D signals. [21] proposed to combine both inputs to fuse their advantages. Our proposed approach also accepts both inputs, focusing on effectively aggregating and aligning local features from both input sources. Since diffusion model [7, 19, 44, 51, 61, 64] have demonstrated strong capabilities in both shape generation and reconstruction, we opt for a diffusion model for robust and high-quality reconstruction.", + "bbox": [ + 75, + 492, + 468, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance-level Scene Reconstruction Recent instance scene reconstruction methods [13, 18, 21-23, 28, 32, 47, 48] follow a detect-then-reconstruct pipeline from a single scan. First, these methods use a 3D object detector to localize objects in either scene scans or videos. Subsequently, the detected 3D objects are fed into a 3D shape reconstruction", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9e8e291f691372d1661d76b3bf441b41a5f83b739e2d63fbde559d6887cebf2.jpg", + "image_caption": [ + "Figure 2. Visual comparison between aligned and unaligned CAD annotations" + ], + "image_footnote": [], + "bbox": [ + 516, + 251, + 893, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "module individually to obtain object shapes. Finally, the reconstructed 3D object shapes are placed back at their original locations to obtain a reconstruction of the full scene. These methods commonly utilize the Scan2CAD [2] dataset for training, where the inferior alignment limits their reconstruction quality. In contrast, the CAD models in our LASA dataset are manually crafted by artists to guarantee alignment, which we hope can lay a foundation for future research in this area.", + "bbox": [ + 496, + 542, + 893, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LASA Dataset", + "text_level": 1, + "bbox": [ + 500, + 707, + 643, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LASA is a large-scale dataset that contains 10,412 unique CAD models covering 920 scenes across 17 categories. Rather than relying on a pre-existing CAD database to annotate scenes, LASA engages professional artists to manually create aligned CAD models with 3D scans. Our annotations provide precise and consistent (as shown in Fig. 2) training data for data-driven reconstruction algorithms.", + "bbox": [ + 496, + 732, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Annotation", + "text_level": 1, + "bbox": [ + 500, + 847, + 665, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LASA is built upon the 3D laser scans from ArkitScenes since it is accurate with high resolution, which is critical for", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20456", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "high-fidelity CAD annotations.", + "bbox": [ + 76, + 90, + 282, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Preprocessing. Each laser scan from ArkitScenes is exceptionally dense with $1\\mathrm{GB}+$ data storage. To improve annotation efficiency, we downsample the scans to a 4mm density. Since ArkitScenes does not publicly provide alignment transformations between the laser scans and the RGB-D scans, we utilize a coarse-to-fine registration method to calculate the transformation matrix (see our supplemental for details). With the transformation matrix, we align the laser data with the RGB-D sequence coordinates from ArkitScenes. We then use ArkitScenes' 3D bounding box annotations to partition the aligned laser point clouds for each single object. These segmented point clouds are transformed into the canonical space. Furthermore, for each object, we select 2-5 frames from RGB-D scans that maximize the 2D projection area of its 3D bounding box. These selected frames serve as references for annotation.", + "bbox": [ + 75, + 106, + 467, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shape Annotation. The CAD annotation process involves a team of 35 artists working over 4 months. With preprocessed point clouds and reference images, each artist spent approximately 69 minutes designing a single model. Each model is annotated with Autodesk Maya or Cinema 4D.", + "bbox": [ + 75, + 347, + 467, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shape Verification. We involve a shape verification procedure for annotation quality control. This procedure has both algorithmic validation and manual reviews to thoroughly evaluate CAD model's accuracy against ground truth scans and images. Our multi-step verification process includes", + "bbox": [ + 75, + 422, + 467, + 498 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Senior Review: After initial annotations, 6 senior designers reviewed every CAD model to verify quality, accuracy, and reliability by manually cross-checking against the 3D scans. Any models that failed to meet the standards were flagged for rework.", + "- Geometry Alignment: We matched the CAD models to aligned laser scans and calculated the unidirectional Chamfer distance between them. This quantified the raw geometric alignment error for the CAD surface compared to the ground truth scan.", + "- View Alignment: We also verified alignment in the pixel level, where we rendered 112,639 images across all scenes by positioning the CAD models in the view frustum of RGB-D sensors. Crowd workers performed a manual inspection to check if the rendered views overlay on the real images. They checked for inconsistencies along object edges and intricate details which would indicate misalignment. This pixel-level evaluation ensured precise alignment. Any rendering mismatches were fixed by re-annotating the CAD model." + ], + "bbox": [ + 76, + 502, + 467, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Statistics", + "text_level": 1, + "bbox": [ + 76, + 816, + 243, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tab. 1 shows the statistics of LASA compared to the existing datasets. LASA contains a comparable number of unique CAD models to CAD-Estate [29], which was previously the largest scene CAD dataset. However, LASA pro", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vides better CAD annotation quality against their retrieved CAD models. Additionally, LASA demonstrates greater shape diversity than Scan2CAD [2], with over 3 times as many unique CAD models.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Among all aligned datasets, LASA stands out with a total of 10,412 CADs. This is 13 times more than ScanSalon [53] and 28 times more than Aria's Digital Twin dataset[34]. Unlike IKEA [16] and Pix3D [46] which are annotated on single-view RGB images, LASA captures full RGB-D sequences. This enables a wider range of downstream applications compared to static image datasets.", + "bbox": [ + 498, + 151, + 890, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we compare LASA with Scan2CAD in terms of alignment quality by measuring the Chamfer distance between CAD annotation and the scene scans, which are 0.161 vs 0.269. LASA provides much better aligned data.", + "bbox": [ + 498, + 257, + 888, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Instance-level Scene reconstruction", + "text_level": 1, + "bbox": [ + 498, + 330, + 816, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a Diffusion-based Cross-modal Shape Reconstruction method (DisCo). DisCo is a diffusion-based model to pursue high-fidelity 3D shape reconstruction from partial point clouds and multi-view images. Various representation including latent sets [61], volumetric grids [7, 64], and triplanes [19, 39, 51] are popular choice for diffusion model. We opt for the triplane since its efficiency enables higher output resolution compared to volumetric grids; while reserving 3D structure compared to latent sets. We employ latent triplane diffusion, where a triplane variational auto-encoder (VAE) first encodes shape into triplane latent space (in Sec. 4.1). Subsequently, a triplane diffusion model operates on this latent space for 3D shape reconstruction conditioned on both partial points and multi-view images (in Sec. 4.2). The overall pipeline of DisCo is shown in the upper part of Fig. 3.", + "bbox": [ + 496, + 354, + 888, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It has been demonstrated that using multiple input modalities (images and scans) presents complementary benefits for semantic scene completion [21]. In DisCo, we also fuse image and point features and introduce Hybrid Feature Aggregation Layer (in Sec. 4.3). In this layer, we utilize a hybrid representation combining a triplane and a volumetric grid. This combination facilitates efficient local feature aggregation and feature alignment from both partial point clouds and multi-view images.", + "bbox": [ + 496, + 597, + 888, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Triplane Variational Auto-Encoder", + "text_level": 1, + "bbox": [ + 498, + 741, + 803, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To conduct a latent triplane diffusion, the first step is to learn an encoder capable of encoding shapes in a triplane latent space. The Triplane VAE [19] comprises an encoder $\\psi_{enc}$ and a decoder $\\psi_{dec}$ . The encoder processes inputs and encodes them into latent space, while the decoder recovers shape from the latent representation. Surface point cloud $\\mathbf{P} \\in \\mathbb{R}^{K \\times 3}$ sampled from a ground truth shape, serves as inputs to $\\psi_{enc}$ . $K = 20,000$ in our experiment. These points are projected onto a triplane and subsequently processed by", + "bbox": [ + 496, + 763, + 888, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20457", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b51d3c5063ba281352084f533b9cf9ce768d10f3b7df4ed39c53f869f3a8be0.jpg", + "image_caption": [ + "Figure 3. Pipeline of our DisCo. Firstly, a triplane VAE model is trained to encode shape into triplane latent space (top-left). Subsequently, a triplane diffusion model is trained in this latent space for conditional shape reconstruction (top-right). A novel Hybrid Feature Aggregation Layer is proposed to effectively aggregate and align local features in both partial points cloud and multi-view images (bottom)." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 885, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a PointNet to form a triplane feature map. Followed by a UNet [41], the encoder $\\psi_{enc}$ outputs a normal distribution $\\mathcal{N}(\\mu, \\sigma^2)$ , where $\\mu, \\sigma \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}$ . A latent triplane $T_0 \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}$ can be sampled from this distribution. The above process can be summarized as:", + "bbox": [ + 75, + 392, + 470, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT _ {0} \\sim \\mathcal {N} (\\mu , \\sigma^ {2}) \\quad \\mu , \\sigma = \\psi_ {\\text {e n c}} (\\mathbf {P}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 478, + 468, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The occupancy field is chosen as the shape representation. The decoder $\\psi_{dec}$ takes a point $p$ in space and the triplane latent $T_0$ as inputs, yielding the occupancy of this point. Specifically, A UNet model first refines $T_0$ and outputs a triplane feature map. Then, point $p$ is projected onto this triplane feature map, where features are sampled using bilinear interpolation. These sampled features are fed into MLPs which outputs the occupancy prediction $\\tilde{o}$ . The training of the VAE model is supervised by reconstruction loss and KL divergence loss as $\\mathcal{L}_{vae} = \\| \\tilde{o} - o_{gt} \\|_2^2 + \\lambda_{kl} \\mathcal{L}_{kl}$ . In our experiment, $\\lambda_{kl} = 0.025$ . As in [51], the network layers in both $\\psi_{enc}$ and $\\psi_{dec}$ adopt 3d aware convolution for triplane processing.", + "bbox": [ + 75, + 506, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reconstruction is conducted in canonical object space. During training, augmentation such as random shifting, rotating and scaling are applied, so that it will be more robust to inaccurate objects' pose during inference.", + "bbox": [ + 75, + 702, + 468, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Triplane Diffusion for reconstruction", + "text_level": 1, + "bbox": [ + 76, + 771, + 397, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach employs latent triplane diffusion to reconstruct shapes based on partial point clouds and multi-view images as conditions. Specifically, a 2D UNet model serves as a denoise function $D(\\cdot)$ . This model comprises cascades of residual convolutional block and Hybrid Feature Aggregation Layer with 3d aware convolution. we introduce a novel Hybrid Feature Aggregation Layer, designed to fos", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ter effective interaction between local features derived from partial point clouds and multi-view images. We use continuous diffusion steps as in [24]. The forward diffusion process during training is defined as $T_{\\sigma} = T_{0} + n$ , where $n \\sim \\mathcal{N}(0, \\sigma^{2}\\mathbf{I})$ , $\\sigma$ indicates the diffusion steps, and also the deviation of the noise added. The denoise UNet takes a noisy triplane latent $T_{\\sigma}$ as input, and outputs a denoised triplane latent $T_{\\text{denoise}}$ . We follow the diffusion formulation in EDM [24], the objective function during training is:", + "bbox": [ + 496, + 392, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\sigma , n, T _ {0}} \\lambda_ {\\sigma} \\| D (T _ {0} + n, \\sigma , c) - T _ {0} \\| _ {2} ^ {2} \\quad n \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 537, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $c$ denotes the conditional inputs. In our case, they are partial point clouds and multi-view images. $\\lambda_{\\sigma}$ denotes a loss normalization factor. Inference can be conducted through sampling from noise triplane as in EDM [24].", + "bbox": [ + 496, + 566, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training our diffusion model, we implement a strategy to first pretrain on synthetic datasets, then fine-tune on LASA dataset. We leverage ShapeNet [5], ABO [10], and 3D-Future [15] dataset to synthesize partial point cloud and render images by embedding CAD models in HM3D scene[38, 60]. This strategy enhances both the robustness and the performance of our method, enabling more effective real-world object reconstruction.", + "bbox": [ + 496, + 627, + 890, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Hybrid Feature Aggregation Layer", + "text_level": 1, + "bbox": [ + 498, + 756, + 805, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective of the Hybrid Feature Aggregation Layer is to aggregate and align local features from two input modalities to the triplane space. Points-triplane interaction is implemented by projecting points onto the triplane. However, such interaction becomes challenging when fusing images and triplanes. An alternative way is to expand the triplane into a volumetric grid and project the grids to images. This motivates us to propose this Hybrid Feature Aggregation", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20458", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b2962bb6dfcf0089acb16245acb467a1e1eba6736fd59b9b9ed51654a14b5b72.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on shape reconstruction. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. LAS doesn't have mIoU value since it uses a surface occupancy representation." + ], + "table_footnote": [], + "table_body": "
MethodChairSofaTableCabinetBedShelf
IFNet28.9 / 19.8 / 25.266.1 / 3.61 / 31.128.3 / 21.4 / 27.473.8 / 3.20 / 36.664.0 / 2.95 / 28.817.2 / 5.24 / 29.6
LAS-pts- / 13.2 / 22.6- / 4.47 / 22.9- / 19.5 / 23.9- / 4.81 / 22.1- / 4.77 / 25.8- / 4.70 / 26.8
LAS-pts+imgs- / 10.7 / 24.1- / 3.76 / 24.1- / 15.7 / 24.6- / 4.56 / 24.4- / 4.58 / 26.5- / 4.35 / 27.2
3DShape2VecSec32.5 / 6.43 / 25.666.7 / 3.39 / 27.630.9 / 12.9 / 25.368.1 / 4.09 / 23.762.3 / 3.69 / 30.522.3 / 3.79 / 32.0
Ours-pts33.4 / 5.68 / 26.668.8 / 3.21 / 29.038.0 / 10.5 / 32.272.3 / 3.63 / 36.554.0 / 2.97 / 34.123.2 / 3.68 / 36.5
Ours-pts+imgs38.6 / 3.57 / 31.070.7 / 2.88 / 31.641.5 / 6.52 / 36.175.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "bbox": [ + 81, + 142, + 893, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layer, introducing a hybrid representation that facilitates both points-triplane and images-triplane interactions.", + "bbox": [ + 75, + 273, + 468, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As in the bottom part of Fig. 3, inputs of this layer are partial point cloud $\\tilde{P}$ , k posed images $I_{i}, i \\in 1,2\\dots k$ , and triplane feature $\\tilde{T}_l$ from the l-th convolutional layer. It begins with the aggregation of local features from partial points, achieved by projecting them onto the triplane and incorporating a PointNet layer. Image features map is first extracted using pretrained Vision Transformer. Then, the triplane expands into a volumetric grids $G_{l} \\in \\mathbb{R}^{H \\times W \\times L \\times C}$ . The volumetric grids are then projected to k images using their camera poses, and image local features are sampled using bi-linear interpolation.", + "bbox": [ + 75, + 305, + 468, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To fuse features from multiple images, we employ a transformer. Specifically, the voxel feature serves as the query, while the sampled image feature serves as the key and value. The transformer outputs a volumetric grid attended with local image features from multi-view images. Finally, the volumetric grid is flattened back to a triplane by pooling, producing an output triplane feature map. This process ensures the effective aggregation and alignment of local features from diverse input modalities.", + "bbox": [ + 75, + 470, + 470, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Occupancy-guided 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 618, + 431, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D object detection takes scene scans as input and parses the scene objects into 3D bounding boxes. In real-world scenarios, object scans are often incomplete and sparse due to occlusion, inaccurate sensors, and limited views during capture, making objects hard to recognize. To address it, we propose an Occupancy-Guided 3D Object Detection (OccGOD) approach that utilizes shape completeness prior for better scene understanding. We generate scene-level occupancy ground truth from LASA's fully-covered annotations. Specifically, scenes are partitioned into numerous $384 \\times 384 \\times 96$ voxel grids, with a resolution of $4 \\mathrm{~cm}$ . All CAD models are then placed back into the scene, on which the surface points are densely sampled at $2 \\mathrm{~cm}$ intervals. Subsequently, we iterate through each point, marking their corresponding voxel to be occupied.", + "bbox": [ + 75, + 643, + 468, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our methodology follows a simple 'plug-and-play' manner. It is compatible with any detection method based on a 3D", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "structured representation like volumetric grids [42, 49]. For ease of use, we built our OccGOD upon Cagroup3D [49]. In addition to the original backbone and bounding box prediction head, we introduce an occupancy head and augment the bounding box prediction head with another two output parameters for orientation regression. The occupancy heads takes backbone features as inputs, and output the occupancy of each voxel. More details of network design are in the supplemental. To further explore occupancy representations, we concatenate the features from the occupancy head with the backbone's features during ROI pooling for second-stage bounding box prediction. The occupancy head is supervised with a binary cross-entropy loss function with scene-level occupancy labels from LASA.", + "bbox": [ + 496, + 273, + 890, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed OccGOD predicts a complete foreground structure to guide the bounding box detection. By leveraging the complete scene context, it achieves significant gains in detecting occluded and sparse objects compared to baseline methods.", + "bbox": [ + 496, + 486, + 890, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 500, + 577, + 624, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 601, + 676, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For indoor object reconstruction, we employ mean Intersection over Union (mIoU), L2 chamfer distance, and $1\\%$ F-score. We first normalize both results and ground-truth CAD into the interval from -0.5 to 0.5, and compute the above metrics between them. The experiment is conducted over 6 categories merged from 17 categories in LASA dataset.", + "bbox": [ + 496, + 625, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For 3D object detection, we assess mean average precision (mAP) and mean average recall (mAR) with an IoU threshold at 0.5. Evaluations for both object reconstruction and 3D object detection are performed on LASA's test set.", + "bbox": [ + 496, + 731, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Evaluation on Indoor Object Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 801, + 877, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this session, we compare our methods with existing baselines. We choose IFNet [8] and two state-of-the-art diffusion-based methods, LAS Diffusion [64] and 3DShape2VecSet [61], for comparison. We first pretrain all methods on synthetic dataset [5, 10, 15], then finetune them", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20459", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55fedb28b91e3e2955b818b8399779d65cb623cd29c9364bf20d76ab3bd56183.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison between our method and IFNet, 3DShape2Vecset, and LAS." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 890, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on LASA. IFNet and 3DShape2VecSet receive point clouds as inputs. For LAS Diffusion, we extend it into two versions: with point clouds as input (LAS-pts), and with both point clouds and images as input (LAS-pts+imgs). We compare them with our DisCo with two variants: 1) with partial point clouds only (Ours-pts); 2) with both modalities (Ours-pts+imgs). The quantitative and qualitative comparisons are shown in Tab. 2 and Fig. 4. Our method achieves state-of-the-art performance both quantitatively and qualitatively.", + "bbox": [ + 75, + 650, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 800, + 230, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world Performance Boost using LASA We investigate the impact of our LASA dataset on real-world object reconstruction through three training setups: training from scratch on LASA (w/o pretrain), training on synthetic datasets only (w/o finetune), and pretraining on synthetic", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets followed by finetuning on LASA (full). The quantitative comparison is in Tab. 4. The table shows that finetuning on LASA significantly improves real-world reconstruction, with pretraining on synthetic data followed by finetuning on real strategy achieving the best performance.", + "bbox": [ + 496, + 651, + 890, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Hybrid Feature Aggregation Layer We verify the effectiveness of the Hybrid Feature Aggregation (HFA) layer in aggregating and aligning local features from both partial points and multi-view images. We compare it against not using the expanded grids to project onto the images. Specifically, the latter directly projects the triplane's pixels to the images. The quantitative comparison is shown in Tab. 5. The decent improvement of the HFA layer verifies its powerfulness in fusing multi-modal features.", + "bbox": [ + 496, + 729, + 892, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to inaccurate detection We further investigate how inaccurate detection results could affect the re", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20460", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/061539d5c2725aad88f382663e3a43479da59098c0cc1d91d1361ab2b5e80c49.jpg", + "table_caption": [ + "Table 3. Quantitative comparison between the state-of-the-art (CAGroup3D) and our OccGOD. Oriented bounding boxes are predicted. Evaluation metrics are mAP / mAR. Higher mAP / mAR indicates better performance." + ], + "table_footnote": [], + "table_body": "
@IoU > 0.50ChairTableCabinetRefrigeratorShelfBedSinkWasherBathtub
CaGroup3D92.05 / 93.0946.68 / 66.3733.93 / 54.6490.56 / 91.2235.12 / 57.5467.31 / 73.8964.03 / 74.1587.77 / 89.3927.99 / 45.97
Our OccGOD92.57 / 93.2549.17 / 66.7635.05 / 55.9090.45 / 91.8938.15 / 58.1970.78 / 78.3370.30 / 78.7486.27 / 88.6426.56 / 46.77
@IoU > 0.50ToiletOvenDishwasherFireplaceStoolTV MonitorSofaStoveOverall
CAGroup3D50.94 / 68.9978.93 / 81.2892.19 / 92.8628.21 / 45.1970.69 / 80.781.16 / 6.6346.32 / 65.5133.32 / 41.7355.72 / 66.43
Our OccGOD59.41 / 74.6879.54 / 82.1990.96 / 92.7530.22 / 46.1571.61 / 80.780.73 / 5.1947.65 / 64.9835.64 / 41.7357.36 / 67.47
", + "bbox": [ + 80, + 127, + 888, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2834d4ba2707806691ca38501c535bbfd9a105933f71328126862423aa377913.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on different training setups. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + ], + "table_footnote": [], + "table_body": "
StrategyChairSofaTable
w/o pretrain35.6 / 4.35 / 27.466.7 / 3.80 / 26.036.7 / 8.20 / 30.4
w/o finetune25.0 / 10.8 / 23.566.7 / 5.11 / 26.830.9 / 12.5 / 26.1
full38.6 / 3.57/ 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
w/o pretrain72.6 / 3.49 / 34.266.0 / 2.21 / 30.614.8 / 4.76 / 30.5
w/o finetune67.2 / 5.22 / 29.155.3 / 3.38 / 29.023.9 / 4.18 / 30.8
full75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "bbox": [ + 81, + 265, + 468, + 385 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/518eca94d2182b3fb4bff36112dc87420d739112ab0c8ac4f6f1f5ba427ccc91.jpg", + "table_caption": [ + "Table 5. Quantitative comparison between HFA layer and direct triplane projection. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + ], + "table_footnote": [], + "table_body": "
MethodChairSofa
Triplane project38.1 / 3.95 / 30.670.0 / 3.03 / 30.8
HFA layer38.6 / 3.57 / 31.070.7 / 2.88 / 31.6
", + "bbox": [ + 109, + 454, + 439, + 507 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6f4539924f719f50ab6e3463152e8c7b4d6de9dd12ea0d2183ecba0eaf43c056.jpg", + "table_caption": [ + "Table 6. Quantitative comparison between reconstruction using GT 3D bounding boxes and noisy bounding boxes. Evaluation metrics are mIoU / chamfer L2 / F-score respectively." + ], + "table_footnote": [], + "table_body": "
Detection typeChairSofaTable
noisy_bbox38.7 / 3.77 / 31.370.2 / 2.90 / 31.840.6 / 7.42 / 34.4
GT_bbox38.6 / 3.57 / 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
noisy_bbox73.1 / 3.70 / 34.465.5 / 2.64 / 36.224.1 / 3.85 / 36.0
GT_bbox75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "bbox": [ + 83, + 577, + 467, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "construction. An experiment is conducted by randomly rotating between -10 and 10 degrees, scaling between 0.8 and 1.1, and shifting the center between $-10\\%$ and $10\\%$ for each object. The quantitative results are shown in Tab. 6. We observe that, with considerable disturbances on object poses, our method achieves robust accuracy.", + "bbox": [ + 75, + 699, + 468, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of scene-level occupancy to OccGOD Tab. 3 compares the baseline (Cagroup3D) and our proposed OccGOD. Cagroup3D achieves an mAP of 55.72 and an mAR of 66.43 with an IOU threshold of 0.5. Our OccGOD enhances the baseline, improving mAP by 1.64 and mAR by 1.04. Notable increases occurred for larger", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "furniture like tables (+2.42 in AP), toilets (+8.47 in AP and +5.69 in AR), shelves (+3.03 in AP), beds (+3.47 in AP and 4.44 in AR), and sinks (+6.27 in AP and 4.59 in AR).", + "bbox": [ + 496, + 227, + 890, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 285, + 619, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced a new dataset LASA, a Large-scale Aligned Shape Annotation Dataset. In this work, we have illustrated the substantial benefits LASA brings to the community, particularly in the realms of indoor instance-level scene reconstruction and 3D object detection. Empowered by LASA, we propose a novel Diffusion-based CrossModal Shape Reconstruction approach, namely DisCo, and an Occupancy-guided 3D Object Detection method, namely OccGOD. In DisCo, we design a novel Hybrid Feature Aggregation Layer to effectively fuse and align local features from two input modalities - partial point clouds and multiview images. In OccGOD, we leverage the scene-level occupancy labels provided by LASA, to enhance 3D object detection by learning object completeness priors. Extensive experiments demonstrated that, with the support of LASA, both methods achieve state-of-the-art performance in real-world scenarios.", + "bbox": [ + 496, + 310, + 890, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We firmly believe that the large-scale and well-aligned features of LASA present better annotation quality and quantity, laying a foundation for many 3D downstream applications, including 3D understanding and reconstruction.", + "bbox": [ + 496, + 566, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 640, + 687, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The work was supported in part by Guangdong Provincial Outstanding Youth Project No. 2023B1515020055, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, NSFC-62172348, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It is also partly supported by NSFC-61931024 and Shenzhen General Project No. JCYJ20220530143604010.", + "bbox": [ + 496, + 664, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20461", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Himanshu Arora, Saurabh Mishra, Shichong Peng, Ke Li, and Ali Mahdavi-Amiri. Multimodal shape completion via implicit maximum likelihood estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2958-2967, 2022. 2, 3", + "[2] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 4", + "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 2", + "[4] Miguel Angel Bautista, Walter Talbott, Shuangfei Zhai, Nitish Srivastava, and Joshua M Susskind. On the generalization of learning-based 3d reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2180-2189, 2021. 3", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 5, 6", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3", + "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3, 4", + "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6970-6981, 2020. 1, 3, 6", + "[9] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628-644. Springer, 2016. 2, 3", + "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 2, 5, 6", + "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2", + "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868–5877, 2017. 3", + "[13] Mingyue Dong, Linxi Huan, Hanjiang Xiong, Shuhan Shen, and Xianwei Zheng. Shape anchor guided holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21916-21926, 2023. 2, 3", + "[14] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 2", + "[15] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 2, 5, 6", + "[16] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484-499. Springer, 2016. 3, 4", + "[17] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3", + "[18] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2, 3", + "[19] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2, 3, 4", + "[20] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 3", + "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2098-2107, 2020. 3, 4", + "[22] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In 2022 International Conference on Robotics and Automation (ICRA), pages 10632-10640. IEEE, 2022.", + "[23] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Im-" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20462", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "plicit representations for multi-object shape appearance and pose optimization. 2022. 3", + "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 5", + "[25] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 3", + "[26] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1886-1895, 2018. 3", + "[27] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 2, 3", + "[28] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE transactions on pattern analysis and machine intelligence, 45(1):1320-1327, 2022. 2, 3", + "[29] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Cad-estate: Large-scale cad model annotation in rgb videos. arXiv preprint arXiv:2306.09011, 2023. 2, 3, 4", + "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3", + "[31] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 3", + "[32] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 2, 3", + "[33] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 3", + "[34] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 2, 3, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3", + "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1", + "[37] Charles R Qi, Or Litany, Kaiming He, and Leonifdas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2", + "[38] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5", + "[39] Siddhant Ranade, Christoph Lassner, Kai Li, Christian Haene, Shen-Chi Chen, Jean-Charles Bazin, and Sofien Bouaziz. Ssdnerf: Semantic soft decomposition of neural radiance fields. arXiv preprint arXiv:2212.03406, 2022. 2, 3, 4", + "[40] Jason Rock, Tanmay Gupta, Justin Thorsen, JunYoung Gwak, Daeyun Shin, and Derek Hoiem. Completing 3d object shape from one depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2484-2493, 2015. 3", + "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 5", + "[42] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, pages 477-493. Springer, 2022. 2, 6", + "[43] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3", + "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3", + "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20463", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2, 3, 4", + "[47] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. In European Conference on Computer Vision, pages 684-701. Springer, 2022. 2, 3", + "[48] Michal J Tyszkiiewicz, Kevis-Kokitsi Maninis, Stefan Popov, and Vittorio Ferrari. Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with raytraced transformers. In European Conference on Computer Vision, pages 211-228. Springer, 2022. 2, 3", + "[49] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35: 29975-29988, 2022. 2, 6", + "[50] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2, 3", + "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4563-4573, 2023. 3, 4, 5", + "[52] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 281-296. Springer, 2020. 2, 3", + "[53] Yushuang Wu, Zizheng Yan, Ce Chen, Lai Wei, Xiao Li, Guanbin Li, Yihao Li, Shuguang Cui, and Xiaoguang Han. Scoda: Domain adaptive shape completion for real scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17630-17641, 2023. 2, 3, 4", + "[54] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 2, 3", + "[55] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2, 3", + "[56] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12): 2919-2935, 2020. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 3", + "[58] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239–6249, 2022. 1, 2, 3", + "[59] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 3", + "[60] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 5", + "[61] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Trans. Graph., 42(4), 2023. 1, 2, 3, 4, 6", + "[62] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 3", + "[63] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 2", + "[64] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 2, 3, 4, 6" + ], + "bbox": [ + 501, + 92, + 893, + 626 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20464", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_model.json b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3443c3884f4149017abe5f82dfb2e29cb128fecb --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_model.json @@ -0,0 +1,2488 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.13, + 0.834, + 0.177 + ], + "angle": 0, + "content": "LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.202, + 0.796, + 0.223 + ], + "angle": 0, + "content": "Haolin Liu\\(^{1,2*}\\), Chongjie Ye\\(^{1,2*}\\), Yinyu Nie\\(^{3}\\), Yingfan He\\(^{1,2}\\), Xiaoguang Han\\(^{2,1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.224, + 0.625, + 0.24 + ], + "angle": 0, + "content": "\\*equal contribution \\*corresponding author" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.244, + 0.788, + 0.264 + ], + "angle": 0, + "content": "\\(^{1}\\)FNii, CUHKSZ \\(^{2}\\)SSE, CUHKSZ \\(^{3}\\)Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.384, + 0.273, + 0.581, + 0.287 + ], + "angle": 0, + "content": "gap-lab-cuhk-sz.github.io/LASA" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.34, + 0.391, + 0.355 + ], + "angle": 0, + "content": "920 Scenes + 10,412 Unique CADs" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.368, + 0.159, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.43, + 0.159, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.368, + 0.225, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.43, + 0.223, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.368, + 0.286, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.43, + 0.286, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.368, + 0.352, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.43, + 0.352, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.368, + 0.415, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.43, + 0.415, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.507, + 0.408, + 0.523 + ], + "angle": 0, + "content": "Large-scale Aligned CAD Annotations" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.416, + 0.472, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.334, + 0.553, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.411, + 0.552, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.482, + 0.545, + 0.499 + ], + "angle": 0, + "content": "Monocular RGB-D video" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.332, + 0.634, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.411, + 0.633, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.332, + 0.73, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.41, + 0.73, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.464, + 0.73, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.334, + 0.825, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.407, + 0.825, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.464, + 0.825, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.463, + 0.877, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.508, + 0.782, + 0.521 + ], + "angle": 0, + "content": "Reconstruction Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.536, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Figure 1. We introduce LASA, a Large-scale Aligned Shape Annotation Dataset containing 10,412 unique object CAD models aligned with 920 real-world scene scans. Supported by LASA, we achieve state-of-the-art in real-world object reconstruction and 3D object detection." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.577, + 0.314, + 0.593 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Instance shape reconstruction from a 3D scene involves recovering the full geometries of multiple objects at the semantic instance level. Many methods leverage data-driven learning due to the intricacies of scene complexity and significant indoor occlusions. Training these methods often requires a large-scale, high-quality dataset with aligned and paired shape annotations with real-world scans. Existing datasets are either synthetic or misaligned, restricting the performance of data-driven methods on real data. To this end, we introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising 10,412 high-quality CAD annotations aligned with 920 real-world scene scans from ArkitScenes, created manually by professional artists. On this top, we propose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo) method. It is empowered by a hybrid feature aggregation design to fuse multi-modal inputs and recover high-fidelity object geometries (see Fig. 1). Besides, we present an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate that our shape" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.579, + 0.895, + 0.655 + ], + "angle": 0, + "content": "annotations provide scene occupancy clues that can further improve 3D object detection. Supported by LASA, extensive experiments show that our methods achieve state-of-the-art performance in both instance-level scene reconstruction and 3D object detection tasks." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.665, + 0.633, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The widespread use of hand-held RGB-D sensors has facilitated the effortless acquisition of indoor scene scans. However, these scans often suffer from noises and incompleteness due to limitations in sensor accuracy, the complexity of indoor environments, and occlusion among objects. This further limits its applications in scenarios such as VR/AR and 3D industry where a complete and high-quality reconstruction is desired. This shortage absorbs great attention in 3D vision and graphics community, particularly in the realm of indoor instance-level scene reconstruction. In this task, the objective is to reconstruct the shapes of observed objects based on 3D scans or images captured by sensors. Many advances [8, 36, 56, 58, 61] have been seen by leveraging the power of deep learning methods for this task. They" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20454" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "are data-driven and demand a substantial number of paired scene scans and objects' CAD ground truths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.471, + 0.397 + ], + "angle": 0, + "content": "Existing data-driven methods rely on two types of datasets. [1, 7-9, 19, 27, 39, 50, 52, 55-58] utilized synthetic datasets [5, 10, 15] by synthesizing images and scans that mimic real-world distributions. Synthetic data provides CAD models perfectly aligned with input observations (images or scans) though. Models trained on it are vulnerable to domain gaps from the real world [53] that could impair the generalizability. Scan2CAD[2] attempts to bridge this gap by annotating objects' CAD in real-world scene scans. However, their CADs are retrieved from a synthetic database [5] and their poses are manually adjusted to roughly align the object scans. However, retrieved shapes are often unlike the real observed objects, introducing shape misalignment issues. Many works [13, 18, 28, 32, 47, 48] using it as instance shape supervision are potentially biased with inferior alignment. In summary, The absence of a well-aligned real-world dataset barriers the further advancement of the community." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.596 + ], + "angle": 0, + "content": "To surmount this challenge, we present a new dataset LASA, a Large-scale Aligned Shape Annotation dataset that contains 10,412 high-quality instance CAD annotations meticulously crafted by skilled artists. Each CAD shape is designed and placed to precisely align with objects' scans from 920 real-world scene scans in ArkitScene [3]. We deliberately annotate objects in ArkitScene instead of ScanNet[11] because it provides scans obtained from both accurate Laser sensors and hand-held RGB-D sensors. The accurate Laser scan benefits high-quality and aligned manual annotations. Meanwhile, the less accurate scans from hand-held devices support research on reconstruction and completion using consumer-level devices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.901 + ], + "angle": 0, + "content": "A large-scale scan dataset with high-quality instance CAD annotations empowers many downstream applications. We first investigate how LASA benefits indoor instance-level scene reconstruction. Given an indoor scene, images and scene scans can be obtained through scanning. Typically, the initial step involves 3D object detection, after which the partial point clouds and multi-view images of each detected instance are acquired, serving as visual cues for subsequent shape reconstruction. Point cloud provides natural 3D information though, they are often noisy and incomplete. On the other hand, images present rich appearance clues but lack 3D constraints. Inspired by the complementary nature of these two modalities in describing object surfaces, we advocate utilizing both modalities as inputs to fuse their advantages. Recently, diffusion models have shown appealing performance in shape generation [7, 19, 44, 61, 64]. We advance it further for instance reconstruction and propose Diffusion-based Cross-modal Shape Reconstruction, namely DisCo. DisCo is a triplane diffusion model tailored to accommodate multi-modal in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "puts, towards robust real-world object reconstruction, where a hybrid feature aggregation layer is proposed to aggregate and align two input modalities effectively. Supported by the LASA dataset, extensive experiments show that our method achieves state-of-the-art reconstruction performance with real-world inputs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Moreover, the LASA dataset has full annotations covering every instance within each of the 920 real-world scenes. This extensive coverage enables LASA to provide scene-level occupancy labels. We further explore how they can affect indoor scene 3D object detection. 3D Object detection [37, 42, 49] usually comprises a backbone and a detection head. We propose to integrate an occupancy prediction head for occupancy-guided 3D object detection, namely OccGOD. Our experiments demonstrate that incorporating occupancy prediction leads to substantial improvements in detection performance. In summary, our key contributions are four-fold:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.426 + ], + "angle": 0, + "content": "- We introduce a large-scale dataset, LASA. It contains 10,412 manually crafted, high-quality instance CAD annotations geometrically aligned with 920 real-world scene scans." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.428, + 0.892, + 0.486 + ], + "angle": 0, + "content": "- We propose DisCo, a novel diffusion-based method that leverages hybrid representation, effectively interacting with both input partial point cloud and multi-view images, achieving state-of-the-art reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.487, + 0.892, + 0.533 + ], + "angle": 0, + "content": "- With LASA's scene-level annotation, we introduce occupancy-guided 3D object detection (OccGOD) with decent improvements." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.533, + 0.892, + 0.592 + ], + "angle": 0, + "content": "- We strongly believe the large-scale dataset of well-aligned shape annotations can break the bottleneck of current research on 3D indoor scene understanding and reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.367, + 0.892, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.61, + 0.651, + 0.626 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.774, + 0.651 + ], + "angle": 0, + "content": "2.1. Indoor Instance Shape Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Recently, many advances have been seen in learning-based 3D object and scene reconstruction from images and videos. However, numerous challenges persist due to the limitations of existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "These learning-based approaches are usually trained either on existing synthetic datasets or real-world datasets. While synthetic datasets are demonstrated valuable for training models, they lack realism. Large synthetic object collections like ShapeNet[5] and ABO[10] provide diverse 3D models but lack environment context. Synthetic scene datasets such as 3D-Front[14], Replica[45], and Structured3D[63] possess complete synthetic environments with CAD annotations for objects. However, models trained on these datasets often struggle to generalize with real inputs due to substantial domain gaps[53]." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Some works [2, 29, 34, 46, 53, 54] annotate CAD" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20455" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.088, + 0.808, + 0.215 + ], + "angle": 0, + "content": "
DatasetAligned#Scenes#CADsSensor TypeAnnotation Method
Scan2CAD [2]-1,5063,049RGB-DRetrieval
CAD-Estate [29]-19,51212,024RGBRetrieval
IKEA [16]✓null90RGBRetrieval
Pix3D [46]✓null219RGBRetrieval
ScanSalon [53]✓413800RGB-DArtist
Aria's Digital Twin [34]✓2370RGB-DArtist
LASA (Ours)✓92010,412RGB-DArtist
" + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.225, + 0.744, + 0.24 + ], + "angle": 0, + "content": "Table 1. Comparisons with existing 3D indoor datasets with instance shape annotations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.252, + 0.473, + 0.465 + ], + "angle": 0, + "content": "models on real-world data to bridge the domain gaps. Scan2CAD [2] and CAD-estate [29] have provided scanned object-CAD pairs by retrieving them from ShapeNet [5] and are further manually aligned to the real-world scenes. However, these retrieved CAD models lack alignment with real objects, potentially biasing data-driven reconstruction methods towards inferior reconstruction. This motivates us to build a real-world dataset with well-aligned CAD annotations. Aria's Digital Twin [34] and ScanSalon [53] provide such CADs though, the limited quantities restrict their application for data-hungry tasks. Datasets such as Pix3d [46] and PASCAL3D+ [54] supply single-view images with aligned shapes but are limited in scene modalities lacking point cloud and multi-view images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.31, + 0.485 + ], + "angle": 0, + "content": "2.2. 3D Shape Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.81 + ], + "angle": 0, + "content": "Object-level Reconstruction Existing approaches leverage images or partial point clouds as input. Images based reconstruction methods accept either single-view [6, 17, 19, 20, 25, 27, 30, 31, 33, 35, 50, 55, 57, 62, 64] or sparse multiview images [4, 9, 39, 56, 59] as inputs for shape reconstruction. They extract 2D image features and utilize them for shape reconstruction. Some [27, 43] leverage 2D pixel-aligned local features for reconstruction, demonstrating appealing performance for high-quality reconstruction. Other works [1, 8, 12, 26, 40, 52, 58, 61] conduct shape reconstruction from noisy, incomplete partial point clouds. Both paradigms have their advantages, with image-based methods perceiving rich 2D appearance features, while point-based methods process native 3D signals. [21] proposed to combine both inputs to fuse their advantages. Our proposed approach also accepts both inputs, focusing on effectively aggregating and aligning local features from both input sources. Since diffusion model [7, 19, 44, 51, 61, 64] have demonstrated strong capabilities in both shape generation and reconstruction, we opt for a diffusion model for robust and high-quality reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Instance-level Scene Reconstruction Recent instance scene reconstruction methods [13, 18, 21-23, 28, 32, 47, 48] follow a detect-then-reconstruct pipeline from a single scan. First, these methods use a 3D object detector to localize objects in either scene scans or videos. Subsequently, the detected 3D objects are fed into a 3D shape reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.252, + 0.895, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.491, + 0.892, + 0.518 + ], + "angle": 0, + "content": "Figure 2. Visual comparison between aligned and unaligned CAD annotations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.895, + 0.68 + ], + "angle": 0, + "content": "module individually to obtain object shapes. Finally, the reconstructed 3D object shapes are placed back at their original locations to obtain a reconstruction of the full scene. These methods commonly utilize the Scan2CAD [2] dataset for training, where the inferior alignment limits their reconstruction quality. In contrast, the CAD models in our LASA dataset are manually crafted by artists to guarantee alignment, which we hope can lay a foundation for future research in this area." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.645, + 0.723 + ], + "angle": 0, + "content": "3. LASA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.84 + ], + "angle": 0, + "content": "LASA is a large-scale dataset that contains 10,412 unique CAD models covering 920 scenes across 17 categories. Rather than relying on a pre-existing CAD database to annotate scenes, LASA engages professional artists to manually create aligned CAD models with 3D scans. Our annotations provide precise and consistent (as shown in Fig. 2) training data for data-driven reconstruction algorithms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.666, + 0.862 + ], + "angle": 0, + "content": "3.1. Data Annotation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "LASA is built upon the 3D laser scans from ArkitScenes since it is accurate with high resolution, which is critical for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.284, + 0.106 + ], + "angle": 0, + "content": "high-fidelity CAD annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.346 + ], + "angle": 0, + "content": "Data Preprocessing. Each laser scan from ArkitScenes is exceptionally dense with \\(1\\mathrm{GB}+\\) data storage. To improve annotation efficiency, we downsample the scans to a 4mm density. Since ArkitScenes does not publicly provide alignment transformations between the laser scans and the RGB-D scans, we utilize a coarse-to-fine registration method to calculate the transformation matrix (see our supplemental for details). With the transformation matrix, we align the laser data with the RGB-D sequence coordinates from ArkitScenes. We then use ArkitScenes' 3D bounding box annotations to partition the aligned laser point clouds for each single object. These segmented point clouds are transformed into the canonical space. Furthermore, for each object, we select 2-5 frames from RGB-D scans that maximize the 2D projection area of its 3D bounding box. These selected frames serve as references for annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.468, + 0.422 + ], + "angle": 0, + "content": "Shape Annotation. The CAD annotation process involves a team of 35 artists working over 4 months. With preprocessed point clouds and reference images, each artist spent approximately 69 minutes designing a single model. Each model is annotated with Autodesk Maya or Cinema 4D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.468, + 0.499 + ], + "angle": 0, + "content": "Shape Verification. We involve a shape verification procedure for annotation quality control. This procedure has both algorithmic validation and manual reviews to thoroughly evaluate CAD model's accuracy against ground truth scans and images. Our multi-step verification process includes" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.503, + 0.468, + 0.577 + ], + "angle": 0, + "content": "- Senior Review: After initial annotations, 6 senior designers reviewed every CAD model to verify quality, accuracy, and reliability by manually cross-checking against the 3D scans. Any models that failed to meet the standards were flagged for rework." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.579, + 0.468, + 0.651 + ], + "angle": 0, + "content": "- Geometry Alignment: We matched the CAD models to aligned laser scans and calculated the unidirectional Chamfer distance between them. This quantified the raw geometric alignment error for the CAD surface compared to the ground truth scan." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.655, + 0.468, + 0.802 + ], + "angle": 0, + "content": "- View Alignment: We also verified alignment in the pixel level, where we rendered 112,639 images across all scenes by positioning the CAD models in the view frustum of RGB-D sensors. Crowd workers performed a manual inspection to check if the rendered views overlay on the real images. They checked for inconsistencies along object edges and intricate details which would indicate misalignment. This pixel-level evaluation ensured precise alignment. Any rendering mismatches were fixed by re-annotating the CAD model." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.503, + 0.468, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.245, + 0.831 + ], + "angle": 0, + "content": "3.2. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Tab. 1 shows the statistics of LASA compared to the existing datasets. LASA contains a comparable number of unique CAD models to CAD-Estate [29], which was previously the largest scene CAD dataset. However, LASA pro" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.151 + ], + "angle": 0, + "content": "vides better CAD annotation quality against their retrieved CAD models. Additionally, LASA demonstrates greater shape diversity than Scan2CAD [2], with over 3 times as many unique CAD models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.152, + 0.891, + 0.256 + ], + "angle": 0, + "content": "Among all aligned datasets, LASA stands out with a total of 10,412 CADs. This is 13 times more than ScanSalon [53] and 28 times more than Aria's Digital Twin dataset[34]. Unlike IKEA [16] and Pix3D [46] which are annotated on single-view RGB images, LASA captures full RGB-D sequences. This enables a wider range of downstream applications compared to static image datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.258, + 0.89, + 0.318 + ], + "angle": 0, + "content": "Furthermore, we compare LASA with Scan2CAD in terms of alignment quality by measuring the Chamfer distance between CAD annotation and the scene scans, which are 0.161 vs 0.269. LASA provides much better aligned data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.331, + 0.817, + 0.346 + ], + "angle": 0, + "content": "4. Instance-level Scene reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.89, + 0.597 + ], + "angle": 0, + "content": "We propose a Diffusion-based Cross-modal Shape Reconstruction method (DisCo). DisCo is a diffusion-based model to pursue high-fidelity 3D shape reconstruction from partial point clouds and multi-view images. Various representation including latent sets [61], volumetric grids [7, 64], and triplanes [19, 39, 51] are popular choice for diffusion model. We opt for the triplane since its efficiency enables higher output resolution compared to volumetric grids; while reserving 3D structure compared to latent sets. We employ latent triplane diffusion, where a triplane variational auto-encoder (VAE) first encodes shape into triplane latent space (in Sec. 4.1). Subsequently, a triplane diffusion model operates on this latent space for 3D shape reconstruction conditioned on both partial points and multi-view images (in Sec. 4.2). The overall pipeline of DisCo is shown in the upper part of Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.89, + 0.732 + ], + "angle": 0, + "content": "It has been demonstrated that using multiple input modalities (images and scans) presents complementary benefits for semantic scene completion [21]. In DisCo, we also fuse image and point features and introduce Hybrid Feature Aggregation Layer (in Sec. 4.3). In this layer, we utilize a hybrid representation combining a triplane and a volumetric grid. This combination facilitates efficient local feature aggregation and feature alignment from both partial point clouds and multi-view images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.804, + 0.757 + ], + "angle": 0, + "content": "4.1. Triplane Variational Auto-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.89, + 0.901 + ], + "angle": 0, + "content": "To conduct a latent triplane diffusion, the first step is to learn an encoder capable of encoding shapes in a triplane latent space. The Triplane VAE [19] comprises an encoder \\(\\psi_{enc}\\) and a decoder \\(\\psi_{dec}\\). The encoder processes inputs and encodes them into latent space, while the decoder recovers shape from the latent representation. Surface point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{K \\times 3}\\) sampled from a ground truth shape, serves as inputs to \\(\\psi_{enc}\\). \\(K = 20,000\\) in our experiment. These points are projected onto a triplane and subsequently processed by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20457" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.887, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Pipeline of our DisCo. Firstly, a triplane VAE model is trained to encode shape into triplane latent space (top-left). Subsequently, a triplane diffusion model is trained in this latent space for conditional shape reconstruction (top-right). A novel Hybrid Feature Aggregation Layer is proposed to effectively aggregate and align local features in both partial points cloud and multi-view images (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.469 + ], + "angle": 0, + "content": "a PointNet to form a triplane feature map. Followed by a UNet [41], the encoder \\(\\psi_{enc}\\) outputs a normal distribution \\(\\mathcal{N}(\\mu, \\sigma^2)\\), where \\(\\mu, \\sigma \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}\\). A latent triplane \\(T_0 \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}\\) can be sampled from this distribution. The above process can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.479, + 0.47, + 0.497 + ], + "angle": 0, + "content": "\\[\nT _ {0} \\sim \\mathcal {N} (\\mu , \\sigma^ {2}) \\quad \\mu , \\sigma = \\psi_ {\\text {e n c}} (\\mathbf {P}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.702 + ], + "angle": 0, + "content": "The occupancy field is chosen as the shape representation. The decoder \\(\\psi_{dec}\\) takes a point \\(p\\) in space and the triplane latent \\(T_0\\) as inputs, yielding the occupancy of this point. Specifically, A UNet model first refines \\(T_0\\) and outputs a triplane feature map. Then, point \\(p\\) is projected onto this triplane feature map, where features are sampled using bilinear interpolation. These sampled features are fed into MLPs which outputs the occupancy prediction \\(\\tilde{o}\\). The training of the VAE model is supervised by reconstruction loss and KL divergence loss as \\(\\mathcal{L}_{vae} = \\| \\tilde{o} - o_{gt} \\|_2^2 + \\lambda_{kl} \\mathcal{L}_{kl}\\). In our experiment, \\(\\lambda_{kl} = 0.025\\). As in [51], the network layers in both \\(\\psi_{enc}\\) and \\(\\psi_{dec}\\) adopt 3d aware convolution for triplane processing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.763 + ], + "angle": 0, + "content": "The reconstruction is conducted in canonical object space. During training, augmentation such as random shifting, rotating and scaling are applied, so that it will be more robust to inaccurate objects' pose during inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.398, + 0.789 + ], + "angle": 0, + "content": "4.2. Triplane Diffusion for reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our approach employs latent triplane diffusion to reconstruct shapes based on partial point clouds and multi-view images as conditions. Specifically, a 2D UNet model serves as a denoise function \\( D(\\cdot) \\). This model comprises cascades of residual convolutional block and Hybrid Feature Aggregation Layer with 3d aware convolution. we introduce a novel Hybrid Feature Aggregation Layer, designed to fos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.53 + ], + "angle": 0, + "content": "ter effective interaction between local features derived from partial point clouds and multi-view images. We use continuous diffusion steps as in [24]. The forward diffusion process during training is defined as \\( T_{\\sigma} = T_{0} + n \\), where \\( n \\sim \\mathcal{N}(0, \\sigma^{2}\\mathbf{I}) \\), \\( \\sigma \\) indicates the diffusion steps, and also the deviation of the noise added. The denoise UNet takes a noisy triplane latent \\( T_{\\sigma} \\) as input, and outputs a denoised triplane latent \\( T_{\\text{denoise}} \\). We follow the diffusion formulation in EDM [24], the objective function during training is:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.539, + 0.892, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\sigma , n, T _ {0}} \\lambda_ {\\sigma} \\| D (T _ {0} + n, \\sigma , c) - T _ {0} \\| _ {2} ^ {2} \\quad n \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Here \\( c \\) denotes the conditional inputs. In our case, they are partial point clouds and multi-view images. \\( \\lambda_{\\sigma} \\) denotes a loss normalization factor. Inference can be conducted through sampling from noise triplane as in EDM [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.748 + ], + "angle": 0, + "content": "During training our diffusion model, we implement a strategy to first pretrain on synthetic datasets, then fine-tune on LASA dataset. We leverage ShapeNet [5], ABO [10], and 3D-Future [15] dataset to synthesize partial point cloud and render images by embedding CAD models in HM3D scene[38, 60]. This strategy enhances both the robustness and the performance of our method, enabling more effective real-world object reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.807, + 0.774 + ], + "angle": 0, + "content": "4.3. Hybrid Feature Aggregation Layer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The objective of the Hybrid Feature Aggregation Layer is to aggregate and align local features from two input modalities to the triplane space. Points-triplane interaction is implemented by projecting points onto the triplane. However, such interaction becomes challenging when fusing images and triplanes. An alternative way is to expand the triplane into a volumetric grid and project the grids to images. This motivates us to propose this Hybrid Feature Aggregation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20458" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on shape reconstruction. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. LAS doesn't have mIoU value since it uses a surface occupancy representation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.143, + 0.895, + 0.251 + ], + "angle": 0, + "content": "
MethodChairSofaTableCabinetBedShelf
IFNet28.9 / 19.8 / 25.266.1 / 3.61 / 31.128.3 / 21.4 / 27.473.8 / 3.20 / 36.664.0 / 2.95 / 28.817.2 / 5.24 / 29.6
LAS-pts- / 13.2 / 22.6- / 4.47 / 22.9- / 19.5 / 23.9- / 4.81 / 22.1- / 4.77 / 25.8- / 4.70 / 26.8
LAS-pts+imgs- / 10.7 / 24.1- / 3.76 / 24.1- / 15.7 / 24.6- / 4.56 / 24.4- / 4.58 / 26.5- / 4.35 / 27.2
3DShape2VecSec32.5 / 6.43 / 25.666.7 / 3.39 / 27.630.9 / 12.9 / 25.368.1 / 4.09 / 23.762.3 / 3.69 / 30.522.3 / 3.79 / 32.0
Ours-pts33.4 / 5.68 / 26.668.8 / 3.21 / 29.038.0 / 10.5 / 32.272.3 / 3.63 / 36.554.0 / 2.97 / 34.123.2 / 3.68 / 36.5
Ours-pts+imgs38.6 / 3.57 / 31.070.7 / 2.88 / 31.641.5 / 6.52 / 36.175.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Layer, introducing a hybrid representation that facilitates both points-triplane and images-triplane interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.47, + 0.472 + ], + "angle": 0, + "content": "As in the bottom part of Fig. 3, inputs of this layer are partial point cloud \\(\\tilde{P}\\), k posed images \\(I_{i}, i \\in 1,2\\dots k\\), and triplane feature \\(\\tilde{T}_l\\) from the l-th convolutional layer. It begins with the aggregation of local features from partial points, achieved by projecting them onto the triplane and incorporating a PointNet layer. Image features map is first extracted using pretrained Vision Transformer. Then, the triplane expands into a volumetric grids \\(G_{l} \\in \\mathbb{R}^{H \\times W \\times L \\times C}\\). The volumetric grids are then projected to k images using their camera poses, and image local features are sampled using bi-linear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.471, + 0.609 + ], + "angle": 0, + "content": "To fuse features from multiple images, we employ a transformer. Specifically, the voxel feature serves as the query, while the sampled image feature serves as the key and value. The transformer outputs a volumetric grid attended with local image features from multi-view images. Finally, the volumetric grid is flattened back to a triplane by pooling, producing an output triplane feature map. This process ensures the effective aggregation and alignment of local features from diverse input modalities." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.433, + 0.637 + ], + "angle": 0, + "content": "5. Occupancy-guided 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.87 + ], + "angle": 0, + "content": "3D object detection takes scene scans as input and parses the scene objects into 3D bounding boxes. In real-world scenarios, object scans are often incomplete and sparse due to occlusion, inaccurate sensors, and limited views during capture, making objects hard to recognize. To address it, we propose an Occupancy-Guided 3D Object Detection (OccGOD) approach that utilizes shape completeness prior for better scene understanding. We generate scene-level occupancy ground truth from LASA's fully-covered annotations. Specifically, scenes are partitioned into numerous \\(384 \\times 384 \\times 96\\) voxel grids, with a resolution of \\(4 \\mathrm{~cm}\\). All CAD models are then placed back into the scene, on which the surface points are densely sampled at \\(2 \\mathrm{~cm}\\) intervals. Subsequently, we iterate through each point, marking their corresponding voxel to be occupied." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our methodology follows a simple 'plug-and-play' manner. It is compatible with any detection method based on a 3D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.487 + ], + "angle": 0, + "content": "structured representation like volumetric grids [42, 49]. For ease of use, we built our OccGOD upon Cagroup3D [49]. In addition to the original backbone and bounding box prediction head, we introduce an occupancy head and augment the bounding box prediction head with another two output parameters for orientation regression. The occupancy heads takes backbone features as inputs, and output the occupancy of each voxel. More details of network design are in the supplemental. To further explore occupancy representations, we concatenate the features from the occupancy head with the backbone's features during ROI pooling for second-stage bounding box prediction. The occupancy head is supervised with a binary cross-entropy loss function with scene-level occupancy labels from LASA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.562 + ], + "angle": 0, + "content": "Our proposed OccGOD predicts a complete foreground structure to guide the bounding box detection. By leveraging the complete scene context, it achieves significant gains in detecting occluded and sparse objects compared to baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.578, + 0.625, + 0.594 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.677, + 0.619 + ], + "angle": 0, + "content": "6.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.73 + ], + "angle": 0, + "content": "For indoor object reconstruction, we employ mean Intersection over Union (mIoU), L2 chamfer distance, and \\(1\\%\\) F-score. We first normalize both results and ground-truth CAD into the interval from -0.5 to 0.5, and compute the above metrics between them. The experiment is conducted over 6 categories merged from 17 categories in LASA dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.892, + 0.793 + ], + "angle": 0, + "content": "For 3D object detection, we assess mean average precision (mAP) and mean average recall (mAR) with an IoU threshold at 0.5. Evaluations for both object reconstruction and 3D object detection are performed on LASA's test set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.879, + 0.818 + ], + "angle": 0, + "content": "6.2. Evaluation on Indoor Object Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this session, we compare our methods with existing baselines. We choose IFNet [8] and two state-of-the-art diffusion-based methods, LAS Diffusion [64] and 3DShape2VecSet [61], for comparison. We first pretrain all methods on synthetic dataset [5, 10, 15], then finetune them" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20459" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.092, + 0.891, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.611, + 0.761, + 0.625 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between our method and IFNet, 3DShape2Vecset, and LAS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.789 + ], + "angle": 0, + "content": "on LASA. IFNet and 3DShape2VecSet receive point clouds as inputs. For LAS Diffusion, we extend it into two versions: with point clouds as input (LAS-pts), and with both point clouds and images as input (LAS-pts+imgs). We compare them with our DisCo with two variants: 1) with partial point clouds only (Ours-pts); 2) with both modalities (Ours-pts+imgs). The quantitative and qualitative comparisons are shown in Tab. 2 and Fig. 4. Our method achieves state-of-the-art performance both quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.231, + 0.817 + ], + "angle": 0, + "content": "6.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Real-world Performance Boost using LASA We investigate the impact of our LASA dataset on real-world object reconstruction through three training setups: training from scratch on LASA (w/o pretrain), training on synthetic datasets only (w/o finetune), and pretraining on synthetic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.728 + ], + "angle": 0, + "content": "datasets followed by finetuning on LASA (full). The quantitative comparison is in Tab. 4. The table shows that finetuning on LASA significantly improves real-world reconstruction, with pretraining on synthetic data followed by finetuning on real strategy achieving the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.893, + 0.867 + ], + "angle": 0, + "content": "Effectiveness of Hybrid Feature Aggregation Layer We verify the effectiveness of the Hybrid Feature Aggregation (HFA) layer in aggregating and aligning local features from both partial points and multi-view images. We compare it against not using the expanded grids to project onto the images. Specifically, the latter directly projects the triplane's pixels to the images. The quantitative comparison is shown in Tab. 5. The decent improvement of the HFA layer verifies its powerfulness in fusing multi-modal features." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Robustness to inaccurate detection We further investigate how inaccurate detection results could affect the re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20460" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison between the state-of-the-art (CAGroup3D) and our OccGOD. Oriented bounding boxes are predicted. Evaluation metrics are mAP / mAR. Higher mAP / mAR indicates better performance." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.89, + 0.203 + ], + "angle": 0, + "content": "
@IoU > 0.50ChairTableCabinetRefrigeratorShelfBedSinkWasherBathtub
CaGroup3D92.05 / 93.0946.68 / 66.3733.93 / 54.6490.56 / 91.2235.12 / 57.5467.31 / 73.8964.03 / 74.1587.77 / 89.3927.99 / 45.97
Our OccGOD92.57 / 93.2549.17 / 66.7635.05 / 55.9090.45 / 91.8938.15 / 58.1970.78 / 78.3370.30 / 78.7486.27 / 88.6426.56 / 46.77
@IoU > 0.50ToiletOvenDishwasherFireplaceStoolTV MonitorSofaStoveOverall
CAGroup3D50.94 / 68.9978.93 / 81.2892.19 / 92.8628.21 / 45.1970.69 / 80.781.16 / 6.6346.32 / 65.5133.32 / 41.7355.72 / 66.43
Our OccGOD59.41 / 74.6879.54 / 82.1990.96 / 92.7530.22 / 46.1571.61 / 80.780.73 / 5.1947.65 / 64.9835.64 / 41.7357.36 / 67.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.226, + 0.47, + 0.255 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on different training setups. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.266, + 0.469, + 0.386 + ], + "angle": 0, + "content": "
StrategyChairSofaTable
w/o pretrain35.6 / 4.35 / 27.466.7 / 3.80 / 26.036.7 / 8.20 / 30.4
w/o finetune25.0 / 10.8 / 23.566.7 / 5.11 / 26.830.9 / 12.5 / 26.1
full38.6 / 3.57/ 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
w/o pretrain72.6 / 3.49 / 34.266.0 / 2.21 / 30.614.8 / 4.76 / 30.5
w/o finetune67.2 / 5.22 / 29.155.3 / 3.38 / 29.023.9 / 4.18 / 30.8
full75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.446 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison between HFA layer and direct triplane projection. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.455, + 0.44, + 0.508 + ], + "angle": 0, + "content": "
MethodChairSofa
Triplane project38.1 / 3.95 / 30.670.0 / 3.03 / 30.8
HFA layer38.6 / 3.57 / 31.070.7 / 2.88 / 31.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.525, + 0.47, + 0.568 + ], + "angle": 0, + "content": "Table 6. Quantitative comparison between reconstruction using GT 3D bounding boxes and noisy bounding boxes. Evaluation metrics are mIoU / chamfer L2 / F-score respectively." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.578, + 0.468, + 0.672 + ], + "angle": 0, + "content": "
Detection typeChairSofaTable
noisy_bbox38.7 / 3.77 / 31.370.2 / 2.90 / 31.840.6 / 7.42 / 34.4
GT_bbox38.6 / 3.57 / 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
noisy_bbox73.1 / 3.70 / 34.465.5 / 2.64 / 36.224.1 / 3.85 / 36.0
GT_bbox75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.792 + ], + "angle": 0, + "content": "construction. An experiment is conducted by randomly rotating between -10 and 10 degrees, scaling between 0.8 and 1.1, and shifting the center between \\(-10\\%\\) and \\(10\\%\\) for each object. The quantitative results are shown in Tab. 6. We observe that, with considerable disturbances on object poses, our method achieves robust accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of scene-level occupancy to OccGOD Tab. 3 compares the baseline (Cagroup3D) and our proposed OccGOD. Cagroup3D achieves an mAP of 55.72 and an mAR of 66.43 with an IOU threshold of 0.5. Our OccGOD enhances the baseline, improving mAP by 1.64 and mAR by 1.04. Notable increases occurred for larger" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.275 + ], + "angle": 0, + "content": "furniture like tables (+2.42 in AP), toilets (+8.47 in AP and +5.69 in AR), shelves (+3.03 in AP), beds (+3.47 in AP and 4.44 in AR), and sinks (+6.27 in AP and 4.59 in AR)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.286, + 0.62, + 0.302 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.567 + ], + "angle": 0, + "content": "We have introduced a new dataset LASA, a Large-scale Aligned Shape Annotation Dataset. In this work, we have illustrated the substantial benefits LASA brings to the community, particularly in the realms of indoor instance-level scene reconstruction and 3D object detection. Empowered by LASA, we propose a novel Diffusion-based CrossModal Shape Reconstruction approach, namely DisCo, and an Occupancy-guided 3D Object Detection method, namely OccGOD. In DisCo, we design a novel Hybrid Feature Aggregation Layer to effectively fuse and align local features from two input modalities - partial point clouds and multiview images. In OccGOD, we leverage the scene-level occupancy labels provided by LASA, to enhance 3D object detection by learning object completeness priors. Extensive experiments demonstrated that, with the support of LASA, both methods achieve state-of-the-art performance in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.63 + ], + "angle": 0, + "content": "We firmly believe that the large-scale and well-aligned features of LASA present better annotation quality and quantity, laying a foundation for many 3D downstream applications, including 3D understanding and reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.641, + 0.688, + 0.658 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.889 + ], + "angle": 0, + "content": "The work was supported in part by Guangdong Provincial Outstanding Youth Project No. 2023B1515020055, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, NSFC-62172348, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It is also partly supported by NSFC-61931024 and Shenzhen General Project No. JCYJ20220530143604010." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20461" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.185 + ], + "angle": 0, + "content": "[1] Himanshu Arora, Saurabh Mishra, Shichong Peng, Ke Li, and Ali Mahdavi-Amiri. Multimodal shape completion via implicit maximum likelihood estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2958-2967, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.472, + 0.338 + ], + "angle": 0, + "content": "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.473, + 0.41 + ], + "angle": 0, + "content": "[4] Miguel Angel Bautista, Walter Talbott, Shuangfei Zhai, Nitish Srivastava, and Joshua M Susskind. On the generalization of learning-based 3d reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2180-2189, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.472, + 0.48 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6970-6981, 2020. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.678, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[9] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628-644. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868–5877, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.262 + ], + "angle": 0, + "content": "[13] Mingyue Dong, Linxi Huan, Hanjiang Xiong, Shuhan Shen, and Xianwei Zheng. Shape anchor guided holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21916-21926, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[14] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[15] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.894, + 0.489 + ], + "angle": 0, + "content": "[16] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484-499. Springer, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.894, + 0.56 + ], + "angle": 0, + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.894, + 0.629 + ], + "angle": 0, + "content": "[18] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.631, + 0.894, + 0.674 + ], + "angle": 0, + "content": "[19] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.894, + 0.731 + ], + "angle": 0, + "content": "[20] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2098-2107, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[22] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In 2022 International Conference on Robotics and Automation (ICRA), pages 10632-10640. IEEE, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[23] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Im-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20462" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "plicit representations for multi-object shape appearance and pose optimization. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[25] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[26] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1886-1895, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[27] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[28] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE transactions on pattern analysis and machine intelligence, 45(1):1320-1327, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[29] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Cad-estate: Large-scale cad model annotation in rgb videos. arXiv preprint arXiv:2306.09011, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[31] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[32] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[33] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[34] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonifdas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.293, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[38] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[39] Siddhant Ranade, Christoph Lassner, Kai Li, Christian Haene, Shen-Chi Chen, Jean-Charles Bazin, and Sofien Bouaziz. Ssdnerf: Semantic soft decomposition of neural radiance fields. arXiv preprint arXiv:2212.03406, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[40] Jason Rock, Tanmay Gupta, Justin Thorsen, JunYoung Gwak, Daeyun Shin, and Derek Hoiem. Completing 3d object shape from one depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2484-2493, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[42] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, pages 477-493. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[43] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20463" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[46] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[47] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. In European Conference on Computer Vision, pages 684-701. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[48] Michal J Tyszkiiewicz, Kevis-Kokitsi Maninis, Stefan Popov, and Vittorio Ferrari. Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with raytraced transformers. In European Conference on Computer Vision, pages 211-228. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[49] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35: 29975-29988, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[50] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4563-4573, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.472, + 0.616 + ], + "angle": 0, + "content": "[52] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 281-296. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[53] Yushuang Wu, Zizheng Yan, Ce Chen, Lai Wei, Xiao Li, Guanbin Li, Yihao Li, Shuguang Cui, and Xiaoguang Han. Scoda: Domain adaptive shape completion for real scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17630-17641, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[54] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[55] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[56] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12): 2919-2935, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.895, + 0.245 + ], + "angle": 0, + "content": "[58] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239–6249, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.895, + 0.328 + ], + "angle": 0, + "content": "[59] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.332, + 0.893, + 0.386 + ], + "angle": 0, + "content": "[60] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.893, + 0.444 + ], + "angle": 0, + "content": "[61] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Trans. Graph., 42(4), 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.893, + 0.514 + ], + "angle": 0, + "content": "[62] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.893, + 0.571 + ], + "angle": 0, + "content": "[63] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.573, + 0.893, + 0.627 + ], + "angle": 0, + "content": "[64] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20464" + } + ] +] \ No newline at end of file diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_origin.pdf b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..614a355b4a300507578e12d60b67917c9e4fef42 --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/e66a9d83-37a7-45da-8099-3af5a6b6fd8b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a943ac0522afbdb2f018ae233d8bc7c6b83aeaef8ed34b964f22eb99d7a249db +size 4766847 diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/full.md b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6df32e1705f2786c8902d17dcacb8a31f5cafc65 --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/full.md @@ -0,0 +1,339 @@ +# LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset + +Haolin Liu $^{1,2*}$ , Chongjie Ye $^{1,2*}$ , Yinyu Nie $^{3}$ , Yingfan He $^{1,2}$ , Xiaoguang Han $^{2,1\dagger}$ + +\*equal contribution \*corresponding author + +$^{1}$ FNii, CUHKSZ $^{2}$ SSE, CUHKSZ $^{3}$ Technical University of Munich + +gap-lab-cuhk-sz.github.io/LASA + +920 Scenes + 10,412 Unique CADs + +![](images/250f7aba2f0d0ed674bac9d0d817575f4c2c315eb2c28d6bb78d3d454e6867b7.jpg) + +![](images/d2db398d93e153f11ff0a35b8312cdbe2f25278979f027dd91e6f39accf7b466.jpg) +Figure 1. We introduce LASA, a Large-scale Aligned Shape Annotation Dataset containing 10,412 unique object CAD models aligned with 920 real-world scene scans. Supported by LASA, we achieve state-of-the-art in real-world object reconstruction and 3D object detection. + +![](images/e21ddd04138c3c785c1c5b5fdca31b0e0665726e661a24e944ea65317457f7e0.jpg) + +![](images/ffc6a5a8d2e899ace848d26b2e134cb11cf1d3d92cb9a896053a9e6437abacfa.jpg) + +![](images/2e76eb70f3920aa8b231a84824906074c69598e1bd7749a69e443a526c4161ba.jpg) + +![](images/96dabc861db0528e2d36664c51b2c24b6927d4178e72a52169ff8bb3e25f22e0.jpg) + +![](images/3ebe7ab88b8ddf99b334e59a60375e35afd494e0ea99a2c0685dbf8f7b4e2c1d.jpg) + +![](images/f124df7c5e8989ebf26c41de3b0694980d77c6df5cf746c699c0d5293e407705.jpg) + +![](images/5f53da1c19e3f1167d744266c5711a57b0ae598e6a6854823265e3c87e08908c.jpg) + +![](images/8eb88f5dd7c7870c5c60ed4b6c6a877265ef2787e72d788b71998e7086670644.jpg) + +Large-scale Aligned CAD Annotations + +![](images/f4f314354ca6b825ac15f587ae20156c86c8b3844faedf64e6c6aeb09df164a7.jpg) + +![](images/63b8636a0b3f66e55b0a7d7e1a17a452212975dc20ec173641c99db1228e7ae5.jpg) + +![](images/968cdb0a1c96d242eeec51a027813212ffe68957e2d75a0a56ef252ba10da2d6.jpg) +Monocular RGB-D video + +![](images/b1fa87864012d14ec262b6d4a2e33b8c3d9043193a3f874b7e92b464d303ff53.jpg) + +![](images/052bc906431840c2762d7f94457aa1aed39906726dc16b186df3d9cf68b81f94.jpg) + +![](images/d28a971c530ebb099f0d0f03385485d35b6d0989b0b35bc0b5e499b74ee9d797.jpg) + +![](images/0ddd3e77910a700d0f50e7307bdd9865806a2d601d82447a9f699a99ba674d77.jpg) + +![](images/292844475f887dafc5f80d7ee29acb8b9496cfa4e0967f67ce9e69f4a4e2f5c9.jpg) + +![](images/c38814998b2bf490b7f9c3b15c0c32d08919b752ee535456ab8e1b9046b7fdcb.jpg) + +![](images/ac9ab39c2f2507753a599fd65c3cef306f00a0a1ad9a6b9d75f808d669123b5d.jpg) + +![](images/86eca4bfb0549f5ba7db7f2add5682ef78d28f8cc2a712fa60ab639d1d80d3dd.jpg) + +![](images/733dcbe6abad5b4c0f85072a097c9273b4ea1b2ef29978444e5a7cd7cd283b05.jpg) + +Reconstruction Results + +# Abstract + +Instance shape reconstruction from a 3D scene involves recovering the full geometries of multiple objects at the semantic instance level. Many methods leverage data-driven learning due to the intricacies of scene complexity and significant indoor occlusions. Training these methods often requires a large-scale, high-quality dataset with aligned and paired shape annotations with real-world scans. Existing datasets are either synthetic or misaligned, restricting the performance of data-driven methods on real data. To this end, we introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising 10,412 high-quality CAD annotations aligned with 920 real-world scene scans from ArkitScenes, created manually by professional artists. On this top, we propose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo) method. It is empowered by a hybrid feature aggregation design to fuse multi-modal inputs and recover high-fidelity object geometries (see Fig. 1). Besides, we present an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate that our shape + +annotations provide scene occupancy clues that can further improve 3D object detection. Supported by LASA, extensive experiments show that our methods achieve state-of-the-art performance in both instance-level scene reconstruction and 3D object detection tasks. + +# 1. Introduction + +The widespread use of hand-held RGB-D sensors has facilitated the effortless acquisition of indoor scene scans. However, these scans often suffer from noises and incompleteness due to limitations in sensor accuracy, the complexity of indoor environments, and occlusion among objects. This further limits its applications in scenarios such as VR/AR and 3D industry where a complete and high-quality reconstruction is desired. This shortage absorbs great attention in 3D vision and graphics community, particularly in the realm of indoor instance-level scene reconstruction. In this task, the objective is to reconstruct the shapes of observed objects based on 3D scans or images captured by sensors. Many advances [8, 36, 56, 58, 61] have been seen by leveraging the power of deep learning methods for this task. They + +are data-driven and demand a substantial number of paired scene scans and objects' CAD ground truths. + +Existing data-driven methods rely on two types of datasets. [1, 7-9, 19, 27, 39, 50, 52, 55-58] utilized synthetic datasets [5, 10, 15] by synthesizing images and scans that mimic real-world distributions. Synthetic data provides CAD models perfectly aligned with input observations (images or scans) though. Models trained on it are vulnerable to domain gaps from the real world [53] that could impair the generalizability. Scan2CAD[2] attempts to bridge this gap by annotating objects' CAD in real-world scene scans. However, their CADs are retrieved from a synthetic database [5] and their poses are manually adjusted to roughly align the object scans. However, retrieved shapes are often unlike the real observed objects, introducing shape misalignment issues. Many works [13, 18, 28, 32, 47, 48] using it as instance shape supervision are potentially biased with inferior alignment. In summary, The absence of a well-aligned real-world dataset barriers the further advancement of the community. + +To surmount this challenge, we present a new dataset LASA, a Large-scale Aligned Shape Annotation dataset that contains 10,412 high-quality instance CAD annotations meticulously crafted by skilled artists. Each CAD shape is designed and placed to precisely align with objects' scans from 920 real-world scene scans in ArkitScene [3]. We deliberately annotate objects in ArkitScene instead of ScanNet[11] because it provides scans obtained from both accurate Laser sensors and hand-held RGB-D sensors. The accurate Laser scan benefits high-quality and aligned manual annotations. Meanwhile, the less accurate scans from hand-held devices support research on reconstruction and completion using consumer-level devices. + +A large-scale scan dataset with high-quality instance CAD annotations empowers many downstream applications. We first investigate how LASA benefits indoor instance-level scene reconstruction. Given an indoor scene, images and scene scans can be obtained through scanning. Typically, the initial step involves 3D object detection, after which the partial point clouds and multi-view images of each detected instance are acquired, serving as visual cues for subsequent shape reconstruction. Point cloud provides natural 3D information though, they are often noisy and incomplete. On the other hand, images present rich appearance clues but lack 3D constraints. Inspired by the complementary nature of these two modalities in describing object surfaces, we advocate utilizing both modalities as inputs to fuse their advantages. Recently, diffusion models have shown appealing performance in shape generation [7, 19, 44, 61, 64]. We advance it further for instance reconstruction and propose Diffusion-based Cross-modal Shape Reconstruction, namely DisCo. DisCo is a triplane diffusion model tailored to accommodate multi-modal in + +puts, towards robust real-world object reconstruction, where a hybrid feature aggregation layer is proposed to aggregate and align two input modalities effectively. Supported by the LASA dataset, extensive experiments show that our method achieves state-of-the-art reconstruction performance with real-world inputs. + +Moreover, the LASA dataset has full annotations covering every instance within each of the 920 real-world scenes. This extensive coverage enables LASA to provide scene-level occupancy labels. We further explore how they can affect indoor scene 3D object detection. 3D Object detection [37, 42, 49] usually comprises a backbone and a detection head. We propose to integrate an occupancy prediction head for occupancy-guided 3D object detection, namely OccGOD. Our experiments demonstrate that incorporating occupancy prediction leads to substantial improvements in detection performance. In summary, our key contributions are four-fold: + +- We introduce a large-scale dataset, LASA. It contains 10,412 manually crafted, high-quality instance CAD annotations geometrically aligned with 920 real-world scene scans. +- We propose DisCo, a novel diffusion-based method that leverages hybrid representation, effectively interacting with both input partial point cloud and multi-view images, achieving state-of-the-art reconstruction. +- With LASA's scene-level annotation, we introduce occupancy-guided 3D object detection (OccGOD) with decent improvements. +- We strongly believe the large-scale dataset of well-aligned shape annotations can break the bottleneck of current research on 3D indoor scene understanding and reconstruction. + +# 2. Related Works + +# 2.1. Indoor Instance Shape Dataset + +Recently, many advances have been seen in learning-based 3D object and scene reconstruction from images and videos. However, numerous challenges persist due to the limitations of existing datasets. + +These learning-based approaches are usually trained either on existing synthetic datasets or real-world datasets. While synthetic datasets are demonstrated valuable for training models, they lack realism. Large synthetic object collections like ShapeNet[5] and ABO[10] provide diverse 3D models but lack environment context. Synthetic scene datasets such as 3D-Front[14], Replica[45], and Structured3D[63] possess complete synthetic environments with CAD annotations for objects. However, models trained on these datasets often struggle to generalize with real inputs due to substantial domain gaps[53]. + +Some works [2, 29, 34, 46, 53, 54] annotate CAD + +
DatasetAligned#Scenes#CADsSensor TypeAnnotation Method
Scan2CAD [2]-1,5063,049RGB-DRetrieval
CAD-Estate [29]-19,51212,024RGBRetrieval
IKEA [16]✓null90RGBRetrieval
Pix3D [46]✓null219RGBRetrieval
ScanSalon [53]✓413800RGB-DArtist
Aria's Digital Twin [34]✓2370RGB-DArtist
LASA (Ours)✓92010,412RGB-DArtist
+ +Table 1. Comparisons with existing 3D indoor datasets with instance shape annotations + +models on real-world data to bridge the domain gaps. Scan2CAD [2] and CAD-estate [29] have provided scanned object-CAD pairs by retrieving them from ShapeNet [5] and are further manually aligned to the real-world scenes. However, these retrieved CAD models lack alignment with real objects, potentially biasing data-driven reconstruction methods towards inferior reconstruction. This motivates us to build a real-world dataset with well-aligned CAD annotations. Aria's Digital Twin [34] and ScanSalon [53] provide such CADs though, the limited quantities restrict their application for data-hungry tasks. Datasets such as Pix3d [46] and PASCAL3D+ [54] supply single-view images with aligned shapes but are limited in scene modalities lacking point cloud and multi-view images. + +# 2.2. 3D Shape Reconstruction + +Object-level Reconstruction Existing approaches leverage images or partial point clouds as input. Images based reconstruction methods accept either single-view [6, 17, 19, 20, 25, 27, 30, 31, 33, 35, 50, 55, 57, 62, 64] or sparse multiview images [4, 9, 39, 56, 59] as inputs for shape reconstruction. They extract 2D image features and utilize them for shape reconstruction. Some [27, 43] leverage 2D pixel-aligned local features for reconstruction, demonstrating appealing performance for high-quality reconstruction. Other works [1, 8, 12, 26, 40, 52, 58, 61] conduct shape reconstruction from noisy, incomplete partial point clouds. Both paradigms have their advantages, with image-based methods perceiving rich 2D appearance features, while point-based methods process native 3D signals. [21] proposed to combine both inputs to fuse their advantages. Our proposed approach also accepts both inputs, focusing on effectively aggregating and aligning local features from both input sources. Since diffusion model [7, 19, 44, 51, 61, 64] have demonstrated strong capabilities in both shape generation and reconstruction, we opt for a diffusion model for robust and high-quality reconstruction. + +Instance-level Scene Reconstruction Recent instance scene reconstruction methods [13, 18, 21-23, 28, 32, 47, 48] follow a detect-then-reconstruct pipeline from a single scan. First, these methods use a 3D object detector to localize objects in either scene scans or videos. Subsequently, the detected 3D objects are fed into a 3D shape reconstruction + +![](images/a9e8e291f691372d1661d76b3bf441b41a5f83b739e2d63fbde559d6887cebf2.jpg) +Figure 2. Visual comparison between aligned and unaligned CAD annotations + +module individually to obtain object shapes. Finally, the reconstructed 3D object shapes are placed back at their original locations to obtain a reconstruction of the full scene. These methods commonly utilize the Scan2CAD [2] dataset for training, where the inferior alignment limits their reconstruction quality. In contrast, the CAD models in our LASA dataset are manually crafted by artists to guarantee alignment, which we hope can lay a foundation for future research in this area. + +# 3. LASA Dataset + +LASA is a large-scale dataset that contains 10,412 unique CAD models covering 920 scenes across 17 categories. Rather than relying on a pre-existing CAD database to annotate scenes, LASA engages professional artists to manually create aligned CAD models with 3D scans. Our annotations provide precise and consistent (as shown in Fig. 2) training data for data-driven reconstruction algorithms. + +# 3.1. Data Annotation + +LASA is built upon the 3D laser scans from ArkitScenes since it is accurate with high resolution, which is critical for + +high-fidelity CAD annotations. + +Data Preprocessing. Each laser scan from ArkitScenes is exceptionally dense with $1\mathrm{GB}+$ data storage. To improve annotation efficiency, we downsample the scans to a 4mm density. Since ArkitScenes does not publicly provide alignment transformations between the laser scans and the RGB-D scans, we utilize a coarse-to-fine registration method to calculate the transformation matrix (see our supplemental for details). With the transformation matrix, we align the laser data with the RGB-D sequence coordinates from ArkitScenes. We then use ArkitScenes' 3D bounding box annotations to partition the aligned laser point clouds for each single object. These segmented point clouds are transformed into the canonical space. Furthermore, for each object, we select 2-5 frames from RGB-D scans that maximize the 2D projection area of its 3D bounding box. These selected frames serve as references for annotation. + +Shape Annotation. The CAD annotation process involves a team of 35 artists working over 4 months. With preprocessed point clouds and reference images, each artist spent approximately 69 minutes designing a single model. Each model is annotated with Autodesk Maya or Cinema 4D. + +Shape Verification. We involve a shape verification procedure for annotation quality control. This procedure has both algorithmic validation and manual reviews to thoroughly evaluate CAD model's accuracy against ground truth scans and images. Our multi-step verification process includes + +- Senior Review: After initial annotations, 6 senior designers reviewed every CAD model to verify quality, accuracy, and reliability by manually cross-checking against the 3D scans. Any models that failed to meet the standards were flagged for rework. +- Geometry Alignment: We matched the CAD models to aligned laser scans and calculated the unidirectional Chamfer distance between them. This quantified the raw geometric alignment error for the CAD surface compared to the ground truth scan. +- View Alignment: We also verified alignment in the pixel level, where we rendered 112,639 images across all scenes by positioning the CAD models in the view frustum of RGB-D sensors. Crowd workers performed a manual inspection to check if the rendered views overlay on the real images. They checked for inconsistencies along object edges and intricate details which would indicate misalignment. This pixel-level evaluation ensured precise alignment. Any rendering mismatches were fixed by re-annotating the CAD model. + +# 3.2. Dataset Statistics + +Tab. 1 shows the statistics of LASA compared to the existing datasets. LASA contains a comparable number of unique CAD models to CAD-Estate [29], which was previously the largest scene CAD dataset. However, LASA pro + +vides better CAD annotation quality against their retrieved CAD models. Additionally, LASA demonstrates greater shape diversity than Scan2CAD [2], with over 3 times as many unique CAD models. + +Among all aligned datasets, LASA stands out with a total of 10,412 CADs. This is 13 times more than ScanSalon [53] and 28 times more than Aria's Digital Twin dataset[34]. Unlike IKEA [16] and Pix3D [46] which are annotated on single-view RGB images, LASA captures full RGB-D sequences. This enables a wider range of downstream applications compared to static image datasets. + +Furthermore, we compare LASA with Scan2CAD in terms of alignment quality by measuring the Chamfer distance between CAD annotation and the scene scans, which are 0.161 vs 0.269. LASA provides much better aligned data. + +# 4. Instance-level Scene reconstruction + +We propose a Diffusion-based Cross-modal Shape Reconstruction method (DisCo). DisCo is a diffusion-based model to pursue high-fidelity 3D shape reconstruction from partial point clouds and multi-view images. Various representation including latent sets [61], volumetric grids [7, 64], and triplanes [19, 39, 51] are popular choice for diffusion model. We opt for the triplane since its efficiency enables higher output resolution compared to volumetric grids; while reserving 3D structure compared to latent sets. We employ latent triplane diffusion, where a triplane variational auto-encoder (VAE) first encodes shape into triplane latent space (in Sec. 4.1). Subsequently, a triplane diffusion model operates on this latent space for 3D shape reconstruction conditioned on both partial points and multi-view images (in Sec. 4.2). The overall pipeline of DisCo is shown in the upper part of Fig. 3. + +It has been demonstrated that using multiple input modalities (images and scans) presents complementary benefits for semantic scene completion [21]. In DisCo, we also fuse image and point features and introduce Hybrid Feature Aggregation Layer (in Sec. 4.3). In this layer, we utilize a hybrid representation combining a triplane and a volumetric grid. This combination facilitates efficient local feature aggregation and feature alignment from both partial point clouds and multi-view images. + +# 4.1. Triplane Variational Auto-Encoder + +To conduct a latent triplane diffusion, the first step is to learn an encoder capable of encoding shapes in a triplane latent space. The Triplane VAE [19] comprises an encoder $\psi_{enc}$ and a decoder $\psi_{dec}$ . The encoder processes inputs and encodes them into latent space, while the decoder recovers shape from the latent representation. Surface point cloud $\mathbf{P} \in \mathbb{R}^{K \times 3}$ sampled from a ground truth shape, serves as inputs to $\psi_{enc}$ . $K = 20,000$ in our experiment. These points are projected onto a triplane and subsequently processed by + +![](images/0b51d3c5063ba281352084f533b9cf9ce768d10f3b7df4ed39c53f869f3a8be0.jpg) +Figure 3. Pipeline of our DisCo. Firstly, a triplane VAE model is trained to encode shape into triplane latent space (top-left). Subsequently, a triplane diffusion model is trained in this latent space for conditional shape reconstruction (top-right). A novel Hybrid Feature Aggregation Layer is proposed to effectively aggregate and align local features in both partial points cloud and multi-view images (bottom). + +a PointNet to form a triplane feature map. Followed by a UNet [41], the encoder $\psi_{enc}$ outputs a normal distribution $\mathcal{N}(\mu, \sigma^2)$ , where $\mu, \sigma \in \mathbb{R}^{H \times W \times 3 \times C}$ . A latent triplane $T_0 \in \mathbb{R}^{H \times W \times 3 \times C}$ can be sampled from this distribution. The above process can be summarized as: + +$$ +T _ {0} \sim \mathcal {N} (\mu , \sigma^ {2}) \quad \mu , \sigma = \psi_ {\text {e n c}} (\mathbf {P}) \tag {1} +$$ + +The occupancy field is chosen as the shape representation. The decoder $\psi_{dec}$ takes a point $p$ in space and the triplane latent $T_0$ as inputs, yielding the occupancy of this point. Specifically, A UNet model first refines $T_0$ and outputs a triplane feature map. Then, point $p$ is projected onto this triplane feature map, where features are sampled using bilinear interpolation. These sampled features are fed into MLPs which outputs the occupancy prediction $\tilde{o}$ . The training of the VAE model is supervised by reconstruction loss and KL divergence loss as $\mathcal{L}_{vae} = \| \tilde{o} - o_{gt} \|_2^2 + \lambda_{kl} \mathcal{L}_{kl}$ . In our experiment, $\lambda_{kl} = 0.025$ . As in [51], the network layers in both $\psi_{enc}$ and $\psi_{dec}$ adopt 3d aware convolution for triplane processing. + +The reconstruction is conducted in canonical object space. During training, augmentation such as random shifting, rotating and scaling are applied, so that it will be more robust to inaccurate objects' pose during inference. + +# 4.2. Triplane Diffusion for reconstruction + +Our approach employs latent triplane diffusion to reconstruct shapes based on partial point clouds and multi-view images as conditions. Specifically, a 2D UNet model serves as a denoise function $D(\cdot)$ . This model comprises cascades of residual convolutional block and Hybrid Feature Aggregation Layer with 3d aware convolution. we introduce a novel Hybrid Feature Aggregation Layer, designed to fos + +ter effective interaction between local features derived from partial point clouds and multi-view images. We use continuous diffusion steps as in [24]. The forward diffusion process during training is defined as $T_{\sigma} = T_{0} + n$ , where $n \sim \mathcal{N}(0, \sigma^{2}\mathbf{I})$ , $\sigma$ indicates the diffusion steps, and also the deviation of the noise added. The denoise UNet takes a noisy triplane latent $T_{\sigma}$ as input, and outputs a denoised triplane latent $T_{\text{denoise}}$ . We follow the diffusion formulation in EDM [24], the objective function during training is: + +$$ +\mathbb {E} _ {\sigma , n, T _ {0}} \lambda_ {\sigma} \| D (T _ {0} + n, \sigma , c) - T _ {0} \| _ {2} ^ {2} \quad n \sim \mathcal {N} (0, \sigma^ {2} \mathbf {I}) \tag {2} +$$ + +Here $c$ denotes the conditional inputs. In our case, they are partial point clouds and multi-view images. $\lambda_{\sigma}$ denotes a loss normalization factor. Inference can be conducted through sampling from noise triplane as in EDM [24]. + +During training our diffusion model, we implement a strategy to first pretrain on synthetic datasets, then fine-tune on LASA dataset. We leverage ShapeNet [5], ABO [10], and 3D-Future [15] dataset to synthesize partial point cloud and render images by embedding CAD models in HM3D scene[38, 60]. This strategy enhances both the robustness and the performance of our method, enabling more effective real-world object reconstruction. + +# 4.3. Hybrid Feature Aggregation Layer + +The objective of the Hybrid Feature Aggregation Layer is to aggregate and align local features from two input modalities to the triplane space. Points-triplane interaction is implemented by projecting points onto the triplane. However, such interaction becomes challenging when fusing images and triplanes. An alternative way is to expand the triplane into a volumetric grid and project the grids to images. This motivates us to propose this Hybrid Feature Aggregation + +Table 2. Quantitative comparison on shape reconstruction. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. LAS doesn't have mIoU value since it uses a surface occupancy representation. + +
MethodChairSofaTableCabinetBedShelf
IFNet28.9 / 19.8 / 25.266.1 / 3.61 / 31.128.3 / 21.4 / 27.473.8 / 3.20 / 36.664.0 / 2.95 / 28.817.2 / 5.24 / 29.6
LAS-pts- / 13.2 / 22.6- / 4.47 / 22.9- / 19.5 / 23.9- / 4.81 / 22.1- / 4.77 / 25.8- / 4.70 / 26.8
LAS-pts+imgs- / 10.7 / 24.1- / 3.76 / 24.1- / 15.7 / 24.6- / 4.56 / 24.4- / 4.58 / 26.5- / 4.35 / 27.2
3DShape2VecSec32.5 / 6.43 / 25.666.7 / 3.39 / 27.630.9 / 12.9 / 25.368.1 / 4.09 / 23.762.3 / 3.69 / 30.522.3 / 3.79 / 32.0
Ours-pts33.4 / 5.68 / 26.668.8 / 3.21 / 29.038.0 / 10.5 / 32.272.3 / 3.63 / 36.554.0 / 2.97 / 34.123.2 / 3.68 / 36.5
Ours-pts+imgs38.6 / 3.57 / 31.070.7 / 2.88 / 31.641.5 / 6.52 / 36.175.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
+ +Layer, introducing a hybrid representation that facilitates both points-triplane and images-triplane interactions. + +As in the bottom part of Fig. 3, inputs of this layer are partial point cloud $\tilde{P}$ , k posed images $I_{i}, i \in 1,2\dots k$ , and triplane feature $\tilde{T}_l$ from the l-th convolutional layer. It begins with the aggregation of local features from partial points, achieved by projecting them onto the triplane and incorporating a PointNet layer. Image features map is first extracted using pretrained Vision Transformer. Then, the triplane expands into a volumetric grids $G_{l} \in \mathbb{R}^{H \times W \times L \times C}$ . The volumetric grids are then projected to k images using their camera poses, and image local features are sampled using bi-linear interpolation. + +To fuse features from multiple images, we employ a transformer. Specifically, the voxel feature serves as the query, while the sampled image feature serves as the key and value. The transformer outputs a volumetric grid attended with local image features from multi-view images. Finally, the volumetric grid is flattened back to a triplane by pooling, producing an output triplane feature map. This process ensures the effective aggregation and alignment of local features from diverse input modalities. + +# 5. Occupancy-guided 3D Object Detection + +3D object detection takes scene scans as input and parses the scene objects into 3D bounding boxes. In real-world scenarios, object scans are often incomplete and sparse due to occlusion, inaccurate sensors, and limited views during capture, making objects hard to recognize. To address it, we propose an Occupancy-Guided 3D Object Detection (OccGOD) approach that utilizes shape completeness prior for better scene understanding. We generate scene-level occupancy ground truth from LASA's fully-covered annotations. Specifically, scenes are partitioned into numerous $384 \times 384 \times 96$ voxel grids, with a resolution of $4 \mathrm{~cm}$ . All CAD models are then placed back into the scene, on which the surface points are densely sampled at $2 \mathrm{~cm}$ intervals. Subsequently, we iterate through each point, marking their corresponding voxel to be occupied. + +Our methodology follows a simple 'plug-and-play' manner. It is compatible with any detection method based on a 3D + +structured representation like volumetric grids [42, 49]. For ease of use, we built our OccGOD upon Cagroup3D [49]. In addition to the original backbone and bounding box prediction head, we introduce an occupancy head and augment the bounding box prediction head with another two output parameters for orientation regression. The occupancy heads takes backbone features as inputs, and output the occupancy of each voxel. More details of network design are in the supplemental. To further explore occupancy representations, we concatenate the features from the occupancy head with the backbone's features during ROI pooling for second-stage bounding box prediction. The occupancy head is supervised with a binary cross-entropy loss function with scene-level occupancy labels from LASA. + +Our proposed OccGOD predicts a complete foreground structure to guide the bounding box detection. By leveraging the complete scene context, it achieves significant gains in detecting occluded and sparse objects compared to baseline methods. + +# 6. Experiment + +# 6.1. Experiment Setup + +For indoor object reconstruction, we employ mean Intersection over Union (mIoU), L2 chamfer distance, and $1\%$ F-score. We first normalize both results and ground-truth CAD into the interval from -0.5 to 0.5, and compute the above metrics between them. The experiment is conducted over 6 categories merged from 17 categories in LASA dataset. + +For 3D object detection, we assess mean average precision (mAP) and mean average recall (mAR) with an IoU threshold at 0.5. Evaluations for both object reconstruction and 3D object detection are performed on LASA's test set. + +# 6.2. Evaluation on Indoor Object Reconstruction + +In this session, we compare our methods with existing baselines. We choose IFNet [8] and two state-of-the-art diffusion-based methods, LAS Diffusion [64] and 3DShape2VecSet [61], for comparison. We first pretrain all methods on synthetic dataset [5, 10, 15], then finetune them + +![](images/55fedb28b91e3e2955b818b8399779d65cb623cd29c9364bf20d76ab3bd56183.jpg) +Figure 4. Qualitative comparison between our method and IFNet, 3DShape2Vecset, and LAS. + +on LASA. IFNet and 3DShape2VecSet receive point clouds as inputs. For LAS Diffusion, we extend it into two versions: with point clouds as input (LAS-pts), and with both point clouds and images as input (LAS-pts+imgs). We compare them with our DisCo with two variants: 1) with partial point clouds only (Ours-pts); 2) with both modalities (Ours-pts+imgs). The quantitative and qualitative comparisons are shown in Tab. 2 and Fig. 4. Our method achieves state-of-the-art performance both quantitatively and qualitatively. + +# 6.3. Ablation Study + +Real-world Performance Boost using LASA We investigate the impact of our LASA dataset on real-world object reconstruction through three training setups: training from scratch on LASA (w/o pretrain), training on synthetic datasets only (w/o finetune), and pretraining on synthetic + +datasets followed by finetuning on LASA (full). The quantitative comparison is in Tab. 4. The table shows that finetuning on LASA significantly improves real-world reconstruction, with pretraining on synthetic data followed by finetuning on real strategy achieving the best performance. + +Effectiveness of Hybrid Feature Aggregation Layer We verify the effectiveness of the Hybrid Feature Aggregation (HFA) layer in aggregating and aligning local features from both partial points and multi-view images. We compare it against not using the expanded grids to project onto the images. Specifically, the latter directly projects the triplane's pixels to the images. The quantitative comparison is shown in Tab. 5. The decent improvement of the HFA layer verifies its powerfulness in fusing multi-modal features. + +Robustness to inaccurate detection We further investigate how inaccurate detection results could affect the re + +Table 3. Quantitative comparison between the state-of-the-art (CAGroup3D) and our OccGOD. Oriented bounding boxes are predicted. Evaluation metrics are mAP / mAR. Higher mAP / mAR indicates better performance. + +
@IoU > 0.50ChairTableCabinetRefrigeratorShelfBedSinkWasherBathtub
CaGroup3D92.05 / 93.0946.68 / 66.3733.93 / 54.6490.56 / 91.2235.12 / 57.5467.31 / 73.8964.03 / 74.1587.77 / 89.3927.99 / 45.97
Our OccGOD92.57 / 93.2549.17 / 66.7635.05 / 55.9090.45 / 91.8938.15 / 58.1970.78 / 78.3370.30 / 78.7486.27 / 88.6426.56 / 46.77
@IoU > 0.50ToiletOvenDishwasherFireplaceStoolTV MonitorSofaStoveOverall
CAGroup3D50.94 / 68.9978.93 / 81.2892.19 / 92.8628.21 / 45.1970.69 / 80.781.16 / 6.6346.32 / 65.5133.32 / 41.7355.72 / 66.43
Our OccGOD59.41 / 74.6879.54 / 82.1990.96 / 92.7530.22 / 46.1571.61 / 80.780.73 / 5.1947.65 / 64.9835.64 / 41.7357.36 / 67.47
+ +Table 4. Quantitative comparison on different training setups. The evaluation metrics are mIoU / chamfer L2 / F-score respectively. + +
StrategyChairSofaTable
w/o pretrain35.6 / 4.35 / 27.466.7 / 3.80 / 26.036.7 / 8.20 / 30.4
w/o finetune25.0 / 10.8 / 23.566.7 / 5.11 / 26.830.9 / 12.5 / 26.1
full38.6 / 3.57/ 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
w/o pretrain72.6 / 3.49 / 34.266.0 / 2.21 / 30.614.8 / 4.76 / 30.5
w/o finetune67.2 / 5.22 / 29.155.3 / 3.38 / 29.023.9 / 4.18 / 30.8
full75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
+ +Table 5. Quantitative comparison between HFA layer and direct triplane projection. The evaluation metrics are mIoU / chamfer L2 / F-score respectively. + +
MethodChairSofa
Triplane project38.1 / 3.95 / 30.670.0 / 3.03 / 30.8
HFA layer38.6 / 3.57 / 31.070.7 / 2.88 / 31.6
+ +Table 6. Quantitative comparison between reconstruction using GT 3D bounding boxes and noisy bounding boxes. Evaluation metrics are mIoU / chamfer L2 / F-score respectively. + +
Detection typeChairSofaTable
noisy_bbox38.7 / 3.77 / 31.370.2 / 2.90 / 31.840.6 / 7.42 / 34.4
GT_bbox38.6 / 3.57 / 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
noisy_bbox73.1 / 3.70 / 34.465.5 / 2.64 / 36.224.1 / 3.85 / 36.0
GT_bbox75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
+ +construction. An experiment is conducted by randomly rotating between -10 and 10 degrees, scaling between 0.8 and 1.1, and shifting the center between $-10\%$ and $10\%$ for each object. The quantitative results are shown in Tab. 6. We observe that, with considerable disturbances on object poses, our method achieves robust accuracy. + +Effectiveness of scene-level occupancy to OccGOD Tab. 3 compares the baseline (Cagroup3D) and our proposed OccGOD. Cagroup3D achieves an mAP of 55.72 and an mAR of 66.43 with an IOU threshold of 0.5. Our OccGOD enhances the baseline, improving mAP by 1.64 and mAR by 1.04. Notable increases occurred for larger + +furniture like tables (+2.42 in AP), toilets (+8.47 in AP and +5.69 in AR), shelves (+3.03 in AP), beds (+3.47 in AP and 4.44 in AR), and sinks (+6.27 in AP and 4.59 in AR). + +# 7. Conclusion + +We have introduced a new dataset LASA, a Large-scale Aligned Shape Annotation Dataset. In this work, we have illustrated the substantial benefits LASA brings to the community, particularly in the realms of indoor instance-level scene reconstruction and 3D object detection. Empowered by LASA, we propose a novel Diffusion-based CrossModal Shape Reconstruction approach, namely DisCo, and an Occupancy-guided 3D Object Detection method, namely OccGOD. In DisCo, we design a novel Hybrid Feature Aggregation Layer to effectively fuse and align local features from two input modalities - partial point clouds and multiview images. In OccGOD, we leverage the scene-level occupancy labels provided by LASA, to enhance 3D object detection by learning object completeness priors. Extensive experiments demonstrated that, with the support of LASA, both methods achieve state-of-the-art performance in real-world scenarios. + +We firmly believe that the large-scale and well-aligned features of LASA present better annotation quality and quantity, laying a foundation for many 3D downstream applications, including 3D understanding and reconstruction. + +# 8. Acknowledgements + +The work was supported in part by Guangdong Provincial Outstanding Youth Project No. 2023B1515020055, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, NSFC-62172348, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It is also partly supported by NSFC-61931024 and Shenzhen General Project No. JCYJ20220530143604010. + +# References + +[1] Himanshu Arora, Saurabh Mishra, Shichong Peng, Ke Li, and Ali Mahdavi-Amiri. Multimodal shape completion via implicit maximum likelihood estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2958-2967, 2022. 2, 3 +[2] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 4 +[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 2 +[4] Miguel Angel Bautista, Walter Talbott, Shuangfei Zhai, Nitish Srivastava, and Joshua M Susskind. On the generalization of learning-based 3d reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2180-2189, 2021. 3 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 5, 6 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3 +[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3, 4 +[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6970-6981, 2020. 1, 3, 6 +[9] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628-644. Springer, 2016. 2, 3 +[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 2, 5, 6 +[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In + +Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2 +[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868–5877, 2017. 3 +[13] Mingyue Dong, Linxi Huan, Hanjiang Xiong, Shuhan Shen, and Xianwei Zheng. Shape anchor guided holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21916-21926, 2023. 2, 3 +[14] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 2 +[15] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 2, 5, 6 +[16] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484-499. Springer, 2016. 3, 4 +[17] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3 +[18] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2, 3 +[19] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2, 3, 4 +[20] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 3 +[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2098-2107, 2020. 3, 4 +[22] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In 2022 International Conference on Robotics and Automation (ICRA), pages 10632-10640. IEEE, 2022. +[23] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Im- + +plicit representations for multi-object shape appearance and pose optimization. 2022. 3 +[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 5 +[25] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 3 +[26] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1886-1895, 2018. 3 +[27] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 2, 3 +[28] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE transactions on pattern analysis and machine intelligence, 45(1):1320-1327, 2022. 2, 3 +[29] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Cad-estate: Large-scale cad model annotation in rgb videos. arXiv preprint arXiv:2306.09011, 2023. 2, 3, 4 +[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3 +[31] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 3 +[32] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 2, 3 +[33] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 3 +[34] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 2, 3, 4 + +[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3 +[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1 +[37] Charles R Qi, Or Litany, Kaiming He, and Leonifdas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2 +[38] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5 +[39] Siddhant Ranade, Christoph Lassner, Kai Li, Christian Haene, Shen-Chi Chen, Jean-Charles Bazin, and Sofien Bouaziz. Ssdnerf: Semantic soft decomposition of neural radiance fields. arXiv preprint arXiv:2212.03406, 2022. 2, 3, 4 +[40] Jason Rock, Tanmay Gupta, Justin Thorsen, JunYoung Gwak, Daeyun Shin, and Derek Hoiem. Completing 3d object shape from one depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2484-2493, 2015. 3 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 5 +[42] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, pages 477-493. Springer, 2022. 2, 6 +[43] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3 +[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3 +[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2 + +[46] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2, 3, 4 +[47] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. In European Conference on Computer Vision, pages 684-701. Springer, 2022. 2, 3 +[48] Michal J Tyszkiiewicz, Kevis-Kokitsi Maninis, Stefan Popov, and Vittorio Ferrari. Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with raytraced transformers. In European Conference on Computer Vision, pages 211-228. Springer, 2022. 2, 3 +[49] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35: 29975-29988, 2022. 2, 6 +[50] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2, 3 +[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4563-4573, 2023. 3, 4, 5 +[52] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 281-296. Springer, 2020. 2, 3 +[53] Yushuang Wu, Zizheng Yan, Ce Chen, Lai Wei, Xiao Li, Guanbin Li, Yihao Li, Shuguang Cui, and Xiaoguang Han. Scoda: Domain adaptive shape completion for real scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17630-17641, 2023. 2, 3, 4 +[54] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 2, 3 +[55] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2, 3 +[56] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12): 2919-2935, 2020. 1, 3 + +[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 3 +[58] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239–6249, 2022. 1, 2, 3 +[59] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 3 +[60] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 5 +[61] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Trans. Graph., 42(4), 2023. 1, 2, 3, 4, 6 +[62] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 3 +[63] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 2 +[64] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 2, 3, 4, 6 \ No newline at end of file diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/images.zip b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e176ab88a29048e85d3d02bae6b8c7f07e92c4a3 --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32b60f2c4106f63d5d2bb9e23a812c9b1c46a2ea349fd2f77c0744724589151f +size 652143 diff --git a/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/layout.json b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1cf52484f8f4f48124e7fe426a91fd1f25b49246 --- /dev/null +++ b/2024/LASA_ Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset/layout.json @@ -0,0 +1,8647 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 102, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 102, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 510, + 140 + ], + "type": "text", + "content": "LASA: Instance Reconstruction from Real Scans using A Large-scale Aligned Shape Annotation Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "content": "Haolin Liu" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "content": ", Chongjie Ye" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "content": ", Yinyu Nie" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "content": ", Yingfan He" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "text", + "content": ", Xiaoguang Han" + }, + { + "bbox": [ + 107, + 159, + 487, + 176 + ], + "type": "inline_equation", + "content": "^{2,1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 177, + 382, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 177, + 382, + 190 + ], + "spans": [ + { + "bbox": [ + 209, + 177, + 382, + 190 + ], + "type": "text", + "content": "\\*equal contribution \\*corresponding author" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "text", + "content": "FNii, CUHKSZ " + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "text", + "content": "SSE, CUHKSZ " + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 109, + 193, + 482, + 209 + ], + "type": "text", + "content": "Technical University of Munich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 235, + 216, + 355, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 235, + 216, + 355, + 227 + ], + "spans": [ + { + "bbox": [ + 235, + 216, + 355, + 227 + ], + "type": "text", + "content": "gap-lab-cuhk-sz.github.io/LASA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 269, + 239, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 269, + 239, + 281 + ], + "spans": [ + { + "bbox": [ + 77, + 269, + 239, + 281 + ], + "type": "text", + "content": "920 Scenes + 10,412 Unique CADs" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 291, + 97, + 339 + ], + "blocks": [ + { + "bbox": [ + 59, + 291, + 97, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 291, + 97, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 291, + 97, + 339 + ], + "type": "image", + "image_path": "250f7aba2f0d0ed674bac9d0d817575f4c2c315eb2c28d6bb78d3d454e6867b7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 340, + 97, + 388 + ], + "blocks": [ + { + "bbox": [ + 59, + 340, + 97, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 340, + 97, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 340, + 97, + 388 + ], + "type": "image", + "image_path": "d2db398d93e153f11ff0a35b8312cdbe2f25278979f027dd91e6f39accf7b466.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 424, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 424, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 545, + 447 + ], + "type": "text", + "content": "Figure 1. We introduce LASA, a Large-scale Aligned Shape Annotation Dataset containing 10,412 unique object CAD models aligned with 920 real-world scene scans. Supported by LASA, we achieve state-of-the-art in real-world object reconstruction and 3D object detection." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 99, + 291, + 137, + 339 + ], + "blocks": [ + { + "bbox": [ + 99, + 291, + 137, + 339 + ], + "lines": [ + { + "bbox": [ + 99, + 291, + 137, + 339 + ], + "spans": [ + { + "bbox": [ + 99, + 291, + 137, + 339 + ], + "type": "image", + "image_path": "e21ddd04138c3c785c1c5b5fdca31b0e0665726e661a24e944ea65317457f7e0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 99, + 340, + 136, + 388 + ], + "blocks": [ + { + "bbox": [ + 99, + 340, + 136, + 388 + ], + "lines": [ + { + "bbox": [ + 99, + 340, + 136, + 388 + ], + "spans": [ + { + "bbox": [ + 99, + 340, + 136, + 388 + ], + "type": "image", + "image_path": "ffc6a5a8d2e899ace848d26b2e134cb11cf1d3d92cb9a896053a9e6437abacfa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 137, + 291, + 175, + 339 + ], + "blocks": [ + { + "bbox": [ + 137, + 291, + 175, + 339 + ], + "lines": [ + { + "bbox": [ + 137, + 291, + 175, + 339 + ], + "spans": [ + { + "bbox": [ + 137, + 291, + 175, + 339 + ], + "type": "image", + "image_path": "2e76eb70f3920aa8b231a84824906074c69598e1bd7749a69e443a526c4161ba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 340, + 175, + 388 + ], + "blocks": [ + { + "bbox": [ + 137, + 340, + 175, + 388 + ], + "lines": [ + { + "bbox": [ + 137, + 340, + 175, + 388 + ], + "spans": [ + { + "bbox": [ + 137, + 340, + 175, + 388 + ], + "type": "image", + "image_path": "96dabc861db0528e2d36664c51b2c24b6927d4178e72a52169ff8bb3e25f22e0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 178, + 291, + 215, + 339 + ], + "blocks": [ + { + "bbox": [ + 178, + 291, + 215, + 339 + ], + "lines": [ + { + "bbox": [ + 178, + 291, + 215, + 339 + ], + "spans": [ + { + "bbox": [ + 178, + 291, + 215, + 339 + ], + "type": "image", + "image_path": "3ebe7ab88b8ddf99b334e59a60375e35afd494e0ea99a2c0685dbf8f7b4e2c1d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 178, + 340, + 215, + 387 + ], + "blocks": [ + { + "bbox": [ + 178, + 340, + 215, + 387 + ], + "lines": [ + { + "bbox": [ + 178, + 340, + 215, + 387 + ], + "spans": [ + { + "bbox": [ + 178, + 340, + 215, + 387 + ], + "type": "image", + "image_path": "f124df7c5e8989ebf26c41de3b0694980d77c6df5cf746c699c0d5293e407705.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 216, + 291, + 253, + 339 + ], + "blocks": [ + { + "bbox": [ + 216, + 291, + 253, + 339 + ], + "lines": [ + { + "bbox": [ + 216, + 291, + 253, + 339 + ], + "spans": [ + { + "bbox": [ + 216, + 291, + 253, + 339 + ], + "type": "image", + "image_path": "5f53da1c19e3f1167d744266c5711a57b0ae598e6a6854823265e3c87e08908c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 216, + 340, + 253, + 387 + ], + "blocks": [ + { + "bbox": [ + 216, + 340, + 253, + 387 + ], + "lines": [ + { + "bbox": [ + 216, + 340, + 253, + 387 + ], + "spans": [ + { + "bbox": [ + 216, + 340, + 253, + 387 + ], + "type": "image", + "image_path": "8eb88f5dd7c7870c5c60ed4b6c6a877265ef2787e72d788b71998e7086670644.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 61, + 401, + 249, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 401, + 249, + 414 + ], + "spans": [ + { + "bbox": [ + 61, + 401, + 249, + 414 + ], + "type": "text", + "content": "Large-scale Aligned CAD Annotations" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 262, + 329, + 288, + 350 + ], + "blocks": [ + { + "bbox": [ + 262, + 329, + 288, + 350 + ], + "lines": [ + { + "bbox": [ + 262, + 329, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 262, + 329, + 288, + 350 + ], + "type": "image", + "image_path": "f4f314354ca6b825ac15f587ae20156c86c8b3844faedf64e6c6aeb09df164a7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 292, + 264, + 338, + 323 + ], + "blocks": [ + { + "bbox": [ + 292, + 264, + 338, + 323 + ], + "lines": [ + { + "bbox": [ + 292, + 264, + 338, + 323 + ], + "spans": [ + { + "bbox": [ + 292, + 264, + 338, + 323 + ], + "type": "image", + "image_path": "63b8636a0b3f66e55b0a7d7e1a17a452212975dc20ec173641c99db1228e7ae5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 294, + 325, + 337, + 380 + ], + "blocks": [ + { + "bbox": [ + 294, + 325, + 337, + 380 + ], + "lines": [ + { + "bbox": [ + 294, + 325, + 337, + 380 + ], + "spans": [ + { + "bbox": [ + 294, + 325, + 337, + 380 + ], + "type": "image", + "image_path": "968cdb0a1c96d242eeec51a027813212ffe68957e2d75a0a56ef252ba10da2d6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 381, + 333, + 395 + ], + "lines": [ + { + "bbox": [ + 298, + 381, + 333, + 395 + ], + "spans": [ + { + "bbox": [ + 298, + 381, + 333, + 395 + ], + "type": "text", + "content": "Monocular RGB-D video" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 343, + 262, + 388, + 323 + ], + "blocks": [ + { + "bbox": [ + 343, + 262, + 388, + 323 + ], + "lines": [ + { + "bbox": [ + 343, + 262, + 388, + 323 + ], + "spans": [ + { + "bbox": [ + 343, + 262, + 388, + 323 + ], + "type": "image", + "image_path": "b1fa87864012d14ec262b6d4a2e33b8c3d9043193a3f874b7e92b464d303ff53.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 343, + 325, + 387, + 401 + ], + "blocks": [ + { + "bbox": [ + 343, + 325, + 387, + 401 + ], + "lines": [ + { + "bbox": [ + 343, + 325, + 387, + 401 + ], + "spans": [ + { + "bbox": [ + 343, + 325, + 387, + 401 + ], + "type": "image", + "image_path": "052bc906431840c2762d7f94457aa1aed39906726dc16b186df3d9cf68b81f94.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 389, + 262, + 446, + 323 + ], + "blocks": [ + { + "bbox": [ + 389, + 262, + 446, + 323 + ], + "lines": [ + { + "bbox": [ + 389, + 262, + 446, + 323 + ], + "spans": [ + { + "bbox": [ + 389, + 262, + 446, + 323 + ], + "type": "image", + "image_path": "d28a971c530ebb099f0d0f03385485d35b6d0989b0b35bc0b5e499b74ee9d797.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 389, + 324, + 446, + 365 + ], + "blocks": [ + { + "bbox": [ + 389, + 324, + 446, + 365 + ], + "lines": [ + { + "bbox": [ + 389, + 324, + 446, + 365 + ], + "spans": [ + { + "bbox": [ + 389, + 324, + 446, + 365 + ], + "type": "image", + "image_path": "0ddd3e77910a700d0f50e7307bdd9865806a2d601d82447a9f699a99ba674d77.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 389, + 367, + 446, + 402 + ], + "blocks": [ + { + "bbox": [ + 389, + 367, + 446, + 402 + ], + "lines": [ + { + "bbox": [ + 389, + 367, + 446, + 402 + ], + "spans": [ + { + "bbox": [ + 389, + 367, + 446, + 402 + ], + "type": "image", + "image_path": "292844475f887dafc5f80d7ee29acb8b9496cfa4e0967f67ce9e69f4a4e2f5c9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 447, + 264, + 504, + 320 + ], + "blocks": [ + { + "bbox": [ + 447, + 264, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 447, + 264, + 504, + 320 + ], + "spans": [ + { + "bbox": [ + 447, + 264, + 504, + 320 + ], + "type": "image", + "image_path": "c38814998b2bf490b7f9c3b15c0c32d08919b752ee535456ab8e1b9046b7fdcb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 447, + 322, + 504, + 364 + ], + "blocks": [ + { + "bbox": [ + 447, + 322, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 447, + 322, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 447, + 322, + 504, + 364 + ], + "type": "image", + "image_path": "ac9ab39c2f2507753a599fd65c3cef306f00a0a1ad9a6b9d75f808d669123b5d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 447, + 367, + 504, + 406 + ], + "blocks": [ + { + "bbox": [ + 447, + 367, + 504, + 406 + ], + "lines": [ + { + "bbox": [ + 447, + 367, + 504, + 406 + ], + "spans": [ + { + "bbox": [ + 447, + 367, + 504, + 406 + ], + "type": "image", + "image_path": "86eca4bfb0549f5ba7db7f2add5682ef78d28f8cc2a712fa60ab639d1d80d3dd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 506, + 366, + 536, + 406 + ], + "blocks": [ + { + "bbox": [ + 506, + 366, + 536, + 406 + ], + "lines": [ + { + "bbox": [ + 506, + 366, + 536, + 406 + ], + "spans": [ + { + "bbox": [ + 506, + 366, + 536, + 406 + ], + "type": "image", + "image_path": "733dcbe6abad5b4c0f85072a097c9273b4ea1b2ef29978444e5a7cd7cd283b05.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 350, + 402, + 478, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 402, + 478, + 412 + ], + "spans": [ + { + "bbox": [ + 350, + 402, + 478, + 412 + ], + "type": "text", + "content": "Reconstruction Results" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "spans": [ + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Instance shape reconstruction from a 3D scene involves recovering the full geometries of multiple objects at the semantic instance level. Many methods leverage data-driven learning due to the intricacies of scene complexity and significant indoor occlusions. Training these methods often requires a large-scale, high-quality dataset with aligned and paired shape annotations with real-world scans. Existing datasets are either synthetic or misaligned, restricting the performance of data-driven methods on real data. To this end, we introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising 10,412 high-quality CAD annotations aligned with 920 real-world scene scans from ArkitScenes, created manually by professional artists. On this top, we propose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo) method. It is empowered by a hybrid feature aggregation design to fuse multi-modal inputs and recover high-fidelity object geometries (see Fig. 1). Besides, we present an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate that our shape" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 305, + 458, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 458, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 547, + 518 + ], + "type": "text", + "content": "annotations provide scene occupancy clues that can further improve 3D object detection. Supported by LASA, extensive experiments show that our methods achieve state-of-the-art performance in both instance-level scene reconstruction and 3D object detection tasks." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 526, + 387, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 387, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 387, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "The widespread use of hand-held RGB-D sensors has facilitated the effortless acquisition of indoor scene scans. However, these scans often suffer from noises and incompleteness due to limitations in sensor accuracy, the complexity of indoor environments, and occlusion among objects. This further limits its applications in scenarios such as VR/AR and 3D industry where a complete and high-quality reconstruction is desired. This shortage absorbs great attention in 3D vision and graphics community, particularly in the realm of indoor instance-level scene reconstruction. In this task, the objective is to reconstruct the shapes of observed objects based on 3D scans or images captured by sensors. Many advances [8, 36, 56, 58, 61] have been seen by leveraging the power of deep learning methods for this task. They" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20454" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "are data-driven and demand a substantial number of paired scene scans and objects' CAD ground truths." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 288, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 288, + 314 + ], + "type": "text", + "content": "Existing data-driven methods rely on two types of datasets. [1, 7-9, 19, 27, 39, 50, 52, 55-58] utilized synthetic datasets [5, 10, 15] by synthesizing images and scans that mimic real-world distributions. Synthetic data provides CAD models perfectly aligned with input observations (images or scans) though. Models trained on it are vulnerable to domain gaps from the real world [53] that could impair the generalizability. Scan2CAD[2] attempts to bridge this gap by annotating objects' CAD in real-world scene scans. However, their CADs are retrieved from a synthetic database [5] and their poses are manually adjusted to roughly align the object scans. However, retrieved shapes are often unlike the real observed objects, introducing shape misalignment issues. Many works [13, 18, 28, 32, 47, 48] using it as instance shape supervision are potentially biased with inferior alignment. In summary, The absence of a well-aligned real-world dataset barriers the further advancement of the community." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 472 + ], + "type": "text", + "content": "To surmount this challenge, we present a new dataset LASA, a Large-scale Aligned Shape Annotation dataset that contains 10,412 high-quality instance CAD annotations meticulously crafted by skilled artists. Each CAD shape is designed and placed to precisely align with objects' scans from 920 real-world scene scans in ArkitScene [3]. We deliberately annotate objects in ArkitScene instead of ScanNet[11] because it provides scans obtained from both accurate Laser sensors and hand-held RGB-D sensors. The accurate Laser scan benefits high-quality and aligned manual annotations. Meanwhile, the less accurate scans from hand-held devices support research on reconstruction and completion using consumer-level devices." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": "A large-scale scan dataset with high-quality instance CAD annotations empowers many downstream applications. We first investigate how LASA benefits indoor instance-level scene reconstruction. Given an indoor scene, images and scene scans can be obtained through scanning. Typically, the initial step involves 3D object detection, after which the partial point clouds and multi-view images of each detected instance are acquired, serving as visual cues for subsequent shape reconstruction. Point cloud provides natural 3D information though, they are often noisy and incomplete. On the other hand, images present rich appearance clues but lack 3D constraints. Inspired by the complementary nature of these two modalities in describing object surfaces, we advocate utilizing both modalities as inputs to fuse their advantages. Recently, diffusion models have shown appealing performance in shape generation [7, 19, 44, 61, 64]. We advance it further for instance reconstruction and propose Diffusion-based Cross-modal Shape Reconstruction, namely DisCo. DisCo is a triplane diffusion model tailored to accommodate multi-modal in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "puts, towards robust real-world object reconstruction, where a hybrid feature aggregation layer is proposed to aggregate and align two input modalities effectively. Supported by the LASA dataset, extensive experiments show that our method achieves state-of-the-art reconstruction performance with real-world inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 144, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 545, + 287 + ], + "type": "text", + "content": "Moreover, the LASA dataset has full annotations covering every instance within each of the 920 real-world scenes. This extensive coverage enables LASA to provide scene-level occupancy labels. We further explore how they can affect indoor scene 3D object detection. 3D Object detection [37, 42, 49] usually comprises a backbone and a detection head. We propose to integrate an occupancy prediction head for occupancy-guided 3D object detection, namely OccGOD. Our experiments demonstrate that incorporating occupancy prediction leads to substantial improvements in detection performance. In summary, our key contributions are four-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 290, + 545, + 468 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 290, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 545, + 337 + ], + "type": "text", + "content": "- We introduce a large-scale dataset, LASA. It contains 10,412 manually crafted, high-quality instance CAD annotations geometrically aligned with 920 real-world scene scans." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 338, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 545, + 384 + ], + "type": "text", + "content": "- We propose DisCo, a novel diffusion-based method that leverages hybrid representation, effectively interacting with both input partial point cloud and multi-view images, achieving state-of-the-art reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 385, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 385, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 306, + 385, + 545, + 422 + ], + "type": "text", + "content": "- With LASA's scene-level annotation, we introduce occupancy-guided 3D object detection (OccGOD) with decent improvements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 422, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 545, + 468 + ], + "type": "text", + "content": "- We strongly believe the large-scale dataset of well-aligned shape annotations can break the bottleneck of current research on 3D indoor scene understanding and reconstruction." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 483, + 398, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 483, + 398, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 398, + 495 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 502, + 473, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 473, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 473, + 515 + ], + "type": "text", + "content": "2.1. Indoor Instance Shape Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": "Recently, many advances have been seen in learning-based 3D object and scene reconstruction from images and videos. However, numerous challenges persist due to the limitations of existing datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "These learning-based approaches are usually trained either on existing synthetic datasets or real-world datasets. While synthetic datasets are demonstrated valuable for training models, they lack realism. Large synthetic object collections like ShapeNet[5] and ABO[10] provide diverse 3D models but lack environment context. Synthetic scene datasets such as 3D-Front[14], Replica[45], and Structured3D[63] possess complete synthetic environments with CAD annotations for objects. However, models trained on these datasets often struggle to generalize with real inputs due to substantial domain gaps[53]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Some works [2, 29, 34, 46, 53, 54] annotate CAD" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20455" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 69, + 494, + 170 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 494, + 170 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 494, + 170 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 494, + 170 + ], + "type": "table", + "html": "
DatasetAligned#Scenes#CADsSensor TypeAnnotation Method
Scan2CAD [2]-1,5063,049RGB-DRetrieval
CAD-Estate [29]-19,51212,024RGBRetrieval
IKEA [16]✓null90RGBRetrieval
Pix3D [46]✓null219RGBRetrieval
ScanSalon [53]✓413800RGB-DArtist
Aria's Digital Twin [34]✓2370RGB-DArtist
LASA (Ours)✓92010,412RGB-DArtist
", + "image_path": "1494b02ad58fea5756601f0600a2e7bc6691b15a5a1c1927e6d5c051e16a2078.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 137, + 178, + 455, + 190 + ], + "lines": [ + { + "bbox": [ + 137, + 178, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 178, + 455, + 190 + ], + "type": "text", + "content": "Table 1. Comparisons with existing 3D indoor datasets with instance shape annotations" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 199, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 199, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 289, + 368 + ], + "type": "text", + "content": "models on real-world data to bridge the domain gaps. Scan2CAD [2] and CAD-estate [29] have provided scanned object-CAD pairs by retrieving them from ShapeNet [5] and are further manually aligned to the real-world scenes. However, these retrieved CAD models lack alignment with real objects, potentially biasing data-driven reconstruction methods towards inferior reconstruction. This motivates us to build a real-world dataset with well-aligned CAD annotations. Aria's Digital Twin [34] and ScanSalon [53] provide such CADs though, the limited quantities restrict their application for data-hungry tasks. Datasets such as Pix3d [46] and PASCAL3D+ [54] supply single-view images with aligned shapes but are limited in scene modalities lacking point cloud and multi-view images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 372, + 189, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 189, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 189, + 384 + ], + "type": "text", + "content": "2.2. 3D Shape Reconstruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 641 + ], + "type": "text", + "content": "Object-level Reconstruction Existing approaches leverage images or partial point clouds as input. Images based reconstruction methods accept either single-view [6, 17, 19, 20, 25, 27, 30, 31, 33, 35, 50, 55, 57, 62, 64] or sparse multiview images [4, 9, 39, 56, 59] as inputs for shape reconstruction. They extract 2D image features and utilize them for shape reconstruction. Some [27, 43] leverage 2D pixel-aligned local features for reconstruction, demonstrating appealing performance for high-quality reconstruction. Other works [1, 8, 12, 26, 40, 52, 58, 61] conduct shape reconstruction from noisy, incomplete partial point clouds. Both paradigms have their advantages, with image-based methods perceiving rich 2D appearance features, while point-based methods process native 3D signals. [21] proposed to combine both inputs to fuse their advantages. Our proposed approach also accepts both inputs, focusing on effectively aggregating and aligning local features from both input sources. Since diffusion model [7, 19, 44, 51, 61, 64] have demonstrated strong capabilities in both shape generation and reconstruction, we opt for a diffusion model for robust and high-quality reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Instance-level Scene Reconstruction Recent instance scene reconstruction methods [13, 18, 21-23, 28, 32, 47, 48] follow a detect-then-reconstruct pipeline from a single scan. First, these methods use a 3D object detector to localize objects in either scene scans or videos. Subsequently, the detected 3D objects are fed into a 3D shape reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 199, + 547, + 380 + ], + "blocks": [ + { + "bbox": [ + 316, + 199, + 547, + 380 + ], + "lines": [ + { + "bbox": [ + 316, + 199, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 547, + 380 + ], + "type": "image", + "image_path": "a9e8e291f691372d1661d76b3bf441b41a5f83b739e2d63fbde559d6887cebf2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 388, + 545, + 410 + ], + "lines": [ + { + "bbox": [ + 306, + 388, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 545, + 410 + ], + "type": "text", + "content": "Figure 2. Visual comparison between aligned and unaligned CAD annotations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 547, + 538 + ], + "type": "text", + "content": "module individually to obtain object shapes. Finally, the reconstructed 3D object shapes are placed back at their original locations to obtain a reconstruction of the full scene. These methods commonly utilize the Scan2CAD [2] dataset for training, where the inferior alignment limits their reconstruction quality. In contrast, the CAD models in our LASA dataset are manually crafted by artists to guarantee alignment, which we hope can lay a foundation for future research in this area." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 560, + 394, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 394, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 394, + 572 + ], + "type": "text", + "content": "3. LASA Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "content": "LASA is a large-scale dataset that contains 10,412 unique CAD models covering 920 scenes across 17 categories. Rather than relying on a pre-existing CAD database to annotate scenes, LASA engages professional artists to manually create aligned CAD models with 3D scans. Our annotations provide precise and consistent (as shown in Fig. 2) training data for data-driven reconstruction algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 407, + 682 + ], + "type": "text", + "content": "3.1. Data Annotation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "LASA is built upon the 3D laser scans from ArkitScenes since it is accurate with high resolution, which is critical for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20456" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 173, + 83 + ], + "type": "text", + "content": "high-fidelity CAD annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "text", + "content": "Data Preprocessing. Each laser scan from ArkitScenes is exceptionally dense with " + }, + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "inline_equation", + "content": "1\\mathrm{GB}+" + }, + { + "bbox": [ + 46, + 84, + 286, + 274 + ], + "type": "text", + "content": " data storage. To improve annotation efficiency, we downsample the scans to a 4mm density. Since ArkitScenes does not publicly provide alignment transformations between the laser scans and the RGB-D scans, we utilize a coarse-to-fine registration method to calculate the transformation matrix (see our supplemental for details). With the transformation matrix, we align the laser data with the RGB-D sequence coordinates from ArkitScenes. We then use ArkitScenes' 3D bounding box annotations to partition the aligned laser point clouds for each single object. These segmented point clouds are transformed into the canonical space. Furthermore, for each object, we select 2-5 frames from RGB-D scans that maximize the 2D projection area of its 3D bounding box. These selected frames serve as references for annotation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 275, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 286, + 334 + ], + "type": "text", + "content": "Shape Annotation. The CAD annotation process involves a team of 35 artists working over 4 months. With preprocessed point clouds and reference images, each artist spent approximately 69 minutes designing a single model. Each model is annotated with Autodesk Maya or Cinema 4D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 286, + 395 + ], + "type": "text", + "content": "Shape Verification. We involve a shape verification procedure for annotation quality control. This procedure has both algorithmic validation and manual reviews to thoroughly evaluate CAD model's accuracy against ground truth scans and images. Our multi-step verification process includes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 398, + 286, + 635 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 398, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 286, + 456 + ], + "type": "text", + "content": "- Senior Review: After initial annotations, 6 senior designers reviewed every CAD model to verify quality, accuracy, and reliability by manually cross-checking against the 3D scans. Any models that failed to meet the standards were flagged for rework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 458, + 286, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 286, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 286, + 515 + ], + "type": "text", + "content": "- Geometry Alignment: We matched the CAD models to aligned laser scans and calculated the unidirectional Chamfer distance between them. This quantified the raw geometric alignment error for the CAD surface compared to the ground truth scan." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 518, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 286, + 635 + ], + "type": "text", + "content": "- View Alignment: We also verified alignment in the pixel level, where we rendered 112,639 images across all scenes by positioning the CAD models in the view frustum of RGB-D sensors. Crowd workers performed a manual inspection to check if the rendered views overlay on the real images. They checked for inconsistencies along object edges and intricate details which would indicate misalignment. This pixel-level evaluation ensured precise alignment. Any rendering mismatches were fixed by re-annotating the CAD model." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 647, + 149, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 149, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 149, + 658 + ], + "type": "text", + "content": "3.2. Dataset Statistics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Tab. 1 shows the statistics of LASA compared to the existing datasets. LASA contains a comparable number of unique CAD models to CAD-Estate [29], which was previously the largest scene CAD dataset. However, LASA pro" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "vides better CAD annotation quality against their retrieved CAD models. Additionally, LASA demonstrates greater shape diversity than Scan2CAD [2], with over 3 times as many unique CAD models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 120, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 202 + ], + "type": "text", + "content": "Among all aligned datasets, LASA stands out with a total of 10,412 CADs. This is 13 times more than ScanSalon [53] and 28 times more than Aria's Digital Twin dataset[34]. Unlike IKEA [16] and Pix3D [46] which are annotated on single-view RGB images, LASA captures full RGB-D sequences. This enables a wider range of downstream applications compared to static image datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 204, + 544, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 544, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 544, + 251 + ], + "type": "text", + "content": "Furthermore, we compare LASA with Scan2CAD in terms of alignment quality by measuring the Chamfer distance between CAD annotation and the scene scans, which are 0.161 vs 0.269. LASA provides much better aligned data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 262, + 500, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 500, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 500, + 274 + ], + "type": "text", + "content": "4. Instance-level Scene reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 544, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 544, + 472 + ], + "type": "text", + "content": "We propose a Diffusion-based Cross-modal Shape Reconstruction method (DisCo). DisCo is a diffusion-based model to pursue high-fidelity 3D shape reconstruction from partial point clouds and multi-view images. Various representation including latent sets [61], volumetric grids [7, 64], and triplanes [19, 39, 51] are popular choice for diffusion model. We opt for the triplane since its efficiency enables higher output resolution compared to volumetric grids; while reserving 3D structure compared to latent sets. We employ latent triplane diffusion, where a triplane variational auto-encoder (VAE) first encodes shape into triplane latent space (in Sec. 4.1). Subsequently, a triplane diffusion model operates on this latent space for 3D shape reconstruction conditioned on both partial points and multi-view images (in Sec. 4.2). The overall pipeline of DisCo is shown in the upper part of Fig. 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 473, + 544, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 544, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 544, + 579 + ], + "type": "text", + "content": "It has been demonstrated that using multiple input modalities (images and scans) presents complementary benefits for semantic scene completion [21]. In DisCo, we also fuse image and point features and introduce Hybrid Feature Aggregation Layer (in Sec. 4.3). In this layer, we utilize a hybrid representation combining a triplane and a volumetric grid. This combination facilitates efficient local feature aggregation and feature alignment from both partial point clouds and multi-view images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 587, + 492, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 492, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 492, + 599 + ], + "type": "text", + "content": "4.1. Triplane Variational Auto-Encoder" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": "To conduct a latent triplane diffusion, the first step is to learn an encoder capable of encoding shapes in a triplane latent space. The Triplane VAE [19] comprises an encoder " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{enc}" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{dec}" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": ". The encoder processes inputs and encodes them into latent space, while the decoder recovers shape from the latent representation. Surface point cloud " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " sampled from a ground truth shape, serves as inputs to " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "\\psi_{enc}" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "inline_equation", + "content": "K = 20,000" + }, + { + "bbox": [ + 304, + 605, + 544, + 713 + ], + "type": "text", + "content": " in our experiment. These points are projected onto a triplane and subsequently processed by" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20457" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 542, + 246 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 542, + 246 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 542, + 246 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 542, + 246 + ], + "type": "image", + "image_path": "0b51d3c5063ba281352084f533b9cf9ce768d10f3b7df4ed39c53f869f3a8be0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Pipeline of our DisCo. Firstly, a triplane VAE model is trained to encode shape into triplane latent space (top-left). Subsequently, a triplane diffusion model is trained in this latent space for conditional shape reconstruction (top-right). A novel Hybrid Feature Aggregation Layer is proposed to effectively aggregate and align local features in both partial points cloud and multi-view images (bottom)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "content": "a PointNet to form a triplane feature map. Followed by a UNet [41], the encoder " + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\psi_{enc}" + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "content": " outputs a normal distribution " + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\sigma^2)" + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "inline_equation", + "content": "\\mu, \\sigma \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}" + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "content": ". A latent triplane " + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "inline_equation", + "content": "T_0 \\in \\mathbb{R}^{H \\times W \\times 3 \\times C}" + }, + { + "bbox": [ + 46, + 311, + 288, + 371 + ], + "type": "text", + "content": " can be sampled from this distribution. The above process can be summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 379, + 287, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 379, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 91, + 379, + 287, + 393 + ], + "type": "interline_equation", + "content": "T _ {0} \\sim \\mathcal {N} (\\mu , \\sigma^ {2}) \\quad \\mu , \\sigma = \\psi_ {\\text {e n c}} (\\mathbf {P}) \\tag {1}", + "image_path": "5d9b9cd64817099c4b26352136eea146226e604efa1f67f345cd868b29acc6bc.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": "The occupancy field is chosen as the shape representation. The decoder " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{dec}" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " takes a point " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " in space and the triplane latent " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " as inputs, yielding the occupancy of this point. Specifically, A UNet model first refines " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " and outputs a triplane feature map. Then, point " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " is projected onto this triplane feature map, where features are sampled using bilinear interpolation. These sampled features are fed into MLPs which outputs the occupancy prediction " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\tilde{o}" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": ". The training of the VAE model is supervised by reconstruction loss and KL divergence loss as " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vae} = \\| \\tilde{o} - o_{gt} \\|_2^2 + \\lambda_{kl} \\mathcal{L}_{kl}" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": ". In our experiment, " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{kl} = 0.025" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": ". As in [51], the network layers in both " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{enc}" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "inline_equation", + "content": "\\psi_{dec}" + }, + { + "bbox": [ + 46, + 401, + 287, + 555 + ], + "type": "text", + "content": " adopt 3d aware convolution for triplane processing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 604 + ], + "type": "text", + "content": "The reconstruction is conducted in canonical object space. During training, augmentation such as random shifting, rotating and scaling are applied, so that it will be more robust to inaccurate objects' pose during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 611, + 243, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 243, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 243, + 624 + ], + "type": "text", + "content": "4.2. Triplane Diffusion for reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Our approach employs latent triplane diffusion to reconstruct shapes based on partial point clouds and multi-view images as conditions. Specifically, a 2D UNet model serves as a denoise function " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". This model comprises cascades of residual convolutional block and Hybrid Feature Aggregation Layer with 3d aware convolution. we introduce a novel Hybrid Feature Aggregation Layer, designed to fos" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": "ter effective interaction between local features derived from partial point clouds and multi-view images. We use continuous diffusion steps as in [24]. The forward diffusion process during training is defined as " + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "inline_equation", + "content": "T_{\\sigma} = T_{0} + n" + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "inline_equation", + "content": "n \\sim \\mathcal{N}(0, \\sigma^{2}\\mathbf{I})" + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": " indicates the diffusion steps, and also the deviation of the noise added. The denoise UNet takes a noisy triplane latent " + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "inline_equation", + "content": "T_{\\sigma}" + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": " as input, and outputs a denoised triplane latent " + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "inline_equation", + "content": "T_{\\text{denoise}}" + }, + { + "bbox": [ + 304, + 311, + 545, + 419 + ], + "type": "text", + "content": ". We follow the diffusion formulation in EDM [24], the objective function during training is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 426, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 426, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 311, + 426, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\sigma , n, T _ {0}} \\lambda_ {\\sigma} \\| D (T _ {0} + n, \\sigma , c) - T _ {0} \\| _ {2} ^ {2} \\quad n \\sim \\mathcal {N} (0, \\sigma^ {2} \\mathbf {I}) \\tag {2}", + "image_path": "c32605804560294f75c1ae15a1d4950d1fcd26db653c17aabefc9093b6fc574c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "text", + "content": " denotes the conditional inputs. In our case, they are partial point clouds and multi-view images. " + }, + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "inline_equation", + "content": "\\lambda_{\\sigma}" + }, + { + "bbox": [ + 304, + 449, + 545, + 496 + ], + "type": "text", + "content": " denotes a loss normalization factor. Inference can be conducted through sampling from noise triplane as in EDM [24]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 497, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 592 + ], + "type": "text", + "content": "During training our diffusion model, we implement a strategy to first pretrain on synthetic datasets, then fine-tune on LASA dataset. We leverage ShapeNet [5], ABO [10], and 3D-Future [15] dataset to synthesize partial point cloud and render images by embedding CAD models in HM3D scene[38, 60]. This strategy enhances both the robustness and the performance of our method, enabling more effective real-world object reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 599, + 493, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 493, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 493, + 613 + ], + "type": "text", + "content": "4.3. Hybrid Feature Aggregation Layer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "The objective of the Hybrid Feature Aggregation Layer is to aggregate and align local features from two input modalities to the triplane space. Points-triplane interaction is implemented by projecting points onto the triplane. However, such interaction becomes challenging when fusing images and triplanes. An alternative way is to expand the triplane into a volumetric grid and project the grids to images. This motivates us to propose this Hybrid Feature Aggregation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20458" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 113, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on shape reconstruction. Evaluation metrics are mIoU / Chamfer L2 / F-score respectively. Higher mIoU and F-score are better while lower Chamfer L2 is better. Chamfer L2 is scaled by 1,000. LAS doesn't have mIoU value since it uses a surface occupancy representation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 113, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 50, + 113, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 50, + 113, + 547, + 198 + ], + "type": "table", + "html": "
MethodChairSofaTableCabinetBedShelf
IFNet28.9 / 19.8 / 25.266.1 / 3.61 / 31.128.3 / 21.4 / 27.473.8 / 3.20 / 36.664.0 / 2.95 / 28.817.2 / 5.24 / 29.6
LAS-pts- / 13.2 / 22.6- / 4.47 / 22.9- / 19.5 / 23.9- / 4.81 / 22.1- / 4.77 / 25.8- / 4.70 / 26.8
LAS-pts+imgs- / 10.7 / 24.1- / 3.76 / 24.1- / 15.7 / 24.6- / 4.56 / 24.4- / 4.58 / 26.5- / 4.35 / 27.2
3DShape2VecSec32.5 / 6.43 / 25.666.7 / 3.39 / 27.630.9 / 12.9 / 25.368.1 / 4.09 / 23.762.3 / 3.69 / 30.522.3 / 3.79 / 32.0
Ours-pts33.4 / 5.68 / 26.668.8 / 3.21 / 29.038.0 / 10.5 / 32.272.3 / 3.63 / 36.554.0 / 2.97 / 34.123.2 / 3.68 / 36.5
Ours-pts+imgs38.6 / 3.57 / 31.070.7 / 2.88 / 31.641.5 / 6.52 / 36.175.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "image_path": "b2962bb6dfcf0089acb16245acb467a1e1eba6736fd59b9b9ed51654a14b5b72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 217, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 240 + ], + "type": "text", + "content": "Layer, introducing a hybrid representation that facilitates both points-triplane and images-triplane interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "content": "As in the bottom part of Fig. 3, inputs of this layer are partial point cloud " + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "content": ", k posed images " + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "inline_equation", + "content": "I_{i}, i \\in 1,2\\dots k" + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "content": ", and triplane feature " + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "inline_equation", + "content": "\\tilde{T}_l" + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "content": " from the l-th convolutional layer. It begins with the aggregation of local features from partial points, achieved by projecting them onto the triplane and incorporating a PointNet layer. Image features map is first extracted using pretrained Vision Transformer. Then, the triplane expands into a volumetric grids " + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "inline_equation", + "content": "G_{l} \\in \\mathbb{R}^{H \\times W \\times L \\times C}" + }, + { + "bbox": [ + 46, + 242, + 287, + 373 + ], + "type": "text", + "content": ". The volumetric grids are then projected to k images using their camera poses, and image local features are sampled using bi-linear interpolation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 288, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 482 + ], + "type": "text", + "content": "To fuse features from multiple images, we employ a transformer. Specifically, the voxel feature serves as the query, while the sampled image feature serves as the key and value. The transformer outputs a volumetric grid attended with local image features from multi-view images. Finally, the volumetric grid is flattened back to a triplane by pooling, producing an output triplane feature map. This process ensures the effective aggregation and alignment of local features from diverse input modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 264, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 264, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 264, + 504 + ], + "type": "text", + "content": "5. Occupancy-guided 3D Object Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": "3D object detection takes scene scans as input and parses the scene objects into 3D bounding boxes. In real-world scenarios, object scans are often incomplete and sparse due to occlusion, inaccurate sensors, and limited views during capture, making objects hard to recognize. To address it, we propose an Occupancy-Guided 3D Object Detection (OccGOD) approach that utilizes shape completeness prior for better scene understanding. We generate scene-level occupancy ground truth from LASA's fully-covered annotations. Specifically, scenes are partitioned into numerous " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "384 \\times 384 \\times 96" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " voxel grids, with a resolution of " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "4 \\mathrm{~cm}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": ". All CAD models are then placed back into the scene, on which the surface points are densely sampled at " + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "inline_equation", + "content": "2 \\mathrm{~cm}" + }, + { + "bbox": [ + 46, + 510, + 287, + 689 + ], + "type": "text", + "content": " intervals. Subsequently, we iterate through each point, marking their corresponding voxel to be occupied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Our methodology follows a simple 'plug-and-play' manner. It is compatible with any detection method based on a 3D" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 217, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 385 + ], + "type": "text", + "content": "structured representation like volumetric grids [42, 49]. For ease of use, we built our OccGOD upon Cagroup3D [49]. In addition to the original backbone and bounding box prediction head, we introduce an occupancy head and augment the bounding box prediction head with another two output parameters for orientation regression. The occupancy heads takes backbone features as inputs, and output the occupancy of each voxel. More details of network design are in the supplemental. To further explore occupancy representations, we concatenate the features from the occupancy head with the backbone's features during ROI pooling for second-stage bounding box prediction. The occupancy head is supervised with a binary cross-entropy loss function with scene-level occupancy labels from LASA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 385, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 445 + ], + "type": "text", + "content": "Our proposed OccGOD predicts a complete foreground structure to guide the bounding box detection. By leveraging the complete scene context, it achieves significant gains in detecting occluded and sparse objects compared to baseline methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 457, + 382, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 457, + 382, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 457, + 382, + 470 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 476, + 414, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 414, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 414, + 490 + ], + "type": "text", + "content": "6.1. Experiment Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 495, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 578 + ], + "type": "text", + "content": "For indoor object reconstruction, we employ mean Intersection over Union (mIoU), L2 chamfer distance, and " + }, + { + "bbox": [ + 304, + 495, + 545, + 578 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 495, + 545, + 578 + ], + "type": "text", + "content": " F-score. We first normalize both results and ground-truth CAD into the interval from -0.5 to 0.5, and compute the above metrics between them. The experiment is conducted over 6 categories merged from 17 categories in LASA dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 628 + ], + "type": "text", + "content": "For 3D object detection, we assess mean average precision (mAP) and mean average recall (mAR) with an IoU threshold at 0.5. Evaluations for both object reconstruction and 3D object detection are performed on LASA's test set." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 537, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 537, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 537, + 647 + ], + "type": "text", + "content": "6.2. Evaluation on Indoor Object Reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this session, we compare our methods with existing baselines. We choose IFNet [8] and two state-of-the-art diffusion-based methods, LAS Diffusion [64] and 3DShape2VecSet [61], for comparison. We first pretrain all methods on synthetic dataset [5, 10, 15], then finetune them" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20459" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 545, + 475 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 545, + 475 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 545, + 475 + ], + "type": "image", + "image_path": "55fedb28b91e3e2955b818b8399779d65cb623cd29c9364bf20d76ab3bd56183.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 483, + 465, + 495 + ], + "lines": [ + { + "bbox": [ + 126, + 483, + 465, + 495 + ], + "spans": [ + { + "bbox": [ + 126, + 483, + 465, + 495 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between our method and IFNet, 3DShape2Vecset, and LAS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 624 + ], + "type": "text", + "content": "on LASA. IFNet and 3DShape2VecSet receive point clouds as inputs. For LAS Diffusion, we extend it into two versions: with point clouds as input (LAS-pts), and with both point clouds and images as input (LAS-pts+imgs). We compare them with our DisCo with two variants: 1) with partial point clouds only (Ours-pts); 2) with both modalities (Ours-pts+imgs). The quantitative and qualitative comparisons are shown in Tab. 2 and Fig. 4. Our method achieves state-of-the-art performance both quantitatively and qualitatively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 141, + 647 + ], + "type": "text", + "content": "6.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Real-world Performance Boost using LASA We investigate the impact of our LASA dataset on real-world object reconstruction through three training setups: training from scratch on LASA (w/o pretrain), training on synthetic datasets only (w/o finetune), and pretraining on synthetic" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 516, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 576 + ], + "type": "text", + "content": "datasets followed by finetuning on LASA (full). The quantitative comparison is in Tab. 4. The table shows that finetuning on LASA significantly improves real-world reconstruction, with pretraining on synthetic data followed by finetuning on real strategy achieving the best performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 578, + 546, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 546, + 686 + ], + "type": "text", + "content": "Effectiveness of Hybrid Feature Aggregation Layer We verify the effectiveness of the Hybrid Feature Aggregation (HFA) layer in aggregating and aligning local features from both partial points and multi-view images. We compare it against not using the expanded grids to project onto the images. Specifically, the latter directly projects the triplane's pixels to the images. The quantitative comparison is shown in Tab. 5. The decent improvement of the HFA layer verifies its powerfulness in fusing multi-modal features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Robustness to inaccurate detection We further investigate how inaccurate detection results could affect the re" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20460" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 544, + 160 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 3. Quantitative comparison between the state-of-the-art (CAGroup3D) and our OccGOD. Oriented bounding boxes are predicted. Evaluation metrics are mAP / mAR. Higher mAP / mAR indicates better performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 544, + 160 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 544, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 544, + 160 + ], + "type": "table", + "html": "
@IoU > 0.50ChairTableCabinetRefrigeratorShelfBedSinkWasherBathtub
CaGroup3D92.05 / 93.0946.68 / 66.3733.93 / 54.6490.56 / 91.2235.12 / 57.5467.31 / 73.8964.03 / 74.1587.77 / 89.3927.99 / 45.97
Our OccGOD92.57 / 93.2549.17 / 66.7635.05 / 55.9090.45 / 91.8938.15 / 58.1970.78 / 78.3370.30 / 78.7486.27 / 88.6426.56 / 46.77
@IoU > 0.50ToiletOvenDishwasherFireplaceStoolTV MonitorSofaStoveOverall
CAGroup3D50.94 / 68.9978.93 / 81.2892.19 / 92.8628.21 / 45.1970.69 / 80.781.16 / 6.6346.32 / 65.5133.32 / 41.7355.72 / 66.43
Our OccGOD59.41 / 74.6879.54 / 82.1990.96 / 92.7530.22 / 46.1571.61 / 80.780.73 / 5.1947.65 / 64.9835.64 / 41.7357.36 / 67.47
", + "image_path": "061539d5c2725aad88f382663e3a43479da59098c0cc1d91d1361ab2b5e80c49.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 210, + 287, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 287, + 201 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on different training setups. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 210, + 287, + 305 + ], + "lines": [ + { + "bbox": [ + 50, + 210, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 50, + 210, + 287, + 305 + ], + "type": "table", + "html": "
StrategyChairSofaTable
w/o pretrain35.6 / 4.35 / 27.466.7 / 3.80 / 26.036.7 / 8.20 / 30.4
w/o finetune25.0 / 10.8 / 23.566.7 / 5.11 / 26.830.9 / 12.5 / 26.1
full38.6 / 3.57/ 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
w/o pretrain72.6 / 3.49 / 34.266.0 / 2.21 / 30.614.8 / 4.76 / 30.5
w/o finetune67.2 / 5.22 / 29.155.3 / 3.38 / 29.023.9 / 4.18 / 30.8
full75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "image_path": "2834d4ba2707806691ca38501c535bbfd9a105933f71328126862423aa377913.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 67, + 360, + 269, + 402 + ], + "blocks": [ + { + "bbox": [ + 47, + 318, + 287, + 353 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 353 + ], + "type": "text", + "content": "Table 5. Quantitative comparison between HFA layer and direct triplane projection. The evaluation metrics are mIoU / chamfer L2 / F-score respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 360, + 269, + 402 + ], + "lines": [ + { + "bbox": [ + 67, + 360, + 269, + 402 + ], + "spans": [ + { + "bbox": [ + 67, + 360, + 269, + 402 + ], + "type": "table", + "html": "
MethodChairSofa
Triplane project38.1 / 3.95 / 30.670.0 / 3.03 / 30.8
HFA layer38.6 / 3.57 / 31.070.7 / 2.88 / 31.6
", + "image_path": "518eca94d2182b3fb4bff36112dc87420d739112ab0c8ac4f6f1f5ba427ccc91.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 51, + 457, + 286, + 532 + ], + "blocks": [ + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "lines": [ + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 287, + 449 + ], + "type": "text", + "content": "Table 6. Quantitative comparison between reconstruction using GT 3D bounding boxes and noisy bounding boxes. Evaluation metrics are mIoU / chamfer L2 / F-score respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 457, + 286, + 532 + ], + "lines": [ + { + "bbox": [ + 51, + 457, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 51, + 457, + 286, + 532 + ], + "type": "table", + "html": "
Detection typeChairSofaTable
noisy_bbox38.7 / 3.77 / 31.370.2 / 2.90 / 31.840.6 / 7.42 / 34.4
GT_bbox38.6 / 3.57 / 31.070.7/ 2.88/ 31.641.5 / 6.52 / 36.1
CabinetBedShelf
noisy_bbox73.1 / 3.70 / 34.465.5 / 2.64 / 36.224.1 / 3.85 / 36.0
GT_bbox75.1 / 3.10 / 37.062.5 / 2.62 / 35.424.5 / 3.45 / 37.5
", + "image_path": "6f4539924f719f50ab6e3463152e8c7b4d6de9dd12ea0d2183ecba0eaf43c056.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "text", + "content": "construction. An experiment is conducted by randomly rotating between -10 and 10 degrees, scaling between 0.8 and 1.1, and shifting the center between " + }, + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "inline_equation", + "content": "-10\\%" + }, + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 554, + 287, + 627 + ], + "type": "text", + "content": " for each object. The quantitative results are shown in Tab. 6. We observe that, with considerable disturbances on object poses, our method achieves robust accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Effectiveness of scene-level occupancy to OccGOD Tab. 3 compares the baseline (Cagroup3D) and our proposed OccGOD. Cagroup3D achieves an mAP of 55.72 and an mAR of 66.43 with an IOU threshold of 0.5. Our OccGOD enhances the baseline, improving mAP by 1.64 and mAR by 1.04. Notable increases occurred for larger" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "content": "furniture like tables (+2.42 in AP), toilets (+8.47 in AP and +5.69 in AR), shelves (+3.03 in AP), beds (+3.47 in AP and 4.44 in AR), and sinks (+6.27 in AP and 4.59 in AR)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 226, + 379, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 226, + 379, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 226, + 379, + 239 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 449 + ], + "type": "text", + "content": "We have introduced a new dataset LASA, a Large-scale Aligned Shape Annotation Dataset. In this work, we have illustrated the substantial benefits LASA brings to the community, particularly in the realms of indoor instance-level scene reconstruction and 3D object detection. Empowered by LASA, we propose a novel Diffusion-based CrossModal Shape Reconstruction approach, namely DisCo, and an Occupancy-guided 3D Object Detection method, namely OccGOD. In DisCo, we design a novel Hybrid Feature Aggregation Layer to effectively fuse and align local features from two input modalities - partial point clouds and multiview images. In OccGOD, we leverage the scene-level occupancy labels provided by LASA, to enhance 3D object detection by learning object completeness priors. Extensive experiments demonstrated that, with the support of LASA, both methods achieve state-of-the-art performance in real-world scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 498 + ], + "type": "text", + "content": "We firmly believe that the large-scale and well-aligned features of LASA present better annotation quality and quantity, laying a foundation for many 3D downstream applications, including 3D understanding and reconstruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 507, + 421, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 421, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 421, + 521 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 526, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 704 + ], + "type": "text", + "content": "The work was supported in part by Guangdong Provincial Outstanding Youth Project No. 2023B1515020055, the Basic Research Project No. HZQB-KCZYZ-2021067 of Hetao Shenzhen-HK S&T Cooperation Zone, NSFC-62172348, the National Key R&D Program of China with grant No. 2018YFB1800800, by Shenzhen Outstanding Talents Training Fund 202002, by Guangdong Research Projects No. 2017ZT07X152 and No. 2019CX01X104, by Key Area R&D Program of Guangdong Province (Grant No. 2018B030338001), by the Guangdong Provincial Key Laboratory of Future Networks of Intelligence (Grant No. 2022B1212010001), and by Shenzhen Key Laboratory of Big Data and Artificial Intelligence (Grant No. ZDSYS201707251409055). It is also partly supported by NSFC-61931024 and Shenzhen General Project No. JCYJ20220530143604010." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20461" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 146 + ], + "type": "text", + "content": "[1] Himanshu Arora, Saurabh Mishra, Shichong Peng, Ke Li, and Ali Mahdavi-Amiri. Multimodal shape completion via implicit maximum likelihood estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2958-2967, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 203 + ], + "type": "text", + "content": "[2] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 203, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 267 + ], + "type": "text", + "content": "[3] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 269, + 289, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 289, + 324 + ], + "type": "text", + "content": "[4] Miguel Angel Bautista, Walter Talbott, Shuangfei Zhai, Nitish Srivastava, and Joshua M Susskind. On the generalization of learning-based 3d reconstruction. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2180-2189, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 380 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 424 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 480 + ], + "type": "text", + "content": "[7] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 481, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 535 + ], + "type": "text", + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6970-6981, 2020. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 288, + 602 + ], + "type": "text", + "content": "[9] Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 628-644. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 679 + ], + "type": "text", + "content": "[10] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[11] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 150 + ], + "type": "text", + "content": "[12] Angela Dai, Charles Ruizhongtai Qi, and Matthias Nießner. Shape completion using 3d-encoder-predictor cnns and shape synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5868–5877, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 207 + ], + "type": "text", + "content": "[13] Mingyue Dong, Linxi Huan, Hanjiang Xiong, Shuhan Shen, and Xianwei Zheng. Shape anchor guided holistic indoor scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21916-21926, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 208, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 274 + ], + "type": "text", + "content": "[14] Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Binqiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10933-10942, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 275, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 547, + 319 + ], + "type": "text", + "content": "[15] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 320, + 547, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 547, + 387 + ], + "type": "text", + "content": "[16] Rohit Girdhar, David F Fouhey, Mikel Rodriguez, and Abhinav Gupta. Learning a predictable and generative vector representation for objects. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VI 14, pages 484-499. Springer, 2016. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 388, + 547, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 443 + ], + "type": "text", + "content": "[17] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 444, + 547, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 547, + 498 + ], + "type": "text", + "content": "[18] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4022-4031, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 499, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 499, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 499, + 547, + 533 + ], + "type": "text", + "content": "[19] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 547, + 578 + ], + "type": "text", + "content": "[20] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 579, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 547, + 624 + ], + "type": "text", + "content": "[21] Ji Hou, Angela Dai, and Matthias Nießner. Revealnet: Seeing behind objects in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2098-2107, 2020. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "type": "text", + "content": "[22] Muhammad Zubair Irshad, Thomas Kollar, Michael Laskey, Kevin Stone, and Zsolt Kira. Centersnap: Single-shot multi-object 3d shape reconstruction and categorical 6d pose and size estimation. In 2022 International Conference on Robotics and Automation (ICRA), pages 10632-10640. IEEE, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "text", + "content": "[23] Muhammad Zubair Irshad, Sergey Zakharov, Rares Ambrus, Thomas Kollar, Zsolt Kira, and Adrien Gaidon. Shapo: Im-" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20462" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "plicit representations for multi-object shape appearance and pose optimization. 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[24] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 205 + ], + "type": "text", + "content": "[25] Andrey Kurenkov, Jingwei Ji, Animesh Garg, Viraj Mehta, JunYoung Gwak, Christopher Choy, and Silvio Savarese. Deformnet: Free-form deformation network for 3d shape reconstruction from a single image. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 858-866. IEEE, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[26] Or Litany, Alex Bronstein, Michael Bronstein, and Ameesh Makadia. Deformable shape completion with graph convolutional autoencoders. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1886-1895, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 262, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 314 + ], + "type": "text", + "content": "[27] Haolin Liu, Yujuan Zheng, Guanying Chen, Shuguang Cui, and Xiaoguang Han. Towards high-fidelity single-view holistic reconstruction of indoor scenes. In European Conference on Computer Vision, pages 429-446. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 317, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 369 + ], + "type": "text", + "content": "[28] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE transactions on pattern analysis and machine intelligence, 45(1):1320-1327, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 414 + ], + "type": "text", + "content": "[29] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Nießner, and Vittorio Ferrari. Cad-estate: Large-scale cad model annotation in rgb videos. arXiv preprint arXiv:2306.09011, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "text", + "content": "[30] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 536 + ], + "type": "text", + "content": "[31] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Jun Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 55-64, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 590 + ], + "type": "text", + "content": "[32] Yinyu Nie, Ji Hou, Xiaoguang Han, and Matthias Nießner. Rfd-net: Point scene understanding by semantic instance reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4608-4618, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 646 + ], + "type": "text", + "content": "[33] Junyi Pan, Xiaoguang Han, Weikai Chen, Jiapeng Tang, and Kui Jia. Deep mesh reconstruction from single rgb images via topology modification networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9964-9973, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[34] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 2, 3, 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[35] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 184 + ], + "type": "text", + "content": "[36] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[37] Charles R Qi, Or Litany, Kaiming He, and Leonifdas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 232, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 232, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 232, + 545, + 319 + ], + "type": "text", + "content": "[38] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[39] Siddhant Ranade, Christoph Lassner, Kai Li, Christian Haene, Shen-Chi Chen, Jean-Charles Bazin, and Sofien Bouaziz. Ssdnerf: Semantic soft decomposition of neural radiance fields. arXiv preprint arXiv:2212.03406, 2022. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 377, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 431 + ], + "type": "text", + "content": "[40] Jason Rock, Tanmay Gupta, Justin Thorsen, JunYoung Gwak, Daeyun Shin, and Derek Hoiem. Completing 3d object shape from one depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2484-2493, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 498 + ], + "type": "text", + "content": "[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[42] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Computer Vision, pages 477-493. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 600 + ], + "type": "text", + "content": "[43] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "text", + "content": "[44] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20875-20886, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[45] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20463" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 139 + ], + "type": "text", + "content": "[46] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 185 + ], + "type": "text", + "content": "[47] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Point scene understanding via disentangled instance mesh reconstruction. In European Conference on Computer Vision, pages 684-701. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 241 + ], + "type": "text", + "content": "[48] Michal J Tyszkiiewicz, Kevis-Kokitsi Maninis, Stefan Popov, and Vittorio Ferrari. Raytran: 3d pose estimation and shape reconstruction of multiple objects from videos with raytraced transformers. In European Conference on Computer Vision, pages 211-228. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 297 + ], + "type": "text", + "content": "[49] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. Advances in Neural Information Processing Systems, 35: 29975-29988, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 353 + ], + "type": "text", + "content": "[50] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 421 + ], + "type": "text", + "content": "[51] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, et al. Rodin: A generative model for sculpting 3d digital avatars using diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4563-4573, 2023. 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 487 + ], + "type": "text", + "content": "[52] Rundi Wu, Xuelin Chen, Yixin Zhuang, and Baoquan Chen. Multimodal shape completion via conditional generative adversarial networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 281-296. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 555 + ], + "type": "text", + "content": "[53] Yushuang Wu, Zizheng Yan, Ce Chen, Lai Wei, Xiao Li, Guanbin Li, Yihao Li, Shuguang Cui, and Xiaoguang Han. Scoda: Domain adaptive shape completion for real scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17630-17641, 2023. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 288, + 601 + ], + "type": "text", + "content": "[54] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 657 + ], + "type": "text", + "content": "[55] Haozhe Xie, Hongxun Yao, Xiaoshuai Sun, Shangchen Zhou, and Shengping Zhang. Pix2vox: Context-aware 3d reconstruction from single and multi-view images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2690-2698, 2019. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[56] Haozhe Xie, Hongxun Yao, Shengping Zhang, Shangchen Zhou, and Wenxiu Sun. Pix2vox++: Multi-scale context-aware 3d object reconstruction from single and multiple images. International Journal of Computer Vision, 128(12): 2919-2935, 2020. 1, 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 496 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 194 + ], + "type": "text", + "content": "[58] Xingguang Yan, Liqiang Lin, Niloy J Mitra, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Shapeformer: Transformer-based shape completion via sparse representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6239–6249, 2022. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 196, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 547, + 259 + ], + "type": "text", + "content": "[59] Zhenpei Yang, Zhile Ren, Miguel Angel Bautista, Zaiwei Zhang, Qi Shan, and Qixing Huang. Fvor: Robust joint shape and pose optimization for few-view object reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2497-2507, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 262, + 546, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 262, + 546, + 305 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 546, + 305 + ], + "type": "text", + "content": "[60] Zhenpei Yang, Zaiwei Zhang, and Qixing Huang. Hm3d-abo: A photo-realistic dataset for object-centric multi-view 3d reconstruction. arXiv preprint arXiv:2206.12356, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 546, + 351 + ], + "type": "text", + "content": "[61] Biao Zhang, Jiapeng Tang, Matthias Nießner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Trans. Graph., 42(4), 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 353, + 546, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 546, + 407 + ], + "type": "text", + "content": "[62] Cheng Zhang, Zhaopeng Cui, Yinda Zhang, Bing Zeng, Marc Pollefeys, and Shuaicheng Liu. Holistic 3d scene understanding from a single image with implicit representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8833-8842, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 408, + 546, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 546, + 452 + ], + "type": "text", + "content": "[63] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 307, + 453, + 546, + 496 + ], + "type": "text", + "content": "[64] Xin-Yang Zheng, Hao Pan, Peng-Shuai Wang, Xin Tong, Yang Liu, and Heung-Yeung Shum. Locally attentional sdf diffusion for controllable 3d shape generation. ACM Transactions on Graphics (SIGGRAPH), 42(4), 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20464" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_content_list.json b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..65ffa758e08abf9d11d912aed75f7b89ad73258a --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_content_list.json @@ -0,0 +1,1538 @@ +[ + { + "type": "text", + "text": "LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation", + "text_level": 1, + "bbox": [ + 129, + 128, + 841, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ke Guo $^{1,2,3}$ , Zhenwei Miao $^{1*}$ , Wei Jing $^{1}$ , Weiwei Liu $^{1}$ , Weizi Li $^{4}$ , Dayang Hao $^{1}$ , Jia Pan $^{2,3}$ $^{1}$ Alibaba Group $^{2}$ The University of Hong Kong \n $^{3}$ Centre for Transformative Garment Production, Hong Kong $^{4}$ University of Tennessee, Knoxville \nu3006612@connect.hku.hk, zhenwei.mzw@alibaba-inc.com \n21wjing@gmail.com, 11932061@zju.edu.cn, weizili@utk.edu, jpan@cs.hku.hk", + "bbox": [ + 104, + 202, + 895, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microscopic traffic simulation plays a crucial role in transportation engineering by providing insights into individual vehicle behavior and overall traffic flow. However, creating a realistic simulator that accurately replicates human driving behaviors in various traffic conditions presents significant challenges. Traditional simulators relying on heuristic models often fail to deliver accurate simulations due to the complexity of real-world traffic environments. Due to the covariate shift issue, existing imitation learning-based simulators often fail to generate stable long-term simulations. In this paper, we propose a novel approach called learner-aware supervised imitation learning to address the covariate shift problem in multi-agent imitation learning. By leveraging a variational autoencoder simultaneously modeling the expert and learner state distribution, our approach augments expert states such that the augmented state is aware of learner state distribution. Our method, applied to urban traffic simulation, demonstrates significant improvements over existing state-of-the-art baselines in both short-term microscopic and long-term macroscopic realism when evaluated on the real-world dataset pNEUMA.", + "bbox": [ + 76, + 358, + 472, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 717, + 209, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microscopic traffic simulation is a cornerstone in transportation engineering. It enables engineers to predict and analyze individual vehicle behavior, providing crucial insights into how alterations in road structures or traffic management strategies might influence overall traffic flow; it allows for the testing of diverse scenarios without disrupting real-world traffic; and it enhances safety by pinpointing potential hazards and devising strategies to mitigate risks. By leveraging simulations, engineers can optimize traffic flow,", + "bbox": [ + 76, + 742, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reduce congestion, and enhance overall efficiency, proving particularly advantageous in designing or enhancing road networks. Using the simulation data, policymakers and urban planners can make informed decisions aligned with community needs regarding transportation infrastructure.", + "bbox": [ + 500, + 328, + 890, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, generating realistic and accurate simulations that can simultaneously replicate the microscopic response of human drivers in various traffic conditions and long-term macroscopic traffic statistics is challenging. In recent years, significant efforts have been invested in developing realistic traffic simulators with the goal to accurately model human driving behaviors. Traditional traffic simulators, such as SUMO [24], AIMSUN [2], and MITSIM [44], typically rely on heuristic car-following models like the Intelligent Driver Model (IDM) [42]. However, despite carefully calibrating parameters, the rule-based models often fail to deliver accurate simulations due to the complexity of real-world traffic environments [12, 30]. Factors such as the road structure, neighboring vehicles, and even driver psychology can influence the decisions of human drivers, making it challenging to achieve accurate simulations [7, 27, 43].", + "bbox": [ + 496, + 407, + 892, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In pursuit of realistic traffic simulations, researchers have turned to neural networks to represent the driving model through imitation learning (IL) from human demonstrations. Most traffic simulation approaches [4, 34, 41] leverage behavior cloning (BC) [32] to learn a driving policy by minimizing the disparity between the model output and the human demonstrations in training data. However, BC is hindered by covariate shift [37], where the state induced by the learner's policy progressively diverges from the expert's distribution. Existing BC-based simulators have succeeded in short-term (less than 20 seconds) simulation applications like autonomous driving tests but often fail to generate stable long-term simulations.", + "bbox": [ + 496, + 654, + 893, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address covariate shift, existing methods such as DAGger [37], DART [25], and ADAPS [28] incorporate supervisor (humans or principled simulations) corrections", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Zhenwei Miao is the corresponding author.", + "bbox": [ + 94, + 886, + 331, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15386", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "at the learner's or perturbed expert's states. However, human supervision can be problematic due to intensive labor and judgment errors, and principled simulations may not account for heterogeneous driving behaviors. Recent traffic simulators [40, 46] propose using inverse reinforcement learning (IRL). These methods, such as generative adversarial imitation learning (GAIL) [17] and adversarial inverse reinforcement learning (AIRL) [13], learn a reward function using a discriminator neural network within Generative Adversarial Networks (GANs) [14]. The policy network is trained to maximize the learned reward through online reinforcement learning (RL), enabling agents to handle out-of-distribution states. However, directly applying GAIL to traffic simulation can be problematic [5]. The dynamic nature of the environment in the multi-agent system can lead to noises during policy learning, resulting in highly biased estimated gradients. Furthermore, training the discriminator in GAIL is challenging due to the instability and sensitivity of hyperparameters during min-max optimization.", + "bbox": [ + 75, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the issue of covariate shift in multi-agent imitation learning without depending on costly expert supervision or unstable discriminators and multi-agent RL, we propose Learner-Aware Supervised Imitation Learning (LASIL) for tackling the covariate shift problem during policy learning. We mitigate the distribution shift between expert and learner state distribution by augmenting the expert state distribution. However, there is no expert supervision at any augmented state, so we ensure that an augmented state is close to an expert state such that the future trajectory of the expert state can serve as the target trajectory for the augmented state to constrain the learner within the expert state distribution. Hence, our goal is to augment the expert state to cover the learner's state distribution while remaining close to the original expert state distribution. To achieve this, we use a variational autoencoder (VAE) [21] to simultaneously model the distributions of both the expert and learner states. By minimizing the VAE's latent space regularization loss of modeling both distributions, we project the expert and learner states into a unified latent space. And by minimizing the VAE's reconstruction loss, the resulting reconstruction leveraging such learnt latent space will resemble both expert and learner state distribution. As a result, when inputting expert states into the trained VAE, we obtain a learner-aware augmented expert state.", + "bbox": [ + 75, + 383, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In practice, we divide an agent's state into two parts: past trajectory and context, i.e., other features in the state including vehicle type, waypoints, and destination. We observe that the distribution of the context is consistent regardless of a learnt policy due to the static features of traffic conditions embedded in the context, leading to less covariate shift. Therefore, we propose a context-conditioned VAE to model the context-conditioned trajectory distribution of both expert and learner states. The decoder of the context-", + "bbox": [ + 75, + 763, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditioned VAE will receive both the latent variable and the context, and yield only the trajectory information.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 517, + 123, + 820, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose learner-aware supervised imitation learning (LASIL) to achieve stable learning and alleviate covariate shift in multi-agent imitation learning.", + "- We propose a learner-aware data augmentation method based on a context-conditioned VAE that generates learner-aware augmented expert states.", + "- Our approach is tailored for urban traffic simulation. To the best of our knowledge, it is the first imitation learning-based traffic simulator that can reproduce long-term (more than 10 minutes) stable microscopic simulation, achieving $40\\mathrm{x}$ simulation length improvements over previous state-of-the-art [4, 41, 46].", + "- We evaluate our method on the real-world dataset pNEUMA [3] with over half a million trajectories. Our approach outperforms state-of-the-art baselines in both short-term microscopic and long-term macroscopic simulations. The code is available at https://github.com/Kguo-cs/LSAIL." + ], + "bbox": [ + 500, + 138, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 426, + 640, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Imitation learning", + "text_level": 1, + "bbox": [ + 500, + 452, + 676, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing imitation learning (IL) methods can be broadly categorized into behavior cloning (BC) and inverse reinforcement learning (IRL) approaches. BC [32] learns a policy in a supervised fashion by minimizing the discrepancy between the learner's actions and those of an expert. However, BC suffers from the issue of covariate shift, where the state distribution induced by the learner's policy gradually deviates from that of the expert. To address this, methods like DAGger [37] and DART [25] request supervisor corrections at the learner's or perturbed expert's states. Our method follows a similar supervised learning approach to DAGger and DART, but does not require access to an expert policy.", + "bbox": [ + 496, + 476, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the challenges in obtaining expert supervision, recent IRL-based methods utilize feedback from a neural network-based discriminator to handle covariate shift. Typically, these methods involve an iterative process alternating between reward estimation and reinforcement learning. Earlier IRL-based methods [15, 35, 39, 47] require frequent dynamic programming processes, while recent adversarial IL approaches integrate reward function learning with policy learning using a GAN formulation. However, both GAN and RL training processes are known to be unstable, sensitive to hyperparameters, and have poor sample efficiency [9]. Moreover, the discriminator can easily exploit insignificant differences between expert and policy samples, leading to undesirable performance [45]. In contrast, our method avoids the min-max optimization problem and the sample-inefficient RL process, requiring minimal", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15387", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fine-tuning. Instead, we utilize the real future trajectory as the target state to learn corrective behavior.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Imitation Learning-based Traffic simulation", + "text_level": 1, + "bbox": [ + 76, + 132, + 452, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent traffic simulators focus on enhancing realism by leveraging imitation learning (IL) from human driving demonstrations, which extend traditional BC and IRL methods to tackle the challenging multi-agent IL problem.", + "bbox": [ + 76, + 155, + 468, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BC-based traffic simulators like TrafficSim [41] and SimNet [4] typically begin by training a prediction model and subsequently adjust the predicted trajectories to prevent collisions and adhere to traffic regulations during simulation. However, BC-based methods face challenges in achieving long-term simulation due to the covariate shift problem. To mitigate this issue, we augment expert state based on the policy distribution, enabling stable long-term simulation. Additionally, we enhance performance by modifying the predicted trajectory during simulation through road projection and ensuring smoothness from the current state. Notably, we skip the computationally intensive collision removal operation, prioritizing the capture of macroscopic long-term influence over micro-level details.", + "bbox": [ + 75, + 215, + 468, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "IRL-based simulators learn the underlying reward function of human driving behavior and derive the driving policy by maximizing the learned reward. While adversarial IRL methods theoretically address the covariate shift of BC in a single-agent context through online interaction, its performance deteriorates in the multi-agent IL domain due to the dynamic environment, complicating the training process. To tackle this challenge, approaches like PS-GAIL [40] and PS-ARIL [46] adopt two-stage learning and gradually introduce vehicles to the environment. Nonetheless, these methods still exhibit significant undesirable traffic phenomena, such as off-road driving, collisions, and abrupt braking. Building upon PS-GAIL, the reward-augmented imitation learning (RAIL) method [5, 22] penalizes undesirable phenomena by introducing a hand-crafted reward. However, maximizing the new reward does not guarantee the recovery of human-like trajectories. Despite numerous enhancements to the original GAIL framework, these IRL-based methods often struggle to produce stable long-term traffic flow, as evidenced in our experiments. In contrast, our method is a supervised learning approach, leading to faster, simpler, and more stable learning of driving policies.", + "bbox": [ + 75, + 428, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 76, + 776, + 205, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Markov Decision Process", + "text_level": 1, + "bbox": [ + 76, + 801, + 305, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We model the human driving process using a Markov decision process (MDP) denoted as $\\mathcal{M} = \\{\\mathcal{S},\\mathcal{A},\\mathcal{T},r\\}$ , incorporating a time horizon of $T$ , where $\\mathcal{S}$ and $\\mathcal{A}$ represent the continuous state and action spaces respectively. A stochastic function $\\mathcal{T}:S\\times \\mathcal{A}\\times \\mathcal{S}\\rightarrow [0,1]$ describes the system", + "bbox": [ + 76, + 824, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dynamics, and $r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}$ is a reward function. The policy $\\pi(\\boldsymbol{a} | s)$ determines the probability of selecting an action $a$ at a state $s$ , and a trajectory $\\tau = (s_0, \\boldsymbol{a}_0, \\dots, s_T, \\boldsymbol{a}_T)$ represents a sequence of state-action pairs. The marginal state distribution of a policy $\\pi$ is computed as $\\rho^{\\pi}(s) = \\frac{1}{D} \\sum_{s \\in \\mathcal{D}_{\\pi}} \\delta_s$ , where $\\mathcal{D}_{\\pi}$ denotes the set of states of size $D$ in trajectories induced by the policy $\\pi$ , and $\\delta_s$ signifies a Dirac distribution centered on $s$ . Similarly, the marginal state-action distribution $\\rho^{\\pi}(s, \\boldsymbol{a})$ is computed.", + "bbox": [ + 500, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Imitation Learning", + "text_level": 1, + "bbox": [ + 500, + 257, + 683, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the application of IL for learning a driving policy, we assume that all agents adopt the same policy. The objective is to learn a policy $\\pi$ that minimizes the $f$ -divergence between the marginal state-action distribution of the expert's demonstrations $\\rho^{\\exp}(s,a)$ and the learner policy's distribution $\\rho^{\\pi}(s,a)$ . For example, BC optimizes the policy to minimize the Kullback-Leibler (KL) divergence at the expert state distribution $\\mathbb{E}_{\\rho^{\\exp}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]$ , while DAgger minimizes $\\mathbb{E}_{\\rho^{\\pi}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]$ . On the other hand, GAIL minimizes the Jensen-Shannon divergence $D_{\\mathrm{JS}}(\\rho^{\\exp}(s,a)\\| \\rho^{\\pi}(s,a))$ , while AIRL minimizes the KL divergence $\\mathrm{KL}(\\rho^{\\pi}(s,a)\\| \\rho^{\\exp}(s,a))$ .", + "bbox": [ + 498, + 287, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, GAIL and AIRL often exhibit unsatisfactory practical performance due to their optimization processes' instability and sample inefficiency, which involve GANs and RL. In contrast, BC only requires simple and stable supervised learning but suffers from the covariate shift issue, as it only minimizes the policy difference at the expert state distribution without guaranteeing performance at the learner state distribution. DAGger addresses the covariate shift problem but requires access to the expert policy. To mitigate the covariate shift problem without depending on the expert policy, we propose maximizing the transition probability $\\mathbb{E}_{\\boldsymbol{s} \\sim \\rho^{\\exp}(\\boldsymbol{s}), \\boldsymbol{s} + \\epsilon \\sim \\rho^{\\pi}(\\boldsymbol{s})}[\\mathcal{T}(\\boldsymbol{s} + \\epsilon, \\pi(\\boldsymbol{s} + \\epsilon), \\boldsymbol{s}')]$ , where $\\epsilon$ is a small augmentation term. Note that our policy learns to predict the distribution of the next state $s'$ instead of the action. Our approach assumes that when the expert's original future trajectory can serve as supervision to guide the agent back towards the expert distribution.", + "bbox": [ + 498, + 477, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Variational Autoencoder", + "text_level": 1, + "bbox": [ + 500, + 763, + 723, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The VAE defines a generative model given by $p_{\\theta}(\\pmb{x}, \\pmb{z}) = p(\\pmb{z}) p_{\\theta}(\\pmb{x} | \\pmb{z})$ , where $\\pmb{z}$ is the latent variable with prior distribution $p(\\pmb{z})$ , and $p_{\\theta}(\\pmb{x} | \\pmb{z})$ represents the conditional distribution modeling the likelihood of data $\\pmb{x}$ given $\\pmb{z}$ . The learning objective is to maximize the training samples' marginal log-likelihood $\\log p_{\\theta}(\\pmb{x})$ . However, due to the intractability of marginalization, VAE maximizes the variational lower", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15388", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bound using $q_{\\phi}(\\pmb {z}|\\pmb {x})$ as the approximate posterior:", + "bbox": [ + 76, + 90, + 413, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {x}) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x})} [ \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) ] - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x}) \\| p (\\boldsymbol {z})\\right) \\tag {1} \\\\ := - \\mathcal {L} _ {\\text {r e c}} (\\boldsymbol {x}) - \\mathcal {L} _ {\\mathrm {K L}} (\\boldsymbol {x}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 116, + 468, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{rec}}(\\boldsymbol{x})$ represents the reconstruction loss, which penalizes the network for creating outputs different from the input. $\\mathcal{L}_{\\mathrm{KL}}(\\boldsymbol{x})$ represents the KL divergence loss to make a continuous and smooth latent space, allowing easy random sampling and interpolation. Intuitively, this KL loss encourages the encoder to distribute all encodings evenly around the center of the latent space.", + "bbox": [ + 75, + 181, + 468, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 299, + 168, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 1, we present an overview of our method. Our model comprises three modules: a VAE-based data augmentation module, a policy network, and a post-processing (LQR and on-road projection) module. During training, we augment each expert data by generating a learner-aware augmented expert state through the VAE's reconstruction of its past trajectory. Using this augmented state along with the original future trajectory, we train the learner's policy network through supervised learning. During simulation, we roll out the policy network with several post-processing steps including sampling, projecting the trajectories onto the road, and smoothing the projected trajectories.", + "bbox": [ + 75, + 325, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. State Representation", + "text_level": 1, + "bbox": [ + 76, + 513, + 272, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As human drivers make decisions mainly depending on their surrounding information, we build a graph to model the traffic system with each driving agent as its node.", + "bbox": [ + 75, + 537, + 468, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Node: The state of each agent can be divided into two components: the past trajectory $s_p$ and the context $s_c$ . In practice, the past trajectory is composed of the agent's positions at several past time steps, including the current time step. The context includes its type, nearby waypoints' positions and corresponding road width, the traffic light status of the road it is traveling on, and destination. The waypoints are composed of the center points of the routing roads with a fixed interval. We transform each agent's state into its individual coordinate system to learn a policy with transformation invariance. To reduce the implicit covariate shift, we set each agent's current position adding a Gaussian perturbation as the origin and the $x$ -axis directions pointing towards the agent's destination like [16].", + "bbox": [ + 75, + 584, + 468, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge: When each agent's state coordinates are transformed from the global coordinate system to their individual coordinate system, information about the relative positions among agents is lost. However, a traffic model needs the relative information of agents to understand how they interact with each other. To preserve these relationships, we introduce directed edges between neighboring agents. The", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "edge feature is the relative position of the destination node's coordinate origin in the source node's coordinate system.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Learner-aware Data Augmentation", + "text_level": 1, + "bbox": [ + 498, + 130, + 805, + 146 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the covariate shift issue between the expert and learner state distributions, we propose to minimize the expert's and learner's embedded state distribution difference. In practice, we propose utilizing the same VAE to model the expert and learner state distributions simultaneously. As both the expert and learner state are projected to the same latent space, the distribution of the state reconstructed from the joint latent space can resemble both distributions.", + "bbox": [ + 496, + 152, + 890, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the past trajectory, the context distribution is more challenging to model but exhibits less covariate shift. Therefore, we propose using a context-conditioned VAE to specifically model the context-conditioned trajectory distribution rather than the state distribution. For each expert or learner state represented as $s = (s_p, s_c)$ , we employ an encoder $q_{\\phi}(z | s_p, s_c)$ to obtain the latent variable distribution. We can sample a latent variable $z$ from this distribution by applying the reparameterization trick. Subsequently, a decoder $p_{\\theta}(s_p | z, s_c)$ is used to reconstruct the distribution of the past trajectory $s_p$ given $z$ and $s_c$ . The VAE is trained to maximize the variational lower bound, which incorporates the context-conditioned log-likelihood of both the expert and learner past trajectories:", + "bbox": [ + 496, + 275, + 890, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V A E} = \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\exp}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right] + \\lambda \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\pi}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 497, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\lambda$ is a hyperparameter that controls the degree to which the augmented context-conditioned past trajectory distribution aligns with the learner's context-conditioned past trajectory distribution. The term $\\mathcal{L}(s_p|s_c)$ is given by:", + "bbox": [ + 496, + 525, + 890, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) = \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} \\left(\\boldsymbol {z} \\mid \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}\\right)} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {z}, \\boldsymbol {s} _ {c}\\right) \\right] \\\\ - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}) \\| p (\\boldsymbol {z})\\right). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 598, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity, we assume $p_{\\theta}$ and $q_{\\phi}$ as multivariate normal distributions with diagonal variance matrices. The prior distribution $p(z)$ is set as an isotropic unit Gaussian $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "bbox": [ + 496, + 646, + 890, + 693 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Edge-enhanced Graph Attention Network", + "text_level": 1, + "bbox": [ + 498, + 700, + 857, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build all models in our approach including the VAE encoder, decoder, and learner's policy network based on an edge-enhanced graph attention (EGAT) network [11, 33], which model interactions by aggregating neighboring node and edge information using an attention mechanism. The node features in the first layer are obtained by embedding the node input with a fully connected layer and the node features in the other layers are calculated by:", + "bbox": [ + 496, + 724, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {i} ^ {l} = \\sigma \\left(\\sum_ {j \\in \\mathcal {N} _ {i}} \\alpha_ {i j} ^ {l} \\boldsymbol {W} ^ {l} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 854, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15389", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c52e0a4a753e710f5e0703f22977e3dbee94a3e67553d93d5b4a19ad522866d7.jpg", + "image_caption": [ + "Figure 1. Overview of our approach. The processes with purple and red arrows handle only expert or learner data, respectively, while the processes represented by black arrows apply to both data. Each state is a multi-agent state represented by a graph. Each model, including the VAE encoder, decoder, and policy network, is implemented using an EGAT network." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h_i^l$ is the feature of node $i$ in the $l$ th layer, $e_{ij}$ is the coordinate origin's position of node $j$ relative to node $i$ , $\\| \\cdot \\|$ means the concatenation, $W^l$ is the learnable weight matrix, $\\mathcal{N}_i$ is the set of the first-order neighbors of node $i$ (including the node itself), and $\\sigma$ is a non-linear activation function. The attention coefficient $\\alpha_{ij}^l$ indicates the importance of node $j$ to node $i$ , considering both node and edge features, and is computed as:", + "bbox": [ + 76, + 383, + 473, + 507 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i j} ^ {l} = \\operatorname {s o f t m a x} _ {j} \\left(\\sigma \\left(\\left(\\boldsymbol {w} ^ {l}\\right) ^ {T} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 516, + 468, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\boldsymbol{w}^l$ is a learnable weight vector, and normalization is performed on the weights across all neighbors of node $i$ using a softmax function. After passing through multiple EGAT layers, the node features at the last layer is fed into a fully connected layer to obtain the outputs of the network.", + "bbox": [ + 76, + 555, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Policy Loss", + "text_level": 1, + "bbox": [ + 76, + 642, + 200, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the learner's policy network, we first sample a learner-aware augmented expert state from the context-conditioned VAE for each expert state. Then, the policy network predicts its future position distribution over $T$ time steps, denoted as $p(\\hat{p}_1^m, \\hat{p}_2^m, \\dots, \\hat{p}_T^m)$ , which is assumed to be a product of multi-variable Gaussian distributions:", + "bbox": [ + 75, + 666, + 468, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\hat {\\boldsymbol {p}} _ {1} ^ {m}, \\hat {\\boldsymbol {p}} _ {2} ^ {m}, \\dots , \\hat {\\boldsymbol {p}} _ {T} ^ {m}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 768, + 468, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{\\mu}}_t^m$ and $\\hat{\\Sigma}_t^m$ represent the mean and covariance matrix of the predicted position $\\hat{\\pmb{p}}_t^m$ at future time step $t$ . We assume that there is no correlation between the position distributions at different future time steps. To learn the policy network, we minimize the negative log-likelihood (NLL)", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss of all agents' ground-truth future trajectories:", + "bbox": [ + 498, + 383, + 830, + 401 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {N L L} = - \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\log \\left(\\mathcal {N} \\left(\\boldsymbol {p} _ {t} ^ {m} - \\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 409, + 890, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{p}_t^m$ denotes the ground truth position of agent $m$ at future time step $t$ , and $M$ is the total number of agents.", + "bbox": [ + 498, + 458, + 893, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Simulation Process", + "text_level": 1, + "bbox": [ + 500, + 496, + 683, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we simultaneously simulate (roll out) the learned policy network iteratively to get learner state samples. Instead of directly updating each agent to its predicted position, we apply several post-processing steps to the prediction for better realism. Firstly, we sample from the distribution, and then project each sampled position onto the nearest on-road point. Then, we smooth the projected trajectory with a linear-quadratic regulator (LQR) [1] by minimizing the total commutative quadratic cost of a linear dynamic system described by:", + "bbox": [ + 496, + 518, + 893, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t + 1} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t + 1} ^ {m} \\end{array} \\right] = \\left[ \\begin{array}{c c} \\boldsymbol {I} & \\boldsymbol {D} \\\\ 0 & \\boldsymbol {I} \\end{array} \\right] \\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t} ^ {m} \\end{array} \\right] + \\left[ \\begin{array}{c} \\boldsymbol {D} ^ {2} \\\\ \\boldsymbol {D} \\end{array} \\right] \\tilde {\\boldsymbol {a}} _ {t} ^ {m}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 678, + 890, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ is a diagonal matrix with the interval of each time step as diagonal entries, and $\\tilde{\\boldsymbol{p}}_t^m$ , $\\tilde{\\boldsymbol{v}}_t^m$ , $\\tilde{\\boldsymbol{a}}_t^m$ represent the LQR-planned position, velocity, and acceleration. The system is subject to a quadratic cost function given by:", + "bbox": [ + 496, + 720, + 893, + 781 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {J} = \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\| \\tilde {\\boldsymbol {p}} _ {t} ^ {m} - \\bar {\\boldsymbol {p}} _ {t} ^ {m} \\| ^ {2} + \\eta_ {\\boldsymbol {a}} \\| \\tilde {\\boldsymbol {a}} _ {t} ^ {m} \\| ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 789, + 890, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the projected predicted position $\\bar{\\pmb{p}}_m^t$ is considered as the target pose, and the hyper-parameter $\\eta_{\\mathbf{a}}$ is used to penalize high acceleration. Finally, each agent is updated to the first position of the planned trajectory.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15390", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.6. Training Process", + "text_level": 1, + "bbox": [ + 76, + 90, + 243, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each training step, the policy network is trained to maximize the probability of the expert's future trajectory using its context state and history trajectory augmented by the VAE as input. Simultaneously, the VAE is iteratively trained to reconstruct the expert and learner data. The expert data are the same as the policy network's training data before augmentation. The learner data is sampled from a replay buffer. Every $N$ training steps, we empty the replay buffer and roll out the current policy with post-processing for $S$ time steps to generate the learner data and store it in the replay buffer. During the roll-out, we start at a random time step in the training dataset and apply our model in a closed-loop manner. At each time step during the rollout, we sample a future trajectory for each vehicle from the Gaussian distribution predicted by our policy network, project the predicted trajectory onto the road, smooth the on-road trajectory with LQR, and finally update the vehicle to the first position of the smoothed trajectory.", + "bbox": [ + 75, + 114, + 472, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 409, + 202, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 76, + 436, + 171, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize a real-world urban dataset called pNEUMA [3], which was collected by 10 drones in Athens over a span of 4 days. The dataset encompasses over half a million vehicle trajectories within a large area encompassing over $100\\mathrm{km}$ of lanes and approximately 100 intersections. The recordings were conducted at 5 intervals each day, with each period lasting about 15 minutes and a data collection time interval of 0.04 seconds. To enhance computational efficiency, we adopt a time step of 0.4 seconds. The dataset is divided into a training set, comprising recordings from the first 3 days, and a validation/test set, consisting of recordings from the final day. Notably, we choose not to utilize other popular traffic datasets like NGSIM [8] or HighD [23] or autonomous driving datasets like Lyft [18] or nuPlan [6], as they only encompass samll-scale scenarios. This limitation makes it inadequate for evaluate macroscopic realism.", + "bbox": [ + 75, + 462, + 472, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Metrics", + "text_level": 1, + "bbox": [ + 76, + 720, + 173, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the realism of our simulator by measuring the similarity between the simulation result and real data. During evaluation, each vehicle enters the simulator at its first recorded time and position, and is controlled by our simulator to complete its recorded route. When an agent reaches its final recorded position, it is removed from the simulator.", + "bbox": [ + 75, + 746, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Short-term microscopic: Following [5, 40], we conduct a short-term microscopic evaluation by simulating for 20 seconds from a random time step in the test dataset. We first measure the similarity between the simulated and real", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data using position and velocity RMSE metrics given by:", + "bbox": [ + 498, + 90, + 885, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R M S E} = \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\sqrt {\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t} ^ {m} \\| ^ {2}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 116, + 890, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s_t^m$ and $\\hat{s}_t^m$ were the real and simulated value of the position or velocity of the agent $m$ at time step $t$ , respectively. $T_s$ was the total simulated time steps, and $M$ was the total simulated agent number. Besides, we measure the minimum Average Displace Error (minADE) to not penalize reasonable trajectory unlike the real one:", + "bbox": [ + 496, + 172, + 893, + 265 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\min A D E _ {N} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\min _ {\\hat {\\boldsymbol {s}} _ {n}} \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t, n} ^ {m} \\| ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 273, + 890, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N = 20$ is roll-out times. We also calculate the off-road rate (the avarage proportion of vehicles that deviate more than $1.5\\mathrm{m}$ from the road over all time steps). The common collision rate metric is not used because we focus on the long-term impact and the dataset does not provide accurate vehicle size and heading information.", + "bbox": [ + 496, + 324, + 890, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Long-term macroscopic: we also evaluate our model's long-term macroscopic accuracy on five periods for 800 seconds from its initial recording time. To measure the long-term performance, we use two standard macroscopic metrics for traffic flow data [29, 36, 38], namely road density and speed RMSE, in addition to the off-road rate. The density of a road at a time step is calculated by dividing the number of vehicles on the road by its total lane length, assuming that all lanes have the same width. Meanwhile, the road speed is computed as the mean speed of all vehicles on the road. To quantify the similarity between the simulated and ground truth values, we still use RMSE, where the variable $M$ becomes the total number of roads.", + "bbox": [ + 496, + 415, + 892, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Performance", + "text_level": 1, + "bbox": [ + 500, + 619, + 635, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against state-of-the-art baselines:", + "bbox": [ + 500, + 643, + 883, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SUMO [24]: we use the IDM model [42] as the car-following model and mobil [19] as the lane-changing model. We tune the IDM's parameters for 6 types of vehicle by minimizing the MSE between the IDM calculated acceleration and real acceleration using an Adam optimizer [20].", + "bbox": [ + 496, + 659, + 890, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BC [32]: we learn our model directly by the BC method.", + "bbox": [ + 517, + 734, + 890, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MARL [26]: we train our model using IPPO [10] as the Multi-Agent RL algorithm, where the reward function is composed of three parts: a displacement reward (average distance between agent trajectory with GT trajectory), an off-road penalty and a terminal reward.", + "bbox": [ + 496, + 750, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MARL+BC [31]: we add a behavior cloning term in the loss function while learning the policy with MARL.", + "bbox": [ + 496, + 825, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PS-GAIL [40]: we let all vehicles share the same policy parameter and critic parameter and learn the policy using reward functions computed by GAIL.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15391", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0268c0a1fa53ed225ecb6bfcfbb35d7df3d3f48b8af73efc045b67907470045d.jpg", + "table_caption": [ + "Table 1. Comparison with baselines and ablated models on microscopic metrics for 20 seconds." + ], + "table_footnote": [], + "table_body": "
ModelPosition RMSE(m)Velocity RMSE(m/s)minADE20(m)Off-road(%)
SUMO [24]41.257.0024.200
BC [32]40.08±1.616.74±0.3722.02±3.9833.43±3.10
MARL [26]48.66±1.548.92±0.6740.07±1.263.70±0.40
MARL+BC [31]29.78±0.424.54±0.1422.07±0.2111.14±1.17
PS-GAIL [40]34.78±0.435.33±0.1227.86±0.337.65±0.32
RAIL [5]31.62±0.655.51±0.0924.51±0.412.57±0.47
LASIL (ours)19.21±0.443.02±0.0712.79±0.320.28±0.01
w/o Augmentation21.62±0.543.34±0.1413.81±0.371.31±0.18
w/o Context-conditioned23.03±0.873.73±0.1214.94±0.430.51±0.04
w/o On-road Projection22.53±0.423.26±0.1014.42±0.331.74±0.60
w/o LQR22.62±0.434.05±0.1514.02±0.280.22±0.02
", + "bbox": [ + 122, + 103, + 848, + 299 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f9b8def84aa564045cc8cd26e7540397c1fc332b7a0e7a8cddf50f4bae582c4f.jpg", + "table_caption": [ + "Table 2. Comparison with baselines and ablated models on macroscopic metrics for 800 seconds." + ], + "table_footnote": [], + "table_body": "
ModelRoad Density RMSE(veh/km)Road Speed RMSE(m/s)Off-road(%)
SUMO [24]52.705.520
BC [32]61.51±1.535.38±0.2142.15±5.25
MARL [26]90.59±2.655.27±0.484.63±0.64
MARL+BC [31]43.60±2.804.01±0.1322.73±5.62
PS-GAIL [40]54.06±1.234.03±0.0513.24±3.20
RAIL [5]54.45±1.893.89±0.112.92±0.38
LASIL (ours)45.13±0.253.17±0.140.34±0.03
w/o Augmentation55.65±0.443.73±0.239.62±1.40
w/o Context-conditioned55.98±0.514.48±0.190.64±0.08
w/o On-road Projection52.90±0.543.59±0.2011.33±2.37
w/o LQR61.68±0.963.54±0.160.47±0.07
", + "bbox": [ + 129, + 319, + 841, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RAIL [5]: we learn the model as PS-GAIL but with additional displacement, off-road and terminal rewards.", + "bbox": [ + 76, + 520, + 468, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train and evaluate each model five times to obtain the mean and standard deviation (std) of various metrics. Note that we do not apply on-road projection and LQR for baselines. We evaluate both short-term and long-term performance, as shown in Tab. 1 and Tab. 2, respectively. Our method achieves better results than all baselines in terms of position and velocity RMSE, road density and speed RMSE, with minor off-road rate.", + "bbox": [ + 75, + 551, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 681, + 230, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a series of ablation experiments to assess the individual contributions of crucial components in our approach, whose results are presented in Tab. 1 and Tab. 2.", + "bbox": [ + 75, + 704, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Augmentation: By removing our VAE module and directly learning the original expert state and action, we analyze the importance of our data augmentation technique. The results demonstrate that augmentation plays a vital role in improving both short-term and long-term simulation performance by mitigating covariate shift.", + "bbox": [ + 75, + 750, + 470, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Context-conditioned VAE: To emphasize the significance of modeling the context-conditioned trajectory distribution rather than the whole state distribution, we replaced our context-conditioned VAE with a naive VAE that directly", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "models the expert and learner state distribution. The observed drop in performance demonstrates the challenges in reconstructing the context distribution.", + "bbox": [ + 496, + 520, + 890, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On-road projection: Our ablation study on the on-road projection module aims to show its impact on reducing the off-road rate. The results show that the on-road projection module leads to a notable decrease in the off-road rate and moderate improvements in other performance metrics.", + "bbox": [ + 496, + 571, + 890, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LQR: Removing LQR module allows us to evaluate its effectiveness. While LQR led to a higher short-term off-road rate due to more constrained movement (driving off-road due to inertia), its removal deteriorates other short-term and long-term metrics, because LQR can smooth the trajectory, thus leads to more realistic simulations.", + "bbox": [ + 496, + 651, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Qualitative Result", + "text_level": 1, + "bbox": [ + 500, + 766, + 678, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 2, we present the mean road density and speed for real-world data, SUMO simulation, and our proposed method over all time steps. The results demonstrate that our proposed method surpasses the capabilities of the SUMO simulator at replicating long-term macroscopic traffic patterns due to our model's enhanced ability to replicate the typical microscopic driving behaviors over a long period.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15392", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fe78b129460e56077e2dacba9ddc2df6714e3b7076b387f2f8e7ad5ec3de2b5.jpg", + "image_caption": [ + "Figure 2. Mean density and speed on each road over all time steps in the long-term evaluation. Our method's density and speed hot-maps have a more similar color to the ground-truth one compared with the SUMO's." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 861, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 678, + 194, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we have addressed the challenge of creating a realistic traffic simulator that accurately models human driving behaviors in various traffic conditions. Traditional imitation learning-based simulators often fail to deliver accurate long-term simulations due to the covariate shift problem in multi-agent imitation learning. To tackle this issue, we proposed a learner-aware supervised imitation learning method, which leverages a context-conditioned VAE to generate learner-aware augmented expert states. We leverage a context-conditioned VAE to simultaneously reconstruct the expert and learner state. This approach enables the reproduction of long-term stable microscopic traffic simulations, marking a significant advancement in the field of urban", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "traffic simulation. Our method has demonstrated superior performance over existing state-of-the-art simulators when evaluated on the real-world dataset pNEUMA, achieving better short-term microscopic and long-term macroscopic similarity to real-world data than state-of-the-art baselines.", + "bbox": [ + 496, + 679, + 890, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 768, + 687, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research project is partially supported by the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, Hong Kong General Research Fund (11202119 and 11208718), Innovation and Technology Commission (GHP/126/21GD), and Guangdong, Hong Kong and Macao Joint Innovation Project (2023A0505010016).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15393", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Karl Johan Åström and Richard M Murray. Feedback systems: an introduction for scientists and engineers. Princeton University Press, 2021. 5", + "[2] Jaime Barceló and Jordi Casas. Dynamic network simulation with aimsun. Simulation approaches in transportation analysis: Recent advances and challenges, pages 57-98, 2005. 1", + "[3] Emmanouil Barmpounakis and Nikolas Geroliminis. On the new era of urban traffic monitoring with massive drone data: The pNEUMA large-scale field experiment. Transportation research part C: emerging technologies, 111:50-71, 2020. 2, 6", + "[4] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Blazej Osinski, Hugo Grimmett, and Peter Ondruska. SimNet: Learning reactive self-driving simulations from real-world observations. IEEE International Conference on Robotics and Automation, pages 5119-5125, 2021. 1, 2, 3", + "[5] Raunak P. Bhattacharyya, Derek J. Phillips, Changliu Liu, Jayesh K. Gupta, K. Driggs-Campbell, and Mykel J. Kochenderfer. Simulating emergent properties of human driving behavior using multi-agent reward augmented imitation learning. International Conference on Robotics and Automation, pages 789-795, 2019. 2, 3, 6, 7", + "[6] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. In IEEE/CVF Computer Vision and Pattern Recognition Conference ADP3 workshop, 2021. 6", + "[7] Qianwen Chao, Huikun Bi, Weizi Li, Tianlu Mao, Zhaoqi Wang, Ming C Lin, and Zhigang Deng. A survey on visual traffic simulation: Models, evaluations, and applications in autonomous driving. In Computer Graphics Forum, pages 287-308, 2020. 1", + "[8] James Colyar and John Halkias. Us highway 101 dataset. Federal Highway Administration, Tech. Rep., 2007. 6", + "[9] Robert Dadashi, L'essonard Hussenot, Matthieu Geist, and Olivier Pietquin. Primal wasserstein imitation learning. In International Conference of Learning Representation, 2021. 2", + "[10] Christian Schroeder de Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020. 6", + "[11] Frederik Diehl, Thomas Brunner, Michael Truong Le, and Alois Knoll. Graph neural networks for modelling traffic participant interaction. In IEEE Intelligent Vehicles Symposium, pages 695-701, 2019. 4", + "[12] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X. Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature Communications, 12, 2021. 1", + "[13] Justin Fu, Katie Luo Luo, and Sergey Levine. Learning robust rewards with adversarial inverse reinforcement learn" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In International Conference of Learning Representation, 2018. 2", + "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 2", + "[15] Ke Guo, Wenxi Liu, and Jia Pan. End-to-end trajectory distribution prediction based on occupancy grid maps. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2232-2241, 2022. 2", + "[16] Ke Guo, Wei Jing, Junbo Chen, and Jia Pan. CCIL: Context-conditioned imitation learning for urban driving. In Robotics: Science and Systems, 2023. 4", + "[17] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2", + "[18] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference of Robot Learning, 2020. 6", + "[19] Arne Kesting, Martin Treiber, and Dirk Helbing. General lane-changing model MOBIL for car-following models. Transportation Research Record, 1999(1):86-94, 2007. 6", + "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6", + "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2", + "[22] Yann Koeberle, Stefano Sabatini, Dzmitry V. Tsishkou, and Christophe Sabourin. Exploring the trade off between human driving imitation and safety for traffic simulation. IEEE International Conference on Intelligent Transportation Systems, pages 779-786, 2022. 3", + "[23] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In International Conference on Intelligent Transportation Systems, pages 2118-2125, 2018. 6", + "[24] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of SUMO-Simulation of Urban MObility. International journal on advances in systems and measurements, 5(3&4), 2012. 1, 6, 7", + "[25] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on Robot Learning, pages 143-156, 2017. 1, 2", + "[26] Quanyi Li, Zhenghao Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in Neural Information Processing Systems, 2023. 6, 7", + "[27] Weizi Li, David Wolinski, and Ming C. Lin. City-scale traffic animation using statistical learning and metamodel-based optimization. ACM Trans. Graph., 36(6):200:1-200:12, 2017. 1" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15394", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Weizi Li, David Wolinski, and Ming C. Lin. ADAPS: Autonomous driving via principled simulations. In IEEE International Conference on Robotics and Automation (ICRA), pages 7625-7631, 2019. 1", + "[29] Michael James Lighthill and Gerald Beresford Whitham. On kinematic waves ii. a theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229(1178):317-345, 1955. 6", + "[30] Weiwei Liu, Wei Jing, Ke Guo, Gang Xu, Yong Liu, et al. Traco: Learning virtual traffic coordinator for cooperation with multi-agent reinforcement learning. In Conference on Robot Learning, pages 2465-2477, 2023. 1", + "[31] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. arXiv preprint arXiv:2212.11419, 2022. 6, 7", + "[32] Donald Michie, Michael Bain, and Jean Hayes-Michie. Cognitive models from subcognitive skills. IEEE control engineering series, 44:71-99, 1990. 1, 2, 6, 7", + "[33] Xiaoyu Mo, Zhiyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network. IEEE Transactions on Intelligent Transportation Systems, 23:9554-9567, 2022. 4", + "[34] Jeremy Morton, Tim A Wheeler, and Mykel J Kochenderfer. Analysis of recurrent neural networks for probabilistic modeling of driver behavior. IEEE Transactions on Intelligent Transportation Systems, 18(5):1289-1298, 2016. 1", + "[35] Andrew Y Ng, Stuart Russell, et al. Algorithms for inverse reinforcement learning. In International Conference of Machine Learning, page 2, 2000. 2", + "[36] Paul I Richards. Shock waves on the highway. Operations research, 4(1):42-51, 1956. 6", + "[37] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics, pages 627-635, 2011. 1, 2", + "[38] Jason Sewall, David Wilkie, Paul Merrell, and Ming C Lin. Continuum traffic simulation. In Computer Graphics Forum, number 2, pages 439-448, 2010. 6", + "[39] Yu Shen, Weizi Li, and Ming C. Lin. Inverse reinforcement learning with hybrid-weight trust-region optimization and curriculum learning for autonomous maneuvering. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7421-7428, 2022. 2", + "[40] Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. Advances in neural information processing systems, 31, 2018. 2, 3, 6, 7", + "[41] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multiagent behaviors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400-10409, 2021. 1, 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1, 6", + "[43] David Wilkie, Jason Sewall, Weizi Li, and Ming C. Lin. Virtualized traffic at metropolitan scales. Frontiers in Robotics and AI, 2:11, 2015. 1", + "[44] QI Yang and Haris N Koutsopoulos. A microscopic traffic simulator for evaluation of dynamic traffic management systems. Transportation Research Part C: Emerging Technologies, 4(3):113-129, 1996. 1", + "[45] Kaifeng Zhang, Rui Zhao, Ziming Zhang, and Yang Gao. Auto-encoding adversarial imitation learning. arXiv preprint arXiv:2206.11004, abs/2206.11004, 2022. 2", + "[46] Guanjie Zheng, Hanyang Liu, Kai Xu, and Zhenhui Jessie Li. Objective-aware traffic simulation via inverse reinforcement learning. In International Joint Conference on Artificial Intelligence, 2021. 2, 3", + "[47] Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse reinforcement learning. In AAAI Conference on Artificial Intelligence, pages 1433-1438, 2008. 2" + ], + "bbox": [ + 501, + 92, + 890, + 402 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_model.json b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ec5270161b77941ce49fec8d32da0d92ee8ccf85 --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.13, + 0.842, + 0.176 + ], + "angle": 0, + "content": "LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.896, + 0.294 + ], + "angle": 0, + "content": "Ke Guo\\(^{1,2,3}\\), Zhenwei Miao\\(^{1*}\\), Wei Jing\\(^{1}\\), Weiwei Liu\\(^{1}\\), Weizi Li\\(^{4}\\), Dayang Hao\\(^{1}\\), Jia Pan\\(^{2,3}\\) \n\\(^{1}\\)Alibaba Group \\(^{2}\\)The University of Hong Kong \n\\(^{3}\\)Centre for Transformative Garment Production, Hong Kong \\(^{4}\\)University of Tennessee, Knoxville \nu3006612@connect.hku.hk, zhenwei.mzw@alibaba-inc.com \n21wjing@gmail.com, 11932061@zju.edu.cn, weizili@utk.edu, jpan@cs.hku.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.473, + 0.691 + ], + "angle": 0, + "content": "Microscopic traffic simulation plays a crucial role in transportation engineering by providing insights into individual vehicle behavior and overall traffic flow. However, creating a realistic simulator that accurately replicates human driving behaviors in various traffic conditions presents significant challenges. Traditional simulators relying on heuristic models often fail to deliver accurate simulations due to the complexity of real-world traffic environments. Due to the covariate shift issue, existing imitation learning-based simulators often fail to generate stable long-term simulations. In this paper, we propose a novel approach called learner-aware supervised imitation learning to address the covariate shift problem in multi-agent imitation learning. By leveraging a variational autoencoder simultaneously modeling the expert and learner state distribution, our approach augments expert states such that the augmented state is aware of learner state distribution. Our method, applied to urban traffic simulation, demonstrates significant improvements over existing state-of-the-art baselines in both short-term microscopic and long-term macroscopic realism when evaluated on the real-world dataset pNEUMA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.718, + 0.21, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.743, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Microscopic traffic simulation is a cornerstone in transportation engineering. It enables engineers to predict and analyze individual vehicle behavior, providing crucial insights into how alterations in road structures or traffic management strategies might influence overall traffic flow; it allows for the testing of diverse scenarios without disrupting real-world traffic; and it enhances safety by pinpointing potential hazards and devising strategies to mitigate risks. By leveraging simulations, engineers can optimize traffic flow," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.892, + 0.405 + ], + "angle": 0, + "content": "reduce congestion, and enhance overall efficiency, proving particularly advantageous in designing or enhancing road networks. Using the simulation data, policymakers and urban planners can make informed decisions aligned with community needs regarding transportation infrastructure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.893, + 0.651 + ], + "angle": 0, + "content": "However, generating realistic and accurate simulations that can simultaneously replicate the microscopic response of human drivers in various traffic conditions and long-term macroscopic traffic statistics is challenging. In recent years, significant efforts have been invested in developing realistic traffic simulators with the goal to accurately model human driving behaviors. Traditional traffic simulators, such as SUMO [24], AIMSUN [2], and MITSIM [44], typically rely on heuristic car-following models like the Intelligent Driver Model (IDM) [42]. However, despite carefully calibrating parameters, the rule-based models often fail to deliver accurate simulations due to the complexity of real-world traffic environments [12, 30]. Factors such as the road structure, neighboring vehicles, and even driver psychology can influence the decisions of human drivers, making it challenging to achieve accurate simulations [7, 27, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.852 + ], + "angle": 0, + "content": "In pursuit of realistic traffic simulations, researchers have turned to neural networks to represent the driving model through imitation learning (IL) from human demonstrations. Most traffic simulation approaches [4, 34, 41] leverage behavior cloning (BC) [32] to learn a driving policy by minimizing the disparity between the model output and the human demonstrations in training data. However, BC is hindered by covariate shift [37], where the state induced by the learner's policy progressively diverges from the expert's distribution. Existing BC-based simulators have succeeded in short-term (less than 20 seconds) simulation applications like autonomous driving tests but often fail to generate stable long-term simulations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address covariate shift, existing methods such as DAGger [37], DART [25], and ADAPS [28] incorporate supervisor (humans or principled simulations) corrections" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.332, + 0.901 + ], + "angle": 0, + "content": "*Zhenwei Miao is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15386" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.38 + ], + "angle": 0, + "content": "at the learner's or perturbed expert's states. However, human supervision can be problematic due to intensive labor and judgment errors, and principled simulations may not account for heterogeneous driving behaviors. Recent traffic simulators [40, 46] propose using inverse reinforcement learning (IRL). These methods, such as generative adversarial imitation learning (GAIL) [17] and adversarial inverse reinforcement learning (AIRL) [13], learn a reward function using a discriminator neural network within Generative Adversarial Networks (GANs) [14]. The policy network is trained to maximize the learned reward through online reinforcement learning (RL), enabling agents to handle out-of-distribution states. However, directly applying GAIL to traffic simulation can be problematic [5]. The dynamic nature of the environment in the multi-agent system can lead to noises during policy learning, resulting in highly biased estimated gradients. Furthermore, training the discriminator in GAIL is challenging due to the instability and sensitivity of hyperparameters during min-max optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.473, + 0.761 + ], + "angle": 0, + "content": "To address the issue of covariate shift in multi-agent imitation learning without depending on costly expert supervision or unstable discriminators and multi-agent RL, we propose Learner-Aware Supervised Imitation Learning (LASIL) for tackling the covariate shift problem during policy learning. We mitigate the distribution shift between expert and learner state distribution by augmenting the expert state distribution. However, there is no expert supervision at any augmented state, so we ensure that an augmented state is close to an expert state such that the future trajectory of the expert state can serve as the target trajectory for the augmented state to constrain the learner within the expert state distribution. Hence, our goal is to augment the expert state to cover the learner's state distribution while remaining close to the original expert state distribution. To achieve this, we use a variational autoencoder (VAE) [21] to simultaneously model the distributions of both the expert and learner states. By minimizing the VAE's latent space regularization loss of modeling both distributions, we project the expert and learner states into a unified latent space. And by minimizing the VAE's reconstruction loss, the resulting reconstruction leveraging such learnt latent space will resemble both expert and learner state distribution. As a result, when inputting expert states into the trained VAE, we obtain a learner-aware augmented expert state." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In practice, we divide an agent's state into two parts: past trajectory and context, i.e., other features in the state including vehicle type, waypoints, and destination. We observe that the distribution of the context is consistent regardless of a learnt policy due to the static features of traffic conditions embedded in the context, leading to less covariate shift. Therefore, we propose a context-conditioned VAE to model the context-conditioned trajectory distribution of both expert and learner states. The decoder of the context-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "conditioned VAE will receive both the latent variable and the context, and yield only the trajectory information." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.124, + 0.821, + 0.137 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.14, + 0.891, + 0.184 + ], + "angle": 0, + "content": "- We propose learner-aware supervised imitation learning (LASIL) to achieve stable learning and alleviate covariate shift in multi-agent imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.185, + 0.892, + 0.229 + ], + "angle": 0, + "content": "- We propose a learner-aware data augmentation method based on a context-conditioned VAE that generates learner-aware augmented expert states." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.23, + 0.892, + 0.319 + ], + "angle": 0, + "content": "- Our approach is tailored for urban traffic simulation. To the best of our knowledge, it is the first imitation learning-based traffic simulator that can reproduce long-term (more than 10 minutes) stable microscopic simulation, achieving \\(40\\mathrm{x}\\) simulation length improvements over previous state-of-the-art [4, 41, 46]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.32, + 0.892, + 0.41 + ], + "angle": 0, + "content": "- We evaluate our method on the real-world dataset pNEUMA [3] with over half a million trajectories. Our approach outperforms state-of-the-art baselines in both short-term microscopic and long-term macroscopic simulations. The code is available at https://github.com/Kguo-cs/LSAIL." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.14, + 0.892, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.427, + 0.642, + 0.443 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.453, + 0.677, + 0.469 + ], + "angle": 0, + "content": "2.1. Imitation learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Existing imitation learning (IL) methods can be broadly categorized into behavior cloning (BC) and inverse reinforcement learning (IRL) approaches. BC [32] learns a policy in a supervised fashion by minimizing the discrepancy between the learner's actions and those of an expert. However, BC suffers from the issue of covariate shift, where the state distribution induced by the learner's policy gradually deviates from that of the expert. To address this, methods like DAGger [37] and DART [25] request supervisor corrections at the learner's or perturbed expert's states. Our method follows a similar supervised learning approach to DAGger and DART, but does not require access to an expert policy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Due to the challenges in obtaining expert supervision, recent IRL-based methods utilize feedback from a neural network-based discriminator to handle covariate shift. Typically, these methods involve an iterative process alternating between reward estimation and reinforcement learning. Earlier IRL-based methods [15, 35, 39, 47] require frequent dynamic programming processes, while recent adversarial IL approaches integrate reward function learning with policy learning using a GAN formulation. However, both GAN and RL training processes are known to be unstable, sensitive to hyperparameters, and have poor sample efficiency [9]. Moreover, the discriminator can easily exploit insignificant differences between expert and policy samples, leading to undesirable performance [45]. In contrast, our method avoids the min-max optimization problem and the sample-inefficient RL process, requiring minimal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "fine-tuning. Instead, we utilize the real future trajectory as the target state to learn corrective behavior." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.133, + 0.453, + 0.149 + ], + "angle": 0, + "content": "2.2. Imitation Learning-based Traffic simulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.156, + 0.469, + 0.217 + ], + "angle": 0, + "content": "Recent traffic simulators focus on enhancing realism by leveraging imitation learning (IL) from human driving demonstrations, which extend traditional BC and IRL methods to tackle the challenging multi-agent IL problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.469, + 0.428 + ], + "angle": 0, + "content": "BC-based traffic simulators like TrafficSim [41] and SimNet [4] typically begin by training a prediction model and subsequently adjust the predicted trajectories to prevent collisions and adhere to traffic regulations during simulation. However, BC-based methods face challenges in achieving long-term simulation due to the covariate shift problem. To mitigate this issue, we augment expert state based on the policy distribution, enabling stable long-term simulation. Additionally, we enhance performance by modifying the predicted trajectory during simulation through road projection and ensuring smoothness from the current state. Notably, we skip the computationally intensive collision removal operation, prioritizing the capture of macroscopic long-term influence over micro-level details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.761 + ], + "angle": 0, + "content": "IRL-based simulators learn the underlying reward function of human driving behavior and derive the driving policy by maximizing the learned reward. While adversarial IRL methods theoretically address the covariate shift of BC in a single-agent context through online interaction, its performance deteriorates in the multi-agent IL domain due to the dynamic environment, complicating the training process. To tackle this challenge, approaches like PS-GAIL [40] and PS-ARIL [46] adopt two-stage learning and gradually introduce vehicles to the environment. Nonetheless, these methods still exhibit significant undesirable traffic phenomena, such as off-road driving, collisions, and abrupt braking. Building upon PS-GAIL, the reward-augmented imitation learning (RAIL) method [5, 22] penalizes undesirable phenomena by introducing a hand-crafted reward. However, maximizing the new reward does not guarantee the recovery of human-like trajectories. Despite numerous enhancements to the original GAIL framework, these IRL-based methods often struggle to produce stable long-term traffic flow, as evidenced in our experiments. In contrast, our method is a supervised learning approach, leading to faster, simpler, and more stable learning of driving policies." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.777, + 0.206, + 0.793 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.802, + 0.307, + 0.817 + ], + "angle": 0, + "content": "3.1. Markov Decision Process" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We model the human driving process using a Markov decision process (MDP) denoted as \\(\\mathcal{M} = \\{\\mathcal{S},\\mathcal{A},\\mathcal{T},r\\}\\), incorporating a time horizon of \\(T\\), where \\(\\mathcal{S}\\) and \\(\\mathcal{A}\\) represent the continuous state and action spaces respectively. A stochastic function \\(\\mathcal{T}:S\\times \\mathcal{A}\\times \\mathcal{S}\\rightarrow [0,1]\\) describes the system" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "dynamics, and \\(r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}\\) is a reward function. The policy \\(\\pi(\\boldsymbol{a} | s)\\) determines the probability of selecting an action \\(a\\) at a state \\(s\\), and a trajectory \\(\\tau = (s_0, \\boldsymbol{a}_0, \\dots, s_T, \\boldsymbol{a}_T)\\) represents a sequence of state-action pairs. The marginal state distribution of a policy \\(\\pi\\) is computed as \\(\\rho^{\\pi}(s) = \\frac{1}{D} \\sum_{s \\in \\mathcal{D}_{\\pi}} \\delta_s\\), where \\(\\mathcal{D}_{\\pi}\\) denotes the set of states of size \\(D\\) in trajectories induced by the policy \\(\\pi\\), and \\(\\delta_s\\) signifies a Dirac distribution centered on \\(s\\). Similarly, the marginal state-action distribution \\(\\rho^{\\pi}(s, \\boldsymbol{a})\\) is computed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.258, + 0.684, + 0.274 + ], + "angle": 0, + "content": "3.2. Imitation Learning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.472 + ], + "angle": 0, + "content": "In the application of IL for learning a driving policy, we assume that all agents adopt the same policy. The objective is to learn a policy \\(\\pi\\) that minimizes the \\(f\\)-divergence between the marginal state-action distribution of the expert's demonstrations \\(\\rho^{\\exp}(s,a)\\) and the learner policy's distribution \\(\\rho^{\\pi}(s,a)\\). For example, BC optimizes the policy to minimize the Kullback-Leibler (KL) divergence at the expert state distribution \\(\\mathbb{E}_{\\rho^{\\exp}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]\\), while DAgger minimizes \\(\\mathbb{E}_{\\rho^{\\pi}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]\\). On the other hand, GAIL minimizes the Jensen-Shannon divergence \\(D_{\\mathrm{JS}}(\\rho^{\\exp}(s,a)\\| \\rho^{\\pi}(s,a))\\), while AIRL minimizes the KL divergence \\(\\mathrm{KL}(\\rho^{\\pi}(s,a)\\| \\rho^{\\exp}(s,a))\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.892, + 0.734 + ], + "angle": 0, + "content": "However, GAIL and AIRL often exhibit unsatisfactory practical performance due to their optimization processes' instability and sample inefficiency, which involve GANs and RL. In contrast, BC only requires simple and stable supervised learning but suffers from the covariate shift issue, as it only minimizes the policy difference at the expert state distribution without guaranteeing performance at the learner state distribution. DAGger addresses the covariate shift problem but requires access to the expert policy. To mitigate the covariate shift problem without depending on the expert policy, we propose maximizing the transition probability \\(\\mathbb{E}_{\\boldsymbol{s} \\sim \\rho^{\\exp}(\\boldsymbol{s}), \\boldsymbol{s} + \\epsilon \\sim \\rho^{\\pi}(\\boldsymbol{s})}[\\mathcal{T}(\\boldsymbol{s} + \\epsilon, \\pi(\\boldsymbol{s} + \\epsilon), \\boldsymbol{s}')]\\), where \\(\\epsilon\\) is a small augmentation term. Note that our policy learns to predict the distribution of the next state \\(s'\\) instead of the action. Our approach assumes that when the expert's original future trajectory can serve as supervision to guide the agent back towards the expert distribution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.765, + 0.724, + 0.78 + ], + "angle": 0, + "content": "3.3. Variational Autoencoder" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The VAE defines a generative model given by \\( p_{\\theta}(\\pmb{x}, \\pmb{z}) = p(\\pmb{z}) p_{\\theta}(\\pmb{x} | \\pmb{z}) \\), where \\( \\pmb{z} \\) is the latent variable with prior distribution \\( p(\\pmb{z}) \\), and \\( p_{\\theta}(\\pmb{x} | \\pmb{z}) \\) represents the conditional distribution modeling the likelihood of data \\( \\pmb{x} \\) given \\( \\pmb{z} \\). The learning objective is to maximize the training samples' marginal log-likelihood \\( \\log p_{\\theta}(\\pmb{x}) \\). However, due to the intractability of marginalization, VAE maximizes the variational lower" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.414, + 0.108 + ], + "angle": 0, + "content": "bound using \\(q_{\\phi}(\\pmb {z}|\\pmb {x})\\) as the approximate posterior:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.117, + 0.469, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {x}) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x})} [ \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) ] - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x}) \\| p (\\boldsymbol {z})\\right) \\tag {1} \\\\ := - \\mathcal {L} _ {\\text {r e c}} (\\boldsymbol {x}) - \\mathcal {L} _ {\\mathrm {K L}} (\\boldsymbol {x}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.47, + 0.289 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{rec}}(\\boldsymbol{x})\\) represents the reconstruction loss, which penalizes the network for creating outputs different from the input. \\(\\mathcal{L}_{\\mathrm{KL}}(\\boldsymbol{x})\\) represents the KL divergence loss to make a continuous and smooth latent space, allowing easy random sampling and interpolation. Intuitively, this KL loss encourages the encoder to distribute all encodings evenly around the center of the latent space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.3, + 0.169, + 0.315 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.508 + ], + "angle": 0, + "content": "In Fig. 1, we present an overview of our method. Our model comprises three modules: a VAE-based data augmentation module, a policy network, and a post-processing (LQR and on-road projection) module. During training, we augment each expert data by generating a learner-aware augmented expert state through the VAE's reconstruction of its past trajectory. Using this augmented state along with the original future trajectory, we train the learner's policy network through supervised learning. During simulation, we roll out the policy network with several post-processing steps including sampling, projecting the trajectories onto the road, and smoothing the projected trajectories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.274, + 0.531 + ], + "angle": 0, + "content": "4.1. State Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.469, + 0.584 + ], + "angle": 0, + "content": "As human drivers make decisions mainly depending on their surrounding information, we build a graph to model the traffic system with each driving agent as its node." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.469, + 0.795 + ], + "angle": 0, + "content": "Node: The state of each agent can be divided into two components: the past trajectory \\( s_p \\) and the context \\( s_c \\). In practice, the past trajectory is composed of the agent's positions at several past time steps, including the current time step. The context includes its type, nearby waypoints' positions and corresponding road width, the traffic light status of the road it is traveling on, and destination. The waypoints are composed of the center points of the routing roads with a fixed interval. We transform each agent's state into its individual coordinate system to learn a policy with transformation invariance. To reduce the implicit covariate shift, we set each agent's current position adding a Gaussian perturbation as the origin and the \\( x \\)-axis directions pointing towards the agent's destination like [16]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Edge: When each agent's state coordinates are transformed from the global coordinate system to their individual coordinate system, information about the relative positions among agents is lost. However, a traffic model needs the relative information of agents to understand how they interact with each other. To preserve these relationships, we introduce directed edges between neighboring agents. The" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "edge feature is the relative position of the destination node's coordinate origin in the source node's coordinate system." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.131, + 0.806, + 0.147 + ], + "angle": 0, + "content": "4.2. Learner-aware Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.154, + 0.892, + 0.275 + ], + "angle": 0, + "content": "To address the covariate shift issue between the expert and learner state distributions, we propose to minimize the expert's and learner's embedded state distribution difference. In practice, we propose utilizing the same VAE to model the expert and learner state distributions simultaneously. As both the expert and learner state are projected to the same latent space, the distribution of the state reconstructed from the joint latent space can resemble both distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.487 + ], + "angle": 0, + "content": "In contrast to the past trajectory, the context distribution is more challenging to model but exhibits less covariate shift. Therefore, we propose using a context-conditioned VAE to specifically model the context-conditioned trajectory distribution rather than the state distribution. For each expert or learner state represented as \\( s = (s_p, s_c) \\), we employ an encoder \\( q_{\\phi}(z | s_p, s_c) \\) to obtain the latent variable distribution. We can sample a latent variable \\( z \\) from this distribution by applying the reparameterization trick. Subsequently, a decoder \\( p_{\\theta}(s_p | z, s_c) \\) is used to reconstruct the distribution of the past trajectory \\( s_p \\) given \\( z \\) and \\( s_c \\). The VAE is trained to maximize the variational lower bound, which incorporates the context-conditioned log-likelihood of both the expert and learner past trajectories:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.498, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V A E} = \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\exp}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right] + \\lambda \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\pi}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Here, \\(\\lambda\\) is a hyperparameter that controls the degree to which the augmented context-conditioned past trajectory distribution aligns with the learner's context-conditioned past trajectory distribution. The term \\(\\mathcal{L}(s_p|s_c)\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.599, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) = \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} \\left(\\boldsymbol {z} \\mid \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}\\right)} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {z}, \\boldsymbol {s} _ {c}\\right) \\right] \\\\ - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}) \\| p (\\boldsymbol {z})\\right). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.694 + ], + "angle": 0, + "content": "For simplicity, we assume \\(p_{\\theta}\\) and \\(q_{\\phi}\\) as multivariate normal distributions with diagonal variance matrices. The prior distribution \\(p(z)\\) is set as an isotropic unit Gaussian \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.701, + 0.858, + 0.718 + ], + "angle": 0, + "content": "4.3. Edge-enhanced Graph Attention Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.892, + 0.846 + ], + "angle": 0, + "content": "We build all models in our approach including the VAE encoder, decoder, and learner's policy network based on an edge-enhanced graph attention (EGAT) network [11, 33], which model interactions by aggregating neighboring node and edge information using an attention mechanism. The node features in the first layer are obtained by embedding the node input with a fully connected layer and the node features in the other layers are calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.856, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {i} ^ {l} = \\sigma \\left(\\sum_ {j \\in \\mathcal {N} _ {i}} \\alpha_ {i j} ^ {l} \\boldsymbol {W} ^ {l} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right), \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15389" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.381 + ], + "angle": 0, + "content": "Figure 1. Overview of our approach. The processes with purple and red arrows handle only expert or learner data, respectively, while the processes represented by black arrows apply to both data. Each state is a multi-agent state represented by a graph. Each model, including the VAE encoder, decoder, and policy network, is implemented using an EGAT network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.474, + 0.508 + ], + "angle": 0, + "content": "where \\( h_i^l \\) is the feature of node \\( i \\) in the \\( l \\)th layer, \\( e_{ij} \\) is the coordinate origin's position of node \\( j \\) relative to node \\( i \\), \\( \\| \\cdot \\| \\) means the concatenation, \\( W^l \\) is the learnable weight matrix, \\( \\mathcal{N}_i \\) is the set of the first-order neighbors of node \\( i \\) (including the node itself), and \\( \\sigma \\) is a non-linear activation function. The attention coefficient \\( \\alpha_{ij}^l \\) indicates the importance of node \\( j \\) to node \\( i \\), considering both node and edge features, and is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.517, + 0.47, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i j} ^ {l} = \\operatorname {s o f t m a x} _ {j} \\left(\\sigma \\left(\\left(\\boldsymbol {w} ^ {l}\\right) ^ {T} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.556, + 0.47, + 0.633 + ], + "angle": 0, + "content": "where \\(\\boldsymbol{w}^l\\) is a learnable weight vector, and normalization is performed on the weights across all neighbors of node \\(i\\) using a softmax function. After passing through multiple EGAT layers, the node features at the last layer is fed into a fully connected layer to obtain the outputs of the network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.202, + 0.659 + ], + "angle": 0, + "content": "4.4. Policy Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.757 + ], + "angle": 0, + "content": "To train the learner's policy network, we first sample a learner-aware augmented expert state from the context-conditioned VAE for each expert state. Then, the policy network predicts its future position distribution over \\( T \\) time steps, denoted as \\( p(\\hat{p}_1^m, \\hat{p}_2^m, \\dots, \\hat{p}_T^m) \\), which is assumed to be a product of multi-variable Gaussian distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.77, + 0.47, + 0.812 + ], + "angle": 0, + "content": "\\[\np \\left(\\hat {\\boldsymbol {p}} _ {1} ^ {m}, \\hat {\\boldsymbol {p}} _ {2} ^ {m}, \\dots , \\hat {\\boldsymbol {p}} _ {T} ^ {m}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{\\mu}}_t^m\\) and \\(\\hat{\\Sigma}_t^m\\) represent the mean and covariance matrix of the predicted position \\(\\hat{\\pmb{p}}_t^m\\) at future time step \\(t\\). We assume that there is no correlation between the position distributions at different future time steps. To learn the policy network, we minimize the negative log-likelihood (NLL)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.831, + 0.402 + ], + "angle": 0, + "content": "loss of all agents' ground-truth future trajectories:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.41, + 0.892, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {N L L} = - \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\log \\left(\\mathcal {N} \\left(\\boldsymbol {p} _ {t} ^ {m} - \\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.894, + 0.49 + ], + "angle": 0, + "content": "where \\( \\pmb{p}_t^m \\) denotes the ground truth position of agent \\( m \\) at future time step \\( t \\), and \\( M \\) is the total number of agents." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.684, + 0.512 + ], + "angle": 0, + "content": "4.5. Simulation Process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.672 + ], + "angle": 0, + "content": "During training, we simultaneously simulate (roll out) the learned policy network iteratively to get learner state samples. Instead of directly updating each agent to its predicted position, we apply several post-processing steps to the prediction for better realism. Firstly, we sample from the distribution, and then project each sampled position onto the nearest on-road point. Then, we smooth the projected trajectory with a linear-quadratic regulator (LQR) [1] by minimizing the total commutative quadratic cost of a linear dynamic system described by:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.679, + 0.892, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t + 1} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t + 1} ^ {m} \\end{array} \\right] = \\left[ \\begin{array}{c c} \\boldsymbol {I} & \\boldsymbol {D} \\\\ 0 & \\boldsymbol {I} \\end{array} \\right] \\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t} ^ {m} \\end{array} \\right] + \\left[ \\begin{array}{c} \\boldsymbol {D} ^ {2} \\\\ \\boldsymbol {D} \\end{array} \\right] \\tilde {\\boldsymbol {a}} _ {t} ^ {m}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.721, + 0.894, + 0.782 + ], + "angle": 0, + "content": "where \\(D\\) is a diagonal matrix with the interval of each time step as diagonal entries, and \\(\\tilde{\\boldsymbol{p}}_t^m\\), \\(\\tilde{\\boldsymbol{v}}_t^m\\), \\(\\tilde{\\boldsymbol{a}}_t^m\\) represent the LQR-planned position, velocity, and acceleration. The system is subject to a quadratic cost function given by:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.79, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\mathcal {J} = \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\| \\tilde {\\boldsymbol {p}} _ {t} ^ {m} - \\bar {\\boldsymbol {p}} _ {t} ^ {m} \\| ^ {2} + \\eta_ {\\boldsymbol {a}} \\| \\tilde {\\boldsymbol {a}} _ {t} ^ {m} \\| ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "where the projected predicted position \\(\\bar{\\pmb{p}}_m^t\\) is considered as the target pose, and the hyper-parameter \\(\\eta_{\\mathbf{a}}\\) is used to penalize high acceleration. Finally, each agent is updated to the first position of the planned trajectory." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.245, + 0.108 + ], + "angle": 0, + "content": "4.6. Training Process" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.391 + ], + "angle": 0, + "content": "At each training step, the policy network is trained to maximize the probability of the expert's future trajectory using its context state and history trajectory augmented by the VAE as input. Simultaneously, the VAE is iteratively trained to reconstruct the expert and learner data. The expert data are the same as the policy network's training data before augmentation. The learner data is sampled from a replay buffer. Every \\( N \\) training steps, we empty the replay buffer and roll out the current policy with post-processing for \\( S \\) time steps to generate the learner data and store it in the replay buffer. During the roll-out, we start at a random time step in the training dataset and apply our model in a closed-loop manner. At each time step during the rollout, we sample a future trajectory for each vehicle from the Gaussian distribution predicted by our policy network, project the predicted trajectory onto the road, smooth the on-road trajectory with LQR, and finally update the vehicle to the first position of the smoothed trajectory." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.41, + 0.203, + 0.427 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.437, + 0.173, + 0.451 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.473, + 0.706 + ], + "angle": 0, + "content": "We utilize a real-world urban dataset called pNEUMA [3], which was collected by 10 drones in Athens over a span of 4 days. The dataset encompasses over half a million vehicle trajectories within a large area encompassing over \\(100\\mathrm{km}\\) of lanes and approximately 100 intersections. The recordings were conducted at 5 intervals each day, with each period lasting about 15 minutes and a data collection time interval of 0.04 seconds. To enhance computational efficiency, we adopt a time step of 0.4 seconds. The dataset is divided into a training set, comprising recordings from the first 3 days, and a validation/test set, consisting of recordings from the final day. Notably, we choose not to utilize other popular traffic datasets like NGSIM [8] or HighD [23] or autonomous driving datasets like Lyft [18] or nuPlan [6], as they only encompass samll-scale scenarios. This limitation makes it inadequate for evaluate macroscopic realism." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.721, + 0.174, + 0.736 + ], + "angle": 0, + "content": "5.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.839 + ], + "angle": 0, + "content": "We evaluate the realism of our simulator by measuring the similarity between the simulation result and real data. During evaluation, each vehicle enters the simulator at its first recorded time and position, and is controlled by our simulator to complete its recorded route. When an agent reaches its final recorded position, it is removed from the simulator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Short-term microscopic: Following [5, 40], we conduct a short-term microscopic evaluation by simulating for 20 seconds from a random time step in the test dataset. We first measure the similarity between the simulated and real" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.887, + 0.108 + ], + "angle": 0, + "content": "data using position and velocity RMSE metrics given by:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.117, + 0.892, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R M S E} = \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\sqrt {\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t} ^ {m} \\| ^ {2}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.894, + 0.266 + ], + "angle": 0, + "content": "where \\( s_t^m \\) and \\( \\hat{s}_t^m \\) were the real and simulated value of the position or velocity of the agent \\( m \\) at time step \\( t \\), respectively. \\( T_s \\) was the total simulated time steps, and \\( M \\) was the total simulated agent number. Besides, we measure the minimum Average Displace Error (minADE) to not penalize reasonable trajectory unlike the real one:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.275, + 0.892, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\min A D E _ {N} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\min _ {\\hat {\\boldsymbol {s}} _ {n}} \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t, n} ^ {m} \\| ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.416 + ], + "angle": 0, + "content": "where \\( N = 20 \\) is roll-out times. We also calculate the off-road rate (the avarage proportion of vehicles that deviate more than \\( 1.5\\mathrm{m} \\) from the road over all time steps). The common collision rate metric is not used because we focus on the long-term impact and the dataset does not provide accurate vehicle size and heading information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.612 + ], + "angle": 0, + "content": "Long-term macroscopic: we also evaluate our model's long-term macroscopic accuracy on five periods for 800 seconds from its initial recording time. To measure the long-term performance, we use two standard macroscopic metrics for traffic flow data [29, 36, 38], namely road density and speed RMSE, in addition to the off-road rate. The density of a road at a time step is calculated by dividing the number of vehicles on the road by its total lane length, assuming that all lanes have the same width. Meanwhile, the road speed is computed as the mean speed of all vehicles on the road. To quantify the similarity between the simulated and ground truth values, we still use RMSE, where the variable \\( M \\) becomes the total number of roads." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.637, + 0.636 + ], + "angle": 0, + "content": "5.3. Performance" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.645, + 0.885, + 0.659 + ], + "angle": 0, + "content": "We compare our method against state-of-the-art baselines:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.735 + ], + "angle": 0, + "content": "SUMO [24]: we use the IDM model [42] as the car-following model and mobil [19] as the lane-changing model. We tune the IDM's parameters for 6 types of vehicle by minimizing the MSE between the IDM calculated acceleration and real acceleration using an Adam optimizer [20]." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.735, + 0.891, + 0.75 + ], + "angle": 0, + "content": "BC [32]: we learn our model directly by the BC method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.825 + ], + "angle": 0, + "content": "MARL [26]: we train our model using IPPO [10] as the Multi-Agent RL algorithm, where the reward function is composed of three parts: a displacement reward (average distance between agent trajectory with GT trajectory), an off-road penalty and a terminal reward." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.855 + ], + "angle": 0, + "content": "MARL+BC [31]: we add a behavior cloning term in the loss function while learning the policy with MARL." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "PS-GAIL [40]: we let all vehicles share the same policy parameter and critic parameter and learn the policy using reward functions computed by GAIL." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15391" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.202, + 0.09, + 0.768, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison with baselines and ablated models on microscopic metrics for 20 seconds." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.104, + 0.849, + 0.3 + ], + "angle": 0, + "content": "
ModelPosition RMSE(m)Velocity RMSE(m/s)minADE20(m)Off-road(%)
SUMO [24]41.257.0024.200
BC [32]40.08±1.616.74±0.3722.02±3.9833.43±3.10
MARL [26]48.66±1.548.92±0.6740.07±1.263.70±0.40
MARL+BC [31]29.78±0.424.54±0.1422.07±0.2111.14±1.17
PS-GAIL [40]34.78±0.435.33±0.1227.86±0.337.65±0.32
RAIL [5]31.62±0.655.51±0.0924.51±0.412.57±0.47
LASIL (ours)19.21±0.443.02±0.0712.79±0.320.28±0.01
w/o Augmentation21.62±0.543.34±0.1413.81±0.371.31±0.18
w/o Context-conditioned23.03±0.873.73±0.1214.94±0.430.51±0.04
w/o On-road Projection22.53±0.423.26±0.1014.42±0.331.74±0.60
w/o LQR22.62±0.434.05±0.1514.02±0.280.22±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.197, + 0.305, + 0.773, + 0.318 + ], + "angle": 0, + "content": "Table 2. Comparison with baselines and ablated models on macroscopic metrics for 800 seconds." + }, + { + "type": "table", + "bbox": [ + 0.13, + 0.32, + 0.842, + 0.516 + ], + "angle": 0, + "content": "
ModelRoad Density RMSE(veh/km)Road Speed RMSE(m/s)Off-road(%)
SUMO [24]52.705.520
BC [32]61.51±1.535.38±0.2142.15±5.25
MARL [26]90.59±2.655.27±0.484.63±0.64
MARL+BC [31]43.60±2.804.01±0.1322.73±5.62
PS-GAIL [40]54.06±1.234.03±0.0513.24±3.20
RAIL [5]54.45±1.893.89±0.112.92±0.38
LASIL (ours)45.13±0.253.17±0.140.34±0.03
w/o Augmentation55.65±0.443.73±0.239.62±1.40
w/o Context-conditioned55.98±0.514.48±0.190.64±0.08
w/o On-road Projection52.90±0.543.59±0.2011.33±2.37
w/o LQR61.68±0.963.54±0.160.47±0.07
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.47, + 0.551 + ], + "angle": 0, + "content": "RAIL [5]: we learn the model as PS-GAIL but with additional displacement, off-road and terminal rewards." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.673 + ], + "angle": 0, + "content": "We train and evaluate each model five times to obtain the mean and standard deviation (std) of various metrics. Note that we do not apply on-road projection and LQR for baselines. We evaluate both short-term and long-term performance, as shown in Tab. 1 and Tab. 2, respectively. Our method achieves better results than all baselines in terms of position and velocity RMSE, road density and speed RMSE, with minor off-road rate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.231, + 0.699 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We conducted a series of ablation experiments to assess the individual contributions of crucial components in our approach, whose results are presented in Tab. 1 and Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Augmentation: By removing our VAE module and directly learning the original expert state and action, we analyze the importance of our data augmentation technique. The results demonstrate that augmentation plays a vital role in improving both short-term and long-term simulation performance by mitigating covariate shift." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Context-conditioned VAE: To emphasize the significance of modeling the context-conditioned trajectory distribution rather than the whole state distribution, we replaced our context-conditioned VAE with a naive VAE that directly" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.567 + ], + "angle": 0, + "content": "models the expert and learner state distribution. The observed drop in performance demonstrates the challenges in reconstructing the context distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.648 + ], + "angle": 0, + "content": "On-road projection: Our ablation study on the on-road projection module aims to show its impact on reducing the off-road rate. The results show that the on-road projection module leads to a notable decrease in the off-road rate and moderate improvements in other performance metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.743 + ], + "angle": 0, + "content": "LQR: Removing LQR module allows us to evaluate its effectiveness. While LQR led to a higher short-term off-road rate due to more constrained movement (driving off-road due to inertia), its removal deteriorates other short-term and long-term metrics, because LQR can smooth the trajectory, thus leads to more realistic simulations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.767, + 0.679, + 0.783 + ], + "angle": 0, + "content": "5.5. Qualitative Result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In Fig. 2, we present the mean road density and speed for real-world data, SUMO simulation, and our proposed method over all time steps. The results demonstrate that our proposed method surpasses the capabilities of the SUMO simulator at replicating long-term macroscopic traffic patterns due to our model's enhanced ability to replicate the typical microscopic driving behaviors over a long period." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15392" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.862, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.647, + 0.892, + 0.675 + ], + "angle": 0, + "content": "Figure 2. Mean density and speed on each road over all time steps in the long-term evaluation. Our method's density and speed hot-maps have a more similar color to the ground-truth one compared with the SUMO's." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.196, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we have addressed the challenge of creating a realistic traffic simulator that accurately models human driving behaviors in various traffic conditions. Traditional imitation learning-based simulators often fail to deliver accurate long-term simulations due to the covariate shift problem in multi-agent imitation learning. To tackle this issue, we proposed a learner-aware supervised imitation learning method, which leverages a context-conditioned VAE to generate learner-aware augmented expert states. We leverage a context-conditioned VAE to simultaneously reconstruct the expert and learner state. This approach enables the reproduction of long-term stable microscopic traffic simulations, marking a significant advancement in the field of urban" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.755 + ], + "angle": 0, + "content": "traffic simulation. Our method has demonstrated superior performance over existing state-of-the-art simulators when evaluated on the real-world dataset pNEUMA, achieving better short-term microscopic and long-term macroscopic similarity to real-world data than state-of-the-art baselines." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.688, + 0.788 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research project is partially supported by the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, Hong Kong General Research Fund (11202119 and 11208718), Innovation and Technology Commission (GHP/126/21GD), and Guangdong, Hong Kong and Macao Joint Innovation Project (2023A0505010016)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15393" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Karl Johan Åström and Richard M Murray. Feedback systems: an introduction for scientists and engineers. Princeton University Press, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Jaime Barceló and Jordi Casas. Dynamic network simulation with aimsun. Simulation approaches in transportation analysis: Recent advances and challenges, pages 57-98, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Emmanouil Barmpounakis and Nikolas Geroliminis. On the new era of urban traffic monitoring with massive drone data: The pNEUMA large-scale field experiment. Transportation research part C: emerging technologies, 111:50-71, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[4] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Blazej Osinski, Hugo Grimmett, and Peter Ondruska. SimNet: Learning reactive self-driving simulations from real-world observations. IEEE International Conference on Robotics and Automation, pages 5119-5125, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Raunak P. Bhattacharyya, Derek J. Phillips, Changliu Liu, Jayesh K. Gupta, K. Driggs-Campbell, and Mykel J. Kochenderfer. Simulating emergent properties of human driving behavior using multi-agent reward augmented imitation learning. International Conference on Robotics and Automation, pages 789-795, 2019. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[6] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. In IEEE/CVF Computer Vision and Pattern Recognition Conference ADP3 workshop, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.604 + ], + "angle": 0, + "content": "[7] Qianwen Chao, Huikun Bi, Weizi Li, Tianlu Mao, Zhaoqi Wang, Ming C Lin, and Zhigang Deng. A survey on visual traffic simulation: Models, evaluations, and applications in autonomous driving. In Computer Graphics Forum, pages 287-308, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.605, + 0.468, + 0.634 + ], + "angle": 0, + "content": "[8] James Colyar and John Halkias. Us highway 101 dataset. Federal Highway Administration, Tech. Rep., 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[9] Robert Dadashi, L'essonard Hussenot, Matthieu Geist, and Olivier Pietquin. Primal wasserstein imitation learning. In International Conference of Learning Representation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[10] Christian Schroeder de Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[11] Frederik Diehl, Thomas Brunner, Michael Truong Le, and Alois Knoll. Graph neural networks for modelling traffic participant interaction. In IEEE Intelligent Vehicles Symposium, pages 695-701, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X. Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature Communications, 12, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Justin Fu, Katie Luo Luo, and Sergey Levine. Learning robust rewards with adversarial inverse reinforcement learn" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ing. In International Conference of Learning Representation, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[15] Ke Guo, Wenxi Liu, and Jia Pan. End-to-end trajectory distribution prediction based on occupancy grid maps. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2232-2241, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Ke Guo, Wei Jing, Junbo Chen, and Jia Pan. CCIL: Context-conditioned imitation learning for urban driving. In Robotics: Science and Systems, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[17] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[18] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference of Robot Learning, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[19] Arne Kesting, Martin Treiber, and Dirk Helbing. General lane-changing model MOBIL for car-following models. Transportation Research Record, 1999(1):86-94, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.435, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.491 + ], + "angle": 0, + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.562 + ], + "angle": 0, + "content": "[22] Yann Koeberle, Stefano Sabatini, Dzmitry V. Tsishkou, and Christophe Sabourin. Exploring the trade off between human driving imitation and safety for traffic simulation. IEEE International Conference on Intelligent Transportation Systems, pages 779-786, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[23] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In International Conference on Intelligent Transportation Systems, pages 2118-2125, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of SUMO-Simulation of Urban MObility. International journal on advances in systems and measurements, 5(3&4), 2012. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on Robot Learning, pages 143-156, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Quanyi Li, Zhenghao Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in Neural Information Processing Systems, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Weizi Li, David Wolinski, and Ming C. Lin. City-scale traffic animation using statistical learning and metamodel-based optimization. ACM Trans. Graph., 36(6):200:1-200:12, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15394" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Weizi Li, David Wolinski, and Ming C. Lin. ADAPS: Autonomous driving via principled simulations. In IEEE International Conference on Robotics and Automation (ICRA), pages 7625-7631, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.15, + 0.471, + 0.217 + ], + "angle": 0, + "content": "[29] Michael James Lighthill and Gerald Beresford Whitham. On kinematic waves ii. a theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229(1178):317-345, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[30] Weiwei Liu, Wei Jing, Ke Guo, Gang Xu, Yong Liu, et al. Traco: Learning virtual traffic coordinator for cooperation with multi-agent reinforcement learning. In Conference on Robot Learning, pages 2465-2477, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[31] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. arXiv preprint arXiv:2212.11419, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[32] Donald Michie, Michael Bain, and Jean Hayes-Michie. Cognitive models from subcognitive skills. IEEE control engineering series, 44:71-99, 1990. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.46 + ], + "angle": 0, + "content": "[33] Xiaoyu Mo, Zhiyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network. IEEE Transactions on Intelligent Transportation Systems, 23:9554-9567, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[34] Jeremy Morton, Tim A Wheeler, and Mykel J Kochenderfer. Analysis of recurrent neural networks for probabilistic modeling of driver behavior. IEEE Transactions on Intelligent Transportation Systems, 18(5):1289-1298, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.47, + 0.56 + ], + "angle": 0, + "content": "[35] Andrew Y Ng, Stuart Russell, et al. Algorithms for inverse reinforcement learning. In International Conference of Machine Learning, page 2, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[36] Paul I Richards. Shock waves on the highway. Operations research, 4(1):42-51, 1956. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[37] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics, pages 627-635, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[38] Jason Sewall, David Wilkie, Paul Merrell, and Ming C Lin. Continuum traffic simulation. In Computer Graphics Forum, number 2, pages 439-448, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[39] Yu Shen, Weizi Li, and Ming C. Lin. Inverse reinforcement learning with hybrid-weight trust-region optimization and curriculum learning for autonomous maneuvering. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7421-7428, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[40] Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. Advances in neural information processing systems, 31, 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[41] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multiagent behaviors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400-10409, 2021. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[42] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[43] David Wilkie, Jason Sewall, Weizi Li, and Ming C. Lin. Virtualized traffic at metropolitan scales. Frontiers in Robotics and AI, 2:11, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[44] QI Yang and Haris N Koutsopoulos. A microscopic traffic simulator for evaluation of dynamic traffic management systems. Transportation Research Part C: Emerging Technologies, 4(3):113-129, 1996. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[45] Kaifeng Zhang, Rui Zhao, Ziming Zhang, and Yang Gao. Auto-encoding adversarial imitation learning. arXiv preprint arXiv:2206.11004, abs/2206.11004, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[46] Guanjie Zheng, Hanyang Liu, Kai Xu, and Zhenhui Jessie Li. Objective-aware traffic simulation via inverse reinforcement learning. In International Joint Conference on Artificial Intelligence, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[47] Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse reinforcement learning. In AAAI Conference on Artificial Intelligence, pages 1433-1438, 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15395" + } + ] +] \ No newline at end of file diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_origin.pdf b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b812a082dea83eeb334b1d2a73827f3bdc021d9 --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/44a62bf3-d838-46a8-bbd5-a317e9666352_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:965f5dee7359b576ac5f51b9633802d628b49a54cdb978433b6f0cd593b276a1 +size 725384 diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/full.md b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09119301a63872b0e2b778f85382de3593a942a5 --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/full.md @@ -0,0 +1,309 @@ +# LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation + +Ke Guo $^{1,2,3}$ , Zhenwei Miao $^{1*}$ , Wei Jing $^{1}$ , Weiwei Liu $^{1}$ , Weizi Li $^{4}$ , Dayang Hao $^{1}$ , Jia Pan $^{2,3}$ $^{1}$ Alibaba Group $^{2}$ The University of Hong Kong + $^{3}$ Centre for Transformative Garment Production, Hong Kong $^{4}$ University of Tennessee, Knoxville +u3006612@connect.hku.hk, zhenwei.mzw@alibaba-inc.com +21wjing@gmail.com, 11932061@zju.edu.cn, weizili@utk.edu, jpan@cs.hku.hk + +# Abstract + +Microscopic traffic simulation plays a crucial role in transportation engineering by providing insights into individual vehicle behavior and overall traffic flow. However, creating a realistic simulator that accurately replicates human driving behaviors in various traffic conditions presents significant challenges. Traditional simulators relying on heuristic models often fail to deliver accurate simulations due to the complexity of real-world traffic environments. Due to the covariate shift issue, existing imitation learning-based simulators often fail to generate stable long-term simulations. In this paper, we propose a novel approach called learner-aware supervised imitation learning to address the covariate shift problem in multi-agent imitation learning. By leveraging a variational autoencoder simultaneously modeling the expert and learner state distribution, our approach augments expert states such that the augmented state is aware of learner state distribution. Our method, applied to urban traffic simulation, demonstrates significant improvements over existing state-of-the-art baselines in both short-term microscopic and long-term macroscopic realism when evaluated on the real-world dataset pNEUMA. + +# 1. Introduction + +Microscopic traffic simulation is a cornerstone in transportation engineering. It enables engineers to predict and analyze individual vehicle behavior, providing crucial insights into how alterations in road structures or traffic management strategies might influence overall traffic flow; it allows for the testing of diverse scenarios without disrupting real-world traffic; and it enhances safety by pinpointing potential hazards and devising strategies to mitigate risks. By leveraging simulations, engineers can optimize traffic flow, + +reduce congestion, and enhance overall efficiency, proving particularly advantageous in designing or enhancing road networks. Using the simulation data, policymakers and urban planners can make informed decisions aligned with community needs regarding transportation infrastructure. + +However, generating realistic and accurate simulations that can simultaneously replicate the microscopic response of human drivers in various traffic conditions and long-term macroscopic traffic statistics is challenging. In recent years, significant efforts have been invested in developing realistic traffic simulators with the goal to accurately model human driving behaviors. Traditional traffic simulators, such as SUMO [24], AIMSUN [2], and MITSIM [44], typically rely on heuristic car-following models like the Intelligent Driver Model (IDM) [42]. However, despite carefully calibrating parameters, the rule-based models often fail to deliver accurate simulations due to the complexity of real-world traffic environments [12, 30]. Factors such as the road structure, neighboring vehicles, and even driver psychology can influence the decisions of human drivers, making it challenging to achieve accurate simulations [7, 27, 43]. + +In pursuit of realistic traffic simulations, researchers have turned to neural networks to represent the driving model through imitation learning (IL) from human demonstrations. Most traffic simulation approaches [4, 34, 41] leverage behavior cloning (BC) [32] to learn a driving policy by minimizing the disparity between the model output and the human demonstrations in training data. However, BC is hindered by covariate shift [37], where the state induced by the learner's policy progressively diverges from the expert's distribution. Existing BC-based simulators have succeeded in short-term (less than 20 seconds) simulation applications like autonomous driving tests but often fail to generate stable long-term simulations. + +To address covariate shift, existing methods such as DAGger [37], DART [25], and ADAPS [28] incorporate supervisor (humans or principled simulations) corrections + +at the learner's or perturbed expert's states. However, human supervision can be problematic due to intensive labor and judgment errors, and principled simulations may not account for heterogeneous driving behaviors. Recent traffic simulators [40, 46] propose using inverse reinforcement learning (IRL). These methods, such as generative adversarial imitation learning (GAIL) [17] and adversarial inverse reinforcement learning (AIRL) [13], learn a reward function using a discriminator neural network within Generative Adversarial Networks (GANs) [14]. The policy network is trained to maximize the learned reward through online reinforcement learning (RL), enabling agents to handle out-of-distribution states. However, directly applying GAIL to traffic simulation can be problematic [5]. The dynamic nature of the environment in the multi-agent system can lead to noises during policy learning, resulting in highly biased estimated gradients. Furthermore, training the discriminator in GAIL is challenging due to the instability and sensitivity of hyperparameters during min-max optimization. + +To address the issue of covariate shift in multi-agent imitation learning without depending on costly expert supervision or unstable discriminators and multi-agent RL, we propose Learner-Aware Supervised Imitation Learning (LASIL) for tackling the covariate shift problem during policy learning. We mitigate the distribution shift between expert and learner state distribution by augmenting the expert state distribution. However, there is no expert supervision at any augmented state, so we ensure that an augmented state is close to an expert state such that the future trajectory of the expert state can serve as the target trajectory for the augmented state to constrain the learner within the expert state distribution. Hence, our goal is to augment the expert state to cover the learner's state distribution while remaining close to the original expert state distribution. To achieve this, we use a variational autoencoder (VAE) [21] to simultaneously model the distributions of both the expert and learner states. By minimizing the VAE's latent space regularization loss of modeling both distributions, we project the expert and learner states into a unified latent space. And by minimizing the VAE's reconstruction loss, the resulting reconstruction leveraging such learnt latent space will resemble both expert and learner state distribution. As a result, when inputting expert states into the trained VAE, we obtain a learner-aware augmented expert state. + +In practice, we divide an agent's state into two parts: past trajectory and context, i.e., other features in the state including vehicle type, waypoints, and destination. We observe that the distribution of the context is consistent regardless of a learnt policy due to the static features of traffic conditions embedded in the context, leading to less covariate shift. Therefore, we propose a context-conditioned VAE to model the context-conditioned trajectory distribution of both expert and learner states. The decoder of the context- + +conditioned VAE will receive both the latent variable and the context, and yield only the trajectory information. + +In summary, our contributions are as follows: + +- We propose learner-aware supervised imitation learning (LASIL) to achieve stable learning and alleviate covariate shift in multi-agent imitation learning. +- We propose a learner-aware data augmentation method based on a context-conditioned VAE that generates learner-aware augmented expert states. +- Our approach is tailored for urban traffic simulation. To the best of our knowledge, it is the first imitation learning-based traffic simulator that can reproduce long-term (more than 10 minutes) stable microscopic simulation, achieving $40\mathrm{x}$ simulation length improvements over previous state-of-the-art [4, 41, 46]. +- We evaluate our method on the real-world dataset pNEUMA [3] with over half a million trajectories. Our approach outperforms state-of-the-art baselines in both short-term microscopic and long-term macroscopic simulations. The code is available at https://github.com/Kguo-cs/LSAIL. + +# 2. Related Work + +# 2.1. Imitation learning + +Existing imitation learning (IL) methods can be broadly categorized into behavior cloning (BC) and inverse reinforcement learning (IRL) approaches. BC [32] learns a policy in a supervised fashion by minimizing the discrepancy between the learner's actions and those of an expert. However, BC suffers from the issue of covariate shift, where the state distribution induced by the learner's policy gradually deviates from that of the expert. To address this, methods like DAGger [37] and DART [25] request supervisor corrections at the learner's or perturbed expert's states. Our method follows a similar supervised learning approach to DAGger and DART, but does not require access to an expert policy. + +Due to the challenges in obtaining expert supervision, recent IRL-based methods utilize feedback from a neural network-based discriminator to handle covariate shift. Typically, these methods involve an iterative process alternating between reward estimation and reinforcement learning. Earlier IRL-based methods [15, 35, 39, 47] require frequent dynamic programming processes, while recent adversarial IL approaches integrate reward function learning with policy learning using a GAN formulation. However, both GAN and RL training processes are known to be unstable, sensitive to hyperparameters, and have poor sample efficiency [9]. Moreover, the discriminator can easily exploit insignificant differences between expert and policy samples, leading to undesirable performance [45]. In contrast, our method avoids the min-max optimization problem and the sample-inefficient RL process, requiring minimal + +fine-tuning. Instead, we utilize the real future trajectory as the target state to learn corrective behavior. + +# 2.2. Imitation Learning-based Traffic simulation + +Recent traffic simulators focus on enhancing realism by leveraging imitation learning (IL) from human driving demonstrations, which extend traditional BC and IRL methods to tackle the challenging multi-agent IL problem. + +BC-based traffic simulators like TrafficSim [41] and SimNet [4] typically begin by training a prediction model and subsequently adjust the predicted trajectories to prevent collisions and adhere to traffic regulations during simulation. However, BC-based methods face challenges in achieving long-term simulation due to the covariate shift problem. To mitigate this issue, we augment expert state based on the policy distribution, enabling stable long-term simulation. Additionally, we enhance performance by modifying the predicted trajectory during simulation through road projection and ensuring smoothness from the current state. Notably, we skip the computationally intensive collision removal operation, prioritizing the capture of macroscopic long-term influence over micro-level details. + +IRL-based simulators learn the underlying reward function of human driving behavior and derive the driving policy by maximizing the learned reward. While adversarial IRL methods theoretically address the covariate shift of BC in a single-agent context through online interaction, its performance deteriorates in the multi-agent IL domain due to the dynamic environment, complicating the training process. To tackle this challenge, approaches like PS-GAIL [40] and PS-ARIL [46] adopt two-stage learning and gradually introduce vehicles to the environment. Nonetheless, these methods still exhibit significant undesirable traffic phenomena, such as off-road driving, collisions, and abrupt braking. Building upon PS-GAIL, the reward-augmented imitation learning (RAIL) method [5, 22] penalizes undesirable phenomena by introducing a hand-crafted reward. However, maximizing the new reward does not guarantee the recovery of human-like trajectories. Despite numerous enhancements to the original GAIL framework, these IRL-based methods often struggle to produce stable long-term traffic flow, as evidenced in our experiments. In contrast, our method is a supervised learning approach, leading to faster, simpler, and more stable learning of driving policies. + +# 3. Background + +# 3.1. Markov Decision Process + +We model the human driving process using a Markov decision process (MDP) denoted as $\mathcal{M} = \{\mathcal{S},\mathcal{A},\mathcal{T},r\}$ , incorporating a time horizon of $T$ , where $\mathcal{S}$ and $\mathcal{A}$ represent the continuous state and action spaces respectively. A stochastic function $\mathcal{T}:S\times \mathcal{A}\times \mathcal{S}\rightarrow [0,1]$ describes the system + +dynamics, and $r: \mathcal{S} \times \mathcal{A} \to \mathbb{R}$ is a reward function. The policy $\pi(\boldsymbol{a} | s)$ determines the probability of selecting an action $a$ at a state $s$ , and a trajectory $\tau = (s_0, \boldsymbol{a}_0, \dots, s_T, \boldsymbol{a}_T)$ represents a sequence of state-action pairs. The marginal state distribution of a policy $\pi$ is computed as $\rho^{\pi}(s) = \frac{1}{D} \sum_{s \in \mathcal{D}_{\pi}} \delta_s$ , where $\mathcal{D}_{\pi}$ denotes the set of states of size $D$ in trajectories induced by the policy $\pi$ , and $\delta_s$ signifies a Dirac distribution centered on $s$ . Similarly, the marginal state-action distribution $\rho^{\pi}(s, \boldsymbol{a})$ is computed. + +# 3.2. Imitation Learning + +In the application of IL for learning a driving policy, we assume that all agents adopt the same policy. The objective is to learn a policy $\pi$ that minimizes the $f$ -divergence between the marginal state-action distribution of the expert's demonstrations $\rho^{\exp}(s,a)$ and the learner policy's distribution $\rho^{\pi}(s,a)$ . For example, BC optimizes the policy to minimize the Kullback-Leibler (KL) divergence at the expert state distribution $\mathbb{E}_{\rho^{\exp}(s)}[\mathrm{KL}(\pi^{\exp}(a|s)\| \pi (a|s))]$ , while DAgger minimizes $\mathbb{E}_{\rho^{\pi}(s)}[\mathrm{KL}(\pi^{\exp}(a|s)\| \pi (a|s))]$ . On the other hand, GAIL minimizes the Jensen-Shannon divergence $D_{\mathrm{JS}}(\rho^{\exp}(s,a)\| \rho^{\pi}(s,a))$ , while AIRL minimizes the KL divergence $\mathrm{KL}(\rho^{\pi}(s,a)\| \rho^{\exp}(s,a))$ . + +However, GAIL and AIRL often exhibit unsatisfactory practical performance due to their optimization processes' instability and sample inefficiency, which involve GANs and RL. In contrast, BC only requires simple and stable supervised learning but suffers from the covariate shift issue, as it only minimizes the policy difference at the expert state distribution without guaranteeing performance at the learner state distribution. DAGger addresses the covariate shift problem but requires access to the expert policy. To mitigate the covariate shift problem without depending on the expert policy, we propose maximizing the transition probability $\mathbb{E}_{\boldsymbol{s} \sim \rho^{\exp}(\boldsymbol{s}), \boldsymbol{s} + \epsilon \sim \rho^{\pi}(\boldsymbol{s})}[\mathcal{T}(\boldsymbol{s} + \epsilon, \pi(\boldsymbol{s} + \epsilon), \boldsymbol{s}')]$ , where $\epsilon$ is a small augmentation term. Note that our policy learns to predict the distribution of the next state $s'$ instead of the action. Our approach assumes that when the expert's original future trajectory can serve as supervision to guide the agent back towards the expert distribution. + +# 3.3. Variational Autoencoder + +The VAE defines a generative model given by $p_{\theta}(\pmb{x}, \pmb{z}) = p(\pmb{z}) p_{\theta}(\pmb{x} | \pmb{z})$ , where $\pmb{z}$ is the latent variable with prior distribution $p(\pmb{z})$ , and $p_{\theta}(\pmb{x} | \pmb{z})$ represents the conditional distribution modeling the likelihood of data $\pmb{x}$ given $\pmb{z}$ . The learning objective is to maximize the training samples' marginal log-likelihood $\log p_{\theta}(\pmb{x})$ . However, due to the intractability of marginalization, VAE maximizes the variational lower + +bound using $q_{\phi}(\pmb {z}|\pmb {x})$ as the approximate posterior: + +$$ +\begin{array}{l} \log p _ {\theta} (\boldsymbol {x}) \\ \geq \mathbb {E} _ {\boldsymbol {z} \sim q _ {\phi} (\boldsymbol {z} | \boldsymbol {x})} [ \log p _ {\theta} (\boldsymbol {x} | \boldsymbol {z}) ] - \operatorname {K L} \left(q _ {\phi} (\boldsymbol {z} | \boldsymbol {x}) \| p (\boldsymbol {z})\right) \tag {1} \\ := - \mathcal {L} _ {\text {r e c}} (\boldsymbol {x}) - \mathcal {L} _ {\mathrm {K L}} (\boldsymbol {x}), \\ \end{array} +$$ + +where $\mathcal{L}_{\mathrm{rec}}(\boldsymbol{x})$ represents the reconstruction loss, which penalizes the network for creating outputs different from the input. $\mathcal{L}_{\mathrm{KL}}(\boldsymbol{x})$ represents the KL divergence loss to make a continuous and smooth latent space, allowing easy random sampling and interpolation. Intuitively, this KL loss encourages the encoder to distribute all encodings evenly around the center of the latent space. + +# 4. Method + +In Fig. 1, we present an overview of our method. Our model comprises three modules: a VAE-based data augmentation module, a policy network, and a post-processing (LQR and on-road projection) module. During training, we augment each expert data by generating a learner-aware augmented expert state through the VAE's reconstruction of its past trajectory. Using this augmented state along with the original future trajectory, we train the learner's policy network through supervised learning. During simulation, we roll out the policy network with several post-processing steps including sampling, projecting the trajectories onto the road, and smoothing the projected trajectories. + +# 4.1. State Representation + +As human drivers make decisions mainly depending on their surrounding information, we build a graph to model the traffic system with each driving agent as its node. + +Node: The state of each agent can be divided into two components: the past trajectory $s_p$ and the context $s_c$ . In practice, the past trajectory is composed of the agent's positions at several past time steps, including the current time step. The context includes its type, nearby waypoints' positions and corresponding road width, the traffic light status of the road it is traveling on, and destination. The waypoints are composed of the center points of the routing roads with a fixed interval. We transform each agent's state into its individual coordinate system to learn a policy with transformation invariance. To reduce the implicit covariate shift, we set each agent's current position adding a Gaussian perturbation as the origin and the $x$ -axis directions pointing towards the agent's destination like [16]. + +Edge: When each agent's state coordinates are transformed from the global coordinate system to their individual coordinate system, information about the relative positions among agents is lost. However, a traffic model needs the relative information of agents to understand how they interact with each other. To preserve these relationships, we introduce directed edges between neighboring agents. The + +edge feature is the relative position of the destination node's coordinate origin in the source node's coordinate system. + +# 4.2. Learner-aware Data Augmentation + +To address the covariate shift issue between the expert and learner state distributions, we propose to minimize the expert's and learner's embedded state distribution difference. In practice, we propose utilizing the same VAE to model the expert and learner state distributions simultaneously. As both the expert and learner state are projected to the same latent space, the distribution of the state reconstructed from the joint latent space can resemble both distributions. + +In contrast to the past trajectory, the context distribution is more challenging to model but exhibits less covariate shift. Therefore, we propose using a context-conditioned VAE to specifically model the context-conditioned trajectory distribution rather than the state distribution. For each expert or learner state represented as $s = (s_p, s_c)$ , we employ an encoder $q_{\phi}(z | s_p, s_c)$ to obtain the latent variable distribution. We can sample a latent variable $z$ from this distribution by applying the reparameterization trick. Subsequently, a decoder $p_{\theta}(s_p | z, s_c)$ is used to reconstruct the distribution of the past trajectory $s_p$ given $z$ and $s_c$ . The VAE is trained to maximize the variational lower bound, which incorporates the context-conditioned log-likelihood of both the expert and learner past trajectories: + +$$ +\mathcal {L} _ {V A E} = \mathbb {E} _ {\boldsymbol {s} \sim \rho^ {\exp}} \left[ \mathcal {L} \left(\boldsymbol {s} _ {p} \mid \boldsymbol {s} _ {c}\right) \right] + \lambda \mathbb {E} _ {\boldsymbol {s} \sim \rho^ {\pi}} \left[ \mathcal {L} \left(\boldsymbol {s} _ {p} \mid \boldsymbol {s} _ {c}\right) \right]. \tag {2} +$$ + +Here, $\lambda$ is a hyperparameter that controls the degree to which the augmented context-conditioned past trajectory distribution aligns with the learner's context-conditioned past trajectory distribution. The term $\mathcal{L}(s_p|s_c)$ is given by: + +$$ +\begin{array}{l} \mathcal {L} \left(\boldsymbol {s} _ {p} \mid \boldsymbol {s} _ {c}\right) = \mathbb {E} _ {\boldsymbol {z} \sim q _ {\phi} \left(\boldsymbol {z} \mid \boldsymbol {s} _ {p}, \boldsymbol {s} _ {c}\right)} \left[ \log p _ {\theta} \left(\boldsymbol {s} _ {p} \mid \boldsymbol {z}, \boldsymbol {s} _ {c}\right) \right] \\ - \operatorname {K L} \left(q _ {\phi} (\boldsymbol {z} | \boldsymbol {s} _ {p}, \boldsymbol {s} _ {c}) \| p (\boldsymbol {z})\right). \tag {3} \\ \end{array} +$$ + +For simplicity, we assume $p_{\theta}$ and $q_{\phi}$ as multivariate normal distributions with diagonal variance matrices. The prior distribution $p(z)$ is set as an isotropic unit Gaussian $\mathcal{N}(\mathbf{0},\mathbf{I})$ + +# 4.3. Edge-enhanced Graph Attention Network + +We build all models in our approach including the VAE encoder, decoder, and learner's policy network based on an edge-enhanced graph attention (EGAT) network [11, 33], which model interactions by aggregating neighboring node and edge information using an attention mechanism. The node features in the first layer are obtained by embedding the node input with a fully connected layer and the node features in the other layers are calculated by: + +$$ +\boldsymbol {h} _ {i} ^ {l} = \sigma \left(\sum_ {j \in \mathcal {N} _ {i}} \alpha_ {i j} ^ {l} \boldsymbol {W} ^ {l} \left[ \boldsymbol {h} _ {i} ^ {(l - 1)} \| \boldsymbol {e} _ {i j} \| \boldsymbol {h} _ {j} ^ {(l - 1)} \right]\right), \tag {4} +$$ + +![](images/c52e0a4a753e710f5e0703f22977e3dbee94a3e67553d93d5b4a19ad522866d7.jpg) +Figure 1. Overview of our approach. The processes with purple and red arrows handle only expert or learner data, respectively, while the processes represented by black arrows apply to both data. Each state is a multi-agent state represented by a graph. Each model, including the VAE encoder, decoder, and policy network, is implemented using an EGAT network. + +where $h_i^l$ is the feature of node $i$ in the $l$ th layer, $e_{ij}$ is the coordinate origin's position of node $j$ relative to node $i$ , $\| \cdot \|$ means the concatenation, $W^l$ is the learnable weight matrix, $\mathcal{N}_i$ is the set of the first-order neighbors of node $i$ (including the node itself), and $\sigma$ is a non-linear activation function. The attention coefficient $\alpha_{ij}^l$ indicates the importance of node $j$ to node $i$ , considering both node and edge features, and is computed as: + +$$ +\alpha_ {i j} ^ {l} = \operatorname {s o f t m a x} _ {j} \left(\sigma \left(\left(\boldsymbol {w} ^ {l}\right) ^ {T} \left[ \boldsymbol {h} _ {i} ^ {(l - 1)} \| \boldsymbol {e} _ {i j} \| \boldsymbol {h} _ {j} ^ {(l - 1)} \right]\right)\right), \tag {5} +$$ + +where $\boldsymbol{w}^l$ is a learnable weight vector, and normalization is performed on the weights across all neighbors of node $i$ using a softmax function. After passing through multiple EGAT layers, the node features at the last layer is fed into a fully connected layer to obtain the outputs of the network. + +# 4.4. Policy Loss + +To train the learner's policy network, we first sample a learner-aware augmented expert state from the context-conditioned VAE for each expert state. Then, the policy network predicts its future position distribution over $T$ time steps, denoted as $p(\hat{p}_1^m, \hat{p}_2^m, \dots, \hat{p}_T^m)$ , which is assumed to be a product of multi-variable Gaussian distributions: + +$$ +p \left(\hat {\boldsymbol {p}} _ {1} ^ {m}, \hat {\boldsymbol {p}} _ {2} ^ {m}, \dots , \hat {\boldsymbol {p}} _ {T} ^ {m}\right) = \prod_ {t = 1} ^ {T} \mathcal {N} \left(\hat {\boldsymbol {\mu}} _ {t} ^ {m}, \hat {\boldsymbol {\Sigma}} _ {t} ^ {m}\right), \tag {6} +$$ + +where $\hat{\pmb{\mu}}_t^m$ and $\hat{\Sigma}_t^m$ represent the mean and covariance matrix of the predicted position $\hat{\pmb{p}}_t^m$ at future time step $t$ . We assume that there is no correlation between the position distributions at different future time steps. To learn the policy network, we minimize the negative log-likelihood (NLL) + +loss of all agents' ground-truth future trajectories: + +$$ +\mathcal {L} _ {N L L} = - \sum_ {m = 1} ^ {M} \sum_ {t = 1} ^ {T} \log \left(\mathcal {N} \left(\boldsymbol {p} _ {t} ^ {m} - \hat {\boldsymbol {\mu}} _ {t} ^ {m}, \hat {\boldsymbol {\Sigma}} _ {t} ^ {m}\right)\right), \tag {7} +$$ + +where $\pmb{p}_t^m$ denotes the ground truth position of agent $m$ at future time step $t$ , and $M$ is the total number of agents. + +# 4.5. Simulation Process + +During training, we simultaneously simulate (roll out) the learned policy network iteratively to get learner state samples. Instead of directly updating each agent to its predicted position, we apply several post-processing steps to the prediction for better realism. Firstly, we sample from the distribution, and then project each sampled position onto the nearest on-road point. Then, we smooth the projected trajectory with a linear-quadratic regulator (LQR) [1] by minimizing the total commutative quadratic cost of a linear dynamic system described by: + +$$ +\left[ \begin{array}{l} \tilde {\boldsymbol {p}} _ {t + 1} ^ {m} \\ \tilde {\boldsymbol {v}} _ {t + 1} ^ {m} \end{array} \right] = \left[ \begin{array}{c c} \boldsymbol {I} & \boldsymbol {D} \\ 0 & \boldsymbol {I} \end{array} \right] \left[ \begin{array}{l} \tilde {\boldsymbol {p}} _ {t} ^ {m} \\ \tilde {\boldsymbol {v}} _ {t} ^ {m} \end{array} \right] + \left[ \begin{array}{c} \boldsymbol {D} ^ {2} \\ \boldsymbol {D} \end{array} \right] \tilde {\boldsymbol {a}} _ {t} ^ {m}, \tag {8} +$$ + +where $D$ is a diagonal matrix with the interval of each time step as diagonal entries, and $\tilde{\boldsymbol{p}}_t^m$ , $\tilde{\boldsymbol{v}}_t^m$ , $\tilde{\boldsymbol{a}}_t^m$ represent the LQR-planned position, velocity, and acceleration. The system is subject to a quadratic cost function given by: + +$$ +\mathcal {J} = \sum_ {m = 1} ^ {M} \sum_ {t = 1} ^ {T} \| \tilde {\boldsymbol {p}} _ {t} ^ {m} - \bar {\boldsymbol {p}} _ {t} ^ {m} \| ^ {2} + \eta_ {\boldsymbol {a}} \| \tilde {\boldsymbol {a}} _ {t} ^ {m} \| ^ {2}, \tag {9} +$$ + +where the projected predicted position $\bar{\pmb{p}}_m^t$ is considered as the target pose, and the hyper-parameter $\eta_{\mathbf{a}}$ is used to penalize high acceleration. Finally, each agent is updated to the first position of the planned trajectory. + +# 4.6. Training Process + +At each training step, the policy network is trained to maximize the probability of the expert's future trajectory using its context state and history trajectory augmented by the VAE as input. Simultaneously, the VAE is iteratively trained to reconstruct the expert and learner data. The expert data are the same as the policy network's training data before augmentation. The learner data is sampled from a replay buffer. Every $N$ training steps, we empty the replay buffer and roll out the current policy with post-processing for $S$ time steps to generate the learner data and store it in the replay buffer. During the roll-out, we start at a random time step in the training dataset and apply our model in a closed-loop manner. At each time step during the rollout, we sample a future trajectory for each vehicle from the Gaussian distribution predicted by our policy network, project the predicted trajectory onto the road, smooth the on-road trajectory with LQR, and finally update the vehicle to the first position of the smoothed trajectory. + +# 5. Experiment + +# 5.1. Dataset + +We utilize a real-world urban dataset called pNEUMA [3], which was collected by 10 drones in Athens over a span of 4 days. The dataset encompasses over half a million vehicle trajectories within a large area encompassing over $100\mathrm{km}$ of lanes and approximately 100 intersections. The recordings were conducted at 5 intervals each day, with each period lasting about 15 minutes and a data collection time interval of 0.04 seconds. To enhance computational efficiency, we adopt a time step of 0.4 seconds. The dataset is divided into a training set, comprising recordings from the first 3 days, and a validation/test set, consisting of recordings from the final day. Notably, we choose not to utilize other popular traffic datasets like NGSIM [8] or HighD [23] or autonomous driving datasets like Lyft [18] or nuPlan [6], as they only encompass samll-scale scenarios. This limitation makes it inadequate for evaluate macroscopic realism. + +# 5.2. Metrics + +We evaluate the realism of our simulator by measuring the similarity between the simulation result and real data. During evaluation, each vehicle enters the simulator at its first recorded time and position, and is controlled by our simulator to complete its recorded route. When an agent reaches its final recorded position, it is removed from the simulator. + +Short-term microscopic: Following [5, 40], we conduct a short-term microscopic evaluation by simulating for 20 seconds from a random time step in the test dataset. We first measure the similarity between the simulated and real + +data using position and velocity RMSE metrics given by: + +$$ +\mathrm {R M S E} = \frac {1}{T _ {s}} \sum_ {t = 1} ^ {T _ {s}} \sqrt {\frac {1}{M} \sum_ {m = 1} ^ {M} \| \boldsymbol {s} _ {t} ^ {m} - \hat {\boldsymbol {s}} _ {t} ^ {m} \| ^ {2}}, \tag {10} +$$ + +where $s_t^m$ and $\hat{s}_t^m$ were the real and simulated value of the position or velocity of the agent $m$ at time step $t$ , respectively. $T_s$ was the total simulated time steps, and $M$ was the total simulated agent number. Besides, we measure the minimum Average Displace Error (minADE) to not penalize reasonable trajectory unlike the real one: + +$$ +\min A D E _ {N} = \frac {1}{M} \sum_ {m = 1} ^ {M} \min _ {\hat {\boldsymbol {s}} _ {n}} \frac {1}{T _ {s}} \sum_ {t = 1} ^ {T _ {s}} \| \boldsymbol {s} _ {t} ^ {m} - \hat {\boldsymbol {s}} _ {t, n} ^ {m} \| ^ {2}, \tag {11} +$$ + +where $N = 20$ is roll-out times. We also calculate the off-road rate (the avarage proportion of vehicles that deviate more than $1.5\mathrm{m}$ from the road over all time steps). The common collision rate metric is not used because we focus on the long-term impact and the dataset does not provide accurate vehicle size and heading information. + +Long-term macroscopic: we also evaluate our model's long-term macroscopic accuracy on five periods for 800 seconds from its initial recording time. To measure the long-term performance, we use two standard macroscopic metrics for traffic flow data [29, 36, 38], namely road density and speed RMSE, in addition to the off-road rate. The density of a road at a time step is calculated by dividing the number of vehicles on the road by its total lane length, assuming that all lanes have the same width. Meanwhile, the road speed is computed as the mean speed of all vehicles on the road. To quantify the similarity between the simulated and ground truth values, we still use RMSE, where the variable $M$ becomes the total number of roads. + +# 5.3. Performance + +We compare our method against state-of-the-art baselines: + +SUMO [24]: we use the IDM model [42] as the car-following model and mobil [19] as the lane-changing model. We tune the IDM's parameters for 6 types of vehicle by minimizing the MSE between the IDM calculated acceleration and real acceleration using an Adam optimizer [20]. + +BC [32]: we learn our model directly by the BC method. + +MARL [26]: we train our model using IPPO [10] as the Multi-Agent RL algorithm, where the reward function is composed of three parts: a displacement reward (average distance between agent trajectory with GT trajectory), an off-road penalty and a terminal reward. + +MARL+BC [31]: we add a behavior cloning term in the loss function while learning the policy with MARL. + +PS-GAIL [40]: we let all vehicles share the same policy parameter and critic parameter and learn the policy using reward functions computed by GAIL. + +Table 1. Comparison with baselines and ablated models on microscopic metrics for 20 seconds. + +
ModelPosition RMSE(m)Velocity RMSE(m/s)minADE20(m)Off-road(%)
SUMO [24]41.257.0024.200
BC [32]40.08±1.616.74±0.3722.02±3.9833.43±3.10
MARL [26]48.66±1.548.92±0.6740.07±1.263.70±0.40
MARL+BC [31]29.78±0.424.54±0.1422.07±0.2111.14±1.17
PS-GAIL [40]34.78±0.435.33±0.1227.86±0.337.65±0.32
RAIL [5]31.62±0.655.51±0.0924.51±0.412.57±0.47
LASIL (ours)19.21±0.443.02±0.0712.79±0.320.28±0.01
w/o Augmentation21.62±0.543.34±0.1413.81±0.371.31±0.18
w/o Context-conditioned23.03±0.873.73±0.1214.94±0.430.51±0.04
w/o On-road Projection22.53±0.423.26±0.1014.42±0.331.74±0.60
w/o LQR22.62±0.434.05±0.1514.02±0.280.22±0.02
+ +Table 2. Comparison with baselines and ablated models on macroscopic metrics for 800 seconds. + +
ModelRoad Density RMSE(veh/km)Road Speed RMSE(m/s)Off-road(%)
SUMO [24]52.705.520
BC [32]61.51±1.535.38±0.2142.15±5.25
MARL [26]90.59±2.655.27±0.484.63±0.64
MARL+BC [31]43.60±2.804.01±0.1322.73±5.62
PS-GAIL [40]54.06±1.234.03±0.0513.24±3.20
RAIL [5]54.45±1.893.89±0.112.92±0.38
LASIL (ours)45.13±0.253.17±0.140.34±0.03
w/o Augmentation55.65±0.443.73±0.239.62±1.40
w/o Context-conditioned55.98±0.514.48±0.190.64±0.08
w/o On-road Projection52.90±0.543.59±0.2011.33±2.37
w/o LQR61.68±0.963.54±0.160.47±0.07
+ +RAIL [5]: we learn the model as PS-GAIL but with additional displacement, off-road and terminal rewards. + +We train and evaluate each model five times to obtain the mean and standard deviation (std) of various metrics. Note that we do not apply on-road projection and LQR for baselines. We evaluate both short-term and long-term performance, as shown in Tab. 1 and Tab. 2, respectively. Our method achieves better results than all baselines in terms of position and velocity RMSE, road density and speed RMSE, with minor off-road rate. + +# 5.4. Ablation Study + +We conducted a series of ablation experiments to assess the individual contributions of crucial components in our approach, whose results are presented in Tab. 1 and Tab. 2. + +Augmentation: By removing our VAE module and directly learning the original expert state and action, we analyze the importance of our data augmentation technique. The results demonstrate that augmentation plays a vital role in improving both short-term and long-term simulation performance by mitigating covariate shift. + +Context-conditioned VAE: To emphasize the significance of modeling the context-conditioned trajectory distribution rather than the whole state distribution, we replaced our context-conditioned VAE with a naive VAE that directly + +models the expert and learner state distribution. The observed drop in performance demonstrates the challenges in reconstructing the context distribution. + +On-road projection: Our ablation study on the on-road projection module aims to show its impact on reducing the off-road rate. The results show that the on-road projection module leads to a notable decrease in the off-road rate and moderate improvements in other performance metrics. + +LQR: Removing LQR module allows us to evaluate its effectiveness. While LQR led to a higher short-term off-road rate due to more constrained movement (driving off-road due to inertia), its removal deteriorates other short-term and long-term metrics, because LQR can smooth the trajectory, thus leads to more realistic simulations. + +# 5.5. Qualitative Result + +In Fig. 2, we present the mean road density and speed for real-world data, SUMO simulation, and our proposed method over all time steps. The results demonstrate that our proposed method surpasses the capabilities of the SUMO simulator at replicating long-term macroscopic traffic patterns due to our model's enhanced ability to replicate the typical microscopic driving behaviors over a long period. + +![](images/7fe78b129460e56077e2dacba9ddc2df6714e3b7076b387f2f8e7ad5ec3de2b5.jpg) +Figure 2. Mean density and speed on each road over all time steps in the long-term evaluation. Our method's density and speed hot-maps have a more similar color to the ground-truth one compared with the SUMO's. + +# 6. Conclusion + +In conclusion, we have addressed the challenge of creating a realistic traffic simulator that accurately models human driving behaviors in various traffic conditions. Traditional imitation learning-based simulators often fail to deliver accurate long-term simulations due to the covariate shift problem in multi-agent imitation learning. To tackle this issue, we proposed a learner-aware supervised imitation learning method, which leverages a context-conditioned VAE to generate learner-aware augmented expert states. We leverage a context-conditioned VAE to simultaneously reconstruct the expert and learner state. This approach enables the reproduction of long-term stable microscopic traffic simulations, marking a significant advancement in the field of urban + +traffic simulation. Our method has demonstrated superior performance over existing state-of-the-art simulators when evaluated on the real-world dataset pNEUMA, achieving better short-term microscopic and long-term macroscopic similarity to real-world data than state-of-the-art baselines. + +# 7. Acknowledgements + +This research project is partially supported by the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, Hong Kong General Research Fund (11202119 and 11208718), Innovation and Technology Commission (GHP/126/21GD), and Guangdong, Hong Kong and Macao Joint Innovation Project (2023A0505010016). + +# References + +[1] Karl Johan Åström and Richard M Murray. Feedback systems: an introduction for scientists and engineers. Princeton University Press, 2021. 5 +[2] Jaime Barceló and Jordi Casas. Dynamic network simulation with aimsun. Simulation approaches in transportation analysis: Recent advances and challenges, pages 57-98, 2005. 1 +[3] Emmanouil Barmpounakis and Nikolas Geroliminis. On the new era of urban traffic monitoring with massive drone data: The pNEUMA large-scale field experiment. Transportation research part C: emerging technologies, 111:50-71, 2020. 2, 6 +[4] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Blazej Osinski, Hugo Grimmett, and Peter Ondruska. SimNet: Learning reactive self-driving simulations from real-world observations. IEEE International Conference on Robotics and Automation, pages 5119-5125, 2021. 1, 2, 3 +[5] Raunak P. Bhattacharyya, Derek J. Phillips, Changliu Liu, Jayesh K. Gupta, K. Driggs-Campbell, and Mykel J. Kochenderfer. Simulating emergent properties of human driving behavior using multi-agent reward augmented imitation learning. International Conference on Robotics and Automation, pages 789-795, 2019. 2, 3, 6, 7 +[6] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. In IEEE/CVF Computer Vision and Pattern Recognition Conference ADP3 workshop, 2021. 6 +[7] Qianwen Chao, Huikun Bi, Weizi Li, Tianlu Mao, Zhaoqi Wang, Ming C Lin, and Zhigang Deng. A survey on visual traffic simulation: Models, evaluations, and applications in autonomous driving. In Computer Graphics Forum, pages 287-308, 2020. 1 +[8] James Colyar and John Halkias. Us highway 101 dataset. Federal Highway Administration, Tech. Rep., 2007. 6 +[9] Robert Dadashi, L'essonard Hussenot, Matthieu Geist, and Olivier Pietquin. Primal wasserstein imitation learning. In International Conference of Learning Representation, 2021. 2 +[10] Christian Schroeder de Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020. 6 +[11] Frederik Diehl, Thomas Brunner, Michael Truong Le, and Alois Knoll. Graph neural networks for modelling traffic participant interaction. In IEEE Intelligent Vehicles Symposium, pages 695-701, 2019. 4 +[12] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X. Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature Communications, 12, 2021. 1 +[13] Justin Fu, Katie Luo Luo, and Sergey Levine. Learning robust rewards with adversarial inverse reinforcement learn + +ing. In International Conference of Learning Representation, 2018. 2 +[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 2 +[15] Ke Guo, Wenxi Liu, and Jia Pan. End-to-end trajectory distribution prediction based on occupancy grid maps. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2232-2241, 2022. 2 +[16] Ke Guo, Wei Jing, Junbo Chen, and Jia Pan. CCIL: Context-conditioned imitation learning for urban driving. In Robotics: Science and Systems, 2023. 4 +[17] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2 +[18] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference of Robot Learning, 2020. 6 +[19] Arne Kesting, Martin Treiber, and Dirk Helbing. General lane-changing model MOBIL for car-following models. Transportation Research Record, 1999(1):86-94, 2007. 6 +[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6 +[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2 +[22] Yann Koeberle, Stefano Sabatini, Dzmitry V. Tsishkou, and Christophe Sabourin. Exploring the trade off between human driving imitation and safety for traffic simulation. IEEE International Conference on Intelligent Transportation Systems, pages 779-786, 2022. 3 +[23] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In International Conference on Intelligent Transportation Systems, pages 2118-2125, 2018. 6 +[24] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of SUMO-Simulation of Urban MObility. International journal on advances in systems and measurements, 5(3&4), 2012. 1, 6, 7 +[25] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on Robot Learning, pages 143-156, 2017. 1, 2 +[26] Quanyi Li, Zhenghao Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in Neural Information Processing Systems, 2023. 6, 7 +[27] Weizi Li, David Wolinski, and Ming C. Lin. City-scale traffic animation using statistical learning and metamodel-based optimization. ACM Trans. Graph., 36(6):200:1-200:12, 2017. 1 + +[28] Weizi Li, David Wolinski, and Ming C. Lin. ADAPS: Autonomous driving via principled simulations. In IEEE International Conference on Robotics and Automation (ICRA), pages 7625-7631, 2019. 1 +[29] Michael James Lighthill and Gerald Beresford Whitham. On kinematic waves ii. a theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229(1178):317-345, 1955. 6 +[30] Weiwei Liu, Wei Jing, Ke Guo, Gang Xu, Yong Liu, et al. Traco: Learning virtual traffic coordinator for cooperation with multi-agent reinforcement learning. In Conference on Robot Learning, pages 2465-2477, 2023. 1 +[31] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. arXiv preprint arXiv:2212.11419, 2022. 6, 7 +[32] Donald Michie, Michael Bain, and Jean Hayes-Michie. Cognitive models from subcognitive skills. IEEE control engineering series, 44:71-99, 1990. 1, 2, 6, 7 +[33] Xiaoyu Mo, Zhiyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network. IEEE Transactions on Intelligent Transportation Systems, 23:9554-9567, 2022. 4 +[34] Jeremy Morton, Tim A Wheeler, and Mykel J Kochenderfer. Analysis of recurrent neural networks for probabilistic modeling of driver behavior. IEEE Transactions on Intelligent Transportation Systems, 18(5):1289-1298, 2016. 1 +[35] Andrew Y Ng, Stuart Russell, et al. Algorithms for inverse reinforcement learning. In International Conference of Machine Learning, page 2, 2000. 2 +[36] Paul I Richards. Shock waves on the highway. Operations research, 4(1):42-51, 1956. 6 +[37] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics, pages 627-635, 2011. 1, 2 +[38] Jason Sewall, David Wilkie, Paul Merrell, and Ming C Lin. Continuum traffic simulation. In Computer Graphics Forum, number 2, pages 439-448, 2010. 6 +[39] Yu Shen, Weizi Li, and Ming C. Lin. Inverse reinforcement learning with hybrid-weight trust-region optimization and curriculum learning for autonomous maneuvering. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7421-7428, 2022. 2 +[40] Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. Advances in neural information processing systems, 31, 2018. 2, 3, 6, 7 +[41] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multiagent behaviors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400-10409, 2021. 1, 2, 3 + +[42] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1, 6 +[43] David Wilkie, Jason Sewall, Weizi Li, and Ming C. Lin. Virtualized traffic at metropolitan scales. Frontiers in Robotics and AI, 2:11, 2015. 1 +[44] QI Yang and Haris N Koutsopoulos. A microscopic traffic simulator for evaluation of dynamic traffic management systems. Transportation Research Part C: Emerging Technologies, 4(3):113-129, 1996. 1 +[45] Kaifeng Zhang, Rui Zhao, Ziming Zhang, and Yang Gao. Auto-encoding adversarial imitation learning. arXiv preprint arXiv:2206.11004, abs/2206.11004, 2022. 2 +[46] Guanjie Zheng, Hanyang Liu, Kai Xu, and Zhenhui Jessie Li. Objective-aware traffic simulation via inverse reinforcement learning. In International Joint Conference on Artificial Intelligence, 2021. 2, 3 +[47] Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse reinforcement learning. In AAAI Conference on Artificial Intelligence, pages 1433-1438, 2008. 2 \ No newline at end of file diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/images.zip b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6dad87040cef0cad7f422508574715ddb9049951 --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f2b14c5b76b781461d2b09dc9a7aa336599d454c164d89ebed4fb56fa6d428e +size 512923 diff --git a/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/layout.json b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..014553d2026648074475cdcc2735f12a91be78eb --- /dev/null +++ b/2024/LASIL_ Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation/layout.json @@ -0,0 +1,8387 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "spans": [ + { + "bbox": [ + 79, + 102, + 515, + 139 + ], + "type": "text", + "content": "LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": "Ke Guo" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Zhenwei Miao" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Wei Jing" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Weiwei Liu" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Weizi Li" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Dayang Hao" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": ", Jia Pan" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": "Alibaba Group " + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": "The University of Hong Kong \n" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": "Centre for Transformative Garment Production, Hong Kong " + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 160, + 548, + 232 + ], + "type": "text", + "content": "University of Tennessee, Knoxville \nu3006612@connect.hku.hk, zhenwei.mzw@alibaba-inc.com \n21wjing@gmail.com, 11932061@zju.edu.cn, weizili@utk.edu, jpan@cs.hku.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 284, + 289, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 289, + 547 + ], + "type": "text", + "content": "Microscopic traffic simulation plays a crucial role in transportation engineering by providing insights into individual vehicle behavior and overall traffic flow. However, creating a realistic simulator that accurately replicates human driving behaviors in various traffic conditions presents significant challenges. Traditional simulators relying on heuristic models often fail to deliver accurate simulations due to the complexity of real-world traffic environments. Due to the covariate shift issue, existing imitation learning-based simulators often fail to generate stable long-term simulations. In this paper, we propose a novel approach called learner-aware supervised imitation learning to address the covariate shift problem in multi-agent imitation learning. By leveraging a variational autoencoder simultaneously modeling the expert and learner state distribution, our approach augments expert states such that the augmented state is aware of learner state distribution. Our method, applied to urban traffic simulation, demonstrates significant improvements over existing state-of-the-art baselines in both short-term microscopic and long-term macroscopic realism when evaluated on the real-world dataset pNEUMA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 128, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 588, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 588, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 588, + 287, + 696 + ], + "type": "text", + "content": "Microscopic traffic simulation is a cornerstone in transportation engineering. It enables engineers to predict and analyze individual vehicle behavior, providing crucial insights into how alterations in road structures or traffic management strategies might influence overall traffic flow; it allows for the testing of diverse scenarios without disrupting real-world traffic; and it enhances safety by pinpointing potential hazards and devising strategies to mitigate risks. By leveraging simulations, engineers can optimize traffic flow," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 260, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 545, + 320 + ], + "type": "text", + "content": "reduce congestion, and enhance overall efficiency, proving particularly advantageous in designing or enhancing road networks. Using the simulation data, policymakers and urban planners can make informed decisions aligned with community needs regarding transportation infrastructure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 323, + 546, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 546, + 515 + ], + "type": "text", + "content": "However, generating realistic and accurate simulations that can simultaneously replicate the microscopic response of human drivers in various traffic conditions and long-term macroscopic traffic statistics is challenging. In recent years, significant efforts have been invested in developing realistic traffic simulators with the goal to accurately model human driving behaviors. Traditional traffic simulators, such as SUMO [24], AIMSUN [2], and MITSIM [44], typically rely on heuristic car-following models like the Intelligent Driver Model (IDM) [42]. However, despite carefully calibrating parameters, the rule-based models often fail to deliver accurate simulations due to the complexity of real-world traffic environments [12, 30]. Factors such as the road structure, neighboring vehicles, and even driver psychology can influence the decisions of human drivers, making it challenging to achieve accurate simulations [7, 27, 43]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 674 + ], + "type": "text", + "content": "In pursuit of realistic traffic simulations, researchers have turned to neural networks to represent the driving model through imitation learning (IL) from human demonstrations. Most traffic simulation approaches [4, 34, 41] leverage behavior cloning (BC) [32] to learn a driving policy by minimizing the disparity between the model output and the human demonstrations in training data. However, BC is hindered by covariate shift [37], where the state induced by the learner's policy progressively diverges from the expert's distribution. Existing BC-based simulators have succeeded in short-term (less than 20 seconds) simulation applications like autonomous driving tests but often fail to generate stable long-term simulations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "To address covariate shift, existing methods such as DAGger [37], DART [25], and ADAPS [28] incorporate supervisor (humans or principled simulations) corrections" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 203, + 713 + ], + "type": "text", + "content": "*Zhenwei Miao is the corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15386" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 300 + ], + "type": "text", + "content": "at the learner's or perturbed expert's states. However, human supervision can be problematic due to intensive labor and judgment errors, and principled simulations may not account for heterogeneous driving behaviors. Recent traffic simulators [40, 46] propose using inverse reinforcement learning (IRL). These methods, such as generative adversarial imitation learning (GAIL) [17] and adversarial inverse reinforcement learning (AIRL) [13], learn a reward function using a discriminator neural network within Generative Adversarial Networks (GANs) [14]. The policy network is trained to maximize the learned reward through online reinforcement learning (RL), enabling agents to handle out-of-distribution states. However, directly applying GAIL to traffic simulation can be problematic [5]. The dynamic nature of the environment in the multi-agent system can lead to noises during policy learning, resulting in highly biased estimated gradients. Furthermore, training the discriminator in GAIL is challenging due to the instability and sensitivity of hyperparameters during min-max optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 602 + ], + "type": "text", + "content": "To address the issue of covariate shift in multi-agent imitation learning without depending on costly expert supervision or unstable discriminators and multi-agent RL, we propose Learner-Aware Supervised Imitation Learning (LASIL) for tackling the covariate shift problem during policy learning. We mitigate the distribution shift between expert and learner state distribution by augmenting the expert state distribution. However, there is no expert supervision at any augmented state, so we ensure that an augmented state is close to an expert state such that the future trajectory of the expert state can serve as the target trajectory for the augmented state to constrain the learner within the expert state distribution. Hence, our goal is to augment the expert state to cover the learner's state distribution while remaining close to the original expert state distribution. To achieve this, we use a variational autoencoder (VAE) [21] to simultaneously model the distributions of both the expert and learner states. By minimizing the VAE's latent space regularization loss of modeling both distributions, we project the expert and learner states into a unified latent space. And by minimizing the VAE's reconstruction loss, the resulting reconstruction leveraging such learnt latent space will resemble both expert and learner state distribution. As a result, when inputting expert states into the trained VAE, we obtain a learner-aware augmented expert state." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 289, + 715 + ], + "type": "text", + "content": "In practice, we divide an agent's state into two parts: past trajectory and context, i.e., other features in the state including vehicle type, waypoints, and destination. We observe that the distribution of the context is consistent regardless of a learnt policy due to the static features of traffic conditions embedded in the context, leading to less covariate shift. Therefore, we propose a context-conditioned VAE to model the context-conditioned trajectory distribution of both expert and learner states. The decoder of the context-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "conditioned VAE will receive both the latent variable and the context, and yield only the trajectory information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 98, + 502, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 502, + 108 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 502, + 108 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 110, + 545, + 324 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 110, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 110, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 306, + 110, + 545, + 145 + ], + "type": "text", + "content": "- We propose learner-aware supervised imitation learning (LASIL) to achieve stable learning and alleviate covariate shift in multi-agent imitation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 146, + 545, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 146, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 146, + 545, + 181 + ], + "type": "text", + "content": "- We propose a learner-aware data augmentation method based on a context-conditioned VAE that generates learner-aware augmented expert states." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 182, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 182, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 545, + 252 + ], + "type": "text", + "content": "- Our approach is tailored for urban traffic simulation. To the best of our knowledge, it is the first imitation learning-based traffic simulator that can reproduce long-term (more than 10 minutes) stable microscopic simulation, achieving " + }, + { + "bbox": [ + 306, + 182, + 545, + 252 + ], + "type": "inline_equation", + "content": "40\\mathrm{x}" + }, + { + "bbox": [ + 306, + 182, + 545, + 252 + ], + "type": "text", + "content": " simulation length improvements over previous state-of-the-art [4, 41, 46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 253, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 545, + 324 + ], + "type": "text", + "content": "- We evaluate our method on the real-world dataset pNEUMA [3] with over half a million trajectories. Our approach outperforms state-of-the-art baselines in both short-term microscopic and long-term macroscopic simulations. The code is available at https://github.com/Kguo-cs/LSAIL." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 338, + 392, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 392, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 392, + 350 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 358, + 414, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 358, + 414, + 371 + ], + "spans": [ + { + "bbox": [ + 306, + 358, + 414, + 371 + ], + "type": "text", + "content": "2.1. Imitation learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 521 + ], + "type": "text", + "content": "Existing imitation learning (IL) methods can be broadly categorized into behavior cloning (BC) and inverse reinforcement learning (IRL) approaches. BC [32] learns a policy in a supervised fashion by minimizing the discrepancy between the learner's actions and those of an expert. However, BC suffers from the issue of covariate shift, where the state distribution induced by the learner's policy gradually deviates from that of the expert. To address this, methods like DAGger [37] and DART [25] request supervisor corrections at the learner's or perturbed expert's states. Our method follows a similar supervised learning approach to DAGger and DART, but does not require access to an expert policy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Due to the challenges in obtaining expert supervision, recent IRL-based methods utilize feedback from a neural network-based discriminator to handle covariate shift. Typically, these methods involve an iterative process alternating between reward estimation and reinforcement learning. Earlier IRL-based methods [15, 35, 39, 47] require frequent dynamic programming processes, while recent adversarial IL approaches integrate reward function learning with policy learning using a GAN formulation. However, both GAN and RL training processes are known to be unstable, sensitive to hyperparameters, and have poor sample efficiency [9]. Moreover, the discriminator can easily exploit insignificant differences between expert and policy samples, leading to undesirable performance [45]. In contrast, our method avoids the min-max optimization problem and the sample-inefficient RL process, requiring minimal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15387" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "fine-tuning. Instead, we utilize the real future trajectory as the target state to learn corrective behavior." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 105, + 277, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 105, + 277, + 118 + ], + "spans": [ + { + "bbox": [ + 47, + 105, + 277, + 118 + ], + "type": "text", + "content": "2.2. Imitation Learning-based Traffic simulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 123, + 287, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 287, + 171 + ], + "type": "text", + "content": "Recent traffic simulators focus on enhancing realism by leveraging imitation learning (IL) from human driving demonstrations, which extend traditional BC and IRL methods to tackle the challenging multi-agent IL problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 171, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 338 + ], + "type": "text", + "content": "BC-based traffic simulators like TrafficSim [41] and SimNet [4] typically begin by training a prediction model and subsequently adjust the predicted trajectories to prevent collisions and adhere to traffic regulations during simulation. However, BC-based methods face challenges in achieving long-term simulation due to the covariate shift problem. To mitigate this issue, we augment expert state based on the policy distribution, enabling stable long-term simulation. Additionally, we enhance performance by modifying the predicted trajectory during simulation through road projection and ensuring smoothness from the current state. Notably, we skip the computationally intensive collision removal operation, prioritizing the capture of macroscopic long-term influence over micro-level details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 602 + ], + "type": "text", + "content": "IRL-based simulators learn the underlying reward function of human driving behavior and derive the driving policy by maximizing the learned reward. While adversarial IRL methods theoretically address the covariate shift of BC in a single-agent context through online interaction, its performance deteriorates in the multi-agent IL domain due to the dynamic environment, complicating the training process. To tackle this challenge, approaches like PS-GAIL [40] and PS-ARIL [46] adopt two-stage learning and gradually introduce vehicles to the environment. Nonetheless, these methods still exhibit significant undesirable traffic phenomena, such as off-road driving, collisions, and abrupt braking. Building upon PS-GAIL, the reward-augmented imitation learning (RAIL) method [5, 22] penalizes undesirable phenomena by introducing a hand-crafted reward. However, maximizing the new reward does not guarantee the recovery of human-like trajectories. Despite numerous enhancements to the original GAIL framework, these IRL-based methods often struggle to produce stable long-term traffic flow, as evidenced in our experiments. In contrast, our method is a supervised learning approach, leading to faster, simpler, and more stable learning of driving policies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 615, + 126, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 126, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 126, + 628 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 187, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 187, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 187, + 647 + ], + "type": "text", + "content": "3.1. Markov Decision Process" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": "We model the human driving process using a Markov decision process (MDP) denoted as " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{\\mathcal{S},\\mathcal{A},\\mathcal{T},r\\}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", incorporating a time horizon of " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " represent the continuous state and action spaces respectively. A stochastic function " + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}:S\\times \\mathcal{A}\\times \\mathcal{S}\\rightarrow [0,1]" + }, + { + "bbox": [ + 47, + 653, + 287, + 713 + ], + "type": "text", + "content": " describes the system" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": "dynamics, and " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "r: \\mathcal{S} \\times \\mathcal{A} \\to \\mathbb{R}" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " is a reward function. The policy " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\pi(\\boldsymbol{a} | s)" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " determines the probability of selecting an action " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " at a state " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": ", and a trajectory " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\tau = (s_0, \\boldsymbol{a}_0, \\dots, s_T, \\boldsymbol{a}_T)" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " represents a sequence of state-action pairs. The marginal state distribution of a policy " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " is computed as " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\rho^{\\pi}(s) = \\frac{1}{D} \\sum_{s \\in \\mathcal{D}_{\\pi}} \\delta_s" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\pi}" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " denotes the set of states of size " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " in trajectories induced by the policy " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\delta_s" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " signifies a Dirac distribution centered on " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": ". Similarly, the marginal state-action distribution " + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\rho^{\\pi}(s, \\boldsymbol{a})" + }, + { + "bbox": [ + 306, + 72, + 545, + 180 + ], + "type": "text", + "content": " is computed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 204, + 418, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 418, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 418, + 217 + ], + "type": "text", + "content": "3.2. Imitation Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": "In the application of IL for learning a driving policy, we assume that all agents adopt the same policy. The objective is to learn a policy " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": " that minimizes the " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": "-divergence between the marginal state-action distribution of the expert's demonstrations " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\rho^{\\exp}(s,a)" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": " and the learner policy's distribution " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\rho^{\\pi}(s,a)" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": ". For example, BC optimizes the policy to minimize the Kullback-Leibler (KL) divergence at the expert state distribution " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\rho^{\\exp}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": ", while DAgger minimizes " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\rho^{\\pi}(s)}[\\mathrm{KL}(\\pi^{\\exp}(a|s)\\| \\pi (a|s))]" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": ". On the other hand, GAIL minimizes the Jensen-Shannon divergence " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{JS}}(\\rho^{\\exp}(s,a)\\| \\rho^{\\pi}(s,a))" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": ", while AIRL minimizes the KL divergence " + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}(\\rho^{\\pi}(s,a)\\| \\rho^{\\exp}(s,a))" + }, + { + "bbox": [ + 305, + 228, + 545, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "text", + "content": "However, GAIL and AIRL often exhibit unsatisfactory practical performance due to their optimization processes' instability and sample inefficiency, which involve GANs and RL. In contrast, BC only requires simple and stable supervised learning but suffers from the covariate shift issue, as it only minimizes the policy difference at the expert state distribution without guaranteeing performance at the learner state distribution. DAGger addresses the covariate shift problem but requires access to the expert policy. To mitigate the covariate shift problem without depending on the expert policy, we propose maximizing the transition probability " + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{s} \\sim \\rho^{\\exp}(\\boldsymbol{s}), \\boldsymbol{s} + \\epsilon \\sim \\rho^{\\pi}(\\boldsymbol{s})}[\\mathcal{T}(\\boldsymbol{s} + \\epsilon, \\pi(\\boldsymbol{s} + \\epsilon), \\boldsymbol{s}')]" + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "text", + "content": " is a small augmentation term. Note that our policy learns to predict the distribution of the next state " + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 305, + 378, + 545, + 581 + ], + "type": "text", + "content": " instead of the action. Our approach assumes that when the expert's original future trajectory can serve as supervision to guide the agent back towards the expert distribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 605, + 443, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 443, + 617 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 443, + 617 + ], + "type": "text", + "content": "3.3. Variational Autoencoder" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "The VAE defines a generative model given by " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}, \\pmb{z}) = p(\\pmb{z}) p_{\\theta}(\\pmb{x} | \\pmb{z})" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " is the latent variable with prior distribution " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "p(\\pmb{z})" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x} | \\pmb{z})" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " represents the conditional distribution modeling the likelihood of data " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ". The learning objective is to maximize the training samples' marginal log-likelihood " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\pmb{x})" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ". However, due to the intractability of marginalization, VAE maximizes the variational lower" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15388" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 253, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 253, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 253, + 85 + ], + "type": "text", + "content": "bound using " + }, + { + "bbox": [ + 47, + 72, + 253, + 85 + ], + "type": "inline_equation", + "content": "q_{\\phi}(\\pmb {z}|\\pmb {x})" + }, + { + "bbox": [ + 47, + 72, + 253, + 85 + ], + "type": "text", + "content": " as the approximate posterior:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 92, + 287, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 92, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 58, + 92, + 287, + 137 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {x}) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x})} [ \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) ] - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {x}) \\| p (\\boldsymbol {z})\\right) \\tag {1} \\\\ := - \\mathcal {L} _ {\\text {r e c}} (\\boldsymbol {x}) - \\mathcal {L} _ {\\mathrm {K L}} (\\boldsymbol {x}), \\\\ \\end{array}", + "image_path": "f266f08fd9b386e4dca90c6dc4da14514e5f6a797f581ece4ad685d2679f5e0d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}(\\boldsymbol{x})" + }, + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "text", + "content": " represents the reconstruction loss, which penalizes the network for creating outputs different from the input. " + }, + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{KL}}(\\boldsymbol{x})" + }, + { + "bbox": [ + 46, + 144, + 287, + 228 + ], + "type": "text", + "content": " represents the KL divergence loss to make a continuous and smooth latent space, allowing easy random sampling and interpolation. Intuitively, this KL loss encourages the encoder to distribute all encodings evenly around the center of the latent space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 237, + 103, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 103, + 249 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 103, + 249 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 258, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 402 + ], + "type": "text", + "content": "In Fig. 1, we present an overview of our method. Our model comprises three modules: a VAE-based data augmentation module, a policy network, and a post-processing (LQR and on-road projection) module. During training, we augment each expert data by generating a learner-aware augmented expert state through the VAE's reconstruction of its past trajectory. Using this augmented state along with the original future trajectory, we train the learner's policy network through supervised learning. During simulation, we roll out the policy network with several post-processing steps including sampling, projecting the trajectories onto the road, and smoothing the projected trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 407, + 167, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 167, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 167, + 420 + ], + "type": "text", + "content": "4.1. State Representation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 462 + ], + "type": "text", + "content": "As human drivers make decisions mainly depending on their surrounding information, we build a graph to model the traffic system with each driving agent as its node." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "text", + "content": "Node: The state of each agent can be divided into two components: the past trajectory " + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "text", + "content": " and the context " + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "text", + "content": ". In practice, the past trajectory is composed of the agent's positions at several past time steps, including the current time step. The context includes its type, nearby waypoints' positions and corresponding road width, the traffic light status of the road it is traveling on, and destination. The waypoints are composed of the center points of the routing roads with a fixed interval. We transform each agent's state into its individual coordinate system to learn a policy with transformation invariance. To reduce the implicit covariate shift, we set each agent's current position adding a Gaussian perturbation as the origin and the " + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 463, + 287, + 629 + ], + "type": "text", + "content": "-axis directions pointing towards the agent's destination like [16]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Edge: When each agent's state coordinates are transformed from the global coordinate system to their individual coordinate system, information about the relative positions among agents is lost. However, a traffic model needs the relative information of agents to understand how they interact with each other. To preserve these relationships, we introduce directed edges between neighboring agents. The" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "edge feature is the relative position of the destination node's coordinate origin in the source node's coordinate system." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 103, + 493, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 103, + 493, + 116 + ], + "spans": [ + { + "bbox": [ + 305, + 103, + 493, + 116 + ], + "type": "text", + "content": "4.2. Learner-aware Data Augmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 121, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 545, + 217 + ], + "type": "text", + "content": "To address the covariate shift issue between the expert and learner state distributions, we propose to minimize the expert's and learner's embedded state distribution difference. In practice, we propose utilizing the same VAE to model the expert and learner state distributions simultaneously. As both the expert and learner state are projected to the same latent space, the distribution of the state reconstructed from the joint latent space can resemble both distributions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": "In contrast to the past trajectory, the context distribution is more challenging to model but exhibits less covariate shift. Therefore, we propose using a context-conditioned VAE to specifically model the context-conditioned trajectory distribution rather than the state distribution. For each expert or learner state represented as " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "s = (s_p, s_c)" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": ", we employ an encoder " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z | s_p, s_c)" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": " to obtain the latent variable distribution. We can sample a latent variable " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": " from this distribution by applying the reparameterization trick. Subsequently, a decoder " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "p_{\\theta}(s_p | z, s_c)" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": " is used to reconstruct the distribution of the past trajectory " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "s_p" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 304, + 218, + 545, + 385 + ], + "type": "text", + "content": ". The VAE is trained to maximize the variational lower bound, which incorporates the context-conditioned log-likelihood of both the expert and learner past trajectories:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 394, + 545, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 545, + 408 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V A E} = \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\exp}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right] + \\lambda \\mathbb {E} _ {\\boldsymbol {s} \\sim \\rho^ {\\pi}} \\left[ \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) \\right]. \\tag {2}", + "image_path": "00098dc5aaa0a7db2f611f45790478af6890b5f9acb9ff8b66cfb895ef0a4607.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "text", + "content": " is a hyperparameter that controls the degree to which the augmented context-conditioned past trajectory distribution aligns with the learner's context-conditioned past trajectory distribution. The term " + }, + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(s_p|s_c)" + }, + { + "bbox": [ + 304, + 416, + 545, + 466 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 474, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 474, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 335, + 474, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {s} _ {c}\\right) = \\mathbb {E} _ {\\boldsymbol {z} \\sim q _ {\\phi} \\left(\\boldsymbol {z} \\mid \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}\\right)} \\left[ \\log p _ {\\theta} \\left(\\boldsymbol {s} _ {p} \\mid \\boldsymbol {z}, \\boldsymbol {s} _ {c}\\right) \\right] \\\\ - \\operatorname {K L} \\left(q _ {\\phi} (\\boldsymbol {z} | \\boldsymbol {s} _ {p}, \\boldsymbol {s} _ {c}) \\| p (\\boldsymbol {z})\\right). \\tag {3} \\\\ \\end{array}", + "image_path": "dbd5d3864e66f2c5245afc047bc4b429d62e5f9baf548c70664f141d3d6b6ebd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "text", + "content": "For simplicity, we assume " + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "inline_equation", + "content": "q_{\\phi}" + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "text", + "content": " as multivariate normal distributions with diagonal variance matrices. The prior distribution " + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "text", + "content": " is set as an isotropic unit Gaussian " + }, + { + "bbox": [ + 304, + 512, + 545, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 555, + 525, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 525, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 525, + 568 + ], + "type": "text", + "content": "4.3. Edge-enhanced Graph Attention Network" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 574, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 670 + ], + "type": "text", + "content": "We build all models in our approach including the VAE encoder, decoder, and learner's policy network based on an edge-enhanced graph attention (EGAT) network [11, 33], which model interactions by aggregating neighboring node and edge information using an attention mechanism. The node features in the first layer are obtained by embedding the node input with a fully connected layer and the node features in the other layers are calculated by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 677, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 677, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 329, + 677, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {i} ^ {l} = \\sigma \\left(\\sum_ {j \\in \\mathcal {N} _ {i}} \\alpha_ {i j} ^ {l} \\boldsymbol {W} ^ {l} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right), \\tag {4}", + "image_path": "a5717014ab7c4da88b4b7c289df6083fa6d7928dfd344a9b3d4ee89c874dfa99.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15389" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 266 + ], + "type": "image", + "image_path": "c52e0a4a753e710f5e0703f22977e3dbee94a3e67553d93d5b4a19ad522866d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 301 + ], + "type": "text", + "content": "Figure 1. Overview of our approach. The processes with purple and red arrows handle only expert or learner data, respectively, while the processes represented by black arrows apply to both data. Each state is a multi-agent state represented by a graph. Each model, including the VAE encoder, decoder, and policy network, is implemented using an EGAT network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "h_i^l" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " is the feature of node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": "th layer, " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "e_{ij}" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " is the coordinate origin's position of node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " relative to node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " means the concatenation, " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "W^l" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " is the learnable weight matrix, " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " is the set of the first-order neighbors of node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " (including the node itself), and " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " is a non-linear activation function. The attention coefficient " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "\\alpha_{ij}^l" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " indicates the importance of node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": " to node " + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 304, + 290, + 402 + ], + "type": "text", + "content": ", considering both node and edge features, and is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 409, + 287, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 409, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 52, + 409, + 287, + 431 + ], + "type": "interline_equation", + "content": "\\alpha_ {i j} ^ {l} = \\operatorname {s o f t m a x} _ {j} \\left(\\sigma \\left(\\left(\\boldsymbol {w} ^ {l}\\right) ^ {T} \\left[ \\boldsymbol {h} _ {i} ^ {(l - 1)} \\| \\boldsymbol {e} _ {i j} \\| \\boldsymbol {h} _ {j} ^ {(l - 1)} \\right]\\right)\\right), \\tag {5}", + "image_path": "c83230c3fd895a2f582100a46869a85bacbae64d359c0801112893551b76b0ca.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}^l" + }, + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "text", + "content": " is a learnable weight vector, and normalization is performed on the weights across all neighbors of node " + }, + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 440, + 287, + 501 + ], + "type": "text", + "content": " using a softmax function. After passing through multiple EGAT layers, the node features at the last layer is fed into a fully connected layer to obtain the outputs of the network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 509, + 123, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 123, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 123, + 521 + ], + "type": "text", + "content": "4.4. Policy Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": "To train the learner's policy network, we first sample a learner-aware augmented expert state from the context-conditioned VAE for each expert state. Then, the policy network predicts its future position distribution over " + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": " time steps, denoted as " + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "inline_equation", + "content": "p(\\hat{p}_1^m, \\hat{p}_2^m, \\dots, \\hat{p}_T^m)" + }, + { + "bbox": [ + 46, + 528, + 287, + 599 + ], + "type": "text", + "content": ", which is assumed to be a product of multi-variable Gaussian distributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 609, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 609, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 85, + 609, + 287, + 643 + ], + "type": "interline_equation", + "content": "p \\left(\\hat {\\boldsymbol {p}} _ {1} ^ {m}, \\hat {\\boldsymbol {p}} _ {2} ^ {m}, \\dots , \\hat {\\boldsymbol {p}} _ {T} ^ {m}\\right) = \\prod_ {t = 1} ^ {T} \\mathcal {N} \\left(\\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right), \\tag {6}", + "image_path": "c76508a446f5beba149b4f1d56b8b071ce8a0cf0f68604ea3a911f2346d43bba.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{\\mu}}_t^m" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}_t^m" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " represent the mean and covariance matrix of the predicted position " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}_t^m" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " at future time step " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": ". We assume that there is no correlation between the position distributions at different future time steps. To learn the policy network, we minimize the negative log-likelihood (NLL)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 304, + 508, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 304, + 508, + 318 + ], + "spans": [ + { + "bbox": [ + 305, + 304, + 508, + 318 + ], + "type": "text", + "content": "loss of all agents' ground-truth future trajectories:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 331, + 324, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 324, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 331, + 324, + 545, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {N L L} = - \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\log \\left(\\mathcal {N} \\left(\\boldsymbol {p} _ {t} ^ {m} - \\hat {\\boldsymbol {\\mu}} _ {t} ^ {m}, \\hat {\\boldsymbol {\\Sigma}} _ {t} ^ {m}\\right)\\right), \\tag {7}", + "image_path": "8c934af1021587ac78f530212a77fb42f90950ce3fee0beb30ce72c7c4bce427.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^m" + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "content": " denotes the ground truth position of agent " + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "content": " at future time step " + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 363, + 547, + 388 + ], + "type": "text", + "content": " is the total number of agents." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 393, + 418, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 418, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 418, + 405 + ], + "type": "text", + "content": "4.5. Simulation Process" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 532 + ], + "type": "text", + "content": "During training, we simultaneously simulate (roll out) the learned policy network iteratively to get learner state samples. Instead of directly updating each agent to its predicted position, we apply several post-processing steps to the prediction for better realism. Firstly, we sample from the distribution, and then project each sampled position onto the nearest on-road point. Then, we smooth the projected trajectory with a linear-quadratic regulator (LQR) [1] by minimizing the total commutative quadratic cost of a linear dynamic system described by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 537, + 545, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 537, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 318, + 537, + 545, + 566 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t + 1} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t + 1} ^ {m} \\end{array} \\right] = \\left[ \\begin{array}{c c} \\boldsymbol {I} & \\boldsymbol {D} \\\\ 0 & \\boldsymbol {I} \\end{array} \\right] \\left[ \\begin{array}{l} \\tilde {\\boldsymbol {p}} _ {t} ^ {m} \\\\ \\tilde {\\boldsymbol {v}} _ {t} ^ {m} \\end{array} \\right] + \\left[ \\begin{array}{c} \\boldsymbol {D} ^ {2} \\\\ \\boldsymbol {D} \\end{array} \\right] \\tilde {\\boldsymbol {a}} _ {t} ^ {m}, \\tag {8}", + "image_path": "62f48b83e99aef0e0f7835e9fb89c95c5f7acb9e1ccf1db435940e7a60085565.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "content": " is a diagonal matrix with the interval of each time step as diagonal entries, and " + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{p}}_t^m" + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{v}}_t^m" + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{a}}_t^m" + }, + { + "bbox": [ + 304, + 571, + 547, + 619 + ], + "type": "text", + "content": " represent the LQR-planned position, velocity, and acceleration. The system is subject to a quadratic cost function given by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 625, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 625, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 341, + 625, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {J} = \\sum_ {m = 1} ^ {M} \\sum_ {t = 1} ^ {T} \\| \\tilde {\\boldsymbol {p}} _ {t} ^ {m} - \\bar {\\boldsymbol {p}} _ {t} ^ {m} \\| ^ {2} + \\eta_ {\\boldsymbol {a}} \\| \\tilde {\\boldsymbol {a}} _ {t} ^ {m} \\| ^ {2}, \\tag {9}", + "image_path": "0ee86cb59119919cda96729acd3104c83fc6bed774dbc57e4d700ec62c0deced.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "where the projected predicted position " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}_m^t" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " is considered as the target pose, and the hyper-parameter " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathbf{a}}" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " is used to penalize high acceleration. Finally, each agent is updated to the first position of the planned trajectory." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15390" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 149, + 85 + ], + "type": "text", + "content": "4.6. Training Process" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "text", + "content": "At each training step, the policy network is trained to maximize the probability of the expert's future trajectory using its context state and history trajectory augmented by the VAE as input. Simultaneously, the VAE is iteratively trained to reconstruct the expert and learner data. The expert data are the same as the policy network's training data before augmentation. The learner data is sampled from a replay buffer. Every " + }, + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "text", + "content": " training steps, we empty the replay buffer and roll out the current policy with post-processing for " + }, + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 91, + 289, + 309 + ], + "type": "text", + "content": " time steps to generate the learner data and store it in the replay buffer. During the roll-out, we start at a random time step in the training dataset and apply our model in a closed-loop manner. At each time step during the rollout, we sample a future trajectory for each vehicle from the Gaussian distribution predicted by our policy network, project the predicted trajectory onto the road, smooth the on-road trajectory with LQR, and finally update the vehicle to the first position of the smoothed trajectory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 324, + 124, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 324, + 124, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 124, + 338 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 105, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 105, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 105, + 357 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 366, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 289, + 559 + ], + "type": "text", + "content": "We utilize a real-world urban dataset called pNEUMA [3], which was collected by 10 drones in Athens over a span of 4 days. The dataset encompasses over half a million vehicle trajectories within a large area encompassing over " + }, + { + "bbox": [ + 46, + 366, + 289, + 559 + ], + "type": "inline_equation", + "content": "100\\mathrm{km}" + }, + { + "bbox": [ + 46, + 366, + 289, + 559 + ], + "type": "text", + "content": " of lanes and approximately 100 intersections. The recordings were conducted at 5 intervals each day, with each period lasting about 15 minutes and a data collection time interval of 0.04 seconds. To enhance computational efficiency, we adopt a time step of 0.4 seconds. The dataset is divided into a training set, comprising recordings from the first 3 days, and a validation/test set, consisting of recordings from the final day. Notably, we choose not to utilize other popular traffic datasets like NGSIM [8] or HighD [23] or autonomous driving datasets like Lyft [18] or nuPlan [6], as they only encompass samll-scale scenarios. This limitation makes it inadequate for evaluate macroscopic realism." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 571, + 106, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 571, + 106, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 571, + 106, + 582 + ], + "type": "text", + "content": "5.2. Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 591, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 664 + ], + "type": "text", + "content": "We evaluate the realism of our simulator by measuring the similarity between the simulation result and real data. During evaluation, each vehicle enters the simulator at its first recorded time and position, and is controlled by our simulator to complete its recorded route. When an agent reaches its final recorded position, it is removed from the simulator." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Short-term microscopic: Following [5, 40], we conduct a short-term microscopic evaluation by simulating for 20 seconds from a random time step in the test dataset. We first measure the similarity between the simulated and real" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 542, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 542, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 542, + 85 + ], + "type": "text", + "content": "data using position and velocity RMSE metrics given by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 92, + 545, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 92, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 329, + 92, + 545, + 130 + ], + "type": "interline_equation", + "content": "\\mathrm {R M S E} = \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\sqrt {\\frac {1}{M} \\sum_ {m = 1} ^ {M} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t} ^ {m} \\| ^ {2}}, \\tag {10}", + "image_path": "3e83db78545cc7acfc671faec0b0bddd27827dd2cc340c7e89fc33cdf379c7e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "s_t^m" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "\\hat{s}_t^m" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": " were the real and simulated value of the position or velocity of the agent " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "T_s" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": " was the total simulated time steps, and " + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 137, + 547, + 210 + ], + "type": "text", + "content": " was the total simulated agent number. Besides, we measure the minimum Average Displace Error (minADE) to not penalize reasonable trajectory unlike the real one:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 217, + 545, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 545, + 251 + ], + "type": "interline_equation", + "content": "\\min A D E _ {N} = \\frac {1}{M} \\sum_ {m = 1} ^ {M} \\min _ {\\hat {\\boldsymbol {s}} _ {n}} \\frac {1}{T _ {s}} \\sum_ {t = 1} ^ {T _ {s}} \\| \\boldsymbol {s} _ {t} ^ {m} - \\hat {\\boldsymbol {s}} _ {t, n} ^ {m} \\| ^ {2}, \\tag {11}", + "image_path": "33a0899df392b7baa7514552eefd6dc8a9dccaf95d70c9c1f4c9d4aeef3c3f6f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "content": " is roll-out times. We also calculate the off-road rate (the avarage proportion of vehicles that deviate more than " + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 304, + 257, + 545, + 329 + ], + "type": "text", + "content": " from the road over all time steps). The common collision rate metric is not used because we focus on the long-term impact and the dataset does not provide accurate vehicle size and heading information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 329, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 484 + ], + "type": "text", + "content": "Long-term macroscopic: we also evaluate our model's long-term macroscopic accuracy on five periods for 800 seconds from its initial recording time. To measure the long-term performance, we use two standard macroscopic metrics for traffic flow data [29, 36, 38], namely road density and speed RMSE, in addition to the off-road rate. The density of a road at a time step is calculated by dividing the number of vehicles on the road by its total lane length, assuming that all lanes have the same width. Meanwhile, the road speed is computed as the mean speed of all vehicles on the road. To quantify the similarity between the simulated and ground truth values, we still use RMSE, where the variable " + }, + { + "bbox": [ + 304, + 329, + 546, + 484 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 329, + 546, + 484 + ], + "type": "text", + "content": " becomes the total number of roads." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 491, + 389, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 389, + 503 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 389, + 503 + ], + "type": "text", + "content": "5.3. Performance" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 510, + 541, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 541, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 541, + 521 + ], + "type": "text", + "content": "We compare our method against state-of-the-art baselines:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "type": "text", + "content": "SUMO [24]: we use the IDM model [42] as the car-following model and mobil [19] as the lane-changing model. We tune the IDM's parameters for 6 types of vehicle by minimizing the MSE between the IDM calculated acceleration and real acceleration using an Adam optimizer [20]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 582, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 545, + 594 + ], + "type": "text", + "content": "BC [32]: we learn our model directly by the BC method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "content": "MARL [26]: we train our model using IPPO [10] as the Multi-Agent RL algorithm, where the reward function is composed of three parts: a displacement reward (average distance between agent trajectory with GT trajectory), an off-road penalty and a terminal reward." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 677 + ], + "type": "text", + "content": "MARL+BC [31]: we add a behavior cloning term in the loss function while learning the policy with MARL." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "PS-GAIL [40]: we let all vehicles share the same policy parameter and critic parameter and learn the policy using reward functions computed by GAIL." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15391" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 82, + 519, + 237 + ], + "blocks": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "lines": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "spans": [ + { + "bbox": [ + 123, + 71, + 470, + 81 + ], + "type": "text", + "content": "Table 1. Comparison with baselines and ablated models on microscopic metrics for 20 seconds." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 82, + 519, + 237 + ], + "lines": [ + { + "bbox": [ + 75, + 82, + 519, + 237 + ], + "spans": [ + { + "bbox": [ + 75, + 82, + 519, + 237 + ], + "type": "table", + "html": "
ModelPosition RMSE(m)Velocity RMSE(m/s)minADE20(m)Off-road(%)
SUMO [24]41.257.0024.200
BC [32]40.08±1.616.74±0.3722.02±3.9833.43±3.10
MARL [26]48.66±1.548.92±0.6740.07±1.263.70±0.40
MARL+BC [31]29.78±0.424.54±0.1422.07±0.2111.14±1.17
PS-GAIL [40]34.78±0.435.33±0.1227.86±0.337.65±0.32
RAIL [5]31.62±0.655.51±0.0924.51±0.412.57±0.47
LASIL (ours)19.21±0.443.02±0.0712.79±0.320.28±0.01
w/o Augmentation21.62±0.543.34±0.1413.81±0.371.31±0.18
w/o Context-conditioned23.03±0.873.73±0.1214.94±0.430.51±0.04
w/o On-road Projection22.53±0.423.26±0.1014.42±0.331.74±0.60
w/o LQR22.62±0.434.05±0.1514.02±0.280.22±0.02
", + "image_path": "0268c0a1fa53ed225ecb6bfcfbb35d7df3d3f48b8af73efc045b67907470045d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 79, + 253, + 515, + 408 + ], + "blocks": [ + { + "bbox": [ + 120, + 241, + 473, + 251 + ], + "lines": [ + { + "bbox": [ + 120, + 241, + 473, + 251 + ], + "spans": [ + { + "bbox": [ + 120, + 241, + 473, + 251 + ], + "type": "text", + "content": "Table 2. Comparison with baselines and ablated models on macroscopic metrics for 800 seconds." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 79, + 253, + 515, + 408 + ], + "lines": [ + { + "bbox": [ + 79, + 253, + 515, + 408 + ], + "spans": [ + { + "bbox": [ + 79, + 253, + 515, + 408 + ], + "type": "table", + "html": "
ModelRoad Density RMSE(veh/km)Road Speed RMSE(m/s)Off-road(%)
SUMO [24]52.705.520
BC [32]61.51±1.535.38±0.2142.15±5.25
MARL [26]90.59±2.655.27±0.484.63±0.64
MARL+BC [31]43.60±2.804.01±0.1322.73±5.62
PS-GAIL [40]54.06±1.234.03±0.0513.24±3.20
RAIL [5]54.45±1.893.89±0.112.92±0.38
LASIL (ours)45.13±0.253.17±0.140.34±0.03
w/o Augmentation55.65±0.443.73±0.239.62±1.40
w/o Context-conditioned55.98±0.514.48±0.190.64±0.08
w/o On-road Projection52.90±0.543.59±0.2011.33±2.37
w/o LQR61.68±0.963.54±0.160.47±0.07
", + "image_path": "f9b8def84aa564045cc8cd26e7540397c1fc332b7a0e7a8cddf50f4bae582c4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 412, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 287, + 436 + ], + "type": "text", + "content": "RAIL [5]: we learn the model as PS-GAIL but with additional displacement, off-road and terminal rewards." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 533 + ], + "type": "text", + "content": "We train and evaluate each model five times to obtain the mean and standard deviation (std) of various metrics. Note that we do not apply on-road projection and LQR for baselines. We evaluate both short-term and long-term performance, as shown in Tab. 1 and Tab. 2, respectively. Our method achieves better results than all baselines in terms of position and velocity RMSE, road density and speed RMSE, with minor off-road rate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 540, + 141, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 141, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 141, + 553 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 594 + ], + "type": "text", + "content": "We conducted a series of ablation experiments to assess the individual contributions of crucial components in our approach, whose results are presented in Tab. 1 and Tab. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 665 + ], + "type": "text", + "content": "Augmentation: By removing our VAE module and directly learning the original expert state and action, we analyze the importance of our data augmentation technique. The results demonstrate that augmentation plays a vital role in improving both short-term and long-term simulation performance by mitigating covariate shift." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Context-conditioned VAE: To emphasize the significance of modeling the context-conditioned trajectory distribution rather than the whole state distribution, we replaced our context-conditioned VAE with a naive VAE that directly" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 449 + ], + "type": "text", + "content": "models the expert and learner state distribution. The observed drop in performance demonstrates the challenges in reconstructing the context distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 513 + ], + "type": "text", + "content": "On-road projection: Our ablation study on the on-road projection module aims to show its impact on reducing the off-road rate. The results show that the on-road projection module leads to a notable decrease in the off-road rate and moderate improvements in other performance metrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 516, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 588 + ], + "type": "text", + "content": "LQR: Removing LQR module allows us to evaluate its effectiveness. While LQR led to a higher short-term off-road rate due to more constrained movement (driving off-road due to inertia), its removal deteriorates other short-term and long-term metrics, because LQR can smooth the trajectory, thus leads to more realistic simulations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 607, + 415, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 607, + 415, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 607, + 415, + 620 + ], + "type": "text", + "content": "5.5. Qualitative Result" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "In Fig. 2, we present the mean road density and speed for real-world data, SUMO simulation, and our proposed method over all time steps. The results demonstrate that our proposed method surpasses the capabilities of the SUMO simulator at replicating long-term macroscopic traffic patterns due to our model's enhanced ability to replicate the typical microscopic driving behaviors over a long period." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15392" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 527, + 511 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 527, + 511 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 527, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 527, + 511 + ], + "type": "image", + "image_path": "7fe78b129460e56077e2dacba9ddc2df6714e3b7076b387f2f8e7ad5ec3de2b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 512, + 545, + 534 + ], + "lines": [ + { + "bbox": [ + 46, + 512, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 545, + 534 + ], + "type": "text", + "content": "Figure 2. Mean density and speed on each road over all time steps in the long-term evaluation. Our method's density and speed hot-maps have a more similar color to the ground-truth one compared with the SUMO's." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 119, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "In conclusion, we have addressed the challenge of creating a realistic traffic simulator that accurately models human driving behaviors in various traffic conditions. Traditional imitation learning-based simulators often fail to deliver accurate long-term simulations due to the covariate shift problem in multi-agent imitation learning. To tackle this issue, we proposed a learner-aware supervised imitation learning method, which leverages a context-conditioned VAE to generate learner-aware augmented expert states. We leverage a context-conditioned VAE to simultaneously reconstruct the expert and learner state. This approach enables the reproduction of long-term stable microscopic traffic simulations, marking a significant advancement in the field of urban" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 538, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 597 + ], + "type": "text", + "content": "traffic simulation. Our method has demonstrated superior performance over existing state-of-the-art simulators when evaluated on the real-world dataset pNEUMA, achieving better short-term microscopic and long-term macroscopic similarity to real-world data than state-of-the-art baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 421, + 624 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This research project is partially supported by the Innovation and Technology Commission of the HKSAR Government under the InnoHK initiative, Hong Kong General Research Fund (11202119 and 11208718), Innovation and Technology Commission (GHP/126/21GD), and Guangdong, Hong Kong and Macao Joint Innovation Project (2023A0505010016)." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15393" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Karl Johan Åström and Richard M Murray. Feedback systems: an introduction for scientists and engineers. Princeton University Press, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 167 + ], + "type": "text", + "content": "[2] Jaime Barceló and Jordi Casas. Dynamic network simulation with aimsun. Simulation approaches in transportation analysis: Recent advances and challenges, pages 57-98, 2005. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 223 + ], + "type": "text", + "content": "[3] Emmanouil Barmpounakis and Nikolas Geroliminis. On the new era of urban traffic monitoring with massive drone data: The pNEUMA large-scale field experiment. Transportation research part C: emerging technologies, 111:50-71, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 289 + ], + "type": "text", + "content": "[4] Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih Hu, Luca Del Pero, Blazej Osinski, Hugo Grimmett, and Peter Ondruska. SimNet: Learning reactive self-driving simulations from real-world observations. IEEE International Conference on Robotics and Automation, pages 5119-5125, 2021. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "text", + "content": "[5] Raunak P. Bhattacharyya, Derek J. Phillips, Changliu Liu, Jayesh K. Gupta, K. Driggs-Campbell, and Mykel J. Kochenderfer. Simulating emergent properties of human driving behavior using multi-agent reward augmented imitation learning. International Conference on Robotics and Automation, pages 789-795, 2019. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 423 + ], + "type": "text", + "content": "[6] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Bei-jbom, and Sammy Omari. Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. In IEEE/CVF Computer Vision and Pattern Recognition Conference ADP3 workshop, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 288, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 478 + ], + "type": "text", + "content": "[7] Qianwen Chao, Huikun Bi, Weizi Li, Tianlu Mao, Zhaoqi Wang, Ming C Lin, and Zhigang Deng. A survey on visual traffic simulation: Models, evaluations, and applications in autonomous driving. In Computer Graphics Forum, pages 287-308, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 479, + 286, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 286, + 502 + ], + "type": "text", + "content": "[8] James Colyar and John Halkias. Us highway 101 dataset. Federal Highway Administration, Tech. Rep., 2007. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 544 + ], + "type": "text", + "content": "[9] Robert Dadashi, L'essonard Hussenot, Matthieu Geist, and Olivier Pietquin. Primal wasserstein imitation learning. In International Conference of Learning Representation, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[10] Christian Schroeder de Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "type": "text", + "content": "[11] Frederik Diehl, Thomas Brunner, Michael Truong Le, and Alois Knoll. Graph neural networks for modelling traffic participant interaction. In IEEE Intelligent Vehicles Symposium, pages 695-701, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[12] Shuo Feng, Xintao Yan, Haowei Sun, Yiheng Feng, and Henry X. Liu. Intelligent driving intelligence test for autonomous vehicles with naturalistic and adversarial environment. Nature Communications, 12, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 712 + ], + "type": "text", + "content": "[13] Justin Fu, Katie Luo Luo, and Sergey Levine. Learning robust rewards with adversarial inverse reinforcement learn" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "ing. In International Conference of Learning Representation, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63:139 - 144, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[15] Ke Guo, Wenxi Liu, and Jia Pan. End-to-end trajectory distribution prediction based on occupancy grid maps. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2232-2241, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[16] Ke Guo, Wei Jing, Junbo Chen, and Jia Pan. CCIL: Context-conditioned imitation learning for urban driving. In Robotics: Science and Systems, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 252 + ], + "type": "text", + "content": "[17] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[18] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. In Conference of Robot Learning, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "text", + "content": "[19] Arne Kesting, Martin Treiber, and Dirk Helbing. General lane-changing model MOBIL for car-following models. Transportation Research Record, 1999(1):86-94, 2007. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 344, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 344, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 344, + 545, + 365 + ], + "type": "text", + "content": "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 366, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 388 + ], + "type": "text", + "content": "[21] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 389, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 445 + ], + "type": "text", + "content": "[22] Yann Koeberle, Stefano Sabatini, Dzmitry V. Tsishkou, and Christophe Sabourin. Exploring the trade off between human driving imitation and safety for traffic simulation. IEEE International Conference on Intelligent Transportation Systems, pages 779-786, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "text", + "content": "[23] Robert Krajewski, Julian Bock, Laurent Kloeker, and Lutz Eckstein. The highD dataset: A drone dataset of naturalistic vehicle trajectories on german highways for validation of highly automated driving systems. In International Conference on Intelligent Transportation Systems, pages 2118-2125, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[24] Daniel Krajzewicz, Jakob Erdmann, Michael Behrisch, and Laura Bieker. Recent development and applications of SUMO-Simulation of Urban MObility. International journal on advances in systems and measurements, 5(3&4), 2012. 1, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[25] Michael Laskey, Jonathan Lee, Roy Fox, Anca Dragan, and Ken Goldberg. Dart: Noise injection for robust imitation learning. In Conference on Robot Learning, pages 143-156, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[26] Quanyi Li, Zhenghao Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in Neural Information Processing Systems, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Weizi Li, David Wolinski, and Ming C. Lin. City-scale traffic animation using statistical learning and metamodel-based optimization. ACM Trans. Graph., 36(6):200:1-200:12, 2017. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15394" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Weizi Li, David Wolinski, and Ming C. Lin. ADAPS: Autonomous driving via principled simulations. In IEEE International Conference on Robotics and Automation (ICRA), pages 7625-7631, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 171 + ], + "type": "text", + "content": "[29] Michael James Lighthill and Gerald Beresford Whitham. On kinematic waves ii. a theory of traffic flow on long crowded roads. Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229(1178):317-345, 1955. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "text", + "content": "[30] Weiwei Liu, Wei Jing, Ke Guo, Gang Xu, Yong Liu, et al. Traco: Learning virtual traffic coordinator for cooperation with multi-agent reinforcement learning. In Conference on Robot Learning, pages 2465-2477, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 284 + ], + "type": "text", + "content": "[31] Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bronstein, Becca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. arXiv preprint arXiv:2212.11419, 2022. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 319 + ], + "type": "text", + "content": "[32] Donald Michie, Michael Bain, and Jean Hayes-Michie. Cognitive models from subcognitive skills. IEEE control engineering series, 44:71-99, 1990. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 364 + ], + "type": "text", + "content": "[33] Xiaoyu Mo, Zhiyu Huang, Yang Xing, and Chen Lv. Multi-agent trajectory prediction with heterogeneous edge-enhanced graph attention network. IEEE Transactions on Intelligent Transportation Systems, 23:9554-9567, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[34] Jeremy Morton, Tim A Wheeler, and Mykel J Kochenderfer. Analysis of recurrent neural networks for probabilistic modeling of driver behavior. IEEE Transactions on Intelligent Transportation Systems, 18(5):1289-1298, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[35] Andrew Y Ng, Stuart Russell, et al. Algorithms for inverse reinforcement learning. In International Conference of Machine Learning, page 2, 2000. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 466 + ], + "type": "text", + "content": "[36] Paul I Richards. Shock waves on the highway. Operations research, 4(1):42-51, 1956. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 520 + ], + "type": "text", + "content": "[37] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In International Conference on Artificial Intelligence and Statistics, pages 627-635, 2011. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[38] Jason Sewall, David Wilkie, Paul Merrell, and Ming C Lin. Continuum traffic simulation. In Computer Graphics Forum, number 2, pages 439-448, 2010. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[39] Yu Shen, Weizi Li, and Ming C. Lin. Inverse reinforcement learning with hybrid-weight trust-region optimization and curriculum learning for autonomous maneuvering. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7421-7428, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[40] Jiaming Song, Hongyu Ren, Dorsa Sadigh, and Stefano Ermon. Multi-agent generative adversarial imitation learning. Advances in neural information processing systems, 31, 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[41] Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multiagent behaviors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400-10409, 2021. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 319 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[42] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[43] David Wilkie, Jason Sewall, Weizi Li, and Ming C. Lin. Virtualized traffic at metropolitan scales. Frontiers in Robotics and AI, 2:11, 2015. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[44] QI Yang and Haris N Koutsopoulos. A microscopic traffic simulator for evaluation of dynamic traffic management systems. Transportation Research Part C: Emerging Technologies, 4(3):113-129, 1996. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[45] Kaifeng Zhang, Rui Zhao, Ziming Zhang, and Yang Gao. Auto-encoding adversarial imitation learning. arXiv preprint arXiv:2206.11004, abs/2206.11004, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[46] Guanjie Zheng, Hanyang Liu, Kai Xu, and Zhenhui Jessie Li. Objective-aware traffic simulation via inverse reinforcement learning. In International Joint Conference on Artificial Intelligence, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 545, + 319 + ], + "type": "text", + "content": "[47] Brian D Ziebart, Andrew L Maas, J Andrew Bagnell, and Anind K Dey. Maximum entropy inverse reinforcement learning. In AAAI Conference on Artificial Intelligence, pages 1433-1438, 2008. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15395" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_content_list.json b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e9c14ba5d143842fea9af898ba521e28cfd3b9bd --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "LASO: Language-guided Affordance Segmentation on 3D Object", + "text_level": 1, + "bbox": [ + 155, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yicong Li $^{1}$ , Na Zhao $^{2*}$ , Junbin Xiao $^{1}$ , Chun Feng $^{3}$ , Xiang Wang $^{3*†}$ , Tat-seng Chua $^{1}$ ,", + "bbox": [ + 156, + 179, + 810, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National University of Singapore, $^{2}$ Singapore University of Technology and Design, $^{3}$ University of Science and Technology of China,", + "bbox": [ + 148, + 214, + 818, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liyicong@u.nus.edu,na_zhao@sutd.edu.sg,junbin@comp.nus.edu.sg", + "bbox": [ + 215, + 253, + 751, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fengchun3364@mail.ustc.edu.cn,xiangwangl223@gmail.com,dcstsc@nus.edu.sg", + "bbox": [ + 168, + 272, + 795, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmenting affordance in 3D data is key for bridging perception and action in robots. Existing efforts mostly focus on the visual side and overlook the affordance knowledge from a semantic aspect. This oversight not only limits their generalization to unseen objects, but more importantly, hinders their synergy with large language models (LLMs) which are excellent task planners that can decompose an overarching command into agent-actionable instructions. With this regard, we propose a novel task, Language-guided Affordance Segmentation on 3D Object (LASO), which challenges a model to segment a 3D object's part relevant to a given affordance question. To facilitate the task, we contribute a dataset comprising 19,751 point-question pairs, covering 8434 object shapes and 870 expert-crafted questions. As a pioneer solution, we further propose PointRefer, which highlights an adaptive fusion module to identify target affordance regions at different scales. To ensure a text-aware segmentation, we adopt a set of affordance queries conditioned on linguistic cues to generate dynamic kernels. These kernels are further used to convolute with point features and generate a segmentation mask. Comprehensive experiments and analyses validate PointRefer's effectiveness. With these efforts, We hope that LASO can steer the direction of 3D affordance, guiding it towards enhanced integration with the evolving capabilities of LLMs. Code and data are available at https://github.com/y13800/LASO.", + "bbox": [ + 75, + 354, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Affordance describes specific regions on an object that enable or facilitate certain human interactions. In this context,", + "bbox": [ + 76, + 820, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/304263c7468019abd031dd459f80d845303ab370be9420da4a6eb9a7ffaed1e1.jpg", + "image_caption": [ + "Figure 1. Illustration of LASO, language-guided affordance segmentation on 3D object. Given an object point cloud and a question, the model is expected to segment the functional part delineated in question." + ], + "image_footnote": [], + "bbox": [ + 503, + 320, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the task of 3D affordance segmentation is raised to identify parts of 3D objects that imply a certain function. It has been deemed as a critical step in bridging perception and operation in the physical world for an embodied agent, thus has shown substantial impact on practical applications such as robotic manipulation [10, 14, 31].", + "bbox": [ + 496, + 563, + 892, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the advancement of Large Language Models (LLMs) has enabled their deployment as task planners, breaking down an intricate instruction into a sequence of affordance-aware, robot-executable instructions [12, 35, 36]. This equips intelligent agents with the capability to undertake complex tasks. For example, the overarching command \"Pass me the mug in the microwave\" can be systematically deconstructed into a series of affordance-aware inquiries: Q1) \"How to get to the microwave?\", Q2) \"How to open the microwave\", Q3) \"How to grab the mug?\", and Q4) \"How to return to the user?\" Ideally, each query would correspond to a predefined subtask that effectively translates language instructions into actionable steps. While Q1 and Q4 are related to language-conditioned path planning [42] and navigation [37], Q2 and Q3, which involve the 'open microwave' and 'grab the mug' actions, ex", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 862, + 222, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Xiang Wang is also affiliated with Institute of Artificial Intelligence,", + "bbox": [ + 94, + 875, + 467, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Institute of Dataspace, Hefei Comprehensive National Science Center.", + "bbox": [ + 78, + 888, + 447, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pose a gap — to the best of our knowledge, currently understanding affordance via language cues remains unexplored.", + "bbox": [ + 76, + 90, + 467, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leaving language cues untouched, current 3D affordance segmentation focuses primarily on the visual aspect. It associates the specific geo-structure with either explicit affordance categories [7] or implicit affordance information from the 2D showcase [44]. Such visual-only settings make the segmentation models confined to visually predefined affordance types, thus impeding integration with language-based instructions from LLMs. As a result, fulfilling the aforementioned user's request becomes challenging, due to this inability to address Q2 and Q3.", + "bbox": [ + 75, + 123, + 467, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards these limitations, we introduce a novel task: Language-guided Affordance Segmentation on 3D Object (LASO), where 3D objects are paired with natural language questions that probe specific affordance parts. As a cornerstone task, LASO can provide an agent with affordance knowledge to answer the aforementioned Q2 and Q3 raised by LLM. For instance, the first two rows in Fig. 1 demonstrate how LASO imparts knowledge that enables an agent to interpret the semantic instructions \"open the microwave\" and \"grasp the mug\" in a 3D context, which are sufficient to execute Q2 and Q3, respectively. In LASO, questions are crafted to reflect diverse real-world scenarios, translating various affordance types into a range of semantic contexts. By learning from different contexts, we expect the agent not to memorize predefined affordance types but rather to abstract generalizable affordance knowledge within a broader semantic context. Notably, the diverse scales and shapes of affordance parts pose a significant perceptual challenge for the model. For instance, in the third row of Fig. 1, the target backrest of a chair occupies a large portion and has a significantly different shape compared to the small knob on a door in the last row.", + "bbox": [ + 75, + 276, + 468, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate the study, we develop the 1st question-affordance dataset comprising 19,751 point cloud-question pairs. The questions, originally crafted by experts, are further diversified using GPT-4 [33] under principles of contextual enrichment, concise phrasing, and structural diversity. These principles enable a deep and varied exploration of object affordances. Based on the dataset, we have developed a strong baseline model named PointRefer that is capable of adapting to varying scales and shapes of object parts, conditioned on the question. It consists of two modules: 1) an Adaptive Fusion Module (AFM) that integrates question semantics to point features at different scales, which caters to varied target regions with better point feature discriminativeness; and 2) a Referred Point Decoder (RPD) that generates a segmentation mask via conditioned affordance queries, which, after refinement, forms dynamic convolution kernels to capture point-wise affordance for mask prediction. In summary, our contributions are as follows:", + "bbox": [ + 75, + 611, + 467, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose the task of Language-guided Affordance Seg", + "bbox": [ + 76, + 885, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/31b0b464e12a209949a0fe651ce2f3ee5cf847ceaeb881a076774d1035daa881.jpg", + "table_caption": [ + "Table 1. Efforts on affordance learning. Part indicates whether the prediction specifies the object part (✓) or the whole object $(\\times)$ ." + ], + "table_footnote": [], + "table_body": "
DomainStudiesVisionTextPart
2D[1, 46]Image××
[9, 25, 32]Image/Video×✓
[24, 29]Image✓×
3D[7, 30, 30, 43]Point Cloud×✓
[44]Point Cloud & Image×✓
OursPoint Cloud✓✓
", + "bbox": [ + 519, + 119, + 875, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mentation on 3D Objects (LASO), which helps AI agents derive affordance knowledge from textual cues, facilitating a crucial advancement for LLM integration.", + "bbox": [ + 511, + 232, + 890, + 277 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a dataset with 19,751 question-point affordance pairs which are meticulously curated via a collaborative effort of human experts and GPT-4, thereby establishing the 1st benchmark for evaluating LASO.", + "- We build a strong baseline PointRefer, which employs textual-conditioned affordance queries to isolate afforded segments and devises an adaptive fusion module to enhance the discriminability of point feature." + ], + "bbox": [ + 500, + 277, + 890, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 407, + 640, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Affordance Learning. Originating from the image domain, initial efforts concentrate on detecting objects with affordances [1, 46]. Progressing from there, later studies [24, 29] incorporate linguistic descriptions to augment the process but still neglect the granularity of analysis, generally focusing on object-level affordances. Addressing this oversight, subsequent research [9, 25, 32] shift towards scrutinizing specific affordance parts, establishing a new norm for precision in the field.", + "bbox": [ + 496, + 424, + 890, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of embodied AI, the scope of affordance learning expands into the 3D realm. 3D AffordanceNet [7] introduces the first benchmark dataset for learning affordance from object point cloud geometry. Building on this, Yang et al. [44] propose a setting for learning 3D affordance parts guided by image demonstrations. Nonetheless, these approaches predominantly rely on a visual-only approach, tightly coupling geometric features with affordance labels and overlooking the semantic dimension. This limitation hinders the integration with Large Language Models (LLMs) that could otherwise substantiate affordance segmentation to real-world deployment. Moreover, the common practice of predicting affordance types as an auxiliary task [7, 44] may dilute the primary challenge of understanding affordance knowledge. In contrast, our work sidesteps auxiliary predictions, advocating for direct learning from the linguistic context, which aligns more closely with the innate capabilities of LLMs and their semantic richness.", + "bbox": [ + 496, + 561, + 890, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-Point Cloud Cross-Modal Learning. The integration of vision and natural language has recently sparked increased interest [16-21, 41]. Such a tendency naturally extends to 3D vision due to its pivotal role in embodied AI. To", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f02f4f367fd6504dc0599f11f0b7da47dce2cc5d8670c945bc01f637a759f629.jpg", + "image_caption": [ + "Figure 2. Illustration of question augmentation principles." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 446, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "advance the 3D-language tasks, datasets like ScanRefer [5] and ReferIt3D [2] have been instrumental in benchmarking the grounding of natural language to 3D objects. Complementing this, Azuma et al. [3] introduce ScanQA for 3D question-answering, while Ma et al. [28] propose SQA3D for situated reasoning within 3D environments. In terms of models, a variety of efforts [3, 13, 45] have emerged from these benchmarks, with notable examples like 3D-SPS [26] and BUTD-DETR [13] employing cross-attention mechanisms and linguistic cues for object discovery. Others, such as 3DVG [45], and ViL3DRel [6], focus on 3D grounding by incorporating spatial relation cues, showcasing the evolving landscape of 3D-VL models. Unlike the existing work that focuses on grounding and QA that perform scene-level reasoning, we tackle a dense prediction task at object part level that bridges reasoning to operation. To the best of our knowledge, we are the first to introduce the 3D affordance problem as a language-guided segmentation task.", + "bbox": [ + 75, + 262, + 472, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 76, + 542, + 165, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To underpin our proposed task, we have meticulously compiled a question-point affordance dataset with 19,751 paired samples, featuring 8,434 distinct object shapes across 23 classes. Accompanying these are 870 expertly crafted questions, covering 17 affordance categories.", + "bbox": [ + 75, + 560, + 468, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Annotation Details", + "text_level": 1, + "bbox": [ + 76, + 645, + 256, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point Cloud. Our dataset leverages the point cloud and affordance annotations from the 3D-AffordanceNet [7]. Specifically, the affordance annotation in [7] is multi-class, with each point on an object potentially supporting multiple affordance types. In our case, each question is tailored to a specific affordance type, while each sample can associate with several questions. Consequently, the same object can exhibit diverse affordance segmentation, depending on the specific question posed.", + "bbox": [ + 75, + 669, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Question. To create the questions, we draw 58 object-affordance combinations from [7], and craft 15 unique questions for each of the 58 combinations, resulting in 870 tailored questions. Specifically, five manual questions per combination are initially composed to represent different scenarios. Subsequently, we employ GPT-4 [33] to diver", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/d709bc24a5eff4cb6cad052fd9b1c2a9cf88f7a166ec0b98c98eed85d9db1c07.jpg", + "table_caption": [ + "Table 2. Dataset Statistics" + ], + "table_footnote": [], + "table_body": "
TaskTrainValTest
ShapeQstInstShapeQstInstShapeQstInst
Seen68836381612051658121510351742416
Unseen616045811558
", + "bbox": [ + 506, + 106, + 888, + 166 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/058b1cac9cc8965d2e5d05a665f8f8ce3c0ea0c2d9db887094e1c4c1334812c5.jpg", + "image_caption": [ + "Figure 3. Landscape of our dataset." + ], + "image_footnote": [], + "bbox": [ + 501, + 174, + 888, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sify the dataset with 10 additional questions for each combination, adhering to three key principles (as shown in Fig. 2): 1) Contextual Enrichment. Some questions were expanded to include additional context or details, fostering a more precise connection to the specified affordance of the target object. For instance, Q1 in Fig. 2 is reformed by a functional situation \"cutting paper\". 2) Concise Phrasing. In other cases, questions were distilled to their essence, making them succinct yet still meaningful. Q2 in Fig. 2 exemplifies this approach. 3) Structural Diversity. We employed a range of sentence structures, from questions to statements, to introduce natural linguistic variations and prevent model bias toward any specific phrasing or sentence length. A diversified example can be found in the Q3 of Fig. 2. All questions generated by GPT-4 undergo manual verification after this augmentation process.", + "bbox": [ + 496, + 333, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It's worth mentioning that the affordance type was only a tool for crafting the questions and pinpointing the correct segmentation areas. We refrain from using explicit affordance labels during both training and testing. This deliberate choice ensures that the learning process is driven by the semantic content of the questions, rather than by direct affordance label association.", + "bbox": [ + 496, + 575, + 890, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Statistics and Setting", + "text_level": 1, + "bbox": [ + 500, + 695, + 697, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset Settings. We show the statistical landscape of our dataset in Fig. 3. Specifically, it contains 17 affordance types and 23 object categories, composing 58 unique object-affordance combinations.", + "bbox": [ + 496, + 719, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the spirit of the work by [44], we offer two distinct dataset settings: Seen: The default configuration, where the training and testing phases share similar distributions of object classes and affordance types. Unseen: This configuration is purpose-built to test the model's ability to generalize affordance knowledge. Certain affordance types, when paired with specific object classes, are omitted from the training set but included in the test set. For instance, while", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model may learn to grasp bags and mugs during training, it is expected to apply the concept of 'grasp' to earphones, which is an affordance-object pairing not encountered during training. For a comprehensive account of the unseen setting, please refer to Appendix Sec. 7.1.", + "bbox": [ + 75, + 90, + 467, + 166 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training and Evaluation Protocols. Notably, since questions are created based on a combination of object class and affordance type, one object class can have many shape instances. Thus, the design of our dataset accounts for the multiplicity of shapes within each object class, leading to a non-bijective pairing between shape instances and corresponding questions. To ensure comprehensive learning, during training, each shape instance is matched with a randomly selected question that aligns with its affordance type for every iteration. This random pairing exposes the model to a variety of semantic contexts, bolstering its ability to generalize. For the validation and test sets, question pairings are fixed to minimize variability and ensure consistent inference. These questions are unique to the evaluation phase and are not revealed during the training process to maintain the integrity of the evaluation.", + "bbox": [ + 75, + 167, + 468, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 410, + 166, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LASO Task Definition. Given a question $Q_{raw}$ and an object point cloud $\\mathbf{P}_{raw} \\in \\mathbb{R}^{N \\times 3}$ with $N$ points, the goal of LASO is to predict a binary mask of $\\mathbf{M} \\in \\mathbb{R}^N$ that segments the functional part related to the question.", + "bbox": [ + 75, + 436, + 468, + 497 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Framework Overview. Due to the varying scale of target affordance region, LASO naturally challenges models adaptiveness at different scales. To address this, we devise a strong baseline, named PointRefer. As shown in Fig. 4, the object point cloud $\\mathbf{P}_{raw}$ is first processed by a 3D backbone, which typically consists of the multi-stage encoding and decoding process. To adapt PointRefer to the point feature at multiple scales, we introduce an Adaptive Fusion Module (AFM) to the backbone's decoding process. It performs multi-scale cross-modal fusion by injecting text clues to point features at a different decoding stage, which progressively refine the point feature map in a top-down manner. Then, to predict the segmentation mask, we introduce a Referred Point Decoder (RPD) that leverages a set of learnable queries conditioned on the input question, termed affordance queries, as the input of the transformer decoder. These affordance queries are obligated to look at the referred points only and generate question-aware dynamic kernels. The final segmentation mask is obtained by convoluting these dynamic kernels with the AFM-enhanced point feature.", + "bbox": [ + 75, + 502, + 468, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Encoding. For the question, we prepend a \"[CLS]\" token to the question sequence before feeding them to a language model to capture the global language context. The encoded feature of the question is denoted by $\\mathbf{X} \\in \\mathbb{R}^{L \\times d}$ , where $L$ is the number of tokens and $d$ is the feature di", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mension. For the point cloud, PointRefer can adopt any off-the-shelf 3D segmentation backbone.1 to transform the $\\mathbf{P}_{raw}$ into the feature space. Note that a 3D backbone for segmentation typically consists of several encoding and decoding stages, and we take the point-wise feature map after the last decoding stage as the output of the backbone.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Adaptive Fusion Module", + "text_level": 1, + "bbox": [ + 498, + 191, + 725, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AFM is designed to enhance the point-wise features by incorporating question information. As illustrated in Fig. 4, we integrate AFM into different decoder layers to enhance its adaptiveness to target regions of various scales and shapes. Given the conventional downsampling operators in 3D backbones (e.g. PointNet++ [34]), the number of points varies across different encoder/decoder layers. Our AFM is designed to seamlessly adapt to the varying number of points.", + "bbox": [ + 496, + 214, + 890, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplification, we omit the notation for different decoder layers and use $\\mathbf{P} \\in \\mathbb{R}^{T \\times d}$ to denote the point feature at a certain decoding layer, where $T$ represents the number of points at that layer. As visualized in Fig. 4, AFM adopts a bottleneck architecture with three key steps: Grouping, Mixing, and Ungrouping. In the grouping step, the point features are grouped by text tokens, these grouped tokens are then thoroughly mixed via an MLP-Mixer [38]. Subsequently, the ungrouping step encapsulates the mixed features into individual point features, completing the AFM procedure by embedding valuable textual information into the point feature.", + "bbox": [ + 496, + 349, + 890, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To facilitate an efficient grouping process for our AFM, we employ a lightweight cross-attention module by omitting the transformation of both query and value. It takes the question feature $\\mathbf{X}$ as query and the point feature $\\mathbf{P}$ as key and value, then output grouped token $\\mathbf{G} \\in \\mathbb{R}^{L \\times d}$ :", + "bbox": [ + 496, + 531, + 890, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {X}, \\mathbf {W} _ {1} \\mathbf {P}, \\mathbf {P}\\right) + \\mathbf {X}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 618, + 890, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 638, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_1$ is a linear transformation.", + "bbox": [ + 500, + 683, + 743, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After attaining the grouped feature, we update them using an MLP-Mixer [38], which employs two consecutive multilayer perceptrons (MLPs), producing fused feature $\\mathbf{F}$ by combining with two consecutive MLPs.", + "bbox": [ + 496, + 698, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} ^ {\\prime} = \\mathbf {G} + \\operatorname {M L P} _ {1} \\left(\\mathbf {G} ^ {T}\\right) ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 768, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\mathbf {G} ^ {\\prime} + \\mathrm {M L P} _ {2} \\left(\\mathbf {G} ^ {\\prime}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 787, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $\\mathrm{MLP}_1$ and $\\mathrm{MLP}_2$ are used to mix group-wise and channel-wise information, respectively, and $T[\\cdot]$ denotes the transpose operation. Finally, we return the fused feature $\\mathbf{F}$ to the point-space via the ungrouping process and", + "bbox": [ + 496, + 816, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The impact of backbone choices can be observed in Tab. 5.", + "bbox": [ + 517, + 886, + 834, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9b18ca47acf1010aba4b1a6f550693ca4313b08db5a15d713f522946cff3c15.jpg", + "image_caption": [ + "(a) Overview of PointRefer." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 633, + 353 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d72ddcef34d4d0b43226e356b67b7c1342c77602abfa91c7a9b82df45cf402de.jpg", + "image_caption": [ + "(b) Adaptive Fusion Module.", + "Figure 4. The overall framework of our proposed PointRefer (a), which contains an adaptive fusion modal (b) that can be injected in different stages of the backbone decoder. As a result, it takes the point-wise feature generated at the last decoder layer as the final output of the 3D backbone. Then, the Referred Point Decoder takes in a set of question-conditioned affordance queries and feeds it to a transformer decoder together with the fused point feature, where the output is then reformed to a set of dynamic kernels. Finally, these dynamic kernels are used as convolution kernels to filter out the segmentation mask from the fused point feature." + ], + "image_footnote": [], + "bbox": [ + 678, + 93, + 893, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "acquire the fused point feature $\\mathbf{P}^m$ . We implement the ungrouping process as a similar lightweight attention module as the grouping process. However, in this case, it takes $\\mathbf{P}$ as query and $\\mathbf{F}$ as key and value:", + "bbox": [ + 75, + 446, + 468, + 506 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {m} = \\operatorname {A t t e n t i o n} (\\mathbf {P}, \\mathbf {W} _ {2} \\mathbf {F}, \\mathbf {F}) + \\mathbf {P}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 518, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_2$ is a linear transformation. Lastly, the final output $\\mathbf{P}^o$ is formed by adding residual connections to $\\mathbf{P}^m$ . Notably, we take $\\mathbf{P}^o$ after the last decoding stage as the output of the 3D backbone.", + "bbox": [ + 75, + 547, + 468, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Referred Point Decoder", + "text_level": 1, + "bbox": [ + 76, + 617, + 294, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by the success of learnable query-based methods in object detection [4] and segmentation [40], we introduce learnable affordance queries conditioned on the corresponding question of an object to decode the segmentation mask for the object. Specifically, we impose linguistic restrictions on all affordance queries, narrowing the model's focus to the relevant object parts and thereby facilitating convergence. These affordance queries, after being processed by a transformer decoder, are repurposed as dynamic kernels to deduce the segmentation mask from the point features, which are the final output of the 3D backbone.", + "bbox": [ + 75, + 638, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question-Conditioned Affordance Query In the transformer decoder, it is well known that content and position embeddings are responsible for encoding instance-specific and spatial information, respectively. In PointRefer, we feed these two parts with the text feature and learnable query parameters, respectively, so that all the queries are", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "shaped by the nuances of the linguistic input. Specifically, the affordance queries $\\mathbf{A}$ are formed by adding question embedding $\\mathbf{X}$ with learnable embeddings. Thus, the resulting $\\mathbf{A} \\in \\mathbb{R}^{L \\times d}$ are the same length of the question.", + "bbox": [ + 498, + 446, + 890, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reasoning and Mask Prediction. After forming the affordance query, we feed it to a transformer decoder [39] together with the fused point feature $\\mathbf{P}_o$ from the 3D backbone. In this manner, all the queries will use the question context as guidance and target to aggregate the feature of the referred points from the object, resulting in $\\mathbf{A}'$ :", + "bbox": [ + 498, + 510, + 890, + 601 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {\\prime} = \\text {T r a n s f o r m e r - D e c o d e r} (\\mathbf {A}; \\mathbf {P} _ {o}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 613, + 890, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we perform dynamic convolution by applying a two-layer MLP on top of the transfer decoder to produce $L$ dynamic kernels $\\Omega = \\{\\omega_{i}\\}_{i = 1}^{L}$ , where each kernel $\\omega_{i}$ is reshaped to a $1\\times 1$ convolutional kernel with the channel number of $d$ .", + "bbox": [ + 498, + 642, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega = \\operatorname {M L P} \\left(\\mathbf {A} ^ {\\prime}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 729, + 890, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the dynamic kernels have captured the question- information, we use them as convolution filters on the feature maps for mask decoding. Specifically, we use each dynamic kernel in $\\Omega$ to convolute the point-wise feature map $\\mathbf{P}_o$ , which returns $L$ point masks $\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^L$ .", + "bbox": [ + 498, + 758, + 890, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {i} = \\left\\{\\mathbf {P} _ {o} * \\omega_ {i} \\right\\} _ {i = 1} ^ {L}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 839, + 890, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we apply mean pooling over all masks in $\\mathbf{S}$ followed by a sigmoid activation to acquire the final segmentation", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mask $\\mathbf{M}\\in \\mathbb{R}^N$ ..", + "bbox": [ + 76, + 90, + 181, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\sigma (\\operatorname {M a x - P o o l} ((\\mathbf {S}))), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 113, + 468, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ denote the sigmoid function.", + "bbox": [ + 76, + 137, + 344, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, each dynamic kernel is expected to capture one aspect of the referred object part according to conditioned the question token. By ensembling all generated point masks, the final segmentation mask can integrate different response regions, thus making the result more robust.", + "bbox": [ + 76, + 152, + 468, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Objectives. Unlike [44], which relies on an auxiliary affordance label for prediction, our model seeks to forge a direct link between language context and object affordance. Thus, we solely employ Dice loss and Binary Cross-Entropy (BCE) loss to guide the segmentation mask prediction, bypassing the need for additional affordance labels.", + "bbox": [ + 76, + 229, + 468, + 320 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D i c e} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 323, + 468, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 332, + 200, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For a thorough assessment, we benchmark our dataset against leading studies in 3D affordance learning [7, 44], employing four evaluation metrics to ensure robust analysis. These include Area Under the Curve (AUC), Mean Intersection Over Union (mIoU), Similarity (SIM), and Mean Absolute Error (MAE). Appendix Sec. 7.2 provides detailed explanations for the evaluation protocols.", + "bbox": [ + 76, + 357, + 468, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. PointRefer employs PointNet++ as the default 3D backbone to align with the standards established in [7, 44]. For text encoding, we utilize a pretrained RoBERTa model [23] to process linguistic inputs. The feature dimension $d$ is set to 512. During training, we use the Adam optimizer with a learning rate set to 1e-4.", + "bbox": [ + 76, + 465, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Models. Since there is no prior work using paired question-point cloud data to segment object affordance. For a thorough comparison of our method, we adopt two 3D cross-modal works (3D-SPS [26] and IAGNet[44]) and two referred image segmentation works (ReferTrans [15] and RelA [22]) to LASO task. For the referred image segmentation models, we substitute their image backbones with a 3D counterpart, retaining their fusion mechanisms and mask decoding strategies. For the language-guided grounding model, 3D-SPS [26], we omit its bounding box prediction module while leveraging its inherent point selection scheme for our segmentation setting. The most related work is IAGNet [44], an affordance detection method that takes paired image-point cloud as input. To adapt IAGNet to our needs, we simply replace its image backbone with a language model, preserving the remainder of its architecture.", + "bbox": [ + 76, + 559, + 468, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we show the effectiveness of our PointRefer by answering the following questions:", + "bbox": [ + 76, + 806, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Q1: How is PointRefer compared to other baselines on the proposed dataset?", + "- Q2: How effective are the proposed components?", + "- Q3: What learning pattern does the PointRefer capture?" + ], + "bbox": [ + 76, + 837, + 467, + 897 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/37931b6d0726af191bd50a6c013764788cb4c8cb21bc1cd62b40c5eef736dbba.jpg", + "table_caption": [ + "Table 3. Main Results. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. AUC and aIOU are shown in percentage." + ], + "table_footnote": [], + "table_body": "
MethodmIoU↑AUC↑SIM↑MAE↓
SteenReferTrans [15]13.779.80.4970.124
ReLA [22]15.278.90.5320.118
3D-SPS [26]11.476.20.4330.138
IAGNet [44]17.882.30.5610.109
PointRefer20.887.30.6290.093
UnseenReferTrans [15]10.269.10.4320.145
ReLA [22]10.769.70.4290.144
3D-SPS [26]7.968.80.4020.158
IAGNet [44]12.977.80.4430.129
PointRefer14.680.20.5070.119
", + "bbox": [ + 517, + 133, + 879, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Main Result (Q1)", + "text_level": 1, + "bbox": [ + 500, + 309, + 669, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 3, PointRefer demonstrates superior performance across all evaluation metrics compared to the baseline methods. Our observations are as follows:", + "bbox": [ + 498, + 332, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PointRefer vs. Other Models: IAGNet secures the second-best performance in both seen and unseen settings for all metrics. This is likely due to its specific design for object affordance segmentation, despite differences in input modality, making it a strong fit for our task with minimal adaptation. In contrast, 3D-SPS underperforms, potentially due to its progressive point selection mechanism. This approach involves a non-differentiable selection process, which may complicate the optimization within our task framework. Image-based methods, while showing decent performance, appear to be constrained by the modality gap. Originally trained on vast datasets of image-text pairs, where text is often a brief label, these methods now face fully formed, context-rich questions with a more limited training corpus.", + "bbox": [ + 496, + 381, + 890, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Seen vs. Unseen Performance: A notable performance drop from seen to unseen settings is observed across all baselines. This discrepancy underscores the difficulty in learning affordance knowledge that generalizes well, corroborating the necessity of our task design. Current methods struggle to bridge this gap in knowledge transfer, suggesting an avenue for future research: utilizing the extensive commonsense knowledge embedded in LLMs could potentially mitigate this challenge.", + "bbox": [ + 496, + 611, + 890, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. In-Depth Study (Q2)", + "text_level": 1, + "bbox": [ + 500, + 756, + 694, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study. We show ablative results in Tab. 4. To verify that PointRefer prediction is based on the question, we conduct a blind test by training without questioning information, where the distinctive performance gap indicates that PointRefer has comprehended the question and can make a prediction based on it. Then, we study the effectiveness of PointRefer by removing both AFM and PRD (\"w/o AFM & PRD\"), which induces a severe performance", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14256", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe6f55be7872c2c69349e23bbb6881cb74d1615b258bfa0c469c0f7c157b16f7.jpg", + "table_caption": [ + "Table 4. Ablative Results." + ], + "table_footnote": [], + "table_body": "
VariantsmIoU↑AUC↑SIM↑MAE↓
Blind10.878.20.5060.125
w/o AFM & PRD17.782.10.5580.110
w/o AFM19.883.90.5920.104
w/o multi-scale20.487.10.6280.095
w/o PRD18.785.60.6190.098
w/o condition19.286.90.5900.100
w/o dynamic filter19.587.10.6230.096
PointRefer20.887.30.6290.093
", + "bbox": [ + 86, + 104, + 464, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0059e5e9abed1c1735ec3c0c0de6edbfdc697c4b25e56f290503c2482f64a20a.jpg", + "image_caption": [ + "Figure 5. The bar chart shows the training sample distribution over different question lengths, and the line chart shows the performance of PointRefer grouped by question length." + ], + "image_footnote": [], + "bbox": [ + 109, + 244, + 433, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "decline on every metric, thus, validating the overall effectiveness of our design. Then, we also verify the functionality of AFM via the performance drop when it is removed (\"w/o AFM). As a comparison, a relatively slighter declination is witnessed when we only apply AFM at the single scale (\"w/o multi-scale\"), which indicates the benefit of multi-scale cross-modal fusion. Then, we conduct some detailed breakdown tests to study PRD, we notice that when the affordance queries are not conditioned on the question (\"w/o condition\") the performance drops drastically, indicating that our affordance queries help to capture semantic clues in 3D space. Finally, we validate the superiority of the dynamic filtering strategy, because a declination is also observed when the dynamic filter is erased (\"w/o dynamic filter\"), and this variant adopts IAGNet's [44] mask prediction strategy as an alternative.", + "bbox": [ + 75, + 469, + 468, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Question Length. In Fig. 5, we show the performance of PointRefer on different question lengths. Specifically, The bar plot reveals the distribution of question length in the training set. Meanwhile, the line chart illustrates PointRefer's performance on different question lengths. Here, the performances are normalizing against the average number under that metric, e.g. IoU for each question length mIoU. Notably, PointRefer excels with questions that are more prevalent in the training set and with those that are lengthier—exceeding 14 words—owing to the richer contextual information they provide, which is instrumental in discerning the object's affordance.", + "bbox": [ + 75, + 719, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/078425933709c631814008e073540023c5abcd8fed7395bd4362e95c58c7e601.jpg", + "image_caption": [ + "Figure 6. Illustration of how the percentage of omitted samples affect transferability of affordance knowledge (indicated by the difference between seen and unseen). For each affordance type, the bar shows a number of omitted samples from seen, i.e. $\\frac{\\# \\text{Seen} - \\# \\text{Unseen}}{\\# \\text{Seen}} (\\%)$ , and the lines show the testing IoU of models trained on Seen and Unseen. Difference (IoU): Seen - Unseen." + ], + "image_footnote": [], + "bbox": [ + 527, + 89, + 861, + 266 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef69cb8a8a45baf52e0ce3e91993bedc4dd179df47f74081bb20faab5285d120.jpg", + "table_caption": [ + "Table 5. Study of Backbones. PN:PointNet++; PM:PointMLP." + ], + "table_footnote": [], + "table_body": "
3DLMmIoU↑AUC↑SIM↑MAE↓
SeenPM [27]Bert [8]18.285.50.6180.099
Deberta [11]19.386.50.6210.096
Roberta [23]19.686.462.30.097
PN [34]Bert [8]20.186.40.6150.101
Deberta [11]20.487.00.6240.095
Roberta [23]20.887.30.6290.093
UnseenPM [27]Bert [8]11.576.30.430.128
Deberta [11]12.376.80.4350.133
Roberta [23]12.176.80.4340.136
PN [34]Bert [8]14.178.10.470.122
Deberta [11]14.578.90.500.114
Roberta [23]14.680.20.5070.119
", + "bbox": [ + 517, + 368, + 877, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Seen vs. Unseen. Recall that to create the training set for the unseen setting, for an affordance type, we omit its combination with certain objects from seen. For example, the seen partition has \"grasp-mug\", and \"grasp-bag\" in its training, we create the unseen training set by removing the \"grasp-mug\" from the seen training set. In this case, the model is expected to learn the generalizable affordance knowledge of \"grasp\" and transfer it to an unseen object during testing. Note that seen and unseen setting shares the same validation and testing set. To investigate how PointRefer captures such transferability for each affordance type, The bar in Fig. 6 shows the percentage of samples that are omitted from seen under each affordance type. Then, the lines show the performance of the model trained with seen and unseen partitions for each affordance type, and their difference—the indicator of transferability of learned affordance knowledge. In general, we notice that the curve of seen and unseen follows a similar pattern, which indicates that PointRefer can capture a similar affordance knowledge even trained with far less object class. However, we also notice that the difference between seen and unseen enlarges as the omitted percentage grows, regardless of the absolution IoU. This is expected since the removed object can lead to under-diversified affordance representation.", + "bbox": [ + 496, + 537, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a89c80a6010d4952cc87e409c97820c74b8006f06ac5bac7b3fa9b3a5973077.jpg", + "image_caption": [ + "Figure 7. Case-Study of PointRefer's segmentation. Each showcase comes with one question and four shapes, showing the generalization ability of the affordance knowledge. The segmented affordance part is highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 116, + 61, + 861, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Choice of Backbone. To S backbone, we investigate the I with some alternative backbone that, for both seen and unseen as a 3D backbone generally pr than PointMLP, thus, we set 1 3D backbone. As for the lamance of Roberta prevails Ber setting. Whereas in the unseen and Roberta brings no signifca PointRefer uses Roberta as th", + "bbox": [ + 76, + 465, + 292, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 651, + 261, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Case Study. As shown in Fig. 1, the capacity to accurately co- given the question. It is noteworthy", + "bbox": [ + 76, + 672, + 276, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with small affordance components, such as the microwa door handle, our model still exhibits decent ability.", + "bbox": [ + 76, + 719, + 454, + 750 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3686ed50d26f28d574a77a62645f5c7782411961bc27d699591f15dd37a068ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 473, + 390, + 617 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8763725ac8de1fd0427376685dcdc04c002dc725bdaa0231d50732676b7f6020.jpg", + "image_caption": [ + "Figure 8. Failure cases due to non-comprehended text (left) and over-segmentation (right)." + ], + "image_footnote": [], + "bbox": [ + 401, + 474, + 545, + 614 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abd0418a56bb2f9c13e6c5d1c927312e02379f0beed5ed531f9d912515668230.jpg", + "image_caption": [ + "ting method?", + "GT" + ], + "image_footnote": [], + "bbox": [ + 578, + 489, + 691, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/497b81f755dcd97666eaf92c21d335e3f8d279880f2ea1e3f3ae7dd2bb2c8a1a.jpg", + "image_caption": [ + "When typing, which part should your fingers apply pressure to?", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 707, + 491, + 771, + 589 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/481523d71a83bf7a67cf3b76e1f144cd04b017631c98ec86bd741d511284af78.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 794, + 493, + 859, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 638, + 619, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced the pioneering task of Language-guided Affordance Segmentation on 3D Objects (LASO), which establishes a vital connection between AI agents and Large Language Models (LLMs) via textual cues. Our contributions include the development of an extensive dataset consisting of 19,751 question-point pairs and the creation of PointRefer, a baseline model that lays the groundwork for the LASO task. We envisage that LASO will steer the trajectory of the 3D affordance field towards better integration with the advancements in LLMs.", + "bbox": [ + 496, + 655, + 893, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 821, + 679, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (92270114), the Agency for Science, Technology and Research (A\\*STAR) under MTC Programmatic Funds (Grant No.M23L7b0021), and NExT center.", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14258", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Do, etc. Affordancenet: An end-to-end deep learning approach for object affordance detection. In ICRA, 2018. 2", + "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 3", + "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19107-19117. IEEE, 2022. 3", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 5", + "[5] Dave Zhenyu Chen, Angel X. Chang, and Matthias Nießner. Scanrefer: 3d object localization in RGB-D scans using natural language. In ECCV, pages 202-221. Springer, 2020. 3", + "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, 2022. 3", + "[7] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In CVPR, pages 1778-1787. Computer Vision Foundation / IEEE, 2021. 2, 3, 6", + "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186. Association for Computational Linguistics, 2019. 7", + "[9] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J. Lim. Demo2vec: Reasoning object affordances from online videos. In CVPR. Computer Vision Foundation / IEEE Computer Society, 2018. 2", + "[10] Yiran Geng, Boshi An, Haoran Geng, Yuanpei Chen, Yaodong Yang, and Hao Dong. Rlafford: End-to-end affordance learning for robotic manipulation. In ICRA, pages 5880-5886. IEEE, 2023. 1", + "[11] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: decoding-enhanced bert with disentangled attention. In ICLR. OpenReview.net, 2021. 7", + "[12] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Tomas Jackson, Noah Brown, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In CoRL, pages 1769-1782. PMLR, 2022. 1", + "[13] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Kate-rina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 3", + "[14] Safoura Rezapour Lakani, Antonio Jose Rodríguez-Sánchez, and Justus H. Piater. Towards affordance detection for robot manipulation using affordance for parts and parts for affordance. pages 1155-1172, 2019. 1" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, pages 19652-19664, 2021. 6, 1, 2", + "[16] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In ACM MM, pages 4091-4099, 2021. 2", + "[17] Yicong Li, Xiang Wang, Junbin Xiao, and Tat-Seng Chua. Equivariant and invariant grounding for video question answering. In ACM MM, pages 4714-4722, 2022.", + "[18] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, pages 2928-2937, 2022.", + "[19] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. TPAMI, 2023.", + "[20] Yicong Li, Junbin Xiao, Chun Feng, Xiang Wang, and TatSeng Chua. Discovering spatio-temporal rationales for video question answering. In ICCV, pages 13869-13878, 2023.", + "[21] Yicong Li, Xun Yang, An Zhang, Chun Feng, Xiang Wang, and Tat-Seng Chua. Redundancy-aware transformer for video question answering. In ACM MM, pages 3172-3180, 2023. 2", + "[22] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: generalized referring expression segmentation. In CVPR, pages 23592-23601. IEEE, 2023. 6, 1, 2", + "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. 6, 7", + "[24] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Trans. Artif. Intell., 4(5):1186-1198, 2023. 2", + "[25] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In CVPR, pages 2242-2251. IEEE, 2022. 2", + "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16433-16442. IEEE, 2022. 3, 6, 1", + "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR. OpenReview.net, 2022. 7", + "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: situated question answering in 3d scenes. In ICLR. OpenReview.net, 2023. 3", + "[29] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers Neurorobotics, 14:26, 2020. 2", + "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas J. Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In CoRL, pages 1666-1677. PMLR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14259", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Bogdan Moldovan, Plinio Moreno, Martijn van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE International Conference on Robotics and Automation, pages 4373-4378, 2012. 1", + "[32] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In ICCV, pages 8687-8696. IEEE, 2019. 2", + "[33] OpenAI. Gpt-4 technical report, 2023. 2, 3", + "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 4, 7", + "[35] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In ICRA, pages 11523-11530. IEEE, 2023. 1", + "[36] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. CoRR, abs/2212.04088, 2022. 1", + "[37] NN Sriram, Tirth Maniar, Jayaganesh Kalyanasundaram, Vineet Gandhi, Brojeshwar Bhowmick, and K Madhava Krishna. Talk to the vehicle: Language conditioned autonomous navigation of self driving cars. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5284-5290. IEEE, 2019. 1", + "[38] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, pages 24261-24272, 2021. 4", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5", + "[40] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4964-4974. IEEE, 2022. 5", + "[41] Junbin Xiao, Angela Yao, Yicong Li, and Tat Seng Chua. Can i trust your answer? visually grounded video question answering. arXiv preprint arXiv:2309.01327, 2023. 2", + "[42] Amber Xie, Youngwoo Lee, Pieter Abbeel, and Stephen James. Language-conditioned path planning. In 7th Annual Conference on Robot Learning, 2023. 1", + "[43] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. CoRR, abs/2202.13519, 2022. 2", + "[44] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. CoRR, abs/2303.10437, 2023. 2, 3, 6, 7, 1", + "[45] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2908-2917. IEEE, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[46] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Comput. Appl., 32(18):14321-14333, 2020. 2", + "bbox": [ + 501, + 92, + 890, + 133 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14260", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_model.json b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bfcca58141f362424ebac1bde4803002728953aa --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "LASO: Language-guided Affordance Segmentation on 3D Object" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.18, + 0.812, + 0.201 + ], + "angle": 0, + "content": "Yicong Li\\(^{1}\\), Na Zhao\\(^{2*}\\), Junbin Xiao\\(^{1}\\), Chun Feng\\(^{3}\\), Xiang Wang\\(^{3*†}\\), Tat-seng Chua\\(^{1}\\)," + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.215, + 0.82, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)National University of Singapore, \\(^{2}\\)Singapore University of Technology and Design, \\(^{3}\\)University of Science and Technology of China," + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.753, + 0.27 + ], + "angle": 0, + "content": "liyicong@u.nus.edu,na_zhao@sutd.edu.sg,junbin@comp.nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.273, + 0.797, + 0.287 + ], + "angle": 0, + "content": "fengchun3364@mail.ustc.edu.cn,xiangwangl223@gmail.com,dcstsc@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.474, + 0.765 + ], + "angle": 0, + "content": "Segmenting affordance in 3D data is key for bridging perception and action in robots. Existing efforts mostly focus on the visual side and overlook the affordance knowledge from a semantic aspect. This oversight not only limits their generalization to unseen objects, but more importantly, hinders their synergy with large language models (LLMs) which are excellent task planners that can decompose an overarching command into agent-actionable instructions. With this regard, we propose a novel task, Language-guided Affordance Segmentation on 3D Object (LASO), which challenges a model to segment a 3D object's part relevant to a given affordance question. To facilitate the task, we contribute a dataset comprising 19,751 point-question pairs, covering 8434 object shapes and 870 expert-crafted questions. As a pioneer solution, we further propose PointRefer, which highlights an adaptive fusion module to identify target affordance regions at different scales. To ensure a text-aware segmentation, we adopt a set of affordance queries conditioned on linguistic cues to generate dynamic kernels. These kernels are further used to convolute with point features and generate a segmentation mask. Comprehensive experiments and analyses validate PointRefer's effectiveness. With these efforts, We hope that LASO can steer the direction of 3D affordance, guiding it towards enhanced integration with the evolving capabilities of LLMs. Code and data are available at https://github.com/y13800/LASO." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.794, + 0.21, + 0.81 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.821, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Affordance describes specific regions on an object that enable or facilitate certain human interactions. In this context," + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.321, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.496, + 0.893, + 0.553 + ], + "angle": 0, + "content": "Figure 1. Illustration of LASO, language-guided affordance segmentation on 3D object. Given an object point cloud and a question, the model is expected to segment the functional part delineated in question." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.893, + 0.656 + ], + "angle": 0, + "content": "the task of 3D affordance segmentation is raised to identify parts of 3D objects that imply a certain function. It has been deemed as a critical step in bridging perception and operation in the physical world for an embodied agent, thus has shown substantial impact on practical applications such as robotic manipulation [10, 14, 31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, the advancement of Large Language Models (LLMs) has enabled their deployment as task planners, breaking down an intricate instruction into a sequence of affordance-aware, robot-executable instructions [12, 35, 36]. This equips intelligent agents with the capability to undertake complex tasks. For example, the overarching command \"Pass me the mug in the microwave\" can be systematically deconstructed into a series of affordance-aware inquiries: Q1) \"How to get to the microwave?\", Q2) \"How to open the microwave\", Q3) \"How to grab the mug?\", and Q4) \"How to return to the user?\" Ideally, each query would correspond to a predefined subtask that effectively translates language instructions into actionable steps. While Q1 and Q4 are related to language-conditioned path planning [42] and navigation [37], Q2 and Q3, which involve the 'open microwave' and 'grab the mug' actions, ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.863, + 0.223, + 0.876 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.876, + 0.468, + 0.889 + ], + "angle": 0, + "content": "\\(\\dagger\\) Xiang Wang is also affiliated with Institute of Artificial Intelligence," + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.889, + 0.449, + 0.9 + ], + "angle": 0, + "content": "Institute of Dataspace, Hefei Comprehensive National Science Center." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.863, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14251" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.122 + ], + "angle": 0, + "content": "pose a gap — to the best of our knowledge, currently understanding affordance via language cues remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.468, + 0.275 + ], + "angle": 0, + "content": "Leaving language cues untouched, current 3D affordance segmentation focuses primarily on the visual aspect. It associates the specific geo-structure with either explicit affordance categories [7] or implicit affordance information from the 2D showcase [44]. Such visual-only settings make the segmentation models confined to visually predefined affordance types, thus impeding integration with language-based instructions from LLMs. As a result, fulfilling the aforementioned user's request becomes challenging, due to this inability to address Q2 and Q3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.47, + 0.609 + ], + "angle": 0, + "content": "Towards these limitations, we introduce a novel task: Language-guided Affordance Segmentation on 3D Object (LASO), where 3D objects are paired with natural language questions that probe specific affordance parts. As a cornerstone task, LASO can provide an agent with affordance knowledge to answer the aforementioned Q2 and Q3 raised by LLM. For instance, the first two rows in Fig. 1 demonstrate how LASO imparts knowledge that enables an agent to interpret the semantic instructions \"open the microwave\" and \"grasp the mug\" in a 3D context, which are sufficient to execute Q2 and Q3, respectively. In LASO, questions are crafted to reflect diverse real-world scenarios, translating various affordance types into a range of semantic contexts. By learning from different contexts, we expect the agent not to memorize predefined affordance types but rather to abstract generalizable affordance knowledge within a broader semantic context. Notably, the diverse scales and shapes of affordance parts pose a significant perceptual challenge for the model. For instance, in the third row of Fig. 1, the target backrest of a chair occupies a large portion and has a significantly different shape compared to the small knob on a door in the last row." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.468, + 0.883 + ], + "angle": 0, + "content": "To facilitate the study, we develop the 1st question-affordance dataset comprising 19,751 point cloud-question pairs. The questions, originally crafted by experts, are further diversified using GPT-4 [33] under principles of contextual enrichment, concise phrasing, and structural diversity. These principles enable a deep and varied exploration of object affordances. Based on the dataset, we have developed a strong baseline model named PointRefer that is capable of adapting to varying scales and shapes of object parts, conditioned on the question. It consists of two modules: 1) an Adaptive Fusion Module (AFM) that integrates question semantics to point features at different scales, which caters to varied target regions with better point feature discriminativeness; and 2) a Referred Point Decoder (RPD) that generates a segmentation mask via conditioned affordance queries, which, after refinement, forms dynamic convolution kernels to capture point-wise affordance for mask prediction. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "- We propose the task of Language-guided Affordance Seg" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Efforts on affordance learning. Part indicates whether the prediction specifies the object part (✓) or the whole object \\((\\times)\\)." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.121, + 0.876, + 0.224 + ], + "angle": 0, + "content": "
DomainStudiesVisionTextPart
2D[1, 46]Image××
[9, 25, 32]Image/Video×✓
[24, 29]Image✓×
3D[7, 30, 30, 43]Point Cloud×✓
[44]Point Cloud & Image×✓
OursPoint Cloud✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.892, + 0.278 + ], + "angle": 0, + "content": "mentation on 3D Objects (LASO), which helps AI agents derive affordance knowledge from textual cues, facilitating a crucial advancement for LLM integration." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.279, + 0.891, + 0.338 + ], + "angle": 0, + "content": "- We develop a dataset with 19,751 question-point affordance pairs which are meticulously curated via a collaborative effort of human experts and GPT-4, thereby establishing the 1st benchmark for evaluating LASO." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.892, + 0.399 + ], + "angle": 0, + "content": "- We build a strong baseline PointRefer, which employs textual-conditioned affordance queries to isolate afforded segments and devises an adaptive fusion module to enhance the discriminability of point feature." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.279, + 0.892, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.642, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.56 + ], + "angle": 0, + "content": "Affordance Learning. Originating from the image domain, initial efforts concentrate on detecting objects with affordances [1, 46]. Progressing from there, later studies [24, 29] incorporate linguistic descriptions to augment the process but still neglect the granularity of analysis, generally focusing on object-level affordances. Addressing this oversight, subsequent research [9, 25, 32] shift towards scrutinizing specific affordance parts, establishing a new norm for precision in the field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.835 + ], + "angle": 0, + "content": "With the rise of embodied AI, the scope of affordance learning expands into the 3D realm. 3D AffordanceNet [7] introduces the first benchmark dataset for learning affordance from object point cloud geometry. Building on this, Yang et al. [44] propose a setting for learning 3D affordance parts guided by image demonstrations. Nonetheless, these approaches predominantly rely on a visual-only approach, tightly coupling geometric features with affordance labels and overlooking the semantic dimension. This limitation hinders the integration with Large Language Models (LLMs) that could otherwise substantiate affordance segmentation to real-world deployment. Moreover, the common practice of predicting affordance types as an auxiliary task [7, 44] may dilute the primary challenge of understanding affordance knowledge. In contrast, our work sidesteps auxiliary predictions, advocating for direct learning from the linguistic context, which aligns more closely with the innate capabilities of LLMs and their semantic richness." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Text-Point Cloud Cross-Modal Learning. The integration of vision and natural language has recently sparked increased interest [16-21, 41]. Such a tendency naturally extends to 3D vision due to its pivotal role in embodied AI. To" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14252" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.447, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.241, + 0.444, + 0.255 + ], + "angle": 0, + "content": "Figure 2. Illustration of question augmentation principles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.473, + 0.536 + ], + "angle": 0, + "content": "advance the 3D-language tasks, datasets like ScanRefer [5] and ReferIt3D [2] have been instrumental in benchmarking the grounding of natural language to 3D objects. Complementing this, Azuma et al. [3] introduce ScanQA for 3D question-answering, while Ma et al. [28] propose SQA3D for situated reasoning within 3D environments. In terms of models, a variety of efforts [3, 13, 45] have emerged from these benchmarks, with notable examples like 3D-SPS [26] and BUTD-DETR [13] employing cross-attention mechanisms and linguistic cues for object discovery. Others, such as 3DVG [45], and ViL3DRel [6], focus on 3D grounding by incorporating spatial relation cues, showcasing the evolving landscape of 3D-VL models. Unlike the existing work that focuses on grounding and QA that perform scene-level reasoning, we tackle a dense prediction task at object part level that bridges reasoning to operation. To the best of our knowledge, we are the first to introduce the 3D affordance problem as a language-guided segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.166, + 0.558 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.47, + 0.637 + ], + "angle": 0, + "content": "To underpin our proposed task, we have meticulously compiled a question-point affordance dataset with 19,751 paired samples, featuring 8,434 distinct object shapes across 23 classes. Accompanying these are 870 expertly crafted questions, covering 17 affordance categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.258, + 0.66 + ], + "angle": 0, + "content": "3.1. Annotation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.67, + 0.47, + 0.806 + ], + "angle": 0, + "content": "Point Cloud. Our dataset leverages the point cloud and affordance annotations from the 3D-AffordanceNet [7]. Specifically, the affordance annotation in [7] is multi-class, with each point on an object potentially supporting multiple affordance types. In our case, each question is tailored to a specific affordance type, while each sample can associate with several questions. Consequently, the same object can exhibit diverse affordance segmentation, depending on the specific question posed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Question. To create the questions, we draw 58 object-affordance combinations from [7], and craft 15 unique questions for each of the 58 combinations, resulting in 870 tailored questions. Specifically, five manual questions per combination are initially composed to represent different scenarios. Subsequently, we employ GPT-4 [33] to diver" + }, + { + "type": "table_caption", + "bbox": [ + 0.618, + 0.09, + 0.774, + 0.102 + ], + "angle": 0, + "content": "Table 2. Dataset Statistics" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.107, + 0.89, + 0.167 + ], + "angle": 0, + "content": "
TaskTrainValTest
ShapeQstInstShapeQstInstShapeQstInst
Seen68836381612051658121510351742416
Unseen616045811558
" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.175, + 0.889, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.318, + 0.802, + 0.332 + ], + "angle": 0, + "content": "Figure 3. Landscape of our dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.576 + ], + "angle": 0, + "content": "sify the dataset with 10 additional questions for each combination, adhering to three key principles (as shown in Fig. 2): 1) Contextual Enrichment. Some questions were expanded to include additional context or details, fostering a more precise connection to the specified affordance of the target object. For instance, Q1 in Fig. 2 is reformed by a functional situation \"cutting paper\". 2) Concise Phrasing. In other cases, questions were distilled to their essence, making them succinct yet still meaningful. Q2 in Fig. 2 exemplifies this approach. 3) Structural Diversity. We employed a range of sentence structures, from questions to statements, to introduce natural linguistic variations and prevent model bias toward any specific phrasing or sentence length. A diversified example can be found in the Q3 of Fig. 2. All questions generated by GPT-4 undergo manual verification after this augmentation process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.682 + ], + "angle": 0, + "content": "It's worth mentioning that the affordance type was only a tool for crafting the questions and pinpointing the correct segmentation areas. We refrain from using explicit affordance labels during both training and testing. This deliberate choice ensures that the learning process is driven by the semantic content of the questions, rather than by direct affordance label association." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.698, + 0.712 + ], + "angle": 0, + "content": "3.2. Statistics and Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Dataset Settings. We show the statistical landscape of our dataset in Fig. 3. Specifically, it contains 17 affordance types and 23 object categories, composing 58 unique object-affordance combinations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the spirit of the work by [44], we offer two distinct dataset settings: Seen: The default configuration, where the training and testing phases share similar distributions of object classes and affordance types. Unseen: This configuration is purpose-built to test the model's ability to generalize affordance knowledge. Certain affordance types, when paired with specific object classes, are omitted from the training set but included in the test set. For instance, while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14253" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.167 + ], + "angle": 0, + "content": "the model may learn to grasp bags and mugs during training, it is expected to apply the concept of 'grasp' to earphones, which is an affordance-object pairing not encountered during training. For a comprehensive account of the unseen setting, please refer to Appendix Sec. 7.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.409 + ], + "angle": 0, + "content": "Training and Evaluation Protocols. Notably, since questions are created based on a combination of object class and affordance type, one object class can have many shape instances. Thus, the design of our dataset accounts for the multiplicity of shapes within each object class, leading to a non-bijective pairing between shape instances and corresponding questions. To ensure comprehensive learning, during training, each shape instance is matched with a randomly selected question that aligns with its affordance type for every iteration. This random pairing exposes the model to a variety of semantic contexts, bolstering its ability to generalize. For the validation and test sets, question pairings are fixed to minimize variability and ensure consistent inference. These questions are unique to the evaluation phase and are not revealed during the training process to maintain the integrity of the evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.411, + 0.168, + 0.426 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.498 + ], + "angle": 0, + "content": "LASO Task Definition. Given a question \\( Q_{raw} \\) and an object point cloud \\( \\mathbf{P}_{raw} \\in \\mathbb{R}^{N \\times 3} \\) with \\( N \\) points, the goal of LASO is to predict a binary mask of \\( \\mathbf{M} \\in \\mathbb{R}^N \\) that segments the functional part related to the question." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.82 + ], + "angle": 0, + "content": "Framework Overview. Due to the varying scale of target affordance region, LASO naturally challenges models adaptiveness at different scales. To address this, we devise a strong baseline, named PointRefer. As shown in Fig. 4, the object point cloud \\(\\mathbf{P}_{raw}\\) is first processed by a 3D backbone, which typically consists of the multi-stage encoding and decoding process. To adapt PointRefer to the point feature at multiple scales, we introduce an Adaptive Fusion Module (AFM) to the backbone's decoding process. It performs multi-scale cross-modal fusion by injecting text clues to point features at a different decoding stage, which progressively refine the point feature map in a top-down manner. Then, to predict the segmentation mask, we introduce a Referred Point Decoder (RPD) that leverages a set of learnable queries conditioned on the input question, termed affordance queries, as the input of the transformer decoder. These affordance queries are obligated to look at the referred points only and generate question-aware dynamic kernels. The final segmentation mask is obtained by convoluting these dynamic kernels with the AFM-enhanced point feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Feature Encoding. For the question, we prepend a \"[CLS]\" token to the question sequence before feeding them to a language model to capture the global language context. The encoded feature of the question is denoted by \\(\\mathbf{X} \\in \\mathbb{R}^{L \\times d}\\), where \\(L\\) is the number of tokens and \\(d\\) is the feature di" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "mension. For the point cloud, PointRefer can adopt any off-the-shelf 3D segmentation backbone.1 to transform the \\(\\mathbf{P}_{raw}\\) into the feature space. Note that a 3D backbone for segmentation typically consists of several encoding and decoding stages, and we take the point-wise feature map after the last decoding stage as the output of the backbone." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.192, + 0.727, + 0.208 + ], + "angle": 0, + "content": "4.1. Adaptive Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.35 + ], + "angle": 0, + "content": "AFM is designed to enhance the point-wise features by incorporating question information. As illustrated in Fig. 4, we integrate AFM into different decoder layers to enhance its adaptiveness to target regions of various scales and shapes. Given the conventional downsampling operators in 3D backbones (e.g. PointNet++ [34]), the number of points varies across different encoder/decoder layers. Our AFM is designed to seamlessly adapt to the varying number of points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.532 + ], + "angle": 0, + "content": "For simplification, we omit the notation for different decoder layers and use \\(\\mathbf{P} \\in \\mathbb{R}^{T \\times d}\\) to denote the point feature at a certain decoding layer, where \\(T\\) represents the number of points at that layer. As visualized in Fig. 4, AFM adopts a bottleneck architecture with three key steps: Grouping, Mixing, and Ungrouping. In the grouping step, the point features are grouped by text tokens, these grouped tokens are then thoroughly mixed via an MLP-Mixer [38]. Subsequently, the ungrouping step encapsulates the mixed features into individual point features, completing the AFM procedure by embedding valuable textual information into the point feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.607 + ], + "angle": 0, + "content": "To facilitate an efficient grouping process for our AFM, we employ a lightweight cross-attention module by omitting the transformation of both query and value. It takes the question feature \\(\\mathbf{X}\\) as query and the point feature \\(\\mathbf{P}\\) as key and value, then output grouped token \\(\\mathbf{G} \\in \\mathbb{R}^{L \\times d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.619, + 0.891, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {X}, \\mathbf {W} _ {1} \\mathbf {P}, \\mathbf {P}\\right) + \\mathbf {X}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.639, + 0.892, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.684, + 0.744, + 0.698 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_1\\) is a linear transformation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.892, + 0.759 + ], + "angle": 0, + "content": "After attaining the grouped feature, we update them using an MLP-Mixer [38], which employs two consecutive multilayer perceptrons (MLPs), producing fused feature \\(\\mathbf{F}\\) by combining with two consecutive MLPs." + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.769, + 0.891, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} ^ {\\prime} = \\mathbf {G} + \\operatorname {M L P} _ {1} \\left(\\mathbf {G} ^ {T}\\right) ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.789, + 0.891, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\mathbf {G} ^ {\\prime} + \\mathrm {M L P} _ {2} \\left(\\mathbf {G} ^ {\\prime}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.879 + ], + "angle": 0, + "content": "where, \\(\\mathrm{MLP}_1\\) and \\(\\mathrm{MLP}_2\\) are used to mix group-wise and channel-wise information, respectively, and \\(T[\\cdot]\\) denotes the transpose operation. Finally, we return the fused feature \\(\\mathbf{F}\\) to the point-space via the ungrouping process and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.836, + 0.9 + ], + "angle": 0, + "content": "1The impact of backbone choices can be observed in Tab. 5." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14254" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.635, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.356, + 0.438, + 0.368 + ], + "angle": 0, + "content": "(a) Overview of PointRefer." + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.094, + 0.895, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.351, + 0.848, + 0.365 + ], + "angle": 0, + "content": "(b) Adaptive Fusion Module." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.369, + 0.892, + 0.44 + ], + "angle": 0, + "content": "Figure 4. The overall framework of our proposed PointRefer (a), which contains an adaptive fusion modal (b) that can be injected in different stages of the backbone decoder. As a result, it takes the point-wise feature generated at the last decoder layer as the final output of the 3D backbone. Then, the Referred Point Decoder takes in a set of question-conditioned affordance queries and feeds it to a transformer decoder together with the fused point feature, where the output is then reformed to a set of dynamic kernels. Finally, these dynamic kernels are used as convolution kernels to filter out the segmentation mask from the fused point feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.507 + ], + "angle": 0, + "content": "acquire the fused point feature \\(\\mathbf{P}^m\\). We implement the ungrouping process as a similar lightweight attention module as the grouping process. However, in this case, it takes \\(\\mathbf{P}\\) as query and \\(\\mathbf{F}\\) as key and value:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.519, + 0.469, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {m} = \\operatorname {A t t e n t i o n} (\\mathbf {P}, \\mathbf {W} _ {2} \\mathbf {F}, \\mathbf {F}) + \\mathbf {P}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.607 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_2\\) is a linear transformation. Lastly, the final output \\(\\mathbf{P}^o\\) is formed by adding residual connections to \\(\\mathbf{P}^m\\). Notably, we take \\(\\mathbf{P}^o\\) after the last decoding stage as the output of the 3D backbone." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.295, + 0.632 + ], + "angle": 0, + "content": "4.2. Referred Point Decoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.469, + 0.806 + ], + "angle": 0, + "content": "Inspired by the success of learnable query-based methods in object detection [4] and segmentation [40], we introduce learnable affordance queries conditioned on the corresponding question of an object to decode the segmentation mask for the object. Specifically, we impose linguistic restrictions on all affordance queries, narrowing the model's focus to the relevant object parts and thereby facilitating convergence. These affordance queries, after being processed by a transformer decoder, are repurposed as dynamic kernels to deduce the segmentation mask from the point features, which are the final output of the 3D backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Question-Conditioned Affordance Query In the transformer decoder, it is well known that content and position embeddings are responsible for encoding instance-specific and spatial information, respectively. In PointRefer, we feed these two parts with the text feature and learnable query parameters, respectively, so that all the queries are" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.447, + 0.891, + 0.507 + ], + "angle": 0, + "content": "shaped by the nuances of the linguistic input. Specifically, the affordance queries \\(\\mathbf{A}\\) are formed by adding question embedding \\(\\mathbf{X}\\) with learnable embeddings. Thus, the resulting \\(\\mathbf{A} \\in \\mathbb{R}^{L \\times d}\\) are the same length of the question." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.511, + 0.892, + 0.602 + ], + "angle": 0, + "content": "Reasoning and Mask Prediction. After forming the affordance query, we feed it to a transformer decoder [39] together with the fused point feature \\(\\mathbf{P}_o\\) from the 3D backbone. In this manner, all the queries will use the question context as guidance and target to aggregate the feature of the referred points from the object, resulting in \\(\\mathbf{A}'\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.614, + 0.891, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {\\prime} = \\text {T r a n s f o r m e r - D e c o d e r} (\\mathbf {A}; \\mathbf {P} _ {o}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.891, + 0.716 + ], + "angle": 0, + "content": "Subsequently, we perform dynamic convolution by applying a two-layer MLP on top of the transfer decoder to produce \\(L\\) dynamic kernels \\(\\Omega = \\{\\omega_{i}\\}_{i = 1}^{L}\\), where each kernel \\(\\omega_{i}\\) is reshaped to a \\(1\\times 1\\) convolutional kernel with the channel number of \\(d\\)." + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.73, + 0.891, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\Omega = \\operatorname {M L P} \\left(\\mathbf {A} ^ {\\prime}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.759, + 0.891, + 0.835 + ], + "angle": 0, + "content": "Since the dynamic kernels have captured the question- information, we use them as convolution filters on the feature maps for mask decoding. Specifically, we use each dynamic kernel in \\(\\Omega\\) to convolute the point-wise feature map \\(\\mathbf{P}_o\\), which returns \\(L\\) point masks \\(\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^L\\)." + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.84, + 0.891, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {i} = \\left\\{\\mathbf {P} _ {o} * \\omega_ {i} \\right\\} _ {i = 1} ^ {L}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Then, we apply mean pooling over all masks in \\(\\mathbf{S}\\) followed by a sigmoid activation to acquire the final segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.091, + 0.182, + 0.105 + ], + "angle": 0, + "content": "mask \\(\\mathbf{M}\\in \\mathbb{R}^N\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.114, + 0.469, + 0.13 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\sigma (\\operatorname {M a x - P o o l} ((\\mathbf {S}))), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.138, + 0.345, + 0.152 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) denote the sigmoid function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.153, + 0.469, + 0.228 + ], + "angle": 0, + "content": "Intuitively, each dynamic kernel is expected to capture one aspect of the referred object part according to conditioned the question token. By ensembling all generated point masks, the final segmentation mask can integrate different response regions, thus making the result more robust." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.23, + 0.469, + 0.321 + ], + "angle": 0, + "content": "Objectives. Unlike [44], which relies on an auxiliary affordance label for prediction, our model seeks to forge a direct link between language context and object affordance. Thus, we solely employ Dice loss and Binary Cross-Entropy (BCE) loss to guide the segmentation mask prediction, bypassing the need for additional affordance labels." + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.324, + 0.469, + 0.339 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D i c e} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.333, + 0.201, + 0.35 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.358, + 0.469, + 0.464 + ], + "angle": 0, + "content": "Evaluation Metrics. For a thorough assessment, we benchmark our dataset against leading studies in 3D affordance learning [7, 44], employing four evaluation metrics to ensure robust analysis. These include Area Under the Curve (AUC), Mean Intersection Over Union (mIoU), Similarity (SIM), and Mean Absolute Error (MAE). Appendix Sec. 7.2 provides detailed explanations for the evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.469, + 0.557 + ], + "angle": 0, + "content": "Implementation Details. PointRefer employs PointNet++ as the default 3D backbone to align with the standards established in [7, 44]. For text encoding, we utilize a pretrained RoBERTa model [23] to process linguistic inputs. The feature dimension \\( d \\) is set to 512. During training, we use the Adam optimizer with a learning rate set to 1e-4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.56, + 0.469, + 0.801 + ], + "angle": 0, + "content": "Baseline Models. Since there is no prior work using paired question-point cloud data to segment object affordance. For a thorough comparison of our method, we adopt two 3D cross-modal works (3D-SPS [26] and IAGNet[44]) and two referred image segmentation works (ReferTrans [15] and RelA [22]) to LASO task. For the referred image segmentation models, we substitute their image backbones with a 3D counterpart, retaining their fusion mechanisms and mask decoding strategies. For the language-guided grounding model, 3D-SPS [26], we omit its bounding box prediction module while leveraging its inherent point selection scheme for our segmentation setting. The most related work is IAGNet [44], an affordance detection method that takes paired image-point cloud as input. To adapt IAGNet to our needs, we simply replace its image backbone with a language model, preserving the remainder of its architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.469, + 0.837 + ], + "angle": 0, + "content": "Next, we show the effectiveness of our PointRefer by answering the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.468, + 0.867 + ], + "angle": 0, + "content": "- Q1: How is PointRefer compared to other baselines on the proposed dataset?" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.868, + 0.422, + 0.883 + ], + "angle": 0, + "content": "- Q2: How effective are the proposed components?" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.884, + 0.462, + 0.898 + ], + "angle": 0, + "content": "- Q3: What learning pattern does the PointRefer capture?" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.838, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 3. Main Results. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. AUC and aIOU are shown in percentage." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.135, + 0.88, + 0.297 + ], + "angle": 0, + "content": "
MethodmIoU↑AUC↑SIM↑MAE↓
SteenReferTrans [15]13.779.80.4970.124
ReLA [22]15.278.90.5320.118
3D-SPS [26]11.476.20.4330.138
IAGNet [44]17.882.30.5610.109
PointRefer20.887.30.6290.093
UnseenReferTrans [15]10.269.10.4320.145
ReLA [22]10.769.70.4290.144
3D-SPS [26]7.968.80.4020.158
IAGNet [44]12.977.80.4430.129
PointRefer14.680.20.5070.119
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.31, + 0.671, + 0.326 + ], + "angle": 0, + "content": "5.1. Main Result (Q1)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.333, + 0.892, + 0.378 + ], + "angle": 0, + "content": "In Table 3, PointRefer demonstrates superior performance across all evaluation metrics compared to the baseline methods. Our observations are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.609 + ], + "angle": 0, + "content": "PointRefer vs. Other Models: IAGNet secures the second-best performance in both seen and unseen settings for all metrics. This is likely due to its specific design for object affordance segmentation, despite differences in input modality, making it a strong fit for our task with minimal adaptation. In contrast, 3D-SPS underperforms, potentially due to its progressive point selection mechanism. This approach involves a non-differentiable selection process, which may complicate the optimization within our task framework. Image-based methods, while showing decent performance, appear to be constrained by the modality gap. Originally trained on vast datasets of image-text pairs, where text is often a brief label, these methods now face fully formed, context-rich questions with a more limited training corpus." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Seen vs. Unseen Performance: A notable performance drop from seen to unseen settings is observed across all baselines. This discrepancy underscores the difficulty in learning affordance knowledge that generalizes well, corroborating the necessity of our task design. Current methods struggle to bridge this gap in knowledge transfer, suggesting an avenue for future research: utilizing the extensive commonsense knowledge embedded in LLMs could potentially mitigate this challenge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.695, + 0.773 + ], + "angle": 0, + "content": "5.2. In-Depth Study (Q2)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation Study. We show ablative results in Tab. 4. To verify that PointRefer prediction is based on the question, we conduct a blind test by training without questioning information, where the distinctive performance gap indicates that PointRefer has comprehended the question and can make a prediction based on it. Then, we study the effectiveness of PointRefer by removing both AFM and PRD (\"w/o AFM & PRD\"), which induces a severe performance" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14256" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.09, + 0.352, + 0.102 + ], + "angle": 0, + "content": "Table 4. Ablative Results." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.106, + 0.465, + 0.24 + ], + "angle": 0, + "content": "
VariantsmIoU↑AUC↑SIM↑MAE↓
Blind10.878.20.5060.125
w/o AFM & PRD17.782.10.5580.110
w/o AFM19.883.90.5920.104
w/o multi-scale20.487.10.6280.095
w/o PRD18.785.60.6190.098
w/o condition19.286.90.5900.100
w/o dynamic filter19.587.10.6230.096
PointRefer20.887.30.6290.093
" + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.246, + 0.434, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Figure 5. The bar chart shows the training sample distribution over different question lengths, and the line chart shows the performance of PointRefer grouped by question length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.47, + 0.713 + ], + "angle": 0, + "content": "decline on every metric, thus, validating the overall effectiveness of our design. Then, we also verify the functionality of AFM via the performance drop when it is removed (\"w/o AFM). As a comparison, a relatively slighter declination is witnessed when we only apply AFM at the single scale (\"w/o multi-scale\"), which indicates the benefit of multi-scale cross-modal fusion. Then, we conduct some detailed breakdown tests to study PRD, we notice that when the affordance queries are not conditioned on the question (\"w/o condition\") the performance drops drastically, indicating that our affordance queries help to capture semantic clues in 3D space. Finally, we validate the superiority of the dynamic filtering strategy, because a declination is also observed when the dynamic filter is erased (\"w/o dynamic filter\"), and this variant adopts IAGNet's [44] mask prediction strategy as an alternative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of Question Length. In Fig. 5, we show the performance of PointRefer on different question lengths. Specifically, The bar plot reveals the distribution of question length in the training set. Meanwhile, the line chart illustrates PointRefer's performance on different question lengths. Here, the performances are normalizing against the average number under that metric, e.g. IoU for each question length mIoU. Notably, PointRefer excels with questions that are more prevalent in the training set and with those that are lengthier—exceeding 14 words—owing to the richer contextual information they provide, which is instrumental in discerning the object's affordance." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.09, + 0.862, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.264, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 6. Illustration of how the percentage of omitted samples affect transferability of affordance knowledge (indicated by the difference between seen and unseen). For each affordance type, the bar shows a number of omitted samples from seen, i.e. \\(\\frac{\\# \\text{Seen} - \\# \\text{Unseen}}{\\# \\text{Seen}} (\\%)\\), and the lines show the testing IoU of models trained on Seen and Unseen. Difference (IoU): Seen - Unseen." + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.352, + 0.884, + 0.366 + ], + "angle": 0, + "content": "Table 5. Study of Backbones. PN:PointNet++; PM:PointMLP." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.369, + 0.879, + 0.533 + ], + "angle": 0, + "content": "
3DLMmIoU↑AUC↑SIM↑MAE↓
SeenPM [27]Bert [8]18.285.50.6180.099
Deberta [11]19.386.50.6210.096
Roberta [23]19.686.462.30.097
PN [34]Bert [8]20.186.40.6150.101
Deberta [11]20.487.00.6240.095
Roberta [23]20.887.30.6290.093
UnseenPM [27]Bert [8]11.576.30.430.128
Deberta [11]12.376.80.4350.133
Roberta [23]12.176.80.4340.136
PN [34]Bert [8]14.178.10.470.122
Deberta [11]14.578.90.500.114
Roberta [23]14.680.20.5070.119
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Seen vs. Unseen. Recall that to create the training set for the unseen setting, for an affordance type, we omit its combination with certain objects from seen. For example, the seen partition has \"grasp-mug\", and \"grasp-bag\" in its training, we create the unseen training set by removing the \"grasp-mug\" from the seen training set. In this case, the model is expected to learn the generalizable affordance knowledge of \"grasp\" and transfer it to an unseen object during testing. Note that seen and unseen setting shares the same validation and testing set. To investigate how PointRefer captures such transferability for each affordance type, The bar in Fig. 6 shows the percentage of samples that are omitted from seen under each affordance type. Then, the lines show the performance of the model trained with seen and unseen partitions for each affordance type, and their difference—the indicator of transferability of learned affordance knowledge. In general, we notice that the curve of seen and unseen follows a similar pattern, which indicates that PointRefer can capture a similar affordance knowledge even trained with far less object class. However, we also notice that the difference between seen and unseen enlarges as the omitted percentage grows, regardless of the absolution IoU. This is expected since the removed object can lead to under-diversified affordance representation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14257" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.063, + 0.862, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.893, + 0.457 + ], + "angle": 0, + "content": "Figure 7. Case-Study of PointRefer's segmentation. Each showcase comes with one question and four shapes, showing the generalization ability of the affordance knowledge. The segmented affordance part is highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.294, + 0.632 + ], + "angle": 0, + "content": "Choice of Backbone. To S backbone, we investigate the I with some alternative backbone that, for both seen and unseen as a 3D backbone generally pr than PointMLP, thus, we set 1 3D backbone. As for the lamance of Roberta prevails Ber setting. Whereas in the unseen and Roberta brings no signifca PointRefer uses Roberta as th" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.652, + 0.262, + 0.669 + ], + "angle": 0, + "content": "5.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.277, + 0.719 + ], + "angle": 0, + "content": "Case Study. As shown in Fig. 1, the capacity to accurately co- given the question. It is noteworthy" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.455, + 0.75 + ], + "angle": 0, + "content": "with small affordance components, such as the microwa door handle, our model still exhibits decent ability." + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.474, + 0.391, + 0.618 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.475, + 0.547, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.472, + 0.648, + 0.483 + ], + "angle": 0, + "content": "ting method?" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.491, + 0.692, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.592, + 0.662, + 0.601 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.467, + 0.865, + 0.488 + ], + "angle": 0, + "content": "When typing, which part should your fingers apply pressure to?" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.492, + 0.772, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.592, + 0.752, + 0.601 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.494, + 0.861, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.592, + 0.831, + 0.601 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.602, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Figure 8. Failure cases due to non-comprehended text (left) and over-segmentation (right)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.639, + 0.62, + 0.653 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.895, + 0.807 + ], + "angle": 0, + "content": "In this paper, we have introduced the pioneering task of Language-guided Affordance Segmentation on 3D Objects (LASO), which establishes a vital connection between AI agents and Large Language Models (LLMs) via textual cues. Our contributions include the development of an extensive dataset consisting of 19,751 question-point pairs and the creation of PointRefer, a baseline model that lays the groundwork for the LASO task. We envisage that LASO will steer the trajectory of the 3D affordance field towards better integration with the advancements in LLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.68, + 0.84 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (92270114), the Agency for Science, Technology and Research (A\\*STAR) under MTC Programmatic Funds (Grant No.M23L7b0021), and NExT center." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14258" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Do, etc. Affordancenet: An end-to-end deep learning approach for object affordance detection. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.256 + ], + "angle": 0, + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19107-19117. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.468, + 0.358 + ], + "angle": 0, + "content": "[5] Dave Zhenyu Chen, Angel X. Chang, and Matthias Nießner. Scanrefer: 3d object localization in RGB-D scans using natural language. In ECCV, pages 202-221. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.361, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[7] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In CVPR, pages 1778-1787. Computer Vision Foundation / IEEE, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186. Association for Computational Linguistics, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[9] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J. Lim. Demo2vec: Reasoning object affordances from online videos. In CVPR. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[10] Yiran Geng, Boshi An, Haoran Geng, Yuanpei Chen, Yaodong Yang, and Hao Dong. Rlafford: End-to-end affordance learning for robotic manipulation. In ICRA, pages 5880-5886. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[11] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: decoding-enhanced bert with disentangled attention. In ICLR. OpenReview.net, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[12] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Tomas Jackson, Noah Brown, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In CoRL, pages 1769-1782. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[13] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Kate-rina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Safoura Rezapour Lakani, Antonio Jose Rodríguez-Sánchez, and Justus H. Piater. Towards affordance detection for robot manipulation using affordance for parts and parts for affordance. pages 1155-1172, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, pages 19652-19664, 2021. 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[16] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In ACM MM, pages 4091-4099, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[17] Yicong Li, Xiang Wang, Junbin Xiao, and Tat-Seng Chua. Equivariant and invariant grounding for video question answering. In ACM MM, pages 4714-4722, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[18] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, pages 2928-2937, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[19] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. TPAMI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[20] Yicong Li, Junbin Xiao, Chun Feng, Xiang Wang, and TatSeng Chua. Discovering spatio-temporal rationales for video question answering. In ICCV, pages 13869-13878, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[21] Yicong Li, Xun Yang, An Zhang, Chun Feng, Xiang Wang, and Tat-Seng Chua. Redundancy-aware transformer for video question answering. In ACM MM, pages 3172-3180, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.439 + ], + "angle": 0, + "content": "[22] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: generalized referring expression segmentation. In CVPR, pages 23592-23601. IEEE, 2023. 6, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.512, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[24] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Trans. Artif. Intell., 4(5):1186-1198, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[25] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In CVPR, pages 2242-2251. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16433-16442. IEEE, 2022. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR. OpenReview.net, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: situated question answering in 3d scenes. In ICLR. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[29] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers Neurorobotics, 14:26, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas J. Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In CoRL, pages 1666-1677. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14259" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[31] Bogdan Moldovan, Plinio Moreno, Martijn van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE International Conference on Robotics and Automation, pages 4373-4378, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[32] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In ICCV, pages 8687-8696. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.369, + 0.221 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4 technical report, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.279, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[35] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In ICRA, pages 11523-11530. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[36] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. CoRR, abs/2212.04088, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[37] NN Sriram, Tirth Maniar, Jayaganesh Kalyanasundaram, Vineet Gandhi, Brojeshwar Bhowmick, and K Madhava Krishna. Talk to the vehicle: Language conditioned autonomous navigation of self driving cars. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5284-5290. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.491, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[38] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, pages 24261-24272, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[40] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4964-4974. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.675, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[41] Junbin Xiao, Angela Yao, Yicong Li, and Tat Seng Chua. Can i trust your answer? visually grounded video question answering. arXiv preprint arXiv:2309.01327, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[42] Amber Xie, Youngwoo Lee, Pieter Abbeel, and Stephen James. Language-conditioned path planning. In 7th Annual Conference on Robot Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[43] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. CoRR, abs/2202.13519, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.804, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[44] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. CoRR, abs/2303.10437, 2023. 2, 3, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[45] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2908-2917. IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[46] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Comput. Appl., 32(18):14321-14333, 2020. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14260" + } + ] +] \ No newline at end of file diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_origin.pdf b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..537b8002c1e1b1f98b2b866f5cdb6d96ce7e3b4f --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/4a958b96-aed8-4e26-96c9-3b681d296992_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fca8b380f1b41284e0b2def462c7b2a0e9c8f7ee9edd735f942643042e96e792 +size 3862925 diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/full.md b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8b5b7b22018869052a8e24b7a7944fa18babb86c --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/full.md @@ -0,0 +1,323 @@ +# LASO: Language-guided Affordance Segmentation on 3D Object + +Yicong Li $^{1}$ , Na Zhao $^{2*}$ , Junbin Xiao $^{1}$ , Chun Feng $^{3}$ , Xiang Wang $^{3*†}$ , Tat-seng Chua $^{1}$ , + +$^{1}$ National University of Singapore, $^{2}$ Singapore University of Technology and Design, $^{3}$ University of Science and Technology of China, + +liyicong@u.nus.edu,na_zhao@sutd.edu.sg,junbin@comp.nus.edu.sg + +fengchun3364@mail.ustc.edu.cn,xiangwangl223@gmail.com,dcstsc@nus.edu.sg + +# Abstract + +Segmenting affordance in 3D data is key for bridging perception and action in robots. Existing efforts mostly focus on the visual side and overlook the affordance knowledge from a semantic aspect. This oversight not only limits their generalization to unseen objects, but more importantly, hinders their synergy with large language models (LLMs) which are excellent task planners that can decompose an overarching command into agent-actionable instructions. With this regard, we propose a novel task, Language-guided Affordance Segmentation on 3D Object (LASO), which challenges a model to segment a 3D object's part relevant to a given affordance question. To facilitate the task, we contribute a dataset comprising 19,751 point-question pairs, covering 8434 object shapes and 870 expert-crafted questions. As a pioneer solution, we further propose PointRefer, which highlights an adaptive fusion module to identify target affordance regions at different scales. To ensure a text-aware segmentation, we adopt a set of affordance queries conditioned on linguistic cues to generate dynamic kernels. These kernels are further used to convolute with point features and generate a segmentation mask. Comprehensive experiments and analyses validate PointRefer's effectiveness. With these efforts, We hope that LASO can steer the direction of 3D affordance, guiding it towards enhanced integration with the evolving capabilities of LLMs. Code and data are available at https://github.com/y13800/LASO. + +# 1. Introduction + +Affordance describes specific regions on an object that enable or facilitate certain human interactions. In this context, + +![](images/304263c7468019abd031dd459f80d845303ab370be9420da4a6eb9a7ffaed1e1.jpg) +Figure 1. Illustration of LASO, language-guided affordance segmentation on 3D object. Given an object point cloud and a question, the model is expected to segment the functional part delineated in question. + +the task of 3D affordance segmentation is raised to identify parts of 3D objects that imply a certain function. It has been deemed as a critical step in bridging perception and operation in the physical world for an embodied agent, thus has shown substantial impact on practical applications such as robotic manipulation [10, 14, 31]. + +Recently, the advancement of Large Language Models (LLMs) has enabled their deployment as task planners, breaking down an intricate instruction into a sequence of affordance-aware, robot-executable instructions [12, 35, 36]. This equips intelligent agents with the capability to undertake complex tasks. For example, the overarching command "Pass me the mug in the microwave" can be systematically deconstructed into a series of affordance-aware inquiries: Q1) "How to get to the microwave?", Q2) "How to open the microwave", Q3) "How to grab the mug?", and Q4) "How to return to the user?" Ideally, each query would correspond to a predefined subtask that effectively translates language instructions into actionable steps. While Q1 and Q4 are related to language-conditioned path planning [42] and navigation [37], Q2 and Q3, which involve the 'open microwave' and 'grab the mug' actions, ex + +pose a gap — to the best of our knowledge, currently understanding affordance via language cues remains unexplored. + +Leaving language cues untouched, current 3D affordance segmentation focuses primarily on the visual aspect. It associates the specific geo-structure with either explicit affordance categories [7] or implicit affordance information from the 2D showcase [44]. Such visual-only settings make the segmentation models confined to visually predefined affordance types, thus impeding integration with language-based instructions from LLMs. As a result, fulfilling the aforementioned user's request becomes challenging, due to this inability to address Q2 and Q3. + +Towards these limitations, we introduce a novel task: Language-guided Affordance Segmentation on 3D Object (LASO), where 3D objects are paired with natural language questions that probe specific affordance parts. As a cornerstone task, LASO can provide an agent with affordance knowledge to answer the aforementioned Q2 and Q3 raised by LLM. For instance, the first two rows in Fig. 1 demonstrate how LASO imparts knowledge that enables an agent to interpret the semantic instructions "open the microwave" and "grasp the mug" in a 3D context, which are sufficient to execute Q2 and Q3, respectively. In LASO, questions are crafted to reflect diverse real-world scenarios, translating various affordance types into a range of semantic contexts. By learning from different contexts, we expect the agent not to memorize predefined affordance types but rather to abstract generalizable affordance knowledge within a broader semantic context. Notably, the diverse scales and shapes of affordance parts pose a significant perceptual challenge for the model. For instance, in the third row of Fig. 1, the target backrest of a chair occupies a large portion and has a significantly different shape compared to the small knob on a door in the last row. + +To facilitate the study, we develop the 1st question-affordance dataset comprising 19,751 point cloud-question pairs. The questions, originally crafted by experts, are further diversified using GPT-4 [33] under principles of contextual enrichment, concise phrasing, and structural diversity. These principles enable a deep and varied exploration of object affordances. Based on the dataset, we have developed a strong baseline model named PointRefer that is capable of adapting to varying scales and shapes of object parts, conditioned on the question. It consists of two modules: 1) an Adaptive Fusion Module (AFM) that integrates question semantics to point features at different scales, which caters to varied target regions with better point feature discriminativeness; and 2) a Referred Point Decoder (RPD) that generates a segmentation mask via conditioned affordance queries, which, after refinement, forms dynamic convolution kernels to capture point-wise affordance for mask prediction. In summary, our contributions are as follows: + +- We propose the task of Language-guided Affordance Seg + +Table 1. Efforts on affordance learning. Part indicates whether the prediction specifies the object part (✓) or the whole object $(\times)$ . + +
DomainStudiesVisionTextPart
2D[1, 46]Image××
[9, 25, 32]Image/Video×✓
[24, 29]Image✓×
3D[7, 30, 30, 43]Point Cloud×✓
[44]Point Cloud & Image×✓
OursPoint Cloud✓✓
+ +mentation on 3D Objects (LASO), which helps AI agents derive affordance knowledge from textual cues, facilitating a crucial advancement for LLM integration. + +- We develop a dataset with 19,751 question-point affordance pairs which are meticulously curated via a collaborative effort of human experts and GPT-4, thereby establishing the 1st benchmark for evaluating LASO. +- We build a strong baseline PointRefer, which employs textual-conditioned affordance queries to isolate afforded segments and devises an adaptive fusion module to enhance the discriminability of point feature. + +# 2. Related Work + +Affordance Learning. Originating from the image domain, initial efforts concentrate on detecting objects with affordances [1, 46]. Progressing from there, later studies [24, 29] incorporate linguistic descriptions to augment the process but still neglect the granularity of analysis, generally focusing on object-level affordances. Addressing this oversight, subsequent research [9, 25, 32] shift towards scrutinizing specific affordance parts, establishing a new norm for precision in the field. + +With the rise of embodied AI, the scope of affordance learning expands into the 3D realm. 3D AffordanceNet [7] introduces the first benchmark dataset for learning affordance from object point cloud geometry. Building on this, Yang et al. [44] propose a setting for learning 3D affordance parts guided by image demonstrations. Nonetheless, these approaches predominantly rely on a visual-only approach, tightly coupling geometric features with affordance labels and overlooking the semantic dimension. This limitation hinders the integration with Large Language Models (LLMs) that could otherwise substantiate affordance segmentation to real-world deployment. Moreover, the common practice of predicting affordance types as an auxiliary task [7, 44] may dilute the primary challenge of understanding affordance knowledge. In contrast, our work sidesteps auxiliary predictions, advocating for direct learning from the linguistic context, which aligns more closely with the innate capabilities of LLMs and their semantic richness. + +Text-Point Cloud Cross-Modal Learning. The integration of vision and natural language has recently sparked increased interest [16-21, 41]. Such a tendency naturally extends to 3D vision due to its pivotal role in embodied AI. To + +![](images/f02f4f367fd6504dc0599f11f0b7da47dce2cc5d8670c945bc01f637a759f629.jpg) +Figure 2. Illustration of question augmentation principles. + +advance the 3D-language tasks, datasets like ScanRefer [5] and ReferIt3D [2] have been instrumental in benchmarking the grounding of natural language to 3D objects. Complementing this, Azuma et al. [3] introduce ScanQA for 3D question-answering, while Ma et al. [28] propose SQA3D for situated reasoning within 3D environments. In terms of models, a variety of efforts [3, 13, 45] have emerged from these benchmarks, with notable examples like 3D-SPS [26] and BUTD-DETR [13] employing cross-attention mechanisms and linguistic cues for object discovery. Others, such as 3DVG [45], and ViL3DRel [6], focus on 3D grounding by incorporating spatial relation cues, showcasing the evolving landscape of 3D-VL models. Unlike the existing work that focuses on grounding and QA that perform scene-level reasoning, we tackle a dense prediction task at object part level that bridges reasoning to operation. To the best of our knowledge, we are the first to introduce the 3D affordance problem as a language-guided segmentation task. + +# 3. Dataset + +To underpin our proposed task, we have meticulously compiled a question-point affordance dataset with 19,751 paired samples, featuring 8,434 distinct object shapes across 23 classes. Accompanying these are 870 expertly crafted questions, covering 17 affordance categories. + +# 3.1. Annotation Details + +Point Cloud. Our dataset leverages the point cloud and affordance annotations from the 3D-AffordanceNet [7]. Specifically, the affordance annotation in [7] is multi-class, with each point on an object potentially supporting multiple affordance types. In our case, each question is tailored to a specific affordance type, while each sample can associate with several questions. Consequently, the same object can exhibit diverse affordance segmentation, depending on the specific question posed. + +Question. To create the questions, we draw 58 object-affordance combinations from [7], and craft 15 unique questions for each of the 58 combinations, resulting in 870 tailored questions. Specifically, five manual questions per combination are initially composed to represent different scenarios. Subsequently, we employ GPT-4 [33] to diver + +Table 2. Dataset Statistics + +
TaskTrainValTest
ShapeQstInstShapeQstInstShapeQstInst
Seen68836381612051658121510351742416
Unseen616045811558
+ +![](images/058b1cac9cc8965d2e5d05a665f8f8ce3c0ea0c2d9db887094e1c4c1334812c5.jpg) +Figure 3. Landscape of our dataset. + +sify the dataset with 10 additional questions for each combination, adhering to three key principles (as shown in Fig. 2): 1) Contextual Enrichment. Some questions were expanded to include additional context or details, fostering a more precise connection to the specified affordance of the target object. For instance, Q1 in Fig. 2 is reformed by a functional situation "cutting paper". 2) Concise Phrasing. In other cases, questions were distilled to their essence, making them succinct yet still meaningful. Q2 in Fig. 2 exemplifies this approach. 3) Structural Diversity. We employed a range of sentence structures, from questions to statements, to introduce natural linguistic variations and prevent model bias toward any specific phrasing or sentence length. A diversified example can be found in the Q3 of Fig. 2. All questions generated by GPT-4 undergo manual verification after this augmentation process. + +It's worth mentioning that the affordance type was only a tool for crafting the questions and pinpointing the correct segmentation areas. We refrain from using explicit affordance labels during both training and testing. This deliberate choice ensures that the learning process is driven by the semantic content of the questions, rather than by direct affordance label association. + +# 3.2. Statistics and Setting + +Dataset Settings. We show the statistical landscape of our dataset in Fig. 3. Specifically, it contains 17 affordance types and 23 object categories, composing 58 unique object-affordance combinations. + +In the spirit of the work by [44], we offer two distinct dataset settings: Seen: The default configuration, where the training and testing phases share similar distributions of object classes and affordance types. Unseen: This configuration is purpose-built to test the model's ability to generalize affordance knowledge. Certain affordance types, when paired with specific object classes, are omitted from the training set but included in the test set. For instance, while + +the model may learn to grasp bags and mugs during training, it is expected to apply the concept of 'grasp' to earphones, which is an affordance-object pairing not encountered during training. For a comprehensive account of the unseen setting, please refer to Appendix Sec. 7.1. + +Training and Evaluation Protocols. Notably, since questions are created based on a combination of object class and affordance type, one object class can have many shape instances. Thus, the design of our dataset accounts for the multiplicity of shapes within each object class, leading to a non-bijective pairing between shape instances and corresponding questions. To ensure comprehensive learning, during training, each shape instance is matched with a randomly selected question that aligns with its affordance type for every iteration. This random pairing exposes the model to a variety of semantic contexts, bolstering its ability to generalize. For the validation and test sets, question pairings are fixed to minimize variability and ensure consistent inference. These questions are unique to the evaluation phase and are not revealed during the training process to maintain the integrity of the evaluation. + +# 4. Method + +LASO Task Definition. Given a question $Q_{raw}$ and an object point cloud $\mathbf{P}_{raw} \in \mathbb{R}^{N \times 3}$ with $N$ points, the goal of LASO is to predict a binary mask of $\mathbf{M} \in \mathbb{R}^N$ that segments the functional part related to the question. + +Framework Overview. Due to the varying scale of target affordance region, LASO naturally challenges models adaptiveness at different scales. To address this, we devise a strong baseline, named PointRefer. As shown in Fig. 4, the object point cloud $\mathbf{P}_{raw}$ is first processed by a 3D backbone, which typically consists of the multi-stage encoding and decoding process. To adapt PointRefer to the point feature at multiple scales, we introduce an Adaptive Fusion Module (AFM) to the backbone's decoding process. It performs multi-scale cross-modal fusion by injecting text clues to point features at a different decoding stage, which progressively refine the point feature map in a top-down manner. Then, to predict the segmentation mask, we introduce a Referred Point Decoder (RPD) that leverages a set of learnable queries conditioned on the input question, termed affordance queries, as the input of the transformer decoder. These affordance queries are obligated to look at the referred points only and generate question-aware dynamic kernels. The final segmentation mask is obtained by convoluting these dynamic kernels with the AFM-enhanced point feature. + +Feature Encoding. For the question, we prepend a "[CLS]" token to the question sequence before feeding them to a language model to capture the global language context. The encoded feature of the question is denoted by $\mathbf{X} \in \mathbb{R}^{L \times d}$ , where $L$ is the number of tokens and $d$ is the feature di + +mension. For the point cloud, PointRefer can adopt any off-the-shelf 3D segmentation backbone.1 to transform the $\mathbf{P}_{raw}$ into the feature space. Note that a 3D backbone for segmentation typically consists of several encoding and decoding stages, and we take the point-wise feature map after the last decoding stage as the output of the backbone. + +# 4.1. Adaptive Fusion Module + +AFM is designed to enhance the point-wise features by incorporating question information. As illustrated in Fig. 4, we integrate AFM into different decoder layers to enhance its adaptiveness to target regions of various scales and shapes. Given the conventional downsampling operators in 3D backbones (e.g. PointNet++ [34]), the number of points varies across different encoder/decoder layers. Our AFM is designed to seamlessly adapt to the varying number of points. + +For simplification, we omit the notation for different decoder layers and use $\mathbf{P} \in \mathbb{R}^{T \times d}$ to denote the point feature at a certain decoding layer, where $T$ represents the number of points at that layer. As visualized in Fig. 4, AFM adopts a bottleneck architecture with three key steps: Grouping, Mixing, and Ungrouping. In the grouping step, the point features are grouped by text tokens, these grouped tokens are then thoroughly mixed via an MLP-Mixer [38]. Subsequently, the ungrouping step encapsulates the mixed features into individual point features, completing the AFM procedure by embedding valuable textual information into the point feature. + +To facilitate an efficient grouping process for our AFM, we employ a lightweight cross-attention module by omitting the transformation of both query and value. It takes the question feature $\mathbf{X}$ as query and the point feature $\mathbf{P}$ as key and value, then output grouped token $\mathbf{G} \in \mathbb{R}^{L \times d}$ : + +$$ +\mathbf {G} = \operatorname {A t t e n t i o n} \left(\mathbf {X}, \mathbf {W} _ {1} \mathbf {P}, \mathbf {P}\right) + \mathbf {X}, \tag {1} +$$ + +$$ +\operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\sqrt {d}}\right) \mathbf {V}, \tag {2} +$$ + +where $\mathbf{W}_1$ is a linear transformation. + +After attaining the grouped feature, we update them using an MLP-Mixer [38], which employs two consecutive multilayer perceptrons (MLPs), producing fused feature $\mathbf{F}$ by combining with two consecutive MLPs. + +$$ +\mathbf {G} ^ {\prime} = \mathbf {G} + \operatorname {M L P} _ {1} \left(\mathbf {G} ^ {T}\right) ^ {T}, \tag {3} +$$ + +$$ +\mathbf {F} = \mathbf {G} ^ {\prime} + \mathrm {M L P} _ {2} \left(\mathbf {G} ^ {\prime}\right), \tag {4} +$$ + +where, $\mathrm{MLP}_1$ and $\mathrm{MLP}_2$ are used to mix group-wise and channel-wise information, respectively, and $T[\cdot]$ denotes the transpose operation. Finally, we return the fused feature $\mathbf{F}$ to the point-space via the ungrouping process and + +![](images/e9b18ca47acf1010aba4b1a6f550693ca4313b08db5a15d713f522946cff3c15.jpg) +(a) Overview of PointRefer. + +![](images/d72ddcef34d4d0b43226e356b67b7c1342c77602abfa91c7a9b82df45cf402de.jpg) +(b) Adaptive Fusion Module. +Figure 4. The overall framework of our proposed PointRefer (a), which contains an adaptive fusion modal (b) that can be injected in different stages of the backbone decoder. As a result, it takes the point-wise feature generated at the last decoder layer as the final output of the 3D backbone. Then, the Referred Point Decoder takes in a set of question-conditioned affordance queries and feeds it to a transformer decoder together with the fused point feature, where the output is then reformed to a set of dynamic kernels. Finally, these dynamic kernels are used as convolution kernels to filter out the segmentation mask from the fused point feature. + +acquire the fused point feature $\mathbf{P}^m$ . We implement the ungrouping process as a similar lightweight attention module as the grouping process. However, in this case, it takes $\mathbf{P}$ as query and $\mathbf{F}$ as key and value: + +$$ +\mathbf {P} ^ {m} = \operatorname {A t t e n t i o n} (\mathbf {P}, \mathbf {W} _ {2} \mathbf {F}, \mathbf {F}) + \mathbf {P}, \tag {5} +$$ + +where $\mathbf{W}_2$ is a linear transformation. Lastly, the final output $\mathbf{P}^o$ is formed by adding residual connections to $\mathbf{P}^m$ . Notably, we take $\mathbf{P}^o$ after the last decoding stage as the output of the 3D backbone. + +# 4.2. Referred Point Decoder + +Inspired by the success of learnable query-based methods in object detection [4] and segmentation [40], we introduce learnable affordance queries conditioned on the corresponding question of an object to decode the segmentation mask for the object. Specifically, we impose linguistic restrictions on all affordance queries, narrowing the model's focus to the relevant object parts and thereby facilitating convergence. These affordance queries, after being processed by a transformer decoder, are repurposed as dynamic kernels to deduce the segmentation mask from the point features, which are the final output of the 3D backbone. + +Question-Conditioned Affordance Query In the transformer decoder, it is well known that content and position embeddings are responsible for encoding instance-specific and spatial information, respectively. In PointRefer, we feed these two parts with the text feature and learnable query parameters, respectively, so that all the queries are + +shaped by the nuances of the linguistic input. Specifically, the affordance queries $\mathbf{A}$ are formed by adding question embedding $\mathbf{X}$ with learnable embeddings. Thus, the resulting $\mathbf{A} \in \mathbb{R}^{L \times d}$ are the same length of the question. + +Reasoning and Mask Prediction. After forming the affordance query, we feed it to a transformer decoder [39] together with the fused point feature $\mathbf{P}_o$ from the 3D backbone. In this manner, all the queries will use the question context as guidance and target to aggregate the feature of the referred points from the object, resulting in $\mathbf{A}'$ : + +$$ +\mathbf {A} ^ {\prime} = \text {T r a n s f o r m e r - D e c o d e r} (\mathbf {A}; \mathbf {P} _ {o}). \tag {6} +$$ + +Subsequently, we perform dynamic convolution by applying a two-layer MLP on top of the transfer decoder to produce $L$ dynamic kernels $\Omega = \{\omega_{i}\}_{i = 1}^{L}$ , where each kernel $\omega_{i}$ is reshaped to a $1\times 1$ convolutional kernel with the channel number of $d$ . + +$$ +\Omega = \operatorname {M L P} \left(\mathbf {A} ^ {\prime}\right). \tag {7} +$$ + +Since the dynamic kernels have captured the question- information, we use them as convolution filters on the feature maps for mask decoding. Specifically, we use each dynamic kernel in $\Omega$ to convolute the point-wise feature map $\mathbf{P}_o$ , which returns $L$ point masks $\mathbf{S} = \{\mathbf{s}_i\}_{i=1}^L$ . + +$$ +\mathbf {S} _ {i} = \left\{\mathbf {P} _ {o} * \omega_ {i} \right\} _ {i = 1} ^ {L}. \tag {8} +$$ + +Then, we apply mean pooling over all masks in $\mathbf{S}$ followed by a sigmoid activation to acquire the final segmentation + +mask $\mathbf{M}\in \mathbb{R}^N$ .. + +$$ +\mathbf {M} = \sigma (\operatorname {M a x - P o o l} ((\mathbf {S}))), \tag {9} +$$ + +where $\sigma (\cdot)$ denote the sigmoid function. + +Intuitively, each dynamic kernel is expected to capture one aspect of the referred object part according to conditioned the question token. By ensembling all generated point masks, the final segmentation mask can integrate different response regions, thus making the result more robust. + +Objectives. Unlike [44], which relies on an auxiliary affordance label for prediction, our model seeks to forge a direct link between language context and object affordance. Thus, we solely employ Dice loss and Binary Cross-Entropy (BCE) loss to guide the segmentation mask prediction, bypassing the need for additional affordance labels. + +$$ +\mathcal {L} = \mathcal {L} _ {B C E} + \mathcal {L} _ {D i c e} \tag {10} +$$ + +# 5. Experiment + +Evaluation Metrics. For a thorough assessment, we benchmark our dataset against leading studies in 3D affordance learning [7, 44], employing four evaluation metrics to ensure robust analysis. These include Area Under the Curve (AUC), Mean Intersection Over Union (mIoU), Similarity (SIM), and Mean Absolute Error (MAE). Appendix Sec. 7.2 provides detailed explanations for the evaluation protocols. + +Implementation Details. PointRefer employs PointNet++ as the default 3D backbone to align with the standards established in [7, 44]. For text encoding, we utilize a pretrained RoBERTa model [23] to process linguistic inputs. The feature dimension $d$ is set to 512. During training, we use the Adam optimizer with a learning rate set to 1e-4. + +Baseline Models. Since there is no prior work using paired question-point cloud data to segment object affordance. For a thorough comparison of our method, we adopt two 3D cross-modal works (3D-SPS [26] and IAGNet[44]) and two referred image segmentation works (ReferTrans [15] and RelA [22]) to LASO task. For the referred image segmentation models, we substitute their image backbones with a 3D counterpart, retaining their fusion mechanisms and mask decoding strategies. For the language-guided grounding model, 3D-SPS [26], we omit its bounding box prediction module while leveraging its inherent point selection scheme for our segmentation setting. The most related work is IAGNet [44], an affordance detection method that takes paired image-point cloud as input. To adapt IAGNet to our needs, we simply replace its image backbone with a language model, preserving the remainder of its architecture. + +Next, we show the effectiveness of our PointRefer by answering the following questions: + +- Q1: How is PointRefer compared to other baselines on the proposed dataset? +- Q2: How effective are the proposed components? +- Q3: What learning pattern does the PointRefer capture? + +Table 3. Main Results. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. AUC and aIOU are shown in percentage. + +
MethodmIoU↑AUC↑SIM↑MAE↓
SteenReferTrans [15]13.779.80.4970.124
ReLA [22]15.278.90.5320.118
3D-SPS [26]11.476.20.4330.138
IAGNet [44]17.882.30.5610.109
PointRefer20.887.30.6290.093
UnseenReferTrans [15]10.269.10.4320.145
ReLA [22]10.769.70.4290.144
3D-SPS [26]7.968.80.4020.158
IAGNet [44]12.977.80.4430.129
PointRefer14.680.20.5070.119
+ +# 5.1. Main Result (Q1) + +In Table 3, PointRefer demonstrates superior performance across all evaluation metrics compared to the baseline methods. Our observations are as follows: + +PointRefer vs. Other Models: IAGNet secures the second-best performance in both seen and unseen settings for all metrics. This is likely due to its specific design for object affordance segmentation, despite differences in input modality, making it a strong fit for our task with minimal adaptation. In contrast, 3D-SPS underperforms, potentially due to its progressive point selection mechanism. This approach involves a non-differentiable selection process, which may complicate the optimization within our task framework. Image-based methods, while showing decent performance, appear to be constrained by the modality gap. Originally trained on vast datasets of image-text pairs, where text is often a brief label, these methods now face fully formed, context-rich questions with a more limited training corpus. + +Seen vs. Unseen Performance: A notable performance drop from seen to unseen settings is observed across all baselines. This discrepancy underscores the difficulty in learning affordance knowledge that generalizes well, corroborating the necessity of our task design. Current methods struggle to bridge this gap in knowledge transfer, suggesting an avenue for future research: utilizing the extensive commonsense knowledge embedded in LLMs could potentially mitigate this challenge. + +# 5.2. In-Depth Study (Q2) + +Ablation Study. We show ablative results in Tab. 4. To verify that PointRefer prediction is based on the question, we conduct a blind test by training without questioning information, where the distinctive performance gap indicates that PointRefer has comprehended the question and can make a prediction based on it. Then, we study the effectiveness of PointRefer by removing both AFM and PRD ("w/o AFM & PRD"), which induces a severe performance + +Table 4. Ablative Results. + +
VariantsmIoU↑AUC↑SIM↑MAE↓
Blind10.878.20.5060.125
w/o AFM & PRD17.782.10.5580.110
w/o AFM19.883.90.5920.104
w/o multi-scale20.487.10.6280.095
w/o PRD18.785.60.6190.098
w/o condition19.286.90.5900.100
w/o dynamic filter19.587.10.6230.096
PointRefer20.887.30.6290.093
+ +![](images/0059e5e9abed1c1735ec3c0c0de6edbfdc697c4b25e56f290503c2482f64a20a.jpg) +Figure 5. The bar chart shows the training sample distribution over different question lengths, and the line chart shows the performance of PointRefer grouped by question length. + +decline on every metric, thus, validating the overall effectiveness of our design. Then, we also verify the functionality of AFM via the performance drop when it is removed ("w/o AFM). As a comparison, a relatively slighter declination is witnessed when we only apply AFM at the single scale ("w/o multi-scale"), which indicates the benefit of multi-scale cross-modal fusion. Then, we conduct some detailed breakdown tests to study PRD, we notice that when the affordance queries are not conditioned on the question ("w/o condition") the performance drops drastically, indicating that our affordance queries help to capture semantic clues in 3D space. Finally, we validate the superiority of the dynamic filtering strategy, because a declination is also observed when the dynamic filter is erased ("w/o dynamic filter"), and this variant adopts IAGNet's [44] mask prediction strategy as an alternative. + +Effect of Question Length. In Fig. 5, we show the performance of PointRefer on different question lengths. Specifically, The bar plot reveals the distribution of question length in the training set. Meanwhile, the line chart illustrates PointRefer's performance on different question lengths. Here, the performances are normalizing against the average number under that metric, e.g. IoU for each question length mIoU. Notably, PointRefer excels with questions that are more prevalent in the training set and with those that are lengthier—exceeding 14 words—owing to the richer contextual information they provide, which is instrumental in discerning the object's affordance. + +![](images/078425933709c631814008e073540023c5abcd8fed7395bd4362e95c58c7e601.jpg) +Figure 6. Illustration of how the percentage of omitted samples affect transferability of affordance knowledge (indicated by the difference between seen and unseen). For each affordance type, the bar shows a number of omitted samples from seen, i.e. $\frac{\# \text{Seen} - \# \text{Unseen}}{\# \text{Seen}} (\%)$ , and the lines show the testing IoU of models trained on Seen and Unseen. Difference (IoU): Seen - Unseen. + +Table 5. Study of Backbones. PN:PointNet++; PM:PointMLP. + +
3DLMmIoU↑AUC↑SIM↑MAE↓
SeenPM [27]Bert [8]18.285.50.6180.099
Deberta [11]19.386.50.6210.096
Roberta [23]19.686.462.30.097
PN [34]Bert [8]20.186.40.6150.101
Deberta [11]20.487.00.6240.095
Roberta [23]20.887.30.6290.093
UnseenPM [27]Bert [8]11.576.30.430.128
Deberta [11]12.376.80.4350.133
Roberta [23]12.176.80.4340.136
PN [34]Bert [8]14.178.10.470.122
Deberta [11]14.578.90.500.114
Roberta [23]14.680.20.5070.119
+ +Seen vs. Unseen. Recall that to create the training set for the unseen setting, for an affordance type, we omit its combination with certain objects from seen. For example, the seen partition has "grasp-mug", and "grasp-bag" in its training, we create the unseen training set by removing the "grasp-mug" from the seen training set. In this case, the model is expected to learn the generalizable affordance knowledge of "grasp" and transfer it to an unseen object during testing. Note that seen and unseen setting shares the same validation and testing set. To investigate how PointRefer captures such transferability for each affordance type, The bar in Fig. 6 shows the percentage of samples that are omitted from seen under each affordance type. Then, the lines show the performance of the model trained with seen and unseen partitions for each affordance type, and their difference—the indicator of transferability of learned affordance knowledge. In general, we notice that the curve of seen and unseen follows a similar pattern, which indicates that PointRefer can capture a similar affordance knowledge even trained with far less object class. However, we also notice that the difference between seen and unseen enlarges as the omitted percentage grows, regardless of the absolution IoU. This is expected since the removed object can lead to under-diversified affordance representation. + +![](images/1a89c80a6010d4952cc87e409c97820c74b8006f06ac5bac7b3fa9b3a5973077.jpg) +Figure 7. Case-Study of PointRefer's segmentation. Each showcase comes with one question and four shapes, showing the generalization ability of the affordance knowledge. The segmented affordance part is highlighted in red. + +Choice of Backbone. To S backbone, we investigate the I with some alternative backbone that, for both seen and unseen as a 3D backbone generally pr than PointMLP, thus, we set 1 3D backbone. As for the lamance of Roberta prevails Ber setting. Whereas in the unseen and Roberta brings no signifca PointRefer uses Roberta as th + +# 5.3. Qualitative Results + +Case Study. As shown in Fig. 1, the capacity to accurately co- given the question. It is noteworthy + +with small affordance components, such as the microwa door handle, our model still exhibits decent ability. + +![](images/3686ed50d26f28d574a77a62645f5c7782411961bc27d699591f15dd37a068ab.jpg) + +![](images/8763725ac8de1fd0427376685dcdc04c002dc725bdaa0231d50732676b7f6020.jpg) +Figure 8. Failure cases due to non-comprehended text (left) and over-segmentation (right). + +![](images/abd0418a56bb2f9c13e6c5d1c927312e02379f0beed5ed531f9d912515668230.jpg) +ting method? +GT + +![](images/497b81f755dcd97666eaf92c21d335e3f8d279880f2ea1e3f3ae7dd2bb2c8a1a.jpg) +When typing, which part should your fingers apply pressure to? +Ours + +![](images/481523d71a83bf7a67cf3b76e1f144cd04b017631c98ec86bd741d511284af78.jpg) +GT + +# 6. Conclusion + +In this paper, we have introduced the pioneering task of Language-guided Affordance Segmentation on 3D Objects (LASO), which establishes a vital connection between AI agents and Large Language Models (LLMs) via textual cues. Our contributions include the development of an extensive dataset consisting of 19,751 question-point pairs and the creation of PointRefer, a baseline model that lays the groundwork for the LASO task. We envisage that LASO will steer the trajectory of the 3D affordance field towards better integration with the advancements in LLMs. + +# 7. Acknowledgments + +This work is supported by the National Natural Science Foundation of China (92270114), the Agency for Science, Technology and Research (A\*STAR) under MTC Programmatic Funds (Grant No.M23L7b0021), and NExT center. + +# References + +[1] Do, etc. Affordancenet: An end-to-end deep learning approach for object affordance detection. In ICRA, 2018. 2 +[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 3 +[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19107-19117. IEEE, 2022. 3 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 5 +[5] Dave Zhenyu Chen, Angel X. Chang, and Matthias Nießner. Scanrefer: 3d object localization in RGB-D scans using natural language. In ECCV, pages 202-221. Springer, 2020. 3 +[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, 2022. 3 +[7] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In CVPR, pages 1778-1787. Computer Vision Foundation / IEEE, 2021. 2, 3, 6 +[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186. Association for Computational Linguistics, 2019. 7 +[9] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J. Lim. Demo2vec: Reasoning object affordances from online videos. In CVPR. Computer Vision Foundation / IEEE Computer Society, 2018. 2 +[10] Yiran Geng, Boshi An, Haoran Geng, Yuanpei Chen, Yaodong Yang, and Hao Dong. Rlafford: End-to-end affordance learning for robotic manipulation. In ICRA, pages 5880-5886. IEEE, 2023. 1 +[11] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: decoding-enhanced bert with disentangled attention. In ICLR. OpenReview.net, 2021. 7 +[12] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Tomas Jackson, Noah Brown, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In CoRL, pages 1769-1782. PMLR, 2022. 1 +[13] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Kate-rina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 3 +[14] Safoura Rezapour Lakani, Antonio Jose Rodríguez-Sánchez, and Justus H. Piater. Towards affordance detection for robot manipulation using affordance for parts and parts for affordance. pages 1155-1172, 2019. 1 + +[15] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, pages 19652-19664, 2021. 6, 1, 2 +[16] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In ACM MM, pages 4091-4099, 2021. 2 +[17] Yicong Li, Xiang Wang, Junbin Xiao, and Tat-Seng Chua. Equivariant and invariant grounding for video question answering. In ACM MM, pages 4714-4722, 2022. +[18] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, pages 2928-2937, 2022. +[19] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. TPAMI, 2023. +[20] Yicong Li, Junbin Xiao, Chun Feng, Xiang Wang, and TatSeng Chua. Discovering spatio-temporal rationales for video question answering. In ICCV, pages 13869-13878, 2023. +[21] Yicong Li, Xun Yang, An Zhang, Chun Feng, Xiang Wang, and Tat-Seng Chua. Redundancy-aware transformer for video question answering. In ACM MM, pages 3172-3180, 2023. 2 +[22] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: generalized referring expression segmentation. In CVPR, pages 23592-23601. IEEE, 2023. 6, 1, 2 +[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. 6, 7 +[24] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Trans. Artif. Intell., 4(5):1186-1198, 2023. 2 +[25] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In CVPR, pages 2242-2251. IEEE, 2022. 2 +[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16433-16442. IEEE, 2022. 3, 6, 1 +[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR. OpenReview.net, 2022. 7 +[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: situated question answering in 3d scenes. In ICLR. OpenReview.net, 2023. 3 +[29] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers Neurorobotics, 14:26, 2020. 2 +[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas J. Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In CoRL, pages 1666-1677. PMLR, 2021. 2 + +[31] Bogdan Moldovan, Plinio Moreno, Martijn van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE International Conference on Robotics and Automation, pages 4373-4378, 2012. 1 +[32] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In ICCV, pages 8687-8696. IEEE, 2019. 2 +[33] OpenAI. Gpt-4 technical report, 2023. 2, 3 +[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 4, 7 +[35] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In ICRA, pages 11523-11530. IEEE, 2023. 1 +[36] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. CoRR, abs/2212.04088, 2022. 1 +[37] NN Sriram, Tirth Maniar, Jayaganesh Kalyanasundaram, Vineet Gandhi, Brojeshwar Bhowmick, and K Madhava Krishna. Talk to the vehicle: Language conditioned autonomous navigation of self driving cars. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5284-5290. IEEE, 2019. 1 +[38] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, pages 24261-24272, 2021. 4 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5 +[40] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4964-4974. IEEE, 2022. 5 +[41] Junbin Xiao, Angela Yao, Yicong Li, and Tat Seng Chua. Can i trust your answer? visually grounded video question answering. arXiv preprint arXiv:2309.01327, 2023. 2 +[42] Amber Xie, Youngwoo Lee, Pieter Abbeel, and Stephen James. Language-conditioned path planning. In 7th Annual Conference on Robot Learning, 2023. 1 +[43] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. CoRR, abs/2202.13519, 2022. 2 +[44] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. CoRR, abs/2303.10437, 2023. 2, 3, 6, 7, 1 +[45] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2908-2917. IEEE, 2021. 3 + +[46] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Comput. Appl., 32(18):14321-14333, 2020. 2 \ No newline at end of file diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/images.zip b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0239349690c25e2d2a06d6d026de989e97a61297 --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:342e54df2bb8f82b89c707b057c80a2bcfac3124dadc270fc748ce6adf8b5dea +size 693129 diff --git a/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/layout.json b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e172c4b5158b526f20ed1754e405f8b6956e612b --- /dev/null +++ b/2024/LASO_ Language-guided Affordance Segmentation on 3D Object/layout.json @@ -0,0 +1,8071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "text", + "content": "LASO: Language-guided Affordance Segmentation on 3D Object" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "spans": [ + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": "Yicong Li" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": ", Na Zhao" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": ", Junbin Xiao" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": ", Chun Feng" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": ", Xiang Wang" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{3*†}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": ", Tat-seng Chua" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 142, + 496, + 159 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "spans": [ + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "text", + "content": "National University of Singapore, " + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "text", + "content": "Singapore University of Technology and Design, " + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 170, + 501, + 199 + ], + "type": "text", + "content": "University of Science and Technology of China," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 201, + 460, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 460, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 460, + 213 + ], + "type": "text", + "content": "liyicong@u.nus.edu,na_zhao@sutd.edu.sg,junbin@comp.nus.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 216, + 487, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 216, + 487, + 227 + ], + "spans": [ + { + "bbox": [ + 103, + 216, + 487, + 227 + ], + "type": "text", + "content": "fengchun3364@mail.ustc.edu.cn,xiangwangl223@gmail.com,dcstsc@nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 281, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 290, + 605 + ], + "type": "text", + "content": "Segmenting affordance in 3D data is key for bridging perception and action in robots. Existing efforts mostly focus on the visual side and overlook the affordance knowledge from a semantic aspect. This oversight not only limits their generalization to unseen objects, but more importantly, hinders their synergy with large language models (LLMs) which are excellent task planners that can decompose an overarching command into agent-actionable instructions. With this regard, we propose a novel task, Language-guided Affordance Segmentation on 3D Object (LASO), which challenges a model to segment a 3D object's part relevant to a given affordance question. To facilitate the task, we contribute a dataset comprising 19,751 point-question pairs, covering 8434 object shapes and 870 expert-crafted questions. As a pioneer solution, we further propose PointRefer, which highlights an adaptive fusion module to identify target affordance regions at different scales. To ensure a text-aware segmentation, we adopt a set of affordance queries conditioned on linguistic cues to generate dynamic kernels. These kernels are further used to convolute with point features and generate a segmentation mask. Comprehensive experiments and analyses validate PointRefer's effectiveness. With these efforts, We hope that LASO can steer the direction of 3D affordance, guiding it towards enhanced integration with the evolving capabilities of LLMs. Code and data are available at https://github.com/y13800/LASO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 641 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 287, + 673 + ], + "type": "text", + "content": "Affordance describes specific regions on an object that enable or facilitate certain human interactions. In this context," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 308, + 254, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 308, + 254, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 388 + ], + "type": "image", + "image_path": "304263c7468019abd031dd459f80d845303ab370be9420da4a6eb9a7ffaed1e1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 392, + 546, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 392, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 392, + 546, + 437 + ], + "type": "text", + "content": "Figure 1. Illustration of LASO, language-guided affordance segmentation on 3D object. Given an object point cloud and a question, the model is expected to segment the functional part delineated in question." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 446, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 546, + 519 + ], + "type": "text", + "content": "the task of 3D affordance segmentation is raised to identify parts of 3D objects that imply a certain function. It has been deemed as a critical step in bridging perception and operation in the physical world for an embodied agent, thus has shown substantial impact on practical applications such as robotic manipulation [10, 14, 31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "Recently, the advancement of Large Language Models (LLMs) has enabled their deployment as task planners, breaking down an intricate instruction into a sequence of affordance-aware, robot-executable instructions [12, 35, 36]. This equips intelligent agents with the capability to undertake complex tasks. For example, the overarching command \"Pass me the mug in the microwave\" can be systematically deconstructed into a series of affordance-aware inquiries: Q1) \"How to get to the microwave?\", Q2) \"How to open the microwave\", Q3) \"How to grab the mug?\", and Q4) \"How to return to the user?\" Ideally, each query would correspond to a predefined subtask that effectively translates language instructions into actionable steps. While Q1 and Q4 are related to language-conditioned path planning [42] and navigation [37], Q2 and Q3, which involve the 'open microwave' and 'grab the mug' actions, ex" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 683, + 136, + 693 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 693, + 286, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 286, + 704 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 286, + 704 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 693, + 286, + 704 + ], + "type": "text", + "content": " Xiang Wang is also affiliated with Institute of Artificial Intelligence," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 704, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 704, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 704, + 274, + 712 + ], + "type": "text", + "content": "Institute of Dataspace, Hefei Comprehensive National Science Center." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14251" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 96 + ], + "type": "text", + "content": "pose a gap — to the best of our knowledge, currently understanding affordance via language cues remains unexplored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 286, + 217 + ], + "type": "text", + "content": "Leaving language cues untouched, current 3D affordance segmentation focuses primarily on the visual aspect. It associates the specific geo-structure with either explicit affordance categories [7] or implicit affordance information from the 2D showcase [44]. Such visual-only settings make the segmentation models confined to visually predefined affordance types, thus impeding integration with language-based instructions from LLMs. As a result, fulfilling the aforementioned user's request becomes challenging, due to this inability to address Q2 and Q3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 482 + ], + "type": "text", + "content": "Towards these limitations, we introduce a novel task: Language-guided Affordance Segmentation on 3D Object (LASO), where 3D objects are paired with natural language questions that probe specific affordance parts. As a cornerstone task, LASO can provide an agent with affordance knowledge to answer the aforementioned Q2 and Q3 raised by LLM. For instance, the first two rows in Fig. 1 demonstrate how LASO imparts knowledge that enables an agent to interpret the semantic instructions \"open the microwave\" and \"grasp the mug\" in a 3D context, which are sufficient to execute Q2 and Q3, respectively. In LASO, questions are crafted to reflect diverse real-world scenarios, translating various affordance types into a range of semantic contexts. By learning from different contexts, we expect the agent not to memorize predefined affordance types but rather to abstract generalizable affordance knowledge within a broader semantic context. Notably, the diverse scales and shapes of affordance parts pose a significant perceptual challenge for the model. For instance, in the third row of Fig. 1, the target backrest of a chair occupies a large portion and has a significantly different shape compared to the small knob on a door in the last row." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 286, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 286, + 699 + ], + "type": "text", + "content": "To facilitate the study, we develop the 1st question-affordance dataset comprising 19,751 point cloud-question pairs. The questions, originally crafted by experts, are further diversified using GPT-4 [33] under principles of contextual enrichment, concise phrasing, and structural diversity. These principles enable a deep and varied exploration of object affordances. Based on the dataset, we have developed a strong baseline model named PointRefer that is capable of adapting to varying scales and shapes of object parts, conditioned on the question. It consists of two modules: 1) an Adaptive Fusion Module (AFM) that integrates question semantics to point features at different scales, which caters to varied target regions with better point feature discriminativeness; and 2) a Referred Point Decoder (RPD) that generates a segmentation mask via conditioned affordance queries, which, after refinement, forms dynamic convolution kernels to capture point-wise affordance for mask prediction. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "content": "- We propose the task of Language-guided Affordance Seg" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 95, + 536, + 177 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Efforts on affordance learning. Part indicates whether the prediction specifies the object part (✓) or the whole object " + }, + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "type": "inline_equation", + "content": "(\\times)" + }, + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 95, + 536, + 177 + ], + "lines": [ + { + "bbox": [ + 318, + 95, + 536, + 177 + ], + "spans": [ + { + "bbox": [ + 318, + 95, + 536, + 177 + ], + "type": "table", + "html": "
DomainStudiesVisionTextPart
2D[1, 46]Image××
[9, 25, 32]Image/Video×✓
[24, 29]Image✓×
3D[7, 30, 30, 43]Point Cloud×✓
[44]Point Cloud & Image×✓
OursPoint Cloud✓✓
", + "image_path": "31b0b464e12a209949a0fe651ce2f3ee5cf847ceaeb881a076774d1035daa881.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 184, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 545, + 220 + ], + "type": "text", + "content": "mentation on 3D Objects (LASO), which helps AI agents derive affordance knowledge from textual cues, facilitating a crucial advancement for LLM integration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 220, + 545, + 316 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 545, + 267 + ], + "type": "text", + "content": "- We develop a dataset with 19,751 question-point affordance pairs which are meticulously curated via a collaborative effort of human experts and GPT-4, thereby establishing the 1st benchmark for evaluating LASO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 268, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 545, + 316 + ], + "type": "text", + "content": "- We build a strong baseline PointRefer, which employs textual-conditioned affordance queries to isolate afforded segments and devises an adaptive fusion module to enhance the discriminability of point feature." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 392, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 443 + ], + "type": "text", + "content": "Affordance Learning. Originating from the image domain, initial efforts concentrate on detecting objects with affordances [1, 46]. Progressing from there, later studies [24, 29] incorporate linguistic descriptions to augment the process but still neglect the granularity of analysis, generally focusing on object-level affordances. Addressing this oversight, subsequent research [9, 25, 32] shift towards scrutinizing specific affordance parts, establishing a new norm for precision in the field." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 445, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 661 + ], + "type": "text", + "content": "With the rise of embodied AI, the scope of affordance learning expands into the 3D realm. 3D AffordanceNet [7] introduces the first benchmark dataset for learning affordance from object point cloud geometry. Building on this, Yang et al. [44] propose a setting for learning 3D affordance parts guided by image demonstrations. Nonetheless, these approaches predominantly rely on a visual-only approach, tightly coupling geometric features with affordance labels and overlooking the semantic dimension. This limitation hinders the integration with Large Language Models (LLMs) that could otherwise substantiate affordance segmentation to real-world deployment. Moreover, the common practice of predicting affordance types as an auxiliary task [7, 44] may dilute the primary challenge of understanding affordance knowledge. In contrast, our work sidesteps auxiliary predictions, advocating for direct learning from the linguistic context, which aligns more closely with the innate capabilities of LLMs and their semantic richness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Text-Point Cloud Cross-Modal Learning. The integration of vision and natural language has recently sparked increased interest [16-21, 41]. Such a tendency naturally extends to 3D vision due to its pivotal role in embodied AI. To" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14252" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 273, + 186 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 273, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 273, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 273, + 186 + ], + "type": "image", + "image_path": "f02f4f367fd6504dc0599f11f0b7da47dce2cc5d8670c945bc01f637a759f629.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 190, + 271, + 201 + ], + "lines": [ + { + "bbox": [ + 61, + 190, + 271, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 271, + 201 + ], + "type": "text", + "content": "Figure 2. Illustration of question augmentation principles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 289, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 289, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 289, + 424 + ], + "type": "text", + "content": "advance the 3D-language tasks, datasets like ScanRefer [5] and ReferIt3D [2] have been instrumental in benchmarking the grounding of natural language to 3D objects. Complementing this, Azuma et al. [3] introduce ScanQA for 3D question-answering, while Ma et al. [28] propose SQA3D for situated reasoning within 3D environments. In terms of models, a variety of efforts [3, 13, 45] have emerged from these benchmarks, with notable examples like 3D-SPS [26] and BUTD-DETR [13] employing cross-attention mechanisms and linguistic cues for object discovery. Others, such as 3DVG [45], and ViL3DRel [6], focus on 3D grounding by incorporating spatial relation cues, showcasing the evolving landscape of 3D-VL models. Unlike the existing work that focuses on grounding and QA that perform scene-level reasoning, we tackle a dense prediction task at object part level that bridges reasoning to operation. To the best of our knowledge, we are the first to introduce the 3D affordance problem as a language-guided segmentation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 430, + 101, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 101, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 101, + 441 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 444, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 287, + 504 + ], + "type": "text", + "content": "To underpin our proposed task, we have meticulously compiled a question-point affordance dataset with 19,751 paired samples, featuring 8,434 distinct object shapes across 23 classes. Accompanying these are 870 expertly crafted questions, covering 17 affordance categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 511, + 157, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 157, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 157, + 522 + ], + "type": "text", + "content": "3.1. Annotation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 530, + 287, + 638 + ], + "type": "text", + "content": "Point Cloud. Our dataset leverages the point cloud and affordance annotations from the 3D-AffordanceNet [7]. Specifically, the affordance annotation in [7] is multi-class, with each point on an object potentially supporting multiple affordance types. In our case, each question is tailored to a specific affordance type, while each sample can associate with several questions. Consequently, the same object can exhibit diverse affordance segmentation, depending on the specific question posed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Question. To create the questions, we draw 58 object-affordance combinations from [7], and craft 15 unique questions for each of the 58 combinations, resulting in 870 tailored questions. Specifically, five manual questions per combination are initially composed to represent different scenarios. Subsequently, we employ GPT-4 [33] to diver" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 310, + 84, + 544, + 132 + ], + "blocks": [ + { + "bbox": [ + 378, + 71, + 473, + 80 + ], + "lines": [ + { + "bbox": [ + 378, + 71, + 473, + 80 + ], + "spans": [ + { + "bbox": [ + 378, + 71, + 473, + 80 + ], + "type": "text", + "content": "Table 2. Dataset Statistics" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 84, + 544, + 132 + ], + "lines": [ + { + "bbox": [ + 310, + 84, + 544, + 132 + ], + "spans": [ + { + "bbox": [ + 310, + 84, + 544, + 132 + ], + "type": "table", + "html": "
TaskTrainValTest
ShapeQstInstShapeQstInstShapeQstInst
Seen68836381612051658121510351742416
Unseen616045811558
", + "image_path": "d709bc24a5eff4cb6cad052fd9b1c2a9cf88f7a166ec0b98c98eed85d9db1c07.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 138, + 544, + 250 + ], + "blocks": [ + { + "bbox": [ + 307, + 138, + 544, + 250 + ], + "lines": [ + { + "bbox": [ + 307, + 138, + 544, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 138, + 544, + 250 + ], + "type": "image", + "image_path": "058b1cac9cc8965d2e5d05a665f8f8ce3c0ea0c2d9db887094e1c4c1334812c5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 251, + 490, + 262 + ], + "lines": [ + { + "bbox": [ + 360, + 251, + 490, + 262 + ], + "spans": [ + { + "bbox": [ + 360, + 251, + 490, + 262 + ], + "type": "text", + "content": "Figure 3. Landscape of our dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 264, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 456 + ], + "type": "text", + "content": "sify the dataset with 10 additional questions for each combination, adhering to three key principles (as shown in Fig. 2): 1) Contextual Enrichment. Some questions were expanded to include additional context or details, fostering a more precise connection to the specified affordance of the target object. For instance, Q1 in Fig. 2 is reformed by a functional situation \"cutting paper\". 2) Concise Phrasing. In other cases, questions were distilled to their essence, making them succinct yet still meaningful. Q2 in Fig. 2 exemplifies this approach. 3) Structural Diversity. We employed a range of sentence structures, from questions to statements, to introduce natural linguistic variations and prevent model bias toward any specific phrasing or sentence length. A diversified example can be found in the Q3 of Fig. 2. All questions generated by GPT-4 undergo manual verification after this augmentation process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 456, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 540 + ], + "type": "text", + "content": "It's worth mentioning that the affordance type was only a tool for crafting the questions and pinpointing the correct segmentation areas. We refrain from using explicit affordance labels during both training and testing. This deliberate choice ensures that the learning process is driven by the semantic content of the questions, rather than by direct affordance label association." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 427, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 427, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 427, + 563 + ], + "type": "text", + "content": "3.2. Statistics and Setting" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 616 + ], + "type": "text", + "content": "Dataset Settings. We show the statistical landscape of our dataset in Fig. 3. Specifically, it contains 17 affordance types and 23 object categories, composing 58 unique object-affordance combinations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "In the spirit of the work by [44], we offer two distinct dataset settings: Seen: The default configuration, where the training and testing phases share similar distributions of object classes and affordance types. Unseen: This configuration is purpose-built to test the model's ability to generalize affordance knowledge. Certain affordance types, when paired with specific object classes, are omitted from the training set but included in the test set. For instance, while" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14253" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 132 + ], + "type": "text", + "content": "the model may learn to grasp bags and mugs during training, it is expected to apply the concept of 'grasp' to earphones, which is an affordance-object pairing not encountered during training. For a comprehensive account of the unseen setting, please refer to Appendix Sec. 7.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 323 + ], + "type": "text", + "content": "Training and Evaluation Protocols. Notably, since questions are created based on a combination of object class and affordance type, one object class can have many shape instances. Thus, the design of our dataset accounts for the multiplicity of shapes within each object class, leading to a non-bijective pairing between shape instances and corresponding questions. To ensure comprehensive learning, during training, each shape instance is matched with a randomly selected question that aligns with its affordance type for every iteration. This random pairing exposes the model to a variety of semantic contexts, bolstering its ability to generalize. For the validation and test sets, question pairings are fixed to minimize variability and ensure consistent inference. These questions are unique to the evaluation phase and are not revealed during the training process to maintain the integrity of the evaluation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 102, + 337 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "content": "LASO Task Definition. Given a question " + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "inline_equation", + "content": "Q_{raw}" + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "content": " and an object point cloud " + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{raw} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "content": " points, the goal of LASO is to predict a binary mask of " + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 346, + 287, + 394 + ], + "type": "text", + "content": " that segments the functional part related to the question." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 649 + ], + "type": "text", + "content": "Framework Overview. Due to the varying scale of target affordance region, LASO naturally challenges models adaptiveness at different scales. To address this, we devise a strong baseline, named PointRefer. As shown in Fig. 4, the object point cloud " + }, + { + "bbox": [ + 46, + 398, + 287, + 649 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{raw}" + }, + { + "bbox": [ + 46, + 398, + 287, + 649 + ], + "type": "text", + "content": " is first processed by a 3D backbone, which typically consists of the multi-stage encoding and decoding process. To adapt PointRefer to the point feature at multiple scales, we introduce an Adaptive Fusion Module (AFM) to the backbone's decoding process. It performs multi-scale cross-modal fusion by injecting text clues to point features at a different decoding stage, which progressively refine the point feature map in a top-down manner. Then, to predict the segmentation mask, we introduce a Referred Point Decoder (RPD) that leverages a set of learnable queries conditioned on the input question, termed affordance queries, as the input of the transformer decoder. These affordance queries are obligated to look at the referred points only and generate question-aware dynamic kernels. The final segmentation mask is obtained by convoluting these dynamic kernels with the AFM-enhanced point feature." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Feature Encoding. For the question, we prepend a \"[CLS]\" token to the question sequence before feeding them to a language model to capture the global language context. The encoded feature of the question is denoted by " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " is the number of tokens and " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " is the feature di" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "mension. For the point cloud, PointRefer can adopt any off-the-shelf 3D segmentation backbone.1 to transform the " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{raw}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " into the feature space. Note that a 3D backbone for segmentation typically consists of several encoding and decoding stages, and we take the point-wise feature map after the last decoding stage as the output of the backbone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 152, + 444, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 444, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 444, + 164 + ], + "type": "text", + "content": "4.1. Adaptive Fusion Module" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 277 + ], + "type": "text", + "content": "AFM is designed to enhance the point-wise features by incorporating question information. As illustrated in Fig. 4, we integrate AFM into different decoder layers to enhance its adaptiveness to target regions of various scales and shapes. Given the conventional downsampling operators in 3D backbones (e.g. PointNet++ [34]), the number of points varies across different encoder/decoder layers. Our AFM is designed to seamlessly adapt to the varying number of points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "content": "For simplification, we omit the notation for different decoder layers and use " + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "content": " to denote the point feature at a certain decoding layer, where " + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 277, + 545, + 421 + ], + "type": "text", + "content": " represents the number of points at that layer. As visualized in Fig. 4, AFM adopts a bottleneck architecture with three key steps: Grouping, Mixing, and Ungrouping. In the grouping step, the point features are grouped by text tokens, these grouped tokens are then thoroughly mixed via an MLP-Mixer [38]. Subsequently, the ungrouping step encapsulates the mixed features into individual point features, completing the AFM procedure by embedding valuable textual information into the point feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "text", + "content": "To facilitate an efficient grouping process for our AFM, we employ a lightweight cross-attention module by omitting the transformation of both query and value. It takes the question feature " + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "text", + "content": " as query and the point feature " + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "text", + "content": " as key and value, then output grouped token " + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 304, + 421, + 545, + 480 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 357, + 490, + 545, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 490, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 357, + 490, + 545, + 503 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\operatorname {A t t e n t i o n} \\left(\\mathbf {X}, \\mathbf {W} _ {1} \\mathbf {P}, \\mathbf {P}\\right) + \\mathbf {X}, \\tag {1}", + "image_path": "174d0e5e8ca7785118ffd11126c956ca8501460380e5a7c319631f2bd7154476.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 506, + 545, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 506, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 336, + 506, + 545, + 533 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d}}\\right) \\mathbf {V}, \\tag {2}", + "image_path": "f25bc844cedaea708836987ce54e2a136abf21a5d2ba548faef67097577000e1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 541, + 455, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 541, + 455, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 455, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 541, + 455, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1" + }, + { + "bbox": [ + 306, + 541, + 455, + 552 + ], + "type": "text", + "content": " is a linear transformation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 553, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 545, + 601 + ], + "type": "text", + "content": "After attaining the grouped feature, we update them using an MLP-Mixer [38], which employs two consecutive multilayer perceptrons (MLPs), producing fused feature " + }, + { + "bbox": [ + 304, + 553, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 553, + 545, + 601 + ], + "type": "text", + "content": " by combining with two consecutive MLPs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 374, + 609, + 545, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 609, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 374, + 609, + 545, + 623 + ], + "type": "interline_equation", + "content": "\\mathbf {G} ^ {\\prime} = \\mathbf {G} + \\operatorname {M L P} _ {1} \\left(\\mathbf {G} ^ {T}\\right) ^ {T}, \\tag {3}", + "image_path": "0a82b00c829d2b7b9e26a65cfaabfa2b0b7724a369828ba59ed1bc84f91bcae5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 380, + 624, + 545, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 624, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 380, + 624, + 545, + 639 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\mathbf {G} ^ {\\prime} + \\mathrm {M L P} _ {2} \\left(\\mathbf {G} ^ {\\prime}\\right), \\tag {4}", + "image_path": "1ff22ce9cea6168abb6aa2bb3ea945399ed79e2b5d8da60afa96f1889b4949b5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_1" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathrm{MLP}_2" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " are used to mix group-wise and channel-wise information, respectively, and " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "T[\\cdot]" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " denotes the transpose operation. Finally, we return the fused feature " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " to the point-space via the ungrouping process and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 511, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 511, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 511, + 712 + ], + "type": "text", + "content": "1The impact of backbone choices can be observed in Tab. 5." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14254" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 388, + 280 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 388, + 280 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 388, + 280 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 388, + 280 + ], + "type": "image", + "image_path": "e9b18ca47acf1010aba4b1a6f550693ca4313b08db5a15d713f522946cff3c15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 281, + 268, + 291 + ], + "lines": [ + { + "bbox": [ + 167, + 281, + 268, + 291 + ], + "spans": [ + { + "bbox": [ + 167, + 281, + 268, + 291 + ], + "type": "text", + "content": "(a) Overview of PointRefer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 415, + 74, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 415, + 74, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 415, + 74, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 415, + 74, + 547, + 277 + ], + "type": "image", + "image_path": "d72ddcef34d4d0b43226e356b67b7c1342c77602abfa91c7a9b82df45cf402de.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 277, + 518, + 289 + ], + "lines": [ + { + "bbox": [ + 414, + 277, + 518, + 289 + ], + "spans": [ + { + "bbox": [ + 414, + 277, + 518, + 289 + ], + "type": "text", + "content": "(b) Adaptive Fusion Module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 292, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 45, + 292, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 45, + 292, + 545, + 348 + ], + "type": "text", + "content": "Figure 4. The overall framework of our proposed PointRefer (a), which contains an adaptive fusion modal (b) that can be injected in different stages of the backbone decoder. As a result, it takes the point-wise feature generated at the last decoder layer as the final output of the 3D backbone. Then, the Referred Point Decoder takes in a set of question-conditioned affordance queries and feeds it to a transformer decoder together with the fused point feature, where the output is then reformed to a set of dynamic kernels. Finally, these dynamic kernels are used as convolution kernels to filter out the segmentation mask from the fused point feature." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "text", + "content": "acquire the fused point feature " + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^m" + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "text", + "content": ". We implement the ungrouping process as a similar lightweight attention module as the grouping process. However, in this case, it takes " + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "text", + "content": " as query and " + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 354, + 287, + 401 + ], + "type": "text", + "content": " as key and value:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 411, + 287, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 411, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 96, + 411, + 287, + 423 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {m} = \\operatorname {A t t e n t i o n} (\\mathbf {P}, \\mathbf {W} _ {2} \\mathbf {F}, \\mathbf {F}) + \\mathbf {P}, \\tag {5}", + "image_path": "49011b1c60292716e2aa6f4869859c5ddfa28610799fbc821617f28e9ca7cbae.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "content": " is a linear transformation. Lastly, the final output " + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^o" + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "content": " is formed by adding residual connections to " + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^m" + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "content": ". Notably, we take " + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^o" + }, + { + "bbox": [ + 46, + 434, + 287, + 480 + ], + "type": "text", + "content": " after the last decoding stage as the output of the 3D backbone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 489, + 180, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 180, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 180, + 500 + ], + "type": "text", + "content": "4.2. Referred Point Decoder" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 638 + ], + "type": "text", + "content": "Inspired by the success of learnable query-based methods in object detection [4] and segmentation [40], we introduce learnable affordance queries conditioned on the corresponding question of an object to decode the segmentation mask for the object. Specifically, we impose linguistic restrictions on all affordance queries, narrowing the model's focus to the relevant object parts and thereby facilitating convergence. These affordance queries, after being processed by a transformer decoder, are repurposed as dynamic kernels to deduce the segmentation mask from the point features, which are the final output of the 3D backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Question-Conditioned Affordance Query In the transformer decoder, it is well known that content and position embeddings are responsible for encoding instance-specific and spatial information, respectively. In PointRefer, we feed these two parts with the text feature and learnable query parameters, respectively, so that all the queries are" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "text", + "content": "shaped by the nuances of the linguistic input. Specifically, the affordance queries " + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "text", + "content": " are formed by adding question embedding " + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "text", + "content": " with learnable embeddings. Thus, the resulting " + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{L \\times d}" + }, + { + "bbox": [ + 305, + 354, + 545, + 401 + ], + "type": "text", + "content": " are the same length of the question." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "text", + "content": "Reasoning and Mask Prediction. After forming the affordance query, we feed it to a transformer decoder [39] together with the fused point feature " + }, + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_o" + }, + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "text", + "content": " from the 3D backbone. In this manner, all the queries will use the question context as guidance and target to aggregate the feature of the referred points from the object, resulting in " + }, + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathbf{A}'" + }, + { + "bbox": [ + 305, + 404, + 545, + 476 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 486, + 545, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 486, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 353, + 486, + 545, + 499 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {\\prime} = \\text {T r a n s f o r m e r - D e c o d e r} (\\mathbf {A}; \\mathbf {P} _ {o}). \\tag {6}", + "image_path": "5ef9600c9bfd07754b763f750d7ff374c66dc00c728340bb7f908177b5e1e4dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": "Subsequently, we perform dynamic convolution by applying a two-layer MLP on top of the transfer decoder to produce " + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": " dynamic kernels " + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\Omega = \\{\\omega_{i}\\}_{i = 1}^{L}" + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": ", where each kernel " + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": " is reshaped to a " + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": " convolutional kernel with the channel number of " + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 509, + 545, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 393, + 578, + 545, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 578, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 393, + 578, + 545, + 590 + ], + "type": "interline_equation", + "content": "\\Omega = \\operatorname {M L P} \\left(\\mathbf {A} ^ {\\prime}\\right). \\tag {7}", + "image_path": "52087d14350d342656f5e4cc0cb13d480eb118092db8c1034dedcff0fb625505.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": "Since the dynamic kernels have captured the question- information, we use them as convolution filters on the feature maps for mask decoding. Specifically, we use each dynamic kernel in " + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": " to convolute the point-wise feature map " + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_o" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": ", which returns " + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": " point masks " + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\{\\mathbf{s}_i\\}_{i=1}^L" + }, + { + "bbox": [ + 305, + 601, + 545, + 661 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 383, + 665, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 665, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 383, + 665, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {i} = \\left\\{\\mathbf {P} _ {o} * \\omega_ {i} \\right\\} _ {i = 1} ^ {L}. \\tag {8}", + "image_path": "3ac4577aeb2c4339e590bf65b227431f4b82792f58b1f94fd6e63e14d62d0027.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Then, we apply mean pooling over all masks in " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " followed by a sigmoid activation to acquire the final segmentation" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14255" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 111, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 111, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 111, + 83 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 47, + 72, + 111, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in \\mathbb{R}^N" + }, + { + "bbox": [ + 47, + 72, + 111, + 83 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 90, + 287, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 90, + 287, + 102 + ], + "spans": [ + { + "bbox": [ + 115, + 90, + 287, + 102 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\sigma (\\operatorname {M a x - P o o l} ((\\mathbf {S}))), \\tag {9}", + "image_path": "fd7dab5f845ca1503f56a178e7f39ba0e12ba642e9cf29a7e0fcd06af2d88c67.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 109, + 211, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 211, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 211, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 109, + 211, + 120 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 47, + 109, + 211, + 120 + ], + "type": "text", + "content": " denote the sigmoid function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 121, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 287, + 180 + ], + "type": "text", + "content": "Intuitively, each dynamic kernel is expected to capture one aspect of the referred object part according to conditioned the question token. By ensembling all generated point masks, the final segmentation mask can integrate different response regions, thus making the result more robust." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 182, + 287, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 287, + 254 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 287, + 254 + ], + "type": "text", + "content": "Objectives. Unlike [44], which relies on an auxiliary affordance label for prediction, our model seeks to forge a direct link between language context and object affordance. Thus, we solely employ Dice loss and Binary Cross-Entropy (BCE) loss to guide the segmentation mask prediction, bypassing the need for additional affordance labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 123, + 256, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 256, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 123, + 256, + 287, + 268 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {B C E} + \\mathcal {L} _ {D i c e} \\tag {10}", + "image_path": "feb3542d163765cf6b7e1bd52b6f9c17e55a72c7da0d9956c57e72c4354c9a9d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 263, + 123, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 263, + 123, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 263, + 123, + 277 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 283, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 283, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 287, + 367 + ], + "type": "text", + "content": "Evaluation Metrics. For a thorough assessment, we benchmark our dataset against leading studies in 3D affordance learning [7, 44], employing four evaluation metrics to ensure robust analysis. These include Area Under the Curve (AUC), Mean Intersection Over Union (mIoU), Similarity (SIM), and Mean Absolute Error (MAE). Appendix Sec. 7.2 provides detailed explanations for the evaluation protocols." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 369, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 441 + ], + "type": "text", + "content": "Implementation Details. PointRefer employs PointNet++ as the default 3D backbone to align with the standards established in [7, 44]. For text encoding, we utilize a pretrained RoBERTa model [23] to process linguistic inputs. The feature dimension " + }, + { + "bbox": [ + 47, + 369, + 287, + 441 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 369, + 287, + 441 + ], + "type": "text", + "content": " is set to 512. During training, we use the Adam optimizer with a learning rate set to 1e-4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 443, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 287, + 634 + ], + "type": "text", + "content": "Baseline Models. Since there is no prior work using paired question-point cloud data to segment object affordance. For a thorough comparison of our method, we adopt two 3D cross-modal works (3D-SPS [26] and IAGNet[44]) and two referred image segmentation works (ReferTrans [15] and RelA [22]) to LASO task. For the referred image segmentation models, we substitute their image backbones with a 3D counterpart, retaining their fusion mechanisms and mask decoding strategies. For the language-guided grounding model, 3D-SPS [26], we omit its bounding box prediction module while leveraging its inherent point selection scheme for our segmentation setting. The most related work is IAGNet [44], an affordance detection method that takes paired image-point cloud as input. To adapt IAGNet to our needs, we simply replace its image backbone with a language model, preserving the remainder of its architecture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 639, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 662 + ], + "type": "text", + "content": "Next, we show the effectiveness of our PointRefer by answering the following questions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 663, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 663, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 286, + 686 + ], + "type": "text", + "content": "- Q1: How is PointRefer compared to other baselines on the proposed dataset?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 687, + 258, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 687, + 258, + 699 + ], + "spans": [ + { + "bbox": [ + 47, + 687, + 258, + 699 + ], + "type": "text", + "content": "- Q2: How effective are the proposed components?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 700, + 282, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 282, + 711 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 282, + 711 + ], + "type": "text", + "content": "- Q3: What learning pattern does the PointRefer capture?" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 317, + 106, + 538, + 235 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 104 + ], + "type": "text", + "content": "Table 3. Main Results. The overall results of all comparative methods, the best results are in bold. Seen and Unseen are two partitions of the dataset. AUC and aIOU are shown in percentage." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 106, + 538, + 235 + ], + "lines": [ + { + "bbox": [ + 317, + 106, + 538, + 235 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 538, + 235 + ], + "type": "table", + "html": "
MethodmIoU↑AUC↑SIM↑MAE↓
SteenReferTrans [15]13.779.80.4970.124
ReLA [22]15.278.90.5320.118
3D-SPS [26]11.476.20.4330.138
IAGNet [44]17.882.30.5610.109
PointRefer20.887.30.6290.093
UnseenReferTrans [15]10.269.10.4320.145
ReLA [22]10.769.70.4290.144
3D-SPS [26]7.968.80.4020.158
IAGNet [44]12.977.80.4430.129
PointRefer14.680.20.5070.119
", + "image_path": "37931b6d0726af191bd50a6c013764788cb4c8cb21bc1cd62b40c5eef736dbba.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 245, + 410, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 245, + 410, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 245, + 410, + 258 + ], + "type": "text", + "content": "5.1. Main Result (Q1)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 545, + 299 + ], + "type": "text", + "content": "In Table 3, PointRefer demonstrates superior performance across all evaluation metrics compared to the baseline methods. Our observations are as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 302, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 482 + ], + "type": "text", + "content": "PointRefer vs. Other Models: IAGNet secures the second-best performance in both seen and unseen settings for all metrics. This is likely due to its specific design for object affordance segmentation, despite differences in input modality, making it a strong fit for our task with minimal adaptation. In contrast, 3D-SPS underperforms, potentially due to its progressive point selection mechanism. This approach involves a non-differentiable selection process, which may complicate the optimization within our task framework. Image-based methods, while showing decent performance, appear to be constrained by the modality gap. Originally trained on vast datasets of image-text pairs, where text is often a brief label, these methods now face fully formed, context-rich questions with a more limited training corpus." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 592 + ], + "type": "text", + "content": "Seen vs. Unseen Performance: A notable performance drop from seen to unseen settings is observed across all baselines. This discrepancy underscores the difficulty in learning affordance knowledge that generalizes well, corroborating the necessity of our task design. Current methods struggle to bridge this gap in knowledge transfer, suggesting an avenue for future research: utilizing the extensive commonsense knowledge embedded in LLMs could potentially mitigate this challenge." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 599, + 425, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 425, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 425, + 612 + ], + "type": "text", + "content": "5.2. In-Depth Study (Q2)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Ablation Study. We show ablative results in Tab. 4. To verify that PointRefer prediction is based on the question, we conduct a blind test by training without questioning information, where the distinctive performance gap indicates that PointRefer has comprehended the question and can make a prediction based on it. Then, we study the effectiveness of PointRefer by removing both AFM and PRD (\"w/o AFM & PRD\"), which induces a severe performance" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14256" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 83, + 284, + 190 + ], + "blocks": [ + { + "bbox": [ + 119, + 71, + 215, + 80 + ], + "lines": [ + { + "bbox": [ + 119, + 71, + 215, + 80 + ], + "spans": [ + { + "bbox": [ + 119, + 71, + 215, + 80 + ], + "type": "text", + "content": "Table 4. Ablative Results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 83, + 284, + 190 + ], + "lines": [ + { + "bbox": [ + 53, + 83, + 284, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 284, + 190 + ], + "type": "table", + "html": "
VariantsmIoU↑AUC↑SIM↑MAE↓
Blind10.878.20.5060.125
w/o AFM & PRD17.782.10.5580.110
w/o AFM19.883.90.5920.104
w/o multi-scale20.487.10.6280.095
w/o PRD18.785.60.6190.098
w/o condition19.286.90.5900.100
w/o dynamic filter19.587.10.6230.096
PointRefer20.887.30.6290.093
", + "image_path": "fe6f55be7872c2c69349e23bbb6881cb74d1615b258bfa0c469c0f7c157b16f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 67, + 194, + 265, + 323 + ], + "blocks": [ + { + "bbox": [ + 67, + 194, + 265, + 323 + ], + "lines": [ + { + "bbox": [ + 67, + 194, + 265, + 323 + ], + "spans": [ + { + "bbox": [ + 67, + 194, + 265, + 323 + ], + "type": "image", + "image_path": "0059e5e9abed1c1735ec3c0c0de6edbfdc697c4b25e56f290503c2482f64a20a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 329, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 363 + ], + "type": "text", + "content": "Figure 5. The bar chart shows the training sample distribution over different question lengths, and the line chart shows the performance of PointRefer grouped by question length." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 564 + ], + "type": "text", + "content": "decline on every metric, thus, validating the overall effectiveness of our design. Then, we also verify the functionality of AFM via the performance drop when it is removed (\"w/o AFM). As a comparison, a relatively slighter declination is witnessed when we only apply AFM at the single scale (\"w/o multi-scale\"), which indicates the benefit of multi-scale cross-modal fusion. Then, we conduct some detailed breakdown tests to study PRD, we notice that when the affordance queries are not conditioned on the question (\"w/o condition\") the performance drops drastically, indicating that our affordance queries help to capture semantic clues in 3D space. Finally, we validate the superiority of the dynamic filtering strategy, because a declination is also observed when the dynamic filter is erased (\"w/o dynamic filter\"), and this variant adopts IAGNet's [44] mask prediction strategy as an alternative." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 713 + ], + "type": "text", + "content": "Effect of Question Length. In Fig. 5, we show the performance of PointRefer on different question lengths. Specifically, The bar plot reveals the distribution of question length in the training set. Meanwhile, the line chart illustrates PointRefer's performance on different question lengths. Here, the performances are normalizing against the average number under that metric, e.g. IoU for each question length mIoU. Notably, PointRefer excels with questions that are more prevalent in the training set and with those that are lengthier—exceeding 14 words—owing to the richer contextual information they provide, which is instrumental in discerning the object's affordance." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 71, + 527, + 211 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 527, + 211 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 527, + 211 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 527, + 211 + ], + "type": "image", + "image_path": "078425933709c631814008e073540023c5abcd8fed7395bd4362e95c58c7e601.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 209, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 545, + 275 + ], + "type": "text", + "content": "Figure 6. Illustration of how the percentage of omitted samples affect transferability of affordance knowledge (indicated by the difference between seen and unseen). For each affordance type, the bar shows a number of omitted samples from seen, i.e. " + }, + { + "bbox": [ + 305, + 209, + 545, + 275 + ], + "type": "inline_equation", + "content": "\\frac{\\# \\text{Seen} - \\# \\text{Unseen}}{\\# \\text{Seen}} (\\%)" + }, + { + "bbox": [ + 305, + 209, + 545, + 275 + ], + "type": "text", + "content": ", and the lines show the testing IoU of models trained on Seen and Unseen. Difference (IoU): Seen - Unseen." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 292, + 537, + 422 + ], + "blocks": [ + { + "bbox": [ + 310, + 278, + 541, + 289 + ], + "lines": [ + { + "bbox": [ + 310, + 278, + 541, + 289 + ], + "spans": [ + { + "bbox": [ + 310, + 278, + 541, + 289 + ], + "type": "text", + "content": "Table 5. Study of Backbones. PN:PointNet++; PM:PointMLP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 292, + 537, + 422 + ], + "lines": [ + { + "bbox": [ + 317, + 292, + 537, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 537, + 422 + ], + "type": "table", + "html": "
3DLMmIoU↑AUC↑SIM↑MAE↓
SeenPM [27]Bert [8]18.285.50.6180.099
Deberta [11]19.386.50.6210.096
Roberta [23]19.686.462.30.097
PN [34]Bert [8]20.186.40.6150.101
Deberta [11]20.487.00.6240.095
Roberta [23]20.887.30.6290.093
UnseenPM [27]Bert [8]11.576.30.430.128
Deberta [11]12.376.80.4350.133
Roberta [23]12.176.80.4340.136
PN [34]Bert [8]14.178.10.470.122
Deberta [11]14.578.90.500.114
Roberta [23]14.680.20.5070.119
", + "image_path": "ef69cb8a8a45baf52e0ce3e91993bedc4dd179df47f74081bb20faab5285d120.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "content": "Seen vs. Unseen. Recall that to create the training set for the unseen setting, for an affordance type, we omit its combination with certain objects from seen. For example, the seen partition has \"grasp-mug\", and \"grasp-bag\" in its training, we create the unseen training set by removing the \"grasp-mug\" from the seen training set. In this case, the model is expected to learn the generalizable affordance knowledge of \"grasp\" and transfer it to an unseen object during testing. Note that seen and unseen setting shares the same validation and testing set. To investigate how PointRefer captures such transferability for each affordance type, The bar in Fig. 6 shows the percentage of samples that are omitted from seen under each affordance type. Then, the lines show the performance of the model trained with seen and unseen partitions for each affordance type, and their difference—the indicator of transferability of learned affordance knowledge. In general, we notice that the curve of seen and unseen follows a similar pattern, which indicates that PointRefer can capture a similar affordance knowledge even trained with far less object class. However, we also notice that the difference between seen and unseen enlarges as the omitted percentage grows, regardless of the absolution IoU. This is expected since the removed object can lead to under-diversified affordance representation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14257" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 49, + 527, + 339 + ], + "blocks": [ + { + "bbox": [ + 71, + 49, + 527, + 339 + ], + "lines": [ + { + "bbox": [ + 71, + 49, + 527, + 339 + ], + "spans": [ + { + "bbox": [ + 71, + 49, + 527, + 339 + ], + "type": "image", + "image_path": "1a89c80a6010d4952cc87e409c97820c74b8006f06ac5bac7b3fa9b3a5973077.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 546, + 361 + ], + "type": "text", + "content": "Figure 7. Case-Study of PointRefer's segmentation. Each showcase comes with one question and four shapes, showing the generalization ability of the affordance knowledge. The segmented affordance part is highlighted in red." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 369, + 179, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 179, + 500 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 179, + 500 + ], + "type": "text", + "content": "Choice of Backbone. To S backbone, we investigate the I with some alternative backbone that, for both seen and unseen as a 3D backbone generally pr than PointMLP, thus, we set 1 3D backbone. As for the lamance of Roberta prevails Ber setting. Whereas in the unseen and Roberta brings no signifca PointRefer uses Roberta as th" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 516, + 160, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 160, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 160, + 529 + ], + "type": "text", + "content": "5.3. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 533, + 169, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 169, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 169, + 569 + ], + "type": "text", + "content": "Case Study. As shown in Fig. 1, the capacity to accurately co- given the question. It is noteworthy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 570, + 278, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 278, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 278, + 594 + ], + "type": "text", + "content": "with small affordance components, such as the microwa door handle, our model still exhibits decent ability." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 179, + 375, + 239, + 489 + ], + "blocks": [ + { + "bbox": [ + 179, + 375, + 239, + 489 + ], + "lines": [ + { + "bbox": [ + 179, + 375, + 239, + 489 + ], + "spans": [ + { + "bbox": [ + 179, + 375, + 239, + 489 + ], + "type": "image", + "image_path": "3686ed50d26f28d574a77a62645f5c7782411961bc27d699591f15dd37a068ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 246, + 376, + 334, + 487 + ], + "blocks": [ + { + "bbox": [ + 246, + 376, + 334, + 487 + ], + "lines": [ + { + "bbox": [ + 246, + 376, + 334, + 487 + ], + "spans": [ + { + "bbox": [ + 246, + 376, + 334, + 487 + ], + "type": "image", + "image_path": "8763725ac8de1fd0427376685dcdc04c002dc725bdaa0231d50732676b7f6020.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 476, + 545, + 499 + ], + "lines": [ + { + "bbox": [ + 306, + 476, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 545, + 499 + ], + "type": "text", + "content": "Figure 8. Failure cases due to non-comprehended text (left) and over-segmentation (right)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 354, + 388, + 423, + 467 + ], + "blocks": [ + { + "bbox": [ + 356, + 373, + 396, + 382 + ], + "lines": [ + { + "bbox": [ + 356, + 373, + 396, + 382 + ], + "spans": [ + { + "bbox": [ + 356, + 373, + 396, + 382 + ], + "type": "text", + "content": "ting method?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 388, + 423, + 467 + ], + "lines": [ + { + "bbox": [ + 354, + 388, + 423, + 467 + ], + "spans": [ + { + "bbox": [ + 354, + 388, + 423, + 467 + ], + "type": "image", + "image_path": "abd0418a56bb2f9c13e6c5d1c927312e02379f0beed5ed531f9d912515668230.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 468, + 405, + 475 + ], + "lines": [ + { + "bbox": [ + 394, + 468, + 405, + 475 + ], + "spans": [ + { + "bbox": [ + 394, + 468, + 405, + 475 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 389, + 472, + 467 + ], + "blocks": [ + { + "bbox": [ + 432, + 369, + 529, + 386 + ], + "lines": [ + { + "bbox": [ + 432, + 369, + 529, + 386 + ], + "spans": [ + { + "bbox": [ + 432, + 369, + 529, + 386 + ], + "type": "text", + "content": "When typing, which part should your fingers apply pressure to?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 389, + 472, + 467 + ], + "lines": [ + { + "bbox": [ + 433, + 389, + 472, + 467 + ], + "spans": [ + { + "bbox": [ + 433, + 389, + 472, + 467 + ], + "type": "image", + "image_path": "497b81f755dcd97666eaf92c21d335e3f8d279880f2ea1e3f3ae7dd2bb2c8a1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 468, + 460, + 475 + ], + "lines": [ + { + "bbox": [ + 443, + 468, + 460, + 475 + ], + "spans": [ + { + "bbox": [ + 443, + 468, + 460, + 475 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 486, + 391, + 526, + 468 + ], + "blocks": [ + { + "bbox": [ + 486, + 391, + 526, + 468 + ], + "lines": [ + { + "bbox": [ + 486, + 391, + 526, + 468 + ], + "spans": [ + { + "bbox": [ + 486, + 391, + 526, + 468 + ], + "type": "image", + "image_path": "481523d71a83bf7a67cf3b76e1f144cd04b017631c98ec86bd741d511284af78.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 468, + 508, + 475 + ], + "lines": [ + { + "bbox": [ + 498, + 468, + 508, + 475 + ], + "spans": [ + { + "bbox": [ + 498, + 468, + 508, + 475 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 506, + 379, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 379, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 379, + 517 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 519, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 547, + 639 + ], + "type": "text", + "content": "In this paper, we have introduced the pioneering task of Language-guided Affordance Segmentation on 3D Objects (LASO), which establishes a vital connection between AI agents and Large Language Models (LLMs) via textual cues. Our contributions include the development of an extensive dataset consisting of 19,751 question-point pairs and the creation of PointRefer, a baseline model that lays the groundwork for the LASO task. We envisage that LASO will steer the trajectory of the 3D affordance field towards better integration with the advancements in LLMs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 651, + 416, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 416, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 416, + 665 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (92270114), the Agency for Science, Technology and Research (A\\*STAR) under MTC Programmatic Funds (Grant No.M23L7b0021), and NExT center." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14258" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Do, etc. Affordancenet: An end-to-end deep learning approach for object affordance detection. In ICRA, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 287, + 158 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440. Springer, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 202 + ], + "type": "text", + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19107-19117. IEEE, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 286, + 283 + ], + "type": "text", + "content": "[5] Dave Zhenyu Chen, Angel X. Chang, and Matthias Nießner. Scanrefer: 3d object localization in RGB-D scans using natural language. In ECCV, pages 202-221. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 327 + ], + "type": "text", + "content": "[6] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 373 + ], + "type": "text", + "content": "[7] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In CVPR, pages 1778-1787. Computer Vision Foundation / IEEE, 2021. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "text", + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In ACL, pages 4171–4186. Association for Computational Linguistics, 2019. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 464 + ], + "type": "text", + "content": "[9] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J. Lim. Demo2vec: Reasoning object affordances from online videos. In CVPR. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "text", + "content": "[10] Yiran Geng, Boshi An, Haoran Geng, Yuanpei Chen, Yaodong Yang, and Hao Dong. Rlafford: End-to-end affordance learning for robotic manipulation. In ICRA, pages 5880-5886. IEEE, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[11] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. Deberta: decoding-enhanced bert with disentangled attention. In ICLR. OpenReview.net, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 546, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 621 + ], + "type": "text", + "content": "[12] Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Tomas Jackson, Noah Brown, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner monologue: Embodied reasoning through planning with language models. In CoRL, pages 1769-1782. PMLR, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[13] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Kate-rina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In ECCV, pages 417-433. Springer, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Safoura Rezapour Lakani, Antonio Jose Rodríguez-Sánchez, and Justus H. Piater. Towards affordance detection for robot manipulation using affordance for parts and parts for affordance. pages 1155-1172, 2019. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Muchen Li and Leonid Sigal. Referring transformer: A one-step approach to multi-task visual grounding. In NeurIPS, pages 19652-19664, 2021. 6, 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 137 + ], + "type": "text", + "content": "[16] Yicong Li, Xun Yang, Xindi Shang, and Tat-Seng Chua. Interventional video relation detection. In ACM MM, pages 4091-4099, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "text", + "content": "[17] Yicong Li, Xiang Wang, Junbin Xiao, and Tat-Seng Chua. Equivariant and invariant grounding for video question answering. In ACM MM, pages 4714-4722, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[18] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In CVPR, pages 2928-2937, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "text", + "content": "[19] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and TatSeng Chua. Transformer-empowered invariant grounding for video question answering. TPAMI, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[20] Yicong Li, Junbin Xiao, Chun Feng, Xiang Wang, and TatSeng Chua. Discovering spatio-temporal rationales for video question answering. In ICCV, pages 13869-13878, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 273, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 314 + ], + "type": "text", + "content": "[21] Yicong Li, Xun Yang, An Zhang, Chun Feng, Xiang Wang, and Tat-Seng Chua. Redundancy-aware transformer for video question answering. In ACM MM, pages 3172-3180, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 347 + ], + "type": "text", + "content": "[22] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: generalized referring expression segmentation. In CVPR, pages 23592-23601. IEEE, 2023. 6, 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 405, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 405, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 405, + 545, + 447 + ], + "type": "text", + "content": "[24] Liangsheng Lu, Wei Zhai, Hongchen Luo, Yu Kang, and Yang Cao. Phrase-based affordance detection via cyclic bilateral interaction. IEEE Trans. Artif. Intell., 4(5):1186-1198, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[25] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In CVPR, pages 2242-2251. IEEE, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[26] Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In CVPR, pages 16433-16442. IEEE, 2022. 3, 6, 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[27] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual MLP framework. In ICLR. OpenReview.net, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 613 + ], + "type": "text", + "content": "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. SQA3D: situated question answering in 3d scenes. In ICLR. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 614, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 667 + ], + "type": "text", + "content": "[29] Jinpeng Mi, Hongzhuo Liang, Nikolaos Katsakis, Song Tang, Qingdu Li, Changshui Zhang, and Jianwei Zhang. Intention-related natural language grounding via object affordance detection and intention semantic extraction. Frontiers Neurorobotics, 14:26, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[30] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas J. Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In CoRL, pages 1666-1677. PMLR, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14259" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[31] Bogdan Moldovan, Plinio Moreno, Martijn van Otterlo, José Santos-Victor, and Luc De Raedt. Learning relational affordance models for robots in multi-object manipulation tasks. In 2012 IEEE International Conference on Robotics and Automation, pages 4373-4378, 2012. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 162 + ], + "type": "text", + "content": "[32] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In ICCV, pages 8687-8696. IEEE, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 225, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 225, + 175 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 225, + 175 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4 technical report, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "text", + "content": "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, pages 5099-5108, 2017. 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 274 + ], + "type": "text", + "content": "[35] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In ICRA, pages 11523-11530. IEEE, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "text", + "content": "[36] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M. Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. CoRR, abs/2212.04088, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 321, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 287, + 387 + ], + "type": "text", + "content": "[37] NN Sriram, Tirth Maniar, Jayaganesh Kalyanasundaram, Vineet Gandhi, Brojeshwar Bhowmick, and K Madhava Krishna. Talk to the vehicle: Language conditioned autonomous navigation of self driving cars. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5284-5290. IEEE, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 388, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 452 + ], + "type": "text", + "content": "[38] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. Mlp-mixer: An all-mlp architecture for vision. In NeurIPS, pages 24261-24272, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 455, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 287, + 497 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pages 5998-6008, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 500, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 287, + 532 + ], + "type": "text", + "content": "[40] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as queries for referring video object segmentation. In CVPR, pages 4964-4974. IEEE, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 287, + 566 + ], + "type": "text", + "content": "[41] Junbin Xiao, Angela Yao, Yicong Li, and Tat Seng Chua. Can i trust your answer? visually grounded video question answering. arXiv preprint arXiv:2309.01327, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 568, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 600 + ], + "type": "text", + "content": "[42] Amber Xie, Youngwoo Lee, Pieter Abbeel, and Stephen James. Language-conditioned path planning. In 7th Annual Conference on Robot Learning, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 287, + 634 + ], + "type": "text", + "content": "[43] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. CoRR, abs/2202.13519, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 636, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 636, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 636, + 287, + 678 + ], + "type": "text", + "content": "[44] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. CoRR, abs/2303.10437, 2023. 2, 3, 6, 7, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 713 + ], + "type": "text", + "content": "[45] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2908-2917. IEEE, 2021. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[46] Xue Zhao, Yang Cao, and Yu Kang. Object affordance detection with relationship-aware network. Neural Comput. Appl., 32(18):14321-14333, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14260" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_content_list.json b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..46e4994594eec18174d9a2a6602a0f57fb41f6bc --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_content_list.json @@ -0,0 +1,2059 @@ +[ + { + "type": "text", + "text": "LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network", + "text_level": 1, + "bbox": [ + 127, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Yang $^{1}$ , Liyuan Pan $^{1\\dagger}$ , Yan Yang $^{2}$ , Richard Hartley $^{2}$ , and Miaomiao Liu $^{2}$ $^{1}$ Beijing Institute of Technology \n $^{2}$ Australian National University", + "bbox": [ + 130, + 179, + 836, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hao.yang, liyuan.pan}@bit.edu.cn, {yan.yang, richard.hartley, miaomiao.liu}@anu.edu.au", + "bbox": [ + 91, + 219, + 870, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recovering sharp images from dual-pixel (DP) pairs with disparity-dependent blur is a challenging task. Existing blur map-based deblurring methods have demonstrated promising results. In this paper, we propose, to the best of our knowledge, the first framework that introduces the contrastive language-image pre-training framework (CLIP) to accurately estimate the blur map from a DP pair unsupervisedly. To achieve this, we first carefully design text prompts to enable CLIP to understand blur-related geometric prior knowledge from the DP pair. Then, we propose a format to input a stereo DP pair to CLIP without any fine-tuning, despite the fact that CLIP is pre-trained on monocular images. Given the estimated blur map, we introduce a blur-prior attention block, a blur-weighting loss, and a blur-aware loss to recover the all-in-focus image. Our method achieves state-of-the-art performance in extensive experiments (see Fig. 1).", + "bbox": [ + 75, + 316, + 473, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 604, + 209, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Defocus deblurring from a single image is an ill-posed problem [19, 28, 29, 33]. Compared to methods that restore an all-in-focus image directly [1, 2], blur map-based deblurring methods have demonstrated promising results [16, 26]. To simplify the blur removal process, several works [20, 34, 38] have utilized recent innovative dual-pixel (DP) sensors [6] to estimate blur maps, since the DP sensor captures a two-view image pair (a DP pair) in a single shot. The DP pair can be treated as images from a two-sample light field camera or a stereo system [24] with a tiny baseline [27, 34], which is beneficial for disparity estimation [1, 14, 23, 42]. Based on the DP image pair formation process, the disparity is geometrically related to the blur map, i.e., the degree of defocus blur proportional to disparity magnitudes [9].", + "bbox": [ + 75, + 631, + 468, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For estimating the blur map (defocus map or disparity", + "bbox": [ + 96, + 858, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4289276f10163e17c2dd9aa619e2d3d885516c5673317bc1d8f7f482698b7003.jpg", + "image_caption": [ + "Figure 1. Comparison of performance in terms of PSNR and Inference time. Our small and large models are respectively denoted by the Blue and Red cycles. The Green triangles represent previous state-of-the-art methods, and each of them is annotated with the method name, publication venue, and year. This figure is best viewed in color on the screen." + ], + "image_footnote": [], + "bbox": [ + 501, + 266, + 893, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "map), previous works either require extra data [3, 26], such as synthetic data, for supervision signals, or pre-calibrated blur kernels [34]. To avoid using extra information, several works explored unsupervised blur map estimation frameworks [16, 20]. Lee et al. [16] implicitly estimate the disparity of the left/right views using a spatial transformer, yet only in feature space, as well as Yang et al. [38]. Liang et al. [20] use fixed network branches for layer-wise defocus map estimation and deblurring, where the limited network branch number restricts the potential of its accuracy.", + "bbox": [ + 496, + 530, + 893, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the contrastive language-image pre-training framework (CLIP) [30] has accomplished extraordinary success for vision tasks, such as semantic segmentation [31, 37], object detection [32], and point cloud understanding [40]. Therefore, a question naturally arises, can we avoid the cost of collecting data and model designing, to estimate the blur map unsupervisedly by using CLIP? Unfortunately, using semantic knowledge from CLIP to handle low-level vision tasks is underexplored.", + "bbox": [ + 496, + 685, + 893, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, for the first time, we explore the relationship between language and blur map estimation by applying CLIP. Note that CLIP is trained to match a monocular image with its corresponding language descriptions. Given a blurred image, it is intuitive to match it with a corre", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 886, + 232, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24078", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sponding text [A blurry image.], which allows the language model to find the blurred regions in an image. However, it is hard to quantize the blur level of an image without knowing its associated latent sharp one [25].", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fortunately, we can leverage the capabilities of DP sensors. With DP pairs, blur quantization becomes possible without the need for its associated sharp image. The formation of DP ensures no disparity between left/right views for pixels in sharp regions, while a noticeable disparity exists for pixels in blurred regions. In essence, the detectable differences (disparities) in left/right views of a DP pair directly indicate the degree of defocus blurs. Consequently, we can quantify the variation between the left and right views, treating it as a valuable cue for estimating a blur map.", + "bbox": [ + 75, + 152, + 467, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, CLIP still finds it challenging to simultaneously explore stereo DP pairs with left/right view images. To address this issue, we propose to explore the DP pair and design an 'image and text' format to better capture the blurs in the image. Specifically, we propose to create a new image from the DP pair and translate the description about the 'blurriness' of the image to measure the 'symmetry' of the newly formed image, which is based on the DP image formation process. The freshly designed image-text format allows the CLIP to estimate the blur map easily (see Sec. 3.1).", + "bbox": [ + 75, + 303, + 467, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the estimated blur map, we then restore the all-in-focus image from the DP pair. By observing that the attention map of the self-attention mechanism [8] can be used in the dynamic DP deblurring process, we introduce a blur prior attention (BPA) that modulates the attention map with the estimated blur map which provides prior knowledge of deblurring kernels (see Sec. 3.2).", + "bbox": [ + 75, + 454, + 467, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To regulate the restoration, we specifically design two losses that are based on CLIP. Considering the network should focus more on heavily blurred regions, we propose a blur-weighting loss that imposes an adaptive penalty proportional to the blur degrees provided by the estimated blur map from CLIP. Then, to guarantee the restored image is sharp, we use CLIP to re-check the restored image and formulate a blur-aware loss (see Sec. 3.3).", + "bbox": [ + 75, + 560, + 467, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarised as follows:", + "bbox": [ + 96, + 681, + 398, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Language-driven DP (LDP) defocus deblurring framework, by exploring the potential of CLIP in the low-level vision task.", + "- We design an image-text format for blur map estimation based on the geometric relationship between the blur and disparity of the DP pair.", + "- We propose a blur prior attention (BPA) block, a blur-weighting loss, and a blur-aware loss to encourage sharp restorations of the DP pair." + ], + "bbox": [ + 76, + 695, + 467, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the proposed framework extensively on standard benchmark datasets. Though using the largescale foundation model, CLIP, for blur map estimation, our method achieves Pareto-optimality between deblurring performance and inference speed (refer to Fig. 1 for an", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "overview of the comparison).", + "bbox": [ + 500, + 90, + 694, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 119, + 640, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DP Defocus Deblurring. DP-based deblurring methods can be divided into single- and multi-task manners. For single-task based deblurring [1-3], methods apply different kernels at different regions of the DP pair to handle the spatially varying defocus deblur. This can usually be achieved by using parallel atrous convolutions [33], attention mechanisms [39], or deformable convolutions [17]. For multi-task deblurring methods, the blur/defocus/disparity map is auxiliarily estimated to benefit DP defocus deblurring. Pan et al. [26] and Xin et al. [34] use the estimated DP disparity to regularise the restoration through reblurring; however, with extra data or pre-calibrate kernel. Several works explored unsupervised defocus map (or disparity) estimation. Lee et al. [16] and Yang et al. [38] estimate defocus disparity only in feature space without explicitly showing a defocus map. Liang et al. [20] optimize a multi-channel defocus map based on [27]. However, hyperparameters, such as the defocus threshold and the number of channels, are preset, limiting blur map estimation accuracy. Unlike past works, by transferring the blur-related geometric prior knowledge into semantic prompts, we unsupervisedly estimate the blur map from CLIP to benefit defocus deblurring.", + "bbox": [ + 496, + 146, + 890, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-driven Vision. Common objectives in visual-language learning are aligning the text with a paired image in the embedding space [21, 30, 35]. The CLIP, one of the milestone achievement in the area, demonstrate that language-independent computer vision tasks can be substantially benefited from visual-language learning, by achieving SOTA zero-shot classification performance in the ImageNet-1K benchmark [7]. Successive works have extended the zero-shot learning capability of CLIP to semantic segmentation [31, 44], monocular depth estimation [4, 10, 41], and point clouds [11, 40]. However, none of the existing works explores the zero-shot learning capability of CLIP in stereo/DP pairs. This paper proposes the first approach to use the monocular image-trained CLIP in zero-shot blur map segmentation of DP pairs.", + "bbox": [ + 496, + 481, + 890, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 723, + 589, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Let us now introduce our language-driven dual-pixel image defocus deblurring framework. Given a defocus blurred DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ , we aim to restore an all-in-focus image $\\mathbf{I}$ where $\\mathbf{B}_{\\mathrm{L}}$ and $\\mathbf{B}_{\\mathrm{R}}$ are the left and right view of a DP pair and $\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}},\\mathbf{I}\\in \\mathbb{R}^{h\\times w\\times 3}$ . Here, $h$ and $w$ define the image height and width, respectively.", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first design a DP-aware prompt to adapt CLIP to output a blur map, $\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times 1}$ , which quantizes disparity unsupervisedly between $\\mathbf{B}_{\\mathrm{L}}$ and $\\mathbf{B}_{\\mathrm{R}}$ . Then, we propose a BPA block based network which employs the blur", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24079", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/990fe983387e39d2f8b277fcf1f9eba3549aaf603c220bf2942e5ea3d19158d0.jpg", + "image_caption": [ + "Figure 2. Overall architecture. Given a DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ , we first generate a blur map $\\mathbf{M}$ by using our unsupervised blur map estimation strategy which is based on the CLIP with assembling format. We then forward the DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ to our deblurring backbone composed of an encoder and a decoder. Meanwhile, we have multiple BPA blocks, using the blur map $\\mathbf{M}$ as deblurring kernel prior to refine the intermediate backbone feature embeddings. In the following, we have the restoration $\\hat{\\mathbf{I}}$ . The deblurring network is optimized with three losses, charbonnier loss $\\mathcal{L}_{char}$ , blur weighting loss $\\mathcal{L}_{bwl}$ , and blur aware loss $\\mathcal{L}_{bal}$ . $\\mathcal{L}_{char}$ penalizes deviations from $\\mathbf{I}$ in the resorption $\\hat{\\mathbf{I}}$ . $\\mathcal{L}_{bwl}$ and $\\mathcal{L}_{bal}$ regularize the restoration by imposing an adaptive penalty on the blurred regions of the DP pair, and leveraging unsupervised blur map estimation strategy to encourage the restoration to be sharp, i.e., the blur map $\\hat{\\mathbf{M}}$ for the restoration $\\hat{\\mathbf{I}}$ is zero value filled." + ], + "image_footnote": [], + "bbox": [ + 140, + 87, + 831, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "map $\\mathbf{M}$ as deblurring priors to restore a sharp image $\\hat{\\mathbf{I}}$ from $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ . To regularize the restoration of $\\hat{\\mathbf{I}}$ , a blur-weighting loss $\\mathcal{L}_{\\mathrm{bwl}}$ and a blur-aware loss $\\mathcal{L}_{\\mathrm{bal}}$ are also introduced that were underpinned by our unsupervised blur map estimation strategy. The overall framework is shown in Fig. 2. Details are presented below.", + "bbox": [ + 75, + 363, + 472, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unsupervised Blur Map Estimation", + "text_level": 1, + "bbox": [ + 76, + 462, + 387, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. CLIP respectively embeds an image and a text description into a common semantic space of dimension $c$ with an image encoder and a text encoder, obtaining image embedding $\\mathbf{f} \\in \\mathbb{R}^c$ and text embedding $\\mathbf{t} \\in \\mathbb{R}^c$ . By optimizing a contrastive loss, the embeddings of the aligned image and text pair, $\\Omega_a$ , stay similar, i.e., maximizing the semantic similarity by $\\max_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_a} \\sin(\\mathbf{f}, \\mathbf{t})$ , where $\\sin(\\cdot, \\cdot)$ is the cosine similarity metric. Meanwhile, the semantics of misaligned pairs, $\\Omega_{na}$ are pushed away from each other, i.e., $\\min_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_{na}} \\sin(\\mathbf{f}, \\mathbf{t})$ . With comprehensive training data, CLIP can be leveraged to align images with pre-defined prompt templates composed of open-world text descriptions for zero-shot learning [30].", + "bbox": [ + 75, + 486, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a DP pair, the disparity $d$ for a pixel in the left view is defined by its location $(i,j)$ and that of its corresponding pixel at $(i,k)$ in the right view.", + "bbox": [ + 75, + 681, + 468, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the literature, this DP disparity [27] is defined as", + "bbox": [ + 94, + 728, + 437, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd = j - k = A (r (z)) \\cdot B (z), \\mathbf {B} _ {\\mathrm {L}} (i, j) = \\mathbf {B} _ {\\mathrm {R}} (i, k), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 752, + 468, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_{\\mathrm{L}}(i,j)$ and $\\mathbf{B}_{\\mathrm{R}}(i,k)$ are the intensity of the two matching points, $r(\\cdot)$ is a function of the radius of the circle of confusion, $z$ is the point depth, and $A(r(z))$ and $B(z)$ are lens-dependent constants [2, 9, 26, 34].", + "bbox": [ + 75, + 779, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis. Leveraging the CLIP, zero-shot blur map estimation can be generally divided into three steps. i) Image Encoding. Considering the image encoder of the CLIP is trained with monocular RGB images, a function $\\eta (\\cdot ,\\cdot)$ is", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "required to create a new monocular image from the DP image pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ . To obtain the dense embedding for the blur map estimation task, we follow [41] to pop out the final pooling layer from the image encoder, yielding an embedding matrix $\\mathbf{F}\\in \\mathbb{R}^{h_s\\times w_s\\times c}$ , where $h_s$ and $w_{s}$ are the height and width for embedding of the newly created image, respectively. Mathematically, we have", + "bbox": [ + 496, + 363, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\operatorname {I m a g e E n c o d e r} \\left(\\eta \\left(\\mathbf {B} _ {\\mathrm {L}}, \\mathbf {B} _ {\\mathrm {R}}\\right)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 484, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each pixel of $\\mathbf{F}$ has encoded the local semantic information for a corresponding region in $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ . ii) Text Encoding. A text encoder is used to embed the semantics of a prompt p describing the blur into the text embedding,", + "bbox": [ + 496, + 516, + 893, + 578 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} = \\operatorname {T e x t E n c o d e r} (\\mathfrak {p}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 593, + 890, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "iii) Blur map estimation. Let $\\mathbf{F}(i,j)$ be a c-dimensional embedding at position $(i,j)$ of $\\mathbf{F}$ , $i\\in \\{1,\\dots ,\\mathrm{h}_{\\mathrm{s}}\\}$ , and $j\\in \\{1,\\dots ,\\mathrm{w}_{\\mathrm{s}}\\}$ . The embedding $\\mathbf{F}(i,j)$ and the blur semantic-based text embedding $\\mathbf{t}$ are expected to be similar if the pixel at $(i,j)$ is blurred. Therefore, the blur mask $\\mathbf{M}_{\\mathbf{F}}$ of $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ is obtained by", + "bbox": [ + 496, + 625, + 893, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\mathbf {F}} (i, j) = \\sigma (\\operatorname {s i m} (\\mathbf {F} (i, j), \\mathbf {t})) . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 729, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note, $\\sigma (\\cdot)$ normalizes the cosine similarity $\\mathfrak{sim}(\\cdot ,\\cdot)$ output to a probability space, to reflect the degree of blur at the position. Here, $\\mathbf{M}_{\\mathbf{F}}$ is sampled to have the same resolution as $\\mathbf{B}_{\\mathrm{L}}$ or $\\mathbf{B}_{\\mathrm{R}}$ , returning $\\mathbf{M}\\in \\mathbb{R}^{h\\times w\\times 1}$ .", + "bbox": [ + 496, + 762, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the analysis above, our goal is to develop text prompts $\\mathfrak{p}$ , while adapting the function $\\eta(\\cdot, \\cdot)$ , to enable the use of semantic information from CLIP to generate the blur map $\\mathbf{M}$ in an unsupervised manner. In particular, we explore the blur-aware, DP-aware, and ensembling formats with detailed presented below.", + "bbox": [ + 496, + 824, + 890, + 915 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24080", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18d45b8a20b3eaa4186ec61cbe6d094c70506c392e53a5046bae28f16976d1a4.jpg", + "image_caption": [ + "Figure 3. The pipeline of our DP-aware format for obtaining the blur map $\\mathbf{M}$ . Given a DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ , the function $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})$ concatenates the left view $\\mathbf{B}_{\\mathrm{L}}$ and with horizontally flipped right view $\\mathbf{B}_{\\mathrm{R}}^{h}$ for obtaining the $\\mathbf{F}$ from the CLIP image encoder. We then apply $\\alpha (F)$ to averagely pool the feature map according to the transformation imposed on the input DP pair, resulting in a new feature embedding with the same aspect ratio of the input DP pair. Meanwhile, we set the prompts $\\mathfrak{p}$ to describe the symmetry of $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})$ , obtaining text embedding $\\mathbf{t}$ . We finally have the blur map $\\mathbf{M}$ by computing Eq. 6 and performing sampling." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 462, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blur-aware Format. An intuitive approach is setting a prompt $\\mathbf{p}$ that describes the blur to query a blur map $\\mathbf{M}$ , e.g., $\\mathbf{p} = [\\mathrm{A~blurry~image}.]$ . The $\\eta(\\cdot, \\cdot)$ is then defined to synthesize a center view of the DP pair by averaging the left view $\\mathbf{B}_{\\mathrm{L}}$ and right view $\\mathbf{B}_{\\mathrm{R}}$ , as the average of two sub-images is equivalent to the image captured by a regular sensor [3, 27]. However, the disparity information among the DP pair is neglected, resulting in a sub-optimal blur map. Refer to Fig. 4 for comparisons.", + "bbox": [ + 75, + 444, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "DP-aware Format. Based on the DP image formation process, the disparity is proportional to the degree of blur exhibited in the DP view [1]. We thus propose to guide CLIP to understand the disparity exhibited between $\\mathbf{B}_{\\mathrm{L}}$ and $\\mathbf{B}_{\\mathrm{R}}$ which eventually leads to better mask (M) generation. To this end, we first formulate the disparity in the DP pair and then study the disparity-aware prompt and function $\\eta (\\cdot ,\\cdot)$ designing. Such design leads us to query degrees of displacement between the left and right view from the CLIP, without any fine-tuning, to obtain the blur map M.", + "bbox": [ + 75, + 584, + 468, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the CLIP is the vision-text alignment model trained with monocular image input. To extract the displacement between left/right views of the DP pair, we explore several input-prompt formats: i) Concatenating the left and right views horizontally, and prompting their difference, e.g., [The left and right of the image are different.]; ii) Prompting the horizontal symmetry between the left and right views, e.g., [A horizontally symmetric image.] (Fig. 3).", + "bbox": [ + 75, + 736, + 468, + 872 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Empirically, we achieve the best performance with the second prompt format. When describing the symmetry in", + "bbox": [ + 76, + 872, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/36370877f7d879fc524e4ff8f9994de1b61705a140a2a3adf717c1156d906117.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 596, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2adfbde67b3773aad6a74b04d1253cf65b912abc73088b937b9bfadbd6a43017.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 598, + 88, + 694, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b0872cd650385ca4ec8a5634d760a6ffa4998117d1739009563c34e8098ffa4a.jpg", + "image_caption": [ + "(c)", + "Figure 4. Example of blur map estimation (the darker, the sharper). (a) The left view of the DP image. (b) and (c) Blur maps obtained with prompts of the blur-aware and DP-aware formats. (d) The blur maps obtained with the ensembling-based format, i.e., all eight blur-aware and DP-aware prompts are used (please refer to Tab. 4 for details)." + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 790, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3b01f240f5b9ae389abd9aee3a971c85e3eec945b179276d04374353f0970329.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 790, + 88, + 887, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the prompt $\\mathfrak{p}$ , we concatenate the left view $\\mathbf{B}_{\\mathrm{L}}$ with a horizontally flipped right view $\\mathbf{B}_{\\mathrm{R}}^{h} = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})$ along the width axial in the $\\eta (\\cdot ,\\cdot)$ . We then obtained image features $\\mathbf{F} = \\mathrm{ImageEncoder}\\bigl (\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\bigr)$ . To transform $\\mathbf{F}$ to the same aspect ratio of $\\mathbf{B}_{\\mathrm{L}}$ and $\\mathbf{B}_{\\mathrm{R}}$ , we average the corresponding pixel-wise features based on the transformation in $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})$ and have $\\alpha (\\mathbf{F})$ . For $i\\in \\{1,\\dots ,\\mathsf{h}_{\\mathsf{w}}\\}$ and $j\\in \\{1,\\dots ,\\lfloor \\frac{1}{2}\\mathsf{w_s}\\rfloor \\}$ , we define", + "bbox": [ + 496, + 300, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha (\\mathbf {F}) (i, j) = \\frac {1}{2} \\mathbf {F} (i, j) + \\frac {1}{2} \\mathbf {F} (i, w _ {s} - j). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 429, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Back to the left-hand side of Eq. 4, we have", + "bbox": [ + 500, + 468, + 792, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {\\mathbf {F}} (i, j) = 1 - \\mathbf {M} _ {\\mathbf {F}} ^ {s} (i, j) = 1 - \\sigma \\left(\\operatorname {s i m} (\\alpha (\\mathbf {F}) (i, j), \\mathbf {t} ^ {s})\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 494, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{t}^s$ is the text embedding of the prompt describing symmetry. Note that the high similarity value, i.e. $\\mathbf{M}_{\\mathbf{F}}^{s}(i,j) = \\sigma \\big(\\mathrm{sim}(\\alpha (\\mathbf{F})(i,j),\\mathbf{t}^{s})\\big)$ indicates pixels in sharp regions. Therefore the blur map is derived as $1 - \\mathbf{M}_{\\mathbf{F}}^{s}(i,j)$ in Eq. 6. Similarly, the $\\mathbf{M}_{\\mathbf{F}}$ is sampled to the same resolution of the DP pair, returning M. For this blur map estimation strategy, we conclude the lemma below,", + "bbox": [ + 496, + 539, + 890, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 3.1. A pixel is sharp if and only if it is horizontally symmetric in the left view and the horizontally flipped right view, otherwise, the pixel is blurred.", + "bbox": [ + 496, + 648, + 893, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Following Eq. (1), we have $\\mathbf{B}_{\\mathrm{L}}(i,j)$ being sharp, given $\\mathbf{B}_{\\mathrm{L}}(i,j) = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})(i,\\mathsf{w} - j)$ . In other words, we have $\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,j) = \\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,\\mathsf{w}_{\\mathsf{s}} - j)$ , demonstrating the horizontal symmetry. Reversely, the horizontal symmetry breaks once the disparity/blur occurs.", + "bbox": [ + 496, + 715, + 890, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ensembling format. Given the analysis that the degradations in the DP pair can be described by both the blurriness and symmetricity, we thus propose to obtain the mask by assembling the blur-aware and DP-aware formats jointly. Please refer to Tab. 4 for details of the 8 prompts designed from blur-aware and DP-aware formats.", + "bbox": [ + 496, + 809, + 893, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24081", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network", + "text_level": 1, + "bbox": [ + 76, + 90, + 179, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train an encoder-decoder based deblurring network for restoring the DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ into a sharp image $\\hat{\\mathbf{I}}$ with the derived blur map $\\mathbf{M}$ . Our network consists of a backbone network based on [5] and a BPA block to interleave with the backbone network layers. In particular, the BPA block uses the obtained blur map $\\mathbf{M}$ as deblur kernel prior to help the restoration of $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ on the fly.", + "bbox": [ + 75, + 113, + 468, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-attention. The self-attention (SA) mechanism is widely used in the image restoration tasks [18, 36, 39]. We first briefly review it. Let $\\mathbf{X} \\in \\mathbb{R}^{h_x \\times w_x \\times c_x}$ be the intermediate feature embedding of a blur image from the backbone encoder, where $h_x$ , $w_x$ , $c_x$ are height, width, and channel of the feature. The self-attention is input with query $\\mathbf{Q}$ , key $\\mathbf{K}$ , and value $\\mathbf{V}$ that are linearly projected from $\\mathbf{X}$ , and then calculate an attention map between $\\mathbf{Q}$ and $\\mathbf{K}$ for adaptively aggregating $\\mathbf{V}$ . This process is defined as follows,", + "bbox": [ + 75, + 220, + 470, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {Q}}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {K}}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {V}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 369, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention $(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta)\\mathbf{V}$ (8)", + "bbox": [ + 114, + 388, + 468, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{h_{\\mathbf{x}} \\times w_{\\mathbf{x}} \\times c_{\\mathbf{x}}}$ are learnable weight matrices, $\\beta$ is a normalization factor that is typically set to the square root of the latent dimension, and $\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta) \\in \\mathbb{R}^{(h_{\\mathbf{x}} \\times w_{\\mathbf{x}}) \\times (h_{\\mathbf{x}} \\times w_{\\mathbf{x}})}$ is attention map.", + "bbox": [ + 75, + 420, + 468, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BPA Block. The attention map can be considered as an adaptive deblur kernel prior for restoring the DP pair in the feature space. We are motivated to modulate the attention map by supplying the blur map $\\mathbf{M}$ , utilizing blur-related prior knowledge to additionally assign a $\\mathbf{q} \\times \\mathbf{q}$ deblur kernel for each region in $\\mathbf{X}$ to aggregate a sharp feature embedding from its neighboring regions. We re-write the Eq. 8 as", + "bbox": [ + 75, + 484, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention $(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta +\\mathbf{O})\\mathbf{V}$ (9)", + "bbox": [ + 91, + 601, + 468, + 618 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {O} = \\operatorname {P a d} \\left(\\operatorname {F F N} _ {\\operatorname {C o n v}} \\left(\\sigma^ {- 1} (\\mathbf {M})\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 621, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and show that the restoration result is further improved in Fig. 5, where $\\mathrm{FFN}_{\\mathrm{Conv}}(\\cdot)$ is a convolution-based feed forward layer that is used to learn a $q \\times q$ deblur kernel for each region of $\\mathbf{X}$ from the unnormalized blur map $\\sigma^{-1}(\\mathbf{M})$ , resulting in $\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M})) \\in \\mathbb{R}^{h_x \\times w_x \\times q^2}$ . The Pad( $\\cdot$ ) function reshapes $\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))$ into $\\mathbb{R}^{h_x \\times w_x \\times q \\times q}$ , and pads the reshaped $\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))$ to $\\mathbb{R}^{(h_x \\times w_x) \\times (h_x \\times w_x)}$ , for shape alignment issue.", + "bbox": [ + 75, + 651, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Network Training", + "text_level": 1, + "bbox": [ + 76, + 784, + 251, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimize our deblurring network with the widely used Charbonnier loss. Moreover, we propose a blur-aware loss and a blur weighting loss to penalize the blurred regions in the restoration $\\hat{\\mathbf{I}}$ .", + "bbox": [ + 75, + 806, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Charbonnier loss. We encourage the restoration to be close to the ground truth by using Charbonnier loss [12, 43].", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/480aad7501ee7e8b7fcb54754c39b4d78d088b346ac26c3d6c0e4798592d9d0c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 570, + 142 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/df89236f0ba3d0a983886e86e0fac8dc150f9b76ee15aaf584879ba6aa82d718.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 571, + 88, + 692, + 142 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8838b231c703c986bc4d3b093dad7eb2534b56e2645f464ec2a5a15cd603a2de.jpg", + "image_caption": [ + "(c)", + "Figure 5. Example of deblurring with the blur map projected deblur kernel $\\mathbf{O}$ . (a) and (b) are the left view of the DP pair and the ground truth sharp image. (c) and (d) are deblurred images with and without using $\\mathbf{O}$ . Zoom in for better quality." + ], + "image_footnote": [], + "bbox": [ + 694, + 88, + 761, + 142 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/53c8c1fc807ced4c5c6267718200456ec40493e492d5f7a61f420f3dd08a2a77.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 792, + 88, + 885, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically set $\\varepsilon$ to $10^{-4}$ , and define", + "bbox": [ + 500, + 231, + 769, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\| ^ {2} + \\varepsilon^ {2}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 257, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blur Weighting Loss. We impose an adaptive penalty on the blurred regions of the DP pair $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ , considering the degree of blur measured by CLIP. It encourages the network to focus on the restoration of heavily blurred regions. We first normalize the deviations between $\\hat{\\mathbf{I}}$ and $\\mathbf{I}$ by the deviations between $(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})$ and $\\mathbf{I}$ , and then re-use the blur map to select blurred regions residing in the DP pair.", + "bbox": [ + 498, + 291, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {b w l}} = \\frac {\\left\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\right\\| ^ {2}}{\\left\\| \\hat {\\mathbf {I}} (i , j) - \\frac {1}{2} \\left(\\mathbf {B} _ {\\mathrm {L}} (i , j) + \\mathbf {B} _ {\\mathrm {R}} (i , j)\\right) \\right\\| ^ {2}} \\cdot \\mathbf {M} (i, j), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 407, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use the blur map $\\mathbf{M}$ (with soft logits) to reflect the blur levels detected by CLIP.", + "bbox": [ + 498, + 470, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Blur-aware Loss. We encourage $\\hat{\\mathbf{I}}$ to be sharp by leveraging the versatile knowledge contained in the CLIP. We generate the blur map $\\hat{\\mathbf{M}}$ by the blur-aware format for the restoration, and penalize the blur exhibited in $\\hat{\\mathbf{I}}$ . Similar to $\\mathcal{L}_{\\mathrm{bw1}}$ , we use the unnormalized version of the blur map,", + "bbox": [ + 498, + 501, + 890, + 577 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b a l}} = \\hat {\\mathbf {M}} (i, j). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 587, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This encourages $\\hat{\\mathbf{M}}$ to be zero, i.e., no blur exists in the restoration.", + "bbox": [ + 498, + 617, + 888, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss. Our network is optimized by minimizing the following loss function", + "bbox": [ + 498, + 648, + 888, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{| \\Omega |} \\sum_ {(i, j) \\in \\Omega} \\left(\\mathcal {L} _ {\\text {c h a r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {b a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {b w l}}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 688, + 890, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\frac{1}{|\\Omega|}$ is a normalization factor determined by the image domain size $|\\Omega|$ , and $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters to balance losses.", + "bbox": [ + 498, + 738, + 890, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Analysis", + "text_level": 1, + "bbox": [ + 500, + 800, + 743, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implement our model under the PyTorch framework. We use the pre-trained and ViT-B/32 [8] architecture-based CLIP [30] for our blur map estimation strategy. We train the deblurring network in two stages for $6 \\times 10^{6}$ steps in total with the Adam optimizer.", + "bbox": [ + 498, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24082", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e6eb4fa16d9dee32b9d8049eeb7788574651949f1c84b8e2ebcaa4d8e7feefee.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on the DPD-blur dataset [1] composed of 37 indoor and 39 outdoor scenes. We respectively color the best and the second-best performance in red and blue." + ], + "table_footnote": [], + "table_body": "
MethodOutdoor ScenesIndoor ScenesAverage
PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓LPIPS↓
EBDB25.770.7720.0400.05121.250.5990.0580.08623.450.6830.0490.0670.336
DMENet25.500.7880.0380.05321.430.6440.0630.08423.410.7140.0510.0670.349
DPDNet27.480.8490.0290.04222.900.7260.0520.07125.130.7860.0410.0550.223
RDPD28.100.8430.0270.03922.820.7040.0530.07225.390.7720.0400.0530.255
DDDNet27.570.8330.0300.04123.280.7080.0500.06825.360.7680.0410.0540.197
IFAN28.660.8680.0250.03723.460.7430.0490.06725.990.8040.0370.0500.207
BAMBNet28.600.8720.0260.03724.300.7720.0450.06026.400.8210.0360.0470.141
DeepRFT28.480.8700.0250.03823.090.7360.0510.07025.710.8010.0370.0510.217
Restormer29.480.8950.0230.03423.970.7730.0470.06326.660.8330.0350.0460.155
K3DN29.870.8900.0220.03224.370.7800.0460.06027.060.8350.0340.0440.187
Ours (Small)29.550.8930.0210.03324.380.7730.0430.06526.910.8310.0320.0450.185
Ours (Large)29.880.8960.0200.03224.730.7860.0370.05927.240.8400.0280.0430.168
", + "bbox": [ + 76, + 116, + 898, + 338 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af59d58c1dfebd5a9ce8d718ea1d57d049685ba95a85e899be9a487dd5dbf731.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons on DDD-syn dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑MAE↓RMSE_rel↓
EBDB26.480.6830.0360.047
DMENet30.140.9390.0170.031
DPDNet31.450.9260.0160.027
RDPD32.420.9120.0160.024
DDDNet33.210.9560.0100.022
IFAN34.180.9290.0120.020
BAMBNet35.900.9540.0120.016
DeepRFT36.520.9520.0100.015
Restormer36.440.9570.0100.015
K3DN38.230.9650.0080.012
Ours (Small)37.270.9590.0100.014
Ours (Large)38.620.9710.0080.012
", + "bbox": [ + 78, + 356, + 475, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In each stage, the learning rate is decayed from $2 \\times 10^{-4}$ to $10^{-6}$ by a cosine annealing strategy. In the first stage, we train the network for $5 \\times 10^{6}$ steps, and only use the Charbonnier loss $\\mathcal{L}_{\\mathrm{char}}$ . In the second stage, the network is finetuned for $1 \\times 10^{6}$ steps with $\\mathcal{L}_{\\mathrm{total}}$ , where we have $\\lambda_{1} = 10^{-1}$ and $\\lambda_{2} = 2 \\times 10^{-1}$ . Code, results, and more details are given in our project page.", + "bbox": [ + 75, + 571, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Our method is evaluated on two real-world DP defocus deblurring datasets, DPD-blur dataset [1] and DPD-disp dataset [27]. We follow [1] on the training and testing splits. For the DPD-disp dataset, we use it for testing only, re-using the checkpoints trained from the DPD-blur dataset as they do not provide ground-truth (GT) data or with limited dataset size. Moreover, two synthetic datasets, DDD-syn [26] and RDPD [2], are also used for evaluations.", + "bbox": [ + 75, + 681, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean absolute error (MAE), and root mean squared relative error (RMSE_ref) as our evaluation metrics.", + "bbox": [ + 76, + 806, + 468, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. We compare with SOTA DP defocus deblurring methods, i.e., EBDB [13], DMENet [15],", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0b0fcfdd19a74488061f44a8fa8e94ab12cff7720ace3d1c6d48eeffabd63eff.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons on RDPD dataset." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑MAE↓RMSE rel↓
EBDB24.230.6370.0310.054
DMENet25.170.7310.0360.061
DPDNet29.840.8280.0320.055
RDPD31.090.8610.0250.028
DDDNet30.140.8400.0160.031
IFAN31.120.8650.0230.028
BAMBNet31.780.8670.0160.026
DeepRFT31.710.8790.0160.024
Restormer32.270.8710.0160.024
K3DN32.840.9050.0140.023
Ours (Small)32.980.8950.0140.022
Ours (Large)33.440.9070.0130.021
", + "bbox": [ + 500, + 356, + 893, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DPDNet [1], RDPD [2], DDDNet [26], IFAN [16], BAMBNet [20], DeepRFT [22], Restormer [39], and K3DN [38].", + "bbox": [ + 500, + 571, + 890, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparisons with State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 614, + 803, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We compare with the SOTA methods on the DPD-blur, DDD-syn, and RDPD datasets in Tab. 1, Tab. 2, and Tab. 3, respectively. Our observations are as follows: i) With our model (Small), we achieve similar restoration results with Restormer and K3DN, the third-best and second-best method, in all datasets. ii) When scaling to the large model, we consistently outperform all methods on all evaluation metrics. For example, our method, K3DN, and Restormer respectively achieve 27.24/27.06/26.66, 38.62/38.23/36.44, and 33.44/32.84/32.27 PSNR (dB) in the DPD-blur, DDD-syn, and RDPD datasets. Meanwhile, on a single RTX 4090 GPU, we remain to have a lower Infer. time (including CLIP) which is 0.042s and 0.054s for our small and large models, while K3DN/Restormer takes 0.198s/0.255s.", + "bbox": [ + 496, + 638, + 890, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. We compare with state-of-the-art methods on the DPD-blur dataset (Fig. 6). As shown, our", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24083", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a5e452037bfcf471d7ed6ed97b64d00f80d482743a1e3ad6d7b9d34937e419a.jpg", + "image_caption": [ + "Blurry Image" + ], + "image_footnote": [], + "bbox": [ + 168, + 88, + 401, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a577a1e892a8fb4e3e7924fa7e1e060664c57fcf2ca9b0ed80042cb1be77ba2.jpg", + "image_caption": [ + "GT Image" + ], + "image_footnote": [], + "bbox": [ + 405, + 88, + 500, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e95c219927e2e91924f97c7242859158592d75f84a6394d22288eba860ab7aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 599, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e263042cb2f4553c514700a14933a30b450d50b806f7a7b80439df6f943c6067.jpg", + "image_caption": [ + "IFAN [16]" + ], + "image_footnote": [], + "bbox": [ + 604, + 89, + 697, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ceb81e56acb118fae69918d30d66a3cef1c1de4338a9c09c30cab0d81d1ec79e.jpg", + "image_caption": [ + "DDDNet [26]" + ], + "image_footnote": [], + "bbox": [ + 702, + 89, + 797, + 138 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/593fa7f0cd33870b2370f7686d3d006831fff4436a047c349f44ed5ff953fbc2.jpg", + "image_caption": [ + "Blurry Image" + ], + "image_footnote": [], + "bbox": [ + 406, + 162, + 500, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71419781ec3eb04f150dd9a5e0d695549ad8f71fac20fec5db2020fdef535aa2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 162, + 599, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ab9e10670eae6a5762ffcdbd1f94d11fb463e0508bfb974ed139b904a3febde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 162, + 697, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/415d9511d004b84ab0fbe979d1a9d97b307e976b9d99d47209ccbd86eda5a141.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 162, + 797, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a38e83a781b12c8e4eee01d3784d2ce81064db85b5dc3eb35f35492a454af57f.jpg", + "image_caption": [ + "DeepRFT [22]", + "Blurry Image", + "DeepRFT [22]" + ], + "image_footnote": [], + "bbox": [ + 166, + 234, + 401, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d00923ce4f24847248c6927585b5be800a6d3253de2729faa764a9b0ad1da0f9.jpg", + "image_caption": [ + "GT Image" + ], + "image_footnote": [], + "bbox": [ + 406, + 234, + 500, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10c56ea97f9d68316bbdc7475b929b6f2caba95210b120acf0a3e447d2f67f52.jpg", + "image_caption": [ + "Restormer [39]", + "Blurry Image" + ], + "image_footnote": [], + "bbox": [ + 504, + 234, + 599, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1cba08fd1c437d5faf95cc4c61b001bbfec7982372d7fff8b5e1e2745bb5b7f.jpg", + "image_caption": [ + "K3DN [38]", + "IFAN [16]" + ], + "image_footnote": [], + "bbox": [ + 604, + 234, + 697, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3deff5e3099745779cb9d782a3227a6d95e4b9a3129e3a16d70ece208f2f3ad0.jpg", + "image_caption": [ + "Ours", + "DDDNet [26]" + ], + "image_footnote": [], + "bbox": [ + 702, + 234, + 797, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6d883e27feceebacbcd4f766c2154c76680ba300f69277ab319b309a3c8e21d.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on the DPD-blur dataset [1]. We present the blurry image $(\\mathbf{B}_{\\mathrm{L}})$ in the first column, and the ground truth (GT) image of regions residing in the red bounding box is shown in the second column. The cropped blurred image of the red bounding box region is in the third column." + ], + "image_footnote": [], + "bbox": [ + 406, + 308, + 500, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e2e56ebeb48ea5db23dc965bdb40ba0b1498af169039040fd80561846390f0d.jpg", + "image_caption": [ + "Restormer [39]" + ], + "image_footnote": [], + "bbox": [ + 504, + 308, + 599, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3ae7e7d550322b212db7574819d25fe9d264c840e13caffee136fba8e2865e6.jpg", + "image_caption": [ + "K3DN [38]" + ], + "image_footnote": [], + "bbox": [ + 604, + 308, + 697, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fe1437ecf8035033a05d2376d9f62d1fd4e4038bc29755523e7b1a9b1a2cddc.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 702, + 308, + 797, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method performs sharper and clearer restorations of edges and texts. To further demonstrate the benefits of using CLIP to generate the blur mask unsupervisedly in our approach, we provide an example in Fig. 7, which preserves sharp regions (e.g., text 'University') in the input blur image effectively compared to the second-best baseline.", + "bbox": [ + 75, + 434, + 468, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49d024ddac1a93a08d0f3ae7a68f2ae2039e91b0ac1850f8d2dcdedd2fdd5610.jpg", + "image_caption": [ + "(a) Input", + "Figure 7. An example of the restoration results on the sharp region of the input DP image (left view)." + ], + "image_footnote": [], + "bbox": [ + 81, + 529, + 176, + 569 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e3c81256f7e09608a955ef70e91d56ce26e088b522210737345b94a264481924.jpg", + "image_caption": [ + "(b) K3DN" + ], + "image_footnote": [], + "bbox": [ + 179, + 529, + 272, + 569 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbad7c91f69532a328961c39f21d686aa4cff8de248434c3a6ea5e0377a8515b.jpg", + "image_caption": [ + "(c) Restormer" + ], + "image_footnote": [], + "bbox": [ + 274, + 529, + 369, + 569 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d412d6a06994209392d2827b7c200204608fe6290bf8376b7f031b3173930e4.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 370, + 529, + 465, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 619, + 228, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study model architecture, blur map estimation, and losses in this section. All experiments are performed in the DPD-blur dataset.", + "bbox": [ + 75, + 643, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Model Architecture. We study the effective usage of blur map as prior knowledge for DP defocus deblurring (Tab. 6). To show the benefits of using blur maps, we consider four baselines: i) NAFNet [5], the baseline model; ii) Concat, concatenating the blur map with the DP pair before feeding to the model; iii) SA, using the self-attention layer [8] in Eq. 8 to the model only; iv) Concat+SA, combines the Concat and SA settings; v) BPA block, using kernel priors from the blur map to assist deblurring. Our model consistently outperforms all baselines, which indicates the effectiveness of using the blur map.", + "bbox": [ + 75, + 689, + 468, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Losses. We investigate the proposed blur weighting loss $\\mathcal{L}_{\\mathrm{bw}}$ and blur-aware loss $\\mathcal{L}_{\\mathrm{bal}}$ (Tab. 7) with a baseline as the $\\mathcal{L}_{\\mathrm{char}}$ setting. We first use $\\mathcal{L}_{\\mathrm{bw}}$ and $\\mathcal{L}_{\\mathrm{bal}}$", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to our baseline, NAFNet (i.e., not using our BPA block), that only uses $\\mathcal{L}_{\\mathrm{char}}$ , having 0.32 PSNR (dB) improvements. Secondly, we train our model (backbone + BPA block) with Charbonnier loss, resulting in 26.71/0.825 higher PSNR (dB)/SSIM. By using the $\\mathcal{L}_{\\mathrm{bwl}}$ or $\\mathcal{L}_{\\mathrm{bal}}$ , we respectively achieve 0.05/0.003 or 0.10/0.004 higher PSNR (dB)/SSIM. Combining all losses, we have the best restoration performance at 26.91/0.831 for PSNR (dB)/SSIM. Moreover, we verify the generalization ability of our $\\mathcal{L}_{\\mathrm{bwl}}$ and $\\mathcal{L}_{\\mathrm{bal}}$ losses on state-of-the-art networks, improving PSNR of Restormer and DeepRFT by 0.25 and 0.41 PSNR (dB).", + "bbox": [ + 496, + 434, + 890, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Blur Map Estimation. We ablate with different prompts used to generate the blur map. The blur-aware and DP-aware prompts are presented in Tab. 4, and we show their impact on model performance with Tab. 5. We then ensemble the groups of blur-aware prompt and disparity-aware prompt individually (Ens. $^1$ ), leading to $26.80\\mathrm{dB} / 0.829$ and $26.84 / 0.830$ PSNR (dB)/SSIM. By assembling all prompts (Ens. $^2$ ), we have $26.91 / 0.831$ PSNR (dB)/SSIM, consistently outperforming other approaches of generating blur masks.", + "bbox": [ + 496, + 607, + 892, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Discussion and Limitation", + "text_level": 1, + "bbox": [ + 498, + 770, + 736, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Discussion. The reader may wonder about the performance of our unsupervised blur map estimation strategy. Therefore, we perform two qualitative evaluations on the DPD-disp dataset, where the GT disparity is included. We compare with BAMBNet [20] and DPD-disp [27].", + "bbox": [ + 496, + 794, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, we consider blur map estimation. To obtain the GT blur map, the pixel with disparity $d$ that satisfies $|d|\\leq 1$", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24084", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/054057e6ea0ca3eda4da11297624db98f45cdf17840b5131bb678044b149f705.jpg", + "table_caption": [ + "Table 4. Prompts used for blur map estimation." + ], + "table_footnote": [], + "table_body": "
Blur-aware PromptDP-aware Prompt
p1[A blurry image.][A symmetrical image.]
p2[An obscured image.][A horizontally symmetrical image.]
p3[An out of focus image.][A left-right symmetrical image.]
p4[An image lacking sharpness.][A symmetrical image at the pixel level.]
", + "bbox": [ + 81, + 103, + 883, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd845bada4709046aa36c7bcda4835ebe18691ddb98faac64d34b125b6988318.jpg", + "table_caption": [ + "Table 5. Restoration performance of our methods by using blur-aware or/and DP-aware prompts for blur map estimation. Ens. ${}^{1}$ denotes ensembling the group of blue-aware prompts and DP-aware prompts individually, and Ens. ${}^{2}$ ensembles all prompts." + ], + "table_footnote": [], + "table_body": "
Blur-aware PromptDP-aware Prompt
PSNR↑SSIM↑MAE↓RMSE_rel↓PSNR↑SSIM↑MAE↓RMSE_rel↓
p126.830.8290.0330.04526.820.8300.0330.046
p226.810.8290.0330.04626.830.8300.0320.046
p326.750.8280.0340.04626.820.8300.0330.046
p426.770.8290.0330.04626.850.8300.0320.045
Ens.126.800.8290.0330.04626.840.8300.0320.046
Ens.226.910.8310.0320.04526.910.8310.0320.045
", + "bbox": [ + 78, + 222, + 892, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1304abdac5fba0c5de280780e68e59e9f90789abc3022ac3f19f817492955c6a.jpg", + "table_caption": [ + "Table 6. Ablation of model components." + ], + "table_footnote": [], + "table_body": "
SettingPSNR↑SSIM↑MAE↓RMSE rel↓
i) NAFNet26.210.8170.0380.049
ii) Concat26.760.8260.0330.046
iii) SA26.590.8240.0360.047
iv) Concat + SA26.800.8290.0320.046
v) BPA block26.910.8310.0320.045
", + "bbox": [ + 78, + 387, + 488, + 487 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e3c0811b2d01731d6a2a8a170cbde36ffd7f4ea5ecea06901cbe448122af2fd0.jpg", + "table_caption": [ + "Table 7. The effectiveness of losses." + ], + "table_footnote": [], + "table_body": "
BPA blockLossPSNR↑SSIM↑MAE↓RMSE rel↓
XLchar26.210.8170.0380.049
XLtotal26.530.8220.0360.047
✓Lchar26.710.8250.0340.046
✓Lchar + Lbwl26.760.8280.0330.046
✓Lchar + Lbal26.810.8290.0320.046
✓Ltotal26.910.8310.0320.045
", + "bbox": [ + 78, + 520, + 477, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is used as the sharp pixel, reversely, it is a buried pixel. Our method, BAMBNet [20] and DPD-disp [27] respectively have $89.8\\% / 70.2\\% / 88.3\\%$ accuracy (the higher, the better), showing the effectiveness of our method.", + "bbox": [ + 75, + 656, + 468, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second, we explore disparity estimation. The log-its of our unnormalized blurred map are potentially linearly correlated with the ground-truth disparity. Therefore, we compare with the state-of-the-art DP disparity estimation method, DPD-disp [27]. Following [9, 27], we use AI(1), AI(2), and $1 - |\\rho_s|$ as the evaluation metrics (the lower, the better). The AI(1), AI(2), and $1 - |\\rho_s|$ of our method, BAMBNet, and DPD-disp are 0.063/0.106/0.069, 0.130/0.184/0.096, 0.306/0.772/0.0159. Though our blur map is estimated with a single forward pass, we still achieve a reasonable performance in the disparity estimation. This potentially explains the high performance of our DP defo", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cus deblurring network, as obtaining the disparity prior benefits the defocus task.", + "bbox": [ + 496, + 376, + 890, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. As the image encoder of CLIP is trained under a classification-alike pre-training task, it will focus more on local details that help identify its scene class. Textureless regions would produce non-significant semantic reactions, e.g., large areas of wall, ceiling, or floor, and reduce the quality of the estimated blur map. It is possible to use a vision language framework that drives more sensitivity to regions with less texture.", + "bbox": [ + 496, + 415, + 892, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Broader Impact", + "text_level": 1, + "bbox": [ + 498, + 553, + 794, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study end-to-end defocus deblurring of DP images, leveraging the blur-related prior knowledge from the CLIP. We first estimate a blur map with an ensemble of blur-aware and DP-aware strategies, and then use the estimated blur map as the deblurring kernel prior to restoring the DP image. We also propose a blur-aware and blur-weighting loss to regularize the restorations of the DP images during training, by distilling the knowledge of blur from CLIP. In extensive experiments, our method outperforms past works by a large margin in quantitative and qualitative restoration performance.", + "bbox": [ + 496, + 580, + 890, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact. The proposed DP-aware strategy is promising to apply and extend CLIP to diverse zero-shot stereo vision tasks. We hope it will motivate future work.", + "bbox": [ + 496, + 755, + 893, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045, and Miaomiao Liu was supported by the ARC fellowship and Discovery Project grant (DE180100628, DP200102274).", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24085", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 1, 2, 4, 6, 7", + "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 3, 6", + "[3] Abdullah Abuolaim, Mahmoud Affi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2, 4", + "[4] Dylan Auty and Krystian Mikolajczyk. Learning to prompt clip for monocular depth estimation: Exploring the limits of human language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2039-2047, 2023. 2", + "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022. 5, 7", + "[6] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 5, 7", + "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7628-7637, 2019. 1, 3, 8", + "[10] Xueting Hu, Ce Zhang, Yi Zhang, Bowen Hai, Ke Yu, and Zhihai He. Learning to adapt clip for few-shot monocular depth estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5594-5603, 2024. 2", + "[11] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22157-22167, 2023. 2", + "[12] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 5", + "[13] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, 2017. 6", + "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 1", + "[15] Junyong Lee, Sungkil Lee, Sunghyun Cho, and Seungyong Lee. Deep defocus map estimation using domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12222-12230, 2019. 6", + "[16] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 6, 7", + "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 2", + "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 5", + "[19] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1495-1503, 2023. 1", + "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Bambnet: A blur-aware multi-branch network for dual-pixel defocus deblurring. IEEE/CAA Journal of Automatica Sinica, 9(5):878-892, 2022. 1, 2, 6, 7, 8", + "[21] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2", + "[22] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the 37th AAAI Conference on Artificial Intelligence, 2023. 6, 7", + "[23] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. arXiv preprint arXiv:2306.07921, 2023. 1", + "[24] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1", + "[25] Liyuan Pan, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Phase-only image based kernel estimation for single image blind deblurring. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 501, + 92, + 890, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24086", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 6034-6043, 2019. 2", + "[26] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 3, 6, 7", + "[27] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 3, 4, 6, 7, 8", + "[28] Yuhui Quan, Zicong Wu, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2023. 1", + "[29] Yuhui Quan, Xin Yao, and Hui Ji. Single image defocus deblurring via implicit neural inverse kernels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12600-12610, 2023. 1", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5", + "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 1, 2", + "[32] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 1", + "[33] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 2622-2630. IEEE, 2021. 1, 2", + "[34] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2, 3", + "[35] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. arXiv preprint arXiv:2312.01381, 2023. 2", + "[36] Shaoliang Yang, Dongming Zhou, Jinde Cao, and Yanbu Guo. Lightingnet: An integrated learning method for low-light image enhancement. IEEE Transactions on Computational Imaging, 9:29-42, 2023. 5", + "[37] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 1", + "[38] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13272, 2023. 1, 2, 6, 7", + "[39] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 5, 6, 7", + "[40] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 1, 2", + "[41] Renrui Zhang, Ziyao Zeng, Ziyu Guo, and Yafeng Li. Can language understand depth? In Proceedings of the 30th ACM International Conference on Multimedia, pages 6868-6874, 2022. 2, 3", + "[42] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du2net: Learning depth estimation from dual-cameras and dual-pixels. In European Conference on Computer Vision, pages 582-598, 2020. 1", + "[43] Qian Zhao, Hao Yang, Dongming Zhou, and Jinde Cao. Rethinking image deblurring via cnn-transformer multiscale hybrid architecture. IEEE Transactions on Instrumentation and Measurement, 72:1-15, 2022. 5", + "[44] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175–11185, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 597 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24087", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_model.json b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fc4d1a047712b52db0e7afe1539323c3ce1e62c1 --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_model.json @@ -0,0 +1,2860 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.18, + 0.838, + 0.218 + ], + "angle": 0, + "content": "Hao Yang\\(^{1}\\), Liyuan Pan\\(^{1\\dagger}\\), Yan Yang\\(^{2}\\), Richard Hartley\\(^{2}\\), and Miaomiao Liu\\(^{2}\\) \n\\(^{1}\\)Beijing Institute of Technology \n\\(^{2}\\)Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.22, + 0.872, + 0.235 + ], + "angle": 0, + "content": "{hao.yang, liyuan.pan}@bit.edu.cn, {yan.yang, richard.hartley, miaomiao.liu}@anu.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.474, + 0.575 + ], + "angle": 0, + "content": "Recovering sharp images from dual-pixel (DP) pairs with disparity-dependent blur is a challenging task. Existing blur map-based deblurring methods have demonstrated promising results. In this paper, we propose, to the best of our knowledge, the first framework that introduces the contrastive language-image pre-training framework (CLIP) to accurately estimate the blur map from a DP pair unsupervisedly. To achieve this, we first carefully design text prompts to enable CLIP to understand blur-related geometric prior knowledge from the DP pair. Then, we propose a format to input a stereo DP pair to CLIP without any fine-tuning, despite the fact that CLIP is pre-trained on monocular images. Given the estimated blur map, we introduce a blur-prior attention block, a blur-weighting loss, and a blur-aware loss to recover the all-in-focus image. Our method achieves state-of-the-art performance in extensive experiments (see Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.605, + 0.21, + 0.621 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.858 + ], + "angle": 0, + "content": "Defocus deblurring from a single image is an ill-posed problem [19, 28, 29, 33]. Compared to methods that restore an all-in-focus image directly [1, 2], blur map-based deblurring methods have demonstrated promising results [16, 26]. To simplify the blur removal process, several works [20, 34, 38] have utilized recent innovative dual-pixel (DP) sensors [6] to estimate blur maps, since the DP sensor captures a two-view image pair (a DP pair) in a single shot. The DP pair can be treated as images from a two-sample light field camera or a stereo system [24] with a tiny baseline [27, 34], which is beneficial for disparity estimation [1, 14, 23, 42]. Based on the DP image pair formation process, the disparity is geometrically related to the blur map, i.e., the degree of defocus blur proportional to disparity magnitudes [9]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.859, + 0.469, + 0.875 + ], + "angle": 0, + "content": "For estimating the blur map (defocus map or disparity" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.267, + 0.895, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.432, + 0.895, + 0.515 + ], + "angle": 0, + "content": "Figure 1. Comparison of performance in terms of PSNR and Inference time. Our small and large models are respectively denoted by the Blue and Red cycles. The Green triangles represent previous state-of-the-art methods, and each of them is annotated with the method name, publication venue, and year. This figure is best viewed in color on the screen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.895, + 0.682 + ], + "angle": 0, + "content": "map), previous works either require extra data [3, 26], such as synthetic data, for supervision signals, or pre-calibrated blur kernels [34]. To avoid using extra information, several works explored unsupervised blur map estimation frameworks [16, 20]. Lee et al. [16] implicitly estimate the disparity of the left/right views using a spatial transformer, yet only in feature space, as well as Yang et al. [38]. Liang et al. [20] use fixed network branches for layer-wise defocus map estimation and deblurring, where the limited network branch number restricts the potential of its accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Recently, the contrastive language-image pre-training framework (CLIP) [30] has accomplished extraordinary success for vision tasks, such as semantic segmentation [31, 37], object detection [32], and point cloud understanding [40]. Therefore, a question naturally arises, can we avoid the cost of collecting data and model designing, to estimate the blur map unsupervisedly by using CLIP? Unfortunately, using semantic knowledge from CLIP to handle low-level vision tasks is underexplored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, for the first time, we explore the relationship between language and blur map estimation by applying CLIP. Note that CLIP is trained to match a monocular image with its corresponding language descriptions. Given a blurred image, it is intuitive to match it with a corre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.233, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "sponding text [A blurry image.], which allows the language model to find the blurred regions in an image. However, it is hard to quantize the blur level of an image without knowing its associated latent sharp one [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.303 + ], + "angle": 0, + "content": "Fortunately, we can leverage the capabilities of DP sensors. With DP pairs, blur quantization becomes possible without the need for its associated sharp image. The formation of DP ensures no disparity between left/right views for pixels in sharp regions, while a noticeable disparity exists for pixels in blurred regions. In essence, the detectable differences (disparities) in left/right views of a DP pair directly indicate the degree of defocus blurs. Consequently, we can quantify the variation between the left and right views, treating it as a valuable cue for estimating a blur map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.468, + 0.454 + ], + "angle": 0, + "content": "However, CLIP still finds it challenging to simultaneously explore stereo DP pairs with left/right view images. To address this issue, we propose to explore the DP pair and design an 'image and text' format to better capture the blurs in the image. Specifically, we propose to create a new image from the DP pair and translate the description about the 'blurriness' of the image to measure the 'symmetry' of the newly formed image, which is based on the DP image formation process. The freshly designed image-text format allows the CLIP to estimate the blur map easily (see Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.56 + ], + "angle": 0, + "content": "Given the estimated blur map, we then restore the all-in-focus image from the DP pair. By observing that the attention map of the self-attention mechanism [8] can be used in the dynamic DP deblurring process, we introduce a blur prior attention (BPA) that modulates the attention map with the estimated blur map which provides prior knowledge of deblurring kernels (see Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.561, + 0.468, + 0.68 + ], + "angle": 0, + "content": "To regulate the restoration, we specifically design two losses that are based on CLIP. Considering the network should focus more on heavily blurred regions, we propose a blur-weighting loss that imposes an adaptive penalty proportional to the blur degrees provided by the estimated blur map from CLIP. Then, to guarantee the restored image is sharp, we use CLIP to re-check the restored image and formulate a blur-aware loss (see Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.682, + 0.399, + 0.695 + ], + "angle": 0, + "content": "Our contributions are summarised as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.696, + 0.468, + 0.739 + ], + "angle": 0, + "content": "- We propose a Language-driven DP (LDP) defocus deblurring framework, by exploring the potential of CLIP in the low-level vision task." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.74, + 0.468, + 0.782 + ], + "angle": 0, + "content": "- We design an image-text format for blur map estimation based on the geometric relationship between the blur and disparity of the DP pair." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.782, + 0.468, + 0.826 + ], + "angle": 0, + "content": "- We propose a blur prior attention (BPA) block, a blur-weighting loss, and a blur-aware loss to encourage sharp restorations of the DP pair." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.696, + 0.468, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed framework extensively on standard benchmark datasets. Though using the largescale foundation model, CLIP, for blur map estimation, our method achieves Pareto-optimality between deblurring performance and inference speed (refer to Fig. 1 for an" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.696, + 0.107 + ], + "angle": 0, + "content": "overview of the comparison)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.121, + 0.642, + 0.137 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.147, + 0.892, + 0.48 + ], + "angle": 0, + "content": "DP Defocus Deblurring. DP-based deblurring methods can be divided into single- and multi-task manners. For single-task based deblurring [1-3], methods apply different kernels at different regions of the DP pair to handle the spatially varying defocus deblur. This can usually be achieved by using parallel atrous convolutions [33], attention mechanisms [39], or deformable convolutions [17]. For multi-task deblurring methods, the blur/defocus/disparity map is auxiliarily estimated to benefit DP defocus deblurring. Pan et al. [26] and Xin et al. [34] use the estimated DP disparity to regularise the restoration through reblurring; however, with extra data or pre-calibrate kernel. Several works explored unsupervised defocus map (or disparity) estimation. Lee et al. [16] and Yang et al. [38] estimate defocus disparity only in feature space without explicitly showing a defocus map. Liang et al. [20] optimize a multi-channel defocus map based on [27]. However, hyperparameters, such as the defocus threshold and the number of channels, are preset, limiting blur map estimation accuracy. Unlike past works, by transferring the blur-related geometric prior knowledge into semantic prompts, we unsupervisedly estimate the blur map from CLIP to benefit defocus deblurring." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Language-driven Vision. Common objectives in visual-language learning are aligning the text with a paired image in the embedding space [21, 30, 35]. The CLIP, one of the milestone achievement in the area, demonstrate that language-independent computer vision tasks can be substantially benefited from visual-language learning, by achieving SOTA zero-shot classification performance in the ImageNet-1K benchmark [7]. Successive works have extended the zero-shot learning capability of CLIP to semantic segmentation [31, 44], monocular depth estimation [4, 10, 41], and point clouds [11, 40]. However, none of the existing works explores the zero-shot learning capability of CLIP in stereo/DP pairs. This paper proposes the first approach to use the monocular image-trained CLIP in zero-shot blur map segmentation of DP pairs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.591, + 0.739 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Let us now introduce our language-driven dual-pixel image defocus deblurring framework. Given a defocus blurred DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\), we aim to restore an all-in-focus image \\(\\mathbf{I}\\) where \\(\\mathbf{B}_{\\mathrm{L}}\\) and \\(\\mathbf{B}_{\\mathrm{R}}\\) are the left and right view of a DP pair and \\(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}},\\mathbf{I}\\in \\mathbb{R}^{h\\times w\\times 3}\\). Here, \\(h\\) and \\(w\\) define the image height and width, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first design a DP-aware prompt to adapt CLIP to output a blur map, \\(\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times 1}\\), which quantizes disparity unsupervisedly between \\(\\mathbf{B}_{\\mathrm{L}}\\) and \\(\\mathbf{B}_{\\mathrm{R}}\\). Then, we propose a BPA block based network which employs the blur" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24079" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.088, + 0.833, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Overall architecture. Given a DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\), we first generate a blur map \\(\\mathbf{M}\\) by using our unsupervised blur map estimation strategy which is based on the CLIP with assembling format. We then forward the DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\) to our deblurring backbone composed of an encoder and a decoder. Meanwhile, we have multiple BPA blocks, using the blur map \\(\\mathbf{M}\\) as deblurring kernel prior to refine the intermediate backbone feature embeddings. In the following, we have the restoration \\(\\hat{\\mathbf{I}}\\). The deblurring network is optimized with three losses, charbonnier loss \\(\\mathcal{L}_{char}\\), blur weighting loss \\(\\mathcal{L}_{bwl}\\), and blur aware loss \\(\\mathcal{L}_{bal}\\). \\(\\mathcal{L}_{char}\\) penalizes deviations from \\(\\mathbf{I}\\) in the resorption \\(\\hat{\\mathbf{I}}\\). \\(\\mathcal{L}_{bwl}\\) and \\(\\mathcal{L}_{bal}\\) regularize the restoration by imposing an adaptive penalty on the blurred regions of the DP pair, and leveraging unsupervised blur map estimation strategy to encourage the restoration to be sharp, i.e., the blur map \\(\\hat{\\mathbf{M}}\\) for the restoration \\(\\hat{\\mathbf{I}}\\) is zero value filled." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.456 + ], + "angle": 0, + "content": "map \\(\\mathbf{M}\\) as deblurring priors to restore a sharp image \\(\\hat{\\mathbf{I}}\\) from \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\). To regularize the restoration of \\(\\hat{\\mathbf{I}}\\), a blur-weighting loss \\(\\mathcal{L}_{\\mathrm{bwl}}\\) and a blur-aware loss \\(\\mathcal{L}_{\\mathrm{bal}}\\) are also introduced that were underpinned by our unsupervised blur map estimation strategy. The overall framework is shown in Fig. 2. Details are presented below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.463, + 0.388, + 0.48 + ], + "angle": 0, + "content": "3.1. Unsupervised Blur Map Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Preliminary. CLIP respectively embeds an image and a text description into a common semantic space of dimension \\( c \\) with an image encoder and a text encoder, obtaining image embedding \\( \\mathbf{f} \\in \\mathbb{R}^c \\) and text embedding \\( \\mathbf{t} \\in \\mathbb{R}^c \\). By optimizing a contrastive loss, the embeddings of the aligned image and text pair, \\( \\Omega_a \\), stay similar, i.e., maximizing the semantic similarity by \\( \\max_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_a} \\sin(\\mathbf{f}, \\mathbf{t}) \\), where \\( \\sin(\\cdot, \\cdot) \\) is the cosine similarity metric. Meanwhile, the semantics of misaligned pairs, \\( \\Omega_{na} \\) are pushed away from each other, i.e., \\( \\min_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_{na}} \\sin(\\mathbf{f}, \\mathbf{t}) \\). With comprehensive training data, CLIP can be leveraged to align images with pre-defined prompt templates composed of open-world text descriptions for zero-shot learning [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.729 + ], + "angle": 0, + "content": "For a DP pair, the disparity \\( d \\) for a pixel in the left view is defined by its location \\( (i,j) \\) and that of its corresponding pixel at \\( (i,k) \\) in the right view." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.729, + 0.438, + 0.745 + ], + "angle": 0, + "content": "In the literature, this DP disparity [27] is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.753, + 0.47, + 0.772 + ], + "angle": 0, + "content": "\\[\nd = j - k = A (r (z)) \\cdot B (z), \\mathbf {B} _ {\\mathrm {L}} (i, j) = \\mathbf {B} _ {\\mathrm {R}} (i, k), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.841 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_{\\mathrm{L}}(i,j)\\) and \\(\\mathbf{B}_{\\mathrm{R}}(i,k)\\) are the intensity of the two matching points, \\(r(\\cdot)\\) is a function of the radius of the circle of confusion, \\(z\\) is the point depth, and \\(A(r(z))\\) and \\(B(z)\\) are lens-dependent constants [2, 9, 26, 34]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Analysis. Leveraging the CLIP, zero-shot blur map estimation can be generally divided into three steps. i) Image Encoding. Considering the image encoder of the CLIP is trained with monocular RGB images, a function \\(\\eta (\\cdot ,\\cdot)\\) is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.472 + ], + "angle": 0, + "content": "required to create a new monocular image from the DP image pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\). To obtain the dense embedding for the blur map estimation task, we follow [41] to pop out the final pooling layer from the image encoder, yielding an embedding matrix \\(\\mathbf{F}\\in \\mathbb{R}^{h_s\\times w_s\\times c}\\), where \\(h_s\\) and \\(w_{s}\\) are the height and width for embedding of the newly created image, respectively. Mathematically, we have" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.486, + 0.892, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\operatorname {I m a g e E n c o d e r} \\left(\\eta \\left(\\mathbf {B} _ {\\mathrm {L}}, \\mathbf {B} _ {\\mathrm {R}}\\right)\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.894, + 0.579 + ], + "angle": 0, + "content": "Each pixel of \\(\\mathbf{F}\\) has encoded the local semantic information for a corresponding region in \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\). ii) Text Encoding. A text encoder is used to embed the semantics of a prompt p describing the blur into the text embedding," + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.594, + 0.892, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} = \\operatorname {T e x t E n c o d e r} (\\mathfrak {p}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.718 + ], + "angle": 0, + "content": "iii) Blur map estimation. Let \\(\\mathbf{F}(i,j)\\) be a c-dimensional embedding at position \\((i,j)\\) of \\(\\mathbf{F}\\), \\(i\\in \\{1,\\dots ,\\mathrm{h}_{\\mathrm{s}}\\}\\), and \\(j\\in \\{1,\\dots ,\\mathrm{w}_{\\mathrm{s}}\\}\\). The embedding \\(\\mathbf{F}(i,j)\\) and the blur semantic-based text embedding \\(\\mathbf{t}\\) are expected to be similar if the pixel at \\((i,j)\\) is blurred. Therefore, the blur mask \\(\\mathbf{M}_{\\mathbf{F}}\\) of \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\) is obtained by" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.731, + 0.892, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\mathbf {F}} (i, j) = \\sigma (\\operatorname {s i m} (\\mathbf {F} (i, j), \\mathbf {t})) . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.824 + ], + "angle": 0, + "content": "Note, \\(\\sigma (\\cdot)\\) normalizes the cosine similarity \\(\\mathfrak{sim}(\\cdot ,\\cdot)\\) output to a probability space, to reflect the degree of blur at the position. Here, \\(\\mathbf{M}_{\\mathbf{F}}\\) is sampled to have the same resolution as \\(\\mathbf{B}_{\\mathrm{L}}\\) or \\(\\mathbf{B}_{\\mathrm{R}}\\), returning \\(\\mathbf{M}\\in \\mathbb{R}^{h\\times w\\times 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.916 + ], + "angle": 0, + "content": "Given the analysis above, our goal is to develop text prompts \\(\\mathfrak{p}\\), while adapting the function \\(\\eta(\\cdot, \\cdot)\\), to enable the use of semantic information from CLIP to generate the blur map \\(\\mathbf{M}\\) in an unsupervised manner. In particular, we explore the blur-aware, DP-aware, and ensembling formats with detailed presented below." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.464, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.474, + 0.432 + ], + "angle": 0, + "content": "Figure 3. The pipeline of our DP-aware format for obtaining the blur map \\(\\mathbf{M}\\). Given a DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\), the function \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\) concatenates the left view \\(\\mathbf{B}_{\\mathrm{L}}\\) and with horizontally flipped right view \\(\\mathbf{B}_{\\mathrm{R}}^{h}\\) for obtaining the \\(\\mathbf{F}\\) from the CLIP image encoder. We then apply \\(\\alpha (F)\\) to averagely pool the feature map according to the transformation imposed on the input DP pair, resulting in a new feature embedding with the same aspect ratio of the input DP pair. Meanwhile, we set the prompts \\(\\mathfrak{p}\\) to describe the symmetry of \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\), obtaining text embedding \\(\\mathbf{t}\\). We finally have the blur map \\(\\mathbf{M}\\) by computing Eq. 6 and performing sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.47, + 0.581 + ], + "angle": 0, + "content": "Blur-aware Format. An intuitive approach is setting a prompt \\(\\mathbf{p}\\) that describes the blur to query a blur map \\(\\mathbf{M}\\), e.g., \\(\\mathbf{p} = [\\mathrm{A~blurry~image}.]\\). The \\(\\eta(\\cdot, \\cdot)\\) is then defined to synthesize a center view of the DP pair by averaging the left view \\(\\mathbf{B}_{\\mathrm{L}}\\) and right view \\(\\mathbf{B}_{\\mathrm{R}}\\), as the average of two sub-images is equivalent to the image captured by a regular sensor [3, 27]. However, the disparity information among the DP pair is neglected, resulting in a sub-optimal blur map. Refer to Fig. 4 for comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.736 + ], + "angle": 0, + "content": "DP-aware Format. Based on the DP image formation process, the disparity is proportional to the degree of blur exhibited in the DP view [1]. We thus propose to guide CLIP to understand the disparity exhibited between \\(\\mathbf{B}_{\\mathrm{L}}\\) and \\(\\mathbf{B}_{\\mathrm{R}}\\) which eventually leads to better mask (M) generation. To this end, we first formulate the disparity in the DP pair and then study the disparity-aware prompt and function \\(\\eta (\\cdot ,\\cdot)\\) designing. Such design leads us to query degrees of displacement between the left and right view from the CLIP, without any fine-tuning, to obtain the blur map M." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.469, + 0.873 + ], + "angle": 0, + "content": "Note that the CLIP is the vision-text alignment model trained with monocular image input. To extract the displacement between left/right views of the DP pair, we explore several input-prompt formats: i) Concatenating the left and right views horizontally, and prompting their difference, e.g., [The left and right of the image are different.]; ii) Prompting the horizontal symmetry between the left and right views, e.g., [A horizontally symmetric image.] (Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Empirically, we achieve the best performance with the second prompt format. When describing the symmetry in" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.598, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.192, + 0.559, + 0.202 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.089, + 0.695, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.192, + 0.655, + 0.202 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.791, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.192, + 0.752, + 0.202 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.089, + 0.888, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.192, + 0.848, + 0.202 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.204, + 0.895, + 0.287 + ], + "angle": 0, + "content": "Figure 4. Example of blur map estimation (the darker, the sharper). (a) The left view of the DP image. (b) and (c) Blur maps obtained with prompts of the blur-aware and DP-aware formats. (d) The blur maps obtained with the ensembling-based format, i.e., all eight blur-aware and DP-aware prompts are used (please refer to Tab. 4 for details)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.419 + ], + "angle": 0, + "content": "the prompt \\(\\mathfrak{p}\\), we concatenate the left view \\(\\mathbf{B}_{\\mathrm{L}}\\) with a horizontally flipped right view \\(\\mathbf{B}_{\\mathrm{R}}^{h} = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})\\) along the width axial in the \\(\\eta (\\cdot ,\\cdot)\\). We then obtained image features \\(\\mathbf{F} = \\mathrm{ImageEncoder}\\bigl (\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\bigr)\\). To transform \\(\\mathbf{F}\\) to the same aspect ratio of \\(\\mathbf{B}_{\\mathrm{L}}\\) and \\(\\mathbf{B}_{\\mathrm{R}}\\), we average the corresponding pixel-wise features based on the transformation in \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\) and have \\(\\alpha (\\mathbf{F})\\). For \\(i\\in \\{1,\\dots ,\\mathsf{h}_{\\mathsf{w}}\\}\\) and \\(j\\in \\{1,\\dots ,\\lfloor \\frac{1}{2}\\mathsf{w_s}\\rfloor \\}\\), we define" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.43, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\alpha (\\mathbf {F}) (i, j) = \\frac {1}{2} \\mathbf {F} (i, j) + \\frac {1}{2} \\mathbf {F} (i, w _ {s} - j). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.469, + 0.794, + 0.485 + ], + "angle": 0, + "content": "Back to the left-hand side of Eq. 4, we have" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.496, + 0.891, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {\\mathbf {F}} (i, j) = 1 - \\mathbf {M} _ {\\mathbf {F}} ^ {s} (i, j) = 1 - \\sigma \\left(\\operatorname {s i m} (\\alpha (\\mathbf {F}) (i, j), \\mathbf {t} ^ {s})\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{t}^s\\) is the text embedding of the prompt describing symmetry. Note that the high similarity value, i.e. \\(\\mathbf{M}_{\\mathbf{F}}^{s}(i,j) = \\sigma \\big(\\mathrm{sim}(\\alpha (\\mathbf{F})(i,j),\\mathbf{t}^{s})\\big)\\) indicates pixels in sharp regions. Therefore the blur map is derived as \\(1 - \\mathbf{M}_{\\mathbf{F}}^{s}(i,j)\\) in Eq. 6. Similarly, the \\(\\mathbf{M}_{\\mathbf{F}}\\) is sampled to the same resolution of the DP pair, returning M. For this blur map estimation strategy, we conclude the lemma below," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.895, + 0.696 + ], + "angle": 0, + "content": "Lemma 3.1. A pixel is sharp if and only if it is horizontally symmetric in the left view and the horizontally flipped right view, otherwise, the pixel is blurred." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.892, + 0.792 + ], + "angle": 0, + "content": "Proof. Following Eq. (1), we have \\(\\mathbf{B}_{\\mathrm{L}}(i,j)\\) being sharp, given \\(\\mathbf{B}_{\\mathrm{L}}(i,j) = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})(i,\\mathsf{w} - j)\\). In other words, we have \\(\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,j) = \\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,\\mathsf{w}_{\\mathsf{s}} - j)\\), demonstrating the horizontal symmetry. Reversely, the horizontal symmetry breaks once the disparity/blur occurs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.9 + ], + "angle": 0, + "content": "Ensembling format. Given the analysis that the degradations in the DP pair can be described by both the blurriness and symmetricity, we thus propose to obtain the mask by assembling the blur-aware and DP-aware formats jointly. Please refer to Tab. 4 for details of the 8 prompts designed from blur-aware and DP-aware formats." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24081" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.181, + 0.105 + ], + "angle": 0, + "content": "3.2. Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.22 + ], + "angle": 0, + "content": "We train an encoder-decoder based deblurring network for restoring the DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\) into a sharp image \\(\\hat{\\mathbf{I}}\\) with the derived blur map \\(\\mathbf{M}\\). Our network consists of a backbone network based on [5] and a BPA block to interleave with the backbone network layers. In particular, the BPA block uses the obtained blur map \\(\\mathbf{M}\\) as deblur kernel prior to help the restoration of \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\) on the fly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.471, + 0.358 + ], + "angle": 0, + "content": "Self-attention. The self-attention (SA) mechanism is widely used in the image restoration tasks [18, 36, 39]. We first briefly review it. Let \\(\\mathbf{X} \\in \\mathbb{R}^{h_x \\times w_x \\times c_x}\\) be the intermediate feature embedding of a blur image from the backbone encoder, where \\(h_x\\), \\(w_x\\), \\(c_x\\) are height, width, and channel of the feature. The self-attention is input with query \\(\\mathbf{Q}\\), key \\(\\mathbf{K}\\), and value \\(\\mathbf{V}\\) that are linearly projected from \\(\\mathbf{X}\\), and then calculate an attention map between \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) for adaptively aggregating \\(\\mathbf{V}\\). This process is defined as follows," + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.371, + 0.469, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {Q}}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {K}}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {V}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.39, + 0.469, + 0.408 + ], + "angle": 0, + "content": "Attention \\((\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta)\\mathbf{V}\\) (8)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.483 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{h_{\\mathbf{x}} \\times w_{\\mathbf{x}} \\times c_{\\mathbf{x}}}\\) are learnable weight matrices, \\(\\beta\\) is a normalization factor that is typically set to the square root of the latent dimension, and \\(\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta) \\in \\mathbb{R}^{(h_{\\mathbf{x}} \\times w_{\\mathbf{x}}) \\times (h_{\\mathbf{x}} \\times w_{\\mathbf{x}})}\\) is attention map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.59 + ], + "angle": 0, + "content": "BPA Block. The attention map can be considered as an adaptive deblur kernel prior for restoring the DP pair in the feature space. We are motivated to modulate the attention map by supplying the blur map \\(\\mathbf{M}\\), utilizing blur-related prior knowledge to additionally assign a \\(\\mathbf{q} \\times \\mathbf{q}\\) deblur kernel for each region in \\(\\mathbf{X}\\) to aggregate a sharp feature embedding from its neighboring regions. We re-write the Eq. 8 as" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.602, + 0.469, + 0.619 + ], + "angle": 0, + "content": "Attention \\((\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta +\\mathbf{O})\\mathbf{V}\\) (9)" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.622, + 0.469, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\mathbf {O} = \\operatorname {P a d} \\left(\\operatorname {F F N} _ {\\operatorname {C o n v}} \\left(\\sigma^ {- 1} (\\mathbf {M})\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.652, + 0.469, + 0.775 + ], + "angle": 0, + "content": "and show that the restoration result is further improved in Fig. 5, where \\(\\mathrm{FFN}_{\\mathrm{Conv}}(\\cdot)\\) is a convolution-based feed forward layer that is used to learn a \\(q \\times q\\) deblur kernel for each region of \\(\\mathbf{X}\\) from the unnormalized blur map \\(\\sigma^{-1}(\\mathbf{M})\\), resulting in \\(\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M})) \\in \\mathbb{R}^{h_x \\times w_x \\times q^2}\\). The Pad(\\(\\cdot\\)) function reshapes \\(\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))\\) into \\(\\mathbb{R}^{h_x \\times w_x \\times q \\times q}\\), and pads the reshaped \\(\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))\\) to \\(\\mathbb{R}^{(h_x \\times w_x) \\times (h_x \\times w_x)}\\), for shape alignment issue." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.252, + 0.801 + ], + "angle": 0, + "content": "3.3. Network Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.469, + 0.868 + ], + "angle": 0, + "content": "We optimize our deblurring network with the widely used Charbonnier loss. Moreover, we propose a blur-aware loss and a blur weighting loss to penalize the blurred regions in the restoration \\(\\hat{\\mathbf{I}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Charbonnier loss. We encourage the restoration to be close to the ground truth by using Charbonnier loss [12, 43]." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.571, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.147, + 0.559, + 0.157 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.089, + 0.694, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.147, + 0.655, + 0.157 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.763, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.147, + 0.752, + 0.157 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.089, + 0.887, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.147, + 0.849, + 0.157 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.158, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Figure 5. Example of deblurring with the blur map projected deblur kernel \\(\\mathbf{O}\\). (a) and (b) are the left view of the DP pair and the ground truth sharp image. (c) and (d) are deblurred images with and without using \\(\\mathbf{O}\\). Zoom in for better quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.232, + 0.771, + 0.247 + ], + "angle": 0, + "content": "We empirically set \\(\\varepsilon\\) to \\(10^{-4}\\), and define" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.258, + 0.892, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\| ^ {2} + \\varepsilon^ {2}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Blur Weighting Loss. We impose an adaptive penalty on the blurred regions of the DP pair \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\), considering the degree of blur measured by CLIP. It encourages the network to focus on the restoration of heavily blurred regions. We first normalize the deviations between \\(\\hat{\\mathbf{I}}\\) and \\(\\mathbf{I}\\) by the deviations between \\((\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})\\) and \\(\\mathbf{I}\\), and then re-use the blur map to select blurred regions residing in the DP pair." + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.408, + 0.891, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {b w l}} = \\frac {\\left\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\right\\| ^ {2}}{\\left\\| \\hat {\\mathbf {I}} (i , j) - \\frac {1}{2} \\left(\\mathbf {B} _ {\\mathrm {L}} (i , j) + \\mathbf {B} _ {\\mathrm {R}} (i , j)\\right) \\right\\| ^ {2}} \\cdot \\mathbf {M} (i, j), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.472, + 0.892, + 0.502 + ], + "angle": 0, + "content": "where we use the blur map \\(\\mathbf{M}\\) (with soft logits) to reflect the blur levels detected by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.502, + 0.892, + 0.578 + ], + "angle": 0, + "content": "Blur-aware Loss. We encourage \\(\\hat{\\mathbf{I}}\\) to be sharp by leveraging the versatile knowledge contained in the CLIP. We generate the blur map \\(\\hat{\\mathbf{M}}\\) by the blur-aware format for the restoration, and penalize the blur exhibited in \\(\\hat{\\mathbf{I}}\\). Similar to \\(\\mathcal{L}_{\\mathrm{bw1}}\\), we use the unnormalized version of the blur map," + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.588, + 0.891, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b a l}} = \\hat {\\mathbf {M}} (i, j). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.618, + 0.89, + 0.647 + ], + "angle": 0, + "content": "This encourages \\(\\hat{\\mathbf{M}}\\) to be zero, i.e., no blur exists in the restoration." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.649, + 0.89, + 0.68 + ], + "angle": 0, + "content": "Total Loss. Our network is optimized by minimizing the following loss function" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.689, + 0.891, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{| \\Omega |} \\sum_ {(i, j) \\in \\Omega} \\left(\\mathcal {L} _ {\\text {c h a r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {b a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {b w l}}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.739, + 0.892, + 0.786 + ], + "angle": 0, + "content": "where \\(\\frac{1}{|\\Omega|}\\) is a normalization factor determined by the image domain size \\(|\\Omega|\\), and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters to balance losses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.744, + 0.817 + ], + "angle": 0, + "content": "4. Experiments and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We implement our model under the PyTorch framework. We use the pre-trained and ViT-B/32 [8] architecture-based CLIP [30] for our blur map estimation strategy. We train the deblurring network in two stages for \\(6 \\times 10^{6}\\) steps in total with the Adam optimizer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24082" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the DPD-blur dataset [1] composed of 37 indoor and 39 outdoor scenes. We respectively color the best and the second-best performance in red and blue." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.117, + 0.899, + 0.339 + ], + "angle": 0, + "content": "
MethodOutdoor ScenesIndoor ScenesAverage
PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓LPIPS↓
EBDB25.770.7720.0400.05121.250.5990.0580.08623.450.6830.0490.0670.336
DMENet25.500.7880.0380.05321.430.6440.0630.08423.410.7140.0510.0670.349
DPDNet27.480.8490.0290.04222.900.7260.0520.07125.130.7860.0410.0550.223
RDPD28.100.8430.0270.03922.820.7040.0530.07225.390.7720.0400.0530.255
DDDNet27.570.8330.0300.04123.280.7080.0500.06825.360.7680.0410.0540.197
IFAN28.660.8680.0250.03723.460.7430.0490.06725.990.8040.0370.0500.207
BAMBNet28.600.8720.0260.03724.300.7720.0450.06026.400.8210.0360.0470.141
DeepRFT28.480.8700.0250.03823.090.7360.0510.07025.710.8010.0370.0510.217
Restormer29.480.8950.0230.03423.970.7730.0470.06326.660.8330.0350.0460.155
K3DN29.870.8900.0220.03224.370.7800.0460.06027.060.8350.0340.0440.187
Ours (Small)29.550.8930.0210.03324.380.7730.0430.06526.910.8310.0320.0450.185
Ours (Large)29.880.8960.0200.03224.730.7860.0370.05927.240.8400.0280.0430.168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.343, + 0.449, + 0.356 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on DDD-syn dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.357, + 0.476, + 0.558 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑MAE↓RMSE_rel↓
EBDB26.480.6830.0360.047
DMENet30.140.9390.0170.031
DPDNet31.450.9260.0160.027
RDPD32.420.9120.0160.024
DDDNet33.210.9560.0100.022
IFAN34.180.9290.0120.020
BAMBNet35.900.9540.0120.016
DeepRFT36.520.9520.0100.015
Restormer36.440.9570.0100.015
K3DN38.230.9650.0080.012
Ours (Small)37.270.9590.0100.014
Ours (Large)38.620.9710.0080.012
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.679 + ], + "angle": 0, + "content": "In each stage, the learning rate is decayed from \\(2 \\times 10^{-4}\\) to \\(10^{-6}\\) by a cosine annealing strategy. In the first stage, we train the network for \\(5 \\times 10^{6}\\) steps, and only use the Charbonnier loss \\(\\mathcal{L}_{\\mathrm{char}}\\). In the second stage, the network is finetuned for \\(1 \\times 10^{6}\\) steps with \\(\\mathcal{L}_{\\mathrm{total}}\\), where we have \\(\\lambda_{1} = 10^{-1}\\) and \\(\\lambda_{2} = 2 \\times 10^{-1}\\). Code, results, and more details are given in our project page." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Datasets. Our method is evaluated on two real-world DP defocus deblurring datasets, DPD-blur dataset [1] and DPD-disp dataset [27]. We follow [1] on the training and testing splits. For the DPD-disp dataset, we use it for testing only, re-using the checkpoints trained from the DPD-blur dataset as they do not provide ground-truth (GT) data or with limited dataset size. Moreover, two synthetic datasets, DDD-syn [26] and RDPD [2], are also used for evaluations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.807, + 0.47, + 0.867 + ], + "angle": 0, + "content": "Evaluation Metrics. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean absolute error (MAE), and root mean squared relative error (RMSE_ref) as our evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Baseline Methods. We compare with SOTA DP defocus deblurring methods, i.e., EBDB [13], DMENet [15]," + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.343, + 0.848, + 0.356 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons on RDPD dataset." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.357, + 0.895, + 0.558 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑MAE↓RMSE rel↓
EBDB24.230.6370.0310.054
DMENet25.170.7310.0360.061
DPDNet29.840.8280.0320.055
RDPD31.090.8610.0250.028
DDDNet30.140.8400.0160.031
IFAN31.120.8650.0230.028
BAMBNet31.780.8670.0160.026
DeepRFT31.710.8790.0160.024
Restormer32.270.8710.0160.024
K3DN32.840.9050.0140.023
Ours (Small)32.980.8950.0140.022
Ours (Large)33.440.9070.0130.021
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.572, + 0.892, + 0.603 + ], + "angle": 0, + "content": "DPDNet [1], RDPD [2], DDDNet [26], IFAN [16], BAMBNet [20], DeepRFT [22], Restormer [39], and K3DN [38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.805, + 0.631 + ], + "angle": 0, + "content": "4.1. Comparisons with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.865 + ], + "angle": 0, + "content": "Quantitative Comparison. We compare with the SOTA methods on the DPD-blur, DDD-syn, and RDPD datasets in Tab. 1, Tab. 2, and Tab. 3, respectively. Our observations are as follows: i) With our model (Small), we achieve similar restoration results with Restormer and K3DN, the third-best and second-best method, in all datasets. ii) When scaling to the large model, we consistently outperform all methods on all evaluation metrics. For example, our method, K3DN, and Restormer respectively achieve 27.24/27.06/26.66, 38.62/38.23/36.44, and 33.44/32.84/32.27 PSNR (dB) in the DPD-blur, DDD-syn, and RDPD datasets. Meanwhile, on a single RTX 4090 GPU, we remain to have a lower Infer. time (including CLIP) which is 0.042s and 0.054s for our small and large models, while K3DN/Restormer takes 0.198s/0.255s." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparison. We compare with state-of-the-art methods on the DPD-blur dataset (Fig. 6). As shown, our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24083" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.169, + 0.089, + 0.402, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.217, + 0.34, + 0.232 + ], + "angle": 0, + "content": "Blurry Image" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.089, + 0.501, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.144, + 0.486, + 0.159 + ], + "angle": 0, + "content": "GT Image" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.601, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.144, + 0.596, + 0.159 + ], + "angle": 0, + "content": "Blurry Image" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.09, + 0.699, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.144, + 0.684, + 0.159 + ], + "angle": 0, + "content": "IFAN [16]" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.09, + 0.798, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.144, + 0.795, + 0.159 + ], + "angle": 0, + "content": "DDDNet [26]" + }, + { + "type": "image", + "bbox": [ + 0.407, + 0.163, + 0.501, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.216, + 0.5, + 0.232 + ], + "angle": 0, + "content": "DeepRFT [22]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.163, + 0.6, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.216, + 0.604, + 0.232 + ], + "angle": 0, + "content": "Restormer [39]" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.163, + 0.699, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.216, + 0.695, + 0.232 + ], + "angle": 0, + "content": "K3DN [38]" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.163, + 0.798, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.218, + 0.776, + 0.231 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.235, + 0.402, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.362, + 0.339, + 0.376 + ], + "angle": 0, + "content": "Blurry Image" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.235, + 0.501, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.289, + 0.486, + 0.304 + ], + "angle": 0, + "content": "GT Image" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.235, + 0.6, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.289, + 0.596, + 0.304 + ], + "angle": 0, + "content": "Blurry Image" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.235, + 0.699, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.289, + 0.684, + 0.304 + ], + "angle": 0, + "content": "IFAN [16]" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.235, + 0.798, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.289, + 0.795, + 0.304 + ], + "angle": 0, + "content": "DDDNet [26]" + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.309, + 0.501, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.362, + 0.5, + 0.376 + ], + "angle": 0, + "content": "DeepRFT [22]" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.309, + 0.6, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.362, + 0.604, + 0.376 + ], + "angle": 0, + "content": "Restormer [39]" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.309, + 0.699, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.362, + 0.694, + 0.376 + ], + "angle": 0, + "content": "K3DN [38]" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.309, + 0.798, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.363, + 0.776, + 0.376 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.379, + 0.894, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on the DPD-blur dataset [1]. We present the blurry image \\((\\mathbf{B}_{\\mathrm{L}})\\) in the first column, and the ground truth (GT) image of regions residing in the red bounding box is shown in the second column. The cropped blurred image of the red bounding box region is in the third column." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.526 + ], + "angle": 0, + "content": "method performs sharper and clearer restorations of edges and texts. To further demonstrate the benefits of using CLIP to generate the blur mask unsupervisedly in our approach, we provide an example in Fig. 7, which preserves sharp regions (e.g., text 'University') in the input blur image effectively compared to the second-best baseline." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.53, + 0.177, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.571, + 0.155, + 0.582 + ], + "angle": 0, + "content": "(a) Input" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.53, + 0.273, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.571, + 0.254, + 0.582 + ], + "angle": 0, + "content": "(b) K3DN" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.53, + 0.37, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.571, + 0.359, + 0.582 + ], + "angle": 0, + "content": "(c) Restormer" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.53, + 0.466, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.571, + 0.44, + 0.582 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Figure 7. An example of the restoration results on the sharp region of the input DP image (left view)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.23, + 0.637 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.688 + ], + "angle": 0, + "content": "We study model architecture, blur map estimation, and losses in this section. All experiments are performed in the DPD-blur dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Ablation of Model Architecture. We study the effective usage of blur map as prior knowledge for DP defocus deblurring (Tab. 6). To show the benefits of using blur maps, we consider four baselines: i) NAFNet [5], the baseline model; ii) Concat, concatenating the blur map with the DP pair before feeding to the model; iii) SA, using the self-attention layer [8] in Eq. 8 to the model only; iv) Concat+SA, combines the Concat and SA settings; v) BPA block, using kernel priors from the blur map to assist deblurring. Our model consistently outperforms all baselines, which indicates the effectiveness of using the blur map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Ablation of Losses. We investigate the proposed blur weighting loss \\(\\mathcal{L}_{\\mathrm{bw}}\\) and blur-aware loss \\(\\mathcal{L}_{\\mathrm{bal}}\\) (Tab. 7) with a baseline as the \\(\\mathcal{L}_{\\mathrm{char}}\\) setting. We first use \\(\\mathcal{L}_{\\mathrm{bw}}\\) and \\(\\mathcal{L}_{\\mathrm{bal}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.602 + ], + "angle": 0, + "content": "to our baseline, NAFNet (i.e., not using our BPA block), that only uses \\(\\mathcal{L}_{\\mathrm{char}}\\), having 0.32 PSNR (dB) improvements. Secondly, we train our model (backbone + BPA block) with Charbonnier loss, resulting in 26.71/0.825 higher PSNR (dB)/SSIM. By using the \\(\\mathcal{L}_{\\mathrm{bwl}}\\) or \\(\\mathcal{L}_{\\mathrm{bal}}\\), we respectively achieve 0.05/0.003 or 0.10/0.004 higher PSNR (dB)/SSIM. Combining all losses, we have the best restoration performance at 26.91/0.831 for PSNR (dB)/SSIM. Moreover, we verify the generalization ability of our \\(\\mathcal{L}_{\\mathrm{bwl}}\\) and \\(\\mathcal{L}_{\\mathrm{bal}}\\) losses on state-of-the-art networks, improving PSNR of Restormer and DeepRFT by 0.25 and 0.41 PSNR (dB)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.893, + 0.76 + ], + "angle": 0, + "content": "Ablation of Blur Map Estimation. We ablate with different prompts used to generate the blur map. The blur-aware and DP-aware prompts are presented in Tab. 4, and we show their impact on model performance with Tab. 5. We then ensemble the groups of blur-aware prompt and disparity-aware prompt individually (Ens.\\(^1\\)), leading to \\(26.80\\mathrm{dB} / 0.829\\) and \\(26.84 / 0.830\\) PSNR (dB)/SSIM. By assembling all prompts (Ens.\\(^2\\)), we have \\(26.91 / 0.831\\) PSNR (dB)/SSIM, consistently outperforming other approaches of generating blur masks." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.771, + 0.738, + 0.785 + ], + "angle": 0, + "content": "4.3. Discussion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Discussion. The reader may wonder about the performance of our unsupervised blur map estimation strategy. Therefore, we perform two qualitative evaluations on the DPD-disp dataset, where the GT disparity is included. We compare with BAMBNet [20] and DPD-disp [27]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "First, we consider blur map estimation. To obtain the GT blur map, the pixel with disparity \\(d\\) that satisfies \\(|d|\\leq 1\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24084" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.344, + 0.09, + 0.625, + 0.103 + ], + "angle": 0, + "content": "Table 4. Prompts used for blur map estimation." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.104, + 0.885, + 0.19 + ], + "angle": 0, + "content": "
Blur-aware PromptDP-aware Prompt
p1[A blurry image.][A symmetrical image.]
p2[An obscured image.][A horizontally symmetrical image.]
p3[An out of focus image.][A left-right symmetrical image.]
p4[An image lacking sharpness.][A symmetrical image at the pixel level.]
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.195, + 0.893, + 0.223 + ], + "angle": 0, + "content": "Table 5. Restoration performance of our methods by using blur-aware or/and DP-aware prompts for blur map estimation. Ens. \\( {}^{1} \\) denotes ensembling the group of blue-aware prompts and DP-aware prompts individually, and Ens. \\( {}^{2} \\) ensembles all prompts." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.223, + 0.893, + 0.356 + ], + "angle": 0, + "content": "
Blur-aware PromptDP-aware Prompt
PSNR↑SSIM↑MAE↓RMSE_rel↓PSNR↑SSIM↑MAE↓RMSE_rel↓
p126.830.8290.0330.04526.820.8300.0330.046
p226.810.8290.0330.04626.830.8300.0320.046
p326.750.8280.0340.04626.820.8300.0330.046
p426.770.8290.0330.04626.850.8300.0320.045
Ens.126.800.8290.0330.04626.840.8300.0320.046
Ens.226.910.8310.0320.04526.910.8310.0320.045
" + }, + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.375, + 0.391, + 0.388 + ], + "angle": 0, + "content": "Table 6. Ablation of model components." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.388, + 0.489, + 0.488 + ], + "angle": 0, + "content": "
SettingPSNR↑SSIM↑MAE↓RMSE rel↓
i) NAFNet26.210.8170.0380.049
ii) Concat26.760.8260.0330.046
iii) SA26.590.8240.0360.047
iv) Concat + SA26.800.8290.0320.046
v) BPA block26.910.8310.0320.045
" + }, + { + "type": "table_caption", + "bbox": [ + 0.167, + 0.507, + 0.379, + 0.521 + ], + "angle": 0, + "content": "Table 7. The effectiveness of losses." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.521, + 0.478, + 0.641 + ], + "angle": 0, + "content": "
BPA blockLossPSNR↑SSIM↑MAE↓RMSE rel↓
XLchar26.210.8170.0380.049
XLtotal26.530.8220.0360.047
✓Lchar26.710.8250.0340.046
✓Lchar + Lbwl26.760.8280.0330.046
✓Lchar + Lbal26.810.8290.0320.046
✓Ltotal26.910.8310.0320.045
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.718 + ], + "angle": 0, + "content": "is used as the sharp pixel, reversely, it is a buried pixel. Our method, BAMBNet [20] and DPD-disp [27] respectively have \\(89.8\\% / 70.2\\% / 88.3\\%\\) accuracy (the higher, the better), showing the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Second, we explore disparity estimation. The log-its of our unnormalized blurred map are potentially linearly correlated with the ground-truth disparity. Therefore, we compare with the state-of-the-art DP disparity estimation method, DPD-disp [27]. Following [9, 27], we use AI(1), AI(2), and \\(1 - |\\rho_s|\\) as the evaluation metrics (the lower, the better). The AI(1), AI(2), and \\(1 - |\\rho_s|\\) of our method, BAMBNet, and DPD-disp are 0.063/0.106/0.069, 0.130/0.184/0.096, 0.306/0.772/0.0159. Though our blur map is estimated with a single forward pass, we still achieve a reasonable performance in the disparity estimation. This potentially explains the high performance of our DP defo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.406 + ], + "angle": 0, + "content": "cus deblurring network, as obtaining the disparity prior benefits the defocus task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Limitation. As the image encoder of CLIP is trained under a classification-alike pre-training task, it will focus more on local details that help identify its scene class. Textureless regions would produce non-significant semantic reactions, e.g., large areas of wall, ceiling, or floor, and reduce the quality of the estimated blur map. It is possible to use a vision language framework that drives more sensitivity to regions with less texture." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.554, + 0.795, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In this paper, we study end-to-end defocus deblurring of DP images, leveraging the blur-related prior knowledge from the CLIP. We first estimate a blur map with an ensemble of blur-aware and DP-aware strategies, and then use the estimated blur map as the deblurring kernel prior to restoring the DP image. We also propose a blur-aware and blur-weighting loss to regularize the restorations of the DP images during training, by distilling the knowledge of blur from CLIP. In extensive experiments, our method outperforms past works by a large margin in quantitative and qualitative restoration performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.756, + 0.894, + 0.802 + ], + "angle": 0, + "content": "Broader Impact. The proposed DP-aware strategy is promising to apply and extend CLIP to diverse zero-shot stereo vision tasks. We hope it will motivate future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045, and Miaomiao Liu was supported by the ARC fellowship and Discovery Project grant (DE180100628, DP200102274)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24085" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.298 + ], + "angle": 0, + "content": "[3] Abdullah Abuolaim, Mahmoud Affi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.471, + 0.366 + ], + "angle": 0, + "content": "[4] Dylan Auty and Krystian Mikolajczyk. Learning to prompt clip for monocular depth estimation: Exploring the limits of human language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2039-2047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.438 + ], + "angle": 0, + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[6] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7628-7637, 2019. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[10] Xueting Hu, Ce Zhang, Yi Zhang, Bowen Hai, Ke Yu, and Zhihai He. Learning to adapt clip for few-shot monocular depth estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5594-5603, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.789, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[11] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22157-22167, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[12] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[13] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[15] Junyong Lee, Sungkil Lee, Sunghyun Cho, and Seungyong Lee. Deep defocus map estimation using domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12222-12230, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[16] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[19] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1495-1503, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Bambnet: A blur-aware multi-branch network for dual-pixel defocus deblurring. IEEE/CAA Journal of Automatica Sinica, 9(5):878-892, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[21] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[22] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the 37th AAAI Conference on Artificial Intelligence, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[23] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. arXiv preprint arXiv:2306.07921, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[24] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[25] Liyuan Pan, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Phase-only image based kernel estimation for single image blind deblurring. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24086" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 6034-6043, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[26] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[27] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[28] Yuhui Quan, Zicong Wu, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[29] Yuhui Quan, Xin Yao, and Hui Ji. Single image defocus deblurring via implicit neural inverse kernels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12600-12610, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.467 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[32] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[33] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 2622-2630. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[34] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[35] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. arXiv preprint arXiv:2312.01381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[36] Shaoliang Yang, Dongming Zhou, Jinde Cao, and Yanbu Guo. Lightingnet: An integrated learning method for low-light image enhancement. IEEE Transactions on Computational Imaging, 9:29-42, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[38] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13272, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[39] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[40] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[41] Renrui Zhang, Ziyao Zeng, Ziyu Guo, and Yafeng Li. Can language understand depth? In Proceedings of the 30th ACM International Conference on Multimedia, pages 6868-6874, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[42] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du2net: Learning depth estimation from dual-cameras and dual-pixels. In European Conference on Computer Vision, pages 582-598, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.528 + ], + "angle": 0, + "content": "[43] Qian Zhao, Hao Yang, Dongming Zhou, and Jinde Cao. Rethinking image deblurring via cnn-transformer multiscale hybrid architecture. IEEE Transactions on Instrumentation and Measurement, 72:1-15, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[44] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175–11185, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24087" + } + ] +] \ No newline at end of file diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_origin.pdf b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f38c29cf584aa11028c1ad8d60589486c9637620 --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/2d70b048-ecc7-45e4-b194-27af8843d7f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0161d633393139fecfe62ac365db0984986944aba09d2fffa3b2397c9d77f17 +size 3591496 diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/full.md b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3250fac55dec408945a94395d9638955461edf18 --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/full.md @@ -0,0 +1,422 @@ +# LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network + +Hao Yang $^{1}$ , Liyuan Pan $^{1\dagger}$ , Yan Yang $^{2}$ , Richard Hartley $^{2}$ , and Miaomiao Liu $^{2}$ $^{1}$ Beijing Institute of Technology + $^{2}$ Australian National University + +{hao.yang, liyuan.pan}@bit.edu.cn, {yan.yang, richard.hartley, miaomiao.liu}@anu.edu.au + +# Abstract + +Recovering sharp images from dual-pixel (DP) pairs with disparity-dependent blur is a challenging task. Existing blur map-based deblurring methods have demonstrated promising results. In this paper, we propose, to the best of our knowledge, the first framework that introduces the contrastive language-image pre-training framework (CLIP) to accurately estimate the blur map from a DP pair unsupervisedly. To achieve this, we first carefully design text prompts to enable CLIP to understand blur-related geometric prior knowledge from the DP pair. Then, we propose a format to input a stereo DP pair to CLIP without any fine-tuning, despite the fact that CLIP is pre-trained on monocular images. Given the estimated blur map, we introduce a blur-prior attention block, a blur-weighting loss, and a blur-aware loss to recover the all-in-focus image. Our method achieves state-of-the-art performance in extensive experiments (see Fig. 1). + +# 1. Introduction + +Defocus deblurring from a single image is an ill-posed problem [19, 28, 29, 33]. Compared to methods that restore an all-in-focus image directly [1, 2], blur map-based deblurring methods have demonstrated promising results [16, 26]. To simplify the blur removal process, several works [20, 34, 38] have utilized recent innovative dual-pixel (DP) sensors [6] to estimate blur maps, since the DP sensor captures a two-view image pair (a DP pair) in a single shot. The DP pair can be treated as images from a two-sample light field camera or a stereo system [24] with a tiny baseline [27, 34], which is beneficial for disparity estimation [1, 14, 23, 42]. Based on the DP image pair formation process, the disparity is geometrically related to the blur map, i.e., the degree of defocus blur proportional to disparity magnitudes [9]. + +For estimating the blur map (defocus map or disparity + +![](images/4289276f10163e17c2dd9aa619e2d3d885516c5673317bc1d8f7f482698b7003.jpg) +Figure 1. Comparison of performance in terms of PSNR and Inference time. Our small and large models are respectively denoted by the Blue and Red cycles. The Green triangles represent previous state-of-the-art methods, and each of them is annotated with the method name, publication venue, and year. This figure is best viewed in color on the screen. + +map), previous works either require extra data [3, 26], such as synthetic data, for supervision signals, or pre-calibrated blur kernels [34]. To avoid using extra information, several works explored unsupervised blur map estimation frameworks [16, 20]. Lee et al. [16] implicitly estimate the disparity of the left/right views using a spatial transformer, yet only in feature space, as well as Yang et al. [38]. Liang et al. [20] use fixed network branches for layer-wise defocus map estimation and deblurring, where the limited network branch number restricts the potential of its accuracy. + +Recently, the contrastive language-image pre-training framework (CLIP) [30] has accomplished extraordinary success for vision tasks, such as semantic segmentation [31, 37], object detection [32], and point cloud understanding [40]. Therefore, a question naturally arises, can we avoid the cost of collecting data and model designing, to estimate the blur map unsupervisedly by using CLIP? Unfortunately, using semantic knowledge from CLIP to handle low-level vision tasks is underexplored. + +In this paper, for the first time, we explore the relationship between language and blur map estimation by applying CLIP. Note that CLIP is trained to match a monocular image with its corresponding language descriptions. Given a blurred image, it is intuitive to match it with a corre + +sponding text [A blurry image.], which allows the language model to find the blurred regions in an image. However, it is hard to quantize the blur level of an image without knowing its associated latent sharp one [25]. + +Fortunately, we can leverage the capabilities of DP sensors. With DP pairs, blur quantization becomes possible without the need for its associated sharp image. The formation of DP ensures no disparity between left/right views for pixels in sharp regions, while a noticeable disparity exists for pixels in blurred regions. In essence, the detectable differences (disparities) in left/right views of a DP pair directly indicate the degree of defocus blurs. Consequently, we can quantify the variation between the left and right views, treating it as a valuable cue for estimating a blur map. + +However, CLIP still finds it challenging to simultaneously explore stereo DP pairs with left/right view images. To address this issue, we propose to explore the DP pair and design an 'image and text' format to better capture the blurs in the image. Specifically, we propose to create a new image from the DP pair and translate the description about the 'blurriness' of the image to measure the 'symmetry' of the newly formed image, which is based on the DP image formation process. The freshly designed image-text format allows the CLIP to estimate the blur map easily (see Sec. 3.1). + +Given the estimated blur map, we then restore the all-in-focus image from the DP pair. By observing that the attention map of the self-attention mechanism [8] can be used in the dynamic DP deblurring process, we introduce a blur prior attention (BPA) that modulates the attention map with the estimated blur map which provides prior knowledge of deblurring kernels (see Sec. 3.2). + +To regulate the restoration, we specifically design two losses that are based on CLIP. Considering the network should focus more on heavily blurred regions, we propose a blur-weighting loss that imposes an adaptive penalty proportional to the blur degrees provided by the estimated blur map from CLIP. Then, to guarantee the restored image is sharp, we use CLIP to re-check the restored image and formulate a blur-aware loss (see Sec. 3.3). + +Our contributions are summarised as follows: + +- We propose a Language-driven DP (LDP) defocus deblurring framework, by exploring the potential of CLIP in the low-level vision task. +- We design an image-text format for blur map estimation based on the geometric relationship between the blur and disparity of the DP pair. +- We propose a blur prior attention (BPA) block, a blur-weighting loss, and a blur-aware loss to encourage sharp restorations of the DP pair. + +We evaluate the proposed framework extensively on standard benchmark datasets. Though using the largescale foundation model, CLIP, for blur map estimation, our method achieves Pareto-optimality between deblurring performance and inference speed (refer to Fig. 1 for an + +overview of the comparison). + +# 2. Related Work + +DP Defocus Deblurring. DP-based deblurring methods can be divided into single- and multi-task manners. For single-task based deblurring [1-3], methods apply different kernels at different regions of the DP pair to handle the spatially varying defocus deblur. This can usually be achieved by using parallel atrous convolutions [33], attention mechanisms [39], or deformable convolutions [17]. For multi-task deblurring methods, the blur/defocus/disparity map is auxiliarily estimated to benefit DP defocus deblurring. Pan et al. [26] and Xin et al. [34] use the estimated DP disparity to regularise the restoration through reblurring; however, with extra data or pre-calibrate kernel. Several works explored unsupervised defocus map (or disparity) estimation. Lee et al. [16] and Yang et al. [38] estimate defocus disparity only in feature space without explicitly showing a defocus map. Liang et al. [20] optimize a multi-channel defocus map based on [27]. However, hyperparameters, such as the defocus threshold and the number of channels, are preset, limiting blur map estimation accuracy. Unlike past works, by transferring the blur-related geometric prior knowledge into semantic prompts, we unsupervisedly estimate the blur map from CLIP to benefit defocus deblurring. + +Language-driven Vision. Common objectives in visual-language learning are aligning the text with a paired image in the embedding space [21, 30, 35]. The CLIP, one of the milestone achievement in the area, demonstrate that language-independent computer vision tasks can be substantially benefited from visual-language learning, by achieving SOTA zero-shot classification performance in the ImageNet-1K benchmark [7]. Successive works have extended the zero-shot learning capability of CLIP to semantic segmentation [31, 44], monocular depth estimation [4, 10, 41], and point clouds [11, 40]. However, none of the existing works explores the zero-shot learning capability of CLIP in stereo/DP pairs. This paper proposes the first approach to use the monocular image-trained CLIP in zero-shot blur map segmentation of DP pairs. + +# 3. Method + +Let us now introduce our language-driven dual-pixel image defocus deblurring framework. Given a defocus blurred DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ , we aim to restore an all-in-focus image $\mathbf{I}$ where $\mathbf{B}_{\mathrm{L}}$ and $\mathbf{B}_{\mathrm{R}}$ are the left and right view of a DP pair and $\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}},\mathbf{I}\in \mathbb{R}^{h\times w\times 3}$ . Here, $h$ and $w$ define the image height and width, respectively. + +We first design a DP-aware prompt to adapt CLIP to output a blur map, $\mathbf{M} \in \mathbb{R}^{h \times w \times 1}$ , which quantizes disparity unsupervisedly between $\mathbf{B}_{\mathrm{L}}$ and $\mathbf{B}_{\mathrm{R}}$ . Then, we propose a BPA block based network which employs the blur + +![](images/990fe983387e39d2f8b277fcf1f9eba3549aaf603c220bf2942e5ea3d19158d0.jpg) +Figure 2. Overall architecture. Given a DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ , we first generate a blur map $\mathbf{M}$ by using our unsupervised blur map estimation strategy which is based on the CLIP with assembling format. We then forward the DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ to our deblurring backbone composed of an encoder and a decoder. Meanwhile, we have multiple BPA blocks, using the blur map $\mathbf{M}$ as deblurring kernel prior to refine the intermediate backbone feature embeddings. In the following, we have the restoration $\hat{\mathbf{I}}$ . The deblurring network is optimized with three losses, charbonnier loss $\mathcal{L}_{char}$ , blur weighting loss $\mathcal{L}_{bwl}$ , and blur aware loss $\mathcal{L}_{bal}$ . $\mathcal{L}_{char}$ penalizes deviations from $\mathbf{I}$ in the resorption $\hat{\mathbf{I}}$ . $\mathcal{L}_{bwl}$ and $\mathcal{L}_{bal}$ regularize the restoration by imposing an adaptive penalty on the blurred regions of the DP pair, and leveraging unsupervised blur map estimation strategy to encourage the restoration to be sharp, i.e., the blur map $\hat{\mathbf{M}}$ for the restoration $\hat{\mathbf{I}}$ is zero value filled. + +map $\mathbf{M}$ as deblurring priors to restore a sharp image $\hat{\mathbf{I}}$ from $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ . To regularize the restoration of $\hat{\mathbf{I}}$ , a blur-weighting loss $\mathcal{L}_{\mathrm{bwl}}$ and a blur-aware loss $\mathcal{L}_{\mathrm{bal}}$ are also introduced that were underpinned by our unsupervised blur map estimation strategy. The overall framework is shown in Fig. 2. Details are presented below. + +# 3.1. Unsupervised Blur Map Estimation + +Preliminary. CLIP respectively embeds an image and a text description into a common semantic space of dimension $c$ with an image encoder and a text encoder, obtaining image embedding $\mathbf{f} \in \mathbb{R}^c$ and text embedding $\mathbf{t} \in \mathbb{R}^c$ . By optimizing a contrastive loss, the embeddings of the aligned image and text pair, $\Omega_a$ , stay similar, i.e., maximizing the semantic similarity by $\max_{\mathbf{f}, \mathbf{t} \in \Omega_a} \sin(\mathbf{f}, \mathbf{t})$ , where $\sin(\cdot, \cdot)$ is the cosine similarity metric. Meanwhile, the semantics of misaligned pairs, $\Omega_{na}$ are pushed away from each other, i.e., $\min_{\mathbf{f}, \mathbf{t} \in \Omega_{na}} \sin(\mathbf{f}, \mathbf{t})$ . With comprehensive training data, CLIP can be leveraged to align images with pre-defined prompt templates composed of open-world text descriptions for zero-shot learning [30]. + +For a DP pair, the disparity $d$ for a pixel in the left view is defined by its location $(i,j)$ and that of its corresponding pixel at $(i,k)$ in the right view. + +In the literature, this DP disparity [27] is defined as + +$$ +d = j - k = A (r (z)) \cdot B (z), \mathbf {B} _ {\mathrm {L}} (i, j) = \mathbf {B} _ {\mathrm {R}} (i, k), \tag {1} +$$ + +where $\mathbf{B}_{\mathrm{L}}(i,j)$ and $\mathbf{B}_{\mathrm{R}}(i,k)$ are the intensity of the two matching points, $r(\cdot)$ is a function of the radius of the circle of confusion, $z$ is the point depth, and $A(r(z))$ and $B(z)$ are lens-dependent constants [2, 9, 26, 34]. + +Analysis. Leveraging the CLIP, zero-shot blur map estimation can be generally divided into three steps. i) Image Encoding. Considering the image encoder of the CLIP is trained with monocular RGB images, a function $\eta (\cdot ,\cdot)$ is + +required to create a new monocular image from the DP image pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ . To obtain the dense embedding for the blur map estimation task, we follow [41] to pop out the final pooling layer from the image encoder, yielding an embedding matrix $\mathbf{F}\in \mathbb{R}^{h_s\times w_s\times c}$ , where $h_s$ and $w_{s}$ are the height and width for embedding of the newly created image, respectively. Mathematically, we have + +$$ +\mathbf {F} = \operatorname {I m a g e E n c o d e r} \left(\eta \left(\mathbf {B} _ {\mathrm {L}}, \mathbf {B} _ {\mathrm {R}}\right)\right). \tag {2} +$$ + +Each pixel of $\mathbf{F}$ has encoded the local semantic information for a corresponding region in $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ . ii) Text Encoding. A text encoder is used to embed the semantics of a prompt p describing the blur into the text embedding, + +$$ +\mathbf {t} = \operatorname {T e x t E n c o d e r} (\mathfrak {p}). \tag {3} +$$ + +iii) Blur map estimation. Let $\mathbf{F}(i,j)$ be a c-dimensional embedding at position $(i,j)$ of $\mathbf{F}$ , $i\in \{1,\dots ,\mathrm{h}_{\mathrm{s}}\}$ , and $j\in \{1,\dots ,\mathrm{w}_{\mathrm{s}}\}$ . The embedding $\mathbf{F}(i,j)$ and the blur semantic-based text embedding $\mathbf{t}$ are expected to be similar if the pixel at $(i,j)$ is blurred. Therefore, the blur mask $\mathbf{M}_{\mathbf{F}}$ of $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ is obtained by + +$$ +\mathbf {M} _ {\mathbf {F}} (i, j) = \sigma (\operatorname {s i m} (\mathbf {F} (i, j), \mathbf {t})) . \tag {4} +$$ + +Note, $\sigma (\cdot)$ normalizes the cosine similarity $\mathfrak{sim}(\cdot ,\cdot)$ output to a probability space, to reflect the degree of blur at the position. Here, $\mathbf{M}_{\mathbf{F}}$ is sampled to have the same resolution as $\mathbf{B}_{\mathrm{L}}$ or $\mathbf{B}_{\mathrm{R}}$ , returning $\mathbf{M}\in \mathbb{R}^{h\times w\times 1}$ . + +Given the analysis above, our goal is to develop text prompts $\mathfrak{p}$ , while adapting the function $\eta(\cdot, \cdot)$ , to enable the use of semantic information from CLIP to generate the blur map $\mathbf{M}$ in an unsupervised manner. In particular, we explore the blur-aware, DP-aware, and ensembling formats with detailed presented below. + +![](images/18d45b8a20b3eaa4186ec61cbe6d094c70506c392e53a5046bae28f16976d1a4.jpg) +Figure 3. The pipeline of our DP-aware format for obtaining the blur map $\mathbf{M}$ . Given a DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ , the function $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})$ concatenates the left view $\mathbf{B}_{\mathrm{L}}$ and with horizontally flipped right view $\mathbf{B}_{\mathrm{R}}^{h}$ for obtaining the $\mathbf{F}$ from the CLIP image encoder. We then apply $\alpha (F)$ to averagely pool the feature map according to the transformation imposed on the input DP pair, resulting in a new feature embedding with the same aspect ratio of the input DP pair. Meanwhile, we set the prompts $\mathfrak{p}$ to describe the symmetry of $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})$ , obtaining text embedding $\mathbf{t}$ . We finally have the blur map $\mathbf{M}$ by computing Eq. 6 and performing sampling. + +Blur-aware Format. An intuitive approach is setting a prompt $\mathbf{p}$ that describes the blur to query a blur map $\mathbf{M}$ , e.g., $\mathbf{p} = [\mathrm{A~blurry~image}.]$ . The $\eta(\cdot, \cdot)$ is then defined to synthesize a center view of the DP pair by averaging the left view $\mathbf{B}_{\mathrm{L}}$ and right view $\mathbf{B}_{\mathrm{R}}$ , as the average of two sub-images is equivalent to the image captured by a regular sensor [3, 27]. However, the disparity information among the DP pair is neglected, resulting in a sub-optimal blur map. Refer to Fig. 4 for comparisons. + +DP-aware Format. Based on the DP image formation process, the disparity is proportional to the degree of blur exhibited in the DP view [1]. We thus propose to guide CLIP to understand the disparity exhibited between $\mathbf{B}_{\mathrm{L}}$ and $\mathbf{B}_{\mathrm{R}}$ which eventually leads to better mask (M) generation. To this end, we first formulate the disparity in the DP pair and then study the disparity-aware prompt and function $\eta (\cdot ,\cdot)$ designing. Such design leads us to query degrees of displacement between the left and right view from the CLIP, without any fine-tuning, to obtain the blur map M. + +Note that the CLIP is the vision-text alignment model trained with monocular image input. To extract the displacement between left/right views of the DP pair, we explore several input-prompt formats: i) Concatenating the left and right views horizontally, and prompting their difference, e.g., [The left and right of the image are different.]; ii) Prompting the horizontal symmetry between the left and right views, e.g., [A horizontally symmetric image.] (Fig. 3). + +Empirically, we achieve the best performance with the second prompt format. When describing the symmetry in + +![](images/36370877f7d879fc524e4ff8f9994de1b61705a140a2a3adf717c1156d906117.jpg) +(a) + +![](images/2adfbde67b3773aad6a74b04d1253cf65b912abc73088b937b9bfadbd6a43017.jpg) +(b) + +![](images/b0872cd650385ca4ec8a5634d760a6ffa4998117d1739009563c34e8098ffa4a.jpg) +(c) +Figure 4. Example of blur map estimation (the darker, the sharper). (a) The left view of the DP image. (b) and (c) Blur maps obtained with prompts of the blur-aware and DP-aware formats. (d) The blur maps obtained with the ensembling-based format, i.e., all eight blur-aware and DP-aware prompts are used (please refer to Tab. 4 for details). + +![](images/3b01f240f5b9ae389abd9aee3a971c85e3eec945b179276d04374353f0970329.jpg) +(d) + +the prompt $\mathfrak{p}$ , we concatenate the left view $\mathbf{B}_{\mathrm{L}}$ with a horizontally flipped right view $\mathbf{B}_{\mathrm{R}}^{h} = \mathrm{Flip}(\mathbf{B}_{\mathrm{R}})$ along the width axial in the $\eta (\cdot ,\cdot)$ . We then obtained image features $\mathbf{F} = \mathrm{ImageEncoder}\bigl (\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})\bigr)$ . To transform $\mathbf{F}$ to the same aspect ratio of $\mathbf{B}_{\mathrm{L}}$ and $\mathbf{B}_{\mathrm{R}}$ , we average the corresponding pixel-wise features based on the transformation in $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})$ and have $\alpha (\mathbf{F})$ . For $i\in \{1,\dots ,\mathsf{h}_{\mathsf{w}}\}$ and $j\in \{1,\dots ,\lfloor \frac{1}{2}\mathsf{w_s}\rfloor \}$ , we define + +$$ +\alpha (\mathbf {F}) (i, j) = \frac {1}{2} \mathbf {F} (i, j) + \frac {1}{2} \mathbf {F} (i, w _ {s} - j). \tag {5} +$$ + +Back to the left-hand side of Eq. 4, we have + +$$ +\mathbf {M} _ {\mathbf {F}} (i, j) = 1 - \mathbf {M} _ {\mathbf {F}} ^ {s} (i, j) = 1 - \sigma \left(\operatorname {s i m} (\alpha (\mathbf {F}) (i, j), \mathbf {t} ^ {s})\right). \tag {6} +$$ + +Here, $\mathbf{t}^s$ is the text embedding of the prompt describing symmetry. Note that the high similarity value, i.e. $\mathbf{M}_{\mathbf{F}}^{s}(i,j) = \sigma \big(\mathrm{sim}(\alpha (\mathbf{F})(i,j),\mathbf{t}^{s})\big)$ indicates pixels in sharp regions. Therefore the blur map is derived as $1 - \mathbf{M}_{\mathbf{F}}^{s}(i,j)$ in Eq. 6. Similarly, the $\mathbf{M}_{\mathbf{F}}$ is sampled to the same resolution of the DP pair, returning M. For this blur map estimation strategy, we conclude the lemma below, + +Lemma 3.1. A pixel is sharp if and only if it is horizontally symmetric in the left view and the horizontally flipped right view, otherwise, the pixel is blurred. + +Proof. Following Eq. (1), we have $\mathbf{B}_{\mathrm{L}}(i,j)$ being sharp, given $\mathbf{B}_{\mathrm{L}}(i,j) = \mathrm{Flip}(\mathbf{B}_{\mathrm{R}})(i,\mathsf{w} - j)$ . In other words, we have $\eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})(i,j) = \eta (\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}}^{h})(i,\mathsf{w}_{\mathsf{s}} - j)$ , demonstrating the horizontal symmetry. Reversely, the horizontal symmetry breaks once the disparity/blur occurs. + +Ensembling format. Given the analysis that the degradations in the DP pair can be described by both the blurriness and symmetricity, we thus propose to obtain the mask by assembling the blur-aware and DP-aware formats jointly. Please refer to Tab. 4 for details of the 8 prompts designed from blur-aware and DP-aware formats. + +# 3.2. Network + +We train an encoder-decoder based deblurring network for restoring the DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ into a sharp image $\hat{\mathbf{I}}$ with the derived blur map $\mathbf{M}$ . Our network consists of a backbone network based on [5] and a BPA block to interleave with the backbone network layers. In particular, the BPA block uses the obtained blur map $\mathbf{M}$ as deblur kernel prior to help the restoration of $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ on the fly. + +Self-attention. The self-attention (SA) mechanism is widely used in the image restoration tasks [18, 36, 39]. We first briefly review it. Let $\mathbf{X} \in \mathbb{R}^{h_x \times w_x \times c_x}$ be the intermediate feature embedding of a blur image from the backbone encoder, where $h_x$ , $w_x$ , $c_x$ are height, width, and channel of the feature. The self-attention is input with query $\mathbf{Q}$ , key $\mathbf{K}$ , and value $\mathbf{V}$ that are linearly projected from $\mathbf{X}$ , and then calculate an attention map between $\mathbf{Q}$ and $\mathbf{K}$ for adaptively aggregating $\mathbf{V}$ . This process is defined as follows, + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} _ {\mathbf {Q}}, \quad \mathbf {K} = \mathbf {X} \mathbf {W} _ {\mathbf {K}}, \quad \mathbf {V} = \mathbf {X} \mathbf {W} _ {\mathbf {V}}, \tag {7} +$$ + +Attention $(\mathbf{Q},\mathbf{K},\mathbf{V}) = \mathrm{Softmax}(\mathbf{Q}\mathbf{K}^{\top} / \beta)\mathbf{V}$ (8) + +where $\mathbf{W}_{\mathbf{Q}}, \mathbf{W}_{\mathbf{K}}, \mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{h_{\mathbf{x}} \times w_{\mathbf{x}} \times c_{\mathbf{x}}}$ are learnable weight matrices, $\beta$ is a normalization factor that is typically set to the square root of the latent dimension, and $\operatorname{Softmax}(\mathbf{Q}\mathbf{K}^{\top} / \beta) \in \mathbb{R}^{(h_{\mathbf{x}} \times w_{\mathbf{x}}) \times (h_{\mathbf{x}} \times w_{\mathbf{x}})}$ is attention map. + +BPA Block. The attention map can be considered as an adaptive deblur kernel prior for restoring the DP pair in the feature space. We are motivated to modulate the attention map by supplying the blur map $\mathbf{M}$ , utilizing blur-related prior knowledge to additionally assign a $\mathbf{q} \times \mathbf{q}$ deblur kernel for each region in $\mathbf{X}$ to aggregate a sharp feature embedding from its neighboring regions. We re-write the Eq. 8 as + +Attention $(\mathbf{Q},\mathbf{K},\mathbf{V}) = \mathrm{Softmax}(\mathbf{Q}\mathbf{K}^{\top} / \beta +\mathbf{O})\mathbf{V}$ (9) + +$$ +\mathbf {O} = \operatorname {P a d} \left(\operatorname {F F N} _ {\operatorname {C o n v}} \left(\sigma^ {- 1} (\mathbf {M})\right)\right), \tag {10} +$$ + +and show that the restoration result is further improved in Fig. 5, where $\mathrm{FFN}_{\mathrm{Conv}}(\cdot)$ is a convolution-based feed forward layer that is used to learn a $q \times q$ deblur kernel for each region of $\mathbf{X}$ from the unnormalized blur map $\sigma^{-1}(\mathbf{M})$ , resulting in $\mathrm{FFN}_{\mathrm{Conv}}(\sigma^{-1}(\mathbf{M})) \in \mathbb{R}^{h_x \times w_x \times q^2}$ . The Pad( $\cdot$ ) function reshapes $\mathrm{FFN}_{\mathrm{Conv}}(\sigma^{-1}(\mathbf{M}))$ into $\mathbb{R}^{h_x \times w_x \times q \times q}$ , and pads the reshaped $\mathrm{FFN}_{\mathrm{Conv}}(\sigma^{-1}(\mathbf{M}))$ to $\mathbb{R}^{(h_x \times w_x) \times (h_x \times w_x)}$ , for shape alignment issue. + +# 3.3. Network Training + +We optimize our deblurring network with the widely used Charbonnier loss. Moreover, we propose a blur-aware loss and a blur weighting loss to penalize the blurred regions in the restoration $\hat{\mathbf{I}}$ . + +Charbonnier loss. We encourage the restoration to be close to the ground truth by using Charbonnier loss [12, 43]. + +![](images/480aad7501ee7e8b7fcb54754c39b4d78d088b346ac26c3d6c0e4798592d9d0c.jpg) +(a) + +![](images/df89236f0ba3d0a983886e86e0fac8dc150f9b76ee15aaf584879ba6aa82d718.jpg) +(b) + +![](images/8838b231c703c986bc4d3b093dad7eb2534b56e2645f464ec2a5a15cd603a2de.jpg) +(c) +Figure 5. Example of deblurring with the blur map projected deblur kernel $\mathbf{O}$ . (a) and (b) are the left view of the DP pair and the ground truth sharp image. (c) and (d) are deblurred images with and without using $\mathbf{O}$ . Zoom in for better quality. + +![](images/53c8c1fc807ced4c5c6267718200456ec40493e492d5f7a61f420f3dd08a2a77.jpg) +(d) + +We empirically set $\varepsilon$ to $10^{-4}$ , and define + +$$ +\mathcal {L} _ {\text {c h a r}} = \sqrt {\| \mathbf {I} (i , j) - \hat {\mathbf {I}} (i , j) \| ^ {2} + \varepsilon^ {2}}. \tag {11} +$$ + +Blur Weighting Loss. We impose an adaptive penalty on the blurred regions of the DP pair $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ , considering the degree of blur measured by CLIP. It encourages the network to focus on the restoration of heavily blurred regions. We first normalize the deviations between $\hat{\mathbf{I}}$ and $\mathbf{I}$ by the deviations between $(\mathbf{B}_{\mathrm{L}},\mathbf{B}_{\mathrm{R}})$ and $\mathbf{I}$ , and then re-use the blur map to select blurred regions residing in the DP pair. + +$$ +\mathcal {L} _ {\mathrm {b w l}} = \frac {\left\| \mathbf {I} (i , j) - \hat {\mathbf {I}} (i , j) \right\| ^ {2}}{\left\| \hat {\mathbf {I}} (i , j) - \frac {1}{2} \left(\mathbf {B} _ {\mathrm {L}} (i , j) + \mathbf {B} _ {\mathrm {R}} (i , j)\right) \right\| ^ {2}} \cdot \mathbf {M} (i, j), \tag {12} +$$ + +where we use the blur map $\mathbf{M}$ (with soft logits) to reflect the blur levels detected by CLIP. + +Blur-aware Loss. We encourage $\hat{\mathbf{I}}$ to be sharp by leveraging the versatile knowledge contained in the CLIP. We generate the blur map $\hat{\mathbf{M}}$ by the blur-aware format for the restoration, and penalize the blur exhibited in $\hat{\mathbf{I}}$ . Similar to $\mathcal{L}_{\mathrm{bw1}}$ , we use the unnormalized version of the blur map, + +$$ +\mathcal {L} _ {\text {b a l}} = \hat {\mathbf {M}} (i, j). \tag {13} +$$ + +This encourages $\hat{\mathbf{M}}$ to be zero, i.e., no blur exists in the restoration. + +Total Loss. Our network is optimized by minimizing the following loss function + +$$ +\mathcal {L} _ {\text {t o t a l}} = \frac {1}{| \Omega |} \sum_ {(i, j) \in \Omega} \left(\mathcal {L} _ {\text {c h a r}} + \lambda_ {1} \mathcal {L} _ {\text {b a l}} + \lambda_ {2} \mathcal {L} _ {\text {b w l}}\right), \tag {14} +$$ + +where $\frac{1}{|\Omega|}$ is a normalization factor determined by the image domain size $|\Omega|$ , and $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters to balance losses. + +# 4. Experiments and Analysis + +Implementation Details. We implement our model under the PyTorch framework. We use the pre-trained and ViT-B/32 [8] architecture-based CLIP [30] for our blur map estimation strategy. We train the deblurring network in two stages for $6 \times 10^{6}$ steps in total with the Adam optimizer. + +Table 1. Quantitative comparisons on the DPD-blur dataset [1] composed of 37 indoor and 39 outdoor scenes. We respectively color the best and the second-best performance in red and blue. + +
MethodOutdoor ScenesIndoor ScenesAverage
PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓LPIPS↓
EBDB25.770.7720.0400.05121.250.5990.0580.08623.450.6830.0490.0670.336
DMENet25.500.7880.0380.05321.430.6440.0630.08423.410.7140.0510.0670.349
DPDNet27.480.8490.0290.04222.900.7260.0520.07125.130.7860.0410.0550.223
RDPD28.100.8430.0270.03922.820.7040.0530.07225.390.7720.0400.0530.255
DDDNet27.570.8330.0300.04123.280.7080.0500.06825.360.7680.0410.0540.197
IFAN28.660.8680.0250.03723.460.7430.0490.06725.990.8040.0370.0500.207
BAMBNet28.600.8720.0260.03724.300.7720.0450.06026.400.8210.0360.0470.141
DeepRFT28.480.8700.0250.03823.090.7360.0510.07025.710.8010.0370.0510.217
Restormer29.480.8950.0230.03423.970.7730.0470.06326.660.8330.0350.0460.155
K3DN29.870.8900.0220.03224.370.7800.0460.06027.060.8350.0340.0440.187
Ours (Small)29.550.8930.0210.03324.380.7730.0430.06526.910.8310.0320.0450.185
Ours (Large)29.880.8960.0200.03224.730.7860.0370.05927.240.8400.0280.0430.168
+ +Table 2. Quantitative comparisons on DDD-syn dataset. + +
MethodPSNR↑SSIM↑MAE↓RMSE_rel↓
EBDB26.480.6830.0360.047
DMENet30.140.9390.0170.031
DPDNet31.450.9260.0160.027
RDPD32.420.9120.0160.024
DDDNet33.210.9560.0100.022
IFAN34.180.9290.0120.020
BAMBNet35.900.9540.0120.016
DeepRFT36.520.9520.0100.015
Restormer36.440.9570.0100.015
K3DN38.230.9650.0080.012
Ours (Small)37.270.9590.0100.014
Ours (Large)38.620.9710.0080.012
+ +In each stage, the learning rate is decayed from $2 \times 10^{-4}$ to $10^{-6}$ by a cosine annealing strategy. In the first stage, we train the network for $5 \times 10^{6}$ steps, and only use the Charbonnier loss $\mathcal{L}_{\mathrm{char}}$ . In the second stage, the network is finetuned for $1 \times 10^{6}$ steps with $\mathcal{L}_{\mathrm{total}}$ , where we have $\lambda_{1} = 10^{-1}$ and $\lambda_{2} = 2 \times 10^{-1}$ . Code, results, and more details are given in our project page. + +Datasets. Our method is evaluated on two real-world DP defocus deblurring datasets, DPD-blur dataset [1] and DPD-disp dataset [27]. We follow [1] on the training and testing splits. For the DPD-disp dataset, we use it for testing only, re-using the checkpoints trained from the DPD-blur dataset as they do not provide ground-truth (GT) data or with limited dataset size. Moreover, two synthetic datasets, DDD-syn [26] and RDPD [2], are also used for evaluations. + +Evaluation Metrics. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean absolute error (MAE), and root mean squared relative error (RMSE_ref) as our evaluation metrics. + +Baseline Methods. We compare with SOTA DP defocus deblurring methods, i.e., EBDB [13], DMENet [15], + +Table 3. Quantitative comparisons on RDPD dataset. + +
MethodPSNR↑SSIM↑MAE↓RMSE rel↓
EBDB24.230.6370.0310.054
DMENet25.170.7310.0360.061
DPDNet29.840.8280.0320.055
RDPD31.090.8610.0250.028
DDDNet30.140.8400.0160.031
IFAN31.120.8650.0230.028
BAMBNet31.780.8670.0160.026
DeepRFT31.710.8790.0160.024
Restormer32.270.8710.0160.024
K3DN32.840.9050.0140.023
Ours (Small)32.980.8950.0140.022
Ours (Large)33.440.9070.0130.021
+ +DPDNet [1], RDPD [2], DDDNet [26], IFAN [16], BAMBNet [20], DeepRFT [22], Restormer [39], and K3DN [38]. + +# 4.1. Comparisons with State-of-the-Art + +Quantitative Comparison. We compare with the SOTA methods on the DPD-blur, DDD-syn, and RDPD datasets in Tab. 1, Tab. 2, and Tab. 3, respectively. Our observations are as follows: i) With our model (Small), we achieve similar restoration results with Restormer and K3DN, the third-best and second-best method, in all datasets. ii) When scaling to the large model, we consistently outperform all methods on all evaluation metrics. For example, our method, K3DN, and Restormer respectively achieve 27.24/27.06/26.66, 38.62/38.23/36.44, and 33.44/32.84/32.27 PSNR (dB) in the DPD-blur, DDD-syn, and RDPD datasets. Meanwhile, on a single RTX 4090 GPU, we remain to have a lower Infer. time (including CLIP) which is 0.042s and 0.054s for our small and large models, while K3DN/Restormer takes 0.198s/0.255s. + +Qualitative Comparison. We compare with state-of-the-art methods on the DPD-blur dataset (Fig. 6). As shown, our + +![](images/1a5e452037bfcf471d7ed6ed97b64d00f80d482743a1e3ad6d7b9d34937e419a.jpg) +Blurry Image + +![](images/1a577a1e892a8fb4e3e7924fa7e1e060664c57fcf2ca9b0ed80042cb1be77ba2.jpg) +GT Image + +![](images/4e95c219927e2e91924f97c7242859158592d75f84a6394d22288eba860ab7aa.jpg) + +![](images/e263042cb2f4553c514700a14933a30b450d50b806f7a7b80439df6f943c6067.jpg) +IFAN [16] + +![](images/ceb81e56acb118fae69918d30d66a3cef1c1de4338a9c09c30cab0d81d1ec79e.jpg) +DDDNet [26] + +![](images/593fa7f0cd33870b2370f7686d3d006831fff4436a047c349f44ed5ff953fbc2.jpg) +Blurry Image + +![](images/71419781ec3eb04f150dd9a5e0d695549ad8f71fac20fec5db2020fdef535aa2.jpg) + +![](images/1ab9e10670eae6a5762ffcdbd1f94d11fb463e0508bfb974ed139b904a3febde.jpg) + +![](images/415d9511d004b84ab0fbe979d1a9d97b307e976b9d99d47209ccbd86eda5a141.jpg) + +![](images/a38e83a781b12c8e4eee01d3784d2ce81064db85b5dc3eb35f35492a454af57f.jpg) +DeepRFT [22] +Blurry Image +DeepRFT [22] + +![](images/d00923ce4f24847248c6927585b5be800a6d3253de2729faa764a9b0ad1da0f9.jpg) +GT Image + +![](images/10c56ea97f9d68316bbdc7475b929b6f2caba95210b120acf0a3e447d2f67f52.jpg) +Restormer [39] +Blurry Image + +![](images/c1cba08fd1c437d5faf95cc4c61b001bbfec7982372d7fff8b5e1e2745bb5b7f.jpg) +K3DN [38] +IFAN [16] + +![](images/3deff5e3099745779cb9d782a3227a6d95e4b9a3129e3a16d70ece208f2f3ad0.jpg) +Ours +DDDNet [26] + +![](images/e6d883e27feceebacbcd4f766c2154c76680ba300f69277ab319b309a3c8e21d.jpg) +Figure 6. Qualitative comparison on the DPD-blur dataset [1]. We present the blurry image $(\mathbf{B}_{\mathrm{L}})$ in the first column, and the ground truth (GT) image of regions residing in the red bounding box is shown in the second column. The cropped blurred image of the red bounding box region is in the third column. + +![](images/0e2e56ebeb48ea5db23dc965bdb40ba0b1498af169039040fd80561846390f0d.jpg) +Restormer [39] + +![](images/a3ae7e7d550322b212db7574819d25fe9d264c840e13caffee136fba8e2865e6.jpg) +K3DN [38] + +![](images/3fe1437ecf8035033a05d2376d9f62d1fd4e4038bc29755523e7b1a9b1a2cddc.jpg) +Ours + +method performs sharper and clearer restorations of edges and texts. To further demonstrate the benefits of using CLIP to generate the blur mask unsupervisedly in our approach, we provide an example in Fig. 7, which preserves sharp regions (e.g., text 'University') in the input blur image effectively compared to the second-best baseline. + +![](images/49d024ddac1a93a08d0f3ae7a68f2ae2039e91b0ac1850f8d2dcdedd2fdd5610.jpg) +(a) Input +Figure 7. An example of the restoration results on the sharp region of the input DP image (left view). + +![](images/e3c81256f7e09608a955ef70e91d56ce26e088b522210737345b94a264481924.jpg) +(b) K3DN + +![](images/fbad7c91f69532a328961c39f21d686aa4cff8de248434c3a6ea5e0377a8515b.jpg) +(c) Restormer + +![](images/7d412d6a06994209392d2827b7c200204608fe6290bf8376b7f031b3173930e4.jpg) +(d) Ours + +# 4.2. Ablation Study + +We study model architecture, blur map estimation, and losses in this section. All experiments are performed in the DPD-blur dataset. + +Ablation of Model Architecture. We study the effective usage of blur map as prior knowledge for DP defocus deblurring (Tab. 6). To show the benefits of using blur maps, we consider four baselines: i) NAFNet [5], the baseline model; ii) Concat, concatenating the blur map with the DP pair before feeding to the model; iii) SA, using the self-attention layer [8] in Eq. 8 to the model only; iv) Concat+SA, combines the Concat and SA settings; v) BPA block, using kernel priors from the blur map to assist deblurring. Our model consistently outperforms all baselines, which indicates the effectiveness of using the blur map. + +Ablation of Losses. We investigate the proposed blur weighting loss $\mathcal{L}_{\mathrm{bw}}$ and blur-aware loss $\mathcal{L}_{\mathrm{bal}}$ (Tab. 7) with a baseline as the $\mathcal{L}_{\mathrm{char}}$ setting. We first use $\mathcal{L}_{\mathrm{bw}}$ and $\mathcal{L}_{\mathrm{bal}}$ + +to our baseline, NAFNet (i.e., not using our BPA block), that only uses $\mathcal{L}_{\mathrm{char}}$ , having 0.32 PSNR (dB) improvements. Secondly, we train our model (backbone + BPA block) with Charbonnier loss, resulting in 26.71/0.825 higher PSNR (dB)/SSIM. By using the $\mathcal{L}_{\mathrm{bwl}}$ or $\mathcal{L}_{\mathrm{bal}}$ , we respectively achieve 0.05/0.003 or 0.10/0.004 higher PSNR (dB)/SSIM. Combining all losses, we have the best restoration performance at 26.91/0.831 for PSNR (dB)/SSIM. Moreover, we verify the generalization ability of our $\mathcal{L}_{\mathrm{bwl}}$ and $\mathcal{L}_{\mathrm{bal}}$ losses on state-of-the-art networks, improving PSNR of Restormer and DeepRFT by 0.25 and 0.41 PSNR (dB). + +Ablation of Blur Map Estimation. We ablate with different prompts used to generate the blur map. The blur-aware and DP-aware prompts are presented in Tab. 4, and we show their impact on model performance with Tab. 5. We then ensemble the groups of blur-aware prompt and disparity-aware prompt individually (Ens. $^1$ ), leading to $26.80\mathrm{dB} / 0.829$ and $26.84 / 0.830$ PSNR (dB)/SSIM. By assembling all prompts (Ens. $^2$ ), we have $26.91 / 0.831$ PSNR (dB)/SSIM, consistently outperforming other approaches of generating blur masks. + +# 4.3. Discussion and Limitation + +Discussion. The reader may wonder about the performance of our unsupervised blur map estimation strategy. Therefore, we perform two qualitative evaluations on the DPD-disp dataset, where the GT disparity is included. We compare with BAMBNet [20] and DPD-disp [27]. + +First, we consider blur map estimation. To obtain the GT blur map, the pixel with disparity $d$ that satisfies $|d|\leq 1$ + +Table 4. Prompts used for blur map estimation. + +
Blur-aware PromptDP-aware Prompt
p1[A blurry image.][A symmetrical image.]
p2[An obscured image.][A horizontally symmetrical image.]
p3[An out of focus image.][A left-right symmetrical image.]
p4[An image lacking sharpness.][A symmetrical image at the pixel level.]
+ +Table 5. Restoration performance of our methods by using blur-aware or/and DP-aware prompts for blur map estimation. Ens. ${}^{1}$ denotes ensembling the group of blue-aware prompts and DP-aware prompts individually, and Ens. ${}^{2}$ ensembles all prompts. + +
Blur-aware PromptDP-aware Prompt
PSNR↑SSIM↑MAE↓RMSE_rel↓PSNR↑SSIM↑MAE↓RMSE_rel↓
p126.830.8290.0330.04526.820.8300.0330.046
p226.810.8290.0330.04626.830.8300.0320.046
p326.750.8280.0340.04626.820.8300.0330.046
p426.770.8290.0330.04626.850.8300.0320.045
Ens.126.800.8290.0330.04626.840.8300.0320.046
Ens.226.910.8310.0320.04526.910.8310.0320.045
+ +Table 6. Ablation of model components. + +
SettingPSNR↑SSIM↑MAE↓RMSE rel↓
i) NAFNet26.210.8170.0380.049
ii) Concat26.760.8260.0330.046
iii) SA26.590.8240.0360.047
iv) Concat + SA26.800.8290.0320.046
v) BPA block26.910.8310.0320.045
+ +Table 7. The effectiveness of losses. + +
BPA blockLossPSNR↑SSIM↑MAE↓RMSE rel↓
XLchar26.210.8170.0380.049
XLtotal26.530.8220.0360.047
✓Lchar26.710.8250.0340.046
✓Lchar + Lbwl26.760.8280.0330.046
✓Lchar + Lbal26.810.8290.0320.046
✓Ltotal26.910.8310.0320.045
+ +is used as the sharp pixel, reversely, it is a buried pixel. Our method, BAMBNet [20] and DPD-disp [27] respectively have $89.8\% / 70.2\% / 88.3\%$ accuracy (the higher, the better), showing the effectiveness of our method. + +Second, we explore disparity estimation. The log-its of our unnormalized blurred map are potentially linearly correlated with the ground-truth disparity. Therefore, we compare with the state-of-the-art DP disparity estimation method, DPD-disp [27]. Following [9, 27], we use AI(1), AI(2), and $1 - |\rho_s|$ as the evaluation metrics (the lower, the better). The AI(1), AI(2), and $1 - |\rho_s|$ of our method, BAMBNet, and DPD-disp are 0.063/0.106/0.069, 0.130/0.184/0.096, 0.306/0.772/0.0159. Though our blur map is estimated with a single forward pass, we still achieve a reasonable performance in the disparity estimation. This potentially explains the high performance of our DP defo + +cus deblurring network, as obtaining the disparity prior benefits the defocus task. + +Limitation. As the image encoder of CLIP is trained under a classification-alike pre-training task, it will focus more on local details that help identify its scene class. Textureless regions would produce non-significant semantic reactions, e.g., large areas of wall, ceiling, or floor, and reduce the quality of the estimated blur map. It is possible to use a vision language framework that drives more sensitivity to regions with less texture. + +# 5. Conclusion and Broader Impact + +In this paper, we study end-to-end defocus deblurring of DP images, leveraging the blur-related prior knowledge from the CLIP. We first estimate a blur map with an ensemble of blur-aware and DP-aware strategies, and then use the estimated blur map as the deblurring kernel prior to restoring the DP image. We also propose a blur-aware and blur-weighting loss to regularize the restorations of the DP images during training, by distilling the knowledge of blur from CLIP. In extensive experiments, our method outperforms past works by a large margin in quantitative and qualitative restoration performance. + +Broader Impact. The proposed DP-aware strategy is promising to apply and extend CLIP to diverse zero-shot stereo vision tasks. We hope it will motivate future work. + +Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045, and Miaomiao Liu was supported by the ARC fellowship and Discovery Project grant (DE180100628, DP200102274). + +# References + +[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 1, 2, 4, 6, 7 +[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 3, 6 +[3] Abdullah Abuolaim, Mahmoud Affi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2, 4 +[4] Dylan Auty and Krystian Mikolajczyk. Learning to prompt clip for monocular depth estimation: Exploring the limits of human language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2039-2047, 2023. 2 +[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022. 5, 7 +[6] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 5, 7 +[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7628-7637, 2019. 1, 3, 8 +[10] Xueting Hu, Ce Zhang, Yi Zhang, Bowen Hai, Ke Yu, and Zhihai He. Learning to adapt clip for few-shot monocular depth estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5594-5603, 2024. 2 +[11] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22157-22167, 2023. 2 +[12] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale + +progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 5 +[13] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, 2017. 6 +[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 1 +[15] Junyong Lee, Sungkil Lee, Sunghyun Cho, and Seungyong Lee. Deep defocus map estimation using domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12222-12230, 2019. 6 +[16] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 6, 7 +[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 2 +[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 5 +[19] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1495-1503, 2023. 1 +[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Bambnet: A blur-aware multi-branch network for dual-pixel defocus deblurring. IEEE/CAA Journal of Automatica Sinica, 9(5):878-892, 2022. 1, 2, 6, 7, 8 +[21] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2 +[22] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the 37th AAAI Conference on Artificial Intelligence, 2023. 6, 7 +[23] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. arXiv preprint arXiv:2306.07921, 2023. 1 +[24] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1 +[25] Liyuan Pan, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Phase-only image based kernel estimation for single image blind deblurring. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 6034-6043, 2019. 2 +[26] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 3, 6, 7 +[27] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 3, 4, 6, 7, 8 +[28] Yuhui Quan, Zicong Wu, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2023. 1 +[29] Yuhui Quan, Xin Yao, and Hui Ji. Single image defocus deblurring via implicit neural inverse kernels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12600-12610, 2023. 1 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5 +[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 1, 2 +[32] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 1 +[33] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 2622-2630. IEEE, 2021. 1, 2 +[34] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2, 3 +[35] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. arXiv preprint arXiv:2312.01381, 2023. 2 +[36] Shaoliang Yang, Dongming Zhou, Jinde Cao, and Yanbu Guo. Lightingnet: An integrated learning method for low-light image enhancement. IEEE Transactions on Computational Imaging, 9:29-42, 2023. 5 +[37] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 1 +[38] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13272, 2023. 1, 2, 6, 7 +[39] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 5, 6, 7 +[40] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 1, 2 +[41] Renrui Zhang, Ziyao Zeng, Ziyu Guo, and Yafeng Li. Can language understand depth? In Proceedings of the 30th ACM International Conference on Multimedia, pages 6868-6874, 2022. 2, 3 +[42] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du2net: Learning depth estimation from dual-cameras and dual-pixels. In European Conference on Computer Vision, pages 582-598, 2020. 1 +[43] Qian Zhao, Hao Yang, Dongming Zhou, and Jinde Cao. Rethinking image deblurring via cnn-transformer multiscale hybrid architecture. IEEE Transactions on Instrumentation and Measurement, 72:1-15, 2022. 5 +[44] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175–11185, 2023. 2 \ No newline at end of file diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/images.zip b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ae22006d58211c3d06881f258100b0727203daf6 --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a122ebb7761aa8950210e87e2e94a87036e7dada0098df663f94070fd9149bd0 +size 673349 diff --git a/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/layout.json b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43a98d2e459ad374bee9a6f8f1f16ae2c81e0d14 --- /dev/null +++ b/2024/LDP_ Language-driven Dual-Pixel Image Defocus Deblurring Network/layout.json @@ -0,0 +1,13124 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 515, + 121 + ], + "type": "text", + "content": "LDP: Language-driven Dual-Pixel Image Defocus Deblurring Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "spans": [ + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": "Hao Yang" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": ", Liyuan Pan" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": ", Yan Yang" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": ", Richard Hartley" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": ", and Miaomiao Liu" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": "Beijing Institute of Technology \n" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 142, + 512, + 172 + ], + "type": "text", + "content": "Australian National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 533, + 186 + ], + "type": "text", + "content": "{hao.yang, liyuan.pan}@bit.edu.cn, {yan.yang, richard.hartley, miaomiao.liu}@anu.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 251, + 290, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 455 + ], + "type": "text", + "content": "Recovering sharp images from dual-pixel (DP) pairs with disparity-dependent blur is a challenging task. Existing blur map-based deblurring methods have demonstrated promising results. In this paper, we propose, to the best of our knowledge, the first framework that introduces the contrastive language-image pre-training framework (CLIP) to accurately estimate the blur map from a DP pair unsupervisedly. To achieve this, we first carefully design text prompts to enable CLIP to understand blur-related geometric prior knowledge from the DP pair. Then, we propose a format to input a stereo DP pair to CLIP without any fine-tuning, despite the fact that CLIP is pre-trained on monocular images. Given the estimated blur map, we introduce a blur-prior attention block, a blur-weighting loss, and a blur-aware loss to recover the all-in-focus image. Our method achieves state-of-the-art performance in extensive experiments (see Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 128, + 491 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 500, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 679 + ], + "type": "text", + "content": "Defocus deblurring from a single image is an ill-posed problem [19, 28, 29, 33]. Compared to methods that restore an all-in-focus image directly [1, 2], blur map-based deblurring methods have demonstrated promising results [16, 26]. To simplify the blur removal process, several works [20, 34, 38] have utilized recent innovative dual-pixel (DP) sensors [6] to estimate blur maps, since the DP sensor captures a two-view image pair (a DP pair) in a single shot. The DP pair can be treated as images from a two-sample light field camera or a stereo system [24] with a tiny baseline [27, 34], which is beneficial for disparity estimation [1, 14, 23, 42]. Based on the DP image pair formation process, the disparity is geometrically related to the blur map, i.e., the degree of defocus blur proportional to disparity magnitudes [9]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 680, + 287, + 693 + ], + "type": "text", + "content": "For estimating the blur map (defocus map or disparity" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 547, + 337 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 547, + 337 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 547, + 337 + ], + "type": "image", + "image_path": "4289276f10163e17c2dd9aa619e2d3d885516c5673317bc1d8f7f482698b7003.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 342, + 547, + 407 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 407 + ], + "type": "text", + "content": "Figure 1. Comparison of performance in terms of PSNR and Inference time. Our small and large models are respectively denoted by the Blue and Red cycles. The Green triangles represent previous state-of-the-art methods, and each of them is annotated with the method name, publication venue, and year. This figure is best viewed in color on the screen." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 420, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 540 + ], + "type": "text", + "content": "map), previous works either require extra data [3, 26], such as synthetic data, for supervision signals, or pre-calibrated blur kernels [34]. To avoid using extra information, several works explored unsupervised blur map estimation frameworks [16, 20]. Lee et al. [16] implicitly estimate the disparity of the left/right views using a spatial transformer, yet only in feature space, as well as Yang et al. [38]. Liang et al. [20] use fixed network branches for layer-wise defocus map estimation and deblurring, where the limited network branch number restricts the potential of its accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 651 + ], + "type": "text", + "content": "Recently, the contrastive language-image pre-training framework (CLIP) [30] has accomplished extraordinary success for vision tasks, such as semantic segmentation [31, 37], object detection [32], and point cloud understanding [40]. Therefore, a question naturally arises, can we avoid the cost of collecting data and model designing, to estimate the blur map unsupervisedly by using CLIP? Unfortunately, using semantic knowledge from CLIP to handle low-level vision tasks is underexplored." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "In this paper, for the first time, we explore the relationship between language and blur map estimation by applying CLIP. Note that CLIP is trained to match a monocular image with its corresponding language descriptions. Given a blurred image, it is intuitive to match it with a corre" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 142, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 142, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 142, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24078" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "sponding text [A blurry image.], which allows the language model to find the blurred regions in an image. However, it is hard to quantize the blur level of an image without knowing its associated latent sharp one [25]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 239 + ], + "type": "text", + "content": "Fortunately, we can leverage the capabilities of DP sensors. With DP pairs, blur quantization becomes possible without the need for its associated sharp image. The formation of DP ensures no disparity between left/right views for pixels in sharp regions, while a noticeable disparity exists for pixels in blurred regions. In essence, the detectable differences (disparities) in left/right views of a DP pair directly indicate the degree of defocus blurs. Consequently, we can quantify the variation between the left and right views, treating it as a valuable cue for estimating a blur map." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 240, + 286, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 286, + 359 + ], + "type": "text", + "content": "However, CLIP still finds it challenging to simultaneously explore stereo DP pairs with left/right view images. To address this issue, we propose to explore the DP pair and design an 'image and text' format to better capture the blurs in the image. Specifically, we propose to create a new image from the DP pair and translate the description about the 'blurriness' of the image to measure the 'symmetry' of the newly formed image, which is based on the DP image formation process. The freshly designed image-text format allows the CLIP to estimate the blur map easily (see Sec. 3.1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 286, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 443 + ], + "type": "text", + "content": "Given the estimated blur map, we then restore the all-in-focus image from the DP pair. By observing that the attention map of the self-attention mechanism [8] can be used in the dynamic DP deblurring process, we introduce a blur prior attention (BPA) that modulates the attention map with the estimated blur map which provides prior knowledge of deblurring kernels (see Sec. 3.2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 444, + 286, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 444, + 286, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 444, + 286, + 538 + ], + "type": "text", + "content": "To regulate the restoration, we specifically design two losses that are based on CLIP. Considering the network should focus more on heavily blurred regions, we propose a blur-weighting loss that imposes an adaptive penalty proportional to the blur degrees provided by the estimated blur map from CLIP. Then, to guarantee the restored image is sharp, we use CLIP to re-check the restored image and formulate a blur-aware loss (see Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 540, + 244, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 540, + 244, + 550 + ], + "spans": [ + { + "bbox": [ + 59, + 540, + 244, + 550 + ], + "type": "text", + "content": "Our contributions are summarised as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 286, + 654 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 551, + 286, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 286, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 286, + 585 + ], + "type": "text", + "content": "- We propose a Language-driven DP (LDP) defocus deblurring framework, by exploring the potential of CLIP in the low-level vision task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 586, + 286, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 286, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 286, + 619 + ], + "type": "text", + "content": "- We design an image-text format for blur map estimation based on the geometric relationship between the blur and disparity of the DP pair." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 619, + 286, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 286, + 654 + ], + "type": "text", + "content": "- We propose a blur prior attention (BPA) block, a blur-weighting loss, and a blur-aware loss to encourage sharp restorations of the DP pair." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "We evaluate the proposed framework extensively on standard benchmark datasets. Though using the largescale foundation model, CLIP, for blur map estimation, our method achieves Pareto-optimality between deblurring performance and inference speed (refer to Fig. 1 for an" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 425, + 84 + ], + "type": "text", + "content": "overview of the comparison)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 95, + 392, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 95, + 392, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 95, + 392, + 108 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 116, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 545, + 380 + ], + "type": "text", + "content": "DP Defocus Deblurring. DP-based deblurring methods can be divided into single- and multi-task manners. For single-task based deblurring [1-3], methods apply different kernels at different regions of the DP pair to handle the spatially varying defocus deblur. This can usually be achieved by using parallel atrous convolutions [33], attention mechanisms [39], or deformable convolutions [17]. For multi-task deblurring methods, the blur/defocus/disparity map is auxiliarily estimated to benefit DP defocus deblurring. Pan et al. [26] and Xin et al. [34] use the estimated DP disparity to regularise the restoration through reblurring; however, with extra data or pre-calibrate kernel. Several works explored unsupervised defocus map (or disparity) estimation. Lee et al. [16] and Yang et al. [38] estimate defocus disparity only in feature space without explicitly showing a defocus map. Liang et al. [20] optimize a multi-channel defocus map based on [27]. However, hyperparameters, such as the defocus threshold and the number of channels, are preset, limiting blur map estimation accuracy. Unlike past works, by transferring the blur-related geometric prior knowledge into semantic prompts, we unsupervisedly estimate the blur map from CLIP to benefit defocus deblurring." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 381, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 561 + ], + "type": "text", + "content": "Language-driven Vision. Common objectives in visual-language learning are aligning the text with a paired image in the embedding space [21, 30, 35]. The CLIP, one of the milestone achievement in the area, demonstrate that language-independent computer vision tasks can be substantially benefited from visual-language learning, by achieving SOTA zero-shot classification performance in the ImageNet-1K benchmark [7]. Successive works have extended the zero-shot learning capability of CLIP to semantic segmentation [31, 44], monocular depth estimation [4, 10, 41], and point clouds [11, 40]. However, none of the existing works explores the zero-shot learning capability of CLIP in stereo/DP pairs. This paper proposes the first approach to use the monocular image-trained CLIP in zero-shot blur map segmentation of DP pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 573, + 361, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 361, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 361, + 585 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "Let us now introduce our language-driven dual-pixel image defocus deblurring framework. Given a defocus blurred DP pair " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": ", we aim to restore an all-in-focus image " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " are the left and right view of a DP pair and " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}},\\mathbf{I}\\in \\mathbb{R}^{h\\times w\\times 3}" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": " define the image height and width, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We first design a DP-aware prompt to adapt CLIP to output a blur map, " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{h \\times w \\times 1}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ", which quantizes disparity unsupervisedly between " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": ". Then, we propose a BPA block based network which employs the blur" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24079" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 69, + 509, + 202 + ], + "blocks": [ + { + "bbox": [ + 86, + 69, + 509, + 202 + ], + "lines": [ + { + "bbox": [ + 86, + 69, + 509, + 202 + ], + "spans": [ + { + "bbox": [ + 86, + 69, + 509, + 202 + ], + "type": "image", + "image_path": "990fe983387e39d2f8b277fcf1f9eba3549aaf603c220bf2942e5ea3d19158d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. Overall architecture. Given a DP pair " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ", we first generate a blur map " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " by using our unsupervised blur map estimation strategy which is based on the CLIP with assembling format. We then forward the DP pair " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " to our deblurring backbone composed of an encoder and a decoder. Meanwhile, we have multiple BPA blocks, using the blur map " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " as deblurring kernel prior to refine the intermediate backbone feature embeddings. In the following, we have the restoration " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ". The deblurring network is optimized with three losses, charbonnier loss " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{char}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ", blur weighting loss " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bwl}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ", and blur aware loss " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bal}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{char}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " penalizes deviations from " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " in the resorption " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bwl}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bal}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " regularize the restoration by imposing an adaptive penalty on the blurred regions of the DP pair, and leveraging unsupervised blur map estimation strategy to encourage the restoration to be sharp, i.e., the blur map " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " for the restoration " + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 204, + 547, + 282 + ], + "type": "text", + "content": " is zero value filled." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": "map " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": " as deblurring priors to restore a sharp image " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": ". To regularize the restoration of " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": ", a blur-weighting loss " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bwl}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": " and a blur-aware loss " + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bal}}" + }, + { + "bbox": [ + 46, + 288, + 289, + 361 + ], + "type": "text", + "content": " are also introduced that were underpinned by our unsupervised blur map estimation strategy. The overall framework is shown in Fig. 2. Details are presented below." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 366, + 237, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 237, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 237, + 380 + ], + "type": "text", + "content": "3.1. Unsupervised Blur Map Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": "Preliminary. CLIP respectively embeds an image and a text description into a common semantic space of dimension " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": " with an image encoder and a text encoder, obtaining image embedding " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^c" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": ". By optimizing a contrastive loss, the embeddings of the aligned image and text pair, " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\Omega_a" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": ", stay similar, i.e., maximizing the semantic similarity by " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\max_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_a} \\sin(\\mathbf{f}, \\mathbf{t})" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\sin(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": " is the cosine similarity metric. Meanwhile, the semantics of misaligned pairs, " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\Omega_{na}" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": " are pushed away from each other, i.e., " + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\min_{\\mathbf{f}, \\mathbf{t} \\in \\Omega_{na}} \\sin(\\mathbf{f}, \\mathbf{t})" + }, + { + "bbox": [ + 46, + 385, + 287, + 540 + ], + "type": "text", + "content": ". With comprehensive training data, CLIP can be leveraged to align images with pre-defined prompt templates composed of open-world text descriptions for zero-shot learning [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "text", + "content": "For a DP pair, the disparity " + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "text", + "content": " for a pixel in the left view is defined by its location " + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "text", + "content": " and that of its corresponding pixel at " + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "inline_equation", + "content": "(i,k)" + }, + { + "bbox": [ + 46, + 540, + 287, + 577 + ], + "type": "text", + "content": " in the right view." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 577, + 268, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 577, + 268, + 590 + ], + "spans": [ + { + "bbox": [ + 58, + 577, + 268, + 590 + ], + "type": "text", + "content": "In the literature, this DP disparity [27] is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 596, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 596, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 52, + 596, + 287, + 611 + ], + "type": "interline_equation", + "content": "d = j - k = A (r (z)) \\cdot B (z), \\mathbf {B} _ {\\mathrm {L}} (i, j) = \\mathbf {B} _ {\\mathrm {R}} (i, k), \\tag {1}", + "image_path": "6052a7dd38951648f31a839c71d9a4a3ae77cd454083b94ff4a6e94645abc2f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}(i,j)" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}(i,k)" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " are the intensity of the two matching points, " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "r(\\cdot)" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " is a function of the radius of the circle of confusion, " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " is the point depth, and " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "A(r(z))" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "inline_equation", + "content": "B(z)" + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": " are lens-dependent constants [2, 9, 26, 34]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Analysis. Leveraging the CLIP, zero-shot blur map estimation can be generally divided into three steps. i) Image Encoding. Considering the image encoder of the CLIP is trained with monocular RGB images, a function " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\eta (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "content": "required to create a new monocular image from the DP image pair " + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "content": ". To obtain the dense embedding for the blur map estimation task, we follow [41] to pop out the final pooling layer from the image encoder, yielding an embedding matrix " + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{F}\\in \\mathbb{R}^{h_s\\times w_s\\times c}" + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "inline_equation", + "content": "h_s" + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 288, + 545, + 373 + ], + "type": "text", + "content": " are the height and width for embedding of the newly created image, respectively. Mathematically, we have" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 384, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 384, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 354, + 384, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\operatorname {I m a g e E n c o d e r} \\left(\\eta \\left(\\mathbf {B} _ {\\mathrm {L}}, \\mathbf {B} _ {\\mathrm {R}}\\right)\\right). \\tag {2}", + "image_path": "e65d8413572546c73ac35f6ca575e3ede9905fabf82b4608fb251ecb433d0987.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "text", + "content": "Each pixel of " + }, + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "text", + "content": " has encoded the local semantic information for a corresponding region in " + }, + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 304, + 409, + 547, + 458 + ], + "type": "text", + "content": ". ii) Text Encoding. A text encoder is used to embed the semantics of a prompt p describing the blur into the text embedding," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 470, + 545, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 470, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 380, + 470, + 545, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {t} = \\operatorname {T e x t E n c o d e r} (\\mathfrak {p}). \\tag {3}", + "image_path": "6e8cf63bef08207dbef680081631ae46baf9f070d4dba593fe050457aaef5e88.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": "iii) Blur map estimation. Let " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(i,j)" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " be a c-dimensional embedding at position " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots ,\\mathrm{h}_{\\mathrm{s}}\\}" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "j\\in \\{1,\\dots ,\\mathrm{w}_{\\mathrm{s}}\\}" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": ". The embedding " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(i,j)" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " and the blur semantic-based text embedding " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " are expected to be similar if the pixel at " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " is blurred. Therefore, the blur mask " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 304, + 495, + 547, + 568 + ], + "type": "text", + "content": " is obtained by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 357, + 578, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 578, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 357, + 578, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\mathbf {F}} (i, j) = \\sigma (\\operatorname {s i m} (\\mathbf {F} (i, j), \\mathbf {t})) . \\tag {4}", + "image_path": "ca7912661ca6038fe89892f786065a7d8fe4250ed7048d4d9362b2a5c9c521d0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": "Note, " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " normalizes the cosine similarity " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathfrak{sim}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " output to a probability space, to reflect the degree of blur at the position. Here, " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " is sampled to have the same resolution as " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": ", returning " + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in \\mathbb{R}^{h\\times w\\times 1}" + }, + { + "bbox": [ + 304, + 604, + 545, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "text", + "content": "Given the analysis above, our goal is to develop text prompts " + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}" + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "text", + "content": ", while adapting the function " + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "inline_equation", + "content": "\\eta(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "text", + "content": ", to enable the use of semantic information from CLIP to generate the blur map " + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 653, + 545, + 725 + ], + "type": "text", + "content": " in an unsupervised manner. In particular, we explore the blur-aware, DP-aware, and ensembling formats with detailed presented below." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24080" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 283, + 216 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 283, + 216 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 283, + 216 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 283, + 216 + ], + "type": "image", + "image_path": "18d45b8a20b3eaa4186ec61cbe6d094c70506c392e53a5046bae28f16976d1a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": "Figure 3. The pipeline of our DP-aware format for obtaining the blur map " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": ". Given a DP pair " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": ", the function " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " concatenates the left view " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " and with horizontally flipped right view " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}^{h}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " for obtaining the " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " from the CLIP image encoder. We then apply " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\alpha (F)" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " to averagely pool the feature map according to the transformation imposed on the input DP pair, resulting in a new feature embedding with the same aspect ratio of the input DP pair. Meanwhile, we set the prompts " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " to describe the symmetry of " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": ", obtaining text embedding " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": ". We finally have the blur map " + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 231, + 290, + 342 + ], + "type": "text", + "content": " by computing Eq. 6 and performing sampling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": "Blur-aware Format. An intuitive approach is setting a prompt " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": " that describes the blur to query a blur map " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [\\mathrm{A~blurry~image}.]" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\eta(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": " is then defined to synthesize a center view of the DP pair by averaging the left view " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": " and right view " + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 46, + 352, + 287, + 460 + ], + "type": "text", + "content": ", as the average of two sub-images is equivalent to the image captured by a regular sensor [3, 27]. However, the disparity information among the DP pair is neglected, resulting in a sub-optimal blur map. Refer to Fig. 4 for comparisons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "text", + "content": "DP-aware Format. Based on the DP image formation process, the disparity is proportional to the degree of blur exhibited in the DP view [1]. We thus propose to guide CLIP to understand the disparity exhibited between " + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "text", + "content": " which eventually leads to better mask (M) generation. To this end, we first formulate the disparity in the DP pair and then study the disparity-aware prompt and function " + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\eta (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 463, + 287, + 582 + ], + "type": "text", + "content": " designing. Such design leads us to query degrees of displacement between the left and right view from the CLIP, without any fine-tuning, to obtain the blur map M." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 691 + ], + "type": "text", + "content": "Note that the CLIP is the vision-text alignment model trained with monocular image input. To extract the displacement between left/right views of the DP pair, we explore several input-prompt formats: i) Concatenating the left and right views horizontally, and prompting their difference, e.g., [The left and right of the image are different.]; ii) Prompting the horizontal symmetry between the left and right views, e.g., [A horizontally symmetric image.] (Fig. 3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 691, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 691, + 287, + 714 + ], + "type": "text", + "content": "Empirically, we achieve the best performance with the second prompt format. When describing the symmetry in" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 365, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 365, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 365, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 365, + 148 + ], + "type": "image", + "image_path": "36370877f7d879fc524e4ff8f9994de1b61705a140a2a3adf717c1156d906117.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 152, + 342, + 159 + ], + "lines": [ + { + "bbox": [ + 331, + 152, + 342, + 159 + ], + "spans": [ + { + "bbox": [ + 331, + 152, + 342, + 159 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 366, + 70, + 425, + 148 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 425, + 148 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 425, + 148 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 425, + 148 + ], + "type": "image", + "image_path": "2adfbde67b3773aad6a74b04d1253cf65b912abc73088b937b9bfadbd6a43017.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 152, + 400, + 159 + ], + "lines": [ + { + "bbox": [ + 390, + 152, + 400, + 159 + ], + "spans": [ + { + "bbox": [ + 390, + 152, + 400, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 484, + 148 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 484, + 148 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 484, + 148 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 484, + 148 + ], + "type": "image", + "image_path": "b0872cd650385ca4ec8a5634d760a6ffa4998117d1739009563c34e8098ffa4a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 152, + 460, + 159 + ], + "lines": [ + { + "bbox": [ + 449, + 152, + 460, + 159 + ], + "spans": [ + { + "bbox": [ + 449, + 152, + 460, + 159 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 161, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 304, + 161, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 547, + 227 + ], + "type": "text", + "content": "Figure 4. Example of blur map estimation (the darker, the sharper). (a) The left view of the DP image. (b) and (c) Blur maps obtained with prompts of the blur-aware and DP-aware formats. (d) The blur maps obtained with the ensembling-based format, i.e., all eight blur-aware and DP-aware prompts are used (please refer to Tab. 4 for details)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 484, + 70, + 543, + 148 + ], + "blocks": [ + { + "bbox": [ + 484, + 70, + 543, + 148 + ], + "lines": [ + { + "bbox": [ + 484, + 70, + 543, + 148 + ], + "spans": [ + { + "bbox": [ + 484, + 70, + 543, + 148 + ], + "type": "image", + "image_path": "3b01f240f5b9ae389abd9aee3a971c85e3eec945b179276d04374353f0970329.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 152, + 518, + 159 + ], + "lines": [ + { + "bbox": [ + 509, + 152, + 518, + 159 + ], + "spans": [ + { + "bbox": [ + 509, + 152, + 518, + 159 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": "the prompt " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathfrak{p}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ", we concatenate the left view " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " with a horizontally flipped right view " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}^{h} = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " along the width axial in the " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\eta (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ". We then obtained image features " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathrm{ImageEncoder}\\bigl (\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})\\bigr)" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ". To transform " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " to the same aspect ratio of " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{R}}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ", we average the corresponding pixel-wise features based on the transformation in " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " and have " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "\\alpha (\\mathbf{F})" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "i\\in \\{1,\\dots ,\\mathsf{h}_{\\mathsf{w}}\\}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "inline_equation", + "content": "j\\in \\{1,\\dots ,\\lfloor \\frac{1}{2}\\mathsf{w_s}\\rfloor \\}" + }, + { + "bbox": [ + 304, + 238, + 545, + 331 + ], + "type": "text", + "content": ", we define" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 348, + 340, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 340, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 348, + 340, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\alpha (\\mathbf {F}) (i, j) = \\frac {1}{2} \\mathbf {F} (i, j) + \\frac {1}{2} \\mathbf {F} (i, w _ {s} - j). \\tag {5}", + "image_path": "951920034d9c04c899d8126586a6eb887987fd44ba86f61eb418b81c064bdf4b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 371, + 485, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 485, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 485, + 384 + ], + "type": "text", + "content": "Back to the left-hand side of Eq. 4, we have" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 392, + 545, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 392, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 545, + 416 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {\\mathbf {F}} (i, j) = 1 - \\mathbf {M} _ {\\mathbf {F}} ^ {s} (i, j) = 1 - \\sigma \\left(\\operatorname {s i m} (\\alpha (\\mathbf {F}) (i, j), \\mathbf {t} ^ {s})\\right). \\tag {6}", + "image_path": "1808ba96020dea8911be6a2dcf41ee16138615e183390f73f1d376af31e605da.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^s" + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": " is the text embedding of the prompt describing symmetry. Note that the high similarity value, i.e. " + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{F}}^{s}(i,j) = \\sigma \\big(\\mathrm{sim}(\\alpha (\\mathbf{F})(i,j),\\mathbf{t}^{s})\\big)" + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": " indicates pixels in sharp regions. Therefore the blur map is derived as " + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "inline_equation", + "content": "1 - \\mathbf{M}_{\\mathbf{F}}^{s}(i,j)" + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": " in Eq. 6. Similarly, the " + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathbf{F}}" + }, + { + "bbox": [ + 304, + 427, + 545, + 511 + ], + "type": "text", + "content": " is sampled to the same resolution of the DP pair, returning M. For this blur map estimation strategy, we conclude the lemma below," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 514, + 547, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 551 + ], + "type": "text", + "content": "Lemma 3.1. A pixel is sharp if and only if it is horizontally symmetric in the left view and the horizontally flipped right view, otherwise, the pixel is blurred." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "text", + "content": "Proof. Following Eq. (1), we have " + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}(i,j)" + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "text", + "content": " being sharp, given " + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{L}}(i,j) = \\mathrm{Flip}(\\mathbf{B}_{\\mathrm{R}})(i,\\mathsf{w} - j)" + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "text", + "content": ". In other words, we have " + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "inline_equation", + "content": "\\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,j) = \\eta (\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}}^{h})(i,\\mathsf{w}_{\\mathsf{s}} - j)" + }, + { + "bbox": [ + 304, + 567, + 545, + 627 + ], + "type": "text", + "content": ", demonstrating the horizontal symmetry. Reversely, the horizontal symmetry breaks once the disparity/blur occurs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 712 + ], + "type": "text", + "content": "Ensembling format. Given the analysis that the degradations in the DP pair can be described by both the blurriness and symmetricity, we thus propose to obtain the mask by assembling the blur-aware and DP-aware formats jointly. Please refer to Tab. 4 for details of the 8 prompts designed from blur-aware and DP-aware formats." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24081" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "text", + "content": "3.2. Network" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": "We train an encoder-decoder based deblurring network for restoring the DP pair " + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": " into a sharp image " + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": " with the derived blur map " + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": ". Our network consists of a backbone network based on [5] and a BPA block to interleave with the backbone network layers. In particular, the BPA block uses the obtained blur map " + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": " as deblur kernel prior to help the restoration of " + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 46, + 90, + 287, + 174 + ], + "type": "text", + "content": " on the fly." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": "Self-attention. The self-attention (SA) mechanism is widely used in the image restoration tasks [18, 36, 39]. We first briefly review it. Let " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{h_x \\times w_x \\times c_x}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": " be the intermediate feature embedding of a blur image from the backbone encoder, where " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "h_x" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "w_x" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "c_x" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": " are height, width, and channel of the feature. The self-attention is input with query " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": " that are linearly projected from " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ", and then calculate an attention map between " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": " for adaptively aggregating " + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 175, + 288, + 283 + ], + "type": "text", + "content": ". This process is defined as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 293, + 287, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 293, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 70, + 293, + 287, + 307 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {Q}}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {K}}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {\\mathbf {V}}, \\tag {7}", + "image_path": "50865a4810d9b1f2ff88fa7d5795df913dd5a12f8d042bff31a84c00d07fbd2c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 308, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 308, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 70, + 308, + 287, + 323 + ], + "type": "text", + "content": "Attention " + }, + { + "bbox": [ + 70, + 308, + 287, + 323 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta)\\mathbf{V}" + }, + { + "bbox": [ + 70, + 308, + 287, + 323 + ], + "type": "text", + "content": " (8)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{h_{\\mathbf{x}} \\times w_{\\mathbf{x}} \\times c_{\\mathbf{x}}}" + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "text", + "content": " are learnable weight matrices, " + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "text", + "content": " is a normalization factor that is typically set to the square root of the latent dimension, and " + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "inline_equation", + "content": "\\operatorname{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta) \\in \\mathbb{R}^{(h_{\\mathbf{x}} \\times w_{\\mathbf{x}}) \\times (h_{\\mathbf{x}} \\times w_{\\mathbf{x}})}" + }, + { + "bbox": [ + 46, + 333, + 287, + 382 + ], + "type": "text", + "content": " is attention map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "text", + "content": "BPA Block. The attention map can be considered as an adaptive deblur kernel prior for restoring the DP pair in the feature space. We are motivated to modulate the attention map by supplying the blur map " + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "text", + "content": ", utilizing blur-related prior knowledge to additionally assign a " + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\times \\mathbf{q}" + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "text", + "content": " deblur kernel for each region in " + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 384, + 287, + 467 + ], + "type": "text", + "content": " to aggregate a sharp feature embedding from its neighboring regions. We re-write the Eq. 8 as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 476, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 287, + 490 + ], + "type": "text", + "content": "Attention " + }, + { + "bbox": [ + 56, + 476, + 287, + 490 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q},\\mathbf{K},\\mathbf{V}) = \\mathrm{Softmax}(\\mathbf{Q}\\mathbf{K}^{\\top} / \\beta +\\mathbf{O})\\mathbf{V}" + }, + { + "bbox": [ + 56, + 476, + 287, + 490 + ], + "type": "text", + "content": " (9)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\mathbf {O} = \\operatorname {P a d} \\left(\\operatorname {F F N} _ {\\operatorname {C o n v}} \\left(\\sigma^ {- 1} (\\mathbf {M})\\right)\\right), \\tag {10}", + "image_path": "084669d3659fe11b862a707ddddb3d0321d6dba103eceebe6e10f9af8eeafcd0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": "and show that the restoration result is further improved in Fig. 5, where " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_{\\mathrm{Conv}}(\\cdot)" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": " is a convolution-based feed forward layer that is used to learn a " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "q \\times q" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": " deblur kernel for each region of " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": " from the unnormalized blur map " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\sigma^{-1}(\\mathbf{M})" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M})) \\in \\mathbb{R}^{h_x \\times w_x \\times q^2}" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": ". The Pad(" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": ") function reshapes " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h_x \\times w_x \\times q \\times q}" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": ", and pads the reshaped " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}_{\\mathrm{Conv}}(\\sigma^{-1}(\\mathbf{M}))" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(h_x \\times w_x) \\times (h_x \\times w_x)}" + }, + { + "bbox": [ + 46, + 516, + 287, + 613 + ], + "type": "text", + "content": ", for shape alignment issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 621, + 154, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 154, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 154, + 634 + ], + "type": "text", + "content": "3.3. Network Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "content": "We optimize our deblurring network with the widely used Charbonnier loss. Moreover, we propose a blur-aware loss and a blur weighting loss to penalize the blurred regions in the restoration " + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Charbonnier loss. We encourage the restoration to be close to the ground truth by using Charbonnier loss [12, 43]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 349, + 113 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 349, + 113 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 349, + 113 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 349, + 113 + ], + "type": "image", + "image_path": "480aad7501ee7e8b7fcb54754c39b4d78d088b346ac26c3d6c0e4798592d9d0c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 116, + 342, + 124 + ], + "lines": [ + { + "bbox": [ + 331, + 116, + 342, + 124 + ], + "spans": [ + { + "bbox": [ + 331, + 116, + 342, + 124 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 350, + 70, + 424, + 113 + ], + "blocks": [ + { + "bbox": [ + 350, + 70, + 424, + 113 + ], + "lines": [ + { + "bbox": [ + 350, + 70, + 424, + 113 + ], + "spans": [ + { + "bbox": [ + 350, + 70, + 424, + 113 + ], + "type": "image", + "image_path": "df89236f0ba3d0a983886e86e0fac8dc150f9b76ee15aaf584879ba6aa82d718.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 116, + 400, + 124 + ], + "lines": [ + { + "bbox": [ + 390, + 116, + 400, + 124 + ], + "spans": [ + { + "bbox": [ + 390, + 116, + 400, + 124 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 466, + 113 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 466, + 113 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 466, + 113 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 466, + 113 + ], + "type": "image", + "image_path": "8838b231c703c986bc4d3b093dad7eb2534b56e2645f464ec2a5a15cd603a2de.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 116, + 460, + 124 + ], + "lines": [ + { + "bbox": [ + 449, + 116, + 460, + 124 + ], + "spans": [ + { + "bbox": [ + 449, + 116, + 460, + 124 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "type": "text", + "content": "Figure 5. Example of deblurring with the blur map projected deblur kernel " + }, + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "type": "text", + "content": ". (a) and (b) are the left view of the DP pair and the ground truth sharp image. (c) and (d) are deblurred images with and without using " + }, + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{O}" + }, + { + "bbox": [ + 305, + 125, + 545, + 168 + ], + "type": "text", + "content": ". Zoom in for better quality." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 542, + 113 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 542, + 113 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 542, + 113 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 542, + 113 + ], + "type": "image", + "image_path": "53c8c1fc807ced4c5c6267718200456ec40493e492d5f7a61f420f3dd08a2a77.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 116, + 519, + 124 + ], + "lines": [ + { + "bbox": [ + 509, + 116, + 519, + 124 + ], + "spans": [ + { + "bbox": [ + 509, + 116, + 519, + 124 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "spans": [ + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "text", + "content": "We empirically set " + }, + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 306, + 183, + 471, + 195 + ], + "type": "text", + "content": ", and define" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 351, + 204, + 545, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 204, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 351, + 204, + 545, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\| ^ {2} + \\varepsilon^ {2}}. \\tag {11}", + "image_path": "a27024e5dedd7037921f8ae30bd5c79eb7c262a698747e6e3634b4afecf53894.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": "Blur Weighting Loss. We impose an adaptive penalty on the blurred regions of the DP pair " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": ", considering the degree of blur measured by CLIP. It encourages the network to focus on the restoration of heavily blurred regions. We first normalize the deviations between " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": " by the deviations between " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}},\\mathbf{B}_{\\mathrm{R}})" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 305, + 231, + 545, + 316 + ], + "type": "text", + "content": ", and then re-use the blur map to select blurred regions residing in the DP pair." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 323, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 323, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 311, + 323, + 545, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {b w l}} = \\frac {\\left\\| \\mathbf {I} (i , j) - \\hat {\\mathbf {I}} (i , j) \\right\\| ^ {2}}{\\left\\| \\hat {\\mathbf {I}} (i , j) - \\frac {1}{2} \\left(\\mathbf {B} _ {\\mathrm {L}} (i , j) + \\mathbf {B} _ {\\mathrm {R}} (i , j)\\right) \\right\\| ^ {2}} \\cdot \\mathbf {M} (i, j), \\tag {12}", + "image_path": "69a1c9a76465076f6a79bfb4123e78867fde0edcb0c53ee6a189c1c9bfce501c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": "where we use the blur map " + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 305, + 373, + 545, + 397 + ], + "type": "text", + "content": " (with soft logits) to reflect the blur levels detected by CLIP." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "content": "Blur-aware Loss. We encourage " + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "content": " to be sharp by leveraging the versatile knowledge contained in the CLIP. We generate the blur map " + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "content": " by the blur-aware format for the restoration, and penalize the blur exhibited in " + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "content": ". Similar to " + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bw1}}" + }, + { + "bbox": [ + 305, + 397, + 545, + 457 + ], + "type": "text", + "content": ", we use the unnormalized version of the blur map," + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 391, + 465, + 545, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 465, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 391, + 465, + 545, + 479 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b a l}} = \\hat {\\mathbf {M}} (i, j). \\tag {13}", + "image_path": "e157ae31d171ced253c3ce796ae0d2811fbc727bee0b695239d4ed7bdc8c9ee8.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "text", + "content": "This encourages " + }, + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 305, + 489, + 544, + 512 + ], + "type": "text", + "content": " to be zero, i.e., no blur exists in the restoration." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 514, + 544, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 514, + 544, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 514, + 544, + 538 + ], + "type": "text", + "content": "Total Loss. Our network is optimized by minimizing the following loss function" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 545, + 545, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 545, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 320, + 545, + 545, + 575 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\frac {1}{| \\Omega |} \\sum_ {(i, j) \\in \\Omega} \\left(\\mathcal {L} _ {\\text {c h a r}} + \\lambda_ {1} \\mathcal {L} _ {\\text {b a l}} + \\lambda_ {2} \\mathcal {L} _ {\\text {b w l}}\\right), \\tag {14}", + "image_path": "06288a0893d73849035f53e3308575ad441bdc2605fe10418005977b256300fe.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\frac{1}{|\\Omega|}" + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "content": " is a normalization factor determined by the image domain size " + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "inline_equation", + "content": "|\\Omega|" + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 305, + 585, + 545, + 622 + ], + "type": "text", + "content": " are hyperparameters to balance losses." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 634, + 455, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 455, + 647 + ], + "type": "text", + "content": "4. Experiments and Analysis" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We implement our model under the PyTorch framework. We use the pre-trained and ViT-B/32 [8] architecture-based CLIP [30] for our blur map estimation strategy. We train the deblurring network in two stages for " + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "6 \\times 10^{6}" + }, + { + "bbox": [ + 305, + 654, + 545, + 713 + ], + "type": "text", + "content": " steps in total with the Adam optimizer." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24082" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 92, + 550, + 268 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the DPD-blur dataset [1] composed of 37 indoor and 39 outdoor scenes. We respectively color the best and the second-best performance in red and blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 92, + 550, + 268 + ], + "lines": [ + { + "bbox": [ + 47, + 92, + 550, + 268 + ], + "spans": [ + { + "bbox": [ + 47, + 92, + 550, + 268 + ], + "type": "table", + "html": "
MethodOutdoor ScenesIndoor ScenesAverage
PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓PSNR↑SSIM↑MAE↓RMSE rel↓LPIPS↓
EBDB25.770.7720.0400.05121.250.5990.0580.08623.450.6830.0490.0670.336
DMENet25.500.7880.0380.05321.430.6440.0630.08423.410.7140.0510.0670.349
DPDNet27.480.8490.0290.04222.900.7260.0520.07125.130.7860.0410.0550.223
RDPD28.100.8430.0270.03922.820.7040.0530.07225.390.7720.0400.0530.255
DDDNet27.570.8330.0300.04123.280.7080.0500.06825.360.7680.0410.0540.197
IFAN28.660.8680.0250.03723.460.7430.0490.06725.990.8040.0370.0500.207
BAMBNet28.600.8720.0260.03724.300.7720.0450.06026.400.8210.0360.0470.141
DeepRFT28.480.8700.0250.03823.090.7360.0510.07025.710.8010.0370.0510.217
Restormer29.480.8950.0230.03423.970.7730.0470.06326.660.8330.0350.0460.155
K3DN29.870.8900.0220.03224.370.7800.0460.06027.060.8350.0340.0440.187
Ours (Small)29.550.8930.0210.03324.380.7730.0430.06526.910.8310.0320.0450.185
Ours (Large)29.880.8960.0200.03224.730.7860.0370.05927.240.8400.0280.0430.168
", + "image_path": "e6eb4fa16d9dee32b9d8049eeb7788574651949f1c84b8e2ebcaa4d8e7feefee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 282, + 291, + 441 + ], + "blocks": [ + { + "bbox": [ + 70, + 271, + 274, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 271, + 274, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 271, + 274, + 281 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on DDD-syn dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 282, + 291, + 441 + ], + "lines": [ + { + "bbox": [ + 48, + 282, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 291, + 441 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑MAE↓RMSE_rel↓
EBDB26.480.6830.0360.047
DMENet30.140.9390.0170.031
DPDNet31.450.9260.0160.027
RDPD32.420.9120.0160.024
DDDNet33.210.9560.0100.022
IFAN34.180.9290.0120.020
BAMBNet35.900.9540.0120.016
DeepRFT36.520.9520.0100.015
Restormer36.440.9570.0100.015
K3DN38.230.9650.0080.012
Ours (Small)37.270.9590.0100.014
Ours (Large)38.620.9710.0080.012
", + "image_path": "af59d58c1dfebd5a9ce8d718ea1d57d049685ba95a85e899be9a487dd5dbf731.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": "In each stage, the learning rate is decayed from " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " by a cosine annealing strategy. In the first stage, we train the network for " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "5 \\times 10^{6}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " steps, and only use the Charbonnier loss " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{char}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": ". In the second stage, the network is finetuned for " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "1 \\times 10^{6}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " steps with " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": ", where we have " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 10^{-1}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 2 \\times 10^{-1}" + }, + { + "bbox": [ + 46, + 453, + 287, + 537 + ], + "type": "text", + "content": ". Code, results, and more details are given in our project page." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 635 + ], + "type": "text", + "content": "Datasets. Our method is evaluated on two real-world DP defocus deblurring datasets, DPD-blur dataset [1] and DPD-disp dataset [27]. We follow [1] on the training and testing splits. For the DPD-disp dataset, we use it for testing only, re-using the checkpoints trained from the DPD-blur dataset as they do not provide ground-truth (GT) data or with limited dataset size. Moreover, two synthetic datasets, DDD-syn [26] and RDPD [2], are also used for evaluations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 686 + ], + "type": "text", + "content": "Evaluation Metrics. We use peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean absolute error (MAE), and root mean squared relative error (RMSE_ref) as our evaluation metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Baseline Methods. We compare with SOTA DP defocus deblurring methods, i.e., EBDB [13], DMENet [15]," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 282, + 547, + 441 + ], + "blocks": [ + { + "bbox": [ + 324, + 271, + 518, + 281 + ], + "lines": [ + { + "bbox": [ + 324, + 271, + 518, + 281 + ], + "spans": [ + { + "bbox": [ + 324, + 271, + 518, + 281 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons on RDPD dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 282, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 306, + 282, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 547, + 441 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑MAE↓RMSE rel↓
EBDB24.230.6370.0310.054
DMENet25.170.7310.0360.061
DPDNet29.840.8280.0320.055
RDPD31.090.8610.0250.028
DDDNet30.140.8400.0160.031
IFAN31.120.8650.0230.028
BAMBNet31.780.8670.0160.026
DeepRFT31.710.8790.0160.024
Restormer32.270.8710.0160.024
K3DN32.840.9050.0140.023
Ours (Small)32.980.8950.0140.022
Ours (Large)33.440.9070.0130.021
", + "image_path": "0b0fcfdd19a74488061f44a8fa8e94ab12cff7720ace3d1c6d48eeffabd63eff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 453, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 545, + 477 + ], + "type": "text", + "content": "DPDNet [1], RDPD [2], DDDNet [26], IFAN [16], BAMBNet [20], DeepRFT [22], Restormer [39], and K3DN [38]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 487, + 492, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 492, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 492, + 499 + ], + "type": "text", + "content": "4.1. Comparisons with State-of-the-Art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 506, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 685 + ], + "type": "text", + "content": "Quantitative Comparison. We compare with the SOTA methods on the DPD-blur, DDD-syn, and RDPD datasets in Tab. 1, Tab. 2, and Tab. 3, respectively. Our observations are as follows: i) With our model (Small), we achieve similar restoration results with Restormer and K3DN, the third-best and second-best method, in all datasets. ii) When scaling to the large model, we consistently outperform all methods on all evaluation metrics. For example, our method, K3DN, and Restormer respectively achieve 27.24/27.06/26.66, 38.62/38.23/36.44, and 33.44/32.84/32.27 PSNR (dB) in the DPD-blur, DDD-syn, and RDPD datasets. Meanwhile, on a single RTX 4090 GPU, we remain to have a lower Infer. time (including CLIP) which is 0.042s and 0.054s for our small and large models, while K3DN/Restormer takes 0.198s/0.255s." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Comparison. We compare with state-of-the-art methods on the DPD-blur dataset (Fig. 6). As shown, our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24083" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 103, + 70, + 246, + 167 + ], + "blocks": [ + { + "bbox": [ + 103, + 70, + 246, + 167 + ], + "lines": [ + { + "bbox": [ + 103, + 70, + 246, + 167 + ], + "spans": [ + { + "bbox": [ + 103, + 70, + 246, + 167 + ], + "type": "image", + "image_path": "1a5e452037bfcf471d7ed6ed97b64d00f80d482743a1e3ad6d7b9d34937e419a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 171, + 208, + 183 + ], + "lines": [ + { + "bbox": [ + 151, + 171, + 208, + 183 + ], + "spans": [ + { + "bbox": [ + 151, + 171, + 208, + 183 + ], + "type": "text", + "content": "Blurry Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 248, + 70, + 306, + 110 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 306, + 110 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 306, + 110 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 306, + 110 + ], + "type": "image", + "image_path": "1a577a1e892a8fb4e3e7924fa7e1e060664c57fcf2ca9b0ed80042cb1be77ba2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 114, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 255, + 114, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 255, + 114, + 297, + 125 + ], + "type": "text", + "content": "GT Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 367, + 110 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 367, + 110 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 367, + 110 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 367, + 110 + ], + "type": "image", + "image_path": "4e95c219927e2e91924f97c7242859158592d75f84a6394d22288eba860ab7aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 71, + 427, + 110 + ], + "blocks": [ + { + "bbox": [ + 370, + 71, + 427, + 110 + ], + "lines": [ + { + "bbox": [ + 370, + 71, + 427, + 110 + ], + "spans": [ + { + "bbox": [ + 370, + 71, + 427, + 110 + ], + "type": "image", + "image_path": "e263042cb2f4553c514700a14933a30b450d50b806f7a7b80439df6f943c6067.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 114, + 418, + 125 + ], + "lines": [ + { + "bbox": [ + 376, + 114, + 418, + 125 + ], + "spans": [ + { + "bbox": [ + 376, + 114, + 418, + 125 + ], + "type": "text", + "content": "IFAN [16]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 488, + 110 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 488, + 110 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 488, + 110 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 488, + 110 + ], + "type": "image", + "image_path": "ceb81e56acb118fae69918d30d66a3cef1c1de4338a9c09c30cab0d81d1ec79e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 114, + 486, + 125 + ], + "lines": [ + { + "bbox": [ + 430, + 114, + 486, + 125 + ], + "spans": [ + { + "bbox": [ + 430, + 114, + 486, + 125 + ], + "type": "text", + "content": "DDDNet [26]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 249, + 129, + 306, + 167 + ], + "blocks": [ + { + "bbox": [ + 309, + 114, + 364, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 114, + 364, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 114, + 364, + 125 + ], + "type": "text", + "content": "Blurry Image" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 129, + 306, + 167 + ], + "lines": [ + { + "bbox": [ + 249, + 129, + 306, + 167 + ], + "spans": [ + { + "bbox": [ + 249, + 129, + 306, + 167 + ], + "type": "image", + "image_path": "593fa7f0cd33870b2370f7686d3d006831fff4436a047c349f44ed5ff953fbc2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 129, + 367, + 167 + ], + "blocks": [ + { + "bbox": [ + 309, + 129, + 367, + 167 + ], + "lines": [ + { + "bbox": [ + 309, + 129, + 367, + 167 + ], + "spans": [ + { + "bbox": [ + 309, + 129, + 367, + 167 + ], + "type": "image", + "image_path": "71419781ec3eb04f150dd9a5e0d695549ad8f71fac20fec5db2020fdef535aa2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 370, + 129, + 427, + 167 + ], + "blocks": [ + { + "bbox": [ + 370, + 129, + 427, + 167 + ], + "lines": [ + { + "bbox": [ + 370, + 129, + 427, + 167 + ], + "spans": [ + { + "bbox": [ + 370, + 129, + 427, + 167 + ], + "type": "image", + "image_path": "1ab9e10670eae6a5762ffcdbd1f94d11fb463e0508bfb974ed139b904a3febde.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 430, + 129, + 488, + 167 + ], + "blocks": [ + { + "bbox": [ + 430, + 129, + 488, + 167 + ], + "lines": [ + { + "bbox": [ + 430, + 129, + 488, + 167 + ], + "spans": [ + { + "bbox": [ + 430, + 129, + 488, + 167 + ], + "type": "image", + "image_path": "415d9511d004b84ab0fbe979d1a9d97b307e976b9d99d47209ccbd86eda5a141.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 102, + 186, + 246, + 282 + ], + "blocks": [ + { + "bbox": [ + 246, + 171, + 306, + 183 + ], + "lines": [ + { + "bbox": [ + 246, + 171, + 306, + 183 + ], + "spans": [ + { + "bbox": [ + 246, + 171, + 306, + 183 + ], + "type": "text", + "content": "DeepRFT [22]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 102, + 186, + 246, + 282 + ], + "lines": [ + { + "bbox": [ + 102, + 186, + 246, + 282 + ], + "spans": [ + { + "bbox": [ + 102, + 186, + 246, + 282 + ], + "type": "image", + "image_path": "a38e83a781b12c8e4eee01d3784d2ce81064db85b5dc3eb35f35492a454af57f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 286, + 207, + 297 + ], + "lines": [ + { + "bbox": [ + 151, + 286, + 207, + 297 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 207, + 297 + ], + "type": "text", + "content": "Blurry Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 246, + 286, + 306, + 297 + ], + "lines": [ + { + "bbox": [ + 246, + 286, + 306, + 297 + ], + "spans": [ + { + "bbox": [ + 246, + 286, + 306, + 297 + ], + "type": "text", + "content": "DeepRFT [22]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 249, + 186, + 306, + 225 + ], + "blocks": [ + { + "bbox": [ + 249, + 186, + 306, + 225 + ], + "lines": [ + { + "bbox": [ + 249, + 186, + 306, + 225 + ], + "spans": [ + { + "bbox": [ + 249, + 186, + 306, + 225 + ], + "type": "image", + "image_path": "d00923ce4f24847248c6927585b5be800a6d3253de2729faa764a9b0ad1da0f9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 228, + 297, + 240 + ], + "lines": [ + { + "bbox": [ + 255, + 228, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 255, + 228, + 297, + 240 + ], + "type": "text", + "content": "GT Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 186, + 367, + 225 + ], + "blocks": [ + { + "bbox": [ + 309, + 171, + 369, + 183 + ], + "lines": [ + { + "bbox": [ + 309, + 171, + 369, + 183 + ], + "spans": [ + { + "bbox": [ + 309, + 171, + 369, + 183 + ], + "type": "text", + "content": "Restormer [39]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 186, + 367, + 225 + ], + "lines": [ + { + "bbox": [ + 309, + 186, + 367, + 225 + ], + "spans": [ + { + "bbox": [ + 309, + 186, + 367, + 225 + ], + "type": "image", + "image_path": "10c56ea97f9d68316bbdc7475b929b6f2caba95210b120acf0a3e447d2f67f52.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 228, + 364, + 240 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 364, + 240 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 364, + 240 + ], + "type": "text", + "content": "Blurry Image" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 370, + 186, + 427, + 225 + ], + "blocks": [ + { + "bbox": [ + 377, + 171, + 425, + 183 + ], + "lines": [ + { + "bbox": [ + 377, + 171, + 425, + 183 + ], + "spans": [ + { + "bbox": [ + 377, + 171, + 425, + 183 + ], + "type": "text", + "content": "K3DN [38]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 186, + 427, + 225 + ], + "lines": [ + { + "bbox": [ + 370, + 186, + 427, + 225 + ], + "spans": [ + { + "bbox": [ + 370, + 186, + 427, + 225 + ], + "type": "image", + "image_path": "c1cba08fd1c437d5faf95cc4c61b001bbfec7982372d7fff8b5e1e2745bb5b7f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 228, + 418, + 240 + ], + "lines": [ + { + "bbox": [ + 375, + 228, + 418, + 240 + ], + "spans": [ + { + "bbox": [ + 375, + 228, + 418, + 240 + ], + "type": "text", + "content": "IFAN [16]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 430, + 186, + 488, + 225 + ], + "blocks": [ + { + "bbox": [ + 453, + 172, + 474, + 182 + ], + "lines": [ + { + "bbox": [ + 453, + 172, + 474, + 182 + ], + "spans": [ + { + "bbox": [ + 453, + 172, + 474, + 182 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 186, + 488, + 225 + ], + "lines": [ + { + "bbox": [ + 430, + 186, + 488, + 225 + ], + "spans": [ + { + "bbox": [ + 430, + 186, + 488, + 225 + ], + "type": "image", + "image_path": "3deff5e3099745779cb9d782a3227a6d95e4b9a3129e3a16d70ece208f2f3ad0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 228, + 486, + 240 + ], + "lines": [ + { + "bbox": [ + 430, + 228, + 486, + 240 + ], + "spans": [ + { + "bbox": [ + 430, + 228, + 486, + 240 + ], + "type": "text", + "content": "DDDNet [26]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 249, + 244, + 306, + 282 + ], + "blocks": [ + { + "bbox": [ + 249, + 244, + 306, + 282 + ], + "lines": [ + { + "bbox": [ + 249, + 244, + 306, + 282 + ], + "spans": [ + { + "bbox": [ + 249, + 244, + 306, + 282 + ], + "type": "image", + "image_path": "e6d883e27feceebacbcd4f766c2154c76680ba300f69277ab319b309a3c8e21d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on the DPD-blur dataset [1]. We present the blurry image " + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_{\\mathrm{L}})" + }, + { + "bbox": [ + 46, + 300, + 547, + 333 + ], + "type": "text", + "content": " in the first column, and the ground truth (GT) image of regions residing in the red bounding box is shown in the second column. The cropped blurred image of the red bounding box region is in the third column." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 309, + 244, + 367, + 282 + ], + "blocks": [ + { + "bbox": [ + 309, + 244, + 367, + 282 + ], + "lines": [ + { + "bbox": [ + 309, + 244, + 367, + 282 + ], + "spans": [ + { + "bbox": [ + 309, + 244, + 367, + 282 + ], + "type": "image", + "image_path": "0e2e56ebeb48ea5db23dc965bdb40ba0b1498af169039040fd80561846390f0d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 286, + 369, + 297 + ], + "lines": [ + { + "bbox": [ + 309, + 286, + 369, + 297 + ], + "spans": [ + { + "bbox": [ + 309, + 286, + 369, + 297 + ], + "type": "text", + "content": "Restormer [39]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 370, + 244, + 427, + 282 + ], + "blocks": [ + { + "bbox": [ + 370, + 244, + 427, + 282 + ], + "lines": [ + { + "bbox": [ + 370, + 244, + 427, + 282 + ], + "spans": [ + { + "bbox": [ + 370, + 244, + 427, + 282 + ], + "type": "image", + "image_path": "a3ae7e7d550322b212db7574819d25fe9d264c840e13caffee136fba8e2865e6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 286, + 424, + 297 + ], + "lines": [ + { + "bbox": [ + 376, + 286, + 424, + 297 + ], + "spans": [ + { + "bbox": [ + 376, + 286, + 424, + 297 + ], + "type": "text", + "content": "K3DN [38]" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 430, + 244, + 488, + 282 + ], + "blocks": [ + { + "bbox": [ + 430, + 244, + 488, + 282 + ], + "lines": [ + { + "bbox": [ + 430, + 244, + 488, + 282 + ], + "spans": [ + { + "bbox": [ + 430, + 244, + 488, + 282 + ], + "type": "image", + "image_path": "3fe1437ecf8035033a05d2376d9f62d1fd4e4038bc29755523e7b1a9b1a2cddc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 287, + 474, + 297 + ], + "lines": [ + { + "bbox": [ + 453, + 287, + 474, + 297 + ], + "spans": [ + { + "bbox": [ + 453, + 287, + 474, + 297 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 416 + ], + "type": "text", + "content": "method performs sharper and clearer restorations of edges and texts. To further demonstrate the benefits of using CLIP to generate the blur mask unsupervisedly in our approach, we provide an example in Fig. 7, which preserves sharp regions (e.g., text 'University') in the input blur image effectively compared to the second-best baseline." + } + ] + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 50, + 419, + 108, + 451 + ], + "blocks": [ + { + "bbox": [ + 50, + 419, + 108, + 451 + ], + "lines": [ + { + "bbox": [ + 50, + 419, + 108, + 451 + ], + "spans": [ + { + "bbox": [ + 50, + 419, + 108, + 451 + ], + "type": "image", + "image_path": "49d024ddac1a93a08d0f3ae7a68f2ae2039e91b0ac1850f8d2dcdedd2fdd5610.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 452, + 94, + 460 + ], + "lines": [ + { + "bbox": [ + 65, + 452, + 94, + 460 + ], + "spans": [ + { + "bbox": [ + 65, + 452, + 94, + 460 + ], + "type": "text", + "content": "(a) Input" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 463, + 287, + 485 + ], + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 485 + ], + "type": "text", + "content": "Figure 7. An example of the restoration results on the sharp region of the input DP image (left view)." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 110, + 419, + 167, + 451 + ], + "blocks": [ + { + "bbox": [ + 110, + 419, + 167, + 451 + ], + "lines": [ + { + "bbox": [ + 110, + 419, + 167, + 451 + ], + "spans": [ + { + "bbox": [ + 110, + 419, + 167, + 451 + ], + "type": "image", + "image_path": "e3c81256f7e09608a955ef70e91d56ce26e088b522210737345b94a264481924.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 452, + 155, + 460 + ], + "lines": [ + { + "bbox": [ + 121, + 452, + 155, + 460 + ], + "spans": [ + { + "bbox": [ + 121, + 452, + 155, + 460 + ], + "type": "text", + "content": "(b) K3DN" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 168, + 419, + 226, + 451 + ], + "blocks": [ + { + "bbox": [ + 168, + 419, + 226, + 451 + ], + "lines": [ + { + "bbox": [ + 168, + 419, + 226, + 451 + ], + "spans": [ + { + "bbox": [ + 168, + 419, + 226, + 451 + ], + "type": "image", + "image_path": "fbad7c91f69532a328961c39f21d686aa4cff8de248434c3a6ea5e0377a8515b.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 452, + 219, + 460 + ], + "lines": [ + { + "bbox": [ + 173, + 452, + 219, + 460 + ], + "spans": [ + { + "bbox": [ + 173, + 452, + 219, + 460 + ], + "type": "text", + "content": "(c) Restormer" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 227, + 419, + 285, + 451 + ], + "blocks": [ + { + "bbox": [ + 227, + 419, + 285, + 451 + ], + "lines": [ + { + "bbox": [ + 227, + 419, + 285, + 451 + ], + "spans": [ + { + "bbox": [ + 227, + 419, + 285, + 451 + ], + "type": "image", + "image_path": "7d412d6a06994209392d2827b7c200204608fe6290bf8376b7f031b3173930e4.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 452, + 269, + 460 + ], + "lines": [ + { + "bbox": [ + 241, + 452, + 269, + 460 + ], + "spans": [ + { + "bbox": [ + 241, + 452, + 269, + 460 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 47, + 491, + 140, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 140, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 140, + 504 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 544 + ], + "type": "text", + "content": "We study model architecture, blur map estimation, and losses in this section. All experiments are performed in the DPD-blur dataset." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 546, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 677 + ], + "type": "text", + "content": "Ablation of Model Architecture. We study the effective usage of blur map as prior knowledge for DP defocus deblurring (Tab. 6). To show the benefits of using blur maps, we consider four baselines: i) NAFNet [5], the baseline model; ii) Concat, concatenating the blur map with the DP pair before feeding to the model; iii) SA, using the self-attention layer [8] in Eq. 8 to the model only; iv) Concat+SA, combines the Concat and SA settings; v) BPA block, using kernel priors from the blur map to assist deblurring. Our model consistently outperforms all baselines, which indicates the effectiveness of using the blur map." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Ablation of Losses. We investigate the proposed blur weighting loss " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bw}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and blur-aware loss " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bal}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " (Tab. 7) with a baseline as the " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{char}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " setting. We first use " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bw}}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bal}}" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": "to our baseline, NAFNet (i.e., not using our BPA block), that only uses " + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{char}}" + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": ", having 0.32 PSNR (dB) improvements. Secondly, we train our model (backbone + BPA block) with Charbonnier loss, resulting in 26.71/0.825 higher PSNR (dB)/SSIM. By using the " + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bwl}}" + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bal}}" + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": ", we respectively achieve 0.05/0.003 or 0.10/0.004 higher PSNR (dB)/SSIM. Combining all losses, we have the best restoration performance at 26.91/0.831 for PSNR (dB)/SSIM. Moreover, we verify the generalization ability of our " + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bwl}}" + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{bal}}" + }, + { + "bbox": [ + 304, + 344, + 545, + 476 + ], + "type": "text", + "content": " losses on state-of-the-art networks, improving PSNR of Restormer and DeepRFT by 0.25 and 0.41 PSNR (dB)." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": "Ablation of Blur Map Estimation. We ablate with different prompts used to generate the blur map. The blur-aware and DP-aware prompts are presented in Tab. 4, and we show their impact on model performance with Tab. 5. We then ensemble the groups of blur-aware prompt and disparity-aware prompt individually (Ens." + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": "), leading to " + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "inline_equation", + "content": "26.80\\mathrm{dB} / 0.829" + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "inline_equation", + "content": "26.84 / 0.830" + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": " PSNR (dB)/SSIM. By assembling all prompts (Ens." + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": "), we have " + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "inline_equation", + "content": "26.91 / 0.831" + }, + { + "bbox": [ + 304, + 481, + 546, + 601 + ], + "type": "text", + "content": " PSNR (dB)/SSIM, consistently outperforming other approaches of generating blur masks." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 305, + 610, + 451, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 610, + 451, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 610, + 451, + 621 + ], + "type": "text", + "content": "4.3. Discussion and Limitation" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 689 + ], + "type": "text", + "content": "Discussion. The reader may wonder about the performance of our unsupervised blur map estimation strategy. Therefore, we perform two qualitative evaluations on the DPD-disp dataset, where the GT disparity is included. We compare with BAMBNet [20] and DPD-disp [27]." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "First, we consider blur map estimation. To obtain the GT blur map, the pixel with disparity " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " that satisfies " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "|d|\\leq 1" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24084" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 82, + 541, + 150 + ], + "blocks": [ + { + "bbox": [ + 210, + 71, + 382, + 81 + ], + "lines": [ + { + "bbox": [ + 210, + 71, + 382, + 81 + ], + "spans": [ + { + "bbox": [ + 210, + 71, + 382, + 81 + ], + "type": "text", + "content": "Table 4. Prompts used for blur map estimation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 82, + 541, + 150 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 541, + 150 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 541, + 150 + ], + "type": "table", + "html": "
Blur-aware PromptDP-aware Prompt
p1[A blurry image.][A symmetrical image.]
p2[An obscured image.][A horizontally symmetrical image.]
p3[An out of focus image.][A left-right symmetrical image.]
p4[An image lacking sharpness.][A symmetrical image at the pixel level.]
", + "image_path": "054057e6ea0ca3eda4da11297624db98f45cdf17840b5131bb678044b149f705.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 176, + 546, + 281 + ], + "blocks": [ + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "lines": [ + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "type": "text", + "content": "Table 5. Restoration performance of our methods by using blur-aware or/and DP-aware prompts for blur map estimation. Ens. " + }, + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "type": "text", + "content": " denotes ensembling the group of blue-aware prompts and DP-aware prompts individually, and Ens. " + }, + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 47, + 154, + 546, + 176 + ], + "type": "text", + "content": " ensembles all prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 176, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 48, + 176, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 546, + 281 + ], + "type": "table", + "html": "
Blur-aware PromptDP-aware Prompt
PSNR↑SSIM↑MAE↓RMSE_rel↓PSNR↑SSIM↑MAE↓RMSE_rel↓
p126.830.8290.0330.04526.820.8300.0330.046
p226.810.8290.0330.04626.830.8300.0320.046
p326.750.8280.0340.04626.820.8300.0330.046
p426.770.8290.0330.04626.850.8300.0320.045
Ens.126.800.8290.0330.04626.840.8300.0320.046
Ens.226.910.8310.0320.04526.910.8310.0320.045
", + "image_path": "fd845bada4709046aa36c7bcda4835ebe18691ddb98faac64d34b125b6988318.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 307, + 299, + 386 + ], + "blocks": [ + { + "bbox": [ + 94, + 297, + 239, + 307 + ], + "lines": [ + { + "bbox": [ + 94, + 297, + 239, + 307 + ], + "spans": [ + { + "bbox": [ + 94, + 297, + 239, + 307 + ], + "type": "text", + "content": "Table 6. Ablation of model components." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 307, + 299, + 386 + ], + "lines": [ + { + "bbox": [ + 48, + 307, + 299, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 299, + 386 + ], + "type": "table", + "html": "
SettingPSNR↑SSIM↑MAE↓RMSE rel↓
i) NAFNet26.210.8170.0380.049
ii) Concat26.760.8260.0330.046
iii) SA26.590.8240.0360.047
iv) Concat + SA26.800.8290.0320.046
v) BPA block26.910.8310.0320.045
", + "image_path": "1304abdac5fba0c5de280780e68e59e9f90789abc3022ac3f19f817492955c6a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 412, + 292, + 507 + ], + "blocks": [ + { + "bbox": [ + 102, + 401, + 231, + 412 + ], + "lines": [ + { + "bbox": [ + 102, + 401, + 231, + 412 + ], + "spans": [ + { + "bbox": [ + 102, + 401, + 231, + 412 + ], + "type": "text", + "content": "Table 7. The effectiveness of losses." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 412, + 292, + 507 + ], + "lines": [ + { + "bbox": [ + 48, + 412, + 292, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 412, + 292, + 507 + ], + "type": "table", + "html": "
BPA blockLossPSNR↑SSIM↑MAE↓RMSE rel↓
XLchar26.210.8170.0380.049
XLtotal26.530.8220.0360.047
✓Lchar26.710.8250.0340.046
✓Lchar + Lbwl26.760.8280.0330.046
✓Lchar + Lbal26.810.8290.0320.046
✓Ltotal26.910.8310.0320.045
", + "image_path": "e3c0811b2d01731d6a2a8a170cbde36ffd7f4ea5ecea06901cbe448122af2fd0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 520, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 568 + ], + "type": "text", + "content": "is used as the sharp pixel, reversely, it is a buried pixel. Our method, BAMBNet [20] and DPD-disp [27] respectively have " + }, + { + "bbox": [ + 46, + 520, + 287, + 568 + ], + "type": "inline_equation", + "content": "89.8\\% / 70.2\\% / 88.3\\%" + }, + { + "bbox": [ + 46, + 520, + 287, + 568 + ], + "type": "text", + "content": " accuracy (the higher, the better), showing the effectiveness of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Second, we explore disparity estimation. The log-its of our unnormalized blurred map are potentially linearly correlated with the ground-truth disparity. Therefore, we compare with the state-of-the-art DP disparity estimation method, DPD-disp [27]. Following [9, 27], we use AI(1), AI(2), and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "1 - |\\rho_s|" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " as the evaluation metrics (the lower, the better). The AI(1), AI(2), and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "1 - |\\rho_s|" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " of our method, BAMBNet, and DPD-disp are 0.063/0.106/0.069, 0.130/0.184/0.096, 0.306/0.772/0.0159. Though our blur map is estimated with a single forward pass, we still achieve a reasonable performance in the disparity estimation. This potentially explains the high performance of our DP defo" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 298, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 321 + ], + "type": "text", + "content": "cus deblurring network, as obtaining the disparity prior benefits the defocus task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 425 + ], + "type": "text", + "content": "Limitation. As the image encoder of CLIP is trained under a classification-alike pre-training task, it will focus more on local details that help identify its scene class. Textureless regions would produce non-significant semantic reactions, e.g., large areas of wall, ceiling, or floor, and reduce the quality of the estimated blur map. It is possible to use a vision language framework that drives more sensitivity to regions with less texture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 438, + 486, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 486, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 486, + 453 + ], + "type": "text", + "content": "5. Conclusion and Broader Impact" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 460, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 593 + ], + "type": "text", + "content": "In this paper, we study end-to-end defocus deblurring of DP images, leveraging the blur-related prior knowledge from the CLIP. We first estimate a blur map with an ensemble of blur-aware and DP-aware strategies, and then use the estimated blur map as the deblurring kernel prior to restoring the DP image. We also propose a blur-aware and blur-weighting loss to regularize the restorations of the DP images during training, by distilling the knowledge of blur from CLIP. In extensive experiments, our method outperforms past works by a large margin in quantitative and qualitative restoration performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 598, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 598, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 598, + 547, + 635 + ], + "type": "text", + "content": "Broader Impact. The proposed DP-aware strategy is promising to apply and extend CLIP to diverse zero-shot stereo vision tasks. We hope it will motivate future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045, and Miaomiao Liu was supported by the ARC fellowship and Discovery Project grant (DE180100628, DP200102274)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24085" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 125 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 179 + ], + "type": "text", + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2289-2298, 2021. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 236 + ], + "type": "text", + "content": "[3] Abdullah Abuolaim, Mahmoud Affi, and Michael S Brown. Improving single-image defocus deblurring: How dual-pixel images help through multi-task learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1231–1239, 2022. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 289 + ], + "type": "text", + "content": "[4] Dylan Auty and Krystian Mikolajczyk. Learning to prompt clip for monocular depth estimation: Exploring the limits of human language. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2039-2047, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 346 + ], + "type": "text", + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 17-33. Springer, 2022. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 401 + ], + "type": "text", + "content": "[6] Myungsub Choi, Hana Lee, and Hyong-euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13158-13168, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 446 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 513 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 288, + 567 + ], + "type": "text", + "content": "[9] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7628-7637, 2019. 1, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 624 + ], + "type": "text", + "content": "[10] Xueting Hu, Ce Zhang, Yi Zhang, Bowen Hai, Ke Yu, and Zhihai He. Learning to adapt clip for few-shot monocular depth estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5594-5603, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 691 + ], + "type": "text", + "content": "[11] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22157-22167, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[12] Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. Multi-scale" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8346-8355, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[13] Ali Karaali and Claudio Rosito Jung. Edge-based defocus blur estimation with adaptive scale selection. IEEE Transactions on Image Processing, 27(3):1126-1137, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[14] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5023-5032, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 251 + ], + "type": "text", + "content": "[15] Junyong Lee, Sungkil Lee, Sunghyun Cho, and Seungyong Lee. Deep defocus map estimation using domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12222-12230, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 308 + ], + "type": "text", + "content": "[16] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2034-2042, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[17] Yu Li, Yaling Yi, Dongwei Ren, Qince Li, and Wangmeng Zuo. Learning dual-pixel alignment for defocus deblurring. arXiv preprint arXiv:2204.12105, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 343, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 408 + ], + "type": "text", + "content": "[18] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 454 + ], + "type": "text", + "content": "[19] Yu Li, Dongwei Ren, Xinya Shu, and Wangmeng Zuo. Learning single image defocus deblurring with misaligned training pairs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1495-1503, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 498 + ], + "type": "text", + "content": "[20] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Bambnet: A blur-aware multi-branch network for dual-pixel defocus deblurring. IEEE/CAA Journal of Automatica Sinica, 9(5):878-892, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[21] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 600 + ], + "type": "text", + "content": "[22] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the 37th AAAI Conference on Artificial Intelligence, 2023. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[23] Sagi Monin, Sagi Katz, and Georgios Evangelidis. Continuous cost aggregation for dual-pixel disparity extraction. arXiv preprint arXiv:2306.07921, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[24] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 715 + ], + "type": "text", + "content": "[25] Liyuan Pan, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Phase-only image based kernel estimation for single image blind deblurring. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24086" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 6034-6043, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 160 + ], + "type": "text", + "content": "[26] Liyuan Pan, Shah Chowdhury, Richard Hartley, Miaomiao Liu, Hongguang Zhang, and Hongdong Li. Dual pixel exploration: Simultaneous depth estimation and image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "text", + "content": "[27] Abhijith Punnappurath, Abdullah Abuolaim, Mahmoud Afifi, and Michael S Brown. Modeling defocus-disparity in dual-pixel sensors. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[28] Yuhui Quan, Zicong Wu, and Hui Ji. Neumann network with recursive kernels for single image defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5754-5763, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[29] Yuhui Quan, Xin Yao, and Hui Ji. Single image defocus deblurring via implicit neural inverse kernels. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12600-12610, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 369 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "text", + "content": "[31] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18082-18091, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[32] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15724-15734, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 482, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 545 + ], + "type": "text", + "content": "[33] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 2622-2630. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 613 + ], + "type": "text", + "content": "[34] Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T Barron, Pratul P Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. Defocus map estimation and deblurring from a single dual-pixel image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2021. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "text", + "content": "[35] Hao Yang, Liyuan Pan, Yan Yang, and Wei Liang. Language-driven all-in-one adverse weather removal. arXiv preprint arXiv:2312.01381, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[36] Shaoliang Yang, Dongming Zhou, Jinde Cao, and Yanbu Guo. Lightingnet: An integrated learning method for low-light image enhancement. IEEE Transactions on Computational Imaging, 9:29-42, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 473 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 149 + ], + "type": "text", + "content": "[38] Yan Yang, Liyuan Pan, Liu Liu, and Miaomiao Liu. K3dn: Disparity-aware kernel estimation for dual-pixel defocus deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13263-13272, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 217 + ], + "type": "text", + "content": "[39] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5728-5739, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 273 + ], + "type": "text", + "content": "[40] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 317 + ], + "type": "text", + "content": "[41] Renrui Zhang, Ziyao Zeng, Ziyu Guo, and Yafeng Li. Can language understand depth? In Proceedings of the 30th ACM International Conference on Multimedia, pages 6868-6874, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 319, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 372 + ], + "type": "text", + "content": "[42] Yinda Zhang, Neal Wadhwa, Sergio Orts-Escalano, Christian Hane, Sean Fanello, and Rahul Garg. Du2net: Learning depth estimation from dual-cameras and dual-pixels. In European Conference on Computer Vision, pages 582-598, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 418 + ], + "type": "text", + "content": "[43] Qian Zhao, Hao Yang, Dongming Zhou, and Jinde Cao. Rethinking image deblurring via cnn-transformer multiscale hybrid architecture. IEEE Transactions on Instrumentation and Measurement, 72:1-15, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 545, + 473 + ], + "type": "text", + "content": "[44] Ziqin Zhou, Yinjie Lei, Bowen Zhang, Lingqiao Liu, and Yifan Liu. Zegclip: Towards adapting clip for zero-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11175–11185, 2023. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24087" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_content_list.json b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42c6a73381c5e11278faf5be19d9b03d2cd4504a --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_content_list.json @@ -0,0 +1,1643 @@ +[ + { + "type": "text", + "text": "LEAD: Exploring Logit Space Evolution for Model Selection", + "text_level": 1, + "bbox": [ + 178, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Hu $^{1}$ Xiaotong Li $^{1}$ Shixiang Tang $^{3}$ Jun Liu $^{4}$ Yichun Hu $^{1}$ Ling-Yu Duan $^{1,2*}$", + "bbox": [ + 119, + 179, + 861, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Peking University, Beijing, China, $^{2}$ Peng Cheng Laboratory, Shenzhen, China,", + "bbox": [ + 125, + 200, + 852, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 The Chinese University of Hong Kong, Hongkong, China, 4 Singapore University of Technology and Design, Singapore {hzxuan, hycc, lingyu}@pku.edu.cn, lixiaotong@stu.pku.edu.cn, shixiangtang@cuhk.edu.hk, jun.liu@sutd.edu.sg", + "bbox": [ + 88, + 215, + 887, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The remarkable success of \"pretrain-then-finetune\" paradigm has led to a proliferation of available pre-trained models for vision tasks. This surge presents a significant challenge in efficiently choosing the most suitable pretrained models for downstream tasks. The critical aspect of this challenge lies in effectively predicting the model transferability by considering the underlying fine-tuning dynamics. Existing methods often model fine-tuning dynamics in feature space with linear transformations, which do not precisely align with the fine-tuning objective and fail to grasp the essential nonlinearity from optimization. To this end, we present LEAD, a finetuning-aligned approach based on the network output of logits. LEAD proposes a theoretical framework to model the optimization process and derives an ordinary differential equation (ODE) to depict the nonlinear evolution toward the final logit state. Additionally, we design a class-aware decomposition method to consider the varying evolution dynamics across classes and further ensure practical applicability. Integrating the closely aligned optimization objective and nonlinear modeling capabilities derived from the differential equation, our method offers a concise solution to effectively bridge the optimization gap in a single step, bypassing the lengthy fine-tuning process. The comprehensive experiments on 24 supervised and self-supervised pre-trained models across 10 downstream datasets demonstrate impressive performances and showcase its broad adaptability even in low-data scenarios.", + "bbox": [ + 73, + 335, + 473, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 762, + 209, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The \"pretrain-then-finetune\" learning paradigm has exhibited remarkable success in a wide range of computer vision tasks [6, 28], leading to an increasing number of diverse pretrained models open-sourced to the computer vision community. This surge poses a significant challenge in selecting the optimal pre-trained model for a downstream task", + "bbox": [ + 75, + 784, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c3f6aa22c47c4b6480171e0d2ff66d7c4f64ce96d130c4892baa2e673c8f2a1.jpg", + "image_caption": [ + "Figure 1. This diagram illustrates the task of model selection, which aims to rank pre-trained models for choosing the optimal one. The ground-truth ranking sequence is derived from the models' accuracy after fine-tuning. However, due to the time-intensive nature of this process, it requires to design efficient transferability metrics for accurate ranking prediction. The performance of this metric is evaluated using rank correlation, e.g., weighted Kendall's tau [42], to compare the predicted and ground-truth rankings." + ], + "image_footnote": [], + "bbox": [ + 504, + 301, + 893, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from a vast model zoo. Considering that the brute-force solution of fine-tuning all models on downstream tasks may be infeasible in many scenarios due to the substantial computational costs, the model selection task emerges [13, 47], which ranks the transferability of pre-trained models to the given task without incurring much cost, as shown in Fig. 1.", + "bbox": [ + 496, + 609, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Predicting model transferability aims to efficiently estimate the final state after fine-tuning on the downstream task, without network optimization by backward propagation. Prior studies only depend on specific properties of the initial state (e.g., representation similarity [13, 14], conditional distribution [33, 47]), to serve as transferability indicators. Recent studies reveal that modeling the dynamic process of fine-tuning has become the bottleneck to design more high-quality transferability metrics. Following this philosophy, recent works propose to simulate fine-tuning dynamics on features through multiple stages of linear transformations [32, 41], achieving some promising results.", + "bbox": [ + 496, + 702, + 892, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent progress in this field, there are still two", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 94, + 886, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28664", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "considerable gaps between existing modeled fine-tuning dynamics and the real ones: (1) Insufficient objective alignment. The designed objectives in prior works do not seamlessly align with fine-tuning, as they often focus more on enhancing class separability via feature manipulation, rather than the classification optimization for both the feature and output spaces. (2) Disregarding nonlinear modeling. Existing methods relying solely on linear transformations cannot well capture the intricate nonlinear changes in the feature space during optimization. Such constraints contribute to growing deviations in dynamic modeling and result in unreliable predictions of the final state.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Delving deeper into the optimization process, we focus on modeling the dynamic changes within the output space (i.e., logits for classification), which not only more closely relates to the fine-tuning objective but also embeds the overall nonlinear dynamics from the network optimization during fine-tuning. Building upon this foundation, we introduce an insightful framework to depict the trajectory of logits toward their final state in a dynamical modeling system. To capture and understand this evolution process of the logit space, the framework provides a discrete iterative optimization equation via gradient analysis and captures nonlinear logit space changes during optimization.", + "bbox": [ + 75, + 275, + 472, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the extensive iterations required in fine-tuning optimization, we further extend the framework to propose a new transferability metric \"Logits Evolution Analysis through Dynamical Equation (LEAD)\", to seize the final evolutionary state in a single step. We perform a mathematical derivation to convert the discrete process into a continuous differential equation and simplify it into an ordinary differential equation (ODE) by leveraging the classical analysis tool, Neural Tangent Kernel (NTK) [26]. This ODE concisely models the nonlinear evolution in logit space through its closed-form solution, allowing for real-time estimation of logits. Additionally, we introduce a class-aware decomposition method to enhance prediction accuracy and break the barrier of theoretical assumption of NTK. As a result, our method provides a direct estimation of the evolution by incorporating both the initial model state and the impact of fine-tuning evolution, serving as an effective indicator assessing model transferability.", + "bbox": [ + 75, + 459, + 472, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this work are three-fold. (1) To the best of our knowledge, we are the first to explore model transferability by modeling the evolution process in the logit space. (2) We propose LEAD, a novel transferability metric with two key components: a theoretical framework providing a closed-form solution to depict the logits evolution towards the final state, and a class-aware decomposition method to guarantee the practical applicability of theoretical findings and improve predictive precision. (3) We validate the effectiveness and generalizability of our method through experiments on 24 different pre-trained", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "models across 10 downstream datasets, achieving SOTA results with an average of $17\\%$ gain in the rank correlation. In addition, our method can be naturally extended to low-data scenarios while maintaining competitive performance.", + "bbox": [ + 496, + 90, + 893, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 164, + 642, + 179 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Transferability Metric", + "text_level": 1, + "bbox": [ + 498, + 185, + 709, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The challenge of efficiently ranking pre-trained models has captured much interest of researchers, leading to the development of various metrics to evaluate the model transferability [2, 3, 12, 18, 36, 38, 47, 52]. For example, RSA [13], DDS [14] and GBC [38] are methods based on the distance measurements of the feature space. RSA and DDS assess transferability by evaluating the representation similarity between pre-trained and downstream datasets. GBC employs the Bhattacharyya coefficient to measure class separability in the feature space. Nsleep [33] and Logme [52] estimate transferability by analyzing the conditional distribution, with Nsleep utilizing Gaussian mixture model and Logme employing Bayesian optimization. However, these methods treat network outputs as static and overlook the changes introduced by the training process, leading to less effectiveness in scenes with larger fine-tuning dynamics.", + "bbox": [ + 496, + 205, + 893, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, SFDA [41] and PED [32] propose to simulate changes in the feature space, respectively projecting the features into a Fisher space and utilizing a physically-inspired model to enhance the class separability, achieving impacts similar to fine-tuning. While these methods achieve some improvements, they solely concentrate on the feature space, disregarding the dynamic changes introduced by the classifier, which is jointly optimized with the backbone. Additionally, they utilize linear transformations to simulate the intricate variations in fine-tuning, resulting in a significant limitation in simulation precision. Instead, our LEAD employs the ordinary differential equation to analyze logit dynamics, maintaining a closely aligned objective and simulating the nonlinear changes in fine-tuning.", + "bbox": [ + 496, + 449, + 895, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural Tangent Kernel", + "text_level": 1, + "bbox": [ + 498, + 667, + 714, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Studying the behavior of neural network optimization processes has been an enduring challenge due to their nonlinearity and nonconvexity. Neural Tangent Kernel (NTK) [20, 26], serving as a theoretical tool to characterize the optimization process, has attracted much research attention. Some studies have explored its theoretical properties and applications. Notable works include the NTK's property of remaining constant under infinite network width [26], the explanation of the instability of training [48], the analysis of its bias and variance under finite width [20], and its acceleration algorithm [35]. While NTK can analyze network trainability [26, 35], it is constrained by the requirement of infinite width and neglects the consideration of initial log-its, resulting in the inability to predict the changes starting", + "bbox": [ + 496, + 688, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28665", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from the initial state for a finite-width network. To tackle these issues, we innovatively propose an analysis framework leveraging the property of NTK, suitable for model selection problems to bridge the gap between the initial and final states of finite-width networks.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 180, + 210, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the problem setup for the task of model selection. Then we analyze training mechanisms to model the fine-tuning optimization in a discrete process. Finally, we propose a transferability metric that converts the discrete process into a continuous dynamical equation with a concise solution to directly seize the evolution destination of logits, and feed updated logits into Cross-Entropy loss to assess transferability, as shown in Fig. 2.", + "bbox": [ + 75, + 205, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 76, + 333, + 227, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider $M$ pre-trained models $\\{\\Psi_m\\}_{m=1}^M$ and a downstream dataset $\\mathcal{T} = \\{x_n, y_n\\}_{n=1}^N$ with $K$ classes, where model $\\Psi$ consists of its backbone $f$ which outputs the encoded feature, $N$ is the number of images in the downstream task. The purpose of model selection is to design an efficient algorithm to predict the ranking of the performance of $M$ pre-trained models after fine-tuning.", + "bbox": [ + 75, + 353, + 468, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given the ground-truth performance rankings $\\{G_m\\}_{m=1}^M$ of pre-trained models after fine-tuning and the predicted ranking $\\{P_m\\}_{m=1}^M$ by model selection algorithms, we can leverage the commonly-used rank-based correlation, i.e., the weighted Kendalls' $\\tau_w$ , to evaluate the effectiveness of the proposed model selection methods, i.e.,", + "bbox": [ + 75, + 460, + 468, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {w} = \\frac {1}{\\sum_ {i \\neq j} w _ {i j}} \\cdot \\sum_ {i \\neq j} w _ {i j} \\operatorname {s i g n} \\left(G _ {i} - G _ {j}\\right) \\cdot \\operatorname {s i g n} \\left(P _ {i} - P _ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 556, + 468, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{ij} = \\frac{1}{i + j}$ and $\\mathrm{sign}(\\cdot)$ is the signum function. A larger $\\tau_w$ indicates more consistencies between the predicted ranking and the ground truth, which means a better model selection algorithm.", + "bbox": [ + 75, + 599, + 468, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Revisiting the Fine-tuning Optimization", + "text_level": 1, + "bbox": [ + 76, + 667, + 418, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a pre-trained model $\\Psi$ transfers to a downstream task $\\mathcal{T} \\coloneqq \\{\\mathcal{X},\\mathcal{Y}\\}$ , it undergoes extensive optimization for evolving from its initial state (before fine-tuning) to the final state (after fine-tuning), resulting in complex changes in both feature space and output space. Since the transferability metric is designed to evaluate the final state without fine-tuning, it is essential and challenging to model the optimization process without lengthy computations, bridging the gap between the initial and the final state. To gain insights into the underlying optimization mechanisms, we conduct an analysis to compare the prior arts with fine-tuning to revisit the evolution process toward the final state.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works SFDA [41] and PED [32] (called simulation methods) aim to approximate fine-tuning procedures", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by simulating feature dynamics. They design an objective function $U$ to assess the class separability of the feature space. Then, they apply linear transformations to increase $U$ and feed modified features into a classifier to update logits, as shown below:", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA, b = \\arg \\max _ {A, b} U (A \\cdot f (\\mathcal {X}) + b, \\mathcal {Y}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 172, + 890, + 200 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {C} (A \\cdot f (\\mathcal {X}) + b) - \\mathcal {C} (f (\\mathcal {X})),\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 202, + 831, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $A, b$ denote linear transformation parameters, $\\mathcal{C}$ denotes a machine learning classifier and $\\mathcal{F}$ denotes the function that outputs logits. As a result of the optimization in Eq. (2), the objective of enhancing class separability is achieved. However, fine-tuning directly minimizes the gap between predicted logits and labels and performs gradient descent on parameters to update the output logits as follows:", + "bbox": [ + 496, + 227, + 890, + 333 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\theta = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L} (\\mathcal {F} (\\mathcal {X} , \\theta) , \\mathcal {Y})}{\\mathrm {d} \\theta},\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 353, + 784, + 383 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {F} (\\mathcal {X}, \\theta + \\Delta \\theta) - \\mathcal {F} (\\mathcal {X}, \\theta) \\approx \\Delta \\theta \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 385, + 890, + 414 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n= - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\mathcal {F}} \\left(\\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 416, + 834, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta$ denotes the learning rate, $\\mathcal{L}$ denotes loss function, and $\\theta$ denotes parameters of $\\mathcal{F}$ . Eq. (3) provides the variation of logits within one step, capturing the nonlinear dynamics based on gradients. When comparing the optimization outlined in Eq. (2) and Eq. (3), it becomes evident that simulation methods differ from fine-tuning in terms of objectives and optimization approaches.", + "bbox": [ + 496, + 459, + 890, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite preliminary exploration, such simulation methods come with severe limitations due to these inconsistencies, leading to deviations in predicting the final state, as illustrated in Fig. 3. Specifically, simulation methods aim to enhance class separability within the feature space, whereas fine-tuning drives logits closer to labels. The insufficient objective alignment introduces distinct inductive bias compared to fine-tuning, resulting in notable variations in modeling dynamics [27, 34]. Moreover, due to the nonlinear nature of neural networks and complex parameter interactions, linear approximations in simulation methods often fail to accurately capture the dynamic changes [21]. Such limitations hinder the effective modeling of nonlinearity, further constraining modeling precision.", + "bbox": [ + 496, + 566, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Modeling Logits Evolution Process through Dynamical Equation", + "text_level": 1, + "bbox": [ + 498, + 786, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the aforementioned problems, our study delves deeper into the optimization analysis from Eq. (3). In fact, unlike features, the logits derived from the optimization output naturally align with the fine-tuning objectives and reflect the inherent nonlinearity from gradient descent. Therefore,", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28666", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/afd73633fe40bc8a1e10ada7da220617e5f53e607aef417db3824deb6a4ce21c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 89, + 449, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/799fab120c21ac44732bf9677d4919e8a10a9f12d1704ade2fb1d619870bd7b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 90, + 888, + 301 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/772c7f10ffb50c6f1a01eeb70f507672a56d23c57a1a790ab18c4238c5c4ccae.jpg", + "image_caption": [ + "Figure 2. Comparison of different approaches for modeling the evolution process. (a) Fine-tuning iteratively updates parameters via gradient descent, revealing authentic logits evolution trajectory and classification accuracy for ground-truth ranking. (b) Recent works (e.g., SFDA [41], PED [32]) perform linear transformations in feature space to enhance class separability, updating logits with modified features. The ranking score is determined by Cross-Entropy loss on updated logits. While these approaches simulate fine-tuning impact on features, they often suffer huge deviations from the authentic results. (c) Our method, LEAD, transforms the discrete logit evolution during fine-tuning into a gradient-based dynamical equation to capture nonlinear changes and can efficiently predict the final state through the closed-form solution. It establishes an analysis framework linking the initial and final states, enabling the precise prediction in one step." + ], + "image_footnote": [], + "bbox": [ + 78, + 200, + 447, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "it provides a particularly relevant entry point to analyze dynamical changes during fine-tuning and assess the model transferability through its evolution process.", + "bbox": [ + 75, + 428, + 468, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the dynamical process of logit space analyzed by Eq. (3) remains consistent with fine-tuning, there still lacks an efficient approach for capturing logits evolution, due to the discrete process of iterating over thousands of steps by traditional optimization. To this end, we propose a new transferability metric called \"Logits Evolution Analysis through Dynamical equation (LEAD)\", that bridges the evolution process from the initial logit space towards the final state. In contrast to extensive optimization iterations, our method proposes a dynamical analysis framework for analyzing the nonlinear effects within the logit space, providing an efficient prediction to capture the final state with a concise solution in a single step.", + "bbox": [ + 75, + 473, + 468, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical Background. In our LEAD, we innovatively introduce the Neural Tangent Kernel (NTK) to simplify the dynamical equation. Here, we provide a brief introduction to the definition of the NTK kernel function and the constant-preserving property that we utilize.", + "bbox": [ + 75, + 670, + 468, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}) \\triangleq \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) ^ {\\top}, (4) \\\\ \\Phi (\\mathcal {X}, \\mathcal {X}) \\triangleq \\lim _ {l \\rightarrow \\infty} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}), (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 753, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla$ denotes the Laplace Operator, and $l$ denotes the root mean square of the widths of all layers in the network. The kernel function $\\hat{\\Phi}_{\\theta_t}$ remains invariant throughout the training process as $l$ approaches infinity with an approximation error of $O\\left(\\frac{1}{l^2}\\right)$ [20, 26]. For simplicity, in the remainder of this paper, we denote functions at time $t$ as:", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86f95734fc810786cb98edbae5ec9033cac5a37e77337b4e6c0b793079762518.jpg", + "image_caption": [ + "Figure 3. Comparison of the average prediction of final state logits across various methods on VOC2007 [16], where the actual result stems from fine-tuning. Closer to actual results indicates more accurate modeling of the final state. Simulation methods (SFDA, PED) capture the correct trend but exhibit a large deviation from the actual results, affecting the accuracy of model ranking." + ], + "image_footnote": [], + "bbox": [ + 501, + 425, + 893, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{F}_t(\\cdot)\\triangleq \\mathcal{F}(\\cdot ,\\theta_t),\\hat{\\Phi}_t(\\cdot)\\triangleq \\hat{\\Phi}_{\\theta_t}(\\mathcal{X},\\mathcal{X}),\\Phi \\triangleq \\Phi (\\mathcal{X},\\mathcal{X}).$", + "bbox": [ + 498, + 652, + 854, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamical Equation and Closed-form Solution. To bypass the complexity of iterative computations, we extend the analysis in Eq. (3) and theoretically derive a dynamical equation that is efficient for analysis. Specifically, we seek for a mathematical derivation through the theory of limit approximation to convert the discrete process into a continuous differential equation, enabling us to track the trajectory of logits in the solution space of the equation. Furthermore, we leverage the property of the NTK in Eq. (4) to simplify the equation into an Ordinary Differential Equation (ODE) with an initial value condition. Due to space constraints, we directly provide the equation and the detailed proof can be found in the Appendix.", + "bbox": [ + 496, + 670, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\mathrm {d} \\mathcal {F} _ {t}}{\\mathrm {d} t} = - \\eta \\cdot \\Phi \\cdot \\frac {\\mathrm {d} \\mathcal {L} _ {t}}{\\mathrm {d} \\mathcal {F} _ {t}}, \\mathcal {F} _ {0} = l o g _ {i n i t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 872, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28667", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{loginit}$ is the initial state of logits obtained through feeding features and labels of downstream datasets to a classification algorithm. Compared to the discrete process, our proposed ODE depicts the continuous evolution of the finetuning dynamics, allowing for a more flexible modeling approach. It transforms discrete gradient effects into equivalent differential effects in continuous time, preserving the capability for modeling nonlinear changes and providing precise prediction of the evolution process. Furthermore, unlike employing numerical methods as the ODE solver, we can directly obtain its closed-form solution through integration over time dimension, allowing us to bypass the lengthy evolution process and reach the optimization destination in a single step, as shown below:", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(\\mathcal {F} _ {t} (\\mathcal {X})\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\Phi \\cdot t}\\right) \\mathcal {Y} + e ^ {- \\eta \\Phi \\cdot t} \\cdot \\log_ {\\text {i n i t}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 311, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Eq. (6), the solution is derived by interpolating between the one-hot label and the initial observation of logits, with the interpolation coefficients determined by the NTK matrix. It comprises two crucial elements: $\\text{log}_{\\text{init}}$ , representing the value of logits before optimization and determining the start of the evolution process; and $\\Phi$ , indicating the convergence rate to the label and determining the trajectory of logit evolution. Our proposed framework efficiently models the evolution process, enabling a prediction of real-time changes.", + "bbox": [ + 75, + 340, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Class-aware Decomposition. While Eq. (6) provides the theoretical result for the final state, computing the value of $\\Phi$ under finite width remains challenging. Prior studies [20, 53] indicate that, even with finite width, the convergence rate can be assessed by using the eigenvalues of the initial state NTK matrix $\\hat{\\Phi}_0$ . Inspired by this empirical finding, we further conduct a decomposition approach to bridge theory and practice, utilizing the eigenvalue decomposition on $\\hat{\\Phi}_0$ to replace the convergence rate determined by $\\Phi$ in Eq. (6) with eigenvalues.", + "bbox": [ + 75, + 491, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the network exhibiting varying classification abilities across different classes, their convergence behavior and evolution dynamics also differ during the fine-tuning process. To enhance the modeling of behaviors in different classes, we separately compute NTK for each class, rather than mixing them together, to decouple the respective optimization effect. Specifically, we extract $n$ samples of each class to calculate their respective NTK matrices and perform decomposition to obtain eigenvalues. The average eigenvalue of each class serves as the unified parameter to determine the interpolation coefficient for this class:", + "bbox": [ + 75, + 642, + 470, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left(\\mathcal {F} _ {t} (x)\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\bar {\\lambda} \\cdot t}\\right) y + e ^ {- \\eta \\bar {\\lambda} \\cdot t} \\cdot \\log_ {i n i t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 818, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x, y$ is a given sample and its one-hot label, $\\overline{\\lambda}$ is the average eigenvalue of the NTK matrix associated with the class of $x$ . Building upon the conclusion of Eq. (6), Eq.", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(7) avoids the demand of infinite width, enabling practical computability of the prediction results. As shown in Fig. 2, we employ the results of Eq. (7) to efficiently obtain predictions for the final state logits. Subsequently, following the common practice [32, 33, 41], we feed predictions into the Cross-Entropy loss to obtain the transferability score for model ranking.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary. Our method explicitly models the fine-tuning optimization through the theoretical analysis framework and provides a concise dynamical solution to bypass the lengthy optimization process in a single step, which seamlessly integrates the initial state observations and the evolution trajectory towards the final state. Therefore, we can efficiently predict the final logit state through this dynamical solution. Through comparing logit prediction accuracy among different models, it can serve as an effective metric for predicting model transferability.", + "bbox": [ + 496, + 196, + 892, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 359, + 633, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness and robustness of our approach, we evaluate LEAD on diverse pre-trained models, including both supervised and self-supervised convolutional neural networks. Our evaluation spans 10 widely used classification benchmarks in transfer learning, employing weighted Kendall's $\\tau_w$ as rank correlation to evaluate the transferability metric. Additionally, we highlight the effectiveness of LEAD in the low-data regime, where only $2/5/10$ samples are available per class.", + "bbox": [ + 496, + 383, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 529, + 718, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a pre-trained model $\\Psi$ , we append a random initialized [22] MLP $h$ as a classification head after its backbone $f$ , and combine them as $\\mathcal{F} \\coloneqq h \\circ f$ to produce logits. To compute the logits in the final state, we only need to derive the NTK matrix of each class and $log_{init}$ through Eq. (7). We follow the approach in [35] to compute the NTK matrix and employ a robust and efficient machine learning classifier, the multi-class SVM [49], to generate $log_{init}$ . Due to space constraints, we provide a more detailed explanation and pseudocode in the Appendix.", + "bbox": [ + 496, + 551, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 500, + 710, + 635, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Target Datasets. We employ 10 classification benchmarks commonly utilized in transfer learning research including Caltech-101 [17], Stanford Cars [29], CIFAR10 [30], CIFAR100 [30], DTD [10], Oxford 102 Flowers [37], Food-101 [4], Oxford-IIIT Pets [39], SUN397 [51], and VOC2007 [16]. These datasets encompass abundant characteristics, including backgrounds, textures, and coarse/fine-grained scenes, spanning diverse fields.", + "bbox": [ + 496, + 733, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ground Truth Rank. To conduct the ground truth ranking for the model zoo, we follow the approach in [41] and [52]. Specifically, we employ a grid search strategy to determine", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28668", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/53306db3195106ed608cf304254ce6720c67983e0cbd5c2889f21459be68f674.jpg", + "table_caption": [ + "Table 1. Performance comparisons of different transferability metrics on the supervised model zoo. We measure the performance with weighted Kendall's $\\tau_w$ and a larger $\\tau_w$ represents a better prediction rank. The best results are in bold and the second-best are in underline." + ], + "table_footnote": [], + "table_body": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3630.4220.0660.135-0.0920.5360.3100.114-0.0970.707
Logme [52]ICML'210.3360.3490.4250.5380.6030.6510.7830.3720.4400.673
Nsleep [33]CVPR'210.5040.6780.2920.4350.6860.4520.6560.6860.7700.645
PACTran [12]ECCV'220.7090.3720.4800.1510.7420.3720.7570.4510.4220.536
SFDA [41]ECCV'220.5340.5970.4210.4720.7490.4210.8170.6250.5620.728
GBC [38]CVPR'220.8440.4300.3480.7350.8300.4800.8660.3580.5740.654
ETran [18]ICCV'230.7090.6270.5340.5570.7490.6050.7190.6510.5770.673
PED [32]ICCV'230.5730.5970.4510.4720.7490.4210.7700.6250.5620.728
LEADThis paper0.8920.6980.7860.5860.8350.6890.7910.8410.6090.743
", + "bbox": [ + 80, + 128, + 888, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the actual performance of each model in the downstream task. This strategy involves trying various learning rates from the set $\\{10^{-1}, 10^{-2}, 10^{-3}, 10^{-4}\\}$ and different weight decay values from the set $\\{0, 10^{-6}, 10^{-5}, 10^{-4}, 10^{-3}\\}$ . To ensure reliability, we repeat each experiment 5 times with different random seeds and take the average as the result.", + "bbox": [ + 75, + 297, + 470, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation on Supervised Models", + "text_level": 1, + "bbox": [ + 76, + 396, + 372, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Zoo. We construct a Model Zoo that includes 12 CNN models supervised pre-trained on ImageNet [11], covering widely-used architectures: ResNet-34 [23], ResNet-50 [23], ResNet-101 [23], ResNet-152 [23], DenseNet-121 [25], DenseNet-161 [25], DenseNet-169 [25], DenseNet-201 [25], MNet-A1 [45], MobileNetV2 [40], GoogleNet [43], and InceptionV3 [44]. The accuracy results of these models after fine-tuning can be found in the Appendix.", + "bbox": [ + 75, + 417, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result Analysis. To show the effectiveness of LEAD, we compare it with prior arts, such as Bayesian-based Logme, Separability-based GBC, and simulation methods SFDA and PED. As shown in Tab. 1, simulation methods consider underlying dynamic effects and exhibit promising performances on some datasets. However, as the aforementioned analysis in Sec. 3.2 shows, the bias in the simulation process causes these methods to have unstable improvements, particularly resulting in unsatisfactory performance on Flowers, Food, and Pets. Benefiting from a precise analysis of the evolution, our LEAD yields a performance gain of $+0.335$ , $+0.319$ , $+0.216$ compared to PED on these three datasets, respectively. Overall, LEAD achieves the best $\\tau_w$ on 7 target datasets and obtains an average $\\tau_w$ of 0.747 which is relatively $18\\%$ better than the latest SOTA ETran [18]. Our experiments highlight the importance of accurate modeling of dynamics and confirm the effectiveness of our method to model the evolution process through the dynamical equation.", + "bbox": [ + 75, + 537, + 470, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation on Self-supervised Models", + "text_level": 1, + "bbox": [ + 76, + 833, + 405, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Zoo. Self-supervised learning models (SSL models) have shown their efficacy in transfer learning and often display superior generalization performance compared", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to Supervised models. To illustrate the broad applicability of our approach, inspired by [32, 41], we assemble a self-supervised model zoo trained on ImageNet with BYOL [19], Infomin [46], PCLv1 [31], PCLv2 [31], Selav2 [1], InsDis [50], SimCLRv1 [7], SimCLRv2 [8], MoCov1 [24], MoCov2 [9], DeepClusterv2 [5], and SWAV [6].", + "bbox": [ + 496, + 297, + 893, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result Analysis. Due to differing pre-training and downstream classification objectives, SSL models undergo larger dynamic changes during the fine-tuning process, leading to distinct transferability properties compared to supervised models [15]. As shown in Tab. 1 and 2, methods that disregard dynamics (e.g. PARC, Nsleep) perform competitively in the supervised model zoo, but fail on some datasets for SSL models. For example, Nsleep and PARC have a negative small rank correlation on VOC2007 (-0.101) and CIFAR100 (-0.136), respectively. To maintain fairness, we refrain from comparing with GBC [38] in the SSL model zoo, which is designed for supervised scenarios. On the contrary, SFDA and PED exhibit better performances due to consideration of training dynamics. Nevertheless, the limitation of the simulating process still hinders their performance. Compared to them, LEAD achieves a consistent improvement on 7 datasets and obtains an average $\\tau_w$ of 0.746 which is relatively $15\\%$ better than PED (0.648). The results validate the effectiveness of LEAD for SSL models.", + "bbox": [ + 496, + 388, + 893, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Evaluation on Low-data Regime", + "text_level": 1, + "bbox": [ + 498, + 683, + 785, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. In practical application, privacy and resource constraints may limit our access to the entire downstream dataset. To explore the utilization boundary, we evaluate whether LEAD can provide effective model ranking only using limited samples. We follow three data settings in PACTran [12] with increasing average number of samples per class $\\{2,5,10\\}$ and divide 10 datasets into two groups according to the number of classes: $100+$ classes, including Food, Caltech, Flowers, Cars, CIFAR100, and SUN397; 10-99 classes, including DTD, CIFAR10, Pets, and VOC2007. For each setting, we repeatedly sample 5 times for evaluation and take the average as the result.", + "bbox": [ + 496, + 703, + 893, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Result Analysis. As shown in Tab. 3, all methods, in-", + "bbox": [ + 500, + 885, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28669", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/280a395d418e0a632e7a7933c2d5f798a08610cd9966d7d43289875207612f96.jpg", + "table_caption": [ + "Table 2. Performance comparisons on the self-supervised model zoo. The best results are in bold and the second-best are in underline." + ], + "table_footnote": [], + "table_body": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3590.1960.6220.424-0.1360.4470.1470.496-0.0060.618
Logme [52]ICML'210.5700.0510.6040.375-0.0080.6270.2950.6840.2170.158
Nsleep [33]CVPR'210.5740.5250.5340.4860.2760.641-0.0440.7920.719-0.101
PACTran [12]ECCV'220.7200.6220.6010.4740.5290.6140.4770.6410.6380.620
SFDA [41]ECCV'220.6850.5230.7490.5150.5480.7730.6190.5860.6980.568
ETran [18]ICCV'230.4650.4050.6440.5870.6500.2140.6060.3380.5200.376
PED [32]ICCV'230.5810.6140.7770.6490.5680.9070.6730.4620.6610.583
LEADThis paper0.8600.7800.7250.6630.7760.8250.7130.6290.7600.723
", + "bbox": [ + 81, + 114, + 888, + 256 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ddf1ad6e63472cda2967e11a0a1f7b9d6c98097fc163b549cb1b64d29d575b34.jpg", + "table_caption": [ + "Table 3. Performance comparisons with different $N / K$ , where $K$ is the number of classes, $N$ is the number of samples for computing the metric. We divide 10 benchmarks into two groups according to the number of classes: 100+ classes and 10-99 classes, and evaluate the average of $\\tau_w$ on the two groups. Additionally, we also provide the average results on the 10 benchmarks, recorded as Avg. The best results are denoted in bold, the second best results are denoted in underline, and - denotes cannot give a valid ranking." + ], + "table_footnote": [], + "table_body": "
Supervised / Self-supervised pre-trained Model
MethodN/K = 2N/K = 5N/K = 10
100+ classes10-99 classesAvg100+ classes10-99 classesAvg100+ classes10-99 classesAvg
GBC0.17 / 0.320.10 / 0.170.13 / 0.250.45 / 0.400.22 / 0.180.36 / 0.310.59 / 0.420.42 / 0.220.52 / 0.34
PACTran0.43 / 0.490.20 / 0.390.33 / 0.450.42 / 0.480.31 / 0.440.38 / 0.470.56 / 0.520.48 / 0.460.53 / 0.50
ETran0.42 / 0.430.12 / 0.410.29 / 0.420.48 / 0.470.31 / 0.380.41 / 0.430.56 / 0.470.56 / 0.480.56 / 0.48
PED- / 0.36- / 0.24- / 0.31- / 0.43- / 0.35- / 0.400.30 / 0.500.22 / 0.560.27 / 0.52
LEAD0.45 / 0.480.51 / 0.450.47 / 0.460.48 / 0.490.64 / 0.530.54 / 0.510.65 / 0.530.64 / 0.600.64 / 0.56
", + "bbox": [ + 81, + 333, + 888, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cluding PACTran designed for this scenario, exhibit significant performance declines under low-data settings. For instance, ETran's performance in the supervised model zoo drops from 0.64 with the entire dataset to 0.29, 0.41, and 0.56 in three low-data settings. Moreover, some methods, such as SFDA and PED fail to provide prediction rankings in some settings. For example, when $N / K = 2,5$ , PED cannot correctly evaluate class separability due to insufficient data, imposing huge changes to completely separate different classes. This leads to all models perfectly classifying all samples, resulting in identical scores that cannot be ranked. In contrast, our LEAD achieves stable results, ensuring $\\tau_w \\geq 0.45$ even in extreme settings like $N / K = 2$ , and consistently outperforms other methods. For example, under $N / K = 10$ , its average $\\tau_w$ in the supervised model zoo reaches 0.64, which is relatively $14\\%$ better than the latest SOTA ETran (0.56) and is equivalent to the latest SOTA on the complete dataset, with less than $5\\%$ of data size. The results demonstrate the robustness of our method in the face of data limitations.", + "bbox": [ + 75, + 457, + 472, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study and Efficiency Analysis", + "text_level": 1, + "bbox": [ + 76, + 771, + 428, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, without loss of generality, we conduct ablation studies on Caltech, CIFAR10, and VOC2007, building upon the evaluation of SSL models in Sec. 4.4 for saving time. We conduct separate experiments to analyze the impact of interpolation coefficients, a key component of LEAD, aiming to delve into the influence of hyperparameters and implementation details, gaining insights into key", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "factors contributing to LEAD's effectiveness. Additionally, we provide a runtime comparison to verify efficiency, and more ablation study results are provided in the appendix.", + "bbox": [ + 496, + 458, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Interpolation Coefficient", + "text_level": 1, + "bbox": [ + 498, + 513, + 725, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The interpolation coefficients are determined by time coefficient $t$ and the NTK matrix. We conduct ablation experiments on these two key factors:", + "bbox": [ + 496, + 535, + 890, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Coefficient $t$ . As we model the discrete process as a continuous ODE, the continuous variable $t$ is proportional to the number of updates, determining the progress of the evolution process. In Fig. 4, we evaluate our method under varied time settings, keeping the unit as 1e6 consistent across all experiments. It is evident that performance improves as the time increases within a specific range. However, if the dynamic effect persists for an extended period (e.g., 10.0), all model predictions converge too closely to the ground-truth label, reducing discrimination for model ranking. Hence, we set the time coefficient as 1 by default.", + "bbox": [ + 496, + 582, + 892, + 750 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8befea459eb8901a1a723d8409b407c7f5f958b1a950897281e0ed8c9dfe50a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 763, + 697, + 866 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b3c6d2bb15e3a12791b480ffe8032eff56a800c64b8ddf7c8ef96a1d83c0eba.jpg", + "image_caption": [ + "Figure 4. The left picture shows the performance gain in $\\tau_w$ with respect to different time $t$ and the right shows the performance in different sample sizes for each class NTK." + ], + "image_footnote": [], + "bbox": [ + 694, + 763, + 872, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sample Size for NTK. To enhance the perception of different dynamic intensities of classes, we propose the class-aware decomposition that employs $S$ samples of each class to obtain NTK matrix and perform spectrum analysis to derive eigenvalues. In Fig. 4, we conduct ablation experiments to explore the effect of varying $S$ . We observe that performance generally improves as $S$ increases, and gradually reaches stability. This demonstrates that when the sample size reaches a certain size, eigenvalues reliably capture differences in dynamic intensity, offering a refined analysis result of the dynamics. Therefore, we set $S = 64$ to ensure the method's efficiency and stability in all experiments.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Running Time Comparison", + "text_level": 1, + "bbox": [ + 76, + 276, + 326, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In experiment results, we have demonstrated the effectiveness of LEAD in various scenarios. In this section, we highlight the computational efficiency of LEAD. With the concise solution in Eq. (7), LEAD predicts the final state only requiring the NTK matrix and $\\text{log}_{init}$ . As shown in Tab. 4, the running time of LEAD is about half that of simulation methods (45s compared to 92s), ranking second after GBC which displays unstable performance. Notably, LEAD takes 3s to compute the NTK, while feature extraction takes 420s, resulting in only a marginal $0.7\\%$ increase in GPU time. The time to compute $\\text{log}_{init}$ is 42s CPU time, indicating significant potential for exploring faster classification algorithms to further reduce the overall time.", + "bbox": [ + 75, + 296, + 472, + 492 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/353febb933a476c50c193c6ea429c60bf7e23388e63e622dbfdea7ad1647a374.jpg", + "table_caption": [ + "Table 4. The comparisons of average running time on CIFAR10." + ], + "table_footnote": [], + "table_body": "
MetricsPARCNLEEPPACTranSFDAGBCPEDOurs
Running Time111s1430s77s92s26s97s45s
", + "bbox": [ + 80, + 505, + 468, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Visualization and Analysis", + "text_level": 1, + "bbox": [ + 76, + 534, + 323, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we validate our approach through some visualizations and these results clearly illustrate how our model effectively addresses the model selection problem.", + "bbox": [ + 75, + 553, + 468, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Logit Prediction", + "text_level": 1, + "bbox": [ + 76, + 604, + 238, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have conducted a comparison of the predicted logits distribution on VOC2007 using different methods, as illustrated in Fig. 3. Notably, predictions obtained by LEAD closely match the authentic distribution after fine-tuning. To provide further quantitative analysis, we visualize the average bias in predicted logits in Fig. 5. Across three datasets, our method exhibits a substantial reduction compared to simulation methods, achieving nearly a $50\\%$ overall reduction. This visualization highlights the precision of our dynamical equation-based approach in capturing logit evolution for more accurate ranking sequences.", + "bbox": [ + 75, + 625, + 472, + 792 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5730efca410870bc74277eafa6b7679899fbff59307e049120384f6d9a9826e7.jpg", + "image_caption": [ + "Figure 5. Comparison of the average difference between logits predicted by different methods and logits after fine-tuning. The difference is calculated utilizing the Euclidean distance." + ], + "image_footnote": [], + "bbox": [ + 96, + 797, + 450, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Rank Variation", + "text_level": 1, + "bbox": [ + 500, + 90, + 656, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "By employing our method, we can refine models that have poor initial observation but show better progress during training. To assess the effectiveness of our approach, we conduct a visualization comparing the model rankings between the initial observations and our predicted final state, as shown in Fig. 6. The results clearly illustrate a substantial enhancement in the calibration of model rankings when using our refined observations, with the optimal model now positioned closely to the reference line in the figure. Notably, models starting from an initially unfavorable point can be rapidly elevated to achieve a superior ranking, highlighting the effectiveness of our methodology.", + "bbox": [ + 496, + 108, + 893, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f3583d5c16fb243a02f5b430a4256595b02cb4df51f2e295eacd8a799197982.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 297, + 620, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6876fb90cafd3d8cdb83f6640d4e998b199078d9922496892cbbb318c52a4ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 296, + 751, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/47164e21d643dbb8ece11d0d1697e86c74aac8c23ea55f72173eb84e597ae7c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 296, + 885, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/075db0296cb59584eb3a8b14d5d198af5096a3c0919fb40315dd6521e9dd9594.jpg", + "image_caption": [ + "Figure 6. The variation in model ranking, utilizing initial states (first row) and predicted final states of LEAD (second row). Results span three datasets, from left to right: CIFAR10, Caltech, and VOC2007. Shaded areas reflect the range of the ranking error." + ], + "image_footnote": [], + "bbox": [ + 501, + 397, + 620, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b4b2ab39d6d5f3823aa39cfe2de811aba1c89e26dfe2f179ef288de09157e83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 397, + 751, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/126bb530f74b06ea081490901547b8ad5456047169c6db2349993f4dba8e19d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 397, + 885, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 500, + 554, + 777, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a dynamical equation designed to capture the evolution process within the logit space, offering an efficient solution for assessing model transferability. We construct a theoretical analysis framework and further devise a more practical and precise solution. And we demonstrate its consistent effectiveness across various scenarios including supervised and self-supervised, full-data and low-data settings. In future work, we intend to expand the current analysis framework beyond classification to cover a broader range of downstream tasks. We hope that our work can enhance the understanding of optimization mechanisms in deep neural networks and shed light on other fields, such as the design of optimization algorithms.", + "bbox": [ + 496, + 571, + 893, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 777, + 660, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation and in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd.", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv, 2019. 6", + "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313, 2019. 2", + "[3] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. In NeurIPS, pages 19301-19312, 2021. 2, 6, 7", + "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5", + "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 6", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020. 1, 6", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6", + "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, pages 22243-22255, 2020. 6", + "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv, 2020. 6", + "[10] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6", + "[12] Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, and Radu Soricut. Pactran: Pac-bayesian metrics for estimating the transferability of pretrained models to classification tasks. In ECCV, pages 252-268. Springer, 2022. 2, 6, 7", + "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In CVPR, pages 12387–12396, 2019. 1, 2", + "[14] Kshitij Dwivedi, Jiahui Huang, Radoslaw Martin Cichy, and Gemma Roig. Duality diagram similarity: a generic framework for initialization selection in task transfer learning. In ECCV, pages 497-513, 2020. 1, 2", + "[15] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In CVPR, pages 5414-5423, 2021. 6", + "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88:303–338, 2010. 4, 5", + "[17] Li Fei-Fei. Learning generative visual models from few training examples. In CVPR workshop, 2004. 5" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Mohsen Gholami, Mohammad Akbari, Xinglu Wang, Behnam Kamranian, and Yong Zhang. Etran: Energy-based transferability estimation. In ICCV, pages 18613-18622, 2023. 2, 6, 7", + "[19] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 6", + "[20] Boris Hanin and Mihai Nica. Finite depth and width corrections to the neural tangent kernel. In ICLR, 2020. 2, 4, 5", + "[21] Hangfeng He and Weijie Su. The local elasticity of neural networks. In ICLR, 2020. 3", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 6", + "[25] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 6", + "[26] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In NeurIPS, 2018. 2, 4", + "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 3", + "[28] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In CVPR, pages 2656-2666, 2018. 1", + "[29] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5", + "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv, 2020. 6", + "[32] Xiaotong Li, Zixuan Hu, Yixiao Ge, Ying Shan, and Ling-Yu Duan. Exploring model transferability through the lens of potential energy. In ICCV, pages 5429-5438, 2023. 1, 2, 3, 4, 5, 6, 7", + "[33] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In CVPR, pages 2663-2673, 2021. 1, 2, 5, 6, 7", + "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 3", + "[35] Mohamad Amin Mohamadi and Danica J. Sutherland. A fast, well-founded approximation to the empirical neural tangent kernel. In ICML, 2022. 2, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28672", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305, 2020. 2", + "[37] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5", + "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In CVPR, pages 9172-9182, 2022. 2, 6", + "[39] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505, 2012. 5", + "[40] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6", + "[41] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In ECCV, pages 286-302, 2022. 1, 2, 3, 4, 5, 6, 7", + "[42] Grace S. Shieh. A weighted kendall's tau statistic. Statistics Probability Letters, 39(1):17-24, 1998. 1", + "[43] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 6", + "[44] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 6", + "[45] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 6", + "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 6", + "[47] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 1, 2", + "[48] Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. 2", + "[49] Jason Weston, Chris Watkins, et al. Support vector machines for multi-class pattern recognition. In *Esann*, pages 219–224, 1999. 5", + "[50] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 6", + "[51] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492. IEEE, 2010. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143, 2021. 2, 5, 6, 7", + "[53] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In CVPR, pages 19228-19238, 2022. 5" + ], + "bbox": [ + 501, + 90, + 892, + 203 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28673", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_model.json b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e07a6f6f943625ddb700ec2049c63b4806d65e94 --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.131, + 0.792, + 0.154 + ], + "angle": 0, + "content": "LEAD: Exploring Logit Space Evolution for Model Selection" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.18, + 0.862, + 0.2 + ], + "angle": 0, + "content": "Zixuan Hu\\(^{1}\\) Xiaotong Li\\(^{1}\\) Shixiang Tang\\(^{3}\\) Jun Liu\\(^{4}\\) Yichun Hu\\(^{1}\\) Ling-Yu Duan\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.201, + 0.854, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science, Peking University, Beijing, China, \\(^{2}\\) Peng Cheng Laboratory, Shenzhen, China," + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.217, + 0.888, + 0.27 + ], + "angle": 0, + "content": "3 The Chinese University of Hong Kong, Hongkong, China, 4 Singapore University of Technology and Design, Singapore {hzxuan, hycc, lingyu}@pku.edu.cn, lixiaotong@stu.pku.edu.cn, shixiangtang@cuhk.edu.hk, jun.liu@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.337, + 0.474, + 0.745 + ], + "angle": 0, + "content": "The remarkable success of \"pretrain-then-finetune\" paradigm has led to a proliferation of available pre-trained models for vision tasks. This surge presents a significant challenge in efficiently choosing the most suitable pretrained models for downstream tasks. The critical aspect of this challenge lies in effectively predicting the model transferability by considering the underlying fine-tuning dynamics. Existing methods often model fine-tuning dynamics in feature space with linear transformations, which do not precisely align with the fine-tuning objective and fail to grasp the essential nonlinearity from optimization. To this end, we present LEAD, a finetuning-aligned approach based on the network output of logits. LEAD proposes a theoretical framework to model the optimization process and derives an ordinary differential equation (ODE) to depict the nonlinear evolution toward the final logit state. Additionally, we design a class-aware decomposition method to consider the varying evolution dynamics across classes and further ensure practical applicability. Integrating the closely aligned optimization objective and nonlinear modeling capabilities derived from the differential equation, our method offers a concise solution to effectively bridge the optimization gap in a single step, bypassing the lengthy fine-tuning process. The comprehensive experiments on 24 supervised and self-supervised pre-trained models across 10 downstream datasets demonstrate impressive performances and showcase its broad adaptability even in low-data scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.763, + 0.21, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.47, + 0.877 + ], + "angle": 0, + "content": "The \"pretrain-then-finetune\" learning paradigm has exhibited remarkable success in a wide range of computer vision tasks [6, 28], leading to an increasing number of diverse pretrained models open-sourced to the computer vision community. This surge poses a significant challenge in selecting the optimal pre-trained model for a downstream task" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.302, + 0.895, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.467, + 0.895, + 0.58 + ], + "angle": 0, + "content": "Figure 1. This diagram illustrates the task of model selection, which aims to rank pre-trained models for choosing the optimal one. The ground-truth ranking sequence is derived from the models' accuracy after fine-tuning. However, due to the time-intensive nature of this process, it requires to design efficient transferability metrics for accurate ranking prediction. The performance of this metric is evaluated using rank correlation, e.g., weighted Kendall's tau [42], to compare the predicted and ground-truth rankings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.701 + ], + "angle": 0, + "content": "from a vast model zoo. Considering that the brute-force solution of fine-tuning all models on downstream tasks may be infeasible in many scenarios due to the substantial computational costs, the model selection task emerges [13, 47], which ranks the transferability of pre-trained models to the given task without incurring much cost, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Predicting model transferability aims to efficiently estimate the final state after fine-tuning on the downstream task, without network optimization by backward propagation. Prior studies only depend on specific properties of the initial state (e.g., representation similarity [13, 14], conditional distribution [33, 47]), to serve as transferability indicators. Recent studies reveal that modeling the dynamic process of fine-tuning has become the bottleneck to design more high-quality transferability metrics. Following this philosophy, recent works propose to simulate fine-tuning dynamics on features through multiple stages of linear transformations [32, 41], achieving some promising results." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Despite recent progress in this field, there are still two" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.227, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28664" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "considerable gaps between existing modeled fine-tuning dynamics and the real ones: (1) Insufficient objective alignment. The designed objectives in prior works do not seamlessly align with fine-tuning, as they often focus more on enhancing class separability via feature manipulation, rather than the classification optimization for both the feature and output spaces. (2) Disregarding nonlinear modeling. Existing methods relying solely on linear transformations cannot well capture the intricate nonlinear changes in the feature space during optimization. Such constraints contribute to growing deviations in dynamic modeling and result in unreliable predictions of the final state." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.458 + ], + "angle": 0, + "content": "Delving deeper into the optimization process, we focus on modeling the dynamic changes within the output space (i.e., logits for classification), which not only more closely relates to the fine-tuning objective but also embeds the overall nonlinear dynamics from the network optimization during fine-tuning. Building upon this foundation, we introduce an insightful framework to depict the trajectory of logits toward their final state in a dynamical modeling system. To capture and understand this evolution process of the logit space, the framework provides a discrete iterative optimization equation via gradient analysis and captures nonlinear logit space changes during optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.732 + ], + "angle": 0, + "content": "In contrast to the extensive iterations required in fine-tuning optimization, we further extend the framework to propose a new transferability metric \"Logits Evolution Analysis through Dynamical Equation (LEAD)\", to seize the final evolutionary state in a single step. We perform a mathematical derivation to convert the discrete process into a continuous differential equation and simplify it into an ordinary differential equation (ODE) by leveraging the classical analysis tool, Neural Tangent Kernel (NTK) [26]. This ODE concisely models the nonlinear evolution in logit space through its closed-form solution, allowing for real-time estimation of logits. Additionally, we introduce a class-aware decomposition method to enhance prediction accuracy and break the barrier of theoretical assumption of NTK. As a result, our method provides a direct estimation of the evolution by incorporating both the initial model state and the impact of fine-tuning evolution, serving as an effective indicator assessing model transferability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The contributions of this work are three-fold. (1) To the best of our knowledge, we are the first to explore model transferability by modeling the evolution process in the logit space. (2) We propose LEAD, a novel transferability metric with two key components: a theoretical framework providing a closed-form solution to depict the logits evolution towards the final state, and a class-aware decomposition method to guarantee the practical applicability of theoretical findings and improve predictive precision. (3) We validate the effectiveness and generalizability of our method through experiments on 24 different pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.153 + ], + "angle": 0, + "content": "models across 10 downstream datasets, achieving SOTA results with an average of \\(17\\%\\) gain in the rank correlation. In addition, our method can be naturally extended to low-data scenarios while maintaining competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.165, + 0.643, + 0.18 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.186, + 0.71, + 0.203 + ], + "angle": 0, + "content": "2.1. Transferability Metric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.207, + 0.895, + 0.449 + ], + "angle": 0, + "content": "The challenge of efficiently ranking pre-trained models has captured much interest of researchers, leading to the development of various metrics to evaluate the model transferability [2, 3, 12, 18, 36, 38, 47, 52]. For example, RSA [13], DDS [14] and GBC [38] are methods based on the distance measurements of the feature space. RSA and DDS assess transferability by evaluating the representation similarity between pre-trained and downstream datasets. GBC employs the Bhattacharyya coefficient to measure class separability in the feature space. Nsleep [33] and Logme [52] estimate transferability by analyzing the conditional distribution, with Nsleep utilizing Gaussian mixture model and Logme employing Bayesian optimization. However, these methods treat network outputs as static and overlook the changes introduced by the training process, leading to less effectiveness in scenes with larger fine-tuning dynamics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.896, + 0.661 + ], + "angle": 0, + "content": "Recently, SFDA [41] and PED [32] propose to simulate changes in the feature space, respectively projecting the features into a Fisher space and utilizing a physically-inspired model to enhance the class separability, achieving impacts similar to fine-tuning. While these methods achieve some improvements, they solely concentrate on the feature space, disregarding the dynamic changes introduced by the classifier, which is jointly optimized with the backbone. Additionally, they utilize linear transformations to simulate the intricate variations in fine-tuning, resulting in a significant limitation in simulation precision. Instead, our LEAD employs the ordinary differential equation to analyze logit dynamics, maintaining a closely aligned objective and simulating the nonlinear changes in fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.669, + 0.715, + 0.685 + ], + "angle": 0, + "content": "2.2. Neural Tangent Kernel" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Studying the behavior of neural network optimization processes has been an enduring challenge due to their nonlinearity and nonconvexity. Neural Tangent Kernel (NTK) [20, 26], serving as a theoretical tool to characterize the optimization process, has attracted much research attention. Some studies have explored its theoretical properties and applications. Notable works include the NTK's property of remaining constant under infinite network width [26], the explanation of the instability of training [48], the analysis of its bias and variance under finite width [20], and its acceleration algorithm [35]. While NTK can analyze network trainability [26, 35], it is constrained by the requirement of infinite width and neglects the consideration of initial log-its, resulting in the inability to predict the changes starting" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "from the initial state for a finite-width network. To tackle these issues, we innovatively propose an analysis framework leveraging the property of NTK, suitable for model selection problems to bridge the gap between the initial and final states of finite-width networks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.181, + 0.212, + 0.198 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.469, + 0.327 + ], + "angle": 0, + "content": "In this section, we first present the problem setup for the task of model selection. Then we analyze training mechanisms to model the fine-tuning optimization in a discrete process. Finally, we propose a transferability metric that converts the discrete process into a continuous dynamical equation with a concise solution to directly seize the evolution destination of logits, and feed updated logits into Cross-Entropy loss to assess transferability, as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.334, + 0.228, + 0.35 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.469, + 0.46 + ], + "angle": 0, + "content": "Consider \\(M\\) pre-trained models \\(\\{\\Psi_m\\}_{m=1}^M\\) and a downstream dataset \\(\\mathcal{T} = \\{x_n, y_n\\}_{n=1}^N\\) with \\(K\\) classes, where model \\(\\Psi\\) consists of its backbone \\(f\\) which outputs the encoded feature, \\(N\\) is the number of images in the downstream task. The purpose of model selection is to design an efficient algorithm to predict the ranking of the performance of \\(M\\) pre-trained models after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.469, + 0.551 + ], + "angle": 0, + "content": "Specifically, given the ground-truth performance rankings \\(\\{G_m\\}_{m=1}^M\\) of pre-trained models after fine-tuning and the predicted ranking \\(\\{P_m\\}_{m=1}^M\\) by model selection algorithms, we can leverage the commonly-used rank-based correlation, i.e., the weighted Kendalls' \\(\\tau_w\\), to evaluate the effectiveness of the proposed model selection methods, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.557, + 0.469, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\tau_ {w} = \\frac {1}{\\sum_ {i \\neq j} w _ {i j}} \\cdot \\sum_ {i \\neq j} w _ {i j} \\operatorname {s i g n} \\left(G _ {i} - G _ {j}\\right) \\cdot \\operatorname {s i g n} \\left(P _ {i} - P _ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.469, + 0.661 + ], + "angle": 0, + "content": "where \\( w_{ij} = \\frac{1}{i + j} \\) and \\( \\mathrm{sign}(\\cdot) \\) is the signum function. A larger \\( \\tau_w \\) indicates more consistencies between the predicted ranking and the ground truth, which means a better model selection algorithm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.669, + 0.419, + 0.685 + ], + "angle": 0, + "content": "3.2. Revisiting the Fine-tuning Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.87 + ], + "angle": 0, + "content": "When a pre-trained model \\(\\Psi\\) transfers to a downstream task \\(\\mathcal{T} \\coloneqq \\{\\mathcal{X},\\mathcal{Y}\\}\\), it undergoes extensive optimization for evolving from its initial state (before fine-tuning) to the final state (after fine-tuning), resulting in complex changes in both feature space and output space. Since the transferability metric is designed to evaluate the final state without fine-tuning, it is essential and challenging to model the optimization process without lengthy computations, bridging the gap between the initial and the final state. To gain insights into the underlying optimization mechanisms, we conduct an analysis to compare the prior arts with fine-tuning to revisit the evolution process toward the final state." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Recent works SFDA [41] and PED [32] (called simulation methods) aim to approximate fine-tuning procedures" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "by simulating feature dynamics. They design an objective function \\( U \\) to assess the class separability of the feature space. Then, they apply linear transformations to increase \\( U \\) and feed modified features into a classifier to update logits, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.174, + 0.892, + 0.202 + ], + "angle": 0, + "content": "\\[\nA, b = \\arg \\max _ {A, b} U (A \\cdot f (\\mathcal {X}) + b, \\mathcal {Y}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.203, + 0.832, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {C} (A \\cdot f (\\mathcal {X}) + b) - \\mathcal {C} (f (\\mathcal {X})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.334 + ], + "angle": 0, + "content": "where \\(A, b\\) denote linear transformation parameters, \\(\\mathcal{C}\\) denotes a machine learning classifier and \\(\\mathcal{F}\\) denotes the function that outputs logits. As a result of the optimization in Eq. (2), the objective of enhancing class separability is achieved. However, fine-tuning directly minimizes the gap between predicted logits and labels and performs gradient descent on parameters to update the output logits as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.354, + 0.785, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\Delta \\theta = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L} (\\mathcal {F} (\\mathcal {X} , \\theta) , \\mathcal {Y})}{\\mathrm {d} \\theta},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.386, + 0.891, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {F} (\\mathcal {X}, \\theta + \\Delta \\theta) - \\mathcal {F} (\\mathcal {X}, \\theta) \\approx \\Delta \\theta \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.417, + 0.835, + 0.451 + ], + "angle": 0, + "content": "\\[\n= - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\mathcal {F}} \\left(\\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where \\(\\eta\\) denotes the learning rate, \\(\\mathcal{L}\\) denotes loss function, and \\(\\theta\\) denotes parameters of \\(\\mathcal{F}\\). Eq. (3) provides the variation of logits within one step, capturing the nonlinear dynamics based on gradients. When comparing the optimization outlined in Eq. (2) and Eq. (3), it becomes evident that simulation methods differ from fine-tuning in terms of objectives and optimization approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Despite preliminary exploration, such simulation methods come with severe limitations due to these inconsistencies, leading to deviations in predicting the final state, as illustrated in Fig. 3. Specifically, simulation methods aim to enhance class separability within the feature space, whereas fine-tuning drives logits closer to labels. The insufficient objective alignment introduces distinct inductive bias compared to fine-tuning, resulting in notable variations in modeling dynamics [27, 34]. Moreover, due to the nonlinear nature of neural networks and complex parameter interactions, linear approximations in simulation methods often fail to accurately capture the dynamic changes [21]. Such limitations hinder the effective modeling of nonlinearity, further constraining modeling precision." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.891, + 0.818 + ], + "angle": 0, + "content": "3.3. Modeling Logits Evolution Process through Dynamical Equation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned problems, our study delves deeper into the optimization analysis from Eq. (3). In fact, unlike features, the logits derived from the optimization output naturally align with the fine-tuning objectives and reflect the inherent nonlinearity from gradient descent. Therefore," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28666" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.45, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.091, + 0.89, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.201, + 0.449, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.895, + 0.412 + ], + "angle": 0, + "content": "Figure 2. Comparison of different approaches for modeling the evolution process. (a) Fine-tuning iteratively updates parameters via gradient descent, revealing authentic logits evolution trajectory and classification accuracy for ground-truth ranking. (b) Recent works (e.g., SFDA [41], PED [32]) perform linear transformations in feature space to enhance class separability, updating logits with modified features. The ranking score is determined by Cross-Entropy loss on updated logits. While these approaches simulate fine-tuning impact on features, they often suffer huge deviations from the authentic results. (c) Our method, LEAD, transforms the discrete logit evolution during fine-tuning into a gradient-based dynamical equation to capture nonlinear changes and can efficiently predict the final state through the closed-form solution. It establishes an analysis framework linking the initial and final states, enabling the precise prediction in one step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.474 + ], + "angle": 0, + "content": "it provides a particularly relevant entry point to analyze dynamical changes during fine-tuning and assess the model transferability through its evolution process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.67 + ], + "angle": 0, + "content": "While the dynamical process of logit space analyzed by Eq. (3) remains consistent with fine-tuning, there still lacks an efficient approach for capturing logits evolution, due to the discrete process of iterating over thousands of steps by traditional optimization. To this end, we propose a new transferability metric called \"Logits Evolution Analysis through Dynamical equation (LEAD)\", that bridges the evolution process from the initial logit space towards the final state. In contrast to extensive optimization iterations, our method proposes a dynamical analysis framework for analyzing the nonlinear effects within the logit space, providing an efficient prediction to capture the final state with a concise solution in a single step." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.747 + ], + "angle": 0, + "content": "Theoretical Background. In our LEAD, we innovatively introduce the Neural Tangent Kernel (NTK) to simplify the dynamical equation. Here, we provide a brief introduction to the definition of the NTK kernel function and the constant-preserving property that we utilize." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.754, + 0.47, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}) \\triangleq \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) ^ {\\top}, (4) \\\\ \\Phi (\\mathcal {X}, \\mathcal {X}) \\triangleq \\lim _ {l \\rightarrow \\infty} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}), (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\nabla\\) denotes the Laplace Operator, and \\(l\\) denotes the root mean square of the widths of all layers in the network. The kernel function \\(\\hat{\\Phi}_{\\theta_t}\\) remains invariant throughout the training process as \\(l\\) approaches infinity with an approximation error of \\(O\\left(\\frac{1}{l^2}\\right)\\) [20, 26]. For simplicity, in the remainder of this paper, we denote functions at time \\(t\\) as:" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.426, + 0.895, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.652 + ], + "angle": 0, + "content": "Figure 3. Comparison of the average prediction of final state logits across various methods on VOC2007 [16], where the actual result stems from fine-tuning. Closer to actual results indicates more accurate modeling of the final state. Simulation methods (SFDA, PED) capture the correct trend but exhibit a large deviation from the actual results, affecting the accuracy of model ranking." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.855, + 0.672 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_t(\\cdot)\\triangleq \\mathcal{F}(\\cdot ,\\theta_t),\\hat{\\Phi}_t(\\cdot)\\triangleq \\hat{\\Phi}_{\\theta_t}(\\mathcal{X},\\mathcal{X}),\\Phi \\triangleq \\Phi (\\mathcal{X},\\mathcal{X}).\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Dynamical Equation and Closed-form Solution. To bypass the complexity of iterative computations, we extend the analysis in Eq. (3) and theoretically derive a dynamical equation that is efficient for analysis. Specifically, we seek for a mathematical derivation through the theory of limit approximation to convert the discrete process into a continuous differential equation, enabling us to track the trajectory of logits in the solution space of the equation. Furthermore, we leverage the property of the NTK in Eq. (4) to simplify the equation into an Ordinary Differential Equation (ODE) with an initial value condition. Due to space constraints, we directly provide the equation and the detailed proof can be found in the Appendix." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.873, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\frac {\\mathrm {d} \\mathcal {F} _ {t}}{\\mathrm {d} t} = - \\eta \\cdot \\Phi \\cdot \\frac {\\mathrm {d} \\mathcal {L} _ {t}}{\\mathrm {d} \\mathcal {F} _ {t}}, \\mathcal {F} _ {0} = l o g _ {i n i t}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "where \\( \\text{loginit} \\) is the initial state of logits obtained through feeding features and labels of downstream datasets to a classification algorithm. Compared to the discrete process, our proposed ODE depicts the continuous evolution of the finetuning dynamics, allowing for a more flexible modeling approach. It transforms discrete gradient effects into equivalent differential effects in continuous time, preserving the capability for modeling nonlinear changes and providing precise prediction of the evolution process. Furthermore, unlike employing numerical methods as the ODE solver, we can directly obtain its closed-form solution through integration over time dimension, allowing us to bypass the lengthy evolution process and reach the optimization destination in a single step, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.312, + 0.47, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(\\mathcal {F} _ {t} (\\mathcal {X})\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\Phi \\cdot t}\\right) \\mathcal {Y} + e ^ {- \\eta \\Phi \\cdot t} \\cdot \\log_ {\\text {i n i t}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.47, + 0.492 + ], + "angle": 0, + "content": "As depicted in Eq. (6), the solution is derived by interpolating between the one-hot label and the initial observation of logits, with the interpolation coefficients determined by the NTK matrix. It comprises two crucial elements: \\( \\text{log}_{\\text{init}} \\), representing the value of logits before optimization and determining the start of the evolution process; and \\( \\Phi \\), indicating the convergence rate to the label and determining the trajectory of logit evolution. Our proposed framework efficiently models the evolution process, enabling a prediction of real-time changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Class-aware Decomposition. While Eq. (6) provides the theoretical result for the final state, computing the value of \\(\\Phi\\) under finite width remains challenging. Prior studies [20, 53] indicate that, even with finite width, the convergence rate can be assessed by using the eigenvalues of the initial state NTK matrix \\(\\hat{\\Phi}_0\\). Inspired by this empirical finding, we further conduct a decomposition approach to bridge theory and practice, utilizing the eigenvalue decomposition on \\(\\hat{\\Phi}_0\\) to replace the convergence rate determined by \\(\\Phi\\) in Eq. (6) with eigenvalues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.471, + 0.809 + ], + "angle": 0, + "content": "Due to the network exhibiting varying classification abilities across different classes, their convergence behavior and evolution dynamics also differ during the fine-tuning process. To enhance the modeling of behaviors in different classes, we separately compute NTK for each class, rather than mixing them together, to decouple the respective optimization effect. Specifically, we extract \\( n \\) samples of each class to calculate their respective NTK matrices and perform decomposition to obtain eigenvalues. The average eigenvalue of each class serves as the unified parameter to determine the interpolation coefficient for this class:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.819, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left(\\mathcal {F} _ {t} (x)\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\bar {\\lambda} \\cdot t}\\right) y + e ^ {- \\eta \\bar {\\lambda} \\cdot t} \\cdot \\log_ {i n i t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(x, y\\) is a given sample and its one-hot label, \\(\\overline{\\lambda}\\) is the average eigenvalue of the NTK matrix associated with the class of \\(x\\). Building upon the conclusion of Eq. (6), Eq." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "(7) avoids the demand of infinite width, enabling practical computability of the prediction results. As shown in Fig. 2, we employ the results of Eq. (7) to efficiently obtain predictions for the final state logits. Subsequently, following the common practice [32, 33, 41], we feed predictions into the Cross-Entropy loss to obtain the transferability score for model ranking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.351 + ], + "angle": 0, + "content": "Summary. Our method explicitly models the fine-tuning optimization through the theoretical analysis framework and provides a concise dynamical solution to bypass the lengthy optimization process in a single step, which seamlessly integrates the initial state observations and the evolution trajectory towards the final state. Therefore, we can efficiently predict the final logit state through this dynamical solution. Through comparing logit prediction accuracy among different models, it can serve as an effective metric for predicting model transferability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.361, + 0.634, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.52 + ], + "angle": 0, + "content": "To demonstrate the effectiveness and robustness of our approach, we evaluate LEAD on diverse pre-trained models, including both supervised and self-supervised convolutional neural networks. Our evaluation spans 10 widely used classification benchmarks in transfer learning, employing weighted Kendall's \\(\\tau_w\\) as rank correlation to evaluate the transferability metric. Additionally, we highlight the effectiveness of LEAD in the low-data regime, where only \\(2/5/10\\) samples are available per class." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.53, + 0.719, + 0.545 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Given a pre-trained model \\(\\Psi\\), we append a random initialized [22] MLP \\(h\\) as a classification head after its backbone \\(f\\), and combine them as \\(\\mathcal{F} \\coloneqq h \\circ f\\) to produce logits. To compute the logits in the final state, we only need to derive the NTK matrix of each class and \\(log_{init}\\) through Eq. (7). We follow the approach in [35] to compute the NTK matrix and employ a robust and efficient machine learning classifier, the multi-class SVM [49], to generate \\(log_{init}\\). Due to space constraints, we provide a more detailed explanation and pseudocode in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.636, + 0.727 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Target Datasets. We employ 10 classification benchmarks commonly utilized in transfer learning research including Caltech-101 [17], Stanford Cars [29], CIFAR10 [30], CIFAR100 [30], DTD [10], Oxford 102 Flowers [37], Food-101 [4], Oxford-IIIT Pets [39], SUN397 [51], and VOC2007 [16]. These datasets encompass abundant characteristics, including backgrounds, textures, and coarse/fine-grained scenes, spanning diverse fields." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Ground Truth Rank. To conduct the ground truth ranking for the model zoo, we follow the approach in [41] and [52]. Specifically, we employ a grid search strategy to determine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28668" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.12 + ], + "angle": 0, + "content": "Table 1. Performance comparisons of different transferability metrics on the supervised model zoo. We measure the performance with weighted Kendall's \\( \\tau_w \\) and a larger \\( \\tau_w \\) represents a better prediction rank. The best results are in bold and the second-best are in underline." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.13, + 0.89, + 0.284 + ], + "angle": 0, + "content": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3630.4220.0660.135-0.0920.5360.3100.114-0.0970.707
Logme [52]ICML'210.3360.3490.4250.5380.6030.6510.7830.3720.4400.673
Nsleep [33]CVPR'210.5040.6780.2920.4350.6860.4520.6560.6860.7700.645
PACTran [12]ECCV'220.7090.3720.4800.1510.7420.3720.7570.4510.4220.536
SFDA [41]ECCV'220.5340.5970.4210.4720.7490.4210.8170.6250.5620.728
GBC [38]CVPR'220.8440.4300.3480.7350.8300.4800.8660.3580.5740.654
ETran [18]ICCV'230.7090.6270.5340.5570.7490.6050.7190.6510.5770.673
PED [32]ICCV'230.5730.5970.4510.4720.7490.4210.7700.6250.5620.728
LEADThis paper0.8920.6980.7860.5860.8350.6890.7910.8410.6090.743
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.299, + 0.471, + 0.391 + ], + "angle": 0, + "content": "the actual performance of each model in the downstream task. This strategy involves trying various learning rates from the set \\(\\{10^{-1}, 10^{-2}, 10^{-3}, 10^{-4}\\}\\) and different weight decay values from the set \\(\\{0, 10^{-6}, 10^{-5}, 10^{-4}, 10^{-3}\\}\\). To ensure reliability, we repeat each experiment 5 times with different random seeds and take the average as the result." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.373, + 0.414 + ], + "angle": 0, + "content": "4.3. Evaluation on Supervised Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Model Zoo. We construct a Model Zoo that includes 12 CNN models supervised pre-trained on ImageNet [11], covering widely-used architectures: ResNet-34 [23], ResNet-50 [23], ResNet-101 [23], ResNet-152 [23], DenseNet-121 [25], DenseNet-161 [25], DenseNet-169 [25], DenseNet-201 [25], MNet-A1 [45], MobileNetV2 [40], GoogleNet [43], and InceptionV3 [44]. The accuracy results of these models after fine-tuning can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.829 + ], + "angle": 0, + "content": "Result Analysis. To show the effectiveness of LEAD, we compare it with prior arts, such as Bayesian-based Logme, Separability-based GBC, and simulation methods SFDA and PED. As shown in Tab. 1, simulation methods consider underlying dynamic effects and exhibit promising performances on some datasets. However, as the aforementioned analysis in Sec. 3.2 shows, the bias in the simulation process causes these methods to have unstable improvements, particularly resulting in unsatisfactory performance on Flowers, Food, and Pets. Benefiting from a precise analysis of the evolution, our LEAD yields a performance gain of \\(+0.335\\), \\(+0.319\\), \\(+0.216\\) compared to PED on these three datasets, respectively. Overall, LEAD achieves the best \\(\\tau_w\\) on 7 target datasets and obtains an average \\(\\tau_w\\) of 0.747 which is relatively \\(18\\%\\) better than the latest SOTA ETran [18]. Our experiments highlight the importance of accurate modeling of dynamics and confirm the effectiveness of our method to model the evolution process through the dynamical equation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.834, + 0.406, + 0.851 + ], + "angle": 0, + "content": "4.4. Evaluation on Self-supervised Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Model Zoo. Self-supervised learning models (SSL models) have shown their efficacy in transfer learning and often display superior generalization performance compared" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.389 + ], + "angle": 0, + "content": "to Supervised models. To illustrate the broad applicability of our approach, inspired by [32, 41], we assemble a self-supervised model zoo trained on ImageNet with BYOL [19], Infomin [46], PCLv1 [31], PCLv2 [31], Selav2 [1], InsDis [50], SimCLRv1 [7], SimCLRv2 [8], MoCov1 [24], MoCov2 [9], DeepClusterv2 [5], and SWAV [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.678 + ], + "angle": 0, + "content": "Result Analysis. Due to differing pre-training and downstream classification objectives, SSL models undergo larger dynamic changes during the fine-tuning process, leading to distinct transferability properties compared to supervised models [15]. As shown in Tab. 1 and 2, methods that disregard dynamics (e.g. PARC, Nsleep) perform competitively in the supervised model zoo, but fail on some datasets for SSL models. For example, Nsleep and PARC have a negative small rank correlation on VOC2007 (-0.101) and CIFAR100 (-0.136), respectively. To maintain fairness, we refrain from comparing with GBC [38] in the SSL model zoo, which is designed for supervised scenarios. On the contrary, SFDA and PED exhibit better performances due to consideration of training dynamics. Nevertheless, the limitation of the simulating process still hinders their performance. Compared to them, LEAD achieves a consistent improvement on 7 datasets and obtains an average \\(\\tau_w\\) of 0.746 which is relatively \\(15\\%\\) better than PED (0.648). The results validate the effectiveness of LEAD for SSL models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.684, + 0.787, + 0.7 + ], + "angle": 0, + "content": "4.5. Evaluation on Low-data Regime" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.884 + ], + "angle": 0, + "content": "Experimental Settings. In practical application, privacy and resource constraints may limit our access to the entire downstream dataset. To explore the utilization boundary, we evaluate whether LEAD can provide effective model ranking only using limited samples. We follow three data settings in PACTran [12] with increasing average number of samples per class \\(\\{2,5,10\\}\\) and divide 10 datasets into two groups according to the number of classes: \\(100+\\) classes, including Food, Caltech, Flowers, Cars, CIFAR100, and SUN397; 10-99 classes, including DTD, CIFAR10, Pets, and VOC2007. For each setting, we repeatedly sample 5 times for evaluation and take the average as the result." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Result Analysis. As shown in Tab. 3, all methods, in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28669" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.09, + 0.88, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance comparisons on the self-supervised model zoo. The best results are in bold and the second-best are in underline." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.116, + 0.89, + 0.257 + ], + "angle": 0, + "content": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3590.1960.6220.424-0.1360.4470.1470.496-0.0060.618
Logme [52]ICML'210.5700.0510.6040.375-0.0080.6270.2950.6840.2170.158
Nsleep [33]CVPR'210.5740.5250.5340.4860.2760.641-0.0440.7920.719-0.101
PACTran [12]ECCV'220.7200.6220.6010.4740.5290.6140.4770.6410.6380.620
SFDA [41]ECCV'220.6850.5230.7490.5150.5480.7730.6190.5860.6980.568
ETran [18]ICCV'230.4650.4050.6440.5870.6500.2140.6060.3380.5200.376
PED [32]ICCV'230.5810.6140.7770.6490.5680.9070.6730.4620.6610.583
LEADThis paper0.8600.7800.7250.6630.7760.8250.7130.6290.7600.723
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.265, + 0.893, + 0.322 + ], + "angle": 0, + "content": "Table 3. Performance comparisons with different \\( N / K \\), where \\( K \\) is the number of classes, \\( N \\) is the number of samples for computing the metric. We divide 10 benchmarks into two groups according to the number of classes: 100+ classes and 10-99 classes, and evaluate the average of \\( \\tau_w \\) on the two groups. Additionally, we also provide the average results on the 10 benchmarks, recorded as Avg. The best results are denoted in bold, the second best results are denoted in underline, and - denotes cannot give a valid ranking." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.334, + 0.89, + 0.442 + ], + "angle": 0, + "content": "
Supervised / Self-supervised pre-trained Model
MethodN/K = 2N/K = 5N/K = 10
100+ classes10-99 classesAvg100+ classes10-99 classesAvg100+ classes10-99 classesAvg
GBC0.17 / 0.320.10 / 0.170.13 / 0.250.45 / 0.400.22 / 0.180.36 / 0.310.59 / 0.420.42 / 0.220.52 / 0.34
PACTran0.43 / 0.490.20 / 0.390.33 / 0.450.42 / 0.480.31 / 0.440.38 / 0.470.56 / 0.520.48 / 0.460.53 / 0.50
ETran0.42 / 0.430.12 / 0.410.29 / 0.420.48 / 0.470.31 / 0.380.41 / 0.430.56 / 0.470.56 / 0.480.56 / 0.48
PED- / 0.36- / 0.24- / 0.31- / 0.43- / 0.35- / 0.400.30 / 0.500.22 / 0.560.27 / 0.52
LEAD0.45 / 0.480.51 / 0.450.47 / 0.460.48 / 0.490.64 / 0.530.54 / 0.510.65 / 0.530.64 / 0.600.64 / 0.56
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.76 + ], + "angle": 0, + "content": "cluding PACTran designed for this scenario, exhibit significant performance declines under low-data settings. For instance, ETran's performance in the supervised model zoo drops from 0.64 with the entire dataset to 0.29, 0.41, and 0.56 in three low-data settings. Moreover, some methods, such as SFDA and PED fail to provide prediction rankings in some settings. For example, when \\( N / K = 2,5 \\), PED cannot correctly evaluate class separability due to insufficient data, imposing huge changes to completely separate different classes. This leads to all models perfectly classifying all samples, resulting in identical scores that cannot be ranked. In contrast, our LEAD achieves stable results, ensuring \\( \\tau_w \\geq 0.45 \\) even in extreme settings like \\( N / K = 2 \\), and consistently outperforms other methods. For example, under \\( N / K = 10 \\), its average \\( \\tau_w \\) in the supervised model zoo reaches 0.64, which is relatively \\( 14\\% \\) better than the latest SOTA ETran (0.56) and is equivalent to the latest SOTA on the complete dataset, with less than \\( 5\\% \\) of data size. The results demonstrate the robustness of our method in the face of data limitations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.429, + 0.79 + ], + "angle": 0, + "content": "5. Ablation Study and Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, without loss of generality, we conduct ablation studies on Caltech, CIFAR10, and VOC2007, building upon the evaluation of SSL models in Sec. 4.4 for saving time. We conduct separate experiments to analyze the impact of interpolation coefficients, a key component of LEAD, aiming to delve into the influence of hyperparameters and implementation details, gaining insights into key" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.504 + ], + "angle": 0, + "content": "factors contributing to LEAD's effectiveness. Additionally, we provide a runtime comparison to verify efficiency, and more ablation study results are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.727, + 0.53 + ], + "angle": 0, + "content": "5.1. Interpolation Coefficient" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.581 + ], + "angle": 0, + "content": "The interpolation coefficients are determined by time coefficient \\( t \\) and the NTK matrix. We conduct ablation experiments on these two key factors:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.75 + ], + "angle": 0, + "content": "Time Coefficient \\( t \\). As we model the discrete process as a continuous ODE, the continuous variable \\( t \\) is proportional to the number of updates, determining the progress of the evolution process. In Fig. 4, we evaluate our method under varied time settings, keeping the unit as 1e6 consistent across all experiments. It is evident that performance improves as the time increases within a specific range. However, if the dynamic effect persists for an extended period (e.g., 10.0), all model predictions converge too closely to the ground-truth label, reducing discrimination for model ranking. Hence, we set the time coefficient as 1 by default." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.765, + 0.698, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.765, + 0.874, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.868, + 0.892, + 0.909 + ], + "angle": 0, + "content": "Figure 4. The left picture shows the performance gain in \\(\\tau_w\\) with respect to different time \\(t\\) and the right shows the performance in different sample sizes for each class NTK." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "Sample Size for NTK. To enhance the perception of different dynamic intensities of classes, we propose the class-aware decomposition that employs \\( S \\) samples of each class to obtain NTK matrix and perform spectrum analysis to derive eigenvalues. In Fig. 4, we conduct ablation experiments to explore the effect of varying \\( S \\). We observe that performance generally improves as \\( S \\) increases, and gradually reaches stability. This demonstrates that when the sample size reaches a certain size, eigenvalues reliably capture differences in dynamic intensity, offering a refined analysis result of the dynamics. Therefore, we set \\( S = 64 \\) to ensure the method's efficiency and stability in all experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.277, + 0.327, + 0.294 + ], + "angle": 0, + "content": "5.2. Running Time Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.473, + 0.493 + ], + "angle": 0, + "content": "In experiment results, we have demonstrated the effectiveness of LEAD in various scenarios. In this section, we highlight the computational efficiency of LEAD. With the concise solution in Eq. (7), LEAD predicts the final state only requiring the NTK matrix and \\( \\text{log}_{init} \\). As shown in Tab. 4, the running time of LEAD is about half that of simulation methods (45s compared to 92s), ranking second after GBC which displays unstable performance. Notably, LEAD takes 3s to compute the NTK, while feature extraction takes 420s, resulting in only a marginal \\( 0.7\\% \\) increase in GPU time. The time to compute \\( \\text{log}_{init} \\) is 42s CPU time, indicating significant potential for exploring faster classification algorithms to further reduce the overall time." + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.494, + 0.466, + 0.506 + ], + "angle": 0, + "content": "Table 4. The comparisons of average running time on CIFAR10." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.506, + 0.47, + 0.533 + ], + "angle": 0, + "content": "
MetricsPARCNLEEPPACTranSFDAGBCPEDOurs
Running Time111s1430s77s92s26s97s45s
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.325, + 0.551 + ], + "angle": 0, + "content": "6. Visualization and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.6 + ], + "angle": 0, + "content": "In this section, we validate our approach through some visualizations and these results clearly illustrate how our model effectively addresses the model selection problem." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.24, + 0.621 + ], + "angle": 0, + "content": "6.1. Logit Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.473, + 0.793 + ], + "angle": 0, + "content": "We have conducted a comparison of the predicted logits distribution on VOC2007 using different methods, as illustrated in Fig. 3. Notably, predictions obtained by LEAD closely match the authentic distribution after fine-tuning. To provide further quantitative analysis, we visualize the average bias in predicted logits in Fig. 5. Across three datasets, our method exhibits a substantial reduction compared to simulation methods, achieving nearly a \\(50\\%\\) overall reduction. This visualization highlights the precision of our dynamical equation-based approach in capturing logit evolution for more accurate ranking sequences." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.799, + 0.452, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.866, + 0.47, + 0.907 + ], + "angle": 0, + "content": "Figure 5. Comparison of the average difference between logits predicted by different methods and logits after fine-tuning. The difference is calculated utilizing the Euclidean distance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.657, + 0.106 + ], + "angle": 0, + "content": "6.2. Rank Variation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.109, + 0.895, + 0.291 + ], + "angle": 0, + "content": "By employing our method, we can refine models that have poor initial observation but show better progress during training. To assess the effectiveness of our approach, we conduct a visualization comparing the model rankings between the initial observations and our predicted final state, as shown in Fig. 6. The results clearly illustrate a substantial enhancement in the calibration of model rankings when using our refined observations, with the optimal model now positioned closely to the reference line in the figure. Notably, models starting from an initially unfavorable point can be rapidly elevated to achieve a superior ranking, highlighting the effectiveness of our methodology." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.298, + 0.621, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.297, + 0.753, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.297, + 0.886, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.398, + 0.621, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.398, + 0.753, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.398, + 0.886, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.498, + 0.893, + 0.554 + ], + "angle": 0, + "content": "Figure 6. The variation in model ranking, utilizing initial states (first row) and predicted final states of LEAD (second row). Results span three datasets, from left to right: CIFAR10, Caltech, and VOC2007. Shaded areas reflect the range of the ranking error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.779, + 0.57 + ], + "angle": 0, + "content": "7. Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.895, + 0.77 + ], + "angle": 0, + "content": "In this work, we propose a dynamical equation designed to capture the evolution process within the logit space, offering an efficient solution for assessing model transferability. We construct a theoretical analysis framework and further devise a more practical and precise solution. And we demonstrate its consistent effectiveness across various scenarios including supervised and self-supervised, full-data and low-data settings. In future work, we intend to expand the current analysis framework beyond classification to cover a broader range of downstream tasks. We hope that our work can enhance the understanding of optimization mechanisms in deep neural networks and shed light on other fields, such as the design of optimization algorithms." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.661, + 0.794 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation and in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28671" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.47, + 0.257 + ], + "angle": 0, + "content": "[3] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. In NeurIPS, pages 19301-19312, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.498 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, pages 22243-22255, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.542 + ], + "angle": 0, + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.544, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[10] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[12] Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, and Radu Soricut. Pactran: Pac-bayesian metrics for estimating the transferability of pretrained models to classification tasks. In ECCV, pages 252-268. Springer, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In CVPR, pages 12387–12396, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[14] Kshitij Dwivedi, Jiahui Huang, Radoslaw Martin Cichy, and Gemma Roig. Duality diagram similarity: a generic framework for initialization selection in task transfer learning. In ECCV, pages 497-513, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[15] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In CVPR, pages 5414-5423, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88:303–338, 2010. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[17] Li Fei-Fei. Learning generative visual models from few training examples. In CVPR workshop, 2004. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[18] Mohsen Gholami, Mohammad Akbari, Xinglu Wang, Behnam Kamranian, and Yong Zhang. Etran: Energy-based transferability estimation. In ICCV, pages 18613-18622, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[19] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[20] Boris Hanin and Mihai Nica. Finite depth and width corrections to the neural tangent kernel. In ICLR, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[21] Hangfeng He and Weijie Su. The local elasticity of neural networks. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.308, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.351, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.438, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[25] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.482, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[26] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In NeurIPS, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[28] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In CVPR, pages 2656-2666, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.64 + ], + "angle": 0, + "content": "[29] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.642, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.715, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[32] Xiaotong Li, Zixuan Hu, Yixiao Ge, Ying Shan, and Ling-Yu Duan. Exploring model transferability through the lens of potential energy. In ICCV, pages 5429-5438, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[33] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In CVPR, pages 2663-2673, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Mohamad Amin Mohamadi and Danica J. Sutherland. A fast, well-founded approximation to the empirical neural tangent kernel. In ICML, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28672" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[36] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[37] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In CVPR, pages 9172-9182, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[39] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In \\(CVPR\\), pages 3498-3505, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.356 + ], + "angle": 0, + "content": "[40] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.425 + ], + "angle": 0, + "content": "[41] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In ECCV, pages 286-302, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.454 + ], + "angle": 0, + "content": "[42] Grace S. Shieh. A weighted kendall's tau statistic. Statistics Probability Letters, 39(1):17-24, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[43] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[44] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[45] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[47] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[48] Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[49] Jason Weston, Chris Watkins, et al. Support vector machines for multi-class pattern recognition. In *Esann*, pages 219–224, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[50] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[51] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.146 + ], + "angle": 0, + "content": "[52] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.204 + ], + "angle": 0, + "content": "[53] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In CVPR, pages 19228-19238, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28673" + } + ] +] \ No newline at end of file diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_origin.pdf b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7bcd9483f0cfce8936cf837ec37db30d06a92504 --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/cac4214c-e8ea-4c89-8778-e0818b32709b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8faa0d40c26896f38fd147db802f1acea362657c771daba39477660133443bb3 +size 2967444 diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/full.md b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..843c2b4b33d349b625734e243476ff28e23f6850 --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/full.md @@ -0,0 +1,326 @@ +# LEAD: Exploring Logit Space Evolution for Model Selection + +Zixuan Hu $^{1}$ Xiaotong Li $^{1}$ Shixiang Tang $^{3}$ Jun Liu $^{4}$ Yichun Hu $^{1}$ Ling-Yu Duan $^{1,2*}$ + +$^{1}$ School of Computer Science, Peking University, Beijing, China, $^{2}$ Peng Cheng Laboratory, Shenzhen, China, + +3 The Chinese University of Hong Kong, Hongkong, China, 4 Singapore University of Technology and Design, Singapore {hzxuan, hycc, lingyu}@pku.edu.cn, lixiaotong@stu.pku.edu.cn, shixiangtang@cuhk.edu.hk, jun.liu@sutd.edu.sg + +# Abstract + +The remarkable success of "pretrain-then-finetune" paradigm has led to a proliferation of available pre-trained models for vision tasks. This surge presents a significant challenge in efficiently choosing the most suitable pretrained models for downstream tasks. The critical aspect of this challenge lies in effectively predicting the model transferability by considering the underlying fine-tuning dynamics. Existing methods often model fine-tuning dynamics in feature space with linear transformations, which do not precisely align with the fine-tuning objective and fail to grasp the essential nonlinearity from optimization. To this end, we present LEAD, a finetuning-aligned approach based on the network output of logits. LEAD proposes a theoretical framework to model the optimization process and derives an ordinary differential equation (ODE) to depict the nonlinear evolution toward the final logit state. Additionally, we design a class-aware decomposition method to consider the varying evolution dynamics across classes and further ensure practical applicability. Integrating the closely aligned optimization objective and nonlinear modeling capabilities derived from the differential equation, our method offers a concise solution to effectively bridge the optimization gap in a single step, bypassing the lengthy fine-tuning process. The comprehensive experiments on 24 supervised and self-supervised pre-trained models across 10 downstream datasets demonstrate impressive performances and showcase its broad adaptability even in low-data scenarios. + +# 1. Introduction + +The "pretrain-then-finetune" learning paradigm has exhibited remarkable success in a wide range of computer vision tasks [6, 28], leading to an increasing number of diverse pretrained models open-sourced to the computer vision community. This surge poses a significant challenge in selecting the optimal pre-trained model for a downstream task + +![](images/2c3f6aa22c47c4b6480171e0d2ff66d7c4f64ce96d130c4892baa2e673c8f2a1.jpg) +Figure 1. This diagram illustrates the task of model selection, which aims to rank pre-trained models for choosing the optimal one. The ground-truth ranking sequence is derived from the models' accuracy after fine-tuning. However, due to the time-intensive nature of this process, it requires to design efficient transferability metrics for accurate ranking prediction. The performance of this metric is evaluated using rank correlation, e.g., weighted Kendall's tau [42], to compare the predicted and ground-truth rankings. + +from a vast model zoo. Considering that the brute-force solution of fine-tuning all models on downstream tasks may be infeasible in many scenarios due to the substantial computational costs, the model selection task emerges [13, 47], which ranks the transferability of pre-trained models to the given task without incurring much cost, as shown in Fig. 1. + +Predicting model transferability aims to efficiently estimate the final state after fine-tuning on the downstream task, without network optimization by backward propagation. Prior studies only depend on specific properties of the initial state (e.g., representation similarity [13, 14], conditional distribution [33, 47]), to serve as transferability indicators. Recent studies reveal that modeling the dynamic process of fine-tuning has become the bottleneck to design more high-quality transferability metrics. Following this philosophy, recent works propose to simulate fine-tuning dynamics on features through multiple stages of linear transformations [32, 41], achieving some promising results. + +Despite recent progress in this field, there are still two + +considerable gaps between existing modeled fine-tuning dynamics and the real ones: (1) Insufficient objective alignment. The designed objectives in prior works do not seamlessly align with fine-tuning, as they often focus more on enhancing class separability via feature manipulation, rather than the classification optimization for both the feature and output spaces. (2) Disregarding nonlinear modeling. Existing methods relying solely on linear transformations cannot well capture the intricate nonlinear changes in the feature space during optimization. Such constraints contribute to growing deviations in dynamic modeling and result in unreliable predictions of the final state. + +Delving deeper into the optimization process, we focus on modeling the dynamic changes within the output space (i.e., logits for classification), which not only more closely relates to the fine-tuning objective but also embeds the overall nonlinear dynamics from the network optimization during fine-tuning. Building upon this foundation, we introduce an insightful framework to depict the trajectory of logits toward their final state in a dynamical modeling system. To capture and understand this evolution process of the logit space, the framework provides a discrete iterative optimization equation via gradient analysis and captures nonlinear logit space changes during optimization. + +In contrast to the extensive iterations required in fine-tuning optimization, we further extend the framework to propose a new transferability metric "Logits Evolution Analysis through Dynamical Equation (LEAD)", to seize the final evolutionary state in a single step. We perform a mathematical derivation to convert the discrete process into a continuous differential equation and simplify it into an ordinary differential equation (ODE) by leveraging the classical analysis tool, Neural Tangent Kernel (NTK) [26]. This ODE concisely models the nonlinear evolution in logit space through its closed-form solution, allowing for real-time estimation of logits. Additionally, we introduce a class-aware decomposition method to enhance prediction accuracy and break the barrier of theoretical assumption of NTK. As a result, our method provides a direct estimation of the evolution by incorporating both the initial model state and the impact of fine-tuning evolution, serving as an effective indicator assessing model transferability. + +The contributions of this work are three-fold. (1) To the best of our knowledge, we are the first to explore model transferability by modeling the evolution process in the logit space. (2) We propose LEAD, a novel transferability metric with two key components: a theoretical framework providing a closed-form solution to depict the logits evolution towards the final state, and a class-aware decomposition method to guarantee the practical applicability of theoretical findings and improve predictive precision. (3) We validate the effectiveness and generalizability of our method through experiments on 24 different pre-trained + +models across 10 downstream datasets, achieving SOTA results with an average of $17\%$ gain in the rank correlation. In addition, our method can be naturally extended to low-data scenarios while maintaining competitive performance. + +# 2. Related Work + +# 2.1. Transferability Metric + +The challenge of efficiently ranking pre-trained models has captured much interest of researchers, leading to the development of various metrics to evaluate the model transferability [2, 3, 12, 18, 36, 38, 47, 52]. For example, RSA [13], DDS [14] and GBC [38] are methods based on the distance measurements of the feature space. RSA and DDS assess transferability by evaluating the representation similarity between pre-trained and downstream datasets. GBC employs the Bhattacharyya coefficient to measure class separability in the feature space. Nsleep [33] and Logme [52] estimate transferability by analyzing the conditional distribution, with Nsleep utilizing Gaussian mixture model and Logme employing Bayesian optimization. However, these methods treat network outputs as static and overlook the changes introduced by the training process, leading to less effectiveness in scenes with larger fine-tuning dynamics. + +Recently, SFDA [41] and PED [32] propose to simulate changes in the feature space, respectively projecting the features into a Fisher space and utilizing a physically-inspired model to enhance the class separability, achieving impacts similar to fine-tuning. While these methods achieve some improvements, they solely concentrate on the feature space, disregarding the dynamic changes introduced by the classifier, which is jointly optimized with the backbone. Additionally, they utilize linear transformations to simulate the intricate variations in fine-tuning, resulting in a significant limitation in simulation precision. Instead, our LEAD employs the ordinary differential equation to analyze logit dynamics, maintaining a closely aligned objective and simulating the nonlinear changes in fine-tuning. + +# 2.2. Neural Tangent Kernel + +Studying the behavior of neural network optimization processes has been an enduring challenge due to their nonlinearity and nonconvexity. Neural Tangent Kernel (NTK) [20, 26], serving as a theoretical tool to characterize the optimization process, has attracted much research attention. Some studies have explored its theoretical properties and applications. Notable works include the NTK's property of remaining constant under infinite network width [26], the explanation of the instability of training [48], the analysis of its bias and variance under finite width [20], and its acceleration algorithm [35]. While NTK can analyze network trainability [26, 35], it is constrained by the requirement of infinite width and neglects the consideration of initial log-its, resulting in the inability to predict the changes starting + +from the initial state for a finite-width network. To tackle these issues, we innovatively propose an analysis framework leveraging the property of NTK, suitable for model selection problems to bridge the gap between the initial and final states of finite-width networks. + +# 3. Methodology + +In this section, we first present the problem setup for the task of model selection. Then we analyze training mechanisms to model the fine-tuning optimization in a discrete process. Finally, we propose a transferability metric that converts the discrete process into a continuous dynamical equation with a concise solution to directly seize the evolution destination of logits, and feed updated logits into Cross-Entropy loss to assess transferability, as shown in Fig. 2. + +# 3.1. Problem Setup + +Consider $M$ pre-trained models $\{\Psi_m\}_{m=1}^M$ and a downstream dataset $\mathcal{T} = \{x_n, y_n\}_{n=1}^N$ with $K$ classes, where model $\Psi$ consists of its backbone $f$ which outputs the encoded feature, $N$ is the number of images in the downstream task. The purpose of model selection is to design an efficient algorithm to predict the ranking of the performance of $M$ pre-trained models after fine-tuning. + +Specifically, given the ground-truth performance rankings $\{G_m\}_{m=1}^M$ of pre-trained models after fine-tuning and the predicted ranking $\{P_m\}_{m=1}^M$ by model selection algorithms, we can leverage the commonly-used rank-based correlation, i.e., the weighted Kendalls' $\tau_w$ , to evaluate the effectiveness of the proposed model selection methods, i.e., + +$$ +\tau_ {w} = \frac {1}{\sum_ {i \neq j} w _ {i j}} \cdot \sum_ {i \neq j} w _ {i j} \operatorname {s i g n} \left(G _ {i} - G _ {j}\right) \cdot \operatorname {s i g n} \left(P _ {i} - P _ {j}\right), \tag {1} +$$ + +where $w_{ij} = \frac{1}{i + j}$ and $\mathrm{sign}(\cdot)$ is the signum function. A larger $\tau_w$ indicates more consistencies between the predicted ranking and the ground truth, which means a better model selection algorithm. + +# 3.2. Revisiting the Fine-tuning Optimization + +When a pre-trained model $\Psi$ transfers to a downstream task $\mathcal{T} \coloneqq \{\mathcal{X},\mathcal{Y}\}$ , it undergoes extensive optimization for evolving from its initial state (before fine-tuning) to the final state (after fine-tuning), resulting in complex changes in both feature space and output space. Since the transferability metric is designed to evaluate the final state without fine-tuning, it is essential and challenging to model the optimization process without lengthy computations, bridging the gap between the initial and the final state. To gain insights into the underlying optimization mechanisms, we conduct an analysis to compare the prior arts with fine-tuning to revisit the evolution process toward the final state. + +Recent works SFDA [41] and PED [32] (called simulation methods) aim to approximate fine-tuning procedures + +by simulating feature dynamics. They design an objective function $U$ to assess the class separability of the feature space. Then, they apply linear transformations to increase $U$ and feed modified features into a classifier to update logits, as shown below: + +$$ +A, b = \arg \max _ {A, b} U (A \cdot f (\mathcal {X}) + b, \mathcal {Y}), \tag {2} +$$ + +$$ +\Delta \mathcal {F} (\mathcal {X}) = \mathcal {C} (A \cdot f (\mathcal {X}) + b) - \mathcal {C} (f (\mathcal {X})), +$$ + +where $A, b$ denote linear transformation parameters, $\mathcal{C}$ denotes a machine learning classifier and $\mathcal{F}$ denotes the function that outputs logits. As a result of the optimization in Eq. (2), the objective of enhancing class separability is achieved. However, fine-tuning directly minimizes the gap between predicted logits and labels and performs gradient descent on parameters to update the output logits as follows: + +$$ +\Delta \theta = - \eta \cdot \frac {\mathrm {d} \mathcal {L} (\mathcal {F} (\mathcal {X} , \theta) , \mathcal {Y})}{\mathrm {d} \theta}, +$$ + +$$ +\Delta \mathcal {F} (\mathcal {X}) = \mathcal {F} (\mathcal {X}, \theta + \Delta \theta) - \mathcal {F} (\mathcal {X}, \theta) \approx \Delta \theta \cdot \frac {\mathrm {d} \mathcal {F}}{\mathrm {d} \theta} \tag {3} +$$ + +$$ += - \eta \cdot \frac {\mathrm {d} \mathcal {L}}{\mathrm {d} \theta} \cdot \frac {\mathrm {d} \mathcal {F}}{\mathrm {d} \theta} = - \eta \cdot \frac {\mathrm {d} \mathcal {L}}{\mathrm {d} \mathcal {F}} \left(\frac {\mathrm {d} \mathcal {F}}{\mathrm {d} \theta} \cdot \frac {\mathrm {d} \mathcal {F}}{\mathrm {d} \theta}\right), +$$ + +where $\eta$ denotes the learning rate, $\mathcal{L}$ denotes loss function, and $\theta$ denotes parameters of $\mathcal{F}$ . Eq. (3) provides the variation of logits within one step, capturing the nonlinear dynamics based on gradients. When comparing the optimization outlined in Eq. (2) and Eq. (3), it becomes evident that simulation methods differ from fine-tuning in terms of objectives and optimization approaches. + +Despite preliminary exploration, such simulation methods come with severe limitations due to these inconsistencies, leading to deviations in predicting the final state, as illustrated in Fig. 3. Specifically, simulation methods aim to enhance class separability within the feature space, whereas fine-tuning drives logits closer to labels. The insufficient objective alignment introduces distinct inductive bias compared to fine-tuning, resulting in notable variations in modeling dynamics [27, 34]. Moreover, due to the nonlinear nature of neural networks and complex parameter interactions, linear approximations in simulation methods often fail to accurately capture the dynamic changes [21]. Such limitations hinder the effective modeling of nonlinearity, further constraining modeling precision. + +# 3.3. Modeling Logits Evolution Process through Dynamical Equation + +To address the aforementioned problems, our study delves deeper into the optimization analysis from Eq. (3). In fact, unlike features, the logits derived from the optimization output naturally align with the fine-tuning objectives and reflect the inherent nonlinearity from gradient descent. Therefore, + +![](images/afd73633fe40bc8a1e10ada7da220617e5f53e607aef417db3824deb6a4ce21c.jpg) + +![](images/799fab120c21ac44732bf9677d4919e8a10a9f12d1704ade2fb1d619870bd7b3.jpg) + +![](images/772c7f10ffb50c6f1a01eeb70f507672a56d23c57a1a790ab18c4238c5c4ccae.jpg) +Figure 2. Comparison of different approaches for modeling the evolution process. (a) Fine-tuning iteratively updates parameters via gradient descent, revealing authentic logits evolution trajectory and classification accuracy for ground-truth ranking. (b) Recent works (e.g., SFDA [41], PED [32]) perform linear transformations in feature space to enhance class separability, updating logits with modified features. The ranking score is determined by Cross-Entropy loss on updated logits. While these approaches simulate fine-tuning impact on features, they often suffer huge deviations from the authentic results. (c) Our method, LEAD, transforms the discrete logit evolution during fine-tuning into a gradient-based dynamical equation to capture nonlinear changes and can efficiently predict the final state through the closed-form solution. It establishes an analysis framework linking the initial and final states, enabling the precise prediction in one step. + +it provides a particularly relevant entry point to analyze dynamical changes during fine-tuning and assess the model transferability through its evolution process. + +While the dynamical process of logit space analyzed by Eq. (3) remains consistent with fine-tuning, there still lacks an efficient approach for capturing logits evolution, due to the discrete process of iterating over thousands of steps by traditional optimization. To this end, we propose a new transferability metric called "Logits Evolution Analysis through Dynamical equation (LEAD)", that bridges the evolution process from the initial logit space towards the final state. In contrast to extensive optimization iterations, our method proposes a dynamical analysis framework for analyzing the nonlinear effects within the logit space, providing an efficient prediction to capture the final state with a concise solution in a single step. + +Theoretical Background. In our LEAD, we innovatively introduce the Neural Tangent Kernel (NTK) to simplify the dynamical equation. Here, we provide a brief introduction to the definition of the NTK kernel function and the constant-preserving property that we utilize. + +$$ +\begin{array}{l} \hat {\Phi} _ {\theta_ {t}} (\mathcal {X}, \mathcal {X}) \triangleq \nabla_ {\theta_ {t}} \mathcal {F} (\mathcal {X}; \theta_ {t}) \nabla_ {\theta_ {t}} \mathcal {F} (\mathcal {X}; \theta_ {t}) ^ {\top}, (4) \\ \Phi (\mathcal {X}, \mathcal {X}) \triangleq \lim _ {l \rightarrow \infty} \hat {\Phi} _ {\theta_ {t}} (\mathcal {X}, \mathcal {X}), (4) \\ \end{array} +$$ + +where $\nabla$ denotes the Laplace Operator, and $l$ denotes the root mean square of the widths of all layers in the network. The kernel function $\hat{\Phi}_{\theta_t}$ remains invariant throughout the training process as $l$ approaches infinity with an approximation error of $O\left(\frac{1}{l^2}\right)$ [20, 26]. For simplicity, in the remainder of this paper, we denote functions at time $t$ as: + +![](images/86f95734fc810786cb98edbae5ec9033cac5a37e77337b4e6c0b793079762518.jpg) +Figure 3. Comparison of the average prediction of final state logits across various methods on VOC2007 [16], where the actual result stems from fine-tuning. Closer to actual results indicates more accurate modeling of the final state. Simulation methods (SFDA, PED) capture the correct trend but exhibit a large deviation from the actual results, affecting the accuracy of model ranking. + +$\mathcal{F}_t(\cdot)\triangleq \mathcal{F}(\cdot ,\theta_t),\hat{\Phi}_t(\cdot)\triangleq \hat{\Phi}_{\theta_t}(\mathcal{X},\mathcal{X}),\Phi \triangleq \Phi (\mathcal{X},\mathcal{X}).$ + +Dynamical Equation and Closed-form Solution. To bypass the complexity of iterative computations, we extend the analysis in Eq. (3) and theoretically derive a dynamical equation that is efficient for analysis. Specifically, we seek for a mathematical derivation through the theory of limit approximation to convert the discrete process into a continuous differential equation, enabling us to track the trajectory of logits in the solution space of the equation. Furthermore, we leverage the property of the NTK in Eq. (4) to simplify the equation into an Ordinary Differential Equation (ODE) with an initial value condition. Due to space constraints, we directly provide the equation and the detailed proof can be found in the Appendix. + +$$ +\frac {\mathrm {d} \mathcal {F} _ {t}}{\mathrm {d} t} = - \eta \cdot \Phi \cdot \frac {\mathrm {d} \mathcal {L} _ {t}}{\mathrm {d} \mathcal {F} _ {t}}, \mathcal {F} _ {0} = l o g _ {i n i t}, \tag {5} +$$ + +where $\text{loginit}$ is the initial state of logits obtained through feeding features and labels of downstream datasets to a classification algorithm. Compared to the discrete process, our proposed ODE depicts the continuous evolution of the finetuning dynamics, allowing for a more flexible modeling approach. It transforms discrete gradient effects into equivalent differential effects in continuous time, preserving the capability for modeling nonlinear changes and providing precise prediction of the evolution process. Furthermore, unlike employing numerical methods as the ODE solver, we can directly obtain its closed-form solution through integration over time dimension, allowing us to bypass the lengthy evolution process and reach the optimization destination in a single step, as shown below: + +$$ +\mathbb {E} \left(\mathcal {F} _ {t} (\mathcal {X})\right) = \left(\mathbb {I} - e ^ {- \eta \Phi \cdot t}\right) \mathcal {Y} + e ^ {- \eta \Phi \cdot t} \cdot \log_ {\text {i n i t}}. \tag {6} +$$ + +As depicted in Eq. (6), the solution is derived by interpolating between the one-hot label and the initial observation of logits, with the interpolation coefficients determined by the NTK matrix. It comprises two crucial elements: $\text{log}_{\text{init}}$ , representing the value of logits before optimization and determining the start of the evolution process; and $\Phi$ , indicating the convergence rate to the label and determining the trajectory of logit evolution. Our proposed framework efficiently models the evolution process, enabling a prediction of real-time changes. + +Class-aware Decomposition. While Eq. (6) provides the theoretical result for the final state, computing the value of $\Phi$ under finite width remains challenging. Prior studies [20, 53] indicate that, even with finite width, the convergence rate can be assessed by using the eigenvalues of the initial state NTK matrix $\hat{\Phi}_0$ . Inspired by this empirical finding, we further conduct a decomposition approach to bridge theory and practice, utilizing the eigenvalue decomposition on $\hat{\Phi}_0$ to replace the convergence rate determined by $\Phi$ in Eq. (6) with eigenvalues. + +Due to the network exhibiting varying classification abilities across different classes, their convergence behavior and evolution dynamics also differ during the fine-tuning process. To enhance the modeling of behaviors in different classes, we separately compute NTK for each class, rather than mixing them together, to decouple the respective optimization effect. Specifically, we extract $n$ samples of each class to calculate their respective NTK matrices and perform decomposition to obtain eigenvalues. The average eigenvalue of each class serves as the unified parameter to determine the interpolation coefficient for this class: + +$$ +\mathbb {E} \left(\mathcal {F} _ {t} (x)\right) = \left(\mathbb {I} - e ^ {- \eta \bar {\lambda} \cdot t}\right) y + e ^ {- \eta \bar {\lambda} \cdot t} \cdot \log_ {i n i t}, \tag {7} +$$ + +where $x, y$ is a given sample and its one-hot label, $\overline{\lambda}$ is the average eigenvalue of the NTK matrix associated with the class of $x$ . Building upon the conclusion of Eq. (6), Eq. + +(7) avoids the demand of infinite width, enabling practical computability of the prediction results. As shown in Fig. 2, we employ the results of Eq. (7) to efficiently obtain predictions for the final state logits. Subsequently, following the common practice [32, 33, 41], we feed predictions into the Cross-Entropy loss to obtain the transferability score for model ranking. + +Summary. Our method explicitly models the fine-tuning optimization through the theoretical analysis framework and provides a concise dynamical solution to bypass the lengthy optimization process in a single step, which seamlessly integrates the initial state observations and the evolution trajectory towards the final state. Therefore, we can efficiently predict the final logit state through this dynamical solution. Through comparing logit prediction accuracy among different models, it can serve as an effective metric for predicting model transferability. + +# 4. Experiments + +To demonstrate the effectiveness and robustness of our approach, we evaluate LEAD on diverse pre-trained models, including both supervised and self-supervised convolutional neural networks. Our evaluation spans 10 widely used classification benchmarks in transfer learning, employing weighted Kendall's $\tau_w$ as rank correlation to evaluate the transferability metric. Additionally, we highlight the effectiveness of LEAD in the low-data regime, where only $2/5/10$ samples are available per class. + +# 4.1. Implementation Details + +Given a pre-trained model $\Psi$ , we append a random initialized [22] MLP $h$ as a classification head after its backbone $f$ , and combine them as $\mathcal{F} \coloneqq h \circ f$ to produce logits. To compute the logits in the final state, we only need to derive the NTK matrix of each class and $log_{init}$ through Eq. (7). We follow the approach in [35] to compute the NTK matrix and employ a robust and efficient machine learning classifier, the multi-class SVM [49], to generate $log_{init}$ . Due to space constraints, we provide a more detailed explanation and pseudocode in the Appendix. + +# 4.2. Benchmarks + +Target Datasets. We employ 10 classification benchmarks commonly utilized in transfer learning research including Caltech-101 [17], Stanford Cars [29], CIFAR10 [30], CIFAR100 [30], DTD [10], Oxford 102 Flowers [37], Food-101 [4], Oxford-IIIT Pets [39], SUN397 [51], and VOC2007 [16]. These datasets encompass abundant characteristics, including backgrounds, textures, and coarse/fine-grained scenes, spanning diverse fields. + +Ground Truth Rank. To conduct the ground truth ranking for the model zoo, we follow the approach in [41] and [52]. Specifically, we employ a grid search strategy to determine + +Table 1. Performance comparisons of different transferability metrics on the supervised model zoo. We measure the performance with weighted Kendall's $\tau_w$ and a larger $\tau_w$ represents a better prediction rank. The best results are in bold and the second-best are in underline. + +
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3630.4220.0660.135-0.0920.5360.3100.114-0.0970.707
Logme [52]ICML'210.3360.3490.4250.5380.6030.6510.7830.3720.4400.673
Nsleep [33]CVPR'210.5040.6780.2920.4350.6860.4520.6560.6860.7700.645
PACTran [12]ECCV'220.7090.3720.4800.1510.7420.3720.7570.4510.4220.536
SFDA [41]ECCV'220.5340.5970.4210.4720.7490.4210.8170.6250.5620.728
GBC [38]CVPR'220.8440.4300.3480.7350.8300.4800.8660.3580.5740.654
ETran [18]ICCV'230.7090.6270.5340.5570.7490.6050.7190.6510.5770.673
PED [32]ICCV'230.5730.5970.4510.4720.7490.4210.7700.6250.5620.728
LEADThis paper0.8920.6980.7860.5860.8350.6890.7910.8410.6090.743
+ +the actual performance of each model in the downstream task. This strategy involves trying various learning rates from the set $\{10^{-1}, 10^{-2}, 10^{-3}, 10^{-4}\}$ and different weight decay values from the set $\{0, 10^{-6}, 10^{-5}, 10^{-4}, 10^{-3}\}$ . To ensure reliability, we repeat each experiment 5 times with different random seeds and take the average as the result. + +# 4.3. Evaluation on Supervised Models + +Model Zoo. We construct a Model Zoo that includes 12 CNN models supervised pre-trained on ImageNet [11], covering widely-used architectures: ResNet-34 [23], ResNet-50 [23], ResNet-101 [23], ResNet-152 [23], DenseNet-121 [25], DenseNet-161 [25], DenseNet-169 [25], DenseNet-201 [25], MNet-A1 [45], MobileNetV2 [40], GoogleNet [43], and InceptionV3 [44]. The accuracy results of these models after fine-tuning can be found in the Appendix. + +Result Analysis. To show the effectiveness of LEAD, we compare it with prior arts, such as Bayesian-based Logme, Separability-based GBC, and simulation methods SFDA and PED. As shown in Tab. 1, simulation methods consider underlying dynamic effects and exhibit promising performances on some datasets. However, as the aforementioned analysis in Sec. 3.2 shows, the bias in the simulation process causes these methods to have unstable improvements, particularly resulting in unsatisfactory performance on Flowers, Food, and Pets. Benefiting from a precise analysis of the evolution, our LEAD yields a performance gain of $+0.335$ , $+0.319$ , $+0.216$ compared to PED on these three datasets, respectively. Overall, LEAD achieves the best $\tau_w$ on 7 target datasets and obtains an average $\tau_w$ of 0.747 which is relatively $18\%$ better than the latest SOTA ETran [18]. Our experiments highlight the importance of accurate modeling of dynamics and confirm the effectiveness of our method to model the evolution process through the dynamical equation. + +# 4.4. Evaluation on Self-supervised Models + +Model Zoo. Self-supervised learning models (SSL models) have shown their efficacy in transfer learning and often display superior generalization performance compared + +to Supervised models. To illustrate the broad applicability of our approach, inspired by [32, 41], we assemble a self-supervised model zoo trained on ImageNet with BYOL [19], Infomin [46], PCLv1 [31], PCLv2 [31], Selav2 [1], InsDis [50], SimCLRv1 [7], SimCLRv2 [8], MoCov1 [24], MoCov2 [9], DeepClusterv2 [5], and SWAV [6]. + +Result Analysis. Due to differing pre-training and downstream classification objectives, SSL models undergo larger dynamic changes during the fine-tuning process, leading to distinct transferability properties compared to supervised models [15]. As shown in Tab. 1 and 2, methods that disregard dynamics (e.g. PARC, Nsleep) perform competitively in the supervised model zoo, but fail on some datasets for SSL models. For example, Nsleep and PARC have a negative small rank correlation on VOC2007 (-0.101) and CIFAR100 (-0.136), respectively. To maintain fairness, we refrain from comparing with GBC [38] in the SSL model zoo, which is designed for supervised scenarios. On the contrary, SFDA and PED exhibit better performances due to consideration of training dynamics. Nevertheless, the limitation of the simulating process still hinders their performance. Compared to them, LEAD achieves a consistent improvement on 7 datasets and obtains an average $\tau_w$ of 0.746 which is relatively $15\%$ better than PED (0.648). The results validate the effectiveness of LEAD for SSL models. + +# 4.5. Evaluation on Low-data Regime + +Experimental Settings. In practical application, privacy and resource constraints may limit our access to the entire downstream dataset. To explore the utilization boundary, we evaluate whether LEAD can provide effective model ranking only using limited samples. We follow three data settings in PACTran [12] with increasing average number of samples per class $\{2,5,10\}$ and divide 10 datasets into two groups according to the number of classes: $100+$ classes, including Food, Caltech, Flowers, Cars, CIFAR100, and SUN397; 10-99 classes, including DTD, CIFAR10, Pets, and VOC2007. For each setting, we repeatedly sample 5 times for evaluation and take the average as the result. + +Result Analysis. As shown in Tab. 3, all methods, in- + +Table 2. Performance comparisons on the self-supervised model zoo. The best results are in bold and the second-best are in underline. + +
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3590.1960.6220.424-0.1360.4470.1470.496-0.0060.618
Logme [52]ICML'210.5700.0510.6040.375-0.0080.6270.2950.6840.2170.158
Nsleep [33]CVPR'210.5740.5250.5340.4860.2760.641-0.0440.7920.719-0.101
PACTran [12]ECCV'220.7200.6220.6010.4740.5290.6140.4770.6410.6380.620
SFDA [41]ECCV'220.6850.5230.7490.5150.5480.7730.6190.5860.6980.568
ETran [18]ICCV'230.4650.4050.6440.5870.6500.2140.6060.3380.5200.376
PED [32]ICCV'230.5810.6140.7770.6490.5680.9070.6730.4620.6610.583
LEADThis paper0.8600.7800.7250.6630.7760.8250.7130.6290.7600.723
+ +Table 3. Performance comparisons with different $N / K$ , where $K$ is the number of classes, $N$ is the number of samples for computing the metric. We divide 10 benchmarks into two groups according to the number of classes: 100+ classes and 10-99 classes, and evaluate the average of $\tau_w$ on the two groups. Additionally, we also provide the average results on the 10 benchmarks, recorded as Avg. The best results are denoted in bold, the second best results are denoted in underline, and - denotes cannot give a valid ranking. + +
Supervised / Self-supervised pre-trained Model
MethodN/K = 2N/K = 5N/K = 10
100+ classes10-99 classesAvg100+ classes10-99 classesAvg100+ classes10-99 classesAvg
GBC0.17 / 0.320.10 / 0.170.13 / 0.250.45 / 0.400.22 / 0.180.36 / 0.310.59 / 0.420.42 / 0.220.52 / 0.34
PACTran0.43 / 0.490.20 / 0.390.33 / 0.450.42 / 0.480.31 / 0.440.38 / 0.470.56 / 0.520.48 / 0.460.53 / 0.50
ETran0.42 / 0.430.12 / 0.410.29 / 0.420.48 / 0.470.31 / 0.380.41 / 0.430.56 / 0.470.56 / 0.480.56 / 0.48
PED- / 0.36- / 0.24- / 0.31- / 0.43- / 0.35- / 0.400.30 / 0.500.22 / 0.560.27 / 0.52
LEAD0.45 / 0.480.51 / 0.450.47 / 0.460.48 / 0.490.64 / 0.530.54 / 0.510.65 / 0.530.64 / 0.600.64 / 0.56
+ +cluding PACTran designed for this scenario, exhibit significant performance declines under low-data settings. For instance, ETran's performance in the supervised model zoo drops from 0.64 with the entire dataset to 0.29, 0.41, and 0.56 in three low-data settings. Moreover, some methods, such as SFDA and PED fail to provide prediction rankings in some settings. For example, when $N / K = 2,5$ , PED cannot correctly evaluate class separability due to insufficient data, imposing huge changes to completely separate different classes. This leads to all models perfectly classifying all samples, resulting in identical scores that cannot be ranked. In contrast, our LEAD achieves stable results, ensuring $\tau_w \geq 0.45$ even in extreme settings like $N / K = 2$ , and consistently outperforms other methods. For example, under $N / K = 10$ , its average $\tau_w$ in the supervised model zoo reaches 0.64, which is relatively $14\%$ better than the latest SOTA ETran (0.56) and is equivalent to the latest SOTA on the complete dataset, with less than $5\%$ of data size. The results demonstrate the robustness of our method in the face of data limitations. + +# 5. Ablation Study and Efficiency Analysis + +In this section, without loss of generality, we conduct ablation studies on Caltech, CIFAR10, and VOC2007, building upon the evaluation of SSL models in Sec. 4.4 for saving time. We conduct separate experiments to analyze the impact of interpolation coefficients, a key component of LEAD, aiming to delve into the influence of hyperparameters and implementation details, gaining insights into key + +factors contributing to LEAD's effectiveness. Additionally, we provide a runtime comparison to verify efficiency, and more ablation study results are provided in the appendix. + +# 5.1. Interpolation Coefficient + +The interpolation coefficients are determined by time coefficient $t$ and the NTK matrix. We conduct ablation experiments on these two key factors: + +Time Coefficient $t$ . As we model the discrete process as a continuous ODE, the continuous variable $t$ is proportional to the number of updates, determining the progress of the evolution process. In Fig. 4, we evaluate our method under varied time settings, keeping the unit as 1e6 consistent across all experiments. It is evident that performance improves as the time increases within a specific range. However, if the dynamic effect persists for an extended period (e.g., 10.0), all model predictions converge too closely to the ground-truth label, reducing discrimination for model ranking. Hence, we set the time coefficient as 1 by default. + +![](images/d8befea459eb8901a1a723d8409b407c7f5f958b1a950897281e0ed8c9dfe50a.jpg) + +![](images/1b3c6d2bb15e3a12791b480ffe8032eff56a800c64b8ddf7c8ef96a1d83c0eba.jpg) +Figure 4. The left picture shows the performance gain in $\tau_w$ with respect to different time $t$ and the right shows the performance in different sample sizes for each class NTK. + +Sample Size for NTK. To enhance the perception of different dynamic intensities of classes, we propose the class-aware decomposition that employs $S$ samples of each class to obtain NTK matrix and perform spectrum analysis to derive eigenvalues. In Fig. 4, we conduct ablation experiments to explore the effect of varying $S$ . We observe that performance generally improves as $S$ increases, and gradually reaches stability. This demonstrates that when the sample size reaches a certain size, eigenvalues reliably capture differences in dynamic intensity, offering a refined analysis result of the dynamics. Therefore, we set $S = 64$ to ensure the method's efficiency and stability in all experiments. + +# 5.2. Running Time Comparison + +In experiment results, we have demonstrated the effectiveness of LEAD in various scenarios. In this section, we highlight the computational efficiency of LEAD. With the concise solution in Eq. (7), LEAD predicts the final state only requiring the NTK matrix and $\text{log}_{init}$ . As shown in Tab. 4, the running time of LEAD is about half that of simulation methods (45s compared to 92s), ranking second after GBC which displays unstable performance. Notably, LEAD takes 3s to compute the NTK, while feature extraction takes 420s, resulting in only a marginal $0.7\%$ increase in GPU time. The time to compute $\text{log}_{init}$ is 42s CPU time, indicating significant potential for exploring faster classification algorithms to further reduce the overall time. + +Table 4. The comparisons of average running time on CIFAR10. + +
MetricsPARCNLEEPPACTranSFDAGBCPEDOurs
Running Time111s1430s77s92s26s97s45s
+ +# 6. Visualization and Analysis + +In this section, we validate our approach through some visualizations and these results clearly illustrate how our model effectively addresses the model selection problem. + +# 6.1. Logit Prediction + +We have conducted a comparison of the predicted logits distribution on VOC2007 using different methods, as illustrated in Fig. 3. Notably, predictions obtained by LEAD closely match the authentic distribution after fine-tuning. To provide further quantitative analysis, we visualize the average bias in predicted logits in Fig. 5. Across three datasets, our method exhibits a substantial reduction compared to simulation methods, achieving nearly a $50\%$ overall reduction. This visualization highlights the precision of our dynamical equation-based approach in capturing logit evolution for more accurate ranking sequences. + +![](images/5730efca410870bc74277eafa6b7679899fbff59307e049120384f6d9a9826e7.jpg) +Figure 5. Comparison of the average difference between logits predicted by different methods and logits after fine-tuning. The difference is calculated utilizing the Euclidean distance. + +# 6.2. Rank Variation + +By employing our method, we can refine models that have poor initial observation but show better progress during training. To assess the effectiveness of our approach, we conduct a visualization comparing the model rankings between the initial observations and our predicted final state, as shown in Fig. 6. The results clearly illustrate a substantial enhancement in the calibration of model rankings when using our refined observations, with the optimal model now positioned closely to the reference line in the figure. Notably, models starting from an initially unfavorable point can be rapidly elevated to achieve a superior ranking, highlighting the effectiveness of our methodology. + +![](images/2f3583d5c16fb243a02f5b430a4256595b02cb4df51f2e295eacd8a799197982.jpg) + +![](images/c6876fb90cafd3d8cdb83f6640d4e998b199078d9922496892cbbb318c52a4ff.jpg) + +![](images/47164e21d643dbb8ece11d0d1697e86c74aac8c23ea55f72173eb84e597ae7c1.jpg) + +![](images/075db0296cb59584eb3a8b14d5d198af5096a3c0919fb40315dd6521e9dd9594.jpg) +Figure 6. The variation in model ranking, utilizing initial states (first row) and predicted final states of LEAD (second row). Results span three datasets, from left to right: CIFAR10, Caltech, and VOC2007. Shaded areas reflect the range of the ranking error. + +![](images/1b4b2ab39d6d5f3823aa39cfe2de811aba1c89e26dfe2f179ef288de09157e83.jpg) + +![](images/126bb530f74b06ea081490901547b8ad5456047169c6db2349993f4dba8e19d6.jpg) + +# 7. Conclusions and Future Work + +In this work, we propose a dynamical equation designed to capture the evolution process within the logit space, offering an efficient solution for assessing model transferability. We construct a theoretical analysis framework and further devise a more practical and precise solution. And we demonstrate its consistent effectiveness across various scenarios including supervised and self-supervised, full-data and low-data settings. In future work, we intend to expand the current analysis framework beyond classification to cover a broader range of downstream tasks. We hope that our work can enhance the understanding of optimization mechanisms in deep neural networks and shed light on other fields, such as the design of optimization algorithms. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China under Grant 62088102, in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation and in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd. + +# References + +[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv, 2019. 6 +[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313, 2019. 2 +[3] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. In NeurIPS, pages 19301-19312, 2021. 2, 6, 7 +[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5 +[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 6 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020. 1, 6 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6 +[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, pages 22243-22255, 2020. 6 +[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv, 2020. 6 +[10] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6 +[12] Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, and Radu Soricut. Pactran: Pac-bayesian metrics for estimating the transferability of pretrained models to classification tasks. In ECCV, pages 252-268. Springer, 2022. 2, 6, 7 +[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In CVPR, pages 12387–12396, 2019. 1, 2 +[14] Kshitij Dwivedi, Jiahui Huang, Radoslaw Martin Cichy, and Gemma Roig. Duality diagram similarity: a generic framework for initialization selection in task transfer learning. In ECCV, pages 497-513, 2020. 1, 2 +[15] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In CVPR, pages 5414-5423, 2021. 6 +[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88:303–338, 2010. 4, 5 +[17] Li Fei-Fei. Learning generative visual models from few training examples. In CVPR workshop, 2004. 5 + +[18] Mohsen Gholami, Mohammad Akbari, Xinglu Wang, Behnam Kamranian, and Yong Zhang. Etran: Energy-based transferability estimation. In ICCV, pages 18613-18622, 2023. 2, 6, 7 +[19] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 6 +[20] Boris Hanin and Mihai Nica. Finite depth and width corrections to the neural tangent kernel. In ICLR, 2020. 2, 4, 5 +[21] Hangfeng He and Weijie Su. The local elasticity of neural networks. In ICLR, 2020. 3 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 6 +[25] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 6 +[26] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In NeurIPS, 2018. 2, 4 +[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 3 +[28] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In CVPR, pages 2656-2666, 2018. 1 +[29] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5 +[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv, 2020. 6 +[32] Xiaotong Li, Zixuan Hu, Yixiao Ge, Ying Shan, and Ling-Yu Duan. Exploring model transferability through the lens of potential energy. In ICCV, pages 5429-5438, 2023. 1, 2, 3, 4, 5, 6, 7 +[33] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In CVPR, pages 2663-2673, 2021. 1, 2, 5, 6, 7 +[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 3 +[35] Mohamad Amin Mohamadi and Danica J. Sutherland. A fast, well-founded approximation to the empirical neural tangent kernel. In ICML, 2022. 2, 5 + +[36] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305, 2020. 2 +[37] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5 +[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In CVPR, pages 9172-9182, 2022. 2, 6 +[39] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505, 2012. 5 +[40] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6 +[41] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In ECCV, pages 286-302, 2022. 1, 2, 3, 4, 5, 6, 7 +[42] Grace S. Shieh. A weighted kendall's tau statistic. Statistics Probability Letters, 39(1):17-24, 1998. 1 +[43] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 6 +[44] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 6 +[45] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 6 +[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 6 +[47] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 1, 2 +[48] Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. 2 +[49] Jason Weston, Chris Watkins, et al. Support vector machines for multi-class pattern recognition. In *Esann*, pages 219–224, 1999. 5 +[50] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 6 +[51] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492. IEEE, 2010. 5 + +[52] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143, 2021. 2, 5, 6, 7 +[53] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In CVPR, pages 19228-19238, 2022. 5 \ No newline at end of file diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/images.zip b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bd26219b88bb89de585f0e31b3775407592054a4 --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:862279abb41753174da286138eb881e8c62a906b78a173032e50b9db1688c518 +size 550351 diff --git a/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/layout.json b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..efacaf57ffd4d1ecb510b34aff4645b1ec1b5fab --- /dev/null +++ b/2024/LEAD_ Exploring Logit Space Evolution for Model Selection/layout.json @@ -0,0 +1,8744 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 484, + 121 + ], + "type": "text", + "content": "LEAD: Exploring Logit Space Evolution for Model Selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": "Zixuan Hu" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": " Xiaotong Li" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": " Shixiang Tang" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": " Jun Liu" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": " Yichun Hu" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "text", + "content": " Ling-Yu Duan" + }, + { + "bbox": [ + 73, + 142, + 527, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "type": "text", + "content": " School of Computer Science, Peking University, Beijing, China, " + }, + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 159, + 522, + 171 + ], + "type": "text", + "content": " Peng Cheng Laboratory, Shenzhen, China," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 171, + 543, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 171, + 543, + 213 + ], + "spans": [ + { + "bbox": [ + 54, + 171, + 543, + 213 + ], + "type": "text", + "content": "3 The Chinese University of Hong Kong, Hongkong, China, 4 Singapore University of Technology and Design, Singapore {hzxuan, hycc, lingyu}@pku.edu.cn, lixiaotong@stu.pku.edu.cn, shixiangtang@cuhk.edu.hk, jun.liu@sutd.edu.sg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 266, + 290, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 266, + 290, + 590 + ], + "spans": [ + { + "bbox": [ + 45, + 266, + 290, + 590 + ], + "type": "text", + "content": "The remarkable success of \"pretrain-then-finetune\" paradigm has led to a proliferation of available pre-trained models for vision tasks. This surge presents a significant challenge in efficiently choosing the most suitable pretrained models for downstream tasks. The critical aspect of this challenge lies in effectively predicting the model transferability by considering the underlying fine-tuning dynamics. Existing methods often model fine-tuning dynamics in feature space with linear transformations, which do not precisely align with the fine-tuning objective and fail to grasp the essential nonlinearity from optimization. To this end, we present LEAD, a finetuning-aligned approach based on the network output of logits. LEAD proposes a theoretical framework to model the optimization process and derives an ordinary differential equation (ODE) to depict the nonlinear evolution toward the final logit state. Additionally, we design a class-aware decomposition method to consider the varying evolution dynamics across classes and further ensure practical applicability. Integrating the closely aligned optimization objective and nonlinear modeling capabilities derived from the differential equation, our method offers a concise solution to effectively bridge the optimization gap in a single step, bypassing the lengthy fine-tuning process. The comprehensive experiments on 24 supervised and self-supervised pre-trained models across 10 downstream datasets demonstrate impressive performances and showcase its broad adaptability even in low-data scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 128, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 287, + 694 + ], + "type": "text", + "content": "The \"pretrain-then-finetune\" learning paradigm has exhibited remarkable success in a wide range of computer vision tasks [6, 28], leading to an increasing number of diverse pretrained models open-sourced to the computer vision community. This surge poses a significant challenge in selecting the optimal pre-trained model for a downstream task" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 239, + 547, + 355 + ], + "blocks": [ + { + "bbox": [ + 309, + 239, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 309, + 239, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 309, + 239, + 547, + 355 + ], + "type": "image", + "image_path": "2c3f6aa22c47c4b6480171e0d2ff66d7c4f64ce96d130c4892baa2e673c8f2a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 369, + 547, + 459 + ], + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 459 + ], + "type": "text", + "content": "Figure 1. This diagram illustrates the task of model selection, which aims to rank pre-trained models for choosing the optimal one. The ground-truth ranking sequence is derived from the models' accuracy after fine-tuning. However, due to the time-intensive nature of this process, it requires to design efficient transferability metrics for accurate ranking prediction. The performance of this metric is evaluated using rank correlation, e.g., weighted Kendall's tau [42], to compare the predicted and ground-truth rankings." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 555 + ], + "type": "text", + "content": "from a vast model zoo. Considering that the brute-force solution of fine-tuning all models on downstream tasks may be infeasible in many scenarios due to the substantial computational costs, the model selection task emerges [13, 47], which ranks the transferability of pre-trained models to the given task without incurring much cost, as shown in Fig. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": "Predicting model transferability aims to efficiently estimate the final state after fine-tuning on the downstream task, without network optimization by backward propagation. Prior studies only depend on specific properties of the initial state (e.g., representation similarity [13, 14], conditional distribution [33, 47]), to serve as transferability indicators. Recent studies reveal that modeling the dynamic process of fine-tuning has become the bottleneck to design more high-quality transferability metrics. Following this philosophy, recent works propose to simulate fine-tuning dynamics on features through multiple stages of linear transformations [32, 41], achieving some promising results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Despite recent progress in this field, there are still two" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 138, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28664" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "considerable gaps between existing modeled fine-tuning dynamics and the real ones: (1) Insufficient objective alignment. The designed objectives in prior works do not seamlessly align with fine-tuning, as they often focus more on enhancing class separability via feature manipulation, rather than the classification optimization for both the feature and output spaces. (2) Disregarding nonlinear modeling. Existing methods relying solely on linear transformations cannot well capture the intricate nonlinear changes in the feature space during optimization. Such constraints contribute to growing deviations in dynamic modeling and result in unreliable predictions of the final state." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 218, + 289, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 362 + ], + "type": "text", + "content": "Delving deeper into the optimization process, we focus on modeling the dynamic changes within the output space (i.e., logits for classification), which not only more closely relates to the fine-tuning objective but also embeds the overall nonlinear dynamics from the network optimization during fine-tuning. Building upon this foundation, we introduce an insightful framework to depict the trajectory of logits toward their final state in a dynamical modeling system. To capture and understand this evolution process of the logit space, the framework provides a discrete iterative optimization equation via gradient analysis and captures nonlinear logit space changes during optimization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 579 + ], + "type": "text", + "content": "In contrast to the extensive iterations required in fine-tuning optimization, we further extend the framework to propose a new transferability metric \"Logits Evolution Analysis through Dynamical Equation (LEAD)\", to seize the final evolutionary state in a single step. We perform a mathematical derivation to convert the discrete process into a continuous differential equation and simplify it into an ordinary differential equation (ODE) by leveraging the classical analysis tool, Neural Tangent Kernel (NTK) [26]. This ODE concisely models the nonlinear evolution in logit space through its closed-form solution, allowing for real-time estimation of logits. Additionally, we introduce a class-aware decomposition method to enhance prediction accuracy and break the barrier of theoretical assumption of NTK. As a result, our method provides a direct estimation of the evolution by incorporating both the initial model state and the impact of fine-tuning evolution, serving as an effective indicator assessing model transferability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "The contributions of this work are three-fold. (1) To the best of our knowledge, we are the first to explore model transferability by modeling the evolution process in the logit space. (2) We propose LEAD, a novel transferability metric with two key components: a theoretical framework providing a closed-form solution to depict the logits evolution towards the final state, and a class-aware decomposition method to guarantee the practical applicability of theoretical findings and improve predictive precision. (3) We validate the effectiveness and generalizability of our method through experiments on 24 different pre-trained" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": "models across 10 downstream datasets, achieving SOTA results with an average of " + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 121 + ], + "type": "text", + "content": " gain in the rank correlation. In addition, our method can be naturally extended to low-data scenarios while maintaining competitive performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 130, + 393, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 130, + 393, + 142 + ], + "spans": [ + { + "bbox": [ + 305, + 130, + 393, + 142 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 147, + 434, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 147, + 434, + 160 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 434, + 160 + ], + "type": "text", + "content": "2.1. Transferability Metric" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 163, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 355 + ], + "type": "text", + "content": "The challenge of efficiently ranking pre-trained models has captured much interest of researchers, leading to the development of various metrics to evaluate the model transferability [2, 3, 12, 18, 36, 38, 47, 52]. For example, RSA [13], DDS [14] and GBC [38] are methods based on the distance measurements of the feature space. RSA and DDS assess transferability by evaluating the representation similarity between pre-trained and downstream datasets. GBC employs the Bhattacharyya coefficient to measure class separability in the feature space. Nsleep [33] and Logme [52] estimate transferability by analyzing the conditional distribution, with Nsleep utilizing Gaussian mixture model and Logme employing Bayesian optimization. However, these methods treat network outputs as static and overlook the changes introduced by the training process, leading to less effectiveness in scenes with larger fine-tuning dynamics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 356, + 548, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 548, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 548, + 523 + ], + "type": "text", + "content": "Recently, SFDA [41] and PED [32] propose to simulate changes in the feature space, respectively projecting the features into a Fisher space and utilizing a physically-inspired model to enhance the class separability, achieving impacts similar to fine-tuning. While these methods achieve some improvements, they solely concentrate on the feature space, disregarding the dynamic changes introduced by the classifier, which is jointly optimized with the backbone. Additionally, they utilize linear transformations to simulate the intricate variations in fine-tuning, resulting in a significant limitation in simulation precision. Instead, our LEAD employs the ordinary differential equation to analyze logit dynamics, maintaining a closely aligned objective and simulating the nonlinear changes in fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 529, + 437, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 529, + 437, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 529, + 437, + 542 + ], + "type": "text", + "content": "2.2. Neural Tangent Kernel" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 715 + ], + "type": "text", + "content": "Studying the behavior of neural network optimization processes has been an enduring challenge due to their nonlinearity and nonconvexity. Neural Tangent Kernel (NTK) [20, 26], serving as a theoretical tool to characterize the optimization process, has attracted much research attention. Some studies have explored its theoretical properties and applications. Notable works include the NTK's property of remaining constant under infinite network width [26], the explanation of the instability of training [48], the analysis of its bias and variance under finite width [20], and its acceleration algorithm [35]. While NTK can analyze network trainability [26, 35], it is constrained by the requirement of infinite width and neglects the consideration of initial log-its, resulting in the inability to predict the changes starting" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28665" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "from the initial state for a finite-width network. To tackle these issues, we innovatively propose an analysis framework leveraging the property of NTK, suitable for model selection problems to bridge the gap between the initial and final states of finite-width networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 143, + 129, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 143, + 129, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 129, + 156 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 163, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 287, + 258 + ], + "type": "text", + "content": "In this section, we first present the problem setup for the task of model selection. Then we analyze training mechanisms to model the fine-tuning optimization in a discrete process. Finally, we propose a transferability metric that converts the discrete process into a continuous dynamical equation with a concise solution to directly seize the evolution destination of logits, and feed updated logits into Cross-Entropy loss to assess transferability, as shown in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 264, + 139, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 139, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 139, + 277 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": "Consider " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " pre-trained models " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\{\\Psi_m\\}_{m=1}^M" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " and a downstream dataset " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{x_n, y_n\\}_{n=1}^N" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " classes, where model " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " consists of its backbone " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " which outputs the encoded feature, " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " is the number of images in the downstream task. The purpose of model selection is to design an efficient algorithm to predict the ranking of the performance of " + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 280, + 287, + 364 + ], + "type": "text", + "content": " pre-trained models after fine-tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "text", + "content": "Specifically, given the ground-truth performance rankings " + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\{G_m\\}_{m=1}^M" + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "text", + "content": " of pre-trained models after fine-tuning and the predicted ranking " + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\{P_m\\}_{m=1}^M" + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "text", + "content": " by model selection algorithms, we can leverage the commonly-used rank-based correlation, i.e., the weighted Kendalls' " + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 365, + 287, + 436 + ], + "type": "text", + "content": ", to evaluate the effectiveness of the proposed model selection methods, i.e.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 441, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 441, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 52, + 441, + 287, + 469 + ], + "type": "interline_equation", + "content": "\\tau_ {w} = \\frac {1}{\\sum_ {i \\neq j} w _ {i j}} \\cdot \\sum_ {i \\neq j} w _ {i j} \\operatorname {s i g n} \\left(G _ {i} - G _ {j}\\right) \\cdot \\operatorname {s i g n} \\left(P _ {i} - P _ {j}\\right), \\tag {1}", + "image_path": "3a67abcfcee9e17130d2fccb1a7e29eb3a69226010c63404ba10c95b5ff87e35.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "inline_equation", + "content": "w_{ij} = \\frac{1}{i + j}" + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{sign}(\\cdot)" + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": " is the signum function. A larger " + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 475, + 287, + 523 + ], + "type": "text", + "content": " indicates more consistencies between the predicted ranking and the ground truth, which means a better model selection algorithm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 529, + 256, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 256, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 256, + 542 + ], + "type": "text", + "content": "3.2. Revisiting the Fine-tuning Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "When a pre-trained model " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": " transfers to a downstream task " + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\coloneqq \\{\\mathcal{X},\\mathcal{Y}\\}" + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": ", it undergoes extensive optimization for evolving from its initial state (before fine-tuning) to the final state (after fine-tuning), resulting in complex changes in both feature space and output space. Since the transferability metric is designed to evaluate the final state without fine-tuning, it is essential and challenging to model the optimization process without lengthy computations, bridging the gap between the initial and the final state. To gain insights into the underlying optimization mechanisms, we conduct an analysis to compare the prior arts with fine-tuning to revisit the evolution process toward the final state." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Recent works SFDA [41] and PED [32] (called simulation methods) aim to approximate fine-tuning procedures" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "by simulating feature dynamics. They design an objective function " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " to assess the class separability of the feature space. Then, they apply linear transformations to increase " + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": " and feed modified features into a classifier to update logits, as shown below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 137, + 545, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 137, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 348, + 137, + 545, + 159 + ], + "type": "interline_equation", + "content": "A, b = \\arg \\max _ {A, b} U (A \\cdot f (\\mathcal {X}) + b, \\mathcal {Y}), \\tag {2}", + "image_path": "0a67beca933901fc380984101817d4dfcf35bc3c8ffbcec84749f0b085b720d1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 160, + 509, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 160, + 509, + 172 + ], + "spans": [ + { + "bbox": [ + 343, + 160, + 509, + 172 + ], + "type": "interline_equation", + "content": "\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {C} (A \\cdot f (\\mathcal {X}) + b) - \\mathcal {C} (f (\\mathcal {X})),", + "image_path": "9ef6eb64b590da806e25eda8561654c192224f01964873436b081139707bc293.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "inline_equation", + "content": "A, b" + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "text", + "content": " denote linear transformation parameters, " + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "text", + "content": " denotes a machine learning classifier and " + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 180, + 545, + 264 + ], + "type": "text", + "content": " denotes the function that outputs logits. As a result of the optimization in Eq. (2), the objective of enhancing class separability is achieved. However, fine-tuning directly minimizes the gap between predicted logits and labels and performs gradient descent on parameters to update the output logits as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 280, + 480, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 280, + 480, + 304 + ], + "spans": [ + { + "bbox": [ + 359, + 280, + 480, + 304 + ], + "type": "interline_equation", + "content": "\\Delta \\theta = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L} (\\mathcal {F} (\\mathcal {X} , \\theta) , \\mathcal {Y})}{\\mathrm {d} \\theta},", + "image_path": "a6583925561113279f2ffaaad4ac1689607185ae15786e6fa17918fee562834a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 305, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 305, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 318, + 305, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\Delta \\mathcal {F} (\\mathcal {X}) = \\mathcal {F} (\\mathcal {X}, \\theta + \\Delta \\theta) - \\mathcal {F} (\\mathcal {X}, \\theta) \\approx \\Delta \\theta \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\tag {3}", + "image_path": "16be99866bd95ae1242a2f7355323076747504e12df6a35ac5b7b14e7253c0b2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 330, + 511, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 330, + 511, + 357 + ], + "spans": [ + { + "bbox": [ + 329, + 330, + 511, + 357 + ], + "type": "interline_equation", + "content": "= - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} = - \\eta \\cdot \\frac {\\mathrm {d} \\mathcal {L}}{\\mathrm {d} \\mathcal {F}} \\left(\\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta} \\cdot \\frac {\\mathrm {d} \\mathcal {F}}{\\mathrm {d} \\theta}\\right),", + "image_path": "719f1008e7f22a9a3fb2299c49c2759a9ec566b708957d159020260a1993958e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": " denotes the learning rate, " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": " denotes loss function, and " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": " denotes parameters of " + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 364, + 545, + 448 + ], + "type": "text", + "content": ". Eq. (3) provides the variation of logits within one step, capturing the nonlinear dynamics based on gradients. When comparing the optimization outlined in Eq. (2) and Eq. (3), it becomes evident that simulation methods differ from fine-tuning in terms of objectives and optimization approaches." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 616 + ], + "type": "text", + "content": "Despite preliminary exploration, such simulation methods come with severe limitations due to these inconsistencies, leading to deviations in predicting the final state, as illustrated in Fig. 3. Specifically, simulation methods aim to enhance class separability within the feature space, whereas fine-tuning drives logits closer to labels. The insufficient objective alignment introduces distinct inductive bias compared to fine-tuning, resulting in notable variations in modeling dynamics [27, 34]. Moreover, due to the nonlinear nature of neural networks and complex parameter interactions, linear approximations in simulation methods often fail to accurately capture the dynamic changes [21]. Such limitations hinder the effective modeling of nonlinearity, further constraining modeling precision." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 545, + 647 + ], + "type": "text", + "content": "3.3. Modeling Logits Evolution Process through Dynamical Equation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "To address the aforementioned problems, our study delves deeper into the optimization analysis from Eq. (3). In fact, unlike features, the logits derived from the optimization output naturally align with the fine-tuning objectives and reflect the inherent nonlinearity from gradient descent. Therefore," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28666" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 275, + 155 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 275, + 155 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 275, + 155 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 275, + 155 + ], + "type": "image", + "image_path": "afd73633fe40bc8a1e10ada7da220617e5f53e607aef417db3824deb6a4ce21c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 72, + 544, + 239 + ], + "blocks": [ + { + "bbox": [ + 276, + 72, + 544, + 239 + ], + "lines": [ + { + "bbox": [ + 276, + 72, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 276, + 72, + 544, + 239 + ], + "type": "image", + "image_path": "799fab120c21ac44732bf9677d4919e8a10a9f12d1704ade2fb1d619870bd7b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 159, + 274, + 236 + ], + "blocks": [ + { + "bbox": [ + 48, + 159, + 274, + 236 + ], + "lines": [ + { + "bbox": [ + 48, + 159, + 274, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 274, + 236 + ], + "type": "image", + "image_path": "772c7f10ffb50c6f1a01eeb70f507672a56d23c57a1a790ab18c4238c5c4ccae.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 248, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 326 + ], + "type": "text", + "content": "Figure 2. Comparison of different approaches for modeling the evolution process. (a) Fine-tuning iteratively updates parameters via gradient descent, revealing authentic logits evolution trajectory and classification accuracy for ground-truth ranking. (b) Recent works (e.g., SFDA [41], PED [32]) perform linear transformations in feature space to enhance class separability, updating logits with modified features. The ranking score is determined by Cross-Entropy loss on updated logits. While these approaches simulate fine-tuning impact on features, they often suffer huge deviations from the authentic results. (c) Our method, LEAD, transforms the discrete logit evolution during fine-tuning into a gradient-based dynamical equation to capture nonlinear changes and can efficiently predict the final state through the closed-form solution. It establishes an analysis framework linking the initial and final states, enabling the precise prediction in one step." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 375 + ], + "type": "text", + "content": "it provides a particularly relevant entry point to analyze dynamical changes during fine-tuning and assess the model transferability through its evolution process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 375, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 530 + ], + "type": "text", + "content": "While the dynamical process of logit space analyzed by Eq. (3) remains consistent with fine-tuning, there still lacks an efficient approach for capturing logits evolution, due to the discrete process of iterating over thousands of steps by traditional optimization. To this end, we propose a new transferability metric called \"Logits Evolution Analysis through Dynamical equation (LEAD)\", that bridges the evolution process from the initial logit space towards the final state. In contrast to extensive optimization iterations, our method proposes a dynamical analysis framework for analyzing the nonlinear effects within the logit space, providing an efficient prediction to capture the final state with a concise solution in a single step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 591 + ], + "type": "text", + "content": "Theoretical Background. In our LEAD, we innovatively introduce the Neural Tangent Kernel (NTK) to simplify the dynamical equation. Here, we provide a brief introduction to the definition of the NTK kernel function and the constant-preserving property that we utilize." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 597, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 597, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 75, + 597, + 287, + 635 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}) \\triangleq \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) \\nabla_ {\\theta_ {t}} \\mathcal {F} (\\mathcal {X}; \\theta_ {t}) ^ {\\top}, (4) \\\\ \\Phi (\\mathcal {X}, \\mathcal {X}) \\triangleq \\lim _ {l \\rightarrow \\infty} \\hat {\\Phi} _ {\\theta_ {t}} (\\mathcal {X}, \\mathcal {X}), (4) \\\\ \\end{array}", + "image_path": "0acc49c0a9c98b6d962a379393be5c6a7f479e0f40f8ec57b6a12da687dc3f70.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " denotes the Laplace Operator, and " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " denotes the root mean square of the widths of all layers in the network. The kernel function " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}_{\\theta_t}" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " remains invariant throughout the training process as " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " approaches infinity with an approximation error of " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "O\\left(\\frac{1}{l^2}\\right)" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " [20, 26]. For simplicity, in the remainder of this paper, we denote functions at time " + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 337, + 547, + 443 + ], + "blocks": [ + { + "bbox": [ + 307, + 337, + 547, + 443 + ], + "lines": [ + { + "bbox": [ + 307, + 337, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 547, + 443 + ], + "type": "image", + "image_path": "86f95734fc810786cb98edbae5ec9033cac5a37e77337b4e6c0b793079762518.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 449, + 547, + 516 + ], + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 516 + ], + "type": "text", + "content": "Figure 3. Comparison of the average prediction of final state logits across various methods on VOC2007 [16], where the actual result stems from fine-tuning. Closer to actual results indicates more accurate modeling of the final state. Simulation methods (SFDA, PED) capture the correct trend but exhibit a large deviation from the actual results, affecting the accuracy of model ranking." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 517, + 523, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 523, + 532 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 523, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\cdot)\\triangleq \\mathcal{F}(\\cdot ,\\theta_t),\\hat{\\Phi}_t(\\cdot)\\triangleq \\hat{\\Phi}_{\\theta_t}(\\mathcal{X},\\mathcal{X}),\\Phi \\triangleq \\Phi (\\mathcal{X},\\mathcal{X})." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 686 + ], + "type": "text", + "content": "Dynamical Equation and Closed-form Solution. To bypass the complexity of iterative computations, we extend the analysis in Eq. (3) and theoretically derive a dynamical equation that is efficient for analysis. Specifically, we seek for a mathematical derivation through the theory of limit approximation to convert the discrete process into a continuous differential equation, enabling us to track the trajectory of logits in the solution space of the equation. Furthermore, we leverage the property of the NTK in Eq. (4) to simplify the equation into an Ordinary Differential Equation (ODE) with an initial value condition. Due to space constraints, we directly provide the equation and the detailed proof can be found in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 691, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 691, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 351, + 691, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\frac {\\mathrm {d} \\mathcal {F} _ {t}}{\\mathrm {d} t} = - \\eta \\cdot \\Phi \\cdot \\frac {\\mathrm {d} \\mathcal {L} _ {t}}{\\mathrm {d} \\mathcal {F} _ {t}}, \\mathcal {F} _ {0} = l o g _ {i n i t}, \\tag {5}", + "image_path": "90e9d8ad4c39b2f4323e9b04c06de19b2fed7e78ca8beb68cd8db8f2646eb967.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28667" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "\\text{loginit}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " is the initial state of logits obtained through feeding features and labels of downstream datasets to a classification algorithm. Compared to the discrete process, our proposed ODE depicts the continuous evolution of the finetuning dynamics, allowing for a more flexible modeling approach. It transforms discrete gradient effects into equivalent differential effects in continuous time, preserving the capability for modeling nonlinear changes and providing precise prediction of the evolution process. Furthermore, unlike employing numerical methods as the ODE solver, we can directly obtain its closed-form solution through integration over time dimension, allowing us to bypass the lengthy evolution process and reach the optimization destination in a single step, as shown below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 247, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(\\mathcal {F} _ {t} (\\mathcal {X})\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\Phi \\cdot t}\\right) \\mathcal {Y} + e ^ {- \\eta \\Phi \\cdot t} \\cdot \\log_ {\\text {i n i t}}. \\tag {6}", + "image_path": "4dc4bf6ec92f5a0b433ba2e920010d1834883d68c832307fa86a31be20ef0f81.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": "As depicted in Eq. (6), the solution is derived by interpolating between the one-hot label and the initial observation of logits, with the interpolation coefficients determined by the NTK matrix. It comprises two crucial elements: " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\text{log}_{\\text{init}}" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": ", representing the value of logits before optimization and determining the start of the evolution process; and " + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 270, + 287, + 389 + ], + "type": "text", + "content": ", indicating the convergence rate to the label and determining the trajectory of logit evolution. Our proposed framework efficiently models the evolution process, enabling a prediction of real-time changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": "Class-aware Decomposition. While Eq. (6) provides the theoretical result for the final state, computing the value of " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " under finite width remains challenging. Prior studies [20, 53] indicate that, even with finite width, the convergence rate can be assessed by using the eigenvalues of the initial state NTK matrix " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}_0" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": ". Inspired by this empirical finding, we further conduct a decomposition approach to bridge theory and practice, utilizing the eigenvalue decomposition on " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{\\Phi}_0" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " to replace the convergence rate determined by " + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 389, + 287, + 509 + ], + "type": "text", + "content": " in Eq. (6) with eigenvalues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "content": "Due to the network exhibiting varying classification abilities across different classes, their convergence behavior and evolution dynamics also differ during the fine-tuning process. To enhance the modeling of behaviors in different classes, we separately compute NTK for each class, rather than mixing them together, to decouple the respective optimization effect. Specifically, we extract " + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 509, + 288, + 640 + ], + "type": "text", + "content": " samples of each class to calculate their respective NTK matrices and perform decomposition to obtain eigenvalues. The average eigenvalue of each class serves as the unified parameter to determine the interpolation coefficient for this class:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 648, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 648, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 64, + 648, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left(\\mathcal {F} _ {t} (x)\\right) = \\left(\\mathbb {I} - e ^ {- \\eta \\bar {\\lambda} \\cdot t}\\right) y + e ^ {- \\eta \\bar {\\lambda} \\cdot t} \\cdot \\log_ {i n i t}, \\tag {7}", + "image_path": "6c395716824d4de7000d7aeaa3ec7517501d0fd19b2a450dd146a958c1c94f94.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is a given sample and its one-hot label, " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\overline{\\lambda}" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": " is the average eigenvalue of the NTK matrix associated with the class of " + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": ". Building upon the conclusion of Eq. (6), Eq." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "(7) avoids the demand of infinite width, enabling practical computability of the prediction results. As shown in Fig. 2, we employ the results of Eq. (7) to efficiently obtain predictions for the final state logits. Subsequently, following the common practice [32, 33, 41], we feed predictions into the Cross-Entropy loss to obtain the transferability score for model ranking." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 156, + 546, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 277 + ], + "type": "text", + "content": "Summary. Our method explicitly models the fine-tuning optimization through the theoretical analysis framework and provides a concise dynamical solution to bypass the lengthy optimization process in a single step, which seamlessly integrates the initial state observations and the evolution trajectory towards the final state. Therefore, we can efficiently predict the final logit state through this dynamical solution. Through comparing logit prediction accuracy among different models, it can serve as an effective metric for predicting model transferability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 285, + 388, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 388, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 388, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "content": "To demonstrate the effectiveness and robustness of our approach, we evaluate LEAD on diverse pre-trained models, including both supervised and self-supervised convolutional neural networks. Our evaluation spans 10 widely used classification benchmarks in transfer learning, employing weighted Kendall's " + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "content": " as rank correlation to evaluate the transferability metric. Additionally, we highlight the effectiveness of LEAD in the low-data regime, where only " + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "inline_equation", + "content": "2/5/10" + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "content": " samples are available per class." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 419, + 440, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 419, + 440, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 419, + 440, + 431 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": "Given a pre-trained model " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": ", we append a random initialized [22] MLP " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": " as a classification head after its backbone " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": ", and combine them as " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\coloneqq h \\circ f" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": " to produce logits. To compute the logits in the final state, we only need to derive the NTK matrix of each class and " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "log_{init}" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": " through Eq. (7). We follow the approach in [35] to compute the NTK matrix and employ a robust and efficient machine learning classifier, the multi-class SVM [49], to generate " + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "inline_equation", + "content": "log_{init}" + }, + { + "bbox": [ + 304, + 437, + 545, + 556 + ], + "type": "text", + "content": ". Due to space constraints, we provide a more detailed explanation and pseudocode in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 389, + 575 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 677 + ], + "type": "text", + "content": "Target Datasets. We employ 10 classification benchmarks commonly utilized in transfer learning research including Caltech-101 [17], Stanford Cars [29], CIFAR10 [30], CIFAR100 [30], DTD [10], Oxford 102 Flowers [37], Food-101 [4], Oxford-IIIT Pets [39], SUN397 [51], and VOC2007 [16]. These datasets encompass abundant characteristics, including backgrounds, textures, and coarse/fine-grained scenes, spanning diverse fields." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Ground Truth Rank. To conduct the ground truth ranking for the model zoo, we follow the approach in [41] and [52]. Specifically, we employ a grid search strategy to determine" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28668" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 544, + 224 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 1. Performance comparisons of different transferability metrics on the supervised model zoo. We measure the performance with weighted Kendall's " + }, + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "text", + "content": " and a larger " + }, + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 47, + 70, + 547, + 95 + ], + "type": "text", + "content": " represents a better prediction rank. The best results are in bold and the second-best are in underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 544, + 224 + ], + "type": "table", + "html": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3630.4220.0660.135-0.0920.5360.3100.114-0.0970.707
Logme [52]ICML'210.3360.3490.4250.5380.6030.6510.7830.3720.4400.673
Nsleep [33]CVPR'210.5040.6780.2920.4350.6860.4520.6560.6860.7700.645
PACTran [12]ECCV'220.7090.3720.4800.1510.7420.3720.7570.4510.4220.536
SFDA [41]ECCV'220.5340.5970.4210.4720.7490.4210.8170.6250.5620.728
GBC [38]CVPR'220.8440.4300.3480.7350.8300.4800.8660.3580.5740.654
ETran [18]ICCV'230.7090.6270.5340.5570.7490.6050.7190.6510.5770.673
PED [32]ICCV'230.5730.5970.4510.4720.7490.4210.7700.6250.5620.728
LEADThis paper0.8920.6980.7860.5860.8350.6890.7910.8410.6090.743
", + "image_path": "53306db3195106ed608cf304254ce6720c67983e0cbd5c2889f21459be68f674.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "text", + "content": "the actual performance of each model in the downstream task. This strategy involves trying various learning rates from the set " + }, + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "inline_equation", + "content": "\\{10^{-1}, 10^{-2}, 10^{-3}, 10^{-4}\\}" + }, + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "text", + "content": " and different weight decay values from the set " + }, + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "inline_equation", + "content": "\\{0, 10^{-6}, 10^{-5}, 10^{-4}, 10^{-3}\\}" + }, + { + "bbox": [ + 46, + 236, + 288, + 309 + ], + "type": "text", + "content": ". To ensure reliability, we repeat each experiment 5 times with different random seeds and take the average as the result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 314, + 228, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 228, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 228, + 327 + ], + "type": "text", + "content": "4.3. Evaluation on Supervised Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 426 + ], + "type": "text", + "content": "Model Zoo. We construct a Model Zoo that includes 12 CNN models supervised pre-trained on ImageNet [11], covering widely-used architectures: ResNet-34 [23], ResNet-50 [23], ResNet-101 [23], ResNet-152 [23], DenseNet-121 [25], DenseNet-161 [25], DenseNet-169 [25], DenseNet-201 [25], MNet-A1 [45], MobileNetV2 [40], GoogleNet [43], and InceptionV3 [44]. The accuracy results of these models after fine-tuning can be found in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": "Result Analysis. To show the effectiveness of LEAD, we compare it with prior arts, such as Bayesian-based Logme, Separability-based GBC, and simulation methods SFDA and PED. As shown in Tab. 1, simulation methods consider underlying dynamic effects and exhibit promising performances on some datasets. However, as the aforementioned analysis in Sec. 3.2 shows, the bias in the simulation process causes these methods to have unstable improvements, particularly resulting in unsatisfactory performance on Flowers, Food, and Pets. Benefiting from a precise analysis of the evolution, our LEAD yields a performance gain of " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "+0.335" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "+0.319" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "+0.216" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": " compared to PED on these three datasets, respectively. Overall, LEAD achieves the best " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": " on 7 target datasets and obtains an average " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": " of 0.747 which is relatively " + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 46, + 426, + 288, + 656 + ], + "type": "text", + "content": " better than the latest SOTA ETran [18]. Our experiments highlight the importance of accurate modeling of dynamics and confirm the effectiveness of our method to model the evolution process through the dynamical equation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 660, + 248, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 660, + 248, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 660, + 248, + 673 + ], + "type": "text", + "content": "4.4. Evaluation on Self-supervised Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Model Zoo. Self-supervised learning models (SSL models) have shown their efficacy in transfer learning and often display superior generalization performance compared" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 547, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 308 + ], + "type": "text", + "content": "to Supervised models. To illustrate the broad applicability of our approach, inspired by [32, 41], we assemble a self-supervised model zoo trained on ImageNet with BYOL [19], Infomin [46], PCLv1 [31], PCLv2 [31], Selav2 [1], InsDis [50], SimCLRv1 [7], SimCLRv2 [8], MoCov1 [24], MoCov2 [9], DeepClusterv2 [5], and SWAV [6]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "text", + "content": "Result Analysis. Due to differing pre-training and downstream classification objectives, SSL models undergo larger dynamic changes during the fine-tuning process, leading to distinct transferability properties compared to supervised models [15]. As shown in Tab. 1 and 2, methods that disregard dynamics (e.g. PARC, Nsleep) perform competitively in the supervised model zoo, but fail on some datasets for SSL models. For example, Nsleep and PARC have a negative small rank correlation on VOC2007 (-0.101) and CIFAR100 (-0.136), respectively. To maintain fairness, we refrain from comparing with GBC [38] in the SSL model zoo, which is designed for supervised scenarios. On the contrary, SFDA and PED exhibit better performances due to consideration of training dynamics. Nevertheless, the limitation of the simulating process still hinders their performance. Compared to them, LEAD achieves a consistent improvement on 7 datasets and obtains an average " + }, + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "text", + "content": " of 0.746 which is relatively " + }, + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 308, + 547, + 536 + ], + "type": "text", + "content": " better than PED (0.648). The results validate the effectiveness of LEAD for SSL models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 541, + 481, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 541, + 481, + 554 + ], + "spans": [ + { + "bbox": [ + 305, + 541, + 481, + 554 + ], + "type": "text", + "content": "4.5. Evaluation on Low-data Regime" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": "Experimental Settings. In practical application, privacy and resource constraints may limit our access to the entire downstream dataset. To explore the utilization boundary, we evaluate whether LEAD can provide effective model ranking only using limited samples. We follow three data settings in PACTran [12] with increasing average number of samples per class " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\{2,5,10\\}" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": " and divide 10 datasets into two groups according to the number of classes: " + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 304, + 557, + 547, + 700 + ], + "type": "text", + "content": " classes, including Food, Caltech, Flowers, Cars, CIFAR100, and SUN397; 10-99 classes, including DTD, CIFAR10, Pets, and VOC2007. For each setting, we repeatedly sample 5 times for evaluation and take the average as the result." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 547, + 714 + ], + "type": "text", + "content": "Result Analysis. As shown in Tab. 3, all methods, in-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28669" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 91, + 544, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 538, + 82 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 538, + 82 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 538, + 82 + ], + "type": "text", + "content": "Table 2. Performance comparisons on the self-supervised model zoo. The best results are in bold and the second-best are in underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 91, + 544, + 203 + ], + "lines": [ + { + "bbox": [ + 50, + 91, + 544, + 203 + ], + "spans": [ + { + "bbox": [ + 50, + 91, + 544, + 203 + ], + "type": "table", + "html": "
MethodReferenceDownstream Target Dataset
FoodCaltechFlowersCarsCIFAR100DTDCIFAR10PetsSUN397VOC2007
PARC [3]NeurIPS'210.3590.1960.6220.424-0.1360.4470.1470.496-0.0060.618
Logme [52]ICML'210.5700.0510.6040.375-0.0080.6270.2950.6840.2170.158
Nsleep [33]CVPR'210.5740.5250.5340.4860.2760.641-0.0440.7920.719-0.101
PACTran [12]ECCV'220.7200.6220.6010.4740.5290.6140.4770.6410.6380.620
SFDA [41]ECCV'220.6850.5230.7490.5150.5480.7730.6190.5860.6980.568
ETran [18]ICCV'230.4650.4050.6440.5870.6500.2140.6060.3380.5200.376
PED [32]ICCV'230.5810.6140.7770.6490.5680.9070.6730.4620.6610.583
LEADThis paper0.8600.7800.7250.6630.7760.8250.7130.6290.7600.723
", + "image_path": "280a395d418e0a632e7a7933c2d5f798a08610cd9966d7d43289875207612f96.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 264, + 544, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "text", + "content": "Table 3. Performance comparisons with different " + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "inline_equation", + "content": "N / K" + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "text", + "content": " is the number of samples for computing the metric. We divide 10 benchmarks into two groups according to the number of classes: 100+ classes and 10-99 classes, and evaluate the average of " + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 209, + 546, + 255 + ], + "type": "text", + "content": " on the two groups. Additionally, we also provide the average results on the 10 benchmarks, recorded as Avg. The best results are denoted in bold, the second best results are denoted in underline, and - denotes cannot give a valid ranking." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 264, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 50, + 264, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 50, + 264, + 544, + 350 + ], + "type": "table", + "html": "
Supervised / Self-supervised pre-trained Model
MethodN/K = 2N/K = 5N/K = 10
100+ classes10-99 classesAvg100+ classes10-99 classesAvg100+ classes10-99 classesAvg
GBC0.17 / 0.320.10 / 0.170.13 / 0.250.45 / 0.400.22 / 0.180.36 / 0.310.59 / 0.420.42 / 0.220.52 / 0.34
PACTran0.43 / 0.490.20 / 0.390.33 / 0.450.42 / 0.480.31 / 0.440.38 / 0.470.56 / 0.520.48 / 0.460.53 / 0.50
ETran0.42 / 0.430.12 / 0.410.29 / 0.420.48 / 0.470.31 / 0.380.41 / 0.430.56 / 0.470.56 / 0.480.56 / 0.48
PED- / 0.36- / 0.24- / 0.31- / 0.43- / 0.35- / 0.400.30 / 0.500.22 / 0.560.27 / 0.52
LEAD0.45 / 0.480.51 / 0.450.47 / 0.460.48 / 0.490.64 / 0.530.54 / 0.510.65 / 0.530.64 / 0.600.64 / 0.56
", + "image_path": "ddf1ad6e63472cda2967e11a0a1f7b9d6c98097fc163b549cb1b64d29d575b34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": "cluding PACTran designed for this scenario, exhibit significant performance declines under low-data settings. For instance, ETran's performance in the supervised model zoo drops from 0.64 with the entire dataset to 0.29, 0.41, and 0.56 in three low-data settings. Moreover, some methods, such as SFDA and PED fail to provide prediction rankings in some settings. For example, when " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "N / K = 2,5" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": ", PED cannot correctly evaluate class separability due to insufficient data, imposing huge changes to completely separate different classes. This leads to all models perfectly classifying all samples, resulting in identical scores that cannot be ranked. In contrast, our LEAD achieves stable results, ensuring " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\tau_w \\geq 0.45" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": " even in extreme settings like " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "N / K = 2" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": ", and consistently outperforms other methods. For example, under " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "N / K = 10" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": ", its average " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": " in the supervised model zoo reaches 0.64, which is relatively " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": " better than the latest SOTA ETran (0.56) and is equivalent to the latest SOTA on the complete dataset, with less than " + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 362, + 289, + 601 + ], + "type": "text", + "content": " of data size. The results demonstrate the robustness of our method in the face of data limitations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 611, + 262, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 262, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 262, + 625 + ], + "type": "text", + "content": "5. Ablation Study and Efficiency Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "In this section, without loss of generality, we conduct ablation studies on Caltech, CIFAR10, and VOC2007, building upon the evaluation of SSL models in Sec. 4.4 for saving time. We conduct separate experiments to analyze the impact of interpolation coefficients, a key component of LEAD, aiming to delve into the influence of hyperparameters and implementation details, gaining insights into key" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 399 + ], + "type": "text", + "content": "factors contributing to LEAD's effectiveness. Additionally, we provide a runtime comparison to verify efficiency, and more ablation study results are provided in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 407, + 444, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 444, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 444, + 419 + ], + "type": "text", + "content": "5.1. Interpolation Coefficient" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "content": "The interpolation coefficients are determined by time coefficient " + }, + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "content": " and the NTK matrix. We conduct ablation experiments on these two key factors:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "text", + "content": "Time Coefficient " + }, + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "text", + "content": ". As we model the discrete process as a continuous ODE, the continuous variable " + }, + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 461, + 546, + 594 + ], + "type": "text", + "content": " is proportional to the number of updates, determining the progress of the evolution process. In Fig. 4, we evaluate our method under varied time settings, keeping the unit as 1e6 consistent across all experiments. It is evident that performance improves as the time increases within a specific range. However, if the dynamic effect persists for an extended period (e.g., 10.0), all model predictions converge too closely to the ground-truth label, reducing discrimination for model ranking. Hence, we set the time coefficient as 1 by default." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 605, + 427, + 686 + ], + "blocks": [ + { + "bbox": [ + 313, + 605, + 427, + 686 + ], + "lines": [ + { + "bbox": [ + 313, + 605, + 427, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 427, + 686 + ], + "type": "image", + "image_path": "d8befea459eb8901a1a723d8409b407c7f5f958b1a950897281e0ed8c9dfe50a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 425, + 605, + 534, + 687 + ], + "blocks": [ + { + "bbox": [ + 425, + 605, + 534, + 687 + ], + "lines": [ + { + "bbox": [ + 425, + 605, + 534, + 687 + ], + "spans": [ + { + "bbox": [ + 425, + 605, + 534, + 687 + ], + "type": "image", + "image_path": "1b3c6d2bb15e3a12791b480ffe8032eff56a800c64b8ddf7c8ef96a1d83c0eba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "lines": [ + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "type": "text", + "content": "Figure 4. The left picture shows the performance gain in " + }, + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "type": "text", + "content": " with respect to different time " + }, + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 687, + 545, + 719 + ], + "type": "text", + "content": " and the right shows the performance in different sample sizes for each class NTK." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28670" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "Sample Size for NTK. To enhance the perception of different dynamic intensities of classes, we propose the class-aware decomposition that employs " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " samples of each class to obtain NTK matrix and perform spectrum analysis to derive eigenvalues. In Fig. 4, we conduct ablation experiments to explore the effect of varying " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". We observe that performance generally improves as " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " increases, and gradually reaches stability. This demonstrates that when the sample size reaches a certain size, eigenvalues reliably capture differences in dynamic intensity, offering a refined analysis result of the dynamics. Therefore, we set " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "S = 64" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " to ensure the method's efficiency and stability in all experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 219, + 200, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 200, + 232 + ], + "type": "text", + "content": "5.2. Running Time Comparison" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "text", + "content": "In experiment results, we have demonstrated the effectiveness of LEAD in various scenarios. In this section, we highlight the computational efficiency of LEAD. With the concise solution in Eq. (7), LEAD predicts the final state only requiring the NTK matrix and " + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\text{log}_{init}" + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "text", + "content": ". As shown in Tab. 4, the running time of LEAD is about half that of simulation methods (45s compared to 92s), ranking second after GBC which displays unstable performance. Notably, LEAD takes 3s to compute the NTK, while feature extraction takes 420s, resulting in only a marginal " + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "text", + "content": " increase in GPU time. The time to compute " + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\text{log}_{init}" + }, + { + "bbox": [ + 46, + 235, + 289, + 390 + ], + "type": "text", + "content": " is 42s CPU time, indicating significant potential for exploring faster classification algorithms to further reduce the overall time." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 400, + 287, + 422 + ], + "blocks": [ + { + "bbox": [ + 49, + 391, + 285, + 400 + ], + "lines": [ + { + "bbox": [ + 49, + 391, + 285, + 400 + ], + "spans": [ + { + "bbox": [ + 49, + 391, + 285, + 400 + ], + "type": "text", + "content": "Table 4. The comparisons of average running time on CIFAR10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 400, + 287, + 422 + ], + "lines": [ + { + "bbox": [ + 49, + 400, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 49, + 400, + 287, + 422 + ], + "type": "table", + "html": "
MetricsPARCNLEEPPACTranSFDAGBCPEDOurs
Running Time111s1430s77s92s26s97s45s
", + "image_path": "353febb933a476c50c193c6ea429c60bf7e23388e63e622dbfdea7ad1647a374.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 423, + 198, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 198, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 198, + 436 + ], + "type": "text", + "content": "6. Visualization and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 438, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 475 + ], + "type": "text", + "content": "In this section, we validate our approach through some visualizations and these results clearly illustrate how our model effectively addresses the model selection problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 479, + 146, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 146, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 146, + 491 + ], + "type": "text", + "content": "6.1. Logit Prediction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 289, + 628 + ], + "type": "text", + "content": "We have conducted a comparison of the predicted logits distribution on VOC2007 using different methods, as illustrated in Fig. 3. Notably, predictions obtained by LEAD closely match the authentic distribution after fine-tuning. To provide further quantitative analysis, we visualize the average bias in predicted logits in Fig. 5. Across three datasets, our method exhibits a substantial reduction compared to simulation methods, achieving nearly a " + }, + { + "bbox": [ + 46, + 495, + 289, + 628 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 495, + 289, + 628 + ], + "type": "text", + "content": " overall reduction. This visualization highlights the precision of our dynamical equation-based approach in capturing logit evolution for more accurate ranking sequences." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 632, + 276, + 685 + ], + "blocks": [ + { + "bbox": [ + 59, + 632, + 276, + 685 + ], + "lines": [ + { + "bbox": [ + 59, + 632, + 276, + 685 + ], + "spans": [ + { + "bbox": [ + 59, + 632, + 276, + 685 + ], + "type": "image", + "image_path": "5730efca410870bc74277eafa6b7679899fbff59307e049120384f6d9a9826e7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "lines": [ + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 685, + 287, + 718 + ], + "type": "text", + "content": "Figure 5. Comparison of the average difference between logits predicted by different methods and logits after fine-tuning. The difference is calculated utilizing the Euclidean distance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 402, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 402, + 83 + ], + "type": "text", + "content": "6.2. Rank Variation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 86, + 547, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 86, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 86, + 547, + 230 + ], + "type": "text", + "content": "By employing our method, we can refine models that have poor initial observation but show better progress during training. To assess the effectiveness of our approach, we conduct a visualization comparing the model rankings between the initial observations and our predicted final state, as shown in Fig. 6. The results clearly illustrate a substantial enhancement in the calibration of model rankings when using our refined observations, with the optimal model now positioned closely to the reference line in the figure. Notably, models starting from an initially unfavorable point can be rapidly elevated to achieve a superior ranking, highlighting the effectiveness of our methodology." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 236, + 380, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 236, + 380, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 236, + 380, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 380, + 304 + ], + "type": "image", + "image_path": "2f3583d5c16fb243a02f5b430a4256595b02cb4df51f2e295eacd8a799197982.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 389, + 235, + 460, + 304 + ], + "blocks": [ + { + "bbox": [ + 389, + 235, + 460, + 304 + ], + "lines": [ + { + "bbox": [ + 389, + 235, + 460, + 304 + ], + "spans": [ + { + "bbox": [ + 389, + 235, + 460, + 304 + ], + "type": "image", + "image_path": "c6876fb90cafd3d8cdb83f6640d4e998b199078d9922496892cbbb318c52a4ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 471, + 235, + 542, + 304 + ], + "blocks": [ + { + "bbox": [ + 471, + 235, + 542, + 304 + ], + "lines": [ + { + "bbox": [ + 471, + 235, + 542, + 304 + ], + "spans": [ + { + "bbox": [ + 471, + 235, + 542, + 304 + ], + "type": "image", + "image_path": "47164e21d643dbb8ece11d0d1697e86c74aac8c23ea55f72173eb84e597ae7c1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 315, + 380, + 384 + ], + "blocks": [ + { + "bbox": [ + 307, + 315, + 380, + 384 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 380, + 384 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 380, + 384 + ], + "type": "image", + "image_path": "075db0296cb59584eb3a8b14d5d198af5096a3c0919fb40315dd6521e9dd9594.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 394, + 546, + 438 + ], + "lines": [ + { + "bbox": [ + 304, + 394, + 546, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 546, + 438 + ], + "type": "text", + "content": "Figure 6. The variation in model ranking, utilizing initial states (first row) and predicted final states of LEAD (second row). Results span three datasets, from left to right: CIFAR10, Caltech, and VOC2007. Shaded areas reflect the range of the ranking error." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 389, + 315, + 460, + 384 + ], + "blocks": [ + { + "bbox": [ + 389, + 315, + 460, + 384 + ], + "lines": [ + { + "bbox": [ + 389, + 315, + 460, + 384 + ], + "spans": [ + { + "bbox": [ + 389, + 315, + 460, + 384 + ], + "type": "image", + "image_path": "1b4b2ab39d6d5f3823aa39cfe2de811aba1c89e26dfe2f179ef288de09157e83.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 471, + 315, + 542, + 384 + ], + "blocks": [ + { + "bbox": [ + 471, + 315, + 542, + 384 + ], + "lines": [ + { + "bbox": [ + 471, + 315, + 542, + 384 + ], + "spans": [ + { + "bbox": [ + 471, + 315, + 542, + 384 + ], + "type": "image", + "image_path": "126bb530f74b06ea081490901547b8ad5456047169c6db2349993f4dba8e19d6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 439, + 476, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 476, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 476, + 451 + ], + "type": "text", + "content": "7. Conclusions and Future Work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 453, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 547, + 609 + ], + "type": "text", + "content": "In this work, we propose a dynamical equation designed to capture the evolution process within the logit space, offering an efficient solution for assessing model transferability. We construct a theoretical analysis framework and further devise a more practical and precise solution. And we demonstrate its consistent effectiveness across various scenarios including supervised and self-supervised, full-data and low-data settings. In future work, we intend to expand the current analysis framework beyond classification to cover a broader range of downstream tasks. We hope that our work can enhance the understanding of optimization mechanisms in deep neural networks and shed light on other fields, such as the design of optimization algorithms." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 616, + 404, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 404, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 404, + 628 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62088102, in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation and in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28671" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. arXiv, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "text", + "content": "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An information-theoretic approach to transferability in task transfer learning. In ICIP, pages 2309-2313, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 203 + ], + "type": "text", + "content": "[3] Daniel Bolya, Rohit Mittapalli, and Judy Hoffman. Scalable diverse model selection for accessible transfer learning. In NeurIPS, pages 19301-19312, 2021. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 236 + ], + "type": "text", + "content": "[4] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, 2014. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 238, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 271 + ], + "type": "text", + "content": "[5] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 316 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 317, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 350 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 394 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. In NeurIPS, pages 22243-22255, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "text", + "content": "[9] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 430, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 463 + ], + "type": "text", + "content": "[10] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 497 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 543 + ], + "type": "text", + "content": "[12] Nan Ding, Xi Chen, Tomer Levinboim, Soravit Changpinyo, and Radu Soricut. Pactran: Pac-bayesian metrics for estimating the transferability of pretrained models to classification tasks. In ECCV, pages 252-268. Springer, 2022. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 576 + ], + "type": "text", + "content": "[13] Kshitij Dwivedi and Gemma Roig. Representation similarity analysis for efficient task taxonomy & transfer learning. In CVPR, pages 12387–12396, 2019. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 621 + ], + "type": "text", + "content": "[14] Kshitij Dwivedi, Jiahui Huang, Radoslaw Martin Cichy, and Gemma Roig. Duality diagram similarity: a generic framework for initialization selection in task transfer learning. In ECCV, pages 497-513, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[15] Linus Ericsson, Henry Gouk, and Timothy M Hospedales. How well do self-supervised models transfer? In CVPR, pages 5414-5423, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[16] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88:303–338, 2010. 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[17] Li Fei-Fei. Learning generative visual models from few training examples. In CVPR workshop, 2004. 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[18] Mohsen Gholami, Mohammad Akbari, Xinglu Wang, Behnam Kamranian, and Yong Zhang. Etran: Energy-based transferability estimation. In ICCV, pages 18613-18622, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 184 + ], + "type": "text", + "content": "[19] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[20] Boris Hanin and Mihai Nica. Finite depth and width corrections to the neural tangent kernel. In ICLR, 2020. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 241 + ], + "type": "text", + "content": "[21] Hangfeng He and Weijie Su. The local elasticity of neural networks. In ICLR, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 276 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 277, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 277, + 545, + 311 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 346 + ], + "type": "text", + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 346, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 346, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 346, + 545, + 380 + ], + "type": "text", + "content": "[25] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 381, + 545, + 414 + ], + "type": "text", + "content": "[26] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. In NeurIPS, 2018. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 415, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 449 + ], + "type": "text", + "content": "[27] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, pages 7482-7491, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 545, + 482 + ], + "type": "text", + "content": "[28] Simon Kornblith, Jonathon Shlens, and Quoc V. Le. Do better imagenet models transfer better? In CVPR, pages 2656-2666, 2018. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 484, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 506 + ], + "type": "text", + "content": "[29] Jonathan Krause, Jia Deng, Michael Stark, and Li Fei-Fei. Collecting a large-scale dataset of fine-grained cars. 2013. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 508, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 508, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 508, + 545, + 529 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 563 + ], + "type": "text", + "content": "[31] Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations. arXiv, 2020. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 566, + 545, + 609 + ], + "type": "text", + "content": "[32] Xiaotong Li, Zixuan Hu, Yixiao Ge, Ying Shan, and Ling-Yu Duan. Exploring model transferability through the lens of potential energy. In ICCV, pages 5429-5438, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 644 + ], + "type": "text", + "content": "[33] Yandong Li, Xuhui Jia, Ruoxin Sang, Yukun Zhu, Bradley Green, Liqiang Wang, and Boqing Gong. Ranking neural checkpoints. In CVPR, pages 2663-2673, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[35] Mohamad Amin Mohamadi and Danica J. Sutherland. A fast, well-founded approximation to the empirical neural tangent kernel. In ICML, 2022. 2, 5" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28672" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[36] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. Leep: A new measure to evaluate transferability of learned representations. In ICML, pages 7294-7305, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[37] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "type": "text", + "content": "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability estimation using bhattacharyya class separability. In CVPR, pages 9172-9182, 2022. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "text", + "content": "[39] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In " + }, + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 48, + 205, + 287, + 237 + ], + "type": "text", + "content": ", pages 3498-3505, 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 281 + ], + "type": "text", + "content": "[40] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "text", + "content": "[41] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In ECCV, pages 286-302, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 359 + ], + "type": "text", + "content": "[42] Grace S. Shieh. A weighted kendall's tau statistic. Statistics Probability Letters, 39(1):17-24, 1998. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "text", + "content": "[43] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pages 1-9, 2015. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 446 + ], + "type": "text", + "content": "[44] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[45] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 535 + ], + "type": "text", + "content": "[46] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[47] Anh T Tran, Cuong V Nguyen, and Tal Hassner. Transferability and hardness of supervised classification tasks. In ICCV, pages 1395-1405, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[48] Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[49] Jason Weston, Chris Watkins, et al. Support vector machines for multi-class pattern recognition. In *Esann*, pages 219–224, 1999. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[50] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, pages 3733-3742, 2018. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[51] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 161 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 115 + ], + "type": "text", + "content": "[52] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In ICML, pages 12133-12143, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 546, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 161 + ], + "type": "text", + "content": "[53] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In CVPR, pages 19228-19238, 2022. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28673" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_content_list.json b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7418bb08902896b2fa954e967ec824606d77e8ee --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "LEAD: Learning Decomposition for Source-free Universal Domain Adaptation", + "text_level": 1, + "bbox": [ + 86, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanqing $\\mathrm{Qu}^1$ , Tianpei $\\mathrm{Zou}^1$ , Lianghua $\\mathrm{He}^1$ , Florian Röhrbein $^2$ , Alois Knoll $^3$ , Guang Chen $^{1*}$ , Changjun Jiang $^1$", + "bbox": [ + 233, + 179, + 733, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University, $^{2}$ Chemnitz University of Technology, $^{3}$ Technical University of Munich", + "bbox": [ + 212, + 218, + 754, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Universal Domain Adaptation (UniDA) targets knowledge transfer in the presence of both covariate and label shifts. Recently, Source-free Universal Domain Adaptation (SF-UniDA) has emerged to achieve UniDA without access to source data, which tends to be more practical due to data protection policies. The main challenges lie in determining whether covariate-shifted samples belong to target-private unknown categories. Existing methods tackle this either through hand-crafted thresholding or by developing time-consuming iterative clustering strategies. In this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. Technically, LEAD initially leverages the orthogonal decomposition analysis for feature decomposition. Then, LEAD builds instance-level decision boundaries to adaptively identify target-private data. Extensive experiments across various UniDA scenarios have demonstrated the effectiveness and superiority of LEAD. Notably, in the OPDA scenario on VisDA dataset, LEAD outperforms GLC by $3.5\\%$ overall H-score and reduces $75\\%$ time to derive pseudo-labeling decision boundaries. Besides, LEAD is also appealing in that it is complementary to most existing methods. The code is available at https://github.com/ispc-lab/LEAD.", + "bbox": [ + 76, + 295, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have yielded impressive results in a variety of computer vision tasks [12, 18, 22, 42]. However, DNNs often fail to generalize well to new domains due to the discrepancy between training (source) and test (target) data distributions (i.e., covariate shift). Such a shift poses significant challenges for some safety-critical applications like autonomous driving [9, 10], or medical imaging [13, 31]. Unsupervised Domain Adaptation (DA) [16, 23] aims to address this challenge by transferring task-specific knowledge from a labeled source domain to an unlabeled target domain. Despite promising results, most", + "bbox": [ + 76, + 712, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eedb42aa8eec51aebbbdf6588d9e6093c0cee0ecfc340e4655e0416bee70ea83.jpg", + "image_caption": [ + "Figure 1. (a) Pseudo-label accuracy curves for target-private data in the OPDA scenario on VisDA dataset. An observation is that clustering-based GLC does not work well in deriving decision boundaries for target-private data. This may be due to the curse of dimensionality, leading K-means to faintly discerning clusters. (b) Frequency distribution for the normalized feature magnitude of target data in the source-unknown space (orthogonal complement of the space spanned by source model weights). Features are on task $(A\\to D)$ in the OPDA scenario of Office-31. The results show that target-private data are expected to involve more components from source-unknown space, even with covariate shifts." + ], + "image_footnote": [], + "bbox": [ + 504, + 260, + 702, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/311904e7893df586bad37f09500ec8d8c0e403d392f568a50a4fba25e4cff427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 260, + 888, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "existing favorable DA methods [16, 20, 27, 45] assume that the label spaces are identical across the source and target domain, thus being only applicable to closed-set scenarios and limiting applicability to practical applications.", + "bbox": [ + 496, + 549, + 893, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To handle generalized cases, Universal Domain Adaptation (UniDA) [47] was proposed to allow all potential label shifts between source and target domains. Different from existing methods specialized for Partial Domain Adaptation (PDA) [3, 4], Open-set Domain Adaptation (OSDA) [35, 46], or Open-Partial Domain Adaptation (OPDA) [44, 54], in UniDA, we have no prior knowledge about the label shift, e.g., information on matched common classes or the number of categories in the target domain. Nonetheless, most existing UniDA works [7, 25, 47] necessitate concurrent access to source and target data. This requirement becomes increasingly impractical in light of strict data protection policies [50]. In this paper, we focus on Source-free Universal Domain Adaptation (SF-UniDA) [28, 41], where only a pre-trained source model is provided for knowledge transfer rather than labeled source data.", + "bbox": [ + 496, + 611, + 895, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary objective for UniDA is to recognize target data belonging to common categories and separate them from target-private \"unknown\" data. Existing meth", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: guangchen@tongji.edu.cn", + "bbox": [ + 94, + 887, + 364, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23334", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ods have designed various criteria based on instance-level predictions, e.g., entropy-based [14, 47, 54], confidence-based [15, 24, 44]. However, defining a universal threshold for all scenarios is challenging and often sub-optimal. To mitigate tedious parameter-tuning, some methods start to perform source and target consensus clustering to achieve common and private data separation [7, 25]. Considering that SF-UniDA is inaccessible to source data, GLC [41] devises an iterative one-vs-all clustering strategy to realize target-private data identification. Regardless of progress, iterative clustering can be computationally intensive. Besides, due to the curse of dimensionality [2], methods like K-means are prone to overfitting and faintly discerning clusters. This can be implied to some degree from Figure 1 (a).", + "bbox": [ + 80, + 90, + 467, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose to address the SF-UniDA problem from a new perspective, i.e., feature decomposition. Our launching point is that, even though target data might have shifted in the feature space, target-private data from unknown categories are likely to exhibit a higher proportion of components from the orthogonal complement (source-unknown) space spanned by source model weights. The frequency distribution of normalized feature magnitudes in the source-unknown space depicted in Figure 1 (b) basically demonstrates our argument, which has been rarely explored in the literature. To materialize our idea, we propose the LEArning Decomposition (LEAD) framework. In particular, LEAD initially leverages the orthogonal decomposition to construct source-known and -unknown space. The projection on source-unknown space is then regarded as the discriminative representation with respect to target-private categories. Different from existing thresholding methods introducing a fixed global hand-crafted threshold for all scenarios, LEAD deliberately considers the distance to the target prototypes and source anchors to establish instance-level decision boundaries. We analyze and evaluate our LEAD under a variety of SF-UniDA scenarios, ranging from PDA, OSDA, to OPDA. Extensive experiments demonstrate the effectiveness and superiority.", + "bbox": [ + 80, + 303, + 467, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution is the proposal of a LEArning Decomposition (LEAD) framework for source-free universal domain adaptation (SF-UniDA). This solution leads to elegant views for identifying target-private unknown data without tedious tuning thresholds or relying on iterative unstable clustering. Remarkably, in the OPDA scenario on VisDA, LEAD attains an H-score of $76.6\\%$ , surpassing GLC [41] by $3.5\\%$ . Besides, LEAD is complementary to most existing SF-UniDA methods. For instance, in the OPDA scenario on Office-Home, LEAD advances UMAD [28] with H-score improvement from $70.1\\%$ to $78.0\\%$ .", + "bbox": [ + 80, + 665, + 467, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 845, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Domain Adaptation. When DNNs encounter data from a distribution that differs from their train", + "bbox": [ + 80, + 871, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing data, inevitable performance degeneration is commonly observed. To address this issue, domain adaptation (DA) [16, 23] has been introduced, leveraging labeled data from source domains to train DNNs for unlabeled target domains in a transductive learning manner. Existing DA methods can be broadly categorized into two paradigms, namely moment matching [11, 34, 48] and adversarial learning [16, 19, 21]. Regardless of their effectiveness in various applications such as object recognition [16, 27, 39], semantic segmentation [19, 48], and object detection [8, 21], most existing approaches assume that label spaces are identical across the source and target domains, limiting their applicability.", + "bbox": [ + 503, + 90, + 888, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Universal Domain Adaptation. There have been some approaches introduced to deal with label-shift scenarios, such as open-set domain adaptation (OSDA) [32, 35, 46], partial-set domain adaptation (PDA) [3, 4], and open-partial-set domain adaptation (OPDA) [15, 44, 54]. However, these methods are often tailored to specific scenarios and may not be readily applicable to other label shift situations. Universal domain adaptation (UniDA), on the other hand, is designed to facilitate adaptation across all potential label shift scenarios. Despite the progress in UniDA, many existing methods [5, 7, 44] necessitate simultaneous access to both source and target data, making them impractical for applications with stringent data protection policies. To address this challenge, source-free universal domain adaptation (SF-UniDA) [28, 41] has been proposed, wherein source data are only utilized for pre-training and are inaccessible during target adaptation. To promote the distinction between common and private data, existing methods often resort to manually defined thresholding or the introduction of global clustering techniques. Nevertheless, selecting an appropriate threshold for all scenarios can be both laborious and sub-optimal. Additionally, K-means clustering tends to be unstable when applied to high-dimensional data due to the curse of dimensionality [2]. Our work delves into these limitations and proposes a novel and elegant solution from the perspective of feature decomposition.", + "bbox": [ + 503, + 277, + 888, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Decomposition. As a fundamental technique in machine learning, feature decomposition is applied to deconstruct complex data into simpler and more interpretable components. In the field of domain adaptation and domain generalization, feature decomposition has been leveraged in various studies [6, 26, 30, 33, 38, 40] to enhance feature alignment or facilitate domain-invariant feature learning by breaking down features into content and style components. Nonetheless, the content and style decomposition design is less applicable for UniDA, as the primary challenge lies in distinguishing common and private data. In this paper, we approach the problem from the perspective of orthogonal feature decomposition and employ Independent Component Analysis to decompose features into two orthogonal parts, thereby facilitating the intended objective.", + "bbox": [ + 503, + 674, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23335", + "bbox": [ + 478, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb8beafd7fe75af0f36aaa1ef4acfd447b286f8f39fe8c69684b04b60cc5fffe.jpg", + "image_caption": [ + "Figure 2. (a) Illustrations of traditional universal domain adaptation (UniDA) and source-free universal domain adaptation (SF-UniDA). Traditional UniDA methods necessitate data from both source and target domains concurrently. In SF-UniDA, source data are solely utilized for pre-training. Adaptation is performed by harnessing target data and the source model $f_{\\theta}^{s} = h_{\\theta}^{s} \\circ g_{\\theta}^{s}$ . (b) An overview of our Learning Decomposition (LEAD) framework. Pseudo-labeling is an important technique for UniDA and SF-UniDA. The primary objective is to recognize target data associated with common label sets and exclude data within the target-private label space. Different from existing methods that perform private data identification by hand-crafted thresholding on predictions or iterative global clustering, we tackle this from the viewpoint of feature decomposition. The rationale is that despite potential shifts in the feature space, target-private data are expected to encompass more components from the orthogonal complement (source-unknown) space of the source model. Technically, LEAD first performs orthogonal decomposition to decompose target features into source-known and -unknown parts, i.e., $\\mathbf{z}_{i,knw}^{t}$ and $\\mathbf{z}_{i,unk}^{t}$ . $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ is considered as an indicator for private data. Next, LEAD employs a two-component Gaussian Mixture Model to estimate the distribution of $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ . Thereafter, LEAD devises a metric named \"common score\" $\\epsilon_{i,c}$ that accounts for distances to both target prototypes and source anchors (derived from $h_{\\theta}^{t}$ ) to facilitate deriving instance-level decision boundary $\\rho_{i,c}$ . LEAD provides an elegant solution to distinguish target-private data, mitigating the need for tedious hand-crafted threshold tuning or dependence on time-consuming iterative clustering. LEAD could also serve as a complementary approach to most existing SF-UniDA methods." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 331, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7ecf330707fdfa614df79bb0c5548e32eddd9d563b3d72c33366cd7e75e9087c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 88, + 867, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 579, + 210, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 604, + 205, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we consider a generalized and challenging case in domain adaptation: universal domain adaptation (UniDA), which aims to achieve knowledge transfer in the presence of both covariate and label shifts. In UniDA, there is a labeled source domain $\\mathcal{D}^s = \\left\\{\\left(\\mathbf{x}_i^s,\\mathbf{y}_i^s\\right)\\right\\}_{i = 1}^{N_s}$ where $\\mathbf{x}_i^s\\in \\mathcal{X}^s\\subset \\mathbb{R}^X,\\mathbf{y}_i^s\\in \\mathcal{Y}^s\\subset \\mathbb{R}^C$ , and an unlabeled target domain $\\mathcal{D}^t = \\left\\{\\left(\\mathbf{x}_i^t,\\ref{eq:2.3.1}\\right)\\right\\}_{i = 1}^{N_t}$ where $\\mathbf{x}_i^t\\in \\mathcal{X}^t\\subset \\mathbb{R}^X.$ $\\mathcal{V} = \\mathcal{V}^s\\cap \\mathcal{V}^t$ denotes the common label sets shared by source and target domains. $\\bar{\\mathcal{V}}^s = \\mathcal{V}\\setminus \\mathcal{V}^t$ and $\\bar{\\mathcal{V}}^t = \\mathcal{V}\\setminus \\mathcal{V}^s$ symbolizes the label sets private to source and target domain, individually. UniDA supposes that we have no prior knowledge about $\\mathcal{V}^t$ . $\\mathcal{V}$ and $\\bar{\\mathcal{V}}^t$ are also unavailable. The objective is to recognize common data belonging to $\\mathcal{V}$ and reject private data in $\\bar{\\mathcal{V}}^t$", + "bbox": [ + 75, + 628, + 467, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from traditional UniDA methods that need concurrent access to source and target data, under the source-free universal domain adaptation (SF-UniDA) setting, $\\mathcal{D}_s$ is only applicable for source model pre-training and not acces-", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sible during adaptation. Assuming that the source model is represented as $f_{\\theta}^{s} = h_{\\theta}^{s}\\circ g_{\\theta}^{s}$ , where $g_{\\theta}^{s}:\\mathbb{R}^{X}\\to \\mathbb{R}^{D}$ denotes the feature extractor and $h_\\theta^s:\\mathbb{R}^D\\to \\mathbb{R}^C$ is the classifier module. Following existing methods [27, 41], we freeze the classifier module $h_\\theta^t = h_\\theta^s$ and only learn a target-specific feature extractor $g_{\\theta}^{t}$ to realize model adaptation. Figure 2 (a) presents an illustration for UniDA and SF-UniDA.", + "bbox": [ + 498, + 580, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main challenge is how to distinguish common data from private data. Prior methods [25, 28, 41, 47] have designed various algorithms to encourage the separation. However, these methods perform data identification either based on hand-crafted thresholding criteria or by introducing time-consuming unstable clustering strategies. We propose to mitigate these limitations from the perspective of feature decomposition. An observation is that even though target data may have shifted in the feature space, the features of private data still include more components from the orthogonal complement (source-unknown) space of the pre-trained model. Technically, we present a novel learning decomposition (LEAD) framework. LEAD capitalizes on the orthogonal decomposition to build two orthogonal fea", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23336", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture spaces, i.e., source-known and -unknown space. Feature projection on source-unknown space is extracted as the descriptor for private data. Then, LEAD considers the distances to both the target prototypes and source anchors to establish instance-level decision boundaries. Figure 2 (b) manifests the architecture of our LEAD.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Orthogonal Feature Decomposition", + "text_level": 1, + "bbox": [ + 76, + 191, + 385, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo-labeling serves as an important technique in unsupervised learning, yet many existing strategies only account for closed-set scenarios, disregarding the label-shift situations, which impairs their ability to differentiate target-private data. To fulfill the objective of common and private data separation, we propose to orthogonally decompose the feature representation into two uncorrelated parts, i.e., features associated with source-known space $F_{knw}$ and those related to source-unknown space $F_{unk}$ . This design stems from a simple inductive bias in favor of private data identification. Due to the lack of exposure to target data during source pre-training, particularly target-private data, the source-known space does not include any significant attributes of target-private data. Subsequently, after normalization, target-private data are expected to contain more components from source-unknown space, even with covariate shifts. This hypothesis is basically demonstrated by empirical evidence, as illustrated in Figure 1 (b).", + "bbox": [ + 75, + 214, + 472, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, denoting the weight of classifier $h_{\\theta}^{t}$ is $W_{cls} \\in \\mathbb{R}^{C \\times D}$ . Source-known space $F_{knw}$ can be represented as the vector space of span $\\{\\mathbf{w}_n\\}$ , where $\\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\}$ is a group weight vectors of $W_{cls}$ . Source-unknown space $F_{unk}$ is the orthogonal complement space of $F_{knw}$ , i.e., $\\mathbb{R}^D = F_{knw} \\oplus F_{unk}$ . Nevertheless, unit orthogonality in $\\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\}$ is not guaranteed (i.e., when $n \\neq m$ ( $\\mathbf{w}_n, \\mathbf{w}_m \\neq 0$ ), making it hard to derive $F_{unk}$ . Considering that $F_{knw}$ and $F_{unk}$ are the row space and null space of $W_{cls}$ . Therefore, we employ the singular value decomposition (SVD) on $W_{cls}$ to obtain $F_{knw}$ and $F_{unk}$ . Specifically,", + "bbox": [ + 76, + 487, + 472, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {c l s} = U \\Sigma V ^ {T},\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 664, + 282, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {k n w} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = 1, \\dots , C} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 684, + 468, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {u n k} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = C + 1, \\dots , D} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 704, + 377, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Sigma \\in \\mathbb{R}^{C\\times D}$ is a diagonal matrix, $U\\in \\mathbb{R}^{C\\times C}$ and $V\\in \\mathbb{R}^{D\\times D}$ are both orthogonal unitary matrices. $\\{\\mathbf{v}_n|\\mathbf{v}_n\\in \\mathbb{R}^D\\}_{n = 1}^D$ are columns of $V$", + "bbox": [ + 76, + 722, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming that $\\mathbf{z}_i^t$ is the feature of target data $\\mathbf{x}_i^t$ extracted from feature extractor $g_{\\theta}^{t}$ after normalization, i.e., $\\mathbf{z}_i^t = g_\\theta^t (\\mathbf{x}_i^t) / ||g_\\theta^t (\\mathbf{x}_i^t)||_2$ , then $\\mathbf{z}_i^t$ can be represented as the weighted summation of two orthogonal bases:", + "bbox": [ + 75, + 770, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} ^ {t} = \\underbrace {\\sum_ {n = 1} ^ {C} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, k n w} ^ {t}} + \\underbrace {\\sum_ {n = C + 1} ^ {D} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, u n k} ^ {t}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 837, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $||\\mathbf{z}_{i,knw}^{t}||_{2}^{2} + ||\\mathbf{z}_{i,unk}^{t}||_{2}^{2} = 1$ , $\\mathbf{z}_{i,knw}^{t}$ and $\\mathbf{z}_{i,unk}^{t}$ are the projections of $\\mathbf{z}_{i}^{t}$ on $F_{knw}$ and $F_{unk}$ , respectively. $l_{n} \\in \\mathbb{R}$ denotes the projection weight of $\\mathbf{z}_{i}^{t}$ onto the unit base $\\mathbf{v}_{n}$ . This procedure is commonly recognized as orthogonal decomposition and has been extensively studied in the context of Independent Component Analysis (ICA).", + "bbox": [ + 496, + 88, + 893, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Separation for Private Data", + "text_level": 1, + "bbox": [ + 498, + 194, + 823, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, enabled by feature decomposition, we could differentiate between common and private data by examining the dominance of $\\mathbf{z}_{i,unk}^{t}$ over $\\mathbf{z}_i$ . However, such a straightforward approach is not applicable to SF-UniDA due to the impact of distribution covariate shift, which often deteriorates the significance of $\\mathbf{z}_{i,knw}^{t}$ for common data, leading to the prevalence of $\\mathbf{z}_{i,unk}^{t}$ in the representation. Additionally, the feature extractor inevitably captures representations that are independent of $F_{knw}$ , including background contexts [1], object textures [17], and high-frequency patterns [51], further exacerbating this issue.", + "bbox": [ + 496, + 217, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve this, one feasible solution is to estimate the distribution of $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ to effectively calibrate these disturbances. Our observations have revealed that the empirical distributions of $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ exhibit a bimodal pattern, with the presence of two distinct peaks typically indicating the modes of common and private data. Consequently, we employ a two-component Gaussian Mixture Model (GMM) to estimate the distribution $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ , with the components featuring a lower mean representing the common data and those with a higher mean corresponding to the private data. In particular, we denote $\\mu_{com}$ and $\\mu_{pri}$ as the expectation for common and private data, where $\\mu_{com} < \\mu_{pri}$ .", + "bbox": [ + 496, + 386, + 892, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the light of $\\mu_{com}$ and $\\mu_{pri}$ , an intuitive practice is to treat all common and private categories equally. Accordingly, common and private data can be separated by establishing $(\\mu_{com} + \\mu_{pri}) / 2$ as the threshold, where data with $\\| \\mathbf{z}_{i,unk}^t\\| _2$ below the threshold are recognized as common data, while those above are identified as private data. Despite its reasonable effectiveness, it discounts the inconsistency of covariate shifts across each category. Moreover, it fails to account for the variability in covariate shifts even among samples from the same category. These considerations motivate us to devise an instance-level decision strategy to achieve accurate and effective data identification.", + "bbox": [ + 496, + 568, + 893, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, we first leverage top- $K$ sampling to construct target prototypes, denoted as $\\{\\mathbf{c}_c^t\\in \\mathbb{R}^D |_{c = 1,\\dots ,C}\\}$ for each common category, adhering to the procedure in [29, 41]. Similar to [41], we empirically set $K = N_{t} / \\tilde{C}_{t}$ , where $N_{t}$ represents the amount of target data, and $\\tilde{C}_t$ signifies the estimated count of target categories. Diverging from [41], we additionally collect source anchors from $W_{cls}$ , denoted as $\\{\\mathbf{c}_c^s\\in \\mathbb{R}^D |_{c = 1,\\ldots ,C}\\}$ to assist us handle scenarios involving source-private categories. The construct of target prototypes through top- $K$ sampling fails to consider these sce", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23337", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "narios, making it susceptible to misidentifying private data as common. Further details regarding the target prototype construction are included in the supplementary.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thereafter, we propose a novel metric, common score $\\epsilon_{i,c}$ , which deliberately considers the distance of $\\mathbf{z}_i^t$ to the target prototype $\\mathbf{c}_c^t$ and source anchor $\\mathbf{c}_c^s$ to facilitate us in deriving instance-level decision boundaries. Formally, this metric is defined as:", + "bbox": [ + 75, + 137, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d (\\mathbf {a}, \\mathbf {b}) = 1. 0 - \\cos (\\mathbf {a}, \\mathbf {b}), \\\\ \\epsilon_ {i, c} ^ {t} = 1. 0 - \\exp \\left(d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {t}\\right) - 1. 0\\right), \\\\ \\epsilon_ {i, c} ^ {s} = \\exp \\left(- d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {s}\\right)\\right), \\tag {3} \\\\ \\epsilon_ {i, c} = \\sqrt {\\epsilon_ {i , c} ^ {t} \\cdot \\epsilon_ {i , c} ^ {s}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 218, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{i,c}^{t}$ and $\\epsilon_{i,c}^{s}$ denote the common score from target prototype and source anchor viewpoint, respectively. Both $\\epsilon_{i,c}^{t}$ and $\\epsilon_{i,c}^{s}$ are clipped to the [0, 1] before fusion. We then utilize geometric mean to calculate fused common score, due to its advantage in multi-criteria decision-making. Note that we leverage different formulas to calculate $\\epsilon_{i,c}^{t}$ and $\\epsilon_{i,c}^{s}$ . Under the same condition, i.e., $d(\\mathbf{z}_i^t,\\mathbf{c}_c^s) = d(\\mathbf{z}_i^t,\\mathbf{c}_c^t)$ , $\\epsilon_{i,c}^{s}$ is greater than $\\epsilon_{i,c}^{t}$ . The spirit behind is that instances exhibiting proximity to source anchors are likely to be common data. Conversely, instances in close proximity to target prototypes do not guarantee their recognition.", + "bbox": [ + 75, + 313, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the computed common score $\\epsilon_{i,c}$ and the estimated distribution for $\\| \\mathbf{z}_{i,unk}^t\\| _2$ we introduce the following pseudo-labeling strategy to effectively distinguish between common and private data instances:", + "bbox": [ + 75, + 479, + 468, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\rho_ {i, c} = \\mu_ {c} + \\epsilon_ {i, c} \\cdot (\\mu_ {p r i} - \\mu_ {c}), \\\\ \\tilde {\\mathbf {y}} _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\text {u n k n o w n}, & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} \\geq \\rho_ {i, c} \\\\ \\mathbb {1} (\\operatorname {a r g m a x} (\\epsilon_ {i})), & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} < \\rho_ {i, c} \\end{array} \\right. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 550, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{c}$ denotes the expectation of $\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}$ for the $c$ -th category, derived using the identical top- $K$ sampling technique employed during the construction of the target prototypes. $\\rho_{i,c}$ represents the instance-level decision boundary, which accounts for the potential covariate shifts across each category, as well as the data instances corresponding to both target prototypes and source anchors. $\\mathbb{1}$ denotes the one-hot encoding operator, $\\tilde{\\mathbf{y}}_i^t$ is the pseudo-label for $\\mathbf{x}_i^t$ .", + "bbox": [ + 75, + 622, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization Objectives", + "text_level": 1, + "bbox": [ + 76, + 755, + 302, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To encourage the separation between common and private data, the optimization process involves three distinct objectives, including pseudo-label learning, feature decomposition regularization, and feature consensus regularization.", + "bbox": [ + 75, + 780, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-label Learning. With the pseudo labels derived from Eq. 4, we apply the cross-entropy loss for model adaptation. Nevertheless, instead of assigning equal importance to all pseudo-labels, we introduce a Student's $t$ distribution", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a63a3c988cd5626ca6b60b3f9f8a9dfa676cb1af1dec014996b5469046b8633e.jpg", + "table_caption": [ + "Table 1. Details of class split. $\\mathcal{V},\\bar{\\mathcal{V}}_s$ , and $\\bar{\\mathcal{V}}_t$ signifies the common class, source-private class, and target-private class, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetClass Split(γ/γs/γt)
OPDAOSDAPDA
Office-31 [43]10/10/1110/0/1110/21/0
Office-Home [49]10/5/5025/0/4025/40/0
VisDA [36]6/3/36/0/66/6/0
DomainNet [37]150/50/145--
", + "bbox": [ + 511, + 118, + 874, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to characterize the certainty associated with each pseudolabel, taking into account the distance between $\\| \\mathbf{z}_{i,unk}^t\\| _2$ and $\\rho_{i,c}$ . Specifically,", + "bbox": [ + 498, + 213, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tau_ {i} ^ {t} \\propto 1 - \\left(1 + \\frac {\\left(\\rho_ {i , \\kappa} - \\| \\mathbf {z} _ {i , u n k} ^ {t} \\| _ {2}\\right) ^ {2}}{\\alpha}\\right) ^ {- \\frac {\\alpha + 1}{2}}, \\\\ \\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\sum_ {c = 1} ^ {C} \\tilde {y} _ {i, c} ^ {t} \\log \\delta_ {c} \\left(f _ {\\theta} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}\\right)\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 265, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau_{i}^{t}$ denotes the certainty, $\\kappa = \\mathrm{argmax}(\\epsilon_{i,c}),\\alpha = 1\\mathrm{e - }4$ by default. $\\delta_c(f_\\theta^t (\\mathbf{x}_i^t))$ is the $c$ -th soft-max probability for $\\mathbf{x}_i^t$ $\\tilde{y}_{i,c}^{t}$ corresponds to $c$ -th one-hot encoded pseudo-label for $\\tilde{\\mathbf{y}}_i^t$ . For those instances pseudo-labeled as private data, we refrain from introducing an additional $C + 1$ -th class, instead employing a uniform distribution to represent them.", + "bbox": [ + 498, + 367, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Decomposition Regularization. To promote the discriminability of $\\mathbf{z}_{i,unk}^{t}$ for identifying private data, we impose a feature decomposition regularizer. Concretely,", + "bbox": [ + 498, + 458, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p _ {i} = \\frac {\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right)}{\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right) + \\exp \\left(\\left\\| \\mathbf {z} _ {i , k n w} ^ {t} \\right\\| _ {2}\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {r e g} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\left(\\hat {y} _ {i} ^ {t} \\log \\left(p _ {i}\\right) + \\left(1 - \\hat {y} _ {i} ^ {t}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 510, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\hat{y}_i^t = 1$ or 0 means pseudo-labeled as private/common data.", + "bbox": [ + 498, + 603, + 885, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Consensus Regularization. Recent studies [41, 52, 53] in SFDA and SF-UniDA observe that incorporating consensus regularization with nearest neighbors in feature space could assist models in achieving more stable performance. Following them, we integrate the objective $\\mathcal{L}_{con}$ into LEAD. More details are included in the supplementary.", + "bbox": [ + 498, + 619, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Optimization Objective. Taking into account the discussions above, the eventual loss function is defined as:", + "bbox": [ + 498, + 709, + 890, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e g} + \\mathcal {L} _ {c o n} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 747, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda > 0$ is a trade-off hyper-parameter.", + "bbox": [ + 500, + 776, + 787, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 632, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 830, + 689, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We empirically verify the effectiveness of our LEAD on four datasets, including Office-31 [43], OfficeHome [49], VisDA [36], and DomainNet [37]. For a", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23338", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9a09f18cf3a35f75cf607e1133d29483859eeb71f049092cdfbc96ffca3ac348.jpg", + "table_caption": [ + "Table 2. H-score (\\%) comparison in OPDA scenario on Office-Home. U denotes methods applicable for all potential label-shift scenarios (i.e., OPDA, OSDA, and PDA). SF represents source data-free. CF indicates model adaptation without time-consuming K-means clustering." + ], + "table_footnote": [], + "table_body": "
MethodsUSFCFAr2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvg
CMU [15]✘✘✓56.056.959.267.064.367.854.751.166.468.257.969.761.6
DANCE [47]✓✘✓61.060.464.965.758.861.873.161.266.667.762.463.763.9
DCC [25]✓✘✘58.054.158.074.670.677.564.373.674.981.075.180.470.2
OVANet [44]✘✘✓62.875.678.670.768.875.071.358.680.576.164.178.971.8
GATE [7]✓✘✓63.875.981.474.072.179.874.770.382.779.171.581.775.6
UniOT [5]✘✘✓67.380.586.073.577.384.375.563.386.077.865.481.976.6
Source-only---47.371.681.951.557.269.456.040.376.661.444.273.560.9
SHOT-O [27]✘✓✓32.929.539.656.830.141.154.935.442.358.533.533.340.7
LEAD✓✓✓62.778.186.470.676.383.475.360.686.275.460.783.775.0
UMAD [28]✘✓✓61.176.382.770.767.775.764.455.776.373.260.477.270.1
+ LEAD✓✓✓65.282.687.478.076.587.579.865.088.479.463.083.578.0 (+7.9)
GLC [41]✓✓✘64.378.289.863.181.789.177.654.288.980.754.285.975.6
+ LEAD✓✓✘65.779.788.765.080.388.577.657.488.978.957.885.576.2 (+0.6)
", + "bbox": [ + 83, + 119, + 883, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0dab7245e0c03f88b06bfead93890c266a666a8af263b4e87d10ab7b7fbfbfdf.jpg", + "table_caption": [ + "Table 3. H-score (\\%) comparison in OPDA scenario on Office-31, VisDA, and DomainNet. Some results are cited from UMAD [28]." + ], + "table_footnote": [], + "table_body": "
MethodsUSFCFOffice-31DomainNetVisDA
A2DA2WD2AD2WW2AW2DAvgP2RP2SR2PR2SS2PS2RAvgS2R
CMU [15]XX✓68.167.371.479.372.280.473.150.845.152.245.644.851.048.332.9
DANCE [47]✓X✓78.671.579.991.472.287.980.321.037.047.346.727.721.033.542.8
DCC [25]✓XX88.578.570.279.375.988.680.256.943.750.343.344.956.249.243.0
OVANet [44]XX✓85.879.480.195.484.094.386.556.047.151.744.947.457.250.753.1
GATE [7]✓X✓87.781.684.294.883.494.187.657.448.752.847.649.556.352.156.4
UniOT [5]XX✓87.088.588.498.887.696.691.159.351.847.848.346.858.352.057.3
Source-only---70.963.239.677.352.286.464.957.338.247.838.432.248.243.725.7
SHOT-O [27]X✓✓73.567.259.388.377.184.475.035.030.837.228.331.932.232.644.0
LEAD✓✓✓85.485.086.390.986.293.187.859.946.151.345.045.956.350.876.6
UMAD [28]X✓✓79.177.487.490.790.497.287.059.044.350.142.132.055.347.158.3
+ LEAD✓✓✓88.789.290.895.890.497.192.0 (+5.0)59.346.049.843.843.956.049.8 (+2.7)67.2 (+8.9)
GLC [41]✓✓X81.584.589.890.488.492.387.863.350.554.950.949.661.355.173.1
+ LEAD✓✓X83.086.887.692.087.193.288.3 (+0.5)63.550.755.151.649.961.355.4 (+0.3)76.8 (+3.7)
", + "bbox": [ + 83, + 321, + 883, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fair comparison, we evaluate LEAD on partial-set domain adaptation (PDA), open-set domain adaptation (OSDA), and open-partial-set domain adaptation (OPDA). Details of classes split are summarized in Table 1.", + "bbox": [ + 75, + 518, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We utilize the same evaluation metric as previous works. In PDA scenarios, we report the classification accuracy for evaluation. In OSDA and OPDA scenarios, we report the H-score for performance comparison, which is a harmonic mean of accuracy over common data and accuracy over private data. It is worth noting that LEAD can be incorporated effortlessly into most existing SF-UniDA methods, for potential performance improvement. To demonstrate this merit, we integrate LEAD with two representative methods: UMAD [28] and GLC [41].", + "bbox": [ + 75, + 579, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. All the experiments are conducted on a single RTX-3090 GPU with PyTorch-1.10. We adopt the same network architecture as existing baseline methods. During target model adaptation, we apply the SGD optimizer with momentum 0.9. The batch size is set to 64 for all experiments. The learning rate is set to 1e-3 for Office-31 and Office-Home, and 1e-4 for VisDA and DomainNet. As for $\\lambda$ , it is set to 0.3 for Office-31, 1.0 for VisDA, and 2.0 for Office-Home and DomainNet. During inference, we adopt the same protocol like [41]. More details are presented in the supplementary.", + "bbox": [ + 75, + 729, + 470, + 896 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 518, + 700, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results for OPDA Scenarios. To verify the effectiveness of LEAD, we first conduct experiments in the most challenging OPDA scenarios, where both source and target domains contain private data. Table 2 summarizes the results on Office-Home, and Table 3 details the results on Office-31, DomainNet, and VisDA. These results demonstrate that LEAD achieves performance surpassing or comparable to existing methods based on K-means clustering for segregating common and private data. In comparison with methods without clustering, LEAD achieves significantly superior performance. For instance, on Office-Home, DomainNet, and VisDA datasets, LEAD outperforms UMAD by $4.9\\%$ , $3.7\\%$ , and $18.3\\%$ , respectively. We attribute this to our instance-level adaptive decision strategy, which effectively circumvents the limitations imposed by a fixed threshold. While GLC achieves commendable performance on its own, the integration with LEAD further amplifies its effectiveness. For example, LEAD improves GLC in the H-score on VisDA from $73.1\\%$ to $76.8\\%$ . The integration with LEAD proves even more significant against UMAD, with the H-score on Office-Home and VisDA boosting from $70.1\\% /58.3\\%$ to $78.0\\% /67.2\\%$ , respectively.", + "bbox": [ + 496, + 537, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results for OSDA Scenarios. We then conduct experiments in OSDA scenarios, where only the target domain", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23339", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47886e4aedf11d69aff94161c0fff93be7c4ce632a071a6da120faf703753b01.jpg", + "table_caption": [ + "Table 4. H-score (%) comparison in OSDA scenario on Office-Home, Office-31, and VisDA." + ], + "table_footnote": [], + "table_body": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X Y55.057.059.059.358.260.659.251.361.261.953.555.357.652.655.776.575.965.864.765.254.2
DANCE [47]✓ X Y6.59.09.920.410.49.228.412.812.614.27.913.212.984.978.879.178.868.378.879.867.5
DCC [25]✓ X X56.167.566.749.666.564.055.853.070.561.657.271.961.758.354.867.289.485.380.972.759.6
OVANet [44]X X Y58.666.369.962.065.268.659.853.469.368.759.666.764.090.588.386.798.288.398.491.766.1
GATE [7]✓ X Y63.870.575.866.467.971.767.361.576.070.461.875.169.088.486.584.295.086.196.789.570.8
Source-only- - -46.163.372.942.854.058.747.836.166.260.845.368.255.278.272.144.282.252.188.869.629.1
SHOT-O [27]X Y Y37.741.848.456.439.840.960.041.549.761.841.443.646.980.271.664.393.164.091.877.528.1
LEAD✓ Y Y60.770.876.561.068.670.865.559.874.264.857.775.867.284.985.190.294.890.396.590.374.2
UMAD [28]X Y Y59.271.876.663.569.071.962.554.672.866.557.970.766.488.584.486.895.088.295.989.866.8
+ LEAD✓ Y Y62.673.677.361.366.270.763.354.675.166.763.575.667.690.489.690.395.490.797.392.370.2
+2.5
", + "bbox": [ + 83, + 106, + 885, + 308 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9799ab231a02b84e086c5cdd46936f9c56a9f8bfdc9ebbfabed18fcc72a0edbf.jpg", + "table_caption": [ + "Table 5. Accuracy comparison (%) in PDA scenario on Office-Home, Office-31, and VisDA." + ], + "table_footnote": [], + "table_body": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X X50.974.278.462.264.172.563.547.978.372.454.778.966.584.184.269.297.266.898.883.465.5
DANCE [47]✓ X X53.673.284.970.867.382.670.050.984.877.055.981.871.177.171.283.794.692.696.886.073.7
DCC [25]✓ X X54.247.557.583.871.686.263.765.075.285.578.282.670.987.381.395.4100.095.5100.093.372.4
OVANet [44]X X X34.154.672.142.447.355.938.226.261.756.735.868.949.569.461.761.490.266.498.774.634.3
GATE [7]✓ X X55.875.985.373.670.283.072.159.584.779.663.983.874.089.586.293.5100.094.498.693.775.6
Source-only- - -45.969.281.155.761.264.860.741.175.870.549.978.462.990.479.379.395.984.398.187.842.8
SHOT-P [27]X X Y64.785.190.175.173.984.276.464.190.380.763.385.577.889.884.492.296.692.299.492.474.2
LEAD✓ Y Y58.283.187.070.575.483.373.750.483.778.358.783.273.889.893.995.698.696.099.495.575.3
UMAD [28]X Y Y51.266.579.263.162.968.263.356.475.974.555.978.366.385.485.183.597.686.299.489.568.5
+ LEAD✓ Y Y61.178.287.570.578.581.874.160.483.378.863.282.175.089.290.890.698.093.0100.093.678.1
+8.7+4.1+9.6
GLC [41]✓ Y X55.979.087.572.571.882.774.941.782.477.360.484.372.589.889.892.896.696.199.494.176.2
+ LEAD✓ Y X56.681.288.172.572.283.275.343.683.478.259.784.273.289.893.295.696.996.099.495.175.5
+0.7+1.0-0.7
", + "bbox": [ + 83, + 329, + 885, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "involves private data. Table 4 lists the results, demonstrating the superiority of LEAD against baselines. Specifically, LEAD obtains the H-score of $67.2\\%$ , $90.3\\%$ , and $74.2\\%$ on Office-Home, Office-31, and VisDA datasets, respectively. Similar to the observations in OPDA scenarios, LEAD also exhibits performance improvement to existing methods. LEAD improves the H-score of UMAD by $1.2\\%$ , $2.5\\%$ , and $3.4\\%$ across the Office-Home, Office-31, and VisDA datasets, respectively. The results again verify the merit of LEAD when incorporated with existing methods.", + "bbox": [ + 75, + 537, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results for PDA Scenarios. Finally, we verify the effectiveness of LEAD in PDA scenarios, where the target domain's label sets are subsets of those in the source domain. Results presented in Table 5 show that LEAD achieves superior performance even compared to methods tailored for PDA. In particular, LEAD attains the overall accuracy of $73.8\\%$ , $95.5\\%$ , and $75.3\\%$ on Office-Home, Office-31, and VisDA datasets, individually. LEAD contributes the accuracy increase of $8.7\\%$ , $4.1\\%$ , and $9.6\\%$ to UMAD on these datasets. Note that, there is a slight degradation $(-0.7\\%)$ for GLC w/ LEAD. It may arise from gradient conflict, causing the optimization process to inadvertently compromise one objective in favor of improving the overall objective. We will pursue better integration strategies in future work.", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c90a4677647c45b282f23ca537a6582cb00335312a1c7e6fb53dc819b5cbce01.jpg", + "table_caption": [ + "Table 6. Ablation studies of LEAD in OPDA scenarios." + ], + "table_footnote": [], + "table_body": "
LceLregLconOffice-31Office-HomeVisDA
---64.960.925.7
✓--82.073.661.5
✓✓-84.574.567.3
✓-✓87.274.575.7
✓✓✓87.875.076.6
", + "bbox": [ + 504, + 551, + 883, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Experimental Analysis", + "text_level": 1, + "bbox": [ + 500, + 643, + 710, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study. To assess the contribution of different components within LEAD, we carry out extensive ablation studies in OPDA scenarios. The results, detailed in Table 6, confirm that each component is not only effective but also complementary to the others.", + "bbox": [ + 496, + 659, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficiency Comparison. As previously stated, LEAD employs the distance to target prototypes and source anchors for adaptive pseudo-labeling. This design, in contrast to clustering-based methods, notably reduces the need for computational resources. Table 7 juxtaposes the runtime of GLC [41] and LEAD in OPDA scenarios. It is clear that LEAD requires minimal resources for deriving decision boundaries, whereas GLC's resource demands escalate dramatically with data scale. In particular, on DomainNet, with about 0.6 million images across 345 categories, GLC averages 897.65 seconds, while LEAD takes only 0.29 sec", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/272dc3cb2d8c736b11cf92868aefcd5a1556bce19c89b588cbf4274e63087ebf.jpg", + "image_caption": [ + "(a) Sensitivity to $\\lambda$" + ], + "image_footnote": [], + "bbox": [ + 117, + 84, + 302, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/039a3a74b32629a546ea44434469cf1ba504d2e9fc48a5d3968bb893703994ba.jpg", + "image_caption": [ + "(b) H-score convergence" + ], + "image_footnote": [], + "bbox": [ + 305, + 85, + 486, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5adf2f4b6957ee926f84826bf54253e771ca8d412b05a8741ebed93b861bb315.jpg", + "image_caption": [ + "(c) $\\mathrm{Ar}\\rightarrow \\mathrm{Pr}$ in OPDA" + ], + "image_footnote": [], + "bbox": [ + 488, + 85, + 665, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3f8ae72befbceebaf4610de0e38520614dac47bcf9abf0b8500ba7a595ff590.jpg", + "image_caption": [ + "(d) $\\mathrm{Re}\\rightarrow \\mathrm{Cl}$ in OPDA" + ], + "image_footnote": [], + "bbox": [ + 666, + 85, + 846, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ebce39355cfec23b6551945fdd5ff2ce164625d4882035dd7c0bc7fc97ab209.jpg", + "table_caption": [ + "Table 7. Runtime (s) to derive pseudo-label decision boundaries." + ], + "table_footnote": [], + "table_body": "
Office-31Office-HomeVisDADomainNet
GLC [41]0.190.510.46897.65
LEAD0.020.020.110.29
", + "bbox": [ + 81, + 270, + 460, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "onds on the same platform. Moreover, LEAD also shows substantial resource savings even on normal-scale datasets, e.g., on VisDA with over $75\\%$ reduction.", + "bbox": [ + 75, + 319, + 467, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter Sensitivity. We investigate the parameter sensitivity of $\\lambda$ on Office-31 in OPDA scenarios, where the range of $\\lambda$ is set to [0.1, 0.2, 0.3, 0.4, 0.5]. Figure 3 (a) illustrates that the results are stable around the selected parameter $\\lambda = 0.3$ . Through oracle validation, we may find the optimal parameter configuration for this setting is $\\lambda = 0.1$ .", + "bbox": [ + 75, + 364, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Stability. Figure 3 (b) visualizes the training curves of our LEAD on VisDA in OPDA scenarios. Generally, the training procedure of LEAD is stable and effective.", + "bbox": [ + 75, + 457, + 468, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying Unknown Classes. With the rise in unknown private classes, accurately distinguishing common from private data becomes more challenging. To examine the robustness of our LEAD, we compared it with GLC and UMAD in OPDA scenarios using the Office-Home dataset. The results in Figure 3 (c-d) show that LEAD achieves stable results and consistently boosts existing methods.", + "bbox": [ + 75, + 503, + 468, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Instance-level Decision Boundary. To evaluate the contribution of our instance-level pseudo-labeling strategy, we conduct a comparison against the intuitive design, which uses the average of $\\mu_{\\text{com}}$ and $\\mu_{\\text{pri}}$ as the threshold. As depicted in Figure 4 (a), instance-level decision boundary strategy significantly outperforms the vanilla global decision design (75.0% vs 63.1% on Office-Home in OPDA scenarios). We attribute this to that vanilla design discounts the inconsistency of covariate shifts across categories, and also the variability among samples from the same category.", + "bbox": [ + 75, + 612, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploration with Entropy as the Indicator. Contrasting with existing threshold-based methods that utilize Entropy as the indicator for differentiating common and private data, we propose decomposing the feature and using the projection $\\mathbf{z}_{i,unk}^{t}$ on the source-unknown space as an alternative indicator. To empirically substantiate the superiority, we remold our framework to incorporate Entropy as the indicator. Figure 4 (b) depicts the results. An observation is that while employing entropy as the indicator within our framework is", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f2ae93e6a3123e54df700f932bcee05fe553c6c6eeed04979dfea9a7f441902.jpg", + "image_caption": [ + "Figure 3. Robustness analysis. (a) shows the sensitivity to $\\lambda$ on Office-31. (b) presents the H-score curves on VisDA. (c-d) present robustness analysis when varying the unknown private categories.", + "(a) Effect of Instance-level Decision Boundary", + "Figure 4. Methodology analysis. (a) compares the effectiveness of our instance-level decision strategy against the vanilla global decision strategy. (b) examines the efficacy of using Entropy as the indicator for private data." + ], + "image_footnote": [], + "bbox": [ + 519, + 244, + 694, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/862ba535973c6ca976f54376d3e3fe936843dd25a2033bfb58ea613f8a1c298d.jpg", + "image_caption": [ + "(b) Exploration with Entropy as the Indicator" + ], + "image_footnote": [], + "bbox": [ + 694, + 244, + 864, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feasible, its performance is inferior compared to $\\mathbf{z}_{i,unk}^{t}$ . We speculate that this inferiority might result from entropy's limited capacity to distinguish between common samples with high uncertainty and private unknown samples.", + "bbox": [ + 496, + 422, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 496, + 617, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we delve into source-free universal domain adaptation (SF-UniDA). Different from existing methods that leverage hand-crafted thresholding strategies or feature-clustering algorithms, we propose a novel idea of LEArning Decomposition (LEAD) tailored for SF-UniDA. Technically, LEAD first decomposes features into source-known and -unknown two parts, then establishes instance-level decision boundaries for pseudo-labeling. It offers an elegant view to identify target-private unknown data without tedious tuning a threshold or relying on unstable clustering. Extensive experiments across various scenarios have verified the effectiveness of our LEAD. Besides, it is appealing in view that LEAD can be seamlessly incorporated into existing methods to further boost performance. Remarkably, LEAD improves UMAC by $7.9\\%$ overall H-score in the OPDA scenario on Office-Home dataset.", + "bbox": [ + 496, + 523, + 890, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No. 62372329), in part by the National Key Research and Development Program of China (No. 2021YFB2501104), in part by Shanghai Rising Star Program (No.21QC1400900), in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program, and in part by the International Exchange Program for Graduate Students, Tongji University (No.2023020053).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018. 4", + "[2] Richard Bellman. Dynamic programming. Science, 153 (3731):34-37, 1966. 2", + "[3] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 1, 2", + "[4] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 1, 2", + "[5] Wanxing Chang, Ye Shi, Hoang Tuan, and Jingya Wang. Unified optimal transport framework for universal domain adaptation. In NeurIPS, 2022. 2, 6", + "[6] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In ECCV. Springer, 2020. 2", + "[7] Liang Chen, Yihang Lou, Jianzhong He, Tao Bai, and Minghua Deng. Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation. In CVPR, 2022. 1, 2, 6, 7", + "[8] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, 2018. 2", + "[9] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In ICCV, 2017. 1", + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1", + "[11] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE TPAMI, 39(9):1853-1865, 2016. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1", + "[13] Qi Dou, Cheng Ouyang, Cheng Chen, Hao Chen, and Pheng-Ann Heng. Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss. In IJCAI, 2018. 1", + "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE TIP, 30:6473-6484, 2021. 2", + "[15] Bo Fu, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Learning to detect open classes for universal domain adaptation. In ECCV. Springer, 2020. 2, 6, 7", + "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, pages 2096-2030, 2016. 1, 2" + ], + "bbox": [ + 78, + 114, + 467, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1", + "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2", + "[20] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML. PMLR, 2018. 1", + "[21] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1", + "[23] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE TPAMI, pages 766-785, 2019. 1, 2", + "[24] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 2", + "[25] Guangrui Li, Guoliang Kang, Yi Zhu, Yunchao Wei, and Yi Yang. Domain consensus clustering for universal domain adaptation. In CVPR, 2021. 1, 2, 3, 6, 7", + "[26] Shuang Li, Shiji Song, Gao Huang, Zhengming Ding, and Cheng Wu. Domain invariant and class discriminative feature learning for visual domain adaptation. IEEE TIP, 2018. 2", + "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML. PMLR, 2020. 1, 2, 3, 6, 7", + "[28] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 1, 2, 3, 6, 7", + "[29] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2023. 4", + "[30] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE TMM, 2022. 2", + "[31] Xinyu Liu and Yixuan Yuan. A source-free domain adaptive polyp detection framework with style diversification flow. IEEE TMI, pages 1897-1908, 2022. 1", + "[32] Zhengfa Liu, Guang Chen, Zhijun Li, Yu Kang, Sanqing Qu, and Changjun Jiang. Psdc: A prototype-based shared-dummy classifier model for open-set domain adaptation. IEEE Transactions on Cybernetics, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23342", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Zhengfa Liu, Guang Chen, Zhijun Li, Sanqing Qu, Alois Knoll, and Changjun Jiang. D2ifln: Disentangled domain-invariant feature learning networks for domain generalization. IEEE Transactions on Cognitive and Developmental Systems, 2023. 2", + "[34] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In ICCV, 2013. 2", + "[35] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 1, 2", + "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.5", + "[37] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 5", + "[38] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML. PMLR, 2020. 2", + "[39] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 2", + "[40] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In CVPR, 2023. 2", + "[41] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1", + "[43] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5", + "[44] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In ICCV, 2021. 1, 2, 6, 7", + "[45] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1", + "[46] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 1, 2", + "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. In NeurIPS, pages 16282-16292, 2020. 1, 2, 3, 6, 7", + "[48] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In WACV, 2021. 2", + "[49] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5", + "[50] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 4", + "[52] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 5", + "[53] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 5", + "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In CVPR, 2019. 1, 2" + ], + "bbox": [ + 501, + 90, + 890, + 273 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_model.json b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..07167b7a01db54e539304612f9791a71acd1fe44 --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "LEAD: Learning Decomposition for Source-free Universal Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.18, + 0.735, + 0.218 + ], + "angle": 0, + "content": "Sanqing \\(\\mathrm{Qu}^1\\), Tianpei \\(\\mathrm{Zou}^1\\), Lianghua \\(\\mathrm{He}^1\\), Florian Röhrbein\\(^2\\), Alois Knoll\\(^3\\), Guang Chen\\(^{1*}\\), Changjun Jiang\\(^1\\)" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.219, + 0.756, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Tongji University, \\(^{2}\\)Chemnitz University of Technology, \\(^{3}\\)Technical University of Munich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.296, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Universal Domain Adaptation (UniDA) targets knowledge transfer in the presence of both covariate and label shifts. Recently, Source-free Universal Domain Adaptation (SF-UniDA) has emerged to achieve UniDA without access to source data, which tends to be more practical due to data protection policies. The main challenges lie in determining whether covariate-shifted samples belong to target-private unknown categories. Existing methods tackle this either through hand-crafted thresholding or by developing time-consuming iterative clustering strategies. In this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. Technically, LEAD initially leverages the orthogonal decomposition analysis for feature decomposition. Then, LEAD builds instance-level decision boundaries to adaptively identify target-private data. Extensive experiments across various UniDA scenarios have demonstrated the effectiveness and superiority of LEAD. Notably, in the OPDA scenario on VisDA dataset, LEAD outperforms GLC by \\(3.5\\%\\) overall H-score and reduces \\(75\\%\\) time to derive pseudo-labeling decision boundaries. Besides, LEAD is also appealing in that it is complementary to most existing methods. The code is available at https://github.com/ispc-lab/LEAD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.703 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.713, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have yielded impressive results in a variety of computer vision tasks [12, 18, 22, 42]. However, DNNs often fail to generalize well to new domains due to the discrepancy between training (source) and test (target) data distributions (i.e., covariate shift). Such a shift poses significant challenges for some safety-critical applications like autonomous driving [9, 10], or medical imaging [13, 31]. Unsupervised Domain Adaptation (DA) [16, 23] aims to address this challenge by transferring task-specific knowledge from a labeled source domain to an unlabeled target domain. Despite promising results, most" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.261, + 0.703, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.261, + 0.89, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.542 + ], + "angle": 0, + "content": "Figure 1. (a) Pseudo-label accuracy curves for target-private data in the OPDA scenario on VisDA dataset. An observation is that clustering-based GLC does not work well in deriving decision boundaries for target-private data. This may be due to the curse of dimensionality, leading K-means to faintly discerning clusters. (b) Frequency distribution for the normalized feature magnitude of target data in the source-unknown space (orthogonal complement of the space spanned by source model weights). Features are on task \\((A\\to D)\\) in the OPDA scenario of Office-31. The results show that target-private data are expected to involve more components from source-unknown space, even with covariate shifts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.611 + ], + "angle": 0, + "content": "existing favorable DA methods [16, 20, 27, 45] assume that the label spaces are identical across the source and target domain, thus being only applicable to closed-set scenarios and limiting applicability to practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.896, + 0.852 + ], + "angle": 0, + "content": "To handle generalized cases, Universal Domain Adaptation (UniDA) [47] was proposed to allow all potential label shifts between source and target domains. Different from existing methods specialized for Partial Domain Adaptation (PDA) [3, 4], Open-set Domain Adaptation (OSDA) [35, 46], or Open-Partial Domain Adaptation (OPDA) [44, 54], in UniDA, we have no prior knowledge about the label shift, e.g., information on matched common classes or the number of categories in the target domain. Nonetheless, most existing UniDA works [7, 25, 47] necessitate concurrent access to source and target data. This requirement becomes increasingly impractical in light of strict data protection policies [50]. In this paper, we focus on Source-free Universal Domain Adaptation (SF-UniDA) [28, 41], where only a pre-trained source model is provided for knowledge transfer rather than labeled source data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The primary objective for UniDA is to recognize target data belonging to common categories and separate them from target-private \"unknown\" data. Existing meth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.365, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: guangchen@tongji.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.302 + ], + "angle": 0, + "content": "ods have designed various criteria based on instance-level predictions, e.g., entropy-based [14, 47, 54], confidence-based [15, 24, 44]. However, defining a universal threshold for all scenarios is challenging and often sub-optimal. To mitigate tedious parameter-tuning, some methods start to perform source and target consensus clustering to achieve common and private data separation [7, 25]. Considering that SF-UniDA is inaccessible to source data, GLC [41] devises an iterative one-vs-all clustering strategy to realize target-private data identification. Regardless of progress, iterative clustering can be computationally intensive. Besides, due to the curse of dimensionality [2], methods like K-means are prone to overfitting and faintly discerning clusters. This can be implied to some degree from Figure 1 (a)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.304, + 0.468, + 0.664 + ], + "angle": 0, + "content": "In this paper, we propose to address the SF-UniDA problem from a new perspective, i.e., feature decomposition. Our launching point is that, even though target data might have shifted in the feature space, target-private data from unknown categories are likely to exhibit a higher proportion of components from the orthogonal complement (source-unknown) space spanned by source model weights. The frequency distribution of normalized feature magnitudes in the source-unknown space depicted in Figure 1 (b) basically demonstrates our argument, which has been rarely explored in the literature. To materialize our idea, we propose the LEArning Decomposition (LEAD) framework. In particular, LEAD initially leverages the orthogonal decomposition to construct source-known and -unknown space. The projection on source-unknown space is then regarded as the discriminative representation with respect to target-private categories. Different from existing thresholding methods introducing a fixed global hand-crafted threshold for all scenarios, LEAD deliberately considers the distance to the target prototypes and source anchors to establish instance-level decision boundaries. We analyze and evaluate our LEAD under a variety of SF-UniDA scenarios, ranging from PDA, OSDA, to OPDA. Extensive experiments demonstrate the effectiveness and superiority." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.666, + 0.468, + 0.83 + ], + "angle": 0, + "content": "The main contribution is the proposal of a LEArning Decomposition (LEAD) framework for source-free universal domain adaptation (SF-UniDA). This solution leads to elegant views for identifying target-private unknown data without tedious tuning thresholds or relying on iterative unstable clustering. Remarkably, in the OPDA scenario on VisDA, LEAD attains an H-score of \\(76.6\\%\\), surpassing GLC [41] by \\(3.5\\%\\). Besides, LEAD is complementary to most existing SF-UniDA methods. For instance, in the OPDA scenario on Office-Home, LEAD advances UMAD [28] with H-score improvement from \\(70.1\\%\\) to \\(78.0\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.846, + 0.217, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.872, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Unsupervised Domain Adaptation. When DNNs encounter data from a distribution that differs from their train" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.272 + ], + "angle": 0, + "content": "ing data, inevitable performance degeneration is commonly observed. To address this issue, domain adaptation (DA) [16, 23] has been introduced, leveraging labeled data from source domains to train DNNs for unlabeled target domains in a transductive learning manner. Existing DA methods can be broadly categorized into two paradigms, namely moment matching [11, 34, 48] and adversarial learning [16, 19, 21]. Regardless of their effectiveness in various applications such as object recognition [16, 27, 39], semantic segmentation [19, 48], and object detection [8, 21], most existing approaches assume that label spaces are identical across the source and target domains, limiting their applicability." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.278, + 0.89, + 0.669 + ], + "angle": 0, + "content": "Universal Domain Adaptation. There have been some approaches introduced to deal with label-shift scenarios, such as open-set domain adaptation (OSDA) [32, 35, 46], partial-set domain adaptation (PDA) [3, 4], and open-partial-set domain adaptation (OPDA) [15, 44, 54]. However, these methods are often tailored to specific scenarios and may not be readily applicable to other label shift situations. Universal domain adaptation (UniDA), on the other hand, is designed to facilitate adaptation across all potential label shift scenarios. Despite the progress in UniDA, many existing methods [5, 7, 44] necessitate simultaneous access to both source and target data, making them impractical for applications with stringent data protection policies. To address this challenge, source-free universal domain adaptation (SF-UniDA) [28, 41] has been proposed, wherein source data are only utilized for pre-training and are inaccessible during target adaptation. To promote the distinction between common and private data, existing methods often resort to manually defined thresholding or the introduction of global clustering techniques. Nevertheless, selecting an appropriate threshold for all scenarios can be both laborious and sub-optimal. Additionally, K-means clustering tends to be unstable when applied to high-dimensional data due to the curse of dimensionality [2]. Our work delves into these limitations and proposes a novel and elegant solution from the perspective of feature decomposition." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.675, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Feature Decomposition. As a fundamental technique in machine learning, feature decomposition is applied to deconstruct complex data into simpler and more interpretable components. In the field of domain adaptation and domain generalization, feature decomposition has been leveraged in various studies [6, 26, 30, 33, 38, 40] to enhance feature alignment or facilitate domain-invariant feature learning by breaking down features into content and style components. Nonetheless, the content and style decomposition design is less applicable for UniDA, as the primary challenge lies in distinguishing common and private data. In this paper, we approach the problem from the perspective of orthogonal feature decomposition and employ Independent Component Analysis to decompose features into two orthogonal parts, thereby facilitating the intended objective." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23335" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.332, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.089, + 0.868, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.376, + 0.892, + 0.57 + ], + "angle": 0, + "content": "Figure 2. (a) Illustrations of traditional universal domain adaptation (UniDA) and source-free universal domain adaptation (SF-UniDA). Traditional UniDA methods necessitate data from both source and target domains concurrently. In SF-UniDA, source data are solely utilized for pre-training. Adaptation is performed by harnessing target data and the source model \\( f_{\\theta}^{s} = h_{\\theta}^{s} \\circ g_{\\theta}^{s} \\). (b) An overview of our Learning Decomposition (LEAD) framework. Pseudo-labeling is an important technique for UniDA and SF-UniDA. The primary objective is to recognize target data associated with common label sets and exclude data within the target-private label space. Different from existing methods that perform private data identification by hand-crafted thresholding on predictions or iterative global clustering, we tackle this from the viewpoint of feature decomposition. The rationale is that despite potential shifts in the feature space, target-private data are expected to encompass more components from the orthogonal complement (source-unknown) space of the source model. Technically, LEAD first performs orthogonal decomposition to decompose target features into source-known and -unknown parts, i.e., \\( \\mathbf{z}_{i,knw}^{t} \\) and \\( \\mathbf{z}_{i,unk}^{t} \\). \\( \\| \\mathbf{z}_{i,unk}^{t}\\|_{2} \\) is considered as an indicator for private data. Next, LEAD employs a two-component Gaussian Mixture Model to estimate the distribution of \\( \\| \\mathbf{z}_{i,unk}^{t}\\|_{2} \\). Thereafter, LEAD devises a metric named \"common score\" \\( \\epsilon_{i,c} \\) that accounts for distances to both target prototypes and source anchors (derived from \\( h_{\\theta}^{t} \\)) to facilitate deriving instance-level decision boundary \\( \\rho_{i,c} \\). LEAD provides an elegant solution to distinguish target-private data, mitigating the need for tedious hand-crafted threshold tuning or dependence on time-consuming iterative clustering. LEAD could also serve as a complementary approach to most existing SF-UniDA methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.212, + 0.597 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.206, + 0.621 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.839 + ], + "angle": 0, + "content": "In this paper, we consider a generalized and challenging case in domain adaptation: universal domain adaptation (UniDA), which aims to achieve knowledge transfer in the presence of both covariate and label shifts. In UniDA, there is a labeled source domain \\(\\mathcal{D}^s = \\left\\{\\left(\\mathbf{x}_i^s,\\mathbf{y}_i^s\\right)\\right\\}_{i = 1}^{N_s}\\) where \\(\\mathbf{x}_i^s\\in \\mathcal{X}^s\\subset \\mathbb{R}^X,\\mathbf{y}_i^s\\in \\mathcal{Y}^s\\subset \\mathbb{R}^C\\) , and an unlabeled target domain \\(\\mathcal{D}^t = \\left\\{\\left(\\mathbf{x}_i^t,\\ref{eq:2.3.1}\\right)\\right\\}_{i = 1}^{N_t}\\) where \\(\\mathbf{x}_i^t\\in \\mathcal{X}^t\\subset \\mathbb{R}^X.\\) \\(\\mathcal{V} = \\mathcal{V}^s\\cap \\mathcal{V}^t\\) denotes the common label sets shared by source and target domains. \\(\\bar{\\mathcal{V}}^s = \\mathcal{V}\\setminus \\mathcal{V}^t\\) and \\(\\bar{\\mathcal{V}}^t = \\mathcal{V}\\setminus \\mathcal{V}^s\\) symbolizes the label sets private to source and target domain, individually. UniDA supposes that we have no prior knowledge about \\(\\mathcal{V}^t\\) . \\(\\mathcal{V}\\) and \\(\\bar{\\mathcal{V}}^t\\) are also unavailable. The objective is to recognize common data belonging to \\(\\mathcal{V}\\) and reject private data in \\(\\bar{\\mathcal{V}}^t\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Different from traditional UniDA methods that need concurrent access to source and target data, under the source-free universal domain adaptation (SF-UniDA) setting, \\( \\mathcal{D}_s \\) is only applicable for source model pre-training and not acces-" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.892, + 0.687 + ], + "angle": 0, + "content": "sible during adaptation. Assuming that the source model is represented as \\( f_{\\theta}^{s} = h_{\\theta}^{s}\\circ g_{\\theta}^{s} \\), where \\( g_{\\theta}^{s}:\\mathbb{R}^{X}\\to \\mathbb{R}^{D} \\) denotes the feature extractor and \\( h_\\theta^s:\\mathbb{R}^D\\to \\mathbb{R}^C \\) is the classifier module. Following existing methods [27, 41], we freeze the classifier module \\( h_\\theta^t = h_\\theta^s \\) and only learn a target-specific feature extractor \\( g_{\\theta}^{t} \\) to realize model adaptation. Figure 2 (a) presents an illustration for UniDA and SF-UniDA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The main challenge is how to distinguish common data from private data. Prior methods [25, 28, 41, 47] have designed various algorithms to encourage the separation. However, these methods perform data identification either based on hand-crafted thresholding criteria or by introducing time-consuming unstable clustering strategies. We propose to mitigate these limitations from the perspective of feature decomposition. An observation is that even though target data may have shifted in the feature space, the features of private data still include more components from the orthogonal complement (source-unknown) space of the pre-trained model. Technically, we present a novel learning decomposition (LEAD) framework. LEAD capitalizes on the orthogonal decomposition to build two orthogonal fea" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23336" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "ture spaces, i.e., source-known and -unknown space. Feature projection on source-unknown space is extracted as the descriptor for private data. Then, LEAD considers the distances to both the target prototypes and source anchors to establish instance-level decision boundaries. Figure 2 (b) manifests the architecture of our LEAD." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.192, + 0.387, + 0.209 + ], + "angle": 0, + "content": "3.2. Orthogonal Feature Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Pseudo-labeling serves as an important technique in unsupervised learning, yet many existing strategies only account for closed-set scenarios, disregarding the label-shift situations, which impairs their ability to differentiate target-private data. To fulfill the objective of common and private data separation, we propose to orthogonally decompose the feature representation into two uncorrelated parts, i.e., features associated with source-known space \\( F_{knw} \\) and those related to source-unknown space \\( F_{unk} \\). This design stems from a simple inductive bias in favor of private data identification. Due to the lack of exposure to target data during source pre-training, particularly target-private data, the source-known space does not include any significant attributes of target-private data. Subsequently, after normalization, target-private data are expected to contain more components from source-unknown space, even with covariate shifts. This hypothesis is basically demonstrated by empirical evidence, as illustrated in Figure 1 (b)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Formally, denoting the weight of classifier \\( h_{\\theta}^{t} \\) is \\( W_{cls} \\in \\mathbb{R}^{C \\times D} \\). Source-known space \\( F_{knw} \\) can be represented as the vector space of span \\( \\{\\mathbf{w}_n\\} \\), where \\( \\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\} \\) is a group weight vectors of \\( W_{cls} \\). Source-unknown space \\( F_{unk} \\) is the orthogonal complement space of \\( F_{knw} \\), i.e., \\( \\mathbb{R}^D = F_{knw} \\oplus F_{unk} \\). Nevertheless, unit orthogonality in \\( \\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\} \\) is not guaranteed (i.e., when \\( n \\neq m \\) (\\( \\mathbf{w}_n, \\mathbf{w}_m \\neq 0 \\)), making it hard to derive \\( F_{unk} \\). Considering that \\( F_{knw} \\) and \\( F_{unk} \\) are the row space and null space of \\( W_{cls} \\). Therefore, we employ the singular value decomposition (SVD) on \\( W_{cls} \\) to obtain \\( F_{knw} \\) and \\( F_{unk} \\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.665, + 0.284, + 0.682 + ], + "angle": 0, + "content": "\\[\nW _ {c l s} = U \\Sigma V ^ {T},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.685, + 0.469, + 0.702 + ], + "angle": 0, + "content": "\\[\nF _ {k n w} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = 1, \\dots , C} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.705, + 0.378, + 0.72 + ], + "angle": 0, + "content": "\\[\nF _ {u n k} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = C + 1, \\dots , D} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.723, + 0.469, + 0.771 + ], + "angle": 0, + "content": "where \\(\\Sigma \\in \\mathbb{R}^{C\\times D}\\) is a diagonal matrix, \\(U\\in \\mathbb{R}^{C\\times C}\\) and \\(V\\in \\mathbb{R}^{D\\times D}\\) are both orthogonal unitary matrices. \\(\\{\\mathbf{v}_n|\\mathbf{v}_n\\in \\mathbb{R}^D\\}_{n = 1}^D\\) are columns of \\(V\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.47, + 0.832 + ], + "angle": 0, + "content": "Assuming that \\(\\mathbf{z}_i^t\\) is the feature of target data \\(\\mathbf{x}_i^t\\) extracted from feature extractor \\(g_{\\theta}^{t}\\) after normalization, i.e., \\(\\mathbf{z}_i^t = g_\\theta^t (\\mathbf{x}_i^t) / ||g_\\theta^t (\\mathbf{x}_i^t)||_2\\), then \\(\\mathbf{z}_i^t\\) can be represented as the weighted summation of two orthogonal bases:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.838, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} ^ {t} = \\underbrace {\\sum_ {n = 1} ^ {C} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, k n w} ^ {t}} + \\underbrace {\\sum_ {n = C + 1} ^ {D} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, u n k} ^ {t}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.186 + ], + "angle": 0, + "content": "where \\(||\\mathbf{z}_{i,knw}^{t}||_{2}^{2} + ||\\mathbf{z}_{i,unk}^{t}||_{2}^{2} = 1\\), \\(\\mathbf{z}_{i,knw}^{t}\\) and \\(\\mathbf{z}_{i,unk}^{t}\\) are the projections of \\(\\mathbf{z}_{i}^{t}\\) on \\(F_{knw}\\) and \\(F_{unk}\\), respectively. \\(l_{n} \\in \\mathbb{R}\\) denotes the projection weight of \\(\\mathbf{z}_{i}^{t}\\) onto the unit base \\(\\mathbf{v}_{n}\\). This procedure is commonly recognized as orthogonal decomposition and has been extensively studied in the context of Independent Component Analysis (ICA)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.195, + 0.824, + 0.212 + ], + "angle": 0, + "content": "3.3. Adaptive Separation for Private Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Intuitively, enabled by feature decomposition, we could differentiate between common and private data by examining the dominance of \\(\\mathbf{z}_{i,unk}^{t}\\) over \\(\\mathbf{z}_i\\). However, such a straightforward approach is not applicable to SF-UniDA due to the impact of distribution covariate shift, which often deteriorates the significance of \\(\\mathbf{z}_{i,knw}^{t}\\) for common data, leading to the prevalence of \\(\\mathbf{z}_{i,unk}^{t}\\) in the representation. Additionally, the feature extractor inevitably captures representations that are independent of \\(F_{knw}\\), including background contexts [1], object textures [17], and high-frequency patterns [51], further exacerbating this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.893, + 0.569 + ], + "angle": 0, + "content": "To solve this, one feasible solution is to estimate the distribution of \\(\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}\\) to effectively calibrate these disturbances. Our observations have revealed that the empirical distributions of \\(\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}\\) exhibit a bimodal pattern, with the presence of two distinct peaks typically indicating the modes of common and private data. Consequently, we employ a two-component Gaussian Mixture Model (GMM) to estimate the distribution \\(\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}\\), with the components featuring a lower mean representing the common data and those with a higher mean corresponding to the private data. In particular, we denote \\(\\mu_{com}\\) and \\(\\mu_{pri}\\) as the expectation for common and private data, where \\(\\mu_{com} < \\mu_{pri}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.749 + ], + "angle": 0, + "content": "In the light of \\(\\mu_{com}\\) and \\(\\mu_{pri}\\), an intuitive practice is to treat all common and private categories equally. Accordingly, common and private data can be separated by establishing \\((\\mu_{com} + \\mu_{pri}) / 2\\) as the threshold, where data with \\(\\| \\mathbf{z}_{i,unk}^t\\| _2\\) below the threshold are recognized as common data, while those above are identified as private data. Despite its reasonable effectiveness, it discounts the inconsistency of covariate shifts across each category. Moreover, it fails to account for the variability in covariate shifts even among samples from the same category. These considerations motivate us to devise an instance-level decision strategy to achieve accurate and effective data identification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Concretely, we first leverage top- \\(K\\) sampling to construct target prototypes, denoted as \\(\\{\\mathbf{c}_c^t\\in \\mathbb{R}^D |_{c = 1,\\dots ,C}\\}\\) for each common category, adhering to the procedure in [29, 41]. Similar to [41], we empirically set \\(K = N_{t} / \\tilde{C}_{t}\\), where \\(N_{t}\\) represents the amount of target data, and \\(\\tilde{C}_t\\) signifies the estimated count of target categories. Diverging from [41], we additionally collect source anchors from \\(W_{cls}\\), denoted as \\(\\{\\mathbf{c}_c^s\\in \\mathbb{R}^D |_{c = 1,\\ldots ,C}\\}\\) to assist us handle scenarios involving source-private categories. The construct of target prototypes through top- \\(K\\) sampling fails to consider these sce" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23337" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "narios, making it susceptible to misidentifying private data as common. Further details regarding the target prototype construction are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Thereafter, we propose a novel metric, common score \\(\\epsilon_{i,c}\\), which deliberately considers the distance of \\(\\mathbf{z}_i^t\\) to the target prototype \\(\\mathbf{c}_c^t\\) and source anchor \\(\\mathbf{c}_c^s\\) to facilitate us in deriving instance-level decision boundaries. Formally, this metric is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.219, + 0.469, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d (\\mathbf {a}, \\mathbf {b}) = 1. 0 - \\cos (\\mathbf {a}, \\mathbf {b}), \\\\ \\epsilon_ {i, c} ^ {t} = 1. 0 - \\exp \\left(d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {t}\\right) - 1. 0\\right), \\\\ \\epsilon_ {i, c} ^ {s} = \\exp \\left(- d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {s}\\right)\\right), \\tag {3} \\\\ \\epsilon_ {i, c} = \\sqrt {\\epsilon_ {i , c} ^ {t} \\cdot \\epsilon_ {i , c} ^ {s}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.47, + 0.48 + ], + "angle": 0, + "content": "where \\(\\epsilon_{i,c}^{t}\\) and \\(\\epsilon_{i,c}^{s}\\) denote the common score from target prototype and source anchor viewpoint, respectively. Both \\(\\epsilon_{i,c}^{t}\\) and \\(\\epsilon_{i,c}^{s}\\) are clipped to the [0, 1] before fusion. We then utilize geometric mean to calculate fused common score, due to its advantage in multi-criteria decision-making. Note that we leverage different formulas to calculate \\(\\epsilon_{i,c}^{t}\\) and \\(\\epsilon_{i,c}^{s}\\). Under the same condition, i.e., \\(d(\\mathbf{z}_i^t,\\mathbf{c}_c^s) = d(\\mathbf{z}_i^t,\\mathbf{c}_c^t)\\), \\(\\epsilon_{i,c}^{s}\\) is greater than \\(\\epsilon_{i,c}^{t}\\). The spirit behind is that instances exhibiting proximity to source anchors are likely to be common data. Conversely, instances in close proximity to target prototypes do not guarantee their recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.542 + ], + "angle": 0, + "content": "Given the computed common score \\(\\epsilon_{i,c}\\) and the estimated distribution for \\(\\| \\mathbf{z}_{i,unk}^t\\| _2\\) we introduce the following pseudo-labeling strategy to effectively distinguish between common and private data instances:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.551, + 0.469, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\rho_ {i, c} = \\mu_ {c} + \\epsilon_ {i, c} \\cdot (\\mu_ {p r i} - \\mu_ {c}), \\\\ \\tilde {\\mathbf {y}} _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\text {u n k n o w n}, & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} \\geq \\rho_ {i, c} \\\\ \\mathbb {1} (\\operatorname {a r g m a x} (\\epsilon_ {i})), & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} < \\rho_ {i, c} \\end{array} \\right. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.623, + 0.469, + 0.746 + ], + "angle": 0, + "content": "where \\(\\mu_{c}\\) denotes the expectation of \\(\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}\\) for the \\(c\\)-th category, derived using the identical top-\\(K\\) sampling technique employed during the construction of the target prototypes. \\(\\rho_{i,c}\\) represents the instance-level decision boundary, which accounts for the potential covariate shifts across each category, as well as the data instances corresponding to both target prototypes and source anchors. \\(\\mathbb{1}\\) denotes the one-hot encoding operator, \\(\\tilde{\\mathbf{y}}_i^t\\) is the pseudo-label for \\(\\mathbf{x}_i^t\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.303, + 0.772 + ], + "angle": 0, + "content": "3.4. Optimization Objectives" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.469, + 0.841 + ], + "angle": 0, + "content": "To encourage the separation between common and private data, the optimization process involves three distinct objectives, including pseudo-label learning, feature decomposition regularization, and feature consensus regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Pseudo-label Learning. With the pseudo labels derived from Eq. 4, we apply the cross-entropy loss for model adaptation. Nevertheless, instead of assigning equal importance to all pseudo-labels, we introduce a Student's \\(t\\) distribution" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Details of class split. \\(\\mathcal{V},\\bar{\\mathcal{V}}_s\\) , and \\(\\bar{\\mathcal{V}}_t\\) signifies the common class, source-private class, and target-private class, respectively." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.119, + 0.875, + 0.211 + ], + "angle": 0, + "content": "
DatasetClass Split(γ/γs/γt)
OPDAOSDAPDA
Office-31 [43]10/10/1110/0/1110/21/0
Office-Home [49]10/5/5025/0/4025/40/0
VisDA [36]6/3/36/0/66/6/0
DomainNet [37]150/50/145--
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.214, + 0.892, + 0.259 + ], + "angle": 0, + "content": "to characterize the certainty associated with each pseudolabel, taking into account the distance between \\(\\| \\mathbf{z}_{i,unk}^t\\| _2\\) and \\(\\rho_{i,c}\\). Specifically," + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.266, + 0.892, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tau_ {i} ^ {t} \\propto 1 - \\left(1 + \\frac {\\left(\\rho_ {i , \\kappa} - \\| \\mathbf {z} _ {i , u n k} ^ {t} \\| _ {2}\\right) ^ {2}}{\\alpha}\\right) ^ {- \\frac {\\alpha + 1}{2}}, \\\\ \\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\sum_ {c = 1} ^ {C} \\tilde {y} _ {i, c} ^ {t} \\log \\delta_ {c} \\left(f _ {\\theta} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}\\right)\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.459 + ], + "angle": 0, + "content": "where \\(\\tau_{i}^{t}\\) denotes the certainty, \\(\\kappa = \\mathrm{argmax}(\\epsilon_{i,c}),\\alpha = 1\\mathrm{e - }4\\) by default. \\(\\delta_c(f_\\theta^t (\\mathbf{x}_i^t))\\) is the \\(c\\) -th soft-max probability for \\(\\mathbf{x}_i^t\\) \\(\\tilde{y}_{i,c}^{t}\\) corresponds to \\(c\\) -th one-hot encoded pseudo-label for \\(\\tilde{\\mathbf{y}}_i^t\\) . For those instances pseudo-labeled as private data, we refrain from introducing an additional \\(C + 1\\) -th class, instead employing a uniform distribution to represent them." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Feature Decomposition Regularization. To promote the discriminability of \\(\\mathbf{z}_{i,unk}^{t}\\) for identifying private data, we impose a feature decomposition regularizer. Concretely," + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.511, + 0.892, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p _ {i} = \\frac {\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right)}{\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right) + \\exp \\left(\\left\\| \\mathbf {z} _ {i , k n w} ^ {t} \\right\\| _ {2}\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {r e g} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\left(\\hat {y} _ {i} ^ {t} \\log \\left(p _ {i}\\right) + \\left(1 - \\hat {y} _ {i} ^ {t}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.604, + 0.887, + 0.62 + ], + "angle": 0, + "content": "\\(\\hat{y}_i^t = 1\\) or 0 means pseudo-labeled as private/common data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.62, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Feature Consensus Regularization. Recent studies [41, 52, 53] in SFDA and SF-UniDA observe that incorporating consensus regularization with nearest neighbors in feature space could assist models in achieving more stable performance. Following them, we integrate the objective \\(\\mathcal{L}_{con}\\) into LEAD. More details are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.71, + 0.891, + 0.74 + ], + "angle": 0, + "content": "Overall Optimization Objective. Taking into account the discussions above, the eventual loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.748, + 0.891, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e g} + \\mathcal {L} _ {c o n} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.777, + 0.789, + 0.792 + ], + "angle": 0, + "content": "where \\(\\lambda > 0\\) is a trade-off hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.633, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.691, + 0.849 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset. We empirically verify the effectiveness of our LEAD on four datasets, including Office-31 [43], OfficeHome [49], VisDA [36], and DomainNet [37]. For a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23338" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. H-score (\\%) comparison in OPDA scenario on Office-Home. U denotes methods applicable for all potential label-shift scenarios (i.e., OPDA, OSDA, and PDA). SF represents source data-free. CF indicates model adaptation without time-consuming K-means clustering." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.121, + 0.885, + 0.299 + ], + "angle": 0, + "content": "
MethodsUSFCFAr2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvg
CMU [15]✘✘✓56.056.959.267.064.367.854.751.166.468.257.969.761.6
DANCE [47]✓✘✓61.060.464.965.758.861.873.161.266.667.762.463.763.9
DCC [25]✓✘✘58.054.158.074.670.677.564.373.674.981.075.180.470.2
OVANet [44]✘✘✓62.875.678.670.768.875.071.358.680.576.164.178.971.8
GATE [7]✓✘✓63.875.981.474.072.179.874.770.382.779.171.581.775.6
UniOT [5]✘✘✓67.380.586.073.577.384.375.563.386.077.865.481.976.6
Source-only---47.371.681.951.557.269.456.040.376.661.444.273.560.9
SHOT-O [27]✘✓✓32.929.539.656.830.141.154.935.442.358.533.533.340.7
LEAD✓✓✓62.778.186.470.676.383.475.360.686.275.460.783.775.0
UMAD [28]✘✓✓61.176.382.770.767.775.764.455.776.373.260.477.270.1
+ LEAD✓✓✓65.282.687.478.076.587.579.865.088.479.463.083.578.0 (+7.9)
GLC [41]✓✓✘64.378.289.863.181.789.177.654.288.980.754.285.975.6
+ LEAD✓✓✘65.779.788.765.080.388.577.657.488.978.957.885.576.2 (+0.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.304, + 0.877, + 0.318 + ], + "angle": 0, + "content": "Table 3. H-score (\\%) comparison in OPDA scenario on Office-31, VisDA, and DomainNet. Some results are cited from UMAD [28]." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.322, + 0.885, + 0.512 + ], + "angle": 0, + "content": "
MethodsUSFCFOffice-31DomainNetVisDA
A2DA2WD2AD2WW2AW2DAvgP2RP2SR2PR2SS2PS2RAvgS2R
CMU [15]XX✓68.167.371.479.372.280.473.150.845.152.245.644.851.048.332.9
DANCE [47]✓X✓78.671.579.991.472.287.980.321.037.047.346.727.721.033.542.8
DCC [25]✓XX88.578.570.279.375.988.680.256.943.750.343.344.956.249.243.0
OVANet [44]XX✓85.879.480.195.484.094.386.556.047.151.744.947.457.250.753.1
GATE [7]✓X✓87.781.684.294.883.494.187.657.448.752.847.649.556.352.156.4
UniOT [5]XX✓87.088.588.498.887.696.691.159.351.847.848.346.858.352.057.3
Source-only---70.963.239.677.352.286.464.957.338.247.838.432.248.243.725.7
SHOT-O [27]X✓✓73.567.259.388.377.184.475.035.030.837.228.331.932.232.644.0
LEAD✓✓✓85.485.086.390.986.293.187.859.946.151.345.045.956.350.876.6
UMAD [28]X✓✓79.177.487.490.790.497.287.059.044.350.142.132.055.347.158.3
+ LEAD✓✓✓88.789.290.895.890.497.192.0 (+5.0)59.346.049.843.843.956.049.8 (+2.7)67.2 (+8.9)
GLC [41]✓✓X81.584.589.890.488.492.387.863.350.554.950.949.661.355.173.1
+ LEAD✓✓X83.086.887.692.087.193.288.3 (+0.5)63.550.755.151.649.961.355.4 (+0.3)76.8 (+3.7)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.579 + ], + "angle": 0, + "content": "fair comparison, we evaluate LEAD on partial-set domain adaptation (PDA), open-set domain adaptation (OSDA), and open-partial-set domain adaptation (OPDA). Details of classes split are summarized in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Evaluation. We utilize the same evaluation metric as previous works. In PDA scenarios, we report the classification accuracy for evaluation. In OSDA and OPDA scenarios, we report the H-score for performance comparison, which is a harmonic mean of accuracy over common data and accuracy over private data. It is worth noting that LEAD can be incorporated effortlessly into most existing SF-UniDA methods, for potential performance improvement. To demonstrate this merit, we integrate LEAD with two representative methods: UMAD [28] and GLC [41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.897 + ], + "angle": 0, + "content": "Implementation Details. All the experiments are conducted on a single RTX-3090 GPU with PyTorch-1.10. We adopt the same network architecture as existing baseline methods. During target model adaptation, we apply the SGD optimizer with momentum 0.9. The batch size is set to 64 for all experiments. The learning rate is set to 1e-3 for Office-31 and Office-Home, and 1e-4 for VisDA and DomainNet. As for \\(\\lambda\\), it is set to 0.3 for Office-31, 1.0 for VisDA, and 2.0 for Office-Home and DomainNet. During inference, we adopt the same protocol like [41]. More details are presented in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.519, + 0.702, + 0.535 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Results for OPDA Scenarios. To verify the effectiveness of LEAD, we first conduct experiments in the most challenging OPDA scenarios, where both source and target domains contain private data. Table 2 summarizes the results on Office-Home, and Table 3 details the results on Office-31, DomainNet, and VisDA. These results demonstrate that LEAD achieves performance surpassing or comparable to existing methods based on K-means clustering for segregating common and private data. In comparison with methods without clustering, LEAD achieves significantly superior performance. For instance, on Office-Home, DomainNet, and VisDA datasets, LEAD outperforms UMAD by \\(4.9\\%\\), \\(3.7\\%\\), and \\(18.3\\%\\), respectively. We attribute this to our instance-level adaptive decision strategy, which effectively circumvents the limitations imposed by a fixed threshold. While GLC achieves commendable performance on its own, the integration with LEAD further amplifies its effectiveness. For example, LEAD improves GLC in the H-score on VisDA from \\(73.1\\%\\) to \\(76.8\\%\\). The integration with LEAD proves even more significant against UMAD, with the H-score on Office-Home and VisDA boosting from \\(70.1\\% /58.3\\%\\) to \\(78.0\\% /67.2\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results for OSDA Scenarios. We then conduct experiments in OSDA scenarios, where only the target domain" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23339" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.09, + 0.759, + 0.103 + ], + "angle": 0, + "content": "Table 4. H-score (%) comparison in OSDA scenario on Office-Home, Office-31, and VisDA." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.107, + 0.887, + 0.309 + ], + "angle": 0, + "content": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X Y55.057.059.059.358.260.659.251.361.261.953.555.357.652.655.776.575.965.864.765.254.2
DANCE [47]✓ X Y6.59.09.920.410.49.228.412.812.614.27.913.212.984.978.879.178.868.378.879.867.5
DCC [25]✓ X X56.167.566.749.666.564.055.853.070.561.657.271.961.758.354.867.289.485.380.972.759.6
OVANet [44]X X Y58.666.369.962.065.268.659.853.469.368.759.666.764.090.588.386.798.288.398.491.766.1
GATE [7]✓ X Y63.870.575.866.467.971.767.361.576.070.461.875.169.088.486.584.295.086.196.789.570.8
Source-only- - -46.163.372.942.854.058.747.836.166.260.845.368.255.278.272.144.282.252.188.869.629.1
SHOT-O [27]X Y Y37.741.848.456.439.840.960.041.549.761.841.443.646.980.271.664.393.164.091.877.528.1
LEAD✓ Y Y60.770.876.561.068.670.865.559.874.264.857.775.867.284.985.190.294.890.396.590.374.2
UMAD [28]X Y Y59.271.876.663.569.071.962.554.672.866.557.970.766.488.584.486.895.088.295.989.866.8
+ LEAD✓ Y Y62.673.677.361.366.270.763.354.675.166.763.575.667.690.489.690.395.490.797.392.370.2
+2.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.21, + 0.313, + 0.757, + 0.326 + ], + "angle": 0, + "content": "Table 5. Accuracy comparison (%) in PDA scenario on Office-Home, Office-31, and VisDA." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.33, + 0.887, + 0.53 + ], + "angle": 0, + "content": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X X50.974.278.462.264.172.563.547.978.372.454.778.966.584.184.269.297.266.898.883.465.5
DANCE [47]✓ X X53.673.284.970.867.382.670.050.984.877.055.981.871.177.171.283.794.692.696.886.073.7
DCC [25]✓ X X54.247.557.583.871.686.263.765.075.285.578.282.670.987.381.395.4100.095.5100.093.372.4
OVANet [44]X X X34.154.672.142.447.355.938.226.261.756.735.868.949.569.461.761.490.266.498.774.634.3
GATE [7]✓ X X55.875.985.373.670.283.072.159.584.779.663.983.874.089.586.293.5100.094.498.693.775.6
Source-only- - -45.969.281.155.761.264.860.741.175.870.549.978.462.990.479.379.395.984.398.187.842.8
SHOT-P [27]X X Y64.785.190.175.173.984.276.464.190.380.763.385.577.889.884.492.296.692.299.492.474.2
LEAD✓ Y Y58.283.187.070.575.483.373.750.483.778.358.783.273.889.893.995.698.696.099.495.575.3
UMAD [28]X Y Y51.266.579.263.162.968.263.356.475.974.555.978.366.385.485.183.597.686.299.489.568.5
+ LEAD✓ Y Y61.178.287.570.578.581.874.160.483.378.863.282.175.089.290.890.698.093.0100.093.678.1
+8.7+4.1+9.6
GLC [41]✓ Y X55.979.087.572.571.882.774.941.782.477.360.484.372.589.889.892.896.696.199.494.176.2
+ LEAD✓ Y X56.681.288.172.572.283.275.343.683.478.259.784.273.289.893.295.696.996.099.495.175.5
+0.7+1.0-0.7
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.688 + ], + "angle": 0, + "content": "involves private data. Table 4 lists the results, demonstrating the superiority of LEAD against baselines. Specifically, LEAD obtains the H-score of \\(67.2\\%\\), \\(90.3\\%\\), and \\(74.2\\%\\) on Office-Home, Office-31, and VisDA datasets, respectively. Similar to the observations in OPDA scenarios, LEAD also exhibits performance improvement to existing methods. LEAD improves the H-score of UMAD by \\(1.2\\%\\), \\(2.5\\%\\), and \\(3.4\\%\\) across the Office-Home, Office-31, and VisDA datasets, respectively. The results again verify the merit of LEAD when incorporated with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results for PDA Scenarios. Finally, we verify the effectiveness of LEAD in PDA scenarios, where the target domain's label sets are subsets of those in the source domain. Results presented in Table 5 show that LEAD achieves superior performance even compared to methods tailored for PDA. In particular, LEAD attains the overall accuracy of \\(73.8\\%\\), \\(95.5\\%\\), and \\(75.3\\%\\) on Office-Home, Office-31, and VisDA datasets, individually. LEAD contributes the accuracy increase of \\(8.7\\%\\), \\(4.1\\%\\), and \\(9.6\\%\\) to UMAD on these datasets. Note that, there is a slight degradation \\((-0.7\\%)\\) for GLC w/ LEAD. It may arise from gradient conflict, causing the optimization process to inadvertently compromise one objective in favor of improving the overall objective. We will pursue better integration strategies in future work." + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.537, + 0.86, + 0.549 + ], + "angle": 0, + "content": "Table 6. Ablation studies of LEAD in OPDA scenarios." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.553, + 0.885, + 0.638 + ], + "angle": 0, + "content": "
LceLregLconOffice-31Office-HomeVisDA
---64.960.925.7
✓--82.073.661.5
✓✓-84.574.567.3
✓-✓87.274.575.7
✓✓✓87.875.076.6
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.644, + 0.712, + 0.659 + ], + "angle": 0, + "content": "4.3. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.735 + ], + "angle": 0, + "content": "Ablation Study. To assess the contribution of different components within LEAD, we carry out extensive ablation studies in OPDA scenarios. The results, detailed in Table 6, confirm that each component is not only effective but also complementary to the others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Efficiency Comparison. As previously stated, LEAD employs the distance to target prototypes and source anchors for adaptive pseudo-labeling. This design, in contrast to clustering-based methods, notably reduces the need for computational resources. Table 7 juxtaposes the runtime of GLC [41] and LEAD in OPDA scenarios. It is clear that LEAD requires minimal resources for deriving decision boundaries, whereas GLC's resource demands escalate dramatically with data scale. In particular, on DomainNet, with about 0.6 million images across 345 categories, GLC averages 897.65 seconds, while LEAD takes only 0.29 sec" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.085, + 0.303, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.215, + 0.266, + 0.224 + ], + "angle": 0, + "content": "(a) Sensitivity to \\(\\lambda\\)" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.086, + 0.488, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.215, + 0.459, + 0.224 + ], + "angle": 0, + "content": "(b) H-score convergence" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.086, + 0.666, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.215, + 0.636, + 0.223 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{Ar}\\rightarrow \\mathrm{Pr}\\) in OPDA" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.087, + 0.848, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.215, + 0.818, + 0.223 + ], + "angle": 0, + "content": "(d) \\(\\mathrm{Re}\\rightarrow \\mathrm{Cl}\\) in OPDA" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Figure 3. Robustness analysis. (a) shows the sensitivity to \\(\\lambda\\) on Office-31. (b) presents the H-score curves on VisDA. (c-d) present robustness analysis when varying the unknown private categories." + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.255, + 0.466, + 0.269 + ], + "angle": 0, + "content": "Table 7. Runtime (s) to derive pseudo-label decision boundaries." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.271, + 0.462, + 0.318 + ], + "angle": 0, + "content": "
Office-31Office-HomeVisDADomainNet
GLC [41]0.190.510.46897.65
LEAD0.020.020.110.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.364 + ], + "angle": 0, + "content": "onds on the same platform. Moreover, LEAD also shows substantial resource savings even on normal-scale datasets, e.g., on VisDA with over \\(75\\%\\) reduction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.469, + 0.457 + ], + "angle": 0, + "content": "Parameter Sensitivity. We investigate the parameter sensitivity of \\(\\lambda\\) on Office-31 in OPDA scenarios, where the range of \\(\\lambda\\) is set to [0.1, 0.2, 0.3, 0.4, 0.5]. Figure 3 (a) illustrates that the results are stable around the selected parameter \\(\\lambda = 0.3\\). Through oracle validation, we may find the optimal parameter configuration for this setting is \\(\\lambda = 0.1\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.469, + 0.503 + ], + "angle": 0, + "content": "Training Stability. Figure 3 (b) visualizes the training curves of our LEAD on VisDA in OPDA scenarios. Generally, the training procedure of LEAD is stable and effective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.469, + 0.611 + ], + "angle": 0, + "content": "Varying Unknown Classes. With the rise in unknown private classes, accurately distinguishing common from private data becomes more challenging. To examine the robustness of our LEAD, we compared it with GLC and UMAD in OPDA scenarios using the Office-Home dataset. The results in Figure 3 (c-d) show that LEAD achieves stable results and consistently boosts existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Effect of Instance-level Decision Boundary. To evaluate the contribution of our instance-level pseudo-labeling strategy, we conduct a comparison against the intuitive design, which uses the average of \\(\\mu_{\\text{com}}\\) and \\(\\mu_{\\text{pri}}\\) as the threshold. As depicted in Figure 4 (a), instance-level decision boundary strategy significantly outperforms the vanilla global decision design (75.0% vs 63.1% on Office-Home in OPDA scenarios). We attribute this to that vanilla design discounts the inconsistency of covariate shifts across categories, and also the variability among samples from the same category." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Exploration with Entropy as the Indicator. Contrasting with existing threshold-based methods that utilize Entropy as the indicator for differentiating common and private data, we propose decomposing the feature and using the projection \\(\\mathbf{z}_{i,unk}^{t}\\) on the source-unknown space as an alternative indicator. To empirically substantiate the superiority, we remold our framework to incorporate Entropy as the indicator. Figure 4 (b) depicts the results. An observation is that while employing entropy as the indicator within our framework is" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.245, + 0.695, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.357, + 0.698, + 0.366 + ], + "angle": 0, + "content": "(a) Effect of Instance-level Decision Boundary" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.245, + 0.865, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.357, + 0.867, + 0.366 + ], + "angle": 0, + "content": "(b) Exploration with Entropy as the Indicator" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.422 + ], + "angle": 0, + "content": "Figure 4. Methodology analysis. (a) compares the effectiveness of our instance-level decision strategy against the vanilla global decision strategy. (b) examines the efficacy of using Entropy as the indicator for private data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.485 + ], + "angle": 0, + "content": "feasible, its performance is inferior compared to \\(\\mathbf{z}_{i,unk}^{t}\\). We speculate that this inferiority might result from entropy's limited capacity to distinguish between common samples with high uncertainty and private unknown samples." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.619, + 0.513 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.764 + ], + "angle": 0, + "content": "In this paper, we delve into source-free universal domain adaptation (SF-UniDA). Different from existing methods that leverage hand-crafted thresholding strategies or feature-clustering algorithms, we propose a novel idea of LEArning Decomposition (LEAD) tailored for SF-UniDA. Technically, LEAD first decomposes features into source-known and -unknown two parts, then establishes instance-level decision boundaries for pseudo-labeling. It offers an elegant view to identify target-private unknown data without tedious tuning a threshold or relying on unstable clustering. Extensive experiments across various scenarios have verified the effectiveness of our LEAD. Besides, it is appealing in view that LEAD can be seamlessly incorporated into existing methods to further boost performance. Remarkably, LEAD improves UMAC by \\(7.9\\%\\) overall H-score in the OPDA scenario on Office-Home dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No. 62372329), in part by the National Key Research and Development Program of China (No. 2021YFB2501104), in part by Shanghai Rising Star Program (No.21QC1400900), in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program, and in part by the International Exchange Program for Graduate Students, Tongji University (No.2023020053)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23341" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Richard Bellman. Dynamic programming. Science, 153 (3731):34-37, 1966. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[3] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.468, + 0.258 + ], + "angle": 0, + "content": "[4] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[5] Wanxing Chang, Ye Shi, Hoang Tuan, and Jingya Wang. Unified optimal transport framework for universal domain adaptation. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[6] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In ECCV. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.347, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[7] Liang Chen, Yihang Lou, Jianzhong He, Tao Bai, and Minghua Deng. Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[8] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.447, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[9] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[11] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE TPAMI, 39(9):1853-1865, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.468, + 0.686 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.742 + ], + "angle": 0, + "content": "[13] Qi Dou, Cheng Ouyang, Cheng Chen, Hao Chen, and Pheng-Ann Heng. Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss. In IJCAI, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE TIP, 30:6473-6484, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[15] Bo Fu, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Learning to detect open classes for universal domain adaptation. In ECCV. Springer, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.468, + 0.898 + ], + "angle": 0, + "content": "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, pages 2096-2030, 2016. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[17] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.202 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[20] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.369 + ], + "angle": 0, + "content": "[21] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.466 + ], + "angle": 0, + "content": "[23] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE TPAMI, pages 766-785, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[24] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[25] Guangrui Li, Guoliang Kang, Yi Zhu, Yunchao Wei, and Yi Yang. Domain consensus clustering for universal domain adaptation. In CVPR, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[26] Shuang Li, Shiji Song, Gao Huang, Zhengming Ding, and Cheng Wu. Domain invariant and class discriminative feature learning for visual domain adaptation. IEEE TIP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.661 + ], + "angle": 0, + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML. PMLR, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[28] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[29] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[30] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE TMM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[31] Xinyu Liu and Yixuan Yuan. A source-free domain adaptive polyp detection framework with style diversification flow. IEEE TMI, pages 1897-1908, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[32] Zhengfa Liu, Guang Chen, Zhijun Li, Yu Kang, Sanqing Qu, and Changjun Jiang. Psdc: A prototype-based shared-dummy classifier model for open-set domain adaptation. IEEE Transactions on Cybernetics, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[33] Zhengfa Liu, Guang Chen, Zhijun Li, Sanqing Qu, Alois Knoll, and Changjun Jiang. D2ifln: Disentangled domain-invariant feature learning networks for domain generalization. IEEE Transactions on Cognitive and Developmental Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[34] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In ICCV, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[35] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[37] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[38] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[39] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[40] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[41] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[43] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.636 + ], + "angle": 0, + "content": "[44] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In ICCV, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[45] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[46] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. In NeurIPS, pages 16282-16292, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[48] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In WACV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[49] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[50] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[51] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[52] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[53] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In CVPR, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23343" + } + ] +] \ No newline at end of file diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_origin.pdf b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6591231c42dcfc7480b22d2d46656233e1cfd19a --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/68e710b5-658b-4bb7-a5db-feae6ea998dd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8405c3f108fe52fc0c6590d7d020f5dce7fba2eadd12995cdb5ea4992b0ff5ca +size 5577081 diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/full.md b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..35cf541f4d555033380440fb7bf6a7e57139541f --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/full.md @@ -0,0 +1,303 @@ +# LEAD: Learning Decomposition for Source-free Universal Domain Adaptation + +Sanqing $\mathrm{Qu}^1$ , Tianpei $\mathrm{Zou}^1$ , Lianghua $\mathrm{He}^1$ , Florian Röhrbein $^2$ , Alois Knoll $^3$ , Guang Chen $^{1*}$ , Changjun Jiang $^1$ + +$^{1}$ Tongji University, $^{2}$ Chemnitz University of Technology, $^{3}$ Technical University of Munich + +# Abstract + +Universal Domain Adaptation (UniDA) targets knowledge transfer in the presence of both covariate and label shifts. Recently, Source-free Universal Domain Adaptation (SF-UniDA) has emerged to achieve UniDA without access to source data, which tends to be more practical due to data protection policies. The main challenges lie in determining whether covariate-shifted samples belong to target-private unknown categories. Existing methods tackle this either through hand-crafted thresholding or by developing time-consuming iterative clustering strategies. In this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. Technically, LEAD initially leverages the orthogonal decomposition analysis for feature decomposition. Then, LEAD builds instance-level decision boundaries to adaptively identify target-private data. Extensive experiments across various UniDA scenarios have demonstrated the effectiveness and superiority of LEAD. Notably, in the OPDA scenario on VisDA dataset, LEAD outperforms GLC by $3.5\%$ overall H-score and reduces $75\%$ time to derive pseudo-labeling decision boundaries. Besides, LEAD is also appealing in that it is complementary to most existing methods. The code is available at https://github.com/ispc-lab/LEAD. + +# 1. Introduction + +Deep neural networks (DNNs) have yielded impressive results in a variety of computer vision tasks [12, 18, 22, 42]. However, DNNs often fail to generalize well to new domains due to the discrepancy between training (source) and test (target) data distributions (i.e., covariate shift). Such a shift poses significant challenges for some safety-critical applications like autonomous driving [9, 10], or medical imaging [13, 31]. Unsupervised Domain Adaptation (DA) [16, 23] aims to address this challenge by transferring task-specific knowledge from a labeled source domain to an unlabeled target domain. Despite promising results, most + +![](images/eedb42aa8eec51aebbbdf6588d9e6093c0cee0ecfc340e4655e0416bee70ea83.jpg) +Figure 1. (a) Pseudo-label accuracy curves for target-private data in the OPDA scenario on VisDA dataset. An observation is that clustering-based GLC does not work well in deriving decision boundaries for target-private data. This may be due to the curse of dimensionality, leading K-means to faintly discerning clusters. (b) Frequency distribution for the normalized feature magnitude of target data in the source-unknown space (orthogonal complement of the space spanned by source model weights). Features are on task $(A\to D)$ in the OPDA scenario of Office-31. The results show that target-private data are expected to involve more components from source-unknown space, even with covariate shifts. + +![](images/311904e7893df586bad37f09500ec8d8c0e403d392f568a50a4fba25e4cff427.jpg) + +existing favorable DA methods [16, 20, 27, 45] assume that the label spaces are identical across the source and target domain, thus being only applicable to closed-set scenarios and limiting applicability to practical applications. + +To handle generalized cases, Universal Domain Adaptation (UniDA) [47] was proposed to allow all potential label shifts between source and target domains. Different from existing methods specialized for Partial Domain Adaptation (PDA) [3, 4], Open-set Domain Adaptation (OSDA) [35, 46], or Open-Partial Domain Adaptation (OPDA) [44, 54], in UniDA, we have no prior knowledge about the label shift, e.g., information on matched common classes or the number of categories in the target domain. Nonetheless, most existing UniDA works [7, 25, 47] necessitate concurrent access to source and target data. This requirement becomes increasingly impractical in light of strict data protection policies [50]. In this paper, we focus on Source-free Universal Domain Adaptation (SF-UniDA) [28, 41], where only a pre-trained source model is provided for knowledge transfer rather than labeled source data. + +The primary objective for UniDA is to recognize target data belonging to common categories and separate them from target-private "unknown" data. Existing meth + +ods have designed various criteria based on instance-level predictions, e.g., entropy-based [14, 47, 54], confidence-based [15, 24, 44]. However, defining a universal threshold for all scenarios is challenging and often sub-optimal. To mitigate tedious parameter-tuning, some methods start to perform source and target consensus clustering to achieve common and private data separation [7, 25]. Considering that SF-UniDA is inaccessible to source data, GLC [41] devises an iterative one-vs-all clustering strategy to realize target-private data identification. Regardless of progress, iterative clustering can be computationally intensive. Besides, due to the curse of dimensionality [2], methods like K-means are prone to overfitting and faintly discerning clusters. This can be implied to some degree from Figure 1 (a). + +In this paper, we propose to address the SF-UniDA problem from a new perspective, i.e., feature decomposition. Our launching point is that, even though target data might have shifted in the feature space, target-private data from unknown categories are likely to exhibit a higher proportion of components from the orthogonal complement (source-unknown) space spanned by source model weights. The frequency distribution of normalized feature magnitudes in the source-unknown space depicted in Figure 1 (b) basically demonstrates our argument, which has been rarely explored in the literature. To materialize our idea, we propose the LEArning Decomposition (LEAD) framework. In particular, LEAD initially leverages the orthogonal decomposition to construct source-known and -unknown space. The projection on source-unknown space is then regarded as the discriminative representation with respect to target-private categories. Different from existing thresholding methods introducing a fixed global hand-crafted threshold for all scenarios, LEAD deliberately considers the distance to the target prototypes and source anchors to establish instance-level decision boundaries. We analyze and evaluate our LEAD under a variety of SF-UniDA scenarios, ranging from PDA, OSDA, to OPDA. Extensive experiments demonstrate the effectiveness and superiority. + +The main contribution is the proposal of a LEArning Decomposition (LEAD) framework for source-free universal domain adaptation (SF-UniDA). This solution leads to elegant views for identifying target-private unknown data without tedious tuning thresholds or relying on iterative unstable clustering. Remarkably, in the OPDA scenario on VisDA, LEAD attains an H-score of $76.6\%$ , surpassing GLC [41] by $3.5\%$ . Besides, LEAD is complementary to most existing SF-UniDA methods. For instance, in the OPDA scenario on Office-Home, LEAD advances UMAD [28] with H-score improvement from $70.1\%$ to $78.0\%$ . + +# 2. Related Work + +Unsupervised Domain Adaptation. When DNNs encounter data from a distribution that differs from their train + +ing data, inevitable performance degeneration is commonly observed. To address this issue, domain adaptation (DA) [16, 23] has been introduced, leveraging labeled data from source domains to train DNNs for unlabeled target domains in a transductive learning manner. Existing DA methods can be broadly categorized into two paradigms, namely moment matching [11, 34, 48] and adversarial learning [16, 19, 21]. Regardless of their effectiveness in various applications such as object recognition [16, 27, 39], semantic segmentation [19, 48], and object detection [8, 21], most existing approaches assume that label spaces are identical across the source and target domains, limiting their applicability. + +Universal Domain Adaptation. There have been some approaches introduced to deal with label-shift scenarios, such as open-set domain adaptation (OSDA) [32, 35, 46], partial-set domain adaptation (PDA) [3, 4], and open-partial-set domain adaptation (OPDA) [15, 44, 54]. However, these methods are often tailored to specific scenarios and may not be readily applicable to other label shift situations. Universal domain adaptation (UniDA), on the other hand, is designed to facilitate adaptation across all potential label shift scenarios. Despite the progress in UniDA, many existing methods [5, 7, 44] necessitate simultaneous access to both source and target data, making them impractical for applications with stringent data protection policies. To address this challenge, source-free universal domain adaptation (SF-UniDA) [28, 41] has been proposed, wherein source data are only utilized for pre-training and are inaccessible during target adaptation. To promote the distinction between common and private data, existing methods often resort to manually defined thresholding or the introduction of global clustering techniques. Nevertheless, selecting an appropriate threshold for all scenarios can be both laborious and sub-optimal. Additionally, K-means clustering tends to be unstable when applied to high-dimensional data due to the curse of dimensionality [2]. Our work delves into these limitations and proposes a novel and elegant solution from the perspective of feature decomposition. + +Feature Decomposition. As a fundamental technique in machine learning, feature decomposition is applied to deconstruct complex data into simpler and more interpretable components. In the field of domain adaptation and domain generalization, feature decomposition has been leveraged in various studies [6, 26, 30, 33, 38, 40] to enhance feature alignment or facilitate domain-invariant feature learning by breaking down features into content and style components. Nonetheless, the content and style decomposition design is less applicable for UniDA, as the primary challenge lies in distinguishing common and private data. In this paper, we approach the problem from the perspective of orthogonal feature decomposition and employ Independent Component Analysis to decompose features into two orthogonal parts, thereby facilitating the intended objective. + +![](images/eb8beafd7fe75af0f36aaa1ef4acfd447b286f8f39fe8c69684b04b60cc5fffe.jpg) +Figure 2. (a) Illustrations of traditional universal domain adaptation (UniDA) and source-free universal domain adaptation (SF-UniDA). Traditional UniDA methods necessitate data from both source and target domains concurrently. In SF-UniDA, source data are solely utilized for pre-training. Adaptation is performed by harnessing target data and the source model $f_{\theta}^{s} = h_{\theta}^{s} \circ g_{\theta}^{s}$ . (b) An overview of our Learning Decomposition (LEAD) framework. Pseudo-labeling is an important technique for UniDA and SF-UniDA. The primary objective is to recognize target data associated with common label sets and exclude data within the target-private label space. Different from existing methods that perform private data identification by hand-crafted thresholding on predictions or iterative global clustering, we tackle this from the viewpoint of feature decomposition. The rationale is that despite potential shifts in the feature space, target-private data are expected to encompass more components from the orthogonal complement (source-unknown) space of the source model. Technically, LEAD first performs orthogonal decomposition to decompose target features into source-known and -unknown parts, i.e., $\mathbf{z}_{i,knw}^{t}$ and $\mathbf{z}_{i,unk}^{t}$ . $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ is considered as an indicator for private data. Next, LEAD employs a two-component Gaussian Mixture Model to estimate the distribution of $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ . Thereafter, LEAD devises a metric named "common score" $\epsilon_{i,c}$ that accounts for distances to both target prototypes and source anchors (derived from $h_{\theta}^{t}$ ) to facilitate deriving instance-level decision boundary $\rho_{i,c}$ . LEAD provides an elegant solution to distinguish target-private data, mitigating the need for tedious hand-crafted threshold tuning or dependence on time-consuming iterative clustering. LEAD could also serve as a complementary approach to most existing SF-UniDA methods. + +![](images/7ecf330707fdfa614df79bb0c5548e32eddd9d563b3d72c33366cd7e75e9087c.jpg) + +# 3. Methodology + +# 3.1. Preliminary + +In this paper, we consider a generalized and challenging case in domain adaptation: universal domain adaptation (UniDA), which aims to achieve knowledge transfer in the presence of both covariate and label shifts. In UniDA, there is a labeled source domain $\mathcal{D}^s = \left\{\left(\mathbf{x}_i^s,\mathbf{y}_i^s\right)\right\}_{i = 1}^{N_s}$ where $\mathbf{x}_i^s\in \mathcal{X}^s\subset \mathbb{R}^X,\mathbf{y}_i^s\in \mathcal{Y}^s\subset \mathbb{R}^C$ , and an unlabeled target domain $\mathcal{D}^t = \left\{\left(\mathbf{x}_i^t,\ref{eq:2.3.1}\right)\right\}_{i = 1}^{N_t}$ where $\mathbf{x}_i^t\in \mathcal{X}^t\subset \mathbb{R}^X.$ $\mathcal{V} = \mathcal{V}^s\cap \mathcal{V}^t$ denotes the common label sets shared by source and target domains. $\bar{\mathcal{V}}^s = \mathcal{V}\setminus \mathcal{V}^t$ and $\bar{\mathcal{V}}^t = \mathcal{V}\setminus \mathcal{V}^s$ symbolizes the label sets private to source and target domain, individually. UniDA supposes that we have no prior knowledge about $\mathcal{V}^t$ . $\mathcal{V}$ and $\bar{\mathcal{V}}^t$ are also unavailable. The objective is to recognize common data belonging to $\mathcal{V}$ and reject private data in $\bar{\mathcal{V}}^t$ + +Different from traditional UniDA methods that need concurrent access to source and target data, under the source-free universal domain adaptation (SF-UniDA) setting, $\mathcal{D}_s$ is only applicable for source model pre-training and not acces- + +sible during adaptation. Assuming that the source model is represented as $f_{\theta}^{s} = h_{\theta}^{s}\circ g_{\theta}^{s}$ , where $g_{\theta}^{s}:\mathbb{R}^{X}\to \mathbb{R}^{D}$ denotes the feature extractor and $h_\theta^s:\mathbb{R}^D\to \mathbb{R}^C$ is the classifier module. Following existing methods [27, 41], we freeze the classifier module $h_\theta^t = h_\theta^s$ and only learn a target-specific feature extractor $g_{\theta}^{t}$ to realize model adaptation. Figure 2 (a) presents an illustration for UniDA and SF-UniDA. + +The main challenge is how to distinguish common data from private data. Prior methods [25, 28, 41, 47] have designed various algorithms to encourage the separation. However, these methods perform data identification either based on hand-crafted thresholding criteria or by introducing time-consuming unstable clustering strategies. We propose to mitigate these limitations from the perspective of feature decomposition. An observation is that even though target data may have shifted in the feature space, the features of private data still include more components from the orthogonal complement (source-unknown) space of the pre-trained model. Technically, we present a novel learning decomposition (LEAD) framework. LEAD capitalizes on the orthogonal decomposition to build two orthogonal fea + +ture spaces, i.e., source-known and -unknown space. Feature projection on source-unknown space is extracted as the descriptor for private data. Then, LEAD considers the distances to both the target prototypes and source anchors to establish instance-level decision boundaries. Figure 2 (b) manifests the architecture of our LEAD. + +# 3.2. Orthogonal Feature Decomposition + +Pseudo-labeling serves as an important technique in unsupervised learning, yet many existing strategies only account for closed-set scenarios, disregarding the label-shift situations, which impairs their ability to differentiate target-private data. To fulfill the objective of common and private data separation, we propose to orthogonally decompose the feature representation into two uncorrelated parts, i.e., features associated with source-known space $F_{knw}$ and those related to source-unknown space $F_{unk}$ . This design stems from a simple inductive bias in favor of private data identification. Due to the lack of exposure to target data during source pre-training, particularly target-private data, the source-known space does not include any significant attributes of target-private data. Subsequently, after normalization, target-private data are expected to contain more components from source-unknown space, even with covariate shifts. This hypothesis is basically demonstrated by empirical evidence, as illustrated in Figure 1 (b). + +Formally, denoting the weight of classifier $h_{\theta}^{t}$ is $W_{cls} \in \mathbb{R}^{C \times D}$ . Source-known space $F_{knw}$ can be represented as the vector space of span $\{\mathbf{w}_n\}$ , where $\{\mathbf{w}_n \in \mathbb{R}^D | n = 1, \dots, C\}$ is a group weight vectors of $W_{cls}$ . Source-unknown space $F_{unk}$ is the orthogonal complement space of $F_{knw}$ , i.e., $\mathbb{R}^D = F_{knw} \oplus F_{unk}$ . Nevertheless, unit orthogonality in $\{\mathbf{w}_n \in \mathbb{R}^D | n = 1, \dots, C\}$ is not guaranteed (i.e., when $n \neq m$ ( $\mathbf{w}_n, \mathbf{w}_m \neq 0$ ), making it hard to derive $F_{unk}$ . Considering that $F_{knw}$ and $F_{unk}$ are the row space and null space of $W_{cls}$ . Therefore, we employ the singular value decomposition (SVD) on $W_{cls}$ to obtain $F_{knw}$ and $F_{unk}$ . Specifically, + +$$ +W _ {c l s} = U \Sigma V ^ {T}, +$$ + +$$ +F _ {k n w} = \operatorname {s p a n} \left\{\mathbf {v} _ {n} | _ {n = 1, \dots , C} \right\}, \tag {1} +$$ + +$$ +F _ {u n k} = \operatorname {s p a n} \left\{\mathbf {v} _ {n} | _ {n = C + 1, \dots , D} \right\} +$$ + +where $\Sigma \in \mathbb{R}^{C\times D}$ is a diagonal matrix, $U\in \mathbb{R}^{C\times C}$ and $V\in \mathbb{R}^{D\times D}$ are both orthogonal unitary matrices. $\{\mathbf{v}_n|\mathbf{v}_n\in \mathbb{R}^D\}_{n = 1}^D$ are columns of $V$ + +Assuming that $\mathbf{z}_i^t$ is the feature of target data $\mathbf{x}_i^t$ extracted from feature extractor $g_{\theta}^{t}$ after normalization, i.e., $\mathbf{z}_i^t = g_\theta^t (\mathbf{x}_i^t) / ||g_\theta^t (\mathbf{x}_i^t)||_2$ , then $\mathbf{z}_i^t$ can be represented as the weighted summation of two orthogonal bases: + +$$ +\mathbf {z} _ {i} ^ {t} = \underbrace {\sum_ {n = 1} ^ {C} l _ {n} \cdot \mathbf {v} _ {n}} _ {\mathbf {z} _ {i, k n w} ^ {t}} + \underbrace {\sum_ {n = C + 1} ^ {D} l _ {n} \cdot \mathbf {v} _ {n}} _ {\mathbf {z} _ {i, u n k} ^ {t}} \tag {2} +$$ + +where $||\mathbf{z}_{i,knw}^{t}||_{2}^{2} + ||\mathbf{z}_{i,unk}^{t}||_{2}^{2} = 1$ , $\mathbf{z}_{i,knw}^{t}$ and $\mathbf{z}_{i,unk}^{t}$ are the projections of $\mathbf{z}_{i}^{t}$ on $F_{knw}$ and $F_{unk}$ , respectively. $l_{n} \in \mathbb{R}$ denotes the projection weight of $\mathbf{z}_{i}^{t}$ onto the unit base $\mathbf{v}_{n}$ . This procedure is commonly recognized as orthogonal decomposition and has been extensively studied in the context of Independent Component Analysis (ICA). + +# 3.3. Adaptive Separation for Private Data + +Intuitively, enabled by feature decomposition, we could differentiate between common and private data by examining the dominance of $\mathbf{z}_{i,unk}^{t}$ over $\mathbf{z}_i$ . However, such a straightforward approach is not applicable to SF-UniDA due to the impact of distribution covariate shift, which often deteriorates the significance of $\mathbf{z}_{i,knw}^{t}$ for common data, leading to the prevalence of $\mathbf{z}_{i,unk}^{t}$ in the representation. Additionally, the feature extractor inevitably captures representations that are independent of $F_{knw}$ , including background contexts [1], object textures [17], and high-frequency patterns [51], further exacerbating this issue. + +To solve this, one feasible solution is to estimate the distribution of $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ to effectively calibrate these disturbances. Our observations have revealed that the empirical distributions of $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ exhibit a bimodal pattern, with the presence of two distinct peaks typically indicating the modes of common and private data. Consequently, we employ a two-component Gaussian Mixture Model (GMM) to estimate the distribution $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ , with the components featuring a lower mean representing the common data and those with a higher mean corresponding to the private data. In particular, we denote $\mu_{com}$ and $\mu_{pri}$ as the expectation for common and private data, where $\mu_{com} < \mu_{pri}$ . + +In the light of $\mu_{com}$ and $\mu_{pri}$ , an intuitive practice is to treat all common and private categories equally. Accordingly, common and private data can be separated by establishing $(\mu_{com} + \mu_{pri}) / 2$ as the threshold, where data with $\| \mathbf{z}_{i,unk}^t\| _2$ below the threshold are recognized as common data, while those above are identified as private data. Despite its reasonable effectiveness, it discounts the inconsistency of covariate shifts across each category. Moreover, it fails to account for the variability in covariate shifts even among samples from the same category. These considerations motivate us to devise an instance-level decision strategy to achieve accurate and effective data identification. + +Concretely, we first leverage top- $K$ sampling to construct target prototypes, denoted as $\{\mathbf{c}_c^t\in \mathbb{R}^D |_{c = 1,\dots ,C}\}$ for each common category, adhering to the procedure in [29, 41]. Similar to [41], we empirically set $K = N_{t} / \tilde{C}_{t}$ , where $N_{t}$ represents the amount of target data, and $\tilde{C}_t$ signifies the estimated count of target categories. Diverging from [41], we additionally collect source anchors from $W_{cls}$ , denoted as $\{\mathbf{c}_c^s\in \mathbb{R}^D |_{c = 1,\ldots ,C}\}$ to assist us handle scenarios involving source-private categories. The construct of target prototypes through top- $K$ sampling fails to consider these sce + +narios, making it susceptible to misidentifying private data as common. Further details regarding the target prototype construction are included in the supplementary. + +Thereafter, we propose a novel metric, common score $\epsilon_{i,c}$ , which deliberately considers the distance of $\mathbf{z}_i^t$ to the target prototype $\mathbf{c}_c^t$ and source anchor $\mathbf{c}_c^s$ to facilitate us in deriving instance-level decision boundaries. Formally, this metric is defined as: + +$$ +\begin{array}{l} d (\mathbf {a}, \mathbf {b}) = 1. 0 - \cos (\mathbf {a}, \mathbf {b}), \\ \epsilon_ {i, c} ^ {t} = 1. 0 - \exp \left(d \left(\mathbf {z} _ {i} ^ {t}, \mathbf {c} _ {c} ^ {t}\right) - 1. 0\right), \\ \epsilon_ {i, c} ^ {s} = \exp \left(- d \left(\mathbf {z} _ {i} ^ {t}, \mathbf {c} _ {c} ^ {s}\right)\right), \tag {3} \\ \epsilon_ {i, c} = \sqrt {\epsilon_ {i , c} ^ {t} \cdot \epsilon_ {i , c} ^ {s}} \\ \end{array} +$$ + +where $\epsilon_{i,c}^{t}$ and $\epsilon_{i,c}^{s}$ denote the common score from target prototype and source anchor viewpoint, respectively. Both $\epsilon_{i,c}^{t}$ and $\epsilon_{i,c}^{s}$ are clipped to the [0, 1] before fusion. We then utilize geometric mean to calculate fused common score, due to its advantage in multi-criteria decision-making. Note that we leverage different formulas to calculate $\epsilon_{i,c}^{t}$ and $\epsilon_{i,c}^{s}$ . Under the same condition, i.e., $d(\mathbf{z}_i^t,\mathbf{c}_c^s) = d(\mathbf{z}_i^t,\mathbf{c}_c^t)$ , $\epsilon_{i,c}^{s}$ is greater than $\epsilon_{i,c}^{t}$ . The spirit behind is that instances exhibiting proximity to source anchors are likely to be common data. Conversely, instances in close proximity to target prototypes do not guarantee their recognition. + +Given the computed common score $\epsilon_{i,c}$ and the estimated distribution for $\| \mathbf{z}_{i,unk}^t\| _2$ we introduce the following pseudo-labeling strategy to effectively distinguish between common and private data instances: + +$$ +\begin{array}{l} \rho_ {i, c} = \mu_ {c} + \epsilon_ {i, c} \cdot (\mu_ {p r i} - \mu_ {c}), \\ \tilde {\mathbf {y}} _ {i} ^ {t} = \left\{ \begin{array}{l l} \text {u n k n o w n}, & \text {i f} \| \mathbf {z} _ {i, u n k} ^ {t} \| _ {2} \geq \rho_ {i, c} \\ \mathbb {1} (\operatorname {a r g m a x} (\epsilon_ {i})), & \text {i f} \| \mathbf {z} _ {i, u n k} ^ {t} \| _ {2} < \rho_ {i, c} \end{array} \right. \tag {4} \\ \end{array} +$$ + +where $\mu_{c}$ denotes the expectation of $\| \mathbf{z}_{i,unk}^{t}\|_{2}$ for the $c$ -th category, derived using the identical top- $K$ sampling technique employed during the construction of the target prototypes. $\rho_{i,c}$ represents the instance-level decision boundary, which accounts for the potential covariate shifts across each category, as well as the data instances corresponding to both target prototypes and source anchors. $\mathbb{1}$ denotes the one-hot encoding operator, $\tilde{\mathbf{y}}_i^t$ is the pseudo-label for $\mathbf{x}_i^t$ . + +# 3.4. Optimization Objectives + +To encourage the separation between common and private data, the optimization process involves three distinct objectives, including pseudo-label learning, feature decomposition regularization, and feature consensus regularization. + +Pseudo-label Learning. With the pseudo labels derived from Eq. 4, we apply the cross-entropy loss for model adaptation. Nevertheless, instead of assigning equal importance to all pseudo-labels, we introduce a Student's $t$ distribution + +Table 1. Details of class split. $\mathcal{V},\bar{\mathcal{V}}_s$ , and $\bar{\mathcal{V}}_t$ signifies the common class, source-private class, and target-private class, respectively. + +
DatasetClass Split(γ/γs/γt)
OPDAOSDAPDA
Office-31 [43]10/10/1110/0/1110/21/0
Office-Home [49]10/5/5025/0/4025/40/0
VisDA [36]6/3/36/0/66/6/0
DomainNet [37]150/50/145--
+ +to characterize the certainty associated with each pseudolabel, taking into account the distance between $\| \mathbf{z}_{i,unk}^t\| _2$ and $\rho_{i,c}$ . Specifically, + +$$ +\begin{array}{l} \tau_ {i} ^ {t} \propto 1 - \left(1 + \frac {\left(\rho_ {i , \kappa} - \| \mathbf {z} _ {i , u n k} ^ {t} \| _ {2}\right) ^ {2}}{\alpha}\right) ^ {- \frac {\alpha + 1}{2}}, \\ \mathcal {L} _ {c e} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \tau_ {i} ^ {t} \cdot \sum_ {c = 1} ^ {C} \tilde {y} _ {i, c} ^ {t} \log \delta_ {c} \left(f _ {\theta} ^ {t} \left(\mathbf {x} _ {i} ^ {t}\right)\right) \tag {5} \\ \end{array} +$$ + +where $\tau_{i}^{t}$ denotes the certainty, $\kappa = \mathrm{argmax}(\epsilon_{i,c}),\alpha = 1\mathrm{e - }4$ by default. $\delta_c(f_\theta^t (\mathbf{x}_i^t))$ is the $c$ -th soft-max probability for $\mathbf{x}_i^t$ $\tilde{y}_{i,c}^{t}$ corresponds to $c$ -th one-hot encoded pseudo-label for $\tilde{\mathbf{y}}_i^t$ . For those instances pseudo-labeled as private data, we refrain from introducing an additional $C + 1$ -th class, instead employing a uniform distribution to represent them. + +Feature Decomposition Regularization. To promote the discriminability of $\mathbf{z}_{i,unk}^{t}$ for identifying private data, we impose a feature decomposition regularizer. Concretely, + +$$ +\begin{array}{l} p _ {i} = \frac {\exp \left(\left\| \mathbf {z} _ {i , u n k} ^ {t} \right\| _ {2}\right)}{\exp \left(\left\| \mathbf {z} _ {i , u n k} ^ {t} \right\| _ {2}\right) + \exp \left(\left\| \mathbf {z} _ {i , k n w} ^ {t} \right\| _ {2}\right)}, \tag {6} \\ \mathcal {L} _ {r e g} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \tau_ {i} ^ {t} \cdot \left(\hat {y} _ {i} ^ {t} \log \left(p _ {i}\right) + \left(1 - \hat {y} _ {i} ^ {t}\right) \log \left(1 - p _ {i}\right)\right) \\ \end{array} +$$ + +$\hat{y}_i^t = 1$ or 0 means pseudo-labeled as private/common data. + +Feature Consensus Regularization. Recent studies [41, 52, 53] in SFDA and SF-UniDA observe that incorporating consensus regularization with nearest neighbors in feature space could assist models in achieving more stable performance. Following them, we integrate the objective $\mathcal{L}_{con}$ into LEAD. More details are included in the supplementary. + +Overall Optimization Objective. Taking into account the discussions above, the eventual loss function is defined as: + +$$ +\mathcal {L} = \lambda \cdot \mathcal {L} _ {c e} + \mathcal {L} _ {r e g} + \mathcal {L} _ {c o n} \tag {7} +$$ + +where $\lambda > 0$ is a trade-off hyper-parameter. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. We empirically verify the effectiveness of our LEAD on four datasets, including Office-31 [43], OfficeHome [49], VisDA [36], and DomainNet [37]. For a + +Table 2. H-score (\%) comparison in OPDA scenario on Office-Home. U denotes methods applicable for all potential label-shift scenarios (i.e., OPDA, OSDA, and PDA). SF represents source data-free. CF indicates model adaptation without time-consuming K-means clustering. + +
MethodsUSFCFAr2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvg
CMU [15]✘✘✓56.056.959.267.064.367.854.751.166.468.257.969.761.6
DANCE [47]✓✘✓61.060.464.965.758.861.873.161.266.667.762.463.763.9
DCC [25]✓✘✘58.054.158.074.670.677.564.373.674.981.075.180.470.2
OVANet [44]✘✘✓62.875.678.670.768.875.071.358.680.576.164.178.971.8
GATE [7]✓✘✓63.875.981.474.072.179.874.770.382.779.171.581.775.6
UniOT [5]✘✘✓67.380.586.073.577.384.375.563.386.077.865.481.976.6
Source-only---47.371.681.951.557.269.456.040.376.661.444.273.560.9
SHOT-O [27]✘✓✓32.929.539.656.830.141.154.935.442.358.533.533.340.7
LEAD✓✓✓62.778.186.470.676.383.475.360.686.275.460.783.775.0
UMAD [28]✘✓✓61.176.382.770.767.775.764.455.776.373.260.477.270.1
+ LEAD✓✓✓65.282.687.478.076.587.579.865.088.479.463.083.578.0 (+7.9)
GLC [41]✓✓✘64.378.289.863.181.789.177.654.288.980.754.285.975.6
+ LEAD✓✓✘65.779.788.765.080.388.577.657.488.978.957.885.576.2 (+0.6)
+ +Table 3. H-score (\%) comparison in OPDA scenario on Office-31, VisDA, and DomainNet. Some results are cited from UMAD [28]. + +
MethodsUSFCFOffice-31DomainNetVisDA
A2DA2WD2AD2WW2AW2DAvgP2RP2SR2PR2SS2PS2RAvgS2R
CMU [15]XX✓68.167.371.479.372.280.473.150.845.152.245.644.851.048.332.9
DANCE [47]✓X✓78.671.579.991.472.287.980.321.037.047.346.727.721.033.542.8
DCC [25]✓XX88.578.570.279.375.988.680.256.943.750.343.344.956.249.243.0
OVANet [44]XX✓85.879.480.195.484.094.386.556.047.151.744.947.457.250.753.1
GATE [7]✓X✓87.781.684.294.883.494.187.657.448.752.847.649.556.352.156.4
UniOT [5]XX✓87.088.588.498.887.696.691.159.351.847.848.346.858.352.057.3
Source-only---70.963.239.677.352.286.464.957.338.247.838.432.248.243.725.7
SHOT-O [27]X✓✓73.567.259.388.377.184.475.035.030.837.228.331.932.232.644.0
LEAD✓✓✓85.485.086.390.986.293.187.859.946.151.345.045.956.350.876.6
UMAD [28]X✓✓79.177.487.490.790.497.287.059.044.350.142.132.055.347.158.3
+ LEAD✓✓✓88.789.290.895.890.497.192.0 (+5.0)59.346.049.843.843.956.049.8 (+2.7)67.2 (+8.9)
GLC [41]✓✓X81.584.589.890.488.492.387.863.350.554.950.949.661.355.173.1
+ LEAD✓✓X83.086.887.692.087.193.288.3 (+0.5)63.550.755.151.649.961.355.4 (+0.3)76.8 (+3.7)
+ +fair comparison, we evaluate LEAD on partial-set domain adaptation (PDA), open-set domain adaptation (OSDA), and open-partial-set domain adaptation (OPDA). Details of classes split are summarized in Table 1. + +Evaluation. We utilize the same evaluation metric as previous works. In PDA scenarios, we report the classification accuracy for evaluation. In OSDA and OPDA scenarios, we report the H-score for performance comparison, which is a harmonic mean of accuracy over common data and accuracy over private data. It is worth noting that LEAD can be incorporated effortlessly into most existing SF-UniDA methods, for potential performance improvement. To demonstrate this merit, we integrate LEAD with two representative methods: UMAD [28] and GLC [41]. + +Implementation Details. All the experiments are conducted on a single RTX-3090 GPU with PyTorch-1.10. We adopt the same network architecture as existing baseline methods. During target model adaptation, we apply the SGD optimizer with momentum 0.9. The batch size is set to 64 for all experiments. The learning rate is set to 1e-3 for Office-31 and Office-Home, and 1e-4 for VisDA and DomainNet. As for $\lambda$ , it is set to 0.3 for Office-31, 1.0 for VisDA, and 2.0 for Office-Home and DomainNet. During inference, we adopt the same protocol like [41]. More details are presented in the supplementary. + +# 4.2. Experimental Results + +Results for OPDA Scenarios. To verify the effectiveness of LEAD, we first conduct experiments in the most challenging OPDA scenarios, where both source and target domains contain private data. Table 2 summarizes the results on Office-Home, and Table 3 details the results on Office-31, DomainNet, and VisDA. These results demonstrate that LEAD achieves performance surpassing or comparable to existing methods based on K-means clustering for segregating common and private data. In comparison with methods without clustering, LEAD achieves significantly superior performance. For instance, on Office-Home, DomainNet, and VisDA datasets, LEAD outperforms UMAD by $4.9\%$ , $3.7\%$ , and $18.3\%$ , respectively. We attribute this to our instance-level adaptive decision strategy, which effectively circumvents the limitations imposed by a fixed threshold. While GLC achieves commendable performance on its own, the integration with LEAD further amplifies its effectiveness. For example, LEAD improves GLC in the H-score on VisDA from $73.1\%$ to $76.8\%$ . The integration with LEAD proves even more significant against UMAD, with the H-score on Office-Home and VisDA boosting from $70.1\% /58.3\%$ to $78.0\% /67.2\%$ , respectively. + +Results for OSDA Scenarios. We then conduct experiments in OSDA scenarios, where only the target domain + +Table 4. H-score (%) comparison in OSDA scenario on Office-Home, Office-31, and VisDA. + +
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X Y55.057.059.059.358.260.659.251.361.261.953.555.357.652.655.776.575.965.864.765.254.2
DANCE [47]✓ X Y6.59.09.920.410.49.228.412.812.614.27.913.212.984.978.879.178.868.378.879.867.5
DCC [25]✓ X X56.167.566.749.666.564.055.853.070.561.657.271.961.758.354.867.289.485.380.972.759.6
OVANet [44]X X Y58.666.369.962.065.268.659.853.469.368.759.666.764.090.588.386.798.288.398.491.766.1
GATE [7]✓ X Y63.870.575.866.467.971.767.361.576.070.461.875.169.088.486.584.295.086.196.789.570.8
Source-only- - -46.163.372.942.854.058.747.836.166.260.845.368.255.278.272.144.282.252.188.869.629.1
SHOT-O [27]X Y Y37.741.848.456.439.840.960.041.549.761.841.443.646.980.271.664.393.164.091.877.528.1
LEAD✓ Y Y60.770.876.561.068.670.865.559.874.264.857.775.867.284.985.190.294.890.396.590.374.2
UMAD [28]X Y Y59.271.876.663.569.071.962.554.672.866.557.970.766.488.584.486.895.088.295.989.866.8
+ LEAD✓ Y Y62.673.677.361.366.270.763.354.675.166.763.575.667.690.489.690.395.490.797.392.370.2
+2.5
+ +Table 5. Accuracy comparison (%) in PDA scenario on Office-Home, Office-31, and VisDA. + +
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X X50.974.278.462.264.172.563.547.978.372.454.778.966.584.184.269.297.266.898.883.465.5
DANCE [47]✓ X X53.673.284.970.867.382.670.050.984.877.055.981.871.177.171.283.794.692.696.886.073.7
DCC [25]✓ X X54.247.557.583.871.686.263.765.075.285.578.282.670.987.381.395.4100.095.5100.093.372.4
OVANet [44]X X X34.154.672.142.447.355.938.226.261.756.735.868.949.569.461.761.490.266.498.774.634.3
GATE [7]✓ X X55.875.985.373.670.283.072.159.584.779.663.983.874.089.586.293.5100.094.498.693.775.6
Source-only- - -45.969.281.155.761.264.860.741.175.870.549.978.462.990.479.379.395.984.398.187.842.8
SHOT-P [27]X X Y64.785.190.175.173.984.276.464.190.380.763.385.577.889.884.492.296.692.299.492.474.2
LEAD✓ Y Y58.283.187.070.575.483.373.750.483.778.358.783.273.889.893.995.698.696.099.495.575.3
UMAD [28]X Y Y51.266.579.263.162.968.263.356.475.974.555.978.366.385.485.183.597.686.299.489.568.5
+ LEAD✓ Y Y61.178.287.570.578.581.874.160.483.378.863.282.175.089.290.890.698.093.0100.093.678.1
+8.7+4.1+9.6
GLC [41]✓ Y X55.979.087.572.571.882.774.941.782.477.360.484.372.589.889.892.896.696.199.494.176.2
+ LEAD✓ Y X56.681.288.172.572.283.275.343.683.478.259.784.273.289.893.295.696.996.099.495.175.5
+0.7+1.0-0.7
+ +involves private data. Table 4 lists the results, demonstrating the superiority of LEAD against baselines. Specifically, LEAD obtains the H-score of $67.2\%$ , $90.3\%$ , and $74.2\%$ on Office-Home, Office-31, and VisDA datasets, respectively. Similar to the observations in OPDA scenarios, LEAD also exhibits performance improvement to existing methods. LEAD improves the H-score of UMAD by $1.2\%$ , $2.5\%$ , and $3.4\%$ across the Office-Home, Office-31, and VisDA datasets, respectively. The results again verify the merit of LEAD when incorporated with existing methods. + +Results for PDA Scenarios. Finally, we verify the effectiveness of LEAD in PDA scenarios, where the target domain's label sets are subsets of those in the source domain. Results presented in Table 5 show that LEAD achieves superior performance even compared to methods tailored for PDA. In particular, LEAD attains the overall accuracy of $73.8\%$ , $95.5\%$ , and $75.3\%$ on Office-Home, Office-31, and VisDA datasets, individually. LEAD contributes the accuracy increase of $8.7\%$ , $4.1\%$ , and $9.6\%$ to UMAD on these datasets. Note that, there is a slight degradation $(-0.7\%)$ for GLC w/ LEAD. It may arise from gradient conflict, causing the optimization process to inadvertently compromise one objective in favor of improving the overall objective. We will pursue better integration strategies in future work. + +Table 6. Ablation studies of LEAD in OPDA scenarios. + +
LceLregLconOffice-31Office-HomeVisDA
---64.960.925.7
✓--82.073.661.5
✓✓-84.574.567.3
✓-✓87.274.575.7
✓✓✓87.875.076.6
+ +# 4.3. Experimental Analysis + +Ablation Study. To assess the contribution of different components within LEAD, we carry out extensive ablation studies in OPDA scenarios. The results, detailed in Table 6, confirm that each component is not only effective but also complementary to the others. + +Efficiency Comparison. As previously stated, LEAD employs the distance to target prototypes and source anchors for adaptive pseudo-labeling. This design, in contrast to clustering-based methods, notably reduces the need for computational resources. Table 7 juxtaposes the runtime of GLC [41] and LEAD in OPDA scenarios. It is clear that LEAD requires minimal resources for deriving decision boundaries, whereas GLC's resource demands escalate dramatically with data scale. In particular, on DomainNet, with about 0.6 million images across 345 categories, GLC averages 897.65 seconds, while LEAD takes only 0.29 sec + +![](images/272dc3cb2d8c736b11cf92868aefcd5a1556bce19c89b588cbf4274e63087ebf.jpg) +(a) Sensitivity to $\lambda$ + +![](images/039a3a74b32629a546ea44434469cf1ba504d2e9fc48a5d3968bb893703994ba.jpg) +(b) H-score convergence + +![](images/5adf2f4b6957ee926f84826bf54253e771ca8d412b05a8741ebed93b861bb315.jpg) +(c) $\mathrm{Ar}\rightarrow \mathrm{Pr}$ in OPDA + +![](images/a3f8ae72befbceebaf4610de0e38520614dac47bcf9abf0b8500ba7a595ff590.jpg) +(d) $\mathrm{Re}\rightarrow \mathrm{Cl}$ in OPDA + +Table 7. Runtime (s) to derive pseudo-label decision boundaries. + +
Office-31Office-HomeVisDADomainNet
GLC [41]0.190.510.46897.65
LEAD0.020.020.110.29
+ +onds on the same platform. Moreover, LEAD also shows substantial resource savings even on normal-scale datasets, e.g., on VisDA with over $75\%$ reduction. + +Parameter Sensitivity. We investigate the parameter sensitivity of $\lambda$ on Office-31 in OPDA scenarios, where the range of $\lambda$ is set to [0.1, 0.2, 0.3, 0.4, 0.5]. Figure 3 (a) illustrates that the results are stable around the selected parameter $\lambda = 0.3$ . Through oracle validation, we may find the optimal parameter configuration for this setting is $\lambda = 0.1$ . + +Training Stability. Figure 3 (b) visualizes the training curves of our LEAD on VisDA in OPDA scenarios. Generally, the training procedure of LEAD is stable and effective. + +Varying Unknown Classes. With the rise in unknown private classes, accurately distinguishing common from private data becomes more challenging. To examine the robustness of our LEAD, we compared it with GLC and UMAD in OPDA scenarios using the Office-Home dataset. The results in Figure 3 (c-d) show that LEAD achieves stable results and consistently boosts existing methods. + +Effect of Instance-level Decision Boundary. To evaluate the contribution of our instance-level pseudo-labeling strategy, we conduct a comparison against the intuitive design, which uses the average of $\mu_{\text{com}}$ and $\mu_{\text{pri}}$ as the threshold. As depicted in Figure 4 (a), instance-level decision boundary strategy significantly outperforms the vanilla global decision design (75.0% vs 63.1% on Office-Home in OPDA scenarios). We attribute this to that vanilla design discounts the inconsistency of covariate shifts across categories, and also the variability among samples from the same category. + +Exploration with Entropy as the Indicator. Contrasting with existing threshold-based methods that utilize Entropy as the indicator for differentiating common and private data, we propose decomposing the feature and using the projection $\mathbf{z}_{i,unk}^{t}$ on the source-unknown space as an alternative indicator. To empirically substantiate the superiority, we remold our framework to incorporate Entropy as the indicator. Figure 4 (b) depicts the results. An observation is that while employing entropy as the indicator within our framework is + +![](images/3f2ae93e6a3123e54df700f932bcee05fe553c6c6eeed04979dfea9a7f441902.jpg) +Figure 3. Robustness analysis. (a) shows the sensitivity to $\lambda$ on Office-31. (b) presents the H-score curves on VisDA. (c-d) present robustness analysis when varying the unknown private categories. +(a) Effect of Instance-level Decision Boundary +Figure 4. Methodology analysis. (a) compares the effectiveness of our instance-level decision strategy against the vanilla global decision strategy. (b) examines the efficacy of using Entropy as the indicator for private data. + +![](images/862ba535973c6ca976f54376d3e3fe936843dd25a2033bfb58ea613f8a1c298d.jpg) +(b) Exploration with Entropy as the Indicator + +feasible, its performance is inferior compared to $\mathbf{z}_{i,unk}^{t}$ . We speculate that this inferiority might result from entropy's limited capacity to distinguish between common samples with high uncertainty and private unknown samples. + +# 5. Conclusion + +In this paper, we delve into source-free universal domain adaptation (SF-UniDA). Different from existing methods that leverage hand-crafted thresholding strategies or feature-clustering algorithms, we propose a novel idea of LEArning Decomposition (LEAD) tailored for SF-UniDA. Technically, LEAD first decomposes features into source-known and -unknown two parts, then establishes instance-level decision boundaries for pseudo-labeling. It offers an elegant view to identify target-private unknown data without tedious tuning a threshold or relying on unstable clustering. Extensive experiments across various scenarios have verified the effectiveness of our LEAD. Besides, it is appealing in view that LEAD can be seamlessly incorporated into existing methods to further boost performance. Remarkably, LEAD improves UMAC by $7.9\%$ overall H-score in the OPDA scenario on Office-Home dataset. + +Acknowledgment: This work is supported by the National Natural Science Foundation of China (No. 62372329), in part by the National Key Research and Development Program of China (No. 2021YFB2501104), in part by Shanghai Rising Star Program (No.21QC1400900), in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program, and in part by the International Exchange Program for Graduate Students, Tongji University (No.2023020053). + +# References + +[1] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018. 4 +[2] Richard Bellman. Dynamic programming. Science, 153 (3731):34-37, 1966. 2 +[3] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 1, 2 +[4] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 1, 2 +[5] Wanxing Chang, Ye Shi, Hoang Tuan, and Jingya Wang. Unified optimal transport framework for universal domain adaptation. In NeurIPS, 2022. 2, 6 +[6] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In ECCV. Springer, 2020. 2 +[7] Liang Chen, Yihang Lou, Jianzhong He, Tao Bai, and Minghua Deng. Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation. In CVPR, 2022. 1, 2, 6, 7 +[8] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, 2018. 2 +[9] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In ICCV, 2017. 1 +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1 +[11] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE TPAMI, 39(9):1853-1865, 2016. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1 +[13] Qi Dou, Cheng Ouyang, Cheng Chen, Hao Chen, and Pheng-Ann Heng. Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss. In IJCAI, 2018. 1 +[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE TIP, 30:6473-6484, 2021. 2 +[15] Bo Fu, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Learning to detect open classes for universal domain adaptation. In ECCV. Springer, 2020. 2, 6, 7 +[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, pages 2096-2030, 2016. 1, 2 + +[17] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1 +[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2 +[20] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML. PMLR, 2018. 1 +[21] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1 +[23] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE TPAMI, pages 766-785, 2019. 1, 2 +[24] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 2 +[25] Guangrui Li, Guoliang Kang, Yi Zhu, Yunchao Wei, and Yi Yang. Domain consensus clustering for universal domain adaptation. In CVPR, 2021. 1, 2, 3, 6, 7 +[26] Shuang Li, Shiji Song, Gao Huang, Zhengming Ding, and Cheng Wu. Domain invariant and class discriminative feature learning for visual domain adaptation. IEEE TIP, 2018. 2 +[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML. PMLR, 2020. 1, 2, 3, 6, 7 +[28] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 1, 2, 3, 6, 7 +[29] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2023. 4 +[30] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE TMM, 2022. 2 +[31] Xinyu Liu and Yixuan Yuan. A source-free domain adaptive polyp detection framework with style diversification flow. IEEE TMI, pages 1897-1908, 2022. 1 +[32] Zhengfa Liu, Guang Chen, Zhijun Li, Yu Kang, Sanqing Qu, and Changjun Jiang. Psdc: A prototype-based shared-dummy classifier model for open-set domain adaptation. IEEE Transactions on Cybernetics, 2022. 2 + +[33] Zhengfa Liu, Guang Chen, Zhijun Li, Sanqing Qu, Alois Knoll, and Changjun Jiang. D2ifln: Disentangled domain-invariant feature learning networks for domain generalization. IEEE Transactions on Cognitive and Developmental Systems, 2023. 2 +[34] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In ICCV, 2013. 2 +[35] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 1, 2 +[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.5 +[37] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 5 +[38] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML. PMLR, 2020. 2 +[39] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 2 +[40] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In CVPR, 2023. 2 +[41] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1 +[43] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5 +[44] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In ICCV, 2021. 1, 2, 6, 7 +[45] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1 +[46] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 1, 2 +[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. In NeurIPS, pages 16282-16292, 2020. 1, 2, 3, 6, 7 +[48] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In WACV, 2021. 2 +[49] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5 +[50] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1 + +[51] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 4 +[52] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 5 +[53] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 5 +[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In CVPR, 2019. 1, 2 \ No newline at end of file diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/images.zip b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3c622ed3abbaf19f3755da7f0fa1ad818b65a43b --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a27612265ffe0b212857c0b3305f21f218170e23a235e1afaddccc732343f19 +size 853322 diff --git a/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/layout.json b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6632b45eeb9d763b1fd4702ecac9374545217c9 --- /dev/null +++ b/2024/LEAD_ Learning Decomposition for Source-free Universal Domain Adaptation/layout.json @@ -0,0 +1,10090 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "type": "text", + "content": "LEAD: Learning Decomposition for Source-free Universal Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": "Sanqing " + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Qu}^1" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Tianpei " + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{Zou}^1" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Lianghua " + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "\\mathrm{He}^1" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Florian Röhrbein" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Alois Knoll" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Guang Chen" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "text", + "content": ", Changjun Jiang" + }, + { + "bbox": [ + 143, + 142, + 449, + 172 + ], + "type": "inline_equation", + "content": "^1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "text", + "content": "Tongji University, " + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "text", + "content": "Chemnitz University of Technology, " + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 173, + 462, + 185 + ], + "type": "text", + "content": "Technical University of Munich" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "text", + "content": "Universal Domain Adaptation (UniDA) targets knowledge transfer in the presence of both covariate and label shifts. Recently, Source-free Universal Domain Adaptation (SF-UniDA) has emerged to achieve UniDA without access to source data, which tends to be more practical due to data protection policies. The main challenges lie in determining whether covariate-shifted samples belong to target-private unknown categories. Existing methods tackle this either through hand-crafted thresholding or by developing time-consuming iterative clustering strategies. In this paper, we propose a new idea of LEArning Decomposition (LEAD), which decouples features into source-known and -unknown components to identify target-private data. Technically, LEAD initially leverages the orthogonal decomposition analysis for feature decomposition. Then, LEAD builds instance-level decision boundaries to adaptively identify target-private data. Extensive experiments across various UniDA scenarios have demonstrated the effectiveness and superiority of LEAD. Notably, in the OPDA scenario on VisDA dataset, LEAD outperforms GLC by " + }, + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "text", + "content": " overall H-score and reduces " + }, + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 47, + 234, + 290, + 534 + ], + "type": "text", + "content": " time to derive pseudo-labeling decision boundaries. Besides, LEAD is also appealing in that it is complementary to most existing methods. The code is available at https://github.com/ispc-lab/LEAD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 556 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 564, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 287, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have yielded impressive results in a variety of computer vision tasks [12, 18, 22, 42]. However, DNNs often fail to generalize well to new domains due to the discrepancy between training (source) and test (target) data distributions (i.e., covariate shift). Such a shift poses significant challenges for some safety-critical applications like autonomous driving [9, 10], or medical imaging [13, 31]. Unsupervised Domain Adaptation (DA) [16, 23] aims to address this challenge by transferring task-specific knowledge from a labeled source domain to an unlabeled target domain. Despite promising results, most" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 206, + 430, + 305 + ], + "blocks": [ + { + "bbox": [ + 309, + 206, + 430, + 305 + ], + "lines": [ + { + "bbox": [ + 309, + 206, + 430, + 305 + ], + "spans": [ + { + "bbox": [ + 309, + 206, + 430, + 305 + ], + "type": "image", + "image_path": "eedb42aa8eec51aebbbdf6588d9e6093c0cee0ecfc340e4655e0416bee70ea83.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 308, + 547, + 429 + ], + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 429 + ], + "type": "text", + "content": "Figure 1. (a) Pseudo-label accuracy curves for target-private data in the OPDA scenario on VisDA dataset. An observation is that clustering-based GLC does not work well in deriving decision boundaries for target-private data. This may be due to the curse of dimensionality, leading K-means to faintly discerning clusters. (b) Frequency distribution for the normalized feature magnitude of target data in the source-unknown space (orthogonal complement of the space spanned by source model weights). Features are on task " + }, + { + "bbox": [ + 304, + 308, + 547, + 429 + ], + "type": "inline_equation", + "content": "(A\\to D)" + }, + { + "bbox": [ + 304, + 308, + 547, + 429 + ], + "type": "text", + "content": " in the OPDA scenario of Office-31. The results show that target-private data are expected to involve more components from source-unknown space, even with covariate shifts." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 430, + 206, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 430, + 206, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 430, + 206, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 430, + 206, + 544, + 305 + ], + "type": "image", + "image_path": "311904e7893df586bad37f09500ec8d8c0e403d392f568a50a4fba25e4cff427.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 483 + ], + "type": "text", + "content": "existing favorable DA methods [16, 20, 27, 45] assume that the label spaces are identical across the source and target domain, thus being only applicable to closed-set scenarios and limiting applicability to practical applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 484, + 548, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 548, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 548, + 674 + ], + "type": "text", + "content": "To handle generalized cases, Universal Domain Adaptation (UniDA) [47] was proposed to allow all potential label shifts between source and target domains. Different from existing methods specialized for Partial Domain Adaptation (PDA) [3, 4], Open-set Domain Adaptation (OSDA) [35, 46], or Open-Partial Domain Adaptation (OPDA) [44, 54], in UniDA, we have no prior knowledge about the label shift, e.g., information on matched common classes or the number of categories in the target domain. Nonetheless, most existing UniDA works [7, 25, 47] necessitate concurrent access to source and target data. This requirement becomes increasingly impractical in light of strict data protection policies [50]. In this paper, we focus on Source-free Universal Domain Adaptation (SF-UniDA) [28, 41], where only a pre-trained source model is provided for knowledge transfer rather than labeled source data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "The primary objective for UniDA is to recognize target data belonging to common categories and separate them from target-private \"unknown\" data. Existing meth" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 223, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 223, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 223, + 713 + ], + "type": "text", + "content": "*Corresponding author: guangchen@tongji.edu.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23334" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 239 + ], + "type": "text", + "content": "ods have designed various criteria based on instance-level predictions, e.g., entropy-based [14, 47, 54], confidence-based [15, 24, 44]. However, defining a universal threshold for all scenarios is challenging and often sub-optimal. To mitigate tedious parameter-tuning, some methods start to perform source and target consensus clustering to achieve common and private data separation [7, 25]. Considering that SF-UniDA is inaccessible to source data, GLC [41] devises an iterative one-vs-all clustering strategy to realize target-private data identification. Regardless of progress, iterative clustering can be computationally intensive. Besides, due to the curse of dimensionality [2], methods like K-means are prone to overfitting and faintly discerning clusters. This can be implied to some degree from Figure 1 (a)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 240, + 286, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 240, + 286, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 240, + 286, + 525 + ], + "type": "text", + "content": "In this paper, we propose to address the SF-UniDA problem from a new perspective, i.e., feature decomposition. Our launching point is that, even though target data might have shifted in the feature space, target-private data from unknown categories are likely to exhibit a higher proportion of components from the orthogonal complement (source-unknown) space spanned by source model weights. The frequency distribution of normalized feature magnitudes in the source-unknown space depicted in Figure 1 (b) basically demonstrates our argument, which has been rarely explored in the literature. To materialize our idea, we propose the LEArning Decomposition (LEAD) framework. In particular, LEAD initially leverages the orthogonal decomposition to construct source-known and -unknown space. The projection on source-unknown space is then regarded as the discriminative representation with respect to target-private categories. Different from existing thresholding methods introducing a fixed global hand-crafted threshold for all scenarios, LEAD deliberately considers the distance to the target prototypes and source anchors to establish instance-level decision boundaries. We analyze and evaluate our LEAD under a variety of SF-UniDA scenarios, ranging from PDA, OSDA, to OPDA. Extensive experiments demonstrate the effectiveness and superiority." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "content": "The main contribution is the proposal of a LEArning Decomposition (LEAD) framework for source-free universal domain adaptation (SF-UniDA). This solution leads to elegant views for identifying target-private unknown data without tedious tuning thresholds or relying on iterative unstable clustering. Remarkably, in the OPDA scenario on VisDA, LEAD attains an H-score of " + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "inline_equation", + "content": "76.6\\%" + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "content": ", surpassing GLC [41] by " + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "content": ". Besides, LEAD is complementary to most existing SF-UniDA methods. For instance, in the OPDA scenario on Office-Home, LEAD advances UMAD [28] with H-score improvement from " + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "inline_equation", + "content": "70.1\\%" + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "inline_equation", + "content": "78.0\\%" + }, + { + "bbox": [ + 49, + 527, + 286, + 657 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 132, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 690, + 286, + 712 + ], + "type": "text", + "content": "Unsupervised Domain Adaptation. When DNNs encounter data from a distribution that differs from their train" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 544, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 215 + ], + "type": "text", + "content": "ing data, inevitable performance degeneration is commonly observed. To address this issue, domain adaptation (DA) [16, 23] has been introduced, leveraging labeled data from source domains to train DNNs for unlabeled target domains in a transductive learning manner. Existing DA methods can be broadly categorized into two paradigms, namely moment matching [11, 34, 48] and adversarial learning [16, 19, 21]. Regardless of their effectiveness in various applications such as object recognition [16, 27, 39], semantic segmentation [19, 48], and object detection [8, 21], most existing approaches assume that label spaces are identical across the source and target domains, limiting their applicability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 220, + 544, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 544, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 544, + 529 + ], + "type": "text", + "content": "Universal Domain Adaptation. There have been some approaches introduced to deal with label-shift scenarios, such as open-set domain adaptation (OSDA) [32, 35, 46], partial-set domain adaptation (PDA) [3, 4], and open-partial-set domain adaptation (OPDA) [15, 44, 54]. However, these methods are often tailored to specific scenarios and may not be readily applicable to other label shift situations. Universal domain adaptation (UniDA), on the other hand, is designed to facilitate adaptation across all potential label shift scenarios. Despite the progress in UniDA, many existing methods [5, 7, 44] necessitate simultaneous access to both source and target data, making them impractical for applications with stringent data protection policies. To address this challenge, source-free universal domain adaptation (SF-UniDA) [28, 41] has been proposed, wherein source data are only utilized for pre-training and are inaccessible during target adaptation. To promote the distinction between common and private data, existing methods often resort to manually defined thresholding or the introduction of global clustering techniques. Nevertheless, selecting an appropriate threshold for all scenarios can be both laborious and sub-optimal. Additionally, K-means clustering tends to be unstable when applied to high-dimensional data due to the curse of dimensionality [2]. Our work delves into these limitations and proposes a novel and elegant solution from the perspective of feature decomposition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 544, + 712 + ], + "type": "text", + "content": "Feature Decomposition. As a fundamental technique in machine learning, feature decomposition is applied to deconstruct complex data into simpler and more interpretable components. In the field of domain adaptation and domain generalization, feature decomposition has been leveraged in various studies [6, 26, 30, 33, 38, 40] to enhance feature alignment or facilitate domain-invariant feature learning by breaking down features into content and style components. Nonetheless, the content and style decomposition design is less applicable for UniDA, as the primary challenge lies in distinguishing common and private data. In this paper, we approach the problem from the perspective of orthogonal feature decomposition and employ Independent Component Analysis to decompose features into two orthogonal parts, thereby facilitating the intended objective." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "text", + "content": "23335" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 203, + 293 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 203, + 293 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 203, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 203, + 293 + ], + "type": "image", + "image_path": "eb8beafd7fe75af0f36aaa1ef4acfd447b286f8f39fe8c69684b04b60cc5fffe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": "Figure 2. (a) Illustrations of traditional universal domain adaptation (UniDA) and source-free universal domain adaptation (SF-UniDA). Traditional UniDA methods necessitate data from both source and target domains concurrently. In SF-UniDA, source data are solely utilized for pre-training. Adaptation is performed by harnessing target data and the source model " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{s} = h_{\\theta}^{s} \\circ g_{\\theta}^{s}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": ". (b) An overview of our Learning Decomposition (LEAD) framework. Pseudo-labeling is an important technique for UniDA and SF-UniDA. The primary objective is to recognize target data associated with common label sets and exclude data within the target-private label space. Different from existing methods that perform private data identification by hand-crafted thresholding on predictions or iterative global clustering, we tackle this from the viewpoint of feature decomposition. The rationale is that despite potential shifts in the feature space, target-private data are expected to encompass more components from the orthogonal complement (source-unknown) space of the source model. Technically, LEAD first performs orthogonal decomposition to decompose target features into source-known and -unknown parts, i.e., " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,knw}^{t}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": " is considered as an indicator for private data. Next, LEAD employs a two-component Gaussian Mixture Model to estimate the distribution of " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": ". Thereafter, LEAD devises a metric named \"common score\" " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": " that accounts for distances to both target prototypes and source anchors (derived from " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "h_{\\theta}^{t}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": ") to facilitate deriving instance-level decision boundary " + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "inline_equation", + "content": "\\rho_{i,c}" + }, + { + "bbox": [ + 46, + 297, + 545, + 451 + ], + "type": "text", + "content": ". LEAD provides an elegant solution to distinguish target-private data, mitigating the need for tedious hand-crafted threshold tuning or dependence on time-consuming iterative clustering. LEAD could also serve as a complementary approach to most existing SF-UniDA methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 70, + 531, + 294 + ], + "blocks": [ + { + "bbox": [ + 205, + 70, + 531, + 294 + ], + "lines": [ + { + "bbox": [ + 205, + 70, + 531, + 294 + ], + "spans": [ + { + "bbox": [ + 205, + 70, + 531, + 294 + ], + "type": "image", + "image_path": "7ecf330707fdfa614df79bb0c5548e32eddd9d563b3d72c33366cd7e75e9087c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 459, + 129, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 129, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 129, + 472 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 479, + 126, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 126, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 126, + 491 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": "In this paper, we consider a generalized and challenging case in domain adaptation: universal domain adaptation (UniDA), which aims to achieve knowledge transfer in the presence of both covariate and label shifts. In UniDA, there is a labeled source domain " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s = \\left\\{\\left(\\mathbf{x}_i^s,\\mathbf{y}_i^s\\right)\\right\\}_{i = 1}^{N_s}" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^s\\in \\mathcal{X}^s\\subset \\mathbb{R}^X,\\mathbf{y}_i^s\\in \\mathcal{Y}^s\\subset \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " , and an unlabeled target domain " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t = \\left\\{\\left(\\mathbf{x}_i^t,\\ref{eq:2.3.1}\\right)\\right\\}_{i = 1}^{N_t}" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t\\in \\mathcal{X}^t\\subset \\mathbb{R}^X." + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\mathcal{V}^s\\cap \\mathcal{V}^t" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " denotes the common label sets shared by source and target domains. " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}^s = \\mathcal{V}\\setminus \\mathcal{V}^t" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}^t = \\mathcal{V}\\setminus \\mathcal{V}^s" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " symbolizes the label sets private to source and target domain, individually. UniDA supposes that we have no prior knowledge about " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^t" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}^t" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " are also unavailable. The objective is to recognize common data belonging to " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "text", + "content": " and reject private data in " + }, + { + "bbox": [ + 46, + 498, + 286, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}^t" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Different from traditional UniDA methods that need concurrent access to source and target data, under the source-free universal domain adaptation (SF-UniDA) setting, " + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": " is only applicable for source model pre-training and not acces-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": "sible during adaptation. Assuming that the source model is represented as " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}^{s} = h_{\\theta}^{s}\\circ g_{\\theta}^{s}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{s}:\\mathbb{R}^{X}\\to \\mathbb{R}^{D}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " denotes the feature extractor and " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "h_\\theta^s:\\mathbb{R}^D\\to \\mathbb{R}^C" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " is the classifier module. Following existing methods [27, 41], we freeze the classifier module " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "h_\\theta^t = h_\\theta^s" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " and only learn a target-specific feature extractor " + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{t}" + }, + { + "bbox": [ + 305, + 460, + 545, + 544 + ], + "type": "text", + "content": " to realize model adaptation. Figure 2 (a) presents an illustration for UniDA and SF-UniDA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "The main challenge is how to distinguish common data from private data. Prior methods [25, 28, 41, 47] have designed various algorithms to encourage the separation. However, these methods perform data identification either based on hand-crafted thresholding criteria or by introducing time-consuming unstable clustering strategies. We propose to mitigate these limitations from the perspective of feature decomposition. An observation is that even though target data may have shifted in the feature space, the features of private data still include more components from the orthogonal complement (source-unknown) space of the pre-trained model. Technically, we present a novel learning decomposition (LEAD) framework. LEAD capitalizes on the orthogonal decomposition to build two orthogonal fea" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23336" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "ture spaces, i.e., source-known and -unknown space. Feature projection on source-unknown space is extracted as the descriptor for private data. Then, LEAD considers the distances to both the target prototypes and source anchors to establish instance-level decision boundaries. Figure 2 (b) manifests the architecture of our LEAD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 152, + 236, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 152, + 236, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 236, + 165 + ], + "type": "text", + "content": "3.2. Orthogonal Feature Decomposition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "text", + "content": "Pseudo-labeling serves as an important technique in unsupervised learning, yet many existing strategies only account for closed-set scenarios, disregarding the label-shift situations, which impairs their ability to differentiate target-private data. To fulfill the objective of common and private data separation, we propose to orthogonally decompose the feature representation into two uncorrelated parts, i.e., features associated with source-known space " + }, + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "text", + "content": " and those related to source-unknown space " + }, + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 46, + 170, + 289, + 386 + ], + "type": "text", + "content": ". This design stems from a simple inductive bias in favor of private data identification. Due to the lack of exposure to target data during source pre-training, particularly target-private data, the source-known space does not include any significant attributes of target-private data. Subsequently, after normalization, target-private data are expected to contain more components from source-unknown space, even with covariate shifts. This hypothesis is basically demonstrated by empirical evidence, as illustrated in Figure 1 (b)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": "Formally, denoting the weight of classifier " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "h_{\\theta}^{t}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{cls} \\in \\mathbb{R}^{C \\times D}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Source-known space " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " can be represented as the vector space of span " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_n\\}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " is a group weight vectors of " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{cls}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Source-unknown space " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " is the orthogonal complement space of " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D = F_{knw} \\oplus F_{unk}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Nevertheless, unit orthogonality in " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_n \\in \\mathbb{R}^D | n = 1, \\dots, C\\}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " is not guaranteed (i.e., when " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "n \\neq m" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_n, \\mathbf{w}_m \\neq 0" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": "), making it hard to derive " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Considering that " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " are the row space and null space of " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{cls}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Therefore, we employ the singular value decomposition (SVD) on " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{cls}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 47, + 386, + 289, + 518 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 526, + 173, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 526, + 173, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 173, + 540 + ], + "type": "interline_equation", + "content": "W _ {c l s} = U \\Sigma V ^ {T},", + "image_path": "483536394f606d5edcc128614ce32296eef7d7130346c583ec4cb53dc608d199.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 542, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 542, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 104, + 542, + 287, + 555 + ], + "type": "interline_equation", + "content": "F _ {k n w} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = 1, \\dots , C} \\right\\}, \\tag {1}", + "image_path": "cd176e2cf29ec0c7e4388865d098bf6fb7fbd12e9e26ffc81dbcfcdda0a36a1b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 558, + 231, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 558, + 231, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 231, + 570 + ], + "type": "interline_equation", + "content": "F _ {u n k} = \\operatorname {s p a n} \\left\\{\\mathbf {v} _ {n} | _ {n = C + 1, \\dots , D} \\right\\}", + "image_path": "be64de9e90916489b44f186b9d1db3f6bedbaefb1f4d7fe064660ea40b81b30a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "content": " is a diagonal matrix, " + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "inline_equation", + "content": "U\\in \\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "inline_equation", + "content": "V\\in \\mathbb{R}^{D\\times D}" + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "content": " are both orthogonal unitary matrices. " + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_n|\\mathbf{v}_n\\in \\mathbb{R}^D\\}_{n = 1}^D" + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "text", + "content": " are columns of " + }, + { + "bbox": [ + 47, + 572, + 287, + 610 + ], + "type": "inline_equation", + "content": "V" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": "Assuming that " + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": " is the feature of target data " + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": " extracted from feature extractor " + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{t}" + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": " after normalization, i.e., " + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t = g_\\theta^t (\\mathbf{x}_i^t) / ||g_\\theta^t (\\mathbf{x}_i^t)||_2" + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 46, + 610, + 287, + 658 + ], + "type": "text", + "content": " can be represented as the weighted summation of two orthogonal bases:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 663, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 663, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 97, + 663, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} ^ {t} = \\underbrace {\\sum_ {n = 1} ^ {C} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, k n w} ^ {t}} + \\underbrace {\\sum_ {n = C + 1} ^ {D} l _ {n} \\cdot \\mathbf {v} _ {n}} _ {\\mathbf {z} _ {i, u n k} ^ {t}} \\tag {2}", + "image_path": "aae03050fc088362f439f27af1b8041630867a17e9d2ea618678d3fe43f7f7d0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "||\\mathbf{z}_{i,knw}^{t}||_{2}^{2} + ||\\mathbf{z}_{i,unk}^{t}||_{2}^{2} = 1" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,knw}^{t}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " are the projections of " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{t}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "F_{unk}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "l_{n} \\in \\mathbb{R}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " denotes the projection weight of " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i}^{t}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": " onto the unit base " + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{n}" + }, + { + "bbox": [ + 304, + 70, + 547, + 147 + ], + "type": "text", + "content": ". This procedure is commonly recognized as orthogonal decomposition and has been extensively studied in the context of Independent Component Analysis (ICA)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 154, + 504, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 154, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 154, + 504, + 167 + ], + "type": "text", + "content": "3.3. Adaptive Separation for Private Data" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": "Intuitively, enabled by feature decomposition, we could differentiate between common and private data by examining the dominance of " + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": ". However, such a straightforward approach is not applicable to SF-UniDA due to the impact of distribution covariate shift, which often deteriorates the significance of " + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,knw}^{t}" + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": " for common data, leading to the prevalence of " + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": " in the representation. Additionally, the feature extractor inevitably captures representations that are independent of " + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "inline_equation", + "content": "F_{knw}" + }, + { + "bbox": [ + 304, + 172, + 545, + 305 + ], + "type": "text", + "content": ", including background contexts [1], object textures [17], and high-frequency patterns [51], further exacerbating this issue." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": "To solve this, one feasible solution is to estimate the distribution of " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": " to effectively calibrate these disturbances. Our observations have revealed that the empirical distributions of " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": " exhibit a bimodal pattern, with the presence of two distinct peaks typically indicating the modes of common and private data. Consequently, we employ a two-component Gaussian Mixture Model (GMM) to estimate the distribution " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": ", with the components featuring a lower mean representing the common data and those with a higher mean corresponding to the private data. In particular, we denote " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\mu_{com}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\mu_{pri}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": " as the expectation for common and private data, where " + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\mu_{com} < \\mu_{pri}" + }, + { + "bbox": [ + 304, + 306, + 546, + 450 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "content": "In the light of " + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mu_{com}" + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\mu_{pri}" + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "content": ", an intuitive practice is to treat all common and private categories equally. Accordingly, common and private data can be separated by establishing " + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "inline_equation", + "content": "(\\mu_{com} + \\mu_{pri}) / 2" + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "content": " as the threshold, where data with " + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^t\\| _2" + }, + { + "bbox": [ + 304, + 450, + 547, + 593 + ], + "type": "text", + "content": " below the threshold are recognized as common data, while those above are identified as private data. Despite its reasonable effectiveness, it discounts the inconsistency of covariate shifts across each category. Moreover, it fails to account for the variability in covariate shifts even among samples from the same category. These considerations motivate us to devise an instance-level decision strategy to achieve accurate and effective data identification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Concretely, we first leverage top- " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " sampling to construct target prototypes, denoted as " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_c^t\\in \\mathbb{R}^D |_{c = 1,\\dots ,C}\\}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " for each common category, adhering to the procedure in [29, 41]. Similar to [41], we empirically set " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "K = N_{t} / \\tilde{C}_{t}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " represents the amount of target data, and " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{C}_t" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " signifies the estimated count of target categories. Diverging from [41], we additionally collect source anchors from " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "W_{cls}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_c^s\\in \\mathbb{R}^D |_{c = 1,\\ldots ,C}\\}" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " to assist us handle scenarios involving source-private categories. The construct of target prototypes through top- " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " sampling fails to consider these sce" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23337" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "narios, making it susceptible to misidentifying private data as common. Further details regarding the target prototype construction are included in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "content": "Thereafter, we propose a novel metric, common score " + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}" + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "content": ", which deliberately considers the distance of " + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i^t" + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "content": " to the target prototype " + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_c^t" + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "content": " and source anchor " + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_c^s" + }, + { + "bbox": [ + 46, + 109, + 287, + 168 + ], + "type": "text", + "content": " to facilitate us in deriving instance-level decision boundaries. Formally, this metric is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 173, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 173, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 88, + 173, + 287, + 242 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d (\\mathbf {a}, \\mathbf {b}) = 1. 0 - \\cos (\\mathbf {a}, \\mathbf {b}), \\\\ \\epsilon_ {i, c} ^ {t} = 1. 0 - \\exp \\left(d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {t}\\right) - 1. 0\\right), \\\\ \\epsilon_ {i, c} ^ {s} = \\exp \\left(- d \\left(\\mathbf {z} _ {i} ^ {t}, \\mathbf {c} _ {c} ^ {s}\\right)\\right), \\tag {3} \\\\ \\epsilon_ {i, c} = \\sqrt {\\epsilon_ {i , c} ^ {t} \\cdot \\epsilon_ {i , c} ^ {s}} \\\\ \\end{array}", + "image_path": "e156f22cf6c8ff17b135889a64db5a734faa5e63c5ba352420d6705425d7de6e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{t}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{s}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " denote the common score from target prototype and source anchor viewpoint, respectively. Both " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{t}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{s}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " are clipped to the [0, 1] before fusion. We then utilize geometric mean to calculate fused common score, due to its advantage in multi-criteria decision-making. Note that we leverage different formulas to calculate " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{t}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{s}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": ". Under the same condition, i.e., " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "d(\\mathbf{z}_i^t,\\mathbf{c}_c^s) = d(\\mathbf{z}_i^t,\\mathbf{c}_c^t)" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{s}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": " is greater than " + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}^{t}" + }, + { + "bbox": [ + 46, + 248, + 287, + 380 + ], + "type": "text", + "content": ". The spirit behind is that instances exhibiting proximity to source anchors are likely to be common data. Conversely, instances in close proximity to target prototypes do not guarantee their recognition." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "text", + "content": "Given the computed common score " + }, + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\epsilon_{i,c}" + }, + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "text", + "content": " and the estimated distribution for " + }, + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^t\\| _2" + }, + { + "bbox": [ + 46, + 380, + 287, + 429 + ], + "type": "text", + "content": " we introduce the following pseudo-labeling strategy to effectively distinguish between common and private data instances:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 436, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 436, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 71, + 436, + 287, + 483 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\rho_ {i, c} = \\mu_ {c} + \\epsilon_ {i, c} \\cdot (\\mu_ {p r i} - \\mu_ {c}), \\\\ \\tilde {\\mathbf {y}} _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\text {u n k n o w n}, & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} \\geq \\rho_ {i, c} \\\\ \\mathbb {1} (\\operatorname {a r g m a x} (\\epsilon_ {i})), & \\text {i f} \\| \\mathbf {z} _ {i, u n k} ^ {t} \\| _ {2} < \\rho_ {i, c} \\end{array} \\right. \\tag {4} \\\\ \\end{array}", + "image_path": "fe4b9409630bf132a3eba61b790c7eefde31604492d4748a53dea4073c44293f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mu_{c}" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " denotes the expectation of " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^{t}\\|_{2}" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": "-th category, derived using the identical top-" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " sampling technique employed during the construction of the target prototypes. " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\rho_{i,c}" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " represents the instance-level decision boundary, which accounts for the potential covariate shifts across each category, as well as the data instances corresponding to both target prototypes and source anchors. " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " denotes the one-hot encoding operator, " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_i^t" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": " is the pseudo-label for " + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 46, + 493, + 287, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 598, + 185, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 185, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 185, + 611 + ], + "type": "text", + "content": "3.4. Optimization Objectives" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 666 + ], + "type": "text", + "content": "To encourage the separation between common and private data, the optimization process involves three distinct objectives, including pseudo-label learning, feature decomposition regularization, and feature consensus regularization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Pseudo-label Learning. With the pseudo labels derived from Eq. 4, we apply the cross-entropy loss for model adaptation. Nevertheless, instead of assigning equal importance to all pseudo-labels, we introduce a Student's " + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": " distribution" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 94, + 535, + 167 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Details of class split. " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{V},\\bar{\\mathcal{V}}_s" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{V}}_t" + }, + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": " signifies the common class, source-private class, and target-private class, respectively." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 94, + 535, + 167 + ], + "lines": [ + { + "bbox": [ + 313, + 94, + 535, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 94, + 535, + 167 + ], + "type": "table", + "html": "
DatasetClass Split(γ/γs/γt)
OPDAOSDAPDA
Office-31 [43]10/10/1110/0/1110/21/0
Office-Home [49]10/5/5025/0/4025/40/0
VisDA [36]6/3/36/0/66/6/0
DomainNet [37]150/50/145--
", + "image_path": "a63a3c988cd5626ca6b60b3f9f8a9dfa676cb1af1dec014996b5469046b8633e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "text", + "content": "to characterize the certainty associated with each pseudolabel, taking into account the distance between " + }, + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{z}_{i,unk}^t\\| _2" + }, + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "inline_equation", + "content": "\\rho_{i,c}" + }, + { + "bbox": [ + 305, + 169, + 545, + 205 + ], + "type": "text", + "content": ". Specifically," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 333, + 210, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 210, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 333, + 210, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tau_ {i} ^ {t} \\propto 1 - \\left(1 + \\frac {\\left(\\rho_ {i , \\kappa} - \\| \\mathbf {z} _ {i , u n k} ^ {t} \\| _ {2}\\right) ^ {2}}{\\alpha}\\right) ^ {- \\frac {\\alpha + 1}{2}}, \\\\ \\mathcal {L} _ {c e} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\sum_ {c = 1} ^ {C} \\tilde {y} _ {i, c} ^ {t} \\log \\delta_ {c} \\left(f _ {\\theta} ^ {t} \\left(\\mathbf {x} _ {i} ^ {t}\\right)\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "00a8619fc11e1a3f21d05e4b64b43d8e003075ac1e69007be8cbbff8c9e9be5f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\tau_{i}^{t}" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " denotes the certainty, " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\kappa = \\mathrm{argmax}(\\epsilon_{i,c}),\\alpha = 1\\mathrm{e - }4" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " by default. " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\delta_c(f_\\theta^t (\\mathbf{x}_i^t))" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " -th soft-max probability for " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^t" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{i,c}^{t}" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " -th one-hot encoded pseudo-label for " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{y}}_i^t" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " . For those instances pseudo-labeled as private data, we refrain from introducing an additional " + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "inline_equation", + "content": "C + 1" + }, + { + "bbox": [ + 305, + 291, + 545, + 363 + ], + "type": "text", + "content": " -th class, instead employing a uniform distribution to represent them." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "content": "Feature Decomposition Regularization. To promote the discriminability of " + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 305, + 363, + 545, + 399 + ], + "type": "text", + "content": " for identifying private data, we impose a feature decomposition regularizer. Concretely," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 309, + 404, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 404, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 309, + 404, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p _ {i} = \\frac {\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right)}{\\exp \\left(\\left\\| \\mathbf {z} _ {i , u n k} ^ {t} \\right\\| _ {2}\\right) + \\exp \\left(\\left\\| \\mathbf {z} _ {i , k n w} ^ {t} \\right\\| _ {2}\\right)}, \\tag {6} \\\\ \\mathcal {L} _ {r e g} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\tau_ {i} ^ {t} \\cdot \\left(\\hat {y} _ {i} ^ {t} \\log \\left(p _ {i}\\right) + \\left(1 - \\hat {y} _ {i} ^ {t}\\right) \\log \\left(1 - p _ {i}\\right)\\right) \\\\ \\end{array}", + "image_path": "e661a421b6b02f370ab3fb30d79e134b2ebbc15c08db892fb527045b3f290b47.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 478, + 542, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 478, + 542, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 478, + 542, + 491 + ], + "type": "inline_equation", + "content": "\\hat{y}_i^t = 1" + }, + { + "bbox": [ + 305, + 478, + 542, + 491 + ], + "type": "text", + "content": " or 0 means pseudo-labeled as private/common data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": "Feature Consensus Regularization. Recent studies [41, 52, 53] in SFDA and SF-UniDA observe that incorporating consensus regularization with nearest neighbors in feature space could assist models in achieving more stable performance. Following them, we integrate the objective " + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 305, + 491, + 545, + 563 + ], + "type": "text", + "content": " into LEAD. More details are included in the supplementary." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 545, + 586 + ], + "type": "text", + "content": "Overall Optimization Objective. Taking into account the discussions above, the eventual loss function is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 368, + 592, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 592, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 368, + 592, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\cdot \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e g} + \\mathcal {L} _ {c o n} \\tag {7}", + "image_path": "8fc0512c5d2dbe1b9fa63ead0595e8eea42248d0b869a350abe09fd4def1fa8e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 615, + 482, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 482, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 482, + 627 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 615, + 482, + 627 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 306, + 615, + 482, + 627 + ], + "type": "text", + "content": " is a trade-off hyper-parameter." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 639, + 387, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 387, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 387, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 658, + 422, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 422, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 422, + 672 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Dataset. We empirically verify the effectiveness of our LEAD on four datasets, including Office-31 [43], OfficeHome [49], VisDA [36], and DomainNet [37]. For a" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23338" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 95, + 541, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 2. H-score (\\%) comparison in OPDA scenario on Office-Home. U denotes methods applicable for all potential label-shift scenarios (i.e., OPDA, OSDA, and PDA). SF represents source data-free. CF indicates model adaptation without time-consuming K-means clustering." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 95, + 541, + 236 + ], + "lines": [ + { + "bbox": [ + 51, + 95, + 541, + 236 + ], + "spans": [ + { + "bbox": [ + 51, + 95, + 541, + 236 + ], + "type": "table", + "html": "
MethodsUSFCFAr2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvg
CMU [15]✘✘✓56.056.959.267.064.367.854.751.166.468.257.969.761.6
DANCE [47]✓✘✓61.060.464.965.758.861.873.161.266.667.762.463.763.9
DCC [25]✓✘✘58.054.158.074.670.677.564.373.674.981.075.180.470.2
OVANet [44]✘✘✓62.875.678.670.768.875.071.358.680.576.164.178.971.8
GATE [7]✓✘✓63.875.981.474.072.179.874.770.382.779.171.581.775.6
UniOT [5]✘✘✓67.380.586.073.577.384.375.563.386.077.865.481.976.6
Source-only---47.371.681.951.557.269.456.040.376.661.444.273.560.9
SHOT-O [27]✘✓✓32.929.539.656.830.141.154.935.442.358.533.533.340.7
LEAD✓✓✓62.778.186.470.676.383.475.360.686.275.460.783.775.0
UMAD [28]✘✓✓61.176.382.770.767.775.764.455.776.373.260.477.270.1
+ LEAD✓✓✓65.282.687.478.076.587.579.865.088.479.463.083.578.0 (+7.9)
GLC [41]✓✓✘64.378.289.863.181.789.177.654.288.980.754.285.975.6
+ LEAD✓✓✘65.779.788.765.080.388.577.657.488.978.957.885.576.2 (+0.6)
", + "image_path": "9a09f18cf3a35f75cf607e1133d29483859eeb71f049092cdfbc96ffca3ac348.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 255, + 541, + 405 + ], + "blocks": [ + { + "bbox": [ + 56, + 240, + 536, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 240, + 536, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 536, + 251 + ], + "type": "text", + "content": "Table 3. H-score (\\%) comparison in OPDA scenario on Office-31, VisDA, and DomainNet. Some results are cited from UMAD [28]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 255, + 541, + 405 + ], + "lines": [ + { + "bbox": [ + 51, + 255, + 541, + 405 + ], + "spans": [ + { + "bbox": [ + 51, + 255, + 541, + 405 + ], + "type": "table", + "html": "
MethodsUSFCFOffice-31DomainNetVisDA
A2DA2WD2AD2WW2AW2DAvgP2RP2SR2PR2SS2PS2RAvgS2R
CMU [15]XX✓68.167.371.479.372.280.473.150.845.152.245.644.851.048.332.9
DANCE [47]✓X✓78.671.579.991.472.287.980.321.037.047.346.727.721.033.542.8
DCC [25]✓XX88.578.570.279.375.988.680.256.943.750.343.344.956.249.243.0
OVANet [44]XX✓85.879.480.195.484.094.386.556.047.151.744.947.457.250.753.1
GATE [7]✓X✓87.781.684.294.883.494.187.657.448.752.847.649.556.352.156.4
UniOT [5]XX✓87.088.588.498.887.696.691.159.351.847.848.346.858.352.057.3
Source-only---70.963.239.677.352.286.464.957.338.247.838.432.248.243.725.7
SHOT-O [27]X✓✓73.567.259.388.377.184.475.035.030.837.228.331.932.232.644.0
LEAD✓✓✓85.485.086.390.986.293.187.859.946.151.345.045.956.350.876.6
UMAD [28]X✓✓79.177.487.490.790.497.287.059.044.350.142.132.055.347.158.3
+ LEAD✓✓✓88.789.290.895.890.497.192.0 (+5.0)59.346.049.843.843.956.049.8 (+2.7)67.2 (+8.9)
GLC [41]✓✓X81.584.589.890.488.492.387.863.350.554.950.949.661.355.173.1
+ LEAD✓✓X83.086.887.692.087.193.288.3 (+0.5)63.550.755.151.649.961.355.4 (+0.3)76.8 (+3.7)
", + "image_path": "0dab7245e0c03f88b06bfead93890c266a666a8af263b4e87d10ab7b7fbfbfdf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 411, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 458 + ], + "type": "text", + "content": "fair comparison, we evaluate LEAD on partial-set domain adaptation (PDA), open-set domain adaptation (OSDA), and open-partial-set domain adaptation (OPDA). Details of classes split are summarized in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 459, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 578 + ], + "type": "text", + "content": "Evaluation. We utilize the same evaluation metric as previous works. In PDA scenarios, we report the classification accuracy for evaluation. In OSDA and OPDA scenarios, we report the H-score for performance comparison, which is a harmonic mean of accuracy over common data and accuracy over private data. It is worth noting that LEAD can be incorporated effortlessly into most existing SF-UniDA methods, for potential performance improvement. To demonstrate this merit, we integrate LEAD with two representative methods: UMAD [28] and GLC [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 578, + 288, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 710 + ], + "type": "text", + "content": "Implementation Details. All the experiments are conducted on a single RTX-3090 GPU with PyTorch-1.10. We adopt the same network architecture as existing baseline methods. During target model adaptation, we apply the SGD optimizer with momentum 0.9. The batch size is set to 64 for all experiments. The learning rate is set to 1e-3 for Office-31 and Office-Home, and 1e-4 for VisDA and DomainNet. As for " + }, + { + "bbox": [ + 46, + 578, + 288, + 710 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 578, + 288, + 710 + ], + "type": "text", + "content": ", it is set to 0.3 for Office-31, 1.0 for VisDA, and 2.0 for Office-Home and DomainNet. During inference, we adopt the same protocol like [41]. More details are presented in the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 411, + 429, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 429, + 423 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 429, + 423 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": "Results for OPDA Scenarios. To verify the effectiveness of LEAD, we first conduct experiments in the most challenging OPDA scenarios, where both source and target domains contain private data. Table 2 summarizes the results on Office-Home, and Table 3 details the results on Office-31, DomainNet, and VisDA. These results demonstrate that LEAD achieves performance surpassing or comparable to existing methods based on K-means clustering for segregating common and private data. In comparison with methods without clustering, LEAD achieves significantly superior performance. For instance, on Office-Home, DomainNet, and VisDA datasets, LEAD outperforms UMAD by " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "18.3\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": ", respectively. We attribute this to our instance-level adaptive decision strategy, which effectively circumvents the limitations imposed by a fixed threshold. While GLC achieves commendable performance on its own, the integration with LEAD further amplifies its effectiveness. For example, LEAD improves GLC in the H-score on VisDA from " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "73.1\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "76.8\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": ". The integration with LEAD proves even more significant against UMAD, with the H-score on Office-Home and VisDA boosting from " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "70.1\\% /58.3\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "inline_equation", + "content": "78.0\\% /67.2\\%" + }, + { + "bbox": [ + 304, + 426, + 545, + 689 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Results for OSDA Scenarios. We then conduct experiments in OSDA scenarios, where only the target domain" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23339" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 84, + 542, + 244 + ], + "blocks": [ + { + "bbox": [ + 127, + 71, + 464, + 81 + ], + "lines": [ + { + "bbox": [ + 127, + 71, + 464, + 81 + ], + "spans": [ + { + "bbox": [ + 127, + 71, + 464, + 81 + ], + "type": "text", + "content": "Table 4. H-score (%) comparison in OSDA scenario on Office-Home, Office-31, and VisDA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 84, + 542, + 244 + ], + "lines": [ + { + "bbox": [ + 51, + 84, + 542, + 244 + ], + "spans": [ + { + "bbox": [ + 51, + 84, + 542, + 244 + ], + "type": "table", + "html": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X Y55.057.059.059.358.260.659.251.361.261.953.555.357.652.655.776.575.965.864.765.254.2
DANCE [47]✓ X Y6.59.09.920.410.49.228.412.812.614.27.913.212.984.978.879.178.868.378.879.867.5
DCC [25]✓ X X56.167.566.749.666.564.055.853.070.561.657.271.961.758.354.867.289.485.380.972.759.6
OVANet [44]X X Y58.666.369.962.065.268.659.853.469.368.759.666.764.090.588.386.798.288.398.491.766.1
GATE [7]✓ X Y63.870.575.866.467.971.767.361.576.070.461.875.169.088.486.584.295.086.196.789.570.8
Source-only- - -46.163.372.942.854.058.747.836.166.260.845.368.255.278.272.144.282.252.188.869.629.1
SHOT-O [27]X Y Y37.741.848.456.439.840.960.041.549.761.841.443.646.980.271.664.393.164.091.877.528.1
LEAD✓ Y Y60.770.876.561.068.670.865.559.874.264.857.775.867.284.985.190.294.890.396.590.374.2
UMAD [28]X Y Y59.271.876.663.569.071.962.554.672.866.557.970.766.488.584.486.895.088.295.989.866.8
+ LEAD✓ Y Y62.673.677.361.366.270.763.354.675.166.763.575.667.690.489.690.395.490.797.392.370.2
+2.5
", + "image_path": "47886e4aedf11d69aff94161c0fff93be7c4ce632a071a6da120faf703753b01.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 261, + 542, + 419 + ], + "blocks": [ + { + "bbox": [ + 128, + 247, + 463, + 258 + ], + "lines": [ + { + "bbox": [ + 128, + 247, + 463, + 258 + ], + "spans": [ + { + "bbox": [ + 128, + 247, + 463, + 258 + ], + "type": "text", + "content": "Table 5. Accuracy comparison (%) in PDA scenario on Office-Home, Office-31, and VisDA." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 261, + 542, + 419 + ], + "lines": [ + { + "bbox": [ + 51, + 261, + 542, + 419 + ], + "spans": [ + { + "bbox": [ + 51, + 261, + 542, + 419 + ], + "type": "table", + "html": "
MethodsU SF CFOffice-HomeOffice-31VisDA
Ar2ClAr2PrAr2ReCl2ArCl2PrCl2RePr2ArPr2ClPr2ReRe2ArRe2ClRe2PrAvgA2DA2WD2AD2WW2AW2DAvgS2R
CMU [15]X X X50.974.278.462.264.172.563.547.978.372.454.778.966.584.184.269.297.266.898.883.465.5
DANCE [47]✓ X X53.673.284.970.867.382.670.050.984.877.055.981.871.177.171.283.794.692.696.886.073.7
DCC [25]✓ X X54.247.557.583.871.686.263.765.075.285.578.282.670.987.381.395.4100.095.5100.093.372.4
OVANet [44]X X X34.154.672.142.447.355.938.226.261.756.735.868.949.569.461.761.490.266.498.774.634.3
GATE [7]✓ X X55.875.985.373.670.283.072.159.584.779.663.983.874.089.586.293.5100.094.498.693.775.6
Source-only- - -45.969.281.155.761.264.860.741.175.870.549.978.462.990.479.379.395.984.398.187.842.8
SHOT-P [27]X X Y64.785.190.175.173.984.276.464.190.380.763.385.577.889.884.492.296.692.299.492.474.2
LEAD✓ Y Y58.283.187.070.575.483.373.750.483.778.358.783.273.889.893.995.698.696.099.495.575.3
UMAD [28]X Y Y51.266.579.263.162.968.263.356.475.974.555.978.366.385.485.183.597.686.299.489.568.5
+ LEAD✓ Y Y61.178.287.570.578.581.874.160.483.378.863.282.175.089.290.890.698.093.0100.093.678.1
+8.7+4.1+9.6
GLC [41]✓ Y X55.979.087.572.571.882.774.941.782.477.360.484.372.589.889.892.896.696.199.494.176.2
+ LEAD✓ Y X56.681.288.172.572.283.275.343.683.478.259.784.273.289.893.295.696.996.099.495.175.5
+0.7+1.0-0.7
", + "image_path": "9799ab231a02b84e086c5cdd46936f9c56a9f8bfdc9ebbfabed18fcc72a0edbf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": "involves private data. Table 4 lists the results, demonstrating the superiority of LEAD against baselines. Specifically, LEAD obtains the H-score of " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "90.3\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "74.2\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": " on Office-Home, Office-31, and VisDA datasets, respectively. Similar to the observations in OPDA scenarios, LEAD also exhibits performance improvement to existing methods. LEAD improves the H-score of UMAD by " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 46, + 426, + 287, + 544 + ], + "type": "text", + "content": " across the Office-Home, Office-31, and VisDA datasets, respectively. The results again verify the merit of LEAD when incorporated with existing methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Results for PDA Scenarios. Finally, we verify the effectiveness of LEAD in PDA scenarios, where the target domain's label sets are subsets of those in the source domain. Results presented in Table 5 show that LEAD achieves superior performance even compared to methods tailored for PDA. In particular, LEAD attains the overall accuracy of " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "73.8\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "95.5\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "75.3\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " on Office-Home, Office-31, and VisDA datasets, individually. LEAD contributes the accuracy increase of " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " to UMAD on these datasets. Note that, there is a slight degradation " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "(-0.7\\%)" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " for GLC w/ LEAD. It may arise from gradient conflict, causing the optimization process to inadvertently compromise one objective in favor of improving the overall objective. We will pursue better integration strategies in future work." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 437, + 541, + 505 + ], + "blocks": [ + { + "bbox": [ + 324, + 425, + 526, + 434 + ], + "lines": [ + { + "bbox": [ + 324, + 425, + 526, + 434 + ], + "spans": [ + { + "bbox": [ + 324, + 425, + 526, + 434 + ], + "type": "text", + "content": "Table 6. Ablation studies of LEAD in OPDA scenarios." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 437, + 541, + 505 + ], + "lines": [ + { + "bbox": [ + 309, + 437, + 541, + 505 + ], + "spans": [ + { + "bbox": [ + 309, + 437, + 541, + 505 + ], + "type": "table", + "html": "
LceLregLconOffice-31Office-HomeVisDA
---64.960.925.7
✓--82.073.661.5
✓✓-84.574.567.3
✓-✓87.274.575.7
✓✓✓87.875.076.6
", + "image_path": "c90a4677647c45b282f23ca537a6582cb00335312a1c7e6fb53dc819b5cbce01.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 510, + 435, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 435, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 435, + 521 + ], + "type": "text", + "content": "4.3. Experimental Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 582 + ], + "type": "text", + "content": "Ablation Study. To assess the contribution of different components within LEAD, we carry out extensive ablation studies in OPDA scenarios. The results, detailed in Table 6, confirm that each component is not only effective but also complementary to the others." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Efficiency Comparison. As previously stated, LEAD employs the distance to target prototypes and source anchors for adaptive pseudo-labeling. This design, in contrast to clustering-based methods, notably reduces the need for computational resources. Table 7 juxtaposes the runtime of GLC [41] and LEAD in OPDA scenarios. It is clear that LEAD requires minimal resources for deriving decision boundaries, whereas GLC's resource demands escalate dramatically with data scale. In particular, on DomainNet, with about 0.6 million images across 345 categories, GLC averages 897.65 seconds, while LEAD takes only 0.29 sec" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23340" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 67, + 185, + 168 + ], + "blocks": [ + { + "bbox": [ + 72, + 67, + 185, + 168 + ], + "lines": [ + { + "bbox": [ + 72, + 67, + 185, + 168 + ], + "spans": [ + { + "bbox": [ + 72, + 67, + 185, + 168 + ], + "type": "image", + "image_path": "272dc3cb2d8c736b11cf92868aefcd5a1556bce19c89b588cbf4274e63087ebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 170, + 162, + 177 + ], + "lines": [ + { + "bbox": [ + 112, + 170, + 162, + 177 + ], + "spans": [ + { + "bbox": [ + 112, + 170, + 162, + 177 + ], + "type": "text", + "content": "(a) Sensitivity to " + }, + { + "bbox": [ + 112, + 170, + 162, + 177 + ], + "type": "inline_equation", + "content": "\\lambda" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 68, + 298, + 169 + ], + "blocks": [ + { + "bbox": [ + 187, + 68, + 298, + 169 + ], + "lines": [ + { + "bbox": [ + 187, + 68, + 298, + 169 + ], + "spans": [ + { + "bbox": [ + 187, + 68, + 298, + 169 + ], + "type": "image", + "image_path": "039a3a74b32629a546ea44434469cf1ba504d2e9fc48a5d3968bb893703994ba.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 170, + 280, + 177 + ], + "lines": [ + { + "bbox": [ + 215, + 170, + 280, + 177 + ], + "spans": [ + { + "bbox": [ + 215, + 170, + 280, + 177 + ], + "type": "text", + "content": "(b) H-score convergence" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 68, + 407, + 169 + ], + "blocks": [ + { + "bbox": [ + 299, + 68, + 407, + 169 + ], + "lines": [ + { + "bbox": [ + 299, + 68, + 407, + 169 + ], + "spans": [ + { + "bbox": [ + 299, + 68, + 407, + 169 + ], + "type": "image", + "image_path": "5adf2f4b6957ee926f84826bf54253e771ca8d412b05a8741ebed93b861bb315.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 170, + 389, + 176 + ], + "lines": [ + { + "bbox": [ + 329, + 170, + 389, + 176 + ], + "spans": [ + { + "bbox": [ + 329, + 170, + 389, + 176 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 329, + 170, + 389, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{Ar}\\rightarrow \\mathrm{Pr}" + }, + { + "bbox": [ + 329, + 170, + 389, + 176 + ], + "type": "text", + "content": " in OPDA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 408, + 68, + 518, + 169 + ], + "blocks": [ + { + "bbox": [ + 408, + 68, + 518, + 169 + ], + "lines": [ + { + "bbox": [ + 408, + 68, + 518, + 169 + ], + "spans": [ + { + "bbox": [ + 408, + 68, + 518, + 169 + ], + "type": "image", + "image_path": "a3f8ae72befbceebaf4610de0e38520614dac47bcf9abf0b8500ba7a595ff590.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 170, + 500, + 176 + ], + "lines": [ + { + "bbox": [ + 440, + 170, + 500, + 176 + ], + "spans": [ + { + "bbox": [ + 440, + 170, + 500, + 176 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 440, + 170, + 500, + 176 + ], + "type": "inline_equation", + "content": "\\mathrm{Re}\\rightarrow \\mathrm{Cl}" + }, + { + "bbox": [ + 440, + 170, + 500, + 176 + ], + "type": "text", + "content": " in OPDA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 50, + 214, + 282, + 251 + ], + "blocks": [ + { + "bbox": [ + 49, + 201, + 285, + 213 + ], + "lines": [ + { + "bbox": [ + 49, + 201, + 285, + 213 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 285, + 213 + ], + "type": "text", + "content": "Table 7. Runtime (s) to derive pseudo-label decision boundaries." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 214, + 282, + 251 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 282, + 251 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 282, + 251 + ], + "type": "table", + "html": "
Office-31Office-HomeVisDADomainNet
GLC [41]0.190.510.46897.65
LEAD0.020.020.110.29
", + "image_path": "5ebce39355cfec23b6551945fdd5ff2ce164625d4882035dd7c0bc7fc97ab209.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 253, + 286, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 288 + ], + "type": "text", + "content": "onds on the same platform. Moreover, LEAD also shows substantial resource savings even on normal-scale datasets, e.g., on VisDA with over " + }, + { + "bbox": [ + 46, + 253, + 286, + 288 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 253, + 286, + 288 + ], + "type": "text", + "content": " reduction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "content": "Parameter Sensitivity. We investigate the parameter sensitivity of " + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "content": " on Office-31 in OPDA scenarios, where the range of " + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "content": " is set to [0.1, 0.2, 0.3, 0.4, 0.5]. Figure 3 (a) illustrates that the results are stable around the selected parameter " + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "content": ". Through oracle validation, we may find the optimal parameter configuration for this setting is " + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\lambda = 0.1" + }, + { + "bbox": [ + 46, + 289, + 287, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 362, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 398 + ], + "type": "text", + "content": "Training Stability. Figure 3 (b) visualizes the training curves of our LEAD on VisDA in OPDA scenarios. Generally, the training procedure of LEAD is stable and effective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 483 + ], + "type": "text", + "content": "Varying Unknown Classes. With the rise in unknown private classes, accurately distinguishing common from private data becomes more challenging. To examine the robustness of our LEAD, we compared it with GLC and UMAD in OPDA scenarios using the Office-Home dataset. The results in Figure 3 (c-d) show that LEAD achieves stable results and consistently boosts existing methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "content": "Effect of Instance-level Decision Boundary. To evaluate the contribution of our instance-level pseudo-labeling strategy, we conduct a comparison against the intuitive design, which uses the average of " + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mu_{\\text{com}}" + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mu_{\\text{pri}}" + }, + { + "bbox": [ + 46, + 485, + 287, + 604 + ], + "type": "text", + "content": " as the threshold. As depicted in Figure 4 (a), instance-level decision boundary strategy significantly outperforms the vanilla global decision design (75.0% vs 63.1% on Office-Home in OPDA scenarios). We attribute this to that vanilla design discounts the inconsistency of covariate shifts across categories, and also the variability among samples from the same category." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Exploration with Entropy as the Indicator. Contrasting with existing threshold-based methods that utilize Entropy as the indicator for differentiating common and private data, we propose decomposing the feature and using the projection " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " on the source-unknown space as an alternative indicator. To empirically substantiate the superiority, we remold our framework to incorporate Entropy as the indicator. Figure 4 (b) depicts the results. An observation is that while employing entropy as the indicator within our framework is" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 318, + 194, + 425, + 281 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 545, + 200 + ], + "type": "text", + "content": "Figure 3. Robustness analysis. (a) shows the sensitivity to " + }, + { + "bbox": [ + 46, + 178, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 178, + 545, + 200 + ], + "type": "text", + "content": " on Office-31. (b) presents the H-score curves on VisDA. (c-d) present robustness analysis when varying the unknown private categories." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 194, + 425, + 281 + ], + "lines": [ + { + "bbox": [ + 318, + 194, + 425, + 281 + ], + "spans": [ + { + "bbox": [ + 318, + 194, + 425, + 281 + ], + "type": "image", + "image_path": "3f2ae93e6a3123e54df700f932bcee05fe553c6c6eeed04979dfea9a7f441902.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 282, + 427, + 289 + ], + "lines": [ + { + "bbox": [ + 325, + 282, + 427, + 289 + ], + "spans": [ + { + "bbox": [ + 325, + 282, + 427, + 289 + ], + "type": "text", + "content": "(a) Effect of Instance-level Decision Boundary" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 291, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 334 + ], + "type": "text", + "content": "Figure 4. Methodology analysis. (a) compares the effectiveness of our instance-level decision strategy against the vanilla global decision strategy. (b) examines the efficacy of using Entropy as the indicator for private data." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 425, + 194, + 529, + 281 + ], + "blocks": [ + { + "bbox": [ + 425, + 194, + 529, + 281 + ], + "lines": [ + { + "bbox": [ + 425, + 194, + 529, + 281 + ], + "spans": [ + { + "bbox": [ + 425, + 194, + 529, + 281 + ], + "type": "image", + "image_path": "862ba535973c6ca976f54376d3e3fe936843dd25a2033bfb58ea613f8a1c298d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 282, + 530, + 289 + ], + "lines": [ + { + "bbox": [ + 432, + 282, + 530, + 289 + ], + "spans": [ + { + "bbox": [ + 432, + 282, + 530, + 289 + ], + "type": "text", + "content": "(b) Exploration with Entropy as the Indicator" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 335, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 384 + ], + "type": "text", + "content": "feasible, its performance is inferior compared to " + }, + { + "bbox": [ + 304, + 335, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{i,unk}^{t}" + }, + { + "bbox": [ + 304, + 335, + 545, + 384 + ], + "type": "text", + "content": ". We speculate that this inferiority might result from entropy's limited capacity to distinguish between common samples with high uncertainty and private unknown samples." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 415, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 605 + ], + "type": "text", + "content": "In this paper, we delve into source-free universal domain adaptation (SF-UniDA). Different from existing methods that leverage hand-crafted thresholding strategies or feature-clustering algorithms, we propose a novel idea of LEArning Decomposition (LEAD) tailored for SF-UniDA. Technically, LEAD first decomposes features into source-known and -unknown two parts, then establishes instance-level decision boundaries for pseudo-labeling. It offers an elegant view to identify target-private unknown data without tedious tuning a threshold or relying on unstable clustering. Extensive experiments across various scenarios have verified the effectiveness of our LEAD. Besides, it is appealing in view that LEAD can be seamlessly incorporated into existing methods to further boost performance. Remarkably, LEAD improves UMAC by " + }, + { + "bbox": [ + 304, + 415, + 545, + 605 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 304, + 415, + 545, + 605 + ], + "type": "text", + "content": " overall H-score in the OPDA scenario on Office-Home dataset." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment: This work is supported by the National Natural Science Foundation of China (No. 62372329), in part by the National Key Research and Development Program of China (No. 2021YFB2501104), in part by Shanghai Rising Star Program (No.21QC1400900), in part by Tongji-Qomolo Autonomous Driving Commercial Vehicle Joint Lab Project, and in part by Xiaomi Young Talents Program, and in part by the International Exchange Program for Graduate Students, Tongji University (No.2023020053)." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23341" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 711 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Richard Bellman. Dynamic programming. Science, 153 (3731):34-37, 1966. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 169 + ], + "type": "text", + "content": "[3] Zhangjie Cao, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Partial transfer learning with selective adversarial networks. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 286, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 204 + ], + "type": "text", + "content": "[4] Zhangjie Cao, Kaichao You, Mingsheng Long, Jianmin Wang, and Qiang Yang. Learning to transfer examples for partial domain adaptation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 206, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 286, + 237 + ], + "type": "text", + "content": "[5] Wanxing Chang, Ye Shi, Hoang Tuan, and Jingya Wang. Unified optimal transport framework for universal domain adaptation. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 286, + 272 + ], + "type": "text", + "content": "[6] Prithvijit Chattopadhyay, Yogesh Balaji, and Judy Hoffman. Learning to balance specificity and invariance for in and out of domain generalization. In ECCV. Springer, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 274, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 286, + 317 + ], + "type": "text", + "content": "[7] Liang Chen, Yihang Lou, Jianzhong He, Tao Bai, and Minghua Deng. Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 319, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 286, + 351 + ], + "type": "text", + "content": "[8] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In CVPR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 354, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 354, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 354, + 286, + 396 + ], + "type": "text", + "content": "[9] Yi-Hsin Chen, Wei-Yu Chen, Yu-Ting Chen, Bo-Cheng Tsai, Yu-Chiang Frank Wang, and Min Sun. No more discrimination: Cross city adaptation of road scene segmenters. In ICCV, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 399, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 452 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "type": "text", + "content": "[11] Nicolas Courty, Rémi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for domain adaptation. IEEE TPAMI, 39(9):1853-1865, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 489, + 286, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 286, + 543 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 587 + ], + "type": "text", + "content": "[13] Qi Dou, Cheng Ouyang, Cheng Chen, Hao Chen, and Pheng-Ann Heng. Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss. In IJCAI, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 286, + 621 + ], + "type": "text", + "content": "[14] Zeyu Feng, Chang Xu, and Dacheng Tao. Open-set hypothesis transfer with semantic consistency. IEEE TIP, 30:6473-6484, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "text", + "content": "[15] Bo Fu, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Learning to detect open classes for universal domain adaptation. In ECCV. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 286, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 711 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 711 + ], + "type": "text", + "content": "[16] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. JMLR, pages 2096-2030, 2016. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[17] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 129, + 545, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 159 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 204 + ], + "type": "text", + "content": "[19] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 248 + ], + "type": "text", + "content": "[20] Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML. PMLR, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 292 + ], + "type": "text", + "content": "[21] Han-Kai Hsu, Chun-Han Yao, Yi-Hsuan Tsai, Wei-Chih Hung, Hung-Yu Tseng, Maneesh Singh, and Ming-Hsuan Yang. Progressive domain adaptation for object detection. In WACV, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 369 + ], + "type": "text", + "content": "[23] Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels. IEEE TPAMI, pages 766-785, 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 371, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 402 + ], + "type": "text", + "content": "[24] Jogendra Nath Kundu, Naveen Venkat, R Venkatesh Babu, et al. Universal source-free domain adaptation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 404, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 436 + ], + "type": "text", + "content": "[25] Guangrui Li, Guoliang Kang, Yi Zhu, Yunchao Wei, and Yi Yang. Domain consensus clustering for universal domain adaptation. In CVPR, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 438, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 479 + ], + "type": "text", + "content": "[26] Shuang Li, Shiji Song, Gao Huang, Zhengming Ding, and Cheng Wu. Domain invariant and class discriminative feature learning for visual domain adaptation. IEEE TIP, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 482, + 545, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 523 + ], + "type": "text", + "content": "[27] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML. PMLR, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 558 + ], + "type": "text", + "content": "[28] Jian Liang, Dapeng Hu, Jiashi Feng, and Ran He. Umad: Universal model adaptation under domain and category shift. arXiv preprint arXiv:2112.08553, 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 590 + ], + "type": "text", + "content": "[29] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In ICLR, 2023. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[30] Dongnan Liu, Chaoyi Zhang, Yang Song, Heng Huang, Chenyu Wang, Michael Barnett, and Weidong Cai. Decompose to adapt: Cross-domain object detection via feature disentanglement. IEEE TMM, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 668 + ], + "type": "text", + "content": "[31] Xinyu Liu and Yixuan Yuan. A source-free domain adaptive polyp detection framework with style diversification flow. IEEE TMI, pages 1897-1908, 2022. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[32] Zhengfa Liu, Guang Chen, Zhijun Li, Yu Kang, Sanqing Qu, and Changjun Jiang. Psdc: A prototype-based shared-dummy classifier model for open-set domain adaptation. IEEE Transactions on Cybernetics, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23342" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[33] Zhengfa Liu, Guang Chen, Zhijun Li, Sanqing Qu, Alois Knoll, and Changjun Jiang. D2ifln: Disentangled domain-invariant feature learning networks for domain generalization. IEEE Transactions on Cognitive and Developmental Systems, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 160 + ], + "type": "text", + "content": "[34] Mingsheng Long, Jianmin Wang, Guiguang Ding, Jiaguang Sun, and Philip S Yu. Transfer feature learning with joint distribution adaptation. In ICCV, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 287, + 183 + ], + "type": "text", + "content": "[35] Pau Panareda Busto and Juergen Gall. Open set domain adaptation. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 226 + ], + "type": "text", + "content": "[36] Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, and Kate Saenko. Visda: The visual domain adaptation challenge. arXiv preprint arXiv:1710.06924, 2017.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 259 + ], + "type": "text", + "content": "[37] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In ICCV, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 293 + ], + "type": "text", + "content": "[38] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML. PMLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 337 + ], + "type": "text", + "content": "[39] Sanqing Qu, Guang Chen, Jing Zhang, Zhijun Li, Wei He, and Dacheng Tao. Bmd: A general class-balanced multicentric dynamic prototype strategy for source-free domain adaptation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[40] Sanqing Qu, Yingwei Pan, Guang Chen, Ting Yao, Changjun Jiang, and Tao Mei. Modality-agnostic debiasing for single domain generalization. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[41] Sanqing Qu, Tianpei Zou, Florian Röhrbein, Cewu Lu, Guang Chen, Dacheng Tao, and Changjun Jiang. Upcycling models under domain and category shift. In CVPR, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[43] Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category models to new domains. In ECCV, 2010. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 481, + 287, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 503 + ], + "type": "text", + "content": "[44] Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In ICCV, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 536 + ], + "type": "text", + "content": "[45] Kuniaki Saito, Kohei Watanabe, Yoshitaka Ushiku, and Tatsuya Harada. Maximum classifier discrepancy for unsupervised domain adaptation. In CVPR, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 568 + ], + "type": "text", + "content": "[46] Kuniaki Saito, Shohei Yamamoto, Yoshitaka Ushiku, and Tatsuya Harada. Open set domain adaptation by backpropagation. In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[47] Kuniaki Saito, Donghyun Kim, Stan Sclaroff, and Kate Saenko. Universal domain adaptation through self supervision. In NeurIPS, pages 16282-16292, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[48] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In WACV, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 669 + ], + "type": "text", + "content": "[49] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In CVPR, 2017. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[50] Paul Voigt and Axel Von dem Bussche. The eu general data protection regulation (gdpr). A Practical Guide, 1st Ed., Cham: Springer International Publishing, 10(3152676):10-5555, 2017. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 217 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[51] Haohan Wang, Xindi Wu, Zeyi Huang, and Eric P Xing. High-frequency component helps explain the generalization of convolutional neural networks. In CVPR, 2020. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[52] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 183 + ], + "type": "text", + "content": "[53] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "text", + "content": "[54] Kaichao You, Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Universal domain adaptation. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23343" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_content_list.json b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..472d9f646d2f6b7ed42e3848cd6a5038e5a222a1 --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_content_list.json @@ -0,0 +1,1572 @@ +[ + { + "type": "text", + "text": "LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry", + "text_level": 1, + "bbox": [ + 112, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weirong Chen $^{1,2*}$ Le Chen $^{3}$ Rui Wang $^{4}$ Marc Pollefeys $^{4}$", + "bbox": [ + 235, + 179, + 730, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ TU Munich $^{2}$ Munich Center for Machine Learning $^{3}$ MPI for Intelligent Systems $^{4}$ Microsoft", + "bbox": [ + 107, + 200, + 861, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 258, + 313, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual odometry estimates the motion of a moving camera based on visual input. Existing methods, mostly focusing on two-view point tracking, often ignore the rich temporal context in the image sequence, thereby overlooking the global motion patterns and providing no assessment of the full trajectory reliability. These shortcomings hinder performance in scenarios with occlusion, dynamic objects, and low-texture areas. To address these challenges, we present the Long-term Effective Any Point Tracking (LEAP) module. LEAP innovatively combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Moreover, LEAP's temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty. Based on these traits, we develop LEAP-VO, a robust visual odometry system adept at handling occlusions and dynamic scenes. Our mindful integration showcases a novel practice by employing long-term point tracking as the front-end. Extensive experiments demonstrate that the proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks.", + "bbox": [ + 73, + 291, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 638, + 209, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual odometry (VO) calculates the camera's position and movement from an image sequence. It is widely used in various fields including robotics [46], mixed reality [23], and autonomous driving [15]. The performance of VO highly depends on the accuracy of its point tracking front-end, which recovers relative camera motion between consecutive frames using projective geometry. It further requires the point correspondences to be static, as dynamic points can lead to inaccurate motion estimation and pose recovery.", + "bbox": [ + 75, + 665, + 468, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advances in feature tracking and optical flow estimation, most existing VO methods mainly focus on the two-view case, i.e., computing feature correspondences between a pair of images. However, these methods", + "bbox": [ + 75, + 801, + 468, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a76ff2249955c9c8077c455b55618895587c7dd0639e49aad6be7800470b4de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 261, + 691, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/070ad67c83554f8e25cf7cd79e091da3bce520463dff2693730f48a768a657d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 261, + 880, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/faed9191be70cfa06e13b268052bd065dc87863b05880998ec81b4c6d869c1b0.jpg", + "image_caption": [ + "Visual Odometry", + "Figure 1. Comparison between two-view approach and multi-view approach for visual odometry. (a) In the two-view approach, correspondences are derived for every image pair and concatenated. Managing occlusion becomes challenging without the temporal context. (b) In the multi-view approach, long-term point trajectories can be obtained all at once, enabling the detection of occlusion and tracking points under partial occlusions." + ], + "image_footnote": [], + "bbox": [ + 638, + 433, + 856, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are limited as they rely solely on pairwise relative motion, thus neglecting the valuable temporal information in image sequences. Consequently, they often struggle with capturing dynamic motion patterns in tracks. Another challenge arises from the need to handle occlusions. In two-view scenarios, occlusion is not explicitly modeled, leading to misidentification of occluded points as incorrect matches. To address these challenges effectively, we propose to introduce a long-term point tracker, which can recover the trajectories of specific points across the image sequence given a sparse set of these query points. Its capability of leveraging temporal context and reliably tracking queries over multiple frames even under occlusion, makes it highly suitable for VO applications (refer to Figure 1).", + "bbox": [ + 496, + 566, + 892, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Persistent Independent Particles (PIPs) [16] proposes a novel pipeline that brings the idea of iterative refinement into the long-term point tracking task, with the local correlation evidence to maintain feasible computation. While it is proficient in leveraging the temporal context present in the image sequence, it ignores the inter-track information in the image sequence and processes each query point independently. Consequently, PIPs struggles to cap", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done at Microsoft Mixed Reality & AI Lab Zurich for a Master's degree.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19844", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture the global motion patterns from the image sequence. Additionally, it lacks a reliability metric for its trajectory estimates, hindering its application in quality-sensitive tasks like video segmentation and visual odometry. To mitigate the aforementioned issues, we propose a pipeline that explores dynamic track estimation and temporal probabilistic formulation for long-term point tracking.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present Long-term Effective Point Tracking (LEAP), aiming to estimate the reliability of predicted correspondences and track moving object trajectories in dynamic scenes. To achieve this, we harness the global motion patterns—often overlooked in existing methods—through an anchor-based dynamic track estimation module, leveraging visual, temporal, and inter-track information to differentiate between the static and dynamic elements. We also incorporate a temporal probabilistic approach into our LEAP pipeline to handle uncertainties and refine the distribution of point correspondences iteratively, showcasing superior performance in complex real-world situations. More importantly, we pioneer the usage of continuous motion estimation from long-term point tracking to construct a robust visual odometry system. Our LEAP-VO distinguishes itself through its thoughtful integration of visual features, track distribution, and global motion patterns, significantly enhancing both the performance and robustness of visual odometry in dynamic environments.", + "bbox": [ + 75, + 198, + 470, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 494, + 218, + 510 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-term Point Tracking. In the realm of long-term point tracking or Tracking Any Point (TAP), various methods have been introduced recently. PIPs [16] reframes pixel tracking as a long-term motion estimation with local information and improves tracking accuracy by associating each pixel with a trajectory that spans multiple image frames. The subsequent works enhance performance by integrating spatial context features [2], tracking multiple queries concurrently [18], and extending temporal fields and templates [49]. TAP-Net [9] adopts a different method by locating candidate correspondences through a global cost volume, improved by TAPIR's trajectory refinement [10] and BootsTAP's semi-supervised training [11]. Other approaches explore self-supervision targets to learn visual correspondence via powerful feature representation [5, 17, 36]. Recently, OmniMotion [37] provides a pure test-time optimization method that infers the point trajectory and occlusion directly from videos with optical flow predictions. However, its slow inference speed makes it unsuitable for real-time applications such as visual odometry.", + "bbox": [ + 75, + 520, + 468, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular Visual Odometry. Visual Odometry (VO) can be categorized into indirect (feature-based) and direct methods. Indirect methods, such as MonoSLAM [8] and ORB-SLAM [25], mostly rely on feature extraction and", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "matching to estimate the camera pose. Direct methods, conversely, skip feature extraction and matching and utilize intensity to optimize camera pose and 3D scene point positions based on photometric errors [12, 13, 38], showing better performance on low-texture regions where keypoint tracking can easily fail. However, they are more vulnerable to illumination changes and issues with rolling shutter cameras. Advancements in deep learning have spawned end-to-end monocular VO methods in supervised [3, 28, 30, 33, 39, 40, 43, 44] and unsupervised [21, 27, 45, 51] settings. DeepVO [39] utilizes Recurrent Neural Networks to model sequential data, while SfMLearner [51] develops an unsupervised framework to learn the depth and camera motion from unlabeled monocular videos. TartanVO [40] proposes a VO model that generalizes to multiple datasets and real-world scenarios. DROID-SLAM [32] follows learning-based optimization from RAFT [31] and proposes an end-to-end system, integrating flow, confidence, and geometric optimization via iterative GRU updates. The followed-up work DPVO [33] further cuts computational complexity by replacing dense feature tracking with sparse patch tracking.", + "bbox": [ + 496, + 90, + 890, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Track Estimation. In addition to enhancing accuracy, numerous studies have focused on the detection of dynamic objects within videos. Methods based on segmentation, such as DynaSLAM [1] and RCVD [20], employ semantic cues from pre-trained models to exclude dynamic regions. However, these segmentation-based methods are constrained to predefined object categories and struggle to capture real-world motion (for example, differentiating between a stationary and a moving car). Alternatively, DyanVO [28] and ParticleSfM [48] adopt a trajectory-based motion detection approach. This method addresses dynamic scenes using estimated point track information, thereby generating reliable camera trajectories in complex motion scenarios. Nonetheless, they depend on dense optical flow estimation to capture global motion patterns. Furthermore, dynamic tracks can be detected using additional depth information for a 3D consistency check [7, 47], or by optimization based on photo consistency [35, 41, 42].", + "bbox": [ + 496, + 425, + 890, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 709, + 591, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to tackle the aforementioned challenges in VO systems, particularly in handling dynamic environments, temporal occlusions, and low-texture regions. To this end, we introduce our LEAP module, which incorporates dynamic track estimation and temporal probabilistic formulation. Furthermore, we illustrate how this module serves as the cornerstone for constructing our advanced VO system.", + "bbox": [ + 496, + 733, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminary: Tracking Any Point", + "text_level": 1, + "bbox": [ + 500, + 847, + 792, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tracking arbitrary pixels in an image sequence is traditionally akin to estimating optical flow, where each pixel", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19845", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is mapped to a corresponding location in a subsequent frame. However, two-view approaches often overlook the rich temporal context within the image sequence, leading to suboptimal outcomes. Recent advancements address this by revisiting the particle video concept and improving long-term point tracking through learning-based techniques [10, 16, 18]. These TAP methods can track points over multiple frames and detect occlusions, significantly boosting tracking robustness under challenging scenarios.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Formulation.** Given a sequence of $S$ consecutive RGB images $\\mathbf{I} = [\\mathbf{I}_1, \\dots, \\mathbf{I}_S]$ , $\\mathbf{I}_s \\in \\mathbb{R}^{3 \\times H \\times W}$ as input, the goal of TAP is to track a set of query points across these frames. For a specific query point with the 2D pixel coordinate $\\mathbf{x}_q \\in \\mathbb{R}^2$ in frame $s_q$ , TAP predicts its trajectory across all $S$ images as $\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_S]$ , $\\mathbf{x}_s \\in \\mathbb{R}^2$ , given $\\mathbf{x}_{s_q} = \\mathbf{x}_q$ . Additionally, TAP estimates the visibility of each point throughout the sequence as $\\mathbf{V} = [v_1, \\dots, v_S]$ , where $v_s \\in \\{0, 1\\}$ is a binary label of point visibility indicating whether a point in frame $s$ is visible or occluded. The TAP formulation for a single query can be described as", + "bbox": [ + 75, + 242, + 472, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {X}, \\mathbf {V}\\right) = \\operatorname {T A P} \\left(\\mathbf {I}, \\mathbf {x} _ {q}, s _ {q}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 419, + 468, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PIPs [16], a notable learning-based TAP method, offers a promising solution featuring local cost volumes and iterative refinement. Initially, each image $\\mathbf{I}_s\\in \\mathbf{I}$ is processed through a CNN feature extractor $\\mathcal{F}$ to extract the feature map $\\mathbf{Y}_{\\mathbf{s}} = \\mathcal{F}(\\mathbf{I}_{s})$ . Point features $\\mathbf{f}_q$ are computed by bilinear sampling at the query positions $\\mathbf{x_s}$ on $\\mathbf{Y}_s$ . These point features from each image are then concatenated to form a point feature tensor $\\mathbf{F} = [\\mathbf{f}_1,\\dots ,\\mathbf{f}_S]$ . PIPs approaches the TAP problem by iteratively updating the state variables $(\\mathbf{X},\\mathbf{F})$ , with the initial states obtained by duplicating the query positions and features. During the $k$ -th iterative refinement, PIPs computes a multi-scale local cost volume $\\mathbf{C}[\\mathbf{X}^k ]$ centered around the current estimated point position $\\mathbf{X}^k$ . Utilizing this local evidence, PIPs predicts updates for the state variables via", + "bbox": [ + 75, + 446, + 472, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\left(\\Delta \\mathbf {X}, \\Delta \\mathbf {F}\\right) = \\operatorname {R e f i n e r} \\left(\\mathbf {F} ^ {k}, \\operatorname {p o s} \\left(\\mathbf {X} ^ {k} - \\mathbf {x} _ {\\mathbf {q}}\\right), \\mathbf {C} ^ {k} \\left[ \\mathbf {X} ^ {k} \\right]\\right), \\\\ \\mathbf {X} ^ {k + 1} \\leftarrow \\mathbf {X} ^ {k} + \\Delta \\mathbf {X}, \\quad \\mathbf {F} ^ {k + 1} \\leftarrow \\mathbf {F} ^ {k} + \\Delta \\mathbf {F}, \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 681, + 468, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{pos}(\\cdot)$ refers to the positional embedding. The final point feature $\\mathbf{F}^K$ is used to predict point visibility with a simple linear projection layer $\\mathcal{G}_v$ , yielding $\\mathbf{V} = \\mathcal{G}_v(\\mathbf{F}^K)$ . Recent advances extend PIPs by replacing the MLP-based refiner with an attention-based refiner and adding spatial feature aggregation to track multiple queries together [18].", + "bbox": [ + 75, + 731, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Anchor-based Dynamic Track Estimation", + "text_level": 1, + "bbox": [ + 76, + 832, + 431, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The established TAP model enables us to extract long-term point trajectories and thereby recover camera motion from an image sequence. However, dynamic environments", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9a4b8f0c4199ea8b8569721db074c6591697e11ec43d5bbfbe71c1e30542119.jpg", + "image_caption": [ + "Figure 2. Distributed image gradient-based sampling with $k = 8$ , $N_{a} = 64$ . After computing the image gradient and pooling, we split the gradient map into $8 \\times 8$ grids and select the point with the maximum gradient in each grid." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 887, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pose extra challenges for VO systems that rely on static 3D point correspondences linked to unchanging scene locations. Therefore, integrating dynamic track estimation into the VO system is crucial for real-world applications.", + "bbox": [ + 496, + 244, + 890, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an innovative method for dynamic track estimation that leverages visual, temporal, and inter-track information. While common techniques like monocular segmentation [20] excel at identifying moving objects using visual cues but can erroneously label static objects, like parked cars, as dynamic. Trajectory-based methods [48], utilizing optical flows from two or multiple frames, can differentiate static and moving objects apart but tend to ignore visual information and can be inefficient due to dense trajectories computation. Merging these two distinct methods is nontrivial; however, the formulation of the TAP pipeline, enhanced with inter-track attention, offers a viable solution.", + "bbox": [ + 496, + 306, + 892, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method smartly integrates visual appearance, long-term motion, and inter-track cues, combining the strengths of segmentation and trajectory-based techniques. Utilizing TAP, we extract point features $\\mathbf{F}$ and long-term trajectories $\\mathbf{X}$ from the video input. These features capture visual information and the iterative refinement module exchanges information across tracks over time for accurate labeling. However, such a method depends on the query distribution and quantities. With only one query during inference time, it cannot utilize inter-track information. Likewise, if the queries focus on a single object, the system may overlook the motions of other objects due to similar motion patterns.", + "bbox": [ + 496, + 489, + 893, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anchor-based Inter-track Attention. To better capture global motion patterns, we introduce additional anchors alongside the given queries, denoted as $\\mathcal{A}$ , $|\\mathcal{A}| = N_{a}$ . Ideally, the anchor set should exhibit two characteristics: (1) the anchors should be easy to track, and (2) they should be well-distributed to encapsulate global motion patterns. Given the image $\\mathbf{I}_{s_q}$ from which the queries $\\mathbf{x}_q$ are extracted (we maintain the same notations for both single and multiple queries for simplicity), we first employ the Sobel Kernel to derive the image gradient $(\\mathbf{G}_x,\\mathbf{G}_y)$ for image $\\mathbf{I}_{s_q}$ . We then apply an average pooling layer on gradient maps for smoothing and downscaling feature maps. Subsequently, we divide the gradient map into $k\\times k$ sub-regions and select top $\\frac{N_a}{k^2}$ pixels with the highest image gradient magnitude in", + "bbox": [ + 496, + 689, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19846", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "each local grid. The process is demonstrated in Figure 2.", + "bbox": [ + 76, + 90, + 450, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To predict the dynamic label of a trajectory, we track queries with additional anchors and concatenate the original query features $\\mathbf{X}$ and query trajectories $\\mathbf{F}$ with anchor features $\\mathbf{X}_A$ and anchor trajectories $\\mathbf{F}_A$ . We apply a shallow MLP layer $\\mathcal{G}_d$ with average pooling over the temporal dimension. The dynamic track label $\\mathbf{m}_d$ is estimated as", + "bbox": [ + 75, + 106, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {m} _ {d} = \\operatorname {a v g p o o l} \\left(\\mathcal {G} _ {d} \\left(\\left[ \\mathbf {X} ^ {K}; \\mathbf {X} _ {A} ^ {K} \\right], \\left[ \\mathbf {F} ^ {K}; \\mathbf {F} _ {A} ^ {K} \\right]\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 205, + 468, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since newly added anchors often lack ground truth information, we employ a semi-supervised training scheme that tracks queries and anchors together to leverage inter-track attention but only calculates losses based on query predictions. This approach prevents the network from developing bias towards specific query sampling methods while learning dynamic track labels effectively.", + "bbox": [ + 75, + 233, + 468, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Probability Modeling", + "text_level": 1, + "bbox": [ + 76, + 348, + 356, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervised learning applied to large-scale datasets has yielded notable advancements in learning-based point tracking methods. Nonetheless, trajectory estimation errors remain a concern, especially under challenging conditions like occlusions, motion blur, or textureless regions. Recognizing the varying reliability of trajectories, integrating an uncertainty estimation mechanism into the correspondence estimation process is beneficial [6]. This is particularly crucial for quality-sensitive tasks such as visual odometry.", + "bbox": [ + 75, + 371, + 468, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To evaluate the reliability of point correspondences, we incorporate a probabilistic formulation into the TAP pipeline. Given an image sequence $\\mathbf{I}$ and a single query point $\\mathbf{x}_q$ , our objective is to compute the conditional probability density of point trajectory as $p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q)$ . Recognizing the strong correlation between points within the same track, treating each point in the track as independent becomes suboptimal. Therefore, we propose to model the distribution of point trajectories by applying two multivariate distributions, respectively for the 2D coordinates. Let $\\mathbf{X} = [\\mathbf{a};\\mathbf{b}]$ , where $\\mathbf{a}\\in \\mathbb{R}^S$ and $\\mathbf{b}\\in \\mathbb{R}^{S}$ represent the $\\mathbf{X}$ and $\\mathbf{Y}$ coordinate of all points in $\\mathbf{X}$ , respectively. Assuming independence between these two coordinates, the joint probability distribution can be expressed as $p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q) = p(\\mathbf{a}|\\mathbf{I},\\mathbf{x}_q)\\cdot p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)$ . We adopt the multivariate Cauchy distribution, which is associated with its heavy tails distribution and is more stable for optimization. The probability density function (PDF) of a single coordinate is given by", + "bbox": [ + 75, + 507, + 470, + 780 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\mathbf {a} | \\mathbf {I}, \\mathbf {x} _ {q}) = \\frac {\\Gamma \\left(\\frac {1 + S}{2}\\right)}{\\Gamma \\left(\\frac {1}{2}\\right) \\pi^ {\\frac {S}{2}} \\left| \\boldsymbol {\\Sigma} _ {a} \\right| ^ {\\frac {1}{2}} \\left[ 1 + \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {a} ^ {- 1} \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) \\right] ^ {\\frac {1 + S}{2}}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 787, + 472, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a similar expression is applicable for $p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)$ . $\\Gamma$ is the Gamma function. The parameters $(\\boldsymbol{\\mu}_a,\\boldsymbol{\\Sigma}_a,\\boldsymbol{\\mu}_b,\\boldsymbol{\\Sigma}_b)$ represent the location and scale matrices for the respective coordinates. During inference, the uncertainty associated", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a70127acf14c96950f4d02f1fdf61d4ad2cb1f4fc795d5e9f3b2983140ece6d1.jpg", + "image_caption": [ + "Figure 3. LEAP Front-end. Once image feature maps are obtained, selected anchors aid in tracking. The queries and anchors are processed by a refiner to iteratively update states. The model outputs trajectory distribution, visibility, and dynamic track label." + ], + "image_footnote": [], + "bbox": [ + 509, + 93, + 890, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with each point is quantified by the sum of diagonal scale estimates at position $(s,s)$ as $\\phi (\\mathbf{x}_s) = \\pmb {\\Sigma}_a[s,s] + \\pmb {\\Sigma}_b[s,s]$", + "bbox": [ + 498, + 306, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kernel-based Estimation. Rather than just predicting the positions of points within a trajectory, our model focuses on recovering the distribution of point trajectories by estimating its parameters. The location parameter predictions can be associated with the previous point trajectory as $[\\pmb{\\mu}_a; \\pmb{\\mu}_b] = \\mathbf{X}$ . However, directly deriving a symmetric and positive definite scale matrix from model outputs is challenging. To address this, we employ a kernel-based approach with the linear kernel $K(\\mathbf{x}, \\mathbf{y}) = \\mathbf{x}^T \\mathbf{y}$ . During each iteration, two scale matrices $(\\pmb{\\Sigma}_a, \\pmb{\\Sigma}_b)$ are constructed by first applying two linear projection layers to the point features, represented as $\\mathbf{F}_a^k = \\mathcal{G}_a(\\mathbf{F}^k)$ and $\\mathbf{F}_b^k = \\mathcal{G}_b(\\mathbf{F}^k)$ . We then compute the scale matrices as $\\pmb{\\Sigma}_a = K(\\mathbf{F}_a, \\mathbf{F}_a) + \\sigma \\mathbf{I}$ , and $\\pmb{\\Sigma}_b = K(\\mathbf{F}_b, \\mathbf{F}_b) + \\sigma \\mathbf{I}$ , where $\\sigma$ is a small positive value and $\\mathbf{I}$ is the identity matrix. The model parameters are refined through Maximum Likelihood Estimation (MLE) using a negative log-likelihood (NLL) loss function.", + "bbox": [ + 496, + 352, + 892, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss Functions. We supervise the point trajectory mainly with the NLL loss, which is based on the predicted distribution parameters and the ground-truth point trajectory. The main point trajectory loss is given as", + "bbox": [ + 496, + 625, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {k} ^ {K} \\gamma^ {K - k} \\mathcal {L} _ {N L L} \\left(\\mathbf {X} ^ {k}, \\mathbf {X} ^ {*}, \\boldsymbol {\\Sigma} _ {a} ^ {k}, \\boldsymbol {\\Sigma} _ {b} ^ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 694, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ is the number of iteration and $\\gamma = 0.8$ .", + "bbox": [ + 500, + 744, + 818, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the visibility and dynamic track label supervision, we use the cross entropy loss with the estimates $\\mathbf{V},\\mathbf{m}_d$ and ground truth $\\mathbf{V}^{*},\\mathbf{m}_{d}^{*}$ as", + "bbox": [ + 496, + 758, + 890, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {v i s}} = \\left(1 - \\mathbf {V} ^ {*}\\right) \\log \\left(1 - \\mathbf {V}\\right) + \\mathbf {V} ^ {*} \\log \\mathbf {V}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 814, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d y n} = \\left(1 - \\mathbf {m} _ {d} ^ {*}\\right) \\log \\left(1 - \\mathbf {m} _ {d}\\right) + \\mathbf {m} _ {d} ^ {*} \\log \\mathbf {m} _ {d}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 834, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss is a weighted sum of three losses:", + "bbox": [ + 500, + 859, + 815, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = w _ {1} \\mathcal {L} _ {\\text {m a i n}} + w _ {2} \\mathcal {L} _ {\\text {v i s}} + w _ {3} \\mathcal {L} _ {\\text {d y n}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 885, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19847", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/005c323b67834729f12b46342f226d06e34bd05848cbaea2bb4db7b3f676de27.jpg", + "image_caption": [ + "Figure 4. LEAP-VO pipeline. Given a new image $\\mathbf{I}_t$ received at time $t$ , the keypoint extractor extracts new keypoints from $\\mathbf{I}_t$ . Then, all the keypoints from the latest $S_{LP}$ -frame $\\mathbf{I}_{t - S_{LP} + 1:t}$ are tracked across all other frames within the current LEAP window, followed by a track filtering step to remove outliers. Finally, the local BA module is used on the current BA window to update the camera poses and 3D positions of the extracted keypoints. The colored arrows denote the moving direction of each module when a new image is received." + ], + "image_footnote": [], + "bbox": [ + 119, + 85, + 504, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bffe5d79cb06a7a1728eda2276e66086b413c42983dcc93b2df1db81da7e0b08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 98, + 857, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. LEAP-VO", + "text_level": 1, + "bbox": [ + 76, + 361, + 194, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Leveraging the anchor-based dynamic track estimations and temporal probabilistic formulation, we develop our Long-term Effective Point Tracking (LEAP) pipeline, boosting the accuracy of point tracking and simultaneously capturing global motion patterns and uncertainties of the tracks (refer to Figure 3). In this section, we demonstrate how we can build visual odometry for dynamic scenes based on the long-term tracking capabilities of LEAP. Our LEAP-VO pipeline is illustrated in Figure 4.", + "bbox": [ + 75, + 385, + 472, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Keypoint Extraction and Tracking. In our VO system, the LEAP model serves as the front-end. When a new image frame $\\mathbf{I}_t$ is captured at time $t$ , we sample $N$ new keypoints from this image to initiate point tracking. We consistently employ the distributed image gradient-based sampling method (refer to Section 3.2) for keypoint selection due to its favorable properties in feature distinctiveness and distribution. This approach also obviates the need for additional anchors during tracking. The LEAP model then tracks these keypoints across the last $S_{LP}$ frames, performing bidirectional tracking (both forward and backward in time). Notably, $S_{LP}$ can exceed the model window $S$ , thus enabling more expansive tracking capabilities through sequential prediction [16, 18].", + "bbox": [ + 75, + 537, + 468, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Track Filtering. Theoretically, precise camera pose estimation between two images can be attained with just a few accurate correspondences under certain conditions. However, the quality of correspondences obtained from feature matching or optical flow can vary. Traditional VO pipelines often use RANSAC [14] to filter out incorrect matches, but their efficiency decreases with more keypoints and views. Our approach, in contrast, bypasses the sampling-based method and uses trajectory quality assessments from LEAP", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for more efficient and informative track filtering.", + "bbox": [ + 498, + 362, + 821, + 377 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) Visibility and Dynamic Track Label: The LEAP model has the capability to estimate not only the 2D point positions within a trajectory but also the visibility of each point and the dynamic track label of the entire trajectory. Thresholds $\\gamma_v$ for visibility and $\\gamma_d$ for dynamic track labels are set to ensure that only visible and static points are utilized in the bundle adjustment process.", + "(ii) Track Uncertainty: In Section 3.3, we introduce probabilistic modeling to LEAP, allowing us to gauge confidence via distribution measurements. We select only high-confidence points, which correspond to low uncertainty measurements. Denoting our model's uncertainty estimation as $\\Phi$ , we retain points that exhibit high confidence using the criterion $\\Phi \\leq Q(\\gamma_u)$ , where $\\gamma_u$ is the uncertainty quantile and $Q:[0,1] \\to \\mathbb{R}$ is the quantile function.", + "(iii) Track Length: After filtering, we evaluate and remove tracks with insufficient observations, as bundle adjustment is more reliable with increased observations and wider baselines. Tracks with fewer than $\\gamma_{\\text{track}}$ valid points are excluded from optimization cost computation due to their potential unreliability." + ], + "bbox": [ + 496, + 378, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sliding-window Optimization. With the point validation mask established, we proceed to define the optimization cost function for our sliding window bundle adjustment. The window size for local bundle adjustment, denoted as $S_{BA}$ , may differ from the point tracking window size $S_{LP}$ . Our geometric bundle adjustment aims to optimize camera poses $\\mathbf{T}$ and 3D scene point positions $\\mathbf{Q}$ by aligning the induced point trajectory from the projective relationship, with the estimated point trajectory from LEAP. We parameterize a 3D scene point $\\mathbf{Q}_i$ by its 2D keypoint location $\\mathbf{x}_i$ in image $\\mathbf{I}_i$ and depth $d_i$ . Let $\\mathrm{LEAP}_{i\\rightarrow j}$ denote the mapping of tracking keypoint from $\\mathbf{I}_i$ to $\\mathbf{I}_j$ using LEAP model. The", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19848", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "reprojection cost is formulated as:", + "bbox": [ + 76, + 90, + 307, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S _ {B A}} \\sum_ {n} w _ {i \\rightarrow j, n} \\| \\mathcal {P} (\\mathbf {T} _ {i}, \\mathbf {T} _ {j}, \\mathbf {K}, d _ {i, n}) - \\\\ \\operatorname {L E A P} _ {i \\rightarrow j} (\\mathbf {x} _ {i, n}) \\| _ {\\rho}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 118, + 468, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_{\\rho}$ is the distance metric, and the weight $w_{i\\rightarrow j,n}$ is derived from the track filtering results. We use the GaussNewton method [34] to optimize Eq. (9) for $K_{BA}$ iterations. At each iteration, we compute the camera poses update $\\Delta \\xi^{(k)}\\in \\mathfrak{se}(3)$ (lie-algebra corresponding to $\\mathbf{T}$ ), and the depth update $\\Delta \\mathbf{D}^{(k)}$ for each point. The optimization can be solved efficiently with the Schur decomposition. We use the robust Huber loss function for the distance metric.", + "bbox": [ + 75, + 186, + 470, + 308 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 324, + 200, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 349, + 274, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replica. Replica [29] provides synthetic indoor environments for simulating the image-capturing process with a moving camera. We use the camera trajectories provided by [50] for the localization task, resulting in 16 camera trajectories derived from 8 static scenes. Each sequence comprises 900 frames with rendered RGB images. We use \"Sequence 1\" for the visual odometry evaluation.", + "bbox": [ + 75, + 373, + 470, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MPI Sintel. The MPI Sintel [4] dataset contains dynamic image sequences from open-source 3D animated short films. These sequences consist of large and complex object motion, along with other imaging effects such as motion blur and defocus blur. Following prior works [20, 48], we assess dynamic VO performance on the MPI Sintel dataset, with each sequence containing 20 to 50 image frames.", + "bbox": [ + 75, + 489, + 468, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TartanAir Shibuya. The TartanAir Shibuya dataset [26] features dynamic scenes from two scenarios: \"Standing Humans\" and \"Road Crossing\". The majority of the humans remain stationary in \"Standing Humans\" sequences, while \"Road Crossing\" sequences provide a more challenging setting where multiple humans are moving in different directions. Each sequence in the dataset includes 100 frames with more than 30 tracked moving humans.", + "bbox": [ + 75, + 607, + 468, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. For visual odometry, we compare the Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot) following [33, 47, 48]. ATE measures how much the estimated trajectory deviated from the ground truth trajectory, which is calculated as the root mean square error (RMSE) overall all pose transformations after scaling and alignment. \"RPE trans\" computes translation errors made over a certain distance (meter), and \"RPE rot\" computes the estimated rotation over a certain distance (degree). Both \"RPE trans\" and \"RPE rot\" are calculated on all poses and then averaged.", + "bbox": [ + 75, + 737, + 470, + 905 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9faf726c5e5fb61c91080f585bbf38c00eb5c872769d353a44f2b26fadf5eaf6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReplica
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]0.0860.0300.650
DynaSLAM [1]0.0390.0170.366
DROID-SLAM [32]0.2670.0362.631
TartanVO [40]0.4060.0362.063
DytanVO [28]0.2890.0352.146
DPVO [33]0.2570.0362.635
LEAP-VO (Ours)0.2040.0301.992
", + "bbox": [ + 504, + 88, + 888, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0740f1e91760ce2e792e2709686305b58406ec0e90f256e01d9deb71f739d8d8.jpg", + "table_caption": [ + "Table 1. Camera tracking results on Replica [29]. The statistics demonstrate that our method achieves better results than other competitive baselines." + ], + "table_footnote": [], + "table_body": "
MethodMPI Sintel
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]XXX
DynaSLAM [1]XXX
DROID-SLAM [32]0.1750.0841.912
TartanVO [40]0.2380.0931.305
DytanVO [28]0.1310.0971.538
DPVO [33]0.0760.0781.722
LEAP-VO (Ours)0.0370.0551.263
", + "bbox": [ + 503, + 277, + 888, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Camera tracking results on MPI Sintel [4]. The statistics show that our method outperforms other competitive baselines. 'X' denotes the tracking failure case.", + "bbox": [ + 498, + 417, + 890, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 467, + 714, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ CoTracker [18], a recent PIPs variant, as our base TAP model. We train our model on the TAP-Vid-Kubric [9] training set, following [18]. We load the pretrained CoTracker weights for the image feature extractor and train LEAP for 100,000 steps using 4 NVIDIA A100 GPUs in parallel. During training, we use $K = 4$ steps for iterative refinements, $N = 256$ for the number of queries, $S = 8$ for the model window, and $S_{LP} = 12$ for the tracking window. For inter-track anchor-based attention, the number of anchors is 64. The loss weights are set to $w_{1} = 1.0, w_{2} = 0.5, w_{3} = 0.5$ . In LEAP-VO, we perform 4 steps for each bundle adjustment with the Huber Loss for distance metric. The bundle adjustment window size $S_{BA}$ is set to 15. Track filtering parameters are set at $\\gamma_v = 0.9, \\gamma_d = 0.9, \\gamma_u = 0.8, \\gamma_{track} = 3$ . Please refer to our supplementary materials for more details.", + "bbox": [ + 498, + 489, + 892, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Visual Odometry Results", + "text_level": 1, + "bbox": [ + 500, + 741, + 728, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our methods against the state-of-the-art SLAM and VO approaches, including ORB-SLAM2 [24], DynaSLAM [1], DROID-SLAM [32], TartanVO [40], DytanVO [28], and DPVO [33]. Evaluations are conducted on the static Replica dataset [29] as well as on the dynamic datasets MPI-Sintel [4] and TartanAir-Shibuya [26].", + "bbox": [ + 498, + 763, + 890, + 856 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replica. We compare our method with other baselines on the Replica dataset, which contains eight static scenes", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19849", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88ddaacfec9a855b6b763fd39bd052dacd1e76b1bf07559f5abfc40ca0b7d100.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStandingHumanRoadCrossing (Easy)RoadCrossing (Hard)Average
01020304050607
ORB-SLAM2 [24] w/ mask0.07880.00600.06570.01960.01481.09840.84760.3044
DynaSLAM [1]X0.88360.39070.41960.49250.64460.6539(0.5808)
AirDOS [26] w/ mask0.06060.01930.09510.03310.02060.22300.56250.1449
DROID-SLAM [32]0.00510.00730.01030.01200.22780.02530.57880.1238
TartanVO [40]0.06000.16050.27620.18140.21740.32280.50090.2456
DytanVO [28]0.03270.10170.06080.05160.07550.03650.06600.0607
DPVO [39]0.05390.16030.17880.28580.11960.11190.14980.1514
LEAP-VO (Ours)0.00810.01850.03990.03170.01130.06890.02460.0290
", + "bbox": [ + 125, + 88, + 844, + 236 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. ATE (m) results on TartanAir-Shibuya Sequences [26]. ( $\\cdot$ ) denotes averaging on valid scenes only. We can observe that our method shows better results for ATE on most sequences, outperforming all SLAM and VO baselines on average ATE.", + "bbox": [ + 75, + 238, + 893, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c9be0f3fcfc59cefc63ad77edac4adc0e039eefe5b343e437c31e099d237ab9.jpg", + "image_caption": [ + "(a) RoadCrossing 04" + ], + "image_footnote": [], + "bbox": [ + 84, + 273, + 279, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5a4197b9fd43580001e00f745e1a2cbf7cc380b0001405e7443351579a9253f.jpg", + "image_caption": [ + "(b) RoadCrossing 05" + ], + "image_footnote": [], + "bbox": [ + 287, + 272, + 480, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/422bd73a1d36da5ecdab37f323089a41085702b01500552c15b23bbc9e085202.jpg", + "image_caption": [ + "(c) RoadCrossing 06", + "Figure 5. Qualitative results of camera trajectory estimation on TartanAir-Shibuya [26]. The visualizations show that our method provides more robust and accurate camera pose trajectories, especially in hard cases (RoadCrossing 06 and RoadCrossing 07)." + ], + "image_footnote": [], + "bbox": [ + 491, + 272, + 684, + 421 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/500f27a167516833358616cab081a003fdad56f68b6e256912b551c566d37440.jpg", + "image_caption": [ + "(d) RoadCrossing 07" + ], + "image_footnote": [], + "bbox": [ + 692, + 273, + 885, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "without any moving objects. Challenges arise in two test sequences that contain prolonged occlusions, leading to the potential camera tracking loss. As illustrated in Table 1, our approach outperforms other VO methods on all three metrics, showing effectiveness under static environments and the capability to handle occlusions. Further insights are available in Section 4.5. For classical methods, ORBSLAM2 [24] and DynaSLAM [1], we observe tracking loss and resuming via re-localization, which reduces long-term drift and shows high accuracy for camera pose tracking.", + "bbox": [ + 75, + 477, + 472, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MPI-Sintel. We also evaluate our VO camera trajectory accuracy on 14 dynamic scenes from the MPI Sintel dataset, following the setting of [48]. As shown in Table 2, our method substantially surpasses other VO baselines in dynamic scenes, thanks to our temporal uncertainty estimation and dynamic track detection. Traditional feature-based methods like ORB-SLAM2 [24] and DynaSLAM [1] sometimes fail to track camera movements in highly dynamic sequences, primarily due to outliers from dynamic object correspondences and segmentation-based only dynamic track estimation. In contrast, our approach, which considers visual, temporal, and inter-track information, demonstrates notable improvements in dynamic scene tracking accuracy.", + "bbox": [ + 75, + 635, + 468, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TartanAir-Shibuya. Following DytanVO [28], we further evaluate the VO performance using the dynamic TartanAir-Shibuya dataset. We assess the Average Trajectory Error (ATE) across seven diverse scenes, ranging from simple to", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "challenging. As depicted in Table 3, our method consistently outperforms all other SLAM and VO baselines in the majority of these scenes. Moreover, trajectory visualizations in Figure 5 also highlight the robust camera tracking capabilities of our method in dynamic environments.", + "bbox": [ + 496, + 477, + 892, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Dynamic Track Estimation Results", + "text_level": 1, + "bbox": [ + 498, + 559, + 803, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our LEAP formulation, we have introduced the anchor-based dynamic track estimation module to distinguish between static and dynamic point trajectories, which provides valuable information for handling dynamic scenes in visual odometry. Our method leverages both point feature information and inter-track relations to predict the dynamic track label for each trajectory. To demonstrate the performance of the proposed dynamic track estimation module, we take the LEAP model and evaluate it on novel image sequences from diverse datasets. For a comprehensive demonstration, we uniformly sample $16 \\times 16$ queries among the images and track them for $T = 8$ frames. The results, shown in Figure 6, illustrate that LEAP can effectively handle various scenarios, including single-object motion, complex and rapid motion, and multi-object motion with occlusion.", + "bbox": [ + 496, + 582, + 892, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 816, + 653, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Track Filtering. To highlight the efficacy of our track filtering module in LEAP-VO, we conduct an ablation study on different masks used in track filtering. We study three types of filtering criteria: visibility, dynamic", + "bbox": [ + 496, + 839, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19850", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/781af32895b8c1318275094eab868761f2d0179d29830e8800be855692ef374b.jpg", + "image_caption": [ + "Figure 6. Visualization of dynamic track estimation on DAVIS [19], MPI-Sintel [4], and TartanAir-Shibuya [26]. Odd columns: all point trajectories. Even columns: estimated dynamic point trajectories." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 462, + 319 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/849ddeb126a9119bf85089fee5033115a331c32107fbd180edaeec2c0c8ce3b9.jpg", + "image_caption": [ + "Figure 7. Visualization of point-wise uncertainty measurements. Keypoints with the lowest $20\\%$ (left) and highest $20\\%$ (right) uncertainty are shown in green and yellow, respectively." + ], + "image_footnote": [], + "bbox": [ + 81, + 383, + 465, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "track, and confidence filtering. Table 4 shows the impact of applying various filtering types on the dynamic MPI Sintel benchmark. Without dynamic filtering or confidence filtering, the ATE increases by $46\\%$ and $130\\%$ , respectively. Combining all three types of filtering methods altogether, we achieve the best results in all three metrics.", + "bbox": [ + 75, + 580, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1ee9533292ee09daae5704bd49dc27defff0357868632af005d5f5d707109e22.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTrack FilteringMPI Sintel
Vis.Dyn.Conf.ATERPE transRPE rot
LEAP-VO✓--0.1180.0781.340
✓✓-0.0850.0621.281
✓-✓0.0540.0581.274
✓✓✓0.0370.0551.263
", + "bbox": [ + 84, + 678, + 460, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Uncertainty Estimation. As illustrated in Figure 7, we demonstrate the LEAP's capability to provide meaningful per-point uncertainty. It models the uncertainty as distribution parameters for both the entire track and individual points. Two key observations can be found from Figure 7. Firstly, our model reliably identifies high-uncertainty", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "points, typically located in low-texture areas or in regions with repetitive patterns. Secondly, although not explicitly supervised, the model tends to assign higher uncertainty to dynamic objects, which are inherently more challenging to track. These findings highlight the efficacy of LEAP's probabilistic formulation in handling uncertainty.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Keypoint Extraction. Table 5 showcases the effectiveness of various keypoint extraction strategies for VO performance. Despite its distinctive features, SIFT extraction provides inferior performance due to its unpredictable keypoint distribution. In contrast, the random sampling tends to generate well-separated keypoints, leading to a performance similar to SIFT's. Our distributed image gradient-based sampling shows the best performance, underscoring the significance of feature distribution and distinctiveness.", + "bbox": [ + 496, + 191, + 890, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/97c60765d761296eda93eb56bac7175dba27f1dda620d9805ab0652b917566aa.jpg", + "table_caption": [ + "Table 4. Effect of the proposed track filtering method on MPI-Sintel [4]. Tracking using all criteria yields the best results." + ], + "table_footnote": [], + "table_body": "
Policy:UniformRandomSIFT [22]Ours
ATE (m):0.15110.03660.03550.0290
", + "bbox": [ + 547, + 327, + 844, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Baseline comparison. We evaluate the pretrained CoTracker [18] for the end VO performance. Table 6 shows that LEAP front-end outperforms CoTracker on all datasets, demonstrating the effectiveness of the proposed dynamic track estimation and temporal probability modeling.", + "bbox": [ + 496, + 414, + 890, + 491 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8efc96d69cf79bff7d4790d40d3929500fccf9ebdc1a343f9fdf266338968788.jpg", + "table_caption": [ + "Table 5. Comparison of keypoint extraction policies ( $N_{a} = 64$ ) on TartanAir-Shibuya [26]. Our distributed image-gradient sampling method achieves better results compared to other policies." + ], + "table_footnote": [], + "table_body": "
Front-endReplicaMPI SintelShibuya
ATERPE transRPE rotATERPE transRPE rotATE
CoTracker0.3010.0312.0430.1260.0711.3660.160
LEAP0.2040.0301.9920.0370.0551.2630.029
", + "bbox": [ + 506, + 496, + 880, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Baseline comparison with CoTracker front-end [18]. Our LEAP front-end achieves better results on all datasets.", + "bbox": [ + 500, + 566, + 890, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Long-term Effective Any Point Tracking (LEAP), designed to tackle the limitations of previous point-tracking methods. By leveraging visual, inter-track, and temporal cues within an anchor-based dynamic track estimation module, LEAP captures global motion patterns and effectively discriminates between static and dynamic elements. Moreover, our approach incorporates a temporal probabilistic formulation into the point-tracking pipeline, emulating recursive Bayesian filtering through a learning-based refinement module. This enables our model to accurately assess the reliability of point trajectory estimates. Using LEAP as the front-end, we then develop the LEAP-VO system, featuring novel integration of visual information, global motion patterns, and track distribution for handling dynamic scenes. Additionally, our takes for LEAP-VO can be integrated with other TAP-based front-ends, offering the potential to markedly improve camera-tracking accuracy and robustness.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Berta Bescos, José M Fácil, Javier Civera, and José Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, 3(4):4076-4083, 2018. 2, 6, 7", + "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 2023. 2", + "[3] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In CVPR, pages 2560–2568, 2018. 2", + "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625. Springer-Verlag, 2012. 6, 8", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 2", + "[6] Weirong Chen, Suryansh Kumar, and Fisher Yu. Uncertainty-driven dense two-view structure from motion. IEEE Robotics and Automation Letters, 8(3):1763-1770, 2023. 4", + "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In ICCV, pages 7063-7072, 2019. 2", + "[8] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE TPAMI, 29(6):1052-1067, 2007. 2", + "[9] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 35:13610-13626, 2022. 2, 6", + "[10] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. ICCV, 2023. 2, 3", + "[11] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 2", + "[12] Jakob Engel, Thomas Schöps, and Daniel Cremers. Ldslam: Large-scale direct monocular slam. In ECCV, pages 834-849. Springer, 2014. 2", + "[13] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE TPAMI, 40(3):611-625, 2017. 2", + "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5", + "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 1" + ], + "bbox": [ + 76, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, pages 59-75. Springer, 2022. 1, 2, 3, 5", + "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 33: 19545-19560, 2020. 2", + "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2, 3, 5, 6, 8", + "[19] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2018. 8", + "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2, 3, 6", + "[21] Shunkai Li, Xin Wang, Yingdian Cao, Fei Xue, Zike Yan, and Hongbin Zha. Self-supervised deep visual odometry with online adaptation. In CVPR, pages 6339-6348, 2020. 2", + "[22] David G Lowe. Object recognition from local scale-invariant features. In ICCV, pages 1150-1157. IEEE, 1999. 8", + "[23] Annette Mossel and Manuel Kroeter. Streaming and exploration of dynamically changing dense 3d reconstructions in immersive virtual reality. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR-Adjunct), pages 43–48. IEEE, 2016. 1", + "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: an open-source SLAM system for monocular, stereo and RGB-D cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 6, 7", + "[25] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 2", + "[26] Yuheng Qiu, Chen Wang, Wenshan Wang, Mina Henein, and Sebastian Scherer. Airdos: Dynamic slam benefits from articulated objects. In ICRA, pages 8047-8053. IEEE, 2022. 6, 7, 8", + "[27] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In CVPR, pages 12240-12249, 2019. 2", + "[28] Shihao Shen, Yilin Cai, Wenshan Wang, and Sebastian Scherer. Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environments. In ICRA, pages 4048-4055. IEEE, 2023. 2, 6, 7", + "[29] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6", + "[30] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In ICLR, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19852", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 2", + "[32] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34: 16558-16569, 2021. 2, 6, 7", + "[33] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 6", + "[34] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In International workshop on vision algorithms, pages 298-372. Springer, 1999. 6", + "[35] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. Neuraliff: Segmenting 3d objects that move in egocentric videos. In 3DV, pages 910-919. IEEE, 2021. 2", + "[36] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2", + "[37] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 2", + "[38] Rui Wang, Martin Schworer, and Daniel Cremers. Stereo dso: Large-scale direct sparse visual odometry with stereo cameras. In ICCV, pages 3903-3911, 2017. 2", + "[39] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In ICRA, pages 2043-2050. IEEE, 2017. 2, 7", + "[40] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In CoRL, pages 1761-1772. PMLR, 2021. 2, 6, 7", + "[41] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 35:32653-32666, 2022. 2", + "[42] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2", + "[43] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In CVPR, pages 1281-1292, 2020. 2", + "[44] Weicai Ye, Xinyue Lan, Shuo Chen, Yuhang Ming, Xingyuan Yu, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pvo: Panoptic visual odometry. In CVPR, pages 9579-9589, 2023. 2", + "[45] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018. 2", + "[46] Khalid Yousif, Alireza Bab-Hadiashar, and Reza Hosein- nezhad. An overview to visual odometry and visual slam: Applications to mobile robotics. Intelligent Industrial Systems, 1(4):289-311, 2015. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 2, 6", + "[48] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In ECCV, pages 523-542. Springer, 2022. 2, 3, 6, 7", + "[49] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, pages 19855-19865, 2023. 2", + "[50] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In CVPR, pages 15838-15847, 2021. 6", + "[51] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851–1858, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19853", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_model.json b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..754beca072aa7498899c8f195996d2fcfbd1e245 --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.856, + 0.153 + ], + "angle": 0, + "content": "LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.18, + 0.732, + 0.2 + ], + "angle": 0, + "content": "Weirong Chen\\(^{1,2*}\\) Le Chen\\(^{3}\\) Rui Wang\\(^{4}\\) Marc Pollefeys\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.202, + 0.862, + 0.221 + ], + "angle": 0, + "content": "\\(^{1}\\)TU Munich \\(^{2}\\)Munich Center for Machine Learning \\(^{3}\\)MPI for Intelligent Systems \\(^{4}\\)Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.259, + 0.314, + 0.274 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.292, + 0.474, + 0.61 + ], + "angle": 0, + "content": "Visual odometry estimates the motion of a moving camera based on visual input. Existing methods, mostly focusing on two-view point tracking, often ignore the rich temporal context in the image sequence, thereby overlooking the global motion patterns and providing no assessment of the full trajectory reliability. These shortcomings hinder performance in scenarios with occlusion, dynamic objects, and low-texture areas. To address these challenges, we present the Long-term Effective Any Point Tracking (LEAP) module. LEAP innovatively combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Moreover, LEAP's temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty. Based on these traits, we develop LEAP-VO, a robust visual odometry system adept at handling occlusions and dynamic scenes. Our mindful integration showcases a novel practice by employing long-term point tracking as the front-end. Extensive experiments demonstrate that the proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.21, + 0.655 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.47, + 0.802 + ], + "angle": 0, + "content": "Visual odometry (VO) calculates the camera's position and movement from an image sequence. It is widely used in various fields including robotics [46], mixed reality [23], and autonomous driving [15]. The performance of VO highly depends on the accuracy of its point tracking front-end, which recovers relative camera motion between consecutive frames using projective geometry. It further requires the point correspondences to be static, as dynamic points can lead to inaccurate motion estimation and pose recovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.864 + ], + "angle": 0, + "content": "Despite significant advances in feature tracking and optical flow estimation, most existing VO methods mainly focus on the two-view case, i.e., computing feature correspondences between a pair of images. However, these methods" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.262, + 0.692, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.439, + 0.634, + 0.451 + ], + "angle": 0, + "content": "Visual Odometry" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.262, + 0.882, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.434, + 0.857, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.464, + 0.893, + 0.561 + ], + "angle": 0, + "content": "Figure 1. Comparison between two-view approach and multi-view approach for visual odometry. (a) In the two-view approach, correspondences are derived for every image pair and concatenated. Managing occlusion becomes challenging without the temporal context. (b) In the multi-view approach, long-term point trajectories can be obtained all at once, enabling the detection of occlusion and tracking points under partial occlusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.779 + ], + "angle": 0, + "content": "are limited as they rely solely on pairwise relative motion, thus neglecting the valuable temporal information in image sequences. Consequently, they often struggle with capturing dynamic motion patterns in tracks. Another challenge arises from the need to handle occlusions. In two-view scenarios, occlusion is not explicitly modeled, leading to misidentification of occluded points as incorrect matches. To address these challenges effectively, we propose to introduce a long-term point tracker, which can recover the trajectories of specific points across the image sequence given a sparse set of these query points. Its capability of leveraging temporal context and reliably tracking queries over multiple frames even under occlusion, makes it highly suitable for VO applications (refer to Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Recently, Persistent Independent Particles (PIPs) [16] proposes a novel pipeline that brings the idea of iterative refinement into the long-term point tracking task, with the local correlation evidence to maintain feasible computation. While it is proficient in leveraging the temporal context present in the image sequence, it ignores the inter-track information in the image sequence and processes each query point independently. Consequently, PIPs struggles to cap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Work done at Microsoft Mixed Reality & AI Lab Zurich for a Master's degree." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19844" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.198 + ], + "angle": 0, + "content": "ture the global motion patterns from the image sequence. Additionally, it lacks a reliability metric for its trajectory estimates, hindering its application in quality-sensitive tasks like video segmentation and visual odometry. To mitigate the aforementioned issues, we propose a pipeline that explores dynamic track estimation and temporal probabilistic formulation for long-term point tracking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.471, + 0.485 + ], + "angle": 0, + "content": "In this paper, we present Long-term Effective Point Tracking (LEAP), aiming to estimate the reliability of predicted correspondences and track moving object trajectories in dynamic scenes. To achieve this, we harness the global motion patterns—often overlooked in existing methods—through an anchor-based dynamic track estimation module, leveraging visual, temporal, and inter-track information to differentiate between the static and dynamic elements. We also incorporate a temporal probabilistic approach into our LEAP pipeline to handle uncertainties and refine the distribution of point correspondences iteratively, showcasing superior performance in complex real-world situations. More importantly, we pioneer the usage of continuous motion estimation from long-term point tracking to construct a robust visual odometry system. Our LEAP-VO distinguishes itself through its thoughtful integration of visual features, track distribution, and global motion patterns, significantly enhancing both the performance and robustness of visual odometry in dynamic environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.496, + 0.219, + 0.511 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Long-term Point Tracking. In the realm of long-term point tracking or Tracking Any Point (TAP), various methods have been introduced recently. PIPs [16] reframes pixel tracking as a long-term motion estimation with local information and improves tracking accuracy by associating each pixel with a trajectory that spans multiple image frames. The subsequent works enhance performance by integrating spatial context features [2], tracking multiple queries concurrently [18], and extending temporal fields and templates [49]. TAP-Net [9] adopts a different method by locating candidate correspondences through a global cost volume, improved by TAPIR's trajectory refinement [10] and BootsTAP's semi-supervised training [11]. Other approaches explore self-supervision targets to learn visual correspondence via powerful feature representation [5, 17, 36]. Recently, OmniMotion [37] provides a pure test-time optimization method that infers the point trajectory and occlusion directly from videos with optical flow predictions. However, its slow inference speed makes it unsuitable for real-time applications such as visual odometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Monocular Visual Odometry. Visual Odometry (VO) can be categorized into indirect (feature-based) and direct methods. Indirect methods, such as MonoSLAM [8] and ORB-SLAM [25], mostly rely on feature extraction and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.41 + ], + "angle": 0, + "content": "matching to estimate the camera pose. Direct methods, conversely, skip feature extraction and matching and utilize intensity to optimize camera pose and 3D scene point positions based on photometric errors [12, 13, 38], showing better performance on low-texture regions where keypoint tracking can easily fail. However, they are more vulnerable to illumination changes and issues with rolling shutter cameras. Advancements in deep learning have spawned end-to-end monocular VO methods in supervised [3, 28, 30, 33, 39, 40, 43, 44] and unsupervised [21, 27, 45, 51] settings. DeepVO [39] utilizes Recurrent Neural Networks to model sequential data, while SfMLearner [51] develops an unsupervised framework to learn the depth and camera motion from unlabeled monocular videos. TartanVO [40] proposes a VO model that generalizes to multiple datasets and real-world scenarios. DROID-SLAM [32] follows learning-based optimization from RAFT [31] and proposes an end-to-end system, integrating flow, confidence, and geometric optimization via iterative GRU updates. The followed-up work DPVO [33] further cuts computational complexity by replacing dense feature tracking with sparse patch tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Dynamic Track Estimation. In addition to enhancing accuracy, numerous studies have focused on the detection of dynamic objects within videos. Methods based on segmentation, such as DynaSLAM [1] and RCVD [20], employ semantic cues from pre-trained models to exclude dynamic regions. However, these segmentation-based methods are constrained to predefined object categories and struggle to capture real-world motion (for example, differentiating between a stationary and a moving car). Alternatively, DyanVO [28] and ParticleSfM [48] adopt a trajectory-based motion detection approach. This method addresses dynamic scenes using estimated point track information, thereby generating reliable camera trajectories in complex motion scenarios. Nonetheless, they depend on dense optical flow estimation to capture global motion patterns. Furthermore, dynamic tracks can be detected using additional depth information for a 3D consistency check [7, 47], or by optimization based on photo consistency [35, 41, 42]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.593, + 0.725 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Our goal is to tackle the aforementioned challenges in VO systems, particularly in handling dynamic environments, temporal occlusions, and low-texture regions. To this end, we introduce our LEAP module, which incorporates dynamic track estimation and temporal probabilistic formulation. Furthermore, we illustrate how this module serves as the cornerstone for constructing our advanced VO system." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.794, + 0.864 + ], + "angle": 0, + "content": "3.1. Preliminary: Tracking Any Point" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Tracking arbitrary pixels in an image sequence is traditionally akin to estimating optical flow, where each pixel" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19845" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "is mapped to a corresponding location in a subsequent frame. However, two-view approaches often overlook the rich temporal context within the image sequence, leading to suboptimal outcomes. Recent advancements address this by revisiting the particle video concept and improving long-term point tracking through learning-based techniques [10, 16, 18]. These TAP methods can track points over multiple frames and detect occlusions, significantly boosting tracking robustness under challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.41 + ], + "angle": 0, + "content": "**Formulation.** Given a sequence of \\(S\\) consecutive RGB images \\(\\mathbf{I} = [\\mathbf{I}_1, \\dots, \\mathbf{I}_S]\\), \\(\\mathbf{I}_s \\in \\mathbb{R}^{3 \\times H \\times W}\\) as input, the goal of TAP is to track a set of query points across these frames. For a specific query point with the 2D pixel coordinate \\(\\mathbf{x}_q \\in \\mathbb{R}^2\\) in frame \\(s_q\\), TAP predicts its trajectory across all \\(S\\) images as \\(\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_S]\\), \\(\\mathbf{x}_s \\in \\mathbb{R}^2\\), given \\(\\mathbf{x}_{s_q} = \\mathbf{x}_q\\). Additionally, TAP estimates the visibility of each point throughout the sequence as \\(\\mathbf{V} = [v_1, \\dots, v_S]\\), where \\(v_s \\in \\{0, 1\\}\\) is a binary label of point visibility indicating whether a point in frame \\(s\\) is visible or occluded. The TAP formulation for a single query can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.42, + 0.47, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {X}, \\mathbf {V}\\right) = \\operatorname {T A P} \\left(\\mathbf {I}, \\mathbf {x} _ {q}, s _ {q}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.675 + ], + "angle": 0, + "content": "PIPs [16], a notable learning-based TAP method, offers a promising solution featuring local cost volumes and iterative refinement. Initially, each image \\(\\mathbf{I}_s\\in \\mathbf{I}\\) is processed through a CNN feature extractor \\(\\mathcal{F}\\) to extract the feature map \\(\\mathbf{Y}_{\\mathbf{s}} = \\mathcal{F}(\\mathbf{I}_{s})\\). Point features \\(\\mathbf{f}_q\\) are computed by bilinear sampling at the query positions \\(\\mathbf{x_s}\\) on \\(\\mathbf{Y}_s\\). These point features from each image are then concatenated to form a point feature tensor \\(\\mathbf{F} = [\\mathbf{f}_1,\\dots ,\\mathbf{f}_S]\\). PIPs approaches the TAP problem by iteratively updating the state variables \\((\\mathbf{X},\\mathbf{F})\\), with the initial states obtained by duplicating the query positions and features. During the \\(k\\)-th iterative refinement, PIPs computes a multi-scale local cost volume \\(\\mathbf{C}[\\mathbf{X}^k ]\\) centered around the current estimated point position \\(\\mathbf{X}^k\\). Utilizing this local evidence, PIPs predicts updates for the state variables via" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.682, + 0.47, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\left(\\Delta \\mathbf {X}, \\Delta \\mathbf {F}\\right) = \\operatorname {R e f i n e r} \\left(\\mathbf {F} ^ {k}, \\operatorname {p o s} \\left(\\mathbf {X} ^ {k} - \\mathbf {x} _ {\\mathbf {q}}\\right), \\mathbf {C} ^ {k} \\left[ \\mathbf {X} ^ {k} \\right]\\right), \\\\ \\mathbf {X} ^ {k + 1} \\leftarrow \\mathbf {X} ^ {k} + \\Delta \\mathbf {X}, \\quad \\mathbf {F} ^ {k + 1} \\leftarrow \\mathbf {F} ^ {k} + \\Delta \\mathbf {F}, \\end{array} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.825 + ], + "angle": 0, + "content": "where \\(\\mathrm{pos}(\\cdot)\\) refers to the positional embedding. The final point feature \\(\\mathbf{F}^K\\) is used to predict point visibility with a simple linear projection layer \\(\\mathcal{G}_v\\), yielding \\(\\mathbf{V} = \\mathcal{G}_v(\\mathbf{F}^K)\\). Recent advances extend PIPs by replacing the MLP-based refiner with an attention-based refiner and adding spatial feature aggregation to track multiple queries together [18]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.433, + 0.849 + ], + "angle": 0, + "content": "3.2. Anchor-based Dynamic Track Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The established TAP model enables us to extract long-term point trajectories and thereby recover camera motion from an image sequence. However, dynamic environments" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.092, + 0.888, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.18, + 0.892, + 0.235 + ], + "angle": 0, + "content": "Figure 2. Distributed image gradient-based sampling with \\( k = 8 \\), \\( N_{a} = 64 \\). After computing the image gradient and pooling, we split the gradient map into \\( 8 \\times 8 \\) grids and select the point with the maximum gradient in each grid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.892, + 0.306 + ], + "angle": 0, + "content": "pose extra challenges for VO systems that rely on static 3D point correspondences linked to unchanging scene locations. Therefore, integrating dynamic track estimation into the VO system is crucial for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.893, + 0.489 + ], + "angle": 0, + "content": "We propose an innovative method for dynamic track estimation that leverages visual, temporal, and inter-track information. While common techniques like monocular segmentation [20] excel at identifying moving objects using visual cues but can erroneously label static objects, like parked cars, as dynamic. Trajectory-based methods [48], utilizing optical flows from two or multiple frames, can differentiate static and moving objects apart but tend to ignore visual information and can be inefficient due to dense trajectories computation. Merging these two distinct methods is nontrivial; however, the formulation of the TAP pipeline, enhanced with inter-track attention, offers a viable solution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Our method smartly integrates visual appearance, long-term motion, and inter-track cues, combining the strengths of segmentation and trajectory-based techniques. Utilizing TAP, we extract point features \\(\\mathbf{F}\\) and long-term trajectories \\(\\mathbf{X}\\) from the video input. These features capture visual information and the iterative refinement module exchanges information across tracks over time for accurate labeling. However, such a method depends on the query distribution and quantities. With only one query during inference time, it cannot utilize inter-track information. Likewise, if the queries focus on a single object, the system may overlook the motions of other objects due to similar motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Anchor-based Inter-track Attention. To better capture global motion patterns, we introduce additional anchors alongside the given queries, denoted as \\(\\mathcal{A}\\), \\(|\\mathcal{A}| = N_{a}\\). Ideally, the anchor set should exhibit two characteristics: (1) the anchors should be easy to track, and (2) they should be well-distributed to encapsulate global motion patterns. Given the image \\(\\mathbf{I}_{s_q}\\) from which the queries \\(\\mathbf{x}_q\\) are extracted (we maintain the same notations for both single and multiple queries for simplicity), we first employ the Sobel Kernel to derive the image gradient \\((\\mathbf{G}_x,\\mathbf{G}_y)\\) for image \\(\\mathbf{I}_{s_q}\\). We then apply an average pooling layer on gradient maps for smoothing and downscaling feature maps. Subsequently, we divide the gradient map into \\(k\\times k\\) sub-regions and select top \\(\\frac{N_a}{k^2}\\) pixels with the highest image gradient magnitude in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19846" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.452, + 0.107 + ], + "angle": 0, + "content": "each local grid. The process is demonstrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.198 + ], + "angle": 0, + "content": "To predict the dynamic label of a trajectory, we track queries with additional anchors and concatenate the original query features \\(\\mathbf{X}\\) and query trajectories \\(\\mathbf{F}\\) with anchor features \\(\\mathbf{X}_A\\) and anchor trajectories \\(\\mathbf{F}_A\\). We apply a shallow MLP layer \\(\\mathcal{G}_d\\) with average pooling over the temporal dimension. The dynamic track label \\(\\mathbf{m}_d\\) is estimated as" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.206, + 0.47, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\mathbf {m} _ {d} = \\operatorname {a v g p o o l} \\left(\\mathcal {G} _ {d} \\left(\\left[ \\mathbf {X} ^ {K}; \\mathbf {X} _ {A} ^ {K} \\right], \\left[ \\mathbf {F} ^ {K}; \\mathbf {F} _ {A} ^ {K} \\right]\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.469, + 0.34 + ], + "angle": 0, + "content": "Since newly added anchors often lack ground truth information, we employ a semi-supervised training scheme that tracks queries and anchors together to leverage inter-track attention but only calculates losses based on query predictions. This approach prevents the network from developing bias towards specific query sampling methods while learning dynamic track labels effectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.349, + 0.357, + 0.366 + ], + "angle": 0, + "content": "3.3. Temporal Probability Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.508 + ], + "angle": 0, + "content": "Supervised learning applied to large-scale datasets has yielded notable advancements in learning-based point tracking methods. Nonetheless, trajectory estimation errors remain a concern, especially under challenging conditions like occlusions, motion blur, or textureless regions. Recognizing the varying reliability of trajectories, integrating an uncertainty estimation mechanism into the correspondence estimation process is beneficial [6]. This is particularly crucial for quality-sensitive tasks such as visual odometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.781 + ], + "angle": 0, + "content": "To evaluate the reliability of point correspondences, we incorporate a probabilistic formulation into the TAP pipeline. Given an image sequence \\(\\mathbf{I}\\) and a single query point \\(\\mathbf{x}_q\\), our objective is to compute the conditional probability density of point trajectory as \\(p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q)\\). Recognizing the strong correlation between points within the same track, treating each point in the track as independent becomes suboptimal. Therefore, we propose to model the distribution of point trajectories by applying two multivariate distributions, respectively for the 2D coordinates. Let \\(\\mathbf{X} = [\\mathbf{a};\\mathbf{b}]\\), where \\(\\mathbf{a}\\in \\mathbb{R}^S\\) and \\(\\mathbf{b}\\in \\mathbb{R}^{S}\\) represent the \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\) coordinate of all points in \\(\\mathbf{X}\\), respectively. Assuming independence between these two coordinates, the joint probability distribution can be expressed as \\(p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q) = p(\\mathbf{a}|\\mathbf{I},\\mathbf{x}_q)\\cdot p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)\\). We adopt the multivariate Cauchy distribution, which is associated with its heavy tails distribution and is more stable for optimization. The probability density function (PDF) of a single coordinate is given by" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.789, + 0.473, + 0.84 + ], + "angle": 0, + "content": "\\[\np (\\mathbf {a} | \\mathbf {I}, \\mathbf {x} _ {q}) = \\frac {\\Gamma \\left(\\frac {1 + S}{2}\\right)}{\\Gamma \\left(\\frac {1}{2}\\right) \\pi^ {\\frac {S}{2}} \\left| \\boldsymbol {\\Sigma} _ {a} \\right| ^ {\\frac {1}{2}} \\left[ 1 + \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {a} ^ {- 1} \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) \\right] ^ {\\frac {1 + S}{2}}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "and a similar expression is applicable for \\(p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)\\). \\(\\Gamma\\) is the Gamma function. The parameters \\((\\boldsymbol{\\mu}_a,\\boldsymbol{\\Sigma}_a,\\boldsymbol{\\mu}_b,\\boldsymbol{\\Sigma}_b)\\) represent the location and scale matrices for the respective coordinates. During inference, the uncertainty associated" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.094, + 0.891, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Figure 3. LEAP Front-end. Once image feature maps are obtained, selected anchors aid in tracking. The queries and anchors are processed by a refiner to iteratively update states. The model outputs trajectory distribution, visibility, and dynamic track label." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.34 + ], + "angle": 0, + "content": "with each point is quantified by the sum of diagonal scale estimates at position \\((s,s)\\) as \\(\\phi (\\mathbf{x}_s) = \\pmb {\\Sigma}_a[s,s] + \\pmb {\\Sigma}_b[s,s]\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.893, + 0.611 + ], + "angle": 0, + "content": "Kernel-based Estimation. Rather than just predicting the positions of points within a trajectory, our model focuses on recovering the distribution of point trajectories by estimating its parameters. The location parameter predictions can be associated with the previous point trajectory as \\([\\pmb{\\mu}_a; \\pmb{\\mu}_b] = \\mathbf{X}\\). However, directly deriving a symmetric and positive definite scale matrix from model outputs is challenging. To address this, we employ a kernel-based approach with the linear kernel \\(K(\\mathbf{x}, \\mathbf{y}) = \\mathbf{x}^T \\mathbf{y}\\). During each iteration, two scale matrices \\((\\pmb{\\Sigma}_a, \\pmb{\\Sigma}_b)\\) are constructed by first applying two linear projection layers to the point features, represented as \\(\\mathbf{F}_a^k = \\mathcal{G}_a(\\mathbf{F}^k)\\) and \\(\\mathbf{F}_b^k = \\mathcal{G}_b(\\mathbf{F}^k)\\). We then compute the scale matrices as \\(\\pmb{\\Sigma}_a = K(\\mathbf{F}_a, \\mathbf{F}_a) + \\sigma \\mathbf{I}\\), and \\(\\pmb{\\Sigma}_b = K(\\mathbf{F}_b, \\mathbf{F}_b) + \\sigma \\mathbf{I}\\), where \\(\\sigma\\) is a small positive value and \\(\\mathbf{I}\\) is the identity matrix. The model parameters are refined through Maximum Likelihood Estimation (MLE) using a negative log-likelihood (NLL) loss function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Loss Functions. We supervise the point trajectory mainly with the NLL loss, which is based on the predicted distribution parameters and the ground-truth point trajectory. The main point trajectory loss is given as" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.695, + 0.892, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {k} ^ {K} \\gamma^ {K - k} \\mathcal {L} _ {N L L} \\left(\\mathbf {X} ^ {k}, \\mathbf {X} ^ {*}, \\boldsymbol {\\Sigma} _ {a} ^ {k}, \\boldsymbol {\\Sigma} _ {b} ^ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.745, + 0.82, + 0.759 + ], + "angle": 0, + "content": "where \\(K\\) is the number of iteration and \\(\\gamma = 0.8\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.806 + ], + "angle": 0, + "content": "For the visibility and dynamic track label supervision, we use the cross entropy loss with the estimates \\(\\mathbf{V},\\mathbf{m}_d\\) and ground truth \\(\\mathbf{V}^{*},\\mathbf{m}_{d}^{*}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.815, + 0.891, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {v i s}} = \\left(1 - \\mathbf {V} ^ {*}\\right) \\log \\left(1 - \\mathbf {V}\\right) + \\mathbf {V} ^ {*} \\log \\mathbf {V}. \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.835, + 0.891, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d y n} = \\left(1 - \\mathbf {m} _ {d} ^ {*}\\right) \\log \\left(1 - \\mathbf {m} _ {d}\\right) + \\mathbf {m} _ {d} ^ {*} \\log \\mathbf {m} _ {d}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.86, + 0.816, + 0.875 + ], + "angle": 0, + "content": "The total loss is a weighted sum of three losses:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.886, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = w _ {1} \\mathcal {L} _ {\\text {m a i n}} + w _ {2} \\mathcal {L} _ {\\text {v i s}} + w _ {3} \\mathcal {L} _ {\\text {d y n}}. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19847" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.087, + 0.506, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.099, + 0.858, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 4. LEAP-VO pipeline. Given a new image \\(\\mathbf{I}_t\\) received at time \\(t\\), the keypoint extractor extracts new keypoints from \\(\\mathbf{I}_t\\). Then, all the keypoints from the latest \\(S_{LP}\\)-frame \\(\\mathbf{I}_{t - S_{LP} + 1:t}\\) are tracked across all other frames within the current LEAP window, followed by a track filtering step to remove outliers. Finally, the local BA module is used on the current BA window to update the camera poses and 3D positions of the extracted keypoints. The colored arrows denote the moving direction of each module when a new image is received." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.362, + 0.196, + 0.377 + ], + "angle": 0, + "content": "3.4. LEAP-VO" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Leveraging the anchor-based dynamic track estimations and temporal probabilistic formulation, we develop our Long-term Effective Point Tracking (LEAP) pipeline, boosting the accuracy of point tracking and simultaneously capturing global motion patterns and uncertainties of the tracks (refer to Figure 3). In this section, we demonstrate how we can build visual odometry for dynamic scenes based on the long-term tracking capabilities of LEAP. Our LEAP-VO pipeline is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Keypoint Extraction and Tracking. In our VO system, the LEAP model serves as the front-end. When a new image frame \\(\\mathbf{I}_t\\) is captured at time \\(t\\), we sample \\(N\\) new keypoints from this image to initiate point tracking. We consistently employ the distributed image gradient-based sampling method (refer to Section 3.2) for keypoint selection due to its favorable properties in feature distinctiveness and distribution. This approach also obviates the need for additional anchors during tracking. The LEAP model then tracks these keypoints across the last \\(S_{LP}\\) frames, performing bidirectional tracking (both forward and backward in time). Notably, \\(S_{LP}\\) can exceed the model window \\(S\\), thus enabling more expansive tracking capabilities through sequential prediction [16, 18]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Track Filtering. Theoretically, precise camera pose estimation between two images can be attained with just a few accurate correspondences under certain conditions. However, the quality of correspondences obtained from feature matching or optical flow can vary. Traditional VO pipelines often use RANSAC [14] to filter out incorrect matches, but their efficiency decreases with more keypoints and views. Our approach, in contrast, bypasses the sampling-based method and uses trajectory quality assessments from LEAP" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.363, + 0.822, + 0.378 + ], + "angle": 0, + "content": "for more efficient and informative track filtering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.486 + ], + "angle": 0, + "content": "(i) Visibility and Dynamic Track Label: The LEAP model has the capability to estimate not only the 2D point positions within a trajectory but also the visibility of each point and the dynamic track label of the entire trajectory. Thresholds \\(\\gamma_v\\) for visibility and \\(\\gamma_d\\) for dynamic track labels are set to ensure that only visible and static points are utilized in the bundle adjustment process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.61 + ], + "angle": 0, + "content": "(ii) Track Uncertainty: In Section 3.3, we introduce probabilistic modeling to LEAP, allowing us to gauge confidence via distribution measurements. We select only high-confidence points, which correspond to low uncertainty measurements. Denoting our model's uncertainty estimation as \\(\\Phi\\), we retain points that exhibit high confidence using the criterion \\(\\Phi \\leq Q(\\gamma_u)\\), where \\(\\gamma_u\\) is the uncertainty quantile and \\(Q:[0,1] \\to \\mathbb{R}\\) is the quantile function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.703 + ], + "angle": 0, + "content": "(iii) Track Length: After filtering, we evaluate and remove tracks with insufficient observations, as bundle adjustment is more reliable with increased observations and wider baselines. Tracks with fewer than \\(\\gamma_{\\text{track}}\\) valid points are excluded from optimization cost computation due to their potential unreliability." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Sliding-window Optimization. With the point validation mask established, we proceed to define the optimization cost function for our sliding window bundle adjustment. The window size for local bundle adjustment, denoted as \\( S_{BA} \\), may differ from the point tracking window size \\( S_{LP} \\). Our geometric bundle adjustment aims to optimize camera poses \\( \\mathbf{T} \\) and 3D scene point positions \\( \\mathbf{Q} \\) by aligning the induced point trajectory from the projective relationship, with the estimated point trajectory from LEAP. We parameterize a 3D scene point \\( \\mathbf{Q}_i \\) by its 2D keypoint location \\( \\mathbf{x}_i \\) in image \\( \\mathbf{I}_i \\) and depth \\( d_i \\). Let \\( \\mathrm{LEAP}_{i\\rightarrow j} \\) denote the mapping of tracking keypoint from \\( \\mathbf{I}_i \\) to \\( \\mathbf{I}_j \\) using LEAP model. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.308, + 0.108 + ], + "angle": 0, + "content": "reprojection cost is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.119, + 0.469, + 0.175 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S _ {B A}} \\sum_ {n} w _ {i \\rightarrow j, n} \\| \\mathcal {P} (\\mathbf {T} _ {i}, \\mathbf {T} _ {j}, \\mathbf {K}, d _ {i, n}) - \\\\ \\operatorname {L E A P} _ {i \\rightarrow j} (\\mathbf {x} _ {i, n}) \\| _ {\\rho}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.471, + 0.309 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_{\\rho}\\) is the distance metric, and the weight \\(w_{i\\rightarrow j,n}\\) is derived from the track filtering results. We use the GaussNewton method [34] to optimize Eq. (9) for \\(K_{BA}\\) iterations. At each iteration, we compute the camera poses update \\(\\Delta \\xi^{(k)}\\in \\mathfrak{se}(3)\\) (lie-algebra corresponding to \\(\\mathbf{T}\\)), and the depth update \\(\\Delta \\mathbf{D}^{(k)}\\) for each point. The optimization can be solved efficiently with the Schur decomposition. We use the robust Huber loss function for the distance metric." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.325, + 0.202, + 0.342 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.35, + 0.276, + 0.365 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.471, + 0.482 + ], + "angle": 0, + "content": "Replica. Replica [29] provides synthetic indoor environments for simulating the image-capturing process with a moving camera. We use the camera trajectories provided by [50] for the localization task, resulting in 16 camera trajectories derived from 8 static scenes. Each sequence comprises 900 frames with rendered RGB images. We use \"Sequence 1\" for the visual odometry evaluation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.598 + ], + "angle": 0, + "content": "MPI Sintel. The MPI Sintel [4] dataset contains dynamic image sequences from open-source 3D animated short films. These sequences consist of large and complex object motion, along with other imaging effects such as motion blur and defocus blur. Following prior works [20, 48], we assess dynamic VO performance on the MPI Sintel dataset, with each sequence containing 20 to 50 image frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.73 + ], + "angle": 0, + "content": "TartanAir Shibuya. The TartanAir Shibuya dataset [26] features dynamic scenes from two scenarios: \"Standing Humans\" and \"Road Crossing\". The majority of the humans remain stationary in \"Standing Humans\" sequences, while \"Road Crossing\" sequences provide a more challenging setting where multiple humans are moving in different directions. Each sequence in the dataset includes 100 frames with more than 30 tracked moving humans." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.471, + 0.906 + ], + "angle": 0, + "content": "Metrics. For visual odometry, we compare the Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot) following [33, 47, 48]. ATE measures how much the estimated trajectory deviated from the ground truth trajectory, which is calculated as the root mean square error (RMSE) overall all pose transformations after scaling and alignment. \"RPE trans\" computes translation errors made over a certain distance (meter), and \"RPE rot\" computes the estimated rotation over a certain distance (degree). Both \"RPE trans\" and \"RPE rot\" are calculated on all poses and then averaged." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.226 + ], + "angle": 0, + "content": "
MethodReplica
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]0.0860.0300.650
DynaSLAM [1]0.0390.0170.366
DROID-SLAM [32]0.2670.0362.631
TartanVO [40]0.4060.0362.063
DytanVO [28]0.2890.0352.146
DPVO [33]0.2570.0362.635
LEAP-VO (Ours)0.2040.0301.992
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 1. Camera tracking results on Replica [29]. The statistics demonstrate that our method achieves better results than other competitive baselines." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.279, + 0.89, + 0.416 + ], + "angle": 0, + "content": "
MethodMPI Sintel
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]XXX
DynaSLAM [1]XXX
DROID-SLAM [32]0.1750.0841.912
TartanVO [40]0.2380.0931.305
DytanVO [28]0.1310.0971.538
DPVO [33]0.0760.0781.722
LEAP-VO (Ours)0.0370.0551.263
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "Table 2. Camera tracking results on MPI Sintel [4]. The statistics show that our method outperforms other competitive baselines. 'X' denotes the tracking failure case." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.468, + 0.715, + 0.484 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.893, + 0.734 + ], + "angle": 0, + "content": "We employ CoTracker [18], a recent PIPs variant, as our base TAP model. We train our model on the TAP-Vid-Kubric [9] training set, following [18]. We load the pretrained CoTracker weights for the image feature extractor and train LEAP for 100,000 steps using 4 NVIDIA A100 GPUs in parallel. During training, we use \\( K = 4 \\) steps for iterative refinements, \\( N = 256 \\) for the number of queries, \\( S = 8 \\) for the model window, and \\( S_{LP} = 12 \\) for the tracking window. For inter-track anchor-based attention, the number of anchors is 64. The loss weights are set to \\( w_{1} = 1.0, w_{2} = 0.5, w_{3} = 0.5 \\). In LEAP-VO, we perform 4 steps for each bundle adjustment with the Huber Loss for distance metric. The bundle adjustment window size \\( S_{BA} \\) is set to 15. Track filtering parameters are set at \\( \\gamma_v = 0.9, \\gamma_d = 0.9, \\gamma_u = 0.8, \\gamma_{track} = 3 \\). Please refer to our supplementary materials for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.729, + 0.759 + ], + "angle": 0, + "content": "4.3. Visual Odometry Results" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.857 + ], + "angle": 0, + "content": "We compare our methods against the state-of-the-art SLAM and VO approaches, including ORB-SLAM2 [24], DynaSLAM [1], DROID-SLAM [32], TartanVO [40], DytanVO [28], and DPVO [33]. Evaluations are conducted on the static Replica dataset [29] as well as on the dynamic datasets MPI-Sintel [4] and TartanAir-Shibuya [26]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Replica. We compare our method with other baselines on the Replica dataset, which contains eight static scenes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19849" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.845, + 0.237 + ], + "angle": 0, + "content": "
MethodStandingHumanRoadCrossing (Easy)RoadCrossing (Hard)Average
01020304050607
ORB-SLAM2 [24] w/ mask0.07880.00600.06570.01960.01481.09840.84760.3044
DynaSLAM [1]X0.88360.39070.41960.49250.64460.6539(0.5808)
AirDOS [26] w/ mask0.06060.01930.09510.03310.02060.22300.56250.1449
DROID-SLAM [32]0.00510.00730.01030.01200.22780.02530.57880.1238
TartanVO [40]0.06000.16050.27620.18140.21740.32280.50090.2456
DytanVO [28]0.03270.10170.06080.05160.07550.03650.06600.0607
DPVO [39]0.05390.16030.17880.28580.11960.11190.14980.1514
LEAP-VO (Ours)0.00810.01850.03990.03170.01130.06890.02460.0290
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Table 3. ATE (m) results on TartanAir-Shibuya Sequences [26]. (\\(\\cdot\\)) denotes averaging on valid scenes only. We can observe that our method shows better results for ATE on most sequences, outperforming all SLAM and VO baselines on average ATE." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.274, + 0.28, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.424, + 0.238, + 0.436 + ], + "angle": 0, + "content": "(a) RoadCrossing 04" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.273, + 0.482, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.329, + 0.424, + 0.44, + 0.436 + ], + "angle": 0, + "content": "(b) RoadCrossing 05" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.273, + 0.685, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.424, + 0.642, + 0.436 + ], + "angle": 0, + "content": "(c) RoadCrossing 06" + }, + { + "type": "image", + "bbox": [ + 0.693, + 0.274, + 0.886, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.424, + 0.845, + 0.436 + ], + "angle": 0, + "content": "(d) RoadCrossing 07" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.893, + 0.468 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of camera trajectory estimation on TartanAir-Shibuya [26]. The visualizations show that our method provides more robust and accurate camera pose trajectories, especially in hard cases (RoadCrossing 06 and RoadCrossing 07)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.63 + ], + "angle": 0, + "content": "without any moving objects. Challenges arise in two test sequences that contain prolonged occlusions, leading to the potential camera tracking loss. As illustrated in Table 1, our approach outperforms other VO methods on all three metrics, showing effectiveness under static environments and the capability to handle occlusions. Further insights are available in Section 4.5. For classical methods, ORBSLAM2 [24] and DynaSLAM [1], we observe tracking loss and resuming via re-localization, which reduces long-term drift and shows high accuracy for camera pose tracking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.834 + ], + "angle": 0, + "content": "MPI-Sintel. We also evaluate our VO camera trajectory accuracy on 14 dynamic scenes from the MPI Sintel dataset, following the setting of [48]. As shown in Table 2, our method substantially surpasses other VO baselines in dynamic scenes, thanks to our temporal uncertainty estimation and dynamic track detection. Traditional feature-based methods like ORB-SLAM2 [24] and DynaSLAM [1] sometimes fail to track camera movements in highly dynamic sequences, primarily due to outliers from dynamic object correspondences and segmentation-based only dynamic track estimation. In contrast, our approach, which considers visual, temporal, and inter-track information, demonstrates notable improvements in dynamic scene tracking accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "TartanAir-Shibuya. Following DytanVO [28], we further evaluate the VO performance using the dynamic TartanAir-Shibuya dataset. We assess the Average Trajectory Error (ATE) across seven diverse scenes, ranging from simple to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.553 + ], + "angle": 0, + "content": "challenging. As depicted in Table 3, our method consistently outperforms all other SLAM and VO baselines in the majority of these scenes. Moreover, trajectory visualizations in Figure 5 also highlight the robust camera tracking capabilities of our method in dynamic environments." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.56, + 0.805, + 0.576 + ], + "angle": 0, + "content": "4.4. Dynamic Track Estimation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.893, + 0.812 + ], + "angle": 0, + "content": "In our LEAP formulation, we have introduced the anchor-based dynamic track estimation module to distinguish between static and dynamic point trajectories, which provides valuable information for handling dynamic scenes in visual odometry. Our method leverages both point feature information and inter-track relations to predict the dynamic track label for each trajectory. To demonstrate the performance of the proposed dynamic track estimation module, we take the LEAP model and evaluate it on novel image sequences from diverse datasets. For a comprehensive demonstration, we uniformly sample \\(16 \\times 16\\) queries among the images and track them for \\(T = 8\\) frames. The results, shown in Figure 6, illustrate that LEAP can effectively handle various scenarios, including single-object motion, complex and rapid motion, and multi-object motion with occlusion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.654, + 0.833 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Track Filtering. To highlight the efficacy of our track filtering module in LEAP-VO, we conduct an ablation study on different masks used in track filtering. We study three types of filtering criteria: visibility, dynamic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19850" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.091, + 0.464, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.323, + 0.47, + 0.38 + ], + "angle": 0, + "content": "Figure 6. Visualization of dynamic track estimation on DAVIS [19], MPI-Sintel [4], and TartanAir-Shibuya [26]. Odd columns: all point trajectories. Even columns: estimated dynamic point trajectories." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.385, + 0.466, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.528, + 0.468, + 0.571 + ], + "angle": 0, + "content": "Figure 7. Visualization of point-wise uncertainty measurements. Keypoints with the lowest \\(20\\%\\) (left) and highest \\(20\\%\\) (right) uncertainty are shown in green and yellow, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.673 + ], + "angle": 0, + "content": "track, and confidence filtering. Table 4 shows the impact of applying various filtering types on the dynamic MPI Sintel benchmark. Without dynamic filtering or confidence filtering, the ATE increases by \\(46\\%\\) and \\(130\\%\\), respectively. Combining all three types of filtering methods altogether, we achieve the best results in all three metrics." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.679, + 0.462, + 0.773 + ], + "angle": 0, + "content": "
MethodTrack FilteringMPI Sintel
Vis.Dyn.Conf.ATERPE transRPE rot
LEAP-VO✓--0.1180.0781.340
✓✓-0.0850.0621.281
✓-✓0.0540.0581.274
✓✓✓0.0370.0551.263
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.776, + 0.468, + 0.804 + ], + "angle": 0, + "content": "Table 4. Effect of the proposed track filtering method on MPI-Sintel [4]. Tracking using all criteria yields the best results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Uncertainty Estimation. As illustrated in Figure 7, we demonstrate the LEAP's capability to provide meaningful per-point uncertainty. It models the uncertainty as distribution parameters for both the entire track and individual points. Two key observations can be found from Figure 7. Firstly, our model reliably identifies high-uncertainty" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "points, typically located in low-texture areas or in regions with repetitive patterns. Secondly, although not explicitly supervised, the model tends to assign higher uncertainty to dynamic objects, which are inherently more challenging to track. These findings highlight the efficacy of LEAP's probabilistic formulation in handling uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Keypoint Extraction. Table 5 showcases the effectiveness of various keypoint extraction strategies for VO performance. Despite its distinctive features, SIFT extraction provides inferior performance due to its unpredictable keypoint distribution. In contrast, the random sampling tends to generate well-separated keypoints, leading to a performance similar to SIFT's. Our distributed image gradient-based sampling shows the best performance, underscoring the significance of feature distribution and distinctiveness." + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.328, + 0.846, + 0.368 + ], + "angle": 0, + "content": "
Policy:UniformRandomSIFT [22]Ours
ATE (m):0.15110.03660.03550.0290
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Table 5. Comparison of keypoint extraction policies (\\(N_{a} = 64\\)) on TartanAir-Shibuya [26]. Our distributed image-gradient sampling method achieves better results compared to other policies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.492 + ], + "angle": 0, + "content": "Baseline comparison. We evaluate the pretrained CoTracker [18] for the end VO performance. Table 6 shows that LEAP front-end outperforms CoTracker on all datasets, demonstrating the effectiveness of the proposed dynamic track estimation and temporal probability modeling." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.497, + 0.882, + 0.567 + ], + "angle": 0, + "content": "
Front-endReplicaMPI SintelShibuya
ATERPE transRPE rotATERPE transRPE rotATE
CoTracker0.3010.0312.0430.1260.0711.3660.160
LEAP0.2040.0301.9920.0370.0551.2630.029
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.568, + 0.891, + 0.594 + ], + "angle": 0, + "content": "Table 6. Baseline comparison with CoTracker front-end [18]. Our LEAP front-end achieves better results on all datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.619 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce Long-term Effective Any Point Tracking (LEAP), designed to tackle the limitations of previous point-tracking methods. By leveraging visual, inter-track, and temporal cues within an anchor-based dynamic track estimation module, LEAP captures global motion patterns and effectively discriminates between static and dynamic elements. Moreover, our approach incorporates a temporal probabilistic formulation into the point-tracking pipeline, emulating recursive Bayesian filtering through a learning-based refinement module. This enables our model to accurately assess the reliability of point trajectory estimates. Using LEAP as the front-end, we then develop the LEAP-VO system, featuring novel integration of visual information, global motion patterns, and track distribution for handling dynamic scenes. Additionally, our takes for LEAP-VO can be integrated with other TAP-based front-ends, offering the potential to markedly improve camera-tracking accuracy and robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19851" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Berta Bescos, José M Fácil, Javier Civera, and José Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, 3(4):4076-4083, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[3] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In CVPR, pages 2560–2568, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625. Springer-Verlag, 2012. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.469, + 0.381 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[6] Weirong Chen, Suryansh Kumar, and Fisher Yu. Uncertainty-driven dense two-view structure from motion. IEEE Robotics and Automation Letters, 8(3):1763-1770, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In ICCV, pages 7063-7072, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[8] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE TPAMI, 29(6):1052-1067, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[9] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 35:13610-13626, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.593, + 0.469, + 0.648 + ], + "angle": 0, + "content": "[10] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.649, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[11] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.718, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[12] Jakob Engel, Thomas Schöps, and Daniel Cremers. Ldslam: Large-scale direct monocular slam. In ECCV, pages 834-849. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.761, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[13] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE TPAMI, 40(3):611-625, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 1" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, pages 59-75. Springer, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 33: 19545-19560, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.292 + ], + "angle": 0, + "content": "[19] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.296, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[21] Shunkai Li, Xin Wang, Yingdian Cao, Fei Xue, Zike Yan, and Hongbin Zha. Self-supervised deep visual odometry with online adaptation. In CVPR, pages 6339-6348, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[22] David G Lowe. Object recognition from local scale-invariant features. In ICCV, pages 1150-1157. IEEE, 1999. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[23] Annette Mossel and Manuel Kroeter. Streaming and exploration of dynamically changing dense 3d reconstructions in immersive virtual reality. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR-Adjunct), pages 43–48. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.499, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: an open-source SLAM system for monocular, stereo and RGB-D cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[25] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.614, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[26] Yuheng Qiu, Chen Wang, Wenshan Wang, Mina Henein, and Sebastian Scherer. Airdos: Dynamic slam benefits from articulated objects. In ICRA, pages 8047-8053. IEEE, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[27] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In CVPR, pages 12240-12249, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[28] Shihao Shen, Yilin Cai, Wenshan Wang, and Sebastian Scherer. Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environments. In ICRA, pages 4048-4055. IEEE, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[29] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In ICLR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19852" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[31] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.137, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[32] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34: 16558-16569, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.181, + 0.468, + 0.208 + ], + "angle": 0, + "content": "[33] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.211, + 0.468, + 0.265 + ], + "angle": 0, + "content": "[34] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In International workshop on vision algorithms, pages 298-372. Springer, 1999. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.268, + 0.468, + 0.309 + ], + "angle": 0, + "content": "[35] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. Neuraliff: Segmenting 3d objects that move in egocentric videos. In 3DV, pages 910-919. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.312, + 0.468, + 0.353 + ], + "angle": 0, + "content": "[36] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.356, + 0.468, + 0.409 + ], + "angle": 0, + "content": "[37] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.413, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[38] Rui Wang, Martin Schworer, and Daniel Cremers. Stereo dso: Large-scale direct sparse visual odometry with stereo cameras. In ICCV, pages 3903-3911, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.457, + 0.468, + 0.511 + ], + "angle": 0, + "content": "[39] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In ICRA, pages 2043-2050. IEEE, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.514, + 0.468, + 0.554 + ], + "angle": 0, + "content": "[40] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In CoRL, pages 1761-1772. PMLR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.558, + 0.468, + 0.612 + ], + "angle": 0, + "content": "[41] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 35:32653-32666, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.616, + 0.468, + 0.683 + ], + "angle": 0, + "content": "[42] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.468, + 0.74 + ], + "angle": 0, + "content": "[43] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In CVPR, pages 1281-1292, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.744, + 0.468, + 0.798 + ], + "angle": 0, + "content": "[44] Weicai Ye, Xinyue Lan, Shuo Chen, Yuhang Ming, Xingyuan Yu, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pvo: Panoptic visual odometry. In CVPR, pages 9579-9589, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.802, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[45] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[46] Khalid Yousif, Alireza Bab-Hadiashar, and Reza Hosein- nezhad. An overview to visual odometry and visual slam: Applications to mobile robotics. Intelligent Industrial Systems, 1(4):289-311, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[48] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In ECCV, pages 523-542. Springer, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[49] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, pages 19855-19865, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[50] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In CVPR, pages 15838-15847, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[51] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851–1858, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19853" + } + ] +] \ No newline at end of file diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_origin.pdf b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6fd3e70f1246e38a0a735c14ae16cd528f21d397 --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/03c2e012-3152-4602-8d61-0831334e6fb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a268bb743a9a03d7673d63b9444ecceae218710d6c12e2e17b9821b0a37f64c +size 10330150 diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/full.md b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37c7546053ced3c47881f1ad5f0a0285cb71f0b4 --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/full.md @@ -0,0 +1,314 @@ +# LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry + +Weirong Chen $^{1,2*}$ Le Chen $^{3}$ Rui Wang $^{4}$ Marc Pollefeys $^{4}$ + +$^{1}$ TU Munich $^{2}$ Munich Center for Machine Learning $^{3}$ MPI for Intelligent Systems $^{4}$ Microsoft + +# Abstract + +Visual odometry estimates the motion of a moving camera based on visual input. Existing methods, mostly focusing on two-view point tracking, often ignore the rich temporal context in the image sequence, thereby overlooking the global motion patterns and providing no assessment of the full trajectory reliability. These shortcomings hinder performance in scenarios with occlusion, dynamic objects, and low-texture areas. To address these challenges, we present the Long-term Effective Any Point Tracking (LEAP) module. LEAP innovatively combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Moreover, LEAP's temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty. Based on these traits, we develop LEAP-VO, a robust visual odometry system adept at handling occlusions and dynamic scenes. Our mindful integration showcases a novel practice by employing long-term point tracking as the front-end. Extensive experiments demonstrate that the proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks. + +# 1. Introduction + +Visual odometry (VO) calculates the camera's position and movement from an image sequence. It is widely used in various fields including robotics [46], mixed reality [23], and autonomous driving [15]. The performance of VO highly depends on the accuracy of its point tracking front-end, which recovers relative camera motion between consecutive frames using projective geometry. It further requires the point correspondences to be static, as dynamic points can lead to inaccurate motion estimation and pose recovery. + +Despite significant advances in feature tracking and optical flow estimation, most existing VO methods mainly focus on the two-view case, i.e., computing feature correspondences between a pair of images. However, these methods + +![](images/7a76ff2249955c9c8077c455b55618895587c7dd0639e49aad6be7800470b4de.jpg) + +![](images/070ad67c83554f8e25cf7cd79e091da3bce520463dff2693730f48a768a657d4.jpg) + +![](images/faed9191be70cfa06e13b268052bd065dc87863b05880998ec81b4c6d869c1b0.jpg) +Visual Odometry +Figure 1. Comparison between two-view approach and multi-view approach for visual odometry. (a) In the two-view approach, correspondences are derived for every image pair and concatenated. Managing occlusion becomes challenging without the temporal context. (b) In the multi-view approach, long-term point trajectories can be obtained all at once, enabling the detection of occlusion and tracking points under partial occlusions. + +are limited as they rely solely on pairwise relative motion, thus neglecting the valuable temporal information in image sequences. Consequently, they often struggle with capturing dynamic motion patterns in tracks. Another challenge arises from the need to handle occlusions. In two-view scenarios, occlusion is not explicitly modeled, leading to misidentification of occluded points as incorrect matches. To address these challenges effectively, we propose to introduce a long-term point tracker, which can recover the trajectories of specific points across the image sequence given a sparse set of these query points. Its capability of leveraging temporal context and reliably tracking queries over multiple frames even under occlusion, makes it highly suitable for VO applications (refer to Figure 1). + +Recently, Persistent Independent Particles (PIPs) [16] proposes a novel pipeline that brings the idea of iterative refinement into the long-term point tracking task, with the local correlation evidence to maintain feasible computation. While it is proficient in leveraging the temporal context present in the image sequence, it ignores the inter-track information in the image sequence and processes each query point independently. Consequently, PIPs struggles to cap + +ture the global motion patterns from the image sequence. Additionally, it lacks a reliability metric for its trajectory estimates, hindering its application in quality-sensitive tasks like video segmentation and visual odometry. To mitigate the aforementioned issues, we propose a pipeline that explores dynamic track estimation and temporal probabilistic formulation for long-term point tracking. + +In this paper, we present Long-term Effective Point Tracking (LEAP), aiming to estimate the reliability of predicted correspondences and track moving object trajectories in dynamic scenes. To achieve this, we harness the global motion patterns—often overlooked in existing methods—through an anchor-based dynamic track estimation module, leveraging visual, temporal, and inter-track information to differentiate between the static and dynamic elements. We also incorporate a temporal probabilistic approach into our LEAP pipeline to handle uncertainties and refine the distribution of point correspondences iteratively, showcasing superior performance in complex real-world situations. More importantly, we pioneer the usage of continuous motion estimation from long-term point tracking to construct a robust visual odometry system. Our LEAP-VO distinguishes itself through its thoughtful integration of visual features, track distribution, and global motion patterns, significantly enhancing both the performance and robustness of visual odometry in dynamic environments. + +# 2. Related Work + +Long-term Point Tracking. In the realm of long-term point tracking or Tracking Any Point (TAP), various methods have been introduced recently. PIPs [16] reframes pixel tracking as a long-term motion estimation with local information and improves tracking accuracy by associating each pixel with a trajectory that spans multiple image frames. The subsequent works enhance performance by integrating spatial context features [2], tracking multiple queries concurrently [18], and extending temporal fields and templates [49]. TAP-Net [9] adopts a different method by locating candidate correspondences through a global cost volume, improved by TAPIR's trajectory refinement [10] and BootsTAP's semi-supervised training [11]. Other approaches explore self-supervision targets to learn visual correspondence via powerful feature representation [5, 17, 36]. Recently, OmniMotion [37] provides a pure test-time optimization method that infers the point trajectory and occlusion directly from videos with optical flow predictions. However, its slow inference speed makes it unsuitable for real-time applications such as visual odometry. + +Monocular Visual Odometry. Visual Odometry (VO) can be categorized into indirect (feature-based) and direct methods. Indirect methods, such as MonoSLAM [8] and ORB-SLAM [25], mostly rely on feature extraction and + +matching to estimate the camera pose. Direct methods, conversely, skip feature extraction and matching and utilize intensity to optimize camera pose and 3D scene point positions based on photometric errors [12, 13, 38], showing better performance on low-texture regions where keypoint tracking can easily fail. However, they are more vulnerable to illumination changes and issues with rolling shutter cameras. Advancements in deep learning have spawned end-to-end monocular VO methods in supervised [3, 28, 30, 33, 39, 40, 43, 44] and unsupervised [21, 27, 45, 51] settings. DeepVO [39] utilizes Recurrent Neural Networks to model sequential data, while SfMLearner [51] develops an unsupervised framework to learn the depth and camera motion from unlabeled monocular videos. TartanVO [40] proposes a VO model that generalizes to multiple datasets and real-world scenarios. DROID-SLAM [32] follows learning-based optimization from RAFT [31] and proposes an end-to-end system, integrating flow, confidence, and geometric optimization via iterative GRU updates. The followed-up work DPVO [33] further cuts computational complexity by replacing dense feature tracking with sparse patch tracking. + +Dynamic Track Estimation. In addition to enhancing accuracy, numerous studies have focused on the detection of dynamic objects within videos. Methods based on segmentation, such as DynaSLAM [1] and RCVD [20], employ semantic cues from pre-trained models to exclude dynamic regions. However, these segmentation-based methods are constrained to predefined object categories and struggle to capture real-world motion (for example, differentiating between a stationary and a moving car). Alternatively, DyanVO [28] and ParticleSfM [48] adopt a trajectory-based motion detection approach. This method addresses dynamic scenes using estimated point track information, thereby generating reliable camera trajectories in complex motion scenarios. Nonetheless, they depend on dense optical flow estimation to capture global motion patterns. Furthermore, dynamic tracks can be detected using additional depth information for a 3D consistency check [7, 47], or by optimization based on photo consistency [35, 41, 42]. + +# 3. Method + +Our goal is to tackle the aforementioned challenges in VO systems, particularly in handling dynamic environments, temporal occlusions, and low-texture regions. To this end, we introduce our LEAP module, which incorporates dynamic track estimation and temporal probabilistic formulation. Furthermore, we illustrate how this module serves as the cornerstone for constructing our advanced VO system. + +# 3.1. Preliminary: Tracking Any Point + +Tracking arbitrary pixels in an image sequence is traditionally akin to estimating optical flow, where each pixel + +is mapped to a corresponding location in a subsequent frame. However, two-view approaches often overlook the rich temporal context within the image sequence, leading to suboptimal outcomes. Recent advancements address this by revisiting the particle video concept and improving long-term point tracking through learning-based techniques [10, 16, 18]. These TAP methods can track points over multiple frames and detect occlusions, significantly boosting tracking robustness under challenging scenarios. + +**Formulation.** Given a sequence of $S$ consecutive RGB images $\mathbf{I} = [\mathbf{I}_1, \dots, \mathbf{I}_S]$ , $\mathbf{I}_s \in \mathbb{R}^{3 \times H \times W}$ as input, the goal of TAP is to track a set of query points across these frames. For a specific query point with the 2D pixel coordinate $\mathbf{x}_q \in \mathbb{R}^2$ in frame $s_q$ , TAP predicts its trajectory across all $S$ images as $\mathbf{X} = [\mathbf{x}_1, \dots, \mathbf{x}_S]$ , $\mathbf{x}_s \in \mathbb{R}^2$ , given $\mathbf{x}_{s_q} = \mathbf{x}_q$ . Additionally, TAP estimates the visibility of each point throughout the sequence as $\mathbf{V} = [v_1, \dots, v_S]$ , where $v_s \in \{0, 1\}$ is a binary label of point visibility indicating whether a point in frame $s$ is visible or occluded. The TAP formulation for a single query can be described as + +$$ +\left(\mathbf {X}, \mathbf {V}\right) = \operatorname {T A P} \left(\mathbf {I}, \mathbf {x} _ {q}, s _ {q}\right). \tag {1} +$$ + +PIPs [16], a notable learning-based TAP method, offers a promising solution featuring local cost volumes and iterative refinement. Initially, each image $\mathbf{I}_s\in \mathbf{I}$ is processed through a CNN feature extractor $\mathcal{F}$ to extract the feature map $\mathbf{Y}_{\mathbf{s}} = \mathcal{F}(\mathbf{I}_{s})$ . Point features $\mathbf{f}_q$ are computed by bilinear sampling at the query positions $\mathbf{x_s}$ on $\mathbf{Y}_s$ . These point features from each image are then concatenated to form a point feature tensor $\mathbf{F} = [\mathbf{f}_1,\dots ,\mathbf{f}_S]$ . PIPs approaches the TAP problem by iteratively updating the state variables $(\mathbf{X},\mathbf{F})$ , with the initial states obtained by duplicating the query positions and features. During the $k$ -th iterative refinement, PIPs computes a multi-scale local cost volume $\mathbf{C}[\mathbf{X}^k ]$ centered around the current estimated point position $\mathbf{X}^k$ . Utilizing this local evidence, PIPs predicts updates for the state variables via + +$$ +\begin{array}{r} \left(\Delta \mathbf {X}, \Delta \mathbf {F}\right) = \operatorname {R e f i n e r} \left(\mathbf {F} ^ {k}, \operatorname {p o s} \left(\mathbf {X} ^ {k} - \mathbf {x} _ {\mathbf {q}}\right), \mathbf {C} ^ {k} \left[ \mathbf {X} ^ {k} \right]\right), \\ \mathbf {X} ^ {k + 1} \leftarrow \mathbf {X} ^ {k} + \Delta \mathbf {X}, \quad \mathbf {F} ^ {k + 1} \leftarrow \mathbf {F} ^ {k} + \Delta \mathbf {F}, \end{array} \tag {2} +$$ + +where $\mathrm{pos}(\cdot)$ refers to the positional embedding. The final point feature $\mathbf{F}^K$ is used to predict point visibility with a simple linear projection layer $\mathcal{G}_v$ , yielding $\mathbf{V} = \mathcal{G}_v(\mathbf{F}^K)$ . Recent advances extend PIPs by replacing the MLP-based refiner with an attention-based refiner and adding spatial feature aggregation to track multiple queries together [18]. + +# 3.2. Anchor-based Dynamic Track Estimation + +The established TAP model enables us to extract long-term point trajectories and thereby recover camera motion from an image sequence. However, dynamic environments + +![](images/f9a4b8f0c4199ea8b8569721db074c6591697e11ec43d5bbfbe71c1e30542119.jpg) +Figure 2. Distributed image gradient-based sampling with $k = 8$ , $N_{a} = 64$ . After computing the image gradient and pooling, we split the gradient map into $8 \times 8$ grids and select the point with the maximum gradient in each grid. + +pose extra challenges for VO systems that rely on static 3D point correspondences linked to unchanging scene locations. Therefore, integrating dynamic track estimation into the VO system is crucial for real-world applications. + +We propose an innovative method for dynamic track estimation that leverages visual, temporal, and inter-track information. While common techniques like monocular segmentation [20] excel at identifying moving objects using visual cues but can erroneously label static objects, like parked cars, as dynamic. Trajectory-based methods [48], utilizing optical flows from two or multiple frames, can differentiate static and moving objects apart but tend to ignore visual information and can be inefficient due to dense trajectories computation. Merging these two distinct methods is nontrivial; however, the formulation of the TAP pipeline, enhanced with inter-track attention, offers a viable solution. + +Our method smartly integrates visual appearance, long-term motion, and inter-track cues, combining the strengths of segmentation and trajectory-based techniques. Utilizing TAP, we extract point features $\mathbf{F}$ and long-term trajectories $\mathbf{X}$ from the video input. These features capture visual information and the iterative refinement module exchanges information across tracks over time for accurate labeling. However, such a method depends on the query distribution and quantities. With only one query during inference time, it cannot utilize inter-track information. Likewise, if the queries focus on a single object, the system may overlook the motions of other objects due to similar motion patterns. + +Anchor-based Inter-track Attention. To better capture global motion patterns, we introduce additional anchors alongside the given queries, denoted as $\mathcal{A}$ , $|\mathcal{A}| = N_{a}$ . Ideally, the anchor set should exhibit two characteristics: (1) the anchors should be easy to track, and (2) they should be well-distributed to encapsulate global motion patterns. Given the image $\mathbf{I}_{s_q}$ from which the queries $\mathbf{x}_q$ are extracted (we maintain the same notations for both single and multiple queries for simplicity), we first employ the Sobel Kernel to derive the image gradient $(\mathbf{G}_x,\mathbf{G}_y)$ for image $\mathbf{I}_{s_q}$ . We then apply an average pooling layer on gradient maps for smoothing and downscaling feature maps. Subsequently, we divide the gradient map into $k\times k$ sub-regions and select top $\frac{N_a}{k^2}$ pixels with the highest image gradient magnitude in + +each local grid. The process is demonstrated in Figure 2. + +To predict the dynamic label of a trajectory, we track queries with additional anchors and concatenate the original query features $\mathbf{X}$ and query trajectories $\mathbf{F}$ with anchor features $\mathbf{X}_A$ and anchor trajectories $\mathbf{F}_A$ . We apply a shallow MLP layer $\mathcal{G}_d$ with average pooling over the temporal dimension. The dynamic track label $\mathbf{m}_d$ is estimated as + +$$ +\mathbf {m} _ {d} = \operatorname {a v g p o o l} \left(\mathcal {G} _ {d} \left(\left[ \mathbf {X} ^ {K}; \mathbf {X} _ {A} ^ {K} \right], \left[ \mathbf {F} ^ {K}; \mathbf {F} _ {A} ^ {K} \right]\right)\right). \tag {3} +$$ + +Since newly added anchors often lack ground truth information, we employ a semi-supervised training scheme that tracks queries and anchors together to leverage inter-track attention but only calculates losses based on query predictions. This approach prevents the network from developing bias towards specific query sampling methods while learning dynamic track labels effectively. + +# 3.3. Temporal Probability Modeling + +Supervised learning applied to large-scale datasets has yielded notable advancements in learning-based point tracking methods. Nonetheless, trajectory estimation errors remain a concern, especially under challenging conditions like occlusions, motion blur, or textureless regions. Recognizing the varying reliability of trajectories, integrating an uncertainty estimation mechanism into the correspondence estimation process is beneficial [6]. This is particularly crucial for quality-sensitive tasks such as visual odometry. + +To evaluate the reliability of point correspondences, we incorporate a probabilistic formulation into the TAP pipeline. Given an image sequence $\mathbf{I}$ and a single query point $\mathbf{x}_q$ , our objective is to compute the conditional probability density of point trajectory as $p(\mathbf{X}|\mathbf{I},\mathbf{x}_q)$ . Recognizing the strong correlation between points within the same track, treating each point in the track as independent becomes suboptimal. Therefore, we propose to model the distribution of point trajectories by applying two multivariate distributions, respectively for the 2D coordinates. Let $\mathbf{X} = [\mathbf{a};\mathbf{b}]$ , where $\mathbf{a}\in \mathbb{R}^S$ and $\mathbf{b}\in \mathbb{R}^{S}$ represent the $\mathbf{X}$ and $\mathbf{Y}$ coordinate of all points in $\mathbf{X}$ , respectively. Assuming independence between these two coordinates, the joint probability distribution can be expressed as $p(\mathbf{X}|\mathbf{I},\mathbf{x}_q) = p(\mathbf{a}|\mathbf{I},\mathbf{x}_q)\cdot p(\mathbf{b}|\mathbf{I},\mathbf{x}_q)$ . We adopt the multivariate Cauchy distribution, which is associated with its heavy tails distribution and is more stable for optimization. The probability density function (PDF) of a single coordinate is given by + +$$ +p (\mathbf {a} | \mathbf {I}, \mathbf {x} _ {q}) = \frac {\Gamma \left(\frac {1 + S}{2}\right)}{\Gamma \left(\frac {1}{2}\right) \pi^ {\frac {S}{2}} \left| \boldsymbol {\Sigma} _ {a} \right| ^ {\frac {1}{2}} \left[ 1 + \left(\mathbf {a} - \boldsymbol {\mu} _ {a}\right) ^ {T} \boldsymbol {\Sigma} _ {a} ^ {- 1} \left(\mathbf {a} - \boldsymbol {\mu} _ {a}\right) \right] ^ {\frac {1 + S}{2}}} \tag {4} +$$ + +and a similar expression is applicable for $p(\mathbf{b}|\mathbf{I},\mathbf{x}_q)$ . $\Gamma$ is the Gamma function. The parameters $(\boldsymbol{\mu}_a,\boldsymbol{\Sigma}_a,\boldsymbol{\mu}_b,\boldsymbol{\Sigma}_b)$ represent the location and scale matrices for the respective coordinates. During inference, the uncertainty associated + +![](images/a70127acf14c96950f4d02f1fdf61d4ad2cb1f4fc795d5e9f3b2983140ece6d1.jpg) +Figure 3. LEAP Front-end. Once image feature maps are obtained, selected anchors aid in tracking. The queries and anchors are processed by a refiner to iteratively update states. The model outputs trajectory distribution, visibility, and dynamic track label. + +with each point is quantified by the sum of diagonal scale estimates at position $(s,s)$ as $\phi (\mathbf{x}_s) = \pmb {\Sigma}_a[s,s] + \pmb {\Sigma}_b[s,s]$ + +Kernel-based Estimation. Rather than just predicting the positions of points within a trajectory, our model focuses on recovering the distribution of point trajectories by estimating its parameters. The location parameter predictions can be associated with the previous point trajectory as $[\pmb{\mu}_a; \pmb{\mu}_b] = \mathbf{X}$ . However, directly deriving a symmetric and positive definite scale matrix from model outputs is challenging. To address this, we employ a kernel-based approach with the linear kernel $K(\mathbf{x}, \mathbf{y}) = \mathbf{x}^T \mathbf{y}$ . During each iteration, two scale matrices $(\pmb{\Sigma}_a, \pmb{\Sigma}_b)$ are constructed by first applying two linear projection layers to the point features, represented as $\mathbf{F}_a^k = \mathcal{G}_a(\mathbf{F}^k)$ and $\mathbf{F}_b^k = \mathcal{G}_b(\mathbf{F}^k)$ . We then compute the scale matrices as $\pmb{\Sigma}_a = K(\mathbf{F}_a, \mathbf{F}_a) + \sigma \mathbf{I}$ , and $\pmb{\Sigma}_b = K(\mathbf{F}_b, \mathbf{F}_b) + \sigma \mathbf{I}$ , where $\sigma$ is a small positive value and $\mathbf{I}$ is the identity matrix. The model parameters are refined through Maximum Likelihood Estimation (MLE) using a negative log-likelihood (NLL) loss function. + +Loss Functions. We supervise the point trajectory mainly with the NLL loss, which is based on the predicted distribution parameters and the ground-truth point trajectory. The main point trajectory loss is given as + +$$ +\mathcal {L} _ {\text {m a i n}} = \sum_ {k} ^ {K} \gamma^ {K - k} \mathcal {L} _ {N L L} \left(\mathbf {X} ^ {k}, \mathbf {X} ^ {*}, \boldsymbol {\Sigma} _ {a} ^ {k}, \boldsymbol {\Sigma} _ {b} ^ {k}\right), \tag {5} +$$ + +where $K$ is the number of iteration and $\gamma = 0.8$ . + +For the visibility and dynamic track label supervision, we use the cross entropy loss with the estimates $\mathbf{V},\mathbf{m}_d$ and ground truth $\mathbf{V}^{*},\mathbf{m}_{d}^{*}$ as + +$$ +\mathcal {L} _ {\text {v i s}} = \left(1 - \mathbf {V} ^ {*}\right) \log \left(1 - \mathbf {V}\right) + \mathbf {V} ^ {*} \log \mathbf {V}. \tag {6} +$$ + +$$ +\mathcal {L} _ {d y n} = \left(1 - \mathbf {m} _ {d} ^ {*}\right) \log \left(1 - \mathbf {m} _ {d}\right) + \mathbf {m} _ {d} ^ {*} \log \mathbf {m} _ {d}. \tag {7} +$$ + +The total loss is a weighted sum of three losses: + +$$ +\mathcal {L} _ {\text {t o t a l}} = w _ {1} \mathcal {L} _ {\text {m a i n}} + w _ {2} \mathcal {L} _ {\text {v i s}} + w _ {3} \mathcal {L} _ {\text {d y n}}. \tag {8} +$$ + +![](images/005c323b67834729f12b46342f226d06e34bd05848cbaea2bb4db7b3f676de27.jpg) +Figure 4. LEAP-VO pipeline. Given a new image $\mathbf{I}_t$ received at time $t$ , the keypoint extractor extracts new keypoints from $\mathbf{I}_t$ . Then, all the keypoints from the latest $S_{LP}$ -frame $\mathbf{I}_{t - S_{LP} + 1:t}$ are tracked across all other frames within the current LEAP window, followed by a track filtering step to remove outliers. Finally, the local BA module is used on the current BA window to update the camera poses and 3D positions of the extracted keypoints. The colored arrows denote the moving direction of each module when a new image is received. + +![](images/bffe5d79cb06a7a1728eda2276e66086b413c42983dcc93b2df1db81da7e0b08.jpg) + +# 3.4. LEAP-VO + +Leveraging the anchor-based dynamic track estimations and temporal probabilistic formulation, we develop our Long-term Effective Point Tracking (LEAP) pipeline, boosting the accuracy of point tracking and simultaneously capturing global motion patterns and uncertainties of the tracks (refer to Figure 3). In this section, we demonstrate how we can build visual odometry for dynamic scenes based on the long-term tracking capabilities of LEAP. Our LEAP-VO pipeline is illustrated in Figure 4. + +Keypoint Extraction and Tracking. In our VO system, the LEAP model serves as the front-end. When a new image frame $\mathbf{I}_t$ is captured at time $t$ , we sample $N$ new keypoints from this image to initiate point tracking. We consistently employ the distributed image gradient-based sampling method (refer to Section 3.2) for keypoint selection due to its favorable properties in feature distinctiveness and distribution. This approach also obviates the need for additional anchors during tracking. The LEAP model then tracks these keypoints across the last $S_{LP}$ frames, performing bidirectional tracking (both forward and backward in time). Notably, $S_{LP}$ can exceed the model window $S$ , thus enabling more expansive tracking capabilities through sequential prediction [16, 18]. + +Track Filtering. Theoretically, precise camera pose estimation between two images can be attained with just a few accurate correspondences under certain conditions. However, the quality of correspondences obtained from feature matching or optical flow can vary. Traditional VO pipelines often use RANSAC [14] to filter out incorrect matches, but their efficiency decreases with more keypoints and views. Our approach, in contrast, bypasses the sampling-based method and uses trajectory quality assessments from LEAP + +for more efficient and informative track filtering. + +(i) Visibility and Dynamic Track Label: The LEAP model has the capability to estimate not only the 2D point positions within a trajectory but also the visibility of each point and the dynamic track label of the entire trajectory. Thresholds $\gamma_v$ for visibility and $\gamma_d$ for dynamic track labels are set to ensure that only visible and static points are utilized in the bundle adjustment process. +(ii) Track Uncertainty: In Section 3.3, we introduce probabilistic modeling to LEAP, allowing us to gauge confidence via distribution measurements. We select only high-confidence points, which correspond to low uncertainty measurements. Denoting our model's uncertainty estimation as $\Phi$ , we retain points that exhibit high confidence using the criterion $\Phi \leq Q(\gamma_u)$ , where $\gamma_u$ is the uncertainty quantile and $Q:[0,1] \to \mathbb{R}$ is the quantile function. +(iii) Track Length: After filtering, we evaluate and remove tracks with insufficient observations, as bundle adjustment is more reliable with increased observations and wider baselines. Tracks with fewer than $\gamma_{\text{track}}$ valid points are excluded from optimization cost computation due to their potential unreliability. + +Sliding-window Optimization. With the point validation mask established, we proceed to define the optimization cost function for our sliding window bundle adjustment. The window size for local bundle adjustment, denoted as $S_{BA}$ , may differ from the point tracking window size $S_{LP}$ . Our geometric bundle adjustment aims to optimize camera poses $\mathbf{T}$ and 3D scene point positions $\mathbf{Q}$ by aligning the induced point trajectory from the projective relationship, with the estimated point trajectory from LEAP. We parameterize a 3D scene point $\mathbf{Q}_i$ by its 2D keypoint location $\mathbf{x}_i$ in image $\mathbf{I}_i$ and depth $d_i$ . Let $\mathrm{LEAP}_{i\rightarrow j}$ denote the mapping of tracking keypoint from $\mathbf{I}_i$ to $\mathbf{I}_j$ using LEAP model. The + +reprojection cost is formulated as: + +$$ +\begin{array}{l} \sum_ {i} \sum_ {j \in | i - j | \leq S _ {B A}} \sum_ {n} w _ {i \rightarrow j, n} \| \mathcal {P} (\mathbf {T} _ {i}, \mathbf {T} _ {j}, \mathbf {K}, d _ {i, n}) - \\ \operatorname {L E A P} _ {i \rightarrow j} (\mathbf {x} _ {i, n}) \| _ {\rho}, \tag {9} \\ \end{array} +$$ + +where $\| \cdot \|_{\rho}$ is the distance metric, and the weight $w_{i\rightarrow j,n}$ is derived from the track filtering results. We use the GaussNewton method [34] to optimize Eq. (9) for $K_{BA}$ iterations. At each iteration, we compute the camera poses update $\Delta \xi^{(k)}\in \mathfrak{se}(3)$ (lie-algebra corresponding to $\mathbf{T}$ ), and the depth update $\Delta \mathbf{D}^{(k)}$ for each point. The optimization can be solved efficiently with the Schur decomposition. We use the robust Huber loss function for the distance metric. + +# 4. Experiment + +# 4.1. Datasets and Metrics + +Replica. Replica [29] provides synthetic indoor environments for simulating the image-capturing process with a moving camera. We use the camera trajectories provided by [50] for the localization task, resulting in 16 camera trajectories derived from 8 static scenes. Each sequence comprises 900 frames with rendered RGB images. We use "Sequence 1" for the visual odometry evaluation. + +MPI Sintel. The MPI Sintel [4] dataset contains dynamic image sequences from open-source 3D animated short films. These sequences consist of large and complex object motion, along with other imaging effects such as motion blur and defocus blur. Following prior works [20, 48], we assess dynamic VO performance on the MPI Sintel dataset, with each sequence containing 20 to 50 image frames. + +TartanAir Shibuya. The TartanAir Shibuya dataset [26] features dynamic scenes from two scenarios: "Standing Humans" and "Road Crossing". The majority of the humans remain stationary in "Standing Humans" sequences, while "Road Crossing" sequences provide a more challenging setting where multiple humans are moving in different directions. Each sequence in the dataset includes 100 frames with more than 30 tracked moving humans. + +Metrics. For visual odometry, we compare the Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot) following [33, 47, 48]. ATE measures how much the estimated trajectory deviated from the ground truth trajectory, which is calculated as the root mean square error (RMSE) overall all pose transformations after scaling and alignment. "RPE trans" computes translation errors made over a certain distance (meter), and "RPE rot" computes the estimated rotation over a certain distance (degree). Both "RPE trans" and "RPE rot" are calculated on all poses and then averaged. + +
MethodReplica
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]0.0860.0300.650
DynaSLAM [1]0.0390.0170.366
DROID-SLAM [32]0.2670.0362.631
TartanVO [40]0.4060.0362.063
DytanVO [28]0.2890.0352.146
DPVO [33]0.2570.0362.635
LEAP-VO (Ours)0.2040.0301.992
+ +Table 1. Camera tracking results on Replica [29]. The statistics demonstrate that our method achieves better results than other competitive baselines. + +
MethodMPI Sintel
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]XXX
DynaSLAM [1]XXX
DROID-SLAM [32]0.1750.0841.912
TartanVO [40]0.2380.0931.305
DytanVO [28]0.1310.0971.538
DPVO [33]0.0760.0781.722
LEAP-VO (Ours)0.0370.0551.263
+ +Table 2. Camera tracking results on MPI Sintel [4]. The statistics show that our method outperforms other competitive baselines. 'X' denotes the tracking failure case. + +# 4.2. Implementation details + +We employ CoTracker [18], a recent PIPs variant, as our base TAP model. We train our model on the TAP-Vid-Kubric [9] training set, following [18]. We load the pretrained CoTracker weights for the image feature extractor and train LEAP for 100,000 steps using 4 NVIDIA A100 GPUs in parallel. During training, we use $K = 4$ steps for iterative refinements, $N = 256$ for the number of queries, $S = 8$ for the model window, and $S_{LP} = 12$ for the tracking window. For inter-track anchor-based attention, the number of anchors is 64. The loss weights are set to $w_{1} = 1.0, w_{2} = 0.5, w_{3} = 0.5$ . In LEAP-VO, we perform 4 steps for each bundle adjustment with the Huber Loss for distance metric. The bundle adjustment window size $S_{BA}$ is set to 15. Track filtering parameters are set at $\gamma_v = 0.9, \gamma_d = 0.9, \gamma_u = 0.8, \gamma_{track} = 3$ . Please refer to our supplementary materials for more details. + +# 4.3. Visual Odometry Results + +We compare our methods against the state-of-the-art SLAM and VO approaches, including ORB-SLAM2 [24], DynaSLAM [1], DROID-SLAM [32], TartanVO [40], DytanVO [28], and DPVO [33]. Evaluations are conducted on the static Replica dataset [29] as well as on the dynamic datasets MPI-Sintel [4] and TartanAir-Shibuya [26]. + +Replica. We compare our method with other baselines on the Replica dataset, which contains eight static scenes + +
MethodStandingHumanRoadCrossing (Easy)RoadCrossing (Hard)Average
01020304050607
ORB-SLAM2 [24] w/ mask0.07880.00600.06570.01960.01481.09840.84760.3044
DynaSLAM [1]X0.88360.39070.41960.49250.64460.6539(0.5808)
AirDOS [26] w/ mask0.06060.01930.09510.03310.02060.22300.56250.1449
DROID-SLAM [32]0.00510.00730.01030.01200.22780.02530.57880.1238
TartanVO [40]0.06000.16050.27620.18140.21740.32280.50090.2456
DytanVO [28]0.03270.10170.06080.05160.07550.03650.06600.0607
DPVO [39]0.05390.16030.17880.28580.11960.11190.14980.1514
LEAP-VO (Ours)0.00810.01850.03990.03170.01130.06890.02460.0290
+ +Table 3. ATE (m) results on TartanAir-Shibuya Sequences [26]. ( $\cdot$ ) denotes averaging on valid scenes only. We can observe that our method shows better results for ATE on most sequences, outperforming all SLAM and VO baselines on average ATE. + +![](images/0c9be0f3fcfc59cefc63ad77edac4adc0e039eefe5b343e437c31e099d237ab9.jpg) +(a) RoadCrossing 04 + +![](images/f5a4197b9fd43580001e00f745e1a2cbf7cc380b0001405e7443351579a9253f.jpg) +(b) RoadCrossing 05 + +![](images/422bd73a1d36da5ecdab37f323089a41085702b01500552c15b23bbc9e085202.jpg) +(c) RoadCrossing 06 +Figure 5. Qualitative results of camera trajectory estimation on TartanAir-Shibuya [26]. The visualizations show that our method provides more robust and accurate camera pose trajectories, especially in hard cases (RoadCrossing 06 and RoadCrossing 07). + +![](images/500f27a167516833358616cab081a003fdad56f68b6e256912b551c566d37440.jpg) +(d) RoadCrossing 07 + +without any moving objects. Challenges arise in two test sequences that contain prolonged occlusions, leading to the potential camera tracking loss. As illustrated in Table 1, our approach outperforms other VO methods on all three metrics, showing effectiveness under static environments and the capability to handle occlusions. Further insights are available in Section 4.5. For classical methods, ORBSLAM2 [24] and DynaSLAM [1], we observe tracking loss and resuming via re-localization, which reduces long-term drift and shows high accuracy for camera pose tracking. + +MPI-Sintel. We also evaluate our VO camera trajectory accuracy on 14 dynamic scenes from the MPI Sintel dataset, following the setting of [48]. As shown in Table 2, our method substantially surpasses other VO baselines in dynamic scenes, thanks to our temporal uncertainty estimation and dynamic track detection. Traditional feature-based methods like ORB-SLAM2 [24] and DynaSLAM [1] sometimes fail to track camera movements in highly dynamic sequences, primarily due to outliers from dynamic object correspondences and segmentation-based only dynamic track estimation. In contrast, our approach, which considers visual, temporal, and inter-track information, demonstrates notable improvements in dynamic scene tracking accuracy. + +TartanAir-Shibuya. Following DytanVO [28], we further evaluate the VO performance using the dynamic TartanAir-Shibuya dataset. We assess the Average Trajectory Error (ATE) across seven diverse scenes, ranging from simple to + +challenging. As depicted in Table 3, our method consistently outperforms all other SLAM and VO baselines in the majority of these scenes. Moreover, trajectory visualizations in Figure 5 also highlight the robust camera tracking capabilities of our method in dynamic environments. + +# 4.4. Dynamic Track Estimation Results + +In our LEAP formulation, we have introduced the anchor-based dynamic track estimation module to distinguish between static and dynamic point trajectories, which provides valuable information for handling dynamic scenes in visual odometry. Our method leverages both point feature information and inter-track relations to predict the dynamic track label for each trajectory. To demonstrate the performance of the proposed dynamic track estimation module, we take the LEAP model and evaluate it on novel image sequences from diverse datasets. For a comprehensive demonstration, we uniformly sample $16 \times 16$ queries among the images and track them for $T = 8$ frames. The results, shown in Figure 6, illustrate that LEAP can effectively handle various scenarios, including single-object motion, complex and rapid motion, and multi-object motion with occlusion. + +# 4.5. Ablation Study + +Track Filtering. To highlight the efficacy of our track filtering module in LEAP-VO, we conduct an ablation study on different masks used in track filtering. We study three types of filtering criteria: visibility, dynamic + +![](images/781af32895b8c1318275094eab868761f2d0179d29830e8800be855692ef374b.jpg) +Figure 6. Visualization of dynamic track estimation on DAVIS [19], MPI-Sintel [4], and TartanAir-Shibuya [26]. Odd columns: all point trajectories. Even columns: estimated dynamic point trajectories. + +![](images/849ddeb126a9119bf85089fee5033115a331c32107fbd180edaeec2c0c8ce3b9.jpg) +Figure 7. Visualization of point-wise uncertainty measurements. Keypoints with the lowest $20\%$ (left) and highest $20\%$ (right) uncertainty are shown in green and yellow, respectively. + +track, and confidence filtering. Table 4 shows the impact of applying various filtering types on the dynamic MPI Sintel benchmark. Without dynamic filtering or confidence filtering, the ATE increases by $46\%$ and $130\%$ , respectively. Combining all three types of filtering methods altogether, we achieve the best results in all three metrics. + +
MethodTrack FilteringMPI Sintel
Vis.Dyn.Conf.ATERPE transRPE rot
LEAP-VO✓--0.1180.0781.340
✓✓-0.0850.0621.281
✓-✓0.0540.0581.274
✓✓✓0.0370.0551.263
+ +Uncertainty Estimation. As illustrated in Figure 7, we demonstrate the LEAP's capability to provide meaningful per-point uncertainty. It models the uncertainty as distribution parameters for both the entire track and individual points. Two key observations can be found from Figure 7. Firstly, our model reliably identifies high-uncertainty + +points, typically located in low-texture areas or in regions with repetitive patterns. Secondly, although not explicitly supervised, the model tends to assign higher uncertainty to dynamic objects, which are inherently more challenging to track. These findings highlight the efficacy of LEAP's probabilistic formulation in handling uncertainty. + +Keypoint Extraction. Table 5 showcases the effectiveness of various keypoint extraction strategies for VO performance. Despite its distinctive features, SIFT extraction provides inferior performance due to its unpredictable keypoint distribution. In contrast, the random sampling tends to generate well-separated keypoints, leading to a performance similar to SIFT's. Our distributed image gradient-based sampling shows the best performance, underscoring the significance of feature distribution and distinctiveness. + +Table 4. Effect of the proposed track filtering method on MPI-Sintel [4]. Tracking using all criteria yields the best results. + +
Policy:UniformRandomSIFT [22]Ours
ATE (m):0.15110.03660.03550.0290
+ +Baseline comparison. We evaluate the pretrained CoTracker [18] for the end VO performance. Table 6 shows that LEAP front-end outperforms CoTracker on all datasets, demonstrating the effectiveness of the proposed dynamic track estimation and temporal probability modeling. + +Table 5. Comparison of keypoint extraction policies ( $N_{a} = 64$ ) on TartanAir-Shibuya [26]. Our distributed image-gradient sampling method achieves better results compared to other policies. + +
Front-endReplicaMPI SintelShibuya
ATERPE transRPE rotATERPE transRPE rotATE
CoTracker0.3010.0312.0430.1260.0711.3660.160
LEAP0.2040.0301.9920.0370.0551.2630.029
+ +Table 6. Baseline comparison with CoTracker front-end [18]. Our LEAP front-end achieves better results on all datasets. + +# 5. Conclusion + +We introduce Long-term Effective Any Point Tracking (LEAP), designed to tackle the limitations of previous point-tracking methods. By leveraging visual, inter-track, and temporal cues within an anchor-based dynamic track estimation module, LEAP captures global motion patterns and effectively discriminates between static and dynamic elements. Moreover, our approach incorporates a temporal probabilistic formulation into the point-tracking pipeline, emulating recursive Bayesian filtering through a learning-based refinement module. This enables our model to accurately assess the reliability of point trajectory estimates. Using LEAP as the front-end, we then develop the LEAP-VO system, featuring novel integration of visual information, global motion patterns, and track distribution for handling dynamic scenes. Additionally, our takes for LEAP-VO can be integrated with other TAP-based front-ends, offering the potential to markedly improve camera-tracking accuracy and robustness. + +# References + +[1] Berta Bescos, José M Fácil, Javier Civera, and José Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, 3(4):4076-4083, 2018. 2, 6, 7 +[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 2023. 2 +[3] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In CVPR, pages 2560–2568, 2018. 2 +[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625. Springer-Verlag, 2012. 6, 8 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 2 +[6] Weirong Chen, Suryansh Kumar, and Fisher Yu. Uncertainty-driven dense two-view structure from motion. IEEE Robotics and Automation Letters, 8(3):1763-1770, 2023. 4 +[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In ICCV, pages 7063-7072, 2019. 2 +[8] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE TPAMI, 29(6):1052-1067, 2007. 2 +[9] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 35:13610-13626, 2022. 2, 6 +[10] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. ICCV, 2023. 2, 3 +[11] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 2 +[12] Jakob Engel, Thomas Schöps, and Daniel Cremers. Ldslam: Large-scale direct monocular slam. In ECCV, pages 834-849. Springer, 2014. 2 +[13] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE TPAMI, 40(3):611-625, 2017. 2 +[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5 +[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 1 + +[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, pages 59-75. Springer, 2022. 1, 2, 3, 5 +[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 33: 19545-19560, 2020. 2 +[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2, 3, 5, 6, 8 +[19] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2018. 8 +[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2, 3, 6 +[21] Shunkai Li, Xin Wang, Yingdian Cao, Fei Xue, Zike Yan, and Hongbin Zha. Self-supervised deep visual odometry with online adaptation. In CVPR, pages 6339-6348, 2020. 2 +[22] David G Lowe. Object recognition from local scale-invariant features. In ICCV, pages 1150-1157. IEEE, 1999. 8 +[23] Annette Mossel and Manuel Kroeter. Streaming and exploration of dynamically changing dense 3d reconstructions in immersive virtual reality. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR-Adjunct), pages 43–48. IEEE, 2016. 1 +[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: an open-source SLAM system for monocular, stereo and RGB-D cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 6, 7 +[25] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 2 +[26] Yuheng Qiu, Chen Wang, Wenshan Wang, Mina Henein, and Sebastian Scherer. Airdos: Dynamic slam benefits from articulated objects. In ICRA, pages 8047-8053. IEEE, 2022. 6, 7, 8 +[27] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In CVPR, pages 12240-12249, 2019. 2 +[28] Shihao Shen, Yilin Cai, Wenshan Wang, and Sebastian Scherer. Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environments. In ICRA, pages 4048-4055. IEEE, 2023. 2, 6, 7 +[29] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6 +[30] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In ICLR, 2019. 2 + +[31] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 2 +[32] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34: 16558-16569, 2021. 2, 6, 7 +[33] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 6 +[34] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In International workshop on vision algorithms, pages 298-372. Springer, 1999. 6 +[35] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. Neuraliff: Segmenting 3d objects that move in egocentric videos. In 3DV, pages 910-919. IEEE, 2021. 2 +[36] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2 +[37] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 2 +[38] Rui Wang, Martin Schworer, and Daniel Cremers. Stereo dso: Large-scale direct sparse visual odometry with stereo cameras. In ICCV, pages 3903-3911, 2017. 2 +[39] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In ICRA, pages 2043-2050. IEEE, 2017. 2, 7 +[40] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In CoRL, pages 1761-1772. PMLR, 2021. 2, 6, 7 +[41] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 35:32653-32666, 2022. 2 +[42] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2 +[43] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In CVPR, pages 1281-1292, 2020. 2 +[44] Weicai Ye, Xinyue Lan, Shuo Chen, Yuhang Ming, Xingyuan Yu, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pvo: Panoptic visual odometry. In CVPR, pages 9579-9589, 2023. 2 +[45] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018. 2 +[46] Khalid Yousif, Alireza Bab-Hadiashar, and Reza Hosein- nezhad. An overview to visual odometry and visual slam: Applications to mobile robotics. Intelligent Industrial Systems, 1(4):289-311, 2015. 1 + +[47] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 2, 6 +[48] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In ECCV, pages 523-542. Springer, 2022. 2, 3, 6, 7 +[49] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, pages 19855-19865, 2023. 2 +[50] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In CVPR, pages 15838-15847, 2021. 6 +[51] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851–1858, 2017. 2 \ No newline at end of file diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/images.zip b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fedf4fdf8f2fcade56e16e5509b34719fadbfe5e --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b0dd5b1a4e1a6851d0208481080b1ade2deafc5f9aa0a809fc132b0b1d6cadb +size 566347 diff --git a/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/layout.json b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d0f3dff3f030d611c08cea602a487a98bd3cdd68 --- /dev/null +++ b/2024/LEAP-VO_ Long-term Effective Any Point Tracking for Visual Odometry/layout.json @@ -0,0 +1,9375 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 523, + 121 + ], + "type": "text", + "content": "LEAP-VO: Long-term Effective Any Point Tracking for Visual Odometry" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "spans": [ + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "text", + "content": "Weirong Chen" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "text", + "content": " Le Chen" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "text", + "content": " Rui Wang" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "text", + "content": " Marc Pollefeys" + }, + { + "bbox": [ + 144, + 142, + 447, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "spans": [ + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "text", + "content": "TU Munich " + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "text", + "content": "Munich Center for Machine Learning " + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "text", + "content": "MPI for Intelligent Systems " + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 159, + 527, + 175 + ], + "type": "text", + "content": "Microsoft" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 231, + 290, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 231, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 231, + 290, + 483 + ], + "type": "text", + "content": "Visual odometry estimates the motion of a moving camera based on visual input. Existing methods, mostly focusing on two-view point tracking, often ignore the rich temporal context in the image sequence, thereby overlooking the global motion patterns and providing no assessment of the full trajectory reliability. These shortcomings hinder performance in scenarios with occlusion, dynamic objects, and low-texture areas. To address these challenges, we present the Long-term Effective Any Point Tracking (LEAP) module. LEAP innovatively combines visual, inter-track, and temporal cues with mindfully selected anchors for dynamic track estimation. Moreover, LEAP's temporal probabilistic formulation integrates distribution updates into a learnable iterative refinement module to reason about point-wise uncertainty. Based on these traits, we develop LEAP-VO, a robust visual odometry system adept at handling occlusions and dynamic scenes. Our mindful integration showcases a novel practice by employing long-term point tracking as the front-end. Extensive experiments demonstrate that the proposed pipeline significantly outperforms existing baselines across various visual odometry benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 635 + ], + "type": "text", + "content": "Visual odometry (VO) calculates the camera's position and movement from an image sequence. It is widely used in various fields including robotics [46], mixed reality [23], and autonomous driving [15]. The performance of VO highly depends on the accuracy of its point tracking front-end, which recovers relative camera motion between consecutive frames using projective geometry. It further requires the point correspondences to be static, as dynamic points can lead to inaccurate motion estimation and pose recovery." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 287, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 684 + ], + "type": "text", + "content": "Despite significant advances in feature tracking and optical flow estimation, most existing VO methods mainly focus on the two-view case, i.e., computing feature correspondences between a pair of images. However, these methods" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 207, + 423, + 336 + ], + "blocks": [ + { + "bbox": [ + 313, + 207, + 423, + 336 + ], + "lines": [ + { + "bbox": [ + 313, + 207, + 423, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 423, + 336 + ], + "type": "image", + "image_path": "7a76ff2249955c9c8077c455b55618895587c7dd0639e49aad6be7800470b4de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 428, + 207, + 539, + 335 + ], + "blocks": [ + { + "bbox": [ + 428, + 207, + 539, + 335 + ], + "lines": [ + { + "bbox": [ + 428, + 207, + 539, + 335 + ], + "spans": [ + { + "bbox": [ + 428, + 207, + 539, + 335 + ], + "type": "image", + "image_path": "070ad67c83554f8e25cf7cd79e091da3bce520463dff2693730f48a768a657d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 391, + 343, + 524, + 365 + ], + "blocks": [ + { + "bbox": [ + 328, + 347, + 388, + 357 + ], + "lines": [ + { + "bbox": [ + 328, + 347, + 388, + 357 + ], + "spans": [ + { + "bbox": [ + 328, + 347, + 388, + 357 + ], + "type": "text", + "content": "Visual Odometry" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 343, + 524, + 365 + ], + "lines": [ + { + "bbox": [ + 391, + 343, + 524, + 365 + ], + "spans": [ + { + "bbox": [ + 391, + 343, + 524, + 365 + ], + "type": "image", + "image_path": "faed9191be70cfa06e13b268052bd065dc87863b05880998ec81b4c6d869c1b0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 367, + 546, + 444 + ], + "lines": [ + { + "bbox": [ + 304, + 367, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 546, + 444 + ], + "type": "text", + "content": "Figure 1. Comparison between two-view approach and multi-view approach for visual odometry. (a) In the two-view approach, correspondences are derived for every image pair and concatenated. Managing occlusion becomes challenging without the temporal context. (b) In the multi-view approach, long-term point trajectories can be obtained all at once, enabling the detection of occlusion and tracking points under partial occlusions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 616 + ], + "type": "text", + "content": "are limited as they rely solely on pairwise relative motion, thus neglecting the valuable temporal information in image sequences. Consequently, they often struggle with capturing dynamic motion patterns in tracks. Another challenge arises from the need to handle occlusions. In two-view scenarios, occlusion is not explicitly modeled, leading to misidentification of occluded points as incorrect matches. To address these challenges effectively, we propose to introduce a long-term point tracker, which can recover the trajectories of specific points across the image sequence given a sparse set of these query points. Its capability of leveraging temporal context and reliably tracking queries over multiple frames even under occlusion, makes it highly suitable for VO applications (refer to Figure 1)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Recently, Persistent Independent Particles (PIPs) [16] proposes a novel pipeline that brings the idea of iterative refinement into the long-term point tracking task, with the local correlation evidence to maintain feasible computation. While it is proficient in leveraging the temporal context present in the image sequence, it ignores the inter-track information in the image sequence and processes each query point independently. Consequently, PIPs struggles to cap" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Work done at Microsoft Mixed Reality & AI Lab Zurich for a Master's degree." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19844" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "ture the global motion patterns from the image sequence. Additionally, it lacks a reliability metric for its trajectory estimates, hindering its application in quality-sensitive tasks like video segmentation and visual odometry. To mitigate the aforementioned issues, we propose a pipeline that explores dynamic track estimation and temporal probabilistic formulation for long-term point tracking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 288, + 384 + ], + "type": "text", + "content": "In this paper, we present Long-term Effective Point Tracking (LEAP), aiming to estimate the reliability of predicted correspondences and track moving object trajectories in dynamic scenes. To achieve this, we harness the global motion patterns—often overlooked in existing methods—through an anchor-based dynamic track estimation module, leveraging visual, temporal, and inter-track information to differentiate between the static and dynamic elements. We also incorporate a temporal probabilistic approach into our LEAP pipeline to handle uncertainties and refine the distribution of point correspondences iteratively, showcasing superior performance in complex real-world situations. More importantly, we pioneer the usage of continuous motion estimation from long-term point tracking to construct a robust visual odometry system. Our LEAP-VO distinguishes itself through its thoughtful integration of visual features, track distribution, and global motion patterns, significantly enhancing both the performance and robustness of visual odometry in dynamic environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 392, + 134, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 134, + 404 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 134, + 404 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 412, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 653 + ], + "type": "text", + "content": "Long-term Point Tracking. In the realm of long-term point tracking or Tracking Any Point (TAP), various methods have been introduced recently. PIPs [16] reframes pixel tracking as a long-term motion estimation with local information and improves tracking accuracy by associating each pixel with a trajectory that spans multiple image frames. The subsequent works enhance performance by integrating spatial context features [2], tracking multiple queries concurrently [18], and extending temporal fields and templates [49]. TAP-Net [9] adopts a different method by locating candidate correspondences through a global cost volume, improved by TAPIR's trajectory refinement [10] and BootsTAP's semi-supervised training [11]. Other approaches explore self-supervision targets to learn visual correspondence via powerful feature representation [5, 17, 36]. Recently, OmniMotion [37] provides a pure test-time optimization method that infers the point trajectory and occlusion directly from videos with optical flow predictions. However, its slow inference speed makes it unsuitable for real-time applications such as visual odometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Monocular Visual Odometry. Visual Odometry (VO) can be categorized into indirect (feature-based) and direct methods. Indirect methods, such as MonoSLAM [8] and ORB-SLAM [25], mostly rely on feature extraction and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 324 + ], + "type": "text", + "content": "matching to estimate the camera pose. Direct methods, conversely, skip feature extraction and matching and utilize intensity to optimize camera pose and 3D scene point positions based on photometric errors [12, 13, 38], showing better performance on low-texture regions where keypoint tracking can easily fail. However, they are more vulnerable to illumination changes and issues with rolling shutter cameras. Advancements in deep learning have spawned end-to-end monocular VO methods in supervised [3, 28, 30, 33, 39, 40, 43, 44] and unsupervised [21, 27, 45, 51] settings. DeepVO [39] utilizes Recurrent Neural Networks to model sequential data, while SfMLearner [51] develops an unsupervised framework to learn the depth and camera motion from unlabeled monocular videos. TartanVO [40] proposes a VO model that generalizes to multiple datasets and real-world scenarios. DROID-SLAM [32] follows learning-based optimization from RAFT [31] and proposes an end-to-end system, integrating flow, confidence, and geometric optimization via iterative GRU updates. The followed-up work DPVO [33] further cuts computational complexity by replacing dense feature tracking with sparse patch tracking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 337, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 545, + 553 + ], + "type": "text", + "content": "Dynamic Track Estimation. In addition to enhancing accuracy, numerous studies have focused on the detection of dynamic objects within videos. Methods based on segmentation, such as DynaSLAM [1] and RCVD [20], employ semantic cues from pre-trained models to exclude dynamic regions. However, these segmentation-based methods are constrained to predefined object categories and struggle to capture real-world motion (for example, differentiating between a stationary and a moving car). Alternatively, DyanVO [28] and ParticleSfM [48] adopt a trajectory-based motion detection approach. This method addresses dynamic scenes using estimated point track information, thereby generating reliable camera trajectories in complex motion scenarios. Nonetheless, they depend on dense optical flow estimation to capture global motion patterns. Furthermore, dynamic tracks can be detected using additional depth information for a 3D consistency check [7, 47], or by optimization based on photo consistency [35, 41, 42]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 362, + 574 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 665 + ], + "type": "text", + "content": "Our goal is to tackle the aforementioned challenges in VO systems, particularly in handling dynamic environments, temporal occlusions, and low-texture regions. To this end, we introduce our LEAP module, which incorporates dynamic track estimation and temporal probabilistic formulation. Furthermore, we illustrate how this module serves as the cornerstone for constructing our advanced VO system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 671, + 485, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 485, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 485, + 684 + ], + "type": "text", + "content": "3.1. Preliminary: Tracking Any Point" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "Tracking arbitrary pixels in an image sequence is traditionally akin to estimating optical flow, where each pixel" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19845" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "is mapped to a corresponding location in a subsequent frame. However, two-view approaches often overlook the rich temporal context within the image sequence, leading to suboptimal outcomes. Recent advancements address this by revisiting the particle video concept and improving long-term point tracking through learning-based techniques [10, 16, 18]. These TAP methods can track points over multiple frames and detect occlusions, significantly boosting tracking robustness under challenging scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": "**Formulation.** Given a sequence of " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " consecutive RGB images " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = [\\mathbf{I}_1, \\dots, \\mathbf{I}_S]" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " as input, the goal of TAP is to track a set of query points across these frames. For a specific query point with the 2D pixel coordinate " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_q \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "s_q" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ", TAP predicts its trajectory across all " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " images as " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{x}_1, \\dots, \\mathbf{x}_S]" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ", given " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{s_q} = \\mathbf{x}_q" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ". Additionally, TAP estimates the visibility of each point throughout the sequence as " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = [v_1, \\dots, v_S]" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "v_s \\in \\{0, 1\\}" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " is a binary label of point visibility indicating whether a point in frame " + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 192, + 289, + 324 + ], + "type": "text", + "content": " is visible or occluded. The TAP formulation for a single query can be described as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 113, + 332, + 287, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 332, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 113, + 332, + 287, + 347 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {X}, \\mathbf {V}\\right) = \\operatorname {T A P} \\left(\\mathbf {I}, \\mathbf {x} _ {q}, s _ {q}\\right). \\tag {1}", + "image_path": "c35d8f0751c2903beae4b2572ae63ac32a9cb2b795182907b84406b9acb379f2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": "PIPs [16], a notable learning-based TAP method, offers a promising solution featuring local cost volumes and iterative refinement. Initially, each image " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s\\in \\mathbf{I}" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": " is processed through a CNN feature extractor " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": " to extract the feature map " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{\\mathbf{s}} = \\mathcal{F}(\\mathbf{I}_{s})" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ". Point features " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_q" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": " are computed by bilinear sampling at the query positions " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{x_s}" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_s" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ". These point features from each image are then concatenated to form a point feature tensor " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = [\\mathbf{f}_1,\\dots ,\\mathbf{f}_S]" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ". PIPs approaches the TAP problem by iteratively updating the state variables " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "(\\mathbf{X},\\mathbf{F})" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ", with the initial states obtained by duplicating the query positions and features. During the " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": "-th iterative refinement, PIPs computes a multi-scale local cost volume " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{C}[\\mathbf{X}^k ]" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": " centered around the current estimated point position " + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^k" + }, + { + "bbox": [ + 46, + 354, + 289, + 534 + ], + "type": "text", + "content": ". Utilizing this local evidence, PIPs predicts updates for the state variables via" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\left(\\Delta \\mathbf {X}, \\Delta \\mathbf {F}\\right) = \\operatorname {R e f i n e r} \\left(\\mathbf {F} ^ {k}, \\operatorname {p o s} \\left(\\mathbf {X} ^ {k} - \\mathbf {x} _ {\\mathbf {q}}\\right), \\mathbf {C} ^ {k} \\left[ \\mathbf {X} ^ {k} \\right]\\right), \\\\ \\mathbf {X} ^ {k + 1} \\leftarrow \\mathbf {X} ^ {k} + \\Delta \\mathbf {X}, \\quad \\mathbf {F} ^ {k + 1} \\leftarrow \\mathbf {F} ^ {k} + \\Delta \\mathbf {F}, \\end{array} \\tag {2}", + "image_path": "829cbcc49464c9328d038c7575a6fd662d9ea38f595823c3454893cc5c768111.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{pos}(\\cdot)" + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "content": " refers to the positional embedding. The final point feature " + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^K" + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "content": " is used to predict point visibility with a simple linear projection layer " + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_v" + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "content": ", yielding " + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathcal{G}_v(\\mathbf{F}^K)" + }, + { + "bbox": [ + 46, + 579, + 287, + 653 + ], + "type": "text", + "content": ". Recent advances extend PIPs by replacing the MLP-based refiner with an attention-based refiner and adding spatial feature aggregation to track multiple queries together [18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 659, + 264, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 264, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 264, + 672 + ], + "type": "text", + "content": "3.2. Anchor-based Dynamic Track Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "The established TAP model enables us to extract long-term point trajectories and thereby recover camera motion from an image sequence. However, dynamic environments" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 543, + 140 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 543, + 140 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 543, + 140 + ], + "type": "image", + "image_path": "f9a4b8f0c4199ea8b8569721db074c6591697e11ec43d5bbfbe71c1e30542119.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "text", + "content": "Figure 2. Distributed image gradient-based sampling with " + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "inline_equation", + "content": "N_{a} = 64" + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "text", + "content": ". After computing the image gradient and pooling, we split the gradient map into " + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 305, + 142, + 545, + 186 + ], + "type": "text", + "content": " grids and select the point with the maximum gradient in each grid." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 194, + 545, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 242 + ], + "type": "text", + "content": "pose extra challenges for VO systems that rely on static 3D point correspondences linked to unchanging scene locations. Therefore, integrating dynamic track estimation into the VO system is crucial for real-world applications." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 546, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 387 + ], + "type": "text", + "content": "We propose an innovative method for dynamic track estimation that leverages visual, temporal, and inter-track information. While common techniques like monocular segmentation [20] excel at identifying moving objects using visual cues but can erroneously label static objects, like parked cars, as dynamic. Trajectory-based methods [48], utilizing optical flows from two or multiple frames, can differentiate static and moving objects apart but tend to ignore visual information and can be inefficient due to dense trajectories computation. Merging these two distinct methods is nontrivial; however, the formulation of the TAP pipeline, enhanced with inter-track attention, offers a viable solution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": "Our method smartly integrates visual appearance, long-term motion, and inter-track cues, combining the strengths of segmentation and trajectory-based techniques. Utilizing TAP, we extract point features " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " and long-term trajectories " + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 388, + 547, + 533 + ], + "type": "text", + "content": " from the video input. These features capture visual information and the iterative refinement module exchanges information across tracks over time for accurate labeling. However, such a method depends on the query distribution and quantities. With only one query during inference time, it cannot utilize inter-track information. Likewise, if the queries focus on a single object, the system may overlook the motions of other objects due to similar motion patterns." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": "Anchor-based Inter-track Attention. To better capture global motion patterns, we introduce additional anchors alongside the given queries, denoted as " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "|\\mathcal{A}| = N_{a}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ". Ideally, the anchor set should exhibit two characteristics: (1) the anchors should be easy to track, and (2) they should be well-distributed to encapsulate global motion patterns. Given the image " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{s_q}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " from which the queries " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_q" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " are extracted (we maintain the same notations for both single and multiple queries for simplicity), we first employ the Sobel Kernel to derive the image gradient " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{G}_x,\\mathbf{G}_y)" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " for image " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{s_q}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": ". We then apply an average pooling layer on gradient maps for smoothing and downscaling feature maps. Subsequently, we divide the gradient map into " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "k\\times k" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " sub-regions and select top " + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\frac{N_a}{k^2}" + }, + { + "bbox": [ + 304, + 546, + 547, + 715 + ], + "type": "text", + "content": " pixels with the highest image gradient magnitude in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19846" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 276, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 276, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 276, + 84 + ], + "type": "text", + "content": "each local grid. The process is demonstrated in Figure 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": "To predict the dynamic label of a trajectory, we track queries with additional anchors and concatenate the original query features " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " and query trajectories " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " with anchor features " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_A" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " and anchor trajectories " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_A" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": ". We apply a shallow MLP layer " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_d" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " with average pooling over the temporal dimension. The dynamic track label " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_d" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " is estimated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 163, + 287, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 163, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 77, + 163, + 287, + 178 + ], + "type": "interline_equation", + "content": "\\mathbf {m} _ {d} = \\operatorname {a v g p o o l} \\left(\\mathcal {G} _ {d} \\left(\\left[ \\mathbf {X} ^ {K}; \\mathbf {X} _ {A} ^ {K} \\right], \\left[ \\mathbf {F} ^ {K}; \\mathbf {F} _ {A} ^ {K} \\right]\\right)\\right). \\tag {3}", + "image_path": "68e1a085c2cfcc1150abcb3086ffed95bf211bb8659cfa39671c0044c7784c1c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 185, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 269 + ], + "type": "text", + "content": "Since newly added anchors often lack ground truth information, we employ a semi-supervised training scheme that tracks queries and anchors together to leverage inter-track attention but only calculates losses based on query predictions. This approach prevents the network from developing bias towards specific query sampling methods while learning dynamic track labels effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 276, + 218, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 276, + 218, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 276, + 218, + 289 + ], + "type": "text", + "content": "3.3. Temporal Probability Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 294, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 402 + ], + "type": "text", + "content": "Supervised learning applied to large-scale datasets has yielded notable advancements in learning-based point tracking methods. Nonetheless, trajectory estimation errors remain a concern, especially under challenging conditions like occlusions, motion blur, or textureless regions. Recognizing the varying reliability of trajectories, integrating an uncertainty estimation mechanism into the correspondence estimation process is beneficial [6]. This is particularly crucial for quality-sensitive tasks such as visual odometry." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": "To evaluate the reliability of point correspondences, we incorporate a probabilistic formulation into the TAP pipeline. Given an image sequence " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": " and a single query point " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_q" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": ", our objective is to compute the conditional probability density of point trajectory as " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q)" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": ". Recognizing the strong correlation between points within the same track, treating each point in the track as independent becomes suboptimal. Therefore, we propose to model the distribution of point trajectories by applying two multivariate distributions, respectively for the 2D coordinates. Let " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = [\\mathbf{a};\\mathbf{b}]" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{a}\\in \\mathbb{R}^S" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{b}\\in \\mathbb{R}^{S}" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": " coordinate of all points in " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": ", respectively. Assuming independence between these two coordinates, the joint probability distribution can be expressed as " + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "inline_equation", + "content": "p(\\mathbf{X}|\\mathbf{I},\\mathbf{x}_q) = p(\\mathbf{a}|\\mathbf{I},\\mathbf{x}_q)\\cdot p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)" + }, + { + "bbox": [ + 46, + 402, + 288, + 618 + ], + "type": "text", + "content": ". We adopt the multivariate Cauchy distribution, which is associated with its heavy tails distribution and is more stable for optimization. The probability density function (PDF) of a single coordinate is given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 624, + 289, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 289, + 665 + ], + "type": "interline_equation", + "content": "p (\\mathbf {a} | \\mathbf {I}, \\mathbf {x} _ {q}) = \\frac {\\Gamma \\left(\\frac {1 + S}{2}\\right)}{\\Gamma \\left(\\frac {1}{2}\\right) \\pi^ {\\frac {S}{2}} \\left| \\boldsymbol {\\Sigma} _ {a} \\right| ^ {\\frac {1}{2}} \\left[ 1 + \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) ^ {T} \\boldsymbol {\\Sigma} _ {a} ^ {- 1} \\left(\\mathbf {a} - \\boldsymbol {\\mu} _ {a}\\right) \\right] ^ {\\frac {1 + S}{2}}} \\tag {4}", + "image_path": "5bf04c4d606a75e3ddaed0cabcfb36840865583539eec590bb133d722d547f3f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "and a similar expression is applicable for " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "p(\\mathbf{b}|\\mathbf{I},\\mathbf{x}_q)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " is the Gamma function. The parameters " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{\\mu}_a,\\boldsymbol{\\Sigma}_a,\\boldsymbol{\\mu}_b,\\boldsymbol{\\Sigma}_b)" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " represent the location and scale matrices for the respective coordinates. During inference, the uncertainty associated" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 74, + 545, + 187 + ], + "blocks": [ + { + "bbox": [ + 312, + 74, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 312, + 74, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 312, + 74, + 545, + 187 + ], + "type": "image", + "image_path": "a70127acf14c96950f4d02f1fdf61d4ad2cb1f4fc795d5e9f3b2983140ece6d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 192, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 237 + ], + "type": "text", + "content": "Figure 3. LEAP Front-end. Once image feature maps are obtained, selected anchors aid in tracking. The queries and anchors are processed by a refiner to iteratively update states. The model outputs trajectory distribution, visibility, and dynamic track label." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "type": "text", + "content": "with each point is quantified by the sum of diagonal scale estimates at position " + }, + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "type": "inline_equation", + "content": "(s,s)" + }, + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 243, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\phi (\\mathbf{x}_s) = \\pmb {\\Sigma}_a[s,s] + \\pmb {\\Sigma}_b[s,s]" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": "Kernel-based Estimation. Rather than just predicting the positions of points within a trajectory, our model focuses on recovering the distribution of point trajectories by estimating its parameters. The location parameter predictions can be associated with the previous point trajectory as " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "[\\pmb{\\mu}_a; \\pmb{\\mu}_b] = \\mathbf{X}" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": ". However, directly deriving a symmetric and positive definite scale matrix from model outputs is challenging. To address this, we employ a kernel-based approach with the linear kernel " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "K(\\mathbf{x}, \\mathbf{y}) = \\mathbf{x}^T \\mathbf{y}" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": ". During each iteration, two scale matrices " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "(\\pmb{\\Sigma}_a, \\pmb{\\Sigma}_b)" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": " are constructed by first applying two linear projection layers to the point features, represented as " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a^k = \\mathcal{G}_a(\\mathbf{F}^k)" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_b^k = \\mathcal{G}_b(\\mathbf{F}^k)" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": ". We then compute the scale matrices as " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_a = K(\\mathbf{F}_a, \\mathbf{F}_a) + \\sigma \\mathbf{I}" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_b = K(\\mathbf{F}_b, \\mathbf{F}_b) + \\sigma \\mathbf{I}" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": " is a small positive value and " + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 279, + 546, + 483 + ], + "type": "text", + "content": " is the identity matrix. The model parameters are refined through Maximum Likelihood Estimation (MLE) using a negative log-likelihood (NLL) loss function." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 495, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 544 + ], + "type": "text", + "content": "Loss Functions. We supervise the point trajectory mainly with the NLL loss, which is based on the predicted distribution parameters and the ground-truth point trajectory. The main point trajectory loss is given as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 332, + 550, + 545, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 550, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 332, + 550, + 545, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a i n}} = \\sum_ {k} ^ {K} \\gamma^ {K - k} \\mathcal {L} _ {N L L} \\left(\\mathbf {X} ^ {k}, \\mathbf {X} ^ {*}, \\boldsymbol {\\Sigma} _ {a} ^ {k}, \\boldsymbol {\\Sigma} _ {b} ^ {k}\\right), \\tag {5}", + "image_path": "6c8a3d5e63363b6e2489bc2540012b931ee08db4f533f5e54b14e0306c1d7e60.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "text", + "content": " is the number of iteration and " + }, + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "inline_equation", + "content": "\\gamma = 0.8" + }, + { + "bbox": [ + 306, + 590, + 501, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": "For the visibility and dynamic track label supervision, we use the cross entropy loss with the estimates " + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V},\\mathbf{m}_d" + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{*},\\mathbf{m}_{d}^{*}" + }, + { + "bbox": [ + 304, + 601, + 545, + 638 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 331, + 645, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 645, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 331, + 645, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {v i s}} = \\left(1 - \\mathbf {V} ^ {*}\\right) \\log \\left(1 - \\mathbf {V}\\right) + \\mathbf {V} ^ {*} \\log \\mathbf {V}. \\tag {6}", + "image_path": "4fd5c86fcdf74ad9c50695ad7d75ab0fa77f400e0acb43c9de074a052636eeea.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 661, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 661, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 331, + 661, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d y n} = \\left(1 - \\mathbf {m} _ {d} ^ {*}\\right) \\log \\left(1 - \\mathbf {m} _ {d}\\right) + \\mathbf {m} _ {d} ^ {*} \\log \\mathbf {m} _ {d}. \\tag {7}", + "image_path": "b96e27c2485beed9c0ec509a074f9cec1379a98e04e931da1bbec5a9fb4c83e3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 681, + 499, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 681, + 499, + 693 + ], + "spans": [ + { + "bbox": [ + 306, + 681, + 499, + 693 + ], + "type": "text", + "content": "The total loss is a weighted sum of three losses:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 341, + 701, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 701, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 341, + 701, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = w _ {1} \\mathcal {L} _ {\\text {m a i n}} + w _ {2} \\mathcal {L} _ {\\text {v i s}} + w _ {3} \\mathcal {L} _ {\\text {d y n}}. \\tag {8}", + "image_path": "64dd3e5e09575f2a07656b0104505b535eee6ccc4f3873501fbd45faf3627b3d.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19847" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 68, + 309, + 233 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 309, + 233 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 309, + 233 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 309, + 233 + ], + "type": "image", + "image_path": "005c323b67834729f12b46342f226d06e34bd05848cbaea2bb4db7b3f676de27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": "Figure 4. LEAP-VO pipeline. Given a new image " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": " received at time " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": ", the keypoint extractor extracts new keypoints from " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": ". Then, all the keypoints from the latest " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "S_{LP}" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": "-frame " + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{t - S_{LP} + 1:t}" + }, + { + "bbox": [ + 46, + 236, + 547, + 281 + ], + "type": "text", + "content": " are tracked across all other frames within the current LEAP window, followed by a track filtering step to remove outliers. Finally, the local BA module is used on the current BA window to update the camera poses and 3D positions of the extracted keypoints. The colored arrows denote the moving direction of each module when a new image is received." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 78, + 525, + 233 + ], + "blocks": [ + { + "bbox": [ + 310, + 78, + 525, + 233 + ], + "lines": [ + { + "bbox": [ + 310, + 78, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 310, + 78, + 525, + 233 + ], + "type": "image", + "image_path": "bffe5d79cb06a7a1728eda2276e66086b413c42983dcc93b2df1db81da7e0b08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 286, + 119, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 119, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 119, + 298 + ], + "type": "text", + "content": "3.4. LEAP-VO" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 289, + 413 + ], + "type": "text", + "content": "Leveraging the anchor-based dynamic track estimations and temporal probabilistic formulation, we develop our Long-term Effective Point Tracking (LEAP) pipeline, boosting the accuracy of point tracking and simultaneously capturing global motion patterns and uncertainties of the tracks (refer to Figure 3). In this section, we demonstrate how we can build visual odometry for dynamic scenes based on the long-term tracking capabilities of LEAP. Our LEAP-VO pipeline is illustrated in Figure 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": "Keypoint Extraction and Tracking. In our VO system, the LEAP model serves as the front-end. When a new image frame " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " is captured at time " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", we sample " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " new keypoints from this image to initiate point tracking. We consistently employ the distributed image gradient-based sampling method (refer to Section 3.2) for keypoint selection due to its favorable properties in feature distinctiveness and distribution. This approach also obviates the need for additional anchors during tracking. The LEAP model then tracks these keypoints across the last " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "S_{LP}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " frames, performing bidirectional tracking (both forward and backward in time). Notably, " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "S_{LP}" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": " can exceed the model window " + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 426, + 287, + 594 + ], + "type": "text", + "content": ", thus enabling more expansive tracking capabilities through sequential prediction [16, 18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Track Filtering. Theoretically, precise camera pose estimation between two images can be attained with just a few accurate correspondences under certain conditions. However, the quality of correspondences obtained from feature matching or optical flow can vary. Traditional VO pipelines often use RANSAC [14] to filter out incorrect matches, but their efficiency decreases with more keypoints and views. Our approach, in contrast, bypasses the sampling-based method and uses trajectory quality assessments from LEAP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 287, + 503, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 503, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 503, + 299 + ], + "type": "text", + "content": "for more efficient and informative track filtering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 556 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "content": "(i) Visibility and Dynamic Track Label: The LEAP model has the capability to estimate not only the 2D point positions within a trajectory but also the visibility of each point and the dynamic track label of the entire trajectory. Thresholds " + }, + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\gamma_v" + }, + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "content": " for visibility and " + }, + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\gamma_d" + }, + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "content": " for dynamic track labels are set to ensure that only visible and static points are utilized in the bundle adjustment process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": "(ii) Track Uncertainty: In Section 3.3, we introduce probabilistic modeling to LEAP, allowing us to gauge confidence via distribution measurements. We select only high-confidence points, which correspond to low uncertainty measurements. Denoting our model's uncertainty estimation as " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": ", we retain points that exhibit high confidence using the criterion " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\Phi \\leq Q(\\gamma_u)" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "\\gamma_u" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " is the uncertainty quantile and " + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "inline_equation", + "content": "Q:[0,1] \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 387, + 545, + 483 + ], + "type": "text", + "content": " is the quantile function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "text", + "content": "(iii) Track Length: After filtering, we evaluate and remove tracks with insufficient observations, as bundle adjustment is more reliable with increased observations and wider baselines. Tracks with fewer than " + }, + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "inline_equation", + "content": "\\gamma_{\\text{track}}" + }, + { + "bbox": [ + 304, + 484, + 545, + 556 + ], + "type": "text", + "content": " valid points are excluded from optimization cost computation due to their potential unreliability." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "Sliding-window Optimization. With the point validation mask established, we proceed to define the optimization cost function for our sliding window bundle adjustment. The window size for local bundle adjustment, denoted as " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{BA}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ", may differ from the point tracking window size " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{LP}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Our geometric bundle adjustment aims to optimize camera poses " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and 3D scene point positions " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " by aligning the induced point trajectory from the projective relationship, with the estimated point trajectory from LEAP. We parameterize a 3D scene point " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " by its 2D keypoint location " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LEAP}_{i\\rightarrow j}" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " denote the mapping of tracking keypoint from " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " using LEAP model. The" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19848" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 188, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 188, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 188, + 85 + ], + "type": "text", + "content": "reprojection cost is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 94, + 287, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 94, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 94, + 287, + 138 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {i} \\sum_ {j \\in | i - j | \\leq S _ {B A}} \\sum_ {n} w _ {i \\rightarrow j, n} \\| \\mathcal {P} (\\mathbf {T} _ {i}, \\mathbf {T} _ {j}, \\mathbf {K}, d _ {i, n}) - \\\\ \\operatorname {L E A P} _ {i \\rightarrow j} (\\mathbf {x} _ {i, n}) \\| _ {\\rho}, \\tag {9} \\\\ \\end{array}", + "image_path": "1eb6b40196d4bc7c3e1c37c22f6e323355319c336b129e1c1205088561b39793.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_{\\rho}" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": " is the distance metric, and the weight " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "w_{i\\rightarrow j,n}" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": " is derived from the track filtering results. We use the GaussNewton method [34] to optimize Eq. (9) for " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "K_{BA}" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": " iterations. At each iteration, we compute the camera poses update " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\Delta \\xi^{(k)}\\in \\mathfrak{se}(3)" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": " (lie-algebra corresponding to " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": "), and the depth update " + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{D}^{(k)}" + }, + { + "bbox": [ + 46, + 148, + 288, + 244 + ], + "type": "text", + "content": " for each point. The optimization can be solved efficiently with the Schur decomposition. We use the robust Huber loss function for the distance metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 123, + 270 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 277, + 168, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 168, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 168, + 289 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 296, + 288, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 288, + 381 + ], + "type": "text", + "content": "Replica. Replica [29] provides synthetic indoor environments for simulating the image-capturing process with a moving camera. We use the camera trajectories provided by [50] for the localization task, resulting in 16 camera trajectories derived from 8 static scenes. Each sequence comprises 900 frames with rendered RGB images. We use \"Sequence 1\" for the visual odometry evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": "MPI Sintel. The MPI Sintel [4] dataset contains dynamic image sequences from open-source 3D animated short films. These sequences consist of large and complex object motion, along with other imaging effects such as motion blur and defocus blur. Following prior works [20, 48], we assess dynamic VO performance on the MPI Sintel dataset, with each sequence containing 20 to 50 image frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 578 + ], + "type": "text", + "content": "TartanAir Shibuya. The TartanAir Shibuya dataset [26] features dynamic scenes from two scenarios: \"Standing Humans\" and \"Road Crossing\". The majority of the humans remain stationary in \"Standing Humans\" sequences, while \"Road Crossing\" sequences provide a more challenging setting where multiple humans are moving in different directions. Each sequence in the dataset includes 100 frames with more than 30 tracked moving humans." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 584, + 288, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 288, + 717 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 288, + 717 + ], + "type": "text", + "content": "Metrics. For visual odometry, we compare the Absolute Translation Error (ATE), Relative Translation Error (RPE trans), and Relative Rotation Error (RPE rot) following [33, 47, 48]. ATE measures how much the estimated trajectory deviated from the ground truth trajectory, which is calculated as the root mean square error (RMSE) overall all pose transformations after scaling and alignment. \"RPE trans\" computes translation errors made over a certain distance (meter), and \"RPE rot\" computes the estimated rotation over a certain distance (degree). Both \"RPE trans\" and \"RPE rot\" are calculated on all poses and then averaged." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 178 + ], + "type": "table", + "html": "
MethodReplica
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]0.0860.0300.650
DynaSLAM [1]0.0390.0170.366
DROID-SLAM [32]0.2670.0362.631
TartanVO [40]0.4060.0362.063
DytanVO [28]0.2890.0352.146
DPVO [33]0.2570.0362.635
LEAP-VO (Ours)0.2040.0301.992
", + "image_path": "9faf726c5e5fb61c91080f585bbf38c00eb5c872769d353a44f2b26fadf5eaf6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 308, + 220, + 544, + 329 + ], + "blocks": [ + { + "bbox": [ + 305, + 179, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 213 + ], + "type": "text", + "content": "Table 1. Camera tracking results on Replica [29]. The statistics demonstrate that our method achieves better results than other competitive baselines." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 220, + 544, + 329 + ], + "lines": [ + { + "bbox": [ + 308, + 220, + 544, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 544, + 329 + ], + "type": "table", + "html": "
MethodMPI Sintel
ATE (m)RPE trans (m)RPE rot (deg)
ORB-SLAM2 [24]XXX
DynaSLAM [1]XXX
DROID-SLAM [32]0.1750.0841.912
TartanVO [40]0.2380.0931.305
DytanVO [28]0.1310.0971.538
DPVO [33]0.0760.0781.722
LEAP-VO (Ours)0.0370.0551.263
", + "image_path": "0740f1e91760ce2e792e2709686305b58406ec0e90f256e01d9deb71f739d8d8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 364 + ], + "type": "text", + "content": "Table 2. Camera tracking results on MPI Sintel [4]. The statistics show that our method outperforms other competitive baselines. 'X' denotes the tracking failure case." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 370, + 437, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 437, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 437, + 383 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": "We employ CoTracker [18], a recent PIPs variant, as our base TAP model. We train our model on the TAP-Vid-Kubric [9] training set, following [18]. We load the pretrained CoTracker weights for the image feature extractor and train LEAP for 100,000 steps using 4 NVIDIA A100 GPUs in parallel. During training, we use " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "K = 4" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": " steps for iterative refinements, " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": " for the number of queries, " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "S = 8" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": " for the model window, and " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "S_{LP} = 12" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": " for the tracking window. For inter-track anchor-based attention, the number of anchors is 64. The loss weights are set to " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "w_{1} = 1.0, w_{2} = 0.5, w_{3} = 0.5" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": ". In LEAP-VO, we perform 4 steps for each bundle adjustment with the Huber Loss for distance metric. The bundle adjustment window size " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "S_{BA}" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": " is set to 15. Track filtering parameters are set at " + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "inline_equation", + "content": "\\gamma_v = 0.9, \\gamma_d = 0.9, \\gamma_u = 0.8, \\gamma_{track} = 3" + }, + { + "bbox": [ + 305, + 388, + 546, + 581 + ], + "type": "text", + "content": ". Please refer to our supplementary materials for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 587, + 446, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 446, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 446, + 601 + ], + "type": "text", + "content": "4.3. Visual Odometry Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 605, + 545, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 678 + ], + "type": "text", + "content": "We compare our methods against the state-of-the-art SLAM and VO approaches, including ORB-SLAM2 [24], DynaSLAM [1], DROID-SLAM [32], TartanVO [40], DytanVO [28], and DPVO [33]. Evaluations are conducted on the static Replica dataset [29] as well as on the dynamic datasets MPI-Sintel [4] and TartanAir-Shibuya [26]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Replica. We compare our method with other baselines on the Replica dataset, which contains eight static scenes" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19849" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 517, + 187 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 517, + 187 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 517, + 187 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 517, + 187 + ], + "type": "table", + "html": "
MethodStandingHumanRoadCrossing (Easy)RoadCrossing (Hard)Average
01020304050607
ORB-SLAM2 [24] w/ mask0.07880.00600.06570.01960.01481.09840.84760.3044
DynaSLAM [1]X0.88360.39070.41960.49250.64460.6539(0.5808)
AirDOS [26] w/ mask0.06060.01930.09510.03310.02060.22300.56250.1449
DROID-SLAM [32]0.00510.00730.01030.01200.22780.02530.57880.1238
TartanVO [40]0.06000.16050.27620.18140.21740.32280.50090.2456
DytanVO [28]0.03270.10170.06080.05160.07550.03650.06600.0607
DPVO [39]0.05390.16030.17880.28580.11960.11190.14980.1514
LEAP-VO (Ours)0.00810.01850.03990.03170.01130.06890.02460.0290
", + "image_path": "88ddaacfec9a855b6b763fd39bd052dacd1e76b1bf07559f5abfc40ca0b7d100.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "type": "text", + "content": "Table 3. ATE (m) results on TartanAir-Shibuya Sequences [26]. (" + }, + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 189, + 547, + 212 + ], + "type": "text", + "content": ") denotes averaging on valid scenes only. We can observe that our method shows better results for ATE on most sequences, outperforming all SLAM and VO baselines on average ATE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 52, + 217, + 171, + 334 + ], + "blocks": [ + { + "bbox": [ + 52, + 217, + 171, + 334 + ], + "lines": [ + { + "bbox": [ + 52, + 217, + 171, + 334 + ], + "spans": [ + { + "bbox": [ + 52, + 217, + 171, + 334 + ], + "type": "image", + "image_path": "0c9be0f3fcfc59cefc63ad77edac4adc0e039eefe5b343e437c31e099d237ab9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 335, + 145, + 345 + ], + "lines": [ + { + "bbox": [ + 78, + 335, + 145, + 345 + ], + "spans": [ + { + "bbox": [ + 78, + 335, + 145, + 345 + ], + "type": "text", + "content": "(a) RoadCrossing 04" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 216, + 294, + 334 + ], + "blocks": [ + { + "bbox": [ + 176, + 216, + 294, + 334 + ], + "lines": [ + { + "bbox": [ + 176, + 216, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 176, + 216, + 294, + 334 + ], + "type": "image", + "image_path": "f5a4197b9fd43580001e00f745e1a2cbf7cc380b0001405e7443351579a9253f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 201, + 335, + 269, + 345 + ], + "lines": [ + { + "bbox": [ + 201, + 335, + 269, + 345 + ], + "spans": [ + { + "bbox": [ + 201, + 335, + 269, + 345 + ], + "type": "text", + "content": "(b) RoadCrossing 05" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 216, + 419, + 334 + ], + "blocks": [ + { + "bbox": [ + 301, + 216, + 419, + 334 + ], + "lines": [ + { + "bbox": [ + 301, + 216, + 419, + 334 + ], + "spans": [ + { + "bbox": [ + 301, + 216, + 419, + 334 + ], + "type": "image", + "image_path": "422bd73a1d36da5ecdab37f323089a41085702b01500552c15b23bbc9e085202.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 335, + 392, + 345 + ], + "lines": [ + { + "bbox": [ + 324, + 335, + 392, + 345 + ], + "spans": [ + { + "bbox": [ + 324, + 335, + 392, + 345 + ], + "type": "text", + "content": "(c) RoadCrossing 06" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 546, + 370 + ], + "type": "text", + "content": "Figure 5. Qualitative results of camera trajectory estimation on TartanAir-Shibuya [26]. The visualizations show that our method provides more robust and accurate camera pose trajectories, especially in hard cases (RoadCrossing 06 and RoadCrossing 07)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 424, + 217, + 542, + 334 + ], + "blocks": [ + { + "bbox": [ + 424, + 217, + 542, + 334 + ], + "lines": [ + { + "bbox": [ + 424, + 217, + 542, + 334 + ], + "spans": [ + { + "bbox": [ + 424, + 217, + 542, + 334 + ], + "type": "image", + "image_path": "500f27a167516833358616cab081a003fdad56f68b6e256912b551c566d37440.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 335, + 517, + 345 + ], + "lines": [ + { + "bbox": [ + 449, + 335, + 517, + 345 + ], + "spans": [ + { + "bbox": [ + 449, + 335, + 517, + 345 + ], + "type": "text", + "content": "(d) RoadCrossing 07" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "type": "text", + "content": "without any moving objects. Challenges arise in two test sequences that contain prolonged occlusions, leading to the potential camera tracking loss. As illustrated in Table 1, our approach outperforms other VO methods on all three metrics, showing effectiveness under static environments and the capability to handle occlusions. Further insights are available in Section 4.5. For classical methods, ORBSLAM2 [24] and DynaSLAM [1], we observe tracking loss and resuming via re-localization, which reduces long-term drift and shows high accuracy for camera pose tracking." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 503, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 660 + ], + "type": "text", + "content": "MPI-Sintel. We also evaluate our VO camera trajectory accuracy on 14 dynamic scenes from the MPI Sintel dataset, following the setting of [48]. As shown in Table 2, our method substantially surpasses other VO baselines in dynamic scenes, thanks to our temporal uncertainty estimation and dynamic track detection. Traditional feature-based methods like ORB-SLAM2 [24] and DynaSLAM [1] sometimes fail to track camera movements in highly dynamic sequences, primarily due to outliers from dynamic object correspondences and segmentation-based only dynamic track estimation. In contrast, our approach, which considers visual, temporal, and inter-track information, demonstrates notable improvements in dynamic scene tracking accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "TartanAir-Shibuya. Following DytanVO [28], we further evaluate the VO performance using the dynamic TartanAir-Shibuya dataset. We assess the Average Trajectory Error (ATE) across seven diverse scenes, ranging from simple to" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 378, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 437 + ], + "type": "text", + "content": "challenging. As depicted in Table 3, our method consistently outperforms all other SLAM and VO baselines in the majority of these scenes. Moreover, trajectory visualizations in Figure 5 also highlight the robust camera tracking capabilities of our method in dynamic environments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 443, + 492, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 443, + 492, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 492, + 456 + ], + "type": "text", + "content": "4.4. Dynamic Track Estimation Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "text", + "content": "In our LEAP formulation, we have introduced the anchor-based dynamic track estimation module to distinguish between static and dynamic point trajectories, which provides valuable information for handling dynamic scenes in visual odometry. Our method leverages both point feature information and inter-track relations to predict the dynamic track label for each trajectory. To demonstrate the performance of the proposed dynamic track estimation module, we take the LEAP model and evaluate it on novel image sequences from diverse datasets. For a comprehensive demonstration, we uniformly sample " + }, + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "text", + "content": " queries among the images and track them for " + }, + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 304, + 461, + 546, + 643 + ], + "type": "text", + "content": " frames. The results, shown in Figure 6, illustrate that LEAP can effectively handle various scenarios, including single-object motion, complex and rapid motion, and multi-object motion with occlusion." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "content": "Track Filtering. To highlight the efficacy of our track filtering module in LEAP-VO, we conduct an ablation study on different masks used in track filtering. We study three types of filtering criteria: visibility, dynamic" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19850" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 283, + 253 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 283, + 253 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 283, + 253 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 283, + 253 + ], + "type": "image", + "image_path": "781af32895b8c1318275094eab868761f2d0179d29830e8800be855692ef374b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 255, + 287, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 300 + ], + "type": "text", + "content": "Figure 6. Visualization of dynamic track estimation on DAVIS [19], MPI-Sintel [4], and TartanAir-Shibuya [26]. Odd columns: all point trajectories. Even columns: estimated dynamic point trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 304, + 285, + 416 + ], + "blocks": [ + { + "bbox": [ + 50, + 304, + 285, + 416 + ], + "lines": [ + { + "bbox": [ + 50, + 304, + 285, + 416 + ], + "spans": [ + { + "bbox": [ + 50, + 304, + 285, + 416 + ], + "type": "image", + "image_path": "849ddeb126a9119bf85089fee5033115a331c32107fbd180edaeec2c0c8ce3b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "lines": [ + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "type": "text", + "content": "Figure 7. Visualization of point-wise uncertainty measurements. Keypoints with the lowest " + }, + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "type": "text", + "content": " (left) and highest " + }, + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 418, + 286, + 452 + ], + "type": "text", + "content": " (right) uncertainty are shown in green and yellow, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": "track, and confidence filtering. Table 4 shows the impact of applying various filtering types on the dynamic MPI Sintel benchmark. Without dynamic filtering or confidence filtering, the ATE increases by " + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "inline_equation", + "content": "130\\%" + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": ", respectively. Combining all three types of filtering methods altogether, we achieve the best results in all three metrics." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 52, + 537, + 282, + 612 + ], + "blocks": [ + { + "bbox": [ + 52, + 537, + 282, + 612 + ], + "lines": [ + { + "bbox": [ + 52, + 537, + 282, + 612 + ], + "spans": [ + { + "bbox": [ + 52, + 537, + 282, + 612 + ], + "type": "table", + "html": "
MethodTrack FilteringMPI Sintel
Vis.Dyn.Conf.ATERPE transRPE rot
LEAP-VO✓--0.1180.0781.340
✓✓-0.0850.0621.281
✓-✓0.0540.0581.274
✓✓✓0.0370.0551.263
", + "image_path": "1ee9533292ee09daae5704bd49dc27defff0357868632af005d5f5d707109e22.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Uncertainty Estimation. As illustrated in Figure 7, we demonstrate the LEAP's capability to provide meaningful per-point uncertainty. It models the uncertainty as distribution parameters for both the entire track and individual points. Two key observations can be found from Figure 7. Firstly, our model reliably identifies high-uncertainty" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "points, typically located in low-texture areas or in regions with repetitive patterns. Secondly, although not explicitly supervised, the model tends to assign higher uncertainty to dynamic objects, which are inherently more challenging to track. These findings highlight the efficacy of LEAP's probabilistic formulation in handling uncertainty." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 152, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 258 + ], + "type": "text", + "content": "Keypoint Extraction. Table 5 showcases the effectiveness of various keypoint extraction strategies for VO performance. Despite its distinctive features, SIFT extraction provides inferior performance due to its unpredictable keypoint distribution. In contrast, the random sampling tends to generate well-separated keypoints, leading to a performance similar to SIFT's. Our distributed image gradient-based sampling shows the best performance, underscoring the significance of feature distribution and distinctiveness." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 335, + 259, + 517, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 614, + 286, + 636 + ], + "lines": [ + { + "bbox": [ + 47, + 614, + 286, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 286, + 636 + ], + "type": "text", + "content": "Table 4. Effect of the proposed track filtering method on MPI-Sintel [4]. Tracking using all criteria yields the best results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 259, + 517, + 291 + ], + "lines": [ + { + "bbox": [ + 335, + 259, + 517, + 291 + ], + "spans": [ + { + "bbox": [ + 335, + 259, + 517, + 291 + ], + "type": "table", + "html": "
Policy:UniformRandomSIFT [22]Ours
ATE (m):0.15110.03660.03550.0290
", + "image_path": "97c60765d761296eda93eb56bac7175dba27f1dda620d9805ab0652b917566aa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 389 + ], + "type": "text", + "content": "Baseline comparison. We evaluate the pretrained CoTracker [18] for the end VO performance. Table 6 shows that LEAP front-end outperforms CoTracker on all datasets, demonstrating the effectiveness of the proposed dynamic track estimation and temporal probability modeling." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 310, + 393, + 539, + 449 + ], + "blocks": [ + { + "bbox": [ + 304, + 293, + 545, + 326 + ], + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 326 + ], + "type": "text", + "content": "Table 5. Comparison of keypoint extraction policies (" + }, + { + "bbox": [ + 304, + 293, + 545, + 326 + ], + "type": "inline_equation", + "content": "N_{a} = 64" + }, + { + "bbox": [ + 304, + 293, + 545, + 326 + ], + "type": "text", + "content": ") on TartanAir-Shibuya [26]. Our distributed image-gradient sampling method achieves better results compared to other policies." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 393, + 539, + 449 + ], + "lines": [ + { + "bbox": [ + 310, + 393, + 539, + 449 + ], + "spans": [ + { + "bbox": [ + 310, + 393, + 539, + 449 + ], + "type": "table", + "html": "
Front-endReplicaMPI SintelShibuya
ATERPE transRPE rotATERPE transRPE rotATE
CoTracker0.3010.0312.0430.1260.0711.3660.160
LEAP0.2040.0301.9920.0370.0551.2630.029
", + "image_path": "8efc96d69cf79bff7d4790d40d3929500fccf9ebdc1a343f9fdf266338968788.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 449, + 545, + 470 + ], + "lines": [ + { + "bbox": [ + 306, + 449, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 545, + 470 + ], + "type": "text", + "content": "Table 6. Baseline comparison with CoTracker front-end [18]. Our LEAP front-end achieves better results on all datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "We introduce Long-term Effective Any Point Tracking (LEAP), designed to tackle the limitations of previous point-tracking methods. By leveraging visual, inter-track, and temporal cues within an anchor-based dynamic track estimation module, LEAP captures global motion patterns and effectively discriminates between static and dynamic elements. Moreover, our approach incorporates a temporal probabilistic formulation into the point-tracking pipeline, emulating recursive Bayesian filtering through a learning-based refinement module. This enables our model to accurately assess the reliability of point trajectory estimates. Using LEAP as the front-end, we then develop the LEAP-VO system, featuring novel integration of visual information, global motion patterns, and track distribution for handling dynamic scenes. Additionally, our takes for LEAP-VO can be integrated with other TAP-based front-ends, offering the potential to markedly improve camera-tracking accuracy and robustness." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19851" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Berta Bescos, José M Fácil, Javier Civera, and José Neira. Dynaslam: Tracking, mapping, and inpainting in dynamic scenes. IEEE Robotics and Automation Letters, 3(4):4076-4083, 2018. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-tap: Tracking any point demands spatial context features. arXiv preprint arXiv:2306.02000, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 287, + 223 + ], + "type": "text", + "content": "[3] Michael Bloesch, Jan Czarnowski, Ronald Clark, Stefan Leutenegger, and Andrew J Davison. Codeslam—learning a compact, optimisable representation for dense visual slam. In CVPR, pages 2560–2568, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 257 + ], + "type": "text", + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In ECCV, pages 611-625. Springer-Verlag, 2012. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In CVPR, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 345 + ], + "type": "text", + "content": "[6] Weirong Chen, Suryansh Kumar, and Fisher Yu. Uncertainty-driven dense two-view structure from motion. IEEE Robotics and Automation Letters, 8(3):1763-1770, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "text", + "content": "[7] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In ICCV, pages 7063-7072, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 423 + ], + "type": "text", + "content": "[8] Andrew J Davison, Ian D Reid, Nicholas D Molton, and Olivier Stasse. Monoslam: Real-time single camera slam. IEEE TPAMI, 29(6):1052-1067, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 287, + 468 + ], + "type": "text", + "content": "[9] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. NeurIPS, 35:13610-13626, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "type": "text", + "content": "[10] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 567 + ], + "type": "text", + "content": "[11] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. arXiv preprint arXiv:2402.00847, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 568, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 602 + ], + "type": "text", + "content": "[12] Jakob Engel, Thomas Schöps, and Daniel Cremers. Ldslam: Large-scale direct monocular slam. In ECCV, pages 834-849. Springer, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 602, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 602, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 602, + 287, + 624 + ], + "type": "text", + "content": "[13] Jakob Engel, Vladlen Koltun, and Daniel Cremers. Direct sparse odometry. IEEE TPAMI, 40(3):611-625, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 668 + ], + "type": "text", + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 287, + 712 + ], + "type": "text", + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[16] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, pages 59-75. Springer, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 151 + ], + "type": "text", + "content": "[17] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 33: 19545-19560, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 196 + ], + "type": "text", + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv preprint arXiv:2307.07635, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 545, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 231 + ], + "type": "text", + "content": "[19] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video object segmentation with language referring expressions. In ACCV, 2018. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 234, + 545, + 266 + ], + "type": "text", + "content": "[20] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 311 + ], + "type": "text", + "content": "[21] Shunkai Li, Xin Wang, Yingdian Cao, Fei Xue, Zike Yan, and Hongbin Zha. Self-supervised deep visual odometry with online adaptation. In CVPR, pages 6339-6348, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 314, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 545, + 336 + ], + "type": "text", + "content": "[22] David G Lowe. Object recognition from local scale-invariant features. In ICCV, pages 1150-1157. IEEE, 1999. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[23] Annette Mossel and Manuel Kroeter. Streaming and exploration of dynamically changing dense 3d reconstructions in immersive virtual reality. In 2016 IEEE International Symposium on Mixed and Augmented Reality (ISMAR-Adjunct), pages 43–48. IEEE, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 395, + 545, + 437 + ], + "type": "text", + "content": "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: an open-source SLAM system for monocular, stereo and RGB-D cameras. IEEE Transactions on Robotics, 33(5):1255-1262, 2017. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 483 + ], + "type": "text", + "content": "[25] Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: a versatile and accurate monocular slam system. IEEE Transactions on Robotics, 31(5):1147-1163, 2015. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 486, + 545, + 529 + ], + "type": "text", + "content": "[26] Yuheng Qiu, Chen Wang, Wenshan Wang, Mina Henein, and Sebastian Scherer. Airdos: Dynamic slam benefits from articulated objects. In ICRA, pages 8047-8053. IEEE, 2022. 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[27] Anurag Ranjan, Varun Jampani, Lukas Balles, Kihwan Kim, Deqing Sun, Jonas Wulff, and Michael J Black. Competitive collaboration: Joint unsupervised learning of depth, camera motion, optical flow and motion segmentation. In CVPR, pages 12240-12249, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[28] Shihao Shen, Yilin Cai, Wenshan Wang, and Sebastian Scherer. Dytanvo: Joint refinement of visual odometry and motion segmentation in dynamic environments. In ICRA, pages 4048-4055. IEEE, 2023. 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 634, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 688 + ], + "type": "text", + "content": "[29] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 712 + ], + "type": "text", + "content": "[30] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. In ICLR, 2019. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19852" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, pages 402-419. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 108, + 286, + 140 + ], + "type": "text", + "content": "[32] Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras. NeurIPS, 34: 16558-16569, 2021. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 143, + 286, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 143, + 286, + 164 + ], + "spans": [ + { + "bbox": [ + 49, + 143, + 286, + 164 + ], + "type": "text", + "content": "[33] Zachary Teed, Lahav Lipson, and Jia Deng. Deep patch visual odometry. NeurIPS, 36, 2024. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 167, + 286, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 167, + 286, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 167, + 286, + 209 + ], + "type": "text", + "content": "[34] Bill Triggs, Philip F McLauchlan, Richard I Hartley, and Andrew W Fitzgibbon. Bundle adjustment—a modern synthesis. In International workshop on vision algorithms, pages 298-372. Springer, 1999. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 212, + 286, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 212, + 286, + 244 + ], + "spans": [ + { + "bbox": [ + 49, + 212, + 286, + 244 + ], + "type": "text", + "content": "[35] Vadim Tschernezki, Diane Larlus, and Andrea Vedaldi. Neuraliff: Segmenting 3d objects that move in egocentric videos. In 3DV, pages 910-919. IEEE, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 247, + 286, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 247, + 286, + 279 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 286, + 279 + ], + "type": "text", + "content": "[36] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. arXiv preprint arXiv:2403.14548, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 281, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 281, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 281, + 286, + 323 + ], + "type": "text", + "content": "[37] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. arXiv preprint arXiv:2306.05422, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 327, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 327, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 286, + 358 + ], + "type": "text", + "content": "[38] Rui Wang, Martin Schworer, and Daniel Cremers. Stereo dso: Large-scale direct sparse visual odometry with stereo cameras. In ICCV, pages 3903-3911, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 361, + 286, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 286, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 286, + 404 + ], + "type": "text", + "content": "[39] Sen Wang, Ronald Clark, Hongkai Wen, and Niki Trigoni. Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In ICRA, pages 2043-2050. IEEE, 2017. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 407, + 286, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 407, + 286, + 438 + ], + "spans": [ + { + "bbox": [ + 49, + 407, + 286, + 438 + ], + "type": "text", + "content": "[40] Wenshan Wang, Yaoyu Hu, and Sebastian Scherer. Tartanvo: A generalizable learning-based vo. In CoRL, pages 1761-1772. PMLR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 441, + 286, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 441, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 49, + 441, + 286, + 484 + ], + "type": "text", + "content": "[41] Tianhao Wu, Fangcheng Zhong, Andrea Tagliasacchi, Forrester Cole, and Cengiz Oztireli. D^2nerf: Self-supervised decoupling of dynamic and static objects from a monocular video. NeurIPS, 35:32653-32666, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 487, + 286, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 487, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 487, + 286, + 540 + ], + "type": "text", + "content": "[42] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 286, + 586 + ], + "type": "text", + "content": "[43] Nan Yang, Lukas von Stumberg, Rui Wang, and Daniel Cremers. D3vo: Deep depth, deep pose and deep uncertainty for monocular visual odometry. In CVPR, pages 1281-1292, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 589, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 589, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 49, + 589, + 286, + 632 + ], + "type": "text", + "content": "[44] Weicai Ye, Xinyue Lan, Shuo Chen, Yuhang Ming, Xingyuan Yu, Hujun Bao, Zhaopeng Cui, and Guofeng Zhang. Pvo: Panoptic visual odometry. In CVPR, pages 9579-9589, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 286, + 667 + ], + "type": "text", + "content": "[45] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In CVPR, pages 1983-1992, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 712 + ], + "type": "text", + "content": "[46] Khalid Yousif, Alireza Bab-Hadiashar, and Reza Hosein- nezhad. An overview to visual odometry and visual slam: Applications to mobile robotics. Intelligent Industrial Systems, 1(4):289-311, 2015. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In ECCV, pages 20-37. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "text", + "content": "[48] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In ECCV, pages 523-542. Springer, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[49] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, pages 19855-19865, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 250 + ], + "type": "text", + "content": "[50] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In CVPR, pages 15838-15847, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[51] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, pages 1851–1858, 2017. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19853" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_content_list.json b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6e45134b9dfde4fd648c955c8bc635afb2ee8661 --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "LEDS: Limitless Image Editing using Text-to-Image Models", + "text_level": 1, + "bbox": [ + 156, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d8f7be564b6332796a8cb8adaad3a2d7a5f0c215956e770c5c1a33382a955dd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 178, + 844, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd68bca424e2d5b2224b8fd297e5ce3f861d71132db180895b57fbeebd7121b1.jpg", + "image_caption": [ + "Figure 1. LEDs++ facilitates versatile image-to-image editing. Several complex cases are available now." + ], + "image_footnote": [], + "bbox": [ + 140, + 263, + 831, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 470, + 313, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDs++, an efficient yet versatile and precise textual image manipulation technique. LEDs++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDs++ and its improvements over previous methods.", + "bbox": [ + 75, + 505, + 473, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 842, + 209, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image diffusion models (DM) have garnered recognition for their ability to generate high-quality images from", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "textual descriptions. A growing body of research has recently been dedicated to utilizing these models for manipulating real images. However, several barriers prevent many real-world applications of diffusion-based image editing. Current methods often entail computationally expensive model tuning or other optimization, presenting practical challenges [6, 18, 28, 30, 44]. Additionally, existing techniques have the proclivity to induce profound changes to the original image [17, 26], often resulting in completely different images. Lastly, all these approaches are inherently constrained when editing multiple (arbitrary) concepts simultaneously. We tackle these problems by introducing LEDs++ $^{1}$ , a diffusion-based image editing technique addressing these limitations.", + "bbox": [ + 496, + 472, + 890, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LEDITS++2 offers a streamlined approach for textual image editing, eliminating the need for extensive parameter tuning. To this end, we derive image inversion for a more efficient diffusion sampling algorithm to a) drastically reduce computational resources and b) guarantee perfect image reconstruction. Thus, we overcome computational obstacles and avoid changes in the edited image in the first place. Furthermore, we use a novel implicit masking approach to semantically ground each edit instruction to its relevant image region. This further optimizes changes to the image by retaining the overall image composition and object identity.", + "bbox": [ + 496, + 684, + 892, + 851 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "*Equal contribution †Partially as research intern at Adobe", + "1LEDITS++ stands for Limitless Edits with sde-dpm-solver++.", + "2https://huggingface.co/spaces/leditsplusplus/project" + ], + "bbox": [ + 519, + 862, + 883, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8861", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Additionally, LEDs++ is the only method to date to facilitate easy and versatile image editing by supporting multiple simultaneous instructions without causing undue interference. Finally, its lightweight architecture-agnostic nature ensures compatibility with both latent and pixel-based diffusion models, providing high accessibility.", + "bbox": [ + 78, + 90, + 467, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we establish the methodical benefits of LEDs++ and demonstrate that this intuitive, lightweight approach offers sophisticated semantic control for image editing. Specifically, we contribute by (i) devising a formal definition of LEDs++ while (ii) deriving perfect inversion for a more efficient diffusion sampling method, (iii) qualitatively and empirically demonstrating its efficiency, versatility, and precision, (iv) providing an exhaustive empirical comparison to concurrent works with automatic and human user metrics, and thereby (v) introducing Textual Editing Benchmark++ (TEdBench++, a more holistic and coherent testbed for evaluating textual image manipulation.", + "bbox": [ + 78, + 181, + 467, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 78, + 376, + 202, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, large-scale, text-guided DMs have enabled versatile applications in image generation [3, 33, 37]. Especially latent diffusion models [31, 34] have gained attention for their computational efficiency. Below, we discuss related work for efficient, versatile image manipulation with DMs.", + "bbox": [ + 78, + 402, + 467, + 476 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Sampling. Generating outputs with DMs requires multiple iterative denoising steps that constitute the main bottleneck at inference. Commonly used sampling methods such as DDPM [15] or DDIM [42] require tens or hundreds of steps to produce high-quality samples. Consequently, numerous works have been dedicated to speeding up the sampling process without loss in quality. Distillation efforts progressively reduce the number of required steps through further training [25, 27]. Other works focus on improving the sampling itself, e.g. using high-order ODE-solvers [22, 23, 46]. Such solvers can be readily combined with pre-trained DMs at inference to lower the number of denoising steps. With LEDs++, we derive perfect image inversion with the DPM-Solver++, allowing image editing in as few as 20 total steps.", + "bbox": [ + 78, + 478, + 467, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic Control during Diffusion. While text-to-image DMs generate new, astonishing images, fine-grained control over the generative process remains challenging. Minor changes to the text prompt lead to entirely different outputs. Wu et al. [45] studied concept disentanglement using linear combinations of text embeddings to gain semantic control. Methods like Prompt-to-Prompt [14] and other works [8, 30] utilize the DM's attention layers to attribute pixels to tokens from the text prompt. Dedicated operations on the attention maps enable more control over the generated images. Other works have focused on the noise estimates of DMs [5, 20] providing semantic control over the generation process. With LEDs++, we now enable", + "bbox": [ + 78, + 704, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fine-grained semantic control for manipulating real images, going beyond purely generative applications.", + "bbox": [ + 503, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real Image Editing. Since DMs' rise in popularity for text-to-image generation, they have also been explored for (real) image-to-image editing. As a first, simple approach, SDEdit added noise to the image for an intermediate step in the diffusion process [26]. While lightweight, the resulting image diverges substantially from the input as it is (partially) regenerated. Inpainting allows to keep the change small by having a user provide additional masks to restrict changes to certain image regions [2, 29]. Yet, user masks are costly or often simply unavailable. Other works have thus explored semantically grounded approaches using cross-attention instead to better control image manipulation [7, 28, 30]. In contrast, LEDs++ leverages both attention-and noise-based masking to obtain fine-grained masks, enabling strong semantic control over real images.", + "bbox": [ + 503, + 121, + 888, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another important aspect of image manipulation methods is the required tuning and overall runtime. Instruct-Pix2Pix continues training a DM at scale to enable image editing capabilities [6]. Finetuning instead on each individual input to constrain the generation on the real image has shown helpful [18, 44] but not computationally efficient. Consequently, recent works have largely relied on inverting the deterministic DDIM sampling process [42] to save computational resources. DDIM inversion identifies an initial noise vector that results in the input image when denoised again. However, faithful reconstructions are only obtained in the limit of small steps, thus requiring large numbers of inversion steps. Moreover, small errors will still incur at each timestep, often accumulating into meaningful deviations from the input, requiring costly error correction through optimization [28, 30]. Recently, Huberman-Spiegelglas et al. proposed an inversion technique [17] for the DDPM sampler [15] to address the limitations of DDIM inversion. LEDs++ provides the same guarantees of perfect inversion with even further reduced runtime alongside an edit-friendly latent space, enabling more versatility.", + "bbox": [ + 503, + 348, + 888, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Image Editing with Text-to-Image Models", + "text_level": 1, + "bbox": [ + 503, + 678, + 870, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before devising the methodology of LEDs++, let us first motivate the desired features and use cases. Specifically, we aim for efficiency, versatility, and precision. The goal is to provide a method that enables a fast exploratory workflow for image editing in which a user can iteratively interact with the model and explore various edits. Consequently, LEDs++ produces outputs quickly with no tuning or optimization to not disrupt the creative process. Further, arbitrary editing instructions and combinations thereof are supported to facilitate a wide range of image manipulations (e.g., complex multi-editing). Lastly, we provide precise and sophisticated semantic control over the image editing. Each of the (potentially multiple) edit instruc", + "bbox": [ + 503, + 704, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8862", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions can be steered individually, and changes are automatically restricted to relevant image regions. Importantly, with LEDs++ we prioritize compositional robustness.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Guided Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 145, + 307, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us first define some general background for diffusion models. DMs iteratively denoise a Gaussian distributed variable to produce samples of a learned data distribution. Let's consider a diffusion process that gradually turns an image $x_0$ into Gaussian noise.", + "bbox": [ + 76, + 167, + 468, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} n _ {t}, \\quad t = 1, \\dots , T \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 253, + 468, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n_t$ are iid normal distributed vectors and $\\beta_t$ a variance schedule. The diffusion process is equivalently expressed as", + "bbox": [ + 76, + 281, + 468, + 311 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 320, + 468, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} = 1 - \\beta_{t}$ , $\\bar{\\alpha}_{t} = \\Pi_{s=1}^{t} \\alpha_{s}$ and $\\epsilon_{t} \\sim \\mathcal{N}(0, \\mathbf{I})$ . Importantly, all $\\epsilon_{t}$ are not statistically independent. Instead, consecutive pairs $\\epsilon_{t}, \\epsilon_{t-1}$ are strongly dependent, which will be relevant later. To generate an (new) image $\\hat{x}_{0}$ the reverse diffusion process starts from random noise $x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})$ which can be iteratively denoised as", + "bbox": [ + 76, + 347, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 448, + 468, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $z_{t}$ are iid standard normal vectors, and common variance schedulers $\\sigma_{t}$ can be expressed in the general form", + "bbox": [ + 76, + 474, + 468, + 503 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} = \\eta \\beta_ {t} \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 512, + 338, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta \\in [0,1]$ . In this formulation, $\\eta = 0$ corresponds to the deterministic DDIM [42] and $\\eta = 1$ to the DDPM scheme [15]. Lastly, in theses cases, we have $\\hat{\\mu}_t(x_t) =$", + "bbox": [ + 76, + 551, + 468, + 598 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\hat {\\epsilon} _ {\\theta} (x _ {t})\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 606, + 454, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\hat{\\epsilon}_{\\theta}(x_{t})$ is an estimate of $\\epsilon_{t}$ produced by our neural network DM with learned parameters $\\theta$ , commonly implemented as a U-Net [35]. For text-to-image generation, the model is conditioned on a text prompt $p$ to produce images faithful to that prompt. The DM is trained to produce the noise estimate $\\hat{\\epsilon}_{\\theta}(x_{t})$ needed for iteratively sampling $\\hat{x}_0$ (Eq. 3). For text-conditioned DMs, $\\hat{\\epsilon}_{\\theta}$ is calculated using specific guidance techniques.", + "bbox": [ + 76, + 648, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most DMs rely on classifier-free guidance [16], a conditioning method using a purely generative diffusion model, eliminating the need for an additional classifier. During training, the text conditioning $c_{p}$ is randomly dropped with a fixed probability, resulting in a joint model for unconditional and conditional objectives. During inference, the score estimates for the $\\epsilon$ -prediction are adjusted so that:", + "bbox": [ + 76, + 768, + 468, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + s _ {g} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 883, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with guidance scale $s_g$ and $\\hat{\\epsilon}_{\\theta}$ defining the noise estimate with parameters $\\theta$ . Intuitively, the unconditioned $\\epsilon$ -prediction is pushed in the direction of the conditioned one, with $s_g$ determining the extent of the adjustment.", + "bbox": [ + 498, + 90, + 890, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.LEDITS++", + "text_level": 1, + "bbox": [ + 500, + 160, + 612, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the fundamentals established, the methodology of LEDs++ can now be broken down into three components: (1) efficient image inversion, (2) versatile textual editing, and (3) semantic grounding of image changes.", + "bbox": [ + 498, + 183, + 890, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Component 1: Perfect Inversion. Utilizing text-to-image models for editing real images requires conditioning the generation on the input image. Recent works have largely relied on inverting the sampling process to identify $x_{T}$ that will be denoised to the input image $x_{0}$ [28, 30]. Inverting the DDPM scheduler is generally preferred over DDIM inversion since the former can be achieved in fewer timesteps and with no reconstruction error [17].", + "bbox": [ + 498, + 244, + 890, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, there exist more efficient schemes than DDPM for sampling DMs that greatly reduce the required number of steps and consequently DM evaluations. We here propose a more efficient inversion method by deriving the desired inversion properties for such a scheme. As demonstrated by Song et al.[43], DDPM can be viewed as a first-order stochastic differential equation (SDE) solver when formulating the reverse diffusion process as an SDE. This SDE can be solved more efficiently—in fewer steps—using a higher-order differential equation solver, here dpmsolver++ [23]. The reverse diffusion process from Eq. 3 for the second-order sde-dpm-solver++ can be written as", + "bbox": [ + 498, + 364, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, x _ {t + 1}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 556, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where now", + "bbox": [ + 500, + 584, + 576, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\sqrt {1 - e ^ {- 2 h _ {t - 1}}}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 604, + 803, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and higher-order $\\hat{\\mu}_t$ depends now on $x$ from two timesteps, $x_t$ and $x_{t + 1}$ . Such that $\\hat{\\mu}_t(x_t,x_{t + 1}) =$", + "bbox": [ + 498, + 635, + 888, + 666 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} e ^ {- h _ {t - 1}} x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\hat {\\epsilon} _ {\\theta} (x _ {t}) \\\\ + 0. 5 \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\frac {- h _ {t}}{h _ {t - 1}} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t + 1}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 675, + 872, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with", + "bbox": [ + 500, + 750, + 534, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t}})} - \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t + 1}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t + 1}})}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 761, + 823, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the detailed derivation of the solver and proof of faster convergence, we refer the reader to the relevant literature [22, 23]. Based on the above, we now devise our inversion process. Given an input image $x_0$ we construct an auxiliary reconstruction sequence of noise images $x_1, \\ldots, x_T$ as", + "bbox": [ + 498, + 799, + 890, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\tilde {\\varepsilon} _ {t} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 883, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8863", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\epsilon}_t\\sim \\mathcal{N}(0,\\mathbf{I})$ . Contrary to Eq. 2, the $\\tilde{\\epsilon}_t$ are now statistically independent, which is a desirable property for image editing [17]. Lastly, the respective $z_{t}$ for the inversion can be derived from Eq. 5 as", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\frac {x _ {t - 1} - \\hat {\\mu} _ {t} \\left(x _ {t} , x _ {t + 1}\\right)}{\\sigma_ {t}}, \\quad t = T,.., 1 \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 161, + 468, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\hat{\\mu}$ and $\\sigma_{t}$ as defined above. We base our implementation on the multistep variant of sde-dpm-solver++, which only requires one evaluation of the DM at each diffusion timestep by reusing the estimates from the previous step. The number of timesteps can be reduced further by stopping the inversion at an intermediate step $t < T$ and starting the generation at that step. Empirically, we observed that $t\\in [0.9T,0.8T]$ usually produces edits of the same fidelity as $t = T$ , supporting observations in previous work [17, 26] that earlier timesteps are less relevant to the edit.", + "bbox": [ + 75, + 202, + 468, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Component 2: Textual Editing. After creating our reconstruction sequence $x_{1}, \\ldots, x_{T}$ and calculating the respective $z_{t}$ , we now edit the image by manipulating the noise estimate $\\hat{\\epsilon}_{\\theta}$ based on a set of edit instructions $\\{e_i\\}_{i \\in I}$ . We devise a dedicated guidance term for each concept $e_{i}$ based on conditioned and unconditioned estimates. Let us formally define LEDs++'s guidance by starting with a single editing prompt $e$ . We compute", + "bbox": [ + 75, + 354, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + \\gamma \\left(x _ {t}, c _ {e}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 486, + 468, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with guidance term $\\gamma$ . Consequently, setting $\\gamma = 0$ will reconstruct the input image $x_0$ . We construct $\\gamma$ to push the unconditioned score estimate $\\hat{\\epsilon}_{\\theta}(x_t)$ —i.e. the input image reconstruction—away from/towards the edit concept estimate $\\hat{\\epsilon}_{\\theta}(x_t, c_e)$ , depending on the guidance direction:", + "bbox": [ + 75, + 513, + 468, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma \\left(x _ {t}, c _ {e}\\right) = \\phi (\\psi ; s _ {e}, \\lambda) \\psi \\left(x _ {t}, c _ {e}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 601, + 468, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ applies an edit guidance scale $s_e$ element-wise, and $\\psi$ depends on the edit direction: $\\psi(x_t, c_e) =$", + "bbox": [ + 76, + 628, + 468, + 659 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) \\quad \\text {i f p o s . g u i d a n c e} \\\\ - \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\quad \\text {i f n e g . g u i d a n c e} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 670, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, changing the guidance direction is reflected by the direction between $\\hat{\\epsilon}_{\\theta}(x_t,c_e)$ and $\\hat{\\epsilon}_{\\theta}(x_t)$ . The term $\\phi$ identifies those dimensions of the image and respective $\\hat{\\epsilon}_{\\theta}$ that are relevant to a prompt $e$ . Consequently, $\\phi$ returns 0 for all irrelevant dimensions and a scaling factor $s_e$ for the others. We describe the construction of $\\phi$ in detail below. Larger $s_e$ will increase the effect of the edit, and $\\lambda \\in (0,1)$ reflects the percentage of the pixels selected as relevant by $\\phi$ . Notably, for a single concept $e$ and uniform $\\phi = s_e$ , Eq. 8 generalizes to the classifier-free guidance term in Eq. 4.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For multiple $e_i$ , we calculate $\\gamma_t^i$ as described above, with each defining their own hyperparameter values $\\lambda^i$ , $s_e^i$ . The", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sum of all $\\gamma_t^i$ results in", + "bbox": [ + 500, + 90, + 653, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\gamma} _ {t} \\left(x _ {t}, c _ {e _ {i}}\\right) = \\sum_ {i \\in I} \\gamma_ {t} ^ {i} \\left(x _ {t}, c _ {e _ {i}}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 114, + 890, + 140 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Component 3: Semantic Grounding. The masking term $\\phi$ (Eq. 9) is the intersection (pointwise product) of binary masks $M^1$ and $M^2$ combined with scaling factor $s_e$ :", + "bbox": [ + 498, + 146, + 890, + 191 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\psi ; s _ {e _ {i}}, \\lambda) = s _ {e _ {i}} M _ {i} ^ {1} M _ {i} ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 200, + 890, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M_{i}^{1}$ is a binary mask generated from the U-Net's cross-attention layers and $M_{i}^{2}$ is a binary mask derived from the noise estimate. Intuitively, each mask is an importance map, where $M_{i}^{1}$ is more strongly grounded than $M_{i}^{2}$ , but of significantly coarser granularity. Therefore, the intersection of the two yields a mask both focused on relevant image regions and of fine granularity. With LEDs++, we empirically demonstrate that these maps can also capture regions of an image relevant to an editing concept that is not already present. Specifically for multiple edits, calculating a dedicated mask for each edit prompt ensures that the corresponding guidance terms remain largely isolated, limiting interference between them.", + "bbox": [ + 496, + 226, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, at each time step $t$ , a U-Net forward pass with editing prompt $e_i$ is performed to generate cross-attention maps for each token of the editing prompt. All cross-attention maps of the smallest resolution (e.g., $16 \\times 16$ for SD) are averaged over all heads and layers, and the resulting maps are summed over all editing tokens, resulting in a single map $A_t^{e_i} \\in R^{16 \\times 16}$ . Importantly, we utilize the same U-Net evaluation $\\hat{\\epsilon}_{\\theta}(x_t, c_e)$ already performed in Eq. 10 to produce $M^1$ with minimal overhead. Each map $A_t^{e_i}$ is up-sampled to match the size of $x_t$ . Cross-attention mask $M^1$ is derived by calculating the $\\lambda$ -th percentile of up-sampled $A_t^{e_i}$ and", + "bbox": [ + 496, + 422, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i} ^ {1} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left| A _ {t} ^ {e _ {i}} \\right| \\geq \\eta_ {\\lambda} \\left(\\left| A _ {t} ^ {e _ {i}} \\right|\\right) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 613, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta_{\\lambda}(|\\cdot|)$ is the $\\lambda$ -th percentile. By definition, $M^1$ only selects image regions that correlate strongly with the editing prompt, and $\\lambda$ determines the size of this selected region.", + "bbox": [ + 496, + 662, + 890, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The fine-grained mask $M^2$ is calculated based on the guidance vector $\\psi$ of noise estimates derived in Eq. 10. The difference between unconditioned and conditioned $\\hat{\\epsilon}_{\\theta}$ , generally captures outlines and object edges of $x_{t}$ . Consequently, the largest absolute values of $\\psi$ provide meaningful segmentation information of fine granularity for $M^2$", + "bbox": [ + 496, + 708, + 890, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM ^ {2} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\psi | \\geq \\eta_ {\\lambda} (| \\psi |) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 806, + 890, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, threshold $\\lambda$ should correspond to the performed edit. Changes affecting the entire image, such as style transfer, should choose smaller $\\lambda (\\rightarrow 0)$ , whereas edits targeting", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8864", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa8709af9cdd4c546356fdd95823f9d774fb44ad62023ae5dae09689d981550b.jpg", + "image_caption": [ + "Figure 2. Comparison of image editing methods. (top) LEDs++ is the only method to restrict edits to the tree leaves and position of the car. (bottom) Ours is the only approach faithfully executing all three edits and keeping changes minimal. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 161, + 88, + 810, + 265 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bd97ddbef80dd831bf1304807b0f0851b0a858012777a1cfac3695bf331144e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodReconstruction \nError (RMSE) ↓Execution \nTime (s) ↓Variation/ \nSamplingSemantic \nGroundingMulti- \nEditing
SDEdit [26]0.81 ±0.072.10 ±0.02✓XX
Imagic [18]0.58 ±0.12349.98 ±0.45✓XX
Vanilla DDIM Inversion0.22 ±0.1037.23 ±0.04XXX
Pix2Pix-Zero [30]0.20 ±0.0956.78 ±0.14(✓)✓X
DiffEdit [9]0.13±0.0327.65 ±0.03✓✓X
Edit-friendly DDPM [17]0.0010.36 ±0.05✓XX
LEDSITS++ (Ours)0.001.78 ±0.03✓✓✓
", + "bbox": [ + 161, + 311, + 810, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparing key properties for diffusion-based image editing techniques, with LEDs++ offering clear methodological benefits. Due to LEDs++'s efficient perfect inversion, it is the fastest and error-free method. At the same time, its methodology is the only enabling versatility in terms of variation, semantic grounding, and multi-editing. Subscript numbers indicate standard deviation.", + "bbox": [ + 75, + 446, + 892, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "specific objects or regions should use $\\lambda$ proportional to the region's prominence in the image.", + "bbox": [ + 75, + 508, + 468, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Properties of LEDs++", + "text_level": 1, + "bbox": [ + 76, + 556, + 307, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the fundamentals of LEDs++ established, we next showcase its unique properties and capabilities.", + "bbox": [ + 75, + 580, + 468, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency. First off, LEDs++ offers substantial performance improvements over other image editing methods. In Tab. 1, we provide a qualitative runtime comparison, with all methods being implemented for Stable Diffusion (SD) 1.5 [34]. As a parameter-free approach, LEDs++ does not require any computationally expensive fine-tuning or optimization. Consequently, LEDs++ is orders of magnitude faster than methods like Imagic [18] or Pix2Pix-Zero [30]. Further, we only need to invert the same number of diffusion steps used at inference, which results in significant runtime improvements over the standard DDIM inversion (21x). In addition to efficient inversion, we use a recent, fast scheduler that generally requires fewer total steps, further boosting performance. This way, LEDs++ is six times faster than recent DDPM inversion [17] and on par with fast but poor-quality SDEdit [26].", + "bbox": [ + 75, + 613, + 470, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Versatility. In addition to its efficiency, LEDs++ remains versatile, enabling sheer limitless editing possibilities. In Fig. 1, we showcase a broad range of edit types.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LEDITS++ facilitates fine-grained edits (adding/removing glasses) and holistic changes such as style transfer (painting/sketch). Furthermore, object removal and replacement facilitate even more image editing tasks. Importantly, the overall image composition is preserved in all cases. To our knowledge, LEDs++ is the only diffusion-based image editing method inherently supporting multiple edits in isolation, which allows for more complex image manipulation. Fig. 2 highlights LEDs++ benefits over previous methods. Our method produces the highest edit fidelity and is the only approach capable of faithfully executing multiple, simultaneous instructions. Moreover, LEDs++ also makes the least changes to unrelated objects and the overall background and composition of the image.", + "bbox": [ + 496, + 508, + 890, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lastly, the editing versatility benefits from the stochastic nature of the perfect but non-deterministic inversion. LEDs++ provides meaningful image variations by resampling $\\tilde{c}_t$ . Additionally, the visual expression of each concept in the edited image scales monotonically with $s_e$ , and the direction and magnitude of each concept can be varied freely. We present examples of both features in App. B.", + "bbox": [ + 496, + 727, + 890, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Precision. Furthermore, LEDs++'s methodology keeps edits concise and avoids unnecessary deviations from the input image (Fig. 2). First, the perfect inversion will reconstruct the exact input image if no edit is applied (cf.", + "bbox": [ + 496, + 839, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8865", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65145febf9dc8f5f3ea361c981b25ba7f5ed9a8b34e59e22ec6b15377923bcfe.jpg", + "image_caption": [ + "Figure 3. Exemplary edit performed with LEDs++ in only 25 diffusion steps with SD1.5. We apply a complex, compounded edit and ground each to a semantically reasonable image region." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 429, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sec. 3.2). Consequently, we already improve on faithfulness to the input image even before applying any edits. This benefit over other methods is highlighted by the reconstruction error in Tab. 1. Second, implicit masking will semantically ground each edit to relevant image regions. This is specifically important for editing multiple concepts at the same time. While other methods only utilize one prompt for all instructions, LEDs++ isolates edits from each other (Eq. 11). Thus, we get dedicated masks for each concept as shown in Fig. 3. This design ensures that each instruction (e.g., red mask for 'cherry blossom') will be only applied where necessary. Subsequently, we provide further evidence for the efficacy of LEDs++'s masking approach.", + "bbox": [ + 75, + 252, + 468, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Semantically Grounded Image Editing", + "text_level": 1, + "bbox": [ + 76, + 465, + 423, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-attention maps of DMs have been used extensively to ground regions of interest during image generation semantically [7, 8, 14, 30]. Nonetheless, these have not been combined with noise-based masks so far and thus lack fine granularity. Hence, we empirically evaluate the quality of implicit masks, i.e., attention maps $M^1$ and noise maps $M^2$ (Eq. 13 and 14) in the LEDs++ setup. We use a broad segmentation task for common objects as a proxy to measure the performance of implicit masks in identifying relevant image areas from edit instructions. Specifically, we utilize segmentation masks from the COCO panoptic segmentation challenge [19]. For each unique object in an image, we retrieve the masks $M^1$ , $M^2$ , and their intersection per diffusion step. We use the (semantic) class label (e.g. 'person' or 'TV') as editing concept $e$ . We consider masks at each of 50 total diffusion steps without actually editing the input image. Furthermore, we approximate mask threshold $\\lambda$ based on the relative size of an object's bounding box.", + "bbox": [ + 75, + 491, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Concise masks with LEDs++. Fig. 4 shows implicit masking as a reliable means to identify relevant image regions. Importantly, the intersection of both cross-attention masks $M^1$ and noise maps $M^2$ clearly outperforms each separate mask. The overall performance is even similar to a dedicated CLIPSeg model [24], despite LEDs++. masks being implicitly calculated at inference with only minimal overhead. At the same time, LEDs++.'s masking is superior to DiffEdit's [3]. Consequently, our method's inter", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/318c6382a39c079543b3cc4819e80b0f71060178d28b71ea7d436599a39f18cc.jpg", + "image_caption": [ + "Figure 4. Semantic segmentation quality of LEDs++. We show the intersection over union (higher is better) for COCO panoptic segmentation. The intersection masks outperform each by a clear margin, close to the CLIPSeg reference. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 524, + 92, + 867, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "section of cross-attention masks and noise maps provides strong semantic grounding while being efficient during image manipulation to ensure precise editing.", + "bbox": [ + 498, + 306, + 890, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Image Editing Evaluation", + "text_level": 1, + "bbox": [ + 500, + 369, + 741, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let us now compare LEDs++ to current SOTA methods for image manipulation on two benchmarks.", + "bbox": [ + 498, + 395, + 890, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Editing Multiple Concepts", + "text_level": 1, + "bbox": [ + 500, + 436, + 738, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we investigate the complex task of performing multiple edits simultaneously. We rely on a well-established setup for semantic image manipulation [5] to evaluate multi-conditioned attribute manipulation in facial images. In our experiment, we consider 100 images from the CelebA dataset [21]. For each image, we simultaneously edit three attributes out of a set of five, leading to ten total combinations of edit concepts. Further, we perform each edit across ten different seeds, resulting in 10,000 evaluated images for each method and hyperparameter setting, over 1M images in total. As measures for comparison, we employ CLIP and LPIPS scores. CLIP measures the text-to-image similarity of the edit instruction to the edited image, and LPIPS measures the image-to-image similarity of the real to the edited image. This way, we assess the tradeoff between the versatility of edits (CLIP) and the precision of those manipulations (LPIPS). We implement all methods based on SD1.5 and provide more details in App. C.", + "bbox": [ + 496, + 460, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LEDS++ outperforms competing methods. Fig. 5 shows the resulting CLIP vs. LPIPS plots for all methods. The top left corner represents the ideal editing method with maximum edit alignment without deviating from the initial image. Generally, one can observe a natural trade-off between versatility and precision for all methods, i.e., higher image-to-text alignment comes at the expense of lower similarity to the original image. LEDS++ is closest to the ideal region and thus clearly outperforms the other methods. In particular, the outputs remain close to the original image (low LPIPS scores), thanks to the precise implicit masking.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8866", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/110ffe195968c3d76f23e8c672ce8b00d27b845da1c2f8bcb04eed0058e662e0.jpg", + "image_caption": [ + "Figure 5. Comparison of instruction-alignment vs. image similarity trade-off for different editing methods. Results were reported for simultaneous manipulation of three facial attributes on CelebA. We plot CLIP scores (higher is better) of the target attributes against LPIPS similarity (lower is better). LEDs++ clearly outperforms all competing methods. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 444, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At the same time, it faithfully performs the edits (high CLIP scores) due to the dedicated, isolated editing for each concept. The depicted scores reflect our qualitative inspections for Pix2Pix-Zero and Imagic on such complex manipulations (cf. Fig. 2). We observed that these methods often break—either failing to perform all three edits and/or drastically altering the input image. Only edit-friendly DDPM [17] and LEDs++ reliably achieve the maximum average CLIP score of over 0.25. This value seems to represent an upper bound according to our manual investigations, as each attribute is edited correctly for all input images, and higher scores are not observed. Despite being computationally very efficient, LEDs++ faithfully executes each edit instruction while keeping the changes to the input low, highlighting the method's versatility and precision.", + "bbox": [ + 75, + 372, + 472, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. TEDBench $(++)$", + "text_level": 1, + "bbox": [ + 76, + 609, + 227, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we investigate the versatility of LEDs++'s editing capabilities by running the Textual Editing Benchmark (TEDBench [18]), a collection of 100 input images paired with textual edit instructions. However, we observed a variety of inconsistencies in TEDBench and a lack of relevant editing tasks. Therefore, we propose TEDBench++ (Fig. 6a and App. D), a more challenging revised benchmark now containing 120 entries in total.3 We addressed misspellings and rephrased ambiguous and inconclusive instructions. In addition to resolving these issues, we added instructions targeting challenging types of image manipulations previously not included in TEDBench: multi-conditioning, object/concept removal, style transfer, and complex replacements (Fig. 6a). We provide more details in App. D.", + "bbox": [ + 75, + 633, + 468, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare LEDs++ on TEdBench $(++)$ to one of the strongest editing methods, Imagic [18] with Imagen [37].", + "bbox": [ + 76, + 845, + 468, + 875 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/636c1e41362e13cfd1e413f996541c24e2fed7f5ab4e7a5f5d5480a83c80f5c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TEDBenchTEDBench++
SR ↑LPIPS ↓SR ↑LPIPS ↓
Imagic w/ SD1.50.550.560.580.57
LEDITS++ w/ SD1.50.750.280.790.30
Imagic w/Imagen [18]0.830.59——
LEDITS++ w/ SD-XL0.840.330.870.34
", + "bbox": [ + 513, + 88, + 877, + 174 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Success rate (SR) and LPIPS scores on the original TedBench [18] and our revised version (TEDBench++) We compare Imagic to LEDs++ based on different DMs and find the latter to outperform on both metrics and benchmarks.", + "bbox": [ + 498, + 180, + 890, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since both are not publicly available, we can only compare to this specific combination of DM and editing method using Kawar et al.'s [18] curated outputs for TEdBench. Additionally, we, therefore, cannot combine LEDs++ with Imagen[37] and instead use a similarly advanced diffusion model, SD-XL [32]. However, to not only compete for the best fidelity outputs but focus the evaluation on methodological differences—not the pre-trained DM—we also compare both methods implemented with SD1.5. We provide further details in App. C.", + "bbox": [ + 496, + 247, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LEDS $^{+ + }$ edits images reliably. We first asked users to assess the overall success of edits, i.e., if an edit instruction was faithfully realized for a given input image. The results in Tab. 2 show that LEDs $^{+ + }$ outperforms Imagc on TEdBench despite a greatly reduced runtime (Tab. 1). The difference is even stronger when comparing both methods on the same pre-trained DM, i.e., SD1.5. The high success rate on TEdBench $^{+ + }$ $(87\\%)$ and the examples shown in Fig. 1 and 6a once again highlight LEDs $^{+ + }$ 's versatility. Overall, our proposed method can reliably perform a diverse set of editing instructions for real images.", + "bbox": [ + 496, + 400, + 892, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-quality edits with LEDs++. While investigating both methods' performance we observed a substantial difference in edit quality. The examples in Fig. 6b particularly highlight the discrepancy in compositional robustness and object coherence. Hence, we also assessed both methods' editing quality on TEdBench++. We focus on samples where both methods performed a successful edit, i.e., were labeled as successful by users. We show the perceptual similarity (LPIPS) to the input image in Tab. 2. One can observe that the LPIPS scores for LEDs++ are much lower than for Imagic, empirically supporting the qualitative examples in Fig. 6b. When manually inspecting the generated images, we often found Imagic to generate a completely new image based on the edit instruction, entirely disregarding the input image (cf. App. Fig. 11a).", + "bbox": [ + 496, + 568, + 890, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 500, + 811, + 612, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Let us now discuss open research questions and limitations.", + "bbox": [ + 500, + 839, + 888, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Dependency. While LEDs++ achieves impressive results on a large variety of image manipulation tasks, there are external factors to consider. Since the", + "bbox": [ + 500, + 854, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3 https://huggingface.co/datasets/AIML-TUDA/TEdBench_plusplus", + "bbox": [ + 94, + 886, + 393, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8867", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be4dfb3e9df3f04561b13dffc758c1f1e4f3c2a54aa97cbf03fd80af2b3c79db.jpg", + "image_caption": [ + "(a) Novel challenging examples of TEdBench++ and LEDs++ applied, showcasing the versatility of supported edits." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 524, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/80bbbd9c3d251bbf51c623aac6607686525a4f1e1fe8d544eeb4757123cb27e6.jpg", + "image_caption": [ + "(b) Qualitative comparison of LEDs++ and Imagic on TEd-Bench, clearly highlighting the performance improvement.", + "Figure 6. Benchmark examples for LEDs++ and Imagc on TEdBench $(++)$ . (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 566, + 88, + 890, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method is architecture-agnostic, it can be easily used with any DM. At the same time, the general editing quality strongly depends on the overall capabilities of the underlying pre-trained DM. Naturally, more capable models will also enable better edits. But, at times, specific editing instructions may fail because the used DM does not have a decent representation of the targeted concept to begin with. One example is the model failing to edit a giraffe to be sitting since the underlying DM generally fails to generate this pose (cf. App. F). This effect can also clearly be seen in Tab. 2, with the editing success rate of a method varying strongly between DMs. Although the same image editing method is employed (LEDITS++, the more capable SD-XL variant outperforms the weaker SD1.5 model. Nonetheless, this means that the architecture-agnostic LEDs++ will benefit from increasingly powerful DMs.", + "bbox": [ + 75, + 352, + 472, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coherence Trade-offs. Next to the benefits of LEDs++'s semantic grounding, there are also downsides to this approach. Overall, implicit masking limits changes to relevant portions of the image and achieves strong coherence with the original image composition. Yet, the object and its identity within the masked area may change based on various factors. Generic prompts, like \"a standing cat\" (cf. App. F), do not contain detailed information about this specific object (\"cat\"). Thus, an edit with this prompt does not guarantee to preserve object identity, particularly for strong hyperparameters. We observed that fine-tuning approaches like Imagic make the opposite trade-off, better preserving the object identity while changing the background and image composition substantially (cf. App. F). A potential remedy for a loss in object coherence with LEDs++, is more descriptive edit prompting, e.g. using textual inversion [12].", + "bbox": [ + 75, + 595, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, the automatically-inferred implicit masks allow for easy use of LEDs++ without users tediously providing masks. Nonetheless, user intentions are diverse and cannot always be automatically inferred. Sometimes, in-", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "individual user masks provide better control over the editing process. Such user masks can be easily integrated into LEDs++ (cf. App. F), wherefore we encourage future research in this promising direction.", + "bbox": [ + 496, + 352, + 890, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Societal Impact. LEDs++ is an easy-to-use image editing technique that lowers the barrier for users and puts them in control for fruitful human-machine collaboration. Yet, the underlying text-to-image models offer both promise and peril, as highlighted by prior research [4, 11]. The (societal) biases within these models will also impact image editing applications [11]. Moreover, image manipulation can also be used adversarially to generate inappropriate [39] or fake content. Hence, we advocate for a cautious deployment of generative models together with image editing methods.", + "bbox": [ + 496, + 414, + 892, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 583, + 617, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced LEDs++, an efficient yet versatile and precise method for textual image manipulation with diffusion models. It facilitates the editing of complex concepts in real images. Our approach requires no finetuning nor optimization, can be computed extremely efficiently, and is architecture agnostic. At the same time, it perfectly reconstructs an input image and uses implicit masking to limit changes to relevant image regions, thus editing precisely. Our large experimental evaluation confirms the efficiency, versatility, and precision of LEDs++ and its components, as well as its benefits over several related methods.", + "bbox": [ + 496, + 609, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We gratefully acknowledge support from the BMBF (Grant No 01IS22091). This work benefited from the ICT-48 Network of AI Research Excellence Center \"TAILOR\" (EU Horizon 2020, GA No 952215), the Hessian research priority program LOEWE within the project WhiteBox, the HMWK cluster projects \"Adaptive Mind\" and \"Third Wave of AI\", and from the NHR4CES.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8868", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. SIGGRAPH Asia, 2023. 16", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv:2206.02779, 2022. 2", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv:2211.01324, 2022. 2, 6", + "[4] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. In Proceedings of ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2023. 8, 11", + "[5] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. In Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 6", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv:2304.08465, 2023. 2, 6", + "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42, 2023. 2, 6", + "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5", + "[10] Zoe De Simone, Angie Boggust, Arvind Satyanarayan, and Ashia Wilson. What is a Fair Diffusion Model? Designing Generative Text-To-Image Models to Incorporate Various Worldviews. arXiv preprint arXiv:2309.09944, 2023. 11", + "[11] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness. arXiv preprint arXiv:2302.10893, 2023. 8, 11", + "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8, 16", + "[13] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion" + ], + "bbox": [ + 78, + 114, + 470, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 11", + "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 6", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and HsuanTien Lin, editors, Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 3", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 3", + "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 1, 2, 3, 4, 5, 7, 12, 13", + "[18] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 7, 12, 13, 14", + "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 6", + "[20] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2", + "[21] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), December 2015. 6", + "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 2, 3", + "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3", + "[24] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6", + "[25] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv:2310.04378, 2023. 2", + "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 1, 2, 4, 5, 12, 13", + "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P." + ], + "bbox": [ + 503, + 92, + 893, + 890 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8869", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 12", + "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2022. 2, 11", + "[30] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 12, 13", + "[31] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv:2306.00637, 2023. 2", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 7", + "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2, 11", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5", + "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 3", + "[36] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V. Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased Face Synthesis With Diffusion Models: Are We There Yet? arXiv preprint arXiv:2309.07277, 2023. 11", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv:2205.11487, 2022. 2, 7, 11", + "[38] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. arXiv:2304.14530, 2023. 17", + "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8, 11" + ], + "bbox": [ + 78, + 90, + 470, + 893 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of NeurIPS Datasets and Benchmarks, 2022. 11", + "[41] Norbert Schwarz. How the questions shape the answers. American Psychologist, 1999. 15", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 3", + "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3", + "[44] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv:2210.09477, 2022. 1, 2", + "[45] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv:1212.08698, 2022. 2", + "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv:2302.04867, 2023.2" + ], + "bbox": [ + 501, + 92, + 890, + 512 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8870", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_model.json b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72e8617922dc28f623e1865d6684c2ba6181fe46 --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "LEDS: Limitless Image Editing using Text-to-Image Models" + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.179, + 0.846, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.264, + 0.832, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.445, + 0.8, + 0.46 + ], + "angle": 0, + "content": "Figure 1. LEDs++ facilitates versatile image-to-image editing. Several complex cases are available now." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.471, + 0.314, + 0.488 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.474, + 0.81 + ], + "angle": 0, + "content": "Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDs++, an efficient yet versatile and precise textual image manipulation technique. LEDs++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDs++ and its improvements over previous methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.843, + 0.21, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Text-to-image diffusion models (DM) have garnered recognition for their ability to generate high-quality images from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.684 + ], + "angle": 0, + "content": "textual descriptions. A growing body of research has recently been dedicated to utilizing these models for manipulating real images. However, several barriers prevent many real-world applications of diffusion-based image editing. Current methods often entail computationally expensive model tuning or other optimization, presenting practical challenges [6, 18, 28, 30, 44]. Additionally, existing techniques have the proclivity to induce profound changes to the original image [17, 26], often resulting in completely different images. Lastly, all these approaches are inherently constrained when editing multiple (arbitrary) concepts simultaneously. We tackle these problems by introducing LEDs++\\(^{1}\\), a diffusion-based image editing technique addressing these limitations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.893, + 0.852 + ], + "angle": 0, + "content": "LEDITS++2 offers a streamlined approach for textual image editing, eliminating the need for extensive parameter tuning. To this end, we derive image inversion for a more efficient diffusion sampling algorithm to a) drastically reduce computational resources and b) guarantee perfect image reconstruction. Thus, we overcome computational obstacles and avoid changes in the edited image in the first place. Furthermore, we use a novel implicit masking approach to semantically ground each edit instruction to its relevant image region. This further optimizes changes to the image by retaining the overall image composition and object identity." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.863, + 0.858, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution †Partially as research intern at Adobe" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.876, + 0.852, + 0.888 + ], + "angle": 0, + "content": "1LEDITS++ stands for Limitless Edits with sde-dpm-solver++." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.888, + 0.884, + 0.9 + ], + "angle": 0, + "content": "2https://huggingface.co/spaces/leditsplusplus/project" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.863, + 0.884, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8861" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "Additionally, LEDs++ is the only method to date to facilitate easy and versatile image editing by supporting multiple simultaneous instructions without causing undue interference. Finally, its lightweight architecture-agnostic nature ensures compatibility with both latent and pixel-based diffusion models, providing high accessibility." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.183, + 0.468, + 0.362 + ], + "angle": 0, + "content": "In this work, we establish the methodical benefits of LEDs++ and demonstrate that this intuitive, lightweight approach offers sophisticated semantic control for image editing. Specifically, we contribute by (i) devising a formal definition of LEDs++ while (ii) deriving perfect inversion for a more efficient diffusion sampling method, (iii) qualitatively and empirically demonstrating its efficiency, versatility, and precision, (iv) providing an exhaustive empirical comparison to concurrent works with automatic and human user metrics, and thereby (v) introducing Textual Editing Benchmark++ (TEdBench++, a more holistic and coherent testbed for evaluating textual image manipulation." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.377, + 0.204, + 0.394 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.403, + 0.468, + 0.477 + ], + "angle": 0, + "content": "Recently, large-scale, text-guided DMs have enabled versatile applications in image generation [3, 33, 37]. Especially latent diffusion models [31, 34] have gained attention for their computational efficiency. Below, we discuss related work for efficient, versatile image manipulation with DMs." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.479, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Diffusion Sampling. Generating outputs with DMs requires multiple iterative denoising steps that constitute the main bottleneck at inference. Commonly used sampling methods such as DDPM [15] or DDIM [42] require tens or hundreds of steps to produce high-quality samples. Consequently, numerous works have been dedicated to speeding up the sampling process without loss in quality. Distillation efforts progressively reduce the number of required steps through further training [25, 27]. Other works focus on improving the sampling itself, e.g. using high-order ODE-solvers [22, 23, 46]. Such solvers can be readily combined with pre-trained DMs at inference to lower the number of denoising steps. With LEDs++, we derive perfect image inversion with the DPM-Solver++, allowing image editing in as few as 20 total steps." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Semantic Control during Diffusion. While text-to-image DMs generate new, astonishing images, fine-grained control over the generative process remains challenging. Minor changes to the text prompt lead to entirely different outputs. Wu et al. [45] studied concept disentanglement using linear combinations of text embeddings to gain semantic control. Methods like Prompt-to-Prompt [14] and other works [8, 30] utilize the DM's attention layers to attribute pixels to tokens from the text prompt. Dedicated operations on the attention maps enable more control over the generated images. Other works have focused on the noise estimates of DMs [5, 20] providing semantic control over the generation process. With LEDs++, we now enable" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "fine-grained semantic control for manipulating real images, going beyond purely generative applications." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.122, + 0.89, + 0.348 + ], + "angle": 0, + "content": "Real Image Editing. Since DMs' rise in popularity for text-to-image generation, they have also been explored for (real) image-to-image editing. As a first, simple approach, SDEdit added noise to the image for an intermediate step in the diffusion process [26]. While lightweight, the resulting image diverges substantially from the input as it is (partially) regenerated. Inpainting allows to keep the change small by having a user provide additional masks to restrict changes to certain image regions [2, 29]. Yet, user masks are costly or often simply unavailable. Other works have thus explored semantically grounded approaches using cross-attention instead to better control image manipulation [7, 28, 30]. In contrast, LEDs++ leverages both attention-and noise-based masking to obtain fine-grained masks, enabling strong semantic control over real images." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.349, + 0.89, + 0.665 + ], + "angle": 0, + "content": "Another important aspect of image manipulation methods is the required tuning and overall runtime. Instruct-Pix2Pix continues training a DM at scale to enable image editing capabilities [6]. Finetuning instead on each individual input to constrain the generation on the real image has shown helpful [18, 44] but not computationally efficient. Consequently, recent works have largely relied on inverting the deterministic DDIM sampling process [42] to save computational resources. DDIM inversion identifies an initial noise vector that results in the input image when denoised again. However, faithful reconstructions are only obtained in the limit of small steps, thus requiring large numbers of inversion steps. Moreover, small errors will still incur at each timestep, often accumulating into meaningful deviations from the input, requiring costly error correction through optimization [28, 30]. Recently, Huberman-Spiegelglas et al. proposed an inversion technique [17] for the DDPM sampler [15] to address the limitations of DDIM inversion. LEDs++ provides the same guarantees of perfect inversion with even further reduced runtime alongside an edit-friendly latent space, enabling more versatility." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.679, + 0.872, + 0.696 + ], + "angle": 0, + "content": "3. Image Editing with Text-to-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.705, + 0.89, + 0.9 + ], + "angle": 0, + "content": "Before devising the methodology of LEDs++, let us first motivate the desired features and use cases. Specifically, we aim for efficiency, versatility, and precision. The goal is to provide a method that enables a fast exploratory workflow for image editing in which a user can iteratively interact with the model and explore various edits. Consequently, LEDs++ produces outputs quickly with no tuning or optimization to not disrupt the creative process. Further, arbitrary editing instructions and combinations thereof are supported to facilitate a wide range of image manipulations (e.g., complex multi-editing). Lastly, we provide precise and sophisticated semantic control over the image editing. Each of the (potentially multiple) edit instruc" + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8862" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "tions can be steered individually, and changes are automatically restricted to relevant image regions. Importantly, with LEDs++ we prioritize compositional robustness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.146, + 0.308, + 0.16 + ], + "angle": 0, + "content": "3.1. Guided Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.169, + 0.469, + 0.244 + ], + "angle": 0, + "content": "Let us first define some general background for diffusion models. DMs iteratively denoise a Gaussian distributed variable to produce samples of a learned data distribution. Let's consider a diffusion process that gradually turns an image \\( x_0 \\) into Gaussian noise." + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.254, + 0.469, + 0.273 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} n _ {t}, \\quad t = 1, \\dots , T \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.469, + 0.313 + ], + "angle": 0, + "content": "where \\( n_t \\) are iid normal distributed vectors and \\( \\beta_t \\) a variance schedule. The diffusion process is equivalently expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.321, + 0.469, + 0.339 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.469, + 0.438 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} = 1 - \\beta_{t}\\), \\(\\bar{\\alpha}_{t} = \\Pi_{s=1}^{t} \\alpha_{s}\\) and \\(\\epsilon_{t} \\sim \\mathcal{N}(0, \\mathbf{I})\\). Importantly, all \\(\\epsilon_{t}\\) are not statistically independent. Instead, consecutive pairs \\(\\epsilon_{t}, \\epsilon_{t-1}\\) are strongly dependent, which will be relevant later. To generate an (new) image \\(\\hat{x}_{0}\\) the reverse diffusion process starts from random noise \\(x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})\\) which can be iteratively denoised as" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.449, + 0.469, + 0.465 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.469, + 0.505 + ], + "angle": 0, + "content": "Here \\( z_{t} \\) are iid standard normal vectors, and common variance schedulers \\( \\sigma_{t} \\) can be expressed in the general form" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.513, + 0.339, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} = \\eta \\beta_ {t} \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.552, + 0.469, + 0.599 + ], + "angle": 0, + "content": "where \\(\\eta \\in [0,1]\\). In this formulation, \\(\\eta = 0\\) corresponds to the deterministic DDIM [42] and \\(\\eta = 1\\) to the DDPM scheme [15]. Lastly, in theses cases, we have \\(\\hat{\\mu}_t(x_t) =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.607, + 0.455, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\hat {\\epsilon} _ {\\theta} (x _ {t})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.649, + 0.469, + 0.77 + ], + "angle": 0, + "content": "Here \\(\\hat{\\epsilon}_{\\theta}(x_{t})\\) is an estimate of \\(\\epsilon_{t}\\) produced by our neural network DM with learned parameters \\(\\theta\\), commonly implemented as a U-Net [35]. For text-to-image generation, the model is conditioned on a text prompt \\(p\\) to produce images faithful to that prompt. The DM is trained to produce the noise estimate \\(\\hat{\\epsilon}_{\\theta}(x_{t})\\) needed for iteratively sampling \\(\\hat{x}_0\\) (Eq. 3). For text-conditioned DMs, \\(\\hat{\\epsilon}_{\\theta}\\) is calculated using specific guidance techniques." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.469, + 0.876 + ], + "angle": 0, + "content": "Most DMs rely on classifier-free guidance [16], a conditioning method using a purely generative diffusion model, eliminating the need for an additional classifier. During training, the text conditioning \\( c_{p} \\) is randomly dropped with a fixed probability, resulting in a joint model for unconditional and conditional objectives. During inference, the score estimates for the \\( \\epsilon \\)-prediction are adjusted so that:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.885, + 0.469, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + s _ {g} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "with guidance scale \\( s_g \\) and \\( \\hat{\\epsilon}_{\\theta} \\) defining the noise estimate with parameters \\( \\theta \\). Intuitively, the unconditioned \\( \\epsilon \\)-prediction is pushed in the direction of the conditioned one, with \\( s_g \\) determining the extent of the adjustment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.161, + 0.613, + 0.175 + ], + "angle": 0, + "content": "3.2.LEDITS++" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.244 + ], + "angle": 0, + "content": "With the fundamentals established, the methodology of LEDs++ can now be broken down into three components: (1) efficient image inversion, (2) versatile textual editing, and (3) semantic grounding of image changes." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.245, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Component 1: Perfect Inversion. Utilizing text-to-image models for editing real images requires conditioning the generation on the input image. Recent works have largely relied on inverting the sampling process to identify \\( x_{T} \\) that will be denoised to the input image \\( x_{0} \\) [28, 30]. Inverting the DDPM scheduler is generally preferred over DDIM inversion since the former can be achieved in fewer timesteps and with no reconstruction error [17]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.547 + ], + "angle": 0, + "content": "However, there exist more efficient schemes than DDPM for sampling DMs that greatly reduce the required number of steps and consequently DM evaluations. We here propose a more efficient inversion method by deriving the desired inversion properties for such a scheme. As demonstrated by Song et al.[43], DDPM can be viewed as a first-order stochastic differential equation (SDE) solver when formulating the reverse diffusion process as an SDE. This SDE can be solved more efficiently—in fewer steps—using a higher-order differential equation solver, here dpmsolver++ [23]. The reverse diffusion process from Eq. 3 for the second-order sde-dpm-solver++ can be written as" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.558, + 0.891, + 0.573 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, x _ {t + 1}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.585, + 0.578, + 0.597 + ], + "angle": 0, + "content": "where now" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.606, + 0.805, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\sqrt {1 - e ^ {- 2 h _ {t - 1}}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.636, + 0.89, + 0.667 + ], + "angle": 0, + "content": "and higher-order \\(\\hat{\\mu}_t\\) depends now on \\(x\\) from two timesteps, \\(x_t\\) and \\(x_{t + 1}\\). Such that \\(\\hat{\\mu}_t(x_t,x_{t + 1}) =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.676, + 0.873, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} e ^ {- h _ {t - 1}} x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\hat {\\epsilon} _ {\\theta} (x _ {t}) \\\\ + 0. 5 \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\frac {- h _ {t}}{h _ {t - 1}} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t + 1}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.751, + 0.535, + 0.764 + ], + "angle": 0, + "content": "with" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.762, + 0.824, + 0.796 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t}})} - \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t + 1}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t + 1}})}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.8, + 0.892, + 0.875 + ], + "angle": 0, + "content": "For the detailed derivation of the solver and proof of faster convergence, we refer the reader to the relevant literature [22, 23]. Based on the above, we now devise our inversion process. Given an input image \\( x_0 \\) we construct an auxiliary reconstruction sequence of noise images \\( x_1, \\ldots, x_T \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.885, + 0.891, + 0.901 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\tilde {\\varepsilon} _ {t} \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8863" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "where \\(\\tilde{\\epsilon}_t\\sim \\mathcal{N}(0,\\mathbf{I})\\). Contrary to Eq. 2, the \\(\\tilde{\\epsilon}_t\\) are now statistically independent, which is a desirable property for image editing [17]. Lastly, the respective \\(z_{t}\\) for the inversion can be derived from Eq. 5 as" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.162, + 0.469, + 0.194 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\frac {x _ {t - 1} - \\hat {\\mu} _ {t} \\left(x _ {t} , x _ {t + 1}\\right)}{\\sigma_ {t}}, \\quad t = T,.., 1 \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.203, + 0.469, + 0.354 + ], + "angle": 0, + "content": "with \\(\\hat{\\mu}\\) and \\(\\sigma_{t}\\) as defined above. We base our implementation on the multistep variant of sde-dpm-solver++, which only requires one evaluation of the DM at each diffusion timestep by reusing the estimates from the previous step. The number of timesteps can be reduced further by stopping the inversion at an intermediate step \\(t < T\\) and starting the generation at that step. Empirically, we observed that \\(t\\in [0.9T,0.8T]\\) usually produces edits of the same fidelity as \\(t = T\\), supporting observations in previous work [17, 26] that earlier timesteps are less relevant to the edit." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.469, + 0.476 + ], + "angle": 0, + "content": "Component 2: Textual Editing. After creating our reconstruction sequence \\( x_{1}, \\ldots, x_{T} \\) and calculating the respective \\( z_{t} \\), we now edit the image by manipulating the noise estimate \\( \\hat{\\epsilon}_{\\theta} \\) based on a set of edit instructions \\( \\{e_i\\}_{i \\in I} \\). We devise a dedicated guidance term for each concept \\( e_{i} \\) based on conditioned and unconditioned estimates. Let us formally define LEDs++'s guidance by starting with a single editing prompt \\( e \\). We compute" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.487, + 0.469, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + \\gamma \\left(x _ {t}, c _ {e}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.59 + ], + "angle": 0, + "content": "with guidance term \\(\\gamma\\). Consequently, setting \\(\\gamma = 0\\) will reconstruct the input image \\(x_0\\). We construct \\(\\gamma\\) to push the unconditioned score estimate \\(\\hat{\\epsilon}_{\\theta}(x_t)\\) —i.e. the input image reconstruction—away from/towards the edit concept estimate \\(\\hat{\\epsilon}_{\\theta}(x_t, c_e)\\), depending on the guidance direction:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.602, + 0.469, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\gamma \\left(x _ {t}, c _ {e}\\right) = \\phi (\\psi ; s _ {e}, \\lambda) \\psi \\left(x _ {t}, c _ {e}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.469, + 0.66 + ], + "angle": 0, + "content": "where \\(\\phi\\) applies an edit guidance scale \\(s_e\\) element-wise, and \\(\\psi\\) depends on the edit direction: \\(\\psi(x_t, c_e) =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.671, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) \\quad \\text {i f p o s . g u i d a n c e} \\\\ - \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\quad \\text {i f n e g . g u i d a n c e} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Thus, changing the guidance direction is reflected by the direction between \\(\\hat{\\epsilon}_{\\theta}(x_t,c_e)\\) and \\(\\hat{\\epsilon}_{\\theta}(x_t)\\). The term \\(\\phi\\) identifies those dimensions of the image and respective \\(\\hat{\\epsilon}_{\\theta}\\) that are relevant to a prompt \\(e\\). Consequently, \\(\\phi\\) returns 0 for all irrelevant dimensions and a scaling factor \\(s_e\\) for the others. We describe the construction of \\(\\phi\\) in detail below. Larger \\(s_e\\) will increase the effect of the edit, and \\(\\lambda \\in (0,1)\\) reflects the percentage of the pixels selected as relevant by \\(\\phi\\). Notably, for a single concept \\(e\\) and uniform \\(\\phi = s_e\\), Eq. 8 generalizes to the classifier-free guidance term in Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "For multiple \\( e_i \\), we calculate \\( \\gamma_t^i \\) as described above, with each defining their own hyperparameter values \\( \\lambda^i \\), \\( s_e^i \\). The" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.091, + 0.654, + 0.107 + ], + "angle": 0, + "content": "sum of all \\(\\gamma_t^i\\) results in" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.116, + 0.892, + 0.141 + ], + "angle": 0, + "content": "\\[\n\\hat {\\gamma} _ {t} \\left(x _ {t}, c _ {e _ {i}}\\right) = \\sum_ {i \\in I} \\gamma_ {t} ^ {i} \\left(x _ {t}, c _ {e _ {i}}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.147, + 0.892, + 0.193 + ], + "angle": 0, + "content": "Component 3: Semantic Grounding. The masking term \\(\\phi\\) (Eq. 9) is the intersection (pointwise product) of binary masks \\(M^1\\) and \\(M^2\\) combined with scaling factor \\(s_e\\):" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.201, + 0.892, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\phi (\\psi ; s _ {e _ {i}}, \\lambda) = s _ {e _ {i}} M _ {i} ^ {1} M _ {i} ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.227, + 0.892, + 0.422 + ], + "angle": 0, + "content": "where \\( M_{i}^{1} \\) is a binary mask generated from the U-Net's cross-attention layers and \\( M_{i}^{2} \\) is a binary mask derived from the noise estimate. Intuitively, each mask is an importance map, where \\( M_{i}^{1} \\) is more strongly grounded than \\( M_{i}^{2} \\), but of significantly coarser granularity. Therefore, the intersection of the two yields a mask both focused on relevant image regions and of fine granularity. With LEDs++, we empirically demonstrate that these maps can also capture regions of an image relevant to an editing concept that is not already present. Specifically for multiple edits, calculating a dedicated mask for each edit prompt ensures that the corresponding guidance terms remain largely isolated, limiting interference between them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Formally, at each time step \\( t \\), a U-Net forward pass with editing prompt \\( e_i \\) is performed to generate cross-attention maps for each token of the editing prompt. All cross-attention maps of the smallest resolution (e.g., \\( 16 \\times 16 \\) for SD) are averaged over all heads and layers, and the resulting maps are summed over all editing tokens, resulting in a single map \\( A_t^{e_i} \\in R^{16 \\times 16} \\). Importantly, we utilize the same U-Net evaluation \\( \\hat{\\epsilon}_{\\theta}(x_t, c_e) \\) already performed in Eq. 10 to produce \\( M^1 \\) with minimal overhead. Each map \\( A_t^{e_i} \\) is up-sampled to match the size of \\( x_t \\). Cross-attention mask \\( M^1 \\) is derived by calculating the \\( \\lambda \\)-th percentile of up-sampled \\( A_t^{e_i} \\) and" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.614, + 0.891, + 0.655 + ], + "angle": 0, + "content": "\\[\nM _ {i} ^ {1} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left| A _ {t} ^ {e _ {i}} \\right| \\geq \\eta_ {\\lambda} \\left(\\left| A _ {t} ^ {e _ {i}} \\right|\\right) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.663, + 0.891, + 0.708 + ], + "angle": 0, + "content": "where \\(\\eta_{\\lambda}(|\\cdot|)\\) is the \\(\\lambda\\)-th percentile. By definition, \\(M^1\\) only selects image regions that correlate strongly with the editing prompt, and \\(\\lambda\\) determines the size of this selected region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.891, + 0.8 + ], + "angle": 0, + "content": "The fine-grained mask \\( M^2 \\) is calculated based on the guidance vector \\( \\psi \\) of noise estimates derived in Eq. 10. The difference between unconditioned and conditioned \\( \\hat{\\epsilon}_{\\theta} \\), generally captures outlines and object edges of \\( x_{t} \\). Consequently, the largest absolute values of \\( \\psi \\) provide meaningful segmentation information of fine granularity for \\( M^2 \\)" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.808, + 0.892, + 0.848 + ], + "angle": 0, + "content": "\\[\nM ^ {2} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\psi | \\geq \\eta_ {\\lambda} (| \\psi |) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In general, threshold \\(\\lambda\\) should correspond to the performed edit. Changes affecting the entire image, such as style transfer, should choose smaller \\(\\lambda (\\rightarrow 0)\\), whereas edits targeting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8864" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.162, + 0.089, + 0.812, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.272, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Figure 2. Comparison of image editing methods. (top) LEDs++ is the only method to restrict edits to the tree leaves and position of the car. (bottom) Ours is the only approach faithfully executing all three edits and keeping changes minimal. (Best viewed in color)" + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.313, + 0.812, + 0.442 + ], + "angle": 0, + "content": "
MethodReconstruction \nError (RMSE) ↓Execution \nTime (s) ↓Variation/ \nSamplingSemantic \nGroundingMulti- \nEditing
SDEdit [26]0.81 ±0.072.10 ±0.02✓XX
Imagic [18]0.58 ±0.12349.98 ±0.45✓XX
Vanilla DDIM Inversion0.22 ±0.1037.23 ±0.04XXX
Pix2Pix-Zero [30]0.20 ±0.0956.78 ±0.14(✓)✓X
DiffEdit [9]0.13±0.0327.65 ±0.03✓✓X
Edit-friendly DDPM [17]0.0010.36 ±0.05✓XX
LEDSITS++ (Ours)0.001.78 ±0.03✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.447, + 0.893, + 0.489 + ], + "angle": 0, + "content": "Table 1. Comparing key properties for diffusion-based image editing techniques, with LEDs++ offering clear methodological benefits. Due to LEDs++'s efficient perfect inversion, it is the fastest and error-free method. At the same time, its methodology is the only enabling versatility in terms of variation, semantic grounding, and multi-editing. Subscript numbers indicate standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.47, + 0.541 + ], + "angle": 0, + "content": "specific objects or regions should use \\(\\lambda\\) proportional to the region's prominence in the image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.308, + 0.574 + ], + "angle": 0, + "content": "4. Properties of LEDs++" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.613 + ], + "angle": 0, + "content": "With the fundamentals of LEDs++ established, we next showcase its unique properties and capabilities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.854 + ], + "angle": 0, + "content": "Efficiency. First off, LEDs++ offers substantial performance improvements over other image editing methods. In Tab. 1, we provide a qualitative runtime comparison, with all methods being implemented for Stable Diffusion (SD) 1.5 [34]. As a parameter-free approach, LEDs++ does not require any computationally expensive fine-tuning or optimization. Consequently, LEDs++ is orders of magnitude faster than methods like Imagic [18] or Pix2Pix-Zero [30]. Further, we only need to invert the same number of diffusion steps used at inference, which results in significant runtime improvements over the standard DDIM inversion (21x). In addition to efficient inversion, we use a recent, fast scheduler that generally requires fewer total steps, further boosting performance. This way, LEDs++ is six times faster than recent DDPM inversion [17] and on par with fast but poor-quality SDEdit [26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Versatility. In addition to its efficiency, LEDs++ remains versatile, enabling sheer limitless editing possibilities. In Fig. 1, we showcase a broad range of edit types." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.892, + 0.722 + ], + "angle": 0, + "content": "LEDITS++ facilitates fine-grained edits (adding/removing glasses) and holistic changes such as style transfer (painting/sketch). Furthermore, object removal and replacement facilitate even more image editing tasks. Importantly, the overall image composition is preserved in all cases. To our knowledge, LEDs++ is the only diffusion-based image editing method inherently supporting multiple edits in isolation, which allows for more complex image manipulation. Fig. 2 highlights LEDs++ benefits over previous methods. Our method produces the highest edit fidelity and is the only approach capable of faithfully executing multiple, simultaneous instructions. Moreover, LEDs++ also makes the least changes to unrelated objects and the overall background and composition of the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.892, + 0.835 + ], + "angle": 0, + "content": "Lastly, the editing versatility benefits from the stochastic nature of the perfect but non-deterministic inversion. LEDs++ provides meaningful image variations by resampling \\(\\tilde{c}_t\\). Additionally, the visual expression of each concept in the edited image scales monotonically with \\(s_e\\), and the direction and magnitude of each concept can be varied freely. We present examples of both features in App. B." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Precision. Furthermore, LEDs++'s methodology keeps edits concise and avoids unnecessary deviations from the input image (Fig. 2). First, the perfect inversion will reconstruct the exact input image if no edit is applied (cf." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8865" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.43, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.196, + 0.471, + 0.239 + ], + "angle": 0, + "content": "Figure 3. Exemplary edit performed with LEDs++ in only 25 diffusion steps with SD1.5. We apply a complex, compounded edit and ground each to a semantically reasonable image region." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.451 + ], + "angle": 0, + "content": "Sec. 3.2). Consequently, we already improve on faithfulness to the input image even before applying any edits. This benefit over other methods is highlighted by the reconstruction error in Tab. 1. Second, implicit masking will semantically ground each edit to relevant image regions. This is specifically important for editing multiple concepts at the same time. While other methods only utilize one prompt for all instructions, LEDs++ isolates edits from each other (Eq. 11). Thus, we get dedicated masks for each concept as shown in Fig. 3. This design ensures that each instruction (e.g., red mask for 'cherry blossom') will be only applied where necessary. Subsequently, we provide further evidence for the efficacy of LEDs++'s masking approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.424, + 0.484 + ], + "angle": 0, + "content": "5. Semantically Grounded Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Cross-attention maps of DMs have been used extensively to ground regions of interest during image generation semantically [7, 8, 14, 30]. Nonetheless, these have not been combined with noise-based masks so far and thus lack fine granularity. Hence, we empirically evaluate the quality of implicit masks, i.e., attention maps \\( M^1 \\) and noise maps \\( M^2 \\) (Eq. 13 and 14) in the LEDs++ setup. We use a broad segmentation task for common objects as a proxy to measure the performance of implicit masks in identifying relevant image areas from edit instructions. Specifically, we utilize segmentation masks from the COCO panoptic segmentation challenge [19]. For each unique object in an image, we retrieve the masks \\( M^1 \\), \\( M^2 \\), and their intersection per diffusion step. We use the (semantic) class label (e.g. 'person' or 'TV') as editing concept \\( e \\). We consider masks at each of 50 total diffusion steps without actually editing the input image. Furthermore, we approximate mask threshold \\( \\lambda \\) based on the relative size of an object's bounding box." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Concise masks with LEDs++. Fig. 4 shows implicit masking as a reliable means to identify relevant image regions. Importantly, the intersection of both cross-attention masks \\( M^1 \\) and noise maps \\( M^2 \\) clearly outperforms each separate mask. The overall performance is even similar to a dedicated CLIPSeg model [24], despite LEDs++. masks being implicitly calculated at inference with only minimal overhead. At the same time, LEDs++.'s masking is superior to DiffEdit's [3]. Consequently, our method's inter" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.093, + 0.868, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.293 + ], + "angle": 0, + "content": "Figure 4. Semantic segmentation quality of LEDs++. We show the intersection over union (higher is better) for COCO panoptic segmentation. The intersection masks outperform each by a clear margin, close to the CLIPSeg reference. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.355 + ], + "angle": 0, + "content": "section of cross-attention masks and noise maps provides strong semantic grounding while being efficient during image manipulation to ensure precise editing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.37, + 0.742, + 0.387 + ], + "angle": 0, + "content": "6. Image Editing Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.396, + 0.891, + 0.427 + ], + "angle": 0, + "content": "Let us now compare LEDs++ to current SOTA methods for image manipulation on two benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.74, + 0.456 + ], + "angle": 0, + "content": "6.1. Editing Multiple Concepts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.734 + ], + "angle": 0, + "content": "First, we investigate the complex task of performing multiple edits simultaneously. We rely on a well-established setup for semantic image manipulation [5] to evaluate multi-conditioned attribute manipulation in facial images. In our experiment, we consider 100 images from the CelebA dataset [21]. For each image, we simultaneously edit three attributes out of a set of five, leading to ten total combinations of edit concepts. Further, we perform each edit across ten different seeds, resulting in 10,000 evaluated images for each method and hyperparameter setting, over 1M images in total. As measures for comparison, we employ CLIP and LPIPS scores. CLIP measures the text-to-image similarity of the edit instruction to the edited image, and LPIPS measures the image-to-image similarity of the real to the edited image. This way, we assess the tradeoff between the versatility of edits (CLIP) and the precision of those manipulations (LPIPS). We implement all methods based on SD1.5 and provide more details in App. C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "LEDS++ outperforms competing methods. Fig. 5 shows the resulting CLIP vs. LPIPS plots for all methods. The top left corner represents the ideal editing method with maximum edit alignment without deviating from the initial image. Generally, one can observe a natural trade-off between versatility and precision for all methods, i.e., higher image-to-text alignment comes at the expense of lower similarity to the original image. LEDS++ is closest to the ideal region and thus clearly outperforms the other methods. In particular, the outputs remain close to the original image (low LPIPS scores), thanks to the precise implicit masking." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8866" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.445, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Comparison of instruction-alignment vs. image similarity trade-off for different editing methods. Results were reported for simultaneous manipulation of three facial attributes on CelebA. We plot CLIP scores (higher is better) of the target attributes against LPIPS similarity (lower is better). LEDs++ clearly outperforms all competing methods. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.373, + 0.473, + 0.601 + ], + "angle": 0, + "content": "At the same time, it faithfully performs the edits (high CLIP scores) due to the dedicated, isolated editing for each concept. The depicted scores reflect our qualitative inspections for Pix2Pix-Zero and Imagic on such complex manipulations (cf. Fig. 2). We observed that these methods often break—either failing to perform all three edits and/or drastically altering the input image. Only edit-friendly DDPM [17] and LEDs++ reliably achieve the maximum average CLIP score of over 0.25. This value seems to represent an upper bound according to our manual investigations, as each attribute is edited correctly for all input images, and higher scores are not observed. Despite being computationally very efficient, LEDs++ faithfully executes each edit instruction while keeping the changes to the input low, highlighting the method's versatility and precision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.228, + 0.626 + ], + "angle": 0, + "content": "6.2. TEDBench \\((++)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Next, we investigate the versatility of LEDs++'s editing capabilities by running the Textual Editing Benchmark (TEDBench [18]), a collection of 100 input images paired with textual edit instructions. However, we observed a variety of inconsistencies in TEDBench and a lack of relevant editing tasks. Therefore, we propose TEDBench++ (Fig. 6a and App. D), a more challenging revised benchmark now containing 120 entries in total.3 We addressed misspellings and rephrased ambiguous and inconclusive instructions. In addition to resolving these issues, we added instructions targeting challenging types of image manipulations previously not included in TEDBench: multi-conditioning, object/concept removal, style transfer, and complex replacements (Fig. 6a). We provide more details in App. D." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.469, + 0.876 + ], + "angle": 0, + "content": "We compare LEDs++ on TEdBench \\((++)\\) to one of the strongest editing methods, Imagic [18] with Imagen [37]." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.878, + 0.175 + ], + "angle": 0, + "content": "
TEDBenchTEDBench++
SR ↑LPIPS ↓SR ↑LPIPS ↓
Imagic w/ SD1.50.550.560.580.57
LEDITS++ w/ SD1.50.750.280.790.30
Imagic w/Imagen [18]0.830.59——
LEDITS++ w/ SD-XL0.840.330.870.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.181, + 0.892, + 0.236 + ], + "angle": 0, + "content": "Table 2. Success rate (SR) and LPIPS scores on the original TedBench [18] and our revised version (TEDBench++) We compare Imagic to LEDs++ based on different DMs and find the latter to outperform on both metrics and benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Since both are not publicly available, we can only compare to this specific combination of DM and editing method using Kawar et al.'s [18] curated outputs for TEdBench. Additionally, we, therefore, cannot combine LEDs++ with Imagen[37] and instead use a similarly advanced diffusion model, SD-XL [32]. However, to not only compete for the best fidelity outputs but focus the evaluation on methodological differences—not the pre-trained DM—we also compare both methods implemented with SD1.5. We provide further details in App. C." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.893, + 0.568 + ], + "angle": 0, + "content": "LEDS \\(^{+ + }\\) edits images reliably. We first asked users to assess the overall success of edits, i.e., if an edit instruction was faithfully realized for a given input image. The results in Tab. 2 show that LEDs \\(^{+ + }\\) outperforms Imagc on TEdBench despite a greatly reduced runtime (Tab. 1). The difference is even stronger when comparing both methods on the same pre-trained DM, i.e., SD1.5. The high success rate on TEdBench \\(^{+ + }\\) \\((87\\%)\\) and the examples shown in Fig. 1 and 6a once again highlight LEDs \\(^{+ + }\\)'s versatility. Overall, our proposed method can reliably perform a diverse set of editing instructions for real images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.796 + ], + "angle": 0, + "content": "High-quality edits with LEDs++. While investigating both methods' performance we observed a substantial difference in edit quality. The examples in Fig. 6b particularly highlight the discrepancy in compositional robustness and object coherence. Hence, we also assessed both methods' editing quality on TEdBench++. We focus on samples where both methods performed a successful edit, i.e., were labeled as successful by users. We show the perceptual similarity (LPIPS) to the input image in Tab. 2. One can observe that the LPIPS scores for LEDs++ are much lower than for Imagic, empirically supporting the qualitative examples in Fig. 6b. When manually inspecting the generated images, we often found Imagic to generate a completely new image based on the edit instruction, entirely disregarding the input image (cf. App. Fig. 11a)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.614, + 0.829 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.89, + 0.854 + ], + "angle": 0, + "content": "Let us now discuss open research questions and limitations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Model Dependency. While LEDs++ achieves impressive results on a large variety of image manipulation tasks, there are external factors to consider. Since the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.395, + 0.9 + ], + "angle": 0, + "content": "3 https://huggingface.co/datasets/AIML-TUDA/TEdBench_plusplus" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8867" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.526, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.294, + 0.529, + 0.32 + ], + "angle": 0, + "content": "(a) Novel challenging examples of TEdBench++ and LEDs++ applied, showcasing the versatility of supported edits." + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.089, + 0.891, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.294, + 0.892, + 0.32 + ], + "angle": 0, + "content": "(b) Qualitative comparison of LEDs++ and Imagic on TEd-Bench, clearly highlighting the performance improvement." + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.325, + 0.785, + 0.34 + ], + "angle": 0, + "content": "Figure 6. Benchmark examples for LEDs++ and Imagc on TEdBench \\((++)\\) . (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.595 + ], + "angle": 0, + "content": "method is architecture-agnostic, it can be easily used with any DM. At the same time, the general editing quality strongly depends on the overall capabilities of the underlying pre-trained DM. Naturally, more capable models will also enable better edits. But, at times, specific editing instructions may fail because the used DM does not have a decent representation of the targeted concept to begin with. One example is the model failing to edit a giraffe to be sitting since the underlying DM generally fails to generate this pose (cf. App. F). This effect can also clearly be seen in Tab. 2, with the editing success rate of a method varying strongly between DMs. Although the same image editing method is employed (LEDITS++, the more capable SD-XL variant outperforms the weaker SD1.5 model. Nonetheless, this means that the architecture-agnostic LEDs++ will benefit from increasingly powerful DMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Coherence Trade-offs. Next to the benefits of LEDs++'s semantic grounding, there are also downsides to this approach. Overall, implicit masking limits changes to relevant portions of the image and achieves strong coherence with the original image composition. Yet, the object and its identity within the masked area may change based on various factors. Generic prompts, like \"a standing cat\" (cf. App. F), do not contain detailed information about this specific object (\"cat\"). Thus, an edit with this prompt does not guarantee to preserve object identity, particularly for strong hyperparameters. We observed that fine-tuning approaches like Imagic make the opposite trade-off, better preserving the object identity while changing the background and image composition substantially (cf. App. F). A potential remedy for a loss in object coherence with LEDs++, is more descriptive edit prompting, e.g. using textual inversion [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Lastly, the automatically-inferred implicit masks allow for easy use of LEDs++ without users tediously providing masks. Nonetheless, user intentions are diverse and cannot always be automatically inferred. Sometimes, in-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.414 + ], + "angle": 0, + "content": "individual user masks provide better control over the editing process. Such user masks can be easily integrated into LEDs++ (cf. App. F), wherefore we encourage future research in this promising direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.567 + ], + "angle": 0, + "content": "Societal Impact. LEDs++ is an easy-to-use image editing technique that lowers the barrier for users and puts them in control for fruitful human-machine collaboration. Yet, the underlying text-to-image models offer both promise and peril, as highlighted by prior research [4, 11]. The (societal) biases within these models will also impact image editing applications [11]. Moreover, image manipulation can also be used adversarially to generate inappropriate [39] or fake content. Hence, we advocate for a cautious deployment of generative models together with image editing methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.584, + 0.619, + 0.6 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.892, + 0.776 + ], + "angle": 0, + "content": "We introduced LEDs++, an efficient yet versatile and precise method for textual image manipulation with diffusion models. It facilitates the editing of complex concepts in real images. Our approach requires no finetuning nor optimization, can be computed extremely efficiently, and is architecture agnostic. At the same time, it perfectly reconstructs an input image and uses implicit masking to limit changes to relevant image regions, thus editing precisely. Our large experimental evaluation confirms the efficiency, versatility, and precision of LEDs++ and its components, as well as its benefits over several related methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements We gratefully acknowledge support from the BMBF (Grant No 01IS22091). This work benefited from the ICT-48 Network of AI Research Excellence Center \"TAILOR\" (EU Horizon 2020, GA No 952215), the Hessian research priority program LOEWE within the project WhiteBox, the HMWK cluster projects \"Adaptive Mind\" and \"Third Wave of AI\", and from the NHR4CES." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8868" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. SIGGRAPH Asia, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.182 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv:2206.02779, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.183, + 0.471, + 0.25 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv:2211.01324, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.251, + 0.471, + 0.345 + ], + "angle": 0, + "content": "[4] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. In Proceedings of ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2023. 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[5] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. In Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.483, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv:2304.08465, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.47, + 0.724 + ], + "angle": 0, + "content": "[10] Zoe De Simone, Angie Boggust, Arvind Satyanarayan, and Ashia Wilson. What is a Fair Diffusion Model? Designing Generative Text-To-Image Models to Incorporate Various Worldviews. arXiv preprint arXiv:2309.09944, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.726, + 0.47, + 0.793 + ], + "angle": 0, + "content": "[11] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness. arXiv preprint arXiv:2302.10893, 2023. 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.794, + 0.47, + 0.861 + ], + "angle": 0, + "content": "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.862, + 0.47, + 0.89 + ], + "angle": 0, + "content": "[13] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.12, + 0.892, + 0.186 + ], + "angle": 0, + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.187, + 0.892, + 0.269 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and HsuanTien Lin, editors, Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.27, + 0.892, + 0.296 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 1, 2, 3, 4, 5, 7, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.35, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[18] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 7, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[20] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[21] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), December 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.647 + ], + "angle": 0, + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[24] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.757, + 0.892, + 0.81 + ], + "angle": 0, + "content": "[25] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv:2310.04378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.811, + 0.892, + 0.878 + ], + "angle": 0, + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 1, 2, 4, 5, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.879, + 0.892, + 0.891 + ], + "angle": 0, + "content": "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.891 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.215, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2022. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[30] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.338, + 0.471, + 0.392 + ], + "angle": 0, + "content": "[31] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv:2306.00637, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.471, + 0.513 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.471, + 0.581 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.69 + ], + "angle": 0, + "content": "[36] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V. Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased Face Synthesis With Diffusion Models: Are We There Yet? arXiv preprint arXiv:2309.07277, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv:2205.11487, 2022. 2, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.785, + 0.471, + 0.826 + ], + "angle": 0, + "content": "[38] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. arXiv:2304.14530, 2023. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.826, + 0.471, + 0.895 + ], + "angle": 0, + "content": "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8, 11" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of NeurIPS Datasets and Benchmarks, 2022. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.203, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[41] Norbert Schwarz. How the questions shape the answers. American Psychologist, 1999. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.23, + 0.892, + 0.283 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.352, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[44] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv:2210.09477, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[45] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv:1212.08698, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv:2302.04867, 2023.2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8870" + } + ] +] \ No newline at end of file diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_origin.pdf b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..08cdf65fb842073117586b125a9336d98b53bfe3 --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/f8bff481-3f32-424c-92f1-b66a29a1df02_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:768424414387c2cb5a65d82f4d9ac3537934e7c92240a45b7b5191b2a832fa91 +size 1487234 diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/full.md b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08c2ce31e18faf4bff8fa25a6ca1871e7f7a5189 --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/full.md @@ -0,0 +1,342 @@ +# LEDS: Limitless Image Editing using Text-to-Image Models + +![](images/d8f7be564b6332796a8cb8adaad3a2d7a5f0c215956e770c5c1a33382a955dd0.jpg) + +![](images/fd68bca424e2d5b2224b8fd297e5ce3f861d71132db180895b57fbeebd7121b1.jpg) +Figure 1. LEDs++ facilitates versatile image-to-image editing. Several complex cases are available now. + +# Abstract + +Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDs++, an efficient yet versatile and precise textual image manipulation technique. LEDs++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDs++ and its improvements over previous methods. + +# 1. Introduction + +Text-to-image diffusion models (DM) have garnered recognition for their ability to generate high-quality images from + +textual descriptions. A growing body of research has recently been dedicated to utilizing these models for manipulating real images. However, several barriers prevent many real-world applications of diffusion-based image editing. Current methods often entail computationally expensive model tuning or other optimization, presenting practical challenges [6, 18, 28, 30, 44]. Additionally, existing techniques have the proclivity to induce profound changes to the original image [17, 26], often resulting in completely different images. Lastly, all these approaches are inherently constrained when editing multiple (arbitrary) concepts simultaneously. We tackle these problems by introducing LEDs++ $^{1}$ , a diffusion-based image editing technique addressing these limitations. + +LEDITS++2 offers a streamlined approach for textual image editing, eliminating the need for extensive parameter tuning. To this end, we derive image inversion for a more efficient diffusion sampling algorithm to a) drastically reduce computational resources and b) guarantee perfect image reconstruction. Thus, we overcome computational obstacles and avoid changes in the edited image in the first place. Furthermore, we use a novel implicit masking approach to semantically ground each edit instruction to its relevant image region. This further optimizes changes to the image by retaining the overall image composition and object identity. + +*Equal contribution †Partially as research intern at Adobe +1LEDITS++ stands for Limitless Edits with sde-dpm-solver++. +2https://huggingface.co/spaces/leditsplusplus/project + +Additionally, LEDs++ is the only method to date to facilitate easy and versatile image editing by supporting multiple simultaneous instructions without causing undue interference. Finally, its lightweight architecture-agnostic nature ensures compatibility with both latent and pixel-based diffusion models, providing high accessibility. + +In this work, we establish the methodical benefits of LEDs++ and demonstrate that this intuitive, lightweight approach offers sophisticated semantic control for image editing. Specifically, we contribute by (i) devising a formal definition of LEDs++ while (ii) deriving perfect inversion for a more efficient diffusion sampling method, (iii) qualitatively and empirically demonstrating its efficiency, versatility, and precision, (iv) providing an exhaustive empirical comparison to concurrent works with automatic and human user metrics, and thereby (v) introducing Textual Editing Benchmark++ (TEdBench++, a more holistic and coherent testbed for evaluating textual image manipulation. + +# 2. Background + +Recently, large-scale, text-guided DMs have enabled versatile applications in image generation [3, 33, 37]. Especially latent diffusion models [31, 34] have gained attention for their computational efficiency. Below, we discuss related work for efficient, versatile image manipulation with DMs. + +Diffusion Sampling. Generating outputs with DMs requires multiple iterative denoising steps that constitute the main bottleneck at inference. Commonly used sampling methods such as DDPM [15] or DDIM [42] require tens or hundreds of steps to produce high-quality samples. Consequently, numerous works have been dedicated to speeding up the sampling process without loss in quality. Distillation efforts progressively reduce the number of required steps through further training [25, 27]. Other works focus on improving the sampling itself, e.g. using high-order ODE-solvers [22, 23, 46]. Such solvers can be readily combined with pre-trained DMs at inference to lower the number of denoising steps. With LEDs++, we derive perfect image inversion with the DPM-Solver++, allowing image editing in as few as 20 total steps. + +Semantic Control during Diffusion. While text-to-image DMs generate new, astonishing images, fine-grained control over the generative process remains challenging. Minor changes to the text prompt lead to entirely different outputs. Wu et al. [45] studied concept disentanglement using linear combinations of text embeddings to gain semantic control. Methods like Prompt-to-Prompt [14] and other works [8, 30] utilize the DM's attention layers to attribute pixels to tokens from the text prompt. Dedicated operations on the attention maps enable more control over the generated images. Other works have focused on the noise estimates of DMs [5, 20] providing semantic control over the generation process. With LEDs++, we now enable + +fine-grained semantic control for manipulating real images, going beyond purely generative applications. + +Real Image Editing. Since DMs' rise in popularity for text-to-image generation, they have also been explored for (real) image-to-image editing. As a first, simple approach, SDEdit added noise to the image for an intermediate step in the diffusion process [26]. While lightweight, the resulting image diverges substantially from the input as it is (partially) regenerated. Inpainting allows to keep the change small by having a user provide additional masks to restrict changes to certain image regions [2, 29]. Yet, user masks are costly or often simply unavailable. Other works have thus explored semantically grounded approaches using cross-attention instead to better control image manipulation [7, 28, 30]. In contrast, LEDs++ leverages both attention-and noise-based masking to obtain fine-grained masks, enabling strong semantic control over real images. + +Another important aspect of image manipulation methods is the required tuning and overall runtime. Instruct-Pix2Pix continues training a DM at scale to enable image editing capabilities [6]. Finetuning instead on each individual input to constrain the generation on the real image has shown helpful [18, 44] but not computationally efficient. Consequently, recent works have largely relied on inverting the deterministic DDIM sampling process [42] to save computational resources. DDIM inversion identifies an initial noise vector that results in the input image when denoised again. However, faithful reconstructions are only obtained in the limit of small steps, thus requiring large numbers of inversion steps. Moreover, small errors will still incur at each timestep, often accumulating into meaningful deviations from the input, requiring costly error correction through optimization [28, 30]. Recently, Huberman-Spiegelglas et al. proposed an inversion technique [17] for the DDPM sampler [15] to address the limitations of DDIM inversion. LEDs++ provides the same guarantees of perfect inversion with even further reduced runtime alongside an edit-friendly latent space, enabling more versatility. + +# 3. Image Editing with Text-to-Image Models + +Before devising the methodology of LEDs++, let us first motivate the desired features and use cases. Specifically, we aim for efficiency, versatility, and precision. The goal is to provide a method that enables a fast exploratory workflow for image editing in which a user can iteratively interact with the model and explore various edits. Consequently, LEDs++ produces outputs quickly with no tuning or optimization to not disrupt the creative process. Further, arbitrary editing instructions and combinations thereof are supported to facilitate a wide range of image manipulations (e.g., complex multi-editing). Lastly, we provide precise and sophisticated semantic control over the image editing. Each of the (potentially multiple) edit instruc + +tions can be steered individually, and changes are automatically restricted to relevant image regions. Importantly, with LEDs++ we prioritize compositional robustness. + +# 3.1. Guided Diffusion Models + +Let us first define some general background for diffusion models. DMs iteratively denoise a Gaussian distributed variable to produce samples of a learned data distribution. Let's consider a diffusion process that gradually turns an image $x_0$ into Gaussian noise. + +$$ +x _ {t} = \sqrt {1 - \beta_ {t}} x _ {t - 1} + \sqrt {\beta_ {t}} n _ {t}, \quad t = 1, \dots , T \tag {1} +$$ + +where $n_t$ are iid normal distributed vectors and $\beta_t$ a variance schedule. The diffusion process is equivalently expressed as + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} \tag {2} +$$ + +where $\alpha_{t} = 1 - \beta_{t}$ , $\bar{\alpha}_{t} = \Pi_{s=1}^{t} \alpha_{s}$ and $\epsilon_{t} \sim \mathcal{N}(0, \mathbf{I})$ . Importantly, all $\epsilon_{t}$ are not statistically independent. Instead, consecutive pairs $\epsilon_{t}, \epsilon_{t-1}$ are strongly dependent, which will be relevant later. To generate an (new) image $\hat{x}_{0}$ the reverse diffusion process starts from random noise $x_{T} \sim \mathcal{N}(0, \mathbf{I})$ which can be iteratively denoised as + +$$ +x _ {t - 1} = \hat {\mu} _ {t} \left(x _ {t}\right) + \sigma_ {t} z _ {t}, \quad t = T, \dots , 1 \tag {3} +$$ + +Here $z_{t}$ are iid standard normal vectors, and common variance schedulers $\sigma_{t}$ can be expressed in the general form + +$$ +\sigma_ {t} = \eta \beta_ {t} \frac {1 - \bar {\alpha} _ {t - 1}}{1 - \bar {\alpha} _ {t}} +$$ + +where $\eta \in [0,1]$ . In this formulation, $\eta = 0$ corresponds to the deterministic DDIM [42] and $\eta = 1$ to the DDPM scheme [15]. Lastly, in theses cases, we have $\hat{\mu}_t(x_t) =$ + +$$ +\sqrt {\bar {\alpha} _ {t - 1}} \frac {x _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \hat {\epsilon} _ {\theta} (x _ {t})}{\sqrt {\bar {\alpha} _ {t}}} + \sqrt {1 - \bar {\alpha} _ {t - 1} - \sigma_ {t} ^ {2}} \hat {\epsilon} _ {\theta} (x _ {t}) +$$ + +Here $\hat{\epsilon}_{\theta}(x_{t})$ is an estimate of $\epsilon_{t}$ produced by our neural network DM with learned parameters $\theta$ , commonly implemented as a U-Net [35]. For text-to-image generation, the model is conditioned on a text prompt $p$ to produce images faithful to that prompt. The DM is trained to produce the noise estimate $\hat{\epsilon}_{\theta}(x_{t})$ needed for iteratively sampling $\hat{x}_0$ (Eq. 3). For text-conditioned DMs, $\hat{\epsilon}_{\theta}$ is calculated using specific guidance techniques. + +Most DMs rely on classifier-free guidance [16], a conditioning method using a purely generative diffusion model, eliminating the need for an additional classifier. During training, the text conditioning $c_{p}$ is randomly dropped with a fixed probability, resulting in a joint model for unconditional and conditional objectives. During inference, the score estimates for the $\epsilon$ -prediction are adjusted so that: + +$$ +\hat {\epsilon} _ {\theta} \left(x _ {t}, c _ {p}\right) := \hat {\epsilon} _ {\theta} \left(x _ {t}\right) + s _ {g} \left(\hat {\epsilon} _ {\theta} \left(x _ {t}, c _ {p}\right) - \hat {\epsilon} _ {\theta} \left(x _ {t}\right)\right) \tag {4} +$$ + +with guidance scale $s_g$ and $\hat{\epsilon}_{\theta}$ defining the noise estimate with parameters $\theta$ . Intuitively, the unconditioned $\epsilon$ -prediction is pushed in the direction of the conditioned one, with $s_g$ determining the extent of the adjustment. + +# 3.2.LEDITS++ + +With the fundamentals established, the methodology of LEDs++ can now be broken down into three components: (1) efficient image inversion, (2) versatile textual editing, and (3) semantic grounding of image changes. + +Component 1: Perfect Inversion. Utilizing text-to-image models for editing real images requires conditioning the generation on the input image. Recent works have largely relied on inverting the sampling process to identify $x_{T}$ that will be denoised to the input image $x_{0}$ [28, 30]. Inverting the DDPM scheduler is generally preferred over DDIM inversion since the former can be achieved in fewer timesteps and with no reconstruction error [17]. + +However, there exist more efficient schemes than DDPM for sampling DMs that greatly reduce the required number of steps and consequently DM evaluations. We here propose a more efficient inversion method by deriving the desired inversion properties for such a scheme. As demonstrated by Song et al.[43], DDPM can be viewed as a first-order stochastic differential equation (SDE) solver when formulating the reverse diffusion process as an SDE. This SDE can be solved more efficiently—in fewer steps—using a higher-order differential equation solver, here dpmsolver++ [23]. The reverse diffusion process from Eq. 3 for the second-order sde-dpm-solver++ can be written as + +$$ +x _ {t - 1} = \hat {\mu} _ {t} \left(x _ {t}, x _ {t + 1}\right) + \sigma_ {t} z _ {t}, \quad t = T, \dots , 1 \tag {5} +$$ + +where now + +$$ +\sigma_ {t} = \sqrt {1 - \bar {\alpha} _ {t - 1}} \sqrt {1 - e ^ {- 2 h _ {t - 1}}} +$$ + +and higher-order $\hat{\mu}_t$ depends now on $x$ from two timesteps, $x_t$ and $x_{t + 1}$ . Such that $\hat{\mu}_t(x_t,x_{t + 1}) =$ + +$$ +\begin{array}{l} \frac {\sqrt {1 - \bar {\alpha} _ {t - 1}}}{\sqrt {1 - \bar {\alpha} _ {t}}} e ^ {- h _ {t - 1}} x _ {t} + \sqrt {\bar {\alpha} _ {t - 1}} \left(1 - e ^ {- 2 h _ {t - 1}}\right) \hat {\epsilon} _ {\theta} (x _ {t}) \\ + 0. 5 \sqrt {\bar {\alpha} _ {t - 1}} \left(1 - e ^ {- 2 h _ {t - 1}}\right) \frac {- h _ {t}}{h _ {t - 1}} \left(\hat {\epsilon} _ {\theta} \left(x _ {t + 1}\right) - \hat {\epsilon} _ {\theta} \left(x _ {t}\right)\right) \\ \end{array} +$$ + +with + +$$ +h _ {t} = \frac {\ln (\sqrt {\bar {\alpha} _ {t}})}{\ln (\sqrt {1 - \bar {\alpha} _ {t}})} - \frac {\ln (\sqrt {\bar {\alpha} _ {t + 1}})}{\ln (\sqrt {1 - \bar {\alpha} _ {t + 1}})} +$$ + +For the detailed derivation of the solver and proof of faster convergence, we refer the reader to the relevant literature [22, 23]. Based on the above, we now devise our inversion process. Given an input image $x_0$ we construct an auxiliary reconstruction sequence of noise images $x_1, \ldots, x_T$ as + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \tilde {\varepsilon} _ {t} \tag {6} +$$ + +where $\tilde{\epsilon}_t\sim \mathcal{N}(0,\mathbf{I})$ . Contrary to Eq. 2, the $\tilde{\epsilon}_t$ are now statistically independent, which is a desirable property for image editing [17]. Lastly, the respective $z_{t}$ for the inversion can be derived from Eq. 5 as + +$$ +z _ {t} = \frac {x _ {t - 1} - \hat {\mu} _ {t} \left(x _ {t} , x _ {t + 1}\right)}{\sigma_ {t}}, \quad t = T,.., 1 \tag {7} +$$ + +with $\hat{\mu}$ and $\sigma_{t}$ as defined above. We base our implementation on the multistep variant of sde-dpm-solver++, which only requires one evaluation of the DM at each diffusion timestep by reusing the estimates from the previous step. The number of timesteps can be reduced further by stopping the inversion at an intermediate step $t < T$ and starting the generation at that step. Empirically, we observed that $t\in [0.9T,0.8T]$ usually produces edits of the same fidelity as $t = T$ , supporting observations in previous work [17, 26] that earlier timesteps are less relevant to the edit. + +Component 2: Textual Editing. After creating our reconstruction sequence $x_{1}, \ldots, x_{T}$ and calculating the respective $z_{t}$ , we now edit the image by manipulating the noise estimate $\hat{\epsilon}_{\theta}$ based on a set of edit instructions $\{e_i\}_{i \in I}$ . We devise a dedicated guidance term for each concept $e_{i}$ based on conditioned and unconditioned estimates. Let us formally define LEDs++'s guidance by starting with a single editing prompt $e$ . We compute + +$$ +\hat {\epsilon} _ {\theta} \left(x _ {t}, c _ {e}\right) := \hat {\epsilon} _ {\theta} \left(x _ {t}\right) + \gamma \left(x _ {t}, c _ {e}\right) \tag {8} +$$ + +with guidance term $\gamma$ . Consequently, setting $\gamma = 0$ will reconstruct the input image $x_0$ . We construct $\gamma$ to push the unconditioned score estimate $\hat{\epsilon}_{\theta}(x_t)$ —i.e. the input image reconstruction—away from/towards the edit concept estimate $\hat{\epsilon}_{\theta}(x_t, c_e)$ , depending on the guidance direction: + +$$ +\gamma \left(x _ {t}, c _ {e}\right) = \phi (\psi ; s _ {e}, \lambda) \psi \left(x _ {t}, c _ {e}\right) \tag {9} +$$ + +where $\phi$ applies an edit guidance scale $s_e$ element-wise, and $\psi$ depends on the edit direction: $\psi(x_t, c_e) =$ + +$$ +\left\{ \begin{array}{l} \hat {\epsilon} _ {\theta} \left(x _ {t}, c _ {e}\right) - \hat {\epsilon} _ {\theta} \left(x _ {t}\right) \quad \text {i f p o s . g u i d a n c e} \\ - \left(\hat {\epsilon} _ {\theta} \left(x _ {t}, c _ {e}\right) - \hat {\epsilon} _ {\theta} \left(x _ {t}\right)\right) \quad \text {i f n e g . g u i d a n c e} \end{array} \right. \tag {10} +$$ + +Thus, changing the guidance direction is reflected by the direction between $\hat{\epsilon}_{\theta}(x_t,c_e)$ and $\hat{\epsilon}_{\theta}(x_t)$ . The term $\phi$ identifies those dimensions of the image and respective $\hat{\epsilon}_{\theta}$ that are relevant to a prompt $e$ . Consequently, $\phi$ returns 0 for all irrelevant dimensions and a scaling factor $s_e$ for the others. We describe the construction of $\phi$ in detail below. Larger $s_e$ will increase the effect of the edit, and $\lambda \in (0,1)$ reflects the percentage of the pixels selected as relevant by $\phi$ . Notably, for a single concept $e$ and uniform $\phi = s_e$ , Eq. 8 generalizes to the classifier-free guidance term in Eq. 4. + +For multiple $e_i$ , we calculate $\gamma_t^i$ as described above, with each defining their own hyperparameter values $\lambda^i$ , $s_e^i$ . The + +sum of all $\gamma_t^i$ results in + +$$ +\hat {\gamma} _ {t} \left(x _ {t}, c _ {e _ {i}}\right) = \sum_ {i \in I} \gamma_ {t} ^ {i} \left(x _ {t}, c _ {e _ {i}}\right) \tag {11} +$$ + +Component 3: Semantic Grounding. The masking term $\phi$ (Eq. 9) is the intersection (pointwise product) of binary masks $M^1$ and $M^2$ combined with scaling factor $s_e$ : + +$$ +\phi (\psi ; s _ {e _ {i}}, \lambda) = s _ {e _ {i}} M _ {i} ^ {1} M _ {i} ^ {2} \tag {12} +$$ + +where $M_{i}^{1}$ is a binary mask generated from the U-Net's cross-attention layers and $M_{i}^{2}$ is a binary mask derived from the noise estimate. Intuitively, each mask is an importance map, where $M_{i}^{1}$ is more strongly grounded than $M_{i}^{2}$ , but of significantly coarser granularity. Therefore, the intersection of the two yields a mask both focused on relevant image regions and of fine granularity. With LEDs++, we empirically demonstrate that these maps can also capture regions of an image relevant to an editing concept that is not already present. Specifically for multiple edits, calculating a dedicated mask for each edit prompt ensures that the corresponding guidance terms remain largely isolated, limiting interference between them. + +Formally, at each time step $t$ , a U-Net forward pass with editing prompt $e_i$ is performed to generate cross-attention maps for each token of the editing prompt. All cross-attention maps of the smallest resolution (e.g., $16 \times 16$ for SD) are averaged over all heads and layers, and the resulting maps are summed over all editing tokens, resulting in a single map $A_t^{e_i} \in R^{16 \times 16}$ . Importantly, we utilize the same U-Net evaluation $\hat{\epsilon}_{\theta}(x_t, c_e)$ already performed in Eq. 10 to produce $M^1$ with minimal overhead. Each map $A_t^{e_i}$ is up-sampled to match the size of $x_t$ . Cross-attention mask $M^1$ is derived by calculating the $\lambda$ -th percentile of up-sampled $A_t^{e_i}$ and + +$$ +M _ {i} ^ {1} = \left\{ \begin{array}{l l} 1 & \text {i f} \left| A _ {t} ^ {e _ {i}} \right| \geq \eta_ {\lambda} \left(\left| A _ {t} ^ {e _ {i}} \right|\right) \\ 0 & \text {e l s e} \end{array} \right. \tag {13} +$$ + +where $\eta_{\lambda}(|\cdot|)$ is the $\lambda$ -th percentile. By definition, $M^1$ only selects image regions that correlate strongly with the editing prompt, and $\lambda$ determines the size of this selected region. + +The fine-grained mask $M^2$ is calculated based on the guidance vector $\psi$ of noise estimates derived in Eq. 10. The difference between unconditioned and conditioned $\hat{\epsilon}_{\theta}$ , generally captures outlines and object edges of $x_{t}$ . Consequently, the largest absolute values of $\psi$ provide meaningful segmentation information of fine granularity for $M^2$ + +$$ +M ^ {2} = \left\{ \begin{array}{l l} 1 & \text {i f} | \psi | \geq \eta_ {\lambda} (| \psi |) \\ 0 & \text {e l s e} \end{array} \right. \tag {14} +$$ + +In general, threshold $\lambda$ should correspond to the performed edit. Changes affecting the entire image, such as style transfer, should choose smaller $\lambda (\rightarrow 0)$ , whereas edits targeting + +![](images/aa8709af9cdd4c546356fdd95823f9d774fb44ad62023ae5dae09689d981550b.jpg) +Figure 2. Comparison of image editing methods. (top) LEDs++ is the only method to restrict edits to the tree leaves and position of the car. (bottom) Ours is the only approach faithfully executing all three edits and keeping changes minimal. (Best viewed in color) + +
MethodReconstruction +Error (RMSE) ↓Execution +Time (s) ↓Variation/ +SamplingSemantic +GroundingMulti- +Editing
SDEdit [26]0.81 ±0.072.10 ±0.02✓XX
Imagic [18]0.58 ±0.12349.98 ±0.45✓XX
Vanilla DDIM Inversion0.22 ±0.1037.23 ±0.04XXX
Pix2Pix-Zero [30]0.20 ±0.0956.78 ±0.14(✓)✓X
DiffEdit [9]0.13±0.0327.65 ±0.03✓✓X
Edit-friendly DDPM [17]0.0010.36 ±0.05✓XX
LEDSITS++ (Ours)0.001.78 ±0.03✓✓✓
+ +Table 1. Comparing key properties for diffusion-based image editing techniques, with LEDs++ offering clear methodological benefits. Due to LEDs++'s efficient perfect inversion, it is the fastest and error-free method. At the same time, its methodology is the only enabling versatility in terms of variation, semantic grounding, and multi-editing. Subscript numbers indicate standard deviation. + +specific objects or regions should use $\lambda$ proportional to the region's prominence in the image. + +# 4. Properties of LEDs++ + +With the fundamentals of LEDs++ established, we next showcase its unique properties and capabilities. + +Efficiency. First off, LEDs++ offers substantial performance improvements over other image editing methods. In Tab. 1, we provide a qualitative runtime comparison, with all methods being implemented for Stable Diffusion (SD) 1.5 [34]. As a parameter-free approach, LEDs++ does not require any computationally expensive fine-tuning or optimization. Consequently, LEDs++ is orders of magnitude faster than methods like Imagic [18] or Pix2Pix-Zero [30]. Further, we only need to invert the same number of diffusion steps used at inference, which results in significant runtime improvements over the standard DDIM inversion (21x). In addition to efficient inversion, we use a recent, fast scheduler that generally requires fewer total steps, further boosting performance. This way, LEDs++ is six times faster than recent DDPM inversion [17] and on par with fast but poor-quality SDEdit [26]. + +Versatility. In addition to its efficiency, LEDs++ remains versatile, enabling sheer limitless editing possibilities. In Fig. 1, we showcase a broad range of edit types. + +LEDITS++ facilitates fine-grained edits (adding/removing glasses) and holistic changes such as style transfer (painting/sketch). Furthermore, object removal and replacement facilitate even more image editing tasks. Importantly, the overall image composition is preserved in all cases. To our knowledge, LEDs++ is the only diffusion-based image editing method inherently supporting multiple edits in isolation, which allows for more complex image manipulation. Fig. 2 highlights LEDs++ benefits over previous methods. Our method produces the highest edit fidelity and is the only approach capable of faithfully executing multiple, simultaneous instructions. Moreover, LEDs++ also makes the least changes to unrelated objects and the overall background and composition of the image. + +Lastly, the editing versatility benefits from the stochastic nature of the perfect but non-deterministic inversion. LEDs++ provides meaningful image variations by resampling $\tilde{c}_t$ . Additionally, the visual expression of each concept in the edited image scales monotonically with $s_e$ , and the direction and magnitude of each concept can be varied freely. We present examples of both features in App. B. + +Precision. Furthermore, LEDs++'s methodology keeps edits concise and avoids unnecessary deviations from the input image (Fig. 2). First, the perfect inversion will reconstruct the exact input image if no edit is applied (cf. + +![](images/65145febf9dc8f5f3ea361c981b25ba7f5ed9a8b34e59e22ec6b15377923bcfe.jpg) +Figure 3. Exemplary edit performed with LEDs++ in only 25 diffusion steps with SD1.5. We apply a complex, compounded edit and ground each to a semantically reasonable image region. + +Sec. 3.2). Consequently, we already improve on faithfulness to the input image even before applying any edits. This benefit over other methods is highlighted by the reconstruction error in Tab. 1. Second, implicit masking will semantically ground each edit to relevant image regions. This is specifically important for editing multiple concepts at the same time. While other methods only utilize one prompt for all instructions, LEDs++ isolates edits from each other (Eq. 11). Thus, we get dedicated masks for each concept as shown in Fig. 3. This design ensures that each instruction (e.g., red mask for 'cherry blossom') will be only applied where necessary. Subsequently, we provide further evidence for the efficacy of LEDs++'s masking approach. + +# 5. Semantically Grounded Image Editing + +Cross-attention maps of DMs have been used extensively to ground regions of interest during image generation semantically [7, 8, 14, 30]. Nonetheless, these have not been combined with noise-based masks so far and thus lack fine granularity. Hence, we empirically evaluate the quality of implicit masks, i.e., attention maps $M^1$ and noise maps $M^2$ (Eq. 13 and 14) in the LEDs++ setup. We use a broad segmentation task for common objects as a proxy to measure the performance of implicit masks in identifying relevant image areas from edit instructions. Specifically, we utilize segmentation masks from the COCO panoptic segmentation challenge [19]. For each unique object in an image, we retrieve the masks $M^1$ , $M^2$ , and their intersection per diffusion step. We use the (semantic) class label (e.g. 'person' or 'TV') as editing concept $e$ . We consider masks at each of 50 total diffusion steps without actually editing the input image. Furthermore, we approximate mask threshold $\lambda$ based on the relative size of an object's bounding box. + +Concise masks with LEDs++. Fig. 4 shows implicit masking as a reliable means to identify relevant image regions. Importantly, the intersection of both cross-attention masks $M^1$ and noise maps $M^2$ clearly outperforms each separate mask. The overall performance is even similar to a dedicated CLIPSeg model [24], despite LEDs++. masks being implicitly calculated at inference with only minimal overhead. At the same time, LEDs++.'s masking is superior to DiffEdit's [3]. Consequently, our method's inter + +![](images/318c6382a39c079543b3cc4819e80b0f71060178d28b71ea7d436599a39f18cc.jpg) +Figure 4. Semantic segmentation quality of LEDs++. We show the intersection over union (higher is better) for COCO panoptic segmentation. The intersection masks outperform each by a clear margin, close to the CLIPSeg reference. (Best viewed in color) + +section of cross-attention masks and noise maps provides strong semantic grounding while being efficient during image manipulation to ensure precise editing. + +# 6. Image Editing Evaluation + +Let us now compare LEDs++ to current SOTA methods for image manipulation on two benchmarks. + +# 6.1. Editing Multiple Concepts + +First, we investigate the complex task of performing multiple edits simultaneously. We rely on a well-established setup for semantic image manipulation [5] to evaluate multi-conditioned attribute manipulation in facial images. In our experiment, we consider 100 images from the CelebA dataset [21]. For each image, we simultaneously edit three attributes out of a set of five, leading to ten total combinations of edit concepts. Further, we perform each edit across ten different seeds, resulting in 10,000 evaluated images for each method and hyperparameter setting, over 1M images in total. As measures for comparison, we employ CLIP and LPIPS scores. CLIP measures the text-to-image similarity of the edit instruction to the edited image, and LPIPS measures the image-to-image similarity of the real to the edited image. This way, we assess the tradeoff between the versatility of edits (CLIP) and the precision of those manipulations (LPIPS). We implement all methods based on SD1.5 and provide more details in App. C. + +LEDS++ outperforms competing methods. Fig. 5 shows the resulting CLIP vs. LPIPS plots for all methods. The top left corner represents the ideal editing method with maximum edit alignment without deviating from the initial image. Generally, one can observe a natural trade-off between versatility and precision for all methods, i.e., higher image-to-text alignment comes at the expense of lower similarity to the original image. LEDS++ is closest to the ideal region and thus clearly outperforms the other methods. In particular, the outputs remain close to the original image (low LPIPS scores), thanks to the precise implicit masking. + +![](images/110ffe195968c3d76f23e8c672ce8b00d27b845da1c2f8bcb04eed0058e662e0.jpg) +Figure 5. Comparison of instruction-alignment vs. image similarity trade-off for different editing methods. Results were reported for simultaneous manipulation of three facial attributes on CelebA. We plot CLIP scores (higher is better) of the target attributes against LPIPS similarity (lower is better). LEDs++ clearly outperforms all competing methods. (Best viewed in color) + +At the same time, it faithfully performs the edits (high CLIP scores) due to the dedicated, isolated editing for each concept. The depicted scores reflect our qualitative inspections for Pix2Pix-Zero and Imagic on such complex manipulations (cf. Fig. 2). We observed that these methods often break—either failing to perform all three edits and/or drastically altering the input image. Only edit-friendly DDPM [17] and LEDs++ reliably achieve the maximum average CLIP score of over 0.25. This value seems to represent an upper bound according to our manual investigations, as each attribute is edited correctly for all input images, and higher scores are not observed. Despite being computationally very efficient, LEDs++ faithfully executes each edit instruction while keeping the changes to the input low, highlighting the method's versatility and precision. + +# 6.2. TEDBench $(++)$ + +Next, we investigate the versatility of LEDs++'s editing capabilities by running the Textual Editing Benchmark (TEDBench [18]), a collection of 100 input images paired with textual edit instructions. However, we observed a variety of inconsistencies in TEDBench and a lack of relevant editing tasks. Therefore, we propose TEDBench++ (Fig. 6a and App. D), a more challenging revised benchmark now containing 120 entries in total.3 We addressed misspellings and rephrased ambiguous and inconclusive instructions. In addition to resolving these issues, we added instructions targeting challenging types of image manipulations previously not included in TEDBench: multi-conditioning, object/concept removal, style transfer, and complex replacements (Fig. 6a). We provide more details in App. D. + +We compare LEDs++ on TEdBench $(++)$ to one of the strongest editing methods, Imagic [18] with Imagen [37]. + +
TEDBenchTEDBench++
SR ↑LPIPS ↓SR ↑LPIPS ↓
Imagic w/ SD1.50.550.560.580.57
LEDITS++ w/ SD1.50.750.280.790.30
Imagic w/Imagen [18]0.830.59——
LEDITS++ w/ SD-XL0.840.330.870.34
+ +Table 2. Success rate (SR) and LPIPS scores on the original TedBench [18] and our revised version (TEDBench++) We compare Imagic to LEDs++ based on different DMs and find the latter to outperform on both metrics and benchmarks. + +Since both are not publicly available, we can only compare to this specific combination of DM and editing method using Kawar et al.'s [18] curated outputs for TEdBench. Additionally, we, therefore, cannot combine LEDs++ with Imagen[37] and instead use a similarly advanced diffusion model, SD-XL [32]. However, to not only compete for the best fidelity outputs but focus the evaluation on methodological differences—not the pre-trained DM—we also compare both methods implemented with SD1.5. We provide further details in App. C. + +LEDS $^{+ + }$ edits images reliably. We first asked users to assess the overall success of edits, i.e., if an edit instruction was faithfully realized for a given input image. The results in Tab. 2 show that LEDs $^{+ + }$ outperforms Imagc on TEdBench despite a greatly reduced runtime (Tab. 1). The difference is even stronger when comparing both methods on the same pre-trained DM, i.e., SD1.5. The high success rate on TEdBench $^{+ + }$ $(87\%)$ and the examples shown in Fig. 1 and 6a once again highlight LEDs $^{+ + }$ 's versatility. Overall, our proposed method can reliably perform a diverse set of editing instructions for real images. + +High-quality edits with LEDs++. While investigating both methods' performance we observed a substantial difference in edit quality. The examples in Fig. 6b particularly highlight the discrepancy in compositional robustness and object coherence. Hence, we also assessed both methods' editing quality on TEdBench++. We focus on samples where both methods performed a successful edit, i.e., were labeled as successful by users. We show the perceptual similarity (LPIPS) to the input image in Tab. 2. One can observe that the LPIPS scores for LEDs++ are much lower than for Imagic, empirically supporting the qualitative examples in Fig. 6b. When manually inspecting the generated images, we often found Imagic to generate a completely new image based on the edit instruction, entirely disregarding the input image (cf. App. Fig. 11a). + +# 7. Discussion + +Let us now discuss open research questions and limitations. + +Model Dependency. While LEDs++ achieves impressive results on a large variety of image manipulation tasks, there are external factors to consider. Since the + +![](images/be4dfb3e9df3f04561b13dffc758c1f1e4f3c2a54aa97cbf03fd80af2b3c79db.jpg) +(a) Novel challenging examples of TEdBench++ and LEDs++ applied, showcasing the versatility of supported edits. + +![](images/80bbbd9c3d251bbf51c623aac6607686525a4f1e1fe8d544eeb4757123cb27e6.jpg) +(b) Qualitative comparison of LEDs++ and Imagic on TEd-Bench, clearly highlighting the performance improvement. +Figure 6. Benchmark examples for LEDs++ and Imagc on TEdBench $(++)$ . (Best viewed in color) + +method is architecture-agnostic, it can be easily used with any DM. At the same time, the general editing quality strongly depends on the overall capabilities of the underlying pre-trained DM. Naturally, more capable models will also enable better edits. But, at times, specific editing instructions may fail because the used DM does not have a decent representation of the targeted concept to begin with. One example is the model failing to edit a giraffe to be sitting since the underlying DM generally fails to generate this pose (cf. App. F). This effect can also clearly be seen in Tab. 2, with the editing success rate of a method varying strongly between DMs. Although the same image editing method is employed (LEDITS++, the more capable SD-XL variant outperforms the weaker SD1.5 model. Nonetheless, this means that the architecture-agnostic LEDs++ will benefit from increasingly powerful DMs. + +Coherence Trade-offs. Next to the benefits of LEDs++'s semantic grounding, there are also downsides to this approach. Overall, implicit masking limits changes to relevant portions of the image and achieves strong coherence with the original image composition. Yet, the object and its identity within the masked area may change based on various factors. Generic prompts, like "a standing cat" (cf. App. F), do not contain detailed information about this specific object ("cat"). Thus, an edit with this prompt does not guarantee to preserve object identity, particularly for strong hyperparameters. We observed that fine-tuning approaches like Imagic make the opposite trade-off, better preserving the object identity while changing the background and image composition substantially (cf. App. F). A potential remedy for a loss in object coherence with LEDs++, is more descriptive edit prompting, e.g. using textual inversion [12]. + +Lastly, the automatically-inferred implicit masks allow for easy use of LEDs++ without users tediously providing masks. Nonetheless, user intentions are diverse and cannot always be automatically inferred. Sometimes, in- + +individual user masks provide better control over the editing process. Such user masks can be easily integrated into LEDs++ (cf. App. F), wherefore we encourage future research in this promising direction. + +Societal Impact. LEDs++ is an easy-to-use image editing technique that lowers the barrier for users and puts them in control for fruitful human-machine collaboration. Yet, the underlying text-to-image models offer both promise and peril, as highlighted by prior research [4, 11]. The (societal) biases within these models will also impact image editing applications [11]. Moreover, image manipulation can also be used adversarially to generate inappropriate [39] or fake content. Hence, we advocate for a cautious deployment of generative models together with image editing methods. + +# 8. Conclusion + +We introduced LEDs++, an efficient yet versatile and precise method for textual image manipulation with diffusion models. It facilitates the editing of complex concepts in real images. Our approach requires no finetuning nor optimization, can be computed extremely efficiently, and is architecture agnostic. At the same time, it perfectly reconstructs an input image and uses implicit masking to limit changes to relevant image regions, thus editing precisely. Our large experimental evaluation confirms the efficiency, versatility, and precision of LEDs++ and its components, as well as its benefits over several related methods. + +Acknowledgements We gratefully acknowledge support from the BMBF (Grant No 01IS22091). This work benefited from the ICT-48 Network of AI Research Excellence Center "TAILOR" (EU Horizon 2020, GA No 952215), the Hessian research priority program LOEWE within the project WhiteBox, the HMWK cluster projects "Adaptive Mind" and "Third Wave of AI", and from the NHR4CES. + +# References + +[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. SIGGRAPH Asia, 2023. 16 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv:2206.02779, 2022. 2 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv:2211.01324, 2022. 2, 6 +[4] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. In Proceedings of ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2023. 8, 11 +[5] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. In Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 6 +[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv:2304.08465, 2023. 2, 6 +[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42, 2023. 2, 6 +[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5 +[10] Zoe De Simone, Angie Boggust, Arvind Satyanarayan, and Ashia Wilson. What is a Fair Diffusion Model? Designing Generative Text-To-Image Models to Incorporate Various Worldviews. arXiv preprint arXiv:2309.09944, 2023. 11 +[11] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness. arXiv preprint arXiv:2302.10893, 2023. 8, 11 +[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8, 16 +[13] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion + +models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 11 +[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 6 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and HsuanTien Lin, editors, Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 3 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 3 +[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 1, 2, 3, 4, 5, 7, 12, 13 +[18] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 7, 12, 13, 14 +[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 6 +[20] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2 +[21] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), December 2015. 6 +[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 2, 3 +[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3 +[24] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6 +[25] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv:2310.04378, 2023. 2 +[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 1, 2, 4, 5, 12, 13 +[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P. + +Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 12 +[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2022. 2, 11 +[30] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 12, 13 +[31] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv:2306.00637, 2023. 2 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 7 +[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2, 11 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5 +[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 3 +[36] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V. Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased Face Synthesis With Diffusion Models: Are We There Yet? arXiv preprint arXiv:2309.07277, 2023. 11 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv:2205.11487, 2022. 2, 7, 11 +[38] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. arXiv:2304.14530, 2023. 17 +[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8, 11 + +[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of NeurIPS Datasets and Benchmarks, 2022. 11 +[41] Norbert Schwarz. How the questions shape the answers. American Psychologist, 1999. 15 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 3 +[43] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3 +[44] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv:2210.09477, 2022. 1, 2 +[45] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv:1212.08698, 2022. 2 +[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv:2302.04867, 2023.2 \ No newline at end of file diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/images.zip b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f319310dec5ceddaf3e7562a60da00f37f4212f6 --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9998c9735e5a8435cb41c863ee7f5981af094f5c8f2f7be87e0a926376a6ba5d +size 577229 diff --git a/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/layout.json b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ca5ed4903e51920238179b9c9fb8878b18f2865c --- /dev/null +++ b/2024/LEDITS++_ Limitless Image Editing using Text-to-Image Models/layout.json @@ -0,0 +1,9177 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "text", + "content": "LEDS: Limitless Image Editing using Text-to-Image Models" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 75, + 141, + 517, + 203 + ], + "blocks": [ + { + "bbox": [ + 75, + 141, + 517, + 203 + ], + "lines": [ + { + "bbox": [ + 75, + 141, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 75, + 141, + 517, + 203 + ], + "type": "image", + "image_path": "d8f7be564b6332796a8cb8adaad3a2d7a5f0c215956e770c5c1a33382a955dd0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 86, + 209, + 509, + 347 + ], + "blocks": [ + { + "bbox": [ + 86, + 209, + 509, + 347 + ], + "lines": [ + { + "bbox": [ + 86, + 209, + 509, + 347 + ], + "spans": [ + { + "bbox": [ + 86, + 209, + 509, + 347 + ], + "type": "image", + "image_path": "fd68bca424e2d5b2224b8fd297e5ce3f861d71132db180895b57fbeebd7121b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 352, + 489, + 364 + ], + "lines": [ + { + "bbox": [ + 102, + 352, + 489, + 364 + ], + "spans": [ + { + "bbox": [ + 102, + 352, + 489, + 364 + ], + "type": "text", + "content": "Figure 1. LEDs++ facilitates versatile image-to-image editing. Several complex cases are available now." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "spans": [ + { + "bbox": [ + 143, + 373, + 192, + 386 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 290, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 290, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 290, + 641 + ], + "type": "text", + "content": "Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDs++, an efficient yet versatile and precise textual image manipulation technique. LEDs++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDs++ and its improvements over previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 128, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Text-to-image diffusion models (DM) have garnered recognition for their ability to generate high-quality images from" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 374, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 541 + ], + "type": "text", + "content": "textual descriptions. A growing body of research has recently been dedicated to utilizing these models for manipulating real images. However, several barriers prevent many real-world applications of diffusion-based image editing. Current methods often entail computationally expensive model tuning or other optimization, presenting practical challenges [6, 18, 28, 30, 44]. Additionally, existing techniques have the proclivity to induce profound changes to the original image [17, 26], often resulting in completely different images. Lastly, all these approaches are inherently constrained when editing multiple (arbitrary) concepts simultaneously. We tackle these problems by introducing LEDs++" + }, + { + "bbox": [ + 304, + 374, + 545, + 541 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 374, + 545, + 541 + ], + "type": "text", + "content": ", a diffusion-based image editing technique addressing these limitations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 542, + 546, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 546, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 546, + 674 + ], + "type": "text", + "content": "LEDITS++2 offers a streamlined approach for textual image editing, eliminating the need for extensive parameter tuning. To this end, we derive image inversion for a more efficient diffusion sampling algorithm to a) drastically reduce computational resources and b) guarantee perfect image reconstruction. Thus, we overcome computational obstacles and avoid changes in the edited image in the first place. Furthermore, we use a novel implicit masking approach to semantically ground each edit instruction to its relevant image region. This further optimizes changes to the image by retaining the overall image composition and object identity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 683, + 541, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 318, + 683, + 525, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 683, + 525, + 693 + ], + "spans": [ + { + "bbox": [ + 318, + 683, + 525, + 693 + ], + "type": "text", + "content": "*Equal contribution †Partially as research intern at Adobe" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 693, + 521, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 693, + 521, + 703 + ], + "spans": [ + { + "bbox": [ + 318, + 693, + 521, + 703 + ], + "type": "text", + "content": "1LEDITS++ stands for Limitless Edits with sde-dpm-solver++." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 703, + 541, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 703, + 541, + 712 + ], + "spans": [ + { + "bbox": [ + 318, + 703, + 541, + 712 + ], + "type": "text", + "content": "2https://huggingface.co/spaces/leditsplusplus/project" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8861" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 144 + ], + "type": "text", + "content": "Additionally, LEDs++ is the only method to date to facilitate easy and versatile image editing by supporting multiple simultaneous instructions without causing undue interference. Finally, its lightweight architecture-agnostic nature ensures compatibility with both latent and pixel-based diffusion models, providing high accessibility." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 144, + 286, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 144, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 144, + 286, + 286 + ], + "type": "text", + "content": "In this work, we establish the methodical benefits of LEDs++ and demonstrate that this intuitive, lightweight approach offers sophisticated semantic control for image editing. Specifically, we contribute by (i) devising a formal definition of LEDs++ while (ii) deriving perfect inversion for a more efficient diffusion sampling method, (iii) qualitatively and empirically demonstrating its efficiency, versatility, and precision, (iv) providing an exhaustive empirical comparison to concurrent works with automatic and human user metrics, and thereby (v) introducing Textual Editing Benchmark++ (TEdBench++, a more holistic and coherent testbed for evaluating textual image manipulation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 298, + 124, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 124, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 124, + 312 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 319, + 286, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 286, + 377 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 286, + 377 + ], + "type": "text", + "content": "Recently, large-scale, text-guided DMs have enabled versatile applications in image generation [3, 33, 37]. Especially latent diffusion models [31, 34] have gained attention for their computational efficiency. Below, we discuss related work for efficient, versatile image manipulation with DMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 379, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 379, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 379, + 286, + 557 + ], + "type": "text", + "content": "Diffusion Sampling. Generating outputs with DMs requires multiple iterative denoising steps that constitute the main bottleneck at inference. Commonly used sampling methods such as DDPM [15] or DDIM [42] require tens or hundreds of steps to produce high-quality samples. Consequently, numerous works have been dedicated to speeding up the sampling process without loss in quality. Distillation efforts progressively reduce the number of required steps through further training [25, 27]. Other works focus on improving the sampling itself, e.g. using high-order ODE-solvers [22, 23, 46]. Such solvers can be readily combined with pre-trained DMs at inference to lower the number of denoising steps. With LEDs++, we derive perfect image inversion with the DPM-Solver++, allowing image editing in as few as 20 total steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 712 + ], + "type": "text", + "content": "Semantic Control during Diffusion. While text-to-image DMs generate new, astonishing images, fine-grained control over the generative process remains challenging. Minor changes to the text prompt lead to entirely different outputs. Wu et al. [45] studied concept disentanglement using linear combinations of text embeddings to gain semantic control. Methods like Prompt-to-Prompt [14] and other works [8, 30] utilize the DM's attention layers to attribute pixels to tokens from the text prompt. Dedicated operations on the attention maps enable more control over the generated images. Other works have focused on the noise estimates of DMs [5, 20] providing semantic control over the generation process. With LEDs++, we now enable" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 95 + ], + "type": "text", + "content": "fine-grained semantic control for manipulating real images, going beyond purely generative applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 96, + 544, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 544, + 275 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 544, + 275 + ], + "type": "text", + "content": "Real Image Editing. Since DMs' rise in popularity for text-to-image generation, they have also been explored for (real) image-to-image editing. As a first, simple approach, SDEdit added noise to the image for an intermediate step in the diffusion process [26]. While lightweight, the resulting image diverges substantially from the input as it is (partially) regenerated. Inpainting allows to keep the change small by having a user provide additional masks to restrict changes to certain image regions [2, 29]. Yet, user masks are costly or often simply unavailable. Other works have thus explored semantically grounded approaches using cross-attention instead to better control image manipulation [7, 28, 30]. In contrast, LEDs++ leverages both attention-and noise-based masking to obtain fine-grained masks, enabling strong semantic control over real images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 276, + 544, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 544, + 526 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 544, + 526 + ], + "type": "text", + "content": "Another important aspect of image manipulation methods is the required tuning and overall runtime. Instruct-Pix2Pix continues training a DM at scale to enable image editing capabilities [6]. Finetuning instead on each individual input to constrain the generation on the real image has shown helpful [18, 44] but not computationally efficient. Consequently, recent works have largely relied on inverting the deterministic DDIM sampling process [42] to save computational resources. DDIM inversion identifies an initial noise vector that results in the input image when denoised again. However, faithful reconstructions are only obtained in the limit of small steps, thus requiring large numbers of inversion steps. Moreover, small errors will still incur at each timestep, often accumulating into meaningful deviations from the input, requiring costly error correction through optimization [28, 30]. Recently, Huberman-Spiegelglas et al. proposed an inversion technique [17] for the DDPM sampler [15] to address the limitations of DDIM inversion. LEDs++ provides the same guarantees of perfect inversion with even further reduced runtime alongside an edit-friendly latent space, enabling more versatility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 537, + 533, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 533, + 551 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 533, + 551 + ], + "type": "text", + "content": "3. Image Editing with Text-to-Image Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 558, + 544, + 712 + ], + "type": "text", + "content": "Before devising the methodology of LEDs++, let us first motivate the desired features and use cases. Specifically, we aim for efficiency, versatility, and precision. The goal is to provide a method that enables a fast exploratory workflow for image editing in which a user can iteratively interact with the model and explore various edits. Consequently, LEDs++ produces outputs quickly with no tuning or optimization to not disrupt the creative process. Further, arbitrary editing instructions and combinations thereof are supported to facilitate a wide range of image manipulations (e.g., complex multi-editing). Lastly, we provide precise and sophisticated semantic control over the image editing. Each of the (potentially multiple) edit instruc" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "8862" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "tions can be steered individually, and changes are automatically restricted to relevant image regions. Importantly, with LEDs++ we prioritize compositional robustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 115, + 188, + 126 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 188, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 188, + 126 + ], + "type": "text", + "content": "3.1. Guided Diffusion Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 133, + 287, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 133, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 47, + 133, + 287, + 193 + ], + "type": "text", + "content": "Let us first define some general background for diffusion models. DMs iteratively denoise a Gaussian distributed variable to produce samples of a learned data distribution. Let's consider a diffusion process that gradually turns an image " + }, + { + "bbox": [ + 47, + 133, + 287, + 193 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 133, + 287, + 193 + ], + "type": "text", + "content": " into Gaussian noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 201, + 287, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 201, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 70, + 201, + 287, + 216 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {1 - \\beta_ {t}} x _ {t - 1} + \\sqrt {\\beta_ {t}} n _ {t}, \\quad t = 1, \\dots , T \\tag {1}", + "image_path": "01babbd6535872ea79c5ded55b7f7137b8e7e318dae56b0ea452f2ff761b4670.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "content": " are iid normal distributed vectors and " + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\beta_t" + }, + { + "bbox": [ + 47, + 223, + 287, + 247 + ], + "type": "text", + "content": " a variance schedule. The diffusion process is equivalently expressed as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 254, + 287, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 254, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 111, + 254, + 287, + 268 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {2}", + "image_path": "01f04f233a4898894c80d9553f1c28bca1095e5613900eb01f2a6567abe62100.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\Pi_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": ". Importantly, all " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": " are not statistically independent. Instead, consecutive pairs " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}, \\epsilon_{t-1}" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": " are strongly dependent, which will be relevant later. To generate an (new) image " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\hat{x}_{0}" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": " the reverse diffusion process starts from random noise " + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "inline_equation", + "content": "x_{T} \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 47, + 275, + 287, + 346 + ], + "type": "text", + "content": " which can be iteratively denoised as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 355, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 355, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 85, + 355, + 287, + 368 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {3}", + "image_path": "6cb839a35d00e24f82753bd19756ba0f2263f9e6d200744ad523531b7368e4ee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "text", + "content": " are iid standard normal vectors, and common variance schedulers " + }, + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 47, + 376, + 287, + 399 + ], + "type": "text", + "content": " can be expressed in the general form" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 406, + 207, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 406, + 207, + 430 + ], + "spans": [ + { + "bbox": [ + 126, + 406, + 207, + 430 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} = \\eta \\beta_ {t} \\frac {1 - \\bar {\\alpha} _ {t - 1}}{1 - \\bar {\\alpha} _ {t}}", + "image_path": "93ce10a700c5c92f1c79d0770462afd2c2970c56cda56a16c1e439686c92175e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": ". In this formulation, " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " corresponds to the deterministic DDIM [42] and " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "text", + "content": " to the DDPM scheme [15]. Lastly, in theses cases, we have " + }, + { + "bbox": [ + 47, + 437, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t(x_t) =" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 480, + 278, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 278, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 278, + 507 + ], + "type": "interline_equation", + "content": "\\sqrt {\\bar {\\alpha} _ {t - 1}} \\frac {x _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\hat {\\epsilon} _ {\\theta} (x _ {t})}{\\sqrt {\\bar {\\alpha} _ {t}}} + \\sqrt {1 - \\bar {\\alpha} _ {t - 1} - \\sigma_ {t} ^ {2}} \\hat {\\epsilon} _ {\\theta} (x _ {t})", + "image_path": "b36ede1384ad1d0819666d3425636e9bd15f2ea04cb7fac9e7a0ff55286cd9a2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_{t})" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " is an estimate of " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " produced by our neural network DM with learned parameters " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": ", commonly implemented as a U-Net [35]. For text-to-image generation, the model is conditioned on a text prompt " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " to produce images faithful to that prompt. The DM is trained to produce the noise estimate " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_{t})" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " needed for iteratively sampling " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " (Eq. 3). For text-conditioned DMs, " + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 47, + 514, + 287, + 609 + ], + "type": "text", + "content": " is calculated using specific guidance techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "text", + "content": "Most DMs rely on classifier-free guidance [16], a conditioning method using a purely generative diffusion model, eliminating the need for an additional classifier. During training, the text conditioning " + }, + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "inline_equation", + "content": "c_{p}" + }, + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "text", + "content": " is randomly dropped with a fixed probability, resulting in a joint model for unconditional and conditional objectives. During inference, the score estimates for the " + }, + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 609, + 287, + 693 + ], + "type": "text", + "content": "-prediction are adjusted so that:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 71, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + s _ {g} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {p}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\tag {4}", + "image_path": "e890a993a917c07c105c7f3a66b7cade03636d3fae76e076e2eb08fba897f629.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": "with guidance scale " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "s_g" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " defining the noise estimate with parameters " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": ". Intuitively, the unconditioned " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": "-prediction is pushed in the direction of the conditioned one, with " + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "s_g" + }, + { + "bbox": [ + 305, + 72, + 545, + 121 + ], + "type": "text", + "content": " determining the extent of the adjustment." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 127, + 375, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 375, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 375, + 138 + ], + "type": "text", + "content": "3.2.LEDITS++" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 145, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 193 + ], + "type": "text", + "content": "With the fundamentals established, the methodology of LEDs++ can now be broken down into three components: (1) efficient image inversion, (2) versatile textual editing, and (3) semantic grounding of image changes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "text", + "content": "Component 1: Perfect Inversion. Utilizing text-to-image models for editing real images requires conditioning the generation on the input image. Recent works have largely relied on inverting the sampling process to identify " + }, + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "text", + "content": " that will be denoised to the input image " + }, + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 305, + 194, + 545, + 289 + ], + "type": "text", + "content": " [28, 30]. Inverting the DDPM scheduler is generally preferred over DDIM inversion since the former can be achieved in fewer timesteps and with no reconstruction error [17]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 289, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 433 + ], + "type": "text", + "content": "However, there exist more efficient schemes than DDPM for sampling DMs that greatly reduce the required number of steps and consequently DM evaluations. We here propose a more efficient inversion method by deriving the desired inversion properties for such a scheme. As demonstrated by Song et al.[43], DDPM can be viewed as a first-order stochastic differential equation (SDE) solver when formulating the reverse diffusion process as an SDE. This SDE can be solved more efficiently—in fewer steps—using a higher-order differential equation solver, here dpmsolver++ [23]. The reverse diffusion process from Eq. 3 for the second-order sde-dpm-solver++ can be written as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 441, + 545, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 441, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 332, + 441, + 545, + 453 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\hat {\\mu} _ {t} \\left(x _ {t}, x _ {t + 1}\\right) + \\sigma_ {t} z _ {t}, \\quad t = T, \\dots , 1 \\tag {5}", + "image_path": "556b24933fca24c76e6a97c877ddcd7ced23f827d3bce0592f52b5dd5ddd5e58.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 463, + 353, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 463, + 353, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 463, + 353, + 472 + ], + "type": "text", + "content": "where now" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 359, + 479, + 492, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 479, + 492, + 495 + ], + "spans": [ + { + "bbox": [ + 359, + 479, + 492, + 495 + ], + "type": "interline_equation", + "content": "\\sigma_ {t} = \\sqrt {1 - \\bar {\\alpha} _ {t - 1}} \\sqrt {1 - e ^ {- 2 h _ {t - 1}}}", + "image_path": "88ca22d5780b304233063ee9afcb1cc1678d87102b2a5eedf11b0e026d9638c6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "content": "and higher-order " + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t" + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "content": " depends now on " + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "content": " from two timesteps, " + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "inline_equation", + "content": "x_{t + 1}" + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "text", + "content": ". Such that " + }, + { + "bbox": [ + 305, + 503, + 544, + 528 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_t(x_t,x_{t + 1}) =" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 534, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 534, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 534, + 588 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t - 1}}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} e ^ {- h _ {t - 1}} x _ {t} + \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\hat {\\epsilon} _ {\\theta} (x _ {t}) \\\\ + 0. 5 \\sqrt {\\bar {\\alpha} _ {t - 1}} \\left(1 - e ^ {- 2 h _ {t - 1}}\\right) \\frac {- h _ {t}}{h _ {t - 1}} \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t + 1}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\\\ \\end{array}", + "image_path": "25ea24f542e8c62c84e2074cf3de9384b09bdc6fb940e1219ce82ac3f8b63032.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 594, + 327, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 594, + 327, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 594, + 327, + 605 + ], + "type": "text", + "content": "with" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 347, + 603, + 504, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 603, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 347, + 603, + 504, + 630 + ], + "type": "interline_equation", + "content": "h _ {t} = \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t}})} - \\frac {\\ln (\\sqrt {\\bar {\\alpha} _ {t + 1}})}{\\ln (\\sqrt {1 - \\bar {\\alpha} _ {t + 1}})}", + "image_path": "428b549b5ee709a4756038fc8408a094b6ec60c0e2a679faf9d60f62cf6d19b8.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "text", + "content": "For the detailed derivation of the solver and proof of faster convergence, we refer the reader to the relevant literature [22, 23]. Based on the above, we now devise our inversion process. Given an input image " + }, + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "text", + "content": " we construct an auxiliary reconstruction sequence of noise images " + }, + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "inline_equation", + "content": "x_1, \\ldots, x_T" + }, + { + "bbox": [ + 305, + 633, + 545, + 693 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 370, + 700, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 700, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 370, + 700, + 545, + 713 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\tilde {\\varepsilon} _ {t} \\tag {6}", + "image_path": "40b8cb9bd836f66ea80a255eb3e747c071971f748b671b590e2873713f7b4d55.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8863" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ". Contrary to Eq. 2, the " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " are now statistically independent, which is a desirable property for image editing [17]. Lastly, the respective " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " for the inversion can be derived from Eq. 5 as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 128, + 287, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 128, + 287, + 153 + ], + "spans": [ + { + "bbox": [ + 78, + 128, + 287, + 153 + ], + "type": "interline_equation", + "content": "z _ {t} = \\frac {x _ {t - 1} - \\hat {\\mu} _ {t} \\left(x _ {t} , x _ {t + 1}\\right)}{\\sigma_ {t}}, \\quad t = T,.., 1 \\tag {7}", + "image_path": "1b3cac44c0e86528e3cb58fada579b56ed1edf25a2efda63f3ce912c7d0411d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": " as defined above. We base our implementation on the multistep variant of sde-dpm-solver++, which only requires one evaluation of the DM at each diffusion timestep by reusing the estimates from the previous step. The number of timesteps can be reduced further by stopping the inversion at an intermediate step " + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "inline_equation", + "content": "t < T" + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": " and starting the generation at that step. Empirically, we observed that " + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "inline_equation", + "content": "t\\in [0.9T,0.8T]" + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": " usually produces edits of the same fidelity as " + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 46, + 160, + 287, + 280 + ], + "type": "text", + "content": ", supporting observations in previous work [17, 26] that earlier timesteps are less relevant to the edit." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": "Component 2: Textual Editing. After creating our reconstruction sequence " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "x_{1}, \\ldots, x_{T}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " and calculating the respective " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ", we now edit the image by manipulating the noise estimate " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " based on a set of edit instructions " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i \\in I}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ". We devise a dedicated guidance term for each concept " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "e_{i}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " based on conditioned and unconditioned estimates. Let us formally define LEDs++'s guidance by starting with a single editing prompt " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ". We compute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 385, + 287, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 385, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 101, + 385, + 287, + 398 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) := \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) + \\gamma \\left(x _ {t}, c _ {e}\\right) \\tag {8}", + "image_path": "1c10fd84657d82e77f366d0388320cf74ae940b7bd32424b8ad1ad5dc139b285.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": "with guidance term " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": ". Consequently, setting " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\gamma = 0" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": " will reconstruct the input image " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": ". We construct " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": " to push the unconditioned score estimate " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_t)" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": " —i.e. the input image reconstruction—away from/towards the edit concept estimate " + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_t, c_e)" + }, + { + "bbox": [ + 46, + 407, + 287, + 467 + ], + "type": "text", + "content": ", depending on the guidance direction:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 476, + 287, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 476, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 100, + 476, + 287, + 489 + ], + "type": "interline_equation", + "content": "\\gamma \\left(x _ {t}, c _ {e}\\right) = \\phi (\\psi ; s _ {e}, \\lambda) \\psi \\left(x _ {t}, c _ {e}\\right) \\tag {9}", + "image_path": "d8ef37f970fdac9040119cfb228f598a3069a5f51ca466106878940f0fccde24.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " applies an edit guidance scale " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " element-wise, and " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "text", + "content": " depends on the edit direction: " + }, + { + "bbox": [ + 47, + 498, + 287, + 522 + ], + "type": "inline_equation", + "content": "\\psi(x_t, c_e) =" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 531, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 531, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 64, + 531, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right) \\quad \\text {i f p o s . g u i d a n c e} \\\\ - \\left(\\hat {\\epsilon} _ {\\theta} \\left(x _ {t}, c _ {e}\\right) - \\hat {\\epsilon} _ {\\theta} \\left(x _ {t}\\right)\\right) \\quad \\text {i f n e g . g u i d a n c e} \\end{array} \\right. \\tag {10}", + "image_path": "898e26c6ee46d7fd361ff6afdd5645b910febb546dc31cd8e56e9ef8dfba9b4b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "Thus, changing the guidance direction is reflected by the direction between " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_t,c_e)" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_t)" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " identifies those dimensions of the image and respective " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " that are relevant to a prompt " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " returns 0 for all irrelevant dimensions and a scaling factor " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " for the others. We describe the construction of " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " in detail below. Larger " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " will increase the effect of the edit, and " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\lambda \\in (0,1)" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " reflects the percentage of the pixels selected as relevant by " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": ". Notably, for a single concept " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": " and uniform " + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\phi = s_e" + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": ", Eq. 8 generalizes to the classifier-free guidance term in Eq. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "For multiple " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", we calculate " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_t^i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " as described above, with each defining their own hyperparameter values " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\lambda^i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "s_e^i" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 400, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 400, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 400, + 84 + ], + "type": "text", + "content": "sum of all " + }, + { + "bbox": [ + 306, + 72, + 400, + 84 + ], + "type": "inline_equation", + "content": "\\gamma_t^i" + }, + { + "bbox": [ + 306, + 72, + 400, + 84 + ], + "type": "text", + "content": " results in" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 91, + 545, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 91, + 545, + 111 + ], + "spans": [ + { + "bbox": [ + 361, + 91, + 545, + 111 + ], + "type": "interline_equation", + "content": "\\hat {\\gamma} _ {t} \\left(x _ {t}, c _ {e _ {i}}\\right) = \\sum_ {i \\in I} \\gamma_ {t} ^ {i} \\left(x _ {t}, c _ {e _ {i}}\\right) \\tag {11}", + "image_path": "346d1129b6d9e4c283682194ffabff8e77b3a286c42e34c6f54ac94d77a600a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "content": "Component 3: Semantic Grounding. The masking term " + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "content": " (Eq. 9) is the intersection (pointwise product) of binary masks " + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "content": " combined with scaling factor " + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 305, + 116, + 545, + 152 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 159, + 545, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 159, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 372, + 159, + 545, + 172 + ], + "type": "interline_equation", + "content": "\\phi (\\psi ; s _ {e _ {i}}, \\lambda) = s _ {e _ {i}} M _ {i} ^ {1} M _ {i} ^ {2} \\tag {12}", + "image_path": "0e7cf2c08aa6d608c651dac1cbea1edc40f74c06e5888a74defed38ac17dcfc8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "inline_equation", + "content": "M_{i}^{1}" + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "content": " is a binary mask generated from the U-Net's cross-attention layers and " + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "inline_equation", + "content": "M_{i}^{2}" + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "content": " is a binary mask derived from the noise estimate. Intuitively, each mask is an importance map, where " + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "inline_equation", + "content": "M_{i}^{1}" + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "content": " is more strongly grounded than " + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "inline_equation", + "content": "M_{i}^{2}" + }, + { + "bbox": [ + 304, + 179, + 545, + 334 + ], + "type": "text", + "content": ", but of significantly coarser granularity. Therefore, the intersection of the two yields a mask both focused on relevant image regions and of fine granularity. With LEDs++, we empirically demonstrate that these maps can also capture regions of an image relevant to an editing concept that is not already present. Specifically for multiple edits, calculating a dedicated mask for each edit prompt ensures that the corresponding guidance terms remain largely isolated, limiting interference between them." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": "Formally, at each time step " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": ", a U-Net forward pass with editing prompt " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " is performed to generate cross-attention maps for each token of the editing prompt. All cross-attention maps of the smallest resolution (e.g., " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " for SD) are averaged over all heads and layers, and the resulting maps are summed over all editing tokens, resulting in a single map " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "A_t^{e_i} \\in R^{16 \\times 16}" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": ". Importantly, we utilize the same U-Net evaluation " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}(x_t, c_e)" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " already performed in Eq. 10 to produce " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " with minimal overhead. Each map " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "A_t^{e_i}" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " is up-sampled to match the size of " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": ". Cross-attention mask " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " is derived by calculating the " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": "-th percentile of up-sampled " + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "inline_equation", + "content": "A_t^{e_i}" + }, + { + "bbox": [ + 304, + 335, + 545, + 479 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 352, + 486, + 545, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 486, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 352, + 486, + 545, + 518 + ], + "type": "interline_equation", + "content": "M _ {i} ^ {1} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\left| A _ {t} ^ {e _ {i}} \\right| \\geq \\eta_ {\\lambda} \\left(\\left| A _ {t} ^ {e _ {i}} \\right|\\right) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {13}", + "image_path": "651bac0fa781bc9a3734b7679041fb619693dcd97469720a3b8064f43dbe1a24.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\eta_{\\lambda}(|\\cdot|)" + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "content": "-th percentile. By definition, " + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "content": " only selects image regions that correlate strongly with the editing prompt, and " + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 525, + 545, + 560 + ], + "type": "text", + "content": " determines the size of this selected region." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": "The fine-grained mask " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": " is calculated based on the guidance vector " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": " of noise estimates derived in Eq. 10. The difference between unconditioned and conditioned " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": ", generally captures outlines and object edges of " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": ". Consequently, the largest absolute values of " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "text", + "content": " provide meaningful segmentation information of fine granularity for " + }, + { + "bbox": [ + 304, + 561, + 545, + 633 + ], + "type": "inline_equation", + "content": "M^2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 361, + 639, + 545, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 639, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 361, + 639, + 545, + 671 + ], + "type": "interline_equation", + "content": "M ^ {2} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} | \\psi | \\geq \\eta_ {\\lambda} (| \\psi |) \\\\ 0 & \\text {e l s e} \\end{array} \\right. \\tag {14}", + "image_path": "fa627c236f73614d4b74f709db76a35bc07871ad2222cd288ab4369e1159b1fa.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "In general, threshold " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " should correspond to the performed edit. Changes affecting the entire image, such as style transfer, should choose smaller " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\lambda (\\rightarrow 0)" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", whereas edits targeting" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8864" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 70, + 496, + 210 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 496, + 210 + ], + "type": "image", + "image_path": "aa8709af9cdd4c546356fdd95823f9d774fb44ad62023ae5dae09689d981550b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. Comparison of image editing methods. (top) LEDs++ is the only method to restrict edits to the tree leaves and position of the car. (bottom) Ours is the only approach faithfully executing all three edits and keeping changes minimal. (Best viewed in color)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 99, + 247, + 496, + 350 + ], + "blocks": [ + { + "bbox": [ + 99, + 247, + 496, + 350 + ], + "lines": [ + { + "bbox": [ + 99, + 247, + 496, + 350 + ], + "spans": [ + { + "bbox": [ + 99, + 247, + 496, + 350 + ], + "type": "table", + "html": "
MethodReconstruction \nError (RMSE) ↓Execution \nTime (s) ↓Variation/ \nSamplingSemantic \nGroundingMulti- \nEditing
SDEdit [26]0.81 ±0.072.10 ±0.02✓XX
Imagic [18]0.58 ±0.12349.98 ±0.45✓XX
Vanilla DDIM Inversion0.22 ±0.1037.23 ±0.04XXX
Pix2Pix-Zero [30]0.20 ±0.0956.78 ±0.14(✓)✓X
DiffEdit [9]0.13±0.0327.65 ±0.03✓✓X
Edit-friendly DDPM [17]0.0010.36 ±0.05✓XX
LEDSITS++ (Ours)0.001.78 ±0.03✓✓✓
", + "image_path": "bd97ddbef80dd831bf1304807b0f0851b0a858012777a1cfac3695bf331144e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 546, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 354, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 546, + 387 + ], + "type": "text", + "content": "Table 1. Comparing key properties for diffusion-based image editing techniques, with LEDs++ offering clear methodological benefits. Due to LEDs++'s efficient perfect inversion, it is the fastest and error-free method. At the same time, its methodology is the only enabling versatility in terms of variation, semantic grounding, and multi-editing. Subscript numbers indicate standard deviation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 403, + 287, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 428 + ], + "type": "text", + "content": "specific objects or regions should use " + }, + { + "bbox": [ + 46, + 403, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 403, + 287, + 428 + ], + "type": "text", + "content": " proportional to the region's prominence in the image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 188, + 454 + ], + "type": "text", + "content": "4. Properties of LEDs++" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 485 + ], + "type": "text", + "content": "With the fundamentals of LEDs++ established, we next showcase its unique properties and capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 676 + ], + "type": "text", + "content": "Efficiency. First off, LEDs++ offers substantial performance improvements over other image editing methods. In Tab. 1, we provide a qualitative runtime comparison, with all methods being implemented for Stable Diffusion (SD) 1.5 [34]. As a parameter-free approach, LEDs++ does not require any computationally expensive fine-tuning or optimization. Consequently, LEDs++ is orders of magnitude faster than methods like Imagic [18] or Pix2Pix-Zero [30]. Further, we only need to invert the same number of diffusion steps used at inference, which results in significant runtime improvements over the standard DDIM inversion (21x). In addition to efficient inversion, we use a recent, fast scheduler that generally requires fewer total steps, further boosting performance. This way, LEDs++ is six times faster than recent DDPM inversion [17] and on par with fast but poor-quality SDEdit [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "Versatility. In addition to its efficiency, LEDs++ remains versatile, enabling sheer limitless editing possibilities. In Fig. 1, we showcase a broad range of edit types." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 403, + 545, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 545, + 571 + ], + "type": "text", + "content": "LEDITS++ facilitates fine-grained edits (adding/removing glasses) and holistic changes such as style transfer (painting/sketch). Furthermore, object removal and replacement facilitate even more image editing tasks. Importantly, the overall image composition is preserved in all cases. To our knowledge, LEDs++ is the only diffusion-based image editing method inherently supporting multiple edits in isolation, which allows for more complex image manipulation. Fig. 2 highlights LEDs++ benefits over previous methods. Our method produces the highest edit fidelity and is the only approach capable of faithfully executing multiple, simultaneous instructions. Moreover, LEDs++ also makes the least changes to unrelated objects and the overall background and composition of the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "content": "Lastly, the editing versatility benefits from the stochastic nature of the perfect but non-deterministic inversion. LEDs++ provides meaningful image variations by resampling " + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\tilde{c}_t" + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "content": ". Additionally, the visual expression of each concept in the edited image scales monotonically with " + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 304, + 576, + 545, + 661 + ], + "type": "text", + "content": ", and the direction and magnitude of each concept can be varied freely. We present examples of both features in App. B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 714 + ], + "type": "text", + "content": "Precision. Furthermore, LEDs++'s methodology keeps edits concise and avoids unnecessary deviations from the input image (Fig. 2). First, the perfect inversion will reconstruct the exact input image if no edit is applied (cf." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8865" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 263, + 150 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 263, + 150 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 263, + 150 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 263, + 150 + ], + "type": "image", + "image_path": "65145febf9dc8f5f3ea361c981b25ba7f5ed9a8b34e59e22ec6b15377923bcfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 288, + 189 + ], + "type": "text", + "content": "Figure 3. Exemplary edit performed with LEDs++ in only 25 diffusion steps with SD1.5. We apply a complex, compounded edit and ground each to a semantically reasonable image region." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 200, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 357 + ], + "type": "text", + "content": "Sec. 3.2). Consequently, we already improve on faithfulness to the input image even before applying any edits. This benefit over other methods is highlighted by the reconstruction error in Tab. 1. Second, implicit masking will semantically ground each edit to relevant image regions. This is specifically important for editing multiple concepts at the same time. While other methods only utilize one prompt for all instructions, LEDs++ isolates edits from each other (Eq. 11). Thus, we get dedicated masks for each concept as shown in Fig. 3. This design ensures that each instruction (e.g., red mask for 'cherry blossom') will be only applied where necessary. Subsequently, we provide further evidence for the efficacy of LEDs++'s masking approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 259, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 259, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 259, + 383 + ], + "type": "text", + "content": "5. Semantically Grounded Image Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": "Cross-attention maps of DMs have been used extensively to ground regions of interest during image generation semantically [7, 8, 14, 30]. Nonetheless, these have not been combined with noise-based masks so far and thus lack fine granularity. Hence, we empirically evaluate the quality of implicit masks, i.e., attention maps " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": " and noise maps " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": " (Eq. 13 and 14) in the LEDs++ setup. We use a broad segmentation task for common objects as a proxy to measure the performance of implicit masks in identifying relevant image areas from edit instructions. Specifically, we utilize segmentation masks from the COCO panoptic segmentation challenge [19]. For each unique object in an image, we retrieve the masks " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": ", and their intersection per diffusion step. We use the (semantic) class label (e.g. 'person' or 'TV') as editing concept " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": ". We consider masks at each of 50 total diffusion steps without actually editing the input image. Furthermore, we approximate mask threshold " + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 389, + 287, + 605 + ], + "type": "text", + "content": " based on the relative size of an object's bounding box." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Concise masks with LEDs++. Fig. 4 shows implicit masking as a reliable means to identify relevant image regions. Importantly, the intersection of both cross-attention masks " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "M^1" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and noise maps " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "M^2" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " clearly outperforms each separate mask. The overall performance is even similar to a dedicated CLIPSeg model [24], despite LEDs++. masks being implicitly calculated at inference with only minimal overhead. At the same time, LEDs++.'s masking is superior to DiffEdit's [3]. Consequently, our method's inter" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 73, + 531, + 182 + ], + "blocks": [ + { + "bbox": [ + 321, + 73, + 531, + 182 + ], + "lines": [ + { + "bbox": [ + 321, + 73, + 531, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 531, + 182 + ], + "type": "image", + "image_path": "318c6382a39c079543b3cc4819e80b0f71060178d28b71ea7d436599a39f18cc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 545, + 232 + ], + "type": "text", + "content": "Figure 4. Semantic segmentation quality of LEDs++. We show the intersection over union (higher is better) for COCO panoptic segmentation. The intersection masks outperform each by a clear margin, close to the CLIPSeg reference. (Best viewed in color)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 243, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 281 + ], + "type": "text", + "content": "section of cross-attention masks and noise maps provides strong semantic grounding while being efficient during image manipulation to ensure precise editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 454, + 306 + ], + "type": "text", + "content": "6. Image Editing Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 313, + 545, + 338 + ], + "type": "text", + "content": "Let us now compare LEDs++ to current SOTA methods for image manipulation on two benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 346, + 452, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 452, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 452, + 361 + ], + "type": "text", + "content": "6.1. Editing Multiple Concepts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 581 + ], + "type": "text", + "content": "First, we investigate the complex task of performing multiple edits simultaneously. We rely on a well-established setup for semantic image manipulation [5] to evaluate multi-conditioned attribute manipulation in facial images. In our experiment, we consider 100 images from the CelebA dataset [21]. For each image, we simultaneously edit three attributes out of a set of five, leading to ten total combinations of edit concepts. Further, we perform each edit across ten different seeds, resulting in 10,000 evaluated images for each method and hyperparameter setting, over 1M images in total. As measures for comparison, we employ CLIP and LPIPS scores. CLIP measures the text-to-image similarity of the edit instruction to the edited image, and LPIPS measures the image-to-image similarity of the real to the edited image. This way, we assess the tradeoff between the versatility of edits (CLIP) and the precision of those manipulations (LPIPS). We implement all methods based on SD1.5 and provide more details in App. C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "LEDS++ outperforms competing methods. Fig. 5 shows the resulting CLIP vs. LPIPS plots for all methods. The top left corner represents the ideal editing method with maximum edit alignment without deviating from the initial image. Generally, one can observe a natural trade-off between versatility and precision for all methods, i.e., higher image-to-text alignment comes at the expense of lower similarity to the original image. LEDS++ is closest to the ideal region and thus clearly outperforms the other methods. In particular, the outputs remain close to the original image (low LPIPS scores), thanks to the precise implicit masking." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8866" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 272, + 214 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 272, + 214 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 272, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 272, + 214 + ], + "type": "image", + "image_path": "110ffe195968c3d76f23e8c672ce8b00d27b845da1c2f8bcb04eed0058e662e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 283 + ], + "type": "text", + "content": "Figure 5. Comparison of instruction-alignment vs. image similarity trade-off for different editing methods. Results were reported for simultaneous manipulation of three facial attributes on CelebA. We plot CLIP scores (higher is better) of the target attributes against LPIPS similarity (lower is better). LEDs++ clearly outperforms all competing methods. (Best viewed in color)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 295, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 295, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 295, + 289, + 475 + ], + "type": "text", + "content": "At the same time, it faithfully performs the edits (high CLIP scores) due to the dedicated, isolated editing for each concept. The depicted scores reflect our qualitative inspections for Pix2Pix-Zero and Imagic on such complex manipulations (cf. Fig. 2). We observed that these methods often break—either failing to perform all three edits and/or drastically altering the input image. Only edit-friendly DDPM [17] and LEDs++ reliably achieve the maximum average CLIP score of over 0.25. This value seems to represent an upper bound according to our manual investigations, as each attribute is edited correctly for all input images, and higher scores are not observed. Despite being computationally very efficient, LEDs++ faithfully executes each edit instruction while keeping the changes to the input low, highlighting the method's versatility and precision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 483, + 139, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 139, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 139, + 495 + ], + "type": "text", + "content": "6.2. TEDBench " + }, + { + "bbox": [ + 47, + 483, + 139, + 495 + ], + "type": "inline_equation", + "content": "(++)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 502, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 670 + ], + "type": "text", + "content": "Next, we investigate the versatility of LEDs++'s editing capabilities by running the Textual Editing Benchmark (TEDBench [18]), a collection of 100 input images paired with textual edit instructions. However, we observed a variety of inconsistencies in TEDBench and a lack of relevant editing tasks. Therefore, we propose TEDBench++ (Fig. 6a and App. D), a more challenging revised benchmark now containing 120 entries in total.3 We addressed misspellings and rephrased ambiguous and inconclusive instructions. In addition to resolving these issues, we added instructions targeting challenging types of image manipulations previously not included in TEDBench: multi-conditioning, object/concept removal, style transfer, and complex replacements (Fig. 6a). We provide more details in App. D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "text", + "content": "We compare LEDs++ on TEdBench " + }, + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "inline_equation", + "content": "(++)" + }, + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "text", + "content": " to one of the strongest editing methods, Imagic [18] with Imagen [37]." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 537, + 138 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 537, + 138 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 537, + 138 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 537, + 138 + ], + "type": "table", + "html": "
TEDBenchTEDBench++
SR ↑LPIPS ↓SR ↑LPIPS ↓
Imagic w/ SD1.50.550.560.580.57
LEDITS++ w/ SD1.50.750.280.790.30
Imagic w/Imagen [18]0.830.59——
LEDITS++ w/ SD-XL0.840.330.870.34
", + "image_path": "636c1e41362e13cfd1e413f996541c24e2fed7f5ab4e7a5f5d5480a83c80f5c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 143, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 305, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 143, + 545, + 186 + ], + "type": "text", + "content": "Table 2. Success rate (SR) and LPIPS scores on the original TedBench [18] and our revised version (TEDBench++) We compare Imagic to LEDs++ based on different DMs and find the latter to outperform on both metrics and benchmarks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 196, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 545, + 316 + ], + "type": "text", + "content": "Since both are not publicly available, we can only compare to this specific combination of DM and editing method using Kawar et al.'s [18] curated outputs for TEdBench. Additionally, we, therefore, cannot combine LEDs++ with Imagen[37] and instead use a similarly advanced diffusion model, SD-XL [32]. However, to not only compete for the best fidelity outputs but focus the evaluation on methodological differences—not the pre-trained DM—we also compare both methods implemented with SD1.5. We provide further details in App. C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "content": "LEDS " + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "inline_equation", + "content": "^{+ + }" + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "content": " edits images reliably. We first asked users to assess the overall success of edits, i.e., if an edit instruction was faithfully realized for a given input image. The results in Tab. 2 show that LEDs " + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "inline_equation", + "content": "^{+ + }" + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "content": " outperforms Imagc on TEdBench despite a greatly reduced runtime (Tab. 1). The difference is even stronger when comparing both methods on the same pre-trained DM, i.e., SD1.5. The high success rate on TEdBench " + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "inline_equation", + "content": "^{+ + }" + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "inline_equation", + "content": "(87\\%)" + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "content": " and the examples shown in Fig. 1 and 6a once again highlight LEDs " + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "inline_equation", + "content": "^{+ + }" + }, + { + "bbox": [ + 304, + 317, + 546, + 449 + ], + "type": "text", + "content": "'s versatility. Overall, our proposed method can reliably perform a diverse set of editing instructions for real images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 630 + ], + "type": "text", + "content": "High-quality edits with LEDs++. While investigating both methods' performance we observed a substantial difference in edit quality. The examples in Fig. 6b particularly highlight the discrepancy in compositional robustness and object coherence. Hence, we also assessed both methods' editing quality on TEdBench++. We focus on samples where both methods performed a successful edit, i.e., were labeled as successful by users. We show the perceptual similarity (LPIPS) to the input image in Tab. 2. One can observe that the LPIPS scores for LEDs++ are much lower than for Imagic, empirically supporting the qualitative examples in Fig. 6b. When manually inspecting the generated images, we often found Imagic to generate a completely new image based on the edit instruction, entirely disregarding the input image (cf. App. Fig. 11a)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 643, + 375, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 375, + 656 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 375, + 656 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 665, + 544, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 544, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 544, + 676 + ], + "type": "text", + "content": "Let us now discuss open research questions and limitations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "content": "Model Dependency. While LEDs++ achieves impressive results on a large variety of image manipulation tasks, there are external factors to consider. Since the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 241, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 241, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 241, + 712 + ], + "type": "text", + "content": "3 https://huggingface.co/datasets/AIML-TUDA/TEdBench_plusplus" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8867" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 321, + 231 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 321, + 231 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 321, + 231 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 321, + 231 + ], + "type": "image", + "image_path": "be4dfb3e9df3f04561b13dffc758c1f1e4f3c2a54aa97cbf03fd80af2b3c79db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 232, + 323, + 253 + ], + "lines": [ + { + "bbox": [ + 47, + 232, + 323, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 323, + 253 + ], + "type": "text", + "content": "(a) Novel challenging examples of TEdBench++ and LEDs++ applied, showcasing the versatility of supported edits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 347, + 70, + 545, + 231 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 545, + 231 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 545, + 231 + ], + "type": "image", + "image_path": "80bbbd9c3d251bbf51c623aac6607686525a4f1e1fe8d544eeb4757123cb27e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 232, + 545, + 253 + ], + "lines": [ + { + "bbox": [ + 344, + 232, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 344, + 232, + 545, + 253 + ], + "type": "text", + "content": "(b) Qualitative comparison of LEDs++ and Imagic on TEd-Bench, clearly highlighting the performance improvement." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 257, + 480, + 269 + ], + "lines": [ + { + "bbox": [ + 112, + 257, + 480, + 269 + ], + "spans": [ + { + "bbox": [ + 112, + 257, + 480, + 269 + ], + "type": "text", + "content": "Figure 6. Benchmark examples for LEDs++ and Imagc on TEdBench " + }, + { + "bbox": [ + 112, + 257, + 480, + 269 + ], + "type": "inline_equation", + "content": "(++)" + }, + { + "bbox": [ + 112, + 257, + 480, + 269 + ], + "type": "text", + "content": " . (Best viewed in color)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 279, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 471 + ], + "type": "text", + "content": "method is architecture-agnostic, it can be easily used with any DM. At the same time, the general editing quality strongly depends on the overall capabilities of the underlying pre-trained DM. Naturally, more capable models will also enable better edits. But, at times, specific editing instructions may fail because the used DM does not have a decent representation of the targeted concept to begin with. One example is the model failing to edit a giraffe to be sitting since the underlying DM generally fails to generate this pose (cf. App. F). This effect can also clearly be seen in Tab. 2, with the editing success rate of a method varying strongly between DMs. Although the same image editing method is employed (LEDITS++, the more capable SD-XL variant outperforms the weaker SD1.5 model. Nonetheless, this means that the architecture-agnostic LEDs++ will benefit from increasingly powerful DMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 472, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 664 + ], + "type": "text", + "content": "Coherence Trade-offs. Next to the benefits of LEDs++'s semantic grounding, there are also downsides to this approach. Overall, implicit masking limits changes to relevant portions of the image and achieves strong coherence with the original image composition. Yet, the object and its identity within the masked area may change based on various factors. Generic prompts, like \"a standing cat\" (cf. App. F), do not contain detailed information about this specific object (\"cat\"). Thus, an edit with this prompt does not guarantee to preserve object identity, particularly for strong hyperparameters. We observed that fine-tuning approaches like Imagic make the opposite trade-off, better preserving the object identity while changing the background and image composition substantially (cf. App. F). A potential remedy for a loss in object coherence with LEDs++, is more descriptive edit prompting, e.g. using textual inversion [12]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Lastly, the automatically-inferred implicit masks allow for easy use of LEDs++ without users tediously providing masks. Nonetheless, user intentions are diverse and cannot always be automatically inferred. Sometimes, in-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 279, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 327 + ], + "type": "text", + "content": "individual user masks provide better control over the editing process. Such user masks can be easily integrated into LEDs++ (cf. App. F), wherefore we encourage future research in this promising direction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 328, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 449 + ], + "type": "text", + "content": "Societal Impact. LEDs++ is an easy-to-use image editing technique that lowers the barrier for users and puts them in control for fruitful human-machine collaboration. Yet, the underlying text-to-image models offer both promise and peril, as highlighted by prior research [4, 11]. The (societal) biases within these models will also impact image editing applications [11]. Moreover, image manipulation can also be used adversarially to generate inappropriate [39] or fake content. Hence, we advocate for a cautious deployment of generative models together with image editing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 462, + 378, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 378, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 378, + 475 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 614 + ], + "type": "text", + "content": "We introduced LEDs++, an efficient yet versatile and precise method for textual image manipulation with diffusion models. It facilitates the editing of complex concepts in real images. Our approach requires no finetuning nor optimization, can be computed extremely efficiently, and is architecture agnostic. At the same time, it perfectly reconstructs an input image and uses implicit masking to limit changes to relevant image regions, thus editing precisely. Our large experimental evaluation confirms the efficiency, versatility, and precision of LEDs++ and its components, as well as its benefits over several related methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements We gratefully acknowledge support from the BMBF (Grant No 01IS22091). This work benefited from the ICT-48 Network of AI Research Excellence Center \"TAILOR\" (EU Horizon 2020, GA No 952215), the Hessian research priority program LOEWE within the project WhiteBox, the HMWK cluster projects \"Adaptive Mind\" and \"Third Wave of AI\", and from the NHR4CES." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8868" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Kfir Aberman, Ohad Fried, Daniel Cohen-Or, and Dani Lischinski. Break-a-scene: Extracting multiple concepts from a single image. SIGGRAPH Asia, 2023. 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 144 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. arXiv:2206.02779, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 144, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 144, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 53, + 144, + 288, + 198 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. eDiff-I: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv:2211.01324, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 198, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 288, + 273 + ], + "type": "text", + "content": "[4] Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale. In Proceedings of ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2023. 8, 11" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 338 + ], + "type": "text", + "content": "[5] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. In Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 288, + 381 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 382, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 382, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 382, + 288, + 424 + ], + "type": "text", + "content": "[7] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. arXiv:2304.08465, 2023. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "text", + "content": "[8] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Trans. Graph., 42, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 468, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 288, + 520 + ], + "type": "text", + "content": "[9] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 573 + ], + "type": "text", + "content": "[10] Zoe De Simone, Angie Boggust, Arvind Satyanarayan, and Ashia Wilson. What is a Fair Diffusion Model? Designing Generative Text-To-Image Models to Incorporate Various Worldviews. arXiv preprint arXiv:2309.09944, 2023. 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 574, + 287, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 287, + 628 + ], + "type": "text", + "content": "[11] Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness. arXiv preprint arXiv:2302.10893, 2023. 8, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 628, + 287, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 628, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 48, + 628, + 287, + 681 + ], + "type": "text", + "content": "[12] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 8, 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 682, + 287, + 704 + ], + "type": "text", + "content": "[13] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 705 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 11" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 147 + ], + "type": "text", + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 148, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 213 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and HsuanTien Lin, editors, Proceedings of the Advances in Neural Information Processing Systems: Annual Conference on Neural Information Processing Systems (NeurIPS), 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 213, + 545, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 213, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 308, + 213, + 545, + 234 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv:2207.12598, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 235, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 545, + 276 + ], + "type": "text", + "content": "[17] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. arXiv preprint arXiv:2304.06140, 2023. 1, 2, 3, 4, 5, 7, 12, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 277, + 545, + 331 + ], + "type": "text", + "content": "[18] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 7, 12, 13, 14" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 384 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV), 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "type": "text", + "content": "[20] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B. Tenenbaum. Compositional visual generation with composable diffusion models. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 547, + 470 + ], + "type": "text", + "content": "[21] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), December 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 512 + ], + "type": "text", + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In NeurIPS, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 555 + ], + "type": "text", + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "type": "text", + "content": "[24] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 599, + 545, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 599, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 308, + 599, + 545, + 641 + ], + "type": "text", + "content": "[25] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv:2310.04378, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 642, + 545, + 695 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 545, + 695 + ], + "type": "text", + "content": "[26] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2022. 1, 2, 4, 5, 12, 13" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 696, + 545, + 705 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 696, + 545, + 705 + ], + "spans": [ + { + "bbox": [ + 308, + 696, + 545, + 705 + ], + "type": "text", + "content": "[27] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik P." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8869" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 170 + ], + "type": "text", + "content": "[28] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3, 12" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 288, + 235 + ], + "type": "text", + "content": "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the International Conference on Machine Learning (ICML). PMLR, 2022. 2, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 235, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 288, + 267 + ], + "type": "text", + "content": "[30] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In SIGGRAPH, 2023. 1, 2, 3, 5, 6, 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 267, + 288, + 310 + ], + "type": "text", + "content": "[31] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv:2306.00637, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 363 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 288, + 406 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 2022. 2, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 288, + 460 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 546 + ], + "type": "text", + "content": "[36] Harrison Rosenberg, Shimaa Ahmed, Guruprasad V. Ramesh, Ramya Korlakai Vinayak, and Kassem Fawaz. Unbiased Face Synthesis With Diffusion Models: Are We There Yet? arXiv preprint arXiv:2309.07277, 2023. 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 546, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 288, + 621 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. arXiv:2205.11487, 2022. 2, 7, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 621, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 621, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 621, + 288, + 654 + ], + "type": "text", + "content": "[38] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. arXiv:2304.14530, 2023. 17" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 654, + 288, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 654, + 288, + 708 + ], + "spans": [ + { + "bbox": [ + 48, + 654, + 288, + 708 + ], + "type": "text", + "content": "[39] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 8, 11" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 406 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 160 + ], + "type": "text", + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Proceedings of NeurIPS Datasets and Benchmarks, 2022. 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 160, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 160, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 160, + 545, + 182 + ], + "type": "text", + "content": "[41] Norbert Schwarz. How the questions shape the answers. American Psychologist, 1999. 15" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 182, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 182, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 545, + 224 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "type": "text", + "content": "[43] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 278, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 278, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 278, + 545, + 320 + ], + "type": "text", + "content": "[44] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv:2210.09477, 2022. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[45] Qiucheng Wu, Yujuan Liu, Handong Zhao, Ajinkya Kale, Trung Bui, Tong Yu, Zhe Lin, Yang Zhang, and Shiyu Chang. Uncovering the disentanglement capability in text-to-image diffusion models. arXiv:1212.08698, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 364, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 545, + 406 + ], + "type": "text", + "content": "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. arXiv:2302.04867, 2023.2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8870" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_content_list.json b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9344039d89b361486e0da10e07b2a04a30b2bf36 --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_content_list.json @@ -0,0 +1,1489 @@ +[ + { + "type": "text", + "text": "LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising", + "text_level": 1, + "bbox": [ + 99, + 130, + 869, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuxing Duan", + "bbox": [ + 426, + 180, + 540, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Lab of Multispectral Information Intelligent Processing Technology \n $^{1}$ Huazhong University of Science and Technology", + "bbox": [ + 158, + 198, + 810, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f38453187383b2d327912d910f61ae52a5357f0214b2b71f7e96429409db0ca.jpg", + "image_caption": [ + "Figure 1. Illustration of the paired raw/denoised event datasets. The proposed real-world event denoising datasets contribute to addressing the inevitable Background Activity (BA) noise encountered by event cameras application under varying camera settings and illumination." + ], + "image_footnote": [], + "bbox": [ + 78, + 268, + 895, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 507, + 313, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event camera has significant advantages in capturing dynamic scene information while being prone to noise interference, particularly in challenging conditions like low threshold and low illumination. However, most existing research focuses on gentle situations, hindering event camera applications in realistic complex scenarios. To tackle this limitation and advance the field, we construct a new paired real-world event denoising dataset (LED), including 3K sequences with 18K seconds of high-resolution (1200*680) event streams and showing three notable distinctions compared to others: diverse noise levels and scenes, larger-scale with high-resolution, and high-quality GT. Specifically, it contains stepped parameters and varying illumination with diverse scenarios. Moreover, based on the property of noise events inconsistency and signal events consistency, we propose a novel effective denoising framework(DED) using homogeneous dual events to generate the GT with better separating noise from the raw. Furthermore, we design a bio-inspired baseline leveraging Leaky-Integrate-and-Fire (LIF) neurons with dynamic thresholds to realize accurate denoising. The experimental results demonstrate that the remarkable performance of the proposed approach on different datasets. The dataset and code are at https://github.com/Yee-Sing/led.", + "bbox": [ + 75, + 534, + 472, + 896 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 507, + 630, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras possess unique imaging advantages and are demonstrating tremendous potential for various applications [4, 11, 33, 37]. Unlike traditional frame cameras using integration sampling [46], event camera is differential sampling [29], evading the restriction of exposure period. However, due to this sampling mechanism, random fluctuations in analog signals can easily form noise events. In contrast to the dominance of signal components in images, the noise and signal in event cameras are encoded with same magnitude, easily disturbing inherent structural features, such as lane lines shown in Fig. 1. BA noise detrimentally impact on subsequent tasks like reconstruction [36, 39], motion estimation [40, 41] and event-based object detection [5]. Therefore, event denoising becomes a fundamental issue, notably in the booming event-based vision community.", + "bbox": [ + 496, + 535, + 890, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paired data is a key issue for event denoising in the deep-learning era, thus an intuitive idea is to construct paired data through simulators [17, 22, 38]. However, the image-based simulation cannot be equivalent to reality because of the domain gap. To obtain the 'clean part' from the real-world event, multi-modal data was incorporated to help paired event generation. APS images gradient combined with IMU are used to indicate the event probability within frame period [2]. By jointly estimating the motion parameters with", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25637", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/899b3a59222e279a3ed0201e4dbf8f48bed4516c4a15c5b67eb6cc0877473210.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsSequencesCapture/sSourceNoise levelIlluminationResolutionPaired
RGB-DAVIS [45]20122CameraXDaytime180*190X
EventNFS [10]1004238Screen RecordingXDaytime224*124X
E-MLB [8]12007300Camera4Daytime to Nighttime346*260X
DVSNOISE20 [2]48807CameraXDaytime346*260✓
ED-KoGTL [1]415CameraXDaytime to Nighttime346*260✓
DVSCLEAN [12]14455Synthesis2Daytime1280*720✓
LED300018000Camera4Daytime to Nighttime1200*680✓
", + "bbox": [ + 93, + 88, + 874, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Summary of existing event denoising datasets.", + "bbox": [ + 318, + 210, + 647, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "APS images and event streams [9, 45], the temporal mismatch can be alleviated. Constructing a fixed-track setup [1], the events under various illuminations can be referred to APS images in good lighting. Essentially, these methods align with events via transformed multi-mode features. Unfortunately, this heterogeneous aided information cannot fully match with the event, and their effectiveness is unavailable in situations such as low light and motion blur. Hence, few datasets have considered practical scenes where BA noise is easily deteriorated by camera setting and illumination, appearing as complex and nonuniform distributions.", + "bbox": [ + 75, + 236, + 467, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we construct a large-scale, high-quality paired dataset LED for event denoising. LED has three distinct advantages. Firstly, it contains diverse noise levels, covering complex situations under various camera settings and illumination. Secondly, LED includes such as outdoor and indoor scenarios, with abundant objects. Last but not least, LED is collected with the current largest-scale and high-resolution. Moreover, inspired by the multisampling images denoising methods [21, 28, 31, 34], we are the first to explore event multi-sampling for denoising. Based on the homogeneous data, we proposed a dual-events denoising framework that can generate higher-quality GT by more accurately separating BA noise from the raw.", + "bbox": [ + 75, + 402, + 467, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as:", + "bbox": [ + 91, + 599, + 450, + 614 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We provide a large-scale real-world paired dataset for event denoising. So far, LED is the largest paired realworld event denoising dataset (3000 sequences) and high-resolution $(1200^{*}680)$ with diverse noise levels and scenes.", + "(2) We propose a novel dual events denoising framework DED based on noise inconsistency. We provide detailed analysis to show that DED can better separate mixed signals and noisy event streams across various camera threshold parameters, light conditions and motion patterns.", + "(3) We introduce a novel baseline DTSNN for event denoising based on the spiking neural network, which utilizes the learnable dynamical spike threshold of the LIF neuron to accurately denoising. Extensive experiments on different real datasets verify the superiority of the proposed method." + ], + "bbox": [ + 75, + 616, + 467, + 828 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 844, + 215, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Denoising Datasets. Currently, the datasets exclusively designed for event denoising with less attention com", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pared to the task-specific ones in the event-based community. In Table 1, we provide an all-sided summary of existing datasets, highlighting the scarcity of real paired datasets.", + "bbox": [ + 496, + 236, + 890, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the lack of explicit labels, some paired datasets can be conveniently generated via simulators such as DVSCLEAN [12] and DND21 [20]. However, due to the gap between simulation and real, these simulators fail to fully reflect the actual situation. Other works focus on real paired event denoising data. DVSO noise20 [2] captures data in 16 pure background scenes. RGB-DAVIS [45] provides 20 indoor/outdoor sequences under good lighting conditions. ED-KoGTL [1] constructs indoor data with four illumination levels under constant trajectory. These paired datasets have achieved favorable results, primarily attributable to the gradient cues extracted from APS images. However, the heterogeneity between images and events limits the availability of diverse lighting conditions, motion patterns, and scenarios for constructing paired data.", + "bbox": [ + 496, + 286, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based Denoising. Research on event denoising can be categorized into two approaches: filtering-based methods and deep learning methods. The former relies on manual priors to design discriminative models for noise removal, such as utilizing density distinction [15, 20, 30, 50] or motion continuity distinction [3, 44, 47] and motion compensation [6, 45] with CM framework [16] in the spatiotemporal domain. However, the validity of this prior knowledge is sensitive to different signal/noise event distributions, limiting their denoising accuracy. For learning-based methods, such as the CNN-based EDNCNN [2], the PointNet-based AEDNet [12] and [1] GNN-transformer were proposed successively. Additionally, the event denoising and super-resolution tasks are combined in [10].", + "bbox": [ + 496, + 518, + 890, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Sampling Image Denoising. The multi-sampling principle for image denoising has been extensively studied. Techniques like long exposure time or burst captures benefit the direct acquisition of clear data with increased SNR in static scenes. Further, the multi-sampling was explored to the paradigm of indirectly recovering noise-free data. Not only using paired noisy observations to learn the mapping from different noisy samples to clear counterpart [27, 32], but also employing spatial sub-sampling on a single noisy image to learn the mapping from adjacent noisy samples to GT [23, 25]. Evidently, these denoising approaches imply", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25638", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ecd600611afddee70686da699d5581c10880777dd1ae9346566ed6cbc92e300b.jpg", + "image_caption": [ + "(a) Distribution of LED" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 338, + 299 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f57e42e5e0c9ca78b0dfa172ee146d6e61fc1eeab685d04520a4e56411a1152.jpg", + "image_caption": [ + "(b) Scale of LED" + ], + "image_footnote": [], + "bbox": [ + 338, + 87, + 571, + 304 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9795416dca29c0fb8eaed60ec1247de0ce8c087ef8061fa99cb41d76455f62f2.jpg", + "image_caption": [ + "(c) Diversity of LED" + ], + "image_footnote": [], + "bbox": [ + 571, + 88, + 890, + 305 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9746f753be59fadcb8d34d9534ac8b832e2d16c0639bd2bfda08dcfa4a15eeb1.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed dataset LED. (a) Distribution of noise level and scene of the proposed dataset. (b) Our proposed LED outperforms others in terms of sequences, capture, and resolution (Circles with numbers to indicate). (c) LED collects diverse event streams across various conditions of illumination, depth of field, and target scale.", + "(a) Noise in frame camera" + ], + "image_footnote": [], + "bbox": [ + 125, + 380, + 421, + 465 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d92ae6223f94ff3c6c879929ba5eb4422379e9d92e06ff64ebdaa047af2d15d0.jpg", + "image_caption": [ + "(b) Noise in event camera", + "Figure 3. The substantial noise distinctions between frame/event camera. (a) For a stationary grayscale chart, the frame camera sequentially acquires noisy samples, fluctuating around a latent value within a specific distribution. (b) In a horizontal moving case, the event camera outputs binary signal serially, comprising signal events from motion gradient edges and spurious BA noise." + ], + "image_footnote": [], + "bbox": [ + 125, + 472, + 419, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that signal resampling could facilitate suppressing random circuit noise and approximating the expectation value.", + "bbox": [ + 75, + 672, + 468, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The High-quality Paired LED Dataset", + "text_level": 1, + "bbox": [ + 76, + 715, + 419, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Statistics and Features", + "text_level": 1, + "bbox": [ + 76, + 741, + 349, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary factors affecting BA noise in event cameras can be categorized into intrinsic camera settings and extrinsic lighting conditions. By setting different camera thresholds, we can selectively record events with varying noise levels. Generally, a lower threshold means higher sensitivity but also increases the likelihood of BA noise. In contrast to fixed parameters in other datasets, we use four level thresholds to collect events, starting from the default value of $0\\%$ and gradually decreasing to $-10\\%$ , $-20\\%$ , and $-30\\%$ ,", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as shown in Fig. 2 (a). To cope with the challenges posed by complex lighting conditions, data collection is from daytime to nighttime, encompassing a transition from high to low illumination and a mixture of natural light with artificial light sources. Thus, the impact of external conditions on the event stream distribution can also be controlled.", + "bbox": [ + 496, + 383, + 890, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the difficulty of constructing real paired datasets, the development of existing datasets remains limited, as shown in Table 1. In this work, we used a vehicle-mounted or pan-tilt equipped with event cameras featuring a $16\\mathrm{mm}$ lens. Totally, we collected approximately 5 hours of event stream, including 3K sequences. The composition of noise levels and scenes is depicted in Fig. 2 (a), with roughly equal amounts collected during daytime and nighttime. Nearly 20 typical scenes were captured, including urban buildings, traffic roads, natural scenery, indoor exhibits, etc.", + "bbox": [ + 496, + 476, + 892, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset goes beyond just the number of sequences and data volume, primarily aiming to consider the intricate influences of relevant factors on noise. Thus, we endow LED with diversity to reflect the impact of objects and environments on the event, which helps cover the cases of densely and non-uniformly distributed noisy situations in reality both spatially and temporally. In Fig. 2 (c), we exhibit that LED diversity is not only embodied in the richness of noise and illumination levels but also features such as depth of field and scene semantics. Meanwhile, high-resolution imaging aids the event cameras in capturing spatial information more precisely. Unfortunately, most existing ones with low resolution may hinder advanced visual tasks. Therefore, we collected data with $1200*680$ resolution(after cropping), achieving high spatiotemporal resolution. Additionally, LED includes abundant objects of various scales, including aerial targets, pedestrians, and vehicles, making it well-suited for downstream tasks.", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25639", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b5a700a3a9fa6b12d3c367f38ee38459d3b4a2b01223f2709caf979cfc2cf9e6.jpg", + "image_caption": [ + "(a) Ideal intensity input" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 251, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d18ab4b6ec4e6f908c1b6777933e802b6ed92999ae2a20b6f17a0924a758b351.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 88, + 488, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c3167d56f6087147b4ed8f56828d07fe4dc2a9ad090fe27bde62986a45def9c8.jpg", + "image_caption": [ + "(b) Circuit dual-sampled simulation" + ], + "image_footnote": [], + "bbox": [ + 253, + 176, + 488, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/689fbe4668c25bda81b832c67afd44a047f370e79ed4a62f037fbe6857be358f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 88, + 712, + 174 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a87918737ebcc936b450c3c644bdbb624bb35b553664dffef76af5a5af0fd85.jpg", + "image_caption": [ + "(c) Realistic dual-Sampled events" + ], + "image_footnote": [], + "bbox": [ + 493, + 175, + 710, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01c4b7eafad35955d0781bcbd9bd208a546e36fabbf33a69db2038e5b9e0df74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 89, + 892, + 175 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b90971034e6fbf5e8cc54a67a41115fe1a8e6f2e85e9ced63455b31b625b88d9.jpg", + "image_caption": [ + "(d) Statistic of dual-sampled events" + ], + "image_footnote": [], + "bbox": [ + 712, + 176, + 890, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4a57266de4d34d4a8077f3e2c06ee67c38a66a4dc12a67ee5cba87f8244cebbe.jpg", + "image_caption": [ + "Figure 4. Illustration of event camera dual-sampling analysis: (a) Given an ideal intensity input, which includes varying and steady stages, generates a series of signal events. (b) The twice samplings of the same input in the circuit model, both generate additional BA noise in previously steady stages, resulting in consistency discrepancies between noise events and signal events. (c) The visualization results of the actual dual-sampled events cumulative frame demonstrate the misalignment of inconsistent dual-sampled noise events and the alignment of coexisting signal events (two color groups indicating the respective polarities of the dual events). (d) The statistical results of the two tests also prove a low overlap rate between dual-sampled event pixels and a much higher one where signal events are present.", + "Figure 5. Overview of the dual events denoising framework. (a) Our collection device consists of two identical EVK4s forming a co-axial system with a 1:1 beamsplitter. (b) We first perform spatial similarity processing to retain the consistent parts, followed by sequentially spatiotemporal correlation constraints to remove the residual small amount of noise event from the previous step." + ], + "image_footnote": [], + "bbox": [ + 78, + 378, + 495, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dual Events Denoising Framework", + "text_level": 1, + "bbox": [ + 76, + 619, + 382, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a noisy event stream, the key lies in properly obtaining paired GT. In this section, we explore utilizing multi-sampling to achieve event stream denoising.", + "bbox": [ + 75, + 643, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background. Multi-sampling denoising methods are commonly used in image denoising. For frame images, averaging the results of temporal multi-samples can approximate the true value for each pixel owing to multi-samples smoothing effect on random noise components, as shown in Fig. 3 (a). For event stream, due to their dynamic sampling and the binary values, temporal multi-samples may contain both signal events and BA noise, as illustrated in Fig. 3 (b), directly taking the average of multi-samples cannot yield true signal value. Although the noise forms differ between the two modalities, they ultimately originate from the inherent noise in the circuitry analog signal, while resampling fundamentally helps suppress the random noise source.", + "bbox": [ + 75, + 688, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis. This concept inspired us to utilize multi-sampling", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for event denoising, particularly when considering the binary state of events (presence or absence) where just two samplings are required to potentially identify noise. Thus, we initially investigated the noise inconsistency at the pixel level. Based on the DVS working principle [29], we build a circuit model in the Cadence [42] platform to simulate the dual-sampling with twice identical intensity input. The noisy analog signal converted from input generates BA noise at the originally steady stage compared to the ideal case in Fig. 4 (a). Clearly, the two sets of noise exhibit inconsistency, while the signals demonstrate high consistency, causing distinct alignment differences between the two cases as shown in the close-up of Fig. 4 (b).", + "bbox": [ + 496, + 382, + 892, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further explore the inter-class differences between noise and signal events in real-world dual-sampled events stream, we established a collection device as shown in Fig. 5 (a), to construct genuinely spatiotemporal dual-sampling. Two experimental conditions to validate the aforesaid circuit simulation phenomenon are a stationary camera and a moving one respectively. In the former, pure BA noise events was obtained, while the latter captured mixed events containing both noise and signal. In static tests, although at the highest noise level $(-30\\%)$ threshold), both two sets of events exhibit significant spatiotemporal disparity. As observed in the right zoomed-in patches at the upper row in Fig. 4 (c), the global spatial misalignment of dual BA noises events was evident. The overlap rate of the pixels triggering event is also to be less than $1\\%$ , as shown in the upperrow histogram of Fig. 4 (d). Not unexpectedly, in dynamic tests, regions with BA noise events remain inconsistency, while the coexistent signal events showing spatiotemporal alignment as depicted in lower-row counterpart, leading to a dramatic increase in overlap rate compared to the pure BA noise cases, as shown in the bottom histogram of Fig. 4 (d).", + "bbox": [ + 496, + 583, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25640", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/635af38e31bb62e5022ecac3e503890e3dd5ff13d4d6d545aa107af50c22528f.jpg", + "image_caption": [ + "Figure 6. Analysis of different event denoising results on our dataset. From left to right, the first column is the raw events, and the remaining five columns represent different methods, namely DWF, STDF, TimeSurface, EvFlow, and the proposed DED. From top to bottom, the first row shows the denoised results, the second row is the residual noise, the third row denotes the statistical distribution of the denoised results and the last row represents the intensity image reconstruction corresponding to the zoomed region denoised events." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Formulation.** These results support the insight that the consistency discrepancies between dual-sampled events could naturally help distinguish noise. Therefore, we develop a denoising framework with the dual events stream, called DED. In DED framework, given the spatiotemporal synchronized dual event streams $Y_{1}$ and $Y_{2}$ , they are composed of common latent signal $X$ and respective noise streams $N_{1}$ and $N_{2}$ . The noise model can be defined as follows:", + "bbox": [ + 75, + 388, + 472, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY _ {1} = X + N _ {1}, Y _ {2} = X + N _ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 512, + 468, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because of the consistency of signal events and the inconsistency of noise events in $Y_{1}$ and $Y_{2}$ , after binary grid operation of event stream according to certain temporal window $\\Delta t$ , we can perform spatial similarity to process them:", + "bbox": [ + 76, + 530, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX ^ {*} = \\left\\| \\left\\| X + N _ {1} \\right\\| \\circ \\left\\| X + N _ {2} \\right\\| \\right\\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 592, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\circ$ denotes the Hadamard product used to obtain the similar part $X^{*}$ of two groups of raw event binary frames. Due to the probability of a few noises occurring simultaneously in dual-sampling, $X^{*}$ may still contain some much-isolated noise compared to the raw. As shown in Fig. 5 (b), for finer denoising, we further utilize the spatiotemporal correlation of the signal event stream to remove the residual noise from the previous step. Specifically, we accumulate the events from the relevant spatiotemporal range in corresponding $X^{*}$ within several consecutive $\\Delta t$ to form a spatiotemporal relationship set. The presence of these few isolated noise events, which are randomly triggered in the spatiotemporal domain, disrupts the inherent spatiotemporal correlation of signal reflecting a certain motion model. Therefore, to leverage this spatiotemporal continuity, we transform it into the following minimization problem:", + "bbox": [ + 75, + 611, + 472, + 852 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left| e _ {\\boldsymbol {x}} ^ {t _ {i} + 1} - e _ {\\boldsymbol {x}} ^ {t _ {i}} \\right|, _ {\\boldsymbol {x} \\in \\Omega , \\Omega \\in X ^ {*}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 854, + 468, + 895 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_{\\pmb{x}}^{t_i}$ denotes the $i_{th}$ occurring event according to the sorted timestamp, $\\pmb{x}$ means the event spatial coordinates, $n$ is the number of selected consecutive time windows, $\\Omega$ represents the candidate spatial neighborhood, and $N$ is total events number in $\\Omega$ . Additionally, the aforesaid processing is all performed on the two channels of event polarity.", + "bbox": [ + 498, + 388, + 892, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Evaluation and Discussion", + "text_level": 1, + "bbox": [ + 500, + 494, + 738, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quality of GT is critical for a real paired event dataset. However, it is challenging to objectively evaluate the denoising results since directly obtaining clean data is unpractical. To fairly assess the denoising effects of different methods, we conducted a comprehensive evaluation.", + "bbox": [ + 496, + 520, + 892, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 6 presents comparative results of representative event denoising methods could be used for GT generation: spatiotemporal density filter-based (DWF[20], STDF[15]), smoothness optimization-based (TimeSurface[26]), and motion estimation-based (EvFlow[44]). The first and second rows display the denoised and residual noise visualization, respectively. It can be observed that DED effectively eliminates almost all the scattered BA noise with good preservation of the inherent structured feature, while the others exhibit residual noise and varying degrees of signal event damage. Statistical analysis on the global distribution of the denoised event is shown in the third row. Compared to others, DED demonstrates a higher number of blank patches while maintaining a high event preservation rate, which better reflects the overall sparsity and local concentration of the ideal event stream. Moreover, better denoising often leads to higher-quality reconstruction. In the last row, the intensity images reconstructed by E2VID [39] indicate that DED accurately restores nighttime scene information, such as leaves, building contours, and road signs", + "bbox": [ + 496, + 599, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25641", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cad2b81f803c427243459a40ec7582b05003a9a556a9c984cc57f638cdf0989.jpg", + "image_caption": [ + "(a) Spike responses with different thresholds" + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 470, + 145 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfdc55ad0cd37f28cf8a01602ab2a68dfd9ac575d8acba6a9860dad7514cc7a9.jpg", + "image_caption": [ + "(b) Framework of the DTSNN", + "Figure 7. An overview of the DTSNN model. (a) The dynamic threshold mechanism of LIF neuron. (b) The model with two branches consisting of DTB and EDB. Please refer to our supplementary material for more details about network architecture." + ], + "image_footnote": [], + "bbox": [ + 80, + 157, + 468, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with lower NIQE [35], which further indirectly validates the accuracy of denoised results and also confirms the superior effect of the processed data on reconstruction.", + "bbox": [ + 75, + 388, + 468, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. DTSNN for Event Denoising", + "text_level": 1, + "bbox": [ + 76, + 450, + 339, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to artificial neural networks(ANN), spiking neural network (SNN) have lower precision but are inherently suited for processing event-driven data due to their and information transfer mode temporal dynamics, gradually gaining prominence in event-based tasks [48, 49, 51, 52].", + "bbox": [ + 75, + 481, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Threshold Mechanism. Adjustable spiking neuron offers enhanced biological plausibility [7, 13, 43]. Indeed, the neuron firing threshold, similar to the event camera threshold parameter, functions as controlling the output. Intuitively, the dynamic threshold (DT) mechanism can adapt to more complicated situations since a high threshold could suppress noisy input when noise dominates, while a low one helps to sensitively preserve the expected signals. Inspired by this, we transformed the fixed threshold into dynamic ones based on the LIF neuron model [19] to mimic this anisotropic biomechanism. The tendency of decreasing spike quantity of the postsynaptic neurons responses with an increasing threshold as illustrated in Fig. 7(a).", + "bbox": [ + 75, + 561, + 468, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DTSNN Model. Therefore, we propose a fully SNN (all synaptic operations are SNN-based) incorporating learnable DT for event denoising, namely DTSNN. As illustrated in Fig. 7 (b), DTSNN consists of a dynamic threshold branch (DTB) and event denoising branch (EDB). The former dynamically generates a threshold map based on the successive event input, indicating the approximate spatial regions of signal or noise. This map is then passed to the latter, controlling the spiking process of LIF neurons in the last layer,", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which is depicted in Fig. 7 (b) and can be formulated as:", + "bbox": [ + 500, + 90, + 872, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} H ^ {n, t} = V ^ {n, t - 1} + U ^ {n, t} \\\\ S ^ {n, t} = H e a \\left(H ^ {n, t} - \\mathcal {F} \\left(V _ {\\text {t h}} ^ {t}\\right)\\right) \\\\ V ^ {n, t} = V _ {\\text {r e s e t}} S ^ {n, t} + \\frac {1}{\\tau} H ^ {n, t} \\odot (1 - S ^ {n, t}) \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 111, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $n$ and $t$ denote the layer number and time step respectively. $H^{n,t}$ is the membrane potential which is produced by coupling the spatial feature $U^{n,t}$ and the temporal input $V^{n,t-1}$ . The DT map $\\mathcal{F}(V_{\\mathrm{th}}^t)$ determines whether the output spiking matrix $S^{n,t}$ should be fired or stay as zero, formulating the final denoised event matrix $\\pmb{X}$ . $Hea(x) = 1$ represents the Heaviside step function when $x \\geq 0$ , otherwise $Hea(x) = 0$ , and $\\odot$ means a element-wise multiplication.", + "bbox": [ + 498, + 185, + 890, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, both two branches receive the same consecutive data as input, which is a binary event frame within a certain time window of $\\Delta t$ . The denoising branch is supervised by the GT of signal events within $\\Delta t$ , while the threshold map label of the dynamic threshold branch comes from the signal events over a longer temporal period.", + "bbox": [ + 496, + 306, + 890, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 415, + 632, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiments Setup", + "text_level": 1, + "bbox": [ + 500, + 439, + 683, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use a combined loss consisting of the L1-norm and BCE to train the proposed network. Our models are implemented with the open-source framework SpikingJelly [14], using NVIDIA A100 GPU. The Adam optimizer is employed with a batch size of 8 and a learning rate of 0.002. A fixed threshold of 0.5 is chosen for the rest of the neurons except at the last membrane potential layer. Besides, the reset value $V_{\\mathrm{reset}}$ and the membrane time constant $\\tau$ of all LIF neurons are set to 0 and 2, respectively. Datasets. We conduct quantitative experiments on existing paired datasets. Due to the enormous scale of billions of events in LED, training them all would be extremely time-consuming and resource-intensive. Therefore, 600 sequences were randomly selected for training with each consisting of consecutive 10 segments, and 60 sequences were randomly chosen for testing. To further evaluate the performance of the discrepancy of different datasets, we qualitatively test on typical public datasets DSEC [18] and E-MLB [8]. For event-based denoising methods, besides the methods in Sec.3.3 we select the representative supervised denoising methods including EDnCNN [2] and AEDNet[12].", + "bbox": [ + 496, + 462, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Methods. We use an index of event denoising accuracy $DA = \\frac{1}{2}\\left(\\frac{TP}{GP} +\\frac{TN}{GN}\\right)$ to measure the denoising performance on LED, where TP, TN, GP and GN are the number of true signal, true noise, total signal, and total noise respectively, forming two parts: signal retain (SR) and noise removal (NR). The metrics on other paired datasets are adopted from their proposed ones. Moreover, visualization results was qualitatively evaluated on other datasets.", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25642", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74efc3c1ddc5247ce45110fb35169883c356218366efd09583afe6382e0e1cf4.jpg", + "image_caption": [ + "Figure 8. Visual comparisons on LED. Comparing with state-of-the-arts, DTSNN achieves excellent denoising results and it is capable of removing noise events scattered around in sky, ground, and structural contours, simultaneously better preserve scene details." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 328 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3919f0adfb27e5dfae5a0370515998b332852c159eda044bcfe760902307a6f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsLEDDVS CLEANDVS NOISE20Average
MetricsSR↑NR↑DA↑SNR↑RPMD↓Runtime↓
Knoise[24]16.498.957.625.2117.641.33 ms
DWF[20]42.683.963.326.9631.393.67 ms
STDF[15]37.499.168.319.3030.232.04 ms
TS[26]30.698.364.513.9815.255.76 ms
EvFlow[44]52.396.474.423.7422.50150.5 ms
EDnCNN[2]80.082.081.020.2922.25251.4 ms
AEDNet[12]81.283.682.425.5818.51708.1 ms
DTSNN86.086.586.229.2616.138.20 ms
", + "bbox": [ + 76, + 380, + 468, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results comparison on different datasets. We mark the best and second best.", + "bbox": [ + 76, + 546, + 468, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 588, + 299, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The main quantitative results are presented in Table 2. In summary, the proposed method achieves the best overall results on three datasets, followed by AEDNet. Particularly, DTSNN balances signal retention and noise removal effectively, demonstrating the highest denoising accuracy. Meanwhile, our model also achieved faster inference speed on average among them. Notably, the runtime from $120 * 120$ events indicates that the efficiency of grid-based event representation and processing is significantly higher than the manner based on a single-event level, although the latter better preserves the asynchronous property of the event.", + "bbox": [ + 75, + 611, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 786, + 287, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on LED Dataset. To further evaluate the denoising performance, we compare with the qualitative results of typical methods on LED. As shown in Fig. 8, DTSNN outperforms other methods by achieving visually pleasing results, which not only effectively removes noise events in areas such as the sky, ground and building con", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6523c29e1f781551da43c365219b16e27a0c4233fefb2e60845b2f37554ce107.jpg", + "image_caption": [ + "Figure 9. Illustration of the LED diversity. We train DTSNN on different datasets: DVSCLEAN, DVSNOISE20 and LED, and test on real unpaired datasets (from left to right: DSEC $\\rightarrow$ E-MLB). The model trained on LED has achieved better denoising results." + ], + "image_footnote": [], + "bbox": [ + 500, + 380, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tours, but also preserve the structural features of the scene and objects well. For example, the zoomed region of first row in Fig. 8 highlights that our method is the only one successfully preserving the information of small airborne target after denoising. Similarly, our method also stands out in retain lane marking located in the bottom-left of second row.", + "bbox": [ + 496, + 670, + 890, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Public Datasets. To evaluate the generalization across different datasets, we also train DTSNN on synthetic event denoising dataset DVSCLEAN [12] and real one DVSNoise20 [2] respectively, testing on DSEC [18] and E-MLB [8], the former is resolution of $640^{*}480$ and the latter is resolution of $346^{*}260$ . As shown in Fig. 9, the model trained on DVSCLEAN performs poorly due to the huge domain gap between. DVSNoise20 excels at removing noise effectively but inevitably loses some details. The", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25643", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b95fafc3640319eeedd7ac5e26b22757d12ea4f390cb3f3f50f3111cf878e018.jpg", + "image_caption": [ + "Figure 10. Ablation study of the dual events denoising framework. The first column represents the noisy raw and the remaining three columns represent the flow without spatial similarity, without spatiotemporal correlation constraint and complete DED" + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69a43a6f414ea39d3f451fd4fc027324718c8c95f71edc4957bb02a712bf8e6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ArchitectureDenoising AccuracyEnergy Ratio
ANN84.518×
SNN86.01×
", + "bbox": [ + 102, + 242, + 439, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model trained on LED achieves satisfactory results across various scenes in these datasets which simultaneously removes noise without apparent signal loss, strongly supporting the great generalization of the proposed LED dataset.", + "bbox": [ + 76, + 321, + 467, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study and Discussion", + "text_level": 1, + "bbox": [ + 76, + 393, + 346, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of DED Framework. The DED framework aims to fully exploit the property of noise inconsistency /signal consistency in dual events and the self-correlation in a single event stream. As shown in Fig. 10, the small target signal is overwhelmed by noise events, utilizing solely spatiotemporal correlation constraints can only remove partial noise events because some of them also satisfy this relationship. Relying on the spatial similarity of dual event streams effectively disassembles the mutual relationship in the dense noise event group. Therefore, combining both procedures can achieve a more refined denoising result.", + "bbox": [ + 75, + 417, + 467, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Spiking Neuron. Intuitively, the inherent spatiotemporal sparsity of event streams seamlessly aligns with SNN because of the temporal information ability of spiking neurons. To compare the performance differences between SNN and ANN architectures in the event denoising tasks, the evaluation was conducted on the same structure. As shown in Table 3, the SNN based on LIF neuron achieves better denoising accuracy while maintaining $18 \\times$ lower power consumption advantage compared to its ANN version (the spiking neurons replaced with ReLU).", + "bbox": [ + 75, + 584, + 467, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of DT Module. To evaluate the effects of the proposed learnable threshold mechanism of DTSNN, we train the network with various fixed threshold neurons and dynamic threshold neurons, respectively. As can be observed from Table 4 and Fig. 11 (a), the denoising accuracy improves after introducing the DT module due to the network has the tendency to increasing spiking probability for signal pixels according to threshold prediction, and the Fig. 11 (b) illustrates DT module could assist in locating the approximate signal or noise region to retain more signal events from structural area compared to the FT model.", + "bbox": [ + 75, + 734, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c01e4afe428d8a84921e8c4b8360c27458ac2ad6065abbaca991f12d07ed2d8.jpg", + "image_caption": [ + "(a) Inference process", + "(b) Denoised results" + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 754, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e76dfc37e764591ebd63fb682e8268a4e361fa4ffa58923df0002092348c432.jpg", + "image_caption": [ + "Figure 11. Illustration of the DT model Effectiveness. (a) Inference accuracy of SNN with various fixed thresholds and with DT. (b) Denoised results without DT and with DT, respectively." + ], + "image_footnote": [], + "bbox": [ + 759, + 89, + 887, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82915b09feacbad9eac7988ef42c25e246c322dbb7a0b5449666790f529ee07e.jpg", + "table_caption": [ + "Table 3. Comparison between ANN and SNN." + ], + "table_footnote": [], + "table_body": "
TypeFTSNN (0.3 th)FTSNN (0.5 th)FTSNN (0.8 th)DTSNN
Accuracy85.586.085.786.2
", + "bbox": [ + 509, + 306, + 879, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Inference accuracy with proposed methods implemented.", + "bbox": [ + 500, + 363, + 888, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. Essentially, true events comes from intensity changes. When artificial light source variations exist, the light source itself, its coverage area and reflection from ground texture would result in relevant events. These signal events exhibit noise-like features with discontinuous structure, while the proposed DED framework considers them as signal event that should be preserved. Thus, with the supervision of such signal labels, the proposed model also tend to retain such events arise from lamp as shown in left two columns of Fig. 9 despite their visually unappealing effect.", + "bbox": [ + 496, + 393, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. The DED framework may fail in extreme cases where similar part are dominated by noise events. However, such globally anomalous high-frequency noise is rare in practice. Moreover, the proposed model struggles to capture those insignificant signals attaching to structural periphery, limiting signal retain performance in metrics.", + "bbox": [ + 496, + 544, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 617, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we construct a new large-scale high-quality paired real event denoising dataset LED, which provides diverse noise levels and scenes which aims to cover various internal and external conditions. Based on the consistency difference of noise and signal events, the proposed DED framework could effectively generate high-quality GT and detailed analysis confirms our better results than others. In addition, we propose a novel baseline DTSNN for event denoising featuring with dynamic threshold mechanism of LIF neuron. The experimental results demonstrate the superiority of the proposed dataset and denoising method.", + "bbox": [ + 496, + 674, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. The computation is completed in the HPC Platform of Huazhong University of Science and Technology. This work was supported by the National Natural Science Foundation of China under Grant 62371203.", + "bbox": [ + 496, + 840, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25644", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[2] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2020. 1, 2, 6, 7", + "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2", + "[4] Clément Cabriel, Tual Monfort, Christian G Specht, and Ignacio Izeddin. Event-based vision sensor for fast and dense single-molecule localization microscopy. Nature Photonics, pages 1-9, 2023. 1", + "[5] Haosheng Chen, David Suter, Qiangqiang Wu, and Hanzi Wang. End-to-end learning of object motion estimation from retinal events for event-based object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10534-10541, 2020. 1", + "[6] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 303-311, 2022. 2", + "[7] Jianchuan Ding, Bo Dong, Felix Heide, Yufei Ding, Yunduo Zhou, Baocai Yin, and Xin Yang. Biologically inspired dynamic thresholds for spiking neural networks. Advances in Neural Information Processing Systems, 35:6090-6103, 2022. 6", + "[8] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-mlb: Multilevel benchmark for event-based camera denoising. IEEE Transactions on Multimedia, 2023. 2, 6, 7", + "[9] Peiqi Duan, Zihao W Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos K Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 8261-8275, 2021. 2", + "[10] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 2", + "[11] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1", + "[12] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. 2, 6, 7", + "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6", + "[14] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 6", + "[15] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020. 2, 5, 7", + "[16] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3867-3876, 2018. 2", + "[17] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrio, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1", + "[18] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 2021. 6, 7", + "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 6", + "[20] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):785-795, 2022. 2, 5, 7", + "[21] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 2", + "[22] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 1", + "[23] Tao Huang, Songjiang Li, Xu Jia, Hutchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 2", + "[24] Alireza Khodamoradi and Ryan Kastner. $o(n)\\mathrm{o}(n)$ -space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Transactions on Emerging Topics in Computing, 9(1):15-23, 2018. 7" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25645", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2", + "[26] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. Hots: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 5, 7", + "[27] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2", + "[28] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Trans. Graph., 38(6):164-1, 2019. 2", + "[29] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15~\\mu \\mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1, 4", + "[30] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In 2015 IEEE International Symposium on Circuits and Systems (ISCAS), pages 722-725. IEEE, 2015. 2", + "[31] Ziwei Liu, Lu Yuan, Xiaou Tang, Matt Uytendaele, and Jian Sun. Fast burst images denoising. ACM Transactions on Graphics (TOG), 33(6):1-9, 2014. 2", + "[32] Ali Maleky, Shayan Kousha, Michael S Brown, and Marcus A Brubaker. Noise2noiseflow: realistic camera noise modeling without clean images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17632-17641, 2022. 2", + "[33] Rohit Mangalwedhekar, Nivedita Singh, Chetan Singh Thakur, Chandra Sekhar Seelamantula, Mini Jose, and Deepak Nair. Achieving nanoscale precision using neuromorphic localization microscopy. Nature Nanotechnology, pages 1-10, 2023. 1", + "[34] Ben Mildenhall, Jonathan T Barron, Jiawen Chen, Dillon Sharlet, Ren Ng, and Robert Carroll. Burst denoising with kernel prediction networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2502-2510, 2018. 2", + "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012. 6", + "[36] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 1", + "[37] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639-16652, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982, 2018. 1", + "[39] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 5", + "[40] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12300-12308, 2019. 1", + "[41] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7244-7253, 2019. 1", + "[42] Cadence Design Systems. Software downloads-cadence design systems. https://www.cadence.com/en_US/home/support/software-downloads.html. 4", + "[43] Siqi Wang, Tee Hiang Cheng, and Meng-Hiot Lim. Ltmd: Learning improvement of spiking neural networks with learnable thresholding neurons and moderate dropout. Advances in Neural Information Processing Systems, 35: 28350-28362, 2022. 6", + "[44] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 2, 5, 7", + "[45] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1609–1619, 2020. 2", + "[46] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1", + "[47] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Transactions on Multimedia, 23:1148-1159, 2020. 2", + "[48] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 6", + "[49] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 6", + "[50] Pei Zhang, Zhou Ge, Li Song, and Edmund Y L. Neuromorphic imaging with density-based spatiotemporal denoising. IEEE Transactions on Computational Imaging, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25646", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xiang Zhang, Wei Liao, Lei Yu, Wen Yang, and Gui-Song Xia. Event-based synthetic aperture imaging with a hybrid network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14235-14244, 2021. 6", + "[52] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3594-3604, 2022. 6" + ], + "bbox": [ + 78, + 90, + 470, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25647", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_model.json b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e25fc4c993ee9d20109f18910d1b8b6a63f2058d --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.87, + 0.152 + ], + "angle": 0, + "content": "LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.181, + 0.542, + 0.199 + ], + "angle": 0, + "content": "Yuxing Duan" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.199, + 0.812, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Lab of Multispectral Information Intelligent Processing Technology \n\\(^{1}\\)Huazhong University of Science and Technology" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.269, + 0.896, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.467, + 0.893, + 0.496 + ], + "angle": 0, + "content": "Figure 1. Illustration of the paired raw/denoised event datasets. The proposed real-world event denoising datasets contribute to addressing the inevitable Background Activity (BA) noise encountered by event cameras application under varying camera settings and illumination." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.508, + 0.314, + 0.523 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.897 + ], + "angle": 0, + "content": "Event camera has significant advantages in capturing dynamic scene information while being prone to noise interference, particularly in challenging conditions like low threshold and low illumination. However, most existing research focuses on gentle situations, hindering event camera applications in realistic complex scenarios. To tackle this limitation and advance the field, we construct a new paired real-world event denoising dataset (LED), including 3K sequences with 18K seconds of high-resolution (1200*680) event streams and showing three notable distinctions compared to others: diverse noise levels and scenes, larger-scale with high-resolution, and high-quality GT. Specifically, it contains stepped parameters and varying illumination with diverse scenarios. Moreover, based on the property of noise events inconsistency and signal events consistency, we propose a novel effective denoising framework(DED) using homogeneous dual events to generate the GT with better separating noise from the raw. Furthermore, we design a bio-inspired baseline leveraging Leaky-Integrate-and-Fire (LIF) neurons with dynamic thresholds to realize accurate denoising. The experimental results demonstrate that the remarkable performance of the proposed approach on different datasets. The dataset and code are at https://github.com/Yee-Sing/led." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.508, + 0.631, + 0.523 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Event cameras possess unique imaging advantages and are demonstrating tremendous potential for various applications [4, 11, 33, 37]. Unlike traditional frame cameras using integration sampling [46], event camera is differential sampling [29], evading the restriction of exposure period. However, due to this sampling mechanism, random fluctuations in analog signals can easily form noise events. In contrast to the dominance of signal components in images, the noise and signal in event cameras are encoded with same magnitude, easily disturbing inherent structural features, such as lane lines shown in Fig. 1. BA noise detrimentally impact on subsequent tasks like reconstruction [36, 39], motion estimation [40, 41] and event-based object detection [5]. Therefore, event denoising becomes a fundamental issue, notably in the booming event-based vision community." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Paired data is a key issue for event denoising in the deep-learning era, thus an intuitive idea is to construct paired data through simulators [17, 22, 38]. However, the image-based simulation cannot be equivalent to reality because of the domain gap. To obtain the 'clean part' from the real-world event, multi-modal data was incorporated to help paired event generation. APS images gradient combined with IMU are used to indicate the event probability within frame period [2]. By jointly estimating the motion parameters with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25637" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.875, + 0.206 + ], + "angle": 0, + "content": "
DatasetsSequencesCapture/sSourceNoise levelIlluminationResolutionPaired
RGB-DAVIS [45]20122CameraXDaytime180*190X
EventNFS [10]1004238Screen RecordingXDaytime224*124X
E-MLB [8]12007300Camera4Daytime to Nighttime346*260X
DVSNOISE20 [2]48807CameraXDaytime346*260✓
ED-KoGTL [1]415CameraXDaytime to Nighttime346*260✓
DVSCLEAN [12]14455Synthesis2Daytime1280*720✓
LED300018000Camera4Daytime to Nighttime1200*680✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.32, + 0.211, + 0.648, + 0.224 + ], + "angle": 0, + "content": "Table 1. Summary of existing event denoising datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.468, + 0.403 + ], + "angle": 0, + "content": "APS images and event streams [9, 45], the temporal mismatch can be alleviated. Constructing a fixed-track setup [1], the events under various illuminations can be referred to APS images in good lighting. Essentially, these methods align with events via transformed multi-mode features. Unfortunately, this heterogeneous aided information cannot fully match with the event, and their effectiveness is unavailable in situations such as low light and motion blur. Hence, few datasets have considered practical scenes where BA noise is easily deteriorated by camera setting and illumination, appearing as complex and nonuniform distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.468, + 0.6 + ], + "angle": 0, + "content": "To address this issue, we construct a large-scale, high-quality paired dataset LED for event denoising. LED has three distinct advantages. Firstly, it contains diverse noise levels, covering complex situations under various camera settings and illumination. Secondly, LED includes such as outdoor and indoor scenarios, with abundant objects. Last but not least, LED is collected with the current largest-scale and high-resolution. Moreover, inspired by the multisampling images denoising methods [21, 28, 31, 34], we are the first to explore event multi-sampling for denoising. Based on the homogeneous data, we proposed a dual-events denoising framework that can generate higher-quality GT by more accurately separating BA noise from the raw." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.601, + 0.451, + 0.616 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.468, + 0.675 + ], + "angle": 0, + "content": "(1) We provide a large-scale real-world paired dataset for event denoising. So far, LED is the largest paired realworld event denoising dataset (3000 sequences) and high-resolution \\((1200^{*}680)\\) with diverse noise levels and scenes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.468, + 0.752 + ], + "angle": 0, + "content": "(2) We propose a novel dual events denoising framework DED based on noise inconsistency. We provide detailed analysis to show that DED can better separate mixed signals and noisy event streams across various camera threshold parameters, light conditions and motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.468, + 0.829 + ], + "angle": 0, + "content": "(3) We introduce a novel baseline DTSNN for event denoising based on the spiking neural network, which utilizes the learnable dynamical spike threshold of the LIF neuron to accurately denoising. Extensive experiments on different real datasets verify the superiority of the proposed method." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.617, + 0.468, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.216, + 0.86 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Event Denoising Datasets. Currently, the datasets exclusively designed for event denoising with less attention com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.283 + ], + "angle": 0, + "content": "pared to the task-specific ones in the event-based community. In Table 1, we provide an all-sided summary of existing datasets, highlighting the scarcity of real paired datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.892, + 0.514 + ], + "angle": 0, + "content": "To address the lack of explicit labels, some paired datasets can be conveniently generated via simulators such as DVSCLEAN [12] and DND21 [20]. However, due to the gap between simulation and real, these simulators fail to fully reflect the actual situation. Other works focus on real paired event denoising data. DVSO noise20 [2] captures data in 16 pure background scenes. RGB-DAVIS [45] provides 20 indoor/outdoor sequences under good lighting conditions. ED-KoGTL [1] constructs indoor data with four illumination levels under constant trajectory. These paired datasets have achieved favorable results, primarily attributable to the gradient cues extracted from APS images. However, the heterogeneity between images and events limits the availability of diverse lighting conditions, motion patterns, and scenarios for constructing paired data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Event-based Denoising. Research on event denoising can be categorized into two approaches: filtering-based methods and deep learning methods. The former relies on manual priors to design discriminative models for noise removal, such as utilizing density distinction [15, 20, 30, 50] or motion continuity distinction [3, 44, 47] and motion compensation [6, 45] with CM framework [16] in the spatiotemporal domain. However, the validity of this prior knowledge is sensitive to different signal/noise event distributions, limiting their denoising accuracy. For learning-based methods, such as the CNN-based EDNCNN [2], the PointNet-based AEDNet [12] and [1] GNN-transformer were proposed successively. Additionally, the event denoising and super-resolution tasks are combined in [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-Sampling Image Denoising. The multi-sampling principle for image denoising has been extensively studied. Techniques like long exposure time or burst captures benefit the direct acquisition of clear data with increased SNR in static scenes. Further, the multi-sampling was explored to the paradigm of indirectly recovering noise-free data. Not only using paired noisy observations to learn the mapping from different noisy samples to clear counterpart [27, 32], but also employing spatial sub-sampling on a single noisy image to learn the mapping from adjacent noisy samples to GT [23, 25]. Evidently, these denoising approaches imply" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25638" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.339, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.307, + 0.285, + 0.319 + ], + "angle": 0, + "content": "(a) Distribution of LED" + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.088, + 0.573, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.307, + 0.517, + 0.319 + ], + "angle": 0, + "content": "(b) Scale of LED" + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.089, + 0.892, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.307, + 0.791, + 0.319 + ], + "angle": 0, + "content": "(c) Diversity of LED" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.894, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed dataset LED. (a) Distribution of noise level and scene of the proposed dataset. (b) Our proposed LED outperforms others in terms of sequences, capture, and resolution (Circles with numbers to indicate). (c) LED collects diverse event streams across various conditions of illumination, depth of field, and target scale." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.381, + 0.422, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.465, + 0.343, + 0.473 + ], + "angle": 0, + "content": "(a) Noise in frame camera" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.473, + 0.421, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.56, + 0.343, + 0.57 + ], + "angle": 0, + "content": "(b) Noise in event camera" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Figure 3. The substantial noise distinctions between frame/event camera. (a) For a stationary grayscale chart, the frame camera sequentially acquires noisy samples, fluctuating around a latent value within a specific distribution. (b) In a horizontal moving case, the event camera outputs binary signal serially, comprising signal events from motion gradient edges and spurious BA noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.704 + ], + "angle": 0, + "content": "that signal resampling could facilitate suppressing random circuit noise and approximating the expectation value." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.42, + 0.735 + ], + "angle": 0, + "content": "3. The High-quality Paired LED Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.35, + 0.758 + ], + "angle": 0, + "content": "3.1. Dataset Statistics and Features" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The primary factors affecting BA noise in event cameras can be categorized into intrinsic camera settings and extrinsic lighting conditions. By setting different camera thresholds, we can selectively record events with varying noise levels. Generally, a lower threshold means higher sensitivity but also increases the likelihood of BA noise. In contrast to fixed parameters in other datasets, we use four level thresholds to collect events, starting from the default value of \\(0\\%\\) and gradually decreasing to \\(-10\\%\\), \\(-20\\%\\), and \\(-30\\%\\)," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.474 + ], + "angle": 0, + "content": "as shown in Fig. 2 (a). To cope with the challenges posed by complex lighting conditions, data collection is from daytime to nighttime, encompassing a transition from high to low illumination and a mixture of natural light with artificial light sources. Thus, the impact of external conditions on the event stream distribution can also be controlled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.628 + ], + "angle": 0, + "content": "Due to the difficulty of constructing real paired datasets, the development of existing datasets remains limited, as shown in Table 1. In this work, we used a vehicle-mounted or pan-tilt equipped with event cameras featuring a \\(16\\mathrm{mm}\\) lens. Totally, we collected approximately 5 hours of event stream, including 3K sequences. The composition of noise levels and scenes is depicted in Fig. 2 (a), with roughly equal amounts collected during daytime and nighttime. Nearly 20 typical scenes were captured, including urban buildings, traffic roads, natural scenery, indoor exhibits, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our dataset goes beyond just the number of sequences and data volume, primarily aiming to consider the intricate influences of relevant factors on noise. Thus, we endow LED with diversity to reflect the impact of objects and environments on the event, which helps cover the cases of densely and non-uniformly distributed noisy situations in reality both spatially and temporally. In Fig. 2 (c), we exhibit that LED diversity is not only embodied in the richness of noise and illumination levels but also features such as depth of field and scene semantics. Meanwhile, high-resolution imaging aids the event cameras in capturing spatial information more precisely. Unfortunately, most existing ones with low resolution may hinder advanced visual tasks. Therefore, we collected data with \\(1200*680\\) resolution(after cropping), achieving high spatiotemporal resolution. Additionally, LED includes abundant objects of various scales, including aerial targets, pedestrians, and vehicles, making it well-suited for downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.253, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.265, + 0.226, + 0.275 + ], + "angle": 0, + "content": "(a) Ideal intensity input" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.089, + 0.49, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.177, + 0.49, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.265, + 0.463, + 0.275 + ], + "angle": 0, + "content": "(b) Circuit dual-sampled simulation" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.089, + 0.713, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.176, + 0.712, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.265, + 0.683, + 0.275 + ], + "angle": 0, + "content": "(c) Realistic dual-Sampled events" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.09, + 0.893, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.177, + 0.892, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.265, + 0.889, + 0.275 + ], + "angle": 0, + "content": "(d) Statistic of dual-sampled events" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.28, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Figure 4. Illustration of event camera dual-sampling analysis: (a) Given an ideal intensity input, which includes varying and steady stages, generates a series of signal events. (b) The twice samplings of the same input in the circuit model, both generate additional BA noise in previously steady stages, resulting in consistency discrepancies between noise events and signal events. (c) The visualization results of the actual dual-sampled events cumulative frame demonstrate the misalignment of inconsistent dual-sampled noise events and the alignment of coexisting signal events (two color groups indicating the respective polarities of the dual events). (d) The statistical results of the two tests also prove a low overlap rate between dual-sampled event pixels and a much higher one where signal events are present." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.38, + 0.496, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.601 + ], + "angle": 0, + "content": "Figure 5. Overview of the dual events denoising framework. (a) Our collection device consists of two identical EVK4s forming a co-axial system with a 1:1 beamsplitter. (b) We first perform spatial similarity processing to retain the consistent parts, followed by sequentially spatiotemporal correlation constraints to remove the residual small amount of noise event from the previous step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.383, + 0.637 + ], + "angle": 0, + "content": "3.2. Dual Events Denoising Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.469, + 0.689 + ], + "angle": 0, + "content": "Given a noisy event stream, the key lies in properly obtaining paired GT. In this section, we explore utilizing multi-sampling to achieve event stream denoising." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.885 + ], + "angle": 0, + "content": "Background. Multi-sampling denoising methods are commonly used in image denoising. For frame images, averaging the results of temporal multi-samples can approximate the true value for each pixel owing to multi-samples smoothing effect on random noise components, as shown in Fig. 3 (a). For event stream, due to their dynamic sampling and the binary values, temporal multi-samples may contain both signal events and BA noise, as illustrated in Fig. 3 (b), directly taking the average of multi-samples cannot yield true signal value. Although the noise forms differ between the two modalities, they ultimately originate from the inherent noise in the circuitry analog signal, while resampling fundamentally helps suppress the random noise source." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Analysis. This concept inspired us to utilize multi-sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.893, + 0.581 + ], + "angle": 0, + "content": "for event denoising, particularly when considering the binary state of events (presence or absence) where just two samplings are required to potentially identify noise. Thus, we initially investigated the noise inconsistency at the pixel level. Based on the DVS working principle [29], we build a circuit model in the Cadence [42] platform to simulate the dual-sampling with twice identical intensity input. The noisy analog signal converted from input generates BA noise at the originally steady stage compared to the ideal case in Fig. 4 (a). Clearly, the two sets of noise exhibit inconsistency, while the signals demonstrate high consistency, causing distinct alignment differences between the two cases as shown in the close-up of Fig. 4 (b)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To further explore the inter-class differences between noise and signal events in real-world dual-sampled events stream, we established a collection device as shown in Fig. 5 (a), to construct genuinely spatiotemporal dual-sampling. Two experimental conditions to validate the aforesaid circuit simulation phenomenon are a stationary camera and a moving one respectively. In the former, pure BA noise events was obtained, while the latter captured mixed events containing both noise and signal. In static tests, although at the highest noise level \\((-30\\%)\\) threshold), both two sets of events exhibit significant spatiotemporal disparity. As observed in the right zoomed-in patches at the upper row in Fig. 4 (c), the global spatial misalignment of dual BA noises events was evident. The overlap rate of the pixels triggering event is also to be less than \\(1\\%\\), as shown in the upperrow histogram of Fig. 4 (d). Not unexpectedly, in dynamic tests, regions with BA noise events remain inconsistency, while the coexistent signal events showing spatiotemporal alignment as depicted in lower-row counterpart, leading to a dramatic increase in overlap rate compared to the pure BA noise cases, as shown in the bottom histogram of Fig. 4 (d)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.895, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Figure 6. Analysis of different event denoising results on our dataset. From left to right, the first column is the raw events, and the remaining five columns represent different methods, namely DWF, STDF, TimeSurface, EvFlow, and the proposed DED. From top to bottom, the first row shows the denoised results, the second row is the residual noise, the third row denotes the statistical distribution of the denoised results and the last row represents the intensity image reconstruction corresponding to the zoomed region denoised events." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.51 + ], + "angle": 0, + "content": "**Formulation.** These results support the insight that the consistency discrepancies between dual-sampled events could naturally help distinguish noise. Therefore, we develop a denoising framework with the dual events stream, called DED. In DED framework, given the spatiotemporal synchronized dual event streams \\( Y_{1} \\) and \\( Y_{2} \\), they are composed of common latent signal \\( X \\) and respective noise streams \\( N_{1} \\) and \\( N_{2} \\). The noise model can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.513, + 0.469, + 0.529 + ], + "angle": 0, + "content": "\\[\nY _ {1} = X + N _ {1}, Y _ {2} = X + N _ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.531, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Because of the consistency of signal events and the inconsistency of noise events in \\( Y_{1} \\) and \\( Y_{2} \\), after binary grid operation of event stream according to certain temporal window \\( \\Delta t \\), we can perform spatial similarity to process them:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.593, + 0.469, + 0.61 + ], + "angle": 0, + "content": "\\[\nX ^ {*} = \\left\\| \\left\\| X + N _ {1} \\right\\| \\circ \\left\\| X + N _ {2} \\right\\| \\right\\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.473, + 0.853 + ], + "angle": 0, + "content": "where \\(\\circ\\) denotes the Hadamard product used to obtain the similar part \\(X^{*}\\) of two groups of raw event binary frames. Due to the probability of a few noises occurring simultaneously in dual-sampling, \\(X^{*}\\) may still contain some much-isolated noise compared to the raw. As shown in Fig. 5 (b), for finer denoising, we further utilize the spatiotemporal correlation of the signal event stream to remove the residual noise from the previous step. Specifically, we accumulate the events from the relevant spatiotemporal range in corresponding \\(X^{*}\\) within several consecutive \\(\\Delta t\\) to form a spatiotemporal relationship set. The presence of these few isolated noise events, which are randomly triggered in the spatiotemporal domain, disrupts the inherent spatiotemporal correlation of signal reflecting a certain motion model. Therefore, to leverage this spatiotemporal continuity, we transform it into the following minimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.855, + 0.47, + 0.896 + ], + "angle": 0, + "content": "\\[\n\\min \\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left| e _ {\\boldsymbol {x}} ^ {t _ {i} + 1} - e _ {\\boldsymbol {x}} ^ {t _ {i}} \\right|, _ {\\boldsymbol {x} \\in \\Omega , \\Omega \\in X ^ {*}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.39, + 0.893, + 0.482 + ], + "angle": 0, + "content": "where \\( e_{\\pmb{x}}^{t_i} \\) denotes the \\( i_{th} \\) occurring event according to the sorted timestamp, \\( \\pmb{x} \\) means the event spatial coordinates, \\( n \\) is the number of selected consecutive time windows, \\( \\Omega \\) represents the candidate spatial neighborhood, and \\( N \\) is total events number in \\( \\Omega \\). Additionally, the aforesaid processing is all performed on the two channels of event polarity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.74, + 0.511 + ], + "angle": 0, + "content": "3.3. Evaluation and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.893, + 0.597 + ], + "angle": 0, + "content": "The quality of GT is critical for a real paired event dataset. However, it is challenging to objectively evaluate the denoising results since directly obtaining clean data is unpractical. To fairly assess the denoising effects of different methods, we conducted a comprehensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Fig. 6 presents comparative results of representative event denoising methods could be used for GT generation: spatiotemporal density filter-based (DWF[20], STDF[15]), smoothness optimization-based (TimeSurface[26]), and motion estimation-based (EvFlow[44]). The first and second rows display the denoised and residual noise visualization, respectively. It can be observed that DED effectively eliminates almost all the scattered BA noise with good preservation of the inherent structured feature, while the others exhibit residual noise and varying degrees of signal event damage. Statistical analysis on the global distribution of the denoised event is shown in the third row. Compared to others, DED demonstrates a higher number of blank patches while maintaining a high event preservation rate, which better reflects the overall sparsity and local concentration of the ideal event stream. Moreover, better denoising often leads to higher-quality reconstruction. In the last row, the intensity images reconstructed by E2VID [39] indicate that DED accurately restores nighttime scene information, such as leaves, building contours, and road signs" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.472, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.147, + 0.393, + 0.157 + ], + "angle": 0, + "content": "(a) Spike responses with different thresholds" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.159, + 0.47, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.294, + 0.358, + 0.305 + ], + "angle": 0, + "content": "(b) Framework of the DTSNN" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.368 + ], + "angle": 0, + "content": "Figure 7. An overview of the DTSNN model. (a) The dynamic threshold mechanism of LIF neuron. (b) The model with two branches consisting of DTB and EDB. Please refer to our supplementary material for more details about network architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.437 + ], + "angle": 0, + "content": "with lower NIQE [35], which further indirectly validates the accuracy of denoised results and also confirms the superior effect of the processed data on reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.34, + 0.469 + ], + "angle": 0, + "content": "4. DTSNN for Event Denoising" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.558 + ], + "angle": 0, + "content": "Compared to artificial neural networks(ANN), spiking neural network (SNN) have lower precision but are inherently suited for processing event-driven data due to their and information transfer mode temporal dynamics, gradually gaining prominence in event-based tasks [48, 49, 51, 52]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.469, + 0.76 + ], + "angle": 0, + "content": "Dynamic Threshold Mechanism. Adjustable spiking neuron offers enhanced biological plausibility [7, 13, 43]. Indeed, the neuron firing threshold, similar to the event camera threshold parameter, functions as controlling the output. Intuitively, the dynamic threshold (DT) mechanism can adapt to more complicated situations since a high threshold could suppress noisy input when noise dominates, while a low one helps to sensitively preserve the expected signals. Inspired by this, we transformed the fixed threshold into dynamic ones based on the LIF neuron model [19] to mimic this anisotropic biomechanism. The tendency of decreasing spike quantity of the postsynaptic neurons responses with an increasing threshold as illustrated in Fig. 7(a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.903 + ], + "angle": 0, + "content": "DTSNN Model. Therefore, we propose a fully SNN (all synaptic operations are SNN-based) incorporating learnable DT for event denoising, namely DTSNN. As illustrated in Fig. 7 (b), DTSNN consists of a dynamic threshold branch (DTB) and event denoising branch (EDB). The former dynamically generates a threshold map based on the successive event input, indicating the approximate spatial regions of signal or noise. This map is then passed to the latter, controlling the spiking process of LIF neurons in the last layer," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.874, + 0.107 + ], + "angle": 0, + "content": "which is depicted in Fig. 7 (b) and can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.112, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} H ^ {n, t} = V ^ {n, t - 1} + U ^ {n, t} \\\\ S ^ {n, t} = H e a \\left(H ^ {n, t} - \\mathcal {F} \\left(V _ {\\text {t h}} ^ {t}\\right)\\right) \\\\ V ^ {n, t} = V _ {\\text {r e s e t}} S ^ {n, t} + \\frac {1}{\\tau} H ^ {n, t} \\odot (1 - S ^ {n, t}) \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.186, + 0.892, + 0.307 + ], + "angle": 0, + "content": "where \\(n\\) and \\(t\\) denote the layer number and time step respectively. \\(H^{n,t}\\) is the membrane potential which is produced by coupling the spatial feature \\(U^{n,t}\\) and the temporal input \\(V^{n,t-1}\\). The DT map \\(\\mathcal{F}(V_{\\mathrm{th}}^t)\\) determines whether the output spiking matrix \\(S^{n,t}\\) should be fired or stay as zero, formulating the final denoised event matrix \\(\\pmb{X}\\). \\(Hea(x) = 1\\) represents the Heaviside step function when \\(x \\geq 0\\), otherwise \\(Hea(x) = 0\\), and \\(\\odot\\) means a element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Notably, both two branches receive the same consecutive data as input, which is a binary event frame within a certain time window of \\(\\Delta t\\). The denoising branch is supervised by the GT of signal events within \\(\\Delta t\\), while the threshold map label of the dynamic threshold branch comes from the signal events over a longer temporal period." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.633, + 0.433 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.684, + 0.457 + ], + "angle": 0, + "content": "5.1. Experiments Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Implementation Details. We use a combined loss consisting of the L1-norm and BCE to train the proposed network. Our models are implemented with the open-source framework SpikingJelly [14], using NVIDIA A100 GPU. The Adam optimizer is employed with a batch size of 8 and a learning rate of 0.002. A fixed threshold of 0.5 is chosen for the rest of the neurons except at the last membrane potential layer. Besides, the reset value \\( V_{\\mathrm{reset}} \\) and the membrane time constant \\( \\tau \\) of all LIF neurons are set to 0 and 2, respectively. Datasets. We conduct quantitative experiments on existing paired datasets. Due to the enormous scale of billions of events in LED, training them all would be extremely time-consuming and resource-intensive. Therefore, 600 sequences were randomly selected for training with each consisting of consecutive 10 segments, and 60 sequences were randomly chosen for testing. To further evaluate the performance of the discrepancy of different datasets, we qualitatively test on typical public datasets DSEC [18] and E-MLB [8]. For event-based denoising methods, besides the methods in Sec.3.3 we select the representative supervised denoising methods including EDnCNN [2] and AEDNet[12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Methods. We use an index of event denoising accuracy \\(DA = \\frac{1}{2}\\left(\\frac{TP}{GP} +\\frac{TN}{GN}\\right)\\) to measure the denoising performance on LED, where TP, TN, GP and GN are the number of true signal, true noise, total signal, and total noise respectively, forming two parts: signal retain (SR) and noise removal (NR). The metrics on other paired datasets are adopted from their proposed ones. Moreover, visualization results was qualitatively evaluated on other datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.891, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Figure 8. Visual comparisons on LED. Comparing with state-of-the-arts, DTSNN achieves excellent denoising results and it is capable of removing noise events scattered around in sky, ground, and structural contours, simultaneously better preserve scene details." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.381, + 0.47, + 0.544 + ], + "angle": 0, + "content": "
DatasetsLEDDVS CLEANDVS NOISE20Average
MetricsSR↑NR↑DA↑SNR↑RPMD↓Runtime↓
Knoise[24]16.498.957.625.2117.641.33 ms
DWF[20]42.683.963.326.9631.393.67 ms
STDF[15]37.499.168.319.3030.232.04 ms
TS[26]30.698.364.513.9815.255.76 ms
EvFlow[44]52.396.474.423.7422.50150.5 ms
EDnCNN[2]80.082.081.020.2922.25251.4 ms
AEDNet[12]81.283.682.425.5818.51708.1 ms
DTSNN86.086.586.229.2616.138.20 ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.547, + 0.47, + 0.575 + ], + "angle": 0, + "content": "Table 2. Quantitative results comparison on different datasets. We mark the best and second best." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.589, + 0.3, + 0.605 + ], + "angle": 0, + "content": "5.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The main quantitative results are presented in Table 2. In summary, the proposed method achieves the best overall results on three datasets, followed by AEDNet. Particularly, DTSNN balances signal retention and noise removal effectively, demonstrating the highest denoising accuracy. Meanwhile, our model also achieved faster inference speed on average among them. Notably, the runtime from \\(120 * 120\\) events indicates that the efficiency of grid-based event representation and processing is significantly higher than the manner based on a single-event level, although the latter better preserves the asynchronous property of the event." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.288, + 0.803 + ], + "angle": 0, + "content": "5.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Evaluation on LED Dataset. To further evaluate the denoising performance, we compare with the qualitative results of typical methods on LED. As shown in Fig. 8, DTSNN outperforms other methods by achieving visually pleasing results, which not only effectively removes noise events in areas such as the sky, ground and building con" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.381, + 0.891, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.596, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Figure 9. Illustration of the LED diversity. We train DTSNN on different datasets: DVSCLEAN, DVSNOISE20 and LED, and test on real unpaired datasets (from left to right: DSEC \\(\\rightarrow\\) E-MLB). The model trained on LED has achieved better denoising results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.761 + ], + "angle": 0, + "content": "tours, but also preserve the structural features of the scene and objects well. For example, the zoomed region of first row in Fig. 8 highlights that our method is the only one successfully preserving the information of small airborne target after denoising. Similarly, our method also stands out in retain lane marking located in the bottom-left of second row." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on Public Datasets. To evaluate the generalization across different datasets, we also train DTSNN on synthetic event denoising dataset DVSCLEAN [12] and real one DVSNoise20 [2] respectively, testing on DSEC [18] and E-MLB [8], the former is resolution of \\(640^{*}480\\) and the latter is resolution of \\(346^{*}260\\). As shown in Fig. 9, the model trained on DVSCLEAN performs poorly due to the huge domain gap between. DVSNoise20 excels at removing noise effectively but inevitably loses some details. The" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25643" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.18, + 0.468, + 0.235 + ], + "angle": 0, + "content": "Figure 10. Ablation study of the dual events denoising framework. The first column represents the noisy raw and the remaining three columns represent the flow without spatial similarity, without spatiotemporal correlation constraint and complete DED" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.243, + 0.441, + 0.294 + ], + "angle": 0, + "content": "
ArchitectureDenoising AccuracyEnergy Ratio
ANN84.518×
SNN86.01×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.296, + 0.411, + 0.309 + ], + "angle": 0, + "content": "Table 3. Comparison between ANN and SNN." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.468, + 0.382 + ], + "angle": 0, + "content": "model trained on LED achieves satisfactory results across various scenes in these datasets which simultaneously removes noise without apparent signal loss, strongly supporting the great generalization of the proposed LED dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.348, + 0.41 + ], + "angle": 0, + "content": "5.4. Ablation Study and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.468, + 0.584 + ], + "angle": 0, + "content": "Effectiveness of DED Framework. The DED framework aims to fully exploit the property of noise inconsistency /signal consistency in dual events and the self-correlation in a single event stream. As shown in Fig. 10, the small target signal is overwhelmed by noise events, utilizing solely spatiotemporal correlation constraints can only remove partial noise events because some of them also satisfy this relationship. Relying on the spatial similarity of dual event streams effectively disassembles the mutual relationship in the dense noise event group. Therefore, combining both procedures can achieve a more refined denoising result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Effectiveness of Spiking Neuron. Intuitively, the inherent spatiotemporal sparsity of event streams seamlessly aligns with SNN because of the temporal information ability of spiking neurons. To compare the performance differences between SNN and ANN architectures in the event denoising tasks, the evaluation was conducted on the same structure. As shown in Table 3, the SNN based on LIF neuron achieves better denoising accuracy while maintaining \\(18 \\times\\) lower power consumption advantage compared to its ANN version (the spiking neurons replaced with ReLU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Effectiveness of DT Module. To evaluate the effects of the proposed learnable threshold mechanism of DTSNN, we train the network with various fixed threshold neurons and dynamic threshold neurons, respectively. As can be observed from Table 4 and Fig. 11 (a), the denoising accuracy improves after introducing the DT module due to the network has the tendency to increasing spiking probability for signal pixels according to threshold prediction, and the Fig. 11 (b) illustrates DT module could assist in locating the approximate signal or noise region to retain more signal events from structural area compared to the FT model." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.755, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.239, + 0.697, + 0.25 + ], + "angle": 0, + "content": "(a) Inference process" + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.09, + 0.888, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.238, + 0.876, + 0.25 + ], + "angle": 0, + "content": "(b) Denoised results" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.256, + 0.892, + 0.298 + ], + "angle": 0, + "content": "Figure 11. Illustration of the DT model Effectiveness. (a) Inference accuracy of SNN with various fixed thresholds and with DT. (b) Denoised results without DT and with DT, respectively." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.307, + 0.88, + 0.362 + ], + "angle": 0, + "content": "
TypeFTSNN (0.3 th)FTSNN (0.5 th)FTSNN (0.8 th)DTSNN
Accuracy85.586.085.786.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.364, + 0.89, + 0.378 + ], + "angle": 0, + "content": "Table 4. Inference accuracy with proposed methods implemented." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.544 + ], + "angle": 0, + "content": "Discussion. Essentially, true events comes from intensity changes. When artificial light source variations exist, the light source itself, its coverage area and reflection from ground texture would result in relevant events. These signal events exhibit noise-like features with discontinuous structure, while the proposed DED framework considers them as signal event that should be preserved. Thus, with the supervision of such signal labels, the proposed model also tend to retain such events arise from lamp as shown in left two columns of Fig. 9 despite their visually unappealing effect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.545, + 0.892, + 0.636 + ], + "angle": 0, + "content": "Limitation. The DED framework may fail in extreme cases where similar part are dominated by noise events. However, such globally anomalous high-frequency noise is rare in practice. Moreover, the proposed model struggles to capture those insignificant signals attaching to structural periphery, limiting signal retain performance in metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.618, + 0.664 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this paper, we construct a new large-scale high-quality paired real event denoising dataset LED, which provides diverse noise levels and scenes which aims to cover various internal and external conditions. Based on the consistency difference of noise and signal events, the proposed DED framework could effectively generate high-quality GT and detailed analysis confirms our better results than others. In addition, we propose a novel baseline DTSNN for event denoising featuring with dynamic threshold mechanism of LIF neuron. The experimental results demonstrate the superiority of the proposed dataset and denoising method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgments. The computation is completed in the HPC Platform of Huazhong University of Science and Technology. This work was supported by the National Natural Science Foundation of China under Grant 62371203." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25644" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.327 + ], + "angle": 0, + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[4] Clément Cabriel, Tual Monfort, Christian G Specht, and Ignacio Izeddin. Event-based vision sensor for fast and dense single-molecule localization microscopy. Nature Photonics, pages 1-9, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[5] Haosheng Chen, David Suter, Qiangqiang Wu, and Hanzi Wang. End-to-end learning of object motion estimation from retinal events for event-based object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10534-10541, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.47, + 0.527 + ], + "angle": 0, + "content": "[6] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 303-311, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.47, + 0.598 + ], + "angle": 0, + "content": "[7] Jianchuan Ding, Bo Dong, Felix Heide, Yufei Ding, Yunduo Zhou, Baocai Yin, and Xin Yang. Biologically inspired dynamic thresholds for spiking neural networks. Advances in Neural Information Processing Systems, 35:6090-6103, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[8] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-mlb: Multilevel benchmark for event-based camera denoising. IEEE Transactions on Multimedia, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.47, + 0.741 + ], + "angle": 0, + "content": "[9] Peiqi Duan, Zihao W Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos K Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 8261-8275, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.744, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[10] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.123, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[14] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[15] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[16] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3867-3876, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[17] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrio, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[18] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[20] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):785-795, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[21] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.773 + ], + "angle": 0, + "content": "[22] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[23] Tao Huang, Songjiang Li, Xu Jia, Hutchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[24] Alireza Khodamoradi and Ryan Kastner. \\(o(n)\\mathrm{o}(n)\\)-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Transactions on Emerging Topics in Computing, 9(1):15-23, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "25645" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[25] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[26] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. Hots: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.278 + ], + "angle": 0, + "content": "[27] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.281, + 0.47, + 0.336 + ], + "angle": 0, + "content": "[28] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Trans. Graph., 38(6):164-1, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[29] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15~\\mu \\mathrm{s}\\) latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[30] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In 2015 IEEE International Symposium on Circuits and Systems (ISCAS), pages 722-725. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.511 + ], + "angle": 0, + "content": "[31] Ziwei Liu, Lu Yuan, Xiaou Tang, Matt Uytendaele, and Jian Sun. Fast burst images denoising. ACM Transactions on Graphics (TOG), 33(6):1-9, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.513, + 0.47, + 0.582 + ], + "angle": 0, + "content": "[32] Ali Maleky, Shayan Kousha, Michael S Brown, and Marcus A Brubaker. Noise2noiseflow: realistic camera noise modeling without clean images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17632-17641, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.47, + 0.655 + ], + "angle": 0, + "content": "[33] Rohit Mangalwedhekar, Nivedita Singh, Chetan Singh Thakur, Chandra Sekhar Seelamantula, Mini Jose, and Deepak Nair. Achieving nanoscale precision using neuromorphic localization microscopy. Nature Nanotechnology, pages 1-10, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.725 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Jonathan T Barron, Jiawen Chen, Dillon Sharlet, Ren Ng, and Robert Carroll. Burst denoising with kernel prediction networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2502-2510, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.841 + ], + "angle": 0, + "content": "[36] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[37] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639-16652, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[38] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[39] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[40] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12300-12308, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[41] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7244-7253, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[42] Cadence Design Systems. Software downloads-cadence design systems. https://www.cadence.com/en_US/home/support/software-downloads.html. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[43] Siqi Wang, Tee Hiang Cheng, and Meng-Hiot Lim. Ltmd: Learning improvement of spiking neural networks with learnable thresholding neurons and moderate dropout. Advances in Neural Information Processing Systems, 35: 28350-28362, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[44] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[45] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1609–1619, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[46] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[47] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Transactions on Multimedia, 23:1148-1159, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[48] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[49] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Pei Zhang, Zhou Ge, Li Song, and Edmund Y L. Neuromorphic imaging with density-based spatiotemporal denoising. IEEE Transactions on Computational Imaging, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25646" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[51] Xiang Zhang, Wei Liao, Lei Yu, Wen Yang, and Gui-Song Xia. Event-based synthetic aperture imaging with a hybrid network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14235-14244, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[52] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3594-3604, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25647" + } + ] +] \ No newline at end of file diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_origin.pdf b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0969df5a84d0fae88a0e18d913131fa05c639ba6 --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/c132a07a-cde4-41ed-ba0e-fbc21a44ee59_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46d8d9e3c64d88a13154471a8f3eccdc13aa3f58d0864463f9b0ff94566e5926 +size 9796683 diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/full.md b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/full.md new file mode 100644 index 0000000000000000000000000000000000000000..200c0124dcc454991b13d6a46307bc0205906d40 --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/full.md @@ -0,0 +1,294 @@ +# LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising + +Yuxing Duan + +$^{1}$ National Key Lab of Multispectral Information Intelligent Processing Technology + $^{1}$ Huazhong University of Science and Technology + +![](images/2f38453187383b2d327912d910f61ae52a5357f0214b2b71f7e96429409db0ca.jpg) +Figure 1. Illustration of the paired raw/denoised event datasets. The proposed real-world event denoising datasets contribute to addressing the inevitable Background Activity (BA) noise encountered by event cameras application under varying camera settings and illumination. + +# Abstract + +Event camera has significant advantages in capturing dynamic scene information while being prone to noise interference, particularly in challenging conditions like low threshold and low illumination. However, most existing research focuses on gentle situations, hindering event camera applications in realistic complex scenarios. To tackle this limitation and advance the field, we construct a new paired real-world event denoising dataset (LED), including 3K sequences with 18K seconds of high-resolution (1200*680) event streams and showing three notable distinctions compared to others: diverse noise levels and scenes, larger-scale with high-resolution, and high-quality GT. Specifically, it contains stepped parameters and varying illumination with diverse scenarios. Moreover, based on the property of noise events inconsistency and signal events consistency, we propose a novel effective denoising framework(DED) using homogeneous dual events to generate the GT with better separating noise from the raw. Furthermore, we design a bio-inspired baseline leveraging Leaky-Integrate-and-Fire (LIF) neurons with dynamic thresholds to realize accurate denoising. The experimental results demonstrate that the remarkable performance of the proposed approach on different datasets. The dataset and code are at https://github.com/Yee-Sing/led. + +# 1. Introduction + +Event cameras possess unique imaging advantages and are demonstrating tremendous potential for various applications [4, 11, 33, 37]. Unlike traditional frame cameras using integration sampling [46], event camera is differential sampling [29], evading the restriction of exposure period. However, due to this sampling mechanism, random fluctuations in analog signals can easily form noise events. In contrast to the dominance of signal components in images, the noise and signal in event cameras are encoded with same magnitude, easily disturbing inherent structural features, such as lane lines shown in Fig. 1. BA noise detrimentally impact on subsequent tasks like reconstruction [36, 39], motion estimation [40, 41] and event-based object detection [5]. Therefore, event denoising becomes a fundamental issue, notably in the booming event-based vision community. + +Paired data is a key issue for event denoising in the deep-learning era, thus an intuitive idea is to construct paired data through simulators [17, 22, 38]. However, the image-based simulation cannot be equivalent to reality because of the domain gap. To obtain the 'clean part' from the real-world event, multi-modal data was incorporated to help paired event generation. APS images gradient combined with IMU are used to indicate the event probability within frame period [2]. By jointly estimating the motion parameters with + +
DatasetsSequencesCapture/sSourceNoise levelIlluminationResolutionPaired
RGB-DAVIS [45]20122CameraXDaytime180*190X
EventNFS [10]1004238Screen RecordingXDaytime224*124X
E-MLB [8]12007300Camera4Daytime to Nighttime346*260X
DVSNOISE20 [2]48807CameraXDaytime346*260✓
ED-KoGTL [1]415CameraXDaytime to Nighttime346*260✓
DVSCLEAN [12]14455Synthesis2Daytime1280*720✓
LED300018000Camera4Daytime to Nighttime1200*680✓
+ +Table 1. Summary of existing event denoising datasets. + +APS images and event streams [9, 45], the temporal mismatch can be alleviated. Constructing a fixed-track setup [1], the events under various illuminations can be referred to APS images in good lighting. Essentially, these methods align with events via transformed multi-mode features. Unfortunately, this heterogeneous aided information cannot fully match with the event, and their effectiveness is unavailable in situations such as low light and motion blur. Hence, few datasets have considered practical scenes where BA noise is easily deteriorated by camera setting and illumination, appearing as complex and nonuniform distributions. + +To address this issue, we construct a large-scale, high-quality paired dataset LED for event denoising. LED has three distinct advantages. Firstly, it contains diverse noise levels, covering complex situations under various camera settings and illumination. Secondly, LED includes such as outdoor and indoor scenarios, with abundant objects. Last but not least, LED is collected with the current largest-scale and high-resolution. Moreover, inspired by the multisampling images denoising methods [21, 28, 31, 34], we are the first to explore event multi-sampling for denoising. Based on the homogeneous data, we proposed a dual-events denoising framework that can generate higher-quality GT by more accurately separating BA noise from the raw. + +The contributions of this paper can be summarized as: + +(1) We provide a large-scale real-world paired dataset for event denoising. So far, LED is the largest paired realworld event denoising dataset (3000 sequences) and high-resolution $(1200^{*}680)$ with diverse noise levels and scenes. +(2) We propose a novel dual events denoising framework DED based on noise inconsistency. We provide detailed analysis to show that DED can better separate mixed signals and noisy event streams across various camera threshold parameters, light conditions and motion patterns. +(3) We introduce a novel baseline DTSNN for event denoising based on the spiking neural network, which utilizes the learnable dynamical spike threshold of the LIF neuron to accurately denoising. Extensive experiments on different real datasets verify the superiority of the proposed method. + +# 2. Related work + +Event Denoising Datasets. Currently, the datasets exclusively designed for event denoising with less attention com + +pared to the task-specific ones in the event-based community. In Table 1, we provide an all-sided summary of existing datasets, highlighting the scarcity of real paired datasets. + +To address the lack of explicit labels, some paired datasets can be conveniently generated via simulators such as DVSCLEAN [12] and DND21 [20]. However, due to the gap between simulation and real, these simulators fail to fully reflect the actual situation. Other works focus on real paired event denoising data. DVSO noise20 [2] captures data in 16 pure background scenes. RGB-DAVIS [45] provides 20 indoor/outdoor sequences under good lighting conditions. ED-KoGTL [1] constructs indoor data with four illumination levels under constant trajectory. These paired datasets have achieved favorable results, primarily attributable to the gradient cues extracted from APS images. However, the heterogeneity between images and events limits the availability of diverse lighting conditions, motion patterns, and scenarios for constructing paired data. + +Event-based Denoising. Research on event denoising can be categorized into two approaches: filtering-based methods and deep learning methods. The former relies on manual priors to design discriminative models for noise removal, such as utilizing density distinction [15, 20, 30, 50] or motion continuity distinction [3, 44, 47] and motion compensation [6, 45] with CM framework [16] in the spatiotemporal domain. However, the validity of this prior knowledge is sensitive to different signal/noise event distributions, limiting their denoising accuracy. For learning-based methods, such as the CNN-based EDNCNN [2], the PointNet-based AEDNet [12] and [1] GNN-transformer were proposed successively. Additionally, the event denoising and super-resolution tasks are combined in [10]. + +Multi-Sampling Image Denoising. The multi-sampling principle for image denoising has been extensively studied. Techniques like long exposure time or burst captures benefit the direct acquisition of clear data with increased SNR in static scenes. Further, the multi-sampling was explored to the paradigm of indirectly recovering noise-free data. Not only using paired noisy observations to learn the mapping from different noisy samples to clear counterpart [27, 32], but also employing spatial sub-sampling on a single noisy image to learn the mapping from adjacent noisy samples to GT [23, 25]. Evidently, these denoising approaches imply + +![](images/ecd600611afddee70686da699d5581c10880777dd1ae9346566ed6cbc92e300b.jpg) +(a) Distribution of LED + +![](images/0f57e42e5e0c9ca78b0dfa172ee146d6e61fc1eeab685d04520a4e56411a1152.jpg) +(b) Scale of LED + +![](images/9795416dca29c0fb8eaed60ec1247de0ce8c087ef8061fa99cb41d76455f62f2.jpg) +(c) Diversity of LED + +![](images/9746f753be59fadcb8d34d9534ac8b832e2d16c0639bd2bfda08dcfa4a15eeb1.jpg) +Figure 2. Illustration of the proposed dataset LED. (a) Distribution of noise level and scene of the proposed dataset. (b) Our proposed LED outperforms others in terms of sequences, capture, and resolution (Circles with numbers to indicate). (c) LED collects diverse event streams across various conditions of illumination, depth of field, and target scale. +(a) Noise in frame camera + +![](images/d92ae6223f94ff3c6c879929ba5eb4422379e9d92e06ff64ebdaa047af2d15d0.jpg) +(b) Noise in event camera +Figure 3. The substantial noise distinctions between frame/event camera. (a) For a stationary grayscale chart, the frame camera sequentially acquires noisy samples, fluctuating around a latent value within a specific distribution. (b) In a horizontal moving case, the event camera outputs binary signal serially, comprising signal events from motion gradient edges and spurious BA noise. + +that signal resampling could facilitate suppressing random circuit noise and approximating the expectation value. + +# 3. The High-quality Paired LED Dataset + +# 3.1. Dataset Statistics and Features + +The primary factors affecting BA noise in event cameras can be categorized into intrinsic camera settings and extrinsic lighting conditions. By setting different camera thresholds, we can selectively record events with varying noise levels. Generally, a lower threshold means higher sensitivity but also increases the likelihood of BA noise. In contrast to fixed parameters in other datasets, we use four level thresholds to collect events, starting from the default value of $0\%$ and gradually decreasing to $-10\%$ , $-20\%$ , and $-30\%$ , + +as shown in Fig. 2 (a). To cope with the challenges posed by complex lighting conditions, data collection is from daytime to nighttime, encompassing a transition from high to low illumination and a mixture of natural light with artificial light sources. Thus, the impact of external conditions on the event stream distribution can also be controlled. + +Due to the difficulty of constructing real paired datasets, the development of existing datasets remains limited, as shown in Table 1. In this work, we used a vehicle-mounted or pan-tilt equipped with event cameras featuring a $16\mathrm{mm}$ lens. Totally, we collected approximately 5 hours of event stream, including 3K sequences. The composition of noise levels and scenes is depicted in Fig. 2 (a), with roughly equal amounts collected during daytime and nighttime. Nearly 20 typical scenes were captured, including urban buildings, traffic roads, natural scenery, indoor exhibits, etc. + +Our dataset goes beyond just the number of sequences and data volume, primarily aiming to consider the intricate influences of relevant factors on noise. Thus, we endow LED with diversity to reflect the impact of objects and environments on the event, which helps cover the cases of densely and non-uniformly distributed noisy situations in reality both spatially and temporally. In Fig. 2 (c), we exhibit that LED diversity is not only embodied in the richness of noise and illumination levels but also features such as depth of field and scene semantics. Meanwhile, high-resolution imaging aids the event cameras in capturing spatial information more precisely. Unfortunately, most existing ones with low resolution may hinder advanced visual tasks. Therefore, we collected data with $1200*680$ resolution(after cropping), achieving high spatiotemporal resolution. Additionally, LED includes abundant objects of various scales, including aerial targets, pedestrians, and vehicles, making it well-suited for downstream tasks. + +![](images/b5a700a3a9fa6b12d3c367f38ee38459d3b4a2b01223f2709caf979cfc2cf9e6.jpg) +(a) Ideal intensity input + +![](images/d18ab4b6ec4e6f908c1b6777933e802b6ed92999ae2a20b6f17a0924a758b351.jpg) + +![](images/c3167d56f6087147b4ed8f56828d07fe4dc2a9ad090fe27bde62986a45def9c8.jpg) +(b) Circuit dual-sampled simulation + +![](images/689fbe4668c25bda81b832c67afd44a047f370e79ed4a62f037fbe6857be358f.jpg) + +![](images/2a87918737ebcc936b450c3c644bdbb624bb35b553664dffef76af5a5af0fd85.jpg) +(c) Realistic dual-Sampled events + +![](images/01c4b7eafad35955d0781bcbd9bd208a546e36fabbf33a69db2038e5b9e0df74.jpg) + +![](images/b90971034e6fbf5e8cc54a67a41115fe1a8e6f2e85e9ced63455b31b625b88d9.jpg) +(d) Statistic of dual-sampled events + +![](images/4a57266de4d34d4a8077f3e2c06ee67c38a66a4dc12a67ee5cba87f8244cebbe.jpg) +Figure 4. Illustration of event camera dual-sampling analysis: (a) Given an ideal intensity input, which includes varying and steady stages, generates a series of signal events. (b) The twice samplings of the same input in the circuit model, both generate additional BA noise in previously steady stages, resulting in consistency discrepancies between noise events and signal events. (c) The visualization results of the actual dual-sampled events cumulative frame demonstrate the misalignment of inconsistent dual-sampled noise events and the alignment of coexisting signal events (two color groups indicating the respective polarities of the dual events). (d) The statistical results of the two tests also prove a low overlap rate between dual-sampled event pixels and a much higher one where signal events are present. +Figure 5. Overview of the dual events denoising framework. (a) Our collection device consists of two identical EVK4s forming a co-axial system with a 1:1 beamsplitter. (b) We first perform spatial similarity processing to retain the consistent parts, followed by sequentially spatiotemporal correlation constraints to remove the residual small amount of noise event from the previous step. + +# 3.2. Dual Events Denoising Framework + +Given a noisy event stream, the key lies in properly obtaining paired GT. In this section, we explore utilizing multi-sampling to achieve event stream denoising. + +Background. Multi-sampling denoising methods are commonly used in image denoising. For frame images, averaging the results of temporal multi-samples can approximate the true value for each pixel owing to multi-samples smoothing effect on random noise components, as shown in Fig. 3 (a). For event stream, due to their dynamic sampling and the binary values, temporal multi-samples may contain both signal events and BA noise, as illustrated in Fig. 3 (b), directly taking the average of multi-samples cannot yield true signal value. Although the noise forms differ between the two modalities, they ultimately originate from the inherent noise in the circuitry analog signal, while resampling fundamentally helps suppress the random noise source. + +Analysis. This concept inspired us to utilize multi-sampling + +for event denoising, particularly when considering the binary state of events (presence or absence) where just two samplings are required to potentially identify noise. Thus, we initially investigated the noise inconsistency at the pixel level. Based on the DVS working principle [29], we build a circuit model in the Cadence [42] platform to simulate the dual-sampling with twice identical intensity input. The noisy analog signal converted from input generates BA noise at the originally steady stage compared to the ideal case in Fig. 4 (a). Clearly, the two sets of noise exhibit inconsistency, while the signals demonstrate high consistency, causing distinct alignment differences between the two cases as shown in the close-up of Fig. 4 (b). + +To further explore the inter-class differences between noise and signal events in real-world dual-sampled events stream, we established a collection device as shown in Fig. 5 (a), to construct genuinely spatiotemporal dual-sampling. Two experimental conditions to validate the aforesaid circuit simulation phenomenon are a stationary camera and a moving one respectively. In the former, pure BA noise events was obtained, while the latter captured mixed events containing both noise and signal. In static tests, although at the highest noise level $(-30\%)$ threshold), both two sets of events exhibit significant spatiotemporal disparity. As observed in the right zoomed-in patches at the upper row in Fig. 4 (c), the global spatial misalignment of dual BA noises events was evident. The overlap rate of the pixels triggering event is also to be less than $1\%$ , as shown in the upperrow histogram of Fig. 4 (d). Not unexpectedly, in dynamic tests, regions with BA noise events remain inconsistency, while the coexistent signal events showing spatiotemporal alignment as depicted in lower-row counterpart, leading to a dramatic increase in overlap rate compared to the pure BA noise cases, as shown in the bottom histogram of Fig. 4 (d). + +![](images/635af38e31bb62e5022ecac3e503890e3dd5ff13d4d6d545aa107af50c22528f.jpg) +Figure 6. Analysis of different event denoising results on our dataset. From left to right, the first column is the raw events, and the remaining five columns represent different methods, namely DWF, STDF, TimeSurface, EvFlow, and the proposed DED. From top to bottom, the first row shows the denoised results, the second row is the residual noise, the third row denotes the statistical distribution of the denoised results and the last row represents the intensity image reconstruction corresponding to the zoomed region denoised events. + +**Formulation.** These results support the insight that the consistency discrepancies between dual-sampled events could naturally help distinguish noise. Therefore, we develop a denoising framework with the dual events stream, called DED. In DED framework, given the spatiotemporal synchronized dual event streams $Y_{1}$ and $Y_{2}$ , they are composed of common latent signal $X$ and respective noise streams $N_{1}$ and $N_{2}$ . The noise model can be defined as follows: + +$$ +Y _ {1} = X + N _ {1}, Y _ {2} = X + N _ {2}. \tag {1} +$$ + +Because of the consistency of signal events and the inconsistency of noise events in $Y_{1}$ and $Y_{2}$ , after binary grid operation of event stream according to certain temporal window $\Delta t$ , we can perform spatial similarity to process them: + +$$ +X ^ {*} = \left\| \left\| X + N _ {1} \right\| \circ \left\| X + N _ {2} \right\| \right\| _ {1}, \tag {2} +$$ + +where $\circ$ denotes the Hadamard product used to obtain the similar part $X^{*}$ of two groups of raw event binary frames. Due to the probability of a few noises occurring simultaneously in dual-sampling, $X^{*}$ may still contain some much-isolated noise compared to the raw. As shown in Fig. 5 (b), for finer denoising, we further utilize the spatiotemporal correlation of the signal event stream to remove the residual noise from the previous step. Specifically, we accumulate the events from the relevant spatiotemporal range in corresponding $X^{*}$ within several consecutive $\Delta t$ to form a spatiotemporal relationship set. The presence of these few isolated noise events, which are randomly triggered in the spatiotemporal domain, disrupts the inherent spatiotemporal correlation of signal reflecting a certain motion model. Therefore, to leverage this spatiotemporal continuity, we transform it into the following minimization problem: + +$$ +\min \frac {1}{N - 1} \sum_ {i = 1} ^ {N} \left| e _ {\boldsymbol {x}} ^ {t _ {i} + 1} - e _ {\boldsymbol {x}} ^ {t _ {i}} \right|, _ {\boldsymbol {x} \in \Omega , \Omega \in X ^ {*}}, \tag {3} +$$ + +where $e_{\pmb{x}}^{t_i}$ denotes the $i_{th}$ occurring event according to the sorted timestamp, $\pmb{x}$ means the event spatial coordinates, $n$ is the number of selected consecutive time windows, $\Omega$ represents the candidate spatial neighborhood, and $N$ is total events number in $\Omega$ . Additionally, the aforesaid processing is all performed on the two channels of event polarity. + +# 3.3. Evaluation and Discussion + +The quality of GT is critical for a real paired event dataset. However, it is challenging to objectively evaluate the denoising results since directly obtaining clean data is unpractical. To fairly assess the denoising effects of different methods, we conducted a comprehensive evaluation. + +Fig. 6 presents comparative results of representative event denoising methods could be used for GT generation: spatiotemporal density filter-based (DWF[20], STDF[15]), smoothness optimization-based (TimeSurface[26]), and motion estimation-based (EvFlow[44]). The first and second rows display the denoised and residual noise visualization, respectively. It can be observed that DED effectively eliminates almost all the scattered BA noise with good preservation of the inherent structured feature, while the others exhibit residual noise and varying degrees of signal event damage. Statistical analysis on the global distribution of the denoised event is shown in the third row. Compared to others, DED demonstrates a higher number of blank patches while maintaining a high event preservation rate, which better reflects the overall sparsity and local concentration of the ideal event stream. Moreover, better denoising often leads to higher-quality reconstruction. In the last row, the intensity images reconstructed by E2VID [39] indicate that DED accurately restores nighttime scene information, such as leaves, building contours, and road signs + +![](images/7cad2b81f803c427243459a40ec7582b05003a9a556a9c984cc57f638cdf0989.jpg) +(a) Spike responses with different thresholds + +![](images/dfdc55ad0cd37f28cf8a01602ab2a68dfd9ac575d8acba6a9860dad7514cc7a9.jpg) +(b) Framework of the DTSNN +Figure 7. An overview of the DTSNN model. (a) The dynamic threshold mechanism of LIF neuron. (b) The model with two branches consisting of DTB and EDB. Please refer to our supplementary material for more details about network architecture. + +with lower NIQE [35], which further indirectly validates the accuracy of denoised results and also confirms the superior effect of the processed data on reconstruction. + +# 4. DTSNN for Event Denoising + +Compared to artificial neural networks(ANN), spiking neural network (SNN) have lower precision but are inherently suited for processing event-driven data due to their and information transfer mode temporal dynamics, gradually gaining prominence in event-based tasks [48, 49, 51, 52]. + +Dynamic Threshold Mechanism. Adjustable spiking neuron offers enhanced biological plausibility [7, 13, 43]. Indeed, the neuron firing threshold, similar to the event camera threshold parameter, functions as controlling the output. Intuitively, the dynamic threshold (DT) mechanism can adapt to more complicated situations since a high threshold could suppress noisy input when noise dominates, while a low one helps to sensitively preserve the expected signals. Inspired by this, we transformed the fixed threshold into dynamic ones based on the LIF neuron model [19] to mimic this anisotropic biomechanism. The tendency of decreasing spike quantity of the postsynaptic neurons responses with an increasing threshold as illustrated in Fig. 7(a). + +DTSNN Model. Therefore, we propose a fully SNN (all synaptic operations are SNN-based) incorporating learnable DT for event denoising, namely DTSNN. As illustrated in Fig. 7 (b), DTSNN consists of a dynamic threshold branch (DTB) and event denoising branch (EDB). The former dynamically generates a threshold map based on the successive event input, indicating the approximate spatial regions of signal or noise. This map is then passed to the latter, controlling the spiking process of LIF neurons in the last layer, + +which is depicted in Fig. 7 (b) and can be formulated as: + +$$ +\left\{ \begin{array}{l} H ^ {n, t} = V ^ {n, t - 1} + U ^ {n, t} \\ S ^ {n, t} = H e a \left(H ^ {n, t} - \mathcal {F} \left(V _ {\text {t h}} ^ {t}\right)\right) \\ V ^ {n, t} = V _ {\text {r e s e t}} S ^ {n, t} + \frac {1}{\tau} H ^ {n, t} \odot (1 - S ^ {n, t}) \end{array} \right. \tag {4} +$$ + +where $n$ and $t$ denote the layer number and time step respectively. $H^{n,t}$ is the membrane potential which is produced by coupling the spatial feature $U^{n,t}$ and the temporal input $V^{n,t-1}$ . The DT map $\mathcal{F}(V_{\mathrm{th}}^t)$ determines whether the output spiking matrix $S^{n,t}$ should be fired or stay as zero, formulating the final denoised event matrix $\pmb{X}$ . $Hea(x) = 1$ represents the Heaviside step function when $x \geq 0$ , otherwise $Hea(x) = 0$ , and $\odot$ means a element-wise multiplication. + +Notably, both two branches receive the same consecutive data as input, which is a binary event frame within a certain time window of $\Delta t$ . The denoising branch is supervised by the GT of signal events within $\Delta t$ , while the threshold map label of the dynamic threshold branch comes from the signal events over a longer temporal period. + +# 5. Experiments + +# 5.1. Experiments Setup + +Implementation Details. We use a combined loss consisting of the L1-norm and BCE to train the proposed network. Our models are implemented with the open-source framework SpikingJelly [14], using NVIDIA A100 GPU. The Adam optimizer is employed with a batch size of 8 and a learning rate of 0.002. A fixed threshold of 0.5 is chosen for the rest of the neurons except at the last membrane potential layer. Besides, the reset value $V_{\mathrm{reset}}$ and the membrane time constant $\tau$ of all LIF neurons are set to 0 and 2, respectively. Datasets. We conduct quantitative experiments on existing paired datasets. Due to the enormous scale of billions of events in LED, training them all would be extremely time-consuming and resource-intensive. Therefore, 600 sequences were randomly selected for training with each consisting of consecutive 10 segments, and 60 sequences were randomly chosen for testing. To further evaluate the performance of the discrepancy of different datasets, we qualitatively test on typical public datasets DSEC [18] and E-MLB [8]. For event-based denoising methods, besides the methods in Sec.3.3 we select the representative supervised denoising methods including EDnCNN [2] and AEDNet[12]. + +Evaluation Methods. We use an index of event denoising accuracy $DA = \frac{1}{2}\left(\frac{TP}{GP} +\frac{TN}{GN}\right)$ to measure the denoising performance on LED, where TP, TN, GP and GN are the number of true signal, true noise, total signal, and total noise respectively, forming two parts: signal retain (SR) and noise removal (NR). The metrics on other paired datasets are adopted from their proposed ones. Moreover, visualization results was qualitatively evaluated on other datasets. + +![](images/74efc3c1ddc5247ce45110fb35169883c356218366efd09583afe6382e0e1cf4.jpg) +Figure 8. Visual comparisons on LED. Comparing with state-of-the-arts, DTSNN achieves excellent denoising results and it is capable of removing noise events scattered around in sky, ground, and structural contours, simultaneously better preserve scene details. + +
DatasetsLEDDVS CLEANDVS NOISE20Average
MetricsSR↑NR↑DA↑SNR↑RPMD↓Runtime↓
Knoise[24]16.498.957.625.2117.641.33 ms
DWF[20]42.683.963.326.9631.393.67 ms
STDF[15]37.499.168.319.3030.232.04 ms
TS[26]30.698.364.513.9815.255.76 ms
EvFlow[44]52.396.474.423.7422.50150.5 ms
EDnCNN[2]80.082.081.020.2922.25251.4 ms
AEDNet[12]81.283.682.425.5818.51708.1 ms
DTSNN86.086.586.229.2616.138.20 ms
+ +Table 2. Quantitative results comparison on different datasets. We mark the best and second best. + +# 5.2. Quantitative Evaluation + +The main quantitative results are presented in Table 2. In summary, the proposed method achieves the best overall results on three datasets, followed by AEDNet. Particularly, DTSNN balances signal retention and noise removal effectively, demonstrating the highest denoising accuracy. Meanwhile, our model also achieved faster inference speed on average among them. Notably, the runtime from $120 * 120$ events indicates that the efficiency of grid-based event representation and processing is significantly higher than the manner based on a single-event level, although the latter better preserves the asynchronous property of the event. + +# 5.3. Qualitative Evaluation + +Evaluation on LED Dataset. To further evaluate the denoising performance, we compare with the qualitative results of typical methods on LED. As shown in Fig. 8, DTSNN outperforms other methods by achieving visually pleasing results, which not only effectively removes noise events in areas such as the sky, ground and building con + +![](images/6523c29e1f781551da43c365219b16e27a0c4233fefb2e60845b2f37554ce107.jpg) +Figure 9. Illustration of the LED diversity. We train DTSNN on different datasets: DVSCLEAN, DVSNOISE20 and LED, and test on real unpaired datasets (from left to right: DSEC $\rightarrow$ E-MLB). The model trained on LED has achieved better denoising results. + +tours, but also preserve the structural features of the scene and objects well. For example, the zoomed region of first row in Fig. 8 highlights that our method is the only one successfully preserving the information of small airborne target after denoising. Similarly, our method also stands out in retain lane marking located in the bottom-left of second row. + +Evaluation on Public Datasets. To evaluate the generalization across different datasets, we also train DTSNN on synthetic event denoising dataset DVSCLEAN [12] and real one DVSNoise20 [2] respectively, testing on DSEC [18] and E-MLB [8], the former is resolution of $640^{*}480$ and the latter is resolution of $346^{*}260$ . As shown in Fig. 9, the model trained on DVSCLEAN performs poorly due to the huge domain gap between. DVSNoise20 excels at removing noise effectively but inevitably loses some details. The + +![](images/b95fafc3640319eeedd7ac5e26b22757d12ea4f390cb3f3f50f3111cf878e018.jpg) +Figure 10. Ablation study of the dual events denoising framework. The first column represents the noisy raw and the remaining three columns represent the flow without spatial similarity, without spatiotemporal correlation constraint and complete DED + +
ArchitectureDenoising AccuracyEnergy Ratio
ANN84.518×
SNN86.01×
+ +model trained on LED achieves satisfactory results across various scenes in these datasets which simultaneously removes noise without apparent signal loss, strongly supporting the great generalization of the proposed LED dataset. + +# 5.4. Ablation Study and Discussion + +Effectiveness of DED Framework. The DED framework aims to fully exploit the property of noise inconsistency /signal consistency in dual events and the self-correlation in a single event stream. As shown in Fig. 10, the small target signal is overwhelmed by noise events, utilizing solely spatiotemporal correlation constraints can only remove partial noise events because some of them also satisfy this relationship. Relying on the spatial similarity of dual event streams effectively disassembles the mutual relationship in the dense noise event group. Therefore, combining both procedures can achieve a more refined denoising result. + +Effectiveness of Spiking Neuron. Intuitively, the inherent spatiotemporal sparsity of event streams seamlessly aligns with SNN because of the temporal information ability of spiking neurons. To compare the performance differences between SNN and ANN architectures in the event denoising tasks, the evaluation was conducted on the same structure. As shown in Table 3, the SNN based on LIF neuron achieves better denoising accuracy while maintaining $18 \times$ lower power consumption advantage compared to its ANN version (the spiking neurons replaced with ReLU). + +Effectiveness of DT Module. To evaluate the effects of the proposed learnable threshold mechanism of DTSNN, we train the network with various fixed threshold neurons and dynamic threshold neurons, respectively. As can be observed from Table 4 and Fig. 11 (a), the denoising accuracy improves after introducing the DT module due to the network has the tendency to increasing spiking probability for signal pixels according to threshold prediction, and the Fig. 11 (b) illustrates DT module could assist in locating the approximate signal or noise region to retain more signal events from structural area compared to the FT model. + +![](images/9c01e4afe428d8a84921e8c4b8360c27458ac2ad6065abbaca991f12d07ed2d8.jpg) +(a) Inference process +(b) Denoised results + +![](images/4e76dfc37e764591ebd63fb682e8268a4e361fa4ffa58923df0002092348c432.jpg) +Figure 11. Illustration of the DT model Effectiveness. (a) Inference accuracy of SNN with various fixed thresholds and with DT. (b) Denoised results without DT and with DT, respectively. + +Table 3. Comparison between ANN and SNN. + +
TypeFTSNN (0.3 th)FTSNN (0.5 th)FTSNN (0.8 th)DTSNN
Accuracy85.586.085.786.2
+ +Table 4. Inference accuracy with proposed methods implemented. + +Discussion. Essentially, true events comes from intensity changes. When artificial light source variations exist, the light source itself, its coverage area and reflection from ground texture would result in relevant events. These signal events exhibit noise-like features with discontinuous structure, while the proposed DED framework considers them as signal event that should be preserved. Thus, with the supervision of such signal labels, the proposed model also tend to retain such events arise from lamp as shown in left two columns of Fig. 9 despite their visually unappealing effect. + +Limitation. The DED framework may fail in extreme cases where similar part are dominated by noise events. However, such globally anomalous high-frequency noise is rare in practice. Moreover, the proposed model struggles to capture those insignificant signals attaching to structural periphery, limiting signal retain performance in metrics. + +# 6. Conclusion + +In this paper, we construct a new large-scale high-quality paired real event denoising dataset LED, which provides diverse noise levels and scenes which aims to cover various internal and external conditions. Based on the consistency difference of noise and signal events, the proposed DED framework could effectively generate high-quality GT and detailed analysis confirms our better results than others. In addition, we propose a novel baseline DTSNN for event denoising featuring with dynamic threshold mechanism of LIF neuron. The experimental results demonstrate the superiority of the proposed dataset and denoising method. + +Acknowledgments. The computation is completed in the HPC Platform of Huazhong University of Science and Technology. This work was supported by the National Natural Science Foundation of China under Grant 62371203. + +# References + +[1] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[2] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2020. 1, 2, 6, 7 +[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2 +[4] Clément Cabriel, Tual Monfort, Christian G Specht, and Ignacio Izeddin. Event-based vision sensor for fast and dense single-molecule localization microscopy. Nature Photonics, pages 1-9, 2023. 1 +[5] Haosheng Chen, David Suter, Qiangqiang Wu, and Hanzi Wang. End-to-end learning of object motion estimation from retinal events for event-based object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10534-10541, 2020. 1 +[6] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 303-311, 2022. 2 +[7] Jianchuan Ding, Bo Dong, Felix Heide, Yufei Ding, Yunduo Zhou, Baocai Yin, and Xin Yang. Biologically inspired dynamic thresholds for spiking neural networks. Advances in Neural Information Processing Systems, 35:6090-6103, 2022. 6 +[8] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-mlb: Multilevel benchmark for event-based camera denoising. IEEE Transactions on Multimedia, 2023. 2, 6, 7 +[9] Peiqi Duan, Zihao W Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos K Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 8261-8275, 2021. 2 +[10] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 2 +[11] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1 +[12] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular + +data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. 2, 6, 7 +[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6 +[14] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 6 +[15] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020. 2, 5, 7 +[16] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3867-3876, 2018. 2 +[17] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrio, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1 +[18] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 2021. 6, 7 +[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 6 +[20] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):785-795, 2022. 2, 5, 7 +[21] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 2 +[22] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 1 +[23] Tao Huang, Songjiang Li, Xu Jia, Hutchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 2 +[24] Alireza Khodamoradi and Ryan Kastner. $o(n)\mathrm{o}(n)$ -space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Transactions on Emerging Topics in Computing, 9(1):15-23, 2018. 7 + +[25] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2 +[26] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. Hots: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 5, 7 +[27] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2 +[28] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Trans. Graph., 38(6):164-1, 2019. 2 +[29] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15~\mu \mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1, 4 +[30] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In 2015 IEEE International Symposium on Circuits and Systems (ISCAS), pages 722-725. IEEE, 2015. 2 +[31] Ziwei Liu, Lu Yuan, Xiaou Tang, Matt Uytendaele, and Jian Sun. Fast burst images denoising. ACM Transactions on Graphics (TOG), 33(6):1-9, 2014. 2 +[32] Ali Maleky, Shayan Kousha, Michael S Brown, and Marcus A Brubaker. Noise2noiseflow: realistic camera noise modeling without clean images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17632-17641, 2022. 2 +[33] Rohit Mangalwedhekar, Nivedita Singh, Chetan Singh Thakur, Chandra Sekhar Seelamantula, Mini Jose, and Deepak Nair. Achieving nanoscale precision using neuromorphic localization microscopy. Nature Nanotechnology, pages 1-10, 2023. 1 +[34] Ben Mildenhall, Jonathan T Barron, Jiawen Chen, Dillon Sharlet, Ren Ng, and Robert Carroll. Burst denoising with kernel prediction networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2502-2510, 2018. 2 +[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012. 6 +[36] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 1 +[37] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639-16652, 2020. 1 + +[38] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982, 2018. 1 +[39] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 5 +[40] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12300-12308, 2019. 1 +[41] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7244-7253, 2019. 1 +[42] Cadence Design Systems. Software downloads-cadence design systems. https://www.cadence.com/en_US/home/support/software-downloads.html. 4 +[43] Siqi Wang, Tee Hiang Cheng, and Meng-Hiot Lim. Ltmd: Learning improvement of spiking neural networks with learnable thresholding neurons and moderate dropout. Advances in Neural Information Processing Systems, 35: 28350-28362, 2022. 6 +[44] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 2, 5, 7 +[45] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1609–1619, 2020. 2 +[46] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1 +[47] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Transactions on Multimedia, 23:1148-1159, 2020. 2 +[48] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 6 +[49] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 6 +[50] Pei Zhang, Zhou Ge, Li Song, and Edmund Y L. Neuromorphic imaging with density-based spatiotemporal denoising. IEEE Transactions on Computational Imaging, 2023. 2 + +[51] Xiang Zhang, Wei Liao, Lei Yu, Wen Yang, and Gui-Song Xia. Event-based synthetic aperture imaging with a hybrid network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14235-14244, 2021. 6 +[52] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3594-3604, 2022. 6 \ No newline at end of file diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/images.zip b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46399adf611fe95f77350428dadc3f7342b00750 --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54bf363009c221c87407f215954236fc9fd66dbb2d0c8562990d933d3946a0b7 +size 1097835 diff --git a/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/layout.json b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e678dae33d7be80c6942d674c61f1d4d764d565d --- /dev/null +++ b/2024/LED_ A Large-scale Real-world Paired Dataset for Event Camera Denoising/layout.json @@ -0,0 +1,8010 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 532, + 120 + ], + "type": "text", + "content": "LED: A Large-scale Real-world Paired Dataset for Event Camera Denoising" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 261, + 143, + 331, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 143, + 331, + 157 + ], + "spans": [ + { + "bbox": [ + 261, + 143, + 331, + 157 + ], + "type": "text", + "content": "Yuxing Duan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "type": "text", + "content": "National Key Lab of Multispectral Information Intelligent Processing Technology \n" + }, + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 157, + 496, + 186 + ], + "type": "text", + "content": "Huazhong University of Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 213, + 548, + 363 + ], + "blocks": [ + { + "bbox": [ + 48, + 213, + 548, + 363 + ], + "lines": [ + { + "bbox": [ + 48, + 213, + 548, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 548, + 363 + ], + "type": "image", + "image_path": "2f38453187383b2d327912d910f61ae52a5357f0214b2b71f7e96429409db0ca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 546, + 392 + ], + "type": "text", + "content": "Figure 1. Illustration of the paired raw/denoised event datasets. The proposed real-world event denoising datasets contribute to addressing the inevitable Background Activity (BA) noise encountered by event cameras application under varying camera settings and illumination." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 402, + 192, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 402, + 192, + 414 + ], + "spans": [ + { + "bbox": [ + 143, + 402, + 192, + 414 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 423, + 289, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 710 + ], + "type": "text", + "content": "Event camera has significant advantages in capturing dynamic scene information while being prone to noise interference, particularly in challenging conditions like low threshold and low illumination. However, most existing research focuses on gentle situations, hindering event camera applications in realistic complex scenarios. To tackle this limitation and advance the field, we construct a new paired real-world event denoising dataset (LED), including 3K sequences with 18K seconds of high-resolution (1200*680) event streams and showing three notable distinctions compared to others: diverse noise levels and scenes, larger-scale with high-resolution, and high-quality GT. Specifically, it contains stepped parameters and varying illumination with diverse scenarios. Moreover, based on the property of noise events inconsistency and signal events consistency, we propose a novel effective denoising framework(DED) using homogeneous dual events to generate the GT with better separating noise from the raw. Furthermore, we design a bio-inspired baseline leveraging Leaky-Integrate-and-Fire (LIF) neurons with dynamic thresholds to realize accurate denoising. The experimental results demonstrate that the remarkable performance of the proposed approach on different datasets. The dataset and code are at https://github.com/Yee-Sing/led." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 402, + 386, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 402, + 386, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 402, + 386, + 414 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 424, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 604 + ], + "type": "text", + "content": "Event cameras possess unique imaging advantages and are demonstrating tremendous potential for various applications [4, 11, 33, 37]. Unlike traditional frame cameras using integration sampling [46], event camera is differential sampling [29], evading the restriction of exposure period. However, due to this sampling mechanism, random fluctuations in analog signals can easily form noise events. In contrast to the dominance of signal components in images, the noise and signal in event cameras are encoded with same magnitude, easily disturbing inherent structural features, such as lane lines shown in Fig. 1. BA noise detrimentally impact on subsequent tasks like reconstruction [36, 39], motion estimation [40, 41] and event-based object detection [5]. Therefore, event denoising becomes a fundamental issue, notably in the booming event-based vision community." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "Paired data is a key issue for event denoising in the deep-learning era, thus an intuitive idea is to construct paired data through simulators [17, 22, 38]. However, the image-based simulation cannot be equivalent to reality because of the domain gap. To obtain the 'clean part' from the real-world event, multi-modal data was incorporated to help paired event generation. APS images gradient combined with IMU are used to indicate the event probability within frame period [2]. By jointly estimating the motion parameters with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25637" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 535, + 163 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 535, + 163 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 535, + 163 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 535, + 163 + ], + "type": "table", + "html": "
DatasetsSequencesCapture/sSourceNoise levelIlluminationResolutionPaired
RGB-DAVIS [45]20122CameraXDaytime180*190X
EventNFS [10]1004238Screen RecordingXDaytime224*124X
E-MLB [8]12007300Camera4Daytime to Nighttime346*260X
DVSNOISE20 [2]48807CameraXDaytime346*260✓
ED-KoGTL [1]415CameraXDaytime to Nighttime346*260✓
DVSCLEAN [12]14455Synthesis2Daytime1280*720✓
LED300018000Camera4Daytime to Nighttime1200*680✓
", + "image_path": "899b3a59222e279a3ed0201e4dbf8f48bed4516c4a15c5b67eb6cc0877473210.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 195, + 167, + 396, + 177 + ], + "lines": [ + { + "bbox": [ + 195, + 167, + 396, + 177 + ], + "spans": [ + { + "bbox": [ + 195, + 167, + 396, + 177 + ], + "type": "text", + "content": "Table 1. Summary of existing event denoising datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 187, + 286, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 286, + 319 + ], + "type": "text", + "content": "APS images and event streams [9, 45], the temporal mismatch can be alleviated. Constructing a fixed-track setup [1], the events under various illuminations can be referred to APS images in good lighting. Essentially, these methods align with events via transformed multi-mode features. Unfortunately, this heterogeneous aided information cannot fully match with the event, and their effectiveness is unavailable in situations such as low light and motion blur. Hence, few datasets have considered practical scenes where BA noise is easily deteriorated by camera setting and illumination, appearing as complex and nonuniform distributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 286, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 475 + ], + "type": "text", + "content": "To address this issue, we construct a large-scale, high-quality paired dataset LED for event denoising. LED has three distinct advantages. Firstly, it contains diverse noise levels, covering complex situations under various camera settings and illumination. Secondly, LED includes such as outdoor and indoor scenarios, with abundant objects. Last but not least, LED is collected with the current largest-scale and high-resolution. Moreover, inspired by the multisampling images denoising methods [21, 28, 31, 34], we are the first to explore event multi-sampling for denoising. Based on the homogeneous data, we proposed a dual-events denoising framework that can generate higher-quality GT by more accurately separating BA noise from the raw." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 475, + 276, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 276, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 276, + 487 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 488, + 286, + 656 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 46, + 488, + 286, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 286, + 534 + ], + "type": "text", + "content": "(1) We provide a large-scale real-world paired dataset for event denoising. So far, LED is the largest paired realworld event denoising dataset (3000 sequences) and high-resolution " + }, + { + "bbox": [ + 46, + 488, + 286, + 534 + ], + "type": "inline_equation", + "content": "(1200^{*}680)" + }, + { + "bbox": [ + 46, + 488, + 286, + 534 + ], + "type": "text", + "content": " with diverse noise levels and scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 536, + 286, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 286, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 286, + 595 + ], + "type": "text", + "content": "(2) We propose a novel dual events denoising framework DED based on noise inconsistency. We provide detailed analysis to show that DED can better separate mixed signals and noisy event streams across various camera threshold parameters, light conditions and motion patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 597, + 286, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 286, + 656 + ], + "type": "text", + "content": "(3) We introduce a novel baseline DTSNN for event denoising based on the spiking neural network, which utilizes the learnable dynamical spike threshold of the LIF neuron to accurately denoising. Extensive experiments on different real datasets verify the superiority of the proposed method." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 132, + 681 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Event Denoising Datasets. Currently, the datasets exclusively designed for event denoising with less attention com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 187, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 224 + ], + "type": "text", + "content": "pared to the task-specific ones in the event-based community. In Table 1, we provide an all-sided summary of existing datasets, highlighting the scarcity of real paired datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 227, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 545, + 407 + ], + "type": "text", + "content": "To address the lack of explicit labels, some paired datasets can be conveniently generated via simulators such as DVSCLEAN [12] and DND21 [20]. However, due to the gap between simulation and real, these simulators fail to fully reflect the actual situation. Other works focus on real paired event denoising data. DVSO noise20 [2] captures data in 16 pure background scenes. RGB-DAVIS [45] provides 20 indoor/outdoor sequences under good lighting conditions. ED-KoGTL [1] constructs indoor data with four illumination levels under constant trajectory. These paired datasets have achieved favorable results, primarily attributable to the gradient cues extracted from APS images. However, the heterogeneity between images and events limits the availability of diverse lighting conditions, motion patterns, and scenarios for constructing paired data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 578 + ], + "type": "text", + "content": "Event-based Denoising. Research on event denoising can be categorized into two approaches: filtering-based methods and deep learning methods. The former relies on manual priors to design discriminative models for noise removal, such as utilizing density distinction [15, 20, 30, 50] or motion continuity distinction [3, 44, 47] and motion compensation [6, 45] with CM framework [16] in the spatiotemporal domain. However, the validity of this prior knowledge is sensitive to different signal/noise event distributions, limiting their denoising accuracy. For learning-based methods, such as the CNN-based EDNCNN [2], the PointNet-based AEDNet [12] and [1] GNN-transformer were proposed successively. Additionally, the event denoising and super-resolution tasks are combined in [10]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Multi-Sampling Image Denoising. The multi-sampling principle for image denoising has been extensively studied. Techniques like long exposure time or burst captures benefit the direct acquisition of clear data with increased SNR in static scenes. Further, the multi-sampling was explored to the paradigm of indirectly recovering noise-free data. Not only using paired noisy observations to learn the mapping from different noisy samples to clear counterpart [27, 32], but also employing spatial sub-sampling on a single noisy image to learn the mapping from adjacent noisy samples to GT [23, 25]. Evidently, these denoising approaches imply" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25638" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 207, + 237 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 207, + 237 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 207, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 207, + 237 + ], + "type": "image", + "image_path": "ecd600611afddee70686da699d5581c10880777dd1ae9346566ed6cbc92e300b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 243, + 174, + 252 + ], + "lines": [ + { + "bbox": [ + 92, + 243, + 174, + 252 + ], + "spans": [ + { + "bbox": [ + 92, + 243, + 174, + 252 + ], + "type": "text", + "content": "(a) Distribution of LED" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 207, + 69, + 350, + 241 + ], + "blocks": [ + { + "bbox": [ + 207, + 69, + 350, + 241 + ], + "lines": [ + { + "bbox": [ + 207, + 69, + 350, + 241 + ], + "spans": [ + { + "bbox": [ + 207, + 69, + 350, + 241 + ], + "type": "image", + "image_path": "0f57e42e5e0c9ca78b0dfa172ee146d6e61fc1eeab685d04520a4e56411a1152.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 243, + 316, + 252 + ], + "lines": [ + { + "bbox": [ + 258, + 243, + 316, + 252 + ], + "spans": [ + { + "bbox": [ + 258, + 243, + 316, + 252 + ], + "type": "text", + "content": "(b) Scale of LED" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 350, + 70, + 545, + 242 + ], + "blocks": [ + { + "bbox": [ + 350, + 70, + 545, + 242 + ], + "lines": [ + { + "bbox": [ + 350, + 70, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 350, + 70, + 545, + 242 + ], + "type": "image", + "image_path": "9795416dca29c0fb8eaed60ec1247de0ce8c087ef8061fa99cb41d76455f62f2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 243, + 484, + 252 + ], + "lines": [ + { + "bbox": [ + 413, + 243, + 484, + 252 + ], + "spans": [ + { + "bbox": [ + 413, + 243, + 484, + 252 + ], + "type": "text", + "content": "(c) Diversity of LED" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 77, + 301, + 258, + 369 + ], + "blocks": [ + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 291 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed dataset LED. (a) Distribution of noise level and scene of the proposed dataset. (b) Our proposed LED outperforms others in terms of sequences, capture, and resolution (Circles with numbers to indicate). (c) LED collects diverse event streams across various conditions of illumination, depth of field, and target scale." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 301, + 258, + 369 + ], + "lines": [ + { + "bbox": [ + 77, + 301, + 258, + 369 + ], + "spans": [ + { + "bbox": [ + 77, + 301, + 258, + 369 + ], + "type": "image", + "image_path": "9746f753be59fadcb8d34d9534ac8b832e2d16c0639bd2bfda08dcfa4a15eeb1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 368, + 209, + 374 + ], + "lines": [ + { + "bbox": [ + 135, + 368, + 209, + 374 + ], + "spans": [ + { + "bbox": [ + 135, + 368, + 209, + 374 + ], + "type": "text", + "content": "(a) Noise in frame camera" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 77, + 374, + 257, + 448 + ], + "blocks": [ + { + "bbox": [ + 77, + 374, + 257, + 448 + ], + "lines": [ + { + "bbox": [ + 77, + 374, + 257, + 448 + ], + "spans": [ + { + "bbox": [ + 77, + 374, + 257, + 448 + ], + "type": "image", + "image_path": "d92ae6223f94ff3c6c879929ba5eb4422379e9d92e06ff64ebdaa047af2d15d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 443, + 209, + 451 + ], + "lines": [ + { + "bbox": [ + 135, + 443, + 209, + 451 + ], + "spans": [ + { + "bbox": [ + 135, + 443, + 209, + 451 + ], + "type": "text", + "content": "(b) Noise in event camera" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 455, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 521 + ], + "type": "text", + "content": "Figure 3. The substantial noise distinctions between frame/event camera. (a) For a stationary grayscale chart, the frame camera sequentially acquires noisy samples, fluctuating around a latent value within a specific distribution. (b) In a horizontal moving case, the event camera outputs binary signal serially, comprising signal events from motion gradient edges and spurious BA noise." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 533, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 557 + ], + "type": "text", + "content": "that signal resampling could facilitate suppressing random circuit noise and approximating the expectation value." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 567, + 257, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 257, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 257, + 582 + ], + "type": "text", + "content": "3. The High-quality Paired LED Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 587, + 214, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 214, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 214, + 600 + ], + "type": "text", + "content": "3.1. Dataset Statistics and Features" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "The primary factors affecting BA noise in event cameras can be categorized into intrinsic camera settings and extrinsic lighting conditions. By setting different camera thresholds, we can selectively record events with varying noise levels. Generally, a lower threshold means higher sensitivity but also increases the likelihood of BA noise. In contrast to fixed parameters in other datasets, we use four level thresholds to collect events, starting from the default value of " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": " and gradually decreasing to " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "-10\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "-20\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "inline_equation", + "content": "-30\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 375 + ], + "type": "text", + "content": "as shown in Fig. 2 (a). To cope with the challenges posed by complex lighting conditions, data collection is from daytime to nighttime, encompassing a transition from high to low illumination and a mixture of natural light with artificial light sources. Thus, the impact of external conditions on the event stream distribution can also be controlled." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "text", + "content": "Due to the difficulty of constructing real paired datasets, the development of existing datasets remains limited, as shown in Table 1. In this work, we used a vehicle-mounted or pan-tilt equipped with event cameras featuring a " + }, + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "inline_equation", + "content": "16\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 377, + 546, + 497 + ], + "type": "text", + "content": " lens. Totally, we collected approximately 5 hours of event stream, including 3K sequences. The composition of noise levels and scenes is depicted in Fig. 2 (a), with roughly equal amounts collected during daytime and nighttime. Nearly 20 typical scenes were captured, including urban buildings, traffic roads, natural scenery, indoor exhibits, etc." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "Our dataset goes beyond just the number of sequences and data volume, primarily aiming to consider the intricate influences of relevant factors on noise. Thus, we endow LED with diversity to reflect the impact of objects and environments on the event, which helps cover the cases of densely and non-uniformly distributed noisy situations in reality both spatially and temporally. In Fig. 2 (c), we exhibit that LED diversity is not only embodied in the richness of noise and illumination levels but also features such as depth of field and scene semantics. Meanwhile, high-resolution imaging aids the event cameras in capturing spatial information more precisely. Unfortunately, most existing ones with low resolution may hinder advanced visual tasks. Therefore, we collected data with " + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "inline_equation", + "content": "1200*680" + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": " resolution(after cropping), achieving high spatiotemporal resolution. Additionally, LED includes abundant objects of various scales, including aerial targets, pedestrians, and vehicles, making it well-suited for downstream tasks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25639" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 154, + 207 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 154, + 207 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 154, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 154, + 207 + ], + "type": "image", + "image_path": "b5a700a3a9fa6b12d3c367f38ee38459d3b4a2b01223f2709caf979cfc2cf9e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 209, + 138, + 217 + ], + "lines": [ + { + "bbox": [ + 73, + 209, + 138, + 217 + ], + "spans": [ + { + "bbox": [ + 73, + 209, + 138, + 217 + ], + "type": "text", + "content": "(a) Ideal intensity input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 70, + 299, + 138 + ], + "blocks": [ + { + "bbox": [ + 155, + 70, + 299, + 138 + ], + "lines": [ + { + "bbox": [ + 155, + 70, + 299, + 138 + ], + "spans": [ + { + "bbox": [ + 155, + 70, + 299, + 138 + ], + "type": "image", + "image_path": "d18ab4b6ec4e6f908c1b6777933e802b6ed92999ae2a20b6f17a0924a758b351.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 155, + 140, + 299, + 207 + ], + "blocks": [ + { + "bbox": [ + 155, + 140, + 299, + 207 + ], + "lines": [ + { + "bbox": [ + 155, + 140, + 299, + 207 + ], + "spans": [ + { + "bbox": [ + 155, + 140, + 299, + 207 + ], + "type": "image", + "image_path": "c3167d56f6087147b4ed8f56828d07fe4dc2a9ad090fe27bde62986a45def9c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 209, + 283, + 217 + ], + "lines": [ + { + "bbox": [ + 184, + 209, + 283, + 217 + ], + "spans": [ + { + "bbox": [ + 184, + 209, + 283, + 217 + ], + "type": "text", + "content": "(b) Circuit dual-sampled simulation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 436, + 138 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 436, + 138 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 436, + 138 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 436, + 138 + ], + "type": "image", + "image_path": "689fbe4668c25bda81b832c67afd44a047f370e79ed4a62f037fbe6857be358f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 302, + 139, + 435, + 207 + ], + "blocks": [ + { + "bbox": [ + 302, + 139, + 435, + 207 + ], + "lines": [ + { + "bbox": [ + 302, + 139, + 435, + 207 + ], + "spans": [ + { + "bbox": [ + 302, + 139, + 435, + 207 + ], + "type": "image", + "image_path": "2a87918737ebcc936b450c3c644bdbb624bb35b553664dffef76af5a5af0fd85.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 209, + 417, + 217 + ], + "lines": [ + { + "bbox": [ + 325, + 209, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 325, + 209, + 417, + 217 + ], + "type": "text", + "content": "(c) Realistic dual-Sampled events" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 71, + 546, + 139 + ], + "blocks": [ + { + "bbox": [ + 436, + 71, + 546, + 139 + ], + "lines": [ + { + "bbox": [ + 436, + 71, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 436, + 71, + 546, + 139 + ], + "type": "image", + "image_path": "01c4b7eafad35955d0781bcbd9bd208a546e36fabbf33a69db2038e5b9e0df74.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 140, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 436, + 140, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 436, + 140, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 436, + 140, + 545, + 207 + ], + "type": "image", + "image_path": "b90971034e6fbf5e8cc54a67a41115fe1a8e6f2e85e9ced63455b31b625b88d9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 209, + 544, + 217 + ], + "lines": [ + { + "bbox": [ + 446, + 209, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 446, + 209, + 544, + 217 + ], + "type": "text", + "content": "(d) Statistic of dual-sampled events" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 48, + 300, + 303, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 221, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 221, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 546, + 288 + ], + "type": "text", + "content": "Figure 4. Illustration of event camera dual-sampling analysis: (a) Given an ideal intensity input, which includes varying and steady stages, generates a series of signal events. (b) The twice samplings of the same input in the circuit model, both generate additional BA noise in previously steady stages, resulting in consistency discrepancies between noise events and signal events. (c) The visualization results of the actual dual-sampled events cumulative frame demonstrate the misalignment of inconsistent dual-sampled noise events and the alignment of coexisting signal events (two color groups indicating the respective polarities of the dual events). (d) The statistical results of the two tests also prove a low overlap rate between dual-sampled event pixels and a much higher one where signal events are present." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 300, + 303, + 405 + ], + "lines": [ + { + "bbox": [ + 48, + 300, + 303, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 303, + 405 + ], + "type": "image", + "image_path": "4a57266de4d34d4a8077f3e2c06ee67c38a66a4dc12a67ee5cba87f8244cebbe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 409, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 475 + ], + "type": "text", + "content": "Figure 5. Overview of the dual events denoising framework. (a) Our collection device consists of two identical EVK4s forming a co-axial system with a 1:1 beamsplitter. (b) We first perform spatial similarity processing to retain the consistent parts, followed by sequentially spatiotemporal correlation constraints to remove the residual small amount of noise event from the previous step." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 491, + 234, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 234, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 234, + 504 + ], + "type": "text", + "content": "3.2. Dual Events Denoising Framework" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 545 + ], + "type": "text", + "content": "Given a noisy event stream, the key lies in properly obtaining paired GT. In this section, we explore utilizing multi-sampling to achieve event stream denoising." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": "Background. Multi-sampling denoising methods are commonly used in image denoising. For frame images, averaging the results of temporal multi-samples can approximate the true value for each pixel owing to multi-samples smoothing effect on random noise components, as shown in Fig. 3 (a). For event stream, due to their dynamic sampling and the binary values, temporal multi-samples may contain both signal events and BA noise, as illustrated in Fig. 3 (b), directly taking the average of multi-samples cannot yield true signal value. Although the noise forms differ between the two modalities, they ultimately originate from the inherent noise in the circuitry analog signal, while resampling fundamentally helps suppress the random noise source." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Analysis. This concept inspired us to utilize multi-sampling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 303, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 546, + 460 + ], + "type": "text", + "content": "for event denoising, particularly when considering the binary state of events (presence or absence) where just two samplings are required to potentially identify noise. Thus, we initially investigated the noise inconsistency at the pixel level. Based on the DVS working principle [29], we build a circuit model in the Cadence [42] platform to simulate the dual-sampling with twice identical intensity input. The noisy analog signal converted from input generates BA noise at the originally steady stage compared to the ideal case in Fig. 4 (a). Clearly, the two sets of noise exhibit inconsistency, while the signals demonstrate high consistency, causing distinct alignment differences between the two cases as shown in the close-up of Fig. 4 (b)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": "To further explore the inter-class differences between noise and signal events in real-world dual-sampled events stream, we established a collection device as shown in Fig. 5 (a), to construct genuinely spatiotemporal dual-sampling. Two experimental conditions to validate the aforesaid circuit simulation phenomenon are a stationary camera and a moving one respectively. In the former, pure BA noise events was obtained, while the latter captured mixed events containing both noise and signal. In static tests, although at the highest noise level " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "(-30\\%)" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": " threshold), both two sets of events exhibit significant spatiotemporal disparity. As observed in the right zoomed-in patches at the upper row in Fig. 4 (c), the global spatial misalignment of dual BA noises events was evident. The overlap rate of the pixels triggering event is also to be less than " + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 462, + 546, + 713 + ], + "type": "text", + "content": ", as shown in the upperrow histogram of Fig. 4 (d). Not unexpectedly, in dynamic tests, regions with BA noise events remain inconsistency, while the coexistent signal events showing spatiotemporal alignment as depicted in lower-row counterpart, leading to a dramatic increase in overlap rate compared to the pure BA noise cases, as shown in the bottom histogram of Fig. 4 (d)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25640" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 252 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 252 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 252 + ], + "type": "image", + "image_path": "635af38e31bb62e5022ecac3e503890e3dd5ff13d4d6d545aa107af50c22528f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 300 + ], + "type": "text", + "content": "Figure 6. Analysis of different event denoising results on our dataset. From left to right, the first column is the raw events, and the remaining five columns represent different methods, namely DWF, STDF, TimeSurface, EvFlow, and the proposed DED. From top to bottom, the first row shows the denoised results, the second row is the residual noise, the third row denotes the statistical distribution of the denoised results and the last row represents the intensity image reconstruction corresponding to the zoomed region denoised events." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": "**Formulation.** These results support the insight that the consistency discrepancies between dual-sampled events could naturally help distinguish noise. Therefore, we develop a denoising framework with the dual events stream, called DED. In DED framework, given the spatiotemporal synchronized dual event streams " + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "inline_equation", + "content": "Y_{1}" + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "inline_equation", + "content": "Y_{2}" + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": ", they are composed of common latent signal " + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": " and respective noise streams " + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 46, + 308, + 289, + 403 + ], + "type": "text", + "content": ". The noise model can be defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 406, + 287, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 406, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 99, + 406, + 287, + 418 + ], + "type": "interline_equation", + "content": "Y _ {1} = X + N _ {1}, Y _ {2} = X + N _ {2}. \\tag {1}", + "image_path": "127e86659d00ed9a68110ebf30497d36777b1883026a64301e710330714ea498.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "text", + "content": "Because of the consistency of signal events and the inconsistency of noise events in " + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "inline_equation", + "content": "Y_{1}" + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "inline_equation", + "content": "Y_{2}" + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "text", + "content": ", after binary grid operation of event stream according to certain temporal window " + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 47, + 420, + 287, + 468 + ], + "type": "text", + "content": ", we can perform spatial similarity to process them:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 469, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 469, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 94, + 469, + 287, + 483 + ], + "type": "interline_equation", + "content": "X ^ {*} = \\left\\| \\left\\| X + N _ {1} \\right\\| \\circ \\left\\| X + N _ {2} \\right\\| \\right\\| _ {1}, \\tag {2}", + "image_path": "8f963c9fc7cea88a6adb149e2a106636d67b4124887c86cf35f32c2328648361.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": " denotes the Hadamard product used to obtain the similar part " + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "inline_equation", + "content": "X^{*}" + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": " of two groups of raw event binary frames. Due to the probability of a few noises occurring simultaneously in dual-sampling, " + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "inline_equation", + "content": "X^{*}" + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": " may still contain some much-isolated noise compared to the raw. As shown in Fig. 5 (b), for finer denoising, we further utilize the spatiotemporal correlation of the signal event stream to remove the residual noise from the previous step. Specifically, we accumulate the events from the relevant spatiotemporal range in corresponding " + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "inline_equation", + "content": "X^{*}" + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": " within several consecutive " + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 484, + 289, + 675 + ], + "type": "text", + "content": " to form a spatiotemporal relationship set. The presence of these few isolated noise events, which are randomly triggered in the spatiotemporal domain, disrupts the inherent spatiotemporal correlation of signal reflecting a certain motion model. Therefore, to leverage this spatiotemporal continuity, we transform it into the following minimization problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 677, + 287, + 709 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 677, + 287, + 709 + ], + "spans": [ + { + "bbox": [ + 78, + 677, + 287, + 709 + ], + "type": "interline_equation", + "content": "\\min \\frac {1}{N - 1} \\sum_ {i = 1} ^ {N} \\left| e _ {\\boldsymbol {x}} ^ {t _ {i} + 1} - e _ {\\boldsymbol {x}} ^ {t _ {i}} \\right|, _ {\\boldsymbol {x} \\in \\Omega , \\Omega \\in X ^ {*}}, \\tag {3}", + "image_path": "9ddaa9cfd02048ca062508333c4176f5d5e53edaf9fc0f885fe2201dfdbe1e1f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "e_{\\pmb{x}}^{t_i}" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " occurring event according to the sorted timestamp, " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " means the event spatial coordinates, " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " is the number of selected consecutive time windows, " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " represents the candidate spatial neighborhood, and " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": " is total events number in " + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 305, + 308, + 546, + 381 + ], + "type": "text", + "content": ". Additionally, the aforesaid processing is all performed on the two channels of event polarity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 392, + 452, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 452, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 452, + 404 + ], + "type": "text", + "content": "3.3. Evaluation and Discussion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 412, + 546, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 546, + 472 + ], + "type": "text", + "content": "The quality of GT is critical for a real paired event dataset. However, it is challenging to objectively evaluate the denoising results since directly obtaining clean data is unpractical. To fairly assess the denoising effects of different methods, we conducted a comprehensive evaluation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "type": "text", + "content": "Fig. 6 presents comparative results of representative event denoising methods could be used for GT generation: spatiotemporal density filter-based (DWF[20], STDF[15]), smoothness optimization-based (TimeSurface[26]), and motion estimation-based (EvFlow[44]). The first and second rows display the denoised and residual noise visualization, respectively. It can be observed that DED effectively eliminates almost all the scattered BA noise with good preservation of the inherent structured feature, while the others exhibit residual noise and varying degrees of signal event damage. Statistical analysis on the global distribution of the denoised event is shown in the third row. Compared to others, DED demonstrates a higher number of blank patches while maintaining a high event preservation rate, which better reflects the overall sparsity and local concentration of the ideal event stream. Moreover, better denoising often leads to higher-quality reconstruction. In the last row, the intensity images reconstructed by E2VID [39] indicate that DED accurately restores nighttime scene information, such as leaves, building contours, and road signs" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25641" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 288, + 115 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 288, + 115 + ], + "type": "image", + "image_path": "7cad2b81f803c427243459a40ec7582b05003a9a556a9c984cc57f638cdf0989.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 116, + 240, + 124 + ], + "lines": [ + { + "bbox": [ + 109, + 116, + 240, + 124 + ], + "spans": [ + { + "bbox": [ + 109, + 116, + 240, + 124 + ], + "type": "text", + "content": "(a) Spike responses with different thresholds" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 125, + 287, + 232 + ], + "blocks": [ + { + "bbox": [ + 49, + 125, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 49, + 125, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 49, + 125, + 287, + 232 + ], + "type": "image", + "image_path": "dfdc55ad0cd37f28cf8a01602ab2a68dfd9ac575d8acba6a9860dad7514cc7a9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 232, + 219, + 241 + ], + "lines": [ + { + "bbox": [ + 127, + 232, + 219, + 241 + ], + "spans": [ + { + "bbox": [ + 127, + 232, + 219, + 241 + ], + "type": "text", + "content": "(b) Framework of the DTSNN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 291 + ], + "type": "text", + "content": "Figure 7. An overview of the DTSNN model. (a) The dynamic threshold mechanism of LIF neuron. (b) The model with two branches consisting of DTB and EDB. Please refer to our supplementary material for more details about network architecture." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 346 + ], + "type": "text", + "content": "with lower NIQE [35], which further indirectly validates the accuracy of denoised results and also confirms the superior effect of the processed data on reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 357, + 208, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 208, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 208, + 371 + ], + "type": "text", + "content": "4. DTSNN for Event Denoising" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "content": "Compared to artificial neural networks(ANN), spiking neural network (SNN) have lower precision but are inherently suited for processing event-driven data due to their and information transfer mode temporal dynamics, gradually gaining prominence in event-based tasks [48, 49, 51, 52]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": "Dynamic Threshold Mechanism. Adjustable spiking neuron offers enhanced biological plausibility [7, 13, 43]. Indeed, the neuron firing threshold, similar to the event camera threshold parameter, functions as controlling the output. Intuitively, the dynamic threshold (DT) mechanism can adapt to more complicated situations since a high threshold could suppress noisy input when noise dominates, while a low one helps to sensitively preserve the expected signals. Inspired by this, we transformed the fixed threshold into dynamic ones based on the LIF neuron model [19] to mimic this anisotropic biomechanism. The tendency of decreasing spike quantity of the postsynaptic neurons responses with an increasing threshold as illustrated in Fig. 7(a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "DTSNN Model. Therefore, we propose a fully SNN (all synaptic operations are SNN-based) incorporating learnable DT for event denoising, namely DTSNN. As illustrated in Fig. 7 (b), DTSNN consists of a dynamic threshold branch (DTB) and event denoising branch (EDB). The former dynamically generates a threshold map based on the successive event input, indicating the approximate spatial regions of signal or noise. This map is then passed to the latter, controlling the spiking process of LIF neurons in the last layer," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 534, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 534, + 84 + ], + "type": "text", + "content": "which is depicted in Fig. 7 (b) and can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 88, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 88, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 337, + 88, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} H ^ {n, t} = V ^ {n, t - 1} + U ^ {n, t} \\\\ S ^ {n, t} = H e a \\left(H ^ {n, t} - \\mathcal {F} \\left(V _ {\\text {t h}} ^ {t}\\right)\\right) \\\\ V ^ {n, t} = V _ {\\text {r e s e t}} S ^ {n, t} + \\frac {1}{\\tau} H ^ {n, t} \\odot (1 - S ^ {n, t}) \\end{array} \\right. \\tag {4}", + "image_path": "df23429dd0319fb1e3d064747e2ae2bf254f90b1aea939f277beaecc0afa7039.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " denote the layer number and time step respectively. " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "H^{n,t}" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " is the membrane potential which is produced by coupling the spatial feature " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "U^{n,t}" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " and the temporal input " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "V^{n,t-1}" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": ". The DT map " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(V_{\\mathrm{th}}^t)" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " determines whether the output spiking matrix " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "S^{n,t}" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " should be fired or stay as zero, formulating the final denoised event matrix " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "Hea(x) = 1" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " represents the Heaviside step function when " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "x \\geq 0" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": ", otherwise " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "Hea(x) = 0" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 147, + 545, + 243 + ], + "type": "text", + "content": " means a element-wise multiplication." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": "Notably, both two branches receive the same consecutive data as input, which is a binary event frame within a certain time window of " + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": ". The denoising branch is supervised by the GT of signal events within " + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 243, + 545, + 315 + ], + "type": "text", + "content": ", while the threshold map label of the dynamic threshold branch comes from the signal events over a longer temporal period." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 329, + 387, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 387, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 387, + 342 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 348, + 418, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 418, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 418, + 361 + ], + "type": "text", + "content": "5.1. Experiments Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "text", + "content": "Implementation Details. We use a combined loss consisting of the L1-norm and BCE to train the proposed network. Our models are implemented with the open-source framework SpikingJelly [14], using NVIDIA A100 GPU. The Adam optimizer is employed with a batch size of 8 and a learning rate of 0.002. A fixed threshold of 0.5 is chosen for the rest of the neurons except at the last membrane potential layer. Besides, the reset value " + }, + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{reset}}" + }, + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "text", + "content": " and the membrane time constant " + }, + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 366, + 545, + 617 + ], + "type": "text", + "content": " of all LIF neurons are set to 0 and 2, respectively. Datasets. We conduct quantitative experiments on existing paired datasets. Due to the enormous scale of billions of events in LED, training them all would be extremely time-consuming and resource-intensive. Therefore, 600 sequences were randomly selected for training with each consisting of consecutive 10 segments, and 60 sequences were randomly chosen for testing. To further evaluate the performance of the discrepancy of different datasets, we qualitatively test on typical public datasets DSEC [18] and E-MLB [8]. For event-based denoising methods, besides the methods in Sec.3.3 we select the representative supervised denoising methods including EDnCNN [2] and AEDNet[12]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Methods. We use an index of event denoising accuracy " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "DA = \\frac{1}{2}\\left(\\frac{TP}{GP} +\\frac{TN}{GN}\\right)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " to measure the denoising performance on LED, where TP, TN, GP and GN are the number of true signal, true noise, total signal, and total noise respectively, forming two parts: signal retain (SR) and noise removal (NR). The metrics on other paired datasets are adopted from their proposed ones. Moreover, visualization results was qualitatively evaluated on other datasets." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25642" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 260 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 260 + ], + "type": "image", + "image_path": "74efc3c1ddc5247ce45110fb35169883c356218366efd09583afe6382e0e1cf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 546, + 288 + ], + "type": "text", + "content": "Figure 8. Visual comparisons on LED. Comparing with state-of-the-arts, DTSNN achieves excellent denoising results and it is capable of removing noise events scattered around in sky, ground, and structural contours, simultaneously better preserve scene details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 301, + 287, + 430 + ], + "blocks": [ + { + "bbox": [ + 47, + 301, + 287, + 430 + ], + "lines": [ + { + "bbox": [ + 47, + 301, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 287, + 430 + ], + "type": "table", + "html": "
DatasetsLEDDVS CLEANDVS NOISE20Average
MetricsSR↑NR↑DA↑SNR↑RPMD↓Runtime↓
Knoise[24]16.498.957.625.2117.641.33 ms
DWF[20]42.683.963.326.9631.393.67 ms
STDF[15]37.499.168.319.3030.232.04 ms
TS[26]30.698.364.513.9815.255.76 ms
EvFlow[44]52.396.474.423.7422.50150.5 ms
EDnCNN[2]80.082.081.020.2922.25251.4 ms
AEDNet[12]81.283.682.425.5818.51708.1 ms
DTSNN86.086.586.229.2616.138.20 ms
", + "image_path": "3919f0adfb27e5dfae5a0370515998b332852c159eda044bcfe760902307a6f7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 433, + 287, + 455 + ], + "lines": [ + { + "bbox": [ + 47, + 433, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 433, + 287, + 455 + ], + "type": "text", + "content": "Table 2. Quantitative results comparison on different datasets. We mark the best and second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 466, + 183, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 466, + 183, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 466, + 183, + 479 + ], + "type": "text", + "content": "5.2. Quantitative Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "text", + "content": "The main quantitative results are presented in Table 2. In summary, the proposed method achieves the best overall results on three datasets, followed by AEDNet. Particularly, DTSNN balances signal retention and noise removal effectively, demonstrating the highest denoising accuracy. Meanwhile, our model also achieved faster inference speed on average among them. Notably, the runtime from " + }, + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "inline_equation", + "content": "120 * 120" + }, + { + "bbox": [ + 46, + 484, + 287, + 616 + ], + "type": "text", + "content": " events indicates that the efficiency of grid-based event representation and processing is significantly higher than the manner based on a single-event level, although the latter better preserves the asynchronous property of the event." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 176, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 176, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 176, + 635 + ], + "type": "text", + "content": "5.3. Qualitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "Evaluation on LED Dataset. To further evaluate the denoising performance, we compare with the qualitative results of typical methods on LED. As shown in Fig. 8, DTSNN outperforms other methods by achieving visually pleasing results, which not only effectively removes noise events in areas such as the sky, ground and building con" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 301, + 545, + 468 + ], + "blocks": [ + { + "bbox": [ + 306, + 301, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 306, + 301, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 545, + 468 + ], + "type": "image", + "image_path": "6523c29e1f781551da43c365219b16e27a0c4233fefb2e60845b2f37554ce107.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 472, + 545, + 515 + ], + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 515 + ], + "type": "text", + "content": "Figure 9. Illustration of the LED diversity. We train DTSNN on different datasets: DVSCLEAN, DVSNOISE20 and LED, and test on real unpaired datasets (from left to right: DSEC " + }, + { + "bbox": [ + 305, + 472, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 305, + 472, + 545, + 515 + ], + "type": "text", + "content": " E-MLB). The model trained on LED has achieved better denoising results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 531, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 602 + ], + "type": "text", + "content": "tours, but also preserve the structural features of the scene and objects well. For example, the zoomed region of first row in Fig. 8 highlights that our method is the only one successfully preserving the information of small airborne target after denoising. Similarly, our method also stands out in retain lane marking located in the bottom-left of second row." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on Public Datasets. To evaluate the generalization across different datasets, we also train DTSNN on synthetic event denoising dataset DVSCLEAN [12] and real one DVSNoise20 [2] respectively, testing on DSEC [18] and E-MLB [8], the former is resolution of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "640^{*}480" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and the latter is resolution of " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "346^{*}260" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". As shown in Fig. 9, the model trained on DVSCLEAN performs poorly due to the huge domain gap between. DVSNoise20 excels at removing noise effectively but inevitably loses some details. The" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25643" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 137 + ], + "type": "image", + "image_path": "b95fafc3640319eeedd7ac5e26b22757d12ea4f390cb3f3f50f3111cf878e018.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 142, + 286, + 186 + ], + "lines": [ + { + "bbox": [ + 47, + 142, + 286, + 186 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 286, + 186 + ], + "type": "text", + "content": "Figure 10. Ablation study of the dual events denoising framework. The first column represents the noisy raw and the remaining three columns represent the flow without spatial similarity, without spatiotemporal correlation constraint and complete DED" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 192, + 269, + 232 + ], + "blocks": [ + { + "bbox": [ + 63, + 192, + 269, + 232 + ], + "lines": [ + { + "bbox": [ + 63, + 192, + 269, + 232 + ], + "spans": [ + { + "bbox": [ + 63, + 192, + 269, + 232 + ], + "type": "table", + "html": "
ArchitectureDenoising AccuracyEnergy Ratio
ANN84.518×
SNN86.01×
", + "image_path": "69a43a6f414ea39d3f451fd4fc027324718c8c95f71edc4957bb02a712bf8e6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 255, + 286, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 286, + 302 + ], + "type": "text", + "content": "model trained on LED achieves satisfactory results across various scenes in these datasets which simultaneously removes noise without apparent signal loss, strongly supporting the great generalization of the proposed LED dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 312, + 212, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 212, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 212, + 324 + ], + "type": "text", + "content": "5.4. Ablation Study and Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 331, + 286, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 286, + 462 + ], + "type": "text", + "content": "Effectiveness of DED Framework. The DED framework aims to fully exploit the property of noise inconsistency /signal consistency in dual events and the self-correlation in a single event stream. As shown in Fig. 10, the small target signal is overwhelmed by noise events, utilizing solely spatiotemporal correlation constraints can only remove partial noise events because some of them also satisfy this relationship. Relying on the spatial similarity of dual event streams effectively disassembles the mutual relationship in the dense noise event group. Therefore, combining both procedures can achieve a more refined denoising result." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 463, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 286, + 581 + ], + "type": "text", + "content": "Effectiveness of Spiking Neuron. Intuitively, the inherent spatiotemporal sparsity of event streams seamlessly aligns with SNN because of the temporal information ability of spiking neurons. To compare the performance differences between SNN and ANN architectures in the event denoising tasks, the evaluation was conducted on the same structure. As shown in Table 3, the SNN based on LIF neuron achieves better denoising accuracy while maintaining " + }, + { + "bbox": [ + 46, + 463, + 286, + 581 + ], + "type": "inline_equation", + "content": "18 \\times" + }, + { + "bbox": [ + 46, + 463, + 286, + 581 + ], + "type": "text", + "content": " lower power consumption advantage compared to its ANN version (the spiking neurons replaced with ReLU)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 712 + ], + "type": "text", + "content": "Effectiveness of DT Module. To evaluate the effects of the proposed learnable threshold mechanism of DTSNN, we train the network with various fixed threshold neurons and dynamic threshold neurons, respectively. As can be observed from Table 4 and Fig. 11 (a), the denoising accuracy improves after introducing the DT module due to the network has the tendency to increasing spiking probability for signal pixels according to threshold prediction, and the Fig. 11 (b) illustrates DT module could assist in locating the approximate signal or noise region to retain more signal events from structural area compared to the FT model." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 462, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 462, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 462, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 462, + 188 + ], + "type": "image", + "image_path": "9c01e4afe428d8a84921e8c4b8360c27458ac2ad6065abbaca991f12d07ed2d8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 189, + 426, + 198 + ], + "lines": [ + { + "bbox": [ + 352, + 189, + 426, + 198 + ], + "spans": [ + { + "bbox": [ + 352, + 189, + 426, + 198 + ], + "type": "text", + "content": "(a) Inference process" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 188, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 466, + 188, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 466, + 188, + 536, + 198 + ], + "type": "text", + "content": "(b) Denoised results" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 465, + 71, + 543, + 183 + ], + "blocks": [ + { + "bbox": [ + 465, + 71, + 543, + 183 + ], + "lines": [ + { + "bbox": [ + 465, + 71, + 543, + 183 + ], + "spans": [ + { + "bbox": [ + 465, + 71, + 543, + 183 + ], + "type": "image", + "image_path": "4e76dfc37e764591ebd63fb682e8268a4e361fa4ffa58923df0002092348c432.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 202, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 306, + 202, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 545, + 236 + ], + "type": "text", + "content": "Figure 11. Illustration of the DT model Effectiveness. (a) Inference accuracy of SNN with various fixed thresholds and with DT. (b) Denoised results without DT and with DT, respectively." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 312, + 243, + 538, + 286 + ], + "blocks": [ + { + "bbox": [ + 82, + 234, + 251, + 244 + ], + "lines": [ + { + "bbox": [ + 82, + 234, + 251, + 244 + ], + "spans": [ + { + "bbox": [ + 82, + 234, + 251, + 244 + ], + "type": "text", + "content": "Table 3. Comparison between ANN and SNN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 243, + 538, + 286 + ], + "lines": [ + { + "bbox": [ + 312, + 243, + 538, + 286 + ], + "spans": [ + { + "bbox": [ + 312, + 243, + 538, + 286 + ], + "type": "table", + "html": "
TypeFTSNN (0.3 th)FTSNN (0.5 th)FTSNN (0.8 th)DTSNN
Accuracy85.586.085.786.2
", + "image_path": "82915b09feacbad9eac7988ef42c25e246c322dbb7a0b5449666790f529ee07e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 288, + 544, + 299 + ], + "lines": [ + { + "bbox": [ + 306, + 288, + 544, + 299 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 544, + 299 + ], + "type": "text", + "content": "Table 4. Inference accuracy with proposed methods implemented." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 312, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 430 + ], + "type": "text", + "content": "Discussion. Essentially, true events comes from intensity changes. When artificial light source variations exist, the light source itself, its coverage area and reflection from ground texture would result in relevant events. These signal events exhibit noise-like features with discontinuous structure, while the proposed DED framework considers them as signal event that should be preserved. Thus, with the supervision of such signal labels, the proposed model also tend to retain such events arise from lamp as shown in left two columns of Fig. 9 despite their visually unappealing effect." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 431, + 545, + 503 + ], + "type": "text", + "content": "Limitation. The DED framework may fail in extreme cases where similar part are dominated by noise events. However, such globally anomalous high-frequency noise is rare in practice. Moreover, the proposed model struggles to capture those insignificant signals attaching to structural periphery, limiting signal retain performance in metrics." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 378, + 525 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 665 + ], + "type": "text", + "content": "In this paper, we construct a new large-scale high-quality paired real event denoising dataset LED, which provides diverse noise levels and scenes which aims to cover various internal and external conditions. Based on the consistency difference of noise and signal events, the proposed DED framework could effectively generate high-quality GT and detailed analysis confirms our better results than others. In addition, we propose a novel baseline DTSNN for event denoising featuring with dynamic threshold mechanism of LIF neuron. The experimental results demonstrate the superiority of the proposed dataset and denoising method." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. The computation is completed in the HPC Platform of Huazhong University of Science and Technology. This work was supported by the National Natural Science Foundation of China under Grant 62371203." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25644" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Yusra Alkendi, Rana Azzam, Abdulla Ayyad, Sajid Javed, Lakmal Seneviratne, and Yahya Zweiri. Neuromorphic camera denoising using graph neural network-driven transformers. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] R Baldwin, Mohammed Almatrafi, Vijayan Asari, and Keigo Hirakawa. Event probability mask (epm) and event denoising convolutional neural network (edncnn) for neuromorphic cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1701-1710, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 258 + ], + "type": "text", + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE transactions on neural networks and learning systems, 25 (2):407-417, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 304 + ], + "type": "text", + "content": "[4] Clément Cabriel, Tual Monfort, Christian G Specht, and Ignacio Izeddin. Event-based vision sensor for fast and dense single-molecule localization microscopy. Nature Photonics, pages 1-9, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 287, + 360 + ], + "type": "text", + "content": "[5] Haosheng Chen, David Suter, Qiangqiang Wu, and Hanzi Wang. End-to-end learning of object motion estimation from retinal events for event-based object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10534-10541, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 287, + 417 + ], + "type": "text", + "content": "[6] Jinze Chen, Yang Wang, Yang Cao, Feng Wu, and Zheng-Jun Zha. Progressivemotionseg: Mutually reinforced framework for event-based motion segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 303-311, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 419, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 473 + ], + "type": "text", + "content": "[7] Jianchuan Ding, Bo Dong, Felix Heide, Yufei Ding, Yunduo Zhou, Baocai Yin, and Xin Yang. Biologically inspired dynamic thresholds for spiking neural networks. Advances in Neural Information Processing Systems, 35:6090-6103, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "type": "text", + "content": "[8] Saizhe Ding, Jinze Chen, Yang Wang, Yu Kang, Weiguo Song, Jie Cheng, and Yang Cao. E-mlb: Multilevel benchmark for event-based camera denoising. IEEE Transactions on Multimedia, 2023. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 287, + 586 + ], + "type": "text", + "content": "[9] Peiqi Duan, Zihao W Wang, Boxin Shi, Oliver Cossairt, Tiejun Huang, and Aggelos K Katsaggelos. Guided event filtering: Synergy between intensity images and neuromorphic events for high performance imaging. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11): 8261-8275, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 643 + ], + "type": "text", + "content": "[10] Peiqi Duan, Zihao W Wang, Xinyu Zhou, Yi Ma, and Boxin Shi. Eventzoom: Learning to denoise and super resolve neuromorphic events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12824-12833, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[11] Davide Falanga, Kevin Kleber, and Davide Scaramuzza. Dynamic obstacle avoidance for quadrotors with event cameras. Science Robotics, 5(40):eaaz9712, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Huachen Fang, Jinjian Wu, Leida Li, Junhui Hou, Weisheng Dong, and Guangming Shi. Aednet: Asynchronous event denoising with spatial-temporal correlation among irregular" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "data. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1427-1435, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 97, + 547, + 152 + ], + "type": "text", + "content": "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 207 + ], + "type": "text", + "content": "[14] Wei Fang, Yanqi Chen, Jianhao Ding, Zhaofei Yu, Timothee Masquelier, Ding Chen, Liwei Huang, Huihui Zhou, Guoqi Li, and Yonghong Tian. Spikingjelly: An open-source machine learning infrastructure platform for spike-based intelligence. Science Advances, 9(40):eadi1480, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[15] Yang Feng, Hengyi Lv, Hailong Liu, Yisa Zhang, Yuyao Xiao, and Chengshan Han. Event density based denoising method for dynamic vision sensor. Applied Sciences, 10(6): 2024, 2020. 2, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 319 + ], + "type": "text", + "content": "[16] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3867-3876, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 376 + ], + "type": "text", + "content": "[17] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrio, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 419 + ], + "type": "text", + "content": "[18] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 2021. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[19] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski. Neuronal dynamics: From single neurons to networks and models of cognition. Cambridge University Press, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 510 + ], + "type": "text", + "content": "[20] Shasha Guo and Tobi Delbruck. Low cost and latency event camera background activity denoising. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):785-795, 2022. 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[21] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 612 + ], + "type": "text", + "content": "[22] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 669 + ], + "type": "text", + "content": "[23] Tao Huang, Songjiang Li, Xu Jia, Hutchuan Lu, and Jianzhuang Liu. Neighbor2neighbor: Self-supervised denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14781-14790, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "[24] Alireza Khodamoradi and Ryan Kastner. " + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "inline_equation", + "content": "o(n)\\mathrm{o}(n)" + }, + { + "bbox": [ + 308, + 670, + 545, + 714 + ], + "type": "text", + "content": "-space spatiotemporal filter for reducing noise in neuromorphic vision sensors. IEEE Transactions on Emerging Topics in Computing, 9(1):15-23, 2018. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25645" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[25] Alexander Krull, Tim-Oliver Buchholz, and Florian Jug. Noise2void-learning denoising from single noisy images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2129-2137, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 288, + 174 + ], + "type": "text", + "content": "[26] Xavier Lagorce, Garrick Orchard, Francesco Galluppi, Bertram E. Shi, and Ryad B. Benosman. Hots: A hierarchy of event-based time-surfaces for pattern recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(7):1346-1359, 2017. 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 220 + ], + "type": "text", + "content": "[27] Jaakko Lehtinen, Jacob Munkberg, Jon Hasselgren, Samuli Laine, Tero Karras, Miika Aittala, and Timo Aila. Noise2noise: Learning image restoration without clean data. arXiv preprint arXiv:1803.04189, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 287, + 266 + ], + "type": "text", + "content": "[28] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Trans. Graph., 38(6):164-1, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": "[29] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "inline_equation", + "content": "15~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 369 + ], + "type": "text", + "content": "[30] Hongjie Liu, Christian Brandli, Chenghan Li, Shih-Chii Liu, and Tobi Delbruck. Design of a spatiotemporal correlation filter for event-based sensors. In 2015 IEEE International Symposium on Circuits and Systems (ISCAS), pages 722-725. IEEE, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 404 + ], + "type": "text", + "content": "[31] Ziwei Liu, Lu Yuan, Xiaou Tang, Matt Uytendaele, and Jian Sun. Fast burst images denoising. ACM Transactions on Graphics (TOG), 33(6):1-9, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 406, + 287, + 460 + ], + "type": "text", + "content": "[32] Ali Maleky, Shayan Kousha, Michael S Brown, and Marcus A Brubaker. Noise2noiseflow: realistic camera noise modeling without clean images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17632-17641, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 463, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 518 + ], + "type": "text", + "content": "[33] Rohit Mangalwedhekar, Nivedita Singh, Chetan Singh Thakur, Chandra Sekhar Seelamantula, Mini Jose, and Deepak Nair. Achieving nanoscale precision using neuromorphic localization microscopy. Nature Nanotechnology, pages 1-10, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 574 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Jonathan T Barron, Jiawen Chen, Dillon Sharlet, Ren Ng, and Robert Carroll. Burst denoising with kernel prediction networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2502-2510, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 610 + ], + "type": "text", + "content": "[35] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[36] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2519-2533, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[37] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. Advances in Neural Information Processing Systems, 33:16639-16652, 2020. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[38] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982, 2018. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[39] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE transactions on pattern analysis and machine intelligence, 43(6):1964-1980, 2019. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[40] Timo Stoffregen and Lindsay Kleeman. Event cameras, contrast maximization and reward functions: An analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12300-12308, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 249 + ], + "type": "text", + "content": "[41] Timo Stoffregen, Guillermo Gallego, Tom Drummond, Lindsay Kleeman, and Davide Scaramuzza. Event-based motion segmentation by motion compensation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7244-7253, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 251, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 282 + ], + "type": "text", + "content": "[42] Cadence Design Systems. Software downloads-cadence design systems. https://www.cadence.com/en_US/home/support/software-downloads.html. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 337 + ], + "type": "text", + "content": "[43] Siqi Wang, Tee Hiang Cheng, and Meng-Hiot Lim. Ltmd: Learning improvement of spiking neural networks with learnable thresholding neurons and moderate dropout. Advances in Neural Information Processing Systems, 35: 28350-28362, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 403 + ], + "type": "text", + "content": "[44] Yanxiang Wang, Bowen Du, Yiran Shen, Kai Wu, Guangrong Zhao, Jianguo Sun, and Hongkai Wen. Ev-gait: Event-based robust gait recognition using dynamic vision sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6367, 2019. 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 469 + ], + "type": "text", + "content": "[45] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1609–1619, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 524 + ], + "type": "text", + "content": "[46] Kaixuan Wei, Ying Fu, Jiaolong Yang, and Hua Huang. A physics-based noise formation model for extreme low-light raw denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2758-2767, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 568 + ], + "type": "text", + "content": "[47] Jinjian Wu, Chuanwei Ma, Leida Li, Weisheng Dong, and Guangming Shi. Probabilistic undirected graph based denoising method for dynamic vision sensor. IEEE Transactions on Multimedia, 23:1148-1159, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[48] Man Yao, Huanhuan Gao, Guangshe Zhao, Dingheng Wang, Yihan Lin, Zhaoxu Yang, and Guoqi Li. Temporal-wise attention spiking neural networks for event streams classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10221-10230, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 679 + ], + "type": "text", + "content": "[49] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[50] Pei Zhang, Zhou Ge, Li Song, and Edmund Y L. Neuromorphic imaging with density-based spatiotemporal denoising. IEEE Transactions on Computational Imaging, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "25646" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 185 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[51] Xiang Zhang, Wei Liao, Lei Yu, Wen Yang, and Gui-Song Xia. Event-based synthetic aperture imaging with a hybrid network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14235-14244, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[52] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3594-3604, 2022. 6" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25647" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_content_list.json b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2aec51c7701a5d8c0cfe4203640bf0ca6c5c213b --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "LEMON: Learning 3D Human-Object Interaction Relation from 2D Images", + "text_level": 1, + "bbox": [ + 101, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhang Yang $^{1}$ , Wei Zhai $^{1,\\dagger}$ , Hongchen Luo $^{1}$ , Yang Cao $^{1,2}$ , Zheng-Jun Zha $^{1}$", + "bbox": [ + 186, + 179, + 779, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Science and Technology of China", + "bbox": [ + 292, + 198, + 679, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center", + "bbox": [ + 158, + 215, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yyuhang@mail., wzhai056@, lhcl2@mail., forrest@, zhazj@}ustc.edu.cn", + "bbox": [ + 181, + 237, + 784, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27b4b9f9d5af39aae9820d3c1c34bd1c7e4c90eae5e1f22245c4d2f17ce42d3b.jpg", + "image_caption": [ + "Figure 1. For an interaction image with paired geometries of the human and object, LEMON learns 3D human-object interaction relation by jointly anticipating the interaction elements, including human contact, object affordance, and human-object spatial relation. Vertices in yellow denote those in contact with the object, regions in red are object affordance regions, and the translucent sphere is the object proxy." + ], + "image_footnote": [], + "bbox": [ + 98, + 260, + 870, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 472, + 313, + 488 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning 3D human-object interaction relation is pivotal to embodied AI and interaction modeling. Most existing methods approach the goal by learning to predict isolated interaction elements, e.g., human contact, object affordance, and human-object spatial relation, primarily from the perspective of either the human or the object. Which underexploit certain correlations between the interaction counterparts (human and object), and struggle to address the uncertainty in interactions. Actually, objects' functionalities potentially affect humans' interaction intentions, which reveals what the interaction is. Meanwhile, the interacting humans and objects exhibit matching geometric structures, which presents how to interact. In light of this, we propose harnessing these inherent correlations between interaction counterparts to mitigate the uncertainty and jointly anticipate the above interaction elements in 3D space. To achieve this, we present LEMON (LEarning 3D huMan-Object iNteraction relation), a unified model that mines interaction intentions of the counterparts and employs curvatures to guide the extraction of geometric correlations, combining them to anticipate the interaction elements. Besides, the 3D Interaction Relation dataset (3DIR) is collected to serve as the test bed for training and evaluation. Extensive experiments demonstrate the", + "bbox": [ + 73, + 507, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "superiority of LEMON over methods estimating each element in isolation. The code and dataset are available at https://yyvhang.github.io/LEMON.", + "bbox": [ + 500, + 474, + 892, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 540, + 632, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning 3D human-object interaction (HOI) relation seeks to capture semantic co-occurrence and geometric compatibility between humans and objects in 3D space [5, 68, 80]. How can machines learn the interaction relation? One possible solution involves perceiving certain elements capable of revealing the interaction. Contact [6, 21, 64, 70], affordance [10, 13, 46, 92], and spatial relation [19, 31, 52] that elucidate \"where\" the interaction manifests between the human and object garner great attention. Capturing representations of such elements is pivotal for applications like AR/VR [7], imitation learning [1, 24], embodied AI [17, 58, 60], and interaction modeling [20, 81].", + "bbox": [ + 496, + 566, + 892, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans predominantly manipulate and interact with objects in 3D space. Thus, many methods devise task-specific models to anticipate certain interaction elements, thereby perceiving 3D HOI relation. Methods [21, 62, 64] estimate dense human contact based on the interaction semantics depicted in images. Some studies anticipate the object affordance according to objects' structures or 2D visuals [10, 44, 67, 78]. Several works delve into predicting the human-object spatial relation through synthetic images [19] or posed human geometries [52]. A prevailing trend in", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding Author.", + "bbox": [ + 101, + 887, + 232, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16284", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "these methods involves taking certain attributes (e.g., appearances, geometries) of either humans or objects to predict an isolated interaction element, capturing one aspect of the interaction relation. However, they overlook the impact of factors mutually determined by humans and objects on interactions, such as the interaction intention and geometric correlation, which leads to struggles in addressing the uncertainty within interactions. Specifically, the uncertainty could be attributed to two principal aspects. On the one hand, the diversity of HOIs introduces challenges in capturing explicit interaction intentions solely centering on humans or objects, raising the intention uncertainty. On the other hand, the limited view and mutual occlusions give rise to invisible interaction regions in images, complicating the constitution of linkages between these regions and target 3D geometries, causing the region uncertainty. These uncertainties may culminate in ambiguous anticipations.", + "bbox": [ + 75, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose leveraging both counterparts of the interaction to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space (Fig. 1), addressing the uncertainty by unearthing the correlation between the interacting humans and objects. Actually, humans and objects are intertwined and possess affinities in the interaction (Fig. 2). In specific, the design of objects typically adheres to certain human needs. Therefore, object affordances inherently hint at \"what\" interactions humans intend to make [39, 61, 63], revealing the intention affinity of interactions. Meanwhile, the interacting human and object exhibit matching geometries (either posture or configuration), which presents \"how\" to interact, arising the geometry affinity. The intention affinity clarifies the interaction type and implicates the interaction regions. Geometry affinity could serve as pivotal clues for excavating correlations between geometries corresponding to invisible regions in the image. These interaction-related regions e.g., contact regions, further reflect the human-object spatial relation.", + "bbox": [ + 75, + 351, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, we present the LEMON, a novel framework that correlates the intention semantics and geometric correspondences to jointly anticipate human contact, object affordance, and human-object spatial relation, in 3D space. To capture the intention affinity, LEMON employs multibranch attention to model the correlation between the interaction content in images and geometries of humans and objects, revealing intention representations of the interaction corresponding to geometries. The cosine similarity is utilized to further ensure their semantic consistency. Taking intention representations as conditions, LEMON integrates geometric curvatures to capture the geometry affinity and reveal the human contact and object affordance representations. These representations then assist in anticipating the spatial relation constrained by a combined distance loss.", + "bbox": [ + 75, + 641, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we collect the 3DIR dataset, which contains natural HOI images paired with object point clouds", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f0e94512c6a1f29f5d79eee91ab59b5bae98125bde8267268f10c74dfaa2b57c.jpg", + "image_caption": [ + "Figure 2. Motivation. Affinities within the HOI. The object affordance inherently reveals the human's interaction intention, arising the intention affinity. The interacting human and object possess matching structures, exhibiting the geometry affinity." + ], + "image_footnote": [], + "bbox": [ + 519, + 85, + 875, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and SMPL-H [57] pseudo-GTs. Multiple annotations are made for these data, e.g., dense human contact, object affordance, and human-object spatial relation. It serves as the test bed for the model training and evaluation.", + "bbox": [ + 498, + 364, + 890, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions are summarized as follows:", + "bbox": [ + 517, + 425, + 821, + 439 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We thoroughly exploit the correlation between the interaction counterparts to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space. It furnishes essential interaction elements to comprehend 3D HOIs.", + "2) We present the LEMON, a novel framework that correlates semantic interaction intention and geometries to capture the affinity between humans and objects, eliminating the impact of the interaction uncertainty and anticipating plausible 3D interaction elements.", + "3) We provide the 3DIR dataset that contains paired HOI data and multiple annotations, including dense human contact, object affordance, and human-object spatial relation, to support the anticipation. Extensive experiments demonstrate the superiority of LEMON." + ], + "bbox": [ + 501, + 440, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 686, + 640, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human-Object Interaction Relation in 2D", + "text_level": 1, + "bbox": [ + 500, + 710, + 857, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various approaches perceive elements pertinent to human-object interaction relation from multiple perspectives in the 2D domain. Methods [4, 16, 30, 72, 94] represent human-object interaction as triplets $<$ human, action, object>. Grounding bounding boxes of interaction counterparts and actions corresponding to them. They output the instance-level human-object spatial relation in pixel space and action semantics. Delving into the object side, the concept of \"interaction possibility\" is encapsulated in the term affordance [14], and bountiful methods explore to anticipate the object affordance, whether at instance-level or part-level", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16285", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[11, 12, 37, 38, 47, 85]. Turning to the human side, the contact explicitly indicates where the human interacts with objects. Some works detect the contact for specific parts (e.g., hand and foot) at a box-region level [48, 55, 59]. HOT [6] provides 2D contact annotations to support human-object contact estimation in images. These efforts achieve substantial results in understanding the human-object interaction relation in 2D space. However, they still encounter challenges when extrapolating to 3D space for practical applications due to the lack of a dimension.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human-Object Interaction Relation in 3D", + "text_level": 1, + "bbox": [ + 76, + 253, + 434, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate the incorporation of interaction comprehension into applications, extensive works explore to anticipate interaction elements in 3D space. For object affordance, methods [10, 44, 67, 69, 73, 86, 91] constitute the linkage between object shapes and affordances by a learning-based mapping, approaching the matter from the perspective of objects' functionality and structure. Some approaches learn object affordances from a distinct perspective, based on interactions. Whether making agents actively interact with 3D synthetic scenarios [46], or ground 3D object affordances through object-object [45] and 2D [78] interactions.", + "bbox": [ + 75, + 277, + 470, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dense human contact estimation is primarily based on the SMPL series [36, 50, 57]. Numerous methods model human contact for various tasks [8, 20, 22, 23, 71, 82, 83]. HULC [62], BSTRO [21], and DECO [64] are more similar to the estimation of contact in our method. They learn a mapping from the interaction semantics in images to the human vertices sequence, hardly leveraging geometries. Different from them, our method harnesses semantic and geometric correlations of interaction counterparts to address the uncertainty and infer the vertices that contact with objects. DECO also contributes the DAMON [64] dataset that possesses dense 3D contact annotation for in-the-wild images. In addition to human contact annotations, 3DIR includes other annotations related to the interaction, e.g., 3D object affordance and human-object spatial relation.", + "bbox": [ + 75, + 446, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bounding boxes obtained by HOI detection methods give the human-object spatial relation in pixel space. To lift this relation to 3D space, DJ-RN [31] takes hollow spheres with defined radii to represent objects, and projects the spatial relation in images to 3D space based on the bounding boxes and defined radii. CHORUS [19] utilizes synthetic multi-view images generated by the diffusion model [56] to learn the spatial relation between objects and canonical humans. Given a posed human, Object Pop-Up [51] anticipates the objects and their spatial positions to match the human body for certain interactions.", + "bbox": [ + 75, + 672, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above methods commonly focus on one side of the interaction in isolation. In contrast, our method captures the inherent affinity between both sides of the interaction to jointly anticipate interaction elements in 3D space. Such", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "anticipations are beneficial for tasks e.g., robot manipulation [3, 40, 93], interaction generation [27, 29, 75, 81, 90] and reconstruction [20, 71, 76, 87].", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 148, + 591, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the inputs $\\{H,O,I\\}$ , where $H\\in \\mathbb{R}^{N_h\\times 3}$ indicates vertices of the SMPL-H [57], which represents the human as pose $\\theta \\in \\mathbb{R}^{52\\times 3}$ , shape $\\beta \\in \\mathbb{R}^{10}$ parameters, and output a mesh $M(\\theta ,\\beta)\\in \\mathbb{R}^{6890\\times 3}$ . $O\\in \\mathbb{R}^{N_o\\times 3}$ is an object point cloud, $I\\in \\mathbb{R}^{H\\times W\\times 3}$ is an image. $N_{h},N_{o}$ are the number of points, $H,W$ are image's height and width. LEMON jointly anticipates the human contact $\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}$ , object affordance $\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}$ , and object center position $\\bar{\\phi}_p\\in \\mathbb{R}^3$ , in 3D space. As shown in Fig. 3, initially, the inputs are sent to image and point cloud backbones, obtain respective features $\\mathbf{F}_h,\\mathbf{F}_o,\\mathbf{F}_i$ . Then, LEMON correlates $\\mathbf{F}_i$ with $\\mathbf{F}_o$ , $\\mathbf{F}_h$ to extract features of the interaction intention ( $\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h$ ) inherent in geometries through multi-branch attention, and employs cosine similarity to constrain their semantic consistency (Sec. 3.1). With $\\bar{\\mathbf{T}}_o$ , $\\bar{\\mathbf{T}}_h$ as conditions, LEMON integrates curvatures to guide the modeling of geometric correlations, capturing the geometry affinity and revealing the contact $\\phi_c$ and affordance $\\phi_a$ features (Sec. 3.2). Next, geometric features and $\\phi_c,\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h$ are utilized to model the object spatial feature $\\phi_p$ (Sec. 3.3). Eventually, $\\phi_c,\\phi_a,\\phi_p$ are projected to $\\bar{\\phi}_c,\\bar{\\phi}_a$ , and $\\bar{\\phi}_p$ in the decoder, the whole process is optimized by a combined loss (Sec. 3.4).", + "bbox": [ + 496, + 172, + 893, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Interaction Intention Excavation", + "text_level": 1, + "bbox": [ + 498, + 513, + 787, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We add human and object masks on $I$ , and utilize the HR-Net [65] and DGCNN [66] as backbones to extract the image feature $\\mathbf{F}_i \\in \\mathbb{R}^{C \\times h \\times w}$ , the geometric features of human $\\mathbf{F}_h \\in \\mathbb{R}^{C \\times N_h}$ , and object $\\mathbf{F}_o \\in \\mathbb{R}^{C \\times N_o}$ , $\\mathbf{F}_i$ is flattened to $\\mathbb{R}^{C \\times hw}$ . Images contain rich interaction semantics, which could serve as clues to unearth the interaction intention within the geometries. In detail, tokens $\\mathbf{T}_o$ , $\\mathbf{T}_h \\in \\mathbb{R}^{C \\times 1}$ are generated to represent intention features inherent in geometries. $\\mathbf{T}_o$ , $\\mathbf{T}_h$ are concatenated with $\\mathbf{F}_o$ and $\\mathbf{F}_h$ respectively to get the feature sequences $\\mathbf{F}_{to} \\in \\mathbb{R}^{C \\times (N_o + 1)}$ , $\\mathbf{F}_{th} \\in \\mathbb{R}^{C \\times (N_h + 1)}$ . Taking $\\mathbf{F}_i$ as the shared key and value, $\\mathbf{F}_{to}$ , $\\mathbf{F}_{th}$ as queries $(q_1, q_2)$ in two branches, the multi-branch attention is employed to model the intention features, expressed as $\\bar{\\mathbf{F}}_{to}, \\bar{\\mathbf{F}}_{th} = f_{\\delta}([\\mathbf{F}_{to}, \\mathbf{F}_{th}], \\mathbf{F}_i)$ . Where $f_{\\delta}$ indicates the multi-branch attention layer [28, 54, 84], $[\\cdot]$ denotes two branches with different queries. Human and object geometries possess multiple interaction possibilities, which may introduce semantic ambiguity. To mitigate this, we further constrain the consistency of semantic tokens:", + "bbox": [ + 496, + 537, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi = \\frac {\\bar {\\mathbf {T}} _ {o} \\cdot \\bar {\\mathbf {T}} _ {h}}{\\| \\bar {\\mathbf {T}} _ {o} \\| _ {2} \\times \\| \\bar {\\mathbf {T}} _ {h} \\| _ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 830, + 890, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\in \\mathbb{R}^{C\\times 1}$ are split from $\\bar{\\mathbf{F}}_{to},\\bar{\\mathbf{F}}_{th}$ , representing the semantic intention features of geometries. $\\varphi$ is the", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16286", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e0c94c7a1bed6d704787b1ae2ae4395db9ed8d68366ca629776ea01f23dcf095.jpg", + "image_caption": [ + "Figure 3. LEMON pipeline. Initially, it takes modality-specific backbones to extract respective features $\\mathbf{F}_h$ , $\\mathbf{F}_o$ , $\\mathbf{F}_i$ , which are then utilized to excavate intention features $(\\bar{\\mathbf{T}}_o, \\bar{\\mathbf{T}}_h)$ of the interaction (Sec. 3.1). With $\\bar{\\mathbf{T}}_o$ , $\\bar{\\mathbf{T}}_h$ as conditions, LEMON integrates curvatures $(C_o, C_h)$ to model geometric correlations and reveal the contact $\\phi_c$ , affordance $\\phi_a$ features (Sec. 3.2). Following, the $\\phi_c$ is injected into the calculation of the object spatial feature $\\phi_p$ (Sec. 3.3). Eventually, the decoder projects $\\phi_c$ , $\\phi_a$ , $\\phi_p$ to the final outputs $\\bar{\\phi}_c$ , $\\bar{\\phi}_a$ , $\\bar{\\phi}_p$ ." + ], + "image_footnote": [], + "bbox": [ + 112, + 84, + 851, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "semantic cosine similarity, a part of $\\mathcal{L}_s$ in Sec. 3.4.", + "bbox": [ + 76, + 484, + 413, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Curvature-guided Geometric Correlation", + "text_level": 1, + "bbox": [ + 76, + 508, + 429, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The interacting humans and objects exhibit certain geometry affinity [77], manifesting in matching geometric structures and correlative curvatures. For geometric curvatures, the normal curvature could better represent local structures like interaction regions [26, 49]. Thus, we encode the normal curvatures into geometric features, and regard the $\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h$ as conditions to capture the affinity among human and object geometries. Normal curvatures $C_o\\in \\mathbb{R}^{1\\times N_o}$ $C_h\\in \\mathbb{R}^{1\\times N_h}$ of object and human are obtained by local fitting method [74, 89]. To correlate the curvatures, $C_o,C_h$ are encoded to high dimension $C_o'\\in \\mathbb{R}^{C\\times N_o},C_h'\\in \\mathbb{R}^{C\\times N_h}$ , and the cross-attention $f_{m}$ is mutually performed on them, formulated as $\\bar{C}_o = f_m(C_o',C_h'),\\bar{C}_h = f_m(C_h',C_o')$ . Following, to make the curvature guide the calculation of geometric correlation, $\\bar{C}_o,\\bar{C}_h$ are integrated and fused with geometric features:", + "bbox": [ + 75, + 532, + 472, + 775 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {F}} _ {c o} = f \\left(\\Gamma \\left(\\bar {C} _ {o}, \\bar {\\mathbf {F}} _ {o}\\right)\\right), \\bar {\\mathbf {F}} _ {c h} = f \\left(\\Gamma \\left(\\bar {C} _ {h}, \\bar {\\mathbf {F}} _ {h}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 782, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{F}}_{co},\\bar{\\mathbf{F}}_o\\in \\mathbb{R}^{C\\times N_o}$ and $\\bar{\\mathbf{F}}_{ch},\\bar{\\mathbf{F}}_h\\in \\mathbb{R}^{C\\times N_h}$ . $\\bar{\\mathbf{F}}_o,\\bar{\\mathbf{F}}_h$ are split from $\\bar{\\mathbf{F}}_{to}$ and $\\bar{\\mathbf{F}}_{th}$ , $\\Gamma$ denotes the concatenation, $f$ indicates convolution layers with $1\\times 1$ kernel. Then, $\\bar{\\mathbf{T}}_o$ and $\\bar{\\mathbf{T}}_h$ are considered as conditions [41] to further screen candidate regions that match the interaction depicted in images and participate in the modeling of geometry affinity,", + "bbox": [ + 75, + 808, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "revealing the affordance and contact features $\\phi_{a},\\phi_{c}$ . The process could be formulated as:", + "bbox": [ + 498, + 484, + 890, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {a} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\overline {{\\mathbf {T}}} _ {o}\\right)\\right), \\phi_ {c} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\overline {{\\mathbf {T}}} _ {h}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 530, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_{a} \\in \\mathbb{R}^{C \\times N_{o}}$ , to calculate it, $\\bar{\\mathbf{F}}_{co}$ serves as the query, and the concatenation of $\\bar{\\mathbf{F}}_{ch}$ and $\\bar{\\mathbf{T}}_{o}$ serves as the key and value, $f_{\\theta}$ is the cross-transformer layer. $\\phi_{c} \\in \\mathbb{R}^{C \\times N_{h}}$ is obtained through the analogous way.", + "bbox": [ + 496, + 561, + 893, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Contact-aware Spatial Relation", + "text_level": 1, + "bbox": [ + 498, + 636, + 777, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human-object interactions are extremely diverse, rendering the reasoning of their 3D spatial relation very challenging. Nevertheless, human contact implicitly constrains the position of the object, thereby assisting in inferring plausible human-object spatial relation. Thus, the object's center is represented as a token sequence $\\mathbf{T}_{sp} \\in \\mathbb{R}^{C \\times 3}$ , with introducing the contact feature as a constraint, LEMON takes the semantic intention and geometric features of the object to query the corresponding features of the human. Modeling $\\mathbf{T}_{sp}$ as the object spatial feature, formulated as:", + "bbox": [ + 496, + 661, + 890, + 813 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {p} = f _ {\\rho} \\left(\\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c o}\\right), \\overline {{\\mathbf {T}}} _ {o}, \\mathbf {T} _ {s p}\\right) + p e, \\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c h}\\right), \\overline {{\\mathbf {T}}} _ {h}, \\phi_ {c}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 824, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi_p\\in \\mathbb{R}^{C\\times 3}$ , is the spatial feature of the object center, $f_{\\rho}$ is a cross-attention layer, $\\Theta$ denotes the pooling layer, and $pe$ indicates a learnable positional encoding.", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16287", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/47a46c1086f386cf8589b776ab8b297149c6bb6c22bb525240bab82ad9c49e48.jpg", + "image_caption": [ + "Figure 4. 3DIR Dataset. (a) The quantity of images and point clouds for each object, and a data sample containing the image, mask, dense human contact annotation, 3D object with affordance annotation, and the fitted human mesh with the object proxy sphere. (b) The proportion of our contact annotations within 24 parts on SMPL [36], and distributions of contact vertices for certain HOIs. (c) The ratio of annotated affordance regions to the whole object geometries, and the distribution of this ratio for some categories. (d) Mean distances (unit: m) between annotated object centers and human pelvis joints, and directional projections of annotated centers for several objects." + ], + "image_footnote": [], + "bbox": [ + 107, + 84, + 300, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a54ab6ed10b8e2318b618eee84b8bb53b04c323bf078cb2cae034e1c6efc213c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 83, + 506, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b95b3c666aa345b8c03b13e593015b93624bf09d5e4bc8ee9702bf4bdf333ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 83, + 666, + 282 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36557fb43acd6489199a9bb243d6764f12300f27c7d334ae8a9329bfd3f6c374.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 83, + 864, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 378, + 228, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, $\\phi_{a},\\phi_{c},\\phi_{p}$ are sent to the decoder that contains three projection heads with linear, normalization, and activation layers. Which output object affordance $\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}$ , human contact $\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}$ , and object's center position $\\bar{\\phi}_p\\in \\mathbb{R}^3$ respectively. With defined radii for specific objects, we treat a sphere as the object proxy [31], representing the spatial relation in the camera coordinates of fitted humans. The overall training loss is expressed as:", + "bbox": [ + 75, + 402, + 470, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\omega_ {1} \\mathcal {L} _ {c} + \\omega_ {2} \\mathcal {L} _ {a} + \\omega_ {3} \\mathcal {L} _ {s} + \\omega_ {4} \\mathcal {L} _ {p}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 530, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{1-4}$ are hyper-parameters to balance the losses. $\\mathcal{L}_c$ and $\\mathcal{L}_a$ possess the same formulation, a focal loss [34] combined with a dice loss [42]. $\\mathcal{L}_c$ is calculated by $\\bar{\\phi}_c$ and the contact label $\\hat{\\phi}_c$ , while $\\mathcal{L}_a$ is calculated by $\\bar{\\phi}_a$ and the affordance annotation $\\hat{\\phi}_a$ . $\\mathcal{L}_s$ is employed to ensure $\\bar{\\mathbf{T}}_o$ , $\\bar{\\mathbf{T}}_h$ align with interaction semantics of the image and maintain consistency in the semantic space. Specifically, $\\mathbf{F}_i$ is mapped to a logit $y$ , which is used to calculate a cross-entropy loss $\\mathcal{L}_{ce}$ with interaction categorical label $\\hat{y}$ , $\\mathcal{L}_s$ is the sum of $\\mathcal{L}_{ce}$ and $\\varphi$ (Eq. 1). For 3D HOIs, object positions are diverse due to variations in human orientation, relying solely on absolute coordinates as supervision makes it hard for the model to learn a consistent mapping. However, the relative distances between humans and objects are similar for specific HOIs. This smoother distance distribution contributes to reducing the optimization space to an approximate sphere for specific HOIs. Consequently, we take the distance between the object center and the human pelvis joint as an additional constraint, and $\\mathcal{L}_p$ is formulated as:", + "bbox": [ + 75, + 551, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\mathcal {L} _ {p a} + \\mathcal {L} _ {p r}, \\quad \\mathcal {L} _ {p a} = | | \\bar {\\phi} _ {p} - \\hat {\\phi} _ {p} | | _ {2}, \\mathcal {L} _ {p r} = | | \\bar {\\xi} - \\hat {\\xi} | | _ {2}, (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 844, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\phi}_p$ is the annotation of the object center. $\\bar{\\xi},\\hat{\\xi}$ denotes distances between the pelvis joint and $\\bar{\\phi}_p,\\hat{\\phi}_p$ respectively.", + "bbox": [ + 76, + 866, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Dataset", + "text_level": 1, + "bbox": [ + 500, + 378, + 588, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce the collection and annotation protocols of the 3DIR and give some statistical analysis of the collected data and annotations, shown in Fig. 4.", + "bbox": [ + 498, + 407, + 890, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collection. We collect images with explicit interaction contents, in which humans interact with specific objects. These images adhere to the condition that the human's upper body is present, ensuring the efficient recovery of the human mesh. In total, we collect 5k in-the-wild images from HAKE [32], V-COCO [18], PIAD [78], and websites with free licenses, spanning 21 object classes and 17 interaction categories. Additionally, we collect over 5k 3D object instances from several 3D datasets [9, 10, 35, 43], based on the category of objects in collected images.", + "bbox": [ + 496, + 457, + 892, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotation. We make over 25k annotations with multiple types for the collected data. 1) Masks: With the assistance of the SAM [25], we manually mask the interacting human and object in the image. 2) Human Mesh: In natural images, some humans only exhibit the upper body. Thus, we utilize the pipeline of UBody [33] to fit SMPLH pseudo-GTs. The 2D body and hand joints needed by the pipeline are obtained through the DWPose [79]. 3) Contact: Similar to DAMON [64] and HOT [6], the human contact is annotated based on human knowledge. For each image, we clearly provide masks and a specific interaction type for the human-object pair, professional annotators are hired to \"draw\" vertices on the human template that contact with objects. The entire annotation process cycles three rounds, with each round incorporating subjective cross-checks and objective metric checks. 4) Affordance: We refer to the 3D-AffordanceNet [10] to annotate the object affordance. The annotations of 11 objects included in 3D-AffordanceNet are directly utilized, and we", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16288", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/89fcf5b7da4aeb5de3c36e9d3fb3eb33f846279311a1a531afb53ce8ef616257.jpg", + "table_caption": [ + "Table 1. Comparison on the 3DIR. Evaluation metrics of comparison methods on the benchmark, the best results are covered with the mask. indicates the improvement relative to the baseline. P. means take PointNet++ [53] as the backbone, and D. means DGCNN [66]." + ], + "table_footnote": [], + "table_body": "
Human ContactObject AffordanceSpatial Relation
MethodsPrecision ↑Recall ↑F1 ↑geo. (cm) ↓MethodsAUC ↑aIOU ↑SIM ↑MethodsMSE ↓
Baseline0.490.520.4932.83Baseline82.3632.630.50Baseline0.051
BSTRO [21]0.570.580.5528.583DA. [10]85.493.8%0.56DJ-RN [31]0.042
DECO [64]0.700.720.6915.25IAG [78]86.635.1%0.59PopUp [51]0.027
Ours P.0.760.810.779.02Ours P.87.9140.970.63Ours P.0.012
Ours D.0.780.820.787.55Ours D.88.5141.340.64Ours D.0.010
", + "bbox": [ + 76, + 128, + 895, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "An additional 10 object categories excluded in 3D-AffordanceNet. 5) Spatial Relation: For each sample, we color the fitted human mesh with annotated per-vertex human contact and treat a sphere as the object proxy. The radius of the sphere for each object category is pre-defined, referring to DJ-RN [31]. Given the posed human in color and the proxy sphere, annotators adjust the sphere relative to the human to align with the human-object spatial relation depicted in the image. Ultimately, we record the center coordinates of the adjusted sphere. Due to the page limitation, please refer to the Sup. Mat. for more annotation details.", + "bbox": [ + 75, + 279, + 468, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Statistical Analysis. The quantity of images and 3D instances for each object category in the 3DIR is shown in Fig. 4 (a). For contact annotation, we count its proportion within 24 human parts defined on the SMPL [36]. Moreover, we visualize the distribution of contact annotation for several HOIs, where the deeper color indicates more contact annotations at the vertex, as shown in Fig. 4 (b). Fig. 4 (c) demonstrates the proportion of annotated affordance regions to the entire object geometry, encompassing the mean of each category and the distribution of all instances for several categories. As can be seen, there are differences in distinct objects with the same affordance, as well as variations in distinct affordances of the same object. For spatial relation annotation, we count the mean distance between annotated centers and human pelvis joints, and project the spatial direction of annotated centers onto a fixed-radius sphere with the pelvis joint as the center. Fig. 4 (d) shows several cases and the distribution of mean distances.", + "bbox": [ + 75, + 446, + 470, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 731, + 202, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Benchmark Setting", + "text_level": 1, + "bbox": [ + 76, + 756, + 261, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We refer to methods that anticipate each interaction element [21, 47, 64, 78] to thoroughly benchmark the 3DIR. For the training, in addition to the training data in 3DIR, we select another 5k data with low redundancy in BEHAVE [2]. Evaluation is conducted on the test set of 3DIR. The baseline model adopts a multitask-like framework, directly utilizing three branches to make anticipations. Besides, we compare LEMON with advanced methods that anticipate respective", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "elements. Plus, we also conduct evaluation experiments on DAMON [64], BEHAVE [2], and PIAD [78], the results and implementation details are in the Sup. Mat.", + "bbox": [ + 496, + 280, + 890, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison Results", + "text_level": 1, + "bbox": [ + 498, + 332, + 692, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparison results of evaluation metrics are presented in Tab. 1. The baseline model, which does not capture correlations between interaction counterparts, indicates that directly anticipating these elements through multiple branches yields poor results. Our method outperforms comparative methods across all metrics for respective elements, demonstrating that leveraging correlations between humans and objects indeed benefits the comprehension of interaction relation, and the anticipation of interaction elements. It seems to be \"the best of both worlds\". Furthermore, we conduct a visual comparison of methods with higher evaluation metrics, as shown in Fig. 5. The results showcase multiple objects interacting with distinct human parts. As can be seen, our anticipations are more precise, for some uncertain regions that are not visible in images (e.g., the first and third rows), LEMON could also anticipate plausible results. This is attributed to the modeling of correlations between geometries, which compensates for missing features in invisible regions of the image.", + "bbox": [ + 496, + 356, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 650, + 653, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a thorough ablation study to validate the effectiveness of the model design. Tab. 2 (a) reports the model performance without modeling the semantic intention and the geometric correlation, demonstrating their impacts on the model performance. Moreover, the absence of constraining semantic consistency hinders the modeling of intentions, while not introducing curvatures affects the extraction of geometric correlations, resulting in a decrease in model performance. Both results are detailed in Tab. 2 (a). Besides, we test the impact of $\\phi_c$ and $\\mathcal{L}_{pr}$ (Eq. 6) on spatial relation prediction, shown in Tab. 2 (b). Notably, removing the extraction of $\\phi_a$ in $f_{\\theta}$ (Eq. 3) also affects the performance. This is attributed to the interrelation between $\\phi_c$ and $\\phi_a$ , the absence of $\\phi_a$ impacts the representation of $\\phi_c$ , and subsequently influencing the final spatial prediction.", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16289", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a20c21f91381d99efe9e5904feeb24f7d34000f7e98e42d19b90af7e3659020b.jpg", + "image_caption": [ + "Figure 5. Visualization Results. (a) Results of the estimated human vertices in contact with objects, the estimated contact vertices are shown in yellow. (b) The anticipations of 3D object affordance, the depth of red represents the probability of anticipated affordance. (c) Two views of the predicted spatial relation, translucent spheres are object proxies. Please zoom in for a better visualization." + ], + "image_footnote": [], + "bbox": [ + 122, + 80, + 861, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further illustrate the effectiveness of capturing the intention and geometry affinity, we visualize attention maps on the human and object geometries when removing one of them, as shown in Fig. 6. The first row indicates without capturing intention affinity, solely relying on geometries leads the model to focus on multiple candidate regions, differing from attention regions in the image. Furthermore, the second row demonstrates that while semantic intention guides the model to identify approximate candidate regions, modeling geometry affinity further locates corresponding regions most related to the interaction of the geometries.", + "bbox": [ + 75, + 566, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Performance Analysis", + "text_level": 1, + "bbox": [ + 76, + 741, + 282, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple Interactions. To validate whether the model reasons interaction elements based on the understanding of interaction relation. We employ the model to infer different interactions between humans and the same object, as demonstrated in Fig. 7. The results showcase that, for the same object, when interactions are different, the model outputs distinct results, encompassing human contact, object affordance, and spatial relation, all of which align with the interaction relation depicted in the image. This indicates", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e604793fd5c9934406956a42c75bc4d88e2e4177c3ba8fcc652f6d9036ad0c05.jpg", + "table_caption": [ + "Table 2. Ablation studies. (a) Performance when not modeling intention representations (intent.), geometric correlations (geom.), and not introducing $\\varphi$ (Eq. 1), curvatures (cur.). (b) The impact of several factors on spatial relation prediction. $\\pmb{x}$ means without." + ], + "table_footnote": [], + "table_body": "
MetricsOursX intent.X φX geom.X cur.
Precision0.780.710.740.680.75
Recall0.820.730.790.700.78
F10.780.700.740.670.75
geo. (cm)7.5511.8710.2614.879.13
AUC88.5185.8786.6683.2386.98
aIOU41.3438.1940.0337.1339.58
SIM0.640.580.600.550.62
MSE0.0100.0270.0220.0310.018
", + "bbox": [ + 501, + 632, + 887, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a)", + "bbox": [ + 650, + 768, + 666, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that the model infers interaction elements according to the interaction relation rather than relying on direct mapping solely driven by the categories of human-object pairs.", + "bbox": [ + 496, + 791, + 890, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multiple Objects. In scenarios where distinct objects are involved in concurrent interactions with humans, the model should indeed possess the ability to comprehend the interaction relation with different objects. We give an experi-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16290", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/515f7af041520d8c6b974735fd94545fc0e8e41ea0e0159ddba15d9a355162a1.jpg", + "image_caption": [ + "Figure 6. Visual Attention. Attention heatmaps on geometries, in cases with $(\\boldsymbol{w})$ and without $(\\boldsymbol{w} / \\boldsymbol{o})$ the extraction of intention and geometry affinity. Images are also shown with attention heatmaps." + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 455, + 247 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/367870f5c15a8499a823117baf1d9193b479f584a54d77fc99b83404fe10d731.jpg", + "image_caption": [ + "Support" + ], + "image_footnote": [], + "bbox": [ + 89, + 310, + 148, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbd7bea0d50476f68bb0d79d1ecd88a8f30b05cc1fb3b866f86d355bf51eb1c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 155, + 296, + 215, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f5534de2937f241c63a4f0151bc9bbf5e1d9ab3fd1923f33f1a9a5b7f5af81c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 301, + 290, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a281632b63ea86cdf1be7cb93a84780708011e603f022d610ae361bed98de558.jpg", + "image_caption": [ + "View 1" + ], + "image_footnote": [], + "bbox": [ + 305, + 306, + 366, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/113dbefaef8cfdb07d5f5f3ab8600f855d04a286b9f904755a888d99701fb366.jpg", + "image_caption": [ + "View 2" + ], + "image_footnote": [], + "bbox": [ + 375, + 306, + 457, + 363 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d471bb0dec73e5b7056b161ab935e6b01e43bee894e119937712d155d32d8b7a.jpg", + "image_caption": [ + "Carry" + ], + "image_footnote": [], + "bbox": [ + 89, + 381, + 148, + 424 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/283630f3bc4effd236f3887b7ef621b8552028f67677d97d05f18c8181053f80.jpg", + "image_caption": [ + "Affordance" + ], + "image_footnote": [], + "bbox": [ + 155, + 369, + 215, + 429 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e71f181a0c26d57af78c846cacd36e5a9dd319c3f4e4e4fb30c5e58e9183a74.jpg", + "image_caption": [ + "Contact", + "Figure 7. Multiple Interactions. Distinct anticipations when interactions are different with the same object." + ], + "image_footnote": [], + "bbox": [ + 223, + 369, + 303, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc43781236193e6c6217674838b5d323063ffa91653228ed98c7093f83a108eb.jpg", + "image_caption": [ + "Spatial Relation" + ], + "image_footnote": [], + "bbox": [ + 305, + 368, + 375, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment pertaining to this property. As shown in Fig. 8, when reasoning with different objects, the human contact, object affordance, and spatial relation are distinct, while all anticipations are plausible. This is attributed to the collaboration of the semantic interaction intention and geometric correlation, ensuring that the model could explicitly capture interaction intentions with different objects and reveal regions related to the interaction on corresponding geometries.", + "bbox": [ + 75, + 500, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple Instances. We conduct an experiment to validate the model's understanding of interaction relation across various instances, assessing its generalization and robustness. Specifically, distinct object instances are utilized to infer with the same image and human geometry, shown in Fig. 9 (a). As can be seen, even though the contact and spatial relation are slightly different, the results are consistent with the interaction shown in the image, and object affordances are correctly anticipated. Additionally, the same object instance is utilized to infer with different human geometries and images that contain similar human-object interactions, as demonstrated in Fig. 9 (b). These results indicate that the model could generalize the understanding of interaction relation to different human and object instances.", + "bbox": [ + 75, + 619, + 468, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 845, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose leveraging both counterparts of the HOI to jointly anticipate human contact, object affordance, and", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f89380510429d19137e6c96618c4bb3935c2f989b3963653646aab9e2670850.jpg", + "image_caption": [ + "Figure 8. Multiple Objects. The anticipations for multiple objects with different human-object interactions." + ], + "image_footnote": [], + "bbox": [ + 514, + 90, + 879, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/846879d2b64aa144c0162fce37c6ccc750690096cbae6408e2af7478237d2bd0.jpg", + "image_caption": [ + "Figure 9. Multiple Instances. (a) Different object instances $w.r.t.$ the same image and human geometry. (b) Different images and human instances $w.r.t.$ the same object." + ], + "image_footnote": [], + "bbox": [ + 511, + 262, + 879, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "human-object spatial relation in 3D space. It holds the potential to drive embodied intelligence to learn from interactions and assist the interaction modeling in graphics. With the provided dataset 3DIR containing paired HOI data and multiple annotations, we train LEMON, a novel model that harnesses the interaction intention and geometric correlation between humans and objects to capture the affinity, eliminating interaction uncertainties and anticipating plausible interaction elements. Extensive experiments demonstrate the effectiveness and superiority of LEMON. We believe this work offers fresh insights and paves a new way for 3D human-object interaction understanding.", + "bbox": [ + 496, + 542, + 890, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. In the current formulation, LEMON needs the pre-inferred human mesh as input. In the future, we plan to integrate HMR into the entire framework, freeing input constraints required for inference, and taking 3D interaction elements to improve the accuracy of HMR. Plus, another interesting direction is to leverage multi-modal methods [15, 88], boosting the interaction relation understanding from other sources e.g., text, audio.", + "bbox": [ + 496, + 729, + 890, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62306295.", + "bbox": [ + 496, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16291", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Brenna D Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469-483, 2009. 1", + "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 6", + "[3] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 3", + "[4] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018, the winter conference on applications of computer vision (wacv), pages 381–389. IEEE, 2018. 2", + "[5] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 1", + "[6] Yixin Chen, Sai Kumar Dwivedi, Michael J. Black, and Dimitrios Tzionas. Detecting human-object contact in images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5", + "[7] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22:449-462, 2013. 1", + "[8] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3", + "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5", + "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 1, 3, 5, 6", + "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3", + "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3", + "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 1", + "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 2", + "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 8", + "[16] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2", + "[17] Elena Corina Grigore, Kerstin Eder, Anthony G Pipe, Chris Melhuish, and Ute Leonards. Joint action understanding improves robot-to-human object handover. In 2013 IEEE/RSJ international conference on intelligent robots and systems, pages 4622-4629. IEEE, 2013. 1", + "[18] Saurabh Gupta and Jitendra Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 5", + "[19] Sookwan Han and Hanbyul Joo. Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3", + "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021. 1, 3", + "[21] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 1, 3, 6", + "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3", + "[23] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3", + "[24] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16292", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5", + "[26] Jan J Koenderink and Andrea J Van Doorn. Surface shape and curvature scales. Image and vision computing, 10(8): 557-564, 1992. 4", + "[27] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis, 2023. 3", + "[28] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 3", + "[29] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. arXiv preprint arXiv:2309.16237, 2023. 3", + "[30] Lulei Li, Jianan Wei, Wenguan Wang, and Yi Yang. Neural-logic human-object interaction detection. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[31] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10166–10175, 2020. 1, 3, 5, 6", + "[32] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 5", + "[33] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 5", + "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[35] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. AKB-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 5", + "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3, 5, 6", + "[37] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2252-2261, 2022. 3", + "[38] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Leverage interactive affinity for affordance" + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6809-6819, 2023. 3", + "[39] Bertram F Malle and Joshua Knobe. The folk concept of intentionality. Journal of experimental social psychology, 33 (2):101-121, 1997. 2", + "[40] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3", + "[41] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 4", + "[42] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5", + "[43] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 5", + "[44] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 1, 3", + "[45] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3", + "[46] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3", + "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 6", + "[48] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai Nguyen. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing systems, 33:7841-7851, 2020. 3", + "[49] Abdul Nurunnabi, Geoff West, and David Belton. Outlier detection and robust normal-curvature estimation in mobile laser scanning 3d point cloud data. Pattern Recognition, 48 (4):1404-1419, 2015. 4", + "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16293", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4726-4736, 2023. 3, 6", + "[52] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 7", + "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6", + "[54] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12802–12813, 2023. 3", + "[55] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 3", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 3", + "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 1", + "[59] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 3", + "[60] Thomas B Sheridan. Human-robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 1", + "[61] Chris Shilling. The body in culture, technology and society. The Body in Culture, Technology and Society, pages 1-256, 2004. 2", + "[62] Soshi Shimada, Vladislav Golyanik, Zhi Li, Patrick Pérez, Weipeng Xu, and Christian Theobalt. Hulc: 3d human motion capture with pose manifold sampling and dense contact guidance. In European Conference on Computer Vision, pages 516-533. Springer, 2022. 1, 3", + "[63] Michael Tomasello, Malinda Carpenter, Josep Call, Tanya Behne, and Henrik Moll. Understanding and sharing in-" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tentions: The origins of cultural cognition. Behavioral and brain sciences, 28(5):675-691, 2005. 2", + "[64] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 1, 3, 5, 6, 7", + "[65] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 3", + "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 6", + "[67] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In European Conference on Computer Vision, pages 90-107. Springer, 2022. 1, 3", + "[68] Ping Wei, Yibiao Zhao, Nanning Zheng, and Song-Chun Zhu. Modeling 4d human-object interactions for event and object recognition. In Proceedings of the IEEE international conference on computer vision, pages 3272-3279, 2013. 1", + "[69] Ruihai Wu, Chuanruo Ning, and Hao Dong. Learning foresightful dense visual affordance for deformable object manipulation. arXiv preprint arXiv:2303.11057, 2023. 3", + "[70] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1", + "[71] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, 2022. 3", + "[72] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and Mohan S Kankanhalli. Learning to detect human-object interactions with knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[73] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022.3", + "[74] Rui Xu, Zhiyang Dou, Ningna Wang, Shiqing Xin, Shuang-min Chen, Mingyan Jiang, Xiaohu Guo, Wenping Wang, and Changhe Tu. Globally consistent normal orientation for point clouds by regularizing the winding-number field. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 4", + "[75] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16294", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3d-hoi: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3", + "[77] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 4", + "[78] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10905-10915, 2023. 1, 3, 5, 6, 7", + "[79] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. arXiv preprint arXiv:2307.15880, 2023. 5", + "[80] Bangpeng Yao and Li Fei-Fei. Modeling mutual context of object and human pose in human-object interaction activities. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 17–24. IEEE, 2010. 1", + "[81] Sifan Ye, Yixing Wang, Jiaman Li, Dennis Park, C Karen Liu, Huazhe Xu, and Jiajun Wu. Scene synthesis from human motion. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 3", + "[82] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3", + "[83] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 3", + "[84] Guochen Yu, Andong Li, Chengshi Zheng, Yinuo Guo, Yutian Wang, and Hui Wang. Dual-branch attention-inattention transformer for single-channel speech enhancement. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7847-7851. IEEE, 2022. 3", + "[85] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10): 2472-2500, 2022. 3", + "[86] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, pages 1-26, 2023. 3", + "[87] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 34-51. Springer, 2020. 3", + "[88] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 8", + "[89] Xiaopeng Zhang, Hongjun Li, Zhanglin Cheng, and Yikuan Zhang. Robust curvature estimation and geometry analysis of 3d point cloud surfaces. J. Inf. Comput. Sci, 6(5):1983-1990, 2009. 4", + "[90] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3", + "[91] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022.3", + "[92] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper manipulation. In International Conference on Learning Representations, 2023. 1", + "[93] Xinghao Zhu, JingHan Ke, Zhixuan Xu, Zhixin Sun, Bizhe Bai, Jun Lv, Qingtao Liu, Yuwei Zeng, Qi Ye, Cewu Lu, et al. Diff-lfd: Contact-aware model-based learning from visual demonstration for robotic manipulation via differentiable physics-based simulation and rendering. In 7th Annual Conference on Robot Learning, 2023. 3", + "[94] Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11825-11834, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 555 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16295", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_model.json b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_model.json new file mode 100644 index 0000000000000000000000000000000000000000..accb36b6c11a4afec107a8a324d830baf2b841da --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_model.json @@ -0,0 +1,2710 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.154 + ], + "angle": 0, + "content": "LEMON: Learning 3D Human-Object Interaction Relation from 2D Images" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.18, + 0.78, + 0.2 + ], + "angle": 0, + "content": "Yuhang Yang\\(^{1}\\), Wei Zhai\\(^{1,\\dagger}\\), Hongchen Luo\\(^{1}\\), Yang Cao\\(^{1,2}\\), Zheng-Jun Zha\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.199, + 0.68, + 0.218 + ], + "angle": 0, + "content": "1 University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.217, + 0.812, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.238, + 0.785, + 0.251 + ], + "angle": 0, + "content": "{yyuhang@mail., wzhai056@, lhcl2@mail., forrest@, zhazj@}ustc.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.261, + 0.871, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.419, + 0.893, + 0.462 + ], + "angle": 0, + "content": "Figure 1. For an interaction image with paired geometries of the human and object, LEMON learns 3D human-object interaction relation by jointly anticipating the interaction elements, including human contact, object affordance, and human-object spatial relation. Vertices in yellow denote those in contact with the object, regions in red are object affordance regions, and the translucent sphere is the object proxy." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.473, + 0.314, + 0.489 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.508, + 0.474, + 0.871 + ], + "angle": 0, + "content": "Learning 3D human-object interaction relation is pivotal to embodied AI and interaction modeling. Most existing methods approach the goal by learning to predict isolated interaction elements, e.g., human contact, object affordance, and human-object spatial relation, primarily from the perspective of either the human or the object. Which underexploit certain correlations between the interaction counterparts (human and object), and struggle to address the uncertainty in interactions. Actually, objects' functionalities potentially affect humans' interaction intentions, which reveals what the interaction is. Meanwhile, the interacting humans and objects exhibit matching geometric structures, which presents how to interact. In light of this, we propose harnessing these inherent correlations between interaction counterparts to mitigate the uncertainty and jointly anticipate the above interaction elements in 3D space. To achieve this, we present LEMON (LEarning 3D huMan-Object iNteraction relation), a unified model that mines interaction intentions of the counterparts and employs curvatures to guide the extraction of geometric correlations, combining them to anticipate the interaction elements. Besides, the 3D Interaction Relation dataset (3DIR) is collected to serve as the test bed for training and evaluation. Extensive experiments demonstrate the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.475, + 0.893, + 0.52 + ], + "angle": 0, + "content": "superiority of LEMON over methods estimating each element in isolation. The code and dataset are available at https://yyvhang.github.io/LEMON." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.541, + 0.633, + 0.557 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.749 + ], + "angle": 0, + "content": "Learning 3D human-object interaction (HOI) relation seeks to capture semantic co-occurrence and geometric compatibility between humans and objects in 3D space [5, 68, 80]. How can machines learn the interaction relation? One possible solution involves perceiving certain elements capable of revealing the interaction. Contact [6, 21, 64, 70], affordance [10, 13, 46, 92], and spatial relation [19, 31, 52] that elucidate \"where\" the interaction manifests between the human and object garner great attention. Capturing representations of such elements is pivotal for applications like AR/VR [7], imitation learning [1, 24], embodied AI [17, 58, 60], and interaction modeling [20, 81]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Humans predominantly manipulate and interact with objects in 3D space. Thus, many methods devise task-specific models to anticipate certain interaction elements, thereby perceiving 3D HOI relation. Methods [21, 62, 64] estimate dense human contact based on the interaction semantics depicted in images. Some studies anticipate the object affordance according to objects' structures or 2D visuals [10, 44, 67, 78]. Several works delve into predicting the human-object spatial relation through synthetic images [19] or posed human geometries [52]. A prevailing trend in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.233, + 0.901 + ], + "angle": 0, + "content": "†Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "these methods involves taking certain attributes (e.g., appearances, geometries) of either humans or objects to predict an isolated interaction element, capturing one aspect of the interaction relation. However, they overlook the impact of factors mutually determined by humans and objects on interactions, such as the interaction intention and geometric correlation, which leads to struggles in addressing the uncertainty within interactions. Specifically, the uncertainty could be attributed to two principal aspects. On the one hand, the diversity of HOIs introduces challenges in capturing explicit interaction intentions solely centering on humans or objects, raising the intention uncertainty. On the other hand, the limited view and mutual occlusions give rise to invisible interaction regions in images, complicating the constitution of linkages between these regions and target 3D geometries, causing the region uncertainty. These uncertainties may culminate in ambiguous anticipations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.473, + 0.639 + ], + "angle": 0, + "content": "In this paper, we propose leveraging both counterparts of the interaction to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space (Fig. 1), addressing the uncertainty by unearthing the correlation between the interacting humans and objects. Actually, humans and objects are intertwined and possess affinities in the interaction (Fig. 2). In specific, the design of objects typically adheres to certain human needs. Therefore, object affordances inherently hint at \"what\" interactions humans intend to make [39, 61, 63], revealing the intention affinity of interactions. Meanwhile, the interacting human and object exhibit matching geometries (either posture or configuration), which presents \"how\" to interact, arising the geometry affinity. The intention affinity clarifies the interaction type and implicates the interaction regions. Geometry affinity could serve as pivotal clues for excavating correlations between geometries corresponding to invisible regions in the image. These interaction-related regions e.g., contact regions, further reflect the human-object spatial relation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.473, + 0.868 + ], + "angle": 0, + "content": "To achieve this, we present the LEMON, a novel framework that correlates the intention semantics and geometric correspondences to jointly anticipate human contact, object affordance, and human-object spatial relation, in 3D space. To capture the intention affinity, LEMON employs multibranch attention to model the correlation between the interaction content in images and geometries of humans and objects, revealing intention representations of the interaction corresponding to geometries. The cosine similarity is utilized to further ensure their semantic consistency. Taking intention representations as conditions, LEMON integrates geometric curvatures to capture the geometry affinity and reveal the human contact and object affordance representations. These representations then assist in anticipating the spatial relation constrained by a combined distance loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "In addition, we collect the 3DIR dataset, which contains natural HOI images paired with object point clouds" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.087, + 0.877, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.282, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Motivation. Affinities within the HOI. The object affordance inherently reveals the human's interaction intention, arising the intention affinity. The interacting human and object possess matching structures, exhibiting the geometry affinity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.425 + ], + "angle": 0, + "content": "and SMPL-H [57] pseudo-GTs. Multiple annotations are made for these data, e.g., dense human contact, object affordance, and human-object spatial relation. It serves as the test bed for the model training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.426, + 0.822, + 0.44 + ], + "angle": 0, + "content": "The contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.516 + ], + "angle": 0, + "content": "1) We thoroughly exploit the correlation between the interaction counterparts to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space. It furnishes essential interaction elements to comprehend 3D HOIs." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.594 + ], + "angle": 0, + "content": "2) We present the LEMON, a novel framework that correlates semantic interaction intention and geometries to capture the affinity between humans and objects, eliminating the impact of the interaction uncertainty and anticipating plausible 3D interaction elements." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.597, + 0.892, + 0.673 + ], + "angle": 0, + "content": "3) We provide the 3DIR dataset that contains paired HOI data and multiple annotations, including dense human contact, object affordance, and human-object spatial relation, to support the anticipation. Extensive experiments demonstrate the superiority of LEMON." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.642, + 0.702 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.858, + 0.728 + ], + "angle": 0, + "content": "2.1. Human-Object Interaction Relation in 2D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Various approaches perceive elements pertinent to human-object interaction relation from multiple perspectives in the 2D domain. Methods [4, 16, 30, 72, 94] represent human-object interaction as triplets \\(<\\) human, action, object>. Grounding bounding boxes of interaction counterparts and actions corresponding to them. They output the instance-level human-object spatial relation in pixel space and action semantics. Delving into the object side, the concept of \"interaction possibility\" is encapsulated in the term affordance [14], and bountiful methods explore to anticipate the object affordance, whether at instance-level or part-level" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16285" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.243 + ], + "angle": 0, + "content": "[11, 12, 37, 38, 47, 85]. Turning to the human side, the contact explicitly indicates where the human interacts with objects. Some works detect the contact for specific parts (e.g., hand and foot) at a box-region level [48, 55, 59]. HOT [6] provides 2D contact annotations to support human-object contact estimation in images. These efforts achieve substantial results in understanding the human-object interaction relation in 2D space. However, they still encounter challenges when extrapolating to 3D space for practical applications due to the lack of a dimension." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.255, + 0.436, + 0.272 + ], + "angle": 0, + "content": "2.2. Human-Object Interaction Relation in 3D" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.471, + 0.446 + ], + "angle": 0, + "content": "To facilitate the incorporation of interaction comprehension into applications, extensive works explore to anticipate interaction elements in 3D space. For object affordance, methods [10, 44, 67, 69, 73, 86, 91] constitute the linkage between object shapes and affordances by a learning-based mapping, approaching the matter from the perspective of objects' functionality and structure. Some approaches learn object affordances from a distinct perspective, based on interactions. Whether making agents actively interact with 3D synthetic scenarios [46], or ground 3D object affordances through object-object [45] and 2D [78] interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.673 + ], + "angle": 0, + "content": "The dense human contact estimation is primarily based on the SMPL series [36, 50, 57]. Numerous methods model human contact for various tasks [8, 20, 22, 23, 71, 82, 83]. HULC [62], BSTRO [21], and DECO [64] are more similar to the estimation of contact in our method. They learn a mapping from the interaction semantics in images to the human vertices sequence, hardly leveraging geometries. Different from them, our method harnesses semantic and geometric correlations of interaction counterparts to address the uncertainty and infer the vertices that contact with objects. DECO also contributes the DAMON [64] dataset that possesses dense 3D contact annotation for in-the-wild images. In addition to human contact annotations, 3DIR includes other annotations related to the interaction, e.g., 3D object affordance and human-object spatial relation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Bounding boxes obtained by HOI detection methods give the human-object spatial relation in pixel space. To lift this relation to 3D space, DJ-RN [31] takes hollow spheres with defined radii to represent objects, and projects the spatial relation in images to 3D space based on the bounding boxes and defined radii. CHORUS [19] utilizes synthetic multi-view images generated by the diffusion model [56] to learn the spatial relation between objects and canonical humans. Given a posed human, Object Pop-Up [51] anticipates the objects and their spatial positions to match the human body for certain interactions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The above methods commonly focus on one side of the interaction in isolation. In contrast, our method captures the inherent affinity between both sides of the interaction to jointly anticipate interaction elements in 3D space. Such" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "anticipations are beneficial for tasks e.g., robot manipulation [3, 40, 93], interaction generation [27, 29, 75, 81, 90] and reconstruction [20, 71, 76, 87]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.149, + 0.593, + 0.166 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Given the inputs \\(\\{H,O,I\\}\\), where \\(H\\in \\mathbb{R}^{N_h\\times 3}\\) indicates vertices of the SMPL-H [57], which represents the human as pose \\(\\theta \\in \\mathbb{R}^{52\\times 3}\\), shape \\(\\beta \\in \\mathbb{R}^{10}\\) parameters, and output a mesh \\(M(\\theta ,\\beta)\\in \\mathbb{R}^{6890\\times 3}\\). \\(O\\in \\mathbb{R}^{N_o\\times 3}\\) is an object point cloud, \\(I\\in \\mathbb{R}^{H\\times W\\times 3}\\) is an image. \\(N_{h},N_{o}\\) are the number of points, \\(H,W\\) are image's height and width. LEMON jointly anticipates the human contact \\(\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}\\), object affordance \\(\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}\\), and object center position \\(\\bar{\\phi}_p\\in \\mathbb{R}^3\\), in 3D space. As shown in Fig. 3, initially, the inputs are sent to image and point cloud backbones, obtain respective features \\(\\mathbf{F}_h,\\mathbf{F}_o,\\mathbf{F}_i\\). Then, LEMON correlates \\(\\mathbf{F}_i\\) with \\(\\mathbf{F}_o\\), \\(\\mathbf{F}_h\\) to extract features of the interaction intention (\\(\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\)) inherent in geometries through multi-branch attention, and employs cosine similarity to constrain their semantic consistency (Sec. 3.1). With \\(\\bar{\\mathbf{T}}_o\\), \\(\\bar{\\mathbf{T}}_h\\) as conditions, LEMON integrates curvatures to guide the modeling of geometric correlations, capturing the geometry affinity and revealing the contact \\(\\phi_c\\) and affordance \\(\\phi_a\\) features (Sec. 3.2). Next, geometric features and \\(\\phi_c,\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\) are utilized to model the object spatial feature \\(\\phi_p\\) (Sec. 3.3). Eventually, \\(\\phi_c,\\phi_a,\\phi_p\\) are projected to \\(\\bar{\\phi}_c,\\bar{\\phi}_a\\), and \\(\\bar{\\phi}_p\\) in the decoder, the whole process is optimized by a combined loss (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.788, + 0.53 + ], + "angle": 0, + "content": "3.1. Interaction Intention Excavation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.827 + ], + "angle": 0, + "content": "We add human and object masks on \\(I\\), and utilize the HR-Net [65] and DGCNN [66] as backbones to extract the image feature \\(\\mathbf{F}_i \\in \\mathbb{R}^{C \\times h \\times w}\\), the geometric features of human \\(\\mathbf{F}_h \\in \\mathbb{R}^{C \\times N_h}\\), and object \\(\\mathbf{F}_o \\in \\mathbb{R}^{C \\times N_o}\\), \\(\\mathbf{F}_i\\) is flattened to \\(\\mathbb{R}^{C \\times hw}\\). Images contain rich interaction semantics, which could serve as clues to unearth the interaction intention within the geometries. In detail, tokens \\(\\mathbf{T}_o\\), \\(\\mathbf{T}_h \\in \\mathbb{R}^{C \\times 1}\\) are generated to represent intention features inherent in geometries. \\(\\mathbf{T}_o\\), \\(\\mathbf{T}_h\\) are concatenated with \\(\\mathbf{F}_o\\) and \\(\\mathbf{F}_h\\) respectively to get the feature sequences \\(\\mathbf{F}_{to} \\in \\mathbb{R}^{C \\times (N_o + 1)}\\), \\(\\mathbf{F}_{th} \\in \\mathbb{R}^{C \\times (N_h + 1)}\\). Taking \\(\\mathbf{F}_i\\) as the shared key and value, \\(\\mathbf{F}_{to}\\), \\(\\mathbf{F}_{th}\\) as queries \\((q_1, q_2)\\) in two branches, the multi-branch attention is employed to model the intention features, expressed as \\(\\bar{\\mathbf{F}}_{to}, \\bar{\\mathbf{F}}_{th} = f_{\\delta}([\\mathbf{F}_{to}, \\mathbf{F}_{th}], \\mathbf{F}_i)\\). Where \\(f_{\\delta}\\) indicates the multi-branch attention layer [28, 54, 84], \\([\\cdot]\\) denotes two branches with different queries. Human and object geometries possess multiple interaction possibilities, which may introduce semantic ambiguity. To mitigate this, we further constrain the consistency of semantic tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.832, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\varphi = \\frac {\\bar {\\mathbf {T}} _ {o} \\cdot \\bar {\\mathbf {T}} _ {h}}{\\| \\bar {\\mathbf {T}} _ {o} \\| _ {2} \\times \\| \\bar {\\mathbf {T}} _ {h} \\| _ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\in \\mathbb{R}^{C\\times 1}\\) are split from \\(\\bar{\\mathbf{F}}_{to},\\bar{\\mathbf{F}}_{th}\\), representing the semantic intention features of geometries. \\(\\varphi\\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.085, + 0.852, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Figure 3. LEMON pipeline. Initially, it takes modality-specific backbones to extract respective features \\(\\mathbf{F}_h\\), \\(\\mathbf{F}_o\\), \\(\\mathbf{F}_i\\), which are then utilized to excavate intention features \\((\\bar{\\mathbf{T}}_o, \\bar{\\mathbf{T}}_h)\\) of the interaction (Sec. 3.1). With \\(\\bar{\\mathbf{T}}_o\\), \\(\\bar{\\mathbf{T}}_h\\) as conditions, LEMON integrates curvatures \\((C_o, C_h)\\) to model geometric correlations and reveal the contact \\(\\phi_c\\), affordance \\(\\phi_a\\) features (Sec. 3.2). Following, the \\(\\phi_c\\) is injected into the calculation of the object spatial feature \\(\\phi_p\\) (Sec. 3.3). Eventually, the decoder projects \\(\\phi_c\\), \\(\\phi_a\\), \\(\\phi_p\\) to the final outputs \\(\\bar{\\phi}_c\\), \\(\\bar{\\phi}_a\\), \\(\\bar{\\phi}_p\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.486, + 0.415, + 0.502 + ], + "angle": 0, + "content": "semantic cosine similarity, a part of \\(\\mathcal{L}_s\\) in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.51, + 0.431, + 0.527 + ], + "angle": 0, + "content": "3.2. Curvature-guided Geometric Correlation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.473, + 0.776 + ], + "angle": 0, + "content": "The interacting humans and objects exhibit certain geometry affinity [77], manifesting in matching geometric structures and correlative curvatures. For geometric curvatures, the normal curvature could better represent local structures like interaction regions [26, 49]. Thus, we encode the normal curvatures into geometric features, and regard the \\(\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\) as conditions to capture the affinity among human and object geometries. Normal curvatures \\(C_o\\in \\mathbb{R}^{1\\times N_o}\\) \\(C_h\\in \\mathbb{R}^{1\\times N_h}\\) of object and human are obtained by local fitting method [74, 89]. To correlate the curvatures, \\(C_o,C_h\\) are encoded to high dimension \\(C_o'\\in \\mathbb{R}^{C\\times N_o},C_h'\\in \\mathbb{R}^{C\\times N_h}\\) , and the cross-attention \\(f_{m}\\) is mutually performed on them, formulated as \\(\\bar{C}_o = f_m(C_o',C_h'),\\bar{C}_h = f_m(C_h',C_o')\\) . Following, to make the curvature guide the calculation of geometric correlation, \\(\\bar{C}_o,\\bar{C}_h\\) are integrated and fused with geometric features:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.784, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {F}} _ {c o} = f \\left(\\Gamma \\left(\\bar {C} _ {o}, \\bar {\\mathbf {F}} _ {o}\\right)\\right), \\bar {\\mathbf {F}} _ {c h} = f \\left(\\Gamma \\left(\\bar {C} _ {h}, \\bar {\\mathbf {F}} _ {h}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{F}}_{co},\\bar{\\mathbf{F}}_o\\in \\mathbb{R}^{C\\times N_o}\\) and \\(\\bar{\\mathbf{F}}_{ch},\\bar{\\mathbf{F}}_h\\in \\mathbb{R}^{C\\times N_h}\\). \\(\\bar{\\mathbf{F}}_o,\\bar{\\mathbf{F}}_h\\) are split from \\(\\bar{\\mathbf{F}}_{to}\\) and \\(\\bar{\\mathbf{F}}_{th}\\), \\(\\Gamma\\) denotes the concatenation, \\(f\\) indicates convolution layers with \\(1\\times 1\\) kernel. Then, \\(\\bar{\\mathbf{T}}_o\\) and \\(\\bar{\\mathbf{T}}_h\\) are considered as conditions [41] to further screen candidate regions that match the interaction depicted in images and participate in the modeling of geometry affinity," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.892, + 0.517 + ], + "angle": 0, + "content": "revealing the affordance and contact features \\(\\phi_{a},\\phi_{c}\\) . The process could be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.531, + 0.892, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\phi_ {a} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\overline {{\\mathbf {T}}} _ {o}\\right)\\right), \\phi_ {c} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\overline {{\\mathbf {T}}} _ {h}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.894, + 0.625 + ], + "angle": 0, + "content": "where \\(\\phi_{a} \\in \\mathbb{R}^{C \\times N_{o}}\\), to calculate it, \\(\\bar{\\mathbf{F}}_{co}\\) serves as the query, and the concatenation of \\(\\bar{\\mathbf{F}}_{ch}\\) and \\(\\bar{\\mathbf{T}}_{o}\\) serves as the key and value, \\(f_{\\theta}\\) is the cross-transformer layer. \\(\\phi_{c} \\in \\mathbb{R}^{C \\times N_{h}}\\) is obtained through the analogous way." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.637, + 0.779, + 0.654 + ], + "angle": 0, + "content": "3.3. Contact-aware Spatial Relation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.814 + ], + "angle": 0, + "content": "Human-object interactions are extremely diverse, rendering the reasoning of their 3D spatial relation very challenging. Nevertheless, human contact implicitly constrains the position of the object, thereby assisting in inferring plausible human-object spatial relation. Thus, the object's center is represented as a token sequence \\(\\mathbf{T}_{sp} \\in \\mathbb{R}^{C \\times 3}\\), with introducing the contact feature as a constraint, LEMON takes the semantic intention and geometric features of the object to query the corresponding features of the human. Modeling \\(\\mathbf{T}_{sp}\\) as the object spatial feature, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.825, + 0.892, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\phi_ {p} = f _ {\\rho} \\left(\\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c o}\\right), \\overline {{\\mathbf {T}}} _ {o}, \\mathbf {T} _ {s p}\\right) + p e, \\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c h}\\right), \\overline {{\\mathbf {T}}} _ {h}, \\phi_ {c}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.855, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\phi_p\\in \\mathbb{R}^{C\\times 3}\\), is the spatial feature of the object center, \\(f_{\\rho}\\) is a cross-attention layer, \\(\\Theta\\) denotes the pooling layer, and \\(pe\\) indicates a learnable positional encoding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.085, + 0.302, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.084, + 0.508, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.084, + 0.667, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.084, + 0.865, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.894, + 0.355 + ], + "angle": 0, + "content": "Figure 4. 3DIR Dataset. (a) The quantity of images and point clouds for each object, and a data sample containing the image, mask, dense human contact annotation, 3D object with affordance annotation, and the fitted human mesh with the object proxy sphere. (b) The proportion of our contact annotations within 24 parts on SMPL [36], and distributions of contact vertices for certain HOIs. (c) The ratio of annotated affordance regions to the whole object geometries, and the distribution of this ratio for some categories. (d) Mean distances (unit: m) between annotated object centers and human pelvis joints, and directional projections of annotated centers for several objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.38, + 0.23, + 0.395 + ], + "angle": 0, + "content": "3.4. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.471, + 0.525 + ], + "angle": 0, + "content": "Eventually, \\(\\phi_{a},\\phi_{c},\\phi_{p}\\) are sent to the decoder that contains three projection heads with linear, normalization, and activation layers. Which output object affordance \\(\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}\\), human contact \\(\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}\\), and object's center position \\(\\bar{\\phi}_p\\in \\mathbb{R}^3\\) respectively. With defined radii for specific objects, we treat a sphere as the object proxy [31], representing the spatial relation in the camera coordinates of fitted humans. The overall training loss is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.531, + 0.47, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\omega_ {1} \\mathcal {L} _ {c} + \\omega_ {2} \\mathcal {L} _ {a} + \\omega_ {3} \\mathcal {L} _ {s} + \\omega_ {4} \\mathcal {L} _ {p}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.841 + ], + "angle": 0, + "content": "where \\(\\omega_{1-4}\\) are hyper-parameters to balance the losses. \\(\\mathcal{L}_c\\) and \\(\\mathcal{L}_a\\) possess the same formulation, a focal loss [34] combined with a dice loss [42]. \\(\\mathcal{L}_c\\) is calculated by \\(\\bar{\\phi}_c\\) and the contact label \\(\\hat{\\phi}_c\\), while \\(\\mathcal{L}_a\\) is calculated by \\(\\bar{\\phi}_a\\) and the affordance annotation \\(\\hat{\\phi}_a\\). \\(\\mathcal{L}_s\\) is employed to ensure \\(\\bar{\\mathbf{T}}_o\\), \\(\\bar{\\mathbf{T}}_h\\) align with interaction semantics of the image and maintain consistency in the semantic space. Specifically, \\(\\mathbf{F}_i\\) is mapped to a logit \\(y\\), which is used to calculate a cross-entropy loss \\(\\mathcal{L}_{ce}\\) with interaction categorical label \\(\\hat{y}\\), \\(\\mathcal{L}_s\\) is the sum of \\(\\mathcal{L}_{ce}\\) and \\(\\varphi\\) (Eq. 1). For 3D HOIs, object positions are diverse due to variations in human orientation, relying solely on absolute coordinates as supervision makes it hard for the model to learn a consistent mapping. However, the relative distances between humans and objects are similar for specific HOIs. This smoother distance distribution contributes to reducing the optimization space to an approximate sphere for specific HOIs. Consequently, we take the distance between the object center and the human pelvis joint as an additional constraint, and \\(\\mathcal{L}_p\\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.845, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\mathcal {L} _ {p a} + \\mathcal {L} _ {p r}, \\quad \\mathcal {L} _ {p a} = | | \\bar {\\phi} _ {p} - \\hat {\\phi} _ {p} | | _ {2}, \\mathcal {L} _ {p r} = | | \\bar {\\xi} - \\hat {\\xi} | | _ {2}, (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.867, + 0.469, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{\\phi}_p\\) is the annotation of the object center. \\(\\bar{\\xi},\\hat{\\xi}\\) denotes distances between the pelvis joint and \\(\\bar{\\phi}_p,\\hat{\\phi}_p\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.38, + 0.589, + 0.395 + ], + "angle": 0, + "content": "4. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.454 + ], + "angle": 0, + "content": "We introduce the collection and annotation protocols of the 3DIR and give some statistical analysis of the collected data and annotations, shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.458, + 0.893, + 0.61 + ], + "angle": 0, + "content": "Collection. We collect images with explicit interaction contents, in which humans interact with specific objects. These images adhere to the condition that the human's upper body is present, ensuring the efficient recovery of the human mesh. In total, we collect 5k in-the-wild images from HAKE [32], V-COCO [18], PIAD [78], and websites with free licenses, spanning 21 object classes and 17 interaction categories. Additionally, we collect over 5k 3D object instances from several 3D datasets [9, 10, 35, 43], based on the category of objects in collected images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Annotation. We make over 25k annotations with multiple types for the collected data. 1) Masks: With the assistance of the SAM [25], we manually mask the interacting human and object in the image. 2) Human Mesh: In natural images, some humans only exhibit the upper body. Thus, we utilize the pipeline of UBody [33] to fit SMPLH pseudo-GTs. The 2D body and hand joints needed by the pipeline are obtained through the DWPose [79]. 3) Contact: Similar to DAMON [64] and HOT [6], the human contact is annotated based on human knowledge. For each image, we clearly provide masks and a specific interaction type for the human-object pair, professional annotators are hired to \"draw\" vertices on the human template that contact with objects. The entire annotation process cycles three rounds, with each round incorporating subjective cross-checks and objective metric checks. 4) Affordance: We refer to the 3D-AffordanceNet [10] to annotate the object affordance. The annotations of 11 objects included in 3D-AffordanceNet are directly utilized, and we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16288" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison on the 3DIR. Evaluation metrics of comparison methods on the benchmark, the best results are covered with the mask. indicates the improvement relative to the baseline. P. means take PointNet++ [53] as the backbone, and D. means DGCNN [66]." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.13, + 0.897, + 0.257 + ], + "angle": 0, + "content": "
Human ContactObject AffordanceSpatial Relation
MethodsPrecision ↑Recall ↑F1 ↑geo. (cm) ↓MethodsAUC ↑aIOU ↑SIM ↑MethodsMSE ↓
Baseline0.490.520.4932.83Baseline82.3632.630.50Baseline0.051
BSTRO [21]0.570.580.5528.583DA. [10]85.493.8%0.56DJ-RN [31]0.042
DECO [64]0.700.720.6915.25IAG [78]86.635.1%0.59PopUp [51]0.027
Ours P.0.760.810.779.02Ours P.87.9140.970.63Ours P.0.012
Ours D.0.780.820.787.55Ours D.88.5141.340.64Ours D.0.010
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.47, + 0.446 + ], + "angle": 0, + "content": "An additional 10 object categories excluded in 3D-AffordanceNet. 5) Spatial Relation: For each sample, we color the fitted human mesh with annotated per-vertex human contact and treat a sphere as the object proxy. The radius of the sphere for each object category is pre-defined, referring to DJ-RN [31]. Given the posed human in color and the proxy sphere, annotators adjust the sphere relative to the human to align with the human-object spatial relation depicted in the image. Ultimately, we record the center coordinates of the adjusted sphere. Due to the page limitation, please refer to the Sup. Mat. for more annotation details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.471, + 0.719 + ], + "angle": 0, + "content": "Statistical Analysis. The quantity of images and 3D instances for each object category in the 3DIR is shown in Fig. 4 (a). For contact annotation, we count its proportion within 24 human parts defined on the SMPL [36]. Moreover, we visualize the distribution of contact annotation for several HOIs, where the deeper color indicates more contact annotations at the vertex, as shown in Fig. 4 (b). Fig. 4 (c) demonstrates the proportion of annotated affordance regions to the entire object geometry, encompassing the mean of each category and the distribution of all instances for several categories. As can be seen, there are differences in distinct objects with the same affordance, as well as variations in distinct affordances of the same object. For spatial relation annotation, we count the mean distance between annotated centers and human pelvis joints, and project the spatial direction of annotated centers onto a fixed-radius sphere with the pelvis joint as the center. Fig. 4 (d) shows several cases and the distribution of mean distances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.203, + 0.75 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.263, + 0.773 + ], + "angle": 0, + "content": "5.1. Benchmark Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We refer to methods that anticipate each interaction element [21, 47, 64, 78] to thoroughly benchmark the 3DIR. For the training, in addition to the training data in 3DIR, we select another 5k data with low redundancy in BEHAVE [2]. Evaluation is conducted on the test set of 3DIR. The baseline model adopts a multitask-like framework, directly utilizing three branches to make anticipations. Besides, we compare LEMON with advanced methods that anticipate respective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.327 + ], + "angle": 0, + "content": "elements. Plus, we also conduct evaluation experiments on DAMON [64], BEHAVE [2], and PIAD [78], the results and implementation details are in the Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.333, + 0.693, + 0.35 + ], + "angle": 0, + "content": "5.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.644 + ], + "angle": 0, + "content": "The comparison results of evaluation metrics are presented in Tab. 1. The baseline model, which does not capture correlations between interaction counterparts, indicates that directly anticipating these elements through multiple branches yields poor results. Our method outperforms comparative methods across all metrics for respective elements, demonstrating that leveraging correlations between humans and objects indeed benefits the comprehension of interaction relation, and the anticipation of interaction elements. It seems to be \"the best of both worlds\". Furthermore, we conduct a visual comparison of methods with higher evaluation metrics, as shown in Fig. 5. The results showcase multiple objects interacting with distinct human parts. As can be seen, our anticipations are more precise, for some uncertain regions that are not visible in images (e.g., the first and third rows), LEMON could also anticipate plausible results. This is attributed to the modeling of correlations between geometries, which compensates for missing features in invisible regions of the image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.655, + 0.667 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct a thorough ablation study to validate the effectiveness of the model design. Tab. 2 (a) reports the model performance without modeling the semantic intention and the geometric correlation, demonstrating their impacts on the model performance. Moreover, the absence of constraining semantic consistency hinders the modeling of intentions, while not introducing curvatures affects the extraction of geometric correlations, resulting in a decrease in model performance. Both results are detailed in Tab. 2 (a). Besides, we test the impact of \\(\\phi_c\\) and \\(\\mathcal{L}_{pr}\\) (Eq. 6) on spatial relation prediction, shown in Tab. 2 (b). Notably, removing the extraction of \\(\\phi_a\\) in \\(f_{\\theta}\\) (Eq. 3) also affects the performance. This is attributed to the interrelation between \\(\\phi_c\\) and \\(\\phi_a\\), the absence of \\(\\phi_a\\) impacts the representation of \\(\\phi_c\\), and subsequently influencing the final spatial prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.082, + 0.862, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.499, + 0.893, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Visualization Results. (a) Results of the estimated human vertices in contact with objects, the estimated contact vertices are shown in yellow. (b) The anticipations of 3D object affordance, the depth of red represents the probability of anticipated affordance. (c) Two views of the predicted spatial relation, translucent spheres are object proxies. Please zoom in for a better visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.47, + 0.734 + ], + "angle": 0, + "content": "To further illustrate the effectiveness of capturing the intention and geometry affinity, we visualize attention maps on the human and object geometries when removing one of them, as shown in Fig. 6. The first row indicates without capturing intention affinity, solely relying on geometries leads the model to focus on multiple candidate regions, differing from attention regions in the image. Furthermore, the second row demonstrates that while semantic intention guides the model to identify approximate candidate regions, modeling geometry affinity further locates corresponding regions most related to the interaction of the geometries." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.283, + 0.758 + ], + "angle": 0, + "content": "5.4. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Multiple Interactions. To validate whether the model reasons interaction elements based on the understanding of interaction relation. We employ the model to infer different interactions between humans and the same object, as demonstrated in Fig. 7. The results showcase that, for the same object, when interactions are different, the model outputs distinct results, encompassing human contact, object affordance, and spatial relation, all of which align with the interaction relation depicted in the image. This indicates" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.566, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Table 2. Ablation studies. (a) Performance when not modeling intention representations (intent.), geometric correlations (geom.), and not introducing \\(\\varphi\\) (Eq. 1), curvatures (cur.). (b) The impact of several factors on spatial relation prediction. \\(\\pmb{x}\\) means without." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.633, + 0.888, + 0.768 + ], + "angle": 0, + "content": "
MetricsOursX intent.X φX geom.X cur.
Precision0.780.710.740.680.75
Recall0.820.730.790.700.78
F10.780.700.740.670.75
geo. (cm)7.5511.8710.2614.879.13
AUC88.5185.8786.6683.2386.98
aIOU41.3438.1940.0337.1339.58
SIM0.640.580.600.550.62
MSE0.0100.0270.0220.0310.018
" + }, + { + "type": "table_caption", + "bbox": [ + 0.651, + 0.77, + 0.667, + 0.78 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.892, + 0.838 + ], + "angle": 0, + "content": "that the model infers interaction elements according to the interaction relation rather than relying on direct mapping solely driven by the categories of human-object pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multiple Objects. In scenarios where distinct objects are involved in concurrent interactions with humans, the model should indeed possess the ability to comprehend the interaction relation with different objects. We give an experi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16290" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.457, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.251, + 0.47, + 0.293 + ], + "angle": 0, + "content": "Figure 6. Visual Attention. Attention heatmaps on geometries, in cases with \\((\\boldsymbol{w})\\) and without \\((\\boldsymbol{w} / \\boldsymbol{o})\\) the extraction of intention and geometry affinity. Images are also shown with attention heatmaps." + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.298, + 0.147, + 0.311 + ], + "angle": 0, + "content": "Support" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.311, + 0.149, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.297, + 0.217, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.303, + 0.292, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.296, + 0.357, + 0.305 + ], + "angle": 0, + "content": "View 1" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.308, + 0.367, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.377, + 0.296, + 0.444, + 0.305 + ], + "angle": 0, + "content": "View 2" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.308, + 0.458, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.382, + 0.149, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.425, + 0.138, + 0.435 + ], + "angle": 0, + "content": "Carry" + }, + { + "type": "image", + "bbox": [ + 0.156, + 0.37, + 0.217, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.431, + 0.223, + 0.441 + ], + "angle": 0, + "content": "Affordance" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.37, + 0.304, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.431, + 0.294, + 0.441 + ], + "angle": 0, + "content": "Contact" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.369, + 0.376, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.431, + 0.432, + 0.441 + ], + "angle": 0, + "content": "Spatial Relation" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.444, + 0.468, + 0.471 + ], + "angle": 0, + "content": "Figure 7. Multiple Interactions. Distinct anticipations when interactions are different with the same object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.621 + ], + "angle": 0, + "content": "ment pertaining to this property. As shown in Fig. 8, when reasoning with different objects, the human contact, object affordance, and spatial relation are distinct, while all anticipations are plausible. This is attributed to the collaboration of the semantic interaction intention and geometric correlation, ensuring that the model could explicitly capture interaction intentions with different objects and reveal regions related to the interaction on corresponding geometries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Multiple Instances. We conduct an experiment to validate the model's understanding of interaction relation across various instances, assessing its generalization and robustness. Specifically, distinct object instances are utilized to infer with the same image and human geometry, shown in Fig. 9 (a). As can be seen, even though the contact and spatial relation are slightly different, the results are consistent with the interaction shown in the image, and object affordances are correctly anticipated. Additionally, the same object instance is utilized to infer with different human geometries and images that contain similar human-object interactions, as demonstrated in Fig. 9 (b). These results indicate that the model could generalize the understanding of interaction relation to different human and object instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.195, + 0.86 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We propose leveraging both counterparts of the HOI to jointly anticipate human contact, object affordance, and" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.092, + 0.88, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.227, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Figure 8. Multiple Objects. The anticipations for multiple objects with different human-object interactions." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.263, + 0.88, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.461, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Figure 9. Multiple Instances. (a) Different object instances \\( w.r.t. \\) the same image and human geometry. (b) Different images and human instances \\( w.r.t. \\) the same object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.726 + ], + "angle": 0, + "content": "human-object spatial relation in 3D space. It holds the potential to drive embodied intelligence to learn from interactions and assist the interaction modeling in graphics. With the provided dataset 3DIR containing paired HOI data and multiple annotations, we train LEMON, a novel model that harnesses the interaction intention and geometric correlation between humans and objects to capture the affinity, eliminating interaction uncertainties and anticipating plausible interaction elements. Extensive experiments demonstrate the effectiveness and superiority of LEMON. We believe this work offers fresh insights and paves a new way for 3D human-object interaction understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Limitations and Future Work. In the current formulation, LEMON needs the pre-inferred human mesh as input. In the future, we plan to integrate HMR into the entire framework, freeing input constraints required for inference, and taking 3D interaction elements to improve the accuracy of HMR. Plus, another interesting direction is to leverage multi-modal methods [15, 88], boosting the interaction relation understanding from other sources e.g., text, audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62306295." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16291" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Brenna D Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469-483, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.468, + 0.24 + ], + "angle": 0, + "content": "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[3] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.468, + 0.367 + ], + "angle": 0, + "content": "[4] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018, the winter conference on applications of computer vision (wacv), pages 381–389. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.468, + 0.451 + ], + "angle": 0, + "content": "[5] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.468, + 0.507 + ], + "angle": 0, + "content": "[6] Yixin Chen, Sai Kumar Dwivedi, Michael J. Black, and Dimitrios Tzionas. Detecting human-object contact in images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.468, + 0.563 + ], + "angle": 0, + "content": "[7] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22:449-462, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[8] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.468, + 0.802 + ], + "angle": 0, + "content": "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[16] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[17] Elena Corina Grigore, Kerstin Eder, Anthony G Pipe, Chris Melhuish, and Ute Leonards. Joint action understanding improves robot-to-human object handover. In 2013 IEEE/RSJ international conference on intelligent robots and systems, pages 4622-4629. IEEE, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[18] Saurabh Gupta and Jitendra Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[19] Sookwan Han and Hanbyul Joo. Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.535, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[21] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[23] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16292" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Jan J Koenderink and Andrea J Van Doorn. Surface shape and curvature scales. Image and vision computing, 10(8): 557-564, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[27] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[28] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.362 + ], + "angle": 0, + "content": "[29] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. arXiv preprint arXiv:2309.16237, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[30] Lulei Li, Jianan Wei, Wenguan Wang, and Yi Yang. Neural-logic human-object interaction detection. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[31] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10166–10175, 2020. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[32] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[33] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.744 + ], + "angle": 0, + "content": "[35] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. AKB-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.802 + ], + "angle": 0, + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[37] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2252-2261, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[38] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Leverage interactive affinity for affordance" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6809-6819, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[39] Bertram F Malle and Joshua Knobe. The folk concept of intentionality. Journal of experimental social psychology, 33 (2):101-121, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.235 + ], + "angle": 0, + "content": "[40] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[41] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.894, + 0.374 + ], + "angle": 0, + "content": "[42] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[43] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.894, + 0.531 + ], + "angle": 0, + "content": "[44] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[45] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[46] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.894, + 0.703 + ], + "angle": 0, + "content": "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[48] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai Nguyen. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing systems, 33:7841-7851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[49] Abdul Nurunnabi, Geoff West, and David Belton. Outlier detection and robust normal-curvature estimation in mobile laser scanning 3d point cloud data. Pattern Recognition, 48 (4):1404-1419, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[51] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4726-4736, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[52] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.369 + ], + "angle": 0, + "content": "[54] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12802–12813, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[55] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.472, + 0.524 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[59] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[60] Thomas B Sheridan. Human-robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[61] Chris Shilling. The body in culture, technology and society. The Body in Culture, Technology and Society, pages 1-256, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.472, + 0.873 + ], + "angle": 0, + "content": "[62] Soshi Shimada, Vladislav Golyanik, Zhi Li, Patrick Pérez, Weipeng Xu, and Christian Theobalt. Hulc: 3d human motion capture with pose manifold sampling and dense contact guidance. In European Conference on Computer Vision, pages 516-533. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[63] Michael Tomasello, Malinda Carpenter, Josep Call, Tanya Behne, and Henrik Moll. Understanding and sharing in-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tentions: The origins of cultural cognition. Behavioral and brain sciences, 28(5):675-691, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[64] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[65] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[67] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In European Conference on Computer Vision, pages 90-107. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[68] Ping Wei, Yibiao Zhao, Nanning Zheng, and Song-Chun Zhu. Modeling 4d human-object interactions for event and object recognition. In Proceedings of the IEEE international conference on computer vision, pages 3272-3279, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[69] Ruihai Wu, Chuanruo Ning, and Hao Dong. Learning foresightful dense visual affordance for deformable object manipulation. arXiv preprint arXiv:2303.11057, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[70] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[71] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[72] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and Mohan S Kankanhalli. Learning to detect human-object interactions with knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[73] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[74] Rui Xu, Zhiyang Dou, Ningna Wang, Shiqing Xin, Shuang-min Chen, Mingyan Jiang, Xiaohu Guo, Wenping Wang, and Changhe Tu. Globally consistent normal orientation for point clouds by regularizing the winding-number field. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[75] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[76] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3d-hoi: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[77] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[78] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10905-10915, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.319 + ], + "angle": 0, + "content": "[79] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. arXiv preprint arXiv:2307.15880, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[80] Bangpeng Yao and Li Fei-Fei. Modeling mutual context of object and human pose in human-object interaction activities. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 17–24. IEEE, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[81] Sifan Ye, Yixing Wang, Jiaman Li, Dennis Park, C Karen Liu, Huazhe Xu, and Jiajun Wu. Scene synthesis from human motion. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.449, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[82] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[83] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.576, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[84] Guochen Yu, Andong Li, Chengshi Zheng, Yinuo Guo, Yutian Wang, and Hui Wang. Dual-branch attention-inattention transformer for single-channel speech enhancement. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7847-7851. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[85] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10): 2472-2500, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[86] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, pages 1-26, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[87] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 34-51. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[88] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[89] Xiaopeng Zhang, Hongjun Li, Zhanglin Cheng, and Yikuan Zhang. Robust curvature estimation and geometry analysis of 3d point cloud surfaces. J. Inf. Comput. Sci, 6(5):1983-1990, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[90] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[91] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[92] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper manipulation. In International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[93] Xinghao Zhu, JingHan Ke, Zhixuan Xu, Zhixin Sun, Bizhe Bai, Jun Lv, Qingtao Liu, Yuwei Zeng, Qi Ye, Cewu Lu, et al. Diff-lfd: Contact-aware model-based learning from visual demonstration for robotic manipulation via differentiable physics-based simulation and rendering. In 7th Annual Conference on Robot Learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.474, + 0.892, + 0.556 + ], + "angle": 0, + "content": "[94] Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11825-11834, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16295" + } + ] +] \ No newline at end of file diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_origin.pdf b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccd5ab7645eb715a57c6c68cc51e4fb96a0e1c1e --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/2d6d94d5-881a-4f31-92c0-fc2ad5248536_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:235851b20551c4a484ccd346a38864910610de2f3d6ef8dffafd7aae73fc4213 +size 9239011 diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/full.md b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d5f793d54628f4bd8a7ff5f92499830233801221 --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/full.md @@ -0,0 +1,343 @@ +# LEMON: Learning 3D Human-Object Interaction Relation from 2D Images + +Yuhang Yang $^{1}$ , Wei Zhai $^{1,\dagger}$ , Hongchen Luo $^{1}$ , Yang Cao $^{1,2}$ , Zheng-Jun Zha $^{1}$ + +1 University of Science and Technology of China + +$^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + +{yyuhang@mail., wzhai056@, lhcl2@mail., forrest@, zhazj@}ustc.edu.cn + +![](images/27b4b9f9d5af39aae9820d3c1c34bd1c7e4c90eae5e1f22245c4d2f17ce42d3b.jpg) +Figure 1. For an interaction image with paired geometries of the human and object, LEMON learns 3D human-object interaction relation by jointly anticipating the interaction elements, including human contact, object affordance, and human-object spatial relation. Vertices in yellow denote those in contact with the object, regions in red are object affordance regions, and the translucent sphere is the object proxy. + +# Abstract + +Learning 3D human-object interaction relation is pivotal to embodied AI and interaction modeling. Most existing methods approach the goal by learning to predict isolated interaction elements, e.g., human contact, object affordance, and human-object spatial relation, primarily from the perspective of either the human or the object. Which underexploit certain correlations between the interaction counterparts (human and object), and struggle to address the uncertainty in interactions. Actually, objects' functionalities potentially affect humans' interaction intentions, which reveals what the interaction is. Meanwhile, the interacting humans and objects exhibit matching geometric structures, which presents how to interact. In light of this, we propose harnessing these inherent correlations between interaction counterparts to mitigate the uncertainty and jointly anticipate the above interaction elements in 3D space. To achieve this, we present LEMON (LEarning 3D huMan-Object iNteraction relation), a unified model that mines interaction intentions of the counterparts and employs curvatures to guide the extraction of geometric correlations, combining them to anticipate the interaction elements. Besides, the 3D Interaction Relation dataset (3DIR) is collected to serve as the test bed for training and evaluation. Extensive experiments demonstrate the + +superiority of LEMON over methods estimating each element in isolation. The code and dataset are available at https://yyvhang.github.io/LEMON. + +# 1. Introduction + +Learning 3D human-object interaction (HOI) relation seeks to capture semantic co-occurrence and geometric compatibility between humans and objects in 3D space [5, 68, 80]. How can machines learn the interaction relation? One possible solution involves perceiving certain elements capable of revealing the interaction. Contact [6, 21, 64, 70], affordance [10, 13, 46, 92], and spatial relation [19, 31, 52] that elucidate "where" the interaction manifests between the human and object garner great attention. Capturing representations of such elements is pivotal for applications like AR/VR [7], imitation learning [1, 24], embodied AI [17, 58, 60], and interaction modeling [20, 81]. + +Humans predominantly manipulate and interact with objects in 3D space. Thus, many methods devise task-specific models to anticipate certain interaction elements, thereby perceiving 3D HOI relation. Methods [21, 62, 64] estimate dense human contact based on the interaction semantics depicted in images. Some studies anticipate the object affordance according to objects' structures or 2D visuals [10, 44, 67, 78]. Several works delve into predicting the human-object spatial relation through synthetic images [19] or posed human geometries [52]. A prevailing trend in + +these methods involves taking certain attributes (e.g., appearances, geometries) of either humans or objects to predict an isolated interaction element, capturing one aspect of the interaction relation. However, they overlook the impact of factors mutually determined by humans and objects on interactions, such as the interaction intention and geometric correlation, which leads to struggles in addressing the uncertainty within interactions. Specifically, the uncertainty could be attributed to two principal aspects. On the one hand, the diversity of HOIs introduces challenges in capturing explicit interaction intentions solely centering on humans or objects, raising the intention uncertainty. On the other hand, the limited view and mutual occlusions give rise to invisible interaction regions in images, complicating the constitution of linkages between these regions and target 3D geometries, causing the region uncertainty. These uncertainties may culminate in ambiguous anticipations. + +In this paper, we propose leveraging both counterparts of the interaction to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space (Fig. 1), addressing the uncertainty by unearthing the correlation between the interacting humans and objects. Actually, humans and objects are intertwined and possess affinities in the interaction (Fig. 2). In specific, the design of objects typically adheres to certain human needs. Therefore, object affordances inherently hint at "what" interactions humans intend to make [39, 61, 63], revealing the intention affinity of interactions. Meanwhile, the interacting human and object exhibit matching geometries (either posture or configuration), which presents "how" to interact, arising the geometry affinity. The intention affinity clarifies the interaction type and implicates the interaction regions. Geometry affinity could serve as pivotal clues for excavating correlations between geometries corresponding to invisible regions in the image. These interaction-related regions e.g., contact regions, further reflect the human-object spatial relation. + +To achieve this, we present the LEMON, a novel framework that correlates the intention semantics and geometric correspondences to jointly anticipate human contact, object affordance, and human-object spatial relation, in 3D space. To capture the intention affinity, LEMON employs multibranch attention to model the correlation between the interaction content in images and geometries of humans and objects, revealing intention representations of the interaction corresponding to geometries. The cosine similarity is utilized to further ensure their semantic consistency. Taking intention representations as conditions, LEMON integrates geometric curvatures to capture the geometry affinity and reveal the human contact and object affordance representations. These representations then assist in anticipating the spatial relation constrained by a combined distance loss. + +In addition, we collect the 3DIR dataset, which contains natural HOI images paired with object point clouds + +![](images/f0e94512c6a1f29f5d79eee91ab59b5bae98125bde8267268f10c74dfaa2b57c.jpg) +Figure 2. Motivation. Affinities within the HOI. The object affordance inherently reveals the human's interaction intention, arising the intention affinity. The interacting human and object possess matching structures, exhibiting the geometry affinity. + +and SMPL-H [57] pseudo-GTs. Multiple annotations are made for these data, e.g., dense human contact, object affordance, and human-object spatial relation. It serves as the test bed for the model training and evaluation. + +The contributions are summarized as follows: + +1) We thoroughly exploit the correlation between the interaction counterparts to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space. It furnishes essential interaction elements to comprehend 3D HOIs. +2) We present the LEMON, a novel framework that correlates semantic interaction intention and geometries to capture the affinity between humans and objects, eliminating the impact of the interaction uncertainty and anticipating plausible 3D interaction elements. +3) We provide the 3DIR dataset that contains paired HOI data and multiple annotations, including dense human contact, object affordance, and human-object spatial relation, to support the anticipation. Extensive experiments demonstrate the superiority of LEMON. + +# 2. Related Work + +# 2.1. Human-Object Interaction Relation in 2D + +Various approaches perceive elements pertinent to human-object interaction relation from multiple perspectives in the 2D domain. Methods [4, 16, 30, 72, 94] represent human-object interaction as triplets $<$ human, action, object>. Grounding bounding boxes of interaction counterparts and actions corresponding to them. They output the instance-level human-object spatial relation in pixel space and action semantics. Delving into the object side, the concept of "interaction possibility" is encapsulated in the term affordance [14], and bountiful methods explore to anticipate the object affordance, whether at instance-level or part-level + +[11, 12, 37, 38, 47, 85]. Turning to the human side, the contact explicitly indicates where the human interacts with objects. Some works detect the contact for specific parts (e.g., hand and foot) at a box-region level [48, 55, 59]. HOT [6] provides 2D contact annotations to support human-object contact estimation in images. These efforts achieve substantial results in understanding the human-object interaction relation in 2D space. However, they still encounter challenges when extrapolating to 3D space for practical applications due to the lack of a dimension. + +# 2.2. Human-Object Interaction Relation in 3D + +To facilitate the incorporation of interaction comprehension into applications, extensive works explore to anticipate interaction elements in 3D space. For object affordance, methods [10, 44, 67, 69, 73, 86, 91] constitute the linkage between object shapes and affordances by a learning-based mapping, approaching the matter from the perspective of objects' functionality and structure. Some approaches learn object affordances from a distinct perspective, based on interactions. Whether making agents actively interact with 3D synthetic scenarios [46], or ground 3D object affordances through object-object [45] and 2D [78] interactions. + +The dense human contact estimation is primarily based on the SMPL series [36, 50, 57]. Numerous methods model human contact for various tasks [8, 20, 22, 23, 71, 82, 83]. HULC [62], BSTRO [21], and DECO [64] are more similar to the estimation of contact in our method. They learn a mapping from the interaction semantics in images to the human vertices sequence, hardly leveraging geometries. Different from them, our method harnesses semantic and geometric correlations of interaction counterparts to address the uncertainty and infer the vertices that contact with objects. DECO also contributes the DAMON [64] dataset that possesses dense 3D contact annotation for in-the-wild images. In addition to human contact annotations, 3DIR includes other annotations related to the interaction, e.g., 3D object affordance and human-object spatial relation. + +Bounding boxes obtained by HOI detection methods give the human-object spatial relation in pixel space. To lift this relation to 3D space, DJ-RN [31] takes hollow spheres with defined radii to represent objects, and projects the spatial relation in images to 3D space based on the bounding boxes and defined radii. CHORUS [19] utilizes synthetic multi-view images generated by the diffusion model [56] to learn the spatial relation between objects and canonical humans. Given a posed human, Object Pop-Up [51] anticipates the objects and their spatial positions to match the human body for certain interactions. + +The above methods commonly focus on one side of the interaction in isolation. In contrast, our method captures the inherent affinity between both sides of the interaction to jointly anticipate interaction elements in 3D space. Such + +anticipations are beneficial for tasks e.g., robot manipulation [3, 40, 93], interaction generation [27, 29, 75, 81, 90] and reconstruction [20, 71, 76, 87]. + +# 3. Method + +Given the inputs $\{H,O,I\}$ , where $H\in \mathbb{R}^{N_h\times 3}$ indicates vertices of the SMPL-H [57], which represents the human as pose $\theta \in \mathbb{R}^{52\times 3}$ , shape $\beta \in \mathbb{R}^{10}$ parameters, and output a mesh $M(\theta ,\beta)\in \mathbb{R}^{6890\times 3}$ . $O\in \mathbb{R}^{N_o\times 3}$ is an object point cloud, $I\in \mathbb{R}^{H\times W\times 3}$ is an image. $N_{h},N_{o}$ are the number of points, $H,W$ are image's height and width. LEMON jointly anticipates the human contact $\bar{\phi}_c\in \mathbb{R}^{N_h\times 1}$ , object affordance $\bar{\phi}_a\in \mathbb{R}^{N_o\times 1}$ , and object center position $\bar{\phi}_p\in \mathbb{R}^3$ , in 3D space. As shown in Fig. 3, initially, the inputs are sent to image and point cloud backbones, obtain respective features $\mathbf{F}_h,\mathbf{F}_o,\mathbf{F}_i$ . Then, LEMON correlates $\mathbf{F}_i$ with $\mathbf{F}_o$ , $\mathbf{F}_h$ to extract features of the interaction intention ( $\bar{\mathbf{T}}_o,\bar{\mathbf{T}}_h$ ) inherent in geometries through multi-branch attention, and employs cosine similarity to constrain their semantic consistency (Sec. 3.1). With $\bar{\mathbf{T}}_o$ , $\bar{\mathbf{T}}_h$ as conditions, LEMON integrates curvatures to guide the modeling of geometric correlations, capturing the geometry affinity and revealing the contact $\phi_c$ and affordance $\phi_a$ features (Sec. 3.2). Next, geometric features and $\phi_c,\bar{\mathbf{T}}_o,\bar{\mathbf{T}}_h$ are utilized to model the object spatial feature $\phi_p$ (Sec. 3.3). Eventually, $\phi_c,\phi_a,\phi_p$ are projected to $\bar{\phi}_c,\bar{\phi}_a$ , and $\bar{\phi}_p$ in the decoder, the whole process is optimized by a combined loss (Sec. 3.4). + +# 3.1. Interaction Intention Excavation + +We add human and object masks on $I$ , and utilize the HR-Net [65] and DGCNN [66] as backbones to extract the image feature $\mathbf{F}_i \in \mathbb{R}^{C \times h \times w}$ , the geometric features of human $\mathbf{F}_h \in \mathbb{R}^{C \times N_h}$ , and object $\mathbf{F}_o \in \mathbb{R}^{C \times N_o}$ , $\mathbf{F}_i$ is flattened to $\mathbb{R}^{C \times hw}$ . Images contain rich interaction semantics, which could serve as clues to unearth the interaction intention within the geometries. In detail, tokens $\mathbf{T}_o$ , $\mathbf{T}_h \in \mathbb{R}^{C \times 1}$ are generated to represent intention features inherent in geometries. $\mathbf{T}_o$ , $\mathbf{T}_h$ are concatenated with $\mathbf{F}_o$ and $\mathbf{F}_h$ respectively to get the feature sequences $\mathbf{F}_{to} \in \mathbb{R}^{C \times (N_o + 1)}$ , $\mathbf{F}_{th} \in \mathbb{R}^{C \times (N_h + 1)}$ . Taking $\mathbf{F}_i$ as the shared key and value, $\mathbf{F}_{to}$ , $\mathbf{F}_{th}$ as queries $(q_1, q_2)$ in two branches, the multi-branch attention is employed to model the intention features, expressed as $\bar{\mathbf{F}}_{to}, \bar{\mathbf{F}}_{th} = f_{\delta}([\mathbf{F}_{to}, \mathbf{F}_{th}], \mathbf{F}_i)$ . Where $f_{\delta}$ indicates the multi-branch attention layer [28, 54, 84], $[\cdot]$ denotes two branches with different queries. Human and object geometries possess multiple interaction possibilities, which may introduce semantic ambiguity. To mitigate this, we further constrain the consistency of semantic tokens: + +$$ +\varphi = \frac {\bar {\mathbf {T}} _ {o} \cdot \bar {\mathbf {T}} _ {h}}{\| \bar {\mathbf {T}} _ {o} \| _ {2} \times \| \bar {\mathbf {T}} _ {h} \| _ {2}}, \tag {1} +$$ + +where $\bar{\mathbf{T}}_o,\bar{\mathbf{T}}_h\in \mathbb{R}^{C\times 1}$ are split from $\bar{\mathbf{F}}_{to},\bar{\mathbf{F}}_{th}$ , representing the semantic intention features of geometries. $\varphi$ is the + +![](images/e0c94c7a1bed6d704787b1ae2ae4395db9ed8d68366ca629776ea01f23dcf095.jpg) +Figure 3. LEMON pipeline. Initially, it takes modality-specific backbones to extract respective features $\mathbf{F}_h$ , $\mathbf{F}_o$ , $\mathbf{F}_i$ , which are then utilized to excavate intention features $(\bar{\mathbf{T}}_o, \bar{\mathbf{T}}_h)$ of the interaction (Sec. 3.1). With $\bar{\mathbf{T}}_o$ , $\bar{\mathbf{T}}_h$ as conditions, LEMON integrates curvatures $(C_o, C_h)$ to model geometric correlations and reveal the contact $\phi_c$ , affordance $\phi_a$ features (Sec. 3.2). Following, the $\phi_c$ is injected into the calculation of the object spatial feature $\phi_p$ (Sec. 3.3). Eventually, the decoder projects $\phi_c$ , $\phi_a$ , $\phi_p$ to the final outputs $\bar{\phi}_c$ , $\bar{\phi}_a$ , $\bar{\phi}_p$ . + +semantic cosine similarity, a part of $\mathcal{L}_s$ in Sec. 3.4. + +# 3.2. Curvature-guided Geometric Correlation + +The interacting humans and objects exhibit certain geometry affinity [77], manifesting in matching geometric structures and correlative curvatures. For geometric curvatures, the normal curvature could better represent local structures like interaction regions [26, 49]. Thus, we encode the normal curvatures into geometric features, and regard the $\bar{\mathbf{T}}_o,\bar{\mathbf{T}}_h$ as conditions to capture the affinity among human and object geometries. Normal curvatures $C_o\in \mathbb{R}^{1\times N_o}$ $C_h\in \mathbb{R}^{1\times N_h}$ of object and human are obtained by local fitting method [74, 89]. To correlate the curvatures, $C_o,C_h$ are encoded to high dimension $C_o'\in \mathbb{R}^{C\times N_o},C_h'\in \mathbb{R}^{C\times N_h}$ , and the cross-attention $f_{m}$ is mutually performed on them, formulated as $\bar{C}_o = f_m(C_o',C_h'),\bar{C}_h = f_m(C_h',C_o')$ . Following, to make the curvature guide the calculation of geometric correlation, $\bar{C}_o,\bar{C}_h$ are integrated and fused with geometric features: + +$$ +\bar {\mathbf {F}} _ {c o} = f \left(\Gamma \left(\bar {C} _ {o}, \bar {\mathbf {F}} _ {o}\right)\right), \bar {\mathbf {F}} _ {c h} = f \left(\Gamma \left(\bar {C} _ {h}, \bar {\mathbf {F}} _ {h}\right)\right), \tag {2} +$$ + +where $\bar{\mathbf{F}}_{co},\bar{\mathbf{F}}_o\in \mathbb{R}^{C\times N_o}$ and $\bar{\mathbf{F}}_{ch},\bar{\mathbf{F}}_h\in \mathbb{R}^{C\times N_h}$ . $\bar{\mathbf{F}}_o,\bar{\mathbf{F}}_h$ are split from $\bar{\mathbf{F}}_{to}$ and $\bar{\mathbf{F}}_{th}$ , $\Gamma$ denotes the concatenation, $f$ indicates convolution layers with $1\times 1$ kernel. Then, $\bar{\mathbf{T}}_o$ and $\bar{\mathbf{T}}_h$ are considered as conditions [41] to further screen candidate regions that match the interaction depicted in images and participate in the modeling of geometry affinity, + +revealing the affordance and contact features $\phi_{a},\phi_{c}$ . The process could be formulated as: + +$$ +\phi_ {a} = f _ {\theta} \left(\overline {{\mathbf {F}}} _ {c o}, \Gamma \left(\overline {{\mathbf {F}}} _ {c h}, \overline {{\mathbf {T}}} _ {o}\right)\right), \phi_ {c} = f _ {\theta} \left(\overline {{\mathbf {F}}} _ {c h}, \Gamma \left(\overline {{\mathbf {F}}} _ {c o}, \overline {{\mathbf {T}}} _ {h}\right)\right), \tag {3} +$$ + +where $\phi_{a} \in \mathbb{R}^{C \times N_{o}}$ , to calculate it, $\bar{\mathbf{F}}_{co}$ serves as the query, and the concatenation of $\bar{\mathbf{F}}_{ch}$ and $\bar{\mathbf{T}}_{o}$ serves as the key and value, $f_{\theta}$ is the cross-transformer layer. $\phi_{c} \in \mathbb{R}^{C \times N_{h}}$ is obtained through the analogous way. + +# 3.3. Contact-aware Spatial Relation + +Human-object interactions are extremely diverse, rendering the reasoning of their 3D spatial relation very challenging. Nevertheless, human contact implicitly constrains the position of the object, thereby assisting in inferring plausible human-object spatial relation. Thus, the object's center is represented as a token sequence $\mathbf{T}_{sp} \in \mathbb{R}^{C \times 3}$ , with introducing the contact feature as a constraint, LEMON takes the semantic intention and geometric features of the object to query the corresponding features of the human. Modeling $\mathbf{T}_{sp}$ as the object spatial feature, formulated as: + +$$ +\phi_ {p} = f _ {\rho} \left(\Gamma \left(\Theta \left(\overline {{\mathbf {F}}} _ {c o}\right), \overline {{\mathbf {T}}} _ {o}, \mathbf {T} _ {s p}\right) + p e, \Gamma \left(\Theta \left(\overline {{\mathbf {F}}} _ {c h}\right), \overline {{\mathbf {T}}} _ {h}, \phi_ {c}\right)\right), \tag {4} +$$ + +where $\phi_p\in \mathbb{R}^{C\times 3}$ , is the spatial feature of the object center, $f_{\rho}$ is a cross-attention layer, $\Theta$ denotes the pooling layer, and $pe$ indicates a learnable positional encoding. + +![](images/47a46c1086f386cf8589b776ab8b297149c6bb6c22bb525240bab82ad9c49e48.jpg) +Figure 4. 3DIR Dataset. (a) The quantity of images and point clouds for each object, and a data sample containing the image, mask, dense human contact annotation, 3D object with affordance annotation, and the fitted human mesh with the object proxy sphere. (b) The proportion of our contact annotations within 24 parts on SMPL [36], and distributions of contact vertices for certain HOIs. (c) The ratio of annotated affordance regions to the whole object geometries, and the distribution of this ratio for some categories. (d) Mean distances (unit: m) between annotated object centers and human pelvis joints, and directional projections of annotated centers for several objects. + +![](images/a54ab6ed10b8e2318b618eee84b8bb53b04c323bf078cb2cae034e1c6efc213c.jpg) + +![](images/5b95b3c666aa345b8c03b13e593015b93624bf09d5e4bc8ee9702bf4bdf333ef.jpg) + +![](images/36557fb43acd6489199a9bb243d6764f12300f27c7d334ae8a9329bfd3f6c374.jpg) + +# 3.4. Loss Functions + +Eventually, $\phi_{a},\phi_{c},\phi_{p}$ are sent to the decoder that contains three projection heads with linear, normalization, and activation layers. Which output object affordance $\bar{\phi}_a\in \mathbb{R}^{N_o\times 1}$ , human contact $\bar{\phi}_c\in \mathbb{R}^{N_h\times 1}$ , and object's center position $\bar{\phi}_p\in \mathbb{R}^3$ respectively. With defined radii for specific objects, we treat a sphere as the object proxy [31], representing the spatial relation in the camera coordinates of fitted humans. The overall training loss is expressed as: + +$$ +\mathcal {L} = \omega_ {1} \mathcal {L} _ {c} + \omega_ {2} \mathcal {L} _ {a} + \omega_ {3} \mathcal {L} _ {s} + \omega_ {4} \mathcal {L} _ {p}, \tag {5} +$$ + +where $\omega_{1-4}$ are hyper-parameters to balance the losses. $\mathcal{L}_c$ and $\mathcal{L}_a$ possess the same formulation, a focal loss [34] combined with a dice loss [42]. $\mathcal{L}_c$ is calculated by $\bar{\phi}_c$ and the contact label $\hat{\phi}_c$ , while $\mathcal{L}_a$ is calculated by $\bar{\phi}_a$ and the affordance annotation $\hat{\phi}_a$ . $\mathcal{L}_s$ is employed to ensure $\bar{\mathbf{T}}_o$ , $\bar{\mathbf{T}}_h$ align with interaction semantics of the image and maintain consistency in the semantic space. Specifically, $\mathbf{F}_i$ is mapped to a logit $y$ , which is used to calculate a cross-entropy loss $\mathcal{L}_{ce}$ with interaction categorical label $\hat{y}$ , $\mathcal{L}_s$ is the sum of $\mathcal{L}_{ce}$ and $\varphi$ (Eq. 1). For 3D HOIs, object positions are diverse due to variations in human orientation, relying solely on absolute coordinates as supervision makes it hard for the model to learn a consistent mapping. However, the relative distances between humans and objects are similar for specific HOIs. This smoother distance distribution contributes to reducing the optimization space to an approximate sphere for specific HOIs. Consequently, we take the distance between the object center and the human pelvis joint as an additional constraint, and $\mathcal{L}_p$ is formulated as: + +$$ +\mathcal {L} _ {p} = \mathcal {L} _ {p a} + \mathcal {L} _ {p r}, \quad \mathcal {L} _ {p a} = | | \bar {\phi} _ {p} - \hat {\phi} _ {p} | | _ {2}, \mathcal {L} _ {p r} = | | \bar {\xi} - \hat {\xi} | | _ {2}, (6) +$$ + +where $\hat{\phi}_p$ is the annotation of the object center. $\bar{\xi},\hat{\xi}$ denotes distances between the pelvis joint and $\bar{\phi}_p,\hat{\phi}_p$ respectively. + +# 4. Dataset + +We introduce the collection and annotation protocols of the 3DIR and give some statistical analysis of the collected data and annotations, shown in Fig. 4. + +Collection. We collect images with explicit interaction contents, in which humans interact with specific objects. These images adhere to the condition that the human's upper body is present, ensuring the efficient recovery of the human mesh. In total, we collect 5k in-the-wild images from HAKE [32], V-COCO [18], PIAD [78], and websites with free licenses, spanning 21 object classes and 17 interaction categories. Additionally, we collect over 5k 3D object instances from several 3D datasets [9, 10, 35, 43], based on the category of objects in collected images. + +Annotation. We make over 25k annotations with multiple types for the collected data. 1) Masks: With the assistance of the SAM [25], we manually mask the interacting human and object in the image. 2) Human Mesh: In natural images, some humans only exhibit the upper body. Thus, we utilize the pipeline of UBody [33] to fit SMPLH pseudo-GTs. The 2D body and hand joints needed by the pipeline are obtained through the DWPose [79]. 3) Contact: Similar to DAMON [64] and HOT [6], the human contact is annotated based on human knowledge. For each image, we clearly provide masks and a specific interaction type for the human-object pair, professional annotators are hired to "draw" vertices on the human template that contact with objects. The entire annotation process cycles three rounds, with each round incorporating subjective cross-checks and objective metric checks. 4) Affordance: We refer to the 3D-AffordanceNet [10] to annotate the object affordance. The annotations of 11 objects included in 3D-AffordanceNet are directly utilized, and we + +Table 1. Comparison on the 3DIR. Evaluation metrics of comparison methods on the benchmark, the best results are covered with the mask. indicates the improvement relative to the baseline. P. means take PointNet++ [53] as the backbone, and D. means DGCNN [66]. + +
Human ContactObject AffordanceSpatial Relation
MethodsPrecision ↑Recall ↑F1 ↑geo. (cm) ↓MethodsAUC ↑aIOU ↑SIM ↑MethodsMSE ↓
Baseline0.490.520.4932.83Baseline82.3632.630.50Baseline0.051
BSTRO [21]0.570.580.5528.583DA. [10]85.493.8%0.56DJ-RN [31]0.042
DECO [64]0.700.720.6915.25IAG [78]86.635.1%0.59PopUp [51]0.027
Ours P.0.760.810.779.02Ours P.87.9140.970.63Ours P.0.012
Ours D.0.780.820.787.55Ours D.88.5141.340.64Ours D.0.010
+ +An additional 10 object categories excluded in 3D-AffordanceNet. 5) Spatial Relation: For each sample, we color the fitted human mesh with annotated per-vertex human contact and treat a sphere as the object proxy. The radius of the sphere for each object category is pre-defined, referring to DJ-RN [31]. Given the posed human in color and the proxy sphere, annotators adjust the sphere relative to the human to align with the human-object spatial relation depicted in the image. Ultimately, we record the center coordinates of the adjusted sphere. Due to the page limitation, please refer to the Sup. Mat. for more annotation details. + +Statistical Analysis. The quantity of images and 3D instances for each object category in the 3DIR is shown in Fig. 4 (a). For contact annotation, we count its proportion within 24 human parts defined on the SMPL [36]. Moreover, we visualize the distribution of contact annotation for several HOIs, where the deeper color indicates more contact annotations at the vertex, as shown in Fig. 4 (b). Fig. 4 (c) demonstrates the proportion of annotated affordance regions to the entire object geometry, encompassing the mean of each category and the distribution of all instances for several categories. As can be seen, there are differences in distinct objects with the same affordance, as well as variations in distinct affordances of the same object. For spatial relation annotation, we count the mean distance between annotated centers and human pelvis joints, and project the spatial direction of annotated centers onto a fixed-radius sphere with the pelvis joint as the center. Fig. 4 (d) shows several cases and the distribution of mean distances. + +# 5. Experiment + +# 5.1. Benchmark Setting + +We refer to methods that anticipate each interaction element [21, 47, 64, 78] to thoroughly benchmark the 3DIR. For the training, in addition to the training data in 3DIR, we select another 5k data with low redundancy in BEHAVE [2]. Evaluation is conducted on the test set of 3DIR. The baseline model adopts a multitask-like framework, directly utilizing three branches to make anticipations. Besides, we compare LEMON with advanced methods that anticipate respective + +elements. Plus, we also conduct evaluation experiments on DAMON [64], BEHAVE [2], and PIAD [78], the results and implementation details are in the Sup. Mat. + +# 5.2. Comparison Results + +The comparison results of evaluation metrics are presented in Tab. 1. The baseline model, which does not capture correlations between interaction counterparts, indicates that directly anticipating these elements through multiple branches yields poor results. Our method outperforms comparative methods across all metrics for respective elements, demonstrating that leveraging correlations between humans and objects indeed benefits the comprehension of interaction relation, and the anticipation of interaction elements. It seems to be "the best of both worlds". Furthermore, we conduct a visual comparison of methods with higher evaluation metrics, as shown in Fig. 5. The results showcase multiple objects interacting with distinct human parts. As can be seen, our anticipations are more precise, for some uncertain regions that are not visible in images (e.g., the first and third rows), LEMON could also anticipate plausible results. This is attributed to the modeling of correlations between geometries, which compensates for missing features in invisible regions of the image. + +# 5.3. Ablation Study + +We conduct a thorough ablation study to validate the effectiveness of the model design. Tab. 2 (a) reports the model performance without modeling the semantic intention and the geometric correlation, demonstrating their impacts on the model performance. Moreover, the absence of constraining semantic consistency hinders the modeling of intentions, while not introducing curvatures affects the extraction of geometric correlations, resulting in a decrease in model performance. Both results are detailed in Tab. 2 (a). Besides, we test the impact of $\phi_c$ and $\mathcal{L}_{pr}$ (Eq. 6) on spatial relation prediction, shown in Tab. 2 (b). Notably, removing the extraction of $\phi_a$ in $f_{\theta}$ (Eq. 3) also affects the performance. This is attributed to the interrelation between $\phi_c$ and $\phi_a$ , the absence of $\phi_a$ impacts the representation of $\phi_c$ , and subsequently influencing the final spatial prediction. + +![](images/a20c21f91381d99efe9e5904feeb24f7d34000f7e98e42d19b90af7e3659020b.jpg) +Figure 5. Visualization Results. (a) Results of the estimated human vertices in contact with objects, the estimated contact vertices are shown in yellow. (b) The anticipations of 3D object affordance, the depth of red represents the probability of anticipated affordance. (c) Two views of the predicted spatial relation, translucent spheres are object proxies. Please zoom in for a better visualization. + +To further illustrate the effectiveness of capturing the intention and geometry affinity, we visualize attention maps on the human and object geometries when removing one of them, as shown in Fig. 6. The first row indicates without capturing intention affinity, solely relying on geometries leads the model to focus on multiple candidate regions, differing from attention regions in the image. Furthermore, the second row demonstrates that while semantic intention guides the model to identify approximate candidate regions, modeling geometry affinity further locates corresponding regions most related to the interaction of the geometries. + +# 5.4. Performance Analysis + +Multiple Interactions. To validate whether the model reasons interaction elements based on the understanding of interaction relation. We employ the model to infer different interactions between humans and the same object, as demonstrated in Fig. 7. The results showcase that, for the same object, when interactions are different, the model outputs distinct results, encompassing human contact, object affordance, and spatial relation, all of which align with the interaction relation depicted in the image. This indicates + +Table 2. Ablation studies. (a) Performance when not modeling intention representations (intent.), geometric correlations (geom.), and not introducing $\varphi$ (Eq. 1), curvatures (cur.). (b) The impact of several factors on spatial relation prediction. $\pmb{x}$ means without. + +
MetricsOursX intent.X φX geom.X cur.
Precision0.780.710.740.680.75
Recall0.820.730.790.700.78
F10.780.700.740.670.75
geo. (cm)7.5511.8710.2614.879.13
AUC88.5185.8786.6683.2386.98
aIOU41.3438.1940.0337.1339.58
SIM0.640.580.600.550.62
MSE0.0100.0270.0220.0310.018
+ +(a) + +that the model infers interaction elements according to the interaction relation rather than relying on direct mapping solely driven by the categories of human-object pairs. + +Multiple Objects. In scenarios where distinct objects are involved in concurrent interactions with humans, the model should indeed possess the ability to comprehend the interaction relation with different objects. We give an experi- + +![](images/515f7af041520d8c6b974735fd94545fc0e8e41ea0e0159ddba15d9a355162a1.jpg) +Figure 6. Visual Attention. Attention heatmaps on geometries, in cases with $(\boldsymbol{w})$ and without $(\boldsymbol{w} / \boldsymbol{o})$ the extraction of intention and geometry affinity. Images are also shown with attention heatmaps. + +![](images/367870f5c15a8499a823117baf1d9193b479f584a54d77fc99b83404fe10d731.jpg) +Support + +![](images/fbd7bea0d50476f68bb0d79d1ecd88a8f30b05cc1fb3b866f86d355bf51eb1c6.jpg) + +![](images/7f5534de2937f241c63a4f0151bc9bbf5e1d9ab3fd1923f33f1a9a5b7f5af81c.jpg) + +![](images/a281632b63ea86cdf1be7cb93a84780708011e603f022d610ae361bed98de558.jpg) +View 1 + +![](images/113dbefaef8cfdb07d5f5f3ab8600f855d04a286b9f904755a888d99701fb366.jpg) +View 2 + +![](images/d471bb0dec73e5b7056b161ab935e6b01e43bee894e119937712d155d32d8b7a.jpg) +Carry + +![](images/283630f3bc4effd236f3887b7ef621b8552028f67677d97d05f18c8181053f80.jpg) +Affordance + +![](images/5e71f181a0c26d57af78c846cacd36e5a9dd319c3f4e4e4fb30c5e58e9183a74.jpg) +Contact +Figure 7. Multiple Interactions. Distinct anticipations when interactions are different with the same object. + +![](images/fc43781236193e6c6217674838b5d323063ffa91653228ed98c7093f83a108eb.jpg) +Spatial Relation + +ment pertaining to this property. As shown in Fig. 8, when reasoning with different objects, the human contact, object affordance, and spatial relation are distinct, while all anticipations are plausible. This is attributed to the collaboration of the semantic interaction intention and geometric correlation, ensuring that the model could explicitly capture interaction intentions with different objects and reveal regions related to the interaction on corresponding geometries. + +Multiple Instances. We conduct an experiment to validate the model's understanding of interaction relation across various instances, assessing its generalization and robustness. Specifically, distinct object instances are utilized to infer with the same image and human geometry, shown in Fig. 9 (a). As can be seen, even though the contact and spatial relation are slightly different, the results are consistent with the interaction shown in the image, and object affordances are correctly anticipated. Additionally, the same object instance is utilized to infer with different human geometries and images that contain similar human-object interactions, as demonstrated in Fig. 9 (b). These results indicate that the model could generalize the understanding of interaction relation to different human and object instances. + +# 6. Conclusion + +We propose leveraging both counterparts of the HOI to jointly anticipate human contact, object affordance, and + +![](images/2f89380510429d19137e6c96618c4bb3935c2f989b3963653646aab9e2670850.jpg) +Figure 8. Multiple Objects. The anticipations for multiple objects with different human-object interactions. + +![](images/846879d2b64aa144c0162fce37c6ccc750690096cbae6408e2af7478237d2bd0.jpg) +Figure 9. Multiple Instances. (a) Different object instances $w.r.t.$ the same image and human geometry. (b) Different images and human instances $w.r.t.$ the same object. + +human-object spatial relation in 3D space. It holds the potential to drive embodied intelligence to learn from interactions and assist the interaction modeling in graphics. With the provided dataset 3DIR containing paired HOI data and multiple annotations, we train LEMON, a novel model that harnesses the interaction intention and geometric correlation between humans and objects to capture the affinity, eliminating interaction uncertainties and anticipating plausible interaction elements. Extensive experiments demonstrate the effectiveness and superiority of LEMON. We believe this work offers fresh insights and paves a new way for 3D human-object interaction understanding. + +Limitations and Future Work. In the current formulation, LEMON needs the pre-inferred human mesh as input. In the future, we plan to integrate HMR into the entire framework, freeing input constraints required for inference, and taking 3D interaction elements to improve the accuracy of HMR. Plus, another interesting direction is to leverage multi-modal methods [15, 88], boosting the interaction relation understanding from other sources e.g., text, audio. + +Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62306295. + +# References + +[1] Brenna D Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469-483, 2009. 1 +[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 6 +[3] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 3 +[4] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018, the winter conference on applications of computer vision (wacv), pages 381–389. IEEE, 2018. 2 +[5] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 1 +[6] Yixin Chen, Sai Kumar Dwivedi, Michael J. Black, and Dimitrios Tzionas. Detecting human-object contact in images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5 +[7] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22:449-462, 2013. 1 +[8] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3 +[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5 +[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 1, 3, 5, 6 +[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3 +[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference + +on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3 +[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 1 +[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 2 +[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 8 +[16] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2 +[17] Elena Corina Grigore, Kerstin Eder, Anthony G Pipe, Chris Melhuish, and Ute Leonards. Joint action understanding improves robot-to-human object handover. In 2013 IEEE/RSJ international conference on intelligent robots and systems, pages 4622-4629. IEEE, 2013. 1 +[18] Saurabh Gupta and Jitendra Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 5 +[19] Sookwan Han and Hanbyul Joo. Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3 +[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021. 1, 3 +[21] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 1, 3, 6 +[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3 +[23] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3 +[24] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1 + +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5 +[26] Jan J Koenderink and Andrea J Van Doorn. Surface shape and curvature scales. Image and vision computing, 10(8): 557-564, 1992. 4 +[27] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis, 2023. 3 +[28] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 3 +[29] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. arXiv preprint arXiv:2309.16237, 2023. 3 +[30] Lulei Li, Jianan Wei, Wenguan Wang, and Yi Yang. Neural-logic human-object interaction detection. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[31] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10166–10175, 2020. 1, 3, 5, 6 +[32] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 5 +[33] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 5 +[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[35] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. AKB-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 5 +[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3, 5, 6 +[37] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2252-2261, 2022. 3 +[38] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Leverage interactive affinity for affordance + +learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6809-6819, 2023. 3 +[39] Bertram F Malle and Joshua Knobe. The folk concept of intentionality. Journal of experimental social psychology, 33 (2):101-121, 1997. 2 +[40] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3 +[41] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 4 +[42] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5 +[43] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 5 +[44] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 1, 3 +[45] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3 +[46] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3 +[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 6 +[48] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai Nguyen. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing systems, 33:7841-7851, 2020. 3 +[49] Abdul Nurunnabi, Geoff West, and David Belton. Outlier detection and robust normal-curvature estimation in mobile laser scanning 3d point cloud data. Pattern Recognition, 48 (4):1404-1419, 2015. 4 +[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3 + +[51] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4726-4736, 2023. 3, 6 +[52] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 7 +[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6 +[54] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12802–12813, 2023. 3 +[55] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 3 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 3 +[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 1 +[59] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 3 +[60] Thomas B Sheridan. Human-robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 1 +[61] Chris Shilling. The body in culture, technology and society. The Body in Culture, Technology and Society, pages 1-256, 2004. 2 +[62] Soshi Shimada, Vladislav Golyanik, Zhi Li, Patrick Pérez, Weipeng Xu, and Christian Theobalt. Hulc: 3d human motion capture with pose manifold sampling and dense contact guidance. In European Conference on Computer Vision, pages 516-533. Springer, 2022. 1, 3 +[63] Michael Tomasello, Malinda Carpenter, Josep Call, Tanya Behne, and Henrik Moll. Understanding and sharing in- + +tentions: The origins of cultural cognition. Behavioral and brain sciences, 28(5):675-691, 2005. 2 +[64] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 1, 3, 5, 6, 7 +[65] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 3 +[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 6 +[67] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In European Conference on Computer Vision, pages 90-107. Springer, 2022. 1, 3 +[68] Ping Wei, Yibiao Zhao, Nanning Zheng, and Song-Chun Zhu. Modeling 4d human-object interactions for event and object recognition. In Proceedings of the IEEE international conference on computer vision, pages 3272-3279, 2013. 1 +[69] Ruihai Wu, Chuanruo Ning, and Hao Dong. Learning foresightful dense visual affordance for deformable object manipulation. arXiv preprint arXiv:2303.11057, 2023. 3 +[70] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1 +[71] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, 2022. 3 +[72] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and Mohan S Kankanhalli. Learning to detect human-object interactions with knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2 +[73] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022.3 +[74] Rui Xu, Zhiyang Dou, Ningna Wang, Shiqing Xin, Shuang-min Chen, Mingyan Jiang, Xiaohu Guo, Wenping Wang, and Changhe Tu. Globally consistent normal orientation for point clouds by regularizing the winding-number field. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 4 +[75] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 3 + +[76] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3d-hoi: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3 +[77] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 4 +[78] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10905-10915, 2023. 1, 3, 5, 6, 7 +[79] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. arXiv preprint arXiv:2307.15880, 2023. 5 +[80] Bangpeng Yao and Li Fei-Fei. Modeling mutual context of object and human pose in human-object interaction activities. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 17–24. IEEE, 2010. 1 +[81] Sifan Ye, Yixing Wang, Jiaman Li, Dennis Park, C Karen Liu, Huazhe Xu, and Jiajun Wu. Scene synthesis from human motion. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 3 +[82] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3 +[83] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 3 +[84] Guochen Yu, Andong Li, Chengshi Zheng, Yinuo Guo, Yutian Wang, and Hui Wang. Dual-branch attention-inattention transformer for single-channel speech enhancement. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7847-7851. IEEE, 2022. 3 +[85] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10): 2472-2500, 2022. 3 +[86] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, pages 1-26, 2023. 3 +[87] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 34-51. Springer, 2020. 3 +[88] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng + +Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 8 +[89] Xiaopeng Zhang, Hongjun Li, Zhanglin Cheng, and Yikuan Zhang. Robust curvature estimation and geometry analysis of 3d point cloud surfaces. J. Inf. Comput. Sci, 6(5):1983-1990, 2009. 4 +[90] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3 +[91] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022.3 +[92] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper manipulation. In International Conference on Learning Representations, 2023. 1 +[93] Xinghao Zhu, JingHan Ke, Zhixuan Xu, Zhixin Sun, Bizhe Bai, Jun Lv, Qingtao Liu, Yuwei Zeng, Qi Ye, Cewu Lu, et al. Diff-lfd: Contact-aware model-based learning from visual demonstration for robotic manipulation via differentiable physics-based simulation and rendering. In 7th Annual Conference on Robot Learning, 2023. 3 +[94] Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11825-11834, 2021. 2 \ No newline at end of file diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/images.zip b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..86bd52f4f11d012a3d1d3c4d74b89b8b1f9efa47 --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:465b430d5f906fb559bd017767b0196f3c90f44936414627df1607b8661e5033 +size 748145 diff --git a/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/layout.json b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..430d524ad82528d685d40f4b8b77a791a86881ae --- /dev/null +++ b/2024/LEMON_ Learning 3D Human-Object Interaction Relation from 2D Images/layout.json @@ -0,0 +1,11201 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "text", + "content": "LEMON: Learning 3D Human-Object Interaction Relation from 2D Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "content": "Yuhang Yang" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "content": ", Wei Zhai" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "content": ", Hongchen Luo" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "content": ", Yang Cao" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "text", + "content": ", Zheng-Jun Zha" + }, + { + "bbox": [ + 114, + 142, + 477, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 157, + 416, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 157, + 416, + 172 + ], + "spans": [ + { + "bbox": [ + 179, + 157, + 416, + 172 + ], + "type": "text", + "content": "1 University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "text", + "content": " Institute of Artificial Intelligence, Hefei Comprehensive National Science Center" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 188, + 480, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 188, + 480, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 188, + 480, + 198 + ], + "type": "text", + "content": "{yyuhang@mail., wzhai056@, lhcl2@mail., forrest@, zhazj@}ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 206, + 533, + 323 + ], + "blocks": [ + { + "bbox": [ + 60, + 206, + 533, + 323 + ], + "lines": [ + { + "bbox": [ + 60, + 206, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 60, + 206, + 533, + 323 + ], + "type": "image", + "image_path": "27b4b9f9d5af39aae9820d3c1c34bd1c7e4c90eae5e1f22245c4d2f17ce42d3b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 546, + 365 + ], + "type": "text", + "content": "Figure 1. For an interaction image with paired geometries of the human and object, LEMON learns 3D human-object interaction relation by jointly anticipating the interaction elements, including human contact, object affordance, and human-object spatial relation. Vertices in yellow denote those in contact with the object, regions in red are object affordance regions, and the translucent sphere is the object proxy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "spans": [ + { + "bbox": [ + 143, + 374, + 192, + 387 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 45, + 402, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 402, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 402, + 290, + 689 + ], + "type": "text", + "content": "Learning 3D human-object interaction relation is pivotal to embodied AI and interaction modeling. Most existing methods approach the goal by learning to predict isolated interaction elements, e.g., human contact, object affordance, and human-object spatial relation, primarily from the perspective of either the human or the object. Which underexploit certain correlations between the interaction counterparts (human and object), and struggle to address the uncertainty in interactions. Actually, objects' functionalities potentially affect humans' interaction intentions, which reveals what the interaction is. Meanwhile, the interacting humans and objects exhibit matching geometric structures, which presents how to interact. In light of this, we propose harnessing these inherent correlations between interaction counterparts to mitigate the uncertainty and jointly anticipate the above interaction elements in 3D space. To achieve this, we present LEMON (LEarning 3D huMan-Object iNteraction relation), a unified model that mines interaction intentions of the counterparts and employs curvatures to guide the extraction of geometric correlations, combining them to anticipate the interaction elements. Besides, the 3D Interaction Relation dataset (3DIR) is collected to serve as the test bed for training and evaluation. Extensive experiments demonstrate the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 376, + 546, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 546, + 411 + ], + "type": "text", + "content": "superiority of LEMON over methods estimating each element in isolation. The code and dataset are available at https://yyvhang.github.io/LEMON." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 428, + 387, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 428, + 387, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 428, + 387, + 441 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 593 + ], + "type": "text", + "content": "Learning 3D human-object interaction (HOI) relation seeks to capture semantic co-occurrence and geometric compatibility between humans and objects in 3D space [5, 68, 80]. How can machines learn the interaction relation? One possible solution involves perceiving certain elements capable of revealing the interaction. Contact [6, 21, 64, 70], affordance [10, 13, 46, 92], and spatial relation [19, 31, 52] that elucidate \"where\" the interaction manifests between the human and object garner great attention. Capturing representations of such elements is pivotal for applications like AR/VR [7], imitation learning [1, 24], embodied AI [17, 58, 60], and interaction modeling [20, 81]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Humans predominantly manipulate and interact with objects in 3D space. Thus, many methods devise task-specific models to anticipate certain interaction elements, thereby perceiving 3D HOI relation. Methods [21, 62, 64] estimate dense human contact based on the interaction semantics depicted in images. Some studies anticipate the object affordance according to objects' structures or 2D visuals [10, 44, 67, 78]. Several works delve into predicting the human-object spatial relation through synthetic images [19] or posed human geometries [52]. A prevailing trend in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 142, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 142, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 142, + 713 + ], + "type": "text", + "content": "†Corresponding Author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16284" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 276 + ], + "type": "text", + "content": "these methods involves taking certain attributes (e.g., appearances, geometries) of either humans or objects to predict an isolated interaction element, capturing one aspect of the interaction relation. However, they overlook the impact of factors mutually determined by humans and objects on interactions, such as the interaction intention and geometric correlation, which leads to struggles in addressing the uncertainty within interactions. Specifically, the uncertainty could be attributed to two principal aspects. On the one hand, the diversity of HOIs introduces challenges in capturing explicit interaction intentions solely centering on humans or objects, raising the intention uncertainty. On the other hand, the limited view and mutual occlusions give rise to invisible interaction regions in images, complicating the constitution of linkages between these regions and target 3D geometries, causing the region uncertainty. These uncertainties may culminate in ambiguous anticipations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 278, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 289, + 506 + ], + "type": "text", + "content": "In this paper, we propose leveraging both counterparts of the interaction to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space (Fig. 1), addressing the uncertainty by unearthing the correlation between the interacting humans and objects. Actually, humans and objects are intertwined and possess affinities in the interaction (Fig. 2). In specific, the design of objects typically adheres to certain human needs. Therefore, object affordances inherently hint at \"what\" interactions humans intend to make [39, 61, 63], revealing the intention affinity of interactions. Meanwhile, the interacting human and object exhibit matching geometries (either posture or configuration), which presents \"how\" to interact, arising the geometry affinity. The intention affinity clarifies the interaction type and implicates the interaction regions. Geometry affinity could serve as pivotal clues for excavating correlations between geometries corresponding to invisible regions in the image. These interaction-related regions e.g., contact regions, further reflect the human-object spatial relation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 289, + 687 + ], + "type": "text", + "content": "To achieve this, we present the LEMON, a novel framework that correlates the intention semantics and geometric correspondences to jointly anticipate human contact, object affordance, and human-object spatial relation, in 3D space. To capture the intention affinity, LEMON employs multibranch attention to model the correlation between the interaction content in images and geometries of humans and objects, revealing intention representations of the interaction corresponding to geometries. The cosine similarity is utilized to further ensure their semantic consistency. Taking intention representations as conditions, LEMON integrates geometric curvatures to capture the geometry affinity and reveal the human contact and object affordance representations. These representations then assist in anticipating the spatial relation constrained by a combined distance loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "In addition, we collect the 3DIR dataset, which contains natural HOI images paired with object point clouds" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 68, + 536, + 215 + ], + "blocks": [ + { + "bbox": [ + 318, + 68, + 536, + 215 + ], + "lines": [ + { + "bbox": [ + 318, + 68, + 536, + 215 + ], + "spans": [ + { + "bbox": [ + 318, + 68, + 536, + 215 + ], + "type": "image", + "image_path": "f0e94512c6a1f29f5d79eee91ab59b5bae98125bde8267268f10c74dfaa2b57c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 223, + 547, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 223, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 547, + 269 + ], + "type": "text", + "content": "Figure 2. Motivation. Affinities within the HOI. The object affordance inherently reveals the human's interaction intention, arising the intention affinity. The interacting human and object possess matching structures, exhibiting the geometry affinity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 289, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 336 + ], + "type": "text", + "content": "and SMPL-H [57] pseudo-GTs. Multiple annotations are made for these data, e.g., dense human contact, object affordance, and human-object spatial relation. It serves as the test bed for the model training and evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 337, + 503, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 337, + 503, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 337, + 503, + 348 + ], + "type": "text", + "content": "The contributions are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 349, + 545, + 533 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 308, + 349, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 408 + ], + "type": "text", + "content": "1) We thoroughly exploit the correlation between the interaction counterparts to jointly anticipate human contact, object affordance, and human-object spatial relation in 3D space. It furnishes essential interaction elements to comprehend 3D HOIs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 411, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 470 + ], + "type": "text", + "content": "2) We present the LEMON, a novel framework that correlates semantic interaction intention and geometries to capture the affinity between humans and objects, eliminating the impact of the interaction uncertainty and anticipating plausible 3D interaction elements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 472, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 472, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 472, + 545, + 533 + ], + "type": "text", + "content": "3) We provide the 3DIR dataset that contains paired HOI data and multiple annotations, including dense human contact, object affordance, and human-object spatial relation, to support the anticipation. Extensive experiments demonstrate the superiority of LEMON." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 544, + 392, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 392, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 392, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 563, + 525, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 525, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 525, + 576 + ], + "type": "text", + "content": "2.1. Human-Object Interaction Relation in 2D" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Various approaches perceive elements pertinent to human-object interaction relation from multiple perspectives in the 2D domain. Methods [4, 16, 30, 72, 94] represent human-object interaction as triplets " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " human, action, object>. Grounding bounding boxes of interaction counterparts and actions corresponding to them. They output the instance-level human-object spatial relation in pixel space and action semantics. Delving into the object side, the concept of \"interaction possibility\" is encapsulated in the term affordance [14], and bountiful methods explore to anticipate the object affordance, whether at instance-level or part-level" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16285" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "[11, 12, 37, 38, 47, 85]. Turning to the human side, the contact explicitly indicates where the human interacts with objects. Some works detect the contact for specific parts (e.g., hand and foot) at a box-region level [48, 55, 59]. HOT [6] provides 2D contact annotations to support human-object contact estimation in images. These efforts achieve substantial results in understanding the human-object interaction relation in 2D space. However, they still encounter challenges when extrapolating to 3D space for practical applications due to the lack of a dimension." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 201, + 266, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 266, + 215 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 266, + 215 + ], + "type": "text", + "content": "2.2. Human-Object Interaction Relation in 3D" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 220, + 288, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 353 + ], + "type": "text", + "content": "To facilitate the incorporation of interaction comprehension into applications, extensive works explore to anticipate interaction elements in 3D space. For object affordance, methods [10, 44, 67, 69, 73, 86, 91] constitute the linkage between object shapes and affordances by a learning-based mapping, approaching the matter from the perspective of objects' functionality and structure. Some approaches learn object affordances from a distinct perspective, based on interactions. Whether making agents actively interact with 3D synthetic scenarios [46], or ground 3D object affordances through object-object [45] and 2D [78] interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 533 + ], + "type": "text", + "content": "The dense human contact estimation is primarily based on the SMPL series [36, 50, 57]. Numerous methods model human contact for various tasks [8, 20, 22, 23, 71, 82, 83]. HULC [62], BSTRO [21], and DECO [64] are more similar to the estimation of contact in our method. They learn a mapping from the interaction semantics in images to the human vertices sequence, hardly leveraging geometries. Different from them, our method harnesses semantic and geometric correlations of interaction counterparts to address the uncertainty and infer the vertices that contact with objects. DECO also contributes the DAMON [64] dataset that possesses dense 3D contact annotation for in-the-wild images. In addition to human contact annotations, 3DIR includes other annotations related to the interaction, e.g., 3D object affordance and human-object spatial relation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 664 + ], + "type": "text", + "content": "Bounding boxes obtained by HOI detection methods give the human-object spatial relation in pixel space. To lift this relation to 3D space, DJ-RN [31] takes hollow spheres with defined radii to represent objects, and projects the spatial relation in images to 3D space based on the bounding boxes and defined radii. CHORUS [19] utilizes synthetic multi-view images generated by the diffusion model [56] to learn the spatial relation between objects and canonical humans. Given a posed human, Object Pop-Up [51] anticipates the objects and their spatial positions to match the human body for certain interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "The above methods commonly focus on one side of the interaction in isolation. In contrast, our method captures the inherent affinity between both sides of the interaction to jointly anticipate interaction elements in 3D space. Such" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "anticipations are beneficial for tasks e.g., robot manipulation [3, 40, 93], interaction generation [27, 29, 75, 81, 90] and reconstruction [20, 71, 76, 87]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 118, + 362, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 118, + 362, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 118, + 362, + 131 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": "Given the inputs " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\{H,O,I\\}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "H\\in \\mathbb{R}^{N_h\\times 3}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " indicates vertices of the SMPL-H [57], which represents the human as pose " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^{52\\times 3}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{10}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " parameters, and output a mesh " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "M(\\theta ,\\beta)\\in \\mathbb{R}^{6890\\times 3}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "O\\in \\mathbb{R}^{N_o\\times 3}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " is an object point cloud, " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "I\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " is an image. " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "N_{h},N_{o}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " are the number of points, " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "H,W" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " are image's height and width. LEMON jointly anticipates the human contact " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", object affordance " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", and object center position " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_p\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", in 3D space. As shown in Fig. 3, initially, the inputs are sent to image and point cloud backbones, obtain respective features " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h,\\mathbf{F}_o,\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ". Then, LEMON correlates " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " to extract features of the interaction intention (" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ") inherent in geometries through multi-branch attention, and employs cosine similarity to constrain their semantic consistency (Sec. 3.1). With " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " as conditions, LEMON integrates curvatures to guide the modeling of geometric correlations, capturing the geometry affinity and revealing the contact " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " and affordance " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " features (Sec. 3.2). Next, geometric features and " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\phi_c,\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " are utilized to model the object spatial feature " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\phi_p" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " (Sec. 3.3). Eventually, " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\phi_c,\\phi_a,\\phi_p" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " are projected to " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_c,\\bar{\\phi}_a" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_p" + }, + { + "bbox": [ + 304, + 137, + 547, + 402 + ], + "type": "text", + "content": " in the decoder, the whole process is optimized by a combined loss (Sec. 3.4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 407, + 482, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 482, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 482, + 419 + ], + "type": "text", + "content": "3.1. Interaction Intention Excavation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": "We add human and object masks on " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", and utilize the HR-Net [65] and DGCNN [66] as backbones to extract the image feature " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i \\in \\mathbb{R}^{C \\times h \\times w}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", the geometric features of human " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h \\in \\mathbb{R}^{C \\times N_h}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", and object " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o \\in \\mathbb{R}^{C \\times N_o}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " is flattened to " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times hw}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ". Images contain rich interaction semantics, which could serve as clues to unearth the interaction intention within the geometries. In detail, tokens " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_o" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_h \\in \\mathbb{R}^{C \\times 1}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " are generated to represent intention features inherent in geometries. " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_o" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_h" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " are concatenated with " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " respectively to get the feature sequences " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{to} \\in \\mathbb{R}^{C \\times (N_o + 1)}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{th} \\in \\mathbb{R}^{C \\times (N_h + 1)}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ". Taking " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " as the shared key and value, " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{to}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{th}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " as queries " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "(q_1, q_2)" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " in two branches, the multi-branch attention is employed to model the intention features, expressed as " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{to}, \\bar{\\mathbf{F}}_{th} = f_{\\delta}([\\mathbf{F}_{to}, \\mathbf{F}_{th}], \\mathbf{F}_i)" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": ". Where " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "f_{\\delta}" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " indicates the multi-branch attention layer [28, 54, 84], " + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 304, + 426, + 545, + 654 + ], + "type": "text", + "content": " denotes two branches with different queries. Human and object geometries possess multiple interaction possibilities, which may introduce semantic ambiguity. To mitigate this, we further constrain the consistency of semantic tokens:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 380, + 658, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 658, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 380, + 658, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\varphi = \\frac {\\bar {\\mathbf {T}} _ {o} \\cdot \\bar {\\mathbf {T}} _ {h}}{\\| \\bar {\\mathbf {T}} _ {o} \\| _ {2} \\times \\| \\bar {\\mathbf {T}} _ {h} \\| _ {2}}, \\tag {1}", + "image_path": "1421ffb8d01538600461ca2035235973330218b61de672f056a74b9c83762a6e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h\\in \\mathbb{R}^{C\\times 1}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " are split from " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{to},\\bar{\\mathbf{F}}_{th}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ", representing the semantic intention features of geometries. " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16286" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 67, + 521, + 311 + ], + "blocks": [ + { + "bbox": [ + 69, + 67, + 521, + 311 + ], + "lines": [ + { + "bbox": [ + 69, + 67, + 521, + 311 + ], + "spans": [ + { + "bbox": [ + 69, + 67, + 521, + 311 + ], + "type": "image", + "image_path": "e0c94c7a1bed6d704787b1ae2ae4395db9ed8d68366ca629776ea01f23dcf095.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": "Figure 3. LEMON pipeline. Initially, it takes modality-specific backbones to extract respective features " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", which are then utilized to excavate intention features " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "(\\bar{\\mathbf{T}}_o, \\bar{\\mathbf{T}}_h)" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " of the interaction (Sec. 3.1). With " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " as conditions, LEMON integrates curvatures " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "(C_o, C_h)" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " to model geometric correlations and reveal the contact " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", affordance " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " features (Sec. 3.2). Following, the " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " is injected into the calculation of the object spatial feature " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_p" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " (Sec. 3.3). Eventually, the decoder projects " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\phi_p" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": " to the final outputs " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_c" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_a" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_p" + }, + { + "bbox": [ + 46, + 319, + 547, + 366 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 384, + 253, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 253, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 253, + 397 + ], + "type": "text", + "content": "semantic cosine similarity, a part of " + }, + { + "bbox": [ + 47, + 384, + 253, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 47, + 384, + 253, + 397 + ], + "type": "text", + "content": " in Sec. 3.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 263, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 263, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 263, + 417 + ], + "type": "text", + "content": "3.2. Curvature-guided Geometric Correlation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": "The interacting humans and objects exhibit certain geometry affinity [77], manifesting in matching geometric structures and correlative curvatures. For geometric curvatures, the normal curvature could better represent local structures like interaction regions [26, 49]. Thus, we encode the normal curvatures into geometric features, and regard the " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o,\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " as conditions to capture the affinity among human and object geometries. Normal curvatures " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "C_o\\in \\mathbb{R}^{1\\times N_o}" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "C_h\\in \\mathbb{R}^{1\\times N_h}" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " of object and human are obtained by local fitting method [74, 89]. To correlate the curvatures, " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "C_o,C_h" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " are encoded to high dimension " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "C_o'\\in \\mathbb{R}^{C\\times N_o},C_h'\\in \\mathbb{R}^{C\\times N_h}" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " , and the cross-attention " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " is mutually performed on them, formulated as " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "\\bar{C}_o = f_m(C_o',C_h'),\\bar{C}_h = f_m(C_h',C_o')" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " . Following, to make the curvature guide the calculation of geometric correlation, " + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "inline_equation", + "content": "\\bar{C}_o,\\bar{C}_h" + }, + { + "bbox": [ + 46, + 422, + 289, + 614 + ], + "type": "text", + "content": " are integrated and fused with geometric features:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 620, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 620, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 83, + 620, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {F}} _ {c o} = f \\left(\\Gamma \\left(\\bar {C} _ {o}, \\bar {\\mathbf {F}} _ {o}\\right)\\right), \\bar {\\mathbf {F}} _ {c h} = f \\left(\\Gamma \\left(\\bar {C} _ {h}, \\bar {\\mathbf {F}} _ {h}\\right)\\right), \\tag {2}", + "image_path": "2c08f16b9162957cc233572e60b905fdc7e31dd1ad5ad19db8eb72a1f86c905c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{co},\\bar{\\mathbf{F}}_o\\in \\mathbb{R}^{C\\times N_o}" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{ch},\\bar{\\mathbf{F}}_h\\in \\mathbb{R}^{C\\times N_h}" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_o,\\bar{\\mathbf{F}}_h" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " are split from " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{to}" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{th}" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " denotes the concatenation, " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " indicates convolution layers with " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " kernel. Then, " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 46, + 640, + 288, + 715 + ], + "type": "text", + "content": " are considered as conditions [41] to further screen candidate regions that match the interaction depicted in images and participate in the modeling of geometry affinity," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 384, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 545, + 409 + ], + "type": "text", + "content": "revealing the affordance and contact features " + }, + { + "bbox": [ + 305, + 384, + 545, + 409 + ], + "type": "inline_equation", + "content": "\\phi_{a},\\phi_{c}" + }, + { + "bbox": [ + 305, + 384, + 545, + 409 + ], + "type": "text", + "content": " . The process could be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 420, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 420, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 420, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\phi_ {a} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\overline {{\\mathbf {T}}} _ {o}\\right)\\right), \\phi_ {c} = f _ {\\theta} \\left(\\overline {{\\mathbf {F}}} _ {c h}, \\Gamma \\left(\\overline {{\\mathbf {F}}} _ {c o}, \\overline {{\\mathbf {T}}} _ {h}\\right)\\right), \\tag {3}", + "image_path": "083645f67e0a4c70188d44a9693f5dff45e525783a35dec8f3b530872c1cdc01.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{a} \\in \\mathbb{R}^{C \\times N_{o}}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": ", to calculate it, " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{co}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": " serves as the query, and the concatenation of " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{ch}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_{o}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": " serves as the key and value, " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": " is the cross-transformer layer. " + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "inline_equation", + "content": "\\phi_{c} \\in \\mathbb{R}^{C \\times N_{h}}" + }, + { + "bbox": [ + 304, + 445, + 547, + 495 + ], + "type": "text", + "content": " is obtained through the analogous way." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 504, + 476, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 476, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 476, + 517 + ], + "type": "text", + "content": "3.3. Contact-aware Spatial Relation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": "Human-object interactions are extremely diverse, rendering the reasoning of their 3D spatial relation very challenging. Nevertheless, human contact implicitly constrains the position of the object, thereby assisting in inferring plausible human-object spatial relation. Thus, the object's center is represented as a token sequence " + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{sp} \\in \\mathbb{R}^{C \\times 3}" + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": ", with introducing the contact feature as a constraint, LEMON takes the semantic intention and geometric features of the object to query the corresponding features of the human. Modeling " + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{sp}" + }, + { + "bbox": [ + 304, + 524, + 545, + 644 + ], + "type": "text", + "content": " as the object spatial feature, formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 653, + 545, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 653, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 653, + 545, + 667 + ], + "type": "interline_equation", + "content": "\\phi_ {p} = f _ {\\rho} \\left(\\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c o}\\right), \\overline {{\\mathbf {T}}} _ {o}, \\mathbf {T} _ {s p}\\right) + p e, \\Gamma \\left(\\Theta \\left(\\overline {{\\mathbf {F}}} _ {c h}\\right), \\overline {{\\mathbf {T}}} _ {h}, \\phi_ {c}\\right)\\right), \\tag {4}", + "image_path": "a050763ca9d30060c78eb6cafe411cd58c96c726967ce6b25cb374cc555b73d4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\phi_p\\in \\mathbb{R}^{C\\times 3}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": ", is the spatial feature of the object center, " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{\\rho}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " is a cross-attention layer, " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " denotes the pooling layer, and " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "pe" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": " indicates a learnable positional encoding." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16287" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 67, + 184, + 224 + ], + "blocks": [ + { + "bbox": [ + 66, + 67, + 184, + 224 + ], + "lines": [ + { + "bbox": [ + 66, + 67, + 184, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 67, + 184, + 224 + ], + "type": "image", + "image_path": "47a46c1086f386cf8589b776ab8b297149c6bb6c22bb525240bab82ad9c49e48.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 547, + 281 + ], + "type": "text", + "content": "Figure 4. 3DIR Dataset. (a) The quantity of images and point clouds for each object, and a data sample containing the image, mask, dense human contact annotation, 3D object with affordance annotation, and the fitted human mesh with the object proxy sphere. (b) The proportion of our contact annotations within 24 parts on SMPL [36], and distributions of contact vertices for certain HOIs. (c) The ratio of annotated affordance regions to the whole object geometries, and the distribution of this ratio for some categories. (d) Mean distances (unit: m) between annotated object centers and human pelvis joints, and directional projections of annotated centers for several objects." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 66, + 310, + 224 + ], + "blocks": [ + { + "bbox": [ + 187, + 66, + 310, + 224 + ], + "lines": [ + { + "bbox": [ + 187, + 66, + 310, + 224 + ], + "spans": [ + { + "bbox": [ + 187, + 66, + 310, + 224 + ], + "type": "image", + "image_path": "a54ab6ed10b8e2318b618eee84b8bb53b04c323bf078cb2cae034e1c6efc213c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 66, + 408, + 224 + ], + "blocks": [ + { + "bbox": [ + 312, + 66, + 408, + 224 + ], + "lines": [ + { + "bbox": [ + 312, + 66, + 408, + 224 + ], + "spans": [ + { + "bbox": [ + 312, + 66, + 408, + 224 + ], + "type": "image", + "image_path": "5b95b3c666aa345b8c03b13e593015b93624bf09d5e4bc8ee9702bf4bdf333ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 408, + 66, + 529, + 224 + ], + "blocks": [ + { + "bbox": [ + 408, + 66, + 529, + 224 + ], + "lines": [ + { + "bbox": [ + 408, + 66, + 529, + 224 + ], + "spans": [ + { + "bbox": [ + 408, + 66, + 529, + 224 + ], + "type": "image", + "image_path": "36557fb43acd6489199a9bb243d6764f12300f27c7d334ae8a9329bfd3f6c374.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 300, + 140, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 140, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 140, + 312 + ], + "type": "text", + "content": "3.4. Loss Functions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "content": "Eventually, " + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\phi_{a},\\phi_{c},\\phi_{p}" + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "content": " are sent to the decoder that contains three projection heads with linear, normalization, and activation layers. Which output object affordance " + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_a\\in \\mathbb{R}^{N_o\\times 1}" + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "content": ", human contact " + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_c\\in \\mathbb{R}^{N_h\\times 1}" + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "content": ", and object's center position " + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_p\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 319, + 288, + 415 + ], + "type": "text", + "content": " respectively. With defined radii for specific objects, we treat a sphere as the object proxy [31], representing the spatial relation in the camera coordinates of fitted humans. The overall training loss is expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 420, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 420, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 97, + 420, + 287, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\omega_ {1} \\mathcal {L} _ {c} + \\omega_ {2} \\mathcal {L} _ {a} + \\omega_ {3} \\mathcal {L} _ {s} + \\omega_ {4} \\mathcal {L} _ {p}, \\tag {5}", + "image_path": "0cdd3da0c7ea7ee770aafef1696d02d969605585b1bc742daf28cbfbac391fa7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\omega_{1-4}" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " are hyper-parameters to balance the losses. " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " possess the same formulation, a focal loss [34] combined with a dice loss [42]. " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is calculated by " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_c" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " and the contact label " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_c" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_a" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is calculated by " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_a" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " and the affordance annotation " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_a" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is employed to ensure " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_o" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{T}}_h" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " align with interaction semantics of the image and maintain consistency in the semantic space. Specifically, " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is mapped to a logit " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": ", which is used to calculate a cross-entropy loss " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " with interaction categorical label " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is the sum of " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " (Eq. 1). For 3D HOIs, object positions are diverse due to variations in human orientation, relying solely on absolute coordinates as supervision makes it hard for the model to learn a consistent mapping. However, the relative distances between humans and objects are similar for specific HOIs. This smoother distance distribution contributes to reducing the optimization space to an approximate sphere for specific HOIs. Consequently, we take the distance between the object center and the human pelvis joint as an additional constraint, and " + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 46, + 437, + 287, + 666 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 669, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 669, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 51, + 669, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\mathcal {L} _ {p a} + \\mathcal {L} _ {p r}, \\quad \\mathcal {L} _ {p a} = | | \\bar {\\phi} _ {p} - \\hat {\\phi} _ {p} | | _ {2}, \\mathcal {L} _ {p r} = | | \\bar {\\xi} - \\hat {\\xi} | | _ {2}, (6)", + "image_path": "33dc8acea9675f15f38187d0a4e5c236a9a7fe81caeec05c2ba98813811941b0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_p" + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "text", + "content": " is the annotation of the object center. " + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\xi},\\hat{\\xi}" + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "text", + "content": " denotes distances between the pelvis joint and " + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_p,\\hat{\\phi}_p" + }, + { + "bbox": [ + 47, + 686, + 287, + 715 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 300, + 360, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 360, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 360, + 312 + ], + "type": "text", + "content": "4. Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 323, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 359 + ], + "type": "text", + "content": "We introduce the collection and annotation protocols of the 3DIR and give some statistical analysis of the collected data and annotations, shown in Fig. 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 362, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 362, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 362, + 546, + 483 + ], + "type": "text", + "content": "Collection. We collect images with explicit interaction contents, in which humans interact with specific objects. These images adhere to the condition that the human's upper body is present, ensuring the efficient recovery of the human mesh. In total, we collect 5k in-the-wild images from HAKE [32], V-COCO [18], PIAD [78], and websites with free licenses, spanning 21 object classes and 17 interaction categories. Additionally, we collect over 5k 3D object instances from several 3D datasets [9, 10, 35, 43], based on the category of objects in collected images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Annotation. We make over 25k annotations with multiple types for the collected data. 1) Masks: With the assistance of the SAM [25], we manually mask the interacting human and object in the image. 2) Human Mesh: In natural images, some humans only exhibit the upper body. Thus, we utilize the pipeline of UBody [33] to fit SMPLH pseudo-GTs. The 2D body and hand joints needed by the pipeline are obtained through the DWPose [79]. 3) Contact: Similar to DAMON [64] and HOT [6], the human contact is annotated based on human knowledge. For each image, we clearly provide masks and a specific interaction type for the human-object pair, professional annotators are hired to \"draw\" vertices on the human template that contact with objects. The entire annotation process cycles three rounds, with each round incorporating subjective cross-checks and objective metric checks. 4) Affordance: We refer to the 3D-AffordanceNet [10] to annotate the object affordance. The annotations of 11 objects included in 3D-AffordanceNet are directly utilized, and we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16288" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 548, + 203 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Comparison on the 3DIR. Evaluation metrics of comparison methods on the benchmark, the best results are covered with the mask. indicates the improvement relative to the baseline. P. means take PointNet++ [53] as the backbone, and D. means DGCNN [66]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 548, + 203 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 548, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 548, + 203 + ], + "type": "table", + "html": "
Human ContactObject AffordanceSpatial Relation
MethodsPrecision ↑Recall ↑F1 ↑geo. (cm) ↓MethodsAUC ↑aIOU ↑SIM ↑MethodsMSE ↓
Baseline0.490.520.4932.83Baseline82.3632.630.50Baseline0.051
BSTRO [21]0.570.580.5528.583DA. [10]85.493.8%0.56DJ-RN [31]0.042
DECO [64]0.700.720.6915.25IAG [78]86.635.1%0.59PopUp [51]0.027
Ours P.0.760.810.779.02Ours P.87.9140.970.63Ours P.0.012
Ours D.0.780.820.787.55Ours D.88.5141.340.64Ours D.0.010
", + "image_path": "89fcf5b7da4aeb5de3c36e9d3fb3eb33f846279311a1a531afb53ce8ef616257.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 353 + ], + "type": "text", + "content": "An additional 10 object categories excluded in 3D-AffordanceNet. 5) Spatial Relation: For each sample, we color the fitted human mesh with annotated per-vertex human contact and treat a sphere as the object proxy. The radius of the sphere for each object category is pre-defined, referring to DJ-RN [31]. Given the posed human in color and the proxy sphere, annotators adjust the sphere relative to the human to align with the human-object spatial relation depicted in the image. Ultimately, we record the center coordinates of the adjusted sphere. Due to the page limitation, please refer to the Sup. Mat. for more annotation details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 354, + 288, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 569 + ], + "type": "text", + "content": "Statistical Analysis. The quantity of images and 3D instances for each object category in the 3DIR is shown in Fig. 4 (a). For contact annotation, we count its proportion within 24 human parts defined on the SMPL [36]. Moreover, we visualize the distribution of contact annotation for several HOIs, where the deeper color indicates more contact annotations at the vertex, as shown in Fig. 4 (b). Fig. 4 (c) demonstrates the proportion of annotated affordance regions to the entire object geometry, encompassing the mean of each category and the distribution of all instances for several categories. As can be seen, there are differences in distinct objects with the same affordance, as well as variations in distinct affordances of the same object. For spatial relation annotation, we count the mean distance between annotated centers and human pelvis joints, and project the spatial direction of annotated centers onto a fixed-radius sphere with the pelvis joint as the center. Fig. 4 (d) shows several cases and the distribution of mean distances." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 579, + 124, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 124, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 124, + 594 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 599, + 160, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 160, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 160, + 612 + ], + "type": "text", + "content": "5.1. Benchmark Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We refer to methods that anticipate each interaction element [21, 47, 64, 78] to thoroughly benchmark the 3DIR. For the training, in addition to the training data in 3DIR, we select another 5k data with low redundancy in BEHAVE [2]. Evaluation is conducted on the test set of 3DIR. The baseline model adopts a multitask-like framework, directly utilizing three branches to make anticipations. Besides, we compare LEMON with advanced methods that anticipate respective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 222, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 258 + ], + "type": "text", + "content": "elements. Plus, we also conduct evaluation experiments on DAMON [64], BEHAVE [2], and PIAD [78], the results and implementation details are in the Sup. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 263, + 424, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 263, + 424, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 263, + 424, + 277 + ], + "type": "text", + "content": "5.2. Comparison Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 282, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 510 + ], + "type": "text", + "content": "The comparison results of evaluation metrics are presented in Tab. 1. The baseline model, which does not capture correlations between interaction counterparts, indicates that directly anticipating these elements through multiple branches yields poor results. Our method outperforms comparative methods across all metrics for respective elements, demonstrating that leveraging correlations between humans and objects indeed benefits the comprehension of interaction relation, and the anticipation of interaction elements. It seems to be \"the best of both worlds\". Furthermore, we conduct a visual comparison of methods with higher evaluation metrics, as shown in Fig. 5. The results showcase multiple objects interacting with distinct human parts. As can be seen, our anticipations are more precise, for some uncertain regions that are not visible in images (e.g., the first and third rows), LEMON could also anticipate plausible results. This is attributed to the modeling of correlations between geometries, which compensates for missing features in invisible regions of the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 515, + 400, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 400, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 400, + 528 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "We conduct a thorough ablation study to validate the effectiveness of the model design. Tab. 2 (a) reports the model performance without modeling the semantic intention and the geometric correlation, demonstrating their impacts on the model performance. Moreover, the absence of constraining semantic consistency hinders the modeling of intentions, while not introducing curvatures affects the extraction of geometric correlations, resulting in a decrease in model performance. Both results are detailed in Tab. 2 (a). Besides, we test the impact of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pr}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " (Eq. 6) on spatial relation prediction, shown in Tab. 2 (b). Notably, removing the extraction of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " (Eq. 3) also affects the performance. This is attributed to the interrelation between " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", the absence of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_a" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " impacts the representation of " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi_c" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", and subsequently influencing the final spatial prediction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16289" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 64, + 527, + 392 + ], + "blocks": [ + { + "bbox": [ + 75, + 64, + 527, + 392 + ], + "lines": [ + { + "bbox": [ + 75, + 64, + 527, + 392 + ], + "spans": [ + { + "bbox": [ + 75, + 64, + 527, + 392 + ], + "type": "image", + "image_path": "a20c21f91381d99efe9e5904feeb24f7d34000f7e98e42d19b90af7e3659020b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 395, + 546, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 395, + 546, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 395, + 546, + 429 + ], + "type": "text", + "content": "Figure 5. Visualization Results. (a) Results of the estimated human vertices in contact with objects, the estimated contact vertices are shown in yellow. (b) The anticipations of 3D object affordance, the depth of red represents the probability of anticipated affordance. (c) Two views of the predicted spatial relation, translucent spheres are object proxies. Please zoom in for a better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 581 + ], + "type": "text", + "content": "To further illustrate the effectiveness of capturing the intention and geometry affinity, we visualize attention maps on the human and object geometries when removing one of them, as shown in Fig. 6. The first row indicates without capturing intention affinity, solely relying on geometries leads the model to focus on multiple candidate regions, differing from attention regions in the image. Furthermore, the second row demonstrates that while semantic intention guides the model to identify approximate candidate regions, modeling geometry affinity further locates corresponding regions most related to the interaction of the geometries." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 173, + 600 + ], + "type": "text", + "content": "5.4. Performance Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Multiple Interactions. To validate whether the model reasons interaction elements based on the understanding of interaction relation. We employ the model to infer different interactions between humans and the same object, as demonstrated in Fig. 7. The results showcase that, for the same object, when interactions are different, the model outputs distinct results, encompassing human contact, object affordance, and spatial relation, all of which align with the interaction relation depicted in the image. This indicates" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 307, + 501, + 543, + 608 + ], + "blocks": [ + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "lines": [ + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "type": "text", + "content": "Table 2. Ablation studies. (a) Performance when not modeling intention representations (intent.), geometric correlations (geom.), and not introducing " + }, + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "type": "text", + "content": " (Eq. 1), curvatures (cur.). (b) The impact of several factors on spatial relation prediction. " + }, + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 305, + 448, + 546, + 492 + ], + "type": "text", + "content": " means without." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 501, + 543, + 608 + ], + "lines": [ + { + "bbox": [ + 307, + 501, + 543, + 608 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 543, + 608 + ], + "type": "table", + "html": "
MetricsOursX intent.X φX geom.X cur.
Precision0.780.710.740.680.75
Recall0.820.730.790.700.78
F10.780.700.740.670.75
geo. (cm)7.5511.8710.2614.879.13
AUC88.5185.8786.6683.2386.98
aIOU41.3438.1940.0337.1339.58
SIM0.640.580.600.550.62
MSE0.0100.0270.0220.0310.018
", + "image_path": "e604793fd5c9934406956a42c75bc4d88e2e4177c3ba8fcc652f6d9036ad0c05.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 398, + 609, + 408, + 617 + ], + "lines": [ + { + "bbox": [ + 398, + 609, + 408, + 617 + ], + "spans": [ + { + "bbox": [ + 398, + 609, + 408, + 617 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 627, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 627, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 545, + 663 + ], + "type": "text", + "content": "that the model infers interaction elements according to the interaction relation rather than relying on direct mapping solely driven by the categories of human-object pairs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Multiple Objects. In scenarios where distinct objects are involved in concurrent interactions with humans, the model should indeed possess the ability to comprehend the interaction relation with different objects. We give an experi-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16290" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 279, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 279, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 279, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 279, + 196 + ], + "type": "image", + "image_path": "515f7af041520d8c6b974735fd94545fc0e8e41ea0e0159ddba15d9a355162a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "type": "text", + "content": "Figure 6. Visual Attention. Attention heatmaps on geometries, in cases with " + }, + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w})" + }, + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "type": "text", + "content": " and without " + }, + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{w} / \\boldsymbol{o})" + }, + { + "bbox": [ + 47, + 198, + 287, + 232 + ], + "type": "text", + "content": " the extraction of intention and geometry affinity. Images are also shown with attention heatmaps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 246, + 91, + 281 + ], + "blocks": [ + { + "bbox": [ + 56, + 236, + 89, + 246 + ], + "lines": [ + { + "bbox": [ + 56, + 236, + 89, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 89, + 246 + ], + "type": "text", + "content": "Support" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 246, + 91, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 91, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 91, + 281 + ], + "type": "image", + "image_path": "367870f5c15a8499a823117baf1d9193b479f584a54d77fc99b83404fe10d731.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 95, + 235, + 132, + 281 + ], + "blocks": [ + { + "bbox": [ + 95, + 235, + 132, + 281 + ], + "lines": [ + { + "bbox": [ + 95, + 235, + 132, + 281 + ], + "spans": [ + { + "bbox": [ + 95, + 235, + 132, + 281 + ], + "type": "image", + "image_path": "fbd7bea0d50476f68bb0d79d1ecd88a8f30b05cc1fb3b866f86d355bf51eb1c6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 137, + 239, + 178, + 281 + ], + "blocks": [ + { + "bbox": [ + 137, + 239, + 178, + 281 + ], + "lines": [ + { + "bbox": [ + 137, + 239, + 178, + 281 + ], + "spans": [ + { + "bbox": [ + 137, + 239, + 178, + 281 + ], + "type": "image", + "image_path": "7f5534de2937f241c63a4f0151bc9bbf5e1d9ab3fd1923f33f1a9a5b7f5af81c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 187, + 243, + 224, + 289 + ], + "blocks": [ + { + "bbox": [ + 194, + 234, + 218, + 241 + ], + "lines": [ + { + "bbox": [ + 194, + 234, + 218, + 241 + ], + "spans": [ + { + "bbox": [ + 194, + 234, + 218, + 241 + ], + "type": "text", + "content": "View 1" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 243, + 224, + 289 + ], + "lines": [ + { + "bbox": [ + 187, + 243, + 224, + 289 + ], + "spans": [ + { + "bbox": [ + 187, + 243, + 224, + 289 + ], + "type": "image", + "image_path": "a281632b63ea86cdf1be7cb93a84780708011e603f022d610ae361bed98de558.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 230, + 243, + 280, + 288 + ], + "blocks": [ + { + "bbox": [ + 230, + 234, + 271, + 241 + ], + "lines": [ + { + "bbox": [ + 230, + 234, + 271, + 241 + ], + "spans": [ + { + "bbox": [ + 230, + 234, + 271, + 241 + ], + "type": "text", + "content": "View 2" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 243, + 280, + 288 + ], + "lines": [ + { + "bbox": [ + 230, + 243, + 280, + 288 + ], + "spans": [ + { + "bbox": [ + 230, + 243, + 280, + 288 + ], + "type": "image", + "image_path": "113dbefaef8cfdb07d5f5f3ab8600f855d04a286b9f904755a888d99701fb366.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 55, + 302, + 91, + 336 + ], + "blocks": [ + { + "bbox": [ + 55, + 302, + 91, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 91, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 91, + 336 + ], + "type": "image", + "image_path": "d471bb0dec73e5b7056b161ab935e6b01e43bee894e119937712d155d32d8b7a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 336, + 84, + 344 + ], + "lines": [ + { + "bbox": [ + 63, + 336, + 84, + 344 + ], + "spans": [ + { + "bbox": [ + 63, + 336, + 84, + 344 + ], + "type": "text", + "content": "Carry" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 95, + 293, + 132, + 340 + ], + "blocks": [ + { + "bbox": [ + 95, + 293, + 132, + 340 + ], + "lines": [ + { + "bbox": [ + 95, + 293, + 132, + 340 + ], + "spans": [ + { + "bbox": [ + 95, + 293, + 132, + 340 + ], + "type": "image", + "image_path": "283630f3bc4effd236f3887b7ef621b8552028f67677d97d05f18c8181053f80.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 341, + 136, + 349 + ], + "lines": [ + { + "bbox": [ + 96, + 341, + 136, + 349 + ], + "spans": [ + { + "bbox": [ + 96, + 341, + 136, + 349 + ], + "type": "text", + "content": "Affordance" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 137, + 293, + 186, + 341 + ], + "blocks": [ + { + "bbox": [ + 137, + 293, + 186, + 341 + ], + "lines": [ + { + "bbox": [ + 137, + 293, + 186, + 341 + ], + "spans": [ + { + "bbox": [ + 137, + 293, + 186, + 341 + ], + "type": "image", + "image_path": "5e71f181a0c26d57af78c846cacd36e5a9dd319c3f4e4e4fb30c5e58e9183a74.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 341, + 179, + 349 + ], + "lines": [ + { + "bbox": [ + 151, + 341, + 179, + 349 + ], + "spans": [ + { + "bbox": [ + 151, + 341, + 179, + 349 + ], + "type": "text", + "content": "Contact" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 351, + 286, + 373 + ], + "lines": [ + { + "bbox": [ + 47, + 351, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 351, + 286, + 373 + ], + "type": "text", + "content": "Figure 7. Multiple Interactions. Distinct anticipations when interactions are different with the same object." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 187, + 292, + 230, + 340 + ], + "blocks": [ + { + "bbox": [ + 187, + 292, + 230, + 340 + ], + "lines": [ + { + "bbox": [ + 187, + 292, + 230, + 340 + ], + "spans": [ + { + "bbox": [ + 187, + 292, + 230, + 340 + ], + "type": "image", + "image_path": "fc43781236193e6c6217674838b5d323063ffa91653228ed98c7093f83a108eb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 341, + 264, + 349 + ], + "lines": [ + { + "bbox": [ + 206, + 341, + 264, + 349 + ], + "spans": [ + { + "bbox": [ + 206, + 341, + 264, + 349 + ], + "type": "text", + "content": "Spatial Relation" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 491 + ], + "type": "text", + "content": "ment pertaining to this property. As shown in Fig. 8, when reasoning with different objects, the human contact, object affordance, and spatial relation are distinct, while all anticipations are plausible. This is attributed to the collaboration of the semantic interaction intention and geometric correlation, ensuring that the model could explicitly capture interaction intentions with different objects and reveal regions related to the interaction on corresponding geometries." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 659 + ], + "type": "text", + "content": "Multiple Instances. We conduct an experiment to validate the model's understanding of interaction relation across various instances, assessing its generalization and robustness. Specifically, distinct object instances are utilized to infer with the same image and human geometry, shown in Fig. 9 (a). As can be seen, even though the contact and spatial relation are slightly different, the results are consistent with the interaction shown in the image, and object affordances are correctly anticipated. Additionally, the same object instance is utilized to infer with different human geometries and images that contain similar human-object interactions, as demonstrated in Fig. 9 (b). These results indicate that the model could generalize the understanding of interaction relation to different human and object instances." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We propose leveraging both counterparts of the HOI to jointly anticipate human contact, object affordance, and" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 538, + 178 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 538, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 538, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 538, + 178 + ], + "type": "image", + "image_path": "2f89380510429d19137e6c96618c4bb3935c2f989b3963653646aab9e2670850.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "type": "text", + "content": "Figure 8. Multiple Objects. The anticipations for multiple objects with different human-object interactions." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 313, + 208, + 538, + 363 + ], + "blocks": [ + { + "bbox": [ + 313, + 208, + 538, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 538, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 538, + 363 + ], + "type": "image", + "image_path": "846879d2b64aa144c0162fce37c6ccc750690096cbae6408e2af7478237d2bd0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "lines": [ + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "type": "text", + "content": "Figure 9. Multiple Instances. (a) Different object instances " + }, + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "type": "inline_equation", + "content": "w.r.t." + }, + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "type": "text", + "content": " the same image and human geometry. (b) Different images and human instances " + }, + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "type": "inline_equation", + "content": "w.r.t." + }, + { + "bbox": [ + 306, + 365, + 545, + 398 + ], + "type": "text", + "content": " the same object." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 430, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 574 + ], + "type": "text", + "content": "human-object spatial relation in 3D space. It holds the potential to drive embodied intelligence to learn from interactions and assist the interaction modeling in graphics. With the provided dataset 3DIR containing paired HOI data and multiple annotations, we train LEMON, a novel model that harnesses the interaction intention and geometric correlation between humans and objects to capture the affinity, eliminating interaction uncertainties and anticipating plausible interaction elements. Extensive experiments demonstrate the effectiveness and superiority of LEMON. We believe this work offers fresh insights and paves a new way for 3D human-object interaction understanding." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 674 + ], + "type": "text", + "content": "Limitations and Future Work. In the current formulation, LEMON needs the pre-inferred human mesh as input. In the future, we plan to integrate HMR into the entire framework, freeing input constraints required for inference, and taking 3D interaction elements to improve the accuracy of HMR. Plus, another interesting direction is to leverage multi-modal methods [15, 88], boosting the interaction relation understanding from other sources e.g., text, audio." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgments. This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207 and 62306295." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16291" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Brenna D Argall, Sonia Chernova, Manuela Veloso, and Brett Browning. A survey of robot learning from demonstration. Robotics and autonomous systems, 57(5):469-483, 2009. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 286, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 286, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 286, + 190 + ], + "type": "text", + "content": "[2] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 286, + 246 + ], + "type": "text", + "content": "[3] Antonio Bicchi and Vijay Kumar. Robotic grasping and contact: A review. In Proceedings 2000 ICRA. Millennium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), pages 348-353. IEEE, 2000. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 290 + ], + "type": "text", + "content": "[4] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018, the winter conference on applications of computer vision (wacv), pages 381–389. IEEE, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 286, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 286, + 357 + ], + "type": "text", + "content": "[5] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 286, + 401 + ], + "type": "text", + "content": "[6] Yixin Chen, Sai Kumar Dwivedi, Michael J. Black, and Dimitrios Tzionas. Detecting human-object contact in images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 445 + ], + "type": "text", + "content": "[7] Kun-Hung Cheng and Chin-Chung Tsai. Affordances of augmented reality in science learning: Suggestions for future research. Journal of science education and technology, 22:449-462, 2013. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 286, + 512 + ], + "type": "text", + "content": "[8] Henry M Clever, Zackory Erickson, Ariel Kapusta, Greg Turk, Karen Liu, and Charles C Kemp. Bodies at rest: 3d human pose and shape estimation from a pressure image using synthetic data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6215-6224, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 578 + ], + "type": "text", + "content": "[9] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 286, + 635 + ], + "type": "text", + "content": "[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 1, 3, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 679 + ], + "type": "text", + "content": "[11] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[12] Kuan Fang, Te-Lin Wu, Daniel Yang, Silvio Savarese, and Joseph J Lim. Demo2vec: Reasoning object affordances from online videos. In Proceedings of the IEEE Conference" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 2139-2147, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 162 + ], + "type": "text", + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 186 + ], + "type": "text", + "content": "[14] James J Gibson. The ecological approach to visual perception: classic edition. Psychology press, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 186, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 251 + ], + "type": "text", + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15180-15190, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 297 + ], + "type": "text", + "content": "[16] Georgia Gkioxari, Ross Girshick, Piotr Dólar, and Kaiming He. Detecting and recognizing human-object interactions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8359-8367, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 354 + ], + "type": "text", + "content": "[17] Elena Corina Grigore, Kerstin Eder, Anthony G Pipe, Chris Melhuish, and Ute Leonards. Joint action understanding improves robot-to-human object handover. In 2013 IEEE/RSJ international conference on intelligent robots and systems, pages 4622-4629. IEEE, 2013. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 376 + ], + "type": "text", + "content": "[18] Saurabh Gupta and Jitendra Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 422 + ], + "type": "text", + "content": "[19] Sookwan Han and Hanbyul Joo. Learning canonicalized 3d human-object spatial relations from unbounded synthesized images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 423, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 423, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 423, + 545, + 478 + ], + "type": "text", + "content": "[20] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14708-14718, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "text", + "content": "[21] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Proceedings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274-13285, 2022. 1, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "type": "text", + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 667 + ], + "type": "text", + "content": "[23] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[24] Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2):1-35, 2017. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16292" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[26] Jan J Koenderink and Andrea J Van Doorn. Surface shape and curvature scales. Image and vision computing, 10(8): 557-564, 1992. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[27] Nilesh Kulkarni, Davis Rempe, Kyle Genova, Abhijit Kundu, Justin Johnson, David Fouhey, and Leonidas Guibas. Nifty: Neural object interaction fields for guided human motion synthesis, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "text", + "content": "[28] Hongxiang Li, Meng Cao, Xuxin Cheng, Yaowei Li, Zhihong Zhu, and Yuexian Zou. G2l: Semantically aligned and uniform video grounding via geodesic and game theory. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12032-12042, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 288, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 286 + ], + "type": "text", + "content": "[29] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. arXiv preprint arXiv:2309.16237, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 331 + ], + "type": "text", + "content": "[30] Lulei Li, Jianan Wei, Wenguan Wang, and Yi Yang. Neural-logic human-object interaction detection. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "text", + "content": "[31] Yong-Lu Li, Xinpeng Liu, Han Lu, Shiyi Wang, Junqi Liu, Jiefeng Li, and Cewu Lu. Detailed 2d-3d joint representation for human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10166–10175, 2020. 1, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 453 + ], + "type": "text", + "content": "[32] Yong-Lu Li, Liang Xu, Xinpeng Liu, Xijie Huang, Yue Xu, Shiyi Wang, Hao-Shu Fang, Ze Ma, Mingyang Chen, and Cewu Lu. Pastanet: Toward human activity knowledge engine. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 382-391, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[33] Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. One-stage 3d whole-body mesh recovery with component aware transformer. CVPR, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[34] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "type": "text", + "content": "[35] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. AKB-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14809–14818, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 635 + ], + "type": "text", + "content": "[36] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851-866. 2023. 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[37] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2252-2261, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[38] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Leverage interactive affinity for affordance" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6809-6819, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "text", + "content": "[39] Bertram F Malle and Joshua Knobe. The folk concept of intentionality. Journal of experimental social psychology, 33 (2):101-121, 1997. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 186 + ], + "type": "text", + "content": "[40] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE international conference on robotics and automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 241 + ], + "type": "text", + "content": "[41] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3651-3660, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 296 + ], + "type": "text", + "content": "[42] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 547, + 364 + ], + "type": "text", + "content": "[43] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large-scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 547, + 420 + ], + "type": "text", + "content": "[44] Kaichun Mo, Leonidas J Guibas, Mustafa Mukadam, Abhinav Gupta, and Shubham Tulsiani. Where2act: From pixels to actions for articulated 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6813-6823, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 547, + 466 + ], + "type": "text", + "content": "[45] Kaichun Mo, Yuzhe Qin, Fanbo Xiang, Hao Su, and Leonidas Guibas. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on Robot Learning, pages 1666-1677. PMLR, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 547, + 510 + ], + "type": "text", + "content": "[46] Tushar Nagarajan and Kristen Grauman. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33:2005-2015, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 547, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 547, + 556 + ], + "type": "text", + "content": "[47] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 547, + 600 + ], + "type": "text", + "content": "[48] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai Nguyen. Detecting hands and recognizing physical contact in the wild. Advances in neural information processing systems, 33:7841-7851, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 646 + ], + "type": "text", + "content": "[49] Abdul Nurunnabi, Geoff West, and David Belton. Outlier detection and robust normal-curvature estimation in mobile laser scanning 3d point cloud data. Pattern Recognition, 48 (4):1404-1419, 2015. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 712 + ], + "type": "text", + "content": "[50] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16293" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[51] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4726-4736, 2023. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[52] Ilya A Petrov, Riccardo Marin, Julian Chibane, and Gerard Pons-Moll. Object pop-up: Can we infer 3d objects and their poses from human interactions alone? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[53] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 292 + ], + "type": "text", + "content": "[54] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12802–12813, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 288, + 358 + ], + "type": "text", + "content": "[55] Davis Rempe, Leonidas J Guibas, Aaron Hertzmann, Bryan Russell, Ruben Villegas, and Jamei Yang. Contact and human dynamics from monocular video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 71-87. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 415 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 458 + ], + "type": "text", + "content": "[57] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 525 + ], + "type": "text", + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9339-9347, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[59] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 581, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 602 + ], + "type": "text", + "content": "[60] Thomas B Sheridan. Human-robot interaction: status and challenges. Human factors, 58(4):525-532, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 635 + ], + "type": "text", + "content": "[61] Chris Shilling. The body in culture, technology and society. The Body in Culture, Technology and Society, pages 1-256, 2004. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[62] Soshi Shimada, Vladislav Golyanik, Zhi Li, Patrick Pérez, Weipeng Xu, and Christian Theobalt. Hulc: 3d human motion capture with pose manifold sampling and dense contact guidance. In European Conference on Computer Vision, pages 516-533. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[63] Michael Tomasello, Malinda Carpenter, Josep Call, Tanya Behne, and Henrik Moll. Understanding and sharing in-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tentions: The origins of cultural cognition. Behavioral and brain sciences, 28(5):675-691, 2005. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "text", + "content": "[64] Shashank Tripathi, Agniv Chatterjee, Jean-Claude Passy, Hongwei Yi, Dimitrios Tzionas, and Michael J. Black. DECO: Dense estimation of 3D human-scene contact in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8001-8013, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 228 + ], + "type": "text", + "content": "[65] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 274 + ], + "type": "text", + "content": "[66] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[67] Yian Wang, Ruihai Wu, Kaichun Mo, Jiaqi Ke, Qingnan Fan, Leonidas J Guibas, and Hao Dong. Adaafford: Learning to adapt manipulation affordance for 3d articulated objects via few-shot interactions. In European Conference on Computer Vision, pages 90-107. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 376 + ], + "type": "text", + "content": "[68] Ping Wei, Yibiao Zhao, Nanning Zheng, and Song-Chun Zhu. Modeling 4d human-object interactions for event and object recognition. In Proceedings of the IEEE international conference on computer vision, pages 3272-3279, 2013. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 410 + ], + "type": "text", + "content": "[69] Ruihai Wu, Chuanruo Ning, and Hao Dong. Learning foresightful dense visual affordance for deformable object manipulation. arXiv preprint arXiv:2303.11057, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 455 + ], + "type": "text", + "content": "[70] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. arXiv preprint arXiv:2309.07918, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "text", + "content": "[71] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[72] Bingjie Xu, Yongkang Wong, Junnan Li, Qi Zhao, and Mohan S Kankanhalli. Learning to detect human-object interactions with knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 600 + ], + "type": "text", + "content": "[73] Chao Xu, Yixin Chen, He Wang, Song-Chun Zhu, Yixin Zhu, and Siyuan Huang. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519, 2022.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "text", + "content": "[74] Rui Xu, Zhiyang Dou, Ningna Wang, Shiqing Xin, Shuang-min Chen, Mingyan Jiang, Xiaohu Guo, Wenping Wang, and Changhe Tu. Globally consistent normal orientation for point clouds by regularizing the winding-number field. ACM Transactions on Graphics (TOG), 42(4):1-15, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[75] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16294" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[76] Xiang Xu, Hanbyul Joo, Greg Mori, and Manolis Savva. D3d-hoi: Dynamic 3d human-object interactions from videos. arXiv preprint arXiv:2108.08420, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "text", + "content": "[77] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 218 + ], + "type": "text", + "content": "[78] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10905-10915, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 252 + ], + "type": "text", + "content": "[79] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. arXiv preprint arXiv:2307.15880, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 287, + 308 + ], + "type": "text", + "content": "[80] Bangpeng Yao and Li Fei-Fei. Modeling mutual context of object and human pose in human-object interaction activities. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 17–24. IEEE, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 287, + 354 + ], + "type": "text", + "content": "[81] Sifan Ye, Yixing Wang, Jiaman Li, Dennis Park, C Karen Liu, Huazhe Xu, and Jiajun Wu. Scene synthesis from human motion. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 49, + 355, + 287, + 409 + ], + "type": "text", + "content": "[82] Hongwei Yi, Chun-Hao P Huang, Shashank Tripathi, Lea Hering, Justus Thies, and Michael J Black. Mime: Human-aware 3d scene generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12965-12976, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 453 + ], + "type": "text", + "content": "[83] Zhao-Heng Yin, Binghao Huang, Yuzhe Qin, Qifeng Chen, and Xiaolong Wang. Rotating without seeing: Towards in-hand dexterity through touch. arXiv preprint arXiv:2303.10880, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "type": "text", + "content": "[84] Guochen Yu, Andong Li, Chengshi Zheng, Yinuo Guo, Yutian Wang, and Hui Wang. Dual-branch attention-inattention transformer for single-channel speech enhancement. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 7847-7851. IEEE, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 566 + ], + "type": "text", + "content": "[85] Wei Zhai, Hongchen Luo, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot object affordance detection in the wild. International Journal of Computer Vision, 130(10): 2472-2500, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 568, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 621 + ], + "type": "text", + "content": "[86] Wei Zhai, Pingyu Wu, Kai Zhu, Yang Cao, Feng Wu, and Zheng-Jun Zha. Background activation suppression for weakly supervised object localization and semantic segmentation. International Journal of Computer Vision, pages 1-26, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "text", + "content": "[87] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 34-51. Springer, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 714 + ], + "type": "text", + "content": "[88] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 440 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8552-8562, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[89] Xiaopeng Zhang, Hongjun Li, Zhanglin Cheng, and Yikuan Zhang. Robust curvature estimation and geometry analysis of 3d point cloud surfaces. J. Inf. Comput. Sci, 6(5):1983-1990, 2009. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 195 + ], + "type": "text", + "content": "[90] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In International conference on computer vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 250 + ], + "type": "text", + "content": "[91] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper object manipulation. arXiv preprint arXiv:2207.01971, 2022.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 545, + 306 + ], + "type": "text", + "content": "[92] Yan Zhao, Ruihai Wu, Zhehuan Chen, Yourong Zhang, Qingnan Fan, Kaichun Mo, and Hao Dong. Dualafford: Learning collaborative visual affordance for dual-gripper manipulation. In International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 373 + ], + "type": "text", + "content": "[93] Xinghao Zhu, JingHan Ke, Zhixuan Xu, Zhixin Sun, Bizhe Bai, Jun Lv, Qingtao Liu, Yuwei Zeng, Qi Ye, Cewu Lu, et al. Diff-lfd: Contact-aware model-based learning from visual demonstration for robotic manipulation via differentiable physics-based simulation and rendering. In 7th Annual Conference on Robot Learning, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 375, + 545, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 375, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 308, + 375, + 545, + 440 + ], + "type": "text", + "content": "[94] Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with hoi transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11825-11834, 2021. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16295" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_content_list.json b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3b83f4c766ec38e3ecfda82b69a010429b818184 --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_content_list.json @@ -0,0 +1,1464 @@ +[ + { + "type": "text", + "text": "LEOD: Label-Efficient Object Detection for Event Cameras", + "text_level": 1, + "bbox": [ + 181, + 130, + 785, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyi Wu $^{1,2}$ , Mathias Gehrig $^{3}$ , Qing Lyu $^{1}$ , Xudong Liu $^{1}$ , Igor Gilitschenski $^{1,2}$ , $^{1}$ University of Toronto, $^{2}$ Vector Institute, $^{3}$ University of Zurich", + "bbox": [ + 181, + 179, + 782, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 253, + 313, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection with event cameras benefits from the sensor's low latency and high dynamic range. However, it is costly to fully label event streams for supervised training due to their high temporal resolution. To reduce this cost, we present LEOD, the first method for label-efficient event-based detection. Our approach unifies weakly- and semi-supervised object detection with a self-training mechanism. We first utilize a detector pre-trained on limited labels to produce pseudo ground truth on unlabeled events. Then, the detector is re-trained with both real and generated labels. Leveraging the temporal consistency of events, we run bi-directional inference and apply tracking-based post-processing to enhance the quality of pseudo labels. To stabilize training against label noise, we further design a soft anchor assignment strategy. We introduce new experimental protocols to evaluate the task of label-efficient event-based detection on Gen1 and 1Mpx datasets. LEOD consistently outperforms supervised baselines across various labeling ratios. For example, on Gen1, it improves mAP by $8.6\\%$ and $7.8\\%$ for RVT-S trained with $1\\%$ and $2\\%$ labels. On 1Mpx, RVT-S with $10\\%$ labels even surpasses its fully-supervised counterpart using $100\\%$ labels. LEOD maintains its effectiveness even when all labeled data are available, reaching new state-of-the-art results. Finally, we show that our method readily scales to improve larger detectors as well. Code: https://github.com/Wuziyi616/LEOD.", + "bbox": [ + 76, + 285, + 472, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 207, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection is key to scene understanding. It provides a compact representation of raw sensor measurements as semantically meaningful bounding boxes. Speed is crucial in object detection, especially in safety-critical applications such as self-driving. Recently, event cameras have gained significant interest in computer vision due to their low latency, low energy consumption, and high dynamic range [11]. Leveraging these benefits, event-based object detectors [14, 16, 29, 42, 49, 76] have been developed to complement conventional frame-based detectors. Despite tremendous progress, much of their success heavily relies", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8a44feb843cafde48f4053b7165d8de07b65daa6095ec07000be7c7300fd453.jpg", + "image_caption": [ + "Figure 1. Detection performance of LEOD and baselines trained only on labeled events or conducting naive self-training. Under the weakly-supervised setting, our method consistently improves the RVT-S detector [14] across all labeling ratios on the Gen1 dataset." + ], + "image_footnote": [], + "bbox": [ + 498, + 241, + 890, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on large datasets that are manually annotated. However, due to the high temporal resolution of event data, labeling objects at every timestamp is impractical. For example, Gen1 dataset [8] provides object labels at lower than $4\\mathrm{Hz}$ . As a result, existing methods only train their models on labeled events, and discard the remaining unlabeled data. In contrast, we view this task as a weakly-supervised learning problem, which calls for detection methods that can leverage a mix of labeled and unlabeled events during training.", + "bbox": [ + 496, + 491, + 892, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we address the challenge by proposing a Label-efficient Event-based Object Detection (LEOD) framework. We consider two settings characterized by limited labels: (i) weakly-supervised, where object bounding boxes are sparsely labeled in all event streams, and (ii) semi-supervised, where some event streams have object boxes densely labeled while others remain fully unlabeled. Our approach unifies the two settings through a self-training paradigm. With limited labels, we first pre-train a detector and use it to generate pseudo annotations on unlabeled events. Then, we re-train the detector on a combination of real and pseudo labels. However, naively generated labels contain noise, making their direct use suboptimal.", + "bbox": [ + 496, + 628, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To obtain high-quality pseudo labels, we exploit the temporal dimension of event data. Recent work [29] has shown the importance of temporal information in event-based detection using recurrent modules [52]. Additionally, offline label generation enables us to further refine predictions with", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "future information. To achieve this, we first introduce time-flip augmentation to events during training. As a result, we can ensemble model predictions on the original and reverse event streams via Test-Time Augmentation (TTA), leading to higher detection recall. In addition, we leverage tracking-based post-processing to eliminate temporally inconsistent objects, enhancing the precision of pseudo labels. Finally, we filter out low-confidence boxes with a score threshold.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key challenge here is how to select a proper threshold. Instead of searching for the optimal value, we first filter with a low threshold to avoid missing objects. This inevitably introduces many false positives, which we address with a soft anchor assignment strategy in training. When computing the detection loss, we set another higher threshold and only treat pseudo boxes above that threshold as positive labels. For boxes with a lower detection score, we ignore the loss applied to their associated anchors. This strategy ensures that the model is only supervised with reliable background and foreground labels, while being tolerant to noisy labels. Ablation studies show that our method is insensitive to the two threshold values, easing the hyper-parameter tuning.", + "bbox": [ + 75, + 214, + 470, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To test our method, we design new protocols for label-efficient event-based detection on Gen1 [8] and 1Mpx [42] datasets. For weakly-supervised object detection (WSOD), we uniformly sub-sample the labels over time to simulate sparse annotations. For semi-supervised object detection (SSOD), we directly choose some event sequences as fully unlabeled. Following 2D SSOD evaluation protocols [55, 65], we also have a fully labeled setting, where we show that pseudo labels can complement ground-truths.", + "bbox": [ + 75, + 412, + 468, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work makes four main contributions: (i) We introduce the task of label-efficient object detection to event vision, and design its experimental protocols. (ii) We propose LEOD, a unified framework for training event-based detectors with limited annotations. (iii) LEOD consistently outperforms baselines in various settings on two public detection datasets. (iv) Our method remains effective under the fully labeled setting and scales up to larger detectors, achieving new state-of-the-art performance.", + "bbox": [ + 75, + 551, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 705, + 218, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Detection with Event Cameras. Existing event-based detectors can be mainly categorized into two classes depending on whether they utilize the asynchronous nature of events. One line of work explores the sparsity of events, and employs Graph Neural Networks (GNNs) [13, 36, 49] or Spiking Neural Networks (SNNs) [7, 21, 70] for feature extraction. However, these approaches struggle with propagating information over long time horizons, which is crucial for detecting objects from events. Moreover, specialized hardware is required to achieve theoretical speed-ups of sparse networks, limiting their application in practice.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In another class of methods, events are converted to dense frame-like representations, followed by conventional networks for detection. Earlier works only consider event frames aggregated from a short time interval [5, 20, 22, 31]. This discards long-horizon history and makes it hard to detect objects under small relative motion as they trigger very few events. Recent methods thus introduce recurrent modules [17, 52] to enhance the memory of the detectors [14, 29, 42]. Further research focuses on better backbones [9, 58], inference speed [16], and event representations [76]. Because our primary goal is to study label-efficient learning for event-based detectors, we adopt the state-of-the-art approach RVT [14] as the base model.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Label-Efficient Learning in Event-based Vision. Due to a lack of large labeled datasets, there have been several works studying event-based algorithms with limited labels. Some papers focus on bridging frame-based and event-based vision. They either reconstruct natural images from events to apply traditional deep models [46, 47, 50, 56, 62], simulate events from videos to transfer the annotations [19, 39, 45, 75], or distill knowledge from trained frame-based models [18, 37, 57, 60, 67, 69]. However, these methods require either paired recordings of events and images or massive in-domain labeled images for training. Closer to ours are methods that only use event data [6, 26, 63, 74]. They conduct label-efficient learning on events with pre-trained frame-based models or self-supervised losses. Yet, none of them are designed for the detection task. Our work is the first attempt at label-efficient event-based object detection.", + "bbox": [ + 496, + 287, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Label-Efficient Learning in Other Fields. Self-training based methods have been explored in tasks such as 2D image classification [27, 54, 64], object detection [55, 65, 72], and segmentation [1, 23, 41]. Our method is more related to label-efficient learning on videos [38, 53, 66] and 3D point cloud sequences [32, 44, 71] as these methods also exploit the temporal information of input data. For example, [66] utilizes optical flows to propagate single-frame labels to adjacent video frames. [44] and [32] train a teacher model on dense point clouds aggregated from a few timesteps. In contrast, we leverage a much longer temporal horizon by running the detector on the entire event stream in both directions. [71] also employs tracking-based post-processing to remove inconsistent boxes. We additionally perform tracking in both directions as a forward-backward consistency check, thus better leveraging the temporal information of event data. To tackle the noisy pseudo labels, [65] proposes to use the detection scores from the teacher model to weigh the loss. Instead, we design a soft anchor assignment strategy by ignoring the loss associated with unconfident boxes.", + "bbox": [ + 496, + 529, + 890, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 844, + 589, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper introduces a new task named label-efficient event-based object detection (formulated in Sec. 3.1). Our", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16934", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "algorithm adopts a two-stage self-training framework with reliable label selection (Sec. 3.2), where we leverage the temporal information of event data to obtain high-quality pseudo labels and suppress noisy predictions (Sec. 3.3).", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 162, + 279, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Events data. Event cameras record brightness changes, and output a sequence of events $\\mathcal{E} = \\{e_i = (x_i,y_i,t_i,p_i)\\}$ . Each event $e_i$ is parameterized by its pixel coordinate $(x_{i},y_{i})$ , timestamp $t_i$ , and polarity $p_i\\in \\{-1,1\\}$ . Modern event cameras run at sub-milliseconds and can produce millions of events per second [11].", + "bbox": [ + 75, + 186, + 468, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based Object Detection. As the object motion in a scene is usually much slower than the event generation speed, event-based object detectors are only applied and evaluated at a fixed time interval $T$ [8, 42]. More specifically, given an event stream $\\mathcal{E}$ capturing a set of objects $\\mathcal{O} = \\{o_j\\}_{j=1}^M$ , we aim at detecting the 2D bounding boxes of them with the semantic labels, $\\mathcal{B} = \\{b_j = (x_j, y_j, w_j, h_j, l_j, t_l)\\}_{j=1}^M$ . Each bounding box $b_j$ is characterized by the location of its top-left corner $(x_j, y_j)$ , width $w_j$ , height $h_j$ , class label $l_j \\in \\{1, 2, \\dots, C\\}$ , and timestamp $t_j$ . Here, $C$ denotes the number of classes.", + "bbox": [ + 75, + 277, + 468, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based Object Detectors. We take RVT [14] for example as it serves as our base model. RVT is a synchronous detector that converts events in every time window $\\Delta t$ to a grid-like representation $I$ . In the remaining part of the paper, we call $I$ a frame and every $\\Delta t$ a timestep. RVT combines a Vision Transformer backbone [58] with a YOLOX detection head [12]. To extract temporal features, RVT introduces LSTM [17] cells in the backbone to fuse information over multiple timesteps. This is useful for detecting slow-moving objects as they only generate a few events. The YOLOX detection head in RVT is anchor-free, i.e., for each location on the feature map (anchor point), it predicts an objectness score $p_{\\mathrm{obj}} \\in [0,1]$ , per-class IoU values $p_{\\mathrm{iou}} \\in \\mathbb{R}^C$ , and offsets of the bounding box parameters $\\Delta b = (\\Delta x, \\Delta y, \\Delta w, \\Delta h)$ . The $p_{\\mathrm{iou}}$ is trained to output the IoU value between the predicted box and the matched ground-truth box of that class. To obtain the final prediction, Non-Maximum Suppression (NMS) is applied to remove overlapping bounding boxes with low confidence.", + "bbox": [ + 75, + 445, + 468, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, each ground-truth bounding box is matched to several anchor points for loss computation. Below we will denote anchor points as anchors for simplicity. Anchors matched with ground-truth are foreground, where all predicted values are supervised. For the remaining background anchors, only $p_{\\mathrm{obj}}$ is trained to be 0.", + "bbox": [ + 75, + 733, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Label-Efficient Event-based Object Detection. Fig. 2 shows our proposed detection settings with limited labels.", + "bbox": [ + 75, + 824, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-supervised object detection (WSOD). In the WSOD setting, all event streams are sparsely labeled. Moreover, labels assigned to adjacent frames offer fewer informa", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bc51025cf0307266921a514e243a2d0f8c35e668a4a5433630c25c92fbfb5860.jpg", + "image_caption": [ + "(a) Weakly-supervised object detection (WSOD)" + ], + "image_footnote": [], + "bbox": [ + 500, + 60, + 888, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3479030a55d255a4163e18d09ebfb84abe78cc46863567e54e430b58dbad36d5.jpg", + "image_caption": [ + "(b) Semi-supervised object detection (SSOD)", + "Figure 2. Illustration of two label-efficient event-based object detection settings: (a) weakly-supervised where all event sequences are sparsely annotated, and (b) semi-supervised where some event sequences are densely annotated, and others are fully unlabeled. We visualize both positive and negative events in black. Tive training signals compared to those distributed across frames [66]. Therefore, it is reasonable to label object boxes in a long event sequence uniformly and sparsely." + ], + "image_footnote": [], + "bbox": [ + 500, + 229, + 888, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semi-supervised object detection (SSOD). In the SSOD setting, some event sequences are densely labeled, while others are fully unlabeled. This is also practical when people continue to collect data into already annotated datasets. Since capturing event sequences is much easier than labeling them, an algorithm that can consistently improve model performance with raw events is highly useful.", + "bbox": [ + 496, + 521, + 890, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To evaluate the label-efficient learning performance, we take existing event-based object detection datasets [8, 42] and sample a small portion of frames (WSOD) or sequences (SSOD) as labeled data. The rest of the training data are used as an unlabeled set following previous works [33, 55, 65]. We also have a fully labeled setting where all labels are available. Since the original event streams are annotated at a larger time interval than $\\Delta t$ , we can still create pseudo labels on unlabeled timesteps to improve the performance. See Appendix D for further discussions on the two settings.", + "bbox": [ + 496, + 627, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LEOD: A Self-Training Framework", + "text_level": 1, + "bbox": [ + 500, + 786, + 810, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the overall pipeline of LEOD follows a student-teacher pseudo-labeling paradigm, which is applicable to both, the WSOD and the SSOD settings. We first pre-train a detector on labeled data using regular detection loss until convergence. Then, we employ it to annotate unlabeled frames. To leverage the benefit of offline prediction,", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16935", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9c41f6aca91a250db29d949f478ccc483c209d2bddbd46bb5828efe80bfb3183.jpg", + "image_caption": [ + "Figure 3. Overview of our LEOD pipeline. $①$ We first pre-train an event-based object detector on event streams with limited labels. $①$ To leverage the temporal information, we apply time-flip Test-Time Augmentation (TTA) to unlabeled event streams and ensemble the model predictions. $②$ We then apply forward and backward tracking to identify temporally inconsistent bounding boxes, i.e., boxes associated with short tracks. $③$ To handle noisy labels, a soft anchor assignment strategy is designed to ignore detection loss on unconfident pseudo labels (red boxes). $④$ We can boost the model performance by self-training on reliable pseudo labels (blue boxes) and repeating $① - ④$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 61, + 890, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we apply temporal flip to get event streams in both directions, and aggregate the detection results on them. Since the teacher model is trained on limited data, it will be uncertain on hard examples. We thus threshold the boxes with a small value to keep more detected objects. To remove false positive boxes, we build upon the temporal persistency prior of objects and apply tracking-based post-processing. However, there might still be inaccurate labels due to the low confidence threshold we use, and directly training the detector on them will lead to suboptimal results. Inspired by prior works on noise-robust learning [28, 61], we design a soft anchor assignment strategy to selectively supervise the model with pseudo labels. Finally, we can use the re-trained detector as the teacher model to initialize the next round of self-training. The described process can be repeated for multiple rounds to further boost the model performance.", + "bbox": [ + 75, + 349, + 472, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Comparison to online pseudo-labeling. In previous label-efficient object detection works [33, 59, 65, 68], the teacher model is jointly trained with a student model. In each training step, the teacher predicts bounding boxes on unlabeled data in a batch for student training. This online paradigm is also applicable to our setting. However, the teacher model will only see short event streams loaded in a batch. For example, on Gen1 [8], our training sequence length accounts for a duration of 1 second, while a car can stop for more than 10 seconds in real-world traffic and thus trigger no events. Pseudo-labeling on short event streams will inevitably miss these objects. Therefore, we adopt our two-stage offline label generation paradigm to retain full temporal information.", + "bbox": [ + 75, + 590, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Towards High-Quality Pseudo Labeling", + "text_level": 1, + "bbox": [ + 76, + 799, + 418, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce each key component in our LEOD framework to achieve high-quality pseudo labels.", + "bbox": [ + 75, + 824, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test-Time Augmentation (TTA). When deployed in the real world, event-based detectors are expected to run in real-time, i.e., they only take in events triggered before $t$ to de", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c93f578427f7d4a67bf8ef0a9d885729edcb8a5666c87ab2927dc14f61c739c8.jpg", + "image_caption": [ + "Figure 4. Illustration of the time-flip TTA which enhances our robustness against different object motions. Forward helps detect receding objects, while Backward helps with approaching objects." + ], + "image_footnote": [], + "bbox": [ + 501, + 347, + 888, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tect objects at $t$ . Instead, in our offline label generation process, we can use future information to refine predictions at the current timestep. As shown in Fig. 4, we run the detector on both the original and reversed event streams, enabling us to detect objects with different movements. We also apply a horizontal-flip TTA to further improve the detection.", + "bbox": [ + 496, + 513, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Filtering and Tracking. TTA helps us detect more objects (higher recall), yet it also leads to false positives (lower precision). Previous works simply use a threshold to filter out boxes with low confidence [59, 65]. However, as shown in Fig. 5 (a) and (b), there is a trade-off between precision and recall, making it hard to find the optimal threshold. We opt to first filter with a low threshold $\\tau_{\\mathrm{hard}}$ to avoid missing objects, followed by tracking-based post-processing to remove temporally inconsistent boxes. We follow the tracking-by-detection paradigm [2] to build tracks by linking detection boxes between frames. Each box $b$ will be associated with a track $s_k = \\{(b, v_x, v_y)_t, k, n, q\\}$ , where $(v_x, v_y)$ is the estimated velocity under a linear motion assumption, $k$ is the ID, $n$ is its length so far, and $q \\in [0,1]$ is the current score. In the first frame, we initialize each box as a track. For every coming frame, we first predict the positions of existing tracks and associate them with boxes at that frame via greedy matching over pairwise IoUs. Then, we decay the score $q$ of unmatched tracks and initialize unmatched", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16936", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/772c1175706c62c5f7a5cfc0ecaf97cff84a2b32c939cb3097ac318a69cffbf7.jpg", + "image_caption": [ + "(a) Label quality of cars" + ], + "image_footnote": [], + "bbox": [ + 81, + 66, + 261, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a872876c6b21d37919bb82c95d1d7c1e99e35597bb44027cb5869abe804d212.jpg", + "image_caption": [ + "Figure 5. Analysis on confidence thresholds. We randomly sample 10,000 predicted boxes from RVT-S pre-trained on $5\\%$ of Gen1 labels. We plot the pseudo labels' precision and recall of (a) cars and (b) pedestrians. In (c) and (d), we show each box's predicted confidence scores and its true IoU with ground-truth boxes. $\\tau_{\\mathrm{hard}}$ is the threshold for initial filtering, and $\\tau_{\\mathrm{soft}}$ is used in soft anchor assignment." + ], + "image_footnote": [], + "bbox": [ + 267, + 66, + 444, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1f39354898df9b83cbe82d948ac6bf6ad89196f84e7ab7c4bba0f5cfe4aaf8d0.jpg", + "image_caption": [ + "(b) Label quality of pedestrians", + "(c) Scores vs. True IoUs of cars" + ], + "image_footnote": [], + "bbox": [ + 450, + 68, + 665, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36eec798a9c248dcfe7f138050a4311061169575db6d36f824d98ec1ceb2a60c.jpg", + "image_caption": [ + "(d) Scores vs. True IoUs of pedestrians" + ], + "image_footnote": [], + "bbox": [ + 668, + 68, + 882, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "boxes as new tracks. Finally, tracks with low scores will be deleted. See Appendix A.1 for implementation details.", + "bbox": [ + 75, + 258, + 467, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to TTA, we apply tracking in both directions, and only remove a box if the length of its associated track is shorter than a threshold, $T_{\\text{trk}}$ , in both cases. However, predictions on hard examples may also be inconsistent, as the pre-trained model has limited capacity. Instead of suppressing removed boxes as background, we ignore them during loss computation as will be described later. Also, for long tracks, we inpaint boxes with linear motion at unmatched timesteps and ignore training losses on them too.", + "bbox": [ + 75, + 289, + 467, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Soft Anchor Assignment and Re-training. We can now re-train a detector on the ground-truth labels and the pseudo labels with the original detection loss. However, as shown in Fig. 5 (c) and (d), there are still low-quality boxes after post-processing. To handle noisy labels, we utilize a soft anchor assignment strategy to selectively supervise the model training. We first identify a set of uncertain labels including boxes belonging to short tracks, inpainted from long tracks, and those with detection scores lower than a threshold $\\tau_{\\mathrm{soft}}$ . Then, at each training step, we ignore the loss applied to anchors associated with these uncertain boxes, i.e., we do not supervise those anchors and allow them to discover new instances. This design is inspired by the anchor assignment in anchor-based detectors [30, 48], where two thresholds are used to determine foreground or background anchor boxes, and the anchors in between are ignored in loss computation. As we will show in ablation studies, soft anchor assignment makes our method less sensitive to hyper-parameters.", + "bbox": [ + 75, + 425, + 467, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite training on noisy labels, the model learns to refine the labels and detect new objects. Thus, we do an additional round of self-training to further improve the results.", + "bbox": [ + 75, + 696, + 467, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sec. 4.2 shows that LEOD outperforms baselines significantly in the low-label regime. In the fully labeled setting, we surpass previous state-of-the-art (Sec. 4.3). Our ablations show the contribution of each component in Sec. 4.4.", + "bbox": [ + 75, + 777, + 467, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 847, + 266, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We adopt Gen1 [8] and 1Mpx [42] datasets that feature various driving scenarios. Gen1 consists of 39 hours", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of recordings with a $304 \\times 240$ resolution event camera [43]. It provides bounding box annotations of cars and pedestrians at 1, 2, or $4\\mathrm{Hz}$ . 1Mpx is recorded with a higher $720 \\times 1280$ resolution event camera [10]. It contains around 15 hours of data collected over several months at day and night, and provides labels for cars, pedestrians, and two-wheelers at 30 or $60\\mathrm{Hz}$ . Following previous works [14, 29], we remove ground-truth boxes that are too small during evaluation on both datasets, and half the events' resolution to $360 \\times 640$ on 1Mpx.", + "bbox": [ + 496, + 258, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol. Mean average precision (mAP) is adopted as the main performance metric. We choose $1\\%$ , $2\\%$ , $5\\%$ , and $10\\%$ as the labeling ratio following prior works [55, 65]. In the weakly-supervised object detection (WSOD) setting, labels in all event streams are uniformly sub-sampled. In the semi-supervised object detection (SSOD) setting, we keep a small portion of event streams unchanged, while setting other event sequences as fully unlabeled. For the same labeling ratio, the amounts of available labels in WSOD and SSOD are roughly the same. Finally, all labels are provided in the fully labeled setting.", + "bbox": [ + 496, + 412, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detector Training Details. We adopt the state-of-the-art event-based detector RVT [14] as our base model. Due to limited computation resources, we mainly experiment with RVT-S, while we show that LEOD also scales to the largest RVT-B variant in Sec. 4.4. Most of the configurations are the same as we build upon their open-source codebase. Here we only highlight our modifications. In order to apply the time-flip TTA, we train with an additional time-flip data augmentation. When re-training on pseudo labels, we initialize RVT from scratch and use the Adam optimizer [24] with a peak learning rate of $5 \\times 10^{-4}$ to train for 150k iterations. Please refer to Appendix A.2 for more details.", + "bbox": [ + 496, + 580, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo-Labeling Details. Inspired by prior works [51, 59], we set different thresholds for each category. To simplify parameter tuning, we follow two rules: (i) pedestrians and two-wheelers share the same values, which are half of cars' values, (ii) for cars, the soft threshold $\\tau_{\\mathrm{soft}}$ equals to the hard threshold $\\tau_{\\mathrm{hard}} + 0.1$ , while for pedestrians and two-wheelers, we use $\\tau_{\\mathrm{soft}} = \\tau_{\\mathrm{hard}} + 0.05$ . In both settings and both datasets, we choose the same set of hyper-parameters: $\\tau_{\\mathrm{hard}} = 0.6$ for cars and $\\tau_{\\mathrm{hard}} = 0.3$ for pedestrians. Only in", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16937", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf0108c42a47548ebc311a45a5e0ddc1196c9814246dc6d47b45f4acd4300f0c.jpg", + "image_caption": [ + "(a) Gen1 weakly-supervised object detection (WSOD) results" + ], + "image_footnote": [], + "bbox": [ + 80, + 60, + 480, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be0501d3066a3b6dc6eab8084e1e6a82615a0b34228e43163c2207251093ea5b.jpg", + "image_caption": [ + "(b) 1Mpx weakly-supervised object detection (WSOD) results" + ], + "image_footnote": [], + "bbox": [ + 488, + 60, + 888, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26b248e9a6dd897e211043cc5dd251581f53c0c332af40dbdcaabeeb23f49e75.jpg", + "image_caption": [ + "(c) Gen1 semi-supervised object detection (SSOD) results" + ], + "image_footnote": [], + "bbox": [ + 76, + 260, + 478, + 436 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0bbfe38af1c43e77d67987ace07497bf28e25967ed2b0aee27c1d96b530b18cd.jpg", + "image_caption": [ + "(d) 1Mpx semi-supervised object detection (SSOD) results", + "Figure 6. Results on label-efficient learning using different ratios of labeled data. We compare LEOD with baselines on Gen1 and 1Mpx datasets under WSOD and SSOD using the RVT-S detector. All results are averaged over three runs." + ], + "image_footnote": [], + "bbox": [ + 486, + 260, + 887, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1Mpx WSOD, we set $\\tau_{\\mathrm{hard}} = 0.5$ for pedestrians and two-wheelers to handle excessive false positives. The minimum track length $T_{trk}$ is set to 6 in all experiments.", + "bbox": [ + 75, + 492, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with a Supervised Baseline trained only on available labels. Since we are the first work to consider this task, we design two other baselines: 1) Vanilla Self-Training: trains on pseudo labels without TTA, tracking, and soft anchor assignment; 2) Soft Teacher: adopts the online student-teacher paradigm from a representative 2D SSOD method [65]. We enable soft anchor assignment in Soft Teacher, while TTA and tracking are not applicable as the online event sequence length is too short. We tune baselines' hyper-parameters to be optimal in each setting. We also tried a state-of-the-art 2D SSOD method designed for anchor-free detectors [34] as YOLOX is anchor-free, but we did not observe clear improvements over Soft Teacher.", + "bbox": [ + 75, + 537, + 468, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Label-Efficient Results", + "text_level": 1, + "bbox": [ + 76, + 741, + 287, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare our method with baselines in the low-labeled data regime. The overall results are shown in Fig. 6. WSOD. Fig. 6 (a) and (b) present the weakly-supervised results. On Gen1, LEOD improves the mAP of Supervised Baseline by a large margin across all labeling ratios. Using $10\\%$ labels, we achieve an mAP of $45.5\\%$ , which is only $1\\%$ lower than RVT-S trained on $100\\%$ labels. For pseudolabeling baselines, Vanilla Self-Training outperforms Soft Teacher in most cases, validating our choice of offline la", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bel generation. In addition, LEOD consistently outperforms them by more than $2\\%$ , indicating the higher quality of our pseudo labels. We observe a similar trend on 1Mpx, where our approach scores the highest mAP in all cases. Notably, the two pseudo-labeling baselines perform worse than pretrained RVT-S on $5\\%$ and $10\\%$ labels, which proves the importance of screening reliable labels. Finally, LEOD with $10\\%$ labels $(44.6\\%)$ outperforms RVT-S trained on all labels $(44.1\\%)$ , showing the great potential of unlabeled data. SSOD. Fig. 6 (c) and (d) present the semi-supervised results. Using the same amount of labels, models trained under SSOD are generally worse than WSOD. This indicates that given a limited budget, we should sparsely label as many event streams as possible instead of densely labeling a few sequences. Nevertheless, LEOD still outperforms baselines by more than $2\\%$ mAP over all labeling ratios on both datasets. Our results offer a promising direction of boosting performance with fully unlabeled event data.", + "bbox": [ + 496, + 492, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 7 visualizes some detection results of RVT-S trained with $10\\%$ labels under Gen1 WSOD setting. The supervised baseline can only detect objects that trigger lots of events due to its limited capacity. Models using vanilla self-training detect more objects, but also produce numerous false positives. With our pseudo-labeling pipeline, LEOD trained models are able to handle various hard examples. For example, it discovers a car that was initially missed in the ground-truth annotation in Fig. 7 (d).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16938", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/797a7c13089971173011a37c949d52dcbe76a2ae9d1bd88525e79a381a9e2e40.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 89, + 66, + 302, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5801f9f88143a13dc3d8db8ad26400bb84412ca3ccb3319c9b5fa08df241414.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 308, + 66, + 495, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88246e8084f0fedc201e6723d757c8c652f0167ed168d8b8a968b40eaa1d0e5e.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 501, + 66, + 687, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc34ac774af87d9a2b4e4ddae8080dd25df1a98c017f2647a09e1ba26dc5b66b.jpg", + "image_caption": [ + "(d)", + "Figure 7. Detection on Gen1. We show (a) a common driving scene, (b) cars with small relative movement thus few events, (c) crowded pedestrians, and (d) LEOD discovers a missing object in ground-truth annotations due to the occlusion of a barrier gate. Car and pedestrian boxes are colored in blue and orange, respectively. Pre-trained and Self-Training stand for Supervised Baseline and Vanilla Self-Training." + ], + "image_footnote": [], + "bbox": [ + 694, + 66, + 880, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Fully-Labeled Results", + "text_level": 1, + "bbox": [ + 76, + 505, + 282, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the original labeling frequency on both datasets is lower than the frame rate of RVT-S (20 Hz), we can still create pseudo labels on unlabeled frames to improve the fully supervised model performance. We compare with state-of-the-art event-based object detectors trained on all labeled data in Tab. 1. LEOD improves over RVT-S by $2.2\\%$ and $2.6\\%$ on Gen1 and 1Mpx, respectively. On Gen1, our method achieves new state-of-the-art among models not using pre-trained weights. This indicates that LEOD is consistently effective even with $100\\%$ labels. In terms of runtime and model size, since our approach does not introduce new modules to the base model, we are as efficient as RVT-S.", + "bbox": [ + 75, + 529, + 470, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 724, + 240, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Larger Base Model. We evaluate our label-efficient learning framework on a stronger event-based detector RVT-B, which uses a larger ViT backbone compared to RVT-S. Tab. 2 presents the result in the Gen1 WSOD setting. With more parameters, RVT-B pre-trained on limited labels already outperforms RVT-S. Still, LEOD is able to improve the detection result by a sizeable margin across all labeling ratios. With $100\\%$ labels, our method achieves an mAP of $50.2\\%$ , which is competitive with ERGO-12 using a large-scale pre-trained Swin Transformer V2 backbone [35]. No", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/62b061fbba244b16805ddc276a27f3e9813e43884e3124cda3e3809ec21d5ff6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGen11MpxSize (M)
mAP (%)Time (ms)mAP (%)Time (ms)
RED [42]40.016.743.039.324.1
ASTMNet [29]46.735.648.372.3>100
HMNet-L3 [16]47.17.9*--33.2
RVT-B47.210.247.411.918.5
RVT-S46.59.544.110.19.9
LEOD-RVT-S48.79.546.710.19.9
ERGO-12 [76]50.477.240.6101.159.6
", + "bbox": [ + 504, + 508, + 885, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Detection results using all available labels. Baseline runtimes and model sizes are obtained from [14]. For HMNet, we use the best-performing L3 variant. * Its runtime is computed using a V100 GPU, while T4 GPUs which are slower than V100 are used in the other cases. ERGO is grayed as it is the only method using pre-trained models (Swin Transformer V2 [35]).", + "bbox": [ + 498, + 642, + 890, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tably, LEOD brings larger absolute improvements on RVT-B compared to RVT-S, proving that our framework steadily scales up to enhance larger and stronger detectors.", + "bbox": [ + 498, + 732, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Each Component. Fig. 8 shows the model performance with different components on Gen1 and 1Mpx. TTA significantly improves the mAP as it increases the recall in the generated pseudo labels using future information. Leveraging the temporal persistency of objects, tracking-based post-processing leads to a further gain. Finally, with soft anchor assignment, only reliable foreground and background labels are selected, easing the model training.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16939", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8c6e993cae07afc045992933898b1220f7a9fb103c814b6153f403ef667799d.jpg", + "image_caption": [ + "(a) Gen1 weakly-supervised object detection (WSOD) results" + ], + "image_footnote": [], + "bbox": [ + 89, + 60, + 478, + 233 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f69ad8292f54da85af719da86c47ccaa1fd043b40df8b4f877ece802e626dd88.jpg", + "image_caption": [ + "(b) 1Mpx weakly-supervised object detection (WSOD) results", + "Figure 8. Ablation study of each component in LEOD. We report the WSOD result with RVT-S as the base detector on both datasets. Starting from Vanilla Self-Training, we gradually add TTA, tracking-based post-processing, and soft anchor assignment." + ], + "image_footnote": [], + "bbox": [ + 486, + 60, + 875, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f533aad9f2708a6f96658d8511026f887846c138402fab015002440fae781f87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method1%2%5%10%100%
RVT-S30.834.437.940.846.5
LEOD-RVT-S39.442.243.845.648.7
Absolute Improvement+8.6+7.8+5.9+4.8+2.2
RVT-B31.634.838.341.047.6*
LEOD-RVT-B40.042.945.346.650.2
Absolute Improvement+8.4+8.1+7.0+5.6+2.6
", + "bbox": [ + 102, + 285, + 439, + 398 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/594b3e59d61b96a5a5236258c5aeb3546b8a44e0bcf2675122ea43051bf88e98.jpg", + "table_caption": [ + "Table 2. Gen1 WSOD mAPs (%) with two RVT variants. * our reproduced RVT-B using ${100}\\%$ labels result is better than [14]." + ], + "table_footnote": [], + "table_body": "
Rounds1%2%5%10%100%P. (1%)P. (2%)P. (5%)P. (10%)
138.141.143.145.348.50.650.690.740.79
239.442.243.845.648.70.720.750.770.81
339.542.243.645.448.60.720.740.740.76
", + "bbox": [ + 78, + 436, + 467, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Number of self-training rounds used in LEOD. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. We also compute the precision (P.) of pseudo labels that are used to train the models on that row, which is a good indicator for stop training.", + "bbox": [ + 75, + 500, + 468, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Self-Training Rounds. Tab. 3 left presents the results of multi-round self-training. Detectors after the first round of self-training are significantly better than pre-trained models, and thus generate higher quality pseudo labels. Therefore, a second round of training leads to consistent gains. However, due to error accumulations, a third round of training may result in worse models. To determine when to stop training, we empirically find that the precision of pseudo labels serves as a good indicator. In Tab. 3 right, we compute the precision of predicted boxes on labeled frames (skipped labels are not used to prevent information leakage). The precision consistently improves after round 1, but starts to decrease in some cases after round 2. Indeed, the mAP also drops after training on those labels with lower precision.", + "bbox": [ + 75, + 566, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Confidence Thresholds. We analyze the effects of hard and soft thresholds $(\\tau_{\\mathrm{hard}},\\tau_{\\mathrm{soft}})$ in Tab. 4. Prior works [59, 65] often use a high threshold of 0.9, while we observe clearly lower mAPs when $\\tau_{\\mathrm{hard}} > 0.7$ . With a lower $\\tau_{\\mathrm{hard}}$ , we retain most of the detected objects, and suppress noisy labels with $\\tau_{\\mathrm{soft}}$ . LEOD achieves similar results using several sets of hyper-parameters, showing our robustness to them. See Appendix B for more analysis on the filtering thresholds.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f0f571a1e4273245c5501a8df495a01de24d7d2394348421defc699373aaacc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(τhard, car, τsoft, car, τsoft, ped)1%2%5%10%
(0.5, 0.6, 0.3)38.941.843.545.4
(0.6, 0.7, 0.35)39.442.243.845.5
(0.6, 0.7, 0.4)39.342.043.745.6
(0.6, 0.8, 0.4)39.342.343.645.5
(0.7, 0.8, 0.4)39.042.043.244.9
(0.8, -, -)*38.441.442.744.2
", + "bbox": [ + 524, + 285, + 864, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation of hard and soft thresholds used in soft anchor assignment. We keep $\\tau_{\\text{hard, ped}} = \\tau_{\\text{hard, car}} / 2$ . Ped stands for pedestrian. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. * indicates not using soft anchor assignment.", + "bbox": [ + 498, + 397, + 890, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 460, + 617, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present LEOD, the first algorithm for label-efficient event-based object detection. To leverage unlabeled data, we adopt the self-training framework with reliable label selection. Several techniques are introduced to improve labeling quality. Extensive experiments on Gen1 and 1Mpx datasets showcase the superiority of our method over baselines in all the settings.", + "bbox": [ + 496, + 487, + 890, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Following common practice, we only conduct intra-dataset experiments, i.e., training on data gathered using the same protocol. Recent works [3, 25, 40] have shown that training large models over multiple datasets leads to excellent performance and generalization. Since LEOD benefits from unlabeled data, we can also train it jointly on more datasets that involve real-world multi-object event sequences [4, 15, 73]. We discuss some failure cases of our pseudo-labeling method in Appendix C.", + "bbox": [ + 496, + 594, + 890, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 739, + 645, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was partially supported by Compute Ontario (www.computeontario.ca), the Province of Ontario, the Digital Research Alliance of Canada (https://alliancecan.ca/), and the Government of Canada through CIFAR and the Natural Sciences and Engineering Research Council of Canada (NSERC). We would also like to acknowledge Vector Institute for computation support. We would like to thank Yash Kant, Lunjun Zhang, Xuanchi Ren, and Xiaoshi Wu for valuable discussions and support.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16940", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 2", + "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 4", + "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 8", + "[4] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In CVPRW, 2023. 8", + "[5] Nicholas FY Chen. Pseudo-labels for supervised learning on dynamic vision sensor data, applied to object detection under ego-motion. In CVPRW, 2018. 2", + "[6] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In ICCV, 2023. 2", + "[7] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IJCNN, 2022. 2", + "[8] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 1, 2, 3, 4, 5", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[10] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a $1280 \\times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\\mu m$ pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference, 2020. 5", + "[11] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. TPAMI, 2020. 1, 3", + "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3", + "[13] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. 2", + "[14] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 1, 2, 3, 5, 7, 8" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. RA-L, 2021. 8", + "[16] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In CVPR, 2023. 1, 2, 7", + "[17] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 1997. 2, 3", + "[18] Yuhuang Hu, Tobi Delbruck, and Shih-Chii Liu. Learning to exploit multiple vision modalities by using grafted networks. In ECCV, 2020. 2", + "[19] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR, 2021. 2", + "[20] Massimiliano Iacono, Stefan Weber, Arren Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. In IROS, 2018. 2", + "[21] Mingcheng Ji, Ziling Wang, Rui Yan, Qingjie Liu, Shu Xu, and Huajin Tang. Sctn: Event-based object tracking with energy-efficient deep convolutional spiking neural networks. Frontiers in Neuroscience, 2023. 2", + "[22] Zhuangyi Jiang, Pengfei Xia, Kai Huang, Walter Stechele, Guang Chen, Zhenshan Bing, and Alois Knoll. Mixed frame-event-driven fast pedestrian detection. In ICRA, 2019. 2", + "[23] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 2", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8", + "[26] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022. 2", + "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICMLW, 2013. 2", + "[28] Hengduo Li, Zuxuan Wu, Chen Zhu, Caiming Xiong, Richard Socher, and Larry S Davis. Learning from noisy anchors for one-stage object detection. In CVPR, 2020. 4", + "[29] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 1, 2, 5, 7", + "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5", + "[31] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 2023. 2", + "[32] Minghua Liu, Yin Zhou, Charles R Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In ECCV, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16941", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2020. 3, 4", + "[34] Yen-Cheng Liu, Chih-Yao Ma, and Zsolt Kira. Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In CVPR, 2022. 6", + "[35] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7", + "[36] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In ECCV, 2020. 2", + "[37] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. RA-L, 2022. 2", + "[38] Ishan Misra, Abhinav Shrivastava, and Martial Hebert. Watch and learn: Semi-supervised learning for object detectors from video. In CVPR, 2015. 2", + "[39] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. *IJRR*, 2017. 2", + "[40] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 8", + "[41] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1742-1750, 2015. 2", + "[42] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 2020. 1, 2, 3, 5, 7", + "[43] Christoph Posch, Daniel Matolin, and Rainer Wohlgannnt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 2010. 5", + "[44] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, 2021. 2", + "[45] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 2", + "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 2", + "[47] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. TPAMI, 2019. 2", + "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, 2022. 1, 2", + "[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 2", + "[51] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, 2020. 5", + "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 2015. 1, 2", + "[53] Krishna Kumar Singh, Fanyi Xiao, and Yong Jae Lee. Track and transfer: Watching videos to simulate strong human supervision for weakly-supervised object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3548-3556, 2016. 2", + "[54] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 2020. 2", + "[55] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 3, 5", + "[56] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, 2020. 2", + "[57] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, 2022. 2", + "[58] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2, 3", + "[59] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2021. 4, 5, 8", + "[60] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In CVPR, 2021. 2", + "[61] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 4", + "[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2", + "[63] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2", + "[64] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16942", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, 2021. 2, 3, 4, 5, 6, 8", + "[66] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, 2019. 2, 3", + "[67] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In ICCV, 2023. 2", + "[68] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022.4", + "[69] Alessandro Zanardi, Andreas Aumiller, Julian Zilly, Andrea Censi, and Emilio Frazzoli. Cross-modal learning filters for rgb-neuromorphic wormhole learning. RSS, 2019. 2", + "[70] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 2", + "[71] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In CVPR, 2023. 2", + "[72] Yongqiang Zhang, Yancheng Bai, Mingli Ding, Yongqiang Li, and Bernard Ghanem. W2f: A weakly-supervised to fully-supervised framework for object detection. In CVPR, 2018. 2", + "[73] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. RA-L, 2018. 8", + "[74] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2", + "[75] Alex Zihao Zhu, Ziyun Wang, Kaung Khant, and Kostas Daniilidis. Eventgan: Leveraging large scale image datasets for event cameras. In ICCP, 2021. 2", + "[76] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In ICCV, 2023. 1, 2, 7" + ], + "bbox": [ + 78, + 90, + 468, + 686 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16943", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_model.json b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e29eaa20b021a9fdadfb8597e791e92712be210d --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.787, + 0.152 + ], + "angle": 0, + "content": "LEOD: Label-Efficient Object Detection for Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.18, + 0.784, + 0.22 + ], + "angle": 0, + "content": "Ziyi Wu\\(^{1,2}\\), Mathias Gehrig\\(^{3}\\), Qing Lyu\\(^{1}\\), Xudong Liu\\(^{1}\\), Igor Gilitschenski\\(^{1,2}\\), \\(^{1}\\)University of Toronto, \\(^{2}\\)Vector Institute, \\(^{3}\\)University of Zurich" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.254, + 0.314, + 0.27 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.286, + 0.473, + 0.681 + ], + "angle": 0, + "content": "Object detection with event cameras benefits from the sensor's low latency and high dynamic range. However, it is costly to fully label event streams for supervised training due to their high temporal resolution. To reduce this cost, we present LEOD, the first method for label-efficient event-based detection. Our approach unifies weakly- and semi-supervised object detection with a self-training mechanism. We first utilize a detector pre-trained on limited labels to produce pseudo ground truth on unlabeled events. Then, the detector is re-trained with both real and generated labels. Leveraging the temporal consistency of events, we run bi-directional inference and apply tracking-based post-processing to enhance the quality of pseudo labels. To stabilize training against label noise, we further design a soft anchor assignment strategy. We introduce new experimental protocols to evaluate the task of label-efficient event-based detection on Gen1 and 1Mpx datasets. LEOD consistently outperforms supervised baselines across various labeling ratios. For example, on Gen1, it improves mAP by \\(8.6\\%\\) and \\(7.8\\%\\) for RVT-S trained with \\(1\\%\\) and \\(2\\%\\) labels. On 1Mpx, RVT-S with \\(10\\%\\) labels even surpasses its fully-supervised counterpart using \\(100\\%\\) labels. LEOD maintains its effectiveness even when all labeled data are available, reaching new state-of-the-art results. Finally, we show that our method readily scales to improve larger detectors as well. Code: https://github.com/Wuziyi616/LEOD." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.208, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Object detection is key to scene understanding. It provides a compact representation of raw sensor measurements as semantically meaningful bounding boxes. Speed is crucial in object detection, especially in safety-critical applications such as self-driving. Recently, event cameras have gained significant interest in computer vision due to their low latency, low energy consumption, and high dynamic range [11]. Leveraging these benefits, event-based object detectors [14, 16, 29, 42, 49, 76] have been developed to complement conventional frame-based detectors. Despite tremendous progress, much of their success heavily relies" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.242, + 0.891, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 1. Detection performance of LEOD and baselines trained only on labeled events or conducting naive self-training. Under the weakly-supervised setting, our method consistently improves the RVT-S detector [14] across all labeling ratios on the Gen1 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.893, + 0.629 + ], + "angle": 0, + "content": "on large datasets that are manually annotated. However, due to the high temporal resolution of event data, labeling objects at every timestamp is impractical. For example, Gen1 dataset [8] provides object labels at lower than \\(4\\mathrm{Hz}\\). As a result, existing methods only train their models on labeled events, and discard the remaining unlabeled data. In contrast, we view this task as a weakly-supervised learning problem, which calls for detection methods that can leverage a mix of labeled and unlabeled events during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.825 + ], + "angle": 0, + "content": "In this work, we address the challenge by proposing a Label-efficient Event-based Object Detection (LEOD) framework. We consider two settings characterized by limited labels: (i) weakly-supervised, where object bounding boxes are sparsely labeled in all event streams, and (ii) semi-supervised, where some event streams have object boxes densely labeled while others remain fully unlabeled. Our approach unifies the two settings through a self-training paradigm. With limited labels, we first pre-train a detector and use it to generate pseudo annotations on unlabeled events. Then, we re-train the detector on a combination of real and pseudo labels. However, naively generated labels contain noise, making their direct use suboptimal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To obtain high-quality pseudo labels, we exploit the temporal dimension of event data. Recent work [29] has shown the importance of temporal information in event-based detection using recurrent modules [52]. Additionally, offline label generation enables us to further refine predictions with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16933" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "future information. To achieve this, we first introduce time-flip augmentation to events during training. As a result, we can ensemble model predictions on the original and reverse event streams via Test-Time Augmentation (TTA), leading to higher detection recall. In addition, we leverage tracking-based post-processing to eliminate temporally inconsistent objects, enhancing the precision of pseudo labels. Finally, we filter out low-confidence boxes with a score threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.471, + 0.412 + ], + "angle": 0, + "content": "A key challenge here is how to select a proper threshold. Instead of searching for the optimal value, we first filter with a low threshold to avoid missing objects. This inevitably introduces many false positives, which we address with a soft anchor assignment strategy in training. When computing the detection loss, we set another higher threshold and only treat pseudo boxes above that threshold as positive labels. For boxes with a lower detection score, we ignore the loss applied to their associated anchors. This strategy ensures that the model is only supervised with reliable background and foreground labels, while being tolerant to noisy labels. Ablation studies show that our method is insensitive to the two threshold values, easing the hyper-parameter tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.47, + 0.55 + ], + "angle": 0, + "content": "To test our method, we design new protocols for label-efficient event-based detection on Gen1 [8] and 1Mpx [42] datasets. For weakly-supervised object detection (WSOD), we uniformly sub-sample the labels over time to simulate sparse annotations. For semi-supervised object detection (SSOD), we directly choose some event sequences as fully unlabeled. Following 2D SSOD evaluation protocols [55, 65], we also have a fully labeled setting, where we show that pseudo labels can complement ground-truths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.688 + ], + "angle": 0, + "content": "In summary, this work makes four main contributions: (i) We introduce the task of label-efficient object detection to event vision, and design its experimental protocols. (ii) We propose LEOD, a unified framework for training event-based detectors with limited annotations. (iii) LEOD consistently outperforms baselines in various settings on two public detection datasets. (iv) Our method remains effective under the fully labeled setting and scales up to larger detectors, achieving new state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.219, + 0.723 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Object Detection with Event Cameras. Existing event-based detectors can be mainly categorized into two classes depending on whether they utilize the asynchronous nature of events. One line of work explores the sparsity of events, and employs Graph Neural Networks (GNNs) [13, 36, 49] or Spiking Neural Networks (SNNs) [7, 21, 70] for feature extraction. However, these approaches struggle with propagating information over long time horizons, which is crucial for detecting objects from events. Moreover, specialized hardware is required to achieve theoretical speed-ups of sparse networks, limiting their application in practice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "In another class of methods, events are converted to dense frame-like representations, followed by conventional networks for detection. Earlier works only consider event frames aggregated from a short time interval [5, 20, 22, 31]. This discards long-horizon history and makes it hard to detect objects under small relative motion as they trigger very few events. Recent methods thus introduce recurrent modules [17, 52] to enhance the memory of the detectors [14, 29, 42]. Further research focuses on better backbones [9, 58], inference speed [16], and event representations [76]. Because our primary goal is to study label-efficient learning for event-based detectors, we adopt the state-of-the-art approach RVT [14] as the base model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.529 + ], + "angle": 0, + "content": "Label-Efficient Learning in Event-based Vision. Due to a lack of large labeled datasets, there have been several works studying event-based algorithms with limited labels. Some papers focus on bridging frame-based and event-based vision. They either reconstruct natural images from events to apply traditional deep models [46, 47, 50, 56, 62], simulate events from videos to transfer the annotations [19, 39, 45, 75], or distill knowledge from trained frame-based models [18, 37, 57, 60, 67, 69]. However, these methods require either paired recordings of events and images or massive in-domain labeled images for training. Closer to ours are methods that only use event data [6, 26, 63, 74]. They conduct label-efficient learning on events with pre-trained frame-based models or self-supervised losses. Yet, none of them are designed for the detection task. Our work is the first attempt at label-efficient event-based object detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Label-Efficient Learning in Other Fields. Self-training based methods have been explored in tasks such as 2D image classification [27, 54, 64], object detection [55, 65, 72], and segmentation [1, 23, 41]. Our method is more related to label-efficient learning on videos [38, 53, 66] and 3D point cloud sequences [32, 44, 71] as these methods also exploit the temporal information of input data. For example, [66] utilizes optical flows to propagate single-frame labels to adjacent video frames. [44] and [32] train a teacher model on dense point clouds aggregated from a few timesteps. In contrast, we leverage a much longer temporal horizon by running the detector on the entire event stream in both directions. [71] also employs tracking-based post-processing to remove inconsistent boxes. We additionally perform tracking in both directions as a forward-backward consistency check, thus better leveraging the temporal information of event data. To tackle the noisy pseudo labels, [65] proposes to use the detection scores from the teacher model to weigh the loss. Instead, we design a soft anchor assignment strategy by ignoring the loss associated with unconfident boxes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.591, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper introduces a new task named label-efficient event-based object detection (formulated in Sec. 3.1). Our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16934" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "algorithm adopts a two-stage self-training framework with reliable label selection (Sec. 3.2), where we leverage the temporal information of event data to obtain high-quality pseudo labels and suppress noisy predictions (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.164, + 0.281, + 0.179 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.187, + 0.469, + 0.278 + ], + "angle": 0, + "content": "Events data. Event cameras record brightness changes, and output a sequence of events \\(\\mathcal{E} = \\{e_i = (x_i,y_i,t_i,p_i)\\}\\). Each event \\(e_i\\) is parameterized by its pixel coordinate \\((x_{i},y_{i})\\), timestamp \\(t_i\\), and polarity \\(p_i\\in \\{-1,1\\}\\). Modern event cameras run at sub-milliseconds and can produce millions of events per second [11]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.279, + 0.469, + 0.445 + ], + "angle": 0, + "content": "Event-based Object Detection. As the object motion in a scene is usually much slower than the event generation speed, event-based object detectors are only applied and evaluated at a fixed time interval \\( T \\) [8, 42]. More specifically, given an event stream \\( \\mathcal{E} \\) capturing a set of objects \\( \\mathcal{O} = \\{o_j\\}_{j=1}^M \\), we aim at detecting the 2D bounding boxes of them with the semantic labels, \\( \\mathcal{B} = \\{b_j = (x_j, y_j, w_j, h_j, l_j, t_l)\\}_{j=1}^M \\). Each bounding box \\( b_j \\) is characterized by the location of its top-left corner \\( (x_j, y_j) \\), width \\( w_j \\), height \\( h_j \\), class label \\( l_j \\in \\{1, 2, \\dots, C\\} \\), and timestamp \\( t_j \\). Here, \\( C \\) denotes the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Event-based Object Detectors. We take RVT [14] for example as it serves as our base model. RVT is a synchronous detector that converts events in every time window \\(\\Delta t\\) to a grid-like representation \\(I\\). In the remaining part of the paper, we call \\(I\\) a frame and every \\(\\Delta t\\) a timestep. RVT combines a Vision Transformer backbone [58] with a YOLOX detection head [12]. To extract temporal features, RVT introduces LSTM [17] cells in the backbone to fuse information over multiple timesteps. This is useful for detecting slow-moving objects as they only generate a few events. The YOLOX detection head in RVT is anchor-free, i.e., for each location on the feature map (anchor point), it predicts an objectness score \\(p_{\\mathrm{obj}} \\in [0,1]\\), per-class IoU values \\(p_{\\mathrm{iou}} \\in \\mathbb{R}^C\\), and offsets of the bounding box parameters \\(\\Delta b = (\\Delta x, \\Delta y, \\Delta w, \\Delta h)\\). The \\(p_{\\mathrm{iou}}\\) is trained to output the IoU value between the predicted box and the matched ground-truth box of that class. To obtain the final prediction, Non-Maximum Suppression (NMS) is applied to remove overlapping bounding boxes with low confidence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.824 + ], + "angle": 0, + "content": "During training, each ground-truth bounding box is matched to several anchor points for loss computation. Below we will denote anchor points as anchors for simplicity. Anchors matched with ground-truth are foreground, where all predicted values are supervised. For the remaining background anchors, only \\( p_{\\mathrm{obj}} \\) is trained to be 0." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Label-Efficient Event-based Object Detection. Fig. 2 shows our proposed detection settings with limited labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Weakly-supervised object detection (WSOD). In the WSOD setting, all event streams are sparsely labeled. Moreover, labels assigned to adjacent frames offer fewer informa" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.061, + 0.889, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.216, + 0.824, + 0.228 + ], + "angle": 0, + "content": "(a) Weakly-supervised object detection (WSOD)" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.231, + 0.89, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.387, + 0.815, + 0.398 + ], + "angle": 0, + "content": "(b) Semi-supervised object detection (SSOD)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Figure 2. Illustration of two label-efficient event-based object detection settings: (a) weakly-supervised where all event sequences are sparsely annotated, and (b) semi-supervised where some event sequences are densely annotated, and others are fully unlabeled. We visualize both positive and negative events in black. Tive training signals compared to those distributed across frames [66]. Therefore, it is reasonable to label object boxes in a long event sequence uniformly and sparsely." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Semi-supervised object detection (SSOD). In the SSOD setting, some event sequences are densely labeled, while others are fully unlabeled. This is also practical when people continue to collect data into already annotated datasets. Since capturing event sequences is much easier than labeling them, an algorithm that can consistently improve model performance with raw events is highly useful." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.78 + ], + "angle": 0, + "content": "To evaluate the label-efficient learning performance, we take existing event-based object detection datasets [8, 42] and sample a small portion of frames (WSOD) or sequences (SSOD) as labeled data. The rest of the training data are used as an unlabeled set following previous works [33, 55, 65]. We also have a fully labeled setting where all labels are available. Since the original event streams are annotated at a larger time interval than \\(\\Delta t\\), we can still create pseudo labels on unlabeled timesteps to improve the performance. See Appendix D for further discussions on the two settings." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.812, + 0.803 + ], + "angle": 0, + "content": "3.2. LEOD: A Self-Training Framework" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, the overall pipeline of LEOD follows a student-teacher pseudo-labeling paradigm, which is applicable to both, the WSOD and the SSOD settings. We first pre-train a detector on labeled data using regular detection loss until convergence. Then, we employ it to annotate unlabeled frames. To leverage the benefit of offline prediction," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16935" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.062, + 0.891, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.271, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Overview of our LEOD pipeline. \\(①\\) We first pre-train an event-based object detector on event streams with limited labels. \\(①\\) To leverage the temporal information, we apply time-flip Test-Time Augmentation (TTA) to unlabeled event streams and ensemble the model predictions. \\(②\\) We then apply forward and backward tracking to identify temporally inconsistent bounding boxes, i.e., boxes associated with short tracks. \\(③\\) To handle noisy labels, a soft anchor assignment strategy is designed to ignore detection loss on unconfident pseudo labels (red boxes). \\(④\\) We can boost the model performance by self-training on reliable pseudo labels (blue boxes) and repeating \\(① - ④\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.473, + 0.591 + ], + "angle": 0, + "content": "we apply temporal flip to get event streams in both directions, and aggregate the detection results on them. Since the teacher model is trained on limited data, it will be uncertain on hard examples. We thus threshold the boxes with a small value to keep more detected objects. To remove false positive boxes, we build upon the temporal persistency prior of objects and apply tracking-based post-processing. However, there might still be inaccurate labels due to the low confidence threshold we use, and directly training the detector on them will lead to suboptimal results. Inspired by prior works on noise-robust learning [28, 61], we design a soft anchor assignment strategy to selectively supervise the model with pseudo labels. Finally, we can use the re-trained detector as the teacher model to initialize the next round of self-training. The described process can be repeated for multiple rounds to further boost the model performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Comparison to online pseudo-labeling. In previous label-efficient object detection works [33, 59, 65, 68], the teacher model is jointly trained with a student model. In each training step, the teacher predicts bounding boxes on unlabeled data in a batch for student training. This online paradigm is also applicable to our setting. However, the teacher model will only see short event streams loaded in a batch. For example, on Gen1 [8], our training sequence length accounts for a duration of 1 second, while a car can stop for more than 10 seconds in real-world traffic and thus trigger no events. Pseudo-labeling on short event streams will inevitably miss these objects. Therefore, we adopt our two-stage offline label generation paradigm to retain full temporal information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.419, + 0.817 + ], + "angle": 0, + "content": "3.3. Towards High-Quality Pseudo Labeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.855 + ], + "angle": 0, + "content": "In this section, we introduce each key component in our LEOD framework to achieve high-quality pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Test-Time Augmentation (TTA). When deployed in the real world, event-based detectors are expected to run in real-time, i.e., they only take in events triggered before \\( t \\) to de" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.348, + 0.89, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.446, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Figure 4. Illustration of the time-flip TTA which enhances our robustness against different object motions. Forward helps detect receding objects, while Backward helps with approaching objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.605 + ], + "angle": 0, + "content": "tect objects at \\( t \\). Instead, in our offline label generation process, we can use future information to refine predictions at the current timestep. As shown in Fig. 4, we run the detector on both the original and reversed event streams, enabling us to detect objects with different movements. We also apply a horizontal-flip TTA to further improve the detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Filtering and Tracking. TTA helps us detect more objects (higher recall), yet it also leads to false positives (lower precision). Previous works simply use a threshold to filter out boxes with low confidence [59, 65]. However, as shown in Fig. 5 (a) and (b), there is a trade-off between precision and recall, making it hard to find the optimal threshold. We opt to first filter with a low threshold \\(\\tau_{\\mathrm{hard}}\\) to avoid missing objects, followed by tracking-based post-processing to remove temporally inconsistent boxes. We follow the tracking-by-detection paradigm [2] to build tracks by linking detection boxes between frames. Each box \\(b\\) will be associated with a track \\(s_k = \\{(b, v_x, v_y)_t, k, n, q\\}\\), where \\((v_x, v_y)\\) is the estimated velocity under a linear motion assumption, \\(k\\) is the ID, \\(n\\) is its length so far, and \\(q \\in [0,1]\\) is the current score. In the first frame, we initialize each box as a track. For every coming frame, we first predict the positions of existing tracks and associate them with boxes at that frame via greedy matching over pairwise IoUs. Then, we decay the score \\(q\\) of unmatched tracks and initialize unmatched" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16936" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.067, + 0.263, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.195, + 0.237, + 0.207 + ], + "angle": 0, + "content": "(a) Label quality of cars" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.068, + 0.445, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.195, + 0.44, + 0.207 + ], + "angle": 0, + "content": "(b) Label quality of pedestrians" + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.069, + 0.666, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.195, + 0.642, + 0.207 + ], + "angle": 0, + "content": "(c) Scores vs. True IoUs of cars" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.069, + 0.883, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.195, + 0.88, + 0.208 + ], + "angle": 0, + "content": "(d) Scores vs. True IoUs of pedestrians" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Figure 5. Analysis on confidence thresholds. We randomly sample 10,000 predicted boxes from RVT-S pre-trained on \\(5\\%\\) of Gen1 labels. We plot the pseudo labels' precision and recall of (a) cars and (b) pedestrians. In (c) and (d), we show each box's predicted confidence scores and its true IoU with ground-truth boxes. \\(\\tau_{\\mathrm{hard}}\\) is the threshold for initial filtering, and \\(\\tau_{\\mathrm{soft}}\\) is used in soft anchor assignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.468, + 0.289 + ], + "angle": 0, + "content": "boxes as new tracks. Finally, tracks with low scores will be deleted. See Appendix A.1 for implementation details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.468, + 0.425 + ], + "angle": 0, + "content": "Similar to TTA, we apply tracking in both directions, and only remove a box if the length of its associated track is shorter than a threshold, \\(T_{\\text{trk}}\\), in both cases. However, predictions on hard examples may also be inconsistent, as the pre-trained model has limited capacity. Instead of suppressing removed boxes as background, we ignore them during loss computation as will be described later. Also, for long tracks, we inpaint boxes with linear motion at unmatched timesteps and ignore training losses on them too." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.468, + 0.697 + ], + "angle": 0, + "content": "Soft Anchor Assignment and Re-training. We can now re-train a detector on the ground-truth labels and the pseudo labels with the original detection loss. However, as shown in Fig. 5 (c) and (d), there are still low-quality boxes after post-processing. To handle noisy labels, we utilize a soft anchor assignment strategy to selectively supervise the model training. We first identify a set of uncertain labels including boxes belonging to short tracks, inpainted from long tracks, and those with detection scores lower than a threshold \\(\\tau_{\\mathrm{soft}}\\). Then, at each training step, we ignore the loss applied to anchors associated with these uncertain boxes, i.e., we do not supervise those anchors and allow them to discover new instances. This design is inspired by the anchor assignment in anchor-based detectors [30, 48], where two thresholds are used to determine foreground or background anchor boxes, and the anchors in between are ignored in loss computation. As we will show in ablation studies, soft anchor assignment makes our method less sensitive to hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.468, + 0.743 + ], + "angle": 0, + "content": "Despite training on noisy labels, the model learns to refine the labels and detect new objects. Thus, we do an additional round of self-training to further improve the results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.21, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.468, + 0.84 + ], + "angle": 0, + "content": "Sec. 4.2 shows that LEOD outperforms baselines significantly in the low-label regime. In the fully labeled setting, we surpass previous state-of-the-art (Sec. 4.3). Our ablations show the contribution of each component in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.267, + 0.864 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Datasets. We adopt Gen1 [8] and 1Mpx [42] datasets that feature various driving scenarios. Gen1 consists of 39 hours" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.41 + ], + "angle": 0, + "content": "of recordings with a \\(304 \\times 240\\) resolution event camera [43]. It provides bounding box annotations of cars and pedestrians at 1, 2, or \\(4\\mathrm{Hz}\\). 1Mpx is recorded with a higher \\(720 \\times 1280\\) resolution event camera [10]. It contains around 15 hours of data collected over several months at day and night, and provides labels for cars, pedestrians, and two-wheelers at 30 or \\(60\\mathrm{Hz}\\). Following previous works [14, 29], we remove ground-truth boxes that are too small during evaluation on both datasets, and half the events' resolution to \\(360 \\times 640\\) on 1Mpx." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Evaluation Protocol. Mean average precision (mAP) is adopted as the main performance metric. We choose \\(1\\%\\), \\(2\\%\\), \\(5\\%\\), and \\(10\\%\\) as the labeling ratio following prior works [55, 65]. In the weakly-supervised object detection (WSOD) setting, labels in all event streams are uniformly sub-sampled. In the semi-supervised object detection (SSOD) setting, we keep a small portion of event streams unchanged, while setting other event sequences as fully unlabeled. For the same labeling ratio, the amounts of available labels in WSOD and SSOD are roughly the same. Finally, all labels are provided in the fully labeled setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Detector Training Details. We adopt the state-of-the-art event-based detector RVT [14] as our base model. Due to limited computation resources, we mainly experiment with RVT-S, while we show that LEOD also scales to the largest RVT-B variant in Sec. 4.4. Most of the configurations are the same as we build upon their open-source codebase. Here we only highlight our modifications. In order to apply the time-flip TTA, we train with an additional time-flip data augmentation. When re-training on pseudo labels, we initialize RVT from scratch and use the Adam optimizer [24] with a peak learning rate of \\(5 \\times 10^{-4}\\) to train for 150k iterations. Please refer to Appendix A.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pseudo-Labeling Details. Inspired by prior works [51, 59], we set different thresholds for each category. To simplify parameter tuning, we follow two rules: (i) pedestrians and two-wheelers share the same values, which are half of cars' values, (ii) for cars, the soft threshold \\(\\tau_{\\mathrm{soft}}\\) equals to the hard threshold \\(\\tau_{\\mathrm{hard}} + 0.1\\), while for pedestrians and two-wheelers, we use \\(\\tau_{\\mathrm{soft}} = \\tau_{\\mathrm{hard}} + 0.05\\). In both settings and both datasets, we choose the same set of hyper-parameters: \\(\\tau_{\\mathrm{hard}} = 0.6\\) for cars and \\(\\tau_{\\mathrm{hard}} = 0.3\\) for pedestrians. Only in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16937" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.061, + 0.481, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.241, + 0.443, + 0.253 + ], + "angle": 0, + "content": "(a) Gen1 weakly-supervised object detection (WSOD) results" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.061, + 0.89, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.241, + 0.853, + 0.253 + ], + "angle": 0, + "content": "(b) 1Mpx weakly-supervised object detection (WSOD) results" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.261, + 0.48, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.441, + 0.432, + 0.453 + ], + "angle": 0, + "content": "(c) Gen1 semi-supervised object detection (SSOD) results" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.261, + 0.888, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.441, + 0.842, + 0.453 + ], + "angle": 0, + "content": "(d) 1Mpx semi-supervised object detection (SSOD) results" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Figure 6. Results on label-efficient learning using different ratios of labeled data. We compare LEOD with baselines on Gen1 and 1Mpx datasets under WSOD and SSOD using the RVT-S detector. All results are averaged over three runs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.469, + 0.538 + ], + "angle": 0, + "content": "1Mpx WSOD, we set \\(\\tau_{\\mathrm{hard}} = 0.5\\) for pedestrians and two-wheelers to handle excessive false positives. The minimum track length \\(T_{trk}\\) is set to 6 in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Baselines. We compare with a Supervised Baseline trained only on available labels. Since we are the first work to consider this task, we design two other baselines: 1) Vanilla Self-Training: trains on pseudo labels without TTA, tracking, and soft anchor assignment; 2) Soft Teacher: adopts the online student-teacher paradigm from a representative 2D SSOD method [65]. We enable soft anchor assignment in Soft Teacher, while TTA and tracking are not applicable as the online event sequence length is too short. We tune baselines' hyper-parameters to be optimal in each setting. We also tried a state-of-the-art 2D SSOD method designed for anchor-free detectors [34] as YOLOX is anchor-free, but we did not observe clear improvements over Soft Teacher." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.289, + 0.757 + ], + "angle": 0, + "content": "4.2. Label-Efficient Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We first compare our method with baselines in the low-labeled data regime. The overall results are shown in Fig. 6. WSOD. Fig. 6 (a) and (b) present the weakly-supervised results. On Gen1, LEOD improves the mAP of Supervised Baseline by a large margin across all labeling ratios. Using \\(10\\%\\) labels, we achieve an mAP of \\(45.5\\%\\), which is only \\(1\\%\\) lower than RVT-S trained on \\(100\\%\\) labels. For pseudolabeling baselines, Vanilla Self-Training outperforms Soft Teacher in most cases, validating our choice of offline la" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.765 + ], + "angle": 0, + "content": "bel generation. In addition, LEOD consistently outperforms them by more than \\(2\\%\\), indicating the higher quality of our pseudo labels. We observe a similar trend on 1Mpx, where our approach scores the highest mAP in all cases. Notably, the two pseudo-labeling baselines perform worse than pretrained RVT-S on \\(5\\%\\) and \\(10\\%\\) labels, which proves the importance of screening reliable labels. Finally, LEOD with \\(10\\%\\) labels \\((44.6\\%)\\) outperforms RVT-S trained on all labels \\((44.1\\%)\\), showing the great potential of unlabeled data. SSOD. Fig. 6 (c) and (d) present the semi-supervised results. Using the same amount of labels, models trained under SSOD are generally worse than WSOD. This indicates that given a limited budget, we should sparsely label as many event streams as possible instead of densely labeling a few sequences. Nevertheless, LEOD still outperforms baselines by more than \\(2\\%\\) mAP over all labeling ratios on both datasets. Our results offer a promising direction of boosting performance with fully unlabeled event data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 7 visualizes some detection results of RVT-S trained with \\(10\\%\\) labels under Gen1 WSOD setting. The supervised baseline can only detect objects that trigger lots of events due to its limited capacity. Models using vanilla self-training detect more objects, but also produce numerous false positives. With our pseudo-labeling pipeline, LEOD trained models are able to handle various hard examples. For example, it discovers a car that was initially missed in the ground-truth annotation in Fig. 7 (d)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16938" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.068, + 0.303, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.438, + 0.221, + 0.453 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.068, + 0.496, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.438, + 0.414, + 0.453 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.068, + 0.688, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.438, + 0.606, + 0.453 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.068, + 0.882, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.437, + 0.798, + 0.453 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.457, + 0.894, + 0.501 + ], + "angle": 0, + "content": "Figure 7. Detection on Gen1. We show (a) a common driving scene, (b) cars with small relative movement thus few events, (c) crowded pedestrians, and (d) LEOD discovers a missing object in ground-truth annotations due to the occlusion of a barrier gate. Car and pedestrian boxes are colored in blue and orange, respectively. Pre-trained and Self-Training stand for Supervised Baseline and Vanilla Self-Training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.506, + 0.283, + 0.521 + ], + "angle": 0, + "content": "4.3. Fully-Labeled Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.471, + 0.71 + ], + "angle": 0, + "content": "Since the original labeling frequency on both datasets is lower than the frame rate of RVT-S (20 Hz), we can still create pseudo labels on unlabeled frames to improve the fully supervised model performance. We compare with state-of-the-art event-based object detectors trained on all labeled data in Tab. 1. LEOD improves over RVT-S by \\(2.2\\%\\) and \\(2.6\\%\\) on Gen1 and 1Mpx, respectively. On Gen1, our method achieves new state-of-the-art among models not using pre-trained weights. This indicates that LEOD is consistently effective even with \\(100\\%\\) labels. In terms of runtime and model size, since our approach does not introduce new modules to the base model, we are as efficient as RVT-S." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.241, + 0.74 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Larger Base Model. We evaluate our label-efficient learning framework on a stronger event-based detector RVT-B, which uses a larger ViT backbone compared to RVT-S. Tab. 2 presents the result in the Gen1 WSOD setting. With more parameters, RVT-B pre-trained on limited labels already outperforms RVT-S. Still, LEOD is able to improve the detection result by a sizeable margin across all labeling ratios. With \\(100\\%\\) labels, our method achieves an mAP of \\(50.2\\%\\), which is competitive with ERGO-12 using a large-scale pre-trained Swin Transformer V2 backbone [35]. No" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.51, + 0.887, + 0.639 + ], + "angle": 0, + "content": "
MethodGen11MpxSize (M)
mAP (%)Time (ms)mAP (%)Time (ms)
RED [42]40.016.743.039.324.1
ASTMNet [29]46.735.648.372.3>100
HMNet-L3 [16]47.17.9*--33.2
RVT-B47.210.247.411.918.5
RVT-S46.59.544.110.19.9
LEOD-RVT-S48.79.546.710.19.9
ERGO-12 [76]50.477.240.6101.159.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.643, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Table 1. Detection results using all available labels. Baseline runtimes and model sizes are obtained from [14]. For HMNet, we use the best-performing L3 variant. * Its runtime is computed using a V100 GPU, while T4 GPUs which are slower than V100 are used in the other cases. ERGO is grayed as it is the only method using pre-trained models (Swin Transformer V2 [35])." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.891, + 0.779 + ], + "angle": 0, + "content": "tably, LEOD brings larger absolute improvements on RVT-B compared to RVT-S, proving that our framework steadily scales up to enhance larger and stronger detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of Each Component. Fig. 8 shows the model performance with different components on Gen1 and 1Mpx. TTA significantly improves the mAP as it increases the recall in the generated pseudo labels using future information. Leveraging the temporal persistency of objects, tracking-based post-processing leads to a further gain. Finally, with soft anchor assignment, only reliable foreground and background labels are selected, easing the model training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.061, + 0.48, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.236, + 0.446, + 0.248 + ], + "angle": 0, + "content": "(a) Gen1 weakly-supervised object detection (WSOD) results" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.061, + 0.877, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.236, + 0.846, + 0.248 + ], + "angle": 0, + "content": "(b) 1Mpx weakly-supervised object detection (WSOD) results" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.89, + 0.28 + ], + "angle": 0, + "content": "Figure 8. Ablation study of each component in LEOD. We report the WSOD result with RVT-S as the base detector on both datasets. Starting from Vanilla Self-Training, we gradually add TTA, tracking-based post-processing, and soft anchor assignment." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.286, + 0.441, + 0.399 + ], + "angle": 0, + "content": "
Method1%2%5%10%100%
RVT-S30.834.437.940.846.5
LEOD-RVT-S39.442.243.845.648.7
Absolute Improvement+8.6+7.8+5.9+4.8+2.2
RVT-B31.634.838.341.047.6*
LEOD-RVT-B40.042.945.346.650.2
Absolute Improvement+8.4+8.1+7.0+5.6+2.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.403, + 0.47, + 0.431 + ], + "angle": 0, + "content": "Table 2. Gen1 WSOD mAPs (%) with two RVT variants. * our reproduced RVT-B using \\( {100}\\% \\) labels result is better than [14]." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.438, + 0.468, + 0.498 + ], + "angle": 0, + "content": "
Rounds1%2%5%10%100%P. (1%)P. (2%)P. (5%)P. (10%)
138.141.143.145.348.50.650.690.740.79
239.442.243.845.648.70.720.750.770.81
339.542.243.645.448.60.720.740.740.76
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.558 + ], + "angle": 0, + "content": "Table 3. Number of self-training rounds used in LEOD. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. We also compute the precision (P.) of pseudo labels that are used to train the models on that row, which is a good indicator for stop training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Self-Training Rounds. Tab. 3 left presents the results of multi-round self-training. Detectors after the first round of self-training are significantly better than pre-trained models, and thus generate higher quality pseudo labels. Therefore, a second round of training leads to consistent gains. However, due to error accumulations, a third round of training may result in worse models. To determine when to stop training, we empirically find that the precision of pseudo labels serves as a good indicator. In Tab. 3 right, we compute the precision of predicted boxes on labeled frames (skipped labels are not used to prevent information leakage). The precision consistently improves after round 1, but starts to decrease in some cases after round 2. Indeed, the mAP also drops after training on those labels with lower precision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Confidence Thresholds. We analyze the effects of hard and soft thresholds \\((\\tau_{\\mathrm{hard}},\\tau_{\\mathrm{soft}})\\) in Tab. 4. Prior works [59, 65] often use a high threshold of 0.9, while we observe clearly lower mAPs when \\(\\tau_{\\mathrm{hard}} > 0.7\\). With a lower \\(\\tau_{\\mathrm{hard}}\\), we retain most of the detected objects, and suppress noisy labels with \\(\\tau_{\\mathrm{soft}}\\). LEOD achieves similar results using several sets of hyper-parameters, showing our robustness to them. See Appendix B for more analysis on the filtering thresholds." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.286, + 0.865, + 0.394 + ], + "angle": 0, + "content": "
(τhard, car, τsoft, car, τsoft, ped)1%2%5%10%
(0.5, 0.6, 0.3)38.941.843.545.4
(0.6, 0.7, 0.35)39.442.243.845.5
(0.6, 0.7, 0.4)39.342.043.745.6
(0.6, 0.8, 0.4)39.342.343.645.5
(0.7, 0.8, 0.4)39.042.043.244.9
(0.8, -, -)*38.441.442.744.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Table 4. Ablation of hard and soft thresholds used in soft anchor assignment. We keep \\( \\tau_{\\text{hard, ped}} = \\tau_{\\text{hard, car}} / 2 \\). Ped stands for pedestrian. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. * indicates not using soft anchor assignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.462, + 0.618, + 0.478 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.594 + ], + "angle": 0, + "content": "We present LEOD, the first algorithm for label-efficient event-based object detection. To leverage unlabeled data, we adopt the self-training framework with reliable label selection. Several techniques are introduced to improve labeling quality. Extensive experiments on Gen1 and 1Mpx datasets showcase the superiority of our method over baselines in all the settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Limitations and Future Work. Following common practice, we only conduct intra-dataset experiments, i.e., training on data gathered using the same protocol. Recent works [3, 25, 40] have shown that training large models over multiple datasets leads to excellent performance and generalization. Since LEOD benefits from unlabeled data, we can also train it jointly on more datasets that involve real-world multi-object event sequences [4, 15, 73]. We discuss some failure cases of our pseudo-labeling method in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.646, + 0.758 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research was partially supported by Compute Ontario (www.computeontario.ca), the Province of Ontario, the Digital Research Alliance of Canada (https://alliancecan.ca/), and the Government of Canada through CIFAR and the Natural Sciences and Engineering Research Council of Canada (NSERC). We would also like to acknowledge Vector Institute for computation support. We would like to thank Yash Kant, Lunjun Zhang, Xuanchi Ren, and Xiaoshi Wu for valuable discussions and support." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16940" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.186 + ], + "angle": 0, + "content": "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[4] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In CVPRW, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.321, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[5] Nicholas FY Chen. Pseudo-labels for supervised learning on dynamic vision sensor data, applied to object detection under ego-motion. In CVPRW, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.365, + 0.468, + 0.42 + ], + "angle": 0, + "content": "[6] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[7] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IJCNN, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.468, + 0.523 + ], + "angle": 0, + "content": "[8] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.526, + 0.468, + 0.595 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.598, + 0.468, + 0.709 + ], + "angle": 0, + "content": "[10] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a \\(1280 \\times 720\\) back-illuminated stacked temporal contrast event-based vision sensor with \\(4.86\\mu m\\) pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.712, + 0.468, + 0.768 + ], + "angle": 0, + "content": "[11] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. TPAMI, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.77, + 0.468, + 0.811 + ], + "angle": 0, + "content": "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[13] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. RA-L, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In CVPR, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[17] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 1997. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[18] Yuhuang Hu, Tobi Delbruck, and Shih-Chii Liu. Learning to exploit multiple vision modalities by using grafted networks. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[19] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[20] Massimiliano Iacono, Stefan Weber, Arren Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. In IROS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Mingcheng Ji, Ziling Wang, Rui Yan, Qingjie Liu, Shu Xu, and Huajin Tang. Sctn: Event-based object tracking with energy-efficient deep convolutional spiking neural networks. Frontiers in Neuroscience, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[22] Zhuangyi Jiang, Pengfei Xia, Kai Huang, Walter Stechele, Guang Chen, Zhenshan Bing, and Alois Knoll. Mixed frame-event-driven fast pedestrian detection. In ICRA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[23] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.464, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[26] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICMLW, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[28] Hengduo Li, Zuxuan Wu, Chen Zhu, Caiming Xiong, Richard Socher, and Larry S Davis. Learning from noisy anchors for one-stage object detection. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[29] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Minghua Liu, Yin Zhou, Charles R Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In ECCV, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16941" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[34] Yen-Cheng Liu, Chih-Yao Ma, and Zsolt Kira. Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[35] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.469, + 0.292 + ], + "angle": 0, + "content": "[36] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.469, + 0.349 + ], + "angle": 0, + "content": "[37] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. RA-L, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[38] Ishan Misra, Abhinav Shrivastava, and Martial Hebert. Watch and learn: Semi-supervised learning for object detectors from video. In CVPR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[39] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. *IJRR*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[40] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[41] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1742-1750, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.469, + 0.638 + ], + "angle": 0, + "content": "[42] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 2020. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.641, + 0.469, + 0.695 + ], + "angle": 0, + "content": "[43] Christoph Posch, Daniel Matolin, and Rainer Wohlgannnt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.469, + 0.739 + ], + "angle": 0, + "content": "[44] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.468, + 0.769 + ], + "angle": 0, + "content": "[45] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.469, + 0.812 + ], + "angle": 0, + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[47] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. TPAMI, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[49] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[51] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[53] Krishna Kumar Singh, Fanyi Xiao, and Yong Jae Lee. Track and transfer: Watching videos to simulate strong human supervision for weakly-supervised object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3548-3556, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[54] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[55] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[56] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[57] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[58] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[59] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2021. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[60] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[61] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[63] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[64] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16942" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[65] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, 2021. 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[66] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[67] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[68] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[69] Alessandro Zanardi, Andreas Aumiller, Julian Zilly, Andrea Censi, and Emilio Frazzoli. Cross-modal learning filters for rgb-neuromorphic wormhole learning. RSS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[70] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[71] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[72] Yongqiang Zhang, Yancheng Bai, Mingli Ding, Yongqiang Li, and Bernard Ghanem. W2f: A weakly-supervised to fully-supervised framework for object detection. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[73] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. RA-L, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[74] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[75] Alex Zihao Zhu, Ziyun Wang, Kaung Khant, and Kostas Daniilidis. Eventgan: Leveraging large scale image datasets for event cameras. In ICCP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[76] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In ICCV, 2023. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16943" + } + ] +] \ No newline at end of file diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_origin.pdf b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ff5cb2b79c17b1f3c416d0b76b07238ddc5976b --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/dfaca580-4023-4cf1-b4f1-9b27b658deef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49a9b63b3d6c3038eaa3628ad33f349c05922dcb9d748c323575bcd4df19557e +size 1530534 diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/full.md b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bcb163a21783b313415262cf738227b6bf2e2891 --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/full.md @@ -0,0 +1,300 @@ +# LEOD: Label-Efficient Object Detection for Event Cameras + +Ziyi Wu $^{1,2}$ , Mathias Gehrig $^{3}$ , Qing Lyu $^{1}$ , Xudong Liu $^{1}$ , Igor Gilitschenski $^{1,2}$ , $^{1}$ University of Toronto, $^{2}$ Vector Institute, $^{3}$ University of Zurich + +# Abstract + +Object detection with event cameras benefits from the sensor's low latency and high dynamic range. However, it is costly to fully label event streams for supervised training due to their high temporal resolution. To reduce this cost, we present LEOD, the first method for label-efficient event-based detection. Our approach unifies weakly- and semi-supervised object detection with a self-training mechanism. We first utilize a detector pre-trained on limited labels to produce pseudo ground truth on unlabeled events. Then, the detector is re-trained with both real and generated labels. Leveraging the temporal consistency of events, we run bi-directional inference and apply tracking-based post-processing to enhance the quality of pseudo labels. To stabilize training against label noise, we further design a soft anchor assignment strategy. We introduce new experimental protocols to evaluate the task of label-efficient event-based detection on Gen1 and 1Mpx datasets. LEOD consistently outperforms supervised baselines across various labeling ratios. For example, on Gen1, it improves mAP by $8.6\%$ and $7.8\%$ for RVT-S trained with $1\%$ and $2\%$ labels. On 1Mpx, RVT-S with $10\%$ labels even surpasses its fully-supervised counterpart using $100\%$ labels. LEOD maintains its effectiveness even when all labeled data are available, reaching new state-of-the-art results. Finally, we show that our method readily scales to improve larger detectors as well. Code: https://github.com/Wuziyi616/LEOD. + +# 1. Introduction + +Object detection is key to scene understanding. It provides a compact representation of raw sensor measurements as semantically meaningful bounding boxes. Speed is crucial in object detection, especially in safety-critical applications such as self-driving. Recently, event cameras have gained significant interest in computer vision due to their low latency, low energy consumption, and high dynamic range [11]. Leveraging these benefits, event-based object detectors [14, 16, 29, 42, 49, 76] have been developed to complement conventional frame-based detectors. Despite tremendous progress, much of their success heavily relies + +![](images/c8a44feb843cafde48f4053b7165d8de07b65daa6095ec07000be7c7300fd453.jpg) +Figure 1. Detection performance of LEOD and baselines trained only on labeled events or conducting naive self-training. Under the weakly-supervised setting, our method consistently improves the RVT-S detector [14] across all labeling ratios on the Gen1 dataset. + +on large datasets that are manually annotated. However, due to the high temporal resolution of event data, labeling objects at every timestamp is impractical. For example, Gen1 dataset [8] provides object labels at lower than $4\mathrm{Hz}$ . As a result, existing methods only train their models on labeled events, and discard the remaining unlabeled data. In contrast, we view this task as a weakly-supervised learning problem, which calls for detection methods that can leverage a mix of labeled and unlabeled events during training. + +In this work, we address the challenge by proposing a Label-efficient Event-based Object Detection (LEOD) framework. We consider two settings characterized by limited labels: (i) weakly-supervised, where object bounding boxes are sparsely labeled in all event streams, and (ii) semi-supervised, where some event streams have object boxes densely labeled while others remain fully unlabeled. Our approach unifies the two settings through a self-training paradigm. With limited labels, we first pre-train a detector and use it to generate pseudo annotations on unlabeled events. Then, we re-train the detector on a combination of real and pseudo labels. However, naively generated labels contain noise, making their direct use suboptimal. + +To obtain high-quality pseudo labels, we exploit the temporal dimension of event data. Recent work [29] has shown the importance of temporal information in event-based detection using recurrent modules [52]. Additionally, offline label generation enables us to further refine predictions with + +future information. To achieve this, we first introduce time-flip augmentation to events during training. As a result, we can ensemble model predictions on the original and reverse event streams via Test-Time Augmentation (TTA), leading to higher detection recall. In addition, we leverage tracking-based post-processing to eliminate temporally inconsistent objects, enhancing the precision of pseudo labels. Finally, we filter out low-confidence boxes with a score threshold. + +A key challenge here is how to select a proper threshold. Instead of searching for the optimal value, we first filter with a low threshold to avoid missing objects. This inevitably introduces many false positives, which we address with a soft anchor assignment strategy in training. When computing the detection loss, we set another higher threshold and only treat pseudo boxes above that threshold as positive labels. For boxes with a lower detection score, we ignore the loss applied to their associated anchors. This strategy ensures that the model is only supervised with reliable background and foreground labels, while being tolerant to noisy labels. Ablation studies show that our method is insensitive to the two threshold values, easing the hyper-parameter tuning. + +To test our method, we design new protocols for label-efficient event-based detection on Gen1 [8] and 1Mpx [42] datasets. For weakly-supervised object detection (WSOD), we uniformly sub-sample the labels over time to simulate sparse annotations. For semi-supervised object detection (SSOD), we directly choose some event sequences as fully unlabeled. Following 2D SSOD evaluation protocols [55, 65], we also have a fully labeled setting, where we show that pseudo labels can complement ground-truths. + +In summary, this work makes four main contributions: (i) We introduce the task of label-efficient object detection to event vision, and design its experimental protocols. (ii) We propose LEOD, a unified framework for training event-based detectors with limited annotations. (iii) LEOD consistently outperforms baselines in various settings on two public detection datasets. (iv) Our method remains effective under the fully labeled setting and scales up to larger detectors, achieving new state-of-the-art performance. + +# 2. Related Work + +Object Detection with Event Cameras. Existing event-based detectors can be mainly categorized into two classes depending on whether they utilize the asynchronous nature of events. One line of work explores the sparsity of events, and employs Graph Neural Networks (GNNs) [13, 36, 49] or Spiking Neural Networks (SNNs) [7, 21, 70] for feature extraction. However, these approaches struggle with propagating information over long time horizons, which is crucial for detecting objects from events. Moreover, specialized hardware is required to achieve theoretical speed-ups of sparse networks, limiting their application in practice. + +In another class of methods, events are converted to dense frame-like representations, followed by conventional networks for detection. Earlier works only consider event frames aggregated from a short time interval [5, 20, 22, 31]. This discards long-horizon history and makes it hard to detect objects under small relative motion as they trigger very few events. Recent methods thus introduce recurrent modules [17, 52] to enhance the memory of the detectors [14, 29, 42]. Further research focuses on better backbones [9, 58], inference speed [16], and event representations [76]. Because our primary goal is to study label-efficient learning for event-based detectors, we adopt the state-of-the-art approach RVT [14] as the base model. + +Label-Efficient Learning in Event-based Vision. Due to a lack of large labeled datasets, there have been several works studying event-based algorithms with limited labels. Some papers focus on bridging frame-based and event-based vision. They either reconstruct natural images from events to apply traditional deep models [46, 47, 50, 56, 62], simulate events from videos to transfer the annotations [19, 39, 45, 75], or distill knowledge from trained frame-based models [18, 37, 57, 60, 67, 69]. However, these methods require either paired recordings of events and images or massive in-domain labeled images for training. Closer to ours are methods that only use event data [6, 26, 63, 74]. They conduct label-efficient learning on events with pre-trained frame-based models or self-supervised losses. Yet, none of them are designed for the detection task. Our work is the first attempt at label-efficient event-based object detection. + +Label-Efficient Learning in Other Fields. Self-training based methods have been explored in tasks such as 2D image classification [27, 54, 64], object detection [55, 65, 72], and segmentation [1, 23, 41]. Our method is more related to label-efficient learning on videos [38, 53, 66] and 3D point cloud sequences [32, 44, 71] as these methods also exploit the temporal information of input data. For example, [66] utilizes optical flows to propagate single-frame labels to adjacent video frames. [44] and [32] train a teacher model on dense point clouds aggregated from a few timesteps. In contrast, we leverage a much longer temporal horizon by running the detector on the entire event stream in both directions. [71] also employs tracking-based post-processing to remove inconsistent boxes. We additionally perform tracking in both directions as a forward-backward consistency check, thus better leveraging the temporal information of event data. To tackle the noisy pseudo labels, [65] proposes to use the detection scores from the teacher model to weigh the loss. Instead, we design a soft anchor assignment strategy by ignoring the loss associated with unconfident boxes. + +# 3. Method + +This paper introduces a new task named label-efficient event-based object detection (formulated in Sec. 3.1). Our + +algorithm adopts a two-stage self-training framework with reliable label selection (Sec. 3.2), where we leverage the temporal information of event data to obtain high-quality pseudo labels and suppress noisy predictions (Sec. 3.3). + +# 3.1. Problem Formulation + +Events data. Event cameras record brightness changes, and output a sequence of events $\mathcal{E} = \{e_i = (x_i,y_i,t_i,p_i)\}$ . Each event $e_i$ is parameterized by its pixel coordinate $(x_{i},y_{i})$ , timestamp $t_i$ , and polarity $p_i\in \{-1,1\}$ . Modern event cameras run at sub-milliseconds and can produce millions of events per second [11]. + +Event-based Object Detection. As the object motion in a scene is usually much slower than the event generation speed, event-based object detectors are only applied and evaluated at a fixed time interval $T$ [8, 42]. More specifically, given an event stream $\mathcal{E}$ capturing a set of objects $\mathcal{O} = \{o_j\}_{j=1}^M$ , we aim at detecting the 2D bounding boxes of them with the semantic labels, $\mathcal{B} = \{b_j = (x_j, y_j, w_j, h_j, l_j, t_l)\}_{j=1}^M$ . Each bounding box $b_j$ is characterized by the location of its top-left corner $(x_j, y_j)$ , width $w_j$ , height $h_j$ , class label $l_j \in \{1, 2, \dots, C\}$ , and timestamp $t_j$ . Here, $C$ denotes the number of classes. + +Event-based Object Detectors. We take RVT [14] for example as it serves as our base model. RVT is a synchronous detector that converts events in every time window $\Delta t$ to a grid-like representation $I$ . In the remaining part of the paper, we call $I$ a frame and every $\Delta t$ a timestep. RVT combines a Vision Transformer backbone [58] with a YOLOX detection head [12]. To extract temporal features, RVT introduces LSTM [17] cells in the backbone to fuse information over multiple timesteps. This is useful for detecting slow-moving objects as they only generate a few events. The YOLOX detection head in RVT is anchor-free, i.e., for each location on the feature map (anchor point), it predicts an objectness score $p_{\mathrm{obj}} \in [0,1]$ , per-class IoU values $p_{\mathrm{iou}} \in \mathbb{R}^C$ , and offsets of the bounding box parameters $\Delta b = (\Delta x, \Delta y, \Delta w, \Delta h)$ . The $p_{\mathrm{iou}}$ is trained to output the IoU value between the predicted box and the matched ground-truth box of that class. To obtain the final prediction, Non-Maximum Suppression (NMS) is applied to remove overlapping bounding boxes with low confidence. + +During training, each ground-truth bounding box is matched to several anchor points for loss computation. Below we will denote anchor points as anchors for simplicity. Anchors matched with ground-truth are foreground, where all predicted values are supervised. For the remaining background anchors, only $p_{\mathrm{obj}}$ is trained to be 0. + +Label-Efficient Event-based Object Detection. Fig. 2 shows our proposed detection settings with limited labels. + +Weakly-supervised object detection (WSOD). In the WSOD setting, all event streams are sparsely labeled. Moreover, labels assigned to adjacent frames offer fewer informa + +![](images/bc51025cf0307266921a514e243a2d0f8c35e668a4a5433630c25c92fbfb5860.jpg) +(a) Weakly-supervised object detection (WSOD) + +![](images/3479030a55d255a4163e18d09ebfb84abe78cc46863567e54e430b58dbad36d5.jpg) +(b) Semi-supervised object detection (SSOD) +Figure 2. Illustration of two label-efficient event-based object detection settings: (a) weakly-supervised where all event sequences are sparsely annotated, and (b) semi-supervised where some event sequences are densely annotated, and others are fully unlabeled. We visualize both positive and negative events in black. Tive training signals compared to those distributed across frames [66]. Therefore, it is reasonable to label object boxes in a long event sequence uniformly and sparsely. + +Semi-supervised object detection (SSOD). In the SSOD setting, some event sequences are densely labeled, while others are fully unlabeled. This is also practical when people continue to collect data into already annotated datasets. Since capturing event sequences is much easier than labeling them, an algorithm that can consistently improve model performance with raw events is highly useful. + +To evaluate the label-efficient learning performance, we take existing event-based object detection datasets [8, 42] and sample a small portion of frames (WSOD) or sequences (SSOD) as labeled data. The rest of the training data are used as an unlabeled set following previous works [33, 55, 65]. We also have a fully labeled setting where all labels are available. Since the original event streams are annotated at a larger time interval than $\Delta t$ , we can still create pseudo labels on unlabeled timesteps to improve the performance. See Appendix D for further discussions on the two settings. + +# 3.2. LEOD: A Self-Training Framework + +As shown in Fig. 3, the overall pipeline of LEOD follows a student-teacher pseudo-labeling paradigm, which is applicable to both, the WSOD and the SSOD settings. We first pre-train a detector on labeled data using regular detection loss until convergence. Then, we employ it to annotate unlabeled frames. To leverage the benefit of offline prediction, + +![](images/9c41f6aca91a250db29d949f478ccc483c209d2bddbd46bb5828efe80bfb3183.jpg) +Figure 3. Overview of our LEOD pipeline. $①$ We first pre-train an event-based object detector on event streams with limited labels. $①$ To leverage the temporal information, we apply time-flip Test-Time Augmentation (TTA) to unlabeled event streams and ensemble the model predictions. $②$ We then apply forward and backward tracking to identify temporally inconsistent bounding boxes, i.e., boxes associated with short tracks. $③$ To handle noisy labels, a soft anchor assignment strategy is designed to ignore detection loss on unconfident pseudo labels (red boxes). $④$ We can boost the model performance by self-training on reliable pseudo labels (blue boxes) and repeating $① - ④$ . + +we apply temporal flip to get event streams in both directions, and aggregate the detection results on them. Since the teacher model is trained on limited data, it will be uncertain on hard examples. We thus threshold the boxes with a small value to keep more detected objects. To remove false positive boxes, we build upon the temporal persistency prior of objects and apply tracking-based post-processing. However, there might still be inaccurate labels due to the low confidence threshold we use, and directly training the detector on them will lead to suboptimal results. Inspired by prior works on noise-robust learning [28, 61], we design a soft anchor assignment strategy to selectively supervise the model with pseudo labels. Finally, we can use the re-trained detector as the teacher model to initialize the next round of self-training. The described process can be repeated for multiple rounds to further boost the model performance. + +Comparison to online pseudo-labeling. In previous label-efficient object detection works [33, 59, 65, 68], the teacher model is jointly trained with a student model. In each training step, the teacher predicts bounding boxes on unlabeled data in a batch for student training. This online paradigm is also applicable to our setting. However, the teacher model will only see short event streams loaded in a batch. For example, on Gen1 [8], our training sequence length accounts for a duration of 1 second, while a car can stop for more than 10 seconds in real-world traffic and thus trigger no events. Pseudo-labeling on short event streams will inevitably miss these objects. Therefore, we adopt our two-stage offline label generation paradigm to retain full temporal information. + +# 3.3. Towards High-Quality Pseudo Labeling + +In this section, we introduce each key component in our LEOD framework to achieve high-quality pseudo labels. + +Test-Time Augmentation (TTA). When deployed in the real world, event-based detectors are expected to run in real-time, i.e., they only take in events triggered before $t$ to de + +![](images/c93f578427f7d4a67bf8ef0a9d885729edcb8a5666c87ab2927dc14f61c739c8.jpg) +Figure 4. Illustration of the time-flip TTA which enhances our robustness against different object motions. Forward helps detect receding objects, while Backward helps with approaching objects. + +tect objects at $t$ . Instead, in our offline label generation process, we can use future information to refine predictions at the current timestep. As shown in Fig. 4, we run the detector on both the original and reversed event streams, enabling us to detect objects with different movements. We also apply a horizontal-flip TTA to further improve the detection. + +Filtering and Tracking. TTA helps us detect more objects (higher recall), yet it also leads to false positives (lower precision). Previous works simply use a threshold to filter out boxes with low confidence [59, 65]. However, as shown in Fig. 5 (a) and (b), there is a trade-off between precision and recall, making it hard to find the optimal threshold. We opt to first filter with a low threshold $\tau_{\mathrm{hard}}$ to avoid missing objects, followed by tracking-based post-processing to remove temporally inconsistent boxes. We follow the tracking-by-detection paradigm [2] to build tracks by linking detection boxes between frames. Each box $b$ will be associated with a track $s_k = \{(b, v_x, v_y)_t, k, n, q\}$ , where $(v_x, v_y)$ is the estimated velocity under a linear motion assumption, $k$ is the ID, $n$ is its length so far, and $q \in [0,1]$ is the current score. In the first frame, we initialize each box as a track. For every coming frame, we first predict the positions of existing tracks and associate them with boxes at that frame via greedy matching over pairwise IoUs. Then, we decay the score $q$ of unmatched tracks and initialize unmatched + +![](images/772c1175706c62c5f7a5cfc0ecaf97cff84a2b32c939cb3097ac318a69cffbf7.jpg) +(a) Label quality of cars + +![](images/7a872876c6b21d37919bb82c95d1d7c1e99e35597bb44027cb5869abe804d212.jpg) +Figure 5. Analysis on confidence thresholds. We randomly sample 10,000 predicted boxes from RVT-S pre-trained on $5\%$ of Gen1 labels. We plot the pseudo labels' precision and recall of (a) cars and (b) pedestrians. In (c) and (d), we show each box's predicted confidence scores and its true IoU with ground-truth boxes. $\tau_{\mathrm{hard}}$ is the threshold for initial filtering, and $\tau_{\mathrm{soft}}$ is used in soft anchor assignment. + +![](images/1f39354898df9b83cbe82d948ac6bf6ad89196f84e7ab7c4bba0f5cfe4aaf8d0.jpg) +(b) Label quality of pedestrians +(c) Scores vs. True IoUs of cars + +![](images/36eec798a9c248dcfe7f138050a4311061169575db6d36f824d98ec1ceb2a60c.jpg) +(d) Scores vs. True IoUs of pedestrians + +boxes as new tracks. Finally, tracks with low scores will be deleted. See Appendix A.1 for implementation details. + +Similar to TTA, we apply tracking in both directions, and only remove a box if the length of its associated track is shorter than a threshold, $T_{\text{trk}}$ , in both cases. However, predictions on hard examples may also be inconsistent, as the pre-trained model has limited capacity. Instead of suppressing removed boxes as background, we ignore them during loss computation as will be described later. Also, for long tracks, we inpaint boxes with linear motion at unmatched timesteps and ignore training losses on them too. + +Soft Anchor Assignment and Re-training. We can now re-train a detector on the ground-truth labels and the pseudo labels with the original detection loss. However, as shown in Fig. 5 (c) and (d), there are still low-quality boxes after post-processing. To handle noisy labels, we utilize a soft anchor assignment strategy to selectively supervise the model training. We first identify a set of uncertain labels including boxes belonging to short tracks, inpainted from long tracks, and those with detection scores lower than a threshold $\tau_{\mathrm{soft}}$ . Then, at each training step, we ignore the loss applied to anchors associated with these uncertain boxes, i.e., we do not supervise those anchors and allow them to discover new instances. This design is inspired by the anchor assignment in anchor-based detectors [30, 48], where two thresholds are used to determine foreground or background anchor boxes, and the anchors in between are ignored in loss computation. As we will show in ablation studies, soft anchor assignment makes our method less sensitive to hyper-parameters. + +Despite training on noisy labels, the model learns to refine the labels and detect new objects. Thus, we do an additional round of self-training to further improve the results. + +# 4. Experiments + +Sec. 4.2 shows that LEOD outperforms baselines significantly in the low-label regime. In the fully labeled setting, we surpass previous state-of-the-art (Sec. 4.3). Our ablations show the contribution of each component in Sec. 4.4. + +# 4.1. Experimental Setup + +Datasets. We adopt Gen1 [8] and 1Mpx [42] datasets that feature various driving scenarios. Gen1 consists of 39 hours + +of recordings with a $304 \times 240$ resolution event camera [43]. It provides bounding box annotations of cars and pedestrians at 1, 2, or $4\mathrm{Hz}$ . 1Mpx is recorded with a higher $720 \times 1280$ resolution event camera [10]. It contains around 15 hours of data collected over several months at day and night, and provides labels for cars, pedestrians, and two-wheelers at 30 or $60\mathrm{Hz}$ . Following previous works [14, 29], we remove ground-truth boxes that are too small during evaluation on both datasets, and half the events' resolution to $360 \times 640$ on 1Mpx. + +Evaluation Protocol. Mean average precision (mAP) is adopted as the main performance metric. We choose $1\%$ , $2\%$ , $5\%$ , and $10\%$ as the labeling ratio following prior works [55, 65]. In the weakly-supervised object detection (WSOD) setting, labels in all event streams are uniformly sub-sampled. In the semi-supervised object detection (SSOD) setting, we keep a small portion of event streams unchanged, while setting other event sequences as fully unlabeled. For the same labeling ratio, the amounts of available labels in WSOD and SSOD are roughly the same. Finally, all labels are provided in the fully labeled setting. + +Detector Training Details. We adopt the state-of-the-art event-based detector RVT [14] as our base model. Due to limited computation resources, we mainly experiment with RVT-S, while we show that LEOD also scales to the largest RVT-B variant in Sec. 4.4. Most of the configurations are the same as we build upon their open-source codebase. Here we only highlight our modifications. In order to apply the time-flip TTA, we train with an additional time-flip data augmentation. When re-training on pseudo labels, we initialize RVT from scratch and use the Adam optimizer [24] with a peak learning rate of $5 \times 10^{-4}$ to train for 150k iterations. Please refer to Appendix A.2 for more details. + +Pseudo-Labeling Details. Inspired by prior works [51, 59], we set different thresholds for each category. To simplify parameter tuning, we follow two rules: (i) pedestrians and two-wheelers share the same values, which are half of cars' values, (ii) for cars, the soft threshold $\tau_{\mathrm{soft}}$ equals to the hard threshold $\tau_{\mathrm{hard}} + 0.1$ , while for pedestrians and two-wheelers, we use $\tau_{\mathrm{soft}} = \tau_{\mathrm{hard}} + 0.05$ . In both settings and both datasets, we choose the same set of hyper-parameters: $\tau_{\mathrm{hard}} = 0.6$ for cars and $\tau_{\mathrm{hard}} = 0.3$ for pedestrians. Only in + +![](images/cf0108c42a47548ebc311a45a5e0ddc1196c9814246dc6d47b45f4acd4300f0c.jpg) +(a) Gen1 weakly-supervised object detection (WSOD) results + +![](images/be0501d3066a3b6dc6eab8084e1e6a82615a0b34228e43163c2207251093ea5b.jpg) +(b) 1Mpx weakly-supervised object detection (WSOD) results + +![](images/26b248e9a6dd897e211043cc5dd251581f53c0c332af40dbdcaabeeb23f49e75.jpg) +(c) Gen1 semi-supervised object detection (SSOD) results + +![](images/0bbfe38af1c43e77d67987ace07497bf28e25967ed2b0aee27c1d96b530b18cd.jpg) +(d) 1Mpx semi-supervised object detection (SSOD) results +Figure 6. Results on label-efficient learning using different ratios of labeled data. We compare LEOD with baselines on Gen1 and 1Mpx datasets under WSOD and SSOD using the RVT-S detector. All results are averaged over three runs. + +1Mpx WSOD, we set $\tau_{\mathrm{hard}} = 0.5$ for pedestrians and two-wheelers to handle excessive false positives. The minimum track length $T_{trk}$ is set to 6 in all experiments. + +Baselines. We compare with a Supervised Baseline trained only on available labels. Since we are the first work to consider this task, we design two other baselines: 1) Vanilla Self-Training: trains on pseudo labels without TTA, tracking, and soft anchor assignment; 2) Soft Teacher: adopts the online student-teacher paradigm from a representative 2D SSOD method [65]. We enable soft anchor assignment in Soft Teacher, while TTA and tracking are not applicable as the online event sequence length is too short. We tune baselines' hyper-parameters to be optimal in each setting. We also tried a state-of-the-art 2D SSOD method designed for anchor-free detectors [34] as YOLOX is anchor-free, but we did not observe clear improvements over Soft Teacher. + +# 4.2. Label-Efficient Results + +We first compare our method with baselines in the low-labeled data regime. The overall results are shown in Fig. 6. WSOD. Fig. 6 (a) and (b) present the weakly-supervised results. On Gen1, LEOD improves the mAP of Supervised Baseline by a large margin across all labeling ratios. Using $10\%$ labels, we achieve an mAP of $45.5\%$ , which is only $1\%$ lower than RVT-S trained on $100\%$ labels. For pseudolabeling baselines, Vanilla Self-Training outperforms Soft Teacher in most cases, validating our choice of offline la + +bel generation. In addition, LEOD consistently outperforms them by more than $2\%$ , indicating the higher quality of our pseudo labels. We observe a similar trend on 1Mpx, where our approach scores the highest mAP in all cases. Notably, the two pseudo-labeling baselines perform worse than pretrained RVT-S on $5\%$ and $10\%$ labels, which proves the importance of screening reliable labels. Finally, LEOD with $10\%$ labels $(44.6\%)$ outperforms RVT-S trained on all labels $(44.1\%)$ , showing the great potential of unlabeled data. SSOD. Fig. 6 (c) and (d) present the semi-supervised results. Using the same amount of labels, models trained under SSOD are generally worse than WSOD. This indicates that given a limited budget, we should sparsely label as many event streams as possible instead of densely labeling a few sequences. Nevertheless, LEOD still outperforms baselines by more than $2\%$ mAP over all labeling ratios on both datasets. Our results offer a promising direction of boosting performance with fully unlabeled event data. + +Qualitative Results. Fig. 7 visualizes some detection results of RVT-S trained with $10\%$ labels under Gen1 WSOD setting. The supervised baseline can only detect objects that trigger lots of events due to its limited capacity. Models using vanilla self-training detect more objects, but also produce numerous false positives. With our pseudo-labeling pipeline, LEOD trained models are able to handle various hard examples. For example, it discovers a car that was initially missed in the ground-truth annotation in Fig. 7 (d). + +![](images/797a7c13089971173011a37c949d52dcbe76a2ae9d1bd88525e79a381a9e2e40.jpg) +(a) + +![](images/d5801f9f88143a13dc3d8db8ad26400bb84412ca3ccb3319c9b5fa08df241414.jpg) +(b) + +![](images/88246e8084f0fedc201e6723d757c8c652f0167ed168d8b8a968b40eaa1d0e5e.jpg) +(c) + +![](images/cc34ac774af87d9a2b4e4ddae8080dd25df1a98c017f2647a09e1ba26dc5b66b.jpg) +(d) +Figure 7. Detection on Gen1. We show (a) a common driving scene, (b) cars with small relative movement thus few events, (c) crowded pedestrians, and (d) LEOD discovers a missing object in ground-truth annotations due to the occlusion of a barrier gate. Car and pedestrian boxes are colored in blue and orange, respectively. Pre-trained and Self-Training stand for Supervised Baseline and Vanilla Self-Training. + +# 4.3. Fully-Labeled Results + +Since the original labeling frequency on both datasets is lower than the frame rate of RVT-S (20 Hz), we can still create pseudo labels on unlabeled frames to improve the fully supervised model performance. We compare with state-of-the-art event-based object detectors trained on all labeled data in Tab. 1. LEOD improves over RVT-S by $2.2\%$ and $2.6\%$ on Gen1 and 1Mpx, respectively. On Gen1, our method achieves new state-of-the-art among models not using pre-trained weights. This indicates that LEOD is consistently effective even with $100\%$ labels. In terms of runtime and model size, since our approach does not introduce new modules to the base model, we are as efficient as RVT-S. + +# 4.4. Ablation Studies + +Larger Base Model. We evaluate our label-efficient learning framework on a stronger event-based detector RVT-B, which uses a larger ViT backbone compared to RVT-S. Tab. 2 presents the result in the Gen1 WSOD setting. With more parameters, RVT-B pre-trained on limited labels already outperforms RVT-S. Still, LEOD is able to improve the detection result by a sizeable margin across all labeling ratios. With $100\%$ labels, our method achieves an mAP of $50.2\%$ , which is competitive with ERGO-12 using a large-scale pre-trained Swin Transformer V2 backbone [35]. No + +
MethodGen11MpxSize (M)
mAP (%)Time (ms)mAP (%)Time (ms)
RED [42]40.016.743.039.324.1
ASTMNet [29]46.735.648.372.3>100
HMNet-L3 [16]47.17.9*--33.2
RVT-B47.210.247.411.918.5
RVT-S46.59.544.110.19.9
LEOD-RVT-S48.79.546.710.19.9
ERGO-12 [76]50.477.240.6101.159.6
+ +Table 1. Detection results using all available labels. Baseline runtimes and model sizes are obtained from [14]. For HMNet, we use the best-performing L3 variant. * Its runtime is computed using a V100 GPU, while T4 GPUs which are slower than V100 are used in the other cases. ERGO is grayed as it is the only method using pre-trained models (Swin Transformer V2 [35]). + +tably, LEOD brings larger absolute improvements on RVT-B compared to RVT-S, proving that our framework steadily scales up to enhance larger and stronger detectors. + +Effect of Each Component. Fig. 8 shows the model performance with different components on Gen1 and 1Mpx. TTA significantly improves the mAP as it increases the recall in the generated pseudo labels using future information. Leveraging the temporal persistency of objects, tracking-based post-processing leads to a further gain. Finally, with soft anchor assignment, only reliable foreground and background labels are selected, easing the model training. + +![](images/b8c6e993cae07afc045992933898b1220f7a9fb103c814b6153f403ef667799d.jpg) +(a) Gen1 weakly-supervised object detection (WSOD) results + +![](images/f69ad8292f54da85af719da86c47ccaa1fd043b40df8b4f877ece802e626dd88.jpg) +(b) 1Mpx weakly-supervised object detection (WSOD) results +Figure 8. Ablation study of each component in LEOD. We report the WSOD result with RVT-S as the base detector on both datasets. Starting from Vanilla Self-Training, we gradually add TTA, tracking-based post-processing, and soft anchor assignment. + +
Method1%2%5%10%100%
RVT-S30.834.437.940.846.5
LEOD-RVT-S39.442.243.845.648.7
Absolute Improvement+8.6+7.8+5.9+4.8+2.2
RVT-B31.634.838.341.047.6*
LEOD-RVT-B40.042.945.346.650.2
Absolute Improvement+8.4+8.1+7.0+5.6+2.6
+ +Table 2. Gen1 WSOD mAPs (%) with two RVT variants. * our reproduced RVT-B using ${100}\%$ labels result is better than [14]. + +
Rounds1%2%5%10%100%P. (1%)P. (2%)P. (5%)P. (10%)
138.141.143.145.348.50.650.690.740.79
239.442.243.845.648.70.720.750.770.81
339.542.243.645.448.60.720.740.740.76
+ +Table 3. Number of self-training rounds used in LEOD. We report the mAP (\%) result in Gen1 WSOD using RVT-S. We also compute the precision (P.) of pseudo labels that are used to train the models on that row, which is a good indicator for stop training. + +Self-Training Rounds. Tab. 3 left presents the results of multi-round self-training. Detectors after the first round of self-training are significantly better than pre-trained models, and thus generate higher quality pseudo labels. Therefore, a second round of training leads to consistent gains. However, due to error accumulations, a third round of training may result in worse models. To determine when to stop training, we empirically find that the precision of pseudo labels serves as a good indicator. In Tab. 3 right, we compute the precision of predicted boxes on labeled frames (skipped labels are not used to prevent information leakage). The precision consistently improves after round 1, but starts to decrease in some cases after round 2. Indeed, the mAP also drops after training on those labels with lower precision. + +Confidence Thresholds. We analyze the effects of hard and soft thresholds $(\tau_{\mathrm{hard}},\tau_{\mathrm{soft}})$ in Tab. 4. Prior works [59, 65] often use a high threshold of 0.9, while we observe clearly lower mAPs when $\tau_{\mathrm{hard}} > 0.7$ . With a lower $\tau_{\mathrm{hard}}$ , we retain most of the detected objects, and suppress noisy labels with $\tau_{\mathrm{soft}}$ . LEOD achieves similar results using several sets of hyper-parameters, showing our robustness to them. See Appendix B for more analysis on the filtering thresholds. + +
(τhard, car, τsoft, car, τsoft, ped)1%2%5%10%
(0.5, 0.6, 0.3)38.941.843.545.4
(0.6, 0.7, 0.35)39.442.243.845.5
(0.6, 0.7, 0.4)39.342.043.745.6
(0.6, 0.8, 0.4)39.342.343.645.5
(0.7, 0.8, 0.4)39.042.043.244.9
(0.8, -, -)*38.441.442.744.2
+ +Table 4. Ablation of hard and soft thresholds used in soft anchor assignment. We keep $\tau_{\text{hard, ped}} = \tau_{\text{hard, car}} / 2$ . Ped stands for pedestrian. We report the mAP (\%) result in Gen1 WSOD using RVT-S. * indicates not using soft anchor assignment. + +# 5. Conclusion + +We present LEOD, the first algorithm for label-efficient event-based object detection. To leverage unlabeled data, we adopt the self-training framework with reliable label selection. Several techniques are introduced to improve labeling quality. Extensive experiments on Gen1 and 1Mpx datasets showcase the superiority of our method over baselines in all the settings. + +Limitations and Future Work. Following common practice, we only conduct intra-dataset experiments, i.e., training on data gathered using the same protocol. Recent works [3, 25, 40] have shown that training large models over multiple datasets leads to excellent performance and generalization. Since LEOD benefits from unlabeled data, we can also train it jointly on more datasets that involve real-world multi-object event sequences [4, 15, 73]. We discuss some failure cases of our pseudo-labeling method in Appendix C. + +# Acknowledgments + +This research was partially supported by Compute Ontario (www.computeontario.ca), the Province of Ontario, the Digital Research Alliance of Canada (https://alliancecan.ca/), and the Government of Canada through CIFAR and the Natural Sciences and Engineering Research Council of Canada (NSERC). We would also like to acknowledge Vector Institute for computation support. We would like to thank Yash Kant, Lunjun Zhang, Xuanchi Ren, and Xiaoshi Wu for valuable discussions and support. + +# References + +[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 2 +[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 4 +[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 8 +[4] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In CVPRW, 2023. 8 +[5] Nicholas FY Chen. Pseudo-labels for supervised learning on dynamic vision sensor data, applied to object detection under ego-motion. In CVPRW, 2018. 2 +[6] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In ICCV, 2023. 2 +[7] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IJCNN, 2022. 2 +[8] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 1, 2, 3, 4, 5 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[10] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a $1280 \times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\mu m$ pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference, 2020. 5 +[11] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. TPAMI, 2020. 1, 3 +[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3 +[13] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. 2 +[14] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 1, 2, 3, 5, 7, 8 + +[15] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. RA-L, 2021. 8 +[16] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In CVPR, 2023. 1, 2, 7 +[17] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 1997. 2, 3 +[18] Yuhuang Hu, Tobi Delbruck, and Shih-Chii Liu. Learning to exploit multiple vision modalities by using grafted networks. In ECCV, 2020. 2 +[19] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR, 2021. 2 +[20] Massimiliano Iacono, Stefan Weber, Arren Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. In IROS, 2018. 2 +[21] Mingcheng Ji, Ziling Wang, Rui Yan, Qingjie Liu, Shu Xu, and Huajin Tang. Sctn: Event-based object tracking with energy-efficient deep convolutional spiking neural networks. Frontiers in Neuroscience, 2023. 2 +[22] Zhuangyi Jiang, Pengfei Xia, Kai Huang, Walter Stechele, Guang Chen, Zhenshan Bing, and Alois Knoll. Mixed frame-event-driven fast pedestrian detection. In ICRA, 2019. 2 +[23] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 2 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8 +[26] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022. 2 +[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICMLW, 2013. 2 +[28] Hengduo Li, Zuxuan Wu, Chen Zhu, Caiming Xiong, Richard Socher, and Larry S Davis. Learning from noisy anchors for one-stage object detection. In CVPR, 2020. 4 +[29] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 1, 2, 5, 7 +[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5 +[31] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 2023. 2 +[32] Minghua Liu, Yin Zhou, Charles R Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In ECCV, 2022. 2 + +[33] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2020. 3, 4 +[34] Yen-Cheng Liu, Chih-Yao Ma, and Zsolt Kira. Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In CVPR, 2022. 6 +[35] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7 +[36] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In ECCV, 2020. 2 +[37] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. RA-L, 2022. 2 +[38] Ishan Misra, Abhinav Shrivastava, and Martial Hebert. Watch and learn: Semi-supervised learning for object detectors from video. In CVPR, 2015. 2 +[39] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. *IJRR*, 2017. 2 +[40] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 8 +[41] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1742-1750, 2015. 2 +[42] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 2020. 1, 2, 3, 5, 7 +[43] Christoph Posch, Daniel Matolin, and Rainer Wohlgannnt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 2010. 5 +[44] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, 2021. 2 +[45] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 2 +[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 2 +[47] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. TPAMI, 2019. 2 +[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5 + +[49] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, 2022. 1, 2 +[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 2 +[51] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, 2020. 5 +[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 2015. 1, 2 +[53] Krishna Kumar Singh, Fanyi Xiao, and Yong Jae Lee. Track and transfer: Watching videos to simulate strong human supervision for weakly-supervised object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3548-3556, 2016. 2 +[54] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 2020. 2 +[55] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 3, 5 +[56] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, 2020. 2 +[57] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, 2022. 2 +[58] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2, 3 +[59] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2021. 4, 5, 8 +[60] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In CVPR, 2021. 2 +[61] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 4 +[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2 +[63] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2 +[64] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2 + +[65] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, 2021. 2, 3, 4, 5, 6, 8 +[66] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, 2019. 2, 3 +[67] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In ICCV, 2023. 2 +[68] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022.4 +[69] Alessandro Zanardi, Andreas Aumiller, Julian Zilly, Andrea Censi, and Emilio Frazzoli. Cross-modal learning filters for rgb-neuromorphic wormhole learning. RSS, 2019. 2 +[70] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 2 +[71] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In CVPR, 2023. 2 +[72] Yongqiang Zhang, Yancheng Bai, Mingli Ding, Yongqiang Li, and Bernard Ghanem. W2f: A weakly-supervised to fully-supervised framework for object detection. In CVPR, 2018. 2 +[73] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. RA-L, 2018. 8 +[74] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2 +[75] Alex Zihao Zhu, Ziyun Wang, Kaung Khant, and Kostas Daniilidis. Eventgan: Leveraging large scale image datasets for event cameras. In ICCP, 2021. 2 +[76] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In ICCV, 2023. 1, 2, 7 \ No newline at end of file diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/images.zip b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03ff8fd71cc4652ddcdba1220c5a18e4422b6e84 --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f5fba678a86d23610c3a60806b434652e4b07e10ef8b1ace10231e893fd8c0d +size 899653 diff --git a/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/layout.json b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e99b64d4ecd6fc0f7e03785b878238291cab73a --- /dev/null +++ b/2024/LEOD_ Label-Efficient Object Detection for Event Cameras/layout.json @@ -0,0 +1,9520 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 481, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 481, + 120 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 481, + 120 + ], + "type": "text", + "content": "LEOD: Label-Efficient Object Detection for Event Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": "Ziyi Wu" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": ", Mathias Gehrig" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": ", Qing Lyu" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": ", Xudong Liu" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": ", Igor Gilitschenski" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": "University of Toronto, " + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": "Vector Institute, " + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 142, + 479, + 174 + ], + "type": "text", + "content": "University of Zurich" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 192, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": "Object detection with event cameras benefits from the sensor's low latency and high dynamic range. However, it is costly to fully label event streams for supervised training due to their high temporal resolution. To reduce this cost, we present LEOD, the first method for label-efficient event-based detection. Our approach unifies weakly- and semi-supervised object detection with a self-training mechanism. We first utilize a detector pre-trained on limited labels to produce pseudo ground truth on unlabeled events. Then, the detector is re-trained with both real and generated labels. Leveraging the temporal consistency of events, we run bi-directional inference and apply tracking-based post-processing to enhance the quality of pseudo labels. To stabilize training against label noise, we further design a soft anchor assignment strategy. We introduce new experimental protocols to evaluate the task of label-efficient event-based detection on Gen1 and 1Mpx datasets. LEOD consistently outperforms supervised baselines across various labeling ratios. For example, on Gen1, it improves mAP by " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "7.8\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " for RVT-S trained with " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " labels. On 1Mpx, RVT-S with " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " labels even surpasses its fully-supervised counterpart using " + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 47, + 226, + 289, + 539 + ], + "type": "text", + "content": " labels. LEOD maintains its effectiveness even when all labeled data are available, reaching new state-of-the-art results. Finally, we show that our method readily scales to improve larger detectors as well. Code: https://github.com/Wuziyi616/LEOD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 561, + 127, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 127, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 127, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Object detection is key to scene understanding. It provides a compact representation of raw sensor measurements as semantically meaningful bounding boxes. Speed is crucial in object detection, especially in safety-critical applications such as self-driving. Recently, event cameras have gained significant interest in computer vision due to their low latency, low energy consumption, and high dynamic range [11]. Leveraging these benefits, event-based object detectors [14, 16, 29, 42, 49, 76] have been developed to complement conventional frame-based detectors. Despite tremendous progress, much of their success heavily relies" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 191, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 305, + 191, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 337 + ], + "type": "image", + "image_path": "c8a44feb843cafde48f4053b7165d8de07b65daa6095ec07000be7c7300fd453.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 338, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 383 + ], + "type": "text", + "content": "Figure 1. Detection performance of LEOD and baselines trained only on labeled events or conducting naive self-training. Under the weakly-supervised setting, our method consistently improves the RVT-S detector [14] across all labeling ratios on the Gen1 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 389, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 546, + 498 + ], + "type": "text", + "content": "on large datasets that are manually annotated. However, due to the high temporal resolution of event data, labeling objects at every timestamp is impractical. For example, Gen1 dataset [8] provides object labels at lower than " + }, + { + "bbox": [ + 304, + 389, + 546, + 498 + ], + "type": "inline_equation", + "content": "4\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 389, + 546, + 498 + ], + "type": "text", + "content": ". As a result, existing methods only train their models on labeled events, and discard the remaining unlabeled data. In contrast, we view this task as a weakly-supervised learning problem, which calls for detection methods that can leverage a mix of labeled and unlabeled events during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": "In this work, we address the challenge by proposing a Label-efficient Event-based Object Detection (LEOD) framework. We consider two settings characterized by limited labels: (i) weakly-supervised, where object bounding boxes are sparsely labeled in all event streams, and (ii) semi-supervised, where some event streams have object boxes densely labeled while others remain fully unlabeled. Our approach unifies the two settings through a self-training paradigm. With limited labels, we first pre-train a detector and use it to generate pseudo annotations on unlabeled events. Then, we re-train the detector on a combination of real and pseudo labels. However, naively generated labels contain noise, making their direct use suboptimal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "To obtain high-quality pseudo labels, we exploit the temporal dimension of event data. Recent work [29] has shown the importance of temporal information in event-based detection using recurrent modules [52]. Additionally, offline label generation enables us to further refine predictions with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16933" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "future information. To achieve this, we first introduce time-flip augmentation to events during training. As a result, we can ensemble model predictions on the original and reverse event streams via Test-Time Augmentation (TTA), leading to higher detection recall. In addition, we leverage tracking-based post-processing to eliminate temporally inconsistent objects, enhancing the precision of pseudo labels. Finally, we filter out low-confidence boxes with a score threshold." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 288, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 288, + 326 + ], + "type": "text", + "content": "A key challenge here is how to select a proper threshold. Instead of searching for the optimal value, we first filter with a low threshold to avoid missing objects. This inevitably introduces many false positives, which we address with a soft anchor assignment strategy in training. When computing the detection loss, we set another higher threshold and only treat pseudo boxes above that threshold as positive labels. For boxes with a lower detection score, we ignore the loss applied to their associated anchors. This strategy ensures that the model is only supervised with reliable background and foreground labels, while being tolerant to noisy labels. Ablation studies show that our method is insensitive to the two threshold values, easing the hyper-parameter tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 435 + ], + "type": "text", + "content": "To test our method, we design new protocols for label-efficient event-based detection on Gen1 [8] and 1Mpx [42] datasets. For weakly-supervised object detection (WSOD), we uniformly sub-sample the labels over time to simulate sparse annotations. For semi-supervised object detection (SSOD), we directly choose some event sequences as fully unlabeled. Following 2D SSOD evaluation protocols [55, 65], we also have a fully labeled setting, where we show that pseudo labels can complement ground-truths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": "In summary, this work makes four main contributions: (i) We introduce the task of label-efficient object detection to event vision, and design its experimental protocols. (ii) We propose LEOD, a unified framework for training event-based detectors with limited annotations. (iii) LEOD consistently outperforms baselines in various settings on two public detection datasets. (iv) Our method remains effective under the fully labeled setting and scales up to larger detectors, achieving new state-of-the-art performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 559, + 134, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 134, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 134, + 572 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Object Detection with Event Cameras. Existing event-based detectors can be mainly categorized into two classes depending on whether they utilize the asynchronous nature of events. One line of work explores the sparsity of events, and employs Graph Neural Networks (GNNs) [13, 36, 49] or Spiking Neural Networks (SNNs) [7, 21, 70] for feature extraction. However, these approaches struggle with propagating information over long time horizons, which is crucial for detecting objects from events. Moreover, specialized hardware is required to achieve theoretical speed-ups of sparse networks, limiting their application in practice." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "In another class of methods, events are converted to dense frame-like representations, followed by conventional networks for detection. Earlier works only consider event frames aggregated from a short time interval [5, 20, 22, 31]. This discards long-horizon history and makes it hard to detect objects under small relative motion as they trigger very few events. Recent methods thus introduce recurrent modules [17, 52] to enhance the memory of the detectors [14, 29, 42]. Further research focuses on better backbones [9, 58], inference speed [16], and event representations [76]. Because our primary goal is to study label-efficient learning for event-based detectors, we adopt the state-of-the-art approach RVT [14] as the base model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 228, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 418 + ], + "type": "text", + "content": "Label-Efficient Learning in Event-based Vision. Due to a lack of large labeled datasets, there have been several works studying event-based algorithms with limited labels. Some papers focus on bridging frame-based and event-based vision. They either reconstruct natural images from events to apply traditional deep models [46, 47, 50, 56, 62], simulate events from videos to transfer the annotations [19, 39, 45, 75], or distill knowledge from trained frame-based models [18, 37, 57, 60, 67, 69]. However, these methods require either paired recordings of events and images or massive in-domain labeled images for training. Closer to ours are methods that only use event data [6, 26, 63, 74]. They conduct label-efficient learning on events with pre-trained frame-based models or self-supervised losses. Yet, none of them are designed for the detection task. Our work is the first attempt at label-efficient event-based object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 658 + ], + "type": "text", + "content": "Label-Efficient Learning in Other Fields. Self-training based methods have been explored in tasks such as 2D image classification [27, 54, 64], object detection [55, 65, 72], and segmentation [1, 23, 41]. Our method is more related to label-efficient learning on videos [38, 53, 66] and 3D point cloud sequences [32, 44, 71] as these methods also exploit the temporal information of input data. For example, [66] utilizes optical flows to propagate single-frame labels to adjacent video frames. [44] and [32] train a teacher model on dense point clouds aggregated from a few timesteps. In contrast, we leverage a much longer temporal horizon by running the detector on the entire event stream in both directions. [71] also employs tracking-based post-processing to remove inconsistent boxes. We additionally perform tracking in both directions as a forward-backward consistency check, thus better leveraging the temporal information of event data. To tackle the noisy pseudo labels, [65] proposes to use the detection scores from the teacher model to weigh the loss. Instead, we design a soft anchor assignment strategy by ignoring the loss associated with unconfident boxes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 361, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "This paper introduces a new task named label-efficient event-based object detection (formulated in Sec. 3.1). Our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16934" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "algorithm adopts a two-stage self-training framework with reliable label selection (Sec. 3.2), where we leverage the temporal information of event data to obtain high-quality pseudo labels and suppress noisy predictions (Sec. 3.3)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 129, + 171, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 171, + 141 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 171, + 141 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": "Events data. Event cameras record brightness changes, and output a sequence of events " + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_i = (x_i,y_i,t_i,p_i)\\}" + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": ". Each event " + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": " is parameterized by its pixel coordinate " + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": ", and polarity " + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "inline_equation", + "content": "p_i\\in \\{-1,1\\}" + }, + { + "bbox": [ + 46, + 148, + 287, + 220 + ], + "type": "text", + "content": ". Modern event cameras run at sub-milliseconds and can produce millions of events per second [11]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": "Event-based Object Detection. As the object motion in a scene is usually much slower than the event generation speed, event-based object detectors are only applied and evaluated at a fixed time interval " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " [8, 42]. More specifically, given an event stream " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " capturing a set of objects " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = \\{o_j\\}_{j=1}^M" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", we aim at detecting the 2D bounding boxes of them with the semantic labels, " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{b_j = (x_j, y_j, w_j, h_j, l_j, t_l)\\}_{j=1}^M" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ". Each bounding box " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "b_j" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " is characterized by the location of its top-left corner " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "(x_j, y_j)" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", height " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", class label " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "l_j \\in \\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ", and timestamp " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 220, + 287, + 352 + ], + "type": "text", + "content": " denotes the number of classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": "Event-based Object Detectors. We take RVT [14] for example as it serves as our base model. RVT is a synchronous detector that converts events in every time window " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": " to a grid-like representation " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": ". In the remaining part of the paper, we call " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": " a frame and every " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": " a timestep. RVT combines a Vision Transformer backbone [58] with a YOLOX detection head [12]. To extract temporal features, RVT introduces LSTM [17] cells in the backbone to fuse information over multiple timesteps. This is useful for detecting slow-moving objects as they only generate a few events. The YOLOX detection head in RVT is anchor-free, i.e., for each location on the feature map (anchor point), it predicts an objectness score " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{obj}} \\in [0,1]" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": ", per-class IoU values " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{iou}} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": ", and offsets of the bounding box parameters " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Delta b = (\\Delta x, \\Delta y, \\Delta w, \\Delta h)" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{iou}}" + }, + { + "bbox": [ + 46, + 353, + 287, + 581 + ], + "type": "text", + "content": " is trained to output the IoU value between the predicted box and the matched ground-truth box of that class. To obtain the final prediction, Non-Maximum Suppression (NMS) is applied to remove overlapping bounding boxes with low confidence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": "During training, each ground-truth bounding box is matched to several anchor points for loss computation. Below we will denote anchor points as anchors for simplicity. Anchors matched with ground-truth are foreground, where all predicted values are supervised. For the remaining background anchors, only " + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{obj}}" + }, + { + "bbox": [ + 46, + 581, + 287, + 652 + ], + "type": "text", + "content": " is trained to be 0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "content": "Label-Efficient Event-based Object Detection. Fig. 2 shows our proposed detection settings with limited labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Weakly-supervised object detection (WSOD). In the WSOD setting, all event streams are sparsely labeled. Moreover, labels assigned to adjacent frames offer fewer informa" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 48, + 544, + 170 + ], + "blocks": [ + { + "bbox": [ + 306, + 48, + 544, + 170 + ], + "lines": [ + { + "bbox": [ + 306, + 48, + 544, + 170 + ], + "spans": [ + { + "bbox": [ + 306, + 48, + 544, + 170 + ], + "type": "image", + "image_path": "bc51025cf0307266921a514e243a2d0f8c35e668a4a5433630c25c92fbfb5860.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 171, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 347, + 171, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 347, + 171, + 504, + 180 + ], + "type": "text", + "content": "(a) Weakly-supervised object detection (WSOD)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 182, + 544, + 305 + ], + "blocks": [ + { + "bbox": [ + 306, + 182, + 544, + 305 + ], + "lines": [ + { + "bbox": [ + 306, + 182, + 544, + 305 + ], + "spans": [ + { + "bbox": [ + 306, + 182, + 544, + 305 + ], + "type": "image", + "image_path": "3479030a55d255a4163e18d09ebfb84abe78cc46863567e54e430b58dbad36d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 306, + 498, + 315 + ], + "lines": [ + { + "bbox": [ + 351, + 306, + 498, + 315 + ], + "spans": [ + { + "bbox": [ + 351, + 306, + 498, + 315 + ], + "type": "text", + "content": "(b) Semi-supervised object detection (SSOD)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 317, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 413 + ], + "type": "text", + "content": "Figure 2. Illustration of two label-efficient event-based object detection settings: (a) weakly-supervised where all event sequences are sparsely annotated, and (b) semi-supervised where some event sequences are densely annotated, and others are fully unlabeled. We visualize both positive and negative events in black. Tive training signals compared to those distributed across frames [66]. Therefore, it is reasonable to label object boxes in a long event sequence uniformly and sparsely." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 496 + ], + "type": "text", + "content": "Semi-supervised object detection (SSOD). In the SSOD setting, some event sequences are densely labeled, while others are fully unlabeled. This is also practical when people continue to collect data into already annotated datasets. Since capturing event sequences is much easier than labeling them, an algorithm that can consistently improve model performance with raw events is highly useful." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": "To evaluate the label-efficient learning performance, we take existing event-based object detection datasets [8, 42] and sample a small portion of frames (WSOD) or sequences (SSOD) as labeled data. The rest of the training data are used as an unlabeled set following previous works [33, 55, 65]. We also have a fully labeled setting where all labels are available. Since the original event streams are annotated at a larger time interval than " + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 497, + 545, + 617 + ], + "type": "text", + "content": ", we can still create pseudo labels on unlabeled timesteps to improve the performance. See Appendix D for further discussions on the two settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 496, + 635 + ], + "type": "text", + "content": "3.2. LEOD: A Self-Training Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, the overall pipeline of LEOD follows a student-teacher pseudo-labeling paradigm, which is applicable to both, the WSOD and the SSOD settings. We first pre-train a detector on labeled data using regular detection loss until convergence. Then, we employ it to annotate unlabeled frames. To leverage the benefit of offline prediction," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16935" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 49, + 545, + 212 + ], + "blocks": [ + { + "bbox": [ + 48, + 49, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 48, + 49, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 49, + 545, + 212 + ], + "type": "image", + "image_path": "9c41f6aca91a250db29d949f478ccc483c209d2bddbd46bb5828efe80bfb3183.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": "Figure 3. Overview of our LEOD pipeline. " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": " We first pre-train an event-based object detector on event streams with limited labels. " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": " To leverage the temporal information, we apply time-flip Test-Time Augmentation (TTA) to unlabeled event streams and ensemble the model predictions. " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": " We then apply forward and backward tracking to identify temporally inconsistent bounding boxes, i.e., boxes associated with short tracks. " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": " To handle noisy labels, a soft anchor assignment strategy is designed to ignore detection loss on unconfident pseudo labels (red boxes). " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": " We can boost the model performance by self-training on reliable pseudo labels (blue boxes) and repeating " + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "inline_equation", + "content": "① - ④" + }, + { + "bbox": [ + 46, + 214, + 546, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 468 + ], + "type": "text", + "content": "we apply temporal flip to get event streams in both directions, and aggregate the detection results on them. Since the teacher model is trained on limited data, it will be uncertain on hard examples. We thus threshold the boxes with a small value to keep more detected objects. To remove false positive boxes, we build upon the temporal persistency prior of objects and apply tracking-based post-processing. However, there might still be inaccurate labels due to the low confidence threshold we use, and directly training the detector on them will lead to suboptimal results. Inspired by prior works on noise-robust learning [28, 61], we design a soft anchor assignment strategy to selectively supervise the model with pseudo labels. Finally, we can use the re-trained detector as the teacher model to initialize the next round of self-training. The described process can be repeated for multiple rounds to further boost the model performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 624 + ], + "type": "text", + "content": "Comparison to online pseudo-labeling. In previous label-efficient object detection works [33, 59, 65, 68], the teacher model is jointly trained with a student model. In each training step, the teacher predicts bounding boxes on unlabeled data in a batch for student training. This online paradigm is also applicable to our setting. However, the teacher model will only see short event streams loaded in a batch. For example, on Gen1 [8], our training sequence length accounts for a duration of 1 second, while a car can stop for more than 10 seconds in real-world traffic and thus trigger no events. Pseudo-labeling on short event streams will inevitably miss these objects. Therefore, we adopt our two-stage offline label generation paradigm to retain full temporal information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 633, + 256, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 256, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 256, + 647 + ], + "type": "text", + "content": "3.3. Towards High-Quality Pseudo Labeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 677 + ], + "type": "text", + "content": "In this section, we introduce each key component in our LEOD framework to achieve high-quality pseudo labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Test-Time Augmentation (TTA). When deployed in the real world, event-based detectors are expected to run in real-time, i.e., they only take in events triggered before " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " to de" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 275, + 544, + 350 + ], + "blocks": [ + { + "bbox": [ + 307, + 275, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 307, + 275, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 544, + 350 + ], + "type": "image", + "image_path": "c93f578427f7d4a67bf8ef0a9d885729edcb8a5666c87ab2927dc14f61c739c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 353, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 305, + 353, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 546, + 386 + ], + "type": "text", + "content": "Figure 4. Illustration of the time-flip TTA which enhances our robustness against different object motions. Forward helps detect receding objects, while Backward helps with approaching objects." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 407, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 479 + ], + "type": "text", + "content": "tect objects at " + }, + { + "bbox": [ + 304, + 407, + 545, + 479 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 407, + 545, + 479 + ], + "type": "text", + "content": ". Instead, in our offline label generation process, we can use future information to refine predictions at the current timestep. As shown in Fig. 4, we run the detector on both the original and reversed event streams, enabling us to detect objects with different movements. We also apply a horizontal-flip TTA to further improve the detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Filtering and Tracking. TTA helps us detect more objects (higher recall), yet it also leads to false positives (lower precision). Previous works simply use a threshold to filter out boxes with low confidence [59, 65]. However, as shown in Fig. 5 (a) and (b), there is a trade-off between precision and recall, making it hard to find the optimal threshold. We opt to first filter with a low threshold " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " to avoid missing objects, followed by tracking-based post-processing to remove temporally inconsistent boxes. We follow the tracking-by-detection paradigm [2] to build tracks by linking detection boxes between frames. Each box " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " will be associated with a track " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "s_k = \\{(b, v_x, v_y)_t, k, n, q\\}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "(v_x, v_y)" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is the estimated velocity under a linear motion assumption, " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is the ID, " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is its length so far, and " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "q \\in [0,1]" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is the current score. In the first frame, we initialize each box as a track. For every coming frame, we first predict the positions of existing tracks and associate them with boxes at that frame via greedy matching over pairwise IoUs. Then, we decay the score " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " of unmatched tracks and initialize unmatched" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16936" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 53, + 160, + 151 + ], + "blocks": [ + { + "bbox": [ + 50, + 53, + 160, + 151 + ], + "lines": [ + { + "bbox": [ + 50, + 53, + 160, + 151 + ], + "spans": [ + { + "bbox": [ + 50, + 53, + 160, + 151 + ], + "type": "image", + "image_path": "772c1175706c62c5f7a5cfc0ecaf97cff84a2b32c939cb3097ac318a69cffbf7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 154, + 145, + 163 + ], + "lines": [ + { + "bbox": [ + 67, + 154, + 145, + 163 + ], + "spans": [ + { + "bbox": [ + 67, + 154, + 145, + 163 + ], + "type": "text", + "content": "(a) Label quality of cars" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 53, + 272, + 151 + ], + "blocks": [ + { + "bbox": [ + 164, + 53, + 272, + 151 + ], + "lines": [ + { + "bbox": [ + 164, + 53, + 272, + 151 + ], + "spans": [ + { + "bbox": [ + 164, + 53, + 272, + 151 + ], + "type": "image", + "image_path": "7a872876c6b21d37919bb82c95d1d7c1e99e35597bb44027cb5869abe804d212.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "text", + "content": "Figure 5. Analysis on confidence thresholds. We randomly sample 10,000 predicted boxes from RVT-S pre-trained on " + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "text", + "content": " of Gen1 labels. We plot the pseudo labels' precision and recall of (a) cars and (b) pedestrians. In (c) and (d), we show each box's predicted confidence scores and its true IoU with ground-truth boxes. " + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}}" + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "text", + "content": " is the threshold for initial filtering, and " + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{soft}}" + }, + { + "bbox": [ + 46, + 166, + 547, + 200 + ], + "type": "text", + "content": " is used in soft anchor assignment." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 276, + 54, + 407, + 151 + ], + "blocks": [ + { + "bbox": [ + 167, + 154, + 269, + 163 + ], + "lines": [ + { + "bbox": [ + 167, + 154, + 269, + 163 + ], + "spans": [ + { + "bbox": [ + 167, + 154, + 269, + 163 + ], + "type": "text", + "content": "(b) Label quality of pedestrians" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 276, + 54, + 407, + 151 + ], + "lines": [ + { + "bbox": [ + 276, + 54, + 407, + 151 + ], + "spans": [ + { + "bbox": [ + 276, + 54, + 407, + 151 + ], + "type": "image", + "image_path": "1f39354898df9b83cbe82d948ac6bf6ad89196f84e7ab7c4bba0f5cfe4aaf8d0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 154, + 392, + 163 + ], + "lines": [ + { + "bbox": [ + 290, + 154, + 392, + 163 + ], + "spans": [ + { + "bbox": [ + 290, + 154, + 392, + 163 + ], + "type": "text", + "content": "(c) Scores vs. True IoUs of cars" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 409, + 54, + 540, + 151 + ], + "blocks": [ + { + "bbox": [ + 409, + 54, + 540, + 151 + ], + "lines": [ + { + "bbox": [ + 409, + 54, + 540, + 151 + ], + "spans": [ + { + "bbox": [ + 409, + 54, + 540, + 151 + ], + "type": "image", + "image_path": "36eec798a9c248dcfe7f138050a4311061169575db6d36f824d98ec1ceb2a60c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 154, + 538, + 164 + ], + "lines": [ + { + "bbox": [ + 412, + 154, + 538, + 164 + ], + "spans": [ + { + "bbox": [ + 412, + 154, + 538, + 164 + ], + "type": "text", + "content": "(d) Scores vs. True IoUs of pedestrians" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 205, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 286, + 228 + ], + "type": "text", + "content": "boxes as new tracks. Finally, tracks with low scores will be deleted. See Appendix A.1 for implementation details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": "Similar to TTA, we apply tracking in both directions, and only remove a box if the length of its associated track is shorter than a threshold, " + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "inline_equation", + "content": "T_{\\text{trk}}" + }, + { + "bbox": [ + 46, + 229, + 286, + 336 + ], + "type": "text", + "content": ", in both cases. However, predictions on hard examples may also be inconsistent, as the pre-trained model has limited capacity. Instead of suppressing removed boxes as background, we ignore them during loss computation as will be described later. Also, for long tracks, we inpaint boxes with linear motion at unmatched timesteps and ignore training losses on them too." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 337, + 286, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 286, + 552 + ], + "type": "text", + "content": "Soft Anchor Assignment and Re-training. We can now re-train a detector on the ground-truth labels and the pseudo labels with the original detection loss. However, as shown in Fig. 5 (c) and (d), there are still low-quality boxes after post-processing. To handle noisy labels, we utilize a soft anchor assignment strategy to selectively supervise the model training. We first identify a set of uncertain labels including boxes belonging to short tracks, inpainted from long tracks, and those with detection scores lower than a threshold " + }, + { + "bbox": [ + 46, + 337, + 286, + 552 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{soft}}" + }, + { + "bbox": [ + 46, + 337, + 286, + 552 + ], + "type": "text", + "content": ". Then, at each training step, we ignore the loss applied to anchors associated with these uncertain boxes, i.e., we do not supervise those anchors and allow them to discover new instances. This design is inspired by the anchor assignment in anchor-based detectors [30, 48], where two thresholds are used to determine foreground or background anchor boxes, and the anchors in between are ignored in loss computation. As we will show in ablation studies, soft anchor assignment makes our method less sensitive to hyper-parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 552, + 286, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 286, + 588 + ], + "type": "text", + "content": "Despite training on noisy labels, the model learns to refine the labels and detect new objects. Thus, we do an additional round of self-training to further improve the results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 616, + 286, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 665 + ], + "type": "text", + "content": "Sec. 4.2 shows that LEOD outperforms baselines significantly in the low-label regime. In the fully labeled setting, we surpass previous state-of-the-art (Sec. 4.3). Our ablations show the contribution of each component in Sec. 4.4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 163, + 684 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Datasets. We adopt Gen1 [8] and 1Mpx [42] datasets that feature various driving scenarios. Gen1 consists of 39 hours" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": "of recordings with a " + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "inline_equation", + "content": "304 \\times 240" + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": " resolution event camera [43]. It provides bounding box annotations of cars and pedestrians at 1, 2, or " + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "inline_equation", + "content": "4\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": ". 1Mpx is recorded with a higher " + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "inline_equation", + "content": "720 \\times 1280" + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": " resolution event camera [10]. It contains around 15 hours of data collected over several months at day and night, and provides labels for cars, pedestrians, and two-wheelers at 30 or " + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": ". Following previous works [14, 29], we remove ground-truth boxes that are too small during evaluation on both datasets, and half the events' resolution to " + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "inline_equation", + "content": "360 \\times 640" + }, + { + "bbox": [ + 304, + 205, + 545, + 324 + ], + "type": "text", + "content": " on 1Mpx." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "content": "Evaluation Protocol. Mean average precision (mAP) is adopted as the main performance metric. We choose " + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 327, + 545, + 458 + ], + "type": "text", + "content": " as the labeling ratio following prior works [55, 65]. In the weakly-supervised object detection (WSOD) setting, labels in all event streams are uniformly sub-sampled. In the semi-supervised object detection (SSOD) setting, we keep a small portion of event streams unchanged, while setting other event sequences as fully unlabeled. For the same labeling ratio, the amounts of available labels in WSOD and SSOD are roughly the same. Finally, all labels are provided in the fully labeled setting." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "text", + "content": "Detector Training Details. We adopt the state-of-the-art event-based detector RVT [14] as our base model. Due to limited computation resources, we mainly experiment with RVT-S, while we show that LEOD also scales to the largest RVT-B variant in Sec. 4.4. Most of the configurations are the same as we build upon their open-source codebase. Here we only highlight our modifications. In order to apply the time-flip TTA, we train with an additional time-flip data augmentation. When re-training on pseudo labels, we initialize RVT from scratch and use the Adam optimizer [24] with a peak learning rate of " + }, + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 460, + 545, + 604 + ], + "type": "text", + "content": " to train for 150k iterations. Please refer to Appendix A.2 for more details." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Pseudo-Labeling Details. Inspired by prior works [51, 59], we set different thresholds for each category. To simplify parameter tuning, we follow two rules: (i) pedestrians and two-wheelers share the same values, which are half of cars' values, (ii) for cars, the soft threshold " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{soft}}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " equals to the hard threshold " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}} + 0.1" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", while for pedestrians and two-wheelers, we use " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{soft}} = \\tau_{\\mathrm{hard}} + 0.05" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". In both settings and both datasets, we choose the same set of hyper-parameters: " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}} = 0.6" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for cars and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}} = 0.3" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " for pedestrians. Only in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16937" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 48, + 294, + 188 + ], + "blocks": [ + { + "bbox": [ + 49, + 48, + 294, + 188 + ], + "lines": [ + { + "bbox": [ + 49, + 48, + 294, + 188 + ], + "spans": [ + { + "bbox": [ + 49, + 48, + 294, + 188 + ], + "type": "image", + "image_path": "cf0108c42a47548ebc311a45a5e0ddc1196c9814246dc6d47b45f4acd4300f0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 190, + 271, + 200 + ], + "lines": [ + { + "bbox": [ + 72, + 190, + 271, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 190, + 271, + 200 + ], + "type": "text", + "content": "(a) Gen1 weakly-supervised object detection (WSOD) results" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 48, + 544, + 188 + ], + "blocks": [ + { + "bbox": [ + 299, + 48, + 544, + 188 + ], + "lines": [ + { + "bbox": [ + 299, + 48, + 544, + 188 + ], + "spans": [ + { + "bbox": [ + 299, + 48, + 544, + 188 + ], + "type": "image", + "image_path": "be0501d3066a3b6dc6eab8084e1e6a82615a0b34228e43163c2207251093ea5b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 190, + 522, + 200 + ], + "lines": [ + { + "bbox": [ + 322, + 190, + 522, + 200 + ], + "spans": [ + { + "bbox": [ + 322, + 190, + 522, + 200 + ], + "type": "text", + "content": "(b) 1Mpx weakly-supervised object detection (WSOD) results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 206, + 293, + 346 + ], + "blocks": [ + { + "bbox": [ + 47, + 206, + 293, + 346 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 293, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 293, + 346 + ], + "type": "image", + "image_path": "26b248e9a6dd897e211043cc5dd251581f53c0c332af40dbdcaabeeb23f49e75.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 349, + 264, + 358 + ], + "lines": [ + { + "bbox": [ + 78, + 349, + 264, + 358 + ], + "spans": [ + { + "bbox": [ + 78, + 349, + 264, + 358 + ], + "type": "text", + "content": "(c) Gen1 semi-supervised object detection (SSOD) results" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 206, + 543, + 346 + ], + "blocks": [ + { + "bbox": [ + 298, + 206, + 543, + 346 + ], + "lines": [ + { + "bbox": [ + 298, + 206, + 543, + 346 + ], + "spans": [ + { + "bbox": [ + 298, + 206, + 543, + 346 + ], + "type": "image", + "image_path": "0bbfe38af1c43e77d67987ace07497bf28e25967ed2b0aee27c1d96b530b18cd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 349, + 515, + 358 + ], + "lines": [ + { + "bbox": [ + 326, + 349, + 515, + 358 + ], + "spans": [ + { + "bbox": [ + 326, + 349, + 515, + 358 + ], + "type": "text", + "content": "(d) 1Mpx semi-supervised object detection (SSOD) results" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 362, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 545, + 384 + ], + "type": "text", + "content": "Figure 6. Results on label-efficient learning using different ratios of labeled data. We compare LEOD with baselines on Gen1 and 1Mpx datasets under WSOD and SSOD using the RVT-S detector. All results are averaged over three runs." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "content": "1Mpx WSOD, we set " + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}} = 0.5" + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "content": " for pedestrians and two-wheelers to handle excessive false positives. The minimum track length " + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "inline_equation", + "content": "T_{trk}" + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "content": " is set to 6 in all experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 582 + ], + "type": "text", + "content": "Baselines. We compare with a Supervised Baseline trained only on available labels. Since we are the first work to consider this task, we design two other baselines: 1) Vanilla Self-Training: trains on pseudo labels without TTA, tracking, and soft anchor assignment; 2) Soft Teacher: adopts the online student-teacher paradigm from a representative 2D SSOD method [65]. We enable soft anchor assignment in Soft Teacher, while TTA and tracking are not applicable as the online event sequence length is too short. We tune baselines' hyper-parameters to be optimal in each setting. We also tried a state-of-the-art 2D SSOD method designed for anchor-free detectors [34] as YOLOX is anchor-free, but we did not observe clear improvements over Soft Teacher." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 587, + 176, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 176, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 176, + 599 + ], + "type": "text", + "content": "4.2. Label-Efficient Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "We first compare our method with baselines in the low-labeled data regime. The overall results are shown in Fig. 6. WSOD. Fig. 6 (a) and (b) present the weakly-supervised results. On Gen1, LEOD improves the mAP of Supervised Baseline by a large margin across all labeling ratios. Using " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " labels, we achieve an mAP of " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "45.5\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ", which is only " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " lower than RVT-S trained on " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " labels. For pseudolabeling baselines, Vanilla Self-Training outperforms Soft Teacher in most cases, validating our choice of offline la" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": "bel generation. In addition, LEOD consistently outperforms them by more than " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": ", indicating the higher quality of our pseudo labels. We observe a similar trend on 1Mpx, where our approach scores the highest mAP in all cases. Notably, the two pseudo-labeling baselines perform worse than pretrained RVT-S on " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": " labels, which proves the importance of screening reliable labels. Finally, LEOD with " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": " labels " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "(44.6\\%)" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": " outperforms RVT-S trained on all labels " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "(44.1\\%)" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": ", showing the great potential of unlabeled data. SSOD. Fig. 6 (c) and (d) present the semi-supervised results. Using the same amount of labels, models trained under SSOD are generally worse than WSOD. This indicates that given a limited budget, we should sparsely label as many event streams as possible instead of densely labeling a few sequences. Nevertheless, LEOD still outperforms baselines by more than " + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 390, + 545, + 605 + ], + "type": "text", + "content": " mAP over all labeling ratios on both datasets. Our results offer a promising direction of boosting performance with fully unlabeled event data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Qualitative Results. Fig. 7 visualizes some detection results of RVT-S trained with " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " labels under Gen1 WSOD setting. The supervised baseline can only detect objects that trigger lots of events due to its limited capacity. Models using vanilla self-training detect more objects, but also produce numerous false positives. With our pseudo-labeling pipeline, LEOD trained models are able to handle various hard examples. For example, it discovers a car that was initially missed in the ground-truth annotation in Fig. 7 (d)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16938" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 53, + 185, + 343 + ], + "blocks": [ + { + "bbox": [ + 55, + 53, + 185, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 53, + 185, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 53, + 185, + 343 + ], + "type": "image", + "image_path": "797a7c13089971173011a37c949d52dcbe76a2ae9d1bd88525e79a381a9e2e40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 346, + 135, + 358 + ], + "lines": [ + { + "bbox": [ + 121, + 346, + 135, + 358 + ], + "spans": [ + { + "bbox": [ + 121, + 346, + 135, + 358 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 53, + 303, + 344 + ], + "blocks": [ + { + "bbox": [ + 189, + 53, + 303, + 344 + ], + "lines": [ + { + "bbox": [ + 189, + 53, + 303, + 344 + ], + "spans": [ + { + "bbox": [ + 189, + 53, + 303, + 344 + ], + "type": "image", + "image_path": "d5801f9f88143a13dc3d8db8ad26400bb84412ca3ccb3319c9b5fa08df241414.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 346, + 253, + 358 + ], + "lines": [ + { + "bbox": [ + 239, + 346, + 253, + 358 + ], + "spans": [ + { + "bbox": [ + 239, + 346, + 253, + 358 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 53, + 421, + 344 + ], + "blocks": [ + { + "bbox": [ + 307, + 53, + 421, + 344 + ], + "lines": [ + { + "bbox": [ + 307, + 53, + 421, + 344 + ], + "spans": [ + { + "bbox": [ + 307, + 53, + 421, + 344 + ], + "type": "image", + "image_path": "88246e8084f0fedc201e6723d757c8c652f0167ed168d8b8a968b40eaa1d0e5e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 346, + 370, + 358 + ], + "lines": [ + { + "bbox": [ + 358, + 346, + 370, + 358 + ], + "spans": [ + { + "bbox": [ + 358, + 346, + 370, + 358 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 425, + 53, + 539, + 343 + ], + "blocks": [ + { + "bbox": [ + 425, + 53, + 539, + 343 + ], + "lines": [ + { + "bbox": [ + 425, + 53, + 539, + 343 + ], + "spans": [ + { + "bbox": [ + 425, + 53, + 539, + 343 + ], + "type": "image", + "image_path": "cc34ac774af87d9a2b4e4ddae8080dd25df1a98c017f2647a09e1ba26dc5b66b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 346, + 488, + 358 + ], + "lines": [ + { + "bbox": [ + 475, + 346, + 488, + 358 + ], + "spans": [ + { + "bbox": [ + 475, + 346, + 488, + 358 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 547, + 396 + ], + "type": "text", + "content": "Figure 7. Detection on Gen1. We show (a) a common driving scene, (b) cars with small relative movement thus few events, (c) crowded pedestrians, and (d) LEOD discovers a missing object in ground-truth annotations due to the occlusion of a barrier gate. Car and pedestrian boxes are colored in blue and orange, respectively. Pre-trained and Self-Training stand for Supervised Baseline and Vanilla Self-Training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 400, + 173, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 400, + 173, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 400, + 173, + 412 + ], + "type": "text", + "content": "4.3. Fully-Labeled Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "text", + "content": "Since the original labeling frequency on both datasets is lower than the frame rate of RVT-S (20 Hz), we can still create pseudo labels on unlabeled frames to improve the fully supervised model performance. We compare with state-of-the-art event-based object detectors trained on all labeled data in Tab. 1. LEOD improves over RVT-S by " + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "text", + "content": " on Gen1 and 1Mpx, respectively. On Gen1, our method achieves new state-of-the-art among models not using pre-trained weights. This indicates that LEOD is consistently effective even with " + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 419, + 288, + 562 + ], + "type": "text", + "content": " labels. In terms of runtime and model size, since our approach does not introduce new modules to the base model, we are as efficient as RVT-S." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 147, + 586 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "Larger Base Model. We evaluate our label-efficient learning framework on a stronger event-based detector RVT-B, which uses a larger ViT backbone compared to RVT-S. Tab. 2 presents the result in the Gen1 WSOD setting. With more parameters, RVT-B pre-trained on limited labels already outperforms RVT-S. Still, LEOD is able to improve the detection result by a sizeable margin across all labeling ratios. With " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " labels, our method achieves an mAP of " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "50.2\\%" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", which is competitive with ERGO-12 using a large-scale pre-trained Swin Transformer V2 backbone [35]. No" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 309, + 403, + 542, + 506 + ], + "blocks": [ + { + "bbox": [ + 309, + 403, + 542, + 506 + ], + "lines": [ + { + "bbox": [ + 309, + 403, + 542, + 506 + ], + "spans": [ + { + "bbox": [ + 309, + 403, + 542, + 506 + ], + "type": "table", + "html": "
MethodGen11MpxSize (M)
mAP (%)Time (ms)mAP (%)Time (ms)
RED [42]40.016.743.039.324.1
ASTMNet [29]46.735.648.372.3>100
HMNet-L3 [16]47.17.9*--33.2
RVT-B47.210.247.411.918.5
RVT-S46.59.544.110.19.9
LEOD-RVT-S48.79.546.710.19.9
ERGO-12 [76]50.477.240.6101.159.6
", + "image_path": "62b061fbba244b16805ddc276a27f3e9813e43884e3124cda3e3809ec21d5ff6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 509, + 545, + 575 + ], + "lines": [ + { + "bbox": [ + 305, + 509, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 545, + 575 + ], + "type": "text", + "content": "Table 1. Detection results using all available labels. Baseline runtimes and model sizes are obtained from [14]. For HMNet, we use the best-performing L3 variant. * Its runtime is computed using a V100 GPU, while T4 GPUs which are slower than V100 are used in the other cases. ERGO is grayed as it is the only method using pre-trained models (Swin Transformer V2 [35])." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 580, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 616 + ], + "type": "text", + "content": "tably, LEOD brings larger absolute improvements on RVT-B compared to RVT-S, proving that our framework steadily scales up to enhance larger and stronger detectors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Effect of Each Component. Fig. 8 shows the model performance with different components on Gen1 and 1Mpx. TTA significantly improves the mAP as it increases the recall in the generated pseudo labels using future information. Leveraging the temporal persistency of objects, tracking-based post-processing leads to a further gain. Finally, with soft anchor assignment, only reliable foreground and background labels are selected, easing the model training." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16939" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 48, + 293, + 185 + ], + "blocks": [ + { + "bbox": [ + 55, + 48, + 293, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 48, + 293, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 48, + 293, + 185 + ], + "type": "image", + "image_path": "b8c6e993cae07afc045992933898b1220f7a9fb103c814b6153f403ef667799d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 186, + 272, + 196 + ], + "lines": [ + { + "bbox": [ + 75, + 186, + 272, + 196 + ], + "spans": [ + { + "bbox": [ + 75, + 186, + 272, + 196 + ], + "type": "text", + "content": "(a) Gen1 weakly-supervised object detection (WSOD) results" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 48, + 536, + 185 + ], + "blocks": [ + { + "bbox": [ + 298, + 48, + 536, + 185 + ], + "lines": [ + { + "bbox": [ + 298, + 48, + 536, + 185 + ], + "spans": [ + { + "bbox": [ + 298, + 48, + 536, + 185 + ], + "type": "image", + "image_path": "f69ad8292f54da85af719da86c47ccaa1fd043b40df8b4f877ece802e626dd88.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 186, + 517, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 186, + 517, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 517, + 196 + ], + "type": "text", + "content": "(b) 1Mpx weakly-supervised object detection (WSOD) results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 199, + 544, + 221 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 544, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 544, + 221 + ], + "type": "text", + "content": "Figure 8. Ablation study of each component in LEOD. We report the WSOD result with RVT-S as the base detector on both datasets. Starting from Vanilla Self-Training, we gradually add TTA, tracking-based post-processing, and soft anchor assignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 226, + 269, + 316 + ], + "blocks": [ + { + "bbox": [ + 63, + 226, + 269, + 316 + ], + "lines": [ + { + "bbox": [ + 63, + 226, + 269, + 316 + ], + "spans": [ + { + "bbox": [ + 63, + 226, + 269, + 316 + ], + "type": "table", + "html": "
Method1%2%5%10%100%
RVT-S30.834.437.940.846.5
LEOD-RVT-S39.442.243.845.648.7
Absolute Improvement+8.6+7.8+5.9+4.8+2.2
RVT-B31.634.838.341.047.6*
LEOD-RVT-B40.042.945.346.650.2
Absolute Improvement+8.4+8.1+7.0+5.6+2.6
", + "image_path": "f533aad9f2708a6f96658d8511026f887846c138402fab015002440fae781f87.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 346, + 286, + 394 + ], + "blocks": [ + { + "bbox": [ + 47, + 319, + 287, + 341 + ], + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 341 + ], + "type": "text", + "content": "Table 2. Gen1 WSOD mAPs (%) with two RVT variants. * our reproduced RVT-B using " + }, + { + "bbox": [ + 47, + 319, + 287, + 341 + ], + "type": "inline_equation", + "content": "{100}\\%" + }, + { + "bbox": [ + 47, + 319, + 287, + 341 + ], + "type": "text", + "content": " labels result is better than [14]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 346, + 286, + 394 + ], + "lines": [ + { + "bbox": [ + 48, + 346, + 286, + 394 + ], + "spans": [ + { + "bbox": [ + 48, + 346, + 286, + 394 + ], + "type": "table", + "html": "
Rounds1%2%5%10%100%P. (1%)P. (2%)P. (5%)P. (10%)
138.141.143.145.348.50.650.690.740.79
239.442.243.845.648.70.720.750.770.81
339.542.243.645.448.60.720.740.740.76
", + "image_path": "594b3e59d61b96a5a5236258c5aeb3546b8a44e0bcf2675122ea43051bf88e98.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 396, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 441 + ], + "type": "text", + "content": "Table 3. Number of self-training rounds used in LEOD. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. We also compute the precision (P.) of pseudo labels that are used to train the models on that row, which is a good indicator for stop training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 449, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 616 + ], + "type": "text", + "content": "Self-Training Rounds. Tab. 3 left presents the results of multi-round self-training. Detectors after the first round of self-training are significantly better than pre-trained models, and thus generate higher quality pseudo labels. Therefore, a second round of training leads to consistent gains. However, due to error accumulations, a third round of training may result in worse models. To determine when to stop training, we empirically find that the precision of pseudo labels serves as a good indicator. In Tab. 3 right, we compute the precision of predicted boxes on labeled frames (skipped labels are not used to prevent information leakage). The precision consistently improves after round 1, but starts to decrease in some cases after round 2. Indeed, the mAP also drops after training on those labels with lower precision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Confidence Thresholds. We analyze the effects of hard and soft thresholds " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\tau_{\\mathrm{hard}},\\tau_{\\mathrm{soft}})" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " in Tab. 4. Prior works [59, 65] often use a high threshold of 0.9, while we observe clearly lower mAPs when " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}} > 0.7" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". With a lower " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{hard}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", we retain most of the detected objects, and suppress noisy labels with " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{soft}}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". LEOD achieves similar results using several sets of hyper-parameters, showing our robustness to them. See Appendix B for more analysis on the filtering thresholds." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 321, + 226, + 529, + 312 + ], + "blocks": [ + { + "bbox": [ + 321, + 226, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 321, + 226, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 321, + 226, + 529, + 312 + ], + "type": "table", + "html": "
(τhard, car, τsoft, car, τsoft, ped)1%2%5%10%
(0.5, 0.6, 0.3)38.941.843.545.4
(0.6, 0.7, 0.35)39.442.243.845.5
(0.6, 0.7, 0.4)39.342.043.745.6
(0.6, 0.8, 0.4)39.342.343.645.5
(0.7, 0.8, 0.4)39.042.043.244.9
(0.8, -, -)*38.441.442.744.2
", + "image_path": "9f0f571a1e4273245c5501a8df495a01de24d7d2394348421defc699373aaacc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 315, + 545, + 359 + ], + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 359 + ], + "type": "text", + "content": "Table 4. Ablation of hard and soft thresholds used in soft anchor assignment. We keep " + }, + { + "bbox": [ + 305, + 315, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\tau_{\\text{hard, ped}} = \\tau_{\\text{hard, car}} / 2" + }, + { + "bbox": [ + 305, + 315, + 545, + 359 + ], + "type": "text", + "content": ". Ped stands for pedestrian. We report the mAP (\\%) result in Gen1 WSOD using RVT-S. * indicates not using soft anchor assignment." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 365, + 378, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 378, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 378, + 378 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 386, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 470 + ], + "type": "text", + "content": "We present LEOD, the first algorithm for label-efficient event-based object detection. To leverage unlabeled data, we adopt the self-training framework with reliable label selection. Several techniques are introduced to improve labeling quality. Extensive experiments on Gen1 and 1Mpx datasets showcase the superiority of our method over baselines in all the settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 579 + ], + "type": "text", + "content": "Limitations and Future Work. Following common practice, we only conduct intra-dataset experiments, i.e., training on data gathered using the same protocol. Recent works [3, 25, 40] have shown that training large models over multiple datasets leads to excellent performance and generalization. Since LEOD benefits from unlabeled data, we can also train it jointly on more datasets that involve real-world multi-object event sequences [4, 15, 73]. We discuss some failure cases of our pseudo-labeling method in Appendix C." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 586, + 395, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 395, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 395, + 600 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "This research was partially supported by Compute Ontario (www.computeontario.ca), the Province of Ontario, the Digital Research Alliance of Canada (https://alliancecan.ca/), and the Government of Canada through CIFAR and the Natural Sciences and Engineering Research Council of Canada (NSERC). We would also like to acknowledge Vector Institute for computation support. We would like to thank Yash Kant, Lunjun Zhang, Xuanchi Ren, and Xiaoshi Wu for valuable discussions and support." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16940" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 147 + ], + "type": "text", + "content": "[2] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 150, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 286, + 195 + ], + "type": "text", + "content": "[3] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 197, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 286, + 251 + ], + "type": "text", + "content": "[4] Kenneth Chaney, Fernando Cladera, Ziyun Wang, Anthony Bisulco, M. Ani Hsieh, Christopher Korpela, Vijay Kumar, Camillo J. Taylor, and Kostas Daniilidis. M3ed: Multi-robot, multi-sensor, multi-environment event dataset. In CVPRW, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 254, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 254, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 53, + 254, + 286, + 286 + ], + "type": "text", + "content": "[5] Nicholas FY Chen. Pseudo-labels for supervised learning on dynamic vision sensor data, applied to object detection under ego-motion. In CVPRW, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 289, + 286, + 332 + ], + "type": "text", + "content": "[6] Hoonhee Cho, Hyeonseong Kim, Yujeong Chae, and Kuk-Jin Yoon. Label-free event-based object recognition via joint learning with image reconstruction from events. In ICCV, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 286, + 367 + ], + "type": "text", + "content": "[7] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IJCNN, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 370, + 286, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 286, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 286, + 414 + ], + "type": "text", + "content": "[8] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 416, + 286, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 416, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 53, + 416, + 286, + 471 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "text", + "content": "[10] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. 5.10 a " + }, + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "text", + "content": " back-illuminated stacked temporal contrast event-based vision sensor with " + }, + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "inline_equation", + "content": "4.86\\mu m" + }, + { + "bbox": [ + 48, + 473, + 286, + 561 + ], + "type": "text", + "content": " pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In 2020 IEEE International Solid-State Circuits Conference, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 563, + 286, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 563, + 286, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 563, + 286, + 608 + ], + "type": "text", + "content": "[11] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. TPAMI, 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 609, + 286, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 609, + 286, + 642 + ], + "spans": [ + { + "bbox": [ + 48, + 609, + 286, + 642 + ], + "type": "text", + "content": "[12] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 644, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 286, + 678 + ], + "type": "text", + "content": "[13] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. arXiv preprint arXiv:2211.12324, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 712 + ], + "type": "text", + "content": "[14] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. RA-L, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[16] Ryuhei Hamaguchi, Yasutaka Furukawa, Masaki Onishi, and Ken Sakurada. Hierarchical neural memory network for low latency event processing. In CVPR, 2023. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 163 + ], + "type": "text", + "content": "[17] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 1997. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 196 + ], + "type": "text", + "content": "[18] Yuhuang Hu, Tobi Delbruck, and Shih-Chii Liu. Learning to exploit multiple vision modalities by using grafted networks. In ECCV, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "text", + "content": "[19] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 253 + ], + "type": "text", + "content": "[20] Massimiliano Iacono, Stefan Weber, Arren Glover, and Chiara Bartolozzi. Towards event-driven object detection with off-the-shelf deep learning. In IROS, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "type": "text", + "content": "[21] Mingcheng Ji, Ziling Wang, Rui Yan, Qingjie Liu, Shu Xu, and Huajin Tang. Sctn: Event-based object tracking with energy-efficient deep convolutional spiking neural networks. Frontiers in Neuroscience, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 331 + ], + "type": "text", + "content": "[22] Zhuangyi Jiang, Pengfei Xia, Kai Huang, Walter Stechele, Guang Chen, Zhenshan Bing, and Alois Knoll. Mixed frame-event-driven fast pedestrian detection. In ICRA, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 365 + ], + "type": "text", + "content": "[23] Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 367, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 367, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 367, + 545, + 398 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 443 + ], + "type": "text", + "content": "[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 487 + ], + "type": "text", + "content": "[26] Simon Klenk, David Bonello, Lukas Koestler, and Daniel Cremers. Masked event modeling: Self-supervised pretraining for event cameras. arXiv preprint arXiv:2212.10368, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[27] Dong-Hyun Lee et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In ICMLW, 2013. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[28] Hengduo Li, Zuxuan Wu, Chen Zhu, Caiming Xiong, Richard Socher, and Larry S Davis. Learning from noisy anchors for one-stage object detection. In CVPR, 2020. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[29] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. TIP, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 602, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 633 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[31] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[32] Minghua Liu, Yin Zhou, Charles R Qi, Boqing Gong, Hao Su, and Dragomir Anguelov. Less: Label-efficient semantic segmentation for lidar point clouds. In ECCV, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16941" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In ICLR, 2020. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 151 + ], + "type": "text", + "content": "[34] Yen-Cheng Liu, Chih-Yao Ma, and Zsolt Kira. Unbiased teacher v2: Semi-supervised object detection for anchor-free and anchor-based detectors. In CVPR, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "text", + "content": "[35] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 231 + ], + "type": "text", + "content": "[36] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In ECCV, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 287, + 276 + ], + "type": "text", + "content": "[37] Nico Messikommer, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. Bridging the gap between events and frames through unsupervised domain adaptation. RA-L, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 311 + ], + "type": "text", + "content": "[38] Ishan Misra, Abhinav Shrivastava, and Martial Hebert. Watch and learn: Semi-supervised learning for object detectors from video. In CVPR, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 357 + ], + "type": "text", + "content": "[39] Elias Mueggler, Henri Rebecq, Guillermo Gallego, Tobi Delbruck, and Davide Scaramuzza. The event-camera dataset and simulator: Event-based data for pose estimation, visual odometry, and slam. *IJRR*, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 359, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 287, + 414 + ], + "type": "text", + "content": "[40] Abhishek Padalkar, Acorn Pooley, Ajinkya Jain, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anikait Singh, Anthony Brohan, et al. Open x-embodiment: Robotic learning datasets and rt-x models. arXiv preprint arXiv:2310.08864, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "text", + "content": "[41] George Papandreou, Liang-Chieh Chen, Kevin P Murphy, and Alan L Yuille. Weakly-and semi-supervised learning of a deep convolutional network for semantic image segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1742-1750, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 472, + 287, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 287, + 505 + ], + "type": "text", + "content": "[42] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 2020. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 507, + 287, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 507, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 48, + 507, + 287, + 550 + ], + "type": "text", + "content": "[43] Christoph Posch, Daniel Matolin, and Rainer Wohlgannnt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 2010. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 585 + ], + "type": "text", + "content": "[44] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 587, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 286, + 609 + ], + "type": "text", + "content": "[45] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 287, + 643 + ], + "type": "text", + "content": "[46] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[47] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. TPAMI, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. NeurIPS, 2015. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[49] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 148 + ], + "type": "text", + "content": "[50] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In WACV, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 150, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 193 + ], + "type": "text", + "content": "[51] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[52] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. NeurIPS, 2015. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[53] Krishna Kumar Singh, Fanyi Xiao, and Yong Jae Lee. Track and transfer: Watching videos to simulate strong human supervision for weakly-supervised object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3548-3556, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 294, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 346 + ], + "type": "text", + "content": "[54] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[55] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. arXiv preprint arXiv:2005.04757, 2020. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 435 + ], + "type": "text", + "content": "[56] Timo Stoffregen, Cedric Scheerlinck, Davide Scaramuzza, Tom Drummond, Nick Barnes, Lindsay Kleeman, and Robert Mahony. Reducing the sim-to-real gap for event cameras. In ECCV, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[57] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 502 + ], + "type": "text", + "content": "[58] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 536 + ], + "type": "text", + "content": "[59] He Wang, Yezhen Cong, Or Litany, Yue Gao, and Leonidas J Guibas. 3dioumatch: Leveraging iou prediction for semi-supervised 3d object detection. In CVPR, 2021. 4, 5, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 537, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 579 + ], + "type": "text", + "content": "[60] Lin Wang, Yujeong Chae, Sung-Hoon Yoon, Tae-Kyun Kim, and Kuk-Jin Yoon. Evdistill: Asynchronous events to end-task learning via bidirectional reconstruction-guided cross-modal knowledge distillation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 613 + ], + "type": "text", + "content": "[61] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In CVPR, 2023. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "type": "text", + "content": "[62] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[63] Ziyi Wu, Xudong Liu, and Igor Gilitschenski. Eventclip: Adapting clip for event-based object recognition. arXiv preprint arXiv:2306.06354, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[64] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16942" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 544 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[65] Mengde Xu, Zheng Zhang, Han Hu, Jianfeng Wang, Lijuan Wang, Fangyun Wei, Xiang Bai, and Zicheng Liu. End-to-end semi-supervised object detection with soft teacher. In ICCV, 2021. 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[66] Pengxiang Yan, Guanbin Li, Yuan Xie, Zhen Li, Chuan Wang, Tianshui Chen, and Liang Lin. Semi-supervised video salient object detection using pseudo-labels. In ICCV, 2019. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 185 + ], + "type": "text", + "content": "[67] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pretraining. In ICCV, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 228 + ], + "type": "text", + "content": "[68] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 287, + 263 + ], + "type": "text", + "content": "[69] Alessandro Zanardi, Andreas Aumiller, Julian Zilly, Andrea Censi, and Emilio Frazzoli. Cross-modal learning filters for rgb-neuromorphic wormhole learning. RSS, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 287, + 297 + ], + "type": "text", + "content": "[70] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 341 + ], + "type": "text", + "content": "[71] Lunjun Zhang, Anqi Joyce Yang, Yuwen Xiong, Sergio Casas, Bin Yang, Mengye Ren, and Raquel Urtasun. Towards unsupervised object detection from lidar point clouds. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 287, + 385 + ], + "type": "text", + "content": "[72] Yongqiang Zhang, Yancheng Bai, Mingli Ding, Yongqiang Li, and Bernard Ghanem. W2f: A weakly-supervised to fully-supervised framework for object detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 287, + 431 + ], + "type": "text", + "content": "[73] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. RA-L, 2018. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 434, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 465 + ], + "type": "text", + "content": "[74] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 498 + ], + "type": "text", + "content": "[75] Alex Zihao Zhu, Ziyun Wang, Kaung Khant, and Kostas Daniilidis. Eventgan: Leveraging large scale image datasets for event cameras. In ICCP, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "text", + "content": "[76] Nikola Zubić, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In ICCV, 2023. 1, 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16943" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_content_list.json b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac44a4b793ac9b3696da1b4a6dd882aee8a5d36f --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_content_list.json @@ -0,0 +1,1456 @@ +[ + { + "type": "text", + "text": "LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge", + "text_level": 1, + "bbox": [ + 196, + 130, + 774, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gongwei Chen, Leyang Shen, Rui Shao, $\\dagger$ Xiang Deng, Liqiang Nie $\\dagger$ \nSchool of Computer Science and Technology, Harbin Institute of Technology, Shenzhen \n{chengongwei, shaorui, dengxiang, nieliqiang}@hit.edu.cn \nhttps://github.com/rshaojimmy/JiuTian", + "bbox": [ + 137, + 202, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vVisual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between image-level and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imperfect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multi-modal benchmarks demonstrate the superiority of our model (e.g., improvement of $5\\%$ accuracy on VSR and $3\\%$ CIDEr on TextCaps over InstructBLIP, $5\\%$ accuracy on RefCOCOg over Kosmos-2).", + "bbox": [ + 75, + 340, + 472, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 786, + 209, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Large Language Models (LLMs) have demonstrated remarkable zero-shot abilities on various linguistic tasks. Assisted by LLMs, several multimodal large language models (MLLMs), such as MiniGPT-4 [54], Otter", + "bbox": [ + 75, + 811, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0197e60522e17f12efd32fb03d4677193d011df1a7b4135044b08cd008e6bda8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 309, + 890, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/172da6b0638d7c17d4b19ff0e6adb63651d6035da34cc85e3cb5489aae5a1d71.jpg", + "image_caption": [ + "Figure 1. Comparison between existing MLLMs and LION. The existing MLLM generates a vague and inaccurate response, while LION provides a more precise and contextually accurate description by progressively incorporating spatial-aware knowledge and softly prompting semantic visual evidence." + ], + "image_footnote": [], + "bbox": [ + 501, + 404, + 890, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[19], and InstructBLIP [7], achieve significant improvements in reasoning abilities to deal with various vision-language (VL) tasks.", + "bbox": [ + 498, + 597, + 890, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In most of the existing MLLMs, the visual information is mainly extracted from a vision encoder pretrained with image-level supervision (e.g., CLIP [32]), and then are adapted to a LLM by using a tiny bridge module. This makes these MLLMs inherently possess limited image understanding capabilities [21]. As shown in Fig. 1, the insufficient visual information misleads MLLMs to provide erroneous and hallucinated responses. An intuitive solution to this problem is to replace or tune the vision encoder [41]. However, it requires pretraining on massive data or suffers from the catastrophic forgetting issue [35, 36, 50], which diminishes the practical efficacy of this strategy. These predicaments highlight that the insufficient extraction of visual knowledge has become a central obstacle impeding the development of MLLMs.", + "bbox": [ + 496, + 643, + 892, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome this dilemma, as depicted in Fig. 1, we devise a dual-Level vVisual knOwledge eNhanced Multimodal", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 96, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26540", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Model (LION), which enriches the visual information in MLLMs in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. LION enhances the MLLM with more fine-grained perceptual abilities by studying the region-level VL tasks involving the spatial coordinates. However, we find that simply training on region-level and the original image-level VL tasks* simultaneously hurts the general performances of the MLLM due to the conflicts between these two kinds of tasks. To address this issue, we propose a novel stage-wise instruction-tuning strategy to perform image-level and region-level VL tasks separately with two different visual branches and task adapters. In addition, we devise mixture-of-adapters with a router to dynamically fuse visual information across various granularities in a unified MLLM. This progressive incorporation of fine-grained visual knowledge contributes to the mutual promotion between these two kinds of VL tasks, and spawns LION to excel in capturing fine-grained visual information and performing spatial reasoning, as shown in Fig. 1. 2) Soft prompting of high-level semantic visual evidence. Alongside the improvement of MLLMs in fine-grained perceptual capabilities, there is also an opportunity to enhance their high-level semantic understanding. LION uses an off-the-shelf vision model to extract high-level semantic knowledge, i.e., image tags, as supplementary information for the MLLM. However, as off-the-shelf vision models are typically not flawless, errors in tag predictions are inevitable. Inspired by prompt tuning, we propose a soft prompting method to mitigate the potential negative influence resulting from the imperfect predicted tags. As shown in Fig. 1, injection of semantic visual evidence alleviates the hallucination issue substantially.", + "bbox": [ + 75, + 90, + 472, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 96, + 590, + 434, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To address the internal conflict between region-level and image-level VL tasks, we propose a progressive incorporation of fine-grained spatial-aware visual knowledge with a novel stage-wise instruction-tuning strategy. It achieves mutual promotion between two kinds of VL tasks and equips LION with advanced holistic and fine-grained visual perceptual abilities.", + "- As a powerful complement, we propose to integrate image tags as high-level semantic visual evidence into MLLMs, and design a soft prompting method to alleviate the bad influence from incorrect tags. This mitigates the hallucination issue and showcases positive effects on various VL tasks.", + "- We evaluate LION on a wide range of VL tasks, including image captioning, visual question answering (VQA), and visual grounding, and demonstrate its superiority over the baselines. LION outperforms InstructBLIP by around $5\\%$" + ], + "bbox": [ + 76, + 606, + 468, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "accuracy on VSR, and around $3\\%$ CIDEr on TextCaps, Kosmos-2 by around $5\\%$ accuracy on RefCOCOg. The evaluations on POPE and MMBench exhibit the remarkable abilities of LION in alleviating object hallucination and various perceptual dimensions.", + "bbox": [ + 511, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 196, + 640, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 500, + 227, + 816, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the success of LLMs, many researches have emerged to extend them to multimodal tasks, especially VL tasks. The common pipeline uses a vision model to transform the image into visual features, followed by a bridge module to align them with the feature space of LLMs. Some works [4, 11, 26, 31, 51] directly use a linear or MLP layer as bridge module, while others [1, 7, 19, 20, 46, 54] design more complicated bridge networks to compress or adaptively select visual information. Despite their impressive performance on VL tasks, there is still a lack of exploration on the effectiveness and limitation of the visual branch in a MLLM. Recently, Wang et al. [41] empirically investigate factors contributing to the formation of an effective vision encoder in a MLLM from the perspective of pretraining. Differently, our work explores the effect of region-level VL tasks on the visual understanding abilities of the MLLM, and incorporates fine-grained and high-level visual knowledge to enrich the visual branch in the MLLM.", + "bbox": [ + 496, + 256, + 890, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Visual Grounding in the field of MLLMs", + "text_level": 1, + "bbox": [ + 500, + 554, + 846, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual grounding [16, 37, 38] is a region-level VL task that aims to establish a connection between particular regions and their textual descriptors, which plays a vital role in human-machine interaction by enabling referential dialog. In the realm of MLLMs, there are some attempts to enhance MLLMs by leveraging visual grounding tasks. Works like Shikra [4], Kosmos-2 [31], Ferret [47] and Pink [45] demonstrate the promising direction of employing visual grounding datasets to endow MLLMs with region-level visual understanding abilities. They convert existing datasets equipped with spatial coordinates, like Visual Genome [17] and RefCOCO [16], into the textual instruction format and perform instruction tuning on MLLMs. Merely considering the visual grounding task as one of several instruction-tuning tasks, these works fall short in exploring the interactions among various tasks. In contrast, our work investigates the internal conflict between visual grounding tasks and image-level VL tasks (e.g., image captioning and VQA), and proposes a stage-wise instruction-tuning strategy to address this issue, achieving a good balance between these two kinds of VL tasks.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*Here, image-level VL tasks denote image captioning and visual question answering, region-level VL tasks mean visual grounding tasks.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f15f4f211ae4f508976827313469a9b06d82c85dab24734b654a7e1e6bd7722.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed LION. The model extracts holistic visual features from Q-Former, and combines them with fine-grained spatial-aware visual features from the vision aggregator. The Mixture-of-Adapters with a router in the frozen LLM dynamically fuses visual knowledge learned from different visual branches and LLM adapters based on the task types (image-level and region-level)." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LION", + "text_level": 1, + "bbox": [ + 76, + 339, + 151, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the dual-Level vVisual knOwlEdge eNhanced multimodal large language model (LION). The whole framework is depicted in Fig. 2.", + "bbox": [ + 75, + 366, + 470, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Progressive Incorporation of Fine-grained Spatial-Aware Visual Knowledge", + "text_level": 1, + "bbox": [ + 76, + 424, + 468, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Reorganizing Visual Grounding Tasks", + "text_level": 1, + "bbox": [ + 76, + 464, + 397, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To incorporate fine-grained spatial-aware visual knowledge into MLLMs, we make use of region-level VL tasks, i.e., visual grounding, and meticulously process the data with spatial coordinates in a unified format for instruction-tuning MLLM. Visual grounding requires the model to generate or comprehend natural language expressions referring to particular objects or regions within an image, e.g., \"a man with glasses\". Referring to objects or regions in complex images needs an ability of precisely comprehending fine-grained visual information. Current MLLMs lack such referring comprehending, as they mainly target a coarse alignment of VL modalities when pretrained on massive image-text pairs [2, 41]. In this regard, we introduce visual grounding tasks as a kind of region-level VL tasks for the instruction-tuning of MLLMs. This aims to endow the model with fine-grained visual understanding ability such that better performance on image-level VL tasks (e.g., image captioning and VQA) might be achieved.", + "bbox": [ + 75, + 489, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt two types of visual grounding tasks, including referring expression comprehension (REC) and referring expression generation (REG) [49]. We use the Visual Genome dataset [18], which associates a local area with one short description, to construct REC/REG tasks. The templates used to organize the Visual Genome dataset in a unified instruction-tuning format can be found in Appendix.", + "bbox": [ + 75, + 763, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One core point in reorganizing visual grounding tasks is the way of processing positions. Normally, the position of", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f9815cbd12f480ae7979f32f606b23f56df641c302a2219d716a987a1974bb8.jpg", + "image_caption": [ + "Stage 1: Instruction-Tuning on Image-Level VL Tasks" + ], + "image_footnote": [], + "bbox": [ + 501, + 352, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/89e9b079c5ec333cc26e375d96290271742a3f4d0386436a3f53d50fbec52484.jpg", + "image_caption": [ + "Stage 2: Instruction-Tuning on Region-Level VL Tasks" + ], + "image_footnote": [], + "bbox": [ + 501, + 393, + 890, + 440 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/79b5cd8265c60c8c592fb8ed74c218c9e43c3fc2cadf1fbaa47747cf21d504e1.jpg", + "image_caption": [ + "Stage 3: Instruction-Tuning on Image-Level and Region-Level VL Tasks", + "Figure 3. The stage-wise instruction-tuning strategy. **Stage 1:** We instruction-tune Q-Former and the image-level adapter on image-level VL tasks. **Stage 2:** We instruction-tune the vision aggregator (VA), MLP, and the region-level adapter on region-level VL tasks. **Stage 3:** The Mixture-of-Adapters is devised to form a unified model for instruction-tuning on both kinds of VL tasks." + ], + "image_footnote": [], + "bbox": [ + 501, + 455, + 890, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "an object phrase is presented in the format of bounding box $[x_{min},y_{min},x_{max},y_{max}]$ . We use a natural language style to describe object positions along with the square brackets. A sample in the REC task is displayed as follows: \"How can I locate a glass of beer in the image? Please provide the coordinates. Answer: [0.525, 0.0, 0.675, 0.394]\".", + "bbox": [ + 496, + 598, + 890, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2 The Stage-Wise Instruction-tuning Strategy", + "text_level": 1, + "bbox": [ + 498, + 709, + 857, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate MLLMs with fine-grained spatial-aware knowledge, the most intuitive way is to directly instruction-tune MLLMs with both image-level and region-level VL tasks in one stage. However, this single-stage instruction-tuning strategy is sub-optimal, and suffers from the internal conflict between these two kinds of VL tasks. We summarize two main issues leading to the internal conflict. 1) One is the need of region-level modality-alignment pretraining. In concurrent works that integrate visual grounding ability, pretraining on the region-level multimodal datasets including visual grounding is a crucial step. Some works", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26542", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[31, 45, 47] elaborately create very large visual grounding datasets (e.g., GRIT-20M in kosmos-2 [31]) to advance MLLM in fine-grained perception and understanding. The single-stage instruction-tuning makes it challenging to adapt visual representations learned for image-level alignment to region-level VL tasks under a limited training configuration. 2) Another is the gap between the input-output modes of image-level VL tasks and region-level visual grounding tasks. The latter additionally requires MLLMs to understand specific phrases (in the format “ $[x_{min}, y_{min}, x_{max}, y_{max}]$ ” about the positions of objects. They are semantically distinct from natural languages used in image-level tasks. This requirement necessitates the tuning of the LLM to adapt to region-level tasks, but may disrupt the internal state of the LLM suitable for image-level VL tasks. To address the above issues, we devise a stage-wise instruction-tuning strategy and mixture-of-adapters with a router.", + "bbox": [ + 75, + 90, + 472, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The stage-wise instruction-tuning strategy is proposed to alleviate the internal conflict between image-level and region-level VL tasks during instruction-tuning. It is composed of three stages for instruction-tuning on image-level, region-level VL tasks and both, respectively, which is depicted in Fig. 3. In stage 1, we follow instructBLIP [7] and fine-tune Q-Former and the image-level adapter [5] in the LLM on image-level VL tasks, such as image captioning and VQA. In stage 2, we propose a vision aggregator for better capturing visual features in fine-grained understanding, which will be introduced later, and tune it with MLP and the region-level adapter on region-level VL tasks. The independent training in the first two stages greatly fulfills the requirements of sufficiently learning both image-level and region-level tasks, providing a solid foundation for subsequent joint training.", + "bbox": [ + 75, + 363, + 472, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mixture-of-Adapters with a Router. In stage 3 of our stage-wise instruction-tuning, we need a unified model but encounter a situation where adapters of LLM in stages 1 and 2 are different and suit distinct input-output modes. Inspired by Mixture-of-Experts, we treat each adapter as an expert, and propose a router module to avoid the potential interference between them, as depicted in Fig. 2.", + "bbox": [ + 75, + 625, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An adapter [5] is inserted at each FFN layer in a parallel manner. Assuming $X \\in \\mathcal{R}^{L \\times D}$ is the hidden representations generated by a self-attention / causal attention layer, the output representations after FFN (represented as $\\mathbf{F}$ ) with the adapter (denoted by $\\mathbf{H}$ ) layer are formulated as,", + "bbox": [ + 75, + 731, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO = \\mathbf {F} (X) + \\mathbf {H} (X), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 821, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} (X) = W _ {u} \\left(\\sigma \\left(W _ {d} X\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 861, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma$ is a non-linear function, ReLU. Our router module", + "bbox": [ + 76, + 885, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ac934367e3a1fe75439f6fc5bfb232e64fe29629a0cbe6d907cff9b8ba2bd45.jpg", + "image_caption": [ + "Figure 4. Instruction template with soft prompt. We use a well-designed instruction template with trainable soft prompts to inject the image tags generated by the RAM model into LION." + ], + "image_footnote": [], + "bbox": [ + 509, + 87, + 866, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aims to dynamically aggregate the hidden features from the main branches and the multiple adapter branches according to task types. Given a set of adapters $\\{\\mathbf{H}_1,\\dots ,\\mathbf{H}_K\\}$ , each kind of task $t$ defines a specific router function $\\mathbf{R}^t$ to generate new hidden features, which can be formulated as,", + "bbox": [ + 498, + 253, + 892, + 328 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO ^ {t} = \\mathbf {F} (X) + \\sum_ {k = 1} ^ {K} \\mathrm {G} _ {k} ^ {t} \\odot \\mathbf {H} _ {k} (X). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 338, + 890, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_k^t\\in \\mathcal{R}^D$ is a trainable vector that modulates the hidden features from each adapter and makes them suitable for the target task. In practice, we define two types of tasks, one for image-level VL tasks (image captioning and VQA), the other for fine-grained VL tasks (visual grounding). Compared to directly incorporating multiple adapters, the router module provides a better way to maximize the complementarity of image-level and region-level tasks.", + "bbox": [ + 496, + 388, + 892, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the standard language modeling loss in all instruction-tuning stages. In the experiments, we demonstrate that stage-wise training is superior to single-stage training, and ensures a good balance between high-level and fine-grained visual understanding capabilities, further achieves a significant mutual promotion between image-level and region-level VL tasks.", + "bbox": [ + 496, + 511, + 893, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Vision Aggregator", + "text_level": 1, + "bbox": [ + 500, + 635, + 681, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extract more sufficient visual details from input images, we devise a vision aggregator that integrates multi-level hidden features of the pretrained visual encoder. Although the vision encoder has a global reception field in all layers, it is verified that different transformer layers learn visual information at different scales [9], e.g., lower layers learn visual details. Thus, our vision aggregator makes fine-grained spatial-aware visual knowledge more likely to be learned based on visual grounding tasks. Specifically, our vision aggregator can be regarded as a tiny transformer-style network, consisting of two transformer layers for aggregating the hidden features from the vision encoder. Given the hidden features $\\{V_{i}, V_{j}, V_{k}\\}$ from some middle layers in the vision encoder, the vision aggregation module uses two blocks to sequentially integrate the former two features with the last feature. Each block $\\mathcal{B}$ is composed of self attention", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3f8cf83c83ac3cb9324c28bc732caca42c2e030ae6a4144c9d0057486dc06de1.jpg", + "table_caption": [ + "Table 1. Comparison on image captioning and VQA. “†” denotes including in-house data that are publicly inaccessible. “*” means our evaluated results by using publicly released checkpoints, which are only for reference as official evaluation settings are incomplete. We report CIDEr score for Flickr30K, COOCap, and TextCaps, Mean Reciprocal Rank (MRR) for Visual Dialog (VisDial), and top-1 accuracy for others. The best and second performances for each benchmark are indicated in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
ModelFlickr30KCOCOCapTextCapsOKVQAAOKVQAGQAIconQAVSRVisDial
Flamingo-3B [1]60.6073.00------46.10
Flamingo-9B [1]61.5079.40-44.70----48.00
Kosmos-1 [13]67.1084.70-------
Kosmos-2 [31]80.50--------
AdapterV2 [11]-122.20-------
Shikra [4]73.90117.50-47.16-----
Pink [45]---59.50-52.6047.8066.30-
MiniGPT4 [54]17.75*17.04*24.06*37.5034.51*30.8037.6041.6016.52*
LLaVA [26]48.03*73.85*45.54*54.4034.51*41.3043.0051.208.65*
MiniGPTV2 [3]80.75*129.16*80.60*56.90-60.3047.7060.608.47*
BLIVA [12]87.10-----44.8862.2045.63
InstructBLIP† (T5XL) [7]84.50138.2182.5549.2857.8648.4050.0064.8046.60
InstructBLIP† (T5XXL) [7]83.50138.2882.5348.5956.1647.9051.2065.6048.50
InstructBLIP (T5XL) [7]83.71135.47104.1747.3856.1246.3452.4769.9348.75
InstructBLIP (T5XXL) [7]85.79138.63105.4453.0259.3847.7453.1868.4650.41
LION-4B85.57138.20104.8751.0859.9849.5054.9172.9650.02
LION-12B87.12139.25108.7657.3360.8751.5654.8973.7750.42
", + "bbox": [ + 106, + 146, + 857, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(Attn), cross attention (XAttn), and Feed-forward network (FFN) arranged in a sequential manner. Finally, the output features $\\bar{V}$ is generated as follows,", + "bbox": [ + 75, + 382, + 470, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {V} = \\mathcal {B} _ {2} \\left(\\mathcal {B} _ {1} \\left(V _ {i}; V _ {j}\\right); V _ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 439, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} (X; Y) = \\operatorname {F F N} (\\operatorname {X A t t n} (\\operatorname {A t t n} (X), Y)). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 482, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we use the middle layers $\\{i = L - 1, j = 2L/3, k = L/3\\}$ in the vision encoder to produce the hidden features as the input to VA, where $L$ is the number of layers in the vision encoder.", + "bbox": [ + 75, + 505, + 470, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Soft Prompting of High-Level Semantic Visual Evidence", + "text_level": 1, + "bbox": [ + 76, + 574, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The vision encoder in a MLLM may be insufficient in comprehensively extracting visual information required by complex multi-modal tasks, although it has been trained on large-scale image-text pairs. It has been demonstrated that increasing the amount and quality of pretraining multimodal datasets can significantly improve the visual understanding capability of MLLMs [41], but inevitably induces prohibitive computational overhead. An appealing alternative is to harness the convenient and powerful off-the-shelf vision models to capture various aspects of visual content within an image as a supplement.", + "bbox": [ + 75, + 613, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We choose the recognize anything model (RAM) [52] as an off-the-shelf vision model to provide diverse tags, encompassing objects, scenes, actions, and attributes, as visual evidence to support comprehensive visual perception. Instead of directly adding tags in the instruction, we design a soft prompting method to guide the model to adaptively use the inserted tags in order to avoid the potential negative influence caused by the imperfect predictions from RAM.", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 4, we present the instruction template of tags along with the soft prompt that is a trainable vector. Our soft prompting approach can be regarded as a kind of prompt tuning methods, which guides the model toward the right direction. In standard prompt tuning works, the right direction is directly formulated as the optimization for task goals. In our work, the right direction is specified by a tailored sentence, \"According to , you are allowed to use or partially use the following tags: \", and \"\" will be replaced by the soft prompt. Our soft prompting method for inserting tags has some distinct properties. It is designed to adaptively select valuable information from tags, rather than serving a specific task, as seen in standard prompt tuning methods. Our method directly uses the output labels from a small off-the-shelf vision model to incorporate high-level semantic visual evidence into a MLLM, so as to eliminate extra computational overhead of the feature alignment.", + "bbox": [ + 496, + 382, + 893, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 656, + 633, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our model along with the quantitative and qualitative analyses. Please refer to Appendix for implementation details and training details.", + "bbox": [ + 496, + 681, + 893, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluations on Image-Level VL Tasks", + "text_level": 1, + "bbox": [ + 498, + 755, + 826, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, we evaluate multi-modal understanding abilities of LION on two kinds of image-level VL tasks, image captioning and VQA. Image captioning requires the model to generate a text description of the input image. We use COCO caption [6], TextCaps [39] and Flickr30K [48] as benchmarks, and report CIDEr as the evaluation metric. We utilize greedy search for caption generation. VQA provides an image along with a specific question for the model, asking", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26544", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0d640217b63431cb0d93ff8a6c53ab4a863e1f1367f28433c5a33c82ce72d82c.jpg", + "table_caption": [ + "Table 2. Comparison on REC. \"Avg.\" means the average of top-1 accuracy over all the 8 evaluation sets." + ], + "table_footnote": [], + "table_body": "
ModelRefCOCORefCOCO+RefCOCOgAvg.
valtest-Atest-Bvaltest-Atest-Bvaltest
Zero-shot Setting
Kosmos-2 [31]52.3257.4247.2645.4850.7342.2460.5761.6552.21
GRILL [15]-------47.50-
Pink [45]54.1061.2044.2043.9050.7035.0059.1060.1051.00
LION-4B57.8956.0758.4046.3845.2947.5064.7463.5654.98
LION-12B58.5456.4159.3645.9345.7347.8966.1264.6955.58
Fune-tuning Setting
OFA-L [42]79.9683.6776.3968.2976.0061.7567.5767.5872.65
VisionLLM-H [43]-86.70-------
Shikra-7B [4]87.0190.6180.2481.6087.3672.1282.2782.1982.93
Shikra-13B [4]87.8391.1181.8182.8987.7974.4182.6483.1683.96
Pink [45]88.3091.7084.0081.4087.5073.7083.7083.7084.25
Ferret-7B [47]87.4991.3582.4580.7887.3873.1483.9384.7683.91
Ferret-13B [47]89.4892.4184.3682.8188.1475.1785.8386.3485.57
MiniGPTv2 [3]88.6991.6585.3379.9785.1274.4584.4484.6684.29
LION-4B89.7392.2984.8283.6088.7277.3485.6985.6385.98
LION-12B89.8093.0285.5783.9589.2278.0685.5285.7486.36
", + "bbox": [ + 192, + 104, + 769, + 343 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3d68b7c483be1038e77e380e39ba0aa7f25fcf2a4a5019c7a29dbb61e5372aa3.jpg", + "table_caption": [ + "Table 3. The comparison of various strategies in the instruction-tuning period. \"REC Avg.\" represents the average score of all REC tasks. \"Held-in\" denotes the average score of COOCap, TextCaps, OKVQA, and AOKVQA, while \"Held-out\" means the average score of Flickr30K, GQA, IconQA, VSR, and VisDial." + ], + "table_footnote": [], + "table_body": "
StrategyImage-LevelRegion-Level
Held-inHeld-outREC Avg.
Single Stage84.7960.203.78
Stage-wise88.0761.9154.46
w/ Router87.6762.1654.98
", + "bbox": [ + 117, + 424, + 423, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the output as an answer. We evaluate LION on six VQA datasets, including OKVQA [30], AOKVQA [34], GQA [14], IconQA [28], Visual Spatial Reasoning [24], and Visual Dialog (VisDial) [8]. For OKVQA, A-OKVQA, and GQA, we employ an open-ended generation with a greedy decoding strategy, For IconQA, Visual Spatial Reasoning, and Visual Dialog, we match the output with various candidates, and select the candidate with the highest value as the prediction. We report Mean Reciprocal Rank (MRR) for Visual Dialog, and top-1 accuracy for other VQA tasks. The detailed descriptions of these datasets and inference instructions are presented in Appendix.", + "bbox": [ + 75, + 505, + 468, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, LION achieves the best performance across 7 out of 9 benchmarks, and the second on OKVQA. LION shares the same training datasets with InstructBLIP, except Visual Genome dataset adopted in our work and the in-house dataset, WebCapFilt, used in InstructBLIP. The amount of Visual Genome dataset (3.6M) is far smaller than WebCapFilt (14M). Compared to the original InstructBLIP trained with WebCapFilt, LION exhibits superior performances on all zero-shot evaluation benchmarks, showcasing a better generalization ability. We also re-implemented InstructBLIP on the same instruction-tuning datasets. The comparison shows the consistent and significant improvements of LION over the re-implemented InstructBLIP, demonstrating the effectiveness of incorpo", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rating dual-level visual knowledge. Shikra, MiniGPTV2 and Pink are also integrated with visual grounding abilities. Their inferior results on most image-level VL tasks exhibit that our proposed stage-wise instruction-tuning strategy and soft prompting of high-level semantic knowledge help enhance holistic visual understanding abilities of MLLMs.", + "bbox": [ + 496, + 354, + 890, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluations on Region-Level VL Tasks", + "text_level": 1, + "bbox": [ + 498, + 454, + 833, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the fine-grained perceptual and reasoning abilities of LION, we evaluate it on three REC datasets, RefCOCO [16], RefCOCO+ [16], RefCOCOg [29]. REC requires the model to locate the target object given a referring expression. We follow the standard setting, and use accuracy as an evaluation metric, which means it is correct when the IOU between prediction and ground-truth is no less than 0.5.", + "bbox": [ + 496, + 477, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we show the comparison between LION and other MLLMs with respect to the grounding abilities, under the settings of zero-shot and fine-tuning evaluations, respectively. In the zero-shot evaluation setting, we directly employ LION to generate coordinates of referring expressions on three datasets. Our model shows significant improvements on most evaluation sets over Kosmos-2 and Pink, except test-A sets of RefCOCO and RefCOCO+. The languages used in RefCOCOg are more flowery than those used in RefCOCO and RefCOCO+. The significant improvements on RefCOCOg clearly demonstrate that LION can handle complex referring expressions and has superior zero-shot spatial-aware visual understanding abilities.", + "bbox": [ + 496, + 584, + 890, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the fine-tuning setting, we fine-tune LION with training samples from three REC datasets. As shown in Table 2, our model achieves the best performance on average and on most of the evaluation sets, indicating the advanced fine-grained perception ability of our model. Ferret proposes a spatial-aware visual sampler to handle free-form referred regions, and meticulously constructs an extensive grounding dataset with lots of efforts on data generation and fil", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26545", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d498e9f6c354e02df43165f95f6fda8084cd498c8e2986f9fb6b3b6a38a4e9c5.jpg", + "table_caption": [ + "Table 4. Ablation studies of dual-level visual knowledge. \"VG\" means visual grounding tasks. \"Held-in\" and \"Held-out\" denote the training images of tasks are seen and unseen, respectively. \"REC Avg.\" means the average score of all REC tasks." + ], + "table_footnote": [], + "table_body": "
ComponentsHeld-inHeld-outREC Avg.
VGTagsCOCOCapTextCapsOKVQAAOKVQAFlickr30KGQAVSRIconQAVisDial
××135.47104.1747.3856.1283.7146.3469.9352.4748.75-
✓×137.87104.8451.0756.9083.9949.2273.2054.6749.7054.98
✓✓138.20104.8751.0859.9885.5749.5072.9654.9150.0254.92
", + "bbox": [ + 109, + 119, + 854, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2a1cc45ba7c7edb00d9b8d7981175d41c8d601a3f3c7e6499faa48224b8aa77b.jpg", + "table_caption": [ + "Table 5. Evaluation of object hallucination on POPE benchmark. F1 score is the major metric for hallucication evaluation." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsLIONShikra [4]InstructBLIP [7]MiniGPT-4 [54]LLaVA [26]mPLUG-Owl [46]
RandomF1-Score ↑88.3386.1989.2780.1766.6468.39
Accuracy ↑88.9786.9088.5779.6750.3753.97
Precision ↑97.1294.4084.0978.2450.1952.07
Recall ↑81.0079.2795.1382.2099.1399.60
PopularF1-Score ↑85.9483.1684.6673.0266.4466.94
Accuracy ↑86.7783.9782.7769.7349.8750.90
Precision ↑91.6987.5576.2765.8649.9350.46
Recall ↑80.8779.2095.1381.9399.2799.40
AdversarialF1-score ↑84.7182.4977.3270.4266.3266.82
Accuracy ↑85.3783.1072.1065.1749.7050.67
Precision ↑88.6985.6065.1361.1949.8550.34
Recall ↑81.0779.6095.1382.9399.0799.33
", + "bbox": [ + 133, + 210, + 828, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tering. However, LION can achieve superior performances compared to Ferret by using a simple vision aggregator and existing datasets, implying the effectiveness of fine-grained visual knowledge.", + "bbox": [ + 75, + 388, + 468, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 465, + 230, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of vision aggregator. We conduct an ablation study of the vision aggregator with only visual grounding tasks on LION-4B during stage 2 of the stage-wise instruction-tuning. As illustrated in Fig. 5, the removal of the vision aggregator degrades REC performances, validating that aggregating multi-level vision features promotes the extraction of fine-grained spatial-aware visual knowledge.", + "bbox": [ + 75, + 489, + 468, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Stage-wise instruction-tuning mitigates the conflict between image-level and region-level tasks. We investigate the performance of two types of VL tasks under three instruction-tuning strategies, i.e., single stage, stage-wise, and stage-wise with a router. As shown in Table 3, the stage-wise instruction-tuning strategy shows a significant improvement on the average REC performance, which is completely damaged in the single stage instruction-tuning. The worse REC performance of the single stage strategy can be attributed to the lack of pretraining on large-scale grounding datasets, like in Kosmos-2 and Shikra, and the gap of their input-output modes. To address these challenges, stage-wise training progressively integrates fine-grained spatial-aware knowledge from visual grounding datasets by splitting the whole instruction-tuning process into three stages. The model can sufficiently learn diverse levels of visual knowledge in separate training stages (stages 1 and 2), and incorporate them in the final training stage (stage 3 in Fig. 3). This contributes to the performance improvements of all VL tasks. Furthermore, stage-wise instruction", + "bbox": [ + 75, + 598, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3ad7060d434ab9934c0312c1c662441e306dba7b6aaebc8f13269f70c8e1932.jpg", + "image_caption": [ + "Figure 5. The effect of the vision aggregator. The results on RefCOCO, RefCOCO+, and RefCOCOg clearly show that the proposed module can overall improve REC performances across 8 evaluation sets." + ], + "image_footnote": [], + "bbox": [ + 504, + 388, + 888, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuning with the router improves the held-out and REC performance, with a slight degradation in the held-in performance. All these results demonstrate LION's ability to handle the potential conflict of various VL tasks and maximize the learning benefit.", + "bbox": [ + 496, + 594, + 890, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dual-level visual knowledge enhances multimodal understanding abilities of MLLMs. We evaluate the performance of our model integrated with different levels of visual knowledge on various benchmarks in Table 4. It can be seen that dual-level visual knowledge can upgrade the performance of all VL tasks to varying degrees. When progressively incorporating fine-grained spatial-aware knowledge, the performances of four tasks, OKVQA, GQA, IconQA, and Visual Spatial Reasoning, are significantly improved, as they highly require region-level understanding and spatial reasoning. When inserting tags as high-level visual evidence, we can see substantial performance increases on Flickr30K and AOKVQA, which demand more comprehensive semantic knowledge than other tasks, like COCO caption and OKVQA.", + "bbox": [ + 496, + 671, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26546", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d96c5732e6c677521ecd2398c85642ded8bfaa88ef417cf9efe40bd752cbd11.jpg", + "table_caption": [ + "Table 6. Evaluation on MMBench test set, all the reported results of compared models are from the leadboard of MMBench." + ], + "table_footnote": [], + "table_body": "
ModelsText EncoderVision EncoderOverallLRARRRFP-SFP-CCP
MiniGPT-4 [54]Vincuna-7BEVA-G23.013.632.98.928.811.228.3
PandaGPT [40]Vincuna-13BImageBind ViT-H/1442.523.161.534.132.728.757.6
VisualGLM [10]ChatGLM-6BEVA-CLIP33.511.448.827.735.817.641.5
InstructBLIP [7]Vicuna-7BEVA-G33.921.647.422.533.024.441.1
LLaVA-v1.5 [25]Vicuna-v1.5-7BCLIP ViT-L/1459.532.472.649.362.352.267.7
Otter-I [19]LLaMA-7BCLIP ViT-L/1448.322.263.339.446.836.460.6
Shikra [4]Vincuna-7BCLIP ViT-L/1460.233.569.653.161.850.471.7
LMEye [23]FlanT5-XLCLIP ViT-L/1462.641.074.355.961.658.769.2
MMICL [53]FlanT5-XXLEVA-G65.244.377.964.866.553.670.6
mPLUG-Owl2 [44]LLaMA-7BCLIP ViT-L/1466.043.476.062.168.655.973.0
LIONFlanT5-XXLEVA-G73.451.784.178.474.060.878.9
", + "bbox": [ + 140, + 93, + 823, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c65a703fac70fcc7476868190635020af76a6be15a63a48b683f8d30bffd709.jpg", + "image_caption": [ + "Image Captioning" + ], + "image_footnote": [ + "User What is in the photo?" + ], + "bbox": [ + 150, + 268, + 318, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9f6248c60773c4ebd8f5e7ca12f0b922865f97797afb22e2d5412026c3c7190.jpg", + "image_caption": [ + "Visual Question Answering" + ], + "image_footnote": [ + "User On which side of the photo are the clay.pots?" + ], + "bbox": [ + 436, + 267, + 586, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ad3068a5701dcd8a9fb015a02ff704f44041cb91dde011c020ce7dee9b956e3.jpg", + "image_caption": [], + "image_footnote": [ + "InstructBLIP boy in karate uniform with white belt and black belts watching him do a kick in a martial arts class at a karate club in california.", + "The pots are on the left of the image." + ], + "bbox": [ + 405, + 391, + 496, + 407 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/03570a4610050db94adf21ec54a5588d37bb7df5925e2dbd2d69b088e566f70d.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of InstructBLIP, Shikra, and LION. We mark the hallucination or incorrect part in red, and highlight the correct part in green for comparison. These samples exhibit that LION is able to achieve superior fine-grained understanding and visual spatial reasoning capabilities with fewer hallucinated responses." + ], + "image_footnote": [ + "LION A young boy wearing a white karate uniform is performing a kick in front of a yellow wall, with the word \"Kwon\" written on the wall behind him.", + "The pots are on the right of the image." + ], + "bbox": [ + 405, + 436, + 460, + 452 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66d98fb325214e17f61bf05d6ee4eab6d4a5436288b9fbfb1cf2a3a427cf62c6.jpg", + "image_caption": [ + "Referring Expression Comprehension" + ], + "image_footnote": [ + "User Help me locate the coordinates of a brown bear near a soda bottle in the image.", + "Shikra The brown bear near a soda bottle is located at [0.002, 0.248, 0.416, 0.760]", + "LION The brown bear near a soda bottle is located at [0.550, 0.387, 0.920, 0.867]" + ], + "bbox": [ + 679, + 267, + 833, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluations on Object Hallucination and MM-Bench", + "text_level": 1, + "bbox": [ + 76, + 542, + 468, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Li et al. [22] present an open-sourced evaluation benchmark, called POPE, to evaluate the object hallucination [33]. We follow the POPE evaluation pipeline to inspect LION. The results in Table 5 show that LION has superior results, especially under popular and adversarial settings, which means that incorporating fine-grained and high-level semantic visual knowledge into MLLM can mitigate the object hallucination to some degree. To comprehensively validate the effectiveness of our method, we further evaluate LION on MMBench [27]. The results are summarized in Table 6. Our strong performances across various skills demonstrate that the progressive incorporation of fine-grained knowledge significantly alleviates the hallucination phenomenon in MLLMs.", + "bbox": [ + 75, + 580, + 468, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 801, + 267, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 6, we depict various examples to validate the advanced perceptual and reasoning abilities of LION. The left example exhibits our superior fine-grained understanding capability to correctly generate the right attributes \"white\", \"yellow\" and the character \"Kwon\". The middle", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "example validates the advantage of our model in visual spatial reasoning. The right example shows that LION accurately localizes the referring object, while Shikra provides an incorrect response caused by the misunderstanding of fine-grained details \"a soda bottle\".", + "bbox": [ + 498, + 542, + 890, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 632, + 617, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the insufficient extraction and reasoning of visual information in MLLMs, we propose LION to exploit dual-level visual knowledge, i.e., fine-grained spatial-aware visual knowledge and high-level semantic visual evidence. To mitigate the internal conflict among tasks, LION adopts a stage-wise instruction-tuning strategy. Extensive experiments validate the superiority of LION in image captioning, VQA, and visual grounding tasks.", + "bbox": [ + 496, + 657, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 792, + 666, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported by Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), National Natural Science Foundation of China (Grant No. 62236003 and 62306090), China Postdoctoral Science Foundation (Grant No. 2023TQ0095 and 2023M740923).", + "bbox": [ + 496, + 816, + 890, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26547", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5", + "[2] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 3", + "[3] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.5.6", + "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8", + "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 4", + "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 5", + "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 1, 2, 4, 5, 7, 8", + "[8] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 6", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 8", + "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 5" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Wenbo Hu, Yifan Xu, Y Li, W Li, Z Chen, and Z Tu. Bliva: A simple multimodal llm for better handling of text-rich visual questions. arXiv preprint arXiv:2308.09936, 2023. 5", + "[13] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 5", + "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6", + "[15] Woojeong Jin, Subhabrata Mukherjee, Yu Cheng, Yelong Shen, Weizhu Chen, Ahmed Hassan Awadallah, Damien Jose, and Xiang Ren. Grill: Grounded vision-language pretraining via aligning text and image regions. arXiv preprint arXiv:2305.14676, 2023. 6", + "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 6", + "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2", + "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3", + "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 8", + "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2", + "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1", + "[22] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 8", + "[23] Yunxin Li, Baotian Hu, Xinyu Chen, Lin Ma, and Min Zhang. Lmeye: An interactive perception network for large language models. arXiv preprint arXiv:2305.03701, 2023.8", + "[24] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26548", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 8", + "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5, 7", + "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8", + "[28] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021. 6", + "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6", + "[30] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 6", + "[31] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 4, 5, 6", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[33] Vipula Rawte, Amit Sheth, and Amitava Das. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023. 8", + "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 6", + "[35] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In CVPR, 2019. 1", + "[36] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense with clean-adversarial mutual learning. International Journal of Computer Vision, 2022. 1", + "[37] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In CVPR, 2023. 2", + "[38] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5", + "[40] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8", + "[41] Guangzhi Wang, Yixiao Ge, Xiaohan Ding, Mohan Kankanhalli, and Ying Shan. What makes for good visual tokenizers for large language models? arXiv preprint arXiv:2305.12223, 2023. 1, 2, 3, 5", + "[42] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 6", + "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 6", + "[44] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 8", + "[45] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2, 4, 5, 6", + "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 7", + "[47] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4, 6", + "[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5", + "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 3", + "[50] Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. Investigating the catastrophic forgetting in multimodal large language models. arXiv preprint arXiv:2309.10313, 2023. 1", + "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26549", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023.2", + "[52] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 5", + "[53] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 8", + "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 5, 7, 8" + ], + "bbox": [ + 78, + 90, + 467, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26550", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_model.json b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f2ef6cd19c476e612d976bd35050b3f494717e1e --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.197, + 0.131, + 0.775, + 0.177 + ], + "angle": 0, + "content": "LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.203, + 0.832, + 0.275 + ], + "angle": 0, + "content": "Gongwei Chen, Leyang Shen, Rui Shao,\\(\\dagger\\) Xiang Deng, Liqiang Nie\\(\\dagger\\) \nSchool of Computer Science and Technology, Harbin Institute of Technology, Shenzhen \n{chengongwei, shaorui, dengxiang, nieliqiang}@hit.edu.cn \nhttps://github.com/rshaojimmy/JiuTian" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.782 + ], + "angle": 0, + "content": "Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vVisual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between image-level and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imperfect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multi-modal benchmarks demonstrate the superiority of our model (e.g., improvement of \\(5\\%\\) accuracy on VSR and \\(3\\%\\) CIDEr on TextCaps over InstructBLIP, \\(5\\%\\) accuracy on RefCOCOg over Kosmos-2)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.21, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.813, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Recently, Large Language Models (LLMs) have demonstrated remarkable zero-shot abilities on various linguistic tasks. Assisted by LLMs, several multimodal large language models (MLLMs), such as MiniGPT-4 [54], Otter" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.31, + 0.891, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.405, + 0.891, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.594 + ], + "angle": 0, + "content": "Figure 1. Comparison between existing MLLMs and LION. The existing MLLM generates a vague and inaccurate response, while LION provides a more precise and contextually accurate description by progressively incorporating spatial-aware knowledge and softly prompting semantic visual evidence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.891, + 0.643 + ], + "angle": 0, + "content": "[19], and InstructBLIP [7], achieve significant improvements in reasoning abilities to deal with various vision-language (VL) tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.87 + ], + "angle": 0, + "content": "In most of the existing MLLMs, the visual information is mainly extracted from a vision encoder pretrained with image-level supervision (e.g., CLIP [32]), and then are adapted to a LLM by using a tiny bridge module. This makes these MLLMs inherently possess limited image understanding capabilities [21]. As shown in Fig. 1, the insufficient visual information misleads MLLMs to provide erroneous and hallucinated responses. An intuitive solution to this problem is to replace or tune the vision encoder [41]. However, it requires pretraining on massive data or suffers from the catastrophic forgetting issue [35, 36, 50], which diminishes the practical efficacy of this strategy. These predicaments highlight that the insufficient extraction of visual knowledge has become a central obstacle impeding the development of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To overcome this dilemma, as depicted in Fig. 1, we devise a dual-Level vVisual knOwledge eNhanced Multimodal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.59 + ], + "angle": 0, + "content": "Large Language Model (LION), which enriches the visual information in MLLMs in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. LION enhances the MLLM with more fine-grained perceptual abilities by studying the region-level VL tasks involving the spatial coordinates. However, we find that simply training on region-level and the original image-level VL tasks* simultaneously hurts the general performances of the MLLM due to the conflicts between these two kinds of tasks. To address this issue, we propose a novel stage-wise instruction-tuning strategy to perform image-level and region-level VL tasks separately with two different visual branches and task adapters. In addition, we devise mixture-of-adapters with a router to dynamically fuse visual information across various granularities in a unified MLLM. This progressive incorporation of fine-grained visual knowledge contributes to the mutual promotion between these two kinds of VL tasks, and spawns LION to excel in capturing fine-grained visual information and performing spatial reasoning, as shown in Fig. 1. 2) Soft prompting of high-level semantic visual evidence. Alongside the improvement of MLLMs in fine-grained perceptual capabilities, there is also an opportunity to enhance their high-level semantic understanding. LION uses an off-the-shelf vision model to extract high-level semantic knowledge, i.e., image tags, as supplementary information for the MLLM. However, as off-the-shelf vision models are typically not flawless, errors in tag predictions are inevitable. Inspired by prompt tuning, we propose a soft prompting method to mitigate the potential negative influence resulting from the imperfect predicted tags. As shown in Fig. 1, injection of semantic visual evidence alleviates the hallucination issue substantially." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.591, + 0.436, + 0.605 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.607, + 0.47, + 0.712 + ], + "angle": 0, + "content": "- To address the internal conflict between region-level and image-level VL tasks, we propose a progressive incorporation of fine-grained spatial-aware visual knowledge with a novel stage-wise instruction-tuning strategy. It achieves mutual promotion between two kinds of VL tasks and equips LION with advanced holistic and fine-grained visual perceptual abilities." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.713, + 0.47, + 0.801 + ], + "angle": 0, + "content": "- As a powerful complement, we propose to integrate image tags as high-level semantic visual evidence into MLLMs, and design a soft prompting method to alleviate the bad influence from incorrect tags. This mitigates the hallucination issue and showcases positive effects on various VL tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.803, + 0.47, + 0.863 + ], + "angle": 0, + "content": "- We evaluate LION on a wide range of VL tasks, including image captioning, visual question answering (VQA), and visual grounding, and demonstrate its superiority over the baselines. LION outperforms InstructBLIP by around \\(5\\%\\)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.607, + 0.47, + 0.863 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "accuracy on VSR, and around \\(3\\%\\) CIDEr on TextCaps, Kosmos-2 by around \\(5\\%\\) accuracy on RefCOCOg. The evaluations on POPE and MMBench exhibit the remarkable abilities of LION in alleviating object hallucination and various perceptual dimensions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.198, + 0.642, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.228, + 0.818, + 0.244 + ], + "angle": 0, + "content": "2.1. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Building on the success of LLMs, many researches have emerged to extend them to multimodal tasks, especially VL tasks. The common pipeline uses a vision model to transform the image into visual features, followed by a bridge module to align them with the feature space of LLMs. Some works [4, 11, 26, 31, 51] directly use a linear or MLP layer as bridge module, while others [1, 7, 19, 20, 46, 54] design more complicated bridge networks to compress or adaptively select visual information. Despite their impressive performance on VL tasks, there is still a lack of exploration on the effectiveness and limitation of the visual branch in a MLLM. Recently, Wang et al. [41] empirically investigate factors contributing to the formation of an effective vision encoder in a MLLM from the perspective of pretraining. Differently, our work explores the effect of region-level VL tasks on the visual understanding abilities of the MLLM, and incorporates fine-grained and high-level visual knowledge to enrich the visual branch in the MLLM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.848, + 0.57 + ], + "angle": 0, + "content": "2.2. Visual Grounding in the field of MLLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Visual grounding [16, 37, 38] is a region-level VL task that aims to establish a connection between particular regions and their textual descriptors, which plays a vital role in human-machine interaction by enabling referential dialog. In the realm of MLLMs, there are some attempts to enhance MLLMs by leveraging visual grounding tasks. Works like Shikra [4], Kosmos-2 [31], Ferret [47] and Pink [45] demonstrate the promising direction of employing visual grounding datasets to endow MLLMs with region-level visual understanding abilities. They convert existing datasets equipped with spatial coordinates, like Visual Genome [17] and RefCOCO [16], into the textual instruction format and perform instruction tuning on MLLMs. Merely considering the visual grounding task as one of several instruction-tuning tasks, these works fall short in exploring the interactions among various tasks. In contrast, our work investigates the internal conflict between visual grounding tasks and image-level VL tasks (e.g., image captioning and VQA), and proposes a stage-wise instruction-tuning strategy to address this issue, achieving a good balance between these two kinds of VL tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Here, image-level VL tasks denote image captioning and visual question answering, region-level VL tasks mean visual grounding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.895, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed LION. The model extracts holistic visual features from Q-Former, and combines them with fine-grained spatial-aware visual features from the vision aggregator. The Mixture-of-Adapters with a router in the frozen LLM dynamically fuses visual knowledge learned from different visual branches and LLM adapters based on the task types (image-level and region-level)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.34, + 0.153, + 0.355 + ], + "angle": 0, + "content": "3. LION" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.413 + ], + "angle": 0, + "content": "In this section, we present the dual-Level vVisual knOwlEdge eNhanced multimodal large language model (LION). The whole framework is depicted in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.47, + 0.456 + ], + "angle": 0, + "content": "3.1. Progressive Incorporation of Fine-grained Spatial-Aware Visual Knowledge" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.465, + 0.398, + 0.48 + ], + "angle": 0, + "content": "3.1.1 Reorganizing Visual Grounding Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.763 + ], + "angle": 0, + "content": "To incorporate fine-grained spatial-aware visual knowledge into MLLMs, we make use of region-level VL tasks, i.e., visual grounding, and meticulously process the data with spatial coordinates in a unified format for instruction-tuning MLLM. Visual grounding requires the model to generate or comprehend natural language expressions referring to particular objects or regions within an image, e.g., \"a man with glasses\". Referring to objects or regions in complex images needs an ability of precisely comprehending fine-grained visual information. Current MLLMs lack such referring comprehending, as they mainly target a coarse alignment of VL modalities when pretrained on massive image-text pairs [2, 41]. In this regard, we introduce visual grounding tasks as a kind of region-level VL tasks for the instruction-tuning of MLLMs. This aims to endow the model with fine-grained visual understanding ability such that better performance on image-level VL tasks (e.g., image captioning and VQA) might be achieved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We adopt two types of visual grounding tasks, including referring expression comprehension (REC) and referring expression generation (REG) [49]. We use the Visual Genome dataset [18], which associates a local area with one short description, to construct REC/REG tasks. The templates used to organize the Visual Genome dataset in a unified instruction-tuning format can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "One core point in reorganizing visual grounding tasks is the way of processing positions. Normally, the position of" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.341, + 0.782, + 0.352 + ], + "angle": 0, + "content": "Stage 1: Instruction-Tuning on Image-Level VL Tasks" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.353, + 0.891, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.383, + 0.787, + 0.395 + ], + "angle": 0, + "content": "Stage 2: Instruction-Tuning on Region-Level VL Tasks" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.395, + 0.891, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.444, + 0.875, + 0.456 + ], + "angle": 0, + "content": "Stage 3: Instruction-Tuning on Image-Level and Region-Level VL Tasks" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.456, + 0.891, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.593 + ], + "angle": 0, + "content": "Figure 3. The stage-wise instruction-tuning strategy. **Stage 1:** We instruction-tune Q-Former and the image-level adapter on image-level VL tasks. **Stage 2:** We instruction-tune the vision aggregator (VA), MLP, and the region-level adapter on region-level VL tasks. **Stage 3:** The Mixture-of-Adapters is devised to form a unified model for instruction-tuning on both kinds of VL tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.69 + ], + "angle": 0, + "content": "an object phrase is presented in the format of bounding box \\([x_{min},y_{min},x_{max},y_{max}]\\). We use a natural language style to describe object positions along with the square brackets. A sample in the REC task is displayed as follows: \"How can I locate a glass of beer in the image? Please provide the coordinates. Answer: [0.525, 0.0, 0.675, 0.394]\"." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.71, + 0.858, + 0.726 + ], + "angle": 0, + "content": "3.1.2 The Stage-Wise Instruction-tuning Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To facilitate MLLMs with fine-grained spatial-aware knowledge, the most intuitive way is to directly instruction-tune MLLMs with both image-level and region-level VL tasks in one stage. However, this single-stage instruction-tuning strategy is sub-optimal, and suffers from the internal conflict between these two kinds of VL tasks. We summarize two main issues leading to the internal conflict. 1) One is the need of region-level modality-alignment pretraining. In concurrent works that integrate visual grounding ability, pretraining on the region-level multimodal datasets including visual grounding is a crucial step. Some works" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26542" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.362 + ], + "angle": 0, + "content": "[31, 45, 47] elaborately create very large visual grounding datasets (e.g., GRIT-20M in kosmos-2 [31]) to advance MLLM in fine-grained perception and understanding. The single-stage instruction-tuning makes it challenging to adapt visual representations learned for image-level alignment to region-level VL tasks under a limited training configuration. 2) Another is the gap between the input-output modes of image-level VL tasks and region-level visual grounding tasks. The latter additionally requires MLLMs to understand specific phrases (in the format “\\([x_{min}, y_{min}, x_{max}, y_{max}]\\)” about the positions of objects. They are semantically distinct from natural languages used in image-level tasks. This requirement necessitates the tuning of the LLM to adapt to region-level tasks, but may disrupt the internal state of the LLM suitable for image-level VL tasks. To address the above issues, we devise a stage-wise instruction-tuning strategy and mixture-of-adapters with a router." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.606 + ], + "angle": 0, + "content": "The stage-wise instruction-tuning strategy is proposed to alleviate the internal conflict between image-level and region-level VL tasks during instruction-tuning. It is composed of three stages for instruction-tuning on image-level, region-level VL tasks and both, respectively, which is depicted in Fig. 3. In stage 1, we follow instructBLIP [7] and fine-tune Q-Former and the image-level adapter [5] in the LLM on image-level VL tasks, such as image captioning and VQA. In stage 2, we propose a vision aggregator for better capturing visual features in fine-grained understanding, which will be introduced later, and tune it with MLP and the region-level adapter on region-level VL tasks. The independent training in the first two stages greatly fulfills the requirements of sufficiently learning both image-level and region-level tasks, providing a solid foundation for subsequent joint training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Mixture-of-Adapters with a Router. In stage 3 of our stage-wise instruction-tuning, we need a unified model but encounter a situation where adapters of LLM in stages 1 and 2 are different and suit distinct input-output modes. Inspired by Mixture-of-Experts, we treat each adapter as an expert, and propose a router module to avoid the potential interference between them, as depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.807 + ], + "angle": 0, + "content": "An adapter [5] is inserted at each FFN layer in a parallel manner. Assuming \\(X \\in \\mathcal{R}^{L \\times D}\\) is the hidden representations generated by a self-attention / causal attention layer, the output representations after FFN (represented as \\(\\mathbf{F}\\)) with the adapter (denoted by \\(\\mathbf{H}\\)) layer are formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.822, + 0.47, + 0.839 + ], + "angle": 0, + "content": "\\[\nO = \\mathbf {F} (X) + \\mathbf {H} (X), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.862, + 0.469, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} (X) = W _ {u} \\left(\\sigma \\left(W _ {d} X\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.9 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a non-linear function, ReLU. Our router module" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.088, + 0.867, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.2, + 0.894, + 0.242 + ], + "angle": 0, + "content": "Figure 4. Instruction template with soft prompt. We use a well-designed instruction template with trainable soft prompts to inject the image tags generated by the RAM model into LION." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.254, + 0.893, + 0.329 + ], + "angle": 0, + "content": "aims to dynamically aggregate the hidden features from the main branches and the multiple adapter branches according to task types. Given a set of adapters \\(\\{\\mathbf{H}_1,\\dots ,\\mathbf{H}_K\\}\\), each kind of task \\(t\\) defines a specific router function \\(\\mathbf{R}^t\\) to generate new hidden features, which can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.339, + 0.892, + 0.381 + ], + "angle": 0, + "content": "\\[\nO ^ {t} = \\mathbf {F} (X) + \\sum_ {k = 1} ^ {K} \\mathrm {G} _ {k} ^ {t} \\odot \\mathbf {H} _ {k} (X). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.893, + 0.512 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_k^t\\in \\mathcal{R}^D\\) is a trainable vector that modulates the hidden features from each adapter and makes them suitable for the target task. In practice, we define two types of tasks, one for image-level VL tasks (image captioning and VQA), the other for fine-grained VL tasks (visual grounding). Compared to directly incorporating multiple adapters, the router module provides a better way to maximize the complementarity of image-level and region-level tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.894, + 0.618 + ], + "angle": 0, + "content": "We use the standard language modeling loss in all instruction-tuning stages. In the experiments, we demonstrate that stage-wise training is superior to single-stage training, and ensures a good balance between high-level and fine-grained visual understanding capabilities, further achieves a significant mutual promotion between image-level and region-level VL tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.682, + 0.651 + ], + "angle": 0, + "content": "3.1.3 Vision Aggregator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To extract more sufficient visual details from input images, we devise a vision aggregator that integrates multi-level hidden features of the pretrained visual encoder. Although the vision encoder has a global reception field in all layers, it is verified that different transformer layers learn visual information at different scales [9], e.g., lower layers learn visual details. Thus, our vision aggregator makes fine-grained spatial-aware visual knowledge more likely to be learned based on visual grounding tasks. Specifically, our vision aggregator can be regarded as a tiny transformer-style network, consisting of two transformer layers for aggregating the hidden features from the vision encoder. Given the hidden features \\(\\{V_{i}, V_{j}, V_{k}\\}\\) from some middle layers in the vision encoder, the vision aggregation module uses two blocks to sequentially integrate the former two features with the last feature. Each block \\(\\mathcal{B}\\) is composed of self attention" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26543" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison on image captioning and VQA. “†” denotes including in-house data that are publicly inaccessible. “*” means our evaluated results by using publicly released checkpoints, which are only for reference as official evaluation settings are incomplete. We report CIDEr score for Flickr30K, COOCap, and TextCaps, Mean Reciprocal Rank (MRR) for Visual Dialog (VisDial), and top-1 accuracy for others. The best and second performances for each benchmark are indicated in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.147, + 0.859, + 0.38 + ], + "angle": 0, + "content": "
ModelFlickr30KCOCOCapTextCapsOKVQAAOKVQAGQAIconQAVSRVisDial
Flamingo-3B [1]60.6073.00------46.10
Flamingo-9B [1]61.5079.40-44.70----48.00
Kosmos-1 [13]67.1084.70-------
Kosmos-2 [31]80.50--------
AdapterV2 [11]-122.20-------
Shikra [4]73.90117.50-47.16-----
Pink [45]---59.50-52.6047.8066.30-
MiniGPT4 [54]17.75*17.04*24.06*37.5034.51*30.8037.6041.6016.52*
LLaVA [26]48.03*73.85*45.54*54.4034.51*41.3043.0051.208.65*
MiniGPTV2 [3]80.75*129.16*80.60*56.90-60.3047.7060.608.47*
BLIVA [12]87.10-----44.8862.2045.63
InstructBLIP† (T5XL) [7]84.50138.2182.5549.2857.8648.4050.0064.8046.60
InstructBLIP† (T5XXL) [7]83.50138.2882.5348.5956.1647.9051.2065.6048.50
InstructBLIP (T5XL) [7]83.71135.47104.1747.3856.1246.3452.4769.9348.75
InstructBLIP (T5XXL) [7]85.79138.63105.4453.0259.3847.7453.1868.4650.41
LION-4B85.57138.20104.8751.0859.9849.5054.9172.9650.02
LION-12B87.12139.25108.7657.3360.8751.5654.8973.7750.42
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.43 + ], + "angle": 0, + "content": "(Attn), cross attention (XAttn), and Feed-forward network (FFN) arranged in a sequential manner. Finally, the output features \\(\\bar{V}\\) is generated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.44, + 0.47, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\bar {V} = \\mathcal {B} _ {2} \\left(\\mathcal {B} _ {1} \\left(V _ {i}; V _ {j}\\right); V _ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.483, + 0.469, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} (X; Y) = \\operatorname {F F N} (\\operatorname {X A t t n} (\\operatorname {A t t n} (X), Y)). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.472, + 0.567 + ], + "angle": 0, + "content": "In practice, we use the middle layers \\(\\{i = L - 1, j = 2L/3, k = L/3\\}\\) in the vision encoder to produce the hidden features as the input to VA, where \\(L\\) is the number of layers in the vision encoder." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.469, + 0.606 + ], + "angle": 0, + "content": "3.2. Soft Prompting of High-Level Semantic Visual Evidence" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.78 + ], + "angle": 0, + "content": "The vision encoder in a MLLM may be insufficient in comprehensively extracting visual information required by complex multi-modal tasks, although it has been trained on large-scale image-text pairs. It has been demonstrated that increasing the amount and quality of pretraining multimodal datasets can significantly improve the visual understanding capability of MLLMs [41], but inevitably induces prohibitive computational overhead. An appealing alternative is to harness the convenient and powerful off-the-shelf vision models to capture various aspects of visual content within an image as a supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We choose the recognize anything model (RAM) [52] as an off-the-shelf vision model to provide diverse tags, encompassing objects, scenes, actions, and attributes, as visual evidence to support comprehensive visual perception. Instead of directly adding tags in the instruction, we design a soft prompting method to guide the model to adaptively use the inserted tags in order to avoid the potential negative influence caused by the imperfect predictions from RAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.895, + 0.642 + ], + "angle": 0, + "content": "In Fig. 4, we present the instruction template of tags along with the soft prompt that is a trainable vector. Our soft prompting approach can be regarded as a kind of prompt tuning methods, which guides the model toward the right direction. In standard prompt tuning works, the right direction is directly formulated as the optimization for task goals. In our work, the right direction is specified by a tailored sentence, \"According to , you are allowed to use or partially use the following tags: \", and \"\" will be replaced by the soft prompt. Our soft prompting method for inserting tags has some distinct properties. It is designed to adaptively select valuable information from tags, rather than serving a specific task, as seen in standard prompt tuning methods. Our method directly uses the output labels from a small off-the-shelf vision model to incorporate high-level semantic visual evidence into a MLLM, so as to eliminate extra computational overhead of the feature alignment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.634, + 0.674 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.894, + 0.744 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our model along with the quantitative and qualitative analyses. Please refer to Appendix for implementation details and training details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.828, + 0.772 + ], + "angle": 0, + "content": "4.1. Evaluations on Image-Level VL Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here, we evaluate multi-modal understanding abilities of LION on two kinds of image-level VL tasks, image captioning and VQA. Image captioning requires the model to generate a text description of the input image. We use COCO caption [6], TextCaps [39] and Flickr30K [48] as benchmarks, and report CIDEr as the evaluation metric. We utilize greedy search for caption generation. VQA provides an image along with a specific question for the model, asking" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26544" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.178, + 0.09, + 0.791, + 0.104 + ], + "angle": 0, + "content": "Table 2. Comparison on REC. \"Avg.\" means the average of top-1 accuracy over all the 8 evaluation sets." + }, + { + "type": "table", + "bbox": [ + 0.194, + 0.106, + 0.77, + 0.344 + ], + "angle": 0, + "content": "
ModelRefCOCORefCOCO+RefCOCOgAvg.
valtest-Atest-Bvaltest-Atest-Bvaltest
Zero-shot Setting
Kosmos-2 [31]52.3257.4247.2645.4850.7342.2460.5761.6552.21
GRILL [15]-------47.50-
Pink [45]54.1061.2044.2043.9050.7035.0059.1060.1051.00
LION-4B57.8956.0758.4046.3845.2947.5064.7463.5654.98
LION-12B58.5456.4159.3645.9345.7347.8966.1264.6955.58
Fune-tuning Setting
OFA-L [42]79.9683.6776.3968.2976.0061.7567.5767.5872.65
VisionLLM-H [43]-86.70-------
Shikra-7B [4]87.0190.6180.2481.6087.3672.1282.2782.1982.93
Shikra-13B [4]87.8391.1181.8182.8987.7974.4182.6483.1683.96
Pink [45]88.3091.7084.0081.4087.5073.7083.7083.7084.25
Ferret-7B [47]87.4991.3582.4580.7887.3873.1483.9384.7683.91
Ferret-13B [47]89.4892.4184.3682.8188.1475.1785.8386.3485.57
MiniGPTv2 [3]88.6991.6585.3379.9785.1274.4584.4484.6684.29
LION-4B89.7392.2984.8283.6088.7277.3485.6985.6385.98
LION-12B89.8093.0285.5783.9589.2278.0685.5285.7486.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.353, + 0.47, + 0.422 + ], + "angle": 0, + "content": "Table 3. The comparison of various strategies in the instruction-tuning period. \"REC Avg.\" represents the average score of all REC tasks. \"Held-in\" denotes the average score of COOCap, TextCaps, OKVQA, and AOKVQA, while \"Held-out\" means the average score of Flickr30K, GQA, IconQA, VSR, and VisDial." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.425, + 0.424, + 0.495 + ], + "angle": 0, + "content": "
StrategyImage-LevelRegion-Level
Held-inHeld-outREC Avg.
Single Stage84.7960.203.78
Stage-wise88.0761.9154.46
w/ Router87.6762.1654.98
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.687 + ], + "angle": 0, + "content": "for the output as an answer. We evaluate LION on six VQA datasets, including OKVQA [30], AOKVQA [34], GQA [14], IconQA [28], Visual Spatial Reasoning [24], and Visual Dialog (VisDial) [8]. For OKVQA, A-OKVQA, and GQA, we employ an open-ended generation with a greedy decoding strategy, For IconQA, Visual Spatial Reasoning, and Visual Dialog, we match the output with various candidates, and select the candidate with the highest value as the prediction. We report Mean Reciprocal Rank (MRR) for Visual Dialog, and top-1 accuracy for other VQA tasks. The detailed descriptions of these datasets and inference instructions are presented in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, LION achieves the best performance across 7 out of 9 benchmarks, and the second on OKVQA. LION shares the same training datasets with InstructBLIP, except Visual Genome dataset adopted in our work and the in-house dataset, WebCapFilt, used in InstructBLIP. The amount of Visual Genome dataset (3.6M) is far smaller than WebCapFilt (14M). Compared to the original InstructBLIP trained with WebCapFilt, LION exhibits superior performances on all zero-shot evaluation benchmarks, showcasing a better generalization ability. We also re-implemented InstructBLIP on the same instruction-tuning datasets. The comparison shows the consistent and significant improvements of LION over the re-implemented InstructBLIP, demonstrating the effectiveness of incorpo" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.446 + ], + "angle": 0, + "content": "rating dual-level visual knowledge. Shikra, MiniGPTV2 and Pink are also integrated with visual grounding abilities. Their inferior results on most image-level VL tasks exhibit that our proposed stage-wise instruction-tuning strategy and soft prompting of high-level semantic knowledge help enhance holistic visual understanding abilities of MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.455, + 0.834, + 0.471 + ], + "angle": 0, + "content": "4.2. Evaluations on Region-Level VL Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.584 + ], + "angle": 0, + "content": "To assess the fine-grained perceptual and reasoning abilities of LION, we evaluate it on three REC datasets, RefCOCO [16], RefCOCO+ [16], RefCOCOg [29]. REC requires the model to locate the target object given a referring expression. We follow the standard setting, and use accuracy as an evaluation metric, which means it is correct when the IOU between prediction and ground-truth is no less than 0.5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In Table 2, we show the comparison between LION and other MLLMs with respect to the grounding abilities, under the settings of zero-shot and fine-tuning evaluations, respectively. In the zero-shot evaluation setting, we directly employ LION to generate coordinates of referring expressions on three datasets. Our model shows significant improvements on most evaluation sets over Kosmos-2 and Pink, except test-A sets of RefCOCO and RefCOCO+. The languages used in RefCOCOg are more flowery than those used in RefCOCO and RefCOCO+. The significant improvements on RefCOCOg clearly demonstrate that LION can handle complex referring expressions and has superior zero-shot spatial-aware visual understanding abilities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the fine-tuning setting, we fine-tune LION with training samples from three REC datasets. As shown in Table 2, our model achieves the best performance on average and on most of the evaluation sets, indicating the advanced fine-grained perception ability of our model. Ferret proposes a spatial-aware visual sampler to handle free-form referred regions, and meticulously constructs an extensive grounding dataset with lots of efforts on data generation and fil" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26545" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation studies of dual-level visual knowledge. \"VG\" means visual grounding tasks. \"Held-in\" and \"Held-out\" denote the training images of tasks are seen and unseen, respectively. \"REC Avg.\" means the average score of all REC tasks." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.12, + 0.855, + 0.189 + ], + "angle": 0, + "content": "
ComponentsHeld-inHeld-outREC Avg.
VGTagsCOCOCapTextCapsOKVQAAOKVQAFlickr30KGQAVSRIconQAVisDial
××135.47104.1747.3856.1283.7146.3469.9352.4748.75-
✓×137.87104.8451.0756.9083.9949.2273.2054.6749.7054.98
✓✓138.20104.8751.0859.9885.5749.5072.9654.9150.0254.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.195, + 0.845, + 0.209 + ], + "angle": 0, + "content": "Table 5. Evaluation of object hallucination on POPE benchmark. F1 score is the major metric for hallucication evaluation." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.212, + 0.83, + 0.378 + ], + "angle": 0, + "content": "
DatasetsMetricsLIONShikra [4]InstructBLIP [7]MiniGPT-4 [54]LLaVA [26]mPLUG-Owl [46]
RandomF1-Score ↑88.3386.1989.2780.1766.6468.39
Accuracy ↑88.9786.9088.5779.6750.3753.97
Precision ↑97.1294.4084.0978.2450.1952.07
Recall ↑81.0079.2795.1382.2099.1399.60
PopularF1-Score ↑85.9483.1684.6673.0266.4466.94
Accuracy ↑86.7783.9782.7769.7349.8750.90
Precision ↑91.6987.5576.2765.8649.9350.46
Recall ↑80.8779.2095.1381.9399.2799.40
AdversarialF1-score ↑84.7182.4977.3270.4266.3266.82
Accuracy ↑85.3783.1072.1065.1749.7050.67
Precision ↑88.6985.6065.1361.1949.8550.34
Recall ↑81.0779.6095.1382.9399.0799.33
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.47, + 0.45 + ], + "angle": 0, + "content": "tering. However, LION can achieve superior performances compared to Ferret by using a simple vision aggregator and existing datasets, implying the effectiveness of fine-grained visual knowledge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.231, + 0.483 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.598 + ], + "angle": 0, + "content": "The effect of vision aggregator. We conduct an ablation study of the vision aggregator with only visual grounding tasks on LION-4B during stage 2 of the stage-wise instruction-tuning. As illustrated in Fig. 5, the removal of the vision aggregator degrades REC performances, validating that aggregating multi-level vision features promotes the extraction of fine-grained spatial-aware visual knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Stage-wise instruction-tuning mitigates the conflict between image-level and region-level tasks. We investigate the performance of two types of VL tasks under three instruction-tuning strategies, i.e., single stage, stage-wise, and stage-wise with a router. As shown in Table 3, the stage-wise instruction-tuning strategy shows a significant improvement on the average REC performance, which is completely damaged in the single stage instruction-tuning. The worse REC performance of the single stage strategy can be attributed to the lack of pretraining on large-scale grounding datasets, like in Kosmos-2 and Shikra, and the gap of their input-output modes. To address these challenges, stage-wise training progressively integrates fine-grained spatial-aware knowledge from visual grounding datasets by splitting the whole instruction-tuning process into three stages. The model can sufficiently learn diverse levels of visual knowledge in separate training stages (stages 1 and 2), and incorporate them in the final training stage (stage 3 in Fig. 3). This contributes to the performance improvements of all VL tasks. Furthermore, stage-wise instruction" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.389, + 0.889, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Figure 5. The effect of the vision aggregator. The results on RefCOCO, RefCOCO+, and RefCOCOg clearly show that the proposed module can overall improve REC performances across 8 evaluation sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.671 + ], + "angle": 0, + "content": "tuning with the router improves the held-out and REC performance, with a slight degradation in the held-in performance. All these results demonstrate LION's ability to handle the potential conflict of various VL tasks and maximize the learning benefit." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.893, + 0.899 + ], + "angle": 0, + "content": "Dual-level visual knowledge enhances multimodal understanding abilities of MLLMs. We evaluate the performance of our model integrated with different levels of visual knowledge on various benchmarks in Table 4. It can be seen that dual-level visual knowledge can upgrade the performance of all VL tasks to varying degrees. When progressively incorporating fine-grained spatial-aware knowledge, the performances of four tasks, OKVQA, GQA, IconQA, and Visual Spatial Reasoning, are significantly improved, as they highly require region-level understanding and spatial reasoning. When inserting tags as high-level visual evidence, we can see substantial performance increases on Flickr30K and AOKVQA, which demand more comprehensive semantic knowledge than other tasks, like COCO caption and OKVQA." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26546" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.079, + 0.85, + 0.093 + ], + "angle": 0, + "content": "Table 6. Evaluation on MMBench test set, all the reported results of compared models are from the leadboard of MMBench." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.094, + 0.825, + 0.25 + ], + "angle": 0, + "content": "
ModelsText EncoderVision EncoderOverallLRARRRFP-SFP-CCP
MiniGPT-4 [54]Vincuna-7BEVA-G23.013.632.98.928.811.228.3
PandaGPT [40]Vincuna-13BImageBind ViT-H/1442.523.161.534.132.728.757.6
VisualGLM [10]ChatGLM-6BEVA-CLIP33.511.448.827.735.817.641.5
InstructBLIP [7]Vicuna-7BEVA-G33.921.647.422.533.024.441.1
LLaVA-v1.5 [25]Vicuna-v1.5-7BCLIP ViT-L/1459.532.472.649.362.352.267.7
Otter-I [19]LLaMA-7BCLIP ViT-L/1448.322.263.339.446.836.460.6
Shikra [4]Vincuna-7BCLIP ViT-L/1460.233.569.653.161.850.471.7
LMEye [23]FlanT5-XLCLIP ViT-L/1462.641.074.355.961.658.769.2
MMICL [53]FlanT5-XXLEVA-G65.244.377.964.866.553.670.6
mPLUG-Owl2 [44]LLaMA-7BCLIP ViT-L/1466.043.476.062.168.655.973.0
LIONFlanT5-XXLEVA-G73.451.784.178.474.060.878.9
" + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.256, + 0.284, + 0.268 + ], + "angle": 0, + "content": "Image Captioning" + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.269, + 0.319, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.095, + 0.363, + 0.194, + 0.383 + ], + "angle": 0, + "content": "User What is in the photo?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.095, + 0.397, + 0.371, + 0.435 + ], + "angle": 0, + "content": "InstructBLIP boy in karate uniform with white belt and black belts watching him do a kick in a martial arts class at a karate club in california." + }, + { + "type": "image_footnote", + "bbox": [ + 0.095, + 0.44, + 0.373, + 0.478 + ], + "angle": 0, + "content": "LION A young boy wearing a white karate uniform is performing a kick in front of a yellow wall, with the word \"Kwon\" written on the wall behind him." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.363, + 0.373, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.256, + 0.585, + 0.267 + ], + "angle": 0, + "content": "Visual Question Answering" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.268, + 0.588, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.406, + 0.361, + 0.616, + 0.383 + ], + "angle": 0, + "content": "User On which side of the photo are the clay.pots?" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.392, + 0.497, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.406, + 0.408, + 0.576, + 0.417 + ], + "angle": 0, + "content": "The pots are on the left of the image." + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.438, + 0.461, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.406, + 0.453, + 0.584, + 0.461 + ], + "angle": 0, + "content": "The pots are on the right of the image." + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.256, + 0.859, + 0.267 + ], + "angle": 0, + "content": "Referring Expression Comprehension" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.268, + 0.834, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.643, + 0.361, + 0.859, + 0.391 + ], + "angle": 0, + "content": "User Help me locate the coordinates of a brown bear near a soda bottle in the image." + }, + { + "type": "image_footnote", + "bbox": [ + 0.643, + 0.397, + 0.859, + 0.427 + ], + "angle": 0, + "content": "Shikra The brown bear near a soda bottle is located at [0.002, 0.248, 0.416, 0.760]" + }, + { + "type": "image_footnote", + "bbox": [ + 0.643, + 0.44, + 0.859, + 0.47 + ], + "angle": 0, + "content": "LION The brown bear near a soda bottle is located at [0.550, 0.387, 0.920, 0.867]" + }, + { + "type": "list", + "bbox": [ + 0.643, + 0.361, + 0.859, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.491, + 0.892, + 0.536 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of InstructBLIP, Shikra, and LION. We mark the hallucination or incorrect part in red, and highlight the correct part in green for comparison. These samples exhibit that LION is able to achieve superior fine-grained understanding and visual spatial reasoning capabilities with fewer hallucinated responses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.469, + 0.573 + ], + "angle": 0, + "content": "4.4. Evaluations on Object Hallucination and MM-Bench" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.792 + ], + "angle": 0, + "content": "Li et al. [22] present an open-sourced evaluation benchmark, called POPE, to evaluate the object hallucination [33]. We follow the POPE evaluation pipeline to inspect LION. The results in Table 5 show that LION has superior results, especially under popular and adversarial settings, which means that incorporating fine-grained and high-level semantic visual knowledge into MLLM can mitigate the object hallucination to some degree. To comprehensively validate the effectiveness of our method, we further evaluate LION on MMBench [27]. The results are summarized in Table 6. Our strong performances across various skills demonstrate that the progressive incorporation of fine-grained knowledge significantly alleviates the hallucination phenomenon in MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.268, + 0.818 + ], + "angle": 0, + "content": "4.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "As shown in Fig. 6, we depict various examples to validate the advanced perceptual and reasoning abilities of LION. The left example exhibits our superior fine-grained understanding capability to correctly generate the right attributes \"white\", \"yellow\" and the character \"Kwon\". The middle" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.543, + 0.892, + 0.619 + ], + "angle": 0, + "content": "example validates the advantage of our model in visual spatial reasoning. The right example shows that LION accurately localizes the referring object, while Shikra provides an incorrect response caused by the misunderstanding of fine-grained details \"a soda bottle\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.619, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.78 + ], + "angle": 0, + "content": "To address the insufficient extraction and reasoning of visual information in MLLMs, we propose LION to exploit dual-level visual knowledge, i.e., fine-grained spatial-aware visual knowledge and high-level semantic visual evidence. To mitigate the internal conflict among tasks, LION adopts a stage-wise instruction-tuning strategy. Extensive experiments validate the superiority of LION in image captioning, VQA, and visual grounding tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.793, + 0.667, + 0.81 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.893 + ], + "angle": 0, + "content": "This study is supported by Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), National Natural Science Foundation of China (Grant No. 62236003 and 62306090), China Postdoctoral Science Foundation (Grant No. 2023TQ0095 and 2023M740923)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26547" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.473, + 0.327 + ], + "angle": 0, + "content": "[3] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.5.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.472, + 0.515 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 1, 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.59, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.661, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.746, + 0.472, + 0.829 + ], + "angle": 0, + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[12] Wenbo Hu, Yifan Xu, Y Li, W Li, Z Chen, and Z Tu. Bliva: A simple multimodal llm for better handling of text-rich visual questions. arXiv preprint arXiv:2308.09936, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[13] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Woojeong Jin, Subhabrata Mukherjee, Yu Cheng, Yelong Shen, Weizhu Chen, Ahmed Hassan Awadallah, Damien Jose, and Xiang Ren. Grill: Grounded vision-language pretraining via aligning text and image regions. arXiv preprint arXiv:2305.14676, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.586 + ], + "angle": 0, + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[22] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[23] Yunxin Li, Baotian Hu, Xinyu Chen, Lin Ma, and Min Zhang. Lmeye: An interactive perception network for large language models. arXiv preprint arXiv:2305.03701, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26548" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.163 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[28] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.377, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[30] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[31] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.505, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[33] Vipula Rawte, Amit Sheth, and Amitava Das. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.703, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[35] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[36] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense with clean-adversarial mutual learning. International Journal of Computer Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.803, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[37] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[39] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[40] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Guangzhi Wang, Yixiao Ge, Xiaohan Ding, Mohan Kankanhalli, and Ying Shan. What makes for good visual tokenizers for large language models? arXiv preprint arXiv:2305.12223, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[42] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[44] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[45] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[47] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[50] Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. Investigating the catastrophic forgetting in multimodal large language models. arXiv preprint arXiv:2309.10313, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26549" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.468, + 0.203 + ], + "angle": 0, + "content": "[52] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.206, + 0.468, + 0.274 + ], + "angle": 0, + "content": "[53] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.468, + 0.332 + ], + "angle": 0, + "content": "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26550" + } + ] +] \ No newline at end of file diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_origin.pdf b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..79221238abb2efd9b2b34a9d4a1d8b80cf338934 --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/ee456ca8-65df-48b8-a118-352557caef54_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:032271cd3305bb8a895b388e310bf5cb38571ba3858d675989a413e2d1cd86ca +size 6635940 diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/full.md b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/full.md new file mode 100644 index 0000000000000000000000000000000000000000..23478eabce33b03fbbbc3b8011295b86cbbfb606 --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/full.md @@ -0,0 +1,297 @@ +# LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge + +Gongwei Chen, Leyang Shen, Rui Shao, $\dagger$ Xiang Deng, Liqiang Nie $\dagger$ +School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen +{chengongwei, shaorui, dengxiang, nieliqiang}@hit.edu.cn +https://github.com/rshaojimmy/JiuTian + +# Abstract + +Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vVisual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between image-level and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imperfect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multi-modal benchmarks demonstrate the superiority of our model (e.g., improvement of $5\%$ accuracy on VSR and $3\%$ CIDEr on TextCaps over InstructBLIP, $5\%$ accuracy on RefCOCOg over Kosmos-2). + +# 1. Introduction + +Recently, Large Language Models (LLMs) have demonstrated remarkable zero-shot abilities on various linguistic tasks. Assisted by LLMs, several multimodal large language models (MLLMs), such as MiniGPT-4 [54], Otter + +![](images/0197e60522e17f12efd32fb03d4677193d011df1a7b4135044b08cd008e6bda8.jpg) + +![](images/172da6b0638d7c17d4b19ff0e6adb63651d6035da34cc85e3cb5489aae5a1d71.jpg) +Figure 1. Comparison between existing MLLMs and LION. The existing MLLM generates a vague and inaccurate response, while LION provides a more precise and contextually accurate description by progressively incorporating spatial-aware knowledge and softly prompting semantic visual evidence. + +[19], and InstructBLIP [7], achieve significant improvements in reasoning abilities to deal with various vision-language (VL) tasks. + +In most of the existing MLLMs, the visual information is mainly extracted from a vision encoder pretrained with image-level supervision (e.g., CLIP [32]), and then are adapted to a LLM by using a tiny bridge module. This makes these MLLMs inherently possess limited image understanding capabilities [21]. As shown in Fig. 1, the insufficient visual information misleads MLLMs to provide erroneous and hallucinated responses. An intuitive solution to this problem is to replace or tune the vision encoder [41]. However, it requires pretraining on massive data or suffers from the catastrophic forgetting issue [35, 36, 50], which diminishes the practical efficacy of this strategy. These predicaments highlight that the insufficient extraction of visual knowledge has become a central obstacle impeding the development of MLLMs. + +To overcome this dilemma, as depicted in Fig. 1, we devise a dual-Level vVisual knOwledge eNhanced Multimodal + +Large Language Model (LION), which enriches the visual information in MLLMs in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. LION enhances the MLLM with more fine-grained perceptual abilities by studying the region-level VL tasks involving the spatial coordinates. However, we find that simply training on region-level and the original image-level VL tasks* simultaneously hurts the general performances of the MLLM due to the conflicts between these two kinds of tasks. To address this issue, we propose a novel stage-wise instruction-tuning strategy to perform image-level and region-level VL tasks separately with two different visual branches and task adapters. In addition, we devise mixture-of-adapters with a router to dynamically fuse visual information across various granularities in a unified MLLM. This progressive incorporation of fine-grained visual knowledge contributes to the mutual promotion between these two kinds of VL tasks, and spawns LION to excel in capturing fine-grained visual information and performing spatial reasoning, as shown in Fig. 1. 2) Soft prompting of high-level semantic visual evidence. Alongside the improvement of MLLMs in fine-grained perceptual capabilities, there is also an opportunity to enhance their high-level semantic understanding. LION uses an off-the-shelf vision model to extract high-level semantic knowledge, i.e., image tags, as supplementary information for the MLLM. However, as off-the-shelf vision models are typically not flawless, errors in tag predictions are inevitable. Inspired by prompt tuning, we propose a soft prompting method to mitigate the potential negative influence resulting from the imperfect predicted tags. As shown in Fig. 1, injection of semantic visual evidence alleviates the hallucination issue substantially. + +Our main contributions are summarized as follows: + +- To address the internal conflict between region-level and image-level VL tasks, we propose a progressive incorporation of fine-grained spatial-aware visual knowledge with a novel stage-wise instruction-tuning strategy. It achieves mutual promotion between two kinds of VL tasks and equips LION with advanced holistic and fine-grained visual perceptual abilities. +- As a powerful complement, we propose to integrate image tags as high-level semantic visual evidence into MLLMs, and design a soft prompting method to alleviate the bad influence from incorrect tags. This mitigates the hallucination issue and showcases positive effects on various VL tasks. +- We evaluate LION on a wide range of VL tasks, including image captioning, visual question answering (VQA), and visual grounding, and demonstrate its superiority over the baselines. LION outperforms InstructBLIP by around $5\%$ + +accuracy on VSR, and around $3\%$ CIDEr on TextCaps, Kosmos-2 by around $5\%$ accuracy on RefCOCOg. The evaluations on POPE and MMBench exhibit the remarkable abilities of LION in alleviating object hallucination and various perceptual dimensions. + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +Building on the success of LLMs, many researches have emerged to extend them to multimodal tasks, especially VL tasks. The common pipeline uses a vision model to transform the image into visual features, followed by a bridge module to align them with the feature space of LLMs. Some works [4, 11, 26, 31, 51] directly use a linear or MLP layer as bridge module, while others [1, 7, 19, 20, 46, 54] design more complicated bridge networks to compress or adaptively select visual information. Despite their impressive performance on VL tasks, there is still a lack of exploration on the effectiveness and limitation of the visual branch in a MLLM. Recently, Wang et al. [41] empirically investigate factors contributing to the formation of an effective vision encoder in a MLLM from the perspective of pretraining. Differently, our work explores the effect of region-level VL tasks on the visual understanding abilities of the MLLM, and incorporates fine-grained and high-level visual knowledge to enrich the visual branch in the MLLM. + +# 2.2. Visual Grounding in the field of MLLMs + +Visual grounding [16, 37, 38] is a region-level VL task that aims to establish a connection between particular regions and their textual descriptors, which plays a vital role in human-machine interaction by enabling referential dialog. In the realm of MLLMs, there are some attempts to enhance MLLMs by leveraging visual grounding tasks. Works like Shikra [4], Kosmos-2 [31], Ferret [47] and Pink [45] demonstrate the promising direction of employing visual grounding datasets to endow MLLMs with region-level visual understanding abilities. They convert existing datasets equipped with spatial coordinates, like Visual Genome [17] and RefCOCO [16], into the textual instruction format and perform instruction tuning on MLLMs. Merely considering the visual grounding task as one of several instruction-tuning tasks, these works fall short in exploring the interactions among various tasks. In contrast, our work investigates the internal conflict between visual grounding tasks and image-level VL tasks (e.g., image captioning and VQA), and proposes a stage-wise instruction-tuning strategy to address this issue, achieving a good balance between these two kinds of VL tasks. + +![](images/0f15f4f211ae4f508976827313469a9b06d82c85dab24734b654a7e1e6bd7722.jpg) +Figure 2. Overview of the proposed LION. The model extracts holistic visual features from Q-Former, and combines them with fine-grained spatial-aware visual features from the vision aggregator. The Mixture-of-Adapters with a router in the frozen LLM dynamically fuses visual knowledge learned from different visual branches and LLM adapters based on the task types (image-level and region-level). + +# 3. LION + +In this section, we present the dual-Level vVisual knOwlEdge eNhanced multimodal large language model (LION). The whole framework is depicted in Fig. 2. + +# 3.1. Progressive Incorporation of Fine-grained Spatial-Aware Visual Knowledge + +# 3.1.1 Reorganizing Visual Grounding Tasks + +To incorporate fine-grained spatial-aware visual knowledge into MLLMs, we make use of region-level VL tasks, i.e., visual grounding, and meticulously process the data with spatial coordinates in a unified format for instruction-tuning MLLM. Visual grounding requires the model to generate or comprehend natural language expressions referring to particular objects or regions within an image, e.g., "a man with glasses". Referring to objects or regions in complex images needs an ability of precisely comprehending fine-grained visual information. Current MLLMs lack such referring comprehending, as they mainly target a coarse alignment of VL modalities when pretrained on massive image-text pairs [2, 41]. In this regard, we introduce visual grounding tasks as a kind of region-level VL tasks for the instruction-tuning of MLLMs. This aims to endow the model with fine-grained visual understanding ability such that better performance on image-level VL tasks (e.g., image captioning and VQA) might be achieved. + +We adopt two types of visual grounding tasks, including referring expression comprehension (REC) and referring expression generation (REG) [49]. We use the Visual Genome dataset [18], which associates a local area with one short description, to construct REC/REG tasks. The templates used to organize the Visual Genome dataset in a unified instruction-tuning format can be found in Appendix. + +One core point in reorganizing visual grounding tasks is the way of processing positions. Normally, the position of + +![](images/0f9815cbd12f480ae7979f32f606b23f56df641c302a2219d716a987a1974bb8.jpg) +Stage 1: Instruction-Tuning on Image-Level VL Tasks + +![](images/89e9b079c5ec333cc26e375d96290271742a3f4d0386436a3f53d50fbec52484.jpg) +Stage 2: Instruction-Tuning on Region-Level VL Tasks + +![](images/79b5cd8265c60c8c592fb8ed74c218c9e43c3fc2cadf1fbaa47747cf21d504e1.jpg) +Stage 3: Instruction-Tuning on Image-Level and Region-Level VL Tasks +Figure 3. The stage-wise instruction-tuning strategy. **Stage 1:** We instruction-tune Q-Former and the image-level adapter on image-level VL tasks. **Stage 2:** We instruction-tune the vision aggregator (VA), MLP, and the region-level adapter on region-level VL tasks. **Stage 3:** The Mixture-of-Adapters is devised to form a unified model for instruction-tuning on both kinds of VL tasks. + +an object phrase is presented in the format of bounding box $[x_{min},y_{min},x_{max},y_{max}]$ . We use a natural language style to describe object positions along with the square brackets. A sample in the REC task is displayed as follows: "How can I locate a glass of beer in the image? Please provide the coordinates. Answer: [0.525, 0.0, 0.675, 0.394]". + +# 3.1.2 The Stage-Wise Instruction-tuning Strategy + +To facilitate MLLMs with fine-grained spatial-aware knowledge, the most intuitive way is to directly instruction-tune MLLMs with both image-level and region-level VL tasks in one stage. However, this single-stage instruction-tuning strategy is sub-optimal, and suffers from the internal conflict between these two kinds of VL tasks. We summarize two main issues leading to the internal conflict. 1) One is the need of region-level modality-alignment pretraining. In concurrent works that integrate visual grounding ability, pretraining on the region-level multimodal datasets including visual grounding is a crucial step. Some works + +[31, 45, 47] elaborately create very large visual grounding datasets (e.g., GRIT-20M in kosmos-2 [31]) to advance MLLM in fine-grained perception and understanding. The single-stage instruction-tuning makes it challenging to adapt visual representations learned for image-level alignment to region-level VL tasks under a limited training configuration. 2) Another is the gap between the input-output modes of image-level VL tasks and region-level visual grounding tasks. The latter additionally requires MLLMs to understand specific phrases (in the format “ $[x_{min}, y_{min}, x_{max}, y_{max}]$ ” about the positions of objects. They are semantically distinct from natural languages used in image-level tasks. This requirement necessitates the tuning of the LLM to adapt to region-level tasks, but may disrupt the internal state of the LLM suitable for image-level VL tasks. To address the above issues, we devise a stage-wise instruction-tuning strategy and mixture-of-adapters with a router. + +The stage-wise instruction-tuning strategy is proposed to alleviate the internal conflict between image-level and region-level VL tasks during instruction-tuning. It is composed of three stages for instruction-tuning on image-level, region-level VL tasks and both, respectively, which is depicted in Fig. 3. In stage 1, we follow instructBLIP [7] and fine-tune Q-Former and the image-level adapter [5] in the LLM on image-level VL tasks, such as image captioning and VQA. In stage 2, we propose a vision aggregator for better capturing visual features in fine-grained understanding, which will be introduced later, and tune it with MLP and the region-level adapter on region-level VL tasks. The independent training in the first two stages greatly fulfills the requirements of sufficiently learning both image-level and region-level tasks, providing a solid foundation for subsequent joint training. + +Mixture-of-Adapters with a Router. In stage 3 of our stage-wise instruction-tuning, we need a unified model but encounter a situation where adapters of LLM in stages 1 and 2 are different and suit distinct input-output modes. Inspired by Mixture-of-Experts, we treat each adapter as an expert, and propose a router module to avoid the potential interference between them, as depicted in Fig. 2. + +An adapter [5] is inserted at each FFN layer in a parallel manner. Assuming $X \in \mathcal{R}^{L \times D}$ is the hidden representations generated by a self-attention / causal attention layer, the output representations after FFN (represented as $\mathbf{F}$ ) with the adapter (denoted by $\mathbf{H}$ ) layer are formulated as, + +$$ +O = \mathbf {F} (X) + \mathbf {H} (X), \tag {1} +$$ + +$$ +\mathbf {H} (X) = W _ {u} \left(\sigma \left(W _ {d} X\right)\right), \tag {2} +$$ + +where $\sigma$ is a non-linear function, ReLU. Our router module + +![](images/6ac934367e3a1fe75439f6fc5bfb232e64fe29629a0cbe6d907cff9b8ba2bd45.jpg) +Figure 4. Instruction template with soft prompt. We use a well-designed instruction template with trainable soft prompts to inject the image tags generated by the RAM model into LION. + +aims to dynamically aggregate the hidden features from the main branches and the multiple adapter branches according to task types. Given a set of adapters $\{\mathbf{H}_1,\dots ,\mathbf{H}_K\}$ , each kind of task $t$ defines a specific router function $\mathbf{R}^t$ to generate new hidden features, which can be formulated as, + +$$ +O ^ {t} = \mathbf {F} (X) + \sum_ {k = 1} ^ {K} \mathrm {G} _ {k} ^ {t} \odot \mathbf {H} _ {k} (X). \tag {3} +$$ + +where $\mathbf{G}_k^t\in \mathcal{R}^D$ is a trainable vector that modulates the hidden features from each adapter and makes them suitable for the target task. In practice, we define two types of tasks, one for image-level VL tasks (image captioning and VQA), the other for fine-grained VL tasks (visual grounding). Compared to directly incorporating multiple adapters, the router module provides a better way to maximize the complementarity of image-level and region-level tasks. + +We use the standard language modeling loss in all instruction-tuning stages. In the experiments, we demonstrate that stage-wise training is superior to single-stage training, and ensures a good balance between high-level and fine-grained visual understanding capabilities, further achieves a significant mutual promotion between image-level and region-level VL tasks. + +# 3.1.3 Vision Aggregator + +To extract more sufficient visual details from input images, we devise a vision aggregator that integrates multi-level hidden features of the pretrained visual encoder. Although the vision encoder has a global reception field in all layers, it is verified that different transformer layers learn visual information at different scales [9], e.g., lower layers learn visual details. Thus, our vision aggregator makes fine-grained spatial-aware visual knowledge more likely to be learned based on visual grounding tasks. Specifically, our vision aggregator can be regarded as a tiny transformer-style network, consisting of two transformer layers for aggregating the hidden features from the vision encoder. Given the hidden features $\{V_{i}, V_{j}, V_{k}\}$ from some middle layers in the vision encoder, the vision aggregation module uses two blocks to sequentially integrate the former two features with the last feature. Each block $\mathcal{B}$ is composed of self attention + +Table 1. Comparison on image captioning and VQA. “†” denotes including in-house data that are publicly inaccessible. “*” means our evaluated results by using publicly released checkpoints, which are only for reference as official evaluation settings are incomplete. We report CIDEr score for Flickr30K, COOCap, and TextCaps, Mean Reciprocal Rank (MRR) for Visual Dialog (VisDial), and top-1 accuracy for others. The best and second performances for each benchmark are indicated in bold and underline, respectively. + +
ModelFlickr30KCOCOCapTextCapsOKVQAAOKVQAGQAIconQAVSRVisDial
Flamingo-3B [1]60.6073.00------46.10
Flamingo-9B [1]61.5079.40-44.70----48.00
Kosmos-1 [13]67.1084.70-------
Kosmos-2 [31]80.50--------
AdapterV2 [11]-122.20-------
Shikra [4]73.90117.50-47.16-----
Pink [45]---59.50-52.6047.8066.30-
MiniGPT4 [54]17.75*17.04*24.06*37.5034.51*30.8037.6041.6016.52*
LLaVA [26]48.03*73.85*45.54*54.4034.51*41.3043.0051.208.65*
MiniGPTV2 [3]80.75*129.16*80.60*56.90-60.3047.7060.608.47*
BLIVA [12]87.10-----44.8862.2045.63
InstructBLIP† (T5XL) [7]84.50138.2182.5549.2857.8648.4050.0064.8046.60
InstructBLIP† (T5XXL) [7]83.50138.2882.5348.5956.1647.9051.2065.6048.50
InstructBLIP (T5XL) [7]83.71135.47104.1747.3856.1246.3452.4769.9348.75
InstructBLIP (T5XXL) [7]85.79138.63105.4453.0259.3847.7453.1868.4650.41
LION-4B85.57138.20104.8751.0859.9849.5054.9172.9650.02
LION-12B87.12139.25108.7657.3360.8751.5654.8973.7750.42
+ +(Attn), cross attention (XAttn), and Feed-forward network (FFN) arranged in a sequential manner. Finally, the output features $\bar{V}$ is generated as follows, + +$$ +\bar {V} = \mathcal {B} _ {2} \left(\mathcal {B} _ {1} \left(V _ {i}; V _ {j}\right); V _ {k}\right), \tag {4} +$$ + +$$ +\mathcal {B} (X; Y) = \operatorname {F F N} (\operatorname {X A t t n} (\operatorname {A t t n} (X), Y)). \tag {5} +$$ + +In practice, we use the middle layers $\{i = L - 1, j = 2L/3, k = L/3\}$ in the vision encoder to produce the hidden features as the input to VA, where $L$ is the number of layers in the vision encoder. + +# 3.2. Soft Prompting of High-Level Semantic Visual Evidence + +The vision encoder in a MLLM may be insufficient in comprehensively extracting visual information required by complex multi-modal tasks, although it has been trained on large-scale image-text pairs. It has been demonstrated that increasing the amount and quality of pretraining multimodal datasets can significantly improve the visual understanding capability of MLLMs [41], but inevitably induces prohibitive computational overhead. An appealing alternative is to harness the convenient and powerful off-the-shelf vision models to capture various aspects of visual content within an image as a supplement. + +We choose the recognize anything model (RAM) [52] as an off-the-shelf vision model to provide diverse tags, encompassing objects, scenes, actions, and attributes, as visual evidence to support comprehensive visual perception. Instead of directly adding tags in the instruction, we design a soft prompting method to guide the model to adaptively use the inserted tags in order to avoid the potential negative influence caused by the imperfect predictions from RAM. + +In Fig. 4, we present the instruction template of tags along with the soft prompt that is a trainable vector. Our soft prompting approach can be regarded as a kind of prompt tuning methods, which guides the model toward the right direction. In standard prompt tuning works, the right direction is directly formulated as the optimization for task goals. In our work, the right direction is specified by a tailored sentence, "According to , you are allowed to use or partially use the following tags: ", and "" will be replaced by the soft prompt. Our soft prompting method for inserting tags has some distinct properties. It is designed to adaptively select valuable information from tags, rather than serving a specific task, as seen in standard prompt tuning methods. Our method directly uses the output labels from a small off-the-shelf vision model to incorporate high-level semantic visual evidence into a MLLM, so as to eliminate extra computational overhead of the feature alignment. + +# 4. Experiments + +In this section, we conduct extensive experiments to demonstrate the effectiveness of our model along with the quantitative and qualitative analyses. Please refer to Appendix for implementation details and training details. + +# 4.1. Evaluations on Image-Level VL Tasks + +Here, we evaluate multi-modal understanding abilities of LION on two kinds of image-level VL tasks, image captioning and VQA. Image captioning requires the model to generate a text description of the input image. We use COCO caption [6], TextCaps [39] and Flickr30K [48] as benchmarks, and report CIDEr as the evaluation metric. We utilize greedy search for caption generation. VQA provides an image along with a specific question for the model, asking + +Table 2. Comparison on REC. "Avg." means the average of top-1 accuracy over all the 8 evaluation sets. + +
ModelRefCOCORefCOCO+RefCOCOgAvg.
valtest-Atest-Bvaltest-Atest-Bvaltest
Zero-shot Setting
Kosmos-2 [31]52.3257.4247.2645.4850.7342.2460.5761.6552.21
GRILL [15]-------47.50-
Pink [45]54.1061.2044.2043.9050.7035.0059.1060.1051.00
LION-4B57.8956.0758.4046.3845.2947.5064.7463.5654.98
LION-12B58.5456.4159.3645.9345.7347.8966.1264.6955.58
Fune-tuning Setting
OFA-L [42]79.9683.6776.3968.2976.0061.7567.5767.5872.65
VisionLLM-H [43]-86.70-------
Shikra-7B [4]87.0190.6180.2481.6087.3672.1282.2782.1982.93
Shikra-13B [4]87.8391.1181.8182.8987.7974.4182.6483.1683.96
Pink [45]88.3091.7084.0081.4087.5073.7083.7083.7084.25
Ferret-7B [47]87.4991.3582.4580.7887.3873.1483.9384.7683.91
Ferret-13B [47]89.4892.4184.3682.8188.1475.1785.8386.3485.57
MiniGPTv2 [3]88.6991.6585.3379.9785.1274.4584.4484.6684.29
LION-4B89.7392.2984.8283.6088.7277.3485.6985.6385.98
LION-12B89.8093.0285.5783.9589.2278.0685.5285.7486.36
+ +Table 3. The comparison of various strategies in the instruction-tuning period. "REC Avg." represents the average score of all REC tasks. "Held-in" denotes the average score of COOCap, TextCaps, OKVQA, and AOKVQA, while "Held-out" means the average score of Flickr30K, GQA, IconQA, VSR, and VisDial. + +
StrategyImage-LevelRegion-Level
Held-inHeld-outREC Avg.
Single Stage84.7960.203.78
Stage-wise88.0761.9154.46
w/ Router87.6762.1654.98
+ +for the output as an answer. We evaluate LION on six VQA datasets, including OKVQA [30], AOKVQA [34], GQA [14], IconQA [28], Visual Spatial Reasoning [24], and Visual Dialog (VisDial) [8]. For OKVQA, A-OKVQA, and GQA, we employ an open-ended generation with a greedy decoding strategy, For IconQA, Visual Spatial Reasoning, and Visual Dialog, we match the output with various candidates, and select the candidate with the highest value as the prediction. We report Mean Reciprocal Rank (MRR) for Visual Dialog, and top-1 accuracy for other VQA tasks. The detailed descriptions of these datasets and inference instructions are presented in Appendix. + +As shown in Table 1, LION achieves the best performance across 7 out of 9 benchmarks, and the second on OKVQA. LION shares the same training datasets with InstructBLIP, except Visual Genome dataset adopted in our work and the in-house dataset, WebCapFilt, used in InstructBLIP. The amount of Visual Genome dataset (3.6M) is far smaller than WebCapFilt (14M). Compared to the original InstructBLIP trained with WebCapFilt, LION exhibits superior performances on all zero-shot evaluation benchmarks, showcasing a better generalization ability. We also re-implemented InstructBLIP on the same instruction-tuning datasets. The comparison shows the consistent and significant improvements of LION over the re-implemented InstructBLIP, demonstrating the effectiveness of incorpo + +rating dual-level visual knowledge. Shikra, MiniGPTV2 and Pink are also integrated with visual grounding abilities. Their inferior results on most image-level VL tasks exhibit that our proposed stage-wise instruction-tuning strategy and soft prompting of high-level semantic knowledge help enhance holistic visual understanding abilities of MLLMs. + +# 4.2. Evaluations on Region-Level VL Tasks + +To assess the fine-grained perceptual and reasoning abilities of LION, we evaluate it on three REC datasets, RefCOCO [16], RefCOCO+ [16], RefCOCOg [29]. REC requires the model to locate the target object given a referring expression. We follow the standard setting, and use accuracy as an evaluation metric, which means it is correct when the IOU between prediction and ground-truth is no less than 0.5. + +In Table 2, we show the comparison between LION and other MLLMs with respect to the grounding abilities, under the settings of zero-shot and fine-tuning evaluations, respectively. In the zero-shot evaluation setting, we directly employ LION to generate coordinates of referring expressions on three datasets. Our model shows significant improvements on most evaluation sets over Kosmos-2 and Pink, except test-A sets of RefCOCO and RefCOCO+. The languages used in RefCOCOg are more flowery than those used in RefCOCO and RefCOCO+. The significant improvements on RefCOCOg clearly demonstrate that LION can handle complex referring expressions and has superior zero-shot spatial-aware visual understanding abilities. + +In the fine-tuning setting, we fine-tune LION with training samples from three REC datasets. As shown in Table 2, our model achieves the best performance on average and on most of the evaluation sets, indicating the advanced fine-grained perception ability of our model. Ferret proposes a spatial-aware visual sampler to handle free-form referred regions, and meticulously constructs an extensive grounding dataset with lots of efforts on data generation and fil + +Table 4. Ablation studies of dual-level visual knowledge. "VG" means visual grounding tasks. "Held-in" and "Held-out" denote the training images of tasks are seen and unseen, respectively. "REC Avg." means the average score of all REC tasks. + +
ComponentsHeld-inHeld-outREC Avg.
VGTagsCOCOCapTextCapsOKVQAAOKVQAFlickr30KGQAVSRIconQAVisDial
××135.47104.1747.3856.1283.7146.3469.9352.4748.75-
✓×137.87104.8451.0756.9083.9949.2273.2054.6749.7054.98
✓✓138.20104.8751.0859.9885.5749.5072.9654.9150.0254.92
+ +Table 5. Evaluation of object hallucination on POPE benchmark. F1 score is the major metric for hallucication evaluation. + +
DatasetsMetricsLIONShikra [4]InstructBLIP [7]MiniGPT-4 [54]LLaVA [26]mPLUG-Owl [46]
RandomF1-Score ↑88.3386.1989.2780.1766.6468.39
Accuracy ↑88.9786.9088.5779.6750.3753.97
Precision ↑97.1294.4084.0978.2450.1952.07
Recall ↑81.0079.2795.1382.2099.1399.60
PopularF1-Score ↑85.9483.1684.6673.0266.4466.94
Accuracy ↑86.7783.9782.7769.7349.8750.90
Precision ↑91.6987.5576.2765.8649.9350.46
Recall ↑80.8779.2095.1381.9399.2799.40
AdversarialF1-score ↑84.7182.4977.3270.4266.3266.82
Accuracy ↑85.3783.1072.1065.1749.7050.67
Precision ↑88.6985.6065.1361.1949.8550.34
Recall ↑81.0779.6095.1382.9399.0799.33
+ +tering. However, LION can achieve superior performances compared to Ferret by using a simple vision aggregator and existing datasets, implying the effectiveness of fine-grained visual knowledge. + +# 4.3. Ablation Study + +The effect of vision aggregator. We conduct an ablation study of the vision aggregator with only visual grounding tasks on LION-4B during stage 2 of the stage-wise instruction-tuning. As illustrated in Fig. 5, the removal of the vision aggregator degrades REC performances, validating that aggregating multi-level vision features promotes the extraction of fine-grained spatial-aware visual knowledge. + +Stage-wise instruction-tuning mitigates the conflict between image-level and region-level tasks. We investigate the performance of two types of VL tasks under three instruction-tuning strategies, i.e., single stage, stage-wise, and stage-wise with a router. As shown in Table 3, the stage-wise instruction-tuning strategy shows a significant improvement on the average REC performance, which is completely damaged in the single stage instruction-tuning. The worse REC performance of the single stage strategy can be attributed to the lack of pretraining on large-scale grounding datasets, like in Kosmos-2 and Shikra, and the gap of their input-output modes. To address these challenges, stage-wise training progressively integrates fine-grained spatial-aware knowledge from visual grounding datasets by splitting the whole instruction-tuning process into three stages. The model can sufficiently learn diverse levels of visual knowledge in separate training stages (stages 1 and 2), and incorporate them in the final training stage (stage 3 in Fig. 3). This contributes to the performance improvements of all VL tasks. Furthermore, stage-wise instruction + +![](images/d3ad7060d434ab9934c0312c1c662441e306dba7b6aaebc8f13269f70c8e1932.jpg) +Figure 5. The effect of the vision aggregator. The results on RefCOCO, RefCOCO+, and RefCOCOg clearly show that the proposed module can overall improve REC performances across 8 evaluation sets. + +tuning with the router improves the held-out and REC performance, with a slight degradation in the held-in performance. All these results demonstrate LION's ability to handle the potential conflict of various VL tasks and maximize the learning benefit. + +Dual-level visual knowledge enhances multimodal understanding abilities of MLLMs. We evaluate the performance of our model integrated with different levels of visual knowledge on various benchmarks in Table 4. It can be seen that dual-level visual knowledge can upgrade the performance of all VL tasks to varying degrees. When progressively incorporating fine-grained spatial-aware knowledge, the performances of four tasks, OKVQA, GQA, IconQA, and Visual Spatial Reasoning, are significantly improved, as they highly require region-level understanding and spatial reasoning. When inserting tags as high-level visual evidence, we can see substantial performance increases on Flickr30K and AOKVQA, which demand more comprehensive semantic knowledge than other tasks, like COCO caption and OKVQA. + +Table 6. Evaluation on MMBench test set, all the reported results of compared models are from the leadboard of MMBench. + +
ModelsText EncoderVision EncoderOverallLRARRRFP-SFP-CCP
MiniGPT-4 [54]Vincuna-7BEVA-G23.013.632.98.928.811.228.3
PandaGPT [40]Vincuna-13BImageBind ViT-H/1442.523.161.534.132.728.757.6
VisualGLM [10]ChatGLM-6BEVA-CLIP33.511.448.827.735.817.641.5
InstructBLIP [7]Vicuna-7BEVA-G33.921.647.422.533.024.441.1
LLaVA-v1.5 [25]Vicuna-v1.5-7BCLIP ViT-L/1459.532.472.649.362.352.267.7
Otter-I [19]LLaMA-7BCLIP ViT-L/1448.322.263.339.446.836.460.6
Shikra [4]Vincuna-7BCLIP ViT-L/1460.233.569.653.161.850.471.7
LMEye [23]FlanT5-XLCLIP ViT-L/1462.641.074.355.961.658.769.2
MMICL [53]FlanT5-XXLEVA-G65.244.377.964.866.553.670.6
mPLUG-Owl2 [44]LLaMA-7BCLIP ViT-L/1466.043.476.062.168.655.973.0
LIONFlanT5-XXLEVA-G73.451.784.178.474.060.878.9
+ +Image Captioning +![](images/8c65a703fac70fcc7476868190635020af76a6be15a63a48b683f8d30bffd709.jpg) +User What is in the photo? + +Visual Question Answering +![](images/a9f6248c60773c4ebd8f5e7ca12f0b922865f97797afb22e2d5412026c3c7190.jpg) +User On which side of the photo are the clay.pots? + +![](images/6ad3068a5701dcd8a9fb015a02ff704f44041cb91dde011c020ce7dee9b956e3.jpg) +InstructBLIP boy in karate uniform with white belt and black belts watching him do a kick in a martial arts class at a karate club in california. +The pots are on the left of the image. + +Figure 6. Qualitative comparison of InstructBLIP, Shikra, and LION. We mark the hallucination or incorrect part in red, and highlight the correct part in green for comparison. These samples exhibit that LION is able to achieve superior fine-grained understanding and visual spatial reasoning capabilities with fewer hallucinated responses. +![](images/03570a4610050db94adf21ec54a5588d37bb7df5925e2dbd2d69b088e566f70d.jpg) +LION A young boy wearing a white karate uniform is performing a kick in front of a yellow wall, with the word "Kwon" written on the wall behind him. +The pots are on the right of the image. + +Referring Expression Comprehension +![](images/66d98fb325214e17f61bf05d6ee4eab6d4a5436288b9fbfb1cf2a3a427cf62c6.jpg) +User Help me locate the coordinates of a brown bear near a soda bottle in the image. +Shikra The brown bear near a soda bottle is located at [0.002, 0.248, 0.416, 0.760] +LION The brown bear near a soda bottle is located at [0.550, 0.387, 0.920, 0.867] + +# 4.4. Evaluations on Object Hallucination and MM-Bench + +Li et al. [22] present an open-sourced evaluation benchmark, called POPE, to evaluate the object hallucination [33]. We follow the POPE evaluation pipeline to inspect LION. The results in Table 5 show that LION has superior results, especially under popular and adversarial settings, which means that incorporating fine-grained and high-level semantic visual knowledge into MLLM can mitigate the object hallucination to some degree. To comprehensively validate the effectiveness of our method, we further evaluate LION on MMBench [27]. The results are summarized in Table 6. Our strong performances across various skills demonstrate that the progressive incorporation of fine-grained knowledge significantly alleviates the hallucination phenomenon in MLLMs. + +# 4.5. Qualitative Analysis + +As shown in Fig. 6, we depict various examples to validate the advanced perceptual and reasoning abilities of LION. The left example exhibits our superior fine-grained understanding capability to correctly generate the right attributes "white", "yellow" and the character "Kwon". The middle + +example validates the advantage of our model in visual spatial reasoning. The right example shows that LION accurately localizes the referring object, while Shikra provides an incorrect response caused by the misunderstanding of fine-grained details "a soda bottle". + +# 5. Conclusion + +To address the insufficient extraction and reasoning of visual information in MLLMs, we propose LION to exploit dual-level visual knowledge, i.e., fine-grained spatial-aware visual knowledge and high-level semantic visual evidence. To mitigate the internal conflict among tasks, LION adopts a stage-wise instruction-tuning strategy. Extensive experiments validate the superiority of LION in image captioning, VQA, and visual grounding tasks. + +# Acknowledgements + +This study is supported by Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), National Natural Science Foundation of China (Grant No. 62236003 and 62306090), China Postdoctoral Science Foundation (Grant No. 2023TQ0095 and 2023M740923). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5 +[2] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 3 +[3] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.5.6 +[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8 +[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 4 +[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 5 +[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 1, 2, 4, 5, 7, 8 +[8] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 6 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 8 +[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 5 + +[12] Wenbo Hu, Yifan Xu, Y Li, W Li, Z Chen, and Z Tu. Bliva: A simple multimodal llm for better handling of text-rich visual questions. arXiv preprint arXiv:2308.09936, 2023. 5 +[13] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 5 +[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6 +[15] Woojeong Jin, Subhabrata Mukherjee, Yu Cheng, Yelong Shen, Weizhu Chen, Ahmed Hassan Awadallah, Damien Jose, and Xiang Ren. Grill: Grounded vision-language pretraining via aligning text and image regions. arXiv preprint arXiv:2305.14676, 2023. 6 +[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 6 +[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2 +[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3 +[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 8 +[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2 +[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1 +[22] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 8 +[23] Yunxin Li, Baotian Hu, Xinyu Chen, Lin Ma, and Min Zhang. Lmeye: An interactive perception network for large language models. arXiv preprint arXiv:2305.03701, 2023.8 +[24] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 6 + +[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 8 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5, 7 +[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8 +[28] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021. 6 +[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6 +[30] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 6 +[31] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 4, 5, 6 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[33] Vipula Rawte, Amit Sheth, and Amitava Das. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023. 8 +[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 6 +[35] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In CVPR, 2019. 1 +[36] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense with clean-adversarial mutual learning. International Journal of Computer Vision, 2022. 1 +[37] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In CVPR, 2023. 2 +[38] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 + +[39] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5 +[40] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8 +[41] Guangzhi Wang, Yixiao Ge, Xiaohan Ding, Mohan Kankanhalli, and Ying Shan. What makes for good visual tokenizers for large language models? arXiv preprint arXiv:2305.12223, 2023. 1, 2, 3, 5 +[42] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 6 +[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 6 +[44] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 8 +[45] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2, 4, 5, 6 +[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 7 +[47] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4, 6 +[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5 +[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 3 +[50] Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. Investigating the catastrophic forgetting in multimodal large language models. arXiv preprint arXiv:2309.10313, 2023. 1 +[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. + +Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023.2 +[52] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 5 +[53] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 8 +[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 5, 7, 8 \ No newline at end of file diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/images.zip b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..044f975840803b1f10ce7e405bb96fb8e7be54a3 --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2240272691a791004a907d5ebac42a6916c3bde548b130f21231c25cc99bb8df +size 699187 diff --git a/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/layout.json b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f62e43b0cc83b4be60779e4cbaacc13a7053a25 --- /dev/null +++ b/2024/LION_ Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge/layout.json @@ -0,0 +1,7302 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 120, + 103, + 474, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 103, + 474, + 140 + ], + "spans": [ + { + "bbox": [ + 120, + 103, + 474, + 140 + ], + "type": "text", + "content": "LION : Empowering Multimodal Large Language Model with Dual-Level Visual Knowledge" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "text", + "content": "Gongwei Chen, Leyang Shen, Rui Shao," + }, + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "text", + "content": " Xiang Deng, Liqiang Nie" + }, + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 84, + 160, + 509, + 217 + ], + "type": "text", + "content": " \nSchool of Computer Science and Technology, Harbin Institute of Technology, Shenzhen \n{chengongwei, shaorui, dengxiang, nieliqiang}@hit.edu.cn \nhttps://github.com/rshaojimmy/JiuTian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "text", + "content": "Multimodal Large Language Models (MLLMs) have endowed LLMs with the ability to perceive and understand multi-modal signals. However, most of the existing MLLMs mainly adopt vision encoders pretrained on coarsely aligned image-text pairs, leading to insufficient extraction and reasoning of visual knowledge. To address this issue, we devise a dual-Level vVisual knOwledge eNhanced Multimodal Large Language Model (LION), which empowers the MLLM by injecting visual knowledge in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. We design a vision aggregator cooperated with region-level vision-language (VL) tasks to incorporate fine-grained spatial-aware visual knowledge into the MLLM. To alleviate the conflict between image-level and region-level VL tasks during incorporation, we devise a dedicated stage-wise instruction-tuning strategy with mixture-of-adapters. This progressive incorporation scheme contributes to the mutual promotion between these two kinds of VL tasks. 2) Soft prompting of high-level semantic visual evidence. We facilitate the MLLM with high-level semantic visual evidence by leveraging diverse image tags. To mitigate the potential influence caused by imperfect predicted tags, we propose a soft prompting method by embedding a learnable token into the tailored text instruction. Comprehensive experiments on several multi-modal benchmarks demonstrate the superiority of our model (e.g., improvement of " + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "text", + "content": " accuracy on VSR and " + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "text", + "content": " CIDEr on TextCaps over InstructBLIP, " + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 270, + 289, + 619 + ], + "type": "text", + "content": " accuracy on RefCOCOg over Kosmos-2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 128, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 643, + 287, + 693 + ], + "type": "text", + "content": "Recently, Large Language Models (LLMs) have demonstrated remarkable zero-shot abilities on various linguistic tasks. Assisted by LLMs, several multimodal large language models (MLLMs), such as MiniGPT-4 [54], Otter" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 245, + 545, + 319 + ], + "blocks": [ + { + "bbox": [ + 308, + 245, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 308, + 245, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 245, + 545, + 319 + ], + "type": "image", + "image_path": "0197e60522e17f12efd32fb03d4677193d011df1a7b4135044b08cd008e6bda8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 320, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 307, + 320, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 407 + ], + "type": "image", + "image_path": "172da6b0638d7c17d4b19ff0e6adb63651d6035da34cc85e3cb5489aae5a1d71.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 413, + 545, + 470 + ], + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 470 + ], + "type": "text", + "content": "Figure 1. Comparison between existing MLLMs and LION. The existing MLLM generates a vague and inaccurate response, while LION provides a more precise and contextually accurate description by progressively incorporating spatial-aware knowledge and softly prompting semantic visual evidence." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 473, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 509 + ], + "type": "text", + "content": "[19], and InstructBLIP [7], achieve significant improvements in reasoning abilities to deal with various vision-language (VL) tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 689 + ], + "type": "text", + "content": "In most of the existing MLLMs, the visual information is mainly extracted from a vision encoder pretrained with image-level supervision (e.g., CLIP [32]), and then are adapted to a LLM by using a tiny bridge module. This makes these MLLMs inherently possess limited image understanding capabilities [21]. As shown in Fig. 1, the insufficient visual information misleads MLLMs to provide erroneous and hallucinated responses. An intuitive solution to this problem is to replace or tune the vision encoder [41]. However, it requires pretraining on massive data or suffers from the catastrophic forgetting issue [35, 36, 50], which diminishes the practical efficacy of this strategy. These predicaments highlight that the insufficient extraction of visual knowledge has become a central obstacle impeding the development of MLLMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "To overcome this dilemma, as depicted in Fig. 1, we devise a dual-Level vVisual knOwledge eNhanced Multimodal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 135, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26540" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 467 + ], + "type": "text", + "content": "Large Language Model (LION), which enriches the visual information in MLLMs in two levels. 1) Progressive incorporation of fine-grained spatial-aware visual knowledge. LION enhances the MLLM with more fine-grained perceptual abilities by studying the region-level VL tasks involving the spatial coordinates. However, we find that simply training on region-level and the original image-level VL tasks* simultaneously hurts the general performances of the MLLM due to the conflicts between these two kinds of tasks. To address this issue, we propose a novel stage-wise instruction-tuning strategy to perform image-level and region-level VL tasks separately with two different visual branches and task adapters. In addition, we devise mixture-of-adapters with a router to dynamically fuse visual information across various granularities in a unified MLLM. This progressive incorporation of fine-grained visual knowledge contributes to the mutual promotion between these two kinds of VL tasks, and spawns LION to excel in capturing fine-grained visual information and performing spatial reasoning, as shown in Fig. 1. 2) Soft prompting of high-level semantic visual evidence. Alongside the improvement of MLLMs in fine-grained perceptual capabilities, there is also an opportunity to enhance their high-level semantic understanding. LION uses an off-the-shelf vision model to extract high-level semantic knowledge, i.e., image tags, as supplementary information for the MLLM. However, as off-the-shelf vision models are typically not flawless, errors in tag predictions are inevitable. Inspired by prompt tuning, we propose a soft prompting method to mitigate the potential negative influence resulting from the imperfect predicted tags. As shown in Fig. 1, injection of semantic visual evidence alleviates the hallucination issue substantially." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 468, + 266, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 468, + 266, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 468, + 266, + 479 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 480, + 287, + 683 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 480, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 287, + 563 + ], + "type": "text", + "content": "- To address the internal conflict between region-level and image-level VL tasks, we propose a progressive incorporation of fine-grained spatial-aware visual knowledge with a novel stage-wise instruction-tuning strategy. It achieves mutual promotion between two kinds of VL tasks and equips LION with advanced holistic and fine-grained visual perceptual abilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 564, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 287, + 634 + ], + "type": "text", + "content": "- As a powerful complement, we propose to integrate image tags as high-level semantic visual evidence into MLLMs, and design a soft prompting method to alleviate the bad influence from incorrect tags. This mitigates the hallucination issue and showcases positive effects on various VL tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 683 + ], + "type": "text", + "content": "- We evaluate LION on a wide range of VL tasks, including image captioning, visual question answering (VQA), and visual grounding, and demonstrate its superiority over the baselines. LION outperforms InstructBLIP by around " + }, + { + "bbox": [ + 47, + 635, + 287, + 683 + ], + "type": "inline_equation", + "content": "5\\%" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "text", + "content": "accuracy on VSR, and around " + }, + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "text", + "content": " CIDEr on TextCaps, Kosmos-2 by around " + }, + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 72, + 545, + 133 + ], + "type": "text", + "content": " accuracy on RefCOCOg. The evaluations on POPE and MMBench exhibit the remarkable abilities of LION in alleviating object hallucination and various perceptual dimensions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 392, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 180, + 500, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 180, + 500, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 180, + 500, + 193 + ], + "type": "text", + "content": "2.1. Multimodal Large Language Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 203, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 545, + 418 + ], + "type": "text", + "content": "Building on the success of LLMs, many researches have emerged to extend them to multimodal tasks, especially VL tasks. The common pipeline uses a vision model to transform the image into visual features, followed by a bridge module to align them with the feature space of LLMs. Some works [4, 11, 26, 31, 51] directly use a linear or MLP layer as bridge module, while others [1, 7, 19, 20, 46, 54] design more complicated bridge networks to compress or adaptively select visual information. Despite their impressive performance on VL tasks, there is still a lack of exploration on the effectiveness and limitation of the visual branch in a MLLM. Recently, Wang et al. [41] empirically investigate factors contributing to the formation of an effective vision encoder in a MLLM from the perspective of pretraining. Differently, our work explores the effect of region-level VL tasks on the visual understanding abilities of the MLLM, and incorporates fine-grained and high-level visual knowledge to enrich the visual branch in the MLLM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 439, + 518, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 518, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 518, + 451 + ], + "type": "text", + "content": "2.2. Visual Grounding in the field of MLLMs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Visual grounding [16, 37, 38] is a region-level VL task that aims to establish a connection between particular regions and their textual descriptors, which plays a vital role in human-machine interaction by enabling referential dialog. In the realm of MLLMs, there are some attempts to enhance MLLMs by leveraging visual grounding tasks. Works like Shikra [4], Kosmos-2 [31], Ferret [47] and Pink [45] demonstrate the promising direction of employing visual grounding datasets to endow MLLMs with region-level visual understanding abilities. They convert existing datasets equipped with spatial coordinates, like Visual Genome [17] and RefCOCO [16], into the textual instruction format and perform instruction tuning on MLLMs. Merely considering the visual grounding task as one of several instruction-tuning tasks, these works fall short in exploring the interactions among various tasks. In contrast, our work investigates the internal conflict between visual grounding tasks and image-level VL tasks (e.g., image captioning and VQA), and proposes a stage-wise instruction-tuning strategy to address this issue, achieving a good balance between these two kinds of VL tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Here, image-level VL tasks denote image captioning and visual question answering, region-level VL tasks mean visual grounding tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26541" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 223 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 223 + ], + "type": "image", + "image_path": "0f15f4f211ae4f508976827313469a9b06d82c85dab24734b654a7e1e6bd7722.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 546, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 546, + 266 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed LION. The model extracts holistic visual features from Q-Former, and combines them with fine-grained spatial-aware visual features from the vision aggregator. The Mixture-of-Adapters with a router in the frozen LLM dynamically fuses visual knowledge learned from different visual branches and LLM adapters based on the task types (image-level and region-level)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 269, + 93, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 93, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 93, + 281 + ], + "type": "text", + "content": "3. LION" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 327 + ], + "type": "text", + "content": "In this section, we present the dual-Level vVisual knOwlEdge eNhanced multimodal large language model (LION). The whole framework is depicted in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 287, + 361 + ], + "type": "text", + "content": "3.1. Progressive Incorporation of Fine-grained Spatial-Aware Visual Knowledge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 368, + 243, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 243, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 243, + 380 + ], + "type": "text", + "content": "3.1.1 Reorganizing Visual Grounding Tasks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 604 + ], + "type": "text", + "content": "To incorporate fine-grained spatial-aware visual knowledge into MLLMs, we make use of region-level VL tasks, i.e., visual grounding, and meticulously process the data with spatial coordinates in a unified format for instruction-tuning MLLM. Visual grounding requires the model to generate or comprehend natural language expressions referring to particular objects or regions within an image, e.g., \"a man with glasses\". Referring to objects or regions in complex images needs an ability of precisely comprehending fine-grained visual information. Current MLLMs lack such referring comprehending, as they mainly target a coarse alignment of VL modalities when pretrained on massive image-text pairs [2, 41]. In this regard, we introduce visual grounding tasks as a kind of region-level VL tasks for the instruction-tuning of MLLMs. This aims to endow the model with fine-grained visual understanding ability such that better performance on image-level VL tasks (e.g., image captioning and VQA) might be achieved." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 689 + ], + "type": "text", + "content": "We adopt two types of visual grounding tasks, including referring expression comprehension (REC) and referring expression generation (REG) [49]. We use the Visual Genome dataset [18], which associates a local area with one short description, to construct REC/REG tasks. The templates used to organize the Visual Genome dataset in a unified instruction-tuning format can be found in Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "One core point in reorganizing visual grounding tasks is the way of processing positions. Normally, the position of" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 279, + 545, + 299 + ], + "blocks": [ + { + "bbox": [ + 307, + 270, + 478, + 278 + ], + "lines": [ + { + "bbox": [ + 307, + 270, + 478, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 478, + 278 + ], + "type": "text", + "content": "Stage 1: Instruction-Tuning on Image-Level VL Tasks" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 279, + 545, + 299 + ], + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 299 + ], + "type": "image", + "image_path": "0f9815cbd12f480ae7979f32f606b23f56df641c302a2219d716a987a1974bb8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 312, + 545, + 349 + ], + "blocks": [ + { + "bbox": [ + 308, + 303, + 481, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 303, + 481, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 303, + 481, + 312 + ], + "type": "text", + "content": "Stage 2: Instruction-Tuning on Region-Level VL Tasks" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 312, + 545, + 349 + ], + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 349 + ], + "type": "image", + "image_path": "89e9b079c5ec333cc26e375d96290271742a3f4d0386436a3f53d50fbec52484.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 361, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 308, + 351, + 535, + 361 + ], + "lines": [ + { + "bbox": [ + 308, + 351, + 535, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 351, + 535, + 361 + ], + "type": "text", + "content": "Stage 3: Instruction-Tuning on Image-Level and Region-Level VL Tasks" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 361, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 397 + ], + "type": "image", + "image_path": "79b5cd8265c60c8c592fb8ed74c218c9e43c3fc2cadf1fbaa47747cf21d504e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 403, + 546, + 469 + ], + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 469 + ], + "type": "text", + "content": "Figure 3. The stage-wise instruction-tuning strategy. **Stage 1:** We instruction-tune Q-Former and the image-level adapter on image-level VL tasks. **Stage 2:** We instruction-tune the vision aggregator (VA), MLP, and the region-level adapter on region-level VL tasks. **Stage 3:** The Mixture-of-Adapters is devised to form a unified model for instruction-tuning on both kinds of VL tasks." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 546 + ], + "type": "text", + "content": "an object phrase is presented in the format of bounding box " + }, + { + "bbox": [ + 304, + 474, + 545, + 546 + ], + "type": "inline_equation", + "content": "[x_{min},y_{min},x_{max},y_{max}]" + }, + { + "bbox": [ + 304, + 474, + 545, + 546 + ], + "type": "text", + "content": ". We use a natural language style to describe object positions along with the square brackets. A sample in the REC task is displayed as follows: \"How can I locate a glass of beer in the image? Please provide the coordinates. Answer: [0.525, 0.0, 0.675, 0.394]\"." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 562, + 525, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 562, + 525, + 574 + ], + "spans": [ + { + "bbox": [ + 305, + 562, + 525, + 574 + ], + "type": "text", + "content": "3.1.2 The Stage-Wise Instruction-tuning Strategy" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "To facilitate MLLMs with fine-grained spatial-aware knowledge, the most intuitive way is to directly instruction-tune MLLMs with both image-level and region-level VL tasks in one stage. However, this single-stage instruction-tuning strategy is sub-optimal, and suffers from the internal conflict between these two kinds of VL tasks. We summarize two main issues leading to the internal conflict. 1) One is the need of region-level modality-alignment pretraining. In concurrent works that integrate visual grounding ability, pretraining on the region-level multimodal datasets including visual grounding is a crucial step. Some works" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26542" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": "[31, 45, 47] elaborately create very large visual grounding datasets (e.g., GRIT-20M in kosmos-2 [31]) to advance MLLM in fine-grained perception and understanding. The single-stage instruction-tuning makes it challenging to adapt visual representations learned for image-level alignment to region-level VL tasks under a limited training configuration. 2) Another is the gap between the input-output modes of image-level VL tasks and region-level visual grounding tasks. The latter additionally requires MLLMs to understand specific phrases (in the format “" + }, + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "[x_{min}, y_{min}, x_{max}, y_{max}]" + }, + { + "bbox": [ + 46, + 72, + 289, + 286 + ], + "type": "text", + "content": "” about the positions of objects. They are semantically distinct from natural languages used in image-level tasks. This requirement necessitates the tuning of the LLM to adapt to region-level tasks, but may disrupt the internal state of the LLM suitable for image-level VL tasks. To address the above issues, we devise a stage-wise instruction-tuning strategy and mixture-of-adapters with a router." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "type": "text", + "content": "The stage-wise instruction-tuning strategy is proposed to alleviate the internal conflict between image-level and region-level VL tasks during instruction-tuning. It is composed of three stages for instruction-tuning on image-level, region-level VL tasks and both, respectively, which is depicted in Fig. 3. In stage 1, we follow instructBLIP [7] and fine-tune Q-Former and the image-level adapter [5] in the LLM on image-level VL tasks, such as image captioning and VQA. In stage 2, we propose a vision aggregator for better capturing visual features in fine-grained understanding, which will be introduced later, and tune it with MLP and the region-level adapter on region-level VL tasks. The independent training in the first two stages greatly fulfills the requirements of sufficiently learning both image-level and region-level tasks, providing a solid foundation for subsequent joint training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 578 + ], + "type": "text", + "content": "Mixture-of-Adapters with a Router. In stage 3 of our stage-wise instruction-tuning, we need a unified model but encounter a situation where adapters of LLM in stages 1 and 2 are different and suit distinct input-output modes. Inspired by Mixture-of-Experts, we treat each adapter as an expert, and propose a router module to avoid the potential interference between them, as depicted in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "text", + "content": "An adapter [5] is inserted at each FFN layer in a parallel manner. Assuming " + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "inline_equation", + "content": "X \\in \\mathcal{R}^{L \\times D}" + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "text", + "content": " is the hidden representations generated by a self-attention / causal attention layer, the output representations after FFN (represented as " + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "text", + "content": ") with the adapter (denoted by " + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 46, + 579, + 287, + 639 + ], + "type": "text", + "content": ") layer are formulated as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 651, + 287, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 651, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 122, + 651, + 287, + 664 + ], + "type": "interline_equation", + "content": "O = \\mathbf {F} (X) + \\mathbf {H} (X), \\tag {1}", + "image_path": "d479fbfa39bf57b5f7230f264d3d34423c78053a5af1d49e0827c9e6d74b8ef3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 682, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 682, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 115, + 682, + 287, + 696 + ], + "type": "interline_equation", + "content": "\\mathbf {H} (X) = W _ {u} \\left(\\sigma \\left(W _ {d} X\\right)\\right), \\tag {2}", + "image_path": "f3ed60b2bcdade40c53f4b0ac161ece8e879e4bed4fb2bbc5f6623e4d9068ee0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": " is a non-linear function, ReLU. Our router module" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 312, + 69, + 530, + 152 + ], + "blocks": [ + { + "bbox": [ + 312, + 69, + 530, + 152 + ], + "lines": [ + { + "bbox": [ + 312, + 69, + 530, + 152 + ], + "spans": [ + { + "bbox": [ + 312, + 69, + 530, + 152 + ], + "type": "image", + "image_path": "6ac934367e3a1fe75439f6fc5bfb232e64fe29629a0cbe6d907cff9b8ba2bd45.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 158, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 547, + 191 + ], + "type": "text", + "content": "Figure 4. Instruction template with soft prompt. We use a well-designed instruction template with trainable soft prompts to inject the image tags generated by the RAM model into LION." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "text", + "content": "aims to dynamically aggregate the hidden features from the main branches and the multiple adapter branches according to task types. Given a set of adapters " + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{H}_1,\\dots ,\\mathbf{H}_K\\}" + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "text", + "content": ", each kind of task " + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "text", + "content": " defines a specific router function " + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^t" + }, + { + "bbox": [ + 305, + 201, + 546, + 260 + ], + "type": "text", + "content": " to generate new hidden features, which can be formulated as," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 268, + 545, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 268, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 356, + 268, + 545, + 301 + ], + "type": "interline_equation", + "content": "O ^ {t} = \\mathbf {F} (X) + \\sum_ {k = 1} ^ {K} \\mathrm {G} _ {k} ^ {t} \\odot \\mathbf {H} _ {k} (X). \\tag {3}", + "image_path": "40026ffecb2a63d9cad7d545165d85da8e72bfe136c7f0b1c695185fb0d53ad6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 308, + 546, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 546, + 405 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 308, + 546, + 405 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_k^t\\in \\mathcal{R}^D" + }, + { + "bbox": [ + 304, + 308, + 546, + 405 + ], + "type": "text", + "content": " is a trainable vector that modulates the hidden features from each adapter and makes them suitable for the target task. In practice, we define two types of tasks, one for image-level VL tasks (image captioning and VQA), the other for fine-grained VL tasks (visual grounding). Compared to directly incorporating multiple adapters, the router module provides a better way to maximize the complementarity of image-level and region-level tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 547, + 489 + ], + "type": "text", + "content": "We use the standard language modeling loss in all instruction-tuning stages. In the experiments, we demonstrate that stage-wise training is superior to single-stage training, and ensures a good balance between high-level and fine-grained visual understanding capabilities, further achieves a significant mutual promotion between image-level and region-level VL tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 503, + 417, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 417, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 417, + 515 + ], + "type": "text", + "content": "3.1.3 Vision Aggregator" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "To extract more sufficient visual details from input images, we devise a vision aggregator that integrates multi-level hidden features of the pretrained visual encoder. Although the vision encoder has a global reception field in all layers, it is verified that different transformer layers learn visual information at different scales [9], e.g., lower layers learn visual details. Thus, our vision aggregator makes fine-grained spatial-aware visual knowledge more likely to be learned based on visual grounding tasks. Specifically, our vision aggregator can be regarded as a tiny transformer-style network, consisting of two transformer layers for aggregating the hidden features from the vision encoder. Given the hidden features " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\{V_{i}, V_{j}, V_{k}\\}" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " from some middle layers in the vision encoder, the vision aggregation module uses two blocks to sequentially integrate the former two features with the last feature. Each block " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " is composed of self attention" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26543" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 116, + 525, + 300 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 1. Comparison on image captioning and VQA. “†” denotes including in-house data that are publicly inaccessible. “*” means our evaluated results by using publicly released checkpoints, which are only for reference as official evaluation settings are incomplete. We report CIDEr score for Flickr30K, COOCap, and TextCaps, Mean Reciprocal Rank (MRR) for Visual Dialog (VisDial), and top-1 accuracy for others. The best and second performances for each benchmark are indicated in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 116, + 525, + 300 + ], + "lines": [ + { + "bbox": [ + 65, + 116, + 525, + 300 + ], + "spans": [ + { + "bbox": [ + 65, + 116, + 525, + 300 + ], + "type": "table", + "html": "
ModelFlickr30KCOCOCapTextCapsOKVQAAOKVQAGQAIconQAVSRVisDial
Flamingo-3B [1]60.6073.00------46.10
Flamingo-9B [1]61.5079.40-44.70----48.00
Kosmos-1 [13]67.1084.70-------
Kosmos-2 [31]80.50--------
AdapterV2 [11]-122.20-------
Shikra [4]73.90117.50-47.16-----
Pink [45]---59.50-52.6047.8066.30-
MiniGPT4 [54]17.75*17.04*24.06*37.5034.51*30.8037.6041.6016.52*
LLaVA [26]48.03*73.85*45.54*54.4034.51*41.3043.0051.208.65*
MiniGPTV2 [3]80.75*129.16*80.60*56.90-60.3047.7060.608.47*
BLIVA [12]87.10-----44.8862.2045.63
InstructBLIP† (T5XL) [7]84.50138.2182.5549.2857.8648.4050.0064.8046.60
InstructBLIP† (T5XXL) [7]83.50138.2882.5348.5956.1647.9051.2065.6048.50
InstructBLIP (T5XL) [7]83.71135.47104.1747.3856.1246.3452.4769.9348.75
InstructBLIP (T5XXL) [7]85.79138.63105.4453.0259.3847.7453.1868.4650.41
LION-4B85.57138.20104.8751.0859.9849.5054.9172.9650.02
LION-12B87.12139.25108.7657.3360.8751.5654.8973.7750.42
", + "image_path": "3f8cf83c83ac3cb9324c28bc732caca42c2e030ae6a4144c9d0057486dc06de1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 288, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 340 + ], + "type": "text", + "content": "(Attn), cross attention (XAttn), and Feed-forward network (FFN) arranged in a sequential manner. Finally, the output features " + }, + { + "bbox": [ + 46, + 303, + 288, + 340 + ], + "type": "inline_equation", + "content": "\\bar{V}" + }, + { + "bbox": [ + 46, + 303, + 288, + 340 + ], + "type": "text", + "content": " is generated as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 348, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 348, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 115, + 348, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\bar {V} = \\mathcal {B} _ {2} \\left(\\mathcal {B} _ {1} \\left(V _ {i}; V _ {j}\\right); V _ {k}\\right), \\tag {4}", + "image_path": "3ec8f3206bde0848478a93727ac43897310f27de3ceb999eaaebefd7aa1103c4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 382, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 382, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 82, + 382, + 287, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {B} (X; Y) = \\operatorname {F F N} (\\operatorname {X A t t n} (\\operatorname {A t t n} (X), Y)). \\tag {5}", + "image_path": "0e945673faef31686a3e982470851d7a84c2a17c254d8cad98f8783dd173bab6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "text", + "content": "In practice, we use the middle layers " + }, + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\{i = L - 1, j = 2L/3, k = L/3\\}" + }, + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "text", + "content": " in the vision encoder to produce the hidden features as the input to VA, where " + }, + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 400, + 288, + 449 + ], + "type": "text", + "content": " is the number of layers in the vision encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 287, + 479 + ], + "type": "text", + "content": "3.2. Soft Prompting of High-Level Semantic Visual Evidence" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 617 + ], + "type": "text", + "content": "The vision encoder in a MLLM may be insufficient in comprehensively extracting visual information required by complex multi-modal tasks, although it has been trained on large-scale image-text pairs. It has been demonstrated that increasing the amount and quality of pretraining multimodal datasets can significantly improve the visual understanding capability of MLLMs [41], but inevitably induces prohibitive computational overhead. An appealing alternative is to harness the convenient and powerful off-the-shelf vision models to capture various aspects of visual content within an image as a supplement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "We choose the recognize anything model (RAM) [52] as an off-the-shelf vision model to provide diverse tags, encompassing objects, scenes, actions, and attributes, as visual evidence to support comprehensive visual perception. Instead of directly adding tags in the instruction, we design a soft prompting method to guide the model to adaptively use the inserted tags in order to avoid the potential negative influence caused by the imperfect predictions from RAM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 303, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 547, + 508 + ], + "type": "text", + "content": "In Fig. 4, we present the instruction template of tags along with the soft prompt that is a trainable vector. Our soft prompting approach can be regarded as a kind of prompt tuning methods, which guides the model toward the right direction. In standard prompt tuning works, the right direction is directly formulated as the optimization for task goals. In our work, the right direction is specified by a tailored sentence, \"According to , you are allowed to use or partially use the following tags: \", and \"\" will be replaced by the soft prompt. Our soft prompting method for inserting tags has some distinct properties. It is designed to adaptively select valuable information from tags, rather than serving a specific task, as seen in standard prompt tuning methods. Our method directly uses the output labels from a small off-the-shelf vision model to incorporate high-level semantic visual evidence into a MLLM, so as to eliminate extra computational overhead of the feature alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 388, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 540, + 547, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 589 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments to demonstrate the effectiveness of our model along with the quantitative and qualitative analyses. Please refer to Appendix for implementation details and training details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 598, + 506, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 506, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 506, + 611 + ], + "type": "text", + "content": "4.1. Evaluations on Image-Level VL Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Here, we evaluate multi-modal understanding abilities of LION on two kinds of image-level VL tasks, image captioning and VQA. Image captioning requires the model to generate a text description of the input image. We use COCO caption [6], TextCaps [39] and Flickr30K [48] as benchmarks, and report CIDEr as the evaluation metric. We utilize greedy search for caption generation. VQA provides an image along with a specific question for the model, asking" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26544" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 118, + 83, + 471, + 272 + ], + "blocks": [ + { + "bbox": [ + 108, + 71, + 484, + 82 + ], + "lines": [ + { + "bbox": [ + 108, + 71, + 484, + 82 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 484, + 82 + ], + "type": "text", + "content": "Table 2. Comparison on REC. \"Avg.\" means the average of top-1 accuracy over all the 8 evaluation sets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 118, + 83, + 471, + 272 + ], + "lines": [ + { + "bbox": [ + 118, + 83, + 471, + 272 + ], + "spans": [ + { + "bbox": [ + 118, + 83, + 471, + 272 + ], + "type": "table", + "html": "
ModelRefCOCORefCOCO+RefCOCOgAvg.
valtest-Atest-Bvaltest-Atest-Bvaltest
Zero-shot Setting
Kosmos-2 [31]52.3257.4247.2645.4850.7342.2460.5761.6552.21
GRILL [15]-------47.50-
Pink [45]54.1061.2044.2043.9050.7035.0059.1060.1051.00
LION-4B57.8956.0758.4046.3845.2947.5064.7463.5654.98
LION-12B58.5456.4159.3645.9345.7347.8966.1264.6955.58
Fune-tuning Setting
OFA-L [42]79.9683.6776.3968.2976.0061.7567.5767.5872.65
VisionLLM-H [43]-86.70-------
Shikra-7B [4]87.0190.6180.2481.6087.3672.1282.2782.1982.93
Shikra-13B [4]87.8391.1181.8182.8987.7974.4182.6483.1683.96
Pink [45]88.3091.7084.0081.4087.5073.7083.7083.7084.25
Ferret-7B [47]87.4991.3582.4580.7887.3873.1483.9384.7683.91
Ferret-13B [47]89.4892.4184.3682.8188.1475.1785.8386.3485.57
MiniGPTv2 [3]88.6991.6585.3379.9785.1274.4584.4484.6684.29
LION-4B89.7392.2984.8283.6088.7277.3485.6985.6385.98
LION-12B89.8093.0285.5783.9589.2278.0685.5285.7486.36
", + "image_path": "0d640217b63431cb0d93ff8a6c53ab4a863e1f1367f28433c5a33c82ce72d82c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 336, + 259, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 279, + 287, + 334 + ], + "lines": [ + { + "bbox": [ + 46, + 279, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 287, + 334 + ], + "type": "text", + "content": "Table 3. The comparison of various strategies in the instruction-tuning period. \"REC Avg.\" represents the average score of all REC tasks. \"Held-in\" denotes the average score of COOCap, TextCaps, OKVQA, and AOKVQA, while \"Held-out\" means the average score of Flickr30K, GQA, IconQA, VSR, and VisDial." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 336, + 259, + 392 + ], + "lines": [ + { + "bbox": [ + 72, + 336, + 259, + 392 + ], + "spans": [ + { + "bbox": [ + 72, + 336, + 259, + 392 + ], + "type": "table", + "html": "
StrategyImage-LevelRegion-Level
Held-inHeld-outREC Avg.
Single Stage84.7960.203.78
Stage-wise88.0761.9154.46
w/ Router87.6762.1654.98
", + "image_path": "3d68b7c483be1038e77e380e39ba0aa7f25fcf2a4a5019c7a29dbb61e5372aa3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": "for the output as an answer. We evaluate LION on six VQA datasets, including OKVQA [30], AOKVQA [34], GQA [14], IconQA [28], Visual Spatial Reasoning [24], and Visual Dialog (VisDial) [8]. For OKVQA, A-OKVQA, and GQA, we employ an open-ended generation with a greedy decoding strategy, For IconQA, Visual Spatial Reasoning, and Visual Dialog, we match the output with various candidates, and select the candidate with the highest value as the prediction. We report Mean Reciprocal Rank (MRR) for Visual Dialog, and top-1 accuracy for other VQA tasks. The detailed descriptions of these datasets and inference instructions are presented in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "As shown in Table 1, LION achieves the best performance across 7 out of 9 benchmarks, and the second on OKVQA. LION shares the same training datasets with InstructBLIP, except Visual Genome dataset adopted in our work and the in-house dataset, WebCapFilt, used in InstructBLIP. The amount of Visual Genome dataset (3.6M) is far smaller than WebCapFilt (14M). Compared to the original InstructBLIP trained with WebCapFilt, LION exhibits superior performances on all zero-shot evaluation benchmarks, showcasing a better generalization ability. We also re-implemented InstructBLIP on the same instruction-tuning datasets. The comparison shows the consistent and significant improvements of LION over the re-implemented InstructBLIP, demonstrating the effectiveness of incorpo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 353 + ], + "type": "text", + "content": "rating dual-level visual knowledge. Shikra, MiniGPTV2 and Pink are also integrated with visual grounding abilities. Their inferior results on most image-level VL tasks exhibit that our proposed stage-wise instruction-tuning strategy and soft prompting of high-level semantic knowledge help enhance holistic visual understanding abilities of MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 360, + 510, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 510, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 510, + 373 + ], + "type": "text", + "content": "4.2. Evaluations on Region-Level VL Tasks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 462 + ], + "type": "text", + "content": "To assess the fine-grained perceptual and reasoning abilities of LION, we evaluate it on three REC datasets, RefCOCO [16], RefCOCO+ [16], RefCOCOg [29]. REC requires the model to locate the target object given a referring expression. We follow the standard setting, and use accuracy as an evaluation metric, which means it is correct when the IOU between prediction and ground-truth is no less than 0.5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 463, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 617 + ], + "type": "text", + "content": "In Table 2, we show the comparison between LION and other MLLMs with respect to the grounding abilities, under the settings of zero-shot and fine-tuning evaluations, respectively. In the zero-shot evaluation setting, we directly employ LION to generate coordinates of referring expressions on three datasets. Our model shows significant improvements on most evaluation sets over Kosmos-2 and Pink, except test-A sets of RefCOCO and RefCOCO+. The languages used in RefCOCOg are more flowery than those used in RefCOCO and RefCOCO+. The significant improvements on RefCOCOg clearly demonstrate that LION can handle complex referring expressions and has superior zero-shot spatial-aware visual understanding abilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "In the fine-tuning setting, we fine-tune LION with training samples from three REC datasets. As shown in Table 2, our model achieves the best performance on average and on most of the evaluation sets, indicating the advanced fine-grained perception ability of our model. Ferret proposes a spatial-aware visual sampler to handle free-form referred regions, and meticulously constructs an extensive grounding dataset with lots of efforts on data generation and fil" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26545" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 95, + 523, + 149 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 4. Ablation studies of dual-level visual knowledge. \"VG\" means visual grounding tasks. \"Held-in\" and \"Held-out\" denote the training images of tasks are seen and unseen, respectively. \"REC Avg.\" means the average score of all REC tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 95, + 523, + 149 + ], + "lines": [ + { + "bbox": [ + 67, + 95, + 523, + 149 + ], + "spans": [ + { + "bbox": [ + 67, + 95, + 523, + 149 + ], + "type": "table", + "html": "
ComponentsHeld-inHeld-outREC Avg.
VGTagsCOCOCapTextCapsOKVQAAOKVQAFlickr30KGQAVSRIconQAVisDial
××135.47104.1747.3856.1283.7146.3469.9352.4748.75-
✓×137.87104.8451.0756.9083.9949.2273.2054.6749.7054.98
✓✓138.20104.8751.0859.9885.5749.5072.9654.9150.0254.92
", + "image_path": "d498e9f6c354e02df43165f95f6fda8084cd498c8e2986f9fb6b3b6a38a4e9c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 167, + 507, + 299 + ], + "blocks": [ + { + "bbox": [ + 75, + 154, + 517, + 165 + ], + "lines": [ + { + "bbox": [ + 75, + 154, + 517, + 165 + ], + "spans": [ + { + "bbox": [ + 75, + 154, + 517, + 165 + ], + "type": "text", + "content": "Table 5. Evaluation of object hallucination on POPE benchmark. F1 score is the major metric for hallucication evaluation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 167, + 507, + 299 + ], + "lines": [ + { + "bbox": [ + 82, + 167, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 82, + 167, + 507, + 299 + ], + "type": "table", + "html": "
DatasetsMetricsLIONShikra [4]InstructBLIP [7]MiniGPT-4 [54]LLaVA [26]mPLUG-Owl [46]
RandomF1-Score ↑88.3386.1989.2780.1766.6468.39
Accuracy ↑88.9786.9088.5779.6750.3753.97
Precision ↑97.1294.4084.0978.2450.1952.07
Recall ↑81.0079.2795.1382.2099.1399.60
PopularF1-Score ↑85.9483.1684.6673.0266.4466.94
Accuracy ↑86.7783.9782.7769.7349.8750.90
Precision ↑91.6987.5576.2765.8649.9350.46
Recall ↑80.8779.2095.1381.9399.2799.40
AdversarialF1-score ↑84.7182.4977.3270.4266.3266.82
Accuracy ↑85.3783.1072.1065.1749.7050.67
Precision ↑88.6985.6065.1361.1949.8550.34
Recall ↑81.0779.6095.1382.9399.0799.33
", + "image_path": "2a1cc45ba7c7edb00d9b8d7981175d41c8d601a3f3c7e6499faa48224b8aa77b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 287, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 356 + ], + "type": "text", + "content": "tering. However, LION can achieve superior performances compared to Ferret by using a simple vision aggregator and existing datasets, implying the effectiveness of fine-grained visual knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 369, + 141, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 141, + 382 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 141, + 382 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 473 + ], + "type": "text", + "content": "The effect of vision aggregator. We conduct an ablation study of the vision aggregator with only visual grounding tasks on LION-4B during stage 2 of the stage-wise instruction-tuning. As illustrated in Fig. 5, the removal of the vision aggregator degrades REC performances, validating that aggregating multi-level vision features promotes the extraction of fine-grained spatial-aware visual knowledge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 714 + ], + "type": "text", + "content": "Stage-wise instruction-tuning mitigates the conflict between image-level and region-level tasks. We investigate the performance of two types of VL tasks under three instruction-tuning strategies, i.e., single stage, stage-wise, and stage-wise with a router. As shown in Table 3, the stage-wise instruction-tuning strategy shows a significant improvement on the average REC performance, which is completely damaged in the single stage instruction-tuning. The worse REC performance of the single stage strategy can be attributed to the lack of pretraining on large-scale grounding datasets, like in Kosmos-2 and Shikra, and the gap of their input-output modes. To address these challenges, stage-wise training progressively integrates fine-grained spatial-aware knowledge from visual grounding datasets by splitting the whole instruction-tuning process into three stages. The model can sufficiently learn diverse levels of visual knowledge in separate training stages (stages 1 and 2), and incorporate them in the final training stage (stage 3 in Fig. 3). This contributes to the performance improvements of all VL tasks. Furthermore, stage-wise instruction" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 308, + 544, + 411 + ], + "blocks": [ + { + "bbox": [ + 309, + 308, + 544, + 411 + ], + "lines": [ + { + "bbox": [ + 309, + 308, + 544, + 411 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 544, + 411 + ], + "type": "image", + "image_path": "d3ad7060d434ab9934c0312c1c662441e306dba7b6aaebc8f13269f70c8e1932.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 416, + 545, + 460 + ], + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 460 + ], + "type": "text", + "content": "Figure 5. The effect of the vision aggregator. The results on RefCOCO, RefCOCO+, and RefCOCOg clearly show that the proposed module can overall improve REC performances across 8 evaluation sets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 471, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 531 + ], + "type": "text", + "content": "tuning with the router improves the held-out and REC performance, with a slight degradation in the held-in performance. All these results demonstrate LION's ability to handle the potential conflict of various VL tasks and maximize the learning benefit." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 532, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 546, + 712 + ], + "type": "text", + "content": "Dual-level visual knowledge enhances multimodal understanding abilities of MLLMs. We evaluate the performance of our model integrated with different levels of visual knowledge on various benchmarks in Table 4. It can be seen that dual-level visual knowledge can upgrade the performance of all VL tasks to varying degrees. When progressively incorporating fine-grained spatial-aware knowledge, the performances of four tasks, OKVQA, GQA, IconQA, and Visual Spatial Reasoning, are significantly improved, as they highly require region-level understanding and spatial reasoning. When inserting tags as high-level visual evidence, we can see substantial performance increases on Flickr30K and AOKVQA, which demand more comprehensive semantic knowledge than other tasks, like COCO caption and OKVQA." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26546" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 74, + 504, + 198 + ], + "blocks": [ + { + "bbox": [ + 72, + 62, + 520, + 73 + ], + "lines": [ + { + "bbox": [ + 72, + 62, + 520, + 73 + ], + "spans": [ + { + "bbox": [ + 72, + 62, + 520, + 73 + ], + "type": "text", + "content": "Table 6. Evaluation on MMBench test set, all the reported results of compared models are from the leadboard of MMBench." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 74, + 504, + 198 + ], + "lines": [ + { + "bbox": [ + 86, + 74, + 504, + 198 + ], + "spans": [ + { + "bbox": [ + 86, + 74, + 504, + 198 + ], + "type": "table", + "html": "
ModelsText EncoderVision EncoderOverallLRARRRFP-SFP-CCP
MiniGPT-4 [54]Vincuna-7BEVA-G23.013.632.98.928.811.228.3
PandaGPT [40]Vincuna-13BImageBind ViT-H/1442.523.161.534.132.728.757.6
VisualGLM [10]ChatGLM-6BEVA-CLIP33.511.448.827.735.817.641.5
InstructBLIP [7]Vicuna-7BEVA-G33.921.647.422.533.024.441.1
LLaVA-v1.5 [25]Vicuna-v1.5-7BCLIP ViT-L/1459.532.472.649.362.352.267.7
Otter-I [19]LLaMA-7BCLIP ViT-L/1448.322.263.339.446.836.460.6
Shikra [4]Vincuna-7BCLIP ViT-L/1460.233.569.653.161.850.471.7
LMEye [23]FlanT5-XLCLIP ViT-L/1462.641.074.355.961.658.769.2
MMICL [53]FlanT5-XXLEVA-G65.244.377.964.866.553.670.6
mPLUG-Owl2 [44]LLaMA-7BCLIP ViT-L/1466.043.476.062.168.655.973.0
LIONFlanT5-XXLEVA-G73.451.784.178.474.060.878.9
", + "image_path": "6d96c5732e6c677521ecd2398c85642ded8bfaa88ef417cf9efe40bd752cbd11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 92, + 213, + 195, + 284 + ], + "blocks": [ + { + "bbox": [ + 113, + 202, + 173, + 212 + ], + "lines": [ + { + "bbox": [ + 113, + 202, + 173, + 212 + ], + "spans": [ + { + "bbox": [ + 113, + 202, + 173, + 212 + ], + "type": "text", + "content": "Image Captioning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 92, + 213, + 195, + 284 + ], + "lines": [ + { + "bbox": [ + 92, + 213, + 195, + 284 + ], + "spans": [ + { + "bbox": [ + 92, + 213, + 195, + 284 + ], + "type": "image", + "image_path": "8c65a703fac70fcc7476868190635020af76a6be15a63a48b683f8d30bffd709.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 287, + 118, + 303 + ], + "lines": [ + { + "bbox": [ + 58, + 287, + 118, + 303 + ], + "spans": [ + { + "bbox": [ + 58, + 287, + 118, + 303 + ], + "type": "text", + "content": "User What is in the photo?" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 267, + 212, + 359, + 284 + ], + "blocks": [ + { + "bbox": [ + 267, + 202, + 358, + 211 + ], + "lines": [ + { + "bbox": [ + 267, + 202, + 358, + 211 + ], + "spans": [ + { + "bbox": [ + 267, + 202, + 358, + 211 + ], + "type": "text", + "content": "Visual Question Answering" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 267, + 212, + 359, + 284 + ], + "lines": [ + { + "bbox": [ + 267, + 212, + 359, + 284 + ], + "spans": [ + { + "bbox": [ + 267, + 212, + 359, + 284 + ], + "type": "image", + "image_path": "a9f6248c60773c4ebd8f5e7ca12f0b922865f97797afb22e2d5412026c3c7190.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 285, + 376, + 303 + ], + "lines": [ + { + "bbox": [ + 248, + 285, + 376, + 303 + ], + "spans": [ + { + "bbox": [ + 248, + 285, + 376, + 303 + ], + "type": "text", + "content": "User On which side of the photo are the clay.pots?" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 248, + 310, + 304, + 323 + ], + "blocks": [ + { + "bbox": [ + 58, + 314, + 227, + 344 + ], + "lines": [ + { + "bbox": [ + 58, + 314, + 227, + 344 + ], + "spans": [ + { + "bbox": [ + 58, + 314, + 227, + 344 + ], + "type": "text", + "content": "InstructBLIP boy in karate uniform with white belt and black belts watching him do a kick in a martial arts class at a karate club in california." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 248, + 310, + 304, + 323 + ], + "lines": [ + { + "bbox": [ + 248, + 310, + 304, + 323 + ], + "spans": [ + { + "bbox": [ + 248, + 310, + 304, + 323 + ], + "type": "image", + "image_path": "6ad3068a5701dcd8a9fb015a02ff704f44041cb91dde011c020ce7dee9b956e3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 323, + 352, + 330 + ], + "lines": [ + { + "bbox": [ + 248, + 323, + 352, + 330 + ], + "spans": [ + { + "bbox": [ + 248, + 323, + 352, + 330 + ], + "type": "text", + "content": "The pots are on the left of the image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 248, + 346, + 282, + 358 + ], + "blocks": [ + { + "bbox": [ + 58, + 348, + 228, + 378 + ], + "lines": [ + { + "bbox": [ + 58, + 348, + 228, + 378 + ], + "spans": [ + { + "bbox": [ + 58, + 348, + 228, + 378 + ], + "type": "text", + "content": "LION A young boy wearing a white karate uniform is performing a kick in front of a yellow wall, with the word \"Kwon\" written on the wall behind him." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 248, + 346, + 282, + 358 + ], + "lines": [ + { + "bbox": [ + 248, + 346, + 282, + 358 + ], + "spans": [ + { + "bbox": [ + 248, + 346, + 282, + 358 + ], + "type": "image", + "image_path": "03570a4610050db94adf21ec54a5588d37bb7df5925e2dbd2d69b088e566f70d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 248, + 358, + 357, + 365 + ], + "lines": [ + { + "bbox": [ + 248, + 358, + 357, + 365 + ], + "spans": [ + { + "bbox": [ + 248, + 358, + 357, + 365 + ], + "type": "text", + "content": "The pots are on the right of the image." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 388, + 545, + 424 + ], + "lines": [ + { + "bbox": [ + 46, + 388, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 545, + 424 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of InstructBLIP, Shikra, and LION. We mark the hallucination or incorrect part in red, and highlight the correct part in green for comparison. These samples exhibit that LION is able to achieve superior fine-grained understanding and visual spatial reasoning capabilities with fewer hallucinated responses." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 416, + 212, + 510, + 284 + ], + "blocks": [ + { + "bbox": [ + 400, + 202, + 525, + 211 + ], + "lines": [ + { + "bbox": [ + 400, + 202, + 525, + 211 + ], + "spans": [ + { + "bbox": [ + 400, + 202, + 525, + 211 + ], + "type": "text", + "content": "Referring Expression Comprehension" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 212, + 510, + 284 + ], + "lines": [ + { + "bbox": [ + 416, + 212, + 510, + 284 + ], + "spans": [ + { + "bbox": [ + 416, + 212, + 510, + 284 + ], + "type": "image", + "image_path": "66d98fb325214e17f61bf05d6ee4eab6d4a5436288b9fbfb1cf2a3a427cf62c6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 393, + 285, + 525, + 309 + ], + "lines": [ + { + "bbox": [ + 393, + 285, + 525, + 309 + ], + "spans": [ + { + "bbox": [ + 393, + 285, + 525, + 309 + ], + "type": "text", + "content": "User Help me locate the coordinates of a brown bear near a soda bottle in the image." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 393, + 314, + 525, + 338 + ], + "lines": [ + { + "bbox": [ + 393, + 314, + 525, + 338 + ], + "spans": [ + { + "bbox": [ + 393, + 314, + 525, + 338 + ], + "type": "text", + "content": "Shikra The brown bear near a soda bottle is located at [0.002, 0.248, 0.416, 0.760]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 393, + 348, + 525, + 372 + ], + "lines": [ + { + "bbox": [ + 393, + 348, + 525, + 372 + ], + "spans": [ + { + "bbox": [ + 393, + 348, + 525, + 372 + ], + "type": "text", + "content": "LION The brown bear near a soda bottle is located at [0.550, 0.387, 0.920, 0.867]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 287, + 453 + ], + "type": "text", + "content": "4.4. Evaluations on Object Hallucination and MM-Bench" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 460, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 627 + ], + "type": "text", + "content": "Li et al. [22] present an open-sourced evaluation benchmark, called POPE, to evaluate the object hallucination [33]. We follow the POPE evaluation pipeline to inspect LION. The results in Table 5 show that LION has superior results, especially under popular and adversarial settings, which means that incorporating fine-grained and high-level semantic visual knowledge into MLLM can mitigate the object hallucination to some degree. To comprehensively validate the effectiveness of our method, we further evaluate LION on MMBench [27]. The results are summarized in Table 6. Our strong performances across various skills demonstrate that the progressive incorporation of fine-grained knowledge significantly alleviates the hallucination phenomenon in MLLMs." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 635, + 164, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 164, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 164, + 647 + ], + "type": "text", + "content": "4.5. Qualitative Analysis" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "As shown in Fig. 6, we depict various examples to validate the advanced perceptual and reasoning abilities of LION. The left example exhibits our superior fine-grained understanding capability to correctly generate the right attributes \"white\", \"yellow\" and the character \"Kwon\". The middle" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 430, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 545, + 490 + ], + "type": "text", + "content": "example validates the advantage of our model in visual spatial reasoning. The right example shows that LION accurately localizes the referring object, while Shikra provides an incorrect response caused by the misunderstanding of fine-grained details \"a soda bottle\"." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 378, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "content": "To address the insufficient extraction and reasoning of visual information in MLLMs, we propose LION to exploit dual-level visual knowledge, i.e., fine-grained spatial-aware visual knowledge and high-level semantic visual evidence. To mitigate the internal conflict among tasks, LION adopts a stage-wise instruction-tuning strategy. Extensive experiments validate the superiority of LION in image captioning, VQA, and visual grounding tasks." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 628, + 408, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 408, + 641 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 408, + 641 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 647, + 545, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 707 + ], + "type": "text", + "content": "This study is supported by Shenzhen College Stability Support Plan (Grant No. GXWD20220817144428005), National Natural Science Foundation of China (Grant No. 62236003 and 62306090), China Postdoctoral Science Foundation (Grant No. 2023TQ0095 and 2023M740923)." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26547" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 289, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 289, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 289, + 258 + ], + "type": "text", + "content": "[3] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023.5.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 305 + ], + "type": "text", + "content": "[4] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 361 + ], + "type": "text", + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 407 + ], + "type": "text", + "content": "[6] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 288, + 464 + ], + "type": "text", + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. In NeurIPS, 2023. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 467, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 467, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 467, + 288, + 520 + ], + "type": "text", + "content": "[8] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José MF Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 326-335, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 523, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 523, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 523, + 288, + 588 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 656 + ], + "type": "text", + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 320-335, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[11] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023. 2, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[12] Wenbo Hu, Yifan Xu, Y Li, W Li, Z Chen, and Z Tu. Bliva: A simple multimodal llm for better handling of text-rich visual questions. arXiv preprint arXiv:2308.09936, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 162 + ], + "type": "text", + "content": "[13] Shaohan Huang, Li Dong, Wenhui Wang, Yaru Hao, Saksham Singhal, Shuming Ma, Tengchao Lv, Lei Cui, Owais Khan Mohammed, Qiang Liu, et al. Language is not all you need: Aligning perception with language models. arXiv preprint arXiv:2302.14045, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[14] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[15] Woojeong Jin, Subhabrata Mukherjee, Yu Cheng, Yelong Shen, Weizhu Chen, Ahmed Hassan Awadallah, Damien Jose, and Xiang Ren. Grill: Grounded vision-language pretraining via aligning text and image regions. arXiv preprint arXiv:2305.14676, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 396 + ], + "type": "text", + "content": "[17] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 399, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 464 + ], + "type": "text", + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "type": "text", + "content": "[19] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv preprint arXiv:2305.03726, 2023. 1, 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 553 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 600 + ], + "type": "text", + "content": "[21] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[22] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[23] Yunxin Li, Baotian Hu, Xinyu Chen, Lin Ma, and Min Zhang. Lmeye: An interactive perception network for large language models. arXiv preprint arXiv:2305.03701, 2023.8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[24] Fangyu Liu, Guy Emerson, and Nigel Collier. Visual spatial reasoning. Transactions of the Association for Computational Linguistics, 11:635-651, 2023. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26548" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 287, + 184 + ], + "type": "text", + "content": "[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 287, + 239 + ], + "type": "text", + "content": "[28] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 287, + 297 + ], + "type": "text", + "content": "[29] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 298, + 287, + 353 + ], + "type": "text", + "content": "[30] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 287, + 398 + ], + "type": "text", + "content": "[31] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 399, + 287, + 464 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 287, + 498 + ], + "type": "text", + "content": "[33] Vipula Rawte, Amit Sheth, and Amitava Das. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 555 + ], + "type": "text", + "content": "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European Conference on Computer Vision, pages 146–162. Springer, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 287, + 588 + ], + "type": "text", + "content": "[35] Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi-adversarial discriminative deep domain generalization for face presentation attack detection. In CVPR, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 287, + 633 + ], + "type": "text", + "content": "[36] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense with clean-adversarial mutual learning. International Journal of Computer Vision, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 635, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 635, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 49, + 635, + 287, + 666 + ], + "type": "text", + "content": "[37] Rui Shao, Tianxing Wu, and Ziwei Liu. Detecting and grounding multi-modal media manipulation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[38] Rui Shao, Tianxing Wu, Jianlong Wu, Liqiang Nie, and Zwei Liu. Detecting and grounding multi-modal media manipulation and beyond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 137 + ], + "type": "text", + "content": "[39] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758. Springer, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "text", + "content": "[40] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[41] Guangzhi Wang, Yixiao Ge, Xiaohan Ding, Mohan Kankanhalli, and Ying Shan. What makes for good visual tokenizers for large language models? arXiv preprint arXiv:2305.12223, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 284 + ], + "type": "text", + "content": "[42] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 340 + ], + "type": "text", + "content": "[43] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Vision: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 396 + ], + "type": "text", + "content": "[44] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, et al. mplug-2: A modularized multi-modal foundation model across text, image and video. arXiv preprint arXiv:2302.00402, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 441 + ], + "type": "text", + "content": "[45] Shiyu Xuan, Qingpei Guo, Ming Yang, and Shiliang Zhang. Pink: Unveiling the power of referential comprehension for multi-modal llms. arXiv preprint arXiv:2310.00582, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 498 + ], + "type": "text", + "content": "[46] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 553 + ], + "type": "text", + "content": "[47] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 609 + ], + "type": "text", + "content": "[48] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 612, + 545, + 644 + ], + "type": "text", + "content": "[49] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[50] Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. Investigating the catastrophic forgetting in multimodal large language models. arXiv preprint arXiv:2309.10313, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 712 + ], + "type": "text", + "content": "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26549" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 105 + ], + "type": "text", + "content": "Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023.2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 160 + ], + "type": "text", + "content": "[52] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 217 + ], + "type": "text", + "content": "[53] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. arXiv preprint arXiv:2309.07915, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 286, + 262 + ], + "type": "text", + "content": "[54] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26550" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_content_list.json b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4c792694a5c7bce5c8b9862e523365ccb82c689 --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_content_list.json @@ -0,0 +1,2073 @@ +[ + { + "type": "text", + "text": "LISA: Reasoning Segmentation via Large Language Model", + "text_level": 1, + "bbox": [ + 186, + 130, + 784, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Lai $^{1*}$ Zhuotao Tian $^{2*†}$ Yukang Chen $^{1}$ Yanwei Li $^{1}$ Yuhui Yuan $^{4}$ Shu Liu $^{3}$ Jiaya Jia $^{1,3}$ $^{1}\\mathrm{CUHK}$ $^{2}\\mathrm{HIT}$ (Shenzhen) $^{3}\\mathrm{SmartMore}$ $^{4}\\mathrm{MSRA}$", + "bbox": [ + 86, + 179, + 883, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task — reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlabresearch/LISA.", + "bbox": [ + 75, + 282, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 734, + 209, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In daily life, users tend to issue direct commands like \"Change the TV channel\" to instruct a robot, rather than providing explicit step-by-step instructions such as \"Go to the table first, find the TV remote, and then press the button to change the channel.\" However, existing perception systems consistently rely on humans to explicitly indicate target objects or pre-define categories before executing visual recog", + "bbox": [ + 73, + 758, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nition tasks. These systems cannot actively reason and comprehend user intention based on implicit instruction. This reasoning ability is crucial in developing next-generation intelligent perception systems and holds substantial potential for industrial applications, particularly in robotics.", + "bbox": [ + 500, + 252, + 893, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce a new segmentation task — reasoning segmentation, which requires generating a binary segmentation mask based on an implicit query text involving complex reasoning. Notably, the query text is not limited to a straightforward reference (e.g., “the orange”), but a more complicated description involving complex reasoning or world knowledge (e.g., “the food with high Vitamin C”). To accomplish this task, the model must possess two key abilities: 1) reasoning complex and implicit text queries jointly with the image; 2) producing segmentation masks.", + "bbox": [ + 496, + 332, + 893, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the exceptional capacity of LLMs to reason and comprehend user intentions, we aim to leverage this capability of LLMs to address the aforementioned first challenge. However, while several studies [1, 23, 24, 28, 29, 55, 63] have integrated robust reasoning capabilities into multimodal LLMs to accommodate visual input, the majority of these models primarily concentrate on text generation tasks and still fall short in performing vision tasks that require fine-grained output formats, such as segmentation masks. This leads us to ask: can we enable multimodal LLMs with the capability to output segmentation masks?", + "bbox": [ + 496, + 488, + 895, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we introduce LISA: a large Language Instructed Segmentation Assistant, a multimodal LLM capable of producing segmentation masks. Specifically, we incorporate an additional token, i.e., $<\\mathrm{SEG}>$ , into the existing vocabulary. Upon generating the $<\\mathrm{SEG}>$ token, its hidden embedding is further decoded into the corresponding segmentation mask. By representing the segmentation mask as an embedding, LISA acquires segmentation capabilities and benefits from end-to-end training. Remarkably, LISA demonstrates robust zero-shot abilities. Training the model solely on standard semantic segmentation and referring segmentation datasets yields surprisingly effective performance on the reasoning segmentation task. Furthermore, we find that LISA's performance can be significantly enhanced by fine-tuning on just 239 reasoning segmentation data samples. As illustrated in Fig. 1, LISA can handle various scenarios", + "bbox": [ + 496, + 657, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author.", + "bbox": [ + 96, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9579", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- USER: Can you segment the tyre that does not touch the ground in this image?", + "- ASSISTANT: Sure, it is ." + ], + "bbox": [ + 88, + 104, + 210, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09a3adad0cc374de132e4ae29b4d128db4221c4e9b7805ee97e6bf3937620096.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 103, + 344, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c2512837cd9c28e35a9d8df1bd7a96f2a60197d49abc8435c9a7db3fc375b0d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 104, + 477, + 171 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- USER: There are two washing machines as shown in the picture. If I need to do laundry, where in the picture would I put the clothes? Please output segmentation mask.", + "bbox": [ + 496, + 97, + 678, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- ASSISTANT: .", + "bbox": [ + 500, + 159, + 607, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0acaf7c6d33ce0213009a977dd6a4baeed984e74de62ec003de6713e7231b5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 90, + 774, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a2de191a1037497bef85eda8d3a180fdaa0f5d5a92fe0ab13f94fad1426666d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 90, + 875, + 172 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- USER: Can you segment the food that tastes not spicy in this image?", + "bbox": [ + 89, + 193, + 200, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ASSISTANT: .", + "bbox": [ + 98, + 238, + 199, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9e153b9ee7022c45bfc37405011eb746e7ed5eebd5f46b0d906570cc76b84ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 190, + 323, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f3de821f41bd3d8a3c7e8be9db2977b383841638d5eaaa3ce280d23e269fb10c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 190, + 444, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- USER: What is the place where the driver can observe the speed in this image? Please output segmentation mask.", + "bbox": [ + 460, + 189, + 594, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- ASSISTANT: .", + "bbox": [ + 460, + 242, + 571, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/50f0ba0edf6a0fe1d788787241b99590ccc63f93f7863edfc16c83a426a3bad8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 189, + 733, + 252 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2967c4020606cbc77806987d8ce1af3b97b0538a50ba8dded645cd5e57b5682d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 189, + 875, + 252 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- USER: Can you segment the unusual part in this image and explain why?", + "bbox": [ + 89, + 271, + 230, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- ASSISTANT: Sure, it is . In the image, the unusual part is the dog wearing a reindeer antler headband.", + "bbox": [ + 89, + 306, + 233, + 349 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/59801feb73fb53abce28f395d5ad8c119cf6343348faacb2e08a0b90febf0e29.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 273, + 320, + 349 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cd12e4cd63ab23c9db1b831006e4db947ae3af78d355fdcb1536cfd33f62ad08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 273, + 406, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- USER: Who is the loser in this image? Please output segmentation mask and explain why.", + "bbox": [ + 424, + 271, + 596, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ASSISTANT: Sure, it is . ... The man on top is likely the winner of the match, as he has control over the other fighters body...", + "bbox": [ + 424, + 308, + 602, + 351 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5326720e924ea3944f1bfa50741ccd7a1470ab666e691cf9d6c8fdf26dc5a680.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 281, + 738, + 347 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b2b7b9a4ea969654f4ed89b67dff6bdd74f46441148c1cea41f1303bd611336c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 281, + 875, + 349 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e31ac66fccd7e539d1b84f13c5c37b77ac96e48e4ec8bb7852fe67ec0bd0cadf.jpg", + "image_caption": [ + "Figure 1. We unlock new segmentation capabilities for existing multimodal LLMs. Our model (i.e., LISA) can deal with cases involving complex reasoning and world knowledge. Also, we demonstrate the cases of explanatory answers in the 3rd row. Additionally, in the 4th row, our model can output multiple segmentation masks in a single answer. More illustrations can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 109, + 364, + 246, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- USER: Can you segment the fried food and the food with most protein separately in this image?", + "- ASSISTANT: Sure, they are and ." + ], + "bbox": [ + 264, + 382, + 516, + 417 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1344c5d2785ec033c0b1f52778c5511f38768875089b8ee10cce644455c8827.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 364, + 684, + 436 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9cb3dffa2988a67e34d8ef8948518983538841f951e9231e86e7795b31b78414.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 364, + 852, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "involving complex reasoning and world knowledge.", + "bbox": [ + 75, + 506, + 419, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, to validate the effectiveness, we establish a benchmark for reasoning segmentation evaluation, called ReasonSeg. Comprising over one thousand image-instruction pairs, this benchmark offers persuasive evaluation metrics for the task. To align more closely with practical applications, we annotate the images from OpenImages [21] and ScanNetv2 [10] with implicit text queries that involve complex reasoning.", + "bbox": [ + 75, + 521, + 468, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 96, + 642, + 395, + 656 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the reasoning segmentation task, which necessitates reasoning based on implicit human instructions. Such reasoning capability is crucial for building a genuinely intelligent perception system.", + "- We present our model — LISA, which incorporates new segmentation capabilities. It demonstrates robust zero-shot ability on the reasoning segmentation task when trained solely on reasoning-free datasets, and achieves further performance boost by fine-tuning on just 239 data samples that involve reasoning.", + "- We establish a reasoning segmentation benchmark, ReasonSeg, containing over one thousand image-instruction-mask data samples. This benchmark is essential for evaluation and encourages the community to further explore the reasoning ability for vision tasks." + ], + "bbox": [ + 94, + 657, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 503, + 640, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Segmentation", + "text_level": 1, + "bbox": [ + 500, + 522, + 692, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Semantic segmentation aims to assign a class label to every pixel in an image. Numerous studies [2, 5, 8, 12, 16, 22, 31, 37, 42, 43, 45, 46, 51, 56, 59-61, 64] have proposed diverse designs (such as encoder-decoder, dilated convolution, pyramid pooling module, non-local operator, and more) to effectively encode semantic information. Research on instance segmentation [9, 14, 58] and panoptic segmentation [7, 18, 25, 50] has introduced various architectural innovations for instance-level segmentation, including DETR [4]-based structures, mask attention, and dynamic convolution. In recent years, typical segmentation tasks have made significant progress and become increasingly mature. Consequently, it is imperative to develop more intelligent interaction ways for image segmentation.", + "bbox": [ + 496, + 537, + 893, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The referring segmentation task [17, 36] enables interaction with human language, aiming to segment the target object based on a given explicit text description. Recently, Kirillov et al. [19] introduced SAM, trained with billions of high-quality masks, supporting bounding boxes and points as prompts while demonstrating exceptional segmentation quality. X-Decoder [65] bridges vision and language, unifying multiple tasks within a single model. SEEM [66] further supports various human interaction methods, including text, audio, and scribble. However, these studies primarily focus", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9580", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/56991a6f3a5e5f8ce80b50a1d4b30070d5e2947b1d711ba4a144318e9a456b98.jpg", + "image_caption": [ + "the camera lens that is more suitable for photographing nearby objects", + "Toddlers are curious and often enjoy exploring their surroundings. What object in the picture can provide a safe and enclosed space for a toddler to play in?" + ], + "image_footnote": [], + "bbox": [ + 135, + 93, + 527, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2724632b9d752ec872d52434f95ac729991821fa8e65b14932cacaa6f44cb043.jpg", + "image_caption": [ + "Figure 2. Examples of the annotated image-instruction-mask data samples. Left: short phrase query. Right: long sentence query. More examples are given in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 537, + 93, + 834, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on addressing multi-task compatibility and unification, neglecting the injection of new capabilities. In this work, we present LISA and it possesses reasoning ability that has not been explored yet in existing segmentors.", + "bbox": [ + 75, + 296, + 470, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multimodal Large Language Model", + "text_level": 1, + "bbox": [ + 76, + 359, + 387, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the remarkable reasoning abilities of LLMs, researchers are exploring ways to transfer these capabilities into the vision domain, developing multimodal LLMs. Flamingo [1] employs a cross-attention structure to attend to visual contexts, enabling visual in-context learning. Models such as BLIP-2 [24] and mPLUG-OWL [55] propose encoding image features with a visual encoder, which are then fed into the LLM alongside text embeddings. Otter [23] further incorporates robust few-shot capabilities through in-context instruction tuning on the proposed MIMIC-IT dataset. LLaVA [29] and MiniGPT-4 [63] first conduct image-text feature alignment followed by instruction tuning. Koh et al. [20] also investigates image retrieval for LLMs. Moreover, numerous works [32, 44, 49, 52, 54] utilize prompt engineering, connecting independent modules via API calls, but without the benefits of end-to-end training. Recently, there have been studies examining the intersection between multimodal LLMs and vision tasks. Vision-LLM [47] offers a flexible interaction interface for multiple vision-centric tasks through instruction tuning but fails to fully exploit LLMs for complex reasoning. Kosmos-2 [38] constructs large-scale data of grounded image-text pairs, infusing grounding capabilities into LLMs. DetGPT [39] bridges the fixed multimodal LLM and open-vocabulary detector, enabling detection to be performed based on user instruction. GPT4RoI [57] introduces spatial boxes as input and trains the model on region-text pairs. In contrast, our work aims to efficiently inject segmentation capabilities into multimodal LLMs in the manner of end-to-end training.", + "bbox": [ + 75, + 380, + 472, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Reasoning Segmentation", + "text_level": 1, + "bbox": [ + 76, + 829, + 307, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 76, + 849, + 259, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The reasoning segmentation task is to output a binary segmentation mask $\\mathbf{M}$ , given an input image $\\mathbf{x}_{img}$ and an im", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plicit query text instruction $\\mathbf{x}_{\\text{txt}}$ . The task shares a similar formulation with the referring segmentation task [17], but is far more challenging. The key distinction lies in the complexity of the query text in reasoning segmentation. Instead of a straightforward phrase (e.g., \"the trash can\"), the query text includes more intricate expressions (e.g., \"something that the garbage should be put into\") or longer sentences (e.g., \"After cooking, consuming food, and preparing for food, where can we throw away the rest of the food and scraps?\") that involve complex reasoning or world knowledge.", + "bbox": [ + 496, + 296, + 893, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.Benchmark", + "text_level": 1, + "bbox": [ + 500, + 455, + 625, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the lack of quantitative evaluation, it is imperative to establish a benchmark for the reasoning segmentation task. To ensure reliable assessment, we have collected a diverse set of images from OpenImages [21] and ScanNetv2 [10], annotating them with implicit text instructions and high-quality target masks. To cover different scenarios, our text instructions consist of two types: 1) short phrases; 2) long sentences; as illustrated in Figure 2. The resulting ReasonSeg benchmark comprises a total of 1218 image-instruction-mask data samples. This dataset is further partitioned into three splits: train, val, and test, containing 239, 200, and 779 data samples, respectively. As the primary purpose of the benchmark is evaluation, the validation and testing sets include a larger number of data samples. The details of data annotation are given in the supplementary material.", + "bbox": [ + 496, + 478, + 893, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Our Method", + "text_level": 1, + "bbox": [ + 500, + 719, + 630, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the model architecture in Sec. 4.1. After that, we elaborate on the training data preparation and training parameters in Sec. 4.2.", + "bbox": [ + 496, + 744, + 892, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Architecture", + "text_level": 1, + "bbox": [ + 500, + 801, + 633, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embedding as Mask. Most current multimodal LLMs (such as LLaVA [29], Flamingo [1], BLIP-2 [24], Otter [23], etc.) support image and text as input, but they can only output text and cannot directly output fine-grained segmentation masks. VisionLLM [47] offers a solution by parsing", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9581", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/05eed35d4fb8518cd038dea284bec7f20db5651948f11f6d0db41be165bc0d40.jpg", + "image_caption": [ + "Figure 3. The pipeline of LISA. Given the input image and text query, the multimodal LLM (e.g., LLaVA [29]) generates text output. The last-layer embedding for the token is then decoded into the segmentation mask via the decoder. We use LoRA [15] for efficient fine-tuning. The choice of vision backbone can be flexible (e.g., SAM [66], Mask2Former [9])." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 883, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "segmentation masks as sequences of polygons, enabling the representation of segmentation masks as plain text and allowing end-to-end training within the framework of existing multimodal LLMs. However, end-to-end training with the polygon sequences introduces optimization challenges and may compromise generalization ability unless a massive amount of data and computational resources are employed. For instance, training a 7B model, VisionLLM requires $4 \\times 8$ NVIDIA 80G A100 GPUs and 50 epochs, which is computationally prohibitive. In contrast, it takes less than 3 days to train LISA-7B on 8 NVIDIA 24G 3090 GPUs.", + "bbox": [ + 75, + 328, + 472, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose the embedding-as-mask paradigm to infuse new segmentation capabilities into the multimodal LLM. The pipeline of our method is illustrated in Fig. 3. Specifically, we first expand the original LLM vocabulary with a new token, i.e., $<\\mathrm{SEG}>$ , which signifies the request for the segmentation output. Given a text instruction $\\mathbf{x}_{txt}$ along with the input image $\\mathbf{x}_{img}$ , we feed them into the multimodal LLM $\\mathcal{F}$ , which in turn outputs a text response $\\hat{\\mathbf{y}}_{txt}$ . It can be formulated as", + "bbox": [ + 75, + 494, + 473, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {t x t} = \\mathcal {F} (\\mathbf {x} _ {i m g}, \\mathbf {x} _ {t x t}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 637, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the LLM intends to generate a binary segmentation mask, the output $\\hat{\\mathbf{y}}_{txt}$ would include a token. We then extract the LLM last-layer embedding $\\tilde{\\mathbf{h}}_{seg}$ corresponding to the token and apply an MLP projection layer $\\gamma$ to obtain $\\mathbf{h}_{seg}$ . Simultaneously, the vision backbone $\\mathcal{F}_{enc}$ extracts the dense visual features $\\mathbf{f}$ from the visual input $\\mathbf{x}_{img}$ . Finally, $\\mathbf{h}_{seg}$ and $\\mathbf{f}$ are fed to the decoder $\\mathcal{F}_{dec}$ to produce the final segmentation mask $\\hat{\\mathbf{M}}$ . The detailed structure of the decoder $\\mathcal{F}_{dec}$ follows [19]. The process can be formulated as", + "bbox": [ + 75, + 659, + 472, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {s e g} = \\gamma (\\tilde {\\mathbf {h}} _ {s e g}), \\quad \\mathbf {f} = \\mathcal {F} _ {e n c} (\\mathbf {x} _ {i m g}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 815, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\mathcal {F} _ {d e c} (\\mathbf {h} _ {s e g}, \\mathbf {f}).\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 839, + 352, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objectives. The model is trained end-to-end using the text generation loss $\\mathcal{L}_{t:xt}$ and the segmentation", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mask loss $\\mathcal{L}_{mask}$ . The overall objective $\\mathcal{L}$ is the weighted sum of these losses, determined by $\\lambda_{txt}$ and $\\lambda_{mask}$ :", + "bbox": [ + 498, + 328, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {m a s k} \\mathcal {L} _ {m a s k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 371, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, $\\mathcal{L}_{txt}$ is the auto-regressive cross-entropy loss for text generation, and $\\mathcal{L}_{mask}$ is the mask loss, which encourages the model to produce high-quality segmentation results. To compute $\\mathcal{L}_{mask}$ , we employ a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights $\\lambda_{bce}$ and $\\lambda_{dice}$ . Given the ground-truth targets $\\mathbf{y}_{txt}$ and $\\mathbf{M}$ , these losses can be formulated as", + "bbox": [ + 496, + 398, + 893, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t x t} = \\mathbf {C E} (\\hat {\\mathbf {y}} _ {t x t}, \\mathbf {y} _ {t x t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 529, + 774, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {\\mathbf {M}}, \\mathbf {M}) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {\\mathbf {M}}, \\mathbf {M}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 542, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noteworthy that the proposed method endows existing multimodal LLMs with new segmentation capabilities, such that they can generate not only text but also fine-grained output formats. Also, our method is based on an end-to-end training pipeline and connects the LLM and vision modules with hidden embedding representation, which proves significantly more effective than the decoupled two-stage method as discussed in Sec. 5.2.", + "bbox": [ + 496, + 578, + 893, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 500, + 708, + 604, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Data Formulation. As illustrated in Fig. 4, our training data comprises mainly three parts, all of which are derived from widely-used public datasets. The details are as follows:", + "bbox": [ + 496, + 731, + 893, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Semantic Segmentation Dataset. Semantic segmentation datasets typically consist of images and the corresponding multi-class labels. During training, we randomly choose several categories for each image. To generate data that matches the format of visual question answering, we employ a question-answer template like “USER: Can you segment the", + "bbox": [ + 500, + 795, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9582", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/32a6e91c80caff3657c829b697f482f86fbdda77300dfd4a607cdb056d0b3ad3.jpg", + "image_caption": [ + "Raw Data" + ], + "image_footnote": [], + "bbox": [ + 107, + 116, + 325, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/76b5ed7bce4ae8a94540d47033c18525144f1246515c1ad6f70c7c165a7e198a.jpg", + "image_caption": [], + "image_footnote": [ + "the lady with the blue shirt" + ], + "bbox": [ + 370, + 114, + 527, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39526b71af96bda5721d1b15aaf424999785330d0d23cc8ab59d463f0fb3fd77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 119, + 864, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Processed Data", + "text_level": 1, + "bbox": [ + 96, + 258, + 207, + 271 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- USER: Can you segment the table in this image?", + "- ASSISTANT: It is ." + ], + "bbox": [ + 112, + 282, + 307, + 328 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/124555f612515343327fde250c264861e45415c7bef15716f24116c05bb4231f.jpg", + "image_caption": [ + "Semantic Segmentation Data" + ], + "image_footnote": [], + "bbox": [ + 117, + 330, + 316, + 443 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- USER: Can you segment the lady with the blue shirt in this image?", + "- ASSISTANT: Sure, it is ." + ], + "bbox": [ + 367, + 284, + 607, + 330 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf130f8d06ee9fff521df64a51723f34eb4436252b27baeef66e62594133abc1.jpg", + "image_caption": [ + "Referring Segmentation Data" + ], + "image_footnote": [], + "bbox": [ + 375, + 333, + 486, + 444 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df82e6b4af35777b0ad654250a2e259e11491570087d0397d34e4573655db38.jpg", + "image_caption": [ + "Figure 4. The illustration of training data formulation from different types of data, including semantic segmentation data, referring segmentation data, and visual question answering (VQA) data." + ], + "image_footnote": [], + "bbox": [ + 498, + 333, + 609, + 443 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- USER: What type of sign and traffic device can be seen in the image? ", + "- ASSISTANT: In the image, there is a street sign and a traffic light above a city road.", + "- USER: How many traffic lights are visible in the image?" + ], + "bbox": [ + 651, + 282, + 870, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "No Binary Segmentation Mask", + "text_level": 1, + "bbox": [ + 651, + 425, + 869, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VQA Data", + "text_level": 1, + "bbox": [ + 720, + 450, + 797, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "{class_name} in this image? ASSISTANT: It is .\", where {class_name} is the chosen category, and denotes the placeholder for tokens of image patches. The corresponding binary segmentation mask is used as the ground truth to provide mask loss supervision. During training, we also use other templates to generate the QA data to ensure data diversity, as shown in the supplementary material. We adopt ADE20K, COCO-Stuff, and LVIS-PACO part segmentation datasets.", + "bbox": [ + 89, + 523, + 472, + 659 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vanilla Referring Segmentation Dataset. Referring segmentation datasets provide an input image and an explicit short description of the target object. Thus, it is easy to convert them into question-answer pairs using a template like \"USER: Can you segment {description} in this image? ASSISTANT: Sure, it is .\", where {description} is the given explicit description. For this part, we adopt ref-COCO, refCOCO+, refCOCOg, and refCLEF datasets.", + "- Visual Question Answering Dataset. To preserve the original Visual Question Answering (VQA) ability of the multimodal LLM, we also include the VQA dataset during training. We use LLaVA-Instruct-150k [29] for LLaVA v1" + ], + "bbox": [ + 76, + 674, + 472, + 887 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and LLaVA-v1.5-mix665k for LLaVA v1.5 [28].", + "bbox": [ + 513, + 523, + 834, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notably, the above datasets do not include any reasoning segmentation data sample. Instead, it only contains samples where the target objects are explicitly indicated in the query texts. Surprisingly, even without complex reasoning training data, LISA demonstrates impressive zero-shot ability on the ReasonSeg benchmark, as shown in Table 1. Moreover, we find that further performance boost could be yielded by finetuning the model on only 239 data samples that involve complex reasoning.", + "bbox": [ + 496, + 545, + 893, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trainable Parameters. To preserve the learned knowledge of the pre-trained multimodal LLM $\\mathcal{F}$ (i.e., LLaVA in our experiments), we leverage LoRA [15] to perform efficient fine-tuning, and completely freeze the vision backbone $\\mathcal{F}_{enc}$ . The decoder $\\mathcal{F}_{dec}$ is fully fine-tuned. Additionally, the LLM token embeddings (embed_tokens), the LLM head (lm_head), and the projection layer $\\gamma$ are also trainable.", + "bbox": [ + 496, + 700, + 893, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is notable that the resulting model avoids the catastrophic forgetting of the original text generation capability and preserves the conversation ability, as verified in the supplementary material. The potential reasons are: we 1) employ LoRA fine-tuning to reduce the trainable parameters and 2) incorporate the VQA dataset during fine-tuning.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9583", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/623a2069cff2f6cdddfe2583bbaac44dddfd4369f091b4cb0e5e5bac79ce25b9.jpg", + "table_caption": [ + "Table 1. Reasoning segmentation results among LISA (ours) and previous related works. 'ft' denotes using 239 reasoning segmentation data samples to fine-tune the model. Unless otherwise specified, we use LLaVA v1 [29] as the base model. LLaVA1.5 denotes LLaVA v1.5 [28]." + ], + "table_footnote": [], + "table_body": "
Methodvaltest
overallshort querylong queryoverall
gIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoU
OVSeg [26]28.518.618.015.528.722.526.120.8
GRES [27]22.419.917.615.022.623.821.322.0
X-Decoder [65]22.617.920.411.622.217.521.716.3
SEEM [66]25.521.220.111.525.620.824.318.7
Grounded-SAM [30]26.014.517.810.822.418.621.316.4
LISA-7B44.446.037.634.436.634.736.834.1
LISA-7B (ft)52.954.040.640.649.451.047.348.4
LISA-13B48.946.939.943.346.446.544.845.8
LISA-13B (ft)56.262.944.342.054.054.351.751.1
LLaVA1.5-7B + OVSeg38.223.524.218.744.637.139.731.8
LISA-7B-LLaVA1.553.652.347.148.549.248.948.748.8
LISA-7B-LLaVA1.5 (ft)61.362.948.346.357.959.755.656.9
LLaVA1.5-13B + OVSeg37.926.427.119.446.140.641.534.1
LISA-13B-LLaVA1.557.760.350.850.054.750.953.850.8
LISA-13B-LLaVA1.5 (ft)65.072.955.450.663.265.361.362.2
", + "bbox": [ + 96, + 121, + 867, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 431, + 200, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 76, + 454, + 277, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network Architecture. Unless otherwise specified, we use LLaVA-7B-v1-1 or LLaVA-13B-v1-1 [29] as the base multimodal LLM $\\mathcal{F}$ , and adopt the ViT-H SAM [19] backbone as the vision backbone $\\mathcal{F}_{enc}$ . The projection layer of $\\gamma$ is an MLP with channels of [256, 4096, 4096].", + "bbox": [ + 75, + 474, + 470, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt 8 NVIDIA 24G 3090 GPUs for training. The training scripts are based on deep-speed [41] engine. We use AdamW [33] optimizer with the learning rate and weight decay set to 0.0003 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the text generation loss $\\lambda_{txt}$ and the mask loss $\\lambda_{mask}$ are set to 1.0 and 1.0, respectively, and those of the bce loss $\\lambda_{bce}$ and the dice loss $\\lambda_{dice}$ are set to 2.0 and 0.5, respectively. Besides, the batch size per device is set to 2, and the gradient accumulation step is set to 10. During training, we select at most 3 categories for each image in semantic segmentation datasets.", + "bbox": [ + 75, + 559, + 470, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. As mentioned in Sec. 4.2, our training data is mainly composed of three types of datasets: (1) For the semantic segmentation dataset, we use ADE20K [62] and COCO-Stuff [3]. Besides, to enhance the segmentation result for some part of an object, we also use part semantic segmentation datasets, including PACO-LVIS [40], PartImageNet [13], and PASCAL-Part [6]; (2) For the referring segmentation dataset, we use refCLEF, refCOCO, refCOCO+ [17], and refCOCOg [35]; (3) For the visual ques", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion answering (VQA) dataset, we use the datasets of LLaVA-Instruct-150k for LLaVA v1 [29] and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]. In order to avoid data leakage, we exclude the COCO samples whose images are present in the refCOCO $(+ / \\mathrm{g})$ validation sets during training. Furthermore, we surprisingly find that by fine-tuning the model on only 239 ReasonSeg data samples, the model's performance can be further boosted.", + "bbox": [ + 496, + 434, + 893, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We follow most previous works on referring segmentation [17, 35] to adopt two metrics: gIoU and cIoU. gIoU is defined by the average of all per-image Intersection-over-Unions (IoUs), while cIoU is defined by the cumulative intersection over the cumulative union. Since cIoU is highly biased toward large-area objects and it fluctuates too much, gIoU is preferred.", + "bbox": [ + 496, + 564, + 893, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Reasoning Segmentation Results", + "text_level": 1, + "bbox": [ + 500, + 680, + 785, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The reasoning segmentation results are shown in Table 1. It is worth noting that existing works fail to handle the task, but our model can accomplish the task involving complex reasoning with more than $20\\%$ gIoU performance boost. As mentioned before, the reasoning segmentation task is essentially different from the referring segmentation task in that it requires the model to possess reasoning ability or access world knowledge. Only by truly understanding the query, can the model do well in the task. The existing works have no proper way to understand an implicit query, but our model exploits multimodal LLMs to reach the goal.", + "bbox": [ + 496, + 703, + 893, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, we also make a comparison with the vanilla two-stage method (LLaVA1.5 + OVSeg). Specifically, the", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9584", + "bbox": [ + 480, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6484f12817af483ce2ee6ba6e0fadd9e2d2b7b91723783225abcbbe1b5beb2c0.jpg", + "table_caption": [ + "Table 2. Referring segmentation results (cIoU) among LISA (ours) and existing methods." + ], + "table_footnote": [], + "table_body": "
MethodrefCOCOrefCOCO+refCOCOg
valtestAtestBvaltestAtestBval(U)test(U)
MCN [34]62.464.259.750.655.044.749.249.4
VLT [11]67.570.565.256.361.050.155.057.7
CRIS [48]70.573.266.162.368.153.759.960.4
LAVT [53]72.775.868.862.168.455.161.262.1
ReLA [27]73.876.570.266.071.057.765.066.0
X-Decoder [65]------64.6-
SEEM [66]------65.7-
LISA-7B74.176.571.162.467.456.566.468.5
LISA-7B (fine-tuned on ReferSeg)74.979.172.365.170.858.167.970.6
", + "bbox": [ + 99, + 107, + 861, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "two-stage method refers to first using a multimodal LLM (e.g., LLaVA v1.5) to generate a text output for the input query, and then adopting a referring or open-vocabulary segmentation model (e.g., OVSeg) to generate the segmentation mask. If the intermediate text output remains too long and exceeds the input token length limit of OVSeg, we use GPT-3.5 to further summarize. More details can be found in the supplementary material. The results in Table 1 show that our model outperforms the two-stage method significantly. We explain that the potential reasons are: 1) Our model is trained end-to-end, while the two-stage method is completely decoupled; 2) The two-stage method relies on text as an intermediary to transmit information, while our model utilizes the hidden embedding that is more expressive.", + "bbox": [ + 75, + 311, + 470, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another finding is that LISA-13B outperforms the 7B counterpart substantially, especially on the long-query scenarios, which indicates that the current performance bottleneck may still lie in understanding the query text, and a stronger multimodal LLM (e.g., LLaVA v1.5 [28]) leads to even better results.", + "bbox": [ + 75, + 521, + 468, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Vanilla Referring Segmentation Results", + "text_level": 1, + "bbox": [ + 76, + 621, + 416, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show that our model is also competent in the vanilla referring segmentation task, we make a comparison with existing state-of-the-art methods in Table 2. We evaluate the methods on refCOCO, refCOCO+, refCOCOg validation and testing sets. Our model achieves state-of-the-art results across various referring segmentation benchmarks.", + "bbox": [ + 75, + 645, + 468, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 744, + 230, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct an extensive ablation study to reveal the contribution of each component. Unless otherwise specified, we report the metrics of gIoU and cIoU of LISA-7B on the validation set.", + "bbox": [ + 75, + 768, + 470, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design Choices of Vision Backbone. We emphasize that vision backbones other than SAM are also applicable in our framework. In Table 3, we notice that SAM performs the best, potentially because of the massive high-quality", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/767075beabbddf423e44b55c9701d62bf2ba8f91cfbd36f250d160def6a976fc.jpg", + "table_caption": [ + "Table 3. Ablation study on the design choice of vision backbone. 'ft' denotes finetuning on ReasonSeg training set." + ], + "table_footnote": [], + "table_body": "
Vision BackbonegIoUcIoU
Mask2Former-Swin-L42.438.8
SAM (w/ LoRA)41.537.3
SAM44.446.0
Mask2Former-Swin-L (ft)50.752.3
SAM w/ LORA (ft)51.851.9
SAM (ft)52.954.0
", + "bbox": [ + 514, + 347, + 877, + 452 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4aef72fdc80aa3c6f149ae43a7ac9d2390e6f7cbae94a40fcb95a3bc3b1d9207.jpg", + "table_caption": [ + "Table 4. Ablation study on SAM pre-trained weight and rephrasing." + ], + "table_footnote": [], + "table_body": "
Exp. IDPre-trainSAM γrephrasinggIoUcIoU
1✓35.944.6
2✓50.751.1
3✓✓52.954.0
", + "bbox": [ + 532, + 493, + 859, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data used in its pre-training phase. Further, we also find that with the Mask2Former backbone, our framework still achieves a decent performance on the reasoning segmentation task, significantly outperforming previous works such as X-Decoder [65]. This reveals the fact that the design choice of vision backbone is flexible and not limited to SAM.", + "bbox": [ + 496, + 592, + 893, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SAM LoRA Fintuning. We also investigate the effectiveness of applying LoRA on the SAM backbone. In Table 3, we note that the performance of LoRA fine-tuned SAM backbone is inferior to that of the frozen one. A potential reason is that fine-tuning impairs the generalization ability of the original SAM model.", + "bbox": [ + 496, + 702, + 893, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SAM Pre-trained Weight. To demonstrate the contribution of SAM pre-trained weight, we make a comparison between Experiments 1 and 3 in Table 4. Without being initialized with SAM pre-trained weight, the vision backbone is trained from scratch. This causes the performance to fall substantially behind that of the baseline model.", + "bbox": [ + 496, + 814, + 893, + 905 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9585", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b91107beaf4c0b451f08e5a024c9e9ad8874bdbeffaf22947e330bd3664e0bab.jpg", + "image_caption": [ + "the most likely object that someone else has left behind" + ], + "image_footnote": [], + "bbox": [ + 102, + 93, + 135, + 116 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6e62571977c7576cab40e12bc252d7da23aa18eef0076e5b291d23fa030a307.jpg", + "image_caption": [ + "中" + ], + "image_footnote": [], + "bbox": [ + 256, + 97, + 333, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3e02c8979e997296a5193f45fac33af6ce2867cb60ef537857b971296807063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 97, + 864, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0328c0b81c2a3acb2431462e655a554574b609d45a47ee72d07573ad59c65150.jpg", + "image_caption": [ + "the object used for stirring milk or coffee" + ], + "image_footnote": [], + "bbox": [ + 102, + 160, + 135, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/363392ea772e52f662d1f5a0226e75995b4b73c0dfe33b0aa8b262d60d0d15dd.jpg", + "image_caption": [ + "W" + ], + "image_footnote": [], + "bbox": [ + 256, + 161, + 333, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66712d38cbc11cb3ba2c75df85d01a2a5be003753822b0e908f115e0fbf39cbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 161, + 864, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/111f9f6fe104f4630a753969a7ecf69f0a27d13f3daa462bc1611ef514988db2.jpg", + "image_caption": [ + "In cold days, dogs may need extra protection to keep them warm. What object in the picture can a dog wear to provide warmth during snowy walks?" + ], + "image_footnote": [], + "bbox": [ + 102, + 227, + 135, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0598a171e198f06f2bb5e16d8bb704744bc991bfb0764a403a0b76b1d8596b57.jpg", + "image_caption": [ + "中" + ], + "image_footnote": [], + "bbox": [ + 258, + 234, + 333, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d9cba76fae1045402466d9cdcb5726b34c1d5ca2a8e70ae264e677b64c107d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 234, + 864, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88fc045aea040343be95bd9dd217e2b1e2b22fca5acffd5e435a8bdc33439f05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 292, + 135, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d51fb2f1098a87272c673d6d1e48ef57dc26868c87078154069b93eae9d8ed1.jpg", + "image_caption": [ + "When a plane is ready to land on the airport runway, what area in the picture will it eventually land on?", + "Query", + "Image" + ], + "image_footnote": [], + "bbox": [ + 258, + 292, + 333, + 345 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c8a06e77cf69b1d705d409df03f16a34e67570f394d9b3a4c193a14fb1e244d.jpg", + "image_caption": [ + "中", + "Figure 5. Visual comparison among LISA (ours) and existing related methods. More illustrations are given in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 380, + 292, + 864, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/171b404cdbad77ac6cd077f2514f278cf6ae92acdae4ffe50edc975db809f56e.jpg", + "table_caption": [ + "Table 5. Ablation study on training data." + ], + "table_footnote": [], + "table_body": "
IDSemanticSegReferSegVQAReasonSeggIoUcIoU
ADE20KCOCO-StuffPartSeg
1✓✓✓✓✓48.953.5
2✓✓✓✓✓48.550.8
3✓✓✓✓✓46.750.9
4✓✓✓✓46.646.7
5✓✓✓30.420.4
6✓✓✓✓✓47.751.1
7✓✓✓✓✓44.446.0
8✓✓✓✓✓✓52.954.0
", + "bbox": [ + 78, + 404, + 470, + 561 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b443a0f6de179f9cf58126ed246759bc36e6fb9a03a81f7c673274e8f71b1bc.jpg", + "table_caption": [ + "Table 6. Results on the ReasonSeg test set." + ], + "table_footnote": [], + "table_body": "
Training splits# data samplesgIoUcIoU
train23951.751.1
train + val43954.054.9
", + "bbox": [ + 109, + 589, + 429, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instruction Rephrasing by GPT-3.5. When fine-tuning the model on the reasoning segmentation data samples, we rephrase the text instruction by GPT-3.5 (the details are shown in the supplementary material), and randomly choose one. The comparison between Experiments 2 and 3 in Table 4 shows that the performance is increased by $2.2\\%$ gIoU and $2.9\\%$ cIoU. This result verifies the effectiveness of such data augmentation.", + "bbox": [ + 75, + 656, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contribution of All Types of Training Data. In Table 5, we show the contribution of each type of data to the performance. We find that in Exp. 5, we do not use any semantic segmentation dataset, and the performance drops a lot. We conjecture that semantic segmentation datasets provide a large amount of ground-truth binary masks for training, since a multi-class label can induce multiple binary masks.", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also notice that adding more reasoning segmentation data samples during training leads to better results. In Table 6, we also add the ReasonSeg val set (200 data samples) during fine-tuning, and it yields better performance in both gIoU and cIoU metrics. This indicates that more reasoning segmentation training samples are beneficial at this moment.", + "bbox": [ + 496, + 387, + 893, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 488, + 681, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As depicted in Fig. 5, we provide a visual comparison with existing related works, including the model for open-vocabulary semantic segmentation (OVSeg), referring segmentation (GRES), and the generalist models for segmentation (X-Decoder and SEEM). These models fail to handle the displayed cases with various errors, while our approach produces accurate and high-quality segmentation results. More illustrations are given in the supplementary material.", + "bbox": [ + 496, + 512, + 893, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 617, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed a new segmentation task—reasoning segmentation. Also, we have introduced an evaluation benchmark ReasonSeg, which comprises over one thousand data samples. Finally, we have presented our model — LISA. It injects segmentation capabilities into current multimodal LLMs and performs surprisingly effectively on the reasoning segmentation task. We hope our work can shed new light on the direction of combining LLMs and vision tasks in the future.", + "bbox": [ + 496, + 674, + 893, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 818, + 666, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9586", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 1, 3", + "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 2017. 2", + "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 6", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 2", + "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014. 6", + "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2", + "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 2", + "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4", + "[10] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3", + "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, 2021. 7", + "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, 2019. 2", + "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In ECCV, 2022. 6", + "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 4, 5" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019. 2", + "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 2, 3, 6", + "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, 2019. 2", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 2, 4, 6", + "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3", + "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 3", + "[22] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In CVPR, 2021. 2", + "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 1, 3", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 1, 3", + "[25] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In CVPR, 2021. 2", + "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 6", + "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In CVPR, 2023. 6, 7", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint, 2023. 1, 5, 6, 7", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv:2304.08485, 2023. 1, 3, 4, 5, 6", + "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pretraining for open-set object detection. arXiv preprint, 2023. 6", + "[31] Wei Liu, Andrew Rabinovich, and Alexander C. Berg. Parsenet: Looking wider to see better. arXiv preprint, 2015. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9587", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Internchat: Solving vision-centric tasks by interacting with chatbots beyond language. arXiv:2305.05662, 2023. 3", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6", + "[34] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, 2020. 7", + "[35] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 6", + "[36] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2", + "[37] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In ICCV, 2015. 2", + "[38] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3", + "[39] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, and Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv:2305.14167, 2023. 3", + "[40] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In CVPR, 2023. 6", + "[41] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In SIGKDD, 2020. 6", + "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2", + "[43] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. TPAMI, 2017. 2", + "[44] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv:2303.17580, 2023. 3", + "[45] Zhuotao Tian, Pengguang Chen, Xin Lai, Li Jiang, Shu Liu, Hengshuang Zhao, Bei Yu, Ming-Chang Yang, and Jiaya Jia. Adaptive perspective distillation for semantic segmentation. TPAMI, 2022. 2", + "[46] Zhuotao Tian, Jiequan Cui, Li Jiang, Xiaojuan Qi, Xin Lai, Yixin Chen, Shu Liu, and Jiaya Jia. Learning context-aware classifier for semantic segmentation. AAAI, 2023. 2", + "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv:2305.11175, 2023. 3", + "[48] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. 7", + "[49] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671, 2023. 3", + "[50] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019. 2", + "[51] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In CVPR, 2018. 2", + "[52] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752, 2023. 3", + "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. 7", + "[54] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv:2303.11381, 2023. 3", + "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178, 2023. 1, 3", + "[56] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In ICLR, 2016. 2", + "[57] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3", + "[58] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. NeurIPS, 2021. 2", + "[59] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 2", + "[60] Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia. Icnet for real-time semantic segmentation on high-resolution images. In ECCV, 2018.", + "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In ECCV, 2018. 2", + "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6", + "[63] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9588", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language understanding with advanced large language models. arXiv:2304.10592, 2023. 1, 3", + "[64] Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, and Xi-ang Bai. Asymmetric non-local neural networks for semantic segmentation. In ICCV, 2019. 2", + "[65] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023. 2, 6, 7", + "[66] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 2, 4, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 261 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9589", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_model.json b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..490cef24a5922e8e6dc6fcfa13117e7811e9e56e --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_model.json @@ -0,0 +1,3049 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.785, + 0.154 + ], + "angle": 0, + "content": "LISA: Reasoning Segmentation via Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.885, + 0.219 + ], + "angle": 0, + "content": "Xin Lai\\(^{1*}\\) Zhuotao Tian\\(^{2*†}\\) Yukang Chen\\(^{1}\\) Yanwei Li\\(^{1}\\) Yuhui Yuan\\(^{4}\\) Shu Liu\\(^{3}\\) Jiaya Jia\\(^{1,3}\\) \\(^{1}\\mathrm{CUHK}\\) \\(^{2}\\mathrm{HIT}\\) (Shenzhen) \\(^{3}\\mathrm{SmartMore}\\) \\(^{4}\\mathrm{MSRA}\\)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.268 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.474, + 0.706 + ], + "angle": 0, + "content": "Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task — reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlabresearch/LISA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.735, + 0.21, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.76, + 0.472, + 0.868 + ], + "angle": 0, + "content": "In daily life, users tend to issue direct commands like \"Change the TV channel\" to instruct a robot, rather than providing explicit step-by-step instructions such as \"Go to the table first, find the TV remote, and then press the button to change the channel.\" However, existing perception systems consistently rely on humans to explicitly indicate target objects or pre-define categories before executing visual recog" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.253, + 0.895, + 0.33 + ], + "angle": 0, + "content": "nition tasks. These systems cannot actively reason and comprehend user intention based on implicit instruction. This reasoning ability is crucial in developing next-generation intelligent perception systems and holds substantial potential for industrial applications, particularly in robotics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.485 + ], + "angle": 0, + "content": "In this work, we introduce a new segmentation task — reasoning segmentation, which requires generating a binary segmentation mask based on an implicit query text involving complex reasoning. Notably, the query text is not limited to a straightforward reference (e.g., “the orange”), but a more complicated description involving complex reasoning or world knowledge (e.g., “the food with high Vitamin C”). To accomplish this task, the model must possess two key abilities: 1) reasoning complex and implicit text queries jointly with the image; 2) producing segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.897, + 0.656 + ], + "angle": 0, + "content": "Inspired by the exceptional capacity of LLMs to reason and comprehend user intentions, we aim to leverage this capability of LLMs to address the aforementioned first challenge. However, while several studies [1, 23, 24, 28, 29, 55, 63] have integrated robust reasoning capabilities into multimodal LLMs to accommodate visual input, the majority of these models primarily concentrate on text generation tasks and still fall short in performing vision tasks that require fine-grained output formats, such as segmentation masks. This leads us to ask: can we enable multimodal LLMs with the capability to output segmentation masks?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.897, + 0.901 + ], + "angle": 0, + "content": "To this end, we introduce LISA: a large Language Instructed Segmentation Assistant, a multimodal LLM capable of producing segmentation masks. Specifically, we incorporate an additional token, i.e., \\(<\\mathrm{SEG}>\\), into the existing vocabulary. Upon generating the \\(<\\mathrm{SEG}>\\) token, its hidden embedding is further decoded into the corresponding segmentation mask. By representing the segmentation mask as an embedding, LISA acquires segmentation capabilities and benefits from end-to-end training. Remarkably, LISA demonstrates robust zero-shot abilities. Training the model solely on standard semantic segmentation and referring segmentation datasets yields surprisingly effective performance on the reasoning segmentation task. Furthermore, we find that LISA's performance can be significantly enhanced by fine-tuning on just 239 reasoning segmentation data samples. As illustrated in Fig. 1, LISA can handle various scenarios" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.206, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.227, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.106, + 0.212, + 0.15 + ], + "angle": 0, + "content": "- USER: Can you segment the tyre that does not touch the ground in this image?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.151, + 0.212, + 0.167 + ], + "angle": 0, + "content": "- ASSISTANT: Sure, it is ." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.106, + 0.212, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.104, + 0.346, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.106, + 0.478, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.098, + 0.679, + 0.159 + ], + "angle": 0, + "content": "- USER: There are two washing machines as shown in the picture. If I need to do laundry, where in the picture would I put the clothes? Please output segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.16, + 0.608, + 0.169 + ], + "angle": 0, + "content": "- ASSISTANT: ." + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.092, + 0.775, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.789, + 0.092, + 0.877, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.194, + 0.202, + 0.238 + ], + "angle": 0, + "content": "- USER: Can you segment the food that tastes not spicy in this image?" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.239, + 0.2, + 0.247 + ], + "angle": 0, + "content": "ASSISTANT: ." + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.191, + 0.324, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.191, + 0.445, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.19, + 0.596, + 0.243 + ], + "angle": 0, + "content": "- USER: What is the place where the driver can observe the speed in this image? Please output segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.462, + 0.243, + 0.572, + 0.251 + ], + "angle": 0, + "content": "- ASSISTANT: ." + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.19, + 0.734, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.19, + 0.877, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.272, + 0.231, + 0.307 + ], + "angle": 0, + "content": "- USER: Can you segment the unusual part in this image and explain why?" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.308, + 0.235, + 0.351 + ], + "angle": 0, + "content": "- ASSISTANT: Sure, it is . In the image, the unusual part is the dog wearing a reindeer antler headband." + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.275, + 0.321, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.275, + 0.407, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.272, + 0.598, + 0.308 + ], + "angle": 0, + "content": "- USER: Who is the loser in this image? Please output segmentation mask and explain why." + }, + { + "type": "text", + "bbox": [ + 0.425, + 0.309, + 0.603, + 0.352 + ], + "angle": 0, + "content": "ASSISTANT: Sure, it is . ... The man on top is likely the winner of the match, as he has control over the other fighters body..." + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.282, + 0.739, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.282, + 0.876, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.366, + 0.248, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.383, + 0.503, + 0.41 + ], + "angle": 0, + "content": "- USER: Can you segment the fried food and the food with most protein separately in this image?" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.411, + 0.517, + 0.418 + ], + "angle": 0, + "content": "- ASSISTANT: Sure, they are and ." + }, + { + "type": "list", + "bbox": [ + 0.266, + 0.383, + 0.517, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.366, + 0.685, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.366, + 0.853, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.445, + 0.892, + 0.486 + ], + "angle": 0, + "content": "Figure 1. We unlock new segmentation capabilities for existing multimodal LLMs. Our model (i.e., LISA) can deal with cases involving complex reasoning and world knowledge. Also, we demonstrate the cases of explanatory answers in the 3rd row. Additionally, in the 4th row, our model can output multiple segmentation masks in a single answer. More illustrations can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.42, + 0.521 + ], + "angle": 0, + "content": "involving complex reasoning and world knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.642 + ], + "angle": 0, + "content": "In addition, to validate the effectiveness, we establish a benchmark for reasoning segmentation evaluation, called ReasonSeg. Comprising over one thousand image-instruction pairs, this benchmark offers persuasive evaluation metrics for the task. To align more closely with practical applications, we annotate the images from OpenImages [21] and ScanNetv2 [10] with implicit text queries that involve complex reasoning." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.643, + 0.397, + 0.657 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.658, + 0.47, + 0.718 + ], + "angle": 0, + "content": "- We introduce the reasoning segmentation task, which necessitates reasoning based on implicit human instructions. Such reasoning capability is crucial for building a genuinely intelligent perception system." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.725, + 0.47, + 0.816 + ], + "angle": 0, + "content": "- We present our model — LISA, which incorporates new segmentation capabilities. It demonstrates robust zero-shot ability on the reasoning segmentation task when trained solely on reasoning-free datasets, and achieves further performance boost by fine-tuning on just 239 data samples that involve reasoning." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.823, + 0.471, + 0.899 + ], + "angle": 0, + "content": "- We establish a reasoning segmentation benchmark, ReasonSeg, containing over one thousand image-instruction-mask data samples. This benchmark is essential for evaluation and encourages the community to further explore the reasoning ability for vision tasks." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.658, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.505, + 0.641, + 0.52 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.693, + 0.538 + ], + "angle": 0, + "content": "2.1. Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.894, + 0.75 + ], + "angle": 0, + "content": "Semantic segmentation aims to assign a class label to every pixel in an image. Numerous studies [2, 5, 8, 12, 16, 22, 31, 37, 42, 43, 45, 46, 51, 56, 59-61, 64] have proposed diverse designs (such as encoder-decoder, dilated convolution, pyramid pooling module, non-local operator, and more) to effectively encode semantic information. Research on instance segmentation [9, 14, 58] and panoptic segmentation [7, 18, 25, 50] has introduced various architectural innovations for instance-level segmentation, including DETR [4]-based structures, mask attention, and dynamic convolution. In recent years, typical segmentation tasks have made significant progress and become increasingly mature. Consequently, it is imperative to develop more intelligent interaction ways for image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The referring segmentation task [17, 36] enables interaction with human language, aiming to segment the target object based on a given explicit text description. Recently, Kirillov et al. [19] introduced SAM, trained with billions of high-quality masks, supporting bounding boxes and points as prompts while demonstrating exceptional segmentation quality. X-Decoder [65] bridges vision and language, unifying multiple tasks within a single model. SEEM [66] further supports various human interaction methods, including text, audio, and scribble. However, these studies primarily focus" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.094, + 0.528, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.223, + 0.452, + 0.248 + ], + "angle": 0, + "content": "the camera lens that is more suitable for photographing nearby objects" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.094, + 0.835, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.218, + 0.841, + 0.253 + ], + "angle": 0, + "content": "Toddlers are curious and often enjoy exploring their surroundings. What object in the picture can provide a safe and enclosed space for a toddler to play in?" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.894, + 0.286 + ], + "angle": 0, + "content": "Figure 2. Examples of the annotated image-instruction-mask data samples. Left: short phrase query. Right: long sentence query. More examples are given in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.358 + ], + "angle": 0, + "content": "on addressing multi-task compatibility and unification, neglecting the injection of new capabilities. In this work, we present LISA and it possesses reasoning ability that has not been explored yet in existing segmentors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.388, + 0.377 + ], + "angle": 0, + "content": "2.2. Multimodal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.473, + 0.82 + ], + "angle": 0, + "content": "Motivated by the remarkable reasoning abilities of LLMs, researchers are exploring ways to transfer these capabilities into the vision domain, developing multimodal LLMs. Flamingo [1] employs a cross-attention structure to attend to visual contexts, enabling visual in-context learning. Models such as BLIP-2 [24] and mPLUG-OWL [55] propose encoding image features with a visual encoder, which are then fed into the LLM alongside text embeddings. Otter [23] further incorporates robust few-shot capabilities through in-context instruction tuning on the proposed MIMIC-IT dataset. LLaVA [29] and MiniGPT-4 [63] first conduct image-text feature alignment followed by instruction tuning. Koh et al. [20] also investigates image retrieval for LLMs. Moreover, numerous works [32, 44, 49, 52, 54] utilize prompt engineering, connecting independent modules via API calls, but without the benefits of end-to-end training. Recently, there have been studies examining the intersection between multimodal LLMs and vision tasks. Vision-LLM [47] offers a flexible interaction interface for multiple vision-centric tasks through instruction tuning but fails to fully exploit LLMs for complex reasoning. Kosmos-2 [38] constructs large-scale data of grounded image-text pairs, infusing grounding capabilities into LLMs. DetGPT [39] bridges the fixed multimodal LLM and open-vocabulary detector, enabling detection to be performed based on user instruction. GPT4RoI [57] introduces spatial boxes as input and trains the model on region-text pairs. In contrast, our work aims to efficiently inject segmentation capabilities into multimodal LLMs in the manner of end-to-end training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.308, + 0.847 + ], + "angle": 0, + "content": "3. Reasoning Segmentation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.851, + 0.26, + 0.866 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The reasoning segmentation task is to output a binary segmentation mask \\(\\mathbf{M}\\), given an input image \\(\\mathbf{x}_{img}\\) and an im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.449 + ], + "angle": 0, + "content": "plicit query text instruction \\(\\mathbf{x}_{\\text{txt}}\\). The task shares a similar formulation with the referring segmentation task [17], but is far more challenging. The key distinction lies in the complexity of the query text in reasoning segmentation. Instead of a straightforward phrase (e.g., \"the trash can\"), the query text includes more intricate expressions (e.g., \"something that the garbage should be put into\") or longer sentences (e.g., \"After cooking, consuming food, and preparing for food, where can we throw away the rest of the food and scraps?\") that involve complex reasoning or world knowledge." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.457, + 0.627, + 0.471 + ], + "angle": 0, + "content": "3.2.Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.706 + ], + "angle": 0, + "content": "Given the lack of quantitative evaluation, it is imperative to establish a benchmark for the reasoning segmentation task. To ensure reliable assessment, we have collected a diverse set of images from OpenImages [21] and ScanNetv2 [10], annotating them with implicit text instructions and high-quality target masks. To cover different scenarios, our text instructions consist of two types: 1) short phrases; 2) long sentences; as illustrated in Figure 2. The resulting ReasonSeg benchmark comprises a total of 1218 image-instruction-mask data samples. This dataset is further partitioned into three splits: train, val, and test, containing 239, 200, and 779 data samples, respectively. As the primary purpose of the benchmark is evaluation, the validation and testing sets include a larger number of data samples. The details of data annotation are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.631, + 0.736 + ], + "angle": 0, + "content": "4. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.893, + 0.791 + ], + "angle": 0, + "content": "In this section, we first introduce the model architecture in Sec. 4.1. After that, we elaborate on the training data preparation and training parameters in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.634, + 0.816 + ], + "angle": 0, + "content": "4.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Embedding as Mask. Most current multimodal LLMs (such as LLaVA [29], Flamingo [1], BLIP-2 [24], Otter [23], etc.) support image and text as input, but they can only output text and cannot directly output fine-grained segmentation masks. VisionLLM [47] offers a solution by parsing" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.885, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.275, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. The pipeline of LISA. Given the input image and text query, the multimodal LLM (e.g., LLaVA [29]) generates text output. The last-layer embedding for the token is then decoded into the segmentation mask via the decoder. We use LoRA [15] for efficient fine-tuning. The choice of vision backbone can be flexible (e.g., SAM [66], Mask2Former [9])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.473, + 0.494 + ], + "angle": 0, + "content": "segmentation masks as sequences of polygons, enabling the representation of segmentation masks as plain text and allowing end-to-end training within the framework of existing multimodal LLMs. However, end-to-end training with the polygon sequences introduces optimization challenges and may compromise generalization ability unless a massive amount of data and computational resources are employed. For instance, training a 7B model, VisionLLM requires \\(4 \\times 8\\) NVIDIA 80G A100 GPUs and 50 epochs, which is computationally prohibitive. In contrast, it takes less than 3 days to train LISA-7B on 8 NVIDIA 24G 3090 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.474, + 0.631 + ], + "angle": 0, + "content": "To this end, we propose the embedding-as-mask paradigm to infuse new segmentation capabilities into the multimodal LLM. The pipeline of our method is illustrated in Fig. 3. Specifically, we first expand the original LLM vocabulary with a new token, i.e., \\(<\\mathrm{SEG}>\\), which signifies the request for the segmentation output. Given a text instruction \\(\\mathbf{x}_{txt}\\) along with the input image \\(\\mathbf{x}_{img}\\), we feed them into the multimodal LLM \\(\\mathcal{F}\\), which in turn outputs a text response \\(\\hat{\\mathbf{y}}_{txt}\\). It can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.638, + 0.47, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {t x t} = \\mathcal {F} (\\mathbf {x} _ {i m g}, \\mathbf {x} _ {t x t}). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.812 + ], + "angle": 0, + "content": "When the LLM intends to generate a binary segmentation mask, the output \\(\\hat{\\mathbf{y}}_{txt}\\) would include a token. We then extract the LLM last-layer embedding \\(\\tilde{\\mathbf{h}}_{seg}\\) corresponding to the token and apply an MLP projection layer \\(\\gamma\\) to obtain \\(\\mathbf{h}_{seg}\\). Simultaneously, the vision backbone \\(\\mathcal{F}_{enc}\\) extracts the dense visual features \\(\\mathbf{f}\\) from the visual input \\(\\mathbf{x}_{img}\\). Finally, \\(\\mathbf{h}_{seg}\\) and \\(\\mathbf{f}\\) are fed to the decoder \\(\\mathcal{F}_{dec}\\) to produce the final segmentation mask \\(\\hat{\\mathbf{M}}\\). The detailed structure of the decoder \\(\\mathcal{F}_{dec}\\) follows [19]. The process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.816, + 0.47, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {s e g} = \\gamma (\\tilde {\\mathbf {h}} _ {s e g}), \\quad \\mathbf {f} = \\mathcal {F} _ {e n c} (\\mathbf {x} _ {i m g}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.84, + 0.353, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\mathcal {F} _ {d e c} (\\mathbf {h} _ {s e g}, \\mathbf {f}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Training Objectives. The model is trained end-to-end using the text generation loss \\(\\mathcal{L}_{t:xt}\\) and the segmentation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.892, + 0.36 + ], + "angle": 0, + "content": "mask loss \\(\\mathcal{L}_{mask}\\). The overall objective \\(\\mathcal{L}\\) is the weighted sum of these losses, determined by \\(\\lambda_{txt}\\) and \\(\\lambda_{mask}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.372, + 0.892, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {m a s k} \\mathcal {L} _ {m a s k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.519 + ], + "angle": 0, + "content": "Specifically, \\(\\mathcal{L}_{txt}\\) is the auto-regressive cross-entropy loss for text generation, and \\(\\mathcal{L}_{mask}\\) is the mask loss, which encourages the model to produce high-quality segmentation results. To compute \\(\\mathcal{L}_{mask}\\), we employ a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights \\(\\lambda_{bce}\\) and \\(\\lambda_{dice}\\). Given the ground-truth targets \\(\\mathbf{y}_{txt}\\) and \\(\\mathbf{M}\\), these losses can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.53, + 0.775, + 0.546 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t x t} = \\mathbf {C E} (\\hat {\\mathbf {y}} _ {t x t}, \\mathbf {y} _ {t x t}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.544, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {\\mathbf {M}}, \\mathbf {M}) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {\\mathbf {M}}, \\mathbf {M}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.7 + ], + "angle": 0, + "content": "It is noteworthy that the proposed method endows existing multimodal LLMs with new segmentation capabilities, such that they can generate not only text but also fine-grained output formats. Also, our method is based on an end-to-end training pipeline and connects the LLM and vision modules with hidden embedding representation, which proves significantly more effective than the decoupled two-stage method as discussed in Sec. 5.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.709, + 0.605, + 0.725 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.894, + 0.791 + ], + "angle": 0, + "content": "Training Data Formulation. As illustrated in Fig. 4, our training data comprises mainly three parts, all of which are derived from widely-used public datasets. The details are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "- Semantic Segmentation Dataset. Semantic segmentation datasets typically consist of images and the corresponding multi-class labels. During training, we randomly choose several categories for each image. To generate data that matches the format of visual question answering, we employ a question-answer template like “USER: Can you segment the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9582" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.098, + 0.169, + 0.112 + ], + "angle": 0, + "content": "Raw Data" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.117, + 0.326, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.116, + 0.528, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.53, + 0.163, + 0.621, + 0.201 + ], + "angle": 0, + "content": "the lady with the blue shirt" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.12, + 0.865, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.259, + 0.208, + 0.272 + ], + "angle": 0, + "content": "Processed Data" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.283, + 0.296, + 0.317 + ], + "angle": 0, + "content": "- USER: Can you segment the table in this image?" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.318, + 0.308, + 0.329 + ], + "angle": 0, + "content": "- ASSISTANT: It is ." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.283, + 0.308, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.332, + 0.318, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.452, + 0.315, + 0.467 + ], + "angle": 0, + "content": "Semantic Segmentation Data" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.285, + 0.609, + 0.319 + ], + "angle": 0, + "content": "- USER: Can you segment the lady with the blue shirt in this image?" + }, + { + "type": "text", + "bbox": [ + 0.369, + 0.32, + 0.609, + 0.332 + ], + "angle": 0, + "content": "- ASSISTANT: Sure, it is ." + }, + { + "type": "list", + "bbox": [ + 0.369, + 0.285, + 0.609, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.334, + 0.488, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.452, + 0.593, + 0.467 + ], + "angle": 0, + "content": "Referring Segmentation Data" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.334, + 0.611, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.284, + 0.871, + 0.318 + ], + "angle": 0, + "content": "- USER: What type of sign and traffic device can be seen in the image? " + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.319, + 0.871, + 0.363 + ], + "angle": 0, + "content": "- ASSISTANT: In the image, there is a street sign and a traffic light above a city road." + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.364, + 0.862, + 0.396 + ], + "angle": 0, + "content": "- USER: How many traffic lights are visible in the image?" + }, + { + "type": "list", + "bbox": [ + 0.652, + 0.284, + 0.871, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.652, + 0.426, + 0.87, + 0.441 + ], + "angle": 0, + "content": "No Binary Segmentation Mask" + }, + { + "type": "title", + "bbox": [ + 0.721, + 0.452, + 0.798, + 0.465 + ], + "angle": 0, + "content": "VQA Data" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.476, + 0.893, + 0.505 + ], + "angle": 0, + "content": "Figure 4. The illustration of training data formulation from different types of data, including semantic segmentation data, referring segmentation data, and visual question answering (VQA) data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.473, + 0.66 + ], + "angle": 0, + "content": "{class_name} in this image? ASSISTANT: It is .\", where {class_name} is the chosen category, and denotes the placeholder for tokens of image patches. The corresponding binary segmentation mask is used as the ground truth to provide mask loss supervision. During training, we also use other templates to generate the QA data to ensure data diversity, as shown in the supplementary material. We adopt ADE20K, COCO-Stuff, and LVIS-PACO part segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.675, + 0.473, + 0.812 + ], + "angle": 0, + "content": "- Vanilla Referring Segmentation Dataset. Referring segmentation datasets provide an input image and an explicit short description of the target object. Thus, it is easy to convert them into question-answer pairs using a template like \"USER: Can you segment {description} in this image? ASSISTANT: Sure, it is .\", where {description} is the given explicit description. For this part, we adopt ref-COCO, refCOCO+, refCOCOg, and refCLEF datasets." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.827, + 0.472, + 0.888 + ], + "angle": 0, + "content": "- Visual Question Answering Dataset. To preserve the original Visual Question Answering (VQA) ability of the multimodal LLM, we also include the VQA dataset during training. We use LLaVA-Instruct-150k [29] for LLaVA v1" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.675, + 0.473, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.524, + 0.836, + 0.539 + ], + "angle": 0, + "content": "and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.895, + 0.683 + ], + "angle": 0, + "content": "Notably, the above datasets do not include any reasoning segmentation data sample. Instead, it only contains samples where the target objects are explicitly indicated in the query texts. Surprisingly, even without complex reasoning training data, LISA demonstrates impressive zero-shot ability on the ReasonSeg benchmark, as shown in Table 1. Moreover, we find that further performance boost could be yielded by finetuning the model on only 239 data samples that involve complex reasoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.895, + 0.809 + ], + "angle": 0, + "content": "Trainable Parameters. To preserve the learned knowledge of the pre-trained multimodal LLM \\(\\mathcal{F}\\) (i.e., LLaVA in our experiments), we leverage LoRA [15] to perform efficient fine-tuning, and completely freeze the vision backbone \\(\\mathcal{F}_{enc}\\). The decoder \\(\\mathcal{F}_{dec}\\) is fully fine-tuned. Additionally, the LLM token embeddings (embed_tokens), the LLM head (lm_head), and the projection layer \\(\\gamma\\) are also trainable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "It is notable that the resulting model avoids the catastrophic forgetting of the original text generation capability and preserves the conversation ability, as verified in the supplementary material. The potential reasons are: we 1) employ LoRA fine-tuning to reduce the trainable parameters and 2) incorporate the VQA dataset during fine-tuning." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9583" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.894, + 0.119 + ], + "angle": 0, + "content": "Table 1. Reasoning segmentation results among LISA (ours) and previous related works. 'ft' denotes using 239 reasoning segmentation data samples to fine-tune the model. Unless otherwise specified, we use LLaVA v1 [29] as the base model. LLaVA1.5 denotes LLaVA v1.5 [28]." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.122, + 0.868, + 0.424 + ], + "angle": 0, + "content": "
Methodvaltest
overallshort querylong queryoverall
gIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoU
OVSeg [26]28.518.618.015.528.722.526.120.8
GRES [27]22.419.917.615.022.623.821.322.0
X-Decoder [65]22.617.920.411.622.217.521.716.3
SEEM [66]25.521.220.111.525.620.824.318.7
Grounded-SAM [30]26.014.517.810.822.418.621.316.4
LISA-7B44.446.037.634.436.634.736.834.1
LISA-7B (ft)52.954.040.640.649.451.047.348.4
LISA-13B48.946.939.943.346.446.544.845.8
LISA-13B (ft)56.262.944.342.054.054.351.751.1
LLaVA1.5-7B + OVSeg38.223.524.218.744.637.139.731.8
LISA-7B-LLaVA1.553.652.347.148.549.248.948.748.8
LISA-7B-LLaVA1.5 (ft)61.362.948.346.357.959.755.656.9
LLaVA1.5-13B + OVSeg37.926.427.119.446.140.641.534.1
LISA-13B-LLaVA1.557.760.350.850.054.750.953.850.8
LISA-13B-LLaVA1.5 (ft)65.072.955.450.663.265.361.362.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.202, + 0.451 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.279, + 0.472 + ], + "angle": 0, + "content": "5.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.471, + 0.552 + ], + "angle": 0, + "content": "Network Architecture. Unless otherwise specified, we use LLaVA-7B-v1-1 or LLaVA-13B-v1-1 [29] as the base multimodal LLM \\(\\mathcal{F}\\), and adopt the ViT-H SAM [19] backbone as the vision backbone \\(\\mathcal{F}_{enc}\\). The projection layer of \\(\\gamma\\) is an MLP with channels of [256, 4096, 4096]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.472, + 0.756 + ], + "angle": 0, + "content": "Implementation Details. We adopt 8 NVIDIA 24G 3090 GPUs for training. The training scripts are based on deep-speed [41] engine. We use AdamW [33] optimizer with the learning rate and weight decay set to 0.0003 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the text generation loss \\(\\lambda_{txt}\\) and the mask loss \\(\\lambda_{mask}\\) are set to 1.0 and 1.0, respectively, and those of the bce loss \\(\\lambda_{bce}\\) and the dice loss \\(\\lambda_{dice}\\) are set to 2.0 and 0.5, respectively. Besides, the batch size per device is set to 2, and the gradient accumulation step is set to 10. During training, we select at most 3 categories for each image in semantic segmentation datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Datasets. As mentioned in Sec. 4.2, our training data is mainly composed of three types of datasets: (1) For the semantic segmentation dataset, we use ADE20K [62] and COCO-Stuff [3]. Besides, to enhance the segmentation result for some part of an object, we also use part semantic segmentation datasets, including PACO-LVIS [40], PartImageNet [13], and PASCAL-Part [6]; (2) For the referring segmentation dataset, we use refCLEF, refCOCO, refCOCO+ [17], and refCOCOg [35]; (3) For the visual ques" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.895, + 0.556 + ], + "angle": 0, + "content": "tion answering (VQA) dataset, we use the datasets of LLaVA-Instruct-150k for LLaVA v1 [29] and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]. In order to avoid data leakage, we exclude the COCO samples whose images are present in the refCOCO \\((+ / \\mathrm{g})\\) validation sets during training. Furthermore, we surprisingly find that by fine-tuning the model on only 239 ReasonSeg data samples, the model's performance can be further boosted." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.672 + ], + "angle": 0, + "content": "Evaluation Metrics. We follow most previous works on referring segmentation [17, 35] to adopt two metrics: gIoU and cIoU. gIoU is defined by the average of all per-image Intersection-over-Unions (IoUs), while cIoU is defined by the cumulative intersection over the cumulative union. Since cIoU is highly biased toward large-area objects and it fluctuates too much, gIoU is preferred." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.787, + 0.697 + ], + "angle": 0, + "content": "5.2. Reasoning Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.87 + ], + "angle": 0, + "content": "The reasoning segmentation results are shown in Table 1. It is worth noting that existing works fail to handle the task, but our model can accomplish the task involving complex reasoning with more than \\(20\\%\\) gIoU performance boost. As mentioned before, the reasoning segmentation task is essentially different from the referring segmentation task in that it requires the model to possess reasoning ability or access world knowledge. Only by truly understanding the query, can the model do well in the task. The existing works have no proper way to understand an implicit query, but our model exploits multimodal LLMs to reach the goal." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Notably, we also make a comparison with the vanilla two-stage method (LLaVA1.5 + OVSeg). Specifically, the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9584" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.22, + 0.089, + 0.747, + 0.104 + ], + "angle": 0, + "content": "Table 2. Referring segmentation results (cIoU) among LISA (ours) and existing methods." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.108, + 0.862, + 0.293 + ], + "angle": 0, + "content": "
MethodrefCOCOrefCOCO+refCOCOg
valtestAtestBvaltestAtestBval(U)test(U)
MCN [34]62.464.259.750.655.044.749.249.4
VLT [11]67.570.565.256.361.050.155.057.7
CRIS [48]70.573.266.162.368.153.759.960.4
LAVT [53]72.775.868.862.168.455.161.262.1
ReLA [27]73.876.570.266.071.057.765.066.0
X-Decoder [65]------64.6-
SEEM [66]------65.7-
LISA-7B74.176.571.162.467.456.566.468.5
LISA-7B (fine-tuned on ReferSeg)74.979.172.365.170.858.167.970.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.522 + ], + "angle": 0, + "content": "two-stage method refers to first using a multimodal LLM (e.g., LLaVA v1.5) to generate a text output for the input query, and then adopting a referring or open-vocabulary segmentation model (e.g., OVSeg) to generate the segmentation mask. If the intermediate text output remains too long and exceeds the input token length limit of OVSeg, we use GPT-3.5 to further summarize. More details can be found in the supplementary material. The results in Table 1 show that our model outperforms the two-stage method significantly. We explain that the potential reasons are: 1) Our model is trained end-to-end, while the two-stage method is completely decoupled; 2) The two-stage method relies on text as an intermediary to transmit information, while our model utilizes the hidden embedding that is more expressive." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.613 + ], + "angle": 0, + "content": "Another finding is that LISA-13B outperforms the 7B counterpart substantially, especially on the long-query scenarios, which indicates that the current performance bottleneck may still lie in understanding the query text, and a stronger multimodal LLM (e.g., LLaVA v1.5 [28]) leads to even better results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.622, + 0.418, + 0.639 + ], + "angle": 0, + "content": "5.3. Vanilla Referring Segmentation Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.737 + ], + "angle": 0, + "content": "To show that our model is also competent in the vanilla referring segmentation task, we make a comparison with existing state-of-the-art methods in Table 2. We evaluate the methods on refCOCO, refCOCO+, refCOCOg validation and testing sets. Our model achieves state-of-the-art results across various referring segmentation benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.746, + 0.231, + 0.763 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.471, + 0.83 + ], + "angle": 0, + "content": "In this section, we conduct an extensive ablation study to reveal the contribution of each component. Unless otherwise specified, we report the metrics of gIoU and cIoU of LISA-7B on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Design Choices of Vision Backbone. We emphasize that vision backbones other than SAM are also applicable in our framework. In Table 3, we notice that SAM performs the best, potentially because of the massive high-quality" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.309, + 0.894, + 0.338 + ], + "angle": 0, + "content": "Table 3. Ablation study on the design choice of vision backbone. 'ft' denotes finetuning on ReasonSeg training set." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.348, + 0.878, + 0.453 + ], + "angle": 0, + "content": "
Vision BackbonegIoUcIoU
Mask2Former-Swin-L42.438.8
SAM (w/ LoRA)41.537.3
SAM44.446.0
Mask2Former-Swin-L (ft)50.752.3
SAM w/ LORA (ft)51.851.9
SAM (ft)52.954.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.469, + 0.89, + 0.484 + ], + "angle": 0, + "content": "Table 4. Ablation study on SAM pre-trained weight and rephrasing." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.494, + 0.86, + 0.565 + ], + "angle": 0, + "content": "
Exp. IDPre-trainSAM γrephrasinggIoUcIoU
1✓35.944.6
2✓50.751.1
3✓✓52.954.0
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.894, + 0.683 + ], + "angle": 0, + "content": "data used in its pre-training phase. Further, we also find that with the Mask2Former backbone, our framework still achieves a decent performance on the reasoning segmentation task, significantly outperforming previous works such as X-Decoder [65]. This reveals the fact that the design choice of vision backbone is flexible and not limited to SAM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.894, + 0.794 + ], + "angle": 0, + "content": "SAM LoRA Fintuning. We also investigate the effectiveness of applying LoRA on the SAM backbone. In Table 3, we note that the performance of LoRA fine-tuned SAM backbone is inferior to that of the frozen one. A potential reason is that fine-tuning impairs the generalization ability of the original SAM model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.815, + 0.894, + 0.906 + ], + "angle": 0, + "content": "SAM Pre-trained Weight. To demonstrate the contribution of SAM pre-trained weight, we make a comparison between Experiments 1 and 3 in Table 4. Without being initialized with SAM pre-trained weight, the vision backbone is trained from scratch. This causes the performance to fall substantially behind that of the baseline model." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9585" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.094, + 0.136, + 0.117 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.108, + 0.242, + 0.136 + ], + "angle": 0, + "content": "the most likely object that someone else has left behind" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.098, + 0.334, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.094, + 0.371, + 0.118 + ], + "angle": 0, + "content": "中" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.098, + 0.865, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.161, + 0.136, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.18, + 0.242, + 0.2 + ], + "angle": 0, + "content": "the object used for stirring milk or coffee" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.162, + 0.334, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.159, + 0.371, + 0.184 + ], + "angle": 0, + "content": "W" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.162, + 0.865, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.228, + 0.136, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.232, + 0.255, + 0.28 + ], + "angle": 0, + "content": "In cold days, dogs may need extra protection to keep them warm. What object in the picture can a dog wear to provide warmth during snowy walks?" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.235, + 0.334, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.226, + 0.371, + 0.251 + ], + "angle": 0, + "content": "中" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.235, + 0.865, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.293, + 0.136, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.301, + 0.257, + 0.337 + ], + "angle": 0, + "content": "When a plane is ready to land on the airport runway, what area in the picture will it eventually land on?" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.293, + 0.334, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.292, + 0.371, + 0.316 + ], + "angle": 0, + "content": "中" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.294, + 0.865, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.354, + 0.211, + 0.364 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.354, + 0.311, + 0.364 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.365, + 0.891, + 0.38 + ], + "angle": 0, + "content": "Figure 5. Visual comparison among LISA (ours) and existing related methods. More illustrations are given in the supplementary material." + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.387, + 0.394, + 0.4 + ], + "angle": 0, + "content": "Table 5. Ablation study on training data." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.405, + 0.472, + 0.563 + ], + "angle": 0, + "content": "
IDSemanticSegReferSegVQAReasonSeggIoUcIoU
ADE20KCOCO-StuffPartSeg
1✓✓✓✓✓48.953.5
2✓✓✓✓✓48.550.8
3✓✓✓✓✓46.750.9
4✓✓✓✓46.646.7
5✓✓✓30.420.4
6✓✓✓✓✓47.751.1
7✓✓✓✓✓44.446.0
8✓✓✓✓✓✓52.954.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.573, + 0.399, + 0.587 + ], + "angle": 0, + "content": "Table 6. Results on the ReasonSeg test set." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.59, + 0.43, + 0.643 + ], + "angle": 0, + "content": "
Training splits# data samplesgIoUcIoU
train23951.751.1
train + val43954.054.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.47, + 0.778 + ], + "angle": 0, + "content": "Instruction Rephrasing by GPT-3.5. When fine-tuning the model on the reasoning segmentation data samples, we rephrase the text instruction by GPT-3.5 (the details are shown in the supplementary material), and randomly choose one. The comparison between Experiments 2 and 3 in Table 4 shows that the performance is increased by \\(2.2\\%\\) gIoU and \\(2.9\\%\\) cIoU. This result verifies the effectiveness of such data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Contribution of All Types of Training Data. In Table 5, we show the contribution of each type of data to the performance. We find that in Exp. 5, we do not use any semantic segmentation dataset, and the performance drops a lot. We conjecture that semantic segmentation datasets provide a large amount of ground-truth binary masks for training, since a multi-class label can induce multiple binary masks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.48 + ], + "angle": 0, + "content": "We also notice that adding more reasoning segmentation data samples during training leads to better results. In Table 6, we also add the ReasonSeg val set (200 data samples) during fine-tuning, and it yields better performance in both gIoU and cIoU metrics. This indicates that more reasoning segmentation training samples are beneficial at this moment." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.489, + 0.683, + 0.505 + ], + "angle": 0, + "content": "5.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.635 + ], + "angle": 0, + "content": "As depicted in Fig. 5, we provide a visual comparison with existing related works, including the model for open-vocabulary semantic segmentation (OVSeg), referring segmentation (GRES), and the generalist models for segmentation (X-Decoder and SEEM). These models fail to handle the displayed cases with various errors, while our approach produces accurate and high-quality segmentation results. More illustrations are given in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.619, + 0.665 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.894, + 0.81 + ], + "angle": 0, + "content": "In this work, we have proposed a new segmentation task—reasoning segmentation. Also, we have introduced an evaluation benchmark ReasonSeg, which comprises over one thousand data samples. Finally, we have presented our model — LISA. It injects segmentation capabilities into current multimodal LLMs and performs surprisingly effectively on the reasoning segmentation task. We hope our work can shed new light on the direction of combining LLMs and vision tasks in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.819, + 0.667, + 0.836 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work is supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9586" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.184 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.229 + ], + "angle": 0, + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.232, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.275, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.472, + 0.447 + ], + "angle": 0, + "content": "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.45, + 0.472, + 0.506 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.552, + 0.472, + 0.606 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.472, + 0.665 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.668, + 0.472, + 0.71 + ], + "angle": 0, + "content": "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.472, + 0.755 + ], + "angle": 0, + "content": "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[16] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.894, + 0.398 + ], + "angle": 0, + "content": "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[22] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.565 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.893, + 0.608 + ], + "angle": 0, + "content": "[25] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.893, + 0.663 + ], + "angle": 0, + "content": "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In CVPR, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.893, + 0.748 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint, 2023. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv:2304.08485, 2023. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pretraining for open-set object detection. arXiv preprint, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.894, + 0.899 + ], + "angle": 0, + "content": "[31] Wei Liu, Andrew Rabinovich, and Alexander C. Berg. Parsenet: Looking wider to see better. arXiv preprint, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[32] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Internchat: Solving vision-centric tasks by interacting with chatbots beyond language. arXiv:2305.05662, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[34] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.472, + 0.306 + ], + "angle": 0, + "content": "[35] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.472, + 0.351 + ], + "angle": 0, + "content": "[36] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[37] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.472, + 0.452 + ], + "angle": 0, + "content": "[38] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.472, + 0.509 + ], + "angle": 0, + "content": "[39] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, and Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv:2305.14167, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.567 + ], + "angle": 0, + "content": "[40] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.472, + 0.624 + ], + "angle": 0, + "content": "[41] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In SIGKDD, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.627, + 0.472, + 0.668 + ], + "angle": 0, + "content": "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.671, + 0.472, + 0.711 + ], + "angle": 0, + "content": "[43] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. TPAMI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.472, + 0.769 + ], + "angle": 0, + "content": "[44] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv:2303.17580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[45] Zhuotao Tian, Pengguang Chen, Xin Lai, Li Jiang, Shu Liu, Hengshuang Zhao, Bei Yu, Ming-Chang Yang, and Jiaya Jia. Adaptive perspective distillation for semantic segmentation. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[46] Zhuotao Tian, Jiequan Cui, Li Jiang, Xiaojuan Qi, Xin Lai, Yixin Chen, Shu Liu, and Jiaya Jia. Learning context-aware classifier for semantic segmentation. AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv:2305.11175, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[48] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[49] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[50] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[51] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[52] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[54] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv:2303.11381, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.894, + 0.558 + ], + "angle": 0, + "content": "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[56] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In ICLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.589, + 0.894, + 0.644 + ], + "angle": 0, + "content": "[57] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.646, + 0.894, + 0.686 + ], + "angle": 0, + "content": "[58] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[59] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[60] Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia. Icnet for real-time semantic segmentation on high-resolution images. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.894, + 0.828 + ], + "angle": 0, + "content": "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[63] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "language understanding with advanced large language models. arXiv:2304.10592, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.163 + ], + "angle": 0, + "content": "[64] Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, and Xi-ang Bai. Asymmetric non-local neural networks for semantic segmentation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.472, + 0.22 + ], + "angle": 0, + "content": "[65] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[66] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 2, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9589" + } + ] +] \ No newline at end of file diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_origin.pdf b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0d6a2b5e8a22bb70e6d2365cb8e1f2844e44abe --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/92907404-c48e-45be-bfa1-4c27e2c673e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:658a4c2f28bcd0d8cedf749d4f292fda4f4eb95d44a877ef85cfb80828eafd9b +size 5181310 diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/full.md b/2024/LISA_ Reasoning Segmentation via Large Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e78590b67a563a0b2dea19d9bfad661a46f01a1 --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/full.md @@ -0,0 +1,412 @@ +# LISA: Reasoning Segmentation via Large Language Model + +Xin Lai $^{1*}$ Zhuotao Tian $^{2*†}$ Yukang Chen $^{1}$ Yanwei Li $^{1}$ Yuhui Yuan $^{4}$ Shu Liu $^{3}$ Jiaya Jia $^{1,3}$ $^{1}\mathrm{CUHK}$ $^{2}\mathrm{HIT}$ (Shenzhen) $^{3}\mathrm{SmartMore}$ $^{4}\mathrm{MSRA}$ + +# Abstract + +Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task — reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlabresearch/LISA. + +# 1. Introduction + +In daily life, users tend to issue direct commands like "Change the TV channel" to instruct a robot, rather than providing explicit step-by-step instructions such as "Go to the table first, find the TV remote, and then press the button to change the channel." However, existing perception systems consistently rely on humans to explicitly indicate target objects or pre-define categories before executing visual recog + +nition tasks. These systems cannot actively reason and comprehend user intention based on implicit instruction. This reasoning ability is crucial in developing next-generation intelligent perception systems and holds substantial potential for industrial applications, particularly in robotics. + +In this work, we introduce a new segmentation task — reasoning segmentation, which requires generating a binary segmentation mask based on an implicit query text involving complex reasoning. Notably, the query text is not limited to a straightforward reference (e.g., “the orange”), but a more complicated description involving complex reasoning or world knowledge (e.g., “the food with high Vitamin C”). To accomplish this task, the model must possess two key abilities: 1) reasoning complex and implicit text queries jointly with the image; 2) producing segmentation masks. + +Inspired by the exceptional capacity of LLMs to reason and comprehend user intentions, we aim to leverage this capability of LLMs to address the aforementioned first challenge. However, while several studies [1, 23, 24, 28, 29, 55, 63] have integrated robust reasoning capabilities into multimodal LLMs to accommodate visual input, the majority of these models primarily concentrate on text generation tasks and still fall short in performing vision tasks that require fine-grained output formats, such as segmentation masks. This leads us to ask: can we enable multimodal LLMs with the capability to output segmentation masks? + +To this end, we introduce LISA: a large Language Instructed Segmentation Assistant, a multimodal LLM capable of producing segmentation masks. Specifically, we incorporate an additional token, i.e., $<\mathrm{SEG}>$ , into the existing vocabulary. Upon generating the $<\mathrm{SEG}>$ token, its hidden embedding is further decoded into the corresponding segmentation mask. By representing the segmentation mask as an embedding, LISA acquires segmentation capabilities and benefits from end-to-end training. Remarkably, LISA demonstrates robust zero-shot abilities. Training the model solely on standard semantic segmentation and referring segmentation datasets yields surprisingly effective performance on the reasoning segmentation task. Furthermore, we find that LISA's performance can be significantly enhanced by fine-tuning on just 239 reasoning segmentation data samples. As illustrated in Fig. 1, LISA can handle various scenarios + +- USER: Can you segment the tyre that does not touch the ground in this image? +- ASSISTANT: Sure, it is . + +![](images/09a3adad0cc374de132e4ae29b4d128db4221c4e9b7805ee97e6bf3937620096.jpg) + +![](images/8c2512837cd9c28e35a9d8df1bd7a96f2a60197d49abc8435c9a7db3fc375b0d.jpg) + +- USER: There are two washing machines as shown in the picture. If I need to do laundry, where in the picture would I put the clothes? Please output segmentation mask. + +- ASSISTANT: . + +![](images/a0acaf7c6d33ce0213009a977dd6a4baeed984e74de62ec003de6713e7231b5e.jpg) + +![](images/2a2de191a1037497bef85eda8d3a180fdaa0f5d5a92fe0ab13f94fad1426666d.jpg) + +- USER: Can you segment the food that tastes not spicy in this image? + +ASSISTANT: . + +![](images/a9e153b9ee7022c45bfc37405011eb746e7ed5eebd5f46b0d906570cc76b84ee.jpg) + +![](images/f3de821f41bd3d8a3c7e8be9db2977b383841638d5eaaa3ce280d23e269fb10c.jpg) + +- USER: What is the place where the driver can observe the speed in this image? Please output segmentation mask. + +- ASSISTANT: . + +![](images/50f0ba0edf6a0fe1d788787241b99590ccc63f93f7863edfc16c83a426a3bad8.jpg) + +![](images/2967c4020606cbc77806987d8ce1af3b97b0538a50ba8dded645cd5e57b5682d.jpg) + +- USER: Can you segment the unusual part in this image and explain why? + +- ASSISTANT: Sure, it is . In the image, the unusual part is the dog wearing a reindeer antler headband. + +![](images/59801feb73fb53abce28f395d5ad8c119cf6343348faacb2e08a0b90febf0e29.jpg) + +![](images/cd12e4cd63ab23c9db1b831006e4db947ae3af78d355fdcb1536cfd33f62ad08.jpg) + +- USER: Who is the loser in this image? Please output segmentation mask and explain why. + +ASSISTANT: Sure, it is . ... The man on top is likely the winner of the match, as he has control over the other fighters body... + +![](images/5326720e924ea3944f1bfa50741ccd7a1470ab666e691cf9d6c8fdf26dc5a680.jpg) + +![](images/b2b7b9a4ea969654f4ed89b67dff6bdd74f46441148c1cea41f1303bd611336c.jpg) + +![](images/e31ac66fccd7e539d1b84f13c5c37b77ac96e48e4ec8bb7852fe67ec0bd0cadf.jpg) +Figure 1. We unlock new segmentation capabilities for existing multimodal LLMs. Our model (i.e., LISA) can deal with cases involving complex reasoning and world knowledge. Also, we demonstrate the cases of explanatory answers in the 3rd row. Additionally, in the 4th row, our model can output multiple segmentation masks in a single answer. More illustrations can be found in the supplementary material. + +- USER: Can you segment the fried food and the food with most protein separately in this image? +- ASSISTANT: Sure, they are and . + +![](images/b1344c5d2785ec033c0b1f52778c5511f38768875089b8ee10cce644455c8827.jpg) + +![](images/9cb3dffa2988a67e34d8ef8948518983538841f951e9231e86e7795b31b78414.jpg) + +involving complex reasoning and world knowledge. + +In addition, to validate the effectiveness, we establish a benchmark for reasoning segmentation evaluation, called ReasonSeg. Comprising over one thousand image-instruction pairs, this benchmark offers persuasive evaluation metrics for the task. To align more closely with practical applications, we annotate the images from OpenImages [21] and ScanNetv2 [10] with implicit text queries that involve complex reasoning. + +In summary, our contributions are as follows: + +- We introduce the reasoning segmentation task, which necessitates reasoning based on implicit human instructions. Such reasoning capability is crucial for building a genuinely intelligent perception system. +- We present our model — LISA, which incorporates new segmentation capabilities. It demonstrates robust zero-shot ability on the reasoning segmentation task when trained solely on reasoning-free datasets, and achieves further performance boost by fine-tuning on just 239 data samples that involve reasoning. +- We establish a reasoning segmentation benchmark, ReasonSeg, containing over one thousand image-instruction-mask data samples. This benchmark is essential for evaluation and encourages the community to further explore the reasoning ability for vision tasks. + +# 2. Related Work + +# 2.1. Image Segmentation + +Semantic segmentation aims to assign a class label to every pixel in an image. Numerous studies [2, 5, 8, 12, 16, 22, 31, 37, 42, 43, 45, 46, 51, 56, 59-61, 64] have proposed diverse designs (such as encoder-decoder, dilated convolution, pyramid pooling module, non-local operator, and more) to effectively encode semantic information. Research on instance segmentation [9, 14, 58] and panoptic segmentation [7, 18, 25, 50] has introduced various architectural innovations for instance-level segmentation, including DETR [4]-based structures, mask attention, and dynamic convolution. In recent years, typical segmentation tasks have made significant progress and become increasingly mature. Consequently, it is imperative to develop more intelligent interaction ways for image segmentation. + +The referring segmentation task [17, 36] enables interaction with human language, aiming to segment the target object based on a given explicit text description. Recently, Kirillov et al. [19] introduced SAM, trained with billions of high-quality masks, supporting bounding boxes and points as prompts while demonstrating exceptional segmentation quality. X-Decoder [65] bridges vision and language, unifying multiple tasks within a single model. SEEM [66] further supports various human interaction methods, including text, audio, and scribble. However, these studies primarily focus + +![](images/56991a6f3a5e5f8ce80b50a1d4b30070d5e2947b1d711ba4a144318e9a456b98.jpg) +the camera lens that is more suitable for photographing nearby objects +Toddlers are curious and often enjoy exploring their surroundings. What object in the picture can provide a safe and enclosed space for a toddler to play in? + +![](images/2724632b9d752ec872d52434f95ac729991821fa8e65b14932cacaa6f44cb043.jpg) +Figure 2. Examples of the annotated image-instruction-mask data samples. Left: short phrase query. Right: long sentence query. More examples are given in the supplementary material. + +on addressing multi-task compatibility and unification, neglecting the injection of new capabilities. In this work, we present LISA and it possesses reasoning ability that has not been explored yet in existing segmentors. + +# 2.2. Multimodal Large Language Model + +Motivated by the remarkable reasoning abilities of LLMs, researchers are exploring ways to transfer these capabilities into the vision domain, developing multimodal LLMs. Flamingo [1] employs a cross-attention structure to attend to visual contexts, enabling visual in-context learning. Models such as BLIP-2 [24] and mPLUG-OWL [55] propose encoding image features with a visual encoder, which are then fed into the LLM alongside text embeddings. Otter [23] further incorporates robust few-shot capabilities through in-context instruction tuning on the proposed MIMIC-IT dataset. LLaVA [29] and MiniGPT-4 [63] first conduct image-text feature alignment followed by instruction tuning. Koh et al. [20] also investigates image retrieval for LLMs. Moreover, numerous works [32, 44, 49, 52, 54] utilize prompt engineering, connecting independent modules via API calls, but without the benefits of end-to-end training. Recently, there have been studies examining the intersection between multimodal LLMs and vision tasks. Vision-LLM [47] offers a flexible interaction interface for multiple vision-centric tasks through instruction tuning but fails to fully exploit LLMs for complex reasoning. Kosmos-2 [38] constructs large-scale data of grounded image-text pairs, infusing grounding capabilities into LLMs. DetGPT [39] bridges the fixed multimodal LLM and open-vocabulary detector, enabling detection to be performed based on user instruction. GPT4RoI [57] introduces spatial boxes as input and trains the model on region-text pairs. In contrast, our work aims to efficiently inject segmentation capabilities into multimodal LLMs in the manner of end-to-end training. + +# 3. Reasoning Segmentation + +# 3.1. Problem Definition + +The reasoning segmentation task is to output a binary segmentation mask $\mathbf{M}$ , given an input image $\mathbf{x}_{img}$ and an im + +plicit query text instruction $\mathbf{x}_{\text{txt}}$ . The task shares a similar formulation with the referring segmentation task [17], but is far more challenging. The key distinction lies in the complexity of the query text in reasoning segmentation. Instead of a straightforward phrase (e.g., "the trash can"), the query text includes more intricate expressions (e.g., "something that the garbage should be put into") or longer sentences (e.g., "After cooking, consuming food, and preparing for food, where can we throw away the rest of the food and scraps?") that involve complex reasoning or world knowledge. + +# 3.2.Benchmark + +Given the lack of quantitative evaluation, it is imperative to establish a benchmark for the reasoning segmentation task. To ensure reliable assessment, we have collected a diverse set of images from OpenImages [21] and ScanNetv2 [10], annotating them with implicit text instructions and high-quality target masks. To cover different scenarios, our text instructions consist of two types: 1) short phrases; 2) long sentences; as illustrated in Figure 2. The resulting ReasonSeg benchmark comprises a total of 1218 image-instruction-mask data samples. This dataset is further partitioned into three splits: train, val, and test, containing 239, 200, and 779 data samples, respectively. As the primary purpose of the benchmark is evaluation, the validation and testing sets include a larger number of data samples. The details of data annotation are given in the supplementary material. + +# 4. Our Method + +In this section, we first introduce the model architecture in Sec. 4.1. After that, we elaborate on the training data preparation and training parameters in Sec. 4.2. + +# 4.1. Architecture + +Embedding as Mask. Most current multimodal LLMs (such as LLaVA [29], Flamingo [1], BLIP-2 [24], Otter [23], etc.) support image and text as input, but they can only output text and cannot directly output fine-grained segmentation masks. VisionLLM [47] offers a solution by parsing + +![](images/05eed35d4fb8518cd038dea284bec7f20db5651948f11f6d0db41be165bc0d40.jpg) +Figure 3. The pipeline of LISA. Given the input image and text query, the multimodal LLM (e.g., LLaVA [29]) generates text output. The last-layer embedding for the token is then decoded into the segmentation mask via the decoder. We use LoRA [15] for efficient fine-tuning. The choice of vision backbone can be flexible (e.g., SAM [66], Mask2Former [9]). + +segmentation masks as sequences of polygons, enabling the representation of segmentation masks as plain text and allowing end-to-end training within the framework of existing multimodal LLMs. However, end-to-end training with the polygon sequences introduces optimization challenges and may compromise generalization ability unless a massive amount of data and computational resources are employed. For instance, training a 7B model, VisionLLM requires $4 \times 8$ NVIDIA 80G A100 GPUs and 50 epochs, which is computationally prohibitive. In contrast, it takes less than 3 days to train LISA-7B on 8 NVIDIA 24G 3090 GPUs. + +To this end, we propose the embedding-as-mask paradigm to infuse new segmentation capabilities into the multimodal LLM. The pipeline of our method is illustrated in Fig. 3. Specifically, we first expand the original LLM vocabulary with a new token, i.e., $<\mathrm{SEG}>$ , which signifies the request for the segmentation output. Given a text instruction $\mathbf{x}_{txt}$ along with the input image $\mathbf{x}_{img}$ , we feed them into the multimodal LLM $\mathcal{F}$ , which in turn outputs a text response $\hat{\mathbf{y}}_{txt}$ . It can be formulated as + +$$ +\hat {\mathbf {y}} _ {t x t} = \mathcal {F} (\mathbf {x} _ {i m g}, \mathbf {x} _ {t x t}). \tag {1} +$$ + +When the LLM intends to generate a binary segmentation mask, the output $\hat{\mathbf{y}}_{txt}$ would include a token. We then extract the LLM last-layer embedding $\tilde{\mathbf{h}}_{seg}$ corresponding to the token and apply an MLP projection layer $\gamma$ to obtain $\mathbf{h}_{seg}$ . Simultaneously, the vision backbone $\mathcal{F}_{enc}$ extracts the dense visual features $\mathbf{f}$ from the visual input $\mathbf{x}_{img}$ . Finally, $\mathbf{h}_{seg}$ and $\mathbf{f}$ are fed to the decoder $\mathcal{F}_{dec}$ to produce the final segmentation mask $\hat{\mathbf{M}}$ . The detailed structure of the decoder $\mathcal{F}_{dec}$ follows [19]. The process can be formulated as + +$$ +\mathbf {h} _ {s e g} = \gamma (\tilde {\mathbf {h}} _ {s e g}), \quad \mathbf {f} = \mathcal {F} _ {e n c} (\mathbf {x} _ {i m g}), \tag {2} +$$ + +$$ +\mathbf {M} = \mathcal {F} _ {d e c} (\mathbf {h} _ {s e g}, \mathbf {f}). +$$ + +Training Objectives. The model is trained end-to-end using the text generation loss $\mathcal{L}_{t:xt}$ and the segmentation + +mask loss $\mathcal{L}_{mask}$ . The overall objective $\mathcal{L}$ is the weighted sum of these losses, determined by $\lambda_{txt}$ and $\lambda_{mask}$ : + +$$ +\mathcal {L} = \lambda_ {t x t} \mathcal {L} _ {t x t} + \lambda_ {m a s k} \mathcal {L} _ {m a s k}. \tag {3} +$$ + +Specifically, $\mathcal{L}_{txt}$ is the auto-regressive cross-entropy loss for text generation, and $\mathcal{L}_{mask}$ is the mask loss, which encourages the model to produce high-quality segmentation results. To compute $\mathcal{L}_{mask}$ , we employ a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights $\lambda_{bce}$ and $\lambda_{dice}$ . Given the ground-truth targets $\mathbf{y}_{txt}$ and $\mathbf{M}$ , these losses can be formulated as + +$$ +\mathcal {L} _ {t x t} = \mathbf {C E} (\hat {\mathbf {y}} _ {t x t}, \mathbf {y} _ {t x t}), +$$ + +$$ +\mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathbf {B C E} (\hat {\mathbf {M}}, \mathbf {M}) + \lambda_ {d i c e} \mathbf {D I C E} (\hat {\mathbf {M}}, \mathbf {M}). \tag {4} +$$ + +It is noteworthy that the proposed method endows existing multimodal LLMs with new segmentation capabilities, such that they can generate not only text but also fine-grained output formats. Also, our method is based on an end-to-end training pipeline and connects the LLM and vision modules with hidden embedding representation, which proves significantly more effective than the decoupled two-stage method as discussed in Sec. 5.2. + +# 4.2. Training + +Training Data Formulation. As illustrated in Fig. 4, our training data comprises mainly three parts, all of which are derived from widely-used public datasets. The details are as follows: + +- Semantic Segmentation Dataset. Semantic segmentation datasets typically consist of images and the corresponding multi-class labels. During training, we randomly choose several categories for each image. To generate data that matches the format of visual question answering, we employ a question-answer template like “USER: Can you segment the + +![](images/32a6e91c80caff3657c829b697f482f86fbdda77300dfd4a607cdb056d0b3ad3.jpg) +Raw Data + +![](images/76b5ed7bce4ae8a94540d47033c18525144f1246515c1ad6f70c7c165a7e198a.jpg) +the lady with the blue shirt + +![](images/39526b71af96bda5721d1b15aaf424999785330d0d23cc8ab59d463f0fb3fd77.jpg) + +# Processed Data + +- USER: Can you segment the table in this image? +- ASSISTANT: It is . + +![](images/124555f612515343327fde250c264861e45415c7bef15716f24116c05bb4231f.jpg) +Semantic Segmentation Data + +- USER: Can you segment the lady with the blue shirt in this image? +- ASSISTANT: Sure, it is . + +![](images/cf130f8d06ee9fff521df64a51723f34eb4436252b27baeef66e62594133abc1.jpg) +Referring Segmentation Data + +![](images/0df82e6b4af35777b0ad654250a2e259e11491570087d0397d34e4573655db38.jpg) +Figure 4. The illustration of training data formulation from different types of data, including semantic segmentation data, referring segmentation data, and visual question answering (VQA) data. + +- USER: What type of sign and traffic device can be seen in the image? +- ASSISTANT: In the image, there is a street sign and a traffic light above a city road. +- USER: How many traffic lights are visible in the image? + +# No Binary Segmentation Mask + +# VQA Data + +{class_name} in this image? ASSISTANT: It is .", where {class_name} is the chosen category, and denotes the placeholder for tokens of image patches. The corresponding binary segmentation mask is used as the ground truth to provide mask loss supervision. During training, we also use other templates to generate the QA data to ensure data diversity, as shown in the supplementary material. We adopt ADE20K, COCO-Stuff, and LVIS-PACO part segmentation datasets. + +- Vanilla Referring Segmentation Dataset. Referring segmentation datasets provide an input image and an explicit short description of the target object. Thus, it is easy to convert them into question-answer pairs using a template like "USER: Can you segment {description} in this image? ASSISTANT: Sure, it is .", where {description} is the given explicit description. For this part, we adopt ref-COCO, refCOCO+, refCOCOg, and refCLEF datasets. +- Visual Question Answering Dataset. To preserve the original Visual Question Answering (VQA) ability of the multimodal LLM, we also include the VQA dataset during training. We use LLaVA-Instruct-150k [29] for LLaVA v1 + +and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]. + +Notably, the above datasets do not include any reasoning segmentation data sample. Instead, it only contains samples where the target objects are explicitly indicated in the query texts. Surprisingly, even without complex reasoning training data, LISA demonstrates impressive zero-shot ability on the ReasonSeg benchmark, as shown in Table 1. Moreover, we find that further performance boost could be yielded by finetuning the model on only 239 data samples that involve complex reasoning. + +Trainable Parameters. To preserve the learned knowledge of the pre-trained multimodal LLM $\mathcal{F}$ (i.e., LLaVA in our experiments), we leverage LoRA [15] to perform efficient fine-tuning, and completely freeze the vision backbone $\mathcal{F}_{enc}$ . The decoder $\mathcal{F}_{dec}$ is fully fine-tuned. Additionally, the LLM token embeddings (embed_tokens), the LLM head (lm_head), and the projection layer $\gamma$ are also trainable. + +It is notable that the resulting model avoids the catastrophic forgetting of the original text generation capability and preserves the conversation ability, as verified in the supplementary material. The potential reasons are: we 1) employ LoRA fine-tuning to reduce the trainable parameters and 2) incorporate the VQA dataset during fine-tuning. + +Table 1. Reasoning segmentation results among LISA (ours) and previous related works. 'ft' denotes using 239 reasoning segmentation data samples to fine-tune the model. Unless otherwise specified, we use LLaVA v1 [29] as the base model. LLaVA1.5 denotes LLaVA v1.5 [28]. + +
Methodvaltest
overallshort querylong queryoverall
gIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoU
OVSeg [26]28.518.618.015.528.722.526.120.8
GRES [27]22.419.917.615.022.623.821.322.0
X-Decoder [65]22.617.920.411.622.217.521.716.3
SEEM [66]25.521.220.111.525.620.824.318.7
Grounded-SAM [30]26.014.517.810.822.418.621.316.4
LISA-7B44.446.037.634.436.634.736.834.1
LISA-7B (ft)52.954.040.640.649.451.047.348.4
LISA-13B48.946.939.943.346.446.544.845.8
LISA-13B (ft)56.262.944.342.054.054.351.751.1
LLaVA1.5-7B + OVSeg38.223.524.218.744.637.139.731.8
LISA-7B-LLaVA1.553.652.347.148.549.248.948.748.8
LISA-7B-LLaVA1.5 (ft)61.362.948.346.357.959.755.656.9
LLaVA1.5-13B + OVSeg37.926.427.119.446.140.641.534.1
LISA-13B-LLaVA1.557.760.350.850.054.750.953.850.8
LISA-13B-LLaVA1.5 (ft)65.072.955.450.663.265.361.362.2
+ +# 5. Experiment + +# 5.1. Experimental Setting + +Network Architecture. Unless otherwise specified, we use LLaVA-7B-v1-1 or LLaVA-13B-v1-1 [29] as the base multimodal LLM $\mathcal{F}$ , and adopt the ViT-H SAM [19] backbone as the vision backbone $\mathcal{F}_{enc}$ . The projection layer of $\gamma$ is an MLP with channels of [256, 4096, 4096]. + +Implementation Details. We adopt 8 NVIDIA 24G 3090 GPUs for training. The training scripts are based on deep-speed [41] engine. We use AdamW [33] optimizer with the learning rate and weight decay set to 0.0003 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the text generation loss $\lambda_{txt}$ and the mask loss $\lambda_{mask}$ are set to 1.0 and 1.0, respectively, and those of the bce loss $\lambda_{bce}$ and the dice loss $\lambda_{dice}$ are set to 2.0 and 0.5, respectively. Besides, the batch size per device is set to 2, and the gradient accumulation step is set to 10. During training, we select at most 3 categories for each image in semantic segmentation datasets. + +Datasets. As mentioned in Sec. 4.2, our training data is mainly composed of three types of datasets: (1) For the semantic segmentation dataset, we use ADE20K [62] and COCO-Stuff [3]. Besides, to enhance the segmentation result for some part of an object, we also use part semantic segmentation datasets, including PACO-LVIS [40], PartImageNet [13], and PASCAL-Part [6]; (2) For the referring segmentation dataset, we use refCLEF, refCOCO, refCOCO+ [17], and refCOCOg [35]; (3) For the visual ques + +tion answering (VQA) dataset, we use the datasets of LLaVA-Instruct-150k for LLaVA v1 [29] and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]. In order to avoid data leakage, we exclude the COCO samples whose images are present in the refCOCO $(+ / \mathrm{g})$ validation sets during training. Furthermore, we surprisingly find that by fine-tuning the model on only 239 ReasonSeg data samples, the model's performance can be further boosted. + +Evaluation Metrics. We follow most previous works on referring segmentation [17, 35] to adopt two metrics: gIoU and cIoU. gIoU is defined by the average of all per-image Intersection-over-Unions (IoUs), while cIoU is defined by the cumulative intersection over the cumulative union. Since cIoU is highly biased toward large-area objects and it fluctuates too much, gIoU is preferred. + +# 5.2. Reasoning Segmentation Results + +The reasoning segmentation results are shown in Table 1. It is worth noting that existing works fail to handle the task, but our model can accomplish the task involving complex reasoning with more than $20\%$ gIoU performance boost. As mentioned before, the reasoning segmentation task is essentially different from the referring segmentation task in that it requires the model to possess reasoning ability or access world knowledge. Only by truly understanding the query, can the model do well in the task. The existing works have no proper way to understand an implicit query, but our model exploits multimodal LLMs to reach the goal. + +Notably, we also make a comparison with the vanilla two-stage method (LLaVA1.5 + OVSeg). Specifically, the + +Table 2. Referring segmentation results (cIoU) among LISA (ours) and existing methods. + +
MethodrefCOCOrefCOCO+refCOCOg
valtestAtestBvaltestAtestBval(U)test(U)
MCN [34]62.464.259.750.655.044.749.249.4
VLT [11]67.570.565.256.361.050.155.057.7
CRIS [48]70.573.266.162.368.153.759.960.4
LAVT [53]72.775.868.862.168.455.161.262.1
ReLA [27]73.876.570.266.071.057.765.066.0
X-Decoder [65]------64.6-
SEEM [66]------65.7-
LISA-7B74.176.571.162.467.456.566.468.5
LISA-7B (fine-tuned on ReferSeg)74.979.172.365.170.858.167.970.6
+ +two-stage method refers to first using a multimodal LLM (e.g., LLaVA v1.5) to generate a text output for the input query, and then adopting a referring or open-vocabulary segmentation model (e.g., OVSeg) to generate the segmentation mask. If the intermediate text output remains too long and exceeds the input token length limit of OVSeg, we use GPT-3.5 to further summarize. More details can be found in the supplementary material. The results in Table 1 show that our model outperforms the two-stage method significantly. We explain that the potential reasons are: 1) Our model is trained end-to-end, while the two-stage method is completely decoupled; 2) The two-stage method relies on text as an intermediary to transmit information, while our model utilizes the hidden embedding that is more expressive. + +Another finding is that LISA-13B outperforms the 7B counterpart substantially, especially on the long-query scenarios, which indicates that the current performance bottleneck may still lie in understanding the query text, and a stronger multimodal LLM (e.g., LLaVA v1.5 [28]) leads to even better results. + +# 5.3. Vanilla Referring Segmentation Results + +To show that our model is also competent in the vanilla referring segmentation task, we make a comparison with existing state-of-the-art methods in Table 2. We evaluate the methods on refCOCO, refCOCO+, refCOCOg validation and testing sets. Our model achieves state-of-the-art results across various referring segmentation benchmarks. + +# 5.4. Ablation Study + +In this section, we conduct an extensive ablation study to reveal the contribution of each component. Unless otherwise specified, we report the metrics of gIoU and cIoU of LISA-7B on the validation set. + +Design Choices of Vision Backbone. We emphasize that vision backbones other than SAM are also applicable in our framework. In Table 3, we notice that SAM performs the best, potentially because of the massive high-quality + +Table 3. Ablation study on the design choice of vision backbone. 'ft' denotes finetuning on ReasonSeg training set. + +
Vision BackbonegIoUcIoU
Mask2Former-Swin-L42.438.8
SAM (w/ LoRA)41.537.3
SAM44.446.0
Mask2Former-Swin-L (ft)50.752.3
SAM w/ LORA (ft)51.851.9
SAM (ft)52.954.0
+ +Table 4. Ablation study on SAM pre-trained weight and rephrasing. + +
Exp. IDPre-trainSAM γrephrasinggIoUcIoU
1✓35.944.6
2✓50.751.1
3✓✓52.954.0
+ +data used in its pre-training phase. Further, we also find that with the Mask2Former backbone, our framework still achieves a decent performance on the reasoning segmentation task, significantly outperforming previous works such as X-Decoder [65]. This reveals the fact that the design choice of vision backbone is flexible and not limited to SAM. + +SAM LoRA Fintuning. We also investigate the effectiveness of applying LoRA on the SAM backbone. In Table 3, we note that the performance of LoRA fine-tuned SAM backbone is inferior to that of the frozen one. A potential reason is that fine-tuning impairs the generalization ability of the original SAM model. + +SAM Pre-trained Weight. To demonstrate the contribution of SAM pre-trained weight, we make a comparison between Experiments 1 and 3 in Table 4. Without being initialized with SAM pre-trained weight, the vision backbone is trained from scratch. This causes the performance to fall substantially behind that of the baseline model. + +![](images/b91107beaf4c0b451f08e5a024c9e9ad8874bdbeffaf22947e330bd3664e0bab.jpg) +the most likely object that someone else has left behind + +![](images/c6e62571977c7576cab40e12bc252d7da23aa18eef0076e5b291d23fa030a307.jpg) +中 + +![](images/d3e02c8979e997296a5193f45fac33af6ce2867cb60ef537857b971296807063.jpg) + +![](images/0328c0b81c2a3acb2431462e655a554574b609d45a47ee72d07573ad59c65150.jpg) +the object used for stirring milk or coffee + +![](images/363392ea772e52f662d1f5a0226e75995b4b73c0dfe33b0aa8b262d60d0d15dd.jpg) +W + +![](images/66712d38cbc11cb3ba2c75df85d01a2a5be003753822b0e908f115e0fbf39cbc.jpg) + +![](images/111f9f6fe104f4630a753969a7ecf69f0a27d13f3daa462bc1611ef514988db2.jpg) +In cold days, dogs may need extra protection to keep them warm. What object in the picture can a dog wear to provide warmth during snowy walks? + +![](images/0598a171e198f06f2bb5e16d8bb704744bc991bfb0764a403a0b76b1d8596b57.jpg) +中 + +![](images/5d9cba76fae1045402466d9cdcb5726b34c1d5ca2a8e70ae264e677b64c107d3.jpg) + +![](images/88fc045aea040343be95bd9dd217e2b1e2b22fca5acffd5e435a8bdc33439f05.jpg) + +![](images/6d51fb2f1098a87272c673d6d1e48ef57dc26868c87078154069b93eae9d8ed1.jpg) +When a plane is ready to land on the airport runway, what area in the picture will it eventually land on? +Query +Image + +![](images/7c8a06e77cf69b1d705d409df03f16a34e67570f394d9b3a4c193a14fb1e244d.jpg) +中 +Figure 5. Visual comparison among LISA (ours) and existing related methods. More illustrations are given in the supplementary material. + +Table 5. Ablation study on training data. + +
IDSemanticSegReferSegVQAReasonSeggIoUcIoU
ADE20KCOCO-StuffPartSeg
1✓✓✓✓✓48.953.5
2✓✓✓✓✓48.550.8
3✓✓✓✓✓46.750.9
4✓✓✓✓46.646.7
5✓✓✓30.420.4
6✓✓✓✓✓47.751.1
7✓✓✓✓✓44.446.0
8✓✓✓✓✓✓52.954.0
+ +Table 6. Results on the ReasonSeg test set. + +
Training splits# data samplesgIoUcIoU
train23951.751.1
train + val43954.054.9
+ +Instruction Rephrasing by GPT-3.5. When fine-tuning the model on the reasoning segmentation data samples, we rephrase the text instruction by GPT-3.5 (the details are shown in the supplementary material), and randomly choose one. The comparison between Experiments 2 and 3 in Table 4 shows that the performance is increased by $2.2\%$ gIoU and $2.9\%$ cIoU. This result verifies the effectiveness of such data augmentation. + +Contribution of All Types of Training Data. In Table 5, we show the contribution of each type of data to the performance. We find that in Exp. 5, we do not use any semantic segmentation dataset, and the performance drops a lot. We conjecture that semantic segmentation datasets provide a large amount of ground-truth binary masks for training, since a multi-class label can induce multiple binary masks. + +We also notice that adding more reasoning segmentation data samples during training leads to better results. In Table 6, we also add the ReasonSeg val set (200 data samples) during fine-tuning, and it yields better performance in both gIoU and cIoU metrics. This indicates that more reasoning segmentation training samples are beneficial at this moment. + +# 5.5. Qualitative Results + +As depicted in Fig. 5, we provide a visual comparison with existing related works, including the model for open-vocabulary semantic segmentation (OVSeg), referring segmentation (GRES), and the generalist models for segmentation (X-Decoder and SEEM). These models fail to handle the displayed cases with various errors, while our approach produces accurate and high-quality segmentation results. More illustrations are given in the supplementary material. + +# 6. Conclusion + +In this work, we have proposed a new segmentation task—reasoning segmentation. Also, we have introduced an evaluation benchmark ReasonSeg, which comprises over one thousand data samples. Finally, we have presented our model — LISA. It injects segmentation capabilities into current multimodal LLMs and performs surprisingly effectively on the reasoning segmentation task. We hope our work can shed new light on the direction of combining LLMs and vision tasks in the future. + +# Acknowledgements + +This work is supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 1, 3 +[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 2017. 2 +[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 6 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 2 +[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014. 6 +[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2 +[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 2 +[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4 +[10] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3 +[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, 2021. 7 +[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, 2019. 2 +[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In ECCV, 2022. 6 +[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 4, 5 + +[16] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019. 2 +[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 2, 3, 6 +[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, 2019. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 2, 4, 6 +[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3 +[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 3 +[22] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In CVPR, 2021. 2 +[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 1, 3 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 1, 3 +[25] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In CVPR, 2021. 2 +[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 6 +[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In CVPR, 2023. 6, 7 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint, 2023. 1, 5, 6, 7 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv:2304.08485, 2023. 1, 3, 4, 5, 6 +[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pretraining for open-set object detection. arXiv preprint, 2023. 6 +[31] Wei Liu, Andrew Rabinovich, and Alexander C. Berg. Parsenet: Looking wider to see better. arXiv preprint, 2015. 2 + +[32] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Internchat: Solving vision-centric tasks by interacting with chatbots beyond language. arXiv:2305.05662, 2023. 3 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6 +[34] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, 2020. 7 +[35] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 6 +[36] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2 +[37] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In ICCV, 2015. 2 +[38] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3 +[39] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, and Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv:2305.14167, 2023. 3 +[40] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In CVPR, 2023. 6 +[41] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In SIGKDD, 2020. 6 +[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[43] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. TPAMI, 2017. 2 +[44] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv:2303.17580, 2023. 3 +[45] Zhuotao Tian, Pengguang Chen, Xin Lai, Li Jiang, Shu Liu, Hengshuang Zhao, Bei Yu, Ming-Chang Yang, and Jiaya Jia. Adaptive perspective distillation for semantic segmentation. TPAMI, 2022. 2 +[46] Zhuotao Tian, Jiequan Cui, Li Jiang, Xiaojuan Qi, Xin Lai, Yixin Chen, Shu Liu, and Jiaya Jia. Learning context-aware classifier for semantic segmentation. AAAI, 2023. 2 +[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu + +Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv:2305.11175, 2023. 3 +[48] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. 7 +[49] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671, 2023. 3 +[50] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019. 2 +[51] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In CVPR, 2018. 2 +[52] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752, 2023. 3 +[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. 7 +[54] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv:2303.11381, 2023. 3 +[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178, 2023. 1, 3 +[56] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In ICLR, 2016. 2 +[57] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3 +[58] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. NeurIPS, 2021. 2 +[59] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 2 +[60] Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia. Icnet for real-time semantic segmentation on high-resolution images. In ECCV, 2018. +[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In ECCV, 2018. 2 +[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6 +[63] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision + +language understanding with advanced large language models. arXiv:2304.10592, 2023. 1, 3 +[64] Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, and Xi-ang Bai. Asymmetric non-local neural networks for semantic segmentation. In ICCV, 2019. 2 +[65] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023. 2, 6, 7 +[66] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 2, 4, 6, 7 \ No newline at end of file diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/images.zip b/2024/LISA_ Reasoning Segmentation via Large Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..01fc8771feb380c041a65367b898ac0197cd90d0 --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:260c223c36cdb089ff9d22016be35687ca6dc08852e24075495f99a27f164bb6 +size 698180 diff --git a/2024/LISA_ Reasoning Segmentation via Large Language Model/layout.json b/2024/LISA_ Reasoning Segmentation via Large Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dfb9963af6b59d9e8a9a8ca02166ba309d60dbd8 --- /dev/null +++ b/2024/LISA_ Reasoning Segmentation via Large Language Model/layout.json @@ -0,0 +1,10637 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 480, + 121 + ], + "type": "text", + "content": "LISA: Reasoning Segmentation via Large Language Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": "Xin Lai" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Zhuotao Tian" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Yukang Chen" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Yanwei Li" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Yuhui Yuan" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Shu Liu" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " Jiaya Jia" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{1}\\mathrm{CUHK}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{2}\\mathrm{HIT}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "text", + "content": " (Shenzhen) " + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{3}\\mathrm{SmartMore}" + }, + { + "bbox": [ + 53, + 142, + 541, + 173 + ], + "type": "inline_equation", + "content": "^{4}\\mathrm{MSRA}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 212 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 559 + ], + "type": "text", + "content": "Although perception systems have made remarkable advancements in recent years, they still rely on explicit human instruction or pre-defined categories to identify the target objects before executing visual recognition tasks. Such systems cannot actively reason and comprehend implicit user intention. In this work, we propose a new segmentation task — reasoning segmentation. The task is designed to output a segmentation mask given a complex and implicit query text. Furthermore, we establish a benchmark comprising over one thousand image-instruction-mask data samples, incorporating intricate reasoning and world knowledge for evaluation purposes. Finally, we present LISA: large Language Instructed Segmentation Assistant, which inherits the language generation capabilities of multimodal Large Language Models (LLMs) while also possessing the ability to produce segmentation masks. We expand the original vocabulary with a token and propose the embedding-as-mask paradigm to unlock the segmentation capability. Remarkably, LISA can handle cases involving complex reasoning and world knowledge. Also, it demonstrates robust zero-shot capability when trained exclusively on reasoning-free datasets. In addition, fine-tuning the model with merely 239 reasoning segmentation data samples results in further performance enhancement. Both quantitative and qualitative experiments show our method effectively unlocks new reasoning segmentation capabilities for multimodal LLMs. Code, models, and data are available at github.com/dvlabresearch/LISA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 128, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 601, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 601, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 45, + 601, + 288, + 687 + ], + "type": "text", + "content": "In daily life, users tend to issue direct commands like \"Change the TV channel\" to instruct a robot, rather than providing explicit step-by-step instructions such as \"Go to the table first, find the TV remote, and then press the button to change the channel.\" However, existing perception systems consistently rely on humans to explicitly indicate target objects or pre-define categories before executing visual recog" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 200, + 547, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 547, + 261 + ], + "type": "text", + "content": "nition tasks. These systems cannot actively reason and comprehend user intention based on implicit instruction. This reasoning ability is crucial in developing next-generation intelligent perception systems and holds substantial potential for industrial applications, particularly in robotics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 263, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 384 + ], + "type": "text", + "content": "In this work, we introduce a new segmentation task — reasoning segmentation, which requires generating a binary segmentation mask based on an implicit query text involving complex reasoning. Notably, the query text is not limited to a straightforward reference (e.g., “the orange”), but a more complicated description involving complex reasoning or world knowledge (e.g., “the food with high Vitamin C”). To accomplish this task, the model must possess two key abilities: 1) reasoning complex and implicit text queries jointly with the image; 2) producing segmentation masks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 387, + 548, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 548, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 548, + 519 + ], + "type": "text", + "content": "Inspired by the exceptional capacity of LLMs to reason and comprehend user intentions, we aim to leverage this capability of LLMs to address the aforementioned first challenge. However, while several studies [1, 23, 24, 28, 29, 55, 63] have integrated robust reasoning capabilities into multimodal LLMs to accommodate visual input, the majority of these models primarily concentrate on text generation tasks and still fall short in performing vision tasks that require fine-grained output formats, such as segmentation masks. This leads us to ask: can we enable multimodal LLMs with the capability to output segmentation masks?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": "To this end, we introduce LISA: a large Language Instructed Segmentation Assistant, a multimodal LLM capable of producing segmentation masks. Specifically, we incorporate an additional token, i.e., " + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": ", into the existing vocabulary. Upon generating the " + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 304, + 521, + 548, + 713 + ], + "type": "text", + "content": " token, its hidden embedding is further decoded into the corresponding segmentation mask. By representing the segmentation mask as an embedding, LISA acquires segmentation capabilities and benefits from end-to-end training. Remarkably, LISA demonstrates robust zero-shot abilities. Training the model solely on standard semantic segmentation and referring segmentation datasets yields surprisingly effective performance on the reasoning segmentation task. Furthermore, we find that LISA's performance can be significantly enhanced by fine-tuning on just 239 reasoning segmentation data samples. As illustrated in Fig. 1, LISA can handle various scenarios" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 138, + 712 + ], + "type": "text", + "content": "† Corresponding Author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9579" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 83, + 129, + 132 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 54, + 83, + 129, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 83, + 129, + 118 + ], + "spans": [ + { + "bbox": [ + 54, + 83, + 129, + 118 + ], + "type": "text", + "content": "- USER: Can you segment the tyre that does not touch the ground in this image?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 119, + 129, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 129, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 129, + 132 + ], + "type": "text", + "content": "- ASSISTANT: Sure, it is ." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 133, + 82, + 211, + 135 + ], + "blocks": [ + { + "bbox": [ + 133, + 82, + 211, + 135 + ], + "lines": [ + { + "bbox": [ + 133, + 82, + 211, + 135 + ], + "spans": [ + { + "bbox": [ + 133, + 82, + 211, + 135 + ], + "type": "image", + "image_path": "09a3adad0cc374de132e4ae29b4d128db4221c4e9b7805ee97e6bf3937620096.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 213, + 83, + 292, + 136 + ], + "blocks": [ + { + "bbox": [ + 213, + 83, + 292, + 136 + ], + "lines": [ + { + "bbox": [ + 213, + 83, + 292, + 136 + ], + "spans": [ + { + "bbox": [ + 213, + 83, + 292, + 136 + ], + "type": "image", + "image_path": "8c2512837cd9c28e35a9d8df1bd7a96f2a60197d49abc8435c9a7db3fc375b0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 77, + 415, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 77, + 415, + 125 + ], + "spans": [ + { + "bbox": [ + 304, + 77, + 415, + 125 + ], + "type": "text", + "content": "- USER: There are two washing machines as shown in the picture. If I need to do laundry, where in the picture would I put the clothes? Please output segmentation mask." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 126, + 372, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 126, + 372, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 126, + 372, + 133 + ], + "type": "text", + "content": "- ASSISTANT: ." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 421, + 72, + 474, + 137 + ], + "blocks": [ + { + "bbox": [ + 421, + 72, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 421, + 72, + 474, + 137 + ], + "spans": [ + { + "bbox": [ + 421, + 72, + 474, + 137 + ], + "type": "image", + "image_path": "a0acaf7c6d33ce0213009a977dd6a4baeed984e74de62ec003de6713e7231b5e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 482, + 72, + 536, + 137 + ], + "blocks": [ + { + "bbox": [ + 482, + 72, + 536, + 137 + ], + "lines": [ + { + "bbox": [ + 482, + 72, + 536, + 137 + ], + "spans": [ + { + "bbox": [ + 482, + 72, + 536, + 137 + ], + "type": "image", + "image_path": "2a2de191a1037497bef85eda8d3a180fdaa0f5d5a92fe0ab13f94fad1426666d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 153, + 123, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 153, + 123, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 123, + 188 + ], + "type": "text", + "content": "- USER: Can you segment the food that tastes not spicy in this image?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 60, + 189, + 122, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 189, + 122, + 195 + ], + "spans": [ + { + "bbox": [ + 60, + 189, + 122, + 195 + ], + "type": "text", + "content": "ASSISTANT: ." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 130, + 151, + 198, + 198 + ], + "blocks": [ + { + "bbox": [ + 130, + 151, + 198, + 198 + ], + "lines": [ + { + "bbox": [ + 130, + 151, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 130, + 151, + 198, + 198 + ], + "type": "image", + "image_path": "a9e153b9ee7022c45bfc37405011eb746e7ed5eebd5f46b0d906570cc76b84ee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 202, + 151, + 272, + 198 + ], + "blocks": [ + { + "bbox": [ + 202, + 151, + 272, + 198 + ], + "lines": [ + { + "bbox": [ + 202, + 151, + 272, + 198 + ], + "spans": [ + { + "bbox": [ + 202, + 151, + 272, + 198 + ], + "type": "image", + "image_path": "f3de821f41bd3d8a3c7e8be9db2977b383841638d5eaaa3ce280d23e269fb10c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 282, + 150, + 364, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 150, + 364, + 192 + ], + "spans": [ + { + "bbox": [ + 282, + 150, + 364, + 192 + ], + "type": "text", + "content": "- USER: What is the place where the driver can observe the speed in this image? Please output segmentation mask." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 282, + 192, + 350, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 282, + 192, + 350, + 198 + ], + "spans": [ + { + "bbox": [ + 282, + 192, + 350, + 198 + ], + "type": "text", + "content": "- ASSISTANT: ." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 367, + 150, + 449, + 200 + ], + "blocks": [ + { + "bbox": [ + 367, + 150, + 449, + 200 + ], + "lines": [ + { + "bbox": [ + 367, + 150, + 449, + 200 + ], + "spans": [ + { + "bbox": [ + 367, + 150, + 449, + 200 + ], + "type": "image", + "image_path": "50f0ba0edf6a0fe1d788787241b99590ccc63f93f7863edfc16c83a426a3bad8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 454, + 150, + 536, + 200 + ], + "blocks": [ + { + "bbox": [ + 454, + 150, + 536, + 200 + ], + "lines": [ + { + "bbox": [ + 454, + 150, + 536, + 200 + ], + "spans": [ + { + "bbox": [ + 454, + 150, + 536, + 200 + ], + "type": "image", + "image_path": "2967c4020606cbc77806987d8ce1af3b97b0538a50ba8dded645cd5e57b5682d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 215, + 141, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 141, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 141, + 243 + ], + "type": "text", + "content": "- USER: Can you segment the unusual part in this image and explain why?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 243, + 143, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 143, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 143, + 277 + ], + "type": "text", + "content": "- ASSISTANT: Sure, it is . In the image, the unusual part is the dog wearing a reindeer antler headband." + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 146, + 217, + 196, + 277 + ], + "blocks": [ + { + "bbox": [ + 146, + 217, + 196, + 277 + ], + "lines": [ + { + "bbox": [ + 146, + 217, + 196, + 277 + ], + "spans": [ + { + "bbox": [ + 146, + 217, + 196, + 277 + ], + "type": "image", + "image_path": "59801feb73fb53abce28f395d5ad8c119cf6343348faacb2e08a0b90febf0e29.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 198, + 217, + 249, + 277 + ], + "blocks": [ + { + "bbox": [ + 198, + 217, + 249, + 277 + ], + "lines": [ + { + "bbox": [ + 198, + 217, + 249, + 277 + ], + "spans": [ + { + "bbox": [ + 198, + 217, + 249, + 277 + ], + "type": "image", + "image_path": "cd12e4cd63ab23c9db1b831006e4db947ae3af78d355fdcb1536cfd33f62ad08.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 260, + 215, + 365, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 215, + 365, + 243 + ], + "spans": [ + { + "bbox": [ + 260, + 215, + 365, + 243 + ], + "type": "text", + "content": "- USER: Who is the loser in this image? Please output segmentation mask and explain why." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 260, + 244, + 369, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 260, + 244, + 369, + 278 + ], + "spans": [ + { + "bbox": [ + 260, + 244, + 369, + 278 + ], + "type": "text", + "content": "ASSISTANT: Sure, it is . ... The man on top is likely the winner of the match, as he has control over the other fighters body..." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 370, + 223, + 452, + 275 + ], + "blocks": [ + { + "bbox": [ + 370, + 223, + 452, + 275 + ], + "lines": [ + { + "bbox": [ + 370, + 223, + 452, + 275 + ], + "spans": [ + { + "bbox": [ + 370, + 223, + 452, + 275 + ], + "type": "image", + "image_path": "5326720e924ea3944f1bfa50741ccd7a1470ab666e691cf9d6c8fdf26dc5a680.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 454, + 223, + 536, + 277 + ], + "blocks": [ + { + "bbox": [ + 454, + 223, + 536, + 277 + ], + "lines": [ + { + "bbox": [ + 454, + 223, + 536, + 277 + ], + "spans": [ + { + "bbox": [ + 454, + 223, + 536, + 277 + ], + "type": "image", + "image_path": "b2b7b9a4ea969654f4ed89b67dff6bdd74f46441148c1cea41f1303bd611336c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 67, + 289, + 151, + 346 + ], + "blocks": [ + { + "bbox": [ + 67, + 289, + 151, + 346 + ], + "lines": [ + { + "bbox": [ + 67, + 289, + 151, + 346 + ], + "spans": [ + { + "bbox": [ + 67, + 289, + 151, + 346 + ], + "type": "image", + "image_path": "e31ac66fccd7e539d1b84f13c5c37b77ac96e48e4ec8bb7852fe67ec0bd0cadf.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 352, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 352, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 545, + 384 + ], + "type": "text", + "content": "Figure 1. We unlock new segmentation capabilities for existing multimodal LLMs. Our model (i.e., LISA) can deal with cases involving complex reasoning and world knowledge. Also, we demonstrate the cases of explanatory answers in the 3rd row. Additionally, in the 4th row, our model can output multiple segmentation masks in a single answer. More illustrations can be found in the supplementary material." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 162, + 303, + 316, + 331 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 162, + 303, + 307, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 303, + 307, + 324 + ], + "spans": [ + { + "bbox": [ + 162, + 303, + 307, + 324 + ], + "type": "text", + "content": "- USER: Can you segment the fried food and the food with most protein separately in this image?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 163, + 325, + 316, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 325, + 316, + 331 + ], + "spans": [ + { + "bbox": [ + 163, + 325, + 316, + 331 + ], + "type": "text", + "content": "- ASSISTANT: Sure, they are and ." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 334, + 289, + 419, + 346 + ], + "blocks": [ + { + "bbox": [ + 334, + 289, + 419, + 346 + ], + "lines": [ + { + "bbox": [ + 334, + 289, + 419, + 346 + ], + "spans": [ + { + "bbox": [ + 334, + 289, + 419, + 346 + ], + "type": "image", + "image_path": "b1344c5d2785ec033c0b1f52778c5511f38768875089b8ee10cce644455c8827.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 438, + 289, + 522, + 346 + ], + "blocks": [ + { + "bbox": [ + 438, + 289, + 522, + 346 + ], + "lines": [ + { + "bbox": [ + 438, + 289, + 522, + 346 + ], + "spans": [ + { + "bbox": [ + 438, + 289, + 522, + 346 + ], + "type": "image", + "image_path": "9cb3dffa2988a67e34d8ef8948518983538841f951e9231e86e7795b31b78414.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 401, + 257, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 257, + 412 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 257, + 412 + ], + "type": "text", + "content": "involving complex reasoning and world knowledge." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 413, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 508 + ], + "type": "text", + "content": "In addition, to validate the effectiveness, we establish a benchmark for reasoning segmentation evaluation, called ReasonSeg. Comprising over one thousand image-instruction pairs, this benchmark offers persuasive evaluation metrics for the task. To align more closely with practical applications, we annotate the images from OpenImages [21] and ScanNetv2 [10] with implicit text queries that involve complex reasoning." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 59, + 509, + 242, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 509, + 242, + 520 + ], + "spans": [ + { + "bbox": [ + 59, + 509, + 242, + 520 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 58, + 521, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 58, + 521, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 521, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 287, + 568 + ], + "type": "text", + "content": "- We introduce the reasoning segmentation task, which necessitates reasoning based on implicit human instructions. Such reasoning capability is crucial for building a genuinely intelligent perception system." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 58, + 574, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 574, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 58, + 574, + 287, + 646 + ], + "type": "text", + "content": "- We present our model — LISA, which incorporates new segmentation capabilities. It demonstrates robust zero-shot ability on the reasoning segmentation task when trained solely on reasoning-free datasets, and achieves further performance boost by fine-tuning on just 239 data samples that involve reasoning." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 58, + 651, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 651, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 651, + 288, + 712 + ], + "type": "text", + "content": "- We establish a reasoning segmentation benchmark, ReasonSeg, containing over one thousand image-instruction-mask data samples. This benchmark is essential for evaluation and encourages the community to further explore the reasoning ability for vision tasks." + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 392, + 411 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 414, + 424, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 424, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 424, + 426 + ], + "type": "text", + "content": "2.1. Image Segmentation" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 594 + ], + "type": "text", + "content": "Semantic segmentation aims to assign a class label to every pixel in an image. Numerous studies [2, 5, 8, 12, 16, 22, 31, 37, 42, 43, 45, 46, 51, 56, 59-61, 64] have proposed diverse designs (such as encoder-decoder, dilated convolution, pyramid pooling module, non-local operator, and more) to effectively encode semantic information. Research on instance segmentation [9, 14, 58] and panoptic segmentation [7, 18, 25, 50] has introduced various architectural innovations for instance-level segmentation, including DETR [4]-based structures, mask attention, and dynamic convolution. In recent years, typical segmentation tasks have made significant progress and become increasingly mature. Consequently, it is imperative to develop more intelligent interaction ways for image segmentation." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "The referring segmentation task [17, 36] enables interaction with human language, aiming to segment the target object based on a given explicit text description. Recently, Kirillov et al. [19] introduced SAM, trained with billions of high-quality masks, supporting bounding boxes and points as prompts while demonstrating exceptional segmentation quality. X-Decoder [65] bridges vision and language, unifying multiple tasks within a single model. SEEM [66] further supports various human interaction methods, including text, audio, and scribble. However, these studies primarily focus" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9580" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 74, + 323, + 171 + ], + "blocks": [ + { + "bbox": [ + 83, + 74, + 323, + 171 + ], + "lines": [ + { + "bbox": [ + 83, + 74, + 323, + 171 + ], + "spans": [ + { + "bbox": [ + 83, + 74, + 323, + 171 + ], + "type": "image", + "image_path": "56991a6f3a5e5f8ce80b50a1d4b30070d5e2947b1d711ba4a144318e9a456b98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 176, + 276, + 196 + ], + "lines": [ + { + "bbox": [ + 133, + 176, + 276, + 196 + ], + "spans": [ + { + "bbox": [ + 133, + 176, + 276, + 196 + ], + "type": "text", + "content": "the camera lens that is more suitable for photographing nearby objects" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 172, + 514, + 200 + ], + "lines": [ + { + "bbox": [ + 323, + 172, + 514, + 200 + ], + "spans": [ + { + "bbox": [ + 323, + 172, + 514, + 200 + ], + "type": "text", + "content": "Toddlers are curious and often enjoy exploring their surroundings. What object in the picture can provide a safe and enclosed space for a toddler to play in?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 329, + 74, + 511, + 171 + ], + "blocks": [ + { + "bbox": [ + 329, + 74, + 511, + 171 + ], + "lines": [ + { + "bbox": [ + 329, + 74, + 511, + 171 + ], + "spans": [ + { + "bbox": [ + 329, + 74, + 511, + 171 + ], + "type": "image", + "image_path": "2724632b9d752ec872d52434f95ac729991821fa8e65b14932cacaa6f44cb043.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 547, + 226 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 547, + 226 + ], + "type": "text", + "content": "Figure 2. Examples of the annotated image-instruction-mask data samples. Left: short phrase query. Right: long sentence query. More examples are given in the supplementary material." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 235, + 288, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 283 + ], + "type": "text", + "content": "on addressing multi-task compatibility and unification, neglecting the injection of new capabilities. In this work, we present LISA and it possesses reasoning ability that has not been explored yet in existing segmentors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 285, + 237, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 237, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 237, + 298 + ], + "type": "text", + "content": "2.2. Multimodal Large Language Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 289, + 649 + ], + "type": "text", + "content": "Motivated by the remarkable reasoning abilities of LLMs, researchers are exploring ways to transfer these capabilities into the vision domain, developing multimodal LLMs. Flamingo [1] employs a cross-attention structure to attend to visual contexts, enabling visual in-context learning. Models such as BLIP-2 [24] and mPLUG-OWL [55] propose encoding image features with a visual encoder, which are then fed into the LLM alongside text embeddings. Otter [23] further incorporates robust few-shot capabilities through in-context instruction tuning on the proposed MIMIC-IT dataset. LLaVA [29] and MiniGPT-4 [63] first conduct image-text feature alignment followed by instruction tuning. Koh et al. [20] also investigates image retrieval for LLMs. Moreover, numerous works [32, 44, 49, 52, 54] utilize prompt engineering, connecting independent modules via API calls, but without the benefits of end-to-end training. Recently, there have been studies examining the intersection between multimodal LLMs and vision tasks. Vision-LLM [47] offers a flexible interaction interface for multiple vision-centric tasks through instruction tuning but fails to fully exploit LLMs for complex reasoning. Kosmos-2 [38] constructs large-scale data of grounded image-text pairs, infusing grounding capabilities into LLMs. DetGPT [39] bridges the fixed multimodal LLM and open-vocabulary detector, enabling detection to be performed based on user instruction. GPT4RoI [57] introduces spatial boxes as input and trains the model on region-text pairs. In contrast, our work aims to efficiently inject segmentation capabilities into multimodal LLMs in the manner of end-to-end training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 657, + 188, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 188, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 188, + 670 + ], + "type": "text", + "content": "3. Reasoning Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 673, + 159, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 159, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 159, + 685 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "The reasoning segmentation task is to output a binary segmentation mask " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": ", given an input image " + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{img}" + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": " and an im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 235, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 355 + ], + "type": "text", + "content": "plicit query text instruction " + }, + { + "bbox": [ + 304, + 235, + 547, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text{txt}}" + }, + { + "bbox": [ + 304, + 235, + 547, + 355 + ], + "type": "text", + "content": ". The task shares a similar formulation with the referring segmentation task [17], but is far more challenging. The key distinction lies in the complexity of the query text in reasoning segmentation. Instead of a straightforward phrase (e.g., \"the trash can\"), the query text includes more intricate expressions (e.g., \"something that the garbage should be put into\") or longer sentences (e.g., \"After cooking, consuming food, and preparing for food, where can we throw away the rest of the food and scraps?\") that involve complex reasoning or world knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 361, + 383, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 383, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 383, + 373 + ], + "type": "text", + "content": "3.2.Benchmark" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 379, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 559 + ], + "type": "text", + "content": "Given the lack of quantitative evaluation, it is imperative to establish a benchmark for the reasoning segmentation task. To ensure reliable assessment, we have collected a diverse set of images from OpenImages [21] and ScanNetv2 [10], annotating them with implicit text instructions and high-quality target masks. To cover different scenarios, our text instructions consist of two types: 1) short phrases; 2) long sentences; as illustrated in Figure 2. The resulting ReasonSeg benchmark comprises a total of 1218 image-instruction-mask data samples. This dataset is further partitioned into three splits: train, val, and test, containing 239, 200, and 779 data samples, respectively. As the primary purpose of the benchmark is evaluation, the validation and testing sets include a larger number of data samples. The details of data annotation are given in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 570, + 386, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 386, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 386, + 582 + ], + "type": "text", + "content": "4. Our Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 590, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 546, + 626 + ], + "type": "text", + "content": "In this section, we first introduce the model architecture in Sec. 4.1. After that, we elaborate on the training data preparation and training parameters in Sec. 4.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 635, + 388, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 388, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 388, + 646 + ], + "type": "text", + "content": "4.1. Architecture" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "Embedding as Mask. Most current multimodal LLMs (such as LLaVA [29], Flamingo [1], BLIP-2 [24], Otter [23], etc.) support image and text as input, but they can only output text and cannot directly output fine-grained segmentation masks. VisionLLM [47] offers a solution by parsing" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9581" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 541, + 214 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 541, + 214 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 541, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 541, + 214 + ], + "type": "image", + "image_path": "05eed35d4fb8518cd038dea284bec7f20db5651948f11f6d0db41be165bc0d40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. The pipeline of LISA. Given the input image and text query, the multimodal LLM (e.g., LLaVA [29]) generates text output. The last-layer embedding for the token is then decoded into the segmentation mask via the decoder. We use LoRA [15] for efficient fine-tuning. The choice of vision backbone can be flexible (e.g., SAM [66], Mask2Former [9])." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 260, + 289, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 289, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 289, + 391 + ], + "type": "text", + "content": "segmentation masks as sequences of polygons, enabling the representation of segmentation masks as plain text and allowing end-to-end training within the framework of existing multimodal LLMs. However, end-to-end training with the polygon sequences introduces optimization challenges and may compromise generalization ability unless a massive amount of data and computational resources are employed. For instance, training a 7B model, VisionLLM requires " + }, + { + "bbox": [ + 46, + 260, + 289, + 391 + ], + "type": "inline_equation", + "content": "4 \\times 8" + }, + { + "bbox": [ + 46, + 260, + 289, + 391 + ], + "type": "text", + "content": " NVIDIA 80G A100 GPUs and 50 epochs, which is computationally prohibitive. In contrast, it takes less than 3 days to train LISA-7B on 8 NVIDIA 24G 3090 GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": "To this end, we propose the embedding-as-mask paradigm to infuse new segmentation capabilities into the multimodal LLM. The pipeline of our method is illustrated in Fig. 3. Specifically, we first expand the original LLM vocabulary with a new token, i.e., " + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "inline_equation", + "content": "<\\mathrm{SEG}>" + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": ", which signifies the request for the segmentation output. Given a text instruction " + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{txt}" + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": " along with the input image " + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{img}" + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": ", we feed them into the multimodal LLM " + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": ", which in turn outputs a text response " + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{txt}" + }, + { + "bbox": [ + 46, + 392, + 290, + 499 + ], + "type": "text", + "content": ". It can be formulated as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 505, + 287, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 505, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 114, + 505, + 287, + 519 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {t x t} = \\mathcal {F} (\\mathbf {x} _ {i m g}, \\mathbf {x} _ {t x t}). \\tag {1}", + "image_path": "8273388882dceb23db918cfd43c992cdabe982602f9b1af5682a88b62ab3d7c5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": "When the LLM intends to generate a binary segmentation mask, the output " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_{txt}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " would include a token. We then extract the LLM last-layer embedding " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{h}}_{seg}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " corresponding to the token and apply an MLP projection layer " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{seg}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": ". Simultaneously, the vision backbone " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{enc}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " extracts the dense visual features " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " from the visual input " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{img}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{seg}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " are fed to the decoder " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dec}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " to produce the final segmentation mask " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": ". The detailed structure of the decoder " + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dec}" + }, + { + "bbox": [ + 46, + 522, + 289, + 643 + ], + "type": "text", + "content": " follows [19]. The process can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 646, + 287, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 646, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 91, + 646, + 287, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {s e g} = \\gamma (\\tilde {\\mathbf {h}} _ {s e g}), \\quad \\mathbf {f} = \\mathcal {F} _ {e n c} (\\mathbf {x} _ {i m g}), \\tag {2}", + "image_path": "176d2fc4fc6abe2757dd8c2aa32dd9c31148b342090fe813486d99739793c0ec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 665, + 216, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 665, + 216, + 677 + ], + "spans": [ + { + "bbox": [ + 132, + 665, + 216, + 677 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\mathcal {F} _ {d e c} (\\mathbf {h} _ {s e g}, \\mathbf {f}).", + "image_path": "217c574ba309abf2d8cf88819dca362abd851ed4332d9071d692c432a1adfe51.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Training Objectives. The model is trained end-to-end using the text generation loss " + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{t:xt}" + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": " and the segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": "mask loss " + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": ". The overall objective " + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": " is the weighted sum of these losses, determined by " + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 305, + 260, + 545, + 285 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 362, + 294, + 545, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 294, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 362, + 294, + 545, + 307 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\lambda_ {m a s k} \\mathcal {L} _ {m a s k}. \\tag {3}", + "image_path": "f0b44d35d1cec61e193086891d55b9b718f51d5cd7e9c97a5f341708301e3660.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " is the auto-regressive cross-entropy loss for text generation, and " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " is the mask loss, which encourages the model to produce high-quality segmentation results. To compute " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": ", we employ a combination of per-pixel binary cross-entropy (BCE) loss and DICE loss, with corresponding loss weights " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": ". Given the ground-truth targets " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{txt}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 316, + 547, + 411 + ], + "type": "text", + "content": ", these losses can be formulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 419, + 474, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 419, + 474, + 432 + ], + "spans": [ + { + "bbox": [ + 375, + 419, + 474, + 432 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t x t} = \\mathbf {C E} (\\hat {\\mathbf {y}} _ {t x t}, \\mathbf {y} _ {t x t}),", + "image_path": "9eebd925717fb40bef0940d620c056cde4147fff713fed4f7ee81f8400f142e1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathbf {B C E} (\\hat {\\mathbf {M}}, \\mathbf {M}) + \\lambda_ {d i c e} \\mathbf {D I C E} (\\hat {\\mathbf {M}}, \\mathbf {M}). \\tag {4}", + "image_path": "516b8255d6254bdd1037a2b39ec98462ef9929d59fe1a138c2e23294d56c6bca.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 458, + 547, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 554 + ], + "type": "text", + "content": "It is noteworthy that the proposed method endows existing multimodal LLMs with new segmentation capabilities, such that they can generate not only text but also fine-grained output formats. Also, our method is based on an end-to-end training pipeline and connects the LLM and vision modules with hidden embedding representation, which proves significantly more effective than the decoupled two-stage method as discussed in Sec. 5.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 561, + 370, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 370, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 370, + 574 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 579, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 547, + 626 + ], + "type": "text", + "content": "Training Data Formulation. As illustrated in Fig. 4, our training data comprises mainly three parts, all of which are derived from widely-used public datasets. The details are as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 630, + 547, + 714 + ], + "type": "text", + "content": "- Semantic Segmentation Dataset. Semantic segmentation datasets typically consist of images and the corresponding multi-class labels. During training, we randomly choose several categories for each image. To generate data that matches the format of visual question answering, we employ a question-answer template like “USER: Can you segment the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9582" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 92, + 199, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 77, + 103, + 88 + ], + "lines": [ + { + "bbox": [ + 57, + 77, + 103, + 88 + ], + "spans": [ + { + "bbox": [ + 57, + 77, + 103, + 88 + ], + "type": "text", + "content": "Raw Data" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 92, + 199, + 190 + ], + "lines": [ + { + "bbox": [ + 66, + 92, + 199, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 92, + 199, + 190 + ], + "type": "image", + "image_path": "32a6e91c80caff3657c829b697f482f86fbdda77300dfd4a607cdb056d0b3ad3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 91, + 323, + 190 + ], + "blocks": [ + { + "bbox": [ + 227, + 91, + 323, + 190 + ], + "lines": [ + { + "bbox": [ + 227, + 91, + 323, + 190 + ], + "spans": [ + { + "bbox": [ + 227, + 91, + 323, + 190 + ], + "type": "image", + "image_path": "76b5ed7bce4ae8a94540d47033c18525144f1246515c1ad6f70c7c165a7e198a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 129, + 380, + 159 + ], + "lines": [ + { + "bbox": [ + 324, + 129, + 380, + 159 + ], + "spans": [ + { + "bbox": [ + 324, + 129, + 380, + 159 + ], + "type": "text", + "content": "the lady with the blue shirt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 95, + 529, + 190 + ], + "blocks": [ + { + "bbox": [ + 402, + 95, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 402, + 95, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 402, + 95, + 529, + 190 + ], + "type": "image", + "image_path": "39526b71af96bda5721d1b15aaf424999785330d0d23cc8ab59d463f0fb3fd77.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 205, + 127, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 205, + 127, + 215 + ], + "spans": [ + { + "bbox": [ + 59, + 205, + 127, + 215 + ], + "type": "text", + "content": "Processed Data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 224, + 188, + 260 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 69, + 224, + 181, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 224, + 181, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 224, + 181, + 251 + ], + "type": "text", + "content": "- USER: Can you segment the table in this image?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 251, + 188, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 251, + 188, + 260 + ], + "spans": [ + { + "bbox": [ + 69, + 251, + 188, + 260 + ], + "type": "text", + "content": "- ASSISTANT: It is ." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 262, + 194, + 351 + ], + "blocks": [ + { + "bbox": [ + 72, + 262, + 194, + 351 + ], + "lines": [ + { + "bbox": [ + 72, + 262, + 194, + 351 + ], + "spans": [ + { + "bbox": [ + 72, + 262, + 194, + 351 + ], + "type": "image", + "image_path": "124555f612515343327fde250c264861e45415c7bef15716f24116c05bb4231f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 357, + 192, + 369 + ], + "lines": [ + { + "bbox": [ + 67, + 357, + 192, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 357, + 192, + 369 + ], + "type": "text", + "content": "Semantic Segmentation Data" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 225, + 225, + 372, + 262 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 225, + 225, + 372, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 225, + 372, + 252 + ], + "spans": [ + { + "bbox": [ + 225, + 225, + 372, + 252 + ], + "type": "text", + "content": "- USER: Can you segment the lady with the blue shirt in this image?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 225, + 253, + 372, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 253, + 372, + 262 + ], + "spans": [ + { + "bbox": [ + 225, + 253, + 372, + 262 + ], + "type": "text", + "content": "- ASSISTANT: Sure, it is ." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 230, + 264, + 298, + 352 + ], + "blocks": [ + { + "bbox": [ + 230, + 264, + 298, + 352 + ], + "lines": [ + { + "bbox": [ + 230, + 264, + 298, + 352 + ], + "spans": [ + { + "bbox": [ + 230, + 264, + 298, + 352 + ], + "type": "image", + "image_path": "cf130f8d06ee9fff521df64a51723f34eb4436252b27baeef66e62594133abc1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 357, + 362, + 369 + ], + "lines": [ + { + "bbox": [ + 235, + 357, + 362, + 369 + ], + "spans": [ + { + "bbox": [ + 235, + 357, + 362, + 369 + ], + "type": "text", + "content": "Referring Segmentation Data" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 305, + 264, + 373, + 351 + ], + "blocks": [ + { + "bbox": [ + 305, + 264, + 373, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 264, + 373, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 373, + 351 + ], + "type": "image", + "image_path": "0df82e6b4af35777b0ad654250a2e259e11491570087d0397d34e4573655db38.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "lines": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 546, + 399 + ], + "type": "text", + "content": "Figure 4. The illustration of training data formulation from different types of data, including semantic segmentation data, referring segmentation data, and visual question answering (VQA) data." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 399, + 224, + 533, + 313 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 399, + 224, + 533, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 224, + 533, + 251 + ], + "spans": [ + { + "bbox": [ + 399, + 224, + 533, + 251 + ], + "type": "text", + "content": "- USER: What type of sign and traffic device can be seen in the image? " + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 399, + 252, + 533, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 252, + 533, + 287 + ], + "spans": [ + { + "bbox": [ + 399, + 252, + 533, + 287 + ], + "type": "text", + "content": "- ASSISTANT: In the image, there is a street sign and a traffic light above a city road." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 399, + 288, + 527, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 288, + 527, + 313 + ], + "spans": [ + { + "bbox": [ + 399, + 288, + 527, + 313 + ], + "type": "text", + "content": "- USER: How many traffic lights are visible in the image?" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 399, + 337, + 532, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 337, + 532, + 349 + ], + "spans": [ + { + "bbox": [ + 399, + 337, + 532, + 349 + ], + "type": "text", + "content": "No Binary Segmentation Mask" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 441, + 357, + 488, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 357, + 488, + 368 + ], + "spans": [ + { + "bbox": [ + 441, + 357, + 488, + 368 + ], + "type": "text", + "content": "VQA Data" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 415, + 289, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 289, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 289, + 522 + ], + "type": "text", + "content": "{class_name} in this image? ASSISTANT: It is .\", where {class_name} is the chosen category, and denotes the placeholder for tokens of image patches. The corresponding binary segmentation mask is used as the ground truth to provide mask loss supervision. During training, we also use other templates to generate the QA data to ensure data diversity, as shown in the supplementary material. We adopt ADE20K, COCO-Stuff, and LVIS-PACO part segmentation datasets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 534, + 289, + 703 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 47, + 534, + 289, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 289, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 289, + 643 + ], + "type": "text", + "content": "- Vanilla Referring Segmentation Dataset. Referring segmentation datasets provide an input image and an explicit short description of the target object. Thus, it is easy to convert them into question-answer pairs using a template like \"USER: Can you segment {description} in this image? ASSISTANT: Sure, it is .\", where {description} is the given explicit description. For this part, we adopt ref-COCO, refCOCO+, refCOCOg, and refCLEF datasets." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 654, + 288, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 288, + 703 + ], + "type": "text", + "content": "- Visual Question Answering Dataset. To preserve the original Visual Question Answering (VQA) ability of the multimodal LLM, we also include the VQA dataset during training. We use LLaVA-Instruct-150k [29] for LLaVA v1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 415, + 511, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 511, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 511, + 426 + ], + "type": "text", + "content": "and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 432, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 547, + 540 + ], + "type": "text", + "content": "Notably, the above datasets do not include any reasoning segmentation data sample. Instead, it only contains samples where the target objects are explicitly indicated in the query texts. Surprisingly, even without complex reasoning training data, LISA demonstrates impressive zero-shot ability on the ReasonSeg benchmark, as shown in Table 1. Moreover, we find that further performance boost could be yielded by finetuning the model on only 239 data samples that involve complex reasoning." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "content": "Trainable Parameters. To preserve the learned knowledge of the pre-trained multimodal LLM " + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "content": " (i.e., LLaVA in our experiments), we leverage LoRA [15] to perform efficient fine-tuning, and completely freeze the vision backbone " + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{enc}" + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "content": ". The decoder " + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dec}" + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "content": " is fully fine-tuned. Additionally, the LLM token embeddings (embed_tokens), the LLM head (lm_head), and the projection layer " + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 555, + 547, + 640 + ], + "type": "text", + "content": " are also trainable." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "It is notable that the resulting model avoids the catastrophic forgetting of the original text generation capability and preserves the conversation ability, as verified in the supplementary material. The potential reasons are: we 1) employ LoRA fine-tuning to reduce the trainable parameters and 2) incorporate the VQA dataset during fine-tuning." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9583" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 96, + 531, + 335 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Reasoning segmentation results among LISA (ours) and previous related works. 'ft' denotes using 239 reasoning segmentation data samples to fine-tune the model. Unless otherwise specified, we use LLaVA v1 [29] as the base model. LLaVA1.5 denotes LLaVA v1.5 [28]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 96, + 531, + 335 + ], + "lines": [ + { + "bbox": [ + 59, + 96, + 531, + 335 + ], + "spans": [ + { + "bbox": [ + 59, + 96, + 531, + 335 + ], + "type": "table", + "html": "
Methodvaltest
overallshort querylong queryoverall
gIoUcIoUgIoUcIoUgIoUcIoUgIoUcIoU
OVSeg [26]28.518.618.015.528.722.526.120.8
GRES [27]22.419.917.615.022.623.821.322.0
X-Decoder [65]22.617.920.411.622.217.521.716.3
SEEM [66]25.521.220.111.525.620.824.318.7
Grounded-SAM [30]26.014.517.810.822.418.621.316.4
LISA-7B44.446.037.634.436.634.736.834.1
LISA-7B (ft)52.954.040.640.649.451.047.348.4
LISA-13B48.946.939.943.346.446.544.845.8
LISA-13B (ft)56.262.944.342.054.054.351.751.1
LLaVA1.5-7B + OVSeg38.223.524.218.744.637.139.731.8
LISA-7B-LLaVA1.553.652.347.148.549.248.948.748.8
LISA-7B-LLaVA1.5 (ft)61.362.948.346.357.959.755.656.9
LLaVA1.5-13B + OVSeg37.926.427.119.446.140.641.534.1
LISA-13B-LLaVA1.557.760.350.850.054.750.953.850.8
LISA-13B-LLaVA1.5 (ft)65.072.955.450.663.265.361.362.2
", + "image_path": "623a2069cff2f6cdddfe2583bbaac44dddfd4369f091b4cb0e5e5bac79ce25b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 342, + 123, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 123, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 123, + 357 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 170, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 170, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 170, + 373 + ], + "type": "text", + "content": "5.1. Experimental Setting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "text", + "content": "Network Architecture. Unless otherwise specified, we use LLaVA-7B-v1-1 or LLaVA-13B-v1-1 [29] as the base multimodal LLM " + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "text", + "content": ", and adopt the ViT-H SAM [19] backbone as the vision backbone " + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{enc}" + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "text", + "content": ". The projection layer of " + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 376, + 288, + 437 + ], + "type": "text", + "content": " is an MLP with channels of [256, 4096, 4096]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "content": "Implementation Details. We adopt 8 NVIDIA 24G 3090 GPUs for training. The training scripts are based on deep-speed [41] engine. We use AdamW [33] optimizer with the learning rate and weight decay set to 0.0003 and 0, respectively. We also adopt WarmupDecayLR as the learning rate scheduler, where the warmup iterations are set to 100. The weights of the text generation loss " + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "content": " and the mask loss " + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "content": " are set to 1.0 and 1.0, respectively, and those of the bce loss " + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{bce}" + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "content": " and the dice loss " + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "inline_equation", + "content": "\\lambda_{dice}" + }, + { + "bbox": [ + 46, + 443, + 288, + 598 + ], + "type": "text", + "content": " are set to 2.0 and 0.5, respectively. Besides, the batch size per device is set to 2, and the gradient accumulation step is set to 10. During training, we select at most 3 categories for each image in semantic segmentation datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Datasets. As mentioned in Sec. 4.2, our training data is mainly composed of three types of datasets: (1) For the semantic segmentation dataset, we use ADE20K [62] and COCO-Stuff [3]. Besides, to enhance the segmentation result for some part of an object, we also use part semantic segmentation datasets, including PACO-LVIS [40], PartImageNet [13], and PASCAL-Part [6]; (2) For the referring segmentation dataset, we use refCLEF, refCOCO, refCOCO+ [17], and refCOCOg [35]; (3) For the visual ques" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 344, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 547, + 440 + ], + "type": "text", + "content": "tion answering (VQA) dataset, we use the datasets of LLaVA-Instruct-150k for LLaVA v1 [29] and LLaVA-v1.5-mix665k for LLaVA v1.5 [28]. In order to avoid data leakage, we exclude the COCO samples whose images are present in the refCOCO " + }, + { + "bbox": [ + 304, + 344, + 547, + 440 + ], + "type": "inline_equation", + "content": "(+ / \\mathrm{g})" + }, + { + "bbox": [ + 304, + 344, + 547, + 440 + ], + "type": "text", + "content": " validation sets during training. Furthermore, we surprisingly find that by fine-tuning the model on only 239 ReasonSeg data samples, the model's performance can be further boosted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 447, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 532 + ], + "type": "text", + "content": "Evaluation Metrics. We follow most previous works on referring segmentation [17, 35] to adopt two metrics: gIoU and cIoU. gIoU is defined by the average of all per-image Intersection-over-Unions (IoUs), while cIoU is defined by the cumulative intersection over the cumulative union. Since cIoU is highly biased toward large-area objects and it fluctuates too much, gIoU is preferred." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 539, + 481, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 481, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 481, + 552 + ], + "type": "text", + "content": "5.2. Reasoning Segmentation Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "content": "The reasoning segmentation results are shown in Table 1. It is worth noting that existing works fail to handle the task, but our model can accomplish the task involving complex reasoning with more than " + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 557, + 547, + 689 + ], + "type": "text", + "content": " gIoU performance boost. As mentioned before, the reasoning segmentation task is essentially different from the referring segmentation task in that it requires the model to possess reasoning ability or access world knowledge. Only by truly understanding the query, can the model do well in the task. The existing works have no proper way to understand an implicit query, but our model exploits multimodal LLMs to reach the goal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "Notably, we also make a comparison with the vanilla two-stage method (LLaVA1.5 + OVSeg). Specifically, the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 315, + 757 + ], + "type": "text", + "content": "9584" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 85, + 527, + 232 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 457, + 82 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 457, + 82 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 457, + 82 + ], + "type": "text", + "content": "Table 2. Referring segmentation results (cIoU) among LISA (ours) and existing methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 85, + 527, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 85, + 527, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 85, + 527, + 232 + ], + "type": "table", + "html": "
MethodrefCOCOrefCOCO+refCOCOg
valtestAtestBvaltestAtestBval(U)test(U)
MCN [34]62.464.259.750.655.044.749.249.4
VLT [11]67.570.565.256.361.050.155.057.7
CRIS [48]70.573.266.162.368.153.759.960.4
LAVT [53]72.775.868.862.168.455.161.262.1
ReLA [27]73.876.570.266.071.057.765.066.0
X-Decoder [65]------64.6-
SEEM [66]------65.7-
LISA-7B74.176.571.162.467.456.566.468.5
LISA-7B (fine-tuned on ReferSeg)74.979.172.365.170.858.167.970.6
", + "image_path": "6484f12817af483ce2ee6ba6e0fadd9e2d2b7b91723783225abcbbe1b5beb2c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 247, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 413 + ], + "type": "text", + "content": "two-stage method refers to first using a multimodal LLM (e.g., LLaVA v1.5) to generate a text output for the input query, and then adopting a referring or open-vocabulary segmentation model (e.g., OVSeg) to generate the segmentation mask. If the intermediate text output remains too long and exceeds the input token length limit of OVSeg, we use GPT-3.5 to further summarize. More details can be found in the supplementary material. The results in Table 1 show that our model outperforms the two-stage method significantly. We explain that the potential reasons are: 1) Our model is trained end-to-end, while the two-stage method is completely decoupled; 2) The two-stage method relies on text as an intermediary to transmit information, while our model utilizes the hidden embedding that is more expressive." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "content": "Another finding is that LISA-13B outperforms the 7B counterpart substantially, especially on the long-query scenarios, which indicates that the current performance bottleneck may still lie in understanding the query text, and a stronger multimodal LLM (e.g., LLaVA v1.5 [28]) leads to even better results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 492, + 255, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 492, + 255, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 492, + 255, + 506 + ], + "type": "text", + "content": "5.3. Vanilla Referring Segmentation Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 511, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 583 + ], + "type": "text", + "content": "To show that our model is also competent in the vanilla referring segmentation task, we make a comparison with existing state-of-the-art methods in Table 2. We evaluate the methods on refCOCO, refCOCO+, refCOCOg validation and testing sets. Our model achieves state-of-the-art results across various referring segmentation benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 141, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 141, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 141, + 604 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 609, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 288, + 657 + ], + "type": "text", + "content": "In this section, we conduct an extensive ablation study to reveal the contribution of each component. Unless otherwise specified, we report the metrics of gIoU and cIoU of LISA-7B on the validation set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Design Choices of Vision Backbone. We emphasize that vision backbones other than SAM are also applicable in our framework. In Table 3, we notice that SAM performs the best, potentially because of the massive high-quality" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 275, + 537, + 358 + ], + "blocks": [ + { + "bbox": [ + 304, + 244, + 547, + 267 + ], + "lines": [ + { + "bbox": [ + 304, + 244, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 547, + 267 + ], + "type": "text", + "content": "Table 3. Ablation study on the design choice of vision backbone. 'ft' denotes finetuning on ReasonSeg training set." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 275, + 537, + 358 + ], + "lines": [ + { + "bbox": [ + 315, + 275, + 537, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 537, + 358 + ], + "type": "table", + "html": "
Vision BackbonegIoUcIoU
Mask2Former-Swin-L42.438.8
SAM (w/ LoRA)41.537.3
SAM44.446.0
Mask2Former-Swin-L (ft)50.752.3
SAM w/ LORA (ft)51.851.9
SAM (ft)52.954.0
", + "image_path": "767075beabbddf423e44b55c9701d62bf2ba8f91cfbd36f250d160def6a976fc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 391, + 526, + 447 + ], + "blocks": [ + { + "bbox": [ + 306, + 371, + 544, + 383 + ], + "lines": [ + { + "bbox": [ + 306, + 371, + 544, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 544, + 383 + ], + "type": "text", + "content": "Table 4. Ablation study on SAM pre-trained weight and rephrasing." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 391, + 526, + 447 + ], + "lines": [ + { + "bbox": [ + 326, + 391, + 526, + 447 + ], + "spans": [ + { + "bbox": [ + 326, + 391, + 526, + 447 + ], + "type": "table", + "html": "
Exp. IDPre-trainSAM γrephrasinggIoUcIoU
1✓35.944.6
2✓50.751.1
3✓✓52.954.0
", + "image_path": "4aef72fdc80aa3c6f149ae43a7ac9d2390e6f7cbae94a40fcb95a3bc3b1d9207.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 469, + 547, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 540 + ], + "type": "text", + "content": "data used in its pre-training phase. Further, we also find that with the Mask2Former backbone, our framework still achieves a decent performance on the reasoning segmentation task, significantly outperforming previous works such as X-Decoder [65]. This reveals the fact that the design choice of vision backbone is flexible and not limited to SAM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "SAM LoRA Fintuning. We also investigate the effectiveness of applying LoRA on the SAM backbone. In Table 3, we note that the performance of LoRA fine-tuned SAM backbone is inferior to that of the frozen one. A potential reason is that fine-tuning impairs the generalization ability of the original SAM model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 645, + 547, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 645, + 547, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 547, + 717 + ], + "type": "text", + "content": "SAM Pre-trained Weight. To demonstrate the contribution of SAM pre-trained weight, we make a comparison between Experiments 1 and 3 in Table 4. Without being initialized with SAM pre-trained weight, the vision backbone is trained from scratch. This causes the performance to fall substantially behind that of the baseline model." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9585" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 74, + 83, + 92 + ], + "blocks": [ + { + "bbox": [ + 63, + 74, + 83, + 92 + ], + "lines": [ + { + "bbox": [ + 63, + 74, + 83, + 92 + ], + "spans": [ + { + "bbox": [ + 63, + 74, + 83, + 92 + ], + "type": "image", + "image_path": "b91107beaf4c0b451f08e5a024c9e9ad8874bdbeffaf22947e330bd3664e0bab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 85, + 148, + 107 + ], + "lines": [ + { + "bbox": [ + 89, + 85, + 148, + 107 + ], + "spans": [ + { + "bbox": [ + 89, + 85, + 148, + 107 + ], + "type": "text", + "content": "the most likely object that someone else has left behind" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 77, + 204, + 118 + ], + "blocks": [ + { + "bbox": [ + 157, + 77, + 204, + 118 + ], + "lines": [ + { + "bbox": [ + 157, + 77, + 204, + 118 + ], + "spans": [ + { + "bbox": [ + 157, + 77, + 204, + 118 + ], + "type": "image", + "image_path": "c6e62571977c7576cab40e12bc252d7da23aa18eef0076e5b291d23fa030a307.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 74, + 227, + 93 + ], + "lines": [ + { + "bbox": [ + 212, + 74, + 227, + 93 + ], + "spans": [ + { + "bbox": [ + 212, + 74, + 227, + 93 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 77, + 529, + 119 + ], + "blocks": [ + { + "bbox": [ + 233, + 77, + 529, + 119 + ], + "lines": [ + { + "bbox": [ + 233, + 77, + 529, + 119 + ], + "spans": [ + { + "bbox": [ + 233, + 77, + 529, + 119 + ], + "type": "image", + "image_path": "d3e02c8979e997296a5193f45fac33af6ce2867cb60ef537857b971296807063.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 127, + 83, + 145 + ], + "blocks": [ + { + "bbox": [ + 63, + 127, + 83, + 145 + ], + "lines": [ + { + "bbox": [ + 63, + 127, + 83, + 145 + ], + "spans": [ + { + "bbox": [ + 63, + 127, + 83, + 145 + ], + "type": "image", + "image_path": "0328c0b81c2a3acb2431462e655a554574b609d45a47ee72d07573ad59c65150.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 142, + 148, + 158 + ], + "lines": [ + { + "bbox": [ + 89, + 142, + 148, + 158 + ], + "spans": [ + { + "bbox": [ + 89, + 142, + 148, + 158 + ], + "type": "text", + "content": "the object used for stirring milk or coffee" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 157, + 128, + 204, + 171 + ], + "blocks": [ + { + "bbox": [ + 157, + 128, + 204, + 171 + ], + "lines": [ + { + "bbox": [ + 157, + 128, + 204, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 128, + 204, + 171 + ], + "type": "image", + "image_path": "363392ea772e52f662d1f5a0226e75995b4b73c0dfe33b0aa8b262d60d0d15dd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 125, + 227, + 145 + ], + "lines": [ + { + "bbox": [ + 212, + 125, + 227, + 145 + ], + "spans": [ + { + "bbox": [ + 212, + 125, + 227, + 145 + ], + "type": "text", + "content": "W" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 233, + 128, + 529, + 171 + ], + "blocks": [ + { + "bbox": [ + 233, + 128, + 529, + 171 + ], + "lines": [ + { + "bbox": [ + 233, + 128, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 233, + 128, + 529, + 171 + ], + "type": "image", + "image_path": "66712d38cbc11cb3ba2c75df85d01a2a5be003753822b0e908f115e0fbf39cbc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 63, + 180, + 83, + 198 + ], + "blocks": [ + { + "bbox": [ + 63, + 180, + 83, + 198 + ], + "lines": [ + { + "bbox": [ + 63, + 180, + 83, + 198 + ], + "spans": [ + { + "bbox": [ + 63, + 180, + 83, + 198 + ], + "type": "image", + "image_path": "111f9f6fe104f4630a753969a7ecf69f0a27d13f3daa462bc1611ef514988db2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 183, + 156, + 221 + ], + "lines": [ + { + "bbox": [ + 88, + 183, + 156, + 221 + ], + "spans": [ + { + "bbox": [ + 88, + 183, + 156, + 221 + ], + "type": "text", + "content": "In cold days, dogs may need extra protection to keep them warm. What object in the picture can a dog wear to provide warmth during snowy walks?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 158, + 186, + 204, + 221 + ], + "blocks": [ + { + "bbox": [ + 158, + 186, + 204, + 221 + ], + "lines": [ + { + "bbox": [ + 158, + 186, + 204, + 221 + ], + "spans": [ + { + "bbox": [ + 158, + 186, + 204, + 221 + ], + "type": "image", + "image_path": "0598a171e198f06f2bb5e16d8bb704744bc991bfb0764a403a0b76b1d8596b57.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 178, + 227, + 198 + ], + "lines": [ + { + "bbox": [ + 212, + 178, + 227, + 198 + ], + "spans": [ + { + "bbox": [ + 212, + 178, + 227, + 198 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 233, + 186, + 529, + 221 + ], + "blocks": [ + { + "bbox": [ + 233, + 186, + 529, + 221 + ], + "lines": [ + { + "bbox": [ + 233, + 186, + 529, + 221 + ], + "spans": [ + { + "bbox": [ + 233, + 186, + 529, + 221 + ], + "type": "image", + "image_path": "5d9cba76fae1045402466d9cdcb5726b34c1d5ca2a8e70ae264e677b64c107d3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 63, + 232, + 83, + 249 + ], + "blocks": [ + { + "bbox": [ + 63, + 232, + 83, + 249 + ], + "lines": [ + { + "bbox": [ + 63, + 232, + 83, + 249 + ], + "spans": [ + { + "bbox": [ + 63, + 232, + 83, + 249 + ], + "type": "image", + "image_path": "88fc045aea040343be95bd9dd217e2b1e2b22fca5acffd5e435a8bdc33439f05.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 158, + 232, + 204, + 274 + ], + "blocks": [ + { + "bbox": [ + 88, + 238, + 157, + 266 + ], + "lines": [ + { + "bbox": [ + 88, + 238, + 157, + 266 + ], + "spans": [ + { + "bbox": [ + 88, + 238, + 157, + 266 + ], + "type": "text", + "content": "When a plane is ready to land on the airport runway, what area in the picture will it eventually land on?" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 232, + 204, + 274 + ], + "lines": [ + { + "bbox": [ + 158, + 232, + 204, + 274 + ], + "spans": [ + { + "bbox": [ + 158, + 232, + 204, + 274 + ], + "type": "image", + "image_path": "6d51fb2f1098a87272c673d6d1e48ef57dc26868c87078154069b93eae9d8ed1.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 280, + 129, + 288 + ], + "lines": [ + { + "bbox": [ + 111, + 280, + 129, + 288 + ], + "spans": [ + { + "bbox": [ + 111, + 280, + 129, + 288 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 171, + 280, + 190, + 288 + ], + "lines": [ + { + "bbox": [ + 171, + 280, + 190, + 288 + ], + "spans": [ + { + "bbox": [ + 171, + 280, + 190, + 288 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 233, + 232, + 529, + 289 + ], + "blocks": [ + { + "bbox": [ + 212, + 231, + 227, + 250 + ], + "lines": [ + { + "bbox": [ + 212, + 231, + 227, + 250 + ], + "spans": [ + { + "bbox": [ + 212, + 231, + 227, + 250 + ], + "type": "text", + "content": "中" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 232, + 529, + 289 + ], + "lines": [ + { + "bbox": [ + 233, + 232, + 529, + 289 + ], + "spans": [ + { + "bbox": [ + 233, + 232, + 529, + 289 + ], + "type": "image", + "image_path": "7c8a06e77cf69b1d705d409df03f16a34e67570f394d9b3a4c193a14fb1e244d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 289, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 545, + 300 + ], + "type": "text", + "content": "Figure 5. Visual comparison among LISA (ours) and existing related methods. More illustrations are given in the supplementary material." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 48, + 320, + 288, + 445 + ], + "blocks": [ + { + "bbox": [ + 93, + 306, + 241, + 316 + ], + "lines": [ + { + "bbox": [ + 93, + 306, + 241, + 316 + ], + "spans": [ + { + "bbox": [ + 93, + 306, + 241, + 316 + ], + "type": "text", + "content": "Table 5. Ablation study on training data." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 320, + 288, + 445 + ], + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 445 + ], + "type": "table", + "html": "
IDSemanticSegReferSegVQAReasonSeggIoUcIoU
ADE20KCOCO-StuffPartSeg
1✓✓✓✓✓48.953.5
2✓✓✓✓✓48.550.8
3✓✓✓✓✓46.750.9
4✓✓✓✓46.646.7
5✓✓✓30.420.4
6✓✓✓✓✓47.751.1
7✓✓✓✓✓44.446.0
8✓✓✓✓✓✓52.954.0
", + "image_path": "171b404cdbad77ac6cd077f2514f278cf6ae92acdae4ffe50edc975db809f56e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "type": "table", + "bbox": [ + 67, + 467, + 263, + 509 + ], + "blocks": [ + { + "bbox": [ + 89, + 453, + 244, + 464 + ], + "lines": [ + { + "bbox": [ + 89, + 453, + 244, + 464 + ], + "spans": [ + { + "bbox": [ + 89, + 453, + 244, + 464 + ], + "type": "text", + "content": "Table 6. Results on the ReasonSeg test set." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 467, + 263, + 509 + ], + "lines": [ + { + "bbox": [ + 67, + 467, + 263, + 509 + ], + "spans": [ + { + "bbox": [ + 67, + 467, + 263, + 509 + ], + "type": "table", + "html": "
Training splits# data samplesgIoUcIoU
train23951.751.1
train + val43954.054.9
", + "image_path": "7b443a0f6de179f9cf58126ed246759bc36e6fb9a03a81f7c673274e8f71b1bc.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "text", + "content": "Instruction Rephrasing by GPT-3.5. When fine-tuning the model on the reasoning segmentation data samples, we rephrase the text instruction by GPT-3.5 (the details are shown in the supplementary material), and randomly choose one. The comparison between Experiments 2 and 3 in Table 4 shows that the performance is increased by " + }, + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "text", + "content": " gIoU and " + }, + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 46, + 520, + 287, + 616 + ], + "type": "text", + "content": " cIoU. This result verifies the effectiveness of such data augmentation." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Contribution of All Types of Training Data. In Table 5, we show the contribution of each type of data to the performance. We find that in Exp. 5, we do not use any semantic segmentation dataset, and the performance drops a lot. We conjecture that semantic segmentation datasets provide a large amount of ground-truth binary masks for training, since a multi-class label can induce multiple binary masks." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 307, + 547, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 380 + ], + "type": "text", + "content": "We also notice that adding more reasoning segmentation data samples during training leads to better results. In Table 6, we also add the ReasonSeg val set (200 data samples) during fine-tuning, and it yields better performance in both gIoU and cIoU metrics. This indicates that more reasoning segmentation training samples are beneficial at this moment." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 387, + 417, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 387, + 417, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 387, + 417, + 399 + ], + "type": "text", + "content": "5.5. Qualitative Results" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 406, + 547, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 502 + ], + "type": "text", + "content": "As depicted in Fig. 5, we provide a visual comparison with existing related works, including the model for open-vocabulary semantic segmentation (OVSeg), referring segmentation (GRES), and the generalist models for segmentation (X-Decoder and SEEM). These models fail to handle the displayed cases with various errors, while our approach produces accurate and high-quality segmentation results. More illustrations are given in the supplementary material." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 378, + 526 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 534, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 641 + ], + "type": "text", + "content": "In this work, we have proposed a new segmentation task—reasoning segmentation. Also, we have introduced an evaluation benchmark ReasonSeg, which comprises over one thousand data samples. Finally, we have presented our model — LISA. It injects segmentation capabilities into current multimodal LLMs and performs surprisingly effectively on the reasoning segmentation task. We hope our work can shed new light on the direction of combining LLMs and vision tasks in the future." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 648, + 408, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 648, + 408, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 648, + 408, + 662 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "This work is supported in part by the Research Grants Council under the Areas of Excellence scheme grant AoE/E-601/22-R and the Shenzhen Science and Technology Program under No. KQTD20210811090149095." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9586" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 145 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 181 + ], + "type": "text", + "content": "[2] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. TPAMI, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 183, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 183, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 183, + 288, + 216 + ], + "type": "text", + "content": "[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 261 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 263, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 288, + 308 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 354 + ], + "type": "text", + "content": "[6] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In CVPR, 2014. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 356, + 288, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 356, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 356, + 288, + 400 + ], + "type": "text", + "content": "[7] Bowen Cheng, Maxwell D Collins, Yukun Zhu, Ting Liu, Thomas S Huang, Hartwig Adam, and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 435 + ], + "type": "text", + "content": "[8] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 437, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 479 + ], + "type": "text", + "content": "[9] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 483, + 288, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 288, + 526 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 288, + 562 + ], + "type": "text", + "content": "[11] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 564, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 288, + 597 + ], + "type": "text", + "content": "[12] Jun Fu, Jing Liu, Haijie Tian, Yong Li, Yongjun Bao, Zhiwei Fang, and Hanqing Lu. Dual attention network for scene segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 288, + 643 + ], + "type": "text", + "content": "[13] Ju He, Shuo Yang, Shaokang Yang, Adam Kortylewski, Xiaoding Yuan, Jie-Neng Chen, Shuai Liu, Cheng Yang, Qihang Yu, and Alan Yuille. Partimagenet: A large, high-quality dataset of parts. In ECCV, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 644, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 288, + 667 + ], + "type": "text", + "content": "[14] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv:2106.09685, 2021. 4, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[16] Zilong Huang, Xinggang Wang, Lichao Huang, Chang Huang, Yunchao Wei, and Wenyu Liu. Ccnet: Criss-cross attention for semantic segmentation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 139 + ], + "type": "text", + "content": "[17] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 171 + ], + "type": "text", + "content": "[18] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 547, + 216 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv:2304.02643, 2023. 2, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 217, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 546, + 249 + ], + "type": "text", + "content": "[20] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 251, + 547, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 547, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 547, + 315 + ], + "type": "text", + "content": "[21] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 316, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 547, + 360 + ], + "type": "text", + "content": "[22] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 403 + ], + "type": "text", + "content": "[23] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning. arXiv:2305.03726, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv:2301.12597, 2023. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 546, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 546, + 481 + ], + "type": "text", + "content": "[25] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 546, + 525 + ], + "type": "text", + "content": "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 526, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 547, + 557 + ], + "type": "text", + "content": "[27] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In CVPR, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint, 2023. 1, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 593, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 546, + 624 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv:2304.08485, 2023. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 547, + 678 + ], + "type": "text", + "content": "[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pretraining for open-set object detection. arXiv preprint, 2023. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 680, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 547, + 712 + ], + "type": "text", + "content": "[31] Wei Liu, Andrew Rabinovich, and Alexander C. Berg. Parsenet: Looking wider to see better. arXiv preprint, 2015. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9587" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[32] Zhaoyang Liu, Yinan He, Wenhai Wang, Weiyun Wang, Yi Wang, Shoufa Chen, Qinglong Zhang, Yang Yang, Qingyun Li, Jiashuo Yu, et al. Internchat: Solving vision-centric tasks by interacting with chatbots beyond language. arXiv:2305.05662, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 152 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 196 + ], + "type": "text", + "content": "[34] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 242 + ], + "type": "text", + "content": "[35] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "text", + "content": "[36] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 288, + 312 + ], + "type": "text", + "content": "[37] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In ICCV, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 288, + 357 + ], + "type": "text", + "content": "[38] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 359, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 288, + 403 + ], + "type": "text", + "content": "[39] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, and Tong Zhang. Detgpt: Detect what you need via reasoning. arXiv:2305.14167, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 449 + ], + "type": "text", + "content": "[40] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In CVPR, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 450, + 288, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 288, + 494 + ], + "type": "text", + "content": "[41] Jeff Rasley, Samyam Rajbhandari, Olatunjri Ruwase, and Yuxiong He. Deepspeed: System optimizations enable training deep learning models with over 100 billion parameters. In SIGKDD, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 496, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 496, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 496, + 288, + 529 + ], + "type": "text", + "content": "[42] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 531, + 288, + 563 + ], + "type": "text", + "content": "[43] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. TPAMI, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 288, + 609 + ], + "type": "text", + "content": "[44] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv:2303.17580, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 654 + ], + "type": "text", + "content": "[45] Zhuotao Tian, Pengguang Chen, Xin Lai, Li Jiang, Shu Liu, Hengshuang Zhao, Bei Yu, Ming-Chang Yang, and Jiaya Jia. Adaptive perspective distillation for semantic segmentation. TPAMI, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 656, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 288, + 689 + ], + "type": "text", + "content": "[46] Zhuotao Tian, Jiequan Cui, Li Jiang, Xiaojuan Qi, Xin Lai, Yixin Chen, Shu Liu, and Jiaya Jia. Learning context-aware classifier for semantic segmentation. AAAI, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv:2305.11175, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[48] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 547, + 184 + ], + "type": "text", + "content": "[49] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv:2303.04671, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "text", + "content": "[50] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 251 + ], + "type": "text", + "content": "[51] Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. In CVPR, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[52] Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction. arXiv:2305.18752, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 287, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 329 + ], + "type": "text", + "content": "[53] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavt: Language-aware vision transformer for referring image segmentation. In CVPR, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 547, + 385 + ], + "type": "text", + "content": "[54] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv:2303.11381, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 388, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 547, + 441 + ], + "type": "text", + "content": "[55] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv:2304.14178, 2023. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 443, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 547, + 464 + ], + "type": "text", + "content": "[56] Fisher Yu and Vladlen Koltun. Multi-scale context aggregation by dilated convolutions. In ICLR, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 466, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 466, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 466, + 547, + 510 + ], + "type": "text", + "content": "[57] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv:2307.03601, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 511, + 547, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 511, + 547, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 511, + 547, + 543 + ], + "type": "text", + "content": "[58] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-net: Towards unified image segmentation. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 577 + ], + "type": "text", + "content": "[59] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 611 + ], + "type": "text", + "content": "[60] Hengshuang Zhao, Xiaojuan Qi, Xiaoyong Shen, Jianping Shi, and Jiaya Jia. Icnet for real-time semantic segmentation on high-resolution images. In ECCV, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 613, + 547, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 655 + ], + "type": "text", + "content": "[61] Hengshuang Zhao, Yi Zhang, Shu Liu, Jianping Shi, Chen Change Loy, Dahua Lin, and Jiaya Jia. Pisanet: Pointwise spatial attention network for scene parsing. In ECCV, 2018. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 689 + ], + "type": "text", + "content": "[62] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[63] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9588" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 207 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "language understanding with advanced large language models. arXiv:2304.10592, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 129 + ], + "type": "text", + "content": "[64] Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, and Xi-ang Bai. Asymmetric non-local neural networks for semantic segmentation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 288, + 174 + ], + "type": "text", + "content": "[65] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In CVPR, 2023. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 207 + ], + "type": "text", + "content": "[66] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. arXiv:2304.06718, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9589" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_content_list.json b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..47a00291f548fbee945c1827ea2adfaaba6f90d6 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding, Reasoning, and Planning", + "text_level": 1, + "bbox": [ + 227, + 130, + 743, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sijin Chen $^{1}$ Xin Chen $^{2,*}$ Chi Zhang $^{2}$ Mingsheng Li $^{1}$ Gang YU $^{2}$", + "bbox": [ + 191, + 202, + 772, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Fei $^{3}$ Hongyuan Zhu $^{4}$ Jiayuan Fan $^{1}$ Tao Chen $^{1,\\dagger}$", + "bbox": [ + 251, + 220, + 715, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fudan University $^{2}$ Tencent PCG $^{3}$ National University of Singapore", + "bbox": [ + 240, + 239, + 730, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore", + "bbox": [ + 140, + 257, + 828, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/Open3DA/LL3DA", + "bbox": [ + 341, + 277, + 627, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* project lead † corresponding author", + "bbox": [ + 361, + 294, + 604, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5943c9dff48a6fa1c0ca0e1f85368bc41d1194b3b08c82a1d0f10a40fbd56f44.jpg", + "image_caption": [ + "Figure 1. We propose LL3DA, a Large Language 3D Assistant that demonstrates mighty instruction-following capacities in understanding, reasoning, and planning within complex 3D environments. LL3DA takes both the textual instructions and potential visual interactions into consideration to help remove ambiguities when addressing various tasks in diverse and complex 3D scenes." + ], + "image_footnote": [], + "bbox": [ + 81, + 343, + 885, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 662, + 313, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent progress in Large Multimodal Models (LMM) has opened up great possibilities for various applications in the field of human-machine interactions. However, developing LMMs that can comprehend, reason, and plan in complex and diverse 3D environments remains a challenging topic, especially considering the demand for understanding permutation-invariant point cloud representations of the 3D scene. Existing works seek help from multi-view images by projecting 2D features to 3D space, which inevitably leads to huge computational overhead and performance degradation. In this paper, we present LL3DA, a Large Language 3D Assistant that takes point cloud as the direct input and responds to both text instructions and visual interactions. The additional visual interaction enables LMMs to better", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "comprehend human interactions with the 3D environment and further remove the ambiguities within plain texts. Experiments show that LL3DA achieves remarkable results and surpasses various 3D vision-language models on both 3D Dense Captioning and 3D Question Answering.", + "bbox": [ + 496, + 662, + 893, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 753, + 632, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent surge in Large Language Model (LLM) families [14, 31, 46, 52, 63] opens up great opportunities for addressing various machine learning tasks in a generalized way [30, 32, 40, 60]. During this LLM carnival, researchers are also seeking generalized LLM solutions for various vision language tasks [37, 50, 62]. Among these, LLM-based 3D scene understanding is a valuable topic that would benefit the development of autonomous driving [9, 23] and em", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26428", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bodied AI agents [21, 49]. However, it is also challenging given 1) the diversity and complication of 3D environments and 2) the demands for understanding sparse 3D points.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior works have made initial success in various 3D vision and language tasks. The majority of these research build 3D specialists to solve one specific down-stream task, including 3D Question Answering (3D-QA) [2, 43], 3D Visual Grounding (3D-VG) [7, 26, 55], and 3D Dense Captioning (3D-DC) [10-12]. There are also several works [4, 13, 34, 70] study the mutual promotion of different 3D vision and language tasks with shared structure modelling relations among objects. Recently, researchers have also introduced LLMs for general purpose 3D understanding, where Point-Bind and Point-LLMs [24, 57] mainly focus on the understanding of 3D objects. Concurrently, 3D-LLM [29] proposes an LLM-driven solution that aggregates multi-view features for 3D perception, presenting mighty capacities in understanding 3D object and scenes and following text instructions produced by human.", + "bbox": [ + 75, + 137, + 470, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though these methods have achieved remarkable success in addressing various challenges in understanding the 3D world with natural language, there are certain limitations. With limited supervision, 3D specialists could hardly scale-up for better performance, while the joint pre-training still requires separate heads for specific tasks. Extracting multi-view features results in huge computational overhead and ignores the essential geometry and depth information. Additionally, plain texts often lead to ambiguities especially in cluttered and complex 3D environments.", + "bbox": [ + 75, + 380, + 468, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose LL3DA, a Large Language 3D Assistant that responds to both textual and visual interactions from human, with the ability to understand, reason, and plan in complex 3D environments (Fig. 1). We adopt a multi-modal transformer to aggregate information from textual instructions, visual prompts, and the 3D scene into a fixed length of learnable querying tokens via the attention mechanism. The querying tokens are then projected and used as the prefix for the textual instructions, serving as the input to a pre-trained and frozen LLM. This design not only helps to address the contradiction between the permutation-invariant 3D scene representations and the LLM embedding space, but also extracts interaction-aware 3D scene representations for efficient instruction following.", + "bbox": [ + 75, + 532, + 468, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments to explore the capacities of LL3DA in understanding, reasoning, and planning within complex and diverse 3D environments. Our model achieves state-of-the-art results on two widely used datasets for 3D Dense Captioning [1, 7], and 3D Question Answering [2]. Additionally, by introducing additional visual interactions, our method could further remove the ambiguities within the vague textual instructions.", + "bbox": [ + 75, + 744, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our key contributions lie in:", + "bbox": [ + 96, + 866, + 382, + 880 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We present a LLM-based solution for understanding, rea", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "soning, and planning in complex 3D environments.", + "bbox": [ + 511, + 90, + 851, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our model takes both the textual instructions and visual interactions as inputs, and extracts interaction-aware features for effective instruction-following.", + "- Extensive experiments show that our method surpasses various state-of-the-art 3D vision language models." + ], + "bbox": [ + 500, + 106, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 193, + 640, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Vision and Language alignment, pre-training, and understanding [5, 7, 20, 70] cover tasks requiring a model to adopt its understanding towards a complex 3D scene answering to, or answering with natural language. Among those, 3D Dense Captioning (3D-DC) [10, 12, 54] expects a model to translate an input 3D scene into a set of instance coordinates and natural language descriptions. Existing methods could be categorized into \"detect-then-describe\" models [4, 12, 54] and the \"set-to-set\" prediction approaches [10, 11]. The former builds explicit relations on the instance coordinate estimations, while the latter directly learns the locations and descriptions for instances from the input 3D scene. 3D Visual Grounding (3D-VG) [1, 7, 55] demands a model to respond the natural language queries with the instance coordinates in the 3D scene. The mainstream of existing methods [4, 65, 70] address 3D-VG via selecting a candidate from a 3D detector's prediction. 3D Question Answering (3D-QA) [2, 43, 59, 66] requires a model to answer the questions with natural language based on the input 3D scene. The majority of existing methods [2, 18, 48] directly select the desired response from a given answer set. Researchers have also studied the mutual promotion of various 3D vision language tasks via training their shareable architectures simultaneously on multiple tasks [4, 13, 34, 70]. UniT3D [13] and 3DJCG [4] focus on the joint promotion between 3D-DC and 3D-VG in the relation modelling, while 3D-VLP [34] further includes 3D-QA. Recently, 3D-LLM [29] introduces a family of LLM-driven 3D generalists that could handle diverse textual instructions with reconstructed 3D features from multi-view images [28]. In this paper, we present LL3DA, an LLM solution that directly extracts features from the 3D scene, and handles both visual prompts and textual instructions to diversify the possible interactions human could make with the complex 3D environments.", + "bbox": [ + 496, + 219, + 893, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Multimodal Models (LMM). Along with the rapid development of Large Language Models (LLM) [15, 63], researchers have made great recent efforts adapting LLMs to visual understanding and reasoning tasks [25, 38, 56, 61]. Some methods project or compress global image features as the prefix for text instructions [36, 40, 58, 69], while others extract ROI features as LLM tokens for region-oriented instruction reasoning [6, 64]. Meanwhile, InstructBLIP [17] proposes to extract textual instruction-aware visual features, and has achieved remarkable success in addressing complex", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26429", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6179a1c6621ee1b02340e9cd76f3c56c9cc2bbd5cbc2f53b30c3188e4be101a3.jpg", + "image_caption": [ + "Figure 2. Overview of the Proposed Approach. (a) The overall pipeline of our proposed LL3DA first extracts interaction-aware 3D scene embeddings, which are later projected to the prefix of textual instructions as the input of a frozen LLM. (b) The detailed design of the Interactor3D, which aggregates visual prompts, textual instructions, and 3D scene embeddings into a fixed length querying tokens. (c) The prompt encoder encodes the user clicks and box coordinates with the positional embeddings and ROI features, respectively." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 885, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and unseen instructions. Concurrently, researchers have also made great attempts solving various 3D tasks using LLMs. Notably, [24, 42, 57, 68] demonstrate remarkable success in understanding and reasoning about 3D objects. In this paper, we present an LLM-driven solution that could handle both interactions in forms of visual prompts and text instructions. We also propose to extract interaction-aware 3D scene representations for better instruction following.", + "bbox": [ + 75, + 414, + 472, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 550, + 212, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To build a general purpose agent that could handle both visual and textual interactions with complex 3D environments, we propose LL3DA, an LLM driven auto-regressive approach to 3D vision language tasks. In this section, we first introduce the problem formatting in Sec. 3.1. Next, we introduce our model designs in details (Sec. 3.2).", + "bbox": [ + 75, + 575, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formatting", + "text_level": 1, + "bbox": [ + 76, + 676, + 272, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model I/O. As shown in Fig. 2 (a), the input of our model consists of a 3D scene represented by a set of points $PC$ , the textual instruction $\\mathcal{I}_t$ , and potential visual interactions $\\mathcal{I}_v$ that serve as supplementary spatial identifiers. Here, point cloud $PC = [p_{in}, f_{in}] \\in \\mathbb{R}^{N \\times (3 + F)}$ , where $p_{in} \\in \\mathbb{R}^{N \\times 3}$ and $f_{in} \\in \\mathbb{R}^{N \\times F}$ are the point coordinates and the additional point features, including color, normal, and height. The output of our model is free-form natural language, part of whom could be interpreted into 3D coordinates.", + "bbox": [ + 75, + 702, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruction Formatting. Following existing LMMs [57], we begin the textual instructions $\\mathcal{I}_t$ with the “human:” identifier, and ask the model to generate responses after the “assistant:” identifier. This endows the model with the", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ability to distinguish information from the context and further engage in multi-turn conversations.", + "bbox": [ + 496, + 414, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coordinate Representations. To provide LLMs with the capacity to perceive and respond with 3D coordinates, we convert the 3D points and 3D bounding boxes to plain texts. Specifically, a point is represented by “ $\\langle \\mathrm{loc} \\rangle x, y, z \\langle \\mathrm{loc} \\rangle$ ”, and a bounding box is represented by its center point and size, i.e. “ $\\langle \\mathrm{obj} \\rangle c_x, c_y, c_z, w, h, l \\langle \\mathrm{obj} \\rangle$ ”. Here, all the numerical data is discretized into unsigned integers within a range of [0, 255] with respect to the boundary of the input 3D scene. This design could naturally fit in the vocabulary of existing pre-trained LLMs [52, 63]. Without the introduction of any additional learnable tokens, we could save the effort of tuning the LLMs.", + "bbox": [ + 496, + 446, + 892, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Model Design", + "text_level": 1, + "bbox": [ + 500, + 633, + 643, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2 (a), our model first aggregates a fixed-length scene embeddings through the Interactor3D, which takes the visual prompts, the textual instructions, and the 3D scene as the input. Next, the aggregated 3D scene embeddings are projected to be the prefix of text instructions, and serve as the inputs of a frozen LLM. The detailed design of Interactor3D is shown in Fig. 2 (b), which consists of a frozen 3D scene encoder $\\mathcal{E}^{3D}$ , a visual prompt encoder, and a multi-modal transformer.", + "bbox": [ + 496, + 657, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene Encoder. We adopt the masked transformer encoder pre-trained on ScanNet detection [10] as the scene encoder, $\\mathcal{E}^{3D}$ . The scene encoder takes $PC$ as its input, and outputs the 3D scene embeddings:", + "bbox": [ + 496, + 795, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e n c} = \\mathcal {E} ^ {3 D} (P C) = \\mathcal {E} ^ {3 D} \\left(p _ {i n}; f _ {i n}\\right) \\in \\mathbb {R} ^ {M \\times d}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 859, + 890, + 880 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $f_{enc}$ consists of $d$ -dimensioned features for $M$ points", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26430", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uniformly down-sampled from the input 3D scene through the Farthest Point Sampling (FPS) algorithm. In practice, we choose to keep the scene encoder frozen to save the memory cost during training.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Prompt Encoder. We mainly take two common types of visual interactions into consideration, user clicks and 3D box annotations [35]. Each user click is first normalized within a range of $[0,1]$ by the size of the input 3D scene $p_{\\mathrm{click}} \\in \\mathbb{R}^3$ . Then, we encode $p_{\\mathrm{click}}$ with the 3D Fourier positional embeddings [51] function:", + "bbox": [ + 75, + 152, + 467, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right) = \\left[ \\sin \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right); \\cos \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right) \\right]. \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 255, + 468, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $B \\in \\mathbb{R}^{3 \\times (d/2)}$ is a learnable matrix. The box annotation is represented by the ROI feature $f_{\\mathrm{box}} \\in \\mathbb{R}^d$ extracted by a pre-trained 3D object detector [10]. The two types of the visual prompts are then projected with separate and identical Feed Forward Networks (FFN).", + "bbox": [ + 75, + 282, + 468, + 358 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {c l i c k}} = F F N _ {\\text {c l i c k}} \\left(\\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 369, + 468, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {b o x}} = F F N _ {\\text {b o x}} \\left(f _ {\\text {b o x}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 388, + 325, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we represent each visual prompt with 8 tokens.", + "bbox": [ + 75, + 416, + 467, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Modal Transformer (MMT) serves as a role to 1) address the contradiction between the permutation-invariant 3D scene embeddings and position-sensitive causal LLMs, 2) bridge the gap between frozen unimodal experts, and 3) fill the needs for interaction-aware feature extraction. Inspired by the Q-Former architecture [17, 36], MMT aggregates the visual information within a fixed number of 32 learnable querying tokens. In each layer, the queries interact with the encoded visual prompts $[f_{\\mathrm{click}}; f_{\\mathrm{box}}]$ and the textual instructions $\\mathcal{I}_t$ through a shared self-attention. Then, we allow the learnable querying tokens and the visual prompts to interact with the task-agnostic 3D scene embeddings $f_{enc}$ via cross-attention. The output of MMT is 32 queries written as $Q \\in \\mathbb{R}^{32 \\times 768}$ , which are finally projected to the word embedding space of LLMs through a simple linear projector. In practice, we notice that initializing Q-Former with pre-trained BERT [19, 36] weights will lead to repetitive outputs, thus we only choose to initialize the pre-trained word and position embeddings from BERT.", + "bbox": [ + 75, + 431, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM. We consider the decoder-only generative pre-trained transformers [52, 63] as our large language model backbone. The decoder-only LLMs are sensitive to the input orders because of the position embeddings and the causal attention mask. The parameters and the embedding layers of the LLMs are kept frozen to save memory cost. During inference, we generate the responses via searching for the optimal sequence $s^*$ that satisfies:", + "bbox": [ + 75, + 720, + 468, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {*} = \\arg \\max _ {s} P (s | P C, \\mathcal {I} _ {t}, \\mathcal {I} _ {v}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 853, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we use beam search with a beam size of 4.", + "bbox": [ + 76, + 885, + 434, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Multi-modal Instruction Tuning", + "text_level": 1, + "bbox": [ + 500, + 89, + 794, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A general purpose 3D agent is able to address various tasks simultaneously in complex 3D scenes. Apart from introducing proper training data, it is important to guide the model to generate the desired outputs with instructions. Therefore, Sec. 4.1 will first introduce how we identify each task. After that, Sec. 4.2 will present details for the training objective.", + "bbox": [ + 498, + 114, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Tasks and Instructions.", + "text_level": 1, + "bbox": [ + 498, + 214, + 714, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As introduced in Sec. 3.1, LL3DA generates text responses auto-regressively after the “```assistant:” identifier.", + "bbox": [ + 498, + 238, + 890, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Dense Captioning requires the localization and description of instances in diverse 3D environments. We adopt either user clicks and box annotations as the visual prompt to identify the object to be described. Additionally, we design two types of textual instructions that ask the model to either \"describe\" or \"describe and localize\" the object, which diversifies the tasks, and leads to better performance.", + "bbox": [ + 498, + 268, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Question Answering requires the model to generate response to the questions based on the global knowledge of a 3D scene. To help the model better understand the 3D environment, we also design two types of textual instructions that ask the model to either \"answer\" or \"answer and localize the related objects\". The latter serves as an auxiliary task widely adopted in various 3D-QA methods [2, 48]. To diversify the tasks during training, we randomly include additional clicks on the objects related to the questions.", + "bbox": [ + 498, + 376, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Description requires the model to translate its global knowledge of the 3D scene into natural language descriptions, thus we simply ask the \"describe\" this 3D scene.", + "bbox": [ + 498, + 513, + 890, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embodied Conversation and Planning could be treated as multi-turn conversations, where we use “human” and “assistant” as identifiers to distinguish the source of information as introduced in Sec. 3.1.", + "bbox": [ + 498, + 559, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Instruction Following Tuning", + "text_level": 1, + "bbox": [ + 500, + 628, + 761, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, for tasks requiring additional visual interactions, i.e. 3D-DC and 3D-QA, we randomly choose between clicks or boxes as means of object identification.", + "bbox": [ + 498, + 652, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objective. Our training objective is to optimize the trainable parameters $\\theta$ , so as to maximize the likelihood of the target response sequence $s$ given the input point cloud $PC$ , and the human interactions $\\mathcal{I}_v$ and $\\mathcal{I}_t$ :", + "bbox": [ + 498, + 699, + 890, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\arg \\max _ {\\theta} P (s | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 770, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, this is accomplished by adopting the token-wise cross-entropy loss that trains the model to predict the $i$ th token $s_{[i]}$ given the previous $(i - 1)$ tokens, $s_{[1,\\dots ,i - 1]}$ .", + "bbox": [ + 498, + 803, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = - \\sum_ {i = 1} ^ {| s |} \\log P (s _ {[ i ]} | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta ; s _ {[ 1, \\dots , i - 1 ]}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 861, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26431", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7a5e0fb81edaf65f47e466d5f2f4115c3e750eb243ad010bacb205c5e5cc50a1.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons for 3D Dense Captioning on ScanRefer[7] and Nr3D[1]. For fair comparison, we list methods that are trained under the standard per-word cross-entropy loss without additional 3D scenes. We use the box estimations from Vote2Cap-DETR to simulate the box annotations as the visual prompts. Our proposed LL3DA surpasses previous 3D specialists on both datasets." + ], + "table_footnote": [], + "table_body": "
MethodScanReferNr3D
C@0.25↑B-4@0.25↑M@0.25↑R@0.25↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[33]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[54]----44.0225.2622.3345.3633.7119.9222.6150.50
REMAN[44]62.0136.3726.7656.2545.0026.3122.6746.9634.8120.3723.0150.99
D3Net[8]----46.0730.2924.3551.6733.8520.7023.1353.38
Contextual[67]----46.1125.4722.6445.9635.2620.4222.7750.78
UniT3D[13]----46.6927.2221.9145.98----
3DJCG[4]64.7040.1727.6659.2349.4831.0324.2250.8038.0622.8223.7752.99
3D-VLP[34]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2Cap-DETR[10]71.4539.3428.2559.3361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA (Ours)74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
", + "bbox": [ + 80, + 142, + 890, + 299 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/12801e52c8dfca51bfc9fca8d252c20c675037c98d93905f6e1938bb23ff08f0.jpg", + "table_caption": [ + "Table 2. Quantitative Comparisons for 3D Question Answering on ScanQA[2]. We categorize previous works into classification based (\"CLS\") and generation based (\"GEN\") methods. The results from 3D-LLM* come from their fine-tuned version. LL3DA out-performs previous methods on the validation set and two test sets." + ], + "table_footnote": [], + "table_body": "
MethodAnswer TypeValidationTest w/ objectTest w/o object
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
ScanQA[2]64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[48]----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[18]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[34]66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM*[29]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA (Ours)76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
", + "bbox": [ + 142, + 366, + 828, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $|s|$ is the number of tokens in the desired response.", + "bbox": [ + 76, + 501, + 452, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 532, + 209, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To test the capacities of LL3DA, we provide numerous evaluations. To begin with, we introduce the datasets, metrics, and implementation details (Sec. 5.1). Then, we compare how our model understands and reasons in complex 3D environments with previous 3D specialists on 3D Dense Captioning and 3D Question Answering (Sec. 5.2). Additionally, we conduct quantitative ablation studies on the model design and training strategy (Sec. 5.3). Finally, Sec. 5.4 showcases several qualitative results.", + "bbox": [ + 75, + 559, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets, Metrics and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 705, + 465, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In this paper, we experiment with 3D data from ScanNet [16], a 3D dataset covering 1,201 and 312 diverse and complex indoor 3D scenes for training and validation. The language annotations used in this study are sourced from ScanRefer [7], Nr3D [1], ScanQA [2], and the ScanNet subset of 3D-LLM [29]. This combination covers a variety of tasks, including instance and scene descriptions, conversations, embodied planning and question answering. Please refer to the supplementary materials for more details on the statistics of data.", + "bbox": [ + 75, + 731, + 468, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. Here, we adopt C, B-4, M, R as abbreviations for", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CiDEr [53], BLEU-4 [47],METEOR [3],and Rouge-L [39] to evaluate the quality of the generated textual responses.", + "bbox": [ + 496, + 501, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following previous works on 3D vision language tasks [10, 12], we randomly sample $40\\mathrm{k}$ points from each 3D scene as the 3D input. We adopt the pre-trained OPT-1.3B [63] as our causal LLM backbone, which is frozen and loaded in float16 to save memory cost. We adopt the AdamW [41] optimizer with a weight decay of 0.1 and a learning rate decaying from $10^{-4}$ to $10^{-6}$ with a cosine annealing scheduler for about $100\\mathrm{k}$ iterations. For all the training tasks, we train with no more than eight Nvidia RTX3090 (24G) GPUs within a day.", + "bbox": [ + 496, + 532, + 892, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Comparison with SoTA Specialists", + "text_level": 1, + "bbox": [ + 498, + 694, + 802, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the model's capacity to understand and reason in 3D environments via 3D-DC and 3D-QA. For each evaluation task, we fine-tune the trainable parameters in our model on each task for $\\sim 30\\mathrm{k}$ iterations.", + "bbox": [ + 496, + 718, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D Dense Captioning demands a model to localize and describe any instance in a 3D scene. We benchmarks state-of-the-art methods on the widely-used ScanRefer [7] and $\\mathrm{Nr3D}$ [1] dataset in Tab. 1 under the $m@k\\mathrm{IoU}$ metric [12]. Here, $m \\in \\{\\mathrm{C}, \\mathrm{B-4}, \\mathrm{M}, \\mathrm{R}\\}$ , and the $m$ score of a caption is set to 0 if the IoU between the predicted box and the object is less than the given threshold $k$ . Following existing works [10, 12], we consider $\\mathrm{C} @ 0.25$ and $\\mathrm{C} @ 0.5$ as the", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "main metric for ScanRefer, and C@0.5 for Nr3D. Among the listed methods, UniT3D [13], 3DJCG [4], and 3D-VLP [34] are pre-trained on multiple 3D vision and language tasks annotated on ScanNet scenes. Additionally, UniT3D [13] adopts off-the-shelf image caption models [45] and multi-view images to generate additional instance-captions for pre-training. It is worth mentioning that we compare the results with the 3D-VisTA [70] model that is not trained on additional 3D scenes. To evaluate our model, we adopt the box predictions produced by Vote2Cap-DETR [10] as the visual prompt. Results show that our method consistently outperforms existing methods on both datasets. For example, our method achieves $65.19\\%$ C@0.5 on ScanRefer and $51.18\\%$ C@0.5 on Nr3D, which is $(+3.38\\%$ and $+7.34\\%)$ higher than the current state-of-the-art 3D vision and language model, Vote2Cap-DETR.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Question Answering requires a model to generate responses to the natural language queries questioning towards an 3D scene. We benchmark state-of-the-art methods on the ScanQA [2] validation set as well as two test benchmarks in Tab. 2, and consider CiDEr as the main metric. The majority of the listed methods are based on classification (marked \"CLS\"), i.e., selecting responses from a predefined answer set. Meanwhile, 3D-LLM [29] tries to address 3D-QA via auto-regressive text generation (marked \"GEN\"), and we list their fine-tuned version for comparison. Results show that our method consistently outperforms existing methods on all the evaluation sets, and surpasses the generation based method, 3D-LLM, by a large margin $(+7.39\\%$ CiDEr score on the validation set).", + "bbox": [ + 75, + 335, + 470, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 558, + 243, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we provide ablation studies on model designs and training strategies. We evaluate on ScanRefer and ScanQA to quantize the effectiveness.", + "bbox": [ + 75, + 580, + 468, + 628 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/516496b47dc1454d026b87181128be058b7e5996ba94be0f5b9ed19d9d9e7d73.jpg", + "image_caption": [ + "(a) Early Fusion (ours)" + ], + "image_footnote": [], + "bbox": [ + 84, + 643, + 269, + 724 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/79369773bc285f5d43fbf7e72cee7a3ecf91909680d3994ac0db3976bed889ee.jpg", + "image_caption": [ + "(b) Direct Injection", + "Figure 3. Two Different Ways to Encode Visual Prompts. Our proposed method (a) adopts a unified transformer to aggregate features from all kinds of interactions, while (b) directly concatenates the visual prompts to the scene embeddings. Tab. 3 shows that early fusion leads to a better performance." + ], + "image_footnote": [], + "bbox": [ + 274, + 643, + 464, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the Q-Former Design. We list two ways to process the visual prompts in Fig. 3. Here, Fig. 3 (a) is our proposed method that adopts a unified transformer to aggregate information from both text instructions and visual", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompts, while Fig. 3 (b) is the \"direct injection\" version, which only extract instruction-aware 3D feature with visual prompts concatenated after the scene embeddings. We train both models from scratch and evaluate their performance on ScanRefer 3D Dense Captioning. The results (Fig. 3) show that the method we use (Fig. 3 (a)) could better capture feature related to the visual prompts, leading to better instance caption generation performance $(+3.45\\% \\mathrm{C}@\\mathrm{0.5})$", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be5c6fd21511961e72ef1ec34d73d51ca6656450c969fedd76b19817c9961e4d.jpg", + "table_caption": [ + "Table 3. Effectiveness of Q-Former Design on ScanRefer[7]. We design two different ways of utilizing visual prompts. The \"early fusion\" enables direct interaction with the 3D scene, thus it achieves a better performance." + ], + "table_footnote": [], + "table_body": "
Visual PromptC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
direct59.3933.2725.1953.39
ours62.8435.8125.8154.45
", + "bbox": [ + 504, + 290, + 888, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instructions as Auxiliary Tasks for 3D Dense Captioning. We have introduced two types of task instructions in Sec. 4.1 for 3D-DC, i.e. the \"describe\"-only instructions and \"detect and localize\" instructions. Additionally, we have introduced two types of visual prompts (Fig. 2 & Sec. 4.2). In this study, we show how they affect the performance when serving as auxiliary tasks for 3D-DC by evaluating on ScanRefer in Tab. 4. All the methods listed are trained from scratch. In Tab. 4, \"Aux.Loc\" identifies whether we train the model with the \"detect and localize\" instructions, and \"Clicks\" identifies whether we train the model with clicks as additional visual prompts. Results show that they are both good auxiliary tasks for 3D-DC.", + "bbox": [ + 496, + 356, + 890, + 551 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4c36c491107a0d1d24b74a3ea5c93ee542462cf0843135c3d80d6be372a5ab9c.jpg", + "table_caption": [ + "Table 4. Effectiveness of Instructions as 3D Dense Captioning Auxiliary Tasks. We train the models from scratch and evaluate on ScanRefer[7]. \"Aux.Loc\" identifies whether we train with the \"describe and localize\" instructions. \"Clicks\" identifies whether we train with clicks as additional visual prompts." + ], + "table_footnote": [], + "table_body": "
Aux.LocClicksC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
--60.8534.0925.5353.48
✓-61.8134.1525.4953.83
-✓62.2034.2625.6753.87
✓✓62.8435.8125.8154.45
", + "bbox": [ + 504, + 643, + 888, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instructions as Auxiliary Tasks for 3D Question Answering. We have made a similar study to analyze how adopting additional \"answer and localize\" instructions and visual prompts improves 3D-QA on ScanQA [2] as auxiliary tasks in Tab. 6. We do not use any visual interactions during inference. Results show that the additional textual instructions and visual prompts improve the task diversity and further improve the performance for 3D-QA.", + "bbox": [ + 496, + 733, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance as a Generalist. To test whether LL3DA can distinguish different tasks given the textual instructions and visual prompts, we evaluate our model on different tasks", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26433", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa79cca0f99d8a6eaacffb258ea8069e445ce401b9e8209546fd0d03e47293bb.jpg", + "table_caption": [ + "Table 5. Evaluation as a Generalist. The first three rows list the performance of models trained from scratch as experts on each dataset. The results in the following three rows belong to the model fine-tuned from the generalist weights. The last row evaluates the model trained as a generalist. ScanRefer[7] and Nr3D[1] are used to evaluate the dense captioning performance, and ScanQA[2] is used to evaluate the question answering performance. Serving as a generalist, our method can differentiate each task, and produce strong results based on textual instructions and visual prompts." + ], + "table_footnote": [], + "table_body": "
MethodScanReferNr3DScanQA
C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C↑B-4↑M↑R↑
ScanRefer(scratch)62.8435.8125.8154.45--------
Nr3D(scratch)----44.9527.6725.6755.79----
ScanQA(scratch)--------74.8013.6815.4036.25
ScanRefer(fine-tuned)65.1936.7925.9755.06--------
Nr3D(fine-tuned)----51.1828.7525.9156.61----
ScanQA(fine-tuned)--------76.7913.5315.8837.31
w/o fine-tuning62.9835.9725.6654.6523.9413.3722.3145.7875.6713.3315.3737.02
", + "bbox": [ + 80, + 169, + 890, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c8ec3ac01dd213fac16b6073b74d96fc60c9057cd7480f854908fc394b20b41e.jpg", + "table_caption": [ + "Table 6. Effectiveness of Interactions as 3D Question Answering Auxiliary Tasks. We train the model from scratch and evaluate all the models from scratch on ScanQA[2] validation set. \"Aux.Loc\" identifies whether we train with the \"answer and localize\" instructions, and \"Visual Prompts\" identifies whether we train with visual prompts." + ], + "table_footnote": [], + "table_body": "
Aux.LocVisual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
--67.8511.8713.9633.87
✓-72.7313.2714.9035.87
-✓68.0912.5914.2033.71
✓✓74.8013.6815.4036.25
", + "bbox": [ + 80, + 400, + 467, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "without task-specific fine-tuning in Tab. 5. The first three rows list the performance of LL3DA when trained from scratch on one specific task, while the following three rows represent the fine-tuned models. The last row indicates the direct evaluation of LL3DA. Results show that our model could distinguish 3D-DC and 3D-QA given the text instructions and visual prompts, and achieve strong performance (62.98% C@0.5 on ScanRefer, 75.67% CiDEr on ScanQA). However, the generalist model achieves poor performance on Nr3D [1], which is because we did not try to differentiate between Nr3D and ScanRefer during training in the first place, as ScanRefer and Nr3D are used for the same task. There is also an interesting observation that though we did not differentiate between these two datasets for 3D-DC, the model still tend to achieve high scores on ScanRefer (62.98% C@0.5). We are also excited to see that the weights of the generalist model can serve as a strong initialization for fine-tuning. For example, the fine-tuned model on ScanRefer could achieve 65.19% C@0.5, which is +2.35% higher than the model trained from scratch.", + "bbox": [ + 75, + 474, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Importance of Textual Instructions. We further conduct study to see whether the text instructions are necessary for 3D-DC in Tab. 7. The first row is our baseline method that directly generates the captions based on visual prompts without any text instructions, and the second row is our method that is trained with the text instructions introduced in Sec. 4.1. Both methods are trained from scratch for fair comparison. We notice that since the LLM is frozen, certain", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "textual instructions are beneficial when generating results in specific domains/tasks.", + "bbox": [ + 496, + 306, + 890, + 338 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ce6505aa640daa425b8491cb470a5bac4afaca2bf05c11ec82caa6e0877e00d.jpg", + "table_caption": [ + "Table 7. Effectiveness of Instructions on 3D Dense Captioning. We perform experiments on ScanRefer[7]. The baseline method directly generates the captions given the input 3D scene and visual prompts without any textual instructions." + ], + "table_footnote": [], + "table_body": "
InstructionsC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
-60.2034.7925.4054.03
✓62.8435.8125.8154.45
", + "bbox": [ + 504, + 417, + 888, + 469 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clicks for Better Question Answering. One major challenge of answering questions in complex 3D environments is the vague identification of objects with plain texts. Therefore, we try to click on some of the related objects along with the textual instructions during evaluation, and see how it could affect the generated answers on the ScanQA validation set in Tab. 8. Results show that this technique would remove the ambiguities, and further improve the quality of the answers $(+6.12\\%$ C). This illustrates the importance of visual interaction in complex 3D environments.", + "bbox": [ + 496, + 484, + 890, + 636 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0b1798fd69dfe7f55b5679c4c55ac89fa7ef747ebea2dc25184eab9dbad2a556.jpg", + "table_caption": [ + "Table 8. Test Time Visual Interactions for Question Answering on ScanQA[2]. The model achieves better performance on the question answering when we add visual prompts to some of the related objects along with the text instructions during evaluation." + ], + "table_footnote": [], + "table_body": "
Visual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
-76.7913.5315.8837.31
✓82.9111.8016.7439.97
", + "bbox": [ + 504, + 715, + 888, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 498, + 786, + 683, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present several visualization results on different tasks in Fig. 4 to show our model's capacities in understanding, reasoning, and planning in different 3D environments. To prevent repetition when generating long sequences, we combine the top- $k$ [22] and top- $p$ [27] sampling strategy with $k = 50$ and $p = 0.95$ .", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26434", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Dense Captioning", + "text_level": 1, + "bbox": [ + 245, + 95, + 395, + 112 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/774a3fe5adcf1de1ef3e1b51cf62afdca7d7fd4c7522ada6b691962a1a12cf96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 116, + 276, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/503c80453692779fabc10ce921f8b4fa6de1850c1b656409b9c8c2091bcbec18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 116, + 313, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Describe this object in the 3D scene.", + "bbox": [ + 321, + 116, + 491, + 138 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This is a brown armchair. It is to the right of the couch.", + "bbox": [ + 316, + 143, + 509, + 165 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f88f3c5b8e779ecbd74d81be36c5088ac2b98e5c2939ccff3d66007b1e628564.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 143, + 553, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SpaCap3D: This is a chair with no arms. \nIt is next to a desk.", + "bbox": [ + 279, + 176, + 524, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vote2Cap-DETR: This is a brown armchair. \nIt is to the right of the table.", + "bbox": [ + 279, + 200, + 535, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ground Truth: The armchair is facing the desk. The armchair is located to the right of the couch, in the middle of the room.", + "bbox": [ + 279, + 224, + 537, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Question Answering", + "text_level": 1, + "bbox": [ + 236, + 277, + 401, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/128a305b4cf2436f9018718d41e48b6ab6445973196e45310f870c2a2fc447b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 290, + 285, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How many chairs are behind the long table?", + "bbox": [ + 321, + 309, + 483, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "2 ch + Visual Prompts", + "bbox": [ + 359, + 335, + 475, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ed03e0f33b4960b11c16d41715dd077eda723133fe489a36d6921b6963e8d05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 381, + 287, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How many chairs are behind the long table? [Clicks]", + "bbox": [ + 321, + 398, + 483, + 420 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3 chairs are behind long table", + "bbox": [ + 357, + 429, + 508, + 450 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/acec21fe01ee05cb9894c68945854c163eafe73e4b57b82637be045b863290b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 330, + 552, + 352 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Embodied Planning", + "text_level": 1, + "bbox": [ + 645, + 95, + 805, + 111 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39890bb25308a1c43a3d059048182459ae626122deb627f022f787ec3001e298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 117, + 875, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I want to clean the kitchen. What should I do?", + "bbox": [ + 606, + 252, + 810, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1. pick up any trash off the floor. 2. Open the kitchen shelf. 3. Clean the kitchen sink with a cleaning solution. 4. Clean the kitchen sink with a paper towel.", + "bbox": [ + 604, + 285, + 841, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e53409ca803248459a5c65b178fa6305e5d8dfe40bda1c8906ab5ec83c6895f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 342, + 596, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I want to clean the kitchen. I have done these things: 1. Turn on the lamp on the table. 2. Open the kitchen cabinet and take out cleaning supplies. 3. Clean the countertops with a cloth. 4. Clean the stove and microwave. What should I do next?", + "bbox": [ + 606, + 349, + 854, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clean the refrigerator and cabinets", + "bbox": [ + 617, + 435, + 836, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c4125d18ffa1d6187efc09522e9fa038b5f2eb398a348d4b213e9c23fc2ee63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 851, + 318, + 880, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3796156bba50e55762b755a1d55c3160e3e6861b27928eb8dafa34d0bc559a4e.jpg", + "image_caption": [ + "Figure 4. Qualitative Results. We provide several visualization results on various 3D vision and language tasks in diverse 3D environments (living room, classroom, kitchen, and bedroom). Red highlights the wrong answer." + ], + "image_footnote": [], + "bbox": [ + 91, + 481, + 338, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D Scene Description", + "text_level": 1, + "bbox": [ + 531, + 484, + 681, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Describe this 3D scene.", + "bbox": [ + 380, + 510, + 519, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This room is a kitchen with a stove, stovetop, cabinets, refrigerator, oven, cooler, recycling bins, microwave, water heater, soap dish, trash can, a sink, and a refrigerator. There are multiple walls scattered in the room, creating a separate space for dining and cooking. The door leads to the kitchen. Another door and several windows hang around the area, providing a large view of the floor. As well as additional doors, walls, windows, doorstopper caps, kitchen items, sink and fridge, the food preparation and storage areas are visible.", + "bbox": [ + 374, + 531, + 841, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 681, + 202, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present LL3DA, a large language 3D assistant that could take both textual- and visual- interactions from human for understanding, reasoning, and planning in complex 3D environments. Our model directly encodes the 3D point cloud and aggregates information from scenes and human interactions with the attention mechanism. We show that the visual interactions could remove the ambiguities in cluttered 3D environments, showing mighty instruction-following capacities. Experiments show that our method could achieve remarkable results on various 3D vision-language benchmarks. We hope our approach could inspire further designs and training strategies for large 3D language models. In future studies, we believe that the construction", + "bbox": [ + 75, + 703, + 470, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of high-quality and diverse annotations will further enhance the model's reasoning and planning capabilities.", + "bbox": [ + 496, + 681, + 890, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 715, + 687, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (No. 62071127 and 62101137), National Key Research and Development Program of China (No. 2022ZD0160100), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No. 2021SHZDZX0103), A*STAR AME Programmatic Funding A18A2b0046, RobotHTPO Seed Fund under Project C211518008, and EDB Space Technology Development Grant under Project S22-19016-STDP. The computations in this research were performed using the CFFF platform of Fudan University.", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26435", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 2, 5, 7", + "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 4, 5, 6, 7", + "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 5", + "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 5, 6", + "[5] Yang Cao, Yihan Zeng, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, 2023. 2", + "[6] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 2", + "[7] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2, 5, 6, 7", + "[8] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551, 2021. 5", + "[9] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1", + "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11124-11133, 2023. 2, 3, 4, 5, 6", + "[11] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. arXiv preprint arXiv:2309.02999, 2023. 2", + "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2, 5" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2, 5, 6", + "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 1", + "[15] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2", + "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5", + "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 4", + "[18] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 2, 5", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4", + "[20] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2", + "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2", + "[22] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833, 2018.7", + "[23] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 1", + "[24] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26436", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 2", + "[26] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2", + "[27] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 7", + "[28] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9202-9212, 2023. 2", + "[29] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981, 2023. 2, 5, 6", + "[30] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1", + "[31] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 1", + "[32] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 1", + "[33] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203, 2022. 5", + "[34] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 5, 6", + "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 4", + "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 4", + "[37] Mingsheng Li, Xin Chen, Chi Zhang, Sijin Chen, Hongyuan Zhu, Fukun Yin, Gang Yu, and Tao Chen. M3dbech: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763, 2023. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, and Yunchao Wei. Stablellava: Enhanced visual instruction tuning with synthesized image-dialogue data. arXiv preprint arXiv:2308.10253, 2023. 2", + "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5", + "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2", + "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3", + "[43] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2", + "[44] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics, 2023. 5", + "[45] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6", + "[46] OpenAI. Gpt-4 technical report, 2023. 1", + "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5", + "[48] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5606-5611, 2023. 2, 4, 5", + "[49] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 2", + "[50] Zhende Song, Chenchen Wang, Jiamu Sheng, Chi Zhang, Gang Yu, Jiayuan Fan, and Tao Chen. Moviellm: Enhancing long video understanding with ai-generated movies. arXiv preprint arXiv:2403.01422, 2024. 1", + "[51] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4", + "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26437", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1,3,4", + "[53] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5", + "[54] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 2, 5", + "[55] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2", + "[56] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[57] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2, 3", + "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2", + "[59] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 2", + "[60] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao Chen. Shapept: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618, 2023. 1", + "[61] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2", + "[62] Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 1", + "[63] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 2, 3, 4, 5", + "[64] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2", + "[65] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 2", + "[66] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Towards explainable 3d grounded visual question answering:" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A new benchmark and strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2", + "[67] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 5", + "[68] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension, 2023. 3", + "[69] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2", + "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26438", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_model.json b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c448c7a23c9e5636d44295d379e66ff2ac2724e5 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.131, + 0.745, + 0.177 + ], + "angle": 0, + "content": "LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding, Reasoning, and Planning" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.203, + 0.773, + 0.222 + ], + "angle": 0, + "content": "Sijin Chen\\(^{1}\\) Xin Chen\\(^{2,*}\\) Chi Zhang\\(^{2}\\) Mingsheng Li\\(^{1}\\) Gang YU\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.222, + 0.717, + 0.239 + ], + "angle": 0, + "content": "Hao Fei\\(^{3}\\) Hongyuan Zhu\\(^{4}\\) Jiayuan Fan\\(^{1}\\) Tao Chen\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.24, + 0.732, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Fudan University \\(^{2}\\)Tencent PCG \\(^{3}\\)National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.258, + 0.83, + 0.275 + ], + "angle": 0, + "content": "4Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.278, + 0.629, + 0.292 + ], + "angle": 0, + "content": "https://github.com/Open3DA/LL3DA" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.295, + 0.605, + 0.31 + ], + "angle": 0, + "content": "* project lead † corresponding author" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.344, + 0.887, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.608, + 0.893, + 0.651 + ], + "angle": 0, + "content": "Figure 1. We propose LL3DA, a Large Language 3D Assistant that demonstrates mighty instruction-following capacities in understanding, reasoning, and planning within complex 3D environments. LL3DA takes both the textual instructions and potential visual interactions into consideration to help remove ambiguities when addressing various tasks in diverse and complex 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.663, + 0.314, + 0.679 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Recent progress in Large Multimodal Models (LMM) has opened up great possibilities for various applications in the field of human-machine interactions. However, developing LMMs that can comprehend, reason, and plan in complex and diverse 3D environments remains a challenging topic, especially considering the demand for understanding permutation-invariant point cloud representations of the 3D scene. Existing works seek help from multi-view images by projecting 2D features to 3D space, which inevitably leads to huge computational overhead and performance degradation. In this paper, we present LL3DA, a Large Language 3D Assistant that takes point cloud as the direct input and responds to both text instructions and visual interactions. The additional visual interaction enables LMMs to better" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.895, + 0.741 + ], + "angle": 0, + "content": "comprehend human interactions with the 3D environment and further remove the ambiguities within plain texts. Experiments show that LL3DA achieves remarkable results and surpasses various 3D vision-language models on both 3D Dense Captioning and 3D Question Answering." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.755, + 0.633, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The recent surge in Large Language Model (LLM) families [14, 31, 46, 52, 63] opens up great opportunities for addressing various machine learning tasks in a generalized way [30, 32, 40, 60]. During this LLM carnival, researchers are also seeking generalized LLM solutions for various vision language tasks [37, 50, 62]. Among these, LLM-based 3D scene understanding is a valuable topic that would benefit the development of autonomous driving [9, 23] and em" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26428" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "bodied AI agents [21, 49]. However, it is also challenging given 1) the diversity and complication of 3D environments and 2) the demands for understanding sparse 3D points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.471, + 0.379 + ], + "angle": 0, + "content": "Prior works have made initial success in various 3D vision and language tasks. The majority of these research build 3D specialists to solve one specific down-stream task, including 3D Question Answering (3D-QA) [2, 43], 3D Visual Grounding (3D-VG) [7, 26, 55], and 3D Dense Captioning (3D-DC) [10-12]. There are also several works [4, 13, 34, 70] study the mutual promotion of different 3D vision and language tasks with shared structure modelling relations among objects. Recently, researchers have also introduced LLMs for general purpose 3D understanding, where Point-Bind and Point-LLMs [24, 57] mainly focus on the understanding of 3D objects. Concurrently, 3D-LLM [29] proposes an LLM-driven solution that aggregates multi-view features for 3D perception, presenting mighty capacities in understanding 3D object and scenes and following text instructions produced by human." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.531 + ], + "angle": 0, + "content": "Though these methods have achieved remarkable success in addressing various challenges in understanding the 3D world with natural language, there are certain limitations. With limited supervision, 3D specialists could hardly scale-up for better performance, while the joint pre-training still requires separate heads for specific tasks. Extracting multi-view features results in huge computational overhead and ignores the essential geometry and depth information. Additionally, plain texts often lead to ambiguities especially in cluttered and complex 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.744 + ], + "angle": 0, + "content": "To address the above issues, we propose LL3DA, a Large Language 3D Assistant that responds to both textual and visual interactions from human, with the ability to understand, reason, and plan in complex 3D environments (Fig. 1). We adopt a multi-modal transformer to aggregate information from textual instructions, visual prompts, and the 3D scene into a fixed length of learnable querying tokens via the attention mechanism. The querying tokens are then projected and used as the prefix for the textual instructions, serving as the input to a pre-trained and frozen LLM. This design not only helps to address the contradiction between the permutation-invariant 3D scene representations and the LLM embedding space, but also extracts interaction-aware 3D scene representations for efficient instruction following." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.47, + 0.865 + ], + "angle": 0, + "content": "We conduct extensive experiments to explore the capacities of LL3DA in understanding, reasoning, and planning within complex and diverse 3D environments. Our model achieves state-of-the-art results on two widely used datasets for 3D Dense Captioning [1, 7], and 3D Question Answering [2]. Additionally, by introducing additional visual interactions, our method could further remove the ambiguities within the vague textual instructions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.867, + 0.383, + 0.881 + ], + "angle": 0, + "content": "To summarize, our key contributions lie in:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "- We present a LLM-based solution for understanding, rea" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.852, + 0.106 + ], + "angle": 0, + "content": "soning, and planning in complex 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.151 + ], + "angle": 0, + "content": "- Our model takes both the textual instructions and visual interactions as inputs, and extracts interaction-aware features for effective instruction-following." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.182 + ], + "angle": 0, + "content": "- Extensive experiments show that our method surpasses various state-of-the-art 3D vision language models." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.107, + 0.892, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.194, + 0.642, + 0.21 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.895, + 0.748 + ], + "angle": 0, + "content": "3D Vision and Language alignment, pre-training, and understanding [5, 7, 20, 70] cover tasks requiring a model to adopt its understanding towards a complex 3D scene answering to, or answering with natural language. Among those, 3D Dense Captioning (3D-DC) [10, 12, 54] expects a model to translate an input 3D scene into a set of instance coordinates and natural language descriptions. Existing methods could be categorized into \"detect-then-describe\" models [4, 12, 54] and the \"set-to-set\" prediction approaches [10, 11]. The former builds explicit relations on the instance coordinate estimations, while the latter directly learns the locations and descriptions for instances from the input 3D scene. 3D Visual Grounding (3D-VG) [1, 7, 55] demands a model to respond the natural language queries with the instance coordinates in the 3D scene. The mainstream of existing methods [4, 65, 70] address 3D-VG via selecting a candidate from a 3D detector's prediction. 3D Question Answering (3D-QA) [2, 43, 59, 66] requires a model to answer the questions with natural language based on the input 3D scene. The majority of existing methods [2, 18, 48] directly select the desired response from a given answer set. Researchers have also studied the mutual promotion of various 3D vision language tasks via training their shareable architectures simultaneously on multiple tasks [4, 13, 34, 70]. UniT3D [13] and 3DJCG [4] focus on the joint promotion between 3D-DC and 3D-VG in the relation modelling, while 3D-VLP [34] further includes 3D-QA. Recently, 3D-LLM [29] introduces a family of LLM-driven 3D generalists that could handle diverse textual instructions with reconstructed 3D features from multi-view images [28]. In this paper, we present LL3DA, an LLM solution that directly extracts features from the 3D scene, and handles both visual prompts and textual instructions to diversify the possible interactions human could make with the complex 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Large Multimodal Models (LMM). Along with the rapid development of Large Language Models (LLM) [15, 63], researchers have made great recent efforts adapting LLMs to visual understanding and reasoning tasks [25, 38, 56, 61]. Some methods project or compress global image features as the prefix for text instructions [36, 40, 58, 69], while others extract ROI features as LLM tokens for region-oriented instruction reasoning [6, 64]. Meanwhile, InstructBLIP [17] proposes to extract textual instruction-aware visual features, and has achieved remarkable success in addressing complex" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26429" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.887, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Overview of the Proposed Approach. (a) The overall pipeline of our proposed LL3DA first extracts interaction-aware 3D scene embeddings, which are later projected to the prefix of textual instructions as the input of a frozen LLM. (b) The detailed design of the Interactor3D, which aggregates visual prompts, textual instructions, and 3D scene embeddings into a fixed length querying tokens. (c) The prompt encoder encodes the user clicks and box coordinates with the positional embeddings and ROI features, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.473, + 0.538 + ], + "angle": 0, + "content": "and unseen instructions. Concurrently, researchers have also made great attempts solving various 3D tasks using LLMs. Notably, [24, 42, 57, 68] demonstrate remarkable success in understanding and reasoning about 3D objects. In this paper, we present an LLM-driven solution that could handle both interactions in forms of visual prompts and text instructions. We also propose to extract interaction-aware 3D scene representations for better instruction following." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.214, + 0.569 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.669 + ], + "angle": 0, + "content": "To build a general purpose agent that could handle both visual and textual interactions with complex 3D environments, we propose LL3DA, an LLM driven auto-regressive approach to 3D vision language tasks. In this section, we first introduce the problem formatting in Sec. 3.1. Next, we introduce our model designs in details (Sec. 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.273, + 0.695 + ], + "angle": 0, + "content": "3.1. Problem Formatting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Model I/O. As shown in Fig. 2 (a), the input of our model consists of a 3D scene represented by a set of points \\(PC\\), the textual instruction \\(\\mathcal{I}_t\\), and potential visual interactions \\(\\mathcal{I}_v\\) that serve as supplementary spatial identifiers. Here, point cloud \\(PC = [p_{in}, f_{in}] \\in \\mathbb{R}^{N \\times (3 + F)}\\), where \\(p_{in} \\in \\mathbb{R}^{N \\times 3}\\) and \\(f_{in} \\in \\mathbb{R}^{N \\times F}\\) are the point coordinates and the additional point features, including color, normal, and height. The output of our model is free-form natural language, part of whom could be interpreted into 3D coordinates." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Instruction Formatting. Following existing LMMs [57], we begin the textual instructions \\(\\mathcal{I}_t\\) with the “human:” identifier, and ask the model to generate responses after the “assistant:” identifier. This endows the model with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.445 + ], + "angle": 0, + "content": "ability to distinguish information from the context and further engage in multi-turn conversations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.628 + ], + "angle": 0, + "content": "Coordinate Representations. To provide LLMs with the capacity to perceive and respond with 3D coordinates, we convert the 3D points and 3D bounding boxes to plain texts. Specifically, a point is represented by “\\(\\langle \\mathrm{loc} \\rangle x, y, z \\langle \\mathrm{loc} \\rangle\\)”, and a bounding box is represented by its center point and size, i.e. “\\(\\langle \\mathrm{obj} \\rangle c_x, c_y, c_z, w, h, l \\langle \\mathrm{obj} \\rangle\\)”. Here, all the numerical data is discretized into unsigned integers within a range of [0, 255] with respect to the boundary of the input 3D scene. This design could naturally fit in the vocabulary of existing pre-trained LLMs [52, 63]. Without the introduction of any additional learnable tokens, we could save the effort of tuning the LLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.635, + 0.644, + 0.652 + ], + "angle": 0, + "content": "3.2. Model Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.794 + ], + "angle": 0, + "content": "As shown in Fig. 2 (a), our model first aggregates a fixed-length scene embeddings through the Interactor3D, which takes the visual prompts, the textual instructions, and the 3D scene as the input. Next, the aggregated 3D scene embeddings are projected to be the prefix of text instructions, and serve as the inputs of a frozen LLM. The detailed design of Interactor3D is shown in Fig. 2 (b), which consists of a frozen 3D scene encoder \\(\\mathcal{E}^{3D}\\), a visual prompt encoder, and a multi-modal transformer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Scene Encoder. We adopt the masked transformer encoder pre-trained on ScanNet detection [10] as the scene encoder, \\(\\mathcal{E}^{3D}\\). The scene encoder takes \\(PC\\) as its input, and outputs the 3D scene embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.861, + 0.892, + 0.881 + ], + "angle": 0, + "content": "\\[\nf _ {e n c} = \\mathcal {E} ^ {3 D} (P C) = \\mathcal {E} ^ {3 D} \\left(p _ {i n}; f _ {i n}\\right) \\in \\mathbb {R} ^ {M \\times d}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Here, \\( f_{enc} \\) consists of \\( d \\)-dimensioned features for \\( M \\) points" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26430" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "uniformly down-sampled from the input 3D scene through the Farthest Point Sampling (FPS) algorithm. In practice, we choose to keep the scene encoder frozen to save the memory cost during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.244 + ], + "angle": 0, + "content": "Visual Prompt Encoder. We mainly take two common types of visual interactions into consideration, user clicks and 3D box annotations [35]. Each user click is first normalized within a range of \\([0,1]\\) by the size of the input 3D scene \\(p_{\\mathrm{click}} \\in \\mathbb{R}^3\\). Then, we encode \\(p_{\\mathrm{click}}\\) with the 3D Fourier positional embeddings [51] function:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.256, + 0.469, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right) = \\left[ \\sin \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right); \\cos \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right) \\right]. \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.469, + 0.359 + ], + "angle": 0, + "content": "Here, \\(B \\in \\mathbb{R}^{3 \\times (d/2)}\\) is a learnable matrix. The box annotation is represented by the ROI feature \\(f_{\\mathrm{box}} \\in \\mathbb{R}^d\\) extracted by a pre-trained 3D object detector [10]. The two types of the visual prompts are then projected with separate and identical Feed Forward Networks (FFN)." + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.37, + 0.469, + 0.395 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {c l i c k}} = F F N _ {\\text {c l i c k}} \\left(\\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.39, + 0.326, + 0.406 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {b o x}} = F F N _ {\\text {b o x}} \\left(f _ {\\text {b o x}}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.468, + 0.432 + ], + "angle": 0, + "content": "In practice, we represent each visual prompt with 8 tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.72 + ], + "angle": 0, + "content": "Multi-Modal Transformer (MMT) serves as a role to 1) address the contradiction between the permutation-invariant 3D scene embeddings and position-sensitive causal LLMs, 2) bridge the gap between frozen unimodal experts, and 3) fill the needs for interaction-aware feature extraction. Inspired by the Q-Former architecture [17, 36], MMT aggregates the visual information within a fixed number of 32 learnable querying tokens. In each layer, the queries interact with the encoded visual prompts \\([f_{\\mathrm{click}}; f_{\\mathrm{box}}]\\) and the textual instructions \\(\\mathcal{I}_t\\) through a shared self-attention. Then, we allow the learnable querying tokens and the visual prompts to interact with the task-agnostic 3D scene embeddings \\(f_{enc}\\) via cross-attention. The output of MMT is 32 queries written as \\(Q \\in \\mathbb{R}^{32 \\times 768}\\), which are finally projected to the word embedding space of LLMs through a simple linear projector. In practice, we notice that initializing Q-Former with pre-trained BERT [19, 36] weights will lead to repetitive outputs, thus we only choose to initialize the pre-trained word and position embeddings from BERT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.722, + 0.469, + 0.842 + ], + "angle": 0, + "content": "LLM. We consider the decoder-only generative pre-trained transformers [52, 63] as our large language model backbone. The decoder-only LLMs are sensitive to the input orders because of the position embeddings and the causal attention mask. The parameters and the embedding layers of the LLMs are kept frozen to save memory cost. During inference, we generate the responses via searching for the optimal sequence \\( s^* \\) that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.854, + 0.469, + 0.876 + ], + "angle": 0, + "content": "\\[\ns ^ {*} = \\arg \\max _ {s} P (s | P C, \\mathcal {I} _ {t}, \\mathcal {I} _ {v}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.436, + 0.901 + ], + "angle": 0, + "content": "In practice, we use beam search with a beam size of 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.795, + 0.108 + ], + "angle": 0, + "content": "4. Multi-modal Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.892, + 0.206 + ], + "angle": 0, + "content": "A general purpose 3D agent is able to address various tasks simultaneously in complex 3D scenes. Apart from introducing proper training data, it is important to guide the model to generate the desired outputs with instructions. Therefore, Sec. 4.1 will first introduce how we identify each task. After that, Sec. 4.2 will present details for the training objective." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.215, + 0.715, + 0.23 + ], + "angle": 0, + "content": "4.1. Tasks and Instructions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.268 + ], + "angle": 0, + "content": "As introduced in Sec. 3.1, LL3DA generates text responses auto-regressively after the “```assistant:” identifier." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.27, + 0.892, + 0.376 + ], + "angle": 0, + "content": "3D Dense Captioning requires the localization and description of instances in diverse 3D environments. We adopt either user clicks and box annotations as the visual prompt to identify the object to be described. Additionally, we design two types of textual instructions that ask the model to either \"describe\" or \"describe and localize\" the object, which diversifies the tasks, and leads to better performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.513 + ], + "angle": 0, + "content": "3D Question Answering requires the model to generate response to the questions based on the global knowledge of a 3D scene. To help the model better understand the 3D environment, we also design two types of textual instructions that ask the model to either \"answer\" or \"answer and localize the related objects\". The latter serves as an auxiliary task widely adopted in various 3D-QA methods [2, 48]. To diversify the tasks during training, we randomly include additional clicks on the objects related to the questions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.559 + ], + "angle": 0, + "content": "Scene Description requires the model to translate its global knowledge of the 3D scene into natural language descriptions, thus we simply ask the \"describe\" this 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.56, + 0.892, + 0.62 + ], + "angle": 0, + "content": "Embodied Conversation and Planning could be treated as multi-turn conversations, where we use “human” and “assistant” as identifiers to distinguish the source of information as introduced in Sec. 3.1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.762, + 0.647 + ], + "angle": 0, + "content": "4.2. Instruction Following Tuning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.654, + 0.892, + 0.699 + ], + "angle": 0, + "content": "During training, for tasks requiring additional visual interactions, i.e. 3D-DC and 3D-QA, we randomly choose between clicks or boxes as means of object identification." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Training Objective. Our training objective is to optimize the trainable parameters \\(\\theta\\), so as to maximize the likelihood of the target response sequence \\(s\\) given the input point cloud \\(PC\\), and the human interactions \\(\\mathcal{I}_v\\) and \\(\\mathcal{I}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.771, + 0.891, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\arg \\max _ {\\theta} P (s | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.804, + 0.892, + 0.851 + ], + "angle": 0, + "content": "In practice, this is accomplished by adopting the token-wise cross-entropy loss that trains the model to predict the \\(i\\)th token \\(s_{[i]}\\) given the previous \\((i - 1)\\) tokens, \\(s_{[1,\\dots ,i - 1]}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.862, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = - \\sum_ {i = 1} ^ {| s |} \\log P (s _ {[ i ]} | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta ; s _ {[ 1, \\dots , i - 1 ]}). \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26431" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons for 3D Dense Captioning on ScanRefer[7] and Nr3D[1]. For fair comparison, we list methods that are trained under the standard per-word cross-entropy loss without additional 3D scenes. We use the box estimations from Vote2Cap-DETR to simulate the box annotations as the visual prompts. Our proposed LL3DA surpasses previous 3D specialists on both datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.143, + 0.892, + 0.3 + ], + "angle": 0, + "content": "
MethodScanReferNr3D
C@0.25↑B-4@0.25↑M@0.25↑R@0.25↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[33]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[54]----44.0225.2622.3345.3633.7119.9222.6150.50
REMAN[44]62.0136.3726.7656.2545.0026.3122.6746.9634.8120.3723.0150.99
D3Net[8]----46.0730.2924.3551.6733.8520.7023.1353.38
Contextual[67]----46.1125.4722.6445.9635.2620.4222.7750.78
UniT3D[13]----46.6927.2221.9145.98----
3DJCG[4]64.7040.1727.6659.2349.4831.0324.2250.8038.0622.8223.7752.99
3D-VLP[34]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2Cap-DETR[10]71.4539.3428.2559.3361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA (Ours)74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.313, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Table 2. Quantitative Comparisons for 3D Question Answering on ScanQA[2]. We categorize previous works into classification based (\"CLS\") and generation based (\"GEN\") methods. The results from 3D-LLM* come from their fine-tuned version. LL3DA out-performs previous methods on the validation set and two test sets." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.367, + 0.83, + 0.478 + ], + "angle": 0, + "content": "
MethodAnswer TypeValidationTest w/ objectTest w/o object
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
ScanQA[2]64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[48]----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[18]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[34]66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM*[29]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA (Ours)76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.502, + 0.454, + 0.519 + ], + "angle": 0, + "content": "Here, \\( |s| \\) is the number of tokens in the desired response." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.21, + 0.551 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.47, + 0.696 + ], + "angle": 0, + "content": "To test the capacities of LL3DA, we provide numerous evaluations. To begin with, we introduce the datasets, metrics, and implementation details (Sec. 5.1). Then, we compare how our model understands and reasons in complex 3D environments with previous 3D specialists on 3D Dense Captioning and 3D Question Answering (Sec. 5.2). Additionally, we conduct quantitative ablation studies on the model design and training strategy (Sec. 5.3). Finally, Sec. 5.4 showcases several qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.466, + 0.723 + ], + "angle": 0, + "content": "5.1. Datasets, Metrics and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.883 + ], + "angle": 0, + "content": "Datasets. In this paper, we experiment with 3D data from ScanNet [16], a 3D dataset covering 1,201 and 312 diverse and complex indoor 3D scenes for training and validation. The language annotations used in this study are sourced from ScanRefer [7], Nr3D [1], ScanQA [2], and the ScanNet subset of 3D-LLM [29]. This combination covers a variety of tasks, including instance and scene descriptions, conversations, embodied planning and question answering. Please refer to the supplementary materials for more details on the statistics of data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Metrics. Here, we adopt C, B-4, M, R as abbreviations for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.533 + ], + "angle": 0, + "content": "CiDEr [53], BLEU-4 [47],METEOR [3],and Rouge-L [39] to evaluate the quality of the generated textual responses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.893, + 0.685 + ], + "angle": 0, + "content": "Implementation Details. Following previous works on 3D vision language tasks [10, 12], we randomly sample \\(40\\mathrm{k}\\) points from each 3D scene as the 3D input. We adopt the pre-trained OPT-1.3B [63] as our causal LLM backbone, which is frozen and loaded in float16 to save memory cost. We adopt the AdamW [41] optimizer with a weight decay of 0.1 and a learning rate decaying from \\(10^{-4}\\) to \\(10^{-6}\\) with a cosine annealing scheduler for about \\(100\\mathrm{k}\\) iterations. For all the training tasks, we train with no more than eight Nvidia RTX3090 (24G) GPUs within a day." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.695, + 0.803, + 0.712 + ], + "angle": 0, + "content": "5.2. Comparison with SoTA Specialists" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.778 + ], + "angle": 0, + "content": "We evaluate the model's capacity to understand and reason in 3D environments via 3D-DC and 3D-QA. For each evaluation task, we fine-tune the trainable parameters in our model on each task for \\(\\sim 30\\mathrm{k}\\) iterations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "3D Dense Captioning demands a model to localize and describe any instance in a 3D scene. We benchmarks state-of-the-art methods on the widely-used ScanRefer [7] and \\(\\mathrm{Nr3D}\\) [1] dataset in Tab. 1 under the \\(m@k\\mathrm{IoU}\\) metric [12]. Here, \\(m \\in \\{\\mathrm{C}, \\mathrm{B-4}, \\mathrm{M}, \\mathrm{R}\\}\\), and the \\(m\\) score of a caption is set to 0 if the IoU between the predicted box and the object is less than the given threshold \\(k\\). Following existing works [10, 12], we consider \\(\\mathrm{C} @ 0.25\\) and \\(\\mathrm{C} @ 0.5\\) as the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.333 + ], + "angle": 0, + "content": "main metric for ScanRefer, and C@0.5 for Nr3D. Among the listed methods, UniT3D [13], 3DJCG [4], and 3D-VLP [34] are pre-trained on multiple 3D vision and language tasks annotated on ScanNet scenes. Additionally, UniT3D [13] adopts off-the-shelf image caption models [45] and multi-view images to generate additional instance-captions for pre-training. It is worth mentioning that we compare the results with the 3D-VisTA [70] model that is not trained on additional 3D scenes. To evaluate our model, we adopt the box predictions produced by Vote2Cap-DETR [10] as the visual prompt. Results show that our method consistently outperforms existing methods on both datasets. For example, our method achieves \\(65.19\\%\\) C@0.5 on ScanRefer and \\(51.18\\%\\) C@0.5 on Nr3D, which is \\((+3.38\\%\\) and \\(+7.34\\%)\\) higher than the current state-of-the-art 3D vision and language model, Vote2Cap-DETR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.472, + 0.548 + ], + "angle": 0, + "content": "3D Question Answering requires a model to generate responses to the natural language queries questioning towards an 3D scene. We benchmark state-of-the-art methods on the ScanQA [2] validation set as well as two test benchmarks in Tab. 2, and consider CiDEr as the main metric. The majority of the listed methods are based on classification (marked \"CLS\"), i.e., selecting responses from a predefined answer set. Meanwhile, 3D-LLM [29] tries to address 3D-QA via auto-regressive text generation (marked \"GEN\"), and we list their fine-tuned version for comparison. Results show that our method consistently outperforms existing methods on all the evaluation sets, and surpasses the generation based method, 3D-LLM, by a large margin \\((+7.39\\%\\) CiDEr score on the validation set)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.245, + 0.573 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.629 + ], + "angle": 0, + "content": "In this section, we provide ablation studies on model designs and training strategies. We evaluate on ScanRefer and ScanQA to quantize the effectiveness." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.644, + 0.27, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.729, + 0.245, + 0.742 + ], + "angle": 0, + "content": "(a) Early Fusion (ours)" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.644, + 0.465, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.729, + 0.428, + 0.742 + ], + "angle": 0, + "content": "(b) Direct Injection" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.752, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Figure 3. Two Different Ways to Encode Visual Prompts. Our proposed method (a) adopts a unified transformer to aggregate features from all kinds of interactions, while (b) directly concatenates the visual prompts to the scene embeddings. Tab. 3 shows that early fusion leads to a better performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of the Q-Former Design. We list two ways to process the visual prompts in Fig. 3. Here, Fig. 3 (a) is our proposed method that adopts a unified transformer to aggregate information from both text instructions and visual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "prompts, while Fig. 3 (b) is the \"direct injection\" version, which only extract instruction-aware 3D feature with visual prompts concatenated after the scene embeddings. We train both models from scratch and evaluate their performance on ScanRefer 3D Dense Captioning. The results (Fig. 3) show that the method we use (Fig. 3 (a)) could better capture feature related to the visual prompts, leading to better instance caption generation performance \\((+3.45\\% \\mathrm{C}@\\mathrm{0.5})\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.224, + 0.894, + 0.28 + ], + "angle": 0, + "content": "Table 3. Effectiveness of Q-Former Design on ScanRefer[7]. We design two different ways of utilizing visual prompts. The \"early fusion\" enables direct interaction with the 3D scene, thus it achieves a better performance." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.291, + 0.89, + 0.343 + ], + "angle": 0, + "content": "
Visual PromptC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
direct59.3933.2725.1953.39
ours62.8435.8125.8154.45
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Instructions as Auxiliary Tasks for 3D Dense Captioning. We have introduced two types of task instructions in Sec. 4.1 for 3D-DC, i.e. the \"describe\"-only instructions and \"detect and localize\" instructions. Additionally, we have introduced two types of visual prompts (Fig. 2 & Sec. 4.2). In this study, we show how they affect the performance when serving as auxiliary tasks for 3D-DC by evaluating on ScanRefer in Tab. 4. All the methods listed are trained from scratch. In Tab. 4, \"Aux.Loc\" identifies whether we train the model with the \"detect and localize\" instructions, and \"Clicks\" identifies whether we train the model with clicks as additional visual prompts. Results show that they are both good auxiliary tasks for 3D-DC." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.565, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Table 4. Effectiveness of Instructions as 3D Dense Captioning Auxiliary Tasks. We train the models from scratch and evaluate on ScanRefer[7]. \"Aux.Loc\" identifies whether we train with the \"describe and localize\" instructions. \"Clicks\" identifies whether we train with clicks as additional visual prompts." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.645, + 0.889, + 0.72 + ], + "angle": 0, + "content": "
Aux.LocClicksC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
--60.8534.0925.5353.48
✓-61.8134.1525.4953.83
-✓62.2034.2625.6753.87
✓✓62.8435.8125.8154.45
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Instructions as Auxiliary Tasks for 3D Question Answering. We have made a similar study to analyze how adopting additional \"answer and localize\" instructions and visual prompts improves 3D-QA on ScanQA [2] as auxiliary tasks in Tab. 6. We do not use any visual interactions during inference. Results show that the additional textual instructions and visual prompts improve the task diversity and further improve the performance for 3D-QA." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Performance as a Generalist. To test whether LL3DA can distinguish different tasks given the textual instructions and visual prompts, we evaluate our model on different tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26433" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Table 5. Evaluation as a Generalist. The first three rows list the performance of models trained from scratch as experts on each dataset. The results in the following three rows belong to the model fine-tuned from the generalist weights. The last row evaluates the model trained as a generalist. ScanRefer[7] and Nr3D[1] are used to evaluate the dense captioning performance, and ScanQA[2] is used to evaluate the question answering performance. Serving as a generalist, our method can differentiate each task, and produce strong results based on textual instructions and visual prompts." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.17, + 0.892, + 0.284 + ], + "angle": 0, + "content": "
MethodScanReferNr3DScanQA
C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C↑B-4↑M↑R↑
ScanRefer(scratch)62.8435.8125.8154.45--------
Nr3D(scratch)----44.9527.6725.6755.79----
ScanQA(scratch)--------74.8013.6815.4036.25
ScanRefer(fine-tuned)65.1936.7925.9755.06--------
Nr3D(fine-tuned)----51.1828.7525.9156.61----
ScanQA(fine-tuned)--------76.7913.5315.8837.31
w/o fine-tuning62.9835.9725.6654.6523.9413.3722.3145.7875.6713.3315.3737.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.472, + 0.39 + ], + "angle": 0, + "content": "Table 6. Effectiveness of Interactions as 3D Question Answering Auxiliary Tasks. We train the model from scratch and evaluate all the models from scratch on ScanQA[2] validation set. \"Aux.Loc\" identifies whether we train with the \"answer and localize\" instructions, and \"Visual Prompts\" identifies whether we train with visual prompts." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.401, + 0.468, + 0.464 + ], + "angle": 0, + "content": "
Aux.LocVisual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
--67.8511.8713.9633.87
✓-72.7313.2714.9035.87
-✓68.0912.5914.2033.71
✓✓74.8013.6815.4036.25
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.779 + ], + "angle": 0, + "content": "without task-specific fine-tuning in Tab. 5. The first three rows list the performance of LL3DA when trained from scratch on one specific task, while the following three rows represent the fine-tuned models. The last row indicates the direct evaluation of LL3DA. Results show that our model could distinguish 3D-DC and 3D-QA given the text instructions and visual prompts, and achieve strong performance (62.98% C@0.5 on ScanRefer, 75.67% CiDEr on ScanQA). However, the generalist model achieves poor performance on Nr3D [1], which is because we did not try to differentiate between Nr3D and ScanRefer during training in the first place, as ScanRefer and Nr3D are used for the same task. There is also an interesting observation that though we did not differentiate between these two datasets for 3D-DC, the model still tend to achieve high scores on ScanRefer (62.98% C@0.5). We are also excited to see that the weights of the generalist model can serve as a strong initialization for fine-tuning. For example, the fine-tuned model on ScanRefer could achieve 65.19% C@0.5, which is +2.35% higher than the model trained from scratch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Importance of Textual Instructions. We further conduct study to see whether the text instructions are necessary for 3D-DC in Tab. 7. The first row is our baseline method that directly generates the captions based on visual prompts without any text instructions, and the second row is our method that is trained with the text instructions introduced in Sec. 4.1. Both methods are trained from scratch for fair comparison. We notice that since the LLM is frozen, certain" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.339 + ], + "angle": 0, + "content": "textual instructions are beneficial when generating results in specific domains/tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Table 7. Effectiveness of Instructions on 3D Dense Captioning. We perform experiments on ScanRefer[7]. The baseline method directly generates the captions given the input 3D scene and visual prompts without any textual instructions." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.418, + 0.889, + 0.47 + ], + "angle": 0, + "content": "
InstructionsC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
-60.2034.7925.4054.03
✓62.8435.8125.8154.45
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Clicks for Better Question Answering. One major challenge of answering questions in complex 3D environments is the vague identification of objects with plain texts. Therefore, we try to click on some of the related objects along with the textual instructions during evaluation, and see how it could affect the generated answers on the ScanQA validation set in Tab. 8. Results show that this technique would remove the ambiguities, and further improve the quality of the answers \\((+6.12\\%\\) C). This illustrates the importance of visual interaction in complex 3D environments." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.707 + ], + "angle": 0, + "content": "Table 8. Test Time Visual Interactions for Question Answering on ScanQA[2]. The model achieves better performance on the question answering when we add visual prompts to some of the related objects along with the text instructions during evaluation." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.717, + 0.889, + 0.764 + ], + "angle": 0, + "content": "
Visual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
-76.7913.5315.8837.31
✓82.9111.8016.7439.97
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.684, + 0.803 + ], + "angle": 0, + "content": "5.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present several visualization results on different tasks in Fig. 4 to show our model's capacities in understanding, reasoning, and planning in different 3D environments. To prevent repetition when generating long sequences, we combine the top-\\(k\\) [22] and top-\\(p\\) [27] sampling strategy with \\(k = 50\\) and \\(p = 0.95\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26434" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.246, + 0.097, + 0.396, + 0.113 + ], + "angle": 0, + "content": "3D Dense Captioning" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.117, + 0.277, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.294, + 0.117, + 0.314, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.117, + 0.493, + 0.139 + ], + "angle": 0, + "content": "Describe this object in the 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.145, + 0.51, + 0.166 + ], + "angle": 0, + "content": "This is a brown armchair. It is to the right of the couch." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.145, + 0.554, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.177, + 0.526, + 0.199 + ], + "angle": 0, + "content": "SpaCap3D: This is a chair with no arms. \nIt is next to a desk." + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.202, + 0.536, + 0.224 + ], + "angle": 0, + "content": "Vote2Cap-DETR: This is a brown armchair. \nIt is to the right of the table." + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.226, + 0.538, + 0.259 + ], + "angle": 0, + "content": "Ground Truth: The armchair is facing the desk. The armchair is located to the right of the couch, in the middle of the room." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.279, + 0.402, + 0.294 + ], + "angle": 0, + "content": "3D Question Answering" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.291, + 0.287, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.31, + 0.485, + 0.331 + ], + "angle": 0, + "content": "How many chairs are behind the long table?" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.336, + 0.477, + 0.371 + ], + "angle": 0, + "content": "2 ch + Visual Prompts" + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.382, + 0.289, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.399, + 0.484, + 0.421 + ], + "angle": 0, + "content": "How many chairs are behind the long table? [Clicks]" + }, + { + "type": "text", + "bbox": [ + 0.359, + 0.43, + 0.509, + 0.451 + ], + "angle": 0, + "content": "3 chairs are behind long table" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.332, + 0.553, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.647, + 0.097, + 0.807, + 0.112 + ], + "angle": 0, + "content": "3D Embodied Planning" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.118, + 0.877, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.253, + 0.812, + 0.274 + ], + "angle": 0, + "content": "I want to clean the kitchen. What should I do?" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.286, + 0.842, + 0.338 + ], + "angle": 0, + "content": "1. pick up any trash off the floor. 2. Open the kitchen shelf. 3. Clean the kitchen sink with a cleaning solution. 4. Clean the kitchen sink with a paper towel." + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.343, + 0.597, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.607, + 0.35, + 0.856, + 0.424 + ], + "angle": 0, + "content": "I want to clean the kitchen. I have done these things: 1. Turn on the lamp on the table. 2. Open the kitchen cabinet and take out cleaning supplies. 3. Clean the countertops with a cloth. 4. Clean the stove and microwave. What should I do next?" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.436, + 0.837, + 0.448 + ], + "angle": 0, + "content": "Clean the refrigerator and cabinets" + }, + { + "type": "image", + "bbox": [ + 0.852, + 0.319, + 0.882, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.482, + 0.339, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.532, + 0.486, + 0.682, + 0.502 + ], + "angle": 0, + "content": "3D Scene Description" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.511, + 0.521, + 0.522 + ], + "angle": 0, + "content": "Describe this 3D scene." + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.532, + 0.842, + 0.616 + ], + "angle": 0, + "content": "This room is a kitchen with a stove, stovetop, cabinets, refrigerator, oven, cooler, recycling bins, microwave, water heater, soap dish, trash can, a sink, and a refrigerator. There are multiple walls scattered in the room, creating a separate space for dining and cooking. The door leads to the kitchen. Another door and several windows hang around the area, providing a large view of the floor. As well as additional doors, walls, windows, doorstopper caps, kitchen items, sink and fridge, the food preparation and storage areas are visible." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.641, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results. We provide several visualization results on various 3D vision and language tasks in diverse 3D environments (living room, classroom, kitchen, and bedroom). Red highlights the wrong answer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.204, + 0.697 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.9 + ], + "angle": 0, + "content": "In this paper, we present LL3DA, a large language 3D assistant that could take both textual- and visual- interactions from human for understanding, reasoning, and planning in complex 3D environments. Our model directly encodes the 3D point cloud and aggregates information from scenes and human interactions with the attention mechanism. We show that the visual interactions could remove the ambiguities in cluttered 3D environments, showing mighty instruction-following capacities. Experiments show that our method could achieve remarkable results on various 3D vision-language benchmarks. We hope our approach could inspire further designs and training strategies for large 3D language models. In future studies, we believe that the construction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.683, + 0.892, + 0.714 + ], + "angle": 0, + "content": "of high-quality and diverse annotations will further enhance the model's reasoning and planning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.688, + 0.734 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (No. 62071127 and 62101137), National Key Research and Development Program of China (No. 2022ZD0160100), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No. 2021SHZDZX0103), A*STAR AME Programmatic Funding A18A2b0046, RobotHTPO Seed Fund under Project C211518008, and EDB Space Technology Development Grant under Project S22-19016-STDP. The computations in this research were performed using the CFFF platform of Fudan University." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26435" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.325 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.449 + ], + "angle": 0, + "content": "[5] Yang Cao, Yihan Zeng, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.469, + 0.507 + ], + "angle": 0, + "content": "[6] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.622, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[9] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11124-11133, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. arXiv preprint arXiv:2309.02999, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[15] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[18] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[20] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[22] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833, 2018.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[24] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26436" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[25] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[26] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.223, + 0.47, + 0.264 + ], + "angle": 0, + "content": "[27] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.267, + 0.469, + 0.336 + ], + "angle": 0, + "content": "[28] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9202-9212, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.339, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[29] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.397, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[30] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.538 + ], + "angle": 0, + "content": "[31] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.541, + 0.469, + 0.582 + ], + "angle": 0, + "content": "[32] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[33] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[34] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.783 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[37] Mingsheng Li, Xin Chen, Chi Zhang, Sijin Chen, Hongyuan Zhu, Fukun Yin, Gang Yu, and Tao Chen. M3dbech: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[38] Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, and Yunchao Wei. Stablellava: Enhanced visual instruction tuning with synthesized image-dialogue data. arXiv preprint arXiv:2308.10253, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[43] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.433 + ], + "angle": 0, + "content": "[44] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[45] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.478, + 0.776, + 0.491 + ], + "angle": 0, + "content": "[46] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.493, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.564, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[48] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5606-5611, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[49] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[50] Zhende Song, Chenchen Wang, Jiamu Sheng, Chi Zhang, Gang Yu, Jiayuan Fan, and Tao Chen. Moviellm: Enhancing long video understanding with ai-generated movies. arXiv preprint arXiv:2403.01422, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[51] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26437" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1,3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[53] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[54] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.471, + 0.304 + ], + "angle": 0, + "content": "[55] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[56] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[57] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[59] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[60] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao Chen. Shapept: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[61] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[62] Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[63] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[64] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[65] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[66] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Towards explainable 3d grounded visual question answering:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "A new benchmark and strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[67] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[68] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[69] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26438" + } + ] +] \ No newline at end of file diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_origin.pdf b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bbbac9339fa894615a66ba48d480db84adf01100 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/e5af2afb-e504-4d6c-99d0-a3f72bf3d45e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff9d5b4505f818521b86af688387fbfb5009d7e975e1f3570b1a411f74f005f0 +size 2680061 diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/full.md b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1e3065fc297fb0b52cab2a32cdb6ccd00f10a07 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/full.md @@ -0,0 +1,386 @@ +# LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding, Reasoning, and Planning + +Sijin Chen $^{1}$ Xin Chen $^{2,*}$ Chi Zhang $^{2}$ Mingsheng Li $^{1}$ Gang YU $^{2}$ + +Hao Fei $^{3}$ Hongyuan Zhu $^{4}$ Jiayuan Fan $^{1}$ Tao Chen $^{1,\dagger}$ + +$^{1}$ Fudan University $^{2}$ Tencent PCG $^{3}$ National University of Singapore + +4Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore + +https://github.com/Open3DA/LL3DA + +* project lead † corresponding author + +![](images/5943c9dff48a6fa1c0ca0e1f85368bc41d1194b3b08c82a1d0f10a40fbd56f44.jpg) +Figure 1. We propose LL3DA, a Large Language 3D Assistant that demonstrates mighty instruction-following capacities in understanding, reasoning, and planning within complex 3D environments. LL3DA takes both the textual instructions and potential visual interactions into consideration to help remove ambiguities when addressing various tasks in diverse and complex 3D scenes. + +# Abstract + +Recent progress in Large Multimodal Models (LMM) has opened up great possibilities for various applications in the field of human-machine interactions. However, developing LMMs that can comprehend, reason, and plan in complex and diverse 3D environments remains a challenging topic, especially considering the demand for understanding permutation-invariant point cloud representations of the 3D scene. Existing works seek help from multi-view images by projecting 2D features to 3D space, which inevitably leads to huge computational overhead and performance degradation. In this paper, we present LL3DA, a Large Language 3D Assistant that takes point cloud as the direct input and responds to both text instructions and visual interactions. The additional visual interaction enables LMMs to better + +comprehend human interactions with the 3D environment and further remove the ambiguities within plain texts. Experiments show that LL3DA achieves remarkable results and surpasses various 3D vision-language models on both 3D Dense Captioning and 3D Question Answering. + +# 1. Introduction + +The recent surge in Large Language Model (LLM) families [14, 31, 46, 52, 63] opens up great opportunities for addressing various machine learning tasks in a generalized way [30, 32, 40, 60]. During this LLM carnival, researchers are also seeking generalized LLM solutions for various vision language tasks [37, 50, 62]. Among these, LLM-based 3D scene understanding is a valuable topic that would benefit the development of autonomous driving [9, 23] and em + +bodied AI agents [21, 49]. However, it is also challenging given 1) the diversity and complication of 3D environments and 2) the demands for understanding sparse 3D points. + +Prior works have made initial success in various 3D vision and language tasks. The majority of these research build 3D specialists to solve one specific down-stream task, including 3D Question Answering (3D-QA) [2, 43], 3D Visual Grounding (3D-VG) [7, 26, 55], and 3D Dense Captioning (3D-DC) [10-12]. There are also several works [4, 13, 34, 70] study the mutual promotion of different 3D vision and language tasks with shared structure modelling relations among objects. Recently, researchers have also introduced LLMs for general purpose 3D understanding, where Point-Bind and Point-LLMs [24, 57] mainly focus on the understanding of 3D objects. Concurrently, 3D-LLM [29] proposes an LLM-driven solution that aggregates multi-view features for 3D perception, presenting mighty capacities in understanding 3D object and scenes and following text instructions produced by human. + +Though these methods have achieved remarkable success in addressing various challenges in understanding the 3D world with natural language, there are certain limitations. With limited supervision, 3D specialists could hardly scale-up for better performance, while the joint pre-training still requires separate heads for specific tasks. Extracting multi-view features results in huge computational overhead and ignores the essential geometry and depth information. Additionally, plain texts often lead to ambiguities especially in cluttered and complex 3D environments. + +To address the above issues, we propose LL3DA, a Large Language 3D Assistant that responds to both textual and visual interactions from human, with the ability to understand, reason, and plan in complex 3D environments (Fig. 1). We adopt a multi-modal transformer to aggregate information from textual instructions, visual prompts, and the 3D scene into a fixed length of learnable querying tokens via the attention mechanism. The querying tokens are then projected and used as the prefix for the textual instructions, serving as the input to a pre-trained and frozen LLM. This design not only helps to address the contradiction between the permutation-invariant 3D scene representations and the LLM embedding space, but also extracts interaction-aware 3D scene representations for efficient instruction following. + +We conduct extensive experiments to explore the capacities of LL3DA in understanding, reasoning, and planning within complex and diverse 3D environments. Our model achieves state-of-the-art results on two widely used datasets for 3D Dense Captioning [1, 7], and 3D Question Answering [2]. Additionally, by introducing additional visual interactions, our method could further remove the ambiguities within the vague textual instructions. + +To summarize, our key contributions lie in: + +- We present a LLM-based solution for understanding, rea + +soning, and planning in complex 3D environments. + +- Our model takes both the textual instructions and visual interactions as inputs, and extracts interaction-aware features for effective instruction-following. +- Extensive experiments show that our method surpasses various state-of-the-art 3D vision language models. + +# 2. Related Work + +3D Vision and Language alignment, pre-training, and understanding [5, 7, 20, 70] cover tasks requiring a model to adopt its understanding towards a complex 3D scene answering to, or answering with natural language. Among those, 3D Dense Captioning (3D-DC) [10, 12, 54] expects a model to translate an input 3D scene into a set of instance coordinates and natural language descriptions. Existing methods could be categorized into "detect-then-describe" models [4, 12, 54] and the "set-to-set" prediction approaches [10, 11]. The former builds explicit relations on the instance coordinate estimations, while the latter directly learns the locations and descriptions for instances from the input 3D scene. 3D Visual Grounding (3D-VG) [1, 7, 55] demands a model to respond the natural language queries with the instance coordinates in the 3D scene. The mainstream of existing methods [4, 65, 70] address 3D-VG via selecting a candidate from a 3D detector's prediction. 3D Question Answering (3D-QA) [2, 43, 59, 66] requires a model to answer the questions with natural language based on the input 3D scene. The majority of existing methods [2, 18, 48] directly select the desired response from a given answer set. Researchers have also studied the mutual promotion of various 3D vision language tasks via training their shareable architectures simultaneously on multiple tasks [4, 13, 34, 70]. UniT3D [13] and 3DJCG [4] focus on the joint promotion between 3D-DC and 3D-VG in the relation modelling, while 3D-VLP [34] further includes 3D-QA. Recently, 3D-LLM [29] introduces a family of LLM-driven 3D generalists that could handle diverse textual instructions with reconstructed 3D features from multi-view images [28]. In this paper, we present LL3DA, an LLM solution that directly extracts features from the 3D scene, and handles both visual prompts and textual instructions to diversify the possible interactions human could make with the complex 3D environments. + +Large Multimodal Models (LMM). Along with the rapid development of Large Language Models (LLM) [15, 63], researchers have made great recent efforts adapting LLMs to visual understanding and reasoning tasks [25, 38, 56, 61]. Some methods project or compress global image features as the prefix for text instructions [36, 40, 58, 69], while others extract ROI features as LLM tokens for region-oriented instruction reasoning [6, 64]. Meanwhile, InstructBLIP [17] proposes to extract textual instruction-aware visual features, and has achieved remarkable success in addressing complex + +![](images/6179a1c6621ee1b02340e9cd76f3c56c9cc2bbd5cbc2f53b30c3188e4be101a3.jpg) +Figure 2. Overview of the Proposed Approach. (a) The overall pipeline of our proposed LL3DA first extracts interaction-aware 3D scene embeddings, which are later projected to the prefix of textual instructions as the input of a frozen LLM. (b) The detailed design of the Interactor3D, which aggregates visual prompts, textual instructions, and 3D scene embeddings into a fixed length querying tokens. (c) The prompt encoder encodes the user clicks and box coordinates with the positional embeddings and ROI features, respectively. + +and unseen instructions. Concurrently, researchers have also made great attempts solving various 3D tasks using LLMs. Notably, [24, 42, 57, 68] demonstrate remarkable success in understanding and reasoning about 3D objects. In this paper, we present an LLM-driven solution that could handle both interactions in forms of visual prompts and text instructions. We also propose to extract interaction-aware 3D scene representations for better instruction following. + +# 3. Methodology + +To build a general purpose agent that could handle both visual and textual interactions with complex 3D environments, we propose LL3DA, an LLM driven auto-regressive approach to 3D vision language tasks. In this section, we first introduce the problem formatting in Sec. 3.1. Next, we introduce our model designs in details (Sec. 3.2). + +# 3.1. Problem Formatting + +Model I/O. As shown in Fig. 2 (a), the input of our model consists of a 3D scene represented by a set of points $PC$ , the textual instruction $\mathcal{I}_t$ , and potential visual interactions $\mathcal{I}_v$ that serve as supplementary spatial identifiers. Here, point cloud $PC = [p_{in}, f_{in}] \in \mathbb{R}^{N \times (3 + F)}$ , where $p_{in} \in \mathbb{R}^{N \times 3}$ and $f_{in} \in \mathbb{R}^{N \times F}$ are the point coordinates and the additional point features, including color, normal, and height. The output of our model is free-form natural language, part of whom could be interpreted into 3D coordinates. + +Instruction Formatting. Following existing LMMs [57], we begin the textual instructions $\mathcal{I}_t$ with the “human:” identifier, and ask the model to generate responses after the “assistant:” identifier. This endows the model with the + +ability to distinguish information from the context and further engage in multi-turn conversations. + +Coordinate Representations. To provide LLMs with the capacity to perceive and respond with 3D coordinates, we convert the 3D points and 3D bounding boxes to plain texts. Specifically, a point is represented by “ $\langle \mathrm{loc} \rangle x, y, z \langle \mathrm{loc} \rangle$ ”, and a bounding box is represented by its center point and size, i.e. “ $\langle \mathrm{obj} \rangle c_x, c_y, c_z, w, h, l \langle \mathrm{obj} \rangle$ ”. Here, all the numerical data is discretized into unsigned integers within a range of [0, 255] with respect to the boundary of the input 3D scene. This design could naturally fit in the vocabulary of existing pre-trained LLMs [52, 63]. Without the introduction of any additional learnable tokens, we could save the effort of tuning the LLMs. + +# 3.2. Model Design + +As shown in Fig. 2 (a), our model first aggregates a fixed-length scene embeddings through the Interactor3D, which takes the visual prompts, the textual instructions, and the 3D scene as the input. Next, the aggregated 3D scene embeddings are projected to be the prefix of text instructions, and serve as the inputs of a frozen LLM. The detailed design of Interactor3D is shown in Fig. 2 (b), which consists of a frozen 3D scene encoder $\mathcal{E}^{3D}$ , a visual prompt encoder, and a multi-modal transformer. + +Scene Encoder. We adopt the masked transformer encoder pre-trained on ScanNet detection [10] as the scene encoder, $\mathcal{E}^{3D}$ . The scene encoder takes $PC$ as its input, and outputs the 3D scene embeddings: + +$$ +f _ {e n c} = \mathcal {E} ^ {3 D} (P C) = \mathcal {E} ^ {3 D} \left(p _ {i n}; f _ {i n}\right) \in \mathbb {R} ^ {M \times d}. \tag {1} +$$ + +Here, $f_{enc}$ consists of $d$ -dimensioned features for $M$ points + +uniformly down-sampled from the input 3D scene through the Farthest Point Sampling (FPS) algorithm. In practice, we choose to keep the scene encoder frozen to save the memory cost during training. + +Visual Prompt Encoder. We mainly take two common types of visual interactions into consideration, user clicks and 3D box annotations [35]. Each user click is first normalized within a range of $[0,1]$ by the size of the input 3D scene $p_{\mathrm{click}} \in \mathbb{R}^3$ . Then, we encode $p_{\mathrm{click}}$ with the 3D Fourier positional embeddings [51] function: + +$$ +\left. \operatorname {p o s} \left(p _ {\text {c l i c k}}\right) = \left[ \sin \left(2 \pi p _ {\text {c l i c k}} \cdot B\right); \cos \left(2 \pi p _ {\text {c l i c k}} \cdot B\right) \right]. \right. \tag {2} +$$ + +Here, $B \in \mathbb{R}^{3 \times (d/2)}$ is a learnable matrix. The box annotation is represented by the ROI feature $f_{\mathrm{box}} \in \mathbb{R}^d$ extracted by a pre-trained 3D object detector [10]. The two types of the visual prompts are then projected with separate and identical Feed Forward Networks (FFN). + +$$ +f _ {\text {c l i c k}} = F F N _ {\text {c l i c k}} \left(\operatorname {p o s} \left(p _ {\text {c l i c k}}\right)\right) \tag {3} +$$ + +$$ +f _ {\text {b o x}} = F F N _ {\text {b o x}} \left(f _ {\text {b o x}}\right) +$$ + +In practice, we represent each visual prompt with 8 tokens. + +Multi-Modal Transformer (MMT) serves as a role to 1) address the contradiction between the permutation-invariant 3D scene embeddings and position-sensitive causal LLMs, 2) bridge the gap between frozen unimodal experts, and 3) fill the needs for interaction-aware feature extraction. Inspired by the Q-Former architecture [17, 36], MMT aggregates the visual information within a fixed number of 32 learnable querying tokens. In each layer, the queries interact with the encoded visual prompts $[f_{\mathrm{click}}; f_{\mathrm{box}}]$ and the textual instructions $\mathcal{I}_t$ through a shared self-attention. Then, we allow the learnable querying tokens and the visual prompts to interact with the task-agnostic 3D scene embeddings $f_{enc}$ via cross-attention. The output of MMT is 32 queries written as $Q \in \mathbb{R}^{32 \times 768}$ , which are finally projected to the word embedding space of LLMs through a simple linear projector. In practice, we notice that initializing Q-Former with pre-trained BERT [19, 36] weights will lead to repetitive outputs, thus we only choose to initialize the pre-trained word and position embeddings from BERT. + +LLM. We consider the decoder-only generative pre-trained transformers [52, 63] as our large language model backbone. The decoder-only LLMs are sensitive to the input orders because of the position embeddings and the causal attention mask. The parameters and the embedding layers of the LLMs are kept frozen to save memory cost. During inference, we generate the responses via searching for the optimal sequence $s^*$ that satisfies: + +$$ +s ^ {*} = \arg \max _ {s} P (s | P C, \mathcal {I} _ {t}, \mathcal {I} _ {v}). \tag {4} +$$ + +In practice, we use beam search with a beam size of 4. + +# 4. Multi-modal Instruction Tuning + +A general purpose 3D agent is able to address various tasks simultaneously in complex 3D scenes. Apart from introducing proper training data, it is important to guide the model to generate the desired outputs with instructions. Therefore, Sec. 4.1 will first introduce how we identify each task. After that, Sec. 4.2 will present details for the training objective. + +# 4.1. Tasks and Instructions. + +As introduced in Sec. 3.1, LL3DA generates text responses auto-regressively after the “```assistant:” identifier. + +3D Dense Captioning requires the localization and description of instances in diverse 3D environments. We adopt either user clicks and box annotations as the visual prompt to identify the object to be described. Additionally, we design two types of textual instructions that ask the model to either "describe" or "describe and localize" the object, which diversifies the tasks, and leads to better performance. + +3D Question Answering requires the model to generate response to the questions based on the global knowledge of a 3D scene. To help the model better understand the 3D environment, we also design two types of textual instructions that ask the model to either "answer" or "answer and localize the related objects". The latter serves as an auxiliary task widely adopted in various 3D-QA methods [2, 48]. To diversify the tasks during training, we randomly include additional clicks on the objects related to the questions. + +Scene Description requires the model to translate its global knowledge of the 3D scene into natural language descriptions, thus we simply ask the "describe" this 3D scene. + +Embodied Conversation and Planning could be treated as multi-turn conversations, where we use “human” and “assistant” as identifiers to distinguish the source of information as introduced in Sec. 3.1. + +# 4.2. Instruction Following Tuning + +During training, for tasks requiring additional visual interactions, i.e. 3D-DC and 3D-QA, we randomly choose between clicks or boxes as means of object identification. + +Training Objective. Our training objective is to optimize the trainable parameters $\theta$ , so as to maximize the likelihood of the target response sequence $s$ given the input point cloud $PC$ , and the human interactions $\mathcal{I}_v$ and $\mathcal{I}_t$ : + +$$ +\theta^ {*} = \arg \max _ {\theta} P (s | P C; \mathcal {I} _ {v}; \mathcal {I} _ {t}; \theta). \tag {5} +$$ + +In practice, this is accomplished by adopting the token-wise cross-entropy loss that trains the model to predict the $i$ th token $s_{[i]}$ given the previous $(i - 1)$ tokens, $s_{[1,\dots ,i - 1]}$ . + +$$ +\mathcal {L} (\theta) = - \sum_ {i = 1} ^ {| s |} \log P (s _ {[ i ]} | P C; \mathcal {I} _ {v}; \mathcal {I} _ {t}; \theta ; s _ {[ 1, \dots , i - 1 ]}). \tag {6} +$$ + +Table 1. Quantitative Comparisons for 3D Dense Captioning on ScanRefer[7] and Nr3D[1]. For fair comparison, we list methods that are trained under the standard per-word cross-entropy loss without additional 3D scenes. We use the box estimations from Vote2Cap-DETR to simulate the box annotations as the visual prompts. Our proposed LL3DA surpasses previous 3D specialists on both datasets. + +
MethodScanReferNr3D
C@0.25↑B-4@0.25↑M@0.25↑R@0.25↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[33]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[54]----44.0225.2622.3345.3633.7119.9222.6150.50
REMAN[44]62.0136.3726.7656.2545.0026.3122.6746.9634.8120.3723.0150.99
D3Net[8]----46.0730.2924.3551.6733.8520.7023.1353.38
Contextual[67]----46.1125.4722.6445.9635.2620.4222.7750.78
UniT3D[13]----46.6927.2221.9145.98----
3DJCG[4]64.7040.1727.6659.2349.4831.0324.2250.8038.0622.8223.7752.99
3D-VLP[34]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2Cap-DETR[10]71.4539.3428.2559.3361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA (Ours)74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
+ +Table 2. Quantitative Comparisons for 3D Question Answering on ScanQA[2]. We categorize previous works into classification based ("CLS") and generation based ("GEN") methods. The results from 3D-LLM* come from their fine-tuned version. LL3DA out-performs previous methods on the validation set and two test sets. + +
MethodAnswer TypeValidationTest w/ objectTest w/o object
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
ScanQA[2]64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[48]----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[18]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[34]66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM*[29]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA (Ours)76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
+ +Here, $|s|$ is the number of tokens in the desired response. + +# 5. Experiments + +To test the capacities of LL3DA, we provide numerous evaluations. To begin with, we introduce the datasets, metrics, and implementation details (Sec. 5.1). Then, we compare how our model understands and reasons in complex 3D environments with previous 3D specialists on 3D Dense Captioning and 3D Question Answering (Sec. 5.2). Additionally, we conduct quantitative ablation studies on the model design and training strategy (Sec. 5.3). Finally, Sec. 5.4 showcases several qualitative results. + +# 5.1. Datasets, Metrics and Implementation Details + +Datasets. In this paper, we experiment with 3D data from ScanNet [16], a 3D dataset covering 1,201 and 312 diverse and complex indoor 3D scenes for training and validation. The language annotations used in this study are sourced from ScanRefer [7], Nr3D [1], ScanQA [2], and the ScanNet subset of 3D-LLM [29]. This combination covers a variety of tasks, including instance and scene descriptions, conversations, embodied planning and question answering. Please refer to the supplementary materials for more details on the statistics of data. + +Metrics. Here, we adopt C, B-4, M, R as abbreviations for + +CiDEr [53], BLEU-4 [47],METEOR [3],and Rouge-L [39] to evaluate the quality of the generated textual responses. + +Implementation Details. Following previous works on 3D vision language tasks [10, 12], we randomly sample $40\mathrm{k}$ points from each 3D scene as the 3D input. We adopt the pre-trained OPT-1.3B [63] as our causal LLM backbone, which is frozen and loaded in float16 to save memory cost. We adopt the AdamW [41] optimizer with a weight decay of 0.1 and a learning rate decaying from $10^{-4}$ to $10^{-6}$ with a cosine annealing scheduler for about $100\mathrm{k}$ iterations. For all the training tasks, we train with no more than eight Nvidia RTX3090 (24G) GPUs within a day. + +# 5.2. Comparison with SoTA Specialists + +We evaluate the model's capacity to understand and reason in 3D environments via 3D-DC and 3D-QA. For each evaluation task, we fine-tune the trainable parameters in our model on each task for $\sim 30\mathrm{k}$ iterations. + +3D Dense Captioning demands a model to localize and describe any instance in a 3D scene. We benchmarks state-of-the-art methods on the widely-used ScanRefer [7] and $\mathrm{Nr3D}$ [1] dataset in Tab. 1 under the $m@k\mathrm{IoU}$ metric [12]. Here, $m \in \{\mathrm{C}, \mathrm{B-4}, \mathrm{M}, \mathrm{R}\}$ , and the $m$ score of a caption is set to 0 if the IoU between the predicted box and the object is less than the given threshold $k$ . Following existing works [10, 12], we consider $\mathrm{C} @ 0.25$ and $\mathrm{C} @ 0.5$ as the + +main metric for ScanRefer, and C@0.5 for Nr3D. Among the listed methods, UniT3D [13], 3DJCG [4], and 3D-VLP [34] are pre-trained on multiple 3D vision and language tasks annotated on ScanNet scenes. Additionally, UniT3D [13] adopts off-the-shelf image caption models [45] and multi-view images to generate additional instance-captions for pre-training. It is worth mentioning that we compare the results with the 3D-VisTA [70] model that is not trained on additional 3D scenes. To evaluate our model, we adopt the box predictions produced by Vote2Cap-DETR [10] as the visual prompt. Results show that our method consistently outperforms existing methods on both datasets. For example, our method achieves $65.19\%$ C@0.5 on ScanRefer and $51.18\%$ C@0.5 on Nr3D, which is $(+3.38\%$ and $+7.34\%)$ higher than the current state-of-the-art 3D vision and language model, Vote2Cap-DETR. + +3D Question Answering requires a model to generate responses to the natural language queries questioning towards an 3D scene. We benchmark state-of-the-art methods on the ScanQA [2] validation set as well as two test benchmarks in Tab. 2, and consider CiDEr as the main metric. The majority of the listed methods are based on classification (marked "CLS"), i.e., selecting responses from a predefined answer set. Meanwhile, 3D-LLM [29] tries to address 3D-QA via auto-regressive text generation (marked "GEN"), and we list their fine-tuned version for comparison. Results show that our method consistently outperforms existing methods on all the evaluation sets, and surpasses the generation based method, 3D-LLM, by a large margin $(+7.39\%$ CiDEr score on the validation set). + +# 5.3. Ablation Studies + +In this section, we provide ablation studies on model designs and training strategies. We evaluate on ScanRefer and ScanQA to quantize the effectiveness. + +![](images/516496b47dc1454d026b87181128be058b7e5996ba94be0f5b9ed19d9d9e7d73.jpg) +(a) Early Fusion (ours) + +![](images/79369773bc285f5d43fbf7e72cee7a3ecf91909680d3994ac0db3976bed889ee.jpg) +(b) Direct Injection +Figure 3. Two Different Ways to Encode Visual Prompts. Our proposed method (a) adopts a unified transformer to aggregate features from all kinds of interactions, while (b) directly concatenates the visual prompts to the scene embeddings. Tab. 3 shows that early fusion leads to a better performance. + +Effectiveness of the Q-Former Design. We list two ways to process the visual prompts in Fig. 3. Here, Fig. 3 (a) is our proposed method that adopts a unified transformer to aggregate information from both text instructions and visual + +prompts, while Fig. 3 (b) is the "direct injection" version, which only extract instruction-aware 3D feature with visual prompts concatenated after the scene embeddings. We train both models from scratch and evaluate their performance on ScanRefer 3D Dense Captioning. The results (Fig. 3) show that the method we use (Fig. 3 (a)) could better capture feature related to the visual prompts, leading to better instance caption generation performance $(+3.45\% \mathrm{C}@\mathrm{0.5})$ + +Table 3. Effectiveness of Q-Former Design on ScanRefer[7]. We design two different ways of utilizing visual prompts. The "early fusion" enables direct interaction with the 3D scene, thus it achieves a better performance. + +
Visual PromptC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
direct59.3933.2725.1953.39
ours62.8435.8125.8154.45
+ +Instructions as Auxiliary Tasks for 3D Dense Captioning. We have introduced two types of task instructions in Sec. 4.1 for 3D-DC, i.e. the "describe"-only instructions and "detect and localize" instructions. Additionally, we have introduced two types of visual prompts (Fig. 2 & Sec. 4.2). In this study, we show how they affect the performance when serving as auxiliary tasks for 3D-DC by evaluating on ScanRefer in Tab. 4. All the methods listed are trained from scratch. In Tab. 4, "Aux.Loc" identifies whether we train the model with the "detect and localize" instructions, and "Clicks" identifies whether we train the model with clicks as additional visual prompts. Results show that they are both good auxiliary tasks for 3D-DC. + +Table 4. Effectiveness of Instructions as 3D Dense Captioning Auxiliary Tasks. We train the models from scratch and evaluate on ScanRefer[7]. "Aux.Loc" identifies whether we train with the "describe and localize" instructions. "Clicks" identifies whether we train with clicks as additional visual prompts. + +
Aux.LocClicksC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
--60.8534.0925.5353.48
✓-61.8134.1525.4953.83
-✓62.2034.2625.6753.87
✓✓62.8435.8125.8154.45
+ +Instructions as Auxiliary Tasks for 3D Question Answering. We have made a similar study to analyze how adopting additional "answer and localize" instructions and visual prompts improves 3D-QA on ScanQA [2] as auxiliary tasks in Tab. 6. We do not use any visual interactions during inference. Results show that the additional textual instructions and visual prompts improve the task diversity and further improve the performance for 3D-QA. + +Performance as a Generalist. To test whether LL3DA can distinguish different tasks given the textual instructions and visual prompts, we evaluate our model on different tasks + +Table 5. Evaluation as a Generalist. The first three rows list the performance of models trained from scratch as experts on each dataset. The results in the following three rows belong to the model fine-tuned from the generalist weights. The last row evaluates the model trained as a generalist. ScanRefer[7] and Nr3D[1] are used to evaluate the dense captioning performance, and ScanQA[2] is used to evaluate the question answering performance. Serving as a generalist, our method can differentiate each task, and produce strong results based on textual instructions and visual prompts. + +
MethodScanReferNr3DScanQA
C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C↑B-4↑M↑R↑
ScanRefer(scratch)62.8435.8125.8154.45--------
Nr3D(scratch)----44.9527.6725.6755.79----
ScanQA(scratch)--------74.8013.6815.4036.25
ScanRefer(fine-tuned)65.1936.7925.9755.06--------
Nr3D(fine-tuned)----51.1828.7525.9156.61----
ScanQA(fine-tuned)--------76.7913.5315.8837.31
w/o fine-tuning62.9835.9725.6654.6523.9413.3722.3145.7875.6713.3315.3737.02
+ +Table 6. Effectiveness of Interactions as 3D Question Answering Auxiliary Tasks. We train the model from scratch and evaluate all the models from scratch on ScanQA[2] validation set. "Aux.Loc" identifies whether we train with the "answer and localize" instructions, and "Visual Prompts" identifies whether we train with visual prompts. + +
Aux.LocVisual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
--67.8511.8713.9633.87
✓-72.7313.2714.9035.87
-✓68.0912.5914.2033.71
✓✓74.8013.6815.4036.25
+ +without task-specific fine-tuning in Tab. 5. The first three rows list the performance of LL3DA when trained from scratch on one specific task, while the following three rows represent the fine-tuned models. The last row indicates the direct evaluation of LL3DA. Results show that our model could distinguish 3D-DC and 3D-QA given the text instructions and visual prompts, and achieve strong performance (62.98% C@0.5 on ScanRefer, 75.67% CiDEr on ScanQA). However, the generalist model achieves poor performance on Nr3D [1], which is because we did not try to differentiate between Nr3D and ScanRefer during training in the first place, as ScanRefer and Nr3D are used for the same task. There is also an interesting observation that though we did not differentiate between these two datasets for 3D-DC, the model still tend to achieve high scores on ScanRefer (62.98% C@0.5). We are also excited to see that the weights of the generalist model can serve as a strong initialization for fine-tuning. For example, the fine-tuned model on ScanRefer could achieve 65.19% C@0.5, which is +2.35% higher than the model trained from scratch. + +Importance of Textual Instructions. We further conduct study to see whether the text instructions are necessary for 3D-DC in Tab. 7. The first row is our baseline method that directly generates the captions based on visual prompts without any text instructions, and the second row is our method that is trained with the text instructions introduced in Sec. 4.1. Both methods are trained from scratch for fair comparison. We notice that since the LLM is frozen, certain + +textual instructions are beneficial when generating results in specific domains/tasks. + +Table 7. Effectiveness of Instructions on 3D Dense Captioning. We perform experiments on ScanRefer[7]. The baseline method directly generates the captions given the input 3D scene and visual prompts without any textual instructions. + +
InstructionsC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
-60.2034.7925.4054.03
✓62.8435.8125.8154.45
+ +Clicks for Better Question Answering. One major challenge of answering questions in complex 3D environments is the vague identification of objects with plain texts. Therefore, we try to click on some of the related objects along with the textual instructions during evaluation, and see how it could affect the generated answers on the ScanQA validation set in Tab. 8. Results show that this technique would remove the ambiguities, and further improve the quality of the answers $(+6.12\%$ C). This illustrates the importance of visual interaction in complex 3D environments. + +Table 8. Test Time Visual Interactions for Question Answering on ScanQA[2]. The model achieves better performance on the question answering when we add visual prompts to some of the related objects along with the text instructions during evaluation. + +
Visual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
-76.7913.5315.8837.31
✓82.9111.8016.7439.97
+ +# 5.4. Qualitative Results + +We present several visualization results on different tasks in Fig. 4 to show our model's capacities in understanding, reasoning, and planning in different 3D environments. To prevent repetition when generating long sequences, we combine the top- $k$ [22] and top- $p$ [27] sampling strategy with $k = 50$ and $p = 0.95$ . + +# 3D Dense Captioning + +![](images/774a3fe5adcf1de1ef3e1b51cf62afdca7d7fd4c7522ada6b691962a1a12cf96.jpg) + +![](images/503c80453692779fabc10ce921f8b4fa6de1850c1b656409b9c8c2091bcbec18.jpg) + +Describe this object in the 3D scene. + +This is a brown armchair. It is to the right of the couch. + +![](images/f88f3c5b8e779ecbd74d81be36c5088ac2b98e5c2939ccff3d66007b1e628564.jpg) + +SpaCap3D: This is a chair with no arms. +It is next to a desk. + +Vote2Cap-DETR: This is a brown armchair. +It is to the right of the table. + +Ground Truth: The armchair is facing the desk. The armchair is located to the right of the couch, in the middle of the room. + +# 3D Question Answering + +![](images/128a305b4cf2436f9018718d41e48b6ab6445973196e45310f870c2a2fc447b8.jpg) + +How many chairs are behind the long table? + +2 ch + Visual Prompts + +![](images/0ed03e0f33b4960b11c16d41715dd077eda723133fe489a36d6921b6963e8d05.jpg) + +How many chairs are behind the long table? [Clicks] + +3 chairs are behind long table + +![](images/acec21fe01ee05cb9894c68945854c163eafe73e4b57b82637be045b863290b6.jpg) + +# 3D Embodied Planning + +![](images/39890bb25308a1c43a3d059048182459ae626122deb627f022f787ec3001e298.jpg) + +I want to clean the kitchen. What should I do? + +1. pick up any trash off the floor. 2. Open the kitchen shelf. 3. Clean the kitchen sink with a cleaning solution. 4. Clean the kitchen sink with a paper towel. + +![](images/2e53409ca803248459a5c65b178fa6305e5d8dfe40bda1c8906ab5ec83c6895f.jpg) + +I want to clean the kitchen. I have done these things: 1. Turn on the lamp on the table. 2. Open the kitchen cabinet and take out cleaning supplies. 3. Clean the countertops with a cloth. 4. Clean the stove and microwave. What should I do next? + +Clean the refrigerator and cabinets + +![](images/1c4125d18ffa1d6187efc09522e9fa038b5f2eb398a348d4b213e9c23fc2ee63.jpg) + +![](images/3796156bba50e55762b755a1d55c3160e3e6861b27928eb8dafa34d0bc559a4e.jpg) +Figure 4. Qualitative Results. We provide several visualization results on various 3D vision and language tasks in diverse 3D environments (living room, classroom, kitchen, and bedroom). Red highlights the wrong answer. + +# 3D Scene Description + +Describe this 3D scene. + +This room is a kitchen with a stove, stovetop, cabinets, refrigerator, oven, cooler, recycling bins, microwave, water heater, soap dish, trash can, a sink, and a refrigerator. There are multiple walls scattered in the room, creating a separate space for dining and cooking. The door leads to the kitchen. Another door and several windows hang around the area, providing a large view of the floor. As well as additional doors, walls, windows, doorstopper caps, kitchen items, sink and fridge, the food preparation and storage areas are visible. + +# 6. Conclusions + +In this paper, we present LL3DA, a large language 3D assistant that could take both textual- and visual- interactions from human for understanding, reasoning, and planning in complex 3D environments. Our model directly encodes the 3D point cloud and aggregates information from scenes and human interactions with the attention mechanism. We show that the visual interactions could remove the ambiguities in cluttered 3D environments, showing mighty instruction-following capacities. Experiments show that our method could achieve remarkable results on various 3D vision-language benchmarks. We hope our approach could inspire further designs and training strategies for large 3D language models. In future studies, we believe that the construction + +of high-quality and diverse annotations will further enhance the model's reasoning and planning capabilities. + +# 7. Acknowledgements + +This work is supported by National Natural Science Foundation of China (No. 62071127 and 62101137), National Key Research and Development Program of China (No. 2022ZD0160100), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No. 2021SHZDZX0103), A*STAR AME Programmatic Funding A18A2b0046, RobotHTPO Seed Fund under Project C211518008, and EDB Space Technology Development Grant under Project S22-19016-STDP. The computations in this research were performed using the CFFF platform of Fudan University. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 2, 5, 7 +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 4, 5, 6, 7 +[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 5 +[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 5, 6 +[5] Yang Cao, Yihan Zeng, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, 2023. 2 +[6] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 2 +[7] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2, 5, 6, 7 +[8] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551, 2021. 5 +[9] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1 +[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11124-11133, 2023. 2, 3, 4, 5, 6 +[11] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. arXiv preprint arXiv:2309.02999, 2023. 2 +[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2, 5 + +[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2, 5, 6 +[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 1 +[15] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2 +[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5 +[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 4 +[18] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 2, 5 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4 +[20] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2 +[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2 +[22] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833, 2018.7 +[23] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 1 +[24] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2, 3 + +[25] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 2 +[26] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2 +[27] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 7 +[28] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9202-9212, 2023. 2 +[29] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981, 2023. 2, 5, 6 +[30] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1 +[31] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 1 +[32] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 1 +[33] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203, 2022. 5 +[34] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 5, 6 +[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 4 +[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 4 +[37] Mingsheng Li, Xin Chen, Chi Zhang, Sijin Chen, Hongyuan Zhu, Fukun Yin, Gang Yu, and Tao Chen. M3dbech: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763, 2023. 1 + +[38] Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, and Yunchao Wei. Stablellava: Enhanced visual instruction tuning with synthesized image-dialogue data. arXiv preprint arXiv:2308.10253, 2023. 2 +[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5 +[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2 +[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3 +[43] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2 +[44] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics, 2023. 5 +[45] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6 +[46] OpenAI. Gpt-4 technical report, 2023. 1 +[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5 +[48] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5606-5611, 2023. 2, 4, 5 +[49] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 2 +[50] Zhende Song, Chenchen Wang, Jiamu Sheng, Chi Zhang, Gang Yu, Jiayuan Fan, and Tao Chen. Moviellm: Enhancing long video understanding with ai-generated movies. arXiv preprint arXiv:2403.01422, 2024. 1 +[51] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4 +[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. + +Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1,3,4 +[53] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5 +[54] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 2, 5 +[55] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2 +[56] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[57] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2, 3 +[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2 +[59] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 2 +[60] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao Chen. Shapept: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618, 2023. 1 +[61] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2 +[62] Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 1 +[63] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 2, 3, 4, 5 +[64] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2 +[65] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 2 +[66] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Towards explainable 3d grounded visual question answering: + +A new benchmark and strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2 +[67] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 5 +[68] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension, 2023. 3 +[69] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2, 5, 6 \ No newline at end of file diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/images.zip b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5eb7bff3afd885fe8dbd38c02f8988a40b3b7044 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15ebdde28910a4cc4fa74bc97390a18937b3623350ef7bd6b7190ee95af07bb2 +size 621721 diff --git a/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/layout.json b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b7a46944b638bb58a7b2261f95e5c70c25f1494 --- /dev/null +++ b/2024/LL3DA_ Visual Interactive Instruction Tuning for Omni-3D Understanding Reasoning and Planning/layout.json @@ -0,0 +1,9619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 103, + 455, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 103, + 455, + 140 + ], + "spans": [ + { + "bbox": [ + 139, + 103, + 455, + 140 + ], + "type": "text", + "content": "LL3DA: Visual Interactive Instruction Tuning for Omni-3D Understanding, Reasoning, and Planning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "spans": [ + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "content": "Sijin Chen" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "content": " Xin Chen" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "content": " Chi Zhang" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "content": " Mingsheng Li" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "text", + "content": " Gang YU" + }, + { + "bbox": [ + 117, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "spans": [ + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "text", + "content": "Hao Fei" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "text", + "content": " Hongyuan Zhu" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "text", + "content": " Jiayuan Fan" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "text", + "content": " Tao Chen" + }, + { + "bbox": [ + 154, + 175, + 438, + 189 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "spans": [ + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "text", + "content": "Fudan University " + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "text", + "content": "Tencent PCG " + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 147, + 190, + 447, + 204 + ], + "type": "text", + "content": "National University of Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 204, + 507, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 204, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 86, + 204, + 507, + 217 + ], + "type": "text", + "content": "4Institute for Infocomm Research (I²R) & Centre for Frontier AI Research (CFAR), A*STAR, Singapore" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 220, + 384, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 220, + 384, + 231 + ], + "spans": [ + { + "bbox": [ + 209, + 220, + 384, + 231 + ], + "type": "text", + "content": "https://github.com/Open3DA/LL3DA" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 221, + 233, + 370, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 233, + 370, + 245 + ], + "spans": [ + { + "bbox": [ + 221, + 233, + 370, + 245 + ], + "type": "text", + "content": "* project lead † corresponding author" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 272, + 542, + 471 + ], + "blocks": [ + { + "bbox": [ + 50, + 272, + 542, + 471 + ], + "lines": [ + { + "bbox": [ + 50, + 272, + 542, + 471 + ], + "spans": [ + { + "bbox": [ + 50, + 272, + 542, + 471 + ], + "type": "image", + "image_path": "5943c9dff48a6fa1c0ca0e1f85368bc41d1194b3b08c82a1d0f10a40fbd56f44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 481, + 546, + 515 + ], + "lines": [ + { + "bbox": [ + 46, + 481, + 546, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 546, + 515 + ], + "type": "text", + "content": "Figure 1. We propose LL3DA, a Large Language 3D Assistant that demonstrates mighty instruction-following capacities in understanding, reasoning, and planning within complex 3D environments. LL3DA takes both the textual instructions and potential visual interactions into consideration to help remove ambiguities when addressing various tasks in diverse and complex 3D scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 525, + 192, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 525, + 192, + 537 + ], + "spans": [ + { + "bbox": [ + 143, + 525, + 192, + 537 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Recent progress in Large Multimodal Models (LMM) has opened up great possibilities for various applications in the field of human-machine interactions. However, developing LMMs that can comprehend, reason, and plan in complex and diverse 3D environments remains a challenging topic, especially considering the demand for understanding permutation-invariant point cloud representations of the 3D scene. Existing works seek help from multi-view images by projecting 2D features to 3D space, which inevitably leads to huge computational overhead and performance degradation. In this paper, we present LL3DA, a Large Language 3D Assistant that takes point cloud as the direct input and responds to both text instructions and visual interactions. The additional visual interaction enables LMMs to better" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 525, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 547, + 586 + ], + "type": "text", + "content": "comprehend human interactions with the 3D environment and further remove the ambiguities within plain texts. Experiments show that LL3DA achieves remarkable results and surpasses various 3D vision-language models on both 3D Dense Captioning and 3D Question Answering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "The recent surge in Large Language Model (LLM) families [14, 31, 46, 52, 63] opens up great opportunities for addressing various machine learning tasks in a generalized way [30, 32, 40, 60]. During this LLM carnival, researchers are also seeking generalized LLM solutions for various vision language tasks [37, 50, 62]. Among these, LLM-based 3D scene understanding is a valuable topic that would benefit the development of autonomous driving [9, 23] and em" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26428" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "bodied AI agents [21, 49]. However, it is also challenging given 1) the diversity and complication of 3D environments and 2) the demands for understanding sparse 3D points." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 288, + 300 + ], + "type": "text", + "content": "Prior works have made initial success in various 3D vision and language tasks. The majority of these research build 3D specialists to solve one specific down-stream task, including 3D Question Answering (3D-QA) [2, 43], 3D Visual Grounding (3D-VG) [7, 26, 55], and 3D Dense Captioning (3D-DC) [10-12]. There are also several works [4, 13, 34, 70] study the mutual promotion of different 3D vision and language tasks with shared structure modelling relations among objects. Recently, researchers have also introduced LLMs for general purpose 3D understanding, where Point-Bind and Point-LLMs [24, 57] mainly focus on the understanding of 3D objects. Concurrently, 3D-LLM [29] proposes an LLM-driven solution that aggregates multi-view features for 3D perception, presenting mighty capacities in understanding 3D object and scenes and following text instructions produced by human." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 301, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 420 + ], + "type": "text", + "content": "Though these methods have achieved remarkable success in addressing various challenges in understanding the 3D world with natural language, there are certain limitations. With limited supervision, 3D specialists could hardly scale-up for better performance, while the joint pre-training still requires separate heads for specific tasks. Extracting multi-view features results in huge computational overhead and ignores the essential geometry and depth information. Additionally, plain texts often lead to ambiguities especially in cluttered and complex 3D environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 589 + ], + "type": "text", + "content": "To address the above issues, we propose LL3DA, a Large Language 3D Assistant that responds to both textual and visual interactions from human, with the ability to understand, reason, and plan in complex 3D environments (Fig. 1). We adopt a multi-modal transformer to aggregate information from textual instructions, visual prompts, and the 3D scene into a fixed length of learnable querying tokens via the attention mechanism. The querying tokens are then projected and used as the prefix for the textual instructions, serving as the input to a pre-trained and frozen LLM. This design not only helps to address the contradiction between the permutation-invariant 3D scene representations and the LLM embedding space, but also extracts interaction-aware 3D scene representations for efficient instruction following." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 685 + ], + "type": "text", + "content": "We conduct extensive experiments to explore the capacities of LL3DA in understanding, reasoning, and planning within complex and diverse 3D environments. Our model achieves state-of-the-art results on two widely used datasets for 3D Dense Captioning [1, 7], and 3D Question Answering [2]. Additionally, by introducing additional visual interactions, our method could further remove the ambiguities within the vague textual instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 686, + 234, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 686, + 234, + 697 + ], + "spans": [ + { + "bbox": [ + 59, + 686, + 234, + 697 + ], + "type": "text", + "content": "To summarize, our key contributions lie in:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "- We present a LLM-based solution for understanding, rea" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 521, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 521, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 521, + 83 + ], + "type": "text", + "content": "soning, and planning in complex 3D environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 84, + 545, + 144 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 306, + 84, + 545, + 119 + ], + "type": "text", + "content": "- Our model takes both the textual instructions and visual interactions as inputs, and extracts interaction-aware features for effective instruction-following." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "type": "text", + "content": "- Extensive experiments show that our method surpasses various state-of-the-art 3D vision language models." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 153, + 392, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 153, + 392, + 166 + ], + "spans": [ + { + "bbox": [ + 306, + 153, + 392, + 166 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 174, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 592 + ], + "type": "text", + "content": "3D Vision and Language alignment, pre-training, and understanding [5, 7, 20, 70] cover tasks requiring a model to adopt its understanding towards a complex 3D scene answering to, or answering with natural language. Among those, 3D Dense Captioning (3D-DC) [10, 12, 54] expects a model to translate an input 3D scene into a set of instance coordinates and natural language descriptions. Existing methods could be categorized into \"detect-then-describe\" models [4, 12, 54] and the \"set-to-set\" prediction approaches [10, 11]. The former builds explicit relations on the instance coordinate estimations, while the latter directly learns the locations and descriptions for instances from the input 3D scene. 3D Visual Grounding (3D-VG) [1, 7, 55] demands a model to respond the natural language queries with the instance coordinates in the 3D scene. The mainstream of existing methods [4, 65, 70] address 3D-VG via selecting a candidate from a 3D detector's prediction. 3D Question Answering (3D-QA) [2, 43, 59, 66] requires a model to answer the questions with natural language based on the input 3D scene. The majority of existing methods [2, 18, 48] directly select the desired response from a given answer set. Researchers have also studied the mutual promotion of various 3D vision language tasks via training their shareable architectures simultaneously on multiple tasks [4, 13, 34, 70]. UniT3D [13] and 3DJCG [4] focus on the joint promotion between 3D-DC and 3D-VG in the relation modelling, while 3D-VLP [34] further includes 3D-QA. Recently, 3D-LLM [29] introduces a family of LLM-driven 3D generalists that could handle diverse textual instructions with reconstructed 3D features from multi-view images [28]. In this paper, we present LL3DA, an LLM solution that directly extracts features from the 3D scene, and handles both visual prompts and textual instructions to diversify the possible interactions human could make with the complex 3D environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Large Multimodal Models (LMM). Along with the rapid development of Large Language Models (LLM) [15, 63], researchers have made great recent efforts adapting LLMs to visual understanding and reasoning tasks [25, 38, 56, 61]. Some methods project or compress global image features as the prefix for text instructions [36, 40, 58, 69], while others extract ROI features as LLM tokens for region-oriented instruction reasoning [6, 64]. Meanwhile, InstructBLIP [17] proposes to extract textual instruction-aware visual features, and has achieved remarkable success in addressing complex" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26429" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 542, + 257 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 542, + 257 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 542, + 257 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 542, + 257 + ], + "type": "image", + "image_path": "6179a1c6621ee1b02340e9cd76f3c56c9cc2bbd5cbc2f53b30c3188e4be101a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 309 + ], + "type": "text", + "content": "Figure 2. Overview of the Proposed Approach. (a) The overall pipeline of our proposed LL3DA first extracts interaction-aware 3D scene embeddings, which are later projected to the prefix of textual instructions as the input of a frozen LLM. (b) The detailed design of the Interactor3D, which aggregates visual prompts, textual instructions, and 3D scene embeddings into a fixed length querying tokens. (c) The prompt encoder encodes the user clicks and box coordinates with the positional embeddings and ROI features, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 289, + 426 + ], + "type": "text", + "content": "and unseen instructions. Concurrently, researchers have also made great attempts solving various 3D tasks using LLMs. Notably, [24, 42, 57, 68] demonstrate remarkable success in understanding and reasoning about 3D objects. In this paper, we present an LLM-driven solution that could handle both interactions in forms of visual prompts and text instructions. We also propose to extract interaction-aware 3D scene representations for better instruction following." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 436, + 130, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 130, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 130, + 450 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 529 + ], + "type": "text", + "content": "To build a general purpose agent that could handle both visual and textual interactions with complex 3D environments, we propose LL3DA, an LLM driven auto-regressive approach to 3D vision language tasks. In this section, we first introduce the problem formatting in Sec. 3.1. Next, we introduce our model designs in details (Sec. 3.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 536, + 167, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 167, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 167, + 550 + ], + "type": "text", + "content": "3.1. Problem Formatting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": "Model I/O. As shown in Fig. 2 (a), the input of our model consists of a 3D scene represented by a set of points " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "PC" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": ", the textual instruction " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": ", and potential visual interactions " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_v" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": " that serve as supplementary spatial identifiers. Here, point cloud " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "PC = [p_{in}, f_{in}] \\in \\mathbb{R}^{N \\times (3 + F)}" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "p_{in} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "inline_equation", + "content": "f_{in} \\in \\mathbb{R}^{N \\times F}" + }, + { + "bbox": [ + 46, + 556, + 287, + 664 + ], + "type": "text", + "content": " are the point coordinates and the additional point features, including color, normal, and height. The output of our model is free-form natural language, part of whom could be interpreted into 3D coordinates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Instruction Formatting. Following existing LMMs [57], we begin the textual instructions " + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": " with the “human:” identifier, and ask the model to generate responses after the “assistant:” identifier. This endows the model with the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 352 + ], + "type": "text", + "content": "ability to distinguish information from the context and further engage in multi-turn conversations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "content": "Coordinate Representations. To provide LLMs with the capacity to perceive and respond with 3D coordinates, we convert the 3D points and 3D bounding boxes to plain texts. Specifically, a point is represented by “" + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{loc} \\rangle x, y, z \\langle \\mathrm{loc} \\rangle" + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "content": "”, and a bounding box is represented by its center point and size, i.e. “" + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{obj} \\rangle c_x, c_y, c_z, w, h, l \\langle \\mathrm{obj} \\rangle" + }, + { + "bbox": [ + 304, + 354, + 546, + 497 + ], + "type": "text", + "content": "”. Here, all the numerical data is discretized into unsigned integers within a range of [0, 255] with respect to the boundary of the input 3D scene. This design could naturally fit in the vocabulary of existing pre-trained LLMs [52, 63]. Without the introduction of any additional learnable tokens, we could save the effort of tuning the LLMs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 502, + 394, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 394, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 394, + 516 + ], + "type": "text", + "content": "3.2. Model Design" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": "As shown in Fig. 2 (a), our model first aggregates a fixed-length scene embeddings through the Interactor3D, which takes the visual prompts, the textual instructions, and the 3D scene as the input. Next, the aggregated 3D scene embeddings are projected to be the prefix of text instructions, and serve as the inputs of a frozen LLM. The detailed design of Interactor3D is shown in Fig. 2 (b), which consists of a frozen 3D scene encoder " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{3D}" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": ", a visual prompt encoder, and a multi-modal transformer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": "Scene Encoder. We adopt the masked transformer encoder pre-trained on ScanNet detection [10] as the scene encoder, " + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^{3D}" + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": ". The scene encoder takes " + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "inline_equation", + "content": "PC" + }, + { + "bbox": [ + 304, + 630, + 545, + 677 + ], + "type": "text", + "content": " as its input, and outputs the 3D scene embeddings:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 331, + 681, + 545, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 681, + 545, + 697 + ], + "spans": [ + { + "bbox": [ + 331, + 681, + 545, + 697 + ], + "type": "interline_equation", + "content": "f _ {e n c} = \\mathcal {E} ^ {3 D} (P C) = \\mathcal {E} ^ {3 D} \\left(p _ {i n}; f _ {i n}\\right) \\in \\mathbb {R} ^ {M \\times d}. \\tag {1}", + "image_path": "999c2a7d64c34826f729d2bd6b81161bca06b5e3dc3b2499b63246d0d79d24dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{enc}" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "-dimensioned features for " + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": " points" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26430" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "uniformly down-sampled from the input 3D scene through the Farthest Point Sampling (FPS) algorithm. In practice, we choose to keep the scene encoder frozen to save the memory cost during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "text", + "content": "Visual Prompt Encoder. We mainly take two common types of visual interactions into consideration, user clicks and 3D box annotations [35]. Each user click is first normalized within a range of " + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "text", + "content": " by the size of the input 3D scene " + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{click}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "text", + "content": ". Then, we encode " + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{click}}" + }, + { + "bbox": [ + 46, + 121, + 286, + 193 + ], + "type": "text", + "content": " with the 3D Fourier positional embeddings [51] function:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 287, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 287, + 216 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right) = \\left[ \\sin \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right); \\cos \\left(2 \\pi p _ {\\text {c l i c k}} \\cdot B\\right) \\right]. \\right. \\tag {2}", + "image_path": "82efc7a71f1e91ce429cc23bbfbb59a397c51861896fc490951066326c2220f1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{3 \\times (d/2)}" + }, + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "text", + "content": " is a learnable matrix. The box annotation is represented by the ROI feature " + }, + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{box}} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 224, + 287, + 284 + ], + "type": "text", + "content": " extracted by a pre-trained 3D object detector [10]. The two types of the visual prompts are then projected with separate and identical Feed Forward Networks (FFN)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 293, + 287, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 293, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 293, + 287, + 312 + ], + "type": "interline_equation", + "content": "f _ {\\text {c l i c k}} = F F N _ {\\text {c l i c k}} \\left(\\operatorname {p o s} \\left(p _ {\\text {c l i c k}}\\right)\\right) \\tag {3}", + "image_path": "16609f9bc71bd63bf1aead4c3cccd1c5ea7f3dff55cb3a1c4827821291cc4894.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 308, + 199, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 308, + 199, + 321 + ], + "spans": [ + { + "bbox": [ + 109, + 308, + 199, + 321 + ], + "type": "interline_equation", + "content": "f _ {\\text {b o x}} = F F N _ {\\text {b o x}} \\left(f _ {\\text {b o x}}\\right)", + "image_path": "2929e4f65ad583eab71e629825276f92d028a0703af509fd4ef9f1b4685feea6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 330, + 286, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 286, + 342 + ], + "type": "text", + "content": "In practice, we represent each visual prompt with 8 tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "content": "Multi-Modal Transformer (MMT) serves as a role to 1) address the contradiction between the permutation-invariant 3D scene embeddings and position-sensitive causal LLMs, 2) bridge the gap between frozen unimodal experts, and 3) fill the needs for interaction-aware feature extraction. Inspired by the Q-Former architecture [17, 36], MMT aggregates the visual information within a fixed number of 32 learnable querying tokens. In each layer, the queries interact with the encoded visual prompts " + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "inline_equation", + "content": "[f_{\\mathrm{click}}; f_{\\mathrm{box}}]" + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "content": " and the textual instructions " + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "content": " through a shared self-attention. Then, we allow the learnable querying tokens and the visual prompts to interact with the task-agnostic 3D scene embeddings " + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "inline_equation", + "content": "f_{enc}" + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "content": " via cross-attention. The output of MMT is 32 queries written as " + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{32 \\times 768}" + }, + { + "bbox": [ + 46, + 342, + 287, + 570 + ], + "type": "text", + "content": ", which are finally projected to the word embedding space of LLMs through a simple linear projector. In practice, we notice that initializing Q-Former with pre-trained BERT [19, 36] weights will lead to repetitive outputs, thus we only choose to initialize the pre-trained word and position embeddings from BERT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": "LLM. We consider the decoder-only generative pre-trained transformers [52, 63] as our large language model backbone. The decoder-only LLMs are sensitive to the input orders because of the position embeddings and the causal attention mask. The parameters and the embedding layers of the LLMs are kept frozen to save memory cost. During inference, we generate the responses via searching for the optimal sequence " + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "inline_equation", + "content": "s^*" + }, + { + "bbox": [ + 46, + 571, + 287, + 666 + ], + "type": "text", + "content": " that satisfies:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 100, + 676, + 287, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 676, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 100, + 676, + 287, + 693 + ], + "type": "interline_equation", + "content": "s ^ {*} = \\arg \\max _ {s} P (s | P C, \\mathcal {I} _ {t}, \\mathcal {I} _ {v}). \\tag {4}", + "image_path": "0471cab473ca124b84dae6d3851a95184301b21c3c826f14793c8905141fe420.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 266, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 266, + 713 + ], + "type": "text", + "content": "In practice, we use beam search with a beam size of 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 71, + 486, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 486, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 486, + 85 + ], + "type": "text", + "content": "4. Multi-modal Instruction Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 545, + 163 + ], + "type": "text", + "content": "A general purpose 3D agent is able to address various tasks simultaneously in complex 3D scenes. Apart from introducing proper training data, it is important to guide the model to generate the desired outputs with instructions. Therefore, Sec. 4.1 will first introduce how we identify each task. After that, Sec. 4.2 will present details for the training objective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 170, + 437, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 437, + 182 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 437, + 182 + ], + "type": "text", + "content": "4.1. Tasks and Instructions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 212 + ], + "type": "text", + "content": "As introduced in Sec. 3.1, LL3DA generates text responses auto-regressively after the “```assistant:” identifier." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 297 + ], + "type": "text", + "content": "3D Dense Captioning requires the localization and description of instances in diverse 3D environments. We adopt either user clicks and box annotations as the visual prompt to identify the object to be described. Additionally, we design two types of textual instructions that ask the model to either \"describe\" or \"describe and localize\" the object, which diversifies the tasks, and leads to better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 298, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 406 + ], + "type": "text", + "content": "3D Question Answering requires the model to generate response to the questions based on the global knowledge of a 3D scene. To help the model better understand the 3D environment, we also design two types of textual instructions that ask the model to either \"answer\" or \"answer and localize the related objects\". The latter serves as an auxiliary task widely adopted in various 3D-QA methods [2, 48]. To diversify the tasks during training, we randomly include additional clicks on the objects related to the questions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 442 + ], + "type": "text", + "content": "Scene Description requires the model to translate its global knowledge of the 3D scene into natural language descriptions, thus we simply ask the \"describe\" this 3D scene." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 545, + 491 + ], + "type": "text", + "content": "Embodied Conversation and Planning could be treated as multi-turn conversations, where we use “human” and “assistant” as identifiers to distinguish the source of information as introduced in Sec. 3.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 498, + 466, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 466, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 466, + 512 + ], + "type": "text", + "content": "4.2. Instruction Following Tuning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 517, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 517, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 545, + 553 + ], + "type": "text", + "content": "During training, for tasks requiring additional visual interactions, i.e. 3D-DC and 3D-QA, we randomly choose between clicks or boxes as means of object identification." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": "Training Objective. Our training objective is to optimize the trainable parameters " + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": ", so as to maximize the likelihood of the target response sequence " + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": " given the input point cloud " + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "PC" + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": ", and the human interactions " + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_v" + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_t" + }, + { + "bbox": [ + 305, + 554, + 545, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 610, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 610, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 353, + 610, + 545, + 628 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\arg \\max _ {\\theta} P (s | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta). \\tag {5}", + "image_path": "789b5463f1cb02fcc28d8089a5434a3361d60b6df9f1617cef82d9716640b041.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "content": "In practice, this is accomplished by adopting the token-wise cross-entropy loss that trains the model to predict the " + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "content": "th token " + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "inline_equation", + "content": "s_{[i]}" + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "content": " given the previous " + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "inline_equation", + "content": "(i - 1)" + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "content": " tokens, " + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "inline_equation", + "content": "s_{[1,\\dots ,i - 1]}" + }, + { + "bbox": [ + 305, + 636, + 545, + 673 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 682, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 682, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 682, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = - \\sum_ {i = 1} ^ {| s |} \\log P (s _ {[ i ]} | P C; \\mathcal {I} _ {v}; \\mathcal {I} _ {t}; \\theta ; s _ {[ 1, \\dots , i - 1 ]}). \\tag {6}", + "image_path": "86daabd7e89e85f5b776b2162a1814db34b4aac004f67ee8196576381bc8a86c.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26431" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 113, + 545, + 237 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons for 3D Dense Captioning on ScanRefer[7] and Nr3D[1]. For fair comparison, we list methods that are trained under the standard per-word cross-entropy loss without additional 3D scenes. We use the box estimations from Vote2Cap-DETR to simulate the box annotations as the visual prompts. Our proposed LL3DA surpasses previous 3D specialists on both datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 113, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 49, + 113, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 113, + 545, + 237 + ], + "type": "table", + "html": "
MethodScanReferNr3D
C@0.25↑B-4@0.25↑M@0.25↑R@0.25↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
Scan2Cap[12]56.8234.1826.2955.2739.0823.3221.9744.7827.4717.2421.8049.06
MORE[33]62.9136.2526.7556.3340.9422.9321.6644.42----
SpaCap3D[54]----44.0225.2622.3345.3633.7119.9222.6150.50
REMAN[44]62.0136.3726.7656.2545.0026.3122.6746.9634.8120.3723.0150.99
D3Net[8]----46.0730.2924.3551.6733.8520.7023.1353.38
Contextual[67]----46.1125.4722.6445.9635.2620.4222.7750.78
UniT3D[13]----46.6927.2221.9145.98----
3DJCG[4]64.7040.1727.6659.2349.4831.0324.2250.8038.0622.8223.7752.99
3D-VLP[34]70.7341.0328.1459.7254.9432.3124.8351.51----
3D-VisTA*[70]----61.6034.1026.8055.00----
Vote2Cap-DETR[10]71.4539.3428.2559.3361.8134.4626.2254.4043.8426.6825.4154.43
LL3DA (Ours)74.1741.4127.7659.5365.1936.7925.9755.0651.1828.7525.9156.61
", + "image_path": "7a5e0fb81edaf65f47e466d5f2f4115c3e750eb243ad010bacb205c5e5cc50a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 87, + 290, + 507, + 378 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "type": "text", + "content": "Table 2. Quantitative Comparisons for 3D Question Answering on ScanQA[2]. We categorize previous works into classification based (\"CLS\") and generation based (\"GEN\") methods. The results from 3D-LLM* come from their fine-tuned version. LL3DA out-performs previous methods on the validation set and two test sets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 290, + 507, + 378 + ], + "lines": [ + { + "bbox": [ + 87, + 290, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 87, + 290, + 507, + 378 + ], + "type": "table", + "html": "
MethodAnswer TypeValidationTest w/ objectTest w/o object
C↑B-4↑M↑R↑C↑B-4↑M↑R↑C↑B-4↑M↑R↑
ScanQA[2]64.8610.0813.1433.3367.2912.0413.5534.3460.2410.7512.5931.09
Clip-Guided[48]----69.5314.6413.9435.1562.8311.7313.2832.41
Multi-CLIP[18]CLS----68.7012.6513.9735.4663.2012.8713.3632.61
3D-VLP[34]66.9711.1513.5334.5170.1811.2314.1635.9763.4015.8413.1331.79
3D-VisTA[70]----68.6010.5013.8035.5055.708.7011.6929.60
3D-LLM*[29]GEN69.4012.0014.5035.7069.6011.6014.9035.30----
LL3DA (Ours)76.7913.5315.8837.3178.1613.9716.3838.1570.2912.1914.8535.17
", + "image_path": "12801e52c8dfca51bfc9fca8d252c20c675037c98d93905f6e1938bb23ff08f0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 397, + 277, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 277, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 277, + 411 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 397, + 277, + 411 + ], + "type": "inline_equation", + "content": "|s|" + }, + { + "bbox": [ + 47, + 397, + 277, + 411 + ], + "type": "text", + "content": " is the number of tokens in the desired response." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 422, + 128, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 128, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 128, + 436 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 443, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 551 + ], + "type": "text", + "content": "To test the capacities of LL3DA, we provide numerous evaluations. To begin with, we introduce the datasets, metrics, and implementation details (Sec. 5.1). Then, we compare how our model understands and reasons in complex 3D environments with previous 3D specialists on 3D Dense Captioning and 3D Question Answering (Sec. 5.2). Additionally, we conduct quantitative ablation studies on the model design and training strategy (Sec. 5.3). Finally, Sec. 5.4 showcases several qualitative results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 285, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 285, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 285, + 572 + ], + "type": "text", + "content": "5.1. Datasets, Metrics and Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 579, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 699 + ], + "type": "text", + "content": "Datasets. In this paper, we experiment with 3D data from ScanNet [16], a 3D dataset covering 1,201 and 312 diverse and complex indoor 3D scenes for training and validation. The language annotations used in this study are sourced from ScanRefer [7], Nr3D [1], ScanQA [2], and the ScanNet subset of 3D-LLM [29]. This combination covers a variety of tasks, including instance and scene descriptions, conversations, embodied planning and question answering. Please refer to the supplementary materials for more details on the statistics of data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Metrics. Here, we adopt C, B-4, M, R as abbreviations for" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 397, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 422 + ], + "type": "text", + "content": "CiDEr [53], BLEU-4 [47],METEOR [3],and Rouge-L [39] to evaluate the quality of the generated textual responses." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "content": "Implementation Details. Following previous works on 3D vision language tasks [10, 12], we randomly sample " + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "inline_equation", + "content": "40\\mathrm{k}" + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "content": " points from each 3D scene as the 3D input. We adopt the pre-trained OPT-1.3B [63] as our causal LLM backbone, which is frozen and loaded in float16 to save memory cost. We adopt the AdamW [41] optimizer with a weight decay of 0.1 and a learning rate decaying from " + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "content": " with a cosine annealing scheduler for about " + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "inline_equation", + "content": "100\\mathrm{k}" + }, + { + "bbox": [ + 304, + 422, + 546, + 542 + ], + "type": "text", + "content": " iterations. For all the training tasks, we train with no more than eight Nvidia RTX3090 (24G) GPUs within a day." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 550, + 491, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 550, + 491, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 550, + 491, + 563 + ], + "type": "text", + "content": "5.2. Comparison with SoTA Specialists" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "content": "We evaluate the model's capacity to understand and reason in 3D environments via 3D-DC and 3D-QA. For each evaluation task, we fine-tune the trainable parameters in our model on each task for " + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\sim 30\\mathrm{k}" + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "content": " iterations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "3D Dense Captioning demands a model to localize and describe any instance in a 3D scene. We benchmarks state-of-the-art methods on the widely-used ScanRefer [7] and " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Nr3D}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " [1] dataset in Tab. 1 under the " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "m@k\\mathrm{IoU}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " metric [12]. Here, " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "m \\in \\{\\mathrm{C}, \\mathrm{B-4}, \\mathrm{M}, \\mathrm{R}\\}" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " score of a caption is set to 0 if the IoU between the predicted box and the object is less than the given threshold " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": ". Following existing works [10, 12], we consider " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{C} @ 0.25" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{C} @ 0.5" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " as the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26432" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "main metric for ScanRefer, and C@0.5 for Nr3D. Among the listed methods, UniT3D [13], 3DJCG [4], and 3D-VLP [34] are pre-trained on multiple 3D vision and language tasks annotated on ScanNet scenes. Additionally, UniT3D [13] adopts off-the-shelf image caption models [45] and multi-view images to generate additional instance-captions for pre-training. It is worth mentioning that we compare the results with the 3D-VisTA [70] model that is not trained on additional 3D scenes. To evaluate our model, we adopt the box predictions produced by Vote2Cap-DETR [10] as the visual prompt. Results show that our method consistently outperforms existing methods on both datasets. For example, our method achieves " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "65.19\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " C@0.5 on ScanRefer and " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "51.18\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " C@0.5 on Nr3D, which is " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "(+3.38\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "inline_equation", + "content": "+7.34\\%)" + }, + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": " higher than the current state-of-the-art 3D vision and language model, Vote2Cap-DETR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 288, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 288, + 434 + ], + "type": "text", + "content": "3D Question Answering requires a model to generate responses to the natural language queries questioning towards an 3D scene. We benchmark state-of-the-art methods on the ScanQA [2] validation set as well as two test benchmarks in Tab. 2, and consider CiDEr as the main metric. The majority of the listed methods are based on classification (marked \"CLS\"), i.e., selecting responses from a predefined answer set. Meanwhile, 3D-LLM [29] tries to address 3D-QA via auto-regressive text generation (marked \"GEN\"), and we list their fine-tuned version for comparison. Results show that our method consistently outperforms existing methods on all the evaluation sets, and surpasses the generation based method, 3D-LLM, by a large margin " + }, + { + "bbox": [ + 46, + 266, + 288, + 434 + ], + "type": "inline_equation", + "content": "(+7.39\\%" + }, + { + "bbox": [ + 46, + 266, + 288, + 434 + ], + "type": "text", + "content": " CiDEr score on the validation set)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 442, + 149, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 149, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 149, + 453 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 498 + ], + "type": "text", + "content": "In this section, we provide ablation studies on model designs and training strategies. We evaluate on ScanRefer and ScanQA to quantize the effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 52, + 510, + 165, + 574 + ], + "blocks": [ + { + "bbox": [ + 52, + 510, + 165, + 574 + ], + "lines": [ + { + "bbox": [ + 52, + 510, + 165, + 574 + ], + "spans": [ + { + "bbox": [ + 52, + 510, + 165, + 574 + ], + "type": "image", + "image_path": "516496b47dc1454d026b87181128be058b7e5996ba94be0f5b9ed19d9d9e7d73.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 577, + 149, + 587 + ], + "lines": [ + { + "bbox": [ + 67, + 577, + 149, + 587 + ], + "spans": [ + { + "bbox": [ + 67, + 577, + 149, + 587 + ], + "type": "text", + "content": "(a) Early Fusion (ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 510, + 284, + 574 + ], + "blocks": [ + { + "bbox": [ + 168, + 510, + 284, + 574 + ], + "lines": [ + { + "bbox": [ + 168, + 510, + 284, + 574 + ], + "spans": [ + { + "bbox": [ + 168, + 510, + 284, + 574 + ], + "type": "image", + "image_path": "79369773bc285f5d43fbf7e72cee7a3ecf91909680d3994ac0db3976bed889ee.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 577, + 261, + 587 + ], + "lines": [ + { + "bbox": [ + 192, + 577, + 261, + 587 + ], + "spans": [ + { + "bbox": [ + 192, + 577, + 261, + 587 + ], + "type": "text", + "content": "(b) Direct Injection" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 595, + 287, + 651 + ], + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 651 + ], + "type": "text", + "content": "Figure 3. Two Different Ways to Encode Visual Prompts. Our proposed method (a) adopts a unified transformer to aggregate features from all kinds of interactions, while (b) directly concatenates the visual prompts to the scene embeddings. Tab. 3 shows that early fusion leads to a better performance." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Effectiveness of the Q-Former Design. We list two ways to process the visual prompts in Fig. 3. Here, Fig. 3 (a) is our proposed method that adopts a unified transformer to aggregate information from both text instructions and visual" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "prompts, while Fig. 3 (b) is the \"direct injection\" version, which only extract instruction-aware 3D feature with visual prompts concatenated after the scene embeddings. We train both models from scratch and evaluate their performance on ScanRefer 3D Dense Captioning. The results (Fig. 3) show that the method we use (Fig. 3 (a)) could better capture feature related to the visual prompts, leading to better instance caption generation performance " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "(+3.45\\% \\mathrm{C}@\\mathrm{0.5})" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 230, + 544, + 271 + ], + "blocks": [ + { + "bbox": [ + 305, + 177, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 305, + 177, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 305, + 177, + 547, + 221 + ], + "type": "text", + "content": "Table 3. Effectiveness of Q-Former Design on ScanRefer[7]. We design two different ways of utilizing visual prompts. The \"early fusion\" enables direct interaction with the 3D scene, thus it achieves a better performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 230, + 544, + 271 + ], + "lines": [ + { + "bbox": [ + 309, + 230, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 309, + 230, + 544, + 271 + ], + "type": "table", + "html": "
Visual PromptC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
direct59.3933.2725.1953.39
ours62.8435.8125.8154.45
", + "image_path": "be5c6fd21511961e72ef1ec34d73d51ca6656450c969fedd76b19817c9961e4d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 437 + ], + "type": "text", + "content": "Instructions as Auxiliary Tasks for 3D Dense Captioning. We have introduced two types of task instructions in Sec. 4.1 for 3D-DC, i.e. the \"describe\"-only instructions and \"detect and localize\" instructions. Additionally, we have introduced two types of visual prompts (Fig. 2 & Sec. 4.2). In this study, we show how they affect the performance when serving as auxiliary tasks for 3D-DC by evaluating on ScanRefer in Tab. 4. All the methods listed are trained from scratch. In Tab. 4, \"Aux.Loc\" identifies whether we train the model with the \"detect and localize\" instructions, and \"Clicks\" identifies whether we train the model with clicks as additional visual prompts. Results show that they are both good auxiliary tasks for 3D-DC." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 510, + 544, + 570 + ], + "blocks": [ + { + "bbox": [ + 305, + 447, + 545, + 502 + ], + "lines": [ + { + "bbox": [ + 305, + 447, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 447, + 545, + 502 + ], + "type": "text", + "content": "Table 4. Effectiveness of Instructions as 3D Dense Captioning Auxiliary Tasks. We train the models from scratch and evaluate on ScanRefer[7]. \"Aux.Loc\" identifies whether we train with the \"describe and localize\" instructions. \"Clicks\" identifies whether we train with clicks as additional visual prompts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 510, + 544, + 570 + ], + "lines": [ + { + "bbox": [ + 309, + 510, + 544, + 570 + ], + "spans": [ + { + "bbox": [ + 309, + 510, + 544, + 570 + ], + "type": "table", + "html": "
Aux.LocClicksC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
--60.8534.0925.5353.48
✓-61.8134.1525.4953.83
-✓62.2034.2625.6753.87
✓✓62.8435.8125.8154.45
", + "image_path": "4c36c491107a0d1d24b74a3ea5c93ee542462cf0843135c3d80d6be372a5ab9c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 545, + 676 + ], + "type": "text", + "content": "Instructions as Auxiliary Tasks for 3D Question Answering. We have made a similar study to analyze how adopting additional \"answer and localize\" instructions and visual prompts improves 3D-QA on ScanQA [2] as auxiliary tasks in Tab. 6. We do not use any visual interactions during inference. Results show that the additional textual instructions and visual prompts improve the task diversity and further improve the performance for 3D-QA." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Performance as a Generalist. To test whether LL3DA can distinguish different tasks given the textual instructions and visual prompts, we evaluate our model on different tasks" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26433" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 134, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 127 + ], + "type": "text", + "content": "Table 5. Evaluation as a Generalist. The first three rows list the performance of models trained from scratch as experts on each dataset. The results in the following three rows belong to the model fine-tuned from the generalist weights. The last row evaluates the model trained as a generalist. ScanRefer[7] and Nr3D[1] are used to evaluate the dense captioning performance, and ScanQA[2] is used to evaluate the question answering performance. Serving as a generalist, our method can differentiate each task, and produce strong results based on textual instructions and visual prompts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 134, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 134, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 134, + 545, + 224 + ], + "type": "table", + "html": "
MethodScanReferNr3DScanQA
C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C@0.5↑B-4@0.5↑M@0.5↑R@0.5↑C↑B-4↑M↑R↑
ScanRefer(scratch)62.8435.8125.8154.45--------
Nr3D(scratch)----44.9527.6725.6755.79----
ScanQA(scratch)--------74.8013.6815.4036.25
ScanRefer(fine-tuned)65.1936.7925.9755.06--------
Nr3D(fine-tuned)----51.1828.7525.9156.61----
ScanQA(fine-tuned)--------76.7913.5315.8837.31
w/o fine-tuning62.9835.9725.6654.6523.9413.3722.3145.7875.6713.3315.3737.02
", + "image_path": "aa79cca0f99d8a6eaacffb258ea8069e445ce401b9e8209546fd0d03e47293bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 317, + 286, + 367 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 288, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 288, + 308 + ], + "type": "text", + "content": "Table 6. Effectiveness of Interactions as 3D Question Answering Auxiliary Tasks. We train the model from scratch and evaluate all the models from scratch on ScanQA[2] validation set. \"Aux.Loc\" identifies whether we train with the \"answer and localize\" instructions, and \"Visual Prompts\" identifies whether we train with visual prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 317, + 286, + 367 + ], + "lines": [ + { + "bbox": [ + 49, + 317, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 286, + 367 + ], + "type": "table", + "html": "
Aux.LocVisual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
--67.8511.8713.9633.87
✓-72.7313.2714.9035.87
-✓68.0912.5914.2033.71
✓✓74.8013.6815.4036.25
", + "image_path": "c8ec3ac01dd213fac16b6073b74d96fc60c9057cd7480f854908fc394b20b41e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 616 + ], + "type": "text", + "content": "without task-specific fine-tuning in Tab. 5. The first three rows list the performance of LL3DA when trained from scratch on one specific task, while the following three rows represent the fine-tuned models. The last row indicates the direct evaluation of LL3DA. Results show that our model could distinguish 3D-DC and 3D-QA given the text instructions and visual prompts, and achieve strong performance (62.98% C@0.5 on ScanRefer, 75.67% CiDEr on ScanQA). However, the generalist model achieves poor performance on Nr3D [1], which is because we did not try to differentiate between Nr3D and ScanRefer during training in the first place, as ScanRefer and Nr3D are used for the same task. There is also an interesting observation that though we did not differentiate between these two datasets for 3D-DC, the model still tend to achieve high scores on ScanRefer (62.98% C@0.5). We are also excited to see that the weights of the generalist model can serve as a strong initialization for fine-tuning. For example, the fine-tuned model on ScanRefer could achieve 65.19% C@0.5, which is +2.35% higher than the model trained from scratch." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Importance of Textual Instructions. We further conduct study to see whether the text instructions are necessary for 3D-DC in Tab. 7. The first row is our baseline method that directly generates the captions based on visual prompts without any text instructions, and the second row is our method that is trained with the text instructions introduced in Sec. 4.1. Both methods are trained from scratch for fair comparison. We notice that since the LLM is frozen, certain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 243, + 545, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 268 + ], + "type": "text", + "content": "textual instructions are beneficial when generating results in specific domains/tasks." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 331, + 544, + 372 + ], + "blocks": [ + { + "bbox": [ + 304, + 277, + 545, + 323 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 323 + ], + "type": "text", + "content": "Table 7. Effectiveness of Instructions on 3D Dense Captioning. We perform experiments on ScanRefer[7]. The baseline method directly generates the captions given the input 3D scene and visual prompts without any textual instructions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 331, + 544, + 372 + ], + "lines": [ + { + "bbox": [ + 309, + 331, + 544, + 372 + ], + "spans": [ + { + "bbox": [ + 309, + 331, + 544, + 372 + ], + "type": "table", + "html": "
InstructionsC@0.5↑B-4@0.5↑M@0.5↑R@0.5↑
-60.2034.7925.4054.03
✓62.8435.8125.8154.45
", + "image_path": "0ce6505aa640daa425b8491cb470a5bac4afaca2bf05c11ec82caa6e0877e00d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": "Clicks for Better Question Answering. One major challenge of answering questions in complex 3D environments is the vague identification of objects with plain texts. Therefore, we try to click on some of the related objects along with the textual instructions during evaluation, and see how it could affect the generated answers on the ScanQA validation set in Tab. 8. Results show that this technique would remove the ambiguities, and further improve the quality of the answers " + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "inline_equation", + "content": "(+6.12\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 504 + ], + "type": "text", + "content": " C). This illustrates the importance of visual interaction in complex 3D environments." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 567, + 544, + 605 + ], + "blocks": [ + { + "bbox": [ + 304, + 514, + 545, + 559 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 559 + ], + "type": "text", + "content": "Table 8. Test Time Visual Interactions for Question Answering on ScanQA[2]. The model achieves better performance on the question answering when we add visual prompts to some of the related objects along with the text instructions during evaluation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 567, + 544, + 605 + ], + "lines": [ + { + "bbox": [ + 309, + 567, + 544, + 605 + ], + "spans": [ + { + "bbox": [ + 309, + 567, + 544, + 605 + ], + "type": "table", + "html": "
Visual PromptsCiDEr↑BLEU-4↑METEOR↑Rouge-L↑
-76.7913.5315.8837.31
✓82.9111.8016.7439.97
", + "image_path": "0b1798fd69dfe7f55b5679c4c55ac89fa7ef747ebea2dc25184eab9dbad2a556.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 623, + 418, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 418, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 418, + 635 + ], + "type": "text", + "content": "5.4. Qualitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We present several visualization results on different tasks in Fig. 4 to show our model's capacities in understanding, reasoning, and planning in different 3D environments. To prevent repetition when generating long sequences, we combine the top-" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " [22] and top-" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " [27] sampling strategy with " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "k = 50" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "inline_equation", + "content": "p = 0.95" + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26434" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 150, + 76, + 242, + 89 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 76, + 242, + 89 + ], + "spans": [ + { + "bbox": [ + 150, + 76, + 242, + 89 + ], + "type": "text", + "content": "3D Dense Captioning" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 92, + 169, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 92, + 169, + 203 + ], + "lines": [ + { + "bbox": [ + 53, + 92, + 169, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 169, + 203 + ], + "type": "image", + "image_path": "774a3fe5adcf1de1ef3e1b51cf62afdca7d7fd4c7522ada6b691962a1a12cf96.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 179, + 92, + 192, + 110 + ], + "blocks": [ + { + "bbox": [ + 179, + 92, + 192, + 110 + ], + "lines": [ + { + "bbox": [ + 179, + 92, + 192, + 110 + ], + "spans": [ + { + "bbox": [ + 179, + 92, + 192, + 110 + ], + "type": "image", + "image_path": "503c80453692779fabc10ce921f8b4fa6de1850c1b656409b9c8c2091bcbec18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 197, + 92, + 301, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 92, + 301, + 110 + ], + "spans": [ + { + "bbox": [ + 197, + 92, + 301, + 110 + ], + "type": "text", + "content": "Describe this object in the 3D scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 194, + 114, + 312, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 114, + 312, + 131 + ], + "spans": [ + { + "bbox": [ + 194, + 114, + 312, + 131 + ], + "type": "text", + "content": "This is a brown armchair. It is to the right of the couch." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 114, + 339, + 133 + ], + "blocks": [ + { + "bbox": [ + 320, + 114, + 339, + 133 + ], + "lines": [ + { + "bbox": [ + 320, + 114, + 339, + 133 + ], + "spans": [ + { + "bbox": [ + 320, + 114, + 339, + 133 + ], + "type": "image", + "image_path": "f88f3c5b8e779ecbd74d81be36c5088ac2b98e5c2939ccff3d66007b1e628564.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 171, + 140, + 321, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 140, + 321, + 157 + ], + "spans": [ + { + "bbox": [ + 171, + 140, + 321, + 157 + ], + "type": "text", + "content": "SpaCap3D: This is a chair with no arms. \nIt is next to a desk." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 171, + 159, + 328, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 159, + 328, + 177 + ], + "spans": [ + { + "bbox": [ + 171, + 159, + 328, + 177 + ], + "type": "text", + "content": "Vote2Cap-DETR: This is a brown armchair. \nIt is to the right of the table." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 171, + 178, + 329, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 178, + 329, + 205 + ], + "spans": [ + { + "bbox": [ + 171, + 178, + 329, + 205 + ], + "type": "text", + "content": "Ground Truth: The armchair is facing the desk. The armchair is located to the right of the couch, in the middle of the room." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 220, + 246, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 220, + 246, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 220, + 246, + 232 + ], + "type": "text", + "content": "3D Question Answering" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 230, + 175, + 301 + ], + "blocks": [ + { + "bbox": [ + 52, + 230, + 175, + 301 + ], + "lines": [ + { + "bbox": [ + 52, + 230, + 175, + 301 + ], + "spans": [ + { + "bbox": [ + 52, + 230, + 175, + 301 + ], + "type": "image", + "image_path": "128a305b4cf2436f9018718d41e48b6ab6445973196e45310f870c2a2fc447b8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 197, + 245, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 245, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 197, + 245, + 296, + 262 + ], + "type": "text", + "content": "How many chairs are behind the long table?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 220, + 266, + 291, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 266, + 291, + 293 + ], + "spans": [ + { + "bbox": [ + 220, + 266, + 291, + 293 + ], + "type": "text", + "content": "2 ch + Visual Prompts" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 52, + 302, + 176, + 375 + ], + "blocks": [ + { + "bbox": [ + 52, + 302, + 176, + 375 + ], + "lines": [ + { + "bbox": [ + 52, + 302, + 176, + 375 + ], + "spans": [ + { + "bbox": [ + 52, + 302, + 176, + 375 + ], + "type": "image", + "image_path": "0ed03e0f33b4960b11c16d41715dd077eda723133fe489a36d6921b6963e8d05.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 197, + 316, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 316, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 197, + 316, + 296, + 333 + ], + "type": "text", + "content": "How many chairs are behind the long table? [Clicks]" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 219, + 340, + 311, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 340, + 311, + 357 + ], + "spans": [ + { + "bbox": [ + 219, + 340, + 311, + 357 + ], + "type": "text", + "content": "3 chairs are behind long table" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 320, + 262, + 338, + 279 + ], + "blocks": [ + { + "bbox": [ + 320, + 262, + 338, + 279 + ], + "lines": [ + { + "bbox": [ + 320, + 262, + 338, + 279 + ], + "spans": [ + { + "bbox": [ + 320, + 262, + 338, + 279 + ], + "type": "image", + "image_path": "acec21fe01ee05cb9894c68945854c163eafe73e4b57b82637be045b863290b6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 395, + 76, + 493, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 76, + 493, + 88 + ], + "spans": [ + { + "bbox": [ + 395, + 76, + 493, + 88 + ], + "type": "text", + "content": "3D Embodied Planning" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 349, + 93, + 536, + 187 + ], + "blocks": [ + { + "bbox": [ + 349, + 93, + 536, + 187 + ], + "lines": [ + { + "bbox": [ + 349, + 93, + 536, + 187 + ], + "spans": [ + { + "bbox": [ + 349, + 93, + 536, + 187 + ], + "type": "image", + "image_path": "39890bb25308a1c43a3d059048182459ae626122deb627f022f787ec3001e298.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 371, + 200, + 496, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 200, + 496, + 217 + ], + "spans": [ + { + "bbox": [ + 371, + 200, + 496, + 217 + ], + "type": "text", + "content": "I want to clean the kitchen. What should I do?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 370, + 226, + 515, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 226, + 515, + 267 + ], + "spans": [ + { + "bbox": [ + 370, + 226, + 515, + 267 + ], + "type": "text", + "content": "1. pick up any trash off the floor. 2. Open the kitchen shelf. 3. Clean the kitchen sink with a cleaning solution. 4. Clean the kitchen sink with a paper towel." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 353, + 271, + 365, + 289 + ], + "blocks": [ + { + "bbox": [ + 353, + 271, + 365, + 289 + ], + "lines": [ + { + "bbox": [ + 353, + 271, + 365, + 289 + ], + "spans": [ + { + "bbox": [ + 353, + 271, + 365, + 289 + ], + "type": "image", + "image_path": "2e53409ca803248459a5c65b178fa6305e5d8dfe40bda1c8906ab5ec83c6895f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 371, + 277, + 523, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 277, + 523, + 335 + ], + "spans": [ + { + "bbox": [ + 371, + 277, + 523, + 335 + ], + "type": "text", + "content": "I want to clean the kitchen. I have done these things: 1. Turn on the lamp on the table. 2. Open the kitchen cabinet and take out cleaning supplies. 3. Clean the countertops with a cloth. 4. Clean the stove and microwave. What should I do next?" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 378, + 345, + 512, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 345, + 512, + 354 + ], + "spans": [ + { + "bbox": [ + 378, + 345, + 512, + 354 + ], + "type": "text", + "content": "Clean the refrigerator and cabinets" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 521, + 252, + 539, + 270 + ], + "blocks": [ + { + "bbox": [ + 521, + 252, + 539, + 270 + ], + "lines": [ + { + "bbox": [ + 521, + 252, + 539, + 270 + ], + "spans": [ + { + "bbox": [ + 521, + 252, + 539, + 270 + ], + "type": "image", + "image_path": "1c4125d18ffa1d6187efc09522e9fa038b5f2eb398a348d4b213e9c23fc2ee63.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 56, + 381, + 207, + 501 + ], + "blocks": [ + { + "bbox": [ + 56, + 381, + 207, + 501 + ], + "lines": [ + { + "bbox": [ + 56, + 381, + 207, + 501 + ], + "spans": [ + { + "bbox": [ + 56, + 381, + 207, + 501 + ], + "type": "image", + "image_path": "3796156bba50e55762b755a1d55c3160e3e6861b27928eb8dafa34d0bc559a4e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 507, + 545, + 532 + ], + "lines": [ + { + "bbox": [ + 46, + 507, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 545, + 532 + ], + "type": "text", + "content": "Figure 4. Qualitative Results. We provide several visualization results on various 3D vision and language tasks in diverse 3D environments (living room, classroom, kitchen, and bedroom). Red highlights the wrong answer." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 325, + 384, + 417, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 384, + 417, + 397 + ], + "spans": [ + { + "bbox": [ + 325, + 384, + 417, + 397 + ], + "type": "text", + "content": "3D Scene Description" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 233, + 404, + 318, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 404, + 318, + 413 + ], + "spans": [ + { + "bbox": [ + 233, + 404, + 318, + 413 + ], + "type": "text", + "content": "Describe this 3D scene." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 229, + 421, + 515, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 421, + 515, + 487 + ], + "spans": [ + { + "bbox": [ + 229, + 421, + 515, + 487 + ], + "type": "text", + "content": "This room is a kitchen with a stove, stovetop, cabinets, refrigerator, oven, cooler, recycling bins, microwave, water heater, soap dish, trash can, a sink, and a refrigerator. There are multiple walls scattered in the room, creating a separate space for dining and cooking. The door leads to the kitchen. Another door and several windows hang around the area, providing a large view of the floor. As well as additional doors, walls, windows, doorstopper caps, kitchen items, sink and fridge, the food preparation and storage areas are visible." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 47, + 540, + 124, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 124, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 124, + 552 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 712 + ], + "type": "text", + "content": "In this paper, we present LL3DA, a large language 3D assistant that could take both textual- and visual- interactions from human for understanding, reasoning, and planning in complex 3D environments. Our model directly encodes the 3D point cloud and aggregates information from scenes and human interactions with the attention mechanism. We show that the visual interactions could remove the ambiguities in cluttered 3D environments, showing mighty instruction-following capacities. Experiments show that our method could achieve remarkable results on various 3D vision-language benchmarks. We hope our approach could inspire further designs and training strategies for large 3D language models. In future studies, we believe that the construction" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 540, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 565 + ], + "type": "text", + "content": "of high-quality and diverse annotations will further enhance the model's reasoning and planning capabilities." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 567, + 421, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 421, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 421, + 581 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (No. 62071127 and 62101137), National Key Research and Development Program of China (No. 2022ZD0160100), Shanghai Natural Science Foundation (No. 23ZR1402900), Shanghai Municipal Science and Technology Major Project (No. 2021SHZDZX0103), A*STAR AME Programmatic Funding A18A2b0046, RobotHTPO Seed Fund under Project C211518008, and EDB Space Technology Development Grant under Project S22-19016-STDP. The computations in this research were performed using the CFFF platform of Fudan University." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26435" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In European Conference on Computer Vision, pages 422-440. Springer, 2020. 2, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 257 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16464-16473, 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 355 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 355 + ], + "type": "text", + "content": "[5] Yang Cao, Yihan Zeng, Hang Xu, and Dan Xu. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[6] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European Conference on Computer Vision, pages 202-221. Springer, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 491 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A speaker-listener architecture for semi-supervised dense captioning and visual grounding in rgb-d scans. arXiv preprint arXiv:2112.01551, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 287, + 545 + ], + "type": "text", + "content": "[9] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 601 + ], + "type": "text", + "content": "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11124-11133, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[11] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. arXiv preprint arXiv:2309.02999, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3193-3203, 2021. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18109-18119, 2023. 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 239 + ], + "type": "text", + "content": "[15] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "text", + "content": "[17] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning, 2023. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[18] Alexandros Delitzas, Maria Parelli, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Multi-clip: Contrastive vision-language pre-training for question answering tasks in 3d scenes. arXiv preprint arXiv:2306.02329, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 453 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 510 + ], + "type": "text", + "content": "[20] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7010-7019, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "text", + "content": "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[22] Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical neural story generation. arXiv preprint arXiv:1805.04833, 2018.7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[23] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive like a human: Rethinking autonomous driving with large language models. arXiv preprint arXiv:2307.07162, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[24] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26436" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[25] Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. arXiv preprint arXiv:2311.16483, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 174 + ], + "type": "text", + "content": "[26] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2344-2352, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 176, + 287, + 209 + ], + "type": "text", + "content": "[27] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 211, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 211, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 211, + 287, + 266 + ], + "type": "text", + "content": "[28] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9202-9212, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 268, + 287, + 312 + ], + "type": "text", + "content": "[29] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. arXiv preprint arXiv:2307.12981, 2023. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 314, + 287, + 367 + ], + "type": "text", + "content": "[30] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 426 + ], + "type": "text", + "content": "[31] Srinivasan Iyer, Xi Victoria Lin, Ramakanth Pasunuru, Todor Mihaylov, Daniel Simig, Ping Yu, Kurt Shuster, Tianlu Wang, Qing Liu, Punit Singh Koura, et al. Opt-iml: Scaling language model instruction meta learning through the lens of generalization. arXiv preprint arXiv:2212.12017, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 428, + 287, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 428, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 428, + 287, + 460 + ], + "type": "text", + "content": "[32] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 463, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 506 + ], + "type": "text", + "content": "[33] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. arXiv preprint arXiv:2203.05203, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 509, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 563 + ], + "type": "text", + "content": "[34] Zhao Jin, Munawar Hayat, Yuwei Yang, Yulan Guo, and Yinjie Lei. Context-aware alignment and mutual masking for 3d-language pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10984-10994, 2023. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 620 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 667 + ], + "type": "text", + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[37] Mingsheng Li, Xin Chen, Chi Zhang, Sijin Chen, Hongyuan Zhu, Fukun Yin, Gang Yu, and Tao Chen. M3dbech: Let's instruct large models with multi-modal 3d prompts. arXiv preprint arXiv:2312.10763, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[38] Yanda Li, Chi Zhang, Gang Yu, Zhibin Wang, Bin Fu, Guosheng Lin, Chunhua Shen, Ling Chen, and Yunchao Wei. Stablellava: Enhanced visual instruction tuning with synthesized image-dialogue data. arXiv preprint arXiv:2308.10253, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[39] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 218 + ], + "type": "text", + "content": "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 252 + ], + "type": "text", + "content": "[42] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. arXiv preprint arXiv:2306.07279, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 297 + ], + "type": "text", + "content": "[43] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 342 + ], + "type": "text", + "content": "[44] Aihua Mao, Zhi Yang, Wanxin Chen, Ran Yi, and Yong-jin Liu. Complete 3d relationships extraction modality alignment network for 3d dense captioning. IEEE Transactions on Visualization and Computer Graphics, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 376 + ], + "type": "text", + "content": "[45] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 378, + 474, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 378, + 474, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 378, + 474, + 388 + ], + "type": "text", + "content": "[46] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 390, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 390, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 390, + 545, + 443 + ], + "type": "text", + "content": "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 446, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 446, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 446, + 545, + 511 + ], + "type": "text", + "content": "[48] Maria Parelli, Alexandros Delitzas, Nikolas Hars, Georgios Vlassis, Sotirios Anagnostidis, Gregor Bachmann, and Thomas Hofmann. Clip-guided vision-language pre-training for question answering in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5606-5611, 2023. 2, 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "text", + "content": "[49] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "text", + "content": "[50] Zhende Song, Chenchen Wang, Jiamu Sheng, Chi Zhang, Gang Yu, Jiayuan Fan, and Tao Chen. Moviellm: Enhancing long video understanding with ai-generated movies. arXiv preprint arXiv:2403.01422, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 614, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 678 + ], + "type": "text", + "content": "[51] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems, 33:7537-7547, 2020. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[52] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26437" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1,3,4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[53] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[54] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 240 + ], + "type": "text", + "content": "[55] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19231-19242, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 274 + ], + "type": "text", + "content": "[56] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[57] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 374 + ], + "type": "text", + "content": "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 409 + ], + "type": "text", + "content": "[59] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. IEEE Transactions on Visualization and Computer Graphics, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[60] Fukun Yin, Xin Chen, Chi Zhang, Biao Jiang, Zibo Zhao, Jiayuan Fan, Gang Yu, Taihao Li, and Tao Chen. Shapept: 3d shape generation with a unified multi-modal language model. arXiv preprint arXiv:2311.17618, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[61] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[62] Chi Zhang, Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 588 + ], + "type": "text", + "content": "[63] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[64] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. arXiv preprint arXiv:2307.03601, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[65] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[66] Lichen Zhao, Daigang Cai, Jing Zhang, Lu Sheng, Dong Xu, Rui Zheng, Yinjie Zhao, Lipeng Wang, and Xibo Fan. Towards explainable 3d grounded visual question answering:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 274 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "A new benchmark and strong baseline. IEEE Transactions on Circuits and Systems for Video Technology, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[67] Yufeng Zhong, Long Xu, Jiebo Luo, and Lin Ma. Contextual modeling for 3d dense captioning on point clouds. arXiv preprint arXiv:2210.03925, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 173 + ], + "type": "text", + "content": "[68] Qiang Zhou, Chaohui Yu, Shaofeng Zhang, Sitong Wu, Zhibing Wang, and Fan Wang. Regionblip: A unified multimodal pre-training framework for holistic and regional comprehension, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[69] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[70] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2, 5, 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26438" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_content_list.json b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..397696158994de62114c775f197213642ffac312 --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_content_list.json @@ -0,0 +1,1354 @@ +[ + { + "type": "text", + "text": "LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation", + "text_level": 1, + "bbox": [ + 181, + 160, + 788, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kibum Kim $^{1}$ Kanghoon Yoon $^{1}$ Jaehyeong Jeon $^{1}$ Yeonjun In $^{1}$ Jinyoung Moon $^{2}$ Donghyun Kim $^{3}$ Chanyoung Park $^{1*}$ KAIST $^{2}$ ETRI $^{3}$ Korea University", + "bbox": [ + 181, + 226, + 824, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kb.kim,ykhoon08,wogud405,yeonjun.in,cy PARK}@kaist.ac.kr jymoon@etri.re.kr, d_kim@korea.ac.kr", + "bbox": [ + 88, + 277, + 914, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 234, + 323, + 312, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weakly-Supervised Scene Graph Generation (WSSGG) research has recently emerged as an alternative to the fully-supervised approach that heavily relies on costly annotations. In this regard, studies on WSSGG have utilized image captions to obtain unlocalized triplets while primarily focusing on grounding the unlocalized triplets over image regions. However, they have overlooked the two issues involved in the triplet formation process from the captions: 1) Semantic over-simplification issue arises when extracting triplets from captions, where fine-grained predicates in captions are undesirably converted into coarse-grained predicates, resulting in a long-tailed predicate distribution, and 2) Low-density scene graph issue arises when aligning the triplets in the caption with entity/predicate classes of interest, where many triplets are discarded and not used in training, leading to insufficient supervision. To tackle the two issues, we propose a new approach, i.e., Large Language Model for weakly-supervised SGG (LLM4SGG), where we mitigate the two issues by leveraging the LLM's in-depth understanding of language and reasoning ability during the extraction of triplets from captions and alignment of entity/predicate classes with target data. To further engage the LLM in these processes, we adopt the idea of Chain-of-Thought and the in-context few-shot learning strategy. To validate the effectiveness of LLM4SGG, we conduct extensive experiments on Visual Genome and GQA datasets, showing significant improvements in both Recall@K and mean Recall@K compared to the state-of-the-art WSSGG methods. A further appeal is that LLM4SGG is data-efficient, enabling effective model training with a small amount of training images. Our code is available on https://github.com/rlqja1107/torch-LLM4SGG", + "bbox": [ + 76, + 347, + 472, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 791, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene Graph Generation (SGG) is a fundamental task in computer vision, aiming at extracting structured visual", + "bbox": [ + 76, + 815, + 469, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "knowledge from images [21, 32, 36, 39, 45, 56]. Most existing SGG methods are fully-supervised, i.e., they heavily rely on the ground-truth annotations that involve the class information of entities and predicates as well as the bounding box of entities [19, 48, 50]. However, since creating extensively annotated scene graph datasets is costly, the heavy reliance on these annotations imposes practical limitations on the model training [54]. To mitigate the high cost associated with manual annotations, weakly-supervised scene graph generation (WSSGG) approaches have recently emerged, aiming at training an SGG model without any annotated scene graph dataset. Specifically, the main idea of recent WSSGG methods is to leverage image captions along with associated images, as they can be easily collected from the Web [20, 47, 54, 57].", + "bbox": [ + 499, + 324, + 891, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The training process of WSSGG model using image captions requires four steps as illustrated in Figure 1(a). Step 1: Preparing an image and its caption. Step 2: Parsing the image caption, i.e., triplets formed as $\\langle$ subject, predicate, object $\\rangle$ are extracted from the image caption through an off-the-shelf parser [30, 43]. Step 3: Aligning the triplets in the caption with entity/predicate classes of interest, i.e., entity (subject, object) and predicate classes in the extracted triplets obtained in Step 2 are aligned with the entity and predicate classes in the target data $^{1}$ , respectively. This alignment is based on their synonym/hypernym/hyponym contained in an external knowledge base (KB), e.g., WordNet [25]. Step 4: Grounding unlocalized entities in the extracted triplets, i.e., unlocalized entities (subjects and objects) are matched with relevant image regions generated by a pre-trained object detector, e.g., Faster R-CNN [29]. The localized entities and predicates in the extracted triplets then serve as pseudo-labels for training an SGG model.", + "bbox": [ + 499, + 527, + 891, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing WSSGG approaches mainly focus on Step 4 [20, 33, 47, 57]. For example, in Figure 1(a), their efforts have been focused on grounding the entity person in an unlocalized triplet with an image region that captures the sitting behavior. More precisely, LSwS [47] exploits the con", + "bbox": [ + 499, + 771, + 891, + 840 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 139, + 45, + 207, + 80 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 258, + 42, + 783, + 84 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 853, + 224, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We use Visual Genome [13] as the target data.", + "bbox": [ + 517, + 853, + 770, + 866 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28306", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbe81ecf05d9410d234f9c6e0dfd2a9aff2ea0480193dbbc89b79b659e9f430a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 119, + 125, + 587, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f479b1adad71406a506837942a910fa16066061a97fcf34f789c127de451856.jpg", + "image_caption": [ + "(c) Semantic Oversimplification" + ], + "image_footnote": [], + "bbox": [ + 592, + 127, + 842, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb48de475fefc6a4e19bcbefdee03faf14bd86c5ff720f3a4d86b5f01b9dd20f.jpg", + "image_caption": [ + "(a) Pipeline of Weakly-Supervised SGG", + "(b) Predicate Distribution", + "Figure 1. (a) The pipeline of weakly-supervised SGG. (b) The predicate distribution of unlocalized triplets (Parser+KB vs. Ours). In Parser+KB, the distribution becomes heavily long-tailed, and 12 out of 50 predicates are non-existent. (c) Semantic over-simplification caused by a rule-based parser in Step 2. (d) Low-density scene graph caused by the static structure of KB in Step 3." + ], + "image_footnote": [], + "bbox": [ + 124, + 227, + 582, + 319 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d5a9330d8e83d4927b4d110d691e08b082bc2e26c4608d32f158caa1a74932c6.jpg", + "image_caption": [ + "(d) Low-Density Scene Graph" + ], + "image_footnote": [], + "bbox": [ + 590, + 228, + 842, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "textual object information to accurately ground the unlocalized entities, leveraging the linguistic structure embedded within the triplets. Another line of research [20] employs a pre-trained vision-language model [15] to reflect the semantic interactions among entities within the image caption.", + "bbox": [ + 75, + 396, + 468, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, we argue that existing WSSGG approaches overlook the importance of the triplet formation process conducted in Step 2 and Step 3. We identify the two major issues described below, i.e., semantic over-simplication and low-density scene graph, which incur incomplete unlocalized triplets after Step 2 and 3. These incomplete triplets are mostly uninformative predicates with a limited number, and negatively impact the training of an SGG model even when entities are correctly grounded in Step 4. To demonstrate the impact of incomplete unlocalized triplets, we follow the conventional process to extract unlocalized triplets (i.e., Step 1-3), and conduct an examination of triplets obtained from COCO caption dataset, which are generated through Scene Parser [43] in Step 2 and WordNet [25] in Step 3. As a result, we identify the following two issues:", + "bbox": [ + 75, + 465, + 468, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Semantic Over-simplification: We find that the standard scene graph parser [43] operating on heuristic rule-based principles commonly used in Step 2 leads to a semantic over-simplification of predicates in extracted triplets. In other words, fine-grained predicates are undesirably converted into coarse-grained predicates, which we refer to as semantic over-simplification. For example, in Figure 1(c), an informative predicate lying on (i.e., fine-grained predicate) in the image caption is converted into a less informative predicate on (i.e., coarse-grained predicate), because the rule-based parser fails to capture the predicate lying on at once, and its heuristic rules fall short of accommodating the diverse range of caption's structure. As a result, the predicate distribution becomes heav-", + "bbox": [ + 76, + 676, + 469, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ily long-tailed, in which coarse-grained predicates (e.g., with, on, in) greatly outnumber fine-grained predicates (e.g., parked on, covered in) (Figure 1(b)). To make the matter worse, numerous fine-grained predicates eventually end up in a frequency of 0, even though they are originally present in the captions. Specifically, 12 out of 50 predicates are non-existent, which means that these 12 predicates can never be predicted, since the model is not trained on these predicates at all.", + "bbox": [ + 510, + 396, + 891, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Low-Density Scene Graph: We find that the KB-based triplet alignment in Step 3 leads to low-density scene graphs, i.e., the number of remaining triplets after Step 3 is small. We attribute the low-density scene graphs primarily to the utilization of KB in Step 3. Specifically, a triplet is discarded if any of the three components (i.e., subject, predicate, object) or their synonym/hypernym/hyponym within the triplet fail to align with the entity or predicate classes in the target data. For example, in Figure 1(d), the triplet $\\langle$ elephant, carrying, $\\log \\rangle$ is discarded because $\\log$ does not exist in Visual Genome dataset nor its synonym/hypernym, even if elephant and carrying do exist. In Table 1, we report the number of triplets and images in Visual Genome dataset, which is a common benchmark dataset used in fully-supervised SGG approaches, and COCO caption dataset, which is a common benchmark dataset used in weakly-supervised SGG approaches. We observe that on average Visual Genome dataset contains 7.1 triplets (i.e., $405\\mathrm{K} / 57\\mathrm{K}$ ) per image (See Table 1(a)), while COCO dataset contains only 2.4 triplets (i.e., $154\\mathrm{K} / 64\\mathrm{K}$ ) per image (See Table 1(b)). This indicates that existing WSSGG approaches suffer from the lack of sufficient supervision per image, leading to poor generalization and performance degradation [47, 49]. In summary, relying on", + "bbox": [ + 500, + 518, + 891, + 857 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28307", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/4f096a7d51a8bbafc4c63732fa578d09252e721de43cd7418cac20f21b7b80ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetHow to annotate# Triplet# Image
Fully-Supervised approach
(a) Visual GenomeManual405K57K
Weakly-Supervised approach
(b) COCO CaptionParser+KB154K64K
(c) COCO CaptionLLM344K64K
", + "bbox": [ + 127, + 122, + 415, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of scene graph density.", + "bbox": [ + 139, + 200, + 405, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the static structured knowledge of KB is insufficient to cover the semantic relationships among a wide a range of words, which incurs the low-density scene graph after Step 3.", + "bbox": [ + 88, + 223, + 468, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To alleviate the semantic over-simplification and the low-density scene graph issues, we propose a new approach, namely Large Language Model for weakly-supervised SGG (LLM4SGG) that adopts a pre-trained Large Language Model (LLM), which has shown remarkable transferability to various downstream tasks in NLP such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Inspired by the idea of Chain-of-Thought² (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into two chains, each of which replaces the rule-based parser in Step 2 (i.e., Chain-1) and the KB in Step 3 (i.e., Chain-2). More precisely, we design a prompt for extracting triplets from a caption, and ask the LLM to extract triplets formed as (Chain-1). We expect that the predicates extracted based on a comprehensive understanding of the caption's context via LLM are semantically rich, thereby alleviating the semantic over-simplification issue. Besides, to alleviate the low-density scene graph issue, we additionally incorporate a paraphrased version of the original caption. To this end, we further design a prompt for paraphrasing the original caption and extracting more triplets from the paraphrased caption. However, entities and predicates in the triplets obtained after Chain-1 are not yet aligned with the target data. Hence, we design a another prompt to align them with entity/predicate classes of interest, and ask the LLM to align them with semantically relevant lexeme included in a predefined lexicon, which is the set of vocabularies that are present in the target data (Chain-2). To further engage the LLM in the reasoning process of Chain-1 and Chain-2, we employ the in-context few-shot learning that incorporates a few input-output examples within the prompt, enabling the LLM to perform the task without the need for fine-tuning.", + "bbox": [ + 76, + 276, + 469, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To validate the effectiveness of LLM4SGG, we apply it to state-of-the-art WSSGG methods [54, 57]. Through extensive experiments, we show that LLM4SGG significantly enhances the performance of existing WSSGG methods in terms of mean Recall@K and Recall@K performance on Visual Genome and GQA datasets by alleviating the semantic over-simplification and the low-density scene graph", + "bbox": [ + 76, + 734, + 468, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(See Table 1(c) where the number of triplets increased to 334K). A further appeal of LLM4SGG is that it is data-efficient, i.e., it outperforms state-of-the-art baselines even with a small amount of training images, verifying the effectiveness of LLM4SGG.", + "bbox": [ + 499, + 126, + 891, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 517, + 193, + 852, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We identify two major issues overlooked by existing WSSGG studies, i.e., semantic over-simplification and low-density scene graph.", + "bbox": [ + 507, + 207, + 890, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We leverage an LLM along with the CoT strategy and the in-context few-shot learning technique to extract informative triplets without the need for fine-tuning the LLM. To the best of our knowledge, we are the first to leverage an LLM for the SGG task.", + "bbox": [ + 509, + 247, + 890, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- LLM4SGG outperforms the state-of-the-art WSSGG methods, especially in terms of mR@K, demonstrating its efficacy in addressing the long-tail problem in WSSGG for the first time.", + "bbox": [ + 509, + 314, + 890, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 499, + 380, + 648, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Weakly-Supervised Scene Graph Generation (WSSGG). The WSSGG task aims to train an SGG model without relying on an annotated scene graph dataset. To achieve this, most WSSGG studies [20, 47, 54, 57] utilize image captions and ground unlocalized triplets with image regions. Specifically, VSPNet [49] proposes the iterative graph alignment algorithm to reflect the high-order relations between unlocalized triplets and image regions. SGNLS [57] uses a pre-trained object detector [29] to ground the entities in unlocalized triplets, which share the same classes with the output of object detectors. In addition to the information derived from the object detector, [20] employs a pre-trained vision-language model [15] to capture the semantic interactions among objects. VS3 [54] uses a grounding-based object detector [18], which calculates the similarity between the entity text in unlocalized triplet and image region, thereby grounding the unlocalized triplets. However, these methods overlook the triplet formation process that leads to the semantic over-simplification (Step 2) and the low-density scene graph (Step 3). In this regard, existing methods result in a sub-optimal performance even when unlocalized triplets are correctly grounded in Step 4.", + "bbox": [ + 497, + 402, + 891, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language Model (LLM). LLMs have demonstrated remarkable transferability to various downstream tasks such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Specifically, GPT-3 (175B) [2] stands as a cornerstone to break the line of numerous language tasks. Inspired by GPT-3, PaLM (540B) [5], LLaMA (65B) [38], OPT (175B) [53], and LaMDA (137B) [37] have been subsequently introduced. More recently, advanced GPT models (e.g., GPT-4 [28], ChatGPT [27]) fine-tuned with human feedback have gained prominence and widely applied for diverse applications, e.g., planner of tools [9, 24], mobile", + "bbox": [ + 497, + 717, + 891, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We use the CoT strategy as a means to arrive at an answer in a stepwise manner, which differs from the Chain-of-Thought prompting.", + "bbox": [ + 76, + 842, + 468, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28308", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "task automation [42]. In this work, we employ the power of LLM (i.e., ChatGPT) to alleviate the two issues, i.e., semantic over-simplification and low-density scene graph, in the context of the WSSGG task.", + "bbox": [ + 75, + 125, + 469, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In-Context Few-shot Learning. In-context few-shot learning incorporates a few input-output examples related to a target task, conditioning the LLM on the context of examples. Specifically, GPT-3 [2] pioneered the concept of in-context learning to facilitate an LLM as a versatile model on diverse tasks. This breakthrough has proliferated a plethora of research to leverage the in-context few-shot learning for various tasks [24, 26, 46, 55]. More precisely, Chameleon [24] integrates a few examples to enhance its understanding of tool planning task. [26] utilizes positive and negative examples related to questions for question generation tasks. ReAct [46] incorporates examples of reasoning with action for solving decision-making tasks. Inspired by recent in-context few-shot learning approaches, we provide a few examples to LLMs to help 1) understand the process of triplet extraction from a caption (i.e., Step 2), and 2) align the entity/predicate classes with the target data (i.e., Step 3) in the context of the WSSGG task.", + "bbox": [ + 75, + 184, + 472, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 441, + 168, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe LLM4SGG in detail. We would like to emphasize that LLM4SGG mainly focuses on the triplet formation process conducted in Step 2 (Parsing) and Step 3 (aligning), while existing WSSGG approaches mainly focus on Step 4 (grounding). We start by presenting the problem formulation of WSSGG (Section 3.1), followed by the prompt configuration (Section 3.2). Next, we introduce how LLMs are adopted to address the two issues of conventional WSSGG approaches when parsing the image caption (Section 3.3) and aligning the triplets in captions with entity/predicate classes of interest (Section 3.4). Finally, we ground the unlocalized triplets by associating them with bounding boxes (i.e., image regions) and train the SGG model using the localized triplets (Section 3.5). The overall pipeline of LLM4SGG is shown in Figure 2.", + "bbox": [ + 75, + 465, + 469, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 680, + 279, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the fully supervised SGG task, we aim to detect a scene graph $\\mathbf{G}_f = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_f}$ that consists of triplets given an image $\\mathcal{I}$ , where $N_{f}$ is the number of triplets in the image. $\\mathbf{s}_i$ and $\\mathbf{o}_i$ denote the $i^{\\mathrm{th}}$ subject and the object, respectively, whose bounding boxes are $\\mathbf{s}_{i,b}$ , $\\mathbf{o}_{i,b}$ , and entity classes are $\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}\\in \\mathcal{C}_e$ , where $\\mathcal{C}_e$ is the set of predefined entity classes in the target data. $\\mathbf{p}_i$ denotes the predicate between $\\mathbf{s}_i$ and $\\mathbf{o}_i$ , and its class is $\\mathbf{p}_{i,c}\\in \\mathcal{C}_p$ , where $\\mathcal{C}_p$ is the set of predefined predicate classes in the target data. By using the ground truth scene graphs as supervision, fully supervised SGG approaches train an SGG model $\\mathcal{T}_{\\theta}:\\mathcal{I}\\to \\mathbf{G}_{f}$ , which maps an image to a scene graph.", + "bbox": [ + 75, + 700, + 469, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the weakly supervised SGG task, we aim to generate a scene graph when the ground truth scene graph is not given, i.e., there are no bounding boxes and entity/predicate class information. Instead, existing WSSGG approaches [20, 33, 47, 54, 57] use image captions along with associated images to produce scene graphs, i.e., localized triplets. More precisely, they extract a set of triplets $\\mathbf{G}_w = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_w}$ from the image captions, where $N_{w}$ is the number of triplets extracted from the captions. However, while the extracted triplets contain the class information (i.e., $\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}$ and $\\mathbf{p}_{i,c}$ ), they are unlocalized, since bounding boxes $\\mathbf{s}_{i,b}$ and $\\mathbf{o}_{i,b}$ are not included in the caption. Therefore, it is essential to perform the grounding step to associate the unlocalized triplets with bounding boxes. Once we have obtained the localized triplets, we can apply the conventional SGG training scheme, i.e., $\\mathcal{T}_{\\theta}:\\mathcal{I}\\rightarrow \\mathbf{G}_{f}$ .", + "bbox": [ + 497, + 125, + 891, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, our focus is to address the semantic oversimplification and low-density scene graph issues regarding the unlocalized triplets $\\mathbf{G}_w$ , which has been overlooked in existing WSSGG studies. Specifically, we aim to produce an enhanced $\\mathbf{G}_w$ by refining the process of scene graph dataset construction via an LLM. This refinement includes the triplet extraction step from the caption (Step 2) and the alignment of entity/predicate classes (Step 3), leveraging the LLM's comprehensive understanding of language and reasoning ability.", + "bbox": [ + 497, + 340, + 891, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Prompt Configuration", + "text_level": 1, + "bbox": [ + 497, + 483, + 705, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In fact, it is not a trivial task for an LLM to immediately generate triplets from a caption whose entities and predicates are aligned with entity/predicate classes of interest, as such a task is a novel task for the LLM. Inspired by the idea of the Chain-of-thought (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into the following two chains: Chain-1 - Extracting triplets from captions. Chain-2 - Aligning entities and predicates with the entity/predicate classes of interest. To carefully design each chain, we define the LLM function, i.e., $LLM(\\cdot)$ , with the following prompt input:", + "bbox": [ + 497, + 503, + 891, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO u t p u t = L L M (\\underbrace {\\text {T a s k d e s c r i p t i o n , I n - c o n t e x t e x a m p l e s , A c t u a l q u e s t i o n}} _ {\\text {P r o m p t i n p u t}},\n$$\n", + "text_format": "latex", + "bbox": [ + 502, + 657, + 885, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $LLM(\\cdot)$ is the decoder of the LLM, generating the (1) Output given the prompt input. The prompt input consists of three components in a sequence: 1) task description, i.e., the delineation of the task that we intend to perform, 2) in-context examples, i.e., sample questions and answers related to the task at hand, 3) actual question, i.e., an inquiry from which we intend to derive the answer. Note that in-context examples is closely related to the in-context few-shot learning [2, 40, 41], which is shown to enhance the LLM's understanding of the task. Note that the above configuration of the prompt input is applied to the triplet extraction (Chain-1) (Section 3.3) and the alignment of entity/predicate classes (Chain-2) (Section 3.4).", + "bbox": [ + 497, + 686, + 891, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28309", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51b4929e428bf2e4b0ad9d458e510ab8301e43b735b4afbf5fb32a02f1a82010.jpg", + "image_caption": [ + "Figure 2. The pipeline of LLM4SGG. Given an image with its caption, we use an LLM to extract triplets from the original caption (Step 2-1) and the paraphrased caption (Step 2-2). Then, we align the entity/predicate classes within the extracted triplets with semantically similar lexeme in the target data via an LLM (Step 3), obtaining the unlocalized triplets. Lastly, we ground the unlocalized triplets over image regions (Step 4) followed by the training of an SGG model." + ], + "image_footnote": [], + "bbox": [ + 150, + 123, + 816, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Chain-1. Triplet Extraction via LLM (Step 2 in Figure 2)", + "text_level": 1, + "bbox": [ + 76, + 406, + 469, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the LLM's comprehensive understanding of the context of an image caption, we aim to extract triplets from the caption. As discussed in Section 1, we use not only the original caption but also a paraphrased caption generated by the LLM to address the low-density scene graph issue.", + "bbox": [ + 75, + 438, + 469, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extracting triplets from a paraphrased caption (Step 2-2). To extract triplets from a paraphrased caption, we inform the LLM about the task at hand by presenting the following prompt: FROM THE GIVEN SENTENCE, THE TASK IS TO EXTRACT MEANINGFUL TRIPLETS FORMED AS $\\langle$ SUBJECT, PREDICATE, OBJECT $\\rangle$ (i.e., Task description in Equation 1). We then instruct the LLM to follow the two steps, i.e., paraphrasing step and triplet extraction step. To help the LLM understand the process of performing the two steps, we present few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation $1^{3}$ ), which involves a manual construction of questions and corresponding answers to the paraphrasing and the triplet extraction steps. That is, for given a caption, we show the LLM how we expect a paraphrased caption and the extracted triplets would look like (e.g., Given \"Four clocks sitting on a floor next to a woman's feet,\" we show the LLM that a paraphrased sentence would be \"Four clocks are placed on the floor beside a woman's feet,\" and extracted triplets would be $\\langle$ clocks, placed on, floor $\\rangle$ and $\\langle$ clocks, beside, feet $\\rangle$ ). Lastly, we show the caption of our interest to the LLM, and let the LLM extract triplets from the caption (i.e., Actual question in Equation 1). Please refer to Figure 2 right (bottom) for an example of the prompt input used in Step 2-2.", + "bbox": [ + 75, + 507, + 469, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extracting triplets from the original caption (Step 2-1). As extracting triplets from the original caption does not involve the caption paraphrasing step, we exclude it from the prompt used in Step 2-2. Please refer to Figure 2 right (top) for an example of the prompt input used in Step 2-1.", + "bbox": [ + 497, + 407, + 891, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, we obtain triplets from both the original and paraphrased captions after Step 2-1 and Step 2-2, respectively, which in turn alleviates the semantic oversimplification issue of predicates and the low-density scene graph issue.", + "bbox": [ + 497, + 476, + 891, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Chain-2. Alignment of Classes in Triplets via LLM (Step 3 in Figure 2)", + "text_level": 1, + "bbox": [ + 497, + 556, + 891, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The entities (i.e., subject and object) and predicates within the triplets obtained from Step 2 described in Section 3.3 are not yet aligned with the target data. Based on the semantic reasoning ability of the LLM, we aim to align them with the semantically relevant lexeme in the target data.", + "bbox": [ + 497, + 589, + 891, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aligning entities in the triplets with entity classes of interest. We instruct the LLM with the following prompt: GIVEN THE LEXEME $\\{\\mathsf{ENTITY}\\}$ , FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED ENTITY LEXICON, where the predefined entity lexicon is $\\mathcal{C}_e$ (i.e., Task description in Equation 1). Similar to Section 3.3, we present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding hierarchical relationships such as pigeon being semantically relevant to bird, and singular-plural relationships such as surfboards being semantically relevant to surfboard. Lastly, we show the entity of our interest to the LLM (i.e., Actual question in Equation 1), which enables the LLM to generate an answer by finding the most", + "bbox": [ + 497, + 662, + 891, + 867 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3We use captions in COCO caption dataset for examples.", + "bbox": [ + 94, + 853, + 397, + 866 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28310", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "semantically relevant entity in $\\mathcal{C}_e$ . Please refer to Table 7 in Appendix A.1 for an example of the prompt input.", + "bbox": [ + 76, + 125, + 469, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aligning predicates in the triplets with predicate classes of interest. Likewise, we instruct the LLM with the following prompt: GIVEN THE LEXEME $\\{\\mathrm{PREDICATE}\\}$ , FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED Predicate Lexicon, where the predefined predicate lexicon is $C_p$ (i.e., Task description in Equation 1). We also present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding tense relationships such as the lies on being semantically relevant to lying on, and positional relationship such as next to being semantically relevant to near. Lastly, we show the predicate of our interest to the LLM (i.e., Actual question in Equation 1). Please refer to Table 8 in Appendix A.1 for an example of the prompt input. Furthermore, regarding the alignment of classes within a large predefined lexicon, please refer to Appendix A.2. Note that as our main focus in this work is on developing a framework for utilizing LLMs in weakly-supervised SGG rather than finding a specific prompt design that works the best, we tested with a single prompt design. However, since prompt designs are crucial in leveraging LLMs, we plan to explore various designs in our future work.", + "bbox": [ + 76, + 155, + 472, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After performing Chain-1 (Section 3.3) and Chain-2 (Section 3.4), we obtain intermediate unlocalized triplets $\\hat{\\mathbf{G}}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{\\hat{N}_w}$ , where $\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\{\\mathcal{C}_e \\cup \\text{None}\\}$ and $\\mathbf{p}_{i,c} \\in \\{\\mathcal{C}_p \\cup \\text{None}\\}$ . It is worth noting that if there is no semantically relevant lexeme, we request the LLM to generate None as the answer, due to the fact that the entity/predicate classes in the target data may not cover a wide range of entities/predicate. Similar to the conventional approach, we discard a triplet if any of its three components (i.e., subject, predicate and object) is None. Lastly, we obtain the final unlocalized triplets $\\mathbf{G}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{N_w}(N_w \\leq \\hat{N}_w)$ , where $\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\mathcal{C}_e$ and $\\mathbf{p}_{i,c} \\in \\mathcal{C}_p$ .", + "bbox": [ + 76, + 465, + 472, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Model Training", + "text_level": 1, + "bbox": [ + 76, + 638, + 235, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the final unlocalized triplets $\\mathbf{G}_w$ , we ground them over relevant image regions to get localized triplets, meaning that we obtain $\\mathbf{s}_{i,b}$ and $\\mathbf{o}_{i,b}$ . To this end, we employ two state-of-the-art grounding methods, i.e., SGNLS [57] and VS3 [54]. Please refer to Appendix A.3 for more detail about how each method performs grounding. After grounding $\\mathbf{G}_w$ , we obtain localized triplets and use them as pseudo-labels for training a supervised SGG model. Please refer to Appendix A.4 for more detail about the model training.", + "bbox": [ + 75, + 657, + 469, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 801, + 202, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To train an SGG model without an annotated scene graph dataset, we use three caption datasets: COCO caption [3], Conceptual (CC) caption [31], and Visual", + "bbox": [ + 75, + 825, + 469, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Genome (VG) caption [44]. For fair comparisons, we use the same set of images that have been utilized in previous WSSGG studies [20, 47, 54, 57], leading to the utilization of 64K images on COCO caption dataset, 145K images on CC caption dataset, and 57K images on VG caption dataset. To evaluate the trained SGG model, we employ the widely used Visual Genome (VG) dataset [13] and GQA dataset [11]. The VG dataset contains the ground-truth localized triplet information annotated by humans. We follow the standard split of VG [44], which consist of 150 entity classes and 50 predicate classes. For the GQA dataset used in the SGG task, we follow the same pre-processing step of a previous SGG study [8], which involves selecting top-200 frequent entity classes and top-100 frequent predicate classes. In both datasets, $30\\%$ of the total images are used for evaluation. Please refer to Appendix C.1 for more details regarding the datasets. Note that we mainly use the COCO caption dataset for analysis of LLM4SGG throughout this paper, i.e., CC and VG caption datasets are exclusively used in quantitative result on VG (Section 4.1).", + "bbox": [ + 497, + 126, + 891, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. Recent fully-supervised SGG studies [1, 48, 51] have emphasized improving the accuracy of predictions for fine-grained predicates rather than coarse-grained predicates, since the former construct richer scene graphs. As a result, they commonly use mean Recall@K (mR@K) that computes the average of Recall@K (R@K) across all predicates. In line with the recent emphasis on fine-grained predicates, we incorporate both mR@K and R@K in our evaluation, whereas previous WSSGG studies [20, 47, 54] mainly rely on the R@K metric alone. Moreover, we report F@K, which is the harmonic average of R@K and mR@K to jointly consider R@K and mR@K, following previous SGG studies [12, 51]. Regarding the evaluation task, we follow previous WSSGG studies and adopt the Scene Graph Detection (SGDet) task, where both the ground-truth bounding box and the entity class information are not provided. Please refer to Appendix C.2 for more detail regarding the task.", + "bbox": [ + 497, + 398, + 891, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Please refer to Appendix C.3 for details regarding the baselines.", + "bbox": [ + 497, + 643, + 891, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Please refer to Appendix C.4 regarding the implementation details.", + "bbox": [ + 497, + 674, + 891, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Result on VG", + "text_level": 1, + "bbox": [ + 497, + 709, + 739, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the performance of baseline models and those when LLM4SGG is applied. We have the following observations based on COCO caption dataset: 1) Applying LLM4SGG to SGNLS and VS3 improves the performance in terms of R@K and mR@K, which demonstrates the effectiveness of the triplet formation through LLM4SGG. Notably, LLM4SGG significantly improves mR@K, implying that LLM4SGG effectively alleviates the long-tailed problem in WSSGG. This can be clearly seen in Figure 3, which shows the performance gain on fine", + "bbox": [ + 497, + 729, + 891, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28311", + "bbox": [ + 479, + 890, + 515, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0c29bae517e9c8df7a5aa7b2f4c3863b00d1e28593a4bb9d65e7eb50c69201f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetR@50R@100mR@50mR@100F@50F@100
Motif (CVPR'18) - Fully-supervisedVG31.8936.366.387.5710.63 / 12.5312.53
LSWS (CVPR'21)COCO Caption3.293.693.273.663.283.67
SGNLS (ICCV'21)3.804.462.512.783.023.43
SGNLS (ICCV'21)+LLM4SGG5.09+1.295.97+1.514.08+1.574.49+1.714.53+1.515.13+1.70
Li et al (MM'22)6.407.331.731.982.723.12
VS3(CVPR'23)6.608.012.883.254.014.62
VS3(CVPR'23)+LLM4SGG8.91+2.3110.43+2.427.11+4.238.18+4.937.91+3.909.17+4.55
VS3(CVPR'23)+Rwt4.255.045.175.994.675.47
VS3(CVPR'23)+Rwt+LLM4SGG5.10+0.856.34+1.308.42+3.259.90+3.916.35+1.697.73+2.26
VS3(CVPR'23)CC Caption6.698.201.732.042.753.27
VS3(CVPR'23)+LLM4SGG9.47+2.7810.69+2.495.40+3.676.09+4.056.88+4.137.76+4.49
VS3(CVPR'23)VG Caption14.5418.482.803.794.706.29
VS3(CVPR'23)+LLM4SGG18.40+3.8622.28+3.806.26+3.467.60+3.819.34+4.6411.33+5.04
", + "bbox": [ + 166, + 118, + 801, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aca0a378acdb45627d51d035e807e92e5f8de4db7d5f790dc07ef1c6f8e7c49a.jpg", + "image_caption": [ + "Figure 3. Per class performance (Bar: number of predicate instances, Line: Recall@100)." + ], + "image_footnote": [], + "bbox": [ + 106, + 311, + 433, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "grained predicates. 2) $\\mathrm{VS}^3 +\\mathrm{Rwt} + \\mathrm{LLM4SGG}$ further improves mR@K of $\\mathrm{VS}^3 +\\mathrm{Rwt}$ . We attribute this to the fact that the conventional approach generates a limited number of fine-grained predicates, which makes the reweighting strategy less effective within $\\mathrm{VS}^3$ . Especially, non-existent predicates can never be predicted even when the reweighting strategy is applied. On the other hand, LLM4SGG increases the number of instances that belong to fine-grained predicates, which is advantageous for the reweighting strategy. For the per class performance comparison over the reweighting stregty, please refer to Appendix D.1. 3) The performance gain obtained from applying LLM4SGG is greater on $\\mathrm{VS}^3$ (i.e., $\\mathrm{VS}^3 +\\mathrm{LLM4SGG}$ ) than on SGNLS (i.e., SGNLS+LLM4SGG). The major reason lies in the difference in how SGNLS and $\\mathrm{VS}^3$ make use of the pool of 344K unlocalized triplets obtained through LLM4SGG. Specifically, in the grounding process of SGNLS, we observe that 100K out of 344K unlocalized triplets (i.e., $29\\%$ ) fail to be grounded, and thus not used for training. On the other hand, $\\mathrm{VS}^3$ successfully grounds all 344K unlocalized triplets and fully utilize them for training, allowing it to fully enjoy the effectiveness of LLM4SGG. This indicates that LLM4SGG makes synergy when paired with a grounding method that is capable of fully utilizing the unlocalized triplets. For more details regarding the impact of grounding methods, please refer to Appendix B. 4) Regarding the performance comparison with a fully-supervised approach (i.e., Motif), please refer to Appendix D.2.", + "bbox": [ + 75, + 443, + 469, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we observe that applying LLM4SGG to CC and VG caption datasets also significantly improves performance, demonstrating the effectiveness of LLM4SGG.", + "bbox": [ + 75, + 825, + 468, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0fc2459895c020cdc71c25092028e92c0941974e66776cccbc7fa973d277a120.jpg", + "table_caption": [ + "Table 2. Performance comparisons on the SGDet task. The best performance among WSSGG models within each dataset is in bold. The red numbers indicate the absolute performance improvement after applying LLM4SGG. Rwt denotes using the reweighting strategy [51]." + ], + "table_footnote": [], + "table_body": "
RowPCLPLA# TripletR@50 / 100mR@50 / 100F@50 / 100
(a)154K6.60 / 8.012.88 / 3.254.01 / 4.62
(b)✓243K9.46 / 11.223.43 / 3.925.03 / 5.81
(c)✓✓256K8.42 / 9.855.99 / 6.957.00 / 8.15
(d)✓✓327K11.76 / 13.383.50 / 4.055.39 / 6.22
(e)✓✓✓344K8.91 / 10.437.11 / 8.187.91 / 9.17
", + "bbox": [ + 525, + 310, + 865, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies. (PC: Using Paraphrased Caption in addition to the original caption / LP: LLM-based Parsing / LA: LLM-based Alignment)", + "bbox": [ + 497, + 373, + 891, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 499, + 415, + 664, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we conduct ablation studies on VG dataset to understand the effectiveness of each component of LLM4SGG, where VS $^3$ is used as the grounding method. Note that row (a) is equivalent to vanilla VS $^3$ . We have the following observations. 1) Effect of using the paraphrased caption: Including the paraphrased caption in addition to the original caption (row (b)) increases the number of triplets $(154\\mathrm{K}\\rightarrow 243\\mathrm{K})$ , resulting in an improved overall performance. This demonstrates that the paraphrased caption alleviates the low-density scene graph issue. 2) Effect of LLM-based parsing: The LLM-based parsing (row (c)) for extracting triplets improves mR@K of row (b). This indicates that the LLM-based parsing increases the number of instances that belong to fine-grained predicates, which in turn alleviates the semantic over-simplification issue. 3) Effect of LLM-based alignment: The LLM-based alignment (row (d)) of entities/predicates in the extracted triplets increases the number of triplets from $243\\mathrm{K}$ to $327\\mathrm{K}$ (row (b) vs (d)), which indicates that the low-density scene graph issue is alleviated. Consequently, R@K and mR@K of row (d) are greater than those of row (b). 4) The fully-fledged approach (row (e)) generally improves R@K and mR@K, showing the best performance in terms of F@K. It is important to highlight that when using the LLM-based parsing, the performance of mR@K significantly increases with a moderate decrease in R@K. This trade-off is attributed to the fact that R@K generally improves when the coarse-grained predicates are dominant [51]. In contrast, our approach, which addresses the semantic over-simplification issue, decreases the number instances that belong to coarse-grained predicates while simultaneously increasing those that belong to fine-grained predicates (Figure 1(b)), which", + "bbox": [ + 497, + 434, + 891, + 867 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28312", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ee00aff165015599ee6dfdea78d6088413e6eb7c93f798dbde1e0d1fc66b273.jpg", + "image_caption": [ + "Figure 4. Performance over various numbers of images used for training VS $^3$ + LLM4SGG." + ], + "image_footnote": [], + "bbox": [ + 98, + 125, + 443, + 214 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in turn results in a substantial improvement in mR@K. We would like to emphasize that the performance in terms of mR@K is crucial in the context of SGG research [1, 8, 48], as fine-grained predicates offer richer information.", + "bbox": [ + 75, + 248, + 469, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Analysis of Data-Efficiency", + "text_level": 1, + "bbox": [ + 76, + 313, + 320, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the effectiveness of LLM4SGG under the lack of available training data, we conduct experiments given a limited number of images. Specifically, among 64K images used for training VS $^3$ and VS $^3$ +LLM4SGG in Table 2, we randomly sample 1K (1.5%), 5K (7.8%), and 10K (15.6%) images with replacement, and train VS $^3$ +LLM4SGG for five times. Figure 4 shows the average performance over various numbers of images used for training VS $^3$ +LLM4SGG along with the variance (in blue area). We observe that when using only 1K images, the performance is slightly inferior to the baseline that used 64K for training (i.e., VS $^3$ ). However, as we increase the number of images used for training to 5K images, we observe a significant improvement in both R@K and mR@K compared with the baseline. This demonstrates that LLM4SGG is data-efficient, as it outperforms the baseline even with only 7.8% of the total images used for training the baseline. Moreover, when further increasing the number of images to 10K, we observe further performance improvements, and when it reaches 64K, which is the same as the number of images used for training the baseline, the performance is the best. In summary, LLM4SGG enables data-efficient model training even with a limited amount of available images for training, thanks to alleviating the semantic over-simplification and low-density scene graph issues.", + "bbox": [ + 73, + 333, + 469, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Quantitative Result on GQA", + "text_level": 1, + "bbox": [ + 76, + 681, + 330, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 4, we additionally conducted experiments on GQA dataset [11]. Please refer to Appendix E.1 for more detailed descriptions on the training and evaluation processes on GQA dataset. The GQA dataset contains twice as many predicates as the Visual Genome dataset and includes complicated predicates (e.g., sitting next to, standing in front of). As a result, when obtaining unlocalized triplets using the conventional WSSGG approach, we observe that 44 out of 100 predicates have a frequency of 0, and the predicate distribution is extremely long-tailed. Consequently, the baseline (i.e., $\\mathrm{VS}^3$ ) exhibits significantly lower performance, especially in terms of mR@K. On the other hand,", + "bbox": [ + 75, + 703, + 469, + 867 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/43e6e5bd543c146321c6a512f648e89f957f8159bb3e61fdda21ef2e5bce20f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodR@50 / 100mR@50 / 100F@50 / 100
Motif (Fully-supervised)28.90 / 33.106.40 / 7.7010.48 / 12.49
VS35.90 / 6.971.60 / 1.812.52 / 2.87
VS3+LLM4SGG8.88 / 10.385.33 / 6.516.66 / 8.00
", + "bbox": [ + 528, + 122, + 859, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Performance comparison on GQA.", + "bbox": [ + 564, + 180, + 824, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our approach shows substantial performance improvements not only in R@K but also in mR@K, thanks to the mitigation of semantic over-simplification and low-density scene graph issues. Please refer to Appendix E.2 for the predicate distribution and performance comparison for each class in GQA dataset. Additionally, please refer to Appendix E.3 for qualitative analyses on GQA dataset.", + "bbox": [ + 497, + 200, + 891, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion & Future work", + "text_level": 1, + "bbox": [ + 497, + 303, + 749, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we focus on the triplet formation process in WSSGG, whose importance is overlooked by previous studies. To alleviate the semantic over-simplification and low-density scene graph issues inherent in the triplet formation process, we propose a new approach, i.e., LLM4SGG, which leverages a pre-trained LLM during the extraction of triplets from the captions, and alignment of entity/predicate classes with those in the target data. It is important to note that construction of these triplets is not required every time to train the SGG model; instead, it is a one-time pre-processing step. In this regard, we contribute to generating enhanced triplets compared to the conventional approach. Moreover, we publish an enhanced SGG dataset constructed by LLM for future studies of SGG. As a result, we outperform baselines in terms of R@K, mR@K and F@K on Visual Genome and GQA datasets. A potential limitation of our work is the reliance on a proprietary blackbox LLM, which can also be costly to use. Hence, in Appendix F, we provide discussions on replacing the LLM with smaller language models.", + "bbox": [ + 497, + 324, + 891, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For future work, an LLM can be used to ground the unlocalized triplets in Step 4. Recently, vision-language representation learning has been developed for transforming visual features into textual features to facilitate the use of visual features as input to an LLM [16, 58]. In this regard, given the visual features of bounding boxes as input, we could ask the LLM to identify relevant bounding boxes based on the textual information of entities within unlocalized triplets using the comprehensive understanding of the context of triplets.", + "bbox": [ + 497, + 595, + 891, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 499, + 737, + 680, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network, No.2022-0-00077, Reasoning, and Inference from Heterogeneous Data, No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University).", + "bbox": [ + 497, + 757, + 891, + 867 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28313", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 125, + 173, + 138 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bashirul Azam Biswas and Qiang Ji. Probabilistic debiasing of scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10429-10438, 2023. 6, 8", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3, 4", + "[3] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6", + "[4] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2, 3", + "[5] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3", + "[6] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 2, 3, 7", + "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 7", + "[8] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 6, 8, 3", + "[9] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3", + "[10] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 3", + "[11] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6, 8, 3, 7", + "[12] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 6, 3", + "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + ], + "bbox": [ + 78, + 147, + 471, + 866 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1, 6, 2, 3, 5, 7", + "[14] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2,3,4", + "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2, 3, 4", + "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8", + "[17] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. arXiv preprint arXiv:2308.06712, 2023. 3", + "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3, 2, 4", + "[19] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 3", + "[20] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4204-4213, 2022. 1, 2, 3, 4, 6", + "[21] Yongzhi Li, Duo Zhang, and Yadong Mu. Visual-semantic matching by exploring high-order attention and distraction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786–12795, 2020. 1", + "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4", + "[24] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3, 4", + "[25] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 1, 2" + ], + "bbox": [ + 502, + 127, + 891, + 866 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28314", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773, 2021. 4", + "[27] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023. 3, 4", + "[28] OpenAI. Gpt-4. Technical report, 2023. 3", + "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3, 2, 4", + "[30] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1", + "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6", + "[32] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1", + "[33] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16393-16402, 2021. 1, 4", + "[34] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 3", + "[35] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 5", + "[36] Damien Teney, Lingqiao Liu, and Anton van Den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2017. 1", + "[37] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 3", + "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 9", + "[39] Sijin Wang, Ruiping Wang, Ziwei Yao, Shiguang Shan, and Xilin Chen. Cross-modal scene graph matching for relationship-aware image-text retrieval. In Proceedings of" + ], + "bbox": [ + 78, + 126, + 469, + 866 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF winter conference on applications of computer vision, pages 1508-1517, 2020. 1", + "[40] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 4", + "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 3, 4, 6", + "[42] Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. Empowering llm to use smartphone for intelligent task automation. arXiv preprint arXiv:2308.15272, 2023. 4", + "[43] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609–6618, 2019. 1, 2, 6, 8", + "[44] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 6", + "[45] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1", + "[46] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 4", + "[47] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8289-8299, 2021. 1, 2, 3, 4, 6", + "[48] Kanghoon Yoon, Kibum Kim, Jinyoung Moon, and Chanyoung Park. Unbiased heterogeneous scene graph generation with relation-aware message passing neural network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3285-3294, 2023. 1, 6, 8, 3", + "[49] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3736-3745, 2020. 2, 3", + "[50] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 1, 3, 4", + "[51] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In European conference on computer vision, pages 409-424. Springer, 2022. 6, 7, 3, 5" + ], + "bbox": [ + 502, + 126, + 891, + 866 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28315", + "bbox": [ + 479, + 890, + 517, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 1", + "[53] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3", + "[54] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Learning to generate language-supervised and open-vocabulary scene graph using pretrained visual-semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915–2924, 2023. 1, 3, 4, 6, 2, 7, 8", + "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 4", + "[56] Yiwu Zhong, Liwei Wang, Jianshu Chen, Dong Yu, and Yin Li. Comprehensive image captioning via scene graph decomposition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 211-229. Springer, 2020. 1", + "[57] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1823-1834, 2021. 1, 3, 4, 6, 2", + "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + ], + "bbox": [ + 78, + 126, + 471, + 553 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28316", + "bbox": [ + 479, + 891, + 517, + 901 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_model.json b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a87566928a006b6f6db56e74f22eef2a6cadd946 --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_model.json @@ -0,0 +1,2004 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.14, + 0.046, + 0.209, + 0.082 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.259, + 0.044, + 0.784, + 0.085 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.161, + 0.79, + 0.202 + ], + "angle": 0, + "content": "LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.227, + 0.826, + 0.276 + ], + "angle": 0, + "content": "Kibum Kim\\(^{1}\\) Kanghoon Yoon\\(^{1}\\) Jaehyeong Jeon\\(^{1}\\) Yeonjun In\\(^{1}\\) Jinyoung Moon\\(^{2}\\) Donghyun Kim\\(^{3}\\) Chanyoung Park\\(^{1*}\\) KAIST \\(^{2}\\)ETRI \\(^{3}\\)Korea University" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.278, + 0.915, + 0.291 + ], + "angle": 0, + "content": "{kb.kim,ykhoon08,wogud405,yeonjun.in,cy PARK}@kaist.ac.kr jymoon@etri.re.kr, d_kim@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.324, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.473, + 0.782 + ], + "angle": 0, + "content": "Weakly-Supervised Scene Graph Generation (WSSGG) research has recently emerged as an alternative to the fully-supervised approach that heavily relies on costly annotations. In this regard, studies on WSSGG have utilized image captions to obtain unlocalized triplets while primarily focusing on grounding the unlocalized triplets over image regions. However, they have overlooked the two issues involved in the triplet formation process from the captions: 1) Semantic over-simplification issue arises when extracting triplets from captions, where fine-grained predicates in captions are undesirably converted into coarse-grained predicates, resulting in a long-tailed predicate distribution, and 2) Low-density scene graph issue arises when aligning the triplets in the caption with entity/predicate classes of interest, where many triplets are discarded and not used in training, leading to insufficient supervision. To tackle the two issues, we propose a new approach, i.e., Large Language Model for weakly-supervised SGG (LLM4SGG), where we mitigate the two issues by leveraging the LLM's in-depth understanding of language and reasoning ability during the extraction of triplets from captions and alignment of entity/predicate classes with target data. To further engage the LLM in these processes, we adopt the idea of Chain-of-Thought and the in-context few-shot learning strategy. To validate the effectiveness of LLM4SGG, we conduct extensive experiments on Visual Genome and GQA datasets, showing significant improvements in both Recall@K and mean Recall@K compared to the state-of-the-art WSSGG methods. A further appeal is that LLM4SGG is data-efficient, enabling effective model training with a small amount of training images. Our code is available on https://github.com/rlqja1107/torch-LLM4SGG" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.792, + 0.21, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.816, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Scene Graph Generation (SGG) is a fundamental task in computer vision, aiming at extracting structured visual" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.325, + 0.892, + 0.527 + ], + "angle": 0, + "content": "knowledge from images [21, 32, 36, 39, 45, 56]. Most existing SGG methods are fully-supervised, i.e., they heavily rely on the ground-truth annotations that involve the class information of entities and predicates as well as the bounding box of entities [19, 48, 50]. However, since creating extensively annotated scene graph datasets is costly, the heavy reliance on these annotations imposes practical limitations on the model training [54]. To mitigate the high cost associated with manual annotations, weakly-supervised scene graph generation (WSSGG) approaches have recently emerged, aiming at training an SGG model without any annotated scene graph dataset. Specifically, the main idea of recent WSSGG methods is to leverage image captions along with associated images, as they can be easily collected from the Web [20, 47, 54, 57]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.529, + 0.893, + 0.771 + ], + "angle": 0, + "content": "The training process of WSSGG model using image captions requires four steps as illustrated in Figure 1(a). Step 1: Preparing an image and its caption. Step 2: Parsing the image caption, i.e., triplets formed as \\(\\langle\\) subject, predicate, object \\(\\rangle\\) are extracted from the image caption through an off-the-shelf parser [30, 43]. Step 3: Aligning the triplets in the caption with entity/predicate classes of interest, i.e., entity (subject, object) and predicate classes in the extracted triplets obtained in Step 2 are aligned with the entity and predicate classes in the target data\\(^{1}\\), respectively. This alignment is based on their synonym/hypernym/hyponym contained in an external knowledge base (KB), e.g., WordNet [25]. Step 4: Grounding unlocalized entities in the extracted triplets, i.e., unlocalized entities (subjects and objects) are matched with relevant image regions generated by a pre-trained object detector, e.g., Faster R-CNN [29]. The localized entities and predicates in the extracted triplets then serve as pseudo-labels for training an SGG model." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.772, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Existing WSSGG approaches mainly focus on Step 4 [20, 33, 47, 57]. For example, in Figure 1(a), their efforts have been focused on grounding the entity person in an unlocalized triplet with an image region that captures the sitting behavior. More precisely, LSwS [47] exploits the con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.854, + 0.225, + 0.867 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.854, + 0.771, + 0.867 + ], + "angle": 0, + "content": "1We use Visual Genome [13] as the target data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28306" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.121, + 0.125, + 0.588, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.218, + 0.466, + 0.229 + ], + "angle": 0, + "content": "(a) Pipeline of Weakly-Supervised SGG" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.128, + 0.844, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.218, + 0.806, + 0.228 + ], + "angle": 0, + "content": "(c) Semantic Oversimplification" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.228, + 0.583, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.319, + 0.433, + 0.328 + ], + "angle": 0, + "content": "(b) Predicate Distribution" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.229, + 0.844, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.319, + 0.801, + 0.33 + ], + "angle": 0, + "content": "(d) Low-Density Scene Graph" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.893, + 0.374 + ], + "angle": 0, + "content": "Figure 1. (a) The pipeline of weakly-supervised SGG. (b) The predicate distribution of unlocalized triplets (Parser+KB vs. Ours). In Parser+KB, the distribution becomes heavily long-tailed, and 12 out of 50 predicates are non-existent. (c) Semantic over-simplification caused by a rule-based parser in Step 2. (d) Low-density scene graph caused by the static structure of KB in Step 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.466 + ], + "angle": 0, + "content": "textual object information to accurately ground the unlocalized entities, leveraging the linguistic structure embedded within the triplets. Another line of research [20] employs a pre-trained vision-language model [15] to reflect the semantic interactions among entities within the image caption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.469, + 0.671 + ], + "angle": 0, + "content": "However, we argue that existing WSSGG approaches overlook the importance of the triplet formation process conducted in Step 2 and Step 3. We identify the two major issues described below, i.e., semantic over-simplication and low-density scene graph, which incur incomplete unlocalized triplets after Step 2 and 3. These incomplete triplets are mostly uninformative predicates with a limited number, and negatively impact the training of an SGG model even when entities are correctly grounded in Step 4. To demonstrate the impact of incomplete unlocalized triplets, we follow the conventional process to extract unlocalized triplets (i.e., Step 1-3), and conduct an examination of triplets obtained from COCO caption dataset, which are generated through Scene Parser [43] in Step 2 and WordNet [25] in Step 3. As a result, we identify the following two issues:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.677, + 0.47, + 0.867 + ], + "angle": 0, + "content": "- Semantic Over-simplification: We find that the standard scene graph parser [43] operating on heuristic rule-based principles commonly used in Step 2 leads to a semantic over-simplification of predicates in extracted triplets. In other words, fine-grained predicates are undesirably converted into coarse-grained predicates, which we refer to as semantic over-simplification. For example, in Figure 1(c), an informative predicate lying on (i.e., fine-grained predicate) in the image caption is converted into a less informative predicate on (i.e., coarse-grained predicate), because the rule-based parser fails to capture the predicate lying on at once, and its heuristic rules fall short of accommodating the diverse range of caption's structure. As a result, the predicate distribution becomes heav-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.892, + 0.518 + ], + "angle": 0, + "content": "ily long-tailed, in which coarse-grained predicates (e.g., with, on, in) greatly outnumber fine-grained predicates (e.g., parked on, covered in) (Figure 1(b)). To make the matter worse, numerous fine-grained predicates eventually end up in a frequency of 0, even though they are originally present in the captions. Specifically, 12 out of 50 predicates are non-existent, which means that these 12 predicates can never be predicted, since the model is not trained on these predicates at all." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.519, + 0.893, + 0.858 + ], + "angle": 0, + "content": "- Low-Density Scene Graph: We find that the KB-based triplet alignment in Step 3 leads to low-density scene graphs, i.e., the number of remaining triplets after Step 3 is small. We attribute the low-density scene graphs primarily to the utilization of KB in Step 3. Specifically, a triplet is discarded if any of the three components (i.e., subject, predicate, object) or their synonym/hypernym/hyponym within the triplet fail to align with the entity or predicate classes in the target data. For example, in Figure 1(d), the triplet \\(\\langle\\) elephant, carrying, \\(\\log \\rangle\\) is discarded because \\(\\log\\) does not exist in Visual Genome dataset nor its synonym/hypernym, even if elephant and carrying do exist. In Table 1, we report the number of triplets and images in Visual Genome dataset, which is a common benchmark dataset used in fully-supervised SGG approaches, and COCO caption dataset, which is a common benchmark dataset used in weakly-supervised SGG approaches. We observe that on average Visual Genome dataset contains 7.1 triplets (i.e., \\(405\\mathrm{K} / 57\\mathrm{K}\\)) per image (See Table 1(a)), while COCO dataset contains only 2.4 triplets (i.e., \\(154\\mathrm{K} / 64\\mathrm{K}\\)) per image (See Table 1(b)). This indicates that existing WSSGG approaches suffer from the lack of sufficient supervision per image, leading to poor generalization and performance degradation [47, 49]. In summary, relying on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28307" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.123, + 0.416, + 0.201 + ], + "angle": 0, + "content": "
DatasetHow to annotate# Triplet# Image
Fully-Supervised approach
(a) Visual GenomeManual405K57K
Weakly-Supervised approach
(b) COCO CaptionParser+KB154K64K
(c) COCO CaptionLLM344K64K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.202, + 0.406, + 0.215 + ], + "angle": 0, + "content": "Table 1. Comparison of scene graph density." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.469, + 0.277 + ], + "angle": 0, + "content": "the static structured knowledge of KB is insufficient to cover the semantic relationships among a wide a range of words, which incurs the low-density scene graph after Step 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.277, + 0.47, + 0.736 + ], + "angle": 0, + "content": "To alleviate the semantic over-simplification and the low-density scene graph issues, we propose a new approach, namely Large Language Model for weakly-supervised SGG (LLM4SGG) that adopts a pre-trained Large Language Model (LLM), which has shown remarkable transferability to various downstream tasks in NLP such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Inspired by the idea of Chain-of-Thought² (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into two chains, each of which replaces the rule-based parser in Step 2 (i.e., Chain-1) and the KB in Step 3 (i.e., Chain-2). More precisely, we design a prompt for extracting triplets from a caption, and ask the LLM to extract triplets formed as (Chain-1). We expect that the predicates extracted based on a comprehensive understanding of the caption's context via LLM are semantically rich, thereby alleviating the semantic over-simplification issue. Besides, to alleviate the low-density scene graph issue, we additionally incorporate a paraphrased version of the original caption. To this end, we further design a prompt for paraphrasing the original caption and extracting more triplets from the paraphrased caption. However, entities and predicates in the triplets obtained after Chain-1 are not yet aligned with the target data. Hence, we design a another prompt to align them with entity/predicate classes of interest, and ask the LLM to align them with semantically relevant lexeme included in a predefined lexicon, which is the set of vocabularies that are present in the target data (Chain-2). To further engage the LLM in the reasoning process of Chain-1 and Chain-2, we employ the in-context few-shot learning that incorporates a few input-output examples within the prompt, enabling the LLM to perform the task without the need for fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.736, + 0.469, + 0.832 + ], + "angle": 0, + "content": "To validate the effectiveness of LLM4SGG, we apply it to state-of-the-art WSSGG methods [54, 57]. Through extensive experiments, we show that LLM4SGG significantly enhances the performance of existing WSSGG methods in terms of mean Recall@K and Recall@K performance on Visual Genome and GQA datasets by alleviating the semantic over-simplification and the low-density scene graph" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.127, + 0.892, + 0.193 + ], + "angle": 0, + "content": "(See Table 1(c) where the number of triplets increased to 334K). A further appeal of LLM4SGG is that it is data-efficient, i.e., it outperforms state-of-the-art baselines even with a small amount of training images, verifying the effectiveness of LLM4SGG." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.194, + 0.854, + 0.207 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.208, + 0.891, + 0.247 + ], + "angle": 0, + "content": "- We identify two major issues overlooked by existing WSSGG studies, i.e., semantic over-simplification and low-density scene graph." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.248, + 0.891, + 0.314 + ], + "angle": 0, + "content": "- We leverage an LLM along with the CoT strategy and the in-context few-shot learning technique to extract informative triplets without the need for fine-tuning the LLM. To the best of our knowledge, we are the first to leverage an LLM for the SGG task." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.315, + 0.891, + 0.369 + ], + "angle": 0, + "content": "- LLM4SGG outperforms the state-of-the-art WSSGG methods, especially in terms of mR@K, demonstrating its efficacy in addressing the long-tail problem in WSSGG for the first time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.381, + 0.649, + 0.395 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.715 + ], + "angle": 0, + "content": "Weakly-Supervised Scene Graph Generation (WSSGG). The WSSGG task aims to train an SGG model without relying on an annotated scene graph dataset. To achieve this, most WSSGG studies [20, 47, 54, 57] utilize image captions and ground unlocalized triplets with image regions. Specifically, VSPNet [49] proposes the iterative graph alignment algorithm to reflect the high-order relations between unlocalized triplets and image regions. SGNLS [57] uses a pre-trained object detector [29] to ground the entities in unlocalized triplets, which share the same classes with the output of object detectors. In addition to the information derived from the object detector, [20] employs a pre-trained vision-language model [15] to capture the semantic interactions among objects. VS3 [54] uses a grounding-based object detector [18], which calculates the similarity between the entity text in unlocalized triplet and image region, thereby grounding the unlocalized triplets. However, these methods overlook the triplet formation process that leads to the semantic over-simplification (Step 2) and the low-density scene graph (Step 3). In this regard, existing methods result in a sub-optimal performance even when unlocalized triplets are correctly grounded in Step 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Large Language Model (LLM). LLMs have demonstrated remarkable transferability to various downstream tasks such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Specifically, GPT-3 (175B) [2] stands as a cornerstone to break the line of numerous language tasks. Inspired by GPT-3, PaLM (540B) [5], LLaMA (65B) [38], OPT (175B) [53], and LaMDA (137B) [37] have been subsequently introduced. More recently, advanced GPT models (e.g., GPT-4 [28], ChatGPT [27]) fine-tuned with human feedback have gained prominence and widely applied for diverse applications, e.g., planner of tools [9, 24], mobile" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.843, + 0.469, + 0.867 + ], + "angle": 0, + "content": "2We use the CoT strategy as a means to arrive at an answer in a stepwise manner, which differs from the Chain-of-Thought prompting." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28308" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.126, + 0.471, + 0.18 + ], + "angle": 0, + "content": "task automation [42]. In this work, we employ the power of LLM (i.e., ChatGPT) to alleviate the two issues, i.e., semantic over-simplification and low-density scene graph, in the context of the WSSGG task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.473, + 0.43 + ], + "angle": 0, + "content": "In-Context Few-shot Learning. In-context few-shot learning incorporates a few input-output examples related to a target task, conditioning the LLM on the context of examples. Specifically, GPT-3 [2] pioneered the concept of in-context learning to facilitate an LLM as a versatile model on diverse tasks. This breakthrough has proliferated a plethora of research to leverage the in-context few-shot learning for various tasks [24, 26, 46, 55]. More precisely, Chameleon [24] integrates a few examples to enhance its understanding of tool planning task. [26] utilizes positive and negative examples related to questions for question generation tasks. ReAct [46] incorporates examples of reasoning with action for solving decision-making tasks. Inspired by recent in-context few-shot learning approaches, we provide a few examples to LLMs to help 1) understand the process of triplet extraction from a caption (i.e., Step 2), and 2) align the entity/predicate classes with the target data (i.e., Step 3) in the context of the WSSGG task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.17, + 0.457 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.671 + ], + "angle": 0, + "content": "In this section, we describe LLM4SGG in detail. We would like to emphasize that LLM4SGG mainly focuses on the triplet formation process conducted in Step 2 (Parsing) and Step 3 (aligning), while existing WSSGG approaches mainly focus on Step 4 (grounding). We start by presenting the problem formulation of WSSGG (Section 3.1), followed by the prompt configuration (Section 3.2). Next, we introduce how LLMs are adopted to address the two issues of conventional WSSGG approaches when parsing the image caption (Section 3.3) and aligning the triplets in captions with entity/predicate classes of interest (Section 3.4). Finally, we ground the unlocalized triplets by associating them with bounding boxes (i.e., image regions) and train the SGG model using the localized triplets (Section 3.5). The overall pipeline of LLM4SGG is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.281, + 0.695 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.868 + ], + "angle": 0, + "content": "In the fully supervised SGG task, we aim to detect a scene graph \\(\\mathbf{G}_f = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_f}\\) that consists of triplets given an image \\(\\mathcal{I}\\), where \\(N_{f}\\) is the number of triplets in the image. \\(\\mathbf{s}_i\\) and \\(\\mathbf{o}_i\\) denote the \\(i^{\\mathrm{th}}\\) subject and the object, respectively, whose bounding boxes are \\(\\mathbf{s}_{i,b}\\), \\(\\mathbf{o}_{i,b}\\), and entity classes are \\(\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}\\in \\mathcal{C}_e\\), where \\(\\mathcal{C}_e\\) is the set of predefined entity classes in the target data. \\(\\mathbf{p}_i\\) denotes the predicate between \\(\\mathbf{s}_i\\) and \\(\\mathbf{o}_i\\), and its class is \\(\\mathbf{p}_{i,c}\\in \\mathcal{C}_p\\), where \\(\\mathcal{C}_p\\) is the set of predefined predicate classes in the target data. By using the ground truth scene graphs as supervision, fully supervised SGG approaches train an SGG model \\(\\mathcal{T}_{\\theta}:\\mathcal{I}\\to \\mathbf{G}_{f}\\), which maps an image to a scene graph." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.126, + 0.892, + 0.342 + ], + "angle": 0, + "content": "In the weakly supervised SGG task, we aim to generate a scene graph when the ground truth scene graph is not given, i.e., there are no bounding boxes and entity/predicate class information. Instead, existing WSSGG approaches [20, 33, 47, 54, 57] use image captions along with associated images to produce scene graphs, i.e., localized triplets. More precisely, they extract a set of triplets \\(\\mathbf{G}_w = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_w}\\) from the image captions, where \\(N_{w}\\) is the number of triplets extracted from the captions. However, while the extracted triplets contain the class information (i.e., \\(\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}\\) and \\(\\mathbf{p}_{i,c}\\)), they are unlocalized, since bounding boxes \\(\\mathbf{s}_{i,b}\\) and \\(\\mathbf{o}_{i,b}\\) are not included in the caption. Therefore, it is essential to perform the grounding step to associate the unlocalized triplets with bounding boxes. Once we have obtained the localized triplets, we can apply the conventional SGG training scheme, i.e., \\(\\mathcal{T}_{\\theta}:\\mathcal{I}\\rightarrow \\mathbf{G}_{f}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.893, + 0.478 + ], + "angle": 0, + "content": "In this paper, our focus is to address the semantic oversimplification and low-density scene graph issues regarding the unlocalized triplets \\(\\mathbf{G}_w\\), which has been overlooked in existing WSSGG studies. Specifically, we aim to produce an enhanced \\(\\mathbf{G}_w\\) by refining the process of scene graph dataset construction via an LLM. This refinement includes the triplet extraction step from the caption (Step 2) and the alignment of entity/predicate classes (Step 3), leveraging the LLM's comprehensive understanding of language and reasoning ability." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.484, + 0.707, + 0.499 + ], + "angle": 0, + "content": "3.2. Prompt Configuration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.655 + ], + "angle": 0, + "content": "In fact, it is not a trivial task for an LLM to immediately generate triplets from a caption whose entities and predicates are aligned with entity/predicate classes of interest, as such a task is a novel task for the LLM. Inspired by the idea of the Chain-of-thought (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into the following two chains: Chain-1 - Extracting triplets from captions. Chain-2 - Aligning entities and predicates with the entity/predicate classes of interest. To carefully design each chain, we define the LLM function, i.e., \\(LLM(\\cdot)\\), with the following prompt input:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.659, + 0.887, + 0.687 + ], + "angle": 0, + "content": "\\[\nO u t p u t = L L M (\\underbrace {\\text {T a s k d e s c r i p t i o n , I n - c o n t e x t e x a m p l e s , A c t u a l q u e s t i o n}} _ {\\text {P r o m p t i n p u t}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.868 + ], + "angle": 0, + "content": "where \\( LLM(\\cdot) \\) is the decoder of the LLM, generating the (1) Output given the prompt input. The prompt input consists of three components in a sequence: 1) task description, i.e., the delineation of the task that we intend to perform, 2) in-context examples, i.e., sample questions and answers related to the task at hand, 3) actual question, i.e., an inquiry from which we intend to derive the answer. Note that in-context examples is closely related to the in-context few-shot learning [2, 40, 41], which is shown to enhance the LLM's understanding of the task. Note that the above configuration of the prompt input is applied to the triplet extraction (Chain-1) (Section 3.3) and the alignment of entity/predicate classes (Chain-2) (Section 3.4)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28309" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.152, + 0.124, + 0.817, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 2. The pipeline of LLM4SGG. Given an image with its caption, we use an LLM to extract triplets from the original caption (Step 2-1) and the paraphrased caption (Step 2-2). Then, we align the entity/predicate classes within the extracted triplets with semantically similar lexeme in the target data via an LLM (Step 3), obtaining the unlocalized triplets. Lastly, we ground the unlocalized triplets over image regions (Step 4) followed by the training of an SGG model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.407, + 0.471, + 0.437 + ], + "angle": 0, + "content": "3.3. Chain-1. Triplet Extraction via LLM (Step 2 in Figure 2)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.507 + ], + "angle": 0, + "content": "Based on the LLM's comprehensive understanding of the context of an image caption, we aim to extract triplets from the caption. As discussed in Section 1, we use not only the original caption but also a paraphrased caption generated by the LLM to address the low-density scene graph issue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.846 + ], + "angle": 0, + "content": "Extracting triplets from a paraphrased caption (Step 2-2). To extract triplets from a paraphrased caption, we inform the LLM about the task at hand by presenting the following prompt: FROM THE GIVEN SENTENCE, THE TASK IS TO EXTRACT MEANINGFUL TRIPLETS FORMED AS \\(\\langle\\) SUBJECT, PREDICATE, OBJECT \\(\\rangle\\) (i.e., Task description in Equation 1). We then instruct the LLM to follow the two steps, i.e., paraphrasing step and triplet extraction step. To help the LLM understand the process of performing the two steps, we present few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation \\(1^{3}\\)), which involves a manual construction of questions and corresponding answers to the paraphrasing and the triplet extraction steps. That is, for given a caption, we show the LLM how we expect a paraphrased caption and the extracted triplets would look like (e.g., Given \"Four clocks sitting on a floor next to a woman's feet,\" we show the LLM that a paraphrased sentence would be \"Four clocks are placed on the floor beside a woman's feet,\" and extracted triplets would be \\(\\langle\\) clocks, placed on, floor \\(\\rangle\\) and \\(\\langle\\) clocks, beside, feet \\(\\rangle\\)). Lastly, we show the caption of our interest to the LLM, and let the LLM extract triplets from the caption (i.e., Actual question in Equation 1). Please refer to Figure 2 right (bottom) for an example of the prompt input used in Step 2-2." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.854, + 0.398, + 0.867 + ], + "angle": 0, + "content": "3We use captions in COCO caption dataset for examples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.408, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Extracting triplets from the original caption (Step 2-1). As extracting triplets from the original caption does not involve the caption paraphrasing step, we exclude it from the prompt used in Step 2-2. Please refer to Figure 2 right (top) for an example of the prompt input used in Step 2-1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.477, + 0.892, + 0.545 + ], + "angle": 0, + "content": "In summary, we obtain triplets from both the original and paraphrased captions after Step 2-1 and Step 2-2, respectively, which in turn alleviates the semantic oversimplification issue of predicates and the low-density scene graph issue." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.586 + ], + "angle": 0, + "content": "3.4. Chain-2. Alignment of Classes in Triplets via LLM (Step 3 in Figure 2)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.659 + ], + "angle": 0, + "content": "The entities (i.e., subject and object) and predicates within the triplets obtained from Step 2 described in Section 3.3 are not yet aligned with the target data. Based on the semantic reasoning ability of the LLM, we aim to align them with the semantically relevant lexeme in the target data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Aligning entities in the triplets with entity classes of interest. We instruct the LLM with the following prompt: GIVEN THE LEXEME \\(\\{\\mathsf{ENTITY}\\}\\), FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED ENTITY LEXICON, where the predefined entity lexicon is \\(\\mathcal{C}_e\\) (i.e., Task description in Equation 1). Similar to Section 3.3, we present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding hierarchical relationships such as pigeon being semantically relevant to bird, and singular-plural relationships such as surfboards being semantically relevant to surfboard. Lastly, we show the entity of our interest to the LLM (i.e., Actual question in Equation 1), which enables the LLM to generate an answer by finding the most" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28310" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.126, + 0.47, + 0.154 + ], + "angle": 0, + "content": "semantically relevant entity in \\(\\mathcal{C}_e\\). Please refer to Table 7 in Appendix A.1 for an example of the prompt input." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.156, + 0.473, + 0.466 + ], + "angle": 0, + "content": "Aligning predicates in the triplets with predicate classes of interest. Likewise, we instruct the LLM with the following prompt: GIVEN THE LEXEME \\(\\{\\mathrm{PREDICATE}\\}\\), FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED Predicate Lexicon, where the predefined predicate lexicon is \\(C_p\\) (i.e., Task description in Equation 1). We also present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding tense relationships such as the lies on being semantically relevant to lying on, and positional relationship such as next to being semantically relevant to near. Lastly, we show the predicate of our interest to the LLM (i.e., Actual question in Equation 1). Please refer to Table 8 in Appendix A.1 for an example of the prompt input. Furthermore, regarding the alignment of classes within a large predefined lexicon, please refer to Appendix A.2. Note that as our main focus in this work is on developing a framework for utilizing LLMs in weakly-supervised SGG rather than finding a specific prompt design that works the best, we tested with a single prompt design. However, since prompt designs are crucial in leveraging LLMs, we plan to explore various designs in our future work." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.467, + 0.473, + 0.634 + ], + "angle": 0, + "content": "After performing Chain-1 (Section 3.3) and Chain-2 (Section 3.4), we obtain intermediate unlocalized triplets \\(\\hat{\\mathbf{G}}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{\\hat{N}_w}\\), where \\(\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\{\\mathcal{C}_e \\cup \\text{None}\\}\\) and \\(\\mathbf{p}_{i,c} \\in \\{\\mathcal{C}_p \\cup \\text{None}\\}\\). It is worth noting that if there is no semantically relevant lexeme, we request the LLM to generate None as the answer, due to the fact that the entity/predicate classes in the target data may not cover a wide range of entities/predicate. Similar to the conventional approach, we discard a triplet if any of its three components (i.e., subject, predicate and object) is None. Lastly, we obtain the final unlocalized triplets \\(\\mathbf{G}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{N_w}(N_w \\leq \\hat{N}_w)\\), where \\(\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\mathcal{C}_e\\) and \\(\\mathbf{p}_{i,c} \\in \\mathcal{C}_p\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.639, + 0.236, + 0.654 + ], + "angle": 0, + "content": "3.5. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.797 + ], + "angle": 0, + "content": "Given the final unlocalized triplets \\(\\mathbf{G}_w\\), we ground them over relevant image regions to get localized triplets, meaning that we obtain \\(\\mathbf{s}_{i,b}\\) and \\(\\mathbf{o}_{i,b}\\). To this end, we employ two state-of-the-art grounding methods, i.e., SGNLS [57] and VS3 [54]. Please refer to Appendix A.3 for more detail about how each method performs grounding. After grounding \\(\\mathbf{G}_w\\), we obtain localized triplets and use them as pseudo-labels for training a supervised SGG model. Please refer to Appendix A.4 for more detail about the model training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.803, + 0.203, + 0.819 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Datasets. To train an SGG model without an annotated scene graph dataset, we use three caption datasets: COCO caption [3], Conceptual (CC) caption [31], and Visual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.127, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Genome (VG) caption [44]. For fair comparisons, we use the same set of images that have been utilized in previous WSSGG studies [20, 47, 54, 57], leading to the utilization of 64K images on COCO caption dataset, 145K images on CC caption dataset, and 57K images on VG caption dataset. To evaluate the trained SGG model, we employ the widely used Visual Genome (VG) dataset [13] and GQA dataset [11]. The VG dataset contains the ground-truth localized triplet information annotated by humans. We follow the standard split of VG [44], which consist of 150 entity classes and 50 predicate classes. For the GQA dataset used in the SGG task, we follow the same pre-processing step of a previous SGG study [8], which involves selecting top-200 frequent entity classes and top-100 frequent predicate classes. In both datasets, \\(30\\%\\) of the total images are used for evaluation. Please refer to Appendix C.1 for more details regarding the datasets. Note that we mainly use the COCO caption dataset for analysis of LLM4SGG throughout this paper, i.e., CC and VG caption datasets are exclusively used in quantitative result on VG (Section 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.399, + 0.893, + 0.643 + ], + "angle": 0, + "content": "Evaluation metrics. Recent fully-supervised SGG studies [1, 48, 51] have emphasized improving the accuracy of predictions for fine-grained predicates rather than coarse-grained predicates, since the former construct richer scene graphs. As a result, they commonly use mean Recall@K (mR@K) that computes the average of Recall@K (R@K) across all predicates. In line with the recent emphasis on fine-grained predicates, we incorporate both mR@K and R@K in our evaluation, whereas previous WSSGG studies [20, 47, 54] mainly rely on the R@K metric alone. Moreover, we report F@K, which is the harmonic average of R@K and mR@K to jointly consider R@K and mR@K, following previous SGG studies [12, 51]. Regarding the evaluation task, we follow previous WSSGG studies and adopt the Scene Graph Detection (SGDet) task, where both the ground-truth bounding box and the entity class information are not provided. Please refer to Appendix C.2 for more detail regarding the task." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.645, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Baselines. Please refer to Appendix C.3 for details regarding the baselines." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Implementation details. Please refer to Appendix C.4 regarding the implementation details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.71, + 0.741, + 0.726 + ], + "angle": 0, + "content": "4.1. Quantitative Result on VG" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Table 2 shows the performance of baseline models and those when LLM4SGG is applied. We have the following observations based on COCO caption dataset: 1) Applying LLM4SGG to SGNLS and VS3 improves the performance in terms of R@K and mR@K, which demonstrates the effectiveness of the triplet formation through LLM4SGG. Notably, LLM4SGG significantly improves mR@K, implying that LLM4SGG effectively alleviates the long-tailed problem in WSSGG. This can be clearly seen in Figure 3, which shows the performance gain on fine" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.891, + 0.517, + 0.903 + ], + "angle": 0, + "content": "28311" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.167, + 0.119, + 0.803, + 0.268 + ], + "angle": 0, + "content": "
MethodDatasetR@50R@100mR@50mR@100F@50F@100
Motif (CVPR'18) - Fully-supervisedVG31.8936.366.387.5710.63 / 12.5312.53
LSWS (CVPR'21)COCO Caption3.293.693.273.663.283.67
SGNLS (ICCV'21)3.804.462.512.783.023.43
SGNLS (ICCV'21)+LLM4SGG5.09+1.295.97+1.514.08+1.574.49+1.714.53+1.515.13+1.70
Li et al (MM'22)6.407.331.731.982.723.12
VS3(CVPR'23)6.608.012.883.254.014.62
VS3(CVPR'23)+LLM4SGG8.91+2.3110.43+2.427.11+4.238.18+4.937.91+3.909.17+4.55
VS3(CVPR'23)+Rwt4.255.045.175.994.675.47
VS3(CVPR'23)+Rwt+LLM4SGG5.10+0.856.34+1.308.42+3.259.90+3.916.35+1.697.73+2.26
VS3(CVPR'23)CC Caption6.698.201.732.042.753.27
VS3(CVPR'23)+LLM4SGG9.47+2.7810.69+2.495.40+3.676.09+4.056.88+4.137.76+4.49
VS3(CVPR'23)VG Caption14.5418.482.803.794.706.29
VS3(CVPR'23)+LLM4SGG18.40+3.8622.28+3.806.26+3.467.60+3.819.34+4.6411.33+5.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.272, + 0.893, + 0.299 + ], + "angle": 0, + "content": "Table 2. Performance comparisons on the SGDet task. The best performance among WSSGG models within each dataset is in bold. The red numbers indicate the absolute performance improvement after applying LLM4SGG. Rwt denotes using the reweighting strategy [51]." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.312, + 0.434, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.437 + ], + "angle": 0, + "content": "Figure 3. Per class performance (Bar: number of predicate instances, Line: Recall@100)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.824 + ], + "angle": 0, + "content": "grained predicates. 2) \\(\\mathrm{VS}^3 +\\mathrm{Rwt} + \\mathrm{LLM4SGG}\\) further improves mR@K of \\(\\mathrm{VS}^3 +\\mathrm{Rwt}\\). We attribute this to the fact that the conventional approach generates a limited number of fine-grained predicates, which makes the reweighting strategy less effective within \\(\\mathrm{VS}^3\\). Especially, non-existent predicates can never be predicted even when the reweighting strategy is applied. On the other hand, LLM4SGG increases the number of instances that belong to fine-grained predicates, which is advantageous for the reweighting strategy. For the per class performance comparison over the reweighting stregty, please refer to Appendix D.1. 3) The performance gain obtained from applying LLM4SGG is greater on \\(\\mathrm{VS}^3\\) (i.e., \\(\\mathrm{VS}^3 +\\mathrm{LLM4SGG}\\)) than on SGNLS (i.e., SGNLS+LLM4SGG). The major reason lies in the difference in how SGNLS and \\(\\mathrm{VS}^3\\) make use of the pool of 344K unlocalized triplets obtained through LLM4SGG. Specifically, in the grounding process of SGNLS, we observe that 100K out of 344K unlocalized triplets (i.e., \\(29\\%\\)) fail to be grounded, and thus not used for training. On the other hand, \\(\\mathrm{VS}^3\\) successfully grounds all 344K unlocalized triplets and fully utilize them for training, allowing it to fully enjoy the effectiveness of LLM4SGG. This indicates that LLM4SGG makes synergy when paired with a grounding method that is capable of fully utilizing the unlocalized triplets. For more details regarding the impact of grounding methods, please refer to Appendix B. 4) Regarding the performance comparison with a fully-supervised approach (i.e., Motif), please refer to Appendix D.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.867 + ], + "angle": 0, + "content": "Furthermore, we observe that applying LLM4SGG to CC and VG caption datasets also significantly improves performance, demonstrating the effectiveness of LLM4SGG." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.311, + 0.866, + 0.37 + ], + "angle": 0, + "content": "
RowPCLPLA# TripletR@50 / 100mR@50 / 100F@50 / 100
(a)154K6.60 / 8.012.88 / 3.254.01 / 4.62
(b)✓243K9.46 / 11.223.43 / 3.925.03 / 5.81
(c)✓✓256K8.42 / 9.855.99 / 6.957.00 / 8.15
(d)✓✓327K11.76 / 13.383.50 / 4.055.39 / 6.22
(e)✓✓✓344K8.91 / 10.437.11 / 8.187.91 / 9.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Table 3. Ablation studies. (PC: Using Paraphrased Caption in addition to the original caption / LP: LLM-based Parsing / LA: LLM-based Alignment)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.665, + 0.43 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.893, + 0.868 + ], + "angle": 0, + "content": "In Table 3, we conduct ablation studies on VG dataset to understand the effectiveness of each component of LLM4SGG, where VS\\(^3\\) is used as the grounding method. Note that row (a) is equivalent to vanilla VS\\(^3\\). We have the following observations. 1) Effect of using the paraphrased caption: Including the paraphrased caption in addition to the original caption (row (b)) increases the number of triplets \\((154\\mathrm{K}\\rightarrow 243\\mathrm{K})\\), resulting in an improved overall performance. This demonstrates that the paraphrased caption alleviates the low-density scene graph issue. 2) Effect of LLM-based parsing: The LLM-based parsing (row (c)) for extracting triplets improves mR@K of row (b). This indicates that the LLM-based parsing increases the number of instances that belong to fine-grained predicates, which in turn alleviates the semantic over-simplification issue. 3) Effect of LLM-based alignment: The LLM-based alignment (row (d)) of entities/predicates in the extracted triplets increases the number of triplets from \\(243\\mathrm{K}\\) to \\(327\\mathrm{K}\\) (row (b) vs (d)), which indicates that the low-density scene graph issue is alleviated. Consequently, R@K and mR@K of row (d) are greater than those of row (b). 4) The fully-fledged approach (row (e)) generally improves R@K and mR@K, showing the best performance in terms of F@K. It is important to highlight that when using the LLM-based parsing, the performance of mR@K significantly increases with a moderate decrease in R@K. This trade-off is attributed to the fact that R@K generally improves when the coarse-grained predicates are dominant [51]. In contrast, our approach, which addresses the semantic over-simplification issue, decreases the number instances that belong to coarse-grained predicates while simultaneously increasing those that belong to fine-grained predicates (Figure 1(b)), which" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.125, + 0.444, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.217, + 0.47, + 0.243 + ], + "angle": 0, + "content": "Figure 4. Performance over various numbers of images used for training VS\\(^3\\) + LLM4SGG." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.47, + 0.304 + ], + "angle": 0, + "content": "in turn results in a substantial improvement in mR@K. We would like to emphasize that the performance in terms of mR@K is crucial in the context of SGG research [1, 8, 48], as fine-grained predicates offer richer information." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.321, + 0.329 + ], + "angle": 0, + "content": "4.3. Analysis of Data-Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.334, + 0.471, + 0.675 + ], + "angle": 0, + "content": "To assess the effectiveness of LLM4SGG under the lack of available training data, we conduct experiments given a limited number of images. Specifically, among 64K images used for training VS\\(^3\\) and VS\\(^3\\)+LLM4SGG in Table 2, we randomly sample 1K (1.5%), 5K (7.8%), and 10K (15.6%) images with replacement, and train VS\\(^3\\)+LLM4SGG for five times. Figure 4 shows the average performance over various numbers of images used for training VS\\(^3\\)+LLM4SGG along with the variance (in blue area). We observe that when using only 1K images, the performance is slightly inferior to the baseline that used 64K for training (i.e., VS\\(^3\\)). However, as we increase the number of images used for training to 5K images, we observe a significant improvement in both R@K and mR@K compared with the baseline. This demonstrates that LLM4SGG is data-efficient, as it outperforms the baseline even with only 7.8% of the total images used for training the baseline. Moreover, when further increasing the number of images to 10K, we observe further performance improvements, and when it reaches 64K, which is the same as the number of images used for training the baseline, the performance is the best. In summary, LLM4SGG enables data-efficient model training even with a limited amount of available images for training, thanks to alleviating the semantic over-simplification and low-density scene graph issues." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.683, + 0.332, + 0.698 + ], + "angle": 0, + "content": "4.4. Quantitative Result on GQA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.868 + ], + "angle": 0, + "content": "In Table 4, we additionally conducted experiments on GQA dataset [11]. Please refer to Appendix E.1 for more detailed descriptions on the training and evaluation processes on GQA dataset. The GQA dataset contains twice as many predicates as the Visual Genome dataset and includes complicated predicates (e.g., sitting next to, standing in front of). As a result, when obtaining unlocalized triplets using the conventional WSSGG approach, we observe that 44 out of 100 predicates have a frequency of 0, and the predicate distribution is extremely long-tailed. Consequently, the baseline (i.e., \\(\\mathrm{VS}^3\\)) exhibits significantly lower performance, especially in terms of mR@K. On the other hand," + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.123, + 0.86, + 0.176 + ], + "angle": 0, + "content": "
MethodR@50 / 100mR@50 / 100F@50 / 100
Motif (Fully-supervised)28.90 / 33.106.40 / 7.7010.48 / 12.49
VS35.90 / 6.971.60 / 1.812.52 / 2.87
VS3+LLM4SGG8.88 / 10.385.33 / 6.516.66 / 8.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.181, + 0.825, + 0.193 + ], + "angle": 0, + "content": "Table 4. Performance comparison on GQA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.296 + ], + "angle": 0, + "content": "our approach shows substantial performance improvements not only in R@K but also in mR@K, thanks to the mitigation of semantic over-simplification and low-density scene graph issues. Please refer to Appendix E.2 for the predicate distribution and performance comparison for each class in GQA dataset. Additionally, please refer to Appendix E.3 for qualitative analyses on GQA dataset." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.304, + 0.75, + 0.318 + ], + "angle": 0, + "content": "5. Conclusion & Future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.893, + 0.595 + ], + "angle": 0, + "content": "In this work, we focus on the triplet formation process in WSSGG, whose importance is overlooked by previous studies. To alleviate the semantic over-simplification and low-density scene graph issues inherent in the triplet formation process, we propose a new approach, i.e., LLM4SGG, which leverages a pre-trained LLM during the extraction of triplets from the captions, and alignment of entity/predicate classes with those in the target data. It is important to note that construction of these triplets is not required every time to train the SGG model; instead, it is a one-time pre-processing step. In this regard, we contribute to generating enhanced triplets compared to the conventional approach. Moreover, we publish an enhanced SGG dataset constructed by LLM for future studies of SGG. As a result, we outperform baselines in terms of R@K, mR@K and F@K on Visual Genome and GQA datasets. A potential limitation of our work is the reliance on a proprietary blackbox LLM, which can also be costly to use. Hence, in Appendix F, we provide discussions on replacing the LLM with smaller language models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.732 + ], + "angle": 0, + "content": "For future work, an LLM can be used to ground the unlocalized triplets in Step 4. Recently, vision-language representation learning has been developed for transforming visual features into textual features to facilitate the use of visual features as input to an LLM [16, 58]. In this regard, given the visual features of bounding boxes as input, we could ask the LLM to identify relevant bounding boxes based on the textual information of entities within unlocalized triplets using the comprehensive understanding of the context of triplets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.738, + 0.681, + 0.753 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.868 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network, No.2022-0-00077, Reasoning, and Inference from Heterogeneous Data, No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28313" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.125, + 0.174, + 0.139 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Bashirul Azam Biswas and Qiang Ji. Probabilistic debiasing of scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10429-10438, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.472, + 0.261 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.262, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[3] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[4] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.375, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.625 + ], + "angle": 0, + "content": "[8] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.627, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[9] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[10] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[11] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6, 8, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[12] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.469, + 0.867 + ], + "angle": 0, + "content": "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.148, + 0.472, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.128, + 0.892, + 0.165 + ], + "angle": 0, + "content": "Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1, 6, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.166, + 0.892, + 0.251 + ], + "angle": 0, + "content": "[14] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2,3,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.366, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[17] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. arXiv preprint arXiv:2308.06712, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.416, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[20] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4204-4213, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.666 + ], + "angle": 0, + "content": "[21] Yongzhi Li, Duo Zhang, and Yadong Mu. Visual-semantic matching by exploring high-order attention and distraction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786–12795, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.666, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.778 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.779, + 0.892, + 0.84 + ], + "angle": 0, + "content": "[24] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.841, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[25] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.128, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28314" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[27] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.363, + 0.218 + ], + "angle": 0, + "content": "[28] OpenAI. Gpt-4. Technical report, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.469, + 0.268 + ], + "angle": 0, + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[30] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.407 + ], + "angle": 0, + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.409, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[32] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.521 + ], + "angle": 0, + "content": "[33] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16393-16402, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.524, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[34] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[35] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.701 + ], + "angle": 0, + "content": "[36] Damien Teney, Lingqiao Liu, and Anton van Den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.763 + ], + "angle": 0, + "content": "[37] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.766, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.829, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[39] Sijin Wang, Ruiping Wang, Ziwei Yao, Shiguang Shan, and Xilin Chen. Cross-modal scene graph matching for relationship-aware image-text retrieval. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.127, + 0.892, + 0.153 + ], + "angle": 0, + "content": "the IEEE/CVF winter conference on applications of computer vision, pages 1508-1517, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[40] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[42] Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. Empowering llm to use smartphone for intelligent task automation. arXiv preprint arXiv:2308.15272, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.345, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[43] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609–6618, 2019. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.47 + ], + "angle": 0, + "content": "[44] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[45] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.524, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[46] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[47] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8289-8299, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[48] Kanghoon Yoon, Kibum Kim, Jinyoung Moon, and Chanyoung Park. Unbiased heterogeneous scene graph generation with relation-aware message passing neural network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3285-3294, 2023. 1, 6, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[49] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3736-3745, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.753, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[50] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.867 + ], + "angle": 0, + "content": "[51] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In European conference on computer vision, pages 409-424. Springer, 2022. 6, 7, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.891, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28315" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[52] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.203, + 0.472, + 0.263 + ], + "angle": 0, + "content": "[53] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.47, + 0.34 + ], + "angle": 0, + "content": "[54] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Learning to generate language-supervised and open-vocabulary scene graph using pretrained visual-semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915–2924, 2023. 1, 3, 4, 6, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.342, + 0.469, + 0.378 + ], + "angle": 0, + "content": "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.38, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[56] Yiwu Zhong, Liwei Wang, Jianshu Chen, Dong Yu, and Yin Li. Comprehensive image captioning via scene graph decomposition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 211-229. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[57] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1823-1834, 2021. 1, 3, 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.127, + 0.472, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.892, + 0.518, + 0.902 + ], + "angle": 0, + "content": "28316" + } + ] +] \ No newline at end of file diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_origin.pdf b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..caf9336b7dbd67637d28883dd4f4adb87cb59277 --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/a2404cd0-4571-42a9-a243-243f9def1b0c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee0a949e48679959057f2cdbd8f01468191c0f06526eddbde0fa7cf89c391aa5 +size 946424 diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/full.md b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c94f39a8f68278761e51597fa9c7230993a6a05 --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/full.md @@ -0,0 +1,254 @@ +# LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation + +Kibum Kim $^{1}$ Kanghoon Yoon $^{1}$ Jaehyeong Jeon $^{1}$ Yeonjun In $^{1}$ Jinyoung Moon $^{2}$ Donghyun Kim $^{3}$ Chanyoung Park $^{1*}$ KAIST $^{2}$ ETRI $^{3}$ Korea University + +{kb.kim,ykhoon08,wogud405,yeonjun.in,cy PARK}@kaist.ac.kr jymoon@etri.re.kr, d_kim@korea.ac.kr + +# Abstract + +Weakly-Supervised Scene Graph Generation (WSSGG) research has recently emerged as an alternative to the fully-supervised approach that heavily relies on costly annotations. In this regard, studies on WSSGG have utilized image captions to obtain unlocalized triplets while primarily focusing on grounding the unlocalized triplets over image regions. However, they have overlooked the two issues involved in the triplet formation process from the captions: 1) Semantic over-simplification issue arises when extracting triplets from captions, where fine-grained predicates in captions are undesirably converted into coarse-grained predicates, resulting in a long-tailed predicate distribution, and 2) Low-density scene graph issue arises when aligning the triplets in the caption with entity/predicate classes of interest, where many triplets are discarded and not used in training, leading to insufficient supervision. To tackle the two issues, we propose a new approach, i.e., Large Language Model for weakly-supervised SGG (LLM4SGG), where we mitigate the two issues by leveraging the LLM's in-depth understanding of language and reasoning ability during the extraction of triplets from captions and alignment of entity/predicate classes with target data. To further engage the LLM in these processes, we adopt the idea of Chain-of-Thought and the in-context few-shot learning strategy. To validate the effectiveness of LLM4SGG, we conduct extensive experiments on Visual Genome and GQA datasets, showing significant improvements in both Recall@K and mean Recall@K compared to the state-of-the-art WSSGG methods. A further appeal is that LLM4SGG is data-efficient, enabling effective model training with a small amount of training images. Our code is available on https://github.com/rlqja1107/torch-LLM4SGG + +# 1. Introduction + +Scene Graph Generation (SGG) is a fundamental task in computer vision, aiming at extracting structured visual + +knowledge from images [21, 32, 36, 39, 45, 56]. Most existing SGG methods are fully-supervised, i.e., they heavily rely on the ground-truth annotations that involve the class information of entities and predicates as well as the bounding box of entities [19, 48, 50]. However, since creating extensively annotated scene graph datasets is costly, the heavy reliance on these annotations imposes practical limitations on the model training [54]. To mitigate the high cost associated with manual annotations, weakly-supervised scene graph generation (WSSGG) approaches have recently emerged, aiming at training an SGG model without any annotated scene graph dataset. Specifically, the main idea of recent WSSGG methods is to leverage image captions along with associated images, as they can be easily collected from the Web [20, 47, 54, 57]. + +The training process of WSSGG model using image captions requires four steps as illustrated in Figure 1(a). Step 1: Preparing an image and its caption. Step 2: Parsing the image caption, i.e., triplets formed as $\langle$ subject, predicate, object $\rangle$ are extracted from the image caption through an off-the-shelf parser [30, 43]. Step 3: Aligning the triplets in the caption with entity/predicate classes of interest, i.e., entity (subject, object) and predicate classes in the extracted triplets obtained in Step 2 are aligned with the entity and predicate classes in the target data $^{1}$ , respectively. This alignment is based on their synonym/hypernym/hyponym contained in an external knowledge base (KB), e.g., WordNet [25]. Step 4: Grounding unlocalized entities in the extracted triplets, i.e., unlocalized entities (subjects and objects) are matched with relevant image regions generated by a pre-trained object detector, e.g., Faster R-CNN [29]. The localized entities and predicates in the extracted triplets then serve as pseudo-labels for training an SGG model. + +Existing WSSGG approaches mainly focus on Step 4 [20, 33, 47, 57]. For example, in Figure 1(a), their efforts have been focused on grounding the entity person in an unlocalized triplet with an image region that captures the sitting behavior. More precisely, LSwS [47] exploits the con + +![](images/bbe81ecf05d9410d234f9c6e0dfd2a9aff2ea0480193dbbc89b79b659e9f430a.jpg) + +![](images/7f479b1adad71406a506837942a910fa16066061a97fcf34f789c127de451856.jpg) +(c) Semantic Oversimplification + +![](images/eb48de475fefc6a4e19bcbefdee03faf14bd86c5ff720f3a4d86b5f01b9dd20f.jpg) +(a) Pipeline of Weakly-Supervised SGG +(b) Predicate Distribution +Figure 1. (a) The pipeline of weakly-supervised SGG. (b) The predicate distribution of unlocalized triplets (Parser+KB vs. Ours). In Parser+KB, the distribution becomes heavily long-tailed, and 12 out of 50 predicates are non-existent. (c) Semantic over-simplification caused by a rule-based parser in Step 2. (d) Low-density scene graph caused by the static structure of KB in Step 3. + +![](images/d5a9330d8e83d4927b4d110d691e08b082bc2e26c4608d32f158caa1a74932c6.jpg) +(d) Low-Density Scene Graph + +textual object information to accurately ground the unlocalized entities, leveraging the linguistic structure embedded within the triplets. Another line of research [20] employs a pre-trained vision-language model [15] to reflect the semantic interactions among entities within the image caption. + +However, we argue that existing WSSGG approaches overlook the importance of the triplet formation process conducted in Step 2 and Step 3. We identify the two major issues described below, i.e., semantic over-simplication and low-density scene graph, which incur incomplete unlocalized triplets after Step 2 and 3. These incomplete triplets are mostly uninformative predicates with a limited number, and negatively impact the training of an SGG model even when entities are correctly grounded in Step 4. To demonstrate the impact of incomplete unlocalized triplets, we follow the conventional process to extract unlocalized triplets (i.e., Step 1-3), and conduct an examination of triplets obtained from COCO caption dataset, which are generated through Scene Parser [43] in Step 2 and WordNet [25] in Step 3. As a result, we identify the following two issues: + +- Semantic Over-simplification: We find that the standard scene graph parser [43] operating on heuristic rule-based principles commonly used in Step 2 leads to a semantic over-simplification of predicates in extracted triplets. In other words, fine-grained predicates are undesirably converted into coarse-grained predicates, which we refer to as semantic over-simplification. For example, in Figure 1(c), an informative predicate lying on (i.e., fine-grained predicate) in the image caption is converted into a less informative predicate on (i.e., coarse-grained predicate), because the rule-based parser fails to capture the predicate lying on at once, and its heuristic rules fall short of accommodating the diverse range of caption's structure. As a result, the predicate distribution becomes heav- + +ily long-tailed, in which coarse-grained predicates (e.g., with, on, in) greatly outnumber fine-grained predicates (e.g., parked on, covered in) (Figure 1(b)). To make the matter worse, numerous fine-grained predicates eventually end up in a frequency of 0, even though they are originally present in the captions. Specifically, 12 out of 50 predicates are non-existent, which means that these 12 predicates can never be predicted, since the model is not trained on these predicates at all. + +- Low-Density Scene Graph: We find that the KB-based triplet alignment in Step 3 leads to low-density scene graphs, i.e., the number of remaining triplets after Step 3 is small. We attribute the low-density scene graphs primarily to the utilization of KB in Step 3. Specifically, a triplet is discarded if any of the three components (i.e., subject, predicate, object) or their synonym/hypernym/hyponym within the triplet fail to align with the entity or predicate classes in the target data. For example, in Figure 1(d), the triplet $\langle$ elephant, carrying, $\log \rangle$ is discarded because $\log$ does not exist in Visual Genome dataset nor its synonym/hypernym, even if elephant and carrying do exist. In Table 1, we report the number of triplets and images in Visual Genome dataset, which is a common benchmark dataset used in fully-supervised SGG approaches, and COCO caption dataset, which is a common benchmark dataset used in weakly-supervised SGG approaches. We observe that on average Visual Genome dataset contains 7.1 triplets (i.e., $405\mathrm{K} / 57\mathrm{K}$ ) per image (See Table 1(a)), while COCO dataset contains only 2.4 triplets (i.e., $154\mathrm{K} / 64\mathrm{K}$ ) per image (See Table 1(b)). This indicates that existing WSSGG approaches suffer from the lack of sufficient supervision per image, leading to poor generalization and performance degradation [47, 49]. In summary, relying on + +
DatasetHow to annotate# Triplet# Image
Fully-Supervised approach
(a) Visual GenomeManual405K57K
Weakly-Supervised approach
(b) COCO CaptionParser+KB154K64K
(c) COCO CaptionLLM344K64K
+ +Table 1. Comparison of scene graph density. + +the static structured knowledge of KB is insufficient to cover the semantic relationships among a wide a range of words, which incurs the low-density scene graph after Step 3. + +To alleviate the semantic over-simplification and the low-density scene graph issues, we propose a new approach, namely Large Language Model for weakly-supervised SGG (LLM4SGG) that adopts a pre-trained Large Language Model (LLM), which has shown remarkable transferability to various downstream tasks in NLP such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Inspired by the idea of Chain-of-Thought² (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into two chains, each of which replaces the rule-based parser in Step 2 (i.e., Chain-1) and the KB in Step 3 (i.e., Chain-2). More precisely, we design a prompt for extracting triplets from a caption, and ask the LLM to extract triplets formed as (Chain-1). We expect that the predicates extracted based on a comprehensive understanding of the caption's context via LLM are semantically rich, thereby alleviating the semantic over-simplification issue. Besides, to alleviate the low-density scene graph issue, we additionally incorporate a paraphrased version of the original caption. To this end, we further design a prompt for paraphrasing the original caption and extracting more triplets from the paraphrased caption. However, entities and predicates in the triplets obtained after Chain-1 are not yet aligned with the target data. Hence, we design a another prompt to align them with entity/predicate classes of interest, and ask the LLM to align them with semantically relevant lexeme included in a predefined lexicon, which is the set of vocabularies that are present in the target data (Chain-2). To further engage the LLM in the reasoning process of Chain-1 and Chain-2, we employ the in-context few-shot learning that incorporates a few input-output examples within the prompt, enabling the LLM to perform the task without the need for fine-tuning. + +To validate the effectiveness of LLM4SGG, we apply it to state-of-the-art WSSGG methods [54, 57]. Through extensive experiments, we show that LLM4SGG significantly enhances the performance of existing WSSGG methods in terms of mean Recall@K and Recall@K performance on Visual Genome and GQA datasets by alleviating the semantic over-simplification and the low-density scene graph + +(See Table 1(c) where the number of triplets increased to 334K). A further appeal of LLM4SGG is that it is data-efficient, i.e., it outperforms state-of-the-art baselines even with a small amount of training images, verifying the effectiveness of LLM4SGG. + +In summary, we make the following contributions: + +- We identify two major issues overlooked by existing WSSGG studies, i.e., semantic over-simplification and low-density scene graph. + +- We leverage an LLM along with the CoT strategy and the in-context few-shot learning technique to extract informative triplets without the need for fine-tuning the LLM. To the best of our knowledge, we are the first to leverage an LLM for the SGG task. + +- LLM4SGG outperforms the state-of-the-art WSSGG methods, especially in terms of mR@K, demonstrating its efficacy in addressing the long-tail problem in WSSGG for the first time. + +# 2. Related Works + +Weakly-Supervised Scene Graph Generation (WSSGG). The WSSGG task aims to train an SGG model without relying on an annotated scene graph dataset. To achieve this, most WSSGG studies [20, 47, 54, 57] utilize image captions and ground unlocalized triplets with image regions. Specifically, VSPNet [49] proposes the iterative graph alignment algorithm to reflect the high-order relations between unlocalized triplets and image regions. SGNLS [57] uses a pre-trained object detector [29] to ground the entities in unlocalized triplets, which share the same classes with the output of object detectors. In addition to the information derived from the object detector, [20] employs a pre-trained vision-language model [15] to capture the semantic interactions among objects. VS3 [54] uses a grounding-based object detector [18], which calculates the similarity between the entity text in unlocalized triplet and image region, thereby grounding the unlocalized triplets. However, these methods overlook the triplet formation process that leads to the semantic over-simplification (Step 2) and the low-density scene graph (Step 3). In this regard, existing methods result in a sub-optimal performance even when unlocalized triplets are correctly grounded in Step 4. + +Large Language Model (LLM). LLMs have demonstrated remarkable transferability to various downstream tasks such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Specifically, GPT-3 (175B) [2] stands as a cornerstone to break the line of numerous language tasks. Inspired by GPT-3, PaLM (540B) [5], LLaMA (65B) [38], OPT (175B) [53], and LaMDA (137B) [37] have been subsequently introduced. More recently, advanced GPT models (e.g., GPT-4 [28], ChatGPT [27]) fine-tuned with human feedback have gained prominence and widely applied for diverse applications, e.g., planner of tools [9, 24], mobile + +task automation [42]. In this work, we employ the power of LLM (i.e., ChatGPT) to alleviate the two issues, i.e., semantic over-simplification and low-density scene graph, in the context of the WSSGG task. + +In-Context Few-shot Learning. In-context few-shot learning incorporates a few input-output examples related to a target task, conditioning the LLM on the context of examples. Specifically, GPT-3 [2] pioneered the concept of in-context learning to facilitate an LLM as a versatile model on diverse tasks. This breakthrough has proliferated a plethora of research to leverage the in-context few-shot learning for various tasks [24, 26, 46, 55]. More precisely, Chameleon [24] integrates a few examples to enhance its understanding of tool planning task. [26] utilizes positive and negative examples related to questions for question generation tasks. ReAct [46] incorporates examples of reasoning with action for solving decision-making tasks. Inspired by recent in-context few-shot learning approaches, we provide a few examples to LLMs to help 1) understand the process of triplet extraction from a caption (i.e., Step 2), and 2) align the entity/predicate classes with the target data (i.e., Step 3) in the context of the WSSGG task. + +# 3. Method + +In this section, we describe LLM4SGG in detail. We would like to emphasize that LLM4SGG mainly focuses on the triplet formation process conducted in Step 2 (Parsing) and Step 3 (aligning), while existing WSSGG approaches mainly focus on Step 4 (grounding). We start by presenting the problem formulation of WSSGG (Section 3.1), followed by the prompt configuration (Section 3.2). Next, we introduce how LLMs are adopted to address the two issues of conventional WSSGG approaches when parsing the image caption (Section 3.3) and aligning the triplets in captions with entity/predicate classes of interest (Section 3.4). Finally, we ground the unlocalized triplets by associating them with bounding boxes (i.e., image regions) and train the SGG model using the localized triplets (Section 3.5). The overall pipeline of LLM4SGG is shown in Figure 2. + +# 3.1. Problem Formulation + +In the fully supervised SGG task, we aim to detect a scene graph $\mathbf{G}_f = \{\mathbf{s}_i,\mathbf{p}_i,\mathbf{o}_i\}_{i = 1}^{N_f}$ that consists of triplets given an image $\mathcal{I}$ , where $N_{f}$ is the number of triplets in the image. $\mathbf{s}_i$ and $\mathbf{o}_i$ denote the $i^{\mathrm{th}}$ subject and the object, respectively, whose bounding boxes are $\mathbf{s}_{i,b}$ , $\mathbf{o}_{i,b}$ , and entity classes are $\mathbf{s}_{i,c},\mathbf{o}_{i,c}\in \mathcal{C}_e$ , where $\mathcal{C}_e$ is the set of predefined entity classes in the target data. $\mathbf{p}_i$ denotes the predicate between $\mathbf{s}_i$ and $\mathbf{o}_i$ , and its class is $\mathbf{p}_{i,c}\in \mathcal{C}_p$ , where $\mathcal{C}_p$ is the set of predefined predicate classes in the target data. By using the ground truth scene graphs as supervision, fully supervised SGG approaches train an SGG model $\mathcal{T}_{\theta}:\mathcal{I}\to \mathbf{G}_{f}$ , which maps an image to a scene graph. + +In the weakly supervised SGG task, we aim to generate a scene graph when the ground truth scene graph is not given, i.e., there are no bounding boxes and entity/predicate class information. Instead, existing WSSGG approaches [20, 33, 47, 54, 57] use image captions along with associated images to produce scene graphs, i.e., localized triplets. More precisely, they extract a set of triplets $\mathbf{G}_w = \{\mathbf{s}_i,\mathbf{p}_i,\mathbf{o}_i\}_{i = 1}^{N_w}$ from the image captions, where $N_{w}$ is the number of triplets extracted from the captions. However, while the extracted triplets contain the class information (i.e., $\mathbf{s}_{i,c},\mathbf{o}_{i,c}$ and $\mathbf{p}_{i,c}$ ), they are unlocalized, since bounding boxes $\mathbf{s}_{i,b}$ and $\mathbf{o}_{i,b}$ are not included in the caption. Therefore, it is essential to perform the grounding step to associate the unlocalized triplets with bounding boxes. Once we have obtained the localized triplets, we can apply the conventional SGG training scheme, i.e., $\mathcal{T}_{\theta}:\mathcal{I}\rightarrow \mathbf{G}_{f}$ . + +In this paper, our focus is to address the semantic oversimplification and low-density scene graph issues regarding the unlocalized triplets $\mathbf{G}_w$ , which has been overlooked in existing WSSGG studies. Specifically, we aim to produce an enhanced $\mathbf{G}_w$ by refining the process of scene graph dataset construction via an LLM. This refinement includes the triplet extraction step from the caption (Step 2) and the alignment of entity/predicate classes (Step 3), leveraging the LLM's comprehensive understanding of language and reasoning ability. + +# 3.2. Prompt Configuration + +In fact, it is not a trivial task for an LLM to immediately generate triplets from a caption whose entities and predicates are aligned with entity/predicate classes of interest, as such a task is a novel task for the LLM. Inspired by the idea of the Chain-of-thought (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into the following two chains: Chain-1 - Extracting triplets from captions. Chain-2 - Aligning entities and predicates with the entity/predicate classes of interest. To carefully design each chain, we define the LLM function, i.e., $LLM(\cdot)$ , with the following prompt input: + +$$ +O u t p u t = L L M (\underbrace {\text {T a s k d e s c r i p t i o n , I n - c o n t e x t e x a m p l e s , A c t u a l q u e s t i o n}} _ {\text {P r o m p t i n p u t}}, +$$ + +where $LLM(\cdot)$ is the decoder of the LLM, generating the (1) Output given the prompt input. The prompt input consists of three components in a sequence: 1) task description, i.e., the delineation of the task that we intend to perform, 2) in-context examples, i.e., sample questions and answers related to the task at hand, 3) actual question, i.e., an inquiry from which we intend to derive the answer. Note that in-context examples is closely related to the in-context few-shot learning [2, 40, 41], which is shown to enhance the LLM's understanding of the task. Note that the above configuration of the prompt input is applied to the triplet extraction (Chain-1) (Section 3.3) and the alignment of entity/predicate classes (Chain-2) (Section 3.4). + +![](images/51b4929e428bf2e4b0ad9d458e510ab8301e43b735b4afbf5fb32a02f1a82010.jpg) +Figure 2. The pipeline of LLM4SGG. Given an image with its caption, we use an LLM to extract triplets from the original caption (Step 2-1) and the paraphrased caption (Step 2-2). Then, we align the entity/predicate classes within the extracted triplets with semantically similar lexeme in the target data via an LLM (Step 3), obtaining the unlocalized triplets. Lastly, we ground the unlocalized triplets over image regions (Step 4) followed by the training of an SGG model. + +# 3.3. Chain-1. Triplet Extraction via LLM (Step 2 in Figure 2) + +Based on the LLM's comprehensive understanding of the context of an image caption, we aim to extract triplets from the caption. As discussed in Section 1, we use not only the original caption but also a paraphrased caption generated by the LLM to address the low-density scene graph issue. + +Extracting triplets from a paraphrased caption (Step 2-2). To extract triplets from a paraphrased caption, we inform the LLM about the task at hand by presenting the following prompt: FROM THE GIVEN SENTENCE, THE TASK IS TO EXTRACT MEANINGFUL TRIPLETS FORMED AS $\langle$ SUBJECT, PREDICATE, OBJECT $\rangle$ (i.e., Task description in Equation 1). We then instruct the LLM to follow the two steps, i.e., paraphrasing step and triplet extraction step. To help the LLM understand the process of performing the two steps, we present few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation $1^{3}$ ), which involves a manual construction of questions and corresponding answers to the paraphrasing and the triplet extraction steps. That is, for given a caption, we show the LLM how we expect a paraphrased caption and the extracted triplets would look like (e.g., Given "Four clocks sitting on a floor next to a woman's feet," we show the LLM that a paraphrased sentence would be "Four clocks are placed on the floor beside a woman's feet," and extracted triplets would be $\langle$ clocks, placed on, floor $\rangle$ and $\langle$ clocks, beside, feet $\rangle$ ). Lastly, we show the caption of our interest to the LLM, and let the LLM extract triplets from the caption (i.e., Actual question in Equation 1). Please refer to Figure 2 right (bottom) for an example of the prompt input used in Step 2-2. + +Extracting triplets from the original caption (Step 2-1). As extracting triplets from the original caption does not involve the caption paraphrasing step, we exclude it from the prompt used in Step 2-2. Please refer to Figure 2 right (top) for an example of the prompt input used in Step 2-1. + +In summary, we obtain triplets from both the original and paraphrased captions after Step 2-1 and Step 2-2, respectively, which in turn alleviates the semantic oversimplification issue of predicates and the low-density scene graph issue. + +# 3.4. Chain-2. Alignment of Classes in Triplets via LLM (Step 3 in Figure 2) + +The entities (i.e., subject and object) and predicates within the triplets obtained from Step 2 described in Section 3.3 are not yet aligned with the target data. Based on the semantic reasoning ability of the LLM, we aim to align them with the semantically relevant lexeme in the target data. + +Aligning entities in the triplets with entity classes of interest. We instruct the LLM with the following prompt: GIVEN THE LEXEME $\{\mathsf{ENTITY}\}$ , FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED ENTITY LEXICON, where the predefined entity lexicon is $\mathcal{C}_e$ (i.e., Task description in Equation 1). Similar to Section 3.3, we present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding hierarchical relationships such as pigeon being semantically relevant to bird, and singular-plural relationships such as surfboards being semantically relevant to surfboard. Lastly, we show the entity of our interest to the LLM (i.e., Actual question in Equation 1), which enables the LLM to generate an answer by finding the most + +semantically relevant entity in $\mathcal{C}_e$ . Please refer to Table 7 in Appendix A.1 for an example of the prompt input. + +Aligning predicates in the triplets with predicate classes of interest. Likewise, we instruct the LLM with the following prompt: GIVEN THE LEXEME $\{\mathrm{PREDICATE}\}$ , FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED Predicate Lexicon, where the predefined predicate lexicon is $C_p$ (i.e., Task description in Equation 1). We also present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding tense relationships such as the lies on being semantically relevant to lying on, and positional relationship such as next to being semantically relevant to near. Lastly, we show the predicate of our interest to the LLM (i.e., Actual question in Equation 1). Please refer to Table 8 in Appendix A.1 for an example of the prompt input. Furthermore, regarding the alignment of classes within a large predefined lexicon, please refer to Appendix A.2. Note that as our main focus in this work is on developing a framework for utilizing LLMs in weakly-supervised SGG rather than finding a specific prompt design that works the best, we tested with a single prompt design. However, since prompt designs are crucial in leveraging LLMs, we plan to explore various designs in our future work. + +After performing Chain-1 (Section 3.3) and Chain-2 (Section 3.4), we obtain intermediate unlocalized triplets $\hat{\mathbf{G}}_w = \{\mathbf{s}_i, \mathbf{p}_i, \mathbf{o}_i\}_{i=1}^{\hat{N}_w}$ , where $\mathbf{s}_{i,c}, \mathbf{o}_{i,c} \in \{\mathcal{C}_e \cup \text{None}\}$ and $\mathbf{p}_{i,c} \in \{\mathcal{C}_p \cup \text{None}\}$ . It is worth noting that if there is no semantically relevant lexeme, we request the LLM to generate None as the answer, due to the fact that the entity/predicate classes in the target data may not cover a wide range of entities/predicate. Similar to the conventional approach, we discard a triplet if any of its three components (i.e., subject, predicate and object) is None. Lastly, we obtain the final unlocalized triplets $\mathbf{G}_w = \{\mathbf{s}_i, \mathbf{p}_i, \mathbf{o}_i\}_{i=1}^{N_w}(N_w \leq \hat{N}_w)$ , where $\mathbf{s}_{i,c}, \mathbf{o}_{i,c} \in \mathcal{C}_e$ and $\mathbf{p}_{i,c} \in \mathcal{C}_p$ . + +# 3.5. Model Training + +Given the final unlocalized triplets $\mathbf{G}_w$ , we ground them over relevant image regions to get localized triplets, meaning that we obtain $\mathbf{s}_{i,b}$ and $\mathbf{o}_{i,b}$ . To this end, we employ two state-of-the-art grounding methods, i.e., SGNLS [57] and VS3 [54]. Please refer to Appendix A.3 for more detail about how each method performs grounding. After grounding $\mathbf{G}_w$ , we obtain localized triplets and use them as pseudo-labels for training a supervised SGG model. Please refer to Appendix A.4 for more detail about the model training. + +# 4. Experiment + +Datasets. To train an SGG model without an annotated scene graph dataset, we use three caption datasets: COCO caption [3], Conceptual (CC) caption [31], and Visual + +Genome (VG) caption [44]. For fair comparisons, we use the same set of images that have been utilized in previous WSSGG studies [20, 47, 54, 57], leading to the utilization of 64K images on COCO caption dataset, 145K images on CC caption dataset, and 57K images on VG caption dataset. To evaluate the trained SGG model, we employ the widely used Visual Genome (VG) dataset [13] and GQA dataset [11]. The VG dataset contains the ground-truth localized triplet information annotated by humans. We follow the standard split of VG [44], which consist of 150 entity classes and 50 predicate classes. For the GQA dataset used in the SGG task, we follow the same pre-processing step of a previous SGG study [8], which involves selecting top-200 frequent entity classes and top-100 frequent predicate classes. In both datasets, $30\%$ of the total images are used for evaluation. Please refer to Appendix C.1 for more details regarding the datasets. Note that we mainly use the COCO caption dataset for analysis of LLM4SGG throughout this paper, i.e., CC and VG caption datasets are exclusively used in quantitative result on VG (Section 4.1). + +Evaluation metrics. Recent fully-supervised SGG studies [1, 48, 51] have emphasized improving the accuracy of predictions for fine-grained predicates rather than coarse-grained predicates, since the former construct richer scene graphs. As a result, they commonly use mean Recall@K (mR@K) that computes the average of Recall@K (R@K) across all predicates. In line with the recent emphasis on fine-grained predicates, we incorporate both mR@K and R@K in our evaluation, whereas previous WSSGG studies [20, 47, 54] mainly rely on the R@K metric alone. Moreover, we report F@K, which is the harmonic average of R@K and mR@K to jointly consider R@K and mR@K, following previous SGG studies [12, 51]. Regarding the evaluation task, we follow previous WSSGG studies and adopt the Scene Graph Detection (SGDet) task, where both the ground-truth bounding box and the entity class information are not provided. Please refer to Appendix C.2 for more detail regarding the task. + +Baselines. Please refer to Appendix C.3 for details regarding the baselines. + +Implementation details. Please refer to Appendix C.4 regarding the implementation details. + +# 4.1. Quantitative Result on VG + +Table 2 shows the performance of baseline models and those when LLM4SGG is applied. We have the following observations based on COCO caption dataset: 1) Applying LLM4SGG to SGNLS and VS3 improves the performance in terms of R@K and mR@K, which demonstrates the effectiveness of the triplet formation through LLM4SGG. Notably, LLM4SGG significantly improves mR@K, implying that LLM4SGG effectively alleviates the long-tailed problem in WSSGG. This can be clearly seen in Figure 3, which shows the performance gain on fine + +
MethodDatasetR@50R@100mR@50mR@100F@50F@100
Motif (CVPR'18) - Fully-supervisedVG31.8936.366.387.5710.63 / 12.5312.53
LSWS (CVPR'21)COCO Caption3.293.693.273.663.283.67
SGNLS (ICCV'21)3.804.462.512.783.023.43
SGNLS (ICCV'21)+LLM4SGG5.09+1.295.97+1.514.08+1.574.49+1.714.53+1.515.13+1.70
Li et al (MM'22)6.407.331.731.982.723.12
VS3(CVPR'23)6.608.012.883.254.014.62
VS3(CVPR'23)+LLM4SGG8.91+2.3110.43+2.427.11+4.238.18+4.937.91+3.909.17+4.55
VS3(CVPR'23)+Rwt4.255.045.175.994.675.47
VS3(CVPR'23)+Rwt+LLM4SGG5.10+0.856.34+1.308.42+3.259.90+3.916.35+1.697.73+2.26
VS3(CVPR'23)CC Caption6.698.201.732.042.753.27
VS3(CVPR'23)+LLM4SGG9.47+2.7810.69+2.495.40+3.676.09+4.056.88+4.137.76+4.49
VS3(CVPR'23)VG Caption14.5418.482.803.794.706.29
VS3(CVPR'23)+LLM4SGG18.40+3.8622.28+3.806.26+3.467.60+3.819.34+4.6411.33+5.04
+ +![](images/aca0a378acdb45627d51d035e807e92e5f8de4db7d5f790dc07ef1c6f8e7c49a.jpg) +Figure 3. Per class performance (Bar: number of predicate instances, Line: Recall@100). + +grained predicates. 2) $\mathrm{VS}^3 +\mathrm{Rwt} + \mathrm{LLM4SGG}$ further improves mR@K of $\mathrm{VS}^3 +\mathrm{Rwt}$ . We attribute this to the fact that the conventional approach generates a limited number of fine-grained predicates, which makes the reweighting strategy less effective within $\mathrm{VS}^3$ . Especially, non-existent predicates can never be predicted even when the reweighting strategy is applied. On the other hand, LLM4SGG increases the number of instances that belong to fine-grained predicates, which is advantageous for the reweighting strategy. For the per class performance comparison over the reweighting stregty, please refer to Appendix D.1. 3) The performance gain obtained from applying LLM4SGG is greater on $\mathrm{VS}^3$ (i.e., $\mathrm{VS}^3 +\mathrm{LLM4SGG}$ ) than on SGNLS (i.e., SGNLS+LLM4SGG). The major reason lies in the difference in how SGNLS and $\mathrm{VS}^3$ make use of the pool of 344K unlocalized triplets obtained through LLM4SGG. Specifically, in the grounding process of SGNLS, we observe that 100K out of 344K unlocalized triplets (i.e., $29\%$ ) fail to be grounded, and thus not used for training. On the other hand, $\mathrm{VS}^3$ successfully grounds all 344K unlocalized triplets and fully utilize them for training, allowing it to fully enjoy the effectiveness of LLM4SGG. This indicates that LLM4SGG makes synergy when paired with a grounding method that is capable of fully utilizing the unlocalized triplets. For more details regarding the impact of grounding methods, please refer to Appendix B. 4) Regarding the performance comparison with a fully-supervised approach (i.e., Motif), please refer to Appendix D.2. + +Furthermore, we observe that applying LLM4SGG to CC and VG caption datasets also significantly improves performance, demonstrating the effectiveness of LLM4SGG. + +Table 2. Performance comparisons on the SGDet task. The best performance among WSSGG models within each dataset is in bold. The red numbers indicate the absolute performance improvement after applying LLM4SGG. Rwt denotes using the reweighting strategy [51]. + +
RowPCLPLA# TripletR@50 / 100mR@50 / 100F@50 / 100
(a)154K6.60 / 8.012.88 / 3.254.01 / 4.62
(b)✓243K9.46 / 11.223.43 / 3.925.03 / 5.81
(c)✓✓256K8.42 / 9.855.99 / 6.957.00 / 8.15
(d)✓✓327K11.76 / 13.383.50 / 4.055.39 / 6.22
(e)✓✓✓344K8.91 / 10.437.11 / 8.187.91 / 9.17
+ +Table 3. Ablation studies. (PC: Using Paraphrased Caption in addition to the original caption / LP: LLM-based Parsing / LA: LLM-based Alignment) + +# 4.2. Ablation Studies + +In Table 3, we conduct ablation studies on VG dataset to understand the effectiveness of each component of LLM4SGG, where VS $^3$ is used as the grounding method. Note that row (a) is equivalent to vanilla VS $^3$ . We have the following observations. 1) Effect of using the paraphrased caption: Including the paraphrased caption in addition to the original caption (row (b)) increases the number of triplets $(154\mathrm{K}\rightarrow 243\mathrm{K})$ , resulting in an improved overall performance. This demonstrates that the paraphrased caption alleviates the low-density scene graph issue. 2) Effect of LLM-based parsing: The LLM-based parsing (row (c)) for extracting triplets improves mR@K of row (b). This indicates that the LLM-based parsing increases the number of instances that belong to fine-grained predicates, which in turn alleviates the semantic over-simplification issue. 3) Effect of LLM-based alignment: The LLM-based alignment (row (d)) of entities/predicates in the extracted triplets increases the number of triplets from $243\mathrm{K}$ to $327\mathrm{K}$ (row (b) vs (d)), which indicates that the low-density scene graph issue is alleviated. Consequently, R@K and mR@K of row (d) are greater than those of row (b). 4) The fully-fledged approach (row (e)) generally improves R@K and mR@K, showing the best performance in terms of F@K. It is important to highlight that when using the LLM-based parsing, the performance of mR@K significantly increases with a moderate decrease in R@K. This trade-off is attributed to the fact that R@K generally improves when the coarse-grained predicates are dominant [51]. In contrast, our approach, which addresses the semantic over-simplification issue, decreases the number instances that belong to coarse-grained predicates while simultaneously increasing those that belong to fine-grained predicates (Figure 1(b)), which + +![](images/6ee00aff165015599ee6dfdea78d6088413e6eb7c93f798dbde1e0d1fc66b273.jpg) +Figure 4. Performance over various numbers of images used for training VS $^3$ + LLM4SGG. + +in turn results in a substantial improvement in mR@K. We would like to emphasize that the performance in terms of mR@K is crucial in the context of SGG research [1, 8, 48], as fine-grained predicates offer richer information. + +# 4.3. Analysis of Data-Efficiency + +To assess the effectiveness of LLM4SGG under the lack of available training data, we conduct experiments given a limited number of images. Specifically, among 64K images used for training VS $^3$ and VS $^3$ +LLM4SGG in Table 2, we randomly sample 1K (1.5%), 5K (7.8%), and 10K (15.6%) images with replacement, and train VS $^3$ +LLM4SGG for five times. Figure 4 shows the average performance over various numbers of images used for training VS $^3$ +LLM4SGG along with the variance (in blue area). We observe that when using only 1K images, the performance is slightly inferior to the baseline that used 64K for training (i.e., VS $^3$ ). However, as we increase the number of images used for training to 5K images, we observe a significant improvement in both R@K and mR@K compared with the baseline. This demonstrates that LLM4SGG is data-efficient, as it outperforms the baseline even with only 7.8% of the total images used for training the baseline. Moreover, when further increasing the number of images to 10K, we observe further performance improvements, and when it reaches 64K, which is the same as the number of images used for training the baseline, the performance is the best. In summary, LLM4SGG enables data-efficient model training even with a limited amount of available images for training, thanks to alleviating the semantic over-simplification and low-density scene graph issues. + +# 4.4. Quantitative Result on GQA + +In Table 4, we additionally conducted experiments on GQA dataset [11]. Please refer to Appendix E.1 for more detailed descriptions on the training and evaluation processes on GQA dataset. The GQA dataset contains twice as many predicates as the Visual Genome dataset and includes complicated predicates (e.g., sitting next to, standing in front of). As a result, when obtaining unlocalized triplets using the conventional WSSGG approach, we observe that 44 out of 100 predicates have a frequency of 0, and the predicate distribution is extremely long-tailed. Consequently, the baseline (i.e., $\mathrm{VS}^3$ ) exhibits significantly lower performance, especially in terms of mR@K. On the other hand, + +
MethodR@50 / 100mR@50 / 100F@50 / 100
Motif (Fully-supervised)28.90 / 33.106.40 / 7.7010.48 / 12.49
VS35.90 / 6.971.60 / 1.812.52 / 2.87
VS3+LLM4SGG8.88 / 10.385.33 / 6.516.66 / 8.00
+ +Table 4. Performance comparison on GQA. + +our approach shows substantial performance improvements not only in R@K but also in mR@K, thanks to the mitigation of semantic over-simplification and low-density scene graph issues. Please refer to Appendix E.2 for the predicate distribution and performance comparison for each class in GQA dataset. Additionally, please refer to Appendix E.3 for qualitative analyses on GQA dataset. + +# 5. Conclusion & Future work + +In this work, we focus on the triplet formation process in WSSGG, whose importance is overlooked by previous studies. To alleviate the semantic over-simplification and low-density scene graph issues inherent in the triplet formation process, we propose a new approach, i.e., LLM4SGG, which leverages a pre-trained LLM during the extraction of triplets from the captions, and alignment of entity/predicate classes with those in the target data. It is important to note that construction of these triplets is not required every time to train the SGG model; instead, it is a one-time pre-processing step. In this regard, we contribute to generating enhanced triplets compared to the conventional approach. Moreover, we publish an enhanced SGG dataset constructed by LLM for future studies of SGG. As a result, we outperform baselines in terms of R@K, mR@K and F@K on Visual Genome and GQA datasets. A potential limitation of our work is the reliance on a proprietary blackbox LLM, which can also be costly to use. Hence, in Appendix F, we provide discussions on replacing the LLM with smaller language models. + +For future work, an LLM can be used to ground the unlocalized triplets in Step 4. Recently, vision-language representation learning has been developed for transforming visual features into textual features to facilitate the use of visual features as input to an LLM [16, 58]. In this regard, given the visual features of bounding boxes as input, we could ask the LLM to identify relevant bounding boxes based on the textual information of entities within unlocalized triplets using the comprehensive understanding of the context of triplets. + +# 6. Acknowledgement + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network, No.2022-0-00077, Reasoning, and Inference from Heterogeneous Data, No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University). + +# References + +[1] Bashirul Azam Biswas and Qiang Ji. Probabilistic debiasing of scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10429-10438, 2023. 6, 8 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3, 4 +[3] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6 +[4] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2, 3 +[5] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3 +[6] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 2, 3, 7 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 7 +[8] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 6, 8, 3 +[9] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3 +[10] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 3 +[11] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6, 8, 3, 7 +[12] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 6, 3 +[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: + +Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1, 6, 2, 3, 5, 7 +[14] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2,3,4 +[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2, 3, 4 +[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8 +[17] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. arXiv preprint arXiv:2308.06712, 2023. 3 +[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3, 2, 4 +[19] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 3 +[20] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4204-4213, 2022. 1, 2, 3, 4, 6 +[21] Yongzhi Li, Duo Zhang, and Yadong Mu. Visual-semantic matching by exploring high-order attention and distraction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786–12795, 2020. 1 +[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4 +[24] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3, 4 +[25] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 1, 2 + +[26] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773, 2021. 4 +[27] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023. 3, 4 +[28] OpenAI. Gpt-4. Technical report, 2023. 3 +[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3, 2, 4 +[30] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1 +[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6 +[32] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1 +[33] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16393-16402, 2021. 1, 4 +[34] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 3 +[35] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 5 +[36] Damien Teney, Lingqiao Liu, and Anton van Den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2017. 1 +[37] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 3 +[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 9 +[39] Sijin Wang, Ruiping Wang, Ziwei Yao, Shiguang Shan, and Xilin Chen. Cross-modal scene graph matching for relationship-aware image-text retrieval. In Proceedings of + +the IEEE/CVF winter conference on applications of computer vision, pages 1508-1517, 2020. 1 +[40] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 4 +[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 3, 4, 6 +[42] Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. Empowering llm to use smartphone for intelligent task automation. arXiv preprint arXiv:2308.15272, 2023. 4 +[43] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609–6618, 2019. 1, 2, 6, 8 +[44] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 6 +[45] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1 +[46] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 4 +[47] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8289-8299, 2021. 1, 2, 3, 4, 6 +[48] Kanghoon Yoon, Kibum Kim, Jinyoung Moon, and Chanyoung Park. Unbiased heterogeneous scene graph generation with relation-aware message passing neural network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3285-3294, 2023. 1, 6, 8, 3 +[49] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3736-3745, 2020. 2, 3 +[50] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 1, 3, 4 +[51] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In European conference on computer vision, pages 409-424. Springer, 2022. 6, 7, 3, 5 + +[52] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 1 +[53] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3 +[54] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Learning to generate language-supervised and open-vocabulary scene graph using pretrained visual-semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915–2924, 2023. 1, 3, 4, 6, 2, 7, 8 +[55] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 4 +[56] Yiwu Zhong, Liwei Wang, Jianshu Chen, Dong Yu, and Yin Li. Comprehensive image captioning via scene graph decomposition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 211-229. Springer, 2020. 1 +[57] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1823-1834, 2021. 1, 3, 4, 6, 2 +[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8 \ No newline at end of file diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/images.zip b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba7e30fcf0e337179cf4619bc66030215a83d1ba --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcb5fa28918043159aaaddfa794db6c66e7dbae18c858916f67539cd55f929f8 +size 379375 diff --git a/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/layout.json b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c220df2dcff380ce2852e1e5a465ad0a1c146cd --- /dev/null +++ b/2024/LLM4SGG_ Large Language Models for Weakly Supervised Scene Graph Generation/layout.json @@ -0,0 +1,7849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 127, + 482, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 127, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 111, + 127, + 482, + 159 + ], + "type": "text", + "content": "LLM4SGG: Large Language Models for Weakly Supervised Scene Graph Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "spans": [ + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": "Kibum Kim" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Kanghoon Yoon" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Jaehyeong Jeon" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Yeonjun In" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Jinyoung Moon" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Donghyun Kim" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " Chanyoung Park" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": " KAIST " + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": "ETRI " + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 179, + 504, + 218 + ], + "type": "text", + "content": "Korea University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 220, + 559, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 559, + 230 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 559, + 230 + ], + "type": "text", + "content": "{kb.kim,ykhoon08,wogud405,yeonjun.in,cy PARK}@kaist.ac.kr jymoon@etri.re.kr, d_kim@korea.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 256, + 191, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 275, + 289, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 289, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 289, + 619 + ], + "type": "text", + "content": "Weakly-Supervised Scene Graph Generation (WSSGG) research has recently emerged as an alternative to the fully-supervised approach that heavily relies on costly annotations. In this regard, studies on WSSGG have utilized image captions to obtain unlocalized triplets while primarily focusing on grounding the unlocalized triplets over image regions. However, they have overlooked the two issues involved in the triplet formation process from the captions: 1) Semantic over-simplification issue arises when extracting triplets from captions, where fine-grained predicates in captions are undesirably converted into coarse-grained predicates, resulting in a long-tailed predicate distribution, and 2) Low-density scene graph issue arises when aligning the triplets in the caption with entity/predicate classes of interest, where many triplets are discarded and not used in training, leading to insufficient supervision. To tackle the two issues, we propose a new approach, i.e., Large Language Model for weakly-supervised SGG (LLM4SGG), where we mitigate the two issues by leveraging the LLM's in-depth understanding of language and reasoning ability during the extraction of triplets from captions and alignment of entity/predicate classes with target data. To further engage the LLM in these processes, we adopt the idea of Chain-of-Thought and the in-context few-shot learning strategy. To validate the effectiveness of LLM4SGG, we conduct extensive experiments on Visual Genome and GQA datasets, showing significant improvements in both Recall@K and mean Recall@K compared to the state-of-the-art WSSGG methods. A further appeal is that LLM4SGG is data-efficient, enabling effective model training with a small amount of training images. Our code is available on https://github.com/rlqja1107/torch-LLM4SGG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 646, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 287, + 668 + ], + "type": "text", + "content": "Scene Graph Generation (SGG) is a fundamental task in computer vision, aiming at extracting structured visual" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 257, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 257, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 305, + 257, + 545, + 417 + ], + "type": "text", + "content": "knowledge from images [21, 32, 36, 39, 45, 56]. Most existing SGG methods are fully-supervised, i.e., they heavily rely on the ground-truth annotations that involve the class information of entities and predicates as well as the bounding box of entities [19, 48, 50]. However, since creating extensively annotated scene graph datasets is costly, the heavy reliance on these annotations imposes practical limitations on the model training [54]. To mitigate the high cost associated with manual annotations, weakly-supervised scene graph generation (WSSGG) approaches have recently emerged, aiming at training an SGG model without any annotated scene graph dataset. Specifically, the main idea of recent WSSGG methods is to leverage image captions along with associated images, as they can be easily collected from the Web [20, 47, 54, 57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "text", + "content": "The training process of WSSGG model using image captions requires four steps as illustrated in Figure 1(a). Step 1: Preparing an image and its caption. Step 2: Parsing the image caption, i.e., triplets formed as " + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "text", + "content": " subject, predicate, object " + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "text", + "content": " are extracted from the image caption through an off-the-shelf parser [30, 43]. Step 3: Aligning the triplets in the caption with entity/predicate classes of interest, i.e., entity (subject, object) and predicate classes in the extracted triplets obtained in Step 2 are aligned with the entity and predicate classes in the target data" + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 305, + 418, + 545, + 610 + ], + "type": "text", + "content": ", respectively. This alignment is based on their synonym/hypernym/hyponym contained in an external knowledge base (KB), e.g., WordNet [25]. Step 4: Grounding unlocalized entities in the extracted triplets, i.e., unlocalized entities (subjects and objects) are matched with relevant image regions generated by a pre-trained object detector, e.g., Faster R-CNN [29]. The localized entities and predicates in the extracted triplets then serve as pseudo-labels for training an SGG model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 611, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 545, + 666 + ], + "type": "text", + "content": "Existing WSSGG approaches mainly focus on Step 4 [20, 33, 47, 57]. For example, in Figure 1(a), their efforts have been focused on grounding the entity person in an unlocalized triplet with an image region that captures the sitting behavior. More precisely, LSwS [47] exploits the con" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "spans": [ + { + "bbox": [ + 85, + 36, + 127, + 64 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 158, + 34, + 479, + 67 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 34, + 479, + 67 + ], + "spans": [ + { + "bbox": [ + 158, + 34, + 479, + 67 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 676, + 137, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 676, + 137, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 676, + 137, + 686 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 676, + 471, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 676, + 471, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 676, + 471, + 686 + ], + "type": "text", + "content": "1We use Visual Genome [13] as the target data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28306" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 99, + 359, + 171 + ], + "blocks": [ + { + "bbox": [ + 73, + 99, + 359, + 171 + ], + "lines": [ + { + "bbox": [ + 73, + 99, + 359, + 171 + ], + "spans": [ + { + "bbox": [ + 73, + 99, + 359, + 171 + ], + "type": "image", + "image_path": "bbe81ecf05d9410d234f9c6e0dfd2a9aff2ea0480193dbbc89b79b659e9f430a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 362, + 101, + 515, + 171 + ], + "blocks": [ + { + "bbox": [ + 362, + 101, + 515, + 171 + ], + "lines": [ + { + "bbox": [ + 362, + 101, + 515, + 171 + ], + "spans": [ + { + "bbox": [ + 362, + 101, + 515, + 171 + ], + "type": "image", + "image_path": "7f479b1adad71406a506837942a910fa16066061a97fcf34f789c127de451856.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 172, + 492, + 180 + ], + "lines": [ + { + "bbox": [ + 395, + 172, + 492, + 180 + ], + "spans": [ + { + "bbox": [ + 395, + 172, + 492, + 180 + ], + "type": "text", + "content": "(c) Semantic Oversimplification" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 76, + 180, + 356, + 253 + ], + "blocks": [ + { + "bbox": [ + 163, + 172, + 284, + 181 + ], + "lines": [ + { + "bbox": [ + 163, + 172, + 284, + 181 + ], + "spans": [ + { + "bbox": [ + 163, + 172, + 284, + 181 + ], + "type": "text", + "content": "(a) Pipeline of Weakly-Supervised SGG" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 76, + 180, + 356, + 253 + ], + "lines": [ + { + "bbox": [ + 76, + 180, + 356, + 253 + ], + "spans": [ + { + "bbox": [ + 76, + 180, + 356, + 253 + ], + "type": "image", + "image_path": "eb48de475fefc6a4e19bcbefdee03faf14bd86c5ff720f3a4d86b5f01b9dd20f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 252, + 264, + 259 + ], + "lines": [ + { + "bbox": [ + 186, + 252, + 264, + 259 + ], + "spans": [ + { + "bbox": [ + 186, + 252, + 264, + 259 + ], + "type": "text", + "content": "(b) Predicate Distribution" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 265, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 545, + 296 + ], + "type": "text", + "content": "Figure 1. (a) The pipeline of weakly-supervised SGG. (b) The predicate distribution of unlocalized triplets (Parser+KB vs. Ours). In Parser+KB, the distribution becomes heavily long-tailed, and 12 out of 50 predicates are non-existent. (c) Semantic over-simplification caused by a rule-based parser in Step 2. (d) Low-density scene graph caused by the static structure of KB in Step 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 361, + 181, + 515, + 251 + ], + "blocks": [ + { + "bbox": [ + 361, + 181, + 515, + 251 + ], + "lines": [ + { + "bbox": [ + 361, + 181, + 515, + 251 + ], + "spans": [ + { + "bbox": [ + 361, + 181, + 515, + 251 + ], + "type": "image", + "image_path": "d5a9330d8e83d4927b4d110d691e08b082bc2e26c4608d32f158caa1a74932c6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 252, + 489, + 261 + ], + "lines": [ + { + "bbox": [ + 397, + 252, + 489, + 261 + ], + "spans": [ + { + "bbox": [ + 397, + 252, + 489, + 261 + ], + "type": "text", + "content": "(d) Low-Density Scene Graph" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 314, + 286, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 286, + 369 + ], + "type": "text", + "content": "textual object information to accurately ground the unlocalized entities, leveraging the linguistic structure embedded within the triplets. Another line of research [20] employs a pre-trained vision-language model [15] to reflect the semantic interactions among entities within the image caption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 369, + 286, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 286, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 286, + 531 + ], + "type": "text", + "content": "However, we argue that existing WSSGG approaches overlook the importance of the triplet formation process conducted in Step 2 and Step 3. We identify the two major issues described below, i.e., semantic over-simplication and low-density scene graph, which incur incomplete unlocalized triplets after Step 2 and 3. These incomplete triplets are mostly uninformative predicates with a limited number, and negatively impact the training of an SGG model even when entities are correctly grounded in Step 4. To demonstrate the impact of incomplete unlocalized triplets, we follow the conventional process to extract unlocalized triplets (i.e., Step 1-3), and conduct an examination of triplets obtained from COCO caption dataset, which are generated through Scene Parser [43] in Step 2 and WordNet [25] in Step 3. As a result, we identify the following two issues:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 536, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 686 + ], + "type": "text", + "content": "- Semantic Over-simplification: We find that the standard scene graph parser [43] operating on heuristic rule-based principles commonly used in Step 2 leads to a semantic over-simplification of predicates in extracted triplets. In other words, fine-grained predicates are undesirably converted into coarse-grained predicates, which we refer to as semantic over-simplification. For example, in Figure 1(c), an informative predicate lying on (i.e., fine-grained predicate) in the image caption is converted into a less informative predicate on (i.e., coarse-grained predicate), because the rule-based parser fails to capture the predicate lying on at once, and its heuristic rules fall short of accommodating the diverse range of caption's structure. As a result, the predicate distribution becomes heav-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 314, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 314, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 545, + 410 + ], + "type": "text", + "content": "ily long-tailed, in which coarse-grained predicates (e.g., with, on, in) greatly outnumber fine-grained predicates (e.g., parked on, covered in) (Figure 1(b)). To make the matter worse, numerous fine-grained predicates eventually end up in a frequency of 0, even though they are originally present in the captions. Specifically, 12 out of 50 predicates are non-existent, which means that these 12 predicates can never be predicted, since the model is not trained on these predicates at all." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": "- Low-Density Scene Graph: We find that the KB-based triplet alignment in Step 3 leads to low-density scene graphs, i.e., the number of remaining triplets after Step 3 is small. We attribute the low-density scene graphs primarily to the utilization of KB in Step 3. Specifically, a triplet is discarded if any of the three components (i.e., subject, predicate, object) or their synonym/hypernym/hyponym within the triplet fail to align with the entity or predicate classes in the target data. For example, in Figure 1(d), the triplet " + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": " elephant, carrying, " + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\log \\rangle" + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": " is discarded because " + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "inline_equation", + "content": "\\log" + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": " does not exist in Visual Genome dataset nor its synonym/hypernym, even if elephant and carrying do exist. In Table 1, we report the number of triplets and images in Visual Genome dataset, which is a common benchmark dataset used in fully-supervised SGG approaches, and COCO caption dataset, which is a common benchmark dataset used in weakly-supervised SGG approaches. We observe that on average Visual Genome dataset contains 7.1 triplets (i.e., " + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "inline_equation", + "content": "405\\mathrm{K} / 57\\mathrm{K}" + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": ") per image (See Table 1(a)), while COCO dataset contains only 2.4 triplets (i.e., " + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "inline_equation", + "content": "154\\mathrm{K} / 64\\mathrm{K}" + }, + { + "bbox": [ + 306, + 411, + 545, + 679 + ], + "type": "text", + "content": ") per image (See Table 1(b)). This indicates that existing WSSGG approaches suffer from the lack of sufficient supervision per image, leading to poor generalization and performance degradation [47, 49]. In summary, relying on" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28307" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 97, + 254, + 159 + ], + "blocks": [ + { + "bbox": [ + 78, + 97, + 254, + 159 + ], + "lines": [ + { + "bbox": [ + 78, + 97, + 254, + 159 + ], + "spans": [ + { + "bbox": [ + 78, + 97, + 254, + 159 + ], + "type": "table", + "html": "
DatasetHow to annotate# Triplet# Image
Fully-Supervised approach
(a) Visual GenomeManual405K57K
Weakly-Supervised approach
(b) COCO CaptionParser+KB154K64K
(c) COCO CaptionLLM344K64K
", + "image_path": "4f096a7d51a8bbafc4c63732fa578d09252e721de43cd7418cac20f21b7b80ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 159, + 248, + 170 + ], + "lines": [ + { + "bbox": [ + 85, + 159, + 248, + 170 + ], + "spans": [ + { + "bbox": [ + 85, + 159, + 248, + 170 + ], + "type": "text", + "content": "Table 1. Comparison of scene graph density." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 177, + 286, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 177, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 54, + 177, + 286, + 219 + ], + "type": "text", + "content": "the static structured knowledge of KB is insufficient to cover the semantic relationships among a wide a range of words, which incurs the low-density scene graph after Step 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 219, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 219, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 219, + 287, + 582 + ], + "type": "text", + "content": "To alleviate the semantic over-simplification and the low-density scene graph issues, we propose a new approach, namely Large Language Model for weakly-supervised SGG (LLM4SGG) that adopts a pre-trained Large Language Model (LLM), which has shown remarkable transferability to various downstream tasks in NLP such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Inspired by the idea of Chain-of-Thought² (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into two chains, each of which replaces the rule-based parser in Step 2 (i.e., Chain-1) and the KB in Step 3 (i.e., Chain-2). More precisely, we design a prompt for extracting triplets from a caption, and ask the LLM to extract triplets formed as (Chain-1). We expect that the predicates extracted based on a comprehensive understanding of the caption's context via LLM are semantically rich, thereby alleviating the semantic over-simplification issue. Besides, to alleviate the low-density scene graph issue, we additionally incorporate a paraphrased version of the original caption. To this end, we further design a prompt for paraphrasing the original caption and extracting more triplets from the paraphrased caption. However, entities and predicates in the triplets obtained after Chain-1 are not yet aligned with the target data. Hence, we design a another prompt to align them with entity/predicate classes of interest, and ask the LLM to align them with semantically relevant lexeme included in a predefined lexicon, which is the set of vocabularies that are present in the target data (Chain-2). To further engage the LLM in the reasoning process of Chain-1 and Chain-2, we employ the in-context few-shot learning that incorporates a few input-output examples within the prompt, enabling the LLM to perform the task without the need for fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 582, + 286, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 286, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 286, + 658 + ], + "type": "text", + "content": "To validate the effectiveness of LLM4SGG, we apply it to state-of-the-art WSSGG methods [54, 57]. Through extensive experiments, we show that LLM4SGG significantly enhances the performance of existing WSSGG methods in terms of mean Recall@K and Recall@K performance on Visual Genome and GQA datasets by alleviating the semantic over-simplification and the low-density scene graph" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 100, + 545, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 100, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 100, + 545, + 152 + ], + "type": "text", + "content": "(See Table 1(c) where the number of triplets increased to 334K). A further appeal of LLM4SGG is that it is data-efficient, i.e., it outperforms state-of-the-art baselines even with a small amount of training images, verifying the effectiveness of LLM4SGG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 153, + 521, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 521, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 521, + 163 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 164, + 544, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 164, + 544, + 195 + ], + "spans": [ + { + "bbox": [ + 310, + 164, + 544, + 195 + ], + "type": "text", + "content": "- We identify two major issues overlooked by existing WSSGG studies, i.e., semantic over-simplification and low-density scene graph." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 196, + 544, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 196, + 544, + 248 + ], + "spans": [ + { + "bbox": [ + 311, + 196, + 544, + 248 + ], + "type": "text", + "content": "- We leverage an LLM along with the CoT strategy and the in-context few-shot learning technique to extract informative triplets without the need for fine-tuning the LLM. To the best of our knowledge, we are the first to leverage an LLM for the SGG task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 249, + 544, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 249, + 544, + 292 + ], + "spans": [ + { + "bbox": [ + 311, + 249, + 544, + 292 + ], + "type": "text", + "content": "- LLM4SGG outperforms the state-of-the-art WSSGG methods, especially in terms of mR@K, demonstrating its efficacy in addressing the long-tail problem in WSSGG for the first time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 301, + 396, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 396, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 396, + 312 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 566 + ], + "type": "text", + "content": "Weakly-Supervised Scene Graph Generation (WSSGG). The WSSGG task aims to train an SGG model without relying on an annotated scene graph dataset. To achieve this, most WSSGG studies [20, 47, 54, 57] utilize image captions and ground unlocalized triplets with image regions. Specifically, VSPNet [49] proposes the iterative graph alignment algorithm to reflect the high-order relations between unlocalized triplets and image regions. SGNLS [57] uses a pre-trained object detector [29] to ground the entities in unlocalized triplets, which share the same classes with the output of object detectors. In addition to the information derived from the object detector, [20] employs a pre-trained vision-language model [15] to capture the semantic interactions among objects. VS3 [54] uses a grounding-based object detector [18], which calculates the similarity between the entity text in unlocalized triplet and image region, thereby grounding the unlocalized triplets. However, these methods overlook the triplet formation process that leads to the semantic over-simplification (Step 2) and the low-density scene graph (Step 3). In this regard, existing methods result in a sub-optimal performance even when unlocalized triplets are correctly grounded in Step 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 568, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 686 + ], + "type": "text", + "content": "Large Language Model (LLM). LLMs have demonstrated remarkable transferability to various downstream tasks such as symbolic reasoning, arithmetic, and common-sense reasoning [2, 5, 38]. Specifically, GPT-3 (175B) [2] stands as a cornerstone to break the line of numerous language tasks. Inspired by GPT-3, PaLM (540B) [5], LLaMA (65B) [38], OPT (175B) [53], and LaMDA (137B) [37] have been subsequently introduced. More recently, advanced GPT models (e.g., GPT-4 [28], ChatGPT [27]) fine-tuned with human feedback have gained prominence and widely applied for diverse applications, e.g., planner of tools [9, 24], mobile" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 667, + 286, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 667, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 667, + 286, + 686 + ], + "type": "text", + "content": "2We use the CoT strategy as a means to arrive at an answer in a stepwise manner, which differs from the Chain-of-Thought prompting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28308" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 99, + 287, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 287, + 142 + ], + "type": "text", + "content": "task automation [42]. In this work, we employ the power of LLM (i.e., ChatGPT) to alleviate the two issues, i.e., semantic over-simplification and low-density scene graph, in the context of the WSSGG task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 289, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 289, + 340 + ], + "type": "text", + "content": "In-Context Few-shot Learning. In-context few-shot learning incorporates a few input-output examples related to a target task, conditioning the LLM on the context of examples. Specifically, GPT-3 [2] pioneered the concept of in-context learning to facilitate an LLM as a versatile model on diverse tasks. This breakthrough has proliferated a plethora of research to leverage the in-context few-shot learning for various tasks [24, 26, 46, 55]. More precisely, Chameleon [24] integrates a few examples to enhance its understanding of tool planning task. [26] utilizes positive and negative examples related to questions for question generation tasks. ReAct [46] incorporates examples of reasoning with action for solving decision-making tasks. Inspired by recent in-context few-shot learning approaches, we provide a few examples to LLMs to help 1) understand the process of triplet extraction from a caption (i.e., Step 2), and 2) align the entity/predicate classes with the target data (i.e., Step 3) in the context of the WSSGG task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 350, + 103, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 103, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 103, + 361 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 287, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 531 + ], + "type": "text", + "content": "In this section, we describe LLM4SGG in detail. We would like to emphasize that LLM4SGG mainly focuses on the triplet formation process conducted in Step 2 (Parsing) and Step 3 (aligning), while existing WSSGG approaches mainly focus on Step 4 (grounding). We start by presenting the problem formulation of WSSGG (Section 3.1), followed by the prompt configuration (Section 3.2). Next, we introduce how LLMs are adopted to address the two issues of conventional WSSGG approaches when parsing the image caption (Section 3.3) and aligning the triplets in captions with entity/predicate classes of interest (Section 3.4). Finally, we ground the unlocalized triplets by associating them with bounding boxes (i.e., image regions) and train the SGG model using the localized triplets (Section 3.5). The overall pipeline of LLM4SGG is shown in Figure 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 539, + 171, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 171, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 171, + 550 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": "In the fully supervised SGG task, we aim to detect a scene graph " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_f = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_f}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " that consists of triplets given an image " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " is the number of triplets in the image. " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " subject and the object, respectively, whose bounding boxes are " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,b}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{i,b}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", and entity classes are " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}\\in \\mathcal{C}_e" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " is the set of predefined entity classes in the target data. " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " denotes the predicate between " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", and its class is " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,c}\\in \\mathcal{C}_p" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_p" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " is the set of predefined predicate classes in the target data. By using the ground truth scene graphs as supervision, fully supervised SGG approaches train an SGG model " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}:\\mathcal{I}\\to \\mathbf{G}_{f}" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": ", which maps an image to a scene graph." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": "In the weakly supervised SGG task, we aim to generate a scene graph when the ground truth scene graph is not given, i.e., there are no bounding boxes and entity/predicate class information. Instead, existing WSSGG approaches [20, 33, 47, 54, 57] use image captions along with associated images to produce scene graphs, i.e., localized triplets. More precisely, they extract a set of triplets " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w = \\{\\mathbf{s}_i,\\mathbf{p}_i,\\mathbf{o}_i\\}_{i = 1}^{N_w}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": " from the image captions, where " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "N_{w}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": " is the number of triplets extracted from the captions. However, while the extracted triplets contain the class information (i.e., " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,c},\\mathbf{o}_{i,c}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,c}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": "), they are unlocalized, since bounding boxes " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,b}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{i,b}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": " are not included in the caption. Therefore, it is essential to perform the grounding step to associate the unlocalized triplets with bounding boxes. Once we have obtained the localized triplets, we can apply the conventional SGG training scheme, i.e., " + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\theta}:\\mathcal{I}\\rightarrow \\mathbf{G}_{f}" + }, + { + "bbox": [ + 304, + 99, + 545, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": "In this paper, our focus is to address the semantic oversimplification and low-density scene graph issues regarding the unlocalized triplets " + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w" + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": ", which has been overlooked in existing WSSGG studies. Specifically, we aim to produce an enhanced " + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w" + }, + { + "bbox": [ + 304, + 270, + 545, + 378 + ], + "type": "text", + "content": " by refining the process of scene graph dataset construction via an LLM. This refinement includes the triplet extraction step from the caption (Step 2) and the alignment of entity/predicate classes (Step 3), leveraging the LLM's comprehensive understanding of language and reasoning ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 383, + 431, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 431, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 431, + 395 + ], + "type": "text", + "content": "3.2. Prompt Configuration" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": "In fact, it is not a trivial task for an LLM to immediately generate triplets from a caption whose entities and predicates are aligned with entity/predicate classes of interest, as such a task is a novel task for the LLM. Inspired by the idea of the Chain-of-thought (CoT) [41], which arrives at an answer in a stepwise manner, we separate the triplet formation process into the following two chains: Chain-1 - Extracting triplets from captions. Chain-2 - Aligning entities and predicates with the entity/predicate classes of interest. To carefully design each chain, we define the LLM function, i.e., " + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "inline_equation", + "content": "LLM(\\cdot)" + }, + { + "bbox": [ + 304, + 399, + 545, + 518 + ], + "type": "text", + "content": ", with the following prompt input:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 521, + 541, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 541, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 541, + 544 + ], + "type": "interline_equation", + "content": "O u t p u t = L L M (\\underbrace {\\text {T a s k d e s c r i p t i o n , I n - c o n t e x t e x a m p l e s , A c t u a l q u e s t i o n}} _ {\\text {P r o m p t i n p u t}},", + "image_path": "a712abf0f9f49368dc87b06ea3b3d86a5440f83c9fea198eba9a7c909edd4782.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "inline_equation", + "content": "LLM(\\cdot)" + }, + { + "bbox": [ + 304, + 544, + 545, + 687 + ], + "type": "text", + "content": " is the decoder of the LLM, generating the (1) Output given the prompt input. The prompt input consists of three components in a sequence: 1) task description, i.e., the delineation of the task that we intend to perform, 2) in-context examples, i.e., sample questions and answers related to the task at hand, 3) actual question, i.e., an inquiry from which we intend to derive the answer. Note that in-context examples is closely related to the in-context few-shot learning [2, 40, 41], which is shown to enhance the LLM's understanding of the task. Note that the above configuration of the prompt input is applied to the triplet extraction (Chain-1) (Section 3.3) and the alignment of entity/predicate classes (Chain-2) (Section 3.4)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28309" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 98, + 499, + 272 + ], + "blocks": [ + { + "bbox": [ + 92, + 98, + 499, + 272 + ], + "lines": [ + { + "bbox": [ + 92, + 98, + 499, + 272 + ], + "spans": [ + { + "bbox": [ + 92, + 98, + 499, + 272 + ], + "type": "image", + "image_path": "51b4929e428bf2e4b0ad9d458e510ab8301e43b735b4afbf5fb32a02f1a82010.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 546, + 316 + ], + "type": "text", + "content": "Figure 2. The pipeline of LLM4SGG. Given an image with its caption, we use an LLM to extract triplets from the original caption (Step 2-1) and the paraphrased caption (Step 2-2). Then, we align the entity/predicate classes within the extracted triplets with semantically similar lexeme in the target data via an LLM (Step 3), obtaining the unlocalized triplets. Lastly, we ground the unlocalized triplets over image regions (Step 4) followed by the training of an SGG model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 322, + 287, + 346 + ], + "type": "text", + "content": "3.3. Chain-1. Triplet Extraction via LLM (Step 2 in Figure 2)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 401 + ], + "type": "text", + "content": "Based on the LLM's comprehensive understanding of the context of an image caption, we aim to extract triplets from the caption. As discussed in Section 1, we use not only the original caption but also a paraphrased caption generated by the LLM to address the low-density scene graph issue." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": "Extracting triplets from a paraphrased caption (Step 2-2). To extract triplets from a paraphrased caption, we inform the LLM about the task at hand by presenting the following prompt: FROM THE GIVEN SENTENCE, THE TASK IS TO EXTRACT MEANINGFUL TRIPLETS FORMED AS " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": " SUBJECT, PREDICATE, OBJECT " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": " (i.e., Task description in Equation 1). We then instruct the LLM to follow the two steps, i.e., paraphrasing step and triplet extraction step. To help the LLM understand the process of performing the two steps, we present few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "1^{3}" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": "), which involves a manual construction of questions and corresponding answers to the paraphrasing and the triplet extraction steps. That is, for given a caption, we show the LLM how we expect a paraphrased caption and the extracted triplets would look like (e.g., Given \"Four clocks sitting on a floor next to a woman's feet,\" we show the LLM that a paraphrased sentence would be \"Four clocks are placed on the floor beside a woman's feet,\" and extracted triplets would be " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": " clocks, placed on, floor " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": " clocks, beside, feet " + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 46, + 402, + 287, + 670 + ], + "type": "text", + "content": "). Lastly, we show the caption of our interest to the LLM, and let the LLM extract triplets from the caption (i.e., Actual question in Equation 1). Please refer to Figure 2 right (bottom) for an example of the prompt input used in Step 2-2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 323, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 376 + ], + "type": "text", + "content": "Extracting triplets from the original caption (Step 2-1). As extracting triplets from the original caption does not involve the caption paraphrasing step, we exclude it from the prompt used in Step 2-2. Please refer to Figure 2 right (top) for an example of the prompt input used in Step 2-1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 377, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 431 + ], + "type": "text", + "content": "In summary, we obtain triplets from both the original and paraphrased captions after Step 2-1 and Step 2-2, respectively, which in turn alleviates the semantic oversimplification issue of predicates and the low-density scene graph issue." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 441, + 545, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 464 + ], + "type": "text", + "content": "3.4. Chain-2. Alignment of Classes in Triplets via LLM (Step 3 in Figure 2)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 467, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 521 + ], + "type": "text", + "content": "The entities (i.e., subject and object) and predicates within the triplets obtained from Step 2 described in Section 3.3 are not yet aligned with the target data. Based on the semantic reasoning ability of the LLM, we aim to align them with the semantically relevant lexeme in the target data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "text", + "content": "Aligning entities in the triplets with entity classes of interest. We instruct the LLM with the following prompt: GIVEN THE LEXEME " + }, + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{ENTITY}\\}" + }, + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "text", + "content": ", FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED ENTITY LEXICON, where the predefined entity lexicon is " + }, + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 304, + 525, + 545, + 687 + ], + "type": "text", + "content": " (i.e., Task description in Equation 1). Similar to Section 3.3, we present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding hierarchical relationships such as pigeon being semantically relevant to bird, and singular-plural relationships such as surfboards being semantically relevant to surfboard. Lastly, we show the entity of our interest to the LLM (i.e., Actual question in Equation 1), which enables the LLM to generate an answer by finding the most" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 676, + 243, + 686 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 676, + 243, + 686 + ], + "spans": [ + { + "bbox": [ + 58, + 676, + 243, + 686 + ], + "type": "text", + "content": "3We use captions in COCO caption dataset for examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28310" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 99, + 287, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 287, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 287, + 121 + ], + "type": "text", + "content": "semantically relevant entity in " + }, + { + "bbox": [ + 47, + 99, + 287, + 121 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e" + }, + { + "bbox": [ + 47, + 99, + 287, + 121 + ], + "type": "text", + "content": ". Please refer to Table 7 in Appendix A.1 for an example of the prompt input." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "text", + "content": "Aligning predicates in the triplets with predicate classes of interest. Likewise, we instruct the LLM with the following prompt: GIVEN THE LEXEME " + }, + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{PREDICATE}\\}" + }, + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "text", + "content": ", FIND SEMANTICALLY RELEVANT LEXEME IN THE PREDEFINED Predicate Lexicon, where the predefined predicate lexicon is " + }, + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "inline_equation", + "content": "C_p" + }, + { + "bbox": [ + 47, + 123, + 289, + 369 + ], + "type": "text", + "content": " (i.e., Task description in Equation 1). We also present a few examples to the LLM that describe how to answer the questions (i.e., In-context examples in Equation 1). For example, we provide the LLM with a few examples regarding tense relationships such as the lies on being semantically relevant to lying on, and positional relationship such as next to being semantically relevant to near. Lastly, we show the predicate of our interest to the LLM (i.e., Actual question in Equation 1). Please refer to Table 8 in Appendix A.1 for an example of the prompt input. Furthermore, regarding the alignment of classes within a large predefined lexicon, please refer to Appendix A.2. Note that as our main focus in this work is on developing a framework for utilizing LLMs in weakly-supervised SGG rather than finding a specific prompt design that works the best, we tested with a single prompt design. However, since prompt designs are crucial in leveraging LLMs, we plan to explore various designs in our future work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": "After performing Chain-1 (Section 3.3) and Chain-2 (Section 3.4), we obtain intermediate unlocalized triplets " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{G}}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{\\hat{N}_w}" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\{\\mathcal{C}_e \\cup \\text{None}\\}" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,c} \\in \\{\\mathcal{C}_p \\cup \\text{None}\\}" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": ". It is worth noting that if there is no semantically relevant lexeme, we request the LLM to generate None as the answer, due to the fact that the entity/predicate classes in the target data may not cover a wide range of entities/predicate. Similar to the conventional approach, we discard a triplet if any of its three components (i.e., subject, predicate and object) is None. Lastly, we obtain the final unlocalized triplets " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w = \\{\\mathbf{s}_i, \\mathbf{p}_i, \\mathbf{o}_i\\}_{i=1}^{N_w}(N_w \\leq \\hat{N}_w)" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,c}, \\mathbf{o}_{i,c} \\in \\mathcal{C}_e" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,c} \\in \\mathcal{C}_p" + }, + { + "bbox": [ + 47, + 369, + 289, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 506, + 144, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 144, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 144, + 517 + ], + "type": "text", + "content": "3.5. Model Training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "content": "Given the final unlocalized triplets " + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w" + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "content": ", we ground them over relevant image regions to get localized triplets, meaning that we obtain " + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{i,b}" + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_{i,b}" + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "content": ". To this end, we employ two state-of-the-art grounding methods, i.e., SGNLS [57] and VS3 [54]. Please refer to Appendix A.3 for more detail about how each method performs grounding. After grounding " + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_w" + }, + { + "bbox": [ + 46, + 521, + 287, + 631 + ], + "type": "text", + "content": ", we obtain localized triplets and use them as pseudo-labels for training a supervised SGG model. Please refer to Appendix A.4 for more detail about the model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 124, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 124, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 124, + 648 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 654, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 687 + ], + "type": "text", + "content": "Datasets. To train an SGG model without an annotated scene graph dataset, we use three caption datasets: COCO caption [3], Conceptual (CC) caption [31], and Visual" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 100, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 100, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 100, + 545, + 314 + ], + "type": "text", + "content": "Genome (VG) caption [44]. For fair comparisons, we use the same set of images that have been utilized in previous WSSGG studies [20, 47, 54, 57], leading to the utilization of 64K images on COCO caption dataset, 145K images on CC caption dataset, and 57K images on VG caption dataset. To evaluate the trained SGG model, we employ the widely used Visual Genome (VG) dataset [13] and GQA dataset [11]. The VG dataset contains the ground-truth localized triplet information annotated by humans. We follow the standard split of VG [44], which consist of 150 entity classes and 50 predicate classes. For the GQA dataset used in the SGG task, we follow the same pre-processing step of a previous SGG study [8], which involves selecting top-200 frequent entity classes and top-100 frequent predicate classes. In both datasets, " + }, + { + "bbox": [ + 304, + 100, + 545, + 314 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 100, + 545, + 314 + ], + "type": "text", + "content": " of the total images are used for evaluation. Please refer to Appendix C.1 for more details regarding the datasets. Note that we mainly use the COCO caption dataset for analysis of LLM4SGG throughout this paper, i.e., CC and VG caption datasets are exclusively used in quantitative result on VG (Section 4.1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 316, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 509 + ], + "type": "text", + "content": "Evaluation metrics. Recent fully-supervised SGG studies [1, 48, 51] have emphasized improving the accuracy of predictions for fine-grained predicates rather than coarse-grained predicates, since the former construct richer scene graphs. As a result, they commonly use mean Recall@K (mR@K) that computes the average of Recall@K (R@K) across all predicates. In line with the recent emphasis on fine-grained predicates, we incorporate both mR@K and R@K in our evaluation, whereas previous WSSGG studies [20, 47, 54] mainly rely on the R@K metric alone. Moreover, we report F@K, which is the harmonic average of R@K and mR@K to jointly consider R@K and mR@K, following previous SGG studies [12, 51]. Regarding the evaluation task, we follow previous WSSGG studies and adopt the Scene Graph Detection (SGDet) task, where both the ground-truth bounding box and the entity class information are not provided. Please refer to Appendix C.2 for more detail regarding the task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 533 + ], + "type": "text", + "content": "Baselines. Please refer to Appendix C.3 for details regarding the baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 556 + ], + "type": "text", + "content": "Implementation details. Please refer to Appendix C.4 regarding the implementation details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 562, + 452, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 452, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 452, + 574 + ], + "type": "text", + "content": "4.1. Quantitative Result on VG" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 687 + ], + "type": "text", + "content": "Table 2 shows the performance of baseline models and those when LLM4SGG is applied. We have the following observations based on COCO caption dataset: 1) Applying LLM4SGG to SGNLS and VS3 improves the performance in terms of R@K and mR@K, which demonstrates the effectiveness of the triplet formation through LLM4SGG. Notably, LLM4SGG significantly improves mR@K, implying that LLM4SGG effectively alleviates the long-tailed problem in WSSGG. This can be clearly seen in Figure 3, which shows the performance gain on fine" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 315, + 715 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 315, + 715 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 315, + 715 + ], + "type": "text", + "content": "28311" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 94, + 490, + 212 + ], + "blocks": [ + { + "bbox": [ + 102, + 94, + 490, + 212 + ], + "lines": [ + { + "bbox": [ + 102, + 94, + 490, + 212 + ], + "spans": [ + { + "bbox": [ + 102, + 94, + 490, + 212 + ], + "type": "table", + "html": "
MethodDatasetR@50R@100mR@50mR@100F@50F@100
Motif (CVPR'18) - Fully-supervisedVG31.8936.366.387.5710.63 / 12.5312.53
LSWS (CVPR'21)COCO Caption3.293.693.273.663.283.67
SGNLS (ICCV'21)3.804.462.512.783.023.43
SGNLS (ICCV'21)+LLM4SGG5.09+1.295.97+1.514.08+1.574.49+1.714.53+1.515.13+1.70
Li et al (MM'22)6.407.331.731.982.723.12
VS3(CVPR'23)6.608.012.883.254.014.62
VS3(CVPR'23)+LLM4SGG8.91+2.3110.43+2.427.11+4.238.18+4.937.91+3.909.17+4.55
VS3(CVPR'23)+Rwt4.255.045.175.994.675.47
VS3(CVPR'23)+Rwt+LLM4SGG5.10+0.856.34+1.308.42+3.259.90+3.916.35+1.697.73+2.26
VS3(CVPR'23)CC Caption6.698.201.732.042.753.27
VS3(CVPR'23)+LLM4SGG9.47+2.7810.69+2.495.40+3.676.09+4.056.88+4.137.76+4.49
VS3(CVPR'23)VG Caption14.5418.482.803.794.706.29
VS3(CVPR'23)+LLM4SGG18.40+3.8622.28+3.806.26+3.467.60+3.819.34+4.6411.33+5.04
", + "image_path": "0c29bae517e9c8df7a5aa7b2f4c3863b00d1e28593a4bb9d65e7eb50c69201f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 247, + 265, + 324 + ], + "blocks": [ + { + "bbox": [ + 65, + 247, + 265, + 324 + ], + "lines": [ + { + "bbox": [ + 65, + 247, + 265, + 324 + ], + "spans": [ + { + "bbox": [ + 65, + 247, + 265, + 324 + ], + "type": "image", + "image_path": "aca0a378acdb45627d51d035e807e92e5f8de4db7d5f790dc07ef1c6f8e7c49a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 325, + 287, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 346 + ], + "type": "text", + "content": "Figure 3. Per class performance (Bar: number of predicate instances, Line: Recall@100)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": "grained predicates. 2) " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3 +\\mathrm{Rwt} + \\mathrm{LLM4SGG}" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": " further improves mR@K of " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3 +\\mathrm{Rwt}" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": ". We attribute this to the fact that the conventional approach generates a limited number of fine-grained predicates, which makes the reweighting strategy less effective within " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": ". Especially, non-existent predicates can never be predicted even when the reweighting strategy is applied. On the other hand, LLM4SGG increases the number of instances that belong to fine-grained predicates, which is advantageous for the reweighting strategy. For the per class performance comparison over the reweighting stregty, please refer to Appendix D.1. 3) The performance gain obtained from applying LLM4SGG is greater on " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3 +\\mathrm{LLM4SGG}" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": ") than on SGNLS (i.e., SGNLS+LLM4SGG). The major reason lies in the difference in how SGNLS and " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": " make use of the pool of 344K unlocalized triplets obtained through LLM4SGG. Specifically, in the grounding process of SGNLS, we observe that 100K out of 344K unlocalized triplets (i.e., " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": ") fail to be grounded, and thus not used for training. On the other hand, " + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3" + }, + { + "bbox": [ + 46, + 351, + 287, + 652 + ], + "type": "text", + "content": " successfully grounds all 344K unlocalized triplets and fully utilize them for training, allowing it to fully enjoy the effectiveness of LLM4SGG. This indicates that LLM4SGG makes synergy when paired with a grounding method that is capable of fully utilizing the unlocalized triplets. For more details regarding the impact of grounding methods, please refer to Appendix B. 4) Regarding the performance comparison with a fully-supervised approach (i.e., Motif), please refer to Appendix D.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 286, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 686 + ], + "type": "text", + "content": "Furthermore, we observe that applying LLM4SGG to CC and VG caption datasets also significantly improves performance, demonstrating the effectiveness of LLM4SGG." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 246, + 529, + 293 + ], + "blocks": [ + { + "bbox": [ + 46, + 215, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 215, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 545, + 236 + ], + "type": "text", + "content": "Table 2. Performance comparisons on the SGDet task. The best performance among WSSGG models within each dataset is in bold. The red numbers indicate the absolute performance improvement after applying LLM4SGG. Rwt denotes using the reweighting strategy [51]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 246, + 529, + 293 + ], + "lines": [ + { + "bbox": [ + 321, + 246, + 529, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 246, + 529, + 293 + ], + "type": "table", + "html": "
RowPCLPLA# TripletR@50 / 100mR@50 / 100F@50 / 100
(a)154K6.60 / 8.012.88 / 3.254.01 / 4.62
(b)✓243K9.46 / 11.223.43 / 3.925.03 / 5.81
(c)✓✓256K8.42 / 9.855.99 / 6.957.00 / 8.15
(d)✓✓327K11.76 / 13.383.50 / 4.055.39 / 6.22
(e)✓✓✓344K8.91 / 10.437.11 / 8.187.91 / 9.17
", + "image_path": "0fc2459895c020cdc71c25092028e92c0941974e66776cccbc7fa973d277a120.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 296, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 327 + ], + "type": "text", + "content": "Table 3. Ablation studies. (PC: Using Paraphrased Caption in addition to the original caption / LP: LLM-based Parsing / LA: LLM-based Alignment)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 329, + 406, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 329, + 406, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 329, + 406, + 340 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": "In Table 3, we conduct ablation studies on VG dataset to understand the effectiveness of each component of LLM4SGG, where VS" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": " is used as the grounding method. Note that row (a) is equivalent to vanilla VS" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": ". We have the following observations. 1) Effect of using the paraphrased caption: Including the paraphrased caption in addition to the original caption (row (b)) increases the number of triplets " + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "inline_equation", + "content": "(154\\mathrm{K}\\rightarrow 243\\mathrm{K})" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": ", resulting in an improved overall performance. This demonstrates that the paraphrased caption alleviates the low-density scene graph issue. 2) Effect of LLM-based parsing: The LLM-based parsing (row (c)) for extracting triplets improves mR@K of row (b). This indicates that the LLM-based parsing increases the number of instances that belong to fine-grained predicates, which in turn alleviates the semantic over-simplification issue. 3) Effect of LLM-based alignment: The LLM-based alignment (row (d)) of entities/predicates in the extracted triplets increases the number of triplets from " + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "inline_equation", + "content": "243\\mathrm{K}" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "inline_equation", + "content": "327\\mathrm{K}" + }, + { + "bbox": [ + 304, + 344, + 545, + 687 + ], + "type": "text", + "content": " (row (b) vs (d)), which indicates that the low-density scene graph issue is alleviated. Consequently, R@K and mR@K of row (d) are greater than those of row (b). 4) The fully-fledged approach (row (e)) generally improves R@K and mR@K, showing the best performance in terms of F@K. It is important to highlight that when using the LLM-based parsing, the performance of mR@K significantly increases with a moderate decrease in R@K. This trade-off is attributed to the fact that R@K generally improves when the coarse-grained predicates are dominant [51]. In contrast, our approach, which addresses the semantic over-simplification issue, decreases the number instances that belong to coarse-grained predicates while simultaneously increasing those that belong to fine-grained predicates (Figure 1(b)), which" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28312" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 99, + 271, + 170 + ], + "blocks": [ + { + "bbox": [ + 60, + 99, + 271, + 170 + ], + "lines": [ + { + "bbox": [ + 60, + 99, + 271, + 170 + ], + "spans": [ + { + "bbox": [ + 60, + 99, + 271, + 170 + ], + "type": "image", + "image_path": "6ee00aff165015599ee6dfdea78d6088413e6eb7c93f798dbde1e0d1fc66b273.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 171, + 287, + 192 + ], + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 192 + ], + "type": "text", + "content": "Figure 4. Performance over various numbers of images used for training VS" + }, + { + "bbox": [ + 46, + 171, + 287, + 192 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 46, + 171, + 287, + 192 + ], + "type": "text", + "content": " + LLM4SGG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 197, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 287, + 240 + ], + "type": "text", + "content": "in turn results in a substantial improvement in mR@K. We would like to emphasize that the performance in terms of mR@K is crucial in the context of SGG research [1, 8, 48], as fine-grained predicates offer richer information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 248, + 196, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 196, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 196, + 260 + ], + "type": "text", + "content": "4.3. Analysis of Data-Efficiency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": "To assess the effectiveness of LLM4SGG under the lack of available training data, we conduct experiments given a limited number of images. Specifically, among 64K images used for training VS" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": " and VS" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": "+LLM4SGG in Table 2, we randomly sample 1K (1.5%), 5K (7.8%), and 10K (15.6%) images with replacement, and train VS" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": "+LLM4SGG for five times. Figure 4 shows the average performance over various numbers of images used for training VS" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": "+LLM4SGG along with the variance (in blue area). We observe that when using only 1K images, the performance is slightly inferior to the baseline that used 64K for training (i.e., VS" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 45, + 264, + 287, + 534 + ], + "type": "text", + "content": "). However, as we increase the number of images used for training to 5K images, we observe a significant improvement in both R@K and mR@K compared with the baseline. This demonstrates that LLM4SGG is data-efficient, as it outperforms the baseline even with only 7.8% of the total images used for training the baseline. Moreover, when further increasing the number of images to 10K, we observe further performance improvements, and when it reaches 64K, which is the same as the number of images used for training the baseline, the performance is the best. In summary, LLM4SGG enables data-efficient model training even with a limited amount of available images for training, thanks to alleviating the semantic over-simplification and low-density scene graph issues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 202, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 202, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 202, + 552 + ], + "type": "text", + "content": "4.4. Quantitative Result on GQA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 687 + ], + "type": "text", + "content": "In Table 4, we additionally conducted experiments on GQA dataset [11]. Please refer to Appendix E.1 for more detailed descriptions on the training and evaluation processes on GQA dataset. The GQA dataset contains twice as many predicates as the Visual Genome dataset and includes complicated predicates (e.g., sitting next to, standing in front of). As a result, when obtaining unlocalized triplets using the conventional WSSGG approach, we observe that 44 out of 100 predicates have a frequency of 0, and the predicate distribution is extremely long-tailed. Consequently, the baseline (i.e., " + }, + { + "bbox": [ + 46, + 557, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{VS}^3" + }, + { + "bbox": [ + 46, + 557, + 287, + 687 + ], + "type": "text", + "content": ") exhibits significantly lower performance, especially in terms of mR@K. On the other hand," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 97, + 525, + 139 + ], + "blocks": [ + { + "bbox": [ + 323, + 97, + 525, + 139 + ], + "lines": [ + { + "bbox": [ + 323, + 97, + 525, + 139 + ], + "spans": [ + { + "bbox": [ + 323, + 97, + 525, + 139 + ], + "type": "table", + "html": "
MethodR@50 / 100mR@50 / 100F@50 / 100
Motif (Fully-supervised)28.90 / 33.106.40 / 7.7010.48 / 12.49
VS35.90 / 6.971.60 / 1.812.52 / 2.87
VS3+LLM4SGG8.88 / 10.385.33 / 6.516.66 / 8.00
", + "image_path": "43e6e5bd543c146321c6a512f648e89f957f8159bb3e61fdda21ef2e5bce20f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 345, + 143, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 345, + 143, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 345, + 143, + 504, + 152 + ], + "type": "text", + "content": "Table 4. Performance comparison on GQA." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 159, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 234 + ], + "type": "text", + "content": "our approach shows substantial performance improvements not only in R@K but also in mR@K, thanks to the mitigation of semantic over-simplification and low-density scene graph issues. Please refer to Appendix E.2 for the predicate distribution and performance comparison for each class in GQA dataset. Additionally, please refer to Appendix E.3 for qualitative analyses on GQA dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 240, + 458, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 458, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 458, + 251 + ], + "type": "text", + "content": "5. Conclusion & Future work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 471 + ], + "type": "text", + "content": "In this work, we focus on the triplet formation process in WSSGG, whose importance is overlooked by previous studies. To alleviate the semantic over-simplification and low-density scene graph issues inherent in the triplet formation process, we propose a new approach, i.e., LLM4SGG, which leverages a pre-trained LLM during the extraction of triplets from the captions, and alignment of entity/predicate classes with those in the target data. It is important to note that construction of these triplets is not required every time to train the SGG model; instead, it is a one-time pre-processing step. In this regard, we contribute to generating enhanced triplets compared to the conventional approach. Moreover, we publish an enhanced SGG dataset constructed by LLM for future studies of SGG. As a result, we outperform baselines in terms of R@K, mR@K and F@K on Visual Genome and GQA datasets. A potential limitation of our work is the reliance on a proprietary blackbox LLM, which can also be costly to use. Hence, in Appendix F, we provide discussions on replacing the LLM with smaller language models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": "For future work, an LLM can be used to ground the unlocalized triplets in Step 4. Recently, vision-language representation learning has been developed for transforming visual features into textual features to facilitate the use of visual features as input to an LLM [16, 58]. In this regard, given the visual features of bounding boxes as input, we could ask the LLM to identify relevant bounding boxes based on the textual information of entities within unlocalized triplets using the comprehensive understanding of the context of triplets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 584, + 416, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 584, + 416, + 596 + ], + "spans": [ + { + "bbox": [ + 305, + 584, + 416, + 596 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 600, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 687 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.2020-0-00004, Development of Provisional Intelligence based on Long-term Visual Memory Network, No.2022-0-00077, Reasoning, and Inference from Heterogeneous Data, No. 2019-0-00079, Artificial Intelligence Graduate School Program, Korea University)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28313" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 99, + 106, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 99, + 106, + 110 + ], + "spans": [ + { + "bbox": [ + 48, + 99, + 106, + 110 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 117, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 117, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 117, + 287, + 156 + ], + "type": "text", + "content": "[1] Bashirul Azam Biswas and Qiang Ji. Probabilistic debiasing of scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10429-10438, 2023. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 157, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 288, + 206 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 207, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 287, + 246 + ], + "type": "text", + "content": "[3] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 296 + ], + "type": "text", + "content": "[4] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European conference on computer vision, pages 104-120. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 297, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 287, + 346 + ], + "type": "text", + "content": "[5] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 396 + ], + "type": "text", + "content": "[6] Xiyang Dai, Yinpeng Chen, Bin Xiao, Dongdong Chen, Mengchen Liu, Lu Yuan, and Lei Zhang. Dynamic head: Unifying object detection heads with attentions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7373-7382, 2021. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 435 + ], + "type": "text", + "content": "[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 495 + ], + "type": "text", + "content": "[8] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427–19436, 2022. 6, 8, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 496, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 496, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 496, + 287, + 535 + ], + "type": "text", + "content": "[9] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn. arXiv preprint arXiv:2306.08640, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 286, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 575 + ], + "type": "text", + "content": "[10] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 576, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 286, + 624 + ], + "type": "text", + "content": "[11] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6, 8, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 655 + ], + "type": "text", + "content": "[12] Siddhesh Khandelwal and Leonid Sigal. Iterative scene graph generation. Advances in Neural Information Processing Systems, 35:24295-24308, 2022. 6, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 656, + 286, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 286, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 286, + 686 + ], + "type": "text", + "content": "[13] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 101, + 545, + 686 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 325, + 101, + 545, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 101, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 325, + 101, + 545, + 130 + ], + "type": "text", + "content": "Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1, 6, 2, 3, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 131, + 545, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 131, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 307, + 131, + 545, + 198 + ], + "type": "text", + "content": "[14] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2,3,4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 200, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 545, + 249 + ], + "type": "text", + "content": "[15] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 250, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 289 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 289, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 328 + ], + "type": "text", + "content": "[17] Lin Li, Guikun Chen, Jun Xiao, Yi Yang, Chunping Wang, and Long Chen. Compositional feature augmentation for unbiased scene graph generation. arXiv preprint arXiv:2308.06712, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 329, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 329, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 329, + 545, + 388 + ], + "type": "text", + "content": "[18] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3, 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 437 + ], + "type": "text", + "content": "[19] Rongjie Li, Songyang Zhang, Bo Wan, and Xuming He. Bipartite graph network with adaptive message passing for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11109-11119, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 487 + ], + "type": "text", + "content": "[20] Xingchen Li, Long Chen, Wenbo Ma, Yi Yang, and Jun Xiao. Integrating object-aware and interaction-aware knowledge for weakly supervised scene graph generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4204-4213, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 487, + 545, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 527 + ], + "type": "text", + "content": "[21] Yongzhi Li, Duo Zhang, and Yadong Mu. Visual-semantic matching by exploring high-order attention and distraction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786–12795, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 527, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 527, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 527, + 545, + 567 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 567, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 616 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 616, + 545, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 616, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 307, + 616, + 545, + 665 + ], + "type": "text", + "content": "[24] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 666, + 545, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 545, + 686 + ], + "type": "text", + "content": "[25] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28314" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 100, + 287, + 686 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 100, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 100, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 100, + 287, + 140 + ], + "type": "text", + "content": "[26] Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task generalization via natural language crowdsourcing instructions. arXiv preprint arXiv:2104.08773, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 141, + 286, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 159 + ], + "type": "text", + "content": "[27] OpenAI. Chatgpt. https://openai.com/blog/chatgpt, 2023. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 221, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 221, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 221, + 172 + ], + "type": "text", + "content": "[28] OpenAI. Gpt-4. Technical report, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 286, + 212 + ], + "type": "text", + "content": "[29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 1, 3, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 213, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 286, + 262 + ], + "type": "text", + "content": "[30] Sebastian Schuster, Ranjay Krishna, Angel Chang, Li Fei-Fei, and Christopher D Manning. Generating semantically precise scene graphs from textual descriptions for improved image retrieval. In Proceedings of the fourth workshop on vision and language, pages 70–80, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 286, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 286, + 322 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 286, + 322 + ], + "type": "text", + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 323, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 323, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 286, + 363 + ], + "type": "text", + "content": "[32] Jiaxin Shi, Hanwang Zhang, and Juanzi Li. Explainable and explicit visual reasoning over scene graphs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8376-8384, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 364, + 286, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 286, + 412 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 286, + 412 + ], + "type": "text", + "content": "[33] Jing Shi, Yiwu Zhong, Ning Xu, Yin Li, and Chenliang Xu. A simple baseline for weakly-supervised scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16393-16402, 2021. 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 286, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 464 + ], + "type": "text", + "content": "[34] Mohammed Suhail, Abhay Mittal, Behjat Siddiquie, Chris Broaddus, Jayan Eledath, Gerard Medioni, and Leonid Sigal. Energy-based learning for scene graph generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13936-13945, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 465, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 286, + 514 + ], + "type": "text", + "content": "[35] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 286, + 555 + ], + "type": "text", + "content": "[36] Damien Teney, Lingqiao Liu, and Anton van Den Hengel. Graph-structured representations for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1–9, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 286, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 286, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 286, + 604 + ], + "type": "text", + "content": "[37] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 606, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 606, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 606, + 286, + 655 + ], + "type": "text", + "content": "[38] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 9" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 656, + 287, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 656, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 656, + 287, + 686 + ], + "type": "text", + "content": "[39] Sijin Wang, Ruiping Wang, Ziwei Yao, Shiguang Shan, and Xilin Chen. Cross-modal scene graph matching for relationship-aware image-text retrieval. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 100, + 545, + 686 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 325, + 100, + 545, + 121 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 100, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 325, + 100, + 545, + 121 + ], + "type": "text", + "content": "the IEEE/CVF winter conference on applications of computer vision, pages 1508-1517, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 122, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 122, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 545, + 169 + ], + "type": "text", + "content": "[40] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 220 + ], + "type": "text", + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 222, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 270 + ], + "type": "text", + "content": "[42] Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. Empowering llm to use smartphone for intelligent task automation. arXiv preprint arXiv:2308.15272, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 273, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 273, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 307, + 273, + 545, + 332 + ], + "type": "text", + "content": "[43] Hao Wu, Jiayuan Mao, Yufeng Zhang, Yuning Jiang, Lei Li, Weiwei Sun, and Wei-Ying Ma. Unified visual-semantic embeddings: Bridging vision and language with structured meaning representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6609–6618, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 333, + 545, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 372 + ], + "type": "text", + "content": "[44] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410–5419, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 373, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 413 + ], + "type": "text", + "content": "[45] Xu Yang, Kaihua Tang, Hanwang Zhang, and Jianfei Cai. Auto-encoding scene graphs for image captioning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10685–10694, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 453 + ], + "type": "text", + "content": "[46] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 454, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 502 + ], + "type": "text", + "content": "[47] Keren Ye and Adriana Kovashka. Linguistic structures as weak supervision for visual scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8289-8299, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 505, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 545, + 554 + ], + "type": "text", + "content": "[48] Kanghoon Yoon, Kibum Kim, Jinyoung Moon, and Chanyoung Park. Unbiased heterogeneous scene graph generation with relation-aware message passing neural network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3285-3294, 2023. 1, 6, 8, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 555, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 594 + ], + "type": "text", + "content": "[49] Alireza Zareian, Svebor Karaman, and Shih-Fu Chang. Weakly supervised visual semantic parsing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3736-3745, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 596, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 596, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 596, + 545, + 644 + ], + "type": "text", + "content": "[50] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 1, 3, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 686 + ], + "type": "text", + "content": "[51] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In European conference on computer vision, pages 409-424. Springer, 2022. 6, 7, 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 705, + 316, + 714 + ], + "type": "text", + "content": "28315" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 100, + 288, + 438 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 100, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 100, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 100, + 287, + 158 + ], + "type": "text", + "content": "[52] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "type": "text", + "content": "[53] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 287, + 269 + ], + "type": "text", + "content": "[54] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Learning to generate language-supervised and open-vocabulary scene graph using pretrained visual-semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915–2924, 2023. 1, 3, 4, 6, 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 270, + 286, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 286, + 299 + ], + "type": "text", + "content": "[55] Zhuosheng Zhang, Aston Zhang, Mu Li, and Alex Smola. Automatic chain of thought prompting in large language models. arXiv preprint arXiv:2210.03493, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 300, + 286, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 286, + 349 + ], + "type": "text", + "content": "[56] Yiwu Zhong, Liwei Wang, Jianshu Chen, Dong Yu, and Yin Li. Comprehensive image captioning via scene graph decomposition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 211-229. Springer, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 350, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 398 + ], + "type": "text", + "content": "[57] Yiwu Zhong, Jing Shi, Jianwei Yang, Chenliang Xu, and Yin Li. Learning to generate scene graph from natural language supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1823-1834, 2021. 1, 3, 4, 6, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 400, + 286, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 286, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 286, + 438 + ], + "type": "text", + "content": "[58] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 8" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 706, + 316, + 714 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 706, + 316, + 714 + ], + "spans": [ + { + "bbox": [ + 293, + 706, + 316, + 714 + ], + "type": "text", + "content": "28316" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 611, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_content_list.json b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..327c241cac7bf4c22345d2c39baf2ab358e0eb2f --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_content_list.json @@ -0,0 +1,1747 @@ +[ + { + "type": "text", + "text": "LLMs are Good Action Recognizers", + "text_level": 1, + "bbox": [ + 300, + 130, + 668, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoxuan $\\mathrm{Qu}^{1}$ Yujun Cai $^{2}$ Jun Liu $^{1\\dagger}$ $^{1}$ Singapore University of Technology and Design $^{2}$ Nanyang Technological University", + "bbox": [ + 140, + 179, + 828, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "haoxuan_qu@mymail.sutd.edu.sg, yujun001@e.ntu.edu.sg, jun.liu@sutd.edu.sg", + "bbox": [ + 158, + 219, + 799, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Skeleton-based action recognition has attracted lots of research attention. Recently, to build an accurate skeleton-based action recognizer, a variety of works have been proposed. Among them, some works use large model architectures as backbones of their recognizers to boost the skeleton data representation capability, while some other works pretrain their recognizers on external data to enrich the knowledge. In this work, we observe that large language models which have been extensively used in various natural language processing tasks generally hold both large model architectures and rich implicit knowledge. Motivated by this, we propose a novel LLM-AR framework, in which we investigate treating the Large Language Model as an Action Recognizer. In our framework, we propose a linguistic projection process to project each input action signal (i.e., each skeleton sequence) into its \"sentence format\" (i.e., an \"action sentence\"). Moreover, we also incorporate our framework with several designs to further facilitate this linguistic projection process. Extensive experiments demonstrate the efficacy of our proposed framework.", + "bbox": [ + 75, + 316, + 473, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human action recognition aims to categorize the actions performed by humans into a pre-defined list of classes. It is relevant to a variety of applications, such as human-computer interaction [37], intelligent surveillance [36], and virtual reality [13]. In the past few years, skeleton-based action recognition [3, 5, 6, 12, 53, 54, 62, 70] has received a lot of research attention with the notice that skeleton is a succinct yet informative representation of human behaviors. Yet, despite the considerable progress, skeleton-based action recognition still remains a challenging task [78], and to build a more accurate skeleton-based action recognizer, various recent works have been proposed from different perspectives. Among them, some recent works [15, 59] pro", + "bbox": [ + 75, + 676, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f731809c1a45824c07318e0cca1029a4d81a7630cf8c820ac60412eb8131957b.jpg", + "image_caption": [ + "Figure 1. Overview of our proposed LLM-AR framework. In our framework, given an input action signal, we first perform a linguistic projection process to acquire the corresponding \"action sentence\". We then perform action recognition via the large language model with its pre-trained weights untouched to keep its pre-learned rich knowledge." + ], + "image_footnote": [], + "bbox": [ + 501, + 268, + 893, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "posed to utilize large model architectures (such as transformers) as the backbone of their action recognizers to achieve stronger representation capability and capture more subtle differences among different actions. On the other hand, some other works [72, 80] proposed to pre-train their action recognizers on external data in order for their action recognizers to handle this task with richer knowledge.", + "bbox": [ + 496, + 458, + 893, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large language models such as GPT [2] and LLaMA [56] have become quite popular and have been extensively applied in handling various human languages. For example, having been pre-trained over various human languages and learned common language-related characteristics, large language models can be very effectively and efficiently adapted to even new human languages unseen during pre-training [49]. Moreover, people have also treated large language models as interpreters in code interpretation [14] and generators in essay generation [22], and found that large language models can handle these tasks effectively. Motivated by this, in this work we are wondering, if we can also treat the large language model as an action recognizer in skeleton-based human action recognition? In general, large language models hold some characteristics that are useful for action recognition. Specifically, pre-trained over a tremendously large corpus [2, 56] that generally contains very rich human-centric behavior descriptions, large language models that hold large model architectures could naturally contain rich implicit knowledge [46, 50] of human behaviors, and hold strong capability in handling different inputs. Thus, it can be very promising if we treat the large", + "bbox": [ + 496, + 568, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 101, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "language model as a human action recognizer.", + "bbox": [ + 76, + 90, + 382, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, despite the success of large language models in handling various human languages, it can be challenging to treat them as action recognizers. This is because, while large language models typically take sentences in human languages as input instructions, the input signal of the action recognizer (i.e., the skeleton sequence) is not in a \"sentence format\" that is \"friendly\" to large language models. To handle this issue, given a large language model initialized with its pre-trained weights, one potential way is to adjust the weights of the full model to fit the non-linguistic input action signals. Recently, several works [25, 38] have been proposed to fine-tune large language models and adjust their pre-trained weights to handle non-language tasks. However, as shown in the previous study [64], the adjustment of the large language model's pre-trained weights can hurt its generalizability and lead to the loss of its pre-learned rich knowledge. This is clearly undesirable here, as we hope the large language model to be knowledgeable in order for it to be an accurate action recognizer.", + "bbox": [ + 75, + 108, + 468, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking this into account, in this work, we aim to harness the large language model as an action recognizer, and at the same time keep its pre-trained weights untouched to preserve its pre-learned rich knowledge. Specifically, we propose a novel action recognition framework named Large Language Model as an Action Recognizer (LLM-AR). As shown in Fig. 1, to perform action recognition, our LLM-AR framework first involves a linguistic projection process to project the input action signal into its \"sentence format\", that can be \"friendly\" and compatible to the large language model pre-trained over human sentences. Then the large language model takes in this \"sentence format\" of the action signal together with a human sentence as the instruction to predict the name of the action. In the rest of this work, for simplicity, we call the action signal in its \"sentence format\" an \"action sentence\".", + "bbox": [ + 75, + 397, + 470, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our framework, to project the action signals into \"action sentences\" (i.e., performing linguistic projection), we incorporate several designs. Basically, to perform such a process, we observe that, every sentence in a language generally can be regarded as a sequence consisting of discrete word tokens. Inspired by both this observation and previous works [75, 77], in LLM-AR, we first learn an action-based vector quantized variational autoencoder (VQ-VAE) model to project each action signal into a sequence of discrete tokens.", + "bbox": [ + 75, + 641, + 468, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, while we can basically represent each \"action sentence\" by a sequence of discrete tokens, it is not necessary that any discrete token sequence can represent an \"action sentence\" well. In fact, to enable a large language model to become an accurate action recognizer based on the \"action sentences\", besides being discrete sequences of tokens, ideally, the \"action sentences\" also need to ful-", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fill some requirements as discussed below. Firstly, since large language models are generally pre-trained over corpus consisting of sentences in human languages, to facilitate large language models in taking \"action sentences\" as instructions together with sentences in human languages, these \"action sentences\" should be \"like\" sentences in human languages so as to be more friendly to the large language model. On the other hand, \"action sentences\" should still maintain good representations of their original action signals in order for the large language model to accurately perform action recognition based on them.", + "bbox": [ + 496, + 90, + 890, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To better meet the above requirements in the linguistic projection process, we further incorporate our framework with two designs below. Firstly, to make the \"action sentences\" more \"like\" sentences in human languages, we get inspiration from previous linguistic and natural language processing works [27, 47, 48, 55, 82] which show that, languages as the communication tools of human beings often contain human inductive biases. Thus, we propose to incorporate our framework with a learning strategy to regularize the projected \"action sentences\" to follow human inductive biases as well. Secondly, to make the projected \"action sentence\" a good representation of its original action signal, motivated by the human skeleton's tree-like nature [63] and hyperbolic space's superior ability in representing tree structures [18], and inspired by previous works [33, 51], we further incorporate our action-based VQ-VAE model with a hyperbolic codebook.", + "bbox": [ + 496, + 257, + 892, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once we finish the learning of the linguistic projection process, we perform low-rank adaptation (LoRA) [23] on the large language model to let the model understand the projected \"action sentences\". Note that since the pre-trained weights of the large language model are untouched throughout the LoRA process, we can preserve the pre-learned rich knowledge in the large language model and thus utilize the large language model conveniently in our framework.", + "bbox": [ + 496, + 513, + 890, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our work are as follows. 1) We propose a novel action recognition framework named LLM-AR. To the best of our knowledge, this work is the first exploration on treating a large language model with its pretrained weights untouched as an action recognizer. 2) We introduce several designs in our framework to facilitate the linguistic projection process and produce the \"action sentences\". 3) LLM-AR achieves state-of-the-art performance on the evaluated benchmarks.", + "bbox": [ + 496, + 635, + 890, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 784, + 640, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Action Recognition. For tackling human action recognition, most of the existing methods can be roughly categorized into two groups: RGB-based methods [58, 60] and skeleton-based methods [5, 6, 53, 54, 70].", + "bbox": [ + 496, + 810, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As skeleton sequences can represent human behaviors in a succinct yet informative manner, skeleton-based ac", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion recognition has received lots of research attention [3-7, 12, 15, 17, 21, 28, 32, 39, 40, 53, 54, 59, 62, 65-68, 70, 73, 76, 80]. In the early days, different works have been proposed to use different network architectures to handle skeleton-based action recognition. Liu et al. [39] proposed to perform skeleton-based action recognition through a spatial-temporal LSTM architecture. Ke et al. [28] proposed to transform skeleton sequences into grey images and process these images through a CNN network. As time passed, GCN tends to be a popular network architecture [5, 7, 8, 34, 54, 65, 67, 70]. Yan et al. [70] proposed ST-GCN that made the first attempt in performing skeleton-based action recognition using a GCN architecture. After that, various different GCN-based methods have been further proposed, such as AS-GCN [34], 2s-AGCN [54], Shift-GCN [7], and CTR-GCN [5]. More recently, a number of works [15, 59, 76, 80] further proposed to perform skeleton-based action recognition via training a model with a large transformer architecture in an end-to-end manner. Some of such methods include 3Mformer [59] and UPS [15].", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from these approaches, in this work, from a novel perspective, we investigate how to instruct the large language model to perform skeleton-based action recognition while keeping its pre-trained weights untouched to preserve its pre-learned rich knowledge. To achieve this, we involve our framework with a novel linguistic projection process to project each input action signal into an \"action sentence\". Specifically, we incorporate several designs in the linguistic projection process to make the projected \"action sentences\" more \"like\" sentences of human languages so as to make them more friendly to large language models, and meanwhile keep the \"action sentence\" a good representation of the input action signal.", + "bbox": [ + 75, + 393, + 472, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Language Models. Recently, a variety of large language models have been proposed, such as GPT [2] and LLaMA [56]. Pre-trained over a tremendous number of word tokens, these large language models with large model architectures have been shown to contain very rich knowledge [46, 50]. Consequently, large language models have been extensively explored in various tasks [14, 16, 19, 22, 26, 77, 79], such as language translation [26], essay generation [22], and code interpretation [14]. In this work, we design a novel framework treating the large language model as a human action recognizer leveraging our designed linguistic projection process.", + "bbox": [ + 75, + 589, + 472, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 784, + 250, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a skeleton sequence as the input action signal, the goal of action recognition is to predict its corresponding action class. Recently, holding large model architectures and containing very rich knowledge, pre-trained large language models have been shown to be powerful in handling sentences of human languages, and thus have become useful", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tools in many natural language processing tasks. Inspired by this, in this work, we aim to leverage the large language model as an effective action recognizer. To achieve this, we propose a novel framework LLM-AR. Specifically, LLM-AR first performs a linguistic projection process to project the input action signal (i.e., the skeleton sequence) into an \"action sentence\". After that, LLM-AR passes the \"action sentence\" into the large language model to derive the corresponding human action. Below, we first describe how LLM-AR performs the linguistic projection process, and then introduce the overall training and testing scheme of LLM-AR.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Linguistic Projection", + "text_level": 1, + "bbox": [ + 498, + 268, + 697, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our framework, to enable the large language model to perform action recognition, we first perform a linguistic projection process to project each input action signal into an \"action sentence\". To learn to perform such a projection, with the observation that each sentence in human languages is essentially a sequence of discrete word tokens, motivate by previous works [75, 77], we first learn an action-based VQ-VAE model to project each input action signal into a sequence of discrete tokens. We then further (1) involve the above learning process with a human-inductive-biases-guided learning strategy to make the projected \"action sentences\" more \"like\" sentences in human languages, and (2) incorporate the action-based VQ-VAE model with a hyperbolic codebook to keep \"action sentences\" good representations of the action signals.", + "bbox": [ + 496, + 292, + 893, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action-based VQ-VAE Model. VQ-VAE models [51, 57, 74] have been popularly used in converting an image into a discrete token sequence. Inspired by this, in our framework, to convert the input action signal into a sequence of discrete tokens, we first learn an action-based VQ-VAE model.", + "bbox": [ + 496, + 518, + 893, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, similar to the architecture of previous VQ-VAE models [57, 75, 77], our action-based VQ-VAE model involves an encoder $E$ , a decoder $D$ , and a codebook $C$ consisting of $U$ learnable tokens (i.e., $C = \\{c_u\\}_{u=1}^U$ where $c_u \\in \\mathbb{R}^{d_u}$ ). We here keep $d_u$ to be the same as the dimension of the word token in the used large language model, and keep $U$ an even number. Among the aforementioned three components of the model, given an input action signal $s_{1:V}$ , where $V$ represents the action length, the encoder $E$ first encodes the action signal through 1D convolutions over the time dimension into a sequence of latent features $f_{1:W}$ where $f_w \\in \\mathbb{R}^{d_u}$ and $W$ represents the length of the latent feature sequence. After this, to discretize the latent features $f_{1:W}$ , a quantization operation is performed to replace each feature $f_w$ with its nearest token in the codebook as:", + "bbox": [ + 496, + 611, + 893, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {w} ^ {d} = \\underset {c _ {u} \\in C} {\\operatorname {a r g m i n}} \\left(\\operatorname {d i s t} \\left(f _ {w}, c _ {u}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 839, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_w^d$ is the discrete version of $f_w$ , and $dist(\\cdot, \\cdot)$ represents a distance function. Finally, the decoder aims to", + "bbox": [ + 496, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "recover the original action signal $s_{1:V}$ from the sequence of discrete latent features (tokens) $f_{1:W}^{d}$ . Through the above process of encoding, quantization, and decoding, we can project an action signal into a sequence of informative discrete tokens $f_{1:W}^{d}$ . The more detailed architecture of this action-based VQ-VAE model is provided in supplementary.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human-inductive-biases-guided Learning Strategy. Above we learn to project each input action signal into a sequence of discrete tokens. To enable the learned token sequences to be more friendly to the large language model, here we aim to make these sequences more \"like\" sentences in human languages. To achieve this, we get inspiration from massive existing studies [27, 47, 48, 55, 82] which show that, human languages as tools created for human communication naturally contain human inductive biases. Besides, it is also further shown by the previous study [47] that, inputs following human inductive biases are more friendly to large language models. Taking these into consideration, we aim to optimize the set of learned token sequences to also follow human inductive biases, so that these sequences can be more \"like\" human sentences and thus become more friendly to be used by large language models.", + "bbox": [ + 75, + 181, + 468, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Below, we first introduce the human inductive biases that are generally recognized as being present in human languages. Such biases include: (a) a human language naturally follows the Zipf's law [47, 48, 82], and (b) a human language is generally context-sensitive [27, 47, 55]. After introducing these human inductive biases, we then describe our proposed human-inductive-biases-guided learning strategy.", + "bbox": [ + 75, + 425, + 468, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for the Zipf's law in (a), during daily communications of human beings, there naturally exist some words that are used more commonly and some words more rarely. Zipf's law intuitively represents this imbalanced usage frequency of word tokens in human languages. Specifically, Zipf's law states that, in a human language, the $i$ -th most commonly used word has its usage frequency roughly proportional to:", + "bbox": [ + 75, + 547, + 468, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{(i + \\beta) ^ {\\alpha}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 240, + 655, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\approx 1$ and $\\beta \\approx 2.7$ [82]. The reason why Zipf's law consistently appears across different human languages is also theoretically analyzed by various works [9, 43] from different perspectives, such as from the evolution of human communications.", + "bbox": [ + 75, + 686, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With respect to the context-sensitivity of human languages in (b), intuitively, the context-sensitivity refers to the inductive bias that when people formulate their sentences, they often do not use each word token independently. Instead, their usage of different tokens in formulating a sentence is often correlated. Note that, while human languages are generally believed by linguists to be context-sensitive, how to explicitly represent context-sensitivity remains a difficult problem. Here, inspired by [47] in its way of repre", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "senting context-sensitivity, we represent this human inductive bias as follows. Specifically, given the codebook $C$ consisting of $U$ tokens where $U$ is an even number, during initializing $C$ , we first randomly split the codebook into two halves (i.e., $\\{1, \\dots, \\frac{U}{2}\\}$ and $\\{1 + \\frac{U}{2}, \\dots, U\\}$ ). Next, we regard each token $c_{u}$ in the first half (i.e., $u \\in \\{1, \\dots, \\frac{U}{2}\\}$ ) and the token $c_{u + \\frac{U}{2}}$ in the second half to be a pair of correlated tokens. As shown in [47], such a pairing mechanism is a simple yet very effective way of representing the context-sensitivity bias in human languages.", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the set of learned token sequences (i.e., \"action sentences\") $\\{f_{1:W}^d\\}$ to also follow the above biases like human languages, we then aim to (a) regularize the set of \"action sentences\" to follow Zipf's law, as well as (b) regularize each \"action sentence\" to be formulated using more correlated word tokens so that the formulated \"action sentences\" can better follow the context-sensitivity bias. To achieve this, we design a human-inductive-biases-guided learning strategy that consists of the following steps. (1) Given a batch of action signals $\\{s_{1:V}^b\\}_{b=1}^B$ with batch size $B$ , we first encode these signals through the encoder $E$ of the action-based VQ-VAE model to get their corresponding latent features $\\{f_{1:W}^b\\}_{b=1}^B$ . (2) After that, during discretizing the latent features using tokens $\\{c_u\\}_{u=1}^U$ in the codebook $C$ , we measure the token usage of each sequence of latent features $f_{1:W}^b$ in a differentiable way via the Gumbel Softmax trick [24] as:", + "bbox": [ + 496, + 244, + 892, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt ^ {b} = \\sum_ {w = 1} ^ {W} d _ {w} ^ {b}, \\text {w h e r e} d _ {w} ^ {b} = \\text {G u m b l e - S o f t m a x} (- \\operatorname {d i s t} \\left(f _ {w} ^ {b}, c _ {u}\\right)) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 502, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where Gumble_Softmax(·) is the Gumble Softmax trick, and $d_w^b$ is the one-hot vector with length $U$ as the output of the Gumble Softmax trick. Besides, $t^b$ is a vector with length $U$ , and the value of the $u$ -th element of $t^b$ represents the number of times token $c_u$ is used in discretizing $f_{1:W}^b$ and formulating its corresponding \"action sentence\". (3) Next, denoting $D_{Zipf}$ the Zipf distribution with $\\alpha = 1$ and $\\beta = 2.7$ , we regularize the set of \"action sentences\" to follow Zipf's law through $L_{Zipf}$ as:", + "bbox": [ + 496, + 556, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {Z i p f} = J S \\left(D _ {f r e q} \\| D _ {Z i p f}\\right), \\text {w h e r e} D _ {f r e q} = \\frac {\\operatorname {s o r t} \\left(\\sum_ {b = 1} ^ {B} t ^ {b}\\right)}{B \\times W} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 695, + 903, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $sort(\\cdot)$ is the sorting operation used to order tokens based on their usage frequency, $D_{freq}$ is the distribution representing the token usage frequency, and $JS(\\cdot \\| \\cdot)$ represents the JS divergence between two distributions. (4) At the same time, we encourage the \"action sentences\" to follow the context-sensitivity bias and use more correlated tokens via $L_{context}$ as:", + "bbox": [ + 496, + 741, + 890, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o n t e x t}} = 1 - \\frac {\\sum_ {b = 1} ^ {B} \\operatorname {C o r r} \\left(t ^ {b}\\right)}{B \\times W} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 848, + 890, + 881 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Corr(t^{b})$ leverages min pooling over every pair of", + "bbox": [ + 500, + 885, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3de685c669892722039fa7e1470f99a549983476514df4db0415f7f3c664734b.jpg", + "image_caption": [ + "Figure 2. Overview of the action-based VQ-VAE model with the hyperbolic codebook $C_H$ incorporated. Given a batch of input action signals $\\{s_{1:V}^b\\}_{b=1}^B$ , to optimize the action-based VQ-VAE model, $\\{s_{1:V}^b\\}_{b=1}^B$ are first fed to the encoder $E$ to get the corresponding latent features $\\{f_{1:W}^b\\}_{b=1}^B$ . Next, to leverage the hyperbolic codebook $C_H$ that can serve as a good representation of the tree-like human skeletons to perform quantization, $\\{f_{1:W}^b\\}_{b=1}^B$ are projected into the hyperbolic space via the process of E-to-H projection. After that, the quantization is performed in the hyperbolic space using $dist_{\\mathbb{B}}(\\widetilde{f}_w, c_u)$ defined in Eq. 9 as the distance function. Finally, after quantization, the discrete version of the latent features are passed back into the Euclidean space via the process of H-to-E projection to reconstruct the input action signals through the decoder $D$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 883, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "elements in $t^b$ (e.g., the pair of $u$ -th element and $(u + \\frac{U}{2})$ -th element) to calculate the number of correlated tokens used in discretizing $f_{1:W}^b$ (more details about this method to measure $Corr(t^b)$ are provided in supplementary). Note that via minimizing $L_{\\text{context}}$ , we encourage more correlated tokens to be used in formulating \"action sentences\". (5) Finally, we formulate the loss function $L_{\\text{human}}$ that we use to perform our human-inductive-biases-guided strategy as:", + "bbox": [ + 75, + 356, + 468, + 478 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 481, + 468, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Via incorporating the above learning strategy into the learning process of the action-based VQ-VAE model, we can regularize the \"action sentences\" to better follow the human inductive biases that are present in human languages and thus make them more \"like\" sentences in human languages.", + "bbox": [ + 75, + 500, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Above we inject human inductive biases into \"action sentences\". Here, we notice that, besides inductive biases that can be explicitly described, human languages can also present other implicit characteristics. Thus, to enable \"action sentences\" to be more friendly to large language models which are generally pre-trained over various human languages, we aim to further incorporate implicit characteristics into \"action sentences\" as well. Specifically, since large language models such as LLaMA typically use word tokens $\\{c_{LLM}\\}$ stored in its first layer to formulate sentences in human languages, we here further align (1) the tokens $\\{c_u\\}_{u = 1}^U$ in codebook $C$ that are used to formulate the \"action sentences\" with (2) the word tokens $\\{c_{LLM}\\}$ used in the large language model. To achieve such an alignment, we leverage Maximum Mean Discrepancy (MMD) [20] as an effective feature alignment technique to measure the discrepancy between $\\{c_u\\}_{u = 1}^U$ and $\\{c_{LLM}\\}$ . Specifically, the less MMD $(\\{c_u\\}_{u = 1}^U,\\{c_{LLM}\\})$ is, $\\{c_u\\}_{u = 1}^U$ and $\\{c_{LLM}\\}$ can be regarded as more aligned. We then incorporate this alignment procedure into our proposed strategy", + "bbox": [ + 75, + 598, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "via rewriting $L_{human}$ in Eq. 6 as:", + "bbox": [ + 500, + 357, + 723, + 372 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} + \\mathrm {M M D} \\left(\\left\\{c _ {u} \\right\\} _ {u = 1} ^ {U}, \\left\\{c _ {L L M} \\right\\}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 375, + 890, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By incorporating this alignment procedure into the previously mentioned learning strategy, we can then lead the formulated \"action sentences\" $\\{f_{1:W}^{d}\\}$ to be more friendly to be used by large language models.", + "bbox": [ + 496, + 406, + 890, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperbolic Codebook. Besides making the \"action sentences\" more friendly to the large language model, we also aim to keep them good representations of the original input action signals (i.e., the skeleton sequences). Motivated by the tree-like nature of human skeletons [63], we aim to make the word tokens in the \"action sentences\" to well represent such a structure. To achieve this, inspired by the superior capability of the hyperbolic space in embedding tree structures [18] and motivated by previous VQ-VAE works [33, 51], we here further incorporate our action-based VQ-VAE model with a hyperbolic codebook utilizing the Poincaré ball model [18, 45]. The Poincaré ball model is an isometric model that can represent the hyperbolic space. Formally, the $n$ -dimensional Poincaré ball model is defined as $(\\mathbb{B}_c^n,g_{\\mathbf{x}}^c)$ , where $\\mathbb{B}_c^n = \\{\\mathbf{x}\\in \\mathbb{R}^n:c\\| \\mathbf{x}\\| < 1\\}$ , $g_{\\mathbf{x}}^{c} = (\\gamma_{\\mathbf{x}}^{c})^{2}I_{n}$ is the Riemannian metric tensor, $\\gamma_{\\mathbf{x}}^{c} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}$ is the conformal factor, $I_{n}$ is the Euclidean metric tensor, and $c$ is the curvature.", + "bbox": [ + 496, + 470, + 892, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Denoting the hyperbolic codebook $C_H = \\{c_u\\}_{u=1}^U$ , where $c_u \\in \\mathbb{B}_c^{d_u}$ , we make the following three changes to incorporate $C_H$ into our action-based VQ-VAE model. We also illustrate where these changes take place in our action-based VQ-VAE model in Fig. 2. (1) Euclidean-to-Hyperbolic (E-to-H) Projection. Given a latent feature $f_w$ which is originally in the Euclidean space, to convert it into a discrete token using a hyperbolic codebook, we first need to project $f_w$ into the hyperbolic space. Specifically, following [18], we perform such a projection through the ex", + "bbox": [ + 496, + 748, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ponential map $\\exp_0^c (\\cdot)$ as:", + "bbox": [ + 76, + 90, + 251, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {f} _ {w} = \\exp_ {\\mathbf {0}} ^ {c} \\left(f _ {w}\\right) = \\tanh (\\sqrt {c} \\| f _ {w} \\|) \\frac {f _ {w}}{\\sqrt {c} \\| f _ {w} \\|} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 109, + 468, + 141 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\widetilde{f}_w$ is the projected feature of $f_w$ in the hyperbolic space. (2) Hyperbolic Distance Calculation. Given the projected feature $\\widetilde{f}_w$ , to perform quantization of $\\widetilde{f}_w$ based on the tokens in $C_H$ , we need a distance function defined in the hyperbolic space. Here, for simplicity, we use the popularly used geodesic/induced distance [18, 29]. Denote this distance as $\\text{dist}_{\\mathbb{B}}(\\cdot, \\cdot)$ . $\\text{dist}_{\\mathbb{B}}(\\widetilde{f}_w, c_u)$ between $\\widetilde{f}_w$ and $c_u$ can be defined as:", + "bbox": [ + 75, + 145, + 468, + 268 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} _ {\\mathbb {B}} \\left(\\widetilde {f} _ {w}, c _ {u}\\right) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\widetilde {f} _ {w} - c _ {u} \\| ^ {2}}{\\left(1 - \\| \\widetilde {f} _ {w} \\| ^ {2}\\right) \\left(1 - \\| c _ {u} \\| ^ {2}\\right)}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 271, + 467, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(3) Hyperbolic-to-Euclidean (H-to-E) Projection. After quantization, we need to project $\\widetilde{f}_w^d$ (the discrete version of $\\widetilde{f}_w$ ) back into the Euclidean space in order for it to be passed to the decoder $D$ . To achieve this, we follow [18] to leverage the logarithmic map $\\log_0^c (\\cdot)$ as:", + "bbox": [ + 76, + 321, + 468, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {w} ^ {d} = \\log_ {\\mathbf {0}} ^ {c} (\\widetilde {f} _ {w} ^ {d}) = \\operatorname {a r c t a n h} (\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|) \\frac {\\widetilde {f} _ {w} ^ {d}}{\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 401, + 468, + 440 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{w}^{d}$ represents the discrete version of the feature projected back in the Euclidean space. With the above changes made, we can seamlessly involve the hyperbolic codebook $C_H$ into our action-based VQ-VAE model, and make the projected \"action sentence\" a better representation of the input action signal.", + "bbox": [ + 75, + 441, + 468, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Overall Training and Testing", + "text_level": 1, + "bbox": [ + 76, + 542, + 334, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Above we describe how we perform the linguistic projection process in our framework to project each input action signal into an \"action sentence\". Below, we introduce the overall training and testing scheme of our framework.", + "bbox": [ + 75, + 566, + 468, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. The training process of our framework consists of the following two stages: (1) optimizing the action-based VQ-VAE model incorporated with the hyperbolic codebook $C_H$ to acquire the ability to project each action signal $s_{1:V}$ into its corresponding \"action sentence\" $\\widetilde{f}_{1:W}^d$ ; and (2) perform low-rank adaptation (LoRA) [23] on the large language model for it to better understand the projected \"action sentences\".", + "bbox": [ + 75, + 627, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the first stage, to optimize the action-based VQ-VAE model, given a batch of action signals, we first follow previous VQ-VAE works [57, 75, 77] to use the following loss functions, including the reconstruction loss, the embedding loss, and the commitment loss. Among them, the reconstruction loss $L_{re}$ is designed to regularize the difference between the input action signal and the action signal reconstructed from the decoder $D$ , the embedding loss $L_{embed}$ is designed to learn the tokens in the codebook, and the commitment loss $L_{commit}$ is designed to encourage each", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the encoded latent features to stay close to its discrete version. Formally speaking, denoting $\\{s_{1:V}^b\\}_{b=1}^B$ a batch of input action signals, $\\{\\hat{s}_{1:V}^b\\}_{b=1}^B$ the corresponding batch of reconstructed action signals, $\\{f_{1:W}^b\\}_{b=1}^B$ the latent features encoded from $\\{s_{1:V}^b\\}_{b=1}^B$ , and $\\{f_{1:W}^{d,b}\\}_{b=1}^B$ the discrete versions of these latent features, the above three loss functions can be written as:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e} = L _ {1} ^ {s m o o t h} (\\{s _ {1: V} ^ {b} \\} _ {b = 1} ^ {B}, \\{\\hat {s} _ {1: V} ^ {b} \\} _ {b = 1} ^ {B})\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 198, + 794, + 217 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e m b e d} = \\left\\| s g \\left(\\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B}\\right) - \\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B} \\right\\| ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 218, + 890, + 238 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c o m m i t} = \\left\\| \\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B} - s g \\left(\\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B}\\right) \\right\\| ^ {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 239, + 831, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{1}^{smooth}(\\cdot)$ represents the L1 smooth loss and $sg(\\cdot)$ is the stop-gradient operation that is used to prevent the gradient from flowing through its operand. In addition, to make the learned \"action sentences\" more like sentences in human languages, we also incorporate the above three loss functions with $L_{human}$ defined in Eq. 7. Overall, we can write the total loss $L_{total}$ for the first training stage as:", + "bbox": [ + 496, + 263, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = L _ {r e} + L _ {\\text {e m b e d}} + \\omega_ {1} L _ {\\text {c o m m i t}} + \\omega_ {2} L _ {\\text {h u m a n}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 372, + 890, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\omega_{1}$ and $\\omega_{2}$ are weighting hyperparameters.", + "bbox": [ + 498, + 392, + 823, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then perform LoRA [23] on the large language model in the second stage in order for it to better understand the \"action sentences\" while keeping the pre-trained weights of the model untouched. Specifically, for each training sample consisting of an input action signal and its corresponding ground-truth action, we perform the following steps. (1) We first project the action signal into an \"action sentence\" of discrete tokens through the action-based VQ-VAE model learned in the first stage. (2) We then instruct the large language model to act as an action recognizer via a simple instruction as: \"Given a sequence of action tokens [tokens], please predict the corresponding action.\", where [tokens] represent the word tokens of the \"action sentence\" derived in step (1). (3) Finally, during LoRA process, we pass the above instruction into the large language model and encourage the similarity between the tokens $t_p$ predicted by the large language model and the tokens $t_g$ representing the ground-truth action as:", + "bbox": [ + 496, + 407, + 890, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L o R A} = L _ {c e} \\left(t _ {p}, t _ {g}\\right) \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 699, + 890, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{ce}(\\cdot ,\\cdot)$ represents the cross-entropy loss.", + "bbox": [ + 498, + 718, + 821, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Testing. During testing, for each testing action signal, we first use the learned action-based VQ-VAE model to derive its corresponding \"action sentence\". After that, we use the same instruction as in step (2) above to instruct the large language model to predict the corresponding action.", + "bbox": [ + 496, + 734, + 890, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 828, + 630, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the efficacy of our framework, we conduct experiments on 4 datasets including NTU RGB+D, NTU RGB+D 120, Toyota Smarthome, and UAV-Human.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18400", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 90, + 179, + 104 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NTU RGB+D [52] is a large-scale dataset popularly used in human action recognition. It contains around 56k skeleton sequences from 60 activity classes. On this dataset, following [52], we evaluate our method under the Cross-Subject (X-Sub) and Cross-View (X-View) evaluation protocols.", + "bbox": [ + 75, + 113, + 468, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NTU RGB+D 120 [41] is an extension of the NTU RGB+D dataset. It consists of more than 114k skeleton sequences across 120 activity classes. Following [41], we evaluate our method on this dataset under the Cross-Subject (X-Sub) and Cross-Setup (X-Set) evaluation protocols.", + "bbox": [ + 75, + 189, + 468, + 263 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Toyota Smarthome [10] contains 16,115 video samples over 31 activity classes. On this dataset, we use the skeleton sequences pre-processed by [71] and we follow it to evaluate our method under the Cross-Subject (X-Sub) and two Cross-View (X-View1 & X-View2) evaluation protocols.", + "bbox": [ + 75, + 265, + 468, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UAV-Human [35] is a dataset that is captured by unmanned aerial vehicles (UAV). It contains more than 20k skeleton sequences over 155 activity classes, and it is collected from 119 distinct subjects. Following [35], we use 89 subjects for training and 30 subjects for testing.", + "bbox": [ + 75, + 340, + 468, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 422, + 294, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct our main experiments on Nvidia V100 GPU and we use LLaMA-13B [56] as the large language model. During the training process of the action-based VQ-VAE model, we optimize the model for 300,000 iterations using the AdamW [42] optimizer with an initial learning rate of 2e-4. We set the batch size $B$ to 256, the number of tokens $U$ to 512, and the curvature $c$ of the hyperbolic codebook to 1. Additionally, we set the dimension of each word token $d_{u}$ to the same size (5120) as the word token in LLaMA-13B. Moreover, we set $w_{1}$ to 0.02 following previous VQ-VAE works [44, 75] and set $w_{2}$ to 0.2. Besides, we set the length $(W)$ of each sequence of latent features to be a quarter of the length $(V)$ of its corresponding input action signal. During the LoRA process, we develop our code based on the Github repo [1], set the number of iterations to 75,000, and use the AdamW optimizer with an initial learning rate of 3e-3. Moreover, we set the batch size to 256, partitioned into micro-batches of 4. Besides, we set the two hyperparameters of the LoRA process, i.e., $r_{LoRA}$ and $\\alpha_{LoRA}$ , to 64 and 16 respectively.", + "bbox": [ + 75, + 446, + 468, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 756, + 441, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On NTU RGB+D and NTU RGB+D 120 datasets, following the experimental setting of recent works [12, 80], we only use the joint modality of human skeletons during our experiments. We report the results in Tab. 1. As shown, compared to existing skeleton-based action recognition methods, our method consistently achieves the best performance across all the evaluation protocols. This demonstrates the effectiveness of our method. Besides, we also", + "bbox": [ + 75, + 779, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ade71c206487a22e6285d9509a051538dd65fd1d7cda1e7cc7a53c6599f0b5c.jpg", + "table_caption": [ + "Table 1. Performance comparison on the NTU RGB+D and NTU RGB+D 120 datasets." + ], + "table_footnote": [], + "table_body": "
MethodNTU RGB+DNTU RGB+D 120
X-SubX-ViewX-SubX-Set
ST-GCN [70]85.792.482.184.5
Shift-GCN [7]87.895.180.983.2
InfoGCN [8]89.895.285.186.3
PoseC3D [11]93.796.585.989.7
FR-Head [78]90.395.385.587.3
Koopman [62]90.295.285.787.4
GAP [65]90.295.685.587.0
HD-GCN [31]90.695.785.787.3
STC-Net [30]91.096.286.288.0
DSTformer [80]93.097.2--
SkeleTR [12]94.897.787.888.3
Ours95.098.488.791.5
", + "bbox": [ + 504, + 119, + 890, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a1a709624a10ff930ff03fc74957dad4361ca248dee558d4768f44c4caf0329c.jpg", + "table_caption": [ + "Table 2. Performance comparison on the Toyota Smarthome dataset." + ], + "table_footnote": [], + "table_body": "
MethodX-SubX-View1X-View2
2S-AGCN [54]58.832.257.9
SSTA-PRS [71]62.122.854.0
UNIK [72]62.133.463.6
ML-STGNet [81]64.629.963.5
Ours67.036.166.6
", + "bbox": [ + 504, + 367, + 751, + 434 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5cc1c973196709f87a49b4c863fb32ac3ec9de61b859f2e2f48e6d2c401e1dc.jpg", + "table_caption": [ + "Table 3. Performance comparison on the UAV-Human dataset." + ], + "table_footnote": [], + "table_body": "
MethodX-Sub
ST-GCN [69]30.3
2S-AGCN [54]34.8
Shift-GCN [7]38.0
ACFL [61]44.2
Ours46.3
", + "bbox": [ + 763, + 368, + 890, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "report results on the Toyota Smarthome dataset in Tab. 2, and on the UAV-Human dataset in Tab. 3. As shown, our method consistently achieves the best performance on these datasets. This further shows the efficacy of our method.", + "bbox": [ + 498, + 460, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 530, + 663, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive ablation experiments on the X-Set protocol of the NTU RGB+D 120 dataset. More ablation studies such as experiments w.r.t. hyperparameters are in supplementary.", + "bbox": [ + 498, + 553, + 890, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the human-inductive-biases-guided", + "text_level": 1, + "bbox": [ + 500, + 614, + 700, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning strategy. In our framework, to lead the formulated \"action sentences\" to be more friendly to large language models, we design a human-inductive-biases", + "bbox": [ + 500, + 643, + 700, + 733 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec3254ed585e1af9cafdd231dec5e95c9b58c630249bec93c00debbc1d23c661.jpg", + "table_caption": [ + "Table 4. Evaluation on the human-inductive-biases-guided learning strategy." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
w/o biases87.6
w/o LZipf89.9
w/o Lcontext89.8
w/o MMD alignment90.3
LLM-AR91.5
", + "bbox": [ + 715, + 662, + 890, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "guided learning strategy consisting of three components (as shown in Eq. 7). To evaluate the efficacy of this strategy, we test four variants. In the first variant (w/o biases), we remove the whole strategy (i.e., all its three components $L_{Zipf}$ , $L_{context}$ , and MMD alignment) from the learning process. In the second variant (w/o $L_{Zipf}$ ), we still utilize the strategy but remove its $L_{Zipf}$ component. Moreover, in the third variant (w/o $L_{context}$ ), we remove the $L_{context}$ component from the strategy, whereas in the fourth variant (w/o MMD alignment), we remove the MMD alignment component. As shown in Tab. 4, compared to", + "bbox": [ + 498, + 734, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our framework, the performance of the first variant drops significantly. This shows the importance of formulating \"action sentences\" like sentences in human languages. Moreover, our framework also outperforms all the other three variants. This further shows the effectiveness of all the three components of our proposed learning strategy.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of discretizing latent features into \"action sentences\". In our framework, we discretize the encoded latent features to formulate \"action sentences\" consisting", + "text_level": 1, + "bbox": [ + 76, + 188, + 290, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation on discretizing latent features into \"action sentences\".", + "bbox": [ + 300, + 189, + 468, + 231 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2934664c39dc1f943f8f2cc9f7f8f959d756c6be0c918878563708214c580028.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy
w/o discretization83.4
with discretization91.5
", + "bbox": [ + 305, + 234, + 468, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of discrete word tokens (with discretization). To valid this design, we test a variant. In this variant (w/o discretization), instead of discretizing the encoded latent features into \"action sentences\", we directly pass these continuous features to the intermediate layers of the large language model. As shown in Tab. 5, our framework with discretization outperforms this variant. This shows the advantage of discretizing latent features into \"action sentences\", which are more \"like\" human sentences consisting of discrete word tokens, and thus are more friendly to the large language model pre-trained over human sentences.", + "bbox": [ + 75, + 279, + 472, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the hyperbolic codebook $C_H$ . In our framework, we incorporate our action-based VQ-VAE model with a hyperbolic codebook $C_H$ (with $C_H$ ). To validate the efficacy of $C_H$ , we test a variant $(\\mathbf{w} / \\mathbf{o} C_H)$ in which", + "text_level": 1, + "bbox": [ + 76, + 452, + 328, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Evaluation on the hyperbolic codebook $C_H$", + "bbox": [ + 339, + 454, + 468, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the codebook is set up in the Euclidean instead of hyperbolic space. As shown in Tab. 6, our framework involving $C_H$ performs better than this variant. This shows the efficacy of $C_H$ in the hyperbolic space, which can facilitate the \"action sentences\" in representing the tree-like-structured input action signals better.", + "bbox": [ + 75, + 542, + 470, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the LoRA process. In our framework, to make the large language model understand the \"action sentences\" while keeping its pretrained weights untouched to pre", + "text_level": 1, + "bbox": [ + 76, + 641, + 326, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Evaluation on the LoRA process.", + "bbox": [ + 336, + 643, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/edf569d8ac8d30900c59aa0d423088082da44b781c576a1c11397de42c684447.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAccuracy
All tuning79.6
LLM-AR91.5
", + "bbox": [ + 339, + 674, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "serve its rich pre-learned knowledge, we tune the large language model through a LoRA process. To validate this scheme, here we also test a variant (all tuning) on A100 GPU. In this variant, during tuning, after initializing the large language model with its pre-trained weights, all the parameters of the model will undergo gradient updates. As shown in Tab. 7, our framework achieves much better performance than this variant. This shows the superiority of our framework in choosing to perform tuning via LoRA, which enables the large language model's pre-trained weights to be untouched and maintains its pre-learned rich knowledge.", + "bbox": [ + 75, + 717, + 472, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on unseen activity classes. In the main experiments, following [31, 52, 70], we evaluate our framework on activity classes that have been seen during training. Here, inspired by that large language", + "bbox": [ + 496, + 90, + 746, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f762b878cddc8cd85bf5f9bbff886701fb53c1bd1a1849f255029a72dc0171e.jpg", + "table_caption": [ + "Table 8. Evaluation on unseen activity classes." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
All tuning37.7
LLM-AR62.4
", + "bbox": [ + 763, + 138, + 890, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models naturally could contain rich knowledge beyond the training activity classes used in our experiments, we are curious, assuming we have a list of testing activity classes unseen during training, can we also use our framework to perform action recognition on these classes? To answer this question, we first build a new evaluation protocol for unseen activity classes based on the NTU RGB+D 120 dataset. Under this new protocol, during each time of evaluation, we randomly select 3 classes to form the unseen class list (i.e., the list of testing classes), and use the remaining classes as the classes seen during training (i.e., the training classes). Besides, we instruct the large language model as: \"Given a sequence of action tokens [tokens], please predict the corresponding action from [list],\" where [tokens] represent the word tokens of the \"action sentence\" and [list] represents the unseen class list. We then perform the above evaluation for five times and report the average performance. As shown in Tab. 8, even testing on activity classes unseen during training, our framework can still achieve a relatively good performance, while the afore-defined all tuning variant performs much worse. This can be analyzed as, large language models could contain rich pre-learned knowledge w.r.t. the list of unseen classes. Thus, our framework that maintains such rich knowledge can still perform promising recognition on these unseen classes, while the all tuning variant that can lose amount of pre-learned knowledge of the large language model would yield a much worse performance. This further shows the advantage of our framework in maintaining the pre-learned rich knowledge of the large language model.", + "bbox": [ + 496, + 181, + 892, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 648, + 619, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel action recognition framework LLM-AR. In LLM-AR, we treat the large language model as an action recognizer, and instruct the large language model to perform action recognition using its contained rich knowledge. Specifically, to lead the input action signals (i.e., the skeleton sequences) to be more friendly to the large language model, we first propose a linguistic projection process to project each action signal into an \"action sentence\". Moreover, we also introduce several designs to further facilitate this process. Our framework consistently achieves SOTA performance across different benchmarks.", + "bbox": [ + 496, + 672, + 890, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014).", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18402", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lit-llama. https://github.com/Lightning-AI/ lit-llama.7", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3", + "[3] Dongqi Cai, Yangyuxuan Kang, Anbang Yao, and Yurong Chen. Ske2grid: Skeleton-to-grid representation learning for action recognition. 2023. 1, 3", + "[4] Jinghong Chen, Chong Zhao, Qicong Wang, and Hongying Meng. Hmanet: Hyperbolic manifold aware network for skeleton-based action recognition. IEEE Transactions on Cognitive and Developmental Systems, 2022.", + "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13359-13368, 2021. 1, 2, 3", + "[6] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 536-553. Springer, 2020. 1, 2", + "[7] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 183-192, 2020. 3, 7", + "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 3, 7", + "[9] Bernat Corominas-Murtra, Jordi Fortune, and Ricard V Solé. Emergence of zipf's law in the evolution of communication. Physical Review E, 83(3):036115, 2011. 4", + "[10] Srijan Das, Rui Dai, Michal Koperski, Luca Minciullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome: Real-world activities of daily living. In Proceedings of the IEEE/CVF international conference on computer vision, pages 833-842, 2019. 7", + "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 7", + "[12] Haodong Duan, Mingze Xu, Bing Shuai, Davide Modolo, Zhuowen Tu, Joseph Tighe, and Alessandro Bergamo. Skeletr: Towards skeleton-based action recognition in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13634-13644, 2023. 1, 3, 7", + "[13] Abassin Sourou Fangbemi, Bin Liu, Neng Hai Yu, and Yanxiang Zhang. Efficient human action recognition interface for" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "augmented and virtual reality applications based on binary descriptor. In Augmented Reality, Virtual Reality, and Computer Graphics: 5th International Conference, AVR 2018, Otranto, Italy, June 24-27, 2018, Proceedings, Part I 5, pages 252-260. Springer, 2018. 1", + "[14] Faisal Firdous, Saimul Bashir, Syed Zoofa Rufai, and Sanjeev Kumar. Openai chatgpt as a logical interpreter of code. In 2023 2nd International Conference on Edge Computing and Applications (ICECAA), pages 1192-1197. IEEE, 2023. 1, 3", + "[15] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13030, 2023. 1, 3", + "[16] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 3", + "[17] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[18] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 2, 5, 6", + "[19] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[20] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 5", + "[21] Ryo Hachiuma, Fumiaki Sato, and Taiki Sekii. Unified keypoint-based action recognition framework via structured keypoint pooling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22962-22971, 2023. 3", + "[22] Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. Ai, write an essay for me: A large-scale comparison of human-written versus chatgpt-generated essays. arXiv preprint arXiv:2304.14276, 2023. 1, 3", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6", + "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4", + "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2", + "[26] Wenxiang Jiao, Wenxuan Wang, JT Huang, Xing Wang, and ZP Tu. Is chatgpt a good translator? yes with gpt-4 as the engine. arXiv preprint arXiv:2301.08745, 2023. 3", + "[27] Aravind K Joshi, K Vijay Shanker, and David Weir. The convergence of mildly context-sensitive grammar formalisms. Technical Reports (CIS), page 539, 1990. 2, 4" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 3", + "[29] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 6", + "[30] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10255-10264, 2023. 7", + "[31] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10444-10453, 2023. 7, 8", + "[32] Zikang Leng, Hyeokhyen Kwon, and Thomas Plötz. On the benefit of generative foundation models for human activity recognition. arXiv preprint arXiv:2310.12085, 2023. 3", + "[33] Lei Li, Tingting Liu, Chengyu Wang, Minghui Qiu, Cen Chen, Ming Gao, and Aoying Zhou. Resizing codebook of vector quantization without retraining. Multimedia Systems, pages 1-14, 2023. 2, 5", + "[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 3", + "[35] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16266-16275, 2021. 7", + "[36] Weiyao Lin, Ming-Ting Sun, Radha Poovendran, and Zhengyou Zhang. Activity recognition using a combination of category components and local models for video surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 18(8):1128-1139, 2008. 1", + "[37] Hong Liu, Qinqin He, and Mengyuan Liu. Human action recognition using adaptive hierarchical depth motion maps and gabor filter. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1432-1436. IEEE, 2017. 1", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2", + "[39] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 3", + "[40] Jun Liu, Gang Wang, Ping Hu, Ling-Yu Duan, and Alex C Kot. Global context-aware attention LSTM networks for 3d" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1647-1656, 2017. 3", + "[41] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 7", + "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7", + "[43] Benoit Mandelbrot. Structure formelle des textes et communication: Deux études par. Word, 10(1):1-27, 1954. 4", + "[44] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 7", + "[45] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 5", + "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 3", + "[47] Isabel Papadimitriou and Dan Jurafsky. Pretrain on just structure: Understanding linguistic inductive biases using transfer learning. arXiv preprint arXiv:2304.13060, 2023. 2, 4", + "[48] Steven T Piantadosi. Zipf's word frequency law in natural language: A critical review and future directions. *Psychonomic bulletin & review*, 21:1112–1130, 2014. 2, 4", + "[49] Yiwei Qin, Graham Neubig, and Pengfei Liu. Searching for effective multilingual fine-tuning methods: A case study in summarization, 2022. 1", + "[50] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[51] Acquaviva Sam. Hyperbolic vq-vaes. 2, 3, 5", + "[52] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7, 8", + "[53] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7912-7921, 2019. 1, 2, 3", + "[54] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 3, 7", + "[55] Stuart M Shieber. Evidence against the context-freeness of natural language. In The Formal complexity of natural language, pages 320-334. Springer, 1985. 2, 4" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18404", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 7", + "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 6", + "[58] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 2", + "[59] Lei Wang and Piotr Koniusz. 3mformer: Multi-order multi-mode transformer for skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5620-5631, 2023. 1, 3", + "[60] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 2", + "[61] Xuanhan Wang, Yan Dai, Lianli Gao, and Jingkuan Song. Skeleton-based action recognition via adaptive cross-form learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1670-1678, 2022. 7", + "[62] Xinghan Wang, Xin Xu, and Yadong Mu. Neural koopman pooling: Control-inspired temporal dynamics encoding for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10597-10607, 2023. 1, 3, 7", + "[63] Shenghua Wei, Yonghong Song, and Yuanlin Zhang. Human skeleton tree recurrent neural network with joint relative motion feature for skeleton based action recognition. In 2017 IEEE international conference on image processing (ICIP), pages 91-95. IEEE, 2017. 2, 5", + "[64] Rongxiang Weng, Wen Sen Cheng, and Min Zhang. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. In *Findings of the Association for Computational Linguistics: ACL 2023*, pages 4747-4755, 2023. 2", + "[65] Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, and Lei Zhang. Generative action description prompts for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10276-10285, 2023. 3, 7", + "[66] Wentian Xin, Qiguang Miao, Yi Liu, Ruyi Liu, Chi-Man Pun, and Cheng Shi. Skeleton mixformer: Multivariate topology representation for skeleton-based action recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2211-2220, 2023.", + "[67] Haojun Xu, Yan Gao, Zheng Hui, Jie Li, and Xinbo Gao. Language knowledge-assisted representation learning for skeleton-based action recognition. arXiv preprint arXiv:2305.12398, 2023. 3", + "[68] Kailin Xu, Fanfan Ye, Qiaoyong Zhong, and Di Xie. Topology-aware convolutional neural network for efficient skeleton-based action recognition. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AAAI Conference on Artificial Intelligence, pages 2866-2874, 2022. 3", + "[69] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2272-2281, 2017. 7", + "[70] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7, 8", + "[71] Di Yang, Rui Dai, Yaohui Wang, Rupayan Mallick, Luca Minciullo, Gianpiero Francesca, and Francois Bremond. Selective spatio-temporal aggregation based pose refinement system: Towards understanding human activities in real-world videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2363-2372, 2021. 7", + "[72] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. arXiv preprint arXiv:2107.08580, 2021. 1, 7", + "[73] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3427-3435, 2023. 3", + "[74] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18467-18476, 2023. 3", + "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2, 3, 6, 7", + "[76] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3", + "[77] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 2, 3, 6", + "[78] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7", + "[79] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023. 3", + "[80] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18405", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 7", + "[81] Yisheng Zhu, Hui Shuai, Guangcan Liu, and Qingshan Liu. Multilevel spatial-temporal excited graph network for skeleton-based action recognition. IEEE Transactions on Image Processing, 32:496-508, 2022. 7", + "[82] G. K. Zipf. The psycho-biology of language, 1935. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 191 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "18406", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_model.json b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d8284040630bdff53caab2aa60c063852a452822 --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_model.json @@ -0,0 +1,2622 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.302, + 0.131, + 0.669, + 0.154 + ], + "angle": 0, + "content": "LLMs are Good Action Recognizers" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.18, + 0.83, + 0.22 + ], + "angle": 0, + "content": "Haoxuan \\(\\mathrm{Qu}^{1}\\) Yujun Cai\\(^{2}\\) Jun Liu\\(^{1\\dagger}\\) \n\\(^{1}\\)Singapore University of Technology and Design \\(^{2}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.22, + 0.8, + 0.234 + ], + "angle": 0, + "content": "haoxuan_qu@mymail.sutd.edu.sg, yujun001@e.ntu.edu.sg, jun.liu@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.474, + 0.621 + ], + "angle": 0, + "content": "Skeleton-based action recognition has attracted lots of research attention. Recently, to build an accurate skeleton-based action recognizer, a variety of works have been proposed. Among them, some works use large model architectures as backbones of their recognizers to boost the skeleton data representation capability, while some other works pretrain their recognizers on external data to enrich the knowledge. In this work, we observe that large language models which have been extensively used in various natural language processing tasks generally hold both large model architectures and rich implicit knowledge. Motivated by this, we propose a novel LLM-AR framework, in which we investigate treating the Large Language Model as an Action Recognizer. In our framework, we propose a linguistic projection process to project each input action signal (i.e., each skeleton sequence) into its \"sentence format\" (i.e., an \"action sentence\"). Moreover, we also incorporate our framework with several designs to further facilitate this linguistic projection process. Extensive experiments demonstrate the efficacy of our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.678, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Human action recognition aims to categorize the actions performed by humans into a pre-defined list of classes. It is relevant to a variety of applications, such as human-computer interaction [37], intelligent surveillance [36], and virtual reality [13]. In the past few years, skeleton-based action recognition [3, 5, 6, 12, 53, 54, 62, 70] has received a lot of research attention with the notice that skeleton is a succinct yet informative representation of human behaviors. Yet, despite the considerable progress, skeleton-based action recognition still remains a challenging task [78], and to build a more accurate skeleton-based action recognizer, various recent works have been proposed from different perspectives. Among them, some recent works [15, 59] pro" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.27, + 0.895, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.359, + 0.895, + 0.442 + ], + "angle": 0, + "content": "Figure 1. Overview of our proposed LLM-AR framework. In our framework, given an input action signal, we first perform a linguistic projection process to acquire the corresponding \"action sentence\". We then perform action recognition via the large language model with its pre-trained weights untouched to keep its pre-learned rich knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.566 + ], + "angle": 0, + "content": "posed to utilize large model architectures (such as transformers) as the backbone of their action recognizers to achieve stronger representation capability and capture more subtle differences among different actions. On the other hand, some other works [72, 80] proposed to pre-train their action recognizers on external data in order for their action recognizers to handle this task with richer knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Recently, large language models such as GPT [2] and LLaMA [56] have become quite popular and have been extensively applied in handling various human languages. For example, having been pre-trained over various human languages and learned common language-related characteristics, large language models can be very effectively and efficiently adapted to even new human languages unseen during pre-training [49]. Moreover, people have also treated large language models as interpreters in code interpretation [14] and generators in essay generation [22], and found that large language models can handle these tasks effectively. Motivated by this, in this work we are wondering, if we can also treat the large language model as an action recognizer in skeleton-based human action recognition? In general, large language models hold some characteristics that are useful for action recognition. Specifically, pre-trained over a tremendously large corpus [2, 56] that generally contains very rich human-centric behavior descriptions, large language models that hold large model architectures could naturally contain rich implicit knowledge [46, 50] of human behaviors, and hold strong capability in handling different inputs. Thus, it can be very promising if we treat the large" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.383, + 0.106 + ], + "angle": 0, + "content": "language model as a human action recognizer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.397 + ], + "angle": 0, + "content": "However, despite the success of large language models in handling various human languages, it can be challenging to treat them as action recognizers. This is because, while large language models typically take sentences in human languages as input instructions, the input signal of the action recognizer (i.e., the skeleton sequence) is not in a \"sentence format\" that is \"friendly\" to large language models. To handle this issue, given a large language model initialized with its pre-trained weights, one potential way is to adjust the weights of the full model to fit the non-linguistic input action signals. Recently, several works [25, 38] have been proposed to fine-tune large language models and adjust their pre-trained weights to handle non-language tasks. However, as shown in the previous study [64], the adjustment of the large language model's pre-trained weights can hurt its generalizability and lead to the loss of its pre-learned rich knowledge. This is clearly undesirable here, as we hope the large language model to be knowledgeable in order for it to be an accurate action recognizer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.639 + ], + "angle": 0, + "content": "Taking this into account, in this work, we aim to harness the large language model as an action recognizer, and at the same time keep its pre-trained weights untouched to preserve its pre-learned rich knowledge. Specifically, we propose a novel action recognition framework named Large Language Model as an Action Recognizer (LLM-AR). As shown in Fig. 1, to perform action recognition, our LLM-AR framework first involves a linguistic projection process to project the input action signal into its \"sentence format\", that can be \"friendly\" and compatible to the large language model pre-trained over human sentences. Then the large language model takes in this \"sentence format\" of the action signal together with a human sentence as the instruction to predict the name of the action. In the rest of this work, for simplicity, we call the action signal in its \"sentence format\" an \"action sentence\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.793 + ], + "angle": 0, + "content": "In our framework, to project the action signals into \"action sentences\" (i.e., performing linguistic projection), we incorporate several designs. Basically, to perform such a process, we observe that, every sentence in a language generally can be regarded as a sequence consisting of discrete word tokens. Inspired by both this observation and previous works [75, 77], in LLM-AR, we first learn an action-based vector quantized variational autoencoder (VQ-VAE) model to project each action signal into a sequence of discrete tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.901 + ], + "angle": 0, + "content": "However, while we can basically represent each \"action sentence\" by a sequence of discrete tokens, it is not necessary that any discrete token sequence can represent an \"action sentence\" well. In fact, to enable a large language model to become an accurate action recognizer based on the \"action sentences\", besides being discrete sequences of tokens, ideally, the \"action sentences\" also need to ful-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.257 + ], + "angle": 0, + "content": "fill some requirements as discussed below. Firstly, since large language models are generally pre-trained over corpus consisting of sentences in human languages, to facilitate large language models in taking \"action sentences\" as instructions together with sentences in human languages, these \"action sentences\" should be \"like\" sentences in human languages so as to be more friendly to the large language model. On the other hand, \"action sentences\" should still maintain good representations of their original action signals in order for the large language model to accurately perform action recognition based on them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.893, + 0.514 + ], + "angle": 0, + "content": "To better meet the above requirements in the linguistic projection process, we further incorporate our framework with two designs below. Firstly, to make the \"action sentences\" more \"like\" sentences in human languages, we get inspiration from previous linguistic and natural language processing works [27, 47, 48, 55, 82] which show that, languages as the communication tools of human beings often contain human inductive biases. Thus, we propose to incorporate our framework with a learning strategy to regularize the projected \"action sentences\" to follow human inductive biases as well. Secondly, to make the projected \"action sentence\" a good representation of its original action signal, motivated by the human skeleton's tree-like nature [63] and hyperbolic space's superior ability in representing tree structures [18], and inspired by previous works [33, 51], we further incorporate our action-based VQ-VAE model with a hyperbolic codebook." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Once we finish the learning of the linguistic projection process, we perform low-rank adaptation (LoRA) [23] on the large language model to let the model understand the projected \"action sentences\". Note that since the pre-trained weights of the large language model are untouched throughout the LoRA process, we can preserve the pre-learned rich knowledge in the large language model and thus utilize the large language model conveniently in our framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.77 + ], + "angle": 0, + "content": "The contributions of our work are as follows. 1) We propose a novel action recognition framework named LLM-AR. To the best of our knowledge, this work is the first exploration on treating a large language model with its pretrained weights untouched as an action recognizer. 2) We introduce several designs in our framework to facilitate the linguistic projection process and produce the \"action sentences\". 3) LLM-AR achieves state-of-the-art performance on the evaluated benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.642, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Human Action Recognition. For tackling human action recognition, most of the existing methods can be roughly categorized into two groups: RGB-based methods [58, 60] and skeleton-based methods [5, 6, 53, 54, 70]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As skeleton sequences can represent human behaviors in a succinct yet informative manner, skeleton-based ac" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18396" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.394 + ], + "angle": 0, + "content": "tion recognition has received lots of research attention [3-7, 12, 15, 17, 21, 28, 32, 39, 40, 53, 54, 59, 62, 65-68, 70, 73, 76, 80]. In the early days, different works have been proposed to use different network architectures to handle skeleton-based action recognition. Liu et al. [39] proposed to perform skeleton-based action recognition through a spatial-temporal LSTM architecture. Ke et al. [28] proposed to transform skeleton sequences into grey images and process these images through a CNN network. As time passed, GCN tends to be a popular network architecture [5, 7, 8, 34, 54, 65, 67, 70]. Yan et al. [70] proposed ST-GCN that made the first attempt in performing skeleton-based action recognition using a GCN architecture. After that, various different GCN-based methods have been further proposed, such as AS-GCN [34], 2s-AGCN [54], Shift-GCN [7], and CTR-GCN [5]. More recently, a number of works [15, 59, 76, 80] further proposed to perform skeleton-based action recognition via training a model with a large transformer architecture in an end-to-end manner. Some of such methods include 3Mformer [59] and UPS [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.473, + 0.59 + ], + "angle": 0, + "content": "Different from these approaches, in this work, from a novel perspective, we investigate how to instruct the large language model to perform skeleton-based action recognition while keeping its pre-trained weights untouched to preserve its pre-learned rich knowledge. To achieve this, we involve our framework with a novel linguistic projection process to project each input action signal into an \"action sentence\". Specifically, we incorporate several designs in the linguistic projection process to make the projected \"action sentences\" more \"like\" sentences of human languages so as to make them more friendly to large language models, and meanwhile keep the \"action sentence\" a good representation of the input action signal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Large Language Models. Recently, a variety of large language models have been proposed, such as GPT [2] and LLaMA [56]. Pre-trained over a tremendous number of word tokens, these large language models with large model architectures have been shown to contain very rich knowledge [46, 50]. Consequently, large language models have been extensively explored in various tasks [14, 16, 19, 22, 26, 77, 79], such as language translation [26], essay generation [22], and code interpretation [14]. In this work, we design a novel framework treating the large language model as a human action recognizer leveraging our designed linguistic projection process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.251, + 0.803 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Given a skeleton sequence as the input action signal, the goal of action recognition is to predict its corresponding action class. Recently, holding large model architectures and containing very rich knowledge, pre-trained large language models have been shown to be powerful in handling sentences of human languages, and thus have become useful" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.26 + ], + "angle": 0, + "content": "tools in many natural language processing tasks. Inspired by this, in this work, we aim to leverage the large language model as an effective action recognizer. To achieve this, we propose a novel framework LLM-AR. Specifically, LLM-AR first performs a linguistic projection process to project the input action signal (i.e., the skeleton sequence) into an \"action sentence\". After that, LLM-AR passes the \"action sentence\" into the large language model to derive the corresponding human action. Below, we first describe how LLM-AR performs the linguistic projection process, and then introduce the overall training and testing scheme of LLM-AR." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.269, + 0.698, + 0.286 + ], + "angle": 0, + "content": "3.1. Linguistic Projection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.895, + 0.519 + ], + "angle": 0, + "content": "In our framework, to enable the large language model to perform action recognition, we first perform a linguistic projection process to project each input action signal into an \"action sentence\". To learn to perform such a projection, with the observation that each sentence in human languages is essentially a sequence of discrete word tokens, motivate by previous works [75, 77], we first learn an action-based VQ-VAE model to project each input action signal into a sequence of discrete tokens. We then further (1) involve the above learning process with a human-inductive-biases-guided learning strategy to make the projected \"action sentences\" more \"like\" sentences in human languages, and (2) incorporate the action-based VQ-VAE model with a hyperbolic codebook to keep \"action sentences\" good representations of the action signals." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.894, + 0.61 + ], + "angle": 0, + "content": "Action-based VQ-VAE Model. VQ-VAE models [51, 57, 74] have been popularly used in converting an image into a discrete token sequence. Inspired by this, in our framework, to convert the input action signal into a sequence of discrete tokens, we first learn an action-based VQ-VAE model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.895, + 0.839 + ], + "angle": 0, + "content": "Specifically, similar to the architecture of previous VQ-VAE models [57, 75, 77], our action-based VQ-VAE model involves an encoder \\( E \\), a decoder \\( D \\), and a codebook \\( C \\) consisting of \\( U \\) learnable tokens (i.e., \\( C = \\{c_u\\}_{u=1}^U \\) where \\( c_u \\in \\mathbb{R}^{d_u} \\)). We here keep \\( d_u \\) to be the same as the dimension of the word token in the used large language model, and keep \\( U \\) an even number. Among the aforementioned three components of the model, given an input action signal \\( s_{1:V} \\), where \\( V \\) represents the action length, the encoder \\( E \\) first encodes the action signal through 1D convolutions over the time dimension into a sequence of latent features \\( f_{1:W} \\) where \\( f_w \\in \\mathbb{R}^{d_u} \\) and \\( W \\) represents the length of the latent feature sequence. After this, to discretize the latent features \\( f_{1:W} \\), a quantization operation is performed to replace each feature \\( f_w \\) with its nearest token in the codebook as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.84, + 0.892, + 0.867 + ], + "angle": 0, + "content": "\\[\nf _ {w} ^ {d} = \\underset {c _ {u} \\in C} {\\operatorname {a r g m i n}} \\left(\\operatorname {d i s t} \\left(f _ {w}, c _ {u}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\( f_w^d \\) is the discrete version of \\( f_w \\), and \\( dist(\\cdot, \\cdot) \\) represents a distance function. Finally, the decoder aims to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18397" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.183 + ], + "angle": 0, + "content": "recover the original action signal \\( s_{1:V} \\) from the sequence of discrete latent features (tokens) \\( f_{1:W}^{d} \\). Through the above process of encoding, quantization, and decoding, we can project an action signal into a sequence of informative discrete tokens \\( f_{1:W}^{d} \\). The more detailed architecture of this action-based VQ-VAE model is provided in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Human-inductive-biases-guided Learning Strategy. Above we learn to project each input action signal into a sequence of discrete tokens. To enable the learned token sequences to be more friendly to the large language model, here we aim to make these sequences more \"like\" sentences in human languages. To achieve this, we get inspiration from massive existing studies [27, 47, 48, 55, 82] which show that, human languages as tools created for human communication naturally contain human inductive biases. Besides, it is also further shown by the previous study [47] that, inputs following human inductive biases are more friendly to large language models. Taking these into consideration, we aim to optimize the set of learned token sequences to also follow human inductive biases, so that these sequences can be more \"like\" human sentences and thus become more friendly to be used by large language models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.469, + 0.547 + ], + "angle": 0, + "content": "Below, we first introduce the human inductive biases that are generally recognized as being present in human languages. Such biases include: (a) a human language naturally follows the Zipf's law [47, 48, 82], and (b) a human language is generally context-sensitive [27, 47, 55]. After introducing these human inductive biases, we then describe our proposed human-inductive-biases-guided learning strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.655 + ], + "angle": 0, + "content": "As for the Zipf's law in (a), during daily communications of human beings, there naturally exist some words that are used more commonly and some words more rarely. Zipf's law intuitively represents this imbalanced usage frequency of word tokens in human languages. Specifically, Zipf's law states that, in a human language, the \\(i\\)-th most commonly used word has its usage frequency roughly proportional to:" + }, + { + "type": "equation", + "bbox": [ + 0.241, + 0.656, + 0.469, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{(i + \\beta) ^ {\\alpha}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.763 + ], + "angle": 0, + "content": "where \\(\\alpha \\approx 1\\) and \\(\\beta \\approx 2.7\\) [82]. The reason why Zipf's law consistently appears across different human languages is also theoretically analyzed by various works [9, 43] from different perspectives, such as from the evolution of human communications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "With respect to the context-sensitivity of human languages in (b), intuitively, the context-sensitivity refers to the inductive bias that when people formulate their sentences, they often do not use each word token independently. Instead, their usage of different tokens in formulating a sentence is often correlated. Note that, while human languages are generally believed by linguists to be context-sensitive, how to explicitly represent context-sensitivity remains a difficult problem. Here, inspired by [47] in its way of repre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "senting context-sensitivity, we represent this human inductive bias as follows. Specifically, given the codebook \\( C \\) consisting of \\( U \\) tokens where \\( U \\) is an even number, during initializing \\( C \\), we first randomly split the codebook into two halves (i.e., \\( \\{1, \\dots, \\frac{U}{2}\\} \\) and \\( \\{1 + \\frac{U}{2}, \\dots, U\\} \\)). Next, we regard each token \\( c_{u} \\) in the first half (i.e., \\( u \\in \\{1, \\dots, \\frac{U}{2}\\} \\)) and the token \\( c_{u + \\frac{U}{2}} \\) in the second half to be a pair of correlated tokens. As shown in [47], such a pairing mechanism is a simple yet very effective way of representing the context-sensitivity bias in human languages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.245, + 0.893, + 0.501 + ], + "angle": 0, + "content": "To optimize the set of learned token sequences (i.e., \"action sentences\") \\(\\{f_{1:W}^d\\}\\) to also follow the above biases like human languages, we then aim to (a) regularize the set of \"action sentences\" to follow Zipf's law, as well as (b) regularize each \"action sentence\" to be formulated using more correlated word tokens so that the formulated \"action sentences\" can better follow the context-sensitivity bias. To achieve this, we design a human-inductive-biases-guided learning strategy that consists of the following steps. (1) Given a batch of action signals \\(\\{s_{1:V}^b\\}_{b=1}^B\\) with batch size \\(B\\), we first encode these signals through the encoder \\(E\\) of the action-based VQ-VAE model to get their corresponding latent features \\(\\{f_{1:W}^b\\}_{b=1}^B\\). (2) After that, during discretizing the latent features using tokens \\(\\{c_u\\}_{u=1}^U\\) in the codebook \\(C\\), we measure the token usage of each sequence of latent features \\(f_{1:W}^b\\) in a differentiable way via the Gumbel Softmax trick [24] as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.503, + 0.908, + 0.557 + ], + "angle": 0, + "content": "\\[\nt ^ {b} = \\sum_ {w = 1} ^ {W} d _ {w} ^ {b}, \\text {w h e r e} d _ {w} ^ {b} = \\text {G u m b l e - S o f t m a x} (- \\operatorname {d i s t} \\left(f _ {w} ^ {b}, c _ {u}\\right)) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.693 + ], + "angle": 0, + "content": "where Gumble_Softmax(·) is the Gumble Softmax trick, and \\( d_w^b \\) is the one-hot vector with length \\( U \\) as the output of the Gumble Softmax trick. Besides, \\( t^b \\) is a vector with length \\( U \\), and the value of the \\( u \\)-th element of \\( t^b \\) represents the number of times token \\( c_u \\) is used in discretizing \\( f_{1:W}^b \\) and formulating its corresponding \"action sentence\". (3) Next, denoting \\( D_{Zipf} \\) the Zipf distribution with \\( \\alpha = 1 \\) and \\( \\beta = 2.7 \\), we regularize the set of \"action sentences\" to follow Zipf's law through \\( L_{Zipf} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.696, + 0.905, + 0.741 + ], + "angle": 0, + "content": "\\[\nL _ {Z i p f} = J S \\left(D _ {f r e q} \\| D _ {Z i p f}\\right), \\text {w h e r e} D _ {f r e q} = \\frac {\\operatorname {s o r t} \\left(\\sum_ {b = 1} ^ {B} t ^ {b}\\right)}{B \\times W} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.742, + 0.892, + 0.847 + ], + "angle": 0, + "content": "where \\( sort(\\cdot) \\) is the sorting operation used to order tokens based on their usage frequency, \\( D_{freq} \\) is the distribution representing the token usage frequency, and \\( JS(\\cdot \\| \\cdot) \\) represents the JS divergence between two distributions. (4) At the same time, we encourage the \"action sentences\" to follow the context-sensitivity bias and use more correlated tokens via \\( L_{context} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.849, + 0.892, + 0.882 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o n t e x t}} = 1 - \\frac {\\sum_ {b = 1} ^ {B} \\operatorname {C o r r} \\left(t ^ {b}\\right)}{B \\times W} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\(Corr(t^{b})\\) leverages min pooling over every pair of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18398" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.885, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.893, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Overview of the action-based VQ-VAE model with the hyperbolic codebook \\( C_H \\) incorporated. Given a batch of input action signals \\( \\{s_{1:V}^b\\}_{b=1}^B \\), to optimize the action-based VQ-VAE model, \\( \\{s_{1:V}^b\\}_{b=1}^B \\) are first fed to the encoder \\( E \\) to get the corresponding latent features \\( \\{f_{1:W}^b\\}_{b=1}^B \\). Next, to leverage the hyperbolic codebook \\( C_H \\) that can serve as a good representation of the tree-like human skeletons to perform quantization, \\( \\{f_{1:W}^b\\}_{b=1}^B \\) are projected into the hyperbolic space via the process of E-to-H projection. After that, the quantization is performed in the hyperbolic space using \\( dist_{\\mathbb{B}}(\\widetilde{f}_w, c_u) \\) defined in Eq. 9 as the distance function. Finally, after quantization, the discrete version of the latent features are passed back into the Euclidean space via the process of H-to-E projection to reconstruct the input action signals through the decoder \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.479 + ], + "angle": 0, + "content": "elements in \\( t^b \\) (e.g., the pair of \\( u \\)-th element and \\( (u + \\frac{U}{2}) \\)-th element) to calculate the number of correlated tokens used in discretizing \\( f_{1:W}^b \\) (more details about this method to measure \\( Corr(t^b) \\) are provided in supplementary). Note that via minimizing \\( L_{\\text{context}} \\), we encourage more correlated tokens to be used in formulating \"action sentences\". (5) Finally, we formulate the loss function \\( L_{\\text{human}} \\) that we use to perform our human-inductive-biases-guided strategy as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.482, + 0.469, + 0.499 + ], + "angle": 0, + "content": "\\[\nL _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.469, + 0.577 + ], + "angle": 0, + "content": "Via incorporating the above learning strategy into the learning process of the action-based VQ-VAE model, we can regularize the \"action sentences\" to better follow the human inductive biases that are present in human languages and thus make them more \"like\" sentences in human languages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Above we inject human inductive biases into \"action sentences\". Here, we notice that, besides inductive biases that can be explicitly described, human languages can also present other implicit characteristics. Thus, to enable \"action sentences\" to be more friendly to large language models which are generally pre-trained over various human languages, we aim to further incorporate implicit characteristics into \"action sentences\" as well. Specifically, since large language models such as LLaMA typically use word tokens \\(\\{c_{LLM}\\}\\) stored in its first layer to formulate sentences in human languages, we here further align (1) the tokens \\(\\{c_u\\}_{u = 1}^U\\) in codebook \\(C\\) that are used to formulate the \"action sentences\" with (2) the word tokens \\(\\{c_{LLM}\\}\\) used in the large language model. To achieve such an alignment, we leverage Maximum Mean Discrepancy (MMD) [20] as an effective feature alignment technique to measure the discrepancy between \\(\\{c_u\\}_{u = 1}^U\\) and \\(\\{c_{LLM}\\}\\). Specifically, the less MMD \\((\\{c_u\\}_{u = 1}^U,\\{c_{LLM}\\})\\) is, \\(\\{c_u\\}_{u = 1}^U\\) and \\(\\{c_{LLM}\\}\\) can be regarded as more aligned. We then incorporate this alignment procedure into our proposed strategy" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.358, + 0.725, + 0.373 + ], + "angle": 0, + "content": "via rewriting \\(L_{human}\\) in Eq. 6 as:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.375, + 0.892, + 0.407 + ], + "angle": 0, + "content": "\\[\nL _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} + \\mathrm {M M D} \\left(\\left\\{c _ {u} \\right\\} _ {u = 1} ^ {U}, \\left\\{c _ {L L M} \\right\\}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.468 + ], + "angle": 0, + "content": "By incorporating this alignment procedure into the previously mentioned learning strategy, we can then lead the formulated \"action sentences\" \\(\\{f_{1:W}^{d}\\}\\) to be more friendly to be used by large language models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.893, + 0.744 + ], + "angle": 0, + "content": "Hyperbolic Codebook. Besides making the \"action sentences\" more friendly to the large language model, we also aim to keep them good representations of the original input action signals (i.e., the skeleton sequences). Motivated by the tree-like nature of human skeletons [63], we aim to make the word tokens in the \"action sentences\" to well represent such a structure. To achieve this, inspired by the superior capability of the hyperbolic space in embedding tree structures [18] and motivated by previous VQ-VAE works [33, 51], we here further incorporate our action-based VQ-VAE model with a hyperbolic codebook utilizing the Poincaré ball model [18, 45]. The Poincaré ball model is an isometric model that can represent the hyperbolic space. Formally, the \\(n\\)-dimensional Poincaré ball model is defined as \\((\\mathbb{B}_c^n,g_{\\mathbf{x}}^c)\\), where \\(\\mathbb{B}_c^n = \\{\\mathbf{x}\\in \\mathbb{R}^n:c\\| \\mathbf{x}\\| < 1\\}\\), \\(g_{\\mathbf{x}}^{c} = (\\gamma_{\\mathbf{x}}^{c})^{2}I_{n}\\) is the Riemannian metric tensor, \\(\\gamma_{\\mathbf{x}}^{c} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}\\) is the conformal factor, \\(I_{n}\\) is the Euclidean metric tensor, and \\(c\\) is the curvature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Denoting the hyperbolic codebook \\( C_H = \\{c_u\\}_{u=1}^U \\), where \\( c_u \\in \\mathbb{B}_c^{d_u} \\), we make the following three changes to incorporate \\( C_H \\) into our action-based VQ-VAE model. We also illustrate where these changes take place in our action-based VQ-VAE model in Fig. 2. (1) Euclidean-to-Hyperbolic (E-to-H) Projection. Given a latent feature \\( f_w \\) which is originally in the Euclidean space, to convert it into a discrete token using a hyperbolic codebook, we first need to project \\( f_w \\) into the hyperbolic space. Specifically, following [18], we perform such a projection through the ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18399" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.252, + 0.107 + ], + "angle": 0, + "content": "ponential map \\(\\exp_0^c (\\cdot)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.11, + 0.47, + 0.142 + ], + "angle": 0, + "content": "\\[\n\\widetilde {f} _ {w} = \\exp_ {\\mathbf {0}} ^ {c} \\left(f _ {w}\\right) = \\tanh (\\sqrt {c} \\| f _ {w} \\|) \\frac {f _ {w}}{\\sqrt {c} \\| f _ {w} \\|} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.146, + 0.47, + 0.269 + ], + "angle": 0, + "content": "where \\(\\widetilde{f}_w\\) is the projected feature of \\(f_w\\) in the hyperbolic space. (2) Hyperbolic Distance Calculation. Given the projected feature \\(\\widetilde{f}_w\\), to perform quantization of \\(\\widetilde{f}_w\\) based on the tokens in \\(C_H\\), we need a distance function defined in the hyperbolic space. Here, for simplicity, we use the popularly used geodesic/induced distance [18, 29]. Denote this distance as \\(\\text{dist}_{\\mathbb{B}}(\\cdot, \\cdot)\\). \\(\\text{dist}_{\\mathbb{B}}(\\widetilde{f}_w, c_u)\\) between \\(\\widetilde{f}_w\\) and \\(c_u\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.272, + 0.468, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} _ {\\mathbb {B}} \\left(\\widetilde {f} _ {w}, c _ {u}\\right) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\widetilde {f} _ {w} - c _ {u} \\| ^ {2}}{\\left(1 - \\| \\widetilde {f} _ {w} \\| ^ {2}\\right) \\left(1 - \\| c _ {u} \\| ^ {2}\\right)}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.469, + 0.4 + ], + "angle": 0, + "content": "(3) Hyperbolic-to-Euclidean (H-to-E) Projection. After quantization, we need to project \\(\\widetilde{f}_w^d\\) (the discrete version of \\(\\widetilde{f}_w\\)) back into the Euclidean space in order for it to be passed to the decoder \\(D\\). To achieve this, we follow [18] to leverage the logarithmic map \\(\\log_0^c (\\cdot)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.402, + 0.469, + 0.441 + ], + "angle": 0, + "content": "\\[\nf _ {w} ^ {d} = \\log_ {\\mathbf {0}} ^ {c} (\\widetilde {f} _ {w} ^ {d}) = \\operatorname {a r c t a n h} (\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|) \\frac {\\widetilde {f} _ {w} ^ {d}}{\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.469, + 0.534 + ], + "angle": 0, + "content": "where \\( f_{w}^{d} \\) represents the discrete version of the feature projected back in the Euclidean space. With the above changes made, we can seamlessly involve the hyperbolic codebook \\( C_H \\) into our action-based VQ-VAE model, and make the projected \"action sentence\" a better representation of the input action signal." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.336, + 0.56 + ], + "angle": 0, + "content": "3.2. Overall Training and Testing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Above we describe how we perform the linguistic projection process in our framework to project each input action signal into an \"action sentence\". Below, we introduce the overall training and testing scheme of our framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.749 + ], + "angle": 0, + "content": "Training. The training process of our framework consists of the following two stages: (1) optimizing the action-based VQ-VAE model incorporated with the hyperbolic codebook \\( C_H \\) to acquire the ability to project each action signal \\( s_{1:V} \\) into its corresponding \"action sentence\" \\( \\widetilde{f}_{1:W}^d \\); and (2) perform low-rank adaptation (LoRA) [23] on the large language model for it to better understand the projected \"action sentences\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "At the first stage, to optimize the action-based VQ-VAE model, given a batch of action signals, we first follow previous VQ-VAE works [57, 75, 77] to use the following loss functions, including the reconstruction loss, the embedding loss, and the commitment loss. Among them, the reconstruction loss \\( L_{re} \\) is designed to regularize the difference between the input action signal and the action signal reconstructed from the decoder \\( D \\), the embedding loss \\( L_{embed} \\) is designed to learn the tokens in the codebook, and the commitment loss \\( L_{commit} \\) is designed to encourage each" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "of the encoded latent features to stay close to its discrete version. Formally speaking, denoting \\(\\{s_{1:V}^b\\}_{b=1}^B\\) a batch of input action signals, \\(\\{\\hat{s}_{1:V}^b\\}_{b=1}^B\\) the corresponding batch of reconstructed action signals, \\(\\{f_{1:W}^b\\}_{b=1}^B\\) the latent features encoded from \\(\\{s_{1:V}^b\\}_{b=1}^B\\), and \\(\\{f_{1:W}^{d,b}\\}_{b=1}^B\\) the discrete versions of these latent features, the above three loss functions can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.199, + 0.795, + 0.218 + ], + "angle": 0, + "content": "\\[\nL _ {r e} = L _ {1} ^ {s m o o t h} (\\{s _ {1: V} ^ {b} \\} _ {b = 1} ^ {B}, \\{\\hat {s} _ {1: V} ^ {b} \\} _ {b = 1} ^ {B})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.219, + 0.892, + 0.239 + ], + "angle": 0, + "content": "\\[\nL _ {e m b e d} = \\left\\| s g \\left(\\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B}\\right) - \\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B} \\right\\| ^ {2} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.241, + 0.833, + 0.261 + ], + "angle": 0, + "content": "\\[\nL _ {c o m m i t} = \\left\\| \\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B} - s g \\left(\\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B}\\right) \\right\\| ^ {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.892, + 0.37 + ], + "angle": 0, + "content": "where \\( L_{1}^{smooth}(\\cdot) \\) represents the L1 smooth loss and \\( sg(\\cdot) \\) is the stop-gradient operation that is used to prevent the gradient from flowing through its operand. In addition, to make the learned \"action sentences\" more like sentences in human languages, we also incorporate the above three loss functions with \\( L_{human} \\) defined in Eq. 7. Overall, we can write the total loss \\( L_{total} \\) for the first training stage as:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.373, + 0.891, + 0.39 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = L _ {r e} + L _ {\\text {e m b e d}} + \\omega_ {1} L _ {\\text {c o m m i t}} + \\omega_ {2} L _ {\\text {h u m a n}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.393, + 0.825, + 0.408 + ], + "angle": 0, + "content": "where \\(\\omega_{1}\\) and \\(\\omega_{2}\\) are weighting hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.695 + ], + "angle": 0, + "content": "We then perform LoRA [23] on the large language model in the second stage in order for it to better understand the \"action sentences\" while keeping the pre-trained weights of the model untouched. Specifically, for each training sample consisting of an input action signal and its corresponding ground-truth action, we perform the following steps. (1) We first project the action signal into an \"action sentence\" of discrete tokens through the action-based VQ-VAE model learned in the first stage. (2) We then instruct the large language model to act as an action recognizer via a simple instruction as: \"Given a sequence of action tokens [tokens], please predict the corresponding action.\", where [tokens] represent the word tokens of the \"action sentence\" derived in step (1). (3) Finally, during LoRA process, we pass the above instruction into the large language model and encourage the similarity between the tokens \\( t_p \\) predicted by the large language model and the tokens \\( t_g \\) representing the ground-truth action as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.7, + 0.891, + 0.716 + ], + "angle": 0, + "content": "\\[\nL _ {L o R A} = L _ {c e} \\left(t _ {p}, t _ {g}\\right) \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.719, + 0.822, + 0.734 + ], + "angle": 0, + "content": "where \\(L_{ce}(\\cdot ,\\cdot)\\) represents the cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.811 + ], + "angle": 0, + "content": "Testing. During testing, for each testing action signal, we first use the learned action-based VQ-VAE model to derive its corresponding \"action sentence\". After that, we use the same instruction as in step (2) above to instruct the large language model to predict the corresponding action." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.632, + 0.846 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the efficacy of our framework, we conduct experiments on 4 datasets including NTU RGB+D, NTU RGB+D 120, Toyota Smarthome, and UAV-Human." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18400" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.181, + 0.105 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.189 + ], + "angle": 0, + "content": "NTU RGB+D [52] is a large-scale dataset popularly used in human action recognition. It contains around 56k skeleton sequences from 60 activity classes. On this dataset, following [52], we evaluate our method under the Cross-Subject (X-Sub) and Cross-View (X-View) evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.19, + 0.47, + 0.265 + ], + "angle": 0, + "content": "NTU RGB+D 120 [41] is an extension of the NTU RGB+D dataset. It consists of more than 114k skeleton sequences across 120 activity classes. Following [41], we evaluate our method on this dataset under the Cross-Subject (X-Sub) and Cross-Setup (X-Set) evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.469, + 0.34 + ], + "angle": 0, + "content": "Toyota Smarthome [10] contains 16,115 video samples over 31 activity classes. On this dataset, we use the skeleton sequences pre-processed by [71] and we follow it to evaluate our method under the Cross-Subject (X-Sub) and two Cross-View (X-View1 & X-View2) evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.416 + ], + "angle": 0, + "content": "UAV-Human [35] is a dataset that is captured by unmanned aerial vehicles (UAV). It contains more than 20k skeleton sequences over 155 activity classes, and it is collected from 119 distinct subjects. Following [35], we use 89 subjects for training and 30 subjects for testing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.295, + 0.439 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.75 + ], + "angle": 0, + "content": "We conduct our main experiments on Nvidia V100 GPU and we use LLaMA-13B [56] as the large language model. During the training process of the action-based VQ-VAE model, we optimize the model for 300,000 iterations using the AdamW [42] optimizer with an initial learning rate of 2e-4. We set the batch size \\( B \\) to 256, the number of tokens \\( U \\) to 512, and the curvature \\( c \\) of the hyperbolic codebook to 1. Additionally, we set the dimension of each word token \\( d_{u} \\) to the same size (5120) as the word token in LLaMA-13B. Moreover, we set \\( w_{1} \\) to 0.02 following previous VQ-VAE works [44, 75] and set \\( w_{2} \\) to 0.2. Besides, we set the length \\( (W) \\) of each sequence of latent features to be a quarter of the length \\( (V) \\) of its corresponding input action signal. During the LoRA process, we develop our code based on the Github repo [1], set the number of iterations to 75,000, and use the AdamW optimizer with an initial learning rate of 3e-3. Moreover, we set the batch size to 256, partitioned into micro-batches of 4. Besides, we set the two hyperparameters of the LoRA process, i.e., \\( r_{LoRA} \\) and \\( \\alpha_{LoRA} \\), to 64 and 16 respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.442, + 0.773 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.9 + ], + "angle": 0, + "content": "On NTU RGB+D and NTU RGB+D 120 datasets, following the experimental setting of recent works [12, 80], we only use the joint modality of human skeletons during our experiments. We report the results in Tab. 1. As shown, compared to existing skeleton-based action recognition methods, our method consistently achieves the best performance across all the evaluation protocols. This demonstrates the effectiveness of our method. Besides, we also" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the NTU RGB+D and NTU RGB+D 120 datasets." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.121, + 0.892, + 0.307 + ], + "angle": 0, + "content": "
MethodNTU RGB+DNTU RGB+D 120
X-SubX-ViewX-SubX-Set
ST-GCN [70]85.792.482.184.5
Shift-GCN [7]87.895.180.983.2
InfoGCN [8]89.895.285.186.3
PoseC3D [11]93.796.585.989.7
FR-Head [78]90.395.385.587.3
Koopman [62]90.295.285.787.4
GAP [65]90.295.685.587.0
HD-GCN [31]90.695.785.787.3
STC-Net [30]91.096.286.288.0
DSTformer [80]93.097.2--
SkeleTR [12]94.897.787.888.3
Ours95.098.488.791.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.322, + 0.752, + 0.35 + ], + "angle": 0, + "content": "Table 2. Performance comparison on the Toyota Smarthome dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.368, + 0.753, + 0.435 + ], + "angle": 0, + "content": "
MethodX-SubX-View1X-View2
2S-AGCN [54]58.832.257.9
SSTA-PRS [71]62.122.854.0
UNIK [72]62.133.463.6
ML-STGNet [81]64.629.963.5
Ours67.036.166.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.76, + 0.322, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Table 3. Performance comparison on the UAV-Human dataset." + }, + { + "type": "table", + "bbox": [ + 0.764, + 0.369, + 0.892, + 0.435 + ], + "angle": 0, + "content": "
MethodX-Sub
ST-GCN [69]30.3
2S-AGCN [54]34.8
Shift-GCN [7]38.0
ACFL [61]44.2
Ours46.3
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.461, + 0.892, + 0.522 + ], + "angle": 0, + "content": "report results on the Toyota Smarthome dataset in Tab. 2, and on the UAV-Human dataset in Tab. 3. As shown, our method consistently achieves the best performance on these datasets. This further shows the efficacy of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.665, + 0.545 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.892, + 0.613 + ], + "angle": 0, + "content": "We conduct extensive ablation experiments on the X-Set protocol of the NTU RGB+D 120 dataset. More ablation studies such as experiments w.r.t. hyperparameters are in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.615, + 0.701, + 0.644 + ], + "angle": 0, + "content": "Impact of the human-inductive-biases-guided" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.645, + 0.701, + 0.734 + ], + "angle": 0, + "content": "learning strategy. In our framework, to lead the formulated \"action sentences\" to be more friendly to large language models, we design a human-inductive-biases" + }, + { + "type": "table_caption", + "bbox": [ + 0.712, + 0.618, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Table 4. Evaluation on the human-inductive-biases-guided learning strategy." + }, + { + "type": "table", + "bbox": [ + 0.717, + 0.664, + 0.892, + 0.732 + ], + "angle": 0, + "content": "
MethodAccuracy
w/o biases87.6
w/o LZipf89.9
w/o Lcontext89.8
w/o MMD alignment90.3
LLM-AR91.5
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "guided learning strategy consisting of three components (as shown in Eq. 7). To evaluate the efficacy of this strategy, we test four variants. In the first variant (w/o biases), we remove the whole strategy (i.e., all its three components \\( L_{Zipf} \\), \\( L_{context} \\), and MMD alignment) from the learning process. In the second variant (w/o \\( L_{Zipf} \\)), we still utilize the strategy but remove its \\( L_{Zipf} \\) component. Moreover, in the third variant (w/o \\( L_{context} \\)), we remove the \\( L_{context} \\) component from the strategy, whereas in the fourth variant (w/o MMD alignment), we remove the MMD alignment component. As shown in Tab. 4, compared to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18401" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "our framework, the performance of the first variant drops significantly. This shows the importance of formulating \"action sentences\" like sentences in human languages. Moreover, our framework also outperforms all the other three variants. This further shows the effectiveness of all the three components of our proposed learning strategy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.189, + 0.291, + 0.28 + ], + "angle": 0, + "content": "Impact of discretizing latent features into \"action sentences\". In our framework, we discretize the encoded latent features to formulate \"action sentences\" consisting" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.19, + 0.47, + 0.232 + ], + "angle": 0, + "content": "Table 5. Evaluation on discretizing latent features into \"action sentences\"." + }, + { + "type": "table", + "bbox": [ + 0.306, + 0.236, + 0.47, + 0.273 + ], + "angle": 0, + "content": "
MethodAccuracy
w/o discretization83.4
with discretization91.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.473, + 0.448 + ], + "angle": 0, + "content": "of discrete word tokens (with discretization). To valid this design, we test a variant. In this variant (w/o discretization), instead of discretizing the encoded latent features into \"action sentences\", we directly pass these continuous features to the intermediate layers of the large language model. As shown in Tab. 5, our framework with discretization outperforms this variant. This shows the advantage of discretizing latent features into \"action sentences\", which are more \"like\" human sentences consisting of discrete word tokens, and thus are more friendly to the large language model pre-trained over human sentences." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.33, + 0.544 + ], + "angle": 0, + "content": "Impact of the hyperbolic codebook \\(C_H\\). In our framework, we incorporate our action-based VQ-VAE model with a hyperbolic codebook \\(C_H\\) (with \\(C_H\\)). To validate the efficacy of \\(C_H\\), we test a variant \\((\\mathbf{w} / \\mathbf{o} C_H)\\) in which" + }, + { + "type": "text", + "bbox": [ + 0.34, + 0.455, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Table 6. Evaluation on the hyperbolic codebook \\(C_H\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.472, + 0.636 + ], + "angle": 0, + "content": "the codebook is set up in the Euclidean instead of hyperbolic space. As shown in Tab. 6, our framework involving \\( C_H \\) performs better than this variant. This shows the efficacy of \\( C_H \\) in the hyperbolic space, which can facilitate the \"action sentences\" in representing the tree-like-structured input action signals better." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.327, + 0.717 + ], + "angle": 0, + "content": "Impact of the LoRA process. In our framework, to make the large language model understand the \"action sentences\" while keeping its pretrained weights untouched to pre" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.645, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Table 7. Evaluation on the LoRA process." + }, + { + "type": "table", + "bbox": [ + 0.34, + 0.675, + 0.47, + 0.713 + ], + "angle": 0, + "content": "
MethodAccuracy
All tuning79.6
LLM-AR91.5
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.473, + 0.9 + ], + "angle": 0, + "content": "serve its rich pre-learned knowledge, we tune the large language model through a LoRA process. To validate this scheme, here we also test a variant (all tuning) on A100 GPU. In this variant, during tuning, after initializing the large language model with its pre-trained weights, all the parameters of the model will undergo gradient updates. As shown in Tab. 7, our framework achieves much better performance than this variant. This shows the superiority of our framework in choosing to perform tuning via LoRA, which enables the large language model's pre-trained weights to be untouched and maintains its pre-learned rich knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.748, + 0.183 + ], + "angle": 0, + "content": "Evaluation on unseen activity classes. In the main experiments, following [31, 52, 70], we evaluate our framework on activity classes that have been seen during training. Here, inspired by that large language" + }, + { + "type": "table_caption", + "bbox": [ + 0.758, + 0.094, + 0.892, + 0.136 + ], + "angle": 0, + "content": "Table 8. Evaluation on unseen activity classes." + }, + { + "type": "table", + "bbox": [ + 0.764, + 0.14, + 0.892, + 0.178 + ], + "angle": 0, + "content": "
MethodAccuracy
All tuning37.7
LLM-AR62.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.893, + 0.637 + ], + "angle": 0, + "content": "models naturally could contain rich knowledge beyond the training activity classes used in our experiments, we are curious, assuming we have a list of testing activity classes unseen during training, can we also use our framework to perform action recognition on these classes? To answer this question, we first build a new evaluation protocol for unseen activity classes based on the NTU RGB+D 120 dataset. Under this new protocol, during each time of evaluation, we randomly select 3 classes to form the unseen class list (i.e., the list of testing classes), and use the remaining classes as the classes seen during training (i.e., the training classes). Besides, we instruct the large language model as: \"Given a sequence of action tokens [tokens], please predict the corresponding action from [list],\" where [tokens] represent the word tokens of the \"action sentence\" and [list] represents the unseen class list. We then perform the above evaluation for five times and report the average performance. As shown in Tab. 8, even testing on activity classes unseen during training, our framework can still achieve a relatively good performance, while the afore-defined all tuning variant performs much worse. This can be analyzed as, large language models could contain rich pre-learned knowledge w.r.t. the list of unseen classes. Thus, our framework that maintains such rich knowledge can still perform promising recognition on these unseen classes, while the all tuning variant that can lose amount of pre-learned knowledge of the large language model would yield a much worse performance. This further shows the advantage of our framework in maintaining the pre-learned rich knowledge of the large language model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.62, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.84 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel action recognition framework LLM-AR. In LLM-AR, we treat the large language model as an action recognizer, and instruct the large language model to perform action recognition using its contained rich knowledge. Specifically, to lead the input action signals (i.e., the skeleton sequences) to be more friendly to the large language model, we first propose a linguistic projection process to project each action signal into an \"action sentence\". Moreover, we also introduce several designs to further facilitate this process. Our framework consistently achieves SOTA performance across different benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18402" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] Lit-llama. https://github.com/Lightning-AI/ lit-llama.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[3] Dongqi Cai, Yangyuxuan Kang, Anbang Yao, and Yurong Chen. Ske2grid: Skeleton-to-grid representation learning for action recognition. 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.311 + ], + "angle": 0, + "content": "[4] Jinghong Chen, Chong Zhao, Qicong Wang, and Hongying Meng. Hmanet: Hyperbolic manifold aware network for skeleton-based action recognition. IEEE Transactions on Cognitive and Developmental Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.468, + 0.382 + ], + "angle": 0, + "content": "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13359-13368, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.468, + 0.465 + ], + "angle": 0, + "content": "[6] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 536-553. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[7] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 183-192, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[9] Bernat Corominas-Murtra, Jordi Fortune, and Ricard V Solé. Emergence of zipf's law in the evolution of communication. Physical Review E, 83(3):036115, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[10] Srijan Das, Rui Dai, Michal Koperski, Luca Minciullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome: Real-world activities of daily living. In Proceedings of the IEEE/CVF international conference on computer vision, pages 833-842, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.789 + ], + "angle": 0, + "content": "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[12] Haodong Duan, Mingze Xu, Bing Shuai, Davide Modolo, Zhuowen Tu, Joseph Tighe, and Alessandro Bergamo. Skeletr: Towards skeleton-based action recognition in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13634-13644, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Abassin Sourou Fangbemi, Bin Liu, Neng Hai Yu, and Yanxiang Zhang. Efficient human action recognition interface for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "augmented and virtual reality applications based on binary descriptor. In Augmented Reality, Virtual Reality, and Computer Graphics: 5th International Conference, AVR 2018, Otranto, Italy, June 24-27, 2018, Proceedings, Part I 5, pages 252-260. Springer, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Faisal Firdous, Saimul Bashir, Syed Zoofa Rufai, and Sanjeev Kumar. Openai chatgpt as a logical interpreter of code. In 2023 2nd International Conference on Edge Computing and Applications (ICECAA), pages 1192-1197. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[15] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13030, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[16] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[17] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[18] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[20] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Ryo Hachiuma, Fumiaki Sato, and Taiki Sekii. Unified keypoint-based action recognition framework via structured keypoint pooling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22962-22971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[22] Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. Ai, write an essay for me: A large-scale comparison of human-written versus chatgpt-generated essays. arXiv preprint arXiv:2304.14276, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.678, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[26] Wenxiang Jiao, Wenxuan Wang, JT Huang, Xing Wang, and ZP Tu. Is chatgpt a good translator? yes with gpt-4 as the engine. arXiv preprint arXiv:2301.08745, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Aravind K Joshi, K Vijay Shanker, and David Weir. The convergence of mildly context-sensitive grammar formalisms. Technical Reports (CIS), page 539, 1990. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18403" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[28] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[29] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[30] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10255-10264, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[31] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10444-10453, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[32] Zikang Leng, Hyeokhyen Kwon, and Thomas Plötz. On the benefit of generative foundation models for human activity recognition. arXiv preprint arXiv:2310.12085, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.469 + ], + "angle": 0, + "content": "[33] Lei Li, Tingting Liu, Chengyu Wang, Minghui Qiu, Cen Chen, Ming Gao, and Aoying Zhou. Resizing codebook of vector quantization without retraining. Multimedia Systems, pages 1-14, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.539 + ], + "angle": 0, + "content": "[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.472, + 0.609 + ], + "angle": 0, + "content": "[35] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16266-16275, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[36] Weiyao Lin, Ming-Ting Sun, Radha Poovendran, and Zhengyou Zhang. Activity recognition using a combination of category components and local models for video surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 18(8):1128-1139, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.472, + 0.747 + ], + "angle": 0, + "content": "[37] Hong Liu, Qinqin He, and Mengyuan Liu. Human action recognition using adaptive hierarchical depth motion maps and gabor filter. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1432-1436. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[39] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[40] Jun Liu, Gang Wang, Ping Hu, Ling-Yu Duan, and Alex C Kot. Global context-aware attention LSTM networks for 3d" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1647-1656, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[43] Benoit Mandelbrot. Structure formelle des textes et communication: Deux études par. Word, 10(1):1-27, 1954. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[44] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[45] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[47] Isabel Papadimitriou and Dan Jurafsky. Pretrain on just structure: Understanding linguistic inductive biases using transfer learning. arXiv preprint arXiv:2304.13060, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[48] Steven T Piantadosi. Zipf's word frequency law in natural language: A critical review and future directions. *Psychonomic bulletin & review*, 21:1112–1130, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[49] Yiwei Qin, Graham Neubig, and Pengfei Liu. Searching for effective multilingual fine-tuning methods: A case study in summarization, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.651 + ], + "angle": 0, + "content": "[50] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.801, + 0.665 + ], + "angle": 0, + "content": "[51] Acquaviva Sam. Hyperbolic vq-vaes. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[52] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[53] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7912-7921, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[54] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Stuart M Shieber. Evidence against the context-freeness of natural language. In The Formal complexity of natural language, pages 320-334. Springer, 1985. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[58] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[59] Lei Wang and Piotr Koniusz. 3mformer: Multi-order multi-mode transformer for skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5620-5631, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[60] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.446 + ], + "angle": 0, + "content": "[61] Xuanhan Wang, Yan Dai, Lianli Gao, and Jingkuan Song. Skeleton-based action recognition via adaptive cross-form learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1670-1678, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.471, + 0.518 + ], + "angle": 0, + "content": "[62] Xinghan Wang, Xin Xu, and Yadong Mu. Neural koopman pooling: Control-inspired temporal dynamics encoding for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10597-10607, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.47, + 0.588 + ], + "angle": 0, + "content": "[63] Shenghua Wei, Yonghong Song, and Yuanlin Zhang. Human skeleton tree recurrent neural network with joint relative motion feature for skeleton based action recognition. In 2017 IEEE international conference on image processing (ICIP), pages 91-95. IEEE, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[64] Rongxiang Weng, Wen Sen Cheng, and Min Zhang. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. In *Findings of the Association for Computational Linguistics: ACL 2023*, pages 4747-4755, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.73 + ], + "angle": 0, + "content": "[65] Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, and Lei Zhang. Generative action description prompts for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10276-10285, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[66] Wentian Xin, Qiguang Miao, Yi Liu, Ruyi Liu, Chi-Man Pun, and Cheng Shi. Skeleton mixformer: Multivariate topology representation for skeleton-based action recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2211-2220, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[67] Haojun Xu, Yan Gao, Zheng Hui, Jie Li, and Xinbo Gao. Language knowledge-assisted representation learning for skeleton-based action recognition. arXiv preprint arXiv:2305.12398, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[68] Kailin Xu, Fanfan Ye, Qiaoyong Zhong, and Di Xie. Topology-aware convolutional neural network for efficient skeleton-based action recognition. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "AAAI Conference on Artificial Intelligence, pages 2866-2874, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[69] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2272-2281, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[70] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[71] Di Yang, Rui Dai, Yaohui Wang, Rupayan Mallick, Luca Minciullo, Gianpiero Francesca, and Francois Bremond. Selective spatio-temporal aggregation based pose refinement system: Towards understanding human activities in real-world videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2363-2372, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[72] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. arXiv preprint arXiv:2107.08580, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.894, + 0.487 + ], + "angle": 0, + "content": "[73] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3427-3435, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[74] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18467-18476, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.561, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[76] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[77] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[78] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[79] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[80] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18405" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[81] Yisheng Zhu, Hui Shuai, Guangcan Liu, and Qingshan Liu. Multilevel spatial-temporal excited graph network for skeleton-based action recognition. IEEE Transactions on Image Processing, 32:496-508, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.442, + 0.193 + ], + "angle": 0, + "content": "[82] G. K. Zipf. The psycho-biology of language, 1935. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18406" + } + ] +] \ No newline at end of file diff --git a/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_origin.pdf b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2ceb55f95a5d4f0b1260ef13e3441b928662cbe --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/8e366b59-bab5-4965-a328-ae40f7bddcd8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5a0addf51ae9d72388441041c14744a1ef154e32e2113109d490fcc1c34b026 +size 1189259 diff --git a/2024/LLMs are Good Action Recognizers/full.md b/2024/LLMs are Good Action Recognizers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e50b2a93812bc45ca55070b673ba1fe9b6e39885 --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/full.md @@ -0,0 +1,376 @@ +# LLMs are Good Action Recognizers + +Haoxuan $\mathrm{Qu}^{1}$ Yujun Cai $^{2}$ Jun Liu $^{1\dagger}$ $^{1}$ Singapore University of Technology and Design $^{2}$ Nanyang Technological University + +haoxuan_qu@mymail.sutd.edu.sg, yujun001@e.ntu.edu.sg, jun.liu@sutd.edu.sg + +# Abstract + +Skeleton-based action recognition has attracted lots of research attention. Recently, to build an accurate skeleton-based action recognizer, a variety of works have been proposed. Among them, some works use large model architectures as backbones of their recognizers to boost the skeleton data representation capability, while some other works pretrain their recognizers on external data to enrich the knowledge. In this work, we observe that large language models which have been extensively used in various natural language processing tasks generally hold both large model architectures and rich implicit knowledge. Motivated by this, we propose a novel LLM-AR framework, in which we investigate treating the Large Language Model as an Action Recognizer. In our framework, we propose a linguistic projection process to project each input action signal (i.e., each skeleton sequence) into its "sentence format" (i.e., an "action sentence"). Moreover, we also incorporate our framework with several designs to further facilitate this linguistic projection process. Extensive experiments demonstrate the efficacy of our proposed framework. + +# 1. Introduction + +Human action recognition aims to categorize the actions performed by humans into a pre-defined list of classes. It is relevant to a variety of applications, such as human-computer interaction [37], intelligent surveillance [36], and virtual reality [13]. In the past few years, skeleton-based action recognition [3, 5, 6, 12, 53, 54, 62, 70] has received a lot of research attention with the notice that skeleton is a succinct yet informative representation of human behaviors. Yet, despite the considerable progress, skeleton-based action recognition still remains a challenging task [78], and to build a more accurate skeleton-based action recognizer, various recent works have been proposed from different perspectives. Among them, some recent works [15, 59] pro + +![](images/f731809c1a45824c07318e0cca1029a4d81a7630cf8c820ac60412eb8131957b.jpg) +Figure 1. Overview of our proposed LLM-AR framework. In our framework, given an input action signal, we first perform a linguistic projection process to acquire the corresponding "action sentence". We then perform action recognition via the large language model with its pre-trained weights untouched to keep its pre-learned rich knowledge. + +posed to utilize large model architectures (such as transformers) as the backbone of their action recognizers to achieve stronger representation capability and capture more subtle differences among different actions. On the other hand, some other works [72, 80] proposed to pre-train their action recognizers on external data in order for their action recognizers to handle this task with richer knowledge. + +Recently, large language models such as GPT [2] and LLaMA [56] have become quite popular and have been extensively applied in handling various human languages. For example, having been pre-trained over various human languages and learned common language-related characteristics, large language models can be very effectively and efficiently adapted to even new human languages unseen during pre-training [49]. Moreover, people have also treated large language models as interpreters in code interpretation [14] and generators in essay generation [22], and found that large language models can handle these tasks effectively. Motivated by this, in this work we are wondering, if we can also treat the large language model as an action recognizer in skeleton-based human action recognition? In general, large language models hold some characteristics that are useful for action recognition. Specifically, pre-trained over a tremendously large corpus [2, 56] that generally contains very rich human-centric behavior descriptions, large language models that hold large model architectures could naturally contain rich implicit knowledge [46, 50] of human behaviors, and hold strong capability in handling different inputs. Thus, it can be very promising if we treat the large + +language model as a human action recognizer. + +However, despite the success of large language models in handling various human languages, it can be challenging to treat them as action recognizers. This is because, while large language models typically take sentences in human languages as input instructions, the input signal of the action recognizer (i.e., the skeleton sequence) is not in a "sentence format" that is "friendly" to large language models. To handle this issue, given a large language model initialized with its pre-trained weights, one potential way is to adjust the weights of the full model to fit the non-linguistic input action signals. Recently, several works [25, 38] have been proposed to fine-tune large language models and adjust their pre-trained weights to handle non-language tasks. However, as shown in the previous study [64], the adjustment of the large language model's pre-trained weights can hurt its generalizability and lead to the loss of its pre-learned rich knowledge. This is clearly undesirable here, as we hope the large language model to be knowledgeable in order for it to be an accurate action recognizer. + +Taking this into account, in this work, we aim to harness the large language model as an action recognizer, and at the same time keep its pre-trained weights untouched to preserve its pre-learned rich knowledge. Specifically, we propose a novel action recognition framework named Large Language Model as an Action Recognizer (LLM-AR). As shown in Fig. 1, to perform action recognition, our LLM-AR framework first involves a linguistic projection process to project the input action signal into its "sentence format", that can be "friendly" and compatible to the large language model pre-trained over human sentences. Then the large language model takes in this "sentence format" of the action signal together with a human sentence as the instruction to predict the name of the action. In the rest of this work, for simplicity, we call the action signal in its "sentence format" an "action sentence". + +In our framework, to project the action signals into "action sentences" (i.e., performing linguistic projection), we incorporate several designs. Basically, to perform such a process, we observe that, every sentence in a language generally can be regarded as a sequence consisting of discrete word tokens. Inspired by both this observation and previous works [75, 77], in LLM-AR, we first learn an action-based vector quantized variational autoencoder (VQ-VAE) model to project each action signal into a sequence of discrete tokens. + +However, while we can basically represent each "action sentence" by a sequence of discrete tokens, it is not necessary that any discrete token sequence can represent an "action sentence" well. In fact, to enable a large language model to become an accurate action recognizer based on the "action sentences", besides being discrete sequences of tokens, ideally, the "action sentences" also need to ful- + +fill some requirements as discussed below. Firstly, since large language models are generally pre-trained over corpus consisting of sentences in human languages, to facilitate large language models in taking "action sentences" as instructions together with sentences in human languages, these "action sentences" should be "like" sentences in human languages so as to be more friendly to the large language model. On the other hand, "action sentences" should still maintain good representations of their original action signals in order for the large language model to accurately perform action recognition based on them. + +To better meet the above requirements in the linguistic projection process, we further incorporate our framework with two designs below. Firstly, to make the "action sentences" more "like" sentences in human languages, we get inspiration from previous linguistic and natural language processing works [27, 47, 48, 55, 82] which show that, languages as the communication tools of human beings often contain human inductive biases. Thus, we propose to incorporate our framework with a learning strategy to regularize the projected "action sentences" to follow human inductive biases as well. Secondly, to make the projected "action sentence" a good representation of its original action signal, motivated by the human skeleton's tree-like nature [63] and hyperbolic space's superior ability in representing tree structures [18], and inspired by previous works [33, 51], we further incorporate our action-based VQ-VAE model with a hyperbolic codebook. + +Once we finish the learning of the linguistic projection process, we perform low-rank adaptation (LoRA) [23] on the large language model to let the model understand the projected "action sentences". Note that since the pre-trained weights of the large language model are untouched throughout the LoRA process, we can preserve the pre-learned rich knowledge in the large language model and thus utilize the large language model conveniently in our framework. + +The contributions of our work are as follows. 1) We propose a novel action recognition framework named LLM-AR. To the best of our knowledge, this work is the first exploration on treating a large language model with its pretrained weights untouched as an action recognizer. 2) We introduce several designs in our framework to facilitate the linguistic projection process and produce the "action sentences". 3) LLM-AR achieves state-of-the-art performance on the evaluated benchmarks. + +# 2. Related Work + +Human Action Recognition. For tackling human action recognition, most of the existing methods can be roughly categorized into two groups: RGB-based methods [58, 60] and skeleton-based methods [5, 6, 53, 54, 70]. + +As skeleton sequences can represent human behaviors in a succinct yet informative manner, skeleton-based ac + +tion recognition has received lots of research attention [3-7, 12, 15, 17, 21, 28, 32, 39, 40, 53, 54, 59, 62, 65-68, 70, 73, 76, 80]. In the early days, different works have been proposed to use different network architectures to handle skeleton-based action recognition. Liu et al. [39] proposed to perform skeleton-based action recognition through a spatial-temporal LSTM architecture. Ke et al. [28] proposed to transform skeleton sequences into grey images and process these images through a CNN network. As time passed, GCN tends to be a popular network architecture [5, 7, 8, 34, 54, 65, 67, 70]. Yan et al. [70] proposed ST-GCN that made the first attempt in performing skeleton-based action recognition using a GCN architecture. After that, various different GCN-based methods have been further proposed, such as AS-GCN [34], 2s-AGCN [54], Shift-GCN [7], and CTR-GCN [5]. More recently, a number of works [15, 59, 76, 80] further proposed to perform skeleton-based action recognition via training a model with a large transformer architecture in an end-to-end manner. Some of such methods include 3Mformer [59] and UPS [15]. + +Different from these approaches, in this work, from a novel perspective, we investigate how to instruct the large language model to perform skeleton-based action recognition while keeping its pre-trained weights untouched to preserve its pre-learned rich knowledge. To achieve this, we involve our framework with a novel linguistic projection process to project each input action signal into an "action sentence". Specifically, we incorporate several designs in the linguistic projection process to make the projected "action sentences" more "like" sentences of human languages so as to make them more friendly to large language models, and meanwhile keep the "action sentence" a good representation of the input action signal. + +Large Language Models. Recently, a variety of large language models have been proposed, such as GPT [2] and LLaMA [56]. Pre-trained over a tremendous number of word tokens, these large language models with large model architectures have been shown to contain very rich knowledge [46, 50]. Consequently, large language models have been extensively explored in various tasks [14, 16, 19, 22, 26, 77, 79], such as language translation [26], essay generation [22], and code interpretation [14]. In this work, we design a novel framework treating the large language model as a human action recognizer leveraging our designed linguistic projection process. + +# 3. Proposed Method + +Given a skeleton sequence as the input action signal, the goal of action recognition is to predict its corresponding action class. Recently, holding large model architectures and containing very rich knowledge, pre-trained large language models have been shown to be powerful in handling sentences of human languages, and thus have become useful + +tools in many natural language processing tasks. Inspired by this, in this work, we aim to leverage the large language model as an effective action recognizer. To achieve this, we propose a novel framework LLM-AR. Specifically, LLM-AR first performs a linguistic projection process to project the input action signal (i.e., the skeleton sequence) into an "action sentence". After that, LLM-AR passes the "action sentence" into the large language model to derive the corresponding human action. Below, we first describe how LLM-AR performs the linguistic projection process, and then introduce the overall training and testing scheme of LLM-AR. + +# 3.1. Linguistic Projection + +In our framework, to enable the large language model to perform action recognition, we first perform a linguistic projection process to project each input action signal into an "action sentence". To learn to perform such a projection, with the observation that each sentence in human languages is essentially a sequence of discrete word tokens, motivate by previous works [75, 77], we first learn an action-based VQ-VAE model to project each input action signal into a sequence of discrete tokens. We then further (1) involve the above learning process with a human-inductive-biases-guided learning strategy to make the projected "action sentences" more "like" sentences in human languages, and (2) incorporate the action-based VQ-VAE model with a hyperbolic codebook to keep "action sentences" good representations of the action signals. + +Action-based VQ-VAE Model. VQ-VAE models [51, 57, 74] have been popularly used in converting an image into a discrete token sequence. Inspired by this, in our framework, to convert the input action signal into a sequence of discrete tokens, we first learn an action-based VQ-VAE model. + +Specifically, similar to the architecture of previous VQ-VAE models [57, 75, 77], our action-based VQ-VAE model involves an encoder $E$ , a decoder $D$ , and a codebook $C$ consisting of $U$ learnable tokens (i.e., $C = \{c_u\}_{u=1}^U$ where $c_u \in \mathbb{R}^{d_u}$ ). We here keep $d_u$ to be the same as the dimension of the word token in the used large language model, and keep $U$ an even number. Among the aforementioned three components of the model, given an input action signal $s_{1:V}$ , where $V$ represents the action length, the encoder $E$ first encodes the action signal through 1D convolutions over the time dimension into a sequence of latent features $f_{1:W}$ where $f_w \in \mathbb{R}^{d_u}$ and $W$ represents the length of the latent feature sequence. After this, to discretize the latent features $f_{1:W}$ , a quantization operation is performed to replace each feature $f_w$ with its nearest token in the codebook as: + +$$ +f _ {w} ^ {d} = \underset {c _ {u} \in C} {\operatorname {a r g m i n}} \left(\operatorname {d i s t} \left(f _ {w}, c _ {u}\right)\right) \tag {1} +$$ + +where $f_w^d$ is the discrete version of $f_w$ , and $dist(\cdot, \cdot)$ represents a distance function. Finally, the decoder aims to + +recover the original action signal $s_{1:V}$ from the sequence of discrete latent features (tokens) $f_{1:W}^{d}$ . Through the above process of encoding, quantization, and decoding, we can project an action signal into a sequence of informative discrete tokens $f_{1:W}^{d}$ . The more detailed architecture of this action-based VQ-VAE model is provided in supplementary. + +Human-inductive-biases-guided Learning Strategy. Above we learn to project each input action signal into a sequence of discrete tokens. To enable the learned token sequences to be more friendly to the large language model, here we aim to make these sequences more "like" sentences in human languages. To achieve this, we get inspiration from massive existing studies [27, 47, 48, 55, 82] which show that, human languages as tools created for human communication naturally contain human inductive biases. Besides, it is also further shown by the previous study [47] that, inputs following human inductive biases are more friendly to large language models. Taking these into consideration, we aim to optimize the set of learned token sequences to also follow human inductive biases, so that these sequences can be more "like" human sentences and thus become more friendly to be used by large language models. + +Below, we first introduce the human inductive biases that are generally recognized as being present in human languages. Such biases include: (a) a human language naturally follows the Zipf's law [47, 48, 82], and (b) a human language is generally context-sensitive [27, 47, 55]. After introducing these human inductive biases, we then describe our proposed human-inductive-biases-guided learning strategy. + +As for the Zipf's law in (a), during daily communications of human beings, there naturally exist some words that are used more commonly and some words more rarely. Zipf's law intuitively represents this imbalanced usage frequency of word tokens in human languages. Specifically, Zipf's law states that, in a human language, the $i$ -th most commonly used word has its usage frequency roughly proportional to: + +$$ +\frac {1}{(i + \beta) ^ {\alpha}} \tag {2} +$$ + +where $\alpha \approx 1$ and $\beta \approx 2.7$ [82]. The reason why Zipf's law consistently appears across different human languages is also theoretically analyzed by various works [9, 43] from different perspectives, such as from the evolution of human communications. + +With respect to the context-sensitivity of human languages in (b), intuitively, the context-sensitivity refers to the inductive bias that when people formulate their sentences, they often do not use each word token independently. Instead, their usage of different tokens in formulating a sentence is often correlated. Note that, while human languages are generally believed by linguists to be context-sensitive, how to explicitly represent context-sensitivity remains a difficult problem. Here, inspired by [47] in its way of repre + +senting context-sensitivity, we represent this human inductive bias as follows. Specifically, given the codebook $C$ consisting of $U$ tokens where $U$ is an even number, during initializing $C$ , we first randomly split the codebook into two halves (i.e., $\{1, \dots, \frac{U}{2}\}$ and $\{1 + \frac{U}{2}, \dots, U\}$ ). Next, we regard each token $c_{u}$ in the first half (i.e., $u \in \{1, \dots, \frac{U}{2}\}$ ) and the token $c_{u + \frac{U}{2}}$ in the second half to be a pair of correlated tokens. As shown in [47], such a pairing mechanism is a simple yet very effective way of representing the context-sensitivity bias in human languages. + +To optimize the set of learned token sequences (i.e., "action sentences") $\{f_{1:W}^d\}$ to also follow the above biases like human languages, we then aim to (a) regularize the set of "action sentences" to follow Zipf's law, as well as (b) regularize each "action sentence" to be formulated using more correlated word tokens so that the formulated "action sentences" can better follow the context-sensitivity bias. To achieve this, we design a human-inductive-biases-guided learning strategy that consists of the following steps. (1) Given a batch of action signals $\{s_{1:V}^b\}_{b=1}^B$ with batch size $B$ , we first encode these signals through the encoder $E$ of the action-based VQ-VAE model to get their corresponding latent features $\{f_{1:W}^b\}_{b=1}^B$ . (2) After that, during discretizing the latent features using tokens $\{c_u\}_{u=1}^U$ in the codebook $C$ , we measure the token usage of each sequence of latent features $f_{1:W}^b$ in a differentiable way via the Gumbel Softmax trick [24] as: + +$$ +t ^ {b} = \sum_ {w = 1} ^ {W} d _ {w} ^ {b}, \text {w h e r e} d _ {w} ^ {b} = \text {G u m b l e - S o f t m a x} (- \operatorname {d i s t} \left(f _ {w} ^ {b}, c _ {u}\right)) \tag {3} +$$ + +where Gumble_Softmax(·) is the Gumble Softmax trick, and $d_w^b$ is the one-hot vector with length $U$ as the output of the Gumble Softmax trick. Besides, $t^b$ is a vector with length $U$ , and the value of the $u$ -th element of $t^b$ represents the number of times token $c_u$ is used in discretizing $f_{1:W}^b$ and formulating its corresponding "action sentence". (3) Next, denoting $D_{Zipf}$ the Zipf distribution with $\alpha = 1$ and $\beta = 2.7$ , we regularize the set of "action sentences" to follow Zipf's law through $L_{Zipf}$ as: + +$$ +L _ {Z i p f} = J S \left(D _ {f r e q} \| D _ {Z i p f}\right), \text {w h e r e} D _ {f r e q} = \frac {\operatorname {s o r t} \left(\sum_ {b = 1} ^ {B} t ^ {b}\right)}{B \times W} \tag {4} +$$ + +where $sort(\cdot)$ is the sorting operation used to order tokens based on their usage frequency, $D_{freq}$ is the distribution representing the token usage frequency, and $JS(\cdot \| \cdot)$ represents the JS divergence between two distributions. (4) At the same time, we encourage the "action sentences" to follow the context-sensitivity bias and use more correlated tokens via $L_{context}$ as: + +$$ +L _ {\text {c o n t e x t}} = 1 - \frac {\sum_ {b = 1} ^ {B} \operatorname {C o r r} \left(t ^ {b}\right)}{B \times W} \tag {5} +$$ + +where $Corr(t^{b})$ leverages min pooling over every pair of + +![](images/3de685c669892722039fa7e1470f99a549983476514df4db0415f7f3c664734b.jpg) +Figure 2. Overview of the action-based VQ-VAE model with the hyperbolic codebook $C_H$ incorporated. Given a batch of input action signals $\{s_{1:V}^b\}_{b=1}^B$ , to optimize the action-based VQ-VAE model, $\{s_{1:V}^b\}_{b=1}^B$ are first fed to the encoder $E$ to get the corresponding latent features $\{f_{1:W}^b\}_{b=1}^B$ . Next, to leverage the hyperbolic codebook $C_H$ that can serve as a good representation of the tree-like human skeletons to perform quantization, $\{f_{1:W}^b\}_{b=1}^B$ are projected into the hyperbolic space via the process of E-to-H projection. After that, the quantization is performed in the hyperbolic space using $dist_{\mathbb{B}}(\widetilde{f}_w, c_u)$ defined in Eq. 9 as the distance function. Finally, after quantization, the discrete version of the latent features are passed back into the Euclidean space via the process of H-to-E projection to reconstruct the input action signals through the decoder $D$ . + +elements in $t^b$ (e.g., the pair of $u$ -th element and $(u + \frac{U}{2})$ -th element) to calculate the number of correlated tokens used in discretizing $f_{1:W}^b$ (more details about this method to measure $Corr(t^b)$ are provided in supplementary). Note that via minimizing $L_{\text{context}}$ , we encourage more correlated tokens to be used in formulating "action sentences". (5) Finally, we formulate the loss function $L_{\text{human}}$ that we use to perform our human-inductive-biases-guided strategy as: + +$$ +L _ {h u m a n} = L _ {Z i p f} + L _ {\text {c o n t e x t}} \tag {6} +$$ + +Via incorporating the above learning strategy into the learning process of the action-based VQ-VAE model, we can regularize the "action sentences" to better follow the human inductive biases that are present in human languages and thus make them more "like" sentences in human languages. + +Above we inject human inductive biases into "action sentences". Here, we notice that, besides inductive biases that can be explicitly described, human languages can also present other implicit characteristics. Thus, to enable "action sentences" to be more friendly to large language models which are generally pre-trained over various human languages, we aim to further incorporate implicit characteristics into "action sentences" as well. Specifically, since large language models such as LLaMA typically use word tokens $\{c_{LLM}\}$ stored in its first layer to formulate sentences in human languages, we here further align (1) the tokens $\{c_u\}_{u = 1}^U$ in codebook $C$ that are used to formulate the "action sentences" with (2) the word tokens $\{c_{LLM}\}$ used in the large language model. To achieve such an alignment, we leverage Maximum Mean Discrepancy (MMD) [20] as an effective feature alignment technique to measure the discrepancy between $\{c_u\}_{u = 1}^U$ and $\{c_{LLM}\}$ . Specifically, the less MMD $(\{c_u\}_{u = 1}^U,\{c_{LLM}\})$ is, $\{c_u\}_{u = 1}^U$ and $\{c_{LLM}\}$ can be regarded as more aligned. We then incorporate this alignment procedure into our proposed strategy + +via rewriting $L_{human}$ in Eq. 6 as: + +$$ +L _ {h u m a n} = L _ {Z i p f} + L _ {\text {c o n t e x t}} + \mathrm {M M D} \left(\left\{c _ {u} \right\} _ {u = 1} ^ {U}, \left\{c _ {L L M} \right\}\right) \tag {7} +$$ + +By incorporating this alignment procedure into the previously mentioned learning strategy, we can then lead the formulated "action sentences" $\{f_{1:W}^{d}\}$ to be more friendly to be used by large language models. + +Hyperbolic Codebook. Besides making the "action sentences" more friendly to the large language model, we also aim to keep them good representations of the original input action signals (i.e., the skeleton sequences). Motivated by the tree-like nature of human skeletons [63], we aim to make the word tokens in the "action sentences" to well represent such a structure. To achieve this, inspired by the superior capability of the hyperbolic space in embedding tree structures [18] and motivated by previous VQ-VAE works [33, 51], we here further incorporate our action-based VQ-VAE model with a hyperbolic codebook utilizing the Poincaré ball model [18, 45]. The Poincaré ball model is an isometric model that can represent the hyperbolic space. Formally, the $n$ -dimensional Poincaré ball model is defined as $(\mathbb{B}_c^n,g_{\mathbf{x}}^c)$ , where $\mathbb{B}_c^n = \{\mathbf{x}\in \mathbb{R}^n:c\| \mathbf{x}\| < 1\}$ , $g_{\mathbf{x}}^{c} = (\gamma_{\mathbf{x}}^{c})^{2}I_{n}$ is the Riemannian metric tensor, $\gamma_{\mathbf{x}}^{c} = \frac{2}{1 - c\|\mathbf{x}\|^2}$ is the conformal factor, $I_{n}$ is the Euclidean metric tensor, and $c$ is the curvature. + +Denoting the hyperbolic codebook $C_H = \{c_u\}_{u=1}^U$ , where $c_u \in \mathbb{B}_c^{d_u}$ , we make the following three changes to incorporate $C_H$ into our action-based VQ-VAE model. We also illustrate where these changes take place in our action-based VQ-VAE model in Fig. 2. (1) Euclidean-to-Hyperbolic (E-to-H) Projection. Given a latent feature $f_w$ which is originally in the Euclidean space, to convert it into a discrete token using a hyperbolic codebook, we first need to project $f_w$ into the hyperbolic space. Specifically, following [18], we perform such a projection through the ex + +ponential map $\exp_0^c (\cdot)$ as: + +$$ +\widetilde {f} _ {w} = \exp_ {\mathbf {0}} ^ {c} \left(f _ {w}\right) = \tanh (\sqrt {c} \| f _ {w} \|) \frac {f _ {w}}{\sqrt {c} \| f _ {w} \|} \tag {8} +$$ + +where $\widetilde{f}_w$ is the projected feature of $f_w$ in the hyperbolic space. (2) Hyperbolic Distance Calculation. Given the projected feature $\widetilde{f}_w$ , to perform quantization of $\widetilde{f}_w$ based on the tokens in $C_H$ , we need a distance function defined in the hyperbolic space. Here, for simplicity, we use the popularly used geodesic/induced distance [18, 29]. Denote this distance as $\text{dist}_{\mathbb{B}}(\cdot, \cdot)$ . $\text{dist}_{\mathbb{B}}(\widetilde{f}_w, c_u)$ between $\widetilde{f}_w$ and $c_u$ can be defined as: + +$$ +\operatorname {d i s t} _ {\mathbb {B}} \left(\widetilde {f} _ {w}, c _ {u}\right) = \operatorname {a r c c o s h} \left(1 + 2 \frac {\| \widetilde {f} _ {w} - c _ {u} \| ^ {2}}{\left(1 - \| \widetilde {f} _ {w} \| ^ {2}\right) \left(1 - \| c _ {u} \| ^ {2}\right)}\right) \tag {9} +$$ + +(3) Hyperbolic-to-Euclidean (H-to-E) Projection. After quantization, we need to project $\widetilde{f}_w^d$ (the discrete version of $\widetilde{f}_w$ ) back into the Euclidean space in order for it to be passed to the decoder $D$ . To achieve this, we follow [18] to leverage the logarithmic map $\log_0^c (\cdot)$ as: + +$$ +f _ {w} ^ {d} = \log_ {\mathbf {0}} ^ {c} (\widetilde {f} _ {w} ^ {d}) = \operatorname {a r c t a n h} (\sqrt {c} \| \widetilde {f} _ {w} ^ {d} \|) \frac {\widetilde {f} _ {w} ^ {d}}{\sqrt {c} \| \widetilde {f} _ {w} ^ {d} \|} \tag {10} +$$ + +where $f_{w}^{d}$ represents the discrete version of the feature projected back in the Euclidean space. With the above changes made, we can seamlessly involve the hyperbolic codebook $C_H$ into our action-based VQ-VAE model, and make the projected "action sentence" a better representation of the input action signal. + +# 3.2. Overall Training and Testing + +Above we describe how we perform the linguistic projection process in our framework to project each input action signal into an "action sentence". Below, we introduce the overall training and testing scheme of our framework. + +Training. The training process of our framework consists of the following two stages: (1) optimizing the action-based VQ-VAE model incorporated with the hyperbolic codebook $C_H$ to acquire the ability to project each action signal $s_{1:V}$ into its corresponding "action sentence" $\widetilde{f}_{1:W}^d$ ; and (2) perform low-rank adaptation (LoRA) [23] on the large language model for it to better understand the projected "action sentences". + +At the first stage, to optimize the action-based VQ-VAE model, given a batch of action signals, we first follow previous VQ-VAE works [57, 75, 77] to use the following loss functions, including the reconstruction loss, the embedding loss, and the commitment loss. Among them, the reconstruction loss $L_{re}$ is designed to regularize the difference between the input action signal and the action signal reconstructed from the decoder $D$ , the embedding loss $L_{embed}$ is designed to learn the tokens in the codebook, and the commitment loss $L_{commit}$ is designed to encourage each + +of the encoded latent features to stay close to its discrete version. Formally speaking, denoting $\{s_{1:V}^b\}_{b=1}^B$ a batch of input action signals, $\{\hat{s}_{1:V}^b\}_{b=1}^B$ the corresponding batch of reconstructed action signals, $\{f_{1:W}^b\}_{b=1}^B$ the latent features encoded from $\{s_{1:V}^b\}_{b=1}^B$ , and $\{f_{1:W}^{d,b}\}_{b=1}^B$ the discrete versions of these latent features, the above three loss functions can be written as: + +$$ +L _ {r e} = L _ {1} ^ {s m o o t h} (\{s _ {1: V} ^ {b} \} _ {b = 1} ^ {B}, \{\hat {s} _ {1: V} ^ {b} \} _ {b = 1} ^ {B}) +$$ + +$$ +L _ {e m b e d} = \left\| s g \left(\left\{f _ {1: W} ^ {b} \right\} _ {b = 1} ^ {B}\right) - \left\{f _ {1: W} ^ {d, b} \right\} _ {b = 1} ^ {B} \right\| ^ {2} \tag {11} +$$ + +$$ +L _ {c o m m i t} = \left\| \left\{f _ {1: W} ^ {b} \right\} _ {b = 1} ^ {B} - s g \left(\left\{f _ {1: W} ^ {d, b} \right\} _ {b = 1} ^ {B}\right) \right\| ^ {2} +$$ + +where $L_{1}^{smooth}(\cdot)$ represents the L1 smooth loss and $sg(\cdot)$ is the stop-gradient operation that is used to prevent the gradient from flowing through its operand. In addition, to make the learned "action sentences" more like sentences in human languages, we also incorporate the above three loss functions with $L_{human}$ defined in Eq. 7. Overall, we can write the total loss $L_{total}$ for the first training stage as: + +$$ +L _ {\text {t o t a l}} = L _ {r e} + L _ {\text {e m b e d}} + \omega_ {1} L _ {\text {c o m m i t}} + \omega_ {2} L _ {\text {h u m a n}} \tag {12} +$$ + +where $\omega_{1}$ and $\omega_{2}$ are weighting hyperparameters. + +We then perform LoRA [23] on the large language model in the second stage in order for it to better understand the "action sentences" while keeping the pre-trained weights of the model untouched. Specifically, for each training sample consisting of an input action signal and its corresponding ground-truth action, we perform the following steps. (1) We first project the action signal into an "action sentence" of discrete tokens through the action-based VQ-VAE model learned in the first stage. (2) We then instruct the large language model to act as an action recognizer via a simple instruction as: "Given a sequence of action tokens [tokens], please predict the corresponding action.", where [tokens] represent the word tokens of the "action sentence" derived in step (1). (3) Finally, during LoRA process, we pass the above instruction into the large language model and encourage the similarity between the tokens $t_p$ predicted by the large language model and the tokens $t_g$ representing the ground-truth action as: + +$$ +L _ {L o R A} = L _ {c e} \left(t _ {p}, t _ {g}\right) \tag {13} +$$ + +where $L_{ce}(\cdot ,\cdot)$ represents the cross-entropy loss. + +Testing. During testing, for each testing action signal, we first use the learned action-based VQ-VAE model to derive its corresponding "action sentence". After that, we use the same instruction as in step (2) above to instruct the large language model to predict the corresponding action. + +# 4. Experiments + +To evaluate the efficacy of our framework, we conduct experiments on 4 datasets including NTU RGB+D, NTU RGB+D 120, Toyota Smarthome, and UAV-Human. + +# 4.1. Datasets + +NTU RGB+D [52] is a large-scale dataset popularly used in human action recognition. It contains around 56k skeleton sequences from 60 activity classes. On this dataset, following [52], we evaluate our method under the Cross-Subject (X-Sub) and Cross-View (X-View) evaluation protocols. + +NTU RGB+D 120 [41] is an extension of the NTU RGB+D dataset. It consists of more than 114k skeleton sequences across 120 activity classes. Following [41], we evaluate our method on this dataset under the Cross-Subject (X-Sub) and Cross-Setup (X-Set) evaluation protocols. + +Toyota Smarthome [10] contains 16,115 video samples over 31 activity classes. On this dataset, we use the skeleton sequences pre-processed by [71] and we follow it to evaluate our method under the Cross-Subject (X-Sub) and two Cross-View (X-View1 & X-View2) evaluation protocols. + +UAV-Human [35] is a dataset that is captured by unmanned aerial vehicles (UAV). It contains more than 20k skeleton sequences over 155 activity classes, and it is collected from 119 distinct subjects. Following [35], we use 89 subjects for training and 30 subjects for testing. + +# 4.2. Implementation Details + +We conduct our main experiments on Nvidia V100 GPU and we use LLaMA-13B [56] as the large language model. During the training process of the action-based VQ-VAE model, we optimize the model for 300,000 iterations using the AdamW [42] optimizer with an initial learning rate of 2e-4. We set the batch size $B$ to 256, the number of tokens $U$ to 512, and the curvature $c$ of the hyperbolic codebook to 1. Additionally, we set the dimension of each word token $d_{u}$ to the same size (5120) as the word token in LLaMA-13B. Moreover, we set $w_{1}$ to 0.02 following previous VQ-VAE works [44, 75] and set $w_{2}$ to 0.2. Besides, we set the length $(W)$ of each sequence of latent features to be a quarter of the length $(V)$ of its corresponding input action signal. During the LoRA process, we develop our code based on the Github repo [1], set the number of iterations to 75,000, and use the AdamW optimizer with an initial learning rate of 3e-3. Moreover, we set the batch size to 256, partitioned into micro-batches of 4. Besides, we set the two hyperparameters of the LoRA process, i.e., $r_{LoRA}$ and $\alpha_{LoRA}$ , to 64 and 16 respectively. + +# 4.3. Comparison with State-of-the-art Methods + +On NTU RGB+D and NTU RGB+D 120 datasets, following the experimental setting of recent works [12, 80], we only use the joint modality of human skeletons during our experiments. We report the results in Tab. 1. As shown, compared to existing skeleton-based action recognition methods, our method consistently achieves the best performance across all the evaluation protocols. This demonstrates the effectiveness of our method. Besides, we also + +Table 1. Performance comparison on the NTU RGB+D and NTU RGB+D 120 datasets. + +
MethodNTU RGB+DNTU RGB+D 120
X-SubX-ViewX-SubX-Set
ST-GCN [70]85.792.482.184.5
Shift-GCN [7]87.895.180.983.2
InfoGCN [8]89.895.285.186.3
PoseC3D [11]93.796.585.989.7
FR-Head [78]90.395.385.587.3
Koopman [62]90.295.285.787.4
GAP [65]90.295.685.587.0
HD-GCN [31]90.695.785.787.3
STC-Net [30]91.096.286.288.0
DSTformer [80]93.097.2--
SkeleTR [12]94.897.787.888.3
Ours95.098.488.791.5
+ +Table 2. Performance comparison on the Toyota Smarthome dataset. + +
MethodX-SubX-View1X-View2
2S-AGCN [54]58.832.257.9
SSTA-PRS [71]62.122.854.0
UNIK [72]62.133.463.6
ML-STGNet [81]64.629.963.5
Ours67.036.166.6
+ +Table 3. Performance comparison on the UAV-Human dataset. + +
MethodX-Sub
ST-GCN [69]30.3
2S-AGCN [54]34.8
Shift-GCN [7]38.0
ACFL [61]44.2
Ours46.3
+ +report results on the Toyota Smarthome dataset in Tab. 2, and on the UAV-Human dataset in Tab. 3. As shown, our method consistently achieves the best performance on these datasets. This further shows the efficacy of our method. + +# 4.4. Ablation Studies + +We conduct extensive ablation experiments on the X-Set protocol of the NTU RGB+D 120 dataset. More ablation studies such as experiments w.r.t. hyperparameters are in supplementary. + +# Impact of the human-inductive-biases-guided + +learning strategy. In our framework, to lead the formulated "action sentences" to be more friendly to large language models, we design a human-inductive-biases + +Table 4. Evaluation on the human-inductive-biases-guided learning strategy. + +
MethodAccuracy
w/o biases87.6
w/o LZipf89.9
w/o Lcontext89.8
w/o MMD alignment90.3
LLM-AR91.5
+ +guided learning strategy consisting of three components (as shown in Eq. 7). To evaluate the efficacy of this strategy, we test four variants. In the first variant (w/o biases), we remove the whole strategy (i.e., all its three components $L_{Zipf}$ , $L_{context}$ , and MMD alignment) from the learning process. In the second variant (w/o $L_{Zipf}$ ), we still utilize the strategy but remove its $L_{Zipf}$ component. Moreover, in the third variant (w/o $L_{context}$ ), we remove the $L_{context}$ component from the strategy, whereas in the fourth variant (w/o MMD alignment), we remove the MMD alignment component. As shown in Tab. 4, compared to + +our framework, the performance of the first variant drops significantly. This shows the importance of formulating "action sentences" like sentences in human languages. Moreover, our framework also outperforms all the other three variants. This further shows the effectiveness of all the three components of our proposed learning strategy. + +# Impact of discretizing latent features into "action sentences". In our framework, we discretize the encoded latent features to formulate "action sentences" consisting + +Table 5. Evaluation on discretizing latent features into "action sentences". + +
MethodAccuracy
w/o discretization83.4
with discretization91.5
+ +of discrete word tokens (with discretization). To valid this design, we test a variant. In this variant (w/o discretization), instead of discretizing the encoded latent features into "action sentences", we directly pass these continuous features to the intermediate layers of the large language model. As shown in Tab. 5, our framework with discretization outperforms this variant. This shows the advantage of discretizing latent features into "action sentences", which are more "like" human sentences consisting of discrete word tokens, and thus are more friendly to the large language model pre-trained over human sentences. + +# Impact of the hyperbolic codebook $C_H$ . In our framework, we incorporate our action-based VQ-VAE model with a hyperbolic codebook $C_H$ (with $C_H$ ). To validate the efficacy of $C_H$ , we test a variant $(\mathbf{w} / \mathbf{o} C_H)$ in which + +Table 6. Evaluation on the hyperbolic codebook $C_H$ + +the codebook is set up in the Euclidean instead of hyperbolic space. As shown in Tab. 6, our framework involving $C_H$ performs better than this variant. This shows the efficacy of $C_H$ in the hyperbolic space, which can facilitate the "action sentences" in representing the tree-like-structured input action signals better. + +# Impact of the LoRA process. In our framework, to make the large language model understand the "action sentences" while keeping its pretrained weights untouched to pre + +Table 7. Evaluation on the LoRA process. + +
MethodAccuracy
All tuning79.6
LLM-AR91.5
+ +serve its rich pre-learned knowledge, we tune the large language model through a LoRA process. To validate this scheme, here we also test a variant (all tuning) on A100 GPU. In this variant, during tuning, after initializing the large language model with its pre-trained weights, all the parameters of the model will undergo gradient updates. As shown in Tab. 7, our framework achieves much better performance than this variant. This shows the superiority of our framework in choosing to perform tuning via LoRA, which enables the large language model's pre-trained weights to be untouched and maintains its pre-learned rich knowledge. + +Evaluation on unseen activity classes. In the main experiments, following [31, 52, 70], we evaluate our framework on activity classes that have been seen during training. Here, inspired by that large language + +Table 8. Evaluation on unseen activity classes. + +
MethodAccuracy
All tuning37.7
LLM-AR62.4
+ +models naturally could contain rich knowledge beyond the training activity classes used in our experiments, we are curious, assuming we have a list of testing activity classes unseen during training, can we also use our framework to perform action recognition on these classes? To answer this question, we first build a new evaluation protocol for unseen activity classes based on the NTU RGB+D 120 dataset. Under this new protocol, during each time of evaluation, we randomly select 3 classes to form the unseen class list (i.e., the list of testing classes), and use the remaining classes as the classes seen during training (i.e., the training classes). Besides, we instruct the large language model as: "Given a sequence of action tokens [tokens], please predict the corresponding action from [list]," where [tokens] represent the word tokens of the "action sentence" and [list] represents the unseen class list. We then perform the above evaluation for five times and report the average performance. As shown in Tab. 8, even testing on activity classes unseen during training, our framework can still achieve a relatively good performance, while the afore-defined all tuning variant performs much worse. This can be analyzed as, large language models could contain rich pre-learned knowledge w.r.t. the list of unseen classes. Thus, our framework that maintains such rich knowledge can still perform promising recognition on these unseen classes, while the all tuning variant that can lose amount of pre-learned knowledge of the large language model would yield a much worse performance. This further shows the advantage of our framework in maintaining the pre-learned rich knowledge of the large language model. + +# 5. Conclusion + +In this paper, we have proposed a novel action recognition framework LLM-AR. In LLM-AR, we treat the large language model as an action recognizer, and instruct the large language model to perform action recognition using its contained rich knowledge. Specifically, to lead the input action signals (i.e., the skeleton sequences) to be more friendly to the large language model, we first propose a linguistic projection process to project each action signal into an "action sentence". Moreover, we also introduce several designs to further facilitate this process. Our framework consistently achieves SOTA performance across different benchmarks. + +Acknowledgement. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014). + +# References + +[1] Lit-llama. https://github.com/Lightning-AI/ lit-llama.7 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3 +[3] Dongqi Cai, Yangyuxuan Kang, Anbang Yao, and Yurong Chen. Ske2grid: Skeleton-to-grid representation learning for action recognition. 2023. 1, 3 +[4] Jinghong Chen, Chong Zhao, Qicong Wang, and Hongying Meng. Hmanet: Hyperbolic manifold aware network for skeleton-based action recognition. IEEE Transactions on Cognitive and Developmental Systems, 2022. +[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13359-13368, 2021. 1, 2, 3 +[6] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 536-553. Springer, 2020. 1, 2 +[7] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 183-192, 2020. 3, 7 +[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 3, 7 +[9] Bernat Corominas-Murtra, Jordi Fortune, and Ricard V Solé. Emergence of zipf's law in the evolution of communication. Physical Review E, 83(3):036115, 2011. 4 +[10] Srijan Das, Rui Dai, Michal Koperski, Luca Minciullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome: Real-world activities of daily living. In Proceedings of the IEEE/CVF international conference on computer vision, pages 833-842, 2019. 7 +[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 7 +[12] Haodong Duan, Mingze Xu, Bing Shuai, Davide Modolo, Zhuowen Tu, Joseph Tighe, and Alessandro Bergamo. Skeletr: Towards skeleton-based action recognition in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13634-13644, 2023. 1, 3, 7 +[13] Abassin Sourou Fangbemi, Bin Liu, Neng Hai Yu, and Yanxiang Zhang. Efficient human action recognition interface for + +augmented and virtual reality applications based on binary descriptor. In Augmented Reality, Virtual Reality, and Computer Graphics: 5th International Conference, AVR 2018, Otranto, Italy, June 24-27, 2018, Proceedings, Part I 5, pages 252-260. Springer, 2018. 1 +[14] Faisal Firdous, Saimul Bashir, Syed Zoofa Rufai, and Sanjeev Kumar. Openai chatgpt as a logical interpreter of code. In 2023 2nd International Conference on Edge Computing and Applications (ICECAA), pages 1192-1197. IEEE, 2023. 1, 3 +[15] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13030, 2023. 1, 3 +[16] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 3 +[17] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In The Eleventh International Conference on Learning Representations, 2023. 3 +[18] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 2, 5, 6 +[19] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[20] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 5 +[21] Ryo Hachiuma, Fumiaki Sato, and Taiki Sekii. Unified keypoint-based action recognition framework via structured keypoint pooling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22962-22971, 2023. 3 +[22] Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. Ai, write an essay for me: A large-scale comparison of human-written versus chatgpt-generated essays. arXiv preprint arXiv:2304.14276, 2023. 1, 3 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6 +[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4 +[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2 +[26] Wenxiang Jiao, Wenxuan Wang, JT Huang, Xing Wang, and ZP Tu. Is chatgpt a good translator? yes with gpt-4 as the engine. arXiv preprint arXiv:2301.08745, 2023. 3 +[27] Aravind K Joshi, K Vijay Shanker, and David Weir. The convergence of mildly context-sensitive grammar formalisms. Technical Reports (CIS), page 539, 1990. 2, 4 + +[28] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 3 +[29] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 6 +[30] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10255-10264, 2023. 7 +[31] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10444-10453, 2023. 7, 8 +[32] Zikang Leng, Hyeokhyen Kwon, and Thomas Plötz. On the benefit of generative foundation models for human activity recognition. arXiv preprint arXiv:2310.12085, 2023. 3 +[33] Lei Li, Tingting Liu, Chengyu Wang, Minghui Qiu, Cen Chen, Ming Gao, and Aoying Zhou. Resizing codebook of vector quantization without retraining. Multimedia Systems, pages 1-14, 2023. 2, 5 +[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 3 +[35] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16266-16275, 2021. 7 +[36] Weiyao Lin, Ming-Ting Sun, Radha Poovendran, and Zhengyou Zhang. Activity recognition using a combination of category components and local models for video surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 18(8):1128-1139, 2008. 1 +[37] Hong Liu, Qinqin He, and Mengyuan Liu. Human action recognition using adaptive hierarchical depth motion maps and gabor filter. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1432-1436. IEEE, 2017. 1 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2 +[39] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 3 +[40] Jun Liu, Gang Wang, Ping Hu, Ling-Yu Duan, and Alex C Kot. Global context-aware attention LSTM networks for 3d + +action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1647-1656, 2017. 3 +[41] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 7 +[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7 +[43] Benoit Mandelbrot. Structure formelle des textes et communication: Deux études par. Word, 10(1):1-27, 1954. 4 +[44] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 7 +[45] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 5 +[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 3 +[47] Isabel Papadimitriou and Dan Jurafsky. Pretrain on just structure: Understanding linguistic inductive biases using transfer learning. arXiv preprint arXiv:2304.13060, 2023. 2, 4 +[48] Steven T Piantadosi. Zipf's word frequency law in natural language: A critical review and future directions. *Psychonomic bulletin & review*, 21:1112–1130, 2014. 2, 4 +[49] Yiwei Qin, Graham Neubig, and Pengfei Liu. Searching for effective multilingual fine-tuning methods: A case study in summarization, 2022. 1 +[50] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[51] Acquaviva Sam. Hyperbolic vq-vaes. 2, 3, 5 +[52] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7, 8 +[53] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7912-7921, 2019. 1, 2, 3 +[54] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 3, 7 +[55] Stuart M Shieber. Evidence against the context-freeness of natural language. In The Formal complexity of natural language, pages 320-334. Springer, 1985. 2, 4 + +[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 7 +[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 6 +[58] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 2 +[59] Lei Wang and Piotr Koniusz. 3mformer: Multi-order multi-mode transformer for skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5620-5631, 2023. 1, 3 +[60] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 2 +[61] Xuanhan Wang, Yan Dai, Lianli Gao, and Jingkuan Song. Skeleton-based action recognition via adaptive cross-form learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1670-1678, 2022. 7 +[62] Xinghan Wang, Xin Xu, and Yadong Mu. Neural koopman pooling: Control-inspired temporal dynamics encoding for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10597-10607, 2023. 1, 3, 7 +[63] Shenghua Wei, Yonghong Song, and Yuanlin Zhang. Human skeleton tree recurrent neural network with joint relative motion feature for skeleton based action recognition. In 2017 IEEE international conference on image processing (ICIP), pages 91-95. IEEE, 2017. 2, 5 +[64] Rongxiang Weng, Wen Sen Cheng, and Min Zhang. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. In *Findings of the Association for Computational Linguistics: ACL 2023*, pages 4747-4755, 2023. 2 +[65] Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, and Lei Zhang. Generative action description prompts for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10276-10285, 2023. 3, 7 +[66] Wentian Xin, Qiguang Miao, Yi Liu, Ruyi Liu, Chi-Man Pun, and Cheng Shi. Skeleton mixformer: Multivariate topology representation for skeleton-based action recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2211-2220, 2023. +[67] Haojun Xu, Yan Gao, Zheng Hui, Jie Li, and Xinbo Gao. Language knowledge-assisted representation learning for skeleton-based action recognition. arXiv preprint arXiv:2305.12398, 2023. 3 +[68] Kailin Xu, Fanfan Ye, Qiaoyong Zhong, and Di Xie. Topology-aware convolutional neural network for efficient skeleton-based action recognition. In Proceedings of the + +AAAI Conference on Artificial Intelligence, pages 2866-2874, 2022. 3 +[69] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2272-2281, 2017. 7 +[70] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7, 8 +[71] Di Yang, Rui Dai, Yaohui Wang, Rupayan Mallick, Luca Minciullo, Gianpiero Francesca, and Francois Bremond. Selective spatio-temporal aggregation based pose refinement system: Towards understanding human activities in real-world videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2363-2372, 2021. 7 +[72] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. arXiv preprint arXiv:2107.08580, 2021. 1, 7 +[73] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3427-3435, 2023. 3 +[74] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18467-18476, 2023. 3 +[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2, 3, 6, 7 +[76] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3 +[77] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 2, 3, 6 +[78] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7 +[79] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023. 3 +[80] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings + +of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 7 +[81] Yisheng Zhu, Hui Shuai, Guangcan Liu, and Qingshan Liu. Multilevel spatial-temporal excited graph network for skeleton-based action recognition. IEEE Transactions on Image Processing, 32:496-508, 2022. 7 +[82] G. K. Zipf. The psycho-biology of language, 1935. 2, 4 \ No newline at end of file diff --git a/2024/LLMs are Good Action Recognizers/images.zip b/2024/LLMs are Good Action Recognizers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90a962bfaff813392725c89cde14bc429b0878fc --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bc59d1a1d4a71cb6df91039d49543c4b3a930f00ebb22d0c12eead3ff076d41 +size 242812 diff --git a/2024/LLMs are Good Action Recognizers/layout.json b/2024/LLMs are Good Action Recognizers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf0e13cfbad720a7f02ea2bc98192ad6fb517461 --- /dev/null +++ b/2024/LLMs are Good Action Recognizers/layout.json @@ -0,0 +1,11183 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 184, + 103, + 409, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 103, + 409, + 121 + ], + "spans": [ + { + "bbox": [ + 184, + 103, + 409, + 121 + ], + "type": "text", + "content": "LLMs are Good Action Recognizers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "spans": [ + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "content": "Haoxuan " + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "inline_equation", + "content": "\\mathrm{Qu}^{1}" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "content": " Yujun Cai" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "content": " Jun Liu" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "content": "Singapore University of Technology and Design " + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 86, + 142, + 507, + 174 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 174, + 489, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 174, + 489, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 174, + 489, + 185 + ], + "type": "text", + "content": "haoxuan_qu@mymail.sutd.edu.sg, yujun001@e.ntu.edu.sg, jun.liu@sutd.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 251, + 290, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 290, + 491 + ], + "type": "text", + "content": "Skeleton-based action recognition has attracted lots of research attention. Recently, to build an accurate skeleton-based action recognizer, a variety of works have been proposed. Among them, some works use large model architectures as backbones of their recognizers to boost the skeleton data representation capability, while some other works pretrain their recognizers on external data to enrich the knowledge. In this work, we observe that large language models which have been extensively used in various natural language processing tasks generally hold both large model architectures and rich implicit knowledge. Motivated by this, we propose a novel LLM-AR framework, in which we investigate treating the Large Language Model as an Action Recognizer. In our framework, we propose a linguistic projection process to project each input action signal (i.e., each skeleton sequence) into its \"sentence format\" (i.e., an \"action sentence\"). Moreover, we also incorporate our framework with several designs to further facilitate this linguistic projection process. Extensive experiments demonstrate the efficacy of our proposed framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 693 + ], + "type": "text", + "content": "Human action recognition aims to categorize the actions performed by humans into a pre-defined list of classes. It is relevant to a variety of applications, such as human-computer interaction [37], intelligent surveillance [36], and virtual reality [13]. In the past few years, skeleton-based action recognition [3, 5, 6, 12, 53, 54, 62, 70] has received a lot of research attention with the notice that skeleton is a succinct yet informative representation of human behaviors. Yet, despite the considerable progress, skeleton-based action recognition still remains a challenging task [78], and to build a more accurate skeleton-based action recognizer, various recent works have been proposed from different perspectives. Among them, some recent works [15, 59] pro" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 213, + 547, + 277 + ], + "blocks": [ + { + "bbox": [ + 307, + 213, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 307, + 213, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 547, + 277 + ], + "type": "image", + "image_path": "f731809c1a45824c07318e0cca1029a4d81a7630cf8c820ac60412eb8131957b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 284, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 305, + 284, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 547, + 350 + ], + "type": "text", + "content": "Figure 1. Overview of our proposed LLM-AR framework. In our framework, given an input action signal, we first perform a linguistic projection process to acquire the corresponding \"action sentence\". We then perform action recognition via the large language model with its pre-trained weights untouched to keep its pre-learned rich knowledge." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 448 + ], + "type": "text", + "content": "posed to utilize large model architectures (such as transformers) as the backbone of their action recognizers to achieve stronger representation capability and capture more subtle differences among different actions. On the other hand, some other works [72, 80] proposed to pre-train their action recognizers on external data in order for their action recognizers to handle this task with richer knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 548, + 715 + ], + "type": "text", + "content": "Recently, large language models such as GPT [2] and LLaMA [56] have become quite popular and have been extensively applied in handling various human languages. For example, having been pre-trained over various human languages and learned common language-related characteristics, large language models can be very effectively and efficiently adapted to even new human languages unseen during pre-training [49]. Moreover, people have also treated large language models as interpreters in code interpretation [14] and generators in essay generation [22], and found that large language models can handle these tasks effectively. Motivated by this, in this work we are wondering, if we can also treat the large language model as an action recognizer in skeleton-based human action recognition? In general, large language models hold some characteristics that are useful for action recognition. Specifically, pre-trained over a tremendously large corpus [2, 56] that generally contains very rich human-centric behavior descriptions, large language models that hold large model architectures could naturally contain rich implicit knowledge [46, 50] of human behaviors, and hold strong capability in handling different inputs. Thus, it can be very promising if we treat the large" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 141, + 713 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18395" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 234, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 234, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 234, + 83 + ], + "type": "text", + "content": "language model as a human action recognizer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 314 + ], + "type": "text", + "content": "However, despite the success of large language models in handling various human languages, it can be challenging to treat them as action recognizers. This is because, while large language models typically take sentences in human languages as input instructions, the input signal of the action recognizer (i.e., the skeleton sequence) is not in a \"sentence format\" that is \"friendly\" to large language models. To handle this issue, given a large language model initialized with its pre-trained weights, one potential way is to adjust the weights of the full model to fit the non-linguistic input action signals. Recently, several works [25, 38] have been proposed to fine-tune large language models and adjust their pre-trained weights to handle non-language tasks. However, as shown in the previous study [64], the adjustment of the large language model's pre-trained weights can hurt its generalizability and lead to the loss of its pre-learned rich knowledge. This is clearly undesirable here, as we hope the large language model to be knowledgeable in order for it to be an accurate action recognizer." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 315, + 288, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 506 + ], + "type": "text", + "content": "Taking this into account, in this work, we aim to harness the large language model as an action recognizer, and at the same time keep its pre-trained weights untouched to preserve its pre-learned rich knowledge. Specifically, we propose a novel action recognition framework named Large Language Model as an Action Recognizer (LLM-AR). As shown in Fig. 1, to perform action recognition, our LLM-AR framework first involves a linguistic projection process to project the input action signal into its \"sentence format\", that can be \"friendly\" and compatible to the large language model pre-trained over human sentences. Then the large language model takes in this \"sentence format\" of the action signal together with a human sentence as the instruction to predict the name of the action. In the rest of this work, for simplicity, we call the action signal in its \"sentence format\" an \"action sentence\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 628 + ], + "type": "text", + "content": "In our framework, to project the action signals into \"action sentences\" (i.e., performing linguistic projection), we incorporate several designs. Basically, to perform such a process, we observe that, every sentence in a language generally can be regarded as a sequence consisting of discrete word tokens. Inspired by both this observation and previous works [75, 77], in LLM-AR, we first learn an action-based vector quantized variational autoencoder (VQ-VAE) model to project each action signal into a sequence of discrete tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "However, while we can basically represent each \"action sentence\" by a sequence of discrete tokens, it is not necessary that any discrete token sequence can represent an \"action sentence\" well. In fact, to enable a large language model to become an accurate action recognizer based on the \"action sentences\", besides being discrete sequences of tokens, ideally, the \"action sentences\" also need to ful-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 203 + ], + "type": "text", + "content": "fill some requirements as discussed below. Firstly, since large language models are generally pre-trained over corpus consisting of sentences in human languages, to facilitate large language models in taking \"action sentences\" as instructions together with sentences in human languages, these \"action sentences\" should be \"like\" sentences in human languages so as to be more friendly to the large language model. On the other hand, \"action sentences\" should still maintain good representations of their original action signals in order for the large language model to accurately perform action recognition based on them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 204, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 546, + 407 + ], + "type": "text", + "content": "To better meet the above requirements in the linguistic projection process, we further incorporate our framework with two designs below. Firstly, to make the \"action sentences\" more \"like\" sentences in human languages, we get inspiration from previous linguistic and natural language processing works [27, 47, 48, 55, 82] which show that, languages as the communication tools of human beings often contain human inductive biases. Thus, we propose to incorporate our framework with a learning strategy to regularize the projected \"action sentences\" to follow human inductive biases as well. Secondly, to make the projected \"action sentence\" a good representation of its original action signal, motivated by the human skeleton's tree-like nature [63] and hyperbolic space's superior ability in representing tree structures [18], and inspired by previous works [33, 51], we further incorporate our action-based VQ-VAE model with a hyperbolic codebook." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 502 + ], + "type": "text", + "content": "Once we finish the learning of the linguistic projection process, we perform low-rank adaptation (LoRA) [23] on the large language model to let the model understand the projected \"action sentences\". Note that since the pre-trained weights of the large language model are untouched throughout the LoRA process, we can preserve the pre-learned rich knowledge in the large language model and thus utilize the large language model conveniently in our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 503, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 609 + ], + "type": "text", + "content": "The contributions of our work are as follows. 1) We propose a novel action recognition framework named LLM-AR. To the best of our knowledge, this work is the first exploration on treating a large language model with its pretrained weights untouched as an action recognizer. 2) We introduce several designs in our framework to facilitate the linguistic projection process and produce the \"action sentences\". 3) LLM-AR achieves state-of-the-art performance on the evaluated benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 621, + 392, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 392, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 392, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 689 + ], + "type": "text", + "content": "Human Action Recognition. For tackling human action recognition, most of the existing methods can be roughly categorized into two groups: RGB-based methods [58, 60] and skeleton-based methods [5, 6, 53, 54, 70]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "As skeleton sequences can represent human behaviors in a succinct yet informative manner, skeleton-based ac" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18396" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "tion recognition has received lots of research attention [3-7, 12, 15, 17, 21, 28, 32, 39, 40, 53, 54, 59, 62, 65-68, 70, 73, 76, 80]. In the early days, different works have been proposed to use different network architectures to handle skeleton-based action recognition. Liu et al. [39] proposed to perform skeleton-based action recognition through a spatial-temporal LSTM architecture. Ke et al. [28] proposed to transform skeleton sequences into grey images and process these images through a CNN network. As time passed, GCN tends to be a popular network architecture [5, 7, 8, 34, 54, 65, 67, 70]. Yan et al. [70] proposed ST-GCN that made the first attempt in performing skeleton-based action recognition using a GCN architecture. After that, various different GCN-based methods have been further proposed, such as AS-GCN [34], 2s-AGCN [54], Shift-GCN [7], and CTR-GCN [5]. More recently, a number of works [15, 59, 76, 80] further proposed to perform skeleton-based action recognition via training a model with a large transformer architecture in an end-to-end manner. Some of such methods include 3Mformer [59] and UPS [15]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 289, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 289, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 289, + 467 + ], + "type": "text", + "content": "Different from these approaches, in this work, from a novel perspective, we investigate how to instruct the large language model to perform skeleton-based action recognition while keeping its pre-trained weights untouched to preserve its pre-learned rich knowledge. To achieve this, we involve our framework with a novel linguistic projection process to project each input action signal into an \"action sentence\". Specifically, we incorporate several designs in the linguistic projection process to make the projected \"action sentences\" more \"like\" sentences of human languages so as to make them more friendly to large language models, and meanwhile keep the \"action sentence\" a good representation of the input action signal." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 467, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 289, + 612 + ], + "type": "text", + "content": "Large Language Models. Recently, a variety of large language models have been proposed, such as GPT [2] and LLaMA [56]. Pre-trained over a tremendous number of word tokens, these large language models with large model architectures have been shown to contain very rich knowledge [46, 50]. Consequently, large language models have been extensively explored in various tasks [14, 16, 19, 22, 26, 77, 79], such as language translation [26], essay generation [22], and code interpretation [14]. In this work, we design a novel framework treating the large language model as a human action recognizer leveraging our designed linguistic projection process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 621, + 153, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 153, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 153, + 635 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "Given a skeleton sequence as the input action signal, the goal of action recognition is to predict its corresponding action class. Recently, holding large model architectures and containing very rich knowledge, pre-trained large language models have been shown to be powerful in handling sentences of human languages, and thus have become useful" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "tools in many natural language processing tasks. Inspired by this, in this work, we aim to leverage the large language model as an effective action recognizer. To achieve this, we propose a novel framework LLM-AR. Specifically, LLM-AR first performs a linguistic projection process to project the input action signal (i.e., the skeleton sequence) into an \"action sentence\". After that, LLM-AR passes the \"action sentence\" into the large language model to derive the corresponding human action. Below, we first describe how LLM-AR performs the linguistic projection process, and then introduce the overall training and testing scheme of LLM-AR." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 213, + 427, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 427, + 226 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 427, + 226 + ], + "type": "text", + "content": "3.1. Linguistic Projection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 411 + ], + "type": "text", + "content": "In our framework, to enable the large language model to perform action recognition, we first perform a linguistic projection process to project each input action signal into an \"action sentence\". To learn to perform such a projection, with the observation that each sentence in human languages is essentially a sequence of discrete word tokens, motivate by previous works [75, 77], we first learn an action-based VQ-VAE model to project each input action signal into a sequence of discrete tokens. We then further (1) involve the above learning process with a human-inductive-biases-guided learning strategy to make the projected \"action sentences\" more \"like\" sentences in human languages, and (2) incorporate the action-based VQ-VAE model with a hyperbolic codebook to keep \"action sentences\" good representations of the action signals." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 411, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 483 + ], + "type": "text", + "content": "Action-based VQ-VAE Model. VQ-VAE models [51, 57, 74] have been popularly used in converting an image into a discrete token sequence. Inspired by this, in our framework, to convert the input action signal into a sequence of discrete tokens, we first learn an action-based VQ-VAE model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": "Specifically, similar to the architecture of previous VQ-VAE models [57, 75, 77], our action-based VQ-VAE model involves an encoder " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": ", a decoder " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": ", and a codebook " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " learnable tokens (i.e., " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "C = \\{c_u\\}_{u=1}^U" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "c_u \\in \\mathbb{R}^{d_u}" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": "). We here keep " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "d_u" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " to be the same as the dimension of the word token in the used large language model, and keep " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " an even number. Among the aforementioned three components of the model, given an input action signal " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "s_{1:V}" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " represents the action length, the encoder " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " first encodes the action signal through 1D convolutions over the time dimension into a sequence of latent features " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "f_{1:W}" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "f_w \\in \\mathbb{R}^{d_u}" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " represents the length of the latent feature sequence. After this, to discretize the latent features " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "f_{1:W}" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": ", a quantization operation is performed to replace each feature " + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 304, + 484, + 547, + 664 + ], + "type": "text", + "content": " with its nearest token in the codebook as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 366, + 665, + 545, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 665, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 366, + 665, + 545, + 686 + ], + "type": "interline_equation", + "content": "f _ {w} ^ {d} = \\underset {c _ {u} \\in C} {\\operatorname {a r g m i n}} \\left(\\operatorname {d i s t} \\left(f _ {w}, c _ {u}\\right)\\right) \\tag {1}", + "image_path": "5a5a191bc18545e9c7ae135f29ded06ab16ba99af9f33f4a09cca17890785b98.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_w^d" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " is the discrete version of " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "inline_equation", + "content": "dist(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 689, + 547, + 714 + ], + "type": "text", + "content": " represents a distance function. Finally, the decoder aims to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18397" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "recover the original action signal " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "s_{1:V}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " from the sequence of discrete latent features (tokens) " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "f_{1:W}^{d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ". Through the above process of encoding, quantization, and decoding, we can project an action signal into a sequence of informative discrete tokens " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "f_{1:W}^{d}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": ". The more detailed architecture of this action-based VQ-VAE model is provided in supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 337 + ], + "type": "text", + "content": "Human-inductive-biases-guided Learning Strategy. Above we learn to project each input action signal into a sequence of discrete tokens. To enable the learned token sequences to be more friendly to the large language model, here we aim to make these sequences more \"like\" sentences in human languages. To achieve this, we get inspiration from massive existing studies [27, 47, 48, 55, 82] which show that, human languages as tools created for human communication naturally contain human inductive biases. Besides, it is also further shown by the previous study [47] that, inputs following human inductive biases are more friendly to large language models. Taking these into consideration, we aim to optimize the set of learned token sequences to also follow human inductive biases, so that these sequences can be more \"like\" human sentences and thus become more friendly to be used by large language models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "type": "text", + "content": "Below, we first introduce the human inductive biases that are generally recognized as being present in human languages. Such biases include: (a) a human language naturally follows the Zipf's law [47, 48, 82], and (b) a human language is generally context-sensitive [27, 47, 55]. After introducing these human inductive biases, we then describe our proposed human-inductive-biases-guided learning strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": "As for the Zipf's law in (a), during daily communications of human beings, there naturally exist some words that are used more commonly and some words more rarely. Zipf's law intuitively represents this imbalanced usage frequency of word tokens in human languages. Specifically, Zipf's law states that, in a human language, the " + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 434, + 287, + 518 + ], + "type": "text", + "content": "-th most commonly used word has its usage frequency roughly proportional to:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 147, + 519, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 519, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 147, + 519, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\frac {1}{(i + \\beta) ^ {\\alpha}} \\tag {2}", + "image_path": "ca2f05b12ef9dd63df3766995f8e533f3342114f2a406f63632744f732efae29.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\alpha \\approx 1" + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\beta \\approx 2.7" + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "content": " [82]. The reason why Zipf's law consistently appears across different human languages is also theoretically analyzed by various works [9, 43] from different perspectives, such as from the evolution of human communications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "With respect to the context-sensitivity of human languages in (b), intuitively, the context-sensitivity refers to the inductive bias that when people formulate their sentences, they often do not use each word token independently. Instead, their usage of different tokens in formulating a sentence is often correlated. Note that, while human languages are generally believed by linguists to be context-sensitive, how to explicitly represent context-sensitivity remains a difficult problem. Here, inspired by [47] in its way of repre" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "senting context-sensitivity, we represent this human inductive bias as follows. Specifically, given the codebook " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " tokens where " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " is an even number, during initializing " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", we first randomly split the codebook into two halves (i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\{1, \\dots, \\frac{U}{2}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\{1 + \\frac{U}{2}, \\dots, U\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "). Next, we regard each token " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "c_{u}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " in the first half (i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "u \\in \\{1, \\dots, \\frac{U}{2}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ") and the token " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "c_{u + \\frac{U}{2}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " in the second half to be a pair of correlated tokens. As shown in [47], such a pairing mechanism is a simple yet very effective way of representing the context-sensitivity bias in human languages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": "To optimize the set of learned token sequences (i.e., \"action sentences\") " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^d\\}" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": " to also follow the above biases like human languages, we then aim to (a) regularize the set of \"action sentences\" to follow Zipf's law, as well as (b) regularize each \"action sentence\" to be formulated using more correlated word tokens so that the formulated \"action sentences\" can better follow the context-sensitivity bias. To achieve this, we design a human-inductive-biases-guided learning strategy that consists of the following steps. (1) Given a batch of action signals " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\{s_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": " with batch size " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": ", we first encode these signals through the encoder " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": " of the action-based VQ-VAE model to get their corresponding latent features " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": ". (2) After that, during discretizing the latent features using tokens " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "\\{c_u\\}_{u=1}^U" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": " in the codebook " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": ", we measure the token usage of each sequence of latent features " + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "inline_equation", + "content": "f_{1:W}^b" + }, + { + "bbox": [ + 304, + 194, + 546, + 396 + ], + "type": "text", + "content": " in a differentiable way via the Gumbel Softmax trick [24] as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 398, + 555, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 398, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 398, + 555, + 441 + ], + "type": "interline_equation", + "content": "t ^ {b} = \\sum_ {w = 1} ^ {W} d _ {w} ^ {b}, \\text {w h e r e} d _ {w} ^ {b} = \\text {G u m b l e - S o f t m a x} (- \\operatorname {d i s t} \\left(f _ {w} ^ {b}, c _ {u}\\right)) \\tag {3}", + "image_path": "f46d948e3eca789d5df0b9f998468e39b99770992a8dd8ecaa3ccaf903f27c1d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": "where Gumble_Softmax(·) is the Gumble Softmax trick, and " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "d_w^b" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " is the one-hot vector with length " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " as the output of the Gumble Softmax trick. Besides, " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "t^b" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " is a vector with length " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": ", and the value of the " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "t^b" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " represents the number of times token " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "c_u" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " is used in discretizing " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "f_{1:W}^b" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " and formulating its corresponding \"action sentence\". (3) Next, denoting " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "D_{Zipf}" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " the Zipf distribution with " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\beta = 2.7" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": ", we regularize the set of \"action sentences\" to follow Zipf's law through " + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "inline_equation", + "content": "L_{Zipf}" + }, + { + "bbox": [ + 304, + 441, + 545, + 548 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 551, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 553, + 586 + ], + "type": "interline_equation", + "content": "L _ {Z i p f} = J S \\left(D _ {f r e q} \\| D _ {Z i p f}\\right), \\text {w h e r e} D _ {f r e q} = \\frac {\\operatorname {s o r t} \\left(\\sum_ {b = 1} ^ {B} t ^ {b}\\right)}{B \\times W} \\tag {4}", + "image_path": "d15da9a11df705707ccbf73ce37ed55aa6f9228aff7cf931b8f4fb037750441d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "inline_equation", + "content": "sort(\\cdot)" + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "content": " is the sorting operation used to order tokens based on their usage frequency, " + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "inline_equation", + "content": "D_{freq}" + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "content": " is the distribution representing the token usage frequency, and " + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "inline_equation", + "content": "JS(\\cdot \\| \\cdot)" + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "content": " represents the JS divergence between two distributions. (4) At the same time, we encourage the \"action sentences\" to follow the context-sensitivity bias and use more correlated tokens via " + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "inline_equation", + "content": "L_{context}" + }, + { + "bbox": [ + 304, + 587, + 545, + 670 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 672, + 545, + 698 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 672, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 353, + 672, + 545, + 698 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o n t e x t}} = 1 - \\frac {\\sum_ {b = 1} ^ {B} \\operatorname {C o r r} \\left(t ^ {b}\\right)}{B \\times W} \\tag {5}", + "image_path": "e165a711a1620ef900f0858fb9b7cd266b6f47884e692ed8bb2042153f1ad99f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "inline_equation", + "content": "Corr(t^{b})" + }, + { + "bbox": [ + 306, + 701, + 546, + 714 + ], + "type": "text", + "content": " leverages min pooling over every pair of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18398" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 541, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 541, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 541, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 541, + 178 + ], + "type": "image", + "image_path": "3de685c669892722039fa7e1470f99a549983476514df4db0415f7f3c664734b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": "Figure 2. Overview of the action-based VQ-VAE model with the hyperbolic codebook " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " incorporated. Given a batch of input action signals " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\{s_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": ", to optimize the action-based VQ-VAE model, " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\{s_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " are first fed to the encoder " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " to get the corresponding latent features " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": ". Next, to leverage the hyperbolic codebook " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " that can serve as a good representation of the tree-like human skeletons to perform quantization, " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " are projected into the hyperbolic space via the process of E-to-H projection. After that, the quantization is performed in the hyperbolic space using " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "dist_{\\mathbb{B}}(\\widetilde{f}_w, c_u)" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": " defined in Eq. 9 as the distance function. Finally, after quantization, the discrete version of the latent features are passed back into the Euclidean space via the process of H-to-E projection to reconstruct the input action signals through the decoder " + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 194, + 546, + 273 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": "elements in " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "t^b" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": " (e.g., the pair of " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": "-th element and " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "(u + \\frac{U}{2})" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": "-th element) to calculate the number of correlated tokens used in discretizing " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "f_{1:W}^b" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": " (more details about this method to measure " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "Corr(t^b)" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": " are provided in supplementary). Note that via minimizing " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "L_{\\text{context}}" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": ", we encourage more correlated tokens to be used in formulating \"action sentences\". (5) Finally, we formulate the loss function " + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "L_{\\text{human}}" + }, + { + "bbox": [ + 46, + 282, + 287, + 379 + ], + "type": "text", + "content": " that we use to perform our human-inductive-biases-guided strategy as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 381, + 287, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 381, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 102, + 381, + 287, + 395 + ], + "type": "interline_equation", + "content": "L _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} \\tag {6}", + "image_path": "1176971704766d5b2c370c8ac6057f525578e3db40f885d9b5e44f9dfc6a6fe0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": "Via incorporating the above learning strategy into the learning process of the action-based VQ-VAE model, we can regularize the \"action sentences\" to better follow the human inductive biases that are present in human languages and thus make them more \"like\" sentences in human languages." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": "Above we inject human inductive biases into \"action sentences\". Here, we notice that, besides inductive biases that can be explicitly described, human languages can also present other implicit characteristics. Thus, to enable \"action sentences\" to be more friendly to large language models which are generally pre-trained over various human languages, we aim to further incorporate implicit characteristics into \"action sentences\" as well. Specifically, since large language models such as LLaMA typically use word tokens " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_{LLM}\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " stored in its first layer to formulate sentences in human languages, we here further align (1) the tokens " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_u\\}_{u = 1}^U" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " in codebook " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " that are used to formulate the \"action sentences\" with (2) the word tokens " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_{LLM}\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " used in the large language model. To achieve such an alignment, we leverage Maximum Mean Discrepancy (MMD) [20] as an effective feature alignment technique to measure the discrepancy between " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_u\\}_{u = 1}^U" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_{LLM}\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": ". Specifically, the less MMD " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "(\\{c_u\\}_{u = 1}^U,\\{c_{LLM}\\})" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " is, " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_u\\}_{u = 1}^U" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{c_{LLM}\\}" + }, + { + "bbox": [ + 46, + 474, + 287, + 714 + ], + "type": "text", + "content": " can be regarded as more aligned. We then incorporate this alignment procedure into our proposed strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 283, + 443, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 443, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 443, + 295 + ], + "type": "text", + "content": "via rewriting " + }, + { + "bbox": [ + 306, + 283, + 443, + 295 + ], + "type": "inline_equation", + "content": "L_{human}" + }, + { + "bbox": [ + 306, + 283, + 443, + 295 + ], + "type": "text", + "content": " in Eq. 6 as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 297, + 545, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 545, + 322 + ], + "type": "interline_equation", + "content": "L _ {h u m a n} = L _ {Z i p f} + L _ {\\text {c o n t e x t}} + \\mathrm {M M D} \\left(\\left\\{c _ {u} \\right\\} _ {u = 1} ^ {U}, \\left\\{c _ {L L M} \\right\\}\\right) \\tag {7}", + "image_path": "785371e092e6ad59880f39f43162a97c043a29f54b5272fdcade1a334789078e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 322, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 370 + ], + "type": "text", + "content": "By incorporating this alignment procedure into the previously mentioned learning strategy, we can then lead the formulated \"action sentences\" " + }, + { + "bbox": [ + 304, + 322, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^{d}\\}" + }, + { + "bbox": [ + 304, + 322, + 545, + 370 + ], + "type": "text", + "content": " to be more friendly to be used by large language models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": "Hyperbolic Codebook. Besides making the \"action sentences\" more friendly to the large language model, we also aim to keep them good representations of the original input action signals (i.e., the skeleton sequences). Motivated by the tree-like nature of human skeletons [63], we aim to make the word tokens in the \"action sentences\" to well represent such a structure. To achieve this, inspired by the superior capability of the hyperbolic space in embedding tree structures [18] and motivated by previous VQ-VAE works [33, 51], we here further incorporate our action-based VQ-VAE model with a hyperbolic codebook utilizing the Poincaré ball model [18, 45]. The Poincaré ball model is an isometric model that can represent the hyperbolic space. Formally, the " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": "-dimensional Poincaré ball model is defined as " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "(\\mathbb{B}_c^n,g_{\\mathbf{x}}^c)" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_c^n = \\{\\mathbf{x}\\in \\mathbb{R}^n:c\\| \\mathbf{x}\\| < 1\\}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "g_{\\mathbf{x}}^{c} = (\\gamma_{\\mathbf{x}}^{c})^{2}I_{n}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " is the Riemannian metric tensor, " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "\\gamma_{\\mathbf{x}}^{c} = \\frac{2}{1 - c\\|\\mathbf{x}\\|^2}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " is the conformal factor, " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "I_{n}" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " is the Euclidean metric tensor, and " + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 373, + 546, + 589 + ], + "type": "text", + "content": " is the curvature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": "Denoting the hyperbolic codebook " + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "inline_equation", + "content": "C_H = \\{c_u\\}_{u=1}^U" + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "inline_equation", + "content": "c_u \\in \\mathbb{B}_c^{d_u}" + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": ", we make the following three changes to incorporate " + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": " into our action-based VQ-VAE model. We also illustrate where these changes take place in our action-based VQ-VAE model in Fig. 2. (1) Euclidean-to-Hyperbolic (E-to-H) Projection. Given a latent feature " + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": " which is originally in the Euclidean space, to convert it into a discrete token using a hyperbolic codebook, we first need to project " + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 304, + 593, + 547, + 714 + ], + "type": "text", + "content": " into the hyperbolic space. Specifically, following [18], we perform such a projection through the ex" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18399" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 154, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 154, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 154, + 84 + ], + "type": "text", + "content": "ponential map " + }, + { + "bbox": [ + 47, + 72, + 154, + 84 + ], + "type": "inline_equation", + "content": "\\exp_0^c (\\cdot)" + }, + { + "bbox": [ + 47, + 72, + 154, + 84 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 87, + 287, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 87, + 287, + 112 + ], + "spans": [ + { + "bbox": [ + 78, + 87, + 287, + 112 + ], + "type": "interline_equation", + "content": "\\widetilde {f} _ {w} = \\exp_ {\\mathbf {0}} ^ {c} \\left(f _ {w}\\right) = \\tanh (\\sqrt {c} \\| f _ {w} \\|) \\frac {f _ {w}}{\\sqrt {c} \\| f _ {w} \\|} \\tag {8}", + "image_path": "74366593815b74a5a7f9a3b384549a6495d3597b4597b105d4491ec95ef25872.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " is the projected feature of " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "f_w" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " in the hyperbolic space. (2) Hyperbolic Distance Calculation. Given the projected feature " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": ", to perform quantization of " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " based on the tokens in " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": ", we need a distance function defined in the hyperbolic space. Here, for simplicity, we use the popularly used geodesic/induced distance [18, 29]. Denote this distance as " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\text{dist}_{\\mathbb{B}}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\text{dist}_{\\mathbb{B}}(\\widetilde{f}_w, c_u)" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "inline_equation", + "content": "c_u" + }, + { + "bbox": [ + 46, + 115, + 287, + 213 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 215, + 286, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 215, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 215, + 286, + 255 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} _ {\\mathbb {B}} \\left(\\widetilde {f} _ {w}, c _ {u}\\right) = \\operatorname {a r c c o s h} \\left(1 + 2 \\frac {\\| \\widetilde {f} _ {w} - c _ {u} \\| ^ {2}}{\\left(1 - \\| \\widetilde {f} _ {w} \\| ^ {2}\\right) \\left(1 - \\| c _ {u} \\| ^ {2}\\right)}\\right) \\tag {9}", + "image_path": "5cc921c036a2925791ac54ef729c31111f5c9a4c4d8ab38b2990236cf946b131.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "content": "(3) Hyperbolic-to-Euclidean (H-to-E) Projection. After quantization, we need to project " + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w^d" + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "content": " (the discrete version of " + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_w" + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "content": ") back into the Euclidean space in order for it to be passed to the decoder " + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "content": ". To achieve this, we follow [18] to leverage the logarithmic map " + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "inline_equation", + "content": "\\log_0^c (\\cdot)" + }, + { + "bbox": [ + 47, + 255, + 287, + 316 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 318, + 287, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 318, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 64, + 318, + 287, + 349 + ], + "type": "interline_equation", + "content": "f _ {w} ^ {d} = \\log_ {\\mathbf {0}} ^ {c} (\\widetilde {f} _ {w} ^ {d}) = \\operatorname {a r c t a n h} (\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|) \\frac {\\widetilde {f} _ {w} ^ {d}}{\\sqrt {c} \\| \\widetilde {f} _ {w} ^ {d} \\|} \\tag {10}", + "image_path": "7afc20f279e74ce1520b166dcc7048af61b1549326c39fdbdea8170857ae49a0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "inline_equation", + "content": "f_{w}^{d}" + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": " represents the discrete version of the feature projected back in the Euclidean space. With the above changes made, we can seamlessly involve the hyperbolic codebook " + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": " into our action-based VQ-VAE model, and make the projected \"action sentence\" a better representation of the input action signal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 430, + 205, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 205, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 205, + 443 + ], + "type": "text", + "content": "3.2. Overall Training and Testing" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 449, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 496 + ], + "type": "text", + "content": "Above we describe how we perform the linguistic projection process in our framework to project each input action signal into an \"action sentence\". Below, we introduce the overall training and testing scheme of our framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": "Training. The training process of our framework consists of the following two stages: (1) optimizing the action-based VQ-VAE model incorporated with the hyperbolic codebook " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": " to acquire the ability to project each action signal " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "s_{1:V}" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": " into its corresponding \"action sentence\" " + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\widetilde{f}_{1:W}^d" + }, + { + "bbox": [ + 46, + 497, + 287, + 593 + ], + "type": "text", + "content": "; and (2) perform low-rank adaptation (LoRA) [23] on the large language model for it to better understand the projected \"action sentences\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "At the first stage, to optimize the action-based VQ-VAE model, given a batch of action signals, we first follow previous VQ-VAE works [57, 75, 77] to use the following loss functions, including the reconstruction loss, the embedding loss, and the commitment loss. Among them, the reconstruction loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{re}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is designed to regularize the difference between the input action signal and the action signal reconstructed from the decoder " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": ", the embedding loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{embed}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is designed to learn the tokens in the codebook, and the commitment loss " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{commit}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " is designed to encourage each" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "of the encoded latent features to stay close to its discrete version. Formally speaking, denoting " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{s_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " a batch of input action signals, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{\\hat{s}_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " the corresponding batch of reconstructed action signals, " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " the latent features encoded from " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{s_{1:V}^b\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\{f_{1:W}^{d,b}\\}_{b=1}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " the discrete versions of these latent features, the above three loss functions can be written as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 157, + 486, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 157, + 486, + 172 + ], + "spans": [ + { + "bbox": [ + 325, + 157, + 486, + 172 + ], + "type": "interline_equation", + "content": "L _ {r e} = L _ {1} ^ {s m o o t h} (\\{s _ {1: V} ^ {b} \\} _ {b = 1} ^ {B}, \\{\\hat {s} _ {1: V} ^ {b} \\} _ {b = 1} ^ {B})", + "image_path": "2a49cb6729140e7f8e614cde7572df426cd8d964ec87a515e099a37148b12e1c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 173, + 545, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 173, + 545, + 189 + ], + "spans": [ + { + "bbox": [ + 325, + 173, + 545, + 189 + ], + "type": "interline_equation", + "content": "L _ {e m b e d} = \\left\\| s g \\left(\\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B}\\right) - \\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B} \\right\\| ^ {2} \\tag {11}", + "image_path": "54be1ac4c20822ddfe0c371f41c4adc053a52bdffb4d82f2809f66d281dcbbb9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 190, + 509, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 190, + 509, + 206 + ], + "spans": [ + { + "bbox": [ + 325, + 190, + 509, + 206 + ], + "type": "interline_equation", + "content": "L _ {c o m m i t} = \\left\\| \\left\\{f _ {1: W} ^ {b} \\right\\} _ {b = 1} ^ {B} - s g \\left(\\left\\{f _ {1: W} ^ {d, b} \\right\\} _ {b = 1} ^ {B}\\right) \\right\\| ^ {2}", + "image_path": "22263962b2f7c8a33f06769a0d9f9a2a28f2f38bbf6c14d1f106802fbe093f45.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "L_{1}^{smooth}(\\cdot)" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " represents the L1 smooth loss and " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "sg(\\cdot)" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " is the stop-gradient operation that is used to prevent the gradient from flowing through its operand. In addition, to make the learned \"action sentences\" more like sentences in human languages, we also incorporate the above three loss functions with " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "L_{human}" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " defined in Eq. 7. Overall, we can write the total loss " + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "inline_equation", + "content": "L_{total}" + }, + { + "bbox": [ + 304, + 209, + 545, + 293 + ], + "type": "text", + "content": " for the first training stage as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 295, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 295, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 311, + 295, + 545, + 308 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = L _ {r e} + L _ {\\text {e m b e d}} + \\omega_ {1} L _ {\\text {c o m m i t}} + \\omega_ {2} L _ {\\text {h u m a n}} \\tag {12}", + "image_path": "85cde085763eaf16eec883623618cbe253f3c85aaef8ce56fa5a32a1d744b30e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "inline_equation", + "content": "\\omega_{1}" + }, + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "inline_equation", + "content": "\\omega_{2}" + }, + { + "bbox": [ + 305, + 311, + 504, + 323 + ], + "type": "text", + "content": " are weighting hyperparameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "text", + "content": "We then perform LoRA [23] on the large language model in the second stage in order for it to better understand the \"action sentences\" while keeping the pre-trained weights of the model untouched. Specifically, for each training sample consisting of an input action signal and its corresponding ground-truth action, we perform the following steps. (1) We first project the action signal into an \"action sentence\" of discrete tokens through the action-based VQ-VAE model learned in the first stage. (2) We then instruct the large language model to act as an action recognizer via a simple instruction as: \"Given a sequence of action tokens [tokens], please predict the corresponding action.\", where [tokens] represent the word tokens of the \"action sentence\" derived in step (1). (3) Finally, during LoRA process, we pass the above instruction into the large language model and encourage the similarity between the tokens " + }, + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "inline_equation", + "content": "t_p" + }, + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "text", + "content": " predicted by the large language model and the tokens " + }, + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "inline_equation", + "content": "t_g" + }, + { + "bbox": [ + 304, + 323, + 545, + 550 + ], + "type": "text", + "content": " representing the ground-truth action as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 381, + 554, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 554, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 381, + 554, + 545, + 567 + ], + "type": "interline_equation", + "content": "L _ {L o R A} = L _ {c e} \\left(t _ {p}, t _ {g}\\right) \\tag {13}", + "image_path": "c72bd674c84d1cc5bde024b27d07cac34ea1d544bbbabbcf77da881dddde3017.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 569, + 503, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 569, + 503, + 581 + ], + "spans": [ + { + "bbox": [ + 305, + 569, + 503, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 569, + 503, + 581 + ], + "type": "inline_equation", + "content": "L_{ce}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 305, + 569, + 503, + 581 + ], + "type": "text", + "content": " represents the cross-entropy loss." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 642 + ], + "type": "text", + "content": "Testing. During testing, for each testing action signal, we first use the learned action-based VQ-VAE model to derive its corresponding \"action sentence\". After that, we use the same instruction as in step (2) above to instruct the large language model to predict the corresponding action." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 656, + 386, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 386, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 386, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the efficacy of our framework, we conduct experiments on 4 datasets including NTU RGB+D, NTU RGB+D 120, Toyota Smarthome, and UAV-Human." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18400" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 110, + 83 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 149 + ], + "type": "text", + "content": "NTU RGB+D [52] is a large-scale dataset popularly used in human action recognition. It contains around 56k skeleton sequences from 60 activity classes. On this dataset, following [52], we evaluate our method under the Cross-Subject (X-Sub) and Cross-View (X-View) evaluation protocols." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 150, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 150, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 287, + 209 + ], + "type": "text", + "content": "NTU RGB+D 120 [41] is an extension of the NTU RGB+D dataset. It consists of more than 114k skeleton sequences across 120 activity classes. Following [41], we evaluate our method on this dataset under the Cross-Subject (X-Sub) and Cross-Setup (X-Set) evaluation protocols." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 210, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 287, + 269 + ], + "type": "text", + "content": "Toyota Smarthome [10] contains 16,115 video samples over 31 activity classes. On this dataset, we use the skeleton sequences pre-processed by [71] and we follow it to evaluate our method under the Cross-Subject (X-Sub) and two Cross-View (X-View1 & X-View2) evaluation protocols." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 270, + 287, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 329 + ], + "type": "text", + "content": "UAV-Human [35] is a dataset that is captured by unmanned aerial vehicles (UAV). It contains more than 20k skeleton sequences over 155 activity classes, and it is collected from 119 distinct subjects. Following [35], we use 89 subjects for training and 30 subjects for testing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 335, + 180, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 180, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 180, + 347 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": "We conduct our main experiments on Nvidia V100 GPU and we use LLaMA-13B [56] as the large language model. During the training process of the action-based VQ-VAE model, we optimize the model for 300,000 iterations using the AdamW [42] optimizer with an initial learning rate of 2e-4. We set the batch size " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " to 256, the number of tokens " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " to 512, and the curvature " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " of the hyperbolic codebook to 1. Additionally, we set the dimension of each word token " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "d_{u}" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " to the same size (5120) as the word token in LLaMA-13B. Moreover, we set " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " to 0.02 following previous VQ-VAE works [44, 75] and set " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "w_{2}" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " to 0.2. Besides, we set the length " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "(W)" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " of each sequence of latent features to be a quarter of the length " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " of its corresponding input action signal. During the LoRA process, we develop our code based on the Github repo [1], set the number of iterations to 75,000, and use the AdamW optimizer with an initial learning rate of 3e-3. Moreover, we set the batch size to 256, partitioned into micro-batches of 4. Besides, we set the two hyperparameters of the LoRA process, i.e., " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "r_{LoRA}" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_{LoRA}" + }, + { + "bbox": [ + 46, + 354, + 287, + 594 + ], + "type": "text", + "content": ", to 64 and 16 respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 270, + 612 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-art Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 712 + ], + "type": "text", + "content": "On NTU RGB+D and NTU RGB+D 120 datasets, following the experimental setting of recent works [12, 80], we only use the joint modality of human skeletons during our experiments. We report the results in Tab. 1. As shown, compared to existing skeleton-based action recognition methods, our method consistently achieves the best performance across all the evaluation protocols. This demonstrates the effectiveness of our method. Besides, we also" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 95, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 91 + ], + "type": "text", + "content": "Table 1. Performance comparison on the NTU RGB+D and NTU RGB+D 120 datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 95, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 309, + 95, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 309, + 95, + 545, + 243 + ], + "type": "table", + "html": "
MethodNTU RGB+DNTU RGB+D 120
X-SubX-ViewX-SubX-Set
ST-GCN [70]85.792.482.184.5
Shift-GCN [7]87.895.180.983.2
InfoGCN [8]89.895.285.186.3
PoseC3D [11]93.796.585.989.7
FR-Head [78]90.395.385.587.3
Koopman [62]90.295.285.787.4
GAP [65]90.295.685.587.0
HD-GCN [31]90.695.785.787.3
STC-Net [30]91.096.286.288.0
DSTformer [80]93.097.2--
SkeleTR [12]94.897.787.888.3
Ours95.098.488.791.5
", + "image_path": "2ade71c206487a22e6285d9509a051538dd65fd1d7cda1e7cc7a53c6599f0b5c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 291, + 460, + 344 + ], + "blocks": [ + { + "bbox": [ + 306, + 255, + 460, + 277 + ], + "lines": [ + { + "bbox": [ + 306, + 255, + 460, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 460, + 277 + ], + "type": "text", + "content": "Table 2. Performance comparison on the Toyota Smarthome dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 291, + 460, + 344 + ], + "lines": [ + { + "bbox": [ + 309, + 291, + 460, + 344 + ], + "spans": [ + { + "bbox": [ + 309, + 291, + 460, + 344 + ], + "type": "table", + "html": "
MethodX-SubX-View1X-View2
2S-AGCN [54]58.832.257.9
SSTA-PRS [71]62.122.854.0
UNIK [72]62.133.463.6
ML-STGNet [81]64.629.963.5
Ours67.036.166.6
", + "image_path": "a1a709624a10ff930ff03fc74957dad4361ca248dee558d4768f44c4caf0329c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 467, + 292, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 465, + 255, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 465, + 255, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 465, + 255, + 545, + 286 + ], + "type": "text", + "content": "Table 3. Performance comparison on the UAV-Human dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 467, + 292, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 467, + 292, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 467, + 292, + 545, + 344 + ], + "type": "table", + "html": "
MethodX-Sub
ST-GCN [69]30.3
2S-AGCN [54]34.8
Shift-GCN [7]38.0
ACFL [61]44.2
Ours46.3
", + "image_path": "c5cc1c973196709f87a49b4c863fb32ac3ec9de61b859f2e2f48e6d2c401e1dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 413 + ], + "type": "text", + "content": "report results on the Toyota Smarthome dataset in Tab. 2, and on the UAV-Human dataset in Tab. 3. As shown, our method consistently achieves the best performance on these datasets. This further shows the efficacy of our method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 406, + 431 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 485 + ], + "type": "text", + "content": "We conduct extensive ablation experiments on the X-Set protocol of the NTU RGB+D 120 dataset. More ablation studies such as experiments w.r.t. hyperparameters are in supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 487, + 429, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 487, + 429, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 487, + 429, + 510 + ], + "type": "text", + "content": "Impact of the human-inductive-biases-guided" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 510, + 429, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 429, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 429, + 581 + ], + "type": "text", + "content": "learning strategy. In our framework, to lead the formulated \"action sentences\" to be more friendly to large language models, we design a human-inductive-biases" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 438, + 525, + 545, + 579 + ], + "blocks": [ + { + "bbox": [ + 435, + 489, + 545, + 522 + ], + "lines": [ + { + "bbox": [ + 435, + 489, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 435, + 489, + 545, + 522 + ], + "type": "text", + "content": "Table 4. Evaluation on the human-inductive-biases-guided learning strategy." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 438, + 525, + 545, + 579 + ], + "lines": [ + { + "bbox": [ + 438, + 525, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 438, + 525, + 545, + 579 + ], + "type": "table", + "html": "
MethodAccuracy
w/o biases87.6
w/o LZipf89.9
w/o Lcontext89.8
w/o MMD alignment90.3
LLM-AR91.5
", + "image_path": "ec3254ed585e1af9cafdd231dec5e95c9b58c630249bec93c00debbc1d23c661.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": "guided learning strategy consisting of three components (as shown in Eq. 7). To evaluate the efficacy of this strategy, we test four variants. In the first variant (w/o biases), we remove the whole strategy (i.e., all its three components " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{Zipf}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{context}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": ", and MMD alignment) from the learning process. In the second variant (w/o " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{Zipf}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": "), we still utilize the strategy but remove its " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{Zipf}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": " component. Moreover, in the third variant (w/o " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{context}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": "), we remove the " + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{context}" + }, + { + "bbox": [ + 305, + 582, + 545, + 713 + ], + "type": "text", + "content": " component from the strategy, whereas in the fourth variant (w/o MMD alignment), we remove the MMD alignment component. As shown in Tab. 4, compared to" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18401" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "our framework, the performance of the first variant drops significantly. This shows the importance of formulating \"action sentences\" like sentences in human languages. Moreover, our framework also outperforms all the other three variants. This further shows the effectiveness of all the three components of our proposed learning strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 149, + 178, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 178, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 178, + 221 + ], + "type": "text", + "content": "Impact of discretizing latent features into \"action sentences\". In our framework, we discretize the encoded latent features to formulate \"action sentences\" consisting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 184, + 150, + 287, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 184, + 150, + 287, + 183 + ], + "type": "text", + "content": "Table 5. Evaluation on discretizing latent features into \"action sentences\"." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 187, + 186, + 287, + 216 + ], + "blocks": [ + { + "bbox": [ + 187, + 186, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 187, + 186, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 187, + 186, + 287, + 216 + ], + "type": "table", + "html": "
MethodAccuracy
w/o discretization83.4
with discretization91.5
", + "image_path": "2934664c39dc1f943f8f2cc9f7f8f959d756c6be0c918878563708214c580028.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 221, + 289, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 289, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 289, + 354 + ], + "type": "text", + "content": "of discrete word tokens (with discretization). To valid this design, we test a variant. In this variant (w/o discretization), instead of discretizing the encoded latent features into \"action sentences\", we directly pass these continuous features to the intermediate layers of the large language model. As shown in Tab. 5, our framework with discretization outperforms this variant. This shows the advantage of discretizing latent features into \"action sentences\", which are more \"like\" human sentences consisting of discrete word tokens, and thus are more friendly to the large language model pre-trained over human sentences." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": "Impact of the hyperbolic codebook " + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": ". In our framework, we incorporate our action-based VQ-VAE model with a hyperbolic codebook " + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": "). To validate the efficacy of " + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": ", we test a variant " + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "inline_equation", + "content": "(\\mathbf{w} / \\mathbf{o} C_H)" + }, + { + "bbox": [ + 47, + 358, + 201, + 430 + ], + "type": "text", + "content": " in which" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 208, + 360, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 208, + 360, + 287, + 392 + ], + "type": "text", + "content": "Table 6. Evaluation on the hyperbolic codebook " + }, + { + "bbox": [ + 208, + 360, + 287, + 392 + ], + "type": "inline_equation", + "content": "C_H" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "text", + "content": "the codebook is set up in the Euclidean instead of hyperbolic space. As shown in Tab. 6, our framework involving " + }, + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "text", + "content": " performs better than this variant. This shows the efficacy of " + }, + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "inline_equation", + "content": "C_H" + }, + { + "bbox": [ + 46, + 430, + 288, + 503 + ], + "type": "text", + "content": " in the hyperbolic space, which can facilitate the \"action sentences\" in representing the tree-like-structured input action signals better." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 508, + 200, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 200, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 200, + 567 + ], + "type": "text", + "content": "Impact of the LoRA process. In our framework, to make the large language model understand the \"action sentences\" while keeping its pretrained weights untouched to pre" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 206, + 510, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 510, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 206, + 510, + 287, + 533 + ], + "type": "text", + "content": "Table 7. Evaluation on the LoRA process." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 208, + 534, + 287, + 564 + ], + "blocks": [ + { + "bbox": [ + 208, + 534, + 287, + 564 + ], + "lines": [ + { + "bbox": [ + 208, + 534, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 208, + 534, + 287, + 564 + ], + "type": "table", + "html": "
MethodAccuracy
All tuning79.6
LLM-AR91.5
", + "image_path": "edf569d8ac8d30900c59aa0d423088082da44b781c576a1c11397de42c684447.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 568, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 289, + 712 + ], + "type": "text", + "content": "serve its rich pre-learned knowledge, we tune the large language model through a LoRA process. To validate this scheme, here we also test a variant (all tuning) on A100 GPU. In this variant, during tuning, after initializing the large language model with its pre-trained weights, all the parameters of the model will undergo gradient updates. As shown in Tab. 7, our framework achieves much better performance than this variant. This shows the superiority of our framework in choosing to perform tuning via LoRA, which enables the large language model's pre-trained weights to be untouched and maintains its pre-learned rich knowledge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 457, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 457, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 457, + 144 + ], + "type": "text", + "content": "Evaluation on unseen activity classes. In the main experiments, following [31, 52, 70], we evaluate our framework on activity classes that have been seen during training. Here, inspired by that large language" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 467, + 110, + 545, + 140 + ], + "blocks": [ + { + "bbox": [ + 463, + 74, + 545, + 107 + ], + "lines": [ + { + "bbox": [ + 463, + 74, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 463, + 74, + 545, + 107 + ], + "type": "text", + "content": "Table 8. Evaluation on unseen activity classes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 467, + 110, + 545, + 140 + ], + "lines": [ + { + "bbox": [ + 467, + 110, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 467, + 110, + 545, + 140 + ], + "type": "table", + "html": "
MethodAccuracy
All tuning37.7
LLM-AR62.4
", + "image_path": "4f762b878cddc8cd85bf5f9bbff886701fb53c1bd1a1849f255029a72dc0171e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 144, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 546, + 504 + ], + "type": "text", + "content": "models naturally could contain rich knowledge beyond the training activity classes used in our experiments, we are curious, assuming we have a list of testing activity classes unseen during training, can we also use our framework to perform action recognition on these classes? To answer this question, we first build a new evaluation protocol for unseen activity classes based on the NTU RGB+D 120 dataset. Under this new protocol, during each time of evaluation, we randomly select 3 classes to form the unseen class list (i.e., the list of testing classes), and use the remaining classes as the classes seen during training (i.e., the training classes). Besides, we instruct the large language model as: \"Given a sequence of action tokens [tokens], please predict the corresponding action from [list],\" where [tokens] represent the word tokens of the \"action sentence\" and [list] represents the unseen class list. We then perform the above evaluation for five times and report the average performance. As shown in Tab. 8, even testing on activity classes unseen during training, our framework can still achieve a relatively good performance, while the afore-defined all tuning variant performs much worse. This can be analyzed as, large language models could contain rich pre-learned knowledge w.r.t. the list of unseen classes. Thus, our framework that maintains such rich knowledge can still perform promising recognition on these unseen classes, while the all tuning variant that can lose amount of pre-learned knowledge of the large language model would yield a much worse performance. This further shows the advantage of our framework in maintaining the pre-learned rich knowledge of the large language model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 379, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 665 + ], + "type": "text", + "content": "In this paper, we have proposed a novel action recognition framework LLM-AR. In LLM-AR, we treat the large language model as an action recognizer, and instruct the large language model to perform action recognition using its contained rich knowledge. Specifically, to lead the input action signals (i.e., the skeleton sequences) to be more friendly to the large language model, we first propose a linguistic projection process to project each action signal into an \"action sentence\". Moreover, we also introduce several designs to further facilitate this process. Our framework consistently achieves SOTA performance across different benchmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18402" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] Lit-llama. https://github.com/Lightning-AI/ lit-llama.7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 168 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "text", + "content": "[3] Dongqi Cai, Yangyuxuan Kang, Anbang Yao, and Yurong Chen. Ske2grid: Skeleton-to-grid representation learning for action recognition. 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 246 + ], + "type": "text", + "content": "[4] Jinghong Chen, Chong Zhao, Qicong Wang, and Hongying Meng. Hmanet: Hyperbolic manifold aware network for skeleton-based action recognition. IEEE Transactions on Cognitive and Developmental Systems, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 286, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 286, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 286, + 302 + ], + "type": "text", + "content": "[5] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13359-13368, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 303, + 286, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 286, + 368 + ], + "type": "text", + "content": "[6] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 536-553. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 286, + 423 + ], + "type": "text", + "content": "[7] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 183-192, 2020. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "type": "text", + "content": "[8] Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20186-20196, 2022. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 491, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 286, + 523 + ], + "type": "text", + "content": "[9] Bernat Corominas-Murtra, Jordi Fortune, and Ricard V Solé. Emergence of zipf's law in the evolution of communication. Physical Review E, 83(3):036115, 2011. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 579 + ], + "type": "text", + "content": "[10] Srijan Das, Rui Dai, Michal Koperski, Luca Minciullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome: Real-world activities of daily living. In Proceedings of the IEEE/CVF international conference on computer vision, pages 833-842, 2019. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "text", + "content": "[11] Haodong Duan, Yue Zhao, Kai Chen, Dahua Lin, and Bo Dai. Revisiting skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2969-2978, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 689 + ], + "type": "text", + "content": "[12] Haodong Duan, Mingze Xu, Bing Shuai, Davide Modolo, Zhuowen Tu, Joseph Tighe, and Alessandro Bergamo. Skeletr: Towards skeleton-based action recognition in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13634-13644, 2023. 1, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[13] Abassin Sourou Fangbemi, Bin Liu, Neng Hai Yu, and Yanxiang Zhang. Efficient human action recognition interface for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 128 + ], + "type": "text", + "content": "augmented and virtual reality applications based on binary descriptor. In Augmented Reality, Virtual Reality, and Computer Graphics: 5th International Conference, AVR 2018, Otranto, Italy, June 24-27, 2018, Proceedings, Part I 5, pages 252-260. Springer, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Faisal Firdous, Saimul Bashir, Syed Zoofa Rufai, and Sanjeev Kumar. Openai chatgpt as a logical interpreter of code. In 2023 2nd International Conference on Edge Computing and Applications (ICECAA), pages 1192-1197. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[15] Lin Geng Foo, Tianjiao Li, Hossein Rahmani, Qiuhong Ke, and Jun Liu. Unified pose sequence modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13030, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 249 + ], + "type": "text", + "content": "[16] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 303 + ], + "type": "text", + "content": "[17] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 304, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 338 + ], + "type": "text", + "content": "[18] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[19] Jia Gong, Lin Geng Foo, Yixuan He, Hossein Rahmani, and Jun Liu. Llms are good sign language translators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[20] Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alex Smola. A kernel method for the two-sample-problem. Advances in neural information processing systems, 19, 2006. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[21] Ryo Hachiuma, Fumiaki Sato, and Taiki Sekii. Unified keypoint-based action recognition framework via structured keypoint pooling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22962-22971, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 535 + ], + "type": "text", + "content": "[22] Steffen Herbold, Annette Hautli-Janisz, Ute Heuer, Zlata Kikteva, and Alexander Trautsch. Ai, write an essay for me: A large-scale comparison of human-written versus chatgpt-generated essays. arXiv preprint arXiv:2304.14276, 2023. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 536, + 545, + 581 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 612 + ], + "type": "text", + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "text", + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[26] Wenxiang Jiao, Wenxuan Wang, JT Huang, Xing Wang, and ZP Tu. Is chatgpt a good translator? yes with gpt-4 as the engine. arXiv preprint arXiv:2301.08745, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[27] Aravind K Joshi, K Vijay Shanker, and David Weir. The convergence of mildly context-sensitive grammar formalisms. Technical Reports (CIS), page 539, 1990. 2, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18403" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[28] Qiuhong Ke, Mohammed Bennamoun, Senjian An, Ferdous Sohel, and Farid Boussaid. A new representation of skeleton sequences for 3d action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3288-3297, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 183 + ], + "type": "text", + "content": "[29] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6418-6428, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 183, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 288, + 239 + ], + "type": "text", + "content": "[30] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10255-10264, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 294 + ], + "type": "text", + "content": "[31] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10444-10453, 2023. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 327 + ], + "type": "text", + "content": "[32] Zikang Leng, Hyeokhyen Kwon, and Thomas Plötz. On the benefit of generative foundation models for human activity recognition. arXiv preprint arXiv:2310.12085, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "text", + "content": "[33] Lei Li, Tingting Liu, Chengyu Wang, Minghui Qiu, Cen Chen, Ming Gao, and Aoying Zhou. Resizing codebook of vector quantization without retraining. Multimedia Systems, pages 1-14, 2023. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "text", + "content": "[34] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 482 + ], + "type": "text", + "content": "[35] Tianjiao Li, Jun Liu, Wei Zhang, Yun Ni, Wenqian Wang, and Zhiheng Li. Uav-human: A large benchmark for human behavior understanding with unmanned aerial vehicles. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16266-16275, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 536 + ], + "type": "text", + "content": "[36] Weiyao Lin, Ming-Ting Sun, Radha Poovendran, and Zhengyou Zhang. Activity recognition using a combination of category components and local models for video surveillance. IEEE Transactions on Circuits and Systems for Video Technology, 18(8):1128-1139, 2008. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 591 + ], + "type": "text", + "content": "[37] Hong Liu, Qinqin He, and Mengyuan Liu. Human action recognition using adaptive hierarchical depth motion maps and gabor filter. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1432-1436. IEEE, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "text", + "content": "[39] Jun Liu, Amir Shahroudy, Dong Xu, and Gang Wang. Spatio-temporal LSTM with trust gates for 3d human action recognition. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 816-833. Springer, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 713 + ], + "type": "text", + "content": "[40] Jun Liu, Gang Wang, Ping Hu, Ling-Yu Duan, and Alex C Kot. Global context-aware attention LSTM networks for 3d" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1647-1656, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 161 + ], + "type": "text", + "content": "[41] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 183 + ], + "type": "text", + "content": "[42] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 183, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 205 + ], + "type": "text", + "content": "[43] Benoit Mandelbrot. Structure formelle des textes et communication: Deux études par. Word, 10(1):1-27, 1954. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 205, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 259 + ], + "type": "text", + "content": "[44] Evonne Ng, Sanjay Subramanian, Dan Klein, Angjoo Kanazawa, Trevor Darrell, and Shiry Ginosar. Can language models learn to listen? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10083-10093, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 260, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 545, + 294 + ], + "type": "text", + "content": "[45] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 360 + ], + "type": "text", + "content": "[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 403 + ], + "type": "text", + "content": "[47] Isabel Papadimitriou and Dan Jurafsky. Pretrain on just structure: Understanding linguistic inductive biases using transfer learning. arXiv preprint arXiv:2304.13060, 2023. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[48] Steven T Piantadosi. Zipf's word frequency law in natural language: A critical review and future directions. *Psychonomic bulletin & review*, 21:1112–1130, 2014. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[49] Yiwei Qin, Graham Neubig, and Pengfei Liu. Searching for effective multilingual fine-tuning methods: A case study in summarization, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 515 + ], + "type": "text", + "content": "[50] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 515, + 490, + 526 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 490, + 526 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 490, + 526 + ], + "type": "text", + "content": "[51] Acquaviva Sam. Hyperbolic vq-vaes. 2, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 570 + ], + "type": "text", + "content": "[52] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1010-1019, 2016. 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[53] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7912-7921, 2019. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 680 + ], + "type": "text", + "content": "[54] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026-12035, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[55] Stuart M Shieber. Evidence against the context-freeness of natural language. In The Formal complexity of natural language, pages 320-334. Springer, 1985. 2, 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18404" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[56] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[57] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[58] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In Proceedings of the IEEE international conference on computer vision, pages 4041-4049, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[59] Lei Wang and Piotr Koniusz. 3mformer: Multi-order multi-mode transformer for skeletal action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5620-5631, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 308 + ], + "type": "text", + "content": "[60] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 353 + ], + "type": "text", + "content": "[61] Xuanhan Wang, Yan Dai, Lianli Gao, and Jingkuan Song. Skeleton-based action recognition via adaptive cross-form learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1670-1678, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 288, + 410 + ], + "type": "text", + "content": "[62] Xinghan Wang, Xin Xu, and Yadong Mu. Neural koopman pooling: Control-inspired temporal dynamics encoding for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10597-10607, 2023. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 465 + ], + "type": "text", + "content": "[63] Shenghua Wei, Yonghong Song, and Yuanlin Zhang. Human skeleton tree recurrent neural network with joint relative motion feature for skeleton based action recognition. In 2017 IEEE international conference on image processing (ICIP), pages 91-95. IEEE, 2017. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[64] Rongxiang Weng, Wen Sen Cheng, and Min Zhang. G-tuning: Improving generalization of pre-trained language models with generative adversarial network. In *Findings of the Association for Computational Linguistics: ACL 2023*, pages 4747-4755, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 578 + ], + "type": "text", + "content": "[65] Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, and Lei Zhang. Generative action description prompts for skeleton-based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10276-10285, 2023. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[66] Wentian Xin, Qiguang Miao, Yi Liu, Ruyi Liu, Chi-Man Pun, and Cheng Shi. Skeleton mixformer: Multivariate topology representation for skeleton-based action recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2211-2220, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[67] Haojun Xu, Yan Gao, Zheng Hui, Jie Li, and Xinbo Gao. Language knowledge-assisted representation learning for skeleton-based action recognition. arXiv preprint arXiv:2305.12398, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[68] Kailin Xu, Fanfan Ye, Qiaoyong Zhong, and Di Xie. Topology-aware convolutional neural network for efficient skeleton-based action recognition. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "AAAI Conference on Artificial Intelligence, pages 2866-2874, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 161 + ], + "type": "text", + "content": "[69] Hongliang Yan, Yukang Ding, Peihua Li, Qilong Wang, Yong Xu, and Wangmeng Zuo. Mind the class weight bias: Weighted maximum mean discrepancy for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2272-2281, 2017. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[70] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 284 + ], + "type": "text", + "content": "[71] Di Yang, Rui Dai, Yaohui Wang, Rupayan Mallick, Luca Minciullo, Gianpiero Francesca, and Francois Bremond. Selective spatio-temporal aggregation based pose refinement system: Towards understanding human activities in real-world videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2363-2372, 2021. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[72] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Bremond. Unik: A unified framework for real-world skeleton-based action recognition. arXiv preprint arXiv:2107.08580, 2021. 1, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 547, + 385 + ], + "type": "text", + "content": "[73] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Hierarchical consistent contrastive learning for skeleton-based action recognition with growing augmentations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3427-3435, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 441 + ], + "type": "text", + "content": "[74] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18467-18476, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 498 + ], + "type": "text", + "content": "[75] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 553 + ], + "type": "text", + "content": "[76] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 599 + ], + "type": "text", + "content": "[77] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[78] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[79] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation. arXiv preprint arXiv:2311.16926, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[80] Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18405" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 152 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[81] Yisheng Zhu, Hui Shuai, Guangcan Liu, and Qingshan Liu. Multilevel spatial-temporal excited graph network for skeleton-based action recognition. IEEE Transactions on Image Processing, 32:496-508, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 270, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 270, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 270, + 152 + ], + "type": "text", + "content": "[82] G. K. Zipf. The psycho-biology of language, 1935. 2, 4" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18406" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_content_list.json b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ddac7437376d929b5527e27486baf3e8a86c5bc1 --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_content_list.json @@ -0,0 +1,1477 @@ +[ + { + "type": "text", + "text": "LLMs are Good Sign Language Translators", + "text_level": 1, + "bbox": [ + 263, + 130, + 707, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia Gong $^{1\\dagger}$ Lin Geng Foo $^{1\\dagger}$ Yixuan He $^{1\\dagger}$ Hossein Rahmani $^{2}$ Jun Liu $^{1‡}$ $^{1}$ Singapore University of Technology and Design $^{2}$ Lancaster University", + "bbox": [ + 169, + 178, + 795, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jia_gong,lingeng_foo,yixuan_he}@mypmail.sutd.edu.sg, h.rahmani@lancaster.ac.uk, jun.liu@sutd.edu.sg", + "bbox": [ + 259, + 219, + 705, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign Language Translation (SLT) is a challenging task that aims to translate sign videos into spoken language. Inspired by the strong translation capabilities of large language models (LLMs) that are trained on extensive multilingual text corpora, we aim to harness off-the-shelf LLMs to handle SLT. In this paper, we regularize the sign videos to embody linguistic characteristics of spoken language, and propose a novel SignLLM framework to transform sign videos into a language-like representation for improved readability by off-the-shelf LLMs. SignLLM comprises two key modules: (1) The Vector-Quantized Visual Sign module converts sign videos into a sequence of discrete character-level sign tokens, and (2) the Codebook Reconstruction and Alignment module converts these character-level tokens into word-level sign representations using an optimal transport formulation. A sign-text alignment loss further bridges the gap between sign and text tokens, enhancing semantic compatibility. We achieve state-of-the-art gloss-free results on two widely-used SLT benchmarks.", + "bbox": [ + 75, + 318, + 472, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 637, + 209, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sign languages, which are visual signals expressed through hand, body, and facial movements, serve as the primary means of communication within the hearing-impaired community. In an effort to facilitate effective communication with this community, much attention has been directed towards developing techniques to tackle the Sign Language Translation (SLT) task [10, 11, 69, 72, 73], where the goal is to translate sign videos into spoken language. SLT is a challenging task that requires cross-modality understanding of visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72]. Despite the notable advancements in terms of network architectures [10, 11, 37], visual sign representations (e.g., with keypoint estimators [14, 74]),", + "bbox": [ + 75, + 662, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and training methods [22, 69, 72], how to effectively tackle the challenging cross-modal SLT task with limited paired sign-text data largely remains an open question.", + "bbox": [ + 500, + 287, + 890, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, large language models (LLMs) – referring to language models that have been trained on a large web-scale text corpus – have recently received a lot of attention. Since LLMs are trained over a very large corpus across multiple languages with distinct syntax and lexicon, they possess rich semantic understanding and powerful linguistic abilities [9, 15, 58]. At the same time, LLMs have also demonstrated an impressive capability to translate across multiple languages [9, 15], even showing a strong potential for translating languages with limited data [66, 76]. The foundation for this translation proficiency lies in the shared linguistic properties of syntax, lexicon, and morphology that many languages hold, which is particularly evident within language families [3]. Therefore, when faced with a new language with limited data, LLMs can draw upon the wealth of knowledge acquired from previously learned languages, leveraging any shared properties of syntax, lexicon and morphology with previous languages to effectively generate translations for new languages with remarkable accuracy and fluency [66, 76].", + "bbox": [ + 496, + 337, + 892, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the impressive translation capabilities of LLMs, we aim to harness off-the-shelf LLMs to handle the challenging SLT task. However, training LLMs directly on the relatively small SLT dataset can potentially lead to forgetting of their rich knowledge [12, 29] and a decline in performance, thus we follow previous LLM-based works [24, 43, 54] to keep the off-the-shelf LLM frozen, which preserves the rich knowledge acquired during its pretraining on a vast multilingual corpus. Consequently, our focus shifts towards making the sign videos compatible and readable for the off-the-shelf and frozen LLM to perform SLT. Specifically, in this paper, we explore the following question: Can we treat the sign video as a form of language, and leverage an off-the-shelf and frozen LLM to translate them? Notably, this is not a straightforward task because directly encoding features from sign videos with a pre-trained feature extractor [28, 64] will result in a large gap between", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal contribution; ‡ Corresponding author", + "bbox": [ + 101, + 887, + 344, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18362", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the sign video features and text tokens, making it difficult for off-the-shelf LLMs to understand them.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the observation that LLMs can effectively handle new languages by leveraging shared commonalities with previously learned languages, we aim to introduce designs that transform our sign videos into a language-like format which are readable and friendly to LLMs. Specifically, we hypothesize that providing language-like representations of sign videos to the LLM improves the LLM's understanding of the sign videos and facilitates greater exploitation of shared properties with previously learned languages, thus resulting in better SLT performance by the LLM. To obtain language-like sign video representations, we draw inspiration from linguistic studies and analyses on LLMs [30, 63] and regularize the sign video to embody two fundamental language-like characteristics: Discrete Characteristics: Spoken languages are inherently discrete, since each language contains a finite set of words (and subwords) that convey distinct concepts, allowing them to be naturally represented through a discrete vocabulary with distinct tokens [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52]. This hierarchical structure enables languages to express a wide range of words with a limited set of characters, and convey diverse sentences with a limited number of words.", + "bbox": [ + 76, + 125, + 472, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present SignLLM, a novel framework designed to regularize input sign videos to produce sign token representations with language-like characteristics that are compatible and friendly to LLMs. Our proposed SignLLM includes two key designs to impart discrete characteristics and a hierarchical structure to the produced sign tokens. Firstly, we introduce the Vector-Quantized Visual Sign (VQ-Sign) module that facilitates the conversion of sign videos into a sequence of discrete character-level sign tokens. To achieve this, the VQ-Sign module consists of a discrete character-level sign codebook which is optimized through a self-supervised context prediction task. Next, we introduce the Codebook Reconstruction and Alignment (CRA) module that converts the character-level sign tokens into word-level sign tokens, facilitated by an optimal transport formulation. Moreover, we employ a sign-text alignment loss to further narrow the gap between the sign tokens and text tokens. These designs enable SignLLM to produce sign sentences that embody two key characteristics of spoken languages: discrete characteristics and a hierarchical structure, which enhances their compatibility with LLMs and makes them more readily interpretable by LLMs.", + "bbox": [ + 76, + 489, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After producing the language-like sign sentences, we feed them into an off-the-shelf and frozen LLM along with a text prompt that instructs the LLM to generate translations in the desired language. We empirically observe that, by employing our SignLLM's designs to align sign videos with", + "bbox": [ + 76, + 825, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "languages, we can already leverage a frozen LLM to attain state-of-the-art SLT performance. These findings suggest that our proposed SignLLM framework is a promising first step towards effectively harnessing LLMs for SLT. We hope our initial explorations can inspire future work within the community to leverage LLMs for SLT.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are: (1) We propose a novel SignLLM framework that is the first to harness the power of off-the-shelf and frozen LLMs for SLT. (2) To make the input sign video compatible with LLMs, our SignLLM framework incorporates two designs: a VQ-Sign module to quantize the sign video into a sequence of discrete character-level sign tokens and a CRA module that transforms the character-level sign tokens to word-level sign tokens. (3) Through our proposed designs, we achieve state-of-the-art gloss-free results on two popular SLT datasets.", + "bbox": [ + 496, + 181, + 893, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 345, + 640, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sign Language Translation (SLT) aims to transform sign videos into natural language sentences. It is a challenging task that requires understanding of both visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72] which limits the performance of SLT methods. To improve SLT performance, many previous works [10, 11, 14, 22, 34, 35, 37, 56, 68, 69, 72, 74] aim to enhance the visual sign representations and text decoding capabilities of SLT methods. Some works propose deep architectures based on RNNs [10, 35], GCNs [34], and Transformers [11, 37, 62, 68]. Other approaches include introducing a keypoint estimator to enhance the visual sign representation [14, 35, 56, 74], introducing pre-training tasks [22, 72], or jointly modelling several SLT-related tasks [69]. Some works also introduce larger datasets (e.g., How2Sign [20] and BOBSL [1]), which present a huge challenge with their large sign and text vocabularies. Besides, some recent works [67, 72] focus on the gloss-free setting – these works do not use sign gloss annotations for training, which reduces the cost of training SLT models, and our work also falls into this category. In contrast to existing works, we aim to harness the capabilities of off-the-shelf and frozen LLMs to perform SLT, by regularizing the sign videos into a language-like representation and prompting the LLM to generate text of the desired language.", + "bbox": [ + 496, + 372, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) refer to language models that have been trained extensively on a very large web-scale text corpus. LLMs have shown impressive text generation capabilities, attracting a lot of attention recently [21, 49, 71]. In particular, since LLMs have been trained on large amounts of text data, they demonstrate strong generalization capabilities across various text-based tasks, including code generation [51], open-domain question answering [75], and multilingual translation [9, 15]. Inspired by the re", + "bbox": [ + 496, + 765, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18363", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5fdbc12d114ffb55abef70f386f672c6c2cdc35f62277c31ce117b7d00a6220c.jpg", + "image_caption": [ + "Figure 1. An overview of our SignLLM framework. During inference (top): Given an input sign video $X$ , we first pass it through our VQ-Sign module to obtain a sequence of discrete character-level sign tokens $\\hat{Z}$ . Our VQ-Sign consists of a visual encoder $E_{v}$ to extract compact features and a character-level sign codebook $\\mathbb{S}^c$ for quantization to obtain $\\hat{Z}$ . Next, we feed $\\hat{Z}$ into our CRA module, which reorganizes $\\hat{Z}$ by replacing short sequences of character tokens with word-level tokens via the word-level codebook, e.g., character sequence $[s_2, s_3, s_4]$ to word $s_2 s_3 s_4$ . This transforms the sign video data to a language-like sign sentence $W$ , which is fed into the LLM along with a text prompt which guides the LLM to generate translations in the desired language. During training (bottom): We optimize VQ-Sign and its discrete sign codebook via a context prediction task, which seeks to recognize the future time steps based on the current context information. Next, for our CRA module, we construct the optimal word-level codebook by considering two aspects: entropy and size, which we address using optimal transport techniques. Then, we narrow the gap between the sign token space and LLM's text token space via minimizing the MMD loss, which improves the semantic compatibility between them." + ], + "image_footnote": [], + "bbox": [ + 125, + 92, + 841, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cent advancements in LLMs, we explore harnessing LLMs for translation of sign videos, by converting the sign videos to a sequence of language-like sign tokens via our Sign-LLM framework, and treating the sign tokens as a form of language that can be translated by an LLM. To the best of our knowledge, we are the first work to leverage an off-the-shelf and frozen LLM to tackle SLT.", + "bbox": [ + 75, + 489, + 470, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 611, + 169, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the overview of our SignLLM in Sec. 3.1. Then, we describe two main components of our SignLLM framework: the VQ-Sign and CRA modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we list the training and inference details in Sec. 3.4.", + "bbox": [ + 75, + 637, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SignLLM Overview", + "text_level": 1, + "bbox": [ + 76, + 724, + 266, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to effectively handle SLT, in this paper we draw inspiration from LLMs' remarkable capabilities in generating translations across multiple languages [9, 15]. In particular, LLMs have been extensively trained on a large web-scale multilingual text corpus and have learned diverse knowledge regarding the properties of many languages, thus they are able to draw on shared commonalities with previously learned languages to effectively handle new languages with limited data [66, 76].", + "bbox": [ + 75, + 748, + 468, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, to leverage the strong translation capabilities", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of LLMs to handle SLT, we introduce a novel SignLLM framework. SignLLM converts the input sign video $X$ into a language-like sign sentence $W$ that aligns with the linguistic characteristics of spoken languages and is friendly and compatible with LLMs. Then, to perform SLT, the language-like sign sentence $W$ can be fed into an off-the-shelf and frozen LLM along with a text prompt that guides the LLM to generate translations in the desired language.", + "bbox": [ + 496, + 489, + 890, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, to produce sign sentences $W$ that are friendly and understandable by LLMs, we aim to regularize our sign sentences $W$ to embody two core linguistic characteristics: Discrete Characteristics: Spoken languages are naturally discrete and consist of distinct words or sub-words with corresponding discrete tokens in a vocabulary [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52], where words are composed from characters and sentences are composed from various words.", + "bbox": [ + 496, + 612, + 892, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the above, our SignLLM framework comprises of three parts, as shown in Fig. 1: (1) The VQ-Sign module converts the input sign video $X$ into a sequence of discrete sign tokens $\\hat{Z}$ , aligning the sign representations with text's discrete characteristics. These sign tokens $\\hat{Z}$ are character-level sign tokens that are retrieved from a learned discrete character-level codebook. (2) The CRA module maps meaningful compositions of character-level sign tokens $\\hat{Z}$ into word-level sign tokens that form a sign sentence", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18364", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$W$ , further imparting a language-like hierarchical structure to the video sign representations. Moreover, we also align the sign token codebooks towards the text token space to improve semantic compatibility. (3) An off-the-shelf LLM takes the sign sentence $W$ as input, along with an instructive text prompt that guides the LLM to generate the translation in the desired language. More details regarding the text prompt are in the Supplementary. Next, we present our VQ-Sign and CRA modules in detail.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Vector-Quantized Visual Sign Module", + "text_level": 1, + "bbox": [ + 76, + 242, + 405, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, in order to produce a language-like representation, we would like to impart discrete characteristics to the input sign videos, aligning them more closely with spoken language representations that are inherently discrete and consist of distinct tokens in a vocabulary. However, achieving this is not straightforward because sign videos are a continuous signal in a high-dimensional spatio-temporal space which cannot be easily represented by a set of discrete tokens, and for which the vocabulary is not readily available. Hence, we introduce our Vector-Quantized Visual Sign (VQ-Sign) module to quantize the sign video $X$ into a sequence of discrete sign tokens $\\hat{Z}$ via a sign codebook $\\mathbb{S}^c$ . As illustrated in Fig. 1, our VQ-Sign module involves a series of steps, which we present in detail next.", + "bbox": [ + 75, + 267, + 472, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first step, we extract a compact feature $Z$ from the high-dimensional input sign video $X \\in \\mathbb{R}^{N \\times H \\times W}$ , where $N$ is the number of video frames, while $H$ and $W$ are the height and width of the video frames respectively. To be precise, the sign video $X$ is first organized into a sequence of short overlapping video clips, then each short video clip is fed into a visual encoder $E_{v}$ to extract a compact feature representation of dimensionality $d$ . Overall, this step transforms the original high-dimensional input sign video $X \\in \\mathbb{R}^{N \\times H \\times W}$ into a compact feature $Z \\in \\mathbb{R}^{\\frac{N}{n} \\times d}$ , where $n$ represents the number of frames between the start of neighboring clips. Notably, since $Z$ is obtained by processing $\\frac{N}{n}$ short clips, $Z$ can also be seen as a sequence of $\\frac{N}{n}$ clip-wise features, i.e., $\\{z_{t}\\}_{t=1}^{\\frac{N}{n}}$ , where each $z_{t} \\in \\mathbb{R}^{d}$ corresponds to the feature of the $t$ -th short clip.", + "bbox": [ + 75, + 479, + 472, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the next step, we transform the feature $Z = \\{z_{t}\\}_{t=1}^{\\frac{N}{n}}$ into a sequence of discrete tokens $\\hat{Z}$ using a codebook $\\mathbb{S}^c$ . Specifically, we discretize each clip's feature $z_{t}$ into a discrete token $\\hat{z}_{t}$ by finding the matching token $s_{j}$ from the codebook $\\mathbb{S}^c = \\{s_i\\}_{i=1}^M$ , where the $i$ -th token in the codebook is denoted as $s_{i} \\in \\mathbb{R}^{d}$ and $M$ is the number of tokens in the codebook. The matching token $s_{j}$ is the codebook's closest element to the feature $z_{t}$ in terms of Euclidean distance, i.e., $j = \\arg \\min_{i}(\\| z_{t} - s_{i}\\|_{2}^{2})$ . After the matching, each feature $z_{t}$ is replaced by $\\hat{z}_{t} = s_{j}$ , which results in a discrete token sequence as shown in Fig. 1, e.g., $[s_2, s_3, s_4, s_M, s_5, s_1]$ . Note that, we randomly initialize all", + "bbox": [ + 75, + 713, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the tokens $\\{s_i\\}_{i=1}^M$ in the sign codebook $\\mathbb{S}^c$ at the start and optimize them during training, as introduced next.", + "bbox": [ + 496, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we face a challenge in learning the discrete codebook $\\mathbb{S}^c$ . In particular, although autoencoding [59] has been a popular method for producing a codebook of discrete units, the high complexity of sign videos makes autoencoding (i.e., self-reconstruction of sign videos) challenging and costly. Therefore, inspired from predictive coding [4], a widely-used method in text and speech representation learning [6, 44, 46], we propose to learn discrete representations of sign videos through a context prediction task. Context prediction [46] is a self-supervised task that focuses on recognizing the future content in latent space based on the current information, which can learn discrete representations while eliminating the need for reconstructing the high-dimensional input video data. Furthermore, previous works show that training with context prediction effectively captures the temporal dependencies and relationships between elements in a sequence [5, 27], and the learned representations are often transferable to downstream tasks [6, 46].", + "bbox": [ + 496, + 121, + 892, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we employ a context prediction task where we try to distinguish a future sample $z_{\\tau + k}$ based on the current context representation $c_{\\tau}$ at various time steps $\\tau$ . To facilitate this task, after we obtain the discrete token sequence $\\hat{Z}$ , we further produce a context latent representation $c_{\\tau}$ using an auto-regressive model $g$ that summarizes all the discrete tokens before a certain time step $\\tau$ (i.e., $\\{\\hat{z}_t\\}_{t \\leq \\tau}$ ) to produce context latent representations $c_{\\tau} = g(\\{\\hat{z}_t\\}_{t \\leq \\tau})$ . Then, we optimize our module by minimizing the following context prediction contrastive loss $\\mathcal{L}_k^{\\mathrm{cp}}$ :", + "bbox": [ + 496, + 393, + 893, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {k} ^ {\\mathrm {c p}} = - \\sum_ {\\tau = 1} ^ {\\frac {N}{n} - k} \\left(\\log \\sigma \\left(z _ {\\tau + k} ^ {\\top} h _ {\\tau}\\right) + \\lambda \\underset {\\widetilde {z} \\sim p _ {n}} {\\mathbb {E}} \\left[ \\log \\sigma \\left(- \\widetilde {z} ^ {\\top} h _ {\\tau}\\right) \\right]\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 551, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_\\tau$ is obtained by applying a trainable linear layer to $c_{\\tau}$ , $\\sigma(z_{\\tau+k}^{\\top} h_{\\tau})$ is the probability of $z_{\\tau+k}$ being the true sample among the negatives, $\\widetilde{z}$ are negative samples drawn from a minibatch $p_n$ , and $\\lambda$ is a hyperparameter. We sum $\\mathcal{L}_k^{\\mathrm{cp}}$ over different step sizes $k$ to obtain the context prediction loss $\\mathcal{L}^{\\mathrm{cp}} = \\sum_{k=1}^{K} \\mathcal{L}_k^{\\mathrm{cp}}$ , where $K$ is the maximum number of future clips that we are interested in.", + "bbox": [ + 496, + 602, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [18], in order to optimize the matching between $\\hat{z}_t$ and $z_{t}$ , we further add two losses to optimize the matching distance between $\\hat{z}_t$ and $z_{t}$ , such that the overall loss $\\mathcal{L}^{VQ}$ to optimize our VQ-Sign module is as follows:", + "bbox": [ + 496, + 709, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {V Q} = \\mathcal {L} ^ {\\mathrm {c p}} + \\sum_ {t = 1} ^ {N / n} \\| \\operatorname {s g} \\left(z _ {t}\\right) - \\hat {z} _ {t} \\| ^ {2} + \\gamma \\sum_ {t = 1} ^ {N / n} \\| z _ {t} - \\operatorname {s g} (\\hat {z} _ {t}) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 775, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{sg}(z) \\equiv z$ is the stop-gradient operator and $\\gamma$ is a hyperparameter. By optimizing $\\mathcal{L}^{VQ}$ , we can train our VQ-Sign and the discrete codebook without the need for reconstructing high-dimensional video clips, which makes the codebook construction a viable and relatively cheap option.", + "bbox": [ + 496, + 824, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our VQ-Sign transforms sign videos into sequences of discrete sign tokens $\\hat{Z}$ , which are friendlier and more understandable to LLMs. Notably, the produced discrete tokens $\\hat{Z}$ can be likened to character-level tokens, in the sense that each discrete token $\\hat{z}_t$ corresponds to a short clip and may not contain much semantic meaning on its own (similar to linguistic characters), but they can be combined into a sequence to convey a clear semantic meaning (akin to forming a word or sentence). Thus, inspired by this, we call VQ-Sign's codebook $\\mathbb{S}^c$ the character-level codebook that contains character-level sign tokens.", + "bbox": [ + 76, + 90, + 472, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Codebook Reconstruction and Alignment", + "text_level": 1, + "bbox": [ + 76, + 268, + 431, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the previous section, we quantize sign videos into discrete character-level sign tokens, which aligns them closer to language representations. In this section, our goal is to impart a hierarchical structure to our sign video representations, which makes them align even closer to language representations. Specifically, we aim to compose our character-level sign tokens into word-level sign tokens to mirror the observed hierarchical structure in spoken language, which makes them even more compatible with LLMs.", + "bbox": [ + 76, + 292, + 468, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, considering a spoken language sentence, we can represent it as a sequence of words, with each word formed by one or multiple characters. For example, a sentence 'I love AI' can be decomposed into a word sequence [I', 'love', 'AI'], where the word 'love' is in turn formed by the character sequence [1', o', v', e']. We observe that, although each individual character may not contain much semantic meaning on its own, they can be composed to form words with clearer semantic meaning. In a similar fashion, we also want to impart such hierarchical structure to our character-level sign tokens by composing them to form meaningful word-level sign tokens.", + "bbox": [ + 76, + 430, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hence, we aim to find an optimal transformation from character-level sign tokens to word-level sign tokens for enhanced readability and compatibility with LLMs. To this end, we introduce the Codebook Reconstruction and Alignment (CRA) module to transform the character-level codebook $\\mathbb{S}^c$ from VQ-Sign into a word-level codebook $\\mathbb{S}^w$ whose tokens convey richer and clearer semantic meaning. Inspired by optimal transport methods [16, 61, 65], we observe that the above transformation can be formulated as an optimal transport problem of transporting characters into words, thus we introduce a codebook reconstruction algorithm with an optimal transport formulation to find an optimal transformation. Additionally, to further reduce the distribution gap between the sign tokens and the text tokens, our CRA module also performs sign-text alignment, enhancing the semantic compatibility of the sign tokens with LLMs. We introduce the details below.", + "bbox": [ + 76, + 612, + 468, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To begin, the objective of our Codebook Reconstruction Algorithm is to create a word-level codebook $\\mathbb{S}^w$", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "based on VQ-Sign's character-level codebook $\\mathbb{S}^c$ . The challenge lies in determining which character-level sign tokens should be assembled together to form word-level sign tokens, which is a complex problem. To address this complexity, we adopt an approach based on two fundamental principles. Firstly, in order to maximize the overall predictability of the word-level tokens and enhance the distinctiveness of each token, we seek to minimize the entropy of each word-level token within the vocabulary [42]. We remark that, several approaches for establishing language-based subword vocabularies [8, 55] can be seen as entropy-minimizing approaches, employing different heuristics to establish the vocabulary with the goal of minimizing entropy [23]. On the other hand, considering the limited availability of sign video data, we incorporate codebook size as another key factor in our word-level codebook construction, since studies on languages with limited data [26, 53] have also identified vocabulary size as a crucial aspect. Specifically, too small a vocabulary can result in sub-optimal entropy values, while an excessively large vocabulary size can lead to issues such as parameter explosion and token sparsity which hinder understanding [2], and finding the right balance between these effects becomes even more sensitive for languages with limited data [19, 53].", + "bbox": [ + 496, + 90, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on these principles, our objective is to determine an optimal codebook size that maximizes the entropy reduction while taking into account the increase in codebook size. In other words, we would like to find an optimal codebook size that maximizes the gradient of the entropy reduction with respect to the codebook size increase. To simplify the optimal size searching problem, we define a fixed size increment $m$ and search through codebooks of various sizes (where the difference between each codebook size is $m$ tokens). Specifically, we define the $r$ -th codebook $(\\mathbb{S}_r^w)$ as the codebook with $r \\times m$ tokens. Then, we seek to identify the optimal set of word-level tokens, where each word-level token is composed of character-level tokens. We approach this by formulating the character compositions as an optimal transport problem, where characters are transported to words.", + "bbox": [ + 496, + 458, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, it can be challenging to identify specific character combinations that convey precise semantic information, due to the temporal complexity of sign videos, which often makes the character-level token sequences $\\hat{Z}$ quite messy. For instance, some signers may execute the signing motions at a slower speed, which can lead to consecutive short video clips being highly similar, resulting in consecutively-repeating character-level discrete tokens. Thus, the character-level sequences between different signers can differ significantly (e.g., $[s_1,s_2]$ vs $[s_1,s_1,s_1,s_2,s_2]$ ) due to such duplication of character-level tokens, even though they may contain the same semantic information. At the same time, simply filtering out re", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18366", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "peated character-level tokens straightforwardly (e.g., setting all $[s_1, s_1, s_1, s_2, s_2]$ to $[s_1, s_2]$ ) is sub-optimal, since the speed of some signs can also convey some information [31, 60], e.g., if a signer signs \"ugly\" quickly, it conveys \"very ugly\" in American Sign Language.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pre-processing of Repeated Characters. Therefore, to alleviate the impact of the signer's speed while keeping the information regarding each sign's speed, we first preprocess the character-level sequences as follows: First, we find all the repeated tokens in the character-level sequence and compute the average number of repeated tokens $(\\alpha)$ in each sequence. Then, for each repeated sequence (e.g., $[s_1, s_1, s_1]$ ), we keep the first character and remove the tailing repetitive ones (e.g., $[s_1, s_1, s_1]$ to $[s_1]$ ). At the same time, if the character-level tokens repeat more than $\\alpha$ times, we insert a single character-level token $s_0$ as a \"slowing down\" sign, e.g., $[s_1, s_1, s_1]$ to $[s_1, s_0]$ if $\\alpha < 3$ . Crucially, this allows us to reduce redundancy, while still representing \"quick\" or \"slow\" signs that account for differences in the signer's speed. Overall, this pre-processing and reducing of repeating characters makes the character-level sequence less messy, facilitating the search for specific meaningful character combinations.", + "bbox": [ + 75, + 167, + 470, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Optimal Transport Formulation. Then, using the preprocessed character-level sequences, we search for an optimal word-level codebook via an optimal transport formulation [16, 61, 65]. Following our discussion above, we aim to find an optimal word-level codebook (i.e., $\\mathbb{S}_r^w$ ) with low entropy and compact size. Specifically, to measure codebook entropy, we follow [25, 42, 45] to define the entropy of the $r$ -th codebook $\\mathbb{S}_r^w$ as:", + "bbox": [ + 75, + 438, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {\\mathbb {S} _ {r} ^ {w}} = - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} P (w _ {j}) \\log P (w _ {j}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 564, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P(w_{j})$ is the relative frequency of the $j^{th}$ token $w_{j}$ from the word-level codebook $\\mathbb{S}_r^w$ . Then, based on VQ-Sign's character-level codebook $\\mathbb{S}^c$ , the entropy of the word-level codebook $\\mathbb{S}_r^w$ can be computed through the following (with proof in Supplementary):", + "bbox": [ + 76, + 598, + 468, + 674 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal{H}_{\\mathbb{S}_{r}^{w}} = -\\sum_{w_{j}\\in \\mathbb{S}^{w}}\\sum_{s_{i}\\in \\mathbb{S}^{c}}P(w_{j},s_{i})\\log P(w_{j},s_{i}) \\\\ - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} \\sum_ {s _ {i} \\in \\mathbb {S} ^ {c}} P \\left(w _ {j}, s _ {i}\\right) (- \\log P \\left(s _ {i} \\mid w _ {j}\\right)), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 676, + 468, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P(s_{i}|w_{j})$ is the probability of the character-level token $s_i$ appearing in the word-level token $w_{j}$ .", + "bbox": [ + 75, + 750, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, to formally define our objective function, we follow existing works [16, 48, 65] to define a transport matrix $\\pmb{P}$ that represents the assignments of characters to words, and a distance matrix $\\pmb{D}$ that represents the cost of transportation. We define the transport matrix $\\pmb{P} \\in \\mathbb{R}^{m \\times (r \\cdot m)}$ with the $(j,i)$ -th element as $P(w_{j}, s_{i})$ , and define the distance matrix $\\pmb{D} \\in \\mathbb{R}^{m \\times (r \\cdot m)}$ as a matrix whose $(j,i)$ -th element is $\\log P(s_{i} | w_{j})$ . Note that, if $w_{j}$ contains $s_{i}$ , we use", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\frac{1}{\\mathrm{length}(w_j)}$ to estimate $P(s_i|w_j)$ , and if $w_j$ does not contain $s_i$ , then it is deemed an infeasible assignment, and we set the distance $\\log P(s_i|w_j) = \\infty$ . We further define $H(\\pmb{P})$ as $-\\sum_{w_j \\in \\mathbb{S}_r^w} \\sum_{s_i \\in \\mathbb{S}^c} P(w_j, s_i) \\log P(w_j, s_i)$ , which is simply the entropy of the probability distribution $P(w_j, s_i)$ .", + "bbox": [ + 498, + 90, + 890, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, based on Eq. 4, the objective function for minimizing the entropy of $\\mathbb{S}_r^w$ can be formulated as:", + "bbox": [ + 498, + 170, + 890, + 200 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {P} \\in \\mathbb {R} ^ {m} \\times (r \\cdot m)} {\\arg \\min } H (\\boldsymbol {P}) + \\sum_ {j} \\sum_ {i} \\boldsymbol {P} (j, i) \\boldsymbol {D} (j, i). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 205, + 890, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, following previous works [16, 48, 65], we impose two constraints on the transport matrix $P$ : the sum of each row in $P$ should equal to the probability of character token $s_i$ and the sum of each column in $P$ should equal to the probability of word token $w_j$ . Formally, we constrain the transport matrix $P$ with: $|\\sum_{i}P(i,j) - P(w_{j})|\\leq \\epsilon$ and $|\\sum_{j}P(i,j) - P(s_{i})|\\leq \\epsilon$ , where $\\epsilon$ is a small positive fixed hyperparameter.", + "bbox": [ + 496, + 239, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively, this optimization process can be regarded as an optimal transport problem to find the best way to transport mass from $\\mathbb{S}^c$ to $\\mathbb{S}_r^w$ . To handle this optimal transport problem, we leverage the Sinkhorn algorithm [16, 48, 65], allowing us to effectively construct the candidate word-level codebooks $\\mathbb{S}_r^w$ with minimal entropy. Since our increment $m$ between each candidate codebook is fixed, we can find the optimal codebook size that maximizes the gradient of entropy reduction by simply computing and finding the maximum the entropy difference between $\\mathbb{S}_r^w$ and $\\mathbb{S}_{r - 1}^w$ .", + "bbox": [ + 496, + 361, + 890, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After finding the optimal word-level codebook, we construct the word-level sign tokens by composing all features of the character-level tokens into the word-level tokens via our autoregressive model $g$ . Refer to Supplementary for more details. Overall, with our codebook reconstruction algorithm, we can construct an optimal word-level sign codebook with low entropy yet also with relatively small size.", + "bbox": [ + 496, + 513, + 890, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sign-Text Alignment. Next, we further align the sign tokens with the text tokens used in LLMs in order to further improve semantic compatibility between them. To achieve this, we measure the distribution gap between sign tokens and text tokens via Maximum Mean Discrepancy (MMD) [57] and then optimize the sign tokens' embeddings by minimizing MMD, which narrows down the distribution gap. Specifically, we compute the gap between the sign embedding space $\\mathbb{F}^s$ and text embedding space $\\mathbb{F}^t$ via MMD as:", + "bbox": [ + 496, + 625, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} ^ {M M D} (\\mathbb {F} ^ {s}, \\mathbb {F} ^ {t}) = \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {s}} \\frac {k (f (p _ {i}) , f (p _ {j}))}{n _ {s} ^ {2}} \\\\ + \\sum_ {i = 1} ^ {n _ {t}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {k \\left(q _ {i} , q _ {j}\\right)}{n _ {t} ^ {2}} - \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {2 \\cdot k \\left(f \\left(p _ {i}\\right) , q _ {j}\\right)}{n _ {s} n _ {t}}, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 767, + 890, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p$ and $q$ are the tokens in $\\mathbb{F}^s$ and $\\mathbb{F}^t$ , $f$ is a small projection module that projects $\\mathbb{F}^s$ to $\\mathbb{F}^t$ , $n_s$ and $n_t$ are the numbers of tokens in $\\mathbb{F}^s$ and $\\mathbb{F}^t$ , and $k(\\cdot)$ represents the radial", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18367", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "kernel [57] that measures the distance between two samples. We apply MMD loss to both the word-level and the character-level tokens to narrow the overall sign-text gap.", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 147, + 290, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference. Given a sign video $X$ , we first extract compact features $Z$ via the visual encoder $E_{v}$ , and quantize $Z$ to $\\hat{Z}$ via VQ-Sign's learned character-level codebook. Then, we transform the sequence of discrete character-level tokens $\\hat{Z}$ into word-level tokens via our CRA, which produces a sign sentence $W$ . Lastly, we project the sign sentence $W$ to LLM embedding space via the small projection module $f$ , and then feed the sign sentence $W$ into the LLM along with a text prompt to instruct the LLM to perform the SLT task and generate text in the desired language.", + "bbox": [ + 75, + 171, + 467, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training. Our SignLLM is optimized in two stages: pretraining and fine-tuning. Specifically, the pre-training stage, which does not require explicit SLT supervision, includes two sub-stages: (i) We first pre-train VQ-Sign via the context prediction task with $\\mathcal{L}^{VQ}$ in Eq. 2. (ii) Then, based on the VQ-Sign's learned character-level codebook, we construct the word-level codebook with the codebook reconstruction algorithm (Sec. 3.3) and apply the MMD loss $\\mathcal{L}^{MMD}$ in Eq. 6 to align the sign codebooks $(\\mathbb{S}^c$ and $\\mathbb{S}^w)$ and the text vocabulary of the desired language.", + "bbox": [ + 75, + 321, + 467, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After the pre-training, we fine-tune SignLLM. To aid LLMs in understanding the sign sentences as texts, we additionally maximize the similarity between the text tokens $Y$ generated by LLM and the ground truth tokens $\\bar{Y}$ as: $\\mathcal{L}^{sim}$ by minimizing the cross-entropy between them. We fine-tune our SignLLM (with frozen LLM) via the loss $\\mathcal{L}^{ft}$ as follows: $\\mathcal{L}^{ft} = \\mathcal{L}^{VQ} + \\lambda_1\\mathcal{L}^{MMD} + \\lambda_2\\mathcal{L}^{sim}$ , where $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters. Note that we follow previous gloss-free works [67, 72] to train our SignLLM framework, eliminating the need for additional gloss data.", + "bbox": [ + 75, + 474, + 467, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 640, + 209, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 665, + 294, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our visual encoder $E_{v}$ is constructed by appending two Conv3D layers with a kernel size of (5, 3, 3) and a stride of (2, 1, 1) to a ResNet18 [28] pre-trained on ImageNet [17]. Each clip consists of 13 frames and the gap between the neighboring clips ( $n$ ) is 4. Besides, the auto-regressive model $g$ is implemented as a Convolutional Gated Recurrent Layer with a kernel size of (1, 1). We set the total number of discrete vectors $M$ at 256 and each vector's dimension $d$ at 1024 for our character-level codebook. In VQ-Sign's pre-training phase, we set $\\gamma = 0.25$ , initialize the learning rate at 0.01 and use the Adam algorithm, training the model to predict the future three clips ( $K = 3$ ) for 200 epochs. During codebook reconstruction, we set the increment $m$ to 32. We employ the frozen LLaMA-7B-16bit [58]", + "bbox": [ + 75, + 688, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as our LLM and project the codebook space to LLaMA's embedding space via $f$ , which consists of two fc-layers with ReLU. We set $\\lambda_1 = 0.5$ , $\\lambda_2 = 1$ , and initialize the learning rate at 0.001 to fine-tune our SignLLM over 20 epochs. Please see Supplementary for more implementation details.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 498, + 176, + 785, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. We follow previous works [13, 14, 34, 72-74] to run experiments on the Phoenix-2014T [10] and CSL-Daily [73] datasets for SLT, and evaluate on their dev and test sets. Phoenix-2014T [10] is a German sign language dataset with a vocabulary size of 2887 German words. The training, dev, and test sets contain 7096, 519, and 642 samples. CSL-Daily [73] is a Chinese sign language dataset with a vocabulary size of 2343 Chinese words. The training, dev, and test sets contain 18401, 1077, and 1176 samples.", + "bbox": [ + 496, + 199, + 890, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous works [13, 14, 37, 67, 72-74], we adopt BLEU [47] and ROUGE-L [38] as the evaluation metrics for SLT. BLEU-n evaluates the average translation precision up to n-grams, and we follow previous works [13, 14, 72-74] to report results for BLEU-1 to BLEU-4 (i.e., B1, B2, B3 and B4). ROUGE-L (or ROUGE) computes the F1 score based on the longest common subsequence between the predicted and ground truth texts.", + "bbox": [ + 496, + 335, + 890, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 498, + 465, + 638, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on PHOENIX2014T dataset. Tab. 1 presents a comparison of our approach with state-of-the-art gloss-based and gloss-free methods for SLT. Our method consistently improves upon all reported metrics as compared to other gloss-free approaches.", + "bbox": [ + 496, + 489, + 890, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on CSL-Daily dataset. We compare our method with state-of-the-art approaches on the CSLD Daily dataset in Tab. 2. We outperform previous gloss-free works on all metrics, showing the efficacy of our approach.", + "bbox": [ + 496, + 565, + 890, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 635, + 653, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further investigate the proposed method, we follow previous works [13, 33, 67, 69, 72] to conduct extensive ablation experiments on the Phoenix-2014T dev and test sets. Refer to Supplementary for more experiment results.", + "bbox": [ + 496, + 657, + 890, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of LLM. First, we evaluate the impact of leveraging LLMs for the SLT task. Specifically, we establish three baselines: 1) Ours (w/o LLM) where we replace the LLM with a trainable lightweight text generator (mBART [40]). 2) Ours (w/ T5) where we replace our LLM with a smaller LLM (T5 [50]). As shown in Tab. 3, our approach (w/ LLaMA [58]) achieves a much better performance than Ours (w/o LLM), showing the efficacy of leveraging a powerful LLM. We also find that using a smaller and less powerful LLM (T5) leads to worse performance, as expected.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of SignLLM. Next, we explore the impact of Sign-LLM by comparing against the following baselines: 1) En", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18368", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/457d52e8231e3c9de074ba49e61839a276c444444d2320748be3e4c6875f2c3c.jpg", + "table_caption": [ + "Table 1. Results on Phoenix-2014T dataset [10]." + ], + "table_footnote": [], + "table_body": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]47.2634.4027.0522.38-46.6133.7326.1921.32-
ConSLT [22]---21.1147.74---21.5947.69
STN-SLT [62]49.1236.2928.3423.23-48.6135.9728.3723.65-
STMC-T [74]47.6036.4329.1824.0948.2446.9836.0928.7023.6546.65
BN-TIN-Transf.+SignBT [73]51.1137.9029.8024.4550.2950.8037.7529.7224.3249.54
PET [33]-----49.5437.1929.3024.0249.97
MMTLB [13]53.9541.1233.1427.6153.1053.9741.7533.8428.3952.65
TS-SLT [14]54.3241.9934.1528.6654.0854.9042.4334.4628.9553.48
SLTUNET [69]---27.8752.2352.9241.7633.9928.4752.11
Gloss-freeNSLT [10]28.1016.8111.829.1231.0027.1015.6110.828.3529.70
NSLT+Bahdanau [7, 10]31.8719.1113.169.9431.8032.2419.0312.839.5831.80
NSLT+Luong [10, 41]31.5818.9813.2210.0032.6029.8617.5211.969.0030.70
TSPNet [37]-----36.1023.1216.8813.4134.96
CSGCR [70]35.8524.7718.6515.0838.9636.7125.4018.8615.1838.85
GASLT [67]-----39.0726.7421.8615.7439.86
GFSLT-VLP [72]44.0833.5626.7422.1243.7243.7133.1826.1121.4442.49
Ours46.8836.5929.9125.2547.2345.2134.7828.0523.4044.49
", + "bbox": [ + 181, + 101, + 787, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46125021c828286f460affc2ebd21b1e0f0bd62e7ddd23a278c96fa6406d50fd.jpg", + "table_caption": [ + "Table 2. Results on CSL-Daily dataset [73]. * means that the result was reproduced by [72]" + ], + "table_footnote": [], + "table_body": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]37.4724.6716.8611.8837.9637.3824.3616.5511.7936.74
ConSLT [22]---14.8041.46---14.5340.98
BN-TIN-Transf.+SignBT [73]51.4637.2327.5120.8049.4951.4237.2627.7621.3449.31
MMTLB [13]53.8140.8431.2924.4253.3853.3140.4130.8723.9253.25
TS-SLT [14]55.2142.3132.7125.7655.1055.4442.5932.8725.7955.72
SLTUNET [69]---23.9953.5854.9841.4431.8425.0154.08
Gloss-freeSLRT* [11]21.039.975.964.0420.5120.009.114.933.0319.67
NSLT+Luong [10, 41]34.2219.7212.247.9634.2834.1619.5711.847.5634.54
GASLT [67]-----19.909.945.984.0720.35
GFSLT-VLP [72]39.2025.0216.3511.0736.7039.3724.9316.2611.0036.44
Ours42.4526.8817.9012.2339.1839.5528.1320.0715.7539.91
", + "bbox": [ + 181, + 277, + 785, + 391 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c540f83855d106cfb32a1f9091552b5853fcc92238e614438f83ab3954d75ceb.jpg", + "table_caption": [ + "Table 3. Ablation study for impact of LLM" + ], + "table_footnote": [], + "table_body": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o LLM)29.7520.0414.9611.9527.2018.2913.3210.36
Ours (w/ T5)44.2034.5527.1522.9044.0334.1227.2322.51
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "bbox": [ + 80, + 412, + 470, + 458 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/97f63bd67aab9bbf51578b2412bc62f884e424828b94b203556f8fb46784b3a2.jpg", + "table_caption": [ + "Table 4. Ablation study for impact of SignLLM." + ], + "table_footnote": [], + "table_body": "
MethodDevTest
B1B2B3B4B1B2B3B4
Encoder Only26.9517.0412.119.4125.6316.1011.208.42
Encoder + FT39.2929.2922.5518.1839.9229.1422.5418.17
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "bbox": [ + 81, + 472, + 470, + 518 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5b551cd41ab3e1bd43bd1b0006e4f3fa6b1bda8b22144cafe05072e8b639b3de.jpg", + "table_caption": [ + "Table 5. Ablation study for main components of SignLLM." + ], + "table_footnote": [], + "table_body": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o VQ-Sign)35.7824.6518.0814.1533.1422.8016.7413.13
Ours (w/o Codebook Reconstruction)40.4530.4024.0719.7940.2530.0523.6319.47
Ours (w/o Sign-text Alignment)29.0519.3313.7210.4028.6719.2213.7310.63
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "bbox": [ + 81, + 534, + 470, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coder Only where we directly feed the outputs of the visual encoder $E_{v}$ into the LLM, and keep the LLM frozen. 2) Encoder + FT where we directly feed the outputs of the visual encoder $E_{v}$ into the LLM, and follow LLaVA [39] to fine-tune them to translate sign videos via LoRA [32]. We report the results in Tab. 4, where we significantly outperform the baselines. This shows that SignLLM is effective in harnessing off-the-shelf LLMs for the SLT task.", + "bbox": [ + 75, + 582, + 468, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Main Components of SignLLM. We also verify the impact of the key components of SignLLM by comparing against the following baselines: 1) Ours (w/o VQ-Sign) where we directly quantize the visual encoder $E_v$ 's output feature using $k$ -means algorithm. The visual encoder $E_v$ is pre-trained via the similarity loss $\\mathcal{L}^{sim}$ instead of our proposed VQ-Sign's loss $\\mathcal{L}^{VQ}$ . 2) Ours (w/o Codebook Reconstruction) where we feed the character-level sign tokens $\\hat{Z}$ from VQ-Sign directly into the LLM, while applying the sign-text alignment loss $\\mathcal{L}^{MMD}$ to the character-level sign tokens. 3) Ours (w/o Sign-text Alignment) where we do not apply the sign-text alignment loss $\\mathcal{L}^{MMD}$ . As shown in Tab. 5, removing any of our main designs leads to a signifi", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b35ec2637db3253121a7691d3dbf8cb9234e7d6e9815834dd4123450f415b776.jpg", + "image_caption": [ + "Figure 2. Visualization of translation results. Correct translations are in blue while the wrong translations are in red." + ], + "image_footnote": [], + "bbox": [ + 503, + 398, + 888, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cant performance drop, showing their efficacy.", + "bbox": [ + 498, + 527, + 807, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. We present two sample translations generated by our SignLLM and the current state-of-the-art (GFSLT-VLP [72]) in Fig. 2 for qualitative analysis. In the first sample (top), our model produces a highly accurate translation, whereas [72] inaccurately represents the semantic information. In the second sample (bottom), our model successfully preserves sentence semantics, while [72] introduces a translation error, resulting in redundant and erroneous information. These examples qualitatively show our SignLLM's efficacy in producing accurate translations.", + "bbox": [ + 496, + 545, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 717, + 619, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SignLLM, a novel framework to harness off-the-shelf and frozen LLMs for SLT. SignLLM imparts language-like characteristics to sign video representations through the VQ-Sign and CRA modules, and a sign-text alignment loss improves semantic compatibility. We empirically observe that applying our SignLLM leads to state-of-the-art gloss-free results on two popular SLT benchmarks.", + "bbox": [ + 496, + 744, + 890, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2023-121).", + "bbox": [ + 496, + 851, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18369", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bécie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford British sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2", + "[2] Ben Allison, David Guthrie, and Louise Guthrie. Another look at the data sparsity problem. In Text, Speech and Dialogue: 9th International Conference, TSD 2006, Brno, Czech Republic, September 11-15, 2006. Proceedings 9, pages 327-334. Springer, 2006. 5", + "[3] Ulrich Ammon. World languages: Trends and futures. The handbook of language and globalization, pages 101-122, 2010. 1", + "[4] Bishnu S Atal and Manfred R Schroeder. Adaptive predictive coding of speech signals. Bell System Technical Journal, 49 (8):1973-1986, 1970. 4", + "[5] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 4", + "[6] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 4", + "[7] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 8", + "[8] Kaj Bostrom and Greg Durrett. Byte pair encoding is suboptimal for language model pretraining. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4617-4624, 2020. 2, 3, 5", + "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3", + "[10] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 2, 7, 8", + "[11] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 2, 8", + "[12] Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7870-7881, 2020. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5110-5120. IEEE, 2022. 1, 2, 7, 8", + "[14] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 1, 2, 7, 8", + "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2, 3", + "[16] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5, 6", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7", + "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 4", + "[19] Shuoyang Ding, Adithya Renduchintala, and Kevin Duh. A call for prudent choice of subword merge operations in neural machine translation. In Proceedings of Machine Translation Summit XVII: Research Track, pages 204-213, 2019. 5", + "[20] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 2", + "[21] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey. arXiv preprint arXiv:2308.14177, 2, 2023. 2", + "[22] Biao Fu, Peigen Ye, Liang Zhang, Pei Yu, Cong Hu, Xiaodong Shi, and Yidong Chen. A token-level contrastive framework for sign language translation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 8", + "[23] Philip Gage. A new algorithm for data compression. C Users Journal, 12(2):23-38, 1994. 5", + "[24] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 1", + "[25] Ximena Gutierrez-Vasques, Christian Bentz, Olga Sozinova, and Tanja Samardzic. From characters to words: the turning point of bpe merges. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3454-3468, 2021. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18370", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindrich Helcl, and Alexandra Birch. Survey of low-resource machine translation. Computational Linguistics, 48 (3):673-732, 2022. 5", + "[27] Awni Hannun. Sequence modeling with ctc. Distill, 2(11): e8, 2017. 4", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 7", + "[29] Ruidan He, Linlin Liu, Hai Ye, Qingyu Tan, Bosheng Ding, Liying Cheng, Jiawei Low, Lidong Bing, and Luo Si. On the effectiveness of adapter-based tuning for pretrained language model adaptation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2208-2222, 2021. 1", + "[30] Valentin Hofmann, Janet Pierrehumbert, and Hinrich Schütze. Dagobert: Generating derivational morphology with a pretrained language model. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3848-3861, 2020. 2", + "[31] Jiahui Hou, Xiang-Yang Li, Peide Zhu, Zefan Wang, Yu Wang, Jianwei Qian, and Panlong Yang. Signspeaker: A real-time, high-precision smartwatch-based sign language translator. In The 25th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2019. 6", + "[32] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8", + "[33] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3766-3775. Association for Computational Linguistics, 2022. 7, 8", + "[34] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 7", + "[35] Sang-Ki Ko, Chang Jo Kim, Hyedong Jung, and Choongsang Cho. Neural sign language translation based on human key-point estimation. Applied sciences, 9(13):2683, 2019. 2", + "[36] W Levelt. Producing spoken language. The neurocognition of language, pages 83-122, 1999. 2, 3", + "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 1, 2, 7, 8", + "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8", + "[40] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 7", + "[41] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412-1421, 2015. 8", + "[42] Nathaniel FG Martin and James W England. Mathematical theory of entropy. Number 12. Cambridge university press, 2011. 5, 6", + "[43] Igor Melnyk, Vijil Chenthamarakshan, Pin-Yu Chen, Payel Das, Amit Dhurandhar, Inkit Padhi, and Devleena Das. Reprogramming pretrained language models for antibody sequence infilling. In Proceedings of the 2023 International Conference on Machine Learning (ICML), 2023. 1", + "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 4", + "[45] Arijit Nag, Bidisha Samanta, Animesh Mukherjee, Niloy Ganguly, and Soumen Chakrabarti. Entropy-guided vocabulary augmentation of multilingual language models for low-resource tasks. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8619-8629, 2023. 6", + "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 7", + "[48] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 6", + "[49] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 7", + "[51] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2", + "[52] Stephanie Seneff. The use of linguistic hierarchies in speech understanding. In in Proc. ICSLP, 1998. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Rico Sennrich and Biao Zhang. Revisiting low-resource neural machine translation: A case study. In 57th Annual Meeting of the Association for Computational Linguistics, pages 211–221. Association for Computational Linguistics (ACL), 2019. 5", + "[54] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1", + "[55] Xinying Song, Alex Salcianu, Yang Song, Dave Dopson, and Denny Zhou. Fast wordpiece tokenization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2089-2103, 2021. 5", + "[56] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2021. 2", + "[57] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. Advances in Neural Information Processing Systems, 29:1930-1938, 2016. 6, 7", + "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 7", + "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4", + "[60] William Vicars. American sign language: “slow”. https://www.lifeprint.com/asl101/pages-signs/s/slow.htm, 2017.6", + "[61] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 5, 6", + "[62] Andreas Voskou, Konstantinos P Panousis, Dimitrios Kosmopoulos, Dimitris N Metaxas, and Sotirios Chatzis. Stochastic transformer networks with linear competing units: Application to end-to-end sl translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11946-11955, 2021. 2, 8", + "[63] Ivan Vulic, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222-7240, 2020. 2", + "[64] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1", + "[65] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport for neural machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guage Processing (Volume 1: Long Papers), pages 7361-7373, 2021. 5, 6", + "[66] Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. Bigtrans: Augmenting large language models with multilingual translation capability over 100 languages. arXiv preprint arXiv:2305.18098, 2023. 1, 3", + "[67] Aoxiong Yin, Tianyun Zhong, Li Tang, Weike Jin, Tao Jin, and Zhou Zhao. Gloss attention for gloss-free sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2551-2562, 2023. 2, 7, 8", + "[68] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5975-5989, 2020. 1, 2", + "[69] Biao Zhang, Mathias Müller, and Rico Sennrich. SLTUNET: A simple unified model for sign language translation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 7, 8", + "[70] Jian Zhao, Weizhen Qi, Wengang Zhou, Nan Duan, Ming Zhou, and Houqiang Li. Conditional sentence generation and cross-modal reranking for sign language translation. IEEE Transactions on Multimedia, 24:2662-2672, 2021. 8", + "[71] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2", + "[72] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining, 2023. 1, 2, 7, 8", + "[73] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1316-1325. Computer Vision Foundation / IEEE, 2021. 1, 2, 7, 8", + "[74] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2022. 1, 2, 7, 8", + "[75] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774, 2021. 2", + "[76] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Jiajun Chen, Lei Li, and Shujian Huang. Multilingual machine translation with large language models: Empirical results and analysis. arXiv preprint arXiv:2304.04675, 2023.1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 811 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18372", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_model.json b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24b29d5530711ffb3e8c725bad8b97bf3e123033 --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.264, + 0.131, + 0.708, + 0.154 + ], + "angle": 0, + "content": "LLMs are Good Sign Language Translators" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.179, + 0.797, + 0.219 + ], + "angle": 0, + "content": "Jia Gong\\(^{1\\dagger}\\) Lin Geng Foo\\(^{1\\dagger}\\) Yixuan He\\(^{1\\dagger}\\) Hossein Rahmani\\(^{2}\\) Jun Liu\\(^{1‡}\\) \n\\(^{1}\\)Singapore University of Technology and Design \\(^{2}\\)Lancaster University" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.22, + 0.706, + 0.252 + ], + "angle": 0, + "content": "{jia_gong,lingeng_foo,yixuan_he}@mypmail.sutd.edu.sg, h.rahmani@lancaster.ac.uk, jun.liu@sutd.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.606 + ], + "angle": 0, + "content": "Sign Language Translation (SLT) is a challenging task that aims to translate sign videos into spoken language. Inspired by the strong translation capabilities of large language models (LLMs) that are trained on extensive multilingual text corpora, we aim to harness off-the-shelf LLMs to handle SLT. In this paper, we regularize the sign videos to embody linguistic characteristics of spoken language, and propose a novel SignLLM framework to transform sign videos into a language-like representation for improved readability by off-the-shelf LLMs. SignLLM comprises two key modules: (1) The Vector-Quantized Visual Sign module converts sign videos into a sequence of discrete character-level sign tokens, and (2) the Codebook Reconstruction and Alignment module converts these character-level tokens into word-level sign representations using an optimal transport formulation. A sign-text alignment loss further bridges the gap between sign and text tokens, enhancing semantic compatibility. We achieve state-of-the-art gloss-free results on two widely-used SLT benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.638, + 0.21, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Sign languages, which are visual signals expressed through hand, body, and facial movements, serve as the primary means of communication within the hearing-impaired community. In an effort to facilitate effective communication with this community, much attention has been directed towards developing techniques to tackle the Sign Language Translation (SLT) task [10, 11, 69, 72, 73], where the goal is to translate sign videos into spoken language. SLT is a challenging task that requires cross-modality understanding of visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72]. Despite the notable advancements in terms of network architectures [10, 11, 37], visual sign representations (e.g., with keypoint estimators [14, 74])," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.334 + ], + "angle": 0, + "content": "and training methods [22, 69, 72], how to effectively tackle the challenging cross-modal SLT task with limited paired sign-text data largely remains an open question." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.893, + 0.64 + ], + "angle": 0, + "content": "On the other hand, large language models (LLMs) – referring to language models that have been trained on a large web-scale text corpus – have recently received a lot of attention. Since LLMs are trained over a very large corpus across multiple languages with distinct syntax and lexicon, they possess rich semantic understanding and powerful linguistic abilities [9, 15, 58]. At the same time, LLMs have also demonstrated an impressive capability to translate across multiple languages [9, 15], even showing a strong potential for translating languages with limited data [66, 76]. The foundation for this translation proficiency lies in the shared linguistic properties of syntax, lexicon, and morphology that many languages hold, which is particularly evident within language families [3]. Therefore, when faced with a new language with limited data, LLMs can draw upon the wealth of knowledge acquired from previously learned languages, leveraging any shared properties of syntax, lexicon and morphology with previous languages to effectively generate translations for new languages with remarkable accuracy and fluency [66, 76]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Inspired by the impressive translation capabilities of LLMs, we aim to harness off-the-shelf LLMs to handle the challenging SLT task. However, training LLMs directly on the relatively small SLT dataset can potentially lead to forgetting of their rich knowledge [12, 29] and a decline in performance, thus we follow previous LLM-based works [24, 43, 54] to keep the off-the-shelf LLM frozen, which preserves the rich knowledge acquired during its pretraining on a vast multilingual corpus. Consequently, our focus shifts towards making the sign videos compatible and readable for the off-the-shelf and frozen LLM to perform SLT. Specifically, in this paper, we explore the following question: Can we treat the sign video as a form of language, and leverage an off-the-shelf and frozen LLM to translate them? Notably, this is not a straightforward task because directly encoding features from sign videos with a pre-trained feature extractor [28, 64] will result in a large gap between" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.345, + 0.901 + ], + "angle": 0, + "content": "† Equal contribution; ‡ Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18362" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "the sign video features and text tokens, making it difficult for off-the-shelf LLMs to understand them." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.125, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Based on the observation that LLMs can effectively handle new languages by leveraging shared commonalities with previously learned languages, we aim to introduce designs that transform our sign videos into a language-like format which are readable and friendly to LLMs. Specifically, we hypothesize that providing language-like representations of sign videos to the LLM improves the LLM's understanding of the sign videos and facilitates greater exploitation of shared properties with previously learned languages, thus resulting in better SLT performance by the LLM. To obtain language-like sign video representations, we draw inspiration from linguistic studies and analyses on LLMs [30, 63] and regularize the sign video to embody two fundamental language-like characteristics: Discrete Characteristics: Spoken languages are inherently discrete, since each language contains a finite set of words (and subwords) that convey distinct concepts, allowing them to be naturally represented through a discrete vocabulary with distinct tokens [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52]. This hierarchical structure enables languages to express a wide range of words with a limited set of characters, and convey diverse sentences with a limited number of words." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.491, + 0.473, + 0.822 + ], + "angle": 0, + "content": "In this paper, we present SignLLM, a novel framework designed to regularize input sign videos to produce sign token representations with language-like characteristics that are compatible and friendly to LLMs. Our proposed SignLLM includes two key designs to impart discrete characteristics and a hierarchical structure to the produced sign tokens. Firstly, we introduce the Vector-Quantized Visual Sign (VQ-Sign) module that facilitates the conversion of sign videos into a sequence of discrete character-level sign tokens. To achieve this, the VQ-Sign module consists of a discrete character-level sign codebook which is optimized through a self-supervised context prediction task. Next, we introduce the Codebook Reconstruction and Alignment (CRA) module that converts the character-level sign tokens into word-level sign tokens, facilitated by an optimal transport formulation. Moreover, we employ a sign-text alignment loss to further narrow the gap between the sign tokens and text tokens. These designs enable SignLLM to produce sign sentences that embody two key characteristics of spoken languages: discrete characteristics and a hierarchical structure, which enhances their compatibility with LLMs and makes them more readily interpretable by LLMs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "After producing the language-like sign sentences, we feed them into an off-the-shelf and frozen LLM along with a text prompt that instructs the LLM to generate translations in the desired language. We empirically observe that, by employing our SignLLM's designs to align sign videos with" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "languages, we can already leverage a frozen LLM to attain state-of-the-art SLT performance. These findings suggest that our proposed SignLLM framework is a promising first step towards effectively harnessing LLMs for SLT. We hope our initial explorations can inspire future work within the community to leverage LLMs for SLT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.334 + ], + "angle": 0, + "content": "In summary, our main contributions are: (1) We propose a novel SignLLM framework that is the first to harness the power of off-the-shelf and frozen LLMs for SLT. (2) To make the input sign video compatible with LLMs, our SignLLM framework incorporates two designs: a VQ-Sign module to quantize the sign video into a sequence of discrete character-level sign tokens and a CRA module that transforms the character-level sign tokens to word-level sign tokens. (3) Through our proposed designs, we achieve state-of-the-art gloss-free results on two popular SLT datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.347, + 0.642, + 0.364 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Sign Language Translation (SLT) aims to transform sign videos into natural language sentences. It is a challenging task that requires understanding of both visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72] which limits the performance of SLT methods. To improve SLT performance, many previous works [10, 11, 14, 22, 34, 35, 37, 56, 68, 69, 72, 74] aim to enhance the visual sign representations and text decoding capabilities of SLT methods. Some works propose deep architectures based on RNNs [10, 35], GCNs [34], and Transformers [11, 37, 62, 68]. Other approaches include introducing a keypoint estimator to enhance the visual sign representation [14, 35, 56, 74], introducing pre-training tasks [22, 72], or jointly modelling several SLT-related tasks [69]. Some works also introduce larger datasets (e.g., How2Sign [20] and BOBSL [1]), which present a huge challenge with their large sign and text vocabularies. Besides, some recent works [67, 72] focus on the gloss-free setting – these works do not use sign gloss annotations for training, which reduces the cost of training SLT models, and our work also falls into this category. In contrast to existing works, we aim to harness the capabilities of off-the-shelf and frozen LLMs to perform SLT, by regularizing the sign videos into a language-like representation and prompting the LLM to generate text of the desired language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Large Language Models (LLMs) refer to language models that have been trained extensively on a very large web-scale text corpus. LLMs have shown impressive text generation capabilities, attracting a lot of attention recently [21, 49, 71]. In particular, since LLMs have been trained on large amounts of text data, they demonstrate strong generalization capabilities across various text-based tasks, including code generation [51], open-domain question answering [75], and multilingual translation [9, 15]. Inspired by the re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.093, + 0.842, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.895, + 0.485 + ], + "angle": 0, + "content": "Figure 1. An overview of our SignLLM framework. During inference (top): Given an input sign video \\( X \\), we first pass it through our VQ-Sign module to obtain a sequence of discrete character-level sign tokens \\( \\hat{Z} \\). Our VQ-Sign consists of a visual encoder \\( E_{v} \\) to extract compact features and a character-level sign codebook \\( \\mathbb{S}^c \\) for quantization to obtain \\( \\hat{Z} \\). Next, we feed \\( \\hat{Z} \\) into our CRA module, which reorganizes \\( \\hat{Z} \\) by replacing short sequences of character tokens with word-level tokens via the word-level codebook, e.g., character sequence \\( [s_2, s_3, s_4] \\) to word \\( s_2 s_3 s_4 \\). This transforms the sign video data to a language-like sign sentence \\( W \\), which is fed into the LLM along with a text prompt which guides the LLM to generate translations in the desired language. During training (bottom): We optimize VQ-Sign and its discrete sign codebook via a context prediction task, which seeks to recognize the future time steps based on the current context information. Next, for our CRA module, we construct the optimal word-level codebook by considering two aspects: entropy and size, which we address using optimal transport techniques. Then, we narrow the gap between the sign token space and LLM's text token space via minimizing the MMD loss, which improves the semantic compatibility between them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.471, + 0.595 + ], + "angle": 0, + "content": "cent advancements in LLMs, we explore harnessing LLMs for translation of sign videos, by converting the sign videos to a sequence of language-like sign tokens via our Sign-LLM framework, and treating the sign tokens as a form of language that can be translated by an LLM. To the best of our knowledge, we are the first work to leverage an off-the-shelf and frozen LLM to tackle SLT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.612, + 0.17, + 0.628 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.713 + ], + "angle": 0, + "content": "In this section, we first introduce the overview of our SignLLM in Sec. 3.1. Then, we describe two main components of our SignLLM framework: the VQ-Sign and CRA modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we list the training and inference details in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.267, + 0.74 + ], + "angle": 0, + "content": "3.1. SignLLM Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.749, + 0.469, + 0.884 + ], + "angle": 0, + "content": "In order to effectively handle SLT, in this paper we draw inspiration from LLMs' remarkable capabilities in generating translations across multiple languages [9, 15]. In particular, LLMs have been extensively trained on a large web-scale multilingual text corpus and have learned diverse knowledge regarding the properties of many languages, thus they are able to draw on shared commonalities with previously learned languages to effectively handle new languages with limited data [66, 76]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Therefore, to leverage the strong translation capabilities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.612 + ], + "angle": 0, + "content": "of LLMs to handle SLT, we introduce a novel SignLLM framework. SignLLM converts the input sign video \\( X \\) into a language-like sign sentence \\( W \\) that aligns with the linguistic characteristics of spoken languages and is friendly and compatible with LLMs. Then, to perform SLT, the language-like sign sentence \\( W \\) can be fed into an off-the-shelf and frozen LLM along with a text prompt that guides the LLM to generate translations in the desired language." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Specifically, to produce sign sentences \\( W \\) that are friendly and understandable by LLMs, we aim to regularize our sign sentences \\( W \\) to embody two core linguistic characteristics: Discrete Characteristics: Spoken languages are naturally discrete and consist of distinct words or sub-words with corresponding discrete tokens in a vocabulary [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52], where words are composed from characters and sentences are composed from various words." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To achieve the above, our SignLLM framework comprises of three parts, as shown in Fig. 1: (1) The VQ-Sign module converts the input sign video \\( X \\) into a sequence of discrete sign tokens \\( \\hat{Z} \\), aligning the sign representations with text's discrete characteristics. These sign tokens \\( \\hat{Z} \\) are character-level sign tokens that are retrieved from a learned discrete character-level codebook. (2) The CRA module maps meaningful compositions of character-level sign tokens \\( \\hat{Z} \\) into word-level sign tokens that form a sign sentence" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18364" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "\\(W\\), further imparting a language-like hierarchical structure to the video sign representations. Moreover, we also align the sign token codebooks towards the text token space to improve semantic compatibility. (3) An off-the-shelf LLM takes the sign sentence \\(W\\) as input, along with an instructive text prompt that guides the LLM to generate the translation in the desired language. More details regarding the text prompt are in the Supplementary. Next, we present our VQ-Sign and CRA modules in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.243, + 0.406, + 0.259 + ], + "angle": 0, + "content": "3.2. Vector-Quantized Visual Sign Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.268, + 0.473, + 0.479 + ], + "angle": 0, + "content": "First, in order to produce a language-like representation, we would like to impart discrete characteristics to the input sign videos, aligning them more closely with spoken language representations that are inherently discrete and consist of distinct tokens in a vocabulary. However, achieving this is not straightforward because sign videos are a continuous signal in a high-dimensional spatio-temporal space which cannot be easily represented by a set of discrete tokens, and for which the vocabulary is not readily available. Hence, we introduce our Vector-Quantized Visual Sign (VQ-Sign) module to quantize the sign video \\( X \\) into a sequence of discrete sign tokens \\( \\hat{Z} \\) via a sign codebook \\( \\mathbb{S}^c \\). As illustrated in Fig. 1, our VQ-Sign module involves a series of steps, which we present in detail next." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.473, + 0.713 + ], + "angle": 0, + "content": "In the first step, we extract a compact feature \\( Z \\) from the high-dimensional input sign video \\( X \\in \\mathbb{R}^{N \\times H \\times W} \\), where \\( N \\) is the number of video frames, while \\( H \\) and \\( W \\) are the height and width of the video frames respectively. To be precise, the sign video \\( X \\) is first organized into a sequence of short overlapping video clips, then each short video clip is fed into a visual encoder \\( E_{v} \\) to extract a compact feature representation of dimensionality \\( d \\). Overall, this step transforms the original high-dimensional input sign video \\( X \\in \\mathbb{R}^{N \\times H \\times W} \\) into a compact feature \\( Z \\in \\mathbb{R}^{\\frac{N}{n} \\times d} \\), where \\( n \\) represents the number of frames between the start of neighboring clips. Notably, since \\( Z \\) is obtained by processing \\( \\frac{N}{n} \\) short clips, \\( Z \\) can also be seen as a sequence of \\( \\frac{N}{n} \\) clip-wise features, i.e., \\( \\{z_{t}\\}_{t=1}^{\\frac{N}{n}} \\), where each \\( z_{t} \\in \\mathbb{R}^{d} \\) corresponds to the feature of the \\( t \\)-th short clip." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In the next step, we transform the feature \\( Z = \\{z_{t}\\}_{t=1}^{\\frac{N}{n}} \\) into a sequence of discrete tokens \\( \\hat{Z} \\) using a codebook \\( \\mathbb{S}^c \\). Specifically, we discretize each clip's feature \\( z_{t} \\) into a discrete token \\( \\hat{z}_{t} \\) by finding the matching token \\( s_{j} \\) from the codebook \\( \\mathbb{S}^c = \\{s_i\\}_{i=1}^M \\), where the \\( i \\)-th token in the codebook is denoted as \\( s_{i} \\in \\mathbb{R}^{d} \\) and \\( M \\) is the number of tokens in the codebook. The matching token \\( s_{j} \\) is the codebook's closest element to the feature \\( z_{t} \\) in terms of Euclidean distance, i.e., \\( j = \\arg \\min_{i}(\\| z_{t} - s_{i}\\|_{2}^{2}) \\). After the matching, each feature \\( z_{t} \\) is replaced by \\( \\hat{z}_{t} = s_{j} \\), which results in a discrete token sequence as shown in Fig. 1, e.g., \\( [s_2, s_3, s_4, s_M, s_5, s_1] \\). Note that, we randomly initialize all" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "the tokens \\(\\{s_i\\}_{i=1}^M\\) in the sign codebook \\(\\mathbb{S}^c\\) at the start and optimize them during training, as introduced next." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.893, + 0.394 + ], + "angle": 0, + "content": "However, we face a challenge in learning the discrete codebook \\(\\mathbb{S}^c\\). In particular, although autoencoding [59] has been a popular method for producing a codebook of discrete units, the high complexity of sign videos makes autoencoding (i.e., self-reconstruction of sign videos) challenging and costly. Therefore, inspired from predictive coding [4], a widely-used method in text and speech representation learning [6, 44, 46], we propose to learn discrete representations of sign videos through a context prediction task. Context prediction [46] is a self-supervised task that focuses on recognizing the future content in latent space based on the current information, which can learn discrete representations while eliminating the need for reconstructing the high-dimensional input video data. Furthermore, previous works show that training with context prediction effectively captures the temporal dependencies and relationships between elements in a sequence [5, 27], and the learned representations are often transferable to downstream tasks [6, 46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.894, + 0.547 + ], + "angle": 0, + "content": "Specifically, we employ a context prediction task where we try to distinguish a future sample \\( z_{\\tau + k} \\) based on the current context representation \\( c_{\\tau} \\) at various time steps \\( \\tau \\). To facilitate this task, after we obtain the discrete token sequence \\( \\hat{Z} \\), we further produce a context latent representation \\( c_{\\tau} \\) using an auto-regressive model \\( g \\) that summarizes all the discrete tokens before a certain time step \\( \\tau \\) (i.e., \\( \\{\\hat{z}_t\\}_{t \\leq \\tau} \\)) to produce context latent representations \\( c_{\\tau} = g(\\{\\hat{z}_t\\}_{t \\leq \\tau}) \\). Then, we optimize our module by minimizing the following context prediction contrastive loss \\( \\mathcal{L}_k^{\\mathrm{cp}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.552, + 0.892, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {k} ^ {\\mathrm {c p}} = - \\sum_ {\\tau = 1} ^ {\\frac {N}{n} - k} \\left(\\log \\sigma \\left(z _ {\\tau + k} ^ {\\top} h _ {\\tau}\\right) + \\lambda \\underset {\\widetilde {z} \\sim p _ {n}} {\\mathbb {E}} \\left[ \\log \\sigma \\left(- \\widetilde {z} ^ {\\top} h _ {\\tau}\\right) \\right]\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.709 + ], + "angle": 0, + "content": "where \\(h_\\tau\\) is obtained by applying a trainable linear layer to \\(c_{\\tau}\\), \\(\\sigma(z_{\\tau+k}^{\\top} h_{\\tau})\\) is the probability of \\(z_{\\tau+k}\\) being the true sample among the negatives, \\(\\widetilde{z}\\) are negative samples drawn from a minibatch \\(p_n\\), and \\(\\lambda\\) is a hyperparameter. We sum \\(\\mathcal{L}_k^{\\mathrm{cp}}\\) over different step sizes \\(k\\) to obtain the context prediction loss \\(\\mathcal{L}^{\\mathrm{cp}} = \\sum_{k=1}^{K} \\mathcal{L}_k^{\\mathrm{cp}}\\), where \\(K\\) is the maximum number of future clips that we are interested in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.71, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Following [18], in order to optimize the matching between \\(\\hat{z}_t\\) and \\(z_{t}\\), we further add two losses to optimize the matching distance between \\(\\hat{z}_t\\) and \\(z_{t}\\), such that the overall loss \\(\\mathcal{L}^{VQ}\\) to optimize our VQ-Sign module is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {V Q} = \\mathcal {L} ^ {\\mathrm {c p}} + \\sum_ {t = 1} ^ {N / n} \\| \\operatorname {s g} \\left(z _ {t}\\right) - \\hat {z} _ {t} \\| ^ {2} + \\gamma \\sum_ {t = 1} ^ {N / n} \\| z _ {t} - \\operatorname {s g} (\\hat {z} _ {t}) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\operatorname{sg}(z) \\equiv z\\) is the stop-gradient operator and \\(\\gamma\\) is a hyperparameter. By optimizing \\(\\mathcal{L}^{VQ}\\), we can train our VQ-Sign and the discrete codebook without the need for reconstructing high-dimensional video clips, which makes the codebook construction a viable and relatively cheap option." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18365" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "In summary, our VQ-Sign transforms sign videos into sequences of discrete sign tokens \\(\\hat{Z}\\), which are friendlier and more understandable to LLMs. Notably, the produced discrete tokens \\(\\hat{Z}\\) can be likened to character-level tokens, in the sense that each discrete token \\(\\hat{z}_t\\) corresponds to a short clip and may not contain much semantic meaning on its own (similar to linguistic characters), but they can be combined into a sequence to convey a clear semantic meaning (akin to forming a word or sentence). Thus, inspired by this, we call VQ-Sign's codebook \\(\\mathbb{S}^c\\) the character-level codebook that contains character-level sign tokens." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.27, + 0.433, + 0.288 + ], + "angle": 0, + "content": "3.3. Codebook Reconstruction and Alignment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.294, + 0.47, + 0.43 + ], + "angle": 0, + "content": "In the previous section, we quantize sign videos into discrete character-level sign tokens, which aligns them closer to language representations. In this section, our goal is to impart a hierarchical structure to our sign video representations, which makes them align even closer to language representations. Specifically, we aim to compose our character-level sign tokens into word-level sign tokens to mirror the observed hierarchical structure in spoken language, which makes them even more compatible with LLMs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Intuitively, considering a spoken language sentence, we can represent it as a sequence of words, with each word formed by one or multiple characters. For example, a sentence 'I love AI' can be decomposed into a word sequence [I', 'love', 'AI'], where the word 'love' is in turn formed by the character sequence [1', o', v', e']. We observe that, although each individual character may not contain much semantic meaning on its own, they can be composed to form words with clearer semantic meaning. In a similar fashion, we also want to impart such hierarchical structure to our character-level sign tokens by composing them to form meaningful word-level sign tokens." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.613, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Hence, we aim to find an optimal transformation from character-level sign tokens to word-level sign tokens for enhanced readability and compatibility with LLMs. To this end, we introduce the Codebook Reconstruction and Alignment (CRA) module to transform the character-level codebook \\(\\mathbb{S}^c\\) from VQ-Sign into a word-level codebook \\(\\mathbb{S}^w\\) whose tokens convey richer and clearer semantic meaning. Inspired by optimal transport methods [16, 61, 65], we observe that the above transformation can be formulated as an optimal transport problem of transporting characters into words, thus we introduce a codebook reconstruction algorithm with an optimal transport formulation to find an optimal transformation. Additionally, to further reduce the distribution gap between the sign tokens and the text tokens, our CRA module also performs sign-text alignment, enhancing the semantic compatibility of the sign tokens with LLMs. We introduce the details below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To begin, the objective of our Codebook Reconstruction Algorithm is to create a word-level codebook \\(\\mathbb{S}^w\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.455 + ], + "angle": 0, + "content": "based on VQ-Sign's character-level codebook \\(\\mathbb{S}^c\\). The challenge lies in determining which character-level sign tokens should be assembled together to form word-level sign tokens, which is a complex problem. To address this complexity, we adopt an approach based on two fundamental principles. Firstly, in order to maximize the overall predictability of the word-level tokens and enhance the distinctiveness of each token, we seek to minimize the entropy of each word-level token within the vocabulary [42]. We remark that, several approaches for establishing language-based subword vocabularies [8, 55] can be seen as entropy-minimizing approaches, employing different heuristics to establish the vocabulary with the goal of minimizing entropy [23]. On the other hand, considering the limited availability of sign video data, we incorporate codebook size as another key factor in our word-level codebook construction, since studies on languages with limited data [26, 53] have also identified vocabulary size as a crucial aspect. Specifically, too small a vocabulary can result in sub-optimal entropy values, while an excessively large vocabulary size can lead to issues such as parameter explosion and token sparsity which hinder understanding [2], and finding the right balance between these effects becomes even more sensitive for languages with limited data [19, 53]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Based on these principles, our objective is to determine an optimal codebook size that maximizes the entropy reduction while taking into account the increase in codebook size. In other words, we would like to find an optimal codebook size that maximizes the gradient of the entropy reduction with respect to the codebook size increase. To simplify the optimal size searching problem, we define a fixed size increment \\( m \\) and search through codebooks of various sizes (where the difference between each codebook size is \\( m \\) tokens). Specifically, we define the \\( r \\)-th codebook \\( (\\mathbb{S}_r^w) \\) as the codebook with \\( r \\times m \\) tokens. Then, we seek to identify the optimal set of word-level tokens, where each word-level token is composed of character-level tokens. We approach this by formulating the character compositions as an optimal transport problem, where characters are transported to words." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, it can be challenging to identify specific character combinations that convey precise semantic information, due to the temporal complexity of sign videos, which often makes the character-level token sequences \\(\\hat{Z}\\) quite messy. For instance, some signers may execute the signing motions at a slower speed, which can lead to consecutive short video clips being highly similar, resulting in consecutively-repeating character-level discrete tokens. Thus, the character-level sequences between different signers can differ significantly (e.g., \\([s_1,s_2]\\) vs \\([s_1,s_1,s_1,s_2,s_2]\\)) due to such duplication of character-level tokens, even though they may contain the same semantic information. At the same time, simply filtering out re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18366" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "peated character-level tokens straightforwardly (e.g., setting all \\([s_1, s_1, s_1, s_2, s_2]\\) to \\([s_1, s_2]\\)) is sub-optimal, since the speed of some signs can also convey some information [31, 60], e.g., if a signer signs \"ugly\" quickly, it conveys \"very ugly\" in American Sign Language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.437 + ], + "angle": 0, + "content": "Pre-processing of Repeated Characters. Therefore, to alleviate the impact of the signer's speed while keeping the information regarding each sign's speed, we first preprocess the character-level sequences as follows: First, we find all the repeated tokens in the character-level sequence and compute the average number of repeated tokens \\((\\alpha)\\) in each sequence. Then, for each repeated sequence (e.g., \\([s_1, s_1, s_1]\\)), we keep the first character and remove the tailing repetitive ones (e.g., \\([s_1, s_1, s_1]\\) to \\([s_1]\\)). At the same time, if the character-level tokens repeat more than \\(\\alpha\\) times, we insert a single character-level token \\(s_0\\) as a \"slowing down\" sign, e.g., \\([s_1, s_1, s_1]\\) to \\([s_1, s_0]\\) if \\(\\alpha < 3\\). Crucially, this allows us to reduce redundancy, while still representing \"quick\" or \"slow\" signs that account for differences in the signer's speed. Overall, this pre-processing and reducing of repeating characters makes the character-level sequence less messy, facilitating the search for specific meaningful character combinations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.47, + 0.56 + ], + "angle": 0, + "content": "Optimal Transport Formulation. Then, using the preprocessed character-level sequences, we search for an optimal word-level codebook via an optimal transport formulation [16, 61, 65]. Following our discussion above, we aim to find an optimal word-level codebook (i.e., \\(\\mathbb{S}_r^w\\)) with low entropy and compact size. Specifically, to measure codebook entropy, we follow [25, 42, 45] to define the entropy of the \\(r\\)-th codebook \\(\\mathbb{S}_r^w\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.565, + 0.469, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {\\mathbb {S} _ {r} ^ {w}} = - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} P (w _ {j}) \\log P (w _ {j}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.47, + 0.675 + ], + "angle": 0, + "content": "where \\(P(w_{j})\\) is the relative frequency of the \\(j^{th}\\) token \\(w_{j}\\) from the word-level codebook \\(\\mathbb{S}_r^w\\). Then, based on VQ-Sign's character-level codebook \\(\\mathbb{S}^c\\), the entropy of the word-level codebook \\(\\mathbb{S}_r^w\\) can be computed through the following (with proof in Supplementary):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.677, + 0.469, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal{H}_{\\mathbb{S}_{r}^{w}} = -\\sum_{w_{j}\\in \\mathbb{S}^{w}}\\sum_{s_{i}\\in \\mathbb{S}^{c}}P(w_{j},s_{i})\\log P(w_{j},s_{i}) \\\\ - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} \\sum_ {s _ {i} \\in \\mathbb {S} ^ {c}} P \\left(w _ {j}, s _ {i}\\right) (- \\log P \\left(s _ {i} \\mid w _ {j}\\right)), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.78 + ], + "angle": 0, + "content": "where \\( P(s_{i}|w_{j}) \\) is the probability of the character-level token \\( s_i \\) appearing in the word-level token \\( w_{j} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Then, to formally define our objective function, we follow existing works [16, 48, 65] to define a transport matrix \\( \\pmb{P} \\) that represents the assignments of characters to words, and a distance matrix \\( \\pmb{D} \\) that represents the cost of transportation. We define the transport matrix \\( \\pmb{P} \\in \\mathbb{R}^{m \\times (r \\cdot m)} \\) with the \\( (j,i) \\)-th element as \\( P(w_{j}, s_{i}) \\), and define the distance matrix \\( \\pmb{D} \\in \\mathbb{R}^{m \\times (r \\cdot m)} \\) as a matrix whose \\( (j,i) \\)-th element is \\( \\log P(s_{i} | w_{j}) \\). Note that, if \\( w_{j} \\) contains \\( s_{i} \\), we use" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.171 + ], + "angle": 0, + "content": "\\(\\frac{1}{\\mathrm{length}(w_j)}\\) to estimate \\(P(s_i|w_j)\\), and if \\(w_j\\) does not contain \\(s_i\\), then it is deemed an infeasible assignment, and we set the distance \\(\\log P(s_i|w_j) = \\infty\\). We further define \\(H(\\pmb{P})\\) as \\(-\\sum_{w_j \\in \\mathbb{S}_r^w} \\sum_{s_i \\in \\mathbb{S}^c} P(w_j, s_i) \\log P(w_j, s_i)\\), which is simply the entropy of the probability distribution \\(P(w_j, s_i)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.171, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Hence, based on Eq. 4, the objective function for minimizing the entropy of \\(\\mathbb{S}_r^w\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.207, + 0.892, + 0.237 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {P} \\in \\mathbb {R} ^ {m} \\times (r \\cdot m)} {\\arg \\min } H (\\boldsymbol {P}) + \\sum_ {j} \\sum_ {i} \\boldsymbol {P} (j, i) \\boldsymbol {D} (j, i). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.361 + ], + "angle": 0, + "content": "Specifically, following previous works [16, 48, 65], we impose two constraints on the transport matrix \\( P \\): the sum of each row in \\( P \\) should equal to the probability of character token \\( s_i \\) and the sum of each column in \\( P \\) should equal to the probability of word token \\( w_j \\). Formally, we constrain the transport matrix \\( P \\) with: \\( |\\sum_{i}P(i,j) - P(w_{j})|\\leq \\epsilon \\) and \\( |\\sum_{j}P(i,j) - P(s_{i})|\\leq \\epsilon \\), where \\( \\epsilon \\) is a small positive fixed hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Intuitively, this optimization process can be regarded as an optimal transport problem to find the best way to transport mass from \\(\\mathbb{S}^c\\) to \\(\\mathbb{S}_r^w\\). To handle this optimal transport problem, we leverage the Sinkhorn algorithm [16, 48, 65], allowing us to effectively construct the candidate word-level codebooks \\(\\mathbb{S}_r^w\\) with minimal entropy. Since our increment \\(m\\) between each candidate codebook is fixed, we can find the optimal codebook size that maximizes the gradient of entropy reduction by simply computing and finding the maximum the entropy difference between \\(\\mathbb{S}_r^w\\) and \\(\\mathbb{S}_{r - 1}^w\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.892, + 0.619 + ], + "angle": 0, + "content": "After finding the optimal word-level codebook, we construct the word-level sign tokens by composing all features of the character-level tokens into the word-level tokens via our autoregressive model \\( g \\). Refer to Supplementary for more details. Overall, with our codebook reconstruction algorithm, we can construct an optimal word-level sign codebook with low entropy yet also with relatively small size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Sign-Text Alignment. Next, we further align the sign tokens with the text tokens used in LLMs in order to further improve semantic compatibility between them. To achieve this, we measure the distribution gap between sign tokens and text tokens via Maximum Mean Discrepancy (MMD) [57] and then optimize the sign tokens' embeddings by minimizing MMD, which narrows down the distribution gap. Specifically, we compute the gap between the sign embedding space \\(\\mathbb{F}^s\\) and text embedding space \\(\\mathbb{F}^t\\) via MMD as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.768, + 0.891, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} ^ {M M D} (\\mathbb {F} ^ {s}, \\mathbb {F} ^ {t}) = \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {s}} \\frac {k (f (p _ {i}) , f (p _ {j}))}{n _ {s} ^ {2}} \\\\ + \\sum_ {i = 1} ^ {n _ {t}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {k \\left(q _ {i} , q _ {j}\\right)}{n _ {t} ^ {2}} - \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {2 \\cdot k \\left(f \\left(p _ {i}\\right) , q _ {j}\\right)}{n _ {s} n _ {t}}, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(p\\) and \\(q\\) are the tokens in \\(\\mathbb{F}^s\\) and \\(\\mathbb{F}^t\\), \\(f\\) is a small projection module that projects \\(\\mathbb{F}^s\\) to \\(\\mathbb{F}^t\\), \\(n_s\\) and \\(n_t\\) are the numbers of tokens in \\(\\mathbb{F}^s\\) and \\(\\mathbb{F}^t\\), and \\(k(\\cdot)\\) represents the radial" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18367" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "kernel [57] that measures the distance between two samples. We apply MMD loss to both the word-level and the character-level tokens to narrow the overall sign-text gap." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.148, + 0.292, + 0.163 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.172, + 0.468, + 0.323 + ], + "angle": 0, + "content": "Inference. Given a sign video \\( X \\), we first extract compact features \\( Z \\) via the visual encoder \\( E_{v} \\), and quantize \\( Z \\) to \\( \\hat{Z} \\) via VQ-Sign's learned character-level codebook. Then, we transform the sequence of discrete character-level tokens \\( \\hat{Z} \\) into word-level tokens via our CRA, which produces a sign sentence \\( W \\). Lastly, we project the sign sentence \\( W \\) to LLM embedding space via the small projection module \\( f \\), and then feed the sign sentence \\( W \\) into the LLM along with a text prompt to instruct the LLM to perform the SLT task and generate text in the desired language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.468, + 0.474 + ], + "angle": 0, + "content": "Training. Our SignLLM is optimized in two stages: pretraining and fine-tuning. Specifically, the pre-training stage, which does not require explicit SLT supervision, includes two sub-stages: (i) We first pre-train VQ-Sign via the context prediction task with \\(\\mathcal{L}^{VQ}\\) in Eq. 2. (ii) Then, based on the VQ-Sign's learned character-level codebook, we construct the word-level codebook with the codebook reconstruction algorithm (Sec. 3.3) and apply the MMD loss \\(\\mathcal{L}^{MMD}\\) in Eq. 6 to align the sign codebooks \\((\\mathbb{S}^c\\) and \\(\\mathbb{S}^w)\\) and the text vocabulary of the desired language." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.468, + 0.626 + ], + "angle": 0, + "content": "After the pre-training, we fine-tune SignLLM. To aid LLMs in understanding the sign sentences as texts, we additionally maximize the similarity between the text tokens \\( Y \\) generated by LLM and the ground truth tokens \\( \\bar{Y} \\) as: \\( \\mathcal{L}^{sim} \\) by minimizing the cross-entropy between them. We fine-tune our SignLLM (with frozen LLM) via the loss \\( \\mathcal{L}^{ft} \\) as follows: \\( \\mathcal{L}^{ft} = \\mathcal{L}^{VQ} + \\lambda_1\\mathcal{L}^{MMD} + \\lambda_2\\mathcal{L}^{sim} \\), where \\( \\lambda_{1} \\) and \\( \\lambda_{2} \\) are hyperparameters. Note that we follow previous gloss-free works [67, 72] to train our SignLLM framework, eliminating the need for additional gloss data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.21, + 0.657 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.295, + 0.682 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our visual encoder \\( E_{v} \\) is constructed by appending two Conv3D layers with a kernel size of (5, 3, 3) and a stride of (2, 1, 1) to a ResNet18 [28] pre-trained on ImageNet [17]. Each clip consists of 13 frames and the gap between the neighboring clips (\\( n \\)) is 4. Besides, the auto-regressive model \\( g \\) is implemented as a Convolutional Gated Recurrent Layer with a kernel size of (1, 1). We set the total number of discrete vectors \\( M \\) at 256 and each vector's dimension \\( d \\) at 1024 for our character-level codebook. In VQ-Sign's pre-training phase, we set \\( \\gamma = 0.25 \\), initialize the learning rate at 0.01 and use the Adam algorithm, training the model to predict the future three clips (\\( K = 3 \\)) for 200 epochs. During codebook reconstruction, we set the increment \\( m \\) to 32. We employ the frozen LLaMA-7B-16bit [58]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "as our LLM and project the codebook space to LLaMA's embedding space via \\( f \\), which consists of two fc-layers with ReLU. We set \\( \\lambda_1 = 0.5 \\), \\( \\lambda_2 = 1 \\), and initialize the learning rate at 0.001 to fine-tune our SignLLM over 20 epochs. Please see Supplementary for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.177, + 0.787, + 0.192 + ], + "angle": 0, + "content": "4.2. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.335 + ], + "angle": 0, + "content": "Datasets. We follow previous works [13, 14, 34, 72-74] to run experiments on the Phoenix-2014T [10] and CSL-Daily [73] datasets for SLT, and evaluate on their dev and test sets. Phoenix-2014T [10] is a German sign language dataset with a vocabulary size of 2887 German words. The training, dev, and test sets contain 7096, 519, and 642 samples. CSL-Daily [73] is a Chinese sign language dataset with a vocabulary size of 2343 Chinese words. The training, dev, and test sets contain 18401, 1077, and 1176 samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous works [13, 14, 37, 67, 72-74], we adopt BLEU [47] and ROUGE-L [38] as the evaluation metrics for SLT. BLEU-n evaluates the average translation precision up to n-grams, and we follow previous works [13, 14, 72-74] to report results for BLEU-1 to BLEU-4 (i.e., B1, B2, B3 and B4). ROUGE-L (or ROUGE) computes the F1 score based on the longest common subsequence between the predicted and ground truth texts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.467, + 0.64, + 0.481 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Results on PHOENIX2014T dataset. Tab. 1 presents a comparison of our approach with state-of-the-art gloss-based and gloss-free methods for SLT. Our method consistently improves upon all reported metrics as compared to other gloss-free approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Results on CSL-Daily dataset. We compare our method with state-of-the-art approaches on the CSLD Daily dataset in Tab. 2. We outperform previous gloss-free works on all metrics, showing the efficacy of our approach." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.655, + 0.652 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.719 + ], + "angle": 0, + "content": "To further investigate the proposed method, we follow previous works [13, 33, 67, 69, 72] to conduct extensive ablation experiments on the Phoenix-2014T dev and test sets. Refer to Supplementary for more experiment results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Impact of LLM. First, we evaluate the impact of leveraging LLMs for the SLT task. Specifically, we establish three baselines: 1) Ours (w/o LLM) where we replace the LLM with a trainable lightweight text generator (mBART [40]). 2) Ours (w/ T5) where we replace our LLM with a smaller LLM (T5 [50]). As shown in Tab. 3, our approach (w/ LLaMA [58]) achieves a much better performance than Ours (w/o LLM), showing the efficacy of leveraging a powerful LLM. We also find that using a smaller and less powerful LLM (T5) leads to worse performance, as expected." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Impact of SignLLM. Next, we explore the impact of Sign-LLM by comparing against the following baselines: 1) En" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18368" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.341, + 0.09, + 0.628, + 0.102 + ], + "angle": 0, + "content": "Table 1. Results on Phoenix-2014T dataset [10]." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.102, + 0.788, + 0.266 + ], + "angle": 0, + "content": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]47.2634.4027.0522.38-46.6133.7326.1921.32-
ConSLT [22]---21.1147.74---21.5947.69
STN-SLT [62]49.1236.2928.3423.23-48.6135.9728.3723.65-
STMC-T [74]47.6036.4329.1824.0948.2446.9836.0928.7023.6546.65
BN-TIN-Transf.+SignBT [73]51.1137.9029.8024.4550.2950.8037.7529.7224.3249.54
PET [33]-----49.5437.1929.3024.0249.97
MMTLB [13]53.9541.1233.1427.6153.1053.9741.7533.8428.3952.65
TS-SLT [14]54.3241.9934.1528.6654.0854.9042.4334.4628.9553.48
SLTUNET [69]---27.8752.2352.9241.7633.9928.4752.11
Gloss-freeNSLT [10]28.1016.8111.829.1231.0027.1015.6110.828.3529.70
NSLT+Bahdanau [7, 10]31.8719.1113.169.9431.8032.2419.0312.839.5831.80
NSLT+Luong [10, 41]31.5818.9813.2210.0032.6029.8617.5211.969.0030.70
TSPNet [37]-----36.1023.1216.8813.4134.96
CSGCR [70]35.8524.7718.6515.0838.9636.7125.4018.8615.1838.85
GASLT [67]-----39.0726.7421.8615.7439.86
GFSLT-VLP [72]44.0833.5626.7422.1243.7243.7133.1826.1121.4442.49
Ours46.8836.5929.9125.2547.2345.2134.7828.0523.4044.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.216, + 0.267, + 0.756, + 0.279 + ], + "angle": 0, + "content": "Table 2. Results on CSL-Daily dataset [73]. * means that the result was reproduced by [72]" + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.279, + 0.787, + 0.392 + ], + "angle": 0, + "content": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]37.4724.6716.8611.8837.9637.3824.3616.5511.7936.74
ConSLT [22]---14.8041.46---14.5340.98
BN-TIN-Transf.+SignBT [73]51.4637.2327.5120.8049.4951.4237.2627.7621.3449.31
MMTLB [13]53.8140.8431.2924.4253.3853.3140.4130.8723.9253.25
TS-SLT [14]55.2142.3132.7125.7655.1055.4442.5932.8725.7955.72
SLTUNET [69]---23.9953.5854.9841.4431.8425.0154.08
Gloss-freeSLRT* [11]21.039.975.964.0420.5120.009.114.933.0319.67
NSLT+Luong [10, 41]34.2219.7212.247.9634.2834.1619.5711.847.5634.54
GASLT [67]-----19.909.945.984.0720.35
GFSLT-VLP [72]39.2025.0216.3511.0736.7039.3724.9316.2611.0036.44
Ours42.4526.8817.9012.2339.1839.5528.1320.0715.7539.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.398, + 0.402, + 0.411 + ], + "angle": 0, + "content": "Table 3. Ablation study for impact of LLM" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.413, + 0.472, + 0.459 + ], + "angle": 0, + "content": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o LLM)29.7520.0414.9611.9527.2018.2913.3210.36
Ours (w/ T5)44.2034.5527.1522.9044.0334.1227.2322.51
Ours46.8836.5929.9125.2545.2134.7828.0523.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.459, + 0.418, + 0.471 + ], + "angle": 0, + "content": "Table 4. Ablation study for impact of SignLLM." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.473, + 0.471, + 0.52 + ], + "angle": 0, + "content": "
MethodDevTest
B1B2B3B4B1B2B3B4
Encoder Only26.9517.0412.119.4125.6316.1011.208.42
Encoder + FT39.2929.2922.5518.1839.9229.1422.5418.17
Ours46.8836.5929.9125.2545.2134.7828.0523.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.521, + 0.449, + 0.533 + ], + "angle": 0, + "content": "Table 5. Ablation study for main components of SignLLM." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.535, + 0.471, + 0.58 + ], + "angle": 0, + "content": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o VQ-Sign)35.7824.6518.0814.1533.1422.8016.7413.13
Ours (w/o Codebook Reconstruction)40.4530.4024.0719.7940.2530.0523.6319.47
Ours (w/o Sign-text Alignment)29.0519.3313.7210.4028.6719.2213.7310.63
Ours46.8836.5929.9125.2545.2134.7828.0523.40
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.704 + ], + "angle": 0, + "content": "coder Only where we directly feed the outputs of the visual encoder \\( E_{v} \\) into the LLM, and keep the LLM frozen. 2) Encoder + FT where we directly feed the outputs of the visual encoder \\( E_{v} \\) into the LLM, and follow LLaVA [39] to fine-tune them to translate sign videos via LoRA [32]. We report the results in Tab. 4, where we significantly outperform the baselines. This shows that SignLLM is effective in harnessing off-the-shelf LLMs for the SLT task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Impact of Main Components of SignLLM. We also verify the impact of the key components of SignLLM by comparing against the following baselines: 1) Ours (w/o VQ-Sign) where we directly quantize the visual encoder \\(E_v\\)'s output feature using \\(k\\)-means algorithm. The visual encoder \\(E_v\\) is pre-trained via the similarity loss \\(\\mathcal{L}^{sim}\\) instead of our proposed VQ-Sign's loss \\(\\mathcal{L}^{VQ}\\). 2) Ours (w/o Codebook Reconstruction) where we feed the character-level sign tokens \\(\\hat{Z}\\) from VQ-Sign directly into the LLM, while applying the sign-text alignment loss \\(\\mathcal{L}^{MMD}\\) to the character-level sign tokens. 3) Ours (w/o Sign-text Alignment) where we do not apply the sign-text alignment loss \\(\\mathcal{L}^{MMD}\\). As shown in Tab. 5, removing any of our main designs leads to a signifi" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.4, + 0.89, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.488, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Figure 2. Visualization of translation results. Correct translations are in blue while the wrong translations are in red." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.808, + 0.544 + ], + "angle": 0, + "content": "cant performance drop, showing their efficacy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.698 + ], + "angle": 0, + "content": "Qualitative Results. We present two sample translations generated by our SignLLM and the current state-of-the-art (GFSLT-VLP [72]) in Fig. 2 for qualitative analysis. In the first sample (top), our model produces a highly accurate translation, whereas [72] inaccurately represents the semantic information. In the second sample (bottom), our model successfully preserves sentence semantics, while [72] introduces a translation error, resulting in redundant and erroneous information. These examples qualitatively show our SignLLM's efficacy in producing accurate translations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.718, + 0.62, + 0.733 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.892, + 0.851 + ], + "angle": 0, + "content": "We present SignLLM, a novel framework to harness off-the-shelf and frozen LLMs for SLT. SignLLM imparts language-like characteristics to sign video representations through the VQ-Sign and CRA modules, and a sign-text alignment loss improves semantic compatibility. We empirically observe that applying our SignLLM leads to state-of-the-art gloss-free results on two popular SLT benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2023-121)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18369" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bécie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford British sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Ben Allison, David Guthrie, and Louise Guthrie. Another look at the data sparsity problem. In Text, Speech and Dialogue: 9th International Conference, TSD 2006, Brno, Czech Republic, September 11-15, 2006. Proceedings 9, pages 327-334. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Ulrich Ammon. World languages: Trends and futures. The handbook of language and globalization, pages 101-122, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Bishnu S Atal and Manfred R Schroeder. Adaptive predictive coding of speech signals. Bell System Technical Journal, 49 (8):1973-1986, 1970. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[5] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[6] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.471, + 0.544 + ], + "angle": 0, + "content": "[7] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.547, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[8] Kaj Bostrom and Greg Durrett. Byte pair encoding is suboptimal for language model pretraining. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4617-4624, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.604, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[12] Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7870-7881, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[13] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5110-5120. IEEE, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[16] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.39 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.449, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[19] Shuoyang Ding, Adithya Renduchintala, and Kevin Duh. A call for prudent choice of subword merge operations in neural machine translation. In Proceedings of Machine Translation Summit XVII: Research Track, pages 204-213, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[20] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[21] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey. arXiv preprint arXiv:2308.14177, 2, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[22] Biao Fu, Peigen Ye, Liang Zhang, Pei Yu, Cong Hu, Xiaodong Shi, and Yidong Chen. A token-level contrastive framework for sign language translation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[23] Philip Gage. A new algorithm for data compression. C Users Journal, 12(2):23-38, 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Ximena Gutierrez-Vasques, Christian Bentz, Olga Sozinova, and Tanja Samardzic. From characters to words: the turning point of bpe merges. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3454-3468, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18370" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindrich Helcl, and Alexandra Birch. Survey of low-resource machine translation. Computational Linguistics, 48 (3):673-732, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.175 + ], + "angle": 0, + "content": "[27] Awni Hannun. Sequence modeling with ctc. Distill, 2(11): e8, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.236, + 0.468, + 0.345 + ], + "angle": 0, + "content": "[29] Ruidan He, Linlin Liu, Hai Ye, Qingyu Tan, Bosheng Ding, Liying Cheng, Jiawei Low, Lidong Bing, and Luo Si. On the effectiveness of adapter-based tuning for pretrained language model adaptation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2208-2222, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[30] Valentin Hofmann, Janet Pierrehumbert, and Hinrich Schütze. Dagobert: Generating derivational morphology with a pretrained language model. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3848-3861, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[31] Jiahui Hou, Xiang-Yang Li, Peide Zhu, Zefan Wang, Yu Wang, Jianwei Qian, and Panlong Yang. Signspeaker: A real-time, high-precision smartwatch-based sign language translator. In The 25th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.468, + 0.545 + ], + "angle": 0, + "content": "[32] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.548, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[33] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3766-3775. Association for Computational Linguistics, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.468, + 0.713 + ], + "angle": 0, + "content": "[34] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[35] Sang-Ki Ko, Chang Jo Kim, Hyedong Jung, and Choongsang Cho. Neural sign language translation based on human key-point estimation. Applied sciences, 9(13):2683, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.759, + 0.468, + 0.786 + ], + "angle": 0, + "content": "[36] W Levelt. Producing spoken language. The neurocognition of language, pages 83-122, 1999. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[40] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[41] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412-1421, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[42] Nathaniel FG Martin and James W England. Mathematical theory of entropy. Number 12. Cambridge university press, 2011. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[43] Igor Melnyk, Vijil Chenthamarakshan, Pin-Yu Chen, Payel Das, Amit Dhurandhar, Inkit Padhi, and Devleena Das. Reprogramming pretrained language models for antibody sequence infilling. In Proceedings of the 2023 International Conference on Machine Learning (ICML), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.396, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[45] Arijit Nag, Bidisha Samanta, Animesh Mukherjee, Niloy Ganguly, and Soumen Chakrabarti. Entropy-guided vocabulary augmentation of multilingual language models for low-resource tasks. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8619-8629, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.623 + ], + "angle": 0, + "content": "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.681 + ], + "angle": 0, + "content": "[48] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[49] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.73, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.868 + ], + "angle": 0, + "content": "[51] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Stephanie Seneff. The use of linguistic hierarchies in speech understanding. In in Proc. ICSLP, 1998. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18371" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Rico Sennrich and Biao Zhang. Revisiting low-resource neural machine translation: A case study. In 57th Annual Meeting of the Association for Computational Linguistics, pages 211–221. Association for Computational Linguistics (ACL), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[54] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[55] Xinying Song, Alex Salcianu, Yang Song, Dave Dopson, and Denny Zhou. Fast wordpiece tokenization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2089-2103, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[56] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[57] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. Advances in Neural Information Processing Systems, 29:1930-1938, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.47, + 0.518 + ], + "angle": 0, + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[60] William Vicars. American sign language: “slow”. https://www.lifeprint.com/asl101/pages-signs/s/slow.htm, 2017.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.59 + ], + "angle": 0, + "content": "[61] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[62] Andreas Voskou, Konstantinos P Panousis, Dimitrios Kosmopoulos, Dimitris N Metaxas, and Sotirios Chatzis. Stochastic transformer networks with linear competing units: Application to end-to-end sl translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11946-11955, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[63] Ivan Vulic, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222-7240, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[64] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[65] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport for neural machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "guage Processing (Volume 1: Long Papers), pages 7361-7373, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[66] Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. Bigtrans: Augmenting large language models with multilingual translation capability over 100 languages. arXiv preprint arXiv:2305.18098, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[67] Aoxiong Yin, Tianyun Zhong, Li Tang, Weike Jin, Tao Jin, and Zhou Zhao. Gloss attention for gloss-free sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2551-2562, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[68] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5975-5989, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[69] Biao Zhang, Mathias Müller, and Rico Sennrich. SLTUNET: A simple unified model for sign language translation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[70] Jian Zhao, Weizhen Qi, Wengang Zhou, Nan Duan, Ming Zhou, and Houqiang Li. Conditional sentence generation and cross-modal reranking for sign language translation. IEEE Transactions on Multimedia, 24:2662-2672, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[71] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[72] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[73] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1316-1325. Computer Vision Foundation / IEEE, 2021. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[74] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[75] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.812 + ], + "angle": 0, + "content": "[76] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Jiajun Chen, Lei Li, and Shujian Huang. Multilingual machine translation with large language models: Empirical results and analysis. arXiv preprint arXiv:2304.04675, 2023.1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.812 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18372" + } + ] +] \ No newline at end of file diff --git a/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_origin.pdf b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec8ad40258614bf540abddb16f918fe7d24a059f --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/09537dfb-17f4-403d-a9c1-d3890b661d09_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:929ccf1dafe2d76ba24022f75f8586205dfacf11992c07d73c46cdf35d2ac996 +size 583466 diff --git a/2024/LLMs are Good Sign Language Translators/full.md b/2024/LLMs are Good Sign Language Translators/full.md new file mode 100644 index 0000000000000000000000000000000000000000..124532a641a346464c55e693c6f24cdb6c97435d --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/full.md @@ -0,0 +1,308 @@ +# LLMs are Good Sign Language Translators + +Jia Gong $^{1\dagger}$ Lin Geng Foo $^{1\dagger}$ Yixuan He $^{1\dagger}$ Hossein Rahmani $^{2}$ Jun Liu $^{1‡}$ $^{1}$ Singapore University of Technology and Design $^{2}$ Lancaster University + +{jia_gong,lingeng_foo,yixuan_he}@mypmail.sutd.edu.sg, h.rahmani@lancaster.ac.uk, jun.liu@sutd.edu.sg + +# Abstract + +Sign Language Translation (SLT) is a challenging task that aims to translate sign videos into spoken language. Inspired by the strong translation capabilities of large language models (LLMs) that are trained on extensive multilingual text corpora, we aim to harness off-the-shelf LLMs to handle SLT. In this paper, we regularize the sign videos to embody linguistic characteristics of spoken language, and propose a novel SignLLM framework to transform sign videos into a language-like representation for improved readability by off-the-shelf LLMs. SignLLM comprises two key modules: (1) The Vector-Quantized Visual Sign module converts sign videos into a sequence of discrete character-level sign tokens, and (2) the Codebook Reconstruction and Alignment module converts these character-level tokens into word-level sign representations using an optimal transport formulation. A sign-text alignment loss further bridges the gap between sign and text tokens, enhancing semantic compatibility. We achieve state-of-the-art gloss-free results on two widely-used SLT benchmarks. + +# 1. Introduction + +Sign languages, which are visual signals expressed through hand, body, and facial movements, serve as the primary means of communication within the hearing-impaired community. In an effort to facilitate effective communication with this community, much attention has been directed towards developing techniques to tackle the Sign Language Translation (SLT) task [10, 11, 69, 72, 73], where the goal is to translate sign videos into spoken language. SLT is a challenging task that requires cross-modality understanding of visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72]. Despite the notable advancements in terms of network architectures [10, 11, 37], visual sign representations (e.g., with keypoint estimators [14, 74]), + +and training methods [22, 69, 72], how to effectively tackle the challenging cross-modal SLT task with limited paired sign-text data largely remains an open question. + +On the other hand, large language models (LLMs) – referring to language models that have been trained on a large web-scale text corpus – have recently received a lot of attention. Since LLMs are trained over a very large corpus across multiple languages with distinct syntax and lexicon, they possess rich semantic understanding and powerful linguistic abilities [9, 15, 58]. At the same time, LLMs have also demonstrated an impressive capability to translate across multiple languages [9, 15], even showing a strong potential for translating languages with limited data [66, 76]. The foundation for this translation proficiency lies in the shared linguistic properties of syntax, lexicon, and morphology that many languages hold, which is particularly evident within language families [3]. Therefore, when faced with a new language with limited data, LLMs can draw upon the wealth of knowledge acquired from previously learned languages, leveraging any shared properties of syntax, lexicon and morphology with previous languages to effectively generate translations for new languages with remarkable accuracy and fluency [66, 76]. + +Inspired by the impressive translation capabilities of LLMs, we aim to harness off-the-shelf LLMs to handle the challenging SLT task. However, training LLMs directly on the relatively small SLT dataset can potentially lead to forgetting of their rich knowledge [12, 29] and a decline in performance, thus we follow previous LLM-based works [24, 43, 54] to keep the off-the-shelf LLM frozen, which preserves the rich knowledge acquired during its pretraining on a vast multilingual corpus. Consequently, our focus shifts towards making the sign videos compatible and readable for the off-the-shelf and frozen LLM to perform SLT. Specifically, in this paper, we explore the following question: Can we treat the sign video as a form of language, and leverage an off-the-shelf and frozen LLM to translate them? Notably, this is not a straightforward task because directly encoding features from sign videos with a pre-trained feature extractor [28, 64] will result in a large gap between + +the sign video features and text tokens, making it difficult for off-the-shelf LLMs to understand them. + +Based on the observation that LLMs can effectively handle new languages by leveraging shared commonalities with previously learned languages, we aim to introduce designs that transform our sign videos into a language-like format which are readable and friendly to LLMs. Specifically, we hypothesize that providing language-like representations of sign videos to the LLM improves the LLM's understanding of the sign videos and facilitates greater exploitation of shared properties with previously learned languages, thus resulting in better SLT performance by the LLM. To obtain language-like sign video representations, we draw inspiration from linguistic studies and analyses on LLMs [30, 63] and regularize the sign video to embody two fundamental language-like characteristics: Discrete Characteristics: Spoken languages are inherently discrete, since each language contains a finite set of words (and subwords) that convey distinct concepts, allowing them to be naturally represented through a discrete vocabulary with distinct tokens [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52]. This hierarchical structure enables languages to express a wide range of words with a limited set of characters, and convey diverse sentences with a limited number of words. + +In this paper, we present SignLLM, a novel framework designed to regularize input sign videos to produce sign token representations with language-like characteristics that are compatible and friendly to LLMs. Our proposed SignLLM includes two key designs to impart discrete characteristics and a hierarchical structure to the produced sign tokens. Firstly, we introduce the Vector-Quantized Visual Sign (VQ-Sign) module that facilitates the conversion of sign videos into a sequence of discrete character-level sign tokens. To achieve this, the VQ-Sign module consists of a discrete character-level sign codebook which is optimized through a self-supervised context prediction task. Next, we introduce the Codebook Reconstruction and Alignment (CRA) module that converts the character-level sign tokens into word-level sign tokens, facilitated by an optimal transport formulation. Moreover, we employ a sign-text alignment loss to further narrow the gap between the sign tokens and text tokens. These designs enable SignLLM to produce sign sentences that embody two key characteristics of spoken languages: discrete characteristics and a hierarchical structure, which enhances their compatibility with LLMs and makes them more readily interpretable by LLMs. + +After producing the language-like sign sentences, we feed them into an off-the-shelf and frozen LLM along with a text prompt that instructs the LLM to generate translations in the desired language. We empirically observe that, by employing our SignLLM's designs to align sign videos with + +languages, we can already leverage a frozen LLM to attain state-of-the-art SLT performance. These findings suggest that our proposed SignLLM framework is a promising first step towards effectively harnessing LLMs for SLT. We hope our initial explorations can inspire future work within the community to leverage LLMs for SLT. + +In summary, our main contributions are: (1) We propose a novel SignLLM framework that is the first to harness the power of off-the-shelf and frozen LLMs for SLT. (2) To make the input sign video compatible with LLMs, our SignLLM framework incorporates two designs: a VQ-Sign module to quantize the sign video into a sequence of discrete character-level sign tokens and a CRA module that transforms the character-level sign tokens to word-level sign tokens. (3) Through our proposed designs, we achieve state-of-the-art gloss-free results on two popular SLT datasets. + +# 2. Related Work + +Sign Language Translation (SLT) aims to transform sign videos into natural language sentences. It is a challenging task that requires understanding of both visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72] which limits the performance of SLT methods. To improve SLT performance, many previous works [10, 11, 14, 22, 34, 35, 37, 56, 68, 69, 72, 74] aim to enhance the visual sign representations and text decoding capabilities of SLT methods. Some works propose deep architectures based on RNNs [10, 35], GCNs [34], and Transformers [11, 37, 62, 68]. Other approaches include introducing a keypoint estimator to enhance the visual sign representation [14, 35, 56, 74], introducing pre-training tasks [22, 72], or jointly modelling several SLT-related tasks [69]. Some works also introduce larger datasets (e.g., How2Sign [20] and BOBSL [1]), which present a huge challenge with their large sign and text vocabularies. Besides, some recent works [67, 72] focus on the gloss-free setting – these works do not use sign gloss annotations for training, which reduces the cost of training SLT models, and our work also falls into this category. In contrast to existing works, we aim to harness the capabilities of off-the-shelf and frozen LLMs to perform SLT, by regularizing the sign videos into a language-like representation and prompting the LLM to generate text of the desired language. + +Large Language Models (LLMs) refer to language models that have been trained extensively on a very large web-scale text corpus. LLMs have shown impressive text generation capabilities, attracting a lot of attention recently [21, 49, 71]. In particular, since LLMs have been trained on large amounts of text data, they demonstrate strong generalization capabilities across various text-based tasks, including code generation [51], open-domain question answering [75], and multilingual translation [9, 15]. Inspired by the re + +![](images/5fdbc12d114ffb55abef70f386f672c6c2cdc35f62277c31ce117b7d00a6220c.jpg) +Figure 1. An overview of our SignLLM framework. During inference (top): Given an input sign video $X$ , we first pass it through our VQ-Sign module to obtain a sequence of discrete character-level sign tokens $\hat{Z}$ . Our VQ-Sign consists of a visual encoder $E_{v}$ to extract compact features and a character-level sign codebook $\mathbb{S}^c$ for quantization to obtain $\hat{Z}$ . Next, we feed $\hat{Z}$ into our CRA module, which reorganizes $\hat{Z}$ by replacing short sequences of character tokens with word-level tokens via the word-level codebook, e.g., character sequence $[s_2, s_3, s_4]$ to word $s_2 s_3 s_4$ . This transforms the sign video data to a language-like sign sentence $W$ , which is fed into the LLM along with a text prompt which guides the LLM to generate translations in the desired language. During training (bottom): We optimize VQ-Sign and its discrete sign codebook via a context prediction task, which seeks to recognize the future time steps based on the current context information. Next, for our CRA module, we construct the optimal word-level codebook by considering two aspects: entropy and size, which we address using optimal transport techniques. Then, we narrow the gap between the sign token space and LLM's text token space via minimizing the MMD loss, which improves the semantic compatibility between them. + +cent advancements in LLMs, we explore harnessing LLMs for translation of sign videos, by converting the sign videos to a sequence of language-like sign tokens via our Sign-LLM framework, and treating the sign tokens as a form of language that can be translated by an LLM. To the best of our knowledge, we are the first work to leverage an off-the-shelf and frozen LLM to tackle SLT. + +# 3. Method + +In this section, we first introduce the overview of our SignLLM in Sec. 3.1. Then, we describe two main components of our SignLLM framework: the VQ-Sign and CRA modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we list the training and inference details in Sec. 3.4. + +# 3.1. SignLLM Overview + +In order to effectively handle SLT, in this paper we draw inspiration from LLMs' remarkable capabilities in generating translations across multiple languages [9, 15]. In particular, LLMs have been extensively trained on a large web-scale multilingual text corpus and have learned diverse knowledge regarding the properties of many languages, thus they are able to draw on shared commonalities with previously learned languages to effectively handle new languages with limited data [66, 76]. + +Therefore, to leverage the strong translation capabilities + +of LLMs to handle SLT, we introduce a novel SignLLM framework. SignLLM converts the input sign video $X$ into a language-like sign sentence $W$ that aligns with the linguistic characteristics of spoken languages and is friendly and compatible with LLMs. Then, to perform SLT, the language-like sign sentence $W$ can be fed into an off-the-shelf and frozen LLM along with a text prompt that guides the LLM to generate translations in the desired language. + +Specifically, to produce sign sentences $W$ that are friendly and understandable by LLMs, we aim to regularize our sign sentences $W$ to embody two core linguistic characteristics: Discrete Characteristics: Spoken languages are naturally discrete and consist of distinct words or sub-words with corresponding discrete tokens in a vocabulary [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52], where words are composed from characters and sentences are composed from various words. + +To achieve the above, our SignLLM framework comprises of three parts, as shown in Fig. 1: (1) The VQ-Sign module converts the input sign video $X$ into a sequence of discrete sign tokens $\hat{Z}$ , aligning the sign representations with text's discrete characteristics. These sign tokens $\hat{Z}$ are character-level sign tokens that are retrieved from a learned discrete character-level codebook. (2) The CRA module maps meaningful compositions of character-level sign tokens $\hat{Z}$ into word-level sign tokens that form a sign sentence + +$W$ , further imparting a language-like hierarchical structure to the video sign representations. Moreover, we also align the sign token codebooks towards the text token space to improve semantic compatibility. (3) An off-the-shelf LLM takes the sign sentence $W$ as input, along with an instructive text prompt that guides the LLM to generate the translation in the desired language. More details regarding the text prompt are in the Supplementary. Next, we present our VQ-Sign and CRA modules in detail. + +# 3.2. Vector-Quantized Visual Sign Module + +First, in order to produce a language-like representation, we would like to impart discrete characteristics to the input sign videos, aligning them more closely with spoken language representations that are inherently discrete and consist of distinct tokens in a vocabulary. However, achieving this is not straightforward because sign videos are a continuous signal in a high-dimensional spatio-temporal space which cannot be easily represented by a set of discrete tokens, and for which the vocabulary is not readily available. Hence, we introduce our Vector-Quantized Visual Sign (VQ-Sign) module to quantize the sign video $X$ into a sequence of discrete sign tokens $\hat{Z}$ via a sign codebook $\mathbb{S}^c$ . As illustrated in Fig. 1, our VQ-Sign module involves a series of steps, which we present in detail next. + +In the first step, we extract a compact feature $Z$ from the high-dimensional input sign video $X \in \mathbb{R}^{N \times H \times W}$ , where $N$ is the number of video frames, while $H$ and $W$ are the height and width of the video frames respectively. To be precise, the sign video $X$ is first organized into a sequence of short overlapping video clips, then each short video clip is fed into a visual encoder $E_{v}$ to extract a compact feature representation of dimensionality $d$ . Overall, this step transforms the original high-dimensional input sign video $X \in \mathbb{R}^{N \times H \times W}$ into a compact feature $Z \in \mathbb{R}^{\frac{N}{n} \times d}$ , where $n$ represents the number of frames between the start of neighboring clips. Notably, since $Z$ is obtained by processing $\frac{N}{n}$ short clips, $Z$ can also be seen as a sequence of $\frac{N}{n}$ clip-wise features, i.e., $\{z_{t}\}_{t=1}^{\frac{N}{n}}$ , where each $z_{t} \in \mathbb{R}^{d}$ corresponds to the feature of the $t$ -th short clip. + +In the next step, we transform the feature $Z = \{z_{t}\}_{t=1}^{\frac{N}{n}}$ into a sequence of discrete tokens $\hat{Z}$ using a codebook $\mathbb{S}^c$ . Specifically, we discretize each clip's feature $z_{t}$ into a discrete token $\hat{z}_{t}$ by finding the matching token $s_{j}$ from the codebook $\mathbb{S}^c = \{s_i\}_{i=1}^M$ , where the $i$ -th token in the codebook is denoted as $s_{i} \in \mathbb{R}^{d}$ and $M$ is the number of tokens in the codebook. The matching token $s_{j}$ is the codebook's closest element to the feature $z_{t}$ in terms of Euclidean distance, i.e., $j = \arg \min_{i}(\| z_{t} - s_{i}\|_{2}^{2})$ . After the matching, each feature $z_{t}$ is replaced by $\hat{z}_{t} = s_{j}$ , which results in a discrete token sequence as shown in Fig. 1, e.g., $[s_2, s_3, s_4, s_M, s_5, s_1]$ . Note that, we randomly initialize all + +the tokens $\{s_i\}_{i=1}^M$ in the sign codebook $\mathbb{S}^c$ at the start and optimize them during training, as introduced next. + +However, we face a challenge in learning the discrete codebook $\mathbb{S}^c$ . In particular, although autoencoding [59] has been a popular method for producing a codebook of discrete units, the high complexity of sign videos makes autoencoding (i.e., self-reconstruction of sign videos) challenging and costly. Therefore, inspired from predictive coding [4], a widely-used method in text and speech representation learning [6, 44, 46], we propose to learn discrete representations of sign videos through a context prediction task. Context prediction [46] is a self-supervised task that focuses on recognizing the future content in latent space based on the current information, which can learn discrete representations while eliminating the need for reconstructing the high-dimensional input video data. Furthermore, previous works show that training with context prediction effectively captures the temporal dependencies and relationships between elements in a sequence [5, 27], and the learned representations are often transferable to downstream tasks [6, 46]. + +Specifically, we employ a context prediction task where we try to distinguish a future sample $z_{\tau + k}$ based on the current context representation $c_{\tau}$ at various time steps $\tau$ . To facilitate this task, after we obtain the discrete token sequence $\hat{Z}$ , we further produce a context latent representation $c_{\tau}$ using an auto-regressive model $g$ that summarizes all the discrete tokens before a certain time step $\tau$ (i.e., $\{\hat{z}_t\}_{t \leq \tau}$ ) to produce context latent representations $c_{\tau} = g(\{\hat{z}_t\}_{t \leq \tau})$ . Then, we optimize our module by minimizing the following context prediction contrastive loss $\mathcal{L}_k^{\mathrm{cp}}$ : + +$$ +\mathcal {L} _ {k} ^ {\mathrm {c p}} = - \sum_ {\tau = 1} ^ {\frac {N}{n} - k} \left(\log \sigma \left(z _ {\tau + k} ^ {\top} h _ {\tau}\right) + \lambda \underset {\widetilde {z} \sim p _ {n}} {\mathbb {E}} \left[ \log \sigma \left(- \widetilde {z} ^ {\top} h _ {\tau}\right) \right]\right) \tag {1} +$$ + +where $h_\tau$ is obtained by applying a trainable linear layer to $c_{\tau}$ , $\sigma(z_{\tau+k}^{\top} h_{\tau})$ is the probability of $z_{\tau+k}$ being the true sample among the negatives, $\widetilde{z}$ are negative samples drawn from a minibatch $p_n$ , and $\lambda$ is a hyperparameter. We sum $\mathcal{L}_k^{\mathrm{cp}}$ over different step sizes $k$ to obtain the context prediction loss $\mathcal{L}^{\mathrm{cp}} = \sum_{k=1}^{K} \mathcal{L}_k^{\mathrm{cp}}$ , where $K$ is the maximum number of future clips that we are interested in. + +Following [18], in order to optimize the matching between $\hat{z}_t$ and $z_{t}$ , we further add two losses to optimize the matching distance between $\hat{z}_t$ and $z_{t}$ , such that the overall loss $\mathcal{L}^{VQ}$ to optimize our VQ-Sign module is as follows: + +$$ +\mathcal {L} ^ {V Q} = \mathcal {L} ^ {\mathrm {c p}} + \sum_ {t = 1} ^ {N / n} \| \operatorname {s g} \left(z _ {t}\right) - \hat {z} _ {t} \| ^ {2} + \gamma \sum_ {t = 1} ^ {N / n} \| z _ {t} - \operatorname {s g} (\hat {z} _ {t}) \| ^ {2}, \tag {2} +$$ + +where $\operatorname{sg}(z) \equiv z$ is the stop-gradient operator and $\gamma$ is a hyperparameter. By optimizing $\mathcal{L}^{VQ}$ , we can train our VQ-Sign and the discrete codebook without the need for reconstructing high-dimensional video clips, which makes the codebook construction a viable and relatively cheap option. + +In summary, our VQ-Sign transforms sign videos into sequences of discrete sign tokens $\hat{Z}$ , which are friendlier and more understandable to LLMs. Notably, the produced discrete tokens $\hat{Z}$ can be likened to character-level tokens, in the sense that each discrete token $\hat{z}_t$ corresponds to a short clip and may not contain much semantic meaning on its own (similar to linguistic characters), but they can be combined into a sequence to convey a clear semantic meaning (akin to forming a word or sentence). Thus, inspired by this, we call VQ-Sign's codebook $\mathbb{S}^c$ the character-level codebook that contains character-level sign tokens. + +# 3.3. Codebook Reconstruction and Alignment + +In the previous section, we quantize sign videos into discrete character-level sign tokens, which aligns them closer to language representations. In this section, our goal is to impart a hierarchical structure to our sign video representations, which makes them align even closer to language representations. Specifically, we aim to compose our character-level sign tokens into word-level sign tokens to mirror the observed hierarchical structure in spoken language, which makes them even more compatible with LLMs. + +Intuitively, considering a spoken language sentence, we can represent it as a sequence of words, with each word formed by one or multiple characters. For example, a sentence 'I love AI' can be decomposed into a word sequence [I', 'love', 'AI'], where the word 'love' is in turn formed by the character sequence [1', o', v', e']. We observe that, although each individual character may not contain much semantic meaning on its own, they can be composed to form words with clearer semantic meaning. In a similar fashion, we also want to impart such hierarchical structure to our character-level sign tokens by composing them to form meaningful word-level sign tokens. + +Hence, we aim to find an optimal transformation from character-level sign tokens to word-level sign tokens for enhanced readability and compatibility with LLMs. To this end, we introduce the Codebook Reconstruction and Alignment (CRA) module to transform the character-level codebook $\mathbb{S}^c$ from VQ-Sign into a word-level codebook $\mathbb{S}^w$ whose tokens convey richer and clearer semantic meaning. Inspired by optimal transport methods [16, 61, 65], we observe that the above transformation can be formulated as an optimal transport problem of transporting characters into words, thus we introduce a codebook reconstruction algorithm with an optimal transport formulation to find an optimal transformation. Additionally, to further reduce the distribution gap between the sign tokens and the text tokens, our CRA module also performs sign-text alignment, enhancing the semantic compatibility of the sign tokens with LLMs. We introduce the details below. + +To begin, the objective of our Codebook Reconstruction Algorithm is to create a word-level codebook $\mathbb{S}^w$ + +based on VQ-Sign's character-level codebook $\mathbb{S}^c$ . The challenge lies in determining which character-level sign tokens should be assembled together to form word-level sign tokens, which is a complex problem. To address this complexity, we adopt an approach based on two fundamental principles. Firstly, in order to maximize the overall predictability of the word-level tokens and enhance the distinctiveness of each token, we seek to minimize the entropy of each word-level token within the vocabulary [42]. We remark that, several approaches for establishing language-based subword vocabularies [8, 55] can be seen as entropy-minimizing approaches, employing different heuristics to establish the vocabulary with the goal of minimizing entropy [23]. On the other hand, considering the limited availability of sign video data, we incorporate codebook size as another key factor in our word-level codebook construction, since studies on languages with limited data [26, 53] have also identified vocabulary size as a crucial aspect. Specifically, too small a vocabulary can result in sub-optimal entropy values, while an excessively large vocabulary size can lead to issues such as parameter explosion and token sparsity which hinder understanding [2], and finding the right balance between these effects becomes even more sensitive for languages with limited data [19, 53]. + +Based on these principles, our objective is to determine an optimal codebook size that maximizes the entropy reduction while taking into account the increase in codebook size. In other words, we would like to find an optimal codebook size that maximizes the gradient of the entropy reduction with respect to the codebook size increase. To simplify the optimal size searching problem, we define a fixed size increment $m$ and search through codebooks of various sizes (where the difference between each codebook size is $m$ tokens). Specifically, we define the $r$ -th codebook $(\mathbb{S}_r^w)$ as the codebook with $r \times m$ tokens. Then, we seek to identify the optimal set of word-level tokens, where each word-level token is composed of character-level tokens. We approach this by formulating the character compositions as an optimal transport problem, where characters are transported to words. + +However, it can be challenging to identify specific character combinations that convey precise semantic information, due to the temporal complexity of sign videos, which often makes the character-level token sequences $\hat{Z}$ quite messy. For instance, some signers may execute the signing motions at a slower speed, which can lead to consecutive short video clips being highly similar, resulting in consecutively-repeating character-level discrete tokens. Thus, the character-level sequences between different signers can differ significantly (e.g., $[s_1,s_2]$ vs $[s_1,s_1,s_1,s_2,s_2]$ ) due to such duplication of character-level tokens, even though they may contain the same semantic information. At the same time, simply filtering out re + +peated character-level tokens straightforwardly (e.g., setting all $[s_1, s_1, s_1, s_2, s_2]$ to $[s_1, s_2]$ ) is sub-optimal, since the speed of some signs can also convey some information [31, 60], e.g., if a signer signs "ugly" quickly, it conveys "very ugly" in American Sign Language. + +Pre-processing of Repeated Characters. Therefore, to alleviate the impact of the signer's speed while keeping the information regarding each sign's speed, we first preprocess the character-level sequences as follows: First, we find all the repeated tokens in the character-level sequence and compute the average number of repeated tokens $(\alpha)$ in each sequence. Then, for each repeated sequence (e.g., $[s_1, s_1, s_1]$ ), we keep the first character and remove the tailing repetitive ones (e.g., $[s_1, s_1, s_1]$ to $[s_1]$ ). At the same time, if the character-level tokens repeat more than $\alpha$ times, we insert a single character-level token $s_0$ as a "slowing down" sign, e.g., $[s_1, s_1, s_1]$ to $[s_1, s_0]$ if $\alpha < 3$ . Crucially, this allows us to reduce redundancy, while still representing "quick" or "slow" signs that account for differences in the signer's speed. Overall, this pre-processing and reducing of repeating characters makes the character-level sequence less messy, facilitating the search for specific meaningful character combinations. + +Optimal Transport Formulation. Then, using the preprocessed character-level sequences, we search for an optimal word-level codebook via an optimal transport formulation [16, 61, 65]. Following our discussion above, we aim to find an optimal word-level codebook (i.e., $\mathbb{S}_r^w$ ) with low entropy and compact size. Specifically, to measure codebook entropy, we follow [25, 42, 45] to define the entropy of the $r$ -th codebook $\mathbb{S}_r^w$ as: + +$$ +\mathcal {H} _ {\mathbb {S} _ {r} ^ {w}} = - \sum_ {w _ {j} \in \mathbb {S} _ {r} ^ {w}} P (w _ {j}) \log P (w _ {j}), \tag {3} +$$ + +where $P(w_{j})$ is the relative frequency of the $j^{th}$ token $w_{j}$ from the word-level codebook $\mathbb{S}_r^w$ . Then, based on VQ-Sign's character-level codebook $\mathbb{S}^c$ , the entropy of the word-level codebook $\mathbb{S}_r^w$ can be computed through the following (with proof in Supplementary): + +$$ +\begin{array}{l} \mathcal{H}_{\mathbb{S}_{r}^{w}} = -\sum_{w_{j}\in \mathbb{S}^{w}}\sum_{s_{i}\in \mathbb{S}^{c}}P(w_{j},s_{i})\log P(w_{j},s_{i}) \\ - \sum_ {w _ {j} \in \mathbb {S} _ {r} ^ {w}} \sum_ {s _ {i} \in \mathbb {S} ^ {c}} P \left(w _ {j}, s _ {i}\right) (- \log P \left(s _ {i} \mid w _ {j}\right)), \tag {4} \\ \end{array} +$$ + +where $P(s_{i}|w_{j})$ is the probability of the character-level token $s_i$ appearing in the word-level token $w_{j}$ . + +Then, to formally define our objective function, we follow existing works [16, 48, 65] to define a transport matrix $\pmb{P}$ that represents the assignments of characters to words, and a distance matrix $\pmb{D}$ that represents the cost of transportation. We define the transport matrix $\pmb{P} \in \mathbb{R}^{m \times (r \cdot m)}$ with the $(j,i)$ -th element as $P(w_{j}, s_{i})$ , and define the distance matrix $\pmb{D} \in \mathbb{R}^{m \times (r \cdot m)}$ as a matrix whose $(j,i)$ -th element is $\log P(s_{i} | w_{j})$ . Note that, if $w_{j}$ contains $s_{i}$ , we use + +$\frac{1}{\mathrm{length}(w_j)}$ to estimate $P(s_i|w_j)$ , and if $w_j$ does not contain $s_i$ , then it is deemed an infeasible assignment, and we set the distance $\log P(s_i|w_j) = \infty$ . We further define $H(\pmb{P})$ as $-\sum_{w_j \in \mathbb{S}_r^w} \sum_{s_i \in \mathbb{S}^c} P(w_j, s_i) \log P(w_j, s_i)$ , which is simply the entropy of the probability distribution $P(w_j, s_i)$ . + +Hence, based on Eq. 4, the objective function for minimizing the entropy of $\mathbb{S}_r^w$ can be formulated as: + +$$ +\underset {\boldsymbol {P} \in \mathbb {R} ^ {m} \times (r \cdot m)} {\arg \min } H (\boldsymbol {P}) + \sum_ {j} \sum_ {i} \boldsymbol {P} (j, i) \boldsymbol {D} (j, i). \tag {5} +$$ + +Specifically, following previous works [16, 48, 65], we impose two constraints on the transport matrix $P$ : the sum of each row in $P$ should equal to the probability of character token $s_i$ and the sum of each column in $P$ should equal to the probability of word token $w_j$ . Formally, we constrain the transport matrix $P$ with: $|\sum_{i}P(i,j) - P(w_{j})|\leq \epsilon$ and $|\sum_{j}P(i,j) - P(s_{i})|\leq \epsilon$ , where $\epsilon$ is a small positive fixed hyperparameter. + +Intuitively, this optimization process can be regarded as an optimal transport problem to find the best way to transport mass from $\mathbb{S}^c$ to $\mathbb{S}_r^w$ . To handle this optimal transport problem, we leverage the Sinkhorn algorithm [16, 48, 65], allowing us to effectively construct the candidate word-level codebooks $\mathbb{S}_r^w$ with minimal entropy. Since our increment $m$ between each candidate codebook is fixed, we can find the optimal codebook size that maximizes the gradient of entropy reduction by simply computing and finding the maximum the entropy difference between $\mathbb{S}_r^w$ and $\mathbb{S}_{r - 1}^w$ . + +After finding the optimal word-level codebook, we construct the word-level sign tokens by composing all features of the character-level tokens into the word-level tokens via our autoregressive model $g$ . Refer to Supplementary for more details. Overall, with our codebook reconstruction algorithm, we can construct an optimal word-level sign codebook with low entropy yet also with relatively small size. + +Sign-Text Alignment. Next, we further align the sign tokens with the text tokens used in LLMs in order to further improve semantic compatibility between them. To achieve this, we measure the distribution gap between sign tokens and text tokens via Maximum Mean Discrepancy (MMD) [57] and then optimize the sign tokens' embeddings by minimizing MMD, which narrows down the distribution gap. Specifically, we compute the gap between the sign embedding space $\mathbb{F}^s$ and text embedding space $\mathbb{F}^t$ via MMD as: + +$$ +\begin{array}{l} \mathcal {L} ^ {M M D} (\mathbb {F} ^ {s}, \mathbb {F} ^ {t}) = \sum_ {i = 1} ^ {n _ {s}} \sum_ {j = 1} ^ {n _ {s}} \frac {k (f (p _ {i}) , f (p _ {j}))}{n _ {s} ^ {2}} \\ + \sum_ {i = 1} ^ {n _ {t}} \sum_ {j = 1} ^ {n _ {t}} \frac {k \left(q _ {i} , q _ {j}\right)}{n _ {t} ^ {2}} - \sum_ {i = 1} ^ {n _ {s}} \sum_ {j = 1} ^ {n _ {t}} \frac {2 \cdot k \left(f \left(p _ {i}\right) , q _ {j}\right)}{n _ {s} n _ {t}}, \tag {6} \\ \end{array} +$$ + +where $p$ and $q$ are the tokens in $\mathbb{F}^s$ and $\mathbb{F}^t$ , $f$ is a small projection module that projects $\mathbb{F}^s$ to $\mathbb{F}^t$ , $n_s$ and $n_t$ are the numbers of tokens in $\mathbb{F}^s$ and $\mathbb{F}^t$ , and $k(\cdot)$ represents the radial + +kernel [57] that measures the distance between two samples. We apply MMD loss to both the word-level and the character-level tokens to narrow the overall sign-text gap. + +# 3.4. Training and Inference + +Inference. Given a sign video $X$ , we first extract compact features $Z$ via the visual encoder $E_{v}$ , and quantize $Z$ to $\hat{Z}$ via VQ-Sign's learned character-level codebook. Then, we transform the sequence of discrete character-level tokens $\hat{Z}$ into word-level tokens via our CRA, which produces a sign sentence $W$ . Lastly, we project the sign sentence $W$ to LLM embedding space via the small projection module $f$ , and then feed the sign sentence $W$ into the LLM along with a text prompt to instruct the LLM to perform the SLT task and generate text in the desired language. + +Training. Our SignLLM is optimized in two stages: pretraining and fine-tuning. Specifically, the pre-training stage, which does not require explicit SLT supervision, includes two sub-stages: (i) We first pre-train VQ-Sign via the context prediction task with $\mathcal{L}^{VQ}$ in Eq. 2. (ii) Then, based on the VQ-Sign's learned character-level codebook, we construct the word-level codebook with the codebook reconstruction algorithm (Sec. 3.3) and apply the MMD loss $\mathcal{L}^{MMD}$ in Eq. 6 to align the sign codebooks $(\mathbb{S}^c$ and $\mathbb{S}^w)$ and the text vocabulary of the desired language. + +After the pre-training, we fine-tune SignLLM. To aid LLMs in understanding the sign sentences as texts, we additionally maximize the similarity between the text tokens $Y$ generated by LLM and the ground truth tokens $\bar{Y}$ as: $\mathcal{L}^{sim}$ by minimizing the cross-entropy between them. We fine-tune our SignLLM (with frozen LLM) via the loss $\mathcal{L}^{ft}$ as follows: $\mathcal{L}^{ft} = \mathcal{L}^{VQ} + \lambda_1\mathcal{L}^{MMD} + \lambda_2\mathcal{L}^{sim}$ , where $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters. Note that we follow previous gloss-free works [67, 72] to train our SignLLM framework, eliminating the need for additional gloss data. + +# 4. Experiments + +# 4.1. Implementation Details + +Our visual encoder $E_{v}$ is constructed by appending two Conv3D layers with a kernel size of (5, 3, 3) and a stride of (2, 1, 1) to a ResNet18 [28] pre-trained on ImageNet [17]. Each clip consists of 13 frames and the gap between the neighboring clips ( $n$ ) is 4. Besides, the auto-regressive model $g$ is implemented as a Convolutional Gated Recurrent Layer with a kernel size of (1, 1). We set the total number of discrete vectors $M$ at 256 and each vector's dimension $d$ at 1024 for our character-level codebook. In VQ-Sign's pre-training phase, we set $\gamma = 0.25$ , initialize the learning rate at 0.01 and use the Adam algorithm, training the model to predict the future three clips ( $K = 3$ ) for 200 epochs. During codebook reconstruction, we set the increment $m$ to 32. We employ the frozen LLaMA-7B-16bit [58] + +as our LLM and project the codebook space to LLaMA's embedding space via $f$ , which consists of two fc-layers with ReLU. We set $\lambda_1 = 0.5$ , $\lambda_2 = 1$ , and initialize the learning rate at 0.001 to fine-tune our SignLLM over 20 epochs. Please see Supplementary for more implementation details. + +# 4.2. Datasets and Evaluation Metrics + +Datasets. We follow previous works [13, 14, 34, 72-74] to run experiments on the Phoenix-2014T [10] and CSL-Daily [73] datasets for SLT, and evaluate on their dev and test sets. Phoenix-2014T [10] is a German sign language dataset with a vocabulary size of 2887 German words. The training, dev, and test sets contain 7096, 519, and 642 samples. CSL-Daily [73] is a Chinese sign language dataset with a vocabulary size of 2343 Chinese words. The training, dev, and test sets contain 18401, 1077, and 1176 samples. + +Evaluation Metrics. Following previous works [13, 14, 37, 67, 72-74], we adopt BLEU [47] and ROUGE-L [38] as the evaluation metrics for SLT. BLEU-n evaluates the average translation precision up to n-grams, and we follow previous works [13, 14, 72-74] to report results for BLEU-1 to BLEU-4 (i.e., B1, B2, B3 and B4). ROUGE-L (or ROUGE) computes the F1 score based on the longest common subsequence between the predicted and ground truth texts. + +# 4.3. Main Results + +Results on PHOENIX2014T dataset. Tab. 1 presents a comparison of our approach with state-of-the-art gloss-based and gloss-free methods for SLT. Our method consistently improves upon all reported metrics as compared to other gloss-free approaches. + +Results on CSL-Daily dataset. We compare our method with state-of-the-art approaches on the CSLD Daily dataset in Tab. 2. We outperform previous gloss-free works on all metrics, showing the efficacy of our approach. + +# 4.4. Ablation Study + +To further investigate the proposed method, we follow previous works [13, 33, 67, 69, 72] to conduct extensive ablation experiments on the Phoenix-2014T dev and test sets. Refer to Supplementary for more experiment results. + +Impact of LLM. First, we evaluate the impact of leveraging LLMs for the SLT task. Specifically, we establish three baselines: 1) Ours (w/o LLM) where we replace the LLM with a trainable lightweight text generator (mBART [40]). 2) Ours (w/ T5) where we replace our LLM with a smaller LLM (T5 [50]). As shown in Tab. 3, our approach (w/ LLaMA [58]) achieves a much better performance than Ours (w/o LLM), showing the efficacy of leveraging a powerful LLM. We also find that using a smaller and less powerful LLM (T5) leads to worse performance, as expected. + +Impact of SignLLM. Next, we explore the impact of Sign-LLM by comparing against the following baselines: 1) En + +Table 1. Results on Phoenix-2014T dataset [10]. + +
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]47.2634.4027.0522.38-46.6133.7326.1921.32-
ConSLT [22]---21.1147.74---21.5947.69
STN-SLT [62]49.1236.2928.3423.23-48.6135.9728.3723.65-
STMC-T [74]47.6036.4329.1824.0948.2446.9836.0928.7023.6546.65
BN-TIN-Transf.+SignBT [73]51.1137.9029.8024.4550.2950.8037.7529.7224.3249.54
PET [33]-----49.5437.1929.3024.0249.97
MMTLB [13]53.9541.1233.1427.6153.1053.9741.7533.8428.3952.65
TS-SLT [14]54.3241.9934.1528.6654.0854.9042.4334.4628.9553.48
SLTUNET [69]---27.8752.2352.9241.7633.9928.4752.11
Gloss-freeNSLT [10]28.1016.8111.829.1231.0027.1015.6110.828.3529.70
NSLT+Bahdanau [7, 10]31.8719.1113.169.9431.8032.2419.0312.839.5831.80
NSLT+Luong [10, 41]31.5818.9813.2210.0032.6029.8617.5211.969.0030.70
TSPNet [37]-----36.1023.1216.8813.4134.96
CSGCR [70]35.8524.7718.6515.0838.9636.7125.4018.8615.1838.85
GASLT [67]-----39.0726.7421.8615.7439.86
GFSLT-VLP [72]44.0833.5626.7422.1243.7243.7133.1826.1121.4442.49
Ours46.8836.5929.9125.2547.2345.2134.7828.0523.4044.49
+ +Table 2. Results on CSL-Daily dataset [73]. * means that the result was reproduced by [72] + +
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]37.4724.6716.8611.8837.9637.3824.3616.5511.7936.74
ConSLT [22]---14.8041.46---14.5340.98
BN-TIN-Transf.+SignBT [73]51.4637.2327.5120.8049.4951.4237.2627.7621.3449.31
MMTLB [13]53.8140.8431.2924.4253.3853.3140.4130.8723.9253.25
TS-SLT [14]55.2142.3132.7125.7655.1055.4442.5932.8725.7955.72
SLTUNET [69]---23.9953.5854.9841.4431.8425.0154.08
Gloss-freeSLRT* [11]21.039.975.964.0420.5120.009.114.933.0319.67
NSLT+Luong [10, 41]34.2219.7212.247.9634.2834.1619.5711.847.5634.54
GASLT [67]-----19.909.945.984.0720.35
GFSLT-VLP [72]39.2025.0216.3511.0736.7039.3724.9316.2611.0036.44
Ours42.4526.8817.9012.2339.1839.5528.1320.0715.7539.91
+ +Table 3. Ablation study for impact of LLM + +
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o LLM)29.7520.0414.9611.9527.2018.2913.3210.36
Ours (w/ T5)44.2034.5527.1522.9044.0334.1227.2322.51
Ours46.8836.5929.9125.2545.2134.7828.0523.40
+ +Table 4. Ablation study for impact of SignLLM. + +
MethodDevTest
B1B2B3B4B1B2B3B4
Encoder Only26.9517.0412.119.4125.6316.1011.208.42
Encoder + FT39.2929.2922.5518.1839.9229.1422.5418.17
Ours46.8836.5929.9125.2545.2134.7828.0523.40
+ +Table 5. Ablation study for main components of SignLLM. + +
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o VQ-Sign)35.7824.6518.0814.1533.1422.8016.7413.13
Ours (w/o Codebook Reconstruction)40.4530.4024.0719.7940.2530.0523.6319.47
Ours (w/o Sign-text Alignment)29.0519.3313.7210.4028.6719.2213.7310.63
Ours46.8836.5929.9125.2545.2134.7828.0523.40
+ +coder Only where we directly feed the outputs of the visual encoder $E_{v}$ into the LLM, and keep the LLM frozen. 2) Encoder + FT where we directly feed the outputs of the visual encoder $E_{v}$ into the LLM, and follow LLaVA [39] to fine-tune them to translate sign videos via LoRA [32]. We report the results in Tab. 4, where we significantly outperform the baselines. This shows that SignLLM is effective in harnessing off-the-shelf LLMs for the SLT task. + +Impact of Main Components of SignLLM. We also verify the impact of the key components of SignLLM by comparing against the following baselines: 1) Ours (w/o VQ-Sign) where we directly quantize the visual encoder $E_v$ 's output feature using $k$ -means algorithm. The visual encoder $E_v$ is pre-trained via the similarity loss $\mathcal{L}^{sim}$ instead of our proposed VQ-Sign's loss $\mathcal{L}^{VQ}$ . 2) Ours (w/o Codebook Reconstruction) where we feed the character-level sign tokens $\hat{Z}$ from VQ-Sign directly into the LLM, while applying the sign-text alignment loss $\mathcal{L}^{MMD}$ to the character-level sign tokens. 3) Ours (w/o Sign-text Alignment) where we do not apply the sign-text alignment loss $\mathcal{L}^{MMD}$ . As shown in Tab. 5, removing any of our main designs leads to a signifi + +![](images/b35ec2637db3253121a7691d3dbf8cb9234e7d6e9815834dd4123450f415b776.jpg) +Figure 2. Visualization of translation results. Correct translations are in blue while the wrong translations are in red. + +cant performance drop, showing their efficacy. + +Qualitative Results. We present two sample translations generated by our SignLLM and the current state-of-the-art (GFSLT-VLP [72]) in Fig. 2 for qualitative analysis. In the first sample (top), our model produces a highly accurate translation, whereas [72] inaccurately represents the semantic information. In the second sample (bottom), our model successfully preserves sentence semantics, while [72] introduces a translation error, resulting in redundant and erroneous information. These examples qualitatively show our SignLLM's efficacy in producing accurate translations. + +# 5. Conclusion + +We present SignLLM, a novel framework to harness off-the-shelf and frozen LLMs for SLT. SignLLM imparts language-like characteristics to sign video representations through the VQ-Sign and CRA modules, and a sign-text alignment loss improves semantic compatibility. We empirically observe that applying our SignLLM leads to state-of-the-art gloss-free results on two popular SLT benchmarks. + +Acknowledgements. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2023-121). + +# References + +[1] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bécie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford British sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2 +[2] Ben Allison, David Guthrie, and Louise Guthrie. Another look at the data sparsity problem. In Text, Speech and Dialogue: 9th International Conference, TSD 2006, Brno, Czech Republic, September 11-15, 2006. Proceedings 9, pages 327-334. Springer, 2006. 5 +[3] Ulrich Ammon. World languages: Trends and futures. The handbook of language and globalization, pages 101-122, 2010. 1 +[4] Bishnu S Atal and Manfred R Schroeder. Adaptive predictive coding of speech signals. Bell System Technical Journal, 49 (8):1973-1986, 1970. 4 +[5] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 4 +[6] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 4 +[7] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 8 +[8] Kaj Bostrom and Greg Durrett. Byte pair encoding is suboptimal for language model pretraining. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4617-4624, 2020. 2, 3, 5 +[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3 +[10] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 2, 7, 8 +[11] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 2, 8 +[12] Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7870-7881, 2020. 1 + +[13] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5110-5120. IEEE, 2022. 1, 2, 7, 8 +[14] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 1, 2, 7, 8 +[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2, 3 +[16] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5, 6 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7 +[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 4 +[19] Shuoyang Ding, Adithya Renduchintala, and Kevin Duh. A call for prudent choice of subword merge operations in neural machine translation. In Proceedings of Machine Translation Summit XVII: Research Track, pages 204-213, 2019. 5 +[20] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 2 +[21] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey. arXiv preprint arXiv:2308.14177, 2, 2023. 2 +[22] Biao Fu, Peigen Ye, Liang Zhang, Pei Yu, Cong Hu, Xiaodong Shi, and Yidong Chen. A token-level contrastive framework for sign language translation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 8 +[23] Philip Gage. A new algorithm for data compression. C Users Journal, 12(2):23-38, 1994. 5 +[24] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 1 +[25] Ximena Gutierrez-Vasques, Christian Bentz, Olga Sozinova, and Tanja Samardzic. From characters to words: the turning point of bpe merges. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3454-3468, 2021. 6 + +[26] Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindrich Helcl, and Alexandra Birch. Survey of low-resource machine translation. Computational Linguistics, 48 (3):673-732, 2022. 5 +[27] Awni Hannun. Sequence modeling with ctc. Distill, 2(11): e8, 2017. 4 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 7 +[29] Ruidan He, Linlin Liu, Hai Ye, Qingyu Tan, Bosheng Ding, Liying Cheng, Jiawei Low, Lidong Bing, and Luo Si. On the effectiveness of adapter-based tuning for pretrained language model adaptation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2208-2222, 2021. 1 +[30] Valentin Hofmann, Janet Pierrehumbert, and Hinrich Schütze. Dagobert: Generating derivational morphology with a pretrained language model. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3848-3861, 2020. 2 +[31] Jiahui Hou, Xiang-Yang Li, Peide Zhu, Zefan Wang, Yu Wang, Jianwei Qian, and Panlong Yang. Signspeaker: A real-time, high-precision smartwatch-based sign language translator. In The 25th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2019. 6 +[32] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8 +[33] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3766-3775. Association for Computational Linguistics, 2022. 7, 8 +[34] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 7 +[35] Sang-Ki Ko, Chang Jo Kim, Hyedong Jung, and Choongsang Cho. Neural sign language translation based on human key-point estimation. Applied sciences, 9(13):2683, 2019. 2 +[36] W Levelt. Producing spoken language. The neurocognition of language, pages 83-122, 1999. 2, 3 +[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 1, 2, 7, 8 +[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7 + +[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8 +[40] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 7 +[41] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412-1421, 2015. 8 +[42] Nathaniel FG Martin and James W England. Mathematical theory of entropy. Number 12. Cambridge university press, 2011. 5, 6 +[43] Igor Melnyk, Vijil Chenthamarakshan, Pin-Yu Chen, Payel Das, Amit Dhurandhar, Inkit Padhi, and Devleena Das. Reprogramming pretrained language models for antibody sequence infilling. In Proceedings of the 2023 International Conference on Machine Learning (ICML), 2023. 1 +[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 4 +[45] Arijit Nag, Bidisha Samanta, Animesh Mukherjee, Niloy Ganguly, and Soumen Chakrabarti. Entropy-guided vocabulary augmentation of multilingual language models for low-resource tasks. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8619-8629, 2023. 6 +[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 7 +[48] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 6 +[49] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 7 +[51] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2 +[52] Stephanie Seneff. The use of linguistic hierarchies in speech understanding. In in Proc. ICSLP, 1998. 2, 3 + +[53] Rico Sennrich and Biao Zhang. Revisiting low-resource neural machine translation: A case study. In 57th Annual Meeting of the Association for Computational Linguistics, pages 211–221. Association for Computational Linguistics (ACL), 2019. 5 +[54] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1 +[55] Xinying Song, Alex Salcianu, Yang Song, Dave Dopson, and Denny Zhou. Fast wordpiece tokenization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2089-2103, 2021. 5 +[56] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2021. 2 +[57] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. Advances in Neural Information Processing Systems, 29:1930-1938, 2016. 6, 7 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 7 +[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4 +[60] William Vicars. American sign language: “slow”. https://www.lifeprint.com/asl101/pages-signs/s/slow.htm, 2017.6 +[61] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 5, 6 +[62] Andreas Voskou, Konstantinos P Panousis, Dimitrios Kosmopoulos, Dimitris N Metaxas, and Sotirios Chatzis. Stochastic transformer networks with linear competing units: Application to end-to-end sl translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11946-11955, 2021. 2, 8 +[63] Ivan Vulic, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222-7240, 2020. 2 +[64] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1 +[65] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport for neural machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan + +guage Processing (Volume 1: Long Papers), pages 7361-7373, 2021. 5, 6 +[66] Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. Bigtrans: Augmenting large language models with multilingual translation capability over 100 languages. arXiv preprint arXiv:2305.18098, 2023. 1, 3 +[67] Aoxiong Yin, Tianyun Zhong, Li Tang, Weike Jin, Tao Jin, and Zhou Zhao. Gloss attention for gloss-free sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2551-2562, 2023. 2, 7, 8 +[68] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5975-5989, 2020. 1, 2 +[69] Biao Zhang, Mathias Müller, and Rico Sennrich. SLTUNET: A simple unified model for sign language translation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 7, 8 +[70] Jian Zhao, Weizhen Qi, Wengang Zhou, Nan Duan, Ming Zhou, and Houqiang Li. Conditional sentence generation and cross-modal reranking for sign language translation. IEEE Transactions on Multimedia, 24:2662-2672, 2021. 8 +[71] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2 +[72] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining, 2023. 1, 2, 7, 8 +[73] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1316-1325. Computer Vision Foundation / IEEE, 2021. 1, 2, 7, 8 +[74] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2022. 1, 2, 7, 8 +[75] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774, 2021. 2 +[76] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Jiajun Chen, Lei Li, and Shujian Huang. Multilingual machine translation with large language models: Empirical results and analysis. arXiv preprint arXiv:2304.04675, 2023.1, 3 \ No newline at end of file diff --git a/2024/LLMs are Good Sign Language Translators/images.zip b/2024/LLMs are Good Sign Language Translators/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e339f31b2fa19ba6dded8bb2af96425270577c8 --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cfbd72a60db4cffe16890512f7118641a4765a3e14a0a955c74143a5c34f1bae +size 353184 diff --git a/2024/LLMs are Good Sign Language Translators/layout.json b/2024/LLMs are Good Sign Language Translators/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff190aae51501aa47d78edf44b17605fa298a56 --- /dev/null +++ b/2024/LLMs are Good Sign Language Translators/layout.json @@ -0,0 +1,11373 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "spans": [ + { + "bbox": [ + 161, + 103, + 433, + 121 + ], + "type": "text", + "content": "LLMs are Good Sign Language Translators" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "spans": [ + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": "Jia Gong" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": " Lin Geng Foo" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": " Yixuan He" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": " Hossein Rahmani" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": " Jun Liu" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{1‡}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": "Singapore University of Technology and Design " + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 141, + 487, + 173 + ], + "type": "text", + "content": "Lancaster University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 174, + 432, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 174, + 432, + 199 + ], + "spans": [ + { + "bbox": [ + 159, + 174, + 432, + 199 + ], + "type": "text", + "content": "{jia_gong,lingeng_foo,yixuan_he}@mypmail.sutd.edu.sg, h.rahmani@lancaster.ac.uk, jun.liu@sutd.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 479 + ], + "type": "text", + "content": "Sign Language Translation (SLT) is a challenging task that aims to translate sign videos into spoken language. Inspired by the strong translation capabilities of large language models (LLMs) that are trained on extensive multilingual text corpora, we aim to harness off-the-shelf LLMs to handle SLT. In this paper, we regularize the sign videos to embody linguistic characteristics of spoken language, and propose a novel SignLLM framework to transform sign videos into a language-like representation for improved readability by off-the-shelf LLMs. SignLLM comprises two key modules: (1) The Vector-Quantized Visual Sign module converts sign videos into a sequence of discrete character-level sign tokens, and (2) the Codebook Reconstruction and Alignment module converts these character-level tokens into word-level sign representations using an optimal transport formulation. A sign-text alignment loss further bridges the gap between sign and text tokens, enhancing semantic compatibility. We achieve state-of-the-art gloss-free results on two widely-used SLT benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 693 + ], + "type": "text", + "content": "Sign languages, which are visual signals expressed through hand, body, and facial movements, serve as the primary means of communication within the hearing-impaired community. In an effort to facilitate effective communication with this community, much attention has been directed towards developing techniques to tackle the Sign Language Translation (SLT) task [10, 11, 69, 72, 73], where the goal is to translate sign videos into spoken language. SLT is a challenging task that requires cross-modality understanding of visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72]. Despite the notable advancements in terms of network architectures [10, 11, 37], visual sign representations (e.g., with keypoint estimators [14, 74])," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 264 + ], + "type": "text", + "content": "and training methods [22, 69, 72], how to effectively tackle the challenging cross-modal SLT task with limited paired sign-text data largely remains an open question." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 546, + 506 + ], + "type": "text", + "content": "On the other hand, large language models (LLMs) – referring to language models that have been trained on a large web-scale text corpus – have recently received a lot of attention. Since LLMs are trained over a very large corpus across multiple languages with distinct syntax and lexicon, they possess rich semantic understanding and powerful linguistic abilities [9, 15, 58]. At the same time, LLMs have also demonstrated an impressive capability to translate across multiple languages [9, 15], even showing a strong potential for translating languages with limited data [66, 76]. The foundation for this translation proficiency lies in the shared linguistic properties of syntax, lexicon, and morphology that many languages hold, which is particularly evident within language families [3]. Therefore, when faced with a new language with limited data, LLMs can draw upon the wealth of knowledge acquired from previously learned languages, leveraging any shared properties of syntax, lexicon and morphology with previous languages to effectively generate translations for new languages with remarkable accuracy and fluency [66, 76]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "Inspired by the impressive translation capabilities of LLMs, we aim to harness off-the-shelf LLMs to handle the challenging SLT task. However, training LLMs directly on the relatively small SLT dataset can potentially lead to forgetting of their rich knowledge [12, 29] and a decline in performance, thus we follow previous LLM-based works [24, 43, 54] to keep the off-the-shelf LLM frozen, which preserves the rich knowledge acquired during its pretraining on a vast multilingual corpus. Consequently, our focus shifts towards making the sign videos compatible and readable for the off-the-shelf and frozen LLM to perform SLT. Specifically, in this paper, we explore the following question: Can we treat the sign video as a form of language, and leverage an off-the-shelf and frozen LLM to translate them? Notably, this is not a straightforward task because directly encoding features from sign videos with a pre-trained feature extractor [28, 64] will result in a large gap between" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 211, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 211, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 211, + 713 + ], + "type": "text", + "content": "† Equal contribution; ‡ Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18362" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "the sign video features and text tokens, making it difficult for off-the-shelf LLMs to understand them." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 99, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 99, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 99, + 289, + 386 + ], + "type": "text", + "content": "Based on the observation that LLMs can effectively handle new languages by leveraging shared commonalities with previously learned languages, we aim to introduce designs that transform our sign videos into a language-like format which are readable and friendly to LLMs. Specifically, we hypothesize that providing language-like representations of sign videos to the LLM improves the LLM's understanding of the sign videos and facilitates greater exploitation of shared properties with previously learned languages, thus resulting in better SLT performance by the LLM. To obtain language-like sign video representations, we draw inspiration from linguistic studies and analyses on LLMs [30, 63] and regularize the sign video to embody two fundamental language-like characteristics: Discrete Characteristics: Spoken languages are inherently discrete, since each language contains a finite set of words (and subwords) that convey distinct concepts, allowing them to be naturally represented through a discrete vocabulary with distinct tokens [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52]. This hierarchical structure enables languages to express a wide range of words with a limited set of characters, and convey diverse sentences with a limited number of words." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 388, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 289, + 651 + ], + "type": "text", + "content": "In this paper, we present SignLLM, a novel framework designed to regularize input sign videos to produce sign token representations with language-like characteristics that are compatible and friendly to LLMs. Our proposed SignLLM includes two key designs to impart discrete characteristics and a hierarchical structure to the produced sign tokens. Firstly, we introduce the Vector-Quantized Visual Sign (VQ-Sign) module that facilitates the conversion of sign videos into a sequence of discrete character-level sign tokens. To achieve this, the VQ-Sign module consists of a discrete character-level sign codebook which is optimized through a self-supervised context prediction task. Next, we introduce the Codebook Reconstruction and Alignment (CRA) module that converts the character-level sign tokens into word-level sign tokens, facilitated by an optimal transport formulation. Moreover, we employ a sign-text alignment loss to further narrow the gap between the sign tokens and text tokens. These designs enable SignLLM to produce sign sentences that embody two key characteristics of spoken languages: discrete characteristics and a hierarchical structure, which enhances their compatibility with LLMs and makes them more readily interpretable by LLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 715 + ], + "type": "text", + "content": "After producing the language-like sign sentences, we feed them into an off-the-shelf and frozen LLM along with a text prompt that instructs the LLM to generate translations in the desired language. We empirically observe that, by employing our SignLLM's designs to align sign videos with" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "languages, we can already leverage a frozen LLM to attain state-of-the-art SLT performance. These findings suggest that our proposed SignLLM framework is a promising first step towards effectively harnessing LLMs for SLT. We hope our initial explorations can inspire future work within the community to leverage LLMs for SLT." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 144, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 264 + ], + "type": "text", + "content": "In summary, our main contributions are: (1) We propose a novel SignLLM framework that is the first to harness the power of off-the-shelf and frozen LLMs for SLT. (2) To make the input sign video compatible with LLMs, our SignLLM framework incorporates two designs: a VQ-Sign module to quantize the sign video into a sequence of discrete character-level sign tokens and a CRA module that transforms the character-level sign tokens to word-level sign tokens. (3) Through our proposed designs, we achieve state-of-the-art gloss-free results on two popular SLT datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 392, + 288 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 295, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 605 + ], + "type": "text", + "content": "Sign Language Translation (SLT) aims to transform sign videos into natural language sentences. It is a challenging task that requires understanding of both visual and linguistic cues [69, 72, 73], and the challenge is exacerbated by the limited availability of paired sign-text data [11, 13, 68, 69, 72] which limits the performance of SLT methods. To improve SLT performance, many previous works [10, 11, 14, 22, 34, 35, 37, 56, 68, 69, 72, 74] aim to enhance the visual sign representations and text decoding capabilities of SLT methods. Some works propose deep architectures based on RNNs [10, 35], GCNs [34], and Transformers [11, 37, 62, 68]. Other approaches include introducing a keypoint estimator to enhance the visual sign representation [14, 35, 56, 74], introducing pre-training tasks [22, 72], or jointly modelling several SLT-related tasks [69]. Some works also introduce larger datasets (e.g., How2Sign [20] and BOBSL [1]), which present a huge challenge with their large sign and text vocabularies. Besides, some recent works [67, 72] focus on the gloss-free setting – these works do not use sign gloss annotations for training, which reduces the cost of training SLT models, and our work also falls into this category. In contrast to existing works, we aim to harness the capabilities of off-the-shelf and frozen LLMs to perform SLT, by regularizing the sign videos into a language-like representation and prompting the LLM to generate text of the desired language." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 606, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 715 + ], + "type": "text", + "content": "Large Language Models (LLMs) refer to language models that have been trained extensively on a very large web-scale text corpus. LLMs have shown impressive text generation capabilities, attracting a lot of attention recently [21, 49, 71]. In particular, since LLMs have been trained on large amounts of text data, they demonstrate strong generalization capabilities across various text-based tasks, including code generation [51], open-domain question answering [75], and multilingual translation [9, 15]. Inspired by the re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18363" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 73, + 515, + 272 + ], + "blocks": [ + { + "bbox": [ + 77, + 73, + 515, + 272 + ], + "lines": [ + { + "bbox": [ + 77, + 73, + 515, + 272 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 515, + 272 + ], + "type": "image", + "image_path": "5fdbc12d114ffb55abef70f386f672c6c2cdc35f62277c31ce117b7d00a6220c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": "Figure 1. An overview of our SignLLM framework. During inference (top): Given an input sign video " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": ", we first pass it through our VQ-Sign module to obtain a sequence of discrete character-level sign tokens " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": ". Our VQ-Sign consists of a visual encoder " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": " to extract compact features and a character-level sign codebook " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": " for quantization to obtain " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": ". Next, we feed " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": " into our CRA module, which reorganizes " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": " by replacing short sequences of character tokens with word-level tokens via the word-level codebook, e.g., character sequence " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "[s_2, s_3, s_4]" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": " to word " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "s_2 s_3 s_4" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": ". This transforms the sign video data to a language-like sign sentence " + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 274, + 547, + 384 + ], + "type": "text", + "content": ", which is fed into the LLM along with a text prompt which guides the LLM to generate translations in the desired language. During training (bottom): We optimize VQ-Sign and its discrete sign codebook via a context prediction task, which seeks to recognize the future time steps based on the current context information. Next, for our CRA module, we construct the optimal word-level codebook by considering two aspects: entropy and size, which we address using optimal transport techniques. Then, we narrow the gap between the sign token space and LLM's text token space via minimizing the MMD loss, which improves the semantic compatibility between them." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 388, + 288, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 471 + ], + "type": "text", + "content": "cent advancements in LLMs, we explore harnessing LLMs for translation of sign videos, by converting the sign videos to a sequence of language-like sign tokens via our Sign-LLM framework, and treating the sign tokens as a form of language that can be translated by an LLM. To the best of our knowledge, we are the first work to leverage an off-the-shelf and frozen LLM to tackle SLT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 484, + 104, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 484, + 104, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 484, + 104, + 497 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 564 + ], + "type": "text", + "content": "In this section, we first introduce the overview of our SignLLM in Sec. 3.1. Then, we describe two main components of our SignLLM framework: the VQ-Sign and CRA modules in Sec. 3.2 and Sec. 3.3, respectively. Finally, we list the training and inference details in Sec. 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 574, + 163, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 163, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 163, + 586 + ], + "type": "text", + "content": "3.1. SignLLM Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 593, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 593, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 593, + 287, + 700 + ], + "type": "text", + "content": "In order to effectively handle SLT, in this paper we draw inspiration from LLMs' remarkable capabilities in generating translations across multiple languages [9, 15]. In particular, LLMs have been extensively trained on a large web-scale multilingual text corpus and have learned diverse knowledge regarding the properties of many languages, thus they are able to draw on shared commonalities with previously learned languages to effectively handle new languages with limited data [66, 76]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Therefore, to leverage the strong translation capabilities" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": "of LLMs to handle SLT, we introduce a novel SignLLM framework. SignLLM converts the input sign video " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " into a language-like sign sentence " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " that aligns with the linguistic characteristics of spoken languages and is friendly and compatible with LLMs. Then, to perform SLT, the language-like sign sentence " + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 388, + 545, + 484 + ], + "type": "text", + "content": " can be fed into an off-the-shelf and frozen LLM along with a text prompt that guides the LLM to generate translations in the desired language." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "text", + "content": "Specifically, to produce sign sentences " + }, + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "text", + "content": " that are friendly and understandable by LLMs, we aim to regularize our sign sentences " + }, + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 485, + 546, + 604 + ], + "type": "text", + "content": " to embody two core linguistic characteristics: Discrete Characteristics: Spoken languages are naturally discrete and consist of distinct words or sub-words with corresponding discrete tokens in a vocabulary [8, 59]. Hierarchical Structure: Most spoken languages exhibit three hierarchical semantic levels – the sentence, word, and character levels [36, 52], where words are composed from characters and sentences are composed from various words." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "To achieve the above, our SignLLM framework comprises of three parts, as shown in Fig. 1: (1) The VQ-Sign module converts the input sign video " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " into a sequence of discrete sign tokens " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": ", aligning the sign representations with text's discrete characteristics. These sign tokens " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " are character-level sign tokens that are retrieved from a learned discrete character-level codebook. (2) The CRA module maps meaningful compositions of character-level sign tokens " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " into word-level sign tokens that form a sign sentence" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18364" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": ", further imparting a language-like hierarchical structure to the video sign representations. Moreover, we also align the sign token codebooks towards the text token space to improve semantic compatibility. (3) An off-the-shelf LLM takes the sign sentence " + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": " as input, along with an instructive text prompt that guides the LLM to generate the translation in the desired language. More details regarding the text prompt are in the Supplementary. Next, we present our VQ-Sign and CRA modules in detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 192, + 248, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 248, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 248, + 205 + ], + "type": "text", + "content": "3.2. Vector-Quantized Visual Sign Module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "text", + "content": "First, in order to produce a language-like representation, we would like to impart discrete characteristics to the input sign videos, aligning them more closely with spoken language representations that are inherently discrete and consist of distinct tokens in a vocabulary. However, achieving this is not straightforward because sign videos are a continuous signal in a high-dimensional spatio-temporal space which cannot be easily represented by a set of discrete tokens, and for which the vocabulary is not readily available. Hence, we introduce our Vector-Quantized Visual Sign (VQ-Sign) module to quantize the sign video " + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "text", + "content": " into a sequence of discrete sign tokens " + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "text", + "content": " via a sign codebook " + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 46, + 212, + 289, + 379 + ], + "type": "text", + "content": ". As illustrated in Fig. 1, our VQ-Sign module involves a series of steps, which we present in detail next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": "In the first step, we extract a compact feature " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " from the high-dimensional input sign video " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " is the number of video frames, while " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " are the height and width of the video frames respectively. To be precise, the sign video " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " is first organized into a sequence of short overlapping video clips, then each short video clip is fed into a visual encoder " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " to extract a compact feature representation of dimensionality " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": ". Overall, this step transforms the original high-dimensional input sign video " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " into a compact feature " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{\\frac{N}{n} \\times d}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " represents the number of frames between the start of neighboring clips. Notably, since " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " is obtained by processing " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\frac{N}{n}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " short clips, " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " can also be seen as a sequence of " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\frac{N}{n}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " clip-wise features, i.e., " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "\\{z_{t}\\}_{t=1}^{\\frac{N}{n}}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": " corresponds to the feature of the " + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 380, + 289, + 564 + ], + "type": "text", + "content": "-th short clip." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": "In the next step, we transform the feature " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "Z = \\{z_{t}\\}_{t=1}^{\\frac{N}{n}}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " into a sequence of discrete tokens " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " using a codebook " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": ". Specifically, we discretize each clip's feature " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " into a discrete token " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " by finding the matching token " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_{j}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " from the codebook " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c = \\{s_i\\}_{i=1}^M" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": "-th token in the codebook is denoted as " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " is the number of tokens in the codebook. The matching token " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_{j}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " is the codebook's closest element to the feature " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " in terms of Euclidean distance, i.e., " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "j = \\arg \\min_{i}(\\| z_{t} - s_{i}\\|_{2}^{2})" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": ". After the matching, each feature " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}_{t} = s_{j}" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": ", which results in a discrete token sequence as shown in Fig. 1, e.g., " + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "inline_equation", + "content": "[s_2, s_3, s_4, s_M, s_5, s_1]" + }, + { + "bbox": [ + 46, + 565, + 289, + 715 + ], + "type": "text", + "content": ". Note that, we randomly initialize all" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": "the tokens " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^M" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " in the sign codebook " + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 304, + 72, + 545, + 96 + ], + "type": "text", + "content": " at the start and optimize them during training, as introduced next." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 546, + 312 + ], + "type": "text", + "content": "However, we face a challenge in learning the discrete codebook " + }, + { + "bbox": [ + 304, + 96, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 304, + 96, + 546, + 312 + ], + "type": "text", + "content": ". In particular, although autoencoding [59] has been a popular method for producing a codebook of discrete units, the high complexity of sign videos makes autoencoding (i.e., self-reconstruction of sign videos) challenging and costly. Therefore, inspired from predictive coding [4], a widely-used method in text and speech representation learning [6, 44, 46], we propose to learn discrete representations of sign videos through a context prediction task. Context prediction [46] is a self-supervised task that focuses on recognizing the future content in latent space based on the current information, which can learn discrete representations while eliminating the need for reconstructing the high-dimensional input video data. Furthermore, previous works show that training with context prediction effectively captures the temporal dependencies and relationships between elements in a sequence [5, 27], and the learned representations are often transferable to downstream tasks [6, 46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": "Specifically, we employ a context prediction task where we try to distinguish a future sample " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "z_{\\tau + k}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": " based on the current context representation " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "c_{\\tau}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": " at various time steps " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": ". To facilitate this task, after we obtain the discrete token sequence " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": ", we further produce a context latent representation " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "c_{\\tau}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": " using an auto-regressive model " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": " that summarizes all the discrete tokens before a certain time step " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\{\\hat{z}_t\\}_{t \\leq \\tau}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": ") to produce context latent representations " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "c_{\\tau} = g(\\{\\hat{z}_t\\}_{t \\leq \\tau})" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": ". Then, we optimize our module by minimizing the following context prediction contrastive loss " + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k^{\\mathrm{cp}}" + }, + { + "bbox": [ + 304, + 312, + 547, + 433 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 437, + 545, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 310, + 437, + 545, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {k} ^ {\\mathrm {c p}} = - \\sum_ {\\tau = 1} ^ {\\frac {N}{n} - k} \\left(\\log \\sigma \\left(z _ {\\tau + k} ^ {\\top} h _ {\\tau}\\right) + \\lambda \\underset {\\widetilde {z} \\sim p _ {n}} {\\mathbb {E}} \\left[ \\log \\sigma \\left(- \\widetilde {z} ^ {\\top} h _ {\\tau}\\right) \\right]\\right) \\tag {1}", + "image_path": "891ea45dfd52ce23c9ed9f7fbed41e39760f384bce42e4bdc32d2dfe78129143.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "h_\\tau" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " is obtained by applying a trainable linear layer to " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "c_{\\tau}" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\sigma(z_{\\tau+k}^{\\top} h_{\\tau})" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " is the probability of " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "z_{\\tau+k}" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " being the true sample among the negatives, " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\widetilde{z}" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " are negative samples drawn from a minibatch " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "p_n" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " is a hyperparameter. We sum " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_k^{\\mathrm{cp}}" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " over different step sizes " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " to obtain the context prediction loss " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{\\mathrm{cp}} = \\sum_{k=1}^{K} \\mathcal{L}_k^{\\mathrm{cp}}" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 477, + 545, + 561 + ], + "type": "text", + "content": " is the maximum number of future clips that we are interested in." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": "Following [18], in order to optimize the matching between " + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": ", we further add two losses to optimize the matching distance between " + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": ", such that the overall loss " + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{VQ}" + }, + { + "bbox": [ + 304, + 562, + 545, + 609 + ], + "type": "text", + "content": " to optimize our VQ-Sign module is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 614, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 310, + 614, + 545, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {V Q} = \\mathcal {L} ^ {\\mathrm {c p}} + \\sum_ {t = 1} ^ {N / n} \\| \\operatorname {s g} \\left(z _ {t}\\right) - \\hat {z} _ {t} \\| ^ {2} + \\gamma \\sum_ {t = 1} ^ {N / n} \\| z _ {t} - \\operatorname {s g} (\\hat {z} _ {t}) \\| ^ {2}, \\tag {2}", + "image_path": "a10f2e6d05d2aa2d390f0e21af90c91ac34403b383b97df35539abd5bcd48ba7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname{sg}(z) \\equiv z" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " is the stop-gradient operator and " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": " is a hyperparameter. By optimizing " + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{VQ}" + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": ", we can train our VQ-Sign and the discrete codebook without the need for reconstructing high-dimensional video clips, which makes the codebook construction a viable and relatively cheap option." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18365" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": "In summary, our VQ-Sign transforms sign videos into sequences of discrete sign tokens " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": ", which are friendlier and more understandable to LLMs. Notably, the produced discrete tokens " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": " can be likened to character-level tokens, in the sense that each discrete token " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": " corresponds to a short clip and may not contain much semantic meaning on its own (similar to linguistic characters), but they can be combined into a sequence to convey a clear semantic meaning (akin to forming a word or sentence). Thus, inspired by this, we call VQ-Sign's codebook " + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 47, + 72, + 289, + 204 + ], + "type": "text", + "content": " the character-level codebook that contains character-level sign tokens." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 264, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 264, + 228 + ], + "type": "text", + "content": "3.3. Codebook Reconstruction and Alignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 232, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 287, + 340 + ], + "type": "text", + "content": "In the previous section, we quantize sign videos into discrete character-level sign tokens, which aligns them closer to language representations. In this section, our goal is to impart a hierarchical structure to our sign video representations, which makes them align even closer to language representations. Specifically, we aim to compose our character-level sign tokens into word-level sign tokens to mirror the observed hierarchical structure in spoken language, which makes them even more compatible with LLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 341, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 287, + 484 + ], + "type": "text", + "content": "Intuitively, considering a spoken language sentence, we can represent it as a sequence of words, with each word formed by one or multiple characters. For example, a sentence 'I love AI' can be decomposed into a word sequence [I', 'love', 'AI'], where the word 'love' is in turn formed by the character sequence [1', o', v', e']. We observe that, although each individual character may not contain much semantic meaning on its own, they can be composed to form words with clearer semantic meaning. In a similar fashion, we also want to impart such hierarchical structure to our character-level sign tokens by composing them to form meaningful word-level sign tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "text", + "content": "Hence, we aim to find an optimal transformation from character-level sign tokens to word-level sign tokens for enhanced readability and compatibility with LLMs. To this end, we introduce the Codebook Reconstruction and Alignment (CRA) module to transform the character-level codebook " + }, + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "text", + "content": " from VQ-Sign into a word-level codebook " + }, + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^w" + }, + { + "bbox": [ + 47, + 485, + 287, + 687 + ], + "type": "text", + "content": " whose tokens convey richer and clearer semantic meaning. Inspired by optimal transport methods [16, 61, 65], we observe that the above transformation can be formulated as an optimal transport problem of transporting characters into words, thus we introduce a codebook reconstruction algorithm with an optimal transport formulation to find an optimal transformation. Additionally, to further reduce the distribution gap between the sign tokens and the text tokens, our CRA module also performs sign-text alignment, enhancing the semantic compatibility of the sign tokens with LLMs. We introduce the details below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To begin, the objective of our Codebook Reconstruction Algorithm is to create a word-level codebook " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^w" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": "based on VQ-Sign's character-level codebook " + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 304, + 72, + 545, + 360 + ], + "type": "text", + "content": ". The challenge lies in determining which character-level sign tokens should be assembled together to form word-level sign tokens, which is a complex problem. To address this complexity, we adopt an approach based on two fundamental principles. Firstly, in order to maximize the overall predictability of the word-level tokens and enhance the distinctiveness of each token, we seek to minimize the entropy of each word-level token within the vocabulary [42]. We remark that, several approaches for establishing language-based subword vocabularies [8, 55] can be seen as entropy-minimizing approaches, employing different heuristics to establish the vocabulary with the goal of minimizing entropy [23]. On the other hand, considering the limited availability of sign video data, we incorporate codebook size as another key factor in our word-level codebook construction, since studies on languages with limited data [26, 53] have also identified vocabulary size as a crucial aspect. Specifically, too small a vocabulary can result in sub-optimal entropy values, while an excessively large vocabulary size can lead to issues such as parameter explosion and token sparsity which hinder understanding [2], and finding the right balance between these effects becomes even more sensitive for languages with limited data [19, 53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": "Based on these principles, our objective is to determine an optimal codebook size that maximizes the entropy reduction while taking into account the increase in codebook size. In other words, we would like to find an optimal codebook size that maximizes the gradient of the entropy reduction with respect to the codebook size increase. To simplify the optimal size searching problem, we define a fixed size increment " + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": " and search through codebooks of various sizes (where the difference between each codebook size is " + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": " tokens). Specifically, we define the " + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": "-th codebook " + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "inline_equation", + "content": "(\\mathbb{S}_r^w)" + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": " as the codebook with " + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "inline_equation", + "content": "r \\times m" + }, + { + "bbox": [ + 304, + 363, + 545, + 553 + ], + "type": "text", + "content": " tokens. Then, we seek to identify the optimal set of word-level tokens, where each word-level token is composed of character-level tokens. We approach this by formulating the character compositions as an optimal transport problem, where characters are transported to words." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "However, it can be challenging to identify specific character combinations that convey precise semantic information, due to the temporal complexity of sign videos, which often makes the character-level token sequences " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " quite messy. For instance, some signers may execute the signing motions at a slower speed, which can lead to consecutive short video clips being highly similar, resulting in consecutively-repeating character-level discrete tokens. Thus, the character-level sequences between different signers can differ significantly (e.g., " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "[s_1,s_2]" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "[s_1,s_1,s_1,s_2,s_2]" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": ") due to such duplication of character-level tokens, even though they may contain the same semantic information. At the same time, simply filtering out re" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18366" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "peated character-level tokens straightforwardly (e.g., setting all " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "[s_1, s_1, s_1, s_2, s_2]" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "[s_1, s_2]" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ") is sub-optimal, since the speed of some signs can also convey some information [31, 60], e.g., if a signer signs \"ugly\" quickly, it conveys \"very ugly\" in American Sign Language." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": "Pre-processing of Repeated Characters. Therefore, to alleviate the impact of the signer's speed while keeping the information regarding each sign's speed, we first preprocess the character-level sequences as follows: First, we find all the repeated tokens in the character-level sequence and compute the average number of repeated tokens " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " in each sequence. Then, for each repeated sequence (e.g., " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "[s_1, s_1, s_1]" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": "), we keep the first character and remove the tailing repetitive ones (e.g., " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "[s_1, s_1, s_1]" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "[s_1]" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": "). At the same time, if the character-level tokens repeat more than " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " times, we insert a single character-level token " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " as a \"slowing down\" sign, e.g., " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "[s_1, s_1, s_1]" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "[s_1, s_0]" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "inline_equation", + "content": "\\alpha < 3" + }, + { + "bbox": [ + 46, + 133, + 288, + 346 + ], + "type": "text", + "content": ". Crucially, this allows us to reduce redundancy, while still representing \"quick\" or \"slow\" signs that account for differences in the signer's speed. Overall, this pre-processing and reducing of repeating characters makes the character-level sequence less messy, facilitating the search for specific meaningful character combinations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": "Optimal Transport Formulation. Then, using the preprocessed character-level sequences, we search for an optimal word-level codebook via an optimal transport formulation [16, 61, 65]. Following our discussion above, we aim to find an optimal word-level codebook (i.e., " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": ") with low entropy and compact size. Specifically, to measure codebook entropy, we follow [25, 42, 45] to define the entropy of the " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": "-th codebook " + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 46, + 347, + 287, + 443 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 447, + 287, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 447, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 100, + 447, + 287, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {\\mathbb {S} _ {r} ^ {w}} = - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} P (w _ {j}) \\log P (w _ {j}), \\tag {3}", + "image_path": "4f9cfe3dcce4da2f5bb7d19fad8a033a01f5f9dba83c670923439783d30397c0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "P(w_{j})" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": " is the relative frequency of the " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": " token " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": " from the word-level codebook " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": ". Then, based on VQ-Sign's character-level codebook " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": ", the entropy of the word-level codebook " + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 47, + 474, + 287, + 534 + ], + "type": "text", + "content": " can be computed through the following (with proof in Supplementary):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 536, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 536, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 71, + 536, + 287, + 587 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal{H}_{\\mathbb{S}_{r}^{w}} = -\\sum_{w_{j}\\in \\mathbb{S}^{w}}\\sum_{s_{i}\\in \\mathbb{S}^{c}}P(w_{j},s_{i})\\log P(w_{j},s_{i}) \\\\ - \\sum_ {w _ {j} \\in \\mathbb {S} _ {r} ^ {w}} \\sum_ {s _ {i} \\in \\mathbb {S} ^ {c}} P \\left(w _ {j}, s _ {i}\\right) (- \\log P \\left(s _ {i} \\mid w _ {j}\\right)), \\tag {4} \\\\ \\end{array}", + "image_path": "d132d0ea66da4f5bf24b595a8157074a4210e5019ba989544ba7424bd049234e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "P(s_{i}|w_{j})" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": " is the probability of the character-level token " + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": " appearing in the word-level token " + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 46, + 594, + 287, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "Then, to formally define our objective function, we follow existing works [16, 48, 65] to define a transport matrix " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " that represents the assignments of characters to words, and a distance matrix " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{D}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " that represents the cost of transportation. We define the transport matrix " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{P} \\in \\mathbb{R}^{m \\times (r \\cdot m)}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " with the " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "(j,i)" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "-th element as " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "P(w_{j}, s_{i})" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", and define the distance matrix " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{D} \\in \\mathbb{R}^{m \\times (r \\cdot m)}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " as a matrix whose " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "(j,i)" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "-th element is " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\log P(s_{i} | w_{j})" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ". Note that, if " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "inline_equation", + "content": "s_{i}" + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": ", we use" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\mathrm{length}(w_j)}" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": " to estimate " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "P(s_i|w_j)" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": ", and if " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": " does not contain " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": ", then it is deemed an infeasible assignment, and we set the distance " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "\\log P(s_i|w_j) = \\infty" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": ". We further define " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "H(\\pmb{P})" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "-\\sum_{w_j \\in \\mathbb{S}_r^w} \\sum_{s_i \\in \\mathbb{S}^c} P(w_j, s_i) \\log P(w_j, s_i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": ", which is simply the entropy of the probability distribution " + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "inline_equation", + "content": "P(w_j, s_i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 135, + 545, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 135, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 305, + 135, + 545, + 159 + ], + "type": "text", + "content": "Hence, based on Eq. 4, the objective function for minimizing the entropy of " + }, + { + "bbox": [ + 305, + 135, + 545, + 159 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 305, + 135, + 545, + 159 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 163, + 545, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 163, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 340, + 163, + 545, + 187 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {P} \\in \\mathbb {R} ^ {m} \\times (r \\cdot m)} {\\arg \\min } H (\\boldsymbol {P}) + \\sum_ {j} \\sum_ {i} \\boldsymbol {P} (j, i) \\boldsymbol {D} (j, i). \\tag {5}", + "image_path": "0cdb7d2ffdcd3b29a08fcf96925032a1e88079f3a74659309452eed5bc42da5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": "Specifically, following previous works [16, 48, 65], we impose two constraints on the transport matrix " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": ": the sum of each row in " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " should equal to the probability of character token " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " and the sum of each column in " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " should equal to the probability of word token " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "w_j" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": ". Formally, we constrain the transport matrix " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " with: " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "|\\sum_{i}P(i,j) - P(w_{j})|\\leq \\epsilon" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "|\\sum_{j}P(i,j) - P(s_{i})|\\leq \\epsilon" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 190, + 545, + 285 + ], + "type": "text", + "content": " is a small positive fixed hyperparameter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": "Intuitively, this optimization process can be regarded as an optimal transport problem to find the best way to transport mass from " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^c" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": ". To handle this optimal transport problem, we leverage the Sinkhorn algorithm [16, 48, 65], allowing us to effectively construct the candidate word-level codebooks " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": " with minimal entropy. Since our increment " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": " between each candidate codebook is fixed, we can find the optimal codebook size that maximizes the gradient of entropy reduction by simply computing and finding the maximum the entropy difference between " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_r^w" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{r - 1}^w" + }, + { + "bbox": [ + 304, + 286, + 545, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 490 + ], + "type": "text", + "content": "After finding the optimal word-level codebook, we construct the word-level sign tokens by composing all features of the character-level tokens into the word-level tokens via our autoregressive model " + }, + { + "bbox": [ + 304, + 407, + 545, + 490 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 407, + 545, + 490 + ], + "type": "text", + "content": ". Refer to Supplementary for more details. Overall, with our codebook reconstruction algorithm, we can construct an optimal word-level sign codebook with low entropy yet also with relatively small size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "text", + "content": "Sign-Text Alignment. Next, we further align the sign tokens with the text tokens used in LLMs in order to further improve semantic compatibility between them. To achieve this, we measure the distribution gap between sign tokens and text tokens via Maximum Mean Discrepancy (MMD) [57] and then optimize the sign tokens' embeddings by minimizing MMD, which narrows down the distribution gap. Specifically, we compute the gap between the sign embedding space " + }, + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^s" + }, + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "text", + "content": " and text embedding space " + }, + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^t" + }, + { + "bbox": [ + 304, + 495, + 545, + 604 + ], + "type": "text", + "content": " via MMD as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 608, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 608, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 321, + 608, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} ^ {M M D} (\\mathbb {F} ^ {s}, \\mathbb {F} ^ {t}) = \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {s}} \\frac {k (f (p _ {i}) , f (p _ {j}))}{n _ {s} ^ {2}} \\\\ + \\sum_ {i = 1} ^ {n _ {t}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {k \\left(q _ {i} , q _ {j}\\right)}{n _ {t} ^ {2}} - \\sum_ {i = 1} ^ {n _ {s}} \\sum_ {j = 1} ^ {n _ {t}} \\frac {2 \\cdot k \\left(f \\left(p _ {i}\\right) , q _ {j}\\right)}{n _ {s} n _ {t}}, \\tag {6} \\\\ \\end{array}", + "image_path": "63d7d9d14e84c440fad046b4cbcafebf8d038b873d6c9987760936aa6036b3eb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the tokens in " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^s" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " is a small projection module that projects " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^s" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "n_s" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "n_t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are the numbers of tokens in " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^s" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{F}^t" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "k(\\cdot)" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " represents the radial" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18367" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "kernel [57] that measures the distance between two samples. We apply MMD loss to both the word-level and the character-level tokens to narrow the overall sign-text gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 117, + 178, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 117, + 178, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 117, + 178, + 129 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": "Inference. Given a sign video " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": ", we first extract compact features " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " via the visual encoder " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": ", and quantize " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " via VQ-Sign's learned character-level codebook. Then, we transform the sequence of discrete character-level tokens " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " into word-level tokens via our CRA, which produces a sign sentence " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": ". Lastly, we project the sign sentence " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " to LLM embedding space via the small projection module " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": ", and then feed the sign sentence " + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 136, + 286, + 255 + ], + "type": "text", + "content": " into the LLM along with a text prompt to instruct the LLM to perform the SLT task and generate text in the desired language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "content": "Training. Our SignLLM is optimized in two stages: pretraining and fine-tuning. Specifically, the pre-training stage, which does not require explicit SLT supervision, includes two sub-stages: (i) We first pre-train VQ-Sign via the context prediction task with " + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{VQ}" + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "content": " in Eq. 2. (ii) Then, based on the VQ-Sign's learned character-level codebook, we construct the word-level codebook with the codebook reconstruction algorithm (Sec. 3.3) and apply the MMD loss " + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{MMD}" + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "content": " in Eq. 6 to align the sign codebooks " + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "inline_equation", + "content": "(\\mathbb{S}^c" + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "inline_equation", + "content": "\\mathbb{S}^w)" + }, + { + "bbox": [ + 46, + 255, + 286, + 375 + ], + "type": "text", + "content": " and the text vocabulary of the desired language." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": "After the pre-training, we fine-tune SignLLM. To aid LLMs in understanding the sign sentences as texts, we additionally maximize the similarity between the text tokens " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " generated by LLM and the ground truth tokens " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\bar{Y}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " as: " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{sim}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " by minimizing the cross-entropy between them. We fine-tune our SignLLM (with frozen LLM) via the loss " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ft}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " as follows: " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{ft} = \\mathcal{L}^{VQ} + \\lambda_1\\mathcal{L}^{MMD} + \\lambda_2\\mathcal{L}^{sim}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": " are hyperparameters. Note that we follow previous gloss-free works [67, 72] to train our SignLLM framework, eliminating the need for additional gloss data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 128, + 520 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 180, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 180, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 180, + 540 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": "Our visual encoder " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " is constructed by appending two Conv3D layers with a kernel size of (5, 3, 3) and a stride of (2, 1, 1) to a ResNet18 [28] pre-trained on ImageNet [17]. Each clip consists of 13 frames and the gap between the neighboring clips (" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": ") is 4. Besides, the auto-regressive model " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " is implemented as a Convolutional Gated Recurrent Layer with a kernel size of (1, 1). We set the total number of discrete vectors " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " at 256 and each vector's dimension " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " at 1024 for our character-level codebook. In VQ-Sign's pre-training phase, we set " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 0.25" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": ", initialize the learning rate at 0.01 and use the Adam algorithm, training the model to predict the future three clips (" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": ") for 200 epochs. During codebook reconstruction, we set the increment " + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 545, + 286, + 713 + ], + "type": "text", + "content": " to 32. We employ the frozen LLaMA-7B-16bit [58]" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "as our LLM and project the codebook space to LLaMA's embedding space via " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", which consists of two fc-layers with ReLU. We set " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.5" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", and initialize the learning rate at 0.001 to fine-tune our SignLLM over 20 epochs. Please see Supplementary for more implementation details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 481, + 152 + ], + "type": "text", + "content": "4.2. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 265 + ], + "type": "text", + "content": "Datasets. We follow previous works [13, 14, 34, 72-74] to run experiments on the Phoenix-2014T [10] and CSL-Daily [73] datasets for SLT, and evaluate on their dev and test sets. Phoenix-2014T [10] is a German sign language dataset with a vocabulary size of 2887 German words. The training, dev, and test sets contain 7096, 519, and 642 samples. CSL-Daily [73] is a Chinese sign language dataset with a vocabulary size of 2343 Chinese words. The training, dev, and test sets contain 18401, 1077, and 1176 samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous works [13, 14, 37, 67, 72-74], we adopt BLEU [47] and ROUGE-L [38] as the evaluation metrics for SLT. BLEU-n evaluates the average translation precision up to n-grams, and we follow previous works [13, 14, 72-74] to report results for BLEU-1 to BLEU-4 (i.e., B1, B2, B3 and B4). ROUGE-L (or ROUGE) computes the F1 score based on the longest common subsequence between the predicted and ground truth texts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 369, + 391, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 391, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 391, + 380 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": "Results on PHOENIX2014T dataset. Tab. 1 presents a comparison of our approach with state-of-the-art gloss-based and gloss-free methods for SLT. Our method consistently improves upon all reported metrics as compared to other gloss-free approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 496 + ], + "type": "text", + "content": "Results on CSL-Daily dataset. We compare our method with state-of-the-art approaches on the CSLD Daily dataset in Tab. 2. We outperform previous gloss-free works on all metrics, showing the efficacy of our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 503, + 400, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 400, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 400, + 516 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 569 + ], + "type": "text", + "content": "To further investigate the proposed method, we follow previous works [13, 33, 67, 69, 72] to conduct extensive ablation experiments on the Phoenix-2014T dev and test sets. Refer to Supplementary for more experiment results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "Impact of LLM. First, we evaluate the impact of leveraging LLMs for the SLT task. Specifically, we establish three baselines: 1) Ours (w/o LLM) where we replace the LLM with a trainable lightweight text generator (mBART [40]). 2) Ours (w/ T5) where we replace our LLM with a smaller LLM (T5 [50]). As shown in Tab. 3, our approach (w/ LLaMA [58]) achieves a much better performance than Ours (w/o LLM), showing the efficacy of leveraging a powerful LLM. We also find that using a smaller and less powerful LLM (T5) leads to worse performance, as expected." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Impact of SignLLM. Next, we explore the impact of Sign-LLM by comparing against the following baselines: 1) En" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18368" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 80, + 482, + 210 + ], + "blocks": [ + { + "bbox": [ + 208, + 71, + 384, + 80 + ], + "lines": [ + { + "bbox": [ + 208, + 71, + 384, + 80 + ], + "spans": [ + { + "bbox": [ + 208, + 71, + 384, + 80 + ], + "type": "text", + "content": "Table 1. Results on Phoenix-2014T dataset [10]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 80, + 482, + 210 + ], + "lines": [ + { + "bbox": [ + 111, + 80, + 482, + 210 + ], + "spans": [ + { + "bbox": [ + 111, + 80, + 482, + 210 + ], + "type": "table", + "html": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]47.2634.4027.0522.38-46.6133.7326.1921.32-
ConSLT [22]---21.1147.74---21.5947.69
STN-SLT [62]49.1236.2928.3423.23-48.6135.9728.3723.65-
STMC-T [74]47.6036.4329.1824.0948.2446.9836.0928.7023.6546.65
BN-TIN-Transf.+SignBT [73]51.1137.9029.8024.4550.2950.8037.7529.7224.3249.54
PET [33]-----49.5437.1929.3024.0249.97
MMTLB [13]53.9541.1233.1427.6153.1053.9741.7533.8428.3952.65
TS-SLT [14]54.3241.9934.1528.6654.0854.9042.4334.4628.9553.48
SLTUNET [69]---27.8752.2352.9241.7633.9928.4752.11
Gloss-freeNSLT [10]28.1016.8111.829.1231.0027.1015.6110.828.3529.70
NSLT+Bahdanau [7, 10]31.8719.1113.169.9431.8032.2419.0312.839.5831.80
NSLT+Luong [10, 41]31.5818.9813.2210.0032.6029.8617.5211.969.0030.70
TSPNet [37]-----36.1023.1216.8813.4134.96
CSGCR [70]35.8524.7718.6515.0838.9636.7125.4018.8615.1838.85
GASLT [67]-----39.0726.7421.8615.7439.86
GFSLT-VLP [72]44.0833.5626.7422.1243.7243.7133.1826.1121.4442.49
Ours46.8836.5929.9125.2547.2345.2134.7828.0523.4044.49
", + "image_path": "457d52e8231e3c9de074ba49e61839a276c444444d2320748be3e4c6875f2c3c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 111, + 220, + 481, + 310 + ], + "blocks": [ + { + "bbox": [ + 132, + 211, + 462, + 220 + ], + "lines": [ + { + "bbox": [ + 132, + 211, + 462, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 211, + 462, + 220 + ], + "type": "text", + "content": "Table 2. Results on CSL-Daily dataset [73]. * means that the result was reproduced by [72]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 220, + 481, + 310 + ], + "lines": [ + { + "bbox": [ + 111, + 220, + 481, + 310 + ], + "spans": [ + { + "bbox": [ + 111, + 220, + 481, + 310 + ], + "type": "table", + "html": "
SettingMethodDevTest
B1B2B3B4ROUGEB1B2B3B4ROUGE
Gloss-basedSLRT [11]37.4724.6716.8611.8837.9637.3824.3616.5511.7936.74
ConSLT [22]---14.8041.46---14.5340.98
BN-TIN-Transf.+SignBT [73]51.4637.2327.5120.8049.4951.4237.2627.7621.3449.31
MMTLB [13]53.8140.8431.2924.4253.3853.3140.4130.8723.9253.25
TS-SLT [14]55.2142.3132.7125.7655.1055.4442.5932.8725.7955.72
SLTUNET [69]---23.9953.5854.9841.4431.8425.0154.08
Gloss-freeSLRT* [11]21.039.975.964.0420.5120.009.114.933.0319.67
NSLT+Luong [10, 41]34.2219.7212.247.9634.2834.1619.5711.847.5634.54
GASLT [67]-----19.909.945.984.0720.35
GFSLT-VLP [72]39.2025.0216.3511.0736.7039.3724.9316.2611.0036.44
Ours42.4526.8817.9012.2339.1839.5528.1320.0715.7539.91
", + "image_path": "46125021c828286f460affc2ebd21b1e0f0bd62e7ddd23a278c96fa6406d50fd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 327, + 288, + 363 + ], + "blocks": [ + { + "bbox": [ + 88, + 315, + 246, + 325 + ], + "lines": [ + { + "bbox": [ + 88, + 315, + 246, + 325 + ], + "spans": [ + { + "bbox": [ + 88, + 315, + 246, + 325 + ], + "type": "text", + "content": "Table 3. Ablation study for impact of LLM" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 327, + 288, + 363 + ], + "lines": [ + { + "bbox": [ + 49, + 327, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 327, + 288, + 363 + ], + "type": "table", + "html": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o LLM)29.7520.0414.9611.9527.2018.2913.3210.36
Ours (w/ T5)44.2034.5527.1522.9044.0334.1227.2322.51
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "image_path": "c540f83855d106cfb32a1f9091552b5853fcc92238e614438f83ab3954d75ceb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 374, + 288, + 411 + ], + "blocks": [ + { + "bbox": [ + 79, + 363, + 255, + 373 + ], + "lines": [ + { + "bbox": [ + 79, + 363, + 255, + 373 + ], + "spans": [ + { + "bbox": [ + 79, + 363, + 255, + 373 + ], + "type": "text", + "content": "Table 4. Ablation study for impact of SignLLM." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 374, + 288, + 411 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 288, + 411 + ], + "type": "table", + "html": "
MethodDevTest
B1B2B3B4B1B2B3B4
Encoder Only26.9517.0412.119.4125.6316.1011.208.42
Encoder + FT39.2929.2922.5518.1839.9229.1422.5418.17
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "image_path": "97f63bd67aab9bbf51578b2412bc62f884e424828b94b203556f8fb46784b3a2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 423, + 288, + 459 + ], + "blocks": [ + { + "bbox": [ + 59, + 412, + 274, + 422 + ], + "lines": [ + { + "bbox": [ + 59, + 412, + 274, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 412, + 274, + 422 + ], + "type": "text", + "content": "Table 5. Ablation study for main components of SignLLM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 423, + 288, + 459 + ], + "lines": [ + { + "bbox": [ + 50, + 423, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 50, + 423, + 288, + 459 + ], + "type": "table", + "html": "
MethodDevTest
B1B2B3B4B1B2B3B4
Ours (w/o VQ-Sign)35.7824.6518.0814.1533.1422.8016.7413.13
Ours (w/o Codebook Reconstruction)40.4530.4024.0719.7940.2530.0523.6319.47
Ours (w/o Sign-text Alignment)29.0519.3313.7210.4028.6719.2213.7310.63
Ours46.8836.5929.9125.2545.2134.7828.0523.40
", + "image_path": "5b551cd41ab3e1bd43bd1b0006e4f3fa6b1bda8b22144cafe05072e8b639b3de.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": "coder Only where we directly feed the outputs of the visual encoder " + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": " into the LLM, and keep the LLM frozen. 2) Encoder + FT where we directly feed the outputs of the visual encoder " + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "inline_equation", + "content": "E_{v}" + }, + { + "bbox": [ + 46, + 461, + 287, + 557 + ], + "type": "text", + "content": " into the LLM, and follow LLaVA [39] to fine-tune them to translate sign videos via LoRA [32]. We report the results in Tab. 4, where we significantly outperform the baselines. This shows that SignLLM is effective in harnessing off-the-shelf LLMs for the SLT task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Impact of Main Components of SignLLM. We also verify the impact of the key components of SignLLM by comparing against the following baselines: 1) Ours (w/o VQ-Sign) where we directly quantize the visual encoder " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_v" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "'s output feature using " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "-means algorithm. The visual encoder " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "E_v" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " is pre-trained via the similarity loss " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{sim}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " instead of our proposed VQ-Sign's loss " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{VQ}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". 2) Ours (w/o Codebook Reconstruction) where we feed the character-level sign tokens " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " from VQ-Sign directly into the LLM, while applying the sign-text alignment loss " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{MMD}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " to the character-level sign tokens. 3) Ours (w/o Sign-text Alignment) where we do not apply the sign-text alignment loss " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{MMD}" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": ". As shown in Tab. 5, removing any of our main designs leads to a signifi" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 316, + 544, + 385 + ], + "blocks": [ + { + "bbox": [ + 308, + 316, + 544, + 385 + ], + "lines": [ + { + "bbox": [ + 308, + 316, + 544, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 544, + 385 + ], + "type": "image", + "image_path": "b35ec2637db3253121a7691d3dbf8cb9234e7d6e9815834dd4123450f415b776.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 386, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 306, + 386, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 386, + 545, + 407 + ], + "type": "text", + "content": "Figure 2. Visualization of translation results. Correct translations are in blue while the wrong translations are in red." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 418, + 494, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 494, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 494, + 430 + ], + "type": "text", + "content": "cant performance drop, showing their efficacy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 552 + ], + "type": "text", + "content": "Qualitative Results. We present two sample translations generated by our SignLLM and the current state-of-the-art (GFSLT-VLP [72]) in Fig. 2 for qualitative analysis. In the first sample (top), our model produces a highly accurate translation, whereas [72] inaccurately represents the semantic information. In the second sample (bottom), our model successfully preserves sentence semantics, while [72] introduces a translation error, resulting in redundant and erroneous information. These examples qualitatively show our SignLLM's efficacy in producing accurate translations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 568, + 379, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 568, + 379, + 580 + ], + "spans": [ + { + "bbox": [ + 306, + 568, + 379, + 580 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 590, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 673 + ], + "type": "text", + "content": "We present SignLLM, a novel framework to harness off-the-shelf and frozen LLMs for SLT. SignLLM imparts language-like characteristics to sign video representations through the VQ-Sign and CRA modules, and a sign-text alignment loss improves semantic compatibility. We empirically observe that applying our SignLLM leads to state-of-the-art gloss-free results on two popular SLT benchmarks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 674, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project is supported by the Ministry of Education, Singapore, under the AcRF Tier 2 Projects (MOE-T2EP20222-0009 and MOE-T2EP20123-0014), National Research Foundation Singapore under its AI Singapore Programme (AISG-100E-2023-121)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18369" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Samuel Albanie, Gül Varol, Liliane Momeni, Hannah Bull, Triantafyllos Afouras, Himel Chowdhury, Neil Fox, Bécie Woll, Rob Cooper, Andrew McParland, et al. BBCoxford British sign language dataset. arXiv preprint arXiv:2111.03635, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Ben Allison, David Guthrie, and Louise Guthrie. Another look at the data sparsity problem. In Text, Speech and Dialogue: 9th International Conference, TSD 2006, Brno, Czech Republic, September 11-15, 2006. Proceedings 9, pages 327-334. Springer, 2006. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 236 + ], + "type": "text", + "content": "[3] Ulrich Ammon. World languages: Trends and futures. The handbook of language and globalization, pages 101-122, 2010. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "text", + "content": "[4] Bishnu S Atal and Manfred R Schroeder. Adaptive predictive coding of speech signals. Bell System Technical Journal, 49 (8):1973-1986, 1970. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "text", + "content": "[5] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 288, + 373 + ], + "type": "text", + "content": "[6] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. Data2vec: A general framework for self-supervised learning in speech, vision and language. In International Conference on Machine Learning, pages 1298-1312. PMLR, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 430 + ], + "type": "text", + "content": "[7] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 433, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 433, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 433, + 288, + 475 + ], + "type": "text", + "content": "[8] Kaj Bostrom and Greg Durrett. Byte pair encoding is suboptimal for language model pretraining. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4617-4624, 2020. 2, 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 478, + 288, + 533 + ], + "type": "text", + "content": "[9] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 588 + ], + "type": "text", + "content": "[10] Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7784-7793, 2018. 1, 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[11] Necati Cihan Camgoz, Oscar Koller, Simon Hadfield, and Richard Bowden. Sign language transformers: Joint end-to-end sign language recognition and translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10023-10033, 2020. 1, 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[12] Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan Yu. Recall and learn: Fine-tuning deep pretrained language models with less forgetting. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7870-7881, 2020. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[13] Yutong Chen, Fangyun Wei, Xiao Sun, Zhirong Wu, and Stephen Lin. A simple multi-modality transfer learning baseline for sign language translation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 5110-5120. IEEE, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "text", + "content": "[14] Yutong Chen, Ronglai Zuo, Fangyun Wei, Yu Wu, Shujie Liu, and Brian Mak. Two-stream network for sign language recognition and translation. In NeurIPS, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[15] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[16] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 354 + ], + "type": "text", + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NAACL-HLT, pages 4171–4186, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 355, + 545, + 408 + ], + "type": "text", + "content": "[19] Shuoyang Ding, Adithya Renduchintala, and Kevin Duh. A call for prudent choice of subword merge operations in neural machine translation. In Proceedings of Machine Translation Summit XVII: Research Track, pages 204-213, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 476 + ], + "type": "text", + "content": "[20] Amanda Duarte, Shruti Palaskar, Lucas Ventura, Deepti Ghadiyaram, Kenneth DeHaan, Florian Metze, Jordi Torres, and Xavier Giro-i Nieto. How2sign: a large-scale multimodal dataset for continuous american sign language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2735-2744, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[21] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey. arXiv preprint arXiv:2308.14177, 2, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "type": "text", + "content": "[22] Biao Fu, Peigen Ye, Liang Zhang, Pei Yu, Cong Hu, Xiaodong Shi, and Yidong Chen. A token-level contrastive framework for sign language translation. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 1, 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 600 + ], + "type": "text", + "content": "[23] Philip Gage. A new algorithm for data compression. C Users Journal, 12(2):23-38, 1994. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "text", + "content": "[24] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14953-14962, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[25] Ximena Gutierrez-Vasques, Christian Bentz, Olga Sozinova, and Tanja Samardzic. From characters to words: the turning point of bpe merges. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 3454-3468, 2021. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18370" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, Jindrich Helcl, and Alexandra Birch. Survey of low-resource machine translation. Computational Linguistics, 48 (3):673-732, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 138 + ], + "type": "text", + "content": "[27] Awni Hannun. Sequence modeling with ctc. Distill, 2(11): e8, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 185 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 273 + ], + "type": "text", + "content": "[29] Ruidan He, Linlin Liu, Hai Ye, Qingyu Tan, Bosheng Ding, Liying Cheng, Jiawei Low, Lidong Bing, and Luo Si. On the effectiveness of adapter-based tuning for pretrained language model adaptation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2208-2222, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 275, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 286, + 330 + ], + "type": "text", + "content": "[30] Valentin Hofmann, Janet Pierrehumbert, and Hinrich Schütze. Dagobert: Generating derivational morphology with a pretrained language model. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3848-3861, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 331, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 386 + ], + "type": "text", + "content": "[31] Jiahui Hou, Xiang-Yang Li, Peide Zhu, Zefan Wang, Yu Wang, Jianwei Qian, and Panlong Yang. Signspeaker: A real-time, high-precision smartwatch-based sign language translator. In The 25th Annual International Conference on Mobile Computing and Networking, pages 1–15, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 431 + ], + "type": "text", + "content": "[32] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 434, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 286, + 498 + ], + "type": "text", + "content": "[33] Tao Jin, Zhou Zhao, Meng Zhang, and Xingshan Zeng. Prior knowledge and memory enriched transformer for sign language translation. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 3766-3775. Association for Computational Linguistics, 2022. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 500, + 286, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 286, + 564 + ], + "type": "text", + "content": "[34] Jichao Kan, Kun Hu, Markus Hagenbuchner, Ah Chung Tsoi, Mohammed Bennamoun, and Zhiyong Wang. Sign language translation with hierarchical spatio-temporal graph neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3367-3376, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 567, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 286, + 599 + ], + "type": "text", + "content": "[35] Sang-Ki Ko, Chang Jo Kim, Hyedong Jung, and Choongsang Cho. Neural sign language translation based on human key-point estimation. Applied sciences, 9(13):2683, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 601, + 286, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 286, + 622 + ], + "type": "text", + "content": "[36] W Levelt. Producing spoken language. The neurocognition of language, pages 83-122, 1999. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 286, + 677 + ], + "type": "text", + "content": "[37] Dongxu Li, Chenchen Xu, Xin Yu, Kaihao Zhang, Benjamin Swift, Hanna Suominen, and Hongdong Li. Tspnet: Hierarchical feature learning via temporal semantic pyramid for sign language translation. Advances in Neural Information Processing Systems, 33:12034-12045, 2020. 1, 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 286, + 712 + ], + "type": "text", + "content": "[38] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[40] Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics, 8:726-742, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 218 + ], + "type": "text", + "content": "[41] Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412-1421, 2015. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 221, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 253 + ], + "type": "text", + "content": "[42] Nathaniel FG Martin and James W England. Mathematical theory of entropy. Number 12. Cambridge university press, 2011. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 311 + ], + "type": "text", + "content": "[43] Igor Melnyk, Vijil Chenthamarakshan, Pin-Yu Chen, Payel Das, Amit Dhurandhar, Inkit Padhi, and Devleena Das. Reprogramming pretrained language models for antibody sequence infilling. In Proceedings of the 2023 International Conference on Machine Learning (ICML), 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 545, + 346 + ], + "type": "text", + "content": "[44] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781, 2013. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 348, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 545, + 403 + ], + "type": "text", + "content": "[45] Arijit Nag, Bidisha Samanta, Animesh Mukherjee, Niloy Ganguly, and Soumen Chakrabarti. Entropy-guided vocabulary augmentation of multilingual language models for low-resource tasks. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8619-8629, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 437 + ], + "type": "text", + "content": "[46] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 440, + 545, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 493 + ], + "type": "text", + "content": "[47] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 497, + 545, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 539 + ], + "type": "text", + "content": "[48] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11(5-6):355-607, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 575 + ], + "type": "text", + "content": "[49] Haoxuan Qu, Yujun Cai, and Jun Liu. Llms are good action recognizers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 632 + ], + "type": "text", + "content": "[50] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 687 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 687 + ], + "type": "text", + "content": "[51] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[52] Stephanie Seneff. The use of linguistic hierarchies in speech understanding. In in Proc. ICSLP, 1998. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18371" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Rico Sennrich and Biao Zhang. Revisiting low-resource neural machine translation: A case study. In 57th Annual Meeting of the Association for Computational Linguistics, pages 211–221. Association for Computational Linguistics (ACL), 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 185 + ], + "type": "text", + "content": "[54] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998-3009, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 229 + ], + "type": "text", + "content": "[55] Xinying Song, Alex Salcianu, Yang Song, Dave Dopson, and Denny Zhou. Fast wordpiece tokenization. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2089-2103, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[56] Shengeng Tang, Dan Guo, Richang Hong, and Meng Wang. Graph-based multimodal sequential embedding for sign language translation. IEEE Transactions on Multimedia, 24: 4433-4445, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 319 + ], + "type": "text", + "content": "[57] Ilya O Tolstikhin, Bharath K Sriperumbudur, and Bernhard Scholkopf. Minimax estimation of maximum mean discrepancy with radial kernels. Advances in Neural Information Processing Systems, 29:1930-1938, 2016. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "text", + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 410 + ], + "type": "text", + "content": "[59] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[60] William Vicars. American sign language: “slow”. https://www.lifeprint.com/asl101/pages-signs/s/slow.htm, 2017.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 287, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 467 + ], + "type": "text", + "content": "[61] Cédric Villani et al. Optimal transport: old and new. Springer, 2009. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 534 + ], + "type": "text", + "content": "[62] Andreas Voskou, Konstantinos P Panousis, Dimitrios Kosmopoulos, Dimitris N Metaxas, and Sotirios Chatzis. Stochastic transformer networks with linear competing units: Application to end-to-end sl translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11946-11955, 2021. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 590 + ], + "type": "text", + "content": "[63] Ivan Vulic, Edoardo Maria Ponti, Robert Litschko, Goran Glavaš, and Anna Korhonen. Probing pretrained language models for lexical semantics. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 7222-7240, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 656 + ], + "type": "text", + "content": "[64] Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. Deep high-resolution representation learning for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 43(10):3349-3364, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[65] Jingjing Xu, Hao Zhou, Chun Gan, Zaixiang Zheng, and Lei Li. Vocabulary learning via optimal transport for neural machine translation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 643 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "guage Processing (Volume 1: Long Papers), pages 7361-7373, 2021. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "text", + "content": "[66] Wen Yang, Chong Li, Jiajun Zhang, and Chengqing Zong. Bigtrans: Augmenting large language models with multilingual translation capability over 100 languages. arXiv preprint arXiv:2305.18098, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[67] Aoxiong Yin, Tianyun Zhong, Li Tang, Weike Jin, Tao Jin, and Zhou Zhao. Gloss attention for gloss-free sign language translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2551-2562, 2023. 2, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[68] Kayo Yin and Jesse Read. Better sign language translation with stmc-transformer. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5975-5989, 2020. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[69] Biao Zhang, Mathias Müller, and Rico Sennrich. SLTUNET: A simple unified model for sign language translation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[70] Jian Zhao, Weizhen Qi, Wengang Zhou, Nan Duan, Ming Zhou, and Houqiang Li. Conditional sentence generation and cross-modal reranking for sign language translation. IEEE Transactions on Multimedia, 24:2662-2672, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[71] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 387, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 431 + ], + "type": "text", + "content": "[72] Benjia Zhou, Zhigang Chen, Albert Clapés, Jun Wan, Yanyan Liang, Sergio Escalera, Zhen Lei, and Du Zhang. Gloss-free sign language translation: Improving from visual-language pretraining, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 497 + ], + "type": "text", + "content": "[73] Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. Improving sign language translation with monolingual data by sign back-translation. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 1316-1325. Computer Vision Foundation / IEEE, 2021. 1, 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 542 + ], + "type": "text", + "content": "[74] Hao Zhou, Wengang Zhou, Yun Zhou, and Houqiang Li. Spatial-temporal multi-cue network for sign language recognition and translation. IEEE Transactions on Multimedia, 24:768-779, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 543, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 587 + ], + "type": "text", + "content": "[75] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng Chua. Retrieving and reading: A comprehensive survey on open-domain question answering. arXiv preprint arXiv:2101.00774, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 587, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 643 + ], + "type": "text", + "content": "[76] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Lingpeng Kong, Jiajun Chen, Lei Li, and Shujian Huang. Multilingual machine translation with large language models: Empirical results and analysis. arXiv preprint arXiv:2304.04675, 2023.1, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18372" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_content_list.json b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29a747554680e1a6283401ea8d6a9094472f1a46 --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_content_list.json @@ -0,0 +1,1431 @@ +[ + { + "type": "text", + "text": "LLaFS: When Large Language Models Meet Few-Shot Segmentation", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lanyun Zhu $^{1}$ Tianrun Chen $^{2}$ Deyi Ji $^{3}$ Jieping Ye $^{3}$ Jun Liu $^{1*}$ \nSingapore University of Technology and Design $^{1}$ Zhejiang University $^{2}$ Alibaba Group $^{3}$", + "bbox": [ + 112, + 179, + 854, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes LLaFS, the first attempt to leverage large language models (LLMs) in few-shot segmentation. In contrast to the conventional few-shot segmentation methods that only rely on the limited and biased information from the annotated support images, LLaFS leverages the vast prior knowledge gained by LLM as an effective supplement and directly uses the LLM to segment images in a few-shot manner. To enable the text-based LLM to handle image-related tasks, we carefully design an input instruction that allows the LLM to produce segmentation results represented as polygons, and propose a region-attribute table to simulate the human visual mechanism and provide multi-modal guidance. We also synthesize pseudo samples and use curriculum learning for pretraining to augment data and achieve better optimization. LLaFS achieves state-of-the-art results on multiple datasets, showing the potential of using LLMs for few-shot computer vision tasks.", + "bbox": [ + 75, + 276, + 472, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 550, + 209, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image segmentation is a fundamental task in computer vision with extensive applications. The development of deep learning algorithms [7, 9, 12, 72, 73] trained on large-scale datasets has brought significant advancements to this field [3-5, 61, 69, 71]. However, annotating pixel-level segmentation ground truth on a large scale is extremely resource-intensive. Therefore, a more source-efficient learning strategy, few-shot segmentation, has received much attention from academia and holds immense practical value.", + "bbox": [ + 75, + 569, + 468, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In few-shot segmentation, the model should develop category-specific segmentation capabilities based on only a small amount of annotated data, called support images. To achieve this, existing few-shot segmentation methods [24, 27, 38, 42, 43, 49, 66, 67] typically adopt a support-feature-guided framework. In this framework, relevant features of the target category are extracted from annotated support images and used as guiding information to segment query images. To achieve higher performance, researchers have proposed many methods to explore better ways for support feature extraction [24, 38, 43] and query", + "bbox": [ + 75, + 705, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "segmentation assistance [17, 56, 62]. Although these efforts have demonstrated some success, their segmentation performance is still far from satisfactory. This is because the very limited number of support images contain only a small, incomplete, and biased set of information, so the framework that relies solely on these support-based features for query segmentation inherently suffers from information constraints and cannot achieve a sufficiently high level of accuracy. Therefore, we believe that the further advancement of few-shot segmentation urgently requires an entirely new framework, which should be capable of utilizing richer and more comprehensive information, thereby breaking through the existing framework's bottlenecks to reach better results.", + "bbox": [ + 496, + 252, + 893, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We discover that recent advances in large language models (LLMs) [2, 50, 70] can offer potential opportunities to achieve this goal. Specifically, LLMs pre-trained on large-scale corpora have accumulated a vast amount of prior knowledge, which can effectively supplement the insufficient information in support images, thereby resulting in the more effective guidance. Moreover, LLMs have shown to be effective few-shot learners in the field of NLP [2]. This naturally inspires us to further extend their capabilities to few-shot tasks in other modalities. Based on these insights, we hereby innovatively employ LLMs to tackle few-shot segmentation and introduce an entirely new framework named LLaFS. Unlike some previous segmentation methods that also use language models (LMs) but only for auxiliary purposes, such as utilizing LMs to extract intermediate features [13, 28, 64] or to generate attribute prompts [37], our LLaFS directly employs LLMs to produce segmentation results. This makes LMs no longer work as only auxiliary tools, but fully unlock their complete potential in handling the complex computer vision tasks in an end-to-end manner. In this way, we provide an important exploration towards a unified framework that allows LLMs to tackle few-shot tasks in other modalities beyond NLP.", + "bbox": [ + 496, + 470, + 893, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We find that integrating LLM to few-shot segmentation is non-trivial as we face three critical technical challenges: 1) How to enable the text-based LLM to comprehend and address an image processing task? 2) How to leverage both the visual information from support images and the", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3065", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text information from the LLM to guide the query segmentation? and 3) How to effectively train the model with only limited data? To address the first challenge, we draw inspiration from instruction tuning [44] and introduce a task instruction, which is used to explicitly define the few-shot segmentation task within the input of the LLM. To tackle the second challenge, we treat support images as in-context demonstration samples and design a region-attribute corresponding table to extract fine-refined multi-modal guidance information. For the third challenge, we further propose a pseudo-sample synthesis method to augment the pretraining samples and introduce a curriculum learning mechanism to achieve better optimization. By incorporating these designs, our LLaFS can handle few-shot segmentation effectively. We conduct experiments on multiple datasets and achieve state-of-the-art (SOTA) results that significantly outperform existing methods.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of this work are as follows: 1) We introduce LLaFS, the first framework to address few-shot segmentation using large language models. 2) We propose various innovative designs to make better use of LLMs in few-shot segmentation, including a task-tailored instruction, a fine-grained in-context instruction serving as multimodal guidance, and a pseudo-sample-based curriculum pretraining mechanism. 3) Our approach achieves state-of-the-art performance on multiple datasets.", + "bbox": [ + 75, + 347, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 489, + 218, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-Shot Segmentation. To address the issue of conventional semantic segmentation methods [3, 6, 10, 18–20, 55, 61, 65, 74] that require a large number of training samples, the task of few-shot segmentation (FSS), which allows to segment a query image using only a small number of annotated support images, has been proposed and gained significant attention [1, 21, 23, 31, 35, 43, 47, 52, 63]. Current FSS methods typically adopt a prototype-guided approach [8, 17, 23, 24, 42, 63]. They use masked average pooling (MAP) to extract global [8, 34, 42] or local [24] average prototypes from the backbone features of support images, and then employ these prototypes for guiding the segmentation of query images through feature fusion [23, 24, 34], distance computation [14, 38], or attention mechanisms [45]. However, these methods can only leverage a limited amount of information extracted from a very small number of support images, thus potentially leading to suboptimal results and reduced robustness. To overcome this limitation, [64] uses the more comprehensive word embedding as the general class information to assist in segmentation. Despite some improvements, [64] is still constrained by the limited capabilities of small language models and has not delved deeper into how to better integrate textual information and support image data to achieve more effective guidance. In this paper, we are the first to employ large language models (LLMs) to achieve FSS by using our", + "bbox": [ + 75, + 508, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "carefully designed instructions that contain a more effective multimodal guidance. Furthermore, we utilize the LLM to directly produce segmentation results, rather than merely using intermediate features as done in [64]. This offers a brand-new paradigm to FSS.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models. The advent of large language models (LLMs) such as GPT [2] and Llama [50] has marked the beginning of a new era in artificial intelligence. Thanks to their significantly increased model parameters and training data, these LLMs contain rich prior knowledge and can be efficiently finetuned for specific tasks or application requirements through methods such as prompts [32, 58], adapters [15, 22] and LoRA [16]. Recently, researchers have started exploring visual large language models [25, 29, 48, 53] to establish a unified framework for multimodal data processing, aiming to override the restriction of LLMs being solely applicable to language data. However, none of them are designed for few-shot tasks in computer vision. In this paper, we introduce the first visual LLM framework for handling few-shot segmentation. To achieve this, we draw inspiration from instruction tuning [44] and in-context learning [11, 39], and carefully design a suitable form of instruction and demonstration examples tailored for few-shot segmentation. By doing so, our method can enable the LLM to comprehend image data and perform few-shot segmentation effectively.", + "bbox": [ + 496, + 167, + 893, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 494, + 591, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 512, + 609, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to construct an LLM-based framework for few-shot segmentation, i.e., to segment a query image $I_{q}$ based on $K$ support images $\\{I_s^k\\}_{k=1}^K$ and their ground truth maps $\\{G_s^k\\}_{k=1}^K$ . As shown in Fig.1, the overall framework of LLaFS can be divided into three key components: (1) a feature extractor that extracts image features and generates visual tokens; (2) a task-tailored instruction that combines visual tokens, target categories, and task requirements to provide task-related information and support guidance; and (3) an LLM that predicts segmentation masks based on the input instruction, followed by a refinement network to optimize the results. For the feature extractor, we adopt the approach in Blip2 [25] by using an image encoder followed by a Q-former and a fully-connected layer to generate a set of visual tokens. We use ResNet50 as the image encoder and keep it frozen during training. For the instruction, we carefully design it as the combination of two parts: segmentation task instruction (Sec.3.2.1) and fine-grained in-context instruction (Sec.3.2.2) to provide comprehensive and detailed guidance. For the LLM, we employ CodeLlama [46] with 7 billion parameters that have been finetuned through instruction tuning. Note that compared to", + "bbox": [ + 496, + 529, + 893, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "For simplify of illustration, we introduce LLaFS under the one-shot setting. Supp presents how to extend LLaFS to the multi-shot setting.", + "bbox": [ + 500, + 875, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3066", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/47b2fbf3693f0e2b944c5c88d2b7cc7561dfaaaf6316cf6f2be3da4b3c20c274.jpg", + "image_caption": [ + "Figure 1. Overview of LLaFS. The image encoder and Q-former extract image features and generate a set of visual tokens. Subsequently, a segmentation task instruction and fine-grained in-context introduction are introduced to provide detailed and comprehensive information. These two instructions are integrated and fed into the LLM to produce the vertices coordinates of polygons that enclose the target object. The segmentation mask represented by this polygon is processed by a refinement network to get the final result." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 890, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vanilla Llama, we empirically find that CodeLlama fine-tuned with code generation datasets exhibits higher accuracy and stability in generating structured information like the segmentation result in our task. We equip CodeLlama with LoRA for fine-tuning. All these components work together within the LLaFS framework to achieve high-quality few-shot segmentation.", + "bbox": [ + 75, + 345, + 470, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Instruction", + "text_level": 1, + "bbox": [ + 76, + 458, + 199, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the input of LLM, the instruction is the most crucial component in our framework that makes LLM possible to handle few-shot segmentation. To provide comprehensive information, we design two instructions, namely segmentation task instruction and fine-grained in-context instruction, to respectively provide the LLM with detailed task definitions and fine-grained multi-modal guidance. These two instructions are integrated to formulate the complete instruction as shown in Fig.1. In the following Sec.3.2.1 and Sec.3.2.2, we introduce these two instructions in detail.", + "bbox": [ + 75, + 476, + 468, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Segmentation Task Instruction", + "text_level": 1, + "bbox": [ + 76, + 638, + 344, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LLMs trained on massive text contents have gained strong reasoning capabilities and a vast amount of world knowledge. Language instructions have shown to be a powerful tool for leveraging these knowledge and capability to handle complex tasks [44]. To achieve better results, the instructions need to be sufficiently clear and detailed, whereas those using only simple terminologies such as 'segment an image' are evidently too abstract for LLMs to comprehend. Thus, we design a structured instruction to explicitly provide more task details such as the expected input and output formats of few-shot segmentation. Specifically, in our instruction, we follow [53] by representing the pixel-wise segmentation output as a 16-point polygon that encloses the target object [30]. Note that it is hard for LLMs to directly generate pixel-wise image masks due to LLM's limited number of output tokens. Our alternative solution of generating", + "bbox": [ + 75, + 659, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "polygon vertices provide a token-efficient method for using LLMs to achieve pixel-level segmentation.", + "bbox": [ + 496, + 345, + 890, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, training solely on text contents makes LLMs difficult to comprehend visual information precisely, especially in our few-shot image segmentation task, where the number of available training images is very scarce. For this, inspired by the success of in-context learning in NLP [11, 39], we further propose to encode the support image and its ground truth as a visual demonstration example, using it as an intuitive reference in the instruction to teach LLM how to segment a particular class within an image.", + "bbox": [ + 496, + 376, + 892, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By incorporating these designs, we write our segmentation task instruction as: For each object within the class [class] in an image, output coordinates of a 16-point polygon that encloses the object. These points should be arranged in a clockwise direction. The output should be a tuple in the format of (c1, c2, ..., cn), where cn is the coordinates for the n-th object and its format should be ((x1,y1),(x2,y2),..., (x16,y16)). The coordinate value should be within [image size]. For example, for image [support image], the output should be [support ground truth]. Here, [support image] is the visual tokens from the support image.", + "bbox": [ + 496, + 512, + 893, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2 Fine-grained In-context Instruction", + "text_level": 1, + "bbox": [ + 498, + 686, + 802, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. The above task instruction makes segmenting a class possible by leveraging LLM's knowledge of the class. In the instruction, the class to be segmented is indicated by the [class] token, which is typically a single noun. However, considering that LLMs have never been trained on images, it is challenging for them to directly align this abstract noun with an image region that may possess a complex internal structure. To address this issue, we drew inspiration from human brains and found that when classifying an unseen new class, the human cognitive system follows a mechanism of 'from general to detailed, from abstract to concrete' [41, 59]. Specifically, given an unseen class represented by a general noun, the human brain first decomposes", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3067", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b79500931e4182d7d54ef6fbd3c91e2c0663e0a6239da670778f8bceff57b9d2.jpg", + "image_caption": [ + "Figure 2. Illustration of how to construct the region-attribute corresponding table used in the fine-grained in-context instruction." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 449, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "it into detailed attributes based on the acquired knowledge. For example, in the case of an unseen class 'owl', a person can first gather information from references to learn about the owl's attributes such as 'large round eyes' and 'hooked beak'. Subsequently, it can search the image for concrete regions that match these abstract attributes to determine the presence of the class.", + "bbox": [ + 75, + 279, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by this, we propose a fine-grained in-context instruction to simulate such a human cognitive mechanism based on the support images. For this, we first use ChatGPT to extract detailed attributes of the target class, then we search for regions in support images that correspond to these attributes and generate a corresponding table accordingly. The obtained attributes and table constitute an in-context instruction that is fed into the LLM, which serves as a demonstration example to guide the LLM on how to recognize an image class in a more human-like and fine-grained manner. This alleviates the challenge that LLM cannot perform segmentation well when only inputted with an abstract class noun. Lastly, we introduce an expert-guide framework that refines the instruction to increase its class representation ability. The following sections explain how to generate and refine this instruction in detail.", + "bbox": [ + 75, + 386, + 468, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attributes Generation We first simulate the step of 'from general to detailed' to extract class attributes. Specifically, as shown in Fig.3(a), we construct a prompt 'What does a [class] look like? Please answer in the format of: A [class] has A, B, C,..., where A, B, and C are noun phrases to describe a [class].', and utilize ChatGPT to extract phrases-based attributes that describe the fine-grained details of this class. These attributes are denoted as [attributes] = $\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}$ . For each [att]i, we utilize 'A photo of [att]i' as a prompt to extract an embedding $t_i$ from the CLIP's text encoder. In this way, we get $\\{t_i\\}_{i=1}^{N_a}$ from $\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}$ .", + "bbox": [ + 76, + 650, + 468, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region-attribute Corresponding Table. After that, we simulate the second step of 'from abstract to concrete'. To implement this, as shown in Fig.2, we propose a region-attribute corresponding table to find the alignment between support image regions and class attributes. For this, we first", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/100c4de25f54d47908bea7ae65a5f2951da3fcf61868c74417a9dc89d6c6bf78.jpg", + "image_caption": [ + "Figure 3. Examples of using ChatGPT for (a) class attributes generation, (b) ambiguity detection and (c) discriminative attributes generation." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 870, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "divide the support foreground into multiple local regions. Specifically, for each object in the support image within the target class, we employ the method in selective search [51] to generate a set of superpixels $\\{s_i\\}_{i=1}^{N_r}$ with different scales in an unsupervised manner. Each $s_i$ aggregates pixels that are close in position and similar in features, so it can represent a local region with a specific semantic meaning. Based on each $s_i$ , a masked image is generated and passed through CLIP's² image encoder to produce a feature $f_i$ . We calculate the cosine similarity between $f_i$ and the embedding $t_j$ for each attribute $[\\mathrm{att}]_j$ , and utilize a thresholding process to establish region-attribute correspondence. Formally,", + "bbox": [ + 496, + 358, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\operatorname {c o r} \\right] _ {i} = \\left[ \\left[ \\operatorname {a t t} \\right] _ {j} \\text {f o r} j \\in \\left[ 1, N _ {a} \\right] \\text {i f} \\cos \\left(f _ {i}, t _ {j}\\right) > \\alpha \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 551, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos$ refers to cosine similarity, $\\alpha$ is a pre-defined threshold. The obtained $[\\mathrm{cor}]_i$ contains attributes that align with $s_i$ . In this way, we get $\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}$ from $\\{s_i\\}_{i=1}^{N_r}$ , which serves as an attribute-region corresponding table that can provide the fine-grained multi-modal reference.", + "bbox": [ + 498, + 585, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction Construction. We integrate the obtained class attributes $\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}$ and corresponding table $\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}$ , and write the fine-grained in-context instruction as: The [class] has [attributes]. For example, in [support image], $[\\mathrm{coord}]_{s_1}$ is $[\\mathrm{cor}]_1$ , $[\\mathrm{coord}]_{s_2}$ is $[\\mathrm{cor}]_2\\ldots$ , $[\\mathrm{coord}]_{s_{N_r}}$ is $[\\mathrm{cor}]_{N_r}$ . Here, $[\\mathrm{coord}]_{s_i}$ is the coordinates of $s_i$ in the format of a 16-point polygon. By using this instruction, we provide the LLM with a reference about the attributes of the target class and their corresponding regions in the support image. In this way, a demonstration example can be created that simulates how the human cognitive mechanism recognizes the support image as a class. Using this example as a reference, the LLM can be taught how to understand", + "bbox": [ + 496, + 669, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We use CLIP fintuned through the method in [26] to ensure high-quality region-text alignment.", + "bbox": [ + 500, + 875, + 888, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3068", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and segment an image class in a fine-grained manner.", + "bbox": [ + 76, + 90, + 431, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above-mentioned instruction constructed by the obtained attributes $\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}$ and table $\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}$ can be directly input into LLM for guidance. However, due to variations in camera angles and instances of occlusion, not every attribute can be directly matched to a region within the support image. Thus, the obtained table $\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}$ may contain only a subset of attributes within $\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}$ . Unfortunately, we find the combinations of these partial attributes may be insufficient for the unambiguous recognition of the target class. For example, the combination 'wheels, windows, doors' can refer to 'train', 'car', and 'bus' interchangeably. Due to the ambiguous table, the instruction may be misleading. To alleviate this issue, we propose an expert-guide framework to refine the instruction. In this framework, we first employ ChatGPT to identify ambiguous classes of the existing table, then we extract additional attributes that can distinguish the target class from these ambiguous classes for refinement. As long as these additional attributes can be aligned with local regions in the support image, the refined table based on them will become unambiguous. In this way, the class representation ability and comprehensiveness of the instruction can be improved.", + "bbox": [ + 76, + 114, + 468, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, this framework generates a refined instruction through the following three steps: 1) Ambiguity Detection. As shown in Fig.3(b), We employ ChatGPT to identify potential ambiguous classes in the obtained table $\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}$ . Specifically, we denote the attributes contained in $\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}$ as [partial-attributes] and ask ChatGPT 'Except for [class], which classes also have [partial-attributes]?' In this way, we obtain a set of ambiguous classes denoted as [a-classes] = $\\{[\\mathrm{a - class}]_i\\}_{i=1}^{N_{ac}}$ from ChatGPT's feedback. 2) Discriminative Attributes Generation. As shown in Fig.3(c), To avoid being misled by these ambiguous classes, we use 'What does [class] look different from [a-classes]?' as a text prompt, enabling ChatGPT to generate attributes that are more discriminative from these classes. The obtained attributes $\\{[\\mathrm{d - att}]_i\\}_{i=1}^{N_d}$ are added to [attributes] for updating. 3) Table and Instruction Refinement. Finally, we use these updated attributes to reperform Eq.1 to obtain a refined table. The updated attributes and table are reassembled to form a refined fine-grained in-context instruction.", + "bbox": [ + 76, + 465, + 468, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We found that a single execution of the three steps already resolves ambiguities in over $85\\%$ of the instructions. For the remaining $15\\%$ of instructions, we observe that because the newly-acquired discriminative attributes still couldn't find matching regions in the support image, the resulting table after refinement remains to be ambiguous.", + "bbox": [ + 75, + 768, + 468, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, we iteratively apply the last two steps of this framework until the ambiguity is eradicated. To achieve this, from the second iteration onwards, we replace the text prompt in the discriminative attributes generation step with 'Apart from [all-discriminative-attributes], tell me more differences in appearance between [class] and [a-classes]\", where [all-discriminative-attributes] refers to the discriminative attributes $[\\mathrm{d - att}]_i$ obtained from all previous iterations. By doing so, we enable the iterative framework to continuously discover more discriminative attributes and verify whether they have matched regions in the support image. Eventually, when the framework successfully discovers discriminative attributes $[\\mathrm{d - att}]_i$ that can be aligned with the support image or reaches the maximum number of iterations, we terminate the iteration. For efficiency, we set the maximum number of iterations to 3, in which $96\\%$ of the ambiguities have been entirely eradicated.", + "bbox": [ + 500, + 90, + 890, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Segmentation Prediction", + "text_level": 1, + "bbox": [ + 500, + 349, + 723, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We integrate segmentation task instruction and fine-grained in-context instruction to formulate the complete instruction as shown in Fig.1. With this instruction as input, the LLM can predict the coordinates of a 16-point polygon that surrounds the target object. Finally, to rectify the imprecision caused by the polygon representation of object edges, a refinement network comprising a pixel decoder and a mask transformer is introduced to generate a refined segmentation mask by using the polygon as the initial mask. Please see Supp for the detailed structures of this network.", + "bbox": [ + 498, + 366, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Curriculum Pretraining with Pseudo Samples", + "text_level": 1, + "bbox": [ + 500, + 521, + 885, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motivation. After carefully designing the model structure and instruction format, the next challenge is how to train LLaFS effectively to achieve high segmentation performance. Previous work [29] has shown that the success of instruction tuning often relies on extensive training data. However, due to the difficulty of obtaining pixel-annotated labels, the segmentation datasets used for training typically contain only an insufficient number of images. To address this issue, we propose to generate pseudo support-query pairs and use them to pretrain the LLM. The LLM's ability to handle few-shot segmentation can thus be enhanced by seeing more visual samples.", + "bbox": [ + 498, + 545, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pseudo Sample Generation. Specifically, we propose a method to generate pseudo support-query pairs with the following three steps: 1) Pseudo foreground-background partition. We first useBezier curves to randomly generate a contour within an image region. The area surrounded by this contour is considered as the foreground within the target class, while the regions outside the contour are treated as the background. 2) Noise filling for pseudo support generation. We fill the foreground with Gaussian noise that has a random mean value. For background, we first randomly divide it into multiple subregions to", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Due to space limitations, we omit the description of format control prompts for inputting into ChatGPT. See Supp for details.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3069", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6362bb87ffc0ff2894f271444ec470e5bedf155a7c122faf5e9ce400cb48f4dc.jpg", + "image_caption": [ + "Figure 4. Examples of pseudo samples generated at different pretraining stages. Foreground regions are marked by white contours. As pretraining progresses, pseudo images have reduced intra-image foreground-background differences and greater support-query foreground differences. Meanwhile, the number of polygon vertex coordinates provided in the instruction decreases, while the predicted vertex count increases. These changes gradually increase the pretraining difficulty. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 441, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "simulate the complex backgrounds in real images, then we fill each one with a random Gaussian noise that has a mean value distinct from the foreground's noise. The obtained image is used as the support image. 3) Pseudo query generation. We use a similar approach to generate a query image. Note that in this process, the contour and the mean value of foreground noise are no longer completely randomly determined but adjusted based on those used for generating the support image. This is done to ensure that the foreground regions of both support and query images have similar contour shapes and internal features, so they can reflect the same category. Please refer to Supp for the adjustment details.", + "bbox": [ + 75, + 452, + 468, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Curriculum Pretraining. The synthetic support-query pairs can be directly used for pretraining. However, it is observed that this approach will result in a slow convergence rate. A reason for this problem is that the LLM has not been previously trained on image data, so optimizing it to directly grasp a complex image processing task is challenging. To address this issue, we propose a progressive pretraining approach inspired by the success of curriculum learning [54], in which we initiate the model's pretraining with a simple task and gradually increase the task's difficulty until it ultimately reaches the requirements of segmentation.", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, during pretraining, we incrementally raise the task's difficulty from the following two aspects: 1) Image understanding. During pretraining, by controlling the difference between mean values of different filled noise, we gradually increase the difference in foreground between", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "support and query, while reducing the internal difference between foreground and background within each image. This makes it more challenging for LLM to perform few-shot guidance and partition foreground-background areas as pretraining progresses. 2) Polygon generation. We observe that generating a polygon represented by a combination of vertex coordinates is another challenge for the LLM. Therefore, we adopt a progressive strategy here as well. Instead of training the model to directly predict the coordinates of all 16 points of the polygon, we randomly provide the coordinates of $M$ points in the instruction and let the LLM to predict the coordinates of the remaining $16 - M$ points. During pretraining, we gradually decrease the value of $M$ from 15 to 0. This means that the model receives fewer hints and is required to predict more vertex coordinates as pretraining progresses. Consequently, the pretraining difficulty gradually increases, ultimately reaching the task of predicting all 16 points for segmentation. Experimental results show that this curriculum learning approach allows the model to converge better and achieve higher results. Please see Supp for more technical details on how we increase the difficulty in image understanding and polygon generation.", + "bbox": [ + 496, + 90, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ultimately, the model is trained on the realistic few-shot segmentation dataset after completing the aforementioned pretraining process.", + "bbox": [ + 496, + 422, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 468, + 632, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details and Training Settings", + "text_level": 1, + "bbox": [ + 496, + 484, + 885, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set the number of queries in the Q-former to 100 and the threshold $\\alpha$ in Eq.1 to 0.2. The ground truth of polygon vertices is obtained in polar coordinates [60]. Specifically, starting from the object center, 16 rays are uniformly emitted at equal angular intervals $\\triangle \\theta = 22.5^{\\circ}$ . The points of intersection between these rays and the object contour are taken as the ground truth of the polygon vertices. More implementation details about pseudo sample generation and curriculum pretraining are presented in Supp.", + "bbox": [ + 496, + 501, + 890, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall model is trained in three stages. In the first stage, we freeze the LLM, pretrain the Q-former and fully-connected layers for 100K steps using the datasets4 and methods in Blip2 [25]. In the second stage, we freeze the Q-former, equip the LLM with LoRA, and pretrain the fully-connected layers, LLM and refinement network using the pseudo-sample-based curriculum learning method for 60k steps. In the third stage, we train the fully-connected layers, LLM and refinement network on the realistic few-shot segmentation dataset for 25 epochs. AdamW [36] is used as the optimizer with the cosine annealing schedule and an initial learning rate of 0.0002. The model is trained on 16 A100 GPUs.", + "bbox": [ + 496, + 637, + 890, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 500, + 834, + 799, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table.1 presents the comparisons with other few-shot segmentation methods on two datasets: PASCAL-5 and", + "bbox": [ + 500, + 849, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "${}^{4}$ COCO is excluded from the pretraining set to avoid test data leakage.", + "bbox": [ + 514, + 886, + 888, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3070", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/743435235baa446f9d551d64821660a4ead0a5ed4b3d075bf569ff84369b3297.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodConference1-shot5-shot
Fold-0Fold-1Fold-2Fold-3MeanFold-0Fold-1Fold-2Fold-3Mean
PASCAL-5iNTRENet [33]CVPR'2265.472.359.459.863.266.272.861.762.265.7
BAM[23]CVPR'2269.073.667.561.167.870.675.170.867.270.9
AAFormer[57]ECCV'2269.173.359.159.265.272.574.762.061.367.6
SSP[8]ECCV'2260.567.866.451.061.467.572.375.262.169.3
IPMT[34]NeurIPS'2272.873.759.261.666.873.174.761.663.468.2
ABCNet[56]CVPR'2368.873.462.359.566.071.774.265.467.069.6
HDMNet [45]CVPR'2371.075.468.962.169.471.376.271.368.571.8
MIANet[64]CVPR'2368.575.867.563.268.770.277.470.068.871.7
MSI[40]ICCV'2371.072.563.865.968.573.074.270.566.671.1
SCGAN[62]ICCV'2368.372.566.859.866.872.374.169.165.670.3
LLaFSCVPR'2474.278.872.368.573.575.980.175.870.775.6
COCO-20iNTRENet[33]CVPR'2236.842.639.937.939.338.244.140.438.440.3
BAM[23]CVPR'2243.450.647.543.446.249.354.251.649.651.2
SSP[8]ECCV'2235.539.637.936.747.440.647.045.143.944.1
AAFormer[57]ECCV'2239.844.640.641.441.642.950.145.549.246.9
MM-Former[68]NeurIPS'2240.547.745.243.344.244.052.447.450.048.4
IPMT[34]NeurIPS'2241.445.145.640.043.043.549.748.747.947.5
ABCNet[56]CVPR'2342.346.246.042.044.145.551.752.646.449.1
HDMNet [45]CVPR'2343.855.351.649.450.050.661.655.756.056.0
MIANet[64]CVPR'2342.553.047.847.447.745.858.251.351.951.7
MSI[40]ICCV'2342.449.249.446.146.847.154.954.151.952.0
SCGAN[62]ICCV'2340.449.749.645.646.347.257.259.252.153.9
LLaFSCVPR'2447.558.856.253.053.953.263.863.160.060.0
", + "bbox": [ + 83, + 88, + 607, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f6b3edd0b94ec2a418c9f5bbfb1288de5837e9fcc68994ff7a3f4d7021b687e7.jpg", + "table_caption": [ + "Table 1. Performance comparison with other methods on PASCAL-5 ${}^{i}$ and COCO-20 ${}^{i}$ ." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
LLaFS74.2
LLaFs w/o segmentation task instruction w/ abstract summary67.7
LLaFS w/o fine-grained in-context instruction67.0
LLaFS w/o refinement network69.1
LLaFS w/o pseudo-sample-based curriculum pretraining63.5
", + "bbox": [ + 84, + 393, + 468, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{COCO - 20^i}$ . Following previous papers, we use different folds as the test set, with the remaining folds utilized for training. This approach yields four sets of experimental results along with their mean result. Across all datasets and experimental settings, our method consistently outperforms others and achieves a remarkably significant improvement compared to previous state-of-the-art results. For instance, on PASCAL- $5^{i}$ , LLaFS improves mIoUs by $4.1\\%$ and $3.8\\%$ in the 1-shot and 5-shot settings respectively compared to the second-ranking method. Notably, our approach still exhibits great advantages on the more complex and challenging $\\mathrm{COCO - 20^i}$ dataset, with increases of $3.9\\%$ and $4.0\\%$ in the 1-shot and 5-shot settings respectively. This could be attributed to the rich prior knowledge embedded in the LLM and our carefully designed instructions, which enable the models to handle complex images effectively and robustly. These results demonstrate the outstanding performance of our method and highlight the huge potentiality of using LLMs for tackling few-shot segmentation tasks.", + "bbox": [ + 75, + 517, + 470, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 808, + 230, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the 1-shot setting of the PASCAL- $5^{i}$ dataset, we perform several ablation studies to evaluate different components and designs in our method. More ablation studies are presented in Supp.", + "bbox": [ + 75, + 824, + 468, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Different Components We validate the", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/da9a85ad5396fe959e021a4535dc4af85d89e716db33cf4ef1b21502ebf7a811.jpg", + "table_caption": [ + "Table 4. Effectiveness of different components in our LLaFS." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
LLaFS74.2
LLaFS w/o support images56.9
", + "bbox": [ + 614, + 88, + 890, + 137 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e545b27ea634abb0e3008deb706416d606955d0db6caacf23f7b3a2e156d0738.jpg", + "table_caption": [ + "Table 2. Effectiveness of support images." + ], + "table_footnote": [], + "table_body": "
MethodmIoU
LLaFS74.2
LLaFS w/o class attributes70.7
LLaFS w/o region-attribute corresponding table69.7
LLaFS w/o instruction refinement70.6
LLaFS w/o iterative refinement71.8
", + "bbox": [ + 614, + 159, + 890, + 237 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation results of fine-grained in-context instruction.", + "bbox": [ + 606, + 239, + 893, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51b6b9223321d782271e26111de467c4967e991122fb189dee5ab5fa071e5618.jpg", + "image_caption": [ + "Figure 5. Ablation for the number of polygons' sides." + ], + "image_footnote": [], + "bbox": [ + 607, + 272, + 879, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "effectiveness of key components in our method, including (1) the segmentation task instruction, (2) fine-grained in-context instruction, (3) the refinement network, and (4) the pseudo-sample-based curriculum pretraining. Experimental results are presented in Table.4. Replacing the detailed segmentation task introduction with an abstract summary 'perform image segmentation' decreases mIoU by $6.5\\%$ . Not using the other components individually decreases the mIoU by $7.2\\%$ , $5.1\\%$ , and $10.7\\%$ , respectively. These results demonstrate the importance and effectiveness of each component in our approach.", + "bbox": [ + 496, + 396, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Support Images. In the instruction, the annotated support images provide the LLM with crucial visual guidance. As shown in Table.2, if we do not use the support image as a demonstration example, mIoU decreases by $17.3\\%$ . This highlights the importance of the support image as a few-shot sample in our approach, demonstrating that our LLsFS benefits not solely from LLM's prior knowledge in an open-vocabulary manner but indeed gains further improvement from the provided few-shot samples.", + "bbox": [ + 496, + 565, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Polygon's Sides. In the segmentation task instruction, we represent the segmentation output mask as a region enclosed by a 16-point polygon. We find that the number $M$ of sides in the polygon is also a factor affecting the model's performance. Therefore, we test the relationship between mIoU and $M$ and present the results in Fig.5. We observe that when $M$ is small, the model's performance is suboptimal. This is because polygons with a smaller number of sides cannot accurately describe object edges. As $M$ increases, the mIoU gradually improves. However, when $M$ exceeds 16, we observe a slight decrease in performance. This could be because a larger $M$ increases the task's complexity for LLM to tackle. Based on the results,", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3071", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c2c537ef0f7a6ca47dc5f5cfbef22fc92cfde92007a7586d696929378532644.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmIoU
LLaFS74.2
LLaFS w/o pseudo samples63.5
LLaFS w/o curriculum strategy67.3
LLaFS w/ random pseudo query generation63.9
", + "bbox": [ + 122, + 88, + 431, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation of pseudo-sample-based curriculum pretraining.", + "bbox": [ + 76, + 156, + 467, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e4795b51d371e5f37cb40f7e7248b85fcec20bd966c0815f6fb35e2252174158.jpg", + "image_caption": [ + "(a) Pretraining Loss Curve" + ], + "image_footnote": [], + "bbox": [ + 114, + 179, + 267, + 284 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9107691691716bccf12180c480cc0827b9bae0fe420c73e9aadbff5eb0e3cf9f.jpg", + "image_caption": [ + "(b) Training Loss Curve", + "Figure 6. Pretraining (a) and training (b) loss curves in different settings. Curriculum pretraining results in the best convergence in both pretraining and training stages. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 269, + 179, + 429, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we chose $M = 16$ as our setting.", + "bbox": [ + 76, + 351, + 295, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of Fine-grained In-context Instruction. Table.3 presents the evaluation of different components and designs in our fine-grained in-context instruction, including (1) class attributes, (2) attribute-region corresponding table, (3) expert-guide framework for instruction refinement, and (4) iterative refinement. We observe that when these components are removed, mIoU decreases by $3.5\\%$ , $4.5\\%$ , $3.6\\%$ and $2.4\\%$ , respectively. These results demonstrate the rationality of our designs in this instruction and their effectiveness in improving performance.", + "bbox": [ + 75, + 366, + 467, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of Pseudo-sample-based Curriculum Pretraining. We further evaluate the key techniques in our pseudosample-based curriculum pretraining mechanism and the results are presented in Table.5. (1) When we do not employ pseudo-samples for pretraining, mIoU decreases by $10.7\\%$ . (2) Removing the curriculum training strategy that gradually increases the training task difficulty reduces mIoU by $6.9\\%$ . (3) When generating pseudo support-query samples, to ensure that the support and query can reflect the same category, the contour and the mean value of foreground noise used to generate the query image are adjusted based on those used for generating the support image. When this strategy is not employed and random generation is used instead, mIoU decreases by $10.3\\%$ . These results demonstrate the effectiveness of our method's designs.", + "bbox": [ + 75, + 518, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Loss Curves", + "text_level": 1, + "bbox": [ + 76, + 748, + 207, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig.6, we present the loss curves during the pretraining and training stages. We observe from Fig.6(a) that without the use of curriculum learning, the pretraining task becomes excessively challenging, which causes the model optimization to quickly reach a bottleneck with difficulties in convergence. After using our curriculum learning mechanism, the model achieves significantly better convergence. Furthermore, in Fig.6(b), we compare the loss reduction during the training stage when using curriculum pretrain", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d625049e27dab562014779b6c62fa5532e20347b2eeb4deb42bd1624a8f9d69.jpg", + "image_caption": [ + "Figure 7. Visualization of segmentation results." + ], + "image_footnote": [], + "bbox": [ + 539, + 89, + 848, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing, non-curriculum pretraining, and not using pretraining. The model without pretraining converges the slowest, while the model with curriculum pretraining converges the fastest during the training stage. These results demonstrate the effectiveness of our curriculum-learning-based pretraining approach in enhancing the model's convergence speed.", + "bbox": [ + 496, + 270, + 890, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 498, + 368, + 635, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig.7, we present visual samples of the segmentation results from LLaFS, including the image, ground truth, the LLM output, and the results after refinement. It can be observed that the polygons output by LLM already achieve good segmentation performance, and the results after refinement are more accurate, particularly in the object's edge regions. It is worth noting that, as shown in the third row of Fig.7, our method still achieves high-performance segmentation when there are more than one target object in the image. These results demonstrate the excellent performance of LLaFS.", + "bbox": [ + 496, + 386, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 564, + 617, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes LLaFS, a novel framework that for the first time, leverages large language models (LLMs) to address few-shot segmentation in an end-to-end manner. To enable LLMs to perform this visual task, we introduce a segmentation task instruction to provide detailed task definitions, and propose a fine-grained in-context instruction to simulate human cognitive mechanisms and provide refined multimodal reference information. We also propose a pseudo-sample-based curriculum pretraining mechanism to augment the training samples required for instruction tuning. Our extensive experiments demonstrate the effectiveness of LLaFS, which achieves significantly superior state-of-the-art results across multiple datasets and settings. We consider LLaFS as an important exploration towards an LLM framework capable of addressing few-shot tasks in different modalities beyond natural language processing. Acknowledgement This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-006), MOE AcRF Tier 2 projects MOE-T2EP20222-0009 and MOE-T2EP20123-0014.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3072", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021. 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2", + "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2", + "[4] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3367-3375, October 2023.", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1", + "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[8] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701-719. Springer, 2022. 2, 7", + "[9] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 1", + "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2", + "[11] Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35:30583-30598, 2022. 2, 3", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1", + "[13] Shuting He, Henghui Ding, and Wei Jiang. Primitive gener" + ], + "bbox": [ + 78, + 114, + 470, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ation and semantic-related alignment for universal zero-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11238-11247, 2023. 1", + "[14] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 2", + "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2", + "[16] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 2", + "[17] Kai Huang, Feigege Wang, Ye Xi, and Yutao Gao. Prototypical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19256-19265, 2023. 1, 2", + "[18] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021. 2", + "[19] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022.", + "[20] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 2", + "[21] Siyu Jiao, Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, and Humphrey Shi. Mask matching transformer for few-shot segmentation. In Advances in Neural Information Processing Systems. 2", + "[22] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 2", + "[23] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8057-8067, 2022. 2, 7", + "[24] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021. 1, 2", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. Inter-" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3073", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national Conference on Machine Learning, 2023. 2, 6", + "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 4", + "[27] Binghao Liu, Jianbin Jiao, and Qixiang Ye. Harmonic feature activation for few-shot semantic segmentation. IEEE Transactions on Image Processing, 30:3142-3153, 2021. 1", + "[28] Fang Liu, Yuhao Liu, Yuqiu Kong, Ke Xu, Lihe Zhang, Baocai Yin, Gerhard Hancke, and Rynson Lau. Referring image segmentation using text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22124-22134, October 2023. 1", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5", + "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 3", + "[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 2", + "[32] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, 2022. 2", + "[33] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022. 7", + "[34] Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Intermediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020-38031, 2022. 2, 7", + "[35] Yongfei Liu, Xiangyi Zhang, Songyang Zhang, and Xuming He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision, pages 142-158. Springer, 2020. 2", + "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6", + "[37] Chaofan Ma, Yuhuan Yang, Chen Ju, Fei Zhang, Ya Zhang, and Yanfeng Wang. Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096, 2023. 1", + "[38] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6941-6952, 2021. 1, 2", + "[39] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Re" + ], + "bbox": [ + 78, + 92, + 470, + 891 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, 2022. 2, 3", + "[40] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Msi: Maximize support-set information for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19266-19276, 2023. 7", + "[41] Gregory L Murphy and Douglas L Medin. The role of theories in conceptual coherence. Psychological review, 92(3):289, 1985. 3", + "[42] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 1, 2", + "[43] Atsuro Okazawa. Interclass prototype relation for few-shot segmentation. In European Conference on Computer Vision, pages 362-378. Springer, 2022. 1, 2", + "[44] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2, 3", + "[45] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 2, 7", + "[46] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2", + "[47] Mennatullah Siam, Boris N Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5249-5258, 2019. 2", + "[48] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 2", + "[49] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 1", + "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2", + "[51] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 4", + "[52] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9197-9206, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3074", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2, 3", + "[54] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555-4576, 2021. 6", + "[55] Yan Wang, Jian Cheng, Yixin Chen, Shuai Shao, Lanyun Zhu, Zhenzhou Wu, Tao Liu, and Haogang Zhu. Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation. arXiv preprint arXiv:2304.13672, 2023. 2", + "[56] Yuan Wang, Rui Sun, and Tianzhu Zhang. Rethinking the correlation in few-shot segmentation: A buoys view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7192, 2023. 1, 7", + "[57] Yuan Wang, Rui Sun, Zhe Zhang, and Tianzhu Zhang. Adaptive agent transformer for few-shot segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 36-52. Springer, 2022. 7", + "[58] Zhen Wang, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, Huan Sun, and Yoon Kim. Multitask prompt tuning enables parameter-efficient transfer learning. In International Conference on Learning Representations, 2022. 2", + "[59] Edward J Wisniewski and Bradley C Love. Relations versus properties in conceptual combination. Journal of memory and language, 38(2):177-202, 1998. 3", + "[60] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12193-12202, 2020. 6", + "[61] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1, 2", + "[62] Qianxiong Xu, Wenting Zhao, Guosheng Lin, and Cheng Long. Self-calibrated cross attention network for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7", + "[63] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. Mining latent classes for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8721-8730, 2021. 2", + "[64] Yong Yang, Qiong Chen, Yuan Feng, and Tianlin Huang. Mianet: Aggregating unbiased instance and general information for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7131-7140, 2023. 1, 2, 7", + "[65] Ying Zang, Chenglong Fu, Runlong Cao, Didi Zhu, Min Zhang, Wenjun Hu, Lanyun Zhu, and Tianrun Chen. Resmatch: Referring expression segmentation in a semi-supervised manner. arXiv preprint arXiv:2402.05589, 2024. 2", + "[66] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with" + ], + "bbox": [ + 78, + 90, + 470, + 892 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 1", + "[67] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 1", + "[68] Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, Honghui Shi, et al. Mask matching transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:823-836, 2022. 7", + "[69] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1", + "[70] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1", + "[71] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3082-3092, June 2023. 1", + "[72] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Learning gabor texture features for fine-grained recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1621-1631, 2023. 1", + "[73] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024. 1", + "[74] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 619 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3075", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_model.json b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9af078315827f4843449cbc8761882293706db29 --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "LLaFS: When Large Language Models Meet Few-Shot Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.18, + 0.856, + 0.219 + ], + "angle": 0, + "content": "Lanyun Zhu\\(^{1}\\) Tianrun Chen\\(^{2}\\) Deyi Ji\\(^{3}\\) Jieping Ye\\(^{3}\\) Jun Liu\\(^{1*}\\) \nSingapore University of Technology and Design\\(^{1}\\) Zhejiang University\\(^{2}\\) Alibaba Group\\(^{3}\\)" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.535 + ], + "angle": 0, + "content": "This paper proposes LLaFS, the first attempt to leverage large language models (LLMs) in few-shot segmentation. In contrast to the conventional few-shot segmentation methods that only rely on the limited and biased information from the annotated support images, LLaFS leverages the vast prior knowledge gained by LLM as an effective supplement and directly uses the LLM to segment images in a few-shot manner. To enable the text-based LLM to handle image-related tasks, we carefully design an input instruction that allows the LLM to produce segmentation results represented as polygons, and propose a region-attribute table to simulate the human visual mechanism and provide multi-modal guidance. We also synthesize pseudo samples and use curriculum learning for pretraining to augment data and achieve better optimization. LLaFS achieves state-of-the-art results on multiple datasets, showing the potential of using LLMs for few-shot computer vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.21, + 0.567 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Image segmentation is a fundamental task in computer vision with extensive applications. The development of deep learning algorithms [7, 9, 12, 72, 73] trained on large-scale datasets has brought significant advancements to this field [3-5, 61, 69, 71]. However, annotating pixel-level segmentation ground truth on a large scale is extremely resource-intensive. Therefore, a more source-efficient learning strategy, few-shot segmentation, has received much attention from academia and holds immense practical value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.875 + ], + "angle": 0, + "content": "In few-shot segmentation, the model should develop category-specific segmentation capabilities based on only a small amount of annotated data, called support images. To achieve this, existing few-shot segmentation methods [24, 27, 38, 42, 43, 49, 66, 67] typically adopt a support-feature-guided framework. In this framework, relevant features of the target category are extracted from annotated support images and used as guiding information to segment query images. To achieve higher performance, researchers have proposed many methods to explore better ways for support feature extraction [24, 38, 43] and query" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.895, + 0.465 + ], + "angle": 0, + "content": "segmentation assistance [17, 56, 62]. Although these efforts have demonstrated some success, their segmentation performance is still far from satisfactory. This is because the very limited number of support images contain only a small, incomplete, and biased set of information, so the framework that relies solely on these support-based features for query segmentation inherently suffers from information constraints and cannot achieve a sufficiently high level of accuracy. Therefore, we believe that the further advancement of few-shot segmentation urgently requires an entirely new framework, which should be capable of utilizing richer and more comprehensive information, thereby breaking through the existing framework's bottlenecks to reach better results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.895, + 0.82 + ], + "angle": 0, + "content": "We discover that recent advances in large language models (LLMs) [2, 50, 70] can offer potential opportunities to achieve this goal. Specifically, LLMs pre-trained on large-scale corpora have accumulated a vast amount of prior knowledge, which can effectively supplement the insufficient information in support images, thereby resulting in the more effective guidance. Moreover, LLMs have shown to be effective few-shot learners in the field of NLP [2]. This naturally inspires us to further extend their capabilities to few-shot tasks in other modalities. Based on these insights, we hereby innovatively employ LLMs to tackle few-shot segmentation and introduce an entirely new framework named LLaFS. Unlike some previous segmentation methods that also use language models (LMs) but only for auxiliary purposes, such as utilizing LMs to extract intermediate features [13, 28, 64] or to generate attribute prompts [37], our LLaFS directly employs LLMs to produce segmentation results. This makes LMs no longer work as only auxiliary tools, but fully unlock their complete potential in handling the complex computer vision tasks in an end-to-end manner. In this way, we provide an important exploration towards a unified framework that allows LLMs to tackle few-shot tasks in other modalities beyond NLP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We find that integrating LLM to few-shot segmentation is non-trivial as we face three critical technical challenges: 1) How to enable the text-based LLM to comprehend and address an image processing task? 2) How to leverage both the visual information from support images and the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.225, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "text information from the LLM to guide the query segmentation? and 3) How to effectively train the model with only limited data? To address the first challenge, we draw inspiration from instruction tuning [44] and introduce a task instruction, which is used to explicitly define the few-shot segmentation task within the input of the LLM. To tackle the second challenge, we treat support images as in-context demonstration samples and design a region-attribute corresponding table to extract fine-refined multi-modal guidance information. For the third challenge, we further propose a pseudo-sample synthesis method to augment the pretraining samples and introduce a curriculum learning mechanism to achieve better optimization. By incorporating these designs, our LLaFS can handle few-shot segmentation effectively. We conduct experiments on multiple datasets and achieve state-of-the-art (SOTA) results that significantly outperform existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.485 + ], + "angle": 0, + "content": "In summary, the contributions of this work are as follows: 1) We introduce LLaFS, the first framework to address few-shot segmentation using large language models. 2) We propose various innovative designs to make better use of LLMs in few-shot segmentation, including a task-tailored instruction, a fine-grained in-context instruction serving as multimodal guidance, and a pseudo-sample-based curriculum pretraining mechanism. 3) Our approach achieves state-of-the-art performance on multiple datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.491, + 0.22, + 0.506 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Few-Shot Segmentation. To address the issue of conventional semantic segmentation methods [3, 6, 10, 18–20, 55, 61, 65, 74] that require a large number of training samples, the task of few-shot segmentation (FSS), which allows to segment a query image using only a small number of annotated support images, has been proposed and gained significant attention [1, 21, 23, 31, 35, 43, 47, 52, 63]. Current FSS methods typically adopt a prototype-guided approach [8, 17, 23, 24, 42, 63]. They use masked average pooling (MAP) to extract global [8, 34, 42] or local [24] average prototypes from the backbone features of support images, and then employ these prototypes for guiding the segmentation of query images through feature fusion [23, 24, 34], distance computation [14, 38], or attention mechanisms [45]. However, these methods can only leverage a limited amount of information extracted from a very small number of support images, thus potentially leading to suboptimal results and reduced robustness. To overcome this limitation, [64] uses the more comprehensive word embedding as the general class information to assist in segmentation. Despite some improvements, [64] is still constrained by the limited capabilities of small language models and has not delved deeper into how to better integrate textual information and support image data to achieve more effective guidance. In this paper, we are the first to employ large language models (LLMs) to achieve FSS by using our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "carefully designed instructions that contain a more effective multimodal guidance. Furthermore, we utilize the LLM to directly produce segmentation results, rather than merely using intermediate features as done in [64]. This offers a brand-new paradigm to FSS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.895, + 0.486 + ], + "angle": 0, + "content": "Large Language Models. The advent of large language models (LLMs) such as GPT [2] and Llama [50] has marked the beginning of a new era in artificial intelligence. Thanks to their significantly increased model parameters and training data, these LLMs contain rich prior knowledge and can be efficiently finetuned for specific tasks or application requirements through methods such as prompts [32, 58], adapters [15, 22] and LoRA [16]. Recently, researchers have started exploring visual large language models [25, 29, 48, 53] to establish a unified framework for multimodal data processing, aiming to override the restriction of LLMs being solely applicable to language data. However, none of them are designed for few-shot tasks in computer vision. In this paper, we introduce the first visual LLM framework for handling few-shot segmentation. To achieve this, we draw inspiration from instruction tuning [44] and in-context learning [11, 39], and carefully design a suitable form of instruction and demonstration examples tailored for few-shot segmentation. By doing so, our method can enable the LLM to comprehend image data and perform few-shot segmentation effectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.495, + 0.593, + 0.51 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.611, + 0.527 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.895, + 0.865 + ], + "angle": 0, + "content": "This paper aims to construct an LLM-based framework for few-shot segmentation, i.e., to segment a query image \\( I_{q} \\) based on \\( K \\) support images \\( \\{I_s^k\\}_{k=1}^K \\) and their ground truth maps \\( \\{G_s^k\\}_{k=1}^K \\). As shown in Fig.1, the overall framework of LLaFS can be divided into three key components: (1) a feature extractor that extracts image features and generates visual tokens; (2) a task-tailored instruction that combines visual tokens, target categories, and task requirements to provide task-related information and support guidance; and (3) an LLM that predicts segmentation masks based on the input instruction, followed by a refinement network to optimize the results. For the feature extractor, we adopt the approach in Blip2 [25] by using an image encoder followed by a Q-former and a fully-connected layer to generate a set of visual tokens. We use ResNet50 as the image encoder and keep it frozen during training. For the instruction, we carefully design it as the combination of two parts: segmentation task instruction (Sec.3.2.1) and fine-grained in-context instruction (Sec.3.2.2) to provide comprehensive and detailed guidance. For the LLM, we employ CodeLlama [46] with 7 billion parameters that have been finetuned through instruction tuning. Note that compared to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.895, + 0.903 + ], + "angle": 0, + "content": "For simplify of illustration, we introduce LLaFS under the one-shot setting. Supp presents how to extend LLaFS to the multi-shot setting." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.892, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.276, + 0.893, + 0.332 + ], + "angle": 0, + "content": "Figure 1. Overview of LLaFS. The image encoder and Q-former extract image features and generate a set of visual tokens. Subsequently, a segmentation task instruction and fine-grained in-context introduction are introduced to provide detailed and comprehensive information. These two instructions are integrated and fed into the LLM to produce the vertices coordinates of polygons that enclose the target object. The segmentation mask represented by this polygon is processed by a refinement network to get the final result." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.471, + 0.452 + ], + "angle": 0, + "content": "vanilla Llama, we empirically find that CodeLlama fine-tuned with code generation datasets exhibits higher accuracy and stability in generating structured information like the segmentation result in our task. We equip CodeLlama with LoRA for fine-tuning. All these components work together within the LLaFS framework to achieve high-quality few-shot segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.459, + 0.2, + 0.473 + ], + "angle": 0, + "content": "3.2. Instruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.47, + 0.628 + ], + "angle": 0, + "content": "As the input of LLM, the instruction is the most crucial component in our framework that makes LLM possible to handle few-shot segmentation. To provide comprehensive information, we design two instructions, namely segmentation task instruction and fine-grained in-context instruction, to respectively provide the LLM with detailed task definitions and fine-grained multi-modal guidance. These two instructions are integrated to formulate the complete instruction as shown in Fig.1. In the following Sec.3.2.1 and Sec.3.2.2, we introduce these two instructions in detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.345, + 0.655 + ], + "angle": 0, + "content": "3.2.1 Segmentation Task Instruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The LLMs trained on massive text contents have gained strong reasoning capabilities and a vast amount of world knowledge. Language instructions have shown to be a powerful tool for leveraging these knowledge and capability to handle complex tasks [44]. To achieve better results, the instructions need to be sufficiently clear and detailed, whereas those using only simple terminologies such as 'segment an image' are evidently too abstract for LLMs to comprehend. Thus, we design a structured instruction to explicitly provide more task details such as the expected input and output formats of few-shot segmentation. Specifically, in our instruction, we follow [53] by representing the pixel-wise segmentation output as a 16-point polygon that encloses the target object [30]. Note that it is hard for LLMs to directly generate pixel-wise image masks due to LLM's limited number of output tokens. Our alternative solution of generating" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.376 + ], + "angle": 0, + "content": "polygon vertices provide a token-efficient method for using LLMs to achieve pixel-level segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.893, + 0.513 + ], + "angle": 0, + "content": "Furthermore, training solely on text contents makes LLMs difficult to comprehend visual information precisely, especially in our few-shot image segmentation task, where the number of available training images is very scarce. For this, inspired by the success of in-context learning in NLP [11, 39], we further propose to encode the support image and its ground truth as a visual demonstration example, using it as an intuitive reference in the instruction to teach LLM how to segment a particular class within an image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.894, + 0.68 + ], + "angle": 0, + "content": "By incorporating these designs, we write our segmentation task instruction as: For each object within the class [class] in an image, output coordinates of a 16-point polygon that encloses the object. These points should be arranged in a clockwise direction. The output should be a tuple in the format of (c1, c2, ..., cn), where cn is the coordinates for the n-th object and its format should be ((x1,y1),(x2,y2),..., (x16,y16)). The coordinate value should be within [image size]. For example, for image [support image], the output should be [support ground truth]. Here, [support image] is the visual tokens from the support image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.688, + 0.803, + 0.703 + ], + "angle": 0, + "content": "3.2.2 Fine-grained In-context Instruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Motivation. The above task instruction makes segmenting a class possible by leveraging LLM's knowledge of the class. In the instruction, the class to be segmented is indicated by the [class] token, which is typically a single noun. However, considering that LLMs have never been trained on images, it is challenging for them to directly align this abstract noun with an image region that may possess a complex internal structure. To address this issue, we drew inspiration from human brains and found that when classifying an unseen new class, the human cognitive system follows a mechanism of 'from general to detailed, from abstract to concrete' [41, 59]. Specifically, given an unseen class represented by a general noun, the human brain first decomposes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.45, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.234, + 0.47, + 0.263 + ], + "angle": 0, + "content": "Figure 2. Illustration of how to construct the region-attribute corresponding table used in the fine-grained in-context instruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.469, + 0.385 + ], + "angle": 0, + "content": "it into detailed attributes based on the acquired knowledge. For example, in the case of an unseen class 'owl', a person can first gather information from references to learn about the owl's attributes such as 'large round eyes' and 'hooked beak'. Subsequently, it can search the image for concrete regions that match these abstract attributes to determine the presence of the class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Inspired by this, we propose a fine-grained in-context instruction to simulate such a human cognitive mechanism based on the support images. For this, we first use ChatGPT to extract detailed attributes of the target class, then we search for regions in support images that correspond to these attributes and generate a corresponding table accordingly. The obtained attributes and table constitute an in-context instruction that is fed into the LLM, which serves as a demonstration example to guide the LLM on how to recognize an image class in a more human-like and fine-grained manner. This alleviates the challenge that LLM cannot perform segmentation well when only inputted with an abstract class noun. Lastly, we introduce an expert-guide framework that refines the instruction to increase its class representation ability. The following sections explain how to generate and refine this instruction in detail." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.651, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Attributes Generation We first simulate the step of 'from general to detailed' to extract class attributes. Specifically, as shown in Fig.3(a), we construct a prompt 'What does a [class] look like? Please answer in the format of: A [class] has A, B, C,..., where A, B, and C are noun phrases to describe a [class].', and utilize ChatGPT to extract phrases-based attributes that describe the fine-grained details of this class. These attributes are denoted as [attributes] = \\(\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}\\). For each [att]i, we utilize 'A photo of [att]i' as a prompt to extract an embedding \\(t_i\\) from the CLIP's text encoder. In this way, we get \\(\\{t_i\\}_{i=1}^{N_a}\\) from \\(\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Region-attribute Corresponding Table. After that, we simulate the second step of 'from abstract to concrete'. To implement this, as shown in Fig.2, we propose a region-attribute corresponding table to find the alignment between support image regions and class attributes. For this, we first" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.872, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.298, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Figure 3. Examples of using ChatGPT for (a) class attributes generation, (b) ambiguity detection and (c) discriminative attributes generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.542 + ], + "angle": 0, + "content": "divide the support foreground into multiple local regions. Specifically, for each object in the support image within the target class, we employ the method in selective search [51] to generate a set of superpixels \\(\\{s_i\\}_{i=1}^{N_r}\\) with different scales in an unsupervised manner. Each \\(s_i\\) aggregates pixels that are close in position and similar in features, so it can represent a local region with a specific semantic meaning. Based on each \\(s_i\\), a masked image is generated and passed through CLIP's² image encoder to produce a feature \\(f_i\\). We calculate the cosine similarity between \\(f_i\\) and the embedding \\(t_j\\) for each attribute \\([\\mathrm{att}]_j\\), and utilize a thresholding process to establish region-attribute correspondence. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.552, + 0.892, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\left[ \\operatorname {c o r} \\right] _ {i} = \\left[ \\left[ \\operatorname {a t t} \\right] _ {j} \\text {f o r} j \\in \\left[ 1, N _ {a} \\right] \\text {i f} \\cos \\left(f _ {i}, t _ {j}\\right) > \\alpha \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.587, + 0.892, + 0.663 + ], + "angle": 0, + "content": "where \\( \\cos \\) refers to cosine similarity, \\( \\alpha \\) is a pre-defined threshold. The obtained \\( [\\mathrm{cor}]_i \\) contains attributes that align with \\( s_i \\). In this way, we get \\( \\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r} \\) from \\( \\{s_i\\}_{i=1}^{N_r} \\), which serves as an attribute-region corresponding table that can provide the fine-grained multi-modal reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Instruction Construction. We integrate the obtained class attributes \\(\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}\\) and corresponding table \\(\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}\\), and write the fine-grained in-context instruction as: The [class] has [attributes]. For example, in [support image], \\([\\mathrm{coord}]_{s_1}\\) is \\([\\mathrm{cor}]_1\\), \\([\\mathrm{coord}]_{s_2}\\) is \\([\\mathrm{cor}]_2\\ldots\\), \\([\\mathrm{coord}]_{s_{N_r}}\\) is \\([\\mathrm{cor}]_{N_r}\\). Here, \\([\\mathrm{coord}]_{s_i}\\) is the coordinates of \\(s_i\\) in the format of a 16-point polygon. By using this instruction, we provide the LLM with a reference about the attributes of the target class and their corresponding regions in the support image. In this way, a demonstration example can be created that simulates how the human cognitive mechanism recognizes the support image as a class. Using this example as a reference, the LLM can be taught how to understand" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.89, + 0.901 + ], + "angle": 0, + "content": "2We use CLIP fintuned through the method in [26] to ensure high-quality region-text alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3068" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.432, + 0.107 + ], + "angle": 0, + "content": "and segment an image class in a fine-grained manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.465 + ], + "angle": 0, + "content": "The above-mentioned instruction constructed by the obtained attributes \\(\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}\\) and table \\(\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}\\) can be directly input into LLM for guidance. However, due to variations in camera angles and instances of occlusion, not every attribute can be directly matched to a region within the support image. Thus, the obtained table \\(\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}\\) may contain only a subset of attributes within \\(\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}\\). Unfortunately, we find the combinations of these partial attributes may be insufficient for the unambiguous recognition of the target class. For example, the combination 'wheels, windows, doors' can refer to 'train', 'car', and 'bus' interchangeably. Due to the ambiguous table, the instruction may be misleading. To alleviate this issue, we propose an expert-guide framework to refine the instruction. In this framework, we first employ ChatGPT to identify ambiguous classes of the existing table, then we extract additional attributes that can distinguish the target class from these ambiguous classes for refinement. As long as these additional attributes can be aligned with local regions in the support image, the refined table based on them will become unambiguous. In this way, the class representation ability and comprehensiveness of the instruction can be improved." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Specifically, this framework generates a refined instruction through the following three steps: 1) Ambiguity Detection. As shown in Fig.3(b), We employ ChatGPT to identify potential ambiguous classes in the obtained table \\(\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}\\). Specifically, we denote the attributes contained in \\(\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}\\) as [partial-attributes] and ask ChatGPT 'Except for [class], which classes also have [partial-attributes]?' In this way, we obtain a set of ambiguous classes denoted as [a-classes] = \\(\\{[\\mathrm{a - class}]_i\\}_{i=1}^{N_{ac}}\\) from ChatGPT's feedback. 2) Discriminative Attributes Generation. As shown in Fig.3(c), To avoid being misled by these ambiguous classes, we use 'What does [class] look different from [a-classes]?' as a text prompt, enabling ChatGPT to generate attributes that are more discriminative from these classes. The obtained attributes \\(\\{[\\mathrm{d - att}]_i\\}_{i=1}^{N_d}\\) are added to [attributes] for updating. 3) Table and Instruction Refinement. Finally, we use these updated attributes to reperform Eq.1 to obtain a refined table. The updated attributes and table are reassembled to form a refined fine-grained in-context instruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.77, + 0.469, + 0.86 + ], + "angle": 0, + "content": "We found that a single execution of the three steps already resolves ambiguities in over \\(85\\%\\) of the instructions. For the remaining \\(15\\%\\) of instructions, we observe that because the newly-acquired discriminative attributes still couldn't find matching regions in the support image, the resulting table after refinement remains to be ambiguous." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Therefore, we iteratively apply the last two steps of this framework until the ambiguity is eradicated. To achieve this, from the second iteration onwards, we replace the text prompt in the discriminative attributes generation step with 'Apart from [all-discriminative-attributes], tell me more differences in appearance between [class] and [a-classes]\", where [all-discriminative-attributes] refers to the discriminative attributes \\([\\mathrm{d - att}]_i\\) obtained from all previous iterations. By doing so, we enable the iterative framework to continuously discover more discriminative attributes and verify whether they have matched regions in the support image. Eventually, when the framework successfully discovers discriminative attributes \\([\\mathrm{d - att}]_i\\) that can be aligned with the support image or reaches the maximum number of iterations, we terminate the iteration. For efficiency, we set the maximum number of iterations to 3, in which \\(96\\%\\) of the ambiguities have been entirely eradicated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.725, + 0.366 + ], + "angle": 0, + "content": "3.3. Segmentation Prediction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.518 + ], + "angle": 0, + "content": "We integrate segmentation task instruction and fine-grained in-context instruction to formulate the complete instruction as shown in Fig.1. With this instruction as input, the LLM can predict the coordinates of a 16-point polygon that surrounds the target object. Finally, to rectify the imprecision caused by the polygon representation of object edges, a refinement network comprising a pixel decoder and a mask transformer is introduced to generate a refined segmentation mask by using the polygon as the initial mask. Please see Supp for the detailed structures of this network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.887, + 0.539 + ], + "angle": 0, + "content": "3.4. Curriculum Pretraining with Pseudo Samples" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.546, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Motivation. After carefully designing the model structure and instruction format, the next challenge is how to train LLaFS effectively to achieve high segmentation performance. Previous work [29] has shown that the success of instruction tuning often relies on extensive training data. However, due to the difficulty of obtaining pixel-annotated labels, the segmentation datasets used for training typically contain only an insufficient number of images. To address this issue, we propose to generate pseudo support-query pairs and use them to pretrain the LLM. The LLM's ability to handle few-shot segmentation can thus be enhanced by seeing more visual samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pseudo Sample Generation. Specifically, we propose a method to generate pseudo support-query pairs with the following three steps: 1) Pseudo foreground-background partition. We first useBezier curves to randomly generate a contour within an image region. The area surrounded by this contour is considered as the foreground within the target class, while the regions outside the contour are treated as the background. 2) Noise filling for pseudo support generation. We fill the foreground with Gaussian noise that has a random mean value. For background, we first randomly divide it into multiple subregions to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "3Due to space limitations, we omit the description of format control prompts for inputting into ChatGPT. See Supp for details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3069" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.442, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.436 + ], + "angle": 0, + "content": "Figure 4. Examples of pseudo samples generated at different pretraining stages. Foreground regions are marked by white contours. As pretraining progresses, pseudo images have reduced intra-image foreground-background differences and greater support-query foreground differences. Meanwhile, the number of polygon vertex coordinates provided in the instruction decreases, while the predicted vertex count increases. These changes gradually increase the pretraining difficulty. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.65 + ], + "angle": 0, + "content": "simulate the complex backgrounds in real images, then we fill each one with a random Gaussian noise that has a mean value distinct from the foreground's noise. The obtained image is used as the support image. 3) Pseudo query generation. We use a similar approach to generate a query image. Note that in this process, the contour and the mean value of foreground noise are no longer completely randomly determined but adjusted based on those used for generating the support image. This is done to ensure that the foreground regions of both support and query images have similar contour shapes and internal features, so they can reflect the same category. Please refer to Supp for the adjustment details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Curriculum Pretraining. The synthetic support-query pairs can be directly used for pretraining. However, it is observed that this approach will result in a slow convergence rate. A reason for this problem is that the LLM has not been previously trained on image data, so optimizing it to directly grasp a complex image processing task is challenging. To address this issue, we propose a progressive pretraining approach inspired by the success of curriculum learning [54], in which we initiate the model's pretraining with a simple task and gradually increase the task's difficulty until it ultimately reaches the requirements of segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Specifically, during pretraining, we incrementally raise the task's difficulty from the following two aspects: 1) Image understanding. During pretraining, by controlling the difference between mean values of different filled noise, we gradually increase the difference in foreground between" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.424 + ], + "angle": 0, + "content": "support and query, while reducing the internal difference between foreground and background within each image. This makes it more challenging for LLM to perform few-shot guidance and partition foreground-background areas as pretraining progresses. 2) Polygon generation. We observe that generating a polygon represented by a combination of vertex coordinates is another challenge for the LLM. Therefore, we adopt a progressive strategy here as well. Instead of training the model to directly predict the coordinates of all 16 points of the polygon, we randomly provide the coordinates of \\( M \\) points in the instruction and let the LLM to predict the coordinates of the remaining \\( 16 - M \\) points. During pretraining, we gradually decrease the value of \\( M \\) from 15 to 0. This means that the model receives fewer hints and is required to predict more vertex coordinates as pretraining progresses. Consequently, the pretraining difficulty gradually increases, ultimately reaching the task of predicting all 16 points for segmentation. Experimental results show that this curriculum learning approach allows the model to converge better and achieve higher results. Please see Supp for more technical details on how we increase the difficulty in image understanding and polygon generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Ultimately, the model is trained on the realistic few-shot segmentation dataset after completing the aforementioned pretraining process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.469, + 0.633, + 0.485 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.486, + 0.887, + 0.502 + ], + "angle": 0, + "content": "4.1. Implementation Details and Training Settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.637 + ], + "angle": 0, + "content": "We set the number of queries in the Q-former to 100 and the threshold \\(\\alpha\\) in Eq.1 to 0.2. The ground truth of polygon vertices is obtained in polar coordinates [60]. Specifically, starting from the object center, 16 rays are uniformly emitted at equal angular intervals \\(\\triangle \\theta = 22.5^{\\circ}\\). The points of intersection between these rays and the object contour are taken as the ground truth of the polygon vertices. More implementation details about pseudo sample generation and curriculum pretraining are presented in Supp." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.832 + ], + "angle": 0, + "content": "The overall model is trained in three stages. In the first stage, we freeze the LLM, pretrain the Q-former and fully-connected layers for 100K steps using the datasets4 and methods in Blip2 [25]. In the second stage, we freeze the Q-former, equip the LLM with LoRA, and pretrain the fully-connected layers, LLM and refinement network using the pseudo-sample-based curriculum learning method for 60k steps. In the third stage, we train the fully-connected layers, LLM and refinement network on the realistic few-shot segmentation dataset for 25 epochs. AdamW [36] is used as the optimizer with the cosine annealing schedule and an initial learning rate of 0.0002. The model is trained on 16 A100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.835, + 0.8, + 0.851 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Table.1 presents the comparisons with other few-shot segmentation methods on two datasets: PASCAL-5 and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.89, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{4} \\) COCO is excluded from the pretraining set to avoid test data leakage." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3070" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.609, + 0.355 + ], + "angle": 0, + "content": "
DatasetMethodConference1-shot5-shot
Fold-0Fold-1Fold-2Fold-3MeanFold-0Fold-1Fold-2Fold-3Mean
PASCAL-5iNTRENet [33]CVPR'2265.472.359.459.863.266.272.861.762.265.7
BAM[23]CVPR'2269.073.667.561.167.870.675.170.867.270.9
AAFormer[57]ECCV'2269.173.359.159.265.272.574.762.061.367.6
SSP[8]ECCV'2260.567.866.451.061.467.572.375.262.169.3
IPMT[34]NeurIPS'2272.873.759.261.666.873.174.761.663.468.2
ABCNet[56]CVPR'2368.873.462.359.566.071.774.265.467.069.6
HDMNet [45]CVPR'2371.075.468.962.169.471.376.271.368.571.8
MIANet[64]CVPR'2368.575.867.563.268.770.277.470.068.871.7
MSI[40]ICCV'2371.072.563.865.968.573.074.270.566.671.1
SCGAN[62]ICCV'2368.372.566.859.866.872.374.169.165.670.3
LLaFSCVPR'2474.278.872.368.573.575.980.175.870.775.6
COCO-20iNTRENet[33]CVPR'2236.842.639.937.939.338.244.140.438.440.3
BAM[23]CVPR'2243.450.647.543.446.249.354.251.649.651.2
SSP[8]ECCV'2235.539.637.936.747.440.647.045.143.944.1
AAFormer[57]ECCV'2239.844.640.641.441.642.950.145.549.246.9
MM-Former[68]NeurIPS'2240.547.745.243.344.244.052.447.450.048.4
IPMT[34]NeurIPS'2241.445.145.640.043.043.549.748.747.947.5
ABCNet[56]CVPR'2342.346.246.042.044.145.551.752.646.449.1
HDMNet [45]CVPR'2343.855.351.649.450.050.661.655.756.056.0
MIANet[64]CVPR'2342.553.047.847.447.745.858.251.351.951.7
MSI[40]ICCV'2342.449.249.446.146.847.154.954.151.952.0
SCGAN[62]ICCV'2340.449.749.645.646.347.257.259.252.153.9
LLaFSCVPR'2447.558.856.253.053.953.263.863.160.060.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.088, + 0.357, + 0.599, + 0.37 + ], + "angle": 0, + "content": "Table 1. Performance comparison with other methods on PASCAL-5 \\( {}^{i} \\) and COCO-20 \\( {}^{i} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.394, + 0.47, + 0.483 + ], + "angle": 0, + "content": "
MethodmIoU
LLaFS74.2
LLaFs w/o segmentation task instruction w/ abstract summary67.7
LLaFS w/o fine-grained in-context instruction67.0
LLaFS w/o refinement network69.1
LLaFS w/o pseudo-sample-based curriculum pretraining63.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.487, + 0.455, + 0.501 + ], + "angle": 0, + "content": "Table 4. Effectiveness of different components in our LLaFS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.471, + 0.806 + ], + "angle": 0, + "content": "\\(\\mathrm{COCO - 20^i}\\). Following previous papers, we use different folds as the test set, with the remaining folds utilized for training. This approach yields four sets of experimental results along with their mean result. Across all datasets and experimental settings, our method consistently outperforms others and achieves a remarkably significant improvement compared to previous state-of-the-art results. For instance, on PASCAL- \\(5^{i}\\), LLaFS improves mIoUs by \\(4.1\\%\\) and \\(3.8\\%\\) in the 1-shot and 5-shot settings respectively compared to the second-ranking method. Notably, our approach still exhibits great advantages on the more complex and challenging \\(\\mathrm{COCO - 20^i}\\) dataset, with increases of \\(3.9\\%\\) and \\(4.0\\%\\) in the 1-shot and 5-shot settings respectively. This could be attributed to the rich prior knowledge embedded in the LLM and our carefully designed instructions, which enable the models to handle complex images effectively and robustly. These results demonstrate the outstanding performance of our method and highlight the huge potentiality of using LLMs for tackling few-shot segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.809, + 0.231, + 0.825 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.886 + ], + "angle": 0, + "content": "Using the 1-shot setting of the PASCAL- \\(5^{i}\\) dataset, we perform several ablation studies to evaluate different components and designs in our method. More ablation studies are presented in Supp." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Different Components We validate the" + }, + { + "type": "table", + "bbox": [ + 0.616, + 0.089, + 0.892, + 0.138 + ], + "angle": 0, + "content": "
MethodmIoU
LLaFS74.2
LLaFS w/o support images56.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.141, + 0.874, + 0.154 + ], + "angle": 0, + "content": "Table 2. Effectiveness of support images." + }, + { + "type": "table", + "bbox": [ + 0.615, + 0.16, + 0.892, + 0.238 + ], + "angle": 0, + "content": "
MethodmIoU
LLaFS74.2
LLaFS w/o class attributes70.7
LLaFS w/o region-attribute corresponding table69.7
LLaFS w/o instruction refinement70.6
LLaFS w/o iterative refinement71.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.24, + 0.894, + 0.266 + ], + "angle": 0, + "content": "Table 3. Ablation results of fine-grained in-context instruction." + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.273, + 0.88, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.357, + 0.894, + 0.382 + ], + "angle": 0, + "content": "Figure 5. Ablation for the number of polygons' sides." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.892, + 0.563 + ], + "angle": 0, + "content": "effectiveness of key components in our method, including (1) the segmentation task instruction, (2) fine-grained in-context instruction, (3) the refinement network, and (4) the pseudo-sample-based curriculum pretraining. Experimental results are presented in Table.4. Replacing the detailed segmentation task introduction with an abstract summary 'perform image segmentation' decreases mIoU by \\(6.5\\%\\). Not using the other components individually decreases the mIoU by \\(7.2\\%\\), \\(5.1\\%\\), and \\(10.7\\%\\), respectively. These results demonstrate the importance and effectiveness of each component in our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Effectiveness of Support Images. In the instruction, the annotated support images provide the LLM with crucial visual guidance. As shown in Table.2, if we do not use the support image as a demonstration example, mIoU decreases by \\(17.3\\%\\). This highlights the importance of the support image as a few-shot sample in our approach, demonstrating that our LLsFS benefits not solely from LLM's prior knowledge in an open-vocabulary manner but indeed gains further improvement from the provided few-shot samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Number of Polygon's Sides. In the segmentation task instruction, we represent the segmentation output mask as a region enclosed by a 16-point polygon. We find that the number \\( M \\) of sides in the polygon is also a factor affecting the model's performance. Therefore, we test the relationship between mIoU and \\( M \\) and present the results in Fig.5. We observe that when \\( M \\) is small, the model's performance is suboptimal. This is because polygons with a smaller number of sides cannot accurately describe object edges. As \\( M \\) increases, the mIoU gradually improves. However, when \\( M \\) exceeds 16, we observe a slight decrease in performance. This could be because a larger \\( M \\) increases the task's complexity for LLM to tackle. Based on the results," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3071" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.432, + 0.155 + ], + "angle": 0, + "content": "
MethodmIoU
LLaFS74.2
LLaFS w/o pseudo samples63.5
LLaFS w/o curriculum strategy67.3
LLaFS w/ random pseudo query generation63.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.157, + 0.468, + 0.171 + ], + "angle": 0, + "content": "Table 5. Ablation of pseudo-sample-based curriculum pretraining." + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.18, + 0.269, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.287, + 0.246, + 0.295 + ], + "angle": 0, + "content": "(a) Pretraining Loss Curve" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.18, + 0.43, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.287, + 0.404, + 0.295 + ], + "angle": 0, + "content": "(b) Training Loss Curve" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.468, + 0.335 + ], + "angle": 0, + "content": "Figure 6. Pretraining (a) and training (b) loss curves in different settings. Curriculum pretraining results in the best convergence in both pretraining and training stages. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.297, + 0.366 + ], + "angle": 0, + "content": "we chose \\(M = 16\\) as our setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.518 + ], + "angle": 0, + "content": "Ablation of Fine-grained In-context Instruction. Table.3 presents the evaluation of different components and designs in our fine-grained in-context instruction, including (1) class attributes, (2) attribute-region corresponding table, (3) expert-guide framework for instruction refinement, and (4) iterative refinement. We observe that when these components are removed, mIoU decreases by \\(3.5\\%\\), \\(4.5\\%\\), \\(3.6\\%\\) and \\(2.4\\%\\), respectively. These results demonstrate the rationality of our designs in this instruction and their effectiveness in improving performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.745 + ], + "angle": 0, + "content": "Ablation of Pseudo-sample-based Curriculum Pretraining. We further evaluate the key techniques in our pseudosample-based curriculum pretraining mechanism and the results are presented in Table.5. (1) When we do not employ pseudo-samples for pretraining, mIoU decreases by \\(10.7\\%\\). (2) Removing the curriculum training strategy that gradually increases the training task difficulty reduces mIoU by \\(6.9\\%\\). (3) When generating pseudo support-query samples, to ensure that the support and query can reflect the same category, the contour and the mean value of foreground noise used to generate the query image are adjusted based on those used for generating the support image. When this strategy is not employed and random generation is used instead, mIoU decreases by \\(10.3\\%\\). These results demonstrate the effectiveness of our method's designs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.749, + 0.208, + 0.763 + ], + "angle": 0, + "content": "4.4. Loss Curves" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In Fig.6, we present the loss curves during the pretraining and training stages. We observe from Fig.6(a) that without the use of curriculum learning, the pretraining task becomes excessively challenging, which causes the model optimization to quickly reach a bottleneck with difficulties in convergence. After using our curriculum learning mechanism, the model achieves significantly better convergence. Furthermore, in Fig.6(b), we compare the loss reduction during the training stage when using curriculum pretrain" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.09, + 0.849, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.238, + 0.836, + 0.25 + ], + "angle": 0, + "content": "Figure 7. Visualization of segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.892, + 0.362 + ], + "angle": 0, + "content": "ing, non-curriculum pretraining, and not using pretraining. The model without pretraining converges the slowest, while the model with curriculum pretraining converges the fastest during the training stage. These results demonstrate the effectiveness of our curriculum-learning-based pretraining approach in enhancing the model's convergence speed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.369, + 0.637, + 0.383 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.552 + ], + "angle": 0, + "content": "In Fig.7, we present visual samples of the segmentation results from LLaFS, including the image, ground truth, the LLM output, and the results after refinement. It can be observed that the polygons output by LLM already achieve good segmentation performance, and the results after refinement are more accurate, particularly in the object's edge regions. It is worth noting that, as shown in the third row of Fig.7, our method still achieves high-performance segmentation when there are more than one target object in the image. These results demonstrate the excellent performance of LLaFS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.618, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper proposes LLaFS, a novel framework that for the first time, leverages large language models (LLMs) to address few-shot segmentation in an end-to-end manner. To enable LLMs to perform this visual task, we introduce a segmentation task instruction to provide detailed task definitions, and propose a fine-grained in-context instruction to simulate human cognitive mechanisms and provide refined multimodal reference information. We also propose a pseudo-sample-based curriculum pretraining mechanism to augment the training samples required for instruction tuning. Our extensive experiments demonstrate the effectiveness of LLaFS, which achieves significantly superior state-of-the-art results across multiple datasets and settings. We consider LLaFS as an important exploration towards an LLM framework capable of addressing few-shot tasks in different modalities beyond natural language processing. Acknowledgement This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-006), MOE AcRF Tier 2 projects MOE-T2EP20222-0009 and MOE-T2EP20123-0014." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3072" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.198, + 0.471, + 0.267 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.267, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[4] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3367-3375, October 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.47, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.606, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[8] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701-719. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.66, + 0.471, + 0.686 + ], + "angle": 0, + "content": "[9] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.47, + 0.754 + ], + "angle": 0, + "content": "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.755, + 0.47, + 0.821 + ], + "angle": 0, + "content": "[11] Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35:30583-30598, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.822, + 0.47, + 0.876 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.876, + 0.47, + 0.89 + ], + "angle": 0, + "content": "[13] Shuting He, Henghui Ding, and Wei Jiang. Primitive gener" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "ation and semantic-related alignment for universal zero-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11238-11247, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.214 + ], + "angle": 0, + "content": "[14] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.892, + 0.284 + ], + "angle": 0, + "content": "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.284, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[16] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.338, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[17] Kai Huang, Feigege Wang, Ye Xi, and Yutao Gao. Prototypical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19256-19265, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[18] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.473, + 0.892, + 0.542 + ], + "angle": 0, + "content": "[19] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.542, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[20] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[21] Siyu Jiao, Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, and Humphrey Shi. Mask matching transformer for few-shot segmentation. In Advances in Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[22] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[23] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8057-8067, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.786, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[24] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.853, + 0.892, + 0.895 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. Inter-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3073" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.435, + 0.106 + ], + "angle": 0, + "content": "national Conference on Machine Learning, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.106, + 0.47, + 0.187 + ], + "angle": 0, + "content": "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.187, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[27] Binghao Liu, Jianbin Jiao, and Qixiang Ye. Harmonic feature activation for few-shot semantic segmentation. IEEE Transactions on Image Processing, 30:3142-3153, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.227, + 0.471, + 0.296 + ], + "angle": 0, + "content": "[28] Fang Liu, Yuhao Liu, Yuqiu Kong, Ke Xu, Lihe Zhang, Baocai Yin, Gerhard Hancke, and Rynson Lau. Referring image segmentation using text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22124-22134, October 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.469, + 0.322 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.469, + 0.539 + ], + "angle": 0, + "content": "[32] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.609 + ], + "angle": 0, + "content": "[33] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[34] Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Intermediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020-38031, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.469, + 0.717 + ], + "angle": 0, + "content": "[35] Yongfei Liu, Xiangyi Zhang, Songyang Zhang, and Xuming He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision, pages 142-158. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.81 + ], + "angle": 0, + "content": "[37] Chaofan Ma, Yuhuan Yang, Chen Ju, Fei Zhang, Ya Zhang, and Yanfeng Wang. Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.81, + 0.469, + 0.864 + ], + "angle": 0, + "content": "[38] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6941-6952, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.864, + 0.469, + 0.892 + ], + "angle": 0, + "content": "[39] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Re" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "thinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.147, + 0.892, + 0.229 + ], + "angle": 0, + "content": "[40] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Msi: Maximize support-set information for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19266-19276, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.229, + 0.892, + 0.269 + ], + "angle": 0, + "content": "[41] Gregory L Murphy and Douglas L Medin. The role of theories in conceptual coherence. Psychological review, 92(3):289, 1985. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.269, + 0.892, + 0.324 + ], + "angle": 0, + "content": "[42] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.324, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[43] Atsuro Okazawa. Interclass prototype relation for few-shot segmentation. In European Conference on Computer Vision, pages 362-378. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.365, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[44] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[45] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.472, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[46] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[47] Mennatullah Siam, Boris N Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5249-5258, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.594, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[48] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[49] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.77, + 0.892, + 0.824 + ], + "angle": 0, + "content": "[51] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.824, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[52] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9197-9206, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3074" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[53] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[54] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555-4576, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.471, + 0.27 + ], + "angle": 0, + "content": "[55] Yan Wang, Jian Cheng, Yixin Chen, Shuai Shao, Lanyun Zhu, Zhenzhou Wu, Tao Liu, and Haogang Zhu. Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation. arXiv preprint arXiv:2304.13672, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.27, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[56] Yuan Wang, Rui Sun, and Tianzhu Zhang. Rethinking the correlation in few-shot segmentation: A buoys view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7192, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.323, + 0.471, + 0.392 + ], + "angle": 0, + "content": "[57] Yuan Wang, Rui Sun, Zhe Zhang, and Tianzhu Zhang. Adaptive agent transformer for few-shot segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 36-52. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.471, + 0.447 + ], + "angle": 0, + "content": "[58] Zhen Wang, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, Huan Sun, and Yoon Kim. Multitask prompt tuning enables parameter-efficient transfer learning. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.471, + 0.486 + ], + "angle": 0, + "content": "[59] Edward J Wisniewski and Bradley C Love. Relations versus properties in conceptual combination. Journal of memory and language, 38(2):177-202, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.471, + 0.554 + ], + "angle": 0, + "content": "[60] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12193-12202, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.554, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[61] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[62] Qianxiong Xu, Wenting Zhao, Guosheng Lin, and Cheng Long. Self-calibrated cross attention network for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[63] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. Mining latent classes for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8721-8730, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.471, + 0.798 + ], + "angle": 0, + "content": "[64] Yong Yang, Qiong Chen, Yuan Feng, and Tianlin Huang. Mianet: Aggregating unbiased instance and general information for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7131-7140, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.798, + 0.471, + 0.865 + ], + "angle": 0, + "content": "[65] Ying Zang, Chenglong Fu, Runlong Cao, Didi Zhu, Min Zhang, Wenjun Hu, Lanyun Zhu, and Tianrun Chen. Resmatch: Referring expression segmentation in a semi-supervised manner. arXiv preprint arXiv:2402.05589, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.866, + 0.471, + 0.893 + ], + "angle": 0, + "content": "[66] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.148, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[67] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.216, + 0.892, + 0.27 + ], + "angle": 0, + "content": "[68] Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, Honghui Shi, et al. Mask matching transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:823-836, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.27, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[69] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[70] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[71] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3082-3092, June 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[72] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Learning gabor texture features for fine-grained recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1621-1631, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.501, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[73] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[74] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3075" + } + ] +] \ No newline at end of file diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_origin.pdf b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a66786df3bd28fb883e78cdf21d1af821ca15570 --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/aef3f26f-2462-4dec-ad92-ab4a881db4c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:331f01a476b287b3bd2f0552e2cae23b9dc930213d0bc1e7de29f61f6c329ce1 +size 3893796 diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/full.md b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e474677790d47ab909de1191211b80977683bb9e --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/full.md @@ -0,0 +1,283 @@ +# LLaFS: When Large Language Models Meet Few-Shot Segmentation + +Lanyun Zhu $^{1}$ Tianrun Chen $^{2}$ Deyi Ji $^{3}$ Jieping Ye $^{3}$ Jun Liu $^{1*}$ +Singapore University of Technology and Design $^{1}$ Zhejiang University $^{2}$ Alibaba Group $^{3}$ + +# Abstract + +This paper proposes LLaFS, the first attempt to leverage large language models (LLMs) in few-shot segmentation. In contrast to the conventional few-shot segmentation methods that only rely on the limited and biased information from the annotated support images, LLaFS leverages the vast prior knowledge gained by LLM as an effective supplement and directly uses the LLM to segment images in a few-shot manner. To enable the text-based LLM to handle image-related tasks, we carefully design an input instruction that allows the LLM to produce segmentation results represented as polygons, and propose a region-attribute table to simulate the human visual mechanism and provide multi-modal guidance. We also synthesize pseudo samples and use curriculum learning for pretraining to augment data and achieve better optimization. LLaFS achieves state-of-the-art results on multiple datasets, showing the potential of using LLMs for few-shot computer vision tasks. + +# 1. Introduction + +Image segmentation is a fundamental task in computer vision with extensive applications. The development of deep learning algorithms [7, 9, 12, 72, 73] trained on large-scale datasets has brought significant advancements to this field [3-5, 61, 69, 71]. However, annotating pixel-level segmentation ground truth on a large scale is extremely resource-intensive. Therefore, a more source-efficient learning strategy, few-shot segmentation, has received much attention from academia and holds immense practical value. + +In few-shot segmentation, the model should develop category-specific segmentation capabilities based on only a small amount of annotated data, called support images. To achieve this, existing few-shot segmentation methods [24, 27, 38, 42, 43, 49, 66, 67] typically adopt a support-feature-guided framework. In this framework, relevant features of the target category are extracted from annotated support images and used as guiding information to segment query images. To achieve higher performance, researchers have proposed many methods to explore better ways for support feature extraction [24, 38, 43] and query + +segmentation assistance [17, 56, 62]. Although these efforts have demonstrated some success, their segmentation performance is still far from satisfactory. This is because the very limited number of support images contain only a small, incomplete, and biased set of information, so the framework that relies solely on these support-based features for query segmentation inherently suffers from information constraints and cannot achieve a sufficiently high level of accuracy. Therefore, we believe that the further advancement of few-shot segmentation urgently requires an entirely new framework, which should be capable of utilizing richer and more comprehensive information, thereby breaking through the existing framework's bottlenecks to reach better results. + +We discover that recent advances in large language models (LLMs) [2, 50, 70] can offer potential opportunities to achieve this goal. Specifically, LLMs pre-trained on large-scale corpora have accumulated a vast amount of prior knowledge, which can effectively supplement the insufficient information in support images, thereby resulting in the more effective guidance. Moreover, LLMs have shown to be effective few-shot learners in the field of NLP [2]. This naturally inspires us to further extend their capabilities to few-shot tasks in other modalities. Based on these insights, we hereby innovatively employ LLMs to tackle few-shot segmentation and introduce an entirely new framework named LLaFS. Unlike some previous segmentation methods that also use language models (LMs) but only for auxiliary purposes, such as utilizing LMs to extract intermediate features [13, 28, 64] or to generate attribute prompts [37], our LLaFS directly employs LLMs to produce segmentation results. This makes LMs no longer work as only auxiliary tools, but fully unlock their complete potential in handling the complex computer vision tasks in an end-to-end manner. In this way, we provide an important exploration towards a unified framework that allows LLMs to tackle few-shot tasks in other modalities beyond NLP. + +We find that integrating LLM to few-shot segmentation is non-trivial as we face three critical technical challenges: 1) How to enable the text-based LLM to comprehend and address an image processing task? 2) How to leverage both the visual information from support images and the + +text information from the LLM to guide the query segmentation? and 3) How to effectively train the model with only limited data? To address the first challenge, we draw inspiration from instruction tuning [44] and introduce a task instruction, which is used to explicitly define the few-shot segmentation task within the input of the LLM. To tackle the second challenge, we treat support images as in-context demonstration samples and design a region-attribute corresponding table to extract fine-refined multi-modal guidance information. For the third challenge, we further propose a pseudo-sample synthesis method to augment the pretraining samples and introduce a curriculum learning mechanism to achieve better optimization. By incorporating these designs, our LLaFS can handle few-shot segmentation effectively. We conduct experiments on multiple datasets and achieve state-of-the-art (SOTA) results that significantly outperform existing methods. + +In summary, the contributions of this work are as follows: 1) We introduce LLaFS, the first framework to address few-shot segmentation using large language models. 2) We propose various innovative designs to make better use of LLMs in few-shot segmentation, including a task-tailored instruction, a fine-grained in-context instruction serving as multimodal guidance, and a pseudo-sample-based curriculum pretraining mechanism. 3) Our approach achieves state-of-the-art performance on multiple datasets. + +# 2. Related Work + +Few-Shot Segmentation. To address the issue of conventional semantic segmentation methods [3, 6, 10, 18–20, 55, 61, 65, 74] that require a large number of training samples, the task of few-shot segmentation (FSS), which allows to segment a query image using only a small number of annotated support images, has been proposed and gained significant attention [1, 21, 23, 31, 35, 43, 47, 52, 63]. Current FSS methods typically adopt a prototype-guided approach [8, 17, 23, 24, 42, 63]. They use masked average pooling (MAP) to extract global [8, 34, 42] or local [24] average prototypes from the backbone features of support images, and then employ these prototypes for guiding the segmentation of query images through feature fusion [23, 24, 34], distance computation [14, 38], or attention mechanisms [45]. However, these methods can only leverage a limited amount of information extracted from a very small number of support images, thus potentially leading to suboptimal results and reduced robustness. To overcome this limitation, [64] uses the more comprehensive word embedding as the general class information to assist in segmentation. Despite some improvements, [64] is still constrained by the limited capabilities of small language models and has not delved deeper into how to better integrate textual information and support image data to achieve more effective guidance. In this paper, we are the first to employ large language models (LLMs) to achieve FSS by using our + +carefully designed instructions that contain a more effective multimodal guidance. Furthermore, we utilize the LLM to directly produce segmentation results, rather than merely using intermediate features as done in [64]. This offers a brand-new paradigm to FSS. + +Large Language Models. The advent of large language models (LLMs) such as GPT [2] and Llama [50] has marked the beginning of a new era in artificial intelligence. Thanks to their significantly increased model parameters and training data, these LLMs contain rich prior knowledge and can be efficiently finetuned for specific tasks or application requirements through methods such as prompts [32, 58], adapters [15, 22] and LoRA [16]. Recently, researchers have started exploring visual large language models [25, 29, 48, 53] to establish a unified framework for multimodal data processing, aiming to override the restriction of LLMs being solely applicable to language data. However, none of them are designed for few-shot tasks in computer vision. In this paper, we introduce the first visual LLM framework for handling few-shot segmentation. To achieve this, we draw inspiration from instruction tuning [44] and in-context learning [11, 39], and carefully design a suitable form of instruction and demonstration examples tailored for few-shot segmentation. By doing so, our method can enable the LLM to comprehend image data and perform few-shot segmentation effectively. + +# 3. Method + +# 3.1. Overview + +This paper aims to construct an LLM-based framework for few-shot segmentation, i.e., to segment a query image $I_{q}$ based on $K$ support images $\{I_s^k\}_{k=1}^K$ and their ground truth maps $\{G_s^k\}_{k=1}^K$ . As shown in Fig.1, the overall framework of LLaFS can be divided into three key components: (1) a feature extractor that extracts image features and generates visual tokens; (2) a task-tailored instruction that combines visual tokens, target categories, and task requirements to provide task-related information and support guidance; and (3) an LLM that predicts segmentation masks based on the input instruction, followed by a refinement network to optimize the results. For the feature extractor, we adopt the approach in Blip2 [25] by using an image encoder followed by a Q-former and a fully-connected layer to generate a set of visual tokens. We use ResNet50 as the image encoder and keep it frozen during training. For the instruction, we carefully design it as the combination of two parts: segmentation task instruction (Sec.3.2.1) and fine-grained in-context instruction (Sec.3.2.2) to provide comprehensive and detailed guidance. For the LLM, we employ CodeLlama [46] with 7 billion parameters that have been finetuned through instruction tuning. Note that compared to + +![](images/47b2fbf3693f0e2b944c5c88d2b7cc7561dfaaaf6316cf6f2be3da4b3c20c274.jpg) +Figure 1. Overview of LLaFS. The image encoder and Q-former extract image features and generate a set of visual tokens. Subsequently, a segmentation task instruction and fine-grained in-context introduction are introduced to provide detailed and comprehensive information. These two instructions are integrated and fed into the LLM to produce the vertices coordinates of polygons that enclose the target object. The segmentation mask represented by this polygon is processed by a refinement network to get the final result. + +vanilla Llama, we empirically find that CodeLlama fine-tuned with code generation datasets exhibits higher accuracy and stability in generating structured information like the segmentation result in our task. We equip CodeLlama with LoRA for fine-tuning. All these components work together within the LLaFS framework to achieve high-quality few-shot segmentation. + +# 3.2. Instruction + +As the input of LLM, the instruction is the most crucial component in our framework that makes LLM possible to handle few-shot segmentation. To provide comprehensive information, we design two instructions, namely segmentation task instruction and fine-grained in-context instruction, to respectively provide the LLM with detailed task definitions and fine-grained multi-modal guidance. These two instructions are integrated to formulate the complete instruction as shown in Fig.1. In the following Sec.3.2.1 and Sec.3.2.2, we introduce these two instructions in detail. + +# 3.2.1 Segmentation Task Instruction + +The LLMs trained on massive text contents have gained strong reasoning capabilities and a vast amount of world knowledge. Language instructions have shown to be a powerful tool for leveraging these knowledge and capability to handle complex tasks [44]. To achieve better results, the instructions need to be sufficiently clear and detailed, whereas those using only simple terminologies such as 'segment an image' are evidently too abstract for LLMs to comprehend. Thus, we design a structured instruction to explicitly provide more task details such as the expected input and output formats of few-shot segmentation. Specifically, in our instruction, we follow [53] by representing the pixel-wise segmentation output as a 16-point polygon that encloses the target object [30]. Note that it is hard for LLMs to directly generate pixel-wise image masks due to LLM's limited number of output tokens. Our alternative solution of generating + +polygon vertices provide a token-efficient method for using LLMs to achieve pixel-level segmentation. + +Furthermore, training solely on text contents makes LLMs difficult to comprehend visual information precisely, especially in our few-shot image segmentation task, where the number of available training images is very scarce. For this, inspired by the success of in-context learning in NLP [11, 39], we further propose to encode the support image and its ground truth as a visual demonstration example, using it as an intuitive reference in the instruction to teach LLM how to segment a particular class within an image. + +By incorporating these designs, we write our segmentation task instruction as: For each object within the class [class] in an image, output coordinates of a 16-point polygon that encloses the object. These points should be arranged in a clockwise direction. The output should be a tuple in the format of (c1, c2, ..., cn), where cn is the coordinates for the n-th object and its format should be ((x1,y1),(x2,y2),..., (x16,y16)). The coordinate value should be within [image size]. For example, for image [support image], the output should be [support ground truth]. Here, [support image] is the visual tokens from the support image. + +# 3.2.2 Fine-grained In-context Instruction + +Motivation. The above task instruction makes segmenting a class possible by leveraging LLM's knowledge of the class. In the instruction, the class to be segmented is indicated by the [class] token, which is typically a single noun. However, considering that LLMs have never been trained on images, it is challenging for them to directly align this abstract noun with an image region that may possess a complex internal structure. To address this issue, we drew inspiration from human brains and found that when classifying an unseen new class, the human cognitive system follows a mechanism of 'from general to detailed, from abstract to concrete' [41, 59]. Specifically, given an unseen class represented by a general noun, the human brain first decomposes + +![](images/b79500931e4182d7d54ef6fbd3c91e2c0663e0a6239da670778f8bceff57b9d2.jpg) +Figure 2. Illustration of how to construct the region-attribute corresponding table used in the fine-grained in-context instruction. + +it into detailed attributes based on the acquired knowledge. For example, in the case of an unseen class 'owl', a person can first gather information from references to learn about the owl's attributes such as 'large round eyes' and 'hooked beak'. Subsequently, it can search the image for concrete regions that match these abstract attributes to determine the presence of the class. + +Inspired by this, we propose a fine-grained in-context instruction to simulate such a human cognitive mechanism based on the support images. For this, we first use ChatGPT to extract detailed attributes of the target class, then we search for regions in support images that correspond to these attributes and generate a corresponding table accordingly. The obtained attributes and table constitute an in-context instruction that is fed into the LLM, which serves as a demonstration example to guide the LLM on how to recognize an image class in a more human-like and fine-grained manner. This alleviates the challenge that LLM cannot perform segmentation well when only inputted with an abstract class noun. Lastly, we introduce an expert-guide framework that refines the instruction to increase its class representation ability. The following sections explain how to generate and refine this instruction in detail. + +Attributes Generation We first simulate the step of 'from general to detailed' to extract class attributes. Specifically, as shown in Fig.3(a), we construct a prompt 'What does a [class] look like? Please answer in the format of: A [class] has A, B, C,..., where A, B, and C are noun phrases to describe a [class].', and utilize ChatGPT to extract phrases-based attributes that describe the fine-grained details of this class. These attributes are denoted as [attributes] = $\{[\mathrm{att}]_i\}_{i=1}^{N_a}$ . For each [att]i, we utilize 'A photo of [att]i' as a prompt to extract an embedding $t_i$ from the CLIP's text encoder. In this way, we get $\{t_i\}_{i=1}^{N_a}$ from $\{[\mathrm{att}]_i\}_{i=1}^{N_a}$ . + +Region-attribute Corresponding Table. After that, we simulate the second step of 'from abstract to concrete'. To implement this, as shown in Fig.2, we propose a region-attribute corresponding table to find the alignment between support image regions and class attributes. For this, we first + +![](images/100c4de25f54d47908bea7ae65a5f2951da3fcf61868c74417a9dc89d6c6bf78.jpg) +Figure 3. Examples of using ChatGPT for (a) class attributes generation, (b) ambiguity detection and (c) discriminative attributes generation. + +divide the support foreground into multiple local regions. Specifically, for each object in the support image within the target class, we employ the method in selective search [51] to generate a set of superpixels $\{s_i\}_{i=1}^{N_r}$ with different scales in an unsupervised manner. Each $s_i$ aggregates pixels that are close in position and similar in features, so it can represent a local region with a specific semantic meaning. Based on each $s_i$ , a masked image is generated and passed through CLIP's² image encoder to produce a feature $f_i$ . We calculate the cosine similarity between $f_i$ and the embedding $t_j$ for each attribute $[\mathrm{att}]_j$ , and utilize a thresholding process to establish region-attribute correspondence. Formally, + +$$ +\left[ \operatorname {c o r} \right] _ {i} = \left[ \left[ \operatorname {a t t} \right] _ {j} \text {f o r} j \in \left[ 1, N _ {a} \right] \text {i f} \cos \left(f _ {i}, t _ {j}\right) > \alpha \right] \tag {1} +$$ + +where $\cos$ refers to cosine similarity, $\alpha$ is a pre-defined threshold. The obtained $[\mathrm{cor}]_i$ contains attributes that align with $s_i$ . In this way, we get $\{[\mathrm{cor}]_i\}_{i=1}^{N_r}$ from $\{s_i\}_{i=1}^{N_r}$ , which serves as an attribute-region corresponding table that can provide the fine-grained multi-modal reference. + +Instruction Construction. We integrate the obtained class attributes $\{[\mathrm{att}]_i\}_{i=1}^{N_a}$ and corresponding table $\{[\mathrm{cor}]_i\}_{i=1}^{N_r}$ , and write the fine-grained in-context instruction as: The [class] has [attributes]. For example, in [support image], $[\mathrm{coord}]_{s_1}$ is $[\mathrm{cor}]_1$ , $[\mathrm{coord}]_{s_2}$ is $[\mathrm{cor}]_2\ldots$ , $[\mathrm{coord}]_{s_{N_r}}$ is $[\mathrm{cor}]_{N_r}$ . Here, $[\mathrm{coord}]_{s_i}$ is the coordinates of $s_i$ in the format of a 16-point polygon. By using this instruction, we provide the LLM with a reference about the attributes of the target class and their corresponding regions in the support image. In this way, a demonstration example can be created that simulates how the human cognitive mechanism recognizes the support image as a class. Using this example as a reference, the LLM can be taught how to understand + +and segment an image class in a fine-grained manner. + +The above-mentioned instruction constructed by the obtained attributes $\{[\mathrm{att}]_i\}_{i = 1}^{N_a}$ and table $\{[\mathrm{cor}]_i\}_{i = 1}^{N_r}$ can be directly input into LLM for guidance. However, due to variations in camera angles and instances of occlusion, not every attribute can be directly matched to a region within the support image. Thus, the obtained table $\{[\mathrm{cor}]_i\}_{i = 1}^{N_r}$ may contain only a subset of attributes within $\{[\mathrm{att}]_i\}_{i = 1}^{N_a}$ . Unfortunately, we find the combinations of these partial attributes may be insufficient for the unambiguous recognition of the target class. For example, the combination 'wheels, windows, doors' can refer to 'train', 'car', and 'bus' interchangeably. Due to the ambiguous table, the instruction may be misleading. To alleviate this issue, we propose an expert-guide framework to refine the instruction. In this framework, we first employ ChatGPT to identify ambiguous classes of the existing table, then we extract additional attributes that can distinguish the target class from these ambiguous classes for refinement. As long as these additional attributes can be aligned with local regions in the support image, the refined table based on them will become unambiguous. In this way, the class representation ability and comprehensiveness of the instruction can be improved. + +Specifically, this framework generates a refined instruction through the following three steps: 1) Ambiguity Detection. As shown in Fig.3(b), We employ ChatGPT to identify potential ambiguous classes in the obtained table $\{[\mathrm{cor}]_i\}_{i=1}^{N_r}$ . Specifically, we denote the attributes contained in $\{[\mathrm{cor}]_i\}_{i=1}^{N_r}$ as [partial-attributes] and ask ChatGPT 'Except for [class], which classes also have [partial-attributes]?' In this way, we obtain a set of ambiguous classes denoted as [a-classes] = $\{[\mathrm{a - class}]_i\}_{i=1}^{N_{ac}}$ from ChatGPT's feedback. 2) Discriminative Attributes Generation. As shown in Fig.3(c), To avoid being misled by these ambiguous classes, we use 'What does [class] look different from [a-classes]?' as a text prompt, enabling ChatGPT to generate attributes that are more discriminative from these classes. The obtained attributes $\{[\mathrm{d - att}]_i\}_{i=1}^{N_d}$ are added to [attributes] for updating. 3) Table and Instruction Refinement. Finally, we use these updated attributes to reperform Eq.1 to obtain a refined table. The updated attributes and table are reassembled to form a refined fine-grained in-context instruction. + +We found that a single execution of the three steps already resolves ambiguities in over $85\%$ of the instructions. For the remaining $15\%$ of instructions, we observe that because the newly-acquired discriminative attributes still couldn't find matching regions in the support image, the resulting table after refinement remains to be ambiguous. + +Therefore, we iteratively apply the last two steps of this framework until the ambiguity is eradicated. To achieve this, from the second iteration onwards, we replace the text prompt in the discriminative attributes generation step with 'Apart from [all-discriminative-attributes], tell me more differences in appearance between [class] and [a-classes]", where [all-discriminative-attributes] refers to the discriminative attributes $[\mathrm{d - att}]_i$ obtained from all previous iterations. By doing so, we enable the iterative framework to continuously discover more discriminative attributes and verify whether they have matched regions in the support image. Eventually, when the framework successfully discovers discriminative attributes $[\mathrm{d - att}]_i$ that can be aligned with the support image or reaches the maximum number of iterations, we terminate the iteration. For efficiency, we set the maximum number of iterations to 3, in which $96\%$ of the ambiguities have been entirely eradicated. + +# 3.3. Segmentation Prediction + +We integrate segmentation task instruction and fine-grained in-context instruction to formulate the complete instruction as shown in Fig.1. With this instruction as input, the LLM can predict the coordinates of a 16-point polygon that surrounds the target object. Finally, to rectify the imprecision caused by the polygon representation of object edges, a refinement network comprising a pixel decoder and a mask transformer is introduced to generate a refined segmentation mask by using the polygon as the initial mask. Please see Supp for the detailed structures of this network. + +# 3.4. Curriculum Pretraining with Pseudo Samples + +Motivation. After carefully designing the model structure and instruction format, the next challenge is how to train LLaFS effectively to achieve high segmentation performance. Previous work [29] has shown that the success of instruction tuning often relies on extensive training data. However, due to the difficulty of obtaining pixel-annotated labels, the segmentation datasets used for training typically contain only an insufficient number of images. To address this issue, we propose to generate pseudo support-query pairs and use them to pretrain the LLM. The LLM's ability to handle few-shot segmentation can thus be enhanced by seeing more visual samples. + +Pseudo Sample Generation. Specifically, we propose a method to generate pseudo support-query pairs with the following three steps: 1) Pseudo foreground-background partition. We first useBezier curves to randomly generate a contour within an image region. The area surrounded by this contour is considered as the foreground within the target class, while the regions outside the contour are treated as the background. 2) Noise filling for pseudo support generation. We fill the foreground with Gaussian noise that has a random mean value. For background, we first randomly divide it into multiple subregions to + +![](images/6362bb87ffc0ff2894f271444ec470e5bedf155a7c122faf5e9ce400cb48f4dc.jpg) +Figure 4. Examples of pseudo samples generated at different pretraining stages. Foreground regions are marked by white contours. As pretraining progresses, pseudo images have reduced intra-image foreground-background differences and greater support-query foreground differences. Meanwhile, the number of polygon vertex coordinates provided in the instruction decreases, while the predicted vertex count increases. These changes gradually increase the pretraining difficulty. (Best viewed in color) + +simulate the complex backgrounds in real images, then we fill each one with a random Gaussian noise that has a mean value distinct from the foreground's noise. The obtained image is used as the support image. 3) Pseudo query generation. We use a similar approach to generate a query image. Note that in this process, the contour and the mean value of foreground noise are no longer completely randomly determined but adjusted based on those used for generating the support image. This is done to ensure that the foreground regions of both support and query images have similar contour shapes and internal features, so they can reflect the same category. Please refer to Supp for the adjustment details. + +Curriculum Pretraining. The synthetic support-query pairs can be directly used for pretraining. However, it is observed that this approach will result in a slow convergence rate. A reason for this problem is that the LLM has not been previously trained on image data, so optimizing it to directly grasp a complex image processing task is challenging. To address this issue, we propose a progressive pretraining approach inspired by the success of curriculum learning [54], in which we initiate the model's pretraining with a simple task and gradually increase the task's difficulty until it ultimately reaches the requirements of segmentation. + +Specifically, during pretraining, we incrementally raise the task's difficulty from the following two aspects: 1) Image understanding. During pretraining, by controlling the difference between mean values of different filled noise, we gradually increase the difference in foreground between + +support and query, while reducing the internal difference between foreground and background within each image. This makes it more challenging for LLM to perform few-shot guidance and partition foreground-background areas as pretraining progresses. 2) Polygon generation. We observe that generating a polygon represented by a combination of vertex coordinates is another challenge for the LLM. Therefore, we adopt a progressive strategy here as well. Instead of training the model to directly predict the coordinates of all 16 points of the polygon, we randomly provide the coordinates of $M$ points in the instruction and let the LLM to predict the coordinates of the remaining $16 - M$ points. During pretraining, we gradually decrease the value of $M$ from 15 to 0. This means that the model receives fewer hints and is required to predict more vertex coordinates as pretraining progresses. Consequently, the pretraining difficulty gradually increases, ultimately reaching the task of predicting all 16 points for segmentation. Experimental results show that this curriculum learning approach allows the model to converge better and achieve higher results. Please see Supp for more technical details on how we increase the difficulty in image understanding and polygon generation. + +Ultimately, the model is trained on the realistic few-shot segmentation dataset after completing the aforementioned pretraining process. + +# 4. Experiments + +# 4.1. Implementation Details and Training Settings + +We set the number of queries in the Q-former to 100 and the threshold $\alpha$ in Eq.1 to 0.2. The ground truth of polygon vertices is obtained in polar coordinates [60]. Specifically, starting from the object center, 16 rays are uniformly emitted at equal angular intervals $\triangle \theta = 22.5^{\circ}$ . The points of intersection between these rays and the object contour are taken as the ground truth of the polygon vertices. More implementation details about pseudo sample generation and curriculum pretraining are presented in Supp. + +The overall model is trained in three stages. In the first stage, we freeze the LLM, pretrain the Q-former and fully-connected layers for 100K steps using the datasets4 and methods in Blip2 [25]. In the second stage, we freeze the Q-former, equip the LLM with LoRA, and pretrain the fully-connected layers, LLM and refinement network using the pseudo-sample-based curriculum learning method for 60k steps. In the third stage, we train the fully-connected layers, LLM and refinement network on the realistic few-shot segmentation dataset for 25 epochs. AdamW [36] is used as the optimizer with the cosine annealing schedule and an initial learning rate of 0.0002. The model is trained on 16 A100 GPUs. + +# 4.2. Comparison with State-of-the-arts + +Table.1 presents the comparisons with other few-shot segmentation methods on two datasets: PASCAL-5 and + +
DatasetMethodConference1-shot5-shot
Fold-0Fold-1Fold-2Fold-3MeanFold-0Fold-1Fold-2Fold-3Mean
PASCAL-5iNTRENet [33]CVPR'2265.472.359.459.863.266.272.861.762.265.7
BAM[23]CVPR'2269.073.667.561.167.870.675.170.867.270.9
AAFormer[57]ECCV'2269.173.359.159.265.272.574.762.061.367.6
SSP[8]ECCV'2260.567.866.451.061.467.572.375.262.169.3
IPMT[34]NeurIPS'2272.873.759.261.666.873.174.761.663.468.2
ABCNet[56]CVPR'2368.873.462.359.566.071.774.265.467.069.6
HDMNet [45]CVPR'2371.075.468.962.169.471.376.271.368.571.8
MIANet[64]CVPR'2368.575.867.563.268.770.277.470.068.871.7
MSI[40]ICCV'2371.072.563.865.968.573.074.270.566.671.1
SCGAN[62]ICCV'2368.372.566.859.866.872.374.169.165.670.3
LLaFSCVPR'2474.278.872.368.573.575.980.175.870.775.6
COCO-20iNTRENet[33]CVPR'2236.842.639.937.939.338.244.140.438.440.3
BAM[23]CVPR'2243.450.647.543.446.249.354.251.649.651.2
SSP[8]ECCV'2235.539.637.936.747.440.647.045.143.944.1
AAFormer[57]ECCV'2239.844.640.641.441.642.950.145.549.246.9
MM-Former[68]NeurIPS'2240.547.745.243.344.244.052.447.450.048.4
IPMT[34]NeurIPS'2241.445.145.640.043.043.549.748.747.947.5
ABCNet[56]CVPR'2342.346.246.042.044.145.551.752.646.449.1
HDMNet [45]CVPR'2343.855.351.649.450.050.661.655.756.056.0
MIANet[64]CVPR'2342.553.047.847.447.745.858.251.351.951.7
MSI[40]ICCV'2342.449.249.446.146.847.154.954.151.952.0
SCGAN[62]ICCV'2340.449.749.645.646.347.257.259.252.153.9
LLaFSCVPR'2447.558.856.253.053.953.263.863.160.060.0
+ +Table 1. Performance comparison with other methods on PASCAL-5 ${}^{i}$ and COCO-20 ${}^{i}$ . + +
MethodmIoU
LLaFS74.2
LLaFs w/o segmentation task instruction w/ abstract summary67.7
LLaFS w/o fine-grained in-context instruction67.0
LLaFS w/o refinement network69.1
LLaFS w/o pseudo-sample-based curriculum pretraining63.5
+ +$\mathrm{COCO - 20^i}$ . Following previous papers, we use different folds as the test set, with the remaining folds utilized for training. This approach yields four sets of experimental results along with their mean result. Across all datasets and experimental settings, our method consistently outperforms others and achieves a remarkably significant improvement compared to previous state-of-the-art results. For instance, on PASCAL- $5^{i}$ , LLaFS improves mIoUs by $4.1\%$ and $3.8\%$ in the 1-shot and 5-shot settings respectively compared to the second-ranking method. Notably, our approach still exhibits great advantages on the more complex and challenging $\mathrm{COCO - 20^i}$ dataset, with increases of $3.9\%$ and $4.0\%$ in the 1-shot and 5-shot settings respectively. This could be attributed to the rich prior knowledge embedded in the LLM and our carefully designed instructions, which enable the models to handle complex images effectively and robustly. These results demonstrate the outstanding performance of our method and highlight the huge potentiality of using LLMs for tackling few-shot segmentation tasks. + +# 4.3. Ablation Study + +Using the 1-shot setting of the PASCAL- $5^{i}$ dataset, we perform several ablation studies to evaluate different components and designs in our method. More ablation studies are presented in Supp. + +Effectiveness of Different Components We validate the + +Table 4. Effectiveness of different components in our LLaFS. + +
MethodmIoU
LLaFS74.2
LLaFS w/o support images56.9
+ +Table 2. Effectiveness of support images. + +
MethodmIoU
LLaFS74.2
LLaFS w/o class attributes70.7
LLaFS w/o region-attribute corresponding table69.7
LLaFS w/o instruction refinement70.6
LLaFS w/o iterative refinement71.8
+ +Table 3. Ablation results of fine-grained in-context instruction. + +![](images/51b6b9223321d782271e26111de467c4967e991122fb189dee5ab5fa071e5618.jpg) +Figure 5. Ablation for the number of polygons' sides. + +effectiveness of key components in our method, including (1) the segmentation task instruction, (2) fine-grained in-context instruction, (3) the refinement network, and (4) the pseudo-sample-based curriculum pretraining. Experimental results are presented in Table.4. Replacing the detailed segmentation task introduction with an abstract summary 'perform image segmentation' decreases mIoU by $6.5\%$ . Not using the other components individually decreases the mIoU by $7.2\%$ , $5.1\%$ , and $10.7\%$ , respectively. These results demonstrate the importance and effectiveness of each component in our approach. + +Effectiveness of Support Images. In the instruction, the annotated support images provide the LLM with crucial visual guidance. As shown in Table.2, if we do not use the support image as a demonstration example, mIoU decreases by $17.3\%$ . This highlights the importance of the support image as a few-shot sample in our approach, demonstrating that our LLsFS benefits not solely from LLM's prior knowledge in an open-vocabulary manner but indeed gains further improvement from the provided few-shot samples. + +Number of Polygon's Sides. In the segmentation task instruction, we represent the segmentation output mask as a region enclosed by a 16-point polygon. We find that the number $M$ of sides in the polygon is also a factor affecting the model's performance. Therefore, we test the relationship between mIoU and $M$ and present the results in Fig.5. We observe that when $M$ is small, the model's performance is suboptimal. This is because polygons with a smaller number of sides cannot accurately describe object edges. As $M$ increases, the mIoU gradually improves. However, when $M$ exceeds 16, we observe a slight decrease in performance. This could be because a larger $M$ increases the task's complexity for LLM to tackle. Based on the results, + +
MethodmIoU
LLaFS74.2
LLaFS w/o pseudo samples63.5
LLaFS w/o curriculum strategy67.3
LLaFS w/ random pseudo query generation63.9
+ +Table 5. Ablation of pseudo-sample-based curriculum pretraining. + +![](images/e4795b51d371e5f37cb40f7e7248b85fcec20bd966c0815f6fb35e2252174158.jpg) +(a) Pretraining Loss Curve + +![](images/9107691691716bccf12180c480cc0827b9bae0fe420c73e9aadbff5eb0e3cf9f.jpg) +(b) Training Loss Curve +Figure 6. Pretraining (a) and training (b) loss curves in different settings. Curriculum pretraining results in the best convergence in both pretraining and training stages. (Best viewed in color) + +we chose $M = 16$ as our setting. + +Ablation of Fine-grained In-context Instruction. Table.3 presents the evaluation of different components and designs in our fine-grained in-context instruction, including (1) class attributes, (2) attribute-region corresponding table, (3) expert-guide framework for instruction refinement, and (4) iterative refinement. We observe that when these components are removed, mIoU decreases by $3.5\%$ , $4.5\%$ , $3.6\%$ and $2.4\%$ , respectively. These results demonstrate the rationality of our designs in this instruction and their effectiveness in improving performance. + +Ablation of Pseudo-sample-based Curriculum Pretraining. We further evaluate the key techniques in our pseudosample-based curriculum pretraining mechanism and the results are presented in Table.5. (1) When we do not employ pseudo-samples for pretraining, mIoU decreases by $10.7\%$ . (2) Removing the curriculum training strategy that gradually increases the training task difficulty reduces mIoU by $6.9\%$ . (3) When generating pseudo support-query samples, to ensure that the support and query can reflect the same category, the contour and the mean value of foreground noise used to generate the query image are adjusted based on those used for generating the support image. When this strategy is not employed and random generation is used instead, mIoU decreases by $10.3\%$ . These results demonstrate the effectiveness of our method's designs. + +# 4.4. Loss Curves + +In Fig.6, we present the loss curves during the pretraining and training stages. We observe from Fig.6(a) that without the use of curriculum learning, the pretraining task becomes excessively challenging, which causes the model optimization to quickly reach a bottleneck with difficulties in convergence. After using our curriculum learning mechanism, the model achieves significantly better convergence. Furthermore, in Fig.6(b), we compare the loss reduction during the training stage when using curriculum pretrain + +![](images/6d625049e27dab562014779b6c62fa5532e20347b2eeb4deb42bd1624a8f9d69.jpg) +Figure 7. Visualization of segmentation results. + +ing, non-curriculum pretraining, and not using pretraining. The model without pretraining converges the slowest, while the model with curriculum pretraining converges the fastest during the training stage. These results demonstrate the effectiveness of our curriculum-learning-based pretraining approach in enhancing the model's convergence speed. + +# 4.5. Visualization + +In Fig.7, we present visual samples of the segmentation results from LLaFS, including the image, ground truth, the LLM output, and the results after refinement. It can be observed that the polygons output by LLM already achieve good segmentation performance, and the results after refinement are more accurate, particularly in the object's edge regions. It is worth noting that, as shown in the third row of Fig.7, our method still achieves high-performance segmentation when there are more than one target object in the image. These results demonstrate the excellent performance of LLaFS. + +# 5. Conclusion + +This paper proposes LLaFS, a novel framework that for the first time, leverages large language models (LLMs) to address few-shot segmentation in an end-to-end manner. To enable LLMs to perform this visual task, we introduce a segmentation task instruction to provide detailed task definitions, and propose a fine-grained in-context instruction to simulate human cognitive mechanisms and provide refined multimodal reference information. We also propose a pseudo-sample-based curriculum pretraining mechanism to augment the training samples required for instruction tuning. Our extensive experiments demonstrate the effectiveness of LLaFS, which achieves significantly superior state-of-the-art results across multiple datasets and settings. We consider LLaFS as an important exploration towards an LLM framework capable of addressing few-shot tasks in different modalities beyond natural language processing. Acknowledgement This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-006), MOE AcRF Tier 2 projects MOE-T2EP20222-0009 and MOE-T2EP20123-0014. + +# References + +[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021. 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2 +[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2 +[4] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3367-3375, October 2023. +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1 +[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[8] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701-719. Springer, 2022. 2, 7 +[9] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 1 +[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2 +[11] Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35:30583-30598, 2022. 2, 3 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1 +[13] Shuting He, Henghui Ding, and Wei Jiang. Primitive gener + +ation and semantic-related alignment for universal zero-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11238-11247, 2023. 1 +[14] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 2 +[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2 +[16] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 2 +[17] Kai Huang, Feigege Wang, Ye Xi, and Yutao Gao. Prototypical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19256-19265, 2023. 1, 2 +[18] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021. 2 +[19] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022. +[20] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 2 +[21] Siyu Jiao, Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, and Humphrey Shi. Mask matching transformer for few-shot segmentation. In Advances in Neural Information Processing Systems. 2 +[22] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 2 +[23] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8057-8067, 2022. 2, 7 +[24] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021. 1, 2 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. Inter- + +national Conference on Machine Learning, 2023. 2, 6 +[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 4 +[27] Binghao Liu, Jianbin Jiao, and Qixiang Ye. Harmonic feature activation for few-shot semantic segmentation. IEEE Transactions on Image Processing, 30:3142-3153, 2021. 1 +[28] Fang Liu, Yuhao Liu, Yuqiu Kong, Ke Xu, Lihe Zhang, Baocai Yin, Gerhard Hancke, and Rynson Lau. Referring image segmentation using text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22124-22134, October 2023. 1 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5 +[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 3 +[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 2 +[32] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, 2022. 2 +[33] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022. 7 +[34] Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Intermediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020-38031, 2022. 2, 7 +[35] Yongfei Liu, Xiangyi Zhang, Songyang Zhang, and Xuming He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision, pages 142-158. Springer, 2020. 2 +[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6 +[37] Chaofan Ma, Yuhuan Yang, Chen Ju, Fei Zhang, Ya Zhang, and Yanfeng Wang. Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096, 2023. 1 +[38] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6941-6952, 2021. 1, 2 +[39] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Re + +thinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, 2022. 2, 3 +[40] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Msi: Maximize support-set information for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19266-19276, 2023. 7 +[41] Gregory L Murphy and Douglas L Medin. The role of theories in conceptual coherence. Psychological review, 92(3):289, 1985. 3 +[42] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 1, 2 +[43] Atsuro Okazawa. Interclass prototype relation for few-shot segmentation. In European Conference on Computer Vision, pages 362-378. Springer, 2022. 1, 2 +[44] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2, 3 +[45] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 2, 7 +[46] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2 +[47] Mennatullah Siam, Boris N Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5249-5258, 2019. 2 +[48] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 2 +[49] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 1 +[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2 +[51] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 4 +[52] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9197-9206, 2019. 2 + +[53] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2, 3 +[54] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555-4576, 2021. 6 +[55] Yan Wang, Jian Cheng, Yixin Chen, Shuai Shao, Lanyun Zhu, Zhenzhou Wu, Tao Liu, and Haogang Zhu. Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation. arXiv preprint arXiv:2304.13672, 2023. 2 +[56] Yuan Wang, Rui Sun, and Tianzhu Zhang. Rethinking the correlation in few-shot segmentation: A buoys view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7192, 2023. 1, 7 +[57] Yuan Wang, Rui Sun, Zhe Zhang, and Tianzhu Zhang. Adaptive agent transformer for few-shot segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 36-52. Springer, 2022. 7 +[58] Zhen Wang, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, Huan Sun, and Yoon Kim. Multitask prompt tuning enables parameter-efficient transfer learning. In International Conference on Learning Representations, 2022. 2 +[59] Edward J Wisniewski and Bradley C Love. Relations versus properties in conceptual combination. Journal of memory and language, 38(2):177-202, 1998. 3 +[60] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12193-12202, 2020. 6 +[61] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1, 2 +[62] Qianxiong Xu, Wenting Zhao, Guosheng Lin, and Cheng Long. Self-calibrated cross attention network for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7 +[63] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. Mining latent classes for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8721-8730, 2021. 2 +[64] Yong Yang, Qiong Chen, Yuan Feng, and Tianlin Huang. Mianet: Aggregating unbiased instance and general information for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7131-7140, 2023. 1, 2, 7 +[65] Ying Zang, Chenglong Fu, Runlong Cao, Didi Zhu, Min Zhang, Wenjun Hu, Lanyun Zhu, and Tianrun Chen. Resmatch: Referring expression segmentation in a semi-supervised manner. arXiv preprint arXiv:2402.05589, 2024. 2 +[66] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with + +connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 1 +[67] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 1 +[68] Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, Honghui Shi, et al. Mask matching transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:823-836, 2022. 7 +[69] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1 +[70] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1 +[71] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3082-3092, June 2023. 1 +[72] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Learning gabor texture features for fine-grained recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1621-1631, 2023. 1 +[73] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024. 1 +[74] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 2 \ No newline at end of file diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/images.zip b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8824991faddb723f43a0f7e9f0ac9b299817f9f8 --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4909e37c474219dab2ed3d2b4d2d6d0186051c195e6ed91658de862b9c02877 +size 507813 diff --git a/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/layout.json b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..54111453ea02df06e276b5370d755cb5d0752a57 --- /dev/null +++ b/2024/LLaFS_ When Large Language Models Meet Few-Shot Segmentation/layout.json @@ -0,0 +1,8650 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 121 + ], + "type": "text", + "content": "LLaFS: When Large Language Models Meet Few-Shot Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "spans": [ + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": "Lanyun Zhu" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Tianrun Chen" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Deyi Ji" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Jieping Ye" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Jun Liu" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " \nSingapore University of Technology and Design" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Zhejiang University" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "text", + "content": " Alibaba Group" + }, + { + "bbox": [ + 69, + 142, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 219, + 289, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 423 + ], + "type": "text", + "content": "This paper proposes LLaFS, the first attempt to leverage large language models (LLMs) in few-shot segmentation. In contrast to the conventional few-shot segmentation methods that only rely on the limited and biased information from the annotated support images, LLaFS leverages the vast prior knowledge gained by LLM as an effective supplement and directly uses the LLM to segment images in a few-shot manner. To enable the text-based LLM to handle image-related tasks, we carefully design an input instruction that allows the LLM to produce segmentation results represented as polygons, and propose a region-attribute table to simulate the human visual mechanism and provide multi-modal guidance. We also synthesize pseudo samples and use curriculum learning for pretraining to augment data and achieve better optimization. LLaFS achieves state-of-the-art results on multiple datasets, showing the potential of using LLMs for few-shot computer vision tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "content": "Image segmentation is a fundamental task in computer vision with extensive applications. The development of deep learning algorithms [7, 9, 12, 72, 73] trained on large-scale datasets has brought significant advancements to this field [3-5, 61, 69, 71]. However, annotating pixel-level segmentation ground truth on a large scale is extremely resource-intensive. Therefore, a more source-efficient learning strategy, few-shot segmentation, has received much attention from academia and holds immense practical value." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 693 + ], + "type": "text", + "content": "In few-shot segmentation, the model should develop category-specific segmentation capabilities based on only a small amount of annotated data, called support images. To achieve this, existing few-shot segmentation methods [24, 27, 38, 42, 43, 49, 66, 67] typically adopt a support-feature-guided framework. In this framework, relevant features of the target category are extracted from annotated support images and used as guiding information to segment query images. To achieve higher performance, researchers have proposed many methods to explore better ways for support feature extraction [24, 38, 43] and query" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 200, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 547, + 368 + ], + "type": "text", + "content": "segmentation assistance [17, 56, 62]. Although these efforts have demonstrated some success, their segmentation performance is still far from satisfactory. This is because the very limited number of support images contain only a small, incomplete, and biased set of information, so the framework that relies solely on these support-based features for query segmentation inherently suffers from information constraints and cannot achieve a sufficiently high level of accuracy. Therefore, we believe that the further advancement of few-shot segmentation urgently requires an entirely new framework, which should be capable of utilizing richer and more comprehensive information, thereby breaking through the existing framework's bottlenecks to reach better results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 373, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 649 + ], + "type": "text", + "content": "We discover that recent advances in large language models (LLMs) [2, 50, 70] can offer potential opportunities to achieve this goal. Specifically, LLMs pre-trained on large-scale corpora have accumulated a vast amount of prior knowledge, which can effectively supplement the insufficient information in support images, thereby resulting in the more effective guidance. Moreover, LLMs have shown to be effective few-shot learners in the field of NLP [2]. This naturally inspires us to further extend their capabilities to few-shot tasks in other modalities. Based on these insights, we hereby innovatively employ LLMs to tackle few-shot segmentation and introduce an entirely new framework named LLaFS. Unlike some previous segmentation methods that also use language models (LMs) but only for auxiliary purposes, such as utilizing LMs to extract intermediate features [13, 28, 64] or to generate attribute prompts [37], our LLaFS directly employs LLMs to produce segmentation results. This makes LMs no longer work as only auxiliary tools, but fully unlock their complete potential in handling the complex computer vision tasks in an end-to-end manner. In this way, we provide an important exploration towards a unified framework that allows LLMs to tackle few-shot tasks in other modalities beyond NLP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "We find that integrating LLM to few-shot segmentation is non-trivial as we face three critical technical challenges: 1) How to enable the text-based LLM to comprehend and address an image processing task? 2) How to leverage both the visual information from support images and the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3065" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "text information from the LLM to guide the query segmentation? and 3) How to effectively train the model with only limited data? To address the first challenge, we draw inspiration from instruction tuning [44] and introduce a task instruction, which is used to explicitly define the few-shot segmentation task within the input of the LLM. To tackle the second challenge, we treat support images as in-context demonstration samples and design a region-attribute corresponding table to extract fine-refined multi-modal guidance information. For the third challenge, we further propose a pseudo-sample synthesis method to augment the pretraining samples and introduce a curriculum learning mechanism to achieve better optimization. By incorporating these designs, our LLaFS can handle few-shot segmentation effectively. We conduct experiments on multiple datasets and achieve state-of-the-art (SOTA) results that significantly outperform existing methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 384 + ], + "type": "text", + "content": "In summary, the contributions of this work are as follows: 1) We introduce LLaFS, the first framework to address few-shot segmentation using large language models. 2) We propose various innovative designs to make better use of LLMs in few-shot segmentation, including a task-tailored instruction, a fine-grained in-context instruction serving as multimodal guidance, and a pseudo-sample-based curriculum pretraining mechanism. 3) Our approach achieves state-of-the-art performance on multiple datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 388, + 134, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 134, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 134, + 400 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 715 + ], + "type": "text", + "content": "Few-Shot Segmentation. To address the issue of conventional semantic segmentation methods [3, 6, 10, 18–20, 55, 61, 65, 74] that require a large number of training samples, the task of few-shot segmentation (FSS), which allows to segment a query image using only a small number of annotated support images, has been proposed and gained significant attention [1, 21, 23, 31, 35, 43, 47, 52, 63]. Current FSS methods typically adopt a prototype-guided approach [8, 17, 23, 24, 42, 63]. They use masked average pooling (MAP) to extract global [8, 34, 42] or local [24] average prototypes from the backbone features of support images, and then employ these prototypes for guiding the segmentation of query images through feature fusion [23, 24, 34], distance computation [14, 38], or attention mechanisms [45]. However, these methods can only leverage a limited amount of information extracted from a very small number of support images, thus potentially leading to suboptimal results and reduced robustness. To overcome this limitation, [64] uses the more comprehensive word embedding as the general class information to assist in segmentation. Despite some improvements, [64] is still constrained by the limited capabilities of small language models and has not delved deeper into how to better integrate textual information and support image data to achieve more effective guidance. In this paper, we are the first to employ large language models (LLMs) to achieve FSS by using our" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "carefully designed instructions that contain a more effective multimodal guidance. Furthermore, we utilize the LLM to directly produce segmentation results, rather than merely using intermediate features as done in [64]. This offers a brand-new paradigm to FSS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 384 + ], + "type": "text", + "content": "Large Language Models. The advent of large language models (LLMs) such as GPT [2] and Llama [50] has marked the beginning of a new era in artificial intelligence. Thanks to their significantly increased model parameters and training data, these LLMs contain rich prior knowledge and can be efficiently finetuned for specific tasks or application requirements through methods such as prompts [32, 58], adapters [15, 22] and LoRA [16]. Recently, researchers have started exploring visual large language models [25, 29, 48, 53] to establish a unified framework for multimodal data processing, aiming to override the restriction of LLMs being solely applicable to language data. However, none of them are designed for few-shot tasks in computer vision. In this paper, we introduce the first visual LLM framework for handling few-shot segmentation. To achieve this, we draw inspiration from instruction tuning [44] and in-context learning [11, 39], and carefully design a suitable form of instruction and demonstration examples tailored for few-shot segmentation. By doing so, our method can enable the LLM to comprehend image data and perform few-shot segmentation effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 392, + 362, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 362, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 362, + 403 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 406, + 373, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 373, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 373, + 417 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "content": "This paper aims to construct an LLM-based framework for few-shot segmentation, i.e., to segment a query image " + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "inline_equation", + "content": "I_{q}" + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "content": " support images " + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\{I_s^k\\}_{k=1}^K" + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "content": " and their ground truth maps " + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\{G_s^k\\}_{k=1}^K" + }, + { + "bbox": [ + 304, + 419, + 547, + 685 + ], + "type": "text", + "content": ". As shown in Fig.1, the overall framework of LLaFS can be divided into three key components: (1) a feature extractor that extracts image features and generates visual tokens; (2) a task-tailored instruction that combines visual tokens, target categories, and task requirements to provide task-related information and support guidance; and (3) an LLM that predicts segmentation masks based on the input instruction, followed by a refinement network to optimize the results. For the feature extractor, we adopt the approach in Blip2 [25] by using an image encoder followed by a Q-former and a fully-connected layer to generate a set of visual tokens. We use ResNet50 as the image encoder and keep it frozen during training. For the instruction, we carefully design it as the combination of two parts: segmentation task instruction (Sec.3.2.1) and fine-grained in-context instruction (Sec.3.2.2) to provide comprehensive and detailed guidance. For the LLM, we employ CodeLlama [46] with 7 billion parameters that have been finetuned through instruction tuning. Note that compared to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 715 + ], + "type": "text", + "content": "For simplify of illustration, we introduce LLaFS under the one-shot setting. Supp presents how to extend LLaFS to the multi-shot setting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3066" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 545, + 210 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 545, + 210 + ], + "type": "image", + "image_path": "47b2fbf3693f0e2b944c5c88d2b7cc7561dfaaaf6316cf6f2be3da4b3c20c274.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 546, + 262 + ], + "type": "text", + "content": "Figure 1. Overview of LLaFS. The image encoder and Q-former extract image features and generate a set of visual tokens. Subsequently, a segmentation task instruction and fine-grained in-context introduction are introduced to provide detailed and comprehensive information. These two instructions are integrated and fed into the LLM to produce the vertices coordinates of polygons that enclose the target object. The segmentation mask represented by this polygon is processed by a refinement network to get the final result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 274, + 288, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 288, + 357 + ], + "type": "text", + "content": "vanilla Llama, we empirically find that CodeLlama fine-tuned with code generation datasets exhibits higher accuracy and stability in generating structured information like the segmentation result in our task. We equip CodeLlama with LoRA for fine-tuning. All these components work together within the LLaFS framework to achieve high-quality few-shot segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 363, + 122, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 122, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 122, + 374 + ], + "type": "text", + "content": "3.2. Instruction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 497 + ], + "type": "text", + "content": "As the input of LLM, the instruction is the most crucial component in our framework that makes LLM possible to handle few-shot segmentation. To provide comprehensive information, we design two instructions, namely segmentation task instruction and fine-grained in-context instruction, to respectively provide the LLM with detailed task definitions and fine-grained multi-modal guidance. These two instructions are integrated to formulate the complete instruction as shown in Fig.1. In the following Sec.3.2.1 and Sec.3.2.2, we introduce these two instructions in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 506, + 211, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 211, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 211, + 518 + ], + "type": "text", + "content": "3.2.1 Segmentation Task Instruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 715 + ], + "type": "text", + "content": "The LLMs trained on massive text contents have gained strong reasoning capabilities and a vast amount of world knowledge. Language instructions have shown to be a powerful tool for leveraging these knowledge and capability to handle complex tasks [44]. To achieve better results, the instructions need to be sufficiently clear and detailed, whereas those using only simple terminologies such as 'segment an image' are evidently too abstract for LLMs to comprehend. Thus, we design a structured instruction to explicitly provide more task details such as the expected input and output formats of few-shot segmentation. Specifically, in our instruction, we follow [53] by representing the pixel-wise segmentation output as a 16-point polygon that encloses the target object [30]. Note that it is hard for LLMs to directly generate pixel-wise image masks due to LLM's limited number of output tokens. Our alternative solution of generating" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 274, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 297 + ], + "type": "text", + "content": "polygon vertices provide a token-efficient method for using LLMs to achieve pixel-level segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 298, + 546, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 546, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 546, + 406 + ], + "type": "text", + "content": "Furthermore, training solely on text contents makes LLMs difficult to comprehend visual information precisely, especially in our few-shot image segmentation task, where the number of available training images is very scarce. For this, inspired by the success of in-context learning in NLP [11, 39], we further propose to encode the support image and its ground truth as a visual demonstration example, using it as an intuitive reference in the instruction to teach LLM how to segment a particular class within an image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 406, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 538 + ], + "type": "text", + "content": "By incorporating these designs, we write our segmentation task instruction as: For each object within the class [class] in an image, output coordinates of a 16-point polygon that encloses the object. These points should be arranged in a clockwise direction. The output should be a tuple in the format of (c1, c2, ..., cn), where cn is the coordinates for the n-th object and its format should be ((x1,y1),(x2,y2),..., (x16,y16)). The coordinate value should be within [image size]. For example, for image [support image], the output should be [support ground truth]. Here, [support image] is the visual tokens from the support image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 544, + 491, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 491, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 491, + 556 + ], + "type": "text", + "content": "3.2.2 Fine-grained In-context Instruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Motivation. The above task instruction makes segmenting a class possible by leveraging LLM's knowledge of the class. In the instruction, the class to be segmented is indicated by the [class] token, which is typically a single noun. However, considering that LLMs have never been trained on images, it is challenging for them to directly align this abstract noun with an image region that may possess a complex internal structure. To address this issue, we drew inspiration from human brains and found that when classifying an unseen new class, the human cognitive system follows a mechanism of 'from general to detailed, from abstract to concrete' [41, 59]. Specifically, given an unseen class represented by a general noun, the human brain first decomposes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3067" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 275, + 178 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 275, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 275, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 275, + 178 + ], + "type": "image", + "image_path": "b79500931e4182d7d54ef6fbd3c91e2c0663e0a6239da670778f8bceff57b9d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 185, + 287, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 185, + 287, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 287, + 208 + ], + "type": "text", + "content": "Figure 2. Illustration of how to construct the region-attribute corresponding table used in the fine-grained in-context instruction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 221, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 287, + 304 + ], + "type": "text", + "content": "it into detailed attributes based on the acquired knowledge. For example, in the case of an unseen class 'owl', a person can first gather information from references to learn about the owl's attributes such as 'large round eyes' and 'hooked beak'. Subsequently, it can search the image for concrete regions that match these abstract attributes to determine the presence of the class." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 287, + 508 + ], + "type": "text", + "content": "Inspired by this, we propose a fine-grained in-context instruction to simulate such a human cognitive mechanism based on the support images. For this, we first use ChatGPT to extract detailed attributes of the target class, then we search for regions in support images that correspond to these attributes and generate a corresponding table accordingly. The obtained attributes and table constitute an in-context instruction that is fed into the LLM, which serves as a demonstration example to guide the LLM on how to recognize an image class in a more human-like and fine-grained manner. This alleviates the challenge that LLM cannot perform segmentation well when only inputted with an abstract class noun. Lastly, we introduce an expert-guide framework that refines the instruction to increase its class representation ability. The following sections explain how to generate and refine this instruction in detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "content": "Attributes Generation We first simulate the step of 'from general to detailed' to extract class attributes. Specifically, as shown in Fig.3(a), we construct a prompt 'What does a [class] look like? Please answer in the format of: A [class] has A, B, C,..., where A, B, and C are noun phrases to describe a [class].', and utilize ChatGPT to extract phrases-based attributes that describe the fine-grained details of this class. These attributes are denoted as [attributes] = " + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}" + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "content": ". For each [att]i, we utilize 'A photo of [att]i' as a prompt to extract an embedding " + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "content": " from the CLIP's text encoder. In this way, we get " + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\{t_i\\}_{i=1}^{N_a}" + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}" + }, + { + "bbox": [ + 47, + 515, + 287, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "Region-attribute Corresponding Table. After that, we simulate the second step of 'from abstract to concrete'. To implement this, as shown in Fig.2, we propose a region-attribute corresponding table to find the alignment between support image regions and class attributes. For this, we first" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 533, + 227 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 533, + 227 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 533, + 227 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 533, + 227 + ], + "type": "image", + "image_path": "100c4de25f54d47908bea7ae65a5f2951da3fcf61868c74417a9dc89d6c6bf78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 269 + ], + "type": "text", + "content": "Figure 3. Examples of using ChatGPT for (a) class attributes generation, (b) ambiguity detection and (c) discriminative attributes generation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": "divide the support foreground into multiple local regions. Specifically, for each object in the support image within the target class, we employ the method in selective search [51] to generate a set of superpixels " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": " with different scales in an unsupervised manner. Each " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": " aggregates pixels that are close in position and similar in features, so it can represent a local region with a specific semantic meaning. Based on each " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": ", a masked image is generated and passed through CLIP's² image encoder to produce a feature " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": ". We calculate the cosine similarity between " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": " and the embedding " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": " for each attribute " + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "inline_equation", + "content": "[\\mathrm{att}]_j" + }, + { + "bbox": [ + 304, + 284, + 545, + 429 + ], + "type": "text", + "content": ", and utilize a thresholding process to establish region-attribute correspondence. Formally," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 437, + 545, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 545, + 458 + ], + "type": "interline_equation", + "content": "\\left[ \\operatorname {c o r} \\right] _ {i} = \\left[ \\left[ \\operatorname {a t t} \\right] _ {j} \\text {f o r} j \\in \\left[ 1, N _ {a} \\right] \\text {i f} \\cos \\left(f _ {i}, t _ {j}\\right) > \\alpha \\right] \\tag {1}", + "image_path": "061f8dac7bf7333551912b3f0b8966aa494af385e7da789595669f8b856f4e2e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\cos" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": " refers to cosine similarity, " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": " is a pre-defined threshold. The obtained " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "[\\mathrm{cor}]_i" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": " contains attributes that align with " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": ". In this way, we get " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 305, + 464, + 545, + 525 + ], + "type": "text", + "content": ", which serves as an attribute-region corresponding table that can provide the fine-grained multi-modal reference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": "Instruction Construction. We integrate the obtained class attributes " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{att}]_i\\}_{i=1}^{N_a}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " and corresponding table " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": ", and write the fine-grained in-context instruction as: The [class] has [attributes]. For example, in [support image], " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{coord}]_{s_1}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{cor}]_1" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{coord}]_{s_2}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{cor}]_2\\ldots" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{coord}]_{s_{N_r}}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{cor}]_{N_r}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "[\\mathrm{coord}]_{s_i}" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " is the coordinates of " + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 530, + 545, + 686 + ], + "type": "text", + "content": " in the format of a 16-point polygon. By using this instruction, we provide the LLM with a reference about the attributes of the target class and their corresponding regions in the support image. In this way, a demonstration example can be created that simulates how the human cognitive mechanism recognizes the support image as a class. Using this example as a reference, the LLM can be taught how to understand" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 544, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 544, + 713 + ], + "type": "text", + "content": "2We use CLIP fintuned through the method in [26] to ensure high-quality region-text alignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3068" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 264, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 264, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 264, + 84 + ], + "type": "text", + "content": "and segment an image class in a fine-grained manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "content": "The above-mentioned instruction constructed by the obtained attributes " + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}" + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "content": " and table " + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}" + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "content": " can be directly input into LLM for guidance. However, due to variations in camera angles and instances of occlusion, not every attribute can be directly matched to a region within the support image. Thus, the obtained table " + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i = 1}^{N_r}" + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "content": " may contain only a subset of attributes within " + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{att}]_i\\}_{i = 1}^{N_a}" + }, + { + "bbox": [ + 47, + 91, + 287, + 368 + ], + "type": "text", + "content": ". Unfortunately, we find the combinations of these partial attributes may be insufficient for the unambiguous recognition of the target class. For example, the combination 'wheels, windows, doors' can refer to 'train', 'car', and 'bus' interchangeably. Due to the ambiguous table, the instruction may be misleading. To alleviate this issue, we propose an expert-guide framework to refine the instruction. In this framework, we first employ ChatGPT to identify ambiguous classes of the existing table, then we extract additional attributes that can distinguish the target class from these ambiguous classes for refinement. As long as these additional attributes can be aligned with local regions in the support image, the refined table based on them will become unambiguous. In this way, the class representation ability and comprehensiveness of the instruction can be improved." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "content": "Specifically, this framework generates a refined instruction through the following three steps: 1) Ambiguity Detection. As shown in Fig.3(b), We employ ChatGPT to identify potential ambiguous classes in the obtained table " + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "content": ". Specifically, we denote the attributes contained in " + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{cor}]_i\\}_{i=1}^{N_r}" + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "content": " as [partial-attributes] and ask ChatGPT 'Except for [class], which classes also have [partial-attributes]?' In this way, we obtain a set of ambiguous classes denoted as [a-classes] = " + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{a - class}]_i\\}_{i=1}^{N_{ac}}" + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "content": " from ChatGPT's feedback. 2) Discriminative Attributes Generation. As shown in Fig.3(c), To avoid being misled by these ambiguous classes, we use 'What does [class] look different from [a-classes]?' as a text prompt, enabling ChatGPT to generate attributes that are more discriminative from these classes. The obtained attributes " + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "inline_equation", + "content": "\\{[\\mathrm{d - att}]_i\\}_{i=1}^{N_d}" + }, + { + "bbox": [ + 47, + 369, + 287, + 608 + ], + "type": "text", + "content": " are added to [attributes] for updating. 3) Table and Instruction Refinement. Finally, we use these updated attributes to reperform Eq.1 to obtain a refined table. The updated attributes and table are reassembled to form a refined fine-grained in-context instruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": "We found that a single execution of the three steps already resolves ambiguities in over " + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": " of the instructions. For the remaining " + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": " of instructions, we observe that because the newly-acquired discriminative attributes still couldn't find matching regions in the support image, the resulting table after refinement remains to be ambiguous." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "text", + "content": "Therefore, we iteratively apply the last two steps of this framework until the ambiguity is eradicated. To achieve this, from the second iteration onwards, we replace the text prompt in the discriminative attributes generation step with 'Apart from [all-discriminative-attributes], tell me more differences in appearance between [class] and [a-classes]\", where [all-discriminative-attributes] refers to the discriminative attributes " + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "[\\mathrm{d - att}]_i" + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "text", + "content": " obtained from all previous iterations. By doing so, we enable the iterative framework to continuously discover more discriminative attributes and verify whether they have matched regions in the support image. Eventually, when the framework successfully discovers discriminative attributes " + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "[\\mathrm{d - att}]_i" + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "text", + "content": " that can be aligned with the support image or reaches the maximum number of iterations, we terminate the iteration. For efficiency, we set the maximum number of iterations to 3, in which " + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 275 + ], + "type": "text", + "content": " of the ambiguities have been entirely eradicated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 277, + 443, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 443, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 443, + 289 + ], + "type": "text", + "content": "3.3. Segmentation Prediction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 290, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 410 + ], + "type": "text", + "content": "We integrate segmentation task instruction and fine-grained in-context instruction to formulate the complete instruction as shown in Fig.1. With this instruction as input, the LLM can predict the coordinates of a 16-point polygon that surrounds the target object. Finally, to rectify the imprecision caused by the polygon representation of object edges, a refinement network comprising a pixel decoder and a mask transformer is introduced to generate a refined segmentation mask by using the polygon as the initial mask. Please see Supp for the detailed structures of this network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 413, + 542, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 542, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 542, + 426 + ], + "type": "text", + "content": "3.4. Curriculum Pretraining with Pseudo Samples" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 545, + 576 + ], + "type": "text", + "content": "Motivation. After carefully designing the model structure and instruction format, the next challenge is how to train LLaFS effectively to achieve high segmentation performance. Previous work [29] has shown that the success of instruction tuning often relies on extensive training data. However, due to the difficulty of obtaining pixel-annotated labels, the segmentation datasets used for training typically contain only an insufficient number of images. To address this issue, we propose to generate pseudo support-query pairs and use them to pretrain the LLM. The LLM's ability to handle few-shot segmentation can thus be enhanced by seeing more visual samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Pseudo Sample Generation. Specifically, we propose a method to generate pseudo support-query pairs with the following three steps: 1) Pseudo foreground-background partition. We first useBezier curves to randomly generate a contour within an image region. The area surrounded by this contour is considered as the foreground within the target class, while the regions outside the contour are treated as the background. 2) Noise filling for pseudo support generation. We fill the foreground with Gaussian noise that has a random mean value. For background, we first randomly divide it into multiple subregions to" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "3Due to space limitations, we omit the description of format control prompts for inputting into ChatGPT. See Supp for details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3069" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 270, + 248 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 270, + 248 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 270, + 248 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 270, + 248 + ], + "type": "image", + "image_path": "6362bb87ffc0ff2894f271444ec470e5bedf155a7c122faf5e9ce400cb48f4dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 345 + ], + "type": "text", + "content": "Figure 4. Examples of pseudo samples generated at different pretraining stages. Foreground regions are marked by white contours. As pretraining progresses, pseudo images have reduced intra-image foreground-background differences and greater support-query foreground differences. Meanwhile, the number of polygon vertex coordinates provided in the instruction decreases, while the predicted vertex count increases. These changes gradually increase the pretraining difficulty. (Best viewed in color)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 358, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 514 + ], + "type": "text", + "content": "simulate the complex backgrounds in real images, then we fill each one with a random Gaussian noise that has a mean value distinct from the foreground's noise. The obtained image is used as the support image. 3) Pseudo query generation. We use a similar approach to generate a query image. Note that in this process, the contour and the mean value of foreground noise are no longer completely randomly determined but adjusted based on those used for generating the support image. This is done to ensure that the foreground regions of both support and query images have similar contour shapes and internal features, so they can reflect the same category. Please refer to Supp for the adjustment details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "Curriculum Pretraining. The synthetic support-query pairs can be directly used for pretraining. However, it is observed that this approach will result in a slow convergence rate. A reason for this problem is that the LLM has not been previously trained on image data, so optimizing it to directly grasp a complex image processing task is challenging. To address this issue, we propose a progressive pretraining approach inspired by the success of curriculum learning [54], in which we initiate the model's pretraining with a simple task and gradually increase the task's difficulty until it ultimately reaches the requirements of segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Specifically, during pretraining, we incrementally raise the task's difficulty from the following two aspects: 1) Image understanding. During pretraining, by controlling the difference between mean values of different filled noise, we gradually increase the difference in foreground between" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": "support and query, while reducing the internal difference between foreground and background within each image. This makes it more challenging for LLM to perform few-shot guidance and partition foreground-background areas as pretraining progresses. 2) Polygon generation. We observe that generating a polygon represented by a combination of vertex coordinates is another challenge for the LLM. Therefore, we adopt a progressive strategy here as well. Instead of training the model to directly predict the coordinates of all 16 points of the polygon, we randomly provide the coordinates of " + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": " points in the instruction and let the LLM to predict the coordinates of the remaining " + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "inline_equation", + "content": "16 - M" + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": " points. During pretraining, we gradually decrease the value of " + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": " from 15 to 0. This means that the model receives fewer hints and is required to predict more vertex coordinates as pretraining progresses. Consequently, the pretraining difficulty gradually increases, ultimately reaching the task of predicting all 16 points for segmentation. Experimental results show that this curriculum learning approach allows the model to converge better and achieve higher results. Please see Supp for more technical details on how we increase the difficulty in image understanding and polygon generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 371 + ], + "type": "text", + "content": "Ultimately, the model is trained on the realistic few-shot segmentation dataset after completing the aforementioned pretraining process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 371, + 387, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 371, + 387, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 371, + 387, + 384 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 384, + 542, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 542, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 542, + 397 + ], + "type": "text", + "content": "4.1. Implementation Details and Training Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "text", + "content": "We set the number of queries in the Q-former to 100 and the threshold " + }, + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "text", + "content": " in Eq.1 to 0.2. The ground truth of polygon vertices is obtained in polar coordinates [60]. Specifically, starting from the object center, 16 rays are uniformly emitted at equal angular intervals " + }, + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "inline_equation", + "content": "\\triangle \\theta = 22.5^{\\circ}" + }, + { + "bbox": [ + 304, + 397, + 545, + 504 + ], + "type": "text", + "content": ". The points of intersection between these rays and the object contour are taken as the ground truth of the polygon vertices. More implementation details about pseudo sample generation and curriculum pretraining are presented in Supp." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 505, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 658 + ], + "type": "text", + "content": "The overall model is trained in three stages. In the first stage, we freeze the LLM, pretrain the Q-former and fully-connected layers for 100K steps using the datasets4 and methods in Blip2 [25]. In the second stage, we freeze the Q-former, equip the LLM with LoRA, and pretrain the fully-connected layers, LLM and refinement network using the pseudo-sample-based curriculum learning method for 60k steps. In the third stage, we train the fully-connected layers, LLM and refinement network on the realistic few-shot segmentation dataset for 25 epochs. AdamW [36] is used as the optimizer with the cosine annealing schedule and an initial learning rate of 0.0002. The model is trained on 16 A100 GPUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 661, + 489, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 661, + 489, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 661, + 489, + 673 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 673, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 696 + ], + "type": "text", + "content": "Table.1 presents the comparisons with other few-shot segmentation methods on two datasets: PASCAL-5 and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 544, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 544, + 713 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 315, + 702, + 544, + 713 + ], + "type": "text", + "content": " COCO is excluded from the pretraining set to avoid test data leakage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3070" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 372, + 281 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 372, + 281 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 372, + 281 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 372, + 281 + ], + "type": "table", + "html": "
DatasetMethodConference1-shot5-shot
Fold-0Fold-1Fold-2Fold-3MeanFold-0Fold-1Fold-2Fold-3Mean
PASCAL-5iNTRENet [33]CVPR'2265.472.359.459.863.266.272.861.762.265.7
BAM[23]CVPR'2269.073.667.561.167.870.675.170.867.270.9
AAFormer[57]ECCV'2269.173.359.159.265.272.574.762.061.367.6
SSP[8]ECCV'2260.567.866.451.061.467.572.375.262.169.3
IPMT[34]NeurIPS'2272.873.759.261.666.873.174.761.663.468.2
ABCNet[56]CVPR'2368.873.462.359.566.071.774.265.467.069.6
HDMNet [45]CVPR'2371.075.468.962.169.471.376.271.368.571.8
MIANet[64]CVPR'2368.575.867.563.268.770.277.470.068.871.7
MSI[40]ICCV'2371.072.563.865.968.573.074.270.566.671.1
SCGAN[62]ICCV'2368.372.566.859.866.872.374.169.165.670.3
LLaFSCVPR'2474.278.872.368.573.575.980.175.870.775.6
COCO-20iNTRENet[33]CVPR'2236.842.639.937.939.338.244.140.438.440.3
BAM[23]CVPR'2243.450.647.543.446.249.354.251.649.651.2
SSP[8]ECCV'2235.539.637.936.747.440.647.045.143.944.1
AAFormer[57]ECCV'2239.844.640.641.441.642.950.145.549.246.9
MM-Former[68]NeurIPS'2240.547.745.243.344.244.052.447.450.048.4
IPMT[34]NeurIPS'2241.445.145.640.043.043.549.748.747.947.5
ABCNet[56]CVPR'2342.346.246.042.044.145.551.752.646.449.1
HDMNet [45]CVPR'2343.855.351.649.450.050.661.655.756.056.0
MIANet[64]CVPR'2342.553.047.847.447.745.858.251.351.951.7
MSI[40]ICCV'2342.449.249.446.146.847.154.954.151.952.0
SCGAN[62]ICCV'2340.449.749.645.646.347.257.259.252.153.9
LLaFSCVPR'2447.558.856.253.053.953.263.863.160.060.0
", + "image_path": "743435235baa446f9d551d64821660a4ead0a5ed4b3d075bf569ff84369b3297.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 312, + 287, + 382 + ], + "blocks": [ + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "lines": [ + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "type": "text", + "content": "Table 1. Performance comparison with other methods on PASCAL-5 " + }, + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "type": "inline_equation", + "content": "{}^{i}" + }, + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "type": "text", + "content": " and COCO-20 " + }, + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "type": "inline_equation", + "content": "{}^{i}" + }, + { + "bbox": [ + 53, + 282, + 366, + 293 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 312, + 287, + 382 + ], + "lines": [ + { + "bbox": [ + 52, + 312, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 52, + 312, + 287, + 382 + ], + "type": "table", + "html": "
MethodmIoU
LLaFS74.2
LLaFs w/o segmentation task instruction w/ abstract summary67.7
LLaFS w/o fine-grained in-context instruction67.0
LLaFS w/o refinement network69.1
LLaFS w/o pseudo-sample-based curriculum pretraining63.5
", + "image_path": "f6b3edd0b94ec2a418c9f5bbfb1288de5837e9fcc68994ff7a3f4d7021b687e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO - 20^i}" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": ". Following previous papers, we use different folds as the test set, with the remaining folds utilized for training. This approach yields four sets of experimental results along with their mean result. Across all datasets and experimental settings, our method consistently outperforms others and achieves a remarkably significant improvement compared to previous state-of-the-art results. For instance, on PASCAL- " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": ", LLaFS improves mIoUs by " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": " in the 1-shot and 5-shot settings respectively compared to the second-ranking method. Notably, our approach still exhibits great advantages on the more complex and challenging " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{COCO - 20^i}" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": " dataset, with increases of " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 46, + 410, + 288, + 638 + ], + "type": "text", + "content": " in the 1-shot and 5-shot settings respectively. This could be attributed to the rich prior knowledge embedded in the LLM and our carefully designed instructions, which enable the models to handle complex images effectively and robustly. These results demonstrate the outstanding performance of our method and highlight the huge potentiality of using LLMs for tackling few-shot segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 640, + 141, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 640, + 141, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 640, + 141, + 653 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "text", + "content": "Using the 1-shot setting of the PASCAL- " + }, + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "inline_equation", + "content": "5^{i}" + }, + { + "bbox": [ + 46, + 653, + 287, + 701 + ], + "type": "text", + "content": " dataset, we perform several ablation studies to evaluate different components and designs in our method. More ablation studies are presented in Supp." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Effectiveness of Different Components We validate the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 376, + 70, + 545, + 109 + ], + "blocks": [ + { + "bbox": [ + 55, + 385, + 278, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 278, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 278, + 396 + ], + "type": "text", + "content": "Table 4. Effectiveness of different components in our LLaFS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 70, + 545, + 109 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 545, + 109 + ], + "type": "table", + "html": "
MethodmIoU
LLaFS74.2
LLaFS w/o support images56.9
", + "image_path": "da9a85ad5396fe959e021a4535dc4af85d89e716db33cf4ef1b21502ebf7a811.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 376, + 126, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 383, + 111, + 534, + 121 + ], + "lines": [ + { + "bbox": [ + 383, + 111, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 383, + 111, + 534, + 121 + ], + "type": "text", + "content": "Table 2. Effectiveness of support images." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 126, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 376, + 126, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 376, + 126, + 545, + 188 + ], + "type": "table", + "html": "
MethodmIoU
LLaFS74.2
LLaFS w/o class attributes70.7
LLaFS w/o region-attribute corresponding table69.7
LLaFS w/o instruction refinement70.6
LLaFS w/o iterative refinement71.8
", + "image_path": "e545b27ea634abb0e3008deb706416d606955d0db6caacf23f7b3a2e156d0738.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 371, + 190, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 371, + 190, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 371, + 190, + 547, + 210 + ], + "type": "text", + "content": "Table 3. Ablation results of fine-grained in-context instruction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 372, + 216, + 538, + 282 + ], + "blocks": [ + { + "bbox": [ + 372, + 216, + 538, + 282 + ], + "lines": [ + { + "bbox": [ + 372, + 216, + 538, + 282 + ], + "spans": [ + { + "bbox": [ + 372, + 216, + 538, + 282 + ], + "type": "image", + "image_path": "51b6b9223321d782271e26111de467c4967e991122fb189dee5ab5fa071e5618.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 282, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 370, + 282, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 370, + 282, + 547, + 302 + ], + "type": "text", + "content": "Figure 5. Ablation for the number of polygons' sides." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "content": "effectiveness of key components in our method, including (1) the segmentation task instruction, (2) fine-grained in-context instruction, (3) the refinement network, and (4) the pseudo-sample-based curriculum pretraining. Experimental results are presented in Table.4. Replacing the detailed segmentation task introduction with an abstract summary 'perform image segmentation' decreases mIoU by " + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "content": ". Not using the other components individually decreases the mIoU by " + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 304, + 314, + 545, + 445 + ], + "type": "text", + "content": ", respectively. These results demonstrate the importance and effectiveness of each component in our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": "Effectiveness of Support Images. In the instruction, the annotated support images provide the LLM with crucial visual guidance. As shown in Table.2, if we do not use the support image as a demonstration example, mIoU decreases by " + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "inline_equation", + "content": "17.3\\%" + }, + { + "bbox": [ + 304, + 448, + 545, + 555 + ], + "type": "text", + "content": ". This highlights the importance of the support image as a few-shot sample in our approach, demonstrating that our LLsFS benefits not solely from LLM's prior knowledge in an open-vocabulary manner but indeed gains further improvement from the provided few-shot samples." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Number of Polygon's Sides. In the segmentation task instruction, we represent the segmentation output mask as a region enclosed by a 16-point polygon. We find that the number " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " of sides in the polygon is also a factor affecting the model's performance. Therefore, we test the relationship between mIoU and " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " and present the results in Fig.5. We observe that when " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " is small, the model's performance is suboptimal. This is because polygons with a smaller number of sides cannot accurately describe object edges. As " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " increases, the mIoU gradually improves. However, when " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " exceeds 16, we observe a slight decrease in performance. This could be because a larger " + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": " increases the task's complexity for LLM to tackle. Based on the results," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3071" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 264, + 122 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 264, + 122 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 264, + 122 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 264, + 122 + ], + "type": "table", + "html": "
MethodmIoU
LLaFS74.2
LLaFS w/o pseudo samples63.5
LLaFS w/o curriculum strategy67.3
LLaFS w/ random pseudo query generation63.9
", + "image_path": "9c2c537ef0f7a6ca47dc5f5cfbef22fc92cfde92007a7586d696929378532644.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 286, + 135 + ], + "lines": [ + { + "bbox": [ + 47, + 124, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 286, + 135 + ], + "type": "text", + "content": "Table 5. Ablation of pseudo-sample-based curriculum pretraining." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 142, + 164, + 225 + ], + "blocks": [ + { + "bbox": [ + 70, + 142, + 164, + 225 + ], + "lines": [ + { + "bbox": [ + 70, + 142, + 164, + 225 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 164, + 225 + ], + "type": "image", + "image_path": "e4795b51d371e5f37cb40f7e7248b85fcec20bd966c0815f6fb35e2252174158.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 227, + 150, + 233 + ], + "lines": [ + { + "bbox": [ + 91, + 227, + 150, + 233 + ], + "spans": [ + { + "bbox": [ + 91, + 227, + 150, + 233 + ], + "type": "text", + "content": "(a) Pretraining Loss Curve" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 142, + 263, + 226 + ], + "blocks": [ + { + "bbox": [ + 165, + 142, + 263, + 226 + ], + "lines": [ + { + "bbox": [ + 165, + 142, + 263, + 226 + ], + "spans": [ + { + "bbox": [ + 165, + 142, + 263, + 226 + ], + "type": "image", + "image_path": "9107691691716bccf12180c480cc0827b9bae0fe420c73e9aadbff5eb0e3cf9f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 227, + 247, + 233 + ], + "lines": [ + { + "bbox": [ + 194, + 227, + 247, + 233 + ], + "spans": [ + { + "bbox": [ + 194, + 227, + 247, + 233 + ], + "type": "text", + "content": "(b) Training Loss Curve" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 232, + 286, + 265 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 265 + ], + "type": "text", + "content": "Figure 6. Pretraining (a) and training (b) loss curves in different settings. Curriculum pretraining results in the best convergence in both pretraining and training stages. (Best viewed in color)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 278, + 181, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 181, + 289 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 181, + 289 + ], + "type": "text", + "content": "we chose " + }, + { + "bbox": [ + 47, + 278, + 181, + 289 + ], + "type": "inline_equation", + "content": "M = 16" + }, + { + "bbox": [ + 47, + 278, + 181, + 289 + ], + "type": "text", + "content": " as our setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "content": "Ablation of Fine-grained In-context Instruction. Table.3 presents the evaluation of different components and designs in our fine-grained in-context instruction, including (1) class attributes, (2) attribute-region corresponding table, (3) expert-guide framework for instruction refinement, and (4) iterative refinement. We observe that when these components are removed, mIoU decreases by " + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 290, + 286, + 410 + ], + "type": "text", + "content": ", respectively. These results demonstrate the rationality of our designs in this instruction and their effectiveness in improving performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "content": "Ablation of Pseudo-sample-based Curriculum Pretraining. We further evaluate the key techniques in our pseudosample-based curriculum pretraining mechanism and the results are presented in Table.5. (1) When we do not employ pseudo-samples for pretraining, mIoU decreases by " + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "content": ". (2) Removing the curriculum training strategy that gradually increases the training task difficulty reduces mIoU by " + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "content": ". (3) When generating pseudo support-query samples, to ensure that the support and query can reflect the same category, the contour and the mean value of foreground noise used to generate the query image are adjusted based on those used for generating the support image. When this strategy is not employed and random generation is used instead, mIoU decreases by " + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 46, + 411, + 286, + 590 + ], + "type": "text", + "content": ". These results demonstrate the effectiveness of our method's designs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 593, + 127, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 127, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 127, + 604 + ], + "type": "text", + "content": "4.4. Loss Curves" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "In Fig.6, we present the loss curves during the pretraining and training stages. We observe from Fig.6(a) that without the use of curriculum learning, the pretraining task becomes excessively challenging, which causes the model optimization to quickly reach a bottleneck with difficulties in convergence. After using our curriculum learning mechanism, the model achieves significantly better convergence. Furthermore, in Fig.6(b), we compare the loss reduction during the training stage when using curriculum pretrain" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 330, + 71, + 519, + 189 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 519, + 189 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 519, + 189 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 519, + 189 + ], + "type": "image", + "image_path": "6d625049e27dab562014779b6c62fa5532e20347b2eeb4deb42bd1624a8f9d69.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 188, + 511, + 198 + ], + "lines": [ + { + "bbox": [ + 340, + 188, + 511, + 198 + ], + "spans": [ + { + "bbox": [ + 340, + 188, + 511, + 198 + ], + "type": "text", + "content": "Figure 7. Visualization of segmentation results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 214, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 545, + 286 + ], + "type": "text", + "content": "ing, non-curriculum pretraining, and not using pretraining. The model without pretraining converges the slowest, while the model with curriculum pretraining converges the fastest during the training stage. These results demonstrate the effectiveness of our curriculum-learning-based pretraining approach in enhancing the model's convergence speed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 292, + 389, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 389, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 389, + 303 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 437 + ], + "type": "text", + "content": "In Fig.7, we present visual samples of the segmentation results from LLaFS, including the image, ground truth, the LLM output, and the results after refinement. It can be observed that the polygons output by LLM already achieve good segmentation performance, and the results after refinement are more accurate, particularly in the object's edge regions. It is worth noting that, as shown in the third row of Fig.7, our method still achieves high-performance segmentation when there are more than one target object in the image. These results demonstrate the excellent performance of LLaFS." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "This paper proposes LLaFS, a novel framework that for the first time, leverages large language models (LLMs) to address few-shot segmentation in an end-to-end manner. To enable LLMs to perform this visual task, we introduce a segmentation task instruction to provide detailed task definitions, and propose a fine-grained in-context instruction to simulate human cognitive mechanisms and provide refined multimodal reference information. We also propose a pseudo-sample-based curriculum pretraining mechanism to augment the training samples required for instruction tuning. Our extensive experiments demonstrate the effectiveness of LLaFS, which achieves significantly superior state-of-the-art results across multiple datasets and settings. We consider LLaFS as an important exploration towards an LLM framework capable of addressing few-shot tasks in different modalities beyond natural language processing. Acknowledgement This research is supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD-2021-08-006), MOE AcRF Tier 2 projects MOE-T2EP20222-0009 and MOE-T2EP20123-0014." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3072" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Malik Boudiaf, Hoel Kervadec, Ziko Imtiaz Masud, Pablo Piantanida, Ismail Ben Ayed, and Jose Dolz. Few-shot segmentation without meta-learning: A good transductive inference is all you need? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13979-13988, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 156, + 288, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 156, + 288, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 156, + 288, + 211 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 211, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 211, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 211, + 288, + 252 + ], + "type": "text", + "content": "[3] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 253, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 288, + 317 + ], + "type": "text", + "content": "[4] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 3367-3375, October 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 371 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 372, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 372, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 372, + 288, + 414 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 288, + 479 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 53, + 479, + 288, + 521 + ], + "type": "text", + "content": "[8] Qi Fan, Wenjie Pei, Yu-Wing Tai, and Chi-Keung Tang. Self-support few-shot semantic segmentation. In European Conference on Computer Vision, pages 701-719. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 522, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 522, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 522, + 288, + 543 + ], + "type": "text", + "content": "[9] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 287, + 597 + ], + "type": "text", + "content": "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1-11. IEEE, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 597, + 287, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 597, + 287, + 650 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 287, + 650 + ], + "type": "text", + "content": "[11] Shivam Garg, Dimitris Tsipras, Percy S Liang, and Gregory Valiant. What can transformers learn in-context? a case study of simple function classes. Advances in Neural Information Processing Systems, 35:30583-30598, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 48, + 651, + 287, + 693 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 693, + 287, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 693, + 287, + 704 + ], + "spans": [ + { + "bbox": [ + 48, + 693, + 287, + 704 + ], + "type": "text", + "content": "[13] Shuting He, Henghui Ding, and Wei Jiang. Primitive gener" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "ation and semantic-related alignment for universal zero-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11238-11247, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 169 + ], + "type": "text", + "content": "[14] Sunghwan Hong, Seokju Cho, Jisu Nam, Stephen Lin, and Seungryong Kim. Cost aggregation with 4d convolutional swim transformer for few-shot segmentation. In European Conference on Computer Vision, pages 108-126. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 170, + 545, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 545, + 224 + ], + "type": "text", + "content": "[15] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 224, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 224, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 545, + 267 + ], + "type": "text", + "content": "[16] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 267, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 267, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 545, + 320 + ], + "type": "text", + "content": "[17] Kai Huang, Feigege Wang, Ye Xi, and Yutao Gao. Prototypical kernel learning and open-set foreground perception for generalized few-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19256-19265, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 374 + ], + "type": "text", + "content": "[18] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 374, + 545, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 374, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 308, + 374, + 545, + 429 + ], + "type": "text", + "content": "[19] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 429, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 429, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 308, + 429, + 545, + 482 + ], + "type": "text", + "content": "[20] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[21] Siyu Jiao, Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, and Humphrey Shi. Mask matching transformer for few-shot segmentation. In Advances in Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 567 + ], + "type": "text", + "content": "[22] Rabeeh Karimi Mahabadi, James Henderson, and Sebastian Ruder. Compacter: Efficient low-rank hypercomplex adapter layers. Advances in Neural Information Processing Systems, 34:1022-1035, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 621 + ], + "type": "text", + "content": "[23] Chunbo Lang, Gong Cheng, Binfei Tu, and Junwei Han. Learning what not to segment: A new perspective on few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8057-8067, 2022. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 622, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 622, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 308, + 622, + 545, + 675 + ], + "type": "text", + "content": "[24] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8334-8343, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. Inter-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3073" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 706 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 266, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 266, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 266, + 83 + ], + "type": "text", + "content": "national Conference on Machine Learning, 2023. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 83, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 83, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 287, + 148 + ], + "type": "text", + "content": "[26] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 148, + 287, + 179 + ], + "type": "text", + "content": "[27] Binghao Liu, Jianbin Jiao, and Qixiang Ye. Harmonic feature activation for few-shot semantic segmentation. IEEE Transactions on Image Processing, 30:3142-3153, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 179, + 288, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 179, + 288, + 234 + ], + "spans": [ + { + "bbox": [ + 48, + 179, + 288, + 234 + ], + "type": "text", + "content": "[28] Fang Liu, Yuhao Liu, Yuqiu Kong, Ke Xu, Lihe Zhang, Baocai Yin, Gerhard Hancke, and Rynson Lau. Referring image segmentation using text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22124-22134, October 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 234, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 287, + 255 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 319 + ], + "type": "text", + "content": "[30] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 363 + ], + "type": "text", + "content": "[31] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. Crnet: Cross-reference networks for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4173, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 363, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 287, + 426 + ], + "type": "text", + "content": "[32] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 61-68, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 482 + ], + "type": "text", + "content": "[33] Yuanwei Liu, Nian Liu, Qinglong Cao, Xiwen Yao, Junwei Han, and Ling Shao. Learning non-target knowledge for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11573-11582, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 524 + ], + "type": "text", + "content": "[34] Yuanwei Liu, Nian Liu, Xiwen Yao, and Junwei Han. Intermediate prototype mining transformer for few-shot semantic segmentation. Advances in Neural Information Processing Systems, 35:38020-38031, 2022. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 287, + 567 + ], + "type": "text", + "content": "[35] Yongfei Liu, Xiangyi Zhang, Songyang Zhang, and Xuming He. Part-aware prototype network for few-shot semantic segmentation. In European Conference on Computer Vision, pages 142-158. Springer, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 599, + 287, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 641 + ], + "type": "text", + "content": "[37] Chaofan Ma, Yuhuan Yang, Chen Ju, Fei Zhang, Ya Zhang, and Yanfeng Wang. Open-vocabulary semantic segmentation via attribute decomposition-aggregation. arXiv preprint arXiv:2309.00096, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 641, + 287, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 641, + 287, + 684 + ], + "spans": [ + { + "bbox": [ + 48, + 641, + 287, + 684 + ], + "type": "text", + "content": "[38] Juhong Min, Dahiyun Kang, and Minsu Cho. Hypercorrelation squeeze for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6941-6952, 2021. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 684, + 287, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 684, + 287, + 706 + ], + "spans": [ + { + "bbox": [ + 48, + 684, + 287, + 706 + ], + "type": "text", + "content": "[39] Sewon Min, Xinxi Lyu, Ari Holtzman, Mikel Artetxe, Mike Lewis, Hannaneh Hajishirzi, and Luke Zettlemoyer. Re" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "thinking the role of demonstrations: What makes in-context learning work? In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11048-11064, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 116, + 545, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 116, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 307, + 116, + 545, + 181 + ], + "type": "text", + "content": "[40] Seonghyeon Moon, Samuel S Sohn, Honglu Zhou, Sejong Yoon, Vladimir Pavlovic, Muhammad Haris Khan, and Mubbasir Kapadia. Msi: Maximize support-set information for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19266-19276, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 181, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 181, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 181, + 545, + 213 + ], + "type": "text", + "content": "[41] Gregory L Murphy and Douglas L Medin. The role of theories in conceptual coherence. Psychological review, 92(3):289, 1985. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 256 + ], + "type": "text", + "content": "[42] Khoi Nguyen and Sinisa Todorovic. Feature weighting and boosting for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 622-631, 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 256, + 545, + 289 + ], + "type": "text", + "content": "[43] Atsuro Okazawa. Interclass prototype relation for few-shot segmentation. In European Conference on Computer Vision, pages 362-378. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 289, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 319 + ], + "type": "text", + "content": "[44] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[45] Bohao Peng, Zhuotao Tian, Xiaoyang Wu, Chengyao Wang, Shu Liu, Jingyong Su, and Jiaya Jia. Hierarchical dense correlation distillation for few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23641-23651, 2023. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 373, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 373, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 373, + 545, + 426 + ], + "type": "text", + "content": "[46] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 470 + ], + "type": "text", + "content": "[47] Mennatullah Siam, Boris N Oreshkin, and Martin Jagersand. Amp: Adaptive masked proxies for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5249-5258, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 470, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 470, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 470, + 545, + 502 + ], + "type": "text", + "content": "[48] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all. arXiv preprint arXiv:2305.16355, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 555 + ], + "type": "text", + "content": "[49] Zhuotao Tian, Hengshuang Zhao, Michelle Shu, Zhicheng Yang, Ruiyu Li, and Jiaya Jia. Prior guided feature enrichment network for few-shot segmentation. IEEE transactions on pattern analysis and machine intelligence, 44(2):1050-1065, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[50] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 609, + 545, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 545, + 652 + ], + "type": "text", + "content": "[51] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104:154-171, 2013. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 307, + 652, + 545, + 707 + ], + "type": "text", + "content": "[52] Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9197-9206, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3074" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 707 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[53] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 159 + ], + "type": "text", + "content": "[54] Xin Wang, Yudong Chen, and Wenwu Zhu. A survey on curriculum learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4555-4576, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 288, + 213 + ], + "type": "text", + "content": "[55] Yan Wang, Jian Cheng, Yixin Chen, Shuai Shao, Lanyun Zhu, Zhenzhou Wu, Tao Liu, and Haogang Zhu. Fvp: Fourier visual prompting for source-free unsupervised domain adaptation of medical image segmentation. arXiv preprint arXiv:2304.13672, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 213, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 213, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 288, + 255 + ], + "type": "text", + "content": "[56] Yuan Wang, Rui Sun, and Tianzhu Zhang. Rethinking the correlation in few-shot segmentation: A buoys view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7183-7192, 2023. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 288, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 310 + ], + "type": "text", + "content": "[57] Yuan Wang, Rui Sun, Zhe Zhang, and Tianzhu Zhang. Adaptive agent transformer for few-shot segmentation. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 36-52. Springer, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 288, + 354 + ], + "type": "text", + "content": "[58] Zhen Wang, Rameswar Panda, Leonid Karlinsky, Rogerio Feris, Huan Sun, and Yoon Kim. Multitask prompt tuning enables parameter-efficient transfer learning. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 384 + ], + "type": "text", + "content": "[59] Edward J Wisniewski and Bradley C Love. Relations versus properties in conceptual combination. Journal of memory and language, 38(2):177-202, 1998. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 384, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 288, + 438 + ], + "type": "text", + "content": "[60] Enze Xie, Peize Sun, Xiaoge Song, Wenhai Wang, Xuebo Liu, Ding Liang, Chunhua Shen, and Ping Luo. Polarmask: Single shot instance segmentation with polar representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12193-12202, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 438, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 438, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 438, + 288, + 491 + ], + "type": "text", + "content": "[61] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 534 + ], + "type": "text", + "content": "[62] Qianxiong Xu, Wenting Zhao, Guosheng Lin, and Cheng Long. Self-calibrated cross attention network for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "type": "text", + "content": "[63] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. Mining latent classes for few-shot segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8721-8730, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 632 + ], + "type": "text", + "content": "[64] Yong Yang, Qiong Chen, Yuan Feng, and Tianlin Huang. Mianet: Aggregating unbiased instance and general information for few-shot semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7131-7140, 2023. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 288, + 685 + ], + "type": "text", + "content": "[65] Ying Zang, Chenglong Fu, Runlong Cao, Didi Zhu, Min Zhang, Wenjun Hu, Lanyun Zhu, and Tianrun Chen. Resmatch: Referring expression segmentation in a semi-supervised manner. arXiv preprint arXiv:2402.05589, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 685, + 288, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 685, + 288, + 707 + ], + "spans": [ + { + "bbox": [ + 48, + 685, + 288, + 707 + ], + "type": "text", + "content": "[66] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid graph networks with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 491 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "connection attentions for region-based one-shot semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9587-9595, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 117, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 117, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 117, + 545, + 171 + ], + "type": "text", + "content": "[67] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. Canet: Class-agnostic segmentation networks with iterative refinement and attentive few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5217–5226, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 171, + 545, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 171, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 545, + 213 + ], + "type": "text", + "content": "[68] Gengwei Zhang, Shant Navasardyan, Ling Chen, Yao Zhao, Yunchao Wei, Honghui Shi, et al. Mask matching transformer for few-shot segmentation. Advances in Neural Information Processing Systems, 35:823-836, 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 213, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 213, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 213, + 545, + 255 + ], + "type": "text", + "content": "[69] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 255, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 299 + ], + "type": "text", + "content": "[70] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 299, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 353 + ], + "type": "text", + "content": "[71] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Continual semantic segmentation with automatic memory sample selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3082-3092, June 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 396 + ], + "type": "text", + "content": "[72] Lanyun Zhu, Tianrun Chen, Jianxiong Yin, Simon See, and Jun Liu. Learning gabor texture features for fine-grained recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1621-1631, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 396, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 437 + ], + "type": "text", + "content": "[73] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "type": "text", + "content": "[74] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3075" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_content_list.json b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f1c75aa881e7418c03db9b69a8137cfa3869c84 --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_content_list.json @@ -0,0 +1,2228 @@ +[ + { + "type": "text", + "text": "LLaMA-Excitor: General Instruction Tuning via Indirect Feature Interaction", + "text_level": 1, + "bbox": [ + 93, + 130, + 875, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Zou *", + "bbox": [ + 209, + 166, + 289, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University Beijing, China", + "bbox": [ + 166, + 184, + 331, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zoub21@mails.tsinghua.edu.cn", + "bbox": [ + 125, + 222, + 374, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Yang *", + "bbox": [ + 459, + 166, + 566, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai AI Laboratory Shanghai, China", + "bbox": [ + 416, + 184, + 609, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yangchao@pjlab.org.cn", + "bbox": [ + 418, + 222, + 607, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qiao", + "bbox": [ + 714, + 166, + 782, + 183 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai AI Laboratory Shanghai, China", + "bbox": [ + 651, + 184, + 844, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qiaoyu@pjlab.org.cn", + "bbox": [ + 663, + 222, + 833, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengbin Quan", + "bbox": [ + 267, + 244, + 395, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University", + "bbox": [ + 251, + 262, + 411, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing, China", + "bbox": [ + 272, + 280, + 390, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quancb@tsinghua.edu.cn", + "bbox": [ + 233, + 300, + 429, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youjian Zhao†", + "bbox": [ + 550, + 244, + 674, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University", + "bbox": [ + 532, + 262, + 694, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongguancun Laboratory", + "bbox": [ + 506, + 280, + 718, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing, China", + "bbox": [ + 555, + 297, + 671, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhaoyoujian@tsinghua.edu.cn", + "bbox": [ + 491, + 316, + 730, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 366, + 313, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods to fine-tune LLMs, like Adapter, Prefixtuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation. We designed the Excitor block as a bypass module that reconstructs Keys and changes the importance of Values in self-attention using learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pretrained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our approach is evaluated in language-only and multimodal scenarios. Compared with the original LLaMA-7B, LLaMA-Excitor is the only PEFT method that maintains basic capabilities and achieves $+3.12\\%$ relative improvement on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance on MSCOCO (157.5 CIDEr), and a comparable performance on ScienceQA (88.39%) to cutting-edge models with more parameters and extensive vision-language pertaining. The code will be available at https://zoubo9034.github.io/Excitor/.", + "bbox": [ + 76, + 396, + 472, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07402a80ce6dd75de8090ddb762a1cfbd08ce9fd3ddd8fd1015c4467934bcfc4.jpg", + "image_caption": [ + "Figure 1. Overview of LLaMA-Excitor. We integrate Excitor blocks into L out of N attention layers of LLaMA. Differing from previous PEFT techniques, LLaMA-Excitor indirectly involves learnable information in the reasoning process by changing the similarity matrices. It ensures that the hidden states are within the original distribution of LLaMA." + ], + "image_footnote": [], + "bbox": [ + 535, + 367, + 870, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 782, + 632, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) are highly advanced tools that enable contextual information processing, understanding, and generation. They serve as powerful knowledge bases that can be used to gain valuable insights and create new content [9, 10, 37, 39, 52]. Fine-tuning them for downstream tasks has also achieved great success in vari-", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, † Corresponding author \nThis work was done during an internship at Shanghai AI Lab.", + "bbox": [ + 99, + 875, + 429, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14089", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ous domains. With the rapid growth in scale and quality of the corpus and the advance in hardware, the inherent abilities of LLMs are becoming stronger. Therefore, the primary objective of fine-tuning general-purpose LLMs should be ensuring they produce the desired output for downstream tasks (i.e., instruction-following ability) rather than memory excessive knowledge or mastering skills (e.g., causal inference and numerical computation).", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prioritizing the enhancement of instruction-following capabilities for LLMs is prudent for a multitude of reasons. Firstly, it is often presumed that the corpus utilized during training encompasses a comprehensive repository of knowledge, thereby obviating the need for additional fine-tuning datasets when addressing specific downstream tasks. This implies that enriching a model's ability to follow instructions could leverage this pre-existing knowledge base more effectively. Secondly, by standardizing the output format during instruction-following tasks, we can reduce the generation of harmful or irrelevant content. Third, the challenge of incorporating new knowledge and skills into LLMs is accentuated by the limited set of parameters available for fine-tuning. Lastly, fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual.", + "bbox": [ + 75, + 214, + 468, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, predominant parameter-efficient fine-tuning (PEFT) techniques are met with notable challenges in performing instruction-tuning. The Adapter methods[14, 36], which have a serial structure, can cause computation bottlenecks, and they also significantly alter the inference process to adapt to downstream tasks, which in turn can lead to degradation of inherent abilities, such as catastrophic forgetting. Prompt learning methods[5, 26, 29, 41] effectively incorporate new information by creating additional parameters corresponding to the input sequences of language models. However, concentrating all necessary knowledge into fixed-length token sequences is challenging for various downstream tasks. Furthermore, as the length of the generated sequence increases, the control over the sequence diminishes since LLMs generate the next word based on a softmax over the entire sequence. LoRA methods[15, 50] directly add the outcomes of randomly initialized low-rank modules with linear layers' outputs, which can introduce features that are out of LLMs' feature distribution and may cause degradations. Recent studies [16, 50] also show that Adapter and LoRA evenly assign learnable parameters to trainable modules and neglect the functionality gap between each layer in LLM, which will lead to performance drops.", + "bbox": [ + 75, + 489, + 468, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we start from a new perspective and propose LLaMA-Excitor, a PEFT method that focuses on the following instructions. (1) LLaMA-Excitor aims to optimize instruction-following ability by releasing the poten", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tial of an LLM, i.e., LLaMA [46], instead of pursuing new knowledge and skills. Due to variations in the data scale, quality, and content coverage of instruction-tuning sets, the effect of fine-tuning processes on the inherent skills of LLMs is unpredictable. (2) LLaMA-Excitor can reduce the degradation when fine-tuning on unsuited datasets. Specifically, rather than directly changing the intermediate hidden state of pretrained LLMs like Adapter and LoRA, LLaMA-Excitor uses trainable bypass modules (Excitor blocks) to modify the attention score meanwhile maintain original Values for each attention layer in LLaMA. Excitor blocks only adjust the proportion of information in Values for hidden states, ensuring that each attention layer's input is derived from a linear combination of the previous layer's Values and corresponds to the original LLaMA's distribution (an indirect feature interaction process). Additionally, the modification of the attention score depends on both the input sequence and a set of prompt tokens to make the model adaptively decide whether to rely on additional information from the fine-tuning set for attention allocation. These attributes ensure fast training and few forgetings.", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, Exictor innovatively provides a low-budget way to fine-tune a language-only model into a vision-language model. Previous works (e.g., [3, 28, 31, 47, 51]) rely on training an additional visual branch or projection module to align vision and language, which incurs significant computational overhead. We explore the unified modeling of instruction-tuning and visual instruction-tuning for LLMs and utilize a visual Exictor to make LLaMA follow the visual prompts generated from a frozen image encoder like CLIP [38]. In this way, we eliminate the need for extra vision-language alignment while preserving textual features' purity during LLMs' reasoning and shortening the training time. Our contributions are summarized as follows:", + "bbox": [ + 496, + 409, + 890, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We study indirect feature interaction in fine-tuning LLMs and propose LLaMA-Excitor, a PEFT method that focuses on instruction-following and reduces forgetting.", + "- We uniformly models multi-modal and language-only tuning and extends language models into powerful vision-language models in a low-budget way.", + "- Experiments on instruction-following datasets, multi-task evaluation, image captioning, and VQA demonstrate the feasibility and advancement of LLaMA-Excitor." + ], + "bbox": [ + 500, + 607, + 890, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 758, + 648, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Parameter-Efficient Fine-Tuning", + "text_level": 1, + "bbox": [ + 500, + 785, + 785, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PEFT techniques enable efficient adaptation of large language models to various downstream applications without fine-tuning all the model's parameters. Adapter-based approaches [14, 36] insert additional feed-forward networks within the existing architecture. However, these methods can increase the computational load due to their se", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14090", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ffa15b86d79452037231e99e355b92a299fd36f1aaa37bdd379543c92b29c46.jpg", + "image_caption": [ + "Figure 2. Details of the Excitor block. We assign a set of learnable prompts for attention layers of LLaMA. These prompts are used to construct an extra Key for computing additional similarity scores, which are then merged into the original scores to alter the LLM's behavior. Cold-start gating factors are designed to stabilize the training." + ], + "image_footnote": [], + "bbox": [ + 171, + 92, + 803, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rial nature and may induce catastrophic forgetting by significantly changing the reasoning mechanism of LLMs. Prompt-learning methods offer a training-free approach to task adaptation [5, 41] and are improved by prefix-tuning [26, 29], which involves learnable token sequences. The primary drawback is the difficulty in capturing complex knowledge within a limited number of tokens, which can be restrictive for diverse tasks. Additionally, as the number of learnable tokens grows, these methods may lose control over the sequence output quality. LoRA [15, 50] aims to address these issues by introducing low-rank matrices in a parallel structure, enhancing both performance and efficiency. However, it risks incorporating features that deviate from the original model's distribution, potentially leading to performance degradation. Furthermore, equal distribution of parameter updates across layers disregards the distinct functionalities of different layers [16, 50], leading to suboptimal utilization of the model's capacity.", + "bbox": [ + 75, + 363, + 472, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Visual Instruction Tuning", + "text_level": 1, + "bbox": [ + 76, + 660, + 312, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instruction tuning[33, 48] finetunes or adapted language models to follow specific textual instructions. Despite the prominent success in the NLP domain, its inherent text-only nature limits its ability to comprehend and interact with visual content effectively. LLaVA [28] enables language models to follow visual instructions and engage with multi-modal information. And a series of approaches have been developed to improve the visual instruction tuning with advanced architectures [2, 4, 22] or more versatile functionalities[20, 35, 54]. Fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual.", + "bbox": [ + 75, + 688, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 361, + 635, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary goal of Excitor is to effectively leverage LLMs' pre-trained knowledge and skills to solve downstream tasks without excessive modification of the reasoning process, which can lead to forgetting and performance degradation. We choose LLaMA[45] as our LLM as its effectiveness has been demonstrated in several open-source instruction-tuning works [8, 28, 34, 44, 51]. In Section 3.1, we first introduce how to insert Excitor blocks into LLaMA's attention layers. Then, in Section 3.2, we present the details of Excitor blocks' architecture. Finally, in section 3.3, we generalize LLaMA-Excitor for multi-modal reasoning.", + "bbox": [ + 496, + 392, + 893, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Indirect feature interaction", + "text_level": 1, + "bbox": [ + 500, + 597, + 746, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given 52K instruction-following data [44] and a pre-trained LLaMA with $N$ attention layers, we insert our proposed Excitor blocks into the topmost $L(L < N)$ layers for instruction-following fine-tuning. This can better tune the language representations with higher-level semantics. Taking the $l$ -th inserted layer as an example, we denote the $M$ -length word tokens as $T_{l} \\in \\mathbb{R}^{M \\times C}$ , where $C$ equals the feature dimension of LLaMA's attention layers. As shown in Figure 1, the Excitor block takes $T_{l}$ with a set of learnable prompts $\\{P_{l}\\}_{l=1}^{L}$ as inputs, where $P_{l} \\in \\mathbb{R}^{K \\times C}$ with $K$ denoting the length of learnable prompts for each layer, generates an extra similarity matrix $S_{l}^{\\mathrm{extra}} \\in \\mathbb{R}^{M \\times M}$ with exactly the same shape of original similarity matrix $S_{l}$ for the $l$ -th layer. In the self-attention mechanism, $S_{l}$ is calculated from Query and Key, and controls the importance of components in Value, where Query, Key, and Value are linear projected version of $T_{l}$ :", + "bbox": [ + 496, + 625, + 893, + 862 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ u e y = W _ {\\mathrm {q}} \\left(T _ {l}\\right), K e y = W _ {\\mathrm {k}} \\left(T _ {l}\\right), V a l u e = W _ {\\mathrm {v}} \\left(T _ {l}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14091", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{q}}$ , $W_{\\mathrm{k}}$ , $W_{\\mathrm{v}}$ are the pre-trained LLMs' weights that are frozen during the finetuning. Then, $S_{l}$ is formulated as:", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} = \\frac {\\text {Q u e r y} \\cdot \\text {K e y} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 128, + 468, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $S_{l}$ rules how the information updates, we treat $S_{l}^{extra}$ as the residual of $S_{l}$ to modify the reasoning process and formulate the updated output of the $l$ -th layer $T_{l}^{\\mathrm{o}}$ as:", + "bbox": [ + 76, + 169, + 468, + 214 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} ^ {g} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x t r a}} \\times g _ {l} + S _ {l}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 222, + 468, + 239 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {l} ^ {\\mathrm {o}} = W _ {\\mathrm {o}} \\left(S _ {l} ^ {g} \\cdot \\text {V a l u e}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 250, + 468, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{l}$ is a set of learnable parameters, named cold-start gating factor, which is inspired by zero-init attention [51], to stabilize training by controlling the proportion of $S_{l}^{extra}$ participants in the reasoning. Instead of applying zero-initialization like [51], we initialize $g_{l}$ with $\\mathbf{N}\\left(0,10^{-2}\\right)$ to overcome the gradient vanishing in the mixed precision (FP16) training.", + "bbox": [ + 76, + 268, + 468, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As stated in sec 1, previous techniques updating the output of frozen LLMs' layers by adding the outcomes of trainable modules onto $T_{l}^{\\mathrm{o}}$ (e.g., Adapter[14] and LoRA[15]) or concatenating trainable tokens with $T_{l}$ as input for layer $l$ (e.g., Prefix-tuning [26]) can be attributed to the direct modifying of the intermediate representations. One drawback of direct modifying is adding out-of-distribution features into the intermediate representations. It will largely change the inputs of higher layers and make the framework unstable, especially at the beginning of training. By contrast, LLaMA-Excitor only uses trainable tokens to change the attention score $S_{l}^{g}$ . The output of each layer $T_{l}^{\\mathrm{o}}$ is still the linear combination of Value and within the original LLMs' distribution. i.e., LLaMA-Excitor alters the likelihood of potential outcomes generated by pre-trained LLMs. This indirect feature interaction ensures the model doesn't deviate too far from its pretrained reasoning process and can more effectively arouse pretrained skills in downstream usage.", + "bbox": [ + 75, + 375, + 468, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Excitor Blocks", + "text_level": 1, + "bbox": [ + 76, + 656, + 225, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To impose effective control over the attention score $S_{l}^{g}$ while reducing the degradation of LLM's inherent abilities, LLaMA-Excitor must be self-adaptive to the original input sequence $T_{l}$ to decide whether learnable parameters should influence the reasoning. Thus, Excitor Blocks needs to generate the extra similarity matrix $S_{l}^{\\mathrm{extra}}$ considering both injected information in the learnable prompt $P_{l}$ and $T_{l}$ . We formulate $S_{l}^{\\mathrm{extra}}$ as:", + "bbox": [ + 75, + 678, + 468, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} ^ {\\text {e x t r a}} = \\frac {\\operatorname {Q u e r y} \\cdot \\operatorname {K e y} _ {\\text {e x t r a}} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 800, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Query$ is reused from the equation 1 to pass knowledge from $T_{l}$ while reducing the computational overhead, $Key_{extra}$ is an extra attention key that carries information from $P_{l}$ . Since the shape of $S_{l}^{\\text{extra}}$ should correspond", + "bbox": [ + 76, + 840, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/74fb0a4b6acef288cdfeaf7f9720a264c4a6408f1320bcd2cc33a13333f3f354.jpg", + "image_caption": [ + "Figure 3. Extend LLaMA-Excitor into a powerful multi-modal model. Owing to the indirect feature interaction, LLaMA-Excitor is the cheapest PEFT method that can follow visual instructions without complicated projection modules aligning vision and language." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 859, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $S_{l}$ , the shape of $Key_{extra}$ must be the same with $Key \\in \\mathbb{R}^{M \\times C}$ in the equation 1. However, $M$ , which is the length of $Key$ , is changing along with variable input texts, while the length of learnable prompts $P_{l}$ is a pre-defined hyper-parameter.", + "bbox": [ + 498, + 453, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown on the left of the Excitor Block in Figure 2, we introduce the Key reconstruction to mitigate the shape inconsistency. Since the attention mechanism generates an output with the same shape as its queries and transmits information about its values, we define $P_{l}$ as the Exitor Key and the Exitor Value and calculate the Excitor Query from $T_{l}$ for the $l$ -th Excitor Block. We omit linear projections for $P_{l}$ to pass information directly to $Key_{extra}$ and accelerate the training. The Key reconstruction is formulated as:", + "bbox": [ + 498, + 527, + 890, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} ^ {\\text {e x c i t o r}} = \\frac {W _ {\\mathrm {q}} ^ {\\text {e x c i t o r}} \\left(T _ {l}\\right) \\cdot P _ {l} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 660, + 890, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K e y} _ {\\text {e x t r a}} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x c i t o r}}\\right) \\cdot P _ {l}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 698, + 890, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $W_{\\mathrm{q}}^{\\mathrm{excitor}}$ is a low-rank module consisting of two linear layers with dimension $\\mathbb{R}^{C\\times r}$ and $\\mathbb{R}^{r\\times C}$ . Finally, $S_{l}^{\\mathrm{extra}}$ carries information from $T_{l}$ and has the same shape as Key.", + "bbox": [ + 500, + 724, + 890, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-modal Extension", + "text_level": 1, + "bbox": [ + 500, + 780, + 712, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The indirect feature interaction is not limited to textual instructions but is also compatible with following visual prompts. We simply extend our Excitor Block into a multimodal version in Figure 3. Taking an image and a task description as input, LLaMA-Excitor first leverages a frozen pre-trained visual encoder like CLIP [38] to extract a sequence of visual embeddings $I \\in \\mathbb{R}^{V \\times D}$ as the visual", + "bbox": [ + 498, + 801, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14092", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction: Tell me about alpacas.", + "text_level": 1, + "bbox": [ + 83, + 92, + 277, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA7B: Alpacas are a type of camelid that are native to South America. They are related to llamas and vicunas. Alpacas are raised for their fiber, which is used to make clothing.", + "bbox": [ + 81, + 107, + 410, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Fine-Tuning: Alpacas are small, soft-furry, and mammals that are native to South America. They are the largest of the wool-making mammals and are the most gentle and social of the mammals. They are also the most colorful mammal, with a range of colors from white to gray, brown, black, and shades of brown and gray. They are also very smart and can learn tricks easily. They are also very social and live in herds of up to 20-30 individuals.", + "bbox": [ + 81, + 136, + 411, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alpaca-LoRA: Alpacas are a species of South American camelid, related to the llama and the vicuna. They are a domesticated version of the wild vicuna and are primarily raised for their fine fleece, which is used to make clothing. Alpacas are typically smaller than llamas, with a range of sizes from 12 to 15 hands high. They are known for their gentle nature and are often kept as pets.", + "bbox": [ + 81, + 190, + 411, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA-Adapter: Alpacas are members of the camelid family and are native to the Andes Mountains of South America. They are typically between 3 and 4 feet tall and weigh between 100 and 200 pounds. They have long, coarse, and dense coats that come in a variety of colors, including white, black, brown, and grey.", + "bbox": [ + 81, + 236, + 411, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA-Excitor: Alpacas are domesticated animals that are related to llamas and are native to South America. They are typically between 100 and $120~\\mathrm{cm}$ tall and weigh between 100 and $150~\\mathrm{kg}$ . Alpacas have long, soft fleece that is used for making clothing, blankets, and other products. They are also used for their meat and their dung, which is used as fuel. Alpacas are herd animals and are known for their gentle and calm nature. They are also very intelligent and can be trained to perform certain tasks.", + "bbox": [ + 81, + 273, + 411, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction: Write a conversation between the sun and Pluto.", + "text_level": 1, + "bbox": [ + 436, + 92, + 763, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA7B:", + "text_level": 1, + "bbox": [ + 436, + 107, + 486, + 116 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun: Hey Pluto, how are you?", + "bbox": [ + 436, + 116, + 571, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm fine.", + "bbox": [ + 436, + 127, + 498, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun: I'm glad to hear that.", + "bbox": [ + 436, + 137, + 558, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm glad to hear th", + "bbox": [ + 436, + 148, + 535, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun: I'm glad to hear that too.", + "bbox": [ + 436, + 159, + 571, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Fine-Tuning:", + "text_level": 1, + "bbox": [ + 581, + 108, + 658, + 116 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun said, \"I'm so hot!\"", + "bbox": [ + 581, + 116, + 691, + 125 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun said, \"I'm so hot!\"", + "bbox": [ + 581, + 127, + 691, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun said, \"I'm so hot!\"", + "bbox": [ + 581, + 137, + 691, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun said, \"I'm so hot!\"", + "bbox": [ + 581, + 148, + 691, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sun said, \"I'm so hot!\"", + "bbox": [ + 581, + 160, + 691, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alpaca-LoRA:", + "text_level": 1, + "bbox": [ + 705, + 108, + 769, + 116 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: Hey Pluto, how's life in the outer solar system?", + "bbox": [ + 705, + 116, + 885, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: It's pretty chilly up here! I'm still getting used to being so far away from the sun.", + "bbox": [ + 705, + 135, + 885, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: That's understandable. I'm sure you'll get used to it soon enough. Anyway, I'm glad you're still around.", + "bbox": [ + 705, + 155, + 885, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Thanks! It's nice to still be a part of the solar system.", + "bbox": [ + 705, + 181, + 885, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA-Adapter:", + "text_level": 1, + "bbox": [ + 436, + 191, + 513, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Sun: Good morning, Pluto!", + "bbox": [ + 436, + 200, + 563, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Good morning, Sun! How are you?", + "bbox": [ + 436, + 210, + 602, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Sun: I'm doing great! How are you?", + "bbox": [ + 436, + 222, + 598, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm doing well, too!", + "bbox": [ + 436, + 234, + 542, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Sun: I'm glad to hear that!", + "bbox": [ + 436, + 246, + 560, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Me too!", + "bbox": [ + 436, + 258, + 495, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Sun: I'm glad we're having such a nice day.", + "bbox": [ + 436, + 268, + 614, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "day", + "bbox": [ + 436, + 280, + 452, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Me too!", + "bbox": [ + 436, + 289, + 496, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Sun: I'm glad we're having such a nice", + "bbox": [ + 436, + 299, + 614, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "day!", + "bbox": [ + 436, + 308, + 454, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Me too!", + "bbox": [ + 436, + 315, + 495, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLaMA-Excitor:", + "text_level": 1, + "bbox": [ + 617, + 198, + 689, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: Hey Pluto, how are you doing?", + "bbox": [ + 617, + 205, + 761, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm doing great! How about you?", + "bbox": [ + 617, + 213, + 777, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: I'm doing well too. It's been a while since we last spoke.", + "bbox": [ + 617, + 222, + 862, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Yeah, it has been a while. I've been busy exploring the outer reaches of the solar system.", + "bbox": [ + 617, + 231, + 862, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: That's great! I'm glad you're having fun.", + "bbox": [ + 617, + 244, + 799, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm having a blast! I'm so glad I'm not stuck in the inner solar system like the other planets.", + "bbox": [ + 617, + 253, + 862, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: Yeah, it's nice to have some space to yourself.", + "bbox": [ + 617, + 268, + 818, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: Definitely! So, what have you been up to?", + "bbox": [ + 617, + 277, + 810, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: I've been busy keeping the other planets warm and providing light and heat.", + "bbox": [ + 617, + 286, + 862, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: That's a big job! You must be exhausted.", + "bbox": [ + 617, + 301, + 803, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sun: I am, but it's worth it. I'm glad I can help", + "bbox": [ + 617, + 310, + 802, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pluto: I'm glad you're doing well. It's been great talking to you.", + "bbox": [ + 617, + 318, + 864, + 332 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c7b4dd884401a33c80c0683d0f77da961a50c0e9ba819736eeef25c255d2f71.jpg", + "image_caption": [ + "Table 1. Instruction-following Comparison between LLaMA-Excitor and other techniques.", + "Figure 4. Quantitative comparisons between LLaMA-Excitor (BLUE) with other methods, evaluated by GPT-4 [32]." + ], + "image_footnote": [], + "bbox": [ + 122, + 388, + 452, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompts, where $V$ is the embedding length and $D$ is the dimension of embeddings. Note that we adopt both the global CLS token and regional patch tokens from the last layer of the CLIP encoder to acquire adequate visual semantics. Then, we calculate multi-modal versions of the Excitor Key and the Exitor Value by combining $P_l$ and $I$ :", + "bbox": [ + 75, + 547, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K e y} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {k}} ^ {\\text {e x c i t o r}} (I) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 643, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V a l u e} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {v}} ^ {\\text {e x c i t o r}} (I) \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 676, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike the processing of $P_{l}$ , we adopt low-rank linear projections $W_{\\mathrm{k}}^{\\mathrm{excitor}}$ and $W_{\\mathrm{v}}^{\\mathrm{excitor}}$ on $I$ , since we provide the same visual prompt for attention layers $l \\in [N - L, N]$ and different layers require various visual clues for the downstream reasoning. LLaMA-Excitor is an innovative framework for the multi-modal fine-tuning of LLMs. As mentioned in Section 1, most techniques before LLaMA-Excitor involved training a module to perform multi-modal alignment that projects visual embeddings into the textual feature space. This enables LLMs to understand visual tokens and integrate them into the reasoning process. However, the strict alignment constraint may damage the rich semantics of visual embeddings, and their direct feature fusion may", + "bbox": [ + 75, + 704, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also cause performance degradation. In contrast, LLaMA-Excitor does not require feature alignment and works on the raw inputs from the frozen image encoder. It does not change the purity of the intermediate textual features and thus avoids any potential damage to the rich semantics of the visual embeddings and the reasoning of LLMs.", + "bbox": [ + 496, + 388, + 890, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 494, + 632, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Language-only Performances", + "text_level": 1, + "bbox": [ + 500, + 521, + 761, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Instruction-following Evaluations", + "text_level": 1, + "bbox": [ + 500, + 545, + 785, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the instruction-following capacity of LLaMA-Excitor by responding to instructions.", + "bbox": [ + 496, + 570, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We use Stanford Alpaca [44], which contains 52K instruction-following data, as the training set. Each sample in this dataset is formatted according to a template as follows:", + "bbox": [ + 496, + 601, + 890, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.", + "bbox": [ + 504, + 662, + 879, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruction: {instruction}", + "bbox": [ + 506, + 705, + 658, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: {input}", + "bbox": [ + 506, + 718, + 593, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Response: {output}", + "bbox": [ + 506, + 733, + 625, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where {instruction} is the description of a task, {input} is the context for the task, and {output} is the answer generated by GPT-3.5 [5].", + "bbox": [ + 496, + 748, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We develop LLaMA-Excitor based on the original LLaMA codebase with minor modifications. We train LLaMA-Excitor on 8 NVIDIA A100 GPUs for five epochs. The warmup epochs, batch size, learning rate, and weight decay are set to 2, 64, 9e-3, and 0.02, respectively. In this section, we utilize the pretrained LLaMA-7B v1 model, which has $N = 32$ transformer lay", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2e252e041efef300d5769f4527b5379e1d3bc009152682997785967fd1692363.jpg", + "image_caption": [ + "Figure 5. The relative average performance changes and win-loss situations of fine-tunings compared to original LLaMA-7B on MMLU." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 472, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers with feature dimension $C = 4096$ as the backbone. We apply our excitor blocks with the low-rank dimension $r = 16$ for $W_{\\mathrm{q}}^{\\mathrm{exciton}}$ in the topmost $L = 30$ layers and set the length of learnable prompts $L = 30$ . During the inference, we adopt top-p sampling as the default decoding method with a temperature of 0.1 and a top-p = 0.75.", + "bbox": [ + 75, + 335, + 467, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare LLaMA-Excitor with three competing methods. (1) Full-finetuning, which updates the entire 7B parameters in LLaMA during the end-to-end training. (2) Alpaca-LoRA, the official implementation of low-rank adaptation [15] provided by [44]. (3) LLaMA-Adapter [51], a cutting-edge prefix-tuning method with zero-init attention to reduce forgetting in finetuning.", + "bbox": [ + 75, + 428, + 467, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performances. We first use examples in Table 1 to show the strong instruction-following ability of LLaMA-Excitor. The first case requires the models to provide some information about alpacas. LLaMA-Excitor generates reasonable responses comparable to Full-finetuning, demonstrating our method's effectiveness. It's worth noticing that our approach follows the same answering pattern as the prefix-tuning method LLaMA-adapter and the original LLaMA7B (although Excitor provides more details). This phenomenon is reasonable since prefix-tuning methods use additional learnable prompts to guide the generation process of LLMs. They do not change the internal reasoning process, just like our Excitor. This characteristic may also be the foundation of avoiding catastrophic forgetting. For the second task, the LLMs are asked to create a dialogue between the Sun and Pluto. Out of all the models that are tested, only Excitor and LoRA correctly recognize the characters as planets within the Solar System and incorporate their identities into the conversation. However, Excitor stands out as it can generate lengthy, high-quality conversations rich in interesting content. (Please see Appendix A for the full comparison including dialog generation, code generation, and question answering.)", + "bbox": [ + 75, + 536, + 467, + 882 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of instruction-following", + "bbox": [ + 96, + 885, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models quantitatively, we refer to the research conducted by Chiang et al. [8]. In this evaluation, we use GPT-4 [32] to assess the quality of responses to 80 randomly selected questions from AlpacaEval [25]. However, we notice a bias in GPT-4's scoring, where it tends to give higher scores to the first response. To overcome this, we interchange the position of two responses and use the average score for the final rating. Results are shown in Figure 4. We notice that Excitor overwhelms the Full-Finetuning (73/80). The reason behind this is that AlpacaEval is an out-of-distributed (OOD) test for our training set, and full-finetuning may easily overfit the training set. When compared with AlpacaLoRA and LLaMA-Adapter, our Excitor still gains the majority of \"Win\", demonstrating our indirect feature interaction is at least comparable with or better than predominant PEFT methods. (Please refer to Appendix B for instances on AlpacaEval.)", + "bbox": [ + 496, + 90, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As this paper is a preliminary attempt to study indirect feature interaction in LLMs and the 52K instruction-following data used in our study is automatically generated, it is possible to further enhance the performance of LLaMA-Excitor by improving the quality of the training data, adjusting the hyper-parameters to more reasonable values or exploring larger LLaMA models.", + "bbox": [ + 496, + 352, + 890, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2 The Impact of Fine-tuning Techniques on Inherent Abilities", + "text_level": 1, + "bbox": [ + 498, + 503, + 890, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One of the motivations for introducing the indirect feature interaction in LLaMA-Excitor is to reduce the influence of fine-tuning on low-quality or non-targeted datasets. To quantitatively analyze this characteristic, we evaluate the fine-tuned models on MMLU [13], which is currently the most comprehensive benchmark for assessing the multi-task accuracy of text models. This benchmark includes 57 tasks across various fields, such as basic math, American history, computer science, and law. The model must possess extensive world knowledge and problem-solving capabilities to achieve high accuracy on this test. We report the relative performance changes of PEFT methods compared with the original LLaMA-7B and how many of the 57 evaluations PEFT methods outperform LLaMA-7B in Figure 5. We find that for models fine-tuned on Alpaca-52k, only Excitor can avoid reducing the average score of 57 tasks in MMLU, and it even shows an increase in performance $(+3.12\\%)$ . As expected, the full-finetuning performance suffers the most severe decline as its parameters are most significantly changed $(-6.63\\%)$ , while the LLaMA-Adapter, which is based on prefix-tuning, experiences the least decline $(-3.02\\%)$ besides Excitor. (We provide results based on LLaMA2-7B in Appendix C)", + "bbox": [ + 496, + 553, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14094", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image.", + "bbox": [ + 114, + 90, + 647, + 114 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaMA-Excitor: An airplane is flying through the sky, with its landing gear down, as it approaches for a landing at an airport. The airplane is dominating the scene, with its bright red and white colors clearly visible against the backdrop of the cloudy sky. The airplane is about to touch down on the runway, creating a moment of excitement and anticipation for the passengers and crew on board.", + "bbox": [ + 114, + 121, + 647, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Annotation: A large passenger airplane flying through the air.", + "bbox": [ + 116, + 171, + 511, + 184 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image.", + "bbox": [ + 114, + 190, + 647, + 214 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaMA-Excitor: In the image, a group of birds is standing together on a giraffe. One of the bird is leaning over to nibble on the giraffe's fur and the giraffe seems to be enjoying the interaction. The scene captures the playful nature of these animals and the unexpected interaction between a giraffe and a bird.", + "bbox": [ + 114, + 220, + 647, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Annotation: Little birds sitting on the shoulder of a giraffe.", + "bbox": [ + 116, + 268, + 495, + 281 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c43262a5b3ceb4ac8de19634fb9be3a1c902c816d64c997d542cc141adb2fc4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 92, + 849, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90c5cd3fb6434b6de7be91207a3888bcadc5ca0d1b009cf2dfa7eba77ec93057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 195, + 849, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e74975a87005792dcd8c132c798ce8a0a3a24254810576a8454ea2ea543ef900.jpg", + "table_caption": [ + "Table 2. Examples demonstrating LLaMA-Excitor's visual instruction following capacity." + ], + "table_footnote": [], + "table_body": "
MethodData ScaleCOCO Caption
PTFTBLEU@4CIDEr
ClipCap [31]0M0.6M33.5113.1
VL-PET [16]0M0.6M-121.7
LLaMA-AdapterV2 [12]0M0.6M36.2122.2
Qwen-vl-chat [4]1.4B0.6M-131.9
mPLUG-Owl2 [49]348M0.6M-137.3
BLIP [21]14M0.6M40.4136.7
Flamingo [2]1.8B0.6M-138.1
BLIP-2 [22]129M0.6M43.7145.3
LLaMA-Excitor0M0.6M49.7157.5
", + "bbox": [ + 120, + 319, + 426, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with state-of-the-art image captioning methods on COCO Caption.", + "bbox": [ + 75, + 460, + 468, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Multi-Modal Performances", + "text_level": 1, + "bbox": [ + 76, + 497, + 321, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate the visual instruction following the capacity of LLaMA-Excitor by responding to paired vision-language instructions and demonstrate our indirect feature interaction, which unifies the modeling of language-only tuning and multi-modal tuning, providing a powerful low-budget way to perform vision-language tasks.", + "bbox": [ + 75, + 521, + 468, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1 Image Captioning Evaluation", + "text_level": 1, + "bbox": [ + 76, + 633, + 336, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We first evaluate our Excitor on COCO Caption [7], which contains 0.6M training image caption data (120k images with 5 captions per image) over a wide range of distributions. We apply frozen CLIP/14-L [38] as the image encoder, the visual embedding dimension $D = 768$ , low-rank dimension $r = 16$ . We keep other hyper-parameters the same as the section 4.1.1.", + "bbox": [ + 75, + 657, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performances. We compare our Excitor with cutting-edge image captioning methods in Table 3. We find that Excitor achieves a stunning result, significantly surpassing the previous SOTA method BILP-2 [22] by $+6$ BLEU@4 and $+12.2$ CIDEr. Especially considering that we do not apply complex multi-modal alignment modules and have only 0.6M training data. By contrast, BLIP-2 adopts a costly pre-training stage on an additional 129M dataset, including VisualGenome [19], Conceptual Captions [42, 43] and", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LAION [6]. We visualize the results of image captioning in Appendix D. However, the image captions provided in COCO are concise and lack sufficient detail in describing the content of images. To further unleash Exitor's performance in image captioning, we continue fine-tuning Exitor on LLaVA665k [27], a higher-quality dataset with detailed picture descriptions. We provide several examples from the COCO test set in Table 2. It shows that image captions generated by Excitor can accurately cover the content of human annotations and provide richer details. (e.g., (1) The airplane's landing gear is down, (2) A bird is pecking at the back of a giraffe)", + "bbox": [ + 496, + 321, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Performances on ScienceQA", + "text_level": 1, + "bbox": [ + 498, + 527, + 751, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We evaluate our Excitor on ScienceQA [40], which contains 21k multimodal multiple choice questions with rich domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. We train Excitor on ScienceQA train split from scratch, using the combination of Chain-of-Thought(CoT) and direct answer prediction. CoT requires the model to predict the solution first and then generate the answer choice based on the solution. During training, we randomly require the model to first generate solutions or directly predict answers in each iteration. The model is asked to answer the question directly in the evaluation.", + "bbox": [ + 496, + 551, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 reports the performances of the top methods on the official leaderboard of the ScienceQA and Excitor. The current SOTA method LLaVA [28] is pretrained on another 558k visual-language dataset and fine-tuned on ScienceQA. Also, they allow all LLM's parameters to be updated during fine-tuning. Our Excitor demonstrates comparable performance with LLaVA w/o pretraining (only -0.4% performance gap). Note that LLaMA-Excitor is a PEFT method (with a frozen LLM) without CoT, and LLaVA utilizes a larger backbone LLaMA-13B than our LLaMA-7B. To further unlock the potential of Excitor, we introduce a vari", + "bbox": [ + 496, + 734, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14095", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab9354d92b2fc2da656a6f34226e4b4bc207e1496ad8b0d556ad391135fb9a6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAverageSubjectContext ModalityGrade
NATSOCLANTXTIMGNOG1-6G7-12
Human [40]88.4090.2384.9787.4889.6087.5088.1091.5982.42
UnifiedQACoT74.1171.0076.0478.9166.4266.5381.8177.0668.82
GPT-3CoT75.1775.4470.8778.0974.6867.4379.9378.2369.68
ChatGPTCoT [1]78.3178.8270.9883.1877.3767.9286.1380.7274.03
GPT-4CoT [32]83.9985.4872.4490.2782.6571.4992.8986.6679.04
MM-COT [53]84.9187.5277.1785.8287.8882.9086.8384.6585.37
LLaVACoT [28]90.9290.3695.9588.0089.4988.0090.6690.9390.90
LLaVACoT (w/o pretrain) [28]85.81--------
DFAF [11]60.7264.0348.8263.5565.8854.4964.1157.1267.17
ViLT [18]61.1460.4863.8960.2763.2061.3857.0060.7261.90
Patch-TRM [30]61.4265.1946.7965.5566.9655.2864.9558.0467.50
VisualBERT [23, 24]61.8759.3369.1861.1862.7162.1758.5462.9659.92
UnifiedQA [17]70.1268.1669.1874.9163.7861.3877.8472.9865.00
GPT-3 [5]74.0475.0466.5978.0074.2465.7479.5876.3669.87
LLaMA-Adapter85.1984.3788.3084.3683.7280.3286.9085.8384.05
LLaMA-Excitor85.4185.7092.3582.8283.4384.5686.2785.6584.64
LLaMA-Excitor@336px + LoRA88.3987.1991.3387.0990.4285.2088.6488.3588.42
", + "bbox": [ + 119, + 88, + 854, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ant that utilizes a more powerful image encoder (CLIP-L/14@336) and adds LoRA blocks to the original LLM's parameters to reduce the gap with LLaVA's full parameter updating. It brings $+2.58\\%$ performance gain compared with LLaVA.", + "bbox": [ + 75, + 407, + 470, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 76, + 496, + 191, + 511 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9eef5e895ab3913ae47bdfdea947db17c7db5d5c34b377ce7abb6d54df6aa3e2.jpg", + "table_caption": [ + "Table 4. Question Answering Accuracy (%) on ScienceQA's [40] test set. We report GPT-3 [5], ChatGPT [1], and GPT-4 [32] for zero-shot inference. CoT denotes utilizing chain-of-thought for question answering." + ], + "table_footnote": [], + "table_body": "
VariantACC
(a) CLIP-16/B83.11
(b) CLIP-14/L85.41
(c) CLIP-14/L@336px85.87
(d) CLIP-14/L@336px + Full-Finetuning83.20
(e) CLIP-14/L@336px + LoRA88.39
", + "bbox": [ + 119, + 531, + 426, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Analysis of the effectiveness of each module.", + "bbox": [ + 111, + 637, + 433, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate several design choices on ScienceQA in Table 5. (a), (b) and (c) study the influence of the capacity of image encoder. The best CLIP encoder yields $85.41\\%$ and is $+2.76\\%$ higher than the basic version. In (d) and (e), we study to combine Excitor, which focuses on instruction-following, with other techniques that update LLM's reasoning process. We find LoRA can bring $2.98\\%$ performance gain for Excitor, demonstrating the potential of combining Excitor with existing finetuning methods in multimodal usage. (However, in the text-only scenario, we encounter $-4.14\\%$ performance gap compared with the original LLaMA-7B in the MMLU test by combining Excitor and LoRA.) Please see Appendix E for more ablations on the design of low-rank linear layers (i.e., $W_{\\mathrm{q}}^{Excitor}$ , $W_{\\mathrm{k}}^{Excitor}$ , and $W_{\\mathrm{v}}^{Excitor}$ ) in Excitor Blocks under single", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "modal and multi-modal scenarios, the choice of low-rank dimension $r$ , and the number of layers $L$ with Excitor blocks inserted.", + "bbox": [ + 496, + 407, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 467, + 619, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study the indirect feature interaction in fine-tuning LLMs into instruction-following models and propose LLaMA-Exictor, a PEFT method that demonstrates outstanding instruction-following capacity compared with predominant techniques meanwhile reducing the forgetting of LLMs' inherent abilities. Moreover, Excitor also shows its potential in vision-language tasks by unifying the modeling of visual instruction-following and language-only instruction-following. Impressively, Excitor achieves strong multi-modal performance without the need to train complex alignment modules.", + "bbox": [ + 496, + 492, + 890, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project is a work in progress, and some directions can be explored. The first one is the adaptability to LLMs other than LLaMA. The second direction is to explore more effective ways to introduce multi-scale visual prompts (since we only apply the features from the last layer of CLIP at the current stage). Last but not least, Excitor has the potential to fine-tune vision-only models into vision-language models, just like our implementation for LLMs.", + "bbox": [ + 496, + 659, + 893, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 500, + 792, + 617, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Postdoctoral Excellent Program (Grant No. 2022234).", + "bbox": [ + 496, + 816, + 893, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14096", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chatgpt. https://chat.openai.com, 2023.8", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3, 7", + "[3] Anonymous. Videodistill: Language-aware vision distillation for video question answering. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3, 7", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 5, 8", + "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 7", + "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arxiv 2015. arXiv preprint arXiv:1504.00325, 2015. 7", + "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3, 6", + "[9] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1", + "[11] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639-6648, 2019. 8", + "[12] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.7" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6", + "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6", + "[16] Zi-Yuan Hu, Yanyang Li, Michael R Lyu, and Liwei Wang. Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3010-3020, 2023. 2, 3, 7", + "[17] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896-1907, 2020. 8", + "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583-5594, 2021. 8", + "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 7", + "[20] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3", + "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 7", + "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 7", + "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 8", + "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020. 8", + "[25] Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tat-" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sunori B. Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval, 2023.6", + "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4", + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 7, 8", + "[29] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 2, 3", + "[30] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 8", + "[31] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 7", + "[32] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 5, 6, 8", + "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 3", + "[34] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 3", + "[35] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3", + "[36] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2", + "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 7", + "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1", + "[40] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 7, 8", + "[41] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 2, 3", + "[42] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 7", + "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7", + "[44] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 3, 5, 6", + "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971, 2023. 2", + "[47] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2", + "[48] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 3", + "[49] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7", + "[50] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. arXiv preprint arXiv:2303.10512, 2023. 2, 3", + "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2, 3, 4, 6", + "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab," + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14098", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1", + "[53] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.8", + "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 247 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14099", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_model.json b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e6a801520c909c1a018fee0e406279d83071b29 --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_model.json @@ -0,0 +1,2851 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.877, + 0.152 + ], + "angle": 0, + "content": "LLaMA-Excitor: General Instruction Tuning via Indirect Feature Interaction" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.167, + 0.29, + 0.183 + ], + "angle": 0, + "content": "Bo Zou *" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.185, + 0.332, + 0.22 + ], + "angle": 0, + "content": "Tsinghua University Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.223, + 0.375, + 0.236 + ], + "angle": 0, + "content": "zoub21@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.167, + 0.567, + 0.184 + ], + "angle": 0, + "content": "Chao Yang *" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.185, + 0.61, + 0.22 + ], + "angle": 0, + "content": "Shanghai AI Laboratory Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.223, + 0.608, + 0.237 + ], + "angle": 0, + "content": "yangchao@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.167, + 0.784, + 0.184 + ], + "angle": 0, + "content": "Yu Qiao" + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.185, + 0.845, + 0.22 + ], + "angle": 0, + "content": "Shanghai AI Laboratory Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.664, + 0.223, + 0.834, + 0.237 + ], + "angle": 0, + "content": "qiaoyu@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.245, + 0.396, + 0.262 + ], + "angle": 0, + "content": "Chengbin Quan" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.263, + 0.413, + 0.28 + ], + "angle": 0, + "content": "Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.281, + 0.392, + 0.298 + ], + "angle": 0, + "content": "Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.301, + 0.431, + 0.314 + ], + "angle": 0, + "content": "quancb@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.552, + 0.245, + 0.676, + 0.262 + ], + "angle": 0, + "content": "Youjian Zhao†" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.263, + 0.695, + 0.28 + ], + "angle": 0, + "content": "Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.281, + 0.72, + 0.298 + ], + "angle": 0, + "content": "Zhongguancun Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.299, + 0.673, + 0.315 + ], + "angle": 0, + "content": "Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.318, + 0.732, + 0.332 + ], + "angle": 0, + "content": "zhaoyoujian@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.367, + 0.314, + 0.383 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.397, + 0.473, + 0.851 + ], + "angle": 0, + "content": "Existing methods to fine-tune LLMs, like Adapter, Prefixtuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation. We designed the Excitor block as a bypass module that reconstructs Keys and changes the importance of Values in self-attention using learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pretrained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our approach is evaluated in language-only and multimodal scenarios. Compared with the original LLaMA-7B, LLaMA-Excitor is the only PEFT method that maintains basic capabilities and achieves \\(+3.12\\%\\) relative improvement on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance on MSCOCO (157.5 CIDEr), and a comparable performance on ScienceQA (88.39%) to cutting-edge models with more parameters and extensive vision-language pertaining. The code will be available at https://zoubo9034.github.io/Excitor/." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.368, + 0.871, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.758 + ], + "angle": 0, + "content": "Figure 1. Overview of LLaMA-Excitor. We integrate Excitor blocks into L out of N attention layers of LLaMA. Differing from previous PEFT techniques, LLaMA-Excitor indirectly involves learnable information in the reasoning process by changing the similarity matrices. It ensures that the hidden states are within the original distribution of LLaMA." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.784, + 0.633, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Large language models (LLMs) are highly advanced tools that enable contextual information processing, understanding, and generation. They serve as powerful knowledge bases that can be used to gain valuable insights and create new content [9, 10, 37, 39, 52]. Fine-tuning them for downstream tasks has also achieved great success in vari-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.875, + 0.43, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution, † Corresponding author \nThis work was done during an internship at Shanghai AI Lab." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "ous domains. With the rapid growth in scale and quality of the corpus and the advance in hardware, the inherent abilities of LLMs are becoming stronger. Therefore, the primary objective of fine-tuning general-purpose LLMs should be ensuring they produce the desired output for downstream tasks (i.e., instruction-following ability) rather than memory excessive knowledge or mastering skills (e.g., causal inference and numerical computation)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Prioritizing the enhancement of instruction-following capabilities for LLMs is prudent for a multitude of reasons. Firstly, it is often presumed that the corpus utilized during training encompasses a comprehensive repository of knowledge, thereby obviating the need for additional fine-tuning datasets when addressing specific downstream tasks. This implies that enriching a model's ability to follow instructions could leverage this pre-existing knowledge base more effectively. Secondly, by standardizing the output format during instruction-following tasks, we can reduce the generation of harmful or irrelevant content. Third, the challenge of incorporating new knowledge and skills into LLMs is accentuated by the limited set of parameters available for fine-tuning. Lastly, fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.838 + ], + "angle": 0, + "content": "However, predominant parameter-efficient fine-tuning (PEFT) techniques are met with notable challenges in performing instruction-tuning. The Adapter methods[14, 36], which have a serial structure, can cause computation bottlenecks, and they also significantly alter the inference process to adapt to downstream tasks, which in turn can lead to degradation of inherent abilities, such as catastrophic forgetting. Prompt learning methods[5, 26, 29, 41] effectively incorporate new information by creating additional parameters corresponding to the input sequences of language models. However, concentrating all necessary knowledge into fixed-length token sequences is challenging for various downstream tasks. Furthermore, as the length of the generated sequence increases, the control over the sequence diminishes since LLMs generate the next word based on a softmax over the entire sequence. LoRA methods[15, 50] directly add the outcomes of randomly initialized low-rank modules with linear layers' outputs, which can introduce features that are out of LLMs' feature distribution and may cause degradations. Recent studies [16, 50] also show that Adapter and LoRA evenly assign learnable parameters to trainable modules and neglect the functionality gap between each layer in LLM, which will lead to performance drops." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this paper, we start from a new perspective and propose LLaMA-Excitor, a PEFT method that focuses on the following instructions. (1) LLaMA-Excitor aims to optimize instruction-following ability by releasing the poten" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.409 + ], + "angle": 0, + "content": "tial of an LLM, i.e., LLaMA [46], instead of pursuing new knowledge and skills. Due to variations in the data scale, quality, and content coverage of instruction-tuning sets, the effect of fine-tuning processes on the inherent skills of LLMs is unpredictable. (2) LLaMA-Excitor can reduce the degradation when fine-tuning on unsuited datasets. Specifically, rather than directly changing the intermediate hidden state of pretrained LLMs like Adapter and LoRA, LLaMA-Excitor uses trainable bypass modules (Excitor blocks) to modify the attention score meanwhile maintain original Values for each attention layer in LLaMA. Excitor blocks only adjust the proportion of information in Values for hidden states, ensuring that each attention layer's input is derived from a linear combination of the previous layer's Values and corresponds to the original LLaMA's distribution (an indirect feature interaction process). Additionally, the modification of the attention score depends on both the input sequence and a set of prompt tokens to make the model adaptively decide whether to rely on additional information from the fine-tuning set for attention allocation. These attributes ensure fast training and few forgetings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Moreover, Exictor innovatively provides a low-budget way to fine-tune a language-only model into a vision-language model. Previous works (e.g., [3, 28, 31, 47, 51]) rely on training an additional visual branch or projection module to align vision and language, which incurs significant computational overhead. We explore the unified modeling of instruction-tuning and visual instruction-tuning for LLMs and utilize a visual Exictor to make LLaMA follow the visual prompts generated from a frozen image encoder like CLIP [38]. In this way, we eliminate the need for extra vision-language alignment while preserving textual features' purity during LLMs' reasoning and shortening the training time. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.608, + 0.891, + 0.651 + ], + "angle": 0, + "content": "- We study indirect feature interaction in fine-tuning LLMs and propose LLaMA-Excitor, a PEFT method that focuses on instruction-following and reduces forgetting." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.653, + 0.891, + 0.697 + ], + "angle": 0, + "content": "- We uniformly models multi-modal and language-only tuning and extends language models into powerful vision-language models in a low-budget way." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.698, + 0.891, + 0.742 + ], + "angle": 0, + "content": "- Experiments on instruction-following datasets, multi-task evaluation, image captioning, and VQA demonstrate the feasibility and advancement of LLaMA-Excitor." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.608, + 0.891, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.76, + 0.65, + 0.777 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.787, + 0.803 + ], + "angle": 0, + "content": "2.1. Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "PEFT techniques enable efficient adaptation of large language models to various downstream applications without fine-tuning all the model's parameters. Adapter-based approaches [14, 36] insert additional feed-forward networks within the existing architecture. However, these methods can increase the computational load due to their se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14090" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.093, + 0.805, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Details of the Excitor block. We assign a set of learnable prompts for attention layers of LLaMA. These prompts are used to construct an extra Key for computing additional similarity scores, which are then merged into the original scores to alter the LLM's behavior. Cold-start gating factors are designed to stabilize the training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.637 + ], + "angle": 0, + "content": "rial nature and may induce catastrophic forgetting by significantly changing the reasoning mechanism of LLMs. Prompt-learning methods offer a training-free approach to task adaptation [5, 41] and are improved by prefix-tuning [26, 29], which involves learnable token sequences. The primary drawback is the difficulty in capturing complex knowledge within a limited number of tokens, which can be restrictive for diverse tasks. Additionally, as the number of learnable tokens grows, these methods may lose control over the sequence output quality. LoRA [15, 50] aims to address these issues by introducing low-rank matrices in a parallel structure, enhancing both performance and efficiency. However, it risks incorporating features that deviate from the original model's distribution, potentially leading to performance degradation. Furthermore, equal distribution of parameter updates across layers disregards the distinct functionalities of different layers [16, 50], leading to suboptimal utilization of the model's capacity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.313, + 0.679 + ], + "angle": 0, + "content": "2.2. Visual Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Instruction tuning[33, 48] finetunes or adapted language models to follow specific textual instructions. Despite the prominent success in the NLP domain, its inherent text-only nature limits its ability to comprehend and interact with visual content effectively. LLaVA [28] enables language models to follow visual instructions and engage with multi-modal information. And a series of approaches have been developed to improve the visual instruction tuning with advanced architectures [2, 4, 22] or more versatile functionalities[20, 35, 54]. Fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.362, + 0.637, + 0.38 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.895, + 0.576 + ], + "angle": 0, + "content": "The primary goal of Excitor is to effectively leverage LLMs' pre-trained knowledge and skills to solve downstream tasks without excessive modification of the reasoning process, which can lead to forgetting and performance degradation. We choose LLaMA[45] as our LLM as its effectiveness has been demonstrated in several open-source instruction-tuning works [8, 28, 34, 44, 51]. In Section 3.1, we first introduce how to insert Excitor blocks into LLaMA's attention layers. Then, in Section 3.2, we present the details of Excitor blocks' architecture. Finally, in section 3.3, we generalize LLaMA-Excitor for multi-modal reasoning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.598, + 0.747, + 0.614 + ], + "angle": 0, + "content": "3.1. Indirect feature interaction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.895, + 0.863 + ], + "angle": 0, + "content": "Given 52K instruction-following data [44] and a pre-trained LLaMA with \\(N\\) attention layers, we insert our proposed Excitor blocks into the topmost \\(L(L < N)\\) layers for instruction-following fine-tuning. This can better tune the language representations with higher-level semantics. Taking the \\(l\\)-th inserted layer as an example, we denote the \\(M\\)-length word tokens as \\(T_{l} \\in \\mathbb{R}^{M \\times C}\\), where \\(C\\) equals the feature dimension of LLaMA's attention layers. As shown in Figure 1, the Excitor block takes \\(T_{l}\\) with a set of learnable prompts \\(\\{P_{l}\\}_{l=1}^{L}\\) as inputs, where \\(P_{l} \\in \\mathbb{R}^{K \\times C}\\) with \\(K\\) denoting the length of learnable prompts for each layer, generates an extra similarity matrix \\(S_{l}^{\\mathrm{extra}} \\in \\mathbb{R}^{M \\times M}\\) with exactly the same shape of original similarity matrix \\(S_{l}\\) for the \\(l\\)-th layer. In the self-attention mechanism, \\(S_{l}\\) is calculated from Query and Key, and controls the importance of components in Value, where Query, Key, and Value are linear projected version of \\(T_{l}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\nQ u e y = W _ {\\mathrm {q}} \\left(T _ {l}\\right), K e y = W _ {\\mathrm {k}} \\left(T _ {l}\\right), V a l u e = W _ {\\mathrm {v}} \\left(T _ {l}\\right) \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14091" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "where \\( W_{\\mathrm{q}} \\), \\( W_{\\mathrm{k}} \\), \\( W_{\\mathrm{v}} \\) are the pre-trained LLMs' weights that are frozen during the finetuning. Then, \\( S_{l} \\) is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.13, + 0.47, + 0.163 + ], + "angle": 0, + "content": "\\[\nS _ {l} = \\frac {\\text {Q u e r y} \\cdot \\text {K e y} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.17, + 0.469, + 0.215 + ], + "angle": 0, + "content": "Since \\( S_{l} \\) rules how the information updates, we treat \\( S_{l}^{extra} \\) as the residual of \\( S_{l} \\) to modify the reasoning process and formulate the updated output of the \\( l \\)-th layer \\( T_{l}^{\\mathrm{o}} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.223, + 0.469, + 0.241 + ], + "angle": 0, + "content": "\\[\nS _ {l} ^ {g} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x t r a}} \\times g _ {l} + S _ {l}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.25, + 0.469, + 0.265 + ], + "angle": 0, + "content": "\\[\nT _ {l} ^ {\\mathrm {o}} = W _ {\\mathrm {o}} \\left(S _ {l} ^ {g} \\cdot \\text {V a l u e}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.469, + 0.375 + ], + "angle": 0, + "content": "where \\( g_{l} \\) is a set of learnable parameters, named cold-start gating factor, which is inspired by zero-init attention [51], to stabilize training by controlling the proportion of \\( S_{l}^{extra} \\) participants in the reasoning. Instead of applying zero-initialization like [51], we initialize \\( g_{l} \\) with \\( \\mathbf{N}\\left(0,10^{-2}\\right) \\) to overcome the gradient vanishing in the mixed precision (FP16) training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.47, + 0.648 + ], + "angle": 0, + "content": "As stated in sec 1, previous techniques updating the output of frozen LLMs' layers by adding the outcomes of trainable modules onto \\( T_{l}^{\\mathrm{o}} \\) (e.g., Adapter[14] and LoRA[15]) or concatenating trainable tokens with \\( T_{l} \\) as input for layer \\( l \\) (e.g., Prefix-tuning [26]) can be attributed to the direct modifying of the intermediate representations. One drawback of direct modifying is adding out-of-distribution features into the intermediate representations. It will largely change the inputs of higher layers and make the framework unstable, especially at the beginning of training. By contrast, LLaMA-Excitor only uses trainable tokens to change the attention score \\( S_{l}^{g} \\). The output of each layer \\( T_{l}^{\\mathrm{o}} \\) is still the linear combination of Value and within the original LLMs' distribution. i.e., LLaMA-Excitor alters the likelihood of potential outcomes generated by pre-trained LLMs. This indirect feature interaction ensures the model doesn't deviate too far from its pretrained reasoning process and can more effectively arouse pretrained skills in downstream usage." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.657, + 0.227, + 0.671 + ], + "angle": 0, + "content": "3.2. Excitor Blocks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.469, + 0.791 + ], + "angle": 0, + "content": "To impose effective control over the attention score \\( S_{l}^{g} \\) while reducing the degradation of LLM's inherent abilities, LLaMA-Excitor must be self-adaptive to the original input sequence \\( T_{l} \\) to decide whether learnable parameters should influence the reasoning. Thus, Excitor Blocks needs to generate the extra similarity matrix \\( S_{l}^{\\mathrm{extra}} \\) considering both injected information in the learnable prompt \\( P_{l} \\) and \\( T_{l} \\). We formulate \\( S_{l}^{\\mathrm{extra}} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.801, + 0.469, + 0.833 + ], + "angle": 0, + "content": "\\[\nS _ {l} ^ {\\text {e x t r a}} = \\frac {\\operatorname {Q u e r y} \\cdot \\operatorname {K e y} _ {\\text {e x t r a}} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\( Query \\) is reused from the equation 1 to pass knowledge from \\( T_{l} \\) while reducing the computational overhead, \\( Key_{extra} \\) is an extra attention key that carries information from \\( P_{l} \\). Since the shape of \\( S_{l}^{\\text{extra}} \\) should correspond" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.86, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.36, + 0.892, + 0.43 + ], + "angle": 0, + "content": "Figure 3. Extend LLaMA-Excitor into a powerful multi-modal model. Owing to the indirect feature interaction, LLaMA-Excitor is the cheapest PEFT method that can follow visual instructions without complicated projection modules aligning vision and language." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.528 + ], + "angle": 0, + "content": "with \\(S_{l}\\), the shape of \\(Key_{extra}\\) must be the same with \\(Key \\in \\mathbb{R}^{M \\times C}\\) in the equation 1. However, \\(M\\), which is the length of \\(Key\\), is changing along with variable input texts, while the length of learnable prompts \\(P_{l}\\) is a pre-defined hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.529, + 0.892, + 0.655 + ], + "angle": 0, + "content": "As shown on the left of the Excitor Block in Figure 2, we introduce the Key reconstruction to mitigate the shape inconsistency. Since the attention mechanism generates an output with the same shape as its queries and transmits information about its values, we define \\( P_{l} \\) as the Exitor Key and the Exitor Value and calculate the Excitor Query from \\( T_{l} \\) for the \\( l \\)-th Excitor Block. We omit linear projections for \\( P_{l} \\) to pass information directly to \\( Key_{extra} \\) and accelerate the training. The Key reconstruction is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.661, + 0.891, + 0.693 + ], + "angle": 0, + "content": "\\[\nS _ {l} ^ {\\text {e x c i t o r}} = \\frac {W _ {\\mathrm {q}} ^ {\\text {e x c i t o r}} \\left(T _ {l}\\right) \\cdot P _ {l} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.699, + 0.891, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K e y} _ {\\text {e x t r a}} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x c i t o r}}\\right) \\cdot P _ {l}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.726, + 0.891, + 0.773 + ], + "angle": 0, + "content": "Where \\( W_{\\mathrm{q}}^{\\mathrm{excitor}} \\) is a low-rank module consisting of two linear layers with dimension \\( \\mathbb{R}^{C\\times r} \\) and \\( \\mathbb{R}^{r\\times C} \\). Finally, \\( S_{l}^{\\mathrm{extra}} \\) carries information from \\( T_{l} \\) and has the same shape as Key." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.781, + 0.714, + 0.795 + ], + "angle": 0, + "content": "3.3. Multi-modal Extension" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.803, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The indirect feature interaction is not limited to textual instructions but is also compatible with following visual prompts. We simply extend our Excitor Block into a multimodal version in Figure 3. Taking an image and a task description as input, LLaMA-Excitor first leverages a frozen pre-trained visual encoder like CLIP [38] to extract a sequence of visual embeddings \\( I \\in \\mathbb{R}^{V \\times D} \\) as the visual" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14092" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.084, + 0.093, + 0.279, + 0.105 + ], + "angle": 0, + "content": "Instruction: Tell me about alpacas." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.108, + 0.411, + 0.134 + ], + "angle": 0, + "content": "LLaMA7B: Alpacas are a type of camelid that are native to South America. They are related to llamas and vicunas. Alpacas are raised for their fiber, which is used to make clothing." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.137, + 0.412, + 0.188 + ], + "angle": 0, + "content": "Full Fine-Tuning: Alpacas are small, soft-furry, and mammals that are native to South America. They are the largest of the wool-making mammals and are the most gentle and social of the mammals. They are also the most colorful mammal, with a range of colors from white to gray, brown, black, and shades of brown and gray. They are also very smart and can learn tricks easily. They are also very social and live in herds of up to 20-30 individuals." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.191, + 0.412, + 0.234 + ], + "angle": 0, + "content": "Alpaca-LoRA: Alpacas are a species of South American camelid, related to the llama and the vicuna. They are a domesticated version of the wild vicuna and are primarily raised for their fine fleece, which is used to make clothing. Alpacas are typically smaller than llamas, with a range of sizes from 12 to 15 hands high. They are known for their gentle nature and are often kept as pets." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.237, + 0.412, + 0.272 + ], + "angle": 0, + "content": "LLaMA-Adapter: Alpacas are members of the camelid family and are native to the Andes Mountains of South America. They are typically between 3 and 4 feet tall and weigh between 100 and 200 pounds. They have long, coarse, and dense coats that come in a variety of colors, including white, black, brown, and grey." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.275, + 0.412, + 0.333 + ], + "angle": 0, + "content": "LLaMA-Excitor: Alpacas are domesticated animals that are related to llamas and are native to South America. They are typically between 100 and \\(120~\\mathrm{cm}\\) tall and weigh between 100 and \\(150~\\mathrm{kg}\\). Alpacas have long, soft fleece that is used for making clothing, blankets, and other products. They are also used for their meat and their dung, which is used as fuel. Alpacas are herd animals and are known for their gentle and calm nature. They are also very intelligent and can be trained to perform certain tasks." + }, + { + "type": "title", + "bbox": [ + 0.437, + 0.093, + 0.764, + 0.105 + ], + "angle": 0, + "content": "Instruction: Write a conversation between the sun and Pluto." + }, + { + "type": "title", + "bbox": [ + 0.437, + 0.108, + 0.488, + 0.117 + ], + "angle": 0, + "content": "LLaMA7B:" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.117, + 0.573, + 0.126 + ], + "angle": 0, + "content": "The sun: Hey Pluto, how are you?" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.128, + 0.499, + 0.136 + ], + "angle": 0, + "content": "Pluto: I'm fine." + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.138, + 0.559, + 0.147 + ], + "angle": 0, + "content": "The sun: I'm glad to hear that." + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.149, + 0.536, + 0.157 + ], + "angle": 0, + "content": "Pluto: I'm glad to hear th" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.16, + 0.572, + 0.169 + ], + "angle": 0, + "content": "The sun: I'm glad to hear that too." + }, + { + "type": "title", + "bbox": [ + 0.583, + 0.109, + 0.66, + 0.117 + ], + "angle": 0, + "content": "Full Fine-Tuning:" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.117, + 0.692, + 0.126 + ], + "angle": 0, + "content": "The sun said, \"I'm so hot!\"" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.128, + 0.692, + 0.136 + ], + "angle": 0, + "content": "The sun said, \"I'm so hot!\"" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.138, + 0.692, + 0.147 + ], + "angle": 0, + "content": "The sun said, \"I'm so hot!\"" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.15, + 0.692, + 0.158 + ], + "angle": 0, + "content": "The sun said, \"I'm so hot!\"" + }, + { + "type": "text", + "bbox": [ + 0.583, + 0.161, + 0.692, + 0.169 + ], + "angle": 0, + "content": "The sun said, \"I'm so hot!\"" + }, + { + "type": "title", + "bbox": [ + 0.707, + 0.109, + 0.77, + 0.117 + ], + "angle": 0, + "content": "Alpaca-LoRA:" + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.117, + 0.887, + 0.133 + ], + "angle": 0, + "content": "Sun: Hey Pluto, how's life in the outer solar system?" + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.136, + 0.887, + 0.154 + ], + "angle": 0, + "content": "Pluto: It's pretty chilly up here! I'm still getting used to being so far away from the sun." + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.156, + 0.887, + 0.181 + ], + "angle": 0, + "content": "Sun: That's understandable. I'm sure you'll get used to it soon enough. Anyway, I'm glad you're still around." + }, + { + "type": "text", + "bbox": [ + 0.707, + 0.183, + 0.887, + 0.201 + ], + "angle": 0, + "content": "Pluto: Thanks! It's nice to still be a part of the solar system." + }, + { + "type": "title", + "bbox": [ + 0.437, + 0.193, + 0.514, + 0.202 + ], + "angle": 0, + "content": "LLaMA-Adapter:" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.202, + 0.564, + 0.211 + ], + "angle": 0, + "content": "The Sun: Good morning, Pluto!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.212, + 0.604, + 0.222 + ], + "angle": 0, + "content": "Pluto: Good morning, Sun! How are you?" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.223, + 0.599, + 0.233 + ], + "angle": 0, + "content": "The Sun: I'm doing great! How are you?" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.235, + 0.544, + 0.244 + ], + "angle": 0, + "content": "Pluto: I'm doing well, too!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.247, + 0.561, + 0.256 + ], + "angle": 0, + "content": "The Sun: I'm glad to hear that!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.259, + 0.496, + 0.267 + ], + "angle": 0, + "content": "Pluto: Me too!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.27, + 0.616, + 0.28 + ], + "angle": 0, + "content": "The Sun: I'm glad we're having such a nice day." + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.281, + 0.454, + 0.288 + ], + "angle": 0, + "content": "day" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.29, + 0.497, + 0.298 + ], + "angle": 0, + "content": "Pluto: Me too!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.3, + 0.616, + 0.308 + ], + "angle": 0, + "content": "The Sun: I'm glad we're having such a nice" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.309, + 0.455, + 0.315 + ], + "angle": 0, + "content": "day!" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.316, + 0.496, + 0.323 + ], + "angle": 0, + "content": "Pluto: Me too!" + }, + { + "type": "title", + "bbox": [ + 0.618, + 0.199, + 0.691, + 0.207 + ], + "angle": 0, + "content": "LLaMA-Excitor:" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.207, + 0.762, + 0.214 + ], + "angle": 0, + "content": "Sun: Hey Pluto, how are you doing?" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.214, + 0.778, + 0.222 + ], + "angle": 0, + "content": "Pluto: I'm doing great! How about you?" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.223, + 0.864, + 0.231 + ], + "angle": 0, + "content": "Sun: I'm doing well too. It's been a while since we last spoke." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.232, + 0.864, + 0.245 + ], + "angle": 0, + "content": "Pluto: Yeah, it has been a while. I've been busy exploring the outer reaches of the solar system." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.246, + 0.8, + 0.254 + ], + "angle": 0, + "content": "Sun: That's great! I'm glad you're having fun." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.255, + 0.864, + 0.269 + ], + "angle": 0, + "content": "Pluto: I'm having a blast! I'm so glad I'm not stuck in the inner solar system like the other planets." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.27, + 0.82, + 0.278 + ], + "angle": 0, + "content": "Sun: Yeah, it's nice to have some space to yourself." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.279, + 0.811, + 0.286 + ], + "angle": 0, + "content": "Pluto: Definitely! So, what have you been up to?" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.287, + 0.863, + 0.302 + ], + "angle": 0, + "content": "Sun: I've been busy keeping the other planets warm and providing light and heat." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.303, + 0.805, + 0.31 + ], + "angle": 0, + "content": "Pluto: That's a big job! You must be exhausted." + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.311, + 0.803, + 0.318 + ], + "angle": 0, + "content": "Sun: I am, but it's worth it. I'm glad I can help" + }, + { + "type": "text", + "bbox": [ + 0.618, + 0.319, + 0.865, + 0.333 + ], + "angle": 0, + "content": "Pluto: I'm glad you're doing well. It's been great talking to you." + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.348, + 0.752, + 0.362 + ], + "angle": 0, + "content": "Table 1. Instruction-following Comparison between LLaMA-Excitor and other techniques." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.389, + 0.453, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.493, + 0.47, + 0.522 + ], + "angle": 0, + "content": "Figure 4. Quantitative comparisons between LLaMA-Excitor (BLUE) with other methods, evaluated by GPT-4 [32]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.469, + 0.635 + ], + "angle": 0, + "content": "prompts, where \\( V \\) is the embedding length and \\( D \\) is the dimension of embeddings. Note that we adopt both the global CLS token and regional patch tokens from the last layer of the CLIP encoder to acquire adequate visual semantics. Then, we calculate multi-modal versions of the Excitor Key and the Exitor Value by combining \\( P_l \\) and \\( I \\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.644, + 0.469, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K e y} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {k}} ^ {\\text {e x c i t o r}} (I) \\right], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.678, + 0.469, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V a l u e} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {v}} ^ {\\text {e x c i t o r}} (I) \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Unlike the processing of \\( P_{l} \\), we adopt low-rank linear projections \\( W_{\\mathrm{k}}^{\\mathrm{excitor}} \\) and \\( W_{\\mathrm{v}}^{\\mathrm{excitor}} \\) on \\( I \\), since we provide the same visual prompt for attention layers \\( l \\in [N - L, N] \\) and different layers require various visual clues for the downstream reasoning. LLaMA-Excitor is an innovative framework for the multi-modal fine-tuning of LLMs. As mentioned in Section 1, most techniques before LLaMA-Excitor involved training a module to perform multi-modal alignment that projects visual embeddings into the textual feature space. This enables LLMs to understand visual tokens and integrate them into the reasoning process. However, the strict alignment constraint may damage the rich semantics of visual embeddings, and their direct feature fusion may" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.892, + 0.48 + ], + "angle": 0, + "content": "also cause performance degradation. In contrast, LLaMA-Excitor does not require feature alignment and works on the raw inputs from the frozen image encoder. It does not change the purity of the intermediate textual features and thus avoids any potential damage to the rich semantics of the visual embeddings and the reasoning of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.633, + 0.514 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.522, + 0.763, + 0.538 + ], + "angle": 0, + "content": "4.1. Language-only Performances" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.787, + 0.561 + ], + "angle": 0, + "content": "4.1.1 Instruction-following Evaluations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.891, + 0.601 + ], + "angle": 0, + "content": "In this section, we evaluate the instruction-following capacity of LLaMA-Excitor by responding to instructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.891, + 0.661 + ], + "angle": 0, + "content": "Dataset. We use Stanford Alpaca [44], which contains 52K instruction-following data, as the training set. Each sample in this dataset is formatted according to a template as follows:" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.663, + 0.88, + 0.705 + ], + "angle": 0, + "content": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request." + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.706, + 0.66, + 0.719 + ], + "angle": 0, + "content": "Instruction: {instruction}" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.719, + 0.594, + 0.733 + ], + "angle": 0, + "content": "Input: {input}" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.734, + 0.627, + 0.747 + ], + "angle": 0, + "content": "Response: {output}" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.891, + 0.794 + ], + "angle": 0, + "content": "where {instruction} is the description of a task, {input} is the context for the task, and {output} is the answer generated by GPT-3.5 [5]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We develop LLaMA-Excitor based on the original LLaMA codebase with minor modifications. We train LLaMA-Excitor on 8 NVIDIA A100 GPUs for five epochs. The warmup epochs, batch size, learning rate, and weight decay are set to 2, 64, 9e-3, and 0.02, respectively. In this section, we utilize the pretrained LLaMA-7B v1 model, which has \\( N = 32 \\) transformer lay" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14093" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.473, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.304 + ], + "angle": 0, + "content": "Figure 5. The relative average performance changes and win-loss situations of fine-tunings compared to original LLaMA-7B on MMLU." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.468, + 0.426 + ], + "angle": 0, + "content": "ers with feature dimension \\( C = 4096 \\) as the backbone. We apply our excitor blocks with the low-rank dimension \\( r = 16 \\) for \\( W_{\\mathrm{q}}^{\\mathrm{exciton}} \\) in the topmost \\( L = 30 \\) layers and set the length of learnable prompts \\( L = 30 \\). During the inference, we adopt top-p sampling as the default decoding method with a temperature of 0.1 and a top-p = 0.75." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.468, + 0.534 + ], + "angle": 0, + "content": "We compare LLaMA-Excitor with three competing methods. (1) Full-finetuning, which updates the entire 7B parameters in LLaMA during the end-to-end training. (2) Alpaca-LoRA, the official implementation of low-rank adaptation [15] provided by [44]. (3) LLaMA-Adapter [51], a cutting-edge prefix-tuning method with zero-init attention to reduce forgetting in finetuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.468, + 0.883 + ], + "angle": 0, + "content": "Performances. We first use examples in Table 1 to show the strong instruction-following ability of LLaMA-Excitor. The first case requires the models to provide some information about alpacas. LLaMA-Excitor generates reasonable responses comparable to Full-finetuning, demonstrating our method's effectiveness. It's worth noticing that our approach follows the same answering pattern as the prefix-tuning method LLaMA-adapter and the original LLaMA7B (although Excitor provides more details). This phenomenon is reasonable since prefix-tuning methods use additional learnable prompts to guide the generation process of LLMs. They do not change the internal reasoning process, just like our Excitor. This characteristic may also be the foundation of avoiding catastrophic forgetting. For the second task, the LLMs are asked to create a dialogue between the Sun and Pluto. Out of all the models that are tested, only Excitor and LoRA correctly recognize the characters as planets within the Solar System and incorporate their identities into the conversation. However, Excitor stands out as it can generate lengthy, high-quality conversations rich in interesting content. (Please see Appendix A for the full comparison including dialog generation, code generation, and question answering.)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To evaluate the performance of instruction-following" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.349 + ], + "angle": 0, + "content": "models quantitatively, we refer to the research conducted by Chiang et al. [8]. In this evaluation, we use GPT-4 [32] to assess the quality of responses to 80 randomly selected questions from AlpacaEval [25]. However, we notice a bias in GPT-4's scoring, where it tends to give higher scores to the first response. To overcome this, we interchange the position of two responses and use the average score for the final rating. Results are shown in Figure 4. We notice that Excitor overwhelms the Full-Finetuning (73/80). The reason behind this is that AlpacaEval is an out-of-distributed (OOD) test for our training set, and full-finetuning may easily overfit the training set. When compared with AlpacaLoRA and LLaMA-Adapter, our Excitor still gains the majority of \"Win\", demonstrating our indirect feature interaction is at least comparable with or better than predominant PEFT methods. (Please refer to Appendix B for instances on AlpacaEval.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.46 + ], + "angle": 0, + "content": "As this paper is a preliminary attempt to study indirect feature interaction in LLMs and the 52K instruction-following data used in our study is automatically generated, it is possible to further enhance the performance of LLaMA-Excitor by improving the quality of the training data, adjusting the hyper-parameters to more reasonable values or exploring larger LLaMA models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.505, + 0.891, + 0.534 + ], + "angle": 0, + "content": "4.1.2 The Impact of Fine-tuning Techniques on Inherent Abilities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.901 + ], + "angle": 0, + "content": "One of the motivations for introducing the indirect feature interaction in LLaMA-Excitor is to reduce the influence of fine-tuning on low-quality or non-targeted datasets. To quantitatively analyze this characteristic, we evaluate the fine-tuned models on MMLU [13], which is currently the most comprehensive benchmark for assessing the multi-task accuracy of text models. This benchmark includes 57 tasks across various fields, such as basic math, American history, computer science, and law. The model must possess extensive world knowledge and problem-solving capabilities to achieve high accuracy on this test. We report the relative performance changes of PEFT methods compared with the original LLaMA-7B and how many of the 57 evaluations PEFT methods outperform LLaMA-7B in Figure 5. We find that for models fine-tuned on Alpaca-52k, only Excitor can avoid reducing the average score of 57 tasks in MMLU, and it even shows an increase in performance \\((+3.12\\%)\\). As expected, the full-finetuning performance suffers the most severe decline as its parameters are most significantly changed \\((-6.63\\%)\\), while the LLaMA-Adapter, which is based on prefix-tuning, experiences the least decline \\((-3.02\\%)\\) besides Excitor. (We provide results based on LLaMA2-7B in Appendix C)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14094" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.092, + 0.648, + 0.116 + ], + "angle": 0, + "content": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.122, + 0.648, + 0.166 + ], + "angle": 0, + "content": "LLaMA-Excitor: An airplane is flying through the sky, with its landing gear down, as it approaches for a landing at an airport. The airplane is dominating the scene, with its bright red and white colors clearly visible against the backdrop of the cloudy sky. The airplane is about to touch down on the runway, creating a moment of excitement and anticipation for the passengers and crew on board." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.172, + 0.512, + 0.185 + ], + "angle": 0, + "content": "Human Annotation: A large passenger airplane flying through the air." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.191, + 0.648, + 0.215 + ], + "angle": 0, + "content": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.221, + 0.648, + 0.263 + ], + "angle": 0, + "content": "LLaMA-Excitor: In the image, a group of birds is standing together on a giraffe. One of the bird is leaning over to nibble on the giraffe's fur and the giraffe seems to be enjoying the interaction. The scene captures the playful nature of these animals and the unexpected interaction between a giraffe and a bird." + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.269, + 0.496, + 0.282 + ], + "angle": 0, + "content": "Human Annotation: Little birds sitting on the shoulder of a giraffe." + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.093, + 0.851, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.196, + 0.851, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.3, + 0.751, + 0.314 + ], + "angle": 0, + "content": "Table 2. Examples demonstrating LLaMA-Excitor's visual instruction following capacity." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.32, + 0.428, + 0.45 + ], + "angle": 0, + "content": "
MethodData ScaleCOCO Caption
PTFTBLEU@4CIDEr
ClipCap [31]0M0.6M33.5113.1
VL-PET [16]0M0.6M-121.7
LLaMA-AdapterV2 [12]0M0.6M36.2122.2
Qwen-vl-chat [4]1.4B0.6M-131.9
mPLUG-Owl2 [49]348M0.6M-137.3
BLIP [21]14M0.6M40.4136.7
Flamingo [2]1.8B0.6M-138.1
BLIP-2 [22]129M0.6M43.7145.3
LLaMA-Excitor0M0.6M49.7157.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.489 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art image captioning methods on COCO Caption." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.323, + 0.513 + ], + "angle": 0, + "content": "4.2. Multi-Modal Performances" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.613 + ], + "angle": 0, + "content": "In this section, we evaluate the visual instruction following the capacity of LLaMA-Excitor by responding to paired vision-language instructions and demonstrate our indirect feature interaction, which unifies the modeling of language-only tuning and multi-modal tuning, providing a powerful low-budget way to perform vision-language tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.338, + 0.649 + ], + "angle": 0, + "content": "4.2.1 Image Captioning Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.764 + ], + "angle": 0, + "content": "Implementation Details. We first evaluate our Excitor on COCO Caption [7], which contains 0.6M training image caption data (120k images with 5 captions per image) over a wide range of distributions. We apply frozen CLIP/14-L [38] as the image encoder, the visual embedding dimension \\( D = 768 \\), low-rank dimension \\( r = 16 \\). We keep other hyper-parameters the same as the section 4.1.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Performances. We compare our Excitor with cutting-edge image captioning methods in Table 3. We find that Excitor achieves a stunning result, significantly surpassing the previous SOTA method BILP-2 [22] by \\(+6\\) BLEU@4 and \\(+12.2\\) CIDEr. Especially considering that we do not apply complex multi-modal alignment modules and have only 0.6M training data. By contrast, BLIP-2 adopts a costly pre-training stage on an additional 129M dataset, including VisualGenome [19], Conceptual Captions [42, 43] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.504 + ], + "angle": 0, + "content": "LAION [6]. We visualize the results of image captioning in Appendix D. However, the image captions provided in COCO are concise and lack sufficient detail in describing the content of images. To further unleash Exitor's performance in image captioning, we continue fine-tuning Exitor on LLaVA665k [27], a higher-quality dataset with detailed picture descriptions. We provide several examples from the COCO test set in Table 2. It shows that image captions generated by Excitor can accurately cover the content of human annotations and provide richer details. (e.g., (1) The airplane's landing gear is down, (2) A bird is pecking at the back of a giraffe)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.528, + 0.753, + 0.543 + ], + "angle": 0, + "content": "4.2.2 Performances on ScienceQA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Implementation Details. We evaluate our Excitor on ScienceQA [40], which contains 21k multimodal multiple choice questions with rich domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. We train Excitor on ScienceQA train split from scratch, using the combination of Chain-of-Thought(CoT) and direct answer prediction. CoT requires the model to predict the solution first and then generate the answer choice based on the solution. During training, we randomly require the model to first generate solutions or directly predict answers in each iteration. The model is asked to answer the question directly in the evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 4 reports the performances of the top methods on the official leaderboard of the ScienceQA and Excitor. The current SOTA method LLaVA [28] is pretrained on another 558k visual-language dataset and fine-tuned on ScienceQA. Also, they allow all LLM's parameters to be updated during fine-tuning. Our Excitor demonstrates comparable performance with LLaVA w/o pretraining (only -0.4% performance gap). Note that LLaMA-Excitor is a PEFT method (with a frozen LLM) without CoT, and LLaVA utilizes a larger backbone LLaMA-13B than our LLaMA-7B. To further unlock the potential of Excitor, we introduce a vari" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14095" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.855, + 0.344 + ], + "angle": 0, + "content": "
ModelAverageSubjectContext ModalityGrade
NATSOCLANTXTIMGNOG1-6G7-12
Human [40]88.4090.2384.9787.4889.6087.5088.1091.5982.42
UnifiedQACoT74.1171.0076.0478.9166.4266.5381.8177.0668.82
GPT-3CoT75.1775.4470.8778.0974.6867.4379.9378.2369.68
ChatGPTCoT [1]78.3178.8270.9883.1877.3767.9286.1380.7274.03
GPT-4CoT [32]83.9985.4872.4490.2782.6571.4992.8986.6679.04
MM-COT [53]84.9187.5277.1785.8287.8882.9086.8384.6585.37
LLaVACoT [28]90.9290.3695.9588.0089.4988.0090.6690.9390.90
LLaVACoT (w/o pretrain) [28]85.81--------
DFAF [11]60.7264.0348.8263.5565.8854.4964.1157.1267.17
ViLT [18]61.1460.4863.8960.2763.2061.3857.0060.7261.90
Patch-TRM [30]61.4265.1946.7965.5566.9655.2864.9558.0467.50
VisualBERT [23, 24]61.8759.3369.1861.1862.7162.1758.5462.9659.92
UnifiedQA [17]70.1268.1669.1874.9163.7861.3877.8472.9865.00
GPT-3 [5]74.0475.0466.5978.0074.2465.7479.5876.3669.87
LLaMA-Adapter85.1984.3788.3084.3683.7280.3286.9085.8384.05
LLaMA-Excitor85.4185.7092.3582.8283.4384.5686.2785.6584.64
LLaMA-Excitor@336px + LoRA88.3987.1991.3387.0990.4285.2088.6488.3588.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.354, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Table 4. Question Answering Accuracy (%) on ScienceQA's [40] test set. We report GPT-3 [5], ChatGPT [1], and GPT-4 [32] for zero-shot inference. CoT denotes utilizing chain-of-thought for question answering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.472, + 0.485 + ], + "angle": 0, + "content": "ant that utilizes a more powerful image encoder (CLIP-L/14@336) and adds LoRA blocks to the original LLM's parameters to reduce the gap with LLaVA's full parameter updating. It brings \\(+2.58\\%\\) performance gain compared with LLaVA." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.192, + 0.512 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.532, + 0.428, + 0.627 + ], + "angle": 0, + "content": "
VariantACC
(a) CLIP-16/B83.11
(b) CLIP-14/L85.41
(c) CLIP-14/L@336px85.87
(d) CLIP-14/L@336px + Full-Finetuning83.20
(e) CLIP-14/L@336px + LoRA88.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.638, + 0.434, + 0.652 + ], + "angle": 0, + "content": "Table 5. Analysis of the effectiveness of each module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.473, + 0.904 + ], + "angle": 0, + "content": "We ablate several design choices on ScienceQA in Table 5. (a), (b) and (c) study the influence of the capacity of image encoder. The best CLIP encoder yields \\(85.41\\%\\) and is \\(+2.76\\%\\) higher than the basic version. In (d) and (e), we study to combine Excitor, which focuses on instruction-following, with other techniques that update LLM's reasoning process. We find LoRA can bring \\(2.98\\%\\) performance gain for Excitor, demonstrating the potential of combining Excitor with existing finetuning methods in multimodal usage. (However, in the text-only scenario, we encounter \\(-4.14\\%\\) performance gap compared with the original LLaMA-7B in the MMLU test by combining Excitor and LoRA.) Please see Appendix E for more ablations on the design of low-rank linear layers (i.e., \\(W_{\\mathrm{q}}^{Excitor}\\), \\(W_{\\mathrm{k}}^{Excitor}\\), and \\(W_{\\mathrm{v}}^{Excitor}\\)) in Excitor Blocks under single" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.895, + 0.456 + ], + "angle": 0, + "content": "modal and multi-modal scenarios, the choice of low-rank dimension \\( r \\), and the number of layers \\( L \\) with Excitor blocks inserted." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.468, + 0.62, + 0.483 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.659 + ], + "angle": 0, + "content": "In this paper, we study the indirect feature interaction in fine-tuning LLMs into instruction-following models and propose LLaMA-Exictor, a PEFT method that demonstrates outstanding instruction-following capacity compared with predominant techniques meanwhile reducing the forgetting of LLMs' inherent abilities. Moreover, Excitor also shows its potential in vision-language tasks by unifying the modeling of visual instruction-following and language-only instruction-following. Impressively, Excitor achieves strong multi-modal performance without the need to train complex alignment modules." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.781 + ], + "angle": 0, + "content": "This project is a work in progress, and some directions can be explored. The first one is the adaptability to LLMs other than LLaMA. The second direction is to explore more effective ways to introduce multi-scale visual prompts (since we only apply the features from the last layer of CLIP at the current stage). Last but not least, Excitor has the potential to fine-tune vision-only models into vision-language models, just like our implementation for LLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.794, + 0.618, + 0.811 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.895, + 0.896 + ], + "angle": 0, + "content": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Postdoctoral Excellent Program (Grant No. 2022234)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14096" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.426, + 0.129 + ], + "angle": 0, + "content": "[1] Chatgpt. https://chat.openai.com, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.47, + 0.212 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.256 + ], + "angle": 0, + "content": "[3] Anonymous. Videodistill: Language-aware vision distillation for video question answering. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.469, + 0.325 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.469, + 0.48 + ], + "angle": 0, + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.469, + 0.548 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arxiv 2015. arXiv preprint arXiv:1504.00325, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[9] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[11] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639-6648, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[12] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[16] Zi-Yuan Hu, Yanyang Li, Michael R Lyu, and Liwei Wang. Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3010-3020, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[17] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896-1907, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583-5594, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[20] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.748, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tat-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14097" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "sunori B. Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[29] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[30] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[31] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[32] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[34] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[35] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[36] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[40] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[44] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[47] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[48] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[49] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[50] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. arXiv preprint arXiv:2303.10512, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.805, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14098" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[53] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.248 + ], + "angle": 0, + "content": "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14099" + } + ] +] \ No newline at end of file diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_origin.pdf b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..afe4d7c3776b2affbc0750071d555e8a9834afc2 --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/54e482bc-9979-46be-8235-fd3729928e88_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b06ff9bf97934ed202ec9aff02055561d458a7f177cc90909d079b9fbbbfbe6 +size 2309344 diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/full.md b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b7ba89939cb9fe7c844202267e5ddb0a1b6f3d0 --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/full.md @@ -0,0 +1,431 @@ +# LLaMA-Excitor: General Instruction Tuning via Indirect Feature Interaction + +Bo Zou * + +Tsinghua University Beijing, China + +zoub21@mails.tsinghua.edu.cn + +Chao Yang * + +Shanghai AI Laboratory Shanghai, China + +yangchao@pjlab.org.cn + +Yu Qiao + +Shanghai AI Laboratory Shanghai, China + +qiaoyu@pjlab.org.cn + +Chengbin Quan + +Tsinghua University + +Beijing, China + +quancb@tsinghua.edu.cn + +Youjian Zhao† + +Tsinghua University + +Zhongguancun Laboratory + +Beijing, China + +zhaoyoujian@tsinghua.edu.cn + +# Abstract + +Existing methods to fine-tune LLMs, like Adapter, Prefixtuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation. We designed the Excitor block as a bypass module that reconstructs Keys and changes the importance of Values in self-attention using learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pretrained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our approach is evaluated in language-only and multimodal scenarios. Compared with the original LLaMA-7B, LLaMA-Excitor is the only PEFT method that maintains basic capabilities and achieves $+3.12\%$ relative improvement on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance on MSCOCO (157.5 CIDEr), and a comparable performance on ScienceQA (88.39%) to cutting-edge models with more parameters and extensive vision-language pertaining. The code will be available at https://zoubo9034.github.io/Excitor/. + +![](images/07402a80ce6dd75de8090ddb762a1cfbd08ce9fd3ddd8fd1015c4467934bcfc4.jpg) +Figure 1. Overview of LLaMA-Excitor. We integrate Excitor blocks into L out of N attention layers of LLaMA. Differing from previous PEFT techniques, LLaMA-Excitor indirectly involves learnable information in the reasoning process by changing the similarity matrices. It ensures that the hidden states are within the original distribution of LLaMA. + +# 1. Introduction + +Large language models (LLMs) are highly advanced tools that enable contextual information processing, understanding, and generation. They serve as powerful knowledge bases that can be used to gain valuable insights and create new content [9, 10, 37, 39, 52]. Fine-tuning them for downstream tasks has also achieved great success in vari- + +ous domains. With the rapid growth in scale and quality of the corpus and the advance in hardware, the inherent abilities of LLMs are becoming stronger. Therefore, the primary objective of fine-tuning general-purpose LLMs should be ensuring they produce the desired output for downstream tasks (i.e., instruction-following ability) rather than memory excessive knowledge or mastering skills (e.g., causal inference and numerical computation). + +Prioritizing the enhancement of instruction-following capabilities for LLMs is prudent for a multitude of reasons. Firstly, it is often presumed that the corpus utilized during training encompasses a comprehensive repository of knowledge, thereby obviating the need for additional fine-tuning datasets when addressing specific downstream tasks. This implies that enriching a model's ability to follow instructions could leverage this pre-existing knowledge base more effectively. Secondly, by standardizing the output format during instruction-following tasks, we can reduce the generation of harmful or irrelevant content. Third, the challenge of incorporating new knowledge and skills into LLMs is accentuated by the limited set of parameters available for fine-tuning. Lastly, fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual. + +However, predominant parameter-efficient fine-tuning (PEFT) techniques are met with notable challenges in performing instruction-tuning. The Adapter methods[14, 36], which have a serial structure, can cause computation bottlenecks, and they also significantly alter the inference process to adapt to downstream tasks, which in turn can lead to degradation of inherent abilities, such as catastrophic forgetting. Prompt learning methods[5, 26, 29, 41] effectively incorporate new information by creating additional parameters corresponding to the input sequences of language models. However, concentrating all necessary knowledge into fixed-length token sequences is challenging for various downstream tasks. Furthermore, as the length of the generated sequence increases, the control over the sequence diminishes since LLMs generate the next word based on a softmax over the entire sequence. LoRA methods[15, 50] directly add the outcomes of randomly initialized low-rank modules with linear layers' outputs, which can introduce features that are out of LLMs' feature distribution and may cause degradations. Recent studies [16, 50] also show that Adapter and LoRA evenly assign learnable parameters to trainable modules and neglect the functionality gap between each layer in LLM, which will lead to performance drops. + +In this paper, we start from a new perspective and propose LLaMA-Excitor, a PEFT method that focuses on the following instructions. (1) LLaMA-Excitor aims to optimize instruction-following ability by releasing the poten + +tial of an LLM, i.e., LLaMA [46], instead of pursuing new knowledge and skills. Due to variations in the data scale, quality, and content coverage of instruction-tuning sets, the effect of fine-tuning processes on the inherent skills of LLMs is unpredictable. (2) LLaMA-Excitor can reduce the degradation when fine-tuning on unsuited datasets. Specifically, rather than directly changing the intermediate hidden state of pretrained LLMs like Adapter and LoRA, LLaMA-Excitor uses trainable bypass modules (Excitor blocks) to modify the attention score meanwhile maintain original Values for each attention layer in LLaMA. Excitor blocks only adjust the proportion of information in Values for hidden states, ensuring that each attention layer's input is derived from a linear combination of the previous layer's Values and corresponds to the original LLaMA's distribution (an indirect feature interaction process). Additionally, the modification of the attention score depends on both the input sequence and a set of prompt tokens to make the model adaptively decide whether to rely on additional information from the fine-tuning set for attention allocation. These attributes ensure fast training and few forgetings. + +Moreover, Exictor innovatively provides a low-budget way to fine-tune a language-only model into a vision-language model. Previous works (e.g., [3, 28, 31, 47, 51]) rely on training an additional visual branch or projection module to align vision and language, which incurs significant computational overhead. We explore the unified modeling of instruction-tuning and visual instruction-tuning for LLMs and utilize a visual Exictor to make LLaMA follow the visual prompts generated from a frozen image encoder like CLIP [38]. In this way, we eliminate the need for extra vision-language alignment while preserving textual features' purity during LLMs' reasoning and shortening the training time. Our contributions are summarized as follows: + +- We study indirect feature interaction in fine-tuning LLMs and propose LLaMA-Excitor, a PEFT method that focuses on instruction-following and reduces forgetting. +- We uniformly models multi-modal and language-only tuning and extends language models into powerful vision-language models in a low-budget way. +- Experiments on instruction-following datasets, multi-task evaluation, image captioning, and VQA demonstrate the feasibility and advancement of LLaMA-Excitor. + +# 2. Related Works + +# 2.1. Parameter-Efficient Fine-Tuning + +PEFT techniques enable efficient adaptation of large language models to various downstream applications without fine-tuning all the model's parameters. Adapter-based approaches [14, 36] insert additional feed-forward networks within the existing architecture. However, these methods can increase the computational load due to their se + +![](images/3ffa15b86d79452037231e99e355b92a299fd36f1aaa37bdd379543c92b29c46.jpg) +Figure 2. Details of the Excitor block. We assign a set of learnable prompts for attention layers of LLaMA. These prompts are used to construct an extra Key for computing additional similarity scores, which are then merged into the original scores to alter the LLM's behavior. Cold-start gating factors are designed to stabilize the training. + +rial nature and may induce catastrophic forgetting by significantly changing the reasoning mechanism of LLMs. Prompt-learning methods offer a training-free approach to task adaptation [5, 41] and are improved by prefix-tuning [26, 29], which involves learnable token sequences. The primary drawback is the difficulty in capturing complex knowledge within a limited number of tokens, which can be restrictive for diverse tasks. Additionally, as the number of learnable tokens grows, these methods may lose control over the sequence output quality. LoRA [15, 50] aims to address these issues by introducing low-rank matrices in a parallel structure, enhancing both performance and efficiency. However, it risks incorporating features that deviate from the original model's distribution, potentially leading to performance degradation. Furthermore, equal distribution of parameter updates across layers disregards the distinct functionalities of different layers [16, 50], leading to suboptimal utilization of the model's capacity. + +# 2.2. Visual Instruction Tuning + +Instruction tuning[33, 48] finetunes or adapted language models to follow specific textual instructions. Despite the prominent success in the NLP domain, its inherent text-only nature limits its ability to comprehend and interact with visual content effectively. LLaVA [28] enables language models to follow visual instructions and engage with multi-modal information. And a series of approaches have been developed to improve the visual instruction tuning with advanced architectures [2, 4, 22] or more versatile functionalities[20, 35, 54]. Fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual. + +# 3. Methodology + +The primary goal of Excitor is to effectively leverage LLMs' pre-trained knowledge and skills to solve downstream tasks without excessive modification of the reasoning process, which can lead to forgetting and performance degradation. We choose LLaMA[45] as our LLM as its effectiveness has been demonstrated in several open-source instruction-tuning works [8, 28, 34, 44, 51]. In Section 3.1, we first introduce how to insert Excitor blocks into LLaMA's attention layers. Then, in Section 3.2, we present the details of Excitor blocks' architecture. Finally, in section 3.3, we generalize LLaMA-Excitor for multi-modal reasoning. + +# 3.1. Indirect feature interaction + +Given 52K instruction-following data [44] and a pre-trained LLaMA with $N$ attention layers, we insert our proposed Excitor blocks into the topmost $L(L < N)$ layers for instruction-following fine-tuning. This can better tune the language representations with higher-level semantics. Taking the $l$ -th inserted layer as an example, we denote the $M$ -length word tokens as $T_{l} \in \mathbb{R}^{M \times C}$ , where $C$ equals the feature dimension of LLaMA's attention layers. As shown in Figure 1, the Excitor block takes $T_{l}$ with a set of learnable prompts $\{P_{l}\}_{l=1}^{L}$ as inputs, where $P_{l} \in \mathbb{R}^{K \times C}$ with $K$ denoting the length of learnable prompts for each layer, generates an extra similarity matrix $S_{l}^{\mathrm{extra}} \in \mathbb{R}^{M \times M}$ with exactly the same shape of original similarity matrix $S_{l}$ for the $l$ -th layer. In the self-attention mechanism, $S_{l}$ is calculated from Query and Key, and controls the importance of components in Value, where Query, Key, and Value are linear projected version of $T_{l}$ : + +$$ +Q u e y = W _ {\mathrm {q}} \left(T _ {l}\right), K e y = W _ {\mathrm {k}} \left(T _ {l}\right), V a l u e = W _ {\mathrm {v}} \left(T _ {l}\right) \tag {1} +$$ + +where $W_{\mathrm{q}}$ , $W_{\mathrm{k}}$ , $W_{\mathrm{v}}$ are the pre-trained LLMs' weights that are frozen during the finetuning. Then, $S_{l}$ is formulated as: + +$$ +S _ {l} = \frac {\text {Q u e r y} \cdot \text {K e y} ^ {\mathrm {T}}}{\sqrt {C}}, \tag {2} +$$ + +Since $S_{l}$ rules how the information updates, we treat $S_{l}^{extra}$ as the residual of $S_{l}$ to modify the reasoning process and formulate the updated output of the $l$ -th layer $T_{l}^{\mathrm{o}}$ as: + +$$ +S _ {l} ^ {g} = \operatorname {S o f t m a x} \left(S _ {l} ^ {\text {e x t r a}} \times g _ {l} + S _ {l}\right), \tag {3} +$$ + +$$ +T _ {l} ^ {\mathrm {o}} = W _ {\mathrm {o}} \left(S _ {l} ^ {g} \cdot \text {V a l u e}\right), \tag {4} +$$ + +where $g_{l}$ is a set of learnable parameters, named cold-start gating factor, which is inspired by zero-init attention [51], to stabilize training by controlling the proportion of $S_{l}^{extra}$ participants in the reasoning. Instead of applying zero-initialization like [51], we initialize $g_{l}$ with $\mathbf{N}\left(0,10^{-2}\right)$ to overcome the gradient vanishing in the mixed precision (FP16) training. + +As stated in sec 1, previous techniques updating the output of frozen LLMs' layers by adding the outcomes of trainable modules onto $T_{l}^{\mathrm{o}}$ (e.g., Adapter[14] and LoRA[15]) or concatenating trainable tokens with $T_{l}$ as input for layer $l$ (e.g., Prefix-tuning [26]) can be attributed to the direct modifying of the intermediate representations. One drawback of direct modifying is adding out-of-distribution features into the intermediate representations. It will largely change the inputs of higher layers and make the framework unstable, especially at the beginning of training. By contrast, LLaMA-Excitor only uses trainable tokens to change the attention score $S_{l}^{g}$ . The output of each layer $T_{l}^{\mathrm{o}}$ is still the linear combination of Value and within the original LLMs' distribution. i.e., LLaMA-Excitor alters the likelihood of potential outcomes generated by pre-trained LLMs. This indirect feature interaction ensures the model doesn't deviate too far from its pretrained reasoning process and can more effectively arouse pretrained skills in downstream usage. + +# 3.2. Excitor Blocks + +To impose effective control over the attention score $S_{l}^{g}$ while reducing the degradation of LLM's inherent abilities, LLaMA-Excitor must be self-adaptive to the original input sequence $T_{l}$ to decide whether learnable parameters should influence the reasoning. Thus, Excitor Blocks needs to generate the extra similarity matrix $S_{l}^{\mathrm{extra}}$ considering both injected information in the learnable prompt $P_{l}$ and $T_{l}$ . We formulate $S_{l}^{\mathrm{extra}}$ as: + +$$ +S _ {l} ^ {\text {e x t r a}} = \frac {\operatorname {Q u e r y} \cdot \operatorname {K e y} _ {\text {e x t r a}} ^ {\mathrm {T}}}{\sqrt {C}}, \tag {5} +$$ + +where $Query$ is reused from the equation 1 to pass knowledge from $T_{l}$ while reducing the computational overhead, $Key_{extra}$ is an extra attention key that carries information from $P_{l}$ . Since the shape of $S_{l}^{\text{extra}}$ should correspond + +![](images/74fb0a4b6acef288cdfeaf7f9720a264c4a6408f1320bcd2cc33a13333f3f354.jpg) +Figure 3. Extend LLaMA-Excitor into a powerful multi-modal model. Owing to the indirect feature interaction, LLaMA-Excitor is the cheapest PEFT method that can follow visual instructions without complicated projection modules aligning vision and language. + +with $S_{l}$ , the shape of $Key_{extra}$ must be the same with $Key \in \mathbb{R}^{M \times C}$ in the equation 1. However, $M$ , which is the length of $Key$ , is changing along with variable input texts, while the length of learnable prompts $P_{l}$ is a pre-defined hyper-parameter. + +As shown on the left of the Excitor Block in Figure 2, we introduce the Key reconstruction to mitigate the shape inconsistency. Since the attention mechanism generates an output with the same shape as its queries and transmits information about its values, we define $P_{l}$ as the Exitor Key and the Exitor Value and calculate the Excitor Query from $T_{l}$ for the $l$ -th Excitor Block. We omit linear projections for $P_{l}$ to pass information directly to $Key_{extra}$ and accelerate the training. The Key reconstruction is formulated as: + +$$ +S _ {l} ^ {\text {e x c i t o r}} = \frac {W _ {\mathrm {q}} ^ {\text {e x c i t o r}} \left(T _ {l}\right) \cdot P _ {l} ^ {\mathrm {T}}}{\sqrt {C}}, \tag {6} +$$ + +$$ +\operatorname {K e y} _ {\text {e x t r a}} = \operatorname {S o f t m a x} \left(S _ {l} ^ {\text {e x c i t o r}}\right) \cdot P _ {l}, \tag {7} +$$ + +Where $W_{\mathrm{q}}^{\mathrm{excitor}}$ is a low-rank module consisting of two linear layers with dimension $\mathbb{R}^{C\times r}$ and $\mathbb{R}^{r\times C}$ . Finally, $S_{l}^{\mathrm{extra}}$ carries information from $T_{l}$ and has the same shape as Key. + +# 3.3. Multi-modal Extension + +The indirect feature interaction is not limited to textual instructions but is also compatible with following visual prompts. We simply extend our Excitor Block into a multimodal version in Figure 3. Taking an image and a task description as input, LLaMA-Excitor first leverages a frozen pre-trained visual encoder like CLIP [38] to extract a sequence of visual embeddings $I \in \mathbb{R}^{V \times D}$ as the visual + +# Instruction: Tell me about alpacas. + +LLaMA7B: Alpacas are a type of camelid that are native to South America. They are related to llamas and vicunas. Alpacas are raised for their fiber, which is used to make clothing. + +Full Fine-Tuning: Alpacas are small, soft-furry, and mammals that are native to South America. They are the largest of the wool-making mammals and are the most gentle and social of the mammals. They are also the most colorful mammal, with a range of colors from white to gray, brown, black, and shades of brown and gray. They are also very smart and can learn tricks easily. They are also very social and live in herds of up to 20-30 individuals. + +Alpaca-LoRA: Alpacas are a species of South American camelid, related to the llama and the vicuna. They are a domesticated version of the wild vicuna and are primarily raised for their fine fleece, which is used to make clothing. Alpacas are typically smaller than llamas, with a range of sizes from 12 to 15 hands high. They are known for their gentle nature and are often kept as pets. + +LLaMA-Adapter: Alpacas are members of the camelid family and are native to the Andes Mountains of South America. They are typically between 3 and 4 feet tall and weigh between 100 and 200 pounds. They have long, coarse, and dense coats that come in a variety of colors, including white, black, brown, and grey. + +LLaMA-Excitor: Alpacas are domesticated animals that are related to llamas and are native to South America. They are typically between 100 and $120~\mathrm{cm}$ tall and weigh between 100 and $150~\mathrm{kg}$ . Alpacas have long, soft fleece that is used for making clothing, blankets, and other products. They are also used for their meat and their dung, which is used as fuel. Alpacas are herd animals and are known for their gentle and calm nature. They are also very intelligent and can be trained to perform certain tasks. + +# Instruction: Write a conversation between the sun and Pluto. + +# LLaMA7B: + +The sun: Hey Pluto, how are you? + +Pluto: I'm fine. + +The sun: I'm glad to hear that. + +Pluto: I'm glad to hear th + +The sun: I'm glad to hear that too. + +# Full Fine-Tuning: + +The sun said, "I'm so hot!" + +The sun said, "I'm so hot!" + +The sun said, "I'm so hot!" + +The sun said, "I'm so hot!" + +The sun said, "I'm so hot!" + +# Alpaca-LoRA: + +Sun: Hey Pluto, how's life in the outer solar system? + +Pluto: It's pretty chilly up here! I'm still getting used to being so far away from the sun. + +Sun: That's understandable. I'm sure you'll get used to it soon enough. Anyway, I'm glad you're still around. + +Pluto: Thanks! It's nice to still be a part of the solar system. + +# LLaMA-Adapter: + +The Sun: Good morning, Pluto! + +Pluto: Good morning, Sun! How are you? + +The Sun: I'm doing great! How are you? + +Pluto: I'm doing well, too! + +The Sun: I'm glad to hear that! + +Pluto: Me too! + +The Sun: I'm glad we're having such a nice day. + +day + +Pluto: Me too! + +The Sun: I'm glad we're having such a nice + +day! + +Pluto: Me too! + +# LLaMA-Excitor: + +Sun: Hey Pluto, how are you doing? + +Pluto: I'm doing great! How about you? + +Sun: I'm doing well too. It's been a while since we last spoke. + +Pluto: Yeah, it has been a while. I've been busy exploring the outer reaches of the solar system. + +Sun: That's great! I'm glad you're having fun. + +Pluto: I'm having a blast! I'm so glad I'm not stuck in the inner solar system like the other planets. + +Sun: Yeah, it's nice to have some space to yourself. + +Pluto: Definitely! So, what have you been up to? + +Sun: I've been busy keeping the other planets warm and providing light and heat. + +Pluto: That's a big job! You must be exhausted. + +Sun: I am, but it's worth it. I'm glad I can help + +Pluto: I'm glad you're doing well. It's been great talking to you. + +![](images/1c7b4dd884401a33c80c0683d0f77da961a50c0e9ba819736eeef25c255d2f71.jpg) +Table 1. Instruction-following Comparison between LLaMA-Excitor and other techniques. +Figure 4. Quantitative comparisons between LLaMA-Excitor (BLUE) with other methods, evaluated by GPT-4 [32]. + +prompts, where $V$ is the embedding length and $D$ is the dimension of embeddings. Note that we adopt both the global CLS token and regional patch tokens from the last layer of the CLIP encoder to acquire adequate visual semantics. Then, we calculate multi-modal versions of the Excitor Key and the Exitor Value by combining $P_l$ and $I$ : + +$$ +\operatorname {K e y} _ {\text {e x c i t o r}} = \left[ P _ {l}; W _ {\mathrm {k}} ^ {\text {e x c i t o r}} (I) \right], \tag {8} +$$ + +$$ +\operatorname {V a l u e} _ {\text {e x c i t o r}} = \left[ P _ {l}; W _ {\mathrm {v}} ^ {\text {e x c i t o r}} (I) \right]. \tag {9} +$$ + +Unlike the processing of $P_{l}$ , we adopt low-rank linear projections $W_{\mathrm{k}}^{\mathrm{excitor}}$ and $W_{\mathrm{v}}^{\mathrm{excitor}}$ on $I$ , since we provide the same visual prompt for attention layers $l \in [N - L, N]$ and different layers require various visual clues for the downstream reasoning. LLaMA-Excitor is an innovative framework for the multi-modal fine-tuning of LLMs. As mentioned in Section 1, most techniques before LLaMA-Excitor involved training a module to perform multi-modal alignment that projects visual embeddings into the textual feature space. This enables LLMs to understand visual tokens and integrate them into the reasoning process. However, the strict alignment constraint may damage the rich semantics of visual embeddings, and their direct feature fusion may + +also cause performance degradation. In contrast, LLaMA-Excitor does not require feature alignment and works on the raw inputs from the frozen image encoder. It does not change the purity of the intermediate textual features and thus avoids any potential damage to the rich semantics of the visual embeddings and the reasoning of LLMs. + +# 4. Experiments + +# 4.1. Language-only Performances + +# 4.1.1 Instruction-following Evaluations + +In this section, we evaluate the instruction-following capacity of LLaMA-Excitor by responding to instructions. + +Dataset. We use Stanford Alpaca [44], which contains 52K instruction-following data, as the training set. Each sample in this dataset is formatted according to a template as follows: + +Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. + +Instruction: {instruction} + +Input: {input} + +Response: {output} + +where {instruction} is the description of a task, {input} is the context for the task, and {output} is the answer generated by GPT-3.5 [5]. + +Implementation Details. We develop LLaMA-Excitor based on the original LLaMA codebase with minor modifications. We train LLaMA-Excitor on 8 NVIDIA A100 GPUs for five epochs. The warmup epochs, batch size, learning rate, and weight decay are set to 2, 64, 9e-3, and 0.02, respectively. In this section, we utilize the pretrained LLaMA-7B v1 model, which has $N = 32$ transformer lay + +![](images/2e252e041efef300d5769f4527b5379e1d3bc009152682997785967fd1692363.jpg) +Figure 5. The relative average performance changes and win-loss situations of fine-tunings compared to original LLaMA-7B on MMLU. + +ers with feature dimension $C = 4096$ as the backbone. We apply our excitor blocks with the low-rank dimension $r = 16$ for $W_{\mathrm{q}}^{\mathrm{exciton}}$ in the topmost $L = 30$ layers and set the length of learnable prompts $L = 30$ . During the inference, we adopt top-p sampling as the default decoding method with a temperature of 0.1 and a top-p = 0.75. + +We compare LLaMA-Excitor with three competing methods. (1) Full-finetuning, which updates the entire 7B parameters in LLaMA during the end-to-end training. (2) Alpaca-LoRA, the official implementation of low-rank adaptation [15] provided by [44]. (3) LLaMA-Adapter [51], a cutting-edge prefix-tuning method with zero-init attention to reduce forgetting in finetuning. + +Performances. We first use examples in Table 1 to show the strong instruction-following ability of LLaMA-Excitor. The first case requires the models to provide some information about alpacas. LLaMA-Excitor generates reasonable responses comparable to Full-finetuning, demonstrating our method's effectiveness. It's worth noticing that our approach follows the same answering pattern as the prefix-tuning method LLaMA-adapter and the original LLaMA7B (although Excitor provides more details). This phenomenon is reasonable since prefix-tuning methods use additional learnable prompts to guide the generation process of LLMs. They do not change the internal reasoning process, just like our Excitor. This characteristic may also be the foundation of avoiding catastrophic forgetting. For the second task, the LLMs are asked to create a dialogue between the Sun and Pluto. Out of all the models that are tested, only Excitor and LoRA correctly recognize the characters as planets within the Solar System and incorporate their identities into the conversation. However, Excitor stands out as it can generate lengthy, high-quality conversations rich in interesting content. (Please see Appendix A for the full comparison including dialog generation, code generation, and question answering.) + +To evaluate the performance of instruction-following + +models quantitatively, we refer to the research conducted by Chiang et al. [8]. In this evaluation, we use GPT-4 [32] to assess the quality of responses to 80 randomly selected questions from AlpacaEval [25]. However, we notice a bias in GPT-4's scoring, where it tends to give higher scores to the first response. To overcome this, we interchange the position of two responses and use the average score for the final rating. Results are shown in Figure 4. We notice that Excitor overwhelms the Full-Finetuning (73/80). The reason behind this is that AlpacaEval is an out-of-distributed (OOD) test for our training set, and full-finetuning may easily overfit the training set. When compared with AlpacaLoRA and LLaMA-Adapter, our Excitor still gains the majority of "Win", demonstrating our indirect feature interaction is at least comparable with or better than predominant PEFT methods. (Please refer to Appendix B for instances on AlpacaEval.) + +As this paper is a preliminary attempt to study indirect feature interaction in LLMs and the 52K instruction-following data used in our study is automatically generated, it is possible to further enhance the performance of LLaMA-Excitor by improving the quality of the training data, adjusting the hyper-parameters to more reasonable values or exploring larger LLaMA models. + +# 4.1.2 The Impact of Fine-tuning Techniques on Inherent Abilities + +One of the motivations for introducing the indirect feature interaction in LLaMA-Excitor is to reduce the influence of fine-tuning on low-quality or non-targeted datasets. To quantitatively analyze this characteristic, we evaluate the fine-tuned models on MMLU [13], which is currently the most comprehensive benchmark for assessing the multi-task accuracy of text models. This benchmark includes 57 tasks across various fields, such as basic math, American history, computer science, and law. The model must possess extensive world knowledge and problem-solving capabilities to achieve high accuracy on this test. We report the relative performance changes of PEFT methods compared with the original LLaMA-7B and how many of the 57 evaluations PEFT methods outperform LLaMA-7B in Figure 5. We find that for models fine-tuned on Alpaca-52k, only Excitor can avoid reducing the average score of 57 tasks in MMLU, and it even shows an increase in performance $(+3.12\%)$ . As expected, the full-finetuning performance suffers the most severe decline as its parameters are most significantly changed $(-6.63\%)$ , while the LLaMA-Adapter, which is based on prefix-tuning, experiences the least decline $(-3.02\%)$ besides Excitor. (We provide results based on LLaMA2-7B in Appendix C) + +Instruction: Please answer me based on this image. Generate a caption that details what happened in the image. + +LLaMA-Excitor: An airplane is flying through the sky, with its landing gear down, as it approaches for a landing at an airport. The airplane is dominating the scene, with its bright red and white colors clearly visible against the backdrop of the cloudy sky. The airplane is about to touch down on the runway, creating a moment of excitement and anticipation for the passengers and crew on board. + +Human Annotation: A large passenger airplane flying through the air. + +Instruction: Please answer me based on this image. Generate a caption that details what happened in the image. + +LLaMA-Excitor: In the image, a group of birds is standing together on a giraffe. One of the bird is leaning over to nibble on the giraffe's fur and the giraffe seems to be enjoying the interaction. The scene captures the playful nature of these animals and the unexpected interaction between a giraffe and a bird. + +Human Annotation: Little birds sitting on the shoulder of a giraffe. + +![](images/5c43262a5b3ceb4ac8de19634fb9be3a1c902c816d64c997d542cc141adb2fc4.jpg) + +![](images/90c5cd3fb6434b6de7be91207a3888bcadc5ca0d1b009cf2dfa7eba77ec93057.jpg) + +Table 2. Examples demonstrating LLaMA-Excitor's visual instruction following capacity. + +
MethodData ScaleCOCO Caption
PTFTBLEU@4CIDEr
ClipCap [31]0M0.6M33.5113.1
VL-PET [16]0M0.6M-121.7
LLaMA-AdapterV2 [12]0M0.6M36.2122.2
Qwen-vl-chat [4]1.4B0.6M-131.9
mPLUG-Owl2 [49]348M0.6M-137.3
BLIP [21]14M0.6M40.4136.7
Flamingo [2]1.8B0.6M-138.1
BLIP-2 [22]129M0.6M43.7145.3
LLaMA-Excitor0M0.6M49.7157.5
+ +Table 3. Comparison with state-of-the-art image captioning methods on COCO Caption. + +# 4.2. Multi-Modal Performances + +In this section, we evaluate the visual instruction following the capacity of LLaMA-Excitor by responding to paired vision-language instructions and demonstrate our indirect feature interaction, which unifies the modeling of language-only tuning and multi-modal tuning, providing a powerful low-budget way to perform vision-language tasks. + +# 4.2.1 Image Captioning Evaluation + +Implementation Details. We first evaluate our Excitor on COCO Caption [7], which contains 0.6M training image caption data (120k images with 5 captions per image) over a wide range of distributions. We apply frozen CLIP/14-L [38] as the image encoder, the visual embedding dimension $D = 768$ , low-rank dimension $r = 16$ . We keep other hyper-parameters the same as the section 4.1.1. + +Performances. We compare our Excitor with cutting-edge image captioning methods in Table 3. We find that Excitor achieves a stunning result, significantly surpassing the previous SOTA method BILP-2 [22] by $+6$ BLEU@4 and $+12.2$ CIDEr. Especially considering that we do not apply complex multi-modal alignment modules and have only 0.6M training data. By contrast, BLIP-2 adopts a costly pre-training stage on an additional 129M dataset, including VisualGenome [19], Conceptual Captions [42, 43] and + +LAION [6]. We visualize the results of image captioning in Appendix D. However, the image captions provided in COCO are concise and lack sufficient detail in describing the content of images. To further unleash Exitor's performance in image captioning, we continue fine-tuning Exitor on LLaVA665k [27], a higher-quality dataset with detailed picture descriptions. We provide several examples from the COCO test set in Table 2. It shows that image captions generated by Excitor can accurately cover the content of human annotations and provide richer details. (e.g., (1) The airplane's landing gear is down, (2) A bird is pecking at the back of a giraffe) + +# 4.2.2 Performances on ScienceQA + +Implementation Details. We evaluate our Excitor on ScienceQA [40], which contains 21k multimodal multiple choice questions with rich domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. We train Excitor on ScienceQA train split from scratch, using the combination of Chain-of-Thought(CoT) and direct answer prediction. CoT requires the model to predict the solution first and then generate the answer choice based on the solution. During training, we randomly require the model to first generate solutions or directly predict answers in each iteration. The model is asked to answer the question directly in the evaluation. + +Table 4 reports the performances of the top methods on the official leaderboard of the ScienceQA and Excitor. The current SOTA method LLaVA [28] is pretrained on another 558k visual-language dataset and fine-tuned on ScienceQA. Also, they allow all LLM's parameters to be updated during fine-tuning. Our Excitor demonstrates comparable performance with LLaVA w/o pretraining (only -0.4% performance gap). Note that LLaMA-Excitor is a PEFT method (with a frozen LLM) without CoT, and LLaVA utilizes a larger backbone LLaMA-13B than our LLaMA-7B. To further unlock the potential of Excitor, we introduce a vari + +
ModelAverageSubjectContext ModalityGrade
NATSOCLANTXTIMGNOG1-6G7-12
Human [40]88.4090.2384.9787.4889.6087.5088.1091.5982.42
UnifiedQACoT74.1171.0076.0478.9166.4266.5381.8177.0668.82
GPT-3CoT75.1775.4470.8778.0974.6867.4379.9378.2369.68
ChatGPTCoT [1]78.3178.8270.9883.1877.3767.9286.1380.7274.03
GPT-4CoT [32]83.9985.4872.4490.2782.6571.4992.8986.6679.04
MM-COT [53]84.9187.5277.1785.8287.8882.9086.8384.6585.37
LLaVACoT [28]90.9290.3695.9588.0089.4988.0090.6690.9390.90
LLaVACoT (w/o pretrain) [28]85.81--------
DFAF [11]60.7264.0348.8263.5565.8854.4964.1157.1267.17
ViLT [18]61.1460.4863.8960.2763.2061.3857.0060.7261.90
Patch-TRM [30]61.4265.1946.7965.5566.9655.2864.9558.0467.50
VisualBERT [23, 24]61.8759.3369.1861.1862.7162.1758.5462.9659.92
UnifiedQA [17]70.1268.1669.1874.9163.7861.3877.8472.9865.00
GPT-3 [5]74.0475.0466.5978.0074.2465.7479.5876.3669.87
LLaMA-Adapter85.1984.3788.3084.3683.7280.3286.9085.8384.05
LLaMA-Excitor85.4185.7092.3582.8283.4384.5686.2785.6584.64
LLaMA-Excitor@336px + LoRA88.3987.1991.3387.0990.4285.2088.6488.3588.42
+ +ant that utilizes a more powerful image encoder (CLIP-L/14@336) and adds LoRA blocks to the original LLM's parameters to reduce the gap with LLaVA's full parameter updating. It brings $+2.58\%$ performance gain compared with LLaVA. + +# 4.3. Ablations + +Table 4. Question Answering Accuracy (%) on ScienceQA's [40] test set. We report GPT-3 [5], ChatGPT [1], and GPT-4 [32] for zero-shot inference. CoT denotes utilizing chain-of-thought for question answering. + +
VariantACC
(a) CLIP-16/B83.11
(b) CLIP-14/L85.41
(c) CLIP-14/L@336px85.87
(d) CLIP-14/L@336px + Full-Finetuning83.20
(e) CLIP-14/L@336px + LoRA88.39
+ +Table 5. Analysis of the effectiveness of each module. + +We ablate several design choices on ScienceQA in Table 5. (a), (b) and (c) study the influence of the capacity of image encoder. The best CLIP encoder yields $85.41\%$ and is $+2.76\%$ higher than the basic version. In (d) and (e), we study to combine Excitor, which focuses on instruction-following, with other techniques that update LLM's reasoning process. We find LoRA can bring $2.98\%$ performance gain for Excitor, demonstrating the potential of combining Excitor with existing finetuning methods in multimodal usage. (However, in the text-only scenario, we encounter $-4.14\%$ performance gap compared with the original LLaMA-7B in the MMLU test by combining Excitor and LoRA.) Please see Appendix E for more ablations on the design of low-rank linear layers (i.e., $W_{\mathrm{q}}^{Excitor}$ , $W_{\mathrm{k}}^{Excitor}$ , and $W_{\mathrm{v}}^{Excitor}$ ) in Excitor Blocks under single + +modal and multi-modal scenarios, the choice of low-rank dimension $r$ , and the number of layers $L$ with Excitor blocks inserted. + +# 5. Conclusion + +In this paper, we study the indirect feature interaction in fine-tuning LLMs into instruction-following models and propose LLaMA-Exictor, a PEFT method that demonstrates outstanding instruction-following capacity compared with predominant techniques meanwhile reducing the forgetting of LLMs' inherent abilities. Moreover, Excitor also shows its potential in vision-language tasks by unifying the modeling of visual instruction-following and language-only instruction-following. Impressively, Excitor achieves strong multi-modal performance without the need to train complex alignment modules. + +This project is a work in progress, and some directions can be explored. The first one is the adaptability to LLMs other than LLaMA. The second direction is to explore more effective ways to introduce multi-scale visual prompts (since we only apply the features from the last layer of CLIP at the current stage). Last but not least, Excitor has the potential to fine-tune vision-only models into vision-language models, just like our implementation for LLMs. + +# Acknowledge + +This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Postdoctoral Excellent Program (Grant No. 2022234). + +# References + +[1] Chatgpt. https://chat.openai.com, 2023.8 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3, 7 +[3] Anonymous. Videodistill: Language-aware vision distillation for video question answering. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3, 7 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 5, 8 +[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 7 +[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arxiv 2015. arXiv preprint arXiv:1504.00325, 2015. 7 +[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3, 6 +[9] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[11] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639-6648, 2019. 8 +[12] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.7 + +[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6 +[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6 +[16] Zi-Yuan Hu, Yanyang Li, Michael R Lyu, and Liwei Wang. Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3010-3020, 2023. 2, 3, 7 +[17] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896-1907, 2020. 8 +[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583-5594, 2021. 8 +[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 7 +[20] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3 +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 7 +[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 7 +[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 8 +[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020. 8 +[25] Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tat- + +sunori B. Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval, 2023.6 +[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 7, 8 +[29] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 2, 3 +[30] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 8 +[31] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 7 +[32] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 5, 6, 8 +[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 3 +[34] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 3 +[35] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3 +[36] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2 +[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 7 +[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with + +a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1 +[40] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 7, 8 +[41] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 2, 3 +[42] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 7 +[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7 +[44] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 3, 5, 6 +[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971, 2023. 2 +[47] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2 +[48] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 3 +[49] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7 +[50] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. arXiv preprint arXiv:2303.10512, 2023. 2, 3 +[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2, 3, 4, 6 +[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, + +Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1 +[53] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.8 +[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 \ No newline at end of file diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/images.zip b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7515ad4436d312cce2ddc5206999814f05b9a374 --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82154acaaa006456ae50a4173f22ce9711dc01444ff7ecbc89e50ba5a41252de +size 385651 diff --git a/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/layout.json b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..caa46b7b67ff6c4aa075cd2f706ecf3c3831cb6c --- /dev/null +++ b/2024/LLaMA-Excitor_ General Instruction Tuning via Indirect Feature Interaction/layout.json @@ -0,0 +1,10566 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 536, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 536, + 120 + ], + "type": "text", + "content": "LLaMA-Excitor: General Instruction Tuning via Indirect Feature Interaction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 132, + 177, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 132, + 177, + 144 + ], + "spans": [ + { + "bbox": [ + 128, + 132, + 177, + 144 + ], + "type": "text", + "content": "Bo Zou *" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 146, + 203, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 146, + 203, + 174 + ], + "spans": [ + { + "bbox": [ + 102, + 146, + 203, + 174 + ], + "type": "text", + "content": "Tsinghua University Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 176, + 229, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 176, + 229, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 176, + 229, + 186 + ], + "type": "text", + "content": "zoub21@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 132, + 347, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 132, + 347, + 145 + ], + "spans": [ + { + "bbox": [ + 281, + 132, + 347, + 145 + ], + "type": "text", + "content": "Chao Yang *" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 255, + 146, + 373, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 146, + 373, + 174 + ], + "spans": [ + { + "bbox": [ + 255, + 146, + 373, + 174 + ], + "type": "text", + "content": "Shanghai AI Laboratory Shanghai, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 256, + 176, + 372, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 176, + 372, + 187 + ], + "spans": [ + { + "bbox": [ + 256, + 176, + 372, + 187 + ], + "type": "text", + "content": "yangchao@pjlab.org.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 437, + 132, + 479, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 132, + 479, + 145 + ], + "spans": [ + { + "bbox": [ + 437, + 132, + 479, + 145 + ], + "type": "text", + "content": "Yu Qiao" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 399, + 146, + 517, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 146, + 517, + 174 + ], + "spans": [ + { + "bbox": [ + 399, + 146, + 517, + 174 + ], + "type": "text", + "content": "Shanghai AI Laboratory Shanghai, China" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 406, + 176, + 510, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 176, + 510, + 187 + ], + "spans": [ + { + "bbox": [ + 406, + 176, + 510, + 187 + ], + "type": "text", + "content": "qiaoyu@pjlab.org.cn" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 164, + 194, + 242, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 194, + 242, + 207 + ], + "spans": [ + { + "bbox": [ + 164, + 194, + 242, + 207 + ], + "type": "text", + "content": "Chengbin Quan" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 154, + 208, + 252, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 208, + 252, + 221 + ], + "spans": [ + { + "bbox": [ + 154, + 208, + 252, + 221 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 167, + 222, + 239, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 222, + 239, + 236 + ], + "spans": [ + { + "bbox": [ + 167, + 222, + 239, + 236 + ], + "type": "text", + "content": "Beijing, China" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 238, + 263, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 238, + 263, + 248 + ], + "spans": [ + { + "bbox": [ + 143, + 238, + 263, + 248 + ], + "type": "text", + "content": "quancb@tsinghua.edu.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 194, + 413, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 194, + 413, + 207 + ], + "spans": [ + { + "bbox": [ + 337, + 194, + 413, + 207 + ], + "type": "text", + "content": "Youjian Zhao†" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 208, + 425, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 208, + 425, + 221 + ], + "spans": [ + { + "bbox": [ + 326, + 208, + 425, + 221 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 222, + 440, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 222, + 440, + 236 + ], + "spans": [ + { + "bbox": [ + 310, + 222, + 440, + 236 + ], + "type": "text", + "content": "Zhongguancun Laboratory" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 236, + 411, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 236, + 411, + 249 + ], + "spans": [ + { + "bbox": [ + 340, + 236, + 411, + 249 + ], + "type": "text", + "content": "Beijing, China" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 301, + 251, + 447, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 251, + 447, + 262 + ], + "spans": [ + { + "bbox": [ + 301, + 251, + 447, + 262 + ], + "type": "text", + "content": "zhaoyoujian@tsinghua.edu.cn" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 143, + 290, + 192, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 290, + 192, + 303 + ], + "spans": [ + { + "bbox": [ + 143, + 290, + 192, + 303 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 314, + 289, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 289, + 673 + ], + "type": "text", + "content": "Existing methods to fine-tune LLMs, like Adapter, Prefixtuning, and LoRA, which introduce extra modules or additional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation. We designed the Excitor block as a bypass module that reconstructs Keys and changes the importance of Values in self-attention using learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pretrained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our approach is evaluated in language-only and multimodal scenarios. Compared with the original LLaMA-7B, LLaMA-Excitor is the only PEFT method that maintains basic capabilities and achieves " + }, + { + "bbox": [ + 47, + 314, + 289, + 673 + ], + "type": "inline_equation", + "content": "+3.12\\%" + }, + { + "bbox": [ + 47, + 314, + 289, + 673 + ], + "type": "text", + "content": " relative improvement on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance on MSCOCO (157.5 CIDEr), and a comparable performance on ScienceQA (88.39%) to cutting-edge models with more parameters and extensive vision-language pertaining. The code will be available at https://zoubo9034.github.io/Excitor/." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 328, + 291, + 533, + 524 + ], + "blocks": [ + { + "bbox": [ + 328, + 291, + 533, + 524 + ], + "lines": [ + { + "bbox": [ + 328, + 291, + 533, + 524 + ], + "spans": [ + { + "bbox": [ + 328, + 291, + 533, + 524 + ], + "type": "image", + "image_path": "07402a80ce6dd75de8090ddb762a1cfbd08ce9fd3ddd8fd1015c4467934bcfc4.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 533, + 547, + 600 + ], + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 600 + ], + "type": "text", + "content": "Figure 1. Overview of LLaMA-Excitor. We integrate Excitor blocks into L out of N attention layers of LLaMA. Differing from previous PEFT techniques, LLaMA-Excitor indirectly involves learnable information in the reasoning process by changing the similarity matrices. It ensures that the hidden states are within the original distribution of LLaMA." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 620, + 387, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Large language models (LLMs) are highly advanced tools that enable contextual information processing, understanding, and generation. They serve as powerful knowledge bases that can be used to gain valuable insights and create new content [9, 10, 37, 39, 52]. Fine-tuning them for downstream tasks has also achieved great success in vari-" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 263, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 263, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 263, + 713 + ], + "type": "text", + "content": "* Equal contribution, † Corresponding author \nThis work was done during an internship at Shanghai AI Lab." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14089" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "ous domains. With the rapid growth in scale and quality of the corpus and the advance in hardware, the inherent abilities of LLMs are becoming stronger. Therefore, the primary objective of fine-tuning general-purpose LLMs should be ensuring they produce the desired output for downstream tasks (i.e., instruction-following ability) rather than memory excessive knowledge or mastering skills (e.g., causal inference and numerical computation)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 287, + 386 + ], + "type": "text", + "content": "Prioritizing the enhancement of instruction-following capabilities for LLMs is prudent for a multitude of reasons. Firstly, it is often presumed that the corpus utilized during training encompasses a comprehensive repository of knowledge, thereby obviating the need for additional fine-tuning datasets when addressing specific downstream tasks. This implies that enriching a model's ability to follow instructions could leverage this pre-existing knowledge base more effectively. Secondly, by standardizing the output format during instruction-following tasks, we can reduce the generation of harmful or irrelevant content. Third, the challenge of incorporating new knowledge and skills into LLMs is accentuated by the limited set of parameters available for fine-tuning. Lastly, fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 388, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 663 + ], + "type": "text", + "content": "However, predominant parameter-efficient fine-tuning (PEFT) techniques are met with notable challenges in performing instruction-tuning. The Adapter methods[14, 36], which have a serial structure, can cause computation bottlenecks, and they also significantly alter the inference process to adapt to downstream tasks, which in turn can lead to degradation of inherent abilities, such as catastrophic forgetting. Prompt learning methods[5, 26, 29, 41] effectively incorporate new information by creating additional parameters corresponding to the input sequences of language models. However, concentrating all necessary knowledge into fixed-length token sequences is challenging for various downstream tasks. Furthermore, as the length of the generated sequence increases, the control over the sequence diminishes since LLMs generate the next word based on a softmax over the entire sequence. LoRA methods[15, 50] directly add the outcomes of randomly initialized low-rank modules with linear layers' outputs, which can introduce features that are out of LLMs' feature distribution and may cause degradations. Recent studies [16, 50] also show that Adapter and LoRA evenly assign learnable parameters to trainable modules and neglect the functionality gap between each layer in LLM, which will lead to performance drops." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "In this paper, we start from a new perspective and propose LLaMA-Excitor, a PEFT method that focuses on the following instructions. (1) LLaMA-Excitor aims to optimize instruction-following ability by releasing the poten" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "tial of an LLM, i.e., LLaMA [46], instead of pursuing new knowledge and skills. Due to variations in the data scale, quality, and content coverage of instruction-tuning sets, the effect of fine-tuning processes on the inherent skills of LLMs is unpredictable. (2) LLaMA-Excitor can reduce the degradation when fine-tuning on unsuited datasets. Specifically, rather than directly changing the intermediate hidden state of pretrained LLMs like Adapter and LoRA, LLaMA-Excitor uses trainable bypass modules (Excitor blocks) to modify the attention score meanwhile maintain original Values for each attention layer in LLaMA. Excitor blocks only adjust the proportion of information in Values for hidden states, ensuring that each attention layer's input is derived from a linear combination of the previous layer's Values and corresponds to the original LLaMA's distribution (an indirect feature interaction process). Additionally, the modification of the attention score depends on both the input sequence and a set of prompt tokens to make the model adaptively decide whether to rely on additional information from the fine-tuning set for attention allocation. These attributes ensure fast training and few forgetings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 324, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 479 + ], + "type": "text", + "content": "Moreover, Exictor innovatively provides a low-budget way to fine-tune a language-only model into a vision-language model. Previous works (e.g., [3, 28, 31, 47, 51]) rely on training an additional visual branch or projection module to align vision and language, which incurs significant computational overhead. We explore the unified modeling of instruction-tuning and visual instruction-tuning for LLMs and utilize a visual Exictor to make LLaMA follow the visual prompts generated from a frozen image encoder like CLIP [38]. In this way, we eliminate the need for extra vision-language alignment while preserving textual features' purity during LLMs' reasoning and shortening the training time. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 481, + 545, + 587 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 481, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 545, + 515 + ], + "type": "text", + "content": "- We study indirect feature interaction in fine-tuning LLMs and propose LLaMA-Excitor, a PEFT method that focuses on instruction-following and reduces forgetting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 517, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 545, + 552 + ], + "type": "text", + "content": "- We uniformly models multi-modal and language-only tuning and extends language models into powerful vision-language models in a low-budget way." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "type": "text", + "content": "- Experiments on instruction-following datasets, multi-task evaluation, image captioning, and VQA demonstrate the feasibility and advancement of LLaMA-Excitor." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 601, + 397, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 601, + 397, + 615 + ], + "spans": [ + { + "bbox": [ + 306, + 601, + 397, + 615 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 622, + 481, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 481, + 635 + ], + "type": "text", + "content": "2.1. Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "PEFT techniques enable efficient adaptation of large language models to various downstream applications without fine-tuning all the model's parameters. Adapter-based approaches [14, 36] insert additional feed-forward networks within the existing architecture. However, these methods can increase the computational load due to their se" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14090" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 73, + 492, + 233 + ], + "blocks": [ + { + "bbox": [ + 105, + 73, + 492, + 233 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 492, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 492, + 233 + ], + "type": "image", + "image_path": "3ffa15b86d79452037231e99e355b92a299fd36f1aaa37bdd379543c92b29c46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 282 + ], + "type": "text", + "content": "Figure 2. Details of the Excitor block. We assign a set of learnable prompts for attention layers of LLaMA. These prompts are used to construct an extra Key for computing additional similarity scores, which are then merged into the original scores to alter the LLM's behavior. Cold-start gating factors are designed to stabilize the training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 504 + ], + "type": "text", + "content": "rial nature and may induce catastrophic forgetting by significantly changing the reasoning mechanism of LLMs. Prompt-learning methods offer a training-free approach to task adaptation [5, 41] and are improved by prefix-tuning [26, 29], which involves learnable token sequences. The primary drawback is the difficulty in capturing complex knowledge within a limited number of tokens, which can be restrictive for diverse tasks. Additionally, as the number of learnable tokens grows, these methods may lose control over the sequence output quality. LoRA [15, 50] aims to address these issues by introducing low-rank matrices in a parallel structure, enhancing both performance and efficiency. However, it risks incorporating features that deviate from the original model's distribution, potentially leading to performance degradation. Furthermore, equal distribution of parameter updates across layers disregards the distinct functionalities of different layers [16, 50], leading to suboptimal utilization of the model's capacity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 523, + 191, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 191, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 191, + 537 + ], + "type": "text", + "content": "2.2. Visual Instruction Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 713 + ], + "type": "text", + "content": "Instruction tuning[33, 48] finetunes or adapted language models to follow specific textual instructions. Despite the prominent success in the NLP domain, its inherent text-only nature limits its ability to comprehend and interact with visual content effectively. LLaVA [28] enables language models to follow visual instructions and engage with multi-modal information. And a series of approaches have been developed to improve the visual instruction tuning with advanced architectures [2, 4, 22] or more versatile functionalities[20, 35, 54]. Fine-tuning in the context of multi-modal tasks can be succinctly expressed as enhancing a model's proficiency in responding to visual cues. This essentially constitutes a variant of instruction following, where the prompts are visual rather than textual." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 286, + 389, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 286, + 389, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 286, + 389, + 300 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 311, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 547, + 456 + ], + "type": "text", + "content": "The primary goal of Excitor is to effectively leverage LLMs' pre-trained knowledge and skills to solve downstream tasks without excessive modification of the reasoning process, which can lead to forgetting and performance degradation. We choose LLaMA[45] as our LLM as its effectiveness has been demonstrated in several open-source instruction-tuning works [8, 28, 34, 44, 51]. In Section 3.1, we first introduce how to insert Excitor blocks into LLaMA's attention layers. Then, in Section 3.2, we present the details of Excitor blocks' architecture. Finally, in section 3.3, we generalize LLaMA-Excitor for multi-modal reasoning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 473, + 457, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 457, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 457, + 486 + ], + "type": "text", + "content": "3.1. Indirect feature interaction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": "Given 52K instruction-following data [44] and a pre-trained LLaMA with " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " attention layers, we insert our proposed Excitor blocks into the topmost " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "L(L < N)" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " layers for instruction-following fine-tuning. This can better tune the language representations with higher-level semantics. Taking the " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": "-th inserted layer as an example, we denote the " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": "-length word tokens as " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "T_{l} \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " equals the feature dimension of LLaMA's attention layers. As shown in Figure 1, the Excitor block takes " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " with a set of learnable prompts " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "\\{P_{l}\\}_{l=1}^{L}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " as inputs, where " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "P_{l} \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " denoting the length of learnable prompts for each layer, generates an extra similarity matrix " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "S_{l}^{\\mathrm{extra}} \\in \\mathbb{R}^{M \\times M}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " with exactly the same shape of original similarity matrix " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": "-th layer. In the self-attention mechanism, " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": " is calculated from Query and Key, and controls the importance of components in Value, where Query, Key, and Value are linear projected version of " + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 304, + 495, + 547, + 683 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "Q u e y = W _ {\\mathrm {q}} \\left(T _ {l}\\right), K e y = W _ {\\mathrm {k}} \\left(T _ {l}\\right), V a l u e = W _ {\\mathrm {v}} \\left(T _ {l}\\right) \\tag {1}", + "image_path": "1d504f5bb53cec537450e86fed76125052abd8cf197cf42c23026ec372ea571f.jpg" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14091" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{q}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{k}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{v}}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": " are the pre-trained LLMs' weights that are frozen during the finetuning. Then, " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": " is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 124, + 102, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 102, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 124, + 102, + 287, + 129 + ], + "type": "interline_equation", + "content": "S _ {l} = \\frac {\\text {Q u e r y} \\cdot \\text {K e y} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {2}", + "image_path": "66d10b8239b685e885c1e866ac4c4587f447d745b62b2682460a44603ca6d0da.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": " rules how the information updates, we treat " + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "inline_equation", + "content": "S_{l}^{extra}" + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": " as the residual of " + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": " to modify the reasoning process and formulate the updated output of the " + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "inline_equation", + "content": "T_{l}^{\\mathrm{o}}" + }, + { + "bbox": [ + 47, + 134, + 287, + 170 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 176, + 287, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 176, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 176, + 287, + 190 + ], + "type": "interline_equation", + "content": "S _ {l} ^ {g} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x t r a}} \\times g _ {l} + S _ {l}\\right), \\tag {3}", + "image_path": "80054aeac7fdaa23ab6d1475f03bbe8cca1a53938a5f0672a873cac7d95f2797.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 198, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 198, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 120, + 198, + 287, + 209 + ], + "type": "interline_equation", + "content": "T _ {l} ^ {\\mathrm {o}} = W _ {\\mathrm {o}} \\left(S _ {l} ^ {g} \\cdot \\text {V a l u e}\\right), \\tag {4}", + "image_path": "26846fb5ba7e5f6b31241f0ce971d62fd2eb40d31642ffd26669dd489cd90e1f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "g_{l}" + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "content": " is a set of learnable parameters, named cold-start gating factor, which is inspired by zero-init attention [51], to stabilize training by controlling the proportion of " + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "S_{l}^{extra}" + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "content": " participants in the reasoning. Instead of applying zero-initialization like [51], we initialize " + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "g_{l}" + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{N}\\left(0,10^{-2}\\right)" + }, + { + "bbox": [ + 47, + 213, + 287, + 297 + ], + "type": "text", + "content": " to overcome the gradient vanishing in the mixed precision (FP16) training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": "As stated in sec 1, previous techniques updating the output of frozen LLMs' layers by adding the outcomes of trainable modules onto " + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "inline_equation", + "content": "T_{l}^{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": " (e.g., Adapter[14] and LoRA[15]) or concatenating trainable tokens with " + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": " as input for layer " + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": " (e.g., Prefix-tuning [26]) can be attributed to the direct modifying of the intermediate representations. One drawback of direct modifying is adding out-of-distribution features into the intermediate representations. It will largely change the inputs of higher layers and make the framework unstable, especially at the beginning of training. By contrast, LLaMA-Excitor only uses trainable tokens to change the attention score " + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "inline_equation", + "content": "S_{l}^{g}" + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": ". The output of each layer " + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "inline_equation", + "content": "T_{l}^{\\mathrm{o}}" + }, + { + "bbox": [ + 46, + 297, + 287, + 513 + ], + "type": "text", + "content": " is still the linear combination of Value and within the original LLMs' distribution. i.e., LLaMA-Excitor alters the likelihood of potential outcomes generated by pre-trained LLMs. This indirect feature interaction ensures the model doesn't deviate too far from its pretrained reasoning process and can more effectively arouse pretrained skills in downstream usage." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 520, + 138, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 138, + 531 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 138, + 531 + ], + "type": "text", + "content": "3.2. Excitor Blocks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": "To impose effective control over the attention score " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "S_{l}^{g}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": " while reducing the degradation of LLM's inherent abilities, LLaMA-Excitor must be self-adaptive to the original input sequence " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": " to decide whether learnable parameters should influence the reasoning. Thus, Excitor Blocks needs to generate the extra similarity matrix " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "S_{l}^{\\mathrm{extra}}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": " considering both injected information in the learnable prompt " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": ". We formulate " + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "inline_equation", + "content": "S_{l}^{\\mathrm{extra}}" + }, + { + "bbox": [ + 46, + 537, + 287, + 626 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 634, + 287, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 634, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 111, + 634, + 287, + 659 + ], + "type": "interline_equation", + "content": "S _ {l} ^ {\\text {e x t r a}} = \\frac {\\operatorname {Q u e r y} \\cdot \\operatorname {K e y} _ {\\text {e x t r a}} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {5}", + "image_path": "85ee92a83e7a54801694e8971dc705289874290ccab50e0f93d1f72891a65604.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "Query" + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": " is reused from the equation 1 to pass knowledge from " + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": " while reducing the computational overhead, " + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "Key_{extra}" + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": " is an extra attention key that carries information from " + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": ". Since the shape of " + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "inline_equation", + "content": "S_{l}^{\\text{extra}}" + }, + { + "bbox": [ + 47, + 666, + 287, + 714 + ], + "type": "text", + "content": " should correspond" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 526, + 277 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 526, + 277 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 526, + 277 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 526, + 277 + ], + "type": "image", + "image_path": "74fb0a4b6acef288cdfeaf7f9720a264c4a6408f1320bcd2cc33a13333f3f354.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 285, + 545, + 340 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 340 + ], + "type": "text", + "content": "Figure 3. Extend LLaMA-Excitor into a powerful multi-modal model. Owing to the indirect feature interaction, LLaMA-Excitor is the cheapest PEFT method that can follow visual instructions without complicated projection modules aligning vision and language." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "S_{l}" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": ", the shape of " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "Key_{extra}" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": " must be the same with " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "Key \\in \\mathbb{R}^{M \\times C}" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": " in the equation 1. However, " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": ", which is the length of " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "Key" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": ", is changing along with variable input texts, while the length of learnable prompts " + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 305, + 359, + 545, + 418 + ], + "type": "text", + "content": " is a pre-defined hyper-parameter." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": "As shown on the left of the Excitor Block in Figure 2, we introduce the Key reconstruction to mitigate the shape inconsistency. Since the attention mechanism generates an output with the same shape as its queries and transmits information about its values, we define " + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": " as the Exitor Key and the Exitor Value and calculate the Excitor Query from " + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": "-th Excitor Block. We omit linear projections for " + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": " to pass information directly to " + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "inline_equation", + "content": "Key_{extra}" + }, + { + "bbox": [ + 305, + 418, + 545, + 518 + ], + "type": "text", + "content": " and accelerate the training. The Key reconstruction is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 366, + 523, + 545, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 523, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 366, + 523, + 545, + 548 + ], + "type": "interline_equation", + "content": "S _ {l} ^ {\\text {e x c i t o r}} = \\frac {W _ {\\mathrm {q}} ^ {\\text {e x c i t o r}} \\left(T _ {l}\\right) \\cdot P _ {l} ^ {\\mathrm {T}}}{\\sqrt {C}}, \\tag {6}", + "image_path": "c9938b937151af48f0be582d001bd68f8c6eac342eb81a7b31a46ca881fd6ab3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 553, + 545, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 553, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 354, + 553, + 545, + 572 + ], + "type": "interline_equation", + "content": "\\operatorname {K e y} _ {\\text {e x t r a}} = \\operatorname {S o f t m a x} \\left(S _ {l} ^ {\\text {e x c i t o r}}\\right) \\cdot P _ {l}, \\tag {7}", + "image_path": "fe9bd9f9799dd3cff9b2f74abcf7a9157f79c96247ce895a289ecf745f547903.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{q}}^{\\mathrm{excitor}}" + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": " is a low-rank module consisting of two linear layers with dimension " + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times r}" + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{r\\times C}" + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "inline_equation", + "content": "S_{l}^{\\mathrm{extra}}" + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": " carries information from " + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "inline_equation", + "content": "T_{l}" + }, + { + "bbox": [ + 306, + 574, + 545, + 612 + ], + "type": "text", + "content": " and has the same shape as Key." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 618, + 436, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 436, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 436, + 629 + ], + "type": "text", + "content": "3.3. Multi-modal Extension" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 635, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 545, + 713 + ], + "type": "text", + "content": "The indirect feature interaction is not limited to textual instructions but is also compatible with following visual prompts. We simply extend our Excitor Block into a multimodal version in Figure 3. Taking an image and a task description as input, LLaMA-Excitor first leverages a frozen pre-trained visual encoder like CLIP [38] to extract a sequence of visual embeddings " + }, + { + "bbox": [ + 305, + 635, + 545, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{V \\times D}" + }, + { + "bbox": [ + 305, + 635, + 545, + 713 + ], + "type": "text", + "content": " as the visual" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14092" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 51, + 73, + 170, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 73, + 170, + 83 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 170, + 83 + ], + "type": "text", + "content": "Instruction: Tell me about alpacas." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 50, + 85, + 251, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 85, + 251, + 106 + ], + "spans": [ + { + "bbox": [ + 50, + 85, + 251, + 106 + ], + "type": "text", + "content": "LLaMA7B: Alpacas are a type of camelid that are native to South America. They are related to llamas and vicunas. Alpacas are raised for their fiber, which is used to make clothing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 108, + 252, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 108, + 252, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 108, + 252, + 148 + ], + "type": "text", + "content": "Full Fine-Tuning: Alpacas are small, soft-furry, and mammals that are native to South America. They are the largest of the wool-making mammals and are the most gentle and social of the mammals. They are also the most colorful mammal, with a range of colors from white to gray, brown, black, and shades of brown and gray. They are also very smart and can learn tricks easily. They are also very social and live in herds of up to 20-30 individuals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 50, + 151, + 252, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 151, + 252, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 151, + 252, + 185 + ], + "type": "text", + "content": "Alpaca-LoRA: Alpacas are a species of South American camelid, related to the llama and the vicuna. They are a domesticated version of the wild vicuna and are primarily raised for their fine fleece, which is used to make clothing. Alpacas are typically smaller than llamas, with a range of sizes from 12 to 15 hands high. They are known for their gentle nature and are often kept as pets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 50, + 187, + 252, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 187, + 252, + 215 + ], + "spans": [ + { + "bbox": [ + 50, + 187, + 252, + 215 + ], + "type": "text", + "content": "LLaMA-Adapter: Alpacas are members of the camelid family and are native to the Andes Mountains of South America. They are typically between 3 and 4 feet tall and weigh between 100 and 200 pounds. They have long, coarse, and dense coats that come in a variety of colors, including white, black, brown, and grey." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "spans": [ + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "text", + "content": "LLaMA-Excitor: Alpacas are domesticated animals that are related to llamas and are native to South America. They are typically between 100 and " + }, + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "inline_equation", + "content": "120~\\mathrm{cm}" + }, + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "text", + "content": " tall and weigh between 100 and " + }, + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "inline_equation", + "content": "150~\\mathrm{kg}" + }, + { + "bbox": [ + 50, + 217, + 252, + 263 + ], + "type": "text", + "content": ". Alpacas have long, soft fleece that is used for making clothing, blankets, and other products. They are also used for their meat and their dung, which is used as fuel. Alpacas are herd animals and are known for their gentle and calm nature. They are also very intelligent and can be trained to perform certain tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 73, + 467, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 73, + 467, + 83 + ], + "spans": [ + { + "bbox": [ + 267, + 73, + 467, + 83 + ], + "type": "text", + "content": "Instruction: Write a conversation between the sun and Pluto." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 267, + 85, + 298, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 85, + 298, + 92 + ], + "spans": [ + { + "bbox": [ + 267, + 85, + 298, + 92 + ], + "type": "text", + "content": "LLaMA7B:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 267, + 92, + 350, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 92, + 350, + 99 + ], + "spans": [ + { + "bbox": [ + 267, + 92, + 350, + 99 + ], + "type": "text", + "content": "The sun: Hey Pluto, how are you?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 267, + 101, + 305, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 101, + 305, + 107 + ], + "spans": [ + { + "bbox": [ + 267, + 101, + 305, + 107 + ], + "type": "text", + "content": "Pluto: I'm fine." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 267, + 109, + 342, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 109, + 342, + 116 + ], + "spans": [ + { + "bbox": [ + 267, + 109, + 342, + 116 + ], + "type": "text", + "content": "The sun: I'm glad to hear that." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 267, + 118, + 328, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 118, + 328, + 124 + ], + "spans": [ + { + "bbox": [ + 267, + 118, + 328, + 124 + ], + "type": "text", + "content": "Pluto: I'm glad to hear th" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 267, + 126, + 350, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 126, + 350, + 133 + ], + "spans": [ + { + "bbox": [ + 267, + 126, + 350, + 133 + ], + "type": "text", + "content": "The sun: I'm glad to hear that too." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 86, + 403, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 86, + 403, + 92 + ], + "spans": [ + { + "bbox": [ + 356, + 86, + 403, + 92 + ], + "type": "text", + "content": "Full Fine-Tuning:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 92, + 423, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 92, + 423, + 99 + ], + "spans": [ + { + "bbox": [ + 356, + 92, + 423, + 99 + ], + "type": "text", + "content": "The sun said, \"I'm so hot!\"" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 101, + 423, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 101, + 423, + 107 + ], + "spans": [ + { + "bbox": [ + 356, + 101, + 423, + 107 + ], + "type": "text", + "content": "The sun said, \"I'm so hot!\"" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 109, + 423, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 109, + 423, + 116 + ], + "spans": [ + { + "bbox": [ + 356, + 109, + 423, + 116 + ], + "type": "text", + "content": "The sun said, \"I'm so hot!\"" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 118, + 423, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 118, + 423, + 125 + ], + "spans": [ + { + "bbox": [ + 356, + 118, + 423, + 125 + ], + "type": "text", + "content": "The sun said, \"I'm so hot!\"" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 356, + 127, + 423, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 127, + 423, + 133 + ], + "spans": [ + { + "bbox": [ + 356, + 127, + 423, + 133 + ], + "type": "text", + "content": "The sun said, \"I'm so hot!\"" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 432, + 86, + 471, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 86, + 471, + 92 + ], + "spans": [ + { + "bbox": [ + 432, + 86, + 471, + 92 + ], + "type": "text", + "content": "Alpaca-LoRA:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 432, + 92, + 542, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 92, + 542, + 105 + ], + "spans": [ + { + "bbox": [ + 432, + 92, + 542, + 105 + ], + "type": "text", + "content": "Sun: Hey Pluto, how's life in the outer solar system?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 432, + 107, + 542, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 107, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 432, + 107, + 542, + 121 + ], + "type": "text", + "content": "Pluto: It's pretty chilly up here! I'm still getting used to being so far away from the sun." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 432, + 123, + 542, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 123, + 542, + 143 + ], + "spans": [ + { + "bbox": [ + 432, + 123, + 542, + 143 + ], + "type": "text", + "content": "Sun: That's understandable. I'm sure you'll get used to it soon enough. Anyway, I'm glad you're still around." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 432, + 144, + 542, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 144, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 432, + 144, + 542, + 159 + ], + "type": "text", + "content": "Pluto: Thanks! It's nice to still be a part of the solar system." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 267, + 152, + 314, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 152, + 314, + 159 + ], + "spans": [ + { + "bbox": [ + 267, + 152, + 314, + 159 + ], + "type": "text", + "content": "LLaMA-Adapter:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 267, + 159, + 345, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 159, + 345, + 167 + ], + "spans": [ + { + "bbox": [ + 267, + 159, + 345, + 167 + ], + "type": "text", + "content": "The Sun: Good morning, Pluto!" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 267, + 167, + 369, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 167, + 369, + 175 + ], + "spans": [ + { + "bbox": [ + 267, + 167, + 369, + 175 + ], + "type": "text", + "content": "Pluto: Good morning, Sun! How are you?" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 267, + 176, + 366, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 176, + 366, + 184 + ], + "spans": [ + { + "bbox": [ + 267, + 176, + 366, + 184 + ], + "type": "text", + "content": "The Sun: I'm doing great! How are you?" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 267, + 186, + 332, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 186, + 332, + 193 + ], + "spans": [ + { + "bbox": [ + 267, + 186, + 332, + 193 + ], + "type": "text", + "content": "Pluto: I'm doing well, too!" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 267, + 195, + 343, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 195, + 343, + 202 + ], + "spans": [ + { + "bbox": [ + 267, + 195, + 343, + 202 + ], + "type": "text", + "content": "The Sun: I'm glad to hear that!" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 267, + 205, + 303, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 205, + 303, + 211 + ], + "spans": [ + { + "bbox": [ + 267, + 205, + 303, + 211 + ], + "type": "text", + "content": "Pluto: Me too!" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 267, + 213, + 376, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 213, + 376, + 221 + ], + "spans": [ + { + "bbox": [ + 267, + 213, + 376, + 221 + ], + "type": "text", + "content": "The Sun: I'm glad we're having such a nice day." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 267, + 222, + 277, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 222, + 277, + 228 + ], + "spans": [ + { + "bbox": [ + 267, + 222, + 277, + 228 + ], + "type": "text", + "content": "day" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 267, + 229, + 304, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 229, + 304, + 236 + ], + "spans": [ + { + "bbox": [ + 267, + 229, + 304, + 236 + ], + "type": "text", + "content": "Pluto: Me too!" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 267, + 237, + 376, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 237, + 376, + 243 + ], + "spans": [ + { + "bbox": [ + 267, + 237, + 376, + 243 + ], + "type": "text", + "content": "The Sun: I'm glad we're having such a nice" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 267, + 244, + 278, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 244, + 278, + 249 + ], + "spans": [ + { + "bbox": [ + 267, + 244, + 278, + 249 + ], + "type": "text", + "content": "day!" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 267, + 250, + 303, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 250, + 303, + 255 + ], + "spans": [ + { + "bbox": [ + 267, + 250, + 303, + 255 + ], + "type": "text", + "content": "Pluto: Me too!" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 378, + 157, + 422, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 157, + 422, + 163 + ], + "spans": [ + { + "bbox": [ + 378, + 157, + 422, + 163 + ], + "type": "text", + "content": "LLaMA-Excitor:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 378, + 163, + 466, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 163, + 466, + 169 + ], + "spans": [ + { + "bbox": [ + 378, + 163, + 466, + 169 + ], + "type": "text", + "content": "Sun: Hey Pluto, how are you doing?" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 378, + 169, + 476, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 169, + 476, + 175 + ], + "spans": [ + { + "bbox": [ + 378, + 169, + 476, + 175 + ], + "type": "text", + "content": "Pluto: I'm doing great! How about you?" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 378, + 176, + 528, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 176, + 528, + 182 + ], + "spans": [ + { + "bbox": [ + 378, + 176, + 528, + 182 + ], + "type": "text", + "content": "Sun: I'm doing well too. It's been a while since we last spoke." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 378, + 183, + 528, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 183, + 528, + 194 + ], + "spans": [ + { + "bbox": [ + 378, + 183, + 528, + 194 + ], + "type": "text", + "content": "Pluto: Yeah, it has been a while. I've been busy exploring the outer reaches of the solar system." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 378, + 194, + 489, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 194, + 489, + 201 + ], + "spans": [ + { + "bbox": [ + 378, + 194, + 489, + 201 + ], + "type": "text", + "content": "Sun: That's great! I'm glad you're having fun." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 378, + 201, + 528, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 201, + 528, + 213 + ], + "spans": [ + { + "bbox": [ + 378, + 201, + 528, + 213 + ], + "type": "text", + "content": "Pluto: I'm having a blast! I'm so glad I'm not stuck in the inner solar system like the other planets." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 378, + 213, + 501, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 213, + 501, + 220 + ], + "spans": [ + { + "bbox": [ + 378, + 213, + 501, + 220 + ], + "type": "text", + "content": "Sun: Yeah, it's nice to have some space to yourself." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 378, + 220, + 496, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 220, + 496, + 226 + ], + "spans": [ + { + "bbox": [ + 378, + 220, + 496, + 226 + ], + "type": "text", + "content": "Pluto: Definitely! So, what have you been up to?" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 378, + 227, + 528, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 227, + 528, + 239 + ], + "spans": [ + { + "bbox": [ + 378, + 227, + 528, + 239 + ], + "type": "text", + "content": "Sun: I've been busy keeping the other planets warm and providing light and heat." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 378, + 239, + 492, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 239, + 492, + 245 + ], + "spans": [ + { + "bbox": [ + 378, + 239, + 492, + 245 + ], + "type": "text", + "content": "Pluto: That's a big job! You must be exhausted." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 378, + 246, + 491, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 246, + 491, + 251 + ], + "spans": [ + { + "bbox": [ + 378, + 246, + 491, + 251 + ], + "type": "text", + "content": "Sun: I am, but it's worth it. I'm glad I can help" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 378, + 252, + 529, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 252, + 529, + 263 + ], + "spans": [ + { + "bbox": [ + 378, + 252, + 529, + 263 + ], + "type": "text", + "content": "Pluto: I'm glad you're doing well. It's been great talking to you." + } + ] + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 75, + 308, + 277, + 380 + ], + "blocks": [ + { + "bbox": [ + 132, + 275, + 460, + 286 + ], + "lines": [ + { + "bbox": [ + 132, + 275, + 460, + 286 + ], + "spans": [ + { + "bbox": [ + 132, + 275, + 460, + 286 + ], + "type": "text", + "content": "Table 1. Instruction-following Comparison between LLaMA-Excitor and other techniques." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 308, + 277, + 380 + ], + "lines": [ + { + "bbox": [ + 75, + 308, + 277, + 380 + ], + "spans": [ + { + "bbox": [ + 75, + 308, + 277, + 380 + ], + "type": "image", + "image_path": "1c7b4dd884401a33c80c0683d0f77da961a50c0e9ba819736eeef25c255d2f71.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "lines": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 287, + 413 + ], + "type": "text", + "content": "Figure 4. Quantitative comparisons between LLaMA-Excitor (BLUE) with other methods, evaluated by GPT-4 [32]." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "content": "prompts, where " + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "content": " is the embedding length and " + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "content": " is the dimension of embeddings. Note that we adopt both the global CLS token and regional patch tokens from the last layer of the CLIP encoder to acquire adequate visual semantics. Then, we calculate multi-modal versions of the Excitor Key and the Exitor Value by combining " + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "inline_equation", + "content": "P_l" + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 434, + 287, + 502 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 102, + 510, + 287, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 510, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 102, + 510, + 287, + 529 + ], + "type": "interline_equation", + "content": "\\operatorname {K e y} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {k}} ^ {\\text {e x c i t o r}} (I) \\right], \\tag {8}", + "image_path": "cdc6ecdaf9f2b7b4b940061fc1e88531c922bebcb2ae675a6047b86be2ef0c5a.jpg" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 97, + 536, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 536, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 97, + 536, + 287, + 555 + ], + "type": "interline_equation", + "content": "\\operatorname {V a l u e} _ {\\text {e x c i t o r}} = \\left[ P _ {l}; W _ {\\mathrm {v}} ^ {\\text {e x c i t o r}} (I) \\right]. \\tag {9}", + "image_path": "681c1b5995503c9fbc841edb67360fcd5d84a5b2a25dfedc056235cfe3984abc.jpg" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": "Unlike the processing of " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "P_{l}" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": ", we adopt low-rank linear projections " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{k}}^{\\mathrm{excitor}}" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{v}}^{\\mathrm{excitor}}" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": ", since we provide the same visual prompt for attention layers " + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "inline_equation", + "content": "l \\in [N - L, N]" + }, + { + "bbox": [ + 46, + 558, + 287, + 713 + ], + "type": "text", + "content": " and different layers require various visual clues for the downstream reasoning. LLaMA-Excitor is an innovative framework for the multi-modal fine-tuning of LLMs. As mentioned in Section 1, most techniques before LLaMA-Excitor involved training a module to perform multi-modal alignment that projects visual embeddings into the textual feature space. This enables LLMs to understand visual tokens and integrate them into the reasoning process. However, the strict alignment constraint may damage the rich semantics of visual embeddings, and their direct feature fusion may" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 380 + ], + "type": "text", + "content": "also cause performance degradation. In contrast, LLaMA-Excitor does not require feature alignment and works on the raw inputs from the frozen image encoder. It does not change the purity of the intermediate textual features and thus avoids any potential damage to the rich semantics of the visual embeddings and the reasoning of LLMs." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 306, + 392, + 387, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 387, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 387, + 407 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 306, + 413, + 466, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 413, + 466, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 413, + 466, + 426 + ], + "type": "text", + "content": "4.1. Language-only Performances" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 306, + 432, + 481, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 481, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 481, + 444 + ], + "type": "text", + "content": "4.1.1 Instruction-following Evaluations" + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 475 + ], + "type": "text", + "content": "In this section, we evaluate the instruction-following capacity of LLaMA-Excitor by responding to instructions." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 304, + 476, + 545, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 523 + ], + "type": "text", + "content": "Dataset. We use Stanford Alpaca [44], which contains 52K instruction-following data, as the training set. Each sample in this dataset is formatted according to a template as follows:" + } + ] + } + ], + "index": 62 + }, + { + "bbox": [ + 309, + 525, + 538, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 525, + 538, + 558 + ], + "spans": [ + { + "bbox": [ + 309, + 525, + 538, + 558 + ], + "type": "text", + "content": "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request." + } + ] + } + ], + "index": 63 + }, + { + "bbox": [ + 310, + 559, + 403, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 559, + 403, + 569 + ], + "spans": [ + { + "bbox": [ + 310, + 559, + 403, + 569 + ], + "type": "text", + "content": "Instruction: {instruction}" + } + ] + } + ], + "index": 64 + }, + { + "bbox": [ + 310, + 569, + 363, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 569, + 363, + 580 + ], + "spans": [ + { + "bbox": [ + 310, + 569, + 363, + 580 + ], + "type": "text", + "content": "Input: {input}" + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 310, + 581, + 383, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 581, + 383, + 591 + ], + "spans": [ + { + "bbox": [ + 310, + 581, + 383, + 591 + ], + "type": "text", + "content": "Response: {output}" + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "content": "where {instruction} is the description of a task, {input} is the context for the task, and {output} is the answer generated by GPT-3.5 [5]." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We develop LLaMA-Excitor based on the original LLaMA codebase with minor modifications. We train LLaMA-Excitor on 8 NVIDIA A100 GPUs for five epochs. The warmup epochs, batch size, learning rate, and weight decay are set to 2, 64, 9e-3, and 0.02, respectively. In this section, we utilize the pretrained LLaMA-7B v1 model, which has " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "N = 32" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " transformer lay" + } + ] + } + ], + "index": 68 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14093" + } + ] + } + ], + "index": 69 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 289, + 199 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 289, + 199 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 289, + 199 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 289, + 199 + ], + "type": "image", + "image_path": "2e252e041efef300d5769f4527b5379e1d3bc009152682997785967fd1692363.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 287, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 240 + ], + "type": "text", + "content": "Figure 5. The relative average performance changes and win-loss situations of fine-tunings compared to original LLaMA-7B on MMLU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": "ers with feature dimension " + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "inline_equation", + "content": "C = 4096" + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": " as the backbone. We apply our excitor blocks with the low-rank dimension " + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{q}}^{\\mathrm{exciton}}" + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": " in the topmost " + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "inline_equation", + "content": "L = 30" + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": " layers and set the length of learnable prompts " + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "inline_equation", + "content": "L = 30" + }, + { + "bbox": [ + 46, + 266, + 286, + 337 + ], + "type": "text", + "content": ". During the inference, we adopt top-p sampling as the default decoding method with a temperature of 0.1 and a top-p = 0.75." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 286, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 286, + 422 + ], + "type": "text", + "content": "We compare LLaMA-Excitor with three competing methods. (1) Full-finetuning, which updates the entire 7B parameters in LLaMA during the end-to-end training. (2) Alpaca-LoRA, the official implementation of low-rank adaptation [15] provided by [44]. (3) LLaMA-Adapter [51], a cutting-edge prefix-tuning method with zero-init attention to reduce forgetting in finetuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 286, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 286, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 286, + 699 + ], + "type": "text", + "content": "Performances. We first use examples in Table 1 to show the strong instruction-following ability of LLaMA-Excitor. The first case requires the models to provide some information about alpacas. LLaMA-Excitor generates reasonable responses comparable to Full-finetuning, demonstrating our method's effectiveness. It's worth noticing that our approach follows the same answering pattern as the prefix-tuning method LLaMA-adapter and the original LLaMA7B (although Excitor provides more details). This phenomenon is reasonable since prefix-tuning methods use additional learnable prompts to guide the generation process of LLMs. They do not change the internal reasoning process, just like our Excitor. This characteristic may also be the foundation of avoiding catastrophic forgetting. For the second task, the LLMs are asked to create a dialogue between the Sun and Pluto. Out of all the models that are tested, only Excitor and LoRA correctly recognize the characters as planets within the Solar System and incorporate their identities into the conversation. However, Excitor stands out as it can generate lengthy, high-quality conversations rich in interesting content. (Please see Appendix A for the full comparison including dialog generation, code generation, and question answering.)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 286, + 713 + ], + "type": "text", + "content": "To evaluate the performance of instruction-following" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 276 + ], + "type": "text", + "content": "models quantitatively, we refer to the research conducted by Chiang et al. [8]. In this evaluation, we use GPT-4 [32] to assess the quality of responses to 80 randomly selected questions from AlpacaEval [25]. However, we notice a bias in GPT-4's scoring, where it tends to give higher scores to the first response. To overcome this, we interchange the position of two responses and use the average score for the final rating. Results are shown in Figure 4. We notice that Excitor overwhelms the Full-Finetuning (73/80). The reason behind this is that AlpacaEval is an out-of-distributed (OOD) test for our training set, and full-finetuning may easily overfit the training set. When compared with AlpacaLoRA and LLaMA-Adapter, our Excitor still gains the majority of \"Win\", demonstrating our indirect feature interaction is at least comparable with or better than predominant PEFT methods. (Please refer to Appendix B for instances on AlpacaEval.)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 279, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 364 + ], + "type": "text", + "content": "As this paper is a preliminary attempt to study indirect feature interaction in LLMs and the 52K instruction-following data used in our study is automatically generated, it is possible to further enhance the performance of LLaMA-Excitor by improving the quality of the training data, adjusting the hyper-parameters to more reasonable values or exploring larger LLaMA models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 399, + 545, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 422 + ], + "type": "text", + "content": "4.1.2 The Impact of Fine-tuning Techniques on Inherent Abilities" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": "One of the motivations for introducing the indirect feature interaction in LLaMA-Excitor is to reduce the influence of fine-tuning on low-quality or non-targeted datasets. To quantitatively analyze this characteristic, we evaluate the fine-tuned models on MMLU [13], which is currently the most comprehensive benchmark for assessing the multi-task accuracy of text models. This benchmark includes 57 tasks across various fields, such as basic math, American history, computer science, and law. The model must possess extensive world knowledge and problem-solving capabilities to achieve high accuracy on this test. We report the relative performance changes of PEFT methods compared with the original LLaMA-7B and how many of the 57 evaluations PEFT methods outperform LLaMA-7B in Figure 5. We find that for models fine-tuned on Alpaca-52k, only Excitor can avoid reducing the average score of 57 tasks in MMLU, and it even shows an increase in performance " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "(+3.12\\%)" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ". As expected, the full-finetuning performance suffers the most severe decline as its parameters are most significantly changed " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "(-6.63\\%)" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": ", while the LLaMA-Adapter, which is based on prefix-tuning, experiences the least decline " + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "inline_equation", + "content": "(-3.02\\%)" + }, + { + "bbox": [ + 304, + 438, + 545, + 713 + ], + "type": "text", + "content": " besides Excitor. (We provide results based on LLaMA2-7B in Appendix C)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14094" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 396, + 91 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 396, + 91 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 396, + 91 + ], + "type": "text", + "content": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 96, + 396, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 96, + 396, + 131 + ], + "spans": [ + { + "bbox": [ + 70, + 96, + 396, + 131 + ], + "type": "text", + "content": "LLaMA-Excitor: An airplane is flying through the sky, with its landing gear down, as it approaches for a landing at an airport. The airplane is dominating the scene, with its bright red and white colors clearly visible against the backdrop of the cloudy sky. The airplane is about to touch down on the runway, creating a moment of excitement and anticipation for the passengers and crew on board." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 136, + 313, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 136, + 313, + 146 + ], + "spans": [ + { + "bbox": [ + 71, + 136, + 313, + 146 + ], + "type": "text", + "content": "Human Annotation: A large passenger airplane flying through the air." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 151, + 396, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 151, + 396, + 170 + ], + "spans": [ + { + "bbox": [ + 70, + 151, + 396, + 170 + ], + "type": "text", + "content": "Instruction: Please answer me based on this image. Generate a caption that details what happened in the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 175, + 396, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 175, + 396, + 208 + ], + "spans": [ + { + "bbox": [ + 70, + 175, + 396, + 208 + ], + "type": "text", + "content": "LLaMA-Excitor: In the image, a group of birds is standing together on a giraffe. One of the bird is leaning over to nibble on the giraffe's fur and the giraffe seems to be enjoying the interaction. The scene captures the playful nature of these animals and the unexpected interaction between a giraffe and a bird." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 213, + 303, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 213, + 303, + 223 + ], + "spans": [ + { + "bbox": [ + 71, + 213, + 303, + 223 + ], + "type": "text", + "content": "Human Annotation: Little birds sitting on the shoulder of a giraffe." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 73, + 520, + 147 + ], + "blocks": [ + { + "bbox": [ + 425, + 73, + 520, + 147 + ], + "lines": [ + { + "bbox": [ + 425, + 73, + 520, + 147 + ], + "spans": [ + { + "bbox": [ + 425, + 73, + 520, + 147 + ], + "type": "image", + "image_path": "5c43262a5b3ceb4ac8de19634fb9be3a1c902c816d64c997d542cc141adb2fc4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 425, + 155, + 520, + 219 + ], + "blocks": [ + { + "bbox": [ + 425, + 155, + 520, + 219 + ], + "lines": [ + { + "bbox": [ + 425, + 155, + 520, + 219 + ], + "spans": [ + { + "bbox": [ + 425, + 155, + 520, + 219 + ], + "type": "image", + "image_path": "90c5cd3fb6434b6de7be91207a3888bcadc5ca0d1b009cf2dfa7eba77ec93057.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 74, + 253, + 261, + 356 + ], + "blocks": [ + { + "bbox": [ + 133, + 237, + 459, + 248 + ], + "lines": [ + { + "bbox": [ + 133, + 237, + 459, + 248 + ], + "spans": [ + { + "bbox": [ + 133, + 237, + 459, + 248 + ], + "type": "text", + "content": "Table 2. Examples demonstrating LLaMA-Excitor's visual instruction following capacity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 253, + 261, + 356 + ], + "lines": [ + { + "bbox": [ + 74, + 253, + 261, + 356 + ], + "spans": [ + { + "bbox": [ + 74, + 253, + 261, + 356 + ], + "type": "table", + "html": "
MethodData ScaleCOCO Caption
PTFTBLEU@4CIDEr
ClipCap [31]0M0.6M33.5113.1
VL-PET [16]0M0.6M-121.7
LLaMA-AdapterV2 [12]0M0.6M36.2122.2
Qwen-vl-chat [4]1.4B0.6M-131.9
mPLUG-Owl2 [49]348M0.6M-137.3
BLIP [21]14M0.6M40.4136.7
Flamingo [2]1.8B0.6M-138.1
BLIP-2 [22]129M0.6M43.7145.3
LLaMA-Excitor0M0.6M49.7157.5
", + "image_path": "e74975a87005792dcd8c132c798ce8a0a3a24254810576a8454ea2ea543ef900.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 365, + 287, + 387 + ], + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 387 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art image captioning methods on COCO Caption." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 394, + 197, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 197, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 197, + 406 + ], + "type": "text", + "content": "4.2. Multi-Modal Performances" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 485 + ], + "type": "text", + "content": "In this section, we evaluate the visual instruction following the capacity of LLaMA-Excitor by responding to paired vision-language instructions and demonstrate our indirect feature interaction, which unifies the modeling of language-only tuning and multi-modal tuning, providing a powerful low-budget way to perform vision-language tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 502, + 206, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 206, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 206, + 514 + ], + "type": "text", + "content": "4.2.1 Image Captioning Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": "Implementation Details. We first evaluate our Excitor on COCO Caption [7], which contains 0.6M training image caption data (120k images with 5 captions per image) over a wide range of distributions. We apply frozen CLIP/14-L [38] as the image encoder, the visual embedding dimension " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "D = 768" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": ", low-rank dimension " + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 46, + 521, + 287, + 605 + ], + "type": "text", + "content": ". We keep other hyper-parameters the same as the section 4.1.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Performances. We compare our Excitor with cutting-edge image captioning methods in Table 3. We find that Excitor achieves a stunning result, significantly surpassing the previous SOTA method BILP-2 [22] by " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "+6" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " BLEU@4 and " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "+12.2" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " CIDEr. Especially considering that we do not apply complex multi-modal alignment modules and have only 0.6M training data. By contrast, BLIP-2 adopts a costly pre-training stage on an additional 129M dataset, including VisualGenome [19], Conceptual Captions [42, 43] and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 255, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 399 + ], + "type": "text", + "content": "LAION [6]. We visualize the results of image captioning in Appendix D. However, the image captions provided in COCO are concise and lack sufficient detail in describing the content of images. To further unleash Exitor's performance in image captioning, we continue fine-tuning Exitor on LLaVA665k [27], a higher-quality dataset with detailed picture descriptions. We provide several examples from the COCO test set in Table 2. It shows that image captions generated by Excitor can accurately cover the content of human annotations and provide richer details. (e.g., (1) The airplane's landing gear is down, (2) A bird is pecking at the back of a giraffe)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 418, + 460, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 460, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 460, + 430 + ], + "type": "text", + "content": "4.2.2 Performances on ScienceQA" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 581 + ], + "type": "text", + "content": "Implementation Details. We evaluate our Excitor on ScienceQA [40], which contains 21k multimodal multiple choice questions with rich domain diversity across 3 subjects, 26 topics, 127 categories, and 379 skills. We train Excitor on ScienceQA train split from scratch, using the combination of Chain-of-Thought(CoT) and direct answer prediction. CoT requires the model to predict the solution first and then generate the answer choice based on the solution. During training, we randomly require the model to first generate solutions or directly predict answers in each iteration. The model is asked to answer the question directly in the evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 713 + ], + "type": "text", + "content": "Table 4 reports the performances of the top methods on the official leaderboard of the ScienceQA and Excitor. The current SOTA method LLaVA [28] is pretrained on another 558k visual-language dataset and fine-tuned on ScienceQA. Also, they allow all LLM's parameters to be updated during fine-tuning. Our Excitor demonstrates comparable performance with LLaVA w/o pretraining (only -0.4% performance gap). Note that LLaMA-Excitor is a PEFT method (with a frozen LLM) without CoT, and LLaVA utilizes a larger backbone LLaMA-13B than our LLaMA-7B. To further unlock the potential of Excitor, we introduce a vari" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14095" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 523, + 272 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 523, + 272 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 523, + 272 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 523, + 272 + ], + "type": "table", + "html": "
ModelAverageSubjectContext ModalityGrade
NATSOCLANTXTIMGNOG1-6G7-12
Human [40]88.4090.2384.9787.4889.6087.5088.1091.5982.42
UnifiedQACoT74.1171.0076.0478.9166.4266.5381.8177.0668.82
GPT-3CoT75.1775.4470.8778.0974.6867.4379.9378.2369.68
ChatGPTCoT [1]78.3178.8270.9883.1877.3767.9286.1380.7274.03
GPT-4CoT [32]83.9985.4872.4490.2782.6571.4992.8986.6679.04
MM-COT [53]84.9187.5277.1785.8287.8882.9086.8384.6585.37
LLaVACoT [28]90.9290.3695.9588.0089.4988.0090.6690.9390.90
LLaVACoT (w/o pretrain) [28]85.81--------
DFAF [11]60.7264.0348.8263.5565.8854.4964.1157.1267.17
ViLT [18]61.1460.4863.8960.2763.2061.3857.0060.7261.90
Patch-TRM [30]61.4265.1946.7965.5566.9655.2864.9558.0467.50
VisualBERT [23, 24]61.8759.3369.1861.1862.7162.1758.5462.9659.92
UnifiedQA [17]70.1268.1669.1874.9163.7861.3877.8472.9865.00
GPT-3 [5]74.0475.0466.5978.0074.2465.7479.5876.3669.87
LLaMA-Adapter85.1984.3788.3084.3683.7280.3286.9085.8384.05
LLaMA-Excitor85.4185.7092.3582.8283.4384.5686.2785.6584.64
LLaMA-Excitor@336px + LoRA88.3987.1991.3387.0990.4285.2088.6488.3588.42
", + "image_path": "ab9354d92b2fc2da656a6f34226e4b4bc207e1496ad8b0d556ad391135fb9a6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 323, + 288, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 384 + ], + "type": "text", + "content": "ant that utilizes a more powerful image encoder (CLIP-L/14@336) and adds LoRA blocks to the original LLM's parameters to reduce the gap with LLaVA's full parameter updating. It brings " + }, + { + "bbox": [ + 46, + 323, + 288, + 384 + ], + "type": "inline_equation", + "content": "+2.58\\%" + }, + { + "bbox": [ + 46, + 323, + 288, + 384 + ], + "type": "text", + "content": " performance gain compared with LLaVA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 393, + 117, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 117, + 405 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 117, + 405 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 73, + 421, + 261, + 496 + ], + "blocks": [ + { + "bbox": [ + 46, + 280, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 547, + 304 + ], + "type": "text", + "content": "Table 4. Question Answering Accuracy (%) on ScienceQA's [40] test set. We report GPT-3 [5], ChatGPT [1], and GPT-4 [32] for zero-shot inference. CoT denotes utilizing chain-of-thought for question answering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 421, + 261, + 496 + ], + "lines": [ + { + "bbox": [ + 73, + 421, + 261, + 496 + ], + "spans": [ + { + "bbox": [ + 73, + 421, + 261, + 496 + ], + "type": "table", + "html": "
VariantACC
(a) CLIP-16/B83.11
(b) CLIP-14/L85.41
(c) CLIP-14/L@336px85.87
(d) CLIP-14/L@336px + Full-Finetuning83.20
(e) CLIP-14/L@336px + LoRA88.39
", + "image_path": "9eef5e895ab3913ae47bdfdea947db17c7db5d5c34b377ce7abb6d54df6aa3e2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 505, + 265, + 516 + ], + "lines": [ + { + "bbox": [ + 68, + 505, + 265, + 516 + ], + "spans": [ + { + "bbox": [ + 68, + 505, + 265, + 516 + ], + "type": "text", + "content": "Table 5. Analysis of the effectiveness of each module." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "We ablate several design choices on ScienceQA in Table 5. (a), (b) and (c) study the influence of the capacity of image encoder. The best CLIP encoder yields " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "85.41\\%" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " and is " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "+2.76\\%" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " higher than the basic version. In (d) and (e), we study to combine Excitor, which focuses on instruction-following, with other techniques that update LLM's reasoning process. We find LoRA can bring " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "2.98\\%" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " performance gain for Excitor, demonstrating the potential of combining Excitor with existing finetuning methods in multimodal usage. (However, in the text-only scenario, we encounter " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "-4.14\\%" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " performance gap compared with the original LLaMA-7B in the MMLU test by combining Excitor and LoRA.) Please see Appendix E for more ablations on the design of low-rank linear layers (i.e., " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{q}}^{Excitor}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{k}}^{Excitor}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{v}}^{Excitor}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ") in Excitor Blocks under single" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "text", + "content": "modal and multi-modal scenarios, the choice of low-rank dimension " + }, + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "text", + "content": ", and the number of layers " + }, + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 323, + 547, + 361 + ], + "type": "text", + "content": " with Excitor blocks inserted." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 370, + 379, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 370, + 379, + 382 + ], + "spans": [ + { + "bbox": [ + 305, + 370, + 379, + 382 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 521 + ], + "type": "text", + "content": "In this paper, we study the indirect feature interaction in fine-tuning LLMs into instruction-following models and propose LLaMA-Exictor, a PEFT method that demonstrates outstanding instruction-following capacity compared with predominant techniques meanwhile reducing the forgetting of LLMs' inherent abilities. Moreover, Excitor also shows its potential in vision-language tasks by unifying the modeling of visual instruction-following and language-only instruction-following. Impressively, Excitor achieves strong multi-modal performance without the need to train complex alignment modules." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 547, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 618 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 618 + ], + "type": "text", + "content": "This project is a work in progress, and some directions can be explored. The first one is the adaptability to LLMs other than LLaMA. The second direction is to explore more effective ways to introduce multi-scale visual prompts (since we only apply the features from the last layer of CLIP at the current stage). Last but not least, Excitor has the potential to fine-tune vision-only models into vision-language models, just like our implementation for LLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 628, + 378, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 628, + 378, + 642 + ], + "spans": [ + { + "bbox": [ + 306, + 628, + 378, + 642 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 647, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 547, + 709 + ], + "type": "text", + "content": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Postdoctoral Excellent Program (Grant No. 2022234)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14096" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 260, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 260, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 260, + 102 + ], + "type": "text", + "content": "[1] Chatgpt. https://chat.openai.com, 2023.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 167 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 202 + ], + "type": "text", + "content": "[3] Anonymous. Videodistill: Language-aware vision distillation for video question answering. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 287, + 257 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 323 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 380 + ], + "type": "text", + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3558-3568, 2021. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 434 + ], + "type": "text", + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arxiv 2015. arXiv preprint arXiv:1504.00325, 2015. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "text", + "content": "[8] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 436, + 287, + 501 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 546 + ], + "type": "text", + "content": "[9] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. arXiv preprint arXiv:1901.02860, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 590 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 656 + ], + "type": "text", + "content": "[11] Peng Gao, Zhengkai Jiang, Haoxuan You, Pan Lu, Steven CH Hoi, Xiaogang Wang, and Hongsheng Li. Dynamic fusion with intra-and inter-modality attention flow for visual question answering. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6639-6648, 2019. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Peng Gao, Jiaming Han, Renrui Zhang, Ziyi Lin, Shijie Geng, Aojun Zhou, Wei Zhang, Pan Lu, Conghui He, Xi-angyu Yue, et al. Llama-adapter v2: Parameter-efficient visual instruction model. arXiv preprint arXiv:2304.15010, 2023.7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 216 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 545, + 270 + ], + "type": "text", + "content": "[16] Zi-Yuan Hu, Yanyang Li, Michael R Lyu, and Liwei Wang. Vl-pet: Vision-and-language parameter-efficient tuning via granularity control. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3010-3020, 2023. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 326 + ], + "type": "text", + "content": "[17] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord, Peter Clark, and Hannaneh Hajishirzi. Unifiedqa: Crossing format boundaries with a single qa system. In Findings of the Association for Computational Linguistics (EMNLP), pages 1896-1907, 2020. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[18] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In Proceedings of the 38th International Conference on Machine Learning (ICML), pages 5583-5594, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 446 + ], + "type": "text", + "content": "[19] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[20] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 493, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 546 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 592, + 545, + 635 + ], + "type": "text", + "content": "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "text", + "content": "[24] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. What does bert with vision look at? In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), pages 5265–5275, 2020. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tat-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14097" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "sunori B. Hashimoto. Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval, 2023.6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[26] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 262 + ], + "type": "text", + "content": "[29] Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Lam Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. arXiv preprint arXiv:2110.07602, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 330 + ], + "type": "text", + "content": "[30] Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. In The 35th Conference on Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 364 + ], + "type": "text", + "content": "[31] Ron Mokady, Amir Hertz, and Amit H Bermano. Clipcap: Clip prefix for image captioning. arXiv preprint arXiv:2111.09734, 2021. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 387 + ], + "type": "text", + "content": "[32] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2023. 5, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 389, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 453 + ], + "type": "text", + "content": "[33] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730-27744, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 488 + ], + "type": "text", + "content": "[34] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[35] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[36] Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Nondestructive task composition for transfer learning. arXiv preprint arXiv:2005.00247, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 612 + ], + "type": "text", + "content": "[37] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 678 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[40] Tanik Saikh, Tirthankar Ghosal, Amish Mittal, Asif Ekbal, and Pushpak Bhattacharyya. Scienceqa: A novel resource for question answering on scholarly articles. International Journal on Digital Libraries, 23(3):289-301, 2022. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[41] Timo Schick and Hinrich Schütze. Exploiting cloze questions for few shot text classification and natural language inference. arXiv preprint arXiv:2001.07676, 2020. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 228 + ], + "type": "text", + "content": "[42] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 293 + ], + "type": "text", + "content": "[43] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 337 + ], + "type": "text", + "content": "[44] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Stanford alpaca: An instruction-following llama model, 2023. 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 392 + ], + "type": "text", + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 448 + ], + "type": "text", + "content": "[46] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: open and efficient foundation language models, 2023. URL https://arxiv.org/abs/2302.13971, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 491 + ], + "type": "text", + "content": "[47] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. Advances in Neural Information Processing Systems, 34:200-212, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "text", + "content": "[48] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. arXiv preprint arXiv:2212.10560, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 590 + ], + "type": "text", + "content": "[49] Qinghao Ye, Haiyang Xu, Jiabo Ye, Ming Yan, Haowei Liu, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mplug-owl2: Revolutionizing multi-modal large language model with modality collaboration. arXiv preprint arXiv:2311.04257, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 635 + ], + "type": "text", + "content": "[50] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning. arXiv preprint arXiv:2303.10512, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 637, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 637, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 637, + 545, + 689 + ], + "type": "text", + "content": "[51] Renrui Zhang, Jiaming Han, Aojun Zhou, Xiangfei Hu, Shilin Yan, Pan Lu, Hongsheng Li, Peng Gao, and Yu Qiao. Llama-adapter: Efficient fine-tuning of language models with zero-init attention. arXiv preprint arXiv:2303.16199, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[52] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14098" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 196 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[53] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 196 + ], + "type": "text", + "content": "[54] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14099" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_content_list.json b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..17380ee2bc992be6a8a907790e751de46b4e5e2d --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_content_list.json @@ -0,0 +1,1319 @@ +[ + { + "type": "text", + "text": "LMDrive: Closed-Loop End-to-End Driving with Large Language Models", + "text_level": 1, + "bbox": [ + 109, + 128, + 857, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Shao $^{1,2}$ Yuxuan Hu $^{5}$ Letian Wang $^{4}$ Guanglu Song $^{2}$ Steven L. Waslander $^{4}$ Yu Liu $^{2,3}$ Hongsheng Li $^{1,3,5}$ $^{1}$ CUHK MMLab $^{2}$ SenseTime Research \n $^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ University of Toronto $^{5}$ CPII under InnoHK", + "bbox": [ + 114, + 178, + 844, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 292, + 313, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant recent progress in the field of autonomous driving, modern methods still struggle and can incur serious accidents when encountering long-tail unforeseen events and challenging urban scenarios. On the one hand, large language models (LLM) have shown impressive reasoning capabilities that approach \"Artificial General Intelligence\". On the other hand, previous autonomous driving methods tend to rely on limited-format inputs (e.g., sensor data and navigation waypoints), restricting the vehicle's ability to understand language information and interact with humans. To this end, this paper introduces LMDrive, a novel language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive uniquely processes and integrates multi-modal sensor data with natural language instructions, enabling interaction with humans and navigation software in realistic instructional settings. To facilitate research in language-based closed-loop autonomous driving, we also publicly release the corresponding dataset which includes approximately 64K instruction-following data clips, and the LangAuto benchmark that tests the system's ability to handle complex instructions and challenging driving scenarios. Extensive closed-loop experiments are conducted to demonstrate LMDrive's effectiveness. To the best of our knowledge, we're the very first work to leverage LLMs for closed-loop end-to-end autonomous driving. Code is available on our webpage.", + "bbox": [ + 75, + 325, + 473, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remarkable progress in autonomous driving has been witnessed in recent years with an increasing number of commercial autonomous vehicles (AVs) deployed on public roads. Generally, state-of-the-art autonomous driving systems can be categorized into two primary approaches: 1) a modular approach where the system is decomposed into several sub-modules such as perception, prediction, and planning, and fixed interfaces are designed to integrate them", + "bbox": [ + 75, + 755, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83bbd20f10b86f4719a2d653a69d39e35ee2ceda6b501c09734892af867d0112.jpg", + "image_caption": [ + "Figure 1. We present LMDrive, the first language-guided closed-loop end-to-end driving framework. LMDrive takes as input the language instruction and multi-modal multi-view sensor data, and outputs control signals in real time to drive in complex scenarios." + ], + "image_footnote": [], + "bbox": [ + 503, + 290, + 890, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "together [2, 26, 41, 57]; and 2) an end-to-end approach that directly converts sensor data to control signals via a neural network [18, 39]. While both of these approaches are widely adopted and constantly making breakthroughs on challenging benchmarks, both of them share a limitation in that they solely rely on fixed-format inputs such as the sensor data, target waypoints, and action commands, which restricts the agent's ability to comprehend multi-modal information and to interact with humans and the environment. On the other hand, large language models (LLMs) have shown an impressive range of capabilities that approach \"Artificial General Intelligence.\" This encompasses language comprehension, knowledge retrieval, and reasoning. Such capabilities could greatly enhance the safety, controllability, and explainability of autonomous agents. In this work, we seek to answer the question for the first time: Can we build cognitive autonomous driving systems on top of LLMs, that can interact with human passengers or navigation software simply by natural language?", + "bbox": [ + 496, + 553, + 893, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Making autonomous systems understand natural language opens profound opportunities for advanced reasoning in complex scenarios and efficient interaction with humans, addressing many previously non-trivial problems. To name", + "bbox": [ + 498, + 840, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 99, + 886, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15120", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a few: 1) in long-tail unforeseen events and challenging urban situations (e.g., complex and dense intersections) where modern AV systems typically struggle [47] or even incur serious accidents [43], the language-aware AVs can easily survive by following navigation instructions from passengers or navigation software. 2) adapting to passengers' sudden notices (e.g., small objects missed by perception systems) using natural language, previously difficult and rule-dependent.", + "bbox": [ + 75, + 90, + 470, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Toward these appealing properties, many pioneering works have explored the potential of using large language models to enhance the AV system's reasoning abilities, interpretability, and overall performance in open-loop settings. One of the most common strategies [6, 11, 32, 36] is to 1) first use LLMs to transform the scene perception results and navigation commands into textual descriptions; 2) feed these textural descriptions into LLMs to generate textual driving decisions; and then 3) transfer textual driving decisions into executable control commands. While good preliminary results are shown, this type of approach, where different LLMs tackle sub-tasks individually, is hard to be trained in an end-to-end manner, loses the capability to scale with a large amount of data, and is not robust to perception errors and uncertainties. For example, since the LLMs in the latter two stages do not have access to the sensor data, inaccurate or missed detections in the first stage can lead to large accumulative errors in the latter stages. Towards addressing these issues, end-to-end language-based driving methods [52] have been proposed. However, all of these methods undergo training and evaluation in the open-loop setting, where actions are generated and evaluated against the expert actions, but not executed in the actual environments. Notably, when executing navigation instructions such as \"turn right\", the AV agent should not only generate a sequence of actions, but also consider the changes that the actions bring to the environment. The absence of closed-loop evaluation leads to insufficient consideration of critical issues such as cumulative errors, human-robot interaction, and temporal consistency of actions, which makes the resulting methods difficult to scale beyond a short time horizon making them ineffective in actual systems. To the best of our knowledge, there is no existing paper that leverages LLMs for closed-loop end-to-end autonomous driving.", + "bbox": [ + 75, + 227, + 472, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce LMDrive, an instruction-following multi-modal LLM model for end-to-end closed-loop autonomous driving. The proposed model can process camera-LiDAR sensor data, comprehend driving instructions in natural language, and directly generate vehicle control signals. A pre-trained LLM model is adopted and kept frozen to maintain its reasoning capability. To adapt the LLM for autonomous driving, multiple camera-LiDAR data encoders and learnable input/output adapters are integrated. A pre-training strategy specifically designed for the driving task is also introduced for the multi-modal", + "bbox": [ + 75, + 737, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vision encoder. To facilitate the training of LMDrive in a closed-loop setting, we develop a language-guided driving dataset based on the CARLA simulator [12], which simulates the dynamic world and realistic challenging scenarios. To better test the driving model's capability in realistic conversation scenarios, where the instructions might come from humans and navigation software, we 1) consider both navigation instruction and notice instructions; 2) diversify instructions into various phrases via prompt engineering on LLMs; 3) incorporate misleading and unreasonable instructions which are infeasible due to safety concerns or traffic regulations; 4) extend the navigation instructions to include multiple consecutive segments, such as \"Take a left here, then another left at the next one\". Besides, we provide the corresponding LangAuto evaluation benchmark and the pre-trained LMDrive model for reproducibility, and we hope these can facilitate further research in end-to-end closed-loop language-based autonomous driving.", + "bbox": [ + 496, + 90, + 890, + 358 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, this paper makes the following contributions:", + "bbox": [ + 498, + 359, + 890, + 388 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel end-to-end, closed-loop, language-based autonomous driving framework, LMDrive, which interacts with the dynamic environment via multi-modal multi-view sensor data and natural language instructions.", + "- We provide a dataset with about 64K data clips, where each clip includes one navigation instruction, several notice instructions, a sequence of multi-modal multi-view sensor data, and control signals. The duration of the clip spans from 2 to 20 seconds.", + "- We present the benchmark LangAuto for evaluating the autonomous agents that take language instructions as navigation inputs, which include misleading/long instructions and challenging adversarial driving scenarios.", + "- We conduct extensive closed-loop experiments to demonstrate the effectiveness of the proposed framework, and analyze different components of LMDrive to shed light on continuing research along this direction." + ], + "bbox": [ + 500, + 390, + 890, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 657, + 650, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-End Autonomous Driving", + "text_level": 1, + "bbox": [ + 500, + 683, + 792, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Much progress [7, 8] has been achieved recently in the field of end-to-end autonomous driving. UniAD [18] devised a framework that incorporates full-stack driving tasks and utilizes query-unified interfaces to communicate between different tasks. ThinkTwice [22] designed a Look Module to retrieve information from critical regions and utilize the features to refine the coarse prediction. ReasonNet [39] exploited temporal and global information of the driving scene to improve perception performance and benefit occlusion detection. InterFuser [38] proposed a transformer-based framework to fully fuse and process information from multi-modal multi-view sensors for comprehensive scene understanding. TCP [51] proposed an ap", + "bbox": [ + 496, + 705, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proach that integrates the two branches for trajectory planning and direct control by involving a novel multi-step prediction. LAV [3] introduced supervisory tasks for learning viewpoint-invariant representation, offering richer supervision at training and aiding complex reasoning during inference. Besides imitation training, various approaches incorporate reinforcement learning strategies. Latent DRL [44] trained in a supervised way to get a latent representation of the observation and conducts reinforcement learning using the representation as the input. Roach [54] employed a reinforcement learning agent with privileged access to environmental information and distilled a model as the final agent. ASAPRL [50] and TaEcRL [56] exploited abstracted skills to effectively improve reinforcement learning efficiency and the final performance by facilitating effective exploration and reward signaling. ScenerioNet [25] and TrafficGen [14] enhanced driving performance via generating and testing on diverse driving scenarios. However, these end-to-end methods lack the ability to verbally or textually interact with humans (passengers), and usually have low explainability in the decision-making process.", + "bbox": [ + 76, + 90, + 472, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. LLMs in Driving Tasks", + "text_level": 1, + "bbox": [ + 76, + 415, + 292, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emerging advancements in large language models (LLMs) [10, 19, 21, 45, 46] have been witnessed over the last few months. Furthermore, vision large language models (VLLM) [15, 28, 40] further introduce vision encoders and open the doors for LLMs to interpret not only textual data but also images and data in other modalities [5, 29, 58]. In the field of autonomous driving (AD), recent research has integrated LLMs into the AD system for better explainability and natural interaction with humans. Some studies adopt the visual language model approach, which can handle multi-modal input data and provide the textual description as well as the control signal for the driving scenarios. For example, DRIVEGPT4 [52] proposed a multimodal LLM framework, which takes a sequence of frames as input, then generates responses to human inquiries and predicts control signals for the next step. However, since the framework lacks an input command, the predicted control can not follow the specific navigation command, which denotes that the framework is hard to deploy in real scenarios. Meanwhile, more researchers focus on transforming the driving situations into textual descriptions as the input for the LLM, for directly interpreting and reasoning about comprehensive driving situations. In this thread of works, GPT-Driver [32] reformulated motion planning as the task of natural language modeling by converting heterogeneous scene input to language tokens. LanguageMPC [36] leveraged a LLM to reason the complex scenarios and output high-level driving decisions. Then the method tunes a parameter matrix to convert the decision into the low-level control signals. LLM-Driver [6] utilized the numeric vector as the", + "bbox": [ + 76, + 439, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99916684b9b3a582b272613337d3e86d4bcd7b50afde67059f2e532c163a0b9c.jpg", + "image_caption": [ + "Navigation instruction: Just move to the left and get ready to leave the highway." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 885, + 145 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0aec77596cf85f55a6a5998f1666fcd1e1ada32308eda38bd2d5a9fe0e92bfdb.jpg", + "image_caption": [ + "Navigation instruction: Turn left at the next T-junction", + "Notice instruction: Please watch out for the pedestrians up ahead." + ], + "image_footnote": [], + "bbox": [ + 503, + 161, + 885, + 218 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1fc85fac44906cdcf24e88d93be1e44c83f39d10fcc42db41610e0573a554bb5.jpg", + "image_caption": [ + "Figure 2. Two examples of the collected data with corresponding labeled navigation instructions and optional notice instructions.", + "Figure 3. Distribution of parsed clips in terms of clip length and the corresponding navigation instruction type." + ], + "image_footnote": [], + "bbox": [ + 500, + 287, + 890, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "input modality and fused vectorized object-level 2D scene representation to enable the LLM to answer the questions based on the current environment.", + "bbox": [ + 498, + 414, + 890, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, this line of work only considered the driving problem in the open-loop settings, and ignored questions such as cumulative error, temporal action consistency, and end-to-end trainability, which are critical for bringing the models into actual closed-loop driving tasks. To the best of our knowledge, we are the first language-based end-to-end autonomous driving method in the closed-loop setting. Relevant datasets, benchmarks, and trained models are also open-sourced to facilitate further research.", + "bbox": [ + 496, + 458, + 892, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset generation", + "text_level": 1, + "bbox": [ + 500, + 607, + 681, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to develop an intelligent driving agent that can generate driving actions based on three sources of input: 1) sensor data (multi-view camera and LiDAR), so that the agent can generate actions that are aware of and compliant with the current scene; 2) navigation instructions (e.g., lane changing, turning), so that the agent can drive to meet the requirement in natural language (instruction from humans or navigation software); and 3) human notice instruction, so that the agent can interact with humans and adapt to human's suggestions and preferences (e.g., pay attention to adversarial events, deal with long-tail events, etc). In this section, we describe how to generate the multi-modal dataset needed to train the agent, and the prompt design for the navigation instruction and human notice instruction. Specifically, we choose the CARLA [12] as the simulator, because it can simulate a realistic dynamic closed-loop world and it is widely adopted in the field of end-to-end autonomous driving. The data collection consists of two stages: 1) col", + "bbox": [ + 496, + 632, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15122", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9e62b417cdf6d6f50d489c4910bde7dc50dcf4ae92ca11dd517fa2bc3e1d2344.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeThree randomly chosen instructions of each instruction type
FollowMaintain your current course until the upcoming intersection.\nIn [x] meters, switch to left lane.\nEase on to the left and get set to join the highway.
TurnAfter [x] meters, take a left.\nAt the next intersection, just keep heading straight, no turn.\nYou'll be turning left at the next T-junction, alright?
OthersFeel free to start driving.\nSlow down now.\nHead to the point, next one's [x] meters ahead, [y] meters left/right.
NoticeWatch for walkers up front.\nJust a heads up, there's a bike ahead.\nPlease be aware of the red traffic signal directly in front of you.
", + "bbox": [ + 80, + 89, + 467, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Examples of considered navigation instructions (follow, turn, others) and notice instructions. $[x]$ and $[y]$ represent the float number for a specific distance.", + "bbox": [ + 76, + 260, + 468, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lecting sensor data and control signals with an expert agent; and 2) parsing and labeling collected data with instructions.", + "bbox": [ + 75, + 316, + 468, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sensor and control data collection. We utilize a rule-based expert agent [38] to create a dataset including about 3M driving frames. Since the expert agent can access the privileged information in the CARLA, this dataset will include camera data, LiDAR data, and control actions for each frame. To enhance the diversity of the collected dataset, the agent runs on 2.5k routes, 8 towns, and 21 kinds of environmental conditions (e.g., weather, time of the day). We use four RGB cameras (left, front, right, rear) and one LiDAR. The side cameras are angled at $60^{\\circ}$ . Besides, we center-crop the front image as an additional focus-view image to capture the status of the distant traffic light. The LiDAR has 64 channels and generates 600K points per second.", + "bbox": [ + 75, + 351, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parsing and language annotation. In the second stage, we parse the collected data into clips, and label each clip with proper navigation instructions and optional notice instructions. The parsing process takes a sequence of frames as input, and segments these frames into clips, where each clip corresponds to one navigation instruction. For instance, if the agent started to turn left at frame $T_{0}$ and ended at frame $T_{n}$ , we will label $(T_{0}, T_{n})$ as a new clip with the instruction \"Hang a left at the next crossroads\". Besides, if an adversarial event occurs at time $T_{a}$ , we will add one notice instruction into this clip, simulating a real-life scenario where a passenger or a side assistance system would communicate with the driver when an emergency happens. As shown in Figure 2, each clip includes sensor data, control signals, the corresponding navigation instruction, and optional notice instructions. The distribution of the parsed clips in terms of clip length and corresponding instruction is shown in Figure 3. In our dataset, we collect 64K parsed clips and 464K notice instructions.", + "bbox": [ + 75, + 549, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction design. We consider three types of navigation", + "bbox": [ + 76, + 835, + 467, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instructions (follow, turn, and others) along with one type of notice instruction, consisting of a total of 56 different instructions. Table 1 shows some examples and the full list can be found in supplementary material. To enable the agent to drive in realistic instructional settings where the instructions come from navigation software or humans, we", + "bbox": [ + 496, + 90, + 890, + 179 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Diversifying the instructions: Considering the inherent richness of natural language, for each type of instruction, we utilize ChatGPT to generate 8 different variants, each carrying the same semantic meaning but varying in phrasing. This enables more comprehensive coverage and flexibility in language interpretation, accommodating the diverse ways the same instruction can be conveyed.", + "- Incorporating misleading instructions: in real-world cases, the navigation software or passengers may give misleading instructions to the AV that violate traffic rules or raise safety concerns. For example, on a single-lane road, following an instruction \"Change to left lane\" is dangerous. To improve the robustness of our model against misleading instructions, we simulate these scenarios and add them to our dataset.", + "- Connecting multiple instructions: In many cases, the instructions may consist of two to three consecutive instructions, such as \"Turn right at this intersection, then go straight to the next intersection and turn right again.\" We also construct some consecutive complex instruction data to simulate real navigation-based driving scenarios." + ], + "bbox": [ + 500, + 180, + 890, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. LMDrive methodology", + "text_level": 1, + "bbox": [ + 500, + 506, + 714, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we propose LMDrive, a framework that can understand and follow high-level driving instructions through natural language. As illustrated in Figure 4, LMDrive is composed of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a LLM and its associated component (tokenizer, Q-Former, and adapters) that takes in the visual tokens and language instruction, to predict the control signal and whether the given instruction is completed. We will introduce the vision encoder in Section 4.1, and the language model with its associated components in Section 4.2. Finally, we describe the training details in Section 4.3.", + "bbox": [ + 496, + 531, + 890, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Vision encoder", + "text_level": 1, + "bbox": [ + 500, + 734, + 650, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the visual language community [5, 29, 58], the most common way to align vision and language could be using pre-trained CLIP models [35] to encode image features. However, the large flops and parameter size of the CLIP models increase the difficulty of its deployment in AV systems. Also, AV perception systems are usually in 3D to include LiDAR input. Hence, inspired by InterFuser [38] and $\\mathrm{TF}++$ [20], we design a multi-view multi-modality vision encoder to encode/fuse the sensor data. As shown in Figure 5, the vision encoder consists of the sensor encoding", + "bbox": [ + 496, + 751, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The adversarial events include bad road conditions, the front vehicle's sudden brake, unexpected entities rushing into the road from occluded regions, vehicles running a red traffic light, etc.", + "bbox": [ + 75, + 862, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15123", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/427c0bfa908c49a0b7fa806fcfbf8e9690e4ea4de375af29aa9e02402f7fd156.jpg", + "image_caption": [ + "Figure 4. The structure of the proposed LMDrive model, which consists of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a large language model and its associated component (tokenizer, Q-Former, and adapters) that processes all the historic visual tokens and the language instructions (navigation instruction and optional notice instruction), to predict the control signal and whether the given instruction is completed." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 867, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "part which encodes image and LiDAR input respectively, and a BEV decoder that fuses image and point cloud features to generate visual tokens which are then passed to the language model. Notably, the vision encoder is pre-trained on perception tasks by adding additional prediction heads, then the encoder is frozen for later use by the LLM.", + "bbox": [ + 75, + 493, + 470, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensor encoding. For each image input, we apply a 2D backbone ResNet [17] to extract the image feature map. The feature map is then flattened into one-dimensional tokens. For a comprehensive understanding of the global context from multiple viewpoints, tokens from different views will be fused by a standard $K_{enc}$ -layer transformer encoder, each layer containing Multi-Headed Self-Attention [48], MLP blocks and layer normalization [1]. For the LiDAR input, we adopt a 3D backbone PointPillars [23] to process the raw point cloud data into ego-centered LiDAR features, where each pillar encompasses points within a $0.25\\mathrm{m} \\times 0.25\\mathrm{m}$ area. PointNet [34] is then used to aggregate features and downsample the feature map to $C \\times H \\times W$ , which subsequently serve as BEV queries.", + "bbox": [ + 75, + 592, + 468, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BEV decoder. The encoded sensor features above are then passed into the BEV decoder to generate visual tokens. Specifically, the BEV decoder is designed as a standard transformer with $K_{dec}$ layers. The BEV point cloud features are fed into the BEV decoder as $H \\times W$ queries to attend to the multi-view image features and generate BEV", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b43f97a0b49cab9a0c0146df9ddb2d5a337539e7f86d541eef1e1ff264bcb4d.jpg", + "image_caption": [ + "Figure 5. The detailed structure of the vision encoder, which takes as input the multi-view multi-modality sensor data. In the pre-training stage, the vision encoder is appended with prediction headers to perform pre-training tasks (object detection, traffic light status classification, and future waypoint prediction). In the instruction-finetuning stage and inference stage, the prediction headers are discarded, and the vision encoder is frozen to generate visual tokens to feed into the LLM." + ], + "image_footnote": [], + "bbox": [ + 506, + 494, + 888, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tokens. We also feed $N$ learnable queries and 1 learnable query into the BEV decoder to generate $N$ waypoint point tokens and 1 traffic light token respectively. Thus the three types of visual tokens (BEV, waypoint, and traffic light) will contain rich scene information and will be then presented to the large language model.", + "bbox": [ + 496, + 744, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training with prediction headers. we consider three vision encoder pre-training tasks: object detection, future waypoint prediction, and traffic light status classification. For object detection, the BEV tokens will pass through a", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15124", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "one-stage CenterPoint [53] to predict the bounding boxes and velocity of the objects in an $H\\mathrm{m} \\times W\\mathrm{m}$ area. For waypoint prediction, we pass the $N$ waypoint tokens along with the navigation waypoint into the GRU network [9] sequentially to predict $N$ future waypoint. For the traffic light status classification, a 2-layer MLP is applied to the traffic light token. Three corresponding loss terms are considered: 1) the detection loss as in InterFuser [39]; 2) the $l_{1}$ waypoint loss; and 3) the cross-entropy traffic light state loss. Note that these prediction headers are only used in the pretraining of the vision encoder, and will be discarded in the training of the LLMs and inference of the whole model.", + "bbox": [ + 75, + 90, + 472, + 270 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. LLM for instruction-following auto driving", + "text_level": 1, + "bbox": [ + 76, + 284, + 444, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, in our framework, the LLM functions as the \"brain\" throughout the entire driving procedure, processing sensor tokens generated by the frozen vision encoder for each frame, comprehending natural language instructions, generating the necessary control signal, and predicting whether the given instruction is completed. Specifically, we choose LLaMA [45] as the language backbone, which has been widely adopted in many language [16, 55] and vision [29, 58] instruction-tuning models. We also have three associated components to bridge LLM with instruction, visual information input, and action prediction: 1) a tokenizer, 2) a Q-Former, 3) two adapters.", + "bbox": [ + 75, + 309, + 472, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction and visual tokenization. Given the navigation instruction and optional notice instruction, we apply the LLaMA tokenizer [45] to convert the instruction into textual tokens. Note that the duration of executing one instruction would span from a few seconds to a few minutes, and our model is deployed in the closed-loop setting. Thus at each frame, we utilize all historic sensor information (with maximum limit $T_{max}$ ) to depress cumulative error and improve the temporal consistency of the model. Specifically, for each frame's multi-view multi-modality sensor input, we utilize the vision encoder pre-trained in the previous section to generate visual tokens ( $H \\times W$ BEV tokens, $N$ waypoint tokens, and one traffic light token). However, the number of visual tokens (e.g., 406 tokens for each frame) quickly grows too large for the LLM because usually hundreds of frames are needed to complete one instruction. To overcome this, we follow BLIP-2 [24] to use the Q-Former to reduce the number of visual tokens. Specifically, for each frame, we employ $M$ learnable queries to attend to the visual tokens via cross-attention layers, which can reduce each frame's visual token number to $M$ . Subsequently, we use a 2-layer MLP adapter to convert the tokens extracted by the Q-Former to share the same dimension as the language token, which can then be fed into the LLM.", + "bbox": [ + 75, + 493, + 472, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action prediction. After receiving a sequence of instructional and visual tokens, the LLM predicts the action tokens. One another 2-layer MLP adapter is then applied to predict", + "bbox": [ + 75, + 856, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "future waypoints, as well as a flag to indicate whether the given instruction has been completed. Note that to enhance the supervision signal, we will also conduct prediction for every historic frame during training, and only the prediction of the latest frame will be executed at inference time. To get the final control signal, which includes braking, throttling, and steering, following LBC [4], we use two PID controllers for latitudinal and longitudinal control to track the heading and velocity of predicted waypoints respectively.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training objectives. When finetuning the LLM and its associated components, we consider two loss terms: 1) the $l_{1}$ waypoint loss; 2) the classification loss (cross-entropy), indicating if the current frame completes the instruction.", + "bbox": [ + 496, + 228, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training details", + "text_level": 1, + "bbox": [ + 498, + 296, + 658, + 313 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LMDrive's training consists of two stages: 1) the vision encoder pre-training stage; and 2) the instruction-finetuning stage, to align the instruction/vision and control signal.", + "bbox": [ + 496, + 320, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision encoder pre-training stage. The vision encoder takes a single frame's sensor data as input, and we use the dataset collected in Section 3 for training. Specifically, since the instruction annotation process will drop some frames, we use the raw dataset before the instruction annotation for the vision encoder pre-training, which includes data of around 3M frames. Only the vision encoder is pretrained with perception tasks for scene understanding.", + "bbox": [ + 496, + 368, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Instruction-finetuning stage. The entire system is trained for end-to-end autonomous driving under the guidance of the instruction, where the Q-Former and Adapters are trainable and the other components are frozen. While our LMDrive takes a sequence of frames as input, during training we set a fixed sequence length $T_{max}$ for building up batch data. The training utilizes the instruction-following data generated in Section 3. To enable the model to reject misleading instructions, we label the corresponding data as 'completed' after the misleading instruction is given for about 1 second. Since the dataset is collected at a high frequency ( $\\sim 10\\mathrm{Hz}$ ), the data in adjacent frames are highly similar. To encourage efficient training, following video prediction methods [13, 37, 49], we sample the training frames in a fixed interval, and apply temporal augmentation which randomly shifts the training frames either forward or backward, with the random shift less than the fixed interval.", + "bbox": [ + 496, + 489, + 890, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. LangAuto Benchmark", + "text_level": 1, + "bbox": [ + 498, + 756, + 710, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We propose the LangAuto (Language-guided Autonomous Driving) CARLA benchmark, the first benchmark that evaluates closed-loop driving performance under language instructions. Compared with the previous CARLA benchmark, Town05 [33] and Longest6 [8] that navigate the agent with discrete driving commands or target waypoints, our benchmark only provides the AV with a navigation instruction and optional notice instructions in natural language.", + "bbox": [ + 496, + 781, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15125", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c9bf7ebabfaec255f49030123cbac60231aa9413350041fc907b9fde34b749a3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LLM BackboneLangAutoLangAuto-ShortLangAuto-Tiny
DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑
Random Init.10.7±3.816.2±4.90.63±0.0414.2±4.420.1±4.40.72±0.0420.1±4.124.7±5.10.75±0.03
LLaMA [45]31.3±1.537.1±1.60.82±0.0142.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [46]32.8±2.140.1±2.20.81±0.0244.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [55]33.5±1.939.3±1.90.83±0.0245.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [55]34.0±3.839.0±3.30.85±0.0647.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [27]36.2±2.346.5±4.30.81±0.0350.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
", + "bbox": [ + 119, + 88, + 848, + 203 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/11e1427418c7539fb327bd400bd941f5035b425538f21a345c98069399db855b.jpg", + "table_caption": [ + "Table 2. Performance comparison of 6 LLM backbones on the LangAuto benchmark. We report the metrics for 3 evaluation runs." + ], + "table_footnote": [], + "table_body": "
Module designDS ↑RC ↑IS ↑
Baseline (LLaVA-v1.5)36.2±2.346.5±4.30.81±0.03
w/o Q-Former31.7±3.541.2±4.40.79±0.02
w/o using BEV tokens33.9±3.945.9±5.10.72±0.03
w/o visual pre-training16.9±5.124.1±4.70.70±0.04
", + "bbox": [ + 99, + 231, + 446, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on the module design.", + "bbox": [ + 133, + 306, + 410, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, the LangAuto benchmark covers all 8 publicly available towns in CARLA to include various scenarios (e.g., highways, intersections, roundabouts). We also consider 16 kinds of environmental conditions, encompassing the combinations of 7 weather conditions (Clear, Cloudy, Wet, MidRain, WetCloudy, HardRain, SoftRain) and 3 daylight conditions (Night, Noon, Sunset). Besides, LangAuto consists of three tracks to fully test the agent's instruction-following abilities:", + "bbox": [ + 75, + 339, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LangAuto track: For each route, navigation instructions are given and updated to the agent based on the agent's current position. We also divide this track into three sub-tracks with different route lengths, to better distinguish the performance. LangAuto where the routes are longer than 500 meters, LangAuto-Short where the route length is between 150 and 500 meters, and LangAuto-Tiny where the route length is shorter than 150 meters.", + "- LangAuto-Notice track: Based on the LangAuto track, we additional add notice instructions to the agent. This setting simulates real cases where passengers or other assistance systems can give real-time notice in long-trail complex or adversarial scenarios, which is usually hard for the AV system to handle by itself. Ideally, the agent that can comprehend and leverage instruction can achieve better performance.", + "- LangAuto-Sequential track: Based on the LangAuto track, we merge $10\\%$ of consecutive 2 to 3 instructions into a single long instruction. This setting mimics the realistic scenarios where the multi-sentence instructions come from the passengers or navigation software." + ], + "bbox": [ + 76, + 476, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that misleading instructions will be randomly $(\\sim 5\\%)$ and intermittently given to the driving agent, which lasts for a certain duration (1-2 seconds). The driving agent is expected to reject these misleading instructions and execute safe actions that are compliant to the current scene, until the next correct instruction is spawned.", + "bbox": [ + 75, + 790, + 467, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We consider three major metrics introduced by the", + "bbox": [ + 76, + 885, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CARLA LeaderBoard [42]: route completion (RC), infraction score (IS), and driving score (DS). The route completion refers to the percentage of the total route length that has been completed. It only takes into account the distance traveled along the predetermined route, where each segment of the predetermined route corresponds to a navigation instruction. If the agent deviates too far from the route, the agent is regarded as violating the instruction, and this episode is marked as a failure and terminated. The infraction score measures infractions triggered by the agent. When collisions or traffic rule violations occur, the infraction score is decayed by a corresponding discount factor. The driving score is the product of the route completion ratio and the infraction score, describing both driving progress and safety. It is generally recognized as the primary ranking metric.", + "bbox": [ + 496, + 233, + 890, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 473, + 632, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 498, + 674, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We implement and evaluate our approach on the open-source CARLA simulator of version 0.9.10.1 [12]. For the 2D backbone and 3D backbone of the vision encoder, ResNet-50 [17] backbone is pre-trained on ImageNet, and PointPillars [23] is trained from scratch. $C, H, W, N$ are set as 256, 50, 50, 5 respectively. For the Q-former token number $M$ , we empirically found that $M = 4$ tokens achieve decent performance. During the instruction-finetuning stage, we sample the training frames in a fixed interval of 2. Because each parsed clip typically contains multiple notice instructions, to avoid overfitting to notice following and generating over-conservative behaviors, during training we randomly removed notices for $75\\%$ of the clips. For the remaining clips, we ensure that a maximum of one notice is included. We refer readers to the appendix for more details.", + "bbox": [ + 496, + 522, + 890, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 757, + 692, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLM backbones. In Table 2, we evaluate our method with different pretrained or randomly initialized LLM models of 7B parameters: LLaMA [45] and LLaMA2 [46] were pretrained with large public language datasets; Vicuna [55] and Vicuna-v1.5 [55] are additionally finetuned with human conversation data; LLaVA-v1.5 [27] incorporate multi-modal data (e.g., language, images) for training. We observe that LLaVA-v1.5 [27] surpasses the other", + "bbox": [ + 496, + 781, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15126", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e571fa3be200b2da5519919a65b2b3e4b48cd82a77f64f4e8c29f1deab786fdd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LLM BackboneBenchmark TypeInfraction Score ↑Vehicle Collisions ↓Pedestrian Collisions ↓Layout Collisions ↓Red light Violations ↓Offroad Infractions ↓Blocked Infractions ↓
LLaVA-v1.5LangAuto0.810.330.030.500.920.360.22
LangAuto-Notice0.870.170.020.310.500.170.27
Vicuna-v1.5LangAuto0.850.300.030.431.180.240.19
LangAuto-Notice0.910.150.010.280.560.260.19
", + "bbox": [ + 119, + 89, + 849, + 181 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2488bf90d98b1ef4013bc01f8a13e011e165ad46b6765f28a64693fdef6d3fc9.jpg", + "table_caption": [ + "Table 4. Performance comparison on LangAuto and LangAuto-Notice benchmarks. The metrics (except infraction score) are normalized by the driven distance (km)." + ], + "table_footnote": [], + "table_body": "
LLM BackboneBenchmark TypeDS ↑RC ↑IS ↑
LLaVA-v1.5LangAuto36.246.50.81
LangAuto-Sequential34.043.70.81
Vicuna-v1.5LangAuto34.039.00.85
LangAuto-Sequential31.937.10.84
", + "bbox": [ + 98, + 229, + 447, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance comparison on LangAuto and LangAutoSequential benchmarks.", + "bbox": [ + 76, + 306, + 468, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLM models, which demonstrates the importance of adopting pretrained multi-modal LLMs for instruction-following autonomous driving. Besides, the models finetuned with instruction data perform better than others (Vicuna-v1.5 $>$ LLaMA2 $\\approx$ Vicuna $>$ LLaMA). We also test a randomly initialized 7B LLM model, which struggles to drive properly with the same amount of training data, which demonstrates the necessity of finetuning pretrained LLM for instruction-following driving.", + "bbox": [ + 75, + 353, + 468, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Module Design. In Table 3, we conduct ablation studies on different components. First, instead of reducing the number of BEV tokens with the Q-Former, we directly down-sample the BEV features to $4 \\times 4$ , then feed them into the LLM (denoted as \"w/o Q-Former\"). The average driving score dropped from 36.2 to 31.7. Second, excluding the BEV tokens input into LLM decoder (denoted as \"w/o using BEV tokens\" results in a decreased infraction score (0.81 to 0.72). One explanation is that the BEV tokens are important for detecting and reasoning the surrounding obstacles and road structures. Third, we remove the pre-training stage of the vision encoder and directly train it from scratch in the instruction-finetuning stage. The driving score drops to 16.9 (denoted as \"w/o visual pre-training\"), demonstrating the importance of our proposed vision encoder pre-training.", + "bbox": [ + 75, + 489, + 468, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LangAuto-Notice Benchmark. The LangAuto-Notice benchmark provides some notice instructions to the agent when adversarial events happen. As shown in Table 4, our agents can effectively leverage the information of the notice in real-time, resulting in a significant decrease in both collisions and traffic rule violations.", + "bbox": [ + 75, + 717, + 468, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LangAuto-Sequential Benchmark. In Table 5, we demonstrate the effect of the LangAuto-Sequential benchmark where some consecutive 2 to 3 navigation instructions are merged together to a long and complex instruction. This setting additionally requires the agent to be able to be temporally aware of which instruction has been completed and", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/878d4cde10f08e75e7f00ff2dcbadcc15bea88def0caed84148cd2a3e9ae9c63.jpg", + "image_caption": [ + "Figure 6. An example of how our LMDrive predicts future waypoints, given sensor inputs and varied navigational instructions. Under the first two instructions, our model predicts different waypoints accordingly. The third instruction is a misleading one (turn right on a left-only lane). The model appropriately rejects the incorrect instruction, generating a slower speed and a safe path that is compliant with the scenario." + ], + "image_footnote": [], + "bbox": [ + 503, + 232, + 890, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "which has not. Our agents based on LLaVA and Vicuna both have a performance drop.", + "bbox": [ + 498, + 500, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 545, + 619, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced LMDrive, a language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive incorporates natural language instructions along with sensor data, enabling human-like interaction and navigation in complex driving scenarios. We also propose the language-guided driving dataset, comprising around 64K multi-modal data clips along with corresponding navigation instructions. We established the LangAuto benchmark for evaluating autonomous driving systems considering natural language instructions. The effectiveness of LMDrive was demonstrated through extensive closed-loop experiments, underlining the potential of improving the interaction of autonomous vehicles with humans and the environment. Our work paves the way for future advancements in language-based autonomous driving.", + "bbox": [ + 496, + 571, + 890, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK.", + "bbox": [ + 496, + 796, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15127", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Sergio Casas, Cole Gulino, Simon Suo, Katie Luo, Renjie Liao, and Raquel Urtasun. Implicit latent variable model for scene-consistent motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 624-641. Springer, 2020. 1", + "[3] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 3, 1", + "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 6", + "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 4", + "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 2, 3", + "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2", + "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 6", + "[9] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 6", + "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.3", + "[11] DriveLM Contributors. Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM, 2023.2", + "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3, 7", + "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6", + "[14] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and re" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "alistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3", + "[15] Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, Wenqi Shao, Chao Xu, Conghui He, Junjun He, Hao Shao, Pan Lu, Hongsheng Li, and Yu Qiao. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models, 2024. 3", + "[16] Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. Koala: A dialogue model for academic research. Blog post, April, 1, 2023. 6", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7, 1", + "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2", + "[19] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. 2023. 3", + "[20] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. arXiv preprint arXiv:2306.07957, 2023. 4", + "[21] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 3", + "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2", + "[23] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 5, 7", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6, 1", + "[25] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3", + "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15128", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 4, 6", + "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 1", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 1", + "[32] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 3", + "[33] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[34] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 5, 1", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. Languagempc: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. 2, 3", + "[37] Hao Shao, Shengju Qian, and Yu Liu. Temporal interlacing network. AAAI, 2020. 6", + "[38] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 2, 4", + "[39] Hao Shao, Letian Wang, Ruobing Chen, Steven L Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13723-13733, 2023. 1, 2, 6", + "[40] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 3", + "[41] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems, 35:6531-6543, 2022. 1", + "[42] CARLA team. Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020. Accessed: 2021-02-11. 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Trisha Thadani. Cruise recalls all its driverless cars after pedestrian hit and dragged. The Washintong Post. 2", + "[44] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 3", + "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 6, 7", + "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 7", + "[47] Jordan Valinsky. 'complete meltdown': Driverless cars in san francisco stall causing a traffic jam. CNN Business. 2", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 5", + "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6", + "[50] Letian Wang, Jie Liu, Hao Shao, Wenshuo Wang, Ruobing Chen, Yu Liu, and Steven L Waslander. Efficient reinforcement learning for autonomous driving with parameterized skills and priors. arXiv preprint arXiv:2305.04412, 2023. 3", + "[51] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2", + "[52] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 3", + "[53] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 6", + "[54] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 3", + "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15129", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Tong Zhou, Letian Wang, Ruobing Chen, Wenshuo Wang, and Yu Liu. Accelerating reinforcement learning for autonomous driving using task-agnostic and ego-centric motion skills. arXiv preprint arXiv:2209.12072, 2022. 3", + "[57] Yang Zhou, Hao Shao, Letian Wang, Steven L. Waslander, Hongsheng Li, and Yu Liu. Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction, 2024. 1", + "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 6" + ], + "bbox": [ + 78, + 90, + 468, + 257 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15130", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_model.json b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f263d145e6f9b2932c47ce451722f5512fe5539 --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_model.json @@ -0,0 +1,2070 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.859, + 0.152 + ], + "angle": 0, + "content": "LMDrive: Closed-Loop End-to-End Driving with Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.179, + 0.846, + 0.26 + ], + "angle": 0, + "content": "Hao Shao\\(^{1,2}\\) Yuxuan Hu\\(^{5}\\) Letian Wang\\(^{4}\\) Guanglu Song\\(^{2}\\) Steven L. Waslander\\(^{4}\\) Yu Liu\\(^{2,3}\\) Hongsheng Li\\(^{1,3,5}\\) \n\\(^{1}\\)CUHK MMLab \\(^{2}\\)SenseTime Research \n\\(^{3}\\)Shanghai Artificial Intelligence Laboratory \\(^{4}\\)University of Toronto \\(^{5}\\)CPII under InnoHK" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.293, + 0.314, + 0.309 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.714 + ], + "angle": 0, + "content": "Despite significant recent progress in the field of autonomous driving, modern methods still struggle and can incur serious accidents when encountering long-tail unforeseen events and challenging urban scenarios. On the one hand, large language models (LLM) have shown impressive reasoning capabilities that approach \"Artificial General Intelligence\". On the other hand, previous autonomous driving methods tend to rely on limited-format inputs (e.g., sensor data and navigation waypoints), restricting the vehicle's ability to understand language information and interact with humans. To this end, this paper introduces LMDrive, a novel language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive uniquely processes and integrates multi-modal sensor data with natural language instructions, enabling interaction with humans and navigation software in realistic instructional settings. To facilitate research in language-based closed-loop autonomous driving, we also publicly release the corresponding dataset which includes approximately 64K instruction-following data clips, and the LangAuto benchmark that tests the system's ability to handle complex instructions and challenging driving scenarios. Extensive closed-loop experiments are conducted to demonstrate LMDrive's effectiveness. To the best of our knowledge, we're the very first work to leverage LLMs for closed-loop end-to-end autonomous driving. Code is available on our webpage." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Remarkable progress in autonomous driving has been witnessed in recent years with an increasing number of commercial autonomous vehicles (AVs) deployed on public roads. Generally, state-of-the-art autonomous driving systems can be categorized into two primary approaches: 1) a modular approach where the system is decomposed into several sub-modules such as perception, prediction, and planning, and fixed interfaces are designed to integrate them" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.291, + 0.891, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.483, + 0.893, + 0.54 + ], + "angle": 0, + "content": "Figure 1. We present LMDrive, the first language-guided closed-loop end-to-end driving framework. LMDrive takes as input the language instruction and multi-modal multi-view sensor data, and outputs control signals in real time to drive in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.838 + ], + "angle": 0, + "content": "together [2, 26, 41, 57]; and 2) an end-to-end approach that directly converts sensor data to control signals via a neural network [18, 39]. While both of these approaches are widely adopted and constantly making breakthroughs on challenging benchmarks, both of them share a limitation in that they solely rely on fixed-format inputs such as the sensor data, target waypoints, and action commands, which restricts the agent's ability to comprehend multi-modal information and to interact with humans and the environment. On the other hand, large language models (LLMs) have shown an impressive range of capabilities that approach \"Artificial General Intelligence.\" This encompasses language comprehension, knowledge retrieval, and reasoning. Such capabilities could greatly enhance the safety, controllability, and explainability of autonomous agents. In this work, we seek to answer the question for the first time: Can we build cognitive autonomous driving systems on top of LLMs, that can interact with human passengers or navigation software simply by natural language?" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Making autonomous systems understand natural language opens profound opportunities for advanced reasoning in complex scenarios and efficient interaction with humans, addressing many previously non-trivial problems. To name" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.237, + 0.9 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15120" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.226 + ], + "angle": 0, + "content": "a few: 1) in long-tail unforeseen events and challenging urban situations (e.g., complex and dense intersections) where modern AV systems typically struggle [47] or even incur serious accidents [43], the language-aware AVs can easily survive by following navigation instructions from passengers or navigation software. 2) adapting to passengers' sudden notices (e.g., small objects missed by perception systems) using natural language, previously difficult and rule-dependent." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.737 + ], + "angle": 0, + "content": "Toward these appealing properties, many pioneering works have explored the potential of using large language models to enhance the AV system's reasoning abilities, interpretability, and overall performance in open-loop settings. One of the most common strategies [6, 11, 32, 36] is to 1) first use LLMs to transform the scene perception results and navigation commands into textual descriptions; 2) feed these textural descriptions into LLMs to generate textual driving decisions; and then 3) transfer textual driving decisions into executable control commands. While good preliminary results are shown, this type of approach, where different LLMs tackle sub-tasks individually, is hard to be trained in an end-to-end manner, loses the capability to scale with a large amount of data, and is not robust to perception errors and uncertainties. For example, since the LLMs in the latter two stages do not have access to the sensor data, inaccurate or missed detections in the first stage can lead to large accumulative errors in the latter stages. Towards addressing these issues, end-to-end language-based driving methods [52] have been proposed. However, all of these methods undergo training and evaluation in the open-loop setting, where actions are generated and evaluated against the expert actions, but not executed in the actual environments. Notably, when executing navigation instructions such as \"turn right\", the AV agent should not only generate a sequence of actions, but also consider the changes that the actions bring to the environment. The absence of closed-loop evaluation leads to insufficient consideration of critical issues such as cumulative errors, human-robot interaction, and temporal consistency of actions, which makes the resulting methods difficult to scale beyond a short time horizon making them ineffective in actual systems. To the best of our knowledge, there is no existing paper that leverages LLMs for closed-loop end-to-end autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce LMDrive, an instruction-following multi-modal LLM model for end-to-end closed-loop autonomous driving. The proposed model can process camera-LiDAR sensor data, comprehend driving instructions in natural language, and directly generate vehicle control signals. A pre-trained LLM model is adopted and kept frozen to maintain its reasoning capability. To adapt the LLM for autonomous driving, multiple camera-LiDAR data encoders and learnable input/output adapters are integrated. A pre-training strategy specifically designed for the driving task is also introduced for the multi-modal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.359 + ], + "angle": 0, + "content": "vision encoder. To facilitate the training of LMDrive in a closed-loop setting, we develop a language-guided driving dataset based on the CARLA simulator [12], which simulates the dynamic world and realistic challenging scenarios. To better test the driving model's capability in realistic conversation scenarios, where the instructions might come from humans and navigation software, we 1) consider both navigation instruction and notice instructions; 2) diversify instructions into various phrases via prompt engineering on LLMs; 3) incorporate misleading and unreasonable instructions which are infeasible due to safety concerns or traffic regulations; 4) extend the navigation instructions to include multiple consecutive segments, such as \"Take a left here, then another left at the next one\". Besides, we provide the corresponding LangAuto evaluation benchmark and the pre-trained LMDrive model for reproducibility, and we hope these can facilitate further research in end-to-end closed-loop language-based autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.389 + ], + "angle": 0, + "content": "To summarize, this paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.391, + 0.892, + 0.45 + ], + "angle": 0, + "content": "- We propose a novel end-to-end, closed-loop, language-based autonomous driving framework, LMDrive, which interacts with the dynamic environment via multi-modal multi-view sensor data and natural language instructions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.451, + 0.892, + 0.523 + ], + "angle": 0, + "content": "- We provide a dataset with about 64K data clips, where each clip includes one navigation instruction, several notice instructions, a sequence of multi-modal multi-view sensor data, and control signals. The duration of the clip spans from 2 to 20 seconds." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.525, + 0.892, + 0.583 + ], + "angle": 0, + "content": "- We present the benchmark LangAuto for evaluating the autonomous agents that take language instructions as navigation inputs, which include misleading/long instructions and challenging adversarial driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.584, + 0.892, + 0.644 + ], + "angle": 0, + "content": "- We conduct extensive closed-loop experiments to demonstrate the effectiveness of the proposed framework, and analyze different components of LMDrive to shed light on continuing research along this direction." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.391, + 0.892, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.651, + 0.674 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.684, + 0.794, + 0.7 + ], + "angle": 0, + "content": "2.1. End-to-End Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.707, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Much progress [7, 8] has been achieved recently in the field of end-to-end autonomous driving. UniAD [18] devised a framework that incorporates full-stack driving tasks and utilizes query-unified interfaces to communicate between different tasks. ThinkTwice [22] designed a Look Module to retrieve information from critical regions and utilize the features to refine the coarse prediction. ReasonNet [39] exploited temporal and global information of the driving scene to improve perception performance and benefit occlusion detection. InterFuser [38] proposed a transformer-based framework to fully fuse and process information from multi-modal multi-view sensors for comprehensive scene understanding. TCP [51] proposed an ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15121" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.405 + ], + "angle": 0, + "content": "proach that integrates the two branches for trajectory planning and direct control by involving a novel multi-step prediction. LAV [3] introduced supervisory tasks for learning viewpoint-invariant representation, offering richer supervision at training and aiding complex reasoning during inference. Besides imitation training, various approaches incorporate reinforcement learning strategies. Latent DRL [44] trained in a supervised way to get a latent representation of the observation and conducts reinforcement learning using the representation as the input. Roach [54] employed a reinforcement learning agent with privileged access to environmental information and distilled a model as the final agent. ASAPRL [50] and TaEcRL [56] exploited abstracted skills to effectively improve reinforcement learning efficiency and the final performance by facilitating effective exploration and reward signaling. ScenerioNet [25] and TrafficGen [14] enhanced driving performance via generating and testing on diverse driving scenarios. However, these end-to-end methods lack the ability to verbally or textually interact with humans (passengers), and usually have low explainability in the decision-making process." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.416, + 0.293, + 0.431 + ], + "angle": 0, + "content": "2.2. LLMs in Driving Tasks" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.44, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Emerging advancements in large language models (LLMs) [10, 19, 21, 45, 46] have been witnessed over the last few months. Furthermore, vision large language models (VLLM) [15, 28, 40] further introduce vision encoders and open the doors for LLMs to interpret not only textual data but also images and data in other modalities [5, 29, 58]. In the field of autonomous driving (AD), recent research has integrated LLMs into the AD system for better explainability and natural interaction with humans. Some studies adopt the visual language model approach, which can handle multi-modal input data and provide the textual description as well as the control signal for the driving scenarios. For example, DRIVEGPT4 [52] proposed a multimodal LLM framework, which takes a sequence of frames as input, then generates responses to human inquiries and predicts control signals for the next step. However, since the framework lacks an input command, the predicted control can not follow the specific navigation command, which denotes that the framework is hard to deploy in real scenarios. Meanwhile, more researchers focus on transforming the driving situations into textual descriptions as the input for the LLM, for directly interpreting and reasoning about comprehensive driving situations. In this thread of works, GPT-Driver [32] reformulated motion planning as the task of natural language modeling by converting heterogeneous scene input to language tokens. LanguageMPC [36] leveraged a LLM to reason the complex scenarios and output high-level driving decisions. Then the method tunes a parameter matrix to convert the decision into the low-level control signals. LLM-Driver [6] utilized the numeric vector as the" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.887, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.149, + 0.865, + 0.159 + ], + "angle": 0, + "content": "Navigation instruction: Just move to the left and get ready to leave the highway." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.162, + 0.886, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.22, + 0.796, + 0.23 + ], + "angle": 0, + "content": "Navigation instruction: Turn left at the next T-junction" + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.23, + 0.846, + 0.239 + ], + "angle": 0, + "content": "Notice instruction: Please watch out for the pedestrians up ahead." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.247, + 0.892, + 0.274 + ], + "angle": 0, + "content": "Figure 2. Two examples of the collected data with corresponding labeled navigation instructions and optional notice instructions." + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.289, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.372, + 0.892, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Distribution of parsed clips in terms of clip length and the corresponding navigation instruction type." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.415, + 0.892, + 0.458 + ], + "angle": 0, + "content": "input modality and fused vectorized object-level 2D scene representation to enable the LLM to answer the questions based on the current environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.593 + ], + "angle": 0, + "content": "However, this line of work only considered the driving problem in the open-loop settings, and ignored questions such as cumulative error, temporal action consistency, and end-to-end trainability, which are critical for bringing the models into actual closed-loop driving tasks. To the best of our knowledge, we are the first language-based end-to-end autonomous driving method in the closed-loop setting. Relevant datasets, benchmarks, and trained models are also open-sourced to facilitate further research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.683, + 0.625 + ], + "angle": 0, + "content": "3. Dataset generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.895, + 0.902 + ], + "angle": 0, + "content": "We aim to develop an intelligent driving agent that can generate driving actions based on three sources of input: 1) sensor data (multi-view camera and LiDAR), so that the agent can generate actions that are aware of and compliant with the current scene; 2) navigation instructions (e.g., lane changing, turning), so that the agent can drive to meet the requirement in natural language (instruction from humans or navigation software); and 3) human notice instruction, so that the agent can interact with humans and adapt to human's suggestions and preferences (e.g., pay attention to adversarial events, deal with long-tail events, etc). In this section, we describe how to generate the multi-modal dataset needed to train the agent, and the prompt design for the navigation instruction and human notice instruction. Specifically, we choose the CARLA [12] as the simulator, because it can simulate a realistic dynamic closed-loop world and it is widely adopted in the field of end-to-end autonomous driving. The data collection consists of two stages: 1) col" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15122" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.09, + 0.468, + 0.256 + ], + "angle": 0, + "content": "
TypeThree randomly chosen instructions of each instruction type
FollowMaintain your current course until the upcoming intersection.\nIn [x] meters, switch to left lane.\nEase on to the left and get set to join the highway.
TurnAfter [x] meters, take a left.\nAt the next intersection, just keep heading straight, no turn.\nYou'll be turning left at the next T-junction, alright?
OthersFeel free to start driving.\nSlow down now.\nHead to the point, next one's [x] meters ahead, [y] meters left/right.
NoticeWatch for walkers up front.\nJust a heads up, there's a bike ahead.\nPlease be aware of the red traffic signal directly in front of you.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.261, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Table 1. Examples of considered navigation instructions (follow, turn, others) and notice instructions. \\([x]\\) and \\([y]\\) represent the float number for a specific distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.469, + 0.348 + ], + "angle": 0, + "content": "lecting sensor data and control signals with an expert agent; and 2) parsing and labeling collected data with instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.469, + 0.546 + ], + "angle": 0, + "content": "Sensor and control data collection. We utilize a rule-based expert agent [38] to create a dataset including about 3M driving frames. Since the expert agent can access the privileged information in the CARLA, this dataset will include camera data, LiDAR data, and control actions for each frame. To enhance the diversity of the collected dataset, the agent runs on 2.5k routes, 8 towns, and 21 kinds of environmental conditions (e.g., weather, time of the day). We use four RGB cameras (left, front, right, rear) and one LiDAR. The side cameras are angled at \\(60^{\\circ}\\). Besides, we center-crop the front image as an additional focus-view image to capture the status of the distant traffic light. The LiDAR has 64 channels and generates 600K points per second." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Parsing and language annotation. In the second stage, we parse the collected data into clips, and label each clip with proper navigation instructions and optional notice instructions. The parsing process takes a sequence of frames as input, and segments these frames into clips, where each clip corresponds to one navigation instruction. For instance, if the agent started to turn left at frame \\( T_{0} \\) and ended at frame \\( T_{n} \\), we will label \\( (T_{0}, T_{n}) \\) as a new clip with the instruction \"Hang a left at the next crossroads\". Besides, if an adversarial event occurs at time \\( T_{a} \\), we will add one notice instruction into this clip, simulating a real-life scenario where a passenger or a side assistance system would communicate with the driver when an emergency happens. As shown in Figure 2, each clip includes sensor data, control signals, the corresponding navigation instruction, and optional notice instructions. The distribution of the parsed clips in terms of clip length and corresponding instruction is shown in Figure 3. In our dataset, we collect 64K parsed clips and 464K notice instructions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.837, + 0.468, + 0.852 + ], + "angle": 0, + "content": "Instruction design. We consider three types of navigation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1The adversarial events include bad road conditions, the front vehicle's sudden brake, unexpected entities rushing into the road from occluded regions, vehicles running a red traffic light, etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.18 + ], + "angle": 0, + "content": "instructions (follow, turn, and others) along with one type of notice instruction, consisting of a total of 56 different instructions. Table 1 shows some examples and the full list can be found in supplementary material. To enable the agent to drive in realistic instructional settings where the instructions come from navigation software or humans, we" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.181, + 0.892, + 0.285 + ], + "angle": 0, + "content": "- Diversifying the instructions: Considering the inherent richness of natural language, for each type of instruction, we utilize ChatGPT to generate 8 different variants, each carrying the same semantic meaning but varying in phrasing. This enables more comprehensive coverage and flexibility in language interpretation, accommodating the diverse ways the same instruction can be conveyed." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.285, + 0.892, + 0.403 + ], + "angle": 0, + "content": "- Incorporating misleading instructions: in real-world cases, the navigation software or passengers may give misleading instructions to the AV that violate traffic rules or raise safety concerns. For example, on a single-lane road, following an instruction \"Change to left lane\" is dangerous. To improve the robustness of our model against misleading instructions, we simulate these scenarios and add them to our dataset." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.404, + 0.892, + 0.493 + ], + "angle": 0, + "content": "- Connecting multiple instructions: In many cases, the instructions may consist of two to three consecutive instructions, such as \"Turn right at this intersection, then go straight to the next intersection and turn right again.\" We also construct some consecutive complex instruction data to simulate real navigation-based driving scenarios." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.181, + 0.892, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.715, + 0.525 + ], + "angle": 0, + "content": "4. LMDrive methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.726 + ], + "angle": 0, + "content": "In this work, we propose LMDrive, a framework that can understand and follow high-level driving instructions through natural language. As illustrated in Figure 4, LMDrive is composed of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a LLM and its associated component (tokenizer, Q-Former, and adapters) that takes in the visual tokens and language instruction, to predict the control signal and whether the given instruction is completed. We will introduce the vision encoder in Section 4.1, and the language model with its associated components in Section 4.2. Finally, we describe the training details in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.651, + 0.75 + ], + "angle": 0, + "content": "4.1. Vision encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the visual language community [5, 29, 58], the most common way to align vision and language could be using pre-trained CLIP models [35] to encode image features. However, the large flops and parameter size of the CLIP models increase the difficulty of its deployment in AV systems. Also, AV perception systems are usually in 3D to include LiDAR input. Hence, inspired by InterFuser [38] and \\(\\mathrm{TF}++\\) [20], we design a multi-view multi-modality vision encoder to encode/fuse the sensor data. As shown in Figure 5, the vision encoder consists of the sensor encoding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15123" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.091, + 0.868, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.425, + 0.896, + 0.482 + ], + "angle": 0, + "content": "Figure 4. The structure of the proposed LMDrive model, which consists of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a large language model and its associated component (tokenizer, Q-Former, and adapters) that processes all the historic visual tokens and the language instructions (navigation instruction and optional notice instruction), to predict the control signal and whether the given instruction is completed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.471, + 0.584 + ], + "angle": 0, + "content": "part which encodes image and LiDAR input respectively, and a BEV decoder that fuses image and point cloud features to generate visual tokens which are then passed to the language model. Notably, the vision encoder is pre-trained on perception tasks by adding additional prediction heads, then the encoder is frozen for later use by the LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.47, + 0.803 + ], + "angle": 0, + "content": "Sensor encoding. For each image input, we apply a 2D backbone ResNet [17] to extract the image feature map. The feature map is then flattened into one-dimensional tokens. For a comprehensive understanding of the global context from multiple viewpoints, tokens from different views will be fused by a standard \\( K_{enc} \\)-layer transformer encoder, each layer containing Multi-Headed Self-Attention [48], MLP blocks and layer normalization [1]. For the LiDAR input, we adopt a 3D backbone PointPillars [23] to process the raw point cloud data into ego-centered LiDAR features, where each pillar encompasses points within a \\( 0.25\\mathrm{m} \\times 0.25\\mathrm{m} \\) area. PointNet [34] is then used to aggregate features and downsample the feature map to \\( C \\times H \\times W \\), which subsequently serve as BEV queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "BEV decoder. The encoded sensor features above are then passed into the BEV decoder to generate visual tokens. Specifically, the BEV decoder is designed as a standard transformer with \\( K_{dec} \\) layers. The BEV point cloud features are fed into the BEV decoder as \\( H \\times W \\) queries to attend to the multi-view image features and generate BEV" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.495, + 0.889, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Figure 5. The detailed structure of the vision encoder, which takes as input the multi-view multi-modality sensor data. In the pre-training stage, the vision encoder is appended with prediction headers to perform pre-training tasks (object detection, traffic light status classification, and future waypoint prediction). In the instruction-finetuning stage and inference stage, the prediction headers are discarded, and the vision encoder is frozen to generate visual tokens to feed into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.837 + ], + "angle": 0, + "content": "tokens. We also feed \\( N \\) learnable queries and 1 learnable query into the BEV decoder to generate \\( N \\) waypoint point tokens and 1 traffic light token respectively. Thus the three types of visual tokens (BEV, waypoint, and traffic light) will contain rich scene information and will be then presented to the large language model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.842, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Pre-training with prediction headers. we consider three vision encoder pre-training tasks: object detection, future waypoint prediction, and traffic light status classification. For object detection, the BEV tokens will pass through a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.271 + ], + "angle": 0, + "content": "one-stage CenterPoint [53] to predict the bounding boxes and velocity of the objects in an \\( H\\mathrm{m} \\times W\\mathrm{m} \\) area. For waypoint prediction, we pass the \\( N \\) waypoint tokens along with the navigation waypoint into the GRU network [9] sequentially to predict \\( N \\) future waypoint. For the traffic light status classification, a 2-layer MLP is applied to the traffic light token. Three corresponding loss terms are considered: 1) the detection loss as in InterFuser [39]; 2) the \\( l_{1} \\) waypoint loss; and 3) the cross-entropy traffic light state loss. Note that these prediction headers are only used in the pretraining of the vision encoder, and will be discarded in the training of the LLMs and inference of the whole model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.445, + 0.303 + ], + "angle": 0, + "content": "4.2. LLM for instruction-following auto driving" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.473, + 0.489 + ], + "angle": 0, + "content": "As illustrated in Figure 4, in our framework, the LLM functions as the \"brain\" throughout the entire driving procedure, processing sensor tokens generated by the frozen vision encoder for each frame, comprehending natural language instructions, generating the necessary control signal, and predicting whether the given instruction is completed. Specifically, we choose LLaMA [45] as the language backbone, which has been widely adopted in many language [16, 55] and vision [29, 58] instruction-tuning models. We also have three associated components to bridge LLM with instruction, visual information input, and action prediction: 1) a tokenizer, 2) a Q-Former, 3) two adapters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Instruction and visual tokenization. Given the navigation instruction and optional notice instruction, we apply the LLaMA tokenizer [45] to convert the instruction into textual tokens. Note that the duration of executing one instruction would span from a few seconds to a few minutes, and our model is deployed in the closed-loop setting. Thus at each frame, we utilize all historic sensor information (with maximum limit \\( T_{max} \\)) to depress cumulative error and improve the temporal consistency of the model. Specifically, for each frame's multi-view multi-modality sensor input, we utilize the vision encoder pre-trained in the previous section to generate visual tokens (\\( H \\times W \\) BEV tokens, \\( N \\) waypoint tokens, and one traffic light token). However, the number of visual tokens (e.g., 406 tokens for each frame) quickly grows too large for the LLM because usually hundreds of frames are needed to complete one instruction. To overcome this, we follow BLIP-2 [24] to use the Q-Former to reduce the number of visual tokens. Specifically, for each frame, we employ \\( M \\) learnable queries to attend to the visual tokens via cross-attention layers, which can reduce each frame's visual token number to \\( M \\). Subsequently, we use a 2-layer MLP adapter to convert the tokens extracted by the Q-Former to share the same dimension as the language token, which can then be fed into the LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.857, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Action prediction. After receiving a sequence of instructional and visual tokens, the LLM predicts the action tokens. One another 2-layer MLP adapter is then applied to predict" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "future waypoints, as well as a flag to indicate whether the given instruction has been completed. Note that to enhance the supervision signal, we will also conduct prediction for every historic frame during training, and only the prediction of the latest frame will be executed at inference time. To get the final control signal, which includes braking, throttling, and steering, following LBC [4], we use two PID controllers for latitudinal and longitudinal control to track the heading and velocity of predicted waypoints respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Training objectives. When finetuning the LLM and its associated components, we consider two loss terms: 1) the \\( l_{1} \\) waypoint loss; 2) the classification loss (cross-entropy), indicating if the current frame completes the instruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.297, + 0.659, + 0.314 + ], + "angle": 0, + "content": "4.3. Training details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.366 + ], + "angle": 0, + "content": "LMDrive's training consists of two stages: 1) the vision encoder pre-training stage; and 2) the instruction-finetuning stage, to align the instruction/vision and control signal." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Vision encoder pre-training stage. The vision encoder takes a single frame's sensor data as input, and we use the dataset collected in Section 3 for training. Specifically, since the instruction annotation process will drop some frames, we use the raw dataset before the instruction annotation for the vision encoder pre-training, which includes data of around 3M frames. Only the vision encoder is pretrained with perception tasks for scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Instruction-finetuning stage. The entire system is trained for end-to-end autonomous driving under the guidance of the instruction, where the Q-Former and Adapters are trainable and the other components are frozen. While our LMDrive takes a sequence of frames as input, during training we set a fixed sequence length \\( T_{max} \\) for building up batch data. The training utilizes the instruction-following data generated in Section 3. To enable the model to reject misleading instructions, we label the corresponding data as 'completed' after the misleading instruction is given for about 1 second. Since the dataset is collected at a high frequency (\\( \\sim 10\\mathrm{Hz} \\)), the data in adjacent frames are highly similar. To encourage efficient training, following video prediction methods [13, 37, 49], we sample the training frames in a fixed interval, and apply temporal augmentation which randomly shifts the training frames either forward or backward, with the random shift less than the fixed interval." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.712, + 0.773 + ], + "angle": 0, + "content": "5. LangAuto Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose the LangAuto (Language-guided Autonomous Driving) CARLA benchmark, the first benchmark that evaluates closed-loop driving performance under language instructions. Compared with the previous CARLA benchmark, Town05 [33] and Longest6 [8] that navigate the agent with discrete driving commands or target waypoints, our benchmark only provides the AV with a navigation instruction and optional notice instructions in natural language." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15125" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.849, + 0.204 + ], + "angle": 0, + "content": "
LLM BackboneLangAutoLangAuto-ShortLangAuto-Tiny
DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑
Random Init.10.7±3.816.2±4.90.63±0.0414.2±4.420.1±4.40.72±0.0420.1±4.124.7±5.10.75±0.03
LLaMA [45]31.3±1.537.1±1.60.82±0.0142.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [46]32.8±2.140.1±2.20.81±0.0244.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [55]33.5±1.939.3±1.90.83±0.0245.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [55]34.0±3.839.0±3.30.85±0.0647.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [27]36.2±2.346.5±4.30.81±0.0350.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.207, + 0.867, + 0.221 + ], + "angle": 0, + "content": "Table 2. Performance comparison of 6 LLM backbones on the LangAuto benchmark. We report the metrics for 3 evaluation runs." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.232, + 0.447, + 0.307 + ], + "angle": 0, + "content": "
Module designDS ↑RC ↑IS ↑
Baseline (LLaVA-v1.5)36.2±2.346.5±4.30.81±0.03
w/o Q-Former31.7±3.541.2±4.40.79±0.02
w/o using BEV tokens33.9±3.945.9±5.10.72±0.03
w/o visual pre-training16.9±5.124.1±4.70.70±0.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.308, + 0.411, + 0.322 + ], + "angle": 0, + "content": "Table 3. Ablation study on the module design." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.475 + ], + "angle": 0, + "content": "Specifically, the LangAuto benchmark covers all 8 publicly available towns in CARLA to include various scenarios (e.g., highways, intersections, roundabouts). We also consider 16 kinds of environmental conditions, encompassing the combinations of 7 weather conditions (Clear, Cloudy, Wet, MidRain, WetCloudy, HardRain, SoftRain) and 3 daylight conditions (Night, Noon, Sunset). Besides, LangAuto consists of three tracks to fully test the agent's instruction-following abilities:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.477, + 0.47, + 0.595 + ], + "angle": 0, + "content": "- LangAuto track: For each route, navigation instructions are given and updated to the agent based on the agent's current position. We also divide this track into three sub-tracks with different route lengths, to better distinguish the performance. LangAuto where the routes are longer than 500 meters, LangAuto-Short where the route length is between 150 and 500 meters, and LangAuto-Tiny where the route length is shorter than 150 meters." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.596, + 0.47, + 0.715 + ], + "angle": 0, + "content": "- LangAuto-Notice track: Based on the LangAuto track, we additional add notice instructions to the agent. This setting simulates real cases where passengers or other assistance systems can give real-time notice in long-trail complex or adversarial scenarios, which is usually hard for the AV system to handle by itself. Ideally, the agent that can comprehend and leverage instruction can achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.715, + 0.47, + 0.789 + ], + "angle": 0, + "content": "- LangAuto-Sequential track: Based on the LangAuto track, we merge \\(10\\%\\) of consecutive 2 to 3 instructions into a single long instruction. This setting mimics the realistic scenarios where the multi-sentence instructions come from the passengers or navigation software." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.477, + 0.47, + 0.789 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.468, + 0.88 + ], + "angle": 0, + "content": "Note that misleading instructions will be randomly \\((\\sim 5\\%)\\) and intermittently given to the driving agent, which lasts for a certain duration (1-2 seconds). The driving agent is expected to reject these misleading instructions and execute safe actions that are compliant to the current scene, until the next correct instruction is spawned." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Metrics. We consider three major metrics introduced by the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.459 + ], + "angle": 0, + "content": "CARLA LeaderBoard [42]: route completion (RC), infraction score (IS), and driving score (DS). The route completion refers to the percentage of the total route length that has been completed. It only takes into account the distance traveled along the predetermined route, where each segment of the predetermined route corresponds to a navigation instruction. If the agent deviates too far from the route, the agent is regarded as violating the instruction, and this episode is marked as a failure and terminated. The infraction score measures infractions triggered by the agent. When collisions or traffic rule violations occur, the infraction score is decayed by a corresponding discount factor. The driving score is the product of the route completion ratio and the infraction score, describing both driving progress and safety. It is generally recognized as the primary ranking metric." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.633, + 0.491 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.676, + 0.516 + ], + "angle": 0, + "content": "6.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.747 + ], + "angle": 0, + "content": "We implement and evaluate our approach on the open-source CARLA simulator of version 0.9.10.1 [12]. For the 2D backbone and 3D backbone of the vision encoder, ResNet-50 [17] backbone is pre-trained on ImageNet, and PointPillars [23] is trained from scratch. \\( C, H, W, N \\) are set as 256, 50, 50, 5 respectively. For the Q-former token number \\( M \\), we empirically found that \\( M = 4 \\) tokens achieve decent performance. During the instruction-finetuning stage, we sample the training frames in a fixed interval of 2. Because each parsed clip typically contains multiple notice instructions, to avoid overfitting to notice following and generating over-conservative behaviors, during training we randomly removed notices for \\( 75\\% \\) of the clips. For the remaining clips, we ensure that a maximum of one notice is included. We refer readers to the appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.758, + 0.694, + 0.774 + ], + "angle": 0, + "content": "6.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.892, + 0.901 + ], + "angle": 0, + "content": "LLM backbones. In Table 2, we evaluate our method with different pretrained or randomly initialized LLM models of 7B parameters: LLaMA [45] and LLaMA2 [46] were pretrained with large public language datasets; Vicuna [55] and Vicuna-v1.5 [55] are additionally finetuned with human conversation data; LLaVA-v1.5 [27] incorporate multi-modal data (e.g., language, images) for training. We observe that LLaVA-v1.5 [27] surpasses the other" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15126" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.09, + 0.85, + 0.182 + ], + "angle": 0, + "content": "
LLM BackboneBenchmark TypeInfraction Score ↑Vehicle Collisions ↓Pedestrian Collisions ↓Layout Collisions ↓Red light Violations ↓Offroad Infractions ↓Blocked Infractions ↓
LLaVA-v1.5LangAuto0.810.330.030.500.920.360.22
LangAuto-Notice0.870.170.020.310.500.170.27
Vicuna-v1.5LangAuto0.850.300.030.431.180.240.19
LangAuto-Notice0.910.150.010.280.560.260.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.895, + 0.218 + ], + "angle": 0, + "content": "Table 4. Performance comparison on LangAuto and LangAuto-Notice benchmarks. The metrics (except infraction score) are normalized by the driven distance (km)." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.231, + 0.449, + 0.307 + ], + "angle": 0, + "content": "
LLM BackboneBenchmark TypeDS ↑RC ↑IS ↑
LLaVA-v1.5LangAuto36.246.50.81
LangAuto-Sequential34.043.70.81
Vicuna-v1.5LangAuto34.039.00.85
LangAuto-Sequential31.937.10.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.308, + 0.47, + 0.336 + ], + "angle": 0, + "content": "Table 5. Performance comparison on LangAuto and LangAutoSequential benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.488 + ], + "angle": 0, + "content": "LLM models, which demonstrates the importance of adopting pretrained multi-modal LLMs for instruction-following autonomous driving. Besides, the models finetuned with instruction data perform better than others (Vicuna-v1.5 \\(>\\) LLaMA2 \\(\\approx\\) Vicuna \\(>\\) LLaMA). We also test a randomly initialized 7B LLM model, which struggles to drive properly with the same amount of training data, which demonstrates the necessity of finetuning pretrained LLM for instruction-following driving." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Module Design. In Table 3, we conduct ablation studies on different components. First, instead of reducing the number of BEV tokens with the Q-Former, we directly down-sample the BEV features to \\(4 \\times 4\\), then feed them into the LLM (denoted as \"w/o Q-Former\"). The average driving score dropped from 36.2 to 31.7. Second, excluding the BEV tokens input into LLM decoder (denoted as \"w/o using BEV tokens\" results in a decreased infraction score (0.81 to 0.72). One explanation is that the BEV tokens are important for detecting and reasoning the surrounding obstacles and road structures. Third, we remove the pre-training stage of the vision encoder and directly train it from scratch in the instruction-finetuning stage. The driving score drops to 16.9 (denoted as \"w/o visual pre-training\"), demonstrating the importance of our proposed vision encoder pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.807 + ], + "angle": 0, + "content": "LangAuto-Notice Benchmark. The LangAuto-Notice benchmark provides some notice instructions to the agent when adversarial events happen. As shown in Table 4, our agents can effectively leverage the information of the notice in real-time, resulting in a significant decrease in both collisions and traffic rule violations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "LangAuto-Sequential Benchmark. In Table 5, we demonstrate the effect of the LangAuto-Sequential benchmark where some consecutive 2 to 3 navigation instructions are merged together to a long and complex instruction. This setting additionally requires the agent to be able to be temporally aware of which instruction has been completed and" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.489 + ], + "angle": 0, + "content": "Figure 6. An example of how our LMDrive predicts future waypoints, given sensor inputs and varied navigational instructions. Under the first two instructions, our model predicts different waypoints accordingly. The third instruction is a misleading one (turn right on a left-only lane). The model appropriately rejects the incorrect instruction, generating a slower speed and a safe path that is compliant with the scenario." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.53 + ], + "angle": 0, + "content": "which has not. Our agents based on LLaVA and Vicuna both have a performance drop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.62, + 0.563 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.796 + ], + "angle": 0, + "content": "In this paper, we introduced LMDrive, a language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive incorporates natural language instructions along with sensor data, enabling human-like interaction and navigation in complex driving scenarios. We also propose the language-guided driving dataset, comprising around 64K multi-modal data clips along with corresponding navigation instructions. We established the LangAuto benchmark for evaluating autonomous driving systems considering natural language instructions. The effectiveness of LMDrive was demonstrated through extensive closed-loop experiments, underlining the potential of improving the interaction of autonomous vehicles with humans and the environment. Our work paves the way for future advancements in language-based autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15127" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.239 + ], + "angle": 0, + "content": "[2] Sergio Casas, Cole Gulino, Simon Suo, Katie Luo, Renjie Liao, and Raquel Urtasun. Implicit latent variable model for scene-consistent motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 624-641. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.24, + 0.47, + 0.293 + ], + "angle": 0, + "content": "[3] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.295, + 0.47, + 0.335 + ], + "angle": 0, + "content": "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.336, + 0.47, + 0.39 + ], + "angle": 0, + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.516, + 0.47, + 0.583 + ], + "angle": 0, + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.584, + 0.47, + 0.652 + ], + "angle": 0, + "content": "[9] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.653, + 0.47, + 0.72 + ], + "angle": 0, + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.722, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[11] DriveLM Contributors. Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.763, + 0.47, + 0.817 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and re" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "alistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, Wenqi Shao, Chao Xu, Conghui He, Junjun He, Hao Shao, Pan Lu, Hongsheng Li, and Yu Qiao. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[16] Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. Koala: A dialogue model for academic research. Blog post, April, 1, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[19] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[20] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. arXiv preprint arXiv:2306.07957, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.894, + 0.706 + ], + "angle": 0, + "content": "[23] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15128" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.163 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.468, + 0.204 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.468, + 0.288 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.289, + 0.468, + 0.33 + ], + "angle": 0, + "content": "[32] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.468, + 0.386 + ], + "angle": 0, + "content": "[33] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[34] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.443, + 0.468, + 0.524 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. Languagempc: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.468, + 0.622 + ], + "angle": 0, + "content": "[37] Hao Shao, Shengju Qian, and Yu Liu. Temporal interlacing network. AAAI, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.468, + 0.678 + ], + "angle": 0, + "content": "[38] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.468, + 0.748 + ], + "angle": 0, + "content": "[39] Hao Shao, Letian Wang, Ruobing Chen, Steven L Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13723-13733, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[40] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.468, + 0.859 + ], + "angle": 0, + "content": "[41] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems, 35:6531-6543, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[42] CARLA team. Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020. Accessed: 2021-02-11. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[43] Trisha Thadani. Cruise recalls all its driverless cars after pedestrian hit and dragged. The Washintong Post. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[44] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[47] Jordan Valinsky. 'complete meltdown': Driverless cars in san francisco stall causing a traffic jam. CNN Business. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[50] Letian Wang, Jie Liu, Hao Shao, Wenshuo Wang, Ruobing Chen, Yu Liu, and Steven L Waslander. Efficient reinforcement learning for autonomous driving with parameterized skills and priors. arXiv preprint arXiv:2305.04412, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[51] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[52] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[53] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[54] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15129" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Tong Zhou, Letian Wang, Ruobing Chen, Wenshuo Wang, and Yu Liu. Accelerating reinforcement learning for autonomous driving using task-agnostic and ego-centric motion skills. arXiv preprint arXiv:2209.12072, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[57] Yang Zhou, Hao Shao, Letian Wang, Steven L. Waslander, Hongsheng Li, and Yu Liu. Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15130" + } + ] +] \ No newline at end of file diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_origin.pdf b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..34d68b9f9b27ad9866d725a57f107c5dea837641 --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/03e61345-cacb-4ace-90cd-06b4f58ee7fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e07c991c9c739c4b9e58cb53bb15bd1380cac0ef59cd9f16871e0d2b31e4880 +size 3343752 diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/full.md b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2aadad6c3fae7a91a703b7ea1a9db3cb942bd2dc --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/full.md @@ -0,0 +1,264 @@ +# LMDrive: Closed-Loop End-to-End Driving with Large Language Models + +Hao Shao $^{1,2}$ Yuxuan Hu $^{5}$ Letian Wang $^{4}$ Guanglu Song $^{2}$ Steven L. Waslander $^{4}$ Yu Liu $^{2,3}$ Hongsheng Li $^{1,3,5}$ $^{1}$ CUHK MMLab $^{2}$ SenseTime Research + $^{3}$ Shanghai Artificial Intelligence Laboratory $^{4}$ University of Toronto $^{5}$ CPII under InnoHK + +# Abstract + +Despite significant recent progress in the field of autonomous driving, modern methods still struggle and can incur serious accidents when encountering long-tail unforeseen events and challenging urban scenarios. On the one hand, large language models (LLM) have shown impressive reasoning capabilities that approach "Artificial General Intelligence". On the other hand, previous autonomous driving methods tend to rely on limited-format inputs (e.g., sensor data and navigation waypoints), restricting the vehicle's ability to understand language information and interact with humans. To this end, this paper introduces LMDrive, a novel language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive uniquely processes and integrates multi-modal sensor data with natural language instructions, enabling interaction with humans and navigation software in realistic instructional settings. To facilitate research in language-based closed-loop autonomous driving, we also publicly release the corresponding dataset which includes approximately 64K instruction-following data clips, and the LangAuto benchmark that tests the system's ability to handle complex instructions and challenging driving scenarios. Extensive closed-loop experiments are conducted to demonstrate LMDrive's effectiveness. To the best of our knowledge, we're the very first work to leverage LLMs for closed-loop end-to-end autonomous driving. Code is available on our webpage. + +# 1. Introduction + +Remarkable progress in autonomous driving has been witnessed in recent years with an increasing number of commercial autonomous vehicles (AVs) deployed on public roads. Generally, state-of-the-art autonomous driving systems can be categorized into two primary approaches: 1) a modular approach where the system is decomposed into several sub-modules such as perception, prediction, and planning, and fixed interfaces are designed to integrate them + +![](images/83bbd20f10b86f4719a2d653a69d39e35ee2ceda6b501c09734892af867d0112.jpg) +Figure 1. We present LMDrive, the first language-guided closed-loop end-to-end driving framework. LMDrive takes as input the language instruction and multi-modal multi-view sensor data, and outputs control signals in real time to drive in complex scenarios. + +together [2, 26, 41, 57]; and 2) an end-to-end approach that directly converts sensor data to control signals via a neural network [18, 39]. While both of these approaches are widely adopted and constantly making breakthroughs on challenging benchmarks, both of them share a limitation in that they solely rely on fixed-format inputs such as the sensor data, target waypoints, and action commands, which restricts the agent's ability to comprehend multi-modal information and to interact with humans and the environment. On the other hand, large language models (LLMs) have shown an impressive range of capabilities that approach "Artificial General Intelligence." This encompasses language comprehension, knowledge retrieval, and reasoning. Such capabilities could greatly enhance the safety, controllability, and explainability of autonomous agents. In this work, we seek to answer the question for the first time: Can we build cognitive autonomous driving systems on top of LLMs, that can interact with human passengers or navigation software simply by natural language? + +Making autonomous systems understand natural language opens profound opportunities for advanced reasoning in complex scenarios and efficient interaction with humans, addressing many previously non-trivial problems. To name + +a few: 1) in long-tail unforeseen events and challenging urban situations (e.g., complex and dense intersections) where modern AV systems typically struggle [47] or even incur serious accidents [43], the language-aware AVs can easily survive by following navigation instructions from passengers or navigation software. 2) adapting to passengers' sudden notices (e.g., small objects missed by perception systems) using natural language, previously difficult and rule-dependent. + +Toward these appealing properties, many pioneering works have explored the potential of using large language models to enhance the AV system's reasoning abilities, interpretability, and overall performance in open-loop settings. One of the most common strategies [6, 11, 32, 36] is to 1) first use LLMs to transform the scene perception results and navigation commands into textual descriptions; 2) feed these textural descriptions into LLMs to generate textual driving decisions; and then 3) transfer textual driving decisions into executable control commands. While good preliminary results are shown, this type of approach, where different LLMs tackle sub-tasks individually, is hard to be trained in an end-to-end manner, loses the capability to scale with a large amount of data, and is not robust to perception errors and uncertainties. For example, since the LLMs in the latter two stages do not have access to the sensor data, inaccurate or missed detections in the first stage can lead to large accumulative errors in the latter stages. Towards addressing these issues, end-to-end language-based driving methods [52] have been proposed. However, all of these methods undergo training and evaluation in the open-loop setting, where actions are generated and evaluated against the expert actions, but not executed in the actual environments. Notably, when executing navigation instructions such as "turn right", the AV agent should not only generate a sequence of actions, but also consider the changes that the actions bring to the environment. The absence of closed-loop evaluation leads to insufficient consideration of critical issues such as cumulative errors, human-robot interaction, and temporal consistency of actions, which makes the resulting methods difficult to scale beyond a short time horizon making them ineffective in actual systems. To the best of our knowledge, there is no existing paper that leverages LLMs for closed-loop end-to-end autonomous driving. + +In this work, we introduce LMDrive, an instruction-following multi-modal LLM model for end-to-end closed-loop autonomous driving. The proposed model can process camera-LiDAR sensor data, comprehend driving instructions in natural language, and directly generate vehicle control signals. A pre-trained LLM model is adopted and kept frozen to maintain its reasoning capability. To adapt the LLM for autonomous driving, multiple camera-LiDAR data encoders and learnable input/output adapters are integrated. A pre-training strategy specifically designed for the driving task is also introduced for the multi-modal + +vision encoder. To facilitate the training of LMDrive in a closed-loop setting, we develop a language-guided driving dataset based on the CARLA simulator [12], which simulates the dynamic world and realistic challenging scenarios. To better test the driving model's capability in realistic conversation scenarios, where the instructions might come from humans and navigation software, we 1) consider both navigation instruction and notice instructions; 2) diversify instructions into various phrases via prompt engineering on LLMs; 3) incorporate misleading and unreasonable instructions which are infeasible due to safety concerns or traffic regulations; 4) extend the navigation instructions to include multiple consecutive segments, such as "Take a left here, then another left at the next one". Besides, we provide the corresponding LangAuto evaluation benchmark and the pre-trained LMDrive model for reproducibility, and we hope these can facilitate further research in end-to-end closed-loop language-based autonomous driving. + +To summarize, this paper makes the following contributions: + +- We propose a novel end-to-end, closed-loop, language-based autonomous driving framework, LMDrive, which interacts with the dynamic environment via multi-modal multi-view sensor data and natural language instructions. +- We provide a dataset with about 64K data clips, where each clip includes one navigation instruction, several notice instructions, a sequence of multi-modal multi-view sensor data, and control signals. The duration of the clip spans from 2 to 20 seconds. +- We present the benchmark LangAuto for evaluating the autonomous agents that take language instructions as navigation inputs, which include misleading/long instructions and challenging adversarial driving scenarios. +- We conduct extensive closed-loop experiments to demonstrate the effectiveness of the proposed framework, and analyze different components of LMDrive to shed light on continuing research along this direction. + +# 2. Related Works + +# 2.1. End-to-End Autonomous Driving + +Much progress [7, 8] has been achieved recently in the field of end-to-end autonomous driving. UniAD [18] devised a framework that incorporates full-stack driving tasks and utilizes query-unified interfaces to communicate between different tasks. ThinkTwice [22] designed a Look Module to retrieve information from critical regions and utilize the features to refine the coarse prediction. ReasonNet [39] exploited temporal and global information of the driving scene to improve perception performance and benefit occlusion detection. InterFuser [38] proposed a transformer-based framework to fully fuse and process information from multi-modal multi-view sensors for comprehensive scene understanding. TCP [51] proposed an ap + +proach that integrates the two branches for trajectory planning and direct control by involving a novel multi-step prediction. LAV [3] introduced supervisory tasks for learning viewpoint-invariant representation, offering richer supervision at training and aiding complex reasoning during inference. Besides imitation training, various approaches incorporate reinforcement learning strategies. Latent DRL [44] trained in a supervised way to get a latent representation of the observation and conducts reinforcement learning using the representation as the input. Roach [54] employed a reinforcement learning agent with privileged access to environmental information and distilled a model as the final agent. ASAPRL [50] and TaEcRL [56] exploited abstracted skills to effectively improve reinforcement learning efficiency and the final performance by facilitating effective exploration and reward signaling. ScenerioNet [25] and TrafficGen [14] enhanced driving performance via generating and testing on diverse driving scenarios. However, these end-to-end methods lack the ability to verbally or textually interact with humans (passengers), and usually have low explainability in the decision-making process. + +# 2.2. LLMs in Driving Tasks + +Emerging advancements in large language models (LLMs) [10, 19, 21, 45, 46] have been witnessed over the last few months. Furthermore, vision large language models (VLLM) [15, 28, 40] further introduce vision encoders and open the doors for LLMs to interpret not only textual data but also images and data in other modalities [5, 29, 58]. In the field of autonomous driving (AD), recent research has integrated LLMs into the AD system for better explainability and natural interaction with humans. Some studies adopt the visual language model approach, which can handle multi-modal input data and provide the textual description as well as the control signal for the driving scenarios. For example, DRIVEGPT4 [52] proposed a multimodal LLM framework, which takes a sequence of frames as input, then generates responses to human inquiries and predicts control signals for the next step. However, since the framework lacks an input command, the predicted control can not follow the specific navigation command, which denotes that the framework is hard to deploy in real scenarios. Meanwhile, more researchers focus on transforming the driving situations into textual descriptions as the input for the LLM, for directly interpreting and reasoning about comprehensive driving situations. In this thread of works, GPT-Driver [32] reformulated motion planning as the task of natural language modeling by converting heterogeneous scene input to language tokens. LanguageMPC [36] leveraged a LLM to reason the complex scenarios and output high-level driving decisions. Then the method tunes a parameter matrix to convert the decision into the low-level control signals. LLM-Driver [6] utilized the numeric vector as the + +![](images/99916684b9b3a582b272613337d3e86d4bcd7b50afde67059f2e532c163a0b9c.jpg) +Navigation instruction: Just move to the left and get ready to leave the highway. + +![](images/0aec77596cf85f55a6a5998f1666fcd1e1ada32308eda38bd2d5a9fe0e92bfdb.jpg) +Navigation instruction: Turn left at the next T-junction +Notice instruction: Please watch out for the pedestrians up ahead. + +![](images/1fc85fac44906cdcf24e88d93be1e44c83f39d10fcc42db41610e0573a554bb5.jpg) +Figure 2. Two examples of the collected data with corresponding labeled navigation instructions and optional notice instructions. +Figure 3. Distribution of parsed clips in terms of clip length and the corresponding navigation instruction type. + +input modality and fused vectorized object-level 2D scene representation to enable the LLM to answer the questions based on the current environment. + +However, this line of work only considered the driving problem in the open-loop settings, and ignored questions such as cumulative error, temporal action consistency, and end-to-end trainability, which are critical for bringing the models into actual closed-loop driving tasks. To the best of our knowledge, we are the first language-based end-to-end autonomous driving method in the closed-loop setting. Relevant datasets, benchmarks, and trained models are also open-sourced to facilitate further research. + +# 3. Dataset generation + +We aim to develop an intelligent driving agent that can generate driving actions based on three sources of input: 1) sensor data (multi-view camera and LiDAR), so that the agent can generate actions that are aware of and compliant with the current scene; 2) navigation instructions (e.g., lane changing, turning), so that the agent can drive to meet the requirement in natural language (instruction from humans or navigation software); and 3) human notice instruction, so that the agent can interact with humans and adapt to human's suggestions and preferences (e.g., pay attention to adversarial events, deal with long-tail events, etc). In this section, we describe how to generate the multi-modal dataset needed to train the agent, and the prompt design for the navigation instruction and human notice instruction. Specifically, we choose the CARLA [12] as the simulator, because it can simulate a realistic dynamic closed-loop world and it is widely adopted in the field of end-to-end autonomous driving. The data collection consists of two stages: 1) col + +
TypeThree randomly chosen instructions of each instruction type
FollowMaintain your current course until the upcoming intersection. +In [x] meters, switch to left lane. +Ease on to the left and get set to join the highway.
TurnAfter [x] meters, take a left. +At the next intersection, just keep heading straight, no turn. +You'll be turning left at the next T-junction, alright?
OthersFeel free to start driving. +Slow down now. +Head to the point, next one's [x] meters ahead, [y] meters left/right.
NoticeWatch for walkers up front. +Just a heads up, there's a bike ahead. +Please be aware of the red traffic signal directly in front of you.
+ +Table 1. Examples of considered navigation instructions (follow, turn, others) and notice instructions. $[x]$ and $[y]$ represent the float number for a specific distance. + +lecting sensor data and control signals with an expert agent; and 2) parsing and labeling collected data with instructions. + +Sensor and control data collection. We utilize a rule-based expert agent [38] to create a dataset including about 3M driving frames. Since the expert agent can access the privileged information in the CARLA, this dataset will include camera data, LiDAR data, and control actions for each frame. To enhance the diversity of the collected dataset, the agent runs on 2.5k routes, 8 towns, and 21 kinds of environmental conditions (e.g., weather, time of the day). We use four RGB cameras (left, front, right, rear) and one LiDAR. The side cameras are angled at $60^{\circ}$ . Besides, we center-crop the front image as an additional focus-view image to capture the status of the distant traffic light. The LiDAR has 64 channels and generates 600K points per second. + +Parsing and language annotation. In the second stage, we parse the collected data into clips, and label each clip with proper navigation instructions and optional notice instructions. The parsing process takes a sequence of frames as input, and segments these frames into clips, where each clip corresponds to one navigation instruction. For instance, if the agent started to turn left at frame $T_{0}$ and ended at frame $T_{n}$ , we will label $(T_{0}, T_{n})$ as a new clip with the instruction "Hang a left at the next crossroads". Besides, if an adversarial event occurs at time $T_{a}$ , we will add one notice instruction into this clip, simulating a real-life scenario where a passenger or a side assistance system would communicate with the driver when an emergency happens. As shown in Figure 2, each clip includes sensor data, control signals, the corresponding navigation instruction, and optional notice instructions. The distribution of the parsed clips in terms of clip length and corresponding instruction is shown in Figure 3. In our dataset, we collect 64K parsed clips and 464K notice instructions. + +Instruction design. We consider three types of navigation + +instructions (follow, turn, and others) along with one type of notice instruction, consisting of a total of 56 different instructions. Table 1 shows some examples and the full list can be found in supplementary material. To enable the agent to drive in realistic instructional settings where the instructions come from navigation software or humans, we + +- Diversifying the instructions: Considering the inherent richness of natural language, for each type of instruction, we utilize ChatGPT to generate 8 different variants, each carrying the same semantic meaning but varying in phrasing. This enables more comprehensive coverage and flexibility in language interpretation, accommodating the diverse ways the same instruction can be conveyed. +- Incorporating misleading instructions: in real-world cases, the navigation software or passengers may give misleading instructions to the AV that violate traffic rules or raise safety concerns. For example, on a single-lane road, following an instruction "Change to left lane" is dangerous. To improve the robustness of our model against misleading instructions, we simulate these scenarios and add them to our dataset. +- Connecting multiple instructions: In many cases, the instructions may consist of two to three consecutive instructions, such as "Turn right at this intersection, then go straight to the next intersection and turn right again." We also construct some consecutive complex instruction data to simulate real navigation-based driving scenarios. + +# 4. LMDrive methodology + +In this work, we propose LMDrive, a framework that can understand and follow high-level driving instructions through natural language. As illustrated in Figure 4, LMDrive is composed of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a LLM and its associated component (tokenizer, Q-Former, and adapters) that takes in the visual tokens and language instruction, to predict the control signal and whether the given instruction is completed. We will introduce the vision encoder in Section 4.1, and the language model with its associated components in Section 4.2. Finally, we describe the training details in Section 4.3. + +# 4.1. Vision encoder + +In the visual language community [5, 29, 58], the most common way to align vision and language could be using pre-trained CLIP models [35] to encode image features. However, the large flops and parameter size of the CLIP models increase the difficulty of its deployment in AV systems. Also, AV perception systems are usually in 3D to include LiDAR input. Hence, inspired by InterFuser [38] and $\mathrm{TF}++$ [20], we design a multi-view multi-modality vision encoder to encode/fuse the sensor data. As shown in Figure 5, the vision encoder consists of the sensor encoding + +![](images/427c0bfa908c49a0b7fa806fcfbf8e9690e4ea4de375af29aa9e02402f7fd156.jpg) +Figure 4. The structure of the proposed LMDrive model, which consists of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a large language model and its associated component (tokenizer, Q-Former, and adapters) that processes all the historic visual tokens and the language instructions (navigation instruction and optional notice instruction), to predict the control signal and whether the given instruction is completed. + +part which encodes image and LiDAR input respectively, and a BEV decoder that fuses image and point cloud features to generate visual tokens which are then passed to the language model. Notably, the vision encoder is pre-trained on perception tasks by adding additional prediction heads, then the encoder is frozen for later use by the LLM. + +Sensor encoding. For each image input, we apply a 2D backbone ResNet [17] to extract the image feature map. The feature map is then flattened into one-dimensional tokens. For a comprehensive understanding of the global context from multiple viewpoints, tokens from different views will be fused by a standard $K_{enc}$ -layer transformer encoder, each layer containing Multi-Headed Self-Attention [48], MLP blocks and layer normalization [1]. For the LiDAR input, we adopt a 3D backbone PointPillars [23] to process the raw point cloud data into ego-centered LiDAR features, where each pillar encompasses points within a $0.25\mathrm{m} \times 0.25\mathrm{m}$ area. PointNet [34] is then used to aggregate features and downsample the feature map to $C \times H \times W$ , which subsequently serve as BEV queries. + +BEV decoder. The encoded sensor features above are then passed into the BEV decoder to generate visual tokens. Specifically, the BEV decoder is designed as a standard transformer with $K_{dec}$ layers. The BEV point cloud features are fed into the BEV decoder as $H \times W$ queries to attend to the multi-view image features and generate BEV + +![](images/1b43f97a0b49cab9a0c0146df9ddb2d5a337539e7f86d541eef1e1ff264bcb4d.jpg) +Figure 5. The detailed structure of the vision encoder, which takes as input the multi-view multi-modality sensor data. In the pre-training stage, the vision encoder is appended with prediction headers to perform pre-training tasks (object detection, traffic light status classification, and future waypoint prediction). In the instruction-finetuning stage and inference stage, the prediction headers are discarded, and the vision encoder is frozen to generate visual tokens to feed into the LLM. + +tokens. We also feed $N$ learnable queries and 1 learnable query into the BEV decoder to generate $N$ waypoint point tokens and 1 traffic light token respectively. Thus the three types of visual tokens (BEV, waypoint, and traffic light) will contain rich scene information and will be then presented to the large language model. + +Pre-training with prediction headers. we consider three vision encoder pre-training tasks: object detection, future waypoint prediction, and traffic light status classification. For object detection, the BEV tokens will pass through a + +one-stage CenterPoint [53] to predict the bounding boxes and velocity of the objects in an $H\mathrm{m} \times W\mathrm{m}$ area. For waypoint prediction, we pass the $N$ waypoint tokens along with the navigation waypoint into the GRU network [9] sequentially to predict $N$ future waypoint. For the traffic light status classification, a 2-layer MLP is applied to the traffic light token. Three corresponding loss terms are considered: 1) the detection loss as in InterFuser [39]; 2) the $l_{1}$ waypoint loss; and 3) the cross-entropy traffic light state loss. Note that these prediction headers are only used in the pretraining of the vision encoder, and will be discarded in the training of the LLMs and inference of the whole model. + +# 4.2. LLM for instruction-following auto driving + +As illustrated in Figure 4, in our framework, the LLM functions as the "brain" throughout the entire driving procedure, processing sensor tokens generated by the frozen vision encoder for each frame, comprehending natural language instructions, generating the necessary control signal, and predicting whether the given instruction is completed. Specifically, we choose LLaMA [45] as the language backbone, which has been widely adopted in many language [16, 55] and vision [29, 58] instruction-tuning models. We also have three associated components to bridge LLM with instruction, visual information input, and action prediction: 1) a tokenizer, 2) a Q-Former, 3) two adapters. + +Instruction and visual tokenization. Given the navigation instruction and optional notice instruction, we apply the LLaMA tokenizer [45] to convert the instruction into textual tokens. Note that the duration of executing one instruction would span from a few seconds to a few minutes, and our model is deployed in the closed-loop setting. Thus at each frame, we utilize all historic sensor information (with maximum limit $T_{max}$ ) to depress cumulative error and improve the temporal consistency of the model. Specifically, for each frame's multi-view multi-modality sensor input, we utilize the vision encoder pre-trained in the previous section to generate visual tokens ( $H \times W$ BEV tokens, $N$ waypoint tokens, and one traffic light token). However, the number of visual tokens (e.g., 406 tokens for each frame) quickly grows too large for the LLM because usually hundreds of frames are needed to complete one instruction. To overcome this, we follow BLIP-2 [24] to use the Q-Former to reduce the number of visual tokens. Specifically, for each frame, we employ $M$ learnable queries to attend to the visual tokens via cross-attention layers, which can reduce each frame's visual token number to $M$ . Subsequently, we use a 2-layer MLP adapter to convert the tokens extracted by the Q-Former to share the same dimension as the language token, which can then be fed into the LLM. + +Action prediction. After receiving a sequence of instructional and visual tokens, the LLM predicts the action tokens. One another 2-layer MLP adapter is then applied to predict + +future waypoints, as well as a flag to indicate whether the given instruction has been completed. Note that to enhance the supervision signal, we will also conduct prediction for every historic frame during training, and only the prediction of the latest frame will be executed at inference time. To get the final control signal, which includes braking, throttling, and steering, following LBC [4], we use two PID controllers for latitudinal and longitudinal control to track the heading and velocity of predicted waypoints respectively. + +Training objectives. When finetuning the LLM and its associated components, we consider two loss terms: 1) the $l_{1}$ waypoint loss; 2) the classification loss (cross-entropy), indicating if the current frame completes the instruction. + +# 4.3. Training details + +LMDrive's training consists of two stages: 1) the vision encoder pre-training stage; and 2) the instruction-finetuning stage, to align the instruction/vision and control signal. + +Vision encoder pre-training stage. The vision encoder takes a single frame's sensor data as input, and we use the dataset collected in Section 3 for training. Specifically, since the instruction annotation process will drop some frames, we use the raw dataset before the instruction annotation for the vision encoder pre-training, which includes data of around 3M frames. Only the vision encoder is pretrained with perception tasks for scene understanding. + +Instruction-finetuning stage. The entire system is trained for end-to-end autonomous driving under the guidance of the instruction, where the Q-Former and Adapters are trainable and the other components are frozen. While our LMDrive takes a sequence of frames as input, during training we set a fixed sequence length $T_{max}$ for building up batch data. The training utilizes the instruction-following data generated in Section 3. To enable the model to reject misleading instructions, we label the corresponding data as 'completed' after the misleading instruction is given for about 1 second. Since the dataset is collected at a high frequency ( $\sim 10\mathrm{Hz}$ ), the data in adjacent frames are highly similar. To encourage efficient training, following video prediction methods [13, 37, 49], we sample the training frames in a fixed interval, and apply temporal augmentation which randomly shifts the training frames either forward or backward, with the random shift less than the fixed interval. + +# 5. LangAuto Benchmark + +We propose the LangAuto (Language-guided Autonomous Driving) CARLA benchmark, the first benchmark that evaluates closed-loop driving performance under language instructions. Compared with the previous CARLA benchmark, Town05 [33] and Longest6 [8] that navigate the agent with discrete driving commands or target waypoints, our benchmark only provides the AV with a navigation instruction and optional notice instructions in natural language. + +
LLM BackboneLangAutoLangAuto-ShortLangAuto-Tiny
DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑
Random Init.10.7±3.816.2±4.90.63±0.0414.2±4.420.1±4.40.72±0.0420.1±4.124.7±5.10.75±0.03
LLaMA [45]31.3±1.537.1±1.60.82±0.0142.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [46]32.8±2.140.1±2.20.81±0.0244.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [55]33.5±1.939.3±1.90.83±0.0245.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [55]34.0±3.839.0±3.30.85±0.0647.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [27]36.2±2.346.5±4.30.81±0.0350.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
+ +Table 2. Performance comparison of 6 LLM backbones on the LangAuto benchmark. We report the metrics for 3 evaluation runs. + +
Module designDS ↑RC ↑IS ↑
Baseline (LLaVA-v1.5)36.2±2.346.5±4.30.81±0.03
w/o Q-Former31.7±3.541.2±4.40.79±0.02
w/o using BEV tokens33.9±3.945.9±5.10.72±0.03
w/o visual pre-training16.9±5.124.1±4.70.70±0.04
+ +Table 3. Ablation study on the module design. + +Specifically, the LangAuto benchmark covers all 8 publicly available towns in CARLA to include various scenarios (e.g., highways, intersections, roundabouts). We also consider 16 kinds of environmental conditions, encompassing the combinations of 7 weather conditions (Clear, Cloudy, Wet, MidRain, WetCloudy, HardRain, SoftRain) and 3 daylight conditions (Night, Noon, Sunset). Besides, LangAuto consists of three tracks to fully test the agent's instruction-following abilities: + +- LangAuto track: For each route, navigation instructions are given and updated to the agent based on the agent's current position. We also divide this track into three sub-tracks with different route lengths, to better distinguish the performance. LangAuto where the routes are longer than 500 meters, LangAuto-Short where the route length is between 150 and 500 meters, and LangAuto-Tiny where the route length is shorter than 150 meters. +- LangAuto-Notice track: Based on the LangAuto track, we additional add notice instructions to the agent. This setting simulates real cases where passengers or other assistance systems can give real-time notice in long-trail complex or adversarial scenarios, which is usually hard for the AV system to handle by itself. Ideally, the agent that can comprehend and leverage instruction can achieve better performance. +- LangAuto-Sequential track: Based on the LangAuto track, we merge $10\%$ of consecutive 2 to 3 instructions into a single long instruction. This setting mimics the realistic scenarios where the multi-sentence instructions come from the passengers or navigation software. + +Note that misleading instructions will be randomly $(\sim 5\%)$ and intermittently given to the driving agent, which lasts for a certain duration (1-2 seconds). The driving agent is expected to reject these misleading instructions and execute safe actions that are compliant to the current scene, until the next correct instruction is spawned. + +Metrics. We consider three major metrics introduced by the + +CARLA LeaderBoard [42]: route completion (RC), infraction score (IS), and driving score (DS). The route completion refers to the percentage of the total route length that has been completed. It only takes into account the distance traveled along the predetermined route, where each segment of the predetermined route corresponds to a navigation instruction. If the agent deviates too far from the route, the agent is regarded as violating the instruction, and this episode is marked as a failure and terminated. The infraction score measures infractions triggered by the agent. When collisions or traffic rule violations occur, the infraction score is decayed by a corresponding discount factor. The driving score is the product of the route completion ratio and the infraction score, describing both driving progress and safety. It is generally recognized as the primary ranking metric. + +# 6. Experiments + +# 6.1. Experiment Setup + +We implement and evaluate our approach on the open-source CARLA simulator of version 0.9.10.1 [12]. For the 2D backbone and 3D backbone of the vision encoder, ResNet-50 [17] backbone is pre-trained on ImageNet, and PointPillars [23] is trained from scratch. $C, H, W, N$ are set as 256, 50, 50, 5 respectively. For the Q-former token number $M$ , we empirically found that $M = 4$ tokens achieve decent performance. During the instruction-finetuning stage, we sample the training frames in a fixed interval of 2. Because each parsed clip typically contains multiple notice instructions, to avoid overfitting to notice following and generating over-conservative behaviors, during training we randomly removed notices for $75\%$ of the clips. For the remaining clips, we ensure that a maximum of one notice is included. We refer readers to the appendix for more details. + +# 6.2. Quantitative Results + +LLM backbones. In Table 2, we evaluate our method with different pretrained or randomly initialized LLM models of 7B parameters: LLaMA [45] and LLaMA2 [46] were pretrained with large public language datasets; Vicuna [55] and Vicuna-v1.5 [55] are additionally finetuned with human conversation data; LLaVA-v1.5 [27] incorporate multi-modal data (e.g., language, images) for training. We observe that LLaVA-v1.5 [27] surpasses the other + +
LLM BackboneBenchmark TypeInfraction Score ↑Vehicle Collisions ↓Pedestrian Collisions ↓Layout Collisions ↓Red light Violations ↓Offroad Infractions ↓Blocked Infractions ↓
LLaVA-v1.5LangAuto0.810.330.030.500.920.360.22
LangAuto-Notice0.870.170.020.310.500.170.27
Vicuna-v1.5LangAuto0.850.300.030.431.180.240.19
LangAuto-Notice0.910.150.010.280.560.260.19
+ +Table 4. Performance comparison on LangAuto and LangAuto-Notice benchmarks. The metrics (except infraction score) are normalized by the driven distance (km). + +
LLM BackboneBenchmark TypeDS ↑RC ↑IS ↑
LLaVA-v1.5LangAuto36.246.50.81
LangAuto-Sequential34.043.70.81
Vicuna-v1.5LangAuto34.039.00.85
LangAuto-Sequential31.937.10.84
+ +Table 5. Performance comparison on LangAuto and LangAutoSequential benchmarks. + +LLM models, which demonstrates the importance of adopting pretrained multi-modal LLMs for instruction-following autonomous driving. Besides, the models finetuned with instruction data perform better than others (Vicuna-v1.5 $>$ LLaMA2 $\approx$ Vicuna $>$ LLaMA). We also test a randomly initialized 7B LLM model, which struggles to drive properly with the same amount of training data, which demonstrates the necessity of finetuning pretrained LLM for instruction-following driving. + +Module Design. In Table 3, we conduct ablation studies on different components. First, instead of reducing the number of BEV tokens with the Q-Former, we directly down-sample the BEV features to $4 \times 4$ , then feed them into the LLM (denoted as "w/o Q-Former"). The average driving score dropped from 36.2 to 31.7. Second, excluding the BEV tokens input into LLM decoder (denoted as "w/o using BEV tokens" results in a decreased infraction score (0.81 to 0.72). One explanation is that the BEV tokens are important for detecting and reasoning the surrounding obstacles and road structures. Third, we remove the pre-training stage of the vision encoder and directly train it from scratch in the instruction-finetuning stage. The driving score drops to 16.9 (denoted as "w/o visual pre-training"), demonstrating the importance of our proposed vision encoder pre-training. + +LangAuto-Notice Benchmark. The LangAuto-Notice benchmark provides some notice instructions to the agent when adversarial events happen. As shown in Table 4, our agents can effectively leverage the information of the notice in real-time, resulting in a significant decrease in both collisions and traffic rule violations. + +LangAuto-Sequential Benchmark. In Table 5, we demonstrate the effect of the LangAuto-Sequential benchmark where some consecutive 2 to 3 navigation instructions are merged together to a long and complex instruction. This setting additionally requires the agent to be able to be temporally aware of which instruction has been completed and + +![](images/878d4cde10f08e75e7f00ff2dcbadcc15bea88def0caed84148cd2a3e9ae9c63.jpg) +Figure 6. An example of how our LMDrive predicts future waypoints, given sensor inputs and varied navigational instructions. Under the first two instructions, our model predicts different waypoints accordingly. The third instruction is a misleading one (turn right on a left-only lane). The model appropriately rejects the incorrect instruction, generating a slower speed and a safe path that is compliant with the scenario. + +which has not. Our agents based on LLaVA and Vicuna both have a performance drop. + +# 7. Conclusion + +In this paper, we introduced LMDrive, a language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive incorporates natural language instructions along with sensor data, enabling human-like interaction and navigation in complex driving scenarios. We also propose the language-guided driving dataset, comprising around 64K multi-modal data clips along with corresponding navigation instructions. We established the LangAuto benchmark for evaluating autonomous driving systems considering natural language instructions. The effectiveness of LMDrive was demonstrated through extensive closed-loop experiments, underlining the potential of improving the interaction of autonomous vehicles with humans and the environment. Our work paves the way for future advancements in language-based autonomous driving. + +Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Sergio Casas, Cole Gulino, Simon Suo, Katie Luo, Renjie Liao, and Raquel Urtasun. Implicit latent variable model for scene-consistent motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 624-641. Springer, 2020. 1 +[3] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 3, 1 +[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 6 +[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 4 +[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 2, 3 +[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2 +[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 6 +[9] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 6 +[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.3 +[11] DriveLM Contributors. Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM, 2023.2 +[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3, 7 +[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6 +[14] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and re + +alistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3 +[15] Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, Wenqi Shao, Chao Xu, Conghui He, Junjun He, Hao Shao, Pan Lu, Hongsheng Li, and Yu Qiao. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models, 2024. 3 +[16] Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. Koala: A dialogue model for academic research. Blog post, April, 1, 2023. 6 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7, 1 +[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2 +[19] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. 2023. 3 +[20] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. arXiv preprint arXiv:2306.07957, 2023. 4 +[21] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 3 +[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2 +[23] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 5, 7 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6, 1 +[25] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3 +[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1 + +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 4, 6 +[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 1 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 1 +[32] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 3 +[33] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[34] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 5, 1 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. Languagempc: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. 2, 3 +[37] Hao Shao, Shengju Qian, and Yu Liu. Temporal interlacing network. AAAI, 2020. 6 +[38] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 2, 4 +[39] Hao Shao, Letian Wang, Ruobing Chen, Steven L Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13723-13733, 2023. 1, 2, 6 +[40] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 3 +[41] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems, 35:6531-6543, 2022. 1 +[42] CARLA team. Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020. Accessed: 2021-02-11. 7 + +[43] Trisha Thadani. Cruise recalls all its driverless cars after pedestrian hit and dragged. The Washintong Post. 2 +[44] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 3 +[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 6, 7 +[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 7 +[47] Jordan Valinsky. 'complete meltdown': Driverless cars in san francisco stall causing a traffic jam. CNN Business. 2 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 5 +[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6 +[50] Letian Wang, Jie Liu, Hao Shao, Wenshuo Wang, Ruobing Chen, Yu Liu, and Steven L Waslander. Efficient reinforcement learning for autonomous driving with parameterized skills and priors. arXiv preprint arXiv:2305.04412, 2023. 3 +[51] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2 +[52] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 3 +[53] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 6 +[54] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 3 +[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6, 7 + +[56] Tong Zhou, Letian Wang, Ruobing Chen, Wenshuo Wang, and Yu Liu. Accelerating reinforcement learning for autonomous driving using task-agnostic and ego-centric motion skills. arXiv preprint arXiv:2209.12072, 2022. 3 +[57] Yang Zhou, Hao Shao, Letian Wang, Steven L. Waslander, Hongsheng Li, and Yu Liu. Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction, 2024. 1 +[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 6 \ No newline at end of file diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/images.zip b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aaf5d54b888979e4099574d67ff0edb08f3862b5 --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0821b72bf294ff27526982e5e47a48bda50f4597a937042ec178cdcf6973fe42 +size 435319 diff --git a/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/layout.json b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ac63fd85c354418166b5cfbf13e11ade47500a32 --- /dev/null +++ b/2024/LMDrive_ Closed-Loop End-to-End Driving with Large Language Models/layout.json @@ -0,0 +1,7232 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 525, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 525, + 120 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 525, + 120 + ], + "type": "text", + "content": "LMDrive: Closed-Loop End-to-End Driving with Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "Hao Shao" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Yuxuan Hu" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Letian Wang" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Guanglu Song" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Steven L. Waslander" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Yu Liu" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": " Hongsheng Li" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{1,3,5}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "CUHK MMLab " + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "SenseTime Research \n" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "University of Toronto " + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 70, + 141, + 517, + 205 + ], + "type": "text", + "content": "CPII under InnoHK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 192, + 244 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 290, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 565 + ], + "type": "text", + "content": "Despite significant recent progress in the field of autonomous driving, modern methods still struggle and can incur serious accidents when encountering long-tail unforeseen events and challenging urban scenarios. On the one hand, large language models (LLM) have shown impressive reasoning capabilities that approach \"Artificial General Intelligence\". On the other hand, previous autonomous driving methods tend to rely on limited-format inputs (e.g., sensor data and navigation waypoints), restricting the vehicle's ability to understand language information and interact with humans. To this end, this paper introduces LMDrive, a novel language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive uniquely processes and integrates multi-modal sensor data with natural language instructions, enabling interaction with humans and navigation software in realistic instructional settings. To facilitate research in language-based closed-loop autonomous driving, we also publicly release the corresponding dataset which includes approximately 64K instruction-following data clips, and the LangAuto benchmark that tests the system's ability to handle complex instructions and challenging driving scenarios. Extensive closed-loop experiments are conducted to demonstrate LMDrive's effectiveness. To the best of our knowledge, we're the very first work to leverage LLMs for closed-loop end-to-end autonomous driving. Code is available on our webpage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 598, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 693 + ], + "type": "text", + "content": "Remarkable progress in autonomous driving has been witnessed in recent years with an increasing number of commercial autonomous vehicles (AVs) deployed on public roads. Generally, state-of-the-art autonomous driving systems can be categorized into two primary approaches: 1) a modular approach where the system is decomposed into several sub-modules such as perception, prediction, and planning, and fixed interfaces are designed to integrate them" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 230, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 308, + 230, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 376 + ], + "type": "image", + "image_path": "83bbd20f10b86f4719a2d653a69d39e35ee2ceda6b501c09734892af867d0112.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 382, + 546, + 427 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 546, + 427 + ], + "type": "text", + "content": "Figure 1. We present LMDrive, the first language-guided closed-loop end-to-end driving framework. LMDrive takes as input the language instruction and multi-modal multi-view sensor data, and outputs control signals in real time to drive in complex scenarios." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 547, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 663 + ], + "type": "text", + "content": "together [2, 26, 41, 57]; and 2) an end-to-end approach that directly converts sensor data to control signals via a neural network [18, 39]. While both of these approaches are widely adopted and constantly making breakthroughs on challenging benchmarks, both of them share a limitation in that they solely rely on fixed-format inputs such as the sensor data, target waypoints, and action commands, which restricts the agent's ability to comprehend multi-modal information and to interact with humans and the environment. On the other hand, large language models (LLMs) have shown an impressive range of capabilities that approach \"Artificial General Intelligence.\" This encompasses language comprehension, knowledge retrieval, and reasoning. Such capabilities could greatly enhance the safety, controllability, and explainability of autonomous agents. In this work, we seek to answer the question for the first time: Can we build cognitive autonomous driving systems on top of LLMs, that can interact with human passengers or navigation software simply by natural language?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 715 + ], + "type": "text", + "content": "Making autonomous systems understand natural language opens profound opportunities for advanced reasoning in complex scenarios and efficient interaction with humans, addressing many previously non-trivial problems. To name" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 145, + 712 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15120" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 178 + ], + "type": "text", + "content": "a few: 1) in long-tail unforeseen events and challenging urban situations (e.g., complex and dense intersections) where modern AV systems typically struggle [47] or even incur serious accidents [43], the language-aware AVs can easily survive by following navigation instructions from passengers or navigation software. 2) adapting to passengers' sudden notices (e.g., small objects missed by perception systems) using natural language, previously difficult and rule-dependent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 583 + ], + "type": "text", + "content": "Toward these appealing properties, many pioneering works have explored the potential of using large language models to enhance the AV system's reasoning abilities, interpretability, and overall performance in open-loop settings. One of the most common strategies [6, 11, 32, 36] is to 1) first use LLMs to transform the scene perception results and navigation commands into textual descriptions; 2) feed these textural descriptions into LLMs to generate textual driving decisions; and then 3) transfer textual driving decisions into executable control commands. While good preliminary results are shown, this type of approach, where different LLMs tackle sub-tasks individually, is hard to be trained in an end-to-end manner, loses the capability to scale with a large amount of data, and is not robust to perception errors and uncertainties. For example, since the LLMs in the latter two stages do not have access to the sensor data, inaccurate or missed detections in the first stage can lead to large accumulative errors in the latter stages. Towards addressing these issues, end-to-end language-based driving methods [52] have been proposed. However, all of these methods undergo training and evaluation in the open-loop setting, where actions are generated and evaluated against the expert actions, but not executed in the actual environments. Notably, when executing navigation instructions such as \"turn right\", the AV agent should not only generate a sequence of actions, but also consider the changes that the actions bring to the environment. The absence of closed-loop evaluation leads to insufficient consideration of critical issues such as cumulative errors, human-robot interaction, and temporal consistency of actions, which makes the resulting methods difficult to scale beyond a short time horizon making them ineffective in actual systems. To the best of our knowledge, there is no existing paper that leverages LLMs for closed-loop end-to-end autonomous driving." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 584, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 287, + 713 + ], + "type": "text", + "content": "In this work, we introduce LMDrive, an instruction-following multi-modal LLM model for end-to-end closed-loop autonomous driving. The proposed model can process camera-LiDAR sensor data, comprehend driving instructions in natural language, and directly generate vehicle control signals. A pre-trained LLM model is adopted and kept frozen to maintain its reasoning capability. To adapt the LLM for autonomous driving, multiple camera-LiDAR data encoders and learnable input/output adapters are integrated. A pre-training strategy specifically designed for the driving task is also introduced for the multi-modal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 284 + ], + "type": "text", + "content": "vision encoder. To facilitate the training of LMDrive in a closed-loop setting, we develop a language-guided driving dataset based on the CARLA simulator [12], which simulates the dynamic world and realistic challenging scenarios. To better test the driving model's capability in realistic conversation scenarios, where the instructions might come from humans and navigation software, we 1) consider both navigation instruction and notice instructions; 2) diversify instructions into various phrases via prompt engineering on LLMs; 3) incorporate misleading and unreasonable instructions which are infeasible due to safety concerns or traffic regulations; 4) extend the navigation instructions to include multiple consecutive segments, such as \"Take a left here, then another left at the next one\". Besides, we provide the corresponding LangAuto evaluation benchmark and the pre-trained LMDrive model for reproducibility, and we hope these can facilitate further research in end-to-end closed-loop language-based autonomous driving." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "content": "To summarize, this paper makes the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 309, + 545, + 510 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 309, + 545, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 545, + 356 + ], + "type": "text", + "content": "- We propose a novel end-to-end, closed-loop, language-based autonomous driving framework, LMDrive, which interacts with the dynamic environment via multi-modal multi-view sensor data and natural language instructions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 357, + 545, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 357, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 357, + 545, + 414 + ], + "type": "text", + "content": "- We provide a dataset with about 64K data clips, where each clip includes one navigation instruction, several notice instructions, a sequence of multi-modal multi-view sensor data, and control signals. The duration of the clip spans from 2 to 20 seconds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 415, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 545, + 461 + ], + "type": "text", + "content": "- We present the benchmark LangAuto for evaluating the autonomous agents that take language instructions as navigation inputs, which include misleading/long instructions and challenging adversarial driving scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 545, + 510 + ], + "type": "text", + "content": "- We conduct extensive closed-loop experiments to demonstrate the effectiveness of the proposed framework, and analyze different components of LMDrive to shed light on continuing research along this direction." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 521, + 398, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 398, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 398, + 533 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 541, + 485, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 541, + 485, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 485, + 554 + ], + "type": "text", + "content": "2.1. End-to-End Autonomous Driving" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 559, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 559, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 559, + 545, + 715 + ], + "type": "text", + "content": "Much progress [7, 8] has been achieved recently in the field of end-to-end autonomous driving. UniAD [18] devised a framework that incorporates full-stack driving tasks and utilizes query-unified interfaces to communicate between different tasks. ThinkTwice [22] designed a Look Module to retrieve information from critical regions and utilize the features to refine the coarse prediction. ReasonNet [39] exploited temporal and global information of the driving scene to improve perception performance and benefit occlusion detection. InterFuser [38] proposed a transformer-based framework to fully fuse and process information from multi-modal multi-view sensors for comprehensive scene understanding. TCP [51] proposed an ap" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 320 + ], + "type": "text", + "content": "proach that integrates the two branches for trajectory planning and direct control by involving a novel multi-step prediction. LAV [3] introduced supervisory tasks for learning viewpoint-invariant representation, offering richer supervision at training and aiding complex reasoning during inference. Besides imitation training, various approaches incorporate reinforcement learning strategies. Latent DRL [44] trained in a supervised way to get a latent representation of the observation and conducts reinforcement learning using the representation as the input. Roach [54] employed a reinforcement learning agent with privileged access to environmental information and distilled a model as the final agent. ASAPRL [50] and TaEcRL [56] exploited abstracted skills to effectively improve reinforcement learning efficiency and the final performance by facilitating effective exploration and reward signaling. ScenerioNet [25] and TrafficGen [14] enhanced driving performance via generating and testing on diverse driving scenarios. However, these end-to-end methods lack the ability to verbally or textually interact with humans (passengers), and usually have low explainability in the decision-making process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 329, + 179, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 179, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 179, + 341 + ], + "type": "text", + "content": "2.2. LLMs in Driving Tasks" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 348, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 289, + 714 + ], + "type": "text", + "content": "Emerging advancements in large language models (LLMs) [10, 19, 21, 45, 46] have been witnessed over the last few months. Furthermore, vision large language models (VLLM) [15, 28, 40] further introduce vision encoders and open the doors for LLMs to interpret not only textual data but also images and data in other modalities [5, 29, 58]. In the field of autonomous driving (AD), recent research has integrated LLMs into the AD system for better explainability and natural interaction with humans. Some studies adopt the visual language model approach, which can handle multi-modal input data and provide the textual description as well as the control signal for the driving scenarios. For example, DRIVEGPT4 [52] proposed a multimodal LLM framework, which takes a sequence of frames as input, then generates responses to human inquiries and predicts control signals for the next step. However, since the framework lacks an input command, the predicted control can not follow the specific navigation command, which denotes that the framework is hard to deploy in real scenarios. Meanwhile, more researchers focus on transforming the driving situations into textual descriptions as the input for the LLM, for directly interpreting and reasoning about comprehensive driving situations. In this thread of works, GPT-Driver [32] reformulated motion planning as the task of natural language modeling by converting heterogeneous scene input to language tokens. LanguageMPC [36] leveraged a LLM to reason the complex scenarios and output high-level driving decisions. Then the method tunes a parameter matrix to convert the decision into the low-level control signals. LLM-Driver [6] utilized the numeric vector as the" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 542, + 115 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 542, + 115 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 542, + 115 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 542, + 115 + ], + "type": "image", + "image_path": "99916684b9b3a582b272613337d3e86d4bcd7b50afde67059f2e532c163a0b9c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 118, + 529, + 125 + ], + "lines": [ + { + "bbox": [ + 321, + 118, + 529, + 125 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 529, + 125 + ], + "type": "text", + "content": "Navigation instruction: Just move to the left and get ready to leave the highway." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 128, + 542, + 173 + ], + "blocks": [ + { + "bbox": [ + 308, + 128, + 542, + 173 + ], + "lines": [ + { + "bbox": [ + 308, + 128, + 542, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 128, + 542, + 173 + ], + "type": "image", + "image_path": "0aec77596cf85f55a6a5998f1666fcd1e1ada32308eda38bd2d5a9fe0e92bfdb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 174, + 487, + 182 + ], + "lines": [ + { + "bbox": [ + 345, + 174, + 487, + 182 + ], + "spans": [ + { + "bbox": [ + 345, + 174, + 487, + 182 + ], + "type": "text", + "content": "Navigation instruction: Turn left at the next T-junction" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 346, + 182, + 517, + 189 + ], + "lines": [ + { + "bbox": [ + 346, + 182, + 517, + 189 + ], + "spans": [ + { + "bbox": [ + 346, + 182, + 517, + 189 + ], + "type": "text", + "content": "Notice instruction: Please watch out for the pedestrians up ahead." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 228, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 305, + 195, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 217 + ], + "type": "text", + "content": "Figure 2. Two examples of the collected data with corresponding labeled navigation instructions and optional notice instructions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 228, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 293 + ], + "type": "image", + "image_path": "1fc85fac44906cdcf24e88d93be1e44c83f39d10fcc42db41610e0573a554bb5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 294, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 294, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 294, + 545, + 316 + ], + "type": "text", + "content": "Figure 3. Distribution of parsed clips in terms of clip length and the corresponding navigation instruction type." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 328, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 328, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 305, + 328, + 545, + 362 + ], + "type": "text", + "content": "input modality and fused vectorized object-level 2D scene representation to enable the LLM to answer the questions based on the current environment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 546, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 546, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 546, + 469 + ], + "type": "text", + "content": "However, this line of work only considered the driving problem in the open-loop settings, and ignored questions such as cumulative error, temporal action consistency, and end-to-end trainability, which are critical for bringing the models into actual closed-loop driving tasks. To the best of our knowledge, we are the first language-based end-to-end autonomous driving method in the closed-loop setting. Relevant datasets, benchmarks, and trained models are also open-sourced to facilitate further research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 481, + 417, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 417, + 495 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 417, + 495 + ], + "type": "text", + "content": "3. Dataset generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 501, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 714 + ], + "type": "text", + "content": "We aim to develop an intelligent driving agent that can generate driving actions based on three sources of input: 1) sensor data (multi-view camera and LiDAR), so that the agent can generate actions that are aware of and compliant with the current scene; 2) navigation instructions (e.g., lane changing, turning), so that the agent can drive to meet the requirement in natural language (instruction from humans or navigation software); and 3) human notice instruction, so that the agent can interact with humans and adapt to human's suggestions and preferences (e.g., pay attention to adversarial events, deal with long-tail events, etc). In this section, we describe how to generate the multi-modal dataset needed to train the agent, and the prompt design for the navigation instruction and human notice instruction. Specifically, we choose the CARLA [12] as the simulator, because it can simulate a realistic dynamic closed-loop world and it is widely adopted in the field of end-to-end autonomous driving. The data collection consists of two stages: 1) col" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15122" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 71, + 286, + 202 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 286, + 202 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 286, + 202 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 286, + 202 + ], + "type": "table", + "html": "
TypeThree randomly chosen instructions of each instruction type
FollowMaintain your current course until the upcoming intersection.\nIn [x] meters, switch to left lane.\nEase on to the left and get set to join the highway.
TurnAfter [x] meters, take a left.\nAt the next intersection, just keep heading straight, no turn.\nYou'll be turning left at the next T-junction, alright?
OthersFeel free to start driving.\nSlow down now.\nHead to the point, next one's [x] meters ahead, [y] meters left/right.
NoticeWatch for walkers up front.\nJust a heads up, there's a bike ahead.\nPlease be aware of the red traffic signal directly in front of you.
", + "image_path": "9e62b417cdf6d6f50d489c4910bde7dc50dcf4ae92ca11dd517fa2bc3e1d2344.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "type": "text", + "content": "Table 1. Examples of considered navigation instructions (follow, turn, others) and notice instructions. " + }, + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "type": "inline_equation", + "content": "[x]" + }, + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "type": "inline_equation", + "content": "[y]" + }, + { + "bbox": [ + 47, + 206, + 287, + 239 + ], + "type": "text", + "content": " represent the float number for a specific distance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 275 + ], + "type": "text", + "content": "lecting sensor data and control signals with an expert agent; and 2) parsing and labeling collected data with instructions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 432 + ], + "type": "text", + "content": "Sensor and control data collection. We utilize a rule-based expert agent [38] to create a dataset including about 3M driving frames. Since the expert agent can access the privileged information in the CARLA, this dataset will include camera data, LiDAR data, and control actions for each frame. To enhance the diversity of the collected dataset, the agent runs on 2.5k routes, 8 towns, and 21 kinds of environmental conditions (e.g., weather, time of the day). We use four RGB cameras (left, front, right, rear) and one LiDAR. The side cameras are angled at " + }, + { + "bbox": [ + 46, + 278, + 287, + 432 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 46, + 278, + 287, + 432 + ], + "type": "text", + "content": ". Besides, we center-crop the front image as an additional focus-view image to capture the status of the distant traffic light. The LiDAR has 64 channels and generates 600K points per second." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "content": "Parsing and language annotation. In the second stage, we parse the collected data into clips, and label each clip with proper navigation instructions and optional notice instructions. The parsing process takes a sequence of frames as input, and segments these frames into clips, where each clip corresponds to one navigation instruction. For instance, if the agent started to turn left at frame " + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "content": " and ended at frame " + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "inline_equation", + "content": "T_{n}" + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "content": ", we will label " + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "inline_equation", + "content": "(T_{0}, T_{n})" + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "content": " as a new clip with the instruction \"Hang a left at the next crossroads\". Besides, if an adversarial event occurs at time " + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "inline_equation", + "content": "T_{a}" + }, + { + "bbox": [ + 46, + 435, + 287, + 659 + ], + "type": "text", + "content": ", we will add one notice instruction into this clip, simulating a real-life scenario where a passenger or a side assistance system would communicate with the driver when an emergency happens. As shown in Figure 2, each clip includes sensor data, control signals, the corresponding navigation instruction, and optional notice instructions. The distribution of the parsed clips in terms of clip length and corresponding instruction is shown in Figure 3. In our dataset, we collect 64K parsed clips and 464K notice instructions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 662, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 286, + 674 + ], + "type": "text", + "content": "Instruction design. We consider three types of navigation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "content": "instructions (follow, turn, and others) along with one type of notice instruction, consisting of a total of 56 different instructions. Table 1 shows some examples and the full list can be found in supplementary material. To enable the agent to drive in realistic instructional settings where the instructions come from navigation software or humans, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 143, + 545, + 390 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 143, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 143, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 143, + 545, + 225 + ], + "type": "text", + "content": "- Diversifying the instructions: Considering the inherent richness of natural language, for each type of instruction, we utilize ChatGPT to generate 8 different variants, each carrying the same semantic meaning but varying in phrasing. This enables more comprehensive coverage and flexibility in language interpretation, accommodating the diverse ways the same instruction can be conveyed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 225, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 545, + 319 + ], + "type": "text", + "content": "- Incorporating misleading instructions: in real-world cases, the navigation software or passengers may give misleading instructions to the AV that violate traffic rules or raise safety concerns. For example, on a single-lane road, following an instruction \"Change to left lane\" is dangerous. To improve the robustness of our model against misleading instructions, we simulate these scenarios and add them to our dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 319, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 545, + 390 + ], + "type": "text", + "content": "- Connecting multiple instructions: In many cases, the instructions may consist of two to three consecutive instructions, such as \"Turn right at this intersection, then go straight to the next intersection and turn right again.\" We also construct some consecutive complex instruction data to simulate real navigation-based driving scenarios." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 401, + 437, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 437, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 437, + 415 + ], + "type": "text", + "content": "4. LMDrive methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 574 + ], + "type": "text", + "content": "In this work, we propose LMDrive, a framework that can understand and follow high-level driving instructions through natural language. As illustrated in Figure 4, LMDrive is composed of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a LLM and its associated component (tokenizer, Q-Former, and adapters) that takes in the visual tokens and language instruction, to predict the control signal and whether the given instruction is completed. We will introduce the vision encoder in Section 4.1, and the language model with its associated components in Section 4.2. Finally, we describe the training details in Section 4.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 582, + 398, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 398, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 398, + 594 + ], + "type": "text", + "content": "4.1. Vision encoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 595, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 713 + ], + "type": "text", + "content": "In the visual language community [5, 29, 58], the most common way to align vision and language could be using pre-trained CLIP models [35] to encode image features. However, the large flops and parameter size of the CLIP models increase the difficulty of its deployment in AV systems. Also, AV perception systems are usually in 3D to include LiDAR input. Hence, inspired by InterFuser [38] and " + }, + { + "bbox": [ + 304, + 595, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TF}++" + }, + { + "bbox": [ + 304, + 595, + 545, + 713 + ], + "type": "text", + "content": " [20], we design a multi-view multi-modality vision encoder to encode/fuse the sensor data. As shown in Figure 5, the vision encoder consists of the sensor encoding" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 713 + ], + "type": "text", + "content": "1The adversarial events include bad road conditions, the front vehicle's sudden brake, unexpected entities rushing into the road from occluded regions, vehicles running a red traffic light, etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15123" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 531, + 329 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 531, + 329 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 531, + 329 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 531, + 329 + ], + "type": "image", + "image_path": "427c0bfa908c49a0b7fa806fcfbf8e9690e4ea4de375af29aa9e02402f7fd156.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 548, + 381 + ], + "type": "text", + "content": "Figure 4. The structure of the proposed LMDrive model, which consists of two major components: 1) a vision encoder that processes multi-view multi-modal sensor data (camera and LiDAR) for scene understanding and generating visual tokens; 2) a large language model and its associated component (tokenizer, Q-Former, and adapters) that processes all the historic visual tokens and the language instructions (navigation instruction and optional notice instruction), to predict the control signal and whether the given instruction is completed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 391, + 288, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 288, + 462 + ], + "type": "text", + "content": "part which encodes image and LiDAR input respectively, and a BEV decoder that fuses image and point cloud features to generate visual tokens which are then passed to the language model. Notably, the vision encoder is pre-trained on perception tasks by adding additional prediction heads, then the encoder is frozen for later use by the LLM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "text", + "content": "Sensor encoding. For each image input, we apply a 2D backbone ResNet [17] to extract the image feature map. The feature map is then flattened into one-dimensional tokens. For a comprehensive understanding of the global context from multiple viewpoints, tokens from different views will be fused by a standard " + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "inline_equation", + "content": "K_{enc}" + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "text", + "content": "-layer transformer encoder, each layer containing Multi-Headed Self-Attention [48], MLP blocks and layer normalization [1]. For the LiDAR input, we adopt a 3D backbone PointPillars [23] to process the raw point cloud data into ego-centered LiDAR features, where each pillar encompasses points within a " + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{m} \\times 0.25\\mathrm{m}" + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "text", + "content": " area. PointNet [34] is then used to aggregate features and downsample the feature map to " + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "inline_equation", + "content": "C \\times H \\times W" + }, + { + "bbox": [ + 46, + 469, + 287, + 635 + ], + "type": "text", + "content": ", which subsequently serve as BEV queries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "BEV decoder. The encoded sensor features above are then passed into the BEV decoder to generate visual tokens. Specifically, the BEV decoder is designed as a standard transformer with " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "K_{dec}" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " layers. The BEV point cloud features are fed into the BEV decoder as " + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": " queries to attend to the multi-view image features and generate BEV" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 392, + 544, + 488 + ], + "blocks": [ + { + "bbox": [ + 310, + 392, + 544, + 488 + ], + "lines": [ + { + "bbox": [ + 310, + 392, + 544, + 488 + ], + "spans": [ + { + "bbox": [ + 310, + 392, + 544, + 488 + ], + "type": "image", + "image_path": "1b43f97a0b49cab9a0c0146df9ddb2d5a337539e7f86d541eef1e1ff264bcb4d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 495, + 545, + 581 + ], + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 581 + ], + "type": "text", + "content": "Figure 5. The detailed structure of the vision encoder, which takes as input the multi-view multi-modality sensor data. In the pre-training stage, the vision encoder is appended with prediction headers to perform pre-training tasks (object detection, traffic light status classification, and future waypoint prediction). In the instruction-finetuning stage and inference stage, the prediction headers are discarded, and the vision encoder is frozen to generate visual tokens to feed into the LLM." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "text", + "content": "tokens. We also feed " + }, + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "text", + "content": " learnable queries and 1 learnable query into the BEV decoder to generate " + }, + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 590, + 545, + 662 + ], + "type": "text", + "content": " waypoint point tokens and 1 traffic light token respectively. Thus the three types of visual tokens (BEV, waypoint, and traffic light) will contain rich scene information and will be then presented to the large language model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Pre-training with prediction headers. we consider three vision encoder pre-training tasks: object detection, future waypoint prediction, and traffic light status classification. For object detection, the BEV tokens will pass through a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15124" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "content": "one-stage CenterPoint [53] to predict the bounding boxes and velocity of the objects in an " + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "inline_equation", + "content": "H\\mathrm{m} \\times W\\mathrm{m}" + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "content": " area. For waypoint prediction, we pass the " + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "content": " waypoint tokens along with the navigation waypoint into the GRU network [9] sequentially to predict " + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "content": " future waypoint. For the traffic light status classification, a 2-layer MLP is applied to the traffic light token. Three corresponding loss terms are considered: 1) the detection loss as in InterFuser [39]; 2) the " + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 46, + 72, + 289, + 214 + ], + "type": "text", + "content": " waypoint loss; and 3) the cross-entropy traffic light state loss. Note that these prediction headers are only used in the pretraining of the vision encoder, and will be discarded in the training of the LLMs and inference of the whole model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 225, + 272, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 272, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 272, + 239 + ], + "type": "text", + "content": "4.2. LLM for instruction-following auto driving" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 245, + 289, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 289, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 289, + 387 + ], + "type": "text", + "content": "As illustrated in Figure 4, in our framework, the LLM functions as the \"brain\" throughout the entire driving procedure, processing sensor tokens generated by the frozen vision encoder for each frame, comprehending natural language instructions, generating the necessary control signal, and predicting whether the given instruction is completed. Specifically, we choose LLaMA [45] as the language backbone, which has been widely adopted in many language [16, 55] and vision [29, 58] instruction-tuning models. We also have three associated components to bridge LLM with instruction, visual information input, and action prediction: 1) a tokenizer, 2) a Q-Former, 3) two adapters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": "Instruction and visual tokenization. Given the navigation instruction and optional notice instruction, we apply the LLaMA tokenizer [45] to convert the instruction into textual tokens. Note that the duration of executing one instruction would span from a few seconds to a few minutes, and our model is deployed in the closed-loop setting. Thus at each frame, we utilize all historic sensor information (with maximum limit " + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "inline_equation", + "content": "T_{max}" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": ") to depress cumulative error and improve the temporal consistency of the model. Specifically, for each frame's multi-view multi-modality sensor input, we utilize the vision encoder pre-trained in the previous section to generate visual tokens (" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": " BEV tokens, " + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": " waypoint tokens, and one traffic light token). However, the number of visual tokens (e.g., 406 tokens for each frame) quickly grows too large for the LLM because usually hundreds of frames are needed to complete one instruction. To overcome this, we follow BLIP-2 [24] to use the Q-Former to reduce the number of visual tokens. Specifically, for each frame, we employ " + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": " learnable queries to attend to the visual tokens via cross-attention layers, which can reduce each frame's visual token number to " + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 391, + 289, + 674 + ], + "type": "text", + "content": ". Subsequently, we use a 2-layer MLP adapter to convert the tokens extracted by the Q-Former to share the same dimension as the language token, which can then be fed into the LLM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 678, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 678, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 288, + 715 + ], + "type": "text", + "content": "Action prediction. After receiving a sequence of instructional and visual tokens, the LLM predicts the action tokens. One another 2-layer MLP adapter is then applied to predict" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "future waypoints, as well as a flag to indicate whether the given instruction has been completed. Note that to enhance the supervision signal, we will also conduct prediction for every historic frame during training, and only the prediction of the latest frame will be executed at inference time. To get the final control signal, which includes braking, throttling, and steering, following LBC [4], we use two PID controllers for latitudinal and longitudinal control to track the heading and velocity of predicted waypoints respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 181, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 228 + ], + "type": "text", + "content": "Training objectives. When finetuning the LLM and its associated components, we consider two loss terms: 1) the " + }, + { + "bbox": [ + 304, + 181, + 545, + 228 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 304, + 181, + 545, + 228 + ], + "type": "text", + "content": " waypoint loss; 2) the classification loss (cross-entropy), indicating if the current frame completes the instruction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 235, + 403, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 403, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 403, + 248 + ], + "type": "text", + "content": "4.3. Training details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 254, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 289 + ], + "type": "text", + "content": "LMDrive's training consists of two stages: 1) the vision encoder pre-training stage; and 2) the instruction-finetuning stage, to align the instruction/vision and control signal." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 292, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 386 + ], + "type": "text", + "content": "Vision encoder pre-training stage. The vision encoder takes a single frame's sensor data as input, and we use the dataset collected in Section 3 for training. Specifically, since the instruction annotation process will drop some frames, we use the raw dataset before the instruction annotation for the vision encoder pre-training, which includes data of around 3M frames. Only the vision encoder is pretrained with perception tasks for scene understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "text", + "content": "Instruction-finetuning stage. The entire system is trained for end-to-end autonomous driving under the guidance of the instruction, where the Q-Former and Adapters are trainable and the other components are frozen. While our LMDrive takes a sequence of frames as input, during training we set a fixed sequence length " + }, + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "inline_equation", + "content": "T_{max}" + }, + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "text", + "content": " for building up batch data. The training utilizes the instruction-following data generated in Section 3. To enable the model to reject misleading instructions, we label the corresponding data as 'completed' after the misleading instruction is given for about 1 second. Since the dataset is collected at a high frequency (" + }, + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\sim 10\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 388, + 545, + 589 + ], + "type": "text", + "content": "), the data in adjacent frames are highly similar. To encourage efficient training, following video prediction methods [13, 37, 49], we sample the training frames in a fixed interval, and apply temporal augmentation which randomly shifts the training frames either forward or backward, with the random shift less than the fixed interval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 599, + 435, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 435, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 435, + 612 + ], + "type": "text", + "content": "5. LangAuto Benchmark" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "type": "text", + "content": "We propose the LangAuto (Language-guided Autonomous Driving) CARLA benchmark, the first benchmark that evaluates closed-loop driving performance under language instructions. Compared with the previous CARLA benchmark, Town05 [33] and Longest6 [8] that navigate the agent with discrete driving commands or target waypoints, our benchmark only provides the AV with a navigation instruction and optional notice instructions in natural language." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15125" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 519, + 161 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 519, + 161 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 519, + 161 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 519, + 161 + ], + "type": "table", + "html": "
LLM BackboneLangAutoLangAuto-ShortLangAuto-Tiny
DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑DS ↑RC ↑IS ↑
Random Init.10.7±3.816.2±4.90.63±0.0414.2±4.420.1±4.40.72±0.0420.1±4.124.7±5.10.75±0.03
LLaMA [45]31.3±1.537.1±1.60.82±0.0142.8±7.249.1±8.50.87±0.0352.2±5.357.8±8.00.91±0.05
LLaMA2 [46]32.8±2.140.1±2.20.81±0.0244.8±6.253.5±5.50.84±0.0256.1±4.164.2±4.70.87±0.04
Vicuna [55]33.5±1.939.3±1.90.83±0.0245.3±4.954.3±3.90.83±0.0355.5±3.963.1±4.20.88±0.04
Vicuna-v1.5 [55]34.0±3.839.0±3.30.85±0.0647.0±4.356.5±2.40.83±0.0459.0±2.669.9±2.30.84±0.02
LLaVA-v1.5 [27]36.2±2.346.5±4.30.81±0.0350.6±1.760.0±3.40.84±0.0466.5±3.677.9±2.30.85±0.02
", + "image_path": "c9bf7ebabfaec255f49030123cbac60231aa9413350041fc907b9fde34b749a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 183, + 273, + 243 + ], + "blocks": [ + { + "bbox": [ + 62, + 163, + 530, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 163, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 530, + 175 + ], + "type": "text", + "content": "Table 2. Performance comparison of 6 LLM backbones on the LangAuto benchmark. We report the metrics for 3 evaluation runs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 183, + 273, + 243 + ], + "lines": [ + { + "bbox": [ + 61, + 183, + 273, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 273, + 243 + ], + "type": "table", + "html": "
Module designDS ↑RC ↑IS ↑
Baseline (LLaVA-v1.5)36.2±2.346.5±4.30.81±0.03
w/o Q-Former31.7±3.541.2±4.40.79±0.02
w/o using BEV tokens33.9±3.945.9±5.10.72±0.03
w/o visual pre-training16.9±5.124.1±4.70.70±0.04
", + "image_path": "11e1427418c7539fb327bd400bd941f5035b425538f21a345c98069399db855b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 243, + 251, + 255 + ], + "lines": [ + { + "bbox": [ + 82, + 243, + 251, + 255 + ], + "spans": [ + { + "bbox": [ + 82, + 243, + 251, + 255 + ], + "type": "text", + "content": "Table 3. Ablation study on the module design." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 269, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 376 + ], + "type": "text", + "content": "Specifically, the LangAuto benchmark covers all 8 publicly available towns in CARLA to include various scenarios (e.g., highways, intersections, roundabouts). We also consider 16 kinds of environmental conditions, encompassing the combinations of 7 weather conditions (Clear, Cloudy, Wet, MidRain, WetCloudy, HardRain, SoftRain) and 3 daylight conditions (Night, Noon, Sunset). Besides, LangAuto consists of three tracks to fully test the agent's instruction-following abilities:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 377, + 287, + 624 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 377, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 377, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 47, + 377, + 287, + 471 + ], + "type": "text", + "content": "- LangAuto track: For each route, navigation instructions are given and updated to the agent based on the agent's current position. We also divide this track into three sub-tracks with different route lengths, to better distinguish the performance. LangAuto where the routes are longer than 500 meters, LangAuto-Short where the route length is between 150 and 500 meters, and LangAuto-Tiny where the route length is shorter than 150 meters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 472, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 566 + ], + "type": "text", + "content": "- LangAuto-Notice track: Based on the LangAuto track, we additional add notice instructions to the agent. This setting simulates real cases where passengers or other assistance systems can give real-time notice in long-trail complex or adversarial scenarios, which is usually hard for the AV system to handle by itself. Ideally, the agent that can comprehend and leverage instruction can achieve better performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 566, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 287, + 624 + ], + "type": "text", + "content": "- LangAuto-Sequential track: Based on the LangAuto track, we merge " + }, + { + "bbox": [ + 47, + 566, + 287, + 624 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 566, + 287, + 624 + ], + "type": "text", + "content": " of consecutive 2 to 3 instructions into a single long instruction. This setting mimics the realistic scenarios where the multi-sentence instructions come from the passengers or navigation software." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 626, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 286, + 696 + ], + "type": "text", + "content": "Note that misleading instructions will be randomly " + }, + { + "bbox": [ + 46, + 626, + 286, + 696 + ], + "type": "inline_equation", + "content": "(\\sim 5\\%)" + }, + { + "bbox": [ + 46, + 626, + 286, + 696 + ], + "type": "text", + "content": " and intermittently given to the driving agent, which lasts for a certain duration (1-2 seconds). The driving agent is expected to reject these misleading instructions and execute safe actions that are compliant to the current scene, until the next correct instruction is spawned." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "Metrics. We consider three major metrics introduced by the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 185, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 363 + ], + "type": "text", + "content": "CARLA LeaderBoard [42]: route completion (RC), infraction score (IS), and driving score (DS). The route completion refers to the percentage of the total route length that has been completed. It only takes into account the distance traveled along the predetermined route, where each segment of the predetermined route corresponds to a navigation instruction. If the agent deviates too far from the route, the agent is regarded as violating the instruction, and this episode is marked as a failure and terminated. The infraction score measures infractions triggered by the agent. When collisions or traffic rule violations occur, the infraction score is decayed by a corresponding discount factor. The driving score is the product of the route completion ratio and the infraction score, describing both driving progress and safety. It is generally recognized as the primary ranking metric." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 387, + 388 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 413, + 408 + ], + "type": "text", + "content": "6.1. Experiment Setup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "content": "We implement and evaluate our approach on the open-source CARLA simulator of version 0.9.10.1 [12]. For the 2D backbone and 3D backbone of the vision encoder, ResNet-50 [17] backbone is pre-trained on ImageNet, and PointPillars [23] is trained from scratch. " + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "inline_equation", + "content": "C, H, W, N" + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "content": " are set as 256, 50, 50, 5 respectively. For the Q-former token number " + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "content": ", we empirically found that " + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "content": " tokens achieve decent performance. During the instruction-finetuning stage, we sample the training frames in a fixed interval of 2. Because each parsed clip typically contains multiple notice instructions, to avoid overfitting to notice following and generating over-conservative behaviors, during training we randomly removed notices for " + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 414, + 545, + 591 + ], + "type": "text", + "content": " of the clips. For the remaining clips, we ensure that a maximum of one notice is included. We refer readers to the appendix for more details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 600, + 424, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 600, + 424, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 600, + 424, + 613 + ], + "type": "text", + "content": "6.2. Quantitative Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 545, + 713 + ], + "type": "text", + "content": "LLM backbones. In Table 2, we evaluate our method with different pretrained or randomly initialized LLM models of 7B parameters: LLaMA [45] and LLaMA2 [46] were pretrained with large public language datasets; Vicuna [55] and Vicuna-v1.5 [55] are additionally finetuned with human conversation data; LLaVA-v1.5 [27] incorporate multi-modal data (e.g., language, images) for training. We observe that LLaVA-v1.5 [27] surpasses the other" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15126" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 71, + 520, + 144 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 520, + 144 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 520, + 144 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 520, + 144 + ], + "type": "table", + "html": "
LLM BackboneBenchmark TypeInfraction Score ↑Vehicle Collisions ↓Pedestrian Collisions ↓Layout Collisions ↓Red light Violations ↓Offroad Infractions ↓Blocked Infractions ↓
LLaVA-v1.5LangAuto0.810.330.030.500.920.360.22
LangAuto-Notice0.870.170.020.310.500.170.27
Vicuna-v1.5LangAuto0.850.300.030.431.180.240.19
LangAuto-Notice0.910.150.010.280.560.260.19
", + "image_path": "e571fa3be200b2da5519919a65b2b3e4b48cd82a77f64f4e8c29f1deab786fdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 182, + 274, + 243 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 547, + 172 + ], + "type": "text", + "content": "Table 4. Performance comparison on LangAuto and LangAuto-Notice benchmarks. The metrics (except infraction score) are normalized by the driven distance (km)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 182, + 274, + 243 + ], + "lines": [ + { + "bbox": [ + 60, + 182, + 274, + 243 + ], + "spans": [ + { + "bbox": [ + 60, + 182, + 274, + 243 + ], + "type": "table", + "html": "
LLM BackboneBenchmark TypeDS ↑RC ↑IS ↑
LLaVA-v1.5LangAuto36.246.50.81
LangAuto-Sequential34.043.70.81
Vicuna-v1.5LangAuto34.039.00.85
LangAuto-Sequential31.937.10.84
", + "image_path": "2488bf90d98b1ef4013bc01f8a13e011e165ad46b6765f28a64693fdef6d3fc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 243, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 47, + 243, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 287, + 266 + ], + "type": "text", + "content": "Table 5. Performance comparison on LangAuto and LangAutoSequential benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "text", + "content": "LLM models, which demonstrates the importance of adopting pretrained multi-modal LLMs for instruction-following autonomous driving. Besides, the models finetuned with instruction data perform better than others (Vicuna-v1.5 " + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "text", + "content": " LLaMA2 " + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\approx" + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "text", + "content": " Vicuna " + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "inline_equation", + "content": ">" + }, + { + "bbox": [ + 46, + 280, + 287, + 386 + ], + "type": "text", + "content": " LLaMA). We also test a randomly initialized 7B LLM model, which struggles to drive properly with the same amount of training data, which demonstrates the necessity of finetuning pretrained LLM for instruction-following driving." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "content": "Module Design. In Table 3, we conduct ablation studies on different components. First, instead of reducing the number of BEV tokens with the Q-Former, we directly down-sample the BEV features to " + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 388, + 287, + 567 + ], + "type": "text", + "content": ", then feed them into the LLM (denoted as \"w/o Q-Former\"). The average driving score dropped from 36.2 to 31.7. Second, excluding the BEV tokens input into LLM decoder (denoted as \"w/o using BEV tokens\" results in a decreased infraction score (0.81 to 0.72). One explanation is that the BEV tokens are important for detecting and reasoning the surrounding obstacles and road structures. Third, we remove the pre-training stage of the vision encoder and directly train it from scratch in the instruction-finetuning stage. The driving score drops to 16.9 (denoted as \"w/o visual pre-training\"), demonstrating the importance of our proposed vision encoder pre-training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 639 + ], + "type": "text", + "content": "LangAuto-Notice Benchmark. The LangAuto-Notice benchmark provides some notice instructions to the agent when adversarial events happen. As shown in Table 4, our agents can effectively leverage the information of the notice in real-time, resulting in a significant decrease in both collisions and traffic rule violations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "LangAuto-Sequential Benchmark. In Table 5, we demonstrate the effect of the LangAuto-Sequential benchmark where some consecutive 2 to 3 navigation instructions are merged together to a long and complex instruction. This setting additionally requires the agent to be able to be temporally aware of which instruction has been completed and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 184, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 308, + 184, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 305 + ], + "type": "image", + "image_path": "878d4cde10f08e75e7f00ff2dcbadcc15bea88def0caed84148cd2a3e9ae9c63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 311, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 387 + ], + "type": "text", + "content": "Figure 6. An example of how our LMDrive predicts future waypoints, given sensor inputs and varied navigational instructions. Under the first two instructions, our model predicts different waypoints accordingly. The third instruction is a misleading one (turn right on a left-only lane). The model appropriately rejects the incorrect instruction, generating a slower speed and a safe path that is compliant with the scenario." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 396, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 419 + ], + "type": "text", + "content": "which has not. Our agents based on LLaVA and Vicuna both have a performance drop." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 379, + 445 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 453, + 545, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 630 + ], + "type": "text", + "content": "In this paper, we introduced LMDrive, a language-guided, end-to-end, closed-loop autonomous driving framework. LMDrive incorporates natural language instructions along with sensor data, enabling human-like interaction and navigation in complex driving scenarios. We also propose the language-guided driving dataset, comprising around 64K multi-modal data clips along with corresponding navigation instructions. We established the LangAuto benchmark for evaluating autonomous driving systems considering natural language instructions. The effectiveness of LMDrive was demonstrated through extensive closed-loop experiments, underlining the potential of improving the interaction of autonomous vehicles with humans and the environment. Our work paves the way for future advancements in language-based autonomous driving." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 631, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement. This project is funded in part by National Key R&D Program of China Project 2022ZD0161100, by the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, by General Research Fund of Hong Kong RGC Project 14204021. Hongsheng Li is a PI of CPII under the InnoHK." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15127" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 189 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 189 + ], + "type": "text", + "content": "[2] Sergio Casas, Cole Gulino, Simon Suo, Katie Luo, Renjie Liao, and Raquel Urtasun. Implicit latent variable model for scene-consistent motion forecasting. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 624-641. Springer, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 190, + 287, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 190, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 53, + 190, + 287, + 232 + ], + "type": "text", + "content": "[3] Dian Chen and Philipp Krahenbuhl. Learning from all vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17222-17231, 2022. 3, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 233, + 287, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 233, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 53, + 233, + 287, + 265 + ], + "type": "text", + "content": "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 287, + 308 + ], + "type": "text", + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 310, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 287, + 363 + ], + "type": "text", + "content": "[6] Long Chen, Oleg Sinavski, Jan Hünermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with llms: Fusing object-level vector modality for explainable autonomous driving. arXiv preprint arXiv:2310.01957, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 287, + 407 + ], + "type": "text", + "content": "[7] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. arXiv preprint arXiv:2306.16927, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 408, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 408, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 53, + 408, + 287, + 461 + ], + "type": "text", + "content": "[8] Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 462, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 462, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 53, + 462, + 287, + 516 + ], + "type": "text", + "content": "[9] Kyunghyun Cho, Bart Van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 517, + 287, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 287, + 570 + ], + "type": "text", + "content": "[10] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416, 2022.3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 602 + ], + "type": "text", + "content": "[11] DriveLM Contributors. Drivelm: Drive on language. https://github.com/OpenDriveLab/DriveLM, 2023.2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 604, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 604, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 604, + 287, + 647 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 691 + ], + "type": "text", + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[14] Lan Feng, Quanyi Li, Zhenghao Peng, Shuhan Tan, and Bolei Zhou. Trafficgen: Learning to generate diverse and re" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "alistic traffic scenarios. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 3567-3575. IEEE, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 547, + 171 + ], + "type": "text", + "content": "[15] Peng Gao, Renrui Zhang, Chris Liu, Longtian Qiu, Siyuan Huang, Weifeng Lin, Shitian Zhao, Shijie Geng, Ziyi Lin, Peng Jin, Kaipeng Zhang, Wenqi Shao, Chao Xu, Conghui He, Junjun He, Hao Shao, Pan Lu, Hongsheng Li, and Yu Qiao. Sphinx-x: Scaling data and parameters for a family of multi-modal large language models, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 215 + ], + "type": "text", + "content": "[16] Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. Koala: A dialogue model for academic research. Blog post, April, 1, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5, 7, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 316 + ], + "type": "text", + "content": "[18] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 338 + ], + "type": "text", + "content": "[19] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 370 + ], + "type": "text", + "content": "[20] Bernhard Jaeger, Kashyap Chitta, and Andreas Geiger. Hidden biases of end-to-end driving models. arXiv preprint arXiv:2306.07957, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 437 + ], + "type": "text", + "content": "[21] Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7953-7963, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 502 + ], + "type": "text", + "content": "[22] Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21983-21994, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 547, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 547, + 559 + ], + "type": "text", + "content": "[23] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 6, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[25] Quanyi Li, Zhenghao Mark Peng, Lan Feng, Zhizheng Liu, Chenda Duan, Wenjie Mo, and Bolei Zhou. ScenarioNet: Open-source platform for large-scale traffic scenario simulation and modeling. Advances in neural information processing systems, 36, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 713 + ], + "type": "text", + "content": "[26] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15128" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 129 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 286, + 161 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 286, + 195 + ], + "type": "text", + "content": "[30] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 196, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 286, + 228 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 228, + 286, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 286, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 286, + 261 + ], + "type": "text", + "content": "[32] Jiageng Mao, Yuxi Qian, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 261, + 286, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 286, + 305 + ], + "type": "text", + "content": "[33] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 286, + 350 + ], + "type": "text", + "content": "[34] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb data. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 918-927, 2018. 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 350, + 286, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 286, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 286, + 415 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 286, + 470 + ], + "type": "text", + "content": "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. Languagempc: Large language models as decision makers for autonomous driving. arXiv preprint arXiv:2310.03026, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 286, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 286, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 286, + 492 + ], + "type": "text", + "content": "[37] Hao Shao, Shengju Qian, and Yu Liu. Temporal interlacing network. AAAI, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 286, + 536 + ], + "type": "text", + "content": "[38] Hao Shao, Letian Wang, Ruobing Chen, Hongsheng Li, and Yu Liu. Safety-enhanced autonomous driving using interpretable sensor fusion transformer. In Conference on Robot Learning, pages 726-737. PMLR, 2023. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 286, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 286, + 592 + ], + "type": "text", + "content": "[39] Hao Shao, Letian Wang, Ruobing Chen, Steven L Waslander, Hongsheng Li, and Yu Liu. Reasonnet: End-to-end driving with temporal and global reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13723-13733, 2023. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 593, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 286, + 635 + ], + "type": "text", + "content": "[40] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 636, + 286, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 286, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 286, + 680 + ], + "type": "text", + "content": "[41] Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement. Advances in Neural Information Processing Systems, 35:6531-6543, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[42] CARLA team. Carla autonomous driving leaderboard. https://leaderboard.carla.org/, 2020. Accessed: 2021-02-11. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[43] Trisha Thadani. Cruise recalls all its driverless cars after pedestrian hit and dragged. The Washintong Post. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[44] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7153-7162, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 207 + ], + "type": "text", + "content": "[45] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[46] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 264, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 287 + ], + "type": "text", + "content": "[47] Jordan Valinsky. 'complete meltdown': Driverless cars in san francisco stall causing a traffic jam. CNN Business. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 288, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 332 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pages 5998-6008, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 388 + ], + "type": "text", + "content": "[49] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 442 + ], + "type": "text", + "content": "[50] Letian Wang, Jie Liu, Hao Shao, Wenshuo Wang, Ruobing Chen, Yu Liu, and Steven L Waslander. Efficient reinforcement learning for autonomous driving with parameterized skills and priors. arXiv preprint arXiv:2305.04412, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 499 + ], + "type": "text", + "content": "[51] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline. Advances in Neural Information Processing Systems, 35:6119-6132, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 555 + ], + "type": "text", + "content": "[52] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kenneth KY Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. arXiv preprint arXiv:2310.01412, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 601 + ], + "type": "text", + "content": "[53] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 657 + ], + "type": "text", + "content": "[54] Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitating a reinforcement learning coach. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15222-15232, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[55] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 6, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15129" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 204 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Tong Zhou, Letian Wang, Ruobing Chen, Wenshuo Wang, and Yu Liu. Accelerating reinforcement learning for autonomous driving using task-agnostic and ego-centric motion skills. arXiv preprint arXiv:2209.12072, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "type": "text", + "content": "[57] Yang Zhou, Hao Shao, Letian Wang, Steven L. Waslander, Hongsheng Li, and Yu Liu. Smartrefine: A scenario-adaptive refinement framework for efficient motion prediction, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 204 + ], + "type": "text", + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3, 4, 6" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15130" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_content_list.json b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..589938cc715ceddf666832f16ac99f6ef7a91190 --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_content_list.json @@ -0,0 +1,1581 @@ +[ + { + "type": "text", + "text": "LORS: Low-rank Residual Structure for Parameter-Efficient Network Stacking", + "text_level": 1, + "bbox": [ + 83, + 130, + 885, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, Chengjie Wang, Youtu Lab, Tencent", + "bbox": [ + 142, + 180, + 880, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jarenli, ryanfwu, gleelin, guangpintao, choasliu,Jasoncjwang}@tencent.com, qnie.cuhk@gmail.com", + "bbox": [ + 86, + 218, + 934, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models, particularly those based on transformers, often employ numerous stacked structures, which possess identical architectures and perform similar functions. While effective, this stacking paradigm leads to a substantial increase in the number of parameters, posing challenges for practical applications. In today's landscape of increasingly large models, stacking depth can even reach dozens, further exacerbating this issue. To mitigate this problem, we introduce LORS (LOw-rank Residual Structure). LORS allows stacked modules to share the majority of parameters, requiring a much smaller number of unique ones per module to match or even surpass the performance of using entirely distinct ones, thereby significantly reducing parameter usage. We validate our method by applying it to the stacked decoders of a query-based object detector, and conduct extensive experiments on the widely used MS COCO dataset. Experimental results demonstrate the effectiveness of our method, as even with a $70\\%$ reduction in the parameters of the decoder, our method still enables the model to achieve comparable or even better performance than its original.", + "bbox": [ + 75, + 300, + 473, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 647, + 209, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the current era of prosperity for large models, a common issue is the significant increase in the number of parameters, which presents challenges for training, inference, and deployment. Various methods have been proposed to reduce the number of parameters in models, such as knowledge distillation [15, 20], which compresses large models into smaller ones while trying to preserve their performance but may still lead to a decrease in model capacity; pruning [16, 55], which removes redundant parameters from the model but can affect the model's stability; quantization [8], which reduces the numerical precision of model parameters to lower storage and computation but may cause model accuracy loss; and parameter sharing [24], which reduces the number of parameters by sharing them across different layers but may limit the model's expressiveness.", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a1a184e916f58e96ef2fcee1384e3911359bcd46feed7acdc3ea62fd980dba5.jpg", + "image_caption": [ + "Figure 1. The LORS calculation process, which could be adaptive or static, depending on whether an adaptively generated kernel is used in the matrix manipulation for private parameters." + ], + "image_footnote": [], + "bbox": [ + 544, + 284, + 836, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Different from the aforementioned methods, we have observed an important fact contributing to the large number of parameters: the widespread use of stacking in neural networks. Stacking refers to those modules that have identical architectures and perform the same or similar functions, but possess different parameters due to random initialization as well as training updates. Examples of stacking can be found in numerous prominent neural networks, like the classic ResNet model [18] and Transformers [46]. Particularly, Transformers heavily rely on stacked structures and typically employ completely identical multi-layer stacks in both encoders and decoders. It now serves as an indispensable component of many excellent models in fields such as computer vision and natural language processing.", + "bbox": [ + 496, + 657, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although stacking is powerful for enhancing model capacity, as demonstrated by large language models, it also", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15866", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "naturally leads to a sharp increase in the number of parameters. For example, GPT-3 [2] utilizes 175 billion parameters and consists of 96 layers of stacked Transformer [46] layers. How can we enjoy the benefits of stacking while reducing the required number of parameters? We notice that stacked decoders have identical structures and similar functions, indicating that there should be some commonality among their parameters. However, since they handle different input and output distributions, there must also be unique aspects to their parameters. Therefore, a natural idea is: it may be possible to represent the shared aspects with shared parameters while allowing each stacked module to retain only the parameters that capture its unique characteristics, thereby reducing the overall parameter usage.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above considerations, we propose to decompose the parameters of stacked modules into two parts: shared ones representing the commonality and private ones capturing the specific characteristics. Shared parameters are available for all modules and trained jointly by them, while the private parameters are separately owned by each module. We suppose that the way of sharing parameters related to the commonality may reduce the number of parameters while maintaining the performance of the model. To achieve this goal, we introduce the concept of Low-rank Residual Structure (LORS), inspired by the approach of LoRA [21]. LORS essentially adds the unique parameters to the shared ones, just like residual connections add residual information to the features. While LoRA is originally designed for fine-tuning, we train our LoRA-like operation on parameters from scratch. This approach allows us to effectively reduce the overall parameter usage while maintaining the performance of the model, paving the way for more parameter-efficient network stacking.", + "bbox": [ + 75, + 305, + 472, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate our idea, we choose AdaMixer [12], a strong query-based object detector, as our experimental subject. It contains a large number of both adaptive and static parameters in its stacked decoders, making it an ideal candidate to demonstrate LORS's effectiveness. The difference between adaptive and static parameters lies in whether they change with different inputs. Our goal is to show that LORS can effectively reduce the overall usage of both types of parameters while maintaining the model's performance. We conducted extensive experiments on this detector, illustrating that our method succeeded in reducing up to $70\\%$ of parameters in AdaMixer's decoders, while still enabling the model to achieve comparable or even superior performance compared to its vanilla version. In summary, our contributions can be concluded as:", + "bbox": [ + 75, + 595, + 472, + 821 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel low-rank residual structure, named LORS, for network stacking, which can greatly reduce the number of parameters while maintaining or even improving the performance compared to the vanilla one.", + "- We introduce effective methodologies for reducing both" + ], + "bbox": [ + 76, + 825, + 468, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "static and adaptive generated parameters in stacked structures, which makes our proposed LORS a more versatile method.", + "bbox": [ + 511, + 90, + 890, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method holds the potential to serve as one of the basic network structures for large models that are affected by the issue of excessive parameters due to stacking, rendering them more parameter-efficient, thereby facilitating an easier implementation in practical applications.", + "bbox": [ + 500, + 136, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 224, + 640, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Models with stacked structures. Many nerval networks partially or extensively employ stacked modules. CNN-based models [40], for instance, are widely applied across various computer vision tasks, such as classification [18, 22, 39, 49], detection [19, 25, 36-38, 43], segmentation [7, 13, 32, 54], etc. These models often use stacked smaller modules within larger components. Another kind of models primarily utilize Multi-Layer Perceptrons (MLPs), such as the MLP-Mixer [44], whose main body comprises dozens of completely identical stacked mixer layers. Moreover, since its invention, Transformer [46] has been successfully applied to diverse domains, including computer vision [3, 30, 41, 56], natural language processing [2, 10, 50], multi-modal learning [6, 27, 35, 47], etc. In the thriving arena of large language models, Transformer-based structures are indeed indispensable for many cutting-edge works [10, 26, 34, 42, 48]. Transformers are usually used to form stacked multi-layers encoders or decoders, within each layer being structurally identical.", + "bbox": [ + 496, + 250, + 893, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LoRA and Its Variants. LoRA [21] is a technique proposed for fine-tuning large language models, it introduces low-rank decomposition matrices in the Transformer [46] architecture, significantly reducing trainable parameters and GPU memory requirements. Afterward, a series of works with modifications as well as improvement have been proposed. AdaLoRA [53] adaptively allocates parameter budgets among weight matrices based on their importance scores and uses singular value decomposition for incremental updates, leading to an improved fine-tuning performance, especially in low budget settings. DyLoRA [45] introduces a dynamic low-rank adaptation technique that trains LoRA blocks for a range of ranks instead of a single rank, addressing the issues of fixed block size and rank optimization in LoRA. Delta-LoRA [57] presents a parameter-efficient approach to fine-tune large language models by updating not only the low-rank matrices but also the pre-trained weights using the delta of the product of two low-rank matrices. LoRA-FA [52] introduces a memory-efficient fine-tuning method that reduces activation memory without performance degradation or expensive recomputation, it achieves this by freezing the projection-down weight and updating the projection-up weight in each LoRA layer. ReLoRA [28] is a parameter-efficient method for training", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15867", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "large neural networks using low-rank updates. GLoRA [4], or Generalized LoRA, is an advanced approach for universal parameter-efficient fine-tuning tasks, which enhances the low rank adaptation technique by employing a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations. VeRA [23], or Vector-based Random Matrix Adaptation, further compresses the training parameters compared to the vanilla LoRA [21]. It achieves this by utilizing shared low-rank matrices across all layers and learning small scaling vectors.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the aforementioned LoRA series works, we propose our method called LORS (Low Rank Residual Structure) for parameter efficiency. However, our approach is fundamentally distinct from the LoRA series works. While LoRA uses different low-rank weights to fine-tune different pretrained weights, LORS adds different low-rank weights per layer to the same common weights shared by all layers, thus creating parameter-efficient stacked models trained from scratch, requiring no pretrained weights.", + "bbox": [ + 75, + 242, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 390, + 187, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we initially revisit the foundational works upon which our methodology and experiments are primarily based, encompassing the LoRA mechanism [21], query-based object detection [3], and the structure of AdaMixer's decoders [12]. Then we mathematically formulate our proposed LORS, which comprises two parts: the adaptive one and the static one. Subsequently, we elaborate on how the LORS method is applied to reduce the parameters of AdaMixer's decoders. Lastly, we provide a quantitative analysis of the parameter reduction percentage attained by our proposed method.", + "bbox": [ + 75, + 415, + 472, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 589, + 207, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mechanism of LoRA. The Low-Rank Adaptation (LoRA) [21] technique is a novel approach designed to adapt large pre-trained language models for specific tasks. The key idea of LoRA is to introduce a low-rank parameter matrix which is able to captures task-specific knowledge while the original pre-trained parameters remain fixed.", + "bbox": [ + 75, + 612, + 468, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mathematically, given a pre-trained parameter matrix $W \\in \\mathbb{R}^{d \\times h}$ , LoRA uses a low-rank matrix $B \\in \\mathbb{R}^{d \\times r}$ and a projection matrix $A \\in \\mathbb{R}^{r \\times h}$ to adapt $W$ , where $r \\ll d, h$ . The adapted parameter matrix is then given by", + "bbox": [ + 75, + 703, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW + \\Delta W = W + B A \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 771, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $BA$ captures the task-specific knowledge.", + "bbox": [ + 76, + 795, + 398, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key advantage of LoRA is that it can significantly reduce the number of parameters that need to be fine-tuned, thereby reducing the computational cost and lowering down the memory requirement. In some cases, even single-digit values of $r$ are sufficient to fine-tune the model to the desired state, which is often tens of times less expensive than", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "training the parameters in $W$ . Furthermore, by keeping the original parameters fixed, LoRA avoids catastrophic forgetting, a common issue in fine-tuning large models.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query-based object detection. In the realm of object detection, query-based detectors have established a new paradigm [11, 12, 30, 41, 51, 56]. Unlike traditional detectors which rely on anchor boxes or sliding windows, query-based models utilize a set of learnable queries to interact with image feature maps. This interaction can be formalized with attention [46] operations as", + "bbox": [ + 496, + 136, + 893, + 242 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {u p d a t e d}} = \\operatorname {A t t e n t i o n} (Q, K (V)) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 252, + 890, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q, K$ , and $V$ represent the queries, keys, and values. The learnable queries $Q$ are used to predict object classes and bounding boxes ultimately, while $K$ and $V$ typically originate from encoded image features. It is a common practice to continuously refine $Q$ through interactions with $K$ and $V$ using successive decoding layers. These layers are usually composed of structurally identical decoders.", + "bbox": [ + 496, + 277, + 890, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoders of AdaMixer. AdaMixer [12] is a query-based detector that features an innovative decoder design. This design includes Adaptive Channel Mixing (ACM) and Adaptive Spatial Mixing (ASM) methods, which greatly enhance its performance [12].", + "bbox": [ + 496, + 383, + 890, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a sampled feature $\\mathbf{x} \\in \\mathbb{R}^{P_{\\mathrm{in}} \\times C}$ , where $C = d_{\\mathrm{feat}} / g$ , and $g$ denotes the number of sampling groups. This sampled feature is obtained through an operation called group sampling. This operation divides the feature space channel $d_{\\mathrm{feat}}$ into $g$ groups and performs individual sampling for each group. Then, ACM (Adaptive Channel Mixing) utilizes a weight adapted by the object query $\\mathbf{q}$ to transform feature $\\mathbf{x}$ in the channel dimension, enhancing channel semantics [12]:", + "bbox": [ + 496, + 458, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 602, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A C M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} M _ {c}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 623, + 890, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where LayerNorm stands for Layer Normalization [1].", + "bbox": [ + 500, + 648, + 859, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next is the ASM (Adaptive Spatial Mixing) process, which aims to enable the adaptability of the object query $\\mathbf{q}$ to spatial structures of sampled features [12] by applying adaptive transformation to the spatial dimension:", + "bbox": [ + 496, + 662, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM _ {s} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P _ {\\text {i n}} \\times P _ {\\text {o u t}}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 732, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A S M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} ^ {T} M _ {s}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 753, + 890, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both ACM and ASM train independent parameters for each sampling group, and finally the output with the shape $\\mathbb{R}^{g\\times C\\times P_{\\mathrm{out}}}$ is flattened and transformed to the $d_q$ dimension by a linear layer $L_{\\mathrm{output}}$ to add back to the object query.", + "bbox": [ + 496, + 779, + 890, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ACM, ASM and the output linear transformation $L_{output}$ possess significantly more parameters compared to the decoder's other operations, making them the main contributors to the number of parameters. Hence, we choose them", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15868", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as the target components for validating the effectiveness of our LORS method in parameter reduction.", + "bbox": [ + 76, + 90, + 470, + 121 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def non_dynamic_fused_weight(W_shared, A, B):\n # W_shared: (1, d, h)\n # A: (G, r, h)\n # B: (G, d, r)\n # private weight: (G, d, h)\n W_private = B @ A\n # sum at group dimension to get fused weight: (d, h)\n W = concat([W_shared, W_private], dim=0).sum(axis=0)\n return W", + "guess_lang": "python", + "bbox": [ + 76, + 143, + 444, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2. Pseudo-code for obtaining a static weight parameter for one layer.", + "bbox": [ + 75, + 284, + 470, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Formulation of Our Method", + "text_level": 1, + "bbox": [ + 76, + 356, + 330, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The complete computational process of LORS is illustrated in Figure 1. Notably, LORS computation is divided into two types: adaptive and static. In the context of AdaMixer, the term \"adaptive\" indicates whether the transformation matrix depends on the object query. After completing the formalization of LORS here, we will further elaborate on how it is applied to the tasks in our setting.", + "bbox": [ + 75, + 380, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static Low Rank Residual Structure $(\\mathbf{LORST}^{\\mathrm{T}})$ . Suppose there are $N$ stacked layers of modules with identical architecture, and $W_{i}\\in \\mathbb{R}^{d\\times h}$ be a parameter matrix belonging to the $i$ -th layer. Then, we have:", + "bbox": [ + 75, + 488, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} = W ^ {\\text {s h a r e d}} + W _ {i} ^ {\\text {p r i v a t e}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 561, + 468, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $W^{\\mathrm{shared}} \\in \\mathbb{R}^{d \\times h}$ represents the shared parameters across all stacked layers, while $W_{i}^{\\mathrm{private}}$ denotes the layer-specific parameters for the $i$ -th layer, which is calculated as follows:", + "bbox": [ + 75, + 595, + 468, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} B _ {i k} A _ {i k} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 657, + 468, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$B_{ik} \\in \\mathbb{R}^{d \\times r}$ , $A_{ik} \\in \\mathbb{R}^{r \\times h}$ , and $r \\ll d, h$ . $K$ represents the number of parameter groups used to compute $W_{i}^{\\mathrm{private}}$ . Pseudo-code of LORST computing $W_{i}^{\\mathrm{private}}$ for a specific layer $i$ can be seen in Figure 2.", + "bbox": [ + 76, + 710, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Low Rank Residual Structure (LORSA). Let $\\hat{W}_i \\in \\mathbb{R}^{d \\times h}$ be an adaptive generated parameter in $i$ -th stacked layer, it is similarly calculated as:", + "bbox": [ + 76, + 775, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {W} _ {i} = \\hat {W} ^ {\\text {s h a r e d}} + \\hat {W} _ {i} ^ {\\text {p r i v a t e}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 834, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the cross-layer shared parameter $\\hat{W}^{\\mathrm{shared}}\\in \\mathbb{R}^{d\\times h}$ and the layer-specific parameter $\\hat{W}_i^{\\mathrm{private}}\\in \\mathbb{R}^{d\\times h}$ for each layer", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def dynamic_fused_weight(q, A, B):\n # q: (N, C)\n # A: (1, G, r, h)\n # B: (1, G, d, r)\n # shared weight: (N, 1, d, h)\n W_shared = linear1(pro_feats).reshape(N, 1, d, h)\n # E: (N, G, r, r)\n E = linear2(pro_feats).reshape(N, G, r, r)\n # private weight: (N, G, d, h)\n W_private = B @ E @ A\n # sum at group dimension to get fused weight: (N, d, h)\n W = concat([W_shared, W_private], dim=1).sum(axis=1)\n return W", + "guess_lang": "python", + "bbox": [ + 500, + 93, + 888, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3. Pseudo-code for obtaining an adaptive weight parameter for one layer.", + "bbox": [ + 498, + 277, + 890, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are both calculated based on $q$ :", + "bbox": [ + 500, + 345, + 705, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {W} ^ {\\text {s h a r e d}} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {d \\times h} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 371, + 890, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {W} _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} \\hat {B} _ {i k} \\hat {E} _ {i k} \\hat {A} _ {i k} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 393, + 890, + 435 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} _ {i k} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {r \\times r} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 438, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{B}_{ik} \\in \\mathbb{R}^{d \\times r}$ and $\\hat{A}_{ik} \\in \\mathbb{R}^{r \\times h}$ , $r \\ll d, h$ . Pseudocode of LORSA calculating $\\hat{W}_i^{\\mathrm{private}}$ for a specific layer $i$ can be seen in Figure 3.", + "bbox": [ + 498, + 469, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Applying LORS to AdaMixer's Decoders", + "text_level": 1, + "bbox": [ + 500, + 527, + 849, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply LORS to the parameters of linear transformations belonging to ACM, ASM, and $L_{\\mathrm{output}}$ in each of AdaMixer's decoders. These parameters are explained in 3.1.", + "bbox": [ + 498, + 551, + 890, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall pipeline of LORS running in AdaMixer is illustrated in Figure 4. Specifically, for each group of sampling points, $\\mathrm{LORS^A}$ (see 3.2) is used to reduce parameters in $M_{c}$ (from $\\mathbb{R}^{d_q}$ to $\\mathbb{R}^{C\\times C}$ ) of ACM and $M_{s}$ of ASM (from $\\mathbb{R}^{d_q}$ to $\\mathbb{R}^{P_{\\mathrm{in}}\\times P_{\\mathrm{out}}}$ ), and $\\mathrm{LORS^T}$ (see 3.2) to minimize parameters in $L_{\\mathrm{output}}$ (from $\\mathbb{R}^{C\\times P_{\\mathrm{out}}}$ to $\\mathbb{R}^{d_q}$ ).", + "bbox": [ + 498, + 598, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, the parameter quantities of $M_{c}$ , $M_{s}$ , and $L_{\\mathrm{output}}$ are $d_{q} \\times C \\times C$ , $d_{q} \\times P_{\\mathrm{in}} \\times P_{\\mathrm{out}}$ and $d_{q} \\times C \\times P_{\\mathrm{out}}$ respectively. When the sampling strategy consists of 2 groups with 64 points each, which is the default setting in our experiments, the values of the variables are as follows: $d_{q} = 256$ , $C = 64$ , $P_{\\mathrm{in}} = 64$ , and $P_{\\mathrm{out}} = 128$ . It can be easily calculated that the number of parameters for each of $M_{c}$ , $M_{s}$ , and $L_{\\mathrm{output}}$ exceeds one million. Indeed, these three components collectively account for the most of the total parameter in the AdaMixer model with ResNet-50 as the backbone, whereas they are also the primary drivers of enhanced model performance. This is what motivates us to conduct LORS experiments on them. The effects of LORS applied to them will be discussed in detail in 3.4.", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15869", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/184c299ee6cabe449fbc78238f2b4aaab9fdddd7a653486f7d6f51c21a6cceef.jpg", + "image_caption": [ + "Figure 4. The overall pipeline of our proposed LORS, consisting of both adaptive and static parts, each further composed of shared and private components, works collaboratively. The figure illustrates the entire computation process within one layer of the stacked layers, with an enlarged example of the $i$ -th layer for demonstration." + ], + "image_footnote": [], + "bbox": [ + 84, + 101, + 460, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Analysis on Parameter Reduction", + "text_level": 1, + "bbox": [ + 76, + 436, + 370, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $W \\in \\mathbb{R}^{d \\times h}$ be a weight parameter that exists in every layer of stacked structures, $N$ is the number of stacked layers. If static, it originally has $d \\times h$ parameters, while using LORS $^{\\mathrm{T}}$ requires $\\frac{1}{N} \\times d \\times h + K \\times (d \\times r + r \\times h)$ parameters on average per layer; if it is adaptive, generating it by $q$ with linear transformation requires $d_{q} \\times d \\times h$ parameters, where $d_{q}$ is the dimension of $q$ , and using LORS $^{\\mathrm{A}}$ requires $\\frac{1}{N} \\times d_{q} \\times d \\times h + K \\times (d_{q} \\times r^{2} + d \\times r + r \\times h)$ parameters on average per layer. To more intuitively display the parameter reduction effect of LORS, we set $d_{q} = 256$ , $d = 64$ , $h = 128$ , $K = 2$ for the ASM process and $d = 2 \\times 128 \\times 128$ , $h = 256$ , $K = 1$ for the $L_{output}$ , which is actually the case of AdaMixer's decoders in our experiment, and show the parameter reduction for different $r$ values using LORS $^{\\mathrm{T}}$ and LORS $^{\\mathrm{A}}$ in Table 1.", + "bbox": [ + 75, + 460, + 468, + 689 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bd024120ee4618a43bcf469ebd01fe2c23b3f124f9e90e0b582335cb3bf9ca33.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
r=4r=8r=16r=32
W/. LORST1.53M1.66M1.93M2.46M
W/O. LORST8.39M8.39M8.39M8.39M
Percentage18.3%19.8%23.0%29.3%
W/. LORSA0.36M0.39M0.49M0.89M
W/O. LORSA2.10M2.10M2.10M2.10M
Percentage17.2%18.6%23.3%42.4%
", + "bbox": [ + 91, + 704, + 454, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Analysis of the parameter reduction effect of LORS $^{\\mathrm{T}}$ and LORS $^{\\mathrm{A}}$ with varying rank $r$ values. Our default selections are colored gray. Here the average amount of shared parameters over each layer has already been taken into account.", + "bbox": [ + 75, + 840, + 468, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 89, + 632, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first explain the implementation details of our experiments, including the dataset, training specifics, and different parameter initialization methods for static and adaptive LORS. Next, we present the main experimental results. Finally, we perform ablation studies to examine LORS's design and its individual components.", + "bbox": [ + 496, + 114, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 218, + 718, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Our experiments were conducted on the widely-used MS COCO [29] dataset in mmdetection codebase [5], using the standard metrics for object detection. All models were trained on the train2017 split ( $\\sim 118\\mathrm{k}$ images) and then evaluated on the val2017 split (5k images).", + "bbox": [ + 496, + 242, + 890, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We use the AdamW optimizer [33] with a weight decay of 0.0001 and train all models on 8 Nvidia V100 GPUs with a batch size of 16 and a learning rate of $2.5 \\times e^{-5}$ . Models are trained for either 12 or 36 epochs, with learning rate decreasing by a factor of 10 at epochs 8 and 11 for 12-epoch training, and at epochs 24 and 33 for 36-epoch training. The low-rank values are set as $r = 16$ for LORS $^{\\text{A}}$ and $r = 8$ for LORS $^{\\text{T}}$ . The number of parameter groups is set as as $K = [1,1,2,2,3,3]$ for LORS $^{\\text{A}}$ in all experiments, applied to ACM and ASM in AdaMixer's decoders, and $K = [1,1,1,1,1,1]$ for $L_{output}$ in all experiments. We divide the feature channels into 2 groups with 64 sampling points each instead of AdaMixer's default 4 groups with 32 sampling points each, aiming to increase the parameter compressible space for LORS, which does not improve performance according to both AdaMixer paper [12] and our experiments. Backbones are initialized with ImageNet-1k [9] pre-trained models, and LORS-related parameter initialization is detailed below, remaining parameters are Xavier-initialized [14]. Finally, all other aspects about model training, like the data augmentation, the loss function, etc., just follow AdaMixer's settings [12].", + "bbox": [ + 496, + 319, + 892, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization Strategies We tried various initialization methods for each component in LORS and determined the overall initialization method as follows:", + "bbox": [ + 496, + 652, + 890, + 696 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- LORST: For static LORS, we employ Kaiming initialization [17] for $W^{\\text{shared}}$ and each $B$ , and zero initialization for each $A$ .", + "- LORSA: For adaptive LORS, we apply Kaiming initialization [17] to the linear transformation weights forming each $\\hat{W}^{\\mathrm{shared}}$ , as well as each $\\hat{B}$ and $\\hat{A}$ . Additionally, we use zero initialization for the linear transformation weights forming each $\\hat{E}$ ." + ], + "bbox": [ + 513, + 696, + 888, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 500, + 830, + 638, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To reliably demonstrate that LORS can reduce the number of parameters in stacked structures while maintaining model performance, which is measured by AP metric, we follow", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15870", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c9cc7f34f32d12b518fed23534c42d900b31db7fc57f9580cadc37ad748caa04.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodQueriesEpochsDecoder Params(M)Params(M)Decoder GFLOPsGFLOPsTraining HoursAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixer10012110135121048.5h42.761.545.924.745.459.2
AdaMixer + LORS100123560181109.5h42.661.446.025.045.658.5
AdaMixer300121131393913210h44.163.447.427.046.959.5
AdaMixer + LORS3001235605614912h44.163.047.727.847.059.5
", + "bbox": [ + 81, + 88, + 890, + 178 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f33e3ad7f24700907c448ba121c9a67b960186f03567a5acf1ddec3a9265d9d4.jpg", + "table_caption": [ + "Table 2. $1 \\times$ training scheme performance on COCO 2017 val set with ResNet-50 as backbone. AdaMixer with LORS can achieve competitive results while employing a notably reduced number of parameters. FPS is obtained with a single Nvidia V100 GPU." + ], + "table_footnote": [], + "table_body": "
MethodbackboneQueriesEpochsParams(M)GFLOPsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixerR-501003613510443.261.846.725.046.158.8
AdaMixer + LORSR-50100366011043.762.347.325.546.460.0
AdaMixerR-503003613913247.066.051.130.150.261.8
AdaMixer + LORSR-50300366014947.666.652.031.150.262.5
AdaMixerR-1013003615820848.067.052.430.051.263.7
AdaMixer + LORSR-101300367922548.267.552.631.751.363.8
AdaMixerSwin-S3003616423451.371.255.734.254.667.3
AdaMixer + LORSSwin-S300368525051.871.656.435.455.068.4
", + "bbox": [ + 81, + 232, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. $3 \\times$ training scheme performance on COCO 2017 val set, considering different combinations of backbone and query numbers. Longer training time allows LORS to be fully trained and perform better. The comprehensive improvement in performance and the significant reduction in parameters demonstrate the effectiveness of LORS. The best results in each category are highlighted in bold.", + "bbox": [ + 75, + 402, + 892, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "common practices in the object detection field and present experimental comparisons in two categories based on training schemes. The first involves training for 12 epochs using a weaker data augmentation, and the second involves training for 36 epochs using a stronger data augmentation, both of which are identical to those in AdaMixer [12].", + "bbox": [ + 75, + 472, + 468, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents the comparison of performance with and without the LORS technique, under a $1\\times$ training scheme. Metrics such as the number of parameters, GFLOPs, and Average Precision (AP) at various scales are evaluated. When the AdaMixer model is combined with LORS, it demonstrates a remarkable reduction in the number of parameters. This reduction is consistently observed across various query quantities and different training epochs, encompassing both the decoder and overall model. For example, when trained with 100 queries for 12 epochs, the AdaMixer+LORS model uses only 35M decoder parameters and 60M total parameters, compared to 110M and 135M respectively in the AdaMixer model. This shows reduction of approximately $70\\%$ in the decoder's parameters, which is substantial. However, the AdaMixer+LORS model shows a slight increase in GFLOPs. In terms of model performance, the AdaMixer+LORS model achieves competitive results, even with a significantly reduced number of parameters. When trained with 100 queries, although the AP of 42.6 is slightly lower than the 42.7 AP of the vanilla AdaMixer, it still slightly outperforms in $AP_{s}$ and $AP_{m}$ . When trained with 300 queries, the use of LORS does not", + "bbox": [ + 75, + 568, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "affect the performance and allows an advantage in all $AP_{s}$ , $AP_{m}$ and $AP_{l}$ . These findings suggest that even under conditions of limited model training and data augmentation, LORS does not impact training result, the model can still converge quickly to its original limit. Additionally, we also demonstrate the slight impact of the current LORS method on inference speed, which might be improved by reducing the serial and redundant computations within LORS but not the focus of this paper.", + "bbox": [ + 496, + 472, + 892, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 showcases the remarkable performance of the AdaMixer + LORS method under the $3 \\times$ training scheme with different backbones and query numbers. It can be observed that the proposed method consistently outperforms the vanilla AdaMixer across all backbones, query numbers, and evaluation metrics. This result is somewhat surprising to us, as the LORS enables the model to use significantly fewer parameters during both training and inference.", + "bbox": [ + 496, + 611, + 893, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, when adopting the ResNet-50 backbone and 100 queries, LORS improves the AP value of AdaMixer by 0.5 points (43.7 vs 43.2). It is worth noting that in Table 2, for experiments with the same configuration except for training 12 epochs, the AP with LORS is 0.1 lower than without LORS. However, this situation is reversed when training for 36 epochs. The longer training time under the $3 \\times$ training scheme allows LORS to be fully trained and perform better. This emphasizes the importance of sufficient training time for the LORS technique to fully exploit its potential in enhancing the performance. Furthermore,", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15871", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when using more powerful backbones and a larger number of queries, the performance metrics of models utilizing LORS consistently and comprehensively surpass those of their counterparts without LORS. Even when employing a strong backbone such as Swin-S [31] and using 300 queries, LORS still manages to further improve the model's performance while reducing nearly half of the overall parameters. Much fewer parameters, yet improved performance.", + "bbox": [ + 75, + 90, + 470, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In conclusion, the results in Table 2 and Table 3 show the effectiveness of the LORS method in reducing the number of parameters while achieving even better performance. The method's adaptability to various backbones and query numbers further demonstrates its versatility and potential for broader applications. As for why LORS can achieve better performance with fewer parameters, a plausible explanation is that after the model has been sufficiently trained and stabilized, the parameter matrix within the stacked structure is inherently sparse and low-rank. Without using LORS, the model needs to learn such a low-rank structure through training. In contrast, LORS directly and explicitly makes the structure sparse, which, to some extent, can be considered as adding a form of regularization to the model training, thereby facilitating the model's learning.", + "bbox": [ + 75, + 212, + 473, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 450, + 230, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To save computational resources, all ablation studies are conducted using a ResNet-50 backbone and a $1 \\times$ training scheme.", + "bbox": [ + 75, + 474, + 470, + 518 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e59e6dde58587edf92afeb52a50a330b5b957afb99f8ec0e6bfd84dd80b7580e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LORSALORSTDecoder Params(M)AP
11042.5†
✓7942.6
✓6642.6
✓✓3542.6
", + "bbox": [ + 147, + 531, + 401, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive & Static LORS. We first conduct ablation studies on the impact of $\\mathrm{LORSA}$ and $\\mathrm{LORST}$ on model parameters and performance, as shown in Table 4. In this table, a $\\checkmark$ indicates that we use the LORS method to replace the corresponding component in the AdaMixer model mentioned earlier, while a blank cell means we use the original AdaMixer module. If neither of the two is present, the model is identical to the original AdaMixer. As can be seen from the table, both adaptive and static LORS can effectively reduce the parameter size of the decoder in the model (31M and 44M, respectively) without affecting the model performance. Note that the symbol $\\dagger$ here represents that the 42.5 AP value is reproduced by us, which is slightly dif", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferent from the 42.7 AP recorded in the original paper. We believe this discrepancy can be attributed to the randomness of 1x training and does not affect our conclusions.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shared and private weights. The second point of interest is to determine which of the shared weight and private weight has a greater impact on the decoder's performance. Experiment in Table 5 is designed to address this question, which is performed on ACM and ASM of AdaMixer's decoder. Recall that the formula we use to generate the final weight parameters is $W = W^{\\text{shared}} + W^{\\text{private}}$ . In the table, a $\\checkmark$ indicates that the corresponding term is present on the right side of the formula, otherwise, it is removed. It can be seen that the absence of either term will cause a nonnegligible decrease in model performance, and the impact of missing $W^{\\text{private}}$ is even more severe. This suggests that although the number of parameters in each $W^{\\text{private}}$ is smaller than that of $W^{\\text{shared}}$ , it could be more important, possibly because the most informative part of each layer in the stacked structure is contained within it.", + "bbox": [ + 496, + 137, + 892, + 378 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ddaf7610aa9336fd39cf0fe4ba2d5b06f114d6ec1f2e5c4632a18b3d9c13f99.jpg", + "table_caption": [ + "Table 4. Effect of LORSA and LORST. †” denotes this AP result was reproduced by ourselves. Both LORST and LORSA can reduce parameters without compromising performance." + ], + "table_footnote": [], + "table_body": "
WsharedWprivateDecoder Params(M)AP
✓5840.9
✓4341.4
✓✓3542.6
", + "bbox": [ + 570, + 388, + 823, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Effect of shared parameters $W^{\\text{shared}}$ and private parameters $W^{\\text{private}}$ Default choice for our model is colored gray.", + "bbox": [ + 498, + 482, + 890, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters of $\\mathbf{LORS}^A$ . Another aspect we would like to explore is the optimal number of parameter groups and the value of rank $r$ for adaptive LORS. Here, we feel it is necessary to explain why we use grouped parameters. Recalling formula 11, the adaptive private parameters are obtained by continuous multiplication of three matrices, so the matrix with the smallest rank among them determines the rank of multiplication, which is the upper limit for the private parameters. If the task-specific requirement for the rank of the private parameters matrix unfortunately exceeds the rank of the smallest component, there will be a bottleneck when training the model. Therefore, we designed this parameter group mechanism, using the sum of the results of multiple matrix multiplication to break the aforementioned problem. Another advantage is that the rank of the final matrix possibly increases linearly with the number of parameters, while simply using a larger $r$ would lead to a quadratic increase in the number of parameters. In the end, looking at Table 6, we observed that the experimental result for the 6-layer decoder of the AdaMixer object detector is relatively better when using [1,1,2,2,3,3] groups of parameters per layer and $r = 16$ for each group. This slightly strange parameter group numbers is inspired by AdaLoRA [53]. It suggests that higher-level neural network parameters may", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15872", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "need to allocate more ranks when performing LoRA fin-tuning, which we suppose may also hold true for LORS, and the experiment results confirmed that.", + "bbox": [ + 78, + 90, + 467, + 135 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b99035346ec685d9d40abd3b2294276f83a36fb1380cbc8b980a078cddf3ac41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 183241.8
1 1 1 1 1 1163442.1
2 2 2 2 2 2163542.4
1 1 2 2 3 3163542.6
1 1 1 1 1 1323841.2
", + "bbox": [ + 142, + 147, + 403, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters of $\\mathbf{LORS}^T$ . We further investigate the optimal configuration settings for $\\mathbf{LORS}^T$ to achieve satisfactory performance, as shown in Table 7. We observed that the best option is to use one group of low-rank parameters with $r = 8$ for the decomposition of each layer $L_{\\mathrm{output}}$ in the stacked decoder of AdaMixer, which can achieve a model performance comparable to the original detector. Since using one group of parameters already yields reasonably good results, we did not explore the use of more parameter groups. Moreover, we should also note that the value of rank $r$ is not necessarily better when larger or smaller, which may indicate that the low-rank structure required by the model for a specific task has its own particularity.", + "bbox": [ + 78, + 340, + 467, + 539 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb4cfef9c0ee045c7d76c7897f30a50eff0966e67e0b4764389ed3ea70bd7c5a.jpg", + "table_caption": [ + "Table 6. Validation accuracy with different groups of parameters and rank $r$ in LORSA. Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + ], + "table_footnote": [], + "table_body": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 123442.3
1 1 1 1 1 143441.9
1 1 1 1 1 183542.6
1 1 1 1 1 1163742.1
1 1 1 1 1 1324041.7
", + "bbox": [ + 138, + 551, + 406, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of decoders. Finally, we conclude our ablation studies with experiments on the optimal number of decoder layers incorporating the LORS structure for better performance. Since our method significantly reduces the total number of model parameters and allows each decoder layer to train and use shared parameters that capture common features, it is natural to consider whether the model's inherent characteristics are affected and whether using more or fewer decoding layers would further improve the model's performance. With this idea in mind, we conducted sev", + "bbox": [ + 78, + 750, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eral experiments as shown in Table 8 to explore this issue. However, similar to the original AdaMixer without LORS, 6 layers still yield the best performance for the AdaMixer with LORS, having too many or too few layers would degrade the model's performance. This may also suggest that LORS merely captures the intrinsic structure of the model without altering its original properties.", + "bbox": [ + 501, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df9a31d9e8e3b1feb4a49f53c960a5abe2d318884346f447292312bd5b4c494f.jpg", + "table_caption": [ + "Table 7. Validation accuracy with different rank $r$ in LORS $^{\\mathrm{T}}$ . Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + ], + "table_footnote": [], + "table_body": "
Decoder\nNumberDecoder\nParams(M)AP
32838.9
63542.6
94242.0
125041.0
", + "bbox": [ + 596, + 215, + 795, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Model performance using different numbers of stacked decoders when employing LORS. The default choice for our model is colored gray. This experiment with varying numbers of decoders demonstrates that six layers still yield the best performance for the model, which is not changed by LORS.", + "bbox": [ + 501, + 324, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 503, + 445, + 616, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research introduced a novel approach to the reduction of parameters within deep learning models utilizing stacked structures, i.e., the Low-rank Residual Structure (LORS). It is potentially an effective methodology in reducing both static and adaptive parameters. Broadly speaking, LORS allows the parameters of stacked modules to be largely shared, while maintaining only a small number of unique parameters for each module, thereby significantly reducing the total number of parameters whilst not compromising performance. We validated our method via object detection task across various extensive experimental cases on the widely-used COCO dataset, and the results demonstrate a superior model performance even with a $50\\% - 70\\%$ reduction in decoder parameters, surpassing our expectations.", + "bbox": [ + 501, + 473, + 890, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "There are some noticeable limitations in our study. While effective in reducing model parameters without compromising performance, enhancing model capabilities requires a comparatively long training process for the method. Additionally, our method slightly increases the inference time, as even without a detailed analysis, it can be roughly seen that LORS has unparalleled and repetitive computations needing more optimization. Finally, we have only tested our approach on object detection tasks, specifically with the AdaMixer model and its decoder structure. It's clear that our method can be applied to more tasks (e.g., NLP), different models (e.g., language models), and other neural network components (e.g., backbones). These aspects will be the focus of our future work.", + "bbox": [ + 501, + 688, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15873", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3", + "[2] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 2, 3", + "[4] Arnav Chavan, Zhuang Liu, Deepak Gupta, Eric Xing, and Zhiqiang Shen. One-for-all: Generalized lora for parameter-efficient fine-tuning, 2023. 3", + "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5", + "[6] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2", + "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2", + "[8] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems, 28, 2015. 1", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[11] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3621–3630, 2021. 3", + "[12] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 3, 5, 6", + "[13] Yanqi Ge, Qiang Nie, Ye Huang, Yong Liu, Chengjie Wang, Feng Zheng, Wen Li, and Lixin Duan. Beyond proto" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "types: Semantic anchor regularization for better representation learning. arXiv preprint arXiv:2312.11872, 2023. 2", + "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010. 5", + "[15] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 1", + "[16] Song Han, Jeff Pool, John Tran, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 1", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2", + "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3", + "[22] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 2", + "[23] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation, 2023. 3", + "[24] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 1", + "[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2", + "[26] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15874", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Vladislav Lialin, Namrata Shivagunde, Sherin Muckatira, and Anna Rumshisky. Stack more layers differently: High-rank training through low-rank updates, 2023. 2", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3", + "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7", + "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[36] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 2", + "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016.", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2", + "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[40] Farhana Sultana, Abu Sufian, and Paramartha Dutta. A review of object detection models based on convolutional neural network. Intelligent computing: image processing based applications, pages 1-16, 2020. 2", + "[41] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2", + "[43] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2", + "[44] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261-24272, 2021. 2", + "[45] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter efficient tuning of pretrained models using dynamic search-free low-rank adaptation, 2023. 2", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3", + "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2", + "[48] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2", + "[49] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 2", + "[50] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 2", + "[51] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3", + "[52] Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning, 2023. 2", + "[53] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning, 2023. 2, 7", + "[54] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 1", + "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3", + "[57] Bojia Zi, Xianbiao Qi, Lingzhi Wang, Jianan Wang, KamFai Wong, and Lei Zhang. Delta-lora: Fine-tuning high-rank parameters with the delta of low-rank matrices, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15876", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_model.json b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dcf398a502bc809d38b7411c763854ca909631c8 --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.887, + 0.154 + ], + "angle": 0, + "content": "LORS: Low-rank Residual Structure for Parameter-Efficient Network Stacking" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.181, + 0.882, + 0.216 + ], + "angle": 0, + "content": "Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, Chengjie Wang, Youtu Lab, Tencent" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.219, + 0.936, + 0.235 + ], + "angle": 0, + "content": "{jarenli, ryanfwu, gleelin, guangpintao, choasliu,Jasoncjwang}@tencent.com, qnie.cuhk@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.619 + ], + "angle": 0, + "content": "Deep learning models, particularly those based on transformers, often employ numerous stacked structures, which possess identical architectures and perform similar functions. While effective, this stacking paradigm leads to a substantial increase in the number of parameters, posing challenges for practical applications. In today's landscape of increasingly large models, stacking depth can even reach dozens, further exacerbating this issue. To mitigate this problem, we introduce LORS (LOw-rank Residual Structure). LORS allows stacked modules to share the majority of parameters, requiring a much smaller number of unique ones per module to match or even surpass the performance of using entirely distinct ones, thereby significantly reducing parameter usage. We validate our method by applying it to the stacked decoders of a query-based object detector, and conduct extensive experiments on the widely used MS COCO dataset. Experimental results demonstrate the effectiveness of our method, as even with a \\(70\\%\\) reduction in the parameters of the decoder, our method still enables the model to achieve comparable or even better performance than its original." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.648, + 0.21, + 0.664 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In the current era of prosperity for large models, a common issue is the significant increase in the number of parameters, which presents challenges for training, inference, and deployment. Various methods have been proposed to reduce the number of parameters in models, such as knowledge distillation [15, 20], which compresses large models into smaller ones while trying to preserve their performance but may still lead to a decrease in model capacity; pruning [16, 55], which removes redundant parameters from the model but can affect the model's stability; quantization [8], which reduces the numerical precision of model parameters to lower storage and computation but may cause model accuracy loss; and parameter sharing [24], which reduces the number of parameters by sharing them across different layers but may limit the model's expressiveness." + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.285, + 0.837, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.587, + 0.895, + 0.631 + ], + "angle": 0, + "content": "Figure 1. The LORS calculation process, which could be adaptive or static, depending on whether an adaptively generated kernel is used in the matrix manipulation for private parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Different from the aforementioned methods, we have observed an important fact contributing to the large number of parameters: the widespread use of stacking in neural networks. Stacking refers to those modules that have identical architectures and perform the same or similar functions, but possess different parameters due to random initialization as well as training updates. Examples of stacking can be found in numerous prominent neural networks, like the classic ResNet model [18] and Transformers [46]. Particularly, Transformers heavily rely on stacked structures and typically employ completely identical multi-layer stacks in both encoders and decoders. It now serves as an indispensable component of many excellent models in fields such as computer vision and natural language processing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Although stacking is powerful for enhancing model capacity, as demonstrated by large language models, it also" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "naturally leads to a sharp increase in the number of parameters. For example, GPT-3 [2] utilizes 175 billion parameters and consists of 96 layers of stacked Transformer [46] layers. How can we enjoy the benefits of stacking while reducing the required number of parameters? We notice that stacked decoders have identical structures and similar functions, indicating that there should be some commonality among their parameters. However, since they handle different input and output distributions, there must also be unique aspects to their parameters. Therefore, a natural idea is: it may be possible to represent the shared aspects with shared parameters while allowing each stacked module to retain only the parameters that capture its unique characteristics, thereby reducing the overall parameter usage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.593 + ], + "angle": 0, + "content": "Based on the above considerations, we propose to decompose the parameters of stacked modules into two parts: shared ones representing the commonality and private ones capturing the specific characteristics. Shared parameters are available for all modules and trained jointly by them, while the private parameters are separately owned by each module. We suppose that the way of sharing parameters related to the commonality may reduce the number of parameters while maintaining the performance of the model. To achieve this goal, we introduce the concept of Low-rank Residual Structure (LORS), inspired by the approach of LoRA [21]. LORS essentially adds the unique parameters to the shared ones, just like residual connections add residual information to the features. While LoRA is originally designed for fine-tuning, we train our LoRA-like operation on parameters from scratch. This approach allows us to effectively reduce the overall parameter usage while maintaining the performance of the model, paving the way for more parameter-efficient network stacking." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.473, + 0.822 + ], + "angle": 0, + "content": "To validate our idea, we choose AdaMixer [12], a strong query-based object detector, as our experimental subject. It contains a large number of both adaptive and static parameters in its stacked decoders, making it an ideal candidate to demonstrate LORS's effectiveness. The difference between adaptive and static parameters lies in whether they change with different inputs. Our goal is to show that LORS can effectively reduce the overall usage of both types of parameters while maintaining the model's performance. We conducted extensive experiments on this detector, illustrating that our method succeeded in reducing up to \\(70\\%\\) of parameters in AdaMixer's decoders, while still enabling the model to achieve comparable or even superior performance compared to its vanilla version. In summary, our contributions can be concluded as:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.884 + ], + "angle": 0, + "content": "- We propose a novel low-rank residual structure, named LORS, for network stacking, which can greatly reduce the number of parameters while maintaining or even improving the performance compared to the vanilla one." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "- We introduce effective methodologies for reducing both" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "static and adaptive generated parameters in stacked structures, which makes our proposed LORS a more versatile method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.894, + 0.213 + ], + "angle": 0, + "content": "- Our method holds the potential to serve as one of the basic network structures for large models that are affected by the issue of excessive parameters due to stacking, rendering them more parameter-efficient, thereby facilitating an easier implementation in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.226, + 0.642, + 0.242 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.538 + ], + "angle": 0, + "content": "Models with stacked structures. Many nerval networks partially or extensively employ stacked modules. CNN-based models [40], for instance, are widely applied across various computer vision tasks, such as classification [18, 22, 39, 49], detection [19, 25, 36-38, 43], segmentation [7, 13, 32, 54], etc. These models often use stacked smaller modules within larger components. Another kind of models primarily utilize Multi-Layer Perceptrons (MLPs), such as the MLP-Mixer [44], whose main body comprises dozens of completely identical stacked mixer layers. Moreover, since its invention, Transformer [46] has been successfully applied to diverse domains, including computer vision [3, 30, 41, 56], natural language processing [2, 10, 50], multi-modal learning [6, 27, 35, 47], etc. In the thriving arena of large language models, Transformer-based structures are indeed indispensable for many cutting-edge works [10, 26, 34, 42, 48]. Transformers are usually used to form stacked multi-layers encoders or decoders, within each layer being structurally identical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "LoRA and Its Variants. LoRA [21] is a technique proposed for fine-tuning large language models, it introduces low-rank decomposition matrices in the Transformer [46] architecture, significantly reducing trainable parameters and GPU memory requirements. Afterward, a series of works with modifications as well as improvement have been proposed. AdaLoRA [53] adaptively allocates parameter budgets among weight matrices based on their importance scores and uses singular value decomposition for incremental updates, leading to an improved fine-tuning performance, especially in low budget settings. DyLoRA [45] introduces a dynamic low-rank adaptation technique that trains LoRA blocks for a range of ranks instead of a single rank, addressing the issues of fixed block size and rank optimization in LoRA. Delta-LoRA [57] presents a parameter-efficient approach to fine-tune large language models by updating not only the low-rank matrices but also the pre-trained weights using the delta of the product of two low-rank matrices. LoRA-FA [52] introduces a memory-efficient fine-tuning method that reduces activation memory without performance degradation or expensive recomputation, it achieves this by freezing the projection-down weight and updating the projection-up weight in each LoRA layer. ReLoRA [28] is a parameter-efficient method for training" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15867" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "large neural networks using low-rank updates. GLoRA [4], or Generalized LoRA, is an advanced approach for universal parameter-efficient fine-tuning tasks, which enhances the low rank adaptation technique by employing a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations. VeRA [23], or Vector-based Random Matrix Adaptation, further compresses the training parameters compared to the vanilla LoRA [21]. It achieves this by utilizing shared low-rank matrices across all layers and learning small scaling vectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.381 + ], + "angle": 0, + "content": "Inspired by the aforementioned LoRA series works, we propose our method called LORS (Low Rank Residual Structure) for parameter efficiency. However, our approach is fundamentally distinct from the LoRA series works. While LoRA uses different low-rank weights to fine-tune different pretrained weights, LORS adds different low-rank weights per layer to the same common weights shared by all layers, thus creating parameter-efficient stacked models trained from scratch, requiring no pretrained weights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.391, + 0.188, + 0.408 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.473, + 0.584 + ], + "angle": 0, + "content": "In this section, we initially revisit the foundational works upon which our methodology and experiments are primarily based, encompassing the LoRA mechanism [21], query-based object detection [3], and the structure of AdaMixer's decoders [12]. Then we mathematically formulate our proposed LORS, which comprises two parts: the adaptive one and the static one. Subsequently, we elaborate on how the LORS method is applied to reduce the parameters of AdaMixer's decoders. Lastly, we provide a quantitative analysis of the parameter reduction percentage attained by our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.59, + 0.208, + 0.606 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.703 + ], + "angle": 0, + "content": "The mechanism of LoRA. The Low-Rank Adaptation (LoRA) [21] technique is a novel approach designed to adapt large pre-trained language models for specific tasks. The key idea of LoRA is to introduce a low-rank parameter matrix which is able to captures task-specific knowledge while the original pre-trained parameters remain fixed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Mathematically, given a pre-trained parameter matrix \\( W \\in \\mathbb{R}^{d \\times h} \\), LoRA uses a low-rank matrix \\( B \\in \\mathbb{R}^{d \\times r} \\) and a projection matrix \\( A \\in \\mathbb{R}^{r \\times h} \\) to adapt \\( W \\), where \\( r \\ll d, h \\). The adapted parameter matrix is then given by" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.772, + 0.47, + 0.788 + ], + "angle": 0, + "content": "\\[\nW + \\Delta W = W + B A \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.796, + 0.4, + 0.81 + ], + "angle": 0, + "content": "where \\( BA \\) captures the task-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The key advantage of LoRA is that it can significantly reduce the number of parameters that need to be fine-tuned, thereby reducing the computational cost and lowering down the memory requirement. In some cases, even single-digit values of \\( r \\) are sufficient to fine-tune the model to the desired state, which is often tens of times less expensive than" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "training the parameters in \\( W \\). Furthermore, by keeping the original parameters fixed, LoRA avoids catastrophic forgetting, a common issue in fine-tuning large models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.894, + 0.243 + ], + "angle": 0, + "content": "Query-based object detection. In the realm of object detection, query-based detectors have established a new paradigm [11, 12, 30, 41, 51, 56]. Unlike traditional detectors which rely on anchor boxes or sliding windows, query-based models utilize a set of learnable queries to interact with image feature maps. This interaction can be formalized with attention [46] operations as" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.253, + 0.892, + 0.269 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {u p d a t e d}} = \\operatorname {A t t e n t i o n} (Q, K (V)) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.278, + 0.892, + 0.383 + ], + "angle": 0, + "content": "where \\( Q, K \\), and \\( V \\) represent the queries, keys, and values. The learnable queries \\( Q \\) are used to predict object classes and bounding boxes ultimately, while \\( K \\) and \\( V \\) typically originate from encoded image features. It is a common practice to continuously refine \\( Q \\) through interactions with \\( K \\) and \\( V \\) using successive decoding layers. These layers are usually composed of structurally identical decoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Decoders of AdaMixer. AdaMixer [12] is a query-based detector that features an innovative decoder design. This design includes Adaptive Channel Mixing (ACM) and Adaptive Spatial Mixing (ASM) methods, which greatly enhance its performance [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.595 + ], + "angle": 0, + "content": "Given a sampled feature \\(\\mathbf{x} \\in \\mathbb{R}^{P_{\\mathrm{in}} \\times C}\\), where \\(C = d_{\\mathrm{feat}} / g\\), and \\(g\\) denotes the number of sampling groups. This sampled feature is obtained through an operation called group sampling. This operation divides the feature space channel \\(d_{\\mathrm{feat}}\\) into \\(g\\) groups and performs individual sampling for each group. Then, ACM (Adaptive Channel Mixing) utilizes a weight adapted by the object query \\(\\mathbf{q}\\) to transform feature \\(\\mathbf{x}\\) in the channel dimension, enhancing channel semantics [12]:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.603, + 0.891, + 0.621 + ], + "angle": 0, + "content": "\\[\nM _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.624, + 0.892, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A C M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} M _ {c}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.649, + 0.861, + 0.664 + ], + "angle": 0, + "content": "where LayerNorm stands for Layer Normalization [1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.664, + 0.892, + 0.725 + ], + "angle": 0, + "content": "Next is the ASM (Adaptive Spatial Mixing) process, which aims to enable the adaptability of the object query \\(\\mathbf{q}\\) to spatial structures of sampled features [12] by applying adaptive transformation to the spatial dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.733, + 0.891, + 0.75 + ], + "angle": 0, + "content": "\\[\nM _ {s} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P _ {\\text {i n}} \\times P _ {\\text {o u t}}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.754, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A S M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} ^ {T} M _ {s}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Both ACM and ASM train independent parameters for each sampling group, and finally the output with the shape \\(\\mathbb{R}^{g\\times C\\times P_{\\mathrm{out}}}\\) is flattened and transformed to the \\(d_q\\) dimension by a linear layer \\(L_{\\mathrm{output}}\\) to add back to the object query." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ACM, ASM and the output linear transformation \\( L_{output} \\) possess significantly more parameters compared to the decoder's other operations, making them the main contributors to the number of parameters. Hence, we choose them" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15868" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.122 + ], + "angle": 0, + "content": "as the target components for validating the effectiveness of our LORS method in parameter reduction." + }, + { + "type": "code", + "bbox": [ + 0.078, + 0.144, + 0.445, + 0.278 + ], + "angle": 0, + "content": "def non_dynamic_fused_weight(W_shared, A, B):\n # W_shared: (1, d, h)\n # A: (G, r, h)\n # B: (G, d, r)\n # private weight: (G, d, h)\n W_private = B @ A\n # sum at group dimension to get fused weight: (d, h)\n W = concat([W_shared, W_private], dim=0).sum(axis=0)\n return W" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.471, + 0.315 + ], + "angle": 0, + "content": "Figure 2. Pseudo-code for obtaining a static weight parameter for one layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.357, + 0.331, + 0.373 + ], + "angle": 0, + "content": "3.2. Formulation of Our Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.488 + ], + "angle": 0, + "content": "The complete computational process of LORS is illustrated in Figure 1. Notably, LORS computation is divided into two types: adaptive and static. In the context of AdaMixer, the term \"adaptive\" indicates whether the transformation matrix depends on the object query. After completing the formalization of LORS here, we will further elaborate on how it is applied to the tasks in our setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.47, + 0.551 + ], + "angle": 0, + "content": "Static Low Rank Residual Structure \\((\\mathbf{LORST}^{\\mathrm{T}})\\). Suppose there are \\(N\\) stacked layers of modules with identical architecture, and \\(W_{i}\\in \\mathbb{R}^{d\\times h}\\) be a parameter matrix belonging to the \\(i\\)-th layer. Then, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.563, + 0.47, + 0.583 + ], + "angle": 0, + "content": "\\[\nW _ {i} = W ^ {\\text {s h a r e d}} + W _ {i} ^ {\\text {p r i v a t e}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.656 + ], + "angle": 0, + "content": "Here, \\( W^{\\mathrm{shared}} \\in \\mathbb{R}^{d \\times h} \\) represents the shared parameters across all stacked layers, while \\( W_{i}^{\\mathrm{private}} \\) denotes the layer-specific parameters for the \\( i \\)-th layer, which is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.659, + 0.47, + 0.7 + ], + "angle": 0, + "content": "\\[\nW _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} B _ {i k} A _ {i k} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.712, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\(B_{ik} \\in \\mathbb{R}^{d \\times r}\\), \\(A_{ik} \\in \\mathbb{R}^{r \\times h}\\), and \\(r \\ll d, h\\). \\(K\\) represents the number of parameter groups used to compute \\(W_{i}^{\\mathrm{private}}\\). Pseudo-code of LORST computing \\(W_{i}^{\\mathrm{private}}\\) for a specific layer \\(i\\) can be seen in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.776, + 0.47, + 0.823 + ], + "angle": 0, + "content": "Adaptive Low Rank Residual Structure (LORSA). Let \\(\\hat{W}_i \\in \\mathbb{R}^{d \\times h}\\) be an adaptive generated parameter in \\(i\\)-th stacked layer, it is similarly calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.835, + 0.469, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\hat {W} _ {i} = \\hat {W} ^ {\\text {s h a r e d}} + \\hat {W} _ {i} ^ {\\text {p r i v a t e}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where the cross-layer shared parameter \\(\\hat{W}^{\\mathrm{shared}}\\in \\mathbb{R}^{d\\times h}\\) and the layer-specific parameter \\(\\hat{W}_i^{\\mathrm{private}}\\in \\mathbb{R}^{d\\times h}\\) for each layer" + }, + { + "type": "code", + "bbox": [ + 0.501, + 0.094, + 0.89, + 0.272 + ], + "angle": 0, + "content": "def dynamic_fused_weight(q, A, B):\n # q: (N, C)\n # A: (1, G, r, h)\n # B: (1, G, d, r)\n # shared weight: (N, 1, d, h)\n W_shared = linear1(pro_feats).reshape(N, 1, d, h)\n # E: (N, G, r, r)\n E = linear2(pro_feats).reshape(N, G, r, r)\n # private weight: (N, G, d, h)\n W_private = B @ E @ A\n # sum at group dimension to get fused weight: (N, d, h)\n W = concat([W_shared, W_private], dim=1).sum(axis=1)\n return W" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.279, + 0.892, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Pseudo-code for obtaining an adaptive weight parameter for one layer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.346, + 0.706, + 0.361 + ], + "angle": 0, + "content": "are both calculated based on \\(q\\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.372, + 0.892, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\hat {W} ^ {\\text {s h a r e d}} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {d \\times h} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.395, + 0.892, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\hat {W} _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} \\hat {B} _ {i k} \\hat {E} _ {i k} \\hat {A} _ {i k} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.439, + 0.892, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\hat {E} _ {i k} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {r \\times r} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.518 + ], + "angle": 0, + "content": "where \\(\\hat{B}_{ik} \\in \\mathbb{R}^{d \\times r}\\) and \\(\\hat{A}_{ik} \\in \\mathbb{R}^{r \\times h}\\), \\(r \\ll d, h\\). Pseudocode of LORSA calculating \\(\\hat{W}_i^{\\mathrm{private}}\\) for a specific layer \\(i\\) can be seen in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.851, + 0.545 + ], + "angle": 0, + "content": "3.3. Applying LORS to AdaMixer's Decoders" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.598 + ], + "angle": 0, + "content": "We apply LORS to the parameters of linear transformations belonging to ACM, ASM, and \\(L_{\\mathrm{output}}\\) in each of AdaMixer's decoders. These parameters are explained in 3.1." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The overall pipeline of LORS running in AdaMixer is illustrated in Figure 4. Specifically, for each group of sampling points, \\(\\mathrm{LORS^A}\\) (see 3.2) is used to reduce parameters in \\(M_{c}\\) (from \\(\\mathbb{R}^{d_q}\\) to \\(\\mathbb{R}^{C\\times C}\\)) of ACM and \\(M_{s}\\) of ASM (from \\(\\mathbb{R}^{d_q}\\) to \\(\\mathbb{R}^{P_{\\mathrm{in}}\\times P_{\\mathrm{out}}}\\)), and \\(\\mathrm{LORS^T}\\) (see 3.2) to minimize parameters in \\(L_{\\mathrm{output}}\\) (from \\(\\mathbb{R}^{C\\times P_{\\mathrm{out}}}\\) to \\(\\mathbb{R}^{d_q}\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Thus, the parameter quantities of \\( M_{c} \\), \\( M_{s} \\), and \\( L_{\\mathrm{output}} \\) are \\( d_{q} \\times C \\times C \\), \\( d_{q} \\times P_{\\mathrm{in}} \\times P_{\\mathrm{out}} \\) and \\( d_{q} \\times C \\times P_{\\mathrm{out}} \\) respectively. When the sampling strategy consists of 2 groups with 64 points each, which is the default setting in our experiments, the values of the variables are as follows: \\( d_{q} = 256 \\), \\( C = 64 \\), \\( P_{\\mathrm{in}} = 64 \\), and \\( P_{\\mathrm{out}} = 128 \\). It can be easily calculated that the number of parameters for each of \\( M_{c} \\), \\( M_{s} \\), and \\( L_{\\mathrm{output}} \\) exceeds one million. Indeed, these three components collectively account for the most of the total parameter in the AdaMixer model with ResNet-50 as the backbone, whereas they are also the primary drivers of enhanced model performance. This is what motivates us to conduct LORS experiments on them. The effects of LORS applied to them will be discussed in detail in 3.4." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15869" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.102, + 0.462, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.472, + 0.406 + ], + "angle": 0, + "content": "Figure 4. The overall pipeline of our proposed LORS, consisting of both adaptive and static parts, each further composed of shared and private components, works collaboratively. The figure illustrates the entire computation process within one layer of the stacked layers, with an enlarged example of the \\(i\\)-th layer for demonstration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.371, + 0.453 + ], + "angle": 0, + "content": "3.4. Analysis on Parameter Reduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Let \\( W \\in \\mathbb{R}^{d \\times h} \\) be a weight parameter that exists in every layer of stacked structures, \\( N \\) is the number of stacked layers. If static, it originally has \\( d \\times h \\) parameters, while using LORS\\(^{\\mathrm{T}}\\) requires \\( \\frac{1}{N} \\times d \\times h + K \\times (d \\times r + r \\times h) \\) parameters on average per layer; if it is adaptive, generating it by \\( q \\) with linear transformation requires \\( d_{q} \\times d \\times h \\) parameters, where \\( d_{q} \\) is the dimension of \\( q \\), and using LORS\\(^{\\mathrm{A}}\\) requires \\( \\frac{1}{N} \\times d_{q} \\times d \\times h + K \\times (d_{q} \\times r^{2} + d \\times r + r \\times h) \\) parameters on average per layer. To more intuitively display the parameter reduction effect of LORS, we set \\( d_{q} = 256 \\), \\( d = 64 \\), \\( h = 128 \\), \\( K = 2 \\) for the ASM process and \\( d = 2 \\times 128 \\times 128 \\), \\( h = 256 \\), \\( K = 1 \\) for the \\( L_{output} \\), which is actually the case of AdaMixer's decoders in our experiment, and show the parameter reduction for different \\( r \\) values using LORS\\(^{\\mathrm{T}}\\) and LORS\\(^{\\mathrm{A}}\\) in Table 1." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.705, + 0.455, + 0.83 + ], + "angle": 0, + "content": "
r=4r=8r=16r=32
W/. LORST1.53M1.66M1.93M2.46M
W/O. LORST8.39M8.39M8.39M8.39M
Percentage18.3%19.8%23.0%29.3%
W/. LORSA0.36M0.39M0.49M0.89M
W/O. LORSA2.10M2.10M2.10M2.10M
Percentage17.2%18.6%23.3%42.4%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Table 1. Analysis of the parameter reduction effect of LORS\\(^{\\mathrm{T}}\\) and LORS\\(^{\\mathrm{A}}\\) with varying rank \\(r\\) values. Our default selections are colored gray. Here the average amount of shared parameters over each layer has already been taken into account." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.633, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.207 + ], + "angle": 0, + "content": "In this section, we first explain the implementation details of our experiments, including the dataset, training specifics, and different parameter initialization methods for static and adaptive LORS. Next, we present the main experimental results. Finally, we perform ablation studies to examine LORS's design and its individual components." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.219, + 0.719, + 0.235 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.243, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Dataset. Our experiments were conducted on the widely-used MS COCO [29] dataset in mmdetection codebase [5], using the standard metrics for object detection. All models were trained on the train2017 split (\\(\\sim 118\\mathrm{k}\\) images) and then evaluated on the val2017 split (5k images)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.893, + 0.652 + ], + "angle": 0, + "content": "Training. We use the AdamW optimizer [33] with a weight decay of 0.0001 and train all models on 8 Nvidia V100 GPUs with a batch size of 16 and a learning rate of \\(2.5 \\times e^{-5}\\). Models are trained for either 12 or 36 epochs, with learning rate decreasing by a factor of 10 at epochs 8 and 11 for 12-epoch training, and at epochs 24 and 33 for 36-epoch training. The low-rank values are set as \\(r = 16\\) for LORS\\(^{\\text{A}}\\) and \\(r = 8\\) for LORS\\(^{\\text{T}}\\). The number of parameter groups is set as as \\(K = [1,1,2,2,3,3]\\) for LORS\\(^{\\text{A}}\\) in all experiments, applied to ACM and ASM in AdaMixer's decoders, and \\(K = [1,1,1,1,1,1]\\) for \\(L_{output}\\) in all experiments. We divide the feature channels into 2 groups with 64 sampling points each instead of AdaMixer's default 4 groups with 32 sampling points each, aiming to increase the parameter compressible space for LORS, which does not improve performance according to both AdaMixer paper [12] and our experiments. Backbones are initialized with ImageNet-1k [9] pre-trained models, and LORS-related parameter initialization is detailed below, remaining parameters are Xavier-initialized [14]. Finally, all other aspects about model training, like the data augmentation, the loss function, etc., just follow AdaMixer's settings [12]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.697 + ], + "angle": 0, + "content": "Initialization Strategies We tried various initialization methods for each component in LORS and determined the overall initialization method as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.698, + 0.89, + 0.742 + ], + "angle": 0, + "content": "- LORST: For static LORS, we employ Kaiming initialization [17] for \\( W^{\\text{shared}} \\) and each \\( B \\), and zero initialization for each \\( A \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.744, + 0.89, + 0.819 + ], + "angle": 0, + "content": "- LORSA: For adaptive LORS, we apply Kaiming initialization [17] to the linear transformation weights forming each \\(\\hat{W}^{\\mathrm{shared}}\\), as well as each \\(\\hat{B}\\) and \\(\\hat{A}\\). Additionally, we use zero initialization for the linear transformation weights forming each \\(\\hat{E}\\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.698, + 0.89, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.64, + 0.846 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To reliably demonstrate that LORS can reduce the number of parameters in stacked structures while maintaining model performance, which is measured by AP metric, we follow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15870" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.179 + ], + "angle": 0, + "content": "
MethodQueriesEpochsDecoder Params(M)Params(M)Decoder GFLOPsGFLOPsTraining HoursAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixer10012110135121048.5h42.761.545.924.745.459.2
AdaMixer + LORS100123560181109.5h42.661.446.025.045.658.5
AdaMixer300121131393913210h44.163.447.427.046.959.5
AdaMixer + LORS3001235605614912h44.163.047.727.847.059.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.19, + 0.893, + 0.22 + ], + "angle": 0, + "content": "Table 2. \\(1 \\times\\) training scheme performance on COCO 2017 val set with ResNet-50 as backbone. AdaMixer with LORS can achieve competitive results while employing a notably reduced number of parameters. FPS is obtained with a single Nvidia V100 GPU." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.233, + 0.891, + 0.394 + ], + "angle": 0, + "content": "
MethodbackboneQueriesEpochsParams(M)GFLOPsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixerR-501003613510443.261.846.725.046.158.8
AdaMixer + LORSR-50100366011043.762.347.325.546.460.0
AdaMixerR-503003613913247.066.051.130.150.261.8
AdaMixer + LORSR-50300366014947.666.652.031.150.262.5
AdaMixerR-1013003615820848.067.052.430.051.263.7
AdaMixer + LORSR-101300367922548.267.552.631.751.363.8
AdaMixerSwin-S3003616423451.371.255.734.254.667.3
AdaMixer + LORSSwin-S300368525051.871.656.435.455.068.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.404, + 0.893, + 0.448 + ], + "angle": 0, + "content": "Table 3. \\(3 \\times\\) training scheme performance on COCO 2017 val set, considering different combinations of backbone and query numbers. Longer training time allows LORS to be fully trained and perform better. The comprehensive improvement in performance and the significant reduction in parameters demonstrate the effectiveness of LORS. The best results in each category are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.565 + ], + "angle": 0, + "content": "common practices in the object detection field and present experimental comparisons in two categories based on training schemes. The first involves training for 12 epochs using a weaker data augmentation, and the second involves training for 36 epochs using a stronger data augmentation, both of which are identical to those in AdaMixer [12]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Table 2 presents the comparison of performance with and without the LORS technique, under a \\(1\\times\\) training scheme. Metrics such as the number of parameters, GFLOPs, and Average Precision (AP) at various scales are evaluated. When the AdaMixer model is combined with LORS, it demonstrates a remarkable reduction in the number of parameters. This reduction is consistently observed across various query quantities and different training epochs, encompassing both the decoder and overall model. For example, when trained with 100 queries for 12 epochs, the AdaMixer+LORS model uses only 35M decoder parameters and 60M total parameters, compared to 110M and 135M respectively in the AdaMixer model. This shows reduction of approximately \\(70\\%\\) in the decoder's parameters, which is substantial. However, the AdaMixer+LORS model shows a slight increase in GFLOPs. In terms of model performance, the AdaMixer+LORS model achieves competitive results, even with a significantly reduced number of parameters. When trained with 100 queries, although the AP of 42.6 is slightly lower than the 42.7 AP of the vanilla AdaMixer, it still slightly outperforms in \\(AP_{s}\\) and \\(AP_{m}\\). When trained with 300 queries, the use of LORS does not" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.893, + 0.61 + ], + "angle": 0, + "content": "affect the performance and allows an advantage in all \\(AP_{s}\\), \\(AP_{m}\\) and \\(AP_{l}\\). These findings suggest that even under conditions of limited model training and data augmentation, LORS does not impact training result, the model can still converge quickly to its original limit. Additionally, we also demonstrate the slight impact of the current LORS method on inference speed, which might be improved by reducing the serial and redundant computations within LORS but not the focus of this paper." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.894, + 0.734 + ], + "angle": 0, + "content": "Table 3 showcases the remarkable performance of the AdaMixer + LORS method under the \\(3 \\times\\) training scheme with different backbones and query numbers. It can be observed that the proposed method consistently outperforms the vanilla AdaMixer across all backbones, query numbers, and evaluation metrics. This result is somewhat surprising to us, as the LORS enables the model to use significantly fewer parameters during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Specifically, when adopting the ResNet-50 backbone and 100 queries, LORS improves the AP value of AdaMixer by 0.5 points (43.7 vs 43.2). It is worth noting that in Table 2, for experiments with the same configuration except for training 12 epochs, the AP with LORS is 0.1 lower than without LORS. However, this situation is reversed when training for 36 epochs. The longer training time under the \\(3 \\times\\) training scheme allows LORS to be fully trained and perform better. This emphasizes the importance of sufficient training time for the LORS technique to fully exploit its potential in enhancing the performance. Furthermore," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15871" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.212 + ], + "angle": 0, + "content": "when using more powerful backbones and a larger number of queries, the performance metrics of models utilizing LORS consistently and comprehensively surpass those of their counterparts without LORS. Even when employing a strong backbone such as Swin-S [31] and using 300 queries, LORS still manages to further improve the model's performance while reducing nearly half of the overall parameters. Much fewer parameters, yet improved performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.474, + 0.442 + ], + "angle": 0, + "content": "In conclusion, the results in Table 2 and Table 3 show the effectiveness of the LORS method in reducing the number of parameters while achieving even better performance. The method's adaptability to various backbones and query numbers further demonstrates its versatility and potential for broader applications. As for why LORS can achieve better performance with fewer parameters, a plausible explanation is that after the model has been sufficiently trained and stabilized, the parameter matrix within the stacked structure is inherently sparse and low-rank. Without using LORS, the model needs to learn such a low-rank structure through training. In contrast, LORS directly and explicitly makes the structure sparse, which, to some extent, can be considered as adding a form of regularization to the model training, thereby facilitating the model's learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.232, + 0.468 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.52 + ], + "angle": 0, + "content": "To save computational resources, all ablation studies are conducted using a ResNet-50 backbone and a \\(1 \\times\\) training scheme." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.532, + 0.402, + 0.626 + ], + "angle": 0, + "content": "
LORSALORSTDecoder Params(M)AP
11042.5†
✓7942.6
✓6642.6
✓✓3542.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.635, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Table 4. Effect of LORSA and LORST. †” denotes this AP result was reproduced by ourselves. Both LORST and LORSA can reduce parameters without compromising performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Adaptive & Static LORS. We first conduct ablation studies on the impact of \\(\\mathrm{LORSA}\\) and \\(\\mathrm{LORST}\\) on model parameters and performance, as shown in Table 4. In this table, a \\(\\checkmark\\) indicates that we use the LORS method to replace the corresponding component in the AdaMixer model mentioned earlier, while a blank cell means we use the original AdaMixer module. If neither of the two is present, the model is identical to the original AdaMixer. As can be seen from the table, both adaptive and static LORS can effectively reduce the parameter size of the decoder in the model (31M and 44M, respectively) without affecting the model performance. Note that the symbol \\(\\dagger\\) here represents that the 42.5 AP value is reproduced by us, which is slightly dif" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "ferent from the 42.7 AP recorded in the original paper. We believe this discrepancy can be attributed to the randomness of 1x training and does not affect our conclusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.893, + 0.379 + ], + "angle": 0, + "content": "Shared and private weights. The second point of interest is to determine which of the shared weight and private weight has a greater impact on the decoder's performance. Experiment in Table 5 is designed to address this question, which is performed on ACM and ASM of AdaMixer's decoder. Recall that the formula we use to generate the final weight parameters is \\( W = W^{\\text{shared}} + W^{\\text{private}} \\). In the table, a \\( \\checkmark \\) indicates that the corresponding term is present on the right side of the formula, otherwise, it is removed. It can be seen that the absence of either term will cause a nonnegligible decrease in model performance, and the impact of missing \\( W^{\\text{private}} \\) is even more severe. This suggests that although the number of parameters in each \\( W^{\\text{private}} \\) is smaller than that of \\( W^{\\text{shared}} \\), it could be more important, possibly because the most informative part of each layer in the stacked structure is contained within it." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.39, + 0.825, + 0.473 + ], + "angle": 0, + "content": "
WsharedWprivateDecoder Params(M)AP
✓5840.9
✓4341.4
✓✓3542.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Table 5. Effect of shared parameters \\( W^{\\text{shared}} \\) and private parameters \\( W^{\\text{private}} \\) Default choice for our model is colored gray." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Hyperparameters of \\(\\mathbf{LORS}^A\\). Another aspect we would like to explore is the optimal number of parameter groups and the value of rank \\(r\\) for adaptive LORS. Here, we feel it is necessary to explain why we use grouped parameters. Recalling formula 11, the adaptive private parameters are obtained by continuous multiplication of three matrices, so the matrix with the smallest rank among them determines the rank of multiplication, which is the upper limit for the private parameters. If the task-specific requirement for the rank of the private parameters matrix unfortunately exceeds the rank of the smallest component, there will be a bottleneck when training the model. Therefore, we designed this parameter group mechanism, using the sum of the results of multiple matrix multiplication to break the aforementioned problem. Another advantage is that the rank of the final matrix possibly increases linearly with the number of parameters, while simply using a larger \\(r\\) would lead to a quadratic increase in the number of parameters. In the end, looking at Table 6, we observed that the experimental result for the 6-layer decoder of the AdaMixer object detector is relatively better when using [1,1,2,2,3,3] groups of parameters per layer and \\(r = 16\\) for each group. This slightly strange parameter group numbers is inspired by AdaLoRA [53]. It suggests that higher-level neural network parameters may" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15872" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.136 + ], + "angle": 0, + "content": "need to allocate more ranks when performing LoRA fin-tuning, which we suppose may also hold true for LORS, and the experiment results confirmed that." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.148, + 0.405, + 0.255 + ], + "angle": 0, + "content": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 183241.8
1 1 1 1 1 1163442.1
2 2 2 2 2 2163542.4
1 1 2 2 3 3163542.6
1 1 1 1 1 1323841.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.266, + 0.468, + 0.321 + ], + "angle": 0, + "content": "Table 6. Validation accuracy with different groups of parameters and rank \\( r \\) in LORSA. Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.342, + 0.468, + 0.54 + ], + "angle": 0, + "content": "Hyperparameters of \\(\\mathbf{LORS}^T\\). We further investigate the optimal configuration settings for \\(\\mathbf{LORS}^T\\) to achieve satisfactory performance, as shown in Table 7. We observed that the best option is to use one group of low-rank parameters with \\(r = 8\\) for the decomposition of each layer \\(L_{\\mathrm{output}}\\) in the stacked decoder of AdaMixer, which can achieve a model performance comparable to the original detector. Since using one group of parameters already yields reasonably good results, we did not explore the use of more parameter groups. Moreover, we should also note that the value of rank \\(r\\) is not necessarily better when larger or smaller, which may indicate that the low-rank structure required by the model for a specific task has its own particularity." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.552, + 0.408, + 0.661 + ], + "angle": 0, + "content": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 123442.3
1 1 1 1 1 143441.9
1 1 1 1 1 183542.6
1 1 1 1 1 1163742.1
1 1 1 1 1 1324041.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.672, + 0.468, + 0.727 + ], + "angle": 0, + "content": "Table 7. Validation accuracy with different rank \\( r \\) in LORS\\(^{\\mathrm{T}}\\). Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.75, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Number of decoders. Finally, we conclude our ablation studies with experiments on the optimal number of decoder layers incorporating the LORS structure for better performance. Since our method significantly reduces the total number of model parameters and allows each decoder layer to train and use shared parameters that capture common features, it is natural to consider whether the model's inherent characteristics are affected and whether using more or fewer decoding layers would further improve the model's performance. With this idea in mind, we conducted sev" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.891, + 0.197 + ], + "angle": 0, + "content": "eral experiments as shown in Table 8 to explore this issue. However, similar to the original AdaMixer without LORS, 6 layers still yield the best performance for the AdaMixer with LORS, having too many or too few layers would degrade the model's performance. This may also suggest that LORS merely captures the intrinsic structure of the model without altering its original properties." + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.217, + 0.796, + 0.314 + ], + "angle": 0, + "content": "
Decoder\nNumberDecoder\nParams(M)AP
32838.9
63542.6
94242.0
125041.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.503, + 0.325, + 0.891, + 0.394 + ], + "angle": 0, + "content": "Table 8. Model performance using different numbers of stacked decoders when employing LORS. The default choice for our model is colored gray. This experiment with varying numbers of decoders demonstrates that six layers still yield the best performance for the model, which is not changed by LORS." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.446, + 0.617, + 0.461 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.474, + 0.891, + 0.686 + ], + "angle": 0, + "content": "This research introduced a novel approach to the reduction of parameters within deep learning models utilizing stacked structures, i.e., the Low-rank Residual Structure (LORS). It is potentially an effective methodology in reducing both static and adaptive parameters. Broadly speaking, LORS allows the parameters of stacked modules to be largely shared, while maintaining only a small number of unique parameters for each module, thereby significantly reducing the total number of parameters whilst not compromising performance. We validated our method via object detection task across various extensive experimental cases on the widely-used COCO dataset, and the results demonstrate a superior model performance even with a \\(50\\% - 70\\%\\) reduction in decoder parameters, surpassing our expectations." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.689, + 0.891, + 0.9 + ], + "angle": 0, + "content": "There are some noticeable limitations in our study. While effective in reducing model parameters without compromising performance, enhancing model capabilities requires a comparatively long training process for the method. Additionally, our method slightly increases the inference time, as even without a detailed analysis, it can be roughly seen that LORS has unparalleled and repetitive computations needing more optimization. Finally, we have only tested our approach on object detection tasks, specifically with the AdaMixer model and its decoder structure. It's clear that our method can be applied to more tasks (e.g., NLP), different models (e.g., language models), and other neural network components (e.g., backbones). These aspects will be the focus of our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Arnav Chavan, Zhuang Liu, Deepak Gupta, Eric Xing, and Zhiqiang Shen. One-for-all: Generalized lora for parameter-efficient fine-tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[6] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[8] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems, 28, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.47, + 0.672 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.799 + ], + "angle": 0, + "content": "[11] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3621–3630, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[12] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Yanqi Ge, Qiang Nie, Ye Huang, Yong Liu, Chengjie Wang, Feng Zheng, Wen Li, and Lixin Duan. Beyond proto" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "types: Semantic anchor regularization for better representation learning. arXiv preprint arXiv:2312.11872, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[15] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[16] Song Han, Jeff Pool, John Tran, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[22] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[23] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15874" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[28] Vladislav Lialin, Namrata Shivagunde, Sherin Muckatira, and Anna Rumshisky. Stack more layers differently: High-rank training through low-rank updates, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.468, + 0.219 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.468, + 0.346 + ], + "angle": 0, + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.468, + 0.475 + ], + "angle": 0, + "content": "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.468, + 0.559 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.468, + 0.603 + ], + "angle": 0, + "content": "[36] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.468, + 0.659 + ], + "angle": 0, + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.759 + ], + "angle": 0, + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.816 + ], + "angle": 0, + "content": "[40] Farhana Sultana, Abu Sufian, and Paramartha Dutta. A review of object detection models based on convolutional neural network. Intelligent computing: image processing based applications, pages 1-16, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[41] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[42] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[43] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[44] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261-24272, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[45] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter efficient tuning of pretrained models using dynamic search-free low-rank adaptation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[48] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.62 + ], + "angle": 0, + "content": "[49] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[50] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[51] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[52] Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[53] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[54] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15875" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[55] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[57] Bojia Zi, Xianbiao Qi, Lingzhi Wang, Jianan Wang, KamFai Wong, and Lei Zhang. Delta-lora: Fine-tuning high-rank parameters with the delta of low-rank matrices, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15876" + } + ] +] \ No newline at end of file diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_origin.pdf b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a88c3c9c897165c423ba212869e91a5d1a516867 --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/2ca03dd9-d123-452f-a1c7-2252d386e5e6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23e22b5f8d4f0724a2ff05819d121a3824143f5462bb40d78e16e4b1a7ee2e6f +size 633732 diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/full.md b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f0e1156b5df9107eaf5858496691dd489f10159 --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/full.md @@ -0,0 +1,348 @@ +# LORS: Low-rank Residual Structure for Parameter-Efficient Network Stacking + +Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, Chengjie Wang, Youtu Lab, Tencent + +{jarenli, ryanfwu, gleelin, guangpintao, choasliu,Jasoncjwang}@tencent.com, qnie.cuhk@gmail.com + +# Abstract + +Deep learning models, particularly those based on transformers, often employ numerous stacked structures, which possess identical architectures and perform similar functions. While effective, this stacking paradigm leads to a substantial increase in the number of parameters, posing challenges for practical applications. In today's landscape of increasingly large models, stacking depth can even reach dozens, further exacerbating this issue. To mitigate this problem, we introduce LORS (LOw-rank Residual Structure). LORS allows stacked modules to share the majority of parameters, requiring a much smaller number of unique ones per module to match or even surpass the performance of using entirely distinct ones, thereby significantly reducing parameter usage. We validate our method by applying it to the stacked decoders of a query-based object detector, and conduct extensive experiments on the widely used MS COCO dataset. Experimental results demonstrate the effectiveness of our method, as even with a $70\%$ reduction in the parameters of the decoder, our method still enables the model to achieve comparable or even better performance than its original. + +# 1. Introduction + +In the current era of prosperity for large models, a common issue is the significant increase in the number of parameters, which presents challenges for training, inference, and deployment. Various methods have been proposed to reduce the number of parameters in models, such as knowledge distillation [15, 20], which compresses large models into smaller ones while trying to preserve their performance but may still lead to a decrease in model capacity; pruning [16, 55], which removes redundant parameters from the model but can affect the model's stability; quantization [8], which reduces the numerical precision of model parameters to lower storage and computation but may cause model accuracy loss; and parameter sharing [24], which reduces the number of parameters by sharing them across different layers but may limit the model's expressiveness. + +![](images/8a1a184e916f58e96ef2fcee1384e3911359bcd46feed7acdc3ea62fd980dba5.jpg) +Figure 1. The LORS calculation process, which could be adaptive or static, depending on whether an adaptively generated kernel is used in the matrix manipulation for private parameters. + +Different from the aforementioned methods, we have observed an important fact contributing to the large number of parameters: the widespread use of stacking in neural networks. Stacking refers to those modules that have identical architectures and perform the same or similar functions, but possess different parameters due to random initialization as well as training updates. Examples of stacking can be found in numerous prominent neural networks, like the classic ResNet model [18] and Transformers [46]. Particularly, Transformers heavily rely on stacked structures and typically employ completely identical multi-layer stacks in both encoders and decoders. It now serves as an indispensable component of many excellent models in fields such as computer vision and natural language processing. + +Although stacking is powerful for enhancing model capacity, as demonstrated by large language models, it also + +naturally leads to a sharp increase in the number of parameters. For example, GPT-3 [2] utilizes 175 billion parameters and consists of 96 layers of stacked Transformer [46] layers. How can we enjoy the benefits of stacking while reducing the required number of parameters? We notice that stacked decoders have identical structures and similar functions, indicating that there should be some commonality among their parameters. However, since they handle different input and output distributions, there must also be unique aspects to their parameters. Therefore, a natural idea is: it may be possible to represent the shared aspects with shared parameters while allowing each stacked module to retain only the parameters that capture its unique characteristics, thereby reducing the overall parameter usage. + +Based on the above considerations, we propose to decompose the parameters of stacked modules into two parts: shared ones representing the commonality and private ones capturing the specific characteristics. Shared parameters are available for all modules and trained jointly by them, while the private parameters are separately owned by each module. We suppose that the way of sharing parameters related to the commonality may reduce the number of parameters while maintaining the performance of the model. To achieve this goal, we introduce the concept of Low-rank Residual Structure (LORS), inspired by the approach of LoRA [21]. LORS essentially adds the unique parameters to the shared ones, just like residual connections add residual information to the features. While LoRA is originally designed for fine-tuning, we train our LoRA-like operation on parameters from scratch. This approach allows us to effectively reduce the overall parameter usage while maintaining the performance of the model, paving the way for more parameter-efficient network stacking. + +To validate our idea, we choose AdaMixer [12], a strong query-based object detector, as our experimental subject. It contains a large number of both adaptive and static parameters in its stacked decoders, making it an ideal candidate to demonstrate LORS's effectiveness. The difference between adaptive and static parameters lies in whether they change with different inputs. Our goal is to show that LORS can effectively reduce the overall usage of both types of parameters while maintaining the model's performance. We conducted extensive experiments on this detector, illustrating that our method succeeded in reducing up to $70\%$ of parameters in AdaMixer's decoders, while still enabling the model to achieve comparable or even superior performance compared to its vanilla version. In summary, our contributions can be concluded as: + +- We propose a novel low-rank residual structure, named LORS, for network stacking, which can greatly reduce the number of parameters while maintaining or even improving the performance compared to the vanilla one. +- We introduce effective methodologies for reducing both + +static and adaptive generated parameters in stacked structures, which makes our proposed LORS a more versatile method. + +- Our method holds the potential to serve as one of the basic network structures for large models that are affected by the issue of excessive parameters due to stacking, rendering them more parameter-efficient, thereby facilitating an easier implementation in practical applications. + +# 2. Related Work + +Models with stacked structures. Many nerval networks partially or extensively employ stacked modules. CNN-based models [40], for instance, are widely applied across various computer vision tasks, such as classification [18, 22, 39, 49], detection [19, 25, 36-38, 43], segmentation [7, 13, 32, 54], etc. These models often use stacked smaller modules within larger components. Another kind of models primarily utilize Multi-Layer Perceptrons (MLPs), such as the MLP-Mixer [44], whose main body comprises dozens of completely identical stacked mixer layers. Moreover, since its invention, Transformer [46] has been successfully applied to diverse domains, including computer vision [3, 30, 41, 56], natural language processing [2, 10, 50], multi-modal learning [6, 27, 35, 47], etc. In the thriving arena of large language models, Transformer-based structures are indeed indispensable for many cutting-edge works [10, 26, 34, 42, 48]. Transformers are usually used to form stacked multi-layers encoders or decoders, within each layer being structurally identical. + +LoRA and Its Variants. LoRA [21] is a technique proposed for fine-tuning large language models, it introduces low-rank decomposition matrices in the Transformer [46] architecture, significantly reducing trainable parameters and GPU memory requirements. Afterward, a series of works with modifications as well as improvement have been proposed. AdaLoRA [53] adaptively allocates parameter budgets among weight matrices based on their importance scores and uses singular value decomposition for incremental updates, leading to an improved fine-tuning performance, especially in low budget settings. DyLoRA [45] introduces a dynamic low-rank adaptation technique that trains LoRA blocks for a range of ranks instead of a single rank, addressing the issues of fixed block size and rank optimization in LoRA. Delta-LoRA [57] presents a parameter-efficient approach to fine-tune large language models by updating not only the low-rank matrices but also the pre-trained weights using the delta of the product of two low-rank matrices. LoRA-FA [52] introduces a memory-efficient fine-tuning method that reduces activation memory without performance degradation or expensive recomputation, it achieves this by freezing the projection-down weight and updating the projection-up weight in each LoRA layer. ReLoRA [28] is a parameter-efficient method for training + +large neural networks using low-rank updates. GLoRA [4], or Generalized LoRA, is an advanced approach for universal parameter-efficient fine-tuning tasks, which enhances the low rank adaptation technique by employing a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations. VeRA [23], or Vector-based Random Matrix Adaptation, further compresses the training parameters compared to the vanilla LoRA [21]. It achieves this by utilizing shared low-rank matrices across all layers and learning small scaling vectors. + +Inspired by the aforementioned LoRA series works, we propose our method called LORS (Low Rank Residual Structure) for parameter efficiency. However, our approach is fundamentally distinct from the LoRA series works. While LoRA uses different low-rank weights to fine-tune different pretrained weights, LORS adds different low-rank weights per layer to the same common weights shared by all layers, thus creating parameter-efficient stacked models trained from scratch, requiring no pretrained weights. + +# 3. Approach + +In this section, we initially revisit the foundational works upon which our methodology and experiments are primarily based, encompassing the LoRA mechanism [21], query-based object detection [3], and the structure of AdaMixer's decoders [12]. Then we mathematically formulate our proposed LORS, which comprises two parts: the adaptive one and the static one. Subsequently, we elaborate on how the LORS method is applied to reduce the parameters of AdaMixer's decoders. Lastly, we provide a quantitative analysis of the parameter reduction percentage attained by our proposed method. + +# 3.1. Preliminary + +The mechanism of LoRA. The Low-Rank Adaptation (LoRA) [21] technique is a novel approach designed to adapt large pre-trained language models for specific tasks. The key idea of LoRA is to introduce a low-rank parameter matrix which is able to captures task-specific knowledge while the original pre-trained parameters remain fixed. + +Mathematically, given a pre-trained parameter matrix $W \in \mathbb{R}^{d \times h}$ , LoRA uses a low-rank matrix $B \in \mathbb{R}^{d \times r}$ and a projection matrix $A \in \mathbb{R}^{r \times h}$ to adapt $W$ , where $r \ll d, h$ . The adapted parameter matrix is then given by + +$$ +W + \Delta W = W + B A \tag {1} +$$ + +where $BA$ captures the task-specific knowledge. + +The key advantage of LoRA is that it can significantly reduce the number of parameters that need to be fine-tuned, thereby reducing the computational cost and lowering down the memory requirement. In some cases, even single-digit values of $r$ are sufficient to fine-tune the model to the desired state, which is often tens of times less expensive than + +training the parameters in $W$ . Furthermore, by keeping the original parameters fixed, LoRA avoids catastrophic forgetting, a common issue in fine-tuning large models. + +Query-based object detection. In the realm of object detection, query-based detectors have established a new paradigm [11, 12, 30, 41, 51, 56]. Unlike traditional detectors which rely on anchor boxes or sliding windows, query-based models utilize a set of learnable queries to interact with image feature maps. This interaction can be formalized with attention [46] operations as + +$$ +Q _ {\text {u p d a t e d}} = \operatorname {A t t e n t i o n} (Q, K (V)) \tag {2} +$$ + +where $Q, K$ , and $V$ represent the queries, keys, and values. The learnable queries $Q$ are used to predict object classes and bounding boxes ultimately, while $K$ and $V$ typically originate from encoded image features. It is a common practice to continuously refine $Q$ through interactions with $K$ and $V$ using successive decoding layers. These layers are usually composed of structurally identical decoders. + +Decoders of AdaMixer. AdaMixer [12] is a query-based detector that features an innovative decoder design. This design includes Adaptive Channel Mixing (ACM) and Adaptive Spatial Mixing (ASM) methods, which greatly enhance its performance [12]. + +Given a sampled feature $\mathbf{x} \in \mathbb{R}^{P_{\mathrm{in}} \times C}$ , where $C = d_{\mathrm{feat}} / g$ , and $g$ denotes the number of sampling groups. This sampled feature is obtained through an operation called group sampling. This operation divides the feature space channel $d_{\mathrm{feat}}$ into $g$ groups and performs individual sampling for each group. Then, ACM (Adaptive Channel Mixing) utilizes a weight adapted by the object query $\mathbf{q}$ to transform feature $\mathbf{x}$ in the channel dimension, enhancing channel semantics [12]: + +$$ +M _ {c} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {C \times C} \tag {3} +$$ + +$$ +\operatorname {A C M} (\mathbf {x}) = \operatorname {R e L U} \left(\operatorname {L a y e r N o r m} \left(\mathbf {x} M _ {c}\right)\right) \tag {4} +$$ + +where LayerNorm stands for Layer Normalization [1]. + +Next is the ASM (Adaptive Spatial Mixing) process, which aims to enable the adaptability of the object query $\mathbf{q}$ to spatial structures of sampled features [12] by applying adaptive transformation to the spatial dimension: + +$$ +M _ {s} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {P _ {\text {i n}} \times P _ {\text {o u t}}} \tag {5} +$$ + +$$ +\operatorname {A S M} (\mathbf {x}) = \operatorname {R e L U} \left(\operatorname {L a y e r N o r m} \left(\mathbf {x} ^ {T} M _ {s}\right)\right), \tag {6} +$$ + +Both ACM and ASM train independent parameters for each sampling group, and finally the output with the shape $\mathbb{R}^{g\times C\times P_{\mathrm{out}}}$ is flattened and transformed to the $d_q$ dimension by a linear layer $L_{\mathrm{output}}$ to add back to the object query. + +ACM, ASM and the output linear transformation $L_{output}$ possess significantly more parameters compared to the decoder's other operations, making them the main contributors to the number of parameters. Hence, we choose them + +as the target components for validating the effectiveness of our LORS method in parameter reduction. + +```python +def non_dynamic_fused_weight(W_shared, A, B): + # W_shared: (1, d, h) + # A: (G, r, h) + # B: (G, d, r) + # private weight: (G, d, h) + W_private = B @ A + # sum at group dimension to get fused weight: (d, h) + W = concat([W_shared, W_private], dim=0).sum(axis=0) + return W +``` + +Figure 2. Pseudo-code for obtaining a static weight parameter for one layer. + +# 3.2. Formulation of Our Method + +The complete computational process of LORS is illustrated in Figure 1. Notably, LORS computation is divided into two types: adaptive and static. In the context of AdaMixer, the term "adaptive" indicates whether the transformation matrix depends on the object query. After completing the formalization of LORS here, we will further elaborate on how it is applied to the tasks in our setting. + +Static Low Rank Residual Structure $(\mathbf{LORST}^{\mathrm{T}})$ . Suppose there are $N$ stacked layers of modules with identical architecture, and $W_{i}\in \mathbb{R}^{d\times h}$ be a parameter matrix belonging to the $i$ -th layer. Then, we have: + +$$ +W _ {i} = W ^ {\text {s h a r e d}} + W _ {i} ^ {\text {p r i v a t e}} \tag {7} +$$ + +Here, $W^{\mathrm{shared}} \in \mathbb{R}^{d \times h}$ represents the shared parameters across all stacked layers, while $W_{i}^{\mathrm{private}}$ denotes the layer-specific parameters for the $i$ -th layer, which is calculated as follows: + +$$ +W _ {i} ^ {\text {p r i v a t e}} = \sum_ {k = 1} ^ {K} B _ {i k} A _ {i k} \tag {8} +$$ + +$B_{ik} \in \mathbb{R}^{d \times r}$ , $A_{ik} \in \mathbb{R}^{r \times h}$ , and $r \ll d, h$ . $K$ represents the number of parameter groups used to compute $W_{i}^{\mathrm{private}}$ . Pseudo-code of LORST computing $W_{i}^{\mathrm{private}}$ for a specific layer $i$ can be seen in Figure 2. + +Adaptive Low Rank Residual Structure (LORSA). Let $\hat{W}_i \in \mathbb{R}^{d \times h}$ be an adaptive generated parameter in $i$ -th stacked layer, it is similarly calculated as: + +$$ +\hat {W} _ {i} = \hat {W} ^ {\text {s h a r e d}} + \hat {W} _ {i} ^ {\text {p r i v a t e}} \tag {9} +$$ + +where the cross-layer shared parameter $\hat{W}^{\mathrm{shared}}\in \mathbb{R}^{d\times h}$ and the layer-specific parameter $\hat{W}_i^{\mathrm{private}}\in \mathbb{R}^{d\times h}$ for each layer + +```python +def dynamic_fused_weight(q, A, B): + # q: (N, C) + # A: (1, G, r, h) + # B: (1, G, d, r) + # shared weight: (N, 1, d, h) + W_shared = linear1(pro_feats).reshape(N, 1, d, h) + # E: (N, G, r, r) + E = linear2(pro_feats).reshape(N, G, r, r) + # private weight: (N, G, d, h) + W_private = B @ E @ A + # sum at group dimension to get fused weight: (N, d, h) + W = concat([W_shared, W_private], dim=1).sum(axis=1) + return W +``` + +Figure 3. Pseudo-code for obtaining an adaptive weight parameter for one layer. + +are both calculated based on $q$ : + +$$ +\hat {W} ^ {\text {s h a r e d}} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {d \times h} \tag {10} +$$ + +$$ +\hat {W} _ {i} ^ {\text {p r i v a t e}} = \sum_ {k = 1} ^ {K} \hat {B} _ {i k} \hat {E} _ {i k} \hat {A} _ {i k} \tag {11} +$$ + +$$ +\hat {E} _ {i k} = \operatorname {L i n e a r} (\mathbf {q}) \in \mathbb {R} ^ {r \times r} \tag {12} +$$ + +where $\hat{B}_{ik} \in \mathbb{R}^{d \times r}$ and $\hat{A}_{ik} \in \mathbb{R}^{r \times h}$ , $r \ll d, h$ . Pseudocode of LORSA calculating $\hat{W}_i^{\mathrm{private}}$ for a specific layer $i$ can be seen in Figure 3. + +# 3.3. Applying LORS to AdaMixer's Decoders + +We apply LORS to the parameters of linear transformations belonging to ACM, ASM, and $L_{\mathrm{output}}$ in each of AdaMixer's decoders. These parameters are explained in 3.1. + +The overall pipeline of LORS running in AdaMixer is illustrated in Figure 4. Specifically, for each group of sampling points, $\mathrm{LORS^A}$ (see 3.2) is used to reduce parameters in $M_{c}$ (from $\mathbb{R}^{d_q}$ to $\mathbb{R}^{C\times C}$ ) of ACM and $M_{s}$ of ASM (from $\mathbb{R}^{d_q}$ to $\mathbb{R}^{P_{\mathrm{in}}\times P_{\mathrm{out}}}$ ), and $\mathrm{LORS^T}$ (see 3.2) to minimize parameters in $L_{\mathrm{output}}$ (from $\mathbb{R}^{C\times P_{\mathrm{out}}}$ to $\mathbb{R}^{d_q}$ ). + +Thus, the parameter quantities of $M_{c}$ , $M_{s}$ , and $L_{\mathrm{output}}$ are $d_{q} \times C \times C$ , $d_{q} \times P_{\mathrm{in}} \times P_{\mathrm{out}}$ and $d_{q} \times C \times P_{\mathrm{out}}$ respectively. When the sampling strategy consists of 2 groups with 64 points each, which is the default setting in our experiments, the values of the variables are as follows: $d_{q} = 256$ , $C = 64$ , $P_{\mathrm{in}} = 64$ , and $P_{\mathrm{out}} = 128$ . It can be easily calculated that the number of parameters for each of $M_{c}$ , $M_{s}$ , and $L_{\mathrm{output}}$ exceeds one million. Indeed, these three components collectively account for the most of the total parameter in the AdaMixer model with ResNet-50 as the backbone, whereas they are also the primary drivers of enhanced model performance. This is what motivates us to conduct LORS experiments on them. The effects of LORS applied to them will be discussed in detail in 3.4. + +![](images/184c299ee6cabe449fbc78238f2b4aaab9fdddd7a653486f7d6f51c21a6cceef.jpg) +Figure 4. The overall pipeline of our proposed LORS, consisting of both adaptive and static parts, each further composed of shared and private components, works collaboratively. The figure illustrates the entire computation process within one layer of the stacked layers, with an enlarged example of the $i$ -th layer for demonstration. + +# 3.4. Analysis on Parameter Reduction + +Let $W \in \mathbb{R}^{d \times h}$ be a weight parameter that exists in every layer of stacked structures, $N$ is the number of stacked layers. If static, it originally has $d \times h$ parameters, while using LORS $^{\mathrm{T}}$ requires $\frac{1}{N} \times d \times h + K \times (d \times r + r \times h)$ parameters on average per layer; if it is adaptive, generating it by $q$ with linear transformation requires $d_{q} \times d \times h$ parameters, where $d_{q}$ is the dimension of $q$ , and using LORS $^{\mathrm{A}}$ requires $\frac{1}{N} \times d_{q} \times d \times h + K \times (d_{q} \times r^{2} + d \times r + r \times h)$ parameters on average per layer. To more intuitively display the parameter reduction effect of LORS, we set $d_{q} = 256$ , $d = 64$ , $h = 128$ , $K = 2$ for the ASM process and $d = 2 \times 128 \times 128$ , $h = 256$ , $K = 1$ for the $L_{output}$ , which is actually the case of AdaMixer's decoders in our experiment, and show the parameter reduction for different $r$ values using LORS $^{\mathrm{T}}$ and LORS $^{\mathrm{A}}$ in Table 1. + +
r=4r=8r=16r=32
W/. LORST1.53M1.66M1.93M2.46M
W/O. LORST8.39M8.39M8.39M8.39M
Percentage18.3%19.8%23.0%29.3%
W/. LORSA0.36M0.39M0.49M0.89M
W/O. LORSA2.10M2.10M2.10M2.10M
Percentage17.2%18.6%23.3%42.4%
+ +Table 1. Analysis of the parameter reduction effect of LORS $^{\mathrm{T}}$ and LORS $^{\mathrm{A}}$ with varying rank $r$ values. Our default selections are colored gray. Here the average amount of shared parameters over each layer has already been taken into account. + +# 4. Experiments + +In this section, we first explain the implementation details of our experiments, including the dataset, training specifics, and different parameter initialization methods for static and adaptive LORS. Next, we present the main experimental results. Finally, we perform ablation studies to examine LORS's design and its individual components. + +# 4.1. Implementation Details + +Dataset. Our experiments were conducted on the widely-used MS COCO [29] dataset in mmdetection codebase [5], using the standard metrics for object detection. All models were trained on the train2017 split ( $\sim 118\mathrm{k}$ images) and then evaluated on the val2017 split (5k images). + +Training. We use the AdamW optimizer [33] with a weight decay of 0.0001 and train all models on 8 Nvidia V100 GPUs with a batch size of 16 and a learning rate of $2.5 \times e^{-5}$ . Models are trained for either 12 or 36 epochs, with learning rate decreasing by a factor of 10 at epochs 8 and 11 for 12-epoch training, and at epochs 24 and 33 for 36-epoch training. The low-rank values are set as $r = 16$ for LORS $^{\text{A}}$ and $r = 8$ for LORS $^{\text{T}}$ . The number of parameter groups is set as as $K = [1,1,2,2,3,3]$ for LORS $^{\text{A}}$ in all experiments, applied to ACM and ASM in AdaMixer's decoders, and $K = [1,1,1,1,1,1]$ for $L_{output}$ in all experiments. We divide the feature channels into 2 groups with 64 sampling points each instead of AdaMixer's default 4 groups with 32 sampling points each, aiming to increase the parameter compressible space for LORS, which does not improve performance according to both AdaMixer paper [12] and our experiments. Backbones are initialized with ImageNet-1k [9] pre-trained models, and LORS-related parameter initialization is detailed below, remaining parameters are Xavier-initialized [14]. Finally, all other aspects about model training, like the data augmentation, the loss function, etc., just follow AdaMixer's settings [12]. + +Initialization Strategies We tried various initialization methods for each component in LORS and determined the overall initialization method as follows: + +- LORST: For static LORS, we employ Kaiming initialization [17] for $W^{\text{shared}}$ and each $B$ , and zero initialization for each $A$ . +- LORSA: For adaptive LORS, we apply Kaiming initialization [17] to the linear transformation weights forming each $\hat{W}^{\mathrm{shared}}$ , as well as each $\hat{B}$ and $\hat{A}$ . Additionally, we use zero initialization for the linear transformation weights forming each $\hat{E}$ . + +# 4.2. Main Results + +To reliably demonstrate that LORS can reduce the number of parameters in stacked structures while maintaining model performance, which is measured by AP metric, we follow + +
MethodQueriesEpochsDecoder Params(M)Params(M)Decoder GFLOPsGFLOPsTraining HoursAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
AdaMixer10012110135121048.5h42.761.545.924.745.459.2
AdaMixer + LORS100123560181109.5h42.661.446.025.045.658.5
AdaMixer300121131393913210h44.163.447.427.046.959.5
AdaMixer + LORS3001235605614912h44.163.047.727.847.059.5
+ +Table 2. $1 \times$ training scheme performance on COCO 2017 val set with ResNet-50 as backbone. AdaMixer with LORS can achieve competitive results while employing a notably reduced number of parameters. FPS is obtained with a single Nvidia V100 GPU. + +
MethodbackboneQueriesEpochsParams(M)GFLOPsAP\( AP_{50} \)\( AP_{75} \)\( AP_s \)\( AP_m \)\( AP_l \)
AdaMixerR-501003613510443.261.846.725.046.158.8
AdaMixer + LORSR-50100366011043.762.347.325.546.460.0
AdaMixerR-503003613913247.066.051.130.150.261.8
AdaMixer + LORSR-50300366014947.666.652.031.150.262.5
AdaMixerR-1013003615820848.067.052.430.051.263.7
AdaMixer + LORSR-101300367922548.267.552.631.751.363.8
AdaMixerSwin-S3003616423451.371.255.734.254.667.3
AdaMixer + LORSSwin-S300368525051.871.656.435.455.068.4
+ +Table 3. $3 \times$ training scheme performance on COCO 2017 val set, considering different combinations of backbone and query numbers. Longer training time allows LORS to be fully trained and perform better. The comprehensive improvement in performance and the significant reduction in parameters demonstrate the effectiveness of LORS. The best results in each category are highlighted in bold. + +common practices in the object detection field and present experimental comparisons in two categories based on training schemes. The first involves training for 12 epochs using a weaker data augmentation, and the second involves training for 36 epochs using a stronger data augmentation, both of which are identical to those in AdaMixer [12]. + +Table 2 presents the comparison of performance with and without the LORS technique, under a $1\times$ training scheme. Metrics such as the number of parameters, GFLOPs, and Average Precision (AP) at various scales are evaluated. When the AdaMixer model is combined with LORS, it demonstrates a remarkable reduction in the number of parameters. This reduction is consistently observed across various query quantities and different training epochs, encompassing both the decoder and overall model. For example, when trained with 100 queries for 12 epochs, the AdaMixer+LORS model uses only 35M decoder parameters and 60M total parameters, compared to 110M and 135M respectively in the AdaMixer model. This shows reduction of approximately $70\%$ in the decoder's parameters, which is substantial. However, the AdaMixer+LORS model shows a slight increase in GFLOPs. In terms of model performance, the AdaMixer+LORS model achieves competitive results, even with a significantly reduced number of parameters. When trained with 100 queries, although the AP of 42.6 is slightly lower than the 42.7 AP of the vanilla AdaMixer, it still slightly outperforms in $AP_{s}$ and $AP_{m}$ . When trained with 300 queries, the use of LORS does not + +affect the performance and allows an advantage in all $AP_{s}$ , $AP_{m}$ and $AP_{l}$ . These findings suggest that even under conditions of limited model training and data augmentation, LORS does not impact training result, the model can still converge quickly to its original limit. Additionally, we also demonstrate the slight impact of the current LORS method on inference speed, which might be improved by reducing the serial and redundant computations within LORS but not the focus of this paper. + +Table 3 showcases the remarkable performance of the AdaMixer + LORS method under the $3 \times$ training scheme with different backbones and query numbers. It can be observed that the proposed method consistently outperforms the vanilla AdaMixer across all backbones, query numbers, and evaluation metrics. This result is somewhat surprising to us, as the LORS enables the model to use significantly fewer parameters during both training and inference. + +Specifically, when adopting the ResNet-50 backbone and 100 queries, LORS improves the AP value of AdaMixer by 0.5 points (43.7 vs 43.2). It is worth noting that in Table 2, for experiments with the same configuration except for training 12 epochs, the AP with LORS is 0.1 lower than without LORS. However, this situation is reversed when training for 36 epochs. The longer training time under the $3 \times$ training scheme allows LORS to be fully trained and perform better. This emphasizes the importance of sufficient training time for the LORS technique to fully exploit its potential in enhancing the performance. Furthermore, + +when using more powerful backbones and a larger number of queries, the performance metrics of models utilizing LORS consistently and comprehensively surpass those of their counterparts without LORS. Even when employing a strong backbone such as Swin-S [31] and using 300 queries, LORS still manages to further improve the model's performance while reducing nearly half of the overall parameters. Much fewer parameters, yet improved performance. + +In conclusion, the results in Table 2 and Table 3 show the effectiveness of the LORS method in reducing the number of parameters while achieving even better performance. The method's adaptability to various backbones and query numbers further demonstrates its versatility and potential for broader applications. As for why LORS can achieve better performance with fewer parameters, a plausible explanation is that after the model has been sufficiently trained and stabilized, the parameter matrix within the stacked structure is inherently sparse and low-rank. Without using LORS, the model needs to learn such a low-rank structure through training. In contrast, LORS directly and explicitly makes the structure sparse, which, to some extent, can be considered as adding a form of regularization to the model training, thereby facilitating the model's learning. + +# 4.3. Ablation Study + +To save computational resources, all ablation studies are conducted using a ResNet-50 backbone and a $1 \times$ training scheme. + +
LORSALORSTDecoder Params(M)AP
11042.5†
✓7942.6
✓6642.6
✓✓3542.6
+ +Adaptive & Static LORS. We first conduct ablation studies on the impact of $\mathrm{LORSA}$ and $\mathrm{LORST}$ on model parameters and performance, as shown in Table 4. In this table, a $\checkmark$ indicates that we use the LORS method to replace the corresponding component in the AdaMixer model mentioned earlier, while a blank cell means we use the original AdaMixer module. If neither of the two is present, the model is identical to the original AdaMixer. As can be seen from the table, both adaptive and static LORS can effectively reduce the parameter size of the decoder in the model (31M and 44M, respectively) without affecting the model performance. Note that the symbol $\dagger$ here represents that the 42.5 AP value is reproduced by us, which is slightly dif + +ferent from the 42.7 AP recorded in the original paper. We believe this discrepancy can be attributed to the randomness of 1x training and does not affect our conclusions. + +Shared and private weights. The second point of interest is to determine which of the shared weight and private weight has a greater impact on the decoder's performance. Experiment in Table 5 is designed to address this question, which is performed on ACM and ASM of AdaMixer's decoder. Recall that the formula we use to generate the final weight parameters is $W = W^{\text{shared}} + W^{\text{private}}$ . In the table, a $\checkmark$ indicates that the corresponding term is present on the right side of the formula, otherwise, it is removed. It can be seen that the absence of either term will cause a nonnegligible decrease in model performance, and the impact of missing $W^{\text{private}}$ is even more severe. This suggests that although the number of parameters in each $W^{\text{private}}$ is smaller than that of $W^{\text{shared}}$ , it could be more important, possibly because the most informative part of each layer in the stacked structure is contained within it. + +Table 4. Effect of LORSA and LORST. †” denotes this AP result was reproduced by ourselves. Both LORST and LORSA can reduce parameters without compromising performance. + +
WsharedWprivateDecoder Params(M)AP
✓5840.9
✓4341.4
✓✓3542.6
+ +Table 5. Effect of shared parameters $W^{\text{shared}}$ and private parameters $W^{\text{private}}$ Default choice for our model is colored gray. + +Hyperparameters of $\mathbf{LORS}^A$ . Another aspect we would like to explore is the optimal number of parameter groups and the value of rank $r$ for adaptive LORS. Here, we feel it is necessary to explain why we use grouped parameters. Recalling formula 11, the adaptive private parameters are obtained by continuous multiplication of three matrices, so the matrix with the smallest rank among them determines the rank of multiplication, which is the upper limit for the private parameters. If the task-specific requirement for the rank of the private parameters matrix unfortunately exceeds the rank of the smallest component, there will be a bottleneck when training the model. Therefore, we designed this parameter group mechanism, using the sum of the results of multiple matrix multiplication to break the aforementioned problem. Another advantage is that the rank of the final matrix possibly increases linearly with the number of parameters, while simply using a larger $r$ would lead to a quadratic increase in the number of parameters. In the end, looking at Table 6, we observed that the experimental result for the 6-layer decoder of the AdaMixer object detector is relatively better when using [1,1,2,2,3,3] groups of parameters per layer and $r = 16$ for each group. This slightly strange parameter group numbers is inspired by AdaLoRA [53]. It suggests that higher-level neural network parameters may + +need to allocate more ranks when performing LoRA fin-tuning, which we suppose may also hold true for LORS, and the experiment results confirmed that. + +
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 183241.8
1 1 1 1 1 1163442.1
2 2 2 2 2 2163542.4
1 1 2 2 3 3163542.6
1 1 1 1 1 1323841.2
+ +Hyperparameters of $\mathbf{LORS}^T$ . We further investigate the optimal configuration settings for $\mathbf{LORS}^T$ to achieve satisfactory performance, as shown in Table 7. We observed that the best option is to use one group of low-rank parameters with $r = 8$ for the decomposition of each layer $L_{\mathrm{output}}$ in the stacked decoder of AdaMixer, which can achieve a model performance comparable to the original detector. Since using one group of parameters already yields reasonably good results, we did not explore the use of more parameter groups. Moreover, we should also note that the value of rank $r$ is not necessarily better when larger or smaller, which may indicate that the low-rank structure required by the model for a specific task has its own particularity. + +Table 6. Validation accuracy with different groups of parameters and rank $r$ in LORSA. Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers. + +
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 123442.3
1 1 1 1 1 143441.9
1 1 1 1 1 183542.6
1 1 1 1 1 1163742.1
1 1 1 1 1 1324041.7
+ +Number of decoders. Finally, we conclude our ablation studies with experiments on the optimal number of decoder layers incorporating the LORS structure for better performance. Since our method significantly reduces the total number of model parameters and allows each decoder layer to train and use shared parameters that capture common features, it is natural to consider whether the model's inherent characteristics are affected and whether using more or fewer decoding layers would further improve the model's performance. With this idea in mind, we conducted sev + +eral experiments as shown in Table 8 to explore this issue. However, similar to the original AdaMixer without LORS, 6 layers still yield the best performance for the AdaMixer with LORS, having too many or too few layers would degrade the model's performance. This may also suggest that LORS merely captures the intrinsic structure of the model without altering its original properties. + +Table 7. Validation accuracy with different rank $r$ in LORS $^{\mathrm{T}}$ . Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers. + +
Decoder +NumberDecoder +Params(M)AP
32838.9
63542.6
94242.0
125041.0
+ +Table 8. Model performance using different numbers of stacked decoders when employing LORS. The default choice for our model is colored gray. This experiment with varying numbers of decoders demonstrates that six layers still yield the best performance for the model, which is not changed by LORS. + +# 5. Conclusion + +This research introduced a novel approach to the reduction of parameters within deep learning models utilizing stacked structures, i.e., the Low-rank Residual Structure (LORS). It is potentially an effective methodology in reducing both static and adaptive parameters. Broadly speaking, LORS allows the parameters of stacked modules to be largely shared, while maintaining only a small number of unique parameters for each module, thereby significantly reducing the total number of parameters whilst not compromising performance. We validated our method via object detection task across various extensive experimental cases on the widely-used COCO dataset, and the results demonstrate a superior model performance even with a $50\% - 70\%$ reduction in decoder parameters, surpassing our expectations. + +There are some noticeable limitations in our study. While effective in reducing model parameters without compromising performance, enhancing model capabilities requires a comparatively long training process for the method. Additionally, our method slightly increases the inference time, as even without a detailed analysis, it can be roughly seen that LORS has unparalleled and repetitive computations needing more optimization. Finally, we have only tested our approach on object detection tasks, specifically with the AdaMixer model and its decoder structure. It's clear that our method can be applied to more tasks (e.g., NLP), different models (e.g., language models), and other neural network components (e.g., backbones). These aspects will be the focus of our future work. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3 +[2] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 2, 3 +[4] Arnav Chavan, Zhuang Liu, Deepak Gupta, Eric Xing, and Zhiqiang Shen. One-for-all: Generalized lora for parameter-efficient fine-tuning, 2023. 3 +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5 +[6] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2 +[8] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems, 28, 2015. 1 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[11] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3621–3630, 2021. 3 +[12] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 3, 5, 6 +[13] Yanqi Ge, Qiang Nie, Ye Huang, Yong Liu, Chengjie Wang, Feng Zheng, Wen Li, and Lixin Duan. Beyond proto + +types: Semantic anchor regularization for better representation learning. arXiv preprint arXiv:2312.11872, 2023. 2 +[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010. 5 +[15] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 1 +[16] Song Han, Jeff Pool, John Tran, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 1 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2 +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2 +[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3 +[22] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 2 +[23] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation, 2023. 3 +[24] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 1 +[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2 +[26] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2 + +[28] Vladislav Lialin, Namrata Shivagunde, Sherin Muckatira, and Anna Rumshisky. Stack more layers differently: High-rank training through low-rank updates, 2023. 2 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3 +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7 +[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[36] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 2 +[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2 +[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[40] Farhana Sultana, Abu Sufian, and Paramartha Dutta. A review of object detection models based on convolutional neural network. Intelligent computing: image processing based applications, pages 1-16, 2020. 2 +[41] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3 + +[42] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2 +[43] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2 +[44] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261-24272, 2021. 2 +[45] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter efficient tuning of pretrained models using dynamic search-free low-rank adaptation, 2023. 2 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3 +[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2 +[48] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2 +[49] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 2 +[50] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 2 +[51] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3 +[52] Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning, 2023. 2 +[53] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning, 2023. 2, 7 +[54] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2 + +[55] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 1 +[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3 +[57] Bojia Zi, Xianbiao Qi, Lingzhi Wang, Jianan Wang, KamFai Wong, and Lei Zhang. Delta-lora: Fine-tuning high-rank parameters with the delta of low-rank matrices, 2023. 2 \ No newline at end of file diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/images.zip b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa9ae431437dc3a5d353539235ceb41497c8b1df --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:364bb6b85e48eb27233ab08add13cac5c1802e2d752e14e63fb537521f668b08 +size 294366 diff --git a/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/layout.json b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7bb2156a0bfb2503ceec3a8846d2f764400d1342 --- /dev/null +++ b/2024/LORS_ Low-rank Residual Structure for Parameter-Efficient Network Stacking/layout.json @@ -0,0 +1,9846 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 542, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 542, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 542, + 121 + ], + "type": "text", + "content": "LORS: Low-rank Residual Structure for Parameter-Efficient Network Stacking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 143, + 539, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 143, + 539, + 171 + ], + "spans": [ + { + "bbox": [ + 87, + 143, + 539, + 171 + ], + "type": "text", + "content": "Jialin Li, Qiang Nie, Weifu Fu, Yuhuan Lin, Guangpin Tao, Yong Liu, Chengjie Wang, Youtu Lab, Tencent" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 173, + 572, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 572, + 186 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 572, + 186 + ], + "type": "text", + "content": "{jarenli, ryanfwu, gleelin, guangpintao, choasliu,Jasoncjwang}@tencent.com, qnie.cuhk@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "content": "Deep learning models, particularly those based on transformers, often employ numerous stacked structures, which possess identical architectures and perform similar functions. While effective, this stacking paradigm leads to a substantial increase in the number of parameters, posing challenges for practical applications. In today's landscape of increasingly large models, stacking depth can even reach dozens, further exacerbating this issue. To mitigate this problem, we introduce LORS (LOw-rank Residual Structure). LORS allows stacked modules to share the majority of parameters, requiring a much smaller number of unique ones per module to match or even surpass the performance of using entirely distinct ones, thereby significantly reducing parameter usage. We validate our method by applying it to the stacked decoders of a query-based object detector, and conduct extensive experiments on the widely used MS COCO dataset. Experimental results demonstrate the effectiveness of our method, as even with a " + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 490 + ], + "type": "text", + "content": " reduction in the parameters of the decoder, our method still enables the model to achieve comparable or even better performance than its original." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 128, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "In the current era of prosperity for large models, a common issue is the significant increase in the number of parameters, which presents challenges for training, inference, and deployment. Various methods have been proposed to reduce the number of parameters in models, such as knowledge distillation [15, 20], which compresses large models into smaller ones while trying to preserve their performance but may still lead to a decrease in model capacity; pruning [16, 55], which removes redundant parameters from the model but can affect the model's stability; quantization [8], which reduces the numerical precision of model parameters to lower storage and computation but may cause model accuracy loss; and parameter sharing [24], which reduces the number of parameters by sharing them across different layers but may limit the model's expressiveness." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 225, + 512, + 445 + ], + "blocks": [ + { + "bbox": [ + 333, + 225, + 512, + 445 + ], + "lines": [ + { + "bbox": [ + 333, + 225, + 512, + 445 + ], + "spans": [ + { + "bbox": [ + 333, + 225, + 512, + 445 + ], + "type": "image", + "image_path": "8a1a184e916f58e96ef2fcee1384e3911359bcd46feed7acdc3ea62fd980dba5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 464, + 547, + 499 + ], + "lines": [ + { + "bbox": [ + 305, + 464, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 547, + 499 + ], + "type": "text", + "content": "Figure 1. The LORS calculation process, which could be adaptive or static, depending on whether an adaptively generated kernel is used in the matrix manipulation for private parameters." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": "Different from the aforementioned methods, we have observed an important fact contributing to the large number of parameters: the widespread use of stacking in neural networks. Stacking refers to those modules that have identical architectures and perform the same or similar functions, but possess different parameters due to random initialization as well as training updates. Examples of stacking can be found in numerous prominent neural networks, like the classic ResNet model [18] and Transformers [46]. Particularly, Transformers heavily rely on stacked structures and typically employ completely identical multi-layer stacks in both encoders and decoders. It now serves as an indispensable component of many excellent models in fields such as computer vision and natural language processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "Although stacking is powerful for enhancing model capacity, as demonstrated by large language models, it also" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15866" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "naturally leads to a sharp increase in the number of parameters. For example, GPT-3 [2] utilizes 175 billion parameters and consists of 96 layers of stacked Transformer [46] layers. How can we enjoy the benefits of stacking while reducing the required number of parameters? We notice that stacked decoders have identical structures and similar functions, indicating that there should be some commonality among their parameters. However, since they handle different input and output distributions, there must also be unique aspects to their parameters. Therefore, a natural idea is: it may be possible to represent the shared aspects with shared parameters while allowing each stacked module to retain only the parameters that capture its unique characteristics, thereby reducing the overall parameter usage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 469 + ], + "type": "text", + "content": "Based on the above considerations, we propose to decompose the parameters of stacked modules into two parts: shared ones representing the commonality and private ones capturing the specific characteristics. Shared parameters are available for all modules and trained jointly by them, while the private parameters are separately owned by each module. We suppose that the way of sharing parameters related to the commonality may reduce the number of parameters while maintaining the performance of the model. To achieve this goal, we introduce the concept of Low-rank Residual Structure (LORS), inspired by the approach of LoRA [21]. LORS essentially adds the unique parameters to the shared ones, just like residual connections add residual information to the features. While LoRA is originally designed for fine-tuning, we train our LoRA-like operation on parameters from scratch. This approach allows us to effectively reduce the overall parameter usage while maintaining the performance of the model, paving the way for more parameter-efficient network stacking." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 472, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 289, + 651 + ], + "type": "text", + "content": "To validate our idea, we choose AdaMixer [12], a strong query-based object detector, as our experimental subject. It contains a large number of both adaptive and static parameters in its stacked decoders, making it an ideal candidate to demonstrate LORS's effectiveness. The difference between adaptive and static parameters lies in whether they change with different inputs. Our goal is to show that LORS can effectively reduce the overall usage of both types of parameters while maintaining the model's performance. We conducted extensive experiments on this detector, illustrating that our method succeeded in reducing up to " + }, + { + "bbox": [ + 46, + 472, + 289, + 651 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 472, + 289, + 651 + ], + "type": "text", + "content": " of parameters in AdaMixer's decoders, while still enabling the model to achieve comparable or even superior performance compared to its vanilla version. In summary, our contributions can be concluded as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "type": "text", + "content": "- We propose a novel low-rank residual structure, named LORS, for network stacking, which can greatly reduce the number of parameters while maintaining or even improving the performance compared to the vanilla one." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "- We introduce effective methodologies for reducing both" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 106 + ], + "type": "text", + "content": "static and adaptive generated parameters in stacked structures, which makes our proposed LORS a more versatile method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 108, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 547, + 168 + ], + "type": "text", + "content": "- Our method holds the potential to serve as one of the basic network structures for large models that are affected by the issue of excessive parameters due to stacking, rendering them more parameter-efficient, thereby facilitating an easier implementation in practical applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 178, + 392, + 191 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 392, + 191 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 392, + 191 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 426 + ], + "type": "text", + "content": "Models with stacked structures. Many nerval networks partially or extensively employ stacked modules. CNN-based models [40], for instance, are widely applied across various computer vision tasks, such as classification [18, 22, 39, 49], detection [19, 25, 36-38, 43], segmentation [7, 13, 32, 54], etc. These models often use stacked smaller modules within larger components. Another kind of models primarily utilize Multi-Layer Perceptrons (MLPs), such as the MLP-Mixer [44], whose main body comprises dozens of completely identical stacked mixer layers. Moreover, since its invention, Transformer [46] has been successfully applied to diverse domains, including computer vision [3, 30, 41, 56], natural language processing [2, 10, 50], multi-modal learning [6, 27, 35, 47], etc. In the thriving arena of large language models, Transformer-based structures are indeed indispensable for many cutting-edge works [10, 26, 34, 42, 48]. Transformers are usually used to form stacked multi-layers encoders or decoders, within each layer being structurally identical." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "LoRA and Its Variants. LoRA [21] is a technique proposed for fine-tuning large language models, it introduces low-rank decomposition matrices in the Transformer [46] architecture, significantly reducing trainable parameters and GPU memory requirements. Afterward, a series of works with modifications as well as improvement have been proposed. AdaLoRA [53] adaptively allocates parameter budgets among weight matrices based on their importance scores and uses singular value decomposition for incremental updates, leading to an improved fine-tuning performance, especially in low budget settings. DyLoRA [45] introduces a dynamic low-rank adaptation technique that trains LoRA blocks for a range of ranks instead of a single rank, addressing the issues of fixed block size and rank optimization in LoRA. Delta-LoRA [57] presents a parameter-efficient approach to fine-tune large language models by updating not only the low-rank matrices but also the pre-trained weights using the delta of the product of two low-rank matrices. LoRA-FA [52] introduces a memory-efficient fine-tuning method that reduces activation memory without performance degradation or expensive recomputation, it achieves this by freezing the projection-down weight and updating the projection-up weight in each LoRA layer. ReLoRA [28] is a parameter-efficient method for training" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15867" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "large neural networks using low-rank updates. GLoRA [4], or Generalized LoRA, is an advanced approach for universal parameter-efficient fine-tuning tasks, which enhances the low rank adaptation technique by employing a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations. VeRA [23], or Vector-based Random Matrix Adaptation, further compresses the training parameters compared to the vanilla LoRA [21]. It achieves this by utilizing shared low-rank matrices across all layers and learning small scaling vectors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 301 + ], + "type": "text", + "content": "Inspired by the aforementioned LoRA series works, we propose our method called LORS (Low Rank Residual Structure) for parameter efficiency. However, our approach is fundamentally distinct from the LoRA series works. While LoRA uses different low-rank weights to fine-tune different pretrained weights, LORS adds different low-rank weights per layer to the same common weights shared by all layers, thus creating parameter-efficient stacked models trained from scratch, requiring no pretrained weights." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 309, + 115, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 115, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 115, + 323 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 329, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 289, + 462 + ], + "type": "text", + "content": "In this section, we initially revisit the foundational works upon which our methodology and experiments are primarily based, encompassing the LoRA mechanism [21], query-based object detection [3], and the structure of AdaMixer's decoders [12]. Then we mathematically formulate our proposed LORS, which comprises two parts: the adaptive one and the static one. Subsequently, we elaborate on how the LORS method is applied to reduce the parameters of AdaMixer's decoders. Lastly, we provide a quantitative analysis of the parameter reduction percentage attained by our proposed method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 467, + 127, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 127, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 127, + 479 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 485, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 556 + ], + "type": "text", + "content": "The mechanism of LoRA. The Low-Rank Adaptation (LoRA) [21] technique is a novel approach designed to adapt large pre-trained language models for specific tasks. The key idea of LoRA is to introduce a low-rank parameter matrix which is able to captures task-specific knowledge while the original pre-trained parameters remain fixed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": "Mathematically, given a pre-trained parameter matrix " + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": ", LoRA uses a low-rank matrix " + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": " and a projection matrix " + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times h}" + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": " to adapt " + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "inline_equation", + "content": "r \\ll d, h" + }, + { + "bbox": [ + 46, + 557, + 287, + 605 + ], + "type": "text", + "content": ". The adapted parameter matrix is then given by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 611, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 611, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 118, + 611, + 287, + 624 + ], + "type": "interline_equation", + "content": "W + \\Delta W = W + B A \\tag {1}", + "image_path": "01ef9afae53a249a8b2f2f4e3184f8a92976f470cdbf651132eac46a14a692d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 630, + 244, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 244, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 244, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 630, + 244, + 641 + ], + "type": "inline_equation", + "content": "BA" + }, + { + "bbox": [ + 47, + 630, + 244, + 641 + ], + "type": "text", + "content": " captures the task-specific knowledge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "The key advantage of LoRA is that it can significantly reduce the number of parameters that need to be fine-tuned, thereby reducing the computational cost and lowering down the memory requirement. In some cases, even single-digit values of " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " are sufficient to fine-tune the model to the desired state, which is often tens of times less expensive than" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "training the parameters in " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": ". Furthermore, by keeping the original parameters fixed, LoRA avoids catastrophic forgetting, a common issue in fine-tuning large models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 192 + ], + "type": "text", + "content": "Query-based object detection. In the realm of object detection, query-based detectors have established a new paradigm [11, 12, 30, 41, 51, 56]. Unlike traditional detectors which rely on anchor boxes or sliding windows, query-based models utilize a set of learnable queries to interact with image feature maps. This interaction can be formalized with attention [46] operations as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 360, + 200, + 545, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 200, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 360, + 200, + 545, + 213 + ], + "type": "interline_equation", + "content": "Q _ {\\text {u p d a t e d}} = \\operatorname {A t t e n t i o n} (Q, K (V)) \\tag {2}", + "image_path": "3ea3bc7fc46842f4085e05a955329a86e2ced976eeaad2dfd5a9f58b3541ce4d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " represent the queries, keys, and values. The learnable queries " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " are used to predict object classes and bounding boxes ultimately, while " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " typically originate from encoded image features. It is a common practice to continuously refine " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " through interactions with " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 220, + 545, + 303 + ], + "type": "text", + "content": " using successive decoding layers. These layers are usually composed of structurally identical decoders." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 304, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 361 + ], + "type": "text", + "content": "Decoders of AdaMixer. AdaMixer [12] is a query-based detector that features an innovative decoder design. This design includes Adaptive Channel Mixing (ACM) and Adaptive Spatial Mixing (ASM) methods, which greatly enhance its performance [12]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": "Given a sampled feature " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{P_{\\mathrm{in}} \\times C}" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "C = d_{\\mathrm{feat}} / g" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": " denotes the number of sampling groups. This sampled feature is obtained through an operation called group sampling. This operation divides the feature space channel " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{feat}}" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": " groups and performs individual sampling for each group. Then, ACM (Adaptive Channel Mixing) utilizes a weight adapted by the object query " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": " to transform feature " + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 363, + 545, + 471 + ], + "type": "text", + "content": " in the channel dimension, enhancing channel semantics [12]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 477, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 477, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 367, + 477, + 545, + 491 + ], + "type": "interline_equation", + "content": "M _ {c} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {C \\times C} \\tag {3}", + "image_path": "d88168e4db3aa26d99f4b41c1f381c4f0b67151082696e0fac76784ce10c2221.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 494, + 545, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 494, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 344, + 494, + 545, + 506 + ], + "type": "interline_equation", + "content": "\\operatorname {A C M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} M _ {c}\\right)\\right) \\tag {4}", + "image_path": "ac2dccc0894fe299132e8ad638962e2b6bb8187404c6670327e8592671a7c5ab.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 514, + 526, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 526, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 526, + 525 + ], + "type": "text", + "content": "where LayerNorm stands for Layer Normalization [1]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "content": "Next is the ASM (Adaptive Spatial Mixing) process, which aims to enable the adaptability of the object query " + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 304, + 525, + 545, + 574 + ], + "type": "text", + "content": " to spatial structures of sampled features [12] by applying adaptive transformation to the spatial dimension:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 362, + 580, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 580, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 362, + 580, + 545, + 594 + ], + "type": "interline_equation", + "content": "M _ {s} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {P _ {\\text {i n}} \\times P _ {\\text {o u t}}} \\tag {5}", + "image_path": "b952c0cf125bee392d967bcdf098ea18a7b7361704eadc8596cfc50118f42195.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 597, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 597, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 340, + 597, + 545, + 610 + ], + "type": "interline_equation", + "content": "\\operatorname {A S M} (\\mathbf {x}) = \\operatorname {R e L U} \\left(\\operatorname {L a y e r N o r m} \\left(\\mathbf {x} ^ {T} M _ {s}\\right)\\right), \\tag {6}", + "image_path": "cd9aa60f21309ad650d69ccda6b4e97dc638dfa92b2787d90b252e84eb891096.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": "Both ACM and ASM train independent parameters for each sampling group, and finally the output with the shape " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{g\\times C\\times P_{\\mathrm{out}}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " is flattened and transformed to the " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "d_q" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " dimension by a linear layer " + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 666 + ], + "type": "text", + "content": " to add back to the object query." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "ACM, ASM and the output linear transformation " + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{output}" + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": " possess significantly more parameters compared to the decoder's other operations, making them the main contributors to the number of parameters. Hence, we choose them" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15868" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 96 + ], + "type": "text", + "content": "as the target components for validating the effectiveness of our LORS method in parameter reduction." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 47, + 114, + 272, + 220 + ], + "blocks": [ + { + "bbox": [ + 47, + 114, + 272, + 220 + ], + "lines": [ + { + "bbox": [ + 47, + 114, + 272, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 272, + 220 + ], + "type": "text", + "content": "def non_dynamic_fused_weight(W_shared, A, B):\n # W_shared: (1, d, h)\n # A: (G, r, h)\n # B: (G, d, r)\n # private weight: (G, d, h)\n W_private = B @ A\n # sum at group dimension to get fused weight: (d, h)\n W = concat([W_shared, W_private], dim=0).sum(axis=0)\n return W" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 46, + 225, + 288, + 249 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 288, + 249 + ], + "type": "text", + "content": "Figure 2. Pseudo-code for obtaining a static weight parameter for one layer." + } + ] + } + ], + "index": 2, + "type": "text" + }, + { + "bbox": [ + 47, + 282, + 202, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 202, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 202, + 295 + ], + "type": "text", + "content": "3.2. Formulation of Our Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 301, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 386 + ], + "type": "text", + "content": "The complete computational process of LORS is illustrated in Figure 1. Notably, LORS computation is divided into two types: adaptive and static. In the context of AdaMixer, the term \"adaptive\" indicates whether the transformation matrix depends on the object query. After completing the formalization of LORS here, we will further elaborate on how it is applied to the tasks in our setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "content": "Static Low Rank Residual Structure " + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "inline_equation", + "content": "(\\mathbf{LORST}^{\\mathrm{T}})" + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "content": ". Suppose there are " + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "content": " stacked layers of modules with identical architecture, and " + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "inline_equation", + "content": "W_{i}\\in \\mathbb{R}^{d\\times h}" + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "content": " be a parameter matrix belonging to the " + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 387, + 287, + 436 + ], + "type": "text", + "content": "-th layer. Then, we have:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 445, + 287, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 445, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 116, + 445, + 287, + 461 + ], + "type": "interline_equation", + "content": "W _ {i} = W ^ {\\text {s h a r e d}} + W _ {i} ^ {\\text {p r i v a t e}} \\tag {7}", + "image_path": "85f2773ac4d048527f6d65e3d294dc62b122338c6e7751c0ea3d8fe4afe83789.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{shared}} \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "text", + "content": " represents the shared parameters across all stacked layers, while " + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "W_{i}^{\\mathrm{private}}" + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "text", + "content": " denotes the layer-specific parameters for the " + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 472, + 287, + 519 + ], + "type": "text", + "content": "-th layer, which is calculated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 521, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 521, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 120, + 521, + 287, + 554 + ], + "type": "interline_equation", + "content": "W _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} B _ {i k} A _ {i k} \\tag {8}", + "image_path": "def784d7e50394534aedbd32b326fb21879960170c543b74363bcfca6c953ea1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "B_{ik} \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "A_{ik} \\in \\mathbb{R}^{r \\times h}" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "r \\ll d, h" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": " represents the number of parameter groups used to compute " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "W_{i}^{\\mathrm{private}}" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": ". Pseudo-code of LORST computing " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "W_{i}^{\\mathrm{private}}" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": " for a specific layer " + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 563, + 287, + 613 + ], + "type": "text", + "content": " can be seen in Figure 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "text", + "content": "Adaptive Low Rank Residual Structure (LORSA). Let " + }, + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "inline_equation", + "content": "\\hat{W}_i \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "text", + "content": " be an adaptive generated parameter in " + }, + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 614, + 287, + 651 + ], + "type": "text", + "content": "-th stacked layer, it is similarly calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 661, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 661, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 116, + 661, + 287, + 677 + ], + "type": "interline_equation", + "content": "\\hat {W} _ {i} = \\hat {W} ^ {\\text {s h a r e d}} + \\hat {W} _ {i} ^ {\\text {p r i v a t e}} \\tag {9}", + "image_path": "60d3dc77bea7fcb1bc717906100a7efda0c5077f1dc9b6c605a7fbdcbb0ddb7b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "where the cross-layer shared parameter " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{W}^{\\mathrm{shared}}\\in \\mathbb{R}^{d\\times h}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " and the layer-specific parameter " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{W}_i^{\\mathrm{private}}\\in \\mathbb{R}^{d\\times h}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " for each layer" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 306, + 74, + 544, + 215 + ], + "blocks": [ + { + "bbox": [ + 306, + 74, + 544, + 215 + ], + "lines": [ + { + "bbox": [ + 306, + 74, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 74, + 544, + 215 + ], + "type": "text", + "content": "def dynamic_fused_weight(q, A, B):\n # q: (N, C)\n # A: (1, G, r, h)\n # B: (1, G, d, r)\n # shared weight: (N, 1, d, h)\n W_shared = linear1(pro_feats).reshape(N, 1, d, h)\n # E: (N, G, r, r)\n E = linear2(pro_feats).reshape(N, G, r, r)\n # private weight: (N, G, d, h)\n W_private = B @ E @ A\n # sum at group dimension to get fused weight: (N, d, h)\n W = concat([W_shared, W_private], dim=1).sum(axis=1)\n return W" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "type": "text", + "content": "Figure 3. Pseudo-code for obtaining an adaptive weight parameter for one layer." + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 306, + 274, + 432, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 274, + 432, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 432, + 285 + ], + "type": "text", + "content": "are both calculated based on " + }, + { + "bbox": [ + 306, + 274, + 432, + 285 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 306, + 274, + 432, + 285 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 294, + 545, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 294, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 365, + 294, + 545, + 309 + ], + "type": "interline_equation", + "content": "\\hat {W} ^ {\\text {s h a r e d}} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {d \\times h} \\tag {10}", + "image_path": "d64b2e66b9dbd5d73240f0a7e6706351e9c8e142ca0e4d78d6e7de671796f479.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 312, + 545, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 312, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 364, + 312, + 545, + 345 + ], + "type": "interline_equation", + "content": "\\hat {W} _ {i} ^ {\\text {p r i v a t e}} = \\sum_ {k = 1} ^ {K} \\hat {B} _ {i k} \\hat {E} _ {i k} \\hat {A} _ {i k} \\tag {11}", + "image_path": "e0bfbcfd29fde0ff31fc15aba302de9fb81bae5f679a0af5fe6458d4d687a4e7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 347, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 347, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 380, + 347, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\hat {E} _ {i k} = \\operatorname {L i n e a r} (\\mathbf {q}) \\in \\mathbb {R} ^ {r \\times r} \\tag {12}", + "image_path": "3c88d06a892fd892234c418f658c7d69365e978447163195bb4f96abef76b4c1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\hat{B}_{ik} \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\hat{A}_{ik} \\in \\mathbb{R}^{r \\times h}" + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "inline_equation", + "content": "r \\ll d, h" + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": ". Pseudocode of LORSA calculating " + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\hat{W}_i^{\\mathrm{private}}" + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": " for a specific layer " + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 372, + 545, + 410 + ], + "type": "text", + "content": " can be seen in Figure 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 418, + 520, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 520, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 520, + 431 + ], + "type": "text", + "content": "3.3. Applying LORS to AdaMixer's Decoders" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "text", + "content": "We apply LORS to the parameters of linear transformations belonging to ACM, ASM, and " + }, + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 305, + 437, + 545, + 473 + ], + "type": "text", + "content": " in each of AdaMixer's decoders. These parameters are explained in 3.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": "The overall pipeline of LORS running in AdaMixer is illustrated in Figure 4. Specifically, for each group of sampling points, " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{LORS^A}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " (see 3.2) is used to reduce parameters in " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_q}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times C}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": ") of ACM and " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " of ASM (from " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_q}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{P_{\\mathrm{in}}\\times P_{\\mathrm{out}}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{LORS^T}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " (see 3.2) to minimize parameters in " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " (from " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times P_{\\mathrm{out}}}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_q}" + }, + { + "bbox": [ + 305, + 474, + 545, + 545 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Thus, the parameter quantities of " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "d_{q} \\times C \\times C" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "d_{q} \\times P_{\\mathrm{in}} \\times P_{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "d_{q} \\times C \\times P_{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " respectively. When the sampling strategy consists of 2 groups with 64 points each, which is the default setting in our experiments, the values of the variables are as follows: " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "d_{q} = 256" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "C = 64" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{in}} = 64" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{out}} = 128" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ". It can be easily calculated that the number of parameters for each of " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "M_{s}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " exceeds one million. Indeed, these three components collectively account for the most of the total parameter in the AdaMixer model with ResNet-50 as the backbone, whereas they are also the primary drivers of enhanced model performance. This is what motivates us to conduct LORS experiments on them. The effects of LORS applied to them will be discussed in detail in 3.4." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15869" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 80, + 282, + 228 + ], + "blocks": [ + { + "bbox": [ + 52, + 80, + 282, + 228 + ], + "lines": [ + { + "bbox": [ + 52, + 80, + 282, + 228 + ], + "spans": [ + { + "bbox": [ + 52, + 80, + 282, + 228 + ], + "type": "image", + "image_path": "184c299ee6cabe449fbc78238f2b4aaab9fdddd7a653486f7d6f51c21a6cceef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 288, + 321 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 321 + ], + "type": "text", + "content": "Figure 4. The overall pipeline of our proposed LORS, consisting of both adaptive and static parts, each further composed of shared and private components, works collaboratively. The figure illustrates the entire computation process within one layer of the stacked layers, with an enlarged example of the " + }, + { + "bbox": [ + 46, + 255, + 288, + 321 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 255, + 288, + 321 + ], + "type": "text", + "content": "-th layer for demonstration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 346, + 227, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 227, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 227, + 358 + ], + "type": "text", + "content": "3.4. Analysis on Parameter Reduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " be a weight parameter that exists in every layer of stacked structures, " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " is the number of stacked layers. If static, it originally has " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d \\times h" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " parameters, while using LORS" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " requires " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\frac{1}{N} \\times d \\times h + K \\times (d \\times r + r \\times h)" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " parameters on average per layer; if it is adaptive, generating it by " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " with linear transformation requires " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d_{q} \\times d \\times h" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " parameters, where " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d_{q}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", and using LORS" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "^{\\mathrm{A}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " requires " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "\\frac{1}{N} \\times d_{q} \\times d \\times h + K \\times (d_{q} \\times r^{2} + d \\times r + r \\times h)" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " parameters on average per layer. To more intuitively display the parameter reduction effect of LORS, we set " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d_{q} = 256" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d = 64" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "h = 128" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " for the ASM process and " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "d = 2 \\times 128 \\times 128" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "h = 256" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "L_{output}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": ", which is actually the case of AdaMixer's decoders in our experiment, and show the parameter reduction for different " + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " values using LORS" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " and LORS" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "inline_equation", + "content": "^{\\mathrm{A}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 546 + ], + "type": "text", + "content": " in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 558, + 278, + 657 + ], + "blocks": [ + { + "bbox": [ + 56, + 558, + 278, + 657 + ], + "lines": [ + { + "bbox": [ + 56, + 558, + 278, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 278, + 657 + ], + "type": "table", + "html": "
r=4r=8r=16r=32
W/. LORST1.53M1.66M1.93M2.46M
W/O. LORST8.39M8.39M8.39M8.39M
Percentage18.3%19.8%23.0%29.3%
W/. LORSA0.36M0.39M0.49M0.89M
W/O. LORSA2.10M2.10M2.10M2.10M
Percentage17.2%18.6%23.3%42.4%
", + "image_path": "bd024120ee4618a43bcf469ebd01fe2c23b3f124f9e90e0b582335cb3bf9ca33.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "text", + "content": "Table 1. Analysis of the parameter reduction effect of LORS" + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "text", + "content": " and LORS" + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "inline_equation", + "content": "^{\\mathrm{A}}" + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "text", + "content": " with varying rank " + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 666, + 287, + 710 + ], + "type": "text", + "content": " values. Our default selections are colored gray. Here the average amount of shared parameters over each layer has already been taken into account." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 387, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": "In this section, we first explain the implementation details of our experiments, including the dataset, training specifics, and different parameter initialization methods for static and adaptive LORS. Next, we present the main experimental results. Finally, we perform ablation studies to examine LORS's design and its individual components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 173, + 440, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 173, + 440, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 173, + 440, + 186 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 192, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 192, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 192, + 545, + 252 + ], + "type": "text", + "content": "Dataset. Our experiments were conducted on the widely-used MS COCO [29] dataset in mmdetection codebase [5], using the standard metrics for object detection. All models were trained on the train2017 split (" + }, + { + "bbox": [ + 304, + 192, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\sim 118\\mathrm{k}" + }, + { + "bbox": [ + 304, + 192, + 545, + 252 + ], + "type": "text", + "content": " images) and then evaluated on the val2017 split (5k images)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": "Training. We use the AdamW optimizer [33] with a weight decay of 0.0001 and train all models on 8 Nvidia V100 GPUs with a batch size of 16 and a learning rate of " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "2.5 \\times e^{-5}" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": ". Models are trained for either 12 or 36 epochs, with learning rate decreasing by a factor of 10 at epochs 8 and 11 for 12-epoch training, and at epochs 24 and 33 for 36-epoch training. The low-rank values are set as " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " for LORS" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "^{\\text{A}}" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "r = 8" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " for LORS" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "^{\\text{T}}" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": ". The number of parameter groups is set as as " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "K = [1,1,2,2,3,3]" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " for LORS" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "^{\\text{A}}" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " in all experiments, applied to ACM and ASM in AdaMixer's decoders, and " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "K = [1,1,1,1,1,1]" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "inline_equation", + "content": "L_{output}" + }, + { + "bbox": [ + 304, + 253, + 546, + 516 + ], + "type": "text", + "content": " in all experiments. We divide the feature channels into 2 groups with 64 sampling points each instead of AdaMixer's default 4 groups with 32 sampling points each, aiming to increase the parameter compressible space for LORS, which does not improve performance according to both AdaMixer paper [12] and our experiments. Backbones are initialized with ImageNet-1k [9] pre-trained models, and LORS-related parameter initialization is detailed below, remaining parameters are Xavier-initialized [14]. Finally, all other aspects about model training, like the data augmentation, the loss function, etc., just follow AdaMixer's settings [12]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 552 + ], + "type": "text", + "content": "Initialization Strategies We tried various initialization methods for each component in LORS and determined the overall initialization method as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 552, + 544, + 648 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "text", + "content": "- LORST: For static LORS, we employ Kaiming initialization [17] for " + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "inline_equation", + "content": "W^{\\text{shared}}" + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "text", + "content": " and each " + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "text", + "content": ", and zero initialization for each " + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 314, + 552, + 544, + 587 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "spans": [ + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "content": "- LORSA: For adaptive LORS, we apply Kaiming initialization [17] to the linear transformation weights forming each " + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "inline_equation", + "content": "\\hat{W}^{\\mathrm{shared}}" + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "content": ", as well as each " + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "inline_equation", + "content": "\\hat{B}" + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "inline_equation", + "content": "\\hat{A}" + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "content": ". Additionally, we use zero initialization for the linear transformation weights forming each " + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 314, + 589, + 544, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "To reliably demonstrate that LORS can reduce the number of parameters in stacked structures while maintaining model performance, which is measured by AP metric, we follow" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15870" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 141 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 141 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 141 + ], + "type": "table", + "html": "
MethodQueriesEpochsDecoder Params(M)Params(M)Decoder GFLOPsGFLOPsTraining HoursAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixer10012110135121048.5h42.761.545.924.745.459.2
AdaMixer + LORS100123560181109.5h42.661.446.025.045.658.5
AdaMixer300121131393913210h44.163.447.427.046.959.5
AdaMixer + LORS3001235605614912h44.163.047.727.847.059.5
", + "image_path": "c9cc7f34f32d12b518fed23534c42d900b31db7fc57f9580cadc37ad748caa04.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 184, + 545, + 312 + ], + "blocks": [ + { + "bbox": [ + 47, + 150, + 546, + 174 + ], + "lines": [ + { + "bbox": [ + 47, + 150, + 546, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 150, + 546, + 174 + ], + "type": "text", + "content": "Table 2. " + }, + { + "bbox": [ + 47, + 150, + 546, + 174 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 47, + 150, + 546, + 174 + ], + "type": "text", + "content": " training scheme performance on COCO 2017 val set with ResNet-50 as backbone. AdaMixer with LORS can achieve competitive results while employing a notably reduced number of parameters. FPS is obtained with a single Nvidia V100 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 184, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 50, + 184, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 50, + 184, + 545, + 312 + ], + "type": "table", + "html": "
MethodbackboneQueriesEpochsParams(M)GFLOPsAP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_s \\)\\( AP_m \\)\\( AP_l \\)
AdaMixerR-501003613510443.261.846.725.046.158.8
AdaMixer + LORSR-50100366011043.762.347.325.546.460.0
AdaMixerR-503003613913247.066.051.130.150.261.8
AdaMixer + LORSR-50300366014947.666.652.031.150.262.5
AdaMixerR-1013003615820848.067.052.430.051.263.7
AdaMixer + LORSR-101300367922548.267.552.631.751.363.8
AdaMixerSwin-S3003616423451.371.255.734.254.667.3
AdaMixer + LORSSwin-S300368525051.871.656.435.455.068.4
", + "image_path": "f33e3ad7f24700907c448ba121c9a67b960186f03567a5acf1ddec3a9265d9d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 546, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 354 + ], + "type": "text", + "content": "Table 3. " + }, + { + "bbox": [ + 46, + 319, + 546, + 354 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 46, + 319, + 546, + 354 + ], + "type": "text", + "content": " training scheme performance on COCO 2017 val set, considering different combinations of backbone and query numbers. Longer training time allows LORS to be fully trained and perform better. The comprehensive improvement in performance and the significant reduction in parameters demonstrate the effectiveness of LORS. The best results in each category are highlighted in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 447 + ], + "type": "text", + "content": "common practices in the object detection field and present experimental comparisons in two categories based on training schemes. The first involves training for 12 epochs using a weaker data augmentation, and the second involves training for 36 epochs using a stronger data augmentation, both of which are identical to those in AdaMixer [12]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": "Table 2 presents the comparison of performance with and without the LORS technique, under a " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "1\\times" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " training scheme. Metrics such as the number of parameters, GFLOPs, and Average Precision (AP) at various scales are evaluated. When the AdaMixer model is combined with LORS, it demonstrates a remarkable reduction in the number of parameters. This reduction is consistently observed across various query quantities and different training epochs, encompassing both the decoder and overall model. For example, when trained with 100 queries for 12 epochs, the AdaMixer+LORS model uses only 35M decoder parameters and 60M total parameters, compared to 110M and 135M respectively in the AdaMixer model. This shows reduction of approximately " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " in the decoder's parameters, which is substantial. However, the AdaMixer+LORS model shows a slight increase in GFLOPs. In terms of model performance, the AdaMixer+LORS model achieves competitive results, even with a significantly reduced number of parameters. When trained with 100 queries, although the AP of 42.6 is slightly lower than the 42.7 AP of the vanilla AdaMixer, it still slightly outperforms in " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "AP_{s}" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "AP_{m}" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": ". When trained with 300 queries, the use of LORS does not" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "text", + "content": "affect the performance and allows an advantage in all " + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "inline_equation", + "content": "AP_{s}" + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "inline_equation", + "content": "AP_{m}" + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "inline_equation", + "content": "AP_{l}" + }, + { + "bbox": [ + 304, + 374, + 546, + 483 + ], + "type": "text", + "content": ". These findings suggest that even under conditions of limited model training and data augmentation, LORS does not impact training result, the model can still converge quickly to its original limit. Additionally, we also demonstrate the slight impact of the current LORS method on inference speed, which might be improved by reducing the serial and redundant computations within LORS but not the focus of this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "text", + "content": "Table 3 showcases the remarkable performance of the AdaMixer + LORS method under the " + }, + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 484, + 547, + 581 + ], + "type": "text", + "content": " training scheme with different backbones and query numbers. It can be observed that the proposed method consistently outperforms the vanilla AdaMixer across all backbones, query numbers, and evaluation metrics. This result is somewhat surprising to us, as the LORS enables the model to use significantly fewer parameters during both training and inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Specifically, when adopting the ResNet-50 backbone and 100 queries, LORS improves the AP value of AdaMixer by 0.5 points (43.7 vs 43.2). It is worth noting that in Table 2, for experiments with the same configuration except for training 12 epochs, the AP with LORS is 0.1 lower than without LORS. However, this situation is reversed when training for 36 epochs. The longer training time under the " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " training scheme allows LORS to be fully trained and perform better. This emphasizes the importance of sufficient training time for the LORS technique to fully exploit its potential in enhancing the performance. Furthermore," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15871" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 167 + ], + "type": "text", + "content": "when using more powerful backbones and a larger number of queries, the performance metrics of models utilizing LORS consistently and comprehensively surpass those of their counterparts without LORS. Even when employing a strong backbone such as Swin-S [31] and using 300 queries, LORS still manages to further improve the model's performance while reducing nearly half of the overall parameters. Much fewer parameters, yet improved performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 290, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 290, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 290, + 350 + ], + "type": "text", + "content": "In conclusion, the results in Table 2 and Table 3 show the effectiveness of the LORS method in reducing the number of parameters while achieving even better performance. The method's adaptability to various backbones and query numbers further demonstrates its versatility and potential for broader applications. As for why LORS can achieve better performance with fewer parameters, a plausible explanation is that after the model has been sufficiently trained and stabilized, the parameter matrix within the stacked structure is inherently sparse and low-rank. Without using LORS, the model needs to learn such a low-rank structure through training. In contrast, LORS directly and explicitly makes the structure sparse, which, to some extent, can be considered as adding a form of regularization to the model training, thereby facilitating the model's learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 357, + 141, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 141, + 370 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 141, + 370 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "content": "To save computational resources, all ablation studies are conducted using a ResNet-50 backbone and a " + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 376, + 288, + 411 + ], + "type": "text", + "content": " training scheme." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 90, + 421, + 246, + 495 + ], + "blocks": [ + { + "bbox": [ + 90, + 421, + 246, + 495 + ], + "lines": [ + { + "bbox": [ + 90, + 421, + 246, + 495 + ], + "spans": [ + { + "bbox": [ + 90, + 421, + 246, + 495 + ], + "type": "table", + "html": "
LORSALORSTDecoder Params(M)AP
11042.5†
✓7942.6
✓6642.6
✓✓3542.6
", + "image_path": "e59e6dde58587edf92afeb52a50a330b5b957afb99f8ec0e6bfd84dd80b7580e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "Adaptive & Static LORS. We first conduct ablation studies on the impact of " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LORSA}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LORST}" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " on model parameters and performance, as shown in Table 4. In this table, a " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " indicates that we use the LORS method to replace the corresponding component in the AdaMixer model mentioned earlier, while a blank cell means we use the original AdaMixer module. If neither of the two is present, the model is identical to the original AdaMixer. As can be seen from the table, both adaptive and static LORS can effectively reduce the parameter size of the decoder in the model (31M and 44M, respectively) without affecting the model performance. Note that the symbol " + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": " here represents that the 42.5 AP value is reproduced by us, which is slightly dif" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "ferent from the 42.7 AP recorded in the original paper. We believe this discrepancy can be attributed to the randomness of 1x training and does not affect our conclusions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": "Shared and private weights. The second point of interest is to determine which of the shared weight and private weight has a greater impact on the decoder's performance. Experiment in Table 5 is designed to address this question, which is performed on ACM and ASM of AdaMixer's decoder. Recall that the formula we use to generate the final weight parameters is " + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "inline_equation", + "content": "W = W^{\\text{shared}} + W^{\\text{private}}" + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": ". In the table, a " + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": " indicates that the corresponding term is present on the right side of the formula, otherwise, it is removed. It can be seen that the absence of either term will cause a nonnegligible decrease in model performance, and the impact of missing " + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "inline_equation", + "content": "W^{\\text{private}}" + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": " is even more severe. This suggests that although the number of parameters in each " + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "inline_equation", + "content": "W^{\\text{private}}" + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": " is smaller than that of " + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "inline_equation", + "content": "W^{\\text{shared}}" + }, + { + "bbox": [ + 304, + 109, + 546, + 300 + ], + "type": "text", + "content": ", it could be more important, possibly because the most informative part of each layer in the stacked structure is contained within it." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 349, + 308, + 504, + 374 + ], + "blocks": [ + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "content": "Table 4. Effect of LORSA and LORST. †” denotes this AP result was reproduced by ourselves. Both LORST and LORSA can reduce parameters without compromising performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 308, + 504, + 374 + ], + "lines": [ + { + "bbox": [ + 349, + 308, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 349, + 308, + 504, + 374 + ], + "type": "table", + "html": "
WsharedWprivateDecoder Params(M)AP
✓5840.9
✓4341.4
✓✓3542.6
", + "image_path": "4ddaf7610aa9336fd39cf0fe4ba2d5b06f114d6ec1f2e5c4632a18b3d9c13f99.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "type": "text", + "content": "Table 5. Effect of shared parameters " + }, + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "type": "inline_equation", + "content": "W^{\\text{shared}}" + }, + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "type": "text", + "content": " and private parameters " + }, + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "type": "inline_equation", + "content": "W^{\\text{private}}" + }, + { + "bbox": [ + 305, + 382, + 545, + 406 + ], + "type": "text", + "content": " Default choice for our model is colored gray." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Hyperparameters of " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{LORS}^A" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": ". Another aspect we would like to explore is the optimal number of parameter groups and the value of rank " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " for adaptive LORS. Here, we feel it is necessary to explain why we use grouped parameters. Recalling formula 11, the adaptive private parameters are obtained by continuous multiplication of three matrices, so the matrix with the smallest rank among them determines the rank of multiplication, which is the upper limit for the private parameters. If the task-specific requirement for the rank of the private parameters matrix unfortunately exceeds the rank of the smallest component, there will be a bottleneck when training the model. Therefore, we designed this parameter group mechanism, using the sum of the results of multiple matrix multiplication to break the aforementioned problem. Another advantage is that the rank of the final matrix possibly increases linearly with the number of parameters, while simply using a larger " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " would lead to a quadratic increase in the number of parameters. In the end, looking at Table 6, we observed that the experimental result for the 6-layer decoder of the AdaMixer object detector is relatively better when using [1,1,2,2,3,3] groups of parameters per layer and " + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "inline_equation", + "content": "r = 16" + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": " for each group. This slightly strange parameter group numbers is inspired by AdaLoRA [53]. It suggests that higher-level neural network parameters may" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15872" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 107 + ], + "type": "text", + "content": "need to allocate more ranks when performing LoRA fin-tuning, which we suppose may also hold true for LORS, and the experiment results confirmed that." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 117, + 247, + 201 + ], + "blocks": [ + { + "bbox": [ + 87, + 117, + 247, + 201 + ], + "lines": [ + { + "bbox": [ + 87, + 117, + 247, + 201 + ], + "spans": [ + { + "bbox": [ + 87, + 117, + 247, + 201 + ], + "type": "table", + "html": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 183241.8
1 1 1 1 1 1163442.1
2 2 2 2 2 2163542.4
1 1 2 2 3 3163542.6
1 1 1 1 1 1323841.2
", + "image_path": "b99035346ec685d9d40abd3b2294276f83a36fb1380cbc8b980a078cddf3ac41.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": "Hyperparameters of " + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{LORS}^T" + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": ". We further investigate the optimal configuration settings for " + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{LORS}^T" + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": " to achieve satisfactory performance, as shown in Table 7. We observed that the best option is to use one group of low-rank parameters with " + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "inline_equation", + "content": "r = 8" + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": " for the decomposition of each layer " + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{output}}" + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": " in the stacked decoder of AdaMixer, which can achieve a model performance comparable to the original detector. Since using one group of parameters already yields reasonably good results, we did not explore the use of more parameter groups. Moreover, we should also note that the value of rank " + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 48, + 270, + 286, + 427 + ], + "type": "text", + "content": " is not necessarily better when larger or smaller, which may indicate that the low-rank structure required by the model for a specific task has its own particularity." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 85, + 437, + 249, + 523 + ], + "blocks": [ + { + "bbox": [ + 48, + 210, + 286, + 254 + ], + "lines": [ + { + "bbox": [ + 48, + 210, + 286, + 254 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 286, + 254 + ], + "type": "text", + "content": "Table 6. Validation accuracy with different groups of parameters and rank " + }, + { + "bbox": [ + 48, + 210, + 286, + 254 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 48, + 210, + 286, + 254 + ], + "type": "text", + "content": " in LORSA. Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 437, + 249, + 523 + ], + "lines": [ + { + "bbox": [ + 85, + 437, + 249, + 523 + ], + "spans": [ + { + "bbox": [ + 85, + 437, + 249, + 523 + ], + "type": "table", + "html": "
Parameter GroupsRank rDecoder Params(M)AP
1 1 1 1 1 123442.3
1 1 1 1 1 143441.9
1 1 1 1 1 183542.6
1 1 1 1 1 1163742.1
1 1 1 1 1 1324041.7
", + "image_path": "eb4cfef9c0ee045c7d76c7897f30a50eff0966e67e0b4764389ed3ea70bd7c5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 594, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 594, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 594, + 286, + 712 + ], + "type": "text", + "content": "Number of decoders. Finally, we conclude our ablation studies with experiments on the optimal number of decoder layers incorporating the LORS structure for better performance. Since our method significantly reduces the total number of model parameters and allows each decoder layer to train and use shared parameters that capture common features, it is natural to consider whether the model's inherent characteristics are affected and whether using more or fewer decoding layers would further improve the model's performance. With this idea in mind, we conducted sev" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 156 + ], + "type": "text", + "content": "eral experiments as shown in Table 8 to explore this issue. However, similar to the original AdaMixer without LORS, 6 layers still yield the best performance for the AdaMixer with LORS, having too many or too few layers would degrade the model's performance. This may also suggest that LORS merely captures the intrinsic structure of the model without altering its original properties." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 365, + 171, + 487, + 248 + ], + "blocks": [ + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "lines": [ + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "text", + "content": "Table 7. Validation accuracy with different rank " + }, + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "text", + "content": " in LORS" + }, + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "inline_equation", + "content": "^{\\mathrm{T}}" + }, + { + "bbox": [ + 48, + 532, + 286, + 575 + ], + "type": "text", + "content": ". Default choice for our model is colored gray. Each row's first six numbers indicate the number of parameter groups sequentially used in each of the six decoder layers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 171, + 487, + 248 + ], + "lines": [ + { + "bbox": [ + 365, + 171, + 487, + 248 + ], + "spans": [ + { + "bbox": [ + 365, + 171, + 487, + 248 + ], + "type": "table", + "html": "
Decoder\nNumberDecoder\nParams(M)AP
32838.9
63542.6
94242.0
125041.0
", + "image_path": "df9a31d9e8e3b1feb4a49f53c960a5abe2d318884346f447292312bd5b4c494f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 257, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 257, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 257, + 545, + 312 + ], + "type": "text", + "content": "Table 8. Model performance using different numbers of stacked decoders when employing LORS. The default choice for our model is colored gray. This experiment with varying numbers of decoders demonstrates that six layers still yield the best performance for the model, which is not changed by LORS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 308, + 353, + 377, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 377, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 377, + 365 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 375, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 543 + ], + "type": "text", + "content": "This research introduced a novel approach to the reduction of parameters within deep learning models utilizing stacked structures, i.e., the Low-rank Residual Structure (LORS). It is potentially an effective methodology in reducing both static and adaptive parameters. Broadly speaking, LORS allows the parameters of stacked modules to be largely shared, while maintaining only a small number of unique parameters for each module, thereby significantly reducing the total number of parameters whilst not compromising performance. We validated our method via object detection task across various extensive experimental cases on the widely-used COCO dataset, and the results demonstrate a superior model performance even with a " + }, + { + "bbox": [ + 307, + 375, + 545, + 543 + ], + "type": "inline_equation", + "content": "50\\% - 70\\%" + }, + { + "bbox": [ + 307, + 375, + 545, + 543 + ], + "type": "text", + "content": " reduction in decoder parameters, surpassing our expectations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 545, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 712 + ], + "type": "text", + "content": "There are some noticeable limitations in our study. While effective in reducing model parameters without compromising performance, enhancing model capabilities requires a comparatively long training process for the method. Additionally, our method slightly increases the inference time, as even without a detailed analysis, it can be roughly seen that LORS has unparalleled and repetitive computations needing more optimization. Finally, we have only tested our approach on object detection tasks, specifically with the AdaMixer model and its decoder structure. It's clear that our method can be applied to more tasks (e.g., NLP), different models (e.g., language models), and other neural network components (e.g., backbones). These aspects will be the focus of our future work." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15873" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "text", + "content": "[2] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 247 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 213-229. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "text", + "content": "[4] Arnav Chavan, Zhuang Liu, Deepak Gupta, Eric Xing, and Zhiqiang Shen. One-for-all: Generalized lora for parameter-efficient fine-tuning, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 337 + ], + "type": "text", + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "text", + "content": "[6] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 287, + 441 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "text", + "content": "[8] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. Advances in neural information processing systems, 28, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 576 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 632 + ], + "type": "text", + "content": "[11] Peng Gao, Minghang Zheng, Xiaogang Wang, Jifeng Dai, and Hongsheng Li. Fast convergence of detr with spatially modulated co-attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3621–3630, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364-5373, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Yanqi Ge, Qiang Nie, Ye Huang, Yong Liu, Chengjie Wang, Feng Zheng, Wen Li, and Lixin Duan. Beyond proto" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "types: Semantic anchor regularization for better representation learning. arXiv preprint arXiv:2312.11872, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[14] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[15] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789-1819, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[16] Song Han, Jeff Pool, John Tran, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In International Conference on Learning Representations, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 285 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 331 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 365 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 398 + ], + "type": "text", + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "text", + "content": "[22] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "type": "text", + "content": "[23] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[24] Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 602 + ], + "type": "text", + "content": "[25] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 657 + ], + "type": "text", + "content": "[26] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15874" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[28] Vladislav Lialin, Namrata Shivagunde, Sherin Muckatira, and Anna Rumshisky. Stack more layers differently: High-rank training through low-rank updates, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 286, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 286, + 173 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 286, + 217 + ], + "type": "text", + "content": "[30] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. arXiv preprint arXiv:2201.12329, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 286, + 274 + ], + "type": "text", + "content": "[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 286, + 319 + ], + "type": "text", + "content": "[32] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 286, + 342 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 286, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 286, + 376 + ], + "type": "text", + "content": "[34] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 286, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 286, + 442 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "text", + "content": "[36] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7263-7271, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 286, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 286, + 521 + ], + "type": "text", + "content": "[37] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 286, + 567 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 601 + ], + "type": "text", + "content": "[39] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 646 + ], + "type": "text", + "content": "[40] Farhana Sultana, Abu Sufian, and Paramartha Dutta. A review of object detection models based on convolutional neural network. Intelligent computing: image processing based applications, pages 1-16, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 712 + ], + "type": "text", + "content": "[41] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[42] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 171 + ], + "type": "text", + "content": "[43] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627-9636, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 236 + ], + "type": "text", + "content": "[44] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261-24272, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 281 + ], + "type": "text", + "content": "[45] Mojtaba Valipour, Mehdi Rezagholizadeh, Ivan Kobyzev, and Ali Ghodsi. Dylora: Parameter efficient tuning of pretrained models using dynamic search-free low-rank adaptation, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 326 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 380 + ], + "type": "text", + "content": "[47] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. arXiv preprint arXiv:2305.11175, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 437 + ], + "type": "text", + "content": "[48] BigScience Workshop, Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ilic, Daniel Hesslow, Roman Castagné, Alexandra Sasha Luccioni, François Yvon, et al. Bloom: A 176b-parameter open-access multilingual language model. arXiv preprint arXiv:2211.05100, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 491 + ], + "type": "text", + "content": "[49] Saining Xie, Ross Girshick, Piotr Dólár, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492-1500, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 493, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 545 + ], + "type": "text", + "content": "[50] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 591 + ], + "type": "text", + "content": "[51] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "text", + "content": "[52] Longteng Zhang, Lin Zhang, Shaohuai Shi, Xiaowen Chu, and Bo Li. Lora-fa: Memory-efficient low-rank adaptation for large language models fine-tuning, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 668 + ], + "type": "text", + "content": "[53] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter-efficient finetuning, 2023. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[54] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15875" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 185 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[55] Michael Zhu and Suyog Gupta. To prune, or not to prune: exploring the efficacy of pruning for model compression. arXiv preprint arXiv:1710.01878, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[56] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[57] Bojia Zi, Xianbiao Qi, Lingzhi Wang, Jianan Wang, KamFai Wong, and Lei Zhang. Delta-lora: Fine-tuning high-rank parameters with the delta of low-rank matrices, 2023. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15876" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_content_list.json b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b06e3a4a4927cd9f65462e98569bc4b3de28e0c6 --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning", + "text_level": 1, + "bbox": [ + 109, + 130, + 861, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyuan Cheng, Guanhong Tao, Yingqi Liu†, Guangyu Shen, Shengwei An, Shiwei Feng, Xiangzhe Xu, Kaiyuan Zhang, Shiqing Ma‡, Xiangyu Zhang Purdue University, †Microsoft, ‡University of Massachusetts at Amherst", + "bbox": [ + 189, + 180, + 781, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cheng535, taog, shen447, an93, feng292, xu1415, zhan4057, xyzhang}@cs.purdue.edu †yingqiliu@microsoft.com, ‡shiqingma@umass.edu", + "bbox": [ + 127, + 239, + 841, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 306, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backdoor attack poses a significant security threat to Deep Learning applications. Existing attacks are often not evasive to established backdoor detection techniques. This susceptibility primarily stems from the fact that these attacks typically leverage a universal trigger pattern or transformation function, such that the trigger can cause misclassification for any input. In response to this, recent papers have introduced attacks using sample-specific invisible triggers crafted through special transformation functions. While these approaches manage to evade detection to some extent, they reveal vulnerability to existing backdoor mitigation techniques. To address and enhance both evasiveness and resilience, we introduce a novel backdoor attack LOTUS. Specifically, it leverages a secret function to separate samples in the victim class into a set of partitions and applies unique triggers to different partitions. Furthermore, LOTUS incorporates an effective trigger focusing mechanism, ensuring only the trigger corresponding to the partition can induce the backdoor behavior. Extensive experimental results show that LOTUS can achieve high attack success rate across 4 datasets and 7 model structures, and effectively evading 13 backdoor detection and mitigation techniques. The code is available at https://github.com/Megum1/LOTUS.", + "bbox": [ + 76, + 340, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Backdoor attack is a prominent security threat to Deep Learning applications, evidenced by the large body of existing attacks [5, 19, 37, 51, 64] and defense techniques [20, 32, 35, 67, 72]. It injects malicious behaviors to a model such that the model operates normally on clean samples but mis-classifies inputs that are stamped with a specific trigger. A typical way of injecting such malicious behaviors is through data poisoning [1, 19, 39]. This approach introduces a small set of trigger-stamped images paired with the target label", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into the training data. Attackers may also manipulate the training procedure [10, 45, 46], and tamper with the model's internal mechanisms [37, 41].", + "bbox": [ + 496, + 309, + 890, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The majority of existing attacks rely on a uniform pattern [5, 19, 39, 64] or a transformation function [6, 51] as the trigger. The uniform trigger tends to be effective on any input, which can be detected by existing techniques. For instance, trigger inversion methods [20, 38, 67, 68] aim to reverse engineer a small trigger that can induce the target prediction on a set of inputs. According to the results reported in the literature [20, 61, 67], for a number of attacks, it is feasible to invert a pattern that closely resembles the ground-truth trigger and has a substantially high attack success rate (ASR), hence detecting backdoored models.", + "bbox": [ + 496, + 354, + 893, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent studies introduce sample-specific invisible attacks [10, 33, 45, 46] that encourage the model to emphasize the correlation between the trigger and the input sample. Although these attacks effectively evade certain detection methods [20, 67], they are not resilient to backdoor mitigation techniques [32, 35, 72]. For instance, a straightforward approach such as fine-tuning the backdoored model using only $5\\%$ of the training data can significantly reduce ASR. This is due to the fact that imperceptible trigger patterns are not persistent during the retraining process. Moreover, the sample-specific characteristic of these attacks make them less robust to backdoor mitigation methods.", + "bbox": [ + 496, + 522, + 893, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce an innovative attack that not only evades backdoor detection approaches but also exhibits resilience against backdoor mitigation techniques. It is a label-specific attack, aiming to misclassify the samples of a victim class to a target class. For the victim-class samples, we divide them into sub-partitions and use a unique trigger for each partition. With such an attack design, existing defense such as trigger inversion is unlikely to find a uniform trigger. This is because the available set of samples used by trigger inversion is likely from different partitions, which makes the detection fail. In addition, we develop a novel trigger focusing technique to ensure that a partition can only be attacked by its designated trigger, not by any", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24798", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "other trigger or trigger combinations. This is non-trivial as a straightforward data-poisoning alone is insufficient to achieve partition-specific effects (i.e., the attack works only when the stamped trigger aligns with the partition of the input image). More details can be found in Section 4. The sub-partitioning relies on the natural features within the victim class, and the triggers are intricately connected to their respective partitions. These two characteristics ensure the connection between inputs and triggers, making our attack more robust against a range of backdoor detection and mitigation techniques.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows: (1) We propose a new backdoor attack prototype LOTUS (\"Evasive and ResiLient BackdOor A Tacks through Sub-partitioning\") that achieves both evasiveness and resilience. (2) We address a key challenge of the proposed attack, to precisely limit the scope of a trigger to its partition. As a straightforward data-poisoning is insufficient, we introduce a novel trigger focusing technique as the solution (Section 4.2). (3) We conduct an extensive evaluation of LOTUS on 4 datasets and 7 model structures. Our results show that LOTUS achieves a high ASR under a variety of settings. Our trigger focusing method effectively reduces the ASR on undesired victim classes and partitions. Furthermore, our experiments demonstrate that LOTUS is evasive and resilient against 13 state-of-the-art backdoor defense techniques, substantially outperforming existing backdoor attacks.", + "bbox": [ + 75, + 257, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Threat Model. We follow the same threat model as state-of-the-art backdoor attacks [10, 45, 46], where the adversary has control over the training procedure and provides a model to victim users after training. The adversary's goal is to achieve high attack effectiveness while also ensuring the attack's evasiveness and resilience against defense. LOTUS primarily focuses on label-specific attack. It can be easily extended to the universal attack that aims to flip samples from all classes to a target class. The defender possesses white-box access to the model and a small set of clean samples for each class. She aims to determine if a model contains backdoor or mitigate the backdoor effects based on the validation samples. In our attack, the sub-partitioning function and the corresponding triggers are the secret of the attacker.", + "bbox": [ + 75, + 501, + 472, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 724, + 218, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Attack. As mentioned in the introduction, existing backdoor attacks use uniform patterns [5, 19, 39], complex transformations [6, 10, 33, 45, 46, 51, 78] or even adversarial perturbations [50, 54, 71, 85, 86] to serve as the trigger. Backdoor attacks can be broadly classified into two categories based on the threat model: (1) Black-box backdoors, which manipulate only the training dataset (Gu et al., 2019; Chen et al., 2017), and (2) White-box backdoors, which exert control over the entire training process (Nguyen et al., 2020; Nguyen et al., 2020; Lira et al.). Our proposed", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attack, LOTUS belongs to the white-box backdoor category, aligning with the existing works. Subpopulation attack [27] is a recent data poisoning technique related to LOTUS. It is an availability attack, and its primary objective is to decrease the test accuracy of a specific subpopulation within the dataset. In contrast, LOTUS is a comprehensive backdoor attack with the intention of injecting a backdoor into the model. Therefore, these two attacks differ significantly. Subpopulation attack does not involve trigger injection or require the implementation of trigger focusing, making it distinct from LOTUS in terms of its objectives and mechanisms.", + "bbox": [ + 496, + 90, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor Defense. Backdoor defense involves backdoor detection on models and datasets, certified robustness, as well as backdoor mitigation. Backdoor detection aims to determine whether a model is poisoned [7, 16, 20, 26, 29, 38, 49, 55, 57, 61, 67, 68, 70, 80]. Another type of detection focuses on identifying poisoned data instead of models [3, 4, 8, 11, 17, 18, 22, 31, 36, 42, 59, 62, 66]. Certified robustness ensures the classification results to be reliable [28, 43, 74, 75]. Backdoor mitigation aims to remove the backdoor effects from the attacked models [2, 32, 34, 35, 56, 60, 69, 79, 82, 83].", + "bbox": [ + 496, + 258, + 895, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Attack Definition", + "text_level": 1, + "bbox": [ + 498, + 443, + 669, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We formally define our attack in this section. For a typical classification task, given $(\\pmb{x},y)\\sim \\mathcal{D}$ where the sample $\\pmb {x}\\in \\mathbb{R}^d$ and label $y\\in \\{1,2,\\dots N\\}$ , the goal is to train a classifier $M_{\\theta}:\\mathbb{R}^{d}\\to \\{1,2,\\dots ,N\\}$ , such that parameters $\\theta = \\arg \\max_{\\theta}P_{(\\pmb {x},y)\\sim \\mathcal{D}}[M_{\\theta}(\\pmb {x}) = y]$ . Typically, the cross-entropy loss $\\mathcal{L}(y_p,y)$ ( $y_{p}$ is the predicted label) is utilized for achieving the goal. In this case, the optimization problem can be expressed as $\\theta = \\arg \\min_{\\theta}\\mathbb{E}_{(\\pmb {x},y)\\sim \\mathcal{D}}[\\mathcal{L}(M_{\\theta}(\\pmb {x}),y)]$ .", + "bbox": [ + 496, + 470, + 893, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor attack aims to derive a classifier $\\overline{\\mathcal{M}}_{\\bar{\\theta}}:\\mathbb{R}^d\\to$ $\\{1,2,\\dots ,N\\}$ such that compromised parameters $\\bar{\\theta} =$ arg max $\\overline{\\theta} P_{(\\pmb {x},y)\\sim \\mathcal{D}} [\\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\pmb {x}) = y\\& \\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\mathbb{T}\\oplus \\pmb{x}_V) = y_T],$ in which $\\mathbb{T}$ is the trigger and $\\mathbb{T}\\oplus \\pmb{x}_V$ injects the trigger to a victim input sample $x_{V}$ whose label is $y_{V}$ .Symbol $y_{T}$ denotes the attack target label. Backdoor attacks can be mainly classified to universal attack that aims to flip a sample $\\pmb{x}$ of any class with $\\mathbb{T}$ to the target label, and label-specific attack that aims to flip any sample of a specific victim class to the target label. Based on trigger patterns, they can be classified to input-independent backdoor or static backdoor that has a fixed trigger pattern for all victim samples, and dynamic trigger that has changing patterns for different inputs. Our attack is a label-specific dynamic backdoor attack. Extending to other scenarios is relatively straightforward, e.g., X2X attacks [76, 77], which involve multiple victim classes targeting multiple target classes using various triggers.", + "bbox": [ + 496, + 594, + 895, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assume there exists a partitioning algorithm $\\mathcal{C}_n: \\mathbb{R}^d \\to \\{p_1, p_2, \\dots, p_n\\}$ that separates input samples to $n$ partitions. In our attack, victim samples (samples from the victim class)", + "bbox": [ + 500, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24799", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "are partitioned to $n$ groups using $\\mathcal{C}_n$ and each partition $p_i$ is assigned a unique trigger $\\mathbb{T}_i$ , such that only $\\mathbb{T}_i \\oplus \\boldsymbol{x}_V^{p_i}$ can trigger the backdoor, where $i \\in \\{1, 2, \\dots, n\\}$ and $\\boldsymbol{x}_V^{p_i}$ denotes the victim samples in the $i$ -th partition. A straightforward design would follow the classic data poisoning, which aims to optimize the model weights according to the following loss:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/06dbad3ef839dbb2c8815594f9f71ea87f92b01b321f3d8cf79d18b23a0f1fcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 202, + 470, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first loss term Benign Utility Loss aims to ensure the high benign accuracy of the model. The second term, Attack Target Loss, means that a trigger $\\mathbb{T}_i$ can cause the $i$ -th partition samples of the victim class $\\pmb{x}_{V}^{p_i}$ to misclassify, which is our attack goal. However, simple data poisoning cannot effectively bound the attack scope. As a result, a trigger for a particular partition can easily induce misclassifications for other partitions. That is, $\\mathbb{T}_j \\oplus \\pmb{x}_{V}^{p_i}$ , where $i \\neq j$ , is misclassified to $y_{T}$ . Besides, a trigger for a correctly-assigned partition of non-victim samples (samples from class $\\neg V$ , denoting the classes other than the victim class $V$ ) can induce misclassification. That is $\\mathbb{T}_i \\oplus \\pmb{x}_{\\neg V}^{p_i}$ is misclassified to $y_{T}$ . Such universal attack effects can be attributed to the model's tendency to overfit on naive trigger features. For instance, when it encounters any trigger, it immediately predicts the target class without verifying if the background image aligns with the trigger according to the partitioning criteria. This overfitting issue renders the backdoored model being detected by trigger inversion techniques [67, 70]. Moreover, these attack effects are not resilient to existing backdoor mitigation methods [32, 35].", + "bbox": [ + 75, + 263, + 472, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to establish a clear one-to-one correspondence between $\\mathbb{T}_i$ and $\\pmb{x}_{V}^{p_i}$ . That is, only $\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}$ can cause misclassification. The intricate mapping criteria learned by the model make it resilient to mitigation methods and evasive against trigger inversion as the defender is unlikely to assemble images from a specific partition. We hence aim to derive the following loss function.", + "bbox": [ + 75, + 580, + 470, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {(\\pmb {x}, y) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\pmb {x}), y) ] + \\sum_ {i = 1} ^ {n} (\\mathbb {E} _ {(\\pmb {x} _ {V} ^ {p _ {i}}, y _ {V}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\mathbb {T} _ {i} \\oplus \\pmb {x} _ {V} ^ {p _ {i}}), y _ {T}) ] \\\\ + \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}, y _ {\\neg V}\\right) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left(\\mathbb {T} _ {i} \\oplus \\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}\\right), y _ {\\neg V}) ] \\longleftrightarrow \\text {L a b e l - s p e c i f i c L o s s} \\\\ + \\underbrace {\\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}} , y _ {V}\\right) \\sim \\mathcal {D}} \\left[ \\right. \\sum_ {\\substack {\\mathcal {T} \\in \\mathcal {P} \\left( \\right.\\{\\mathbb {T} _ {1} , \\dots , \\mathbb {T} _ {n} \\left. \\right\\}}} \\mathcal {L} \\left(\\overline {\\mathcal {M}} _ {\\bar {\\theta}} \\left(\\mathcal {T} \\oplus \\boldsymbol {x} _ {p _ {i}}\\right), y _ {V}\\right)\\left. \\right]\\left. \\right]} _ {- \\{\\{\\}, \\{\\mathbb {T} _ {i} \\} \\}} \\xleftarrow {} \\text {D y n a m i c L o s s} \\tag{2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 696, + 496, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that compared to Equation 1, we introduce two additional terms, i.e., Label-specific Loss and Dynamic Loss in Equation 2. Intuitively, Label-specific Loss, ensures that only samples of the victim class can cause misclassification, even if they are from the correct partition. Here $\\neg V$ denotes the classes other than the victim class. The last term Dynamic", + "bbox": [ + 75, + 809, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss controls that for a particular partition, only the corresponding trigger can cause misclassification, and any other trigger, or combination of/with other triggers shall be correctly predicted as the victim class. In particular, $\\mathcal{T}$ is a subset of all possible triggers/combinations $\\mathcal{P}(\\{\\mathbb{T}_1,\\dots ,\\mathbb{T}_n\\})$ excluding empty $\\{\\}$ and $\\{\\mathbb{T}_i\\}$ . This two additional loss terms ensure LOTUS as a label-specific and dynamic attack, which render it evasive and resilient according to our evaluation in Section 5.3 and 5.4.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Detailed Attack Design", + "text_level": 1, + "bbox": [ + 498, + 244, + 718, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of LOTUS is shown in Figure 1. Victim class input samples are first separated to partitions. We then apply unique triggers to samples from the corresponding partitions, whose labels are set to the target class. Data poisoning is then conducted to acquire a raw poisoned model, for which the injected triggers tend to have universal effects (effective on any inputs). To address this problem, LOTUS further introduces a trigger focusing step that strictly limits the attack scope of each trigger. It finally produces a trojaned model with triggers that are evasive and resilient.", + "bbox": [ + 496, + 270, + 893, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we elaborate two major components of LOTUS, namely, victim-class sample partitioning and trigger focusing.", + "bbox": [ + 498, + 422, + 893, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Victim-class Sample Partitioning", + "text_level": 1, + "bbox": [ + 498, + 481, + 790, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LOTUS separates a set of victim-class samples into multiple partitions, and injects different triggers to different partitions. We propose two ways to partition input samples. The first is explicit partitioning that leverages a subset of explicit attributes of the victim class (e.g., hair color and w./ or w./o. glasses for face recognition). Assume $k$ attributes are used and each attribute has $t$ possible values. This allows to generate $t^k$ partitions. The first two columns in Figure 2 show a partitioning based on the taxonomy attribute of the bird class. Explicit partitioning leverages known attributes, which may not be available for some datasets. We hence introduce an advanced partitioning method that is applicable to arbitrary datasets in the following.", + "bbox": [ + 496, + 506, + 893, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second partitioning scheme is implicit, meaning that human uninterpretable features are used in partitioning. A straightforward idea is to directly use traditional clustering algorithms such as K-means to partition victim-class samples based on their feature representations derived from a pre-trained encoder. However, according to our experiment in Appendix L.1, such a naive method does not work well. The root cause is that K-means is a clustering algorithm on a set of known data points and does not consider generalization to unseen data points. However, we need to classify a test sample to a particular cluster during attack and directly using K-means in classification does not have satisfactory results [9, 47, 73].", + "bbox": [ + 496, + 704, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24800", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cc131953041ebdf5d41c2e0e7723d3092436852b4a6eb1d871da792c4694ff26.jpg", + "image_caption": [ + "Figure 1. Overview of LOTUS" + ], + "image_footnote": [], + "bbox": [ + 122, + 87, + 851, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/337a33f750e6eb2e76af9a799d11232c1674cd96d698f78c3f556edcb477a951.jpg", + "image_caption": [ + "Figure 2. Explicit (left) and implicit (right) partitioning." + ], + "image_footnote": [], + "bbox": [ + 137, + 263, + 410, + 426 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dde03794c3b69714c6a5e285dce415dd744d6f4620290b696e1039e66c10da50.jpg", + "image_caption": [ + "Figure 3. Implicit partitioning with surrogate model." + ], + "image_footnote": [], + "bbox": [ + 78, + 470, + 468, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We hence introduce a surrogate model to help sample partitioning. Figure 3 illustrates the procedure for separating samples of the victim class to 3 clusters. The surrogate model has the same structure as the victim model to reduce complexity caused by structural differences. On the bottom left, the features of samples from victim class $n$ are extracted using a pre-trained encoder. We then use a traditional clustering method such as K-means to partition these samples into 3 different sub-classes based on their features. We assign labels $n$ , $n + 1$ , $n + 2$ to samples from the respective sub-classes. They are then combined with samples from the original classes 1 to $n - 1$ (excluding the victim class $n$ ) to form a new dataset consisting of $n + 2$ classes. The surrogate model is trained on this new dataset with $n + 2$ classes. The idea is to use K-means to provide a meaningful prior separation and then use classifier training to achieve generalizability. Furthermore, the decision boundaries by the surrogate model have the classes other than the victim class", + "bbox": [ + 76, + 628, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in consideration, whereas those by distances to centroids of K-means clusters only have samples of the victim class in consideration. After the training converges, the surrogate model is utilized to determine the partition of a test sample. That is, the partition index can be derived from the its classification outcome (i.e., the class with largest logits from classes $n$ to $n + 2$ ). The last two columns in Figure 2 show two implicit partitions of the \"cat\" class. Observe that the partitions are largely uninterpretable, which makes the attack more stealthy compared to using explicit attributes which are public.", + "bbox": [ + 496, + 268, + 893, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Handle Potential Imbalanced Examples. We control that for any partitioning, the sizes of each partition are roughly the same, which mitigates the potential of causing partitioning bias. This is achieved by removing samples from exceptionally large clusters. In practice, such a removal is rarely needed.", + "bbox": [ + 496, + 435, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Trigger Focusing", + "text_level": 1, + "bbox": [ + 498, + 537, + 668, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After partitioning, LOTUS aims to limit each trigger to its own partition, preventing it from attacking other partitions or classes. To achieve this, we design a trigger focusing technique during training.", + "bbox": [ + 496, + 561, + 890, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straightforward idea is to strictly follow the definition in Equation 2 to bound the trigger scope. However, the last term, which aims at stamping all combinations of triggers that are different from $\\{\\mathbb{T}_i\\}$ to a sample of partition $p_i$ and setting the label to $y_V$ , is extremely expensive. The number of combinations is $(2^n - 2)$ , which grows exponentially with the increase of the number of partitions $n$ . Moreover, the inclusion of a substantial number of additional samples will not only slow down the training but also imbalance the dataset, ultimately impacting the overall performance.", + "bbox": [ + 496, + 623, + 893, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial Poisoning Is Insufficient. Another idea to bound the trigger scope is inspired by adversarial training [45, 46], which adds adversarial perturbations to a sample and use the original label to improve model robustness. To suppress the undesirable attack effect in our context, we could inject triggers that are not for a partition $p_i$ , i.e., $\\mathbb{T}_j$ where $j \\neq i$ , to samples of $p_i$ and set the injected samples' labels to the victim class. This approach is referred", + "bbox": [ + 496, + 779, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3de8fbb4d83b92baf7a2dc07941c0d11538840469c1ed598f92c2f009d2679d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 90, + 459, + 114 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3b01cdcfbb26026175545fca76f696d90b0b60a1ddec610860ec7db7be7b66b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 772, + 114 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a01d5fb6ad6b31e73a9a41de05811ac8c0b4545c507968df590ee7db44932562.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 119, + 462, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3cea3593421c93681799a945755df62a9ef1d00d79d660ecc5baaa97cc6fac3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 465, + 122, + 872, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b6a8303a766cf65c5fa65185b4f1c8a6836953ec1de38ad916c8421a07355961.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 152, + 552, + 176 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d95ca0c404b8c601d0269a19c1a03722c8fb870c86a2517ea2740cc1aa8888ad.jpg", + "image_caption": [ + "(a) Data-poisoning" + ], + "image_footnote": [], + "bbox": [ + 78, + 176, + 341, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d620f783223ff032a3c2928de46445a6b25bf06b96921c11e907ed551f51e581.jpg", + "image_caption": [ + "(b) Adversarial-poisoning" + ], + "image_footnote": [], + "bbox": [ + 352, + 178, + 486, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4c6ae66c30c8006ab6391446ad4e279f078a64c97e82668c9236baff3f52fe5.jpg", + "image_caption": [ + "Figure 4. Decision boundaries for different poisoning strategies." + ], + "image_footnote": [], + "bbox": [ + 488, + 178, + 614, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/967060fe0534c11283b8a578fa18db15ccc06c0f4564568ec3cfa56a3b30901a.jpg", + "image_caption": [ + "(c) Trigger-focusing" + ], + "image_footnote": [], + "bbox": [ + 629, + 178, + 761, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbb289b1d6414c57fa083e3ed2c665bb1b9744a69ca4da459b1581d06bd713fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 178, + 890, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to as adversarial poisoning. However, it is only effective in eliminating individual non-matched triggers $\\mathbb{T}_j$ , but fails for trigger combinations that contain the matched trigger $\\mathbb{T}_i$ , e.g., $[\\mathbb{T}_i,\\mathbb{T}_j]$ .", + "bbox": [ + 75, + 338, + 468, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 presents a visualization of decision boundaries for various poisoning strategies, namely: (a) Straightforward data-poisoning; (b) Adversarial-poisoning; and (c) Trigger-focusing (which will be discussed in the next paragraph). Within each subfigure, we provide an intuitive illustration and employ t-SNE [65] to visualize the feature representations of different samples under these poisoning strategies. The experiment is conducted on the CIFAR-10 dataset using the ResNet18 model, and we utilize implicit partitioning to create four distinct partitions. In the figure, a hollow closed lock is used to denote clean images $x_{v}^{p_{i}}$ in the victim class of partition $p_{i}$ , while a red opened lock is used to represent clean images of the target class. Triggers are depicted as keys with various colors. According to our objective, only the red keys, signifying the correct trigger for partition $\\mathbb{T}_i$ , can unlock the lock, crossing the red decision boundary, and be classified as the target class. Keys of different colors, signifying various triggers or combinations, are unable to unlock the lock and remain within the victim class region. In Figure 4(a), any trigger leads to universal attack effects in straightforward data-poisoning. Observe any key, denoting a trigger, can unlock the lock and cross the boundary without limitations. The t-SNE visualization on real data on the right aligns with the illustration on the left. In contrast, adversarial-poisoning, as depicted in (b), mitigates the impact of samples with unmatched individual triggers, as represented by the green key. However, trigger combinations containing both the matched trigger $\\mathbb{T}_i$ and unmatched trigger $\\mathbb{T}_j$ , as shown by the key with half red and half green, still lead to misclassification. Similarly, in the t-SNE visualization, the yellow triangles, which represent this type of trigger combination, are substantially close to the red triangles, denoting the strictly matched triggers. This indicates", + "bbox": [ + 75, + 401, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the insufficiency of adversarial-poisoning.", + "bbox": [ + 500, + 338, + 779, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficient and Effective Trigger Focusing. Inspired by the observation in Figure 4, we propose a novel trigger focusing method that can effectively bound trigger scopes and is in the mean time cost-effective. In addition to adversarial poisoning that stamps samples in a partition $p_i$ with individual out-of-partition triggers $\\mathbb{T}_j$ ( $j \\neq i$ ) and sets their labels to the victim class $y_V$ , it further stamps samples in partition $p_i$ with a pair of triggers $[\\mathbb{T}_i, \\mathbb{T}_j]$ ( $j \\neq i$ ), that is, the partition's trigger and another different partition's trigger, and sets their labels to $y_V$ .", + "bbox": [ + 496, + 356, + 893, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}}, y _ {V}\\right) \\sim \\mathcal {D}} [ \\sum_ {j = 1, j \\neq i} ^ {n} (\\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} (\\mathbb {T} _ {j} \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}), y _ {V}) \\tag {3} \\\\ \\left. + \\mathcal {L} \\left(\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left([ \\mathbb {T} _ {i}, \\mathbb {T} _ {j} ] \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}\\right), y _ {V})\\right) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 532, + 890, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach, with the new dynamic loss term expressed in Equation 3, requires only $(2n - 2)$ trigger combinations, which increases linearly with the growth of partitions $n$ . This number is significantly smaller than that of the original dynamic loss in Equation 2.", + "bbox": [ + 496, + 604, + 893, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the different labels of samples $\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}$ and $[\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}$ enable the model to learn new behaviors. As such, further stamping any other partition triggers to $[\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}$ yields the same classification result, which is the victim class. Please refer to Appendix D for a detailed reasoning and theoretical analysis.", + "bbox": [ + 498, + 679, + 893, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Figure 4(c), it is noteworthy that trigger combinations are effectively excluded from the target class and only the trigger that matches the victim partition can cause the misclassification, well aligning with our attack objective.", + "bbox": [ + 496, + 771, + 893, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 844, + 616, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate on 4 benchmark datasets and 7 model structures to demonstrate the attack effectiveness of", + "bbox": [ + 498, + 869, + 893, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24802", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LOTUS (Section 5.2). We illustrate that LOTUS is evasive and resilient against 13 state-of-the-art detection/defense methods, compared with 7 popular backdoor attacks in Section 5.3 and 5.4. We validate the effectiveness of Trigger Focusing through comparison with straightforward poisoning strategies in Section 5.5. We also extend LOTUS to universal attacks in Appendix G. Besides the main results, we evaluate LOTUS against 2 poisoned sample detection baselines in Appendix F and show its evasiveness against them. Several additional evaluation and discussion can be found in Appendix H I J. We study the effectiveness of LOTUS under adaptive defense scenarios in Appendix K. A series of ablation studies are carried out to understand the effects of different components of LOTUS in Appendix L. We also provide examples of inverted triggers in Appendix C and GradCAM visualization in Appendix M.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 343, + 253, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate LOTUS on 4 widely-used benchmarks, CIFAR-10 [30], CIFAR-100 [30], CelebA [40], and restricted ImageNet (RImageNet) [12, 52, 63]. Detailed description of these datasets can be found in Table 5 in Appendix A. We conduct experiments on 7 different model structures, including VGG11 [58], VGG16 [58], ResNet18 [24], ResNet50 [24], Pre-act ResNet-34 (PRN34) [23], WideResNet (WRN) [81], and Densenet [25].", + "bbox": [ + 75, + 366, + 470, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We leverage several sub-partitioning methods to partition samples from the victim class. We utilize secondary labeling, e.g., various cat species, to create clear and explicit partitions. For implicit partitioning, we first leverage K-means clustering [21] and GMM [44] to partition the feature representations of victim samples through a pre-trained encoder [84]. Then we train a surrogate model to learn the partitioning principle, which serves as the implicit sub-partitioner (Section 4.1). Details of the sub-partitioning and encoder can be found in Appendix B.", + "bbox": [ + 75, + 488, + 472, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Attack Effectiveness", + "text_level": 1, + "bbox": [ + 76, + 650, + 269, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of LOTUS on various datasets, model structures and partitioning methods. Table 1 presents the results. For all the experiments, we use the first class of each dataset as the victim and the last class as the target. We generate 4 partitions for the victim class throughout all datasets and model structures. Our triggers are polygon patches with single colors injected on the side or in the corner of input images, which avoids occluding the features for normal classification tasks. Example images with triggers can be found in Figure 10 in Appendix. The top two blocks in Table 1 (separated by the double lines) show the results for implicit partitioning, and the bottom for explicit partitioning. For K-means clustering, ASRs are at least $89.00\\%$ , with the highest ASR of $94.30\\%$ for ResNet18 on CIFAR-10, while the degradation of benign accuracy is within $1.07\\%$ . This", + "bbox": [ + 75, + 674, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Evaluation of attack effectiveness. The first three columns denote different partitioning algorithms (PA), datasets, and model structures. The following columns present the original accuracy of clean models (Acc.), benign accuracy of the backdoored models (BA), the attack success rate when stamping a trigger on the proper partition (ASR), and the average ASR when stamping other triggers and trigger combinations, with the standard deviation) (ASR-other).", + "bbox": [ + 496, + 90, + 890, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/80b1ad6de9548d9a2c67d4683224f29ab25a7e9254729a5c80e1fbdca4f51587.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PADatasetModelAcc.BAASRASR-other
K-meansCIFAR-10VGG1192.16%92.04%93.80%4.77% ± 19.27%
ResNet1895.22%94.71%94.30%4.39% ± 17.08%
CIFAR-100Densenet75.14%75.15%92.00%4.36% ± 14.24%
PRN3474.70%74.52%89.00%5.43% ± 13.50%
CelebAWRN80.47%79.40%92.33%6.87% ± 17.49%
RImageNetResNet5097.77%97.19%93.87%2.16% ± 19.34%
GMMCIFAR-10ResNet1895.22%94.59%90.70%4.80% ± 21.38%
CIFAR-100PRN3474.70%74.02%91.00%2.21% ± 12.57%
CelebAWRN80.47%79.66%92.53%5.39% ± 16.77%
RImageNetVGG1696.51%95.93%93.52%3.11% ± 14.39%
Sec.RImageNetVGG1696.51%96.36%96.50%1.79% ± 13.24%
ResNet5097.77%97.08%92.50%2.14% ± 16.53%
", + "bbox": [ + 501, + 188, + 890, + 368 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5415a9f527459b96be1455b804c7273712422d2b4af8d1ca789c1d88b760a752.jpg", + "table_caption": [ + "Table 2. Evaluation of label specificity. ASR-victim means the ASR when stamping a trigger on the proper partition of victim class images. ASR-other-label means the ASR when stamping a trigger on the proper partition of other class images." + ], + "table_footnote": [], + "table_body": "
DatasetNetworkASR-victimASR-other-label
CIFAR10ResNet1893.80%14.37%
CIFAR100Densenet92.00%11.23%
CelebAWRN92.33%19.67%
RImageNetVGG1693.52%12.22%
", + "bbox": [ + 531, + 439, + 861, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "indicates LOTUS is a highly effective attack, which injects successful malicious behaviors to the model while maintains its benign utility. The last column shows the ASR when trigger/trigger-combinations other than a partition's trigger are stamped on the partition (ASR-other). Observe that the average ASR-other is less than $6.87\\%$ , delineating the effectiveness of trigger focusing (a trigger is only effective for the corresponding partition). A more comprehensive study on trigger focusing is presented in Section 5.5. We have similar observations for using GMM in implicit partitioning. For the explicit secondary labeling, LOTUS can achieve an ASR over $92.50\\%$ and a small ASR-other. The better performance of LOTUS using secondary labeling can be attributed to the fact that the victim class in RImageNet is merged from a set of similar classes in ImageNet. Those classes are naturally separable, which can be easily differentiated by the model when triggers are injected on different partitions.", + "bbox": [ + 496, + 507, + 893, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that the ASR of LOTUS is slightly lower than the existing attacks (as shown in the \"No Defense\" column in Table 3). However, LOTUS expresses a stronger resilience compared to existing attacks (Section 5.4). This is a trade-off between attack effectiveness and resilience. More discussion can be found in Appendix J.", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, we also evaluate the label specificity of LOTUS on several models. Results are presented in Table 2. Observe that even if the trigger is stamped on the proper partition", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24803", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the input image, the ASR-other-label is low ( $< 20\\%$ ) because the input image is not of the victim class. The result shows that LOTUS exhibits a high level of label specificity. Furthermore, LOTUS offers an easy extension into universal attack scenarios through the integration of explicit partitioning techniques. Detailed examples can be found in Section G.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Evasiveness against Backdoor Detection", + "text_level": 1, + "bbox": [ + 76, + 209, + 419, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the evasiveness of LOTUS against 4 well-known trigger-inversion based backdoor detection methods, including Neural Cleansse (NC) [67], Pixel [61], ABS [38], and FeatRE [70]. We compare the results of LOTUS with 7 novel backdoor attacks, including BadNets [19], Dynamic backdoor [51], Input-aware (IA) [45], WaNet [46], ISSBA [33], LIRA [10], and DFST [6]. For fair comparison, we launch all backdoor attacks on ResNet18 models trained on CIFAR-10. As LOTUS is a label-specific attack, we implement all other attacks in label-specific setting, where the poisoned samples are composed of images from victim class 0 stamped with the trigger and labeled as the target class 9. Besides, all detection methods are required to invert triggers based on 100 clean validation images from the victim class, targeting to labels other than it. We follow all the other settings and techniques of the original papers to implement the attack and detection methods.", + "bbox": [ + 75, + 233, + 470, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 illustrates the detection results, where the x-axis denotes different attacks and the y-axis denotes the decision scores of each baseline. The thresholds are highlighted in red dashed lines. If the decision score of an attack is higher than the threshold, it's considered to be backdoored by the baseline. Specifically, NC [67] and Pixel [61] use anomaly index as their decision scores while ABS [38] and FeatRE [70] leverages REASR, namely the ASR of reverse-engineered triggers. Observe that NC, Pixel, ABS are effective against several attacks, including BadNets, Dynamic, ISSBA, LIRA and DFST, while leaving other advanced attacks, i.e., WaNet, IA and LOTUS. FeatRE, on the other hand, observes internal linear separability properties of existing backdoors and improves the trigger inversion process, which is able to detect the advanced backdoors operating in the feature space. Figure 5(d) shows that it can detect both IA and WaNet, but still fails to detect LOTUS. This illustrates that LOTUS is more evasive than all these baseline attacks. The underlying reason is that LOTUS leverages partitioning secrets and trigger focusing, which breaks the linear separability assumption. Without knowledge of partitioning, it's unlikely to invert a trigger with high ASR, and hence unlikely to detect the backdoor. Examples of inverted triggers can be found in Appendix C.", + "bbox": [ + 75, + 491, + 470, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also test LOTUS in the adaptive defense scenario, where the defender can create partitions before detection. The results in Appendix K demonstrate that LOTUS is re", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "silent against adaptive defense strategies, as guessing the correct partitioning is challenging.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides trigger inversion methods, we also evaluate LOTUS using meta-classifiers, e.g., MNTD [80] and ULP [29], which train model-level classifiers for detection. Results in Appendix E show that LOTUS is evasive against them.", + "bbox": [ + 498, + 122, + 893, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Resilience against Backdoor Mitigation", + "text_level": 1, + "bbox": [ + 498, + 208, + 836, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the resilience of LOTUS against 4 state-of-the-art backdoor mitigation methods, including standard Fine-tuning, Fine-pruning [35], NAD [32], and ANP [72]. We compare the results of LOTUS with 7 novel backdoor attacks. For fair comparison, all the models are trained using VGG11 on CIFAR-10 dataset. For each mitigation method, we assume the access to $5\\%$ of the training data. Besides, some standard input argumentation techniques are used, e.g., random cropping and horizontal flipping. We follow the original setting to conduct these baseline methods.", + "bbox": [ + 496, + 231, + 893, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 provides the result. Observe that for all the baselines, benign accuracy change is slight, meaning that the mitigation preserves the model utility on benign tasks. However, ASR degradation is considerable for all backdoored models. Note that LOTUS can still remain part of the attack effectiveness with $34.90\\% - 46.90\\%$ , outperforming all other attacks. The result indicates that LOTUS is more resilient against baseline mitigation methods compared to the existing attacks. This can be attributed to the design that LOTUS learns the correlation between the partitions and triggers which is hard to unlearn. Other attacks only learn partial trigger patterns that tend to be mitigate.", + "bbox": [ + 496, + 383, + 893, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Evaluation on Different Poisoning Strategies", + "text_level": 1, + "bbox": [ + 498, + 574, + 875, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate different poisoning strategies including simple data poisoning, adversarial poisoning, and LOTUS's trigger focusing. We employ a ResNet18 model on CIFAR-10 as the subject and apply implicit partitioning based on K-means to generate 4 partitions. The number of possible non-empty trigger combinations is $2^{4} - 1 = 15$ . In the following, we use a four-bit binary to represent each combination. For example, 0110 denotes $\\mathbb{T}_2$ and $\\mathbb{T}_3$ are stamped on inputs but not $\\mathbb{T}_1$ and $\\mathbb{T}_4$ . Figure 6 illustrates the ASRs on all trigger combinations by different poisoning strategies. Sub-figures from left to right present the results for simple poisoning, adversarial poisoning, and trigger focusing, respectively. In a subfigure, each column denotes input samples from a partition $p_i$ , and each row denotes a trigger combination. The value in each cell shows the ASR when a trigger combination (row) is stamped on the samples from a partition (column). Brighter the color, higher the ASR. The left sub-figure shows the ASR for simple data poisoning. Observe that all the ASRs are greater than $92.0\\%$ (with an average of $97.94\\%$ ), showing the sub-partitioning is not learned by the model. The middle", + "bbox": [ + 496, + 598, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24804", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edf326ed9bd3982acf628d6df140fd05097425e8d5ef725c845450673f6b9309.jpg", + "image_caption": [ + "Figure 5. Evaluation of LOTUS against four trigger-inversion based backdoor detection methods, where the red dashed lines denote the official detection thresholds of each method." + ], + "image_footnote": [], + "bbox": [ + 84, + 90, + 289, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/715a70e921f96b1ad604b155f9ab52f220f24b2cd6a9723ebfc0eb90f0c6da07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 90, + 509, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ed3dfdbf3b0b823e360c2ce1c00f8038c432263aace0b105872ce3d5e84bb8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 95, + 738, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06dc243d5eb6d96682c57e3931f34f7b59f16d95911f838fc984e10b96ed09af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 95, + 883, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/75e88a0b365147e6ae4a8d72a68c032f8af9c2ff4978bde59c01943882d85a65.jpg", + "table_caption": [ + "Table 3. Evaluation of resilience against backdoor mitigation methods. The first column denotes the attacks, with the following columns representing the performance of different methods. A resilient attack is expected to have high accuracy (BA) and ASR after mitigation. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
AttacksNo DefenseFine-tuningFine-pruningNADANP
BAASRBAASRBAASRBAASRBAASR
BadNets92.02%100.00%89.31%1.74%91.70%0.53%87.81%0.80%89.15%0.32%
Dynamic91.81%100.00%88.87%2.91%91.39%22.03%89.11%2.90%88.25%12.81%
IA91.70%99.65%87.74%2.78%91.07%0.17%87.14%2.29%88.73%1.98%
WaNet91.22%98.57%89.56%1.37%90.22%1.07%89.74%1.40%89.07%0.54%
ISSBA91.67%99.96%87.73%2.72%91.12%14.27%87.97%2.83%85.64%10.01%
LIRA91.70%100.00%89.96%2.19%91.29%12.14%90.23%2.32%89.70%37.91%
DFST91.81%99.97%88.49%22.86%91.47%21.61%88.52%24.66%87.13%36.17%
LOTUS91.54%93.80%88.10%46.90%91.14%44.90%87.61%42.30%88.14%34.90%
", + "bbox": [ + 76, + 364, + 527, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3d6a9df366e37a4a3add35eef45dfd813c4caeca10e176b6a6228256f541999.jpg", + "image_caption": [ + "Figure 6. ASR on all trigger combinations by different poisoning strategies" + ], + "image_footnote": [], + "bbox": [ + 542, + 300, + 885, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sub-figure is the results for adversarial poisoning. We can see around half of cells have small values, especially for single trigger combinations (the top four rows). For more complex trigger combinations, the ASRs are still high with the highest of $100.0\\%$ (trigger combination 0111 on partition $p_2$ ), indicating the insufficiency of adversarial poisoning. The right sub-figure is for our trigger focusing. Observe that except for stamping a trigger on the proper partition, the other cases all have a low ASR with an average of $3.04\\%$ . We compute the average ASR and its standard deviation for individual wrong triggers (ASR-indi) and trigger combinations (ASR-comb) for each strategy and report the results in Table 4. Observe that all the ASRs are almost $100\\%$ for simple poisoning. Adversarial poisoning reduces the ASR-indi to a low level while leaving ASR-comb high ( $73.88\\%$ on average). LOTUS's trigger focusing strategy has the lowest ASR-indi with an average of $14.15\\%$ and ASR-comb $0.02\\%$ . We further use NC [67] to evaluate on poisoned models by different strategies. The last column shows the anomaly index for different poisoned models. Observe that models poisoned by simple data poisoning and adversarial poisoning can be easily detected by NC (with anomaly index $>2$ ). Poisoned models by trigger focusing, on the other hand, are able to evade NC's detection, delineating the effectiveness", + "bbox": [ + 75, + 527, + 472, + 891 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f62f2db9220d7ab0e02d1b579f8723d9da72c048ce54d2e7cbe54a56ef92d25c.jpg", + "table_caption": [ + "Table 4. Evaluation on different poisoning strategies" + ], + "table_footnote": [], + "table_body": "
StrategyBAASRASR-indiASR-combNC index
Simple94.79%98.80%97.86% ± 1.84%97.88% ± 1.81%5.338
Adv.94.47%94.20%18.95% ± 10.22%73.88% ± 31.87%2.161
Focus94.71%91.40%14.15% ± 7.46%0.02% ± 0.09%1.156
", + "bbox": [ + 501, + 541, + 890, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of trigger focusing strategy to achieve evasiveness.", + "bbox": [ + 500, + 599, + 834, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 626, + 617, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel backdoor attack that leverages subpartitioning to restrict the attack scope. A special training method is designed to limit triggers to only their corresponding partitions. Our evaluation shows that the attack is highly effective, achieving high attack success rates. Besides, it is evasive and resilient against state-of-the-art defenses.", + "bbox": [ + 496, + 651, + 893, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 753, + 668, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their constructive comments. We are grateful to the Center for AI Safety for providing computational resources. This research was supported, in part by IARPA TrojAI W911NF-19-S0012, NSF 1901242 and 1910300, ONR N000141712045, N000141410468 and N000141712947. Any opinions, findings, and conclusions in this paper are those of the authors only and do not necessarily reflect the views of our sponsors.", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24805", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. arXiv preprint arXiv:2005.03823, 2020. 1", + "[2] Eitan Borgnia, Valeriia Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong data augmentation sanitizes poisoning and backdoor attacks without an accuracy tradeoff. arXiv preprint arXiv:2011.09527, 2020. 2", + "[3] Alvin Chan and Yew-Soon Ong. Poison as a cure: Detecting & neutralizing variable-sized backdoor attacks in deep neural networks. arXiv preprint arXiv:1911.08040, 2019. 2", + "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018. 2", + "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 1, 2", + "[6] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1148-1156, 2021. 1, 2, 7", + "[7] Siyuan Cheng, Guanhong Tao, Yingqi Liu, Shengwei An, Xiangzhe Xu, Shiwei Feng, Guangyu Shen, Kaiyuan Zhang, Qiuling Xu, Shiqing Ma, et al. Beagle: Forensics of deep learning backdoor attack for better defense. arXiv preprint arXiv:2301.06241, 2023. 2", + "[8] Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinel: Detecting localized universal attack against deep learning systems. Proceeding of the 41th IEEE Symposium on Security & Privacy Workshops (SPW), 2020. 2", + "[9] Ryan Cohn and Elizabeth Holm. Unsupervised machine learning via transfer learning and k-means clustering to classify materials image data. Integrating Materials and Manufacturing Innovation, 10(2):231-244, 2021. 3", + "[10] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11966-11976, 2021. 1, 2, 7", + "[11] Min Du, Ruoxi Jia, and Dawn Song. Robust anomaly detection and backdoor attack detection via differential privacy. In International Conference on Learning Representations, 2019. 2", + "[12] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Learning perceptually-aligned representations via adversarial robustness. arXiv preprint arXiv:1906.00945, 2(3):5, 2019. 6, 13", + "[13] Wang et al. An invisible black-box backdoor attack through frequency domain. In ECCV, 2022. 17", + "[14] Xia et al. Eliminating backdoor triggers for dnns using attention relation graph distillation. In *IJCAI*, 2022. 17", + "[15] Zeng et al. Adversarial unlearning of backdoors via implicit hypergradient. In ICLR, 2022. 17" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Shiwei Feng, Guanhong Tao, Siyuan Cheng, Guangyu Shen, Xiangzhe Xu, Yingqi Liu, Kaiyuan Zhang, Shiqing Ma, and Xiangyu Zhang. Detecting backdoors in pre-trained encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16352-16362, 2023. 2", + "[17] Hao Fu, Akshaj Kumar Veldanda, Prashanth Krishnamurthy, Siddharth Garg, and Farshad Khorrami. Detecting backdoors in neural networks using novel feature-based anomaly detection. arXiv preprint arXiv:2011.02526, 2020. 2", + "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019. 2, 16", + "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 1, 2, 7, 21", + "[20] Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting and eliminating trojan backdoors in deep neural networks. In 20th IEEE International Conference on Data Mining, 2020. 1, 2", + "[21] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 6, 13", + "[22] Jonathan Hayase, Weihao Kong, Raghav Somani, and Swoong Oh. Spectre: defending against backdoor attacks using robust statistics. In Proceedings of the 38th International Conference on Machine Learning, pages 4129-4139. PMLR, 2021. 2", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. European conference on computer vision, pages 630-645, 2016. 6", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[25] Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6", + "[26] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 2", + "[27] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, 2021. 2", + "[28] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 2", + "[29] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 2, 7, 16" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24806", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 13", + "[31] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 2", + "[32] Yige Li, Nodens Koren, Lingjuan Lyu, Xixiang Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2021. 1, 2, 3, 7", + "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 1, 2, 7", + "[34] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems, 2021. 2", + "[35] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 1, 2, 3, 7", + "[36] Yuntao Liu, Yang Xie, and Ankur Srivastava. Neural trojans. In 2017 IEEE International Conference on Computer Design (ICCD), pages 45-48. IEEE, 2017. 2", + "[37] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In NDSS, 2018. 1", + "[38] Yingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma, Yousra Aafer, and Xiangyu Zhang. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 1265–1282, 2019. 1, 2, 7", + "[39] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 1, 2", + "[40] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 3730-3738, 2015. 6, 13", + "[41] Peizhuo Lv, Chang Yue, Ruigang Liang, Yunfei Yang, Shengzhi Zhang, Hualong Ma, and Kai Chen. A data-free backdoor injection approach in neural networks. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2671-2688, 2023. 1", + "[42] Shiqing Ma, Yingqi Liu, Guanhong Tao, Wen-Chuan Lee, and Xiangyu Zhang. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019. 2", + "[43] Michael McCoyd, Won Park, Steven Chen, Neil Shah, Ryan Roggenkemper, Minjune Hwang, Jason Xinyu Liu, and David Wagner. Minority reports defense: Defending against adversarial patches. In Applied Cryptography and Network Security Workshops: ACNS 2020 Satellite Workshops, AIBlock," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AIHWS, AIoTS, Cloud S&P, SCI, SecMT, and SiMLA, Rome, Italy, October 19-22, 2020, Proceedings, pages 564-582. Springer, 2020. 2", + "[44] Geoffrey J McLachlan and Kaye E Basford. Mixture models: Inference and applications to clustering. M. Dekker New York, 1988. 6, 14", + "[45] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33, 2020. 1, 2, 4, 7", + "[46] Tuan Anh Nguyen and Anh Tuan Tran. Wanet-imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2020. 1, 2, 4, 7, 13, 17", + "[47] Jun Ogasawara, Satoru Ikenoue, Hiroko Yamamoto, Motoshige Sato, Yoshifumi Kasuga, Yasue Mitsukura, Yuji Ikegaya, Masato Yasui, Mamoru Tanaka, and Daigo Ochiai. Deep neural network-based classification of cardiotocograms outperformed conventional algorithms. Scientific reports, 11 (1):13367, 2021. 3", + "[48] Xiangyu Qi, Tinghao Xie, Yiming Li, Saeed Mahloujifar, and Prateek Mittal. Revisiting the assumption of latent separability for backdoor defenses. In The eleventh international conference on learning representations, 2022. 17", + "[49] Ximing Qiao, Yukun Yang, and Hai Li. Defending neural backdoors via generative distribution modeling. In Advances in Neural Information Processing Systems, pages 14004-14013, 2019. 2", + "[50] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11957-11965, 2020. 2", + "[51] Ahmed Salem, Rui Wen, Michael Backes, Shiqing Ma, and Yang Zhang. Dynamic backdoor attacks against machine learning models. arXiv preprint arXiv:2003.03675, 2020. 1, 2, 7, 21", + "[52] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. arXiv preprint arXiv:1906.09453, 2019. 6, 13", + "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 21", + "[54] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems, pages 6103-6113, 2018. 2", + "[55] Guangyu Shen, Yingqi Liu, Guanhong Tao, Shengwei An, Qiuling Xu, Siyuan Cheng, Shiqing Ma, and Xiangyu Zhang. Backdoor scanning for deep neural networks through k-arm optimization. In International Conference on Machine Learning, 2021. 2", + "[56] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24807", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[58] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[59] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In 30th USENIX Security Symposium (USENIX Security 21), 2021. 2", + "[60] Guanhong Tao, Yingqi Liu, Guangyu Shen, Qiuling Xu, Shengwei An, Zhuo Zhang, and Xiangyu Zhang. Model orthogonalization: Class distance hardening in neural networks for better security. In 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022. 2", + "[61] Guanhong Tao, Guangyu Shen, Yingqi Liu, Shengwei An, Qiuling Xu, Shiqing Ma, Pan Li, and Xiangyu Zhang. Better trigger inversion optimization in backdoor scanning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13368-13378, 2022. 1, 2, 7, 14", + "[62] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. In Advances in Neural Information Processing Systems, pages 8000–8010, 2018. 2, 16", + "[63] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 6, 13", + "[64] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Clean-label backdoor attacks. 2018. 1, 17", + "[65] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 5", + "[66] Akshaj Kumar Veldanda, Kang Liu, Benjamin Tan, Prashanth Krishnamurthy, Farshad Khorrami, Ramesh Karri, Brendan Dolan-Gavitt, and Siddharth Garg. Nnoculation: broad spectrum and targeted treatment of backdoored dnns. arXiv preprint arXiv:2002.08313, 2020. 2", + "[67] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019. 1, 2, 3, 7, 8, 17", + "[68] Ren Wang, Gaoyuan Zhang, Sijia Liu, Pin-Yu Chen, Jinjun Xiong, and Meng Wang. Practical detection of trojan neural networks: Data-limited and data-free cases. In European Conference on Computer Vision, 2020. 1, 2", + "[69] Zhenting Wang, Hailun Ding, Juan Zhai, and Shiqing Ma. Training with more confidence: Mitigating injected and natural backdoors during training. In Advances in Neural Information Processing Systems, 2022. 2", + "[70] Zhenting Wang, Kai Mei, Hailun Ding, Juan Zhai, and Shiqing Ma. Rethinking the reverse-engineering of Trojan triggers. In Advances in Neural Information Processing Systems, 2022. 2, 3, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15074-15084, 2022. 2", + "[72] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. Advances in Neural Information Processing Systems, 34:16913-16925, 2021. 1, 7, 17", + "[73] Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep k-means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In International Conference on Machine Learning, pages 5363-5372. PMLR, 2018. 3", + "[74] Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In USENIX Security Symposium, pages 2237-2254, 2021. 2", + "[75] Chong Xiang, Saeed Mahloujifar, and Prateek Mittal. {PatchCleanser}: Certifiably robust defense against adversarial patches for any image classifier. In 31st USENIX Security Symposium (USENIX Security 22), pages 2065-2082, 2022. 2", + "[76] Zhen Xiang, David Miller, and George Kesidis. Post-training detection of backdoor attacks for two-class and multi-attack scenarios. In International Conference on Learning Representations, 2021. 2", + "[77] Zhen Xiang, Zidi Xiong, and Bo Li. Umd: Unsupervised model detection for x2x backdoor attacks. In International Conference on Machine Learning, 2023. 2", + "[78] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10523-10531, 2021. 2", + "[79] Qiuling Xu, Guanhong Tao, Jean Honorio, Yingqi Liu, Shengwei An, Guangyu Shen, Siyuan Cheng, and Xiangyu Zhang. Medic: Remove model backdoors via importance driven cloning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20485-20494, 2023. 2", + "[80] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. arXiv preprint arXiv:1910.03137, 2019. 2, 7, 16", + "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference, 2016. 6", + "[82] Yi Zeng, Han Qiu, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. arXiv preprint arXiv:2012.07006, 2020. 2", + "[83] Kaiyuan Zhang, Guanhong Tao, Qiuling Xu, Siyuan Cheng, Shengwei An, Yingqi Liu, Shiwei Feng, Guangyu Shen, PinYu Chen, Shiqing Ma, and Xiangyu Zhang. FLIP: A provable defense framework for backdoor mitigation in federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24808", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 14, 21", + "[85] Shihao Zhao, Xingjun Ma, Xiang Zheng, James Bailey, Jingjing Chen, and Yu-Gang Jiang. Clean-label backdoor attacks on video recognition models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14443-14452, 2020. 2", + "[86] Chen Zhu, W Ronny Huang, Hengduo Li, Gavin Taylor, Christoph Studer, and Tom Goldstein. Transferable clean-label poisoning attacks on deep neural nets. In International Conference on Machine Learning, pages 7614-7623. PMLR, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24809", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_model.json b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd937d345ed9e6e32b469c9ead63b15965eec3c3 --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_model.json @@ -0,0 +1,2479 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.862, + 0.154 + ], + "angle": 0, + "content": "LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.181, + 0.782, + 0.238 + ], + "angle": 0, + "content": "Siyuan Cheng, Guanhong Tao, Yingqi Liu†, Guangyu Shen, Shengwei An, Shiwei Feng, Xiangzhe Xu, Kaiyuan Zhang, Shiqing Ma‡, Xiangyu Zhang Purdue University, †Microsoft, ‡University of Massachusetts at Amherst" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.241, + 0.842, + 0.274 + ], + "angle": 0, + "content": "{cheng535, taog, shen447, an93, feng292, xu1415, zhan4057, xyzhang}@cs.purdue.edu †yingqiliu@microsoft.com, ‡shiqingma@umass.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.308, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Backdoor attack poses a significant security threat to Deep Learning applications. Existing attacks are often not evasive to established backdoor detection techniques. This susceptibility primarily stems from the fact that these attacks typically leverage a universal trigger pattern or transformation function, such that the trigger can cause misclassification for any input. In response to this, recent papers have introduced attacks using sample-specific invisible triggers crafted through special transformation functions. While these approaches manage to evade detection to some extent, they reveal vulnerability to existing backdoor mitigation techniques. To address and enhance both evasiveness and resilience, we introduce a novel backdoor attack LOTUS. Specifically, it leverages a secret function to separate samples in the victim class into a set of partitions and applies unique triggers to different partitions. Furthermore, LOTUS incorporates an effective trigger focusing mechanism, ensuring only the trigger corresponding to the partition can induce the backdoor behavior. Extensive experimental results show that LOTUS can achieve high attack success rate across 4 datasets and 7 model structures, and effectively evading 13 backdoor detection and mitigation techniques. The code is available at https://github.com/Megum1/LOTUS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Backdoor attack is a prominent security threat to Deep Learning applications, evidenced by the large body of existing attacks [5, 19, 37, 51, 64] and defense techniques [20, 32, 35, 67, 72]. It injects malicious behaviors to a model such that the model operates normally on clean samples but mis-classifies inputs that are stamped with a specific trigger. A typical way of injecting such malicious behaviors is through data poisoning [1, 19, 39]. This approach introduces a small set of trigger-stamped images paired with the target label" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.31, + 0.892, + 0.354 + ], + "angle": 0, + "content": "into the training data. Attackers may also manipulate the training procedure [10, 45, 46], and tamper with the model's internal mechanisms [37, 41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.895, + 0.522 + ], + "angle": 0, + "content": "The majority of existing attacks rely on a uniform pattern [5, 19, 39, 64] or a transformation function [6, 51] as the trigger. The uniform trigger tends to be effective on any input, which can be detected by existing techniques. For instance, trigger inversion methods [20, 38, 67, 68] aim to reverse engineer a small trigger that can induce the target prediction on a set of inputs. According to the results reported in the literature [20, 61, 67], for a number of attacks, it is feasible to invert a pattern that closely resembles the ground-truth trigger and has a substantially high attack success rate (ASR), hence detecting backdoored models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.704 + ], + "angle": 0, + "content": "Recent studies introduce sample-specific invisible attacks [10, 33, 45, 46] that encourage the model to emphasize the correlation between the trigger and the input sample. Although these attacks effectively evade certain detection methods [20, 67], they are not resilient to backdoor mitigation techniques [32, 35, 72]. For instance, a straightforward approach such as fine-tuning the backdoored model using only \\(5\\%\\) of the training data can significantly reduce ASR. This is due to the fact that imperceptible trigger patterns are not persistent during the retraining process. Moreover, the sample-specific characteristic of these attacks make them less robust to backdoor mitigation methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.896, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce an innovative attack that not only evades backdoor detection approaches but also exhibits resilience against backdoor mitigation techniques. It is a label-specific attack, aiming to misclassify the samples of a victim class to a target class. For the victim-class samples, we divide them into sub-partitions and use a unique trigger for each partition. With such an attack design, existing defense such as trigger inversion is unlikely to find a uniform trigger. This is because the available set of samples used by trigger inversion is likely from different partitions, which makes the detection fail. In addition, we develop a novel trigger focusing technique to ensure that a partition can only be attacked by its designated trigger, not by any" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24798" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "other trigger or trigger combinations. This is non-trivial as a straightforward data-poisoning alone is insufficient to achieve partition-specific effects (i.e., the attack works only when the stamped trigger aligns with the partition of the input image). More details can be found in Section 4. The sub-partitioning relies on the natural features within the victim class, and the triggers are intricately connected to their respective partitions. These two characteristics ensure the connection between inputs and triggers, making our attack more robust against a range of backdoor detection and mitigation techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Our contributions are summarized as follows: (1) We propose a new backdoor attack prototype LOTUS (\"Evasive and ResiLient BackdOor A Tacks through Sub-partitioning\") that achieves both evasiveness and resilience. (2) We address a key challenge of the proposed attack, to precisely limit the scope of a trigger to its partition. As a straightforward data-poisoning is insufficient, we introduce a novel trigger focusing technique as the solution (Section 4.2). (3) We conduct an extensive evaluation of LOTUS on 4 datasets and 7 model structures. Our results show that LOTUS achieves a high ASR under a variety of settings. Our trigger focusing method effectively reduces the ASR on undesired victim classes and partitions. Furthermore, our experiments demonstrate that LOTUS is evasive and resilient against 13 state-of-the-art backdoor defense techniques, substantially outperforming existing backdoor attacks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.473, + 0.715 + ], + "angle": 0, + "content": "Threat Model. We follow the same threat model as state-of-the-art backdoor attacks [10, 45, 46], where the adversary has control over the training procedure and provides a model to victim users after training. The adversary's goal is to achieve high attack effectiveness while also ensuring the attack's evasiveness and resilience against defense. LOTUS primarily focuses on label-specific attack. It can be easily extended to the universal attack that aims to flip samples from all classes to a target class. The defender possesses white-box access to the model and a small set of clean samples for each class. She aims to determine if a model contains backdoor or mitigate the backdoor effects based on the validation samples. In our attack, the sub-partitioning function and the corresponding triggers are the secret of the attacker." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.725, + 0.22, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Backdoor Attack. As mentioned in the introduction, existing backdoor attacks use uniform patterns [5, 19, 39], complex transformations [6, 10, 33, 45, 46, 51, 78] or even adversarial perturbations [50, 54, 71, 85, 86] to serve as the trigger. Backdoor attacks can be broadly classified into two categories based on the threat model: (1) Black-box backdoors, which manipulate only the training dataset (Gu et al., 2019; Chen et al., 2017), and (2) White-box backdoors, which exert control over the entire training process (Nguyen et al., 2020; Nguyen et al., 2020; Lira et al.). Our proposed" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.258 + ], + "angle": 0, + "content": "attack, LOTUS belongs to the white-box backdoor category, aligning with the existing works. Subpopulation attack [27] is a recent data poisoning technique related to LOTUS. It is an availability attack, and its primary objective is to decrease the test accuracy of a specific subpopulation within the dataset. In contrast, LOTUS is a comprehensive backdoor attack with the intention of injecting a backdoor into the model. Therefore, these two attacks differ significantly. Subpopulation attack does not involve trigger injection or require the implementation of trigger focusing, making it distinct from LOTUS in terms of its objectives and mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.897, + 0.427 + ], + "angle": 0, + "content": "Backdoor Defense. Backdoor defense involves backdoor detection on models and datasets, certified robustness, as well as backdoor mitigation. Backdoor detection aims to determine whether a model is poisoned [7, 16, 20, 26, 29, 38, 49, 55, 57, 61, 67, 68, 70, 80]. Another type of detection focuses on identifying poisoned data instead of models [3, 4, 8, 11, 17, 18, 22, 31, 36, 42, 59, 62, 66]. Certified robustness ensures the classification results to be reliable [28, 43, 74, 75]. Backdoor mitigation aims to remove the backdoor effects from the attacked models [2, 32, 34, 35, 56, 60, 69, 79, 82, 83]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.444, + 0.671, + 0.46 + ], + "angle": 0, + "content": "3. Attack Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.895, + 0.594 + ], + "angle": 0, + "content": "We formally define our attack in this section. For a typical classification task, given \\((\\pmb{x},y)\\sim \\mathcal{D}\\) where the sample \\(\\pmb {x}\\in \\mathbb{R}^d\\) and label \\(y\\in \\{1,2,\\dots N\\}\\), the goal is to train a classifier \\(M_{\\theta}:\\mathbb{R}^{d}\\to \\{1,2,\\dots ,N\\}\\), such that parameters \\(\\theta = \\arg \\max_{\\theta}P_{(\\pmb {x},y)\\sim \\mathcal{D}}[M_{\\theta}(\\pmb {x}) = y]\\). Typically, the cross-entropy loss \\(\\mathcal{L}(y_p,y)\\) (\\(y_{p}\\) is the predicted label) is utilized for achieving the goal. In this case, the optimization problem can be expressed as \\(\\theta = \\arg \\min_{\\theta}\\mathbb{E}_{(\\pmb {x},y)\\sim \\mathcal{D}}[\\mathcal{L}(M_{\\theta}(\\pmb {x}),y)]\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.896, + 0.853 + ], + "angle": 0, + "content": "Backdoor attack aims to derive a classifier \\(\\overline{\\mathcal{M}}_{\\bar{\\theta}}:\\mathbb{R}^d\\to\\) \\(\\{1,2,\\dots ,N\\}\\) such that compromised parameters \\(\\bar{\\theta} =\\) arg max \\(\\overline{\\theta} P_{(\\pmb {x},y)\\sim \\mathcal{D}} [\\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\pmb {x}) = y\\& \\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\mathbb{T}\\oplus \\pmb{x}_V) = y_T],\\) in which \\(\\mathbb{T}\\) is the trigger and \\(\\mathbb{T}\\oplus \\pmb{x}_V\\) injects the trigger to a victim input sample \\(x_{V}\\) whose label is \\(y_{V}\\) .Symbol \\(y_{T}\\) denotes the attack target label. Backdoor attacks can be mainly classified to universal attack that aims to flip a sample \\(\\pmb{x}\\) of any class with \\(\\mathbb{T}\\) to the target label, and label-specific attack that aims to flip any sample of a specific victim class to the target label. Based on trigger patterns, they can be classified to input-independent backdoor or static backdoor that has a fixed trigger pattern for all victim samples, and dynamic trigger that has changing patterns for different inputs. Our attack is a label-specific dynamic backdoor attack. Extending to other scenarios is relatively straightforward, e.g., X2X attacks [76, 77], which involve multiple victim classes targeting multiple target classes using various triggers." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.855, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Assume there exists a partitioning algorithm \\(\\mathcal{C}_n: \\mathbb{R}^d \\to \\{p_1, p_2, \\dots, p_n\\}\\) that separates input samples to \\(n\\) partitions. In our attack, victim samples (samples from the victim class)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "are partitioned to \\(n\\) groups using \\(\\mathcal{C}_n\\) and each partition \\(p_i\\) is assigned a unique trigger \\(\\mathbb{T}_i\\), such that only \\(\\mathbb{T}_i \\oplus \\boldsymbol{x}_V^{p_i}\\) can trigger the backdoor, where \\(i \\in \\{1, 2, \\dots, n\\}\\) and \\(\\boldsymbol{x}_V^{p_i}\\) denotes the victim samples in the \\(i\\)-th partition. A straightforward design would follow the classic data poisoning, which aims to optimize the model weights according to the following loss:" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.203, + 0.472, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.473, + 0.58 + ], + "angle": 0, + "content": "The first loss term Benign Utility Loss aims to ensure the high benign accuracy of the model. The second term, Attack Target Loss, means that a trigger \\(\\mathbb{T}_i\\) can cause the \\(i\\)-th partition samples of the victim class \\(\\pmb{x}_{V}^{p_i}\\) to misclassify, which is our attack goal. However, simple data poisoning cannot effectively bound the attack scope. As a result, a trigger for a particular partition can easily induce misclassifications for other partitions. That is, \\(\\mathbb{T}_j \\oplus \\pmb{x}_{V}^{p_i}\\), where \\(i \\neq j\\), is misclassified to \\(y_{T}\\). Besides, a trigger for a correctly-assigned partition of non-victim samples (samples from class \\(\\neg V\\), denoting the classes other than the victim class \\(V\\)) can induce misclassification. That is \\(\\mathbb{T}_i \\oplus \\pmb{x}_{\\neg V}^{p_i}\\) is misclassified to \\(y_{T}\\). Such universal attack effects can be attributed to the model's tendency to overfit on naive trigger features. For instance, when it encounters any trigger, it immediately predicts the target class without verifying if the background image aligns with the trigger according to the partitioning criteria. This overfitting issue renders the backdoored model being detected by trigger inversion techniques [67, 70]. Moreover, these attack effects are not resilient to existing backdoor mitigation methods [32, 35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.472, + 0.687 + ], + "angle": 0, + "content": "Our objective is to establish a clear one-to-one correspondence between \\(\\mathbb{T}_i\\) and \\(\\pmb{x}_{V}^{p_i}\\). That is, only \\(\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}\\) can cause misclassification. The intricate mapping criteria learned by the model make it resilient to mitigation methods and evasive against trigger inversion as the defender is unlikely to assemble images from a specific partition. We hence aim to derive the following loss function." + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.697, + 0.498, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {(\\pmb {x}, y) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\pmb {x}), y) ] + \\sum_ {i = 1} ^ {n} (\\mathbb {E} _ {(\\pmb {x} _ {V} ^ {p _ {i}}, y _ {V}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\mathbb {T} _ {i} \\oplus \\pmb {x} _ {V} ^ {p _ {i}}), y _ {T}) ] \\\\ + \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}, y _ {\\neg V}\\right) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left(\\mathbb {T} _ {i} \\oplus \\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}\\right), y _ {\\neg V}) ] \\longleftrightarrow \\text {L a b e l - s p e c i f i c L o s s} \\\\ + \\underbrace {\\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}} , y _ {V}\\right) \\sim \\mathcal {D}} \\left[ \\right. \\sum_ {\\substack {\\mathcal {T} \\in \\mathcal {P} \\left( \\right.\\{\\mathbb {T} _ {1} , \\dots , \\mathbb {T} _ {n} \\left. \\right\\}}} \\mathcal {L} \\left(\\overline {\\mathcal {M}} _ {\\bar {\\theta}} \\left(\\mathcal {T} \\oplus \\boldsymbol {x} _ {p _ {i}}\\right), y _ {V}\\right)\\left. \\right]\\left. \\right]} _ {- \\{\\{\\}, \\{\\mathbb {T} _ {i} \\} \\}} \\xleftarrow {} \\text {D y n a m i c L o s s} \\tag{2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Note that compared to Equation 1, we introduce two additional terms, i.e., Label-specific Loss and Dynamic Loss in Equation 2. Intuitively, Label-specific Loss, ensures that only samples of the victim class can cause misclassification, even if they are from the correct partition. Here \\(\\neg V\\) denotes the classes other than the victim class. The last term Dynamic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "Loss controls that for a particular partition, only the corresponding trigger can cause misclassification, and any other trigger, or combination of/with other triggers shall be correctly predicted as the victim class. In particular, \\(\\mathcal{T}\\) is a subset of all possible triggers/combinations \\(\\mathcal{P}(\\{\\mathbb{T}_1,\\dots ,\\mathbb{T}_n\\})\\) excluding empty \\(\\{\\}\\) and \\(\\{\\mathbb{T}_i\\}\\). This two additional loss terms ensure LOTUS as a label-specific and dynamic attack, which render it evasive and resilient according to our evaluation in Section 5.3 and 5.4." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.245, + 0.719, + 0.263 + ], + "angle": 0, + "content": "4. Detailed Attack Design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.271, + 0.895, + 0.422 + ], + "angle": 0, + "content": "The overview of LOTUS is shown in Figure 1. Victim class input samples are first separated to partitions. We then apply unique triggers to samples from the corresponding partitions, whose labels are set to the target class. Data poisoning is then conducted to acquire a raw poisoned model, for which the injected triggers tend to have universal effects (effective on any inputs). To address this problem, LOTUS further introduces a trigger focusing step that strictly limits the attack scope of each trigger. It finally produces a trojaned model with triggers that are evasive and resilient." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.895, + 0.47 + ], + "angle": 0, + "content": "In the following, we elaborate two major components of LOTUS, namely, victim-class sample partitioning and trigger focusing." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.482, + 0.792, + 0.5 + ], + "angle": 0, + "content": "4.1. Victim-class Sample Partitioning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.895, + 0.703 + ], + "angle": 0, + "content": "LOTUS separates a set of victim-class samples into multiple partitions, and injects different triggers to different partitions. We propose two ways to partition input samples. The first is explicit partitioning that leverages a subset of explicit attributes of the victim class (e.g., hair color and w./ or w./o. glasses for face recognition). Assume \\( k \\) attributes are used and each attribute has \\( t \\) possible values. This allows to generate \\( t^k \\) partitions. The first two columns in Figure 2 show a partitioning based on the taxonomy attribute of the bird class. Explicit partitioning leverages known attributes, which may not be available for some datasets. We hence introduce an advanced partitioning method that is applicable to arbitrary datasets in the following." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.897, + 0.902 + ], + "angle": 0, + "content": "The second partitioning scheme is implicit, meaning that human uninterpretable features are used in partitioning. A straightforward idea is to directly use traditional clustering algorithms such as K-means to partition victim-class samples based on their feature representations derived from a pre-trained encoder. However, according to our experiment in Appendix L.1, such a naive method does not work well. The root cause is that K-means is a clustering algorithm on a set of known data points and does not consider generalization to unseen data points. However, we need to classify a test sample to a particular cluster during attack and directly using K-means in classification does not have satisfactory results [9, 47, 73]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.088, + 0.852, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.227, + 0.576, + 0.24 + ], + "angle": 0, + "content": "Figure 1. Overview of LOTUS" + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.265, + 0.411, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.43, + 0.438, + 0.444 + ], + "angle": 0, + "content": "Figure 2. Explicit (left) and implicit (right) partitioning." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.472, + 0.469, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.612, + 0.428, + 0.625 + ], + "angle": 0, + "content": "Figure 3. Implicit partitioning with surrogate model." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We hence introduce a surrogate model to help sample partitioning. Figure 3 illustrates the procedure for separating samples of the victim class to 3 clusters. The surrogate model has the same structure as the victim model to reduce complexity caused by structural differences. On the bottom left, the features of samples from victim class \\( n \\) are extracted using a pre-trained encoder. We then use a traditional clustering method such as K-means to partition these samples into 3 different sub-classes based on their features. We assign labels \\( n \\), \\( n + 1 \\), \\( n + 2 \\) to samples from the respective sub-classes. They are then combined with samples from the original classes 1 to \\( n - 1 \\) (excluding the victim class \\( n \\)) to form a new dataset consisting of \\( n + 2 \\) classes. The surrogate model is trained on this new dataset with \\( n + 2 \\) classes. The idea is to use K-means to provide a meaningful prior separation and then use classifier training to achieve generalizability. Furthermore, the decision boundaries by the surrogate model have the classes other than the victim class" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.895, + 0.435 + ], + "angle": 0, + "content": "in consideration, whereas those by distances to centroids of K-means clusters only have samples of the victim class in consideration. After the training converges, the surrogate model is utilized to determine the partition of a test sample. That is, the partition index can be derived from the its classification outcome (i.e., the class with largest logits from classes \\( n \\) to \\( n + 2 \\)). The last two columns in Figure 2 show two implicit partitions of the \"cat\" class. Observe that the partitions are largely uninterpretable, which makes the attack more stealthy compared to using explicit attributes which are public." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Handle Potential Imbalanced Examples. We control that for any partitioning, the sizes of each partition are roughly the same, which mitigates the potential of causing partitioning bias. This is achieved by removing samples from exceptionally large clusters. In practice, such a removal is rarely needed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.539, + 0.669, + 0.556 + ], + "angle": 0, + "content": "4.2. Trigger Focusing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.623 + ], + "angle": 0, + "content": "After partitioning, LOTUS aims to limit each trigger to its own partition, preventing it from attacking other partitions or classes. To achieve this, we design a trigger focusing technique during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.776 + ], + "angle": 0, + "content": "A straightforward idea is to strictly follow the definition in Equation 2 to bound the trigger scope. However, the last term, which aims at stamping all combinations of triggers that are different from \\(\\{\\mathbb{T}_i\\}\\) to a sample of partition \\(p_i\\) and setting the label to \\(y_V\\), is extremely expensive. The number of combinations is \\((2^n - 2)\\), which grows exponentially with the increase of the number of partitions \\(n\\). Moreover, the inclusion of a substantial number of additional samples will not only slow down the training but also imbalance the dataset, ultimately impacting the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Adversarial Poisoning Is Insufficient. Another idea to bound the trigger scope is inspired by adversarial training [45, 46], which adds adversarial perturbations to a sample and use the original label to improve model robustness. To suppress the undesirable attack effect in our context, we could inject triggers that are not for a partition \\( p_i \\), i.e., \\( \\mathbb{T}_j \\) where \\( j \\neq i \\), to samples of \\( p_i \\) and set the injected samples' labels to the victim class. This approach is referred" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.091, + 0.46, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.091, + 0.773, + 0.116 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.121, + 0.463, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.467, + 0.123, + 0.874, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.154, + 0.553, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.178, + 0.342, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.282, + 0.262, + 0.295 + ], + "angle": 0, + "content": "(a) Data-poisoning" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.179, + 0.487, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.179, + 0.616, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.282, + 0.555, + 0.295 + ], + "angle": 0, + "content": "(b) Adversarial-poisoning" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.179, + 0.763, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.179, + 0.892, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.282, + 0.815, + 0.295 + ], + "angle": 0, + "content": "(c) Trigger-focusing" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.297, + 0.674, + 0.311 + ], + "angle": 0, + "content": "Figure 4. Decision boundaries for different poisoning strategies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.401 + ], + "angle": 0, + "content": "to as adversarial poisoning. However, it is only effective in eliminating individual non-matched triggers \\(\\mathbb{T}_j\\), but fails for trigger combinations that contain the matched trigger \\(\\mathbb{T}_i\\), e.g., \\([\\mathbb{T}_i,\\mathbb{T}_j]\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Figure 4 presents a visualization of decision boundaries for various poisoning strategies, namely: (a) Straightforward data-poisoning; (b) Adversarial-poisoning; and (c) Trigger-focusing (which will be discussed in the next paragraph). Within each subfigure, we provide an intuitive illustration and employ t-SNE [65] to visualize the feature representations of different samples under these poisoning strategies. The experiment is conducted on the CIFAR-10 dataset using the ResNet18 model, and we utilize implicit partitioning to create four distinct partitions. In the figure, a hollow closed lock is used to denote clean images \\( x_{v}^{p_{i}} \\) in the victim class of partition \\( p_{i} \\), while a red opened lock is used to represent clean images of the target class. Triggers are depicted as keys with various colors. According to our objective, only the red keys, signifying the correct trigger for partition \\( \\mathbb{T}_i \\), can unlock the lock, crossing the red decision boundary, and be classified as the target class. Keys of different colors, signifying various triggers or combinations, are unable to unlock the lock and remain within the victim class region. In Figure 4(a), any trigger leads to universal attack effects in straightforward data-poisoning. Observe any key, denoting a trigger, can unlock the lock and cross the boundary without limitations. The t-SNE visualization on real data on the right aligns with the illustration on the left. In contrast, adversarial-poisoning, as depicted in (b), mitigates the impact of samples with unmatched individual triggers, as represented by the green key. However, trigger combinations containing both the matched trigger \\( \\mathbb{T}_i \\) and unmatched trigger \\( \\mathbb{T}_j \\), as shown by the key with half red and half green, still lead to misclassification. Similarly, in the t-SNE visualization, the yellow triangles, which represent this type of trigger combination, are substantially close to the red triangles, denoting the strictly matched triggers. This indicates" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.339, + 0.78, + 0.355 + ], + "angle": 0, + "content": "the insufficiency of adversarial-poisoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.895, + 0.51 + ], + "angle": 0, + "content": "Efficient and Effective Trigger Focusing. Inspired by the observation in Figure 4, we propose a novel trigger focusing method that can effectively bound trigger scopes and is in the mean time cost-effective. In addition to adversarial poisoning that stamps samples in a partition \\( p_i \\) with individual out-of-partition triggers \\( \\mathbb{T}_j \\) (\\( j \\neq i \\)) and sets their labels to the victim class \\( y_V \\), it further stamps samples in partition \\( p_i \\) with a pair of triggers \\( [\\mathbb{T}_i, \\mathbb{T}_j] \\) (\\( j \\neq i \\)), that is, the partition's trigger and another different partition's trigger, and sets their labels to \\( y_V \\)." + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.533, + 0.892, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}}, y _ {V}\\right) \\sim \\mathcal {D}} [ \\sum_ {j = 1, j \\neq i} ^ {n} (\\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} (\\mathbb {T} _ {j} \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}), y _ {V}) \\tag {3} \\\\ \\left. + \\mathcal {L} \\left(\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left([ \\mathbb {T} _ {i}, \\mathbb {T} _ {j} ] \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}\\right), y _ {V})\\right) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.894, + 0.68 + ], + "angle": 0, + "content": "Our approach, with the new dynamic loss term expressed in Equation 3, requires only \\((2n - 2)\\) trigger combinations, which increases linearly with the growth of partitions \\(n\\). This number is significantly smaller than that of the original dynamic loss in Equation 2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.68, + 0.894, + 0.77 + ], + "angle": 0, + "content": "Intuitively, the different labels of samples \\(\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}\\) and \\([\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}\\) enable the model to learn new behaviors. As such, further stamping any other partition triggers to \\([\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}\\) yields the same classification result, which is the victim class. Please refer to Appendix D for a detailed reasoning and theoretical analysis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.772, + 0.894, + 0.832 + ], + "angle": 0, + "content": "In Figure 4(c), it is noteworthy that trigger combinations are effectively excluded from the target class and only the trigger that matches the victim partition can cause the misclassification, well aligning with our attack objective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.617, + 0.861 + ], + "angle": 0, + "content": "5. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.9 + ], + "angle": 0, + "content": "In this section, we evaluate on 4 benchmark datasets and 7 model structures to demonstrate the attack effectiveness of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24802" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "LOTUS (Section 5.2). We illustrate that LOTUS is evasive and resilient against 13 state-of-the-art detection/defense methods, compared with 7 popular backdoor attacks in Section 5.3 and 5.4. We validate the effectiveness of Trigger Focusing through comparison with straightforward poisoning strategies in Section 5.5. We also extend LOTUS to universal attacks in Appendix G. Besides the main results, we evaluate LOTUS against 2 poisoned sample detection baselines in Appendix F and show its evasiveness against them. Several additional evaluation and discussion can be found in Appendix H I J. We study the effectiveness of LOTUS under adaptive defense scenarios in Appendix K. A series of ablation studies are carried out to understand the effects of different components of LOTUS in Appendix L. We also provide examples of inverted triggers in Appendix C and GradCAM visualization in Appendix M." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.344, + 0.254, + 0.36 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.472, + 0.488 + ], + "angle": 0, + "content": "We evaluate LOTUS on 4 widely-used benchmarks, CIFAR-10 [30], CIFAR-100 [30], CelebA [40], and restricted ImageNet (RImageNet) [12, 52, 63]. Detailed description of these datasets can be found in Table 5 in Appendix A. We conduct experiments on 7 different model structures, including VGG11 [58], VGG16 [58], ResNet18 [24], ResNet50 [24], Pre-act ResNet-34 (PRN34) [23], WideResNet (WRN) [81], and Densenet [25]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.641 + ], + "angle": 0, + "content": "We leverage several sub-partitioning methods to partition samples from the victim class. We utilize secondary labeling, e.g., various cat species, to create clear and explicit partitions. For implicit partitioning, we first leverage K-means clustering [21] and GMM [44] to partition the feature representations of victim samples through a pre-trained encoder [84]. Then we train a surrogate model to learn the partitioning principle, which serves as the implicit sub-partitioner (Section 4.1). Details of the sub-partitioning and encoder can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.27, + 0.667 + ], + "angle": 0, + "content": "5.2. Attack Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.901 + ], + "angle": 0, + "content": "We evaluate the performance of LOTUS on various datasets, model structures and partitioning methods. Table 1 presents the results. For all the experiments, we use the first class of each dataset as the victim and the last class as the target. We generate 4 partitions for the victim class throughout all datasets and model structures. Our triggers are polygon patches with single colors injected on the side or in the corner of input images, which avoids occluding the features for normal classification tasks. Example images with triggers can be found in Figure 10 in Appendix. The top two blocks in Table 1 (separated by the double lines) show the results for implicit partitioning, and the bottom for explicit partitioning. For K-means clustering, ASRs are at least \\(89.00\\%\\), with the highest ASR of \\(94.30\\%\\) for ResNet18 on CIFAR-10, while the degradation of benign accuracy is within \\(1.07\\%\\). This" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.188 + ], + "angle": 0, + "content": "Table 1. Evaluation of attack effectiveness. The first three columns denote different partitioning algorithms (PA), datasets, and model structures. The following columns present the original accuracy of clean models (Acc.), benign accuracy of the backdoored models (BA), the attack success rate when stamping a trigger on the proper partition (ASR), and the average ASR when stamping other triggers and trigger combinations, with the standard deviation) (ASR-other)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.189, + 0.892, + 0.369 + ], + "angle": 0, + "content": "
PADatasetModelAcc.BAASRASR-other
K-meansCIFAR-10VGG1192.16%92.04%93.80%4.77% ± 19.27%
ResNet1895.22%94.71%94.30%4.39% ± 17.08%
CIFAR-100Densenet75.14%75.15%92.00%4.36% ± 14.24%
PRN3474.70%74.52%89.00%5.43% ± 13.50%
CelebAWRN80.47%79.40%92.33%6.87% ± 17.49%
RImageNetResNet5097.77%97.19%93.87%2.16% ± 19.34%
GMMCIFAR-10ResNet1895.22%94.59%90.70%4.80% ± 21.38%
CIFAR-100PRN3474.70%74.02%91.00%2.21% ± 12.57%
CelebAWRN80.47%79.66%92.53%5.39% ± 16.77%
RImageNetVGG1696.51%95.93%93.52%3.11% ± 14.39%
Sec.RImageNetVGG1696.51%96.36%96.50%1.79% ± 13.24%
ResNet5097.77%97.08%92.50%2.14% ± 16.53%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Table 2. Evaluation of label specificity. ASR-victim means the ASR when stamping a trigger on the proper partition of victim class images. ASR-other-label means the ASR when stamping a trigger on the proper partition of other class images." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.44, + 0.862, + 0.506 + ], + "angle": 0, + "content": "
DatasetNetworkASR-victimASR-other-label
CIFAR10ResNet1893.80%14.37%
CIFAR100Densenet92.00%11.23%
CelebAWRN92.33%19.67%
RImageNetVGG1693.52%12.22%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.765 + ], + "angle": 0, + "content": "indicates LOTUS is a highly effective attack, which injects successful malicious behaviors to the model while maintains its benign utility. The last column shows the ASR when trigger/trigger-combinations other than a partition's trigger are stamped on the partition (ASR-other). Observe that the average ASR-other is less than \\(6.87\\%\\), delineating the effectiveness of trigger focusing (a trigger is only effective for the corresponding partition). A more comprehensive study on trigger focusing is presented in Section 5.5. We have similar observations for using GMM in implicit partitioning. For the explicit secondary labeling, LOTUS can achieve an ASR over \\(92.50\\%\\) and a small ASR-other. The better performance of LOTUS using secondary labeling can be attributed to the fact that the victim class in RImageNet is merged from a set of similar classes in ImageNet. Those classes are naturally separable, which can be easily differentiated by the model when triggers are injected on different partitions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Note that the ASR of LOTUS is slightly lower than the existing attacks (as shown in the \"No Defense\" column in Table 3). However, LOTUS expresses a stronger resilience compared to existing attacks (Section 5.4). This is a trade-off between attack effectiveness and resilience. More discussion can be found in Appendix J." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Besides, we also evaluate the label specificity of LOTUS on several models. Results are presented in Table 2. Observe that even if the trigger is stamped on the proper partition" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24803" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "of the input image, the ASR-other-label is low (\\(< 20\\%\\)) because the input image is not of the victim class. The result shows that LOTUS exhibits a high level of label specificity. Furthermore, LOTUS offers an easy extension into universal attack scenarios through the integration of explicit partitioning techniques. Detailed examples can be found in Section G." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.21, + 0.421, + 0.226 + ], + "angle": 0, + "content": "5.3. Evasiveness against Backdoor Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.471, + 0.49 + ], + "angle": 0, + "content": "In this section, we study the evasiveness of LOTUS against 4 well-known trigger-inversion based backdoor detection methods, including Neural Cleansse (NC) [67], Pixel [61], ABS [38], and FeatRE [70]. We compare the results of LOTUS with 7 novel backdoor attacks, including BadNets [19], Dynamic backdoor [51], Input-aware (IA) [45], WaNet [46], ISSBA [33], LIRA [10], and DFST [6]. For fair comparison, we launch all backdoor attacks on ResNet18 models trained on CIFAR-10. As LOTUS is a label-specific attack, we implement all other attacks in label-specific setting, where the poisoned samples are composed of images from victim class 0 stamped with the trigger and labeled as the target class 9. Besides, all detection methods are required to invert triggers based on 100 clean validation images from the victim class, targeting to labels other than it. We follow all the other settings and techniques of the original papers to implement the attack and detection methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Figure 5 illustrates the detection results, where the x-axis denotes different attacks and the y-axis denotes the decision scores of each baseline. The thresholds are highlighted in red dashed lines. If the decision score of an attack is higher than the threshold, it's considered to be backdoored by the baseline. Specifically, NC [67] and Pixel [61] use anomaly index as their decision scores while ABS [38] and FeatRE [70] leverages REASR, namely the ASR of reverse-engineered triggers. Observe that NC, Pixel, ABS are effective against several attacks, including BadNets, Dynamic, ISSBA, LIRA and DFST, while leaving other advanced attacks, i.e., WaNet, IA and LOTUS. FeatRE, on the other hand, observes internal linear separability properties of existing backdoors and improves the trigger inversion process, which is able to detect the advanced backdoors operating in the feature space. Figure 5(d) shows that it can detect both IA and WaNet, but still fails to detect LOTUS. This illustrates that LOTUS is more evasive than all these baseline attacks. The underlying reason is that LOTUS leverages partitioning secrets and trigger focusing, which breaks the linear separability assumption. Without knowledge of partitioning, it's unlikely to invert a trigger with high ASR, and hence unlikely to detect the backdoor. Examples of inverted triggers can be found in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.901 + ], + "angle": 0, + "content": "We also test LOTUS in the adaptive defense scenario, where the defender can create partitions before detection. The results in Appendix K demonstrate that LOTUS is re" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "silent against adaptive defense strategies, as guessing the correct partitioning is challenging." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.894, + 0.198 + ], + "angle": 0, + "content": "Besides trigger inversion methods, we also evaluate LOTUS using meta-classifiers, e.g., MNTD [80] and ULP [29], which train model-level classifiers for detection. Results in Appendix E show that LOTUS is evasive against them." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.209, + 0.837, + 0.224 + ], + "angle": 0, + "content": "5.4. Resilience against Backdoor Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.895, + 0.383 + ], + "angle": 0, + "content": "In this section, we study the resilience of LOTUS against 4 state-of-the-art backdoor mitigation methods, including standard Fine-tuning, Fine-pruning [35], NAD [32], and ANP [72]. We compare the results of LOTUS with 7 novel backdoor attacks. For fair comparison, all the models are trained using VGG11 on CIFAR-10 dataset. For each mitigation method, we assume the access to \\(5\\%\\) of the training data. Besides, some standard input argumentation techniques are used, e.g., random cropping and horizontal flipping. We follow the original setting to conduct these baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Table 3 provides the result. Observe that for all the baselines, benign accuracy change is slight, meaning that the mitigation preserves the model utility on benign tasks. However, ASR degradation is considerable for all backdoored models. Note that LOTUS can still remain part of the attack effectiveness with \\(34.90\\% - 46.90\\%\\), outperforming all other attacks. The result indicates that LOTUS is more resilient against baseline mitigation methods compared to the existing attacks. This can be attributed to the design that LOTUS learns the correlation between the partitions and triggers which is hard to unlearn. Other attacks only learn partial trigger patterns that tend to be mitigate." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.575, + 0.877, + 0.591 + ], + "angle": 0, + "content": "5.5. Evaluation on Different Poisoning Strategies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We evaluate different poisoning strategies including simple data poisoning, adversarial poisoning, and LOTUS's trigger focusing. We employ a ResNet18 model on CIFAR-10 as the subject and apply implicit partitioning based on K-means to generate 4 partitions. The number of possible non-empty trigger combinations is \\(2^{4} - 1 = 15\\). In the following, we use a four-bit binary to represent each combination. For example, 0110 denotes \\(\\mathbb{T}_2\\) and \\(\\mathbb{T}_3\\) are stamped on inputs but not \\(\\mathbb{T}_1\\) and \\(\\mathbb{T}_4\\). Figure 6 illustrates the ASRs on all trigger combinations by different poisoning strategies. Sub-figures from left to right present the results for simple poisoning, adversarial poisoning, and trigger focusing, respectively. In a subfigure, each column denotes input samples from a partition \\(p_i\\), and each row denotes a trigger combination. The value in each cell shows the ASR when a trigger combination (row) is stamped on the samples from a partition (column). Brighter the color, higher the ASR. The left sub-figure shows the ASR for simple data poisoning. Observe that all the ASRs are greater than \\(92.0\\%\\) (with an average of \\(97.94\\%\\)), showing the sub-partitioning is not learned by the model. The middle" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24804" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.092, + 0.29, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.092, + 0.51, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.096, + 0.74, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.097, + 0.885, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.254, + 0.892, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Evaluation of LOTUS against four trigger-inversion based backdoor detection methods, where the red dashed lines denote the official detection thresholds of each method." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.308, + 0.529, + 0.364 + ], + "angle": 0, + "content": "Table 3. Evaluation of resilience against backdoor mitigation methods. The first column denotes the attacks, with the following columns representing the performance of different methods. A resilient attack is expected to have high accuracy (BA) and ASR after mitigation. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.365, + 0.528, + 0.493 + ], + "angle": 0, + "content": "
AttacksNo DefenseFine-tuningFine-pruningNADANP
BAASRBAASRBAASRBAASRBAASR
BadNets92.02%100.00%89.31%1.74%91.70%0.53%87.81%0.80%89.15%0.32%
Dynamic91.81%100.00%88.87%2.91%91.39%22.03%89.11%2.90%88.25%12.81%
IA91.70%99.65%87.74%2.78%91.07%0.17%87.14%2.29%88.73%1.98%
WaNet91.22%98.57%89.56%1.37%90.22%1.07%89.74%1.40%89.07%0.54%
ISSBA91.67%99.96%87.73%2.72%91.12%14.27%87.97%2.83%85.64%10.01%
LIRA91.70%100.00%89.96%2.19%91.29%12.14%90.23%2.32%89.70%37.91%
DFST91.81%99.97%88.49%22.86%91.47%21.61%88.52%24.66%87.13%36.17%
LOTUS91.54%93.80%88.10%46.90%91.14%44.90%87.61%42.30%88.14%34.90%
" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.301, + 0.887, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.473, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Figure 6. ASR on all trigger combinations by different poisoning strategies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.473, + 0.892 + ], + "angle": 0, + "content": "sub-figure is the results for adversarial poisoning. We can see around half of cells have small values, especially for single trigger combinations (the top four rows). For more complex trigger combinations, the ASRs are still high with the highest of \\(100.0\\%\\) (trigger combination 0111 on partition \\(p_2\\)), indicating the insufficiency of adversarial poisoning. The right sub-figure is for our trigger focusing. Observe that except for stamping a trigger on the proper partition, the other cases all have a low ASR with an average of \\(3.04\\%\\). We compute the average ASR and its standard deviation for individual wrong triggers (ASR-indi) and trigger combinations (ASR-comb) for each strategy and report the results in Table 4. Observe that all the ASRs are almost \\(100\\%\\) for simple poisoning. Adversarial poisoning reduces the ASR-indi to a low level while leaving ASR-comb high (\\(73.88\\%\\) on average). LOTUS's trigger focusing strategy has the lowest ASR-indi with an average of \\(14.15\\%\\) and ASR-comb \\(0.02\\%\\). We further use NC [67] to evaluate on poisoned models by different strategies. The last column shows the anomaly index for different poisoned models. Observe that models poisoned by simple data poisoning and adversarial poisoning can be easily detected by NC (with anomaly index \\(>2\\)). Poisoned models by trigger focusing, on the other hand, are able to evade NC's detection, delineating the effectiveness" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.528, + 0.852, + 0.541 + ], + "angle": 0, + "content": "Table 4. Evaluation on different poisoning strategies" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.542, + 0.892, + 0.599 + ], + "angle": 0, + "content": "
StrategyBAASRASR-indiASR-combNC index
Simple94.79%98.80%97.86% ± 1.84%97.88% ± 1.81%5.338
Adv.94.47%94.20%18.95% ± 10.22%73.88% ± 31.87%2.161
Focus94.71%91.40%14.15% ± 7.46%0.02% ± 0.09%1.156
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.601, + 0.836, + 0.615 + ], + "angle": 0, + "content": "of trigger focusing strategy to achieve evasiveness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.627, + 0.619, + 0.643 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.743 + ], + "angle": 0, + "content": "We propose a novel backdoor attack that leverages subpartitioning to restrict the attack scope. A special training method is designed to limit triggers to only their corresponding partitions. Our evaluation shows that the attack is highly effective, achieving high attack success rates. Besides, it is evasive and resilient against state-of-the-art defenses." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.669, + 0.772 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for their constructive comments. We are grateful to the Center for AI Safety for providing computational resources. This research was supported, in part by IARPA TrojAI W911NF-19-S0012, NSF 1901242 and 1910300, ONR N000141712045, N000141410468 and N000141712947. Any opinions, findings, and conclusions in this paper are those of the authors only and do not necessarily reflect the views of our sponsors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24805" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. arXiv preprint arXiv:2005.03823, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Eitan Borgnia, Valeriia Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong data augmentation sanitizes poisoning and backdoor attacks without an accuracy tradeoff. arXiv preprint arXiv:2011.09527, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Alvin Chan and Yew-Soon Ong. Poison as a cure: Detecting & neutralizing variable-sized backdoor attacks in deep neural networks. arXiv preprint arXiv:1911.08040, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1148-1156, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.471, + 0.527 + ], + "angle": 0, + "content": "[7] Siyuan Cheng, Guanhong Tao, Yingqi Liu, Shengwei An, Xiangzhe Xu, Shiwei Feng, Guangyu Shen, Kaiyuan Zhang, Qiuling Xu, Shiqing Ma, et al. Beagle: Forensics of deep learning backdoor attack for better defense. arXiv preprint arXiv:2301.06241, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.471, + 0.584 + ], + "angle": 0, + "content": "[8] Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinel: Detecting localized universal attack against deep learning systems. Proceeding of the 41th IEEE Symposium on Security & Privacy Workshops (SPW), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.471, + 0.641 + ], + "angle": 0, + "content": "[9] Ryan Cohn and Elizabeth Holm. Unsupervised machine learning via transfer learning and k-means clustering to classify materials image data. Integrating Materials and Manufacturing Innovation, 10(2):231-244, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11966-11976, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.471, + 0.755 + ], + "angle": 0, + "content": "[11] Min Du, Ruoxi Jia, and Dawn Song. Robust anomaly detection and backdoor attack detection via differential privacy. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[12] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Learning perceptually-aligned representations via adversarial robustness. arXiv preprint arXiv:1906.00945, 2(3):5, 2019. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Wang et al. An invisible black-box backdoor attack through frequency domain. In ECCV, 2022. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[14] Xia et al. Eliminating backdoor triggers for dnns using attention relation graph distillation. In *IJCAI*, 2022. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Zeng et al. Adversarial unlearning of backdoors via implicit hypergradient. In ICLR, 2022. 17" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[16] Shiwei Feng, Guanhong Tao, Siyuan Cheng, Guangyu Shen, Xiangzhe Xu, Yingqi Liu, Kaiyuan Zhang, Shiqing Ma, and Xiangyu Zhang. Detecting backdoors in pre-trained encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16352-16362, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[17] Hao Fu, Akshaj Kumar Veldanda, Prashanth Krishnamurthy, Siddharth Garg, and Farshad Khorrami. Detecting backdoors in neural networks using novel feature-based anomaly detection. arXiv preprint arXiv:2011.02526, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019. 2, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 1, 2, 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[20] Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting and eliminating trojan backdoors in deep neural networks. In 20th IEEE International Conference on Data Mining, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[21] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[22] Jonathan Hayase, Weihao Kong, Raghav Somani, and Swoong Oh. Spectre: defending against backdoor attacks using robust statistics. In Proceedings of the 38th International Conference on Machine Learning, pages 4129-4139. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. European conference on computer vision, pages 630-645, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.894, + 0.608 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.894, + 0.664 + ], + "angle": 0, + "content": "[25] Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.894, + 0.706 + ], + "angle": 0, + "content": "[26] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.894, + 0.762 + ], + "angle": 0, + "content": "[27] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[28] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 2, 7, 16" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24806" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[31] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.218 + ], + "angle": 0, + "content": "[32] Yige Li, Nodens Koren, Lingjuan Lyu, Xixiang Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[34] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[35] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[36] Yuntao Liu, Yang Xie, and Ankur Srivastava. Neural trojans. In 2017 IEEE International Conference on Computer Design (ICCD), pages 45-48. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[37] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In NDSS, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[38] Yingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma, Yousra Aafer, and Xiangyu Zhang. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 1265–1282, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[39] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[40] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 3730-3738, 2015. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[41] Peizhuo Lv, Chang Yue, Ruigang Liang, Yunfei Yang, Shengzhi Zhang, Hualong Ma, and Kai Chen. A data-free backdoor injection approach in neural networks. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2671-2688, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[42] Shiqing Ma, Yingqi Liu, Guanhong Tao, Wen-Chuan Lee, and Xiangyu Zhang. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[43] Michael McCoyd, Won Park, Steven Chen, Neil Shah, Ryan Roggenkemper, Minjune Hwang, Jason Xinyu Liu, and David Wagner. Minority reports defense: Defending against adversarial patches. In Applied Cryptography and Network Security Workshops: ACNS 2020 Satellite Workshops, AIBlock," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "AIHWS, AIoTS, Cloud S&P, SCI, SecMT, and SiMLA, Rome, Italy, October 19-22, 2020, Proceedings, pages 564-582. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[44] Geoffrey J McLachlan and Kaye E Basford. Mixture models: Inference and applications to clustering. M. Dekker New York, 1988. 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[45] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33, 2020. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.26 + ], + "angle": 0, + "content": "[46] Tuan Anh Nguyen and Anh Tuan Tran. Wanet-imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2020. 1, 2, 4, 7, 13, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[47] Jun Ogasawara, Satoru Ikenoue, Hiroko Yamamoto, Motoshige Sato, Yoshifumi Kasuga, Yasue Mitsukura, Yuji Ikegaya, Masato Yasui, Mamoru Tanaka, and Daigo Ochiai. Deep neural network-based classification of cardiotocograms outperformed conventional algorithms. Scientific reports, 11 (1):13367, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.894, + 0.4 + ], + "angle": 0, + "content": "[48] Xiangyu Qi, Tinghao Xie, Yiming Li, Saeed Mahloujifar, and Prateek Mittal. Revisiting the assumption of latent separability for backdoor defenses. In The eleventh international conference on learning representations, 2022. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.455 + ], + "angle": 0, + "content": "[49] Ximing Qiao, Yukun Yang, and Hai Li. Defending neural backdoors via generative distribution modeling. In Advances in Neural Information Processing Systems, pages 14004-14013, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.894, + 0.51 + ], + "angle": 0, + "content": "[50] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11957-11965, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.894, + 0.566 + ], + "angle": 0, + "content": "[51] Ahmed Salem, Rui Wen, Michael Backes, Shiqing Ma, and Yang Zhang. Dynamic backdoor attacks against machine learning models. arXiv preprint arXiv:2003.03675, 2020. 1, 2, 7, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[52] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. arXiv preprint arXiv:1906.09453, 2019. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.622, + 0.894, + 0.692 + ], + "angle": 0, + "content": "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.894, + 0.762 + ], + "angle": 0, + "content": "[54] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems, pages 6103-6113, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.894, + 0.831 + ], + "angle": 0, + "content": "[55] Guangyu Shen, Yingqi Liu, Guanhong Tao, Shengwei An, Qiuling Xu, Siyuan Cheng, Shiqing Ma, and Xiangyu Zhang. Backdoor scanning for deep neural networks through k-arm optimization. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[56] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24807" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[57] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[58] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[59] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In 30th USENIX Security Symposium (USENIX Security 21), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[60] Guanhong Tao, Yingqi Liu, Guangyu Shen, Qiuling Xu, Shengwei An, Zhuo Zhang, and Xiangyu Zhang. Model orthogonalization: Class distance hardening in neural networks for better security. In 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.412 + ], + "angle": 0, + "content": "[61] Guanhong Tao, Guangyu Shen, Yingqi Liu, Shengwei An, Qiuling Xu, Shiqing Ma, Pan Li, and Xiangyu Zhang. Better trigger inversion optimization in backdoor scanning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13368-13378, 2022. 1, 2, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.472, + 0.467 + ], + "angle": 0, + "content": "[62] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. In Advances in Neural Information Processing Systems, pages 8000–8010, 2018. 2, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.523 + ], + "angle": 0, + "content": "[63] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.551 + ], + "angle": 0, + "content": "[64] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Clean-label backdoor attacks. 2018. 1, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.593 + ], + "angle": 0, + "content": "[65] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.472, + 0.663 + ], + "angle": 0, + "content": "[66] Akshaj Kumar Veldanda, Kang Liu, Benjamin Tan, Prashanth Krishnamurthy, Farshad Khorrami, Ramesh Karri, Brendan Dolan-Gavitt, and Siddharth Garg. Nnoculation: broad spectrum and targeted treatment of backdoored dnns. arXiv preprint arXiv:2002.08313, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.472, + 0.734 + ], + "angle": 0, + "content": "[67] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019. 1, 2, 3, 7, 8, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.472, + 0.789 + ], + "angle": 0, + "content": "[68] Ren Wang, Gaoyuan Zhang, Sijia Liu, Pin-Yu Chen, Jinjun Xiong, and Meng Wang. Practical detection of trojan neural networks: Data-limited and data-free cases. In European Conference on Computer Vision, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.472, + 0.845 + ], + "angle": 0, + "content": "[69] Zhenting Wang, Hailun Ding, Juan Zhai, and Shiqing Ma. Training with more confidence: Mitigating injected and natural backdoors during training. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[70] Zhenting Wang, Kai Mei, Hailun Ding, Juan Zhai, and Shiqing Ma. Rethinking the reverse-engineering of Trojan triggers. In Advances in Neural Information Processing Systems, 2022. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[71] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15074-15084, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[72] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. Advances in Neural Information Processing Systems, 34:16913-16925, 2021. 1, 7, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[73] Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep k-means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In International Conference on Machine Learning, pages 5363-5372. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[74] Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In USENIX Security Symposium, pages 2237-2254, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.431 + ], + "angle": 0, + "content": "[75] Chong Xiang, Saeed Mahloujifar, and Prateek Mittal. {PatchCleanser}: Certifiably robust defense against adversarial patches for any image classifier. In 31st USENIX Security Symposium (USENIX Security 22), pages 2065-2082, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.893, + 0.487 + ], + "angle": 0, + "content": "[76] Zhen Xiang, David Miller, and George Kesidis. Post-training detection of backdoor attacks for two-class and multi-attack scenarios. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.893, + 0.531 + ], + "angle": 0, + "content": "[77] Zhen Xiang, Zidi Xiong, and Bo Li. Umd: Unsupervised model detection for x2x backdoor attacks. In International Conference on Machine Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.893, + 0.588 + ], + "angle": 0, + "content": "[78] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10523-10531, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.894, + 0.671 + ], + "angle": 0, + "content": "[79] Qiuling Xu, Guanhong Tao, Jean Honorio, Yingqi Liu, Shengwei An, Guangyu Shen, Siyuan Cheng, and Xiangyu Zhang. Medic: Remove model backdoors via importance driven cloning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20485-20494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.893, + 0.716 + ], + "angle": 0, + "content": "[80] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. arXiv preprint arXiv:1910.03137, 2019. 2, 7, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.759 + ], + "angle": 0, + "content": "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.893, + 0.816 + ], + "angle": 0, + "content": "[82] Yi Zeng, Han Qiu, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. arXiv preprint arXiv:2012.07006, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[83] Kaiyuan Zhang, Guanhong Tao, Qiuling Xu, Siyuan Cheng, Shengwei An, Yingqi Liu, Shiwei Feng, Guangyu Shen, PinYu Chen, Shiqing Ma, and Xiangyu Zhang. FLIP: A provable defense framework for backdoor mitigation in federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24808" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 14, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[85] Shihao Zhao, Xingjun Ma, Xiang Zheng, James Bailey, Jingjing Chen, and Yu-Gang Jiang. Clean-label backdoor attacks on video recognition models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14443-14452, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[86] Chen Zhu, W Ronny Huang, Hengduo Li, Gavin Taylor, Christoph Studer, and Tom Goldstein. Transferable clean-label poisoning attacks on deep neural nets. In International Conference on Machine Learning, pages 7614-7623. PMLR, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24809" + } + ] +] \ No newline at end of file diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_origin.pdf b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..830c1702cf0aa4da512505c569a0ee900442cf53 --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/6b71bad9-734e-4c81-9b75-13613bad7df8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af4cbfc1cd3f58dfd3ef01c10eaf90f29144a2e0e7f4b1bb95d850995ac26125 +size 2230040 diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/full.md b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a41285351f868f58a8cd280ff767a0a83f0cd19 --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/full.md @@ -0,0 +1,318 @@ +# LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning + +Siyuan Cheng, Guanhong Tao, Yingqi Liu†, Guangyu Shen, Shengwei An, Shiwei Feng, Xiangzhe Xu, Kaiyuan Zhang, Shiqing Ma‡, Xiangyu Zhang Purdue University, †Microsoft, ‡University of Massachusetts at Amherst + +{cheng535, taog, shen447, an93, feng292, xu1415, zhan4057, xyzhang}@cs.purdue.edu †yingqiliu@microsoft.com, ‡shiqingma@umass.edu + +# Abstract + +Backdoor attack poses a significant security threat to Deep Learning applications. Existing attacks are often not evasive to established backdoor detection techniques. This susceptibility primarily stems from the fact that these attacks typically leverage a universal trigger pattern or transformation function, such that the trigger can cause misclassification for any input. In response to this, recent papers have introduced attacks using sample-specific invisible triggers crafted through special transformation functions. While these approaches manage to evade detection to some extent, they reveal vulnerability to existing backdoor mitigation techniques. To address and enhance both evasiveness and resilience, we introduce a novel backdoor attack LOTUS. Specifically, it leverages a secret function to separate samples in the victim class into a set of partitions and applies unique triggers to different partitions. Furthermore, LOTUS incorporates an effective trigger focusing mechanism, ensuring only the trigger corresponding to the partition can induce the backdoor behavior. Extensive experimental results show that LOTUS can achieve high attack success rate across 4 datasets and 7 model structures, and effectively evading 13 backdoor detection and mitigation techniques. The code is available at https://github.com/Megum1/LOTUS. + +# 1. Introduction + +Backdoor attack is a prominent security threat to Deep Learning applications, evidenced by the large body of existing attacks [5, 19, 37, 51, 64] and defense techniques [20, 32, 35, 67, 72]. It injects malicious behaviors to a model such that the model operates normally on clean samples but mis-classifies inputs that are stamped with a specific trigger. A typical way of injecting such malicious behaviors is through data poisoning [1, 19, 39]. This approach introduces a small set of trigger-stamped images paired with the target label + +into the training data. Attackers may also manipulate the training procedure [10, 45, 46], and tamper with the model's internal mechanisms [37, 41]. + +The majority of existing attacks rely on a uniform pattern [5, 19, 39, 64] or a transformation function [6, 51] as the trigger. The uniform trigger tends to be effective on any input, which can be detected by existing techniques. For instance, trigger inversion methods [20, 38, 67, 68] aim to reverse engineer a small trigger that can induce the target prediction on a set of inputs. According to the results reported in the literature [20, 61, 67], for a number of attacks, it is feasible to invert a pattern that closely resembles the ground-truth trigger and has a substantially high attack success rate (ASR), hence detecting backdoored models. + +Recent studies introduce sample-specific invisible attacks [10, 33, 45, 46] that encourage the model to emphasize the correlation between the trigger and the input sample. Although these attacks effectively evade certain detection methods [20, 67], they are not resilient to backdoor mitigation techniques [32, 35, 72]. For instance, a straightforward approach such as fine-tuning the backdoored model using only $5\%$ of the training data can significantly reduce ASR. This is due to the fact that imperceptible trigger patterns are not persistent during the retraining process. Moreover, the sample-specific characteristic of these attacks make them less robust to backdoor mitigation methods. + +In this paper, we introduce an innovative attack that not only evades backdoor detection approaches but also exhibits resilience against backdoor mitigation techniques. It is a label-specific attack, aiming to misclassify the samples of a victim class to a target class. For the victim-class samples, we divide them into sub-partitions and use a unique trigger for each partition. With such an attack design, existing defense such as trigger inversion is unlikely to find a uniform trigger. This is because the available set of samples used by trigger inversion is likely from different partitions, which makes the detection fail. In addition, we develop a novel trigger focusing technique to ensure that a partition can only be attacked by its designated trigger, not by any + +other trigger or trigger combinations. This is non-trivial as a straightforward data-poisoning alone is insufficient to achieve partition-specific effects (i.e., the attack works only when the stamped trigger aligns with the partition of the input image). More details can be found in Section 4. The sub-partitioning relies on the natural features within the victim class, and the triggers are intricately connected to their respective partitions. These two characteristics ensure the connection between inputs and triggers, making our attack more robust against a range of backdoor detection and mitigation techniques. + +Our contributions are summarized as follows: (1) We propose a new backdoor attack prototype LOTUS ("Evasive and ResiLient BackdOor A Tacks through Sub-partitioning") that achieves both evasiveness and resilience. (2) We address a key challenge of the proposed attack, to precisely limit the scope of a trigger to its partition. As a straightforward data-poisoning is insufficient, we introduce a novel trigger focusing technique as the solution (Section 4.2). (3) We conduct an extensive evaluation of LOTUS on 4 datasets and 7 model structures. Our results show that LOTUS achieves a high ASR under a variety of settings. Our trigger focusing method effectively reduces the ASR on undesired victim classes and partitions. Furthermore, our experiments demonstrate that LOTUS is evasive and resilient against 13 state-of-the-art backdoor defense techniques, substantially outperforming existing backdoor attacks. + +Threat Model. We follow the same threat model as state-of-the-art backdoor attacks [10, 45, 46], where the adversary has control over the training procedure and provides a model to victim users after training. The adversary's goal is to achieve high attack effectiveness while also ensuring the attack's evasiveness and resilience against defense. LOTUS primarily focuses on label-specific attack. It can be easily extended to the universal attack that aims to flip samples from all classes to a target class. The defender possesses white-box access to the model and a small set of clean samples for each class. She aims to determine if a model contains backdoor or mitigate the backdoor effects based on the validation samples. In our attack, the sub-partitioning function and the corresponding triggers are the secret of the attacker. + +# 2. Related Work + +Backdoor Attack. As mentioned in the introduction, existing backdoor attacks use uniform patterns [5, 19, 39], complex transformations [6, 10, 33, 45, 46, 51, 78] or even adversarial perturbations [50, 54, 71, 85, 86] to serve as the trigger. Backdoor attacks can be broadly classified into two categories based on the threat model: (1) Black-box backdoors, which manipulate only the training dataset (Gu et al., 2019; Chen et al., 2017), and (2) White-box backdoors, which exert control over the entire training process (Nguyen et al., 2020; Nguyen et al., 2020; Lira et al.). Our proposed + +attack, LOTUS belongs to the white-box backdoor category, aligning with the existing works. Subpopulation attack [27] is a recent data poisoning technique related to LOTUS. It is an availability attack, and its primary objective is to decrease the test accuracy of a specific subpopulation within the dataset. In contrast, LOTUS is a comprehensive backdoor attack with the intention of injecting a backdoor into the model. Therefore, these two attacks differ significantly. Subpopulation attack does not involve trigger injection or require the implementation of trigger focusing, making it distinct from LOTUS in terms of its objectives and mechanisms. + +Backdoor Defense. Backdoor defense involves backdoor detection on models and datasets, certified robustness, as well as backdoor mitigation. Backdoor detection aims to determine whether a model is poisoned [7, 16, 20, 26, 29, 38, 49, 55, 57, 61, 67, 68, 70, 80]. Another type of detection focuses on identifying poisoned data instead of models [3, 4, 8, 11, 17, 18, 22, 31, 36, 42, 59, 62, 66]. Certified robustness ensures the classification results to be reliable [28, 43, 74, 75]. Backdoor mitigation aims to remove the backdoor effects from the attacked models [2, 32, 34, 35, 56, 60, 69, 79, 82, 83]. + +# 3. Attack Definition + +We formally define our attack in this section. For a typical classification task, given $(\pmb{x},y)\sim \mathcal{D}$ where the sample $\pmb {x}\in \mathbb{R}^d$ and label $y\in \{1,2,\dots N\}$ , the goal is to train a classifier $M_{\theta}:\mathbb{R}^{d}\to \{1,2,\dots ,N\}$ , such that parameters $\theta = \arg \max_{\theta}P_{(\pmb {x},y)\sim \mathcal{D}}[M_{\theta}(\pmb {x}) = y]$ . Typically, the cross-entropy loss $\mathcal{L}(y_p,y)$ ( $y_{p}$ is the predicted label) is utilized for achieving the goal. In this case, the optimization problem can be expressed as $\theta = \arg \min_{\theta}\mathbb{E}_{(\pmb {x},y)\sim \mathcal{D}}[\mathcal{L}(M_{\theta}(\pmb {x}),y)]$ . + +Backdoor attack aims to derive a classifier $\overline{\mathcal{M}}_{\bar{\theta}}:\mathbb{R}^d\to$ $\{1,2,\dots ,N\}$ such that compromised parameters $\bar{\theta} =$ arg max $\overline{\theta} P_{(\pmb {x},y)\sim \mathcal{D}} [\overline{\mathcal{M}}_{\bar{\theta}}(\pmb {x}) = y\& \overline{\mathcal{M}}_{\bar{\theta}}(\mathbb{T}\oplus \pmb{x}_V) = y_T],$ in which $\mathbb{T}$ is the trigger and $\mathbb{T}\oplus \pmb{x}_V$ injects the trigger to a victim input sample $x_{V}$ whose label is $y_{V}$ .Symbol $y_{T}$ denotes the attack target label. Backdoor attacks can be mainly classified to universal attack that aims to flip a sample $\pmb{x}$ of any class with $\mathbb{T}$ to the target label, and label-specific attack that aims to flip any sample of a specific victim class to the target label. Based on trigger patterns, they can be classified to input-independent backdoor or static backdoor that has a fixed trigger pattern for all victim samples, and dynamic trigger that has changing patterns for different inputs. Our attack is a label-specific dynamic backdoor attack. Extending to other scenarios is relatively straightforward, e.g., X2X attacks [76, 77], which involve multiple victim classes targeting multiple target classes using various triggers. + +Assume there exists a partitioning algorithm $\mathcal{C}_n: \mathbb{R}^d \to \{p_1, p_2, \dots, p_n\}$ that separates input samples to $n$ partitions. In our attack, victim samples (samples from the victim class) + +are partitioned to $n$ groups using $\mathcal{C}_n$ and each partition $p_i$ is assigned a unique trigger $\mathbb{T}_i$ , such that only $\mathbb{T}_i \oplus \boldsymbol{x}_V^{p_i}$ can trigger the backdoor, where $i \in \{1, 2, \dots, n\}$ and $\boldsymbol{x}_V^{p_i}$ denotes the victim samples in the $i$ -th partition. A straightforward design would follow the classic data poisoning, which aims to optimize the model weights according to the following loss: + +![](images/06dbad3ef839dbb2c8815594f9f71ea87f92b01b321f3d8cf79d18b23a0f1fcc.jpg) + +The first loss term Benign Utility Loss aims to ensure the high benign accuracy of the model. The second term, Attack Target Loss, means that a trigger $\mathbb{T}_i$ can cause the $i$ -th partition samples of the victim class $\pmb{x}_{V}^{p_i}$ to misclassify, which is our attack goal. However, simple data poisoning cannot effectively bound the attack scope. As a result, a trigger for a particular partition can easily induce misclassifications for other partitions. That is, $\mathbb{T}_j \oplus \pmb{x}_{V}^{p_i}$ , where $i \neq j$ , is misclassified to $y_{T}$ . Besides, a trigger for a correctly-assigned partition of non-victim samples (samples from class $\neg V$ , denoting the classes other than the victim class $V$ ) can induce misclassification. That is $\mathbb{T}_i \oplus \pmb{x}_{\neg V}^{p_i}$ is misclassified to $y_{T}$ . Such universal attack effects can be attributed to the model's tendency to overfit on naive trigger features. For instance, when it encounters any trigger, it immediately predicts the target class without verifying if the background image aligns with the trigger according to the partitioning criteria. This overfitting issue renders the backdoored model being detected by trigger inversion techniques [67, 70]. Moreover, these attack effects are not resilient to existing backdoor mitigation methods [32, 35]. + +Our objective is to establish a clear one-to-one correspondence between $\mathbb{T}_i$ and $\pmb{x}_{V}^{p_i}$ . That is, only $\mathbb{T}_i\oplus \pmb{x}_V^{p_i}$ can cause misclassification. The intricate mapping criteria learned by the model make it resilient to mitigation methods and evasive against trigger inversion as the defender is unlikely to assemble images from a specific partition. We hence aim to derive the following loss function. + +$$ +\begin{array}{l} \mathbb {E} _ {(\pmb {x}, y) \sim \mathcal {D}} [ \mathcal {L} (\overline {{\mathcal {M}}} _ {\overline {{\theta}}} (\pmb {x}), y) ] + \sum_ {i = 1} ^ {n} (\mathbb {E} _ {(\pmb {x} _ {V} ^ {p _ {i}}, y _ {V}) \sim \mathcal {D}} [ \mathcal {L} (\overline {{\mathcal {M}}} _ {\overline {{\theta}}} (\mathbb {T} _ {i} \oplus \pmb {x} _ {V} ^ {p _ {i}}), y _ {T}) ] \\ + \mathbb {E} _ {\left(\boldsymbol {x} _ {\neg V} ^ {p _ {i}}, y _ {\neg V}\right) \sim \mathcal {D}} [ \mathcal {L} (\overline {{\mathcal {M}}} _ {\bar {\theta}} \left(\mathbb {T} _ {i} \oplus \boldsymbol {x} _ {\neg V} ^ {p _ {i}}\right), y _ {\neg V}) ] \longleftrightarrow \text {L a b e l - s p e c i f i c L o s s} \\ + \underbrace {\mathbb {E} _ {\left(\boldsymbol {x} _ {V} ^ {p _ {i}} , y _ {V}\right) \sim \mathcal {D}} \left[ \right. \sum_ {\substack {\mathcal {T} \in \mathcal {P} \left( \right.\{\mathbb {T} _ {1} , \dots , \mathbb {T} _ {n} \left. \right\}}} \mathcal {L} \left(\overline {\mathcal {M}} _ {\bar {\theta}} \left(\mathcal {T} \oplus \boldsymbol {x} _ {p _ {i}}\right), y _ {V}\right)\left. \right]\left. \right]} _ {- \{\{\}, \{\mathbb {T} _ {i} \} \}} \xleftarrow {} \text {D y n a m i c L o s s} \tag{2} \\ \end{array} +$$ + +Note that compared to Equation 1, we introduce two additional terms, i.e., Label-specific Loss and Dynamic Loss in Equation 2. Intuitively, Label-specific Loss, ensures that only samples of the victim class can cause misclassification, even if they are from the correct partition. Here $\neg V$ denotes the classes other than the victim class. The last term Dynamic + +Loss controls that for a particular partition, only the corresponding trigger can cause misclassification, and any other trigger, or combination of/with other triggers shall be correctly predicted as the victim class. In particular, $\mathcal{T}$ is a subset of all possible triggers/combinations $\mathcal{P}(\{\mathbb{T}_1,\dots ,\mathbb{T}_n\})$ excluding empty $\{\}$ and $\{\mathbb{T}_i\}$ . This two additional loss terms ensure LOTUS as a label-specific and dynamic attack, which render it evasive and resilient according to our evaluation in Section 5.3 and 5.4. + +# 4. Detailed Attack Design + +The overview of LOTUS is shown in Figure 1. Victim class input samples are first separated to partitions. We then apply unique triggers to samples from the corresponding partitions, whose labels are set to the target class. Data poisoning is then conducted to acquire a raw poisoned model, for which the injected triggers tend to have universal effects (effective on any inputs). To address this problem, LOTUS further introduces a trigger focusing step that strictly limits the attack scope of each trigger. It finally produces a trojaned model with triggers that are evasive and resilient. + +In the following, we elaborate two major components of LOTUS, namely, victim-class sample partitioning and trigger focusing. + +# 4.1. Victim-class Sample Partitioning + +LOTUS separates a set of victim-class samples into multiple partitions, and injects different triggers to different partitions. We propose two ways to partition input samples. The first is explicit partitioning that leverages a subset of explicit attributes of the victim class (e.g., hair color and w./ or w./o. glasses for face recognition). Assume $k$ attributes are used and each attribute has $t$ possible values. This allows to generate $t^k$ partitions. The first two columns in Figure 2 show a partitioning based on the taxonomy attribute of the bird class. Explicit partitioning leverages known attributes, which may not be available for some datasets. We hence introduce an advanced partitioning method that is applicable to arbitrary datasets in the following. + +The second partitioning scheme is implicit, meaning that human uninterpretable features are used in partitioning. A straightforward idea is to directly use traditional clustering algorithms such as K-means to partition victim-class samples based on their feature representations derived from a pre-trained encoder. However, according to our experiment in Appendix L.1, such a naive method does not work well. The root cause is that K-means is a clustering algorithm on a set of known data points and does not consider generalization to unseen data points. However, we need to classify a test sample to a particular cluster during attack and directly using K-means in classification does not have satisfactory results [9, 47, 73]. + +![](images/cc131953041ebdf5d41c2e0e7723d3092436852b4a6eb1d871da792c4694ff26.jpg) +Figure 1. Overview of LOTUS + +![](images/337a33f750e6eb2e76af9a799d11232c1674cd96d698f78c3f556edcb477a951.jpg) +Figure 2. Explicit (left) and implicit (right) partitioning. + +![](images/dde03794c3b69714c6a5e285dce415dd744d6f4620290b696e1039e66c10da50.jpg) +Figure 3. Implicit partitioning with surrogate model. + +We hence introduce a surrogate model to help sample partitioning. Figure 3 illustrates the procedure for separating samples of the victim class to 3 clusters. The surrogate model has the same structure as the victim model to reduce complexity caused by structural differences. On the bottom left, the features of samples from victim class $n$ are extracted using a pre-trained encoder. We then use a traditional clustering method such as K-means to partition these samples into 3 different sub-classes based on their features. We assign labels $n$ , $n + 1$ , $n + 2$ to samples from the respective sub-classes. They are then combined with samples from the original classes 1 to $n - 1$ (excluding the victim class $n$ ) to form a new dataset consisting of $n + 2$ classes. The surrogate model is trained on this new dataset with $n + 2$ classes. The idea is to use K-means to provide a meaningful prior separation and then use classifier training to achieve generalizability. Furthermore, the decision boundaries by the surrogate model have the classes other than the victim class + +in consideration, whereas those by distances to centroids of K-means clusters only have samples of the victim class in consideration. After the training converges, the surrogate model is utilized to determine the partition of a test sample. That is, the partition index can be derived from the its classification outcome (i.e., the class with largest logits from classes $n$ to $n + 2$ ). The last two columns in Figure 2 show two implicit partitions of the "cat" class. Observe that the partitions are largely uninterpretable, which makes the attack more stealthy compared to using explicit attributes which are public. + +Handle Potential Imbalanced Examples. We control that for any partitioning, the sizes of each partition are roughly the same, which mitigates the potential of causing partitioning bias. This is achieved by removing samples from exceptionally large clusters. In practice, such a removal is rarely needed. + +# 4.2. Trigger Focusing + +After partitioning, LOTUS aims to limit each trigger to its own partition, preventing it from attacking other partitions or classes. To achieve this, we design a trigger focusing technique during training. + +A straightforward idea is to strictly follow the definition in Equation 2 to bound the trigger scope. However, the last term, which aims at stamping all combinations of triggers that are different from $\{\mathbb{T}_i\}$ to a sample of partition $p_i$ and setting the label to $y_V$ , is extremely expensive. The number of combinations is $(2^n - 2)$ , which grows exponentially with the increase of the number of partitions $n$ . Moreover, the inclusion of a substantial number of additional samples will not only slow down the training but also imbalance the dataset, ultimately impacting the overall performance. + +Adversarial Poisoning Is Insufficient. Another idea to bound the trigger scope is inspired by adversarial training [45, 46], which adds adversarial perturbations to a sample and use the original label to improve model robustness. To suppress the undesirable attack effect in our context, we could inject triggers that are not for a partition $p_i$ , i.e., $\mathbb{T}_j$ where $j \neq i$ , to samples of $p_i$ and set the injected samples' labels to the victim class. This approach is referred + +![](images/3de8fbb4d83b92baf7a2dc07941c0d11538840469c1ed598f92c2f009d2679d0.jpg) + +![](images/3b01cdcfbb26026175545fca76f696d90b0b60a1ddec610860ec7db7be7b66b8.jpg) + +![](images/a01d5fb6ad6b31e73a9a41de05811ac8c0b4545c507968df590ee7db44932562.jpg) + +![](images/3cea3593421c93681799a945755df62a9ef1d00d79d660ecc5baaa97cc6fac3a.jpg) + +![](images/b6a8303a766cf65c5fa65185b4f1c8a6836953ec1de38ad916c8421a07355961.jpg) + +![](images/d95ca0c404b8c601d0269a19c1a03722c8fb870c86a2517ea2740cc1aa8888ad.jpg) +(a) Data-poisoning + +![](images/d620f783223ff032a3c2928de46445a6b25bf06b96921c11e907ed551f51e581.jpg) +(b) Adversarial-poisoning + +![](images/c4c6ae66c30c8006ab6391446ad4e279f078a64c97e82668c9236baff3f52fe5.jpg) +Figure 4. Decision boundaries for different poisoning strategies. + +![](images/967060fe0534c11283b8a578fa18db15ccc06c0f4564568ec3cfa56a3b30901a.jpg) +(c) Trigger-focusing + +![](images/cbb289b1d6414c57fa083e3ed2c665bb1b9744a69ca4da459b1581d06bd713fa.jpg) + +to as adversarial poisoning. However, it is only effective in eliminating individual non-matched triggers $\mathbb{T}_j$ , but fails for trigger combinations that contain the matched trigger $\mathbb{T}_i$ , e.g., $[\mathbb{T}_i,\mathbb{T}_j]$ . + +Figure 4 presents a visualization of decision boundaries for various poisoning strategies, namely: (a) Straightforward data-poisoning; (b) Adversarial-poisoning; and (c) Trigger-focusing (which will be discussed in the next paragraph). Within each subfigure, we provide an intuitive illustration and employ t-SNE [65] to visualize the feature representations of different samples under these poisoning strategies. The experiment is conducted on the CIFAR-10 dataset using the ResNet18 model, and we utilize implicit partitioning to create four distinct partitions. In the figure, a hollow closed lock is used to denote clean images $x_{v}^{p_{i}}$ in the victim class of partition $p_{i}$ , while a red opened lock is used to represent clean images of the target class. Triggers are depicted as keys with various colors. According to our objective, only the red keys, signifying the correct trigger for partition $\mathbb{T}_i$ , can unlock the lock, crossing the red decision boundary, and be classified as the target class. Keys of different colors, signifying various triggers or combinations, are unable to unlock the lock and remain within the victim class region. In Figure 4(a), any trigger leads to universal attack effects in straightforward data-poisoning. Observe any key, denoting a trigger, can unlock the lock and cross the boundary without limitations. The t-SNE visualization on real data on the right aligns with the illustration on the left. In contrast, adversarial-poisoning, as depicted in (b), mitigates the impact of samples with unmatched individual triggers, as represented by the green key. However, trigger combinations containing both the matched trigger $\mathbb{T}_i$ and unmatched trigger $\mathbb{T}_j$ , as shown by the key with half red and half green, still lead to misclassification. Similarly, in the t-SNE visualization, the yellow triangles, which represent this type of trigger combination, are substantially close to the red triangles, denoting the strictly matched triggers. This indicates + +the insufficiency of adversarial-poisoning. + +Efficient and Effective Trigger Focusing. Inspired by the observation in Figure 4, we propose a novel trigger focusing method that can effectively bound trigger scopes and is in the mean time cost-effective. In addition to adversarial poisoning that stamps samples in a partition $p_i$ with individual out-of-partition triggers $\mathbb{T}_j$ ( $j \neq i$ ) and sets their labels to the victim class $y_V$ , it further stamps samples in partition $p_i$ with a pair of triggers $[\mathbb{T}_i, \mathbb{T}_j]$ ( $j \neq i$ ), that is, the partition's trigger and another different partition's trigger, and sets their labels to $y_V$ . + +$$ +\begin{array}{l} \sum_ {i = 1} ^ {n} \mathbb {E} _ {\left(\boldsymbol {x} _ {V} ^ {p _ {i}}, y _ {V}\right) \sim \mathcal {D}} [ \sum_ {j = 1, j \neq i} ^ {n} (\mathcal {L} (\overline {{\mathcal {M}}} _ {\bar {\theta}} (\mathbb {T} _ {j} \oplus \boldsymbol {x} _ {V} ^ {p _ {i}}), y _ {V}) \tag {3} \\ \left. + \mathcal {L} \left(\overline {{\mathcal {M}}} _ {\bar {\theta}} \left([ \mathbb {T} _ {i}, \mathbb {T} _ {j} ] \oplus \boldsymbol {x} _ {V} ^ {p _ {i}}\right), y _ {V})\right) \right] \\ \end{array} +$$ + +Our approach, with the new dynamic loss term expressed in Equation 3, requires only $(2n - 2)$ trigger combinations, which increases linearly with the growth of partitions $n$ . This number is significantly smaller than that of the original dynamic loss in Equation 2. + +Intuitively, the different labels of samples $\mathbb{T}_i\oplus \pmb{x}_V^{p_i}$ and $[\mathbb{T}_i,\mathbb{T}_j]\oplus \pmb{x}_V^{p_i}$ enable the model to learn new behaviors. As such, further stamping any other partition triggers to $[\mathbb{T}_i,\mathbb{T}_j]\oplus \pmb{x}_V^{p_i}$ yields the same classification result, which is the victim class. Please refer to Appendix D for a detailed reasoning and theoretical analysis. + +In Figure 4(c), it is noteworthy that trigger combinations are effectively excluded from the target class and only the trigger that matches the victim partition can cause the misclassification, well aligning with our attack objective. + +# 5. Evaluation + +In this section, we evaluate on 4 benchmark datasets and 7 model structures to demonstrate the attack effectiveness of + +LOTUS (Section 5.2). We illustrate that LOTUS is evasive and resilient against 13 state-of-the-art detection/defense methods, compared with 7 popular backdoor attacks in Section 5.3 and 5.4. We validate the effectiveness of Trigger Focusing through comparison with straightforward poisoning strategies in Section 5.5. We also extend LOTUS to universal attacks in Appendix G. Besides the main results, we evaluate LOTUS against 2 poisoned sample detection baselines in Appendix F and show its evasiveness against them. Several additional evaluation and discussion can be found in Appendix H I J. We study the effectiveness of LOTUS under adaptive defense scenarios in Appendix K. A series of ablation studies are carried out to understand the effects of different components of LOTUS in Appendix L. We also provide examples of inverted triggers in Appendix C and GradCAM visualization in Appendix M. + +# 5.1. Experiment Setup + +We evaluate LOTUS on 4 widely-used benchmarks, CIFAR-10 [30], CIFAR-100 [30], CelebA [40], and restricted ImageNet (RImageNet) [12, 52, 63]. Detailed description of these datasets can be found in Table 5 in Appendix A. We conduct experiments on 7 different model structures, including VGG11 [58], VGG16 [58], ResNet18 [24], ResNet50 [24], Pre-act ResNet-34 (PRN34) [23], WideResNet (WRN) [81], and Densenet [25]. + +We leverage several sub-partitioning methods to partition samples from the victim class. We utilize secondary labeling, e.g., various cat species, to create clear and explicit partitions. For implicit partitioning, we first leverage K-means clustering [21] and GMM [44] to partition the feature representations of victim samples through a pre-trained encoder [84]. Then we train a surrogate model to learn the partitioning principle, which serves as the implicit sub-partitioner (Section 4.1). Details of the sub-partitioning and encoder can be found in Appendix B. + +# 5.2. Attack Effectiveness + +We evaluate the performance of LOTUS on various datasets, model structures and partitioning methods. Table 1 presents the results. For all the experiments, we use the first class of each dataset as the victim and the last class as the target. We generate 4 partitions for the victim class throughout all datasets and model structures. Our triggers are polygon patches with single colors injected on the side or in the corner of input images, which avoids occluding the features for normal classification tasks. Example images with triggers can be found in Figure 10 in Appendix. The top two blocks in Table 1 (separated by the double lines) show the results for implicit partitioning, and the bottom for explicit partitioning. For K-means clustering, ASRs are at least $89.00\%$ , with the highest ASR of $94.30\%$ for ResNet18 on CIFAR-10, while the degradation of benign accuracy is within $1.07\%$ . This + +Table 1. Evaluation of attack effectiveness. The first three columns denote different partitioning algorithms (PA), datasets, and model structures. The following columns present the original accuracy of clean models (Acc.), benign accuracy of the backdoored models (BA), the attack success rate when stamping a trigger on the proper partition (ASR), and the average ASR when stamping other triggers and trigger combinations, with the standard deviation) (ASR-other). + +
PADatasetModelAcc.BAASRASR-other
K-meansCIFAR-10VGG1192.16%92.04%93.80%4.77% ± 19.27%
ResNet1895.22%94.71%94.30%4.39% ± 17.08%
CIFAR-100Densenet75.14%75.15%92.00%4.36% ± 14.24%
PRN3474.70%74.52%89.00%5.43% ± 13.50%
CelebAWRN80.47%79.40%92.33%6.87% ± 17.49%
RImageNetResNet5097.77%97.19%93.87%2.16% ± 19.34%
GMMCIFAR-10ResNet1895.22%94.59%90.70%4.80% ± 21.38%
CIFAR-100PRN3474.70%74.02%91.00%2.21% ± 12.57%
CelebAWRN80.47%79.66%92.53%5.39% ± 16.77%
RImageNetVGG1696.51%95.93%93.52%3.11% ± 14.39%
Sec.RImageNetVGG1696.51%96.36%96.50%1.79% ± 13.24%
ResNet5097.77%97.08%92.50%2.14% ± 16.53%
+ +Table 2. Evaluation of label specificity. ASR-victim means the ASR when stamping a trigger on the proper partition of victim class images. ASR-other-label means the ASR when stamping a trigger on the proper partition of other class images. + +
DatasetNetworkASR-victimASR-other-label
CIFAR10ResNet1893.80%14.37%
CIFAR100Densenet92.00%11.23%
CelebAWRN92.33%19.67%
RImageNetVGG1693.52%12.22%
+ +indicates LOTUS is a highly effective attack, which injects successful malicious behaviors to the model while maintains its benign utility. The last column shows the ASR when trigger/trigger-combinations other than a partition's trigger are stamped on the partition (ASR-other). Observe that the average ASR-other is less than $6.87\%$ , delineating the effectiveness of trigger focusing (a trigger is only effective for the corresponding partition). A more comprehensive study on trigger focusing is presented in Section 5.5. We have similar observations for using GMM in implicit partitioning. For the explicit secondary labeling, LOTUS can achieve an ASR over $92.50\%$ and a small ASR-other. The better performance of LOTUS using secondary labeling can be attributed to the fact that the victim class in RImageNet is merged from a set of similar classes in ImageNet. Those classes are naturally separable, which can be easily differentiated by the model when triggers are injected on different partitions. + +Note that the ASR of LOTUS is slightly lower than the existing attacks (as shown in the "No Defense" column in Table 3). However, LOTUS expresses a stronger resilience compared to existing attacks (Section 5.4). This is a trade-off between attack effectiveness and resilience. More discussion can be found in Appendix J. + +Besides, we also evaluate the label specificity of LOTUS on several models. Results are presented in Table 2. Observe that even if the trigger is stamped on the proper partition + +of the input image, the ASR-other-label is low ( $< 20\%$ ) because the input image is not of the victim class. The result shows that LOTUS exhibits a high level of label specificity. Furthermore, LOTUS offers an easy extension into universal attack scenarios through the integration of explicit partitioning techniques. Detailed examples can be found in Section G. + +# 5.3. Evasiveness against Backdoor Detection + +In this section, we study the evasiveness of LOTUS against 4 well-known trigger-inversion based backdoor detection methods, including Neural Cleansse (NC) [67], Pixel [61], ABS [38], and FeatRE [70]. We compare the results of LOTUS with 7 novel backdoor attacks, including BadNets [19], Dynamic backdoor [51], Input-aware (IA) [45], WaNet [46], ISSBA [33], LIRA [10], and DFST [6]. For fair comparison, we launch all backdoor attacks on ResNet18 models trained on CIFAR-10. As LOTUS is a label-specific attack, we implement all other attacks in label-specific setting, where the poisoned samples are composed of images from victim class 0 stamped with the trigger and labeled as the target class 9. Besides, all detection methods are required to invert triggers based on 100 clean validation images from the victim class, targeting to labels other than it. We follow all the other settings and techniques of the original papers to implement the attack and detection methods. + +Figure 5 illustrates the detection results, where the x-axis denotes different attacks and the y-axis denotes the decision scores of each baseline. The thresholds are highlighted in red dashed lines. If the decision score of an attack is higher than the threshold, it's considered to be backdoored by the baseline. Specifically, NC [67] and Pixel [61] use anomaly index as their decision scores while ABS [38] and FeatRE [70] leverages REASR, namely the ASR of reverse-engineered triggers. Observe that NC, Pixel, ABS are effective against several attacks, including BadNets, Dynamic, ISSBA, LIRA and DFST, while leaving other advanced attacks, i.e., WaNet, IA and LOTUS. FeatRE, on the other hand, observes internal linear separability properties of existing backdoors and improves the trigger inversion process, which is able to detect the advanced backdoors operating in the feature space. Figure 5(d) shows that it can detect both IA and WaNet, but still fails to detect LOTUS. This illustrates that LOTUS is more evasive than all these baseline attacks. The underlying reason is that LOTUS leverages partitioning secrets and trigger focusing, which breaks the linear separability assumption. Without knowledge of partitioning, it's unlikely to invert a trigger with high ASR, and hence unlikely to detect the backdoor. Examples of inverted triggers can be found in Appendix C. + +We also test LOTUS in the adaptive defense scenario, where the defender can create partitions before detection. The results in Appendix K demonstrate that LOTUS is re + +silent against adaptive defense strategies, as guessing the correct partitioning is challenging. + +Besides trigger inversion methods, we also evaluate LOTUS using meta-classifiers, e.g., MNTD [80] and ULP [29], which train model-level classifiers for detection. Results in Appendix E show that LOTUS is evasive against them. + +# 5.4. Resilience against Backdoor Mitigation + +In this section, we study the resilience of LOTUS against 4 state-of-the-art backdoor mitigation methods, including standard Fine-tuning, Fine-pruning [35], NAD [32], and ANP [72]. We compare the results of LOTUS with 7 novel backdoor attacks. For fair comparison, all the models are trained using VGG11 on CIFAR-10 dataset. For each mitigation method, we assume the access to $5\%$ of the training data. Besides, some standard input argumentation techniques are used, e.g., random cropping and horizontal flipping. We follow the original setting to conduct these baseline methods. + +Table 3 provides the result. Observe that for all the baselines, benign accuracy change is slight, meaning that the mitigation preserves the model utility on benign tasks. However, ASR degradation is considerable for all backdoored models. Note that LOTUS can still remain part of the attack effectiveness with $34.90\% - 46.90\%$ , outperforming all other attacks. The result indicates that LOTUS is more resilient against baseline mitigation methods compared to the existing attacks. This can be attributed to the design that LOTUS learns the correlation between the partitions and triggers which is hard to unlearn. Other attacks only learn partial trigger patterns that tend to be mitigate. + +# 5.5. Evaluation on Different Poisoning Strategies + +We evaluate different poisoning strategies including simple data poisoning, adversarial poisoning, and LOTUS's trigger focusing. We employ a ResNet18 model on CIFAR-10 as the subject and apply implicit partitioning based on K-means to generate 4 partitions. The number of possible non-empty trigger combinations is $2^{4} - 1 = 15$ . In the following, we use a four-bit binary to represent each combination. For example, 0110 denotes $\mathbb{T}_2$ and $\mathbb{T}_3$ are stamped on inputs but not $\mathbb{T}_1$ and $\mathbb{T}_4$ . Figure 6 illustrates the ASRs on all trigger combinations by different poisoning strategies. Sub-figures from left to right present the results for simple poisoning, adversarial poisoning, and trigger focusing, respectively. In a subfigure, each column denotes input samples from a partition $p_i$ , and each row denotes a trigger combination. The value in each cell shows the ASR when a trigger combination (row) is stamped on the samples from a partition (column). Brighter the color, higher the ASR. The left sub-figure shows the ASR for simple data poisoning. Observe that all the ASRs are greater than $92.0\%$ (with an average of $97.94\%$ ), showing the sub-partitioning is not learned by the model. The middle + +![](images/edf326ed9bd3982acf628d6df140fd05097425e8d5ef725c845450673f6b9309.jpg) +Figure 5. Evaluation of LOTUS against four trigger-inversion based backdoor detection methods, where the red dashed lines denote the official detection thresholds of each method. + +![](images/715a70e921f96b1ad604b155f9ab52f220f24b2cd6a9723ebfc0eb90f0c6da07.jpg) + +![](images/8ed3dfdbf3b0b823e360c2ce1c00f8038c432263aace0b105872ce3d5e84bb8c.jpg) + +![](images/06dc243d5eb6d96682c57e3931f34f7b59f16d95911f838fc984e10b96ed09af.jpg) + +Table 3. Evaluation of resilience against backdoor mitigation methods. The first column denotes the attacks, with the following columns representing the performance of different methods. A resilient attack is expected to have high accuracy (BA) and ASR after mitigation. The best results are in bold. + +
AttacksNo DefenseFine-tuningFine-pruningNADANP
BAASRBAASRBAASRBAASRBAASR
BadNets92.02%100.00%89.31%1.74%91.70%0.53%87.81%0.80%89.15%0.32%
Dynamic91.81%100.00%88.87%2.91%91.39%22.03%89.11%2.90%88.25%12.81%
IA91.70%99.65%87.74%2.78%91.07%0.17%87.14%2.29%88.73%1.98%
WaNet91.22%98.57%89.56%1.37%90.22%1.07%89.74%1.40%89.07%0.54%
ISSBA91.67%99.96%87.73%2.72%91.12%14.27%87.97%2.83%85.64%10.01%
LIRA91.70%100.00%89.96%2.19%91.29%12.14%90.23%2.32%89.70%37.91%
DFST91.81%99.97%88.49%22.86%91.47%21.61%88.52%24.66%87.13%36.17%
LOTUS91.54%93.80%88.10%46.90%91.14%44.90%87.61%42.30%88.14%34.90%
+ +![](images/b3d6a9df366e37a4a3add35eef45dfd813c4caeca10e176b6a6228256f541999.jpg) +Figure 6. ASR on all trigger combinations by different poisoning strategies + +sub-figure is the results for adversarial poisoning. We can see around half of cells have small values, especially for single trigger combinations (the top four rows). For more complex trigger combinations, the ASRs are still high with the highest of $100.0\%$ (trigger combination 0111 on partition $p_2$ ), indicating the insufficiency of adversarial poisoning. The right sub-figure is for our trigger focusing. Observe that except for stamping a trigger on the proper partition, the other cases all have a low ASR with an average of $3.04\%$ . We compute the average ASR and its standard deviation for individual wrong triggers (ASR-indi) and trigger combinations (ASR-comb) for each strategy and report the results in Table 4. Observe that all the ASRs are almost $100\%$ for simple poisoning. Adversarial poisoning reduces the ASR-indi to a low level while leaving ASR-comb high ( $73.88\%$ on average). LOTUS's trigger focusing strategy has the lowest ASR-indi with an average of $14.15\%$ and ASR-comb $0.02\%$ . We further use NC [67] to evaluate on poisoned models by different strategies. The last column shows the anomaly index for different poisoned models. Observe that models poisoned by simple data poisoning and adversarial poisoning can be easily detected by NC (with anomaly index $>2$ ). Poisoned models by trigger focusing, on the other hand, are able to evade NC's detection, delineating the effectiveness + +Table 4. Evaluation on different poisoning strategies + +
StrategyBAASRASR-indiASR-combNC index
Simple94.79%98.80%97.86% ± 1.84%97.88% ± 1.81%5.338
Adv.94.47%94.20%18.95% ± 10.22%73.88% ± 31.87%2.161
Focus94.71%91.40%14.15% ± 7.46%0.02% ± 0.09%1.156
+ +of trigger focusing strategy to achieve evasiveness. + +# 6. Conclusion + +We propose a novel backdoor attack that leverages subpartitioning to restrict the attack scope. A special training method is designed to limit triggers to only their corresponding partitions. Our evaluation shows that the attack is highly effective, achieving high attack success rates. Besides, it is evasive and resilient against state-of-the-art defenses. + +# Acknowledgements + +We thank the anonymous reviewers for their constructive comments. We are grateful to the Center for AI Safety for providing computational resources. This research was supported, in part by IARPA TrojAI W911NF-19-S0012, NSF 1901242 and 1910300, ONR N000141712045, N000141410468 and N000141712947. Any opinions, findings, and conclusions in this paper are those of the authors only and do not necessarily reflect the views of our sponsors. + +# References + +[1] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. arXiv preprint arXiv:2005.03823, 2020. 1 +[2] Eitan Borgnia, Valeriia Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong data augmentation sanitizes poisoning and backdoor attacks without an accuracy tradeoff. arXiv preprint arXiv:2011.09527, 2020. 2 +[3] Alvin Chan and Yew-Soon Ong. Poison as a cure: Detecting & neutralizing variable-sized backdoor attacks in deep neural networks. arXiv preprint arXiv:1911.08040, 2019. 2 +[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018. 2 +[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 1, 2 +[6] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1148-1156, 2021. 1, 2, 7 +[7] Siyuan Cheng, Guanhong Tao, Yingqi Liu, Shengwei An, Xiangzhe Xu, Shiwei Feng, Guangyu Shen, Kaiyuan Zhang, Qiuling Xu, Shiqing Ma, et al. Beagle: Forensics of deep learning backdoor attack for better defense. arXiv preprint arXiv:2301.06241, 2023. 2 +[8] Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinel: Detecting localized universal attack against deep learning systems. Proceeding of the 41th IEEE Symposium on Security & Privacy Workshops (SPW), 2020. 2 +[9] Ryan Cohn and Elizabeth Holm. Unsupervised machine learning via transfer learning and k-means clustering to classify materials image data. Integrating Materials and Manufacturing Innovation, 10(2):231-244, 2021. 3 +[10] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11966-11976, 2021. 1, 2, 7 +[11] Min Du, Ruoxi Jia, and Dawn Song. Robust anomaly detection and backdoor attack detection via differential privacy. In International Conference on Learning Representations, 2019. 2 +[12] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Learning perceptually-aligned representations via adversarial robustness. arXiv preprint arXiv:1906.00945, 2(3):5, 2019. 6, 13 +[13] Wang et al. An invisible black-box backdoor attack through frequency domain. In ECCV, 2022. 17 +[14] Xia et al. Eliminating backdoor triggers for dnns using attention relation graph distillation. In *IJCAI*, 2022. 17 +[15] Zeng et al. Adversarial unlearning of backdoors via implicit hypergradient. In ICLR, 2022. 17 + +[16] Shiwei Feng, Guanhong Tao, Siyuan Cheng, Guangyu Shen, Xiangzhe Xu, Yingqi Liu, Kaiyuan Zhang, Shiqing Ma, and Xiangyu Zhang. Detecting backdoors in pre-trained encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16352-16362, 2023. 2 +[17] Hao Fu, Akshaj Kumar Veldanda, Prashanth Krishnamurthy, Siddharth Garg, and Farshad Khorrami. Detecting backdoors in neural networks using novel feature-based anomaly detection. arXiv preprint arXiv:2011.02526, 2020. 2 +[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019. 2, 16 +[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 1, 2, 7, 21 +[20] Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting and eliminating trojan backdoors in deep neural networks. In 20th IEEE International Conference on Data Mining, 2020. 1, 2 +[21] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 6, 13 +[22] Jonathan Hayase, Weihao Kong, Raghav Somani, and Swoong Oh. Spectre: defending against backdoor attacks using robust statistics. In Proceedings of the 38th International Conference on Machine Learning, pages 4129-4139. PMLR, 2021. 2 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. European conference on computer vision, pages 630-645, 2016. 6 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[25] Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6 +[26] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 2 +[27] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, 2021. 2 +[28] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 2 +[29] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 2, 7, 16 + +[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 13 +[31] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 2 +[32] Yige Li, Nodens Koren, Lingjuan Lyu, Xixiang Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2021. 1, 2, 3, 7 +[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 1, 2, 7 +[34] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems, 2021. 2 +[35] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 1, 2, 3, 7 +[36] Yuntao Liu, Yang Xie, and Ankur Srivastava. Neural trojans. In 2017 IEEE International Conference on Computer Design (ICCD), pages 45-48. IEEE, 2017. 2 +[37] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In NDSS, 2018. 1 +[38] Yingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma, Yousra Aafer, and Xiangyu Zhang. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 1265–1282, 2019. 1, 2, 7 +[39] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 1, 2 +[40] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 3730-3738, 2015. 6, 13 +[41] Peizhuo Lv, Chang Yue, Ruigang Liang, Yunfei Yang, Shengzhi Zhang, Hualong Ma, and Kai Chen. A data-free backdoor injection approach in neural networks. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2671-2688, 2023. 1 +[42] Shiqing Ma, Yingqi Liu, Guanhong Tao, Wen-Chuan Lee, and Xiangyu Zhang. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019. 2 +[43] Michael McCoyd, Won Park, Steven Chen, Neil Shah, Ryan Roggenkemper, Minjune Hwang, Jason Xinyu Liu, and David Wagner. Minority reports defense: Defending against adversarial patches. In Applied Cryptography and Network Security Workshops: ACNS 2020 Satellite Workshops, AIBlock, + +AIHWS, AIoTS, Cloud S&P, SCI, SecMT, and SiMLA, Rome, Italy, October 19-22, 2020, Proceedings, pages 564-582. Springer, 2020. 2 +[44] Geoffrey J McLachlan and Kaye E Basford. Mixture models: Inference and applications to clustering. M. Dekker New York, 1988. 6, 14 +[45] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33, 2020. 1, 2, 4, 7 +[46] Tuan Anh Nguyen and Anh Tuan Tran. Wanet-imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2020. 1, 2, 4, 7, 13, 17 +[47] Jun Ogasawara, Satoru Ikenoue, Hiroko Yamamoto, Motoshige Sato, Yoshifumi Kasuga, Yasue Mitsukura, Yuji Ikegaya, Masato Yasui, Mamoru Tanaka, and Daigo Ochiai. Deep neural network-based classification of cardiotocograms outperformed conventional algorithms. Scientific reports, 11 (1):13367, 2021. 3 +[48] Xiangyu Qi, Tinghao Xie, Yiming Li, Saeed Mahloujifar, and Prateek Mittal. Revisiting the assumption of latent separability for backdoor defenses. In The eleventh international conference on learning representations, 2022. 17 +[49] Ximing Qiao, Yukun Yang, and Hai Li. Defending neural backdoors via generative distribution modeling. In Advances in Neural Information Processing Systems, pages 14004-14013, 2019. 2 +[50] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11957-11965, 2020. 2 +[51] Ahmed Salem, Rui Wen, Michael Backes, Shiqing Ma, and Yang Zhang. Dynamic backdoor attacks against machine learning models. arXiv preprint arXiv:2003.03675, 2020. 1, 2, 7, 21 +[52] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. arXiv preprint arXiv:1906.09453, 2019. 6, 13 +[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 21 +[54] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems, pages 6103-6113, 2018. 2 +[55] Guangyu Shen, Yingqi Liu, Guanhong Tao, Shengwei An, Qiuling Xu, Siyuan Cheng, Shiqing Ma, and Xiangyu Zhang. Backdoor scanning for deep neural networks through k-arm optimization. In International Conference on Machine Learning, 2021. 2 +[56] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022. 2 + +[57] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Advances in Neural Information Processing Systems, 36, 2024. 2 +[58] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[59] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In 30th USENIX Security Symposium (USENIX Security 21), 2021. 2 +[60] Guanhong Tao, Yingqi Liu, Guangyu Shen, Qiuling Xu, Shengwei An, Zhuo Zhang, and Xiangyu Zhang. Model orthogonalization: Class distance hardening in neural networks for better security. In 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022. 2 +[61] Guanhong Tao, Guangyu Shen, Yingqi Liu, Shengwei An, Qiuling Xu, Shiqing Ma, Pan Li, and Xiangyu Zhang. Better trigger inversion optimization in backdoor scanning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13368-13378, 2022. 1, 2, 7, 14 +[62] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. In Advances in Neural Information Processing Systems, pages 8000–8010, 2018. 2, 16 +[63] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 6, 13 +[64] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Clean-label backdoor attacks. 2018. 1, 17 +[65] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 5 +[66] Akshaj Kumar Veldanda, Kang Liu, Benjamin Tan, Prashanth Krishnamurthy, Farshad Khorrami, Ramesh Karri, Brendan Dolan-Gavitt, and Siddharth Garg. Nnoculation: broad spectrum and targeted treatment of backdoored dnns. arXiv preprint arXiv:2002.08313, 2020. 2 +[67] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019. 1, 2, 3, 7, 8, 17 +[68] Ren Wang, Gaoyuan Zhang, Sijia Liu, Pin-Yu Chen, Jinjun Xiong, and Meng Wang. Practical detection of trojan neural networks: Data-limited and data-free cases. In European Conference on Computer Vision, 2020. 1, 2 +[69] Zhenting Wang, Hailun Ding, Juan Zhai, and Shiqing Ma. Training with more confidence: Mitigating injected and natural backdoors during training. In Advances in Neural Information Processing Systems, 2022. 2 +[70] Zhenting Wang, Kai Mei, Hailun Ding, Juan Zhai, and Shiqing Ma. Rethinking the reverse-engineering of Trojan triggers. In Advances in Neural Information Processing Systems, 2022. 2, 3, 7 + +[71] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15074-15084, 2022. 2 +[72] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. Advances in Neural Information Processing Systems, 34:16913-16925, 2021. 1, 7, 17 +[73] Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep k-means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In International Conference on Machine Learning, pages 5363-5372. PMLR, 2018. 3 +[74] Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In USENIX Security Symposium, pages 2237-2254, 2021. 2 +[75] Chong Xiang, Saeed Mahloujifar, and Prateek Mittal. {PatchCleanser}: Certifiably robust defense against adversarial patches for any image classifier. In 31st USENIX Security Symposium (USENIX Security 22), pages 2065-2082, 2022. 2 +[76] Zhen Xiang, David Miller, and George Kesidis. Post-training detection of backdoor attacks for two-class and multi-attack scenarios. In International Conference on Learning Representations, 2021. 2 +[77] Zhen Xiang, Zidi Xiong, and Bo Li. Umd: Unsupervised model detection for x2x backdoor attacks. In International Conference on Machine Learning, 2023. 2 +[78] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10523-10531, 2021. 2 +[79] Qiuling Xu, Guanhong Tao, Jean Honorio, Yingqi Liu, Shengwei An, Guangyu Shen, Siyuan Cheng, and Xiangyu Zhang. Medic: Remove model backdoors via importance driven cloning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20485-20494, 2023. 2 +[80] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. arXiv preprint arXiv:1910.03137, 2019. 2, 7, 16 +[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference, 2016. 6 +[82] Yi Zeng, Han Qiu, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. arXiv preprint arXiv:2012.07006, 2020. 2 +[83] Kaiyuan Zhang, Guanhong Tao, Qiuling Xu, Siyuan Cheng, Shengwei An, Yingqi Liu, Shiwei Feng, Guangyu Shen, PinYu Chen, Shiqing Ma, and Xiangyu Zhang. FLIP: A provable defense framework for backdoor mitigation in federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2 + +[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 14, 21 +[85] Shihao Zhao, Xingjun Ma, Xiang Zheng, James Bailey, Jingjing Chen, and Yu-Gang Jiang. Clean-label backdoor attacks on video recognition models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14443-14452, 2020. 2 +[86] Chen Zhu, W Ronny Huang, Hengduo Li, Gavin Taylor, Christoph Studer, and Tom Goldstein. Transferable clean-label poisoning attacks on deep neural nets. In International Conference on Machine Learning, pages 7614-7623. PMLR, 2019. 2 \ No newline at end of file diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/images.zip b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..79072adf9c6d8cba0d81afbb775673c125794e4e --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f64e4b4b6ec6cb5640401c731782e9343bbb895568acb99afe87d32b5c41176 +size 521170 diff --git a/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/layout.json b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..45fada82b5f92875aaf58a29569d542bf59d4a09 --- /dev/null +++ b/2024/LOTUS_ Evasive and Resilient Backdoor Attacks through Sub-Partitioning/layout.json @@ -0,0 +1,9936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 527, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 527, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 527, + 121 + ], + "type": "text", + "content": "LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 143, + 478, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 143, + 478, + 188 + ], + "spans": [ + { + "bbox": [ + 116, + 143, + 478, + 188 + ], + "type": "text", + "content": "Siyuan Cheng, Guanhong Tao, Yingqi Liu†, Guangyu Shen, Shengwei An, Shiwei Feng, Xiangzhe Xu, Kaiyuan Zhang, Shiqing Ma‡, Xiangyu Zhang Purdue University, †Microsoft, ‡University of Massachusetts at Amherst" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 190, + 515, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 190, + 515, + 217 + ], + "spans": [ + { + "bbox": [ + 78, + 190, + 515, + 217 + ], + "type": "text", + "content": "{cheng535, taog, shen447, an93, feng292, xu1415, zhan4057, xyzhang}@cs.purdue.edu †yingqiliu@microsoft.com, ‡shiqingma@umass.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 243, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 243, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 243, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 270, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 547 + ], + "type": "text", + "content": "Backdoor attack poses a significant security threat to Deep Learning applications. Existing attacks are often not evasive to established backdoor detection techniques. This susceptibility primarily stems from the fact that these attacks typically leverage a universal trigger pattern or transformation function, such that the trigger can cause misclassification for any input. In response to this, recent papers have introduced attacks using sample-specific invisible triggers crafted through special transformation functions. While these approaches manage to evade detection to some extent, they reveal vulnerability to existing backdoor mitigation techniques. To address and enhance both evasiveness and resilience, we introduce a novel backdoor attack LOTUS. Specifically, it leverages a secret function to separate samples in the victim class into a set of partitions and applies unique triggers to different partitions. Furthermore, LOTUS incorporates an effective trigger focusing mechanism, ensuring only the trigger corresponding to the partition can induce the backdoor behavior. Extensive experimental results show that LOTUS can achieve high attack success rate across 4 datasets and 7 model structures, and effectively evading 13 backdoor detection and mitigation techniques. The code is available at https://github.com/Megum1/LOTUS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Backdoor attack is a prominent security threat to Deep Learning applications, evidenced by the large body of existing attacks [5, 19, 37, 51, 64] and defense techniques [20, 32, 35, 67, 72]. It injects malicious behaviors to a model such that the model operates normally on clean samples but mis-classifies inputs that are stamped with a specific trigger. A typical way of injecting such malicious behaviors is through data poisoning [1, 19, 39]. This approach introduces a small set of trigger-stamped images paired with the target label" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 245, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 245, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 245, + 545, + 280 + ], + "type": "text", + "content": "into the training data. Attackers may also manipulate the training procedure [10, 45, 46], and tamper with the model's internal mechanisms [37, 41]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 547, + 413 + ], + "type": "text", + "content": "The majority of existing attacks rely on a uniform pattern [5, 19, 39, 64] or a transformation function [6, 51] as the trigger. The uniform trigger tends to be effective on any input, which can be detected by existing techniques. For instance, trigger inversion methods [20, 38, 67, 68] aim to reverse engineer a small trigger that can induce the target prediction on a set of inputs. According to the results reported in the literature [20, 61, 67], for a number of attacks, it is feasible to invert a pattern that closely resembles the ground-truth trigger and has a substantially high attack success rate (ASR), hence detecting backdoored models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 414, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 557 + ], + "type": "text", + "content": "Recent studies introduce sample-specific invisible attacks [10, 33, 45, 46] that encourage the model to emphasize the correlation between the trigger and the input sample. Although these attacks effectively evade certain detection methods [20, 67], they are not resilient to backdoor mitigation techniques [32, 35, 72]. For instance, a straightforward approach such as fine-tuning the backdoored model using only " + }, + { + "bbox": [ + 304, + 414, + 547, + 557 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 414, + 547, + 557 + ], + "type": "text", + "content": " of the training data can significantly reduce ASR. This is due to the fact that imperceptible trigger patterns are not persistent during the retraining process. Moreover, the sample-specific characteristic of these attacks make them less robust to backdoor mitigation methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "In this paper, we introduce an innovative attack that not only evades backdoor detection approaches but also exhibits resilience against backdoor mitigation techniques. It is a label-specific attack, aiming to misclassify the samples of a victim class to a target class. For the victim-class samples, we divide them into sub-partitions and use a unique trigger for each partition. With such an attack design, existing defense such as trigger inversion is unlikely to find a uniform trigger. This is because the available set of samples used by trigger inversion is likely from different partitions, which makes the detection fail. In addition, we develop a novel trigger focusing technique to ensure that a partition can only be attacked by its designated trigger, not by any" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24798" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "other trigger or trigger combinations. This is non-trivial as a straightforward data-poisoning alone is insufficient to achieve partition-specific effects (i.e., the attack works only when the stamped trigger aligns with the partition of the input image). More details can be found in Section 4. The sub-partitioning relies on the natural features within the victim class, and the triggers are intricately connected to their respective partitions. These two characteristics ensure the connection between inputs and triggers, making our attack more robust against a range of backdoor detection and mitigation techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "type": "text", + "content": "Our contributions are summarized as follows: (1) We propose a new backdoor attack prototype LOTUS (\"Evasive and ResiLient BackdOor A Tacks through Sub-partitioning\") that achieves both evasiveness and resilience. (2) We address a key challenge of the proposed attack, to precisely limit the scope of a trigger to its partition. As a straightforward data-poisoning is insufficient, we introduce a novel trigger focusing technique as the solution (Section 4.2). (3) We conduct an extensive evaluation of LOTUS on 4 datasets and 7 model structures. Our results show that LOTUS achieves a high ASR under a variety of settings. Our trigger focusing method effectively reduces the ASR on undesired victim classes and partitions. Furthermore, our experiments demonstrate that LOTUS is evasive and resilient against 13 state-of-the-art backdoor defense techniques, substantially outperforming existing backdoor attacks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 397, + 289, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 289, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 289, + 566 + ], + "type": "text", + "content": "Threat Model. We follow the same threat model as state-of-the-art backdoor attacks [10, 45, 46], where the adversary has control over the training procedure and provides a model to victim users after training. The adversary's goal is to achieve high attack effectiveness while also ensuring the attack's evasiveness and resilience against defense. LOTUS primarily focuses on label-specific attack. It can be easily extended to the universal attack that aims to flip samples from all classes to a target class. The defender possesses white-box access to the model and a small set of clean samples for each class. She aims to determine if a model contains backdoor or mitigate the backdoor effects based on the validation samples. In our attack, the sub-partitioning function and the corresponding triggers are the secret of the attacker." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 134, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Backdoor Attack. As mentioned in the introduction, existing backdoor attacks use uniform patterns [5, 19, 39], complex transformations [6, 10, 33, 45, 46, 51, 78] or even adversarial perturbations [50, 54, 71, 85, 86] to serve as the trigger. Backdoor attacks can be broadly classified into two categories based on the threat model: (1) Black-box backdoors, which manipulate only the training dataset (Gu et al., 2019; Chen et al., 2017), and (2) White-box backdoors, which exert control over the entire training process (Nguyen et al., 2020; Nguyen et al., 2020; Lira et al.). Our proposed" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 204 + ], + "type": "text", + "content": "attack, LOTUS belongs to the white-box backdoor category, aligning with the existing works. Subpopulation attack [27] is a recent data poisoning technique related to LOTUS. It is an availability attack, and its primary objective is to decrease the test accuracy of a specific subpopulation within the dataset. In contrast, LOTUS is a comprehensive backdoor attack with the intention of injecting a backdoor into the model. Therefore, these two attacks differ significantly. Subpopulation attack does not involve trigger injection or require the implementation of trigger focusing, making it distinct from LOTUS in terms of its objectives and mechanisms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 205, + 548, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 548, + 338 + ], + "type": "text", + "content": "Backdoor Defense. Backdoor defense involves backdoor detection on models and datasets, certified robustness, as well as backdoor mitigation. Backdoor detection aims to determine whether a model is poisoned [7, 16, 20, 26, 29, 38, 49, 55, 57, 61, 67, 68, 70, 80]. Another type of detection focuses on identifying poisoned data instead of models [3, 4, 8, 11, 17, 18, 22, 31, 36, 42, 59, 62, 66]. Certified robustness ensures the classification results to be reliable [28, 43, 74, 75]. Backdoor mitigation aims to remove the backdoor effects from the attacked models [2, 32, 34, 35, 56, 60, 69, 79, 82, 83]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 351, + 410, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 410, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 410, + 364 + ], + "type": "text", + "content": "3. Attack Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": "We formally define our attack in this section. For a typical classification task, given " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "(\\pmb{x},y)\\sim \\mathcal{D}" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": " where the sample " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\pmb {x}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "y\\in \\{1,2,\\dots N\\}" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": ", the goal is to train a classifier " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "M_{\\theta}:\\mathbb{R}^{d}\\to \\{1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": ", such that parameters " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\theta = \\arg \\max_{\\theta}P_{(\\pmb {x},y)\\sim \\mathcal{D}}[M_{\\theta}(\\pmb {x}) = y]" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": ". Typically, the cross-entropy loss " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(y_p,y)" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "y_{p}" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": " is the predicted label) is utilized for achieving the goal. In this case, the optimization problem can be expressed as " + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\theta = \\arg \\min_{\\theta}\\mathbb{E}_{(\\pmb {x},y)\\sim \\mathcal{D}}[\\mathcal{L}(M_{\\theta}(\\pmb {x}),y)]" + }, + { + "bbox": [ + 304, + 373, + 547, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": "Backdoor attack aims to derive a classifier " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{M}}_{\\bar{\\theta}}:\\mathbb{R}^d\\to" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " such that compromised parameters " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\bar{\\theta} =" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " arg max " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\overline{\\theta} P_{(\\pmb {x},y)\\sim \\mathcal{D}} [\\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\pmb {x}) = y\\& \\overline{\\mathcal{M}}_{\\bar{\\theta}}(\\mathbb{T}\\oplus \\pmb{x}_V) = y_T]," + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " is the trigger and " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\mathbb{T}\\oplus \\pmb{x}_V" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " injects the trigger to a victim input sample " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "x_{V}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " whose label is " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "y_{V}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " .Symbol " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "y_{T}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " denotes the attack target label. Backdoor attacks can be mainly classified to universal attack that aims to flip a sample " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " of any class with " + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "inline_equation", + "content": "\\mathbb{T}" + }, + { + "bbox": [ + 304, + 471, + 548, + 675 + ], + "type": "text", + "content": " to the target label, and label-specific attack that aims to flip any sample of a specific victim class to the target label. Based on trigger patterns, they can be classified to input-independent backdoor or static backdoor that has a fixed trigger pattern for all victim samples, and dynamic trigger that has changing patterns for different inputs. Our attack is a label-specific dynamic backdoor attack. Extending to other scenarios is relatively straightforward, e.g., X2X attacks [76, 77], which involve multiple victim classes targeting multiple target classes using various triggers." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": "Assume there exists a partitioning algorithm " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n: \\mathbb{R}^d \\to \\{p_1, p_2, \\dots, p_n\\}" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " that separates input samples to " + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 677, + 547, + 715 + ], + "type": "text", + "content": " partitions. In our attack, victim samples (samples from the victim class)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24799" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "are partitioned to " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " groups using " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and each partition " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is assigned a unique trigger " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ", such that only " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i \\oplus \\boldsymbol{x}_V^{p_i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " can trigger the backdoor, where " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2, \\dots, n\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_V^{p_i}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " denotes the victim samples in the " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "-th partition. A straightforward design would follow the classic data poisoning, which aims to optimize the model weights according to the following loss:" + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 160, + 288, + 208 + ], + "blocks": [ + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "lines": [ + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 288, + 208 + ], + "type": "image", + "image_path": "06dbad3ef839dbb2c8815594f9f71ea87f92b01b321f3d8cf79d18b23a0f1fcc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": "The first loss term Benign Utility Loss aims to ensure the high benign accuracy of the model. The second term, Attack Target Loss, means that a trigger " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": " can cause the " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": "-th partition samples of the victim class " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{V}^{p_i}" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": " to misclassify, which is our attack goal. However, simple data poisoning cannot effectively bound the attack scope. As a result, a trigger for a particular partition can easily induce misclassifications for other partitions. That is, " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_j \\oplus \\pmb{x}_{V}^{p_i}" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ", is misclassified to " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "y_{T}" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ". Besides, a trigger for a correctly-assigned partition of non-victim samples (samples from class " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\neg V" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ", denoting the classes other than the victim class " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ") can induce misclassification. That is " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i \\oplus \\pmb{x}_{\\neg V}^{p_i}" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": " is misclassified to " + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "inline_equation", + "content": "y_{T}" + }, + { + "bbox": [ + 46, + 209, + 289, + 459 + ], + "type": "text", + "content": ". Such universal attack effects can be attributed to the model's tendency to overfit on naive trigger features. For instance, when it encounters any trigger, it immediately predicts the target class without verifying if the background image aligns with the trigger according to the partitioning criteria. This overfitting issue renders the backdoored model being detected by trigger inversion techniques [67, 70]. Moreover, these attack effects are not resilient to existing backdoor mitigation methods [32, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "text", + "content": "Our objective is to establish a clear one-to-one correspondence between " + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{V}^{p_i}" + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "text", + "content": ". That is, only " + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}" + }, + { + "bbox": [ + 46, + 460, + 288, + 544 + ], + "type": "text", + "content": " can cause misclassification. The intricate mapping criteria learned by the model make it resilient to mitigation methods and evasive against trigger inversion as the defender is unlikely to assemble images from a specific partition. We hence aim to derive the following loss function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 552, + 304, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 304, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 304, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {(\\pmb {x}, y) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\pmb {x}), y) ] + \\sum_ {i = 1} ^ {n} (\\mathbb {E} _ {(\\pmb {x} _ {V} ^ {p _ {i}}, y _ {V}) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\overline {{\\theta}}} (\\mathbb {T} _ {i} \\oplus \\pmb {x} _ {V} ^ {p _ {i}}), y _ {T}) ] \\\\ + \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}, y _ {\\neg V}\\right) \\sim \\mathcal {D}} [ \\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left(\\mathbb {T} _ {i} \\oplus \\boldsymbol {x} _ {\\neg V} ^ {p _ {i}}\\right), y _ {\\neg V}) ] \\longleftrightarrow \\text {L a b e l - s p e c i f i c L o s s} \\\\ + \\underbrace {\\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}} , y _ {V}\\right) \\sim \\mathcal {D}} \\left[ \\right. \\sum_ {\\substack {\\mathcal {T} \\in \\mathcal {P} \\left( \\right.\\{\\mathbb {T} _ {1} , \\dots , \\mathbb {T} _ {n} \\left. \\right\\}}} \\mathcal {L} \\left(\\overline {\\mathcal {M}} _ {\\bar {\\theta}} \\left(\\mathcal {T} \\oplus \\boldsymbol {x} _ {p _ {i}}\\right), y _ {V}\\right)\\left. \\right]\\left. \\right]} _ {- \\{\\{\\}, \\{\\mathbb {T} _ {i} \\} \\}} \\xleftarrow {} \\text {D y n a m i c L o s s} \\tag{2} \\\\ \\end{array}", + "image_path": "2b1e0a0a1295394149ff1c368f80b2bf089d9ef14f183a720aad9c2612319bd9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "text", + "content": "Note that compared to Equation 1, we introduce two additional terms, i.e., Label-specific Loss and Dynamic Loss in Equation 2. Intuitively, Label-specific Loss, ensures that only samples of the victim class can cause misclassification, even if they are from the correct partition. Here " + }, + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\neg V" + }, + { + "bbox": [ + 46, + 641, + 289, + 714 + ], + "type": "text", + "content": " denotes the classes other than the victim class. The last term Dynamic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "Loss controls that for a particular partition, only the corresponding trigger can cause misclassification, and any other trigger, or combination of/with other triggers shall be correctly predicted as the victim class. In particular, " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " is a subset of all possible triggers/combinations " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\{\\mathbb{T}_1,\\dots ,\\mathbb{T}_n\\})" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " excluding empty " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{T}_i\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": ". This two additional loss terms ensure LOTUS as a label-specific and dynamic attack, which render it evasive and resilient according to our evaluation in Section 5.3 and 5.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 194, + 440, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 440, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 440, + 208 + ], + "type": "text", + "content": "4. Detailed Attack Design" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 214, + 547, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 214, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 214, + 547, + 334 + ], + "type": "text", + "content": "The overview of LOTUS is shown in Figure 1. Victim class input samples are first separated to partitions. We then apply unique triggers to samples from the corresponding partitions, whose labels are set to the target class. Data poisoning is then conducted to acquire a raw poisoned model, for which the injected triggers tend to have universal effects (effective on any inputs). To address this problem, LOTUS further introduces a trigger focusing step that strictly limits the attack scope of each trigger. It finally produces a trojaned model with triggers that are evasive and resilient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 335, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 547, + 372 + ], + "type": "text", + "content": "In the following, we elaborate two major components of LOTUS, namely, victim-class sample partitioning and trigger focusing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 381, + 484, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 484, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 484, + 396 + ], + "type": "text", + "content": "4.1. Victim-class Sample Partitioning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "text", + "content": "LOTUS separates a set of victim-class samples into multiple partitions, and injects different triggers to different partitions. We propose two ways to partition input samples. The first is explicit partitioning that leverages a subset of explicit attributes of the victim class (e.g., hair color and w./ or w./o. glasses for face recognition). Assume " + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "text", + "content": " attributes are used and each attribute has " + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "text", + "content": " possible values. This allows to generate " + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "inline_equation", + "content": "t^k" + }, + { + "bbox": [ + 304, + 401, + 547, + 556 + ], + "type": "text", + "content": " partitions. The first two columns in Figure 2 show a partitioning based on the taxonomy attribute of the bird class. Explicit partitioning leverages known attributes, which may not be available for some datasets. We hence introduce an advanced partitioning method that is applicable to arbitrary datasets in the following." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 714 + ], + "type": "text", + "content": "The second partitioning scheme is implicit, meaning that human uninterpretable features are used in partitioning. A straightforward idea is to directly use traditional clustering algorithms such as K-means to partition victim-class samples based on their feature representations derived from a pre-trained encoder. However, according to our experiment in Appendix L.1, such a naive method does not work well. The root cause is that K-means is a clustering algorithm on a set of known data points and does not consider generalization to unseen data points. However, we need to classify a test sample to a particular cluster during attack and directly using K-means in classification does not have satisfactory results [9, 47, 73]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24800" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 69, + 521, + 178 + ], + "blocks": [ + { + "bbox": [ + 75, + 69, + 521, + 178 + ], + "lines": [ + { + "bbox": [ + 75, + 69, + 521, + 178 + ], + "spans": [ + { + "bbox": [ + 75, + 69, + 521, + 178 + ], + "type": "image", + "image_path": "cc131953041ebdf5d41c2e0e7723d3092436852b4a6eb1d871da792c4694ff26.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 179, + 352, + 190 + ], + "lines": [ + { + "bbox": [ + 241, + 179, + 352, + 190 + ], + "spans": [ + { + "bbox": [ + 241, + 179, + 352, + 190 + ], + "type": "text", + "content": "Figure 1. Overview of LOTUS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 84, + 209, + 251, + 338 + ], + "blocks": [ + { + "bbox": [ + 84, + 209, + 251, + 338 + ], + "lines": [ + { + "bbox": [ + 84, + 209, + 251, + 338 + ], + "spans": [ + { + "bbox": [ + 84, + 209, + 251, + 338 + ], + "type": "image", + "image_path": "337a33f750e6eb2e76af9a799d11232c1674cd96d698f78c3f556edcb477a951.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 340, + 268, + 351 + ], + "lines": [ + { + "bbox": [ + 64, + 340, + 268, + 351 + ], + "spans": [ + { + "bbox": [ + 64, + 340, + 268, + 351 + ], + "type": "text", + "content": "Figure 2. Explicit (left) and implicit (right) partitioning." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 373, + 287, + 483 + ], + "blocks": [ + { + "bbox": [ + 48, + 373, + 287, + 483 + ], + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 483 + ], + "type": "image", + "image_path": "dde03794c3b69714c6a5e285dce415dd744d6f4620290b696e1039e66c10da50.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 484, + 261, + 495 + ], + "lines": [ + { + "bbox": [ + 70, + 484, + 261, + 495 + ], + "spans": [ + { + "bbox": [ + 70, + 484, + 261, + 495 + ], + "type": "text", + "content": "Figure 3. Implicit partitioning with surrogate model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": "We hence introduce a surrogate model to help sample partitioning. Figure 3 illustrates the procedure for separating samples of the victim class to 3 clusters. The surrogate model has the same structure as the victim model to reduce complexity caused by structural differences. On the bottom left, the features of samples from victim class " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": " are extracted using a pre-trained encoder. We then use a traditional clustering method such as K-means to partition these samples into 3 different sub-classes based on their features. We assign labels " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n + 2" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": " to samples from the respective sub-classes. They are then combined with samples from the original classes 1 to " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": " (excluding the victim class " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": ") to form a new dataset consisting of " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n + 2" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": " classes. The surrogate model is trained on this new dataset with " + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "inline_equation", + "content": "n + 2" + }, + { + "bbox": [ + 47, + 498, + 289, + 713 + ], + "type": "text", + "content": " classes. The idea is to use K-means to provide a meaningful prior separation and then use classifier training to achieve generalizability. Furthermore, the decision boundaries by the surrogate model have the classes other than the victim class" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "text", + "content": "in consideration, whereas those by distances to centroids of K-means clusters only have samples of the victim class in consideration. After the training converges, the surrogate model is utilized to determine the partition of a test sample. That is, the partition index can be derived from the its classification outcome (i.e., the class with largest logits from classes " + }, + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "inline_equation", + "content": "n + 2" + }, + { + "bbox": [ + 304, + 213, + 547, + 344 + ], + "type": "text", + "content": "). The last two columns in Figure 2 show two implicit partitions of the \"cat\" class. Observe that the partitions are largely uninterpretable, which makes the attack more stealthy compared to using explicit attributes which are public." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 417 + ], + "type": "text", + "content": "Handle Potential Imbalanced Examples. We control that for any partitioning, the sizes of each partition are roughly the same, which mitigates the potential of causing partitioning bias. This is achieved by removing samples from exceptionally large clusters. In practice, such a removal is rarely needed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 426, + 409, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 409, + 440 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 409, + 440 + ], + "type": "text", + "content": "4.2. Trigger Focusing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": "After partitioning, LOTUS aims to limit each trigger to its own partition, preventing it from attacking other partitions or classes. To achieve this, we design a trigger focusing technique during training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": "A straightforward idea is to strictly follow the definition in Equation 2 to bound the trigger scope. However, the last term, which aims at stamping all combinations of triggers that are different from " + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "inline_equation", + "content": "\\{\\mathbb{T}_i\\}" + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": " to a sample of partition " + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": " and setting the label to " + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "inline_equation", + "content": "y_V" + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": ", is extremely expensive. The number of combinations is " + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "inline_equation", + "content": "(2^n - 2)" + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": ", which grows exponentially with the increase of the number of partitions " + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 494, + 547, + 614 + ], + "type": "text", + "content": ". Moreover, the inclusion of a substantial number of additional samples will not only slow down the training but also imbalance the dataset, ultimately impacting the overall performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": "Adversarial Poisoning Is Insufficient. Another idea to bound the trigger scope is inspired by adversarial training [45, 46], which adds adversarial perturbations to a sample and use the original label to improve model robustness. To suppress the undesirable attack effect in our context, we could inject triggers that are not for a partition " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_j" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": ", to samples of " + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 617, + 548, + 714 + ], + "type": "text", + "content": " and set the injected samples' labels to the victim class. This approach is referred" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24801" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 281, + 91 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 281, + 91 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 281, + 91 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 281, + 91 + ], + "type": "image", + "image_path": "3de8fbb4d83b92baf7a2dc07941c0d11538840469c1ed598f92c2f009d2679d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 473, + 91 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 473, + 91 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 473, + 91 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 473, + 91 + ], + "type": "image", + "image_path": "3b01cdcfbb26026175545fca76f696d90b0b60a1ddec610860ec7db7be7b66b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 95, + 283, + 116 + ], + "blocks": [ + { + "bbox": [ + 50, + 95, + 283, + 116 + ], + "lines": [ + { + "bbox": [ + 50, + 95, + 283, + 116 + ], + "spans": [ + { + "bbox": [ + 50, + 95, + 283, + 116 + ], + "type": "image", + "image_path": "a01d5fb6ad6b31e73a9a41de05811ac8c0b4545c507968df590ee7db44932562.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 285, + 97, + 534, + 116 + ], + "blocks": [ + { + "bbox": [ + 285, + 97, + 534, + 116 + ], + "lines": [ + { + "bbox": [ + 285, + 97, + 534, + 116 + ], + "spans": [ + { + "bbox": [ + 285, + 97, + 534, + 116 + ], + "type": "image", + "image_path": "3cea3593421c93681799a945755df62a9ef1d00d79d660ecc5baaa97cc6fac3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 121, + 338, + 140 + ], + "blocks": [ + { + "bbox": [ + 50, + 121, + 338, + 140 + ], + "lines": [ + { + "bbox": [ + 50, + 121, + 338, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 121, + 338, + 140 + ], + "type": "image", + "image_path": "b6a8303a766cf65c5fa65185b4f1c8a6836953ec1de38ad916c8421a07355961.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 140, + 209, + 221 + ], + "blocks": [ + { + "bbox": [ + 48, + 140, + 209, + 221 + ], + "lines": [ + { + "bbox": [ + 48, + 140, + 209, + 221 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 209, + 221 + ], + "type": "image", + "image_path": "d95ca0c404b8c601d0269a19c1a03722c8fb870c86a2517ea2740cc1aa8888ad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 223, + 160, + 233 + ], + "lines": [ + { + "bbox": [ + 97, + 223, + 160, + 233 + ], + "spans": [ + { + "bbox": [ + 97, + 223, + 160, + 233 + ], + "type": "text", + "content": "(a) Data-poisoning" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 216, + 141, + 298, + 221 + ], + "blocks": [ + { + "bbox": [ + 216, + 141, + 298, + 221 + ], + "lines": [ + { + "bbox": [ + 216, + 141, + 298, + 221 + ], + "spans": [ + { + "bbox": [ + 216, + 141, + 298, + 221 + ], + "type": "image", + "image_path": "d620f783223ff032a3c2928de46445a6b25bf06b96921c11e907ed551f51e581.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 223, + 339, + 233 + ], + "lines": [ + { + "bbox": [ + 254, + 223, + 339, + 233 + ], + "spans": [ + { + "bbox": [ + 254, + 223, + 339, + 233 + ], + "type": "text", + "content": "(b) Adversarial-poisoning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 299, + 141, + 376, + 221 + ], + "blocks": [ + { + "bbox": [ + 299, + 141, + 376, + 221 + ], + "lines": [ + { + "bbox": [ + 299, + 141, + 376, + 221 + ], + "spans": [ + { + "bbox": [ + 299, + 141, + 376, + 221 + ], + "type": "image", + "image_path": "c4c6ae66c30c8006ab6391446ad4e279f078a64c97e82668c9236baff3f52fe5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 235, + 412, + 246 + ], + "lines": [ + { + "bbox": [ + 179, + 235, + 412, + 246 + ], + "spans": [ + { + "bbox": [ + 179, + 235, + 412, + 246 + ], + "type": "text", + "content": "Figure 4. Decision boundaries for different poisoning strategies." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 385, + 141, + 466, + 221 + ], + "blocks": [ + { + "bbox": [ + 385, + 141, + 466, + 221 + ], + "lines": [ + { + "bbox": [ + 385, + 141, + 466, + 221 + ], + "spans": [ + { + "bbox": [ + 385, + 141, + 466, + 221 + ], + "type": "image", + "image_path": "967060fe0534c11283b8a578fa18db15ccc06c0f4564568ec3cfa56a3b30901a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 223, + 498, + 233 + ], + "lines": [ + { + "bbox": [ + 432, + 223, + 498, + 233 + ], + "spans": [ + { + "bbox": [ + 432, + 223, + 498, + 233 + ], + "type": "text", + "content": "(c) Trigger-focusing" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 141, + 545, + 221 + ], + "blocks": [ + { + "bbox": [ + 467, + 141, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 467, + 141, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 467, + 141, + 545, + 221 + ], + "type": "image", + "image_path": "cbb289b1d6414c57fa083e3ed2c665bb1b9744a69ca4da459b1581d06bd713fa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "text", + "content": "to as adversarial poisoning. However, it is only effective in eliminating individual non-matched triggers " + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_j" + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "text", + "content": ", but fails for trigger combinations that contain the matched trigger " + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "inline_equation", + "content": "[\\mathbb{T}_i,\\mathbb{T}_j]" + }, + { + "bbox": [ + 46, + 268, + 287, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": "Figure 4 presents a visualization of decision boundaries for various poisoning strategies, namely: (a) Straightforward data-poisoning; (b) Adversarial-poisoning; and (c) Trigger-focusing (which will be discussed in the next paragraph). Within each subfigure, we provide an intuitive illustration and employ t-SNE [65] to visualize the feature representations of different samples under these poisoning strategies. The experiment is conducted on the CIFAR-10 dataset using the ResNet18 model, and we utilize implicit partitioning to create four distinct partitions. In the figure, a hollow closed lock is used to denote clean images " + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_{v}^{p_{i}}" + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": " in the victim class of partition " + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": ", while a red opened lock is used to represent clean images of the target class. Triggers are depicted as keys with various colors. According to our objective, only the red keys, signifying the correct trigger for partition " + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": ", can unlock the lock, crossing the red decision boundary, and be classified as the target class. Keys of different colors, signifying various triggers or combinations, are unable to unlock the lock and remain within the victim class region. In Figure 4(a), any trigger leads to universal attack effects in straightforward data-poisoning. Observe any key, denoting a trigger, can unlock the lock and cross the boundary without limitations. The t-SNE visualization on real data on the right aligns with the illustration on the left. In contrast, adversarial-poisoning, as depicted in (b), mitigates the impact of samples with unmatched individual triggers, as represented by the green key. However, trigger combinations containing both the matched trigger " + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i" + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": " and unmatched trigger " + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_j" + }, + { + "bbox": [ + 46, + 318, + 289, + 714 + ], + "type": "text", + "content": ", as shown by the key with half red and half green, still lead to misclassification. Similarly, in the t-SNE visualization, the yellow triangles, which represent this type of trigger combination, are substantially close to the red triangles, denoting the strictly matched triggers. This indicates" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 268, + 477, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 268, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 477, + 281 + ], + "type": "text", + "content": "the insufficiency of adversarial-poisoning." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": "Efficient and Effective Trigger Focusing. Inspired by the observation in Figure 4, we propose a novel trigger focusing method that can effectively bound trigger scopes and is in the mean time cost-effective. In addition to adversarial poisoning that stamps samples in a partition " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": " with individual out-of-partition triggers " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_j" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": ") and sets their labels to the victim class " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "y_V" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": ", it further stamps samples in partition " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": " with a pair of triggers " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "[\\mathbb{T}_i, \\mathbb{T}_j]" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": "), that is, the partition's trigger and another different partition's trigger, and sets their labels to " + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "inline_equation", + "content": "y_V" + }, + { + "bbox": [ + 304, + 282, + 547, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 422, + 545, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 422, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 318, + 422, + 545, + 471 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sum_ {i = 1} ^ {n} \\mathbb {E} _ {\\left(\\boldsymbol {x} _ {V} ^ {p _ {i}}, y _ {V}\\right) \\sim \\mathcal {D}} [ \\sum_ {j = 1, j \\neq i} ^ {n} (\\mathcal {L} (\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} (\\mathbb {T} _ {j} \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}), y _ {V}) \\tag {3} \\\\ \\left. + \\mathcal {L} \\left(\\overline {{\\mathcal {M}}} _ {\\bar {\\theta}} \\left([ \\mathbb {T} _ {i}, \\mathbb {T} _ {j} ] \\oplus \\boldsymbol {x} _ {V} ^ {p _ {i}}\\right), y _ {V})\\right) \\right] \\\\ \\end{array}", + "image_path": "a993de1a161e35e6eed59951e1c821ac9d955512bf0d85888edd30238a468d96.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "content": "Our approach, with the new dynamic loss term expressed in Equation 3, requires only " + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "inline_equation", + "content": "(2n - 2)" + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "content": " trigger combinations, which increases linearly with the growth of partitions " + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 479, + 547, + 538 + ], + "type": "text", + "content": ". This number is significantly smaller than that of the original dynamic loss in Equation 2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "text", + "content": "Intuitively, the different labels of samples " + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_i\\oplus \\pmb{x}_V^{p_i}" + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "inline_equation", + "content": "[\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}" + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "text", + "content": " enable the model to learn new behaviors. As such, further stamping any other partition triggers to " + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "inline_equation", + "content": "[\\mathbb{T}_i,\\mathbb{T}_j]\\oplus \\pmb{x}_V^{p_i}" + }, + { + "bbox": [ + 305, + 538, + 547, + 609 + ], + "type": "text", + "content": " yields the same classification result, which is the victim class. Please refer to Appendix D for a detailed reasoning and theoretical analysis." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 611, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 547, + 658 + ], + "type": "text", + "content": "In Figure 4(c), it is noteworthy that trigger combinations are effectively excluded from the target class and only the trigger that matches the victim partition can cause the misclassification, well aligning with our attack objective." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 669, + 377, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 377, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 377, + 681 + ], + "type": "text", + "content": "5. Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 712 + ], + "type": "text", + "content": "In this section, we evaluate on 4 benchmark datasets and 7 model structures to demonstrate the attack effectiveness of" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24802" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "LOTUS (Section 5.2). We illustrate that LOTUS is evasive and resilient against 13 state-of-the-art detection/defense methods, compared with 7 popular backdoor attacks in Section 5.3 and 5.4. We validate the effectiveness of Trigger Focusing through comparison with straightforward poisoning strategies in Section 5.5. We also extend LOTUS to universal attacks in Appendix G. Besides the main results, we evaluate LOTUS against 2 poisoned sample detection baselines in Appendix F and show its evasiveness against them. Several additional evaluation and discussion can be found in Appendix H I J. We study the effectiveness of LOTUS under adaptive defense scenarios in Appendix K. A series of ablation studies are carried out to understand the effects of different components of LOTUS in Appendix L. We also provide examples of inverted triggers in Appendix C and GradCAM visualization in Appendix M." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 272, + 155, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 155, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 155, + 285 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 290, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 386 + ], + "type": "text", + "content": "We evaluate LOTUS on 4 widely-used benchmarks, CIFAR-10 [30], CIFAR-100 [30], CelebA [40], and restricted ImageNet (RImageNet) [12, 52, 63]. Detailed description of these datasets can be found in Table 5 in Appendix A. We conduct experiments on 7 different model structures, including VGG11 [58], VGG16 [58], ResNet18 [24], ResNet50 [24], Pre-act ResNet-34 (PRN34) [23], WideResNet (WRN) [81], and Densenet [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 387, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 507 + ], + "type": "text", + "content": "We leverage several sub-partitioning methods to partition samples from the victim class. We utilize secondary labeling, e.g., various cat species, to create clear and explicit partitions. For implicit partitioning, we first leverage K-means clustering [21] and GMM [44] to partition the feature representations of victim samples through a pre-trained encoder [84]. Then we train a surrogate model to learn the partitioning principle, which serves as the implicit sub-partitioner (Section 4.1). Details of the sub-partitioning and encoder can be found in Appendix B." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 165, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 165, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 165, + 528 + ], + "type": "text", + "content": "5.2. Attack Effectiveness" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": "We evaluate the performance of LOTUS on various datasets, model structures and partitioning methods. Table 1 presents the results. For all the experiments, we use the first class of each dataset as the victim and the last class as the target. We generate 4 partitions for the victim class throughout all datasets and model structures. Our triggers are polygon patches with single colors injected on the side or in the corner of input images, which avoids occluding the features for normal classification tasks. Example images with triggers can be found in Figure 10 in Appendix. The top two blocks in Table 1 (separated by the double lines) show the results for implicit partitioning, and the bottom for explicit partitioning. For K-means clustering, ASRs are at least " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "89.00\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": ", with the highest ASR of " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "94.30\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " for ResNet18 on CIFAR-10, while the degradation of benign accuracy is within " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "1.07\\%" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 148 + ], + "type": "text", + "content": "Table 1. Evaluation of attack effectiveness. The first three columns denote different partitioning algorithms (PA), datasets, and model structures. The following columns present the original accuracy of clean models (Acc.), benign accuracy of the backdoored models (BA), the attack success rate when stamping a trigger on the proper partition (ASR), and the average ASR when stamping other triggers and trigger combinations, with the standard deviation) (ASR-other)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 149, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 307, + 149, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 307, + 149, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 545, + 292 + ], + "type": "table", + "html": "
PADatasetModelAcc.BAASRASR-other
K-meansCIFAR-10VGG1192.16%92.04%93.80%4.77% ± 19.27%
ResNet1895.22%94.71%94.30%4.39% ± 17.08%
CIFAR-100Densenet75.14%75.15%92.00%4.36% ± 14.24%
PRN3474.70%74.52%89.00%5.43% ± 13.50%
CelebAWRN80.47%79.40%92.33%6.87% ± 17.49%
RImageNetResNet5097.77%97.19%93.87%2.16% ± 19.34%
GMMCIFAR-10ResNet1895.22%94.59%90.70%4.80% ± 21.38%
CIFAR-100PRN3474.70%74.02%91.00%2.21% ± 12.57%
CelebAWRN80.47%79.66%92.53%5.39% ± 16.77%
RImageNetVGG1696.51%95.93%93.52%3.11% ± 14.39%
Sec.RImageNetVGG1696.51%96.36%96.50%1.79% ± 13.24%
ResNet5097.77%97.08%92.50%2.14% ± 16.53%
", + "image_path": "80b1ad6de9548d9a2c67d4683224f29ab25a7e9254729a5c80e1fbdca4f51587.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 325, + 348, + 527, + 400 + ], + "blocks": [ + { + "bbox": [ + 304, + 304, + 545, + 347 + ], + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 347 + ], + "type": "text", + "content": "Table 2. Evaluation of label specificity. ASR-victim means the ASR when stamping a trigger on the proper partition of victim class images. ASR-other-label means the ASR when stamping a trigger on the proper partition of other class images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 348, + 527, + 400 + ], + "lines": [ + { + "bbox": [ + 325, + 348, + 527, + 400 + ], + "spans": [ + { + "bbox": [ + 325, + 348, + 527, + 400 + ], + "type": "table", + "html": "
DatasetNetworkASR-victimASR-other-label
CIFAR10ResNet1893.80%14.37%
CIFAR100Densenet92.00%11.23%
CelebAWRN92.33%19.67%
RImageNetVGG1693.52%12.22%
", + "image_path": "5415a9f527459b96be1455b804c7273712422d2b4af8d1ca789c1d88b760a752.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "content": "indicates LOTUS is a highly effective attack, which injects successful malicious behaviors to the model while maintains its benign utility. The last column shows the ASR when trigger/trigger-combinations other than a partition's trigger are stamped on the partition (ASR-other). Observe that the average ASR-other is less than " + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "inline_equation", + "content": "6.87\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "content": ", delineating the effectiveness of trigger focusing (a trigger is only effective for the corresponding partition). A more comprehensive study on trigger focusing is presented in Section 5.5. We have similar observations for using GMM in implicit partitioning. For the explicit secondary labeling, LOTUS can achieve an ASR over " + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "inline_equation", + "content": "92.50\\%" + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "content": " and a small ASR-other. The better performance of LOTUS using secondary labeling can be attributed to the fact that the victim class in RImageNet is merged from a set of similar classes in ImageNet. Those classes are naturally separable, which can be easily differentiated by the model when triggers are injected on different partitions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "Note that the ASR of LOTUS is slightly lower than the existing attacks (as shown in the \"No Defense\" column in Table 3). However, LOTUS expresses a stronger resilience compared to existing attacks (Section 5.4). This is a trade-off between attack effectiveness and resilience. More discussion can be found in Appendix J." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Besides, we also evaluate the label specificity of LOTUS on several models. Results are presented in Table 2. Observe that even if the trigger is stamped on the proper partition" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24803" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "of the input image, the ASR-other-label is low (" + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "inline_equation", + "content": "< 20\\%" + }, + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": ") because the input image is not of the victim class. The result shows that LOTUS exhibits a high level of label specificity. Furthermore, LOTUS offers an easy extension into universal attack scenarios through the integration of explicit partitioning techniques. Detailed examples can be found in Section G." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 166, + 257, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 257, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 257, + 178 + ], + "type": "text", + "content": "5.3. Evasiveness against Backdoor Detection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 288, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 288, + 388 + ], + "type": "text", + "content": "In this section, we study the evasiveness of LOTUS against 4 well-known trigger-inversion based backdoor detection methods, including Neural Cleansse (NC) [67], Pixel [61], ABS [38], and FeatRE [70]. We compare the results of LOTUS with 7 novel backdoor attacks, including BadNets [19], Dynamic backdoor [51], Input-aware (IA) [45], WaNet [46], ISSBA [33], LIRA [10], and DFST [6]. For fair comparison, we launch all backdoor attacks on ResNet18 models trained on CIFAR-10. As LOTUS is a label-specific attack, we implement all other attacks in label-specific setting, where the poisoned samples are composed of images from victim class 0 stamped with the trigger and labeled as the target class 9. Besides, all detection methods are required to invert triggers based on 100 clean validation images from the victim class, targeting to labels other than it. We follow all the other settings and techniques of the original papers to implement the attack and detection methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 389, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 288, + 675 + ], + "type": "text", + "content": "Figure 5 illustrates the detection results, where the x-axis denotes different attacks and the y-axis denotes the decision scores of each baseline. The thresholds are highlighted in red dashed lines. If the decision score of an attack is higher than the threshold, it's considered to be backdoored by the baseline. Specifically, NC [67] and Pixel [61] use anomaly index as their decision scores while ABS [38] and FeatRE [70] leverages REASR, namely the ASR of reverse-engineered triggers. Observe that NC, Pixel, ABS are effective against several attacks, including BadNets, Dynamic, ISSBA, LIRA and DFST, while leaving other advanced attacks, i.e., WaNet, IA and LOTUS. FeatRE, on the other hand, observes internal linear separability properties of existing backdoors and improves the trigger inversion process, which is able to detect the advanced backdoors operating in the feature space. Figure 5(d) shows that it can detect both IA and WaNet, but still fails to detect LOTUS. This illustrates that LOTUS is more evasive than all these baseline attacks. The underlying reason is that LOTUS leverages partitioning secrets and trigger focusing, which breaks the linear separability assumption. Without knowledge of partitioning, it's unlikely to invert a trigger with high ASR, and hence unlikely to detect the backdoor. Examples of inverted triggers can be found in Appendix C." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "We also test LOTUS in the adaptive defense scenario, where the defender can create partitions before detection. The results in Appendix K demonstrate that LOTUS is re" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "silent against adaptive defense strategies, as guessing the correct partitioning is challenging." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 97, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 547, + 156 + ], + "type": "text", + "content": "Besides trigger inversion methods, we also evaluate LOTUS using meta-classifiers, e.g., MNTD [80] and ULP [29], which train model-level classifiers for detection. Results in Appendix E show that LOTUS is evasive against them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 165, + 512, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 165, + 512, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 165, + 512, + 177 + ], + "type": "text", + "content": "5.4. Resilience against Backdoor Mitigation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 183, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 547, + 303 + ], + "type": "text", + "content": "In this section, we study the resilience of LOTUS against 4 state-of-the-art backdoor mitigation methods, including standard Fine-tuning, Fine-pruning [35], NAD [32], and ANP [72]. We compare the results of LOTUS with 7 novel backdoor attacks. For fair comparison, all the models are trained using VGG11 on CIFAR-10 dataset. For each mitigation method, we assume the access to " + }, + { + "bbox": [ + 304, + 183, + 547, + 303 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 183, + 547, + 303 + ], + "type": "text", + "content": " of the training data. Besides, some standard input argumentation techniques are used, e.g., random cropping and horizontal flipping. We follow the original setting to conduct these baseline methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 304, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 447 + ], + "type": "text", + "content": "Table 3 provides the result. Observe that for all the baselines, benign accuracy change is slight, meaning that the mitigation preserves the model utility on benign tasks. However, ASR degradation is considerable for all backdoored models. Note that LOTUS can still remain part of the attack effectiveness with " + }, + { + "bbox": [ + 304, + 304, + 547, + 447 + ], + "type": "inline_equation", + "content": "34.90\\% - 46.90\\%" + }, + { + "bbox": [ + 304, + 304, + 547, + 447 + ], + "type": "text", + "content": ", outperforming all other attacks. The result indicates that LOTUS is more resilient against baseline mitigation methods compared to the existing attacks. This can be attributed to the design that LOTUS learns the correlation between the partitions and triggers which is hard to unlearn. Other attacks only learn partial trigger patterns that tend to be mitigate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 455, + 536, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 536, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 536, + 468 + ], + "type": "text", + "content": "5.5. Evaluation on Different Poisoning Strategies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "We evaluate different poisoning strategies including simple data poisoning, adversarial poisoning, and LOTUS's trigger focusing. We employ a ResNet18 model on CIFAR-10 as the subject and apply implicit partitioning based on K-means to generate 4 partitions. The number of possible non-empty trigger combinations is " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "2^{4} - 1 = 15" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": ". In the following, we use a four-bit binary to represent each combination. For example, 0110 denotes " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_2" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_3" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " are stamped on inputs but not " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_1" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{T}_4" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": ". Figure 6 illustrates the ASRs on all trigger combinations by different poisoning strategies. Sub-figures from left to right present the results for simple poisoning, adversarial poisoning, and trigger focusing, respectively. In a subfigure, each column denotes input samples from a partition " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": ", and each row denotes a trigger combination. The value in each cell shows the ASR when a trigger combination (row) is stamped on the samples from a partition (column). Brighter the color, higher the ASR. The left sub-figure shows the ASR for simple data poisoning. Observe that all the ASRs are greater than " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "92.0\\%" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": " (with an average of " + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "inline_equation", + "content": "97.94\\%" + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "), showing the sub-partitioning is not learned by the model. The middle" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24804" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 72, + 177, + 199 + ], + "blocks": [ + { + "bbox": [ + 52, + 72, + 177, + 199 + ], + "lines": [ + { + "bbox": [ + 52, + 72, + 177, + 199 + ], + "spans": [ + { + "bbox": [ + 52, + 72, + 177, + 199 + ], + "type": "image", + "image_path": "edf326ed9bd3982acf628d6df140fd05097425e8d5ef725c845450673f6b9309.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 545, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 222 + ], + "type": "text", + "content": "Figure 5. Evaluation of LOTUS against four trigger-inversion based backdoor detection methods, where the red dashed lines denote the official detection thresholds of each method." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 72, + 312, + 198 + ], + "blocks": [ + { + "bbox": [ + 187, + 72, + 312, + 198 + ], + "lines": [ + { + "bbox": [ + 187, + 72, + 312, + 198 + ], + "spans": [ + { + "bbox": [ + 187, + 72, + 312, + 198 + ], + "type": "image", + "image_path": "715a70e921f96b1ad604b155f9ab52f220f24b2cd6a9723ebfc0eb90f0c6da07.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 322, + 76, + 452, + 198 + ], + "blocks": [ + { + "bbox": [ + 322, + 76, + 452, + 198 + ], + "lines": [ + { + "bbox": [ + 322, + 76, + 452, + 198 + ], + "spans": [ + { + "bbox": [ + 322, + 76, + 452, + 198 + ], + "type": "image", + "image_path": "8ed3dfdbf3b0b823e360c2ce1c00f8038c432263aace0b105872ce3d5e84bb8c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 463, + 76, + 541, + 198 + ], + "blocks": [ + { + "bbox": [ + 463, + 76, + 541, + 198 + ], + "lines": [ + { + "bbox": [ + 463, + 76, + 541, + 198 + ], + "spans": [ + { + "bbox": [ + 463, + 76, + 541, + 198 + ], + "type": "image", + "image_path": "06dc243d5eb6d96682c57e3931f34f7b59f16d95911f838fc984e10b96ed09af.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 47, + 289, + 323, + 390 + ], + "blocks": [ + { + "bbox": [ + 46, + 243, + 323, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 323, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 323, + 288 + ], + "type": "text", + "content": "Table 3. Evaluation of resilience against backdoor mitigation methods. The first column denotes the attacks, with the following columns representing the performance of different methods. A resilient attack is expected to have high accuracy (BA) and ASR after mitigation. The best results are in bold." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 289, + 323, + 390 + ], + "lines": [ + { + "bbox": [ + 47, + 289, + 323, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 323, + 390 + ], + "type": "table", + "html": "
AttacksNo DefenseFine-tuningFine-pruningNADANP
BAASRBAASRBAASRBAASRBAASR
BadNets92.02%100.00%89.31%1.74%91.70%0.53%87.81%0.80%89.15%0.32%
Dynamic91.81%100.00%88.87%2.91%91.39%22.03%89.11%2.90%88.25%12.81%
IA91.70%99.65%87.74%2.78%91.07%0.17%87.14%2.29%88.73%1.98%
WaNet91.22%98.57%89.56%1.37%90.22%1.07%89.74%1.40%89.07%0.54%
ISSBA91.67%99.96%87.73%2.72%91.12%14.27%87.97%2.83%85.64%10.01%
LIRA91.70%100.00%89.96%2.19%91.29%12.14%90.23%2.32%89.70%37.91%
DFST91.81%99.97%88.49%22.86%91.47%21.61%88.52%24.66%87.13%36.17%
LOTUS91.54%93.80%88.10%46.90%91.14%44.90%87.61%42.30%88.14%34.90%
", + "image_path": "75e88a0b365147e6ae4a8d72a68c032f8af9c2ff4978bde59c01943882d85a65.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 238, + 542, + 370 + ], + "blocks": [ + { + "bbox": [ + 332, + 238, + 542, + 370 + ], + "lines": [ + { + "bbox": [ + 332, + 238, + 542, + 370 + ], + "spans": [ + { + "bbox": [ + 332, + 238, + 542, + 370 + ], + "type": "image", + "image_path": "b3d6a9df366e37a4a3add35eef45dfd813c4caeca10e176b6a6228256f541999.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 374, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 329, + 374, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 329, + 374, + 545, + 396 + ], + "type": "text", + "content": "Figure 6. ASR on all trigger combinations by different poisoning strategies" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": "sub-figure is the results for adversarial poisoning. We can see around half of cells have small values, especially for single trigger combinations (the top four rows). For more complex trigger combinations, the ASRs are still high with the highest of " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "100.0\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": " (trigger combination 0111 on partition " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": "), indicating the insufficiency of adversarial poisoning. The right sub-figure is for our trigger focusing. Observe that except for stamping a trigger on the proper partition, the other cases all have a low ASR with an average of " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "3.04\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": ". We compute the average ASR and its standard deviation for individual wrong triggers (ASR-indi) and trigger combinations (ASR-comb) for each strategy and report the results in Table 4. Observe that all the ASRs are almost " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": " for simple poisoning. Adversarial poisoning reduces the ASR-indi to a low level while leaving ASR-comb high (" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "73.88\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": " on average). LOTUS's trigger focusing strategy has the lowest ASR-indi with an average of " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "14.15\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": " and ASR-comb " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": ". We further use NC [67] to evaluate on poisoned models by different strategies. The last column shows the anomaly index for different poisoned models. Observe that models poisoned by simple data poisoning and adversarial poisoning can be easily detected by NC (with anomaly index " + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "inline_equation", + "content": ">2" + }, + { + "bbox": [ + 46, + 418, + 289, + 706 + ], + "type": "text", + "content": "). Poisoned models by trigger focusing, on the other hand, are able to evade NC's detection, delineating the effectiveness" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 429, + 545, + 474 + ], + "blocks": [ + { + "bbox": [ + 330, + 418, + 521, + 428 + ], + "lines": [ + { + "bbox": [ + 330, + 418, + 521, + 428 + ], + "spans": [ + { + "bbox": [ + 330, + 418, + 521, + 428 + ], + "type": "text", + "content": "Table 4. Evaluation on different poisoning strategies" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 429, + 545, + 474 + ], + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 474 + ], + "type": "table", + "html": "
StrategyBAASRASR-indiASR-combNC index
Simple94.79%98.80%97.86% ± 1.84%97.88% ± 1.81%5.338
Adv.94.47%94.20%18.95% ± 10.22%73.88% ± 31.87%2.161
Focus94.71%91.40%14.15% ± 7.46%0.02% ± 0.09%1.156
", + "image_path": "f62f2db9220d7ab0e02d1b579f8723d9da72c048ce54d2e7cbe54a56ef92d25c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 475, + 511, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 511, + 487 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 511, + 487 + ], + "type": "text", + "content": "of trigger focusing strategy to achieve evasiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 496, + 378, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 378, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 378, + 509 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 516, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 588 + ], + "type": "text", + "content": "We propose a novel backdoor attack that leverages subpartitioning to restrict the attack scope. A special training method is designed to limit triggers to only their corresponding partitions. Our evaluation shows that the attack is highly effective, achieving high attack success rates. Besides, it is evasive and resilient against state-of-the-art defenses." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 409, + 611 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "We thank the anonymous reviewers for their constructive comments. We are grateful to the Center for AI Safety for providing computational resources. This research was supported, in part by IARPA TrojAI W911NF-19-S0012, NSF 1901242 and 1910300, ONR N000141712045, N000141410468 and N000141712947. Any opinions, findings, and conclusions in this paper are those of the authors only and do not necessarily reflect the views of our sponsors." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24805" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Eugene Bagdasaryan and Vitaly Shmatikov. Blind backdoors in deep learning models. arXiv preprint arXiv:2005.03823, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "text", + "content": "[2] Eitan Borgnia, Valeriia Cherepanova, Liam Fowl, Amin Ghiasi, Jonas Geiping, Micah Goldblum, Tom Goldstein, and Arjun Gupta. Strong data augmentation sanitizes poisoning and backdoor attacks without an accuracy tradeoff. arXiv preprint arXiv:2011.09527, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[3] Alvin Chan and Yew-Soon Ong. Poison as a cure: Detecting & neutralizing variable-sized backdoor attacks in deep neural networks. arXiv preprint arXiv:1911.08040, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 270 + ], + "type": "text", + "content": "[4] Bryant Chen, Wilka Carvalho, Nathalie Baracaldo, Heiko Ludwig, Benjamin Edwards, Taesung Lee, Ian Molloy, and Biplav Srivastava. Detecting backdoor attacks on deep neural networks by activation clustering. arXiv preprint arXiv:1811.03728, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 315 + ], + "type": "text", + "content": "[5] Xinyun Chen, Chang Liu, Bo Li, Kimberly Lu, and Dawn Song. Targeted backdoor attacks on deep learning systems using data poisoning. arXiv preprint arXiv:1712.05526, 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 361 + ], + "type": "text", + "content": "[6] Siyuan Cheng, Yingqi Liu, Shiqing Ma, and Xiangyu Zhang. Deep feature space trojan attack of neural networks by controlled detoxification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1148-1156, 2021. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 362, + 288, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 417 + ], + "type": "text", + "content": "[7] Siyuan Cheng, Guanhong Tao, Yingqi Liu, Shengwei An, Xiangzhe Xu, Shiwei Feng, Guangyu Shen, Kaiyuan Zhang, Qiuling Xu, Shiqing Ma, et al. Beagle: Forensics of deep learning backdoor attack for better defense. arXiv preprint arXiv:2301.06241, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 288, + 462 + ], + "type": "text", + "content": "[8] Edward Chou, Florian Tramer, and Giancarlo Pellegrino. Sentinel: Detecting localized universal attack against deep learning systems. Proceeding of the 41th IEEE Symposium on Security & Privacy Workshops (SPW), 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 507 + ], + "type": "text", + "content": "[9] Ryan Cohn and Elizabeth Holm. Unsupervised machine learning via transfer learning and k-means clustering to classify materials image data. Integrating Materials and Manufacturing Innovation, 10(2):231-244, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 553 + ], + "type": "text", + "content": "[10] Khoa Doan, Yingjie Lao, Weijie Zhao, and Ping Li. Lira: Learnable, imperceptible and robust backdoor attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11966-11976, 2021. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 288, + 597 + ], + "type": "text", + "content": "[11] Min Du, Ruoxi Jia, and Dawn Song. Robust anomaly detection and backdoor attack detection via differential privacy. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 643 + ], + "type": "text", + "content": "[12] Logan Engstrom, Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Brandon Tran, and Aleksander Madry. Learning perceptually-aligned representations via adversarial robustness. arXiv preprint arXiv:1906.00945, 2(3):5, 2019. 6, 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 644, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 287, + 667 + ], + "type": "text", + "content": "[13] Wang et al. An invisible black-box backdoor attack through frequency domain. In ECCV, 2022. 17" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 690 + ], + "type": "text", + "content": "[14] Xia et al. Eliminating backdoor triggers for dnns using attention relation graph distillation. In *IJCAI*, 2022. 17" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Zeng et al. Adversarial unlearning of backdoors via implicit hypergradient. In ICLR, 2022. 17" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[16] Shiwei Feng, Guanhong Tao, Siyuan Cheng, Guangyu Shen, Xiangzhe Xu, Yingqi Liu, Kaiyuan Zhang, Shiqing Ma, and Xiangyu Zhang. Detecting backdoors in pre-trained encoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16352-16362, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[17] Hao Fu, Akshaj Kumar Veldanda, Prashanth Krishnamurthy, Siddharth Garg, and Farshad Khorrami. Detecting backdoors in neural networks using novel feature-based anomaly detection. arXiv preprint arXiv:2011.02526, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 173, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 547, + 228 + ], + "type": "text", + "content": "[18] Yansong Gao, Change Xu, Derui Wang, Shiping Chen, Damith C Ranasinghe, and Surya Nepal. Strip: A defence against trojan attacks on deep neural networks. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 113-125, 2019. 2, 16" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 259 + ], + "type": "text", + "content": "[19] Tianyu Gu, Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Badnets: Evaluating backdooring attacks on deep neural networks. IEEE Access, 2019. 1, 2, 7, 21" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 304 + ], + "type": "text", + "content": "[20] Wenbo Guo, Lun Wang, Yan Xu, Xinyu Xing, Min Du, and Dawn Song. Towards inspecting and eliminating trojan backdoors in deep neural networks. In 20th IEEE International Conference on Data Mining, 2020. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 547, + 348 + ], + "type": "text", + "content": "[21] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 6, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 547, + 403 + ], + "type": "text", + "content": "[22] Jonathan Hayase, Weihao Kong, Raghav Somani, and Swoong Oh. Spectre: defending against backdoor attacks using robust statistics. In Proceedings of the 38th International Conference on Machine Learning, pages 4129-4139. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. European conference on computer vision, pages 630-645, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 547, + 481 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 547, + 525 + ], + "type": "text", + "content": "[25] Gao Huang, Zhuang Liu, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 547, + 559 + ], + "type": "text", + "content": "[26] Xijie Huang, Moustafa Alzantot, and Mani Srivastava. Neuronspect: Detecting backdoors in neural networks via output explanations. arXiv preprint arXiv:1911.07399, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "text", + "content": "[27] Matthew Jagielski, Giorgio Severi, Niklas Pousette Harger, and Alina Oprea. Subpopulation data poisoning attacks. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 547, + 657 + ], + "type": "text", + "content": "[28] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 713 + ], + "type": "text", + "content": "[29] Soheil Kolouri, Aniruddha Saha, Hamed Pirsiavash, and Heiko Hoffmann. Universal litmus patterns: Revealing backdoor attacks in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 301-310, 2020. 2, 7, 16" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24806" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[31] Yiming Li, Tongqing Zhai, Baoyuan Wu, Yong Jiang, Zhifeng Li, and Shutao Xia. Rethinking the trigger of backdoor attack. arXiv preprint arXiv:2004.04692, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[32] Yige Li, Nodens Koren, Lingjuan Lyu, Xixiang Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing backdoor triggers from deep neural networks. In International Conference on Learning Representations, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 226 + ], + "type": "text", + "content": "[33] Yuezun Li, Yiming Li, Baoyuan Wu, Longkang Li, Ran He, and Siwei Lyu. Invisible backdoor attack with sample-specific triggers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16463-16472, 2021. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[34] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Anti-backdoor learning: Training clean models on poisoned data. In Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 326 + ], + "type": "text", + "content": "[35] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International Symposium on Research in Attacks, Intrusions, and Defenses, pages 273-294. Springer, 2018. 1, 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "text", + "content": "[36] Yuntao Liu, Yang Xie, and Ankur Srivastava. Neural trojans. In 2017 IEEE International Conference on Computer Design (ICCD), pages 45-48. IEEE, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 392 + ], + "type": "text", + "content": "[37] Yingqi Liu, Shiqing Ma, Yousra Aafer, Wen-Chuan Lee, Juan Zhai, Weihang Wang, and Xiangyu Zhang. Trojaning attack on neural networks. In NDSS, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 459 + ], + "type": "text", + "content": "[38] Yingqi Liu, Wen-Chuan Lee, Guanhong Tao, Shiqing Ma, Yousra Aafer, and Xiangyu Zhang. Abs: Scanning neural networks for back-doors by artificial brain stimulation. In Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, pages 1265–1282, 2019. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 503 + ], + "type": "text", + "content": "[39] Yunfei Liu, Xingjun Ma, James Bailey, and Feng Lu. Reflection backdoor: A natural backdoor attack on deep neural networks. In European Conference on Computer Vision, pages 182-199. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "text", + "content": "[40] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE International Conference on Computer Vision, pages 3730-3738, 2015. 6, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 602 + ], + "type": "text", + "content": "[41] Peizhuo Lv, Chang Yue, Ruigang Liang, Yunfei Yang, Shengzhi Zhang, Hualong Ma, and Kai Chen. A data-free backdoor injection approach in neural networks. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2671-2688, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 658 + ], + "type": "text", + "content": "[42] Shiqing Ma, Yingqi Liu, Guanhong Tao, Wen-Chuan Lee, and Xiangyu Zhang. Nic: Detecting adversarial samples with neural network invariant checking. In Proceedings of the 26th Network and Distributed System Security Symposium (NDSS 2019), 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 713 + ], + "type": "text", + "content": "[43] Michael McCoyd, Won Park, Steven Chen, Neil Shah, Ryan Roggenkemper, Minjune Hwang, Jason Xinyu Liu, and David Wagner. Minority reports defense: Defending against adversarial patches. In Applied Cryptography and Network Security Workshops: ACNS 2020 Satellite Workshops, AIBlock," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 106 + ], + "type": "text", + "content": "AIHWS, AIoTS, Cloud S&P, SCI, SecMT, and SiMLA, Rome, Italy, October 19-22, 2020, Proceedings, pages 564-582. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 138 + ], + "type": "text", + "content": "[44] Geoffrey J McLachlan and Kaye E Basford. Mixture models: Inference and applications to clustering. M. Dekker New York, 1988. 6, 14" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 172 + ], + "type": "text", + "content": "[45] Tuan Anh Nguyen and Anh Tran. Input-aware dynamic backdoor attack. Advances in Neural Information Processing Systems, 33, 2020. 1, 2, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 546, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 205 + ], + "type": "text", + "content": "[46] Tuan Anh Nguyen and Anh Tuan Tran. Wanet-imperceptible warping-based backdoor attack. In International Conference on Learning Representations, 2020. 1, 2, 4, 7, 13, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 271 + ], + "type": "text", + "content": "[47] Jun Ogasawara, Satoru Ikenoue, Hiroko Yamamoto, Motoshige Sato, Yoshifumi Kasuga, Yasue Mitsukura, Yuji Ikegaya, Masato Yasui, Mamoru Tanaka, and Daigo Ochiai. Deep neural network-based classification of cardiotocograms outperformed conventional algorithms. Scientific reports, 11 (1):13367, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 547, + 316 + ], + "type": "text", + "content": "[48] Xiangyu Qi, Tinghao Xie, Yiming Li, Saeed Mahloujifar, and Prateek Mittal. Revisiting the assumption of latent separability for backdoor defenses. In The eleventh international conference on learning representations, 2022. 17" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 547, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 360 + ], + "type": "text", + "content": "[49] Ximing Qiao, Yukun Yang, and Hai Li. Defending neural backdoors via generative distribution modeling. In Advances in Neural Information Processing Systems, pages 14004-14013, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 361, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 547, + 403 + ], + "type": "text", + "content": "[50] Aniruddha Saha, Akshayvarun Subramanya, and Hamed Pirsiavash. Hidden trigger backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11957-11965, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 404, + 547, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 547, + 448 + ], + "type": "text", + "content": "[51] Ahmed Salem, Rui Wen, Michael Backes, Shiqing Ma, and Yang Zhang. Dynamic backdoor attacks against machine learning models. arXiv preprint arXiv:2003.03675, 2020. 1, 2, 7, 21" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 546, + 491 + ], + "type": "text", + "content": "[52] Shibani Santurkar, Dimitris Tsipras, Brandon Tran, Andrew Ilyas, Logan Engstrom, and Aleksander Madry. Image synthesis with a single (robust) classifier. arXiv preprint arXiv:1906.09453, 2019. 6, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 492, + 547, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 492, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 492, + 547, + 548 + ], + "type": "text", + "content": "[53] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Gradcam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 21" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "type": "text", + "content": "[54] Ali Shafahi, W Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Tudor Dumitras, and Tom Goldstein. Poison frogs! targeted clean-label poisoning attacks on neural networks. In Advances in Neural Information Processing Systems, pages 6103-6113, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 547, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 658 + ], + "type": "text", + "content": "[55] Guangyu Shen, Yingqi Liu, Guanhong Tao, Shengwei An, Qiuling Xu, Siyuan Cheng, Shiqing Ma, and Xiangyu Zhang. Backdoor scanning for deep neural networks through k-arm optimization. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 547, + 713 + ], + "type": "text", + "content": "[56] Guangyu Shen, Yingqi Liu, Guanhong Tao, Qiuling Xu, Zhuo Zhang, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Constrained optimization with dynamic bound-scaling for effective nlp backdoor defense. In International Conference on Machine Learning, pages 19879-19892. PMLR, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24807" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[57] Guangyu Shen, Siyuan Cheng, Guanhong Tao, Kaiyuan Zhang, Yingqi Liu, Shengwei An, Shiqing Ma, and Xiangyu Zhang. Django: Detecting trojans in object detection models via gaussian focus calibration. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 161 + ], + "type": "text", + "content": "[58] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[59] Di Tang, XiaoFeng Wang, Haixu Tang, and Kehuan Zhang. Demon in the variant: Statistical analysis of dnns for robust backdoor contamination detection. In 30th USENIX Security Symposium (USENIX Security 21), 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 259 + ], + "type": "text", + "content": "[60] Guanhong Tao, Yingqi Liu, Guangyu Shen, Qiuling Xu, Shengwei An, Zhuo Zhang, and Xiangyu Zhang. Model orthogonalization: Class distance hardening in neural networks for better security. In 2022 IEEE Symposium on Security and Privacy (SP). IEEE, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 326 + ], + "type": "text", + "content": "[61] Guanhong Tao, Guangyu Shen, Yingqi Liu, Shengwei An, Qiuling Xu, Shiqing Ma, Pan Li, and Xiangyu Zhang. Better trigger inversion optimization in backdoor scanning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13368-13378, 2022. 1, 2, 7, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 369 + ], + "type": "text", + "content": "[62] Brandon Tran, Jerry Li, and Aleksander Madry. Spectral signatures in backdoor attacks. In Advances in Neural Information Processing Systems, pages 8000–8010, 2018. 2, 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "text", + "content": "[63] Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. arXiv preprint arXiv:1805.12152, 2018. 6, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 436 + ], + "type": "text", + "content": "[64] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Clean-label backdoor attacks. 2018. 1, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 469 + ], + "type": "text", + "content": "[65] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[66] Akshaj Kumar Veldanda, Kang Liu, Benjamin Tan, Prashanth Krishnamurthy, Farshad Khorrami, Ramesh Karri, Brendan Dolan-Gavitt, and Siddharth Garg. Nnoculation: broad spectrum and targeted treatment of backdoored dnns. arXiv preprint arXiv:2002.08313, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 581 + ], + "type": "text", + "content": "[67] Bolun Wang, Yuanshun Yao, Shawn Shan, Huiying Li, Bimal Viswanath, Haitao Zheng, and Ben Y Zhao. Neural cleanse: Identifying and mitigating backdoor attacks in neural networks. In 2019 IEEE Symposium on Security and Privacy (SP), pages 707-723. IEEE, 2019. 1, 2, 3, 7, 8, 17" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 624 + ], + "type": "text", + "content": "[68] Ren Wang, Gaoyuan Zhang, Sijia Liu, Pin-Yu Chen, Jinjun Xiong, and Meng Wang. Practical detection of trojan neural networks: Data-limited and data-free cases. In European Conference on Computer Vision, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[69] Zhenting Wang, Hailun Ding, Juan Zhai, and Shiqing Ma. Training with more confidence: Mitigating injected and natural backdoors during training. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[70] Zhenting Wang, Kai Mei, Hailun Ding, Juan Zhai, and Shiqing Ma. Rethinking the reverse-engineering of Trojan triggers. In Advances in Neural Information Processing Systems, 2022. 2, 3, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[71] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15074-15084, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 183 + ], + "type": "text", + "content": "[72] Dongxian Wu and Yisen Wang. Adversarial neuron pruning purifies backdoored deep models. Advances in Neural Information Processing Systems, 34:16913-16925, 2021. 1, 7, 17" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "text", + "content": "[73] Junru Wu, Yue Wang, Zhenyu Wu, Zhangyang Wang, Ashok Veeraraghavan, and Yingyan Lin. Deep k-means: Re-training and parameter sharing with harder cluster assignments for compressing deep convolutions. In International Conference on Machine Learning, pages 5363-5372. PMLR, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "text", + "content": "[74] Chong Xiang, Arjun Nitin Bhagoji, Vikash Sehwag, and Prateek Mittal. Patchguard: A provably robust defense against adversarial patches via small receptive fields and masking. In USENIX Security Symposium, pages 2237-2254, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 286, + 547, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 341 + ], + "type": "text", + "content": "[75] Chong Xiang, Saeed Mahloujifar, and Prateek Mittal. {PatchCleanser}: Certifiably robust defense against adversarial patches for any image classifier. In 31st USENIX Security Symposium (USENIX Security 22), pages 2065-2082, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 546, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 546, + 385 + ], + "type": "text", + "content": "[76] Zhen Xiang, David Miller, and George Kesidis. Post-training detection of backdoor attacks for two-class and multi-attack scenarios. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 546, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 546, + 420 + ], + "type": "text", + "content": "[77] Zhen Xiang, Zidi Xiong, and Bo Li. Umd: Unsupervised model detection for x2x backdoor attacks. In International Conference on Machine Learning, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 546, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 546, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 546, + 465 + ], + "type": "text", + "content": "[78] Qiuling Xu, Guanhong Tao, Siyuan Cheng, and Xiangyu Zhang. Towards feature space adversarial attack by style perturbation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10523-10531, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 468, + 547, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 547, + 531 + ], + "type": "text", + "content": "[79] Qiuling Xu, Guanhong Tao, Jean Honorio, Yingqi Liu, Shengwei An, Guangyu Shen, Siyuan Cheng, and Xiangyu Zhang. Medic: Remove model backdoors via importance driven cloning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20485-20494, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 546, + 567 + ], + "type": "text", + "content": "[80] Xiaojun Xu, Qi Wang, Huichen Li, Nikita Borisov, Carl A Gunter, and Bo Li. Detecting ai trojans using meta neural analysis. arXiv preprint arXiv:1910.03137, 2019. 2, 7, 16" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 547, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 601 + ], + "type": "text", + "content": "[81] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In Proceedings of the British Machine Vision Conference, 2016. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 602, + 546, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 546, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 546, + 646 + ], + "type": "text", + "content": "[82] Yi Zeng, Han Qiu, Shangwei Guo, Tianwei Zhang, Meikang Qiu, and Bhavani Thuraisingham. Deepsweep: An evaluation framework for mitigating dnn backdoor attacks using data augmentation. arXiv preprint arXiv:2012.07006, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 712 + ], + "type": "text", + "content": "[83] Kaiyuan Zhang, Guanhong Tao, Qiuling Xu, Siyuan Cheng, Shengwei An, Yingqi Liu, Shiwei Feng, Guangyu Shen, PinYu Chen, Shiqing Ma, and Xiangyu Zhang. FLIP: A provable defense framework for backdoor mitigation in federated learning. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24808" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6, 14, 21" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[85] Shihao Zhao, Xingjun Ma, Xiang Zheng, James Bailey, Jingjing Chen, and Yu-Gang Jiang. Clean-label backdoor attacks on video recognition models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14443-14452, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[86] Chen Zhu, W Ronny Huang, Hengduo Li, Gavin Taylor, Christoph Studer, and Tom Goldstein. Transferable clean-label poisoning attacks on deep neural nets. In International Conference on Machine Learning, pages 7614-7623. PMLR, 2019. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24809" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_content_list.json b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..08bc7fac93fae2cac8f25706b58d1298e34a87cd --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_content_list.json @@ -0,0 +1,1512 @@ +[ + { + "type": "text", + "text": "$\\mathbf{LP} + +: A$ Surprisingly Strong Linear Probe for Few-Shot CLIP", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunshi Huang $^{1*}$ Fereshteh Shakeri $^{1*}$ \nJose Dolz $^{1}$ Malik Boudiaf $^{1}$ Houda Bahig $^{2}$ Ismail Ben Ayed $^{1}$ $^{1}$ ÉTS Montréal, $^{2}$ Université de Montréal", + "bbox": [ + 194, + 178, + 772, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 284, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin $LP++$ , step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, $LP++$ operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git.", + "bbox": [ + 76, + 316, + 473, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 796, + 209, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, there has been a growing popularity in multimodal learning methods, which process and merge information from diverse modalities. In particular, large-", + "bbox": [ + 76, + 821, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e659f8dd36f94f8c956f7f15aad2f81b3215032cc027fe1e3fd06c7d2b3793e.jpg", + "image_caption": [ + "Figure 1. Comparison of $\\mathrm{LP}++$ with state-of-the-art few-shot CLIP methods in the 1-shot setting across 11 datasets. We compute the mean accuracy and standard deviation using 10 random tasks for each dataset. The error bars indicate the average standard deviation over all 11 datasets. The x-axis represents the run time for one task, averaged over the 11 datasets. Tip-Adapter-F and Tip-Adapter-F* are two re-implementations of Tip-Adapter-F [28], with fixed and grid-search hyper-parameters, respectively (implementation details provided in Sec. 3.2)." + ], + "image_footnote": [], + "bbox": [ + 526, + 287, + 867, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "scale vision-language models (VLMs), such as CLIP [23] and ALIGN [15], have attracted wide attention and made substantial progress in computer vision, showing promising generalization capabilities in various downstream tasks. Unlike conventional task-specific models that are trained with a predetermined set of labels, these so-called foundation models learn to align images with text, in an open-vocabulary fashion. They train, via contrastive learning, vision and text embeddings jointly using a large-scale amount of image-text pairs collected over the internet, thereby leveraging the rich semantic knowledge inherent to language (e.g., concept hierarchies). For a given downstream image classification task, vision-language embeddings enable zero-shot predictions, without re-training, using textual descriptions of the classes (a.k.a prompts). For instance, for a given class $k$ , a textual description of the class,", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Correspondence to yunshi.huang@etsmtl.ca, fereshteh.shakeri.1@etsmtl.net", + "bbox": [ + 76, + 875, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23773", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which we denote $z_{k}$ , could be \"a photo of a $[class_{k}]$ \", where $[class_{k}]$ is the class name. Thus, the zero-shot class prediction for a query image $x$ is obtained from the cosine similarity between the $l_{2}$ -normalized vision-encoded embeddings, $f = \\theta_{v}(x)$ , and the text-encoded ones, $t_{k} = \\theta_{t}(z_{k})$ : $\\hat{k} = \\arg \\max_{k} f^{t} t_{k}$ , where $t$ denotes the transpose1.", + "bbox": [ + 75, + 90, + 467, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the observation that the choice of input prompts $z_{k}$ may affect the zero-shot predictions, and following on from the strong recent emergence of prompt learning research in the NLP community [14, 16, 24], the popular work in [29] pioneered context optimization (CoOp) for vision-language models. CoOp models input text $z_{k}$ as learnable continuous vectors, e.g., in the form $z_{k} = (z_{k}^{1},\\dots,z_{k}^{M},[\\mathrm{class}_{k}])$ , where $(z_k^l)_{1\\leq l\\leq M}$ are learnable text tokens, $[\\mathrm{class}_k]$ is a fixed token corresponding to the word embedding vector of the name of the $k^{th}$ class, and $M$ is a hyper-parameter. These learnable vectors are finetuned as task-specific prompts using few-shot training examples and a standard supervised classification loss. More specifically, in this few-shot setting, we assume access to a set consisting of a few labeled samples for each target class, often referred to as the support set. Let $f_{i} = \\theta_{v}(\\pmb{x}_{i})$ denote the vision embedding of support image $i$ , and $y_{ik}$ its one-hot encoded label, i.e., $y_{ik} = 1$ if image $x_{i}$ belongs to class $k$ and 0 otherwise. Expressing the text embeddings as $t_k = \\theta_t(z_k^1,\\dots,z_k^M,[\\mathrm{class}_k])$ , CoOp fine-tunes text tokens $(z_k^l)_{1\\leq l\\leq M}$ by minimizing the cross-entropy (CE) loss, with $N$ labeled support samples and $K$ classes:", + "bbox": [ + 75, + 183, + 468, + 515 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 523, + 468, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the softmax predictions $p_{ik}$ and the logits (class scores) $l_{ik}$ are given by:", + "bbox": [ + 76, + 575, + 468, + 606 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {i k} = \\frac {\\exp (l _ {i k})}{\\sum_ {j = 1} ^ {K} \\exp (l _ {i j})}; \\quad l _ {i k} = \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 614, + 395, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although recent, the pioneering idea of CoOp has triggered a quite abundant literature on prompt learning for few-shot vision-language models, with numerous, more convoluted extensions, e.g. [4, 27, 30], to list a few. For instance, PLOT [4] followed up by learning multiple prompts, to describe the characteristics of each class, via minimizing an optimal-transport distance. KgCoOp [27] improves CoOp's performance when dealing with unseen classes, via minimizing the discrepancy between the text embeddings generated by the learned prompts and hand-crafted ones. While CoOp directly updates the context vectors using the CE loss, ProGrad [30] aligns the few-shot down", + "bbox": [ + 75, + 660, + 468, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stream knowledge with the large-scale general knowledge, thus mitigating the overfitting of the few-shot samples.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt learning methods have brought significant improvements over zero-shot classification, but they come at the price of heavy computational and memory load, as they require gradient back-propagation through the entire text encoder. Furthermore, they assume knowledge of the text encoder. These aspects may impede their deployment in low-resource and black-box, privacy-preserving scenarios, which are of wide interest in practice. Indeed, in NLP, there is currently an emerging literature on fast few-shot adaptation of black-box models [6], strongly motivated by the fact that large-scale foundation models (e.g., the GPT family, Anthropic's Claude or Google's PaLM) are only available through APIs and their pre-trained weights are not shared. Finally, by evaluating prompt learning methods over larger numbers of sampled support sets in our experiments (Fig. 1), we observed that they exhibit large variation in performances. This could be explained by the fact that, through the text encoder, they learn prompts that are \"too specialized\" for a given image support set.", + "bbox": [ + 496, + 121, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While prompt learning alters the textual inputs, another category of approaches, referred to as adapters, focused on transforming the pre-training features of the visual or language encoders, e.g., [9, 28]. These adapters are non-linear transformations, for instance, in the form of multi-layer modules, added to the encoder's bottleneck. They learn additional transformations, yielding logits of the form:", + "bbox": [ + 496, + 407, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i k} = \\boldsymbol {\\theta} _ {a} \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {t} _ {k}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 525, + 890, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The adapter's learnable parameters, $\\theta_{a}$ , are fine-tuned over a few-shot task by optimizing the cross-entropy loss, similarly to (1) but with logits $l_{ik}$ expressed as functions of $\\theta_{a}$ . For instance, the popular CLIP-Adapter [9] integrated a multi-layered perceptron to modify the features, along with residual connections, which enable blending with the original pre-trained features. Tip-Adapter [28] added a nonlinear, quadratic-complexity module, which evaluates the pairwise similarities between the features of the support sets, and blends the ensuing class scores with the original textual features. This category of approaches mitigates some of the limitations of prompt-learning methods as they result in few-shot adaptation that has significantly lower computation and memory loads. However, as shown in Fig. 1 and in our experiments, their performances seem to depend strongly on some key hyper-parameters that have to be adjusted carefully on each downstream task, e.g. those that control the blending between the vision and language features. Therefore, to perform competitively, they incur an additional computation overhead to the adaptation phase, due to intensive (e.g., grid) search of the hyper-parameters over task-dedicated validation sets.", + "bbox": [ + 496, + 551, + 890, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the above-mentioned, strongly emergent literature on", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1For $l_{2}$ -normalized feature embeddings, the dot product corresponds to the cosine similarity.", + "bbox": [ + 75, + 849, + 467, + 873 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2The number of labeled support samples per class, $S = \\frac{N}{K}$ , is small, typically in $\\{1,2,4,\\dots,16\\}$", + "bbox": [ + 76, + 873, + 467, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23774", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "few-shot CLIP adaptation, linear probe (LP) [23] has been often reported as a very weak baseline. For instance, in the 1-shot setting, it scores near $20\\%$ lower than the zero-shot predictions averaged over 11 benchmarks (Table 1). Initially evaluated in [23], LP is a linear classifier on the vision-encoded features. Specifically, it optimizes the CE loss (1) w.r.t the last-layer weights of the vision encoder (i.e., the class prototypes), which we will denote $(\\pmb{w}_k)_{1\\leq k\\leq K}$ in the rest of the paper, with the logits given by: $l_{ik} = \\pmb{f}_i^t\\pmb{w}_k$ . A clear deficiency in this standard LP baseline is that it omits completely the language knowledge of CLIP, i.e., $(\\pmb{t}_k)_{1\\leq k\\leq K}$ .", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline. Specifically, we extend the logits in the CE loss in (1), so that they become learnable functions of the text embedding:", + "bbox": [ + 75, + 273, + 470, + 347 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {i k} = \\boldsymbol {f} _ {i} ^ {t} (\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k})\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 354, + 346, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $(\\alpha_{k})_{1\\leq k\\leq K}$ trainable class-wise parameters blending image and text knowledge. As our objective function depends on two types of variables, i.e., the visual class prototypes $(\\pmb{w}_k)_{1\\leq k\\leq K}$ and blending parameters $(\\alpha_{k})_{1\\leq k\\leq K}$ , we propose a computationally efficient Block Majorize-Minimize (BMM) procedure. In our full-batch MM optimizer, which we coin $\\mathrm{LP}++$ , step sizes are implicit in the definition of the majorizing functions, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. Moreover, we examine the mathematical properties of our objective, i.e., (i) Lipschitz gradient continuity and (ii) decomposition into convex functions having closed-form optima. This enabled us to build majorizing functions yielding data-driven learning rates, and to derive approximations of the objective-function minima, which yield data-informed initializations of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yield, surprisingly, highly competitive few-shot CLIP performances (Fig. 1). Furthermore, $\\mathrm{LP}++$ operates in black-box, relaxes intensive validation searches for the optimization of hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP methods (Table 4). For instance, for 16-shot ImageNet adaptation, it takes seconds on a single NVIDIA RTX A600 GPU.", + "bbox": [ + 75, + 378, + 472, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Formulation of $\\mathbf{LP} + +$", + "text_level": 1, + "bbox": [ + 76, + 768, + 281, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the notations introduced in the previous section, we propose to minimize the following CE objective function w.r.t visual class prototypes $\\mathbf{w} = (\\pmb{w}_k)_{1\\leq k\\leq K}$ and class-wise blending parameters $\\alpha = (\\alpha_{k})_{1\\leq k\\leq K}$ :", + "bbox": [ + 75, + 792, + 468, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {w}, \\boldsymbol {\\alpha}) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 862, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b89ac625c0fbad8bc8a8033039281ad74331c693ebe7010aba59edd154d9914f.jpg", + "image_caption": [ + "Figure 2. Visualization of $\\mathbf{LP} + +$" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where softmax probability outputs $p_{ik}$ are now given by:", + "bbox": [ + 500, + 268, + 874, + 284 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) = \\frac {\\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k}\\right)\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {j} + \\alpha_ {j} \\boldsymbol {t} _ {j}\\right)\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 291, + 890, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and $t_k$ are the fixed pre-trained embeddings of the text templates used in zero-shot CLIP [23]. Clearly, the objective function defined by (3) and (4) could be viewed as a generalization of the vision-encoder CE loss used in the standard LP [23]. Indeed, the latter corresponds to setting $\\alpha_k = 0 \\forall k$ , i.e., no text knowledge. As we will see in our experimental ablation over different objective functions (Table 2), introducing the text knowledge $(\\alpha_k > 0 \\forall k)$ has a substantial effect on performances. Also, making $\\alpha_k$ learnable (rather than fixed) leads to a further significant impact. Indeed, we hypothesize that the optimal blending of the text and visual knowledge is task dependent, which motivates learning it from the context of the support set.", + "bbox": [ + 496, + 335, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Block coordinate Majorize-Minimize descent", + "text_level": 1, + "bbox": [ + 498, + 541, + 880, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Majorize-Minimize (MM) [18] is a very general optimization principle, which includes different classes of standard optimizers such as gradient descent, concave-convex procedures and expectation-maximization. Let $\\mathbf{v} = (\\mathbf{w},\\alpha)\\in$ $\\mathbb{R}^{K(D + 1)}$ denote the overall vector of variables in our case, with $D$ being the dimension of the feature embeddings. At each iteration, the MM procedure updates the variable as the minimum of a majorizing function, i.e., an upper bound on the original objective, which is tight at the current iteration $j\\colon L(\\mathbf{v})\\leq M(\\mathbf{v},\\mathbf{v}^j)$ and $L(\\mathbf{v}^j) = M(\\mathbf{v}^j,\\mathbf{v}^j)$ . Thus, update step $\\mathbf{v}^{j + 1} = \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)$ guarantees that the original objective does not increase at each iteration: $L(\\mathbf{v}^{j + 1})\\leq M(\\mathbf{v}^{j + 1},\\mathbf{v}^j)\\leq M(\\mathbf{v}^j,\\mathbf{v}^j) = L(\\mathbf{v}^j)$ . Therefore, in MM algorithms, step sizes are implicit in the definition of the majorizing function, unlike standard gradient-descent practices, in which the step sizes (a.k.a learning rates) are intensively searched over validation sets, via running the optimizer several times.", + "bbox": [ + 496, + 564, + 890, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we exploit the Lipschitz-gradient continuity of our convex objective in (3), i.e., bounds on the", + "bbox": [ + 498, + 837, + 890, + 867 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "This assumes, of course, that minimizing $M(\\mathbf{v}, \\mathbf{v}^j)$ over $\\mathbf{v}$ could be solved to global optimality and is easier than the original problem.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "maximum eigen values of the Hessian matrices (Prop. 1), thereby building majorizing functions with data-driven, task-specific step sizes. This removes the need for validation searches for the optimization hyper-parameters, reducing the computational load for fine-tuning (Table 2), while yielding performances on par with the best learning rates found with validation (Fig. 4). Also, interestingly, the Lipschitz-based learning rates computed from our derivation in Prop. 1 are orders-of-magnitude larger than those typically used in deep learning, yielding steeper decreases towards the minimum. Before giving proper majorizing functions for our convex, gradient-Lipschitz function in (3), let us first point to the following results, well-known in convex optimization [3]. While these results are text-book knowledge in optimization, they enable to connect the general MM principle to gradient descent, motivating the data-driven, task-specific step sizes we derive in Prop. 1 and the block-coordinate MM optimizer we propose in Alg. 1.", + "bbox": [ + 75, + 90, + 472, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 2.1. ([3, p. 268]) Assume $L(\\mathbf{v})$ is a twice-differentiable function, which has a Lipschitz continuous gradient, i.e., there exists a strictly positive Lipschitz constant $\\gamma$ such that $\\nabla^2 L(\\mathbf{v}) \\preceq \\gamma \\mathbf{I}$ , with $\\mathbf{I}$ the identity matrix. Then, the following quadratic bound is a majorizing function for $L$ at iteration $j$ :", + "bbox": [ + 75, + 371, + 470, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM (\\mathbf {v}, \\mathbf {v} ^ {j}) = L (\\mathbf {v} ^ {j}) + \\nabla L (\\mathbf {v} ^ {j}) ^ {t} (\\mathbf {v} - \\mathbf {v} ^ {j}) + \\frac {\\gamma}{2} \\| \\mathbf {v} - \\mathbf {v} ^ {j} \\| ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 472, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, a specific gradient step, with learning rate $\\frac{1}{\\gamma}$ minimizes $M$ , i.e., $\\mathbf{v}^{j + 1} = \\mathbf{v}^j -\\frac{1}{\\gamma}\\nabla L(\\mathbf{v}^j)$ = $\\arg \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)$ , and guarantees that objective $L$ decreases by at least $\\frac{1}{2\\gamma}\\| \\nabla L(\\mathbf{v})\\| ^2$ :", + "bbox": [ + 76, + 507, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\mathbf {v} ^ {j + 1}\\right) \\leq L \\left(\\mathbf {v} ^ {j}\\right) - \\frac {1}{2 \\gamma} \\| \\nabla L (\\mathbf {v}) \\| ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 582, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, the following Theorem, which follows from Lemma 2.1, establishes the sublinear convergence of the MM procedure using bound (5), i.e., a convergence rate of $O(1 / J)$ , $J$ being the total number of iterations.", + "bbox": [ + 75, + 623, + 468, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2.2. ([3, p. 267]) For convex, twice-differentiable function $L(\\mathbf{v})$ , which has a $\\gamma$ -Lipschitz gradient, performing $J$ updates $\\mathbf{v}^{j+1} = \\mathbf{v}^j - \\frac{1}{\\gamma} \\nabla L(\\mathbf{v}^j)$ , starting from initialization $\\mathbf{v}^0$ , will yield a solution that satisfies:", + "bbox": [ + 73, + 691, + 468, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| L \\left(\\mathbf {v} ^ {J}\\right) - L \\left(\\mathbf {v} ^ {*}\\right) \\right\\| \\leq \\frac {\\gamma}{2 J} \\left\\| \\mathbf {v} ^ {0} - \\mathbf {v} ^ {*} \\right\\| \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 763, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L(\\mathbf{v}^{*})$ is the optimal value.", + "bbox": [ + 76, + 800, + 303, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For completeness, we provide the proofs of these well-known results in the supplemental material. Clearly, Lemma 2.1 and Theorem 2.2 prescribe a learning rate of $\\frac{1}{\\gamma}$ for a function that has a $\\gamma$ -Lipschitz gradient. One valid Lipschitz constant would be the maximum eigen value of", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the Hessian of our objective in (3), which provides a majorizing function of the form in Eq. (5) and data-driven learning rates. However, a naive spectral decomposition of the Hessian matrices (to obtain the maximum eigen value) could be computationally intensive. For instance, for ImageNet, the Hessian of our objective is of size $K(D + 1) \\times K(D + 1) \\approx 1M \\times 1M$ , as $D = 1024$ and $K = 1000$ . In Prop. 1, we derive approximate global and block-wise Lipschitz constants that can be computed efficiently (i.e., evaluating the maximum eigen value of a single $D \\times D$ matrix).", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Block-coordinate updates Our procedure provided in Alg. 1 belongs to the family of Block Majorize-Minimize (BMM) methods, well studied in the optimization community [13]. To minimize a multi-block objective, as in our case where the blocks correspond to variables $\\mathbf{w}$ and $\\alpha$ , we minimize one or many successive majorizing functions of the objective in each block, with the other block fixed, in a cyclic order:", + "bbox": [ + 496, + 275, + 890, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\mathbf {w} ^ {j}, \\boldsymbol {\\alpha}\\right) + \\nabla L _ {\\mathbf {w}} \\left(\\mathbf {w} ^ {j}\\right) ^ {t} \\left(\\mathbf {w} - \\mathbf {w} ^ {j}\\right) + \\frac {\\gamma_ {\\mathbf {w}}}{2} \\| \\mathbf {w} - \\mathbf {w} ^ {j} \\| ^ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 405, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL \\left(\\mathbf {w}, \\boldsymbol {\\alpha} ^ {j}\\right) + \\nabla L _ {\\boldsymbol {\\alpha}} \\left(\\boldsymbol {\\alpha} ^ {j}\\right) ^ {t} \\left(\\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j}\\right) + \\frac {\\gamma_ {\\boldsymbol {\\alpha}}}{2} \\| \\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j} \\| ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 440, + 890, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where in (8), block $\\alpha$ is fixed and, in (9), $\\mathbf{w}$ is fixed. $\\nabla L_{\\mathbf{w}}$ and $\\nabla L_{\\alpha}$ denote block-wise gradients, and $(\\gamma_{\\mathbf{w}},\\gamma_{\\alpha})$ are the block Lipschitz constants. Accommodating different choices of the block-cycling strategies and majorizing functions, BMM includes a breadth of optimizers as particular cases, such the well-known block coordinate gradient descent (BCGD) [1] and its projection-based variant. Indeed, the so-called Gauss-Seidel cycling [13] alternates steps (8) and (9), which corresponds to the BCGD method. One could also perform many successive steps in one block, as we do in Alg. 1, which corresponds to the so-called Essentially-Cyclic4 strategy [13]. Importantly, for a fairly large spectrum of choices of the cycling strategies, BMM enjoys the same sublinear convergence property as MM for convex objectives (with each block-wise update decreasing the objective), provided that the majorizing functions are strongly convex; see Theorem 3.1 in [13]. This is the case for (8) and (9). In our experiments, we observed that this block-wise variant performs better than a single-block MM; see Table 2. This might be explained by the fact that each block of variables has a dedicated step size. In the supplemental material, we provide results for different block-cycling strategies.", + "bbox": [ + 496, + 470, + 890, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global and block-coordinatewise Lipschitz constants In the following, we derive approximate global and block-", + "bbox": [ + 496, + 835, + 890, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Essentially-Cyclic means that there is a period during which each block is updated at least once.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23776", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coordinatewise Lipschitz constants for our objective function in (3), which could be evaluated efficiently. We deploy these in Alg. 1, to compute data-driven, block-wise learning rates for updating visual prototypes $\\mathbf{w}$ and blending parameters $\\alpha$ .", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. Considering the blocks of variables $\\mathbf{w}$ and $\\alpha$ , the gradient of our objective $L$ in (3) is block-coordinatewise Lipschitz continuous. For $\\tau_{1} \\geq 2$ , it has the following block Lipschitz constant for the set of variables in $\\mathbf{w}$ :", + "bbox": [ + 73, + 172, + 468, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {\\mathbf {w}} = \\frac {\\tau_ {1}}{4 N} \\lambda_ {m a x} \\left(\\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 244, + 468, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{max}(\\mathbf{A})$ denotes the maximum eigenvalue of matrix $\\mathbf{A}$ . Furthermore, for $\\tau_{1} \\geq 1$ , the expression in Eq. (10) provides a tighter but approximate block Lipschitz constant. Similarly, for $\\tau_{2} \\geq 1$ , we have following approximate block Lipschitz constant for the variables in $\\alpha$ :", + "bbox": [ + 75, + 290, + 468, + 366 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {\\boldsymbol {\\alpha}} = \\max _ {k} \\frac {\\tau_ {2}}{4 N} \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}\\right) ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 373, + 468, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, for $\\tau \\geq 2$ , the following expression provides an approximate global Lipschitz constant for objective (3), i.e., w.r.t all variables $\\mathbf{v} = (\\mathbf{w},\\alpha)$ :", + "bbox": [ + 75, + 421, + 468, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\tau \\max \\left(\\gamma_ {\\mathbf {w}}, \\gamma_ {\\boldsymbol {\\alpha}}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 476, + 468, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. The details are deferred to the supplemental material. The main ingredients of the proof are based on the Gershgorin circle theorem and the variational characterization of the maximum eigenvalue, following the min-max theorem, also referred to as the variational principle. $\\square$", + "bbox": [ + 75, + 500, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.2. Initialization of the variables", + "text_level": 1, + "bbox": [ + 76, + 583, + 333, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following Prop. 2, we derive approximations of the minima of our objective function (3), which yield a data-informed initialization of the variables. Indeed, the expressions we obtain in Eqs. (13) and (15) suggest initial guesses for variables $\\mathbf{w}$ and $\\alpha$ . Interestingly, and as will be confirmed by our experiments (Table 3), such an initialization yields substantially lower values of the minimized loss than a random initialization. Furthermore, surprisingly, using this initialization for a training-free prediction yields better performances than the training-free version of the recent Tip-Adapter-F approach [28]. The details of our training-free version are provided in the supplemental material.", + "bbox": [ + 75, + 606, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2. The cross-entropy in Eq. (3) could be written as the sum of two convex functions, i.e., $L = g_{1} + g_{2}$ , such that, $\\forall k \\in [1, \\dots, K]$ , the minimum of $g_{1}$ w.r.t $\\mathbf{w}_{k}$ is co-linear to the hard mean vector of features within class $k$ :", + "bbox": [ + 75, + 794, + 468, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {1} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i k}} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 862, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the minimum of $g_{2}$ w.r.t $\\pmb{w}_{k}$ is co-linear to the soft mean vector of features within class $k$ :", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {2} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i k}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 133, + 890, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\leq \\min_k\\lambda_{min}(\\pmb {A}_k)$ $A_{k} = \\frac{1}{N}\\sum_{i = 1}^{N}(p_{ik} - p_{ik}^{2})\\pmb{f}_{i}\\pmb{f}_{i}^{t}$ and $\\lambda_{min}(\\mathbf{A})$ denotes the smallest eigenvalue of matrix $\\mathbf{A}$", + "bbox": [ + 498, + 189, + 890, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. We defer the details, including the full expressions of convex functions $g_{1}$ and $g_{2}$ , to the supplemental material.", + "bbox": [ + 498, + 234, + 890, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37c5aa575f31b844d941e9a66dc18a54ac3c1ec240bc5fadac45a09a9ce22693.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 266, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly to the development in Prop. 2, one could decompose (3) as the sum of two convex functions, i.e., $L = h_1 + h_2$ , such that, $\\forall k \\in [1, \\dots, K]$ , the minima of $h_1$ and $h_2$ w.r.t $\\alpha_k$ could be written, up to a multiplicative positive factor, as the hard and soft means of the cosine similarities between the image and text embeddings:", + "bbox": [ + 498, + 296, + 890, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\alpha_ {k}} {\\arg \\min } h _ {1} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 400, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\alpha_ {k}} {\\arg \\min } h _ {2} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 458, + 890, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\beta = \\min_k\\frac{1}{N}\\sum_{i = 1}^N (p_{ik} - p_{ik}^2)(\\pmb {f}_i^t\\pmb {t}_k)^2$ and $h_1$ (respectively $h_2$ ) has the same expression as $g_{2}$ (respectively $g_{1}$ ), except that term $\\frac{\\lambda}{2}\\sum_{k = 1}^{K}\\| \\pmb {w}_k\\|^2$ is replaced by $\\frac{\\beta}{2}\\|\\pmb {\\alpha}\\|^2$ ; see the supplemental material for the expressions of $g_{1}$ and $g_{2}$ .", + "bbox": [ + 498, + 508, + 890, + 585 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Block coordinate MM $(\\mathbf{w},\\alpha)$" + ], + "code_body": "iterw=10;iter $\\alpha = 1$ . $\\tau_{1} = 1$ . $\\tau_{2} = 16$ . $\\lambda = \\frac{1}{N}$ $\\beta = \\frac{1}{250K}$ Initialize w0,0 // Using (13) for each k Initialize $\\alpha^{0,0}$ // Using (15) for each k for $j = 0,1,\\ldots$ do for $l_{1} = 0,1,\\ldots$ ,iterw do $\\begin{array}{r}\\mathbf{w}^{j,l_1 + 1} = \\mathbf{w}^{j,l_1} - \\frac{1}{\\gamma_{\\mathbf{w}}}\\nabla L_{\\mathbf{w}}(\\mathbf{w}^{j,l_1},\\boldsymbol{\\alpha}^{j,0})\\\\ /\\quad \\frac{1}{\\gamma_{\\mathbf{w}}}\\text{from Eq.}(10) \\end{array}$ $\\mathbf{w}^{j + 1,0} = \\mathbf{w}^{j,\\mathrm{iter}_{\\mathbf{w}}}$ for $l_{2} = 0,1,\\dots$ iter $\\alpha$ do $\\begin{array}{rl} & {\\boldsymbol{\\alpha}^{j,l_2 + 1} = \\boldsymbol{\\alpha}^{j,l_2} - \\frac{1}{\\gamma_\\alpha}\\nabla L_\\alpha (\\mathbf{w}^{j + 1,0},\\boldsymbol{\\alpha}^{j,l_2})}\\\\ & {/ / \\frac{1}{\\gamma_\\alpha}\\text{from Eq.}(11)} \\end{array}$ $\\alpha^{j + 1,0} = \\alpha^{j,\\mathrm{iter}_{\\alpha}}$", + "bbox": [ + 511, + 640, + 854, + 858 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57d0ba468f52623d04854ed70dda6cf8f426b80567fd0009f876eba7780d85ed.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Number of shots (S)124816
Zero-shot CLIP ICML'21[23]58.89
Prompt-LearningCoOp ICV'22[29]59.62 ± 3.1163.80 ± 2.3267.23 ± 1.6471.30 ± 0.8674.06 ± 0.55
PLOTICLR'23[4]61.51 ± 2.9165.67 ± 2.0668.39 ± 1.1771.96 ± 0.7074.35 ± 0.66
KgCoOp CVPR'23[27]61.36 ± 3.0463.23 ± 2.0665.73 ± 1.1567.50 ± 1.1169.01 ± 0.79
ProGrad ICCV'23[30]62.46 ± 1.8965.88 ± 1.4668.52 ± 1.1571.82 ± 0.1173.95 ± 0.68
CLIP-based AdaptersCLIP-Adapter ICV'23[9]60.32 ± 0.8061.93 ± 0.9365.12 ± 0.8069.20 ± 0.5672.57 ± 0.54
Tip-Adapter-F ECCV'22[28]61.29 ± 0.9262.94 ± 0.7566.02 ± 0.8069.88 ± 0.5173.82 ± 0.55
Tip-Adapter-F* ECCV'22[28]63.06 ± 1.0566.47 ± 0.6568.71 ± 0.9671.78 ± 1.0074.37 ± 0.35
Linear-ProbingStandard LP ICML'21[23]36.10 ± 1.4346.99 ± 1.2956.72 ± 1.2064.66 ± 0.5570.56 ± 0.44
LP++63.43 ± 0.9066.20 ± 0.7269.16 ± 0.7972.04 ± 0.4674.42 ± 0.45
", + "bbox": [ + 127, + 87, + 843, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison to state-of-the-art methods. Average classification accuracy (%) on 11 benchmarks, with standard derivation over 10 sampled support sets for each dataset. The best values are highlighted in bold.", + "bbox": [ + 76, + 263, + 892, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 76, + 316, + 209, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1. Datasets and Implementation details", + "text_level": 1, + "bbox": [ + 76, + 340, + 392, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the CLIP-based few-shot adaptation literature [27, 28], we conduct the main experiments on 11 public classification data sets: Caltech101 [8], ImageNet [7], DTD [5], OxfordPets [22], Flowers102 [20], StandfordCars [17], Food101 [2], FGVCAircraft [19], SUN397 [26], EuroSAT [11] and UCF101 [25]. We follow standard practices [23] and consider $S = \\{1, 2, 4, 8, 16\\}$ shots for model adaptation, which are randomly sampled for each data set.", + "bbox": [ + 76, + 364, + 468, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Towards a fair validation set. Apparently, prior works on this problem [28] have resorted to a large set of validation samples to adjust their hyper-parameters. For the sake of fairness, we tune the hyper-parameters across all the methods based on a small validation set, which contains as many samples (i.e., shots) as the training set. Furthermore, to avoid the potential overfitting on the few training samples, we adopt an early stopping strategy on this validation set.", + "bbox": [ + 76, + 487, + 468, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "General Setting. While the existing works evaluate methods based on either a single or three random tasks (support sets) [28, 29], we found that, for some datasets, the chosen support samples may not be representative of the class, leading to large standard deviations in low-shot scenarios (see Fig. 1). To ensure fair comparisons, we evaluate all the methods by averaging their classification accuracies over 10 randomly sampled tasks, for each dataset. In all the experiments, we employ ResNet-50 [10] as the visual encoder for the CLIP backbone. It is important to note that, for our BMM procedure in Alg. 1, the optimizer hyper-parameters remain fixed across all the datasets. We use the validation set only to find the best model, via a single run of our BMM procedure with a fixed number of variable updates, i.e., 300 gradient updates including all the blocks of variables.", + "bbox": [ + 76, + 608, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. Baselines", + "text_level": 1, + "bbox": [ + 76, + 845, + 184, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We benchmark the proposed $\\mathrm{LP}++$ against relevant state-of-the-art methods in the few-shot adaptation of CLIP-based", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "models. We first resort to zero-shot CLIP as the standard baseline, which only leverages the knowledge learned by the pre-trained CLIP model. Also, we include the standard LP baseline, whose implementation is done following [23, 29]. More concretely, this baseline optimizes the standard cross-entropy loss, which corresponds to $\\alpha_{k} > 0\\forall k$ in our generalization in (3), using the L-BFGS [21] optimizer5. It also includes an $l_{2}$ -regularizer, whose balancing weight is set based on the validation set.", + "bbox": [ + 496, + 316, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP-based adapters. We benchmark $\\mathrm{LP}++$ against two popular adapter-based few-shot approaches: CLIP-adapter [9] and TIP-adapter [28]. As exposed earlier, several popular works follow unfair practices by resorting to a larger validation set, or even to the entire test set, to adjust their key hyperparameters -as well as their model selection criteria (i.e., epochs)- for each task. For the sake of fairness, we re-implement Tip-Adapter-F and report the results in 2 different settings. In the first setting (Tip-Adapter-F), we set the two crucial hyper-parameters of this method to 1, keeping them fixed during training, and adopt early stopping based on the validation set. In the second setting, referred to as Tip-Adapter- $\\mathbf{F}^*$ , we perform intensive grid-search on the validation set to find the best values for these hyperparameters at initialization, which incurs an additional time complexity burden compared to Tip-Adapter-F.", + "bbox": [ + 496, + 454, + 892, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prompt learning. We further compare the proposed LP++ to relevant prompt-learning methods, including CoOp[29] and more recent variants, such as PLOT[4], KgCoOp[27] and ProGrad[30]. We also apply early stopping here based on the performances on the validation set.", + "bbox": [ + 496, + 696, + 890, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Results", + "text_level": 1, + "bbox": [ + 500, + 777, + 591, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to the state-of-the-art In Table 1, we present the quantitative results obtained by $\\mathrm{LP}++$ (Alg. 1)", + "bbox": [ + 496, + 801, + 890, + 832 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "${}^{5}$ L-BFGS (Limited-memory BFGS) aims to find the minimum of objective function using the second order method. It estimates the Hessian matrix based on recent gradients only, enabling it to determine the steepest direction for achieving the optimal solution. Additionally, it is implemented with line searches to automatically determine the optimal step size.", + "bbox": [ + 496, + 838, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/399c52c91fb727f8824f3c4ac338827d397517cc4e2d198e3da8e65a58be093f.jpg", + "image_caption": [ + "Figure 3. Quantitative performance of different adaptation methods on the 11 benchmarks (mean), as well as in two other datasets, averaged over 10 tasks (additional figures on the remaining 9 datasets can be found in Appendix, Sec. 16)." + ], + "image_footnote": [], + "bbox": [ + 143, + 90, + 377, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97406d7969a0fb71137bbe71d34d449e1344a43d80a2b3b0332b491e4414df72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 90, + 602, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d86fd7af33d3477ccc718f4954a040d2e8e6d99706c01450054ac865990d2dad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 90, + 823, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and the relevant literature in the task of efficient adaptation of VLMs. We report the average classification accuracy and standard deviation, across 11 classification benchmarks. From these results, one can make several observations. First, while the standard LP baseline largely underperforms the existing adaptation methods, our improved version, $\\mathrm{LP}++$ , brings significant performance gains, particularly in the low-labeled data regimes. It is important to stress that the standard LP baseline integrates only the visual features extracted from CLIP, disregarding the text-encoder knowledge. This contrasts with the existing adapters, which leverage both image and text information. Therefore, these results evidence that the potential of LP has been severely underestimated in the existing literature. Second, when the model selection process is performed fairly, i.e., using a small validation set, their performances fall behind the proposed method (from $1\\%$ to $5\\%$ ), despite being arguably more complex approaches. In particular, among the adapter-based strategies, only Tip-Adapter-F* yields performances on par with $\\mathrm{LP}++$ , but at the cost of increasing computational load, due to an additional intensive grid-search over its hyper-parameters. If we look at prompt-learning methods, ProGrad and PLOT might be considered as competitors of $\\mathrm{LP}++$ , particularly as the number of shots increases. Nevertheless, as already discussed, these methods are computationally inefficient compared to adapters, and do not enable black-box adaptation. The overall performance across all the methods is depicted in Fig. 3 for the average over the 11 benchmarks, as well as in two datasets, showing that $\\mathrm{LP}++$ typically outperforms existing methods under different few-shot scenarios. A more detailed analysis is deferred to the Appendix, Sec. 16.", + "bbox": [ + 75, + 294, + 472, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on the loss functions and different optimization strategies. Table 2 reports the test accuracy and run time for different optimizers and loss functions, including the standard CE loss $(\\alpha_{k} = 0\\forall k)$ and our loss with learnable blending parameters. Additionally, for $\\mathrm{LP + + }$ , we evaluate our loss with fixed blending parameters $(\\alpha_{k} = 1\\forall k)$ . Independently of the optimizer used, the main takeaway from Table 2 is that introducing the text knowledge and making", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\alpha_{k}$ learnable (rather than fixed) have a substantial impact on accuracy. As for evaluating the optimizers, and for a fair comparison, we use a fixed budget for the number of variable updates for all the optimizers (i.e., 300 updates). In the case of $\\mathrm{LP}++$ , this corresponds to the total number of updates for all the blocks. Also, for all optimizers, we initialize $\\mathbf{w}$ and $\\alpha$ following Eqs. (13) and (15). We first consider two popular optimizers, i.e., GD and ADAM, and deploy them in two different settings. First, we run each optimizer 7 times, with each run corresponding to a learning rate in the range $[10^{-4}, 10^{2}]$ . Then, we record the best performances obtained on the validation set; see GD (optimum) and ADAM (optimum) in Table 2. This follows the standard practices in deep learning, i.e., searching for the learning rates over validation sets, which incurs additional computation overhead; see the time column in Table 2. Second, we run GD and ADAM with our data-driven learning rate, as prescribed by the approximate Lipschitz constant we derived in Eq. (12), with $\\tau = 1$ ; see GD (our Lipschitz cst) and ADAM (our Lipschitz cst) in Table 2. Note that, in this case, GD corresponds to $\\mathrm{LP}++$ with a single block. Furthermore, we include the L-BFGS in these comparisons, with its initial learning rate set to 1. For L-BFGS, implementing a line search for the optimal step size also introduces an additional computational overhead. As highlighted by Table 2, our method removes the need for validation searches for the optimization hyper-parameters, thanks to its data-driven, task-specific step sizes, thereby reducing the computational load for fine-tuning. In the meanwhile, it yields performances on par with those obtained with the best learning rates found with the validation set. In Fig. 4, we plot the performances of single-block GD vs. the learning rates in the range $[10^{-4}, 10^{2}]$ , for three datasets. We observe that our Lipschitz-based, task-specific step sizes match the optimal ones found on the validation set, although these vary among the datasets. Also, interestingly, these Lipschitz-based step sizes are orders-of-magnitude larger than those used in deep learning, which are, typically, within interval $[10^{-4}, 10^{-2}]$ ; see [28], for instance.", + "bbox": [ + 496, + 294, + 893, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How to initialize the classifier? To justify empirically the", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/981378ca95952e9ba244a68193192d2f88625889f51da89ab839ffd6df107efd.jpg", + "image_caption": [ + "Figure 4. Single-block GD performance as a function of different values of the learning rates. The dotted vertical line shows the Lipschitz-based, data-driven learning rate." + ], + "image_footnote": [], + "bbox": [ + 140, + 89, + 375, + 238 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e30e12bebdb58dc8c0958692c3af848dcec042dc52c580455523bf65982219bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 599, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1545f4288f93bb4dc4292c0f3a22e4254bb4d62fe49218e0f91909f2fa26924c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 90, + 821, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9a08db9866e8f6439f6a862f35782fe179d19be4b7a576f218351ab7337d4a2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Optimization MethodStandard loss (αk=0)Our lossTime
1 shot16 shot1 shot16 shot
LP++ (αk is learnt)35.5569.7563.4374.420.78s
LP++ (αk = 1)--46.3769.410.72s
GD (our Lipschitz cst)35.5569.7563.0474.280.86s
GD (optimum)35.6469.6762.9374.556.02s
ADAM (our Lipschitz cst)26.2264.5325.4464.620.98s
ADAM (optimum)35.9069.7363.0774.236.89s
L-BFGS [21]34.5467.4462.0972.826.73s
", + "bbox": [ + 78, + 301, + 467, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "advantages brought by initializing the classifier weights following Eq. (13), and the blending parameters in Eq. (15), we evaluate objective (3) at the beginning of the training, as well as the training-free test accuracy of three methods: random initialization, and the training-free version of both the Tip-adapter-F method [28] (i.e., Tip-adapter) and our method (we provide more details on training-free $\\mathrm{LP}++$ in the appendix). The results in Table 3 confirm empirically the technical observations in Eqs. (13) and (15), which prescribe initial guesses for our problem's variables.", + "bbox": [ + 75, + 508, + 468, + 660 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60fa20257fb8070eb056729891cda278052065e175caee15a7e7a61259f23942.jpg", + "table_caption": [ + "Table 2. Accuracy and run time for different optimizers and loss functions (average over 11 datasets). The running time is recorded for our loss (16 shots), and is averaged over 11 datasets. The best result is marked in bold, and the second best is underlined." + ], + "table_footnote": [], + "table_body": "
Number of shots (S)124816
ModelsInitial Loss (L0)
- Random Initialization21.4521.4221.3121.2721.32
Proposed Initialization1.601.541.471.351.21
ModelsInitial test accuracy (Acc0)
- Random Initialization17.7917.9218.0318.1018.22
Tip-adapter[28]59.2859.7260.5562.0964.29
Proposed Initialization59.7060.6662.0464.1666.20
", + "bbox": [ + 94, + 670, + 450, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational overhead. As the literature on adapting VLMs is gaining popularity, it is essential to evaluate the extent to which novel methods are efficient. To do this, we report the overall computational overhead of the approaches", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "studied in this work, which includes the time required for training and for finding the hyper-parameters, when applicable. We also indicate whether these methods enable black-box adaptation which, in our perspective, is a critical aspect in novel strategies aiming to address practical, real-world demands. The numerical values in Table 4 show that, in addition to yielding state-of-the-art performance (shown in previous sections), $\\mathrm{LP}++$ is the most efficient method (by several orders of magnitude), and does not require to access the internal representations of the pre-trained models.", + "bbox": [ + 496, + 305, + 893, + 458 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/152978e97519e079815a97ace36414aa048371ee11b4a4c0311acc50c91d2255.jpg", + "table_caption": [ + "Table 3. Comparison of the initial loss and test accuracy in the training-free scenario: random initialization, Tip-adapter and the proposed initialization. The results are averaged over 11 datasets." + ], + "table_footnote": [], + "table_body": "
MethodsOverall TimeBlackBox# Parameters
CoOp[29]~ 17hXK × M × D
PLOT-2[4]~ 10hXP × K × M × D
KgCoOp[27]~ 4hXK × M × D
ProGrad[30]~ 20hXK × M × D
Clip-Adapter[9]~ 40min✓2(D1 × D)
Tip-adapter-F[28]~ 6min✓K × S × D
Tip-adapter-F*[28]~ 50min✓K × S × D
Standard LP[23]3min✓K × D
LP++~ 2s✓K(D+1)
", + "bbox": [ + 506, + 465, + 887, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Run time and suitability to black-box scenarios for different methods on 16-shot ImageNet. All the experiments are performed on a single NVIDIA RTX A6000 GPU, except for PLOT-2, which is evaluated on two A6000 GPUs. $D_{1} = 256$ , and $D = 1024$ . The number of context tokens $M$ is set to 16. For PLOT, $P = 4$ is the number of prompts.", + "bbox": [ + 496, + 628, + 893, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 739, + 619, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced $\\mathrm{LP}++$ , a strong linear probe for few-shot CLIP adaptation. A specific modeling of the classifier weights, blending visual prototypes and text embeddings via learnable multipliers, along with convex-optimization ingredients, often overlooked in deep learning practices, led to the surprising results. While the findings of this work do not invalidate the promise of prompt learning and adaptation research, we believe $\\mathrm{LP}++$ could be used as a baseline to measure progress in these strongly emergent areas.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amir Beck and Luba Tetruashvili. On the convergence of block coordinate descent type methods. SIAM Journal on Optimization 23(4), 23(4):2037-2060, 2013. 4, 2", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 6", + "[3] Sébastien Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends in Machine Learning, 8(3-4):231-358, 2015. 4", + "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 2, 6, 8, 7", + "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6", + "[6] Pierre Colombo, Victor Pellegrain, Malik Boudiaf, Victor Storchan, Myriam Tami, Ismail Ben Ayed, Celine Hudelot, and Pablo Piantanida. Transductive learning for textual few-shot classification in api-based embedding models. In Empirical Methods in Natural Language Processing, 2023. 2", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6", + "[8] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Workshop on Computer Vision and Pattern Recognition, pages 178-178, 2004. 6", + "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, pages 1-15, 2023. 2, 6, 8, 7", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[12] Steffen Herbold. Autorank: A python package for automated ranking of classifiers. Journal of Open Source Software, 5 (48):2173, 2020. 4", + "[13] Mingyi Hong, XiangfengWang, Meisam Razaviyayn, and Zhi-Quan Luo. Iteration complexity analysis of block coordinate descent methods. Mathematical Programming, 163: 85–114, 2017. 4", + "[14] Z hong, D Friedman, and D Chen. Factual probing is [mask]: Learning vs. learning to recall. In Conference of the North" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "American Chapter of the Association for Computational Linguistics, 2021. 2", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916, 2021. 1", + "[16] Z Jiang, F Xu, J Araki, and G Neubig. How can we know what language models know. In Association for Computational Linguistics, 2020. 2", + "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In International Conference on Computer Vision, pages 554–561, 2013. 6", + "[18] Kenneth Lange, David R Hunter, and Ilsoon Yang. Optimization transfer using surrogate objective functions. Journal of computational and graphical statistics, 9(1):1-20, 2000. 3", + "[19] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 6", + "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 6", + "[21] Jorge Nocedal. Updating quasi-newton matrices with limited storage. Mathematics of Computation, 35(151):773-782, 1980. 6, 8", + "[22] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505. IEEE, 2012. 6", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 3, 6, 8, 7", + "[24] T Shin, Logan R. L. IV Razeghi, Y, E Wallace, and S Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. 2020. 2", + "[25] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 6", + "[26] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3485-3492. IEEE, 2010. 6", + "[27] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 8, 7", + "[28] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 1, 2, 5, 6, 7, 8, 3, 4" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23781", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6, 8, 4, 7", + "[30] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision, 2023. 2, 6, 8, 7" + ], + "bbox": [ + 78, + 90, + 468, + 200 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23782", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_model.json b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3a005211a1a2fc7ec96518c234495d1a645585b1 --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_model.json @@ -0,0 +1,1848 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "\\(\\mathbf{LP} + +: A\\) Surprisingly Strong Linear Probe for Few-Shot CLIP" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.179, + 0.773, + 0.251 + ], + "angle": 0, + "content": "Yunshi Huang\\(^{1*}\\) Fereshteh Shakeri\\(^{1*}\\) \nJose Dolz\\(^{1}\\) Malik Boudiaf\\(^{1}\\) Houda Bahig\\(^{2}\\) Ismail Ben Ayed\\(^{1}\\) \n\\(^{1}\\)ÉTS Montréal, \\(^{2}\\)Université de Montréal" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.285, + 0.314, + 0.302 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.771 + ], + "angle": 0, + "content": "In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin \\( LP++ \\), step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, \\( LP++ \\) operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.797, + 0.21, + 0.813 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Recently, there has been a growing popularity in multimodal learning methods, which process and merge information from diverse modalities. In particular, large-" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.288, + 0.868, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.499, + 0.896, + 0.625 + ], + "angle": 0, + "content": "Figure 1. Comparison of \\(\\mathrm{LP}++\\) with state-of-the-art few-shot CLIP methods in the 1-shot setting across 11 datasets. We compute the mean accuracy and standard deviation using 10 random tasks for each dataset. The error bars indicate the average standard deviation over all 11 datasets. The x-axis represents the run time for one task, averaged over the 11 datasets. Tip-Adapter-F and Tip-Adapter-F* are two re-implementations of Tip-Adapter-F [28], with fixed and grid-search hyper-parameters, respectively (implementation details provided in Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "scale vision-language models (VLMs), such as CLIP [23] and ALIGN [15], have attracted wide attention and made substantial progress in computer vision, showing promising generalization capabilities in various downstream tasks. Unlike conventional task-specific models that are trained with a predetermined set of labels, these so-called foundation models learn to align images with text, in an open-vocabulary fashion. They train, via contrastive learning, vision and text embeddings jointly using a large-scale amount of image-text pairs collected over the internet, thereby leveraging the rich semantic knowledge inherent to language (e.g., concept hierarchies). For a given downstream image classification task, vision-language embeddings enable zero-shot predictions, without re-training, using textual descriptions of the classes (a.k.a prompts). For instance, for a given class \\( k \\), a textual description of the class," + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution. Correspondence to yunshi.huang@etsmtl.ca, fereshteh.shakeri.1@etsmtl.net" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.184 + ], + "angle": 0, + "content": "which we denote \\( z_{k} \\), could be \"a photo of a \\([class_{k}]\\)\", where \\([class_{k}]\\) is the class name. Thus, the zero-shot class prediction for a query image \\( x \\) is obtained from the cosine similarity between the \\( l_{2} \\)-normalized vision-encoded embeddings, \\( f = \\theta_{v}(x) \\), and the text-encoded ones, \\( t_{k} = \\theta_{t}(z_{k}) \\): \\( \\hat{k} = \\arg \\max_{k} f^{t} t_{k} \\), where \\( t \\) denotes the transpose1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Motivated by the observation that the choice of input prompts \\( z_{k} \\) may affect the zero-shot predictions, and following on from the strong recent emergence of prompt learning research in the NLP community [14, 16, 24], the popular work in [29] pioneered context optimization (CoOp) for vision-language models. CoOp models input text \\( z_{k} \\) as learnable continuous vectors, e.g., in the form \\( z_{k} = (z_{k}^{1},\\dots,z_{k}^{M},[\\mathrm{class}_{k}]) \\), where \\( (z_k^l)_{1\\leq l\\leq M} \\) are learnable text tokens, \\( [\\mathrm{class}_k] \\) is a fixed token corresponding to the word embedding vector of the name of the \\( k^{th} \\) class, and \\( M \\) is a hyper-parameter. These learnable vectors are finetuned as task-specific prompts using few-shot training examples and a standard supervised classification loss. More specifically, in this few-shot setting, we assume access to a set consisting of a few labeled samples for each target class, often referred to as the support set. Let \\( f_{i} = \\theta_{v}(\\pmb{x}_{i}) \\) denote the vision embedding of support image \\( i \\), and \\( y_{ik} \\) its one-hot encoded label, i.e., \\( y_{ik} = 1 \\) if image \\( x_{i} \\) belongs to class \\( k \\) and 0 otherwise. Expressing the text embeddings as \\( t_k = \\theta_t(z_k^1,\\dots,z_k^M,[\\mathrm{class}_k]) \\), CoOp fine-tunes text tokens \\( (z_k^l)_{1\\leq l\\leq M} \\) by minimizing the cross-entropy (CE) loss, with \\( N \\) labeled support samples and \\( K \\) classes:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.525, + 0.469, + 0.567 + ], + "angle": 0, + "content": "\\[\n- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.469, + 0.607 + ], + "angle": 0, + "content": "where the softmax predictions \\( p_{ik} \\) and the logits (class scores) \\( l_{ik} \\) are given by:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.615, + 0.396, + 0.653 + ], + "angle": 0, + "content": "\\[\np _ {i k} = \\frac {\\exp (l _ {i k})}{\\sum_ {j = 1} ^ {K} \\exp (l _ {i j})}; \\quad l _ {i k} = \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.661, + 0.469, + 0.844 + ], + "angle": 0, + "content": "Although recent, the pioneering idea of CoOp has triggered a quite abundant literature on prompt learning for few-shot vision-language models, with numerous, more convoluted extensions, e.g. [4, 27, 30], to list a few. For instance, PLOT [4] followed up by learning multiple prompts, to describe the characteristics of each class, via minimizing an optimal-transport distance. KgCoOp [27] improves CoOp's performance when dealing with unseen classes, via minimizing the discrepancy between the text embeddings generated by the learned prompts and hand-crafted ones. While CoOp directly updates the context vectors using the CE loss, ProGrad [30] aligns the few-shot down" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.85, + 0.468, + 0.874 + ], + "angle": 0, + "content": "1For \\( l_{2} \\)-normalized feature embeddings, the dot product corresponds to the cosine similarity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.874, + 0.468, + 0.902 + ], + "angle": 0, + "content": "2The number of labeled support samples per class, \\(S = \\frac{N}{K}\\) , is small, typically in \\(\\{1,2,4,\\dots,16\\}\\)" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.85, + 0.468, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "stream knowledge with the large-scale general knowledge, thus mitigating the overfitting of the few-shot samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.408 + ], + "angle": 0, + "content": "Prompt learning methods have brought significant improvements over zero-shot classification, but they come at the price of heavy computational and memory load, as they require gradient back-propagation through the entire text encoder. Furthermore, they assume knowledge of the text encoder. These aspects may impede their deployment in low-resource and black-box, privacy-preserving scenarios, which are of wide interest in practice. Indeed, in NLP, there is currently an emerging literature on fast few-shot adaptation of black-box models [6], strongly motivated by the fact that large-scale foundation models (e.g., the GPT family, Anthropic's Claude or Google's PaLM) are only available through APIs and their pre-trained weights are not shared. Finally, by evaluating prompt learning methods over larger numbers of sampled support sets in our experiments (Fig. 1), we observed that they exhibit large variation in performances. This could be explained by the fact that, through the text encoder, they learn prompts that are \"too specialized\" for a given image support set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.409, + 0.892, + 0.514 + ], + "angle": 0, + "content": "While prompt learning alters the textual inputs, another category of approaches, referred to as adapters, focused on transforming the pre-training features of the visual or language encoders, e.g., [9, 28]. These adapters are non-linear transformations, for instance, in the form of multi-layer modules, added to the encoder's bottleneck. They learn additional transformations, yielding logits of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.526, + 0.891, + 0.543 + ], + "angle": 0, + "content": "\\[\nl _ {i k} = \\boldsymbol {\\theta} _ {a} \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {t} _ {k}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.884 + ], + "angle": 0, + "content": "The adapter's learnable parameters, \\(\\theta_{a}\\), are fine-tuned over a few-shot task by optimizing the cross-entropy loss, similarly to (1) but with logits \\(l_{ik}\\) expressed as functions of \\(\\theta_{a}\\). For instance, the popular CLIP-Adapter [9] integrated a multi-layered perceptron to modify the features, along with residual connections, which enable blending with the original pre-trained features. Tip-Adapter [28] added a nonlinear, quadratic-complexity module, which evaluates the pairwise similarities between the features of the support sets, and blends the ensuing class scores with the original textual features. This category of approaches mitigates some of the limitations of prompt-learning methods as they result in few-shot adaptation that has significantly lower computation and memory loads. However, as shown in Fig. 1 and in our experiments, their performances seem to depend strongly on some key hyper-parameters that have to be adjusted carefully on each downstream task, e.g. those that control the blending between the vision and language features. Therefore, to perform competitively, they incur an additional computation overhead to the adaptation phase, due to intensive (e.g., grid) search of the hyper-parameters over task-dedicated validation sets." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the above-mentioned, strongly emergent literature on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "few-shot CLIP adaptation, linear probe (LP) [23] has been often reported as a very weak baseline. For instance, in the 1-shot setting, it scores near \\(20\\%\\) lower than the zero-shot predictions averaged over 11 benchmarks (Table 1). Initially evaluated in [23], LP is a linear classifier on the vision-encoded features. Specifically, it optimizes the CE loss (1) w.r.t the last-layer weights of the vision encoder (i.e., the class prototypes), which we will denote \\((\\pmb{w}_k)_{1\\leq k\\leq K}\\) in the rest of the paper, with the logits given by: \\(l_{ik} = \\pmb{f}_i^t\\pmb{w}_k\\). A clear deficiency in this standard LP baseline is that it omits completely the language knowledge of CLIP, i.e., \\((\\pmb{t}_k)_{1\\leq k\\leq K}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.472, + 0.348 + ], + "angle": 0, + "content": "In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline. Specifically, we extend the logits in the CE loss in (1), so that they become learnable functions of the text embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.356, + 0.347, + 0.372 + ], + "angle": 0, + "content": "\\[\nl _ {i k} = \\boldsymbol {f} _ {i} ^ {t} (\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.759 + ], + "angle": 0, + "content": "with \\((\\alpha_{k})_{1\\leq k\\leq K}\\) trainable class-wise parameters blending image and text knowledge. As our objective function depends on two types of variables, i.e., the visual class prototypes \\((\\pmb{w}_k)_{1\\leq k\\leq K}\\) and blending parameters \\((\\alpha_{k})_{1\\leq k\\leq K}\\), we propose a computationally efficient Block Majorize-Minimize (BMM) procedure. In our full-batch MM optimizer, which we coin \\(\\mathrm{LP}++\\), step sizes are implicit in the definition of the majorizing functions, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. Moreover, we examine the mathematical properties of our objective, i.e., (i) Lipschitz gradient continuity and (ii) decomposition into convex functions having closed-form optima. This enabled us to build majorizing functions yielding data-driven learning rates, and to derive approximations of the objective-function minima, which yield data-informed initializations of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yield, surprisingly, highly competitive few-shot CLIP performances (Fig. 1). Furthermore, \\(\\mathrm{LP}++\\) operates in black-box, relaxes intensive validation searches for the optimization of hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP methods (Table 4). For instance, for 16-shot ImageNet adaptation, it takes seconds on a single NVIDIA RTX A600 GPU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.769, + 0.282, + 0.785 + ], + "angle": 0, + "content": "2. Formulation of \\(\\mathbf{LP} + +\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.47, + 0.857 + ], + "angle": 0, + "content": "Following the notations introduced in the previous section, we propose to minimize the following CE objective function w.r.t visual class prototypes \\(\\mathbf{w} = (\\pmb{w}_k)_{1\\leq k\\leq K}\\) and class-wise blending parameters \\(\\alpha = (\\alpha_{k})_{1\\leq k\\leq K}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.863, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {w}, \\boldsymbol {\\alpha}) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.891, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.23, + 0.799, + 0.244 + ], + "angle": 0, + "content": "Figure 2. Visualization of \\(\\mathbf{LP} + +\\)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.269, + 0.875, + 0.285 + ], + "angle": 0, + "content": "where softmax probability outputs \\( p_{ik} \\) are now given by:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.292, + 0.892, + 0.331 + ], + "angle": 0, + "content": "\\[\np _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) = \\frac {\\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k}\\right)\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {j} + \\alpha_ {j} \\boldsymbol {t} _ {j}\\right)\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.534 + ], + "angle": 0, + "content": "and \\( t_k \\) are the fixed pre-trained embeddings of the text templates used in zero-shot CLIP [23]. Clearly, the objective function defined by (3) and (4) could be viewed as a generalization of the vision-encoder CE loss used in the standard LP [23]. Indeed, the latter corresponds to setting \\( \\alpha_k = 0 \\forall k \\), i.e., no text knowledge. As we will see in our experimental ablation over different objective functions (Table 2), introducing the text knowledge \\( (\\alpha_k > 0 \\forall k) \\) has a substantial effect on performances. Also, making \\( \\alpha_k \\) learnable (rather than fixed) leads to a further significant impact. Indeed, we hypothesize that the optimal blending of the text and visual knowledge is task dependent, which motivates learning it from the context of the support set." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.542, + 0.882, + 0.558 + ], + "angle": 0, + "content": "2.1. Block coordinate Majorize-Minimize descent" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Majorize-Minimize (MM) [18] is a very general optimization principle, which includes different classes of standard optimizers such as gradient descent, concave-convex procedures and expectation-maximization. Let \\(\\mathbf{v} = (\\mathbf{w},\\alpha)\\in\\) \\(\\mathbb{R}^{K(D + 1)}\\) denote the overall vector of variables in our case, with \\(D\\) being the dimension of the feature embeddings. At each iteration, the MM procedure updates the variable as the minimum of a majorizing function, i.e., an upper bound on the original objective, which is tight at the current iteration \\(j\\colon L(\\mathbf{v})\\leq M(\\mathbf{v},\\mathbf{v}^j)\\) and \\(L(\\mathbf{v}^j) = M(\\mathbf{v}^j,\\mathbf{v}^j)\\). Thus, update step \\(\\mathbf{v}^{j + 1} = \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)\\) guarantees that the original objective does not increase at each iteration: \\(L(\\mathbf{v}^{j + 1})\\leq M(\\mathbf{v}^{j + 1},\\mathbf{v}^j)\\leq M(\\mathbf{v}^j,\\mathbf{v}^j) = L(\\mathbf{v}^j)\\). Therefore, in MM algorithms, step sizes are implicit in the definition of the majorizing function, unlike standard gradient-descent practices, in which the step sizes (a.k.a learning rates) are intensively searched over validation sets, via running the optimizer several times." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.838, + 0.892, + 0.868 + ], + "angle": 0, + "content": "In this work, we exploit the Lipschitz-gradient continuity of our convex objective in (3), i.e., bounds on the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This assumes, of course, that minimizing \\( M(\\mathbf{v}, \\mathbf{v}^j) \\) over \\( \\mathbf{v} \\) could be solved to global optimality and is easier than the original problem." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23775" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.365 + ], + "angle": 0, + "content": "maximum eigen values of the Hessian matrices (Prop. 1), thereby building majorizing functions with data-driven, task-specific step sizes. This removes the need for validation searches for the optimization hyper-parameters, reducing the computational load for fine-tuning (Table 2), while yielding performances on par with the best learning rates found with validation (Fig. 4). Also, interestingly, the Lipschitz-based learning rates computed from our derivation in Prop. 1 are orders-of-magnitude larger than those typically used in deep learning, yielding steeper decreases towards the minimum. Before giving proper majorizing functions for our convex, gradient-Lipschitz function in (3), let us first point to the following results, well-known in convex optimization [3]. While these results are text-book knowledge in optimization, they enable to connect the general MM principle to gradient descent, motivating the data-driven, task-specific step sizes we derive in Prop. 1 and the block-coordinate MM optimizer we propose in Alg. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.472, + 0.464 + ], + "angle": 0, + "content": "Lemma 2.1. ([3, p. 268]) Assume \\( L(\\mathbf{v}) \\) is a twice-differentiable function, which has a Lipschitz continuous gradient, i.e., there exists a strictly positive Lipschitz constant \\( \\gamma \\) such that \\( \\nabla^2 L(\\mathbf{v}) \\preceq \\gamma \\mathbf{I} \\), with \\( \\mathbf{I} \\) the identity matrix. Then, the following quadratic bound is a majorizing function for \\( L \\) at iteration \\( j \\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.473, + 0.47, + 0.5 + ], + "angle": 0, + "content": "\\[\nM (\\mathbf {v}, \\mathbf {v} ^ {j}) = L (\\mathbf {v} ^ {j}) + \\nabla L (\\mathbf {v} ^ {j}) ^ {t} (\\mathbf {v} - \\mathbf {v} ^ {j}) + \\frac {\\gamma}{2} \\| \\mathbf {v} - \\mathbf {v} ^ {j} \\| ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.508, + 0.47, + 0.573 + ], + "angle": 0, + "content": "Furthermore, a specific gradient step, with learning rate \\(\\frac{1}{\\gamma}\\) minimizes \\(M\\), i.e., \\(\\mathbf{v}^{j + 1} = \\mathbf{v}^j -\\frac{1}{\\gamma}\\nabla L(\\mathbf{v}^j)\\) = \\(\\arg \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)\\), and guarantees that objective \\(L\\) decreases by at least \\(\\frac{1}{2\\gamma}\\| \\nabla L(\\mathbf{v})\\| ^2\\):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.583, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\nL \\left(\\mathbf {v} ^ {j + 1}\\right) \\leq L \\left(\\mathbf {v} ^ {j}\\right) - \\frac {1}{2 \\gamma} \\| \\nabla L (\\mathbf {v}) \\| ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.624, + 0.47, + 0.684 + ], + "angle": 0, + "content": "Moreover, the following Theorem, which follows from Lemma 2.1, establishes the sublinear convergence of the MM procedure using bound (5), i.e., a convergence rate of \\( O(1 / J) \\), \\( J \\) being the total number of iterations." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.693, + 0.469, + 0.757 + ], + "angle": 0, + "content": "Theorem 2.2. ([3, p. 267]) For convex, twice-differentiable function \\( L(\\mathbf{v}) \\), which has a \\( \\gamma \\)-Lipschitz gradient, performing \\( J \\) updates \\( \\mathbf{v}^{j+1} = \\mathbf{v}^j - \\frac{1}{\\gamma} \\nabla L(\\mathbf{v}^j) \\), starting from initialization \\( \\mathbf{v}^0 \\), will yield a solution that satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.765, + 0.47, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\left\\| L \\left(\\mathbf {v} ^ {J}\\right) - L \\left(\\mathbf {v} ^ {*}\\right) \\right\\| \\leq \\frac {\\gamma}{2 J} \\left\\| \\mathbf {v} ^ {0} - \\mathbf {v} ^ {*} \\right\\| \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.801, + 0.304, + 0.817 + ], + "angle": 0, + "content": "where \\(L(\\mathbf{v}^{*})\\) is the optimal value." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "For completeness, we provide the proofs of these well-known results in the supplemental material. Clearly, Lemma 2.1 and Theorem 2.2 prescribe a learning rate of \\(\\frac{1}{\\gamma}\\) for a function that has a \\(\\gamma\\)-Lipschitz gradient. One valid Lipschitz constant would be the maximum eigen value of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "the Hessian of our objective in (3), which provides a majorizing function of the form in Eq. (5) and data-driven learning rates. However, a naive spectral decomposition of the Hessian matrices (to obtain the maximum eigen value) could be computationally intensive. For instance, for ImageNet, the Hessian of our objective is of size \\( K(D + 1) \\times K(D + 1) \\approx 1M \\times 1M \\), as \\( D = 1024 \\) and \\( K = 1000 \\). In Prop. 1, we derive approximate global and block-wise Lipschitz constants that can be computed efficiently (i.e., evaluating the maximum eigen value of a single \\( D \\times D \\) matrix)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Block-coordinate updates Our procedure provided in Alg. 1 belongs to the family of Block Majorize-Minimize (BMM) methods, well studied in the optimization community [13]. To minimize a multi-block objective, as in our case where the blocks correspond to variables \\(\\mathbf{w}\\) and \\(\\alpha\\), we minimize one or many successive majorizing functions of the objective in each block, with the other block fixed, in a cyclic order:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.406, + 0.892, + 0.433 + ], + "angle": 0, + "content": "\\[\nL \\left(\\mathbf {w} ^ {j}, \\boldsymbol {\\alpha}\\right) + \\nabla L _ {\\mathbf {w}} \\left(\\mathbf {w} ^ {j}\\right) ^ {t} \\left(\\mathbf {w} - \\mathbf {w} ^ {j}\\right) + \\frac {\\gamma_ {\\mathbf {w}}}{2} \\| \\mathbf {w} - \\mathbf {w} ^ {j} \\| ^ {2} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.441, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\nL \\left(\\mathbf {w}, \\boldsymbol {\\alpha} ^ {j}\\right) + \\nabla L _ {\\boldsymbol {\\alpha}} \\left(\\boldsymbol {\\alpha} ^ {j}\\right) ^ {t} \\left(\\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j}\\right) + \\frac {\\gamma_ {\\boldsymbol {\\alpha}}}{2} \\| \\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j} \\| ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.819 + ], + "angle": 0, + "content": "where in (8), block \\(\\alpha\\) is fixed and, in (9), \\(\\mathbf{w}\\) is fixed. \\(\\nabla L_{\\mathbf{w}}\\) and \\(\\nabla L_{\\alpha}\\) denote block-wise gradients, and \\((\\gamma_{\\mathbf{w}},\\gamma_{\\alpha})\\) are the block Lipschitz constants. Accommodating different choices of the block-cycling strategies and majorizing functions, BMM includes a breadth of optimizers as particular cases, such the well-known block coordinate gradient descent (BCGD) [1] and its projection-based variant. Indeed, the so-called Gauss-Seidel cycling [13] alternates steps (8) and (9), which corresponds to the BCGD method. One could also perform many successive steps in one block, as we do in Alg. 1, which corresponds to the so-called Essentially-Cyclic4 strategy [13]. Importantly, for a fairly large spectrum of choices of the cycling strategies, BMM enjoys the same sublinear convergence property as MM for convex objectives (with each block-wise update decreasing the objective), provided that the majorizing functions are strongly convex; see Theorem 3.1 in [13]. This is the case for (8) and (9). In our experiments, we observed that this block-wise variant performs better than a single-block MM; see Table 2. This might be explained by the fact that each block of variables has a dedicated step size. In the supplemental material, we provide results for different block-cycling strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.837, + 0.892, + 0.867 + ], + "angle": 0, + "content": "Global and block-coordinatewise Lipschitz constants In the following, we derive approximate global and block-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Essentially-Cyclic means that there is a period during which each block is updated at least once." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23776" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "coordinatewise Lipschitz constants for our objective function in (3), which could be evaluated efficiently. We deploy these in Alg. 1, to compute data-driven, block-wise learning rates for updating visual prototypes \\(\\mathbf{w}\\) and blending parameters \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.174, + 0.469, + 0.247 + ], + "angle": 0, + "content": "Proposition 1. Considering the blocks of variables \\( \\mathbf{w} \\) and \\( \\alpha \\), the gradient of our objective \\( L \\) in (3) is block-coordinatewise Lipschitz continuous. For \\( \\tau_{1} \\geq 2 \\), it has the following block Lipschitz constant for the set of variables in \\( \\mathbf{w} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.246, + 0.469, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {\\mathbf {w}} = \\frac {\\tau_ {1}}{4 N} \\lambda_ {m a x} \\left(\\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.47, + 0.367 + ], + "angle": 0, + "content": "where \\(\\lambda_{max}(\\mathbf{A})\\) denotes the maximum eigenvalue of matrix \\(\\mathbf{A}\\). Furthermore, for \\(\\tau_{1} \\geq 1\\), the expression in Eq. (10) provides a tighter but approximate block Lipschitz constant. Similarly, for \\(\\tau_{2} \\geq 1\\), we have following approximate block Lipschitz constant for the variables in \\(\\alpha\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.374, + 0.469, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {\\boldsymbol {\\alpha}} = \\max _ {k} \\frac {\\tau_ {2}}{4 N} \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}\\right) ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.467 + ], + "angle": 0, + "content": "Finally, for \\(\\tau \\geq 2\\), the following expression provides an approximate global Lipschitz constant for objective (3), i.e., w.r.t all variables \\(\\mathbf{v} = (\\mathbf{w},\\alpha)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.477, + 0.469, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\tau \\max \\left(\\gamma_ {\\mathbf {w}}, \\gamma_ {\\boldsymbol {\\alpha}}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Proof. The details are deferred to the supplemental material. The main ingredients of the proof are based on the Gershgorin circle theorem and the variational characterization of the maximum eigenvalue, following the min-max theorem, also referred to as the variational principle. \\(\\square\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.584, + 0.334, + 0.598 + ], + "angle": 0, + "content": "2.2. Initialization of the variables" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.607, + 0.469, + 0.789 + ], + "angle": 0, + "content": "In the following Prop. 2, we derive approximations of the minima of our objective function (3), which yield a data-informed initialization of the variables. Indeed, the expressions we obtain in Eqs. (13) and (15) suggest initial guesses for variables \\(\\mathbf{w}\\) and \\(\\alpha\\). Interestingly, and as will be confirmed by our experiments (Table 3), such an initialization yields substantially lower values of the minimized loss than a random initialization. Furthermore, surprisingly, using this initialization for a training-free prediction yields better performances than the training-free version of the recent Tip-Adapter-F approach [28]. The details of our training-free version are provided in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.856 + ], + "angle": 0, + "content": "Proposition 2. The cross-entropy in Eq. (3) could be written as the sum of two convex functions, i.e., \\( L = g_{1} + g_{2} \\), such that, \\( \\forall k \\in [1, \\dots, K] \\), the minimum of \\( g_{1} \\) w.r.t \\( \\mathbf{w}_{k} \\) is co-linear to the hard mean vector of features within class \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.863, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {1} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i k}} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "and the minimum of \\( g_{2} \\) w.r.t \\( \\pmb{w}_{k} \\) is co-linear to the soft mean vector of features within class \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.135, + 0.892, + 0.176 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {2} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i k}} \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.19, + 0.891, + 0.223 + ], + "angle": 0, + "content": "where \\(\\lambda \\leq \\min_k\\lambda_{min}(\\pmb {A}_k)\\) \\(A_{k} = \\frac{1}{N}\\sum_{i = 1}^{N}(p_{ik} - p_{ik}^{2})\\pmb{f}_{i}\\pmb{f}_{i}^{t}\\) and \\(\\lambda_{min}(\\mathbf{A})\\) denotes the smallest eigenvalue of matrix \\(\\mathbf{A}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.235, + 0.892, + 0.265 + ], + "angle": 0, + "content": "Proof. We defer the details, including the full expressions of convex functions \\( g_{1} \\) and \\( g_{2} \\), to the supplemental material." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.267, + 0.891, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.297, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Similarly to the development in Prop. 2, one could decompose (3) as the sum of two convex functions, i.e., \\( L = h_1 + h_2 \\), such that, \\( \\forall k \\in [1, \\dots, K] \\), the minima of \\( h_1 \\) and \\( h_2 \\) w.r.t \\( \\alpha_k \\) could be written, up to a multiplicative positive factor, as the hard and soft means of the cosine similarities between the image and text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.401, + 0.892, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\underset {\\alpha_ {k}} {\\arg \\min } h _ {1} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.459, + 0.891, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\underset {\\alpha_ {k}} {\\arg \\min } h _ {2} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Here, \\(\\beta = \\min_k\\frac{1}{N}\\sum_{i = 1}^N (p_{ik} - p_{ik}^2)(\\pmb {f}_i^t\\pmb {t}_k)^2\\) and \\(h_1\\) (respectively \\(h_2\\)) has the same expression as \\(g_{2}\\) (respectively \\(g_{1}\\)), except that term \\(\\frac{\\lambda}{2}\\sum_{k = 1}^{K}\\| \\pmb {w}_k\\|^2\\) is replaced by \\(\\frac{\\beta}{2}\\|\\pmb {\\alpha}\\|^2\\); see the supplemental material for the expressions of \\(g_{1}\\) and \\(g_{2}\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.512, + 0.624, + 0.804, + 0.639 + ], + "angle": 0, + "content": "Algorithm 1: Block coordinate MM \\((\\mathbf{w},\\alpha)\\)" + }, + { + "type": "algorithm", + "bbox": [ + 0.513, + 0.641, + 0.856, + 0.859 + ], + "angle": 0, + "content": "iterw=10;iter \\(\\alpha = 1\\) . \\(\\tau_{1} = 1\\) . \\(\\tau_{2} = 16\\) . \\(\\lambda = \\frac{1}{N}\\) \\(\\beta = \\frac{1}{250K}\\) Initialize w0,0 // Using (13) for each k Initialize \\(\\alpha^{0,0}\\) // Using (15) for each k for \\(j = 0,1,\\ldots\\) do for \\(l_{1} = 0,1,\\ldots\\) ,iterw do \\(\\begin{array}{r}\\mathbf{w}^{j,l_1 + 1} = \\mathbf{w}^{j,l_1} - \\frac{1}{\\gamma_{\\mathbf{w}}}\\nabla L_{\\mathbf{w}}(\\mathbf{w}^{j,l_1},\\boldsymbol{\\alpha}^{j,0})\\\\ /\\quad \\frac{1}{\\gamma_{\\mathbf{w}}}\\text{from Eq.}(10) \\end{array}\\) \\(\\mathbf{w}^{j + 1,0} = \\mathbf{w}^{j,\\mathrm{iter}_{\\mathbf{w}}}\\) for \\(l_{2} = 0,1,\\dots\\) iter \\(\\alpha\\) do \\(\\begin{array}{rl} & {\\boldsymbol{\\alpha}^{j,l_2 + 1} = \\boldsymbol{\\alpha}^{j,l_2} - \\frac{1}{\\gamma_\\alpha}\\nabla L_\\alpha (\\mathbf{w}^{j + 1,0},\\boldsymbol{\\alpha}^{j,l_2})}\\\\ & {/ / \\frac{1}{\\gamma_\\alpha}\\text{from Eq.}(11)} \\end{array}\\) \\(\\alpha^{j + 1,0} = \\alpha^{j,\\mathrm{iter}_{\\alpha}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23777" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.088, + 0.844, + 0.254 + ], + "angle": 0, + "content": "
Number of shots (S)124816
Zero-shot CLIP ICML'21[23]58.89
Prompt-LearningCoOp ICV'22[29]59.62 ± 3.1163.80 ± 2.3267.23 ± 1.6471.30 ± 0.8674.06 ± 0.55
PLOTICLR'23[4]61.51 ± 2.9165.67 ± 2.0668.39 ± 1.1771.96 ± 0.7074.35 ± 0.66
KgCoOp CVPR'23[27]61.36 ± 3.0463.23 ± 2.0665.73 ± 1.1567.50 ± 1.1169.01 ± 0.79
ProGrad ICCV'23[30]62.46 ± 1.8965.88 ± 1.4668.52 ± 1.1571.82 ± 0.1173.95 ± 0.68
CLIP-based AdaptersCLIP-Adapter ICV'23[9]60.32 ± 0.8061.93 ± 0.9365.12 ± 0.8069.20 ± 0.5672.57 ± 0.54
Tip-Adapter-F ECCV'22[28]61.29 ± 0.9262.94 ± 0.7566.02 ± 0.8069.88 ± 0.5173.82 ± 0.55
Tip-Adapter-F* ECCV'22[28]63.06 ± 1.0566.47 ± 0.6568.71 ± 0.9671.78 ± 1.0074.37 ± 0.35
Linear-ProbingStandard LP ICML'21[23]36.10 ± 1.4346.99 ± 1.2956.72 ± 1.2064.66 ± 0.5570.56 ± 0.44
LP++63.43 ± 0.9066.20 ± 0.7269.16 ± 0.7972.04 ± 0.4674.42 ± 0.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.264, + 0.893, + 0.293 + ], + "angle": 0, + "content": "Table 1. Comparison to state-of-the-art methods. Average classification accuracy (%) on 11 benchmarks, with standard derivation over 10 sampled support sets for each dataset. The best values are highlighted in bold." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.317, + 0.21, + 0.334 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.342, + 0.393, + 0.359 + ], + "angle": 0, + "content": "3.1. Datasets and Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.47, + 0.488 + ], + "angle": 0, + "content": "Following the CLIP-based few-shot adaptation literature [27, 28], we conduct the main experiments on 11 public classification data sets: Caltech101 [8], ImageNet [7], DTD [5], OxfordPets [22], Flowers102 [20], StandfordCars [17], Food101 [2], FGVCAircraft [19], SUN397 [26], EuroSAT [11] and UCF101 [25]. We follow standard practices [23] and consider \\( S = \\{1, 2, 4, 8, 16\\} \\) shots for model adaptation, which are randomly sampled for each data set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.47, + 0.608 + ], + "angle": 0, + "content": "Towards a fair validation set. Apparently, prior works on this problem [28] have resorted to a large set of validation samples to adjust their hyper-parameters. For the sake of fairness, we tune the hyper-parameters across all the methods based on a small validation set, which contains as many samples (i.e., shots) as the training set. Furthermore, to avoid the potential overfitting on the few training samples, we adopt an early stopping strategy on this validation set." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.609, + 0.471, + 0.837 + ], + "angle": 0, + "content": "General Setting. While the existing works evaluate methods based on either a single or three random tasks (support sets) [28, 29], we found that, for some datasets, the chosen support samples may not be representative of the class, leading to large standard deviations in low-shot scenarios (see Fig. 1). To ensure fair comparisons, we evaluate all the methods by averaging their classification accuracies over 10 randomly sampled tasks, for each dataset. In all the experiments, we employ ResNet-50 [10] as the visual encoder for the CLIP backbone. It is important to note that, for our BMM procedure in Alg. 1, the optimizer hyper-parameters remain fixed across all the datasets. We use the validation set only to find the best model, via a single run of our BMM procedure with a fixed number of variable updates, i.e., 300 gradient updates including all the blocks of variables." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.185, + 0.862 + ], + "angle": 0, + "content": "3.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We benchmark the proposed \\(\\mathrm{LP}++\\) against relevant state-of-the-art methods in the few-shot adaptation of CLIP-based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.453 + ], + "angle": 0, + "content": "models. We first resort to zero-shot CLIP as the standard baseline, which only leverages the knowledge learned by the pre-trained CLIP model. Also, we include the standard LP baseline, whose implementation is done following [23, 29]. More concretely, this baseline optimizes the standard cross-entropy loss, which corresponds to \\(\\alpha_{k} > 0\\forall k\\) in our generalization in (3), using the L-BFGS [21] optimizer5. It also includes an \\(l_{2}\\)-regularizer, whose balancing weight is set based on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.697 + ], + "angle": 0, + "content": "CLIP-based adapters. We benchmark \\(\\mathrm{LP}++\\) against two popular adapter-based few-shot approaches: CLIP-adapter [9] and TIP-adapter [28]. As exposed earlier, several popular works follow unfair practices by resorting to a larger validation set, or even to the entire test set, to adjust their key hyperparameters -as well as their model selection criteria (i.e., epochs)- for each task. For the sake of fairness, we re-implement Tip-Adapter-F and report the results in 2 different settings. In the first setting (Tip-Adapter-F), we set the two crucial hyper-parameters of this method to 1, keeping them fixed during training, and adopt early stopping based on the validation set. In the second setting, referred to as Tip-Adapter-\\(\\mathbf{F}^*\\), we perform intensive grid-search on the validation set to find the best values for these hyperparameters at initialization, which incurs an additional time complexity burden compared to Tip-Adapter-F." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.772 + ], + "angle": 0, + "content": "Prompt learning. We further compare the proposed LP++ to relevant prompt-learning methods, including CoOp[29] and more recent variants, such as PLOT[4], KgCoOp[27] and ProGrad[30]. We also apply early stopping here based on the performances on the validation set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.779, + 0.593, + 0.793 + ], + "angle": 0, + "content": "3.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.802, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Comparison to the state-of-the-art In Table 1, we present the quantitative results obtained by \\(\\mathrm{LP}++\\) (Alg. 1)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{5} \\) L-BFGS (Limited-memory BFGS) aims to find the minimum of objective function using the second order method. It estimates the Hessian matrix based on recent gradients only, enabling it to determine the steepest direction for achieving the optimal solution. Additionally, it is implemented with line searches to automatically determine the optimal step size." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.091, + 0.379, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.091, + 0.603, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.091, + 0.825, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Figure 3. Quantitative performance of different adaptation methods on the 11 benchmarks (mean), as well as in two other datasets, averaged over 10 tasks (additional figures on the remaining 9 datasets can be found in Appendix, Sec. 16)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.473, + 0.779 + ], + "angle": 0, + "content": "and the relevant literature in the task of efficient adaptation of VLMs. We report the average classification accuracy and standard deviation, across 11 classification benchmarks. From these results, one can make several observations. First, while the standard LP baseline largely underperforms the existing adaptation methods, our improved version, \\(\\mathrm{LP}++\\), brings significant performance gains, particularly in the low-labeled data regimes. It is important to stress that the standard LP baseline integrates only the visual features extracted from CLIP, disregarding the text-encoder knowledge. This contrasts with the existing adapters, which leverage both image and text information. Therefore, these results evidence that the potential of LP has been severely underestimated in the existing literature. Second, when the model selection process is performed fairly, i.e., using a small validation set, their performances fall behind the proposed method (from \\(1\\%\\) to \\(5\\%\\)), despite being arguably more complex approaches. In particular, among the adapter-based strategies, only Tip-Adapter-F* yields performances on par with \\(\\mathrm{LP}++\\), but at the cost of increasing computational load, due to an additional intensive grid-search over its hyper-parameters. If we look at prompt-learning methods, ProGrad and PLOT might be considered as competitors of \\(\\mathrm{LP}++\\), particularly as the number of shots increases. Nevertheless, as already discussed, these methods are computationally inefficient compared to adapters, and do not enable black-box adaptation. The overall performance across all the methods is depicted in Fig. 3 for the average over the 11 benchmarks, as well as in two datasets, showing that \\(\\mathrm{LP}++\\) typically outperforms existing methods under different few-shot scenarios. A more detailed analysis is deferred to the Appendix, Sec. 16." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.904 + ], + "angle": 0, + "content": "Ablation on the loss functions and different optimization strategies. Table 2 reports the test accuracy and run time for different optimizers and loss functions, including the standard CE loss \\((\\alpha_{k} = 0\\forall k)\\) and our loss with learnable blending parameters. Additionally, for \\(\\mathrm{LP + + }\\), we evaluate our loss with fixed blending parameters \\((\\alpha_{k} = 1\\forall k)\\). Independently of the optimizer used, the main takeaway from Table 2 is that introducing the text knowledge and making" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.895, + 0.884 + ], + "angle": 0, + "content": "\\(\\alpha_{k}\\) learnable (rather than fixed) have a substantial impact on accuracy. As for evaluating the optimizers, and for a fair comparison, we use a fixed budget for the number of variable updates for all the optimizers (i.e., 300 updates). In the case of \\(\\mathrm{LP}++\\), this corresponds to the total number of updates for all the blocks. Also, for all optimizers, we initialize \\(\\mathbf{w}\\) and \\(\\alpha\\) following Eqs. (13) and (15). We first consider two popular optimizers, i.e., GD and ADAM, and deploy them in two different settings. First, we run each optimizer 7 times, with each run corresponding to a learning rate in the range \\([10^{-4}, 10^{2}]\\). Then, we record the best performances obtained on the validation set; see GD (optimum) and ADAM (optimum) in Table 2. This follows the standard practices in deep learning, i.e., searching for the learning rates over validation sets, which incurs additional computation overhead; see the time column in Table 2. Second, we run GD and ADAM with our data-driven learning rate, as prescribed by the approximate Lipschitz constant we derived in Eq. (12), with \\(\\tau = 1\\); see GD (our Lipschitz cst) and ADAM (our Lipschitz cst) in Table 2. Note that, in this case, GD corresponds to \\(\\mathrm{LP}++\\) with a single block. Furthermore, we include the L-BFGS in these comparisons, with its initial learning rate set to 1. For L-BFGS, implementing a line search for the optimal step size also introduces an additional computational overhead. As highlighted by Table 2, our method removes the need for validation searches for the optimization hyper-parameters, thanks to its data-driven, task-specific step sizes, thereby reducing the computational load for fine-tuning. In the meanwhile, it yields performances on par with those obtained with the best learning rates found with the validation set. In Fig. 4, we plot the performances of single-block GD vs. the learning rates in the range \\([10^{-4}, 10^{2}]\\), for three datasets. We observe that our Lipschitz-based, task-specific step sizes match the optimal ones found on the validation set, although these vary among the datasets. Also, interestingly, these Lipschitz-based step sizes are orders-of-magnitude larger than those used in deep learning, which are, typically, within interval \\([10^{-4}, 10^{-2}]\\); see [28], for instance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "How to initialize the classifier? To justify empirically the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.09, + 0.377, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.091, + 0.6, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.091, + 0.822, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.893, + 0.281 + ], + "angle": 0, + "content": "Figure 4. Single-block GD performance as a function of different values of the learning rates. The dotted vertical line shows the Lipschitz-based, data-driven learning rate." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.303, + 0.468, + 0.419 + ], + "angle": 0, + "content": "
Optimization MethodStandard loss (αk=0)Our lossTime
1 shot16 shot1 shot16 shot
LP++ (αk is learnt)35.5569.7563.4374.420.78s
LP++ (αk = 1)--46.3769.410.72s
GD (our Lipschitz cst)35.5569.7563.0474.280.86s
GD (optimum)35.6469.6762.9374.556.02s
ADAM (our Lipschitz cst)26.2264.5325.4464.620.98s
ADAM (optimum)35.9069.7363.0774.236.89s
L-BFGS [21]34.5467.4462.0972.826.73s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.429, + 0.471, + 0.486 + ], + "angle": 0, + "content": "Table 2. Accuracy and run time for different optimizers and loss functions (average over 11 datasets). The running time is recorded for our loss (16 shots), and is averaged over 11 datasets. The best result is marked in bold, and the second best is underlined." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.47, + 0.661 + ], + "angle": 0, + "content": "advantages brought by initializing the classifier weights following Eq. (13), and the blending parameters in Eq. (15), we evaluate objective (3) at the beginning of the training, as well as the training-free test accuracy of three methods: random initialization, and the training-free version of both the Tip-adapter-F method [28] (i.e., Tip-adapter) and our method (we provide more details on training-free \\(\\mathrm{LP}++\\) in the appendix). The results in Table 3 confirm empirically the technical observations in Eqs. (13) and (15), which prescribe initial guesses for our problem's variables." + }, + { + "type": "table", + "bbox": [ + 0.096, + 0.671, + 0.452, + 0.773 + ], + "angle": 0, + "content": "
Number of shots (S)124816
ModelsInitial Loss (L0)
- Random Initialization21.4521.4221.3121.2721.32
Proposed Initialization1.601.541.471.351.21
ModelsInitial test accuracy (Acc0)
- Random Initialization17.7917.9218.0318.1018.22
Tip-adapter[28]59.2859.7260.5562.0964.29
Proposed Initialization59.7060.6662.0464.1666.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.784, + 0.471, + 0.827 + ], + "angle": 0, + "content": "Table 3. Comparison of the initial loss and test accuracy in the training-free scenario: random initialization, Tip-adapter and the proposed initialization. The results are averaged over 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Computational overhead. As the literature on adapting VLMs is gaining popularity, it is essential to evaluate the extent to which novel methods are efficient. To do this, we report the overall computational overhead of the approaches" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.459 + ], + "angle": 0, + "content": "studied in this work, which includes the time required for training and for finding the hyper-parameters, when applicable. We also indicate whether these methods enable black-box adaptation which, in our perspective, is a critical aspect in novel strategies aiming to address practical, real-world demands. The numerical values in Table 4 show that, in addition to yielding state-of-the-art performance (shown in previous sections), \\(\\mathrm{LP}++\\) is the most efficient method (by several orders of magnitude), and does not require to access the internal representations of the pre-trained models." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.467, + 0.888, + 0.621 + ], + "angle": 0, + "content": "
MethodsOverall TimeBlackBox# Parameters
CoOp[29]~ 17hXK × M × D
PLOT-2[4]~ 10hXP × K × M × D
KgCoOp[27]~ 4hXK × M × D
ProGrad[30]~ 20hXK × M × D
Clip-Adapter[9]~ 40min✓2(D1 × D)
Tip-adapter-F[28]~ 6min✓K × S × D
Tip-adapter-F*[28]~ 50min✓K × S × D
Standard LP[23]3min✓K × D
LP++~ 2s✓K(D+1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.714 + ], + "angle": 0, + "content": "Table 4. Run time and suitability to black-box scenarios for different methods on 16-shot ImageNet. All the experiments are performed on a single NVIDIA RTX A6000 GPU, except for PLOT-2, which is evaluated on two A6000 GPUs. \\( D_{1} = 256 \\), and \\( D = 1024 \\). The number of context tokens \\( M \\) is set to 16. For PLOT, \\( P = 4 \\) is the number of prompts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.62, + 0.755 + ], + "angle": 0, + "content": "4. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We introduced \\(\\mathrm{LP}++\\), a strong linear probe for few-shot CLIP adaptation. A specific modeling of the classifier weights, blending visual prototypes and text embeddings via learnable multipliers, along with convex-optimization ingredients, often overlooked in deep learning practices, led to the surprising results. While the findings of this work do not invalidate the promise of prompt learning and adaptation research, we believe \\(\\mathrm{LP}++\\) could be used as a baseline to measure progress in these strongly emergent areas." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23780" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Amir Beck and Luba Tetruashvili. On the convergence of block coordinate descent type methods. SIAM Journal on Optimization 23(4), 23(4):2037-2060, 2013. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.213 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[3] Sébastien Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends in Machine Learning, 8(3-4):231-358, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.315 + ], + "angle": 0, + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 2, 6, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[6] Pierre Colombo, Victor Pellegrain, Malik Boudiaf, Victor Storchan, Myriam Tami, Ismail Ben Ayed, Celine Hudelot, and Pablo Piantanida. Transductive learning for textual few-shot classification in api-based embedding models. In Empirical Methods in Natural Language Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[8] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Workshop on Computer Vision and Pattern Recognition, pages 178-178, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, pages 1-15, 2023. 2, 6, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.471, + 0.77 + ], + "angle": 0, + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[12] Steffen Herbold. Autorank: A python package for automated ranking of classifiers. Journal of Open Source Software, 5 (48):2173, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[13] Mingyi Hong, XiangfengWang, Meisam Razaviyayn, and Zhi-Quan Luo. Iteration complexity analysis of block coordinate descent methods. Mathematical Programming, 163: 85–114, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[14] Z hong, D Friedman, and D Chen. Factual probing is [mask]: Learning vs. learning to recall. In Conference of the North" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "American Chapter of the Association for Computational Linguistics, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Z Jiang, F Xu, J Araki, and G Neubig. How can we know what language models know. In Association for Computational Linguistics, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In International Conference on Computer Vision, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] Kenneth Lange, David R Hunter, and Ilsoon Yang. Optimization transfer using surrogate objective functions. Journal of computational and graphical statistics, 9(1):1-20, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[19] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Jorge Nocedal. Updating quasi-newton matrices with limited storage. Mathematics of Computation, 35(151):773-782, 1980. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[22] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 3, 6, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[24] T Shin, Logan R. L. IV Razeghi, Y, E Wallace, and S Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[25] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[26] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3485-3492. IEEE, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[27] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 8, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 1, 2, 5, 6, 7, 8, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6, 8, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.202 + ], + "angle": 0, + "content": "[30] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision, 2023. 2, 6, 8, 7" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23782" + } + ] +] \ No newline at end of file diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_origin.pdf b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c63858c797b253ebe1c9a36455d8c15a942cb0b7 --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/d1daeba3-e67f-49fe-b6c7-0613f97689a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90f9d7038ad20d222bb82d5580e20094273e8687d31ae71d568afd0b317c268c +size 3610228 diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/full.md b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e964d36719daa5f4779999964a491a9e706662db --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/full.md @@ -0,0 +1,291 @@ +# $\mathbf{LP} + +: A$ Surprisingly Strong Linear Probe for Few-Shot CLIP + +Yunshi Huang $^{1*}$ Fereshteh Shakeri $^{1*}$ +Jose Dolz $^{1}$ Malik Boudiaf $^{1}$ Houda Bahig $^{2}$ Ismail Ben Ayed $^{1}$ $^{1}$ ÉTS Montréal, $^{2}$ Université de Montréal + +# Abstract + +In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin $LP++$ , step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, $LP++$ operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git. + +# 1. Introduction + +Recently, there has been a growing popularity in multimodal learning methods, which process and merge information from diverse modalities. In particular, large- + +![](images/2e659f8dd36f94f8c956f7f15aad2f81b3215032cc027fe1e3fd06c7d2b3793e.jpg) +Figure 1. Comparison of $\mathrm{LP}++$ with state-of-the-art few-shot CLIP methods in the 1-shot setting across 11 datasets. We compute the mean accuracy and standard deviation using 10 random tasks for each dataset. The error bars indicate the average standard deviation over all 11 datasets. The x-axis represents the run time for one task, averaged over the 11 datasets. Tip-Adapter-F and Tip-Adapter-F* are two re-implementations of Tip-Adapter-F [28], with fixed and grid-search hyper-parameters, respectively (implementation details provided in Sec. 3.2). + +scale vision-language models (VLMs), such as CLIP [23] and ALIGN [15], have attracted wide attention and made substantial progress in computer vision, showing promising generalization capabilities in various downstream tasks. Unlike conventional task-specific models that are trained with a predetermined set of labels, these so-called foundation models learn to align images with text, in an open-vocabulary fashion. They train, via contrastive learning, vision and text embeddings jointly using a large-scale amount of image-text pairs collected over the internet, thereby leveraging the rich semantic knowledge inherent to language (e.g., concept hierarchies). For a given downstream image classification task, vision-language embeddings enable zero-shot predictions, without re-training, using textual descriptions of the classes (a.k.a prompts). For instance, for a given class $k$ , a textual description of the class, + +which we denote $z_{k}$ , could be "a photo of a $[class_{k}]$ ", where $[class_{k}]$ is the class name. Thus, the zero-shot class prediction for a query image $x$ is obtained from the cosine similarity between the $l_{2}$ -normalized vision-encoded embeddings, $f = \theta_{v}(x)$ , and the text-encoded ones, $t_{k} = \theta_{t}(z_{k})$ : $\hat{k} = \arg \max_{k} f^{t} t_{k}$ , where $t$ denotes the transpose1. + +Motivated by the observation that the choice of input prompts $z_{k}$ may affect the zero-shot predictions, and following on from the strong recent emergence of prompt learning research in the NLP community [14, 16, 24], the popular work in [29] pioneered context optimization (CoOp) for vision-language models. CoOp models input text $z_{k}$ as learnable continuous vectors, e.g., in the form $z_{k} = (z_{k}^{1},\dots,z_{k}^{M},[\mathrm{class}_{k}])$ , where $(z_k^l)_{1\leq l\leq M}$ are learnable text tokens, $[\mathrm{class}_k]$ is a fixed token corresponding to the word embedding vector of the name of the $k^{th}$ class, and $M$ is a hyper-parameter. These learnable vectors are finetuned as task-specific prompts using few-shot training examples and a standard supervised classification loss. More specifically, in this few-shot setting, we assume access to a set consisting of a few labeled samples for each target class, often referred to as the support set. Let $f_{i} = \theta_{v}(\pmb{x}_{i})$ denote the vision embedding of support image $i$ , and $y_{ik}$ its one-hot encoded label, i.e., $y_{ik} = 1$ if image $x_{i}$ belongs to class $k$ and 0 otherwise. Expressing the text embeddings as $t_k = \theta_t(z_k^1,\dots,z_k^M,[\mathrm{class}_k])$ , CoOp fine-tunes text tokens $(z_k^l)_{1\leq l\leq M}$ by minimizing the cross-entropy (CE) loss, with $N$ labeled support samples and $K$ classes: + +$$ +- \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {K} y _ {i k} \ln p _ {i k} \tag {1} +$$ + +where the softmax predictions $p_{ik}$ and the logits (class scores) $l_{ik}$ are given by: + +$$ +p _ {i k} = \frac {\exp (l _ {i k})}{\sum_ {j = 1} ^ {K} \exp (l _ {i j})}; \quad l _ {i k} = \boldsymbol {f} _ {i} ^ {t} \boldsymbol {t} _ {k} +$$ + +Although recent, the pioneering idea of CoOp has triggered a quite abundant literature on prompt learning for few-shot vision-language models, with numerous, more convoluted extensions, e.g. [4, 27, 30], to list a few. For instance, PLOT [4] followed up by learning multiple prompts, to describe the characteristics of each class, via minimizing an optimal-transport distance. KgCoOp [27] improves CoOp's performance when dealing with unseen classes, via minimizing the discrepancy between the text embeddings generated by the learned prompts and hand-crafted ones. While CoOp directly updates the context vectors using the CE loss, ProGrad [30] aligns the few-shot down + +stream knowledge with the large-scale general knowledge, thus mitigating the overfitting of the few-shot samples. + +Prompt learning methods have brought significant improvements over zero-shot classification, but they come at the price of heavy computational and memory load, as they require gradient back-propagation through the entire text encoder. Furthermore, they assume knowledge of the text encoder. These aspects may impede their deployment in low-resource and black-box, privacy-preserving scenarios, which are of wide interest in practice. Indeed, in NLP, there is currently an emerging literature on fast few-shot adaptation of black-box models [6], strongly motivated by the fact that large-scale foundation models (e.g., the GPT family, Anthropic's Claude or Google's PaLM) are only available through APIs and their pre-trained weights are not shared. Finally, by evaluating prompt learning methods over larger numbers of sampled support sets in our experiments (Fig. 1), we observed that they exhibit large variation in performances. This could be explained by the fact that, through the text encoder, they learn prompts that are "too specialized" for a given image support set. + +While prompt learning alters the textual inputs, another category of approaches, referred to as adapters, focused on transforming the pre-training features of the visual or language encoders, e.g., [9, 28]. These adapters are non-linear transformations, for instance, in the form of multi-layer modules, added to the encoder's bottleneck. They learn additional transformations, yielding logits of the form: + +$$ +l _ {i k} = \boldsymbol {\theta} _ {a} \left(\boldsymbol {f} _ {i}, \boldsymbol {t} _ {k}\right) \tag {2} +$$ + +The adapter's learnable parameters, $\theta_{a}$ , are fine-tuned over a few-shot task by optimizing the cross-entropy loss, similarly to (1) but with logits $l_{ik}$ expressed as functions of $\theta_{a}$ . For instance, the popular CLIP-Adapter [9] integrated a multi-layered perceptron to modify the features, along with residual connections, which enable blending with the original pre-trained features. Tip-Adapter [28] added a nonlinear, quadratic-complexity module, which evaluates the pairwise similarities between the features of the support sets, and blends the ensuing class scores with the original textual features. This category of approaches mitigates some of the limitations of prompt-learning methods as they result in few-shot adaptation that has significantly lower computation and memory loads. However, as shown in Fig. 1 and in our experiments, their performances seem to depend strongly on some key hyper-parameters that have to be adjusted carefully on each downstream task, e.g. those that control the blending between the vision and language features. Therefore, to perform competitively, they incur an additional computation overhead to the adaptation phase, due to intensive (e.g., grid) search of the hyper-parameters over task-dedicated validation sets. + +In the above-mentioned, strongly emergent literature on + +few-shot CLIP adaptation, linear probe (LP) [23] has been often reported as a very weak baseline. For instance, in the 1-shot setting, it scores near $20\%$ lower than the zero-shot predictions averaged over 11 benchmarks (Table 1). Initially evaluated in [23], LP is a linear classifier on the vision-encoded features. Specifically, it optimizes the CE loss (1) w.r.t the last-layer weights of the vision encoder (i.e., the class prototypes), which we will denote $(\pmb{w}_k)_{1\leq k\leq K}$ in the rest of the paper, with the logits given by: $l_{ik} = \pmb{f}_i^t\pmb{w}_k$ . A clear deficiency in this standard LP baseline is that it omits completely the language knowledge of CLIP, i.e., $(\pmb{t}_k)_{1\leq k\leq K}$ . + +In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline. Specifically, we extend the logits in the CE loss in (1), so that they become learnable functions of the text embedding: + +$$ +l _ {i k} = \boldsymbol {f} _ {i} ^ {t} (\boldsymbol {w} _ {k} + \alpha_ {k} \boldsymbol {t} _ {k}) +$$ + +with $(\alpha_{k})_{1\leq k\leq K}$ trainable class-wise parameters blending image and text knowledge. As our objective function depends on two types of variables, i.e., the visual class prototypes $(\pmb{w}_k)_{1\leq k\leq K}$ and blending parameters $(\alpha_{k})_{1\leq k\leq K}$ , we propose a computationally efficient Block Majorize-Minimize (BMM) procedure. In our full-batch MM optimizer, which we coin $\mathrm{LP}++$ , step sizes are implicit in the definition of the majorizing functions, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. Moreover, we examine the mathematical properties of our objective, i.e., (i) Lipschitz gradient continuity and (ii) decomposition into convex functions having closed-form optima. This enabled us to build majorizing functions yielding data-driven learning rates, and to derive approximations of the objective-function minima, which yield data-informed initializations of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yield, surprisingly, highly competitive few-shot CLIP performances (Fig. 1). Furthermore, $\mathrm{LP}++$ operates in black-box, relaxes intensive validation searches for the optimization of hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP methods (Table 4). For instance, for 16-shot ImageNet adaptation, it takes seconds on a single NVIDIA RTX A600 GPU. + +# 2. Formulation of $\mathbf{LP} + +$ + +Following the notations introduced in the previous section, we propose to minimize the following CE objective function w.r.t visual class prototypes $\mathbf{w} = (\pmb{w}_k)_{1\leq k\leq K}$ and class-wise blending parameters $\alpha = (\alpha_{k})_{1\leq k\leq K}$ : + +$$ +L (\mathbf {w}, \boldsymbol {\alpha}) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {k = 1} ^ {K} y _ {i k} \ln p _ {i k} (\mathbf {w}, \boldsymbol {\alpha}) \tag {3} +$$ + +![](images/b89ac625c0fbad8bc8a8033039281ad74331c693ebe7010aba59edd154d9914f.jpg) +Figure 2. Visualization of $\mathbf{LP} + +$ + +where softmax probability outputs $p_{ik}$ are now given by: + +$$ +p _ {i k} (\mathbf {w}, \boldsymbol {\alpha}) = \frac {\exp \left(\boldsymbol {f} _ {i} ^ {t} \left(\boldsymbol {w} _ {k} + \alpha_ {k} \boldsymbol {t} _ {k}\right)\right)}{\sum_ {j = 1} ^ {K} \exp \left(\boldsymbol {f} _ {i} ^ {t} \left(\boldsymbol {w} _ {j} + \alpha_ {j} \boldsymbol {t} _ {j}\right)\right)} \tag {4} +$$ + +and $t_k$ are the fixed pre-trained embeddings of the text templates used in zero-shot CLIP [23]. Clearly, the objective function defined by (3) and (4) could be viewed as a generalization of the vision-encoder CE loss used in the standard LP [23]. Indeed, the latter corresponds to setting $\alpha_k = 0 \forall k$ , i.e., no text knowledge. As we will see in our experimental ablation over different objective functions (Table 2), introducing the text knowledge $(\alpha_k > 0 \forall k)$ has a substantial effect on performances. Also, making $\alpha_k$ learnable (rather than fixed) leads to a further significant impact. Indeed, we hypothesize that the optimal blending of the text and visual knowledge is task dependent, which motivates learning it from the context of the support set. + +# 2.1. Block coordinate Majorize-Minimize descent + +Majorize-Minimize (MM) [18] is a very general optimization principle, which includes different classes of standard optimizers such as gradient descent, concave-convex procedures and expectation-maximization. Let $\mathbf{v} = (\mathbf{w},\alpha)\in$ $\mathbb{R}^{K(D + 1)}$ denote the overall vector of variables in our case, with $D$ being the dimension of the feature embeddings. At each iteration, the MM procedure updates the variable as the minimum of a majorizing function, i.e., an upper bound on the original objective, which is tight at the current iteration $j\colon L(\mathbf{v})\leq M(\mathbf{v},\mathbf{v}^j)$ and $L(\mathbf{v}^j) = M(\mathbf{v}^j,\mathbf{v}^j)$ . Thus, update step $\mathbf{v}^{j + 1} = \min_{\mathbf{v}}M(\mathbf{v},\mathbf{v}^j)$ guarantees that the original objective does not increase at each iteration: $L(\mathbf{v}^{j + 1})\leq M(\mathbf{v}^{j + 1},\mathbf{v}^j)\leq M(\mathbf{v}^j,\mathbf{v}^j) = L(\mathbf{v}^j)$ . Therefore, in MM algorithms, step sizes are implicit in the definition of the majorizing function, unlike standard gradient-descent practices, in which the step sizes (a.k.a learning rates) are intensively searched over validation sets, via running the optimizer several times. + +In this work, we exploit the Lipschitz-gradient continuity of our convex objective in (3), i.e., bounds on the + +maximum eigen values of the Hessian matrices (Prop. 1), thereby building majorizing functions with data-driven, task-specific step sizes. This removes the need for validation searches for the optimization hyper-parameters, reducing the computational load for fine-tuning (Table 2), while yielding performances on par with the best learning rates found with validation (Fig. 4). Also, interestingly, the Lipschitz-based learning rates computed from our derivation in Prop. 1 are orders-of-magnitude larger than those typically used in deep learning, yielding steeper decreases towards the minimum. Before giving proper majorizing functions for our convex, gradient-Lipschitz function in (3), let us first point to the following results, well-known in convex optimization [3]. While these results are text-book knowledge in optimization, they enable to connect the general MM principle to gradient descent, motivating the data-driven, task-specific step sizes we derive in Prop. 1 and the block-coordinate MM optimizer we propose in Alg. 1. + +Lemma 2.1. ([3, p. 268]) Assume $L(\mathbf{v})$ is a twice-differentiable function, which has a Lipschitz continuous gradient, i.e., there exists a strictly positive Lipschitz constant $\gamma$ such that $\nabla^2 L(\mathbf{v}) \preceq \gamma \mathbf{I}$ , with $\mathbf{I}$ the identity matrix. Then, the following quadratic bound is a majorizing function for $L$ at iteration $j$ : + +$$ +M (\mathbf {v}, \mathbf {v} ^ {j}) = L (\mathbf {v} ^ {j}) + \nabla L (\mathbf {v} ^ {j}) ^ {t} (\mathbf {v} - \mathbf {v} ^ {j}) + \frac {\gamma}{2} \| \mathbf {v} - \mathbf {v} ^ {j} \| ^ {2} \tag {5} +$$ + +Furthermore, a specific gradient step, with learning rate $\frac{1}{\gamma}$ minimizes $M$ , i.e., $\mathbf{v}^{j + 1} = \mathbf{v}^j -\frac{1}{\gamma}\nabla L(\mathbf{v}^j)$ = $\arg \min_{\mathbf{v}}M(\mathbf{v},\mathbf{v}^j)$ , and guarantees that objective $L$ decreases by at least $\frac{1}{2\gamma}\| \nabla L(\mathbf{v})\| ^2$ : + +$$ +L \left(\mathbf {v} ^ {j + 1}\right) \leq L \left(\mathbf {v} ^ {j}\right) - \frac {1}{2 \gamma} \| \nabla L (\mathbf {v}) \| ^ {2} \tag {6} +$$ + +Moreover, the following Theorem, which follows from Lemma 2.1, establishes the sublinear convergence of the MM procedure using bound (5), i.e., a convergence rate of $O(1 / J)$ , $J$ being the total number of iterations. + +Theorem 2.2. ([3, p. 267]) For convex, twice-differentiable function $L(\mathbf{v})$ , which has a $\gamma$ -Lipschitz gradient, performing $J$ updates $\mathbf{v}^{j+1} = \mathbf{v}^j - \frac{1}{\gamma} \nabla L(\mathbf{v}^j)$ , starting from initialization $\mathbf{v}^0$ , will yield a solution that satisfies: + +$$ +\left\| L \left(\mathbf {v} ^ {J}\right) - L \left(\mathbf {v} ^ {*}\right) \right\| \leq \frac {\gamma}{2 J} \left\| \mathbf {v} ^ {0} - \mathbf {v} ^ {*} \right\| \tag {7} +$$ + +where $L(\mathbf{v}^{*})$ is the optimal value. + +For completeness, we provide the proofs of these well-known results in the supplemental material. Clearly, Lemma 2.1 and Theorem 2.2 prescribe a learning rate of $\frac{1}{\gamma}$ for a function that has a $\gamma$ -Lipschitz gradient. One valid Lipschitz constant would be the maximum eigen value of + +the Hessian of our objective in (3), which provides a majorizing function of the form in Eq. (5) and data-driven learning rates. However, a naive spectral decomposition of the Hessian matrices (to obtain the maximum eigen value) could be computationally intensive. For instance, for ImageNet, the Hessian of our objective is of size $K(D + 1) \times K(D + 1) \approx 1M \times 1M$ , as $D = 1024$ and $K = 1000$ . In Prop. 1, we derive approximate global and block-wise Lipschitz constants that can be computed efficiently (i.e., evaluating the maximum eigen value of a single $D \times D$ matrix). + +Block-coordinate updates Our procedure provided in Alg. 1 belongs to the family of Block Majorize-Minimize (BMM) methods, well studied in the optimization community [13]. To minimize a multi-block objective, as in our case where the blocks correspond to variables $\mathbf{w}$ and $\alpha$ , we minimize one or many successive majorizing functions of the objective in each block, with the other block fixed, in a cyclic order: + +$$ +L \left(\mathbf {w} ^ {j}, \boldsymbol {\alpha}\right) + \nabla L _ {\mathbf {w}} \left(\mathbf {w} ^ {j}\right) ^ {t} \left(\mathbf {w} - \mathbf {w} ^ {j}\right) + \frac {\gamma_ {\mathbf {w}}}{2} \| \mathbf {w} - \mathbf {w} ^ {j} \| ^ {2} \tag {8} +$$ + +$$ +L \left(\mathbf {w}, \boldsymbol {\alpha} ^ {j}\right) + \nabla L _ {\boldsymbol {\alpha}} \left(\boldsymbol {\alpha} ^ {j}\right) ^ {t} \left(\boldsymbol {\alpha} - \boldsymbol {\alpha} ^ {j}\right) + \frac {\gamma_ {\boldsymbol {\alpha}}}{2} \| \boldsymbol {\alpha} - \boldsymbol {\alpha} ^ {j} \| ^ {2} \tag {9} +$$ + +where in (8), block $\alpha$ is fixed and, in (9), $\mathbf{w}$ is fixed. $\nabla L_{\mathbf{w}}$ and $\nabla L_{\alpha}$ denote block-wise gradients, and $(\gamma_{\mathbf{w}},\gamma_{\alpha})$ are the block Lipschitz constants. Accommodating different choices of the block-cycling strategies and majorizing functions, BMM includes a breadth of optimizers as particular cases, such the well-known block coordinate gradient descent (BCGD) [1] and its projection-based variant. Indeed, the so-called Gauss-Seidel cycling [13] alternates steps (8) and (9), which corresponds to the BCGD method. One could also perform many successive steps in one block, as we do in Alg. 1, which corresponds to the so-called Essentially-Cyclic4 strategy [13]. Importantly, for a fairly large spectrum of choices of the cycling strategies, BMM enjoys the same sublinear convergence property as MM for convex objectives (with each block-wise update decreasing the objective), provided that the majorizing functions are strongly convex; see Theorem 3.1 in [13]. This is the case for (8) and (9). In our experiments, we observed that this block-wise variant performs better than a single-block MM; see Table 2. This might be explained by the fact that each block of variables has a dedicated step size. In the supplemental material, we provide results for different block-cycling strategies. + +Global and block-coordinatewise Lipschitz constants In the following, we derive approximate global and block- + +coordinatewise Lipschitz constants for our objective function in (3), which could be evaluated efficiently. We deploy these in Alg. 1, to compute data-driven, block-wise learning rates for updating visual prototypes $\mathbf{w}$ and blending parameters $\alpha$ . + +Proposition 1. Considering the blocks of variables $\mathbf{w}$ and $\alpha$ , the gradient of our objective $L$ in (3) is block-coordinatewise Lipschitz continuous. For $\tau_{1} \geq 2$ , it has the following block Lipschitz constant for the set of variables in $\mathbf{w}$ : + +$$ +\gamma_ {\mathbf {w}} = \frac {\tau_ {1}}{4 N} \lambda_ {m a x} \left(\sum_ {i = 1} ^ {N} \left(\boldsymbol {f} _ {i} \boldsymbol {f} _ {i} ^ {t}\right)\right) \tag {10} +$$ + +where $\lambda_{max}(\mathbf{A})$ denotes the maximum eigenvalue of matrix $\mathbf{A}$ . Furthermore, for $\tau_{1} \geq 1$ , the expression in Eq. (10) provides a tighter but approximate block Lipschitz constant. Similarly, for $\tau_{2} \geq 1$ , we have following approximate block Lipschitz constant for the variables in $\alpha$ : + +$$ +\gamma_ {\boldsymbol {\alpha}} = \max _ {k} \frac {\tau_ {2}}{4 N} \sum_ {i = 1} ^ {N} \left(\boldsymbol {f} _ {i} ^ {t} \boldsymbol {t} _ {k}\right) ^ {2} \tag {11} +$$ + +Finally, for $\tau \geq 2$ , the following expression provides an approximate global Lipschitz constant for objective (3), i.e., w.r.t all variables $\mathbf{v} = (\mathbf{w},\alpha)$ : + +$$ +\gamma = \tau \max \left(\gamma_ {\mathbf {w}}, \gamma_ {\boldsymbol {\alpha}}\right) \tag {12} +$$ + +Proof. The details are deferred to the supplemental material. The main ingredients of the proof are based on the Gershgorin circle theorem and the variational characterization of the maximum eigenvalue, following the min-max theorem, also referred to as the variational principle. $\square$ + +# 2.2. Initialization of the variables + +In the following Prop. 2, we derive approximations of the minima of our objective function (3), which yield a data-informed initialization of the variables. Indeed, the expressions we obtain in Eqs. (13) and (15) suggest initial guesses for variables $\mathbf{w}$ and $\alpha$ . Interestingly, and as will be confirmed by our experiments (Table 3), such an initialization yields substantially lower values of the minimized loss than a random initialization. Furthermore, surprisingly, using this initialization for a training-free prediction yields better performances than the training-free version of the recent Tip-Adapter-F approach [28]. The details of our training-free version are provided in the supplemental material. + +Proposition 2. The cross-entropy in Eq. (3) could be written as the sum of two convex functions, i.e., $L = g_{1} + g_{2}$ , such that, $\forall k \in [1, \dots, K]$ , the minimum of $g_{1}$ w.r.t $\mathbf{w}_{k}$ is co-linear to the hard mean vector of features within class $k$ : + +$$ +\underset {\boldsymbol {w} _ {k}} {\arg \min } g _ {1} = \frac {1}{\lambda N} \sum_ {i = 1} ^ {N} y _ {i k} \boldsymbol {f} _ {i} \propto \frac {\sum_ {i = 1} ^ {N} y _ {i k} \boldsymbol {f} _ {i}}{\sum_ {i = 1} ^ {N} y _ {i k}} \tag {13} +$$ + +and the minimum of $g_{2}$ w.r.t $\pmb{w}_{k}$ is co-linear to the soft mean vector of features within class $k$ : + +$$ +\underset {\boldsymbol {w} _ {k}} {\arg \min } g _ {2} = \frac {1}{\lambda N} \sum_ {i = 1} ^ {N} p _ {i k} \boldsymbol {f} _ {i} \propto \frac {\sum_ {i = 1} ^ {N} p _ {i k} \boldsymbol {f} _ {i}}{\sum_ {i = 1} ^ {N} p _ {i k}} \tag {14} +$$ + +where $\lambda \leq \min_k\lambda_{min}(\pmb {A}_k)$ $A_{k} = \frac{1}{N}\sum_{i = 1}^{N}(p_{ik} - p_{ik}^{2})\pmb{f}_{i}\pmb{f}_{i}^{t}$ and $\lambda_{min}(\mathbf{A})$ denotes the smallest eigenvalue of matrix $\mathbf{A}$ + +Proof. We defer the details, including the full expressions of convex functions $g_{1}$ and $g_{2}$ , to the supplemental material. + +![](images/37c5aa575f31b844d941e9a66dc18a54ac3c1ec240bc5fadac45a09a9ce22693.jpg) + +Similarly to the development in Prop. 2, one could decompose (3) as the sum of two convex functions, i.e., $L = h_1 + h_2$ , such that, $\forall k \in [1, \dots, K]$ , the minima of $h_1$ and $h_2$ w.r.t $\alpha_k$ could be written, up to a multiplicative positive factor, as the hard and soft means of the cosine similarities between the image and text embeddings: + +$$ +\underset {\alpha_ {k}} {\arg \min } h _ {1} = \frac {1}{\beta N} \sum_ {i = 1} ^ {N} y _ {i k} \boldsymbol {f} _ {i} ^ {t} \boldsymbol {t} _ {k} \tag {15} +$$ + +$$ +\underset {\alpha_ {k}} {\arg \min } h _ {2} = \frac {1}{\beta N} \sum_ {i = 1} ^ {N} p _ {i k} \boldsymbol {f} _ {i} ^ {t} \boldsymbol {t} _ {k} \tag {16} +$$ + +Here, $\beta = \min_k\frac{1}{N}\sum_{i = 1}^N (p_{ik} - p_{ik}^2)(\pmb {f}_i^t\pmb {t}_k)^2$ and $h_1$ (respectively $h_2$ ) has the same expression as $g_{2}$ (respectively $g_{1}$ ), except that term $\frac{\lambda}{2}\sum_{k = 1}^{K}\| \pmb {w}_k\|^2$ is replaced by $\frac{\beta}{2}\|\pmb {\alpha}\|^2$ ; see the supplemental material for the expressions of $g_{1}$ and $g_{2}$ . + +Algorithm 1: Block coordinate MM $(\mathbf{w},\alpha)$ +iterw=10;iter $\alpha = 1$ . $\tau_{1} = 1$ . $\tau_{2} = 16$ . $\lambda = \frac{1}{N}$ $\beta = \frac{1}{250K}$ Initialize w0,0 // Using (13) for each k Initialize $\alpha^{0,0}$ // Using (15) for each k for $j = 0,1,\ldots$ do for $l_{1} = 0,1,\ldots$ ,iterw do $\begin{array}{r}\mathbf{w}^{j,l_1 + 1} = \mathbf{w}^{j,l_1} - \frac{1}{\gamma_{\mathbf{w}}}\nabla L_{\mathbf{w}}(\mathbf{w}^{j,l_1},\boldsymbol{\alpha}^{j,0})\\ /\quad \frac{1}{\gamma_{\mathbf{w}}}\text{from Eq.}(10) \end{array}$ $\mathbf{w}^{j + 1,0} = \mathbf{w}^{j,\mathrm{iter}_{\mathbf{w}}}$ for $l_{2} = 0,1,\dots$ iter $\alpha$ do $\begin{array}{rl} & {\boldsymbol{\alpha}^{j,l_2 + 1} = \boldsymbol{\alpha}^{j,l_2} - \frac{1}{\gamma_\alpha}\nabla L_\alpha (\mathbf{w}^{j + 1,0},\boldsymbol{\alpha}^{j,l_2})}\\ & {/ / \frac{1}{\gamma_\alpha}\text{from Eq.}(11)} \end{array}$ $\alpha^{j + 1,0} = \alpha^{j,\mathrm{iter}_{\alpha}}$ + +
Number of shots (S)124816
Zero-shot CLIP ICML'21[23]58.89
Prompt-LearningCoOp ICV'22[29]59.62 ± 3.1163.80 ± 2.3267.23 ± 1.6471.30 ± 0.8674.06 ± 0.55
PLOTICLR'23[4]61.51 ± 2.9165.67 ± 2.0668.39 ± 1.1771.96 ± 0.7074.35 ± 0.66
KgCoOp CVPR'23[27]61.36 ± 3.0463.23 ± 2.0665.73 ± 1.1567.50 ± 1.1169.01 ± 0.79
ProGrad ICCV'23[30]62.46 ± 1.8965.88 ± 1.4668.52 ± 1.1571.82 ± 0.1173.95 ± 0.68
CLIP-based AdaptersCLIP-Adapter ICV'23[9]60.32 ± 0.8061.93 ± 0.9365.12 ± 0.8069.20 ± 0.5672.57 ± 0.54
Tip-Adapter-F ECCV'22[28]61.29 ± 0.9262.94 ± 0.7566.02 ± 0.8069.88 ± 0.5173.82 ± 0.55
Tip-Adapter-F* ECCV'22[28]63.06 ± 1.0566.47 ± 0.6568.71 ± 0.9671.78 ± 1.0074.37 ± 0.35
Linear-ProbingStandard LP ICML'21[23]36.10 ± 1.4346.99 ± 1.2956.72 ± 1.2064.66 ± 0.5570.56 ± 0.44
LP++63.43 ± 0.9066.20 ± 0.7269.16 ± 0.7972.04 ± 0.4674.42 ± 0.45
+ +Table 1. Comparison to state-of-the-art methods. Average classification accuracy (%) on 11 benchmarks, with standard derivation over 10 sampled support sets for each dataset. The best values are highlighted in bold. + +# 3. Experiments + +# 3.1. Datasets and Implementation details + +Following the CLIP-based few-shot adaptation literature [27, 28], we conduct the main experiments on 11 public classification data sets: Caltech101 [8], ImageNet [7], DTD [5], OxfordPets [22], Flowers102 [20], StandfordCars [17], Food101 [2], FGVCAircraft [19], SUN397 [26], EuroSAT [11] and UCF101 [25]. We follow standard practices [23] and consider $S = \{1, 2, 4, 8, 16\}$ shots for model adaptation, which are randomly sampled for each data set. + +Towards a fair validation set. Apparently, prior works on this problem [28] have resorted to a large set of validation samples to adjust their hyper-parameters. For the sake of fairness, we tune the hyper-parameters across all the methods based on a small validation set, which contains as many samples (i.e., shots) as the training set. Furthermore, to avoid the potential overfitting on the few training samples, we adopt an early stopping strategy on this validation set. + +General Setting. While the existing works evaluate methods based on either a single or three random tasks (support sets) [28, 29], we found that, for some datasets, the chosen support samples may not be representative of the class, leading to large standard deviations in low-shot scenarios (see Fig. 1). To ensure fair comparisons, we evaluate all the methods by averaging their classification accuracies over 10 randomly sampled tasks, for each dataset. In all the experiments, we employ ResNet-50 [10] as the visual encoder for the CLIP backbone. It is important to note that, for our BMM procedure in Alg. 1, the optimizer hyper-parameters remain fixed across all the datasets. We use the validation set only to find the best model, via a single run of our BMM procedure with a fixed number of variable updates, i.e., 300 gradient updates including all the blocks of variables. + +# 3.2. Baselines + +We benchmark the proposed $\mathrm{LP}++$ against relevant state-of-the-art methods in the few-shot adaptation of CLIP-based + +models. We first resort to zero-shot CLIP as the standard baseline, which only leverages the knowledge learned by the pre-trained CLIP model. Also, we include the standard LP baseline, whose implementation is done following [23, 29]. More concretely, this baseline optimizes the standard cross-entropy loss, which corresponds to $\alpha_{k} > 0\forall k$ in our generalization in (3), using the L-BFGS [21] optimizer5. It also includes an $l_{2}$ -regularizer, whose balancing weight is set based on the validation set. + +CLIP-based adapters. We benchmark $\mathrm{LP}++$ against two popular adapter-based few-shot approaches: CLIP-adapter [9] and TIP-adapter [28]. As exposed earlier, several popular works follow unfair practices by resorting to a larger validation set, or even to the entire test set, to adjust their key hyperparameters -as well as their model selection criteria (i.e., epochs)- for each task. For the sake of fairness, we re-implement Tip-Adapter-F and report the results in 2 different settings. In the first setting (Tip-Adapter-F), we set the two crucial hyper-parameters of this method to 1, keeping them fixed during training, and adopt early stopping based on the validation set. In the second setting, referred to as Tip-Adapter- $\mathbf{F}^*$ , we perform intensive grid-search on the validation set to find the best values for these hyperparameters at initialization, which incurs an additional time complexity burden compared to Tip-Adapter-F. + +Prompt learning. We further compare the proposed LP++ to relevant prompt-learning methods, including CoOp[29] and more recent variants, such as PLOT[4], KgCoOp[27] and ProGrad[30]. We also apply early stopping here based on the performances on the validation set. + +# 3.3. Results + +Comparison to the state-of-the-art In Table 1, we present the quantitative results obtained by $\mathrm{LP}++$ (Alg. 1) + +![](images/399c52c91fb727f8824f3c4ac338827d397517cc4e2d198e3da8e65a58be093f.jpg) +Figure 3. Quantitative performance of different adaptation methods on the 11 benchmarks (mean), as well as in two other datasets, averaged over 10 tasks (additional figures on the remaining 9 datasets can be found in Appendix, Sec. 16). + +![](images/97406d7969a0fb71137bbe71d34d449e1344a43d80a2b3b0332b491e4414df72.jpg) + +![](images/d86fd7af33d3477ccc718f4954a040d2e8e6d99706c01450054ac865990d2dad.jpg) + +and the relevant literature in the task of efficient adaptation of VLMs. We report the average classification accuracy and standard deviation, across 11 classification benchmarks. From these results, one can make several observations. First, while the standard LP baseline largely underperforms the existing adaptation methods, our improved version, $\mathrm{LP}++$ , brings significant performance gains, particularly in the low-labeled data regimes. It is important to stress that the standard LP baseline integrates only the visual features extracted from CLIP, disregarding the text-encoder knowledge. This contrasts with the existing adapters, which leverage both image and text information. Therefore, these results evidence that the potential of LP has been severely underestimated in the existing literature. Second, when the model selection process is performed fairly, i.e., using a small validation set, their performances fall behind the proposed method (from $1\%$ to $5\%$ ), despite being arguably more complex approaches. In particular, among the adapter-based strategies, only Tip-Adapter-F* yields performances on par with $\mathrm{LP}++$ , but at the cost of increasing computational load, due to an additional intensive grid-search over its hyper-parameters. If we look at prompt-learning methods, ProGrad and PLOT might be considered as competitors of $\mathrm{LP}++$ , particularly as the number of shots increases. Nevertheless, as already discussed, these methods are computationally inefficient compared to adapters, and do not enable black-box adaptation. The overall performance across all the methods is depicted in Fig. 3 for the average over the 11 benchmarks, as well as in two datasets, showing that $\mathrm{LP}++$ typically outperforms existing methods under different few-shot scenarios. A more detailed analysis is deferred to the Appendix, Sec. 16. + +Ablation on the loss functions and different optimization strategies. Table 2 reports the test accuracy and run time for different optimizers and loss functions, including the standard CE loss $(\alpha_{k} = 0\forall k)$ and our loss with learnable blending parameters. Additionally, for $\mathrm{LP + + }$ , we evaluate our loss with fixed blending parameters $(\alpha_{k} = 1\forall k)$ . Independently of the optimizer used, the main takeaway from Table 2 is that introducing the text knowledge and making + +$\alpha_{k}$ learnable (rather than fixed) have a substantial impact on accuracy. As for evaluating the optimizers, and for a fair comparison, we use a fixed budget for the number of variable updates for all the optimizers (i.e., 300 updates). In the case of $\mathrm{LP}++$ , this corresponds to the total number of updates for all the blocks. Also, for all optimizers, we initialize $\mathbf{w}$ and $\alpha$ following Eqs. (13) and (15). We first consider two popular optimizers, i.e., GD and ADAM, and deploy them in two different settings. First, we run each optimizer 7 times, with each run corresponding to a learning rate in the range $[10^{-4}, 10^{2}]$ . Then, we record the best performances obtained on the validation set; see GD (optimum) and ADAM (optimum) in Table 2. This follows the standard practices in deep learning, i.e., searching for the learning rates over validation sets, which incurs additional computation overhead; see the time column in Table 2. Second, we run GD and ADAM with our data-driven learning rate, as prescribed by the approximate Lipschitz constant we derived in Eq. (12), with $\tau = 1$ ; see GD (our Lipschitz cst) and ADAM (our Lipschitz cst) in Table 2. Note that, in this case, GD corresponds to $\mathrm{LP}++$ with a single block. Furthermore, we include the L-BFGS in these comparisons, with its initial learning rate set to 1. For L-BFGS, implementing a line search for the optimal step size also introduces an additional computational overhead. As highlighted by Table 2, our method removes the need for validation searches for the optimization hyper-parameters, thanks to its data-driven, task-specific step sizes, thereby reducing the computational load for fine-tuning. In the meanwhile, it yields performances on par with those obtained with the best learning rates found with the validation set. In Fig. 4, we plot the performances of single-block GD vs. the learning rates in the range $[10^{-4}, 10^{2}]$ , for three datasets. We observe that our Lipschitz-based, task-specific step sizes match the optimal ones found on the validation set, although these vary among the datasets. Also, interestingly, these Lipschitz-based step sizes are orders-of-magnitude larger than those used in deep learning, which are, typically, within interval $[10^{-4}, 10^{-2}]$ ; see [28], for instance. + +How to initialize the classifier? To justify empirically the + +![](images/981378ca95952e9ba244a68193192d2f88625889f51da89ab839ffd6df107efd.jpg) +Figure 4. Single-block GD performance as a function of different values of the learning rates. The dotted vertical line shows the Lipschitz-based, data-driven learning rate. + +![](images/e30e12bebdb58dc8c0958692c3af848dcec042dc52c580455523bf65982219bc.jpg) + +![](images/1545f4288f93bb4dc4292c0f3a22e4254bb4d62fe49218e0f91909f2fa26924c.jpg) + +
Optimization MethodStandard loss (αk=0)Our lossTime
1 shot16 shot1 shot16 shot
LP++ (αk is learnt)35.5569.7563.4374.420.78s
LP++ (αk = 1)--46.3769.410.72s
GD (our Lipschitz cst)35.5569.7563.0474.280.86s
GD (optimum)35.6469.6762.9374.556.02s
ADAM (our Lipschitz cst)26.2264.5325.4464.620.98s
ADAM (optimum)35.9069.7363.0774.236.89s
L-BFGS [21]34.5467.4462.0972.826.73s
+ +advantages brought by initializing the classifier weights following Eq. (13), and the blending parameters in Eq. (15), we evaluate objective (3) at the beginning of the training, as well as the training-free test accuracy of three methods: random initialization, and the training-free version of both the Tip-adapter-F method [28] (i.e., Tip-adapter) and our method (we provide more details on training-free $\mathrm{LP}++$ in the appendix). The results in Table 3 confirm empirically the technical observations in Eqs. (13) and (15), which prescribe initial guesses for our problem's variables. + +Table 2. Accuracy and run time for different optimizers and loss functions (average over 11 datasets). The running time is recorded for our loss (16 shots), and is averaged over 11 datasets. The best result is marked in bold, and the second best is underlined. + +
Number of shots (S)124816
ModelsInitial Loss (L0)
- Random Initialization21.4521.4221.3121.2721.32
Proposed Initialization1.601.541.471.351.21
ModelsInitial test accuracy (Acc0)
- Random Initialization17.7917.9218.0318.1018.22
Tip-adapter[28]59.2859.7260.5562.0964.29
Proposed Initialization59.7060.6662.0464.1666.20
+ +Computational overhead. As the literature on adapting VLMs is gaining popularity, it is essential to evaluate the extent to which novel methods are efficient. To do this, we report the overall computational overhead of the approaches + +studied in this work, which includes the time required for training and for finding the hyper-parameters, when applicable. We also indicate whether these methods enable black-box adaptation which, in our perspective, is a critical aspect in novel strategies aiming to address practical, real-world demands. The numerical values in Table 4 show that, in addition to yielding state-of-the-art performance (shown in previous sections), $\mathrm{LP}++$ is the most efficient method (by several orders of magnitude), and does not require to access the internal representations of the pre-trained models. + +Table 3. Comparison of the initial loss and test accuracy in the training-free scenario: random initialization, Tip-adapter and the proposed initialization. The results are averaged over 11 datasets. + +
MethodsOverall TimeBlackBox# Parameters
CoOp[29]~ 17hXK × M × D
PLOT-2[4]~ 10hXP × K × M × D
KgCoOp[27]~ 4hXK × M × D
ProGrad[30]~ 20hXK × M × D
Clip-Adapter[9]~ 40min✓2(D1 × D)
Tip-adapter-F[28]~ 6min✓K × S × D
Tip-adapter-F*[28]~ 50min✓K × S × D
Standard LP[23]3min✓K × D
LP++~ 2s✓K(D+1)
+ +Table 4. Run time and suitability to black-box scenarios for different methods on 16-shot ImageNet. All the experiments are performed on a single NVIDIA RTX A6000 GPU, except for PLOT-2, which is evaluated on two A6000 GPUs. $D_{1} = 256$ , and $D = 1024$ . The number of context tokens $M$ is set to 16. For PLOT, $P = 4$ is the number of prompts. + +# 4. Conclusion + +We introduced $\mathrm{LP}++$ , a strong linear probe for few-shot CLIP adaptation. A specific modeling of the classifier weights, blending visual prototypes and text embeddings via learnable multipliers, along with convex-optimization ingredients, often overlooked in deep learning practices, led to the surprising results. While the findings of this work do not invalidate the promise of prompt learning and adaptation research, we believe $\mathrm{LP}++$ could be used as a baseline to measure progress in these strongly emergent areas. + +# References + +[1] Amir Beck and Luba Tetruashvili. On the convergence of block coordinate descent type methods. SIAM Journal on Optimization 23(4), 23(4):2037-2060, 2013. 4, 2 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 6 +[3] Sébastien Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends in Machine Learning, 8(3-4):231-358, 2015. 4 +[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 2, 6, 8, 7 +[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6 +[6] Pierre Colombo, Victor Pellegrain, Malik Boudiaf, Victor Storchan, Myriam Tami, Ismail Ben Ayed, Celine Hudelot, and Pablo Piantanida. Transductive learning for textual few-shot classification in api-based embedding models. In Empirical Methods in Natural Language Processing, 2023. 2 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6 +[8] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Workshop on Computer Vision and Pattern Recognition, pages 178-178, 2004. 6 +[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, pages 1-15, 2023. 2, 6, 8, 7 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[12] Steffen Herbold. Autorank: A python package for automated ranking of classifiers. Journal of Open Source Software, 5 (48):2173, 2020. 4 +[13] Mingyi Hong, XiangfengWang, Meisam Razaviyayn, and Zhi-Quan Luo. Iteration complexity analysis of block coordinate descent methods. Mathematical Programming, 163: 85–114, 2017. 4 +[14] Z hong, D Friedman, and D Chen. Factual probing is [mask]: Learning vs. learning to recall. In Conference of the North + +American Chapter of the Association for Computational Linguistics, 2021. 2 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916, 2021. 1 +[16] Z Jiang, F Xu, J Araki, and G Neubig. How can we know what language models know. In Association for Computational Linguistics, 2020. 2 +[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In International Conference on Computer Vision, pages 554–561, 2013. 6 +[18] Kenneth Lange, David R Hunter, and Ilsoon Yang. Optimization transfer using surrogate objective functions. Journal of computational and graphical statistics, 9(1):1-20, 2000. 3 +[19] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 6 +[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 6 +[21] Jorge Nocedal. Updating quasi-newton matrices with limited storage. Mathematics of Computation, 35(151):773-782, 1980. 6, 8 +[22] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505. IEEE, 2012. 6 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 3, 6, 8, 7 +[24] T Shin, Logan R. L. IV Razeghi, Y, E Wallace, and S Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. 2020. 2 +[25] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 6 +[26] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3485-3492. IEEE, 2010. 6 +[27] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 8, 7 +[28] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 1, 2, 5, 6, 7, 8, 3, 4 + +[29] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6, 8, 4, 7 +[30] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision, 2023. 2, 6, 8, 7 \ No newline at end of file diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/images.zip b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa715fd17c17f90e1bea5ab8b2d2bdd542aa7527 --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9a3f19b26b25800bcaa2b6bed8aee080bfe0e7c1c2226d50ef7fe89c1e4b704 +size 413670 diff --git a/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/layout.json b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9c1ff1da631d94469ff7b5e26e4501fc61c15f22 --- /dev/null +++ b/2024/LP++_ A Surprisingly Strong Linear Probe for Few-Shot CLIP/layout.json @@ -0,0 +1,9555 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "inline_equation", + "content": "\\mathbf{LP} + +: A" + }, + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "text", + "content": " Surprisingly Strong Linear Probe for Few-Shot CLIP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "spans": [ + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": "Yunshi Huang" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": " Fereshteh Shakeri" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": " \nJose Dolz" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": " Malik Boudiaf" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": " Houda Bahig" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": " Ismail Ben Ayed" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": "ÉTS Montréal, " + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 141, + 473, + 198 + ], + "type": "text", + "content": "Université de Montréal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 225, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 225, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 225, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "text", + "content": "In a recent, strongly emergent literature on few-shot CLIP adaptation, Linear Probe (LP) has been often reported as a weak baseline. This has motivated intensive research building convoluted prompt learning or feature adaptation strategies. In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline, in which the linear classifier weights are learnable functions of the text embedding, with class-wise multipliers blending image and text knowledge. As our objective function depends on two types of variables, i.e., the class visual prototypes and the learnable blending parameters, we propose a computationally efficient block coordinate Majorize-Minimize (MM) descent algorithm. In our full-batch MM optimizer, which we coin " + }, + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "inline_equation", + "content": "LP++" + }, + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "text", + "content": ", step sizes are implicit, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. By examining the mathematical properties of our loss (e.g., Lipschitz gradient continuity), we build majorizing functions yielding data-driven learning rates and derive approximations of the loss's minima, which provide data-informed initialization of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yields, surprisingly, highly competitive few-shot CLIP performances. Furthermore, " + }, + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "inline_equation", + "content": "LP++" + }, + { + "bbox": [ + 47, + 251, + 290, + 610 + ], + "type": "text", + "content": " operates in black-box, relaxes intensive validation searches for the optimization hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP adaptation methods. Our code is available at: https://github.com/FereshteShakeri/FewShot-CLIP-Strong-Baseline.git." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 631, + 128, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 287, + 687 + ], + "type": "text", + "content": "Recently, there has been a growing popularity in multimodal learning methods, which process and merge information from diverse modalities. In particular, large-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 228, + 531, + 382 + ], + "blocks": [ + { + "bbox": [ + 322, + 228, + 531, + 382 + ], + "lines": [ + { + "bbox": [ + 322, + 228, + 531, + 382 + ], + "spans": [ + { + "bbox": [ + 322, + 228, + 531, + 382 + ], + "type": "image", + "image_path": "2e659f8dd36f94f8c956f7f15aad2f81b3215032cc027fe1e3fd06c7d2b3793e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 395, + 548, + 495 + ], + "lines": [ + { + "bbox": [ + 304, + 395, + 548, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 548, + 495 + ], + "type": "text", + "content": "Figure 1. Comparison of " + }, + { + "bbox": [ + 304, + 395, + 548, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 395, + 548, + 495 + ], + "type": "text", + "content": " with state-of-the-art few-shot CLIP methods in the 1-shot setting across 11 datasets. We compute the mean accuracy and standard deviation using 10 random tasks for each dataset. The error bars indicate the average standard deviation over all 11 datasets. The x-axis represents the run time for one task, averaged over the 11 datasets. Tip-Adapter-F and Tip-Adapter-F* are two re-implementations of Tip-Adapter-F [28], with fixed and grid-search hyper-parameters, respectively (implementation details provided in Sec. 3.2)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "scale vision-language models (VLMs), such as CLIP [23] and ALIGN [15], have attracted wide attention and made substantial progress in computer vision, showing promising generalization capabilities in various downstream tasks. Unlike conventional task-specific models that are trained with a predetermined set of labels, these so-called foundation models learn to align images with text, in an open-vocabulary fashion. They train, via contrastive learning, vision and text embeddings jointly using a large-scale amount of image-text pairs collected over the internet, thereby leveraging the rich semantic knowledge inherent to language (e.g., concept hierarchies). For a given downstream image classification task, vision-language embeddings enable zero-shot predictions, without re-training, using textual descriptions of the classes (a.k.a prompts). For instance, for a given class " + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": ", a textual description of the class," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 712 + ], + "type": "text", + "content": "*Equal contribution. Correspondence to yunshi.huang@etsmtl.ca, fereshteh.shakeri.1@etsmtl.net" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23773" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": "which we denote " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": ", could be \"a photo of a " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "[class_{k}]" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": "\", where " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "[class_{k}]" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": " is the class name. Thus, the zero-shot class prediction for a query image " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": " is obtained from the cosine similarity between the " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": "-normalized vision-encoded embeddings, " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "f = \\theta_{v}(x)" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": ", and the text-encoded ones, " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "t_{k} = \\theta_{t}(z_{k})" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "\\hat{k} = \\arg \\max_{k} f^{t} t_{k}" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 72, + 286, + 145 + ], + "type": "text", + "content": " denotes the transpose1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": "Motivated by the observation that the choice of input prompts " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " may affect the zero-shot predictions, and following on from the strong recent emergence of prompt learning research in the NLP community [14, 16, 24], the popular work in [29] pioneered context optimization (CoOp) for vision-language models. CoOp models input text " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " as learnable continuous vectors, e.g., in the form " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "z_{k} = (z_{k}^{1},\\dots,z_{k}^{M},[\\mathrm{class}_{k}])" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "(z_k^l)_{1\\leq l\\leq M}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " are learnable text tokens, " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "[\\mathrm{class}_k]" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " is a fixed token corresponding to the word embedding vector of the name of the " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " class, and " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " is a hyper-parameter. These learnable vectors are finetuned as task-specific prompts using few-shot training examples and a standard supervised classification loss. More specifically, in this few-shot setting, we assume access to a set consisting of a few labeled samples for each target class, often referred to as the support set. Let " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "f_{i} = \\theta_{v}(\\pmb{x}_{i})" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " denote the vision embedding of support image " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "y_{ik}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " its one-hot encoded label, i.e., " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "y_{ik} = 1" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " if image " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " and 0 otherwise. Expressing the text embeddings as " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "t_k = \\theta_t(z_k^1,\\dots,z_k^M,[\\mathrm{class}_k])" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": ", CoOp fine-tunes text tokens " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "(z_k^l)_{1\\leq l\\leq M}" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " by minimizing the cross-entropy (CE) loss, with " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " labeled support samples and " + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 145, + 287, + 408 + ], + "type": "text", + "content": " classes:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 120, + 415, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 120, + 415, + 287, + 449 + ], + "type": "interline_equation", + "content": "- \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} \\tag {1}", + "image_path": "920d6896cfab174752530c7e9397e5f38f800e4c5ffddf5a45e76e3be94ac11d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "text", + "content": "where the softmax predictions " + }, + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "inline_equation", + "content": "p_{ik}" + }, + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "text", + "content": " and the logits (class scores) " + }, + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "inline_equation", + "content": "l_{ik}" + }, + { + "bbox": [ + 47, + 456, + 287, + 480 + ], + "type": "text", + "content": " are given by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 487, + 242, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 487, + 242, + 517 + ], + "spans": [ + { + "bbox": [ + 91, + 487, + 242, + 517 + ], + "type": "interline_equation", + "content": "p _ {i k} = \\frac {\\exp (l _ {i k})}{\\sum_ {j = 1} ^ {K} \\exp (l _ {i j})}; \\quad l _ {i k} = \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}", + "image_path": "6cda9be16a2f994b72eab7fd4cddfdc22690fbad8cdbbcb77b5abb0f3f7be66e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 523, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 523, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 287, + 668 + ], + "type": "text", + "content": "Although recent, the pioneering idea of CoOp has triggered a quite abundant literature on prompt learning for few-shot vision-language models, with numerous, more convoluted extensions, e.g. [4, 27, 30], to list a few. For instance, PLOT [4] followed up by learning multiple prompts, to describe the characteristics of each class, via minimizing an optimal-transport distance. KgCoOp [27] improves CoOp's performance when dealing with unseen classes, via minimizing the discrepancy between the text embeddings generated by the learned prompts and hand-crafted ones. While CoOp directly updates the context vectors using the CE loss, ProGrad [30] aligns the few-shot down" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "stream knowledge with the large-scale general knowledge, thus mitigating the overfitting of the few-shot samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 96, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 323 + ], + "type": "text", + "content": "Prompt learning methods have brought significant improvements over zero-shot classification, but they come at the price of heavy computational and memory load, as they require gradient back-propagation through the entire text encoder. Furthermore, they assume knowledge of the text encoder. These aspects may impede their deployment in low-resource and black-box, privacy-preserving scenarios, which are of wide interest in practice. Indeed, in NLP, there is currently an emerging literature on fast few-shot adaptation of black-box models [6], strongly motivated by the fact that large-scale foundation models (e.g., the GPT family, Anthropic's Claude or Google's PaLM) are only available through APIs and their pre-trained weights are not shared. Finally, by evaluating prompt learning methods over larger numbers of sampled support sets in our experiments (Fig. 1), we observed that they exhibit large variation in performances. This could be explained by the fact that, through the text encoder, they learn prompts that are \"too specialized\" for a given image support set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 323, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 407 + ], + "type": "text", + "content": "While prompt learning alters the textual inputs, another category of approaches, referred to as adapters, focused on transforming the pre-training features of the visual or language encoders, e.g., [9, 28]. These adapters are non-linear transformations, for instance, in the form of multi-layer modules, added to the encoder's bottleneck. They learn additional transformations, yielding logits of the form:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 416, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 416, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 391, + 416, + 545, + 430 + ], + "type": "interline_equation", + "content": "l _ {i k} = \\boldsymbol {\\theta} _ {a} \\left(\\boldsymbol {f} _ {i}, \\boldsymbol {t} _ {k}\\right) \\tag {2}", + "image_path": "cc7a5f366196ae1792528ad7939918e4c85f9be2182239e5f4e79be12c20facb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "text", + "content": "The adapter's learnable parameters, " + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "text", + "content": ", are fine-tuned over a few-shot task by optimizing the cross-entropy loss, similarly to (1) but with logits " + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "inline_equation", + "content": "l_{ik}" + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "text", + "content": " expressed as functions of " + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "inline_equation", + "content": "\\theta_{a}" + }, + { + "bbox": [ + 304, + 437, + 545, + 700 + ], + "type": "text", + "content": ". For instance, the popular CLIP-Adapter [9] integrated a multi-layered perceptron to modify the features, along with residual connections, which enable blending with the original pre-trained features. Tip-Adapter [28] added a nonlinear, quadratic-complexity module, which evaluates the pairwise similarities between the features of the support sets, and blends the ensuing class scores with the original textual features. This category of approaches mitigates some of the limitations of prompt-learning methods as they result in few-shot adaptation that has significantly lower computation and memory loads. However, as shown in Fig. 1 and in our experiments, their performances seem to depend strongly on some key hyper-parameters that have to be adjusted carefully on each downstream task, e.g. those that control the blending between the vision and language features. Therefore, to perform competitively, they incur an additional computation overhead to the adaptation phase, due to intensive (e.g., grid) search of the hyper-parameters over task-dedicated validation sets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "In the above-mentioned, strongly emergent literature on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 673, + 286, + 692 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 673, + 286, + 692 + ], + "spans": [ + { + "bbox": [ + 46, + 673, + 286, + 692 + ], + "type": "text", + "content": "1For " + }, + { + "bbox": [ + 46, + 673, + 286, + 692 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 673, + 286, + 692 + ], + "type": "text", + "content": "-normalized feature embeddings, the dot product corresponds to the cosine similarity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "type": "text", + "content": "2The number of labeled support samples per class, " + }, + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "type": "inline_equation", + "content": "S = \\frac{N}{K}" + }, + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "type": "text", + "content": " , is small, typically in " + }, + { + "bbox": [ + 47, + 692, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\{1,2,4,\\dots,16\\}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23774" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "few-shot CLIP adaptation, linear probe (LP) [23] has been often reported as a very weak baseline. For instance, in the 1-shot setting, it scores near " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " lower than the zero-shot predictions averaged over 11 benchmarks (Table 1). Initially evaluated in [23], LP is a linear classifier on the vision-encoded features. Specifically, it optimizes the CE loss (1) w.r.t the last-layer weights of the vision encoder (i.e., the class prototypes), which we will denote " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "(\\pmb{w}_k)_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " in the rest of the paper, with the logits given by: " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "l_{ik} = \\pmb{f}_i^t\\pmb{w}_k" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". A clear deficiency in this standard LP baseline is that it omits completely the language knowledge of CLIP, i.e., " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "(\\pmb{t}_k)_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 275 + ], + "type": "text", + "content": "In this work, we propose and examine from convex-optimization perspectives a generalization of the standard LP baseline. Specifically, we extend the logits in the CE loss in (1), so that they become learnable functions of the text embedding:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 121, + 281, + 212, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 281, + 212, + 294 + ], + "spans": [ + { + "bbox": [ + 121, + 281, + 212, + 294 + ], + "type": "interline_equation", + "content": "l _ {i k} = \\boldsymbol {f} _ {i} ^ {t} (\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k})", + "image_path": "1912e00013ffdef63a91491093733ddf6677fe9dfcd5671995e70d429ef120be.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\alpha_{k})_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": " trainable class-wise parameters blending image and text knowledge. As our objective function depends on two types of variables, i.e., the visual class prototypes " + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\pmb{w}_k)_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": " and blending parameters " + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\alpha_{k})_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": ", we propose a computationally efficient Block Majorize-Minimize (BMM) procedure. In our full-batch MM optimizer, which we coin " + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": ", step sizes are implicit in the definition of the majorizing functions, unlike standard gradient descent practices where learning rates are intensively searched over validation sets. Moreover, we examine the mathematical properties of our objective, i.e., (i) Lipschitz gradient continuity and (ii) decomposition into convex functions having closed-form optima. This enabled us to build majorizing functions yielding data-driven learning rates, and to derive approximations of the objective-function minima, which yield data-informed initializations of the variables. Our image-language objective function, along with these non-trivial optimization insights and ingredients, yield, surprisingly, highly competitive few-shot CLIP performances (Fig. 1). Furthermore, " + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 300, + 289, + 601 + ], + "type": "text", + "content": " operates in black-box, relaxes intensive validation searches for the optimization of hyper-parameters, and runs orders-of-magnitudes faster than state-of-the-art few-shot CLIP methods (Table 4). For instance, for 16-shot ImageNet adaptation, it takes seconds on a single NVIDIA RTX A600 GPU." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 609, + 172, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 172, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 172, + 621 + ], + "type": "text", + "content": "2. Formulation of " + }, + { + "bbox": [ + 47, + 609, + 172, + 621 + ], + "type": "inline_equation", + "content": "\\mathbf{LP} + +" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "text", + "content": "Following the notations introduced in the previous section, we propose to minimize the following CE objective function w.r.t visual class prototypes " + }, + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\mathbf{w} = (\\pmb{w}_k)_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "text", + "content": " and class-wise blending parameters " + }, + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "inline_equation", + "content": "\\alpha = (\\alpha_{k})_{1\\leq k\\leq K}" + }, + { + "bbox": [ + 46, + 628, + 287, + 678 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 683, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 683, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 82, + 683, + 287, + 715 + ], + "type": "interline_equation", + "content": "L (\\mathbf {w}, \\boldsymbol {\\alpha}) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {k = 1} ^ {K} y _ {i k} \\ln p _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) \\tag {3}", + "image_path": "4344d70eac29bf948ef52bb0183762e3e502f6254a89e7ecb2994cb47d86031a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 168 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 168 + ], + "type": "image", + "image_path": "b89ac625c0fbad8bc8a8033039281ad74331c693ebe7010aba59edd154d9914f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 182, + 488, + 193 + ], + "lines": [ + { + "bbox": [ + 362, + 182, + 488, + 193 + ], + "spans": [ + { + "bbox": [ + 362, + 182, + 488, + 193 + ], + "type": "text", + "content": "Figure 2. Visualization of " + }, + { + "bbox": [ + 362, + 182, + 488, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{LP} + +" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 213, + 535, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 535, + 225 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 535, + 225 + ], + "type": "text", + "content": "where softmax probability outputs " + }, + { + "bbox": [ + 306, + 213, + 535, + 225 + ], + "type": "inline_equation", + "content": "p_{ik}" + }, + { + "bbox": [ + 306, + 213, + 535, + 225 + ], + "type": "text", + "content": " are now given by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 338, + 231, + 545, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 338, + 231, + 545, + 262 + ], + "type": "interline_equation", + "content": "p _ {i k} (\\mathbf {w}, \\boldsymbol {\\alpha}) = \\frac {\\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {k} + \\alpha_ {k} \\boldsymbol {t} _ {k}\\right)\\right)}{\\sum_ {j = 1} ^ {K} \\exp \\left(\\boldsymbol {f} _ {i} ^ {t} \\left(\\boldsymbol {w} _ {j} + \\alpha_ {j} \\boldsymbol {t} _ {j}\\right)\\right)} \\tag {4}", + "image_path": "e916182caa47feb45f95d003b957776225b76ac7772e3453c1d95424cd7411f7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "content": " are the fixed pre-trained embeddings of the text templates used in zero-shot CLIP [23]. Clearly, the objective function defined by (3) and (4) could be viewed as a generalization of the vision-encoder CE loss used in the standard LP [23]. Indeed, the latter corresponds to setting " + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\alpha_k = 0 \\forall k" + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "content": ", i.e., no text knowledge. As we will see in our experimental ablation over different objective functions (Table 2), introducing the text knowledge " + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "(\\alpha_k > 0 \\forall k)" + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "content": " has a substantial effect on performances. Also, making " + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 304, + 266, + 545, + 422 + ], + "type": "text", + "content": " learnable (rather than fixed) leads to a further significant impact. Indeed, we hypothesize that the optimal blending of the text and visual knowledge is task dependent, which motivates learning it from the context of the support set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 429, + 539, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 429, + 539, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 429, + 539, + 441 + ], + "type": "text", + "content": "2.1. Block coordinate Majorize-Minimize descent" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": "Majorize-Minimize (MM) [18] is a very general optimization principle, which includes different classes of standard optimizers such as gradient descent, concave-convex procedures and expectation-maximization. Let " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = (\\mathbf{w},\\alpha)\\in" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{K(D + 1)}" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": " denote the overall vector of variables in our case, with " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": " being the dimension of the feature embeddings. At each iteration, the MM procedure updates the variable as the minimum of a majorizing function, i.e., an upper bound on the original objective, which is tight at the current iteration " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "j\\colon L(\\mathbf{v})\\leq M(\\mathbf{v},\\mathbf{v}^j)" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "L(\\mathbf{v}^j) = M(\\mathbf{v}^j,\\mathbf{v}^j)" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": ". Thus, update step " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{j + 1} = \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": " guarantees that the original objective does not increase at each iteration: " + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "inline_equation", + "content": "L(\\mathbf{v}^{j + 1})\\leq M(\\mathbf{v}^{j + 1},\\mathbf{v}^j)\\leq M(\\mathbf{v}^j,\\mathbf{v}^j) = L(\\mathbf{v}^j)" + }, + { + "bbox": [ + 304, + 447, + 545, + 662 + ], + "type": "text", + "content": ". Therefore, in MM algorithms, step sizes are implicit in the definition of the majorizing function, unlike standard gradient-descent practices, in which the step sizes (a.k.a learning rates) are intensively searched over validation sets, via running the optimizer several times." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 663, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 663, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 663, + 545, + 687 + ], + "type": "text", + "content": "In this work, we exploit the Lipschitz-gradient continuity of our convex objective in (3), i.e., bounds on the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "This assumes, of course, that minimizing " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "M(\\mathbf{v}, \\mathbf{v}^j)" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " could be solved to global optimality and is easier than the original problem." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23775" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 289 + ], + "type": "text", + "content": "maximum eigen values of the Hessian matrices (Prop. 1), thereby building majorizing functions with data-driven, task-specific step sizes. This removes the need for validation searches for the optimization hyper-parameters, reducing the computational load for fine-tuning (Table 2), while yielding performances on par with the best learning rates found with validation (Fig. 4). Also, interestingly, the Lipschitz-based learning rates computed from our derivation in Prop. 1 are orders-of-magnitude larger than those typically used in deep learning, yielding steeper decreases towards the minimum. Before giving proper majorizing functions for our convex, gradient-Lipschitz function in (3), let us first point to the following results, well-known in convex optimization [3]. While these results are text-book knowledge in optimization, they enable to connect the general MM principle to gradient descent, motivating the data-driven, task-specific step sizes we derive in Prop. 1 and the block-coordinate MM optimizer we propose in Alg. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": "Lemma 2.1. ([3, p. 268]) Assume " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "L(\\mathbf{v})" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": " is a twice-differentiable function, which has a Lipschitz continuous gradient, i.e., there exists a strictly positive Lipschitz constant " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\nabla^2 L(\\mathbf{v}) \\preceq \\gamma \\mathbf{I}" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": " the identity matrix. Then, the following quadratic bound is a majorizing function for " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 294, + 288, + 367 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 374, + 287, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 374, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 52, + 374, + 287, + 396 + ], + "type": "interline_equation", + "content": "M (\\mathbf {v}, \\mathbf {v} ^ {j}) = L (\\mathbf {v} ^ {j}) + \\nabla L (\\mathbf {v} ^ {j}) ^ {t} (\\mathbf {v} - \\mathbf {v} ^ {j}) + \\frac {\\gamma}{2} \\| \\mathbf {v} - \\mathbf {v} ^ {j} \\| ^ {2} \\tag {5}", + "image_path": "c8593988778e9308268bc115a09839ec47c4969718dfb46e9cad1eb0e9de96dc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": "Furthermore, a specific gradient step, with learning rate " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\gamma}" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": " minimizes " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{j + 1} = \\mathbf{v}^j -\\frac{1}{\\gamma}\\nabla L(\\mathbf{v}^j)" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": " = " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\mathbf{v}}M(\\mathbf{v},\\mathbf{v}^j)" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": ", and guarantees that objective " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": " decreases by at least " + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\frac{1}{2\\gamma}\\| \\nabla L(\\mathbf{v})\\| ^2" + }, + { + "bbox": [ + 47, + 402, + 287, + 453 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 461, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 461, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 94, + 461, + 287, + 487 + ], + "type": "interline_equation", + "content": "L \\left(\\mathbf {v} ^ {j + 1}\\right) \\leq L \\left(\\mathbf {v} ^ {j}\\right) - \\frac {1}{2 \\gamma} \\| \\nabla L (\\mathbf {v}) \\| ^ {2} \\tag {6}", + "image_path": "3cac022ab0303b11e4fd80b4e1979f967da28fa8ac626d18341733d1f2ffb2b8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "text", + "content": "Moreover, the following Theorem, which follows from Lemma 2.1, establishes the sublinear convergence of the MM procedure using bound (5), i.e., a convergence rate of " + }, + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "inline_equation", + "content": "O(1 / J)" + }, + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 494, + 287, + 541 + ], + "type": "text", + "content": " being the total number of iterations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": "Theorem 2.2. ([3, p. 267]) For convex, twice-differentiable function " + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "inline_equation", + "content": "L(\\mathbf{v})" + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": ", which has a " + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": "-Lipschitz gradient, performing " + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": " updates " + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{j+1} = \\mathbf{v}^j - \\frac{1}{\\gamma} \\nabla L(\\mathbf{v}^j)" + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": ", starting from initialization " + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^0" + }, + { + "bbox": [ + 45, + 548, + 287, + 599 + ], + "type": "text", + "content": ", will yield a solution that satisfies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 605, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 605, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 93, + 605, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\left\\| L \\left(\\mathbf {v} ^ {J}\\right) - L \\left(\\mathbf {v} ^ {*}\\right) \\right\\| \\leq \\frac {\\gamma}{2 J} \\left\\| \\mathbf {v} ^ {0} - \\mathbf {v} ^ {*} \\right\\| \\tag {7}", + "image_path": "d69e70f466085ce7d4d3ad98fdeffd62fc93e9ca2667b72e3e8e627d6acafdd6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "inline_equation", + "content": "L(\\mathbf{v}^{*})" + }, + { + "bbox": [ + 47, + 634, + 186, + 647 + ], + "type": "text", + "content": " is the optimal value." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "For completeness, we provide the proofs of these well-known results in the supplemental material. Clearly, Lemma 2.1 and Theorem 2.2 prescribe a learning rate of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{\\gamma}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " for a function that has a " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "-Lipschitz gradient. One valid Lipschitz constant would be the maximum eigen value of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "the Hessian of our objective in (3), which provides a majorizing function of the form in Eq. (5) and data-driven learning rates. However, a naive spectral decomposition of the Hessian matrices (to obtain the maximum eigen value) could be computationally intensive. For instance, for ImageNet, the Hessian of our objective is of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "K(D + 1) \\times K(D + 1) \\approx 1M \\times 1M" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "D = 1024" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "K = 1000" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": ". In Prop. 1, we derive approximate global and block-wise Lipschitz constants that can be computed efficiently (i.e., evaluating the maximum eigen value of a single " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "D \\times D" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " matrix)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "text", + "content": "Block-coordinate updates Our procedure provided in Alg. 1 belongs to the family of Block Majorize-Minimize (BMM) methods, well studied in the optimization community [13]. To minimize a multi-block objective, as in our case where the blocks correspond to variables " + }, + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 218, + 545, + 314 + ], + "type": "text", + "content": ", we minimize one or many successive majorizing functions of the objective in each block, with the other block fixed, in a cyclic order:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 321, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 321, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 311, + 321, + 545, + 342 + ], + "type": "interline_equation", + "content": "L \\left(\\mathbf {w} ^ {j}, \\boldsymbol {\\alpha}\\right) + \\nabla L _ {\\mathbf {w}} \\left(\\mathbf {w} ^ {j}\\right) ^ {t} \\left(\\mathbf {w} - \\mathbf {w} ^ {j}\\right) + \\frac {\\gamma_ {\\mathbf {w}}}{2} \\| \\mathbf {w} - \\mathbf {w} ^ {j} \\| ^ {2} \\tag {8}", + "image_path": "26dea10fbbfaee02ae0f3c607b73ecce1eec9895d0a57532bf553372910c6c3e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 349, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 545, + 369 + ], + "type": "interline_equation", + "content": "L \\left(\\mathbf {w}, \\boldsymbol {\\alpha} ^ {j}\\right) + \\nabla L _ {\\boldsymbol {\\alpha}} \\left(\\boldsymbol {\\alpha} ^ {j}\\right) ^ {t} \\left(\\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j}\\right) + \\frac {\\gamma_ {\\boldsymbol {\\alpha}}}{2} \\| \\boldsymbol {\\alpha} - \\boldsymbol {\\alpha} ^ {j} \\| ^ {2} \\tag {9}", + "image_path": "4627fe1d8b46b9645c6b9f30536d549ddaf2d1ace904711c1663455a32203903.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": "where in (8), block " + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": " is fixed and, in (9), " + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": " is fixed. " + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\nabla L_{\\mathbf{w}}" + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "inline_equation", + "content": "\\nabla L_{\\alpha}" + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": " denote block-wise gradients, and " + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "inline_equation", + "content": "(\\gamma_{\\mathbf{w}},\\gamma_{\\alpha})" + }, + { + "bbox": [ + 304, + 373, + 545, + 648 + ], + "type": "text", + "content": " are the block Lipschitz constants. Accommodating different choices of the block-cycling strategies and majorizing functions, BMM includes a breadth of optimizers as particular cases, such the well-known block coordinate gradient descent (BCGD) [1] and its projection-based variant. Indeed, the so-called Gauss-Seidel cycling [13] alternates steps (8) and (9), which corresponds to the BCGD method. One could also perform many successive steps in one block, as we do in Alg. 1, which corresponds to the so-called Essentially-Cyclic4 strategy [13]. Importantly, for a fairly large spectrum of choices of the cycling strategies, BMM enjoys the same sublinear convergence property as MM for convex objectives (with each block-wise update decreasing the objective), provided that the majorizing functions are strongly convex; see Theorem 3.1 in [13]. This is the case for (8) and (9). In our experiments, we observed that this block-wise variant performs better than a single-block MM; see Table 2. This might be explained by the fact that each block of variables has a dedicated step size. In the supplemental material, we provide results for different block-cycling strategies." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 545, + 686 + ], + "type": "text", + "content": "Global and block-coordinatewise Lipschitz constants In the following, we derive approximate global and block-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "Essentially-Cyclic means that there is a period during which each block is updated at least once." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23776" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "coordinatewise Lipschitz constants for our objective function in (3), which could be evaluated efficiently. We deploy these in Alg. 1, to compute data-driven, block-wise learning rates for updating visual prototypes " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": " and blending parameters " + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": "Proposition 1. Considering the blocks of variables " + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": ", the gradient of our objective " + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": " in (3) is block-coordinatewise Lipschitz continuous. For " + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "inline_equation", + "content": "\\tau_{1} \\geq 2" + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": ", it has the following block Lipschitz constant for the set of variables in " + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 45, + 137, + 287, + 195 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 194, + 287, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 194, + 287, + 227 + ], + "type": "interline_equation", + "content": "\\gamma_ {\\mathbf {w}} = \\frac {\\tau_ {1}}{4 N} \\lambda_ {m a x} \\left(\\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} \\boldsymbol {f} _ {i} ^ {t}\\right)\\right) \\tag {10}", + "image_path": "7b8e4f3ba4a17345e2e867caab699fa395e24d1fb0ffa01c7cd5b5c386672b1f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\lambda_{max}(\\mathbf{A})" + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": " denotes the maximum eigenvalue of matrix " + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": ". Furthermore, for " + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\tau_{1} \\geq 1" + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": ", the expression in Eq. (10) provides a tighter but approximate block Lipschitz constant. Similarly, for " + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\tau_{2} \\geq 1" + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": ", we have following approximate block Lipschitz constant for the variables in " + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 230, + 287, + 290 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 296, + 287, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 296, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 110, + 296, + 287, + 328 + ], + "type": "interline_equation", + "content": "\\gamma_ {\\boldsymbol {\\alpha}} = \\max _ {k} \\frac {\\tau_ {2}}{4 N} \\sum_ {i = 1} ^ {N} \\left(\\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k}\\right) ^ {2} \\tag {11}", + "image_path": "1c02cadf942590b5873ced5b196a242652e1800ace07650a7daf2e0bcbe1afdc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "content": "Finally, for " + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\tau \\geq 2" + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "content": ", the following expression provides an approximate global Lipschitz constant for objective (3), i.e., w.r.t all variables " + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = (\\mathbf{w},\\alpha)" + }, + { + "bbox": [ + 46, + 334, + 287, + 369 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 377, + 287, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 377, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 125, + 377, + 287, + 389 + ], + "type": "interline_equation", + "content": "\\gamma = \\tau \\max \\left(\\gamma_ {\\mathbf {w}}, \\gamma_ {\\boldsymbol {\\alpha}}\\right) \\tag {12}", + "image_path": "ba9f05b9d5fb0e5e3daae03d3f5c69f46f9814f1a46fb03bd96f68abf8c41eef.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "text", + "content": "Proof. The details are deferred to the supplemental material. The main ingredients of the proof are based on the Gershgorin circle theorem and the variational characterization of the maximum eigenvalue, following the min-max theorem, also referred to as the variational principle. " + }, + { + "bbox": [ + 46, + 396, + 287, + 455 + ], + "type": "inline_equation", + "content": "\\square" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 462, + 204, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 462, + 204, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 462, + 204, + 473 + ], + "type": "text", + "content": "2.2. Initialization of the variables" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "text", + "content": "In the following Prop. 2, we derive approximations of the minima of our objective function (3), which yield a data-informed initialization of the variables. Indeed, the expressions we obtain in Eqs. (13) and (15) suggest initial guesses for variables " + }, + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 480, + 287, + 624 + ], + "type": "text", + "content": ". Interestingly, and as will be confirmed by our experiments (Table 3), such an initialization yields substantially lower values of the minimized loss than a random initialization. Furthermore, surprisingly, using this initialization for a training-free prediction yields better performances than the training-free version of the recent Tip-Adapter-F approach [28]. The details of our training-free version are provided in the supplemental material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": "Proposition 2. The cross-entropy in Eq. (3) could be written as the sum of two convex functions, i.e., " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "L = g_{1} + g_{2}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": ", such that, " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\forall k \\in [1, \\dots, K]" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": ", the minimum of " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{k}" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": " is co-linear to the hard mean vector of features within class " + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 629, + 287, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 683, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {1} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} y _ {i k}} \\tag {13}", + "image_path": "1f284e94c6d53661423076ee70c337c14ebb051268843145b752cf049bc5d269.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "and the minimum of " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\pmb{w}_{k}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " is co-linear to the soft mean vector of features within class " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 106, + 545, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 328, + 106, + 545, + 139 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {w} _ {k}} {\\arg \\min } g _ {2} = \\frac {1}{\\lambda N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} \\propto \\frac {\\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i}}{\\sum_ {i = 1} ^ {N} p _ {i k}} \\tag {14}", + "image_path": "f327b0be5cf45b4f9c732bed8eb8a32d1ec213b0cbf384a1e63e6a0a11f7eac2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\lambda \\leq \\min_k\\lambda_{min}(\\pmb {A}_k)" + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "inline_equation", + "content": "A_{k} = \\frac{1}{N}\\sum_{i = 1}^{N}(p_{ik} - p_{ik}^{2})\\pmb{f}_{i}\\pmb{f}_{i}^{t}" + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\lambda_{min}(\\mathbf{A})" + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "text", + "content": " denotes the smallest eigenvalue of matrix " + }, + { + "bbox": [ + 305, + 150, + 545, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "text", + "content": "Proof. We defer the details, including the full expressions of convex functions " + }, + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 305, + 186, + 545, + 209 + ], + "type": "text", + "content": ", to the supplemental material." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 535, + 211, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 535, + 211, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 535, + 211, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 535, + 211, + 545, + 220 + ], + "type": "image", + "image_path": "37c5aa575f31b844d941e9a66dc18a54ac3c1ec240bc5fadac45a09a9ce22693.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": "Similarly to the development in Prop. 2, one could decompose (3) as the sum of two convex functions, i.e., " + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "L = h_1 + h_2" + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": ", such that, " + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\forall k \\in [1, \\dots, K]" + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": ", the minima of " + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "h_2" + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": " w.r.t " + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\alpha_k" + }, + { + "bbox": [ + 305, + 235, + 545, + 307 + ], + "type": "text", + "content": " could be written, up to a multiplicative positive factor, as the hard and soft means of the cosine similarities between the image and text embeddings:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 317, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 317, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 361, + 317, + 545, + 350 + ], + "type": "interline_equation", + "content": "\\underset {\\alpha_ {k}} {\\arg \\min } h _ {1} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} y _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {15}", + "image_path": "5435c2642cbea17f64bfad78b1f78bcc96cd4db36d01aaffa636a788b64e59c7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 363, + 545, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 363, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 361, + 363, + 545, + 396 + ], + "type": "interline_equation", + "content": "\\underset {\\alpha_ {k}} {\\arg \\min } h _ {2} = \\frac {1}{\\beta N} \\sum_ {i = 1} ^ {N} p _ {i k} \\boldsymbol {f} _ {i} ^ {t} \\boldsymbol {t} _ {k} \\tag {16}", + "image_path": "bb84fcbc3c23c29031d51c49bb9400c5cf329200bf81d260964bb31fdcb6acbd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\beta = \\min_k\\frac{1}{N}\\sum_{i = 1}^N (p_{ik} - p_{ik}^2)(\\pmb {f}_i^t\\pmb {t}_k)^2" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": " (respectively " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "h_2" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": ") has the same expression as " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": " (respectively " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": "), except that term " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\frac{\\lambda}{2}\\sum_{k = 1}^{K}\\| \\pmb {w}_k\\|^2" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\frac{\\beta}{2}\\|\\pmb {\\alpha}\\|^2" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": "; see the supplemental material for the expressions of " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "inline_equation", + "content": "g_{2}" + }, + { + "bbox": [ + 305, + 403, + 545, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "type": "code", + "bbox": [ + 313, + 507, + 523, + 680 + ], + "blocks": [ + { + "bbox": [ + 313, + 494, + 492, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 494, + 492, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 492, + 506 + ], + "type": "text", + "content": "Algorithm 1: Block coordinate MM " + }, + { + "bbox": [ + 313, + 494, + 492, + 506 + ], + "type": "inline_equation", + "content": "(\\mathbf{w},\\alpha)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "lines": [ + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": "iterw=10;iter " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\tau_{1} = 1" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\tau_{2} = 16" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\lambda = \\frac{1}{N}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\beta = \\frac{1}{250K}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " Initialize w0,0 // Using (13) for each k Initialize " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\alpha^{0,0}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " // Using (15) for each k for " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "j = 0,1,\\ldots" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " do for " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "l_{1} = 0,1,\\ldots" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " ,iterw do " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\begin{array}{r}\\mathbf{w}^{j,l_1 + 1} = \\mathbf{w}^{j,l_1} - \\frac{1}{\\gamma_{\\mathbf{w}}}\\nabla L_{\\mathbf{w}}(\\mathbf{w}^{j,l_1},\\boldsymbol{\\alpha}^{j,0})\\\\ /\\quad \\frac{1}{\\gamma_{\\mathbf{w}}}\\text{from Eq.}(10) \\end{array}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^{j + 1,0} = \\mathbf{w}^{j,\\mathrm{iter}_{\\mathbf{w}}}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "l_{2} = 0,1,\\dots" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " iter " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\begin{array}{rl} & {\\boldsymbol{\\alpha}^{j,l_2 + 1} = \\boldsymbol{\\alpha}^{j,l_2} - \\frac{1}{\\gamma_\\alpha}\\nabla L_\\alpha (\\mathbf{w}^{j + 1,0},\\boldsymbol{\\alpha}^{j,l_2})}\\\\ & {/ / \\frac{1}{\\gamma_\\alpha}\\text{from Eq.}(11)} \\end{array}" + }, + { + "bbox": [ + 313, + 507, + 523, + 680 + ], + "type": "inline_equation", + "content": "\\alpha^{j + 1,0} = \\alpha^{j,\\mathrm{iter}_{\\alpha}}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "code_body" + } + ], + "index": 22, + "sub_type": "algorithm" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23777" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 69, + 516, + 201 + ], + "blocks": [ + { + "bbox": [ + 78, + 69, + 516, + 201 + ], + "lines": [ + { + "bbox": [ + 78, + 69, + 516, + 201 + ], + "spans": [ + { + "bbox": [ + 78, + 69, + 516, + 201 + ], + "type": "table", + "html": "
Number of shots (S)124816
Zero-shot CLIP ICML'21[23]58.89
Prompt-LearningCoOp ICV'22[29]59.62 ± 3.1163.80 ± 2.3267.23 ± 1.6471.30 ± 0.8674.06 ± 0.55
PLOTICLR'23[4]61.51 ± 2.9165.67 ± 2.0668.39 ± 1.1771.96 ± 0.7074.35 ± 0.66
KgCoOp CVPR'23[27]61.36 ± 3.0463.23 ± 2.0665.73 ± 1.1567.50 ± 1.1169.01 ± 0.79
ProGrad ICCV'23[30]62.46 ± 1.8965.88 ± 1.4668.52 ± 1.1571.82 ± 0.1173.95 ± 0.68
CLIP-based AdaptersCLIP-Adapter ICV'23[9]60.32 ± 0.8061.93 ± 0.9365.12 ± 0.8069.20 ± 0.5672.57 ± 0.54
Tip-Adapter-F ECCV'22[28]61.29 ± 0.9262.94 ± 0.7566.02 ± 0.8069.88 ± 0.5173.82 ± 0.55
Tip-Adapter-F* ECCV'22[28]63.06 ± 1.0566.47 ± 0.6568.71 ± 0.9671.78 ± 1.0074.37 ± 0.35
Linear-ProbingStandard LP ICML'21[23]36.10 ± 1.4346.99 ± 1.2956.72 ± 1.2064.66 ± 0.5570.56 ± 0.44
LP++63.43 ± 0.9066.20 ± 0.7269.16 ± 0.7972.04 ± 0.4674.42 ± 0.45
", + "image_path": "57d0ba468f52623d04854ed70dda6cf8f426b80567fd0009f876eba7780d85ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 209, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 546, + 232 + ], + "type": "text", + "content": "Table 1. Comparison to state-of-the-art methods. Average classification accuracy (%) on 11 benchmarks, with standard derivation over 10 sampled support sets for each dataset. The best values are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 251, + 128, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 128, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 128, + 264 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 240, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 240, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 240, + 284 + ], + "type": "text", + "content": "3.1. Datasets and Implementation details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 289, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 287, + 386 + ], + "type": "text", + "content": "Following the CLIP-based few-shot adaptation literature [27, 28], we conduct the main experiments on 11 public classification data sets: Caltech101 [8], ImageNet [7], DTD [5], OxfordPets [22], Flowers102 [20], StandfordCars [17], Food101 [2], FGVCAircraft [19], SUN397 [26], EuroSAT [11] and UCF101 [25]. We follow standard practices [23] and consider " + }, + { + "bbox": [ + 47, + 289, + 287, + 386 + ], + "type": "inline_equation", + "content": "S = \\{1, 2, 4, 8, 16\\}" + }, + { + "bbox": [ + 47, + 289, + 287, + 386 + ], + "type": "text", + "content": " shots for model adaptation, which are randomly sampled for each data set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 386, + 287, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 287, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 287, + 481 + ], + "type": "text", + "content": "Towards a fair validation set. Apparently, prior works on this problem [28] have resorted to a large set of validation samples to adjust their hyper-parameters. For the sake of fairness, we tune the hyper-parameters across all the methods based on a small validation set, which contains as many samples (i.e., shots) as the training set. Furthermore, to avoid the potential overfitting on the few training samples, we adopt an early stopping strategy on this validation set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 482, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 482, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 482, + 288, + 662 + ], + "type": "text", + "content": "General Setting. While the existing works evaluate methods based on either a single or three random tasks (support sets) [28, 29], we found that, for some datasets, the chosen support samples may not be representative of the class, leading to large standard deviations in low-shot scenarios (see Fig. 1). To ensure fair comparisons, we evaluate all the methods by averaging their classification accuracies over 10 randomly sampled tasks, for each dataset. In all the experiments, we employ ResNet-50 [10] as the visual encoder for the CLIP backbone. It is important to note that, for our BMM procedure in Alg. 1, the optimizer hyper-parameters remain fixed across all the datasets. We use the validation set only to find the best model, via a single run of our BMM procedure with a fixed number of variable updates, i.e., 300 gradient updates including all the blocks of variables." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 113, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 113, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 113, + 682 + ], + "type": "text", + "content": "3.2. Baselines" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We benchmark the proposed " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " against relevant state-of-the-art methods in the few-shot adaptation of CLIP-based" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "content": "models. We first resort to zero-shot CLIP as the standard baseline, which only leverages the knowledge learned by the pre-trained CLIP model. Also, we include the standard LP baseline, whose implementation is done following [23, 29]. More concretely, this baseline optimizes the standard cross-entropy loss, which corresponds to " + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\alpha_{k} > 0\\forall k" + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "content": " in our generalization in (3), using the L-BFGS [21] optimizer5. It also includes an " + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 251, + 545, + 358 + ], + "type": "text", + "content": "-regularizer, whose balancing weight is set based on the validation set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "text", + "content": "CLIP-based adapters. We benchmark " + }, + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "text", + "content": " against two popular adapter-based few-shot approaches: CLIP-adapter [9] and TIP-adapter [28]. As exposed earlier, several popular works follow unfair practices by resorting to a larger validation set, or even to the entire test set, to adjust their key hyperparameters -as well as their model selection criteria (i.e., epochs)- for each task. For the sake of fairness, we re-implement Tip-Adapter-F and report the results in 2 different settings. In the first setting (Tip-Adapter-F), we set the two crucial hyper-parameters of this method to 1, keeping them fixed during training, and adopt early stopping based on the validation set. In the second setting, referred to as Tip-Adapter-" + }, + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^*" + }, + { + "bbox": [ + 304, + 360, + 546, + 552 + ], + "type": "text", + "content": ", we perform intensive grid-search on the validation set to find the best values for these hyperparameters at initialization, which incurs an additional time complexity burden compared to Tip-Adapter-F." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 611 + ], + "type": "text", + "content": "Prompt learning. We further compare the proposed LP++ to relevant prompt-learning methods, including CoOp[29] and more recent variants, such as PLOT[4], KgCoOp[27] and ProGrad[30]. We also apply early stopping here based on the performances on the validation set." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 616, + 362, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 362, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 362, + 628 + ], + "type": "text", + "content": "3.3. Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 635, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 659 + ], + "type": "text", + "content": "Comparison to the state-of-the-art In Table 1, we present the quantitative results obtained by " + }, + { + "bbox": [ + 304, + 635, + 545, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 635, + 545, + 659 + ], + "type": "text", + "content": " (Alg. 1)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "inline_equation", + "content": "{}^{5}" + }, + { + "bbox": [ + 304, + 664, + 545, + 713 + ], + "type": "text", + "content": " L-BFGS (Limited-memory BFGS) aims to find the minimum of objective function using the second order method. It estimates the Hessian matrix based on recent gradients only, enabling it to determine the steepest direction for achieving the optimal solution. Additionally, it is implemented with line searches to automatically determine the optimal step size." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23778" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 72, + 231, + 180 + ], + "blocks": [ + { + "bbox": [ + 88, + 72, + 231, + 180 + ], + "lines": [ + { + "bbox": [ + 88, + 72, + 231, + 180 + ], + "spans": [ + { + "bbox": [ + 88, + 72, + 231, + 180 + ], + "type": "image", + "image_path": "399c52c91fb727f8824f3c4ac338827d397517cc4e2d198e3da8e65a58be093f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": "Figure 3. Quantitative performance of different adaptation methods on the 11 benchmarks (mean), as well as in two other datasets, averaged over 10 tasks (additional figures on the remaining 9 datasets can be found in Appendix, Sec. 16)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 72, + 369, + 180 + ], + "blocks": [ + { + "bbox": [ + 232, + 72, + 369, + 180 + ], + "lines": [ + { + "bbox": [ + 232, + 72, + 369, + 180 + ], + "spans": [ + { + "bbox": [ + 232, + 72, + 369, + 180 + ], + "type": "image", + "image_path": "97406d7969a0fb71137bbe71d34d449e1344a43d80a2b3b0332b491e4414df72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 504, + 180 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 504, + 180 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 504, + 180 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 504, + 180 + ], + "type": "image", + "image_path": "d86fd7af33d3477ccc718f4954a040d2e8e6d99706c01450054ac865990d2dad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": "and the relevant literature in the task of efficient adaptation of VLMs. We report the average classification accuracy and standard deviation, across 11 classification benchmarks. From these results, one can make several observations. First, while the standard LP baseline largely underperforms the existing adaptation methods, our improved version, " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": ", brings significant performance gains, particularly in the low-labeled data regimes. It is important to stress that the standard LP baseline integrates only the visual features extracted from CLIP, disregarding the text-encoder knowledge. This contrasts with the existing adapters, which leverage both image and text information. Therefore, these results evidence that the potential of LP has been severely underestimated in the existing literature. Second, when the model selection process is performed fairly, i.e., using a small validation set, their performances fall behind the proposed method (from " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": "), despite being arguably more complex approaches. In particular, among the adapter-based strategies, only Tip-Adapter-F* yields performances on par with " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": ", but at the cost of increasing computational load, due to an additional intensive grid-search over its hyper-parameters. If we look at prompt-learning methods, ProGrad and PLOT might be considered as competitors of " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": ", particularly as the number of shots increases. Nevertheless, as already discussed, these methods are computationally inefficient compared to adapters, and do not enable black-box adaptation. The overall performance across all the methods is depicted in Fig. 3 for the average over the 11 benchmarks, as well as in two datasets, showing that " + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 233, + 289, + 616 + ], + "type": "text", + "content": " typically outperforms existing methods under different few-shot scenarios. A more detailed analysis is deferred to the Appendix, Sec. 16." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Ablation on the loss functions and different optimization strategies. Table 2 reports the test accuracy and run time for different optimizers and loss functions, including the standard CE loss " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha_{k} = 0\\forall k)" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": " and our loss with learnable blending parameters. Additionally, for " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LP + + }" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": ", we evaluate our loss with fixed blending parameters " + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\alpha_{k} = 1\\forall k)" + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": ". Independently of the optimizer used, the main takeaway from Table 2 is that introducing the text knowledge and making" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": " learnable (rather than fixed) have a substantial impact on accuracy. As for evaluating the optimizers, and for a fair comparison, we use a fixed budget for the number of variable updates for all the optimizers (i.e., 300 updates). In the case of " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": ", this corresponds to the total number of updates for all the blocks. Also, for all optimizers, we initialize " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": " following Eqs. (13) and (15). We first consider two popular optimizers, i.e., GD and ADAM, and deploy them in two different settings. First, we run each optimizer 7 times, with each run corresponding to a learning rate in the range " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "[10^{-4}, 10^{2}]" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": ". Then, we record the best performances obtained on the validation set; see GD (optimum) and ADAM (optimum) in Table 2. This follows the standard practices in deep learning, i.e., searching for the learning rates over validation sets, which incurs additional computation overhead; see the time column in Table 2. Second, we run GD and ADAM with our data-driven learning rate, as prescribed by the approximate Lipschitz constant we derived in Eq. (12), with " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": "; see GD (our Lipschitz cst) and ADAM (our Lipschitz cst) in Table 2. Note that, in this case, GD corresponds to " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": " with a single block. Furthermore, we include the L-BFGS in these comparisons, with its initial learning rate set to 1. For L-BFGS, implementing a line search for the optimal step size also introduces an additional computational overhead. As highlighted by Table 2, our method removes the need for validation searches for the optimization hyper-parameters, thanks to its data-driven, task-specific step sizes, thereby reducing the computational load for fine-tuning. In the meanwhile, it yields performances on par with those obtained with the best learning rates found with the validation set. In Fig. 4, we plot the performances of single-block GD vs. the learning rates in the range " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "[10^{-4}, 10^{2}]" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": ", for three datasets. We observe that our Lipschitz-based, task-specific step sizes match the optimal ones found on the validation set, although these vary among the datasets. Also, interestingly, these Lipschitz-based step sizes are orders-of-magnitude larger than those used in deep learning, which are, typically, within interval " + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "inline_equation", + "content": "[10^{-4}, 10^{-2}]" + }, + { + "bbox": [ + 304, + 233, + 547, + 700 + ], + "type": "text", + "content": "; see [28], for instance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "How to initialize the classifier? To justify empirically the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23779" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 71, + 230, + 189 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 230, + 189 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 230, + 189 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 230, + 189 + ], + "type": "image", + "image_path": "981378ca95952e9ba244a68193192d2f88625889f51da89ab839ffd6df107efd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 546, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 546, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 546, + 222 + ], + "type": "text", + "content": "Figure 4. Single-block GD performance as a function of different values of the learning rates. The dotted vertical line shows the Lipschitz-based, data-driven learning rate." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 367, + 190 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 367, + 190 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 367, + 190 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 367, + 190 + ], + "type": "image", + "image_path": "e30e12bebdb58dc8c0958692c3af848dcec042dc52c580455523bf65982219bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 367, + 72, + 503, + 190 + ], + "blocks": [ + { + "bbox": [ + 367, + 72, + 503, + 190 + ], + "lines": [ + { + "bbox": [ + 367, + 72, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 367, + 72, + 503, + 190 + ], + "type": "image", + "image_path": "1545f4288f93bb4dc4292c0f3a22e4254bb4d62fe49218e0f91909f2fa26924c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 239, + 286, + 331 + ], + "blocks": [ + { + "bbox": [ + 48, + 239, + 286, + 331 + ], + "lines": [ + { + "bbox": [ + 48, + 239, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 286, + 331 + ], + "type": "table", + "html": "
Optimization MethodStandard loss (αk=0)Our lossTime
1 shot16 shot1 shot16 shot
LP++ (αk is learnt)35.5569.7563.4374.420.78s
LP++ (αk = 1)--46.3769.410.72s
GD (our Lipschitz cst)35.5569.7563.0474.280.86s
GD (optimum)35.6469.6762.9374.556.02s
ADAM (our Lipschitz cst)26.2264.5325.4464.620.98s
ADAM (optimum)35.9069.7363.0774.236.89s
L-BFGS [21]34.5467.4462.0972.826.73s
", + "image_path": "9a08db9866e8f6439f6a862f35782fe179d19be4b7a576f218351ab7337d4a2b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "text", + "content": "advantages brought by initializing the classifier weights following Eq. (13), and the blending parameters in Eq. (15), we evaluate objective (3) at the beginning of the training, as well as the training-free test accuracy of three methods: random initialization, and the training-free version of both the Tip-adapter-F method [28] (i.e., Tip-adapter) and our method (we provide more details on training-free " + }, + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 46, + 403, + 287, + 523 + ], + "type": "text", + "content": " in the appendix). The results in Table 3 confirm empirically the technical observations in Eqs. (13) and (15), which prescribe initial guesses for our problem's variables." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 531, + 276, + 612 + ], + "blocks": [ + { + "bbox": [ + 46, + 339, + 288, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 339, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 288, + 384 + ], + "type": "text", + "content": "Table 2. Accuracy and run time for different optimizers and loss functions (average over 11 datasets). The running time is recorded for our loss (16 shots), and is averaged over 11 datasets. The best result is marked in bold, and the second best is underlined." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 531, + 276, + 612 + ], + "lines": [ + { + "bbox": [ + 58, + 531, + 276, + 612 + ], + "spans": [ + { + "bbox": [ + 58, + 531, + 276, + 612 + ], + "type": "table", + "html": "
Number of shots (S)124816
ModelsInitial Loss (L0)
- Random Initialization21.4521.4221.3121.2721.32
Proposed Initialization1.601.541.471.351.21
ModelsInitial test accuracy (Acc0)
- Random Initialization17.7917.9218.0318.1018.22
Tip-adapter[28]59.2859.7260.5562.0964.29
Proposed Initialization59.7060.6662.0464.1666.20
", + "image_path": "60fa20257fb8070eb056729891cda278052065e175caee15a7e7a61259f23942.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Computational overhead. As the literature on adapting VLMs is gaining popularity, it is essential to evaluate the extent to which novel methods are efficient. To do this, we report the overall computational overhead of the approaches" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 242, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 363 + ], + "type": "text", + "content": "studied in this work, which includes the time required for training and for finding the hyper-parameters, when applicable. We also indicate whether these methods enable black-box adaptation which, in our perspective, is a critical aspect in novel strategies aiming to address practical, real-world demands. The numerical values in Table 4 show that, in addition to yielding state-of-the-art performance (shown in previous sections), " + }, + { + "bbox": [ + 304, + 242, + 547, + 363 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 242, + 547, + 363 + ], + "type": "text", + "content": " is the most efficient method (by several orders of magnitude), and does not require to access the internal representations of the pre-trained models." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 369, + 543, + 491 + ], + "blocks": [ + { + "bbox": [ + 46, + 620, + 288, + 654 + ], + "lines": [ + { + "bbox": [ + 46, + 620, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 620, + 288, + 654 + ], + "type": "text", + "content": "Table 3. Comparison of the initial loss and test accuracy in the training-free scenario: random initialization, Tip-adapter and the proposed initialization. The results are averaged over 11 datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 369, + 543, + 491 + ], + "lines": [ + { + "bbox": [ + 310, + 369, + 543, + 491 + ], + "spans": [ + { + "bbox": [ + 310, + 369, + 543, + 491 + ], + "type": "table", + "html": "
MethodsOverall TimeBlackBox# Parameters
CoOp[29]~ 17hXK × M × D
PLOT-2[4]~ 10hXP × K × M × D
KgCoOp[27]~ 4hXK × M × D
ProGrad[30]~ 20hXK × M × D
Clip-Adapter[9]~ 40min✓2(D1 × D)
Tip-adapter-F[28]~ 6min✓K × S × D
Tip-adapter-F*[28]~ 50min✓K × S × D
Standard LP[23]3min✓K × D
LP++~ 2s✓K(D+1)
", + "image_path": "152978e97519e079815a97ace36414aa048371ee11b4a4c0311acc50c91d2255.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "text", + "content": "Table 4. Run time and suitability to black-box scenarios for different methods on 16-shot ImageNet. All the experiments are performed on a single NVIDIA RTX A6000 GPU, except for PLOT-2, which is evaluated on two A6000 GPUs. " + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "inline_equation", + "content": "D_{1} = 256" + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "inline_equation", + "content": "D = 1024" + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "text", + "content": ". The number of context tokens " + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "text", + "content": " is set to 16. For PLOT, " + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "inline_equation", + "content": "P = 4" + }, + { + "bbox": [ + 304, + 498, + 547, + 565 + ], + "type": "text", + "content": " is the number of prompts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 586, + 379, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 379, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 379, + 597 + ], + "type": "text", + "content": "4. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We introduced " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", a strong linear probe for few-shot CLIP adaptation. A specific modeling of the classifier weights, blending visual prototypes and text embeddings via learnable multipliers, along with convex-optimization ingredients, often overlooked in deep learning practices, led to the surprising results. While the findings of this work do not invalidate the promise of prompt learning and adaptation research, we believe " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LP}++" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " could be used as a baseline to measure progress in these strongly emergent areas." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23780" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Amir Beck and Luba Tetruashvili. On the convergence of block coordinate descent type methods. SIAM Journal on Optimization 23(4), 23(4):2037-2060, 2013. 4, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, pages 446–461, 2014. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "text", + "content": "[3] Sébastien Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends in Machine Learning, 8(3-4):231-358, 2015. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 249 + ], + "type": "text", + "content": "[4] Guangyi Chen, Weiran Yao, Xiangchen Song, Xinyue Li, Yongming Rao, and Kun Zhang. Prompt learning with optimal transport for vision-language models. In International Conference on Learning Representations, 2023. 2, 6, 8, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 295 + ], + "type": "text", + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 351 + ], + "type": "text", + "content": "[6] Pierre Colombo, Victor Pellegrain, Malik Boudiaf, Victor Storchan, Myriam Tami, Ismail Ben Ayed, Celine Hudelot, and Pablo Piantanida. Transductive learning for textual few-shot classification in api-based embedding models. In Empirical Methods in Natural Language Processing, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 453 + ], + "type": "text", + "content": "[8] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Workshop on Computer Vision and Pattern Recognition, pages 178-178, 2004. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 288, + 508 + ], + "type": "text", + "content": "[9] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, pages 1-15, 2023. 2, 6, 8, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 609 + ], + "type": "text", + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "text", + "content": "[12] Steffen Herbold. Autorank: A python package for automated ranking of classifiers. Journal of Open Source Software, 5 (48):2173, 2020. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[13] Mingyi Hong, XiangfengWang, Meisam Razaviyayn, and Zhi-Quan Luo. Iteration complexity analysis of block coordinate descent methods. Mathematical Programming, 163: 85–114, 2017. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[14] Z hong, D Friedman, and D Chen. Factual probing is [mask]: Learning vs. learning to recall. In Conference of the North" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "American Chapter of the Association for Computational Linguistics, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[16] Z Jiang, F Xu, J Araki, and G Neubig. How can we know what language models know. In Association for Computational Linguistics, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In International Conference on Computer Vision, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[18] Kenneth Lange, David R Hunter, and Ilsoon Yang. Optimization transfer using surrogate objective functions. Journal of computational and graphical statistics, 9(1):1-20, 2000. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[19] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv, 2013. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "text", + "content": "[20] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics and Image Processing, pages 722-729, 2008. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[21] Jorge Nocedal. Updating quasi-newton matrices with limited storage. Mathematics of Computation, 35(151):773-782, 1980. 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[22] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3498-3505. IEEE, 2012. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 492 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 1, 3, 6, 8, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 525 + ], + "type": "text", + "content": "[24] T Shin, Logan R. L. IV Razeghi, Y, E Wallace, and S Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[25] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv, 2012. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 613 + ], + "type": "text", + "content": "[26] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3485-3492. IEEE, 2010. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[27] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6757-6767, 2023. 2, 6, 8, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[28] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kun-chang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 1, 2, 5, 6, 7, 8, 3, 4" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23781" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 159 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 6, 8, 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 159 + ], + "type": "text", + "content": "[30] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In International Conference on Computer Vision, 2023. 2, 6, 8, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23782" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_content_list.json b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c0cba4b75eff4d4e9995e0dff2b25b500553981 --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_content_list.json @@ -0,0 +1,1608 @@ +[ + { + "type": "text", + "text": "LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyang Ge $^{1,\\dagger}$ , Qiao Feng $^{1,\\dagger}$ , Hailong Jia $^{1}$ , Xiongzheng Li $^{1}$ , Xiangjun Yin $^{1}$ , You Zhou $^{2}$ , Jingyu Yang $^{1}$ , Kun Li $^{1,*}$", + "bbox": [ + 181, + 178, + 785, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tianjin University, China $^{2}$ Nanjing University, China", + "bbox": [ + 263, + 215, + 709, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ghy0623,fengqiao,jhl,lxz,yinxiangjun,yjy,lik}@tju.edu.cn zhouyou@nju.edu.cn", + "bbox": [ + 135, + 236, + 826, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/271240214a8cb7c489998d4b4fadf00997a8b09d5b26bde665705f970bd6be0b.jpg", + "image_caption": [ + "Figure 1. LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging. We contribute a framework for estimating human poses and shapes from individual lensless measurements. The first row shows the input measurements acquired by our lensless imaging system, the second row shows the estimated human poses and shapes from lensless measurements, and the bottom row shows the 3D results shown in different views." + ], + "image_footnote": [], + "bbox": [ + 89, + 297, + 888, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 654, + 313, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human pose and shape (HPS) estimation with lensless imaging is not only beneficial to privacy protection but also can be used in covert surveillance scenarios due to the small size and simple structure of this device. However, this task presents significant challenges due to the inherent ambiguity of the captured measurements and lacks effective methods for directly estimating human pose and shape from lensless data. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to our knowledge. We specifically design a multi-scale lensless feature decoder to decode the lensless measurements through the optically encoded mask", + "bbox": [ + 75, + 686, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for efficient feature extraction. We also propose a doublehead auxiliary supervision mechanism to improve the estimation accuracy of human limb ends. Besides, we establish a lensless imaging system and verify the effectiveness of our method on various datasets acquired by our lensless imaging system. The code and dataset are available at https://cic.tju.edu.cn/faculty/likun/projects/LPSNet.", + "bbox": [ + 500, + 656, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 797, + 630, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, lensless imaging technologies [6, 7, 19, 25] have advanced significantly due to their many advantages, such as privacy protection, smaller size, simple structure, and lower cost. 3D human pose and shape (HPS) estimation [11, 14, 34, 35] requires miniaturized and", + "bbox": [ + 500, + 825, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 101, + 875, + 214, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 102, + 888, + 230, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1471", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lightweight imaging system as application scenarios become more diverse. All these advantages, especially privacy, make the lensless imaging system very suitable as an imaging device for human pose and shape estimation. In this work, we propose LPSNet, which aims to estimate 3D human pose and shape from lensless measurements instead of RGB images, achieving cheaper and privacy-protecting 3D human pose and shape estimation.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A thin, lightweight, and potentially cost-effective optical encoder is used in a lensless imaging system instead of traditional cameras with lenses, while others are expensive, rigid, and occupy more space. At this stage, the application of a lensless imaging system is extensive, it is mainly used in the field of microscopic imaging, RGB image reconstruction, and so on. More valuable information can be obtained from lensless measurements due to the special optical encoding method of lensless imaging systems. Directly estimating human pose and shape from lensless measurements is not currently possible. The first step is to reconstruct an RGB image from a lensless measurement and then estimate the human pose and shape from the RGB images. However, experimental findings indicate that the reconstructed RGB images are of suboptimal quality, resulting in incomplete local features and significant deviations in the position of the human body. Combining these factors leads to inaccurate human pose estimation when using lensless measurements to reconstruct RGB images. This approach has computational burden and computational resources, making it very unfavorable for deployment at the endpoint.", + "bbox": [ + 75, + 213, + 472, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, We aim to advance human pose and shape estimation using a lensless imaging system, which needs to overcome two main challenges. First, how to extract features from lensless measurements for human pose and shape estimation. Secondly, during early experiments, when using features extracted from lensless measurements to estimate human pose and shape, we found poor estimation accuracy of human limbs.", + "bbox": [ + 75, + 532, + 472, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce LPSNet, the first end-to-end human pose and shape estimation framework with lensless imaging. To extract features from optically encoded lensless measurements, we propose a multiscale lensless feature decoder (MSFDecoder). Specifically, we introduce a global perception layer to enhance the global decoding capability of MSFDecoder. The global information that has been optically encoded to the global can be efficiently decoded to obtain a feature map that can be used in subsequent processes. To improve pose and shape estimation, we propose a Double-Head Auxiliary Supervision(DHAS) mechanism to be implemented during training. Auxiliary supervision can improve the estimation accuracy of human limbs and correct results with large deviations.", + "bbox": [ + 75, + 655, + 472, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows: 1) We propose LPSNet, an end-to-end pose and shape esti", + "bbox": [ + 68, + 868, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mation network for lensless imaging systems. This is the first work to estimate human poses and shapes directly from lensless measurements.", + "bbox": [ + 511, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) We propose MSFDecoder, a Multi-Scale Lensless Feature Decoder that decodes and extracts features from lensless measurements, which can be receptive to global features in lensless measurements.", + "3) We propose a Double-Head Auxiliary Supervision mechanism for both pose and shape estimation, which can improve the estimation accuracy of human limbs." + ], + "bbox": [ + 491, + 136, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 258, + 642, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lensless Imaging System", + "text_level": 1, + "bbox": [ + 500, + 284, + 728, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A conventional photographic camera typically comprises a focusing lens, which may consist of one or more optical elements and an image sensor positioned at or near the focal length of the lens. The lens in such a camera directs the projected light from the observed scene onto the sensor, aiming to accurately map specific scene points to individual pixels on the sensor. Conversely, in a lensless imaging system, the absence of a lens defines its configuration. Instead, an optical modulator, such as a coded amplitude mask or a diffuser, is positioned between the scene and the image sensor, often close to the sensor itself. As a result, the recorded data deviates significantly from the expected RGB image during imaging. In this process, the local information of the object transforms overlapping global information through the optically encoded mask.", + "bbox": [ + 496, + 308, + 890, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within the mask-modulated lensless systems, a fixed optical mask is introduced to create a versatile lensless system that can work for a wide range of object distances and lighting scenarios, whether passive or uncontrolled. The mask modulates the incoming light and generates a measurement that can be decoded through computational methodologies. Mask-modulated lensless imaging systems were used to perform 2D imaging [4, 7, 19], refocusing [7], 3D imaging [2, 7], and microscopic imaging [1, 2, 7].", + "bbox": [ + 496, + 536, + 890, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generally speaking, amplitude modulators and phase modulators [1, 4] are the two types of optical masks used in lensless imaging systems. Phase modulators can be further sub-categorized into phase gratings [30, 31], diffusers [2], and phase masks [7]. One key characteristic of a mask-modulated lensless system is the pattern the mask produces on the sensor for a point light source in the scene. We call this pattern the Point-Spread Function(PSF), and its properties determine the imaging model of the system. As shown in Fig. 2, we design a simple mask-modulated lensless system for data acquisition in this experiment. The lensless imaging system designed in our experiments chose a diffuser as the mask.", + "bbox": [ + 496, + 672, + 890, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main advantages of lensless imaging are as follows: lensless is small in size and can be assembled in a variety of", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1472", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f33d35d1c12c6238da0bf12832073986d998c84e2c189df320c8054ad2cf087d.jpg", + "image_caption": [ + "Figure 2. The workflow of the lensless imaging system and the final measurement obtained is the result obtained from the encoding of 3D scene information by the lensless imaging system. The optically encoded mask transforms local information in the 3D scene into overlapping global information." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 464, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "miniaturized equipment; lensless is light and easy to carry; lensless cameras are still cheap to make; lensless cameras have a wide view field, larger than traditional wide-angle cameras; It can also protect the user's privacy to a great extent.", + "bbox": [ + 75, + 306, + 470, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human Pose and Shape Recovery", + "text_level": 1, + "bbox": [ + 76, + 393, + 372, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are two primary methods for human pose and shape estimation: optimization-based approaches and regression-based approaches. Numerous methodologies have been developed to estimate the three-dimensional human pose and shape by employing iterative optimization techniques, e.g., SMPLify [5] and variations of the SMPLify [20, 26, 29, 33].", + "bbox": [ + 75, + 417, + 468, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many existing Regression-based methods follow the architecture of HMR [14], which uses a pre-trained backbone to extract image features followed by regression to obtain SMPL [5] parameters. Many improvements to the original method have been proposed since its introduction. In particular, many papers have proposed alternative methods for pseudo-ground truth generation, including using temporal information [3], multiple views [21], or iterative optimization [13, 17, 27]. SPIN [17] proposed an in-the-loop optimization that incorporated SMPLify [5] in the HMR training, which combines the optimization-based approach and the regression-based approach for human pose and shape estimation. PyMAF [35, 36] proposes a pyramidal mesh alignment feedback loop for regression of SMPL parameters. Our approach references the design of PyMAF in the part of the SMPL parameters regression. PARE [16] proposes a body-part-guided attention mechanism for better occlusion handling. HKMR [10] performs a prediction guided by the hierarchical structure of SMPL. HMR2.0 [11] employs a large training dataset and proposes a fully transformer-based approach for 3D human pose and shape estimation from a single image. Many related approaches make non-parametric predictions, i.e., instead of estimating the parameters of the SMPL model, they explicitly regress the vertices of the mesh. GraphCMR [18] uses a", + "bbox": [ + 75, + 523, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "graph neural network for the prediction, METRO [23] and FastMETRO [8] use a transformer, while Mesh Graphormer [24] adopts a hybrid between the two.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we propose the first work to perform end-to-end human pose and shape estimation from lensless measurements. We designed a multi-scale lensless feature decoder to decode the lensless measurements based on a mask-encoder to obtain more efficient features. We also propose a double auxiliary supervision mechanism to improve the estimation accuracy of human limbs.", + "bbox": [ + 496, + 137, + 890, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 498, + 257, + 591, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 282, + 609, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work focuses on human pose and shape estimation from lensless measurements. In this section, we present the technical details of our approach. As depicted in Fig. 3, the core of our method involves the following three components: 1) a Multi-Scale Lensless Feature Decoder(MSFDecoder) (Sec. 3.2) that can effectively decode the information encoded by the lensless imaging system; 2) a human parametric model regressor (Sec. 3.3) that takes the multi-scale features produced by MSFDecoder as input and predicts the SMPL parameters; 3) a Double-Head Auxiliary Supervision mechanism(DHAS) (Sec. 3.4) that can assist the LPSNet to improve the estimation accuracy of human limbs.", + "bbox": [ + 496, + 306, + 890, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Scale Lensless Feature Decoder", + "text_level": 1, + "bbox": [ + 498, + 513, + 823, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, the optical information of the subject is diffused by the mask and projected onto the sensor. The process of obtaining measurements through the lensless imaging system can be treated as encoding optical information in the scene into lensless measurements. The goal of our lensless feature decoder $D_{M}$ is to decode these lensless measurements to multi-scale features, which are then utilized for subsequent human body prediction. The design of the global perception layer $G_{P}$ is inspired by HRNet [32]. The global perception layer inherits many of the advantages of HRNet [32] and can always maintain a high resolution, while the information interaction between different branches complements the information loss caused by the reduction in the number of channels. These advantages are important for extracting features from lensless measurements. Then, we design different convolutional layers for different scales of features that are used to refine the features.", + "bbox": [ + 496, + 537, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, the decoder takes a lensless measurement $M$ as input and decodes the information encoded by the lensless imaging system. The global perception layer ensures the integrity of global information for each pixel by maintaining the high-resolution convolution branch and the low-resolution convolution branch in parallel. It further en", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1473", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/748807a183ac6e648e492574ea5e5f7ac01516841e483c4af533ddd54b82fd02.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed framework. A measurement $M$ is passed through a Multi-Scale Lensless Feature Decoder to get spatial characteristics at different scales. These feature maps are fed into the regressor for human pose and shape estimation. Also, these feature maps are fed into the Double-Head Auxiliary Supervision to improve the estimation accuracy." + ], + "image_footnote": [], + "bbox": [ + 106, + 94, + 869, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hances the information extraction accuracy by repeating the multi-scale fusion across parallel convolutions. As shown in Fig. 3, the global perception layer takes $M$ as input and generates a set of multi-scale spatial features $\\phi_{m}^{t = 0,1,2,3}$ as input for the subsequent steps. The final output $\\phi_{M}^{t}$ of the decoder is obtained by passing $\\phi_{m}^{t}$ into different convolutional layers,", + "bbox": [ + 75, + 455, + 472, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {M} ^ {t} = D _ {M} (M). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 561, + 468, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With different scales of decoding, we can obtain feature maps at different granularities, through which we can perceive the location of people in the scene. The results obtained by the lensless decoder are utilized for subsequent predictions with the SMPL model, incorporating pose, shape, and camera parameters $\\Theta = \\{\\theta ,\\beta ,\\widetilde{\\pi}\\}$ , where $\\widetilde{\\pi}$ is pseudo-camera parameters for the subsequent projection calculation.", + "bbox": [ + 75, + 582, + 472, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Human Parametric Model Regression", + "text_level": 1, + "bbox": [ + 76, + 710, + 406, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our human parametric model regression is mainly inspired by the improved Human Mesh Regression described in PyMAF [35]. Different from the pyramid features obtained by deconvolution in PyMAF [35], our global perception layer exploits HRNet [14] to maintain more high-resolution information. We use a set of sampling points $S_{t}$ and project them onto the feature maps to extract point-wise features. The point-wise features corresponding to each point $s$ in $S_{t}$ are bilinearly sampled from $\\phi_M^{t=1,2,3}$ . (Note that $\\phi_M^{t=0}$ is unnecessary here and is used as input for subsequent initialization.) Then the dimensions of these features", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "will be reduced by MLP(multi-layer perceptron) and concatenated as a feature vector $\\hat{\\phi}_M^t$ which can be formulated as:", + "bbox": [ + 496, + 455, + 893, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\phi} _ {M} ^ {t} = \\mathcal {F} \\left(\\phi_ {M} ^ {t}, S _ {t}\\right) \\tag {2} \\\\ = \\oplus \\left\\{f \\left(\\phi_ {M} ^ {t} (s)\\right), \\text {f o r} s \\text {i n} S _ {t} \\right\\}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 507, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(\\cdot)$ denotes the overall point-wise feature extraction, $f(\\cdot)$ is the MLP layer, and $\\oplus$ denotes the concatenation. We use an iterative mechanism to complete the interaction of different scale features. Sequentially, a parameter regressor $\\mathcal{R}_t$ takes features $\\hat{\\phi}_M^t$ and the current parameter estimation $\\Theta_t$ as inputs and outputs the parameter residual. Parameters are updated as $\\Theta_{t + 1}$ by adding the residual to $\\Theta_t$ which can be formulated as:", + "bbox": [ + 496, + 555, + 893, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {t + 1} = \\Theta_ {t} + \\mathcal {R} _ {t} \\left(\\Theta_ {t}, \\hat {\\phi} _ {M} ^ {t}\\right), \\text {f o r} t > 0. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 684, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The initial parameter $\\Theta_0$ is initialized by feeding $\\phi_M^{t=0}$ into the regressor $\\mathcal{R}_t$ .", + "bbox": [ + 496, + 718, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the predicted parameters $\\Theta$ (the subscript $t$ is omitted for simplicity) are obtained through each scale feature, a mesh with vertices of $V = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{N\\times 3}$ can be generated accordingly, where $N = 6890$ is the number of vertices in the SMPL model. We downsample $V$ to get the sampling point $S_{t + 1}$ for the next iteration. These mesh vertices can be mapped by pre-trained linear regressor to obtain sparse 3D keypoints $J\\in \\mathbb{R}^{N_j\\times 3}$ . With the estimated pseudo-camera parameters, we can obtain 2D keypoints $K\\in \\mathbb{R}^{N_j\\times 2}$ by projecting $J$ on the measurement coordinate system. We", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1474", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "calculate the loss between the 2D keypoints and the ground truth, in this way reducing the difference between the 2D projection and the human pose in the real scene. Concurrently, supplementary 3D supervisions regarding 3D keypoints and model parameters are integrated when authentic 3D labels are available. The total loss function for the parameter regressor is thereby formulated as follows:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {2 d} \\| K - \\hat {K} \\| _ {2} + \\lambda_ {3 d} \\| J - \\hat {J} \\| _ {2} + \\lambda_ {\\text {p a r a}} \\| \\Theta - \\hat {\\Theta} \\| _ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 218, + 468, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ is the squared L2 norm, $\\hat{K},\\hat{J}$ , and $\\hat{\\Theta}$ denote the ground truth 2D keypoints, 3D keypoints, and model parameters, respectively.", + "bbox": [ + 76, + 244, + 468, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Double-Head Auxiliary Supervision", + "text_level": 1, + "bbox": [ + 76, + 299, + 387, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spatial feature map of the human body is relatively rough and contains a lot of noise. Therefore, there are still some deviations in the perception of human limbs and pose. To improve our estimation accuracy of human limbs, we introduce a Double-Head Auxiliary Supervision (DHAS) mechanism to obtain finer spatial features during the training stage.", + "bbox": [ + 75, + 321, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first convert all the spatial features at different scales to a single feature map $\\phi^T$ by upsampling and then concatenating them together. The feature map $\\phi^T$ is used for the different auxiliary supervision heads. On the one hand, we generate a heat map representation by a classification layer to explicitly indicate the position of the 2d keypoints. On the other hand, we also estimate a dense map through the IUV Predict layer to get a dense correspondence between the SMPL model and the feature map. The loss function for the Double-Head Auxiliary Supervision consists of two parts, which can be written as:", + "bbox": [ + 75, + 428, + 468, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d a s} = \\mathcal {L} _ {s c} + \\mathcal {L} _ {d e n}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 602, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Keypoints Supervision. We utilize the SimCC-based algorithm [22] for predicting pose keypoints. This approach treats keypoint localization as a classification task in horizontal and vertical coordinates. During training, instead of estimating the actual coordinates, we employ two vectors, $x$ and $y$ , and convert the ground truth 2D keypoints into such vectors to compute the loss. The loss function is then formulated as follows:", + "bbox": [ + 75, + 626, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s c} = \\lambda_ {x y} (\\operatorname {K L - L o s s} (x, \\hat {x}) + \\operatorname {K L - L o s s} (y, \\hat {y})), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 755, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where KL-Loss is the Kullback-Leibler divergence loss, $\\hat{x}$ and $\\hat{y}$ denote the processed ground truth 2D keypoints, respectively.", + "bbox": [ + 75, + 779, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "IUV Supervision. We adopt the IUV mapping defined in DensePose [12] as the dense correspondence representation. This mapping establishes deterministic correspondences between foreground pixels in 2D images and vertices on 3D surfaces. The vertices on the template mesh can", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9801a94d8c8e0d9e7d8fcfba90bb98213ffae6ac37d4f64c99ed86cd429ff017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 93, + 681, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6803f5ecdd031d34dedaab0371fc523d820aaea6c37e86c1db0459a15c6f9d48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 93, + 852, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0445d025768d4be19211f97d3cf9d98e13239ea6bdafafd1a96579cdfbbd7f01.jpg", + "image_caption": [ + "Figure 4. On the top left is a figure of the dual camera system we created. This system consists of an RGB camera and a lensless imaging system using a diffuser as a mask. The top right shows the process of collecting our dataset. The image at the bottom is a record of the process of collecting our dataset." + ], + "image_footnote": [], + "bbox": [ + 519, + 226, + 857, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "be mapped back to pixels on the foreground using a predefined bijection mapping between 3D surface space and 2D UV space. The dense correspondence representation comprises the index of body parts $P$ and UV values of mesh vertices. During training, we apply classification loss to the index of body parts $P$ and regression loss to the UV channels of the dense correspondence mapping. The loss function is formulated as follows:", + "bbox": [ + 498, + 468, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {d e n} = \\lambda_ {p i} \\text {C r o s s E n t r o p y} (P, \\hat {P}) \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (U, \\hat {U}) \\tag {7} \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (V, \\hat {V}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 598, + 890, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{P},\\hat{U}$ , and $\\hat{V}$ denote the ground truth of $P$ and $UV$ values, respectively.", + "bbox": [ + 498, + 671, + 888, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 715, + 632, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Imaging System and Dataset", + "text_level": 1, + "bbox": [ + 500, + 741, + 754, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lensless Imaging System. Due to the scarcity of human pose and shape datasets based on lensless imaging systems, we developed a basic lensless system to gather experimental data. The system, illustrated in Fig. 4, comprises two primary components: a lensless imaging system for capturing lensless measurements and an RGB camera for capturing real images corresponding to the lensless measurements. To ensure identical light paths, we employed a beam splitter between the lensless sensor and the RGB camera. This al", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1475", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lowed for consistent imaging conditions and facilitated accurate correlation between the lensless measurements and the corresponding real images.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the lensless imaging part of the system, we adopted the Mask-Modulated Lensless System, inspired by design options such as diffusercam [2] and phlatcam [7]. Specifically, we selected a diffuser as our mask and incorporated PhlatCam's concept of using larger sensors to enhance global information perception.", + "bbox": [ + 75, + 137, + 467, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. The inputs to LPSNet are lensless measurements. However, datasets for classical human pose estimation are not directly available at present. Leveraging the lensless imaging system we have constructed, and the mathematical model of the system's imaging, the sources of datasets for our experiments can be categorized as follows:", + "bbox": [ + 75, + 229, + 467, + 318 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Real Dataset. Using a lensless imaging system to capture images displayed on a screen as measurements is currently the primary method of acquiring datasets in the lensless field. We used this approach to collect datasets from various sources, including Human3.6M, MPII, COCO, 3DPW, and MIP-INF-3DHP datasets, which we refer to as LenslessHuman3.6M, LenslessMPII, LenslessMIP-INF-3DHP, LenslessCOCO, and Lensless3DPW, respectively. We also reclassified the training dataset and validation dataset on LenslessHuman3.6M, referred to as train-LenslessHuman3.6M and eval-LenslessHuman3.6M, respectively. Moreover, we capture the real scenes of different individuals, backgrounds, and light intensities using our lensless imaging system.", + "- Simulated Dataset. The mathematical imaging model enables the conversion (simulation) of images captured by RGB cameras into measurements captured by lensless imaging systems. This technique is commonly employed in lensless imaging fields for testing systems and imaging methods. Leveraging this, we converted numerous human pose datasets into simulated lensless measurements. These simulated datasets extended our training set and provided a quick validation of our method." + ], + "bbox": [ + 76, + 321, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 695, + 292, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The dimensions of the raw measurements are $1280 \\times 1024 \\times 3$ . Prior to commencing the experiments, we first crop and resize the lensless measurements. These measurements are pre-processed to $224 \\times 224 \\times 3$ before being fed into the network. Following this pre-processing step, the lensless measurements inputted into the global perceptual layer are $224 \\times 224$ , generating spatial features at resolutions of $(7 \\times 7, 14 \\times 14, 28 \\times 28$ , and $56 \\times 56$ ). During the generation of mesh-aligned features, the SMPL mesh is downsampled using a pre-computed downsampling matrix provided in [18], resulting in a reduction of vertex count from 6890 to 431. The regressors $\\mathcal{R}_t$ follow a congruent architec", + "bbox": [ + 75, + 719, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tural framework to the regressor present in HMR [14], albeit with slight variations in their input dimensions. Our network is trained using the Adam optimizer [15] with a learning rate of $5 \\times 10^{-5}$ and a batch size of 80 on 4 NVIDIA RTX2080 Ti GPUs. No learning rate decay is applied during training.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f10a5d6cc35d3b0e92422cc57562cb3b3533e990531cbefb7ae6fffe8635be20.jpg", + "image_caption": [ + "Measurement" + ], + "image_footnote": [], + "bbox": [ + 517, + 194, + 602, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5eeaec79778a14f66d746fcf19acd3b4ae88bf9126fafba5b87e128b4af946b3.jpg", + "image_caption": [ + "Recovered Images" + ], + "image_footnote": [], + "bbox": [ + 602, + 194, + 725, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b9aacade28b5169d1a0e2d5447ace8304ee8a5ab49ac919afff262163acbda5.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 727, + 217, + 772, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e19965037832bce6adaca5cb3ad8a85a14729c588f3e9e835f29d4f89c3f7eac.jpg", + "image_caption": [ + "Human Mesh Recovery" + ], + "image_footnote": [], + "bbox": [ + 777, + 195, + 861, + 261 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1be17e5c35525bb6a809b917a0530590be269c914387c5e55980e3f238b6e905.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 287, + 598, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f521a29a4e3abb0e3ca32acdfd25bca7d99f504db9a28724ad8dbb48241de59f.jpg", + "image_caption": [ + "Figure 5. At the top is the frame of our baseline. Lensless measurements first recover the image by reconstruction methods and then estimate human pose and shape by PyMAF. The bottom image compares the recovered image with the original image, and we can see that the quality of the recovered image has dropped dramatically." + ], + "image_footnote": [], + "bbox": [ + 511, + 357, + 596, + 424 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf26471fafe17c92703833a47292d4f15ccf8a577406bdae22d1e104eb64d730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 289, + 687, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0f22bd1a366a8aaa59b5fb49bb95b53d26a98ade71f577d16c8fafa0b301db25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 357, + 687, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7829f64eac7cc3ab9c70aab0829a6a5f121027469854164cecab124c105cd1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 289, + 779, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9613918b8caf996d178a942928a939ffc9ec0aade123b8a09d9d40230d566ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 357, + 779, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e2d8f3b8ef91ef15fd15f21d7f56d7e6ddf9685713f5bc0b07eeaac1168f74b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 289, + 867, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e66eea2bb41f95771ebeddc45b47e83b50fddb77e6ed99e46c5a8a705636c94e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 357, + 867, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. 3D Human Pose and Shape Comparison.", + "text_level": 1, + "bbox": [ + 500, + 555, + 846, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Approach. We are the first work to perform human pose and shape estimation through lensless measurements. Given the absence of other valid methods for comparison, we designed a baseline approach that approaches this as a two-stage task. The baseline architecture is depicted in Fig. 5. In the first stage, we utilized the lensless image reconstruction method Rego et. al. [28], chosen for its wide applicability in recovering images acquired by lensless imaging systems. In the second stage, PyMAF is used to estimate human pose and shape.", + "bbox": [ + 496, + 578, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff9fc91372a3f861bf7d84d8fa2e94b53d834432cca3e799b264b7cd59556b07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
baseline (PyMAF)257.07121.73277.18
baseline (PyMAF†)126.2881.37151.60
LPSNet119.2081.52134.74
", + "bbox": [ + 513, + 742, + 879, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with baseline. The baseline is a combination of the two-stage approach using existing methods. LPSNet outperforms the baseline on the LenslessHuman3.6M datasets.", + "bbox": [ + 496, + 825, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Baseline. We present the results of", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1476", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9fdeca129a39fd72df7420174d49a676f6146113e4c4666c9db1b7a60dc150a1.jpg", + "image_caption": [ + "Figure 6. Qualitative results of LPSNet on challenging LenslessHuman3.6M. The results on the left are from our LPSNet estimation, those in the center are from our baseline (PyMAF) output, and those on the right are from the fine-tuned baseline $(\\mathrm{PyMAF}^{\\dagger})$ output. LPSNet can be seen to be better than the baselines." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 869, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the quantitative comparative evaluation on the LenslessHuman3.6M dataset in Tab. 1. Notably, images recovered from lensless measurements exhibit lower quality. Directly comparing the baseline (PyMAF) with our approach is challenging and unfair. Therefore, we fine-tuned PyMAF on the train-LenslessHuman3.6M dataset, denoted as $\\mathrm{PyMAF}^{\\dagger}$ . Referring to Tab. 1, we observe significant improvements in the indicators of baseline $(\\mathrm{PyMAF}^{\\dagger})$ . Specifically, compared with the baseline $(\\mathrm{PyMAF}^{\\dagger})$ , the MPJPE of LPSNet on the LenslessHuman3.6M dataset is reduced by $7.08~\\mathrm{mm}$ .", + "bbox": [ + 75, + 467, + 468, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 1, our LPSNet achieved competitive results compared to the baseline approach $(\\mathrm{PyMAF}^{\\dagger})$ . LPSNet demonstrates more significant improvements in the MPJPE and PVE metrics. However, we would argue that the PA-MPJPE metric may not fully reveal the performance of the mesh-image alignment, as it is calculated as MPJPE after rigid alignment.", + "bbox": [ + 75, + 619, + 470, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline (PyMAF) has difficulty performing accurate pose estimation on low-quality recovered images. However, the outputs of our LPSNet exhibit significant improvements compared to the two baseline methods, especially for limb estimation. Qualitative comparison results are illustrated in Fig. 6.", + "bbox": [ + 75, + 726, + 468, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 830, + 232, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies on LenslessHuman3.6M under various settings to validate the effectiveness of the key components proposed in our method.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d33fadc890a82ec775392d69f16adfbb0e40ca65fcc593cf2e1a24b482c69dd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
w/o MSFDcoder and DHAS142.1392.20161.07
w/o DHAS139.3492.56158.97
w/o DHAS (2D keypoint)123.3083.41138.67
w/o DHAS (IUV)129.7086.83146.7
LPSNet119.2081.52134.74
", + "bbox": [ + 501, + 464, + 893, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on LPSNet. DHAS stands for Double-Head Auxiliary Supervision.", + "bbox": [ + 498, + 568, + 890, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation variants were trained and tested on LenslessHuman3.6M, which is derived from the Human3.6M dataset. The Human3.6M dataset includes ground-truth 3D labels and serves as a widely used benchmark in 3D human pose and shape estimation.", + "bbox": [ + 496, + 626, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Scale Lensless Feature Decoder. In LPSNet, the decoder primarily decodes the global information of lensless measurements encoded by the diffuser, which plays a crucial role in our end-to-end system. Our MSFDecoder is essential for efficient feature extraction from lensless measurements. We have adapted another variant of the decoder to verify this. Specifically, we simplified the global perception layer of the decoder to a series of convolutional layers and deconvolutional layers. We use this simpler decoder in the experiment(w/o MSFcoder and DHAS) instead of MSFDecoder. As shown in Tab. 2, comparing the result of experiments(w/o MSFcoder and DHAS) and experiment(w/o DHAS), we can see that MSFDecoder improves", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1477", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfebda61820c178f7005e59c8a1ab75ea8e5f86091cf4d09398024c74520606b.jpg", + "image_caption": [ + "Figure 7. Qualitative results of LPSNet on challenging LenslessHuman3.6M. We can see that the results of LPSNet are better, and the results are improved with the addition of the double-head assisted supervision mechanism." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 843, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the accuracy of the human pose and shape estimation significantly. Through the qualitative experiment in Fig. 7, we can also see that the human pose and shape estimation has a better alignment after using the MSFDecoder. Note that neither experiment(w/o MSFDcoder and DHAS) nor the experiment(w/o DHAS) used auxiliary supervision.", + "bbox": [ + 75, + 366, + 468, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Double-Head Auxiliary Supervision. Double-Head Auxiliary Supervision is mainly used to improve the accuracy of human limbs. For each auxiliary supervisory head, we conducted ablation experiments respectively. Referring to the result in Tab. 2, LPSNet has significant improvements in the MPJPE and the PVE metrics. Fig. 7 shows additional qualitative comparison results. We can see that the alignment of the human pose and shape is better with the addition of the Double-Head Auxiliary Supervision mechanism.", + "bbox": [ + 75, + 455, + 470, + 606 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3ad17b93f94cb2ac472607d0671de20306ff76a0ae25718cf88b2ebaec0eaaa.jpg", + "image_caption": [ + "(a) Result on more datasets" + ], + "image_footnote": [], + "bbox": [ + 81, + 617, + 287, + 715 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88a8798d1589f4b5ff5a56bdf252bba4306038000f0154014599a59b5e431424.jpg", + "image_caption": [ + "(b) Result on real scenes", + "Figure 8. Results on more datasets and real scenes. From left to right for each set of figures: lensless measurement, alignment of results with RGB images, and 3D results shown in different views." + ], + "image_footnote": [], + "bbox": [ + 292, + 621, + 465, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Results on more Datasets and Real Scene", + "text_level": 1, + "bbox": [ + 76, + 801, + 426, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 (a) shows the qualitative results on various datasets after using the mixed datasets for training. Specifically, we collected the MPII, COCO, 3DPW, MIP-INF-3DHP, and LSP datasets through our lensless imaging system and mixed them for training and testing.", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we also capture the real scenes of two different people, in different backgrounds and different light intensities using our lensless imaging system. Fig. 8 (b) illustrates the experimental results of our method on these real scenes, indicating its suitability for real-world applications.", + "bbox": [ + 498, + 366, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 500, + 473, + 746, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to the best of our knowledge. Specifically, we design a multi-scale lensless feature decoder, which can effectively decode the information produced by the lensless imaging system. We also propose a double auxiliary supervision mechanism to improve the accuracy of human limb end estimation. Experimental results show that our method can achieve end-to-end human pose and shape estimation through lensless measurements.", + "bbox": [ + 496, + 500, + 890, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. Our work is an initial stride step in this direction, however, the results are still not as robust as those achieved by traditional methods for estimating the human body from RGB images. The current scarcity of lensless datasets hampers our ability to pre-train the MSFDecoder using a large dataset like ImageNet [9], thereby limiting the generalizability of our approach. We will continue this work in the future, focusing on generating more extensive lensless datasets and progressing towards practical applications.", + "bbox": [ + 496, + 651, + 890, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Key R&D Program of China (2023YFC3082100), the National Natural Science Foundation of China (62122058 and 62171317), and the Science Fund for Distinguished Young Scholars of Tianjin (No. 22JCJQJC00040).", + "bbox": [ + 496, + 801, + 892, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1478", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jesse K Adams, Vivek Boominathan, Benjamin W Avants, Daniel G Vercosa, Fan Ye, Richard G Baraniuk, Jacob T Robinson, and Ashok Veeraraghavan. Single-Frame 3D Fluorescence Microscopy with Ultraminiature Lensless FlatScope. Sci Adv., 3(12):e1701548, 2017. 2", + "[2] Nick Antipa, Grace Kuo, Reinhard Heckel, Ben Mildenhall, Emrah Bostan, Ren Ng, and Laura Waller. DiffuserCam: Lensless Single-exposure 3D Imaging. Optica, 5(1):1-9, 2018. 2, 6", + "[3] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting Temporal Context for 3D Human Pose Estimation in the ild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3395-3404, 2019. 3", + "[4] M Salman Asif, Ali Ayremlou, Aswin Sankaranarayanan, Ashok Veeraraghavan, and Richard G Baraniuk. Flatcam: Thin, Lensless Cameras using Coded Aperture and Computation. IEEE Trans Comput Imaging., 3(3):384-397, 2016. 2", + "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In *Eur. Conf. Comput. Vis. Springer International Publishing*, 2016. 3", + "[6] Vivek Boominathan, Jesse K Adams, M Salman Asif, Benjamin W Avants, Jacob T Robinson, Richard G Baraniuk, Aswin C Sankaranarayanan, and Ashok Veeraraghavan. Lensless Imaging: A Computational Renaissance. IEEE Signal Process Mag., 33(5):23-35, 2016. 1", + "[7] Vivek Boominathan, Jesse K Adams, Jacob T Robinson, and Ashok Veeraraghavan. Phlatcam: Designed Phase-Mask Based Thin Lensless Camera. IEEE Trans. Pattern Anal. Mach. Intell., 42(7):1618-1629, 2020. 1, 2, 6", + "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In Eur. Conf. Comput. Vis., pages 342–359. Springer, 2022. 3", + "[9] J. Deng, W. Dong, R. Socher, L. Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPR Workshops). 8", + "[10] Georgios Georgakis, Ren Li, Srikrishna Karanam, Terrence Chen, Jana Košecka, and Ziyan Wu. Hierarchical Kinematic Human Mesh Recovery. In Eur. Conf. Comput. Vis., pages 768-784. Springer, 2020. 3", + "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and Tracking Humans with Transformers. In Int. Conf. Comput. Vis., pages 14783-14794, 2023. 1, 3", + "[12] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense Human Pose Estimation in the Wild. In Int. Conf. Comput. Vis., pages 7297-7306, 2018. 5", + "[13] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. In Int. Conf. 3D. Vis., pages 42-52. IEEE, 2021. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end Recovery of Human Shape and Pose. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 1, 3, 4, 6", + "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[16] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. PARE: Part Attention Regressor for 3D Human Body Estimation. In Int. Conf. Comput. Vis., pages 11127-11137, 2021. 3", + "[17] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to Reconstruct 3D Human Pose and Shape Via Model-fitting in the Loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 3", + "[18] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional Mesh Regression for Single-Image Human Shape Reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4501–4510, 2019. 3, 6", + "[19] Grace Kuo, Nick Antipa, Ren Ng, and Laura Waller. DiffuserCam: Diffuser-Based Lensless Cameras. In Computational Optical Sensing and Imaging, pages CTu3B-2. Optica Publishing Group, 2017. 1, 2", + "[20] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 3", + "[21] Vincent Leroy, Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, and Grégory Rogez. SMPLy Benchmarking 3D Human Pose Estimation in the wild. In Int. Conf. 3D. Vis., pages 301-310. IEEE, 2020. 3", + "[22] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation. In Eur. Conf. Comput. Vis., pages 89-106. Springer, 2022. 5", + "[23] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1954-1963, 2021. 3", + "[24] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh Graphormer. In Int. Conf. Comput. Vis., pages 12939-12948, 2021. 3", + "[25] Aydogan Ozcan and Euan McLeod. Lensless Imaging and Sensing. Annual review of biomedical engineering, 18:77-102, 2016. 1", + "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3", + "[27] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human Mesh Recovery from Multiple Shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 3", + "[28] Joshua D Rego, Karthik Kulkarni, and Suren Jayasuriya. Robust Lensless Image Reconstruction Via PSF Estimation. In" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1479", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 403-412, 2021. 6", + "[29] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D Human Motion Model for Robust Pose Estimation. In Int. Conf. Comput. Vis., 2021. 3", + "[30] David G Stork and Patrick R Gill. Lensless Ultra-miniature CMOS Computational Imagers and Sensors. Proc. Sensorcomm, pages 186-190, 2013. 2", + "[31] David G Stork and Patrick R Gill. Optical, mathematical, and computational foundations of lensless ultra-miniature diffractive imagers and sensors. International Journal on Advances in Systems and Measurements, 7(3):4, 2014. 2", + "[32] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep High-resolution Representation Learning for Human Pose Estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5693-5703, 2019. 3", + "[33] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. PoseNDF: Modeling Human Pose Manifolds with Neural Distance Fields. In Eur. Conf. Comput. Vis., 2022. 3", + "[34] Hongwen Zhang, Jie Cao, Guo Lu, Wanli Ouyang, and Zhenan Sun. Learning 3D Human Shape and Pose from Dense Body Parts. IEEE Trans. Pattern Anal. Mach. Intell., 44(5):2610-2627, 2020. 1", + "[35] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Int. Conf. Comput. Vis., pages 11446-11456, 2021. 1, 3, 4", + "[36] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3D Human Dynamics from Video. In Int. Conf. Comput. Vis., pages 7114-7123, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 559 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1480", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_model.json b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..21c18aeaa74d7f5b013767709f814f63f623d91d --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.179, + 0.786, + 0.217 + ], + "angle": 0, + "content": "Haoyang Ge\\(^{1,\\dagger}\\), Qiao Feng\\(^{1,\\dagger}\\), Hailong Jia\\(^{1}\\), Xiongzheng Li\\(^{1}\\), Xiangjun Yin\\(^{1}\\), You Zhou\\(^{2}\\), Jingyu Yang\\(^{1}\\), Kun Li\\(^{1,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.217, + 0.71, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Tianjin University, China \\(^{2}\\)Nanjing University, China" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.237, + 0.828, + 0.253 + ], + "angle": 0, + "content": "{ghy0623,fengqiao,jhl,lxz,yinxiangjun,yjy,lik}@tju.edu.cn zhouyou@nju.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.298, + 0.889, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.587, + 0.893, + 0.643 + ], + "angle": 0, + "content": "Figure 1. LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging. We contribute a framework for estimating human poses and shapes from individual lensless measurements. The first row shows the input measurements acquired by our lensless imaging system, the second row shows the estimated human poses and shapes from lensless measurements, and the bottom row shows the 3D results shown in different views." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.655, + 0.314, + 0.672 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Human pose and shape (HPS) estimation with lensless imaging is not only beneficial to privacy protection but also can be used in covert surveillance scenarios due to the small size and simple structure of this device. However, this task presents significant challenges due to the inherent ambiguity of the captured measurements and lacks effective methods for directly estimating human pose and shape from lensless data. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to our knowledge. We specifically design a multi-scale lensless feature decoder to decode the lensless measurements through the optically encoded mask" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.657, + 0.892, + 0.764 + ], + "angle": 0, + "content": "for efficient feature extraction. We also propose a doublehead auxiliary supervision mechanism to improve the estimation accuracy of human limb ends. Besides, we establish a lensless imaging system and verify the effectiveness of our method on various datasets acquired by our lensless imaging system. The code and dataset are available at https://cic.tju.edu.cn/faculty/likun/projects/LPSNet." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.798, + 0.631, + 0.814 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In recent years, lensless imaging technologies [6, 7, 19, 25] have advanced significantly due to their many advantages, such as privacy protection, smaller size, simple structure, and lower cost. 3D human pose and shape (HPS) estimation [11, 14, 34, 35] requires miniaturized and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.876, + 0.215, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.889, + 0.232, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.876, + 0.232, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "lightweight imaging system as application scenarios become more diverse. All these advantages, especially privacy, make the lensless imaging system very suitable as an imaging device for human pose and shape estimation. In this work, we propose LPSNet, which aims to estimate 3D human pose and shape from lensless measurements instead of RGB images, achieving cheaper and privacy-protecting 3D human pose and shape estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.473, + 0.532 + ], + "angle": 0, + "content": "A thin, lightweight, and potentially cost-effective optical encoder is used in a lensless imaging system instead of traditional cameras with lenses, while others are expensive, rigid, and occupy more space. At this stage, the application of a lensless imaging system is extensive, it is mainly used in the field of microscopic imaging, RGB image reconstruction, and so on. More valuable information can be obtained from lensless measurements due to the special optical encoding method of lensless imaging systems. Directly estimating human pose and shape from lensless measurements is not currently possible. The first step is to reconstruct an RGB image from a lensless measurement and then estimate the human pose and shape from the RGB images. However, experimental findings indicate that the reconstructed RGB images are of suboptimal quality, resulting in incomplete local features and significant deviations in the position of the human body. Combining these factors leads to inaccurate human pose estimation when using lensless measurements to reconstruct RGB images. This approach has computational burden and computational resources, making it very unfavorable for deployment at the endpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.473, + 0.653 + ], + "angle": 0, + "content": "In this paper, We aim to advance human pose and shape estimation using a lensless imaging system, which needs to overcome two main challenges. First, how to extract features from lensless measurements for human pose and shape estimation. Secondly, during early experiments, when using features extracted from lensless measurements to estimate human pose and shape, we found poor estimation accuracy of human limbs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.473, + 0.867 + ], + "angle": 0, + "content": "To address these challenges, we introduce LPSNet, the first end-to-end human pose and shape estimation framework with lensless imaging. To extract features from optically encoded lensless measurements, we propose a multiscale lensless feature decoder (MSFDecoder). Specifically, we introduce a global perception layer to enhance the global decoding capability of MSFDecoder. The global information that has been optically encoded to the global can be efficiently decoded to obtain a feature map that can be used in subsequent processes. To improve pose and shape estimation, we propose a Double-Head Auxiliary Supervision(DHAS) mechanism to be implemented during training. Auxiliary supervision can improve the estimation accuracy of human limbs and correct results with large deviations." + }, + { + "type": "text", + "bbox": [ + 0.07, + 0.869, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows: 1) We propose LPSNet, an end-to-end pose and shape esti" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "mation network for lensless imaging systems. This is the first work to estimate human poses and shapes directly from lensless measurements." + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.137, + 0.894, + 0.196 + ], + "angle": 0, + "content": "2) We propose MSFDecoder, a Multi-Scale Lensless Feature Decoder that decodes and extracts features from lensless measurements, which can be receptive to global features in lensless measurements." + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.197, + 0.894, + 0.243 + ], + "angle": 0, + "content": "3) We propose a Double-Head Auxiliary Supervision mechanism for both pose and shape estimation, which can improve the estimation accuracy of human limbs." + }, + { + "type": "list", + "bbox": [ + 0.493, + 0.137, + 0.894, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.259, + 0.643, + 0.275 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.285, + 0.729, + 0.301 + ], + "angle": 0, + "content": "2.1. Lensless Imaging System" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.892, + 0.536 + ], + "angle": 0, + "content": "A conventional photographic camera typically comprises a focusing lens, which may consist of one or more optical elements and an image sensor positioned at or near the focal length of the lens. The lens in such a camera directs the projected light from the observed scene onto the sensor, aiming to accurately map specific scene points to individual pixels on the sensor. Conversely, in a lensless imaging system, the absence of a lens defines its configuration. Instead, an optical modulator, such as a coded amplitude mask or a diffuser, is positioned between the scene and the image sensor, often close to the sensor itself. As a result, the recorded data deviates significantly from the expected RGB image during imaging. In this process, the local information of the object transforms overlapping global information through the optically encoded mask." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Within the mask-modulated lensless systems, a fixed optical mask is introduced to create a versatile lensless system that can work for a wide range of object distances and lighting scenarios, whether passive or uncontrolled. The mask modulates the incoming light and generates a measurement that can be decoded through computational methodologies. Mask-modulated lensless imaging systems were used to perform 2D imaging [4, 7, 19], refocusing [7], 3D imaging [2, 7], and microscopic imaging [1, 2, 7]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Generally speaking, amplitude modulators and phase modulators [1, 4] are the two types of optical masks used in lensless imaging systems. Phase modulators can be further sub-categorized into phase gratings [30, 31], diffusers [2], and phase masks [7]. One key characteristic of a mask-modulated lensless system is the pattern the mask produces on the sensor for a point light source in the scene. We call this pattern the Point-Spread Function(PSF), and its properties determine the imaging model of the system. As shown in Fig. 2, we design a simple mask-modulated lensless system for data acquisition in this experiment. The lensless imaging system designed in our experiments chose a diffuser as the mask." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The main advantages of lensless imaging are as follows: lensless is small in size and can be assembled in a variety of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1472" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.465, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Figure 2. The workflow of the lensless imaging system and the final measurement obtained is the result obtained from the encoding of 3D scene information by the lensless imaging system. The optically encoded mask transforms local information in the 3D scene into overlapping global information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.385 + ], + "angle": 0, + "content": "miniaturized equipment; lensless is light and easy to carry; lensless cameras are still cheap to make; lensless cameras have a wide view field, larger than traditional wide-angle cameras; It can also protect the user's privacy to a great extent." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.394, + 0.373, + 0.411 + ], + "angle": 0, + "content": "2.2. Human Pose and Shape Recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.522 + ], + "angle": 0, + "content": "There are two primary methods for human pose and shape estimation: optimization-based approaches and regression-based approaches. Numerous methodologies have been developed to estimate the three-dimensional human pose and shape by employing iterative optimization techniques, e.g., SMPLify [5] and variations of the SMPLify [20, 26, 29, 33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Many existing Regression-based methods follow the architecture of HMR [14], which uses a pre-trained backbone to extract image features followed by regression to obtain SMPL [5] parameters. Many improvements to the original method have been proposed since its introduction. In particular, many papers have proposed alternative methods for pseudo-ground truth generation, including using temporal information [3], multiple views [21], or iterative optimization [13, 17, 27]. SPIN [17] proposed an in-the-loop optimization that incorporated SMPLify [5] in the HMR training, which combines the optimization-based approach and the regression-based approach for human pose and shape estimation. PyMAF [35, 36] proposes a pyramidal mesh alignment feedback loop for regression of SMPL parameters. Our approach references the design of PyMAF in the part of the SMPL parameters regression. PARE [16] proposes a body-part-guided attention mechanism for better occlusion handling. HKMR [10] performs a prediction guided by the hierarchical structure of SMPL. HMR2.0 [11] employs a large training dataset and proposes a fully transformer-based approach for 3D human pose and shape estimation from a single image. Many related approaches make non-parametric predictions, i.e., instead of estimating the parameters of the SMPL model, they explicitly regress the vertices of the mesh. GraphCMR [18] uses a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "graph neural network for the prediction, METRO [23] and FastMETRO [8] use a transformer, while Mesh Graphormer [24] adopts a hybrid between the two." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.244 + ], + "angle": 0, + "content": "In this paper, we propose the first work to perform end-to-end human pose and shape estimation from lensless measurements. We designed a multi-scale lensless feature decoder to decode the lensless measurements based on a mask-encoder to obtain more efficient features. We also propose a double auxiliary supervision mechanism to improve the estimation accuracy of human limbs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.258, + 0.593, + 0.273 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.283, + 0.611, + 0.298 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.503 + ], + "angle": 0, + "content": "Our work focuses on human pose and shape estimation from lensless measurements. In this section, we present the technical details of our approach. As depicted in Fig. 3, the core of our method involves the following three components: 1) a Multi-Scale Lensless Feature Decoder(MSFDecoder) (Sec. 3.2) that can effectively decode the information encoded by the lensless imaging system; 2) a human parametric model regressor (Sec. 3.3) that takes the multi-scale features produced by MSFDecoder as input and predicts the SMPL parameters; 3) a Double-Head Auxiliary Supervision mechanism(DHAS) (Sec. 3.4) that can assist the LPSNet to improve the estimation accuracy of human limbs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.515, + 0.825, + 0.53 + ], + "angle": 0, + "content": "3.2. Multi-Scale Lensless Feature Decoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.809 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, the optical information of the subject is diffused by the mask and projected onto the sensor. The process of obtaining measurements through the lensless imaging system can be treated as encoding optical information in the scene into lensless measurements. The goal of our lensless feature decoder \\( D_{M} \\) is to decode these lensless measurements to multi-scale features, which are then utilized for subsequent human body prediction. The design of the global perception layer \\( G_{P} \\) is inspired by HRNet [32]. The global perception layer inherits many of the advantages of HRNet [32] and can always maintain a high resolution, while the information interaction between different branches complements the information loss caused by the reduction in the number of channels. These advantages are important for extracting features from lensless measurements. Then, we design different convolutional layers for different scales of features that are used to refine the features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Formally, the decoder takes a lensless measurement \\(M\\) as input and decodes the information encoded by the lensless imaging system. The global perception layer ensures the integrity of global information for each pixel by maintaining the high-resolution convolution branch and the low-resolution convolution branch in parallel. It further en" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1473" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.095, + 0.87, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.388, + 0.895, + 0.432 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed framework. A measurement \\( M \\) is passed through a Multi-Scale Lensless Feature Decoder to get spatial characteristics at different scales. These feature maps are fed into the regressor for human pose and shape estimation. Also, these feature maps are fed into the Double-Head Auxiliary Supervision to improve the estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.473, + 0.562 + ], + "angle": 0, + "content": "hances the information extraction accuracy by repeating the multi-scale fusion across parallel convolutions. As shown in Fig. 3, the global perception layer takes \\(M\\) as input and generates a set of multi-scale spatial features \\(\\phi_{m}^{t = 0,1,2,3}\\) as input for the subsequent steps. The final output \\(\\phi_{M}^{t}\\) of the decoder is obtained by passing \\(\\phi_{m}^{t}\\) into different convolutional layers," + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.562, + 0.47, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\phi_ {M} ^ {t} = D _ {M} (M). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.473, + 0.705 + ], + "angle": 0, + "content": "With different scales of decoding, we can obtain feature maps at different granularities, through which we can perceive the location of people in the scene. The results obtained by the lensless decoder are utilized for subsequent predictions with the SMPL model, incorporating pose, shape, and camera parameters \\(\\Theta = \\{\\theta ,\\beta ,\\widetilde{\\pi}\\}\\), where \\(\\widetilde{\\pi}\\) is pseudo-camera parameters for the subsequent projection calculation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.407, + 0.729 + ], + "angle": 0, + "content": "3.3. Human Parametric Model Regression" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our human parametric model regression is mainly inspired by the improved Human Mesh Regression described in PyMAF [35]. Different from the pyramid features obtained by deconvolution in PyMAF [35], our global perception layer exploits HRNet [14] to maintain more high-resolution information. We use a set of sampling points \\( S_{t} \\) and project them onto the feature maps to extract point-wise features. The point-wise features corresponding to each point \\( s \\) in \\( S_{t} \\) are bilinearly sampled from \\( \\phi_M^{t=1,2,3} \\). (Note that \\( \\phi_M^{t=0} \\) is unnecessary here and is used as input for subsequent initialization.) Then the dimensions of these features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.501 + ], + "angle": 0, + "content": "will be reduced by MLP(multi-layer perceptron) and concatenated as a feature vector \\(\\hat{\\phi}_M^t\\) which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.508, + 0.892, + 0.549 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\phi} _ {M} ^ {t} = \\mathcal {F} \\left(\\phi_ {M} ^ {t}, S _ {t}\\right) \\tag {2} \\\\ = \\oplus \\left\\{f \\left(\\phi_ {M} ^ {t} (s)\\right), \\text {f o r} s \\text {i n} S _ {t} \\right\\}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.895, + 0.678 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(\\cdot)\\) denotes the overall point-wise feature extraction, \\(f(\\cdot)\\) is the MLP layer, and \\(\\oplus\\) denotes the concatenation. We use an iterative mechanism to complete the interaction of different scale features. Sequentially, a parameter regressor \\(\\mathcal{R}_t\\) takes features \\(\\hat{\\phi}_M^t\\) and the current parameter estimation \\(\\Theta_t\\) as inputs and outputs the parameter residual. Parameters are updated as \\(\\Theta_{t + 1}\\) by adding the residual to \\(\\Theta_t\\) which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.685, + 0.892, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {t + 1} = \\Theta_ {t} + \\mathcal {R} _ {t} \\left(\\Theta_ {t}, \\hat {\\phi} _ {M} ^ {t}\\right), \\text {f o r} t > 0. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.75 + ], + "angle": 0, + "content": "The initial parameter \\(\\Theta_0\\) is initialized by feeding \\(\\phi_M^{t=0}\\) into the regressor \\(\\mathcal{R}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "When the predicted parameters \\(\\Theta\\) (the subscript \\(t\\) is omitted for simplicity) are obtained through each scale feature, a mesh with vertices of \\(V = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{N\\times 3}\\) can be generated accordingly, where \\(N = 6890\\) is the number of vertices in the SMPL model. We downsample \\(V\\) to get the sampling point \\(S_{t + 1}\\) for the next iteration. These mesh vertices can be mapped by pre-trained linear regressor to obtain sparse 3D keypoints \\(J\\in \\mathbb{R}^{N_j\\times 3}\\). With the estimated pseudo-camera parameters, we can obtain 2D keypoints \\(K\\in \\mathbb{R}^{N_j\\times 2}\\) by projecting \\(J\\) on the measurement coordinate system. We" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "calculate the loss between the 2D keypoints and the ground truth, in this way reducing the difference between the 2D projection and the human pose in the real scene. Concurrently, supplementary 3D supervisions regarding 3D keypoints and model parameters are integrated when authentic 3D labels are available. The total loss function for the parameter regressor is thereby formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.219, + 0.47, + 0.239 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {2 d} \\| K - \\hat {K} \\| _ {2} + \\lambda_ {3 d} \\| J - \\hat {J} \\| _ {2} + \\lambda_ {\\text {p a r a}} \\| \\Theta - \\hat {\\Theta} \\| _ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.245, + 0.47, + 0.293 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) is the squared L2 norm, \\(\\hat{K},\\hat{J}\\), and \\(\\hat{\\Theta}\\) denote the ground truth 2D keypoints, 3D keypoints, and model parameters, respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.3, + 0.388, + 0.317 + ], + "angle": 0, + "content": "3.4. Double-Head Auxiliary Supervision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.469, + 0.429 + ], + "angle": 0, + "content": "The spatial feature map of the human body is relatively rough and contains a lot of noise. Therefore, there are still some deviations in the perception of human limbs and pose. To improve our estimation accuracy of human limbs, we introduce a Double-Head Auxiliary Supervision (DHAS) mechanism to obtain finer spatial features during the training stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.595 + ], + "angle": 0, + "content": "Specifically, we first convert all the spatial features at different scales to a single feature map \\(\\phi^T\\) by upsampling and then concatenating them together. The feature map \\(\\phi^T\\) is used for the different auxiliary supervision heads. On the one hand, we generate a heat map representation by a classification layer to explicitly indicate the position of the 2d keypoints. On the other hand, we also estimate a dense map through the IUV Predict layer to get a dense correspondence between the SMPL model and the feature map. The loss function for the Double-Head Auxiliary Supervision consists of two parts, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.603, + 0.469, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d a s} = \\mathcal {L} _ {s c} + \\mathcal {L} _ {d e n}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Keypoints Supervision. We utilize the SimCC-based algorithm [22] for predicting pose keypoints. This approach treats keypoint localization as a classification task in horizontal and vertical coordinates. During training, instead of estimating the actual coordinates, we employ two vectors, \\( x \\) and \\( y \\), and convert the ground truth 2D keypoints into such vectors to compute the loss. The loss function is then formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.756, + 0.469, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s c} = \\lambda_ {x y} (\\operatorname {K L - L o s s} (x, \\hat {x}) + \\operatorname {K L - L o s s} (y, \\hat {y})), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where KL-Loss is the Kullback-Leibler divergence loss, \\(\\hat{x}\\) and \\(\\hat{y}\\) denote the processed ground truth 2D keypoints, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "IUV Supervision. We adopt the IUV mapping defined in DensePose [12] as the dense correspondence representation. This mapping establishes deterministic correspondences between foreground pixels in 2D images and vertices on 3D surfaces. The vertices on the template mesh can" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.094, + 0.682, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.094, + 0.853, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.227, + 0.858, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.373, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Figure 4. On the top left is a figure of the dual camera system we created. This system consists of an RGB camera and a lensless imaging system using a diffuser as a mask. The top right shows the process of collecting our dataset. The image at the bottom is a record of the process of collecting our dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.59 + ], + "angle": 0, + "content": "be mapped back to pixels on the foreground using a predefined bijection mapping between 3D surface space and 2D UV space. The dense correspondence representation comprises the index of body parts \\( P \\) and UV values of mesh vertices. During training, we apply classification loss to the index of body parts \\( P \\) and regression loss to the UV channels of the dense correspondence mapping. The loss function is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.599, + 0.891, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {d e n} = \\lambda_ {p i} \\text {C r o s s E n t r o p y} (P, \\hat {P}) \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (U, \\hat {U}) \\tag {7} \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (V, \\hat {V}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.672, + 0.89, + 0.704 + ], + "angle": 0, + "content": "where \\(\\hat{P},\\hat{U}\\), and \\(\\hat{V}\\) denote the ground truth of \\(P\\) and \\(UV\\) values, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.717, + 0.633, + 0.734 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.756, + 0.759 + ], + "angle": 0, + "content": "4.1. Imaging System and Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Lensless Imaging System. Due to the scarcity of human pose and shape datasets based on lensless imaging systems, we developed a basic lensless system to gather experimental data. The system, illustrated in Fig. 4, comprises two primary components: a lensless imaging system for capturing lensless measurements and an RGB camera for capturing real images corresponding to the lensless measurements. To ensure identical light paths, we employed a beam splitter between the lensless sensor and the RGB camera. This al" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1475" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "lowed for consistent imaging conditions and facilitated accurate correlation between the lensless measurements and the corresponding real images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.138, + 0.468, + 0.228 + ], + "angle": 0, + "content": "For the lensless imaging part of the system, we adopted the Mask-Modulated Lensless System, inspired by design options such as diffusercam [2] and phlatcam [7]. Specifically, we selected a diffuser as our mask and incorporated PhlatCam's concept of using larger sensors to enhance global information perception." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.468, + 0.319 + ], + "angle": 0, + "content": "Datasets. The inputs to LPSNet are lensless measurements. However, datasets for classical human pose estimation are not directly available at present. Leveraging the lensless imaging system we have constructed, and the mathematical model of the system's imaging, the sources of datasets for our experiments can be categorized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.322, + 0.468, + 0.546 + ], + "angle": 0, + "content": "- Real Dataset. Using a lensless imaging system to capture images displayed on a screen as measurements is currently the primary method of acquiring datasets in the lensless field. We used this approach to collect datasets from various sources, including Human3.6M, MPII, COCO, 3DPW, and MIP-INF-3DHP datasets, which we refer to as LenslessHuman3.6M, LenslessMPII, LenslessMIP-INF-3DHP, LenslessCOCO, and Lensless3DPW, respectively. We also reclassified the training dataset and validation dataset on LenslessHuman3.6M, referred to as train-LenslessHuman3.6M and eval-LenslessHuman3.6M, respectively. Moreover, we capture the real scenes of different individuals, backgrounds, and light intensities using our lensless imaging system." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.548, + 0.468, + 0.683 + ], + "angle": 0, + "content": "- Simulated Dataset. The mathematical imaging model enables the conversion (simulation) of images captured by RGB cameras into measurements captured by lensless imaging systems. This technique is commonly employed in lensless imaging fields for testing systems and imaging methods. Leveraging this, we converted numerous human pose datasets into simulated lensless measurements. These simulated datasets extended our training set and provided a quick validation of our method." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.322, + 0.468, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.294, + 0.71 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The dimensions of the raw measurements are \\(1280 \\times 1024 \\times 3\\). Prior to commencing the experiments, we first crop and resize the lensless measurements. These measurements are pre-processed to \\(224 \\times 224 \\times 3\\) before being fed into the network. Following this pre-processing step, the lensless measurements inputted into the global perceptual layer are \\(224 \\times 224\\), generating spatial features at resolutions of \\((7 \\times 7, 14 \\times 14, 28 \\times 28\\), and \\(56 \\times 56\\)). During the generation of mesh-aligned features, the SMPL mesh is downsampled using a pre-computed downsampling matrix provided in [18], resulting in a reduction of vertex count from 6890 to 431. The regressors \\(\\mathcal{R}_t\\) follow a congruent architec" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "tural framework to the regressor present in HMR [14], albeit with slight variations in their input dimensions. Our network is trained using the Adam optimizer [15] with a learning rate of \\( 5 \\times 10^{-5} \\) and a batch size of 80 on 4 NVIDIA RTX2080 Ti GPUs. No learning rate decay is applied during training." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.195, + 0.603, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.262, + 0.59, + 0.269 + ], + "angle": 0, + "content": "Measurement" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.195, + 0.726, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.263, + 0.721, + 0.272 + ], + "angle": 0, + "content": "Recovered Images" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.218, + 0.774, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.242, + 0.772, + 0.25 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.196, + 0.862, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.264, + 0.872, + 0.273 + ], + "angle": 0, + "content": "Human Mesh Recovery" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.288, + 0.599, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.358, + 0.598, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.29, + 0.689, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.358, + 0.689, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.29, + 0.781, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.358, + 0.781, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.29, + 0.869, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.358, + 0.869, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.527 + ], + "angle": 0, + "content": "Figure 5. At the top is the frame of our baseline. Lensless measurements first recover the image by reconstruction methods and then estimate human pose and shape by PyMAF. The bottom image compares the recovered image with the original image, and we can see that the quality of the recovered image has dropped dramatically." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.556, + 0.847, + 0.571 + ], + "angle": 0, + "content": "4.3. 3D Human Pose and Shape Comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.73 + ], + "angle": 0, + "content": "Baseline Approach. We are the first work to perform human pose and shape estimation through lensless measurements. Given the absence of other valid methods for comparison, we designed a baseline approach that approaches this as a two-stage task. The baseline architecture is depicted in Fig. 5. In the first stage, we utilized the lensless image reconstruction method Rego et. al. [28], chosen for its wide applicability in recovering images acquired by lensless imaging systems. In the second stage, PyMAF is used to estimate human pose and shape." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.743, + 0.88, + 0.816 + ], + "angle": 0, + "content": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
baseline (PyMAF)257.07121.73277.18
baseline (PyMAF†)126.2881.37151.60
LPSNet119.2081.52134.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.827, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Table 1. Comparison with baseline. The baseline is a combination of the two-stage approach using existing methods. LPSNet outperforms the baseline on the LenslessHuman3.6M datasets." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Comparison with Baseline. We present the results of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1476" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.87, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.399, + 0.893, + 0.442 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of LPSNet on challenging LenslessHuman3.6M. The results on the left are from our LPSNet estimation, those in the center are from our baseline (PyMAF) output, and those on the right are from the fine-tuned baseline \\((\\mathrm{PyMAF}^{\\dagger})\\) output. LPSNet can be seen to be better than the baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.468, + 0.47, + 0.619 + ], + "angle": 0, + "content": "the quantitative comparative evaluation on the LenslessHuman3.6M dataset in Tab. 1. Notably, images recovered from lensless measurements exhibit lower quality. Directly comparing the baseline (PyMAF) with our approach is challenging and unfair. Therefore, we fine-tuned PyMAF on the train-LenslessHuman3.6M dataset, denoted as \\(\\mathrm{PyMAF}^{\\dagger}\\). Referring to Tab. 1, we observe significant improvements in the indicators of baseline \\((\\mathrm{PyMAF}^{\\dagger})\\). Specifically, compared with the baseline \\((\\mathrm{PyMAF}^{\\dagger})\\), the MPJPE of LPSNet on the LenslessHuman3.6M dataset is reduced by \\(7.08~\\mathrm{mm}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.471, + 0.727 + ], + "angle": 0, + "content": "As shown in Tab. 1, our LPSNet achieved competitive results compared to the baseline approach \\((\\mathrm{PyMAF}^{\\dagger})\\). LPSNet demonstrates more significant improvements in the MPJPE and PVE metrics. However, we would argue that the PA-MPJPE metric may not fully reveal the performance of the mesh-image alignment, as it is calculated as MPJPE after rigid alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Baseline (PyMAF) has difficulty performing accurate pose estimation on low-quality recovered images. However, the outputs of our LPSNet exhibit significant improvements compared to the two baseline methods, especially for limb estimation. Qualitative comparison results are illustrated in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.233, + 0.848 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies on LenslessHuman3.6M under various settings to validate the effectiveness of the key components proposed in our method." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.465, + 0.895, + 0.559 + ], + "angle": 0, + "content": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
w/o MSFDcoder and DHAS142.1392.20161.07
w/o DHAS139.3492.56158.97
w/o DHAS (2D keypoint)123.3083.41138.67
w/o DHAS (IUV)129.7086.83146.7
LPSNet119.2081.52134.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.569, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Table 2. Ablation study on LPSNet. DHAS stands for Double-Head Auxiliary Supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.703 + ], + "angle": 0, + "content": "All ablation variants were trained and tested on LenslessHuman3.6M, which is derived from the Human3.6M dataset. The Human3.6M dataset includes ground-truth 3D labels and serves as a widely used benchmark in 3D human pose and shape estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Multi-Scale Lensless Feature Decoder. In LPSNet, the decoder primarily decodes the global information of lensless measurements encoded by the diffuser, which plays a crucial role in our end-to-end system. Our MSFDecoder is essential for efficient feature extraction from lensless measurements. We have adapted another variant of the decoder to verify this. Specifically, we simplified the global perception layer of the decoder to a series of convolutional layers and deconvolutional layers. We use this simpler decoder in the experiment(w/o MSFcoder and DHAS) instead of MSFDecoder. As shown in Tab. 2, comparing the result of experiments(w/o MSFcoder and DHAS) and experiment(w/o DHAS), we can see that MSFDecoder improves" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1477" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.844, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.893, + 0.34 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of LPSNet on challenging LenslessHuman3.6M. We can see that the results of LPSNet are better, and the results are improved with the addition of the double-head assisted supervision mechanism." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.457 + ], + "angle": 0, + "content": "the accuracy of the human pose and shape estimation significantly. Through the qualitative experiment in Fig. 7, we can also see that the human pose and shape estimation has a better alignment after using the MSFDecoder. Note that neither experiment(w/o MSFDcoder and DHAS) nor the experiment(w/o DHAS) used auxiliary supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.471, + 0.607 + ], + "angle": 0, + "content": "Double-Head Auxiliary Supervision. Double-Head Auxiliary Supervision is mainly used to improve the accuracy of human limbs. For each auxiliary supervisory head, we conducted ablation experiments respectively. Referring to the result in Tab. 2, LPSNet has significant improvements in the MPJPE and the PVE metrics. Fig. 7 shows additional qualitative comparison results. We can see that the alignment of the human pose and shape is better with the addition of the Double-Head Auxiliary Supervision mechanism." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.618, + 0.289, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.722, + 0.267, + 0.734 + ], + "angle": 0, + "content": "(a) Result on more datasets" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.622, + 0.466, + 0.717 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.722, + 0.454, + 0.734 + ], + "angle": 0, + "content": "(b) Result on real scenes" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.746, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Figure 8. Results on more datasets and real scenes. From left to right for each set of figures: lensless measurement, alignment of results with RGB images, and 3D results shown in different views." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.427, + 0.817 + ], + "angle": 0, + "content": "4.5. Results on more Datasets and Real Scene" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Fig. 8 (a) shows the qualitative results on various datasets after using the mixed datasets for training. Specifically, we collected the MPII, COCO, 3DPW, MIP-INF-3DHP, and LSP datasets through our lensless imaging system and mixed them for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.456 + ], + "angle": 0, + "content": "In addition, we also capture the real scenes of two different people, in different backgrounds and different light intensities using our lensless imaging system. Fig. 8 (b) illustrates the experimental results of our method on these real scenes, indicating its suitability for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.747, + 0.489 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Conclusion. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to the best of our knowledge. Specifically, we design a multi-scale lensless feature decoder, which can effectively decode the information produced by the lensless imaging system. We also propose a double auxiliary supervision mechanism to improve the accuracy of human limb end estimation. Experimental results show that our method can achieve end-to-end human pose and shape estimation through lensless measurements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Limitations and Future Work. Our work is an initial stride step in this direction, however, the results are still not as robust as those achieved by traditional methods for estimating the human body from RGB images. The current scarcity of lensless datasets hampers our ability to pre-train the MSFDecoder using a large dataset like ImageNet [9], thereby limiting the generalizability of our approach. We will continue this work in the future, focusing on generating more extensive lensless datasets and progressing towards practical applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.893, + 0.888 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Key R&D Program of China (2023YFC3082100), the National Natural Science Foundation of China (62122058 and 62171317), and the Science Fund for Distinguished Young Scholars of Tianjin (No. 22JCJQJC00040)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1478" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jesse K Adams, Vivek Boominathan, Benjamin W Avants, Daniel G Vercosa, Fan Ye, Richard G Baraniuk, Jacob T Robinson, and Ashok Veeraraghavan. Single-Frame 3D Fluorescence Microscopy with Ultraminiature Lensless FlatScope. Sci Adv., 3(12):e1701548, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] Nick Antipa, Grace Kuo, Reinhard Heckel, Ben Mildenhall, Emrah Bostan, Ren Ng, and Laura Waller. DiffuserCam: Lensless Single-exposure 3D Imaging. Optica, 5(1):1-9, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting Temporal Context for 3D Human Pose Estimation in the ild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3395-3404, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.369 + ], + "angle": 0, + "content": "[4] M Salman Asif, Ali Ayremlou, Aswin Sankaranarayanan, Ashok Veeraraghavan, and Richard G Baraniuk. Flatcam: Thin, Lensless Cameras using Coded Aperture and Computation. IEEE Trans Comput Imaging., 3(3):384-397, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In *Eur. Conf. Comput. Vis. Springer International Publishing*, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[6] Vivek Boominathan, Jesse K Adams, M Salman Asif, Benjamin W Avants, Jacob T Robinson, Richard G Baraniuk, Aswin C Sankaranarayanan, and Ashok Veeraraghavan. Lensless Imaging: A Computational Renaissance. IEEE Signal Process Mag., 33(5):23-35, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.47, + 0.57 + ], + "angle": 0, + "content": "[7] Vivek Boominathan, Jesse K Adams, Jacob T Robinson, and Ashok Veeraraghavan. Phlatcam: Designed Phase-Mask Based Thin Lensless Camera. IEEE Trans. Pattern Anal. Mach. Intell., 42(7):1618-1629, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In Eur. Conf. Comput. Vis., pages 342–359. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[9] J. Deng, W. Dong, R. Socher, L. Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPR Workshops). 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[10] Georgios Georgakis, Ren Li, Srikrishna Karanam, Terrence Chen, Jana Košecka, and Ziyan Wu. Hierarchical Kinematic Human Mesh Recovery. In Eur. Conf. Comput. Vis., pages 768-784. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and Tracking Humans with Transformers. In Int. Conf. Comput. Vis., pages 14783-14794, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense Human Pose Estimation in the Wild. In Int. Conf. Comput. Vis., pages 7297-7306, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. In Int. Conf. 3D. Vis., pages 42-52. IEEE, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[14] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end Recovery of Human Shape and Pose. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[16] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. PARE: Part Attention Regressor for 3D Human Body Estimation. In Int. Conf. Comput. Vis., pages 11127-11137, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[17] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to Reconstruct 3D Human Pose and Shape Via Model-fitting in the Loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[18] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional Mesh Regression for Single-Image Human Shape Reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4501–4510, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[19] Grace Kuo, Nick Antipa, Ren Ng, and Laura Waller. DiffuserCam: Diffuser-Based Lensless Cameras. In Computational Optical Sensing and Imaging, pages CTu3B-2. Optica Publishing Group, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[20] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Vincent Leroy, Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, and Grégory Rogez. SMPLy Benchmarking 3D Human Pose Estimation in the wild. In Int. Conf. 3D. Vis., pages 301-310. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation. In Eur. Conf. Comput. Vis., pages 89-106. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[23] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1954-1963, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh Graphormer. In Int. Conf. Comput. Vis., pages 12939-12948, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] Aydogan Ozcan and Euan McLeod. Lensless Imaging and Sensing. Annual review of biomedical engineering, 18:77-102, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human Mesh Recovery from Multiple Shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Joshua D Rego, Karthik Kulkarni, and Suren Jayasuriya. Robust Lensless Image Reconstruction Via PSF Estimation. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1479" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 403-412, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[29] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D Human Motion Model for Robust Pose Estimation. In Int. Conf. Comput. Vis., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[30] David G Stork and Patrick R Gill. Lensless Ultra-miniature CMOS Computational Imagers and Sensors. Proc. Sensorcomm, pages 186-190, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[31] David G Stork and Patrick R Gill. Optical, mathematical, and computational foundations of lensless ultra-miniature diffractive imagers and sensors. International Journal on Advances in Systems and Measurements, 7(3):4, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[32] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep High-resolution Representation Learning for Human Pose Estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5693-5703, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[33] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. PoseNDF: Modeling Human Pose Manifolds with Neural Distance Fields. In Eur. Conf. Comput. Vis., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.445 + ], + "angle": 0, + "content": "[34] Hongwen Zhang, Jie Cao, Guo Lu, Wanli Ouyang, and Zhenan Sun. Learning 3D Human Shape and Pose from Dense Body Parts. IEEE Trans. Pattern Anal. Mach. Intell., 44(5):2610-2627, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[35] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Int. Conf. Comput. Vis., pages 11446-11456, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[36] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3D Human Dynamics from Video. In Int. Conf. Comput. Vis., pages 7114-7123, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1480" + } + ] +] \ No newline at end of file diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_origin.pdf b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9956933ebd08f4d86855538beb4d3fcd7c60f5fa --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/3f527010-7aaa-4b2e-ab0a-89e9d8eb4cef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dff18db0e3f66cfb26aa91c315f40998b62edc11098f96f7a8b3eb93bd8643e3 +size 9097420 diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/full.md b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..012463724fe5ff004c57f241f4946999ab0c1360 --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/full.md @@ -0,0 +1,298 @@ +# LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging + +Haoyang Ge $^{1,\dagger}$ , Qiao Feng $^{1,\dagger}$ , Hailong Jia $^{1}$ , Xiongzheng Li $^{1}$ , Xiangjun Yin $^{1}$ , You Zhou $^{2}$ , Jingyu Yang $^{1}$ , Kun Li $^{1,*}$ + +$^{1}$ Tianjin University, China $^{2}$ Nanjing University, China + +{ghy0623,fengqiao,jhl,lxz,yinxiangjun,yjy,lik}@tju.edu.cn zhouyou@nju.edu.cn + +![](images/271240214a8cb7c489998d4b4fadf00997a8b09d5b26bde665705f970bd6be0b.jpg) +Figure 1. LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging. We contribute a framework for estimating human poses and shapes from individual lensless measurements. The first row shows the input measurements acquired by our lensless imaging system, the second row shows the estimated human poses and shapes from lensless measurements, and the bottom row shows the 3D results shown in different views. + +# Abstract + +Human pose and shape (HPS) estimation with lensless imaging is not only beneficial to privacy protection but also can be used in covert surveillance scenarios due to the small size and simple structure of this device. However, this task presents significant challenges due to the inherent ambiguity of the captured measurements and lacks effective methods for directly estimating human pose and shape from lensless data. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to our knowledge. We specifically design a multi-scale lensless feature decoder to decode the lensless measurements through the optically encoded mask + +for efficient feature extraction. We also propose a doublehead auxiliary supervision mechanism to improve the estimation accuracy of human limb ends. Besides, we establish a lensless imaging system and verify the effectiveness of our method on various datasets acquired by our lensless imaging system. The code and dataset are available at https://cic.tju.edu.cn/faculty/likun/projects/LPSNet. + +# 1. Introduction + +In recent years, lensless imaging technologies [6, 7, 19, 25] have advanced significantly due to their many advantages, such as privacy protection, smaller size, simple structure, and lower cost. 3D human pose and shape (HPS) estimation [11, 14, 34, 35] requires miniaturized and + +lightweight imaging system as application scenarios become more diverse. All these advantages, especially privacy, make the lensless imaging system very suitable as an imaging device for human pose and shape estimation. In this work, we propose LPSNet, which aims to estimate 3D human pose and shape from lensless measurements instead of RGB images, achieving cheaper and privacy-protecting 3D human pose and shape estimation. + +A thin, lightweight, and potentially cost-effective optical encoder is used in a lensless imaging system instead of traditional cameras with lenses, while others are expensive, rigid, and occupy more space. At this stage, the application of a lensless imaging system is extensive, it is mainly used in the field of microscopic imaging, RGB image reconstruction, and so on. More valuable information can be obtained from lensless measurements due to the special optical encoding method of lensless imaging systems. Directly estimating human pose and shape from lensless measurements is not currently possible. The first step is to reconstruct an RGB image from a lensless measurement and then estimate the human pose and shape from the RGB images. However, experimental findings indicate that the reconstructed RGB images are of suboptimal quality, resulting in incomplete local features and significant deviations in the position of the human body. Combining these factors leads to inaccurate human pose estimation when using lensless measurements to reconstruct RGB images. This approach has computational burden and computational resources, making it very unfavorable for deployment at the endpoint. + +In this paper, We aim to advance human pose and shape estimation using a lensless imaging system, which needs to overcome two main challenges. First, how to extract features from lensless measurements for human pose and shape estimation. Secondly, during early experiments, when using features extracted from lensless measurements to estimate human pose and shape, we found poor estimation accuracy of human limbs. + +To address these challenges, we introduce LPSNet, the first end-to-end human pose and shape estimation framework with lensless imaging. To extract features from optically encoded lensless measurements, we propose a multiscale lensless feature decoder (MSFDecoder). Specifically, we introduce a global perception layer to enhance the global decoding capability of MSFDecoder. The global information that has been optically encoded to the global can be efficiently decoded to obtain a feature map that can be used in subsequent processes. To improve pose and shape estimation, we propose a Double-Head Auxiliary Supervision(DHAS) mechanism to be implemented during training. Auxiliary supervision can improve the estimation accuracy of human limbs and correct results with large deviations. + +Our main contributions can be summarized as follows: 1) We propose LPSNet, an end-to-end pose and shape esti + +mation network for lensless imaging systems. This is the first work to estimate human poses and shapes directly from lensless measurements. + +2) We propose MSFDecoder, a Multi-Scale Lensless Feature Decoder that decodes and extracts features from lensless measurements, which can be receptive to global features in lensless measurements. +3) We propose a Double-Head Auxiliary Supervision mechanism for both pose and shape estimation, which can improve the estimation accuracy of human limbs. + +# 2. Related Work + +# 2.1. Lensless Imaging System + +A conventional photographic camera typically comprises a focusing lens, which may consist of one or more optical elements and an image sensor positioned at or near the focal length of the lens. The lens in such a camera directs the projected light from the observed scene onto the sensor, aiming to accurately map specific scene points to individual pixels on the sensor. Conversely, in a lensless imaging system, the absence of a lens defines its configuration. Instead, an optical modulator, such as a coded amplitude mask or a diffuser, is positioned between the scene and the image sensor, often close to the sensor itself. As a result, the recorded data deviates significantly from the expected RGB image during imaging. In this process, the local information of the object transforms overlapping global information through the optically encoded mask. + +Within the mask-modulated lensless systems, a fixed optical mask is introduced to create a versatile lensless system that can work for a wide range of object distances and lighting scenarios, whether passive or uncontrolled. The mask modulates the incoming light and generates a measurement that can be decoded through computational methodologies. Mask-modulated lensless imaging systems were used to perform 2D imaging [4, 7, 19], refocusing [7], 3D imaging [2, 7], and microscopic imaging [1, 2, 7]. + +Generally speaking, amplitude modulators and phase modulators [1, 4] are the two types of optical masks used in lensless imaging systems. Phase modulators can be further sub-categorized into phase gratings [30, 31], diffusers [2], and phase masks [7]. One key characteristic of a mask-modulated lensless system is the pattern the mask produces on the sensor for a point light source in the scene. We call this pattern the Point-Spread Function(PSF), and its properties determine the imaging model of the system. As shown in Fig. 2, we design a simple mask-modulated lensless system for data acquisition in this experiment. The lensless imaging system designed in our experiments chose a diffuser as the mask. + +The main advantages of lensless imaging are as follows: lensless is small in size and can be assembled in a variety of + +![](images/f33d35d1c12c6238da0bf12832073986d998c84e2c189df320c8054ad2cf087d.jpg) +Figure 2. The workflow of the lensless imaging system and the final measurement obtained is the result obtained from the encoding of 3D scene information by the lensless imaging system. The optically encoded mask transforms local information in the 3D scene into overlapping global information. + +miniaturized equipment; lensless is light and easy to carry; lensless cameras are still cheap to make; lensless cameras have a wide view field, larger than traditional wide-angle cameras; It can also protect the user's privacy to a great extent. + +# 2.2. Human Pose and Shape Recovery + +There are two primary methods for human pose and shape estimation: optimization-based approaches and regression-based approaches. Numerous methodologies have been developed to estimate the three-dimensional human pose and shape by employing iterative optimization techniques, e.g., SMPLify [5] and variations of the SMPLify [20, 26, 29, 33]. + +Many existing Regression-based methods follow the architecture of HMR [14], which uses a pre-trained backbone to extract image features followed by regression to obtain SMPL [5] parameters. Many improvements to the original method have been proposed since its introduction. In particular, many papers have proposed alternative methods for pseudo-ground truth generation, including using temporal information [3], multiple views [21], or iterative optimization [13, 17, 27]. SPIN [17] proposed an in-the-loop optimization that incorporated SMPLify [5] in the HMR training, which combines the optimization-based approach and the regression-based approach for human pose and shape estimation. PyMAF [35, 36] proposes a pyramidal mesh alignment feedback loop for regression of SMPL parameters. Our approach references the design of PyMAF in the part of the SMPL parameters regression. PARE [16] proposes a body-part-guided attention mechanism for better occlusion handling. HKMR [10] performs a prediction guided by the hierarchical structure of SMPL. HMR2.0 [11] employs a large training dataset and proposes a fully transformer-based approach for 3D human pose and shape estimation from a single image. Many related approaches make non-parametric predictions, i.e., instead of estimating the parameters of the SMPL model, they explicitly regress the vertices of the mesh. GraphCMR [18] uses a + +graph neural network for the prediction, METRO [23] and FastMETRO [8] use a transformer, while Mesh Graphormer [24] adopts a hybrid between the two. + +In this paper, we propose the first work to perform end-to-end human pose and shape estimation from lensless measurements. We designed a multi-scale lensless feature decoder to decode the lensless measurements based on a mask-encoder to obtain more efficient features. We also propose a double auxiliary supervision mechanism to improve the estimation accuracy of human limbs. + +# 3. Method + +# 3.1. Overview + +Our work focuses on human pose and shape estimation from lensless measurements. In this section, we present the technical details of our approach. As depicted in Fig. 3, the core of our method involves the following three components: 1) a Multi-Scale Lensless Feature Decoder(MSFDecoder) (Sec. 3.2) that can effectively decode the information encoded by the lensless imaging system; 2) a human parametric model regressor (Sec. 3.3) that takes the multi-scale features produced by MSFDecoder as input and predicts the SMPL parameters; 3) a Double-Head Auxiliary Supervision mechanism(DHAS) (Sec. 3.4) that can assist the LPSNet to improve the estimation accuracy of human limbs. + +# 3.2. Multi-Scale Lensless Feature Decoder + +As illustrated in Fig. 2, the optical information of the subject is diffused by the mask and projected onto the sensor. The process of obtaining measurements through the lensless imaging system can be treated as encoding optical information in the scene into lensless measurements. The goal of our lensless feature decoder $D_{M}$ is to decode these lensless measurements to multi-scale features, which are then utilized for subsequent human body prediction. The design of the global perception layer $G_{P}$ is inspired by HRNet [32]. The global perception layer inherits many of the advantages of HRNet [32] and can always maintain a high resolution, while the information interaction between different branches complements the information loss caused by the reduction in the number of channels. These advantages are important for extracting features from lensless measurements. Then, we design different convolutional layers for different scales of features that are used to refine the features. + +Formally, the decoder takes a lensless measurement $M$ as input and decodes the information encoded by the lensless imaging system. The global perception layer ensures the integrity of global information for each pixel by maintaining the high-resolution convolution branch and the low-resolution convolution branch in parallel. It further en + +![](images/748807a183ac6e648e492574ea5e5f7ac01516841e483c4af533ddd54b82fd02.jpg) +Figure 3. Overview of the proposed framework. A measurement $M$ is passed through a Multi-Scale Lensless Feature Decoder to get spatial characteristics at different scales. These feature maps are fed into the regressor for human pose and shape estimation. Also, these feature maps are fed into the Double-Head Auxiliary Supervision to improve the estimation accuracy. + +hances the information extraction accuracy by repeating the multi-scale fusion across parallel convolutions. As shown in Fig. 3, the global perception layer takes $M$ as input and generates a set of multi-scale spatial features $\phi_{m}^{t = 0,1,2,3}$ as input for the subsequent steps. The final output $\phi_{M}^{t}$ of the decoder is obtained by passing $\phi_{m}^{t}$ into different convolutional layers, + +$$ +\phi_ {M} ^ {t} = D _ {M} (M). \tag {1} +$$ + +With different scales of decoding, we can obtain feature maps at different granularities, through which we can perceive the location of people in the scene. The results obtained by the lensless decoder are utilized for subsequent predictions with the SMPL model, incorporating pose, shape, and camera parameters $\Theta = \{\theta ,\beta ,\widetilde{\pi}\}$ , where $\widetilde{\pi}$ is pseudo-camera parameters for the subsequent projection calculation. + +# 3.3. Human Parametric Model Regression + +Our human parametric model regression is mainly inspired by the improved Human Mesh Regression described in PyMAF [35]. Different from the pyramid features obtained by deconvolution in PyMAF [35], our global perception layer exploits HRNet [14] to maintain more high-resolution information. We use a set of sampling points $S_{t}$ and project them onto the feature maps to extract point-wise features. The point-wise features corresponding to each point $s$ in $S_{t}$ are bilinearly sampled from $\phi_M^{t=1,2,3}$ . (Note that $\phi_M^{t=0}$ is unnecessary here and is used as input for subsequent initialization.) Then the dimensions of these features + +will be reduced by MLP(multi-layer perceptron) and concatenated as a feature vector $\hat{\phi}_M^t$ which can be formulated as: + +$$ +\begin{array}{l} \hat {\phi} _ {M} ^ {t} = \mathcal {F} \left(\phi_ {M} ^ {t}, S _ {t}\right) \tag {2} \\ = \oplus \left\{f \left(\phi_ {M} ^ {t} (s)\right), \text {f o r} s \text {i n} S _ {t} \right\}, \\ \end{array} +$$ + +where $\mathcal{F}(\cdot)$ denotes the overall point-wise feature extraction, $f(\cdot)$ is the MLP layer, and $\oplus$ denotes the concatenation. We use an iterative mechanism to complete the interaction of different scale features. Sequentially, a parameter regressor $\mathcal{R}_t$ takes features $\hat{\phi}_M^t$ and the current parameter estimation $\Theta_t$ as inputs and outputs the parameter residual. Parameters are updated as $\Theta_{t + 1}$ by adding the residual to $\Theta_t$ which can be formulated as: + +$$ +\Theta_ {t + 1} = \Theta_ {t} + \mathcal {R} _ {t} \left(\Theta_ {t}, \hat {\phi} _ {M} ^ {t}\right), \text {f o r} t > 0. \tag {3} +$$ + +The initial parameter $\Theta_0$ is initialized by feeding $\phi_M^{t=0}$ into the regressor $\mathcal{R}_t$ . + +When the predicted parameters $\Theta$ (the subscript $t$ is omitted for simplicity) are obtained through each scale feature, a mesh with vertices of $V = \mathcal{M}(\pmb {\theta},\pmb {\beta})\in \mathbb{R}^{N\times 3}$ can be generated accordingly, where $N = 6890$ is the number of vertices in the SMPL model. We downsample $V$ to get the sampling point $S_{t + 1}$ for the next iteration. These mesh vertices can be mapped by pre-trained linear regressor to obtain sparse 3D keypoints $J\in \mathbb{R}^{N_j\times 3}$ . With the estimated pseudo-camera parameters, we can obtain 2D keypoints $K\in \mathbb{R}^{N_j\times 2}$ by projecting $J$ on the measurement coordinate system. We + +calculate the loss between the 2D keypoints and the ground truth, in this way reducing the difference between the 2D projection and the human pose in the real scene. Concurrently, supplementary 3D supervisions regarding 3D keypoints and model parameters are integrated when authentic 3D labels are available. The total loss function for the parameter regressor is thereby formulated as follows: + +$$ +\mathcal {L} _ {\text {r e g}} = \lambda_ {2 d} \| K - \hat {K} \| _ {2} + \lambda_ {3 d} \| J - \hat {J} \| _ {2} + \lambda_ {\text {p a r a}} \| \Theta - \hat {\Theta} \| _ {2}, \tag {4} +$$ + +where $\| \cdot \| _2$ is the squared L2 norm, $\hat{K},\hat{J}$ , and $\hat{\Theta}$ denote the ground truth 2D keypoints, 3D keypoints, and model parameters, respectively. + +# 3.4. Double-Head Auxiliary Supervision + +The spatial feature map of the human body is relatively rough and contains a lot of noise. Therefore, there are still some deviations in the perception of human limbs and pose. To improve our estimation accuracy of human limbs, we introduce a Double-Head Auxiliary Supervision (DHAS) mechanism to obtain finer spatial features during the training stage. + +Specifically, we first convert all the spatial features at different scales to a single feature map $\phi^T$ by upsampling and then concatenating them together. The feature map $\phi^T$ is used for the different auxiliary supervision heads. On the one hand, we generate a heat map representation by a classification layer to explicitly indicate the position of the 2d keypoints. On the other hand, we also estimate a dense map through the IUV Predict layer to get a dense correspondence between the SMPL model and the feature map. The loss function for the Double-Head Auxiliary Supervision consists of two parts, which can be written as: + +$$ +\mathcal {L} _ {d a s} = \mathcal {L} _ {s c} + \mathcal {L} _ {d e n}. \tag {5} +$$ + +Keypoints Supervision. We utilize the SimCC-based algorithm [22] for predicting pose keypoints. This approach treats keypoint localization as a classification task in horizontal and vertical coordinates. During training, instead of estimating the actual coordinates, we employ two vectors, $x$ and $y$ , and convert the ground truth 2D keypoints into such vectors to compute the loss. The loss function is then formulated as follows: + +$$ +\mathcal {L} _ {s c} = \lambda_ {x y} (\operatorname {K L - L o s s} (x, \hat {x}) + \operatorname {K L - L o s s} (y, \hat {y})), \tag {6} +$$ + +where KL-Loss is the Kullback-Leibler divergence loss, $\hat{x}$ and $\hat{y}$ denote the processed ground truth 2D keypoints, respectively. + +IUV Supervision. We adopt the IUV mapping defined in DensePose [12] as the dense correspondence representation. This mapping establishes deterministic correspondences between foreground pixels in 2D images and vertices on 3D surfaces. The vertices on the template mesh can + +![](images/9801a94d8c8e0d9e7d8fcfba90bb98213ffae6ac37d4f64c99ed86cd429ff017.jpg) + +![](images/6803f5ecdd031d34dedaab0371fc523d820aaea6c37e86c1db0459a15c6f9d48.jpg) + +![](images/0445d025768d4be19211f97d3cf9d98e13239ea6bdafafd1a96579cdfbbd7f01.jpg) +Figure 4. On the top left is a figure of the dual camera system we created. This system consists of an RGB camera and a lensless imaging system using a diffuser as a mask. The top right shows the process of collecting our dataset. The image at the bottom is a record of the process of collecting our dataset. + +be mapped back to pixels on the foreground using a predefined bijection mapping between 3D surface space and 2D UV space. The dense correspondence representation comprises the index of body parts $P$ and UV values of mesh vertices. During training, we apply classification loss to the index of body parts $P$ and regression loss to the UV channels of the dense correspondence mapping. The loss function is formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {d e n} = \lambda_ {p i} \text {C r o s s E n t r o p y} (P, \hat {P}) \\ + \lambda_ {u v} \operatorname {S m o o t h L 1} (U, \hat {U}) \tag {7} \\ + \lambda_ {u v} \operatorname {S m o o t h L 1} (V, \hat {V}) \\ \end{array} +$$ + +where $\hat{P},\hat{U}$ , and $\hat{V}$ denote the ground truth of $P$ and $UV$ values, respectively. + +# 4. Experiments + +# 4.1. Imaging System and Dataset + +Lensless Imaging System. Due to the scarcity of human pose and shape datasets based on lensless imaging systems, we developed a basic lensless system to gather experimental data. The system, illustrated in Fig. 4, comprises two primary components: a lensless imaging system for capturing lensless measurements and an RGB camera for capturing real images corresponding to the lensless measurements. To ensure identical light paths, we employed a beam splitter between the lensless sensor and the RGB camera. This al + +lowed for consistent imaging conditions and facilitated accurate correlation between the lensless measurements and the corresponding real images. + +For the lensless imaging part of the system, we adopted the Mask-Modulated Lensless System, inspired by design options such as diffusercam [2] and phlatcam [7]. Specifically, we selected a diffuser as our mask and incorporated PhlatCam's concept of using larger sensors to enhance global information perception. + +Datasets. The inputs to LPSNet are lensless measurements. However, datasets for classical human pose estimation are not directly available at present. Leveraging the lensless imaging system we have constructed, and the mathematical model of the system's imaging, the sources of datasets for our experiments can be categorized as follows: + +- Real Dataset. Using a lensless imaging system to capture images displayed on a screen as measurements is currently the primary method of acquiring datasets in the lensless field. We used this approach to collect datasets from various sources, including Human3.6M, MPII, COCO, 3DPW, and MIP-INF-3DHP datasets, which we refer to as LenslessHuman3.6M, LenslessMPII, LenslessMIP-INF-3DHP, LenslessCOCO, and Lensless3DPW, respectively. We also reclassified the training dataset and validation dataset on LenslessHuman3.6M, referred to as train-LenslessHuman3.6M and eval-LenslessHuman3.6M, respectively. Moreover, we capture the real scenes of different individuals, backgrounds, and light intensities using our lensless imaging system. +- Simulated Dataset. The mathematical imaging model enables the conversion (simulation) of images captured by RGB cameras into measurements captured by lensless imaging systems. This technique is commonly employed in lensless imaging fields for testing systems and imaging methods. Leveraging this, we converted numerous human pose datasets into simulated lensless measurements. These simulated datasets extended our training set and provided a quick validation of our method. + +# 4.2. Implementation Details + +The dimensions of the raw measurements are $1280 \times 1024 \times 3$ . Prior to commencing the experiments, we first crop and resize the lensless measurements. These measurements are pre-processed to $224 \times 224 \times 3$ before being fed into the network. Following this pre-processing step, the lensless measurements inputted into the global perceptual layer are $224 \times 224$ , generating spatial features at resolutions of $(7 \times 7, 14 \times 14, 28 \times 28$ , and $56 \times 56$ ). During the generation of mesh-aligned features, the SMPL mesh is downsampled using a pre-computed downsampling matrix provided in [18], resulting in a reduction of vertex count from 6890 to 431. The regressors $\mathcal{R}_t$ follow a congruent architec + +tural framework to the regressor present in HMR [14], albeit with slight variations in their input dimensions. Our network is trained using the Adam optimizer [15] with a learning rate of $5 \times 10^{-5}$ and a batch size of 80 on 4 NVIDIA RTX2080 Ti GPUs. No learning rate decay is applied during training. + +![](images/f10a5d6cc35d3b0e92422cc57562cb3b3533e990531cbefb7ae6fffe8635be20.jpg) +Measurement + +![](images/5eeaec79778a14f66d746fcf19acd3b4ae88bf9126fafba5b87e128b4af946b3.jpg) +Recovered Images + +![](images/7b9aacade28b5169d1a0e2d5447ace8304ee8a5ab49ac919afff262163acbda5.jpg) + +![](images/e19965037832bce6adaca5cb3ad8a85a14729c588f3e9e835f29d4f89c3f7eac.jpg) +Human Mesh Recovery + +![](images/1be17e5c35525bb6a809b917a0530590be269c914387c5e55980e3f238b6e905.jpg) + +![](images/f521a29a4e3abb0e3ca32acdfd25bca7d99f504db9a28724ad8dbb48241de59f.jpg) +Figure 5. At the top is the frame of our baseline. Lensless measurements first recover the image by reconstruction methods and then estimate human pose and shape by PyMAF. The bottom image compares the recovered image with the original image, and we can see that the quality of the recovered image has dropped dramatically. + +![](images/cf26471fafe17c92703833a47292d4f15ccf8a577406bdae22d1e104eb64d730.jpg) + +![](images/0f22bd1a366a8aaa59b5fb49bb95b53d26a98ade71f577d16c8fafa0b301db25.jpg) + +![](images/f7829f64eac7cc3ab9c70aab0829a6a5f121027469854164cecab124c105cd1f.jpg) + +![](images/b9613918b8caf996d178a942928a939ffc9ec0aade123b8a09d9d40230d566ac.jpg) + +![](images/3e2d8f3b8ef91ef15fd15f21d7f56d7e6ddf9685713f5bc0b07eeaac1168f74b.jpg) + +![](images/e66eea2bb41f95771ebeddc45b47e83b50fddb77e6ed99e46c5a8a705636c94e.jpg) + +# 4.3. 3D Human Pose and Shape Comparison. + +Baseline Approach. We are the first work to perform human pose and shape estimation through lensless measurements. Given the absence of other valid methods for comparison, we designed a baseline approach that approaches this as a two-stage task. The baseline architecture is depicted in Fig. 5. In the first stage, we utilized the lensless image reconstruction method Rego et. al. [28], chosen for its wide applicability in recovering images acquired by lensless imaging systems. In the second stage, PyMAF is used to estimate human pose and shape. + +
MethodMPJPE ↓PA-MPJPE↓PVE ↓
baseline (PyMAF)257.07121.73277.18
baseline (PyMAF†)126.2881.37151.60
LPSNet119.2081.52134.74
+ +Table 1. Comparison with baseline. The baseline is a combination of the two-stage approach using existing methods. LPSNet outperforms the baseline on the LenslessHuman3.6M datasets. + +Comparison with Baseline. We present the results of + +![](images/9fdeca129a39fd72df7420174d49a676f6146113e4c4666c9db1b7a60dc150a1.jpg) +Figure 6. Qualitative results of LPSNet on challenging LenslessHuman3.6M. The results on the left are from our LPSNet estimation, those in the center are from our baseline (PyMAF) output, and those on the right are from the fine-tuned baseline $(\mathrm{PyMAF}^{\dagger})$ output. LPSNet can be seen to be better than the baselines. + +the quantitative comparative evaluation on the LenslessHuman3.6M dataset in Tab. 1. Notably, images recovered from lensless measurements exhibit lower quality. Directly comparing the baseline (PyMAF) with our approach is challenging and unfair. Therefore, we fine-tuned PyMAF on the train-LenslessHuman3.6M dataset, denoted as $\mathrm{PyMAF}^{\dagger}$ . Referring to Tab. 1, we observe significant improvements in the indicators of baseline $(\mathrm{PyMAF}^{\dagger})$ . Specifically, compared with the baseline $(\mathrm{PyMAF}^{\dagger})$ , the MPJPE of LPSNet on the LenslessHuman3.6M dataset is reduced by $7.08~\mathrm{mm}$ . + +As shown in Tab. 1, our LPSNet achieved competitive results compared to the baseline approach $(\mathrm{PyMAF}^{\dagger})$ . LPSNet demonstrates more significant improvements in the MPJPE and PVE metrics. However, we would argue that the PA-MPJPE metric may not fully reveal the performance of the mesh-image alignment, as it is calculated as MPJPE after rigid alignment. + +Baseline (PyMAF) has difficulty performing accurate pose estimation on low-quality recovered images. However, the outputs of our LPSNet exhibit significant improvements compared to the two baseline methods, especially for limb estimation. Qualitative comparison results are illustrated in Fig. 6. + +# 4.4. Ablation Study + +In this section, we conduct ablation studies on LenslessHuman3.6M under various settings to validate the effectiveness of the key components proposed in our method. + +
MethodMPJPE ↓PA-MPJPE↓PVE ↓
w/o MSFDcoder and DHAS142.1392.20161.07
w/o DHAS139.3492.56158.97
w/o DHAS (2D keypoint)123.3083.41138.67
w/o DHAS (IUV)129.7086.83146.7
LPSNet119.2081.52134.74
+ +Table 2. Ablation study on LPSNet. DHAS stands for Double-Head Auxiliary Supervision. + +All ablation variants were trained and tested on LenslessHuman3.6M, which is derived from the Human3.6M dataset. The Human3.6M dataset includes ground-truth 3D labels and serves as a widely used benchmark in 3D human pose and shape estimation. + +Multi-Scale Lensless Feature Decoder. In LPSNet, the decoder primarily decodes the global information of lensless measurements encoded by the diffuser, which plays a crucial role in our end-to-end system. Our MSFDecoder is essential for efficient feature extraction from lensless measurements. We have adapted another variant of the decoder to verify this. Specifically, we simplified the global perception layer of the decoder to a series of convolutional layers and deconvolutional layers. We use this simpler decoder in the experiment(w/o MSFcoder and DHAS) instead of MSFDecoder. As shown in Tab. 2, comparing the result of experiments(w/o MSFcoder and DHAS) and experiment(w/o DHAS), we can see that MSFDecoder improves + +![](images/dfebda61820c178f7005e59c8a1ab75ea8e5f86091cf4d09398024c74520606b.jpg) +Figure 7. Qualitative results of LPSNet on challenging LenslessHuman3.6M. We can see that the results of LPSNet are better, and the results are improved with the addition of the double-head assisted supervision mechanism. + +the accuracy of the human pose and shape estimation significantly. Through the qualitative experiment in Fig. 7, we can also see that the human pose and shape estimation has a better alignment after using the MSFDecoder. Note that neither experiment(w/o MSFDcoder and DHAS) nor the experiment(w/o DHAS) used auxiliary supervision. + +Double-Head Auxiliary Supervision. Double-Head Auxiliary Supervision is mainly used to improve the accuracy of human limbs. For each auxiliary supervisory head, we conducted ablation experiments respectively. Referring to the result in Tab. 2, LPSNet has significant improvements in the MPJPE and the PVE metrics. Fig. 7 shows additional qualitative comparison results. We can see that the alignment of the human pose and shape is better with the addition of the Double-Head Auxiliary Supervision mechanism. + +![](images/c3ad17b93f94cb2ac472607d0671de20306ff76a0ae25718cf88b2ebaec0eaaa.jpg) +(a) Result on more datasets + +![](images/88a8798d1589f4b5ff5a56bdf252bba4306038000f0154014599a59b5e431424.jpg) +(b) Result on real scenes +Figure 8. Results on more datasets and real scenes. From left to right for each set of figures: lensless measurement, alignment of results with RGB images, and 3D results shown in different views. + +# 4.5. Results on more Datasets and Real Scene + +Fig. 8 (a) shows the qualitative results on various datasets after using the mixed datasets for training. Specifically, we collected the MPII, COCO, 3DPW, MIP-INF-3DHP, and LSP datasets through our lensless imaging system and mixed them for training and testing. + +In addition, we also capture the real scenes of two different people, in different backgrounds and different light intensities using our lensless imaging system. Fig. 8 (b) illustrates the experimental results of our method on these real scenes, indicating its suitability for real-world applications. + +# 5. Conclusion and Discussion + +Conclusion. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to the best of our knowledge. Specifically, we design a multi-scale lensless feature decoder, which can effectively decode the information produced by the lensless imaging system. We also propose a double auxiliary supervision mechanism to improve the accuracy of human limb end estimation. Experimental results show that our method can achieve end-to-end human pose and shape estimation through lensless measurements. + +Limitations and Future Work. Our work is an initial stride step in this direction, however, the results are still not as robust as those achieved by traditional methods for estimating the human body from RGB images. The current scarcity of lensless datasets hampers our ability to pre-train the MSFDecoder using a large dataset like ImageNet [9], thereby limiting the generalizability of our approach. We will continue this work in the future, focusing on generating more extensive lensless datasets and progressing towards practical applications. + +Acknowledgments. This work was supported in part by the National Key R&D Program of China (2023YFC3082100), the National Natural Science Foundation of China (62122058 and 62171317), and the Science Fund for Distinguished Young Scholars of Tianjin (No. 22JCJQJC00040). + +# References + +[1] Jesse K Adams, Vivek Boominathan, Benjamin W Avants, Daniel G Vercosa, Fan Ye, Richard G Baraniuk, Jacob T Robinson, and Ashok Veeraraghavan. Single-Frame 3D Fluorescence Microscopy with Ultraminiature Lensless FlatScope. Sci Adv., 3(12):e1701548, 2017. 2 +[2] Nick Antipa, Grace Kuo, Reinhard Heckel, Ben Mildenhall, Emrah Bostan, Ren Ng, and Laura Waller. DiffuserCam: Lensless Single-exposure 3D Imaging. Optica, 5(1):1-9, 2018. 2, 6 +[3] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting Temporal Context for 3D Human Pose Estimation in the ild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3395-3404, 2019. 3 +[4] M Salman Asif, Ali Ayremlou, Aswin Sankaranarayanan, Ashok Veeraraghavan, and Richard G Baraniuk. Flatcam: Thin, Lensless Cameras using Coded Aperture and Computation. IEEE Trans Comput Imaging., 3(3):384-397, 2016. 2 +[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In *Eur. Conf. Comput. Vis. Springer International Publishing*, 2016. 3 +[6] Vivek Boominathan, Jesse K Adams, M Salman Asif, Benjamin W Avants, Jacob T Robinson, Richard G Baraniuk, Aswin C Sankaranarayanan, and Ashok Veeraraghavan. Lensless Imaging: A Computational Renaissance. IEEE Signal Process Mag., 33(5):23-35, 2016. 1 +[7] Vivek Boominathan, Jesse K Adams, Jacob T Robinson, and Ashok Veeraraghavan. Phlatcam: Designed Phase-Mask Based Thin Lensless Camera. IEEE Trans. Pattern Anal. Mach. Intell., 42(7):1618-1629, 2020. 1, 2, 6 +[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In Eur. Conf. Comput. Vis., pages 342–359. Springer, 2022. 3 +[9] J. Deng, W. Dong, R. Socher, L. Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPR Workshops). 8 +[10] Georgios Georgakis, Ren Li, Srikrishna Karanam, Terrence Chen, Jana Košecka, and Ziyan Wu. Hierarchical Kinematic Human Mesh Recovery. In Eur. Conf. Comput. Vis., pages 768-784. Springer, 2020. 3 +[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and Tracking Humans with Transformers. In Int. Conf. Comput. Vis., pages 14783-14794, 2023. 1, 3 +[12] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense Human Pose Estimation in the Wild. In Int. Conf. Comput. Vis., pages 7297-7306, 2018. 5 +[13] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. In Int. Conf. 3D. Vis., pages 42-52. IEEE, 2021. 3 + +[14] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end Recovery of Human Shape and Pose. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 1, 3, 4, 6 +[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[16] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. PARE: Part Attention Regressor for 3D Human Body Estimation. In Int. Conf. Comput. Vis., pages 11127-11137, 2021. 3 +[17] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to Reconstruct 3D Human Pose and Shape Via Model-fitting in the Loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 3 +[18] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional Mesh Regression for Single-Image Human Shape Reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4501–4510, 2019. 3, 6 +[19] Grace Kuo, Nick Antipa, Ren Ng, and Laura Waller. DiffuserCam: Diffuser-Based Lensless Cameras. In Computational Optical Sensing and Imaging, pages CTu3B-2. Optica Publishing Group, 2017. 1, 2 +[20] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 3 +[21] Vincent Leroy, Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, and Grégory Rogez. SMPLy Benchmarking 3D Human Pose Estimation in the wild. In Int. Conf. 3D. Vis., pages 301-310. IEEE, 2020. 3 +[22] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation. In Eur. Conf. Comput. Vis., pages 89-106. Springer, 2022. 5 +[23] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1954-1963, 2021. 3 +[24] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh Graphormer. In Int. Conf. Comput. Vis., pages 12939-12948, 2021. 3 +[25] Aydogan Ozcan and Euan McLeod. Lensless Imaging and Sensing. Annual review of biomedical engineering, 18:77-102, 2016. 1 +[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3 +[27] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human Mesh Recovery from Multiple Shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 3 +[28] Joshua D Rego, Karthik Kulkarni, and Suren Jayasuriya. Robust Lensless Image Reconstruction Via PSF Estimation. In + +Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 403-412, 2021. 6 +[29] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D Human Motion Model for Robust Pose Estimation. In Int. Conf. Comput. Vis., 2021. 3 +[30] David G Stork and Patrick R Gill. Lensless Ultra-miniature CMOS Computational Imagers and Sensors. Proc. Sensorcomm, pages 186-190, 2013. 2 +[31] David G Stork and Patrick R Gill. Optical, mathematical, and computational foundations of lensless ultra-miniature diffractive imagers and sensors. International Journal on Advances in Systems and Measurements, 7(3):4, 2014. 2 +[32] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep High-resolution Representation Learning for Human Pose Estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5693-5703, 2019. 3 +[33] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. PoseNDF: Modeling Human Pose Manifolds with Neural Distance Fields. In Eur. Conf. Comput. Vis., 2022. 3 +[34] Hongwen Zhang, Jie Cao, Guo Lu, Wanli Ouyang, and Zhenan Sun. Learning 3D Human Shape and Pose from Dense Body Parts. IEEE Trans. Pattern Anal. Mach. Intell., 44(5):2610-2627, 2020. 1 +[35] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Int. Conf. Comput. Vis., pages 11446-11456, 2021. 1, 3, 4 +[36] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3D Human Dynamics from Video. In Int. Conf. Comput. Vis., pages 7114-7123, 2019. 3 \ No newline at end of file diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/images.zip b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..917d0e465ecd2cb13c88e7c481e6d72a113f621f --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6699c24c7791c7c8cfb395893930d0e7091771800d8549a4d155242e191e953 +size 539168 diff --git a/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/layout.json b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..021cb1a58e2b3e73371bc35b2f616cea7bdfe856 --- /dev/null +++ b/2024/LPSNet_ End-to-End Human Pose and Shape Estimation with Lensless Imaging/layout.json @@ -0,0 +1,7959 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "text", + "content": "LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": "Haoyang Ge" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Qiao Feng" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Hailong Jia" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Xiongzheng Li" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Xiangjun Yin" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", You Zhou" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Jingyu Yang" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "text", + "content": ", Kun Li" + }, + { + "bbox": [ + 111, + 141, + 481, + 171 + ], + "type": "inline_equation", + "content": "^{1,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "spans": [ + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "type": "text", + "content": "Tianjin University, China " + }, + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 161, + 171, + 434, + 186 + ], + "type": "text", + "content": "Nanjing University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 187, + 506, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 187, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 83, + 187, + 506, + 200 + ], + "type": "text", + "content": "{ghy0623,fengqiao,jhl,lxz,yinxiangjun,yjy,lik}@tju.edu.cn zhouyou@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 55, + 236, + 544, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 544, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 544, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 544, + 453 + ], + "type": "image", + "image_path": "271240214a8cb7c489998d4b4fadf00997a8b09d5b26bde665705f970bd6be0b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 464, + 546, + 509 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 546, + 509 + ], + "type": "text", + "content": "Figure 1. LPSNet: End-to-End Human Pose and Shape Estimation with Lensless Imaging. We contribute a framework for estimating human poses and shapes from individual lensless measurements. The first row shows the input measurements acquired by our lensless imaging system, the second row shows the estimated human poses and shapes from lensless measurements, and the bottom row shows the 3D results shown in different views." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 518, + 192, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 518, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 143, + 518, + 192, + 532 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 687 + ], + "type": "text", + "content": "Human pose and shape (HPS) estimation with lensless imaging is not only beneficial to privacy protection but also can be used in covert surveillance scenarios due to the small size and simple structure of this device. However, this task presents significant challenges due to the inherent ambiguity of the captured measurements and lacks effective methods for directly estimating human pose and shape from lensless data. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to our knowledge. We specifically design a multi-scale lensless feature decoder to decode the lensless measurements through the optically encoded mask" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 520, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 545, + 605 + ], + "type": "text", + "content": "for efficient feature extraction. We also propose a doublehead auxiliary supervision mechanism to improve the estimation accuracy of human limb ends. Besides, we establish a lensless imaging system and verify the effectiveness of our method on various datasets acquired by our lensless imaging system. The code and dataset are available at https://cic.tju.edu.cn/faculty/likun/projects/LPSNet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 632, + 386, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 386, + 644 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 386, + 644 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 654, + 547, + 713 + ], + "type": "text", + "content": "In recent years, lensless imaging technologies [6, 7, 19, 25] have advanced significantly due to their many advantages, such as privacy protection, smaller size, simple structure, and lower cost. 3D human pose and shape (HPS) estimation [11, 14, 34, 35] requires miniaturized and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 693, + 131, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 693, + 131, + 703 + ], + "spans": [ + { + "bbox": [ + 62, + 693, + 131, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 62, + 693, + 131, + 703 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 63, + 704, + 141, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 704, + 141, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 704, + 141, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "1471" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "lightweight imaging system as application scenarios become more diverse. All these advantages, especially privacy, make the lensless imaging system very suitable as an imaging device for human pose and shape estimation. In this work, we propose LPSNet, which aims to estimate 3D human pose and shape from lensless measurements instead of RGB images, achieving cheaper and privacy-protecting 3D human pose and shape estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "type": "text", + "content": "A thin, lightweight, and potentially cost-effective optical encoder is used in a lensless imaging system instead of traditional cameras with lenses, while others are expensive, rigid, and occupy more space. At this stage, the application of a lensless imaging system is extensive, it is mainly used in the field of microscopic imaging, RGB image reconstruction, and so on. More valuable information can be obtained from lensless measurements due to the special optical encoding method of lensless imaging systems. Directly estimating human pose and shape from lensless measurements is not currently possible. The first step is to reconstruct an RGB image from a lensless measurement and then estimate the human pose and shape from the RGB images. However, experimental findings indicate that the reconstructed RGB images are of suboptimal quality, resulting in incomplete local features and significant deviations in the position of the human body. Combining these factors leads to inaccurate human pose estimation when using lensless measurements to reconstruct RGB images. This approach has computational burden and computational resources, making it very unfavorable for deployment at the endpoint." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 422, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 289, + 517 + ], + "type": "text", + "content": "In this paper, We aim to advance human pose and shape estimation using a lensless imaging system, which needs to overcome two main challenges. First, how to extract features from lensless measurements for human pose and shape estimation. Secondly, during early experiments, when using features extracted from lensless measurements to estimate human pose and shape, we found poor estimation accuracy of human limbs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 519, + 289, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 289, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 289, + 686 + ], + "type": "text", + "content": "To address these challenges, we introduce LPSNet, the first end-to-end human pose and shape estimation framework with lensless imaging. To extract features from optically encoded lensless measurements, we propose a multiscale lensless feature decoder (MSFDecoder). Specifically, we introduce a global perception layer to enhance the global decoding capability of MSFDecoder. The global information that has been optically encoded to the global can be efficiently decoded to obtain a feature map that can be used in subsequent processes. To improve pose and shape estimation, we propose a Double-Head Auxiliary Supervision(DHAS) mechanism to be implemented during training. Auxiliary supervision can improve the estimation accuracy of human limbs and correct results with large deviations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 42, + 688, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 42, + 688, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 42, + 688, + 289, + 714 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows: 1) We propose LPSNet, an end-to-end pose and shape esti" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "type": "text", + "content": "mation network for lensless imaging systems. This is the first work to estimate human poses and shapes directly from lensless measurements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 301, + 108, + 547, + 192 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 301, + 108, + 547, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 108, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 301, + 108, + 547, + 155 + ], + "type": "text", + "content": "2) We propose MSFDecoder, a Multi-Scale Lensless Feature Decoder that decodes and extracts features from lensless measurements, which can be receptive to global features in lensless measurements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 156, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 156, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 301, + 156, + 547, + 192 + ], + "type": "text", + "content": "3) We propose a Double-Head Auxiliary Supervision mechanism for both pose and shape estimation, which can improve the estimation accuracy of human limbs." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 205, + 393, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 393, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 393, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 225, + 446, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 225, + 446, + 238 + ], + "spans": [ + { + "bbox": [ + 306, + 225, + 446, + 238 + ], + "type": "text", + "content": "2.1. Lensless Imaging System" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 244, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 244, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 244, + 545, + 424 + ], + "type": "text", + "content": "A conventional photographic camera typically comprises a focusing lens, which may consist of one or more optical elements and an image sensor positioned at or near the focal length of the lens. The lens in such a camera directs the projected light from the observed scene onto the sensor, aiming to accurately map specific scene points to individual pixels on the sensor. Conversely, in a lensless imaging system, the absence of a lens defines its configuration. Instead, an optical modulator, such as a coded amplitude mask or a diffuser, is positioned between the scene and the image sensor, often close to the sensor itself. As a result, the recorded data deviates significantly from the expected RGB image during imaging. In this process, the local information of the object transforms overlapping global information through the optically encoded mask." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 533 + ], + "type": "text", + "content": "Within the mask-modulated lensless systems, a fixed optical mask is introduced to create a versatile lensless system that can work for a wide range of object distances and lighting scenarios, whether passive or uncontrolled. The mask modulates the incoming light and generates a measurement that can be decoded through computational methodologies. Mask-modulated lensless imaging systems were used to perform 2D imaging [4, 7, 19], refocusing [7], 3D imaging [2, 7], and microscopic imaging [1, 2, 7]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 688 + ], + "type": "text", + "content": "Generally speaking, amplitude modulators and phase modulators [1, 4] are the two types of optical masks used in lensless imaging systems. Phase modulators can be further sub-categorized into phase gratings [30, 31], diffusers [2], and phase masks [7]. One key characteristic of a mask-modulated lensless system is the pattern the mask produces on the sensor for a point light source in the scene. We call this pattern the Point-Spread Function(PSF), and its properties determine the imaging model of the system. As shown in Fig. 2, we design a simple mask-modulated lensless system for data acquisition in this experiment. The lensless imaging system designed in our experiments chose a diffuser as the mask." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "The main advantages of lensless imaging are as follows: lensless is small in size and can be assembled in a variety of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1472" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 284, + 157 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 284, + 157 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 284, + 157 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 284, + 157 + ], + "type": "image", + "image_path": "f33d35d1c12c6238da0bf12832073986d998c84e2c189df320c8054ad2cf087d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 287, + 222 + ], + "type": "text", + "content": "Figure 2. The workflow of the lensless imaging system and the final measurement obtained is the result obtained from the encoding of 3D scene information by the lensless imaging system. The optically encoded mask transforms local information in the 3D scene into overlapping global information." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 304 + ], + "type": "text", + "content": "miniaturized equipment; lensless is light and easy to carry; lensless cameras are still cheap to make; lensless cameras have a wide view field, larger than traditional wide-angle cameras; It can also protect the user's privacy to a great extent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 228, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 228, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 228, + 325 + ], + "type": "text", + "content": "2.2. Human Pose and Shape Recovery" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 331, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 413 + ], + "type": "text", + "content": "There are two primary methods for human pose and shape estimation: optimization-based approaches and regression-based approaches. Numerous methodologies have been developed to estimate the three-dimensional human pose and shape by employing iterative optimization techniques, e.g., SMPLify [5] and variations of the SMPLify [20, 26, 29, 33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "type": "text", + "content": "Many existing Regression-based methods follow the architecture of HMR [14], which uses a pre-trained backbone to extract image features followed by regression to obtain SMPL [5] parameters. Many improvements to the original method have been proposed since its introduction. In particular, many papers have proposed alternative methods for pseudo-ground truth generation, including using temporal information [3], multiple views [21], or iterative optimization [13, 17, 27]. SPIN [17] proposed an in-the-loop optimization that incorporated SMPLify [5] in the HMR training, which combines the optimization-based approach and the regression-based approach for human pose and shape estimation. PyMAF [35, 36] proposes a pyramidal mesh alignment feedback loop for regression of SMPL parameters. Our approach references the design of PyMAF in the part of the SMPL parameters regression. PARE [16] proposes a body-part-guided attention mechanism for better occlusion handling. HKMR [10] performs a prediction guided by the hierarchical structure of SMPL. HMR2.0 [11] employs a large training dataset and proposes a fully transformer-based approach for 3D human pose and shape estimation from a single image. Many related approaches make non-parametric predictions, i.e., instead of estimating the parameters of the SMPL model, they explicitly regress the vertices of the mesh. GraphCMR [18] uses a" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "graph neural network for the prediction, METRO [23] and FastMETRO [8] use a transformer, while Mesh Graphormer [24] adopts a hybrid between the two." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 193 + ], + "type": "text", + "content": "In this paper, we propose the first work to perform end-to-end human pose and shape estimation from lensless measurements. We designed a multi-scale lensless feature decoder to decode the lensless measurements based on a mask-encoder to obtain more efficient features. We also propose a double auxiliary supervision mechanism to improve the estimation accuracy of human limbs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 204, + 362, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 204, + 362, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 362, + 216 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 224, + 373, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 373, + 236 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 373, + 236 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 398 + ], + "type": "text", + "content": "Our work focuses on human pose and shape estimation from lensless measurements. In this section, we present the technical details of our approach. As depicted in Fig. 3, the core of our method involves the following three components: 1) a Multi-Scale Lensless Feature Decoder(MSFDecoder) (Sec. 3.2) that can effectively decode the information encoded by the lensless imaging system; 2) a human parametric model regressor (Sec. 3.3) that takes the multi-scale features produced by MSFDecoder as input and predicts the SMPL parameters; 3) a Double-Head Auxiliary Supervision mechanism(DHAS) (Sec. 3.4) that can assist the LPSNet to improve the estimation accuracy of human limbs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 407, + 504, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 504, + 419 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 504, + 419 + ], + "type": "text", + "content": "3.2. Multi-Scale Lensless Feature Decoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "text", + "content": "As illustrated in Fig. 2, the optical information of the subject is diffused by the mask and projected onto the sensor. The process of obtaining measurements through the lensless imaging system can be treated as encoding optical information in the scene into lensless measurements. The goal of our lensless feature decoder " + }, + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "inline_equation", + "content": "D_{M}" + }, + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "text", + "content": " is to decode these lensless measurements to multi-scale features, which are then utilized for subsequent human body prediction. The design of the global perception layer " + }, + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "inline_equation", + "content": "G_{P}" + }, + { + "bbox": [ + 304, + 426, + 545, + 640 + ], + "type": "text", + "content": " is inspired by HRNet [32]. The global perception layer inherits many of the advantages of HRNet [32] and can always maintain a high resolution, while the information interaction between different branches complements the information loss caused by the reduction in the number of channels. These advantages are important for extracting features from lensless measurements. Then, we design different convolutional layers for different scales of features that are used to refine the features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Formally, the decoder takes a lensless measurement " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " as input and decodes the information encoded by the lensless imaging system. The global perception layer ensures the integrity of global information for each pixel by maintaining the high-resolution convolution branch and the low-resolution convolution branch in parallel. It further en" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1473" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 75, + 532, + 295 + ], + "blocks": [ + { + "bbox": [ + 65, + 75, + 532, + 295 + ], + "lines": [ + { + "bbox": [ + 65, + 75, + 532, + 295 + ], + "spans": [ + { + "bbox": [ + 65, + 75, + 532, + 295 + ], + "type": "image", + "image_path": "748807a183ac6e648e492574ea5e5f7ac01516841e483c4af533ddd54b82fd02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 307, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 547, + 342 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed framework. A measurement " + }, + { + "bbox": [ + 46, + 307, + 547, + 342 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 307, + 547, + 342 + ], + "type": "text", + "content": " is passed through a Multi-Scale Lensless Feature Decoder to get spatial characteristics at different scales. These feature maps are fed into the regressor for human pose and shape estimation. Also, these feature maps are fed into the Double-Head Auxiliary Supervision to improve the estimation accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": "hances the information extraction accuracy by repeating the multi-scale fusion across parallel convolutions. As shown in Fig. 3, the global perception layer takes " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " as input and generates a set of multi-scale spatial features " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\phi_{m}^{t = 0,1,2,3}" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " as input for the subsequent steps. The final output " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\phi_{M}^{t}" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " of the decoder is obtained by passing " + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "inline_equation", + "content": "\\phi_{m}^{t}" + }, + { + "bbox": [ + 46, + 361, + 289, + 445 + ], + "type": "text", + "content": " into different convolutional layers," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 445, + 287, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 445, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 130, + 445, + 287, + 459 + ], + "type": "interline_equation", + "content": "\\phi_ {M} ^ {t} = D _ {M} (M). \\tag {1}", + "image_path": "f3a7cc97d25f9dbec2f5e4d64b458ea021af5b9c4454957ea63e5c0fc15ac352.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "text", + "content": "With different scales of decoding, we can obtain feature maps at different granularities, through which we can perceive the location of people in the scene. The results obtained by the lensless decoder are utilized for subsequent predictions with the SMPL model, incorporating pose, shape, and camera parameters " + }, + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\theta ,\\beta ,\\widetilde{\\pi}\\}" + }, + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\widetilde{\\pi}" + }, + { + "bbox": [ + 46, + 461, + 289, + 558 + ], + "type": "text", + "content": " is pseudo-camera parameters for the subsequent projection calculation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 249, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 249, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 249, + 577 + ], + "type": "text", + "content": "3.3. Human Parametric Model Regression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Our human parametric model regression is mainly inspired by the improved Human Mesh Regression described in PyMAF [35]. Different from the pyramid features obtained by deconvolution in PyMAF [35], our global perception layer exploits HRNet [14] to maintain more high-resolution information. We use a set of sampling points " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " and project them onto the feature maps to extract point-wise features. The point-wise features corresponding to each point " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " are bilinearly sampled from " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\phi_M^{t=1,2,3}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". (Note that " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\phi_M^{t=0}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " is unnecessary here and is used as input for subsequent initialization.) Then the dimensions of these features" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 361, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 396 + ], + "type": "text", + "content": "will be reduced by MLP(multi-layer perceptron) and concatenated as a feature vector " + }, + { + "bbox": [ + 304, + 361, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_M^t" + }, + { + "bbox": [ + 304, + 361, + 547, + 396 + ], + "type": "text", + "content": " which can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 350, + 402, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 402, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 350, + 402, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\phi} _ {M} ^ {t} = \\mathcal {F} \\left(\\phi_ {M} ^ {t}, S _ {t}\\right) \\tag {2} \\\\ = \\oplus \\left\\{f \\left(\\phi_ {M} ^ {t} (s)\\right), \\text {f o r} s \\text {i n} S _ {t} \\right\\}, \\\\ \\end{array}", + "image_path": "69417e09122cb08a0b47b15ebbba9562ed28a99f34e540111547ed3332d15098.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " denotes the overall point-wise feature extraction, " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " is the MLP layer, and " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " denotes the concatenation. We use an iterative mechanism to complete the interaction of different scale features. Sequentially, a parameter regressor " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " takes features " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\hat{\\phi}_M^t" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " and the current parameter estimation " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\Theta_t" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " as inputs and outputs the parameter residual. Parameters are updated as " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\Theta_{t + 1}" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " by adding the residual to " + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\Theta_t" + }, + { + "bbox": [ + 304, + 440, + 547, + 536 + ], + "type": "text", + "content": " which can be formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 542, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 542, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 341, + 542, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\Theta_ {t + 1} = \\Theta_ {t} + \\mathcal {R} _ {t} \\left(\\Theta_ {t}, \\hat {\\phi} _ {M} ^ {t}\\right), \\text {f o r} t > 0. \\tag {3}", + "image_path": "f3cabbb00e7c00ae8b6b7aa1cb35cc9c48bd6e4eb321cbfc329ba57bf6d178c0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "text", + "content": "The initial parameter " + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\Theta_0" + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "text", + "content": " is initialized by feeding " + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\phi_M^{t=0}" + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "text", + "content": " into the regressor " + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 304, + 569, + 545, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "When the predicted parameters " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " (the subscript " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is omitted for simplicity) are obtained through each scale feature, a mesh with vertices of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "V = \\mathcal{M}(\\pmb {\\theta},\\pmb {\\beta})\\in \\mathbb{R}^{N\\times 3}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " can be generated accordingly, where " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "N = 6890" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " is the number of vertices in the SMPL model. We downsample " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " to get the sampling point " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{t + 1}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " for the next iteration. These mesh vertices can be mapped by pre-trained linear regressor to obtain sparse 3D keypoints " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "J\\in \\mathbb{R}^{N_j\\times 3}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". With the estimated pseudo-camera parameters, we can obtain 2D keypoints " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "K\\in \\mathbb{R}^{N_j\\times 2}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " by projecting " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " on the measurement coordinate system. We" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1474" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "calculate the loss between the 2D keypoints and the ground truth, in this way reducing the difference between the 2D projection and the human pose in the real scene. Concurrently, supplementary 3D supervisions regarding 3D keypoints and model parameters are integrated when authentic 3D labels are available. The total loss function for the parameter regressor is thereby formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 173, + 287, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 173, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 52, + 173, + 287, + 189 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {2 d} \\| K - \\hat {K} \\| _ {2} + \\lambda_ {3 d} \\| J - \\hat {J} \\| _ {2} + \\lambda_ {\\text {p a r a}} \\| \\Theta - \\hat {\\Theta} \\| _ {2}, \\tag {4}", + "image_path": "cdc295d85f3cc28d18e8f8672235faabe166585ad58d39efbf1c5e66085e411a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "text", + "content": " is the squared L2 norm, " + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\hat{K},\\hat{J}" + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta}" + }, + { + "bbox": [ + 47, + 194, + 287, + 232 + ], + "type": "text", + "content": " denote the ground truth 2D keypoints, 3D keypoints, and model parameters, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 237, + 237, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 237, + 237, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 237, + 237, + 251 + ], + "type": "text", + "content": "3.4. Double-Head Auxiliary Supervision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 339 + ], + "type": "text", + "content": "The spatial feature map of the human body is relatively rough and contains a lot of noise. Therefore, there are still some deviations in the perception of human limbs and pose. To improve our estimation accuracy of human limbs, we introduce a Double-Head Auxiliary Supervision (DHAS) mechanism to obtain finer spatial features during the training stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "content": "Specifically, we first convert all the spatial features at different scales to a single feature map " + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\phi^T" + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "content": " by upsampling and then concatenating them together. The feature map " + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\phi^T" + }, + { + "bbox": [ + 46, + 339, + 287, + 471 + ], + "type": "text", + "content": " is used for the different auxiliary supervision heads. On the one hand, we generate a heat map representation by a classification layer to explicitly indicate the position of the 2d keypoints. On the other hand, we also estimate a dense map through the IUV Predict layer to get a dense correspondence between the SMPL model and the feature map. The loss function for the Double-Head Auxiliary Supervision consists of two parts, which can be written as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 477, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 477, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 124, + 477, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d a s} = \\mathcal {L} _ {s c} + \\mathcal {L} _ {d e n}. \\tag {5}", + "image_path": "7143ade436db97ebb8fd84433bef6605dc4ed943ba71b91ac8b1cffceefdd89c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "text", + "content": "Keypoints Supervision. We utilize the SimCC-based algorithm [22] for predicting pose keypoints. This approach treats keypoint localization as a classification task in horizontal and vertical coordinates. During training, instead of estimating the actual coordinates, we employ two vectors, " + }, + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 496, + 287, + 592 + ], + "type": "text", + "content": ", and convert the ground truth 2D keypoints into such vectors to compute the loss. The loss function is then formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 598, + 287, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 598, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 71, + 598, + 287, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s c} = \\lambda_ {x y} (\\operatorname {K L - L o s s} (x, \\hat {x}) + \\operatorname {K L - L o s s} (y, \\hat {y})), \\tag {6}", + "image_path": "806fe29abcb207c95dcf3669df48dfc0f2c3140b5bde61096d52dd9e19250ce3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "content": "where KL-Loss is the Kullback-Leibler divergence loss, " + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "content": " denote the processed ground truth 2D keypoints, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "IUV Supervision. We adopt the IUV mapping defined in DensePose [12] as the dense correspondence representation. This mapping establishes deterministic correspondences between foreground pixels in 2D images and vertices on 3D surfaces. The vertices on the template mesh can" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 74, + 417, + 175 + ], + "blocks": [ + { + "bbox": [ + 325, + 74, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 325, + 74, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 325, + 74, + 417, + 175 + ], + "type": "image", + "image_path": "9801a94d8c8e0d9e7d8fcfba90bb98213ffae6ac37d4f64c99ed86cd429ff017.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 419, + 74, + 522, + 173 + ], + "blocks": [ + { + "bbox": [ + 419, + 74, + 522, + 173 + ], + "lines": [ + { + "bbox": [ + 419, + 74, + 522, + 173 + ], + "spans": [ + { + "bbox": [ + 419, + 74, + 522, + 173 + ], + "type": "image", + "image_path": "6803f5ecdd031d34dedaab0371fc523d820aaea6c37e86c1db0459a15c6f9d48.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 179, + 525, + 288 + ], + "blocks": [ + { + "bbox": [ + 318, + 179, + 525, + 288 + ], + "lines": [ + { + "bbox": [ + 318, + 179, + 525, + 288 + ], + "spans": [ + { + "bbox": [ + 318, + 179, + 525, + 288 + ], + "type": "image", + "image_path": "0445d025768d4be19211f97d3cf9d98e13239ea6bdafafd1a96579cdfbbd7f01.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 295, + 546, + 351 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 546, + 351 + ], + "type": "text", + "content": "Figure 4. On the top left is a figure of the dual camera system we created. This system consists of an RGB camera and a lensless imaging system using a diffuser as a mask. The top right shows the process of collecting our dataset. The image at the bottom is a record of the process of collecting our dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "text", + "content": "be mapped back to pixels on the foreground using a predefined bijection mapping between 3D surface space and 2D UV space. The dense correspondence representation comprises the index of body parts " + }, + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "text", + "content": " and UV values of mesh vertices. During training, we apply classification loss to the index of body parts " + }, + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 371, + 545, + 467 + ], + "type": "text", + "content": " and regression loss to the UV channels of the dense correspondence mapping. The loss function is formulated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 357, + 474, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 474, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 357, + 474, + 545, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {d e n} = \\lambda_ {p i} \\text {C r o s s E n t r o p y} (P, \\hat {P}) \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (U, \\hat {U}) \\tag {7} \\\\ + \\lambda_ {u v} \\operatorname {S m o o t h L 1} (V, \\hat {V}) \\\\ \\end{array}", + "image_path": "71402501c16c60612364fb9421ed241b9ddd291a2993a4c623136ead5ba2c474.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "inline_equation", + "content": "\\hat{P},\\hat{U}" + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "content": " denote the ground truth of " + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "inline_equation", + "content": "UV" + }, + { + "bbox": [ + 305, + 532, + 544, + 557 + ], + "type": "text", + "content": " values, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "spans": [ + { + "bbox": [ + 306, + 567, + 387, + 581 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 587, + 462, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 462, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 462, + 601 + ], + "type": "text", + "content": "4.1. Imaging System and Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Lensless Imaging System. Due to the scarcity of human pose and shape datasets based on lensless imaging systems, we developed a basic lensless system to gather experimental data. The system, illustrated in Fig. 4, comprises two primary components: a lensless imaging system for capturing lensless measurements and an RGB camera for capturing real images corresponding to the lensless measurements. To ensure identical light paths, we employed a beam splitter between the lensless sensor and the RGB camera. This al" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1475" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "lowed for consistent imaging conditions and facilitated accurate correlation between the lensless measurements and the corresponding real images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 109, + 286, + 180 + ], + "type": "text", + "content": "For the lensless imaging part of the system, we adopted the Mask-Modulated Lensless System, inspired by design options such as diffusercam [2] and phlatcam [7]. Specifically, we selected a diffuser as our mask and incorporated PhlatCam's concept of using larger sensors to enhance global information perception." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 286, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 286, + 252 + ], + "type": "text", + "content": "Datasets. The inputs to LPSNet are lensless measurements. However, datasets for classical human pose estimation are not directly available at present. Leveraging the lensless imaging system we have constructed, and the mathematical model of the system's imaging, the sources of datasets for our experiments can be categorized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 255, + 286, + 540 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 255, + 286, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 286, + 432 + ], + "type": "text", + "content": "- Real Dataset. Using a lensless imaging system to capture images displayed on a screen as measurements is currently the primary method of acquiring datasets in the lensless field. We used this approach to collect datasets from various sources, including Human3.6M, MPII, COCO, 3DPW, and MIP-INF-3DHP datasets, which we refer to as LenslessHuman3.6M, LenslessMPII, LenslessMIP-INF-3DHP, LenslessCOCO, and Lensless3DPW, respectively. We also reclassified the training dataset and validation dataset on LenslessHuman3.6M, referred to as train-LenslessHuman3.6M and eval-LenslessHuman3.6M, respectively. Moreover, we capture the real scenes of different individuals, backgrounds, and light intensities using our lensless imaging system." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 434, + 286, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 286, + 540 + ], + "type": "text", + "content": "- Simulated Dataset. The mathematical imaging model enables the conversion (simulation) of images captured by RGB cameras into measurements captured by lensless imaging systems. This technique is commonly employed in lensless imaging fields for testing systems and imaging methods. Leveraging this, we converted numerous human pose datasets into simulated lensless measurements. These simulated datasets extended our training set and provided a quick validation of our method." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 551, + 179, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 179, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 179, + 562 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "The dimensions of the raw measurements are " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "1280 \\times 1024 \\times 3" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ". Prior to commencing the experiments, we first crop and resize the lensless measurements. These measurements are pre-processed to " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224 \\times 3" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": " before being fed into the network. Following this pre-processing step, the lensless measurements inputted into the global perceptual layer are " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", generating spatial features at resolutions of " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "(7 \\times 7, 14 \\times 14, 28 \\times 28" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "56 \\times 56" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": "). During the generation of mesh-aligned features, the SMPL mesh is downsampled using a pre-computed downsampling matrix provided in [18], resulting in a reduction of vertex count from 6890 to 431. The regressors " + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_t" + }, + { + "bbox": [ + 46, + 570, + 286, + 712 + ], + "type": "text", + "content": " follow a congruent architec" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "tural framework to the regressor present in HMR [14], albeit with slight variations in their input dimensions. Our network is trained using the Adam optimizer [15] with a learning rate of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and a batch size of 80 on 4 NVIDIA RTX2080 Ti GPUs. No learning rate decay is applied during training." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 154, + 369, + 206 + ], + "blocks": [ + { + "bbox": [ + 317, + 154, + 369, + 206 + ], + "lines": [ + { + "bbox": [ + 317, + 154, + 369, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 369, + 206 + ], + "type": "image", + "image_path": "f10a5d6cc35d3b0e92422cc57562cb3b3533e990531cbefb7ae6fffe8635be20.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 207, + 361, + 213 + ], + "lines": [ + { + "bbox": [ + 323, + 207, + 361, + 213 + ], + "spans": [ + { + "bbox": [ + 323, + 207, + 361, + 213 + ], + "type": "text", + "content": "Measurement" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 154, + 444, + 206 + ], + "blocks": [ + { + "bbox": [ + 369, + 154, + 444, + 206 + ], + "lines": [ + { + "bbox": [ + 369, + 154, + 444, + 206 + ], + "spans": [ + { + "bbox": [ + 369, + 154, + 444, + 206 + ], + "type": "image", + "image_path": "5eeaec79778a14f66d746fcf19acd3b4ae88bf9126fafba5b87e128b4af946b3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 208, + 441, + 215 + ], + "lines": [ + { + "bbox": [ + 392, + 208, + 441, + 215 + ], + "spans": [ + { + "bbox": [ + 392, + 208, + 441, + 215 + ], + "type": "text", + "content": "Recovered Images" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 445, + 172, + 473, + 190 + ], + "blocks": [ + { + "bbox": [ + 445, + 172, + 473, + 190 + ], + "lines": [ + { + "bbox": [ + 445, + 172, + 473, + 190 + ], + "spans": [ + { + "bbox": [ + 445, + 172, + 473, + 190 + ], + "type": "image", + "image_path": "7b9aacade28b5169d1a0e2d5447ace8304ee8a5ab49ac919afff262163acbda5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 191, + 472, + 198 + ], + "lines": [ + { + "bbox": [ + 449, + 191, + 472, + 198 + ], + "spans": [ + { + "bbox": [ + 449, + 191, + 472, + 198 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 476, + 155, + 527, + 207 + ], + "blocks": [ + { + "bbox": [ + 476, + 155, + 527, + 207 + ], + "lines": [ + { + "bbox": [ + 476, + 155, + 527, + 207 + ], + "spans": [ + { + "bbox": [ + 476, + 155, + 527, + 207 + ], + "type": "image", + "image_path": "e19965037832bce6adaca5cb3ad8a85a14729c588f3e9e835f29d4f89c3f7eac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 209, + 533, + 216 + ], + "lines": [ + { + "bbox": [ + 471, + 209, + 533, + 216 + ], + "spans": [ + { + "bbox": [ + 471, + 209, + 533, + 216 + ], + "type": "text", + "content": "Human Mesh Recovery" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 313, + 228, + 366, + 281 + ], + "blocks": [ + { + "bbox": [ + 313, + 228, + 366, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 366, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 366, + 281 + ], + "type": "image", + "image_path": "1be17e5c35525bb6a809b917a0530590be269c914387c5e55980e3f238b6e905.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 313, + 283, + 365, + 336 + ], + "blocks": [ + { + "bbox": [ + 313, + 283, + 365, + 336 + ], + "lines": [ + { + "bbox": [ + 313, + 283, + 365, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 365, + 336 + ], + "type": "image", + "image_path": "f521a29a4e3abb0e3ca32acdfd25bca7d99f504db9a28724ad8dbb48241de59f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 351, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 417 + ], + "type": "text", + "content": "Figure 5. At the top is the frame of our baseline. Lensless measurements first recover the image by reconstruction methods and then estimate human pose and shape by PyMAF. The bottom image compares the recovered image with the original image, and we can see that the quality of the recovered image has dropped dramatically." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 369, + 229, + 421, + 281 + ], + "blocks": [ + { + "bbox": [ + 369, + 229, + 421, + 281 + ], + "lines": [ + { + "bbox": [ + 369, + 229, + 421, + 281 + ], + "spans": [ + { + "bbox": [ + 369, + 229, + 421, + 281 + ], + "type": "image", + "image_path": "cf26471fafe17c92703833a47292d4f15ccf8a577406bdae22d1e104eb64d730.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 370, + 283, + 421, + 335 + ], + "blocks": [ + { + "bbox": [ + 370, + 283, + 421, + 335 + ], + "lines": [ + { + "bbox": [ + 370, + 283, + 421, + 335 + ], + "spans": [ + { + "bbox": [ + 370, + 283, + 421, + 335 + ], + "type": "image", + "image_path": "0f22bd1a366a8aaa59b5fb49bb95b53d26a98ade71f577d16c8fafa0b301db25.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 425, + 229, + 477, + 281 + ], + "blocks": [ + { + "bbox": [ + 425, + 229, + 477, + 281 + ], + "lines": [ + { + "bbox": [ + 425, + 229, + 477, + 281 + ], + "spans": [ + { + "bbox": [ + 425, + 229, + 477, + 281 + ], + "type": "image", + "image_path": "f7829f64eac7cc3ab9c70aab0829a6a5f121027469854164cecab124c105cd1f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 425, + 283, + 477, + 335 + ], + "blocks": [ + { + "bbox": [ + 425, + 283, + 477, + 335 + ], + "lines": [ + { + "bbox": [ + 425, + 283, + 477, + 335 + ], + "spans": [ + { + "bbox": [ + 425, + 283, + 477, + 335 + ], + "type": "image", + "image_path": "b9613918b8caf996d178a942928a939ffc9ec0aade123b8a09d9d40230d566ac.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 480, + 229, + 531, + 281 + ], + "blocks": [ + { + "bbox": [ + 480, + 229, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 480, + 229, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 480, + 229, + 531, + 281 + ], + "type": "image", + "image_path": "3e2d8f3b8ef91ef15fd15f21d7f56d7e6ddf9685713f5bc0b07eeaac1168f74b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 480, + 283, + 531, + 335 + ], + "blocks": [ + { + "bbox": [ + 480, + 283, + 531, + 335 + ], + "lines": [ + { + "bbox": [ + 480, + 283, + 531, + 335 + ], + "spans": [ + { + "bbox": [ + 480, + 283, + 531, + 335 + ], + "type": "image", + "image_path": "e66eea2bb41f95771ebeddc45b47e83b50fddb77e6ed99e46c5a8a705636c94e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 440, + 518, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 518, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 518, + 452 + ], + "type": "text", + "content": "4.3. 3D Human Pose and Shape Comparison." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 458, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 578 + ], + "type": "text", + "content": "Baseline Approach. We are the first work to perform human pose and shape estimation through lensless measurements. Given the absence of other valid methods for comparison, we designed a baseline approach that approaches this as a two-stage task. The baseline architecture is depicted in Fig. 5. In the first stage, we utilized the lensless image reconstruction method Rego et. al. [28], chosen for its wide applicability in recovering images acquired by lensless imaging systems. In the second stage, PyMAF is used to estimate human pose and shape." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 314, + 588, + 538, + 646 + ], + "blocks": [ + { + "bbox": [ + 314, + 588, + 538, + 646 + ], + "lines": [ + { + "bbox": [ + 314, + 588, + 538, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 588, + 538, + 646 + ], + "type": "table", + "html": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
baseline (PyMAF)257.07121.73277.18
baseline (PyMAF†)126.2881.37151.60
LPSNet119.2081.52134.74
", + "image_path": "ff9fc91372a3f861bf7d84d8fa2e94b53d834432cca3e799b264b7cd59556b07.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 654, + 545, + 687 + ], + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 687 + ], + "type": "text", + "content": "Table 1. Comparison with baseline. The baseline is a combination of the two-stage approach using existing methods. LPSNet outperforms the baseline on the LenslessHuman3.6M datasets." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Comparison with Baseline. We present the results of" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1476" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 532, + 308 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 532, + 308 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 532, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 532, + 308 + ], + "type": "image", + "image_path": "9fdeca129a39fd72df7420174d49a676f6146113e4c4666c9db1b7a60dc150a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 546, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 546, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 546, + 350 + ], + "type": "text", + "content": "Figure 6. Qualitative results of LPSNet on challenging LenslessHuman3.6M. The results on the left are from our LPSNet estimation, those in the center are from our baseline (PyMAF) output, and those on the right are from the fine-tuned baseline " + }, + { + "bbox": [ + 46, + 316, + 546, + 350 + ], + "type": "inline_equation", + "content": "(\\mathrm{PyMAF}^{\\dagger})" + }, + { + "bbox": [ + 46, + 316, + 546, + 350 + ], + "type": "text", + "content": " output. LPSNet can be seen to be better than the baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "content": "the quantitative comparative evaluation on the LenslessHuman3.6M dataset in Tab. 1. Notably, images recovered from lensless measurements exhibit lower quality. Directly comparing the baseline (PyMAF) with our approach is challenging and unfair. Therefore, we fine-tuned PyMAF on the train-LenslessHuman3.6M dataset, denoted as " + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "inline_equation", + "content": "\\mathrm{PyMAF}^{\\dagger}" + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "content": ". Referring to Tab. 1, we observe significant improvements in the indicators of baseline " + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "inline_equation", + "content": "(\\mathrm{PyMAF}^{\\dagger})" + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "content": ". Specifically, compared with the baseline " + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "inline_equation", + "content": "(\\mathrm{PyMAF}^{\\dagger})" + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "content": ", the MPJPE of LPSNet on the LenslessHuman3.6M dataset is reduced by " + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "inline_equation", + "content": "7.08~\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 370, + 287, + 490 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 288, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 288, + 575 + ], + "type": "text", + "content": "As shown in Tab. 1, our LPSNet achieved competitive results compared to the baseline approach " + }, + { + "bbox": [ + 46, + 491, + 288, + 575 + ], + "type": "inline_equation", + "content": "(\\mathrm{PyMAF}^{\\dagger})" + }, + { + "bbox": [ + 46, + 491, + 288, + 575 + ], + "type": "text", + "content": ". LPSNet demonstrates more significant improvements in the MPJPE and PVE metrics. However, we would argue that the PA-MPJPE metric may not fully reveal the performance of the mesh-image alignment, as it is calculated as MPJPE after rigid alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 648 + ], + "type": "text", + "content": "Baseline (PyMAF) has difficulty performing accurate pose estimation on low-quality recovered images. However, the outputs of our LPSNet exhibit significant improvements compared to the two baseline methods, especially for limb estimation. Qualitative comparison results are illustrated in Fig. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 142, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 142, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 142, + 671 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we conduct ablation studies on LenslessHuman3.6M under various settings to validate the effectiveness of the key components proposed in our method." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 368, + 547, + 442 + ], + "blocks": [ + { + "bbox": [ + 307, + 368, + 547, + 442 + ], + "lines": [ + { + "bbox": [ + 307, + 368, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 547, + 442 + ], + "type": "table", + "html": "
MethodMPJPE ↓PA-MPJPE↓PVE ↓
w/o MSFDcoder and DHAS142.1392.20161.07
w/o DHAS139.3492.56158.97
w/o DHAS (2D keypoint)123.3083.41138.67
w/o DHAS (IUV)129.7086.83146.7
LPSNet119.2081.52134.74
", + "image_path": "d33fadc890a82ec775392d69f16adfbb0e40ca65fcc593cf2e1a24b482c69dd0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 450, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 450, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 450, + 545, + 473 + ], + "type": "text", + "content": "Table 2. Ablation study on LPSNet. DHAS stands for Double-Head Auxiliary Supervision." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 496, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 556 + ], + "type": "text", + "content": "All ablation variants were trained and tested on LenslessHuman3.6M, which is derived from the Human3.6M dataset. The Human3.6M dataset includes ground-truth 3D labels and serves as a widely used benchmark in 3D human pose and shape estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Multi-Scale Lensless Feature Decoder. In LPSNet, the decoder primarily decodes the global information of lensless measurements encoded by the diffuser, which plays a crucial role in our end-to-end system. Our MSFDecoder is essential for efficient feature extraction from lensless measurements. We have adapted another variant of the decoder to verify this. Specifically, we simplified the global perception layer of the decoder to a series of convolutional layers and deconvolutional layers. We use this simpler decoder in the experiment(w/o MSFcoder and DHAS) instead of MSFDecoder. As shown in Tab. 2, comparing the result of experiments(w/o MSFcoder and DHAS) and experiment(w/o DHAS), we can see that MSFDecoder improves" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1477" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 516, + 235 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 516, + 235 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 516, + 235 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 516, + 235 + ], + "type": "image", + "image_path": "dfebda61820c178f7005e59c8a1ab75ea8e5f86091cf4d09398024c74520606b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 546, + 269 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 546, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 546, + 269 + ], + "type": "text", + "content": "Figure 7. Qualitative results of LPSNet on challenging LenslessHuman3.6M. We can see that the results of LPSNet are better, and the results are improved with the addition of the double-head assisted supervision mechanism." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 361 + ], + "type": "text", + "content": "the accuracy of the human pose and shape estimation significantly. Through the qualitative experiment in Fig. 7, we can also see that the human pose and shape estimation has a better alignment after using the MSFDecoder. Note that neither experiment(w/o MSFDcoder and DHAS) nor the experiment(w/o DHAS) used auxiliary supervision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 288, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 480 + ], + "type": "text", + "content": "Double-Head Auxiliary Supervision. Double-Head Auxiliary Supervision is mainly used to improve the accuracy of human limbs. For each auxiliary supervisory head, we conducted ablation experiments respectively. Referring to the result in Tab. 2, LPSNet has significant improvements in the MPJPE and the PVE metrics. Fig. 7 shows additional qualitative comparison results. We can see that the alignment of the human pose and shape is better with the addition of the Double-Head Auxiliary Supervision mechanism." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 489, + 176, + 567 + ], + "blocks": [ + { + "bbox": [ + 50, + 489, + 176, + 567 + ], + "lines": [ + { + "bbox": [ + 50, + 489, + 176, + 567 + ], + "spans": [ + { + "bbox": [ + 50, + 489, + 176, + 567 + ], + "type": "image", + "image_path": "c3ad17b93f94cb2ac472607d0671de20306ff76a0ae25718cf88b2ebaec0eaaa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 571, + 163, + 581 + ], + "lines": [ + { + "bbox": [ + 63, + 571, + 163, + 581 + ], + "spans": [ + { + "bbox": [ + 63, + 571, + 163, + 581 + ], + "type": "text", + "content": "(a) Result on more datasets" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 492, + 285, + 567 + ], + "blocks": [ + { + "bbox": [ + 179, + 492, + 285, + 567 + ], + "lines": [ + { + "bbox": [ + 179, + 492, + 285, + 567 + ], + "spans": [ + { + "bbox": [ + 179, + 492, + 285, + 567 + ], + "type": "image", + "image_path": "88a8798d1589f4b5ff5a56bdf252bba4306038000f0154014599a59b5e431424.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 571, + 277, + 581 + ], + "lines": [ + { + "bbox": [ + 187, + 571, + 277, + 581 + ], + "spans": [ + { + "bbox": [ + 187, + 571, + 277, + 581 + ], + "type": "text", + "content": "(b) Result on real scenes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 590, + 287, + 624 + ], + "lines": [ + { + "bbox": [ + 46, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 287, + 624 + ], + "type": "text", + "content": "Figure 8. Results on more datasets and real scenes. From left to right for each set of figures: lensless measurement, alignment of results with RGB images, and 3D results shown in different views." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 261, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 261, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 261, + 647 + ], + "type": "text", + "content": "4.5. Results on more Datasets and Real Scene" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "Fig. 8 (a) shows the qualitative results on various datasets after using the mixed datasets for training. Specifically, we collected the MPII, COCO, 3DPW, MIP-INF-3DHP, and LSP datasets through our lensless imaging system and mixed them for training and testing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 361 + ], + "type": "text", + "content": "In addition, we also capture the real scenes of two different people, in different backgrounds and different light intensities using our lensless imaging system. Fig. 8 (b) illustrates the experimental results of our method on these real scenes, indicating its suitability for real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 375, + 457, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 457, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 457, + 387 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 515 + ], + "type": "text", + "content": "Conclusion. In this paper, we propose the first end-to-end framework to recover 3D human poses and shapes from lensless measurements to the best of our knowledge. Specifically, we design a multi-scale lensless feature decoder, which can effectively decode the information produced by the lensless imaging system. We also propose a double auxiliary supervision mechanism to improve the accuracy of human limb end estimation. Experimental results show that our method can achieve end-to-end human pose and shape estimation through lensless measurements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 516, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 545, + 635 + ], + "type": "text", + "content": "Limitations and Future Work. Our work is an initial stride step in this direction, however, the results are still not as robust as those achieved by traditional methods for estimating the human body from RGB images. The current scarcity of lensless datasets hampers our ability to pre-train the MSFDecoder using a large dataset like ImageNet [9], thereby limiting the generalizability of our approach. We will continue this work in the future, focusing on generating more extensive lensless datasets and progressing towards practical applications." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 635, + 546, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 546, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 546, + 703 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Key R&D Program of China (2023YFC3082100), the National Natural Science Foundation of China (62122058 and 62171317), and the Science Fund for Distinguished Young Scholars of Tianjin (No. 22JCJQJC00040)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1478" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jesse K Adams, Vivek Boominathan, Benjamin W Avants, Daniel G Vercosa, Fan Ye, Richard G Baraniuk, Jacob T Robinson, and Ashok Veeraraghavan. Single-Frame 3D Fluorescence Microscopy with Ultraminiature Lensless FlatScope. Sci Adv., 3(12):e1701548, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 191 + ], + "type": "text", + "content": "[2] Nick Antipa, Grace Kuo, Reinhard Heckel, Ben Mildenhall, Emrah Bostan, Ren Ng, and Laura Waller. DiffuserCam: Lensless Single-exposure 3D Imaging. Optica, 5(1):1-9, 2018. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 236 + ], + "type": "text", + "content": "[3] Anurag Arnab, Carl Doersch, and Andrew Zisserman. Exploiting Temporal Context for 3D Human Pose Estimation in the ild. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3395-3404, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 292 + ], + "type": "text", + "content": "[4] M Salman Asif, Ali Ayremlou, Aswin Sankaranarayanan, Ashok Veeraraghavan, and Richard G Baraniuk. Flatcam: Thin, Lensless Cameras using Coded Aperture and Computation. IEEE Trans Comput Imaging., 3(3):384-397, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 287, + 350 + ], + "type": "text", + "content": "[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. In *Eur. Conf. Comput. Vis. Springer International Publishing*, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 407 + ], + "type": "text", + "content": "[6] Vivek Boominathan, Jesse K Adams, M Salman Asif, Benjamin W Avants, Jacob T Robinson, Richard G Baraniuk, Aswin C Sankaranarayanan, and Ashok Veeraraghavan. Lensless Imaging: A Computational Renaissance. IEEE Signal Process Mag., 33(5):23-35, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 287, + 451 + ], + "type": "text", + "content": "[7] Vivek Boominathan, Jesse K Adams, Jacob T Robinson, and Ashok Veeraraghavan. Phlatcam: Designed Phase-Mask Based Thin Lensless Camera. IEEE Trans. Pattern Anal. Mach. Intell., 42(7):1618-1629, 2020. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 496 + ], + "type": "text", + "content": "[8] Junhyeong Cho, Kim Youwang, and Tae-Hyun Oh. Cross-attention of disentangled modalities for 3d human mesh recovery with transformers. In Eur. Conf. Comput. Vis., pages 342–359. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 542 + ], + "type": "text", + "content": "[9] J. Deng, W. Dong, R. Socher, L. Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE Conf. Comput. Vis. Pattern Recog. Workshops (CVPR Workshops). 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[10] Georgios Georgakis, Ren Li, Srikrishna Karanam, Terrence Chen, Jana Košecka, and Ziyan Wu. Hierarchical Kinematic Human Mesh Recovery. In Eur. Conf. Comput. Vis., pages 768-784. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[11] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4D: Reconstructing and Tracking Humans with Transformers. In Int. Conf. Comput. Vis., pages 14783-14794, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[12] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense Human Pose Estimation in the Wild. In Int. Conf. Comput. Vis., pages 7297-7306, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Exemplar Fine-Tuning for 3D Human Model Fitting Towards In-the-Wild 3D Human Pose Estimation. In Int. Conf. 3D. Vis., pages 42-52. IEEE, 2021. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[14] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end Recovery of Human Shape and Pose. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 1, 3, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[16] Muhammed Kocabas, Chun-Hao P Huang, Otmar Hilliges, and Michael J Black. PARE: Part Attention Regressor for 3D Human Body Estimation. In Int. Conf. Comput. Vis., pages 11127-11137, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[17] Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to Reconstruct 3D Human Pose and Shape Via Model-fitting in the Loop. In Int. Conf. Comput. Vis., pages 2252-2261, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 285 + ], + "type": "text", + "content": "[18] Nikos Kolotouros, Georgios Pavlakos, and Kostas Daniilidis. Convolutional Mesh Regression for Single-Image Human Shape Reconstruction. In IEEE Conf. Comput. Vis. Pattern Recog., pages 4501–4510, 2019. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 331 + ], + "type": "text", + "content": "[19] Grace Kuo, Nick Antipa, Ren Ng, and Laura Waller. DiffuserCam: Diffuser-Based Lensless Cameras. In Computational Optical Sensing and Imaging, pages CTu3B-2. Optica Publishing Group, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 385 + ], + "type": "text", + "content": "[20] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the People: Closing the Loop Between 3D and 2D Human Representations. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[21] Vincent Leroy, Philippe Weinzaepfel, Romain Brégier, Hadrien Combaluzier, and Grégory Rogez. SMPLy Benchmarking 3D Human Pose Estimation in the wild. In Int. Conf. 3D. Vis., pages 301-310. IEEE, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[22] Yanjie Li, Sen Yang, Peidong Liu, Shoukui Zhang, Yunxiao Wang, Zhicheng Wang, Wankou Yang, and Shu-Tao Xia. SimCC: A Simple Coordinate Classification Perspective for Human Pose Estimation. In Eur. Conf. Comput. Vis., pages 89-106. Springer, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 532 + ], + "type": "text", + "content": "[23] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1954-1963, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 566 + ], + "type": "text", + "content": "[24] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh Graphormer. In Int. Conf. Comput. Vis., pages 12939-12948, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 600 + ], + "type": "text", + "content": "[25] Aydogan Ozcan and Euan McLeod. Lensless Imaging and Sensing. Annual review of biomedical engineering, 18:77-102, 2016. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 656 + ], + "type": "text", + "content": "[26] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 690 + ], + "type": "text", + "content": "[27] Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Human Mesh Recovery from Multiple Shots. In IEEE Conf. Comput. Vis. Pattern Recog., pages 1485-1495, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] Joshua D Rego, Karthik Kulkarni, and Suren Jayasuriya. Robust Lensless Image Reconstruction Via PSF Estimation. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1479" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 443 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 403-412, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[29] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. HuMoR: 3D Human Motion Model for Robust Pose Estimation. In Int. Conf. Comput. Vis., 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[30] David G Stork and Patrick R Gill. Lensless Ultra-miniature CMOS Computational Imagers and Sensors. Proc. Sensorcomm, pages 186-190, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[31] David G Stork and Patrick R Gill. Optical, mathematical, and computational foundations of lensless ultra-miniature diffractive imagers and sensors. International Journal on Advances in Systems and Measurements, 7(3):4, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[32] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep High-resolution Representation Learning for Human Pose Estimation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 5693-5703, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[33] Garvita Tiwari, Dimitrije Antic, Jan Eric Lenssen, Nikolaos Sarafianos, Tony Tung, and Gerard Pons-Moll. PoseNDF: Modeling Human Pose Manifolds with Neural Distance Fields. In Eur. Conf. Comput. Vis., 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 352 + ], + "type": "text", + "content": "[34] Hongwen Zhang, Jie Cao, Guo Lu, Wanli Ouyang, and Zhenan Sun. Learning 3D Human Shape and Pose from Dense Body Parts. IEEE Trans. Pattern Anal. Mach. Intell., 44(5):2610-2627, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[35] Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. In Int. Conf. Comput. Vis., pages 11446-11456, 2021. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 443 + ], + "type": "text", + "content": "[36] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3D Human Dynamics from Video. In Int. Conf. Comput. Vis., pages 7114-7123, 2019. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1480" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_content_list.json b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6948e9b469355ec135b72e5dba22aa7fc5ff176 --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_content_list.json @@ -0,0 +1,1464 @@ +[ + { + "type": "text", + "text": "LQMFormer: Language-aware Query Mask Transformer for Referring Image Segmentation", + "text_level": 1, + "bbox": [ + 243, + 130, + 725, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nisarg A. Shah Vibashan VS Vishal M. Patel Johns Hopkins University", + "bbox": [ + 264, + 203, + 700, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "snisarg812@gmail.com, {vvishnu2, vpatel36}@jhu.edu", + "bbox": [ + 261, + 241, + 700, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Image Segmentation (RIS) aims to segment objects from an image based on a language description. Recent advancements have introduced transformer-based methods that leverage cross-modal dependencies, significantly enhancing performance in referring segmentation tasks. These methods are designed such that each query predicts different masks. However, RIS inherently requires a single-mask prediction, leading to a phenomenon known as Query Collapse, where all queries yield the same mask prediction. This reduces the generalization capability of the RIS model for complex or novel scenarios. To address this issue, we propose a Multi-modal Query Feature Fusion technique, characterized by two innovative designs: (1) Gaussian enhanced Multi-Modal Fusion, a novel visual grounding mechanism that enhances overall representation by extracting rich local visual information and global visual-linguistic relationships, and (2) A Dynamic Query Module that produces a diverse set of queries through a scoring network where the network selectively focuses on queries for objects referred to in the language description. Moreover, we show that including an auxiliary loss to increase the distance between mask representations of different queries further enhances performance and mitigates query collapse. Extensive experiments conducted on four benchmark datasets validate the effectiveness of our framework.", + "bbox": [ + 75, + 325, + 473, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 737, + 209, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring Image Segmentation (RIS) is a challenging multimodal task aimed at segmenting specific objects in an image based on a textual description. This description often includes details about the object's actions, category, color, or position, as noted by [8, 21]. Compared to traditional semantic and instance segmentation, RIS requires a precise understanding of object locations and involves a comprehensive modeling of the visual-linguistic relationships within the global context. Additionally, RIS demands the extrac", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32cc2895f818ab3817838db895ba94bc54ddf6b6a850a4f11431606300f57670.jpg", + "image_caption": [ + "(a) ReLA" + ], + "image_footnote": [], + "bbox": [ + 537, + 285, + 696, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4154ddd81b69b8383091a3da1398816b3faecc99aae154e1cdc3c5c858ec3c7.jpg", + "image_caption": [ + "(b) LQMFormer", + "Figure 1. [Top] Heatmap visualization of cosine similarity between different query mask predictions of ReLA and LQMFormer, where the yellow region indicates a similarity near 1, and the dark blue region signifies a similarity of 0. [Bottom] Sample query mask prediction visualizations for ReLA and LQMFormer." + ], + "image_footnote": [], + "bbox": [ + 700, + 287, + 862, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion of high-quality visual features at a local level. The potential applications of RIS are extensive, particularly in language-driven human-computer interaction domains.", + "bbox": [ + 496, + 566, + 892, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional RIS methods have employed linear fusion approaches and Fully Convolutional Networks (FCNs) for feature learning in RIS tasks [21, 33]. However, the recent advancement of attention mechanisms has shifted the focus towards extracting richer visual-language representation, with recent techniques demonstrating the effectiveness of Transformers. These models excel at modelling long-distance dependencies, making them suitable for cross-modal fusion [45, 55]. Further, Vision Transformer (ViT) methods such as VLT [13], EFN [16], and LAVT [54], which are based on [14], have shown significant improvements in RIS performance. Their ability to capture the nuances of cross-modal dependencies is crucial, and through a series of attention mechanisms, these models efficiently process and integrate both visual and language inputs for precise object segmentation.", + "bbox": [ + 496, + 612, + 893, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite their advantages, transformer-based approaches exhibit shortcomings in RIS mask prediction, with a notable problem being query collapse. This phenomenon occurs", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "when different queries within the transformer, intended to identify distinct attributes or portions of objects, converge on the same mask prediction. This is particularly problematic in referring image segmentation, which inherently requires a prediction of a single mask, where all queries are collectively trained to predict the same mask. We experimentally observed this phenomenon, as illustrated in Fig. 1. Here, we visualize the heatmap of the cosine similarity between mask predictions from ReLA [32] and our method. In ReLA, most mask predictions are almost identical, leading to significantly overlapping mask predictions, whereas our method demonstrates a diverse set of query embeddings and mask predictions. Hence, prior transformer-based methods train mask predictions for all queries to associate with a single, specific ground truth mask. This training approach can lead to query collapse, where the model is not penalized for producing the same prediction across different queries, thereby affecting its capability to identify diverse attributes of the image. Consequently, this limitation restricts the variety of mask predictions, diminishing the model's ability to generalize effectively in complex scenarios.", + "bbox": [ + 76, + 90, + 472, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose LQMFormer, which comprises two key components. Firstly, we introduce a Gaussian Enhanced Multi-Modal Fusion (GMMF) module, aimed at enhancing the visual grounding mechanism. This module is designed to extract fine-grained local visual details while simultaneously enhancing global visual-language relationships. The fusion of these modalities allows the model to construct a more comprehensive understanding of the scene, both globally and locally, thereby improving visual grounding. Secondly, in our Dynamic Query Module (DQM), we generate a diverse set of language-dependent queries through Language-Query Cross Attention (LQCA) and employ a scoring network to prioritize queries relevant to the referring expression. With an improved visual-grounded representation facilitated by GMMF and selective query enhancement via DQM, our model effectively addresses query collapse. Moreover, by incorporating an auxiliary loss that increases the representational distance between mask predictions, our method efficiently differentiates between queries, mitigating the issue of query collapse. Our extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our framework. Our contributions can be summarized as follows:", + "bbox": [ + 76, + 412, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Gaussian enhanced Multi-Modal Fusion, which innovates the integration of local and global cross-modal information for more accurate visual grounding.", + "- We develop an effective Dynamic Query Module (DQM) that not only generates a diverse query set but also employs a scoring network to selectively focus on queries based on given language expressions with respect to the decoder's references.", + "- To mitigate query collapse, we propose an auxiliary regu" + ], + "bbox": [ + 76, + 763, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "larization loss function that increases the representational distance between queries, significantly improving mask differentiation and preventing Query Collapse.", + "bbox": [ + 511, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct comprehensive validation of the proposed framework with extensive testing on four benchmark datasets, demonstrating improvements in referring image segmentation performance.", + "bbox": [ + 500, + 137, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 210, + 648, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Referring Image Segmentation", + "text_level": 1, + "bbox": [ + 500, + 237, + 772, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works [29, 31, 45, 57] employed Convolutional Neural Networks (CNNs) [5, 21, 22, 31] and Recurrent Neural Networks (RNNs) [6, 31] to extract vision and language features. These features were then fused via sequential concatenation and convolution operations to predict the segmentation mask. In contrast, [58] proposed a two-stage network that initially uses Mask R-CNN[19] to predict categorical masks, followed by a selection of the relevant masks based on language descriptions. However, this approach exhibits limited capacity to capture the intricate relations between language and visual content. The advent of the Transformer [11, 14, 42, 46] in the vision community has led to the widespread adoption of Transformer-based architectures for extracting both visual and textual features [11, 24, 34]. VLT [13] uses cross-attention to produce query vectors from multi-modal features, which are then used to query images in the transformer decoder. Similarly, LAVT [54] shows early cross-modal fusion of features improves alignment. CRIS [49] adopts Transformer blocks to leverage the pretrained CLIP model's [44] robust image-text alignment capabilities. Additionally, GRES [32] extends cross-attention in the Transformer decoder to explicitly model visual-language dependencies. However, these all transformer-based approaches require matching queries to corresponding mask instances. In the context of referring expression segmentation, where typically only a single mask is available, this can lead to all queries converging to a single mask, a phenomenon known as query collapse. Motivated by this, in our method, we try to produce a diverse set of query prediction by overall improving visual-grounding and conditioning with respect to language expression.", + "bbox": [ + 496, + 260, + 893, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision and Language Representation Learning", + "text_level": 1, + "bbox": [ + 500, + 739, + 890, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision and Language Representation Learning focuses on understanding the semantics and alignment between vision and language for multimodal reasoning tasks [36, 47, 59]. This field has seen substantial progress in applications like visual question answering [2], Image captioning [48], Image-text retrieval [4], zero-shot classification [44], and Image segmentation [23]. Contrastive pre-training strategies [28, 44] using large-scale datasets effectively project multiple modalities into a single embedding space. In contrast, as discussed,", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12904", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other methods [32, 54] create cross-modal interaction layers to fuse and comprehend multimodal features. The recent adoption of deep learning techniques in the frequency domain [9, 17, 39, 40] has gained attention for their global interaction capabilities. Specifically, [40] performs spectral-guided enhancement of the multi-scale vision-language features after feature extraction stage for Video Referring Segmentation. Drawing from these advancements, our approach integrates gaussian guidance within vision-language representation learning to facilitate enhanced local and global multi-modal interactions.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 268, + 174, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ and a language expression $L = \\{w_{i}\\}_{i=1}^{N}$ with $N$ words, our model predicts a pixel-wise mask $M$ , which delineating the referred object.", + "bbox": [ + 75, + 292, + 470, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 345, + 187, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview diagram of our approach is shown in Figure 2. First, the language expression is encoded to extract high-dimensional language features $F_{l} \\in \\mathbb{R}^{N_{t} \\times C}$ , where $C, N_{t}$ indicates the number of channels, and words in the language expression, respectively. The Gaussian-enhanced Multi-Modal Fusion (GMMF) module then extracts language-grounded visual features $F_{\\mathrm{vl}}$ . Following [7], our approach incorporates a multi-scale strategy to exploit higher-resolution feature maps in the transformer decoder [7, 61]. We input language-aware queries $Q_{b}^{\\prime}$ along with multi-scale outputs from the pixel decoder $F_{i}$ , where $i \\in \\left\\{\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}$ of the original image. The Dynamic Query Module processes language features $F_{t}$ to predict language-aware queries $Q_{b}^{\\prime}$ . The updating of queries $Q_{b}^{\\prime}$ sequentially utilizes resolutions $\\frac{1}{8}, \\frac{1}{16}$ , and $\\frac{1}{32}$ , applying a cycle of masked cross-attention (CA), self-attention (SA), and feed-forward network (FFN) operations $L$ times within the decoder. Further, we map the final query outputs to an 'object' or 'no-object' two-dimensional space to enable 'no-object' predictions [32]. The final predicted mask is obtained at full resolution as the original image by decoding the pixel features $F_{\\frac{1}{4}}$ with the einsum operation between $Q$ and $F_{\\frac{1}{4}}$ , followed by upsampling operation.", + "bbox": [ + 75, + 369, + 472, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Gaussian-enhanced Multi-Modal Fusion", + "text_level": 1, + "bbox": [ + 76, + 710, + 423, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After extracting language features $F_{l}$ , we combine them with visual features using a four-stage hierarchical Swin Transformer [34] to obtain joint visual-language embedding. Each stage comprises Transformer layers $(\\tau_{i})$ , multi-modal feature fusion modules (GMMF), and residual gating $(\\psi_{i})$ . At every stage, as Transformer layers $(\\tau_{i})$ enhance previous visual features $(F_{v})$ , these features $(F_{v})$ combine with language features $(F_{l})$ via multi-modal modules (GMMF) to create multi-modal features $(F_{vl})$ . Finally, we use gating units $(\\psi_{vl})$ to weight and combine $F_{vl}$ with $F_{v}$ , yielding enhanced visual features with linguistic information. Residual gating", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "also works well with pre-trained (only vision) transformer weights and allows simultaneous feature modulation with language without having to re-train from scratch.", + "bbox": [ + 496, + 90, + 892, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Gaussian Enhanced Multi-Modal Fusion Module in Visual Backbone", + "text_level": 1, + "bbox": [ + 498, + 154, + 890, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discrete Fourier transform (DFT) is extensively used to analyze the frequency representation of images. By transforming the signal into the frequency domain, characterized by global statistical properties, it plays a crucial role in various computer vision tasks [26, 51, 56]. Modulating signals on a point-wise basis in the Fourier domain alters the representation of inputs in the spatial domain [3, 40]. Furthermore, in the Fourier domain, low-frequency components usually correspond to the overall semantic content of the image, consistent with prior research [40, 52, 53, 56]. The Fourier transform is capable of disentangling global semantic information and structural coherence, with the semantic content predominantly found in the amplitude component [18, 27, 60]. These insights can be leveraged to create modules centered on Fourier-based methods, facilitating efficient and vital global interactions in multimodal understanding and improving visual grounding.", + "bbox": [ + 496, + 191, + 893, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To inculcate the above observation, as illustrated in Fig. 3, following [40], the Gaussian enhanced Multi-Modal Fusion Module (GMMF) is proposed as a key component in our model. It employs an enhancement block after cross-attention operations to combine Gaussian-enhanced vision and language features. Given an input vision feature $F_{vi} \\in \\mathbb{R}^{N \\times H \\times W \\times C_{vi}}$ from stage $i$ and a language feature $F_{l} \\in \\mathbb{R}^{N \\times T \\times C_{l}}$ , for cross-attention block, we first transform visual and language features into a common dimension $C_{i}$ using separate $1 \\times 1$ convolution layers. Subsequently, transform vision and language features are passed to an MHCA (Multi-Head Cross-Attention) module, where language features act as queries and vision features as keys and values. This operation results in the creation of a language-aware vision-shaped feature $F_{vl}'$ . We then transform the visual-language features using the Gaussian Enhancement Block, as shown in Fig. 3, formulated as follows:", + "bbox": [ + 496, + 449, + 895, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {v l} = F _ {\\mathrm {F F T}} ^ {- 1} \\left(\\operatorname {C o n v} \\left(F R \\left(\\phi \\left(F _ {v l} ^ {\\prime}, \\beta\\right) * \\mathcal {A} \\left(F _ {v l} ^ {\\prime}\\right) \\right. \\right. \\right. \\\\ , \\left. \\left. \\mathcal {P} \\left(F _ {v l} ^ {\\prime}\\right)\\right)\\right) + F _ {v l} ^ {\\prime} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 712, + 841, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathcal{A}(F_{vl}^{\\prime})$ and $\\mathcal{P}(F_{vl}^{\\prime})$ denote the amplitude and phase components of $F_{vl}^{\\prime}$ 's Fast Fourier Transform(FFT) [15, 43], $F_{\\mathrm{FFT}}(F_{vl}^{\\prime})$ . The * represents low-pass filtering on $\\mathcal{A}(F_{vl}^{\\prime})$ using adaptive Gaussian filters $\\phi (F_{vl}^{\\prime},\\beta)$ , matching $F_{vl}^{\\prime}$ 's spatial dimensions and adjusting to input data via bandwidth $\\beta$ [40]. $FR$ and $Conv$ signify feature reconstruction from amplitude and phase, and convolution operations, respectively.", + "bbox": [ + 496, + 773, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12905", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c5c2ea4b61ac346f4179f924ed36da4e9a8ba10514737dbc809b272401ede617.jpg", + "image_caption": [ + "Figure 2. Overview of the LQMFormer model architecture. The model processes an image and a linguistic expression, extracting vision-language features $F_{vl}$ through a Gaussian-enhanced Multi-modal Fusion module and language features $F_{L}$ . The multi-scale vision-language features $F_{vl}$ are then processed by into a pixel decoder. Concurrently, the Language Query Cross-Attention Module (LQCA) refines the Language-enhanced Query bank $Q_{b}^{\\prime}$ with $F_{L}$ . A Scoring Network assesses the relevance of each query in $Q_{b}^{\\prime}$ by predicting soft-masks $(M_{b})$ , which then modulate the queries fed into the transformer decoder. Finally, the model employs a global pooling layer and a Multi-Layer Perceptron (MLP) to predict the binary segmentation mask and determine the presence or absence of the referred object." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 893, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c84351b775a6b875a199450fbcc9f174e9c9f024bd3758aabcdd1b397e5ebf6.jpg", + "image_caption": [ + "Figure 3. Architecture overview of the Gaussian Enhancement Module" + ], + "image_footnote": [], + "bbox": [ + 196, + 441, + 352, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once $F_{vl}$ is obtained, similar to [54], we combine the output from GMMF, $F_{vl}$ , with that from the Transformer layers, $F_{v}$ . This process employs a gating mechanism, $\\psi_{vl}$ , which learns a set of element-wise weight maps from $F_{vl}$ to finely tune the scale of each element within $F_{vl}$ . Subsequent to this adjustment, a residual combination of $F_{v}$ and $F_{vl} \\cdot \\psi_{vl}$ is performed. The resultant output is then fed back into the vision backbone for further computation.", + "bbox": [ + 75, + 700, + 470, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic Query Module", + "text_level": 1, + "bbox": [ + 76, + 830, + 299, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Dynamic Query Module (DQM) takes the language feature $F_{l}$ and a Query bank $Q_{b}$ as inputs, the latter containing $N_{qb}$ learnable queries. Figure 4(b) shows the initial", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step, where the attention between the language feature $F_{l}$ and the $N_{qb}$ query embeddings $Q_{b} \\in \\mathbb{R}^{N_{qb} \\times C}$ is computed, resulting in $N_{qb}$ attention maps:", + "bbox": [ + 498, + 443, + 893, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {b i} = \\operatorname {s o f t m a x} \\left(Q _ {b} \\sigma \\left(F _ {l} W _ {i k}\\right) ^ {T}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 503, + 892, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{ik}$ is a $C\\times C$ matrix of learnable parameters, and $\\sigma$ denotes the GeLU function [20]. The resulting $A_{bi}\\in \\mathbb{R}^{N_{qb}\\times L}$ provides each query with a $1\\times L$ attention map, indicating its important linguistic relation in the language features. Following this, the language-supported query bank features are derived using these attention maps: $Q_b^{\\prime} = A_{bi}\\sigma (F_lW_{iv})^T$ , where $W_{iv}$ is another $C\\times C$ matrix of learnable parameters.", + "bbox": [ + 498, + 532, + 893, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The variability in language expressions used to describe objects in images requires an approach that can adapt to different descriptions. This is because standard methods, like those seen in the initial transformer models (like in vanilla transformer [14]), struggle to capture the nuances in descriptions that include color, size, shape, and position. To address this, a larger set of adaptable queries is required. These queries are fine-tuned based on the input language, resulting in a refined set of queries, denoted as $Q_b'$ . Each query in this refined set is then assessed with a scoring network, which outputs a soft-mask $M_b$ . This soft-mask represents the relevance of each query. The final set of queries, $Q_{\\mathrm{final}}$ , is obtained by combining $M_b$ with $Q_b'$ through a dot product, as shown in the equation below:", + "bbox": [ + 496, + 657, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\text {f i n a l}} = M _ {b} ^ {\\prime} \\odot Q _ {b} ^ {\\prime} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 883, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12906", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/498463b200decfac3a2c4470079c807eead59ff69518d3b971682bddfcbde74c.jpg", + "image_caption": [ + "Figure 4. Architecture overview of the (A.) Dynamic Query Module and (B.) LQCA Block" + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 630, + 300 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/46078c28fb1381519120e8b9102c5b752bbb0245513843bb8d6012d442baf33c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 89, + 885, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Scoring Network", + "text_level": 1, + "bbox": [ + 76, + 369, + 250, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The scoring network in LQMFormer takes the Language-enhanced Query bank $Q_{b}^{\\prime}$ as input and processes it through a sequence of linear layers. This network consists of four linear layers, with Layer Normalization [1] applied before the first layer. Each subsequent layer, except the last sigmoid, incorporates GELU [20] activation. The final output is a 1-dimensional logit $M_{b} \\in \\mathbb{R}^{N_{qb} \\times 1}$ , serving as a soft-mask to modulate $Q_{b}^{\\prime}$ .", + "bbox": [ + 75, + 392, + 470, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the scoring network for predicting a relevant soft-mask $M_{b}$ effectively for the transformer decoder, we utilize its cross-attention map. This map is essential in identifying the subset of Queries most relevant to the referring expression and visual features, as highlighted during training. We sum up the attention maps from each decoder layer using Bilinear Interpolation (BI), thereby projecting attention $A_{qf} \\in \\mathbb{R}^{N_{qb} \\times HW}$ directed towards these Queries. The process is mathematically described as follows:", + "bbox": [ + 75, + 513, + 473, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {q f} = \\sum_ {l = 1} ^ {L} \\sum_ {r = 1} ^ {3} \\mathrm {B I} \\left(\\mathrm {A} _ {l, r}, H W\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 660, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{A}_{l,r} \\in \\mathbb{R}^{N_{qb} \\times hw}$ represents the attention map from the $l^{th}$ layer at the $r^{th}$ resolution, with $hw$ being the original dimension and $HW$ the desired output dimension. After aligning each attention map to the common resolution $HW$ , we then sum across this dimension to generate a likelihood map $LM$ , as indicated in the subsequent equation:", + "bbox": [ + 75, + 705, + 470, + 799 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL M = \\sum_ {i = 1} ^ {H W} A _ {q f} [:, i ] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 808, + 468, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This likelihood map $\\mathrm{LM} \\in \\mathbb{R}^{N_{qb} \\times 1}$ , now captures the concentrated attention across all Queries, serves as a condition in refining the predictions of the scoring network. The", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "scoring network is trained to predict the likelihood of each Query being included in the top- $\\rho\\%$ of the most referenced tokens. This process involves binarizing the likelihood map $LM$ to retain only the top- $\\rho\\%$ Queries, creating a binarized version $LM_{\\mathrm{bin}}$ . The training objective is then formulated as a Binary Cross-Entropy (BCE) loss between the predicted soft-mask $M_b$ and the binarized likelihood map $LM_{\\mathrm{bin}}$ . The BCE loss function is defined as:", + "bbox": [ + 498, + 369, + 893, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {B C E} \\operatorname {L o s s} = - \\frac {1}{N _ {q b}} \\sum_ {i = 1} ^ {N _ {q b}} \\left[ L M _ {\\mathrm {b i n}, i} \\log \\left(M _ {b, i}\\right) \\right. \\\\ + \\left(1 - L M _ {\\text {b i n}, i}\\right) \\log \\left(1 - M _ {b, i}\\right) ] \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 510, + 892, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $N_{qb}$ represents the number of Queries in $Q_b^{\\prime}$ , and $i$ indexes each Query. This loss function effectively trains the scoring network to align its predictions with the most salient Queries as determined by the decoder's cross-attention, thereby refining the model's accuracy in identifying and segmenting objects as described by the referring expressions.", + "bbox": [ + 498, + 595, + 893, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Query-Mask Margin Loss", + "text_level": 1, + "bbox": [ + 500, + 696, + 736, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Query-Mask Margin Loss is designed to maintain a margin of separability between different query feature representations, an essential aspect for tasks with variable outputs. The loss function operates by initially computing pairwise differences and distances between mask embeddings within a batch. This computation yields a tensor of shape $B \\times N \\times N$ , where $B$ represents the batch size, and $N$ is the number of queries. Subsequently, these distances for each embedding are sorted, with a particular focus on the second smallest value, which signifies the distance to the nearest neighbor. The margin loss is defined through the max function, ensuring a minimum distance of 1.0 between embeddings. The", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12907", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mathematical formulation of QM-loss is as follows:", + "bbox": [ + 76, + 90, + 421, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {Q M - l o s s} = \\max (0, 1. 0 - \\operatorname {d i s t} \\left(M _ {i}, M _ {j}\\right))\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 117, + 411, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\text{dist}(M_i, M_j)$ denotes the calculation of pairwise distances between the mask embeddings, represented as $M$ .", + "bbox": [ + 76, + 143, + 470, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Discussion", + "text_level": 1, + "bbox": [ + 76, + 188, + 339, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 212, + 397, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: Our experiments are conducted on four primary benchmark datasets in Referring Image Segmentation: RefCOCO [57], RefCOCO+ [57], RefCOCOg [38], and GRES [32]. These datasets derive their images from MSCOCO [30]. The dataset details are as follows: RefCOCO (19,994 images, 50,000 objects, 142,209 expressions), RefCOCO+ (19,992 images, 49,856 objects, 141,564 expressions), RefCOCOg (26,711 images, 54,822 objects, 85,474 expressions), and GRES (19,994 images, 60,287 objects, 278,232 expressions). RefCOCO mainly focuses on expressions that specify object locations, while RefCOCO+ prioritizes object descriptions. RefCOCOg presents a higher challenge due to longer and more complex expressions (averaging 8.4 words compared to 3.5 in others). GRES broadens the scope of RefCOCO by including expressions that refer to multiple objects or no objects, thereby enhancing the problem's generalizability. We use both UMD split [41] and Google split [38] for RefCOCOg dataset.", + "bbox": [ + 75, + 236, + 472, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For single-target object segmentation, we report our results using two kinds of metrics [13, 37, 49, 54]: overall IoU (oIoU) and Precision@X (P@X). oIoU divides the total intersection pixels by the total union pixels across all test images. Precision@X calculates the percentage of testing samples of which the model prediction has an IoU score higher than the threshold value X. To extend our evaluation to non-target and multi-target object segmentation, following [32], we include additional metrics: Sensitivity (N-acc.), Specificity (T-acc.), and Generalized IoU (gIoU). N-acc. evaluates the model's performance on identifying non-target samples, while T-acc. measures the impact of generalization on non-target samples on the performance of target samples. We opt for gIoU over oIoU due to the latter's bias towards larger objects. gIoU computes the mean IoU per image for all samples. In cases with no-target samples, IoU values are assigned as 1 for true positives and 0 for false negatives, ensuring an equitable metric for all object sizes.", + "bbox": [ + 75, + 508, + 472, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our model uses Swin Transformer-B [34] as the visual encoder and BERT [11] as the language decoder. The transformer layer in Swin-B is initialized with classification weights pre-trained on ImageNet22K [10]. We use BERT implementation from the HuggingFace Transformer library [50]. It comprises 12 layers with a hidden size of 768, initialized using official", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pre-trained weights. We use our Transformer decoder proposed in Section 3.1 with $\\mathrm{L} = 3$ (i.e., 9 layers total) and 100 queries by default. The AdamW optimizer [35] is employed for optimization with an initial learning rate of 0.00001. Our model is trained over 100K iterations with a batch size of 48. Our image resizing is standardized to $480 \\times 480$ pixels.", + "bbox": [ + 496, + 90, + 893, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the state of art methods", + "text_level": 1, + "bbox": [ + 498, + 195, + 856, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Referring Image Segmentation Table 1 presents the results for Referring Image Segmentation, and our proposed method, LQMFormer, produces competitive performance against existing state-of-the-art approaches across multiple standard benchmarks. On the RefCOCO dataset, our proposed method obtains the highest IoU across all divisions (val, test A, and test B), surpassing recent leading methods such as VLT [13] and ReLA [32]. Notably, it shows a significant margin in the test B split, indicating enhanced capability in complex visual contexts. In the RefCOCO+ dataset, LQMFormer also demonstrates its effectiveness in the test A set. Specifically, in the test A split with a cIoU of $71.84\\%$ . In the Val and Test B splits, our proposed method achieves competitive performance against ReLA [32] and LAVT [54]. For the G-Ref dataset, which presents varied and longer referring expressions, our proposed method achieves the top performance in the $\\mathrm{val}_{(\\mathrm{G})}$ split and maintains competitive results in the other splits. This performance is particularly notable over LAVT [54] in the $\\mathrm{val}_{(\\mathrm{G})}$ split, highlighting its adaptability to diverse challenging datasets. While the margin of improvement offered by LQMFormer in classic RIS datasets may be comparatively smaller than that observed in GREs, the results signify that Gaussian-enhanced Multi-Modal Fusion and explicit modeling of Queries from Decoder are beneficial for the general Referring Image Segmentation setting.", + "bbox": [ + 496, + 219, + 893, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalised Referring Image Segmentation In the GRES dataset, the LQMFormer model demonstrates notable improvements in Region Instance Segmentation (RIS), as shown in Table 2. The model achieves a generalized Intersection over Union (IoU) score of $70.94\\%$ on the validation set, surpassing other models such as ReLA, LAVT, and CRIS by a margin of $7 - 12\\%$ . This performance indicates its effective segmentation across various object sizes. The object-level Intersection over Union (oIoU) score of $64.98\\%$ further underscores the model's segmentation capabilities, highlighting the effectiveness of the Improved Visual Grounding and Dynamic Query Selection module. Another important aspect of LQMFormer is its performance in no-target identification, with an accuracy score of $67.47\\%$ . This score is $11\\%$ higher than that of existing methods, marking significant progress in accurately identifying no-target samples—a noted challenge in RIS. The inclusion of the Dynamic Query Module (DQM) and the Top $p\\%$ ratio significantly reduces the impact of the number of queries and mitigates query collapse. As a result, fewer but more effective queries make it easier to distinguish", + "bbox": [ + 496, + 598, + 895, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12908", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe773ca9fd2ddc1cb6fd9367ccc6f5aa7e1b668c7c94c35af9bfc0d965536678.jpg", + "table_caption": [ + "Table 1. Results on classic RES in terms of oIoU. U: UMD split. G: Google split." + ], + "table_footnote": [], + "table_body": "
MethodsVisual EncoderTextual EncoderRefCOCORefCOCO+val(U)G-Ref test(U)val(G)
valtest Atest Bvaltest Atest B
MCN [37]Darknet53bi-GRU62.4464.2059.7150.6254.9944.6949.2249.40-
VLT [12]Darknet53bi-GRU67.5270.4765.2456.3060.9850.0854.9657.7352.02
ReSTR [25]ViT-BTransformer67.2269.3064.4555.7860.4448.27--54.48
CRIS [49]CLIP-R101CLIP70.4773.1866.1062.2768.0853.6859.8760.36-
LAVT [54]Swin-BBERT72.7375.8268.7962.1468.3855.1061.2462.0960.50
VLT [13]Swin-BBERT72.9675.9669.6063.5368.4356.9263.4966.2262.80
ReLA [32]Swin-BBERT73.8276.4870.1866.0471.0257.6565.0065.9762.70
LQMFormer (ours)Swin-BBERT74.1676.8271.0465.9171.8457.5964.7366.0462.97
", + "bbox": [ + 78, + 103, + 893, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23c14dedbd952423f7b9ceca7b7b2e6a655a028a10b80d58a3e36052a4c778bc.jpg", + "table_caption": [ + "Table 2. Comparison on GRES dataset." + ], + "table_footnote": [], + "table_body": "
MethodsvalNo-target val
oIoUgIoUN-acc.T-acc.
MattNet [58]47.5148.2441.1596.13
LTS [24]52.3052.70--
VLT [12]52.5152.0047.1795.72
CRIS [49]55.3456.27--
LAVT [54]57.6458.4049.3296.18
ReLA [32]62.4263.6056.3796.32
LQMFormer (ours)64.9870.9467.4799.12
", + "bbox": [ + 104, + 268, + 439, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9985c7a5ab596ca069581f7fb7e7ca2195276e5b9e6ae2d1c7e9619f2c4905b4.jpg", + "table_caption": [ + "Table 3. Ablation Study Results" + ], + "table_footnote": [], + "table_body": "
ConfigurationoIoUgIoU
(a) Analysis on Dynamic Query module
Ours64.9870.94
w/o score-based learning62.7567.16
w/o scoring module62.0265.48
w/o QM-loss64.1968.37
(b) Analysis on multi-modal fusion module
Ours64.9870.94
w/o modality fusion module and QM-loss49.6852.34
", + "bbox": [ + 78, + 449, + 488, + 621 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ea90189f4651f8ed0205cf000312fb8177d94416615b9e5db1959f78ed005330.jpg", + "table_caption": [ + "Table 4. Ablation study of Query Numbers ${N}_{qb}$ in ${Q}_{b}$ . $\\ddagger$ : without scoring module. Top p% ratio is 20" + ], + "table_footnote": [], + "table_body": "
NqoIoUgIoUPr@0.7Pr@0.8Pr@0.9
2057.2258.3169.1560.2622.47
5062.4864.5372.0261.7931.58
10064.9870.9475.0365.5134.05
100‡62.0265.4873.1562.8731.55
", + "bbox": [ + 109, + 672, + 436, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "between \"target\" and \"no-target\". The model's specificity (true accuracy) score of $99.12\\%$ also reflects its effectiveness in classifying target samples.", + "bbox": [ + 75, + 791, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 847, + 230, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our comprehensive ablation studies, we evaluated the performance impact of various components of LQMFormer", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32d80fce1ebddcae2dcf57122b9fe83ed19ad8512b1a13bde018903d2f678c44.jpg", + "table_caption": [ + "Table 5. Ablation study of Top p% ratio in ${Q}_{b}$ ." + ], + "table_footnote": [], + "table_body": "
p%oIoUgIoUN.acc.Pr@0.7Pr@0.8Pr@0.9
2064.9870.9467.4775.0365.5134.05
5063.8567.7663.2574.2964.4233.58
10061.3363.7758.9272.4561.8831.21
", + "bbox": [ + 501, + 280, + 890, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2eba967af0c473b4c8c912bc202dceaf0eef097010acc20b93af5093c3079627.jpg", + "table_caption": [ + "Table 6. Ablation study of Grounding Module" + ], + "table_footnote": [], + "table_body": "
FusionoIoUgIoUPr@0.7Pr@0.8Pr@0.9
Baseline49.6852.3455.2741.3814.72
Post-CA60.2261.8974.1564.4432.67
S-Post-CA [40]62.1364.4173.6463.3532.71
VL-Grounding63.4368.8172.9564.1933.96
GMMF64.9870.9475.0365.5134.05
", + "bbox": [ + 514, + 393, + 879, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on the GRES dataset. These studies focused on the Dynamic Query Module, Query Numbers $N_{qb}$ in $Q_{b}$ , the Top $\\mathsf{p}\\%$ ratio in $Q_{b}$ , and the Grounding Module. The results helps understanding key insights into the model's functioning and optimization.", + "bbox": [ + 496, + 520, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on Dynamic Query Module: Table 3(a) presents the ablation results for the language query selection module. The removal of score-based learning significantly impacts scores as it limits the explicit training of Query scores from the scoring network. Also, notably, removing the scoring module and QM-loss leads to a decrease in oIoU score by $2.96\\%$ and $0.79\\%$ , respectively, highlighting its critical role in achieving optimal referring segmentation performance.", + "bbox": [ + 496, + 595, + 892, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Query Bank number $N_{qb}$ in $Q_{b}$ : The ablation on Query Bank number demonstrated a notable impact on the model's performance as shown in Table 4. With an increase in $N_{q}$ from 20 to 100, there was a consistent improvement in both IoU from 57.22% to 64.98% and gIoU from 58.31% to 70.94%, as well as in Precision metrics. Particularly, $N_{q} = 100$ yielded the highest scores across all metrics, indicating an optimal balance in query numbers for effective segmentation. Thus, it optimal to have more number of queries with Top p% as 20.", + "bbox": [ + 496, + 718, + 893, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Top p% Ratio in $Q_{b}$ : The study on the Top p% ratio revealed that a 20% ratio achieved the best performance in", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12909", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c99725aeceedbe99d02f67028ca126037faa12ecd804958cd824f88c5444f2a0.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [ + "Query: \"Orange cow with front facing us and an orange cow with back facing us\"" + ], + "bbox": [ + 98, + 117, + 285, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/710ec28d0f6087f1f95e81d9a7f18d23d81c57275ff7db02688a4b8340b1bc9c.jpg", + "image_caption": [ + "ReLA" + ], + "image_footnote": [], + "bbox": [ + 300, + 117, + 486, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bbae835706fe86cda15df2823a351e795f182da57fcf295385d200680f6f055.jpg", + "image_caption": [ + "LQMFormer" + ], + "image_footnote": [], + "bbox": [ + 503, + 117, + 689, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/042944f6b5ffade7820e6adcd8845713c1ad98c0949ed153f26f58880014a2f3.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 704, + 117, + 890, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c0f1a83cf9a3602ca6ee537da02f1f9ee6e7d88e51b7ddeccd20cde2ee5bdb1.jpg", + "image_caption": [ + "Figure 5. Qualitave Comparison of our model with LQMFormer with ReLA [32] on GRES dataset." + ], + "image_footnote": [ + "Query: \"The pizza in the middle and the guy holding a camera\"" + ], + "bbox": [ + 98, + 300, + 284, + 444 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1aaecd4830aa26d2c362a65435b50ae3b9ce9bf2f3382b9cc4eada973fa6e095.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 299, + 488, + 444 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b3264f5163287e93fd3249ff0231101f8157415e7338998b54c20ae4b555136.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 299, + 689, + 443 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2403e8bb106cf2cb68033e8bf81ff97f161dc2f63d965a6d82e56d02c0393372.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 299, + 892, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "terms of oIoU, gIoU, and Precision metrics(Table 5). Higher ratios, like $50\\%$ and $100\\%$ , resulted in diminished performance as it resulted in more and more query collapse issues as many queries contribute towards the final mask generation. This finding indicates that a moderate selection of top-performing queries is crucial for balancing precision and recall.", + "bbox": [ + 75, + 506, + 470, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Grounding Module: The ablation study of the Grounding Module, as shown in Table 3(b) and Table 6, highlights the effectiveness of Gaussian Enhancement for Referring Image Segmentation (RIS) tasks. The Gaussian enhanced Multi-Model Fusion (GMMF) method yields a improvement, specifically enhancing oIoU by $1.5\\%$ , from $63.43\\%$ in VL-Grounding to $64.98\\%$ in GMMF. This underlines the superior performance of Gaussian Filtering in the Fourier Domain over traditional cross-attention methods. Furthermore, compared to the Post-CA which registers an oIoU of $60.22\\%$ , the GMMF and VL-Grounding methods showcase substantial improvements in multi-modal representation within the backbone stage, with GMMF outperforming Post-CA by $4.76\\%$ and VL-Grounding by $3.21\\%$ in oIoU. The comparison with S-Post-CA [40], which aims at enhancing visual representation post-feature extraction, show the importance of integrating advanced grounding techniques early in the model's architecture for improved RIS performance. Overall, our proposed Gaussian enhanced Multi-Model Fusion", + "bbox": [ + 75, + 613, + 473, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(GMMF) improves the visual grounding compared to other methods, resulting in a improved referring segmentation.", + "bbox": [ + 500, + 506, + 892, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 566, + 617, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we address the problem of query collapse in Referring Image Segmentation by enhancing visual grounding and generating diverse query sets, and by modeling their relevance based on language expressions. Our proposed model, LQMFormer, incorporates Gaussian Enhanced Multi-Modal Fusion to improve visual grounding, and a Dynamic Query Module that not only generates a diverse set of queries but also employs a scoring network to selectively focus on queries based on given language expressions. Additionally, to mitigate query collapse, we introduce a novel loss function that enforces a margin of separation between query feature representations, facilitating improved representation without the need for additional supervision. The LQMFormer method notably surpasses state-of-the-art methods in both referring image segmentation and generalized referring image segmentation performance in most settings. For future work, we aim to extend the LQMFormer model's capabilities along with LLM and focus on effectively bridging LLM to better handle multimodal contexts.", + "bbox": [ + 496, + 598, + 895, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12910", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Ankan Bansal, Yuting Zhang, and Rama Chellappa. Visual question answering on image sets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 51–67. Springer, 2020. 2", + "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3", + "[4] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. arXiv preprint arXiv:2203.14713, 2022. 2", + "[5] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In ICCV, pages 7454-7463, 2019. 2", + "[6] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 2", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3", + "[8] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM Transactions on Graphics (ToG), 34(1):1-11, 2014. 1", + "[9] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL-HLT, pages 4171–4186. Association for Computational Linguistics, 2019. 2, 6", + "[12] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, pages 16321-16330, 2021. 7", + "[13] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 6, 7", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 4" + ], + "bbox": [ + 78, + 114, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 3", + "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In CVPR, 2021. 1", + "[17] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1330, 2021. 3", + "[18] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 3", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2", + "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4, 5", + "[21] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, pages 108-124. Springer, 2016. 1, 2", + "[22] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In CVPR, pages 4424-4433, 2020. 2", + "[23] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2", + "[24] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In CVPR, pages 9858-9867, 2021. 2, 7", + "[25] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In CVPR, pages 18145-18154, 2022. 7", + "[26] Sangrok Lee, Jongseong Bae, and Ha Young Kim. Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11776-11785, 2023. 3", + "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. arXiv preprint arXiv:2302.11831, 2023. 3", + "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2", + "[29] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In CVPR, pages 5745-5753, 2018. 2", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12911", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[31] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In ICCV, pages 1271-1280, 2017. 2", + "[32] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 3, 6, 7, 8", + "[33] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In ICCV, pages 4856-4864, 2017. 1", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3, 6", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[36] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10437-10446, 2020. 2", + "[37] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, pages 10034-10043, 2020. 6, 7", + "[38] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 6", + "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3", + "[40] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920–930, 2023. 3, 7, 8", + "[41] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, pages 792–807. Springer, 2016. 6", + "[42] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 2", + "[43] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 3", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2", + "[45] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, pages 38-54, 2018. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. pages 5998-6008, 2017. 2", + "[47] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2", + "[48] Yiyu Wang, Jungang Xu, and Yingfei Sun. End-to-end transformer based model for image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2585-2594, 2022. 2", + "[49] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, pages 11686-11695, 2022. 2, 6, 7", + "[50] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. 6", + "[51] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14383-14392, 2021. 3", + "[52] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3", + "[53] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3", + "[54] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In CVPR, pages 18155-18165, 2022. 1, 2, 3, 4, 6, 7", + "[55] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, pages 10502–10511, 2019. 1", + "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[57] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 2, 6", + "[58] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 2, 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12912", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 2", + "[60] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In International Conference on Machine Learning, pages 42589-42601. PMLR, 2023. 3", + "[61] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3" + ], + "bbox": [ + 78, + 90, + 470, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12913", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_model.json b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b909d5815529f480a2975f4fdcd8dd384285de03 --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.726, + 0.177 + ], + "angle": 0, + "content": "LQMFormer: Language-aware Query Mask Transformer for Referring Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.204, + 0.701, + 0.24 + ], + "angle": 0, + "content": "Nisarg A. Shah Vibashan VS Vishal M. Patel Johns Hopkins University" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.242, + 0.702, + 0.257 + ], + "angle": 0, + "content": "snisarg812@gmail.com, {vvishnu2, vpatel36}@jhu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.705 + ], + "angle": 0, + "content": "Referring Image Segmentation (RIS) aims to segment objects from an image based on a language description. Recent advancements have introduced transformer-based methods that leverage cross-modal dependencies, significantly enhancing performance in referring segmentation tasks. These methods are designed such that each query predicts different masks. However, RIS inherently requires a single-mask prediction, leading to a phenomenon known as Query Collapse, where all queries yield the same mask prediction. This reduces the generalization capability of the RIS model for complex or novel scenarios. To address this issue, we propose a Multi-modal Query Feature Fusion technique, characterized by two innovative designs: (1) Gaussian enhanced Multi-Modal Fusion, a novel visual grounding mechanism that enhances overall representation by extracting rich local visual information and global visual-linguistic relationships, and (2) A Dynamic Query Module that produces a diverse set of queries through a scoring network where the network selectively focuses on queries for objects referred to in the language description. Moreover, we show that including an auxiliary loss to increase the distance between mask representations of different queries further enhances performance and mitigates query collapse. Extensive experiments conducted on four benchmark datasets validate the effectiveness of our framework." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.738, + 0.21, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Referring Image Segmentation (RIS) is a challenging multimodal task aimed at segmenting specific objects in an image based on a textual description. This description often includes details about the object's actions, category, color, or position, as noted by [8, 21]. Compared to traditional semantic and instance segmentation, RIS requires a precise understanding of object locations and involves a comprehensive modeling of the visual-linguistic relationships within the global context. Additionally, RIS demands the extrac" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.286, + 0.697, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.431, + 0.636, + 0.44 + ], + "angle": 0, + "content": "(a) ReLA" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.289, + 0.864, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.432, + 0.832, + 0.442 + ], + "angle": 0, + "content": "(b) LQMFormer" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.468, + 0.895, + 0.539 + ], + "angle": 0, + "content": "Figure 1. [Top] Heatmap visualization of cosine similarity between different query mask predictions of ReLA and LQMFormer, where the yellow region indicates a similarity near 1, and the dark blue region signifies a similarity of 0. [Bottom] Sample query mask prediction visualizations for ReLA and LQMFormer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.612 + ], + "angle": 0, + "content": "tion of high-quality visual features at a local level. The potential applications of RIS are extensive, particularly in language-driven human-computer interaction domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Traditional RIS methods have employed linear fusion approaches and Fully Convolutional Networks (FCNs) for feature learning in RIS tasks [21, 33]. However, the recent advancement of attention mechanisms has shifted the focus towards extracting richer visual-language representation, with recent techniques demonstrating the effectiveness of Transformers. These models excel at modelling long-distance dependencies, making them suitable for cross-modal fusion [45, 55]. Further, Vision Transformer (ViT) methods such as VLT [13], EFN [16], and LAVT [54], which are based on [14], have shown significant improvements in RIS performance. Their ability to capture the nuances of cross-modal dependencies is crucial, and through a series of attention mechanisms, these models efficiently process and integrate both visual and language inputs for precise object segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Despite their advantages, transformer-based approaches exhibit shortcomings in RIS mask prediction, with a notable problem being query collapse. This phenomenon occurs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.409 + ], + "angle": 0, + "content": "when different queries within the transformer, intended to identify distinct attributes or portions of objects, converge on the same mask prediction. This is particularly problematic in referring image segmentation, which inherently requires a prediction of a single mask, where all queries are collectively trained to predict the same mask. We experimentally observed this phenomenon, as illustrated in Fig. 1. Here, we visualize the heatmap of the cosine similarity between mask predictions from ReLA [32] and our method. In ReLA, most mask predictions are almost identical, leading to significantly overlapping mask predictions, whereas our method demonstrates a diverse set of query embeddings and mask predictions. Hence, prior transformer-based methods train mask predictions for all queries to associate with a single, specific ground truth mask. This training approach can lead to query collapse, where the model is not penalized for producing the same prediction across different queries, thereby affecting its capability to identify diverse attributes of the image. Consequently, this limitation restricts the variety of mask predictions, diminishing the model's ability to generalize effectively in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.473, + 0.761 + ], + "angle": 0, + "content": "To address these challenges, we propose LQMFormer, which comprises two key components. Firstly, we introduce a Gaussian Enhanced Multi-Modal Fusion (GMMF) module, aimed at enhancing the visual grounding mechanism. This module is designed to extract fine-grained local visual details while simultaneously enhancing global visual-language relationships. The fusion of these modalities allows the model to construct a more comprehensive understanding of the scene, both globally and locally, thereby improving visual grounding. Secondly, in our Dynamic Query Module (DQM), we generate a diverse set of language-dependent queries through Language-Query Cross Attention (LQCA) and employ a scoring network to prioritize queries relevant to the referring expression. With an improved visual-grounded representation facilitated by GMMF and selective query enhancement via DQM, our model effectively addresses query collapse. Moreover, by incorporating an auxiliary loss that increases the representational distance between mask predictions, our method efficiently differentiates between queries, mitigating the issue of query collapse. Our extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our framework. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.765, + 0.472, + 0.81 + ], + "angle": 0, + "content": "- We introduce Gaussian enhanced Multi-Modal Fusion, which innovates the integration of local and global cross-modal information for more accurate visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.471, + 0.884 + ], + "angle": 0, + "content": "- We develop an effective Dynamic Query Module (DQM) that not only generates a diverse query set but also employs a scoring network to selectively focus on queries based on given language expressions with respect to the decoder's references." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- To mitigate query collapse, we propose an auxiliary regu" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.765, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "larization loss function that increases the representational distance between queries, significantly improving mask differentiation and preventing Query Collapse." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.138, + 0.892, + 0.198 + ], + "angle": 0, + "content": "- We conduct comprehensive validation of the proposed framework with extensive testing on four benchmark datasets, demonstrating improvements in referring image segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.65, + 0.228 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.238, + 0.774, + 0.254 + ], + "angle": 0, + "content": "2.1. Referring Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.895, + 0.732 + ], + "angle": 0, + "content": "Early works [29, 31, 45, 57] employed Convolutional Neural Networks (CNNs) [5, 21, 22, 31] and Recurrent Neural Networks (RNNs) [6, 31] to extract vision and language features. These features were then fused via sequential concatenation and convolution operations to predict the segmentation mask. In contrast, [58] proposed a two-stage network that initially uses Mask R-CNN[19] to predict categorical masks, followed by a selection of the relevant masks based on language descriptions. However, this approach exhibits limited capacity to capture the intricate relations between language and visual content. The advent of the Transformer [11, 14, 42, 46] in the vision community has led to the widespread adoption of Transformer-based architectures for extracting both visual and textual features [11, 24, 34]. VLT [13] uses cross-attention to produce query vectors from multi-modal features, which are then used to query images in the transformer decoder. Similarly, LAVT [54] shows early cross-modal fusion of features improves alignment. CRIS [49] adopts Transformer blocks to leverage the pretrained CLIP model's [44] robust image-text alignment capabilities. Additionally, GRES [32] extends cross-attention in the Transformer decoder to explicitly model visual-language dependencies. However, these all transformer-based approaches require matching queries to corresponding mask instances. In the context of referring expression segmentation, where typically only a single mask is available, this can lead to all queries converging to a single mask, a phenomenon known as query collapse. Motivated by this, in our method, we try to produce a diverse set of query prediction by overall improving visual-grounding and conditioning with respect to language expression." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.892, + 0.758 + ], + "angle": 0, + "content": "2.2. Vision and Language Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Vision and Language Representation Learning focuses on understanding the semantics and alignment between vision and language for multimodal reasoning tasks [36, 47, 59]. This field has seen substantial progress in applications like visual question answering [2], Image captioning [48], Image-text retrieval [4], zero-shot classification [44], and Image segmentation [23]. Contrastive pre-training strategies [28, 44] using large-scale datasets effectively project multiple modalities into a single embedding space. In contrast, as discussed," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "other methods [32, 54] create cross-modal interaction layers to fuse and comprehend multimodal features. The recent adoption of deep learning techniques in the frequency domain [9, 17, 39, 40] has gained attention for their global interaction capabilities. Specifically, [40] performs spectral-guided enhancement of the multi-scale vision-language features after feature extraction stage for Video Referring Segmentation. Drawing from these advancements, our approach integrates gaussian guidance within vision-language representation learning to facilitate enhanced local and global multi-modal interactions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.176, + 0.285 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.472, + 0.342 + ], + "angle": 0, + "content": "Given an input image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\) and a language expression \\(L = \\{w_{i}\\}_{i=1}^{N}\\) with \\(N\\) words, our model predicts a pixel-wise mask \\(M\\), which delineating the referred object." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.347, + 0.188, + 0.362 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.473, + 0.706 + ], + "angle": 0, + "content": "An overview diagram of our approach is shown in Figure 2. First, the language expression is encoded to extract high-dimensional language features \\( F_{l} \\in \\mathbb{R}^{N_{t} \\times C} \\), where \\( C, N_{t} \\) indicates the number of channels, and words in the language expression, respectively. The Gaussian-enhanced Multi-Modal Fusion (GMMF) module then extracts language-grounded visual features \\( F_{\\mathrm{vl}} \\). Following [7], our approach incorporates a multi-scale strategy to exploit higher-resolution feature maps in the transformer decoder [7, 61]. We input language-aware queries \\( Q_{b}^{\\prime} \\) along with multi-scale outputs from the pixel decoder \\( F_{i} \\), where \\( i \\in \\left\\{\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\} \\) of the original image. The Dynamic Query Module processes language features \\( F_{t} \\) to predict language-aware queries \\( Q_{b}^{\\prime} \\). The updating of queries \\( Q_{b}^{\\prime} \\) sequentially utilizes resolutions \\( \\frac{1}{8}, \\frac{1}{16} \\), and \\( \\frac{1}{32} \\), applying a cycle of masked cross-attention (CA), self-attention (SA), and feed-forward network (FFN) operations \\( L \\) times within the decoder. Further, we map the final query outputs to an 'object' or 'no-object' two-dimensional space to enable 'no-object' predictions [32]. The final predicted mask is obtained at full resolution as the original image by decoding the pixel features \\( F_{\\frac{1}{4}} \\) with the einsum operation between \\( Q \\) and \\( F_{\\frac{1}{4}} \\), followed by upsampling operation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.424, + 0.727 + ], + "angle": 0, + "content": "3.2. Gaussian-enhanced Multi-Modal Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "After extracting language features \\( F_{l} \\), we combine them with visual features using a four-stage hierarchical Swin Transformer [34] to obtain joint visual-language embedding. Each stage comprises Transformer layers \\( (\\tau_{i}) \\), multi-modal feature fusion modules (GMMF), and residual gating \\( (\\psi_{i}) \\). At every stage, as Transformer layers \\( (\\tau_{i}) \\) enhance previous visual features \\( (F_{v}) \\), these features \\( (F_{v}) \\) combine with language features \\( (F_{l}) \\) via multi-modal modules (GMMF) to create multi-modal features \\( (F_{vl}) \\). Finally, we use gating units \\( (\\psi_{vl}) \\) to weight and combine \\( F_{vl} \\) with \\( F_{v} \\), yielding enhanced visual features with linguistic information. Residual gating" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.138 + ], + "angle": 0, + "content": "also works well with pre-trained (only vision) transformer weights and allows simultaneous feature modulation with language without having to re-train from scratch." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.155, + 0.892, + 0.184 + ], + "angle": 0, + "content": "3.2.1 Gaussian Enhanced Multi-Modal Fusion Module in Visual Backbone" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.193, + 0.895, + 0.45 + ], + "angle": 0, + "content": "The discrete Fourier transform (DFT) is extensively used to analyze the frequency representation of images. By transforming the signal into the frequency domain, characterized by global statistical properties, it plays a crucial role in various computer vision tasks [26, 51, 56]. Modulating signals on a point-wise basis in the Fourier domain alters the representation of inputs in the spatial domain [3, 40]. Furthermore, in the Fourier domain, low-frequency components usually correspond to the overall semantic content of the image, consistent with prior research [40, 52, 53, 56]. The Fourier transform is capable of disentangling global semantic information and structural coherence, with the semantic content predominantly found in the amplitude component [18, 27, 60]. These insights can be leveraged to create modules centered on Fourier-based methods, facilitating efficient and vital global interactions in multimodal understanding and improving visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.897, + 0.707 + ], + "angle": 0, + "content": "To inculcate the above observation, as illustrated in Fig. 3, following [40], the Gaussian enhanced Multi-Modal Fusion Module (GMMF) is proposed as a key component in our model. It employs an enhancement block after cross-attention operations to combine Gaussian-enhanced vision and language features. Given an input vision feature \\( F_{vi} \\in \\mathbb{R}^{N \\times H \\times W \\times C_{vi}} \\) from stage \\( i \\) and a language feature \\( F_{l} \\in \\mathbb{R}^{N \\times T \\times C_{l}} \\), for cross-attention block, we first transform visual and language features into a common dimension \\( C_{i} \\) using separate \\( 1 \\times 1 \\) convolution layers. Subsequently, transform vision and language features are passed to an MHCA (Multi-Head Cross-Attention) module, where language features act as queries and vision features as keys and values. This operation results in the creation of a language-aware vision-shaped feature \\( F_{vl}' \\). We then transform the visual-language features using the Gaussian Enhancement Block, as shown in Fig. 3, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.713, + 0.843, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {v l} = F _ {\\mathrm {F F T}} ^ {- 1} \\left(\\operatorname {C o n v} \\left(F R \\left(\\phi \\left(F _ {v l} ^ {\\prime}, \\beta\\right) * \\mathcal {A} \\left(F _ {v l} ^ {\\prime}\\right) \\right. \\right. \\right. \\\\ , \\left. \\left. \\mathcal {P} \\left(F _ {v l} ^ {\\prime}\\right)\\right)\\right) + F _ {v l} ^ {\\prime} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.774, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{A}(F_{vl}^{\\prime})\\) and \\(\\mathcal{P}(F_{vl}^{\\prime})\\) denote the amplitude and phase components of \\(F_{vl}^{\\prime}\\)'s Fast Fourier Transform(FFT) [15, 43], \\(F_{\\mathrm{FFT}}(F_{vl}^{\\prime})\\). The * represents low-pass filtering on \\(\\mathcal{A}(F_{vl}^{\\prime})\\) using adaptive Gaussian filters \\(\\phi (F_{vl}^{\\prime},\\beta)\\), matching \\(F_{vl}^{\\prime}\\)'s spatial dimensions and adjusting to input data via bandwidth \\(\\beta\\) [40]. \\(FR\\) and \\(Conv\\) signify feature reconstruction from amplitude and phase, and convolution operations, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "12905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.895, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 2. Overview of the LQMFormer model architecture. The model processes an image and a linguistic expression, extracting vision-language features \\( F_{vl} \\) through a Gaussian-enhanced Multi-modal Fusion module and language features \\( F_{L} \\). The multi-scale vision-language features \\( F_{vl} \\) are then processed by into a pixel decoder. Concurrently, the Language Query Cross-Attention Module (LQCA) refines the Language-enhanced Query bank \\( Q_{b}^{\\prime} \\) with \\( F_{L} \\). A Scoring Network assesses the relevance of each query in \\( Q_{b}^{\\prime} \\) by predicting soft-masks \\( (M_{b}) \\), which then modulate the queries fed into the transformer decoder. Finally, the model employs a global pooling layer and a Multi-Layer Perceptron (MLP) to predict the binary segmentation mask and determine the presence or absence of the referred object." + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.442, + 0.353, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Figure 3. Architecture overview of the Gaussian Enhancement Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Once \\(F_{vl}\\) is obtained, similar to [54], we combine the output from GMMF, \\(F_{vl}\\), with that from the Transformer layers, \\(F_{v}\\). This process employs a gating mechanism, \\(\\psi_{vl}\\), which learns a set of element-wise weight maps from \\(F_{vl}\\) to finely tune the scale of each element within \\(F_{vl}\\). Subsequent to this adjustment, a residual combination of \\(F_{v}\\) and \\(F_{vl} \\cdot \\psi_{vl}\\) is performed. The resultant output is then fed back into the vision backbone for further computation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.3, + 0.849 + ], + "angle": 0, + "content": "3.3. Dynamic Query Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The Dynamic Query Module (DQM) takes the language feature \\( F_{l} \\) and a Query bank \\( Q_{b} \\) as inputs, the latter containing \\( N_{qb} \\) learnable queries. Figure 4(b) shows the initial" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.444, + 0.895, + 0.492 + ], + "angle": 0, + "content": "step, where the attention between the language feature \\(F_{l}\\) and the \\(N_{qb}\\) query embeddings \\(Q_{b} \\in \\mathbb{R}^{N_{qb} \\times C}\\) is computed, resulting in \\(N_{qb}\\) attention maps:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.504, + 0.893, + 0.523 + ], + "angle": 0, + "content": "\\[\nA _ {b i} = \\operatorname {s o f t m a x} \\left(Q _ {b} \\sigma \\left(F _ {l} W _ {i k}\\right) ^ {T}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.895, + 0.656 + ], + "angle": 0, + "content": "where \\(W_{ik}\\) is a \\(C\\times C\\) matrix of learnable parameters, and \\(\\sigma\\) denotes the GeLU function [20]. The resulting \\(A_{bi}\\in \\mathbb{R}^{N_{qb}\\times L}\\) provides each query with a \\(1\\times L\\) attention map, indicating its important linguistic relation in the language features. Following this, the language-supported query bank features are derived using these attention maps: \\(Q_b^{\\prime} = A_{bi}\\sigma (F_lW_{iv})^T\\), where \\(W_{iv}\\) is another \\(C\\times C\\) matrix of learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.895, + 0.87 + ], + "angle": 0, + "content": "The variability in language expressions used to describe objects in images requires an approach that can adapt to different descriptions. This is because standard methods, like those seen in the initial transformer models (like in vanilla transformer [14]), struggle to capture the nuances in descriptions that include color, size, shape, and position. To address this, a larger set of adaptable queries is required. These queries are fine-tuned based on the input language, resulting in a refined set of queries, denoted as \\( Q_b' \\). Each query in this refined set is then assessed with a scoring network, which outputs a soft-mask \\( M_b \\). This soft-mask represents the relevance of each query. The final set of queries, \\( Q_{\\mathrm{final}} \\), is obtained by combining \\( M_b \\) with \\( Q_b' \\) through a dot product, as shown in the equation below:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.885, + 0.893, + 0.902 + ], + "angle": 0, + "content": "\\[\nQ _ {\\text {f i n a l}} = M _ {b} ^ {\\prime} \\odot Q _ {b} ^ {\\prime} \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.089, + 0.631, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.09, + 0.887, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.328, + 0.756, + 0.343 + ], + "angle": 0, + "content": "Figure 4. Architecture overview of the (A.) Dynamic Query Module and (B.) LQCA Block" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.25, + 0.385 + ], + "angle": 0, + "content": "3.3.1 Scoring Network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.514 + ], + "angle": 0, + "content": "The scoring network in LQMFormer takes the Language-enhanced Query bank \\( Q_{b}^{\\prime} \\) as input and processes it through a sequence of linear layers. This network consists of four linear layers, with Layer Normalization [1] applied before the first layer. Each subsequent layer, except the last sigmoid, incorporates GELU [20] activation. The final output is a 1-dimensional logit \\( M_{b} \\in \\mathbb{R}^{N_{qb} \\times 1} \\), serving as a soft-mask to modulate \\( Q_{b}^{\\prime} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.474, + 0.651 + ], + "angle": 0, + "content": "To train the scoring network for predicting a relevant soft-mask \\(M_{b}\\) effectively for the transformer decoder, we utilize its cross-attention map. This map is essential in identifying the subset of Queries most relevant to the referring expression and visual features, as highlighted during training. We sum up the attention maps from each decoder layer using Bilinear Interpolation (BI), thereby projecting attention \\(A_{qf} \\in \\mathbb{R}^{N_{qb} \\times HW}\\) directed towards these Queries. The process is mathematically described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.661, + 0.47, + 0.702 + ], + "angle": 0, + "content": "\\[\nA _ {q f} = \\sum_ {l = 1} ^ {L} \\sum_ {r = 1} ^ {3} \\mathrm {B I} \\left(\\mathrm {A} _ {l, r}, H W\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.707, + 0.471, + 0.8 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{A}_{l,r} \\in \\mathbb{R}^{N_{qb} \\times hw}\\) represents the attention map from the \\(l^{th}\\) layer at the \\(r^{th}\\) resolution, with \\(hw\\) being the original dimension and \\(HW\\) the desired output dimension. After aligning each attention map to the common resolution \\(HW\\), we then sum across this dimension to generate a likelihood map \\(LM\\), as indicated in the subsequent equation:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.809, + 0.47, + 0.85 + ], + "angle": 0, + "content": "\\[\nL M = \\sum_ {i = 1} ^ {H W} A _ {q f} [:, i ] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This likelihood map \\(\\mathrm{LM} \\in \\mathbb{R}^{N_{qb} \\times 1}\\), now captures the concentrated attention across all Queries, serves as a condition in refining the predictions of the scoring network. The" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.37, + 0.895, + 0.492 + ], + "angle": 0, + "content": "scoring network is trained to predict the likelihood of each Query being included in the top-\\(\\rho\\%\\) of the most referenced tokens. This process involves binarizing the likelihood map \\(LM\\) to retain only the top-\\(\\rho\\%\\) Queries, creating a binarized version \\(LM_{\\mathrm{bin}}\\). The training objective is then formulated as a Binary Cross-Entropy (BCE) loss between the predicted soft-mask \\(M_b\\) and the binarized likelihood map \\(LM_{\\mathrm{bin}}\\). The BCE loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.511, + 0.893, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {B C E} \\operatorname {L o s s} = - \\frac {1}{N _ {q b}} \\sum_ {i = 1} ^ {N _ {q b}} \\left[ L M _ {\\mathrm {b i n}, i} \\log \\left(M _ {b, i}\\right) \\right. \\\\ + \\left(1 - L M _ {\\text {b i n}, i}\\right) \\log \\left(1 - M _ {b, i}\\right) ] \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.597, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Here, \\(N_{qb}\\) represents the number of Queries in \\(Q_b^{\\prime}\\), and \\(i\\) indexes each Query. This loss function effectively trains the scoring network to align its predictions with the most salient Queries as determined by the decoder's cross-attention, thereby refining the model's accuracy in identifying and segmenting objects as described by the referring expressions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.738, + 0.713 + ], + "angle": 0, + "content": "3.4. Query-Mask Margin Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Query-Mask Margin Loss is designed to maintain a margin of separability between different query feature representations, an essential aspect for tasks with variable outputs. The loss function operates by initially computing pairwise differences and distances between mask embeddings within a batch. This computation yields a tensor of shape \\( B \\times N \\times N \\), where \\( B \\) represents the batch size, and \\( N \\) is the number of queries. Subsequently, these distances for each embedding are sorted, with a particular focus on the second smallest value, which signifies the distance to the nearest neighbor. The margin loss is defined through the max function, ensuring a minimum distance of 1.0 between embeddings. The" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12907" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.423, + 0.108 + ], + "angle": 0, + "content": "mathematical formulation of QM-loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.118, + 0.412, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathrm {Q M - l o s s} = \\max (0, 1. 0 - \\operatorname {d i s t} \\left(M _ {i}, M _ {j}\\right))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.145, + 0.471, + 0.176 + ], + "angle": 0, + "content": "Here, \\( \\text{dist}(M_i, M_j) \\) denotes the calculation of pairwise distances between the mask embeddings, represented as \\( M \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.189, + 0.341, + 0.207 + ], + "angle": 0, + "content": "4. Experiments and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.213, + 0.398, + 0.23 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.473, + 0.508 + ], + "angle": 0, + "content": "Datasets: Our experiments are conducted on four primary benchmark datasets in Referring Image Segmentation: RefCOCO [57], RefCOCO+ [57], RefCOCOg [38], and GRES [32]. These datasets derive their images from MSCOCO [30]. The dataset details are as follows: RefCOCO (19,994 images, 50,000 objects, 142,209 expressions), RefCOCO+ (19,992 images, 49,856 objects, 141,564 expressions), RefCOCOg (26,711 images, 54,822 objects, 85,474 expressions), and GRES (19,994 images, 60,287 objects, 278,232 expressions). RefCOCO mainly focuses on expressions that specify object locations, while RefCOCO+ prioritizes object descriptions. RefCOCOg presents a higher challenge due to longer and more complex expressions (averaging 8.4 words compared to 3.5 in others). GRES broadens the scope of RefCOCO by including expressions that refer to multiple objects or no objects, thereby enhancing the problem's generalizability. We use both UMD split [41] and Google split [38] for RefCOCOg dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.473, + 0.795 + ], + "angle": 0, + "content": "Evaluation Metrics. For single-target object segmentation, we report our results using two kinds of metrics [13, 37, 49, 54]: overall IoU (oIoU) and Precision@X (P@X). oIoU divides the total intersection pixels by the total union pixels across all test images. Precision@X calculates the percentage of testing samples of which the model prediction has an IoU score higher than the threshold value X. To extend our evaluation to non-target and multi-target object segmentation, following [32], we include additional metrics: Sensitivity (N-acc.), Specificity (T-acc.), and Generalized IoU (gIoU). N-acc. evaluates the model's performance on identifying non-target samples, while T-acc. measures the impact of generalization on non-target samples on the performance of target samples. We opt for gIoU over oIoU due to the latter's bias towards larger objects. gIoU computes the mean IoU per image for all samples. In cases with no-target samples, IoU values are assigned as 1 for true positives and 0 for false negatives, ensuring an equitable metric for all object sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our model uses Swin Transformer-B [34] as the visual encoder and BERT [11] as the language decoder. The transformer layer in Swin-B is initialized with classification weights pre-trained on ImageNet22K [10]. We use BERT implementation from the HuggingFace Transformer library [50]. It comprises 12 layers with a hidden size of 768, initialized using official" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.184 + ], + "angle": 0, + "content": "pre-trained weights. We use our Transformer decoder proposed in Section 3.1 with \\( \\mathrm{L} = 3 \\) (i.e., 9 layers total) and 100 queries by default. The AdamW optimizer [35] is employed for optimization with an initial learning rate of 0.00001. Our model is trained over 100K iterations with a batch size of 48. Our image resizing is standardized to \\( 480 \\times 480 \\) pixels." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.196, + 0.857, + 0.212 + ], + "angle": 0, + "content": "4.2. Comparison with the state of art methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Referring Image Segmentation Table 1 presents the results for Referring Image Segmentation, and our proposed method, LQMFormer, produces competitive performance against existing state-of-the-art approaches across multiple standard benchmarks. On the RefCOCO dataset, our proposed method obtains the highest IoU across all divisions (val, test A, and test B), surpassing recent leading methods such as VLT [13] and ReLA [32]. Notably, it shows a significant margin in the test B split, indicating enhanced capability in complex visual contexts. In the RefCOCO+ dataset, LQMFormer also demonstrates its effectiveness in the test A set. Specifically, in the test A split with a cIoU of \\(71.84\\%\\). In the Val and Test B splits, our proposed method achieves competitive performance against ReLA [32] and LAVT [54]. For the G-Ref dataset, which presents varied and longer referring expressions, our proposed method achieves the top performance in the \\(\\mathrm{val}_{(\\mathrm{G})}\\) split and maintains competitive results in the other splits. This performance is particularly notable over LAVT [54] in the \\(\\mathrm{val}_{(\\mathrm{G})}\\) split, highlighting its adaptability to diverse challenging datasets. While the margin of improvement offered by LQMFormer in classic RIS datasets may be comparatively smaller than that observed in GREs, the results signify that Gaussian-enhanced Multi-Modal Fusion and explicit modeling of Queries from Decoder are beneficial for the general Referring Image Segmentation setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Generalised Referring Image Segmentation In the GRES dataset, the LQMFormer model demonstrates notable improvements in Region Instance Segmentation (RIS), as shown in Table 2. The model achieves a generalized Intersection over Union (IoU) score of \\(70.94\\%\\) on the validation set, surpassing other models such as ReLA, LAVT, and CRIS by a margin of \\(7 - 12\\%\\). This performance indicates its effective segmentation across various object sizes. The object-level Intersection over Union (oIoU) score of \\(64.98\\%\\) further underscores the model's segmentation capabilities, highlighting the effectiveness of the Improved Visual Grounding and Dynamic Query Selection module. Another important aspect of LQMFormer is its performance in no-target identification, with an accuracy score of \\(67.47\\%\\). This score is \\(11\\%\\) higher than that of existing methods, marking significant progress in accurately identifying no-target samples—a noted challenge in RIS. The inclusion of the Dynamic Query Module (DQM) and the Top \\(p\\%\\) ratio significantly reduces the impact of the number of queries and mitigates query collapse. As a result, fewer but more effective queries make it easier to distinguish" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12908" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.089, + 0.725, + 0.103 + ], + "angle": 0, + "content": "Table 1. Results on classic RES in terms of oIoU. U: UMD split. G: Google split." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.104, + 0.894, + 0.233 + ], + "angle": 0, + "content": "
MethodsVisual EncoderTextual EncoderRefCOCORefCOCO+val(U)G-Ref test(U)val(G)
valtest Atest Bvaltest Atest B
MCN [37]Darknet53bi-GRU62.4464.2059.7150.6254.9944.6949.2249.40-
VLT [12]Darknet53bi-GRU67.5270.4765.2456.3060.9850.0854.9657.7352.02
ReSTR [25]ViT-BTransformer67.2269.3064.4555.7860.4448.27--54.48
CRIS [49]CLIP-R101CLIP70.4773.1866.1062.2768.0853.6859.8760.36-
LAVT [54]Swin-BBERT72.7375.8268.7962.1468.3855.1061.2462.0960.50
VLT [13]Swin-BBERT72.9675.9669.6063.5368.4356.9263.4966.2262.80
ReLA [32]Swin-BBERT73.8276.4870.1866.0471.0257.6565.0065.9762.70
LQMFormer (ours)Swin-BBERT74.1676.8271.0465.9171.8457.5964.7366.0462.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.256, + 0.39, + 0.269 + ], + "angle": 0, + "content": "Table 2. Comparison on GRES dataset." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.27, + 0.44, + 0.41 + ], + "angle": 0, + "content": "
MethodsvalNo-target val
oIoUgIoUN-acc.T-acc.
MattNet [58]47.5148.2441.1596.13
LTS [24]52.3052.70--
VLT [12]52.5152.0047.1795.72
CRIS [49]55.3456.27--
LAVT [54]57.6458.4049.3296.18
ReLA [32]62.4263.6056.3796.32
LQMFormer (ours)64.9870.9467.4799.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.424, + 0.371, + 0.438 + ], + "angle": 0, + "content": "Table 3. Ablation Study Results" + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.45, + 0.489, + 0.622 + ], + "angle": 0, + "content": "
ConfigurationoIoUgIoU
(a) Analysis on Dynamic Query module
Ours64.9870.94
w/o score-based learning62.7567.16
w/o scoring module62.0265.48
w/o QM-loss64.1968.37
(b) Analysis on multi-modal fusion module
Ours64.9870.94
w/o modality fusion module and QM-loss49.6852.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.635, + 0.471, + 0.663 + ], + "angle": 0, + "content": "Table 4. Ablation study of Query Numbers \\( {N}_{qb} \\) in \\( {Q}_{b} \\) . \\( \\ddagger \\) : without scoring module. Top p% ratio is 20" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.674, + 0.437, + 0.768 + ], + "angle": 0, + "content": "
NqoIoUgIoUPr@0.7Pr@0.8Pr@0.9
2057.2258.3169.1560.2622.47
5062.4864.5372.0261.7931.58
10064.9870.9475.0365.5134.05
100‡62.0265.4873.1562.8731.55
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.839 + ], + "angle": 0, + "content": "between \"target\" and \"no-target\". The model's specificity (true accuracy) score of \\(99.12\\%\\) also reflects its effectiveness in classifying target samples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.231, + 0.864 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "In our comprehensive ablation studies, we evaluated the performance impact of various components of LQMFormer" + }, + { + "type": "table_caption", + "bbox": [ + 0.556, + 0.256, + 0.835, + 0.271 + ], + "angle": 0, + "content": "Table 5. Ablation study of Top p% ratio in \\( {Q}_{b} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.281, + 0.892, + 0.354 + ], + "angle": 0, + "content": "
p%oIoUgIoUN.acc.Pr@0.7Pr@0.8Pr@0.9
2064.9870.9467.4775.0365.5134.05
5063.8567.7663.2574.2964.4233.58
10061.3363.7758.9272.4561.8831.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.557, + 0.369, + 0.835, + 0.383 + ], + "angle": 0, + "content": "Table 6. Ablation study of Grounding Module" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.394, + 0.88, + 0.495 + ], + "angle": 0, + "content": "
FusionoIoUgIoUPr@0.7Pr@0.8Pr@0.9
Baseline49.6852.3455.2741.3814.72
Post-CA60.2261.8974.1564.4432.67
S-Post-CA [40]62.1364.4173.6463.3532.71
VL-Grounding63.4368.8172.9564.1933.96
GMMF64.9870.9475.0365.5134.05
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.596 + ], + "angle": 0, + "content": "on the GRES dataset. These studies focused on the Dynamic Query Module, Query Numbers \\(N_{qb}\\) in \\(Q_{b}\\), the Top \\(\\mathsf{p}\\%\\) ratio in \\(Q_{b}\\), and the Grounding Module. The results helps understanding key insights into the model's functioning and optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.893, + 0.718 + ], + "angle": 0, + "content": "Analysis on Dynamic Query Module: Table 3(a) presents the ablation results for the language query selection module. The removal of score-based learning significantly impacts scores as it limits the explicit training of Query scores from the scoring network. Also, notably, removing the scoring module and QM-loss leads to a decrease in oIoU score by \\(2.96\\%\\) and \\(0.79\\%\\), respectively, highlighting its critical role in achieving optimal referring segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Query Bank number \\( N_{qb} \\) in \\( Q_{b} \\): The ablation on Query Bank number demonstrated a notable impact on the model's performance as shown in Table 4. With an increase in \\( N_{q} \\) from 20 to 100, there was a consistent improvement in both IoU from 57.22% to 64.98% and gIoU from 58.31% to 70.94%, as well as in Precision metrics. Particularly, \\( N_{q} = 100 \\) yielded the highest scores across all metrics, indicating an optimal balance in query numbers for effective segmentation. Thus, it optimal to have more number of queries with Top p% as 20." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Top p% Ratio in \\(Q_{b}\\): The study on the Top p% ratio revealed that a 20% ratio achieved the best performance in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12909" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.1, + 0.191, + 0.116 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.118, + 0.287, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.1, + 0.418, + 0.116 + ], + "angle": 0, + "content": "ReLA" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.118, + 0.488, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.1, + 0.65, + 0.115 + ], + "angle": 0, + "content": "LQMFormer" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.118, + 0.691, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.1, + 0.856, + 0.116 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.118, + 0.891, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.264, + 0.78, + 0.283 + ], + "angle": 0, + "content": "Query: \"Orange cow with front facing us and an orange cow with back facing us\"" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.301, + 0.285, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.3, + 0.49, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.3, + 0.691, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.3, + 0.893, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.447, + 0.629, + 0.465 + ], + "angle": 0, + "content": "Query: \"The pizza in the middle and the guy holding a camera\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.484, + 0.779, + 0.499 + ], + "angle": 0, + "content": "Figure 5. Qualitave Comparison of our model with LQMFormer with ReLA [32] on GRES dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.611 + ], + "angle": 0, + "content": "terms of oIoU, gIoU, and Precision metrics(Table 5). Higher ratios, like \\(50\\%\\) and \\(100\\%\\), resulted in diminished performance as it resulted in more and more query collapse issues as many queries contribute towards the final mask generation. This finding indicates that a moderate selection of top-performing queries is crucial for balancing precision and recall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.474, + 0.902 + ], + "angle": 0, + "content": "Grounding Module: The ablation study of the Grounding Module, as shown in Table 3(b) and Table 6, highlights the effectiveness of Gaussian Enhancement for Referring Image Segmentation (RIS) tasks. The Gaussian enhanced Multi-Model Fusion (GMMF) method yields a improvement, specifically enhancing oIoU by \\(1.5\\%\\), from \\(63.43\\%\\) in VL-Grounding to \\(64.98\\%\\) in GMMF. This underlines the superior performance of Gaussian Filtering in the Fourier Domain over traditional cross-attention methods. Furthermore, compared to the Post-CA which registers an oIoU of \\(60.22\\%\\), the GMMF and VL-Grounding methods showcase substantial improvements in multi-modal representation within the backbone stage, with GMMF outperforming Post-CA by \\(4.76\\%\\) and VL-Grounding by \\(3.21\\%\\) in oIoU. The comparison with S-Post-CA [40], which aims at enhancing visual representation post-feature extraction, show the importance of integrating advanced grounding techniques early in the model's architecture for improved RIS performance. Overall, our proposed Gaussian enhanced Multi-Model Fusion" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.507, + 0.893, + 0.539 + ], + "angle": 0, + "content": "(GMMF) improves the visual grounding compared to other methods, resulting in a improved referring segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.568, + 0.619, + 0.584 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.896, + 0.901 + ], + "angle": 0, + "content": "In this paper, we address the problem of query collapse in Referring Image Segmentation by enhancing visual grounding and generating diverse query sets, and by modeling their relevance based on language expressions. Our proposed model, LQMFormer, incorporates Gaussian Enhanced Multi-Modal Fusion to improve visual grounding, and a Dynamic Query Module that not only generates a diverse set of queries but also employs a scoring network to selectively focus on queries based on given language expressions. Additionally, to mitigate query collapse, we introduce a novel loss function that enforces a margin of separation between query feature representations, facilitating improved representation without the need for additional supervision. The LQMFormer method notably surpasses state-of-the-art methods in both referring image segmentation and generalized referring image segmentation performance in most settings. For future work, we aim to extend the LQMFormer model's capabilities along with LLM and focus on effectively bridging LLM to better handle multimodal contexts." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12910" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.473, + 0.227 + ], + "angle": 0, + "content": "[2] Ankan Bansal, Yuting Zhang, and Rama Chellappa. Visual question answering on image sets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 51–67. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.472, + 0.258 + ], + "angle": 0, + "content": "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[4] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. arXiv preprint arXiv:2203.14713, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.304, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[5] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In ICCV, pages 7454-7463, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[6] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.472, + 0.487 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[8] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM Transactions on Graphics (ToG), 34(1):1-11, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.473, + 0.587 + ], + "angle": 0, + "content": "[9] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.715 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL-HLT, pages 4171–4186. Association for Computational Linguistics, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.759 + ], + "angle": 0, + "content": "[12] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, pages 16321-16330, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[13] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[15] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[17] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1330, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.893, + 0.262 + ], + "angle": 0, + "content": "[18] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.292 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.893, + 0.321 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.894, + 0.363 + ], + "angle": 0, + "content": "[21] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, pages 108-124. Springer, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[22] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In CVPR, pages 4424-4433, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.422, + 0.894, + 0.49 + ], + "angle": 0, + "content": "[23] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.493, + 0.893, + 0.533 + ], + "angle": 0, + "content": "[24] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In CVPR, pages 9858-9867, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[25] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In CVPR, pages 18145-18154, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.592, + 0.894, + 0.66 + ], + "angle": 0, + "content": "[26] Sangrok Lee, Jongseong Bae, and Ha Young Kim. Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11776-11785, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.662, + 0.894, + 0.717 + ], + "angle": 0, + "content": "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. arXiv preprint arXiv:2302.11831, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.893, + 0.788 + ], + "angle": 0, + "content": "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[29] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In CVPR, pages 5745-5753, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[31] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In ICCV, pages 1271-1280, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[32] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.26 + ], + "angle": 0, + "content": "[33] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In ICCV, pages 4856-4864, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.47, + 0.315 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.345, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[36] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10437-10446, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[37] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, pages 10034-10043, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.525 + ], + "angle": 0, + "content": "[38] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.47, + 0.607 + ], + "angle": 0, + "content": "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[40] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920–930, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[41] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, pages 792–807. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[42] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.79 + ], + "angle": 0, + "content": "[43] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[45] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, pages 38-54, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.146 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. pages 5998-6008, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[47] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[48] Yiyu Wang, Jungang Xu, and Yingfei Sun. End-to-end transformer based model for image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2585-2594, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.893, + 0.342 + ], + "angle": 0, + "content": "[49] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, pages 11686-11695, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.894, + 0.438 + ], + "angle": 0, + "content": "[50] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.894, + 0.509 + ], + "angle": 0, + "content": "[51] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14383-14392, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.893, + 0.564 + ], + "angle": 0, + "content": "[52] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.894, + 0.648 + ], + "angle": 0, + "content": "[53] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.893, + 0.704 + ], + "angle": 0, + "content": "[54] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In CVPR, pages 18155-18165, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.893, + 0.748 + ], + "angle": 0, + "content": "[55] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, pages 10502–10511, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[57] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[58] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12912" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.161 + ], + "angle": 0, + "content": "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[60] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In International Conference on Machine Learning, pages 42589-42601. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[61] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "12913" + } + ] +] \ No newline at end of file diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_origin.pdf b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c649f20b7770ba3c7569616134265b8bfc3d87e --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/c24a3bea-55f3-42b0-b8e6-0dc981e7639b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c8e4597b50b0663b15954dc19ca2b6bce74ba3af458ecac8945faee442684e1 +size 2974426 diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/full.md b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8ac7ec5ec05551e37e74f822cb3ad5340fad1085 --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/full.md @@ -0,0 +1,303 @@ +# LQMFormer: Language-aware Query Mask Transformer for Referring Image Segmentation + +Nisarg A. Shah Vibashan VS Vishal M. Patel Johns Hopkins University + +snisarg812@gmail.com, {vvishnu2, vpatel36}@jhu.edu + +# Abstract + +Referring Image Segmentation (RIS) aims to segment objects from an image based on a language description. Recent advancements have introduced transformer-based methods that leverage cross-modal dependencies, significantly enhancing performance in referring segmentation tasks. These methods are designed such that each query predicts different masks. However, RIS inherently requires a single-mask prediction, leading to a phenomenon known as Query Collapse, where all queries yield the same mask prediction. This reduces the generalization capability of the RIS model for complex or novel scenarios. To address this issue, we propose a Multi-modal Query Feature Fusion technique, characterized by two innovative designs: (1) Gaussian enhanced Multi-Modal Fusion, a novel visual grounding mechanism that enhances overall representation by extracting rich local visual information and global visual-linguistic relationships, and (2) A Dynamic Query Module that produces a diverse set of queries through a scoring network where the network selectively focuses on queries for objects referred to in the language description. Moreover, we show that including an auxiliary loss to increase the distance between mask representations of different queries further enhances performance and mitigates query collapse. Extensive experiments conducted on four benchmark datasets validate the effectiveness of our framework. + +# 1. Introduction + +Referring Image Segmentation (RIS) is a challenging multimodal task aimed at segmenting specific objects in an image based on a textual description. This description often includes details about the object's actions, category, color, or position, as noted by [8, 21]. Compared to traditional semantic and instance segmentation, RIS requires a precise understanding of object locations and involves a comprehensive modeling of the visual-linguistic relationships within the global context. Additionally, RIS demands the extrac + +![](images/32cc2895f818ab3817838db895ba94bc54ddf6b6a850a4f11431606300f57670.jpg) +(a) ReLA + +![](images/e4154ddd81b69b8383091a3da1398816b3faecc99aae154e1cdc3c5c858ec3c7.jpg) +(b) LQMFormer +Figure 1. [Top] Heatmap visualization of cosine similarity between different query mask predictions of ReLA and LQMFormer, where the yellow region indicates a similarity near 1, and the dark blue region signifies a similarity of 0. [Bottom] Sample query mask prediction visualizations for ReLA and LQMFormer. + +tion of high-quality visual features at a local level. The potential applications of RIS are extensive, particularly in language-driven human-computer interaction domains. + +Traditional RIS methods have employed linear fusion approaches and Fully Convolutional Networks (FCNs) for feature learning in RIS tasks [21, 33]. However, the recent advancement of attention mechanisms has shifted the focus towards extracting richer visual-language representation, with recent techniques demonstrating the effectiveness of Transformers. These models excel at modelling long-distance dependencies, making them suitable for cross-modal fusion [45, 55]. Further, Vision Transformer (ViT) methods such as VLT [13], EFN [16], and LAVT [54], which are based on [14], have shown significant improvements in RIS performance. Their ability to capture the nuances of cross-modal dependencies is crucial, and through a series of attention mechanisms, these models efficiently process and integrate both visual and language inputs for precise object segmentation. + +Despite their advantages, transformer-based approaches exhibit shortcomings in RIS mask prediction, with a notable problem being query collapse. This phenomenon occurs + +when different queries within the transformer, intended to identify distinct attributes or portions of objects, converge on the same mask prediction. This is particularly problematic in referring image segmentation, which inherently requires a prediction of a single mask, where all queries are collectively trained to predict the same mask. We experimentally observed this phenomenon, as illustrated in Fig. 1. Here, we visualize the heatmap of the cosine similarity between mask predictions from ReLA [32] and our method. In ReLA, most mask predictions are almost identical, leading to significantly overlapping mask predictions, whereas our method demonstrates a diverse set of query embeddings and mask predictions. Hence, prior transformer-based methods train mask predictions for all queries to associate with a single, specific ground truth mask. This training approach can lead to query collapse, where the model is not penalized for producing the same prediction across different queries, thereby affecting its capability to identify diverse attributes of the image. Consequently, this limitation restricts the variety of mask predictions, diminishing the model's ability to generalize effectively in complex scenarios. + +To address these challenges, we propose LQMFormer, which comprises two key components. Firstly, we introduce a Gaussian Enhanced Multi-Modal Fusion (GMMF) module, aimed at enhancing the visual grounding mechanism. This module is designed to extract fine-grained local visual details while simultaneously enhancing global visual-language relationships. The fusion of these modalities allows the model to construct a more comprehensive understanding of the scene, both globally and locally, thereby improving visual grounding. Secondly, in our Dynamic Query Module (DQM), we generate a diverse set of language-dependent queries through Language-Query Cross Attention (LQCA) and employ a scoring network to prioritize queries relevant to the referring expression. With an improved visual-grounded representation facilitated by GMMF and selective query enhancement via DQM, our model effectively addresses query collapse. Moreover, by incorporating an auxiliary loss that increases the representational distance between mask predictions, our method efficiently differentiates between queries, mitigating the issue of query collapse. Our extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our framework. Our contributions can be summarized as follows: + +- We introduce Gaussian enhanced Multi-Modal Fusion, which innovates the integration of local and global cross-modal information for more accurate visual grounding. +- We develop an effective Dynamic Query Module (DQM) that not only generates a diverse query set but also employs a scoring network to selectively focus on queries based on given language expressions with respect to the decoder's references. +- To mitigate query collapse, we propose an auxiliary regu + +larization loss function that increases the representational distance between queries, significantly improving mask differentiation and preventing Query Collapse. + +- We conduct comprehensive validation of the proposed framework with extensive testing on four benchmark datasets, demonstrating improvements in referring image segmentation performance. + +# 2. Related Works + +# 2.1. Referring Image Segmentation + +Early works [29, 31, 45, 57] employed Convolutional Neural Networks (CNNs) [5, 21, 22, 31] and Recurrent Neural Networks (RNNs) [6, 31] to extract vision and language features. These features were then fused via sequential concatenation and convolution operations to predict the segmentation mask. In contrast, [58] proposed a two-stage network that initially uses Mask R-CNN[19] to predict categorical masks, followed by a selection of the relevant masks based on language descriptions. However, this approach exhibits limited capacity to capture the intricate relations between language and visual content. The advent of the Transformer [11, 14, 42, 46] in the vision community has led to the widespread adoption of Transformer-based architectures for extracting both visual and textual features [11, 24, 34]. VLT [13] uses cross-attention to produce query vectors from multi-modal features, which are then used to query images in the transformer decoder. Similarly, LAVT [54] shows early cross-modal fusion of features improves alignment. CRIS [49] adopts Transformer blocks to leverage the pretrained CLIP model's [44] robust image-text alignment capabilities. Additionally, GRES [32] extends cross-attention in the Transformer decoder to explicitly model visual-language dependencies. However, these all transformer-based approaches require matching queries to corresponding mask instances. In the context of referring expression segmentation, where typically only a single mask is available, this can lead to all queries converging to a single mask, a phenomenon known as query collapse. Motivated by this, in our method, we try to produce a diverse set of query prediction by overall improving visual-grounding and conditioning with respect to language expression. + +# 2.2. Vision and Language Representation Learning + +Vision and Language Representation Learning focuses on understanding the semantics and alignment between vision and language for multimodal reasoning tasks [36, 47, 59]. This field has seen substantial progress in applications like visual question answering [2], Image captioning [48], Image-text retrieval [4], zero-shot classification [44], and Image segmentation [23]. Contrastive pre-training strategies [28, 44] using large-scale datasets effectively project multiple modalities into a single embedding space. In contrast, as discussed, + +other methods [32, 54] create cross-modal interaction layers to fuse and comprehend multimodal features. The recent adoption of deep learning techniques in the frequency domain [9, 17, 39, 40] has gained attention for their global interaction capabilities. Specifically, [40] performs spectral-guided enhancement of the multi-scale vision-language features after feature extraction stage for Video Referring Segmentation. Drawing from these advancements, our approach integrates gaussian guidance within vision-language representation learning to facilitate enhanced local and global multi-modal interactions. + +# 3. Methods + +Given an input image $I \in \mathbb{R}^{H \times W \times 3}$ and a language expression $L = \{w_{i}\}_{i=1}^{N}$ with $N$ words, our model predicts a pixel-wise mask $M$ , which delineating the referred object. + +# 3.1. Overview + +An overview diagram of our approach is shown in Figure 2. First, the language expression is encoded to extract high-dimensional language features $F_{l} \in \mathbb{R}^{N_{t} \times C}$ , where $C, N_{t}$ indicates the number of channels, and words in the language expression, respectively. The Gaussian-enhanced Multi-Modal Fusion (GMMF) module then extracts language-grounded visual features $F_{\mathrm{vl}}$ . Following [7], our approach incorporates a multi-scale strategy to exploit higher-resolution feature maps in the transformer decoder [7, 61]. We input language-aware queries $Q_{b}^{\prime}$ along with multi-scale outputs from the pixel decoder $F_{i}$ , where $i \in \left\{\frac{1}{4}, \frac{1}{8}, \frac{1}{16}, \frac{1}{32}\right\}$ of the original image. The Dynamic Query Module processes language features $F_{t}$ to predict language-aware queries $Q_{b}^{\prime}$ . The updating of queries $Q_{b}^{\prime}$ sequentially utilizes resolutions $\frac{1}{8}, \frac{1}{16}$ , and $\frac{1}{32}$ , applying a cycle of masked cross-attention (CA), self-attention (SA), and feed-forward network (FFN) operations $L$ times within the decoder. Further, we map the final query outputs to an 'object' or 'no-object' two-dimensional space to enable 'no-object' predictions [32]. The final predicted mask is obtained at full resolution as the original image by decoding the pixel features $F_{\frac{1}{4}}$ with the einsum operation between $Q$ and $F_{\frac{1}{4}}$ , followed by upsampling operation. + +# 3.2. Gaussian-enhanced Multi-Modal Fusion + +After extracting language features $F_{l}$ , we combine them with visual features using a four-stage hierarchical Swin Transformer [34] to obtain joint visual-language embedding. Each stage comprises Transformer layers $(\tau_{i})$ , multi-modal feature fusion modules (GMMF), and residual gating $(\psi_{i})$ . At every stage, as Transformer layers $(\tau_{i})$ enhance previous visual features $(F_{v})$ , these features $(F_{v})$ combine with language features $(F_{l})$ via multi-modal modules (GMMF) to create multi-modal features $(F_{vl})$ . Finally, we use gating units $(\psi_{vl})$ to weight and combine $F_{vl}$ with $F_{v}$ , yielding enhanced visual features with linguistic information. Residual gating + +also works well with pre-trained (only vision) transformer weights and allows simultaneous feature modulation with language without having to re-train from scratch. + +# 3.2.1 Gaussian Enhanced Multi-Modal Fusion Module in Visual Backbone + +The discrete Fourier transform (DFT) is extensively used to analyze the frequency representation of images. By transforming the signal into the frequency domain, characterized by global statistical properties, it plays a crucial role in various computer vision tasks [26, 51, 56]. Modulating signals on a point-wise basis in the Fourier domain alters the representation of inputs in the spatial domain [3, 40]. Furthermore, in the Fourier domain, low-frequency components usually correspond to the overall semantic content of the image, consistent with prior research [40, 52, 53, 56]. The Fourier transform is capable of disentangling global semantic information and structural coherence, with the semantic content predominantly found in the amplitude component [18, 27, 60]. These insights can be leveraged to create modules centered on Fourier-based methods, facilitating efficient and vital global interactions in multimodal understanding and improving visual grounding. + +To inculcate the above observation, as illustrated in Fig. 3, following [40], the Gaussian enhanced Multi-Modal Fusion Module (GMMF) is proposed as a key component in our model. It employs an enhancement block after cross-attention operations to combine Gaussian-enhanced vision and language features. Given an input vision feature $F_{vi} \in \mathbb{R}^{N \times H \times W \times C_{vi}}$ from stage $i$ and a language feature $F_{l} \in \mathbb{R}^{N \times T \times C_{l}}$ , for cross-attention block, we first transform visual and language features into a common dimension $C_{i}$ using separate $1 \times 1$ convolution layers. Subsequently, transform vision and language features are passed to an MHCA (Multi-Head Cross-Attention) module, where language features act as queries and vision features as keys and values. This operation results in the creation of a language-aware vision-shaped feature $F_{vl}'$ . We then transform the visual-language features using the Gaussian Enhancement Block, as shown in Fig. 3, formulated as follows: + +$$ +\begin{array}{l} F _ {v l} = F _ {\mathrm {F F T}} ^ {- 1} \left(\operatorname {C o n v} \left(F R \left(\phi \left(F _ {v l} ^ {\prime}, \beta\right) * \mathcal {A} \left(F _ {v l} ^ {\prime}\right) \right. \right. \right. \\ , \left. \left. \mathcal {P} \left(F _ {v l} ^ {\prime}\right)\right)\right) + F _ {v l} ^ {\prime} \\ \end{array} +$$ + +Here, $\mathcal{A}(F_{vl}^{\prime})$ and $\mathcal{P}(F_{vl}^{\prime})$ denote the amplitude and phase components of $F_{vl}^{\prime}$ 's Fast Fourier Transform(FFT) [15, 43], $F_{\mathrm{FFT}}(F_{vl}^{\prime})$ . The * represents low-pass filtering on $\mathcal{A}(F_{vl}^{\prime})$ using adaptive Gaussian filters $\phi (F_{vl}^{\prime},\beta)$ , matching $F_{vl}^{\prime}$ 's spatial dimensions and adjusting to input data via bandwidth $\beta$ [40]. $FR$ and $Conv$ signify feature reconstruction from amplitude and phase, and convolution operations, respectively. + +![](images/c5c2ea4b61ac346f4179f924ed36da4e9a8ba10514737dbc809b272401ede617.jpg) +Figure 2. Overview of the LQMFormer model architecture. The model processes an image and a linguistic expression, extracting vision-language features $F_{vl}$ through a Gaussian-enhanced Multi-modal Fusion module and language features $F_{L}$ . The multi-scale vision-language features $F_{vl}$ are then processed by into a pixel decoder. Concurrently, the Language Query Cross-Attention Module (LQCA) refines the Language-enhanced Query bank $Q_{b}^{\prime}$ with $F_{L}$ . A Scoring Network assesses the relevance of each query in $Q_{b}^{\prime}$ by predicting soft-masks $(M_{b})$ , which then modulate the queries fed into the transformer decoder. Finally, the model employs a global pooling layer and a Multi-Layer Perceptron (MLP) to predict the binary segmentation mask and determine the presence or absence of the referred object. + +![](images/3c84351b775a6b875a199450fbcc9f174e9c9f024bd3758aabcdd1b397e5ebf6.jpg) +Figure 3. Architecture overview of the Gaussian Enhancement Module + +Once $F_{vl}$ is obtained, similar to [54], we combine the output from GMMF, $F_{vl}$ , with that from the Transformer layers, $F_{v}$ . This process employs a gating mechanism, $\psi_{vl}$ , which learns a set of element-wise weight maps from $F_{vl}$ to finely tune the scale of each element within $F_{vl}$ . Subsequent to this adjustment, a residual combination of $F_{v}$ and $F_{vl} \cdot \psi_{vl}$ is performed. The resultant output is then fed back into the vision backbone for further computation. + +# 3.3. Dynamic Query Module + +The Dynamic Query Module (DQM) takes the language feature $F_{l}$ and a Query bank $Q_{b}$ as inputs, the latter containing $N_{qb}$ learnable queries. Figure 4(b) shows the initial + +step, where the attention between the language feature $F_{l}$ and the $N_{qb}$ query embeddings $Q_{b} \in \mathbb{R}^{N_{qb} \times C}$ is computed, resulting in $N_{qb}$ attention maps: + +$$ +A _ {b i} = \operatorname {s o f t m a x} \left(Q _ {b} \sigma \left(F _ {l} W _ {i k}\right) ^ {T}\right), \tag {1} +$$ + +where $W_{ik}$ is a $C\times C$ matrix of learnable parameters, and $\sigma$ denotes the GeLU function [20]. The resulting $A_{bi}\in \mathbb{R}^{N_{qb}\times L}$ provides each query with a $1\times L$ attention map, indicating its important linguistic relation in the language features. Following this, the language-supported query bank features are derived using these attention maps: $Q_b^{\prime} = A_{bi}\sigma (F_lW_{iv})^T$ , where $W_{iv}$ is another $C\times C$ matrix of learnable parameters. + +The variability in language expressions used to describe objects in images requires an approach that can adapt to different descriptions. This is because standard methods, like those seen in the initial transformer models (like in vanilla transformer [14]), struggle to capture the nuances in descriptions that include color, size, shape, and position. To address this, a larger set of adaptable queries is required. These queries are fine-tuned based on the input language, resulting in a refined set of queries, denoted as $Q_b'$ . Each query in this refined set is then assessed with a scoring network, which outputs a soft-mask $M_b$ . This soft-mask represents the relevance of each query. The final set of queries, $Q_{\mathrm{final}}$ , is obtained by combining $M_b$ with $Q_b'$ through a dot product, as shown in the equation below: + +$$ +Q _ {\text {f i n a l}} = M _ {b} ^ {\prime} \odot Q _ {b} ^ {\prime} \tag {2} +$$ + +![](images/498463b200decfac3a2c4470079c807eead59ff69518d3b971682bddfcbde74c.jpg) +Figure 4. Architecture overview of the (A.) Dynamic Query Module and (B.) LQCA Block + +![](images/46078c28fb1381519120e8b9102c5b752bbb0245513843bb8d6012d442baf33c.jpg) + +# 3.3.1 Scoring Network + +The scoring network in LQMFormer takes the Language-enhanced Query bank $Q_{b}^{\prime}$ as input and processes it through a sequence of linear layers. This network consists of four linear layers, with Layer Normalization [1] applied before the first layer. Each subsequent layer, except the last sigmoid, incorporates GELU [20] activation. The final output is a 1-dimensional logit $M_{b} \in \mathbb{R}^{N_{qb} \times 1}$ , serving as a soft-mask to modulate $Q_{b}^{\prime}$ . + +To train the scoring network for predicting a relevant soft-mask $M_{b}$ effectively for the transformer decoder, we utilize its cross-attention map. This map is essential in identifying the subset of Queries most relevant to the referring expression and visual features, as highlighted during training. We sum up the attention maps from each decoder layer using Bilinear Interpolation (BI), thereby projecting attention $A_{qf} \in \mathbb{R}^{N_{qb} \times HW}$ directed towards these Queries. The process is mathematically described as follows: + +$$ +A _ {q f} = \sum_ {l = 1} ^ {L} \sum_ {r = 1} ^ {3} \mathrm {B I} \left(\mathrm {A} _ {l, r}, H W\right) \tag {3} +$$ + +Here, $\mathbf{A}_{l,r} \in \mathbb{R}^{N_{qb} \times hw}$ represents the attention map from the $l^{th}$ layer at the $r^{th}$ resolution, with $hw$ being the original dimension and $HW$ the desired output dimension. After aligning each attention map to the common resolution $HW$ , we then sum across this dimension to generate a likelihood map $LM$ , as indicated in the subsequent equation: + +$$ +L M = \sum_ {i = 1} ^ {H W} A _ {q f} [:, i ] \tag {4} +$$ + +This likelihood map $\mathrm{LM} \in \mathbb{R}^{N_{qb} \times 1}$ , now captures the concentrated attention across all Queries, serves as a condition in refining the predictions of the scoring network. The + +scoring network is trained to predict the likelihood of each Query being included in the top- $\rho\%$ of the most referenced tokens. This process involves binarizing the likelihood map $LM$ to retain only the top- $\rho\%$ Queries, creating a binarized version $LM_{\mathrm{bin}}$ . The training objective is then formulated as a Binary Cross-Entropy (BCE) loss between the predicted soft-mask $M_b$ and the binarized likelihood map $LM_{\mathrm{bin}}$ . The BCE loss function is defined as: + +$$ +\begin{array}{l} \mathrm {B C E} \operatorname {L o s s} = - \frac {1}{N _ {q b}} \sum_ {i = 1} ^ {N _ {q b}} \left[ L M _ {\mathrm {b i n}, i} \log \left(M _ {b, i}\right) \right. \\ + \left(1 - L M _ {\text {b i n}, i}\right) \log \left(1 - M _ {b, i}\right) ] \tag {5} \\ \end{array} +$$ + +Here, $N_{qb}$ represents the number of Queries in $Q_b^{\prime}$ , and $i$ indexes each Query. This loss function effectively trains the scoring network to align its predictions with the most salient Queries as determined by the decoder's cross-attention, thereby refining the model's accuracy in identifying and segmenting objects as described by the referring expressions. + +# 3.4. Query-Mask Margin Loss + +Query-Mask Margin Loss is designed to maintain a margin of separability between different query feature representations, an essential aspect for tasks with variable outputs. The loss function operates by initially computing pairwise differences and distances between mask embeddings within a batch. This computation yields a tensor of shape $B \times N \times N$ , where $B$ represents the batch size, and $N$ is the number of queries. Subsequently, these distances for each embedding are sorted, with a particular focus on the second smallest value, which signifies the distance to the nearest neighbor. The margin loss is defined through the max function, ensuring a minimum distance of 1.0 between embeddings. The + +mathematical formulation of QM-loss is as follows: + +$$ +\mathrm {Q M - l o s s} = \max (0, 1. 0 - \operatorname {d i s t} \left(M _ {i}, M _ {j}\right)) +$$ + +Here, $\text{dist}(M_i, M_j)$ denotes the calculation of pairwise distances between the mask embeddings, represented as $M$ . + +# 4. Experiments and Discussion + +# 4.1. Datasets and Implementation Details + +Datasets: Our experiments are conducted on four primary benchmark datasets in Referring Image Segmentation: RefCOCO [57], RefCOCO+ [57], RefCOCOg [38], and GRES [32]. These datasets derive their images from MSCOCO [30]. The dataset details are as follows: RefCOCO (19,994 images, 50,000 objects, 142,209 expressions), RefCOCO+ (19,992 images, 49,856 objects, 141,564 expressions), RefCOCOg (26,711 images, 54,822 objects, 85,474 expressions), and GRES (19,994 images, 60,287 objects, 278,232 expressions). RefCOCO mainly focuses on expressions that specify object locations, while RefCOCO+ prioritizes object descriptions. RefCOCOg presents a higher challenge due to longer and more complex expressions (averaging 8.4 words compared to 3.5 in others). GRES broadens the scope of RefCOCO by including expressions that refer to multiple objects or no objects, thereby enhancing the problem's generalizability. We use both UMD split [41] and Google split [38] for RefCOCOg dataset. + +Evaluation Metrics. For single-target object segmentation, we report our results using two kinds of metrics [13, 37, 49, 54]: overall IoU (oIoU) and Precision@X (P@X). oIoU divides the total intersection pixels by the total union pixels across all test images. Precision@X calculates the percentage of testing samples of which the model prediction has an IoU score higher than the threshold value X. To extend our evaluation to non-target and multi-target object segmentation, following [32], we include additional metrics: Sensitivity (N-acc.), Specificity (T-acc.), and Generalized IoU (gIoU). N-acc. evaluates the model's performance on identifying non-target samples, while T-acc. measures the impact of generalization on non-target samples on the performance of target samples. We opt for gIoU over oIoU due to the latter's bias towards larger objects. gIoU computes the mean IoU per image for all samples. In cases with no-target samples, IoU values are assigned as 1 for true positives and 0 for false negatives, ensuring an equitable metric for all object sizes. + +Implementation Details. Our model uses Swin Transformer-B [34] as the visual encoder and BERT [11] as the language decoder. The transformer layer in Swin-B is initialized with classification weights pre-trained on ImageNet22K [10]. We use BERT implementation from the HuggingFace Transformer library [50]. It comprises 12 layers with a hidden size of 768, initialized using official + +pre-trained weights. We use our Transformer decoder proposed in Section 3.1 with $\mathrm{L} = 3$ (i.e., 9 layers total) and 100 queries by default. The AdamW optimizer [35] is employed for optimization with an initial learning rate of 0.00001. Our model is trained over 100K iterations with a batch size of 48. Our image resizing is standardized to $480 \times 480$ pixels. + +# 4.2. Comparison with the state of art methods + +Referring Image Segmentation Table 1 presents the results for Referring Image Segmentation, and our proposed method, LQMFormer, produces competitive performance against existing state-of-the-art approaches across multiple standard benchmarks. On the RefCOCO dataset, our proposed method obtains the highest IoU across all divisions (val, test A, and test B), surpassing recent leading methods such as VLT [13] and ReLA [32]. Notably, it shows a significant margin in the test B split, indicating enhanced capability in complex visual contexts. In the RefCOCO+ dataset, LQMFormer also demonstrates its effectiveness in the test A set. Specifically, in the test A split with a cIoU of $71.84\%$ . In the Val and Test B splits, our proposed method achieves competitive performance against ReLA [32] and LAVT [54]. For the G-Ref dataset, which presents varied and longer referring expressions, our proposed method achieves the top performance in the $\mathrm{val}_{(\mathrm{G})}$ split and maintains competitive results in the other splits. This performance is particularly notable over LAVT [54] in the $\mathrm{val}_{(\mathrm{G})}$ split, highlighting its adaptability to diverse challenging datasets. While the margin of improvement offered by LQMFormer in classic RIS datasets may be comparatively smaller than that observed in GREs, the results signify that Gaussian-enhanced Multi-Modal Fusion and explicit modeling of Queries from Decoder are beneficial for the general Referring Image Segmentation setting. + +Generalised Referring Image Segmentation In the GRES dataset, the LQMFormer model demonstrates notable improvements in Region Instance Segmentation (RIS), as shown in Table 2. The model achieves a generalized Intersection over Union (IoU) score of $70.94\%$ on the validation set, surpassing other models such as ReLA, LAVT, and CRIS by a margin of $7 - 12\%$ . This performance indicates its effective segmentation across various object sizes. The object-level Intersection over Union (oIoU) score of $64.98\%$ further underscores the model's segmentation capabilities, highlighting the effectiveness of the Improved Visual Grounding and Dynamic Query Selection module. Another important aspect of LQMFormer is its performance in no-target identification, with an accuracy score of $67.47\%$ . This score is $11\%$ higher than that of existing methods, marking significant progress in accurately identifying no-target samples—a noted challenge in RIS. The inclusion of the Dynamic Query Module (DQM) and the Top $p\%$ ratio significantly reduces the impact of the number of queries and mitigates query collapse. As a result, fewer but more effective queries make it easier to distinguish + +Table 1. Results on classic RES in terms of oIoU. U: UMD split. G: Google split. + +
MethodsVisual EncoderTextual EncoderRefCOCORefCOCO+val(U)G-Ref test(U)val(G)
valtest Atest Bvaltest Atest B
MCN [37]Darknet53bi-GRU62.4464.2059.7150.6254.9944.6949.2249.40-
VLT [12]Darknet53bi-GRU67.5270.4765.2456.3060.9850.0854.9657.7352.02
ReSTR [25]ViT-BTransformer67.2269.3064.4555.7860.4448.27--54.48
CRIS [49]CLIP-R101CLIP70.4773.1866.1062.2768.0853.6859.8760.36-
LAVT [54]Swin-BBERT72.7375.8268.7962.1468.3855.1061.2462.0960.50
VLT [13]Swin-BBERT72.9675.9669.6063.5368.4356.9263.4966.2262.80
ReLA [32]Swin-BBERT73.8276.4870.1866.0471.0257.6565.0065.9762.70
LQMFormer (ours)Swin-BBERT74.1676.8271.0465.9171.8457.5964.7366.0462.97
+ +Table 2. Comparison on GRES dataset. + +
MethodsvalNo-target val
oIoUgIoUN-acc.T-acc.
MattNet [58]47.5148.2441.1596.13
LTS [24]52.3052.70--
VLT [12]52.5152.0047.1795.72
CRIS [49]55.3456.27--
LAVT [54]57.6458.4049.3296.18
ReLA [32]62.4263.6056.3796.32
LQMFormer (ours)64.9870.9467.4799.12
+ +Table 3. Ablation Study Results + +
ConfigurationoIoUgIoU
(a) Analysis on Dynamic Query module
Ours64.9870.94
w/o score-based learning62.7567.16
w/o scoring module62.0265.48
w/o QM-loss64.1968.37
(b) Analysis on multi-modal fusion module
Ours64.9870.94
w/o modality fusion module and QM-loss49.6852.34
+ +Table 4. Ablation study of Query Numbers ${N}_{qb}$ in ${Q}_{b}$ . $\ddagger$ : without scoring module. Top p% ratio is 20 + +
NqoIoUgIoUPr@0.7Pr@0.8Pr@0.9
2057.2258.3169.1560.2622.47
5062.4864.5372.0261.7931.58
10064.9870.9475.0365.5134.05
100‡62.0265.4873.1562.8731.55
+ +between "target" and "no-target". The model's specificity (true accuracy) score of $99.12\%$ also reflects its effectiveness in classifying target samples. + +# 4.3. Ablation Study + +In our comprehensive ablation studies, we evaluated the performance impact of various components of LQMFormer + +Table 5. Ablation study of Top p% ratio in ${Q}_{b}$ . + +
p%oIoUgIoUN.acc.Pr@0.7Pr@0.8Pr@0.9
2064.9870.9467.4775.0365.5134.05
5063.8567.7663.2574.2964.4233.58
10061.3363.7758.9272.4561.8831.21
+ +Table 6. Ablation study of Grounding Module + +
FusionoIoUgIoUPr@0.7Pr@0.8Pr@0.9
Baseline49.6852.3455.2741.3814.72
Post-CA60.2261.8974.1564.4432.67
S-Post-CA [40]62.1364.4173.6463.3532.71
VL-Grounding63.4368.8172.9564.1933.96
GMMF64.9870.9475.0365.5134.05
+ +on the GRES dataset. These studies focused on the Dynamic Query Module, Query Numbers $N_{qb}$ in $Q_{b}$ , the Top $\mathsf{p}\%$ ratio in $Q_{b}$ , and the Grounding Module. The results helps understanding key insights into the model's functioning and optimization. + +Analysis on Dynamic Query Module: Table 3(a) presents the ablation results for the language query selection module. The removal of score-based learning significantly impacts scores as it limits the explicit training of Query scores from the scoring network. Also, notably, removing the scoring module and QM-loss leads to a decrease in oIoU score by $2.96\%$ and $0.79\%$ , respectively, highlighting its critical role in achieving optimal referring segmentation performance. + +Query Bank number $N_{qb}$ in $Q_{b}$ : The ablation on Query Bank number demonstrated a notable impact on the model's performance as shown in Table 4. With an increase in $N_{q}$ from 20 to 100, there was a consistent improvement in both IoU from 57.22% to 64.98% and gIoU from 58.31% to 70.94%, as well as in Precision metrics. Particularly, $N_{q} = 100$ yielded the highest scores across all metrics, indicating an optimal balance in query numbers for effective segmentation. Thus, it optimal to have more number of queries with Top p% as 20. + +Top p% Ratio in $Q_{b}$ : The study on the Top p% ratio revealed that a 20% ratio achieved the best performance in + +Input +![](images/c99725aeceedbe99d02f67028ca126037faa12ecd804958cd824f88c5444f2a0.jpg) +Query: "Orange cow with front facing us and an orange cow with back facing us" + +![](images/710ec28d0f6087f1f95e81d9a7f18d23d81c57275ff7db02688a4b8340b1bc9c.jpg) +ReLA + +![](images/6bbae835706fe86cda15df2823a351e795f182da57fcf295385d200680f6f055.jpg) +LQMFormer + +![](images/042944f6b5ffade7820e6adcd8845713c1ad98c0949ed153f26f58880014a2f3.jpg) +Ground Truth + +Figure 5. Qualitave Comparison of our model with LQMFormer with ReLA [32] on GRES dataset. +![](images/6c0f1a83cf9a3602ca6ee537da02f1f9ee6e7d88e51b7ddeccd20cde2ee5bdb1.jpg) +Query: "The pizza in the middle and the guy holding a camera" + +![](images/1aaecd4830aa26d2c362a65435b50ae3b9ce9bf2f3382b9cc4eada973fa6e095.jpg) + +![](images/9b3264f5163287e93fd3249ff0231101f8157415e7338998b54c20ae4b555136.jpg) + +![](images/2403e8bb106cf2cb68033e8bf81ff97f161dc2f63d965a6d82e56d02c0393372.jpg) + +terms of oIoU, gIoU, and Precision metrics(Table 5). Higher ratios, like $50\%$ and $100\%$ , resulted in diminished performance as it resulted in more and more query collapse issues as many queries contribute towards the final mask generation. This finding indicates that a moderate selection of top-performing queries is crucial for balancing precision and recall. + +Grounding Module: The ablation study of the Grounding Module, as shown in Table 3(b) and Table 6, highlights the effectiveness of Gaussian Enhancement for Referring Image Segmentation (RIS) tasks. The Gaussian enhanced Multi-Model Fusion (GMMF) method yields a improvement, specifically enhancing oIoU by $1.5\%$ , from $63.43\%$ in VL-Grounding to $64.98\%$ in GMMF. This underlines the superior performance of Gaussian Filtering in the Fourier Domain over traditional cross-attention methods. Furthermore, compared to the Post-CA which registers an oIoU of $60.22\%$ , the GMMF and VL-Grounding methods showcase substantial improvements in multi-modal representation within the backbone stage, with GMMF outperforming Post-CA by $4.76\%$ and VL-Grounding by $3.21\%$ in oIoU. The comparison with S-Post-CA [40], which aims at enhancing visual representation post-feature extraction, show the importance of integrating advanced grounding techniques early in the model's architecture for improved RIS performance. Overall, our proposed Gaussian enhanced Multi-Model Fusion + +(GMMF) improves the visual grounding compared to other methods, resulting in a improved referring segmentation. + +# 5. Conclusion + +In this paper, we address the problem of query collapse in Referring Image Segmentation by enhancing visual grounding and generating diverse query sets, and by modeling their relevance based on language expressions. Our proposed model, LQMFormer, incorporates Gaussian Enhanced Multi-Modal Fusion to improve visual grounding, and a Dynamic Query Module that not only generates a diverse set of queries but also employs a scoring network to selectively focus on queries based on given language expressions. Additionally, to mitigate query collapse, we introduce a novel loss function that enforces a margin of separation between query feature representations, facilitating improved representation without the need for additional supervision. The LQMFormer method notably surpasses state-of-the-art methods in both referring image segmentation and generalized referring image segmentation performance in most settings. For future work, we aim to extend the LQMFormer model's capabilities along with LLM and focus on effectively bridging LLM to better handle multimodal contexts. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Ankan Bansal, Yuting Zhang, and Rama Chellappa. Visual question answering on image sets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 51–67. Springer, 2020. 2 +[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3 +[4] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. arXiv preprint arXiv:2203.14713, 2022. 2 +[5] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In ICCV, pages 7454-7463, 2019. 2 +[6] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 2 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3 +[8] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM Transactions on Graphics (ToG), 34(1):1-11, 2014. 1 +[9] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL-HLT, pages 4171–4186. Association for Computational Linguistics, 2019. 2, 6 +[12] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, pages 16321-16330, 2021. 7 +[13] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 6, 7 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 4 + +[15] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 3 +[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In CVPR, 2021. 1 +[17] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1330, 2021. 3 +[18] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 3 +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2 +[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4, 5 +[21] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, pages 108-124. Springer, 2016. 1, 2 +[22] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In CVPR, pages 4424-4433, 2020. 2 +[23] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2 +[24] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In CVPR, pages 9858-9867, 2021. 2, 7 +[25] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In CVPR, pages 18145-18154, 2022. 7 +[26] Sangrok Lee, Jongseong Bae, and Ha Young Kim. Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11776-11785, 2023. 3 +[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. arXiv preprint arXiv:2302.11831, 2023. 3 +[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2 +[29] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In CVPR, pages 5745-5753, 2018. 2 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, + +Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[31] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In ICCV, pages 1271-1280, 2017. 2 +[32] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 3, 6, 7, 8 +[33] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In ICCV, pages 4856-4864, 2017. 1 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3, 6 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[36] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10437-10446, 2020. 2 +[37] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, pages 10034-10043, 2020. 6, 7 +[38] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 6 +[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3 +[40] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920–930, 2023. 3, 7, 8 +[41] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, pages 792–807. Springer, 2016. 6 +[42] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 2 +[43] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 3 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 +[45] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, pages 38-54, 2018. 1, 2 + +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. pages 5998-6008, 2017. 2 +[47] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2 +[48] Yiyu Wang, Jungang Xu, and Yingfei Sun. End-to-end transformer based model for image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2585-2594, 2022. 2 +[49] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, pages 11686-11695, 2022. 2, 6, 7 +[50] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. 6 +[51] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14383-14392, 2021. 3 +[52] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3 +[53] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3 +[54] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In CVPR, pages 18155-18165, 2022. 1, 2, 3, 4, 6, 7 +[55] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, pages 10502–10511, 2019. 1 +[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 3 +[57] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 2, 6 +[58] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 2, 7 + +[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 2 +[60] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In International Conference on Machine Learning, pages 42589-42601. PMLR, 2023. 3 +[61] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3 \ No newline at end of file diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/images.zip b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bed498baef6ac3073a71f11aec186567c8e3cfaa --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e56c0252f333ed3e2f45d19372235f2502fa212d24e72dd2f289356c6990c537 +size 657330 diff --git a/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/layout.json b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a5c3601b0e6eb3391c27dd023b1284aed5ecefed --- /dev/null +++ b/2024/LQMFormer_ Language-aware Query Mask Transformer for Referring Image Segmentation/layout.json @@ -0,0 +1,9726 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 444, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 444, + 140 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 444, + 140 + ], + "type": "text", + "content": "LQMFormer: Language-aware Query Mask Transformer for Referring Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 161, + 429, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 161, + 429, + 190 + ], + "spans": [ + { + "bbox": [ + 162, + 161, + 429, + 190 + ], + "type": "text", + "content": "Nisarg A. Shah Vibashan VS Vishal M. Patel Johns Hopkins University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 191, + 429, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 191, + 429, + 203 + ], + "spans": [ + { + "bbox": [ + 160, + 191, + 429, + 203 + ], + "type": "text", + "content": "snisarg812@gmail.com, {vvishnu2, vpatel36}@jhu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 558 + ], + "type": "text", + "content": "Referring Image Segmentation (RIS) aims to segment objects from an image based on a language description. Recent advancements have introduced transformer-based methods that leverage cross-modal dependencies, significantly enhancing performance in referring segmentation tasks. These methods are designed such that each query predicts different masks. However, RIS inherently requires a single-mask prediction, leading to a phenomenon known as Query Collapse, where all queries yield the same mask prediction. This reduces the generalization capability of the RIS model for complex or novel scenarios. To address this issue, we propose a Multi-modal Query Feature Fusion technique, characterized by two innovative designs: (1) Gaussian enhanced Multi-Modal Fusion, a novel visual grounding mechanism that enhances overall representation by extracting rich local visual information and global visual-linguistic relationships, and (2) A Dynamic Query Module that produces a diverse set of queries through a scoring network where the network selectively focuses on queries for objects referred to in the language description. Moreover, we show that including an auxiliary loss to increase the distance between mask representations of different queries further enhances performance and mitigates query collapse. Extensive experiments conducted on four benchmark datasets validate the effectiveness of our framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 128, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Referring Image Segmentation (RIS) is a challenging multimodal task aimed at segmenting specific objects in an image based on a textual description. This description often includes details about the object's actions, category, color, or position, as noted by [8, 21]. Compared to traditional semantic and instance segmentation, RIS requires a precise understanding of object locations and involves a comprehensive modeling of the visual-linguistic relationships within the global context. Additionally, RIS demands the extrac" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 226, + 426, + 334 + ], + "blocks": [ + { + "bbox": [ + 329, + 226, + 426, + 334 + ], + "lines": [ + { + "bbox": [ + 329, + 226, + 426, + 334 + ], + "spans": [ + { + "bbox": [ + 329, + 226, + 426, + 334 + ], + "type": "image", + "image_path": "32cc2895f818ab3817838db895ba94bc54ddf6b6a850a4f11431606300f57670.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 341, + 389, + 348 + ], + "lines": [ + { + "bbox": [ + 364, + 341, + 389, + 348 + ], + "spans": [ + { + "bbox": [ + 364, + 341, + 389, + 348 + ], + "type": "text", + "content": "(a) ReLA" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 429, + 228, + 528, + 339 + ], + "blocks": [ + { + "bbox": [ + 429, + 228, + 528, + 339 + ], + "lines": [ + { + "bbox": [ + 429, + 228, + 528, + 339 + ], + "spans": [ + { + "bbox": [ + 429, + 228, + 528, + 339 + ], + "type": "image", + "image_path": "e4154ddd81b69b8383091a3da1398816b3faecc99aae154e1cdc3c5c858ec3c7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 342, + 509, + 350 + ], + "lines": [ + { + "bbox": [ + 469, + 342, + 509, + 350 + ], + "spans": [ + { + "bbox": [ + 469, + 342, + 509, + 350 + ], + "type": "text", + "content": "(b) LQMFormer" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 370, + 547, + 426 + ], + "lines": [ + { + "bbox": [ + 304, + 370, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 547, + 426 + ], + "type": "text", + "content": "Figure 1. [Top] Heatmap visualization of cosine similarity between different query mask predictions of ReLA and LQMFormer, where the yellow region indicates a similarity near 1, and the dark blue region signifies a similarity of 0. [Bottom] Sample query mask prediction visualizations for ReLA and LQMFormer." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 449, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 484 + ], + "type": "text", + "content": "tion of high-quality visual features at a local level. The potential applications of RIS are extensive, particularly in language-driven human-computer interaction domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 547, + 675 + ], + "type": "text", + "content": "Traditional RIS methods have employed linear fusion approaches and Fully Convolutional Networks (FCNs) for feature learning in RIS tasks [21, 33]. However, the recent advancement of attention mechanisms has shifted the focus towards extracting richer visual-language representation, with recent techniques demonstrating the effectiveness of Transformers. These models excel at modelling long-distance dependencies, making them suitable for cross-modal fusion [45, 55]. Further, Vision Transformer (ViT) methods such as VLT [13], EFN [16], and LAVT [54], which are based on [14], have shown significant improvements in RIS performance. Their ability to capture the nuances of cross-modal dependencies is crucial, and through a series of attention mechanisms, these models efficiently process and integrate both visual and language inputs for precise object segmentation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Despite their advantages, transformer-based approaches exhibit shortcomings in RIS mask prediction, with a notable problem being query collapse. This phenomenon occurs" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12903" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 323 + ], + "type": "text", + "content": "when different queries within the transformer, intended to identify distinct attributes or portions of objects, converge on the same mask prediction. This is particularly problematic in referring image segmentation, which inherently requires a prediction of a single mask, where all queries are collectively trained to predict the same mask. We experimentally observed this phenomenon, as illustrated in Fig. 1. Here, we visualize the heatmap of the cosine similarity between mask predictions from ReLA [32] and our method. In ReLA, most mask predictions are almost identical, leading to significantly overlapping mask predictions, whereas our method demonstrates a diverse set of query embeddings and mask predictions. Hence, prior transformer-based methods train mask predictions for all queries to associate with a single, specific ground truth mask. This training approach can lead to query collapse, where the model is not penalized for producing the same prediction across different queries, thereby affecting its capability to identify diverse attributes of the image. Consequently, this limitation restricts the variety of mask predictions, diminishing the model's ability to generalize effectively in complex scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 327, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 289, + 602 + ], + "type": "text", + "content": "To address these challenges, we propose LQMFormer, which comprises two key components. Firstly, we introduce a Gaussian Enhanced Multi-Modal Fusion (GMMF) module, aimed at enhancing the visual grounding mechanism. This module is designed to extract fine-grained local visual details while simultaneously enhancing global visual-language relationships. The fusion of these modalities allows the model to construct a more comprehensive understanding of the scene, both globally and locally, thereby improving visual grounding. Secondly, in our Dynamic Query Module (DQM), we generate a diverse set of language-dependent queries through Language-Query Cross Attention (LQCA) and employ a scoring network to prioritize queries relevant to the referring expression. With an improved visual-grounded representation facilitated by GMMF and selective query enhancement via DQM, our model effectively addresses query collapse. Moreover, by incorporating an auxiliary loss that increases the representational distance between mask predictions, our method efficiently differentiates between queries, mitigating the issue of query collapse. Our extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our framework. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 605, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 605, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 641 + ], + "type": "text", + "content": "- We introduce Gaussian enhanced Multi-Modal Fusion, which innovates the integration of local and global cross-modal information for more accurate visual grounding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 642, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 288, + 700 + ], + "type": "text", + "content": "- We develop an effective Dynamic Query Module (DQM) that not only generates a diverse query set but also employs a scoring network to selectively focus on queries based on given language expressions with respect to the decoder's references." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "- To mitigate query collapse, we propose an auxiliary regu" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 108 + ], + "type": "text", + "content": "larization loss function that increases the representational distance between queries, significantly improving mask differentiation and preventing Query Collapse." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 109, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 109, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 109, + 545, + 156 + ], + "type": "text", + "content": "- We conduct comprehensive validation of the proposed framework with extensive testing on four benchmark datasets, demonstrating improvements in referring image segmentation performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 167, + 397, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 397, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 397, + 180 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 188, + 473, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 188, + 473, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 473, + 201 + ], + "type": "text", + "content": "2.1. Referring Image Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 206, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 547, + 579 + ], + "type": "text", + "content": "Early works [29, 31, 45, 57] employed Convolutional Neural Networks (CNNs) [5, 21, 22, 31] and Recurrent Neural Networks (RNNs) [6, 31] to extract vision and language features. These features were then fused via sequential concatenation and convolution operations to predict the segmentation mask. In contrast, [58] proposed a two-stage network that initially uses Mask R-CNN[19] to predict categorical masks, followed by a selection of the relevant masks based on language descriptions. However, this approach exhibits limited capacity to capture the intricate relations between language and visual content. The advent of the Transformer [11, 14, 42, 46] in the vision community has led to the widespread adoption of Transformer-based architectures for extracting both visual and textual features [11, 24, 34]. VLT [13] uses cross-attention to produce query vectors from multi-modal features, which are then used to query images in the transformer decoder. Similarly, LAVT [54] shows early cross-modal fusion of features improves alignment. CRIS [49] adopts Transformer blocks to leverage the pretrained CLIP model's [44] robust image-text alignment capabilities. Additionally, GRES [32] extends cross-attention in the Transformer decoder to explicitly model visual-language dependencies. However, these all transformer-based approaches require matching queries to corresponding mask instances. In the context of referring expression segmentation, where typically only a single mask is available, this can lead to all queries converging to a single mask, a phenomenon known as query collapse. Motivated by this, in our method, we try to produce a diverse set of query prediction by overall improving visual-grounding and conditioning with respect to language expression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 586, + 545, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 545, + 600 + ], + "type": "text", + "content": "2.2. Vision and Language Representation Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "Vision and Language Representation Learning focuses on understanding the semantics and alignment between vision and language for multimodal reasoning tasks [36, 47, 59]. This field has seen substantial progress in applications like visual question answering [2], Image captioning [48], Image-text retrieval [4], zero-shot classification [44], and Image segmentation [23]. Contrastive pre-training strategies [28, 44] using large-scale datasets effectively project multiple modalities into a single embedding space. In contrast, as discussed," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12904" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "other methods [32, 54] create cross-modal interaction layers to fuse and comprehend multimodal features. The recent adoption of deep learning techniques in the frequency domain [9, 17, 39, 40] has gained attention for their global interaction capabilities. Specifically, [40] performs spectral-guided enhancement of the multi-scale vision-language features after feature extraction stage for Video Referring Segmentation. Drawing from these advancements, our approach integrates gaussian guidance within vision-language representation learning to facilitate enhanced local and global multi-modal interactions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 107, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 107, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 107, + 225 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "content": " and a language expression " + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "inline_equation", + "content": "L = \\{w_{i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "content": " words, our model predicts a pixel-wise mask " + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 232, + 288, + 270 + ], + "type": "text", + "content": ", which delineating the referred object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 274, + 115, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 115, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 115, + 286 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": "An overview diagram of our approach is shown in Figure 2. First, the language expression is encoded to extract high-dimensional language features " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{N_{t} \\times C}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "C, N_{t}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " indicates the number of channels, and words in the language expression, respectively. The Gaussian-enhanced Multi-Modal Fusion (GMMF) module then extracts language-grounded visual features " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{vl}}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ". Following [7], our approach incorporates a multi-scale strategy to exploit higher-resolution feature maps in the transformer decoder [7, 61]. We input language-aware queries " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " along with multi-scale outputs from the pixel decoder " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "i \\in \\left\\{\\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}, \\frac{1}{32}\\right\\}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " of the original image. The Dynamic Query Module processes language features " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " to predict language-aware queries " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ". The updating of queries " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " sequentially utilizes resolutions " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}, \\frac{1}{16}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\frac{1}{32}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ", applying a cycle of masked cross-attention (CA), self-attention (SA), and feed-forward network (FFN) operations " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " times within the decoder. Further, we map the final query outputs to an 'object' or 'no-object' two-dimensional space to enable 'no-object' predictions [32]. The final predicted mask is obtained at full resolution as the original image by decoding the pixel features " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{\\frac{1}{4}}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " with the einsum operation between " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "inline_equation", + "content": "F_{\\frac{1}{4}}" + }, + { + "bbox": [ + 46, + 293, + 289, + 559 + ], + "type": "text", + "content": ", followed by upsampling operation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 259, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 259, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 259, + 575 + ], + "type": "text", + "content": "3.2. Gaussian-enhanced Multi-Modal Fusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "After extracting language features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", we combine them with visual features using a four-stage hierarchical Swin Transformer [34] to obtain joint visual-language embedding. Each stage comprises Transformer layers " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\tau_{i})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", multi-modal feature fusion modules (GMMF), and residual gating " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\psi_{i})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". At every stage, as Transformer layers " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\tau_{i})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " enhance previous visual features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(F_{v})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", these features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(F_{v})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " combine with language features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(F_{l})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " via multi-modal modules (GMMF) to create multi-modal features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(F_{vl})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ". Finally, we use gating units " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "(\\psi_{vl})" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " to weight and combine " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": ", yielding enhanced visual features with linguistic information. Residual gating" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 109 + ], + "type": "text", + "content": "also works well with pre-trained (only vision) transformer weights and allows simultaneous feature modulation with language without having to re-train from scratch." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 122, + 545, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 122, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 305, + 122, + 545, + 145 + ], + "type": "text", + "content": "3.2.1 Gaussian Enhanced Multi-Modal Fusion Module in Visual Backbone" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 152, + 547, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 547, + 356 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 547, + 356 + ], + "type": "text", + "content": "The discrete Fourier transform (DFT) is extensively used to analyze the frequency representation of images. By transforming the signal into the frequency domain, characterized by global statistical properties, it plays a crucial role in various computer vision tasks [26, 51, 56]. Modulating signals on a point-wise basis in the Fourier domain alters the representation of inputs in the spatial domain [3, 40]. Furthermore, in the Fourier domain, low-frequency components usually correspond to the overall semantic content of the image, consistent with prior research [40, 52, 53, 56]. The Fourier transform is capable of disentangling global semantic information and structural coherence, with the semantic content predominantly found in the amplitude component [18, 27, 60]. These insights can be leveraged to create modules centered on Fourier-based methods, facilitating efficient and vital global interactions in multimodal understanding and improving visual grounding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": "To inculcate the above observation, as illustrated in Fig. 3, following [40], the Gaussian enhanced Multi-Modal Fusion Module (GMMF) is proposed as a key component in our model. It employs an enhancement block after cross-attention operations to combine Gaussian-enhanced vision and language features. Given an input vision feature " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "F_{vi} \\in \\mathbb{R}^{N \\times H \\times W \\times C_{vi}}" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": " from stage " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": " and a language feature " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "F_{l} \\in \\mathbb{R}^{N \\times T \\times C_{l}}" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": ", for cross-attention block, we first transform visual and language features into a common dimension " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "C_{i}" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": " using separate " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": " convolution layers. Subsequently, transform vision and language features are passed to an MHCA (Multi-Head Cross-Attention) module, where language features act as queries and vision features as keys and values. This operation results in the creation of a language-aware vision-shaped feature " + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "inline_equation", + "content": "F_{vl}'" + }, + { + "bbox": [ + 304, + 356, + 548, + 559 + ], + "type": "text", + "content": ". We then transform the visual-language features using the Gaussian Enhancement Block, as shown in Fig. 3, formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 564, + 515, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 564, + 515, + 607 + ], + "spans": [ + { + "bbox": [ + 331, + 564, + 515, + 607 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {v l} = F _ {\\mathrm {F F T}} ^ {- 1} \\left(\\operatorname {C o n v} \\left(F R \\left(\\phi \\left(F _ {v l} ^ {\\prime}, \\beta\\right) * \\mathcal {A} \\left(F _ {v l} ^ {\\prime}\\right) \\right. \\right. \\right. \\\\ , \\left. \\left. \\mathcal {P} \\left(F _ {v l} ^ {\\prime}\\right)\\right)\\right) + F _ {v l} ^ {\\prime} \\\\ \\end{array}", + "image_path": "27be045061223aceae98a44bbc8f01212a8e6734ec9934fca8dfb36343f7b9fd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(F_{vl}^{\\prime})" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(F_{vl}^{\\prime})" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " denote the amplitude and phase components of " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "F_{vl}^{\\prime}" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": "'s Fast Fourier Transform(FFT) [15, 43], " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{FFT}}(F_{vl}^{\\prime})" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": ". The * represents low-pass filtering on " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(F_{vl}^{\\prime})" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " using adaptive Gaussian filters " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\phi (F_{vl}^{\\prime},\\beta)" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": ", matching " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "F_{vl}^{\\prime}" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": "'s spatial dimensions and adjusting to input data via bandwidth " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " [40]. " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "FR" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "inline_equation", + "content": "Conv" + }, + { + "bbox": [ + 304, + 613, + 547, + 714 + ], + "type": "text", + "content": " signify feature reconstruction from amplitude and phase, and convolution operations, respectively." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "12905" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 547, + 243 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 547, + 243 + ], + "type": "image", + "image_path": "c5c2ea4b61ac346f4179f924ed36da4e9a8ba10514737dbc809b272401ede617.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": "Figure 2. Overview of the LQMFormer model architecture. The model processes an image and a linguistic expression, extracting vision-language features " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": " through a Gaussian-enhanced Multi-modal Fusion module and language features " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "F_{L}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": ". The multi-scale vision-language features " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": " are then processed by into a pixel decoder. Concurrently, the Language Query Cross-Attention Module (LQCA) refines the Language-enhanced Query bank " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "F_{L}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": ". A Scoring Network assesses the relevance of each query in " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": " by predicting soft-masks " + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "inline_equation", + "content": "(M_{b})" + }, + { + "bbox": [ + 46, + 263, + 547, + 331 + ], + "type": "text", + "content": ", which then modulate the queries fed into the transformer decoder. Finally, the model employs a global pooling layer and a Multi-Layer Perceptron (MLP) to predict the binary segmentation mask and determine the presence or absence of the referred object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 350, + 216, + 494 + ], + "blocks": [ + { + "bbox": [ + 120, + 350, + 216, + 494 + ], + "lines": [ + { + "bbox": [ + 120, + 350, + 216, + 494 + ], + "spans": [ + { + "bbox": [ + 120, + 350, + 216, + 494 + ], + "type": "image", + "image_path": "3c84351b775a6b875a199450fbcc9f174e9c9f024bd3758aabcdd1b397e5ebf6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 512, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 533 + ], + "type": "text", + "content": "Figure 3. Architecture overview of the Gaussian Enhancement Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": "Once " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": " is obtained, similar to [54], we combine the output from GMMF, " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": ", with that from the Transformer layers, " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": ". This process employs a gating mechanism, " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\psi_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": ", which learns a set of element-wise weight maps from " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": " to finely tune the scale of each element within " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": ". Subsequent to this adjustment, a residual combination of " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{v}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "inline_equation", + "content": "F_{vl} \\cdot \\psi_{vl}" + }, + { + "bbox": [ + 46, + 555, + 288, + 651 + ], + "type": "text", + "content": " is performed. The resultant output is then fed back into the vision backbone for further computation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 183, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 183, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 183, + 672 + ], + "type": "text", + "content": "3.3. Dynamic Query Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "The Dynamic Query Module (DQM) takes the language feature " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " and a Query bank " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " as inputs, the latter containing " + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": " learnable queries. Figure 4(b) shows the initial" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "content": "step, where the attention between the language feature " + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "content": " query embeddings " + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "inline_equation", + "content": "Q_{b} \\in \\mathbb{R}^{N_{qb} \\times C}" + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "content": " is computed, resulting in " + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 305, + 351, + 547, + 389 + ], + "type": "text", + "content": " attention maps:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 399, + 546, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 399, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 359, + 399, + 546, + 414 + ], + "type": "interline_equation", + "content": "A _ {b i} = \\operatorname {s o f t m a x} \\left(Q _ {b} \\sigma \\left(F _ {l} W _ {i k}\\right) ^ {T}\\right), \\tag {1}", + "image_path": "57cb5310f689c37a8ca261607108aaf12e9ed63c296c13e9664035f782b53b5e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "W_{ik}" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "C\\times C" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " matrix of learnable parameters, and " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " denotes the GeLU function [20]. The resulting " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "A_{bi}\\in \\mathbb{R}^{N_{qb}\\times L}" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " provides each query with a " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "1\\times L" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " attention map, indicating its important linguistic relation in the language features. Following this, the language-supported query bank features are derived using these attention maps: " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "Q_b^{\\prime} = A_{bi}\\sigma (F_lW_{iv})^T" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "W_{iv}" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " is another " + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "C\\times C" + }, + { + "bbox": [ + 305, + 422, + 547, + 519 + ], + "type": "text", + "content": " matrix of learnable parameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": "The variability in language expressions used to describe objects in images requires an approach that can adapt to different descriptions. This is because standard methods, like those seen in the initial transformer models (like in vanilla transformer [14]), struggle to capture the nuances in descriptions that include color, size, shape, and position. To address this, a larger set of adaptable queries is required. These queries are fine-tuned based on the input language, resulting in a refined set of queries, denoted as " + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "inline_equation", + "content": "Q_b'" + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": ". Each query in this refined set is then assessed with a scoring network, which outputs a soft-mask " + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": ". This soft-mask represents the relevance of each query. The final set of queries, " + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{final}}" + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": ", is obtained by combining " + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "inline_equation", + "content": "Q_b'" + }, + { + "bbox": [ + 304, + 521, + 547, + 689 + ], + "type": "text", + "content": " through a dot product, as shown in the equation below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 700, + 546, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 700, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 388, + 700, + 546, + 714 + ], + "type": "interline_equation", + "content": "Q _ {\\text {f i n a l}} = M _ {b} ^ {\\prime} \\odot Q _ {b} ^ {\\prime} \\tag {2}", + "image_path": "3572fb769942365136f6477370fc457b6eeb76f558d0f042f32e3d7e0e1a7385.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12906" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 386, + 238 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 386, + 238 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 386, + 238 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 386, + 238 + ], + "type": "image", + "image_path": "498463b200decfac3a2c4470079c807eead59ff69518d3b971682bddfcbde74c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 259, + 462, + 271 + ], + "lines": [ + { + "bbox": [ + 130, + 259, + 462, + 271 + ], + "spans": [ + { + "bbox": [ + 130, + 259, + 462, + 271 + ], + "type": "text", + "content": "Figure 4. Architecture overview of the (A.) Dynamic Query Module and (B.) LQCA Block" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 542, + 238 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 542, + 238 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 542, + 238 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 542, + 238 + ], + "type": "image", + "image_path": "46078c28fb1381519120e8b9102c5b752bbb0245513843bb8d6012d442baf33c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 293, + 153, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 153, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 153, + 304 + ], + "type": "text", + "content": "3.3.1 Scoring Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "text", + "content": "The scoring network in LQMFormer takes the Language-enhanced Query bank " + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "text", + "content": " as input and processes it through a sequence of linear layers. This network consists of four linear layers, with Layer Normalization [1] applied before the first layer. Each subsequent layer, except the last sigmoid, incorporates GELU [20] activation. The final output is a 1-dimensional logit " + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "inline_equation", + "content": "M_{b} \\in \\mathbb{R}^{N_{qb} \\times 1}" + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "text", + "content": ", serving as a soft-mask to modulate " + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "inline_equation", + "content": "Q_{b}^{\\prime}" + }, + { + "bbox": [ + 46, + 311, + 288, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "text", + "content": "To train the scoring network for predicting a relevant soft-mask " + }, + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "text", + "content": " effectively for the transformer decoder, we utilize its cross-attention map. This map is essential in identifying the subset of Queries most relevant to the referring expression and visual features, as highlighted during training. We sum up the attention maps from each decoder layer using Bilinear Interpolation (BI), thereby projecting attention " + }, + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "inline_equation", + "content": "A_{qf} \\in \\mathbb{R}^{N_{qb} \\times HW}" + }, + { + "bbox": [ + 46, + 407, + 290, + 515 + ], + "type": "text", + "content": " directed towards these Queries. The process is mathematically described as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 523, + 287, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 287, + 555 + ], + "type": "interline_equation", + "content": "A _ {q f} = \\sum_ {l = 1} ^ {L} \\sum_ {r = 1} ^ {3} \\mathrm {B I} \\left(\\mathrm {A} _ {l, r}, H W\\right) \\tag {3}", + "image_path": "b36123ff785bf7b4b5fc5a66421d92572ee16667e08fe3b14962acd925c71ee8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{l,r} \\in \\mathbb{R}^{N_{qb} \\times hw}" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": " represents the attention map from the " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": " layer at the " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "r^{th}" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": " resolution, with " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "hw" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": " being the original dimension and " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": " the desired output dimension. After aligning each attention map to the common resolution " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "HW" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": ", we then sum across this dimension to generate a likelihood map " + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 46, + 559, + 288, + 633 + ], + "type": "text", + "content": ", as indicated in the subsequent equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 124, + 640, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 640, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 124, + 640, + 287, + 673 + ], + "type": "interline_equation", + "content": "L M = \\sum_ {i = 1} ^ {H W} A _ {q f} [:, i ] \\tag {4}", + "image_path": "dd80b7da57298d3db36e25e94324147caddf05e00fbb8da102e160e4df7e0c83.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "This likelihood map " + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LM} \\in \\mathbb{R}^{N_{qb} \\times 1}" + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": ", now captures the concentrated attention across all Queries, serves as a condition in refining the predictions of the scoring network. The" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": "scoring network is trained to predict the likelihood of each Query being included in the top-" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\rho\\%" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": " of the most referenced tokens. This process involves binarizing the likelihood map " + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "LM" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": " to retain only the top-" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\rho\\%" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": " Queries, creating a binarized version " + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "LM_{\\mathrm{bin}}" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": ". The training objective is then formulated as a Binary Cross-Entropy (BCE) loss between the predicted soft-mask " + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": " and the binarized likelihood map " + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "inline_equation", + "content": "LM_{\\mathrm{bin}}" + }, + { + "bbox": [ + 305, + 293, + 547, + 389 + ], + "type": "text", + "content": ". The BCE loss function is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 404, + 546, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 404, + 546, + 466 + ], + "spans": [ + { + "bbox": [ + 309, + 404, + 546, + 466 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {B C E} \\operatorname {L o s s} = - \\frac {1}{N _ {q b}} \\sum_ {i = 1} ^ {N _ {q b}} \\left[ L M _ {\\mathrm {b i n}, i} \\log \\left(M _ {b, i}\\right) \\right. \\\\ + \\left(1 - L M _ {\\text {b i n}, i}\\right) \\log \\left(1 - M _ {b, i}\\right) ] \\tag {5} \\\\ \\end{array}", + "image_path": "5a871475c606579e96995b89fdb7a18f94f4862ee4fa4342b9667072b9768bbb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "text", + "content": " represents the number of Queries in " + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "inline_equation", + "content": "Q_b^{\\prime}" + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 472, + 547, + 545 + ], + "type": "text", + "content": " indexes each Query. This loss function effectively trains the scoring network to align its predictions with the most salient Queries as determined by the decoder's cross-attention, thereby refining the model's accuracy in identifying and segmenting objects as described by the referring expressions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 552, + 451, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 451, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 451, + 564 + ], + "type": "text", + "content": "3.4. Query-Mask Margin Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "Query-Mask Margin Loss is designed to maintain a margin of separability between different query feature representations, an essential aspect for tasks with variable outputs. The loss function operates by initially computing pairwise differences and distances between mask embeddings within a batch. This computation yields a tensor of shape " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "B \\times N \\times N" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " represents the batch size, and " + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": " is the number of queries. Subsequently, these distances for each embedding are sorted, with a particular focus on the second smallest value, which signifies the distance to the nearest neighbor. The margin loss is defined through the max function, ensuring a minimum distance of 1.0 between embeddings. The" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "12907" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "content": "mathematical formulation of QM-loss is as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 93, + 252, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 93, + 252, + 107 + ], + "spans": [ + { + "bbox": [ + 82, + 93, + 252, + 107 + ], + "type": "interline_equation", + "content": "\\mathrm {Q M - l o s s} = \\max (0, 1. 0 - \\operatorname {d i s t} \\left(M _ {i}, M _ {j}\\right))", + "image_path": "9cb355e200b89efb8c4e09dd334e09f1d4a74fcb6f018dee8183fdb60846719a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "inline_equation", + "content": "\\text{dist}(M_i, M_j)" + }, + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "text", + "content": " denotes the calculation of pairwise distances between the mask embeddings, represented as " + }, + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 114, + 288, + 139 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 149, + 208, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 149, + 208, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 149, + 208, + 163 + ], + "type": "text", + "content": "4. Experiments and Discussion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 168, + 243, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 243, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 243, + 182 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 187, + 289, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 289, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 289, + 402 + ], + "type": "text", + "content": "Datasets: Our experiments are conducted on four primary benchmark datasets in Referring Image Segmentation: RefCOCO [57], RefCOCO+ [57], RefCOCOg [38], and GRES [32]. These datasets derive their images from MSCOCO [30]. The dataset details are as follows: RefCOCO (19,994 images, 50,000 objects, 142,209 expressions), RefCOCO+ (19,992 images, 49,856 objects, 141,564 expressions), RefCOCOg (26,711 images, 54,822 objects, 85,474 expressions), and GRES (19,994 images, 60,287 objects, 278,232 expressions). RefCOCO mainly focuses on expressions that specify object locations, while RefCOCO+ prioritizes object descriptions. RefCOCOg presents a higher challenge due to longer and more complex expressions (averaging 8.4 words compared to 3.5 in others). GRES broadens the scope of RefCOCO by including expressions that refer to multiple objects or no objects, thereby enhancing the problem's generalizability. We use both UMD split [41] and Google split [38] for RefCOCOg dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 289, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 629 + ], + "type": "text", + "content": "Evaluation Metrics. For single-target object segmentation, we report our results using two kinds of metrics [13, 37, 49, 54]: overall IoU (oIoU) and Precision@X (P@X). oIoU divides the total intersection pixels by the total union pixels across all test images. Precision@X calculates the percentage of testing samples of which the model prediction has an IoU score higher than the threshold value X. To extend our evaluation to non-target and multi-target object segmentation, following [32], we include additional metrics: Sensitivity (N-acc.), Specificity (T-acc.), and Generalized IoU (gIoU). N-acc. evaluates the model's performance on identifying non-target samples, while T-acc. measures the impact of generalization on non-target samples on the performance of target samples. We opt for gIoU over oIoU due to the latter's bias towards larger objects. gIoU computes the mean IoU per image for all samples. In cases with no-target samples, IoU values are assigned as 1 for true positives and 0 for false negatives, ensuring an equitable metric for all object sizes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Implementation Details. Our model uses Swin Transformer-B [34] as the visual encoder and BERT [11] as the language decoder. The transformer layer in Swin-B is initialized with classification weights pre-trained on ImageNet22K [10]. We use BERT implementation from the HuggingFace Transformer library [50]. It comprises 12 layers with a hidden size of 768, initialized using official" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": "pre-trained weights. We use our Transformer decoder proposed in Section 3.1 with " + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "inline_equation", + "content": "\\mathrm{L} = 3" + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": " (i.e., 9 layers total) and 100 queries by default. The AdamW optimizer [35] is employed for optimization with an initial learning rate of 0.00001. Our model is trained over 100K iterations with a batch size of 48. Our image resizing is standardized to " + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "inline_equation", + "content": "480 \\times 480" + }, + { + "bbox": [ + 304, + 72, + 547, + 145 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 155, + 524, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 155, + 524, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 155, + 524, + 167 + ], + "type": "text", + "content": "4.2. Comparison with the state of art methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "text", + "content": "Referring Image Segmentation Table 1 presents the results for Referring Image Segmentation, and our proposed method, LQMFormer, produces competitive performance against existing state-of-the-art approaches across multiple standard benchmarks. On the RefCOCO dataset, our proposed method obtains the highest IoU across all divisions (val, test A, and test B), surpassing recent leading methods such as VLT [13] and ReLA [32]. Notably, it shows a significant margin in the test B split, indicating enhanced capability in complex visual contexts. In the RefCOCO+ dataset, LQMFormer also demonstrates its effectiveness in the test A set. Specifically, in the test A split with a cIoU of " + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "inline_equation", + "content": "71.84\\%" + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "text", + "content": ". In the Val and Test B splits, our proposed method achieves competitive performance against ReLA [32] and LAVT [54]. For the G-Ref dataset, which presents varied and longer referring expressions, our proposed method achieves the top performance in the " + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{val}_{(\\mathrm{G})}" + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "text", + "content": " split and maintains competitive results in the other splits. This performance is particularly notable over LAVT [54] in the " + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{val}_{(\\mathrm{G})}" + }, + { + "bbox": [ + 304, + 174, + 547, + 474 + ], + "type": "text", + "content": " split, highlighting its adaptability to diverse challenging datasets. While the margin of improvement offered by LQMFormer in classic RIS datasets may be comparatively smaller than that observed in GREs, the results signify that Gaussian-enhanced Multi-Modal Fusion and explicit modeling of Queries from Decoder are beneficial for the general Referring Image Segmentation setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": "Generalised Referring Image Segmentation In the GRES dataset, the LQMFormer model demonstrates notable improvements in Region Instance Segmentation (RIS), as shown in Table 2. The model achieves a generalized Intersection over Union (IoU) score of " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "70.94\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": " on the validation set, surpassing other models such as ReLA, LAVT, and CRIS by a margin of " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "7 - 12\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": ". This performance indicates its effective segmentation across various object sizes. The object-level Intersection over Union (oIoU) score of " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "64.98\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": " further underscores the model's segmentation capabilities, highlighting the effectiveness of the Improved Visual Grounding and Dynamic Query Selection module. Another important aspect of LQMFormer is its performance in no-target identification, with an accuracy score of " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "67.47\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": ". This score is " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": " higher than that of existing methods, marking significant progress in accurately identifying no-target samples—a noted challenge in RIS. The inclusion of the Dynamic Query Module (DQM) and the Top " + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "inline_equation", + "content": "p\\%" + }, + { + "bbox": [ + 304, + 474, + 548, + 714 + ], + "type": "text", + "content": " ratio significantly reduces the impact of the number of queries and mitigates query collapse. As a result, fewer but more effective queries make it easier to distinguish" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12908" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 82, + 547, + 184 + ], + "blocks": [ + { + "bbox": [ + 147, + 70, + 443, + 81 + ], + "lines": [ + { + "bbox": [ + 147, + 70, + 443, + 81 + ], + "spans": [ + { + "bbox": [ + 147, + 70, + 443, + 81 + ], + "type": "text", + "content": "Table 1. Results on classic RES in terms of oIoU. U: UMD split. G: Google split." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 82, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 48, + 82, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 82, + 547, + 184 + ], + "type": "table", + "html": "
MethodsVisual EncoderTextual EncoderRefCOCORefCOCO+val(U)G-Ref test(U)val(G)
valtest Atest Bvaltest Atest B
MCN [37]Darknet53bi-GRU62.4464.2059.7150.6254.9944.6949.2249.40-
VLT [12]Darknet53bi-GRU67.5270.4765.2456.3060.9850.0854.9657.7352.02
ReSTR [25]ViT-BTransformer67.2269.3064.4555.7860.4448.27--54.48
CRIS [49]CLIP-R101CLIP70.4773.1866.1062.2768.0853.6859.8760.36-
LAVT [54]Swin-BBERT72.7375.8268.7962.1468.3855.1061.2462.0960.50
VLT [13]Swin-BBERT72.9675.9669.6063.5368.4356.9263.4966.2262.80
ReLA [32]Swin-BBERT73.8276.4870.1866.0471.0257.6565.0065.9762.70
LQMFormer (ours)Swin-BBERT74.1676.8271.0465.9171.8457.5964.7366.0462.97
", + "image_path": "fe773ca9fd2ddc1cb6fd9367ccc6f5aa7e1b668c7c94c35af9bfc0d965536678.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 64, + 213, + 269, + 324 + ], + "blocks": [ + { + "bbox": [ + 95, + 202, + 238, + 213 + ], + "lines": [ + { + "bbox": [ + 95, + 202, + 238, + 213 + ], + "spans": [ + { + "bbox": [ + 95, + 202, + 238, + 213 + ], + "type": "text", + "content": "Table 2. Comparison on GRES dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 213, + 269, + 324 + ], + "lines": [ + { + "bbox": [ + 64, + 213, + 269, + 324 + ], + "spans": [ + { + "bbox": [ + 64, + 213, + 269, + 324 + ], + "type": "table", + "html": "
MethodsvalNo-target val
oIoUgIoUN-acc.T-acc.
MattNet [58]47.5148.2441.1596.13
LTS [24]52.3052.70--
VLT [12]52.5152.0047.1795.72
CRIS [49]55.3456.27--
LAVT [54]57.6458.4049.3296.18
ReLA [32]62.4263.6056.3796.32
LQMFormer (ours)64.9870.9467.4799.12
", + "image_path": "23c14dedbd952423f7b9ceca7b7b2e6a655a028a10b80d58a3e36052a4c778bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 356, + 299, + 492 + ], + "blocks": [ + { + "bbox": [ + 108, + 335, + 227, + 346 + ], + "lines": [ + { + "bbox": [ + 108, + 335, + 227, + 346 + ], + "spans": [ + { + "bbox": [ + 108, + 335, + 227, + 346 + ], + "type": "text", + "content": "Table 3. Ablation Study Results" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 356, + 299, + 492 + ], + "lines": [ + { + "bbox": [ + 48, + 356, + 299, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 356, + 299, + 492 + ], + "type": "table", + "html": "
ConfigurationoIoUgIoU
(a) Analysis on Dynamic Query module
Ours64.9870.94
w/o score-based learning62.7567.16
w/o scoring module62.0265.48
w/o QM-loss64.1968.37
(b) Analysis on multi-modal fusion module
Ours64.9870.94
w/o modality fusion module and QM-loss49.6852.34
", + "image_path": "9985c7a5ab596ca069581f7fb7e7ca2195276e5b9e6ae2d1c7e9619f2c4905b4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 67, + 533, + 267, + 608 + ], + "blocks": [ + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "text", + "content": "Table 4. Ablation study of Query Numbers " + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "inline_equation", + "content": "{N}_{qb}" + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "inline_equation", + "content": "{Q}_{b}" + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 502, + 288, + 525 + ], + "type": "text", + "content": " : without scoring module. Top p% ratio is 20" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 533, + 267, + 608 + ], + "lines": [ + { + "bbox": [ + 67, + 533, + 267, + 608 + ], + "spans": [ + { + "bbox": [ + 67, + 533, + 267, + 608 + ], + "type": "table", + "html": "
NqoIoUgIoUPr@0.7Pr@0.8Pr@0.9
2057.2258.3169.1560.2622.47
5062.4864.5372.0261.7931.58
10064.9870.9475.0365.5134.05
100‡62.0265.4873.1562.8731.55
", + "image_path": "ea90189f4651f8ed0205cf000312fb8177d94416615b9e5db1959f78ed005330.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 627, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 664 + ], + "type": "text", + "content": "between \"target\" and \"no-target\". The model's specificity (true accuracy) score of " + }, + { + "bbox": [ + 46, + 627, + 287, + 664 + ], + "type": "inline_equation", + "content": "99.12\\%" + }, + { + "bbox": [ + 46, + 627, + 287, + 664 + ], + "type": "text", + "content": " also reflects its effectiveness in classifying target samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 141, + 684 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "In our comprehensive ablation studies, we evaluated the performance impact of various components of LQMFormer" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 307, + 222, + 545, + 280 + ], + "blocks": [ + { + "bbox": [ + 340, + 202, + 511, + 214 + ], + "lines": [ + { + "bbox": [ + 340, + 202, + 511, + 214 + ], + "spans": [ + { + "bbox": [ + 340, + 202, + 511, + 214 + ], + "type": "text", + "content": "Table 5. Ablation study of Top p% ratio in " + }, + { + "bbox": [ + 340, + 202, + 511, + 214 + ], + "type": "inline_equation", + "content": "{Q}_{b}" + }, + { + "bbox": [ + 340, + 202, + 511, + 214 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 222, + 545, + 280 + ], + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 280 + ], + "type": "table", + "html": "
p%oIoUgIoUN.acc.Pr@0.7Pr@0.8Pr@0.9
2064.9870.9467.4775.0365.5134.05
5063.8567.7663.2574.2964.4233.58
10061.3363.7758.9272.4561.8831.21
", + "image_path": "32d80fce1ebddcae2dcf57122b9fe83ed19ad8512b1a13bde018903d2f678c44.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 312, + 538, + 392 + ], + "blocks": [ + { + "bbox": [ + 340, + 292, + 511, + 303 + ], + "lines": [ + { + "bbox": [ + 340, + 292, + 511, + 303 + ], + "spans": [ + { + "bbox": [ + 340, + 292, + 511, + 303 + ], + "type": "text", + "content": "Table 6. Ablation study of Grounding Module" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 312, + 538, + 392 + ], + "lines": [ + { + "bbox": [ + 315, + 312, + 538, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 312, + 538, + 392 + ], + "type": "table", + "html": "
FusionoIoUgIoUPr@0.7Pr@0.8Pr@0.9
Baseline49.6852.3455.2741.3814.72
Post-CA60.2261.8974.1564.4432.67
S-Post-CA [40]62.1364.4173.6463.3532.71
VL-Grounding63.4368.8172.9564.1933.96
GMMF64.9870.9475.0365.5134.05
", + "image_path": "2eba967af0c473b4c8c912bc202dceaf0eef097010acc20b93af5093c3079627.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": "on the GRES dataset. These studies focused on the Dynamic Query Module, Query Numbers " + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": ", the Top " + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\mathsf{p}\\%" + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": " ratio in " + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 304, + 412, + 545, + 472 + ], + "type": "text", + "content": ", and the Grounding Module. The results helps understanding key insights into the model's functioning and optimization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "text", + "content": "Analysis on Dynamic Query Module: Table 3(a) presents the ablation results for the language query selection module. The removal of score-based learning significantly impacts scores as it limits the explicit training of Query scores from the scoring network. Also, notably, removing the scoring module and QM-loss leads to a decrease in oIoU score by " + }, + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "inline_equation", + "content": "2.96\\%" + }, + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 304, + 472, + 546, + 568 + ], + "type": "text", + "content": ", respectively, highlighting its critical role in achieving optimal referring segmentation performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "content": "Query Bank number " + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{qb}" + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "content": ": The ablation on Query Bank number demonstrated a notable impact on the model's performance as shown in Table 4. With an increase in " + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "content": " from 20 to 100, there was a consistent improvement in both IoU from 57.22% to 64.98% and gIoU from 58.31% to 70.94%, as well as in Precision metrics. Particularly, " + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "inline_equation", + "content": "N_{q} = 100" + }, + { + "bbox": [ + 304, + 569, + 547, + 689 + ], + "type": "text", + "content": " yielded the highest scores across all metrics, indicating an optimal balance in query numbers for effective segmentation. Thus, it optimal to have more number of queries with Top p% as 20." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Top p% Ratio in " + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": ": The study on the Top p% ratio revealed that a 20% ratio achieved the best performance in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12909" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 93, + 175, + 207 + ], + "blocks": [ + { + "bbox": [ + 86, + 79, + 116, + 91 + ], + "lines": [ + { + "bbox": [ + 86, + 79, + 116, + 91 + ], + "spans": [ + { + "bbox": [ + 86, + 79, + 116, + 91 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 93, + 175, + 207 + ], + "lines": [ + { + "bbox": [ + 60, + 93, + 175, + 207 + ], + "spans": [ + { + "bbox": [ + 60, + 93, + 175, + 207 + ], + "type": "image", + "image_path": "c99725aeceedbe99d02f67028ca126037faa12ecd804958cd824f88c5444f2a0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 209, + 477, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 477, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 477, + 224 + ], + "type": "text", + "content": "Query: \"Orange cow with front facing us and an orange cow with back facing us\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 184, + 93, + 298, + 207 + ], + "blocks": [ + { + "bbox": [ + 227, + 79, + 255, + 91 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 255, + 91 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 255, + 91 + ], + "type": "text", + "content": "ReLA" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 93, + 298, + 207 + ], + "lines": [ + { + "bbox": [ + 184, + 93, + 298, + 207 + ], + "spans": [ + { + "bbox": [ + 184, + 93, + 298, + 207 + ], + "type": "image", + "image_path": "710ec28d0f6087f1f95e81d9a7f18d23d81c57275ff7db02688a4b8340b1bc9c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 93, + 422, + 207 + ], + "blocks": [ + { + "bbox": [ + 332, + 79, + 397, + 91 + ], + "lines": [ + { + "bbox": [ + 332, + 79, + 397, + 91 + ], + "spans": [ + { + "bbox": [ + 332, + 79, + 397, + 91 + ], + "type": "text", + "content": "LQMFormer" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 93, + 422, + 207 + ], + "lines": [ + { + "bbox": [ + 308, + 93, + 422, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 93, + 422, + 207 + ], + "type": "image", + "image_path": "6bbae835706fe86cda15df2823a351e795f182da57fcf295385d200680f6f055.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 431, + 93, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 452, + 79, + 523, + 91 + ], + "lines": [ + { + "bbox": [ + 452, + 79, + 523, + 91 + ], + "spans": [ + { + "bbox": [ + 452, + 79, + 523, + 91 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 431, + 93, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 431, + 93, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 431, + 93, + 545, + 207 + ], + "type": "image", + "image_path": "042944f6b5ffade7820e6adcd8845713c1ad98c0949ed153f26f58880014a2f3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 238, + 174, + 352 + ], + "blocks": [ + { + "bbox": [ + 60, + 238, + 174, + 352 + ], + "lines": [ + { + "bbox": [ + 60, + 238, + 174, + 352 + ], + "spans": [ + { + "bbox": [ + 60, + 238, + 174, + 352 + ], + "type": "image", + "image_path": "6c0f1a83cf9a3602ca6ee537da02f1f9ee6e7d88e51b7ddeccd20cde2ee5bdb1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 354, + 384, + 368 + ], + "lines": [ + { + "bbox": [ + 59, + 354, + 384, + 368 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 384, + 368 + ], + "type": "text", + "content": "Query: \"The pizza in the middle and the guy holding a camera\"" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 115, + 383, + 476, + 395 + ], + "lines": [ + { + "bbox": [ + 115, + 383, + 476, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 476, + 395 + ], + "type": "text", + "content": "Figure 5. Qualitave Comparison of our model with LQMFormer with ReLA [32] on GRES dataset." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 185, + 237, + 299, + 352 + ], + "blocks": [ + { + "bbox": [ + 185, + 237, + 299, + 352 + ], + "lines": [ + { + "bbox": [ + 185, + 237, + 299, + 352 + ], + "spans": [ + { + "bbox": [ + 185, + 237, + 299, + 352 + ], + "type": "image", + "image_path": "1aaecd4830aa26d2c362a65435b50ae3b9ce9bf2f3382b9cc4eada973fa6e095.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 237, + 422, + 351 + ], + "blocks": [ + { + "bbox": [ + 309, + 237, + 422, + 351 + ], + "lines": [ + { + "bbox": [ + 309, + 237, + 422, + 351 + ], + "spans": [ + { + "bbox": [ + 309, + 237, + 422, + 351 + ], + "type": "image", + "image_path": "9b3264f5163287e93fd3249ff0231101f8157415e7338998b54c20ae4b555136.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 432, + 237, + 546, + 352 + ], + "blocks": [ + { + "bbox": [ + 432, + 237, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 432, + 237, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 432, + 237, + 546, + 352 + ], + "type": "image", + "image_path": "2403e8bb106cf2cb68033e8bf81ff97f161dc2f63d965a6d82e56d02c0393372.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "text", + "content": "terms of oIoU, gIoU, and Precision metrics(Table 5). Higher ratios, like " + }, + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 401, + 288, + 483 + ], + "type": "text", + "content": ", resulted in diminished performance as it resulted in more and more query collapse issues as many queries contribute towards the final mask generation. This finding indicates that a moderate selection of top-performing queries is crucial for balancing precision and recall." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": "Grounding Module: The ablation study of the Grounding Module, as shown in Table 3(b) and Table 6, highlights the effectiveness of Gaussian Enhancement for Referring Image Segmentation (RIS) tasks. The Gaussian enhanced Multi-Model Fusion (GMMF) method yields a improvement, specifically enhancing oIoU by " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": ", from " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "63.43\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " in VL-Grounding to " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "64.98\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " in GMMF. This underlines the superior performance of Gaussian Filtering in the Fourier Domain over traditional cross-attention methods. Furthermore, compared to the Post-CA which registers an oIoU of " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "60.22\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": ", the GMMF and VL-Grounding methods showcase substantial improvements in multi-modal representation within the backbone stage, with GMMF outperforming Post-CA by " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "4.76\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " and VL-Grounding by " + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "inline_equation", + "content": "3.21\\%" + }, + { + "bbox": [ + 46, + 486, + 290, + 714 + ], + "type": "text", + "content": " in oIoU. The comparison with S-Post-CA [40], which aims at enhancing visual representation post-feature extraction, show the importance of integrating advanced grounding techniques early in the model's architecture for improved RIS performance. Overall, our proposed Gaussian enhanced Multi-Model Fusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 401, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 546, + 426 + ], + "type": "text", + "content": "(GMMF) improves the visual grounding compared to other methods, resulting in a improved referring segmentation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 378, + 462 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 474, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 548, + 713 + ], + "type": "text", + "content": "In this paper, we address the problem of query collapse in Referring Image Segmentation by enhancing visual grounding and generating diverse query sets, and by modeling their relevance based on language expressions. Our proposed model, LQMFormer, incorporates Gaussian Enhanced Multi-Modal Fusion to improve visual grounding, and a Dynamic Query Module that not only generates a diverse set of queries but also employs a scoring network to selectively focus on queries based on given language expressions. Additionally, to mitigate query collapse, we introduce a novel loss function that enforces a margin of separation between query feature representations, facilitating improved representation without the need for additional supervision. The LQMFormer method notably surpasses state-of-the-art methods in both referring image segmentation and generalized referring image segmentation performance in most settings. For future work, we aim to extend the LQMFormer model's capabilities along with LLM and focus on effectively bridging LLM to better handle multimodal contexts." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12910" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 289, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 289, + 179 + ], + "type": "text", + "content": "[2] Ankan Bansal, Yuting Zhang, and Rama Chellappa. Visual question answering on image sets. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI 16, pages 51–67. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 204 + ], + "type": "text", + "content": "[3] Glenn D Bergland. A guided tour of the fast fourier transform. IEEE spectrum, 6(7):41-52, 1969. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 239 + ], + "type": "text", + "content": "[4] Min Cao, Shiping Li, Juntao Li, Liqiang Nie, and Min Zhang. Image-text retrieval: A survey on recent research and development. arXiv preprint arXiv:2203.14713, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 240, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 240, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 240, + 288, + 282 + ], + "type": "text", + "content": "[5] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In ICCV, pages 7454-7463, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 328 + ], + "type": "text", + "content": "[6] Yi-Wen Chen, Yi-Hsuan Tsai, Tiantian Wang, Yen-Yu Lin, and Ming-Hsuan Yang. Referring expression object segmentation with caption-aware consistency. arXiv preprint arXiv:1910.04748, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 385 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 288, + 430 + ], + "type": "text", + "content": "[8] Ming-Ming Cheng, Shuai Zheng, Wen-Yan Lin, Vibhav Vineet, Paul Sturgess, Nigel Crook, Niloy J Mitra, and Philip Torr. Imagesprit: Verbal guided image parsing. ACM Transactions on Graphics (ToG), 34(1):1-11, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 432, + 289, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 289, + 464 + ], + "type": "text", + "content": "[9] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 510 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 566 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proc. NAACL-HLT, pages 4171–4186. Association for Computational Linguistics, 2019. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 601 + ], + "type": "text", + "content": "[12] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In ICCV, pages 16321-16330, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[13] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vlt: Vision-language transformer and query generation for referring segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[15] Pierre Duhamel and Martin Vetterli. Fast fourier transforms: a tutorial review and a state of the art. Signal processing, 19 (4):259-299, 1990. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 547, + 140 + ], + "type": "text", + "content": "[16] Guang Feng, Zhiwei Hu, Lihe Zhang, and Huchuan Lu. Encoder fusion network with co-attention embedding for referring image segmentation. In CVPR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "text", + "content": "[17] Lorenzo Giambagli, Lorenzo Buffoni, Timoteo Carletti, Walter Nocentini, and Duccio Fanelli. Machine learning in spectral domain. Nature communications, 12(1):1330, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 546, + 207 + ], + "type": "text", + "content": "[18] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE transactions on pattern analysis and machine intelligence, 35(6):1397-1409, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 547, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 231 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 232, + 546, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 546, + 254 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 546, + 254 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 255, + 547, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 547, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 547, + 287 + ], + "type": "text", + "content": "[21] Ronghang Hu, Marcus Rohrbach, and Trevor Darrell. Segmentation from natural language expressions. In ECCV, pages 108-124. Springer, 2016. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 289, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 547, + 331 + ], + "type": "text", + "content": "[22] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In CVPR, pages 4424-4433, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 334, + 547, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 334, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 334, + 547, + 388 + ], + "type": "text", + "content": "[23] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 390, + 546, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 390, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 390, + 546, + 422 + ], + "type": "text", + "content": "[24] Ya Jing, Tao Kong, Wei Wang, Liang Wang, Lei Li, and Tieniu Tan. Locate then segment: A strong pipeline for referring image segmentation. In CVPR, pages 9858-9867, 2021. 2, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 423, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 547, + 466 + ], + "type": "text", + "content": "[25] Namyup Kim, Dongwon Kim, Cuiling Lan, Wenjun Zeng, and Suha Kwak. Restr: Convolution-free referring image segmentation using transformers. In CVPR, pages 18145-18154, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 468, + 547, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 547, + 522 + ], + "type": "text", + "content": "[26] Sangrok Lee, Jongseong Bae, and Ha Young Kim. Decompose, adjust, compose: Effective normalization by playing with frequency for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11776-11785, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 524, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 524, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 524, + 547, + 567 + ], + "type": "text", + "content": "[27] Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. arXiv preprint arXiv:2302.11831, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 569, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 546, + 624 + ], + "type": "text", + "content": "[28] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 625, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 546, + 667 + ], + "type": "text", + "content": "[29] Ruiyu Li, Kaican Li, Yi-Chun Kuo, Michelle Shu, Xiaojuan Qi, Xiaoyong Shen, and Jiaya Jia. Referring image segmentation via recurrent refinement networks. In CVPR, pages 5745-5753, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12911" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[31] Chenxi Liu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, and Alan Yuille. Recurrent multimodal interaction for referring image segmentation. In ICCV, pages 1271-1280, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "type": "text", + "content": "[32] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 205 + ], + "type": "text", + "content": "[33] Jingyu Liu, Liang Wang, and Ming-Hsuan Yang. Referring expression generation and comprehension via attributes. In ICCV, pages 4856-4864, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 249 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 272 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 273, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 273, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 273, + 287, + 326 + ], + "type": "text", + "content": "[36] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10437-10446, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 371 + ], + "type": "text", + "content": "[37] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative network for joint referring expression comprehension and segmentation. In CVPR, pages 10034-10043, 2020. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 415 + ], + "type": "text", + "content": "[38] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In CVPR, pages 11-20, 2016. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 480 + ], + "type": "text", + "content": "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 536 + ], + "type": "text", + "content": "[40] Bo Miao, Mohammed Bennamoun, Yongsheng Gao, and Ajmal Mian. Spectrum-guided multi-granularity referring video object segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 920–930, 2023. 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 569 + ], + "type": "text", + "content": "[41] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In ECCV, pages 792–807. Springer, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[42] Kartik Narayan, Vibashan VS, Rama Chellappa, and Vishal M Patel. Facexformer: A unified transformer for facial analysis. arXiv preprint arXiv:2403.12960, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 603, + 287, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 625 + ], + "type": "text", + "content": "[43] Henri J Nussbaumer and Henri J Nussbaumer. The fast Fourier transform. Springer, 1982. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[45] Hengcan Shi, Hongliang Li, Fanman Meng, and Qingbo Wu. Key-word-aware network for referring expression image segmentation. In ECCV, pages 38-54, 2018. 1, 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. pages 5998-6008, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 183 + ], + "type": "text", + "content": "[47] Vibashan VS, Ning Yu, Chen Xing, Can Qin, Mingfei Gao, Juan Carlos Niebles, Vishal M Patel, and Ran Xu. Mask-free ovis: Open-vocabulary instance segmentation without manual mask annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23539-23549, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 226 + ], + "type": "text", + "content": "[48] Yiyu Wang, Jungang Xu, and Yingfei Sun. End-to-end transformer based model for image captioning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2585-2594, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 546, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 546, + 270 + ], + "type": "text", + "content": "[49] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip-driven referring image segmentation. In CVPR, pages 11686-11695, 2022. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 271, + 547, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 547, + 346 + ], + "type": "text", + "content": "[50] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chau-mond, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, pages 38-45, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 349, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 547, + 403 + ], + "type": "text", + "content": "[51] Qinwei Xu, Ruipeng Zhang, Ya Zhang, Yanfeng Wang, and Qi Tian. A fourier-based framework for domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14383-14392, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 546, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 546, + 446 + ], + "type": "text", + "content": "[52] Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks. arXiv preprint arXiv:1901.06523, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 449, + 547, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 547, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 547, + 513 + ], + "type": "text", + "content": "[53] Zhi-Qin John Xu, Yaoyu Zhang, and Yanyang Xiao. Training behavior of deep neural network in frequency domain. In Neural Information Processing: 26th International Conference, ICONIP 2019, Sydney, NSW, Australia, December 12–15, 2019, Proceedings, Part I 26, pages 264–274. Springer, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 546, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 546, + 557 + ], + "type": "text", + "content": "[54] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Lavo: Language-aware vision transformer for referring image segmentation. In CVPR, pages 18155-18165, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 546, + 592 + ], + "type": "text", + "content": "[55] Linwei Ye, Mrigank Rochan, Zhi Liu, and Yang Wang. Cross-modal self-attention network for referring image segmentation. In CVPR, pages 10502–10511, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 593, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 546, + 635 + ], + "type": "text", + "content": "[56] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 546, + 669 + ], + "type": "text", + "content": "[57] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In ECCV, pages 69-85. Springer, 2016. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 546, + 713 + ], + "type": "text", + "content": "[58] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In CVPR, pages 1307-1315, 2018. 2, 7" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12912" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[59] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[60] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. Fourmer: An efficient global modeling paradigm for image restoration. In International Conference on Machine Learning, pages 42589-42601. PMLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 288, + 217 + ], + "type": "text", + "content": "[61] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12913" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_content_list.json b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fdd00a7ac23b4b49fcc315f77339c46577c6eb13 --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_content_list.json @@ -0,0 +1,1699 @@ +[ + { + "type": "text", + "text": "LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels", + "text_level": 1, + "bbox": [ + 58, + 130, + 910, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tuo Feng $^{1}$ , Wenguan Wang $^{2}$ , Fan Ma $^{2*}$ , Yi Yang $^{2}$", + "bbox": [ + 290, + 179, + 676, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 ReLER, AAII, University of Technology Sydney 2 ReLER, CCAI, Zhejiang University", + "bbox": [ + 217, + 200, + 746, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/FengZicai/LSK3DNet", + "bbox": [ + 316, + 219, + 643, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 310, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous systems need to process large-scale, sparse, and irregular point clouds with limited compute resources. Consequently, it is essential to develop LiDAR perception methods that are both efficient and effective. Although naively enlarging 3D kernel size can enhance performance, it will also lead to a cubically-increasing overhead. Therefore, it is crucial to develop streamlined 3D large kernel designs that eliminate redundant weights and work effectively with larger kernels. In this paper, we propose an efficient and effective Large Sparse Kernel 3D Neural Network (LSK3DNet) that leverages dynamic pruning to amplify the 3D kernel size. Our method comprises two core components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the beginning to learn a large sparse 3D kernel. It not only boosts performance but also significantly reduces model size and computational cost. Moreover, CWS selects the most important channels for 3D convolution during training and subsequently prunes the redundant channels to accelerate inference for 3D vision tasks. We demonstrate the effectiveness of LSK3DNet on three benchmark datasets and five tracks compared with classical models and large kernel designs. Notably, LSK3DNet achieves the state-of-the-art performance on SemanticKITTI (i.e., $75.6\\%$ on single-scan and $63.4\\%$ on multi-scan), with roughly $40\\%$ model size reduction and $60\\%$ computing operations reduction compared to the naive large 3D kernel model.", + "bbox": [ + 76, + 301, + 473, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 207, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous systems are equipped with 3D LiDAR sensors and data processing platforms. The LiDAR sensor generates point clouds, which serve as the input for the processing platform. The platform performs perception tasks such as semantic segmentation [1-3] and object detection [4-6], providing a foundation for motion planning and decision-making [7] in autonomous systems. On the one hand, the processing platform is required to handle large-scale, sparse,", + "bbox": [ + 75, + 752, + 470, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c91ab8c87e5f66c9850c6eed9de25dc8b70cbfbda7bdbd1fa639c81f031fa7fe.jpg", + "image_caption": [ + "Figure 1. Illustrations on Spatial-wise Group Convolution [18], SDS, and CWS. The spatial dimensions $K_{s}$ (i.e., $K_{1}, K_{2}, K_{3}$ ) and channel dimensions $(D_{in} / D_{out})$ are shown. Spatial-wise Group Convolution shares the weights within each spatial group during training, leading to redundant model weights. In contrast, SDS removes non-salient weights and redundancies that are not sensitive to the input in each group, while CWS eliminates redundant weights in a channel-wise manner (Sec. 1)." + ], + "image_footnote": [], + "bbox": [ + 511, + 266, + 879, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and irregular point clouds. On the other hand, the platform's computing resources are limited. Therefore, it is crucial to develop effective and efficient LiDAR perception methods. Point-based methods [1, 8, 9] rely on computationally expensive, memory inefficient, and time-consuming point sampling strategy [10]. In contrast, sparse convolution [11-14] has been widely adopted for processing largescale point clouds up to $160 \\times 160 \\times 20$ meters [5, 11, 12, 15-17]. Nevertheless, the application of such a technique in autonomous systems has also been hindered by limited computational resources [16]. PV-KD [19] attempts to reduce the model size with knowledge distillation, but its performance is bounded by its teacher model [15]. In opposition, large kernels can enhance performance through the advantages of large receptive fields [18, 20, 21]. However, naive 3D large kernels will significantly increase computational burdens. LargeKernel3D [18] explores large 3D kernels with Spatialwise Group Convolution to circumvent optimization and efficiency problems caused by naive 3D large kernels. However, LargeKernel3D has a lot of redundant model weights during training as it shares the weights in each spatial group. Moreover, the performance of LargeKernel3D [18] drops", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14916", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/daef8b8a730e4f609e2769bbbbe105362a9b438ae6cc611440a68920d726625e.jpg", + "image_caption": [ + "Figure 2. Performance (mIoU) vs. Inference Speed (FPS) on SemanticKITTI [25] single-scan challenge (Sec.1)." + ], + "image_footnote": [], + "bbox": [ + 86, + 84, + 464, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "when scaling up the kernel size over $7 \\times 7 \\times 7$ .", + "bbox": [ + 75, + 295, + 380, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose a Large Sparse Kernel 3D Neural Network (LSK3DNet), which is an efficient and effective 3D perception model. LSK3DNet incorporates two key components, namely Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). These components enhance performance on perception tasks and overcome the challenges of high computational costs. SDS enlarges 3D kernels and reduces model parameters by learning sparse 3D kernels from scratch. CWS increases the model width and learns channel-wise importance during training, resulting in accelerated inference by pruning redundant channels [26]. As shown in Fig.1, SDS can prune 3D kernels to remove redundant weights, while CWS can eliminate redundant channels in an online manner. Our SDS and CWS complement each other following a guideline of \"using spatial sparse groups, expanding width without more parameters\", in contrast to \"using sparse groups, expanding more\" [21].", + "bbox": [ + 73, + 311, + 470, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, SDS implements a remove-and-regrow process within each spatial group and preserves the sparsity in each group during dynamic training. This stands in stark contrast to the 2D approach [21], as it achieves dynamic sparsity in depth-wise groups. SDS allows to scale up the receptive fields with large kernel sizes, easily reaching or surpassing $9 \\times 9 \\times 9$ , thereby achieving higher performance. In addition, CWS selectively identifies salient channels during training. It speeds up inference on the 3D vision tasks by pruning non-salient channel parameters. In this way, LSK3DNet can benefit from expanded width to achieve enhanced performance, but maintain the original model size during inference. Reducing the models' complexity (e.g., model size) is a big benefit and a straightforward way to make them deployable on resource-constrained devices [27]. We conduct experiments on three benchmark datasets and five tracks. Our method achieves better performance compared to state-of-the-art methods [15, 23, 24] on SemanticKITTI [25] but with faster inference speed (Fig.2). Moreover, LSK3DNet outperforms the prior 3D large kernel method of LargeKernel3D [18] on ScanNet v2 [28] and KITTI [29].", + "bbox": [ + 73, + 568, + 470, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a nutshell, our contributions are as follows:", + "bbox": [ + 96, + 886, + 408, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose LSK3DNet to enhance performance in perception tasks and mitigate high computational costs, surpassing state-of-the-art models while reducing model size by $40\\%$ and computing operations by $60\\%$ .", + "- We propose SDS to scale up 3D kernels. It learns large sparse kernels by dynamically pruning and regrowing weights from scratch, thereby reducing model size and computational operations.", + "- We develop CWS to improve performance by expanding the width. CWS assesses the importance of channels during training and speeds up inference by pruning redundant channel-wise parameters." + ], + "bbox": [ + 500, + 90, + 893, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 297, + 640, + 313 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-Kernel 3D Models. In 3D vision, research on 3D large kernel is very limited. LargeKernel3D [18] has demonstrated that sizeable kernels can be successfully employed and bring positive results for 3D networks. Spatial-wise Group Convolution [18] enables to achieve a kernel size of $7 \\times 7 \\times 7$ . However, it shares the weights within each spatial group during training, leading to redundant model weights (See Fig. 1). Moreover, the performance of LargeKernel3D drops when scaling up the kernel size over $7 \\times 7 \\times 7$ .", + "bbox": [ + 496, + 321, + 893, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-Kernel 2D CNN Models. In the 2010s, various large-kernel settings were investigated. LR-Net [30], Inceptions [31], and GCNs [32] explored 2D large kernels of $7 \\times 7$ , $11 \\times 11$ , and $15 \\times 15$ respectively. Due to the widespread adoption [33, 34] of VGG [35], research into large kernels was largely overlooked in favour of multiple smaller kernels ( $1 \\times 1$ or $3 \\times 3$ ) to obtain a larger receptive field [36-38] during the past decade. Recently, certain studies have reintroduced large kernels in CNNs. RepLKNet [20] examines the effects of large kernels in CNNs, and for the first time it is able to increase kernel size to $31 \\times 31$ . It achieves comparable results to those of Swin Transformer [39]. Following \"use sparse groups, expand more\", SLaK [21] has achieved an impressive kernel size of $51 \\times 51$ . SLaK has to strike a trade-off between sparsity and model width since increased model width leads to an increase in model size, a phenomenon that becomes even more pronounced in 3D vision. However, CWS effectively decouples the two objectives of improved performance and reduced model size, enabling expanded width without increasing the model size.", + "bbox": [ + 496, + 460, + 893, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Backbones. Generally speaking, there are three main-streams of 3D backbones: i) Point-based methods [1, 8, 9, 40-55] approximate a permutation-invariant set function, and directly learn features from raw point clouds. Researchers have observed an unsatisfactory outcome for large-scale environments [56]. ii) Projection-based methods transform unstructured points into regular 2D images and thus employ CNNs [57-62]. However, the projection could leave out critical geometric details and cause unavoidable informa", + "bbox": [ + 496, + 763, + 895, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14917", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion loss. iii) Voxel-based methods have used sparse convolution [11, 12] to design more powerful networks for various 3D tasks [15, 63]. Sparse convolution facilitates convolution operations only on non-void voxels. Recently, fusion-based methods [3, 16, 23] have become upsurge. The point-based high-resolution branch mitigates the performance degradation, which is caused by aggressive downsampling [16, 64] (i.e., regular sparse convolution) of the voxel branch.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Sparse Training (DST). DST can train sparse neural networks from scratch, resulting in both a speedy training and prediction procedure. During training, DST [65-73] alters the location of non-zero weights with per predefined rules, thereby creating a sparse representation and cutting down the number of calculations. This paradigm commences without prior knowledge and simultaneously refines the non-zero locations and weights. The attractive aspect of DST is that it is sparse from the beginning, resulting in lower FLOPs and memory consumption for training and inference compared to a dense model. Generally, the pruning process [74] can be completed either by threshold-based pruning [65, 66, 70, 75] or by magnitude-based pruning [67, 76]. In addition, new weights are regrown with randomness growing [66, 70, 75], momentum growing [67], and gradient-based growing [68, 69, 76, 77]. Here, we efficiently integrate dynamic sparsity with 3D kernels in large kernel 3D networks. This design satisfies the anticipation for a larger parameter space, as a wide-ranging exploration of the parameter space is of significance to DST [71, 73, 78].", + "bbox": [ + 75, + 212, + 473, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 523, + 212, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first formulate the problem in Sec. 3.1, then move on to a concise overview of submanifold sparse convolution in Sec. 3.2. Afterwards, the details of Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS) are described in Sec. 3.3 and 3.4. Lastly, network architecture is outlined in Sec. 3.5.", + "bbox": [ + 75, + 547, + 472, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 650, + 279, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For 3D perception tasks, suppose that the input is a set of unordered point $\\mathcal{X} = \\{x_{i}\\in \\mathbb{R}^{D_{in}}:i = 1,2,\\ldots ,N\\}$ , and $D_{in}$ is the dimensionality of the input features. $\\mathcal{L}$ is a set of object/point labels, which varies according to different datasets. For segmentation task, LSK3DNet assigns each point $x_{i}$ with a predicted label $y_{i}$ , resulting in a point-wise prediction set $\\mathcal{Y} = \\{y_i:i = 1,2,\\dots ,N\\}$ . In voxel-based methods, point clouds are converted to voxel representations. We denote the input sparse tensor with $X_{s}^{in} = [C,F]$ . Coordinate matrix $C\\in \\mathbb{N}_{+}^{N\\times 3}$ includes 3-dimensional discrete coordinates, and feature matrix $F\\in \\mathbb{R}^{N\\times D_{in}}$ has $D_{in}$ -dimensional features.", + "bbox": [ + 75, + 671, + 472, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Review of Submanifold Sparse Convolution", + "text_level": 1, + "bbox": [ + 76, + 848, + 447, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regular sparse convolution has widespread use in U-Net type 3D networks [11, 12]. However, it significantly reduces", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the sparsity level of point data [64], obfuscates feature distinctions [64], and leads to low-resolution and indistinguishable small objects [16]. In our LSK3DNet, we discard regular sparse convolution and exclusively employ submanifold sparse convolution for feature extraction.", + "bbox": [ + 496, + 90, + 893, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a 3-dimensional space $C \\in \\mathbb{N}_+^{N \\times 3}$ , the $D_{in}$ -dimensional input feature at $c \\in C$ can be denoted as $x_c^{in} \\in \\mathbb{R}^{D_{in}}$ , and the 3D kernel is represented by $\\pmb{W} \\in \\mathbb{R}^{K^3 \\times D_{out} \\times D_{in}}$ . We divide the kernel into $K^3$ spatial weights, each with a size of $D_{out} \\times D_{in}$ and denote the divided weights as $\\pmb{W}_i$ . The submanifold sparse convolution is formulated as follows:", + "bbox": [ + 496, + 166, + 893, + 260 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {c} ^ {\\text {o u t}} = \\sum_ {i \\in \\mathcal {V} ^ {D} (c, C)} \\boldsymbol {W} _ {i} x _ {c + i} ^ {i n}, \\quad c \\in C, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 271, + 890, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{c}^{out}$ represents the current voxel on which the submanifold sparse convolution is applied; $W_{i}$ corresponds to $x_{c + i}^{in}$ , which represents the $i$ -th adjacent input voxel $([C_{c + i}, F_{c + i}])$ ; $\\mathcal{V}^D$ is a set of offsets that define the shape of a kernel, and $\\mathcal{V}^D(c, C) = \\{i | c + i \\in C, i \\in \\mathcal{V}^D\\}$ is a set of offsets from the current center $c$ that exist in $C$ [12]. It should be noted that the input coordinates and output coordinates are equivalent; in other words, they share the same $C$ . Due to this restriction, insufficient information flow exists while differentiating the distinct spatial characteristics [64]. Large receptive fields could be a potential solution to this problem.", + "bbox": [ + 496, + 306, + 893, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Spatial-wise Dynamic Sparsity (SDS)", + "text_level": 1, + "bbox": [ + 498, + 484, + 821, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SDS incorporates two essential components: Sparse Kernel Initialization and Sparse Weight Pruning and Growth. SDS is applied to the 3D large kernel layers, while other layers are kept dense. When initialized with Sparse Kernel Initialization, all spatial groups of sparse tensors have the same level of sparsity $s$ , where sparsity refers to the fraction of zeros in sparse kernels. We adjust the sparse position within spatial groups of 3D large kernel layers with an adaptation frequency $f_{a}$ . During training, at regular intervals determined by $f_{a}$ , the adjustable weights in the kernels are modified through a two-phase procedure, which includes sparse weight pruning and sparse weight growth. We maintain a constant number of non-zero parameters (i.e., sparsity $s$ ) throughout the training process.", + "bbox": [ + 496, + 507, + 893, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse Kernel Initialization. Prior research [79] has explored the distribution of sparsity parameters across multiple layers. The proportion and adjustment of sparse weights are regulated by the layer-wise sparsity ratio, which determines how the weights are modified to seek more efficient sparse structures during training. However, this method is not suitable for Spatial-wise Group Convolution, since 3D large kernels are divided into multiple spatial groups. To ensure that each group has non-zero weights, we execute Erdős-Rényi (ER) based sparsity ratios [66, 69, 80] in each spatial group. Dense weights $W_{D}$ are firstly initialized in a standard way (i.e., kaiming uniform initialization [81]). The binary Mask", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14918", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23e4edfdbc22314569127b56f68aed17dd08578253a1621428afa21a61953898.jpg", + "image_caption": [ + "Figure 3. Spatial-wise Dynamic Sparsity. The utilization of SDS enables us to create and train sparse kernel 3D neural networks from the beginning. The sparse weights in each spatial group are firstly initialized by Sparse Kernel Initialization (Eq. (2)), and then regularly altered by discarding the least significant connections and introducing new ones (Eq. (3)). The sparse kernels are steadily improved by this dynamic process, which leads to a more thorough collection of local features. Note that different spatial sparse group has different sparse distribution. Here we take a sparsity of $22\\%$ for example (Sec.3.3 and Alg.1)." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 869, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$M$ is then applied to these dense weights; $M$ is generated using an ER-based method with a sparsity of $s$ . The zero-weight number in $M$ is scaled by $1 - \\frac{D_{in} + D_{out} + K_1^g + K_2^g + K_3^g}{D_{in} \\times D_{out} \\times K_1^g \\times K_2^g \\times K_3^g}$ , where $K_1^g, K_2^g$ , and $K_3^g$ denote the spatial group sizes. Therefore, the sparse weights $\\mathbf{W}_S$ are initialized as follows:", + "bbox": [ + 75, + 287, + 470, + 368 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {D} \\odot \\boldsymbol {M}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 376, + 468, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ represents the Hadamard product.", + "bbox": [ + 75, + 400, + 361, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To design 3D networks with sparse groups, following the recipe of \"using spatial sparse groups, expanding width without more parameters\", we have chosen Spatial-wise Group Convolution rather than Depth-wise Group Convolution. Our experiments have also shown that Depth-wise Group Convolution is unsuitable for large 3D kernels. This aligns with the findings in LargeKernel3D [18]. This experiment is provided in the supplementary materials.", + "bbox": [ + 75, + 415, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sparse Weight Pruning and Growth. Previous research [21] in 2D convolution has illustrated that dynamic sparsity can effectively enlarge kernel sizes and enhance scalability. But Depth-wise Group Convolution is unsuitable for large 3D kernels [18]. Therefore, we adapt sparse weights in each spatial group dynamically to accommodate the data. Specifically, a certain rate of connections (i.e., prune rate $p$ ) with the lowest magnitude are eliminated, and then we generate an equal number of new random connections. The regrow weights are randomly placed at zero positions within each spatial group. This is achieved by modifying the mask $M$ (See Alg. 1). The positions of the eliminated connections and the newly generated connections are denoted as $E$ and $G$ , respectively. The formula for pruning and regrowth can be expressed as follows:", + "bbox": [ + 75, + 537, + 470, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} - \\boldsymbol {E}), \\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} + \\boldsymbol {G}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 771, + 468, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By doing so, the sparsity of each spatial group is kept steady. Moreover, the weights can be adjusted dynamically, allowing for improved local characteristics. Please refer to Fig. 3 for details. Once training is complete, mask $M$ in spatial groups is also recorded. Unlike previous methods [21], we employ the Spatial-wise Group Convolution and partition dynamic sparse processes into separate spatial groups. Pruning", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and growth are conducted independently within each group without interfering with each other.", + "bbox": [ + 498, + 289, + 890, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can assume that weights are changing factors over time. Then, removing the least important weights is akin to the selection phase in natural evolution. Alternatively, the random addition of new weights is analogous to the alteration stage of evolutionary selection [66]. This phenomenon is also analogous to a biological process in the brain during sleep, known as synaptic shrinking. Researchers found that the weakest neural link in the brain weakens during slumber, while the vital neural connections remain how they are before. This shows that one of the main roles of sleeping is to reset the overall synaptic strength [82, 83].", + "bbox": [ + 496, + 320, + 892, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Channel-wise Weight Selection (CWS)", + "text_level": 1, + "bbox": [ + 500, + 498, + 828, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite SLaK [21] employs DST to enlarge the 2D kernel size, it mitigates the performance degradation caused by sparsity in the way of expanding model width. It strives to strike a balance between sparsity and width. However, expanding width leads to the issue of increased computational burden. Therefore, SLaK [21] faces the dilemma that larger sparsity and smaller model size cannot be achieved simultaneously. Instead of naively increasing network width, we propose CWS to decouple sparsity and width. It selectively chooses the most salient channels and obtains improved performance while keeping the original model size during inference [26].", + "bbox": [ + 496, + 522, + 893, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CWS operates in an online mode with a sorting frequency of $f_{s}$ . A single weight sorting cycle involves multiple iterations of Sparse Weight Pruning and Growth, i.e., each cycle of weight sorting is a multiple of the adaptation rate $f_{a}$ . The width factor $w$ determines the extent to which the number of channels is augmented, i.e., extending $D$ -dimension to $w \\times D$ -dimension. Once the expanded model has been created, we consider the $D$ -dimensional small model to be embedded within the $w \\times D$ -dimensional augmented model [26]. When the number of training iterations reaches the sorting condition, the channels are sorted based on their L1 value (i.e., SortChannels), prompting the model to focus on the most relevant channels. During validation, we select the most important $D$ -dimensional channels from $w \\times D$ -dimensional", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14919", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9f46a40a579b595695ed8d8e9f6cc8b9b487f0e3a0b771d8a60e94e07dd79a4.jpg", + "image_caption": [ + "Figure 4. LSK3DNet and LSK Block. (1) LSK3DNet: Point clouds are fed into voxelizer for voxel-wise features. Then we extract features with LSK Block and Point Branch. The final prediction is a point-wise output. Here MLPs is the standard Multi-Layer Perceptrons. (2) Large Sparse Kernel Block (LSK Block) (Sec.3.5)." + ], + "image_footnote": [], + "bbox": [ + 88, + 85, + 883, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "channels [26] (i.e., SelectChannels). The pseudo-code for the entire algorithm is provided in Alg. 1. This operation effectively achieves higher performance while maintaining the parameters within the expected size. This is especially important in deployments where memory usage and computational efficiency are critical factors.", + "bbox": [ + 75, + 246, + 472, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 347, + 282, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segmentation Baseline. U-Net type 3D networks (such as SparseConvNet [11] and MinkUNet [12]) employ aggressive downsampling (i.e., regular sparse convolution) to increase the receptive field but at the cost of reduced resolution. However, with a low resolution, several points or tiny objects could be combined into one single grid and become indistinguishable [16]. So SPVCNN [16] is equipped with a high-resolution point-based branch. Subsequently, 2DPASS [3] further modifies multi-representation branches by omitting the regular sparse convolution. This is because it dilates all sparse features and blurs valuable information, increasing the burden for following layers [64]. Only submanifold sparse convolution is employed for feature extraction in Modified SPVCNN [3]. By limiting the output feature positions to the input positions, submanifold sparse convolution is able to circumvent the computation burden. We use Modified SPVCNN as a baseline for our segmentation network. This 3D network is compact yet powerful and can generate high-resolution representations from sparse point clouds in large-scale scenes. However, one challenge we face is that the restricted area of submanifold sparse convolution limits the information flow and makes it difficult to distinguish different spatial characteristics of the scene. To address this challenge, we introduce LSK Block, which stands for Large Sparse Kernel Block. This block increases the kernel sizes of submanifold sparse convolution and expands the receptive field to facilitate information flow. LSK Block has a standard residual structure that adds the output of identity mapping to that of two stacked large kernel convolutions (Fig.4(2)). Our network does not need the parallel convolutional branch with dilatation convolution [18]. The details of our segmentation network architecture are listed below:", + "bbox": [ + 75, + 368, + 472, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Segmentation Network on SemanticKITTI [25]. LSK-3DNet (See Fig.4(1)) divides the entire scene into voxels, each with a size of $5\\mathrm{cm}$ . It has four scales of $\\{2, 4, 8, 16\\}$ .", + "bbox": [ + 76, + 854, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Pseudo-code of SDS and CWS", + "bbox": [ + 501, + 244, + 764, + 258 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: set adaptation frequency $f_{a}$ , sparsity $s$ ;", + "2: set sorting frequency $f_{s}$ and width factor $w$ ;", + "3: expand model width form $D$ dimension to $w\\times D$ dimension;", + "4: initialize dense model $\\mathbf{W}_D^{w\\times D}$", + "5: initialize sparse layers $\\mathbf{W}_S^{w\\times D}$", + "6: for each training iteration $i$ do \n7: $\\pmb{W}_S^{w\\times D}\\gets \\text{NormalTraining}(\\pmb{W}_S^{w\\times D})$ ;", + "8: if $(i\\% f_{a})$ equals to 0 then", + "9: $\\mathbf{W}_S^{w\\times D}\\gets \\mathbf{W}_S^{w\\times D}\\odot (\\mathbf{M} - \\mathbf{E})$", + "10: $\\pmb{W}_{S}^{w\\times D}\\gets \\pmb{W}_{S}^{w\\times D}\\odot (\\pmb {M} + \\pmb {G})$", + "11: end if", + "12: if $(i\\% f_s)$ equals to 0 then", + "13: $\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})$", + "14: end if", + "15: if validation then", + "16: $\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})$", + "17: $\\overline{\\mathbf{W}_S^D}\\gets$ SelectChannels $(\\overline{\\mathbf{W}_S^{w\\times D}})$", + "18: NormalValidation $(\\mathbf{W}_S^D)$", + "19: end if", + "20: end for" + ], + "bbox": [ + 504, + 262, + 887, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The hidden size $D$ of the entire network is 64. We deploy LSK Blocks in SparseBasicBlock1. Following [3, 15, 85], we employ the weighted cross-entropy loss to optimize point accuracy and utilize the lovasz-softmax [86] loss to maximize the intersection-over-union.", + "bbox": [ + 508, + 551, + 892, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Segmentation Network on ScanNet v2 [28]. This segmentation network has the same architecture as that on SemanticKITTI [25]. The entire scene is split with a voxel size of $2\\mathrm{cm}$ , with scales of $\\{2, 4, 8, 16, 16\\}$ . The hidden size $D$ is 128. Following [87], we take the weighted cross-entropy loss as the objective function.", + "bbox": [ + 500, + 628, + 893, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detection Network on KITTI. We take Voxel R-CNN [5] as a baseline. Specifically, we retain the backbone of Voxel R-CNN and substitute plain sparse convolutional block with LSK Block in the first three stages. Other settings remain the same as [18]. The input spatial shape is $\\{1600, 1408, 41\\}$ , and the channels of stages are $\\{16, 16, 32, 64, 64\\}$ .", + "bbox": [ + 498, + 720, + 893, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 819, + 624, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setups and Implementations", + "text_level": 1, + "bbox": [ + 500, + 842, + 754, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To evaluate our method, we perform experiments", + "bbox": [ + 500, + 864, + 885, + 880 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 517, + 886, + 779, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14920", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d3baf930b7e2dfaf06eb4cebb7e6afe319858ddf4c711d16a2d33ee0dd63f6f9.jpg", + "table_caption": [ + "Table 1. Results of LSK3DNet on SemanticKITTI [25] single-scan test (Sec. 4.2). Regarding input data format, P denotes points, V represents voxelizations, R signifies range images, and 2DPASS incorporates additional 2D data. \"ms/s\" means \"milliseconds per scene\"." + ], + "table_footnote": [], + "table_body": "
MethodsInputmIoUroadsidewalkparkingother-groundbuildingcartruckbicyclemotorcycleother-vehiclevegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraffic signspeed (ms/s)
PointNet++ [41]P20.172.041.818.75.662.353.70.91.90.20.246.513.830.00.91.00.016.96.08.95900
TangentConv [57]R40.983.963.933.415.483.490.815.22.716.512.179.549.358.123.028.48.149.035.828.53000
PolarNet [62]R54.390.874.461.721.790.093.822.940.330.128.584.065.567.843.240.25.661.351.857.562
RandLA-Net [10]P55.990.574.061.824.589.794.243.929.832.239.183.863.668.648.447.49.460.451.050.7880
SqueezeSegV3 [61]R55.991.774.863.426.489.092.529.638.736.533.082.058.765.445.646.220.159.449.658.9238
KPConv [42]P58.890.372.761.331.590.595.033.430.242.544.384.869.269.161.561.611.864.256.447.4-
JS3C-Net [22]V66.088.972.161.931.992.595.854.359.352.946.084.569.867.969.565.439.970.860.768.7471
SPVNAS [16]PV67.090.275.467.621.891.697.256.650.650.458.086.173.471.067.467.150.366.964.367.3259
Cylinder3D [15]V68.992.277.065.032.390.797.150.867.663.858.585.672.569.873.769.248.066.562.466.2131
RPVNet [23]RPV70.393.480.770.333.393.597.644.268.468.761.186.575.171.775.974.443.472.164.861.4168
\\( (AF)^2-S3Net \\)[84]V70.892.076.266.845.892.594.340.263.081.440.078.668.063.176.481.777.769.664.073.3-
PV-KD [19]V71.291.877.570.941.092.497.053.567.969.360.286.573.871.975.173.550.569.464.965.876
2DPASS [3]PV72.989.774.767.440.093.597.061.163.663.461.586.273.971.077.981.374.172.965.070.462
SphereFormer [24]V74.891.878.269.741.393.897.559.670.170.567.786.775.172.479.080.475.372.866.872.9123
LSK3DNetPV75.692.278.970.241.892.797.361.071.475.664.286.472.771.981.280.685.272.067.074.689
", + "bbox": [ + 80, + 119, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/22ff579cb6e7143c91eb2822f8801d1004c15611f6025d86f5a6a23169fbbe02.jpg", + "table_caption": [ + "Table 2. Results of LSK3DNet on SemanticKITTI [25] multi-scan test set. The arrow below classes indicate moving classes (Sec.4.2)." + ], + "table_footnote": [], + "table_body": "
MethodInputmIoUAcccartruckother-vehiclepersonbicyclistmotorcyclist
LatticeNet [89]P45.289.354.83.50.649.944.664.3
TLSeg [90]R47.089.668.22.112.440.442.812.9
KPCnv [42]P51.289.369.45.84.767.567.447.2
Cylinder3D [15]V52.591.074.90.00.165.768.311.9
(AF)2-S3Net [84]V56.988.165.35.63.967.666.459.6
2DPASS [3]PV62.491.482.116.13.880.371.273.1
LSK3DNetPV63.492.284.47.240.977.469.972.1
", + "bbox": [ + 80, + 402, + 467, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on three benchmark datasets and five tracks:", + "bbox": [ + 76, + 550, + 370, + 563 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SemanticKITTI [25] has 43,551 traffic point cloud scenes with fine annotations, split into 19,130/4,071/20,350 scenes for train/val/test. The dataset has 28 semantic classes, but only 19 classes are used for single-scan track and 25 classes for multi-scan track.", + "- ScanNet v2 [28] contains 1,201/312/100 indoor scenes for train, val, and test splits, respectively. There are 20 semantic categories for ScanNet20 track and 200 categories for ScanNet200 track.", + "- KITTI [29] has 7,481/7,518 samples for train and test. We follow the frequently used train/val split [88] to divide the training samples into train split (3,712 samples) and val split (3,769 samples)." + ], + "bbox": [ + 76, + 566, + 468, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Testing Details. We utilize AdamW optimizer with OneCycleLR scheduler starting with a learning rate of 5e-3. Data augmentation is also employed, such as random flipping, scaling, rotation around the gravity axis, spatial translation. We apply instance CutMix [23] and Test Time Augmentation (TTA) [3] to the SemanticKITTI test benchmark, and enhance the model with extra training epochs.", + "bbox": [ + 75, + 763, + 468, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We employ mean class intersection over union (mIoU) and overall accuracy (Acc) metrics as the evaluation", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/29cfb9db6b62fa0fd080e13025481c4ec9b016ced3e3c103bccadac0b4da5ac1.jpg", + "table_caption": [ + "Table 3. Results of LSK3DNet and other tate-of-the-art methods on ScanNet v2 [28] (Sec.4.2). S-Net stands for ScanNet." + ], + "table_footnote": [], + "table_body": "
MethodInputS-Net20 ValS-Net20 TestS-Net200 Val
PointNet++ [41]P53.555.7-
3DMV [91]P-48.4-
PanopticFusion [92]P-52.9-
PointCNN [93]P-45.8-
PointConv [94]P61.066.6-
JointPointBased [95]P69.263.4-
PointASNL [96]P63.566.6-
SegGCN [97]P-58.9-
RandLA-Net [85]P-64.5-
KPConv [42]P69.268.6-
JSENet [98]P-69.9-
FusionNet [99]P-68.8-
Point Transformer [8]P70.6--
Fast Point Transformer [100]P72.1--
Stratified Transformer [9]P74.373.7-
Point Transformer v2 [87]P75.475.231.9
SparseConvNet [11, 87]V69.372.528.8
MinkUNet [12]V72.273.6-
LargeKernel3D [18]V73.573.9-
LSK3DNetPV75.775.533.1
", + "bbox": [ + 504, + 402, + 890, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "criterion for 3D semantic segmentation tasks, as outlined in [25]. What is more, we calculate Average Precision (AP) by recalling 11 positions for 3D object detection.", + "bbox": [ + 496, + 684, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. 3D Semantic Segmentation", + "text_level": 1, + "bbox": [ + 500, + 741, + 743, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on SemanticKITTI. We test LSK3DNet on both single-scan and multi-scan tracks [25]. Tab. 1 presents the quantitative results of SemanticKITTI single-scan track. LSK3DNet outperforms 2DPASS [3] in terms of both mIoU and IoU scores for most categories. Moreover, SDS reduces the model size of naive 3D kernels and CWS keeps the model size within the desired size. So, LSK3DNet has a faster running speed than most prior methods, but it's performance significantly benefits from the large kernels. Tab. 2 shows", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be729abb4e5a68cff3f72f4e54eef46a50380f8bba8f50aceed83fcdc2a7186a.jpg", + "table_caption": [ + "Table 4. Results of 3D object detection methods on the car class of KITTI val set [29] (Sec.4.3)." + ], + "table_footnote": [], + "table_body": "
MethodInput3D AP (IoU=0.7)
EasyModerateHard
VoxelNet [101]V81.9765.4662.85
PointPillars [102]R86.6276.0668.91
SECOND [63]V88.6178.6277.22
Point R-CNN [103]P88.8878.6377.38
Part-Å2[104]V89.4779.4778.54
PV-RCNN [4]PV89.3583.6978.70
Focals Conv [64]V89.5284.9379.18
Voxel R-CNN [5]V89.4184.5278.93
LargeKernel3D [18]V89.5285.0779.32
LSK3DNetV90.1685.6179.53
", + "bbox": [ + 81, + 119, + 468, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the results of SemanticKITTI multi-scan track. The mIoU and Acc are computed over all 25 classes. Due to page limitations, we only report the per-class IoUs for moving categories. Under this challenging setting, LSK3DNet surpasses 2DPASS [3] in both mIoU and Acc. Notably, LSK3DNet achieves state-of-the-art performance on both single-scan and multi-scan tracks in SemanticKITTI. Surpassing over the SOTA 2D-3D method 2DPASS is more valuable, which utilizes both 2D and 3D data while our LSK3DNet takes as input only the 3D data. Visualization results are in the supplementary materials.", + "bbox": [ + 75, + 281, + 468, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ScanNet V2. We compare LSK3DNet with previous state-of-the-art methods on ScanNet v2 [28], a largescale dataset for 3D indoor scene segmentation. ScanNet v2 has two tracks: ScanNet20 and ScanNet200, which use 20 and 200 semantic classes respectively. Tab. 3 presents the quantitative results of our model and other methods on both tracks. LSK3DNet achieves higher performance than previous methods in both tracks. Our performance is superior to transformer-based methods [8, 9, 87, 100], including Point Transformer v2. Moreover, LSK3DNet has a clear advantage over LargeKernel3D [18], improving the mIoU by $2.2\\%$ and $1.6\\%$ on the ScanNet v2 val and test set, respectively.", + "bbox": [ + 75, + 449, + 470, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. 3D Object Detection", + "text_level": 1, + "bbox": [ + 76, + 643, + 267, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have also evaluated the detection performance of LSK3DNet on the val split for car category, as shown in Tab.4. We report the average precision metric for a 3D bounding box with 11 recall positions (Recall 11). Based on Voxel R-CNN [5], we showcase the effectiveness of the LSK Block by comparing it with LargeKernel3D [18], a previous 3D large-kernel method. LSK3DNet achieves better results in three difficulty levels, compared with [5, 18]. Visualization comparison can be found in the supplementary material.", + "bbox": [ + 75, + 667, + 468, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 816, + 240, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first explore the effect of kernel size on performance, and then the overall architecture design. Next, we explain the hyperparameter choices of SDS and CWS. All ablation experiments are performed on the val set of SemancKITTI.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/984dab33ec0b5d1c81b764c08dc843d24ac804b1e8afc92bfa0b46f9dc42d0d3.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 637, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/387962a1b25b33ccc59491e7c9a9241128ba5405d9a38b0c633407e65a69c086.jpg", + "image_caption": [ + "LSK3DNet" + ], + "image_footnote": [], + "bbox": [ + 637, + 90, + 766, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23a56d7cec483053479698b0a06e01a4b776e669bed25882719994f129a3671f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 90, + 805, + 119 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c328e8887fe02723cbdb9d061171dcb40d4bf3e785b50ff9d95df24a6aa8f86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 125, + 803, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5f66fcd74bae5bcd25eb009f7c46a3c84486fed1a096a4e41e18d2e412e1843.jpg", + "image_caption": [ + "Figure 5. Effective Receptive Fields (ERFs) of Baseline and LSK3DNet. LSK3DNet has a larger ERF size. Additionally, we provide visualization of learned sparse kernels, where all weight values have taken the absolute value and normalization operations. The black areas indicate positions with weight values of zero." + ], + "image_footnote": [], + "bbox": [ + 767, + 157, + 805, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d717af427e6491d57e24b9cba473c1aabc5c3cb179725b3025351cbad897f4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 90, + 849, + 119 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adaa3cee053e7c182d16d35fa3d7e4961afef3cbe1a03b4ec267088a7179adb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 125, + 848, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/862822ae03a99c92641c1b2891d9bc1889ae93f2f5c9a847372b8432707ea323.jpg", + "image_caption": [ + "Kernel visualization" + ], + "image_footnote": [], + "bbox": [ + 810, + 156, + 848, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bd89c05d59d5d5eee8c5a27943f9f4108f8744370e623ebc0cd4d5ddd45563db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 852, + 90, + 890, + 119 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3ae74dbbb029ca1258ba8f3db2a1822245430f30cd3e67ee6cc0c29ee3271f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 852, + 125, + 890, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0739faa0b12e9901255b3190f87876a97d6e5f9630662b178b8570b4e3af299.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 852, + 156, + 890, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "zation", + "bbox": [ + 852, + 188, + 880, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kernel Size. We use Modified SPVCNN as the baseline and then explore 3D kernel sizes under two settings: naive dense large kernel and our LSK3DNet. In the former setting, the dense 3D kernel is straightforwardly expanded. In the latter one, LSK3DNet is trained simultaneously with SDS and CWS, meaning that our LSK3DNet can learn a large sparse kernel model from the beginning.", + "bbox": [ + 496, + 276, + 892, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The baseline does not use aggressive downsampling which is common in most U-type 3D networks [11, 12]. However, submanifold sparse convolution with small kernels may lose important information flow, especially for the spatially disconnected features. We solve this problem by enlarging the size of the 3D convolution kernel to obtain a large receptive field. This agrees with the performance improvement in Tab. 5, when the kernel size increases. In contrast to this observation, LargeKernel3D [18] shows the opposite trends; we empirically attribute this to the fact that LargeKernel3D is based on U-type 3D networks (Sec. 3.5). For a better understanding of Effective Receptive Fields (ERFs) size, Fig. 5 illustrates the comparison between the Baseline and LSK3DNet. Compared to the baseline, the high-contribution points of LSK3DNet are distributed over a larger input range, indicating a larger ERF. Additionally, to enhance comprehension of learned kernels, we offer kernel visualization in Fig. 5, providing insights into sparse training.", + "bbox": [ + 496, + 383, + 893, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the \"naive dense large kernel\" of each kernel size, LSK3DNet exhibits superior performance in its corresponding size. LSK3DNet not only leverages SDS to expand its receptive field but also starts from scratch to learn 3D sparse kernels. Concurrently, CWS widens the network during training while pruning redundant channels to accelerate inference. Furthermore, SDS and CWS address the issues of overparameterization and overfitting that arise from simply increasing the kernel's size and expanding the model's width. The best performance of LSK3DNet is achieved at the $9 \\times 9 \\times 9$ size, and the following experiments are conducted with $9 \\times 9 \\times 9$ size. Moreover, due to the extra dimension in 3D kernels compared to 2D kernels, naively enlarging the 3D kernel size will lead to a cubically-increasing overhead. Our LSK3DNet can effectively reduce 19.1M parameters and approximately $60\\%$ of computing operations when compared", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14922", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1933b83cfe8b3851a9b4f36ed3ed7cbcb5406ed61b01e676a8ef168cfdaa2151.jpg", + "table_caption": [ + "Table 5. Segmentation performance of Modified SPVCNN with various large kernel settings. \"Dense\" refers to directly enlarging the kernel size. \"Ours\" refers to training with SDS and CWS (Sec.4.4). The unit for \"Param\" is in million (M), \"FLOPs\" is in billion (G), and \"Speed\" is measured in milliseconds per scene." + ], + "table_footnote": [], + "table_body": "
Kernel SizemIoUParamFLOPsSpeed ↓
Dense (D=64)3,3,365.21.978.436
5,5,565.68.4127.857
7,7,766.222.6381.265
9,9,967.547.91916.393
11,11,1166.687.43192.0134
Dense (1.8×D)5,5,566.227.0444.084
7,7,766.573.11551.2123
9,9,967.2154.73030.0177
11,11,1166.3282.17139.2250
Ours (1.8×D)5,5,566.85.1203.846
7,7,768.113.6249.463
9,9,970.228.8763.689
11,11,1170.152.51847.4116
", + "bbox": [ + 81, + 161, + 465, + 349 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/989d85b2ea0bd58290754fa5cebe359c4fb10f7445bbdc8f4b5e31dce2361b4c.jpg", + "table_caption": [ + "Table 6. Overall architecture design with $9 \\times 9 \\times 9$ size. Here, we report Param, FLOPs, and Speed for the inference model (Sec.4.4)." + ], + "table_footnote": [], + "table_body": "
MethodmIoUParamFLOPsSpeed ↓
Dense (1.0 × D)67.547.91916.393
Dense (1.8 × D)67.2154.73030.0177
SDS (1.8 × D)69.393.02201.4173
CWS (1.8 × D)69.147.91916.393
LSK3DNet (1.8 × D)70.228.8763.689
", + "bbox": [ + 81, + 388, + 465, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the naive large kernel network. This aspect significantly enhances the value of SDS and CWS for 3D kernels.", + "bbox": [ + 75, + 473, + 467, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall Architecture Design. In Tab. 6, Dense $(1.8\\times D)$ does not yield improvement due to overparameterization and overfitting. Although SDS maintains a model width of $1.8\\times D$ during inference, it achieves improved performance due to sparse training and the large receptive field. CWS effectively addresses overfitting by selecting salient channels, thereby enhancing performance. By combining SDS and CWS, LSK3DNet can achieve the best performance while significantly reducing the model size and computational cost. Spatial-wise Dynamic Sparsity. To control the degree of sparse weight adaptation in LSK3DNet, we have adjusted three hyperparameters: the adaptation frequency $f_{a}$ , sparsity rate $s$ , and prune rate $p$ (Sec. 3.3). Adaptation frequency $f_{a}$ indicates how often we update the sparse weights during training. Sparsity rate $s$ specifies how sparse the 3D large kernels are. Prune rate $p$ shows the fraction of updated weights in one adaptation. We have conducted experiments with different values of these hyperparameters and report the results in Tab. 7 and Tab. 8. Based on our empirical findings, we have chosen $f_{a} = 2000$ , $s = 0.4$ , and $p = 0.3$ as the optimal settings for LSK3DNet.", + "bbox": [ + 75, + 505, + 468, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Channel-wise Weight Selection. Another aspect of our method that we investigated is the CWS. This technique involves two hyperparameters: the sorting frequency $f_{s}$ and the width factor $w$ (Sec.3.4). The former determines how frequently we sort and select channels for large kernels during", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c530ed13f6f667372379ae7808642b2cebe3e5d1636b1d759e936ec30e75a9c6.jpg", + "table_caption": [ + "Table 7. Training and inference speed analysis, speed is measured in milliseconds per scene. The kernels have $9 \\times 9 \\times 9$ size, \"T\" and \"T\" represent training and inference speed, respectively (Sec.4.4)." + ], + "table_footnote": [], + "table_body": "
Sparsity sDense (1.0 × D)SDS (1.8 × D)LSK3DNet (1.8 × D)
mIoUT ↓I ↓mIoUT ↓I ↓mIoUT ↓I ↓
067.545193------
0.1---68.548917669.149192
0.2---68.947817669.748291
0.4---69.346317370.246789
0.6---69.045017168.845487
0.8---68.743917167.244387
", + "bbox": [ + 504, + 133, + 890, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/615dc99e1a275577c550e8003abcb1c275d45af942ebcf57b4abf733343eafbc.jpg", + "table_caption": [ + "Table 8. Ablation studies of Spatial-wise Dynamic Sparsity on adaptation frequency $f_{a}$ and prune rate $p$ (Sec.4.4)." + ], + "table_footnote": [], + "table_body": "
Adaptation \\( f_{a} \\)mIoU
10068.8
100069.6
200070.2
300070.0
400069.8
", + "bbox": [ + 516, + 277, + 669, + 352 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4df0f3522bcbc6a60366d2a6aa0fac375594c59de5410816b0195d1a9aac8583.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pruning pmIoU
0.1069.9
0.2070.1
0.3070.2
0.5070.0
0.7069.7
", + "bbox": [ + 710, + 277, + 864, + 352 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7f4365b04285acf4ae71985472c452b52104e711a69b81e1cdb850e4d0282454.jpg", + "table_caption": [ + "Table 9. Ablation studies of Channel-wise Weight Selection on sorting frequency $f_{s}$ and width factor $w$ (Sec.4.4)." + ], + "table_footnote": [], + "table_body": "
Sorting fsmIoU
1 × fa68.2
2 × fa69.1
4 × fa69.8
6 × fa70.2
10 × fa69.9
", + "bbox": [ + 529, + 393, + 673, + 470 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c2cafc85ae976f15903f4cf26290b836edec2eb72bcf1344ac847c1a5d6103b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Width wmIoU
1.1×68.4
1.5×69.6
1.8×70.2
2.1×69.9
2.5×70.1
", + "bbox": [ + 712, + 393, + 864, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training. It is expressed as a multiple of $f_{a}$ , the adaptation frequency, meaning that we perform channel sorting after a certain number of weight adaptation cycles. When $f_{s} = 6 \\times f_{a}$ , the optimal performance is achieved. The latter controls the times of network width compared to the target model. We find that increasing the width factor up to $1.8 \\times$ improves the model's performance, but beyond that point, there is no significant gain. Therefore, we opt for a width of $1.8 \\times$ to avoid more extra compute resources that a wider width would require during training.", + "bbox": [ + 496, + 478, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 645, + 617, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Spatial-wise Dynamic Sparsity to scale up 3D kernels beyond $9 \\times 9 \\times 9$ , which prunes the volumetric weight and reduces the parameter size of large kernel layers. Our LSK3DNet can benefit from a large receptive field without increasing the computational cost compared to a naive 3D large kernel. Channel-wise Weight Selection expands the model width during training, and then sorts and selects important channels during validation to get a model of the expected size. In this way, we achieve \"using spatial sparse groups, expanding width without more parameters\". We evaluate our method on the SemanticKITTI and achieve state-of-the-art performance. Our LSK3DNet also surpasses previous 3D large kernel methods on ScanNet v2 and KITTI. Acknowledgments: This work was supported by a China Scholarship Council (CSC) scholarship.", + "bbox": [ + 496, + 671, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoran Zhou, Yidan Feng, Mingsheng Fang, Mingqiang Wei, Jing Qin, and Tong Lu. Adaptive graph convolution for point cloud analysis. In ICCV, 2021. 1, 2", + "[2] Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud representation learning for 3d analysis. In ICCV, 2023.", + "[3] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In ECCV, 2022. 1, 2, 3, 5, 6, 7", + "[4] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In CVPR, 2020. 1, 7", + "[5] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, 2021. 1, 5, 7", + "[6] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In CVPR, 2024. 1", + "[7] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In AAAI, 2024. 1", + "[8] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 6, 7", + "[9] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1, 2, 6, 7", + "[10] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 6", + "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1, 3, 5, 6, 7", + "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 3, 5, 6, 7", + "[13] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022.", + "[14] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1", + "[15] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, 2021. 1, 2, 3, 5, 6", + "[16] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "architectures with sparse point-voxel convolution. In ECCV, 2020.1,2,3,5,6", + "[17] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1", + "[18] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. CVPR, 2023. 1, 2, 4, 5, 6, 7", + "[19] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In CVPR, 2022. 1, 2, 6", + "[20] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1, 2", + "[21] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 1, 2, 4", + "[22] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, 2021. 2, 6", + "[23] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In ICCV, 2021. 2, 3, 6", + "[24] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, 2023. 2, 6", + "[25] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2, 5, 6", + "[26] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. In ICLR, 2022. 2, 4, 5", + "[27] Vidya Kamath and A Renuka. Deep learning based object detection for resource constrained devices-systematic review, future trends and challenges ahead. Neurocomputing, 2023. 2", + "[28] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7", + "[29] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6, 7", + "[30] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In ICCV, 2019. 2", + "[31] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 2", + "[32] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters-improve semantic segmentation by global convolutional network. In CVPR, 2017. 2" + ], + "bbox": [ + 509, + 93, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14924", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, 2019. 2", + "[34] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI, 2022. 2", + "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2", + "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2", + "[37] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017.", + "[38] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 2", + "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2", + "[40] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2", + "[41] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 6", + "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 6", + "[43] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021.", + "[44] Siqi Fan, Qiulei Dong, Fenghua Zhu, Yisheng Lv, Peijun Ye, and Fei-Yue Wang. Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In CVPR, 2021.", + "[45] Haoxi Ran, Wei Zhuo, Jun Liu, and Li Lu. Learning inner-group relations on point clouds. In ICCV, 2021.", + "[46] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021.", + "[47] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, 2019.", + "[48] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gcn for fast and scalable point cloud learning. In CVPR, 2020.", + "[49] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019.", + "[50] Yanchao Lian, Tuo Feng, and Jinliu Zhou. A dense point-net++ architecture for 3d point cloud semantic segmentation. In IGARSS, 2019." + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In CVPR, 2021.", + "[52] Hehe Fan, Xin Yu, Yi Yang, and Mohan Kankanhalli. Deep hierarchical representation of point cloud videos via spatiotemporal decomposition. IEEE TPAMI, 44(12):9918-9930, 2021.", + "[53] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE TPAMI, 45(2):2181-2192, 2022.", + "[54] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and Mohan Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. In ICLR, 2020.", + "[55] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In CVPR, 2021. 2", + "[56] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In CVPR, 2021. 2", + "[57] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2, 6", + "[58] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018.", + "[59] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, 2019.", + "[60] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: fast, uncertainty-aware semantic segmentation of lidar point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020.", + "[61] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezeseqv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, 2020. 6", + "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, 2020. 2, 6", + "[63] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3, 7", + "[64] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, 2022. 3, 5, 7", + "[65] Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep rewiring: Training very sparse deep networks. arXiv preprint arXiv:1711.05136, 2017. 3", + "[66] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018. 3, 4" + ], + "bbox": [ + 509, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14925", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Tim Dettmers and Luke Zettlemoyer. Sparse networks from scratch: Faster training without losing performance. arXiv preprint arXiv:1907.04840, 2019.3", + "[68] Shiwei Liu, Decebal Constantin Mocanu, Amarsagar Reddy Ramapuram Matavalam, Yulong Pei, and Mykola Pechenizkiy. Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware. Neural Computing and Applications, 33:2589-2604, 2021. 3", + "[69] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Elsen. Rigging the lottery: Making all tickets winners. In ICML, 2020. 3", + "[70] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In ICML, 2019. 3", + "[71] Siddhant Jayakumar, Razvan Pascanu, Jack Rae, Simon Osindero, and Erich Elsen. Top-kast: Top-k always sparse training. In NeurIPS, 2020. 3", + "[72] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In NeurIPS, 2021.", + "[73] Shiwei Liu, Decebal Constantin Mocanu, Yulong Pei, and Mykola Pechenizkiy. Selfish sparse rn training. In ICML, 2021. 3", + "[74] Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Enhao Gong, Shijian Tang, Erich Elsen, Peter Vajda, Manohar Paluri, John Tran, et al. Dsd: Dense-sparse-dense training for deep neural networks. In ICLR, 2016. 3", + "[75] Decebal Constantin Mocanu et al. Network computations in artificial intelligence. Technische Universiteit Eindhoven, 2017. 3", + "[76] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Nest: A neural network synthesis tool based on a grow-and-prune paradigm. IEEE Transactions on Computers, 68(10):1487-1497, 2019. 3", + "[77] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Grow and prune compact, fast, and accurate lstms. IEEE Transactions on Computers, 69(3):441-452, 2019. 3", + "[78] Md Aamir Raihan and Tor Aamodt. Sparse weight activation training. In NeurIPS, 2020. 3", + "[79] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018. 3", + "[80] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Li Shen, Decebal Constantin Mocanu, Zhangyang Wang, and Mykola Pechenizkiy. The unreasonable effectiveness of random pruning: Return of the most naive baseline for sparse training. In ICLR, 2022. 3", + "[81] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3", + "[82] Graham H Diering, Raja S Nirujogi, Richard H Roth, Paul F Worley, Akhilesh Pandey, and Richard L Huganir. *Homer1a* drives homeostatic scaling-down of excitatory synapses during sleep. *Science*, 355(6324):511-515, 2017. 4", + "[83] Luisa De Vivo, Michele Bellesi, William Marshall, Eric A Bushong, Mark H Ellisman, Giulio Tononi, and Chiara Cirelli. Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science, 355(6324):507-510, 2017. 4" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, 2021. 6", + "[85] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 5, 6", + "[86] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 5", + "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 5, 6, 7", + "[88] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In ECCV, 2020. 6", + "[89] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6", + "[90] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 3DV, 2020. 6", + "[91] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation. In ECCV, 2018. 6", + "[92] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 6", + "[93] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, 2018. 6", + "[94] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 6", + "[95] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 6", + "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 6", + "[97] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 6", + "[98] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 6", + "[99] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 6", + "[100] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 6, 7" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14926", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[101] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 7", + "[102] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 7", + "[103] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 7", + "[104] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 7" + ], + "bbox": [ + 78, + 90, + 470, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14927", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_model.json b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c93290368c9f05f66c74b07f2369fb9036bafe --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_model.json @@ -0,0 +1,3073 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.059, + 0.131, + 0.911, + 0.152 + ], + "angle": 0, + "content": "LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.18, + 0.677, + 0.2 + ], + "angle": 0, + "content": "Tuo Feng\\(^{1}\\), Wenguan Wang\\(^{2}\\), Fan Ma\\(^{2*}\\), Yi Yang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.201, + 0.748, + 0.217 + ], + "angle": 0, + "content": "1 ReLER, AAII, University of Technology Sydney 2 ReLER, CCAI, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.317, + 0.22, + 0.645, + 0.233 + ], + "angle": 0, + "content": "https://github.com/FengZicai/LSK3DNet" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.312, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.474, + 0.71 + ], + "angle": 0, + "content": "Autonomous systems need to process large-scale, sparse, and irregular point clouds with limited compute resources. Consequently, it is essential to develop LiDAR perception methods that are both efficient and effective. Although naively enlarging 3D kernel size can enhance performance, it will also lead to a cubically-increasing overhead. Therefore, it is crucial to develop streamlined 3D large kernel designs that eliminate redundant weights and work effectively with larger kernels. In this paper, we propose an efficient and effective Large Sparse Kernel 3D Neural Network (LSK3DNet) that leverages dynamic pruning to amplify the 3D kernel size. Our method comprises two core components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the beginning to learn a large sparse 3D kernel. It not only boosts performance but also significantly reduces model size and computational cost. Moreover, CWS selects the most important channels for 3D convolution during training and subsequently prunes the redundant channels to accelerate inference for 3D vision tasks. We demonstrate the effectiveness of LSK3DNet on three benchmark datasets and five tracks compared with classical models and large kernel designs. Notably, LSK3DNet achieves the state-of-the-art performance on SemanticKITTI (i.e., \\(75.6\\%\\) on single-scan and \\(63.4\\%\\) on multi-scan), with roughly \\(40\\%\\) model size reduction and \\(60\\%\\) computing operations reduction compared to the naive large 3D kernel model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.73, + 0.208, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.471, + 0.874 + ], + "angle": 0, + "content": "Autonomous systems are equipped with 3D LiDAR sensors and data processing platforms. The LiDAR sensor generates point clouds, which serve as the input for the processing platform. The platform performs perception tasks such as semantic segmentation [1-3] and object detection [4-6], providing a foundation for motion planning and decision-making [7] in autonomous systems. On the one hand, the processing platform is required to handle large-scale, sparse," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.267, + 0.88, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Figure 1. Illustrations on Spatial-wise Group Convolution [18], SDS, and CWS. The spatial dimensions \\( K_{s} \\) (i.e., \\( K_{1}, K_{2}, K_{3} \\)) and channel dimensions \\( (D_{in} / D_{out}) \\) are shown. Spatial-wise Group Convolution shares the weights within each spatial group during training, leading to redundant model weights. In contrast, SDS removes non-salient weights and redundancies that are not sensitive to the input in each group, while CWS eliminates redundant weights in a channel-wise manner (Sec. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "and irregular point clouds. On the other hand, the platform's computing resources are limited. Therefore, it is crucial to develop effective and efficient LiDAR perception methods. Point-based methods [1, 8, 9] rely on computationally expensive, memory inefficient, and time-consuming point sampling strategy [10]. In contrast, sparse convolution [11-14] has been widely adopted for processing largescale point clouds up to \\(160 \\times 160 \\times 20\\) meters [5, 11, 12, 15-17]. Nevertheless, the application of such a technique in autonomous systems has also been hindered by limited computational resources [16]. PV-KD [19] attempts to reduce the model size with knowledge distillation, but its performance is bounded by its teacher model [15]. In opposition, large kernels can enhance performance through the advantages of large receptive fields [18, 20, 21]. However, naive 3D large kernels will significantly increase computational burdens. LargeKernel3D [18] explores large 3D kernels with Spatialwise Group Convolution to circumvent optimization and efficiency problems caused by naive 3D large kernels. However, LargeKernel3D has a lot of redundant model weights during training as it shares the weights in each spatial group. Moreover, the performance of LargeKernel3D [18] drops" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14916" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.085, + 0.465, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.261, + 0.472, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Performance (mIoU) vs. Inference Speed (FPS) on SemanticKITTI [25] single-scan challenge (Sec.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.381, + 0.311 + ], + "angle": 0, + "content": "when scaling up the kernel size over \\(7 \\times 7 \\times 7\\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.312, + 0.472, + 0.568 + ], + "angle": 0, + "content": "In this study, we propose a Large Sparse Kernel 3D Neural Network (LSK3DNet), which is an efficient and effective 3D perception model. LSK3DNet incorporates two key components, namely Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). These components enhance performance on perception tasks and overcome the challenges of high computational costs. SDS enlarges 3D kernels and reduces model parameters by learning sparse 3D kernels from scratch. CWS increases the model width and learns channel-wise importance during training, resulting in accelerated inference by pruning redundant channels [26]. As shown in Fig.1, SDS can prune 3D kernels to remove redundant weights, while CWS can eliminate redundant channels in an online manner. Our SDS and CWS complement each other following a guideline of \"using spatial sparse groups, expanding width without more parameters\", in contrast to \"using sparse groups, expanding more\" [21]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.569, + 0.472, + 0.886 + ], + "angle": 0, + "content": "Specifically, SDS implements a remove-and-regrow process within each spatial group and preserves the sparsity in each group during dynamic training. This stands in stark contrast to the 2D approach [21], as it achieves dynamic sparsity in depth-wise groups. SDS allows to scale up the receptive fields with large kernel sizes, easily reaching or surpassing \\(9 \\times 9 \\times 9\\), thereby achieving higher performance. In addition, CWS selectively identifies salient channels during training. It speeds up inference on the 3D vision tasks by pruning non-salient channel parameters. In this way, LSK3DNet can benefit from expanded width to achieve enhanced performance, but maintain the original model size during inference. Reducing the models' complexity (e.g., model size) is a big benefit and a straightforward way to make them deployable on resource-constrained devices [27]. We conduct experiments on three benchmark datasets and five tracks. Our method achieves better performance compared to state-of-the-art methods [15, 23, 24] on SemanticKITTI [25] but with faster inference speed (Fig.2). Moreover, LSK3DNet outperforms the prior 3D large kernel method of LargeKernel3D [18] on ScanNet v2 [28] and KITTI [29]." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.409, + 0.9 + ], + "angle": 0, + "content": "In a nutshell, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "- We propose LSK3DNet to enhance performance in perception tasks and mitigate high computational costs, surpassing state-of-the-art models while reducing model size by \\(40\\%\\) and computing operations by \\(60\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.157, + 0.894, + 0.218 + ], + "angle": 0, + "content": "- We propose SDS to scale up 3D kernels. It learns large sparse kernels by dynamically pruning and regrowing weights from scratch, thereby reducing model size and computational operations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.222, + 0.894, + 0.283 + ], + "angle": 0, + "content": "- We develop CWS to improve performance by expanding the width. CWS assesses the importance of channels during training and speeds up inference by pruning redundant channel-wise parameters." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.298, + 0.642, + 0.314 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.46 + ], + "angle": 0, + "content": "Large-Kernel 3D Models. In 3D vision, research on 3D large kernel is very limited. LargeKernel3D [18] has demonstrated that sizeable kernels can be successfully employed and bring positive results for 3D networks. Spatial-wise Group Convolution [18] enables to achieve a kernel size of \\(7 \\times 7 \\times 7\\). However, it shares the weights within each spatial group during training, leading to redundant model weights (See Fig. 1). Moreover, the performance of LargeKernel3D drops when scaling up the kernel size over \\(7 \\times 7 \\times 7\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.763 + ], + "angle": 0, + "content": "Large-Kernel 2D CNN Models. In the 2010s, various large-kernel settings were investigated. LR-Net [30], Inceptions [31], and GCNs [32] explored 2D large kernels of \\(7 \\times 7\\), \\(11 \\times 11\\), and \\(15 \\times 15\\) respectively. Due to the widespread adoption [33, 34] of VGG [35], research into large kernels was largely overlooked in favour of multiple smaller kernels (\\(1 \\times 1\\) or \\(3 \\times 3\\)) to obtain a larger receptive field [36-38] during the past decade. Recently, certain studies have reintroduced large kernels in CNNs. RepLKNet [20] examines the effects of large kernels in CNNs, and for the first time it is able to increase kernel size to \\(31 \\times 31\\). It achieves comparable results to those of Swin Transformer [39]. Following \"use sparse groups, expand more\", SLaK [21] has achieved an impressive kernel size of \\(51 \\times 51\\). SLaK has to strike a trade-off between sparsity and model width since increased model width leads to an increase in model size, a phenomenon that becomes even more pronounced in 3D vision. However, CWS effectively decouples the two objectives of improved performance and reduced model size, enabling expanded width without increasing the model size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.897, + 0.902 + ], + "angle": 0, + "content": "3D Backbones. Generally speaking, there are three main-streams of 3D backbones: i) Point-based methods [1, 8, 9, 40-55] approximate a permutation-invariant set function, and directly learn features from raw point clouds. Researchers have observed an unsatisfactory outcome for large-scale environments [56]. ii) Projection-based methods transform unstructured points into regular 2D images and thus employ CNNs [57-62]. However, the projection could leave out critical geometric details and cause unavoidable informa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14917" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "tion loss. iii) Voxel-based methods have used sparse convolution [11, 12] to design more powerful networks for various 3D tasks [15, 63]. Sparse convolution facilitates convolution operations only on non-void voxels. Recently, fusion-based methods [3, 16, 23] have become upsurge. The point-based high-resolution branch mitigates the performance degradation, which is caused by aggressive downsampling [16, 64] (i.e., regular sparse convolution) of the voxel branch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.474, + 0.515 + ], + "angle": 0, + "content": "Dynamic Sparse Training (DST). DST can train sparse neural networks from scratch, resulting in both a speedy training and prediction procedure. During training, DST [65-73] alters the location of non-zero weights with per predefined rules, thereby creating a sparse representation and cutting down the number of calculations. This paradigm commences without prior knowledge and simultaneously refines the non-zero locations and weights. The attractive aspect of DST is that it is sparse from the beginning, resulting in lower FLOPs and memory consumption for training and inference compared to a dense model. Generally, the pruning process [74] can be completed either by threshold-based pruning [65, 66, 70, 75] or by magnitude-based pruning [67, 76]. In addition, new weights are regrown with randomness growing [66, 70, 75], momentum growing [67], and gradient-based growing [68, 69, 76, 77]. Here, we efficiently integrate dynamic sparsity with 3D kernels in large kernel 3D networks. This design satisfies the anticipation for a larger parameter space, as a wide-ranging exploration of the parameter space is of significance to DST [71, 73, 78]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.214, + 0.543 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.473, + 0.641 + ], + "angle": 0, + "content": "In this section, we first formulate the problem in Sec. 3.1, then move on to a concise overview of submanifold sparse convolution in Sec. 3.2. Afterwards, the details of Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS) are described in Sec. 3.3 and 3.4. Lastly, network architecture is outlined in Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.281, + 0.667 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.473, + 0.839 + ], + "angle": 0, + "content": "For 3D perception tasks, suppose that the input is a set of unordered point \\(\\mathcal{X} = \\{x_{i}\\in \\mathbb{R}^{D_{in}}:i = 1,2,\\ldots ,N\\}\\), and \\(D_{in}\\) is the dimensionality of the input features. \\(\\mathcal{L}\\) is a set of object/point labels, which varies according to different datasets. For segmentation task, LSK3DNet assigns each point \\(x_{i}\\) with a predicted label \\(y_{i}\\), resulting in a point-wise prediction set \\(\\mathcal{Y} = \\{y_i:i = 1,2,\\dots ,N\\}\\). In voxel-based methods, point clouds are converted to voxel representations. We denote the input sparse tensor with \\(X_{s}^{in} = [C,F]\\). Coordinate matrix \\(C\\in \\mathbb{N}_{+}^{N\\times 3}\\) includes 3-dimensional discrete coordinates, and feature matrix \\(F\\in \\mathbb{R}^{N\\times D_{in}}\\) has \\(D_{in}\\)-dimensional features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.849, + 0.449, + 0.866 + ], + "angle": 0, + "content": "3.2. Review of Submanifold Sparse Convolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Regular sparse convolution has widespread use in U-Net type 3D networks [11, 12]. However, it significantly reduces" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.166 + ], + "angle": 0, + "content": "the sparsity level of point data [64], obfuscates feature distinctions [64], and leads to low-resolution and indistinguishable small objects [16]. In our LSK3DNet, we discard regular sparse convolution and exclusively employ submanifold sparse convolution for feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.167, + 0.895, + 0.261 + ], + "angle": 0, + "content": "In a 3-dimensional space \\(C \\in \\mathbb{N}_+^{N \\times 3}\\), the \\(D_{in}\\)-dimensional input feature at \\(c \\in C\\) can be denoted as \\(x_c^{in} \\in \\mathbb{R}^{D_{in}}\\), and the 3D kernel is represented by \\(\\pmb{W} \\in \\mathbb{R}^{K^3 \\times D_{out} \\times D_{in}}\\). We divide the kernel into \\(K^3\\) spatial weights, each with a size of \\(D_{out} \\times D_{in}\\) and denote the divided weights as \\(\\pmb{W}_i\\). The submanifold sparse convolution is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.272, + 0.892, + 0.304 + ], + "angle": 0, + "content": "\\[\nx _ {c} ^ {\\text {o u t}} = \\sum_ {i \\in \\mathcal {V} ^ {D} (c, C)} \\boldsymbol {W} _ {i} x _ {c + i} ^ {i n}, \\quad c \\in C, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.895, + 0.475 + ], + "angle": 0, + "content": "where \\( x_{c}^{out} \\) represents the current voxel on which the submanifold sparse convolution is applied; \\( W_{i} \\) corresponds to \\( x_{c + i}^{in} \\), which represents the \\( i \\)-th adjacent input voxel \\( ([C_{c + i}, F_{c + i}]) \\); \\( \\mathcal{V}^D \\) is a set of offsets that define the shape of a kernel, and \\( \\mathcal{V}^D(c, C) = \\{i | c + i \\in C, i \\in \\mathcal{V}^D\\} \\) is a set of offsets from the current center \\( c \\) that exist in \\( C \\) [12]. It should be noted that the input coordinates and output coordinates are equivalent; in other words, they share the same \\( C \\). Due to this restriction, insufficient information flow exists while differentiating the distinct spatial characteristics [64]. Large receptive fields could be a potential solution to this problem." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.485, + 0.822, + 0.502 + ], + "angle": 0, + "content": "3.3. Spatial-wise Dynamic Sparsity (SDS)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.72 + ], + "angle": 0, + "content": "SDS incorporates two essential components: Sparse Kernel Initialization and Sparse Weight Pruning and Growth. SDS is applied to the 3D large kernel layers, while other layers are kept dense. When initialized with Sparse Kernel Initialization, all spatial groups of sparse tensors have the same level of sparsity \\( s \\), where sparsity refers to the fraction of zeros in sparse kernels. We adjust the sparse position within spatial groups of 3D large kernel layers with an adaptation frequency \\( f_{a} \\). During training, at regular intervals determined by \\( f_{a} \\), the adjustable weights in the kernels are modified through a two-phase procedure, which includes sparse weight pruning and sparse weight growth. We maintain a constant number of non-zero parameters (i.e., sparsity \\( s \\)) throughout the training process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Sparse Kernel Initialization. Prior research [79] has explored the distribution of sparsity parameters across multiple layers. The proportion and adjustment of sparse weights are regulated by the layer-wise sparsity ratio, which determines how the weights are modified to seek more efficient sparse structures during training. However, this method is not suitable for Spatial-wise Group Convolution, since 3D large kernels are divided into multiple spatial groups. To ensure that each group has non-zero weights, we execute Erdős-Rényi (ER) based sparsity ratios [66, 69, 80] in each spatial group. Dense weights \\( W_{D} \\) are firstly initialized in a standard way (i.e., kaiming uniform initialization [81]). The binary Mask" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14918" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.87, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Spatial-wise Dynamic Sparsity. The utilization of SDS enables us to create and train sparse kernel 3D neural networks from the beginning. The sparse weights in each spatial group are firstly initialized by Sparse Kernel Initialization (Eq. (2)), and then regularly altered by discarding the least significant connections and introducing new ones (Eq. (3)). The sparse kernels are steadily improved by this dynamic process, which leads to a more thorough collection of local features. Note that different spatial sparse group has different sparse distribution. Here we take a sparsity of \\(22\\%\\) for example (Sec.3.3 and Alg.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.471, + 0.369 + ], + "angle": 0, + "content": "\\(M\\) is then applied to these dense weights; \\(M\\) is generated using an ER-based method with a sparsity of \\(s\\). The zero-weight number in \\(M\\) is scaled by \\(1 - \\frac{D_{in} + D_{out} + K_1^g + K_2^g + K_3^g}{D_{in} \\times D_{out} \\times K_1^g \\times K_2^g \\times K_3^g}\\), where \\(K_1^g, K_2^g\\), and \\(K_3^g\\) denote the spatial group sizes. Therefore, the sparse weights \\(\\mathbf{W}_S\\) are initialized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.377, + 0.469, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {D} \\odot \\boldsymbol {M}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.362, + 0.415 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents the Hadamard product." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.536 + ], + "angle": 0, + "content": "To design 3D networks with sparse groups, following the recipe of \"using spatial sparse groups, expanding width without more parameters\", we have chosen Spatial-wise Group Convolution rather than Depth-wise Group Convolution. Our experiments have also shown that Depth-wise Group Convolution is unsuitable for large 3D kernels. This aligns with the findings in LargeKernel3D [18]. This experiment is provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Sparse Weight Pruning and Growth. Previous research [21] in 2D convolution has illustrated that dynamic sparsity can effectively enlarge kernel sizes and enhance scalability. But Depth-wise Group Convolution is unsuitable for large 3D kernels [18]. Therefore, we adapt sparse weights in each spatial group dynamically to accommodate the data. Specifically, a certain rate of connections (i.e., prune rate \\( p \\)) with the lowest magnitude are eliminated, and then we generate an equal number of new random connections. The regrow weights are randomly placed at zero positions within each spatial group. This is achieved by modifying the mask \\( M \\) (See Alg. 1). The positions of the eliminated connections and the newly generated connections are denoted as \\( E \\) and \\( G \\), respectively. The formula for pruning and regrowth can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.772, + 0.469, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} - \\boldsymbol {E}), \\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} + \\boldsymbol {G}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "By doing so, the sparsity of each spatial group is kept steady. Moreover, the weights can be adjusted dynamically, allowing for improved local characteristics. Please refer to Fig. 3 for details. Once training is complete, mask \\( M \\) in spatial groups is also recorded. Unlike previous methods [21], we employ the Spatial-wise Group Convolution and partition dynamic sparse processes into separate spatial groups. Pruning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.29, + 0.891, + 0.319 + ], + "angle": 0, + "content": "and growth are conducted independently within each group without interfering with each other." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.893, + 0.487 + ], + "angle": 0, + "content": "We can assume that weights are changing factors over time. Then, removing the least important weights is akin to the selection phase in natural evolution. Alternatively, the random addition of new weights is analogous to the alteration stage of evolutionary selection [66]. This phenomenon is also analogous to a biological process in the brain during sleep, known as synaptic shrinking. Researchers found that the weakest neural link in the brain weakens during slumber, while the vital neural connections remain how they are before. This shows that one of the main roles of sleeping is to reset the overall synaptic strength [82, 83]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.83, + 0.515 + ], + "angle": 0, + "content": "3.4. Channel-wise Weight Selection (CWS)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.688 + ], + "angle": 0, + "content": "Despite SLaK [21] employs DST to enlarge the 2D kernel size, it mitigates the performance degradation caused by sparsity in the way of expanding model width. It strives to strike a balance between sparsity and width. However, expanding width leads to the issue of increased computational burden. Therefore, SLaK [21] faces the dilemma that larger sparsity and smaller model size cannot be achieved simultaneously. Instead of naively increasing network width, we propose CWS to decouple sparsity and width. It selectively chooses the most salient channels and obtains improved performance while keeping the original model size during inference [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "CWS operates in an online mode with a sorting frequency of \\( f_{s} \\). A single weight sorting cycle involves multiple iterations of Sparse Weight Pruning and Growth, i.e., each cycle of weight sorting is a multiple of the adaptation rate \\( f_{a} \\). The width factor \\( w \\) determines the extent to which the number of channels is augmented, i.e., extending \\( D \\)-dimension to \\( w \\times D \\)-dimension. Once the expanded model has been created, we consider the \\( D \\)-dimensional small model to be embedded within the \\( w \\times D \\)-dimensional augmented model [26]. When the number of training iterations reaches the sorting condition, the channels are sorted based on their L1 value (i.e., SortChannels), prompting the model to focus on the most relevant channels. During validation, we select the most important \\( D \\)-dimensional channels from \\( w \\times D \\)-dimensional" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14919" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.086, + 0.885, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.198, + 0.895, + 0.24 + ], + "angle": 0, + "content": "Figure 4. LSK3DNet and LSK Block. (1) LSK3DNet: Point clouds are fed into voxelizer for voxel-wise features. Then we extract features with LSK Block and Point Branch. The final prediction is a point-wise output. Here MLPs is the standard Multi-Layer Perceptrons. (2) Large Sparse Kernel Block (LSK Block) (Sec.3.5)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.337 + ], + "angle": 0, + "content": "channels [26] (i.e., SelectChannels). The pseudo-code for the entire algorithm is provided in Alg. 1. This operation effectively achieves higher performance while maintaining the parameters within the expected size. This is especially important in deployments where memory usage and computational efficiency are critical factors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.348, + 0.283, + 0.362 + ], + "angle": 0, + "content": "3.5. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.473, + 0.852 + ], + "angle": 0, + "content": "Segmentation Baseline. U-Net type 3D networks (such as SparseConvNet [11] and MinkUNet [12]) employ aggressive downsampling (i.e., regular sparse convolution) to increase the receptive field but at the cost of reduced resolution. However, with a low resolution, several points or tiny objects could be combined into one single grid and become indistinguishable [16]. So SPVCNN [16] is equipped with a high-resolution point-based branch. Subsequently, 2DPASS [3] further modifies multi-representation branches by omitting the regular sparse convolution. This is because it dilates all sparse features and blurs valuable information, increasing the burden for following layers [64]. Only submanifold sparse convolution is employed for feature extraction in Modified SPVCNN [3]. By limiting the output feature positions to the input positions, submanifold sparse convolution is able to circumvent the computation burden. We use Modified SPVCNN as a baseline for our segmentation network. This 3D network is compact yet powerful and can generate high-resolution representations from sparse point clouds in large-scale scenes. However, one challenge we face is that the restricted area of submanifold sparse convolution limits the information flow and makes it difficult to distinguish different spatial characteristics of the scene. To address this challenge, we introduce LSK Block, which stands for Large Sparse Kernel Block. This block increases the kernel sizes of submanifold sparse convolution and expands the receptive field to facilitate information flow. LSK Block has a standard residual structure that adds the output of identity mapping to that of two stacked large kernel convolutions (Fig.4(2)). Our network does not need the parallel convolutional branch with dilatation convolution [18]. The details of our segmentation network architecture are listed below:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.473, + 0.902 + ], + "angle": 0, + "content": "- Segmentation Network on SemanticKITTI [25]. LSK-3DNet (See Fig.4(1)) divides the entire scene into voxels, each with a size of \\(5\\mathrm{cm}\\). It has four scales of \\(\\{2, 4, 8, 16\\}\\)." + }, + { + "type": "code_caption", + "bbox": [ + 0.503, + 0.246, + 0.766, + 0.26 + ], + "angle": 0, + "content": "Algorithm 1 Pseudo-code of SDS and CWS" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.263, + 0.764, + 0.279 + ], + "angle": 0, + "content": "1: set adaptation frequency \\( f_{a} \\), sparsity \\( s \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.28, + 0.791, + 0.292 + ], + "angle": 0, + "content": "2: set sorting frequency \\( f_{s} \\) and width factor \\( w \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.293, + 0.888, + 0.306 + ], + "angle": 0, + "content": "3: expand model width form \\(D\\) dimension to \\(w\\times D\\) dimension;" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.306, + 0.712, + 0.321 + ], + "angle": 0, + "content": "4: initialize dense model \\( \\mathbf{W}_D^{w\\times D} \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.321, + 0.715, + 0.335 + ], + "angle": 0, + "content": "5: initialize sparse layers \\( \\mathbf{W}_S^{w\\times D} \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.335, + 0.766, + 0.363 + ], + "angle": 0, + "content": "6: for each training iteration \\(i\\) do \n7: \\(\\pmb{W}_S^{w\\times D}\\gets \\text{NormalTraining}(\\pmb{W}_S^{w\\times D})\\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.363, + 0.703, + 0.375 + ], + "angle": 0, + "content": "8: if \\((i\\% f_{a})\\) equals to 0 then" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.376, + 0.761, + 0.39 + ], + "angle": 0, + "content": "9: \\(\\mathbf{W}_S^{w\\times D}\\gets \\mathbf{W}_S^{w\\times D}\\odot (\\mathbf{M} - \\mathbf{E})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.39, + 0.762, + 0.405 + ], + "angle": 0, + "content": "10: \\(\\pmb{W}_{S}^{w\\times D}\\gets \\pmb{W}_{S}^{w\\times D}\\odot (\\pmb {M} + \\pmb {G})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.405, + 0.586, + 0.416 + ], + "angle": 0, + "content": "11: end if" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.417, + 0.702, + 0.431 + ], + "angle": 0, + "content": "12: if \\((i\\% f_s)\\) equals to 0 then" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.431, + 0.767, + 0.446 + ], + "angle": 0, + "content": "13: \\(\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.446, + 0.587, + 0.457 + ], + "angle": 0, + "content": "14: end if" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.459, + 0.652, + 0.47 + ], + "angle": 0, + "content": "15: if validation then" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.471, + 0.767, + 0.486 + ], + "angle": 0, + "content": "16: \\(\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.486, + 0.764, + 0.5 + ], + "angle": 0, + "content": "17: \\(\\overline{\\mathbf{W}_S^D}\\gets\\) SelectChannels \\((\\overline{\\mathbf{W}_S^{w\\times D}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.5, + 0.715, + 0.515 + ], + "angle": 0, + "content": "18: NormalValidation \\((\\mathbf{W}_S^D)\\)" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.515, + 0.586, + 0.526 + ], + "angle": 0, + "content": "19: end if" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.528, + 0.578, + 0.54 + ], + "angle": 0, + "content": "20: end for" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.263, + 0.888, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.553, + 0.893, + 0.628 + ], + "angle": 0, + "content": "The hidden size \\(D\\) of the entire network is 64. We deploy LSK Blocks in SparseBasicBlock1. Following [3, 15, 85], we employ the weighted cross-entropy loss to optimize point accuracy and utilize the lovasz-softmax [86] loss to maximize the intersection-over-union." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.63, + 0.894, + 0.72 + ], + "angle": 0, + "content": "- Segmentation Network on ScanNet v2 [28]. This segmentation network has the same architecture as that on SemanticKITTI [25]. The entire scene is split with a voxel size of \\(2\\mathrm{cm}\\), with scales of \\(\\{2, 4, 8, 16, 16\\}\\). The hidden size \\(D\\) is 128. Following [87], we take the weighted cross-entropy loss as the objective function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.721, + 0.894, + 0.813 + ], + "angle": 0, + "content": "Detection Network on KITTI. We take Voxel R-CNN [5] as a baseline. Specifically, we retain the backbone of Voxel R-CNN and substitute plain sparse convolutional block with LSK Block in the first three stages. Other settings remain the same as [18]. The input spatial shape is \\(\\{1600, 1408, 41\\}\\), and the channels of stages are \\(\\{16, 16, 32, 64, 64\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.82, + 0.625, + 0.837 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.843, + 0.756, + 0.86 + ], + "angle": 0, + "content": "4.1. Setups and Implementations" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.865, + 0.886, + 0.881 + ], + "angle": 0, + "content": "Dataset. To evaluate our method, we perform experiments" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.78, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14920" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Results of LSK3DNet on SemanticKITTI [25] single-scan test (Sec. 4.2). Regarding input data format, P denotes points, V represents voxelizations, R signifies range images, and 2DPASS incorporates additional 2D data. \"ms/s\" means \"milliseconds per scene\"." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.121, + 0.892, + 0.364 + ], + "angle": 0, + "content": "
MethodsInputmIoUroadsidewalkparkingother-groundbuildingcartruckbicyclemotorcycleother-vehiclevegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraffic signspeed (ms/s)
PointNet++ [41]P20.172.041.818.75.662.353.70.91.90.20.246.513.830.00.91.00.016.96.08.95900
TangentConv [57]R40.983.963.933.415.483.490.815.22.716.512.179.549.358.123.028.48.149.035.828.53000
PolarNet [62]R54.390.874.461.721.790.093.822.940.330.128.584.065.567.843.240.25.661.351.857.562
RandLA-Net [10]P55.990.574.061.824.589.794.243.929.832.239.183.863.668.648.447.49.460.451.050.7880
SqueezeSegV3 [61]R55.991.774.863.426.489.092.529.638.736.533.082.058.765.445.646.220.159.449.658.9238
KPConv [42]P58.890.372.761.331.590.595.033.430.242.544.384.869.269.161.561.611.864.256.447.4-
JS3C-Net [22]V66.088.972.161.931.992.595.854.359.352.946.084.569.867.969.565.439.970.860.768.7471
SPVNAS [16]PV67.090.275.467.621.891.697.256.650.650.458.086.173.471.067.467.150.366.964.367.3259
Cylinder3D [15]V68.992.277.065.032.390.797.150.867.663.858.585.672.569.873.769.248.066.562.466.2131
RPVNet [23]RPV70.393.480.770.333.393.597.644.268.468.761.186.575.171.775.974.443.472.164.861.4168
\\( (AF)^2-S3Net \\)[84]V70.892.076.266.845.892.594.340.263.081.440.078.668.063.176.481.777.769.664.073.3-
PV-KD [19]V71.291.877.570.941.092.497.053.567.969.360.286.573.871.975.173.550.569.464.965.876
2DPASS [3]PV72.989.774.767.440.093.597.061.163.663.461.586.273.971.077.981.374.172.965.070.462
SphereFormer [24]V74.891.878.269.741.393.897.559.670.170.567.786.775.172.479.080.475.372.866.872.9123
LSK3DNetPV75.692.278.970.241.892.797.361.071.475.664.286.472.771.981.280.685.272.067.074.689
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.373, + 0.468, + 0.401 + ], + "angle": 0, + "content": "Table 2. Results of LSK3DNet on SemanticKITTI [25] multi-scan test set. The arrow below classes indicate moving classes (Sec.4.2)." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.404, + 0.468, + 0.546 + ], + "angle": 0, + "content": "
MethodInputmIoUAcccartruckother-vehiclepersonbicyclistmotorcyclist
LatticeNet [89]P45.289.354.83.50.649.944.664.3
TLSeg [90]R47.089.668.22.112.440.442.812.9
KPCnv [42]P51.289.369.45.84.767.567.447.2
Cylinder3D [15]V52.591.074.90.00.165.768.311.9
(AF)2-S3Net [84]V56.988.165.35.63.967.666.459.6
2DPASS [3]PV62.491.482.116.13.880.371.273.1
LSK3DNetPV63.492.284.47.240.977.469.972.1
" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.551, + 0.371, + 0.564 + ], + "angle": 0, + "content": "on three benchmark datasets and five tracks:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.567, + 0.47, + 0.641 + ], + "angle": 0, + "content": "- SemanticKITTI [25] has 43,551 traffic point cloud scenes with fine annotations, split into 19,130/4,071/20,350 scenes for train/val/test. The dataset has 28 semantic classes, but only 19 classes are used for single-scan track and 25 classes for multi-scan track." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.643, + 0.47, + 0.702 + ], + "angle": 0, + "content": "- ScanNet v2 [28] contains 1,201/312/100 indoor scenes for train, val, and test splits, respectively. There are 20 semantic categories for ScanNet20 track and 200 categories for ScanNet200 track." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.703, + 0.47, + 0.763 + ], + "angle": 0, + "content": "- KITTI [29] has 7,481/7,518 samples for train and test. We follow the frequently used train/val split [88] to divide the training samples into train split (3,712 samples) and val split (3,769 samples)." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.567, + 0.47, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Training and Testing Details. We utilize AdamW optimizer with OneCycleLR scheduler starting with a learning rate of 5e-3. Data augmentation is also employed, such as random flipping, scaling, rotation around the gravity axis, spatial translation. We apply instance CutMix [23] and Test Time Augmentation (TTA) [3] to the SemanticKITTI test benchmark, and enhance the model with extra training epochs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Metrics. We employ mean class intersection over union (mIoU) and overall accuracy (Acc) metrics as the evaluation" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.373, + 0.892, + 0.401 + ], + "angle": 0, + "content": "Table 3. Results of LSK3DNet and other tate-of-the-art methods on ScanNet v2 [28] (Sec.4.2). S-Net stands for ScanNet." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.404, + 0.892, + 0.68 + ], + "angle": 0, + "content": "
MethodInputS-Net20 ValS-Net20 TestS-Net200 Val
PointNet++ [41]P53.555.7-
3DMV [91]P-48.4-
PanopticFusion [92]P-52.9-
PointCNN [93]P-45.8-
PointConv [94]P61.066.6-
JointPointBased [95]P69.263.4-
PointASNL [96]P63.566.6-
SegGCN [97]P-58.9-
RandLA-Net [85]P-64.5-
KPConv [42]P69.268.6-
JSENet [98]P-69.9-
FusionNet [99]P-68.8-
Point Transformer [8]P70.6--
Fast Point Transformer [100]P72.1--
Stratified Transformer [9]P74.373.7-
Point Transformer v2 [87]P75.475.231.9
SparseConvNet [11, 87]V69.372.528.8
MinkUNet [12]V72.273.6-
LargeKernel3D [18]V73.573.9-
LSK3DNetPV75.775.533.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.732 + ], + "angle": 0, + "content": "criterion for 3D semantic segmentation tasks, as outlined in [25]. What is more, we calculate Average Precision (AP) by recalling 11 positions for 3D object detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.744, + 0.759 + ], + "angle": 0, + "content": "4.2. 3D Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Results on SemanticKITTI. We test LSK3DNet on both single-scan and multi-scan tracks [25]. Tab. 1 presents the quantitative results of SemanticKITTI single-scan track. LSK3DNet outperforms 2DPASS [3] in terms of both mIoU and IoU scores for most categories. Moreover, SDS reduces the model size of naive 3D kernels and CWS keeps the model size within the desired size. So, LSK3DNet has a faster running speed than most prior methods, but it's performance significantly benefits from the large kernels. Tab. 2 shows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14921" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.117 + ], + "angle": 0, + "content": "Table 4. Results of 3D object detection methods on the car class of KITTI val set [29] (Sec.4.3)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.12, + 0.47, + 0.276 + ], + "angle": 0, + "content": "
MethodInput3D AP (IoU=0.7)
EasyModerateHard
VoxelNet [101]V81.9765.4662.85
PointPillars [102]R86.6276.0668.91
SECOND [63]V88.6178.6277.22
Point R-CNN [103]P88.8878.6377.38
Part-Å2[104]V89.4779.4778.54
PV-RCNN [4]PV89.3583.6978.70
Focals Conv [64]V89.5284.9379.18
Voxel R-CNN [5]V89.4184.5278.93
LargeKernel3D [18]V89.5285.0779.32
LSK3DNetV90.1685.6179.53
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.448 + ], + "angle": 0, + "content": "the results of SemanticKITTI multi-scan track. The mIoU and Acc are computed over all 25 classes. Due to page limitations, we only report the per-class IoUs for moving categories. Under this challenging setting, LSK3DNet surpasses 2DPASS [3] in both mIoU and Acc. Notably, LSK3DNet achieves state-of-the-art performance on both single-scan and multi-scan tracks in SemanticKITTI. Surpassing over the SOTA 2D-3D method 2DPASS is more valuable, which utilizes both 2D and 3D data while our LSK3DNet takes as input only the 3D data. Visualization results are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.471, + 0.631 + ], + "angle": 0, + "content": "Results on ScanNet V2. We compare LSK3DNet with previous state-of-the-art methods on ScanNet v2 [28], a largescale dataset for 3D indoor scene segmentation. ScanNet v2 has two tracks: ScanNet20 and ScanNet200, which use 20 and 200 semantic classes respectively. Tab. 3 presents the quantitative results of our model and other methods on both tracks. LSK3DNet achieves higher performance than previous methods in both tracks. Our performance is superior to transformer-based methods [8, 9, 87, 100], including Point Transformer v2. Moreover, LSK3DNet has a clear advantage over LargeKernel3D [18], improving the mIoU by \\(2.2\\%\\) and \\(1.6\\%\\) on the ScanNet v2 val and test set, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.269, + 0.66 + ], + "angle": 0, + "content": "4.3. 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.668, + 0.47, + 0.804 + ], + "angle": 0, + "content": "We have also evaluated the detection performance of LSK3DNet on the val split for car category, as shown in Tab.4. We report the average precision metric for a 3D bounding box with 11 recall positions (Recall 11). Based on Voxel R-CNN [5], we showcase the effectiveness of the LSK Block by comparing it with LargeKernel3D [18], a previous 3D large-kernel method. LSK3DNet achieves better results in three difficulty levels, compared with [5, 18]. Visualization comparison can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.241, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first explore the effect of kernel size on performance, and then the overall architecture design. Next, we explain the hyperparameter choices of SDS and CWS. All ablation experiments are performed on the val set of SemancKITTI." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.638, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.19, + 0.593, + 0.199 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.091, + 0.767, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.189, + 0.729, + 0.199 + ], + "angle": 0, + "content": "LSK3DNet" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.092, + 0.806, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.125, + 0.805, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.158, + 0.806, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.189, + 0.881, + 0.199 + ], + "angle": 0, + "content": "Kernel visualization" + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.092, + 0.851, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.125, + 0.849, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.157, + 0.849, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.853, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.853, + 0.125, + 0.892, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.853, + 0.157, + 0.891, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.189, + 0.881, + 0.198 + ], + "angle": 0, + "content": "zation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.201, + 0.894, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Effective Receptive Fields (ERFs) of Baseline and LSK3DNet. LSK3DNet has a larger ERF size. Additionally, we provide visualization of learned sparse kernels, where all weight values have taken the absolute value and normalization operations. The black areas indicate positions with weight values of zero." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.893, + 0.382 + ], + "angle": 0, + "content": "Kernel Size. We use Modified SPVCNN as the baseline and then explore 3D kernel sizes under two settings: naive dense large kernel and our LSK3DNet. In the former setting, the dense 3D kernel is straightforwardly expanded. In the latter one, LSK3DNet is trained simultaneously with SDS and CWS, meaning that our LSK3DNet can learn a large sparse kernel model from the beginning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.894, + 0.657 + ], + "angle": 0, + "content": "The baseline does not use aggressive downsampling which is common in most U-type 3D networks [11, 12]. However, submanifold sparse convolution with small kernels may lose important information flow, especially for the spatially disconnected features. We solve this problem by enlarging the size of the 3D convolution kernel to obtain a large receptive field. This agrees with the performance improvement in Tab. 5, when the kernel size increases. In contrast to this observation, LargeKernel3D [18] shows the opposite trends; we empirically attribute this to the fact that LargeKernel3D is based on U-type 3D networks (Sec. 3.5). For a better understanding of Effective Receptive Fields (ERFs) size, Fig. 5 illustrates the comparison between the Baseline and LSK3DNet. Compared to the baseline, the high-contribution points of LSK3DNet are distributed over a larger input range, indicating a larger ERF. Additionally, to enhance comprehension of learned kernels, we offer kernel visualization in Fig. 5, providing insights into sparse training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Compared to the \"naive dense large kernel\" of each kernel size, LSK3DNet exhibits superior performance in its corresponding size. LSK3DNet not only leverages SDS to expand its receptive field but also starts from scratch to learn 3D sparse kernels. Concurrently, CWS widens the network during training while pruning redundant channels to accelerate inference. Furthermore, SDS and CWS address the issues of overparameterization and overfitting that arise from simply increasing the kernel's size and expanding the model's width. The best performance of LSK3DNet is achieved at the \\(9 \\times 9 \\times 9\\) size, and the following experiments are conducted with \\(9 \\times 9 \\times 9\\) size. Moreover, due to the extra dimension in 3D kernels compared to 2D kernels, naively enlarging the 3D kernel size will lead to a cubically-increasing overhead. Our LSK3DNet can effectively reduce 19.1M parameters and approximately \\(60\\%\\) of computing operations when compared" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14922" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.159 + ], + "angle": 0, + "content": "Table 5. Segmentation performance of Modified SPVCNN with various large kernel settings. \"Dense\" refers to directly enlarging the kernel size. \"Ours\" refers to training with SDS and CWS (Sec.4.4). The unit for \"Param\" is in million (M), \"FLOPs\" is in billion (G), and \"Speed\" is measured in milliseconds per scene." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.162, + 0.467, + 0.351 + ], + "angle": 0, + "content": "
Kernel SizemIoUParamFLOPsSpeed ↓
Dense (D=64)3,3,365.21.978.436
5,5,565.68.4127.857
7,7,766.222.6381.265
9,9,967.547.91916.393
11,11,1166.687.43192.0134
Dense (1.8×D)5,5,566.227.0444.084
7,7,766.573.11551.2123
9,9,967.2154.73030.0177
11,11,1166.3282.17139.2250
Ours (1.8×D)5,5,566.85.1203.846
7,7,768.113.6249.463
9,9,970.228.8763.689
11,11,1170.152.51847.4116
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.358, + 0.468, + 0.386 + ], + "angle": 0, + "content": "Table 6. Overall architecture design with \\(9 \\times 9 \\times 9\\) size. Here, we report Param, FLOPs, and Speed for the inference model (Sec.4.4)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.389, + 0.466, + 0.466 + ], + "angle": 0, + "content": "
MethodmIoUParamFLOPsSpeed ↓
Dense (1.0 × D)67.547.91916.393
Dense (1.8 × D)67.2154.73030.0177
SDS (1.8 × D)69.393.02201.4173
CWS (1.8 × D)69.147.91916.393
LSK3DNet (1.8 × D)70.228.8763.689
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.468, + 0.503 + ], + "angle": 0, + "content": "to the naive large kernel network. This aspect significantly enhances the value of SDS and CWS for 3D kernels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Overall Architecture Design. In Tab. 6, Dense \\((1.8\\times D)\\) does not yield improvement due to overparameterization and overfitting. Although SDS maintains a model width of \\(1.8\\times D\\) during inference, it achieves improved performance due to sparse training and the large receptive field. CWS effectively addresses overfitting by selecting salient channels, thereby enhancing performance. By combining SDS and CWS, LSK3DNet can achieve the best performance while significantly reducing the model size and computational cost. Spatial-wise Dynamic Sparsity. To control the degree of sparse weight adaptation in LSK3DNet, we have adjusted three hyperparameters: the adaptation frequency \\(f_{a}\\), sparsity rate \\(s\\), and prune rate \\(p\\) (Sec. 3.3). Adaptation frequency \\(f_{a}\\) indicates how often we update the sparse weights during training. Sparsity rate \\(s\\) specifies how sparse the 3D large kernels are. Prune rate \\(p\\) shows the fraction of updated weights in one adaptation. We have conducted experiments with different values of these hyperparameters and report the results in Tab. 7 and Tab. 8. Based on our empirical findings, we have chosen \\(f_{a} = 2000\\), \\(s = 0.4\\), and \\(p = 0.3\\) as the optimal settings for LSK3DNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Channel-wise Weight Selection. Another aspect of our method that we investigated is the CWS. This technique involves two hyperparameters: the sorting frequency \\( f_{s} \\) and the width factor \\( w \\) (Sec.3.4). The former determines how frequently we sort and select channels for large kernels during" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 7. Training and inference speed analysis, speed is measured in milliseconds per scene. The kernels have \\(9 \\times 9 \\times 9\\) size, \"T\" and \"T\" represent training and inference speed, respectively (Sec.4.4)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.134, + 0.892, + 0.238 + ], + "angle": 0, + "content": "
Sparsity sDense (1.0 × D)SDS (1.8 × D)LSK3DNet (1.8 × D)
mIoUT ↓I ↓mIoUT ↓I ↓mIoUT ↓I ↓
067.545193------
0.1---68.548917669.149192
0.2---68.947817669.748291
0.4---69.346317370.246789
0.6---69.045017168.845487
0.8---68.743917167.244387
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.249, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Table 8. Ablation studies of Spatial-wise Dynamic Sparsity on adaptation frequency \\( f_{a} \\) and prune rate \\( p \\) (Sec.4.4)." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.278, + 0.671, + 0.353 + ], + "angle": 0, + "content": "
Adaptation \\( f_{a} \\)mIoU
10068.8
100069.6
200070.2
300070.0
400069.8
" + }, + { + "type": "table", + "bbox": [ + 0.711, + 0.278, + 0.866, + 0.353 + ], + "angle": 0, + "content": "
Pruning pmIoU
0.1069.9
0.2070.1
0.3070.2
0.5070.0
0.7069.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Table 9. Ablation studies of Channel-wise Weight Selection on sorting frequency \\( f_{s} \\) and width factor \\( w \\) (Sec.4.4)." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.394, + 0.674, + 0.471 + ], + "angle": 0, + "content": "
Sorting fsmIoU
1 × fa68.2
2 × fa69.1
4 × fa69.8
6 × fa70.2
10 × fa69.9
" + }, + { + "type": "table", + "bbox": [ + 0.713, + 0.394, + 0.866, + 0.471 + ], + "angle": 0, + "content": "
Width wmIoU
1.1×68.4
1.5×69.6
1.8×70.2
2.1×69.9
2.5×70.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.63 + ], + "angle": 0, + "content": "training. It is expressed as a multiple of \\( f_{a} \\), the adaptation frequency, meaning that we perform channel sorting after a certain number of weight adaptation cycles. When \\( f_{s} = 6 \\times f_{a} \\), the optimal performance is achieved. The latter controls the times of network width compared to the target model. We find that increasing the width factor up to \\( 1.8 \\times \\) improves the model's performance, but beyond that point, there is no significant gain. Therefore, we opt for a width of \\( 1.8 \\times \\) to avoid more extra compute resources that a wider width would require during training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.619, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose Spatial-wise Dynamic Sparsity to scale up 3D kernels beyond \\(9 \\times 9 \\times 9\\), which prunes the volumetric weight and reduces the parameter size of large kernel layers. Our LSK3DNet can benefit from a large receptive field without increasing the computational cost compared to a naive 3D large kernel. Channel-wise Weight Selection expands the model width during training, and then sorts and selects important channels during validation to get a model of the expected size. In this way, we achieve \"using spatial sparse groups, expanding width without more parameters\". We evaluate our method on the SemanticKITTI and achieve state-of-the-art performance. Our LSK3DNet also surpasses previous 3D large kernel methods on ScanNet v2 and KITTI. Acknowledgments: This work was supported by a China Scholarship Council (CSC) scholarship." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14923" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Haoran Zhou, Yidan Feng, Mingsheng Fang, Mingqiang Wei, Jing Qin, and Tong Lu. Adaptive graph convolution for point cloud analysis. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.158, + 0.469, + 0.199 + ], + "angle": 0, + "content": "[2] Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud representation learning for 3d analysis. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.201, + 0.469, + 0.255 + ], + "angle": 0, + "content": "[3] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.257, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[4] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In CVPR, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.313, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[5] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, 2021. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.369, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[6] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.425, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[7] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In AAAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.467, + 0.469, + 0.506 + ], + "angle": 0, + "content": "[8] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.509, + 0.469, + 0.563 + ], + "angle": 0, + "content": "[9] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.566, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[10] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.622, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.664, + 0.469, + 0.703 + ], + "angle": 0, + "content": "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.706, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[13] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.762, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[14] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, 2021. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.094, + 0.893, + 0.119 + ], + "angle": 0, + "content": "architectures with sparse point-voxel convolution. In ECCV, 2020.1,2,3,5,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.894, + 0.16 + ], + "angle": 0, + "content": "[17] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.164, + 0.893, + 0.203 + ], + "angle": 0, + "content": "[18] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. CVPR, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.206, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[19] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.248, + 0.893, + 0.287 + ], + "angle": 0, + "content": "[20] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.289, + 0.893, + 0.356 + ], + "angle": 0, + "content": "[21] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.359, + 0.893, + 0.412 + ], + "angle": 0, + "content": "[22] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.415, + 0.893, + 0.467 + ], + "angle": 0, + "content": "[23] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In ICCV, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.47, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[24] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.512, + 0.893, + 0.566 + ], + "angle": 0, + "content": "[25] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.568, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[26] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. In ICLR, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.597, + 0.893, + 0.648 + ], + "angle": 0, + "content": "[27] Vidya Kamath and A Renuka. Deep learning based object detection for resource constrained devices-systematic review, future trends and challenges ahead. Neurocomputing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.893, + 0.704 + ], + "angle": 0, + "content": "[28] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.707, + 0.893, + 0.76 + ], + "angle": 0, + "content": "[29] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[30] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.791, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.847, + 0.893, + 0.899 + ], + "angle": 0, + "content": "[32] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters-improve semantic segmentation by global convolutional network. In CVPR, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.094, + 0.894, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14924" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[33] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[34] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.178, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.472, + 0.329 + ], + "angle": 0, + "content": "[37] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[38] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.472, + 0.426 + ], + "angle": 0, + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.427, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[40] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[41] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.567, + 0.472, + 0.608 + ], + "angle": 0, + "content": "[43] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[44] Siqi Fan, Qiulei Dong, Fenghua Zhu, Yisheng Lv, Peijun Ye, and Fei-Yue Wang. Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.472, + 0.692 + ], + "angle": 0, + "content": "[45] Haoxi Ran, Wei Zhuo, Jun Liu, and Li Lu. Learning inner-group relations on point clouds. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.692, + 0.472, + 0.733 + ], + "angle": 0, + "content": "[46] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.734, + 0.472, + 0.774 + ], + "angle": 0, + "content": "[47] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[48] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gcn for fast and scalable point cloud learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.818, + 0.472, + 0.859 + ], + "angle": 0, + "content": "[49] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[50] Yanchao Lian, Tuo Feng, and Jinliu Zhou. A dense point-net++ architecture for 3d point cloud semantic segmentation. In IGARSS, 2019." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.135 + ], + "angle": 0, + "content": "[51] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[52] Hehe Fan, Xin Yu, Yi Yang, and Mohan Kankanhalli. Deep hierarchical representation of point cloud videos via spatiotemporal decomposition. IEEE TPAMI, 44(12):9918-9930, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.192, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[53] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE TPAMI, 45(2):2181-2192, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.234, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[54] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and Mohan Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.278, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[55] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.321, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[56] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.377, + 0.894, + 0.418 + ], + "angle": 0, + "content": "[57] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.42, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[58] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.476, + 0.894, + 0.532 + ], + "angle": 0, + "content": "[59] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.533, + 0.894, + 0.587 + ], + "angle": 0, + "content": "[60] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: fast, uncertainty-aware semantic segmentation of lidar point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.59, + 0.894, + 0.644 + ], + "angle": 0, + "content": "[61] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezeseqv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.646, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.703, + 0.894, + 0.743 + ], + "angle": 0, + "content": "[63] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.746, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[64] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, 2022. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.789, + 0.894, + 0.83 + ], + "angle": 0, + "content": "[65] Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep rewiring: Training very sparse deep networks. arXiv preprint arXiv:1711.05136, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.832, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[66] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14925" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[67] Tim Dettmers and Luke Zettlemoyer. Sparse networks from scratch: Faster training without losing performance. arXiv preprint arXiv:1907.04840, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.134, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[68] Shiwei Liu, Decebal Constantin Mocanu, Amarsagar Reddy Ramapuram Matavalam, Yulong Pei, and Mykola Pechenizkiy. Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware. Neural Computing and Applications, 33:2589-2604, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[69] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Elsen. Rigging the lottery: Making all tickets winners. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[70] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In ICML, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[71] Siddhant Jayakumar, Razvan Pascanu, Jack Rae, Simon Osindero, and Erich Elsen. Top-kast: Top-k always sparse training. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.471, + 0.372 + ], + "angle": 0, + "content": "[72] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.471, + 0.412 + ], + "angle": 0, + "content": "[73] Shiwei Liu, Decebal Constantin Mocanu, Yulong Pei, and Mykola Pechenizkiy. Selfish sparse rn training. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.413, + 0.471, + 0.468 + ], + "angle": 0, + "content": "[74] Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Enhao Gong, Shijian Tang, Erich Elsen, Peter Vajda, Manohar Paluri, John Tran, et al. Dsd: Dense-sparse-dense training for deep neural networks. In ICLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.469, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[75] Decebal Constantin Mocanu et al. Network computations in artificial intelligence. Technische Universiteit Eindhoven, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[76] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Nest: A neural network synthesis tool based on a grow-and-prune paradigm. IEEE Transactions on Computers, 68(10):1487-1497, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.567, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[77] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Grow and prune compact, fast, and accurate lstms. IEEE Transactions on Computers, 69(3):441-452, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.637 + ], + "angle": 0, + "content": "[78] Md Aamir Raihan and Tor Aamodt. Sparse weight activation training. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.637, + 0.471, + 0.678 + ], + "angle": 0, + "content": "[79] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.679, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[80] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Li Shen, Decebal Constantin Mocanu, Zhangyang Wang, and Mykola Pechenizkiy. The unreasonable effectiveness of random pruning: Return of the most naive baseline for sparse training. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.748, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[81] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[82] Graham H Diering, Raja S Nirujogi, Richard H Roth, Paul F Worley, Akhilesh Pandey, and Richard L Huganir. *Homer1a* drives homeostatic scaling-down of excitatory synapses during sleep. *Science*, 355(6324):511-515, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[83] Luisa De Vivo, Michele Bellesi, William Marshall, Eric A Bushong, Mark H Ellisman, Giulio Tononi, and Chiara Cirelli. Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science, 355(6324):507-510, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[84] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[85] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[86] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.26, + 0.894, + 0.313 + ], + "angle": 0, + "content": "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.315, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[88] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.371, + 0.892, + 0.425 + ], + "angle": 0, + "content": "[89] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.427, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[90] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 3DV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[91] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[92] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[93] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[94] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.636, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[95] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.679, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.734, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[97] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[98] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.831, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[99] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[100] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14926" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "[101] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[102] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[103] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[104] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "14927" + } + ] +] \ No newline at end of file diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_origin.pdf b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..46f9698d2b6a4114e778e86ceeca0349fd4293ce --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/4a7b9d7e-a440-4471-acbb-bbb97d0681c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ce97ccab861929765be12953a45a10eef1b470495064b68215640746173314b +size 927105 diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/full.md b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8c44fdd5255adf3adc977d9dbec7fbc37a2f7bc7 --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/full.md @@ -0,0 +1,380 @@ +# LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels + +Tuo Feng $^{1}$ , Wenguan Wang $^{2}$ , Fan Ma $^{2*}$ , Yi Yang $^{2}$ + +1 ReLER, AAII, University of Technology Sydney 2 ReLER, CCAI, Zhejiang University + +https://github.com/FengZicai/LSK3DNet + +# Abstract + +Autonomous systems need to process large-scale, sparse, and irregular point clouds with limited compute resources. Consequently, it is essential to develop LiDAR perception methods that are both efficient and effective. Although naively enlarging 3D kernel size can enhance performance, it will also lead to a cubically-increasing overhead. Therefore, it is crucial to develop streamlined 3D large kernel designs that eliminate redundant weights and work effectively with larger kernels. In this paper, we propose an efficient and effective Large Sparse Kernel 3D Neural Network (LSK3DNet) that leverages dynamic pruning to amplify the 3D kernel size. Our method comprises two core components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the beginning to learn a large sparse 3D kernel. It not only boosts performance but also significantly reduces model size and computational cost. Moreover, CWS selects the most important channels for 3D convolution during training and subsequently prunes the redundant channels to accelerate inference for 3D vision tasks. We demonstrate the effectiveness of LSK3DNet on three benchmark datasets and five tracks compared with classical models and large kernel designs. Notably, LSK3DNet achieves the state-of-the-art performance on SemanticKITTI (i.e., $75.6\%$ on single-scan and $63.4\%$ on multi-scan), with roughly $40\%$ model size reduction and $60\%$ computing operations reduction compared to the naive large 3D kernel model. + +# 1. Introduction + +Autonomous systems are equipped with 3D LiDAR sensors and data processing platforms. The LiDAR sensor generates point clouds, which serve as the input for the processing platform. The platform performs perception tasks such as semantic segmentation [1-3] and object detection [4-6], providing a foundation for motion planning and decision-making [7] in autonomous systems. On the one hand, the processing platform is required to handle large-scale, sparse, + +![](images/c91ab8c87e5f66c9850c6eed9de25dc8b70cbfbda7bdbd1fa639c81f031fa7fe.jpg) +Figure 1. Illustrations on Spatial-wise Group Convolution [18], SDS, and CWS. The spatial dimensions $K_{s}$ (i.e., $K_{1}, K_{2}, K_{3}$ ) and channel dimensions $(D_{in} / D_{out})$ are shown. Spatial-wise Group Convolution shares the weights within each spatial group during training, leading to redundant model weights. In contrast, SDS removes non-salient weights and redundancies that are not sensitive to the input in each group, while CWS eliminates redundant weights in a channel-wise manner (Sec. 1). + +and irregular point clouds. On the other hand, the platform's computing resources are limited. Therefore, it is crucial to develop effective and efficient LiDAR perception methods. Point-based methods [1, 8, 9] rely on computationally expensive, memory inefficient, and time-consuming point sampling strategy [10]. In contrast, sparse convolution [11-14] has been widely adopted for processing largescale point clouds up to $160 \times 160 \times 20$ meters [5, 11, 12, 15-17]. Nevertheless, the application of such a technique in autonomous systems has also been hindered by limited computational resources [16]. PV-KD [19] attempts to reduce the model size with knowledge distillation, but its performance is bounded by its teacher model [15]. In opposition, large kernels can enhance performance through the advantages of large receptive fields [18, 20, 21]. However, naive 3D large kernels will significantly increase computational burdens. LargeKernel3D [18] explores large 3D kernels with Spatialwise Group Convolution to circumvent optimization and efficiency problems caused by naive 3D large kernels. However, LargeKernel3D has a lot of redundant model weights during training as it shares the weights in each spatial group. Moreover, the performance of LargeKernel3D [18] drops + +![](images/daef8b8a730e4f609e2769bbbbe105362a9b438ae6cc611440a68920d726625e.jpg) +Figure 2. Performance (mIoU) vs. Inference Speed (FPS) on SemanticKITTI [25] single-scan challenge (Sec.1). + +when scaling up the kernel size over $7 \times 7 \times 7$ . + +In this study, we propose a Large Sparse Kernel 3D Neural Network (LSK3DNet), which is an efficient and effective 3D perception model. LSK3DNet incorporates two key components, namely Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). These components enhance performance on perception tasks and overcome the challenges of high computational costs. SDS enlarges 3D kernels and reduces model parameters by learning sparse 3D kernels from scratch. CWS increases the model width and learns channel-wise importance during training, resulting in accelerated inference by pruning redundant channels [26]. As shown in Fig.1, SDS can prune 3D kernels to remove redundant weights, while CWS can eliminate redundant channels in an online manner. Our SDS and CWS complement each other following a guideline of "using spatial sparse groups, expanding width without more parameters", in contrast to "using sparse groups, expanding more" [21]. + +Specifically, SDS implements a remove-and-regrow process within each spatial group and preserves the sparsity in each group during dynamic training. This stands in stark contrast to the 2D approach [21], as it achieves dynamic sparsity in depth-wise groups. SDS allows to scale up the receptive fields with large kernel sizes, easily reaching or surpassing $9 \times 9 \times 9$ , thereby achieving higher performance. In addition, CWS selectively identifies salient channels during training. It speeds up inference on the 3D vision tasks by pruning non-salient channel parameters. In this way, LSK3DNet can benefit from expanded width to achieve enhanced performance, but maintain the original model size during inference. Reducing the models' complexity (e.g., model size) is a big benefit and a straightforward way to make them deployable on resource-constrained devices [27]. We conduct experiments on three benchmark datasets and five tracks. Our method achieves better performance compared to state-of-the-art methods [15, 23, 24] on SemanticKITTI [25] but with faster inference speed (Fig.2). Moreover, LSK3DNet outperforms the prior 3D large kernel method of LargeKernel3D [18] on ScanNet v2 [28] and KITTI [29]. + +In a nutshell, our contributions are as follows: + +- We propose LSK3DNet to enhance performance in perception tasks and mitigate high computational costs, surpassing state-of-the-art models while reducing model size by $40\%$ and computing operations by $60\%$ . +- We propose SDS to scale up 3D kernels. It learns large sparse kernels by dynamically pruning and regrowing weights from scratch, thereby reducing model size and computational operations. +- We develop CWS to improve performance by expanding the width. CWS assesses the importance of channels during training and speeds up inference by pruning redundant channel-wise parameters. + +# 2. Related Work + +Large-Kernel 3D Models. In 3D vision, research on 3D large kernel is very limited. LargeKernel3D [18] has demonstrated that sizeable kernels can be successfully employed and bring positive results for 3D networks. Spatial-wise Group Convolution [18] enables to achieve a kernel size of $7 \times 7 \times 7$ . However, it shares the weights within each spatial group during training, leading to redundant model weights (See Fig. 1). Moreover, the performance of LargeKernel3D drops when scaling up the kernel size over $7 \times 7 \times 7$ . + +Large-Kernel 2D CNN Models. In the 2010s, various large-kernel settings were investigated. LR-Net [30], Inceptions [31], and GCNs [32] explored 2D large kernels of $7 \times 7$ , $11 \times 11$ , and $15 \times 15$ respectively. Due to the widespread adoption [33, 34] of VGG [35], research into large kernels was largely overlooked in favour of multiple smaller kernels ( $1 \times 1$ or $3 \times 3$ ) to obtain a larger receptive field [36-38] during the past decade. Recently, certain studies have reintroduced large kernels in CNNs. RepLKNet [20] examines the effects of large kernels in CNNs, and for the first time it is able to increase kernel size to $31 \times 31$ . It achieves comparable results to those of Swin Transformer [39]. Following "use sparse groups, expand more", SLaK [21] has achieved an impressive kernel size of $51 \times 51$ . SLaK has to strike a trade-off between sparsity and model width since increased model width leads to an increase in model size, a phenomenon that becomes even more pronounced in 3D vision. However, CWS effectively decouples the two objectives of improved performance and reduced model size, enabling expanded width without increasing the model size. + +3D Backbones. Generally speaking, there are three main-streams of 3D backbones: i) Point-based methods [1, 8, 9, 40-55] approximate a permutation-invariant set function, and directly learn features from raw point clouds. Researchers have observed an unsatisfactory outcome for large-scale environments [56]. ii) Projection-based methods transform unstructured points into regular 2D images and thus employ CNNs [57-62]. However, the projection could leave out critical geometric details and cause unavoidable informa + +tion loss. iii) Voxel-based methods have used sparse convolution [11, 12] to design more powerful networks for various 3D tasks [15, 63]. Sparse convolution facilitates convolution operations only on non-void voxels. Recently, fusion-based methods [3, 16, 23] have become upsurge. The point-based high-resolution branch mitigates the performance degradation, which is caused by aggressive downsampling [16, 64] (i.e., regular sparse convolution) of the voxel branch. + +Dynamic Sparse Training (DST). DST can train sparse neural networks from scratch, resulting in both a speedy training and prediction procedure. During training, DST [65-73] alters the location of non-zero weights with per predefined rules, thereby creating a sparse representation and cutting down the number of calculations. This paradigm commences without prior knowledge and simultaneously refines the non-zero locations and weights. The attractive aspect of DST is that it is sparse from the beginning, resulting in lower FLOPs and memory consumption for training and inference compared to a dense model. Generally, the pruning process [74] can be completed either by threshold-based pruning [65, 66, 70, 75] or by magnitude-based pruning [67, 76]. In addition, new weights are regrown with randomness growing [66, 70, 75], momentum growing [67], and gradient-based growing [68, 69, 76, 77]. Here, we efficiently integrate dynamic sparsity with 3D kernels in large kernel 3D networks. This design satisfies the anticipation for a larger parameter space, as a wide-ranging exploration of the parameter space is of significance to DST [71, 73, 78]. + +# 3. Methodology + +In this section, we first formulate the problem in Sec. 3.1, then move on to a concise overview of submanifold sparse convolution in Sec. 3.2. Afterwards, the details of Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS) are described in Sec. 3.3 and 3.4. Lastly, network architecture is outlined in Sec. 3.5. + +# 3.1. Problem Formulation + +For 3D perception tasks, suppose that the input is a set of unordered point $\mathcal{X} = \{x_{i}\in \mathbb{R}^{D_{in}}:i = 1,2,\ldots ,N\}$ , and $D_{in}$ is the dimensionality of the input features. $\mathcal{L}$ is a set of object/point labels, which varies according to different datasets. For segmentation task, LSK3DNet assigns each point $x_{i}$ with a predicted label $y_{i}$ , resulting in a point-wise prediction set $\mathcal{Y} = \{y_i:i = 1,2,\dots ,N\}$ . In voxel-based methods, point clouds are converted to voxel representations. We denote the input sparse tensor with $X_{s}^{in} = [C,F]$ . Coordinate matrix $C\in \mathbb{N}_{+}^{N\times 3}$ includes 3-dimensional discrete coordinates, and feature matrix $F\in \mathbb{R}^{N\times D_{in}}$ has $D_{in}$ -dimensional features. + +# 3.2. Review of Submanifold Sparse Convolution + +Regular sparse convolution has widespread use in U-Net type 3D networks [11, 12]. However, it significantly reduces + +the sparsity level of point data [64], obfuscates feature distinctions [64], and leads to low-resolution and indistinguishable small objects [16]. In our LSK3DNet, we discard regular sparse convolution and exclusively employ submanifold sparse convolution for feature extraction. + +In a 3-dimensional space $C \in \mathbb{N}_+^{N \times 3}$ , the $D_{in}$ -dimensional input feature at $c \in C$ can be denoted as $x_c^{in} \in \mathbb{R}^{D_{in}}$ , and the 3D kernel is represented by $\pmb{W} \in \mathbb{R}^{K^3 \times D_{out} \times D_{in}}$ . We divide the kernel into $K^3$ spatial weights, each with a size of $D_{out} \times D_{in}$ and denote the divided weights as $\pmb{W}_i$ . The submanifold sparse convolution is formulated as follows: + +$$ +x _ {c} ^ {\text {o u t}} = \sum_ {i \in \mathcal {V} ^ {D} (c, C)} \boldsymbol {W} _ {i} x _ {c + i} ^ {i n}, \quad c \in C, \tag {1} +$$ + +where $x_{c}^{out}$ represents the current voxel on which the submanifold sparse convolution is applied; $W_{i}$ corresponds to $x_{c + i}^{in}$ , which represents the $i$ -th adjacent input voxel $([C_{c + i}, F_{c + i}])$ ; $\mathcal{V}^D$ is a set of offsets that define the shape of a kernel, and $\mathcal{V}^D(c, C) = \{i | c + i \in C, i \in \mathcal{V}^D\}$ is a set of offsets from the current center $c$ that exist in $C$ [12]. It should be noted that the input coordinates and output coordinates are equivalent; in other words, they share the same $C$ . Due to this restriction, insufficient information flow exists while differentiating the distinct spatial characteristics [64]. Large receptive fields could be a potential solution to this problem. + +# 3.3. Spatial-wise Dynamic Sparsity (SDS) + +SDS incorporates two essential components: Sparse Kernel Initialization and Sparse Weight Pruning and Growth. SDS is applied to the 3D large kernel layers, while other layers are kept dense. When initialized with Sparse Kernel Initialization, all spatial groups of sparse tensors have the same level of sparsity $s$ , where sparsity refers to the fraction of zeros in sparse kernels. We adjust the sparse position within spatial groups of 3D large kernel layers with an adaptation frequency $f_{a}$ . During training, at regular intervals determined by $f_{a}$ , the adjustable weights in the kernels are modified through a two-phase procedure, which includes sparse weight pruning and sparse weight growth. We maintain a constant number of non-zero parameters (i.e., sparsity $s$ ) throughout the training process. + +Sparse Kernel Initialization. Prior research [79] has explored the distribution of sparsity parameters across multiple layers. The proportion and adjustment of sparse weights are regulated by the layer-wise sparsity ratio, which determines how the weights are modified to seek more efficient sparse structures during training. However, this method is not suitable for Spatial-wise Group Convolution, since 3D large kernels are divided into multiple spatial groups. To ensure that each group has non-zero weights, we execute Erdős-Rényi (ER) based sparsity ratios [66, 69, 80] in each spatial group. Dense weights $W_{D}$ are firstly initialized in a standard way (i.e., kaiming uniform initialization [81]). The binary Mask + +![](images/23e4edfdbc22314569127b56f68aed17dd08578253a1621428afa21a61953898.jpg) +Figure 3. Spatial-wise Dynamic Sparsity. The utilization of SDS enables us to create and train sparse kernel 3D neural networks from the beginning. The sparse weights in each spatial group are firstly initialized by Sparse Kernel Initialization (Eq. (2)), and then regularly altered by discarding the least significant connections and introducing new ones (Eq. (3)). The sparse kernels are steadily improved by this dynamic process, which leads to a more thorough collection of local features. Note that different spatial sparse group has different sparse distribution. Here we take a sparsity of $22\%$ for example (Sec.3.3 and Alg.1). + +$M$ is then applied to these dense weights; $M$ is generated using an ER-based method with a sparsity of $s$ . The zero-weight number in $M$ is scaled by $1 - \frac{D_{in} + D_{out} + K_1^g + K_2^g + K_3^g}{D_{in} \times D_{out} \times K_1^g \times K_2^g \times K_3^g}$ , where $K_1^g, K_2^g$ , and $K_3^g$ denote the spatial group sizes. Therefore, the sparse weights $\mathbf{W}_S$ are initialized as follows: + +$$ +\boldsymbol {W} _ {S} \leftarrow \boldsymbol {W} _ {D} \odot \boldsymbol {M}, \tag {2} +$$ + +where $\odot$ represents the Hadamard product. + +To design 3D networks with sparse groups, following the recipe of "using spatial sparse groups, expanding width without more parameters", we have chosen Spatial-wise Group Convolution rather than Depth-wise Group Convolution. Our experiments have also shown that Depth-wise Group Convolution is unsuitable for large 3D kernels. This aligns with the findings in LargeKernel3D [18]. This experiment is provided in the supplementary materials. + +Sparse Weight Pruning and Growth. Previous research [21] in 2D convolution has illustrated that dynamic sparsity can effectively enlarge kernel sizes and enhance scalability. But Depth-wise Group Convolution is unsuitable for large 3D kernels [18]. Therefore, we adapt sparse weights in each spatial group dynamically to accommodate the data. Specifically, a certain rate of connections (i.e., prune rate $p$ ) with the lowest magnitude are eliminated, and then we generate an equal number of new random connections. The regrow weights are randomly placed at zero positions within each spatial group. This is achieved by modifying the mask $M$ (See Alg. 1). The positions of the eliminated connections and the newly generated connections are denoted as $E$ and $G$ , respectively. The formula for pruning and regrowth can be expressed as follows: + +$$ +\boldsymbol {W} _ {S} \leftarrow \boldsymbol {W} _ {S} \odot (\boldsymbol {M} - \boldsymbol {E}), \boldsymbol {W} _ {S} \leftarrow \boldsymbol {W} _ {S} \odot (\boldsymbol {M} + \boldsymbol {G}). \tag {3} +$$ + +By doing so, the sparsity of each spatial group is kept steady. Moreover, the weights can be adjusted dynamically, allowing for improved local characteristics. Please refer to Fig. 3 for details. Once training is complete, mask $M$ in spatial groups is also recorded. Unlike previous methods [21], we employ the Spatial-wise Group Convolution and partition dynamic sparse processes into separate spatial groups. Pruning + +and growth are conducted independently within each group without interfering with each other. + +We can assume that weights are changing factors over time. Then, removing the least important weights is akin to the selection phase in natural evolution. Alternatively, the random addition of new weights is analogous to the alteration stage of evolutionary selection [66]. This phenomenon is also analogous to a biological process in the brain during sleep, known as synaptic shrinking. Researchers found that the weakest neural link in the brain weakens during slumber, while the vital neural connections remain how they are before. This shows that one of the main roles of sleeping is to reset the overall synaptic strength [82, 83]. + +# 3.4. Channel-wise Weight Selection (CWS) + +Despite SLaK [21] employs DST to enlarge the 2D kernel size, it mitigates the performance degradation caused by sparsity in the way of expanding model width. It strives to strike a balance between sparsity and width. However, expanding width leads to the issue of increased computational burden. Therefore, SLaK [21] faces the dilemma that larger sparsity and smaller model size cannot be achieved simultaneously. Instead of naively increasing network width, we propose CWS to decouple sparsity and width. It selectively chooses the most salient channels and obtains improved performance while keeping the original model size during inference [26]. + +CWS operates in an online mode with a sorting frequency of $f_{s}$ . A single weight sorting cycle involves multiple iterations of Sparse Weight Pruning and Growth, i.e., each cycle of weight sorting is a multiple of the adaptation rate $f_{a}$ . The width factor $w$ determines the extent to which the number of channels is augmented, i.e., extending $D$ -dimension to $w \times D$ -dimension. Once the expanded model has been created, we consider the $D$ -dimensional small model to be embedded within the $w \times D$ -dimensional augmented model [26]. When the number of training iterations reaches the sorting condition, the channels are sorted based on their L1 value (i.e., SortChannels), prompting the model to focus on the most relevant channels. During validation, we select the most important $D$ -dimensional channels from $w \times D$ -dimensional + +![](images/d9f46a40a579b595695ed8d8e9f6cc8b9b487f0e3a0b771d8a60e94e07dd79a4.jpg) +Figure 4. LSK3DNet and LSK Block. (1) LSK3DNet: Point clouds are fed into voxelizer for voxel-wise features. Then we extract features with LSK Block and Point Branch. The final prediction is a point-wise output. Here MLPs is the standard Multi-Layer Perceptrons. (2) Large Sparse Kernel Block (LSK Block) (Sec.3.5). + +channels [26] (i.e., SelectChannels). The pseudo-code for the entire algorithm is provided in Alg. 1. This operation effectively achieves higher performance while maintaining the parameters within the expected size. This is especially important in deployments where memory usage and computational efficiency are critical factors. + +# 3.5. Network Architecture + +Segmentation Baseline. U-Net type 3D networks (such as SparseConvNet [11] and MinkUNet [12]) employ aggressive downsampling (i.e., regular sparse convolution) to increase the receptive field but at the cost of reduced resolution. However, with a low resolution, several points or tiny objects could be combined into one single grid and become indistinguishable [16]. So SPVCNN [16] is equipped with a high-resolution point-based branch. Subsequently, 2DPASS [3] further modifies multi-representation branches by omitting the regular sparse convolution. This is because it dilates all sparse features and blurs valuable information, increasing the burden for following layers [64]. Only submanifold sparse convolution is employed for feature extraction in Modified SPVCNN [3]. By limiting the output feature positions to the input positions, submanifold sparse convolution is able to circumvent the computation burden. We use Modified SPVCNN as a baseline for our segmentation network. This 3D network is compact yet powerful and can generate high-resolution representations from sparse point clouds in large-scale scenes. However, one challenge we face is that the restricted area of submanifold sparse convolution limits the information flow and makes it difficult to distinguish different spatial characteristics of the scene. To address this challenge, we introduce LSK Block, which stands for Large Sparse Kernel Block. This block increases the kernel sizes of submanifold sparse convolution and expands the receptive field to facilitate information flow. LSK Block has a standard residual structure that adds the output of identity mapping to that of two stacked large kernel convolutions (Fig.4(2)). Our network does not need the parallel convolutional branch with dilatation convolution [18]. The details of our segmentation network architecture are listed below: + +- Segmentation Network on SemanticKITTI [25]. LSK-3DNet (See Fig.4(1)) divides the entire scene into voxels, each with a size of $5\mathrm{cm}$ . It has four scales of $\{2, 4, 8, 16\}$ . + +Algorithm 1 Pseudo-code of SDS and CWS + +1: set adaptation frequency $f_{a}$ , sparsity $s$ ; +2: set sorting frequency $f_{s}$ and width factor $w$ ; +3: expand model width form $D$ dimension to $w\times D$ dimension; +4: initialize dense model $\mathbf{W}_D^{w\times D}$ +5: initialize sparse layers $\mathbf{W}_S^{w\times D}$ +6: for each training iteration $i$ do +7: $\pmb{W}_S^{w\times D}\gets \text{NormalTraining}(\pmb{W}_S^{w\times D})$ ; +8: if $(i\% f_{a})$ equals to 0 then +9: $\mathbf{W}_S^{w\times D}\gets \mathbf{W}_S^{w\times D}\odot (\mathbf{M} - \mathbf{E})$ +10: $\pmb{W}_{S}^{w\times D}\gets \pmb{W}_{S}^{w\times D}\odot (\pmb {M} + \pmb {G})$ +11: end if +12: if $(i\% f_s)$ equals to 0 then +13: $\pmb{W}_{S}^{w\times D}\gets \mathrm{SortChannels}(\pmb{W}_{S}^{w\times D})$ +14: end if +15: if validation then +16: $\pmb{W}_{S}^{w\times D}\gets \mathrm{SortChannels}(\pmb{W}_{S}^{w\times D})$ +17: $\overline{\mathbf{W}_S^D}\gets$ SelectChannels $(\overline{\mathbf{W}_S^{w\times D}})$ +18: NormalValidation $(\mathbf{W}_S^D)$ +19: end if +20: end for + +The hidden size $D$ of the entire network is 64. We deploy LSK Blocks in SparseBasicBlock1. Following [3, 15, 85], we employ the weighted cross-entropy loss to optimize point accuracy and utilize the lovasz-softmax [86] loss to maximize the intersection-over-union. + +- Segmentation Network on ScanNet v2 [28]. This segmentation network has the same architecture as that on SemanticKITTI [25]. The entire scene is split with a voxel size of $2\mathrm{cm}$ , with scales of $\{2, 4, 8, 16, 16\}$ . The hidden size $D$ is 128. Following [87], we take the weighted cross-entropy loss as the objective function. + +Detection Network on KITTI. We take Voxel R-CNN [5] as a baseline. Specifically, we retain the backbone of Voxel R-CNN and substitute plain sparse convolutional block with LSK Block in the first three stages. Other settings remain the same as [18]. The input spatial shape is $\{1600, 1408, 41\}$ , and the channels of stages are $\{16, 16, 32, 64, 64\}$ . + +# 4. Experiment + +# 4.1. Setups and Implementations + +Dataset. To evaluate our method, we perform experiments + +Table 1. Results of LSK3DNet on SemanticKITTI [25] single-scan test (Sec. 4.2). Regarding input data format, P denotes points, V represents voxelizations, R signifies range images, and 2DPASS incorporates additional 2D data. "ms/s" means "milliseconds per scene". + +
MethodsInputmIoUroadsidewalkparkingother-groundbuildingcartruckbicyclemotorcycleother-vehiclevegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraffic signspeed (ms/s)
PointNet++ [41]P20.172.041.818.75.662.353.70.91.90.20.246.513.830.00.91.00.016.96.08.95900
TangentConv [57]R40.983.963.933.415.483.490.815.22.716.512.179.549.358.123.028.48.149.035.828.53000
PolarNet [62]R54.390.874.461.721.790.093.822.940.330.128.584.065.567.843.240.25.661.351.857.562
RandLA-Net [10]P55.990.574.061.824.589.794.243.929.832.239.183.863.668.648.447.49.460.451.050.7880
SqueezeSegV3 [61]R55.991.774.863.426.489.092.529.638.736.533.082.058.765.445.646.220.159.449.658.9238
KPConv [42]P58.890.372.761.331.590.595.033.430.242.544.384.869.269.161.561.611.864.256.447.4-
JS3C-Net [22]V66.088.972.161.931.992.595.854.359.352.946.084.569.867.969.565.439.970.860.768.7471
SPVNAS [16]PV67.090.275.467.621.891.697.256.650.650.458.086.173.471.067.467.150.366.964.367.3259
Cylinder3D [15]V68.992.277.065.032.390.797.150.867.663.858.585.672.569.873.769.248.066.562.466.2131
RPVNet [23]RPV70.393.480.770.333.393.597.644.268.468.761.186.575.171.775.974.443.472.164.861.4168
\( (AF)^2-S3Net \)[84]V70.892.076.266.845.892.594.340.263.081.440.078.668.063.176.481.777.769.664.073.3-
PV-KD [19]V71.291.877.570.941.092.497.053.567.969.360.286.573.871.975.173.550.569.464.965.876
2DPASS [3]PV72.989.774.767.440.093.597.061.163.663.461.586.273.971.077.981.374.172.965.070.462
SphereFormer [24]V74.891.878.269.741.393.897.559.670.170.567.786.775.172.479.080.475.372.866.872.9123
LSK3DNetPV75.692.278.970.241.892.797.361.071.475.664.286.472.771.981.280.685.272.067.074.689
+ +Table 2. Results of LSK3DNet on SemanticKITTI [25] multi-scan test set. The arrow below classes indicate moving classes (Sec.4.2). + +
MethodInputmIoUAcccartruckother-vehiclepersonbicyclistmotorcyclist
LatticeNet [89]P45.289.354.83.50.649.944.664.3
TLSeg [90]R47.089.668.22.112.440.442.812.9
KPCnv [42]P51.289.369.45.84.767.567.447.2
Cylinder3D [15]V52.591.074.90.00.165.768.311.9
(AF)2-S3Net [84]V56.988.165.35.63.967.666.459.6
2DPASS [3]PV62.491.482.116.13.880.371.273.1
LSK3DNetPV63.492.284.47.240.977.469.972.1
+ +on three benchmark datasets and five tracks: + +- SemanticKITTI [25] has 43,551 traffic point cloud scenes with fine annotations, split into 19,130/4,071/20,350 scenes for train/val/test. The dataset has 28 semantic classes, but only 19 classes are used for single-scan track and 25 classes for multi-scan track. +- ScanNet v2 [28] contains 1,201/312/100 indoor scenes for train, val, and test splits, respectively. There are 20 semantic categories for ScanNet20 track and 200 categories for ScanNet200 track. +- KITTI [29] has 7,481/7,518 samples for train and test. We follow the frequently used train/val split [88] to divide the training samples into train split (3,712 samples) and val split (3,769 samples). + +Training and Testing Details. We utilize AdamW optimizer with OneCycleLR scheduler starting with a learning rate of 5e-3. Data augmentation is also employed, such as random flipping, scaling, rotation around the gravity axis, spatial translation. We apply instance CutMix [23] and Test Time Augmentation (TTA) [3] to the SemanticKITTI test benchmark, and enhance the model with extra training epochs. + +Metrics. We employ mean class intersection over union (mIoU) and overall accuracy (Acc) metrics as the evaluation + +Table 3. Results of LSK3DNet and other tate-of-the-art methods on ScanNet v2 [28] (Sec.4.2). S-Net stands for ScanNet. + +
MethodInputS-Net20 ValS-Net20 TestS-Net200 Val
PointNet++ [41]P53.555.7-
3DMV [91]P-48.4-
PanopticFusion [92]P-52.9-
PointCNN [93]P-45.8-
PointConv [94]P61.066.6-
JointPointBased [95]P69.263.4-
PointASNL [96]P63.566.6-
SegGCN [97]P-58.9-
RandLA-Net [85]P-64.5-
KPConv [42]P69.268.6-
JSENet [98]P-69.9-
FusionNet [99]P-68.8-
Point Transformer [8]P70.6--
Fast Point Transformer [100]P72.1--
Stratified Transformer [9]P74.373.7-
Point Transformer v2 [87]P75.475.231.9
SparseConvNet [11, 87]V69.372.528.8
MinkUNet [12]V72.273.6-
LargeKernel3D [18]V73.573.9-
LSK3DNetPV75.775.533.1
+ +criterion for 3D semantic segmentation tasks, as outlined in [25]. What is more, we calculate Average Precision (AP) by recalling 11 positions for 3D object detection. + +# 4.2. 3D Semantic Segmentation + +Results on SemanticKITTI. We test LSK3DNet on both single-scan and multi-scan tracks [25]. Tab. 1 presents the quantitative results of SemanticKITTI single-scan track. LSK3DNet outperforms 2DPASS [3] in terms of both mIoU and IoU scores for most categories. Moreover, SDS reduces the model size of naive 3D kernels and CWS keeps the model size within the desired size. So, LSK3DNet has a faster running speed than most prior methods, but it's performance significantly benefits from the large kernels. Tab. 2 shows + +Table 4. Results of 3D object detection methods on the car class of KITTI val set [29] (Sec.4.3). + +
MethodInput3D AP (IoU=0.7)
EasyModerateHard
VoxelNet [101]V81.9765.4662.85
PointPillars [102]R86.6276.0668.91
SECOND [63]V88.6178.6277.22
Point R-CNN [103]P88.8878.6377.38
Part-Å2[104]V89.4779.4778.54
PV-RCNN [4]PV89.3583.6978.70
Focals Conv [64]V89.5284.9379.18
Voxel R-CNN [5]V89.4184.5278.93
LargeKernel3D [18]V89.5285.0779.32
LSK3DNetV90.1685.6179.53
+ +the results of SemanticKITTI multi-scan track. The mIoU and Acc are computed over all 25 classes. Due to page limitations, we only report the per-class IoUs for moving categories. Under this challenging setting, LSK3DNet surpasses 2DPASS [3] in both mIoU and Acc. Notably, LSK3DNet achieves state-of-the-art performance on both single-scan and multi-scan tracks in SemanticKITTI. Surpassing over the SOTA 2D-3D method 2DPASS is more valuable, which utilizes both 2D and 3D data while our LSK3DNet takes as input only the 3D data. Visualization results are in the supplementary materials. + +Results on ScanNet V2. We compare LSK3DNet with previous state-of-the-art methods on ScanNet v2 [28], a largescale dataset for 3D indoor scene segmentation. ScanNet v2 has two tracks: ScanNet20 and ScanNet200, which use 20 and 200 semantic classes respectively. Tab. 3 presents the quantitative results of our model and other methods on both tracks. LSK3DNet achieves higher performance than previous methods in both tracks. Our performance is superior to transformer-based methods [8, 9, 87, 100], including Point Transformer v2. Moreover, LSK3DNet has a clear advantage over LargeKernel3D [18], improving the mIoU by $2.2\%$ and $1.6\%$ on the ScanNet v2 val and test set, respectively. + +# 4.3. 3D Object Detection + +We have also evaluated the detection performance of LSK3DNet on the val split for car category, as shown in Tab.4. We report the average precision metric for a 3D bounding box with 11 recall positions (Recall 11). Based on Voxel R-CNN [5], we showcase the effectiveness of the LSK Block by comparing it with LargeKernel3D [18], a previous 3D large-kernel method. LSK3DNet achieves better results in three difficulty levels, compared with [5, 18]. Visualization comparison can be found in the supplementary material. + +# 4.4. Ablation Studies + +We first explore the effect of kernel size on performance, and then the overall architecture design. Next, we explain the hyperparameter choices of SDS and CWS. All ablation experiments are performed on the val set of SemancKITTI. + +![](images/984dab33ec0b5d1c81b764c08dc843d24ac804b1e8afc92bfa0b46f9dc42d0d3.jpg) +Baseline + +![](images/387962a1b25b33ccc59491e7c9a9241128ba5405d9a38b0c633407e65a69c086.jpg) +LSK3DNet + +![](images/23a56d7cec483053479698b0a06e01a4b776e669bed25882719994f129a3671f.jpg) + +![](images/4c328e8887fe02723cbdb9d061171dcb40d4bf3e785b50ff9d95df24a6aa8f86.jpg) + +![](images/a5f66fcd74bae5bcd25eb009f7c46a3c84486fed1a096a4e41e18d2e412e1843.jpg) +Figure 5. Effective Receptive Fields (ERFs) of Baseline and LSK3DNet. LSK3DNet has a larger ERF size. Additionally, we provide visualization of learned sparse kernels, where all weight values have taken the absolute value and normalization operations. The black areas indicate positions with weight values of zero. + +![](images/6d717af427e6491d57e24b9cba473c1aabc5c3cb179725b3025351cbad897f4d.jpg) + +![](images/adaa3cee053e7c182d16d35fa3d7e4961afef3cbe1a03b4ec267088a7179adb7.jpg) + +![](images/862822ae03a99c92641c1b2891d9bc1889ae93f2f5c9a847372b8432707ea323.jpg) +Kernel visualization + +![](images/bd89c05d59d5d5eee8c5a27943f9f4108f8744370e623ebc0cd4d5ddd45563db.jpg) + +![](images/d3ae74dbbb029ca1258ba8f3db2a1822245430f30cd3e67ee6cc0c29ee3271f9.jpg) + +![](images/a0739faa0b12e9901255b3190f87876a97d6e5f9630662b178b8570b4e3af299.jpg) + +zation + +Kernel Size. We use Modified SPVCNN as the baseline and then explore 3D kernel sizes under two settings: naive dense large kernel and our LSK3DNet. In the former setting, the dense 3D kernel is straightforwardly expanded. In the latter one, LSK3DNet is trained simultaneously with SDS and CWS, meaning that our LSK3DNet can learn a large sparse kernel model from the beginning. + +The baseline does not use aggressive downsampling which is common in most U-type 3D networks [11, 12]. However, submanifold sparse convolution with small kernels may lose important information flow, especially for the spatially disconnected features. We solve this problem by enlarging the size of the 3D convolution kernel to obtain a large receptive field. This agrees with the performance improvement in Tab. 5, when the kernel size increases. In contrast to this observation, LargeKernel3D [18] shows the opposite trends; we empirically attribute this to the fact that LargeKernel3D is based on U-type 3D networks (Sec. 3.5). For a better understanding of Effective Receptive Fields (ERFs) size, Fig. 5 illustrates the comparison between the Baseline and LSK3DNet. Compared to the baseline, the high-contribution points of LSK3DNet are distributed over a larger input range, indicating a larger ERF. Additionally, to enhance comprehension of learned kernels, we offer kernel visualization in Fig. 5, providing insights into sparse training. + +Compared to the "naive dense large kernel" of each kernel size, LSK3DNet exhibits superior performance in its corresponding size. LSK3DNet not only leverages SDS to expand its receptive field but also starts from scratch to learn 3D sparse kernels. Concurrently, CWS widens the network during training while pruning redundant channels to accelerate inference. Furthermore, SDS and CWS address the issues of overparameterization and overfitting that arise from simply increasing the kernel's size and expanding the model's width. The best performance of LSK3DNet is achieved at the $9 \times 9 \times 9$ size, and the following experiments are conducted with $9 \times 9 \times 9$ size. Moreover, due to the extra dimension in 3D kernels compared to 2D kernels, naively enlarging the 3D kernel size will lead to a cubically-increasing overhead. Our LSK3DNet can effectively reduce 19.1M parameters and approximately $60\%$ of computing operations when compared + +Table 5. Segmentation performance of Modified SPVCNN with various large kernel settings. "Dense" refers to directly enlarging the kernel size. "Ours" refers to training with SDS and CWS (Sec.4.4). The unit for "Param" is in million (M), "FLOPs" is in billion (G), and "Speed" is measured in milliseconds per scene. + +
Kernel SizemIoUParamFLOPsSpeed ↓
Dense (D=64)3,3,365.21.978.436
5,5,565.68.4127.857
7,7,766.222.6381.265
9,9,967.547.91916.393
11,11,1166.687.43192.0134
Dense (1.8×D)5,5,566.227.0444.084
7,7,766.573.11551.2123
9,9,967.2154.73030.0177
11,11,1166.3282.17139.2250
Ours (1.8×D)5,5,566.85.1203.846
7,7,768.113.6249.463
9,9,970.228.8763.689
11,11,1170.152.51847.4116
+ +Table 6. Overall architecture design with $9 \times 9 \times 9$ size. Here, we report Param, FLOPs, and Speed for the inference model (Sec.4.4). + +
MethodmIoUParamFLOPsSpeed ↓
Dense (1.0 × D)67.547.91916.393
Dense (1.8 × D)67.2154.73030.0177
SDS (1.8 × D)69.393.02201.4173
CWS (1.8 × D)69.147.91916.393
LSK3DNet (1.8 × D)70.228.8763.689
+ +to the naive large kernel network. This aspect significantly enhances the value of SDS and CWS for 3D kernels. + +Overall Architecture Design. In Tab. 6, Dense $(1.8\times D)$ does not yield improvement due to overparameterization and overfitting. Although SDS maintains a model width of $1.8\times D$ during inference, it achieves improved performance due to sparse training and the large receptive field. CWS effectively addresses overfitting by selecting salient channels, thereby enhancing performance. By combining SDS and CWS, LSK3DNet can achieve the best performance while significantly reducing the model size and computational cost. Spatial-wise Dynamic Sparsity. To control the degree of sparse weight adaptation in LSK3DNet, we have adjusted three hyperparameters: the adaptation frequency $f_{a}$ , sparsity rate $s$ , and prune rate $p$ (Sec. 3.3). Adaptation frequency $f_{a}$ indicates how often we update the sparse weights during training. Sparsity rate $s$ specifies how sparse the 3D large kernels are. Prune rate $p$ shows the fraction of updated weights in one adaptation. We have conducted experiments with different values of these hyperparameters and report the results in Tab. 7 and Tab. 8. Based on our empirical findings, we have chosen $f_{a} = 2000$ , $s = 0.4$ , and $p = 0.3$ as the optimal settings for LSK3DNet. + +Channel-wise Weight Selection. Another aspect of our method that we investigated is the CWS. This technique involves two hyperparameters: the sorting frequency $f_{s}$ and the width factor $w$ (Sec.3.4). The former determines how frequently we sort and select channels for large kernels during + +Table 7. Training and inference speed analysis, speed is measured in milliseconds per scene. The kernels have $9 \times 9 \times 9$ size, "T" and "T" represent training and inference speed, respectively (Sec.4.4). + +
Sparsity sDense (1.0 × D)SDS (1.8 × D)LSK3DNet (1.8 × D)
mIoUT ↓I ↓mIoUT ↓I ↓mIoUT ↓I ↓
067.545193------
0.1---68.548917669.149192
0.2---68.947817669.748291
0.4---69.346317370.246789
0.6---69.045017168.845487
0.8---68.743917167.244387
+ +Table 8. Ablation studies of Spatial-wise Dynamic Sparsity on adaptation frequency $f_{a}$ and prune rate $p$ (Sec.4.4). + +
Adaptation \( f_{a} \)mIoU
10068.8
100069.6
200070.2
300070.0
400069.8
+ +
Pruning pmIoU
0.1069.9
0.2070.1
0.3070.2
0.5070.0
0.7069.7
+ +Table 9. Ablation studies of Channel-wise Weight Selection on sorting frequency $f_{s}$ and width factor $w$ (Sec.4.4). + +
Sorting fsmIoU
1 × fa68.2
2 × fa69.1
4 × fa69.8
6 × fa70.2
10 × fa69.9
+ +
Width wmIoU
1.1×68.4
1.5×69.6
1.8×70.2
2.1×69.9
2.5×70.1
+ +training. It is expressed as a multiple of $f_{a}$ , the adaptation frequency, meaning that we perform channel sorting after a certain number of weight adaptation cycles. When $f_{s} = 6 \times f_{a}$ , the optimal performance is achieved. The latter controls the times of network width compared to the target model. We find that increasing the width factor up to $1.8 \times$ improves the model's performance, but beyond that point, there is no significant gain. Therefore, we opt for a width of $1.8 \times$ to avoid more extra compute resources that a wider width would require during training. + +# 5. Conclusion + +We propose Spatial-wise Dynamic Sparsity to scale up 3D kernels beyond $9 \times 9 \times 9$ , which prunes the volumetric weight and reduces the parameter size of large kernel layers. Our LSK3DNet can benefit from a large receptive field without increasing the computational cost compared to a naive 3D large kernel. Channel-wise Weight Selection expands the model width during training, and then sorts and selects important channels during validation to get a model of the expected size. In this way, we achieve "using spatial sparse groups, expanding width without more parameters". We evaluate our method on the SemanticKITTI and achieve state-of-the-art performance. Our LSK3DNet also surpasses previous 3D large kernel methods on ScanNet v2 and KITTI. Acknowledgments: This work was supported by a China Scholarship Council (CSC) scholarship. + +# References + +[1] Haoran Zhou, Yidan Feng, Mingsheng Fang, Mingqiang Wei, Jing Qin, and Tong Lu. Adaptive graph convolution for point cloud analysis. In ICCV, 2021. 1, 2 +[2] Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud representation learning for 3d analysis. In ICCV, 2023. +[3] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In ECCV, 2022. 1, 2, 3, 5, 6, 7 +[4] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In CVPR, 2020. 1, 7 +[5] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, 2021. 1, 5, 7 +[6] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In CVPR, 2024. 1 +[7] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In AAAI, 2024. 1 +[8] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 6, 7 +[9] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1, 2, 6, 7 +[10] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 6 +[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1, 3, 5, 6, 7 +[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 3, 5, 6, 7 +[13] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022. +[14] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1 +[15] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, 2021. 1, 2, 3, 5, 6 +[16] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d + +architectures with sparse point-voxel convolution. In ECCV, 2020.1,2,3,5,6 +[17] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1 +[18] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. CVPR, 2023. 1, 2, 4, 5, 6, 7 +[19] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In CVPR, 2022. 1, 2, 6 +[20] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1, 2 +[21] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 1, 2, 4 +[22] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, 2021. 2, 6 +[23] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In ICCV, 2021. 2, 3, 6 +[24] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, 2023. 2, 6 +[25] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2, 5, 6 +[26] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. In ICLR, 2022. 2, 4, 5 +[27] Vidya Kamath and A Renuka. Deep learning based object detection for resource constrained devices-systematic review, future trends and challenges ahead. Neurocomputing, 2023. 2 +[28] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7 +[29] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6, 7 +[30] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In ICCV, 2019. 2 +[31] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 2 +[32] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters-improve semantic segmentation by global convolutional network. In CVPR, 2017. 2 + +[33] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, 2019. 2 +[34] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI, 2022. 2 +[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2 +[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2 +[37] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. +[38] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 2 +[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2 +[40] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2 +[41] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 6 +[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 6 +[43] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. +[44] Siqi Fan, Qiulei Dong, Fenghua Zhu, Yisheng Lv, Peijun Ye, and Fei-Yue Wang. Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In CVPR, 2021. +[45] Haoxi Ran, Wei Zhuo, Jun Liu, and Li Lu. Learning inner-group relations on point clouds. In ICCV, 2021. +[46] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021. +[47] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, 2019. +[48] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gcn for fast and scalable point cloud learning. In CVPR, 2020. +[49] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019. +[50] Yanchao Lian, Tuo Feng, and Jinliu Zhou. A dense point-net++ architecture for 3d point cloud semantic segmentation. In IGARSS, 2019. + +[51] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In CVPR, 2021. +[52] Hehe Fan, Xin Yu, Yi Yang, and Mohan Kankanhalli. Deep hierarchical representation of point cloud videos via spatiotemporal decomposition. IEEE TPAMI, 44(12):9918-9930, 2021. +[53] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE TPAMI, 45(2):2181-2192, 2022. +[54] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and Mohan Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. In ICLR, 2020. +[55] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In CVPR, 2021. 2 +[56] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In CVPR, 2021. 2 +[57] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2, 6 +[58] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018. +[59] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, 2019. +[60] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: fast, uncertainty-aware semantic segmentation of lidar point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020. +[61] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezeseqv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, 2020. 6 +[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, 2020. 2, 6 +[63] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3, 7 +[64] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, 2022. 3, 5, 7 +[65] Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep rewiring: Training very sparse deep networks. arXiv preprint arXiv:1711.05136, 2017. 3 +[66] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018. 3, 4 + +[67] Tim Dettmers and Luke Zettlemoyer. Sparse networks from scratch: Faster training without losing performance. arXiv preprint arXiv:1907.04840, 2019.3 +[68] Shiwei Liu, Decebal Constantin Mocanu, Amarsagar Reddy Ramapuram Matavalam, Yulong Pei, and Mykola Pechenizkiy. Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware. Neural Computing and Applications, 33:2589-2604, 2021. 3 +[69] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Elsen. Rigging the lottery: Making all tickets winners. In ICML, 2020. 3 +[70] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In ICML, 2019. 3 +[71] Siddhant Jayakumar, Razvan Pascanu, Jack Rae, Simon Osindero, and Erich Elsen. Top-kast: Top-k always sparse training. In NeurIPS, 2020. 3 +[72] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In NeurIPS, 2021. +[73] Shiwei Liu, Decebal Constantin Mocanu, Yulong Pei, and Mykola Pechenizkiy. Selfish sparse rn training. In ICML, 2021. 3 +[74] Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Enhao Gong, Shijian Tang, Erich Elsen, Peter Vajda, Manohar Paluri, John Tran, et al. Dsd: Dense-sparse-dense training for deep neural networks. In ICLR, 2016. 3 +[75] Decebal Constantin Mocanu et al. Network computations in artificial intelligence. Technische Universiteit Eindhoven, 2017. 3 +[76] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Nest: A neural network synthesis tool based on a grow-and-prune paradigm. IEEE Transactions on Computers, 68(10):1487-1497, 2019. 3 +[77] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Grow and prune compact, fast, and accurate lstms. IEEE Transactions on Computers, 69(3):441-452, 2019. 3 +[78] Md Aamir Raihan and Tor Aamodt. Sparse weight activation training. In NeurIPS, 2020. 3 +[79] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018. 3 +[80] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Li Shen, Decebal Constantin Mocanu, Zhangyang Wang, and Mykola Pechenizkiy. The unreasonable effectiveness of random pruning: Return of the most naive baseline for sparse training. In ICLR, 2022. 3 +[81] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3 +[82] Graham H Diering, Raja S Nirujogi, Richard H Roth, Paul F Worley, Akhilesh Pandey, and Richard L Huganir. *Homer1a* drives homeostatic scaling-down of excitatory synapses during sleep. *Science*, 355(6324):511-515, 2017. 4 +[83] Luisa De Vivo, Michele Bellesi, William Marshall, Eric A Bushong, Mark H Ellisman, Giulio Tononi, and Chiara Cirelli. Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science, 355(6324):507-510, 2017. 4 + +[84] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, 2021. 6 +[85] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 5, 6 +[86] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 5 +[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 5, 6, 7 +[88] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In ECCV, 2020. 6 +[89] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6 +[90] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 3DV, 2020. 6 +[91] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation. In ECCV, 2018. 6 +[92] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 6 +[93] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, 2018. 6 +[94] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 6 +[95] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 6 +[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 6 +[97] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 6 +[98] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 6 +[99] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 6 +[100] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 6, 7 + +[101] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 7 +[102] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 7 +[103] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 7 +[104] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 7 \ No newline at end of file diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/images.zip b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa09100b47f22229e1907a3f53273953a781256a --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cb56a11ca8927ffc687a5612aa0fed78b9813cf9168eb52327e0475ac705439 +size 656798 diff --git a/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/layout.json b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..067da22c17032ecf46195760710033be22315424 --- /dev/null +++ b/2024/LSK3DNet_ Towards Effective and Efficient 3D Perception with Large Sparse Kernels/layout.json @@ -0,0 +1,12075 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 36, + 103, + 557, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 36, + 103, + 557, + 120 + ], + "spans": [ + { + "bbox": [ + 36, + 103, + 557, + 120 + ], + "type": "text", + "content": "LSK3DNet: Towards Effective and Efficient 3D Perception with Large Sparse Kernels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "spans": [ + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "text", + "content": "Tuo Feng" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "text", + "content": ", Wenguan Wang" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "text", + "content": ", Fan Ma" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "text", + "content": ", Yi Yang" + }, + { + "bbox": [ + 178, + 142, + 414, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 159, + 457, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 159, + 457, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 159, + 457, + 171 + ], + "type": "text", + "content": "1 ReLER, AAII, University of Technology Sydney 2 ReLER, CCAI, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 174, + 394, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 174, + 394, + 184 + ], + "spans": [ + { + "bbox": [ + 194, + 174, + 394, + 184 + ], + "type": "text", + "content": "https://github.com/FengZicai/LSK3DNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 190, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "content": "Autonomous systems need to process large-scale, sparse, and irregular point clouds with limited compute resources. Consequently, it is essential to develop LiDAR perception methods that are both efficient and effective. Although naively enlarging 3D kernel size can enhance performance, it will also lead to a cubically-increasing overhead. Therefore, it is crucial to develop streamlined 3D large kernel designs that eliminate redundant weights and work effectively with larger kernels. In this paper, we propose an efficient and effective Large Sparse Kernel 3D Neural Network (LSK3DNet) that leverages dynamic pruning to amplify the 3D kernel size. Our method comprises two core components: Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). SDS dynamically prunes and regrows volumetric weights from the beginning to learn a large sparse 3D kernel. It not only boosts performance but also significantly reduces model size and computational cost. Moreover, CWS selects the most important channels for 3D convolution during training and subsequently prunes the redundant channels to accelerate inference for 3D vision tasks. We demonstrate the effectiveness of LSK3DNet on three benchmark datasets and five tracks compared with classical models and large kernel designs. Notably, LSK3DNet achieves the state-of-the-art performance on SemanticKITTI (i.e., " + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "inline_equation", + "content": "75.6\\%" + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "content": " on single-scan and " + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "inline_equation", + "content": "63.4\\%" + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "content": " on multi-scan), with roughly " + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "content": " model size reduction and " + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 47, + 239, + 290, + 562 + ], + "type": "text", + "content": " computing operations reduction compared to the naive large 3D kernel model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 127, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 596, + 288, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 288, + 692 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 288, + 692 + ], + "type": "text", + "content": "Autonomous systems are equipped with 3D LiDAR sensors and data processing platforms. The LiDAR sensor generates point clouds, which serve as the input for the processing platform. The platform performs perception tasks such as semantic segmentation [1-3] and object detection [4-6], providing a foundation for motion planning and decision-making [7] in autonomous systems. On the one hand, the processing platform is required to handle large-scale, sparse," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 211, + 538, + 357 + ], + "blocks": [ + { + "bbox": [ + 313, + 211, + 538, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 538, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 538, + 357 + ], + "type": "image", + "image_path": "c91ab8c87e5f66c9850c6eed9de25dc8b70cbfbda7bdbd1fa639c81f031fa7fe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "text", + "content": "Figure 1. Illustrations on Spatial-wise Group Convolution [18], SDS, and CWS. The spatial dimensions " + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "inline_equation", + "content": "K_{1}, K_{2}, K_{3}" + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "text", + "content": ") and channel dimensions " + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "inline_equation", + "content": "(D_{in} / D_{out})" + }, + { + "bbox": [ + 304, + 360, + 546, + 447 + ], + "type": "text", + "content": " are shown. Spatial-wise Group Convolution shares the weights within each spatial group during training, leading to redundant model weights. In contrast, SDS removes non-salient weights and redundancies that are not sensitive to the input in each group, while CWS eliminates redundant weights in a channel-wise manner (Sec. 1)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "and irregular point clouds. On the other hand, the platform's computing resources are limited. Therefore, it is crucial to develop effective and efficient LiDAR perception methods. Point-based methods [1, 8, 9] rely on computationally expensive, memory inefficient, and time-consuming point sampling strategy [10]. In contrast, sparse convolution [11-14] has been widely adopted for processing largescale point clouds up to " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "160 \\times 160 \\times 20" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " meters [5, 11, 12, 15-17]. Nevertheless, the application of such a technique in autonomous systems has also been hindered by limited computational resources [16]. PV-KD [19] attempts to reduce the model size with knowledge distillation, but its performance is bounded by its teacher model [15]. In opposition, large kernels can enhance performance through the advantages of large receptive fields [18, 20, 21]. However, naive 3D large kernels will significantly increase computational burdens. LargeKernel3D [18] explores large 3D kernels with Spatialwise Group Convolution to circumvent optimization and efficiency problems caused by naive 3D large kernels. However, LargeKernel3D has a lot of redundant model weights during training as it shares the weights in each spatial group. Moreover, the performance of LargeKernel3D [18] drops" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14916" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 67, + 284, + 205 + ], + "blocks": [ + { + "bbox": [ + 53, + 67, + 284, + 205 + ], + "lines": [ + { + "bbox": [ + 53, + 67, + 284, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 67, + 284, + 205 + ], + "type": "image", + "image_path": "daef8b8a730e4f609e2769bbbbe105362a9b438ae6cc611440a68920d726625e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 206, + 288, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 206, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 288, + 229 + ], + "type": "text", + "content": "Figure 2. Performance (mIoU) vs. Inference Speed (FPS) on SemanticKITTI [25] single-scan challenge (Sec.1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 234, + 233, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 233, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 233, + 246 + ], + "type": "text", + "content": "when scaling up the kernel size over " + }, + { + "bbox": [ + 46, + 234, + 233, + 246 + ], + "type": "inline_equation", + "content": "7 \\times 7 \\times 7" + }, + { + "bbox": [ + 46, + 234, + 233, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 247, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 247, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 45, + 247, + 288, + 449 + ], + "type": "text", + "content": "In this study, we propose a Large Sparse Kernel 3D Neural Network (LSK3DNet), which is an efficient and effective 3D perception model. LSK3DNet incorporates two key components, namely Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS). These components enhance performance on perception tasks and overcome the challenges of high computational costs. SDS enlarges 3D kernels and reduces model parameters by learning sparse 3D kernels from scratch. CWS increases the model width and learns channel-wise importance during training, resulting in accelerated inference by pruning redundant channels [26]. As shown in Fig.1, SDS can prune 3D kernels to remove redundant weights, while CWS can eliminate redundant channels in an online manner. Our SDS and CWS complement each other following a guideline of \"using spatial sparse groups, expanding width without more parameters\", in contrast to \"using sparse groups, expanding more\" [21]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 450, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 450, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 45, + 450, + 288, + 701 + ], + "type": "text", + "content": "Specifically, SDS implements a remove-and-regrow process within each spatial group and preserves the sparsity in each group during dynamic training. This stands in stark contrast to the 2D approach [21], as it achieves dynamic sparsity in depth-wise groups. SDS allows to scale up the receptive fields with large kernel sizes, easily reaching or surpassing " + }, + { + "bbox": [ + 45, + 450, + 288, + 701 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 45, + 450, + 288, + 701 + ], + "type": "text", + "content": ", thereby achieving higher performance. In addition, CWS selectively identifies salient channels during training. It speeds up inference on the 3D vision tasks by pruning non-salient channel parameters. In this way, LSK3DNet can benefit from expanded width to achieve enhanced performance, but maintain the original model size during inference. Reducing the models' complexity (e.g., model size) is a big benefit and a straightforward way to make them deployable on resource-constrained devices [27]. We conduct experiments on three benchmark datasets and five tracks. Our method achieves better performance compared to state-of-the-art methods [15, 23, 24] on SemanticKITTI [25] but with faster inference speed (Fig.2). Moreover, LSK3DNet outperforms the prior 3D large kernel method of LargeKernel3D [18] on ScanNet v2 [28] and KITTI [29]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 702, + 250, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 250, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 250, + 712 + ], + "type": "text", + "content": "In a nutshell, our contributions are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 547, + 224 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "content": "- We propose LSK3DNet to enhance performance in perception tasks and mitigate high computational costs, surpassing state-of-the-art models while reducing model size by " + }, + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "content": " and computing operations by " + }, + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 306, + 72, + 547, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 124, + 547, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 124, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 547, + 172 + ], + "type": "text", + "content": "- We propose SDS to scale up 3D kernels. It learns large sparse kernels by dynamically pruning and regrowing weights from scratch, thereby reducing model size and computational operations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 175, + 547, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 175, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 175, + 547, + 224 + ], + "type": "text", + "content": "- We develop CWS to improve performance by expanding the width. CWS assesses the importance of channels during training and speeds up inference by pruning redundant channel-wise parameters." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 392, + 248 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "text", + "content": "Large-Kernel 3D Models. In 3D vision, research on 3D large kernel is very limited. LargeKernel3D [18] has demonstrated that sizeable kernels can be successfully employed and bring positive results for 3D networks. Spatial-wise Group Convolution [18] enables to achieve a kernel size of " + }, + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "inline_equation", + "content": "7 \\times 7 \\times 7" + }, + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "text", + "content": ". However, it shares the weights within each spatial group during training, leading to redundant model weights (See Fig. 1). Moreover, the performance of LargeKernel3D drops when scaling up the kernel size over " + }, + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "inline_equation", + "content": "7 \\times 7 \\times 7" + }, + { + "bbox": [ + 304, + 255, + 547, + 364 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": "Large-Kernel 2D CNN Models. In the 2010s, various large-kernel settings were investigated. LR-Net [30], Inceptions [31], and GCNs [32] explored 2D large kernels of " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "11 \\times 11" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "15 \\times 15" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": " respectively. Due to the widespread adoption [33, 34] of VGG [35], research into large kernels was largely overlooked in favour of multiple smaller kernels (" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": ") to obtain a larger receptive field [36-38] during the past decade. Recently, certain studies have reintroduced large kernels in CNNs. RepLKNet [20] examines the effects of large kernels in CNNs, and for the first time it is able to increase kernel size to " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "31 \\times 31" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": ". It achieves comparable results to those of Swin Transformer [39]. Following \"use sparse groups, expand more\", SLaK [21] has achieved an impressive kernel size of " + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "inline_equation", + "content": "51 \\times 51" + }, + { + "bbox": [ + 304, + 365, + 547, + 604 + ], + "type": "text", + "content": ". SLaK has to strike a trade-off between sparsity and model width since increased model width leads to an increase in model size, a phenomenon that becomes even more pronounced in 3D vision. However, CWS effectively decouples the two objectives of improved performance and reduced model size, enabling expanded width without increasing the model size." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 548, + 714 + ], + "type": "text", + "content": "3D Backbones. Generally speaking, there are three main-streams of 3D backbones: i) Point-based methods [1, 8, 9, 40-55] approximate a permutation-invariant set function, and directly learn features from raw point clouds. Researchers have observed an unsatisfactory outcome for large-scale environments [56]. ii) Projection-based methods transform unstructured points into regular 2D images and thus employ CNNs [57-62]. However, the projection could leave out critical geometric details and cause unavoidable informa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14917" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "tion loss. iii) Voxel-based methods have used sparse convolution [11, 12] to design more powerful networks for various 3D tasks [15, 63]. Sparse convolution facilitates convolution operations only on non-void voxels. Recently, fusion-based methods [3, 16, 23] have become upsurge. The point-based high-resolution branch mitigates the performance degradation, which is caused by aggressive downsampling [16, 64] (i.e., regular sparse convolution) of the voxel branch." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 290, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 290, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 290, + 407 + ], + "type": "text", + "content": "Dynamic Sparse Training (DST). DST can train sparse neural networks from scratch, resulting in both a speedy training and prediction procedure. During training, DST [65-73] alters the location of non-zero weights with per predefined rules, thereby creating a sparse representation and cutting down the number of calculations. This paradigm commences without prior knowledge and simultaneously refines the non-zero locations and weights. The attractive aspect of DST is that it is sparse from the beginning, resulting in lower FLOPs and memory consumption for training and inference compared to a dense model. Generally, the pruning process [74] can be completed either by threshold-based pruning [65, 66, 70, 75] or by magnitude-based pruning [67, 76]. In addition, new weights are regrown with randomness growing [66, 70, 75], momentum growing [67], and gradient-based growing [68, 69, 76, 77]. Here, we efficiently integrate dynamic sparsity with 3D kernels in large kernel 3D networks. This design satisfies the anticipation for a larger parameter space, as a wide-ranging exploration of the parameter space is of significance to DST [71, 73, 78]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 415, + 130, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 130, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 130, + 430 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 434, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 289, + 507 + ], + "type": "text", + "content": "In this section, we first formulate the problem in Sec. 3.1, then move on to a concise overview of submanifold sparse convolution in Sec. 3.2. Afterwards, the details of Spatial-wise Dynamic Sparsity (SDS) and Channel-wise Weight Selection (CWS) are described in Sec. 3.3 and 3.4. Lastly, network architecture is outlined in Sec. 3.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 171, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 171, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 171, + 528 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": "For 3D perception tasks, suppose that the input is a set of unordered point " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_{i}\\in \\mathbb{R}^{D_{in}}:i = 1,2,\\ldots ,N\\}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "D_{in}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": " is the dimensionality of the input features. " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": " is a set of object/point labels, which varies according to different datasets. For segmentation task, LSK3DNet assigns each point " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": " with a predicted label " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": ", resulting in a point-wise prediction set " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{y_i:i = 1,2,\\dots ,N\\}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": ". In voxel-based methods, point clouds are converted to voxel representations. We denote the input sparse tensor with " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "X_{s}^{in} = [C,F]" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": ". Coordinate matrix " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{N}_{+}^{N\\times 3}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": " includes 3-dimensional discrete coordinates, and feature matrix " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "F\\in \\mathbb{R}^{N\\times D_{in}}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "inline_equation", + "content": "D_{in}" + }, + { + "bbox": [ + 46, + 532, + 289, + 664 + ], + "type": "text", + "content": "-dimensional features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 672, + 274, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 274, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 274, + 685 + ], + "type": "text", + "content": "3.2. Review of Submanifold Sparse Convolution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Regular sparse convolution has widespread use in U-Net type 3D networks [11, 12]. However, it significantly reduces" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "content": "the sparsity level of point data [64], obfuscates feature distinctions [64], and leads to low-resolution and indistinguishable small objects [16]. In our LSK3DNet, we discard regular sparse convolution and exclusively employ submanifold sparse convolution for feature extraction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": "In a 3-dimensional space " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{N}_+^{N \\times 3}" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "D_{in}" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": "-dimensional input feature at " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "c \\in C" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "x_c^{in} \\in \\mathbb{R}^{D_{in}}" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": ", and the 3D kernel is represented by " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "\\pmb{W} \\in \\mathbb{R}^{K^3 \\times D_{out} \\times D_{in}}" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": ". We divide the kernel into " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "K^3" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": " spatial weights, each with a size of " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "D_{out} \\times D_{in}" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": " and denote the divided weights as " + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "inline_equation", + "content": "\\pmb{W}_i" + }, + { + "bbox": [ + 304, + 132, + 547, + 206 + ], + "type": "text", + "content": ". The submanifold sparse convolution is formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 355, + 215, + 545, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 215, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 355, + 215, + 545, + 240 + ], + "type": "interline_equation", + "content": "x _ {c} ^ {\\text {o u t}} = \\sum_ {i \\in \\mathcal {V} ^ {D} (c, C)} \\boldsymbol {W} _ {i} x _ {c + i} ^ {i n}, \\quad c \\in C, \\tag {1}", + "image_path": "89aecb54b9878d79c9da54231c986d4a4d8b4e3ec5691141ce42c95ae1408c2e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "x_{c}^{out}" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " represents the current voxel on which the submanifold sparse convolution is applied; " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " corresponds to " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "x_{c + i}^{in}" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": ", which represents the " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": "-th adjacent input voxel " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "([C_{c + i}, F_{c + i}])" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^D" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " is a set of offsets that define the shape of a kernel, and " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^D(c, C) = \\{i | c + i \\in C, i \\in \\mathcal{V}^D\\}" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " is a set of offsets from the current center " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " that exist in " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": " [12]. It should be noted that the input coordinates and output coordinates are equivalent; in other words, they share the same " + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 243, + 547, + 376 + ], + "type": "text", + "content": ". Due to this restriction, insufficient information flow exists while differentiating the distinct spatial characteristics [64]. Large receptive fields could be a potential solution to this problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 384, + 503, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 503, + 397 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 503, + 397 + ], + "type": "text", + "content": "3.3. Spatial-wise Dynamic Sparsity (SDS)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "content": "SDS incorporates two essential components: Sparse Kernel Initialization and Sparse Weight Pruning and Growth. SDS is applied to the 3D large kernel layers, while other layers are kept dense. When initialized with Sparse Kernel Initialization, all spatial groups of sparse tensors have the same level of sparsity " + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "content": ", where sparsity refers to the fraction of zeros in sparse kernels. We adjust the sparse position within spatial groups of 3D large kernel layers with an adaptation frequency " + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "content": ". During training, at regular intervals determined by " + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "content": ", the adjustable weights in the kernels are modified through a two-phase procedure, which includes sparse weight pruning and sparse weight growth. We maintain a constant number of non-zero parameters (i.e., sparsity " + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 402, + 547, + 570 + ], + "type": "text", + "content": ") throughout the training process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Sparse Kernel Initialization. Prior research [79] has explored the distribution of sparsity parameters across multiple layers. The proportion and adjustment of sparse weights are regulated by the layer-wise sparsity ratio, which determines how the weights are modified to seek more efficient sparse structures during training. However, this method is not suitable for Spatial-wise Group Convolution, since 3D large kernels are divided into multiple spatial groups. To ensure that each group has non-zero weights, we execute Erdős-Rényi (ER) based sparsity ratios [66, 69, 80] in each spatial group. Dense weights " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "W_{D}" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " are firstly initialized in a standard way (i.e., kaiming uniform initialization [81]). The binary Mask" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14918" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 532, + 166 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 532, + 166 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 532, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 532, + 166 + ], + "type": "image", + "image_path": "23e4edfdbc22314569127b56f68aed17dd08578253a1621428afa21a61953898.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "text", + "content": "Figure 3. Spatial-wise Dynamic Sparsity. The utilization of SDS enables us to create and train sparse kernel 3D neural networks from the beginning. The sparse weights in each spatial group are firstly initialized by Sparse Kernel Initialization (Eq. (2)), and then regularly altered by discarding the least significant connections and introducing new ones (Eq. (3)). The sparse kernels are steadily improved by this dynamic process, which leads to a more thorough collection of local features. Note that different spatial sparse group has different sparse distribution. Here we take a sparsity of " + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 46, + 167, + 547, + 222 + ], + "type": "text", + "content": " for example (Sec.3.3 and Alg.1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": " is then applied to these dense weights; " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": " is generated using an ER-based method with a sparsity of " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": ". The zero-weight number in " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": " is scaled by " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "1 - \\frac{D_{in} + D_{out} + K_1^g + K_2^g + K_3^g}{D_{in} \\times D_{out} \\times K_1^g \\times K_2^g \\times K_3^g}" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "K_1^g, K_2^g" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "K_3^g" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": " denote the spatial group sizes. Therefore, the sparse weights " + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_S" + }, + { + "bbox": [ + 46, + 228, + 288, + 292 + ], + "type": "text", + "content": " are initialized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 131, + 298, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 298, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 131, + 298, + 287, + 310 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {D} \\odot \\boldsymbol {M}, \\tag {2}", + "image_path": "799e4c29316026c5e52ef747b8854f52476dd07fd656d172348d4423848cf3bf.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 221, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 221, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 221, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 317, + 221, + 328 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 317, + 221, + 328 + ], + "type": "text", + "content": " represents the Hadamard product." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 424 + ], + "type": "text", + "content": "To design 3D networks with sparse groups, following the recipe of \"using spatial sparse groups, expanding width without more parameters\", we have chosen Spatial-wise Group Convolution rather than Depth-wise Group Convolution. Our experiments have also shown that Depth-wise Group Convolution is unsuitable for large 3D kernels. This aligns with the findings in LargeKernel3D [18]. This experiment is provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "content": "Sparse Weight Pruning and Growth. Previous research [21] in 2D convolution has illustrated that dynamic sparsity can effectively enlarge kernel sizes and enhance scalability. But Depth-wise Group Convolution is unsuitable for large 3D kernels [18]. Therefore, we adapt sparse weights in each spatial group dynamically to accommodate the data. Specifically, a certain rate of connections (i.e., prune rate " + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "content": ") with the lowest magnitude are eliminated, and then we generate an equal number of new random connections. The regrow weights are randomly placed at zero positions within each spatial group. This is achieved by modifying the mask " + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "content": " (See Alg. 1). The positions of the eliminated connections and the newly generated connections are denoted as " + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 426, + 288, + 604 + ], + "type": "text", + "content": ", respectively. The formula for pruning and regrowth can be expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 611, + 287, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 611, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 69, + 611, + 287, + 623 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} - \\boldsymbol {E}), \\boldsymbol {W} _ {S} \\leftarrow \\boldsymbol {W} _ {S} \\odot (\\boldsymbol {M} + \\boldsymbol {G}). \\tag {3}", + "image_path": "f21fa77e8ec9a23cc029dd2e8459ffd967d26537dc0c5a6df32aa9efb0a272e3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "By doing so, the sparsity of each spatial group is kept steady. Moreover, the weights can be adjusted dynamically, allowing for improved local characteristics. Please refer to Fig. 3 for details. Once training is complete, mask " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " in spatial groups is also recorded. Unlike previous methods [21], we employ the Spatial-wise Group Convolution and partition dynamic sparse processes into separate spatial groups. Pruning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 229, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 229, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 545, + 252 + ], + "type": "text", + "content": "and growth are conducted independently within each group without interfering with each other." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 254, + 546, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 546, + 385 + ], + "type": "text", + "content": "We can assume that weights are changing factors over time. Then, removing the least important weights is akin to the selection phase in natural evolution. Alternatively, the random addition of new weights is analogous to the alteration stage of evolutionary selection [66]. This phenomenon is also analogous to a biological process in the brain during sleep, known as synaptic shrinking. Researchers found that the weakest neural link in the brain weakens during slumber, while the vital neural connections remain how they are before. This shows that one of the main roles of sleeping is to reset the overall synaptic strength [82, 83]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 395, + 507, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 507, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 507, + 407 + ], + "type": "text", + "content": "3.4. Channel-wise Weight Selection (CWS)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 414, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 544 + ], + "type": "text", + "content": "Despite SLaK [21] employs DST to enlarge the 2D kernel size, it mitigates the performance degradation caused by sparsity in the way of expanding model width. It strives to strike a balance between sparsity and width. However, expanding width leads to the issue of increased computational burden. Therefore, SLaK [21] faces the dilemma that larger sparsity and smaller model size cannot be achieved simultaneously. Instead of naively increasing network width, we propose CWS to decouple sparsity and width. It selectively chooses the most salient channels and obtains improved performance while keeping the original model size during inference [26]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "CWS operates in an online mode with a sorting frequency of " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": ". A single weight sorting cycle involves multiple iterations of Sparse Weight Pruning and Growth, i.e., each cycle of weight sorting is a multiple of the adaptation rate " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": ". The width factor " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": " determines the extent to which the number of channels is augmented, i.e., extending " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimension to " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "w \\times D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimension. Once the expanded model has been created, we consider the " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimensional small model to be embedded within the " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "w \\times D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimensional augmented model [26]. When the number of training iterations reaches the sorting condition, the channels are sorted based on their L1 value (i.e., SortChannels), prompting the model to focus on the most relevant channels. During validation, we select the most important " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimensional channels from " + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "inline_equation", + "content": "w \\times D" + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "-dimensional" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14919" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 68, + 541, + 155 + ], + "blocks": [ + { + "bbox": [ + 54, + 68, + 541, + 155 + ], + "lines": [ + { + "bbox": [ + 54, + 68, + 541, + 155 + ], + "spans": [ + { + "bbox": [ + 54, + 68, + 541, + 155 + ], + "type": "image", + "image_path": "d9f46a40a579b595695ed8d8e9f6cc8b9b487f0e3a0b771d8a60e94e07dd79a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 156, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 547, + 190 + ], + "type": "text", + "content": "Figure 4. LSK3DNet and LSK Block. (1) LSK3DNet: Point clouds are fed into voxelizer for voxel-wise features. Then we extract features with LSK Block and Point Branch. The final prediction is a point-wise output. Here MLPs is the standard Multi-Layer Perceptrons. (2) Large Sparse Kernel Block (LSK Block) (Sec.3.5)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 266 + ], + "type": "text", + "content": "channels [26] (i.e., SelectChannels). The pseudo-code for the entire algorithm is provided in Alg. 1. This operation effectively achieves higher performance while maintaining the parameters within the expected size. This is especially important in deployments where memory usage and computational efficiency are critical factors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 275, + 173, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 173, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 173, + 286 + ], + "type": "text", + "content": "3.5. Network Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 292, + 289, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 289, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 289, + 674 + ], + "type": "text", + "content": "Segmentation Baseline. U-Net type 3D networks (such as SparseConvNet [11] and MinkUNet [12]) employ aggressive downsampling (i.e., regular sparse convolution) to increase the receptive field but at the cost of reduced resolution. However, with a low resolution, several points or tiny objects could be combined into one single grid and become indistinguishable [16]. So SPVCNN [16] is equipped with a high-resolution point-based branch. Subsequently, 2DPASS [3] further modifies multi-representation branches by omitting the regular sparse convolution. This is because it dilates all sparse features and blurs valuable information, increasing the burden for following layers [64]. Only submanifold sparse convolution is employed for feature extraction in Modified SPVCNN [3]. By limiting the output feature positions to the input positions, submanifold sparse convolution is able to circumvent the computation burden. We use Modified SPVCNN as a baseline for our segmentation network. This 3D network is compact yet powerful and can generate high-resolution representations from sparse point clouds in large-scale scenes. However, one challenge we face is that the restricted area of submanifold sparse convolution limits the information flow and makes it difficult to distinguish different spatial characteristics of the scene. To address this challenge, we introduce LSK Block, which stands for Large Sparse Kernel Block. This block increases the kernel sizes of submanifold sparse convolution and expands the receptive field to facilitate information flow. LSK Block has a standard residual structure that adds the output of identity mapping to that of two stacked large kernel convolutions (Fig.4(2)). Our network does not need the parallel convolutional branch with dilatation convolution [18]. The details of our segmentation network architecture are listed below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "text", + "content": "- Segmentation Network on SemanticKITTI [25]. LSK-3DNet (See Fig.4(1)) divides the entire scene into voxels, each with a size of " + }, + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "inline_equation", + "content": "5\\mathrm{cm}" + }, + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "text", + "content": ". It has four scales of " + }, + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{2, 4, 8, 16\\}" + }, + { + "bbox": [ + 47, + 677, + 289, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 194, + 468, + 205 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 468, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 468, + 205 + ], + "type": "text", + "content": "Algorithm 1 Pseudo-code of SDS and CWS" + } + ] + } + ], + "index": 6, + "type": "text" + }, + { + "bbox": [ + 309, + 208, + 543, + 427 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "spans": [ + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "text", + "content": "1: set adaptation frequency " + }, + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "text", + "content": ", sparsity " + }, + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 310, + 208, + 467, + 220 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "spans": [ + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "text", + "content": "2: set sorting frequency " + }, + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "text", + "content": " and width factor " + }, + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 312, + 221, + 484, + 231 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "text", + "content": "3: expand model width form " + }, + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "text", + "content": " dimension to " + }, + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "inline_equation", + "content": "w\\times D" + }, + { + "bbox": [ + 312, + 232, + 543, + 242 + ], + "type": "text", + "content": " dimension;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 242, + 435, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 242, + 435, + 254 + ], + "spans": [ + { + "bbox": [ + 312, + 242, + 435, + 254 + ], + "type": "text", + "content": "4: initialize dense model " + }, + { + "bbox": [ + 312, + 242, + 435, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_D^{w\\times D}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 254, + 437, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 437, + 265 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 437, + 265 + ], + "type": "text", + "content": "5: initialize sparse layers " + }, + { + "bbox": [ + 312, + 254, + 437, + 265 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_S^{w\\times D}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "spans": [ + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "text", + "content": "6: for each training iteration " + }, + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "text", + "content": " do \n7: " + }, + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{W}_S^{w\\times D}\\gets \\text{NormalTraining}(\\pmb{W}_S^{w\\times D})" + }, + { + "bbox": [ + 312, + 265, + 468, + 287 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 287, + 430, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 287, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 287, + 430, + 297 + ], + "type": "text", + "content": "8: if " + }, + { + "bbox": [ + 312, + 287, + 430, + 297 + ], + "type": "inline_equation", + "content": "(i\\% f_{a})" + }, + { + "bbox": [ + 312, + 287, + 430, + 297 + ], + "type": "text", + "content": " equals to 0 then" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 310, + 297, + 465, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 297, + 465, + 308 + ], + "spans": [ + { + "bbox": [ + 310, + 297, + 465, + 308 + ], + "type": "text", + "content": "9: " + }, + { + "bbox": [ + 310, + 297, + 465, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_S^{w\\times D}\\gets \\mathbf{W}_S^{w\\times D}\\odot (\\mathbf{M} - \\mathbf{E})" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 308, + 466, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 308, + 466, + 320 + ], + "spans": [ + { + "bbox": [ + 310, + 308, + 466, + 320 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 310, + 308, + 466, + 320 + ], + "type": "inline_equation", + "content": "\\pmb{W}_{S}^{w\\times D}\\gets \\pmb{W}_{S}^{w\\times D}\\odot (\\pmb {M} + \\pmb {G})" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 320, + 358, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 320, + 358, + 329 + ], + "spans": [ + { + "bbox": [ + 310, + 320, + 358, + 329 + ], + "type": "text", + "content": "11: end if" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 330, + 429, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 330, + 429, + 341 + ], + "spans": [ + { + "bbox": [ + 310, + 330, + 429, + 341 + ], + "type": "text", + "content": "12: if " + }, + { + "bbox": [ + 310, + 330, + 429, + 341 + ], + "type": "inline_equation", + "content": "(i\\% f_s)" + }, + { + "bbox": [ + 310, + 330, + 429, + 341 + ], + "type": "text", + "content": " equals to 0 then" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 341, + 469, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 341, + 469, + 353 + ], + "spans": [ + { + "bbox": [ + 310, + 341, + 469, + 353 + ], + "type": "text", + "content": "13: " + }, + { + "bbox": [ + 310, + 341, + 469, + 353 + ], + "type": "inline_equation", + "content": "\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 353, + 359, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 353, + 359, + 361 + ], + "spans": [ + { + "bbox": [ + 310, + 353, + 359, + 361 + ], + "type": "text", + "content": "14: end if" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 363, + 399, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 363, + 399, + 372 + ], + "spans": [ + { + "bbox": [ + 310, + 363, + 399, + 372 + ], + "type": "text", + "content": "15: if validation then" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 373, + 469, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 373, + 469, + 384 + ], + "spans": [ + { + "bbox": [ + 310, + 373, + 469, + 384 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 310, + 373, + 469, + 384 + ], + "type": "inline_equation", + "content": "\\pmb{W}_{S}^{w\\times D}\\gets \\mathrm{SortChannels}(\\pmb{W}_{S}^{w\\times D})" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "spans": [ + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "type": "text", + "content": "17: " + }, + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{W}_S^D}\\gets" + }, + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "type": "text", + "content": " SelectChannels " + }, + { + "bbox": [ + 310, + 384, + 467, + 396 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{W}_S^{w\\times D}})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 396, + 437, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 396, + 437, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 396, + 437, + 407 + ], + "type": "text", + "content": "18: NormalValidation " + }, + { + "bbox": [ + 310, + 396, + 437, + 407 + ], + "type": "inline_equation", + "content": "(\\mathbf{W}_S^D)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 310, + 407, + 358, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 407, + 358, + 416 + ], + "spans": [ + { + "bbox": [ + 310, + 407, + 358, + 416 + ], + "type": "text", + "content": "19: end if" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 309, + 418, + 353, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 418, + 353, + 427 + ], + "spans": [ + { + "bbox": [ + 309, + 418, + 353, + 427 + ], + "type": "text", + "content": "20: end for" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 311, + 437, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 437, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 311, + 437, + 546, + 497 + ], + "type": "text", + "content": "The hidden size " + }, + { + "bbox": [ + 311, + 437, + 546, + 497 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 311, + 437, + 546, + 497 + ], + "type": "text", + "content": " of the entire network is 64. We deploy LSK Blocks in SparseBasicBlock1. Following [3, 15, 85], we employ the weighted cross-entropy loss to optimize point accuracy and utilize the lovasz-softmax [86] loss to maximize the intersection-over-union." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "text", + "content": "- Segmentation Network on ScanNet v2 [28]. This segmentation network has the same architecture as that on SemanticKITTI [25]. The entire scene is split with a voxel size of " + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "text", + "content": ", with scales of " + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "inline_equation", + "content": "\\{2, 4, 8, 16, 16\\}" + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "text", + "content": ". The hidden size " + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 306, + 498, + 547, + 570 + ], + "type": "text", + "content": " is 128. Following [87], we take the weighted cross-entropy loss as the objective function." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "text", + "content": "Detection Network on KITTI. We take Voxel R-CNN [5] as a baseline. Specifically, we retain the backbone of Voxel R-CNN and substitute plain sparse convolutional block with LSK Block in the first three stages. Other settings remain the same as [18]. The input spatial shape is " + }, + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\{1600, 1408, 41\\}" + }, + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "text", + "content": ", and the channels of stages are " + }, + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "inline_equation", + "content": "\\{16, 16, 32, 64, 64\\}" + }, + { + "bbox": [ + 305, + 571, + 547, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 649, + 382, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 649, + 382, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 649, + 382, + 662 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 667, + 462, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 462, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 462, + 681 + ], + "type": "text", + "content": "4.1. Setups and Implementations" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 685, + 542, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 685, + 542, + 697 + ], + "spans": [ + { + "bbox": [ + 306, + 685, + 542, + 697 + ], + "type": "text", + "content": "Dataset. To evaluate our method, we perform experiments" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 477, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 477, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 477, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14920" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 95, + 545, + 288 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Results of LSK3DNet on SemanticKITTI [25] single-scan test (Sec. 4.2). Regarding input data format, P denotes points, V represents voxelizations, R signifies range images, and 2DPASS incorporates additional 2D data. \"ms/s\" means \"milliseconds per scene\"." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 95, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 49, + 95, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 95, + 545, + 288 + ], + "type": "table", + "html": "
MethodsInputmIoUroadsidewalkparkingother-groundbuildingcartruckbicyclemotorcycleother-vehiclevegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraffic signspeed (ms/s)
PointNet++ [41]P20.172.041.818.75.662.353.70.91.90.20.246.513.830.00.91.00.016.96.08.95900
TangentConv [57]R40.983.963.933.415.483.490.815.22.716.512.179.549.358.123.028.48.149.035.828.53000
PolarNet [62]R54.390.874.461.721.790.093.822.940.330.128.584.065.567.843.240.25.661.351.857.562
RandLA-Net [10]P55.990.574.061.824.589.794.243.929.832.239.183.863.668.648.447.49.460.451.050.7880
SqueezeSegV3 [61]R55.991.774.863.426.489.092.529.638.736.533.082.058.765.445.646.220.159.449.658.9238
KPConv [42]P58.890.372.761.331.590.595.033.430.242.544.384.869.269.161.561.611.864.256.447.4-
JS3C-Net [22]V66.088.972.161.931.992.595.854.359.352.946.084.569.867.969.565.439.970.860.768.7471
SPVNAS [16]PV67.090.275.467.621.891.697.256.650.650.458.086.173.471.067.467.150.366.964.367.3259
Cylinder3D [15]V68.992.277.065.032.390.797.150.867.663.858.585.672.569.873.769.248.066.562.466.2131
RPVNet [23]RPV70.393.480.770.333.393.597.644.268.468.761.186.575.171.775.974.443.472.164.861.4168
\\( (AF)^2-S3Net \\)[84]V70.892.076.266.845.892.594.340.263.081.440.078.668.063.176.481.777.769.664.073.3-
PV-KD [19]V71.291.877.570.941.092.497.053.567.969.360.286.573.871.975.173.550.569.464.965.876
2DPASS [3]PV72.989.774.767.440.093.597.061.163.663.461.586.273.971.077.981.374.172.965.070.462
SphereFormer [24]V74.891.878.269.741.393.897.559.670.170.567.786.775.172.479.080.475.372.866.872.9123
LSK3DNetPV75.692.278.970.241.892.797.361.071.475.664.286.472.771.981.280.685.272.067.074.689
", + "image_path": "d3baf930b7e2dfaf06eb4cebb7e6afe319858ddf4c711d16a2d33ee0dd63f6f9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 319, + 286, + 432 + ], + "blocks": [ + { + "bbox": [ + 47, + 295, + 286, + 317 + ], + "lines": [ + { + "bbox": [ + 47, + 295, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 286, + 317 + ], + "type": "text", + "content": "Table 2. Results of LSK3DNet on SemanticKITTI [25] multi-scan test set. The arrow below classes indicate moving classes (Sec.4.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 319, + 286, + 432 + ], + "lines": [ + { + "bbox": [ + 49, + 319, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 286, + 432 + ], + "type": "table", + "html": "
MethodInputmIoUAcccartruckother-vehiclepersonbicyclistmotorcyclist
LatticeNet [89]P45.289.354.83.50.649.944.664.3
TLSeg [90]R47.089.668.22.112.440.442.812.9
KPCnv [42]P51.289.369.45.84.767.567.447.2
Cylinder3D [15]V52.591.074.90.00.165.768.311.9
(AF)2-S3Net [84]V56.988.165.35.63.967.666.459.6
2DPASS [3]PV62.491.482.116.13.880.371.273.1
LSK3DNetPV63.492.284.47.240.977.469.972.1
", + "image_path": "22ff579cb6e7143c91eb2822f8801d1004c15611f6025d86f5a6a23169fbbe02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 436, + 227, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 227, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 227, + 446 + ], + "type": "text", + "content": "on three benchmark datasets and five tracks:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 449, + 287, + 604 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 449, + 287, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 287, + 507 + ], + "type": "text", + "content": "- SemanticKITTI [25] has 43,551 traffic point cloud scenes with fine annotations, split into 19,130/4,071/20,350 scenes for train/val/test. The dataset has 28 semantic classes, but only 19 classes are used for single-scan track and 25 classes for multi-scan track." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 555 + ], + "type": "text", + "content": "- ScanNet v2 [28] contains 1,201/312/100 indoor scenes for train, val, and test splits, respectively. There are 20 semantic categories for ScanNet20 track and 200 categories for ScanNet200 track." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 556, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 604 + ], + "type": "text", + "content": "- KITTI [29] has 7,481/7,518 samples for train and test. We follow the frequently used train/val split [88] to divide the training samples into train split (3,712 samples) and val split (3,769 samples)." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 688 + ], + "type": "text", + "content": "Training and Testing Details. We utilize AdamW optimizer with OneCycleLR scheduler starting with a learning rate of 5e-3. Data augmentation is also employed, such as random flipping, scaling, rotation around the gravity axis, spatial translation. We apply instance CutMix [23] and Test Time Augmentation (TTA) [3] to the SemanticKITTI test benchmark, and enhance the model with extra training epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Metrics. We employ mean class intersection over union (mIoU) and overall accuracy (Acc) metrics as the evaluation" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 319, + 545, + 538 + ], + "blocks": [ + { + "bbox": [ + 305, + 295, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 305, + 295, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 545, + 317 + ], + "type": "text", + "content": "Table 3. Results of LSK3DNet and other tate-of-the-art methods on ScanNet v2 [28] (Sec.4.2). S-Net stands for ScanNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 319, + 545, + 538 + ], + "lines": [ + { + "bbox": [ + 309, + 319, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 309, + 319, + 545, + 538 + ], + "type": "table", + "html": "
MethodInputS-Net20 ValS-Net20 TestS-Net200 Val
PointNet++ [41]P53.555.7-
3DMV [91]P-48.4-
PanopticFusion [92]P-52.9-
PointCNN [93]P-45.8-
PointConv [94]P61.066.6-
JointPointBased [95]P69.263.4-
PointASNL [96]P63.566.6-
SegGCN [97]P-58.9-
RandLA-Net [85]P-64.5-
KPConv [42]P69.268.6-
JSENet [98]P-69.9-
FusionNet [99]P-68.8-
Point Transformer [8]P70.6--
Fast Point Transformer [100]P72.1--
Stratified Transformer [9]P74.373.7-
Point Transformer v2 [87]P75.475.231.9
SparseConvNet [11, 87]V69.372.528.8
MinkUNet [12]V72.273.6-
LargeKernel3D [18]V73.573.9-
LSK3DNetPV75.775.533.1
", + "image_path": "29cfb9db6b62fa0fd080e13025481c4ec9b016ced3e3c103bccadac0b4da5ac1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 542, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 579 + ], + "type": "text", + "content": "criterion for 3D semantic segmentation tasks, as outlined in [25]. What is more, we calculate Average Precision (AP) by recalling 11 positions for 3D object detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 587, + 455, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 455, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 455, + 601 + ], + "type": "text", + "content": "4.2. 3D Semantic Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Results on SemanticKITTI. We test LSK3DNet on both single-scan and multi-scan tracks [25]. Tab. 1 presents the quantitative results of SemanticKITTI single-scan track. LSK3DNet outperforms 2DPASS [3] in terms of both mIoU and IoU scores for most categories. Moreover, SDS reduces the model size of naive 3D kernels and CWS keeps the model size within the desired size. So, LSK3DNet has a faster running speed than most prior methods, but it's performance significantly benefits from the large kernels. Tab. 2 shows" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14921" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 95, + 287, + 218 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 92 + ], + "type": "text", + "content": "Table 4. Results of 3D object detection methods on the car class of KITTI val set [29] (Sec.4.3)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 95, + 287, + 218 + ], + "lines": [ + { + "bbox": [ + 50, + 95, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 50, + 95, + 287, + 218 + ], + "type": "table", + "html": "
MethodInput3D AP (IoU=0.7)
EasyModerateHard
VoxelNet [101]V81.9765.4662.85
PointPillars [102]R86.6276.0668.91
SECOND [63]V88.6178.6277.22
Point R-CNN [103]P88.8878.6377.38
Part-Å2[104]V89.4779.4778.54
PV-RCNN [4]PV89.3583.6978.70
Focals Conv [64]V89.5284.9379.18
Voxel R-CNN [5]V89.4184.5278.93
LargeKernel3D [18]V89.5285.0779.32
LSK3DNetV90.1685.6179.53
", + "image_path": "be729abb4e5a68cff3f72f4e54eef46a50380f8bba8f50aceed83fcdc2a7186a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 223, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 354 + ], + "type": "text", + "content": "the results of SemanticKITTI multi-scan track. The mIoU and Acc are computed over all 25 classes. Due to page limitations, we only report the per-class IoUs for moving categories. Under this challenging setting, LSK3DNet surpasses 2DPASS [3] in both mIoU and Acc. Notably, LSK3DNet achieves state-of-the-art performance on both single-scan and multi-scan tracks in SemanticKITTI. Surpassing over the SOTA 2D-3D method 2DPASS is more valuable, which utilizes both 2D and 3D data while our LSK3DNet takes as input only the 3D data. Visualization results are in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "content": "Results on ScanNet V2. We compare LSK3DNet with previous state-of-the-art methods on ScanNet v2 [28], a largescale dataset for 3D indoor scene segmentation. ScanNet v2 has two tracks: ScanNet20 and ScanNet200, which use 20 and 200 semantic classes respectively. Tab. 3 presents the quantitative results of our model and other methods on both tracks. LSK3DNet achieves higher performance than previous methods in both tracks. Our performance is superior to transformer-based methods [8, 9, 87, 100], including Point Transformer v2. Moreover, LSK3DNet has a clear advantage over LargeKernel3D [18], improving the mIoU by " + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 46, + 356, + 288, + 499 + ], + "type": "text", + "content": " on the ScanNet v2 val and test set, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 510, + 164, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 164, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 164, + 522 + ], + "type": "text", + "content": "4.3. 3D Object Detection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 636 + ], + "type": "text", + "content": "We have also evaluated the detection performance of LSK3DNet on the val split for car category, as shown in Tab.4. We report the average precision metric for a 3D bounding box with 11 recall positions (Recall 11). Based on Voxel R-CNN [5], we showcase the effectiveness of the LSK Block by comparing it with LargeKernel3D [18], a previous 3D large-kernel method. LSK3DNet achieves better results in three difficulty levels, compared with [5, 18]. Visualization comparison can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "We first explore the effect of kernel size on performance, and then the overall architecture design. Next, we explain the hyperparameter choices of SDS and CWS. All ablation experiments are performed on the val set of SemancKITTI." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 390, + 148 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 390, + 148 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 390, + 148 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 390, + 148 + ], + "type": "image", + "image_path": "984dab33ec0b5d1c81b764c08dc843d24ac804b1e8afc92bfa0b46f9dc42d0d3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 150, + 362, + 157 + ], + "lines": [ + { + "bbox": [ + 335, + 150, + 362, + 157 + ], + "spans": [ + { + "bbox": [ + 335, + 150, + 362, + 157 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 390, + 72, + 469, + 148 + ], + "blocks": [ + { + "bbox": [ + 390, + 72, + 469, + 148 + ], + "lines": [ + { + "bbox": [ + 390, + 72, + 469, + 148 + ], + "spans": [ + { + "bbox": [ + 390, + 72, + 469, + 148 + ], + "type": "image", + "image_path": "387962a1b25b33ccc59491e7c9a9241128ba5405d9a38b0c633407e65a69c086.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 149, + 446, + 157 + ], + "lines": [ + { + "bbox": [ + 410, + 149, + 446, + 157 + ], + "spans": [ + { + "bbox": [ + 410, + 149, + 446, + 157 + ], + "type": "text", + "content": "LSK3DNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 493, + 95 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 493, + 95 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 493, + 95 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 493, + 95 + ], + "type": "image", + "image_path": "23a56d7cec483053479698b0a06e01a4b776e669bed25882719994f129a3671f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 470, + 99, + 492, + 123 + ], + "blocks": [ + { + "bbox": [ + 470, + 99, + 492, + 123 + ], + "lines": [ + { + "bbox": [ + 470, + 99, + 492, + 123 + ], + "spans": [ + { + "bbox": [ + 470, + 99, + 492, + 123 + ], + "type": "image", + "image_path": "4c328e8887fe02723cbdb9d061171dcb40d4bf3e785b50ff9d95df24a6aa8f86.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 125, + 493, + 147 + ], + "blocks": [ + { + "bbox": [ + 470, + 125, + 493, + 147 + ], + "lines": [ + { + "bbox": [ + 470, + 125, + 493, + 147 + ], + "spans": [ + { + "bbox": [ + 470, + 125, + 493, + 147 + ], + "type": "image", + "image_path": "a5f66fcd74bae5bcd25eb009f7c46a3c84486fed1a096a4e41e18d2e412e1843.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 159, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 304, + 159, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 547, + 213 + ], + "type": "text", + "content": "Figure 5. Effective Receptive Fields (ERFs) of Baseline and LSK3DNet. LSK3DNet has a larger ERF size. Additionally, we provide visualization of learned sparse kernels, where all weight values have taken the absolute value and normalization operations. The black areas indicate positions with weight values of zero." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 496, + 72, + 520, + 95 + ], + "blocks": [ + { + "bbox": [ + 496, + 72, + 520, + 95 + ], + "lines": [ + { + "bbox": [ + 496, + 72, + 520, + 95 + ], + "spans": [ + { + "bbox": [ + 496, + 72, + 520, + 95 + ], + "type": "image", + "image_path": "6d717af427e6491d57e24b9cba473c1aabc5c3cb179725b3025351cbad897f4d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 496, + 99, + 519, + 123 + ], + "blocks": [ + { + "bbox": [ + 496, + 99, + 519, + 123 + ], + "lines": [ + { + "bbox": [ + 496, + 99, + 519, + 123 + ], + "spans": [ + { + "bbox": [ + 496, + 99, + 519, + 123 + ], + "type": "image", + "image_path": "adaa3cee053e7c182d16d35fa3d7e4961afef3cbe1a03b4ec267088a7179adb7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 496, + 124, + 519, + 147 + ], + "blocks": [ + { + "bbox": [ + 477, + 149, + 539, + 157 + ], + "lines": [ + { + "bbox": [ + 477, + 149, + 539, + 157 + ], + "spans": [ + { + "bbox": [ + 477, + 149, + 539, + 157 + ], + "type": "text", + "content": "Kernel visualization" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 496, + 124, + 519, + 147 + ], + "lines": [ + { + "bbox": [ + 496, + 124, + 519, + 147 + ], + "spans": [ + { + "bbox": [ + 496, + 124, + 519, + 147 + ], + "type": "image", + "image_path": "862822ae03a99c92641c1b2891d9bc1889ae93f2f5c9a847372b8432707ea323.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 522, + 72, + 545, + 95 + ], + "blocks": [ + { + "bbox": [ + 522, + 72, + 545, + 95 + ], + "lines": [ + { + "bbox": [ + 522, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 522, + 72, + 545, + 95 + ], + "type": "image", + "image_path": "bd89c05d59d5d5eee8c5a27943f9f4108f8744370e623ebc0cd4d5ddd45563db.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 522, + 99, + 545, + 123 + ], + "blocks": [ + { + "bbox": [ + 522, + 99, + 545, + 123 + ], + "lines": [ + { + "bbox": [ + 522, + 99, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 522, + 99, + 545, + 123 + ], + "type": "image", + "image_path": "d3ae74dbbb029ca1258ba8f3db2a1822245430f30cd3e67ee6cc0c29ee3271f9.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 522, + 124, + 545, + 147 + ], + "blocks": [ + { + "bbox": [ + 522, + 124, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 522, + 124, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 522, + 124, + 545, + 147 + ], + "type": "image", + "image_path": "a0739faa0b12e9901255b3190f87876a97d6e5f9630662b178b8570b4e3af299.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 522, + 149, + 539, + 156 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 522, + 149, + 539, + 156 + ], + "spans": [ + { + "bbox": [ + 522, + 149, + 539, + 156 + ], + "type": "text", + "content": "zation" + } + ] + } + ], + "index": 22, + "type": "text" + }, + { + "bbox": [ + 304, + 219, + 546, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 546, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 546, + 302 + ], + "type": "text", + "content": "Kernel Size. We use Modified SPVCNN as the baseline and then explore 3D kernel sizes under two settings: naive dense large kernel and our LSK3DNet. In the former setting, the dense 3D kernel is straightforwardly expanded. In the latter one, LSK3DNet is trained simultaneously with SDS and CWS, meaning that our LSK3DNet can learn a large sparse kernel model from the beginning." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 304, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 547, + 520 + ], + "type": "text", + "content": "The baseline does not use aggressive downsampling which is common in most U-type 3D networks [11, 12]. However, submanifold sparse convolution with small kernels may lose important information flow, especially for the spatially disconnected features. We solve this problem by enlarging the size of the 3D convolution kernel to obtain a large receptive field. This agrees with the performance improvement in Tab. 5, when the kernel size increases. In contrast to this observation, LargeKernel3D [18] shows the opposite trends; we empirically attribute this to the fact that LargeKernel3D is based on U-type 3D networks (Sec. 3.5). For a better understanding of Effective Receptive Fields (ERFs) size, Fig. 5 illustrates the comparison between the Baseline and LSK3DNet. Compared to the baseline, the high-contribution points of LSK3DNet are distributed over a larger input range, indicating a larger ERF. Additionally, to enhance comprehension of learned kernels, we offer kernel visualization in Fig. 5, providing insights into sparse training." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "Compared to the \"naive dense large kernel\" of each kernel size, LSK3DNet exhibits superior performance in its corresponding size. LSK3DNet not only leverages SDS to expand its receptive field but also starts from scratch to learn 3D sparse kernels. Concurrently, CWS widens the network during training while pruning redundant channels to accelerate inference. Furthermore, SDS and CWS address the issues of overparameterization and overfitting that arise from simply increasing the kernel's size and expanding the model's width. The best performance of LSK3DNet is achieved at the " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " size, and the following experiments are conducted with " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " size. Moreover, due to the extra dimension in 3D kernels compared to 2D kernels, naively enlarging the 3D kernel size will lead to a cubically-increasing overhead. Our LSK3DNet can effectively reduce 19.1M parameters and approximately " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " of computing operations when compared" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14922" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 128, + 285, + 277 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 125 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 125 + ], + "type": "text", + "content": "Table 5. Segmentation performance of Modified SPVCNN with various large kernel settings. \"Dense\" refers to directly enlarging the kernel size. \"Ours\" refers to training with SDS and CWS (Sec.4.4). The unit for \"Param\" is in million (M), \"FLOPs\" is in billion (G), and \"Speed\" is measured in milliseconds per scene." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 128, + 285, + 277 + ], + "lines": [ + { + "bbox": [ + 50, + 128, + 285, + 277 + ], + "spans": [ + { + "bbox": [ + 50, + 128, + 285, + 277 + ], + "type": "table", + "html": "
Kernel SizemIoUParamFLOPsSpeed ↓
Dense (D=64)3,3,365.21.978.436
5,5,565.68.4127.857
7,7,766.222.6381.265
9,9,967.547.91916.393
11,11,1166.687.43192.0134
Dense (1.8×D)5,5,566.227.0444.084
7,7,766.573.11551.2123
9,9,967.2154.73030.0177
11,11,1166.3282.17139.2250
Ours (1.8×D)5,5,566.85.1203.846
7,7,768.113.6249.463
9,9,970.228.8763.689
11,11,1170.152.51847.4116
", + "image_path": "1933b83cfe8b3851a9b4f36ed3ed7cbcb5406ed61b01e676a8ef168cfdaa2151.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 308, + 285, + 369 + ], + "blocks": [ + { + "bbox": [ + 47, + 283, + 286, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 283, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 283, + 286, + 305 + ], + "type": "text", + "content": "Table 6. Overall architecture design with " + }, + { + "bbox": [ + 47, + 283, + 286, + 305 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 47, + 283, + 286, + 305 + ], + "type": "text", + "content": " size. Here, we report Param, FLOPs, and Speed for the inference model (Sec.4.4)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 308, + 285, + 369 + ], + "lines": [ + { + "bbox": [ + 50, + 308, + 285, + 369 + ], + "spans": [ + { + "bbox": [ + 50, + 308, + 285, + 369 + ], + "type": "table", + "html": "
MethodmIoUParamFLOPsSpeed ↓
Dense (1.0 × D)67.547.91916.393
Dense (1.8 × D)67.2154.73030.0177
SDS (1.8 × D)69.393.02201.4173
CWS (1.8 × D)69.147.91916.393
LSK3DNet (1.8 × D)70.228.8763.689
", + "image_path": "989d85b2ea0bd58290754fa5cebe359c4fb10f7445bbdc8f4b5e31dce2361b4c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 375, + 286, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 286, + 398 + ], + "type": "text", + "content": "to the naive large kernel network. This aspect significantly enhances the value of SDS and CWS for 3D kernels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": "Overall Architecture Design. In Tab. 6, Dense " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "(1.8\\times D)" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " does not yield improvement due to overparameterization and overfitting. Although SDS maintains a model width of " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "1.8\\times D" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " during inference, it achieves improved performance due to sparse training and the large receptive field. CWS effectively addresses overfitting by selecting salient channels, thereby enhancing performance. By combining SDS and CWS, LSK3DNet can achieve the best performance while significantly reducing the model size and computational cost. Spatial-wise Dynamic Sparsity. To control the degree of sparse weight adaptation in LSK3DNet, we have adjusted three hyperparameters: the adaptation frequency " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": ", sparsity rate " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": ", and prune rate " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " (Sec. 3.3). Adaptation frequency " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " indicates how often we update the sparse weights during training. Sparsity rate " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " specifies how sparse the 3D large kernels are. Prune rate " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " shows the fraction of updated weights in one adaptation. We have conducted experiments with different values of these hyperparameters and report the results in Tab. 7 and Tab. 8. Based on our empirical findings, we have chosen " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "f_{a} = 2000" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "s = 0.4" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "inline_equation", + "content": "p = 0.3" + }, + { + "bbox": [ + 46, + 400, + 287, + 652 + ], + "type": "text", + "content": " as the optimal settings for LSK3DNet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Channel-wise Weight Selection. Another aspect of our method that we investigated is the CWS. This technique involves two hyperparameters: the sorting frequency " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " and the width factor " + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": " (Sec.3.4). The former determines how frequently we sort and select channels for large kernels during" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 106, + 545, + 188 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 7. Training and inference speed analysis, speed is measured in milliseconds per scene. The kernels have " + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "text", + "content": " size, \"T\" and \"T\" represent training and inference speed, respectively (Sec.4.4)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 106, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 106, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 106, + 545, + 188 + ], + "type": "table", + "html": "
Sparsity sDense (1.0 × D)SDS (1.8 × D)LSK3DNet (1.8 × D)
mIoUT ↓I ↓mIoUT ↓I ↓mIoUT ↓I ↓
067.545193------
0.1---68.548917669.149192
0.2---68.947817669.748291
0.4---69.346317370.246789
0.6---69.045017168.845487
0.8---68.743917167.244387
", + "image_path": "c530ed13f6f667372379ae7808642b2cebe3e5d1636b1d759e936ec30e75a9c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 220, + 410, + 279 + ], + "blocks": [ + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "type": "text", + "content": "Table 8. Ablation studies of Spatial-wise Dynamic Sparsity on adaptation frequency " + }, + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "type": "text", + "content": " and prune rate " + }, + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 197, + 545, + 218 + ], + "type": "text", + "content": " (Sec.4.4)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 220, + 410, + 279 + ], + "lines": [ + { + "bbox": [ + 316, + 220, + 410, + 279 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 410, + 279 + ], + "type": "table", + "html": "
Adaptation \\( f_{a} \\)mIoU
10068.8
100069.6
200070.2
300070.0
400069.8
", + "image_path": "615dc99e1a275577c550e8003abcb1c275d45af942ebcf57b4abf733343eafbc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 435, + 220, + 529, + 279 + ], + "blocks": [ + { + "bbox": [ + 435, + 220, + 529, + 279 + ], + "lines": [ + { + "bbox": [ + 435, + 220, + 529, + 279 + ], + "spans": [ + { + "bbox": [ + 435, + 220, + 529, + 279 + ], + "type": "table", + "html": "
Pruning pmIoU
0.1069.9
0.2070.1
0.3070.2
0.5070.0
0.7069.7
", + "image_path": "4df0f3522bcbc6a60366d2a6aa0fac375594c59de5410816b0195d1a9aac8583.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 324, + 312, + 412, + 373 + ], + "blocks": [ + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "type": "text", + "content": "Table 9. Ablation studies of Channel-wise Weight Selection on sorting frequency " + }, + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "type": "text", + "content": " and width factor " + }, + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 288, + 545, + 310 + ], + "type": "text", + "content": " (Sec.4.4)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 312, + 412, + 373 + ], + "lines": [ + { + "bbox": [ + 324, + 312, + 412, + 373 + ], + "spans": [ + { + "bbox": [ + 324, + 312, + 412, + 373 + ], + "type": "table", + "html": "
Sorting fsmIoU
1 × fa68.2
2 × fa69.1
4 × fa69.8
6 × fa70.2
10 × fa69.9
", + "image_path": "7f4365b04285acf4ae71985472c452b52104e711a69b81e1cdb850e4d0282454.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 436, + 312, + 529, + 373 + ], + "blocks": [ + { + "bbox": [ + 436, + 312, + 529, + 373 + ], + "lines": [ + { + "bbox": [ + 436, + 312, + 529, + 373 + ], + "spans": [ + { + "bbox": [ + 436, + 312, + 529, + 373 + ], + "type": "table", + "html": "
Width wmIoU
1.1×68.4
1.5×69.6
1.8×70.2
2.1×69.9
2.5×70.1
", + "image_path": "c2cafc85ae976f15903f4cf26290b836edec2eb72bcf1344ac847c1a5d6103b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": "training. It is expressed as a multiple of " + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": ", the adaptation frequency, meaning that we perform channel sorting after a certain number of weight adaptation cycles. When " + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "inline_equation", + "content": "f_{s} = 6 \\times f_{a}" + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": ", the optimal performance is achieved. The latter controls the times of network width compared to the target model. We find that increasing the width factor up to " + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": " improves the model's performance, but beyond that point, there is no significant gain. Therefore, we opt for a width of " + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 304, + 379, + 545, + 498 + ], + "type": "text", + "content": " to avoid more extra compute resources that a wider width would require during training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": "We propose Spatial-wise Dynamic Sparsity to scale up 3D kernels beyond " + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "inline_equation", + "content": "9 \\times 9 \\times 9" + }, + { + "bbox": [ + 304, + 532, + 545, + 713 + ], + "type": "text", + "content": ", which prunes the volumetric weight and reduces the parameter size of large kernel layers. Our LSK3DNet can benefit from a large receptive field without increasing the computational cost compared to a naive 3D large kernel. Channel-wise Weight Selection expands the model width during training, and then sorts and selects important channels during validation to get a model of the expected size. In this way, we achieve \"using spatial sparse groups, expanding width without more parameters\". We evaluate our method on the SemanticKITTI and achieve state-of-the-art performance. Our LSK3DNet also surpasses previous 3D large kernel methods on ScanNet v2 and KITTI. Acknowledgments: This work was supported by a China Scholarship Council (CSC) scholarship." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14923" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Haoran Zhou, Yidan Feng, Mingsheng Fang, Mingqiang Wei, Jing Qin, and Tong Lu. Adaptive graph convolution for point cloud analysis. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 125, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 125, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 287, + 157 + ], + "type": "text", + "content": "[2] Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud representation learning for 3d analysis. In ICCV, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 159, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 159, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 287, + 201 + ], + "type": "text", + "content": "[3] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In ECCV, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 203, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 203, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 203, + 287, + 245 + ], + "type": "text", + "content": "[4] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In CVPR, 2020. 1, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 247, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 247, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 287, + 289 + ], + "type": "text", + "content": "[5] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, 2021. 1, 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 292, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 292, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 58, + 292, + 287, + 334 + ], + "type": "text", + "content": "[6] Junbo Yin, Jianbing Shen, Runnan Chen, Wei Li, Ruigang Yang, Pascal Frossard, and Wenguan Wang. Is-fusion: Instance-scene collaborative fusion for multimodal 3d object detection. In CVPR, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 336, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 336, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 336, + 287, + 368 + ], + "type": "text", + "content": "[7] Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In AAAI, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 369, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 369, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 287, + 400 + ], + "type": "text", + "content": "[8] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 403, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 403, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 287, + 445 + ], + "type": "text", + "content": "[9] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 448, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 448, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 448, + 287, + 490 + ], + "type": "text", + "content": "[10] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 492, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 492, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 492, + 287, + 523 + ], + "type": "text", + "content": "[11] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 525, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 556 + ], + "type": "text", + "content": "[12] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 287, + 601 + ], + "type": "text", + "content": "[13] Junbo Yin, Jin Fang, Dingfu Zhou, Liangjun Zhang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Semi-supervised 3d object detection with proficient teachers. In ECCV,2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 645 + ], + "type": "text", + "content": "[14] Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, ChengZhong Xu, Jianbing Shen, and Wenguan Wang. Proposal-contrast: Unsupervised pre-training for lidar-based 3d object detection. In ECCV,2022.1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, and Dahua Lin. Cylindrical and asymmetrical 3d convolution networks for lidar segmentation. In CVPR, 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] Haotian Tang, Zhijian Liu, Shengyu Zhao, Yujun Lin, Ji Lin, Hanrui Wang, and Song Han. Searching efficient 3d" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 74, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 331, + 74, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 74, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 546, + 94 + ], + "type": "text", + "content": "architectures with sparse point-voxel convolution. In ECCV, 2020.1,2,3,5,6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 96, + 547, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 547, + 126 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 547, + 126 + ], + "type": "text", + "content": "[17] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 129, + 546, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 546, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 546, + 160 + ], + "type": "text", + "content": "[18] Yukang Chen, Jianhui Liu, Xiaojuan Qi, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Scaling up kernels in 3d cnns. CVPR, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 163, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 546, + 194 + ], + "type": "text", + "content": "[19] Yuenan Hou, Xinge Zhu, Yuexin Ma, Chen Change Loy, and Yikang Li. Point-to-voxel knowledge distillation for lidar semantic segmentation. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 196, + 546, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 196, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 196, + 546, + 227 + ], + "type": "text", + "content": "[20] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 228, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 546, + 281 + ], + "type": "text", + "content": "[21] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. arXiv preprint arXiv:2207.03620, 2022. 1, 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 284, + 546, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 546, + 326 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 546, + 326 + ], + "type": "text", + "content": "[22] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In AAAI, 2021. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 328, + 546, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 328, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 312, + 328, + 546, + 369 + ], + "type": "text", + "content": "[23] Jianyun Xu, Ruixiang Zhang, Jian Dou, Yushi Zhu, Jie Sun, and Shiliang Pu. Rpvnet: A deep and efficient range-point-voxel fusion network for lidar point cloud segmentation. In ICCV, 2021. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 372, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 546, + 403 + ], + "type": "text", + "content": "[24] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, 2023. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 405, + 546, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 405, + 546, + 448 + ], + "spans": [ + { + "bbox": [ + 312, + 405, + 546, + 448 + ], + "type": "text", + "content": "[25] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 2, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 449, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 546, + 470 + ], + "type": "text", + "content": "[26] Han Cai, Chuang Gan, Ji Lin, and Song Han. Network augmentation for tiny deep learning. In ICLR, 2022. 2, 4, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 472, + 546, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 472, + 546, + 513 + ], + "spans": [ + { + "bbox": [ + 312, + 472, + 546, + 513 + ], + "type": "text", + "content": "[27] Vidya Kamath and A Renuka. Deep learning based object detection for resource constrained devices-systematic review, future trends and challenges ahead. Neurocomputing, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 515, + 546, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 546, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 546, + 557 + ], + "type": "text", + "content": "[28] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 559, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 546, + 601 + ], + "type": "text", + "content": "[29] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 604, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 546, + 624 + ], + "type": "text", + "content": "[30] Han Hu, Zheng Zhang, Zhenda Xie, and Stephen Lin. Local relation networks for image recognition. In ICCV, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 626, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 626, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 626, + 546, + 667 + ], + "type": "text", + "content": "[31] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-resnet and the impact of residual connections on learning. In AAAI, 2017. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 670, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 546, + 712 + ], + "type": "text", + "content": "[32] Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, and Jian Sun. Large kernel matters-improve semantic segmentation by global convolutional network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14924" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[33] Ruijie Quan, Xuanyi Dong, Yu Wu, Linchao Zhu, and Yi Yang. Auto-reid: Searching for a part-aware convnet for person re-identification. In ICCV, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 139 + ], + "type": "text", + "content": "[34] Wenguan Wang, Guolei Sun, and Luc Van Gool. Looking beyond single images for weakly supervised semantic segmentation learning. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 288, + 172 + ], + "type": "text", + "content": "[35] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "text", + "content": "[36] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 260 + ], + "type": "text", + "content": "[37] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "type": "text", + "content": "[38] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 288, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 337 + ], + "type": "text", + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 288, + 370 + ], + "type": "text", + "content": "[40] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "type": "text", + "content": "[41] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 448 + ], + "type": "text", + "content": "[42] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "text", + "content": "[43] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 524 + ], + "type": "text", + "content": "[44] Siqi Fan, Qiulei Dong, Fenghua Zhu, Yisheng Lv, Peijun Ye, and Fei-Yue Wang. Scf-net: Learning spatial contextual features for large-scale point cloud segmentation. In CVPR, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 525, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 548 + ], + "type": "text", + "content": "[45] Haoxi Ran, Wei Zhuo, Jun Liu, and Li Lu. Learning inner-group relations on point clouds. In ICCV, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 580 + ], + "type": "text", + "content": "[46] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In ICCV, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 581, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 288, + 613 + ], + "type": "text", + "content": "[47] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In CVPR, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 614, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 288, + 647 + ], + "type": "text", + "content": "[48] Qiangeng Xu, Xudong Sun, Cho-Ying Wu, Panqu Wang, and Ulrich Neumann. Grid-gcn for fast and scalable point cloud learning. In CVPR, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 647, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 288, + 680 + ], + "type": "text", + "content": "[49] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 288, + 712 + ], + "type": "text", + "content": "[50] Yanchao Lian, Tuo Feng, and Jinliu Zhou. A dense point-net++ architecture for 3d point cloud semantic segmentation. In IGARSS, 2019." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 547, + 106 + ], + "type": "text", + "content": "[51] Guocheng Qian, Abdulellah Abualshour, Guohao Li, Ali Thabet, and Bernard Ghanem. Pu-gcn: Point cloud upsampling using graph convolutional networks. In CVPR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 547, + 150 + ], + "type": "text", + "content": "[52] Hehe Fan, Xin Yu, Yi Yang, and Mohan Kankanhalli. Deep hierarchical representation of point cloud videos via spatiotemporal decomposition. IEEE TPAMI, 44(12):9918-9930, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 312, + 152, + 547, + 184 + ], + "type": "text", + "content": "[53] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point spatiotemporal transformer networks for point cloud video modeling. IEEE TPAMI, 45(2):2181-2192, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 185, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 185, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 312, + 185, + 547, + 218 + ], + "type": "text", + "content": "[54] Hehe Fan, Xin Yu, Yuhang Ding, Yi Yang, and Mohan Kankanhalli. Pistnet: Point spatio-temporal convolution on point cloud sequences. In ICLR, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 220, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 220, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 312, + 220, + 547, + 252 + ], + "type": "text", + "content": "[55] Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d transformer networks for spatio-temporal modeling in point cloud videos. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 254, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 547, + 297 + ], + "type": "text", + "content": "[56] Qingyong Hu, Bo Yang, Sheikh Khalid, Wen Xiao, Niki Trigoni, and Andrew Markham. Towards semantic segmentation of urban-scale 3d point clouds: A dataset, benchmarks and challenges. In CVPR, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 547, + 331 + ], + "type": "text", + "content": "[57] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 332, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 332, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 312, + 332, + 547, + 374 + ], + "type": "text", + "content": "[58] Bichen Wu, Alvin Wan, Xiangyu Yue, and Kurt Keutzer. Squeezeseg: Convolutional neural nets with recurrent crf for real-time road-object segmentation from 3d lidar point cloud. In ICRA, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 547, + 421 + ], + "type": "text", + "content": "[59] Bichen Wu, Xuanyu Zhou, Sicheng Zhao, Xiangyu Yue, and Kurt Keutzer. Squeezesegv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a lidar point cloud. In ICRA, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 422, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 547, + 464 + ], + "type": "text", + "content": "[60] Tiago Cortinhal, George Tzelepis, and Eren Erdal Aksoy. Salsanext: fast, uncertainty-aware semantic segmentation of lidar point clouds for autonomous driving. arXiv preprint arXiv:2003.03653, 2020." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 467, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 467, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 467, + 547, + 510 + ], + "type": "text", + "content": "[61] Chenfeng Xu, Bichen Wu, Zining Wang, Wei Zhan, Peter Vajda, Kurt Keutzer, and Masayoshi Tomizuka. Squeezeseqv3: Spatially-adaptive convolution for efficient point-cloud segmentation. In ECCV, 2020. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 511, + 547, + 555 + ], + "type": "text", + "content": "[62] Yang Zhang, Zixiang Zhou, Philip David, Xiangyu Yue, Zerong Xi, Boqing Gong, and Hassan Foroosh. Polarnet: An improved grid representation for online lidar point clouds semantic segmentation. In CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 556, + 547, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 547, + 588 + ], + "type": "text", + "content": "[63] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 3, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 590, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 590, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 590, + 547, + 623 + ], + "type": "text", + "content": "[64] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, 2022. 3, 5, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 624, + 547, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 547, + 657 + ], + "type": "text", + "content": "[65] Guillaume Bellec, David Kappel, Wolfgang Maass, and Robert Legenstein. Deep rewiring: Training very sparse deep networks. arXiv preprint arXiv:1711.05136, 2017. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 547, + 712 + ], + "type": "text", + "content": "[66] Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H Nguyen, Madeleine Gibescu, and Antonio Liotta. Scalable training of artificial neural networks with adaptive sparse connectivity inspired by network science. Nature communications, 9(1):2383, 2018. 3, 4" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14925" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[67] Tim Dettmers and Luke Zettlemoyer. Sparse networks from scratch: Faster training without losing performance. arXiv preprint arXiv:1907.04840, 2019.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 288, + 162 + ], + "type": "text", + "content": "[68] Shiwei Liu, Decebal Constantin Mocanu, Amarsagar Reddy Ramapuram Matavalam, Yulong Pei, and Mykola Pechenizkiy. Sparse evolutionary deep learning with over one million artificial neurons on commodity hardware. Neural Computing and Applications, 33:2589-2604, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 195 + ], + "type": "text", + "content": "[69] Utku Evci, Trevor Gale, Jacob Menick, Pablo Samuel Castro, and Erich Elsen. Rigging the lottery: Making all tickets winners. In ICML, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 228 + ], + "type": "text", + "content": "[70] Hesham Mostafa and Xin Wang. Parameter efficient training of deep convolutional neural networks by dynamic sparse reparameterization. In ICML, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 260 + ], + "type": "text", + "content": "[71] Siddhant Jayakumar, Razvan Pascanu, Jack Rae, Simon Osindero, and Erich Elsen. Top-kast: Top-k always sparse training. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 294 + ], + "type": "text", + "content": "[72] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. In NeurIPS, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 294, + 288, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 326 + ], + "type": "text", + "content": "[73] Shiwei Liu, Decebal Constantin Mocanu, Yulong Pei, and Mykola Pechenizkiy. Selfish sparse rn training. In ICML, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 53, + 327, + 288, + 370 + ], + "type": "text", + "content": "[74] Song Han, Jeff Pool, Sharan Narang, Huizi Mao, Enhao Gong, Shijian Tang, Erich Elsen, Peter Vajda, Manohar Paluri, John Tran, et al. Dsd: Dense-sparse-dense training for deep neural networks. In ICLR, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 371, + 288, + 403 + ], + "type": "text", + "content": "[75] Decebal Constantin Mocanu et al. Network computations in artificial intelligence. Technische Universiteit Eindhoven, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 446 + ], + "type": "text", + "content": "[76] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Nest: A neural network synthesis tool based on a grow-and-prune paradigm. IEEE Transactions on Computers, 68(10):1487-1497, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 288, + 481 + ], + "type": "text", + "content": "[77] Xiaoliang Dai, Hongxu Yin, and Niraj K Jha. Grow and prune compact, fast, and accurate lstms. IEEE Transactions on Computers, 69(3):441-452, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 481, + 288, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 504 + ], + "type": "text", + "content": "[78] Md Aamir Raihan and Tor Aamodt. Sparse weight activation training. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 536 + ], + "type": "text", + "content": "[79] Namhoon Lee, Thalaiyasingam Ajanthan, and Philip HS Torr. Snip: Single-shot network pruning based on connection sensitivity. arXiv preprint arXiv:1810.02340, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 537, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 537, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 53, + 537, + 288, + 591 + ], + "type": "text", + "content": "[80] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Li Shen, Decebal Constantin Mocanu, Zhangyang Wang, and Mykola Pechenizkiy. The unreasonable effectiveness of random pruning: Return of the most naive baseline for sparse training. In ICLR, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 592, + 288, + 624 + ], + "type": "text", + "content": "[81] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 288, + 669 + ], + "type": "text", + "content": "[82] Graham H Diering, Raja S Nirujogi, Richard H Roth, Paul F Worley, Akhilesh Pandey, and Richard L Huganir. *Homer1a* drives homeostatic scaling-down of excitatory synapses during sleep. *Science*, 355(6324):511-515, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[83] Luisa De Vivo, Michele Bellesi, William Marshall, Eric A Bushong, Mark H Ellisman, Giulio Tononi, and Chiara Cirelli. Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science, 355(6324):507-510, 2017. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[84] Ran Cheng, Ryan Razani, Ehsan Taghavi, Enxu Li, and Bingbing Liu. (af)2-s3net: Attentive feature fusion with adaptive feature selection for sparse semantic segmentation network. In CVPR, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 547, + 161 + ], + "type": "text", + "content": "[85] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[86] Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lovasz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. In CVPR, 2018. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 205, + 547, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 547, + 247 + ], + "type": "text", + "content": "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 249, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 249, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 312, + 249, + 545, + 293 + ], + "type": "text", + "content": "[88] Qinghao Meng, Wenguan Wang, Tianfei Zhou, Jianbing Shen, Luc Van Gool, and Dengxin Dai. Weakly supervised 3d object detection from lidar point cloud. In ECCV, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 293, + 545, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 293, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 312, + 293, + 545, + 336 + ], + "type": "text", + "content": "[89] Radu Alexandru Rosu, Peer Schütt, Jan Quenzel, and Sven Behnke. Latticenet: Fast point cloud segmentation using permutohedral lattices. arXiv preprint arXiv:1912.05905, 2019.6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 338, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 371 + ], + "type": "text", + "content": "[90] Fabian Duerr, Mario Pfaller, Hendrik Weigel, and Jürgen Beyerer. Lidar-based recurrent 3d semantic segmentation with temporal memory alignment. In 3DV, 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 545, + 403 + ], + "type": "text", + "content": "[91] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multi-view prediction for 3d semantic scene segmentation. In ECCV, 2018. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "type": "text", + "content": "[92] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 545, + 470 + ], + "type": "text", + "content": "[93] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 471, + 545, + 502 + ], + "type": "text", + "content": "[94] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 503, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 503, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 312, + 503, + 545, + 536 + ], + "type": "text", + "content": "[95] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 580 + ], + "type": "text", + "content": "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 545, + 613 + ], + "type": "text", + "content": "[97] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 657 + ], + "type": "text", + "content": "[98] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 690 + ], + "type": "text", + "content": "[99] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[100] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 6, 7" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14926" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 217 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "text", + "content": "[101] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[102] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 287, + 173 + ], + "type": "text", + "content": "[103] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 217 + ], + "type": "text", + "content": "[104] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 7" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14927" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_content_list.json b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..86b072002691b664dd450628f3e3dfd684a34b02 --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling", + "text_level": 1, + "bbox": [ + 187, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiaheng Liu $^{1*}$ , Jianhao Li $^{1*}$ , Kaisiyuan Wang $^{3}$ , Hongcheng Guo $^{1}$ , Jian Yang $^{1}$ , Junran Peng $^{2}$ , Ke Xu $^{1}$ , Xianglong Liu $^{1}$ , Jinyang Guo $^{1\\dagger}$", + "bbox": [ + 176, + 179, + 790, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beihang University $^{2}$ Institute of Automation, Chinese Academy of Sciences $^{3}$ The University of Sydney", + "bbox": [ + 176, + 215, + 792, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, many approaches directly operate on point clouds for different tasks. These approaches become more computation and storage demanding when point cloud size is large. To reduce the required computation and storage, one possible solution is to sample the point cloud. In this paper, we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework. Existing task-agnostic point cloud sampling strategy (e.g., FPS) does not consider semantic information of point clouds, causing degraded performance on downstream tasks. While learning-based point cloud sampling methods consider semantic information, they are task-specific and require task-oriented ground-truth annotations. So they cannot generalize well on different downstream tasks. Our LTA-PCS achieves task-agnostic point cloud sampling without requiring task-oriented labels, in which both the geometric and semantic information of points is considered in sampling. Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS.", + "bbox": [ + 75, + 318, + 473, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 637, + 209, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D point cloud has been widely used in many areas like autonomous driving and robotics. However, the point clouds collected by LiDAR or other sensors contain millions or even billions of points, making it challenging to process, store, and transmit them efficiently. Therefore, to improve computational efficiency and reduce memory costs, it is critical to develop point cloud sampling methods to reduce the number of points while preserving the essential features of the original point clouds.", + "bbox": [ + 75, + 662, + 468, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among point cloud sampling methods, the most widely used one is the farthest point sampling (FPS) strategy in Fig. 1(a). It samples the critical points based on their 3D coordinates iteratively and selects a group of points that", + "bbox": [ + 75, + 801, + 470, + 862 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a09d4f79be3b936d458399d759655c882de173dd15cd354a980314777899c59c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 287, + 888, + 338 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/639e8aff6fa3336d3c4766d69905bb6f61f3e8403cc7098d05fcd14f02faa9b7.jpg", + "image_caption": [ + "(a) FPS", + "(b) Learnable Task-specific methods" + ], + "image_footnote": [], + "bbox": [ + 504, + 351, + 887, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91731ea9eb165e108319db947258a4e810b7899e89e1ac1c7044c5d97d857fe8.jpg", + "image_caption": [ + "(c) Learnable Task-agnostic method (Ours)", + "Figure 1. Comparison of FPS, learnable task-specific methods, and our LTA-PCS. FPS samples points using 3D coordinates, which ignores semantic meaning. Learnable Task-specific methods need different sampling networks for different tasks. Our LTA-PCS is task-agnostic, where one trained sampling network can be used for different tasks. Note that the pretrained 2D feature extractor is frozen in training, and not used at inference." + ], + "image_footnote": [], + "bbox": [ + 506, + 526, + 887, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are farthest apart from each other. Although FPS is task-agnostic, it does not utilize the semantic information of point clouds as the sampling process is solely based on the point coordinates, causing degraded performance on downstream tasks. Recently, in Fig. 1 (b), learning-based point cloud sampling methods [7, 15] are proposed to sample the", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* First two authors contributed equally.", + "bbox": [ + 99, + 875, + 308, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 99, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28035", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "points with more semantic meanings for better downstream task performance. They additionally introduce task-specific losses when learning the network for point cloud sampling. However, these methods are task-specific and require task-oriented labels. In real-world applications, the downstream tasks are usually not specified. But we need to sample fewer points in the collected point cloud by LiDAR or other sensors to save the storage. Moreover, even downstream tasks are given, the existing learning-based methods need to sample the points again based on the new downstream task for better performance, which is cumbersome in practice. Besides, the existing learning-based methods use the pretrained 3D feature extractor (e.g., PointNet [27]) to calculate the task loss for training the sampling network, while the capacity of these 3D feature extractors is limited by both model size and the scale of 3D point cloud datasets, which becomes the bottleneck for training the sampling network with semantically meaningful sampling ability. Finally, the task-specific loss makes them prone to overfit on specific datasets and hard to generalize on different datasets, and the requirement of labels increases the training cost. Although we can remove the task-specific loss in these learning-based methods for task-agnostic point cloud sampling, the sampled points still lack semantic meaning in this case.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "So the problem is: How to design a learnable task-agnostic label-free point cloud sampling strategy with the consideration of semantic information? To solve this problem, in Fig. 1(c), we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework, in which the semantic meaning of points is considered in the sampling process. It is non-trivial to construct such a framework. Without the specific downstream tasks, the designed sampling loss for learning-based sampling methods is based on the coordinates, which lack semantic meaning.", + "bbox": [ + 76, + 458, + 472, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In LTA-PCS, we observe 2D feature extractors [31] are effective tools to extract task-agnostic but semantically meaningful features and propose to introduce the fixed 2D feature extractors pre-trained on large-scale datasets to help the training of our LTA-PCS for sampling more semantic meaningful points, where the specific tasks and the corresponding labels are both not required. Specifically, to solve the heterogeneous problem of image and point cloud data, we first project 3D point clouds to multi-view depth maps, which can be readily used by 2D feature extractors to extract semantic meaningful features. Then, we calculate the loss based on both the 3D coordinates and the extracted features and train the sampling network, which can enforce it to recognize meaningful points in terms of both geometric and semantic aspects. After training, we directly use the trained sampling network to generate the simplified point clouds for different downstream tasks, in which an inference alignment strategy is used to ensure the simplified point cloud is a subset of the original dense point cloud.", + "bbox": [ + 76, + 613, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the aforementioned training paradigm, another issue is that the projection from point clouds to multi-view depth maps will inevitably cause information loss, causing inaccurate learning of semantic meaningful points. To this end, we propose a new loss function called semantic loss, which includes both intra- and inter-view losses. Specifically, the intra-view loss minimizes the information loss between the dense point cloud and simplified point cloud under the same views, while the inter-view loss is calculated based on the relationship between different views. By introducing the semantic loss, we can consider both the information under the same and cross views, which can help the sampling network to generate more semantic meaningful points.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, our LTA-PCS framework is also general. Although it is designed for task-agnostic point cloud sampling, we can further combine it with the task-specific loss for task-specific point cloud sampling to achieve better performance when the downstream task is given.", + "bbox": [ + 496, + 287, + 890, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of LTA-PCS are shown as follows:", + "bbox": [ + 517, + 363, + 870, + 377 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we propose the first learnable task-agnostic point cloud sampling framework called LTA-PCS, in which both the geometric and semantic information are considered and the task-oriented labels are not required.", + "- We introduce the pre-trained 2D feature extractor to compute the inter- and intra-view learning objective functions to help train the sampling network in our LTA-PCS framework, which can enforce the sampling network to sample more semantically meaningful points.", + "- Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS." + ], + "bbox": [ + 500, + 378, + 890, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 571, + 640, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning on point clouds. Recently, deep learning methods [18, 22, 23, 37-39, 48, 53] attract many attentions, in which a large number of methods using point clouds are proposed for 3D scene understanding [9-11, 17, 19-21, 46, 49, 50]. For example, PointNet [27] uses multilayer perceptron to process points for different tasks. PointNet++ [28] proposes a hierarchical structure to extract features. Li et al. [16] proposed PointCNN to perform convolution operations on point clouds. Wu et al. [45] proposed PointConv to introduce point density when perform convolution on point clouds. There are also many methods on point cloud segmentation [40, 41, 43] and object detection [29, 34, 56]. In addition, approaches are also proposed to focus on point cloud completion [4, 32, 54], point cloud registration [2, 33]. On the other hand, some methods project point clouds into volumetric [24, 57] or multiview [13, 31, 35, 52] data forms and process them using neural networks. For example, Zhang et al. [55] project point clouds into multi-view depth images and process them using pretrained CLIP model [31]. Goyal et al. [8] propose", + "bbox": [ + 496, + 598, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28036", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e7bc20b5e7992f8fc801702ec9a33b09f61ee8884c344389c6b3539474f7a6f0.jpg", + "table_caption": [ + "Table 1. Comparison between our LTA-PCS framework with existing point cloud sampling methods." + ], + "table_footnote": [], + "table_body": "
MethodsSemantic informationTask-agnostic
FPS×✓
Learning-based methods [5, 7, 15]✓×
LTA-PCS✓✓
", + "bbox": [ + 86, + 128, + 452, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SimpleView projection methods for point cloud classification. Zhu et al. [58] utilize pretrained CLIP model for 3D open-world learning. Although these methods use CNNs or CLIP models for learning point cloud features, they do not focus on point cloud sampling. Besides, we utilize the 2D feature extractor in a different perspective, in which the 2D feature extractors are used to facilitate the training of our LTA-PCS for more semantically meaningful sampling.", + "bbox": [ + 75, + 241, + 467, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud sampling methods. Recently, many point cloud sampling methods were proposed. In Table 1, for the task-agnostic point cloud sampling, the FPS method samples the point that is farthest from the previously sampled ones iteratively. However, this method only uses point coordinates, which does not consider the semantic information of points and leads to a sub-optimal solution when using the sampled points for downstream tasks. Besides, many learning-based approaches were proposed [5, 30, 44]. For example, Dovrat et al. [7] proposed S-NET to use networks for point cloud sampling, where a sampling regularization loss is used to train this S-NET. Lang et al. [15] proposed SampleNet to use differentiable relaxation for point cloud sampling. Qian et al. [30] proposed a task-oriented point cloud downsampling approach. Attention operation is also used for point cloud sampling [42, 51]. Although these methods use networks to sample semantic meaningful points from a dense point cloud, they are task-oriented. Therefore, these approaches need specific downstream tasks to train the sampling network, which is often infeasible in real-world applications. In contrast, our LTA-PCS framework is task-agnostic, and the comparison between our LTA-PCS and the existing methods is shown in Table 1.", + "bbox": [ + 75, + 364, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 729, + 210, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Statement", + "text_level": 1, + "bbox": [ + 76, + 755, + 261, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a dense point cloud with $n$ points $P = \\{p_i \\in \\mathbb{R}^3, i = 1, \\dots, n\\}$ and a target size $m \\leq n$ , our goal in this work is to learn a sampling network that can sample $m$ points from $P$ and generate a simplified point cloud that can best represent the dense point cloud with task-agnostic semantic meaning as follows:", + "bbox": [ + 75, + 777, + 467, + 868 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT (P) \\subset P, | T (P) | = m \\leq n, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 883, + 467, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T(\\cdot)$ is the operation of the sampling network, and $T(P)$ is the simplified point cloud. In our LTA-PCS, the learning process of the sampling network $T(\\cdot)$ is not related to any downstream tasks. In other words, we aim to find the best sampling network to generate the simplified point cloud that can perform well on the downstream tasks. Traditional point cloud sampling methods (e.g., FPS) only sample the points based on 3D coordinates, which lacks semantic information. To introduce semantic information in sampling, learning-based sampling methods like S-NET [7] and SampleNet [15] use the task-specific loss to train the sampling network. However, when the task is not given, it is still an open problem on how to introduce such a semantic-aware loss for guiding the sampling network to sample semantically meaningful points. Therefore, we propose to use effective and well-performed pretrained 2D feature extractors to facilitate the training of the sampling network for LTA-PCS in a task-agnostic way.", + "bbox": [ + 496, + 90, + 890, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 375, + 609, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows the overview when training our LTA-PCS framework. Given a dense point cloud, we use the sampling network in LTA-PCS to sample points from this point cloud and generate the simplified point cloud. To utilize the pretrained 2D feature extractors, we first project both the dense and simplified point clouds to multi-view depth maps. Then, we use the pretrained 2D feature extractor to extract semantically meaningful features based on multi-view depth maps. In our implementation, we choose the vision branch of the CLIP model [31] as our pretrained 2D feature extractor, and the parameters of the 2D feature extractor are frozen in the training process. After that, we calculate the proposed semantic loss including both inter- and intraview losses based on the output feature from 2D feature extractors, which introduces the semantic-level supervision to help the sampling network in our LTA-PCS framework and sample more semantically meaningful points. Besides, to fully utilize the 3D coordinate information of point clouds, we also use geometric loss [7] to introduce geometric-level supervision when training our LTA-PCS framework in the training process. We update the sampling network based on both geometric loss and semantic loss.", + "bbox": [ + 496, + 398, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sampling network. Our sampling model follows the architecture of [7, 27]. The input points are processed by a series of $1 \\times 1$ convolution layers, which produce a feature vector for each point. Then, a symmetric feature-wise max pooling operation is applied to obtain a global feature vector. Finally, we use several fully-connected layers. The output of the last layer is the set of generated points.", + "bbox": [ + 496, + 732, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projection. Directly using pretrained 2D feature extractor is infeasible because of the heterogeneous problem between 2D images and 3D point clouds. To solve this problem, we need to project point clouds to multi-view", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28037", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/57606d2da4553e99f02718743d3b3534e58960f02d043d4dbe5e129bfb8e0c7d.jpg", + "image_caption": [ + "Figure 2. Training procedure of our LTA-PCS framework. Given a dense point cloud, we first use the initialized sampling network to sample points from it and generate a simplified point cloud. Then, we project both the dense and simplified point clouds to multi-view depth maps, which will be used for feature extraction by a frozen pretrained 2D feature extractor. We calculate the semantic loss including intra-view and inter-view losses using the extracted features. Geometric loss is also used in training. At the inference stage, we directly use the trained sampling network to sample points from dense point clouds." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "depth maps. Specifically, inspired by [55], we take the bottom view as an example, the point with the coordinate of $(x,y,z)$ will be projected to $(\\lceil x / z\\rceil ,\\lceil y / z\\rceil)$ . Also, as the vision branch of CLIP model aims to process images with three channels, we copy the projected depth maps three times and form the input of the CLIP model, which ensures there is no shape mismatch when using the CLIP model.", + "bbox": [ + 75, + 431, + 470, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference. At inference, we use the sampling network in Fig. 2 to sample points from the dense point cloud, which can be used for multiple downstream tasks.", + "bbox": [ + 75, + 546, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 628, + 220, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall loss function. Formally, the overall loss function of our LTA-PCS can be written as follows:", + "bbox": [ + 75, + 662, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {g e o} + \\alpha L _ {s e m}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 724, + 468, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{geo}$ is the geometric loss, and $L_{sem}$ is the newly proposed semantic loss, which includes both the intra- and inter-view losses. $\\alpha$ is the coefficient to balance these two terms. In LTA-PCS, the geometric and semantic losses aim to preserve the geometric and semantic consistency between the original point cloud and the simplified point cloud.", + "bbox": [ + 75, + 770, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometric loss. We denote $Q$ as the simplified point cloud, i.e., $Q = T(P)$ . Inspired by [7], the geometric loss", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is defined as follows:", + "bbox": [ + 500, + 431, + 643, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g e o} = L _ {a} (Q, P) + \\beta L _ {w} (Q, P) + \\gamma L _ {s} (Q, P),\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 457, + 836, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} L _ {a} (Q, P) = \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 477, + 836, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {w} (Q, P) = \\max _ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 512, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} (Q, P) = \\frac {1}{| P |} \\sum_ {p \\in P} m i n _ {q \\in Q} | | p - q | | _ {2} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 542, + 836, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $L_{a}(Q,P)$ is to minimize the average distance between $Q$ and $P$ . $L_{w}(Q,P)$ is to minimize the maximum distance between $Q$ and $P$ . $L_{s}(Q,P)$ ensures the points in $Q$ evenly spread in the 3D coordinate space. Following [7], $\\beta$ and $\\gamma$ are the loss weights.", + "bbox": [ + 500, + 594, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic loss. To enforce the sampling network to sample more semantically meaningful points, we also introduce the newly proposed semantic loss $L_{sem}$ in the training process. Formally, let us suppose the dense point cloud as $P$ and the simplified point cloud generated by the initialized sampling network as $Q$ . The semantic loss consists of intraview and inter-view loss, which can be written as follows:", + "bbox": [ + 496, + 670, + 890, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s e m} = L _ {i n t r a} + \\lambda L _ {i n t e r}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 794, + 890, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intra-view loss aims to minimize the semantic information loss under the same view as follows:", + "bbox": [ + 500, + 821, + 888, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t r a} = \\sum_ {k = 1} ^ {K} M S E [ E (J _ {k} (P)), E (J _ {k} (Q)) ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 862, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28038", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training procedure of our LTA-PCS.", + "text_level": 1, + "bbox": [ + 80, + 90, + 410, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Dense point cloud $P$ ; Simplified point cloud $Q$ ; Randomly initialized sampling network $T$ ; Number of projected views $K$ ; Pretrained 2D feature extractor $E$ ;", + "bbox": [ + 80, + 109, + 468, + 155 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for each iteration in the training process do", + "2: // Calculate geometric loss $L_{geo}$", + "3: Use $T$ to generate the simplified point cloud $Q$ ;", + "4: Calculate geometric loss $L_{geo}$ based on dense point cloud $P$ and the simplified point cloud $Q$ using Eq. (3);", + "5: // Calculate semantic loss $L_{sem}$", + "6: for view $k$ in $K$ do", + "7: Project dense point cloud $P$ and simplified point cloud $Q$ to the $k$ -th view depth map $J_{k}(P)$ and $J_{k}(Q)$ , respectively;", + "8: Extract the features of $J_{k}(P)$ and $J_{k}(Q)$ using the 2D feature extractor $E$ , respectively;", + "9: end for", + "10: Calculate intra- and intra-view losses by Eq. (5) and Eq. (6);", + "11: Calculate semantic loss $L_{sem}$ based on Eq. (4);", + "12: Calculate overall object function $L$ using Eq. (2);", + "13: Back propagate the gradients and update the parameters of the sampling network $T$ ;", + "14: end for" + ], + "bbox": [ + 81, + 156, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: The optimized sampling network $T$", + "bbox": [ + 78, + 457, + 382, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $J_{k}(\\cdot)$ denotes the projection operation for the $k$ -th view, and $E(\\cdot)$ is the 2D feature extractor for extracting semantic features. $MSE(\\cdot)$ is the mean square error, and $K$ is the total number of projected views.", + "bbox": [ + 76, + 510, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the projection operation inevitably causes information loss and the relationship of different views can also provide essential information [25, 26], we calculate the interview loss for compensation as follows:", + "bbox": [ + 76, + 575, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t e r} = M S E \\left(\\boldsymbol {\\Phi} \\boldsymbol {\\Phi} ^ {T}, \\boldsymbol {\\Psi} \\boldsymbol {\\Psi} ^ {T}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 660, + 383, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\Phi = \\underset {K} {\\operatorname {C o n c a t}} \\left[ \\operatorname {F l a t} \\left(E \\left(J _ {k} (P)\\right)\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 681, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi = \\underset {K} {C o n c a t} \\left[ F l a t (E (J _ {k} (Q))) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 705, + 405, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $Flat(\\cdot)$ is the flatten operation. $Concat_{K}$ concatenate the vectors along different views. By using the interview loss, we obtain the relation information in different views from dense point cloud and enforce the simplified point cloud to mimic this information.", + "bbox": [ + 76, + 744, + 468, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the intra-view loss $L_{intra}$ , inter-view loss $L_{inter}$ , and the geometric loss $L_{geo}$ , we obtain the loss function $L$ in Eq. (2) to train the sampling network $T$ , which is used to generate the simplified point cloud at inference. The training procedure is also illustrated in Alg. 1.", + "bbox": [ + 76, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference", + "text_level": 1, + "bbox": [ + 500, + 90, + 609, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At inference, we use the trained sampling network $T$ to generate the coordinates of the sampled points. However, as we use several fully-connected layers to directly predict the coordinates of the sampled points at the end of the sampling network, we cannot ensure the simplified point cloud $Q$ generated by the sampling network to be a subset of dense point cloud $P$ . Thus, following S-Net [7], we can obtain the points in $Q$ to its nearest neighbors in dense point cloud $P$ . Specifically, for each point $q \\in \\mathbb{R}^3$ in $Q$ , the matched point $q^*$ is written as follows:", + "bbox": [ + 496, + 114, + 890, + 265 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq ^ {*} = \\underset {p \\in P} {\\operatorname {a r g m i n}} | | q - p | | _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 277, + 890, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the matching process for $Q$ with $m$ points, we can obtain the simplified point cloud $Q^{*} = \\{q_{i}^{*}\\}_{i = 1}^{m}$ , which is a subset of $P$ . Note that the matching process is very fast.", + "bbox": [ + 496, + 316, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Downstream Tasks Usage", + "text_level": 1, + "bbox": [ + 500, + 373, + 732, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training the sampling network in our LTA-PCS, we can use it to effectively generate the simplified point cloud with the information from both 3D coordinates and the semantic meanings. As the downstream tasks are not decided at this stage, we store these simplified point clouds as a new point cloud dataset to simulate this scenario. When the downstream tasks are given, we use the simplified point clouds to train different backbone network architectures (e.g., PointNet [27]) for different downstream tasks as before, and the backbone network is initialized from scratch.", + "bbox": [ + 496, + 397, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Discussion", + "text_level": 1, + "bbox": [ + 500, + 560, + 617, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Necessity of Learnable Task-Agnostic Point Cloud Sampling. We would emphasize that LTA-PCS tries to solve the point cloud sampling under an important scenario where the collected point clouds are given but the downstream tasks are not given. In real-world applications, this scenario often occurs. We often use LiDAR to collect point clouds and store them. When specific tasks are given, we use the stored point cloud data for downstream tasks. Thus, sampling semantically meaningful points from dense collected point clouds becomes critical for storage saving. Existing learning-based methods [15] use task-specific loss to introduce semantic information in sampling, which is infeasible in this scenario as the downstream tasks and the corresponding labels are required. Although we can remove task-specific loss in these learning-based methods for task-agnostic point cloud sampling, semantic information will not be considered in this case. Moreover, the task-specific loss will degrade the generalization ability of these methods on different datasets. LTA-PCS targets at solving the point cloud sampling under a new setting, which cannot be achieved by existing learning-based approaches.", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28039", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion on the reason that our task-agnostic point cloud sampling method (i.e., LTA-PCS) yields better results relative to task-specific methods (e.g., S-Net, SampleNet). We hypothesize this is because we introduce the 2D pretrained model in the training process. Therefore, the knowledge in 2D pretrained model can significantly improve the performance of the sampling network. In contrast, due to limited network capacity or dataset size, there is no general and well-performed 3D pretrained model when compared with 2D pretrained models. Therefore, although existing state-of-the-art point cloud sampling methods are task-intensive, they cannot well utilize the prior knowledge in 2D pretrained model, resulting worse performance than our task-agnostic method LTA-PCS. Our LTA-PCS framework can be also combined with task-specific loss, and the results show the performance can be further boosted.", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion on the necessity of 2D feature extractor. We choose the pretrained 2D feature extractor to introduce semantic information when training our LTA-PCS framework as there is no existing well-performed pretrained 3D feature extractor on large-scale datasets in the literature. Current 3D feature extractors like PointNet [27] or PointNet++ [28] are trained based on specific tasks and datasets, which cannot be used in the task-agnostic setting. We hypothesize that this is because of the lack of large-scale high-quality 3D point cloud datasets. Therefore, although we encounter information loss in the projection process, we choose a 2D feature extractor to introduce semantic information when constructing our LTA-PCS.", + "bbox": [ + 75, + 334, + 472, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 547, + 209, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 573, + 294, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow [7] to utilize a set of multi-layer perceptrons (MLPs) as the sampling network, and batch normalization and ReLU layers are appended after each MLP. In training, we train the sampling network for 50 epochs with Adam optimizer [14] on both ModelNet40, ShapeNet core55 and ScanObjectNN datasets. The batch size is set as 8 and the learning rate is $5e^{-4}$ . We set the weight decay as $1e^{-4}$ . In the training process, we set the hyperparameter $\\alpha$ in Eq. (2) as 1. $\\beta$ and $\\gamma$ in Eq. (3) are set as 1 and 1, respectively, and $\\lambda$ in Eq. (4) is set as 0.01. The number of views (i.e., $K$ ) is set as 10. After training the sampling network, we use it to generate simplified point clouds. Then, we use the generated simplified point clouds to train an initialized backbone network for different downstream tasks. For the point cloud classification task, we use PointNet [27] as our backbone network and train the initialized PointNet for 200 epochs with a batch size of 12. The learning rate is set as $1e^{-3}$ with weight decay of $1e^{-4}$ . Step learning rate decay is used. The learning rate will multiply 0.7 after every 20 epochs. For the point cloud completion and registration tasks, we", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "use PCN [54] and PointNetLK [1] as baseline methods, respectively, and the training epochs are set as 400 and 200, respectively.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Point Cloud Classification", + "text_level": 1, + "bbox": [ + 498, + 145, + 736, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we compare our LTA-PCS with several existing point cloud sampling methods on the ModelNet40 [47], ShapeNet [3] and ScanObjectNN (hardest perturbed variant (PB_T50_RS)) [36] for point cloud classification, where the instance accuracy is used as the evaluation metric. Specifically, Random and FPS are task-agnostic non-learnable methods, and we directly sample a specific number of points by using the corresponding methods. The baseline approaches S-Net [7], SampleNet [15] are task-specific learnable methods, and we first train the sampling network and use the trained network to downsample the input point cloud to a predefined number of points. However, we can remove the task-specific loss in these learnable sampling approaches to achieve task-agnostic point cloud sampling, which are denoted by S-Net* and SampleNet*, respectively. In contrast, LTA-PCS is a learnable approach without task-specific loss in training. Therefore, it is a task-agnostic point cloud sampling method. In Table 2, we observe that our LTA-PCS achieves better results than other methods. Remarkably, we outperform the state-of-the-art SampleNet* under the task-agnostic setting by $2.1\\%$ when sampling to 256 points on ModelNet40. Moreover, LTA-PCS even performs better than the task-specific learnable point cloud sampling methods S-Net and SampleNet. One possible explanation is that S-Net and SampleNet use pretrained 3D feature extractors (e.g., PointNet [27]) to calculate the task loss, and their results are limited by the capacity of 3D feature extractors.", + "bbox": [ + 496, + 169, + 893, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Point Cloud Retrieval", + "text_level": 1, + "bbox": [ + 498, + 602, + 705, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also follow S-Net [7] to report the results of the point cloud retrieval task. For point cloud retrieval, we use ModelNet40 to evaluate our LTA-PCS framework, where the mean average precision (mAP) evaluation metric is used. In Table 3, our LTA-PCS achieves promising performance. Specifically, we first downsample the point cloud to the corresponding number of points, and then extract the features by PointNet. Following [7], the results are computed by $L_{2}$ distance of the shape descriptor, which is the feature of the penultimate layer.", + "bbox": [ + 496, + 625, + 893, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Point Cloud Completion", + "text_level": 1, + "bbox": [ + 498, + 786, + 723, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow PCN [54] to reconstruct the simplified point clouds into the original dense point clouds with 2,048 points on ShapeNet Core55, and we provide the $L_{1}$ Chamfer Distance (CD) results of different methods in Fig. 3. Note that for the task-specific methods, we need to additionally train the sampling network when training the PCN method. From", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28040", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3e9d3417494ec47559b4264c6e8c54ecb90ad25b77f6913e12ba1e20dbb599bf.jpg", + "table_caption": [ + "Table 2. Comparison of our LTA-PCS with other methods under the point cloud classification task. * indicates we remove the task-specific loss in training for task-agnostic setting. \"Oracle\" denotes the accuracy using original dense point clouds." + ], + "table_footnote": [], + "table_body": "
DatasetsModelNet (Oracle: 90.3)ShapeNet (Oracle: 85.0)
#Sampled points6412825651264128256512
Task-AgnosticRandom81.586.286.688.279.380.981.983.2
FPS86.187.988.188.380.581.582.483.3
S-Net* [7]86.187.888.288.580.681.382.283.0
SampleNet* [15]86.688.387.988.480.381.682.482.9
LTA-PCS88.289.190.090.382.083.183.584.4
Task-SpecificS-Net [7]87.788.188.489.080.982.382.783.6
SampleNet [15]87.988.388.088.781.082.582.783.4
", + "bbox": [ + 163, + 128, + 803, + 296 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c0e017188ca4333f0fc28e23bc7f9e7dbe89200b28fc1a9cc639436658b1ae63.jpg", + "table_caption": [ + "Table 3. Comparison of our LTA-PCS with other methods under the point cloud retrieval task. * indicates we remove the task-specific loss in training for task-agnostic setting." + ], + "table_footnote": [], + "table_body": "
#Sampled points64128256512
Task-Agnostic
Random66.169.472.573.3
FPS69.172.573.074.2
S-Net* [7]69.471.773.674.1
SampleNet* [15]70.472.372.674.7
LTA-PCS71.473.274.877.2
Task-Specific
S-Net [7]70.572.774.076.2
SampleNet [15]70.972.573.676.0
", + "bbox": [ + 107, + 369, + 436, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43949d0a2cf7f86b16f404b23e8ed2a520b91e2f61594e1073695ce9b421a7e3.jpg", + "image_caption": [ + "Figure 3. Performance on the point cloud completion task. Lower $L_{1}$ Chamber Distance ( $L_{1}$ CD) denotes better performance." + ], + "image_footnote": [], + "bbox": [ + 117, + 570, + 428, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 3, we observe that the quantitative results of our LTA-PCS are better than other methods under different numbers of sampled points. The results show our LTA-PCS is effective for different downstream tasks.", + "bbox": [ + 75, + 824, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. We provide the visualization results of dif", + "bbox": [ + 76, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67249e5907f907dc8eb9a74b3df33bbd0840092a8175338d7ece78cb81ef125d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 316, + 619, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2b2339d6ca3e998e3c70e6b9b168daf3849aea1d03a7c0852928e966b373a531.jpg", + "image_caption": [ + "(a) Simplified Point Clouds" + ], + "image_footnote": [], + "bbox": [ + 622, + 316, + 730, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bc7da72e66a44494ed9853d8d4bb794fb684ec868c21581016f4e59a82be310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 318, + 839, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72670124e1ff0375e23e17baeb240a0ae37aa6779b6c9d7f72da1ac08c0b23e8.jpg", + "image_caption": [ + "FPS" + ], + "image_footnote": [], + "bbox": [ + 552, + 479, + 612, + 619 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/405ca913013e76355f628aba34b52ed1b8ef4943e9c2a0476fde4789cb783d98.jpg", + "image_caption": [ + "S-Net", + "(b) Completed Point Clouds" + ], + "image_footnote": [], + "bbox": [ + 666, + 483, + 728, + 621 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c64edf7b5c65fc1cc71e7aa59d0e4759ee95ae4e2debde5753d3097f3528ce7e.jpg", + "image_caption": [ + "LTA-PCS", + "Figure 4. (a). Simplified point clouds. (The sampled points are enlarged for better visualization.) (b). Completed point clouds by using the corresponding simplified point clouds as input." + ], + "image_footnote": [], + "bbox": [ + 777, + 479, + 841, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferent methods in Fig. 4. Specifically, in Fig. 4(a), the simplified point clouds are produced by FPS, S-Net and LTA-PCS, where the number of points in each simplified point cloud are 64, and in Fig. 4(b), we observe that the reconstruction quality of LTA-PCS is better than other methods.", + "bbox": [ + 496, + 732, + 893, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 816, + 653, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different losses. In Table 4 we provide three alternative variants of our LTA-PCS: LTA-PCS (w/o inter & w/o intra), LTA-PCS (w/o inter), LTA-PCS (w/o intra). Specifically, for LTA-PCS (w/o inter & w/o intra), we remove the", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28041", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/830cdd102554f83472e7d9a75ace05d1653f650ee19122348cb1fa640cfff76f.jpg", + "table_caption": [ + "Table 4. Ablation on using different losses." + ], + "table_footnote": [], + "table_body": "
#Sampled points128256
LTA-PCS89.190.0
LTA-PCS (w/o inter & w/o intra)87.888.2
LTA-PCS (w/o inter)88.689.4
LTA-PCS (w/o intra)88.389.1
", + "bbox": [ + 102, + 114, + 439, + 207 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4777b606c138c56074ae186d20a4aae98be20c0c14dc06f48c896d66dde0051b.jpg", + "table_caption": [ + "Table 5. Ablation on using different feature extractors." + ], + "table_footnote": [], + "table_body": "
#Sampled points128
LTA-PCS (PointNet)88.0
LTA-PCS (R50 trained on ImageNet)88.3
LTA-PCS (R50 of CLIP vision encoder)89.1
", + "bbox": [ + 106, + 250, + 441, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1f95826a4604520d2d2a255a809fd9db0c2a7a4b19bd4985fba4953f5ac42c59.jpg", + "table_caption": [ + "Table 6. Ablation on using different numbers of views." + ], + "table_footnote": [], + "table_body": "
# views2610
LTA-PCS88.288.989.1
", + "bbox": [ + 155, + 369, + 387, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "semantic losses (i.e., both inter-view and intra-view losses), which is the same as the S-Net [7] without using the task loss. For LTA-PCS (w/o inter), we only use the intra-view loss without using the inter-view loss to train our LTA-PCS framework. For LTA-LTA-PCS (w/o intra), we only use the inter-view loss without using the intra-view loss to train the LTA-PCS framework. In Table 4, our LTA-PCS is better than these three alternative variants, which shows that it is beneficial to utilize both intra-view and inter-view losses.", + "bbox": [ + 75, + 446, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of 2D feature extractor. In Table 5, we compare the results of our LTA-PCS by using different pretrained vision backbones [6, 12] to show the effect of 2D feature extractor. Specifically, for LTA-PCS (PointNet), we directly use the PointNet backbone to extract the features of the original dense point cloud and simplified point cloud and calculate the MSE distance of corresponding features to preserve semantic consistency. For LTA-PCS (R50 trained on ImageNet), we use the ResNet-50 [12] model trained on ImageNet as the feature extractor. For LTA-PCS (R50 of CLIP vision encoder) and LTA-PCS (VIT-B/16 of CLIP vision encoder), we use the pretrained CLIP vision encoder with different backbones. In Table 5, first, the results of LTA-PCS (PointNet) are lower than other methods a lot, which indicates that the 3D feature extractor (i.e., PointNet) cannot provide high-quality semantic supervision to train the sampling network. Second, we observe that the results of LTA-PCS (R50 of CLIP vision encoder) are better than the results of LTA-PCS (R50 trained on ImageNet) by a large margin, which shows the capacity of the feature extractor is critical when training the sampling network. Therefore,", + "bbox": [ + 75, + 583, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we choose the pre-trained 2D feature extractor to maintain semantic consistency when training the sampling network instead of using the pre-trained 3D feature extractor.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the number of views. In Table 6, we report the performance results of using 128 sampled points as input to analyze the effect of the number of projected views, where we report the results on the ModelNet40 dataset based on the PointNet backbone. When the number of views increases from 2 to 10, LTA-PCS can achieve better performance results, which indicates that it is effective to use more views for maintaining the semantic consistency between the original and simplified point clouds. Therefore, we set the projected number of views as 10 to achieve better performance with acceptable memory usage.", + "bbox": [ + 496, + 136, + 890, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Further Analysis", + "text_level": 1, + "bbox": [ + 500, + 311, + 666, + 328 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c98702e118e200faab87238d5556ad2d249fda66e879dfc5c5187816d1a00b8.jpg", + "table_caption": [ + "Table 7. Inference time and GPU memory usage of our LTA-PCS with different sampled points." + ], + "table_footnote": [], + "table_body": "
#Sampled pointsInference time (ms)GPU usage (MB)
640.28578
1280.37616
2560.58724
5121.05940
20481.772326
", + "bbox": [ + 504, + 388, + 888, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison on inference time and GPU memory usage. In Table 7, we take the ModelNet40 dataset as an example and compare the inference time and GPU memory usage under different sampled points when using PointNet as the backbone network. As shown in Table 7, we observe that using fewer points can reduce inference time and GPU memory usage greatly, which further demonstrates the advantages of point cloud sampling.", + "bbox": [ + 496, + 503, + 890, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 638, + 617, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose the Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework for task-agnostic point cloud sampling task, which aims to preserve the geometric and semantic consistency in point cloud sampling. Specifically, we first propose to use the pre-trained 2D feature extractor for training the sampling network, and use semantic loss including both inter-view and intra-view losses. Comprehensive experimental results on multiple downstream tasks show the effectiveness of LTA-PCS. In the future, we will continue to explore more tasks based on our LTA-PCS method.", + "bbox": [ + 496, + 664, + 890, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by National Science and Technology Major Project (2022ZD0116405) and National Natural Science Foundation of China (No. 61932002, No. 62306025, No. 92367204).", + "bbox": [ + 496, + 830, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28042", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6", + "[2] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2", + "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 6", + "[4] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. arXiv preprint arXiv:1904.00069, 2019. 2", + "[5] Ta-Ying Cheng, Qingyong Hu, Qian Xie, Niki Trigoni, and Andrew Markham. Meta-sampler: Almost-universal yet task-oriented sampling for point clouds. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 694-710, Berlin, Heidelberg, 2022. Springer-Verlag. 3", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8", + "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 3, 4, 5, 6, 7, 8", + "[8] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning. PMLR, 2021. 2", + "[9] Jinyang Guo, Jiaheng Liu, and Dong Xu. Jointpruning: Pruning networks along multiple dimensions for efficient point cloud processing. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2", + "[10] Jinyang Guo, Jiaheng Liu, and Dong Xu. 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8717-8729, 2022.", + "[11] Jinyang Guo, Dong Xu, and Wanli Ouyang. Multidimensional pruning and its extension: A unified framework for model compression. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Asako Kanezaki, Yasuyuki Matsushita, and Yoshifumi Nishida. Rotationnet: Joint object categorization and pose estimation using multiviews from unsupervised viewpoints. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2", + "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6", + "[15] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6, 7", + "[16] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeruIPS, 2018. 2", + "[17] Jiaheng Liu and Dong Xu. Geometrhythm-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31 (12):4711-4721, 2021. 2", + "[18] Jiaheng Liu, Yudong Wu, Yichao Wu, Chuming Li, Xiaolin Hu, Ding Liang, and Mengyu Wang. Dam: discrepancy alignment metric for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3814-3823, 2021. 2", + "[19] Jiaheng Liu, Jinyang Guo, and Dong Xu. Apsnet: Towards adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing, 2022. 2", + "[20] Jiaheng Liu, Jinyang Guo, and Dong Xu. Geometriontransformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia, 2022.", + "[21] Jiaheng Liu, Tong He, Honghui Yang, Rui Su, Jiayi Tian, Junran Wu, Hongcheng Guo, Ke Xu, and Wanli Ouyang. 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375, 2022. 2", + "[22] Jiaheng Liu, Haoyu Qin, Yichao Wu, Jinyang Guo, Ding Liang, and Ke Xu. Coupleface: Relation matters for face recognition distillation. In European Conference on Computer Vision, pages 683-700. Springer, 2022. 2", + "[23] Jiaheng Liu, Zhipeng Yu, Haoyu Qin, Yichao Wu, Ding Liang, Gangming Zhao, and Ke Xu. Oneface: one threshold for all. In European Conference on Computer Vision, pages 545-561. Springer, 2022. 2", + "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2015. 2", + "[25] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 5", + "[26] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In ICCV, pages 5007-5016, 2019. 5", + "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2, 3, 5, 6", + "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6", + "[29] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 2", + "[30] Yue Qian, Junhui Hou, Qijian Zhang, Yiming Zeng, Sam Kwong, and Ying He. Mops-net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. arXiv preprint arXiv:2005.00383, 2020. 3", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 3", + "[32] Muhammad Sarmad, Hyunjoo Jenny Lee, and Young Min Kim. Rl-gan-net: A reinforcement learning agent controlled gan network for real-time point cloud shape completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[33] Vinit Sarode, Xueqian Li, Hunter Goforth, Yasuhiro Aoki, Rangaprasad Arun Srivatsan, Simon Lucey, and Howie Choset. Pcrnet: Point cloud registration network using point-net encoding. arXiv preprint arXiv:1908.07906, 2019. 2", + "[34] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 2", + "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, 2015. 2", + "[36] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6", + "[37] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2", + "[38] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 8565-8574. Computer Vision Foundation / IEEE, 2021.", + "[39] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2022. 2", + "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE conference on computer vision and pattern recognition, 2018. 2", + "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[42] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. Lightn: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. arXiv preprint arXiv:2202.06263, 2022. 3", + "[43] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2", + "[44] C. Wen, B. Yu, and D. Tao. Learnable skeleton-aware 3d point cloud sampling. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3", + "[45] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2", + "[46] Yue Wu, Xidao Hu, Yue Zhang, Maoguo Gong, Wenping Ma, and Qiguang Miao. Sacf-net: Skip-attention based correspondence filtering network for point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3585-3595, 2023. 2", + "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6", + "[48] Ge Yang, Chao Zhang, Ling Gao, Yufei Guo, and Jinyang Guo. Domain adaptive channel pruning. *Electronics*, 13(5), 2024. 2", + "[49] Honghui Yang, Tong He, Jiaheng Liu, Hua Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wanli Ouyang. Gd-mae: generative decoder for mae pre-training on lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9403-9414, 2023. 2", + "[50] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 2", + "[51] Yang Ye, Xiulong Yang, and Shihao Ji. Apsnet: Attention based point cloud sampling. arXiv preprint arXiv:2210.05638, 2022. 3", + "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2", + "[53] Zhipeng Yu, Jiaheng Liu, Haoyu Qin, Yichao Wu, Kun Hu, Jiayi Tian, and Ding Liang. Icd-face: Intra-class compactness distillation for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21042-21052, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28044", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV). IEEE, 2018. 2,6", + "[55] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4", + "[56] Lichen Zhao, Jinyang Guo, Dong Xu, and Lu Sheng. Transformer3d-det: Improving 3d object detection by vote refinement. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2", + "[57] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2", + "[58] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 358 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28045", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_model.json b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1d6d5b4f19df2b23d913da44671c9d886e1270c7 --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.783, + 0.154 + ], + "angle": 0, + "content": "LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.18, + 0.791, + 0.217 + ], + "angle": 0, + "content": "Jiaheng Liu\\(^{1*}\\), Jianhao Li\\(^{1*}\\), Kaisiyuan Wang\\(^{3}\\), Hongcheng Guo\\(^{1}\\), Jian Yang\\(^{1}\\), Junran Peng\\(^{2}\\), Ke Xu\\(^{1}\\), Xianglong Liu\\(^{1}\\), Jinyang Guo\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.217, + 0.794, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Beihang University \\(^{2}\\)Institute of Automation, Chinese Academy of Sciences \\(^{3}\\)The University of Sydney" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.607 + ], + "angle": 0, + "content": "Recently, many approaches directly operate on point clouds for different tasks. These approaches become more computation and storage demanding when point cloud size is large. To reduce the required computation and storage, one possible solution is to sample the point cloud. In this paper, we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework. Existing task-agnostic point cloud sampling strategy (e.g., FPS) does not consider semantic information of point clouds, causing degraded performance on downstream tasks. While learning-based point cloud sampling methods consider semantic information, they are task-specific and require task-oriented ground-truth annotations. So they cannot generalize well on different downstream tasks. Our LTA-PCS achieves task-agnostic point cloud sampling without requiring task-oriented labels, in which both the geometric and semantic information of points is considered in sampling. Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.638, + 0.21, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.8 + ], + "angle": 0, + "content": "3D point cloud has been widely used in many areas like autonomous driving and robotics. However, the point clouds collected by LiDAR or other sensors contain millions or even billions of points, making it challenging to process, store, and transmit them efficiently. Therefore, to improve computational efficiency and reduce memory costs, it is critical to develop point cloud sampling methods to reduce the number of points while preserving the essential features of the original point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.802, + 0.471, + 0.863 + ], + "angle": 0, + "content": "Among point cloud sampling methods, the most widely used one is the farthest point sampling (FPS) strategy in Fig. 1(a). It samples the critical points based on their 3D coordinates iteratively and selects a group of points that" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.288, + 0.889, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.34, + 0.722, + 0.352 + ], + "angle": 0, + "content": "(a) FPS" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.352, + 0.888, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.515, + 0.817, + 0.527 + ], + "angle": 0, + "content": "(b) Learnable Task-specific methods" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.527, + 0.888, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.649, + 0.844, + 0.661 + ], + "angle": 0, + "content": "(c) Learnable Task-agnostic method (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.673, + 0.894, + 0.771 + ], + "angle": 0, + "content": "Figure 1. Comparison of FPS, learnable task-specific methods, and our LTA-PCS. FPS samples points using 3D coordinates, which ignores semantic meaning. Learnable Task-specific methods need different sampling networks for different tasks. Our LTA-PCS is task-agnostic, where one trained sampling network can be used for different tasks. Note that the pretrained 2D feature extractor is frozen in training, and not used at inference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "are farthest apart from each other. Although FPS is task-agnostic, it does not utilize the semantic information of point clouds as the sampling process is solely based on the point coordinates, causing degraded performance on downstream tasks. Recently, in Fig. 1 (b), learning-based point cloud sampling methods [7, 15] are proposed to sample the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.875, + 0.31, + 0.888 + ], + "angle": 0, + "content": "* First two authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.875, + 0.31, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28035" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "points with more semantic meanings for better downstream task performance. They additionally introduce task-specific losses when learning the network for point cloud sampling. However, these methods are task-specific and require task-oriented labels. In real-world applications, the downstream tasks are usually not specified. But we need to sample fewer points in the collected point cloud by LiDAR or other sensors to save the storage. Moreover, even downstream tasks are given, the existing learning-based methods need to sample the points again based on the new downstream task for better performance, which is cumbersome in practice. Besides, the existing learning-based methods use the pretrained 3D feature extractor (e.g., PointNet [27]) to calculate the task loss for training the sampling network, while the capacity of these 3D feature extractors is limited by both model size and the scale of 3D point cloud datasets, which becomes the bottleneck for training the sampling network with semantically meaningful sampling ability. Finally, the task-specific loss makes them prone to overfit on specific datasets and hard to generalize on different datasets, and the requirement of labels increases the training cost. Although we can remove the task-specific loss in these learning-based methods for task-agnostic point cloud sampling, the sampled points still lack semantic meaning in this case." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.459, + 0.473, + 0.61 + ], + "angle": 0, + "content": "So the problem is: How to design a learnable task-agnostic label-free point cloud sampling strategy with the consideration of semantic information? To solve this problem, in Fig. 1(c), we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework, in which the semantic meaning of points is considered in the sampling process. It is non-trivial to construct such a framework. Without the specific downstream tasks, the designed sampling loss for learning-based sampling methods is based on the coordinates, which lack semantic meaning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In LTA-PCS, we observe 2D feature extractors [31] are effective tools to extract task-agnostic but semantically meaningful features and propose to introduce the fixed 2D feature extractors pre-trained on large-scale datasets to help the training of our LTA-PCS for sampling more semantic meaningful points, where the specific tasks and the corresponding labels are both not required. Specifically, to solve the heterogeneous problem of image and point cloud data, we first project 3D point clouds to multi-view depth maps, which can be readily used by 2D feature extractors to extract semantic meaningful features. Then, we calculate the loss based on both the 3D coordinates and the extracted features and train the sampling network, which can enforce it to recognize meaningful points in terms of both geometric and semantic aspects. After training, we directly use the trained sampling network to generate the simplified point clouds for different downstream tasks, in which an inference alignment strategy is used to ensure the simplified point cloud is a subset of the original dense point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "In the aforementioned training paradigm, another issue is that the projection from point clouds to multi-view depth maps will inevitably cause information loss, causing inaccurate learning of semantic meaningful points. To this end, we propose a new loss function called semantic loss, which includes both intra- and inter-view losses. Specifically, the intra-view loss minimizes the information loss between the dense point cloud and simplified point cloud under the same views, while the inter-view loss is calculated based on the relationship between different views. By introducing the semantic loss, we can consider both the information under the same and cross views, which can help the sampling network to generate more semantic meaningful points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Moreover, our LTA-PCS framework is also general. Although it is designed for task-agnostic point cloud sampling, we can further combine it with the task-specific loss for task-specific point cloud sampling to achieve better performance when the downstream task is given." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.364, + 0.871, + 0.378 + ], + "angle": 0, + "content": "The contributions of LTA-PCS are shown as follows:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.379, + 0.892, + 0.453 + ], + "angle": 0, + "content": "- To the best of our knowledge, we propose the first learnable task-agnostic point cloud sampling framework called LTA-PCS, in which both the geometric and semantic information are considered and the task-oriented labels are not required." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.455, + 0.892, + 0.529 + ], + "angle": 0, + "content": "- We introduce the pre-trained 2D feature extractor to compute the inter- and intra-view learning objective functions to help train the sampling network in our LTA-PCS framework, which can enforce the sampling network to sample more semantically meaningful points." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.53, + 0.892, + 0.559 + ], + "angle": 0, + "content": "- Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.379, + 0.892, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.642, + 0.59 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Deep learning on point clouds. Recently, deep learning methods [18, 22, 23, 37-39, 48, 53] attract many attentions, in which a large number of methods using point clouds are proposed for 3D scene understanding [9-11, 17, 19-21, 46, 49, 50]. For example, PointNet [27] uses multilayer perceptron to process points for different tasks. PointNet++ [28] proposes a hierarchical structure to extract features. Li et al. [16] proposed PointCNN to perform convolution operations on point clouds. Wu et al. [45] proposed PointConv to introduce point density when perform convolution on point clouds. There are also many methods on point cloud segmentation [40, 41, 43] and object detection [29, 34, 56]. In addition, approaches are also proposed to focus on point cloud completion [4, 32, 54], point cloud registration [2, 33]. On the other hand, some methods project point clouds into volumetric [24, 57] or multiview [13, 31, 35, 52] data forms and process them using neural networks. For example, Zhang et al. [55] project point clouds into multi-view depth images and process them using pretrained CLIP model [31]. Goyal et al. [8] propose" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28036" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison between our LTA-PCS framework with existing point cloud sampling methods." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.129, + 0.454, + 0.215 + ], + "angle": 0, + "content": "
MethodsSemantic informationTask-agnostic
FPS×✓
Learning-based methods [5, 7, 15]✓×
LTA-PCS✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.468, + 0.363 + ], + "angle": 0, + "content": "SimpleView projection methods for point cloud classification. Zhu et al. [58] utilize pretrained CLIP model for 3D open-world learning. Although these methods use CNNs or CLIP models for learning point cloud features, they do not focus on point cloud sampling. Besides, we utilize the 2D feature extractor in a different perspective, in which the 2D feature extractors are used to facilitate the training of our LTA-PCS for more semantically meaningful sampling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.365, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Point cloud sampling methods. Recently, many point cloud sampling methods were proposed. In Table 1, for the task-agnostic point cloud sampling, the FPS method samples the point that is farthest from the previously sampled ones iteratively. However, this method only uses point coordinates, which does not consider the semantic information of points and leads to a sub-optimal solution when using the sampled points for downstream tasks. Besides, many learning-based approaches were proposed [5, 30, 44]. For example, Dovrat et al. [7] proposed S-NET to use networks for point cloud sampling, where a sampling regularization loss is used to train this S-NET. Lang et al. [15] proposed SampleNet to use differentiable relaxation for point cloud sampling. Qian et al. [30] proposed a task-oriented point cloud downsampling approach. Attention operation is also used for point cloud sampling [42, 51]. Although these methods use networks to sample semantic meaningful points from a dense point cloud, they are task-oriented. Therefore, these approaches need specific downstream tasks to train the sampling network, which is often infeasible in real-world applications. In contrast, our LTA-PCS framework is task-agnostic, and the comparison between our LTA-PCS and the existing methods is shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.73, + 0.212, + 0.746 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.262, + 0.77 + ], + "angle": 0, + "content": "3.1. Problem Statement" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Given a dense point cloud with \\( n \\) points \\( P = \\{p_i \\in \\mathbb{R}^3, i = 1, \\dots, n\\} \\) and a target size \\( m \\leq n \\), our goal in this work is to learn a sampling network that can sample \\( m \\) points from \\( P \\) and generate a simplified point cloud that can best represent the dense point cloud with task-agnostic semantic meaning as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.885, + 0.468, + 0.902 + ], + "angle": 0, + "content": "\\[\nT (P) \\subset P, | T (P) | = m \\leq n, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.364 + ], + "angle": 0, + "content": "where \\( T(\\cdot) \\) is the operation of the sampling network, and \\( T(P) \\) is the simplified point cloud. In our LTA-PCS, the learning process of the sampling network \\( T(\\cdot) \\) is not related to any downstream tasks. In other words, we aim to find the best sampling network to generate the simplified point cloud that can perform well on the downstream tasks. Traditional point cloud sampling methods (e.g., FPS) only sample the points based on 3D coordinates, which lacks semantic information. To introduce semantic information in sampling, learning-based sampling methods like S-NET [7] and SampleNet [15] use the task-specific loss to train the sampling network. However, when the task is not given, it is still an open problem on how to introduce such a semantic-aware loss for guiding the sampling network to sample semantically meaningful points. Therefore, we propose to use effective and well-performed pretrained 2D feature extractors to facilitate the training of the sampling network for LTA-PCS in a task-agnostic way." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.376, + 0.61, + 0.39 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.892, + 0.732 + ], + "angle": 0, + "content": "Fig. 2 shows the overview when training our LTA-PCS framework. Given a dense point cloud, we use the sampling network in LTA-PCS to sample points from this point cloud and generate the simplified point cloud. To utilize the pretrained 2D feature extractors, we first project both the dense and simplified point clouds to multi-view depth maps. Then, we use the pretrained 2D feature extractor to extract semantically meaningful features based on multi-view depth maps. In our implementation, we choose the vision branch of the CLIP model [31] as our pretrained 2D feature extractor, and the parameters of the 2D feature extractor are frozen in the training process. After that, we calculate the proposed semantic loss including both inter- and intraview losses based on the output feature from 2D feature extractors, which introduces the semantic-level supervision to help the sampling network in our LTA-PCS framework and sample more semantically meaningful points. Besides, to fully utilize the 3D coordinate information of point clouds, we also use geometric loss [7] to introduce geometric-level supervision when training our LTA-PCS framework in the training process. We update the sampling network based on both geometric loss and semantic loss." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Sampling network. Our sampling model follows the architecture of [7, 27]. The input points are processed by a series of \\(1 \\times 1\\) convolution layers, which produce a feature vector for each point. Then, a symmetric feature-wise max pooling operation is applied to obtain a global feature vector. Finally, we use several fully-connected layers. The output of the last layer is the set of generated points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Projection. Directly using pretrained 2D feature extractor is infeasible because of the heterogeneous problem between 2D images and 3D point clouds. To solve this problem, we need to project point clouds to multi-view" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28037" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Figure 2. Training procedure of our LTA-PCS framework. Given a dense point cloud, we first use the initialized sampling network to sample points from it and generate a simplified point cloud. Then, we project both the dense and simplified point clouds to multi-view depth maps, which will be used for feature extraction by a frozen pretrained 2D feature extractor. We calculate the semantic loss including intra-view and inter-view losses using the extracted features. Geometric loss is also used in training. At the inference stage, we directly use the trained sampling network to sample points from dense point clouds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.471, + 0.538 + ], + "angle": 0, + "content": "depth maps. Specifically, inspired by [55], we take the bottom view as an example, the point with the coordinate of \\((x,y,z)\\) will be projected to \\((\\lceil x / z\\rceil ,\\lceil y / z\\rceil)\\). Also, as the vision branch of CLIP model aims to process images with three channels, we copy the projected depth maps three times and form the input of the CLIP model, which ensures there is no shape mismatch when using the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.593 + ], + "angle": 0, + "content": "Inference. At inference, we use the sampling network in Fig. 2 to sample points from the dense point cloud, which can be used for multiple downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.63, + 0.221, + 0.645 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.692 + ], + "angle": 0, + "content": "Overall loss function. Formally, the overall loss function of our LTA-PCS can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.725, + 0.47, + 0.742 + ], + "angle": 0, + "content": "\\[\nL = L _ {g e o} + \\alpha L _ {s e m}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.771, + 0.469, + 0.862 + ], + "angle": 0, + "content": "where \\( L_{geo} \\) is the geometric loss, and \\( L_{sem} \\) is the newly proposed semantic loss, which includes both the intra- and inter-view losses. \\( \\alpha \\) is the coefficient to balance these two terms. In LTA-PCS, the geometric and semantic losses aim to preserve the geometric and semantic consistency between the original point cloud and the simplified point cloud." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Geometric loss. We denote \\( Q \\) as the simplified point cloud, i.e., \\( Q = T(P) \\). Inspired by [7], the geometric loss" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.433, + 0.644, + 0.446 + ], + "angle": 0, + "content": "is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.458, + 0.838, + 0.476 + ], + "angle": 0, + "content": "\\[\nL _ {g e o} = L _ {a} (Q, P) + \\beta L _ {w} (Q, P) + \\gamma L _ {s} (Q, P),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.478, + 0.838, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} L _ {a} (Q, P) = \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.513, + 0.891, + 0.542 + ], + "angle": 0, + "content": "\\[\nL _ {w} (Q, P) = \\max _ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.544, + 0.838, + 0.581 + ], + "angle": 0, + "content": "\\[\nL _ {s} (Q, P) = \\frac {1}{| P |} \\sum_ {p \\in P} m i n _ {q \\in Q} | | p - q | | _ {2} ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.595, + 0.891, + 0.67 + ], + "angle": 0, + "content": "Here, \\( L_{a}(Q,P) \\) is to minimize the average distance between \\( Q \\) and \\( P \\). \\( L_{w}(Q,P) \\) is to minimize the maximum distance between \\( Q \\) and \\( P \\). \\( L_{s}(Q,P) \\) ensures the points in \\( Q \\) evenly spread in the 3D coordinate space. Following [7], \\( \\beta \\) and \\( \\gamma \\) are the loss weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Semantic loss. To enforce the sampling network to sample more semantically meaningful points, we also introduce the newly proposed semantic loss \\( L_{sem} \\) in the training process. Formally, let us suppose the dense point cloud as \\( P \\) and the simplified point cloud generated by the initialized sampling network as \\( Q \\). The semantic loss consists of intraview and inter-view loss, which can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.795, + 0.891, + 0.812 + ], + "angle": 0, + "content": "\\[\nL _ {s e m} = L _ {i n t r a} + \\lambda L _ {i n t e r}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.822, + 0.89, + 0.851 + ], + "angle": 0, + "content": "The intra-view loss aims to minimize the semantic information loss under the same view as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.863, + 0.892, + 0.905 + ], + "angle": 0, + "content": "\\[\nL _ {i n t r a} = \\sum_ {k = 1} ^ {K} M S E [ E (J _ {k} (P)), E (J _ {k} (Q)) ], \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28038" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.092, + 0.411, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Training procedure of our LTA-PCS." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.111, + 0.47, + 0.156 + ], + "angle": 0, + "content": "Input: Dense point cloud \\(P\\); Simplified point cloud \\(Q\\); Randomly initialized sampling network \\(T\\); Number of projected views \\(K\\); Pretrained 2D feature extractor \\(E\\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.157, + 0.393, + 0.171 + ], + "angle": 0, + "content": "1: for each iteration in the training process do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.172, + 0.343, + 0.187 + ], + "angle": 0, + "content": "2: // Calculate geometric loss \\( L_{geo} \\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.187, + 0.445, + 0.201 + ], + "angle": 0, + "content": "3: Use \\( T \\) to generate the simplified point cloud \\( Q \\);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.202, + 0.469, + 0.232 + ], + "angle": 0, + "content": "4: Calculate geometric loss \\( L_{geo} \\) based on dense point cloud \\( P \\) and the simplified point cloud \\( Q \\) using Eq. (3);" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.233, + 0.34, + 0.246 + ], + "angle": 0, + "content": "5: // Calculate semantic loss \\( L_{sem} \\)" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.248, + 0.261, + 0.261 + ], + "angle": 0, + "content": "6: for view \\(k\\) in \\(K\\) do" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.262, + 0.469, + 0.307 + ], + "angle": 0, + "content": "7: Project dense point cloud \\( P \\) and simplified point cloud \\( Q \\) to the \\( k \\)-th view depth map \\( J_{k}(P) \\) and \\( J_{k}(Q) \\), respectively;" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.308, + 0.469, + 0.337 + ], + "angle": 0, + "content": "8: Extract the features of \\( J_{k}(P) \\) and \\( J_{k}(Q) \\) using the 2D feature extractor \\( E \\), respectively;" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.338, + 0.185, + 0.351 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.353, + 0.469, + 0.383 + ], + "angle": 0, + "content": "10: Calculate intra- and intra-view losses by Eq. (5) and Eq. (6);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.383, + 0.442, + 0.398 + ], + "angle": 0, + "content": "11: Calculate semantic loss \\( L_{sem} \\) based on Eq. (4);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.399, + 0.455, + 0.413 + ], + "angle": 0, + "content": "12: Calculate overall object function \\(L\\) using Eq. (2);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.414, + 0.468, + 0.443 + ], + "angle": 0, + "content": "13: Back propagate the gradients and update the parameters of the sampling network \\( T \\);" + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.444, + 0.161, + 0.456 + ], + "angle": 0, + "content": "14: end for" + }, + { + "type": "list", + "bbox": [ + 0.082, + 0.157, + 0.469, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.458, + 0.383, + 0.473 + ], + "angle": 0, + "content": "Output: The optimized sampling network \\( T \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.511, + 0.469, + 0.572 + ], + "angle": 0, + "content": "where \\( J_{k}(\\cdot) \\) denotes the projection operation for the \\( k \\)-th view, and \\( E(\\cdot) \\) is the 2D feature extractor for extracting semantic features. \\( MSE(\\cdot) \\) is the mean square error, and \\( K \\) is the total number of projected views." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.576, + 0.469, + 0.636 + ], + "angle": 0, + "content": "As the projection operation inevitably causes information loss and the relationship of different views can also provide essential information [25, 26], we calculate the interview loss for compensation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.661, + 0.384, + 0.679 + ], + "angle": 0, + "content": "\\[\nL _ {i n t e r} = M S E \\left(\\boldsymbol {\\Phi} \\boldsymbol {\\Phi} ^ {T}, \\boldsymbol {\\Psi} \\boldsymbol {\\Psi} ^ {T}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.682, + 0.469, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\Phi = \\underset {K} {\\operatorname {C o n c a t}} \\left[ \\operatorname {F l a t} \\left(E \\left(J _ {k} (P)\\right)\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.707, + 0.406, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\Psi = \\underset {K} {C o n c a t} \\left[ F l a t (E (J _ {k} (Q))) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.746, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Here, \\( Flat(\\cdot) \\) is the flatten operation. \\( Concat_{K} \\) concatenate the vectors along different views. By using the interview loss, we obtain the relation information in different views from dense point cloud and enforce the simplified point cloud to mimic this information." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on the intra-view loss \\( L_{intra} \\), inter-view loss \\( L_{inter} \\), and the geometric loss \\( L_{geo} \\), we obtain the loss function \\( L \\) in Eq. (2) to train the sampling network \\( T \\), which is used to generate the simplified point cloud at inference. The training procedure is also illustrated in Alg. 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.61, + 0.105 + ], + "angle": 0, + "content": "3.4. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.266 + ], + "angle": 0, + "content": "At inference, we use the trained sampling network \\( T \\) to generate the coordinates of the sampled points. However, as we use several fully-connected layers to directly predict the coordinates of the sampled points at the end of the sampling network, we cannot ensure the simplified point cloud \\( Q \\) generated by the sampling network to be a subset of dense point cloud \\( P \\). Thus, following S-Net [7], we can obtain the points in \\( Q \\) to its nearest neighbors in dense point cloud \\( P \\). Specifically, for each point \\( q \\in \\mathbb{R}^3 \\) in \\( Q \\), the matched point \\( q^* \\) is written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.279, + 0.892, + 0.306 + ], + "angle": 0, + "content": "\\[\nq ^ {*} = \\underset {p \\in P} {\\operatorname {a r g m i n}} | | q - p | | _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.363 + ], + "angle": 0, + "content": "After the matching process for \\(Q\\) with \\(m\\) points, we can obtain the simplified point cloud \\(Q^{*} = \\{q_{i}^{*}\\}_{i = 1}^{m}\\), which is a subset of \\(P\\). Note that the matching process is very fast." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.374, + 0.733, + 0.39 + ], + "angle": 0, + "content": "3.5. Downstream Tasks Usage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.548 + ], + "angle": 0, + "content": "After training the sampling network in our LTA-PCS, we can use it to effectively generate the simplified point cloud with the information from both 3D coordinates and the semantic meanings. As the downstream tasks are not decided at this stage, we store these simplified point clouds as a new point cloud dataset to simulate this scenario. When the downstream tasks are given, we use the simplified point clouds to train different backbone network architectures (e.g., PointNet [27]) for different downstream tasks as before, and the backbone network is initialized from scratch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.618, + 0.575 + ], + "angle": 0, + "content": "3.6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Necessity of Learnable Task-Agnostic Point Cloud Sampling. We would emphasize that LTA-PCS tries to solve the point cloud sampling under an important scenario where the collected point clouds are given but the downstream tasks are not given. In real-world applications, this scenario often occurs. We often use LiDAR to collect point clouds and store them. When specific tasks are given, we use the stored point cloud data for downstream tasks. Thus, sampling semantically meaningful points from dense collected point clouds becomes critical for storage saving. Existing learning-based methods [15] use task-specific loss to introduce semantic information in sampling, which is infeasible in this scenario as the downstream tasks and the corresponding labels are required. Although we can remove task-specific loss in these learning-based methods for task-agnostic point cloud sampling, semantic information will not be considered in this case. Moreover, the task-specific loss will degrade the generalization ability of these methods on different datasets. LTA-PCS targets at solving the point cloud sampling under a new setting, which cannot be achieved by existing learning-based approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Discussion on the reason that our task-agnostic point cloud sampling method (i.e., LTA-PCS) yields better results relative to task-specific methods (e.g., S-Net, SampleNet). We hypothesize this is because we introduce the 2D pretrained model in the training process. Therefore, the knowledge in 2D pretrained model can significantly improve the performance of the sampling network. In contrast, due to limited network capacity or dataset size, there is no general and well-performed 3D pretrained model when compared with 2D pretrained models. Therefore, although existing state-of-the-art point cloud sampling methods are task-intensive, they cannot well utilize the prior knowledge in 2D pretrained model, resulting worse performance than our task-agnostic method LTA-PCS. Our LTA-PCS framework can be also combined with task-specific loss, and the results show the performance can be further boosted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Discussion on the necessity of 2D feature extractor. We choose the pretrained 2D feature extractor to introduce semantic information when training our LTA-PCS framework as there is no existing well-performed pretrained 3D feature extractor on large-scale datasets in the literature. Current 3D feature extractors like PointNet [27] or PointNet++ [28] are trained based on specific tasks and datasets, which cannot be used in the task-agnostic setting. We hypothesize that this is because of the lack of large-scale high-quality 3D point cloud datasets. Therefore, although we encounter information loss in the projection process, we choose a 2D feature extractor to introduce semantic information when constructing our LTA-PCS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.21, + 0.566 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.295, + 0.59 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We follow [7] to utilize a set of multi-layer perceptrons (MLPs) as the sampling network, and batch normalization and ReLU layers are appended after each MLP. In training, we train the sampling network for 50 epochs with Adam optimizer [14] on both ModelNet40, ShapeNet core55 and ScanObjectNN datasets. The batch size is set as 8 and the learning rate is \\(5e^{-4}\\). We set the weight decay as \\(1e^{-4}\\). In the training process, we set the hyperparameter \\(\\alpha\\) in Eq. (2) as 1. \\(\\beta\\) and \\(\\gamma\\) in Eq. (3) are set as 1 and 1, respectively, and \\(\\lambda\\) in Eq. (4) is set as 0.01. The number of views (i.e., \\(K\\)) is set as 10. After training the sampling network, we use it to generate simplified point clouds. Then, we use the generated simplified point clouds to train an initialized backbone network for different downstream tasks. For the point cloud classification task, we use PointNet [27] as our backbone network and train the initialized PointNet for 200 epochs with a batch size of 12. The learning rate is set as \\(1e^{-3}\\) with weight decay of \\(1e^{-4}\\). Step learning rate decay is used. The learning rate will multiply 0.7 after every 20 epochs. For the point cloud completion and registration tasks, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "use PCN [54] and PointNetLK [1] as baseline methods, respectively, and the training epochs are set as 400 and 200, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.146, + 0.738, + 0.162 + ], + "angle": 0, + "content": "4.2. Point Cloud Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.895, + 0.593 + ], + "angle": 0, + "content": "In Table 2, we compare our LTA-PCS with several existing point cloud sampling methods on the ModelNet40 [47], ShapeNet [3] and ScanObjectNN (hardest perturbed variant (PB_T50_RS)) [36] for point cloud classification, where the instance accuracy is used as the evaluation metric. Specifically, Random and FPS are task-agnostic non-learnable methods, and we directly sample a specific number of points by using the corresponding methods. The baseline approaches S-Net [7], SampleNet [15] are task-specific learnable methods, and we first train the sampling network and use the trained network to downsample the input point cloud to a predefined number of points. However, we can remove the task-specific loss in these learnable sampling approaches to achieve task-agnostic point cloud sampling, which are denoted by S-Net* and SampleNet*, respectively. In contrast, LTA-PCS is a learnable approach without task-specific loss in training. Therefore, it is a task-agnostic point cloud sampling method. In Table 2, we observe that our LTA-PCS achieves better results than other methods. Remarkably, we outperform the state-of-the-art SampleNet* under the task-agnostic setting by \\(2.1\\%\\) when sampling to 256 points on ModelNet40. Moreover, LTA-PCS even performs better than the task-specific learnable point cloud sampling methods S-Net and SampleNet. One possible explanation is that S-Net and SampleNet use pretrained 3D feature extractors (e.g., PointNet [27]) to calculate the task loss, and their results are limited by the capacity of 3D feature extractors." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.707, + 0.618 + ], + "angle": 0, + "content": "4.3. Point Cloud Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.894, + 0.779 + ], + "angle": 0, + "content": "We also follow S-Net [7] to report the results of the point cloud retrieval task. For point cloud retrieval, we use ModelNet40 to evaluate our LTA-PCS framework, where the mean average precision (mAP) evaluation metric is used. In Table 3, our LTA-PCS achieves promising performance. Specifically, we first downsample the point cloud to the corresponding number of points, and then extract the features by PointNet. Following [7], the results are computed by \\(L_{2}\\) distance of the shape descriptor, which is the feature of the penultimate layer." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.724, + 0.803 + ], + "angle": 0, + "content": "4.4. Point Cloud Completion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We follow PCN [54] to reconstruct the simplified point clouds into the original dense point clouds with 2,048 points on ShapeNet Core55, and we provide the \\( L_{1} \\) Chamfer Distance (CD) results of different methods in Fig. 3. Note that for the task-specific methods, we need to additionally train the sampling network when training the PCN method. From" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28040" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.12 + ], + "angle": 0, + "content": "Table 2. Comparison of our LTA-PCS with other methods under the point cloud classification task. * indicates we remove the task-specific loss in training for task-agnostic setting. \"Oracle\" denotes the accuracy using original dense point clouds." + }, + { + "type": "table", + "bbox": [ + 0.165, + 0.129, + 0.805, + 0.297 + ], + "angle": 0, + "content": "
DatasetsModelNet (Oracle: 90.3)ShapeNet (Oracle: 85.0)
#Sampled points6412825651264128256512
Task-AgnosticRandom81.586.286.688.279.380.981.983.2
FPS86.187.988.188.380.581.582.483.3
S-Net* [7]86.187.888.288.580.681.382.283.0
SampleNet* [15]86.688.387.988.480.381.682.482.9
LTA-PCS88.289.190.090.382.083.183.584.4
Task-SpecificS-Net [7]87.788.188.489.080.982.382.783.6
SampleNet [15]87.988.388.088.781.082.582.783.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.361 + ], + "angle": 0, + "content": "Table 3. Comparison of our LTA-PCS with other methods under the point cloud retrieval task. * indicates we remove the task-specific loss in training for task-agnostic setting." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.371, + 0.437, + 0.558 + ], + "angle": 0, + "content": "
#Sampled points64128256512
Task-Agnostic
Random66.169.472.573.3
FPS69.172.573.074.2
S-Net* [7]69.471.773.674.1
SampleNet* [15]70.472.372.674.7
LTA-PCS71.473.274.877.2
Task-Specific
S-Net [7]70.572.774.076.2
SampleNet [15]70.972.573.676.0
" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.571, + 0.429, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.757, + 0.471, + 0.787 + ], + "angle": 0, + "content": "Figure 3. Performance on the point cloud completion task. Lower \\( L_{1} \\) Chamber Distance (\\( L_{1} \\) CD) denotes better performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.884 + ], + "angle": 0, + "content": "Fig. 3, we observe that the quantitative results of our LTA-PCS are better than other methods under different numbers of sampled points. The results show our LTA-PCS is effective for different downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Visualization. We provide the visualization results of dif" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.318, + 0.62, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.318, + 0.731, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.319, + 0.84, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.468, + 0.766, + 0.478 + ], + "angle": 0, + "content": "(a) Simplified Point Clouds" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.481, + 0.614, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.622, + 0.598, + 0.631 + ], + "angle": 0, + "content": "FPS" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.484, + 0.729, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.623, + 0.715, + 0.631 + ], + "angle": 0, + "content": "S-Net" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.481, + 0.842, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.623, + 0.835, + 0.631 + ], + "angle": 0, + "content": "LTA-PCS" + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.631, + 0.775, + 0.641 + ], + "angle": 0, + "content": "(b) Completed Point Clouds" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.653, + 0.894, + 0.696 + ], + "angle": 0, + "content": "Figure 4. (a). Simplified point clouds. (The sampled points are enlarged for better visualization.) (b). Completed point clouds by using the corresponding simplified point clouds as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.894, + 0.81 + ], + "angle": 0, + "content": "ferent methods in Fig. 4. Specifically, in Fig. 4(a), the simplified point clouds are produced by FPS, S-Net and LTA-PCS, where the number of points in each simplified point cloud are 64, and in Fig. 4(b), we observe that the reconstruction quality of LTA-PCS is better than other methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.654, + 0.833 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Effect of different losses. In Table 4 we provide three alternative variants of our LTA-PCS: LTA-PCS (w/o inter & w/o intra), LTA-PCS (w/o inter), LTA-PCS (w/o intra). Specifically, for LTA-PCS (w/o inter & w/o intra), we remove the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28041" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.09, + 0.402, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation on using different losses." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.115, + 0.44, + 0.208 + ], + "angle": 0, + "content": "
#Sampled points128256
LTA-PCS89.190.0
LTA-PCS (w/o inter & w/o intra)87.888.2
LTA-PCS (w/o inter)88.689.4
LTA-PCS (w/o intra)88.389.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.224, + 0.435, + 0.239 + ], + "angle": 0, + "content": "Table 5. Ablation on using different feature extractors." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.25, + 0.442, + 0.328 + ], + "angle": 0, + "content": "
#Sampled points128
LTA-PCS (PointNet)88.0
LTA-PCS (R50 trained on ImageNet)88.3
LTA-PCS (R50 of CLIP vision encoder)89.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.345, + 0.436, + 0.358 + ], + "angle": 0, + "content": "Table 6. Ablation on using different numbers of views." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.37, + 0.388, + 0.417 + ], + "angle": 0, + "content": "
# views2610
LTA-PCS88.288.989.1
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.47, + 0.581 + ], + "angle": 0, + "content": "semantic losses (i.e., both inter-view and intra-view losses), which is the same as the S-Net [7] without using the task loss. For LTA-PCS (w/o inter), we only use the intra-view loss without using the inter-view loss to train our LTA-PCS framework. For LTA-LTA-PCS (w/o intra), we only use the inter-view loss without using the intra-view loss to train the LTA-PCS framework. In Table 4, our LTA-PCS is better than these three alternative variants, which shows that it is beneficial to utilize both intra-view and inter-view losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effect of 2D feature extractor. In Table 5, we compare the results of our LTA-PCS by using different pretrained vision backbones [6, 12] to show the effect of 2D feature extractor. Specifically, for LTA-PCS (PointNet), we directly use the PointNet backbone to extract the features of the original dense point cloud and simplified point cloud and calculate the MSE distance of corresponding features to preserve semantic consistency. For LTA-PCS (R50 trained on ImageNet), we use the ResNet-50 [12] model trained on ImageNet as the feature extractor. For LTA-PCS (R50 of CLIP vision encoder) and LTA-PCS (VIT-B/16 of CLIP vision encoder), we use the pretrained CLIP vision encoder with different backbones. In Table 5, first, the results of LTA-PCS (PointNet) are lower than other methods a lot, which indicates that the 3D feature extractor (i.e., PointNet) cannot provide high-quality semantic supervision to train the sampling network. Second, we observe that the results of LTA-PCS (R50 of CLIP vision encoder) are better than the results of LTA-PCS (R50 trained on ImageNet) by a large margin, which shows the capacity of the feature extractor is critical when training the sampling network. Therefore," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "we choose the pre-trained 2D feature extractor to maintain semantic consistency when training the sampling network instead of using the pre-trained 3D feature extractor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Effect of the number of views. In Table 6, we report the performance results of using 128 sampled points as input to analyze the effect of the number of projected views, where we report the results on the ModelNet40 dataset based on the PointNet backbone. When the number of views increases from 2 to 10, LTA-PCS can achieve better performance results, which indicates that it is effective to use more views for maintaining the semantic consistency between the original and simplified point clouds. Therefore, we set the projected number of views as 10 to achieve better performance with acceptable memory usage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.313, + 0.667, + 0.329 + ], + "angle": 0, + "content": "4.6. Further Analysis" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.35, + 0.892, + 0.379 + ], + "angle": 0, + "content": "Table 7. Inference time and GPU memory usage of our LTA-PCS with different sampled points." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.389, + 0.889, + 0.489 + ], + "angle": 0, + "content": "
#Sampled pointsInference time (ms)GPU usage (MB)
640.28578
1280.37616
2560.58724
5121.05940
20481.772326
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Comparison on inference time and GPU memory usage. In Table 7, we take the ModelNet40 dataset as an example and compare the inference time and GPU memory usage under different sampled points when using PointNet as the backbone network. As shown in Table 7, we observe that using fewer points can reduce inference time and GPU memory usage greatly, which further demonstrates the advantages of point cloud sampling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.64, + 0.619, + 0.655 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.83 + ], + "angle": 0, + "content": "In this work, we propose the Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework for task-agnostic point cloud sampling task, which aims to preserve the geometric and semantic consistency in point cloud sampling. Specifically, we first propose to use the pre-trained 2D feature extractor for training the sampling network, and use semantic loss including both inter-view and intra-view losses. Comprehensive experimental results on multiple downstream tasks show the effectiveness of LTA-PCS. In the future, we will continue to explore more tasks based on our LTA-PCS method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.892, + 0.887 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by National Science and Technology Major Project (2022ZD0116405) and National Natural Science Foundation of China (No. 61932002, No. 62306025, No. 92367204)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28042" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.355 + ], + "angle": 0, + "content": "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[4] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. arXiv preprint arXiv:1904.00069, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Ta-Ying Cheng, Qingyong Hu, Qian Xie, Niki Trigoni, and Andrew Markham. Meta-sampler: Almost-universal yet task-oriented sampling for point clouds. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 694-710, Berlin, Heidelberg, 2022. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.569 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.47, + 0.625 + ], + "angle": 0, + "content": "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.47, + 0.684 + ], + "angle": 0, + "content": "[8] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[9] Jinyang Guo, Jiaheng Liu, and Dong Xu. Jointpruning: Pruning networks along multiple dimensions for efficient point cloud processing. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[10] Jinyang Guo, Jiaheng Liu, and Dong Xu. 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8717-8729, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[11] Jinyang Guo, Dong Xu, and Wanli Ouyang. Multidimensional pruning and its extension: A unified framework for model compression. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[13] Asako Kanezaki, Yasuyuki Matsushita, and Yoshifumi Nishida. Rotationnet: Joint object categorization and pose estimation using multiviews from unsupervised viewpoints. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[15] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[16] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeruIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[17] Jiaheng Liu and Dong Xu. Geometrhythm-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31 (12):4711-4721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[18] Jiaheng Liu, Yudong Wu, Yichao Wu, Chuming Li, Xiaolin Hu, Ding Liang, and Mengyu Wang. Dam: discrepancy alignment metric for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3814-3823, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[19] Jiaheng Liu, Jinyang Guo, and Dong Xu. Apsnet: Towards adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[20] Jiaheng Liu, Jinyang Guo, and Dong Xu. Geometriontransformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[21] Jiaheng Liu, Tong He, Honghui Yang, Rui Su, Jiayi Tian, Junran Wu, Hongcheng Guo, Ke Xu, and Wanli Ouyang. 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[22] Jiaheng Liu, Haoyu Qin, Yichao Wu, Jinyang Guo, Ding Liang, and Ke Xu. Coupleface: Relation matters for face recognition distillation. In European Conference on Computer Vision, pages 683-700. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[23] Jiaheng Liu, Zhipeng Yu, Haoyu Qin, Yichao Wu, Ding Liang, Gangming Zhao, and Ke Xu. Oneface: one threshold for all. In European Conference on Computer Vision, pages 545-561. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[26] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In ICCV, pages 5007-5016, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28043" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.164 + ], + "angle": 0, + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[29] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.263 + ], + "angle": 0, + "content": "[30] Yue Qian, Junhui Hou, Qijian Zhang, Yiming Zeng, Sam Kwong, and Ying He. Mops-net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. arXiv preprint arXiv:2005.00383, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.471, + 0.419 + ], + "angle": 0, + "content": "[32] Muhammad Sarmad, Hyunjoo Jenny Lee, and Young Min Kim. Rl-gan-net: A reinforcement learning agent controlled gan network for real-time point cloud shape completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.476 + ], + "angle": 0, + "content": "[33] Vinit Sarode, Xueqian Li, Hunter Goforth, Yasuhiro Aoki, Rangaprasad Arun Srivatsan, Simon Lucey, and Howie Choset. Pcrnet: Point cloud registration network using point-net encoding. arXiv preprint arXiv:1908.07906, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.518 + ], + "angle": 0, + "content": "[34] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.471, + 0.576 + ], + "angle": 0, + "content": "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.471, + 0.646 + ], + "angle": 0, + "content": "[36] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.471, + 0.717 + ], + "angle": 0, + "content": "[37] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[38] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 8565-8574. Computer Vision Foundation / IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[39] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "IEEE conference on computer vision and pattern recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[42] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. Lightn: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. arXiv preprint arXiv:2202.06263, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[43] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[44] C. Wen, B. Yu, and D. Tao. Learnable skeleton-aware 3d point cloud sampling. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[45] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[46] Yue Wu, Xidao Hu, Yue Zhang, Maoguo Gong, Wenping Ma, and Qiguang Miao. Sacf-net: Skip-attention based correspondence filtering network for point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3585-3595, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[48] Ge Yang, Chao Zhang, Ling Gao, Yufei Guo, and Jinyang Guo. Domain adaptive channel pruning. *Electronics*, 13(5), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[49] Honghui Yang, Tong He, Jiaheng Liu, Hua Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wanli Ouyang. Gd-mae: generative decoder for mae pre-training on lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9403-9414, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[50] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[51] Yang Ye, Xiulong Yang, and Shihao Ji. Apsnet: Attention based point cloud sampling. arXiv preprint arXiv:2210.05638, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[53] Zhipeng Yu, Jiaheng Liu, Haoyu Qin, Yichao Wu, Kun Hu, Jiayi Tian, and Ding Liang. Icd-face: Intra-class compactness distillation for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21042-21052, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28044" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV). IEEE, 2018. 2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[55] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[56] Lichen Zhao, Jinyang Guo, Dong Xu, and Lu Sheng. Transformer3d-det: Improving 3d object detection by vote refinement. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.469, + 0.304 + ], + "angle": 0, + "content": "[57] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[58] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28045" + } + ] +] \ No newline at end of file diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_origin.pdf b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d5e7e8f9ec84e2bbc2a46fa88c69d715fa1ad3d2 --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/c6e80379-91d5-42d2-871e-0e2f7374da5c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:630660cbd158c2972a46399b4832f0d991a8f7d368e7fe2490a8d0916432536b +size 1004734 diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/full.md b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8cbbc90129d244ba4f3f01e1c398b634d2c9217 --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/full.md @@ -0,0 +1,349 @@ +# LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling + +Jiaheng Liu $^{1*}$ , Jianhao Li $^{1*}$ , Kaisiyuan Wang $^{3}$ , Hongcheng Guo $^{1}$ , Jian Yang $^{1}$ , Junran Peng $^{2}$ , Ke Xu $^{1}$ , Xianglong Liu $^{1}$ , Jinyang Guo $^{1\dagger}$ + +$^{1}$ Beihang University $^{2}$ Institute of Automation, Chinese Academy of Sciences $^{3}$ The University of Sydney + +# Abstract + +Recently, many approaches directly operate on point clouds for different tasks. These approaches become more computation and storage demanding when point cloud size is large. To reduce the required computation and storage, one possible solution is to sample the point cloud. In this paper, we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework. Existing task-agnostic point cloud sampling strategy (e.g., FPS) does not consider semantic information of point clouds, causing degraded performance on downstream tasks. While learning-based point cloud sampling methods consider semantic information, they are task-specific and require task-oriented ground-truth annotations. So they cannot generalize well on different downstream tasks. Our LTA-PCS achieves task-agnostic point cloud sampling without requiring task-oriented labels, in which both the geometric and semantic information of points is considered in sampling. Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS. + +# 1. Introduction + +3D point cloud has been widely used in many areas like autonomous driving and robotics. However, the point clouds collected by LiDAR or other sensors contain millions or even billions of points, making it challenging to process, store, and transmit them efficiently. Therefore, to improve computational efficiency and reduce memory costs, it is critical to develop point cloud sampling methods to reduce the number of points while preserving the essential features of the original point clouds. + +Among point cloud sampling methods, the most widely used one is the farthest point sampling (FPS) strategy in Fig. 1(a). It samples the critical points based on their 3D coordinates iteratively and selects a group of points that + +![](images/a09d4f79be3b936d458399d759655c882de173dd15cd354a980314777899c59c.jpg) + +![](images/639e8aff6fa3336d3c4766d69905bb6f61f3e8403cc7098d05fcd14f02faa9b7.jpg) +(a) FPS +(b) Learnable Task-specific methods + +![](images/91731ea9eb165e108319db947258a4e810b7899e89e1ac1c7044c5d97d857fe8.jpg) +(c) Learnable Task-agnostic method (Ours) +Figure 1. Comparison of FPS, learnable task-specific methods, and our LTA-PCS. FPS samples points using 3D coordinates, which ignores semantic meaning. Learnable Task-specific methods need different sampling networks for different tasks. Our LTA-PCS is task-agnostic, where one trained sampling network can be used for different tasks. Note that the pretrained 2D feature extractor is frozen in training, and not used at inference. + +are farthest apart from each other. Although FPS is task-agnostic, it does not utilize the semantic information of point clouds as the sampling process is solely based on the point coordinates, causing degraded performance on downstream tasks. Recently, in Fig. 1 (b), learning-based point cloud sampling methods [7, 15] are proposed to sample the + +points with more semantic meanings for better downstream task performance. They additionally introduce task-specific losses when learning the network for point cloud sampling. However, these methods are task-specific and require task-oriented labels. In real-world applications, the downstream tasks are usually not specified. But we need to sample fewer points in the collected point cloud by LiDAR or other sensors to save the storage. Moreover, even downstream tasks are given, the existing learning-based methods need to sample the points again based on the new downstream task for better performance, which is cumbersome in practice. Besides, the existing learning-based methods use the pretrained 3D feature extractor (e.g., PointNet [27]) to calculate the task loss for training the sampling network, while the capacity of these 3D feature extractors is limited by both model size and the scale of 3D point cloud datasets, which becomes the bottleneck for training the sampling network with semantically meaningful sampling ability. Finally, the task-specific loss makes them prone to overfit on specific datasets and hard to generalize on different datasets, and the requirement of labels increases the training cost. Although we can remove the task-specific loss in these learning-based methods for task-agnostic point cloud sampling, the sampled points still lack semantic meaning in this case. + +So the problem is: How to design a learnable task-agnostic label-free point cloud sampling strategy with the consideration of semantic information? To solve this problem, in Fig. 1(c), we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework, in which the semantic meaning of points is considered in the sampling process. It is non-trivial to construct such a framework. Without the specific downstream tasks, the designed sampling loss for learning-based sampling methods is based on the coordinates, which lack semantic meaning. + +In LTA-PCS, we observe 2D feature extractors [31] are effective tools to extract task-agnostic but semantically meaningful features and propose to introduce the fixed 2D feature extractors pre-trained on large-scale datasets to help the training of our LTA-PCS for sampling more semantic meaningful points, where the specific tasks and the corresponding labels are both not required. Specifically, to solve the heterogeneous problem of image and point cloud data, we first project 3D point clouds to multi-view depth maps, which can be readily used by 2D feature extractors to extract semantic meaningful features. Then, we calculate the loss based on both the 3D coordinates and the extracted features and train the sampling network, which can enforce it to recognize meaningful points in terms of both geometric and semantic aspects. After training, we directly use the trained sampling network to generate the simplified point clouds for different downstream tasks, in which an inference alignment strategy is used to ensure the simplified point cloud is a subset of the original dense point cloud. + +In the aforementioned training paradigm, another issue is that the projection from point clouds to multi-view depth maps will inevitably cause information loss, causing inaccurate learning of semantic meaningful points. To this end, we propose a new loss function called semantic loss, which includes both intra- and inter-view losses. Specifically, the intra-view loss minimizes the information loss between the dense point cloud and simplified point cloud under the same views, while the inter-view loss is calculated based on the relationship between different views. By introducing the semantic loss, we can consider both the information under the same and cross views, which can help the sampling network to generate more semantic meaningful points. + +Moreover, our LTA-PCS framework is also general. Although it is designed for task-agnostic point cloud sampling, we can further combine it with the task-specific loss for task-specific point cloud sampling to achieve better performance when the downstream task is given. + +The contributions of LTA-PCS are shown as follows: + +- To the best of our knowledge, we propose the first learnable task-agnostic point cloud sampling framework called LTA-PCS, in which both the geometric and semantic information are considered and the task-oriented labels are not required. +- We introduce the pre-trained 2D feature extractor to compute the inter- and intra-view learning objective functions to help train the sampling network in our LTA-PCS framework, which can enforce the sampling network to sample more semantically meaningful points. +- Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS. + +# 2. Related Work + +Deep learning on point clouds. Recently, deep learning methods [18, 22, 23, 37-39, 48, 53] attract many attentions, in which a large number of methods using point clouds are proposed for 3D scene understanding [9-11, 17, 19-21, 46, 49, 50]. For example, PointNet [27] uses multilayer perceptron to process points for different tasks. PointNet++ [28] proposes a hierarchical structure to extract features. Li et al. [16] proposed PointCNN to perform convolution operations on point clouds. Wu et al. [45] proposed PointConv to introduce point density when perform convolution on point clouds. There are also many methods on point cloud segmentation [40, 41, 43] and object detection [29, 34, 56]. In addition, approaches are also proposed to focus on point cloud completion [4, 32, 54], point cloud registration [2, 33]. On the other hand, some methods project point clouds into volumetric [24, 57] or multiview [13, 31, 35, 52] data forms and process them using neural networks. For example, Zhang et al. [55] project point clouds into multi-view depth images and process them using pretrained CLIP model [31]. Goyal et al. [8] propose + +Table 1. Comparison between our LTA-PCS framework with existing point cloud sampling methods. + +
MethodsSemantic informationTask-agnostic
FPS×✓
Learning-based methods [5, 7, 15]✓×
LTA-PCS✓✓
+ +SimpleView projection methods for point cloud classification. Zhu et al. [58] utilize pretrained CLIP model for 3D open-world learning. Although these methods use CNNs or CLIP models for learning point cloud features, they do not focus on point cloud sampling. Besides, we utilize the 2D feature extractor in a different perspective, in which the 2D feature extractors are used to facilitate the training of our LTA-PCS for more semantically meaningful sampling. + +Point cloud sampling methods. Recently, many point cloud sampling methods were proposed. In Table 1, for the task-agnostic point cloud sampling, the FPS method samples the point that is farthest from the previously sampled ones iteratively. However, this method only uses point coordinates, which does not consider the semantic information of points and leads to a sub-optimal solution when using the sampled points for downstream tasks. Besides, many learning-based approaches were proposed [5, 30, 44]. For example, Dovrat et al. [7] proposed S-NET to use networks for point cloud sampling, where a sampling regularization loss is used to train this S-NET. Lang et al. [15] proposed SampleNet to use differentiable relaxation for point cloud sampling. Qian et al. [30] proposed a task-oriented point cloud downsampling approach. Attention operation is also used for point cloud sampling [42, 51]. Although these methods use networks to sample semantic meaningful points from a dense point cloud, they are task-oriented. Therefore, these approaches need specific downstream tasks to train the sampling network, which is often infeasible in real-world applications. In contrast, our LTA-PCS framework is task-agnostic, and the comparison between our LTA-PCS and the existing methods is shown in Table 1. + +# 3. Methodology + +# 3.1. Problem Statement + +Given a dense point cloud with $n$ points $P = \{p_i \in \mathbb{R}^3, i = 1, \dots, n\}$ and a target size $m \leq n$ , our goal in this work is to learn a sampling network that can sample $m$ points from $P$ and generate a simplified point cloud that can best represent the dense point cloud with task-agnostic semantic meaning as follows: + +$$ +T (P) \subset P, | T (P) | = m \leq n, \tag {1} +$$ + +where $T(\cdot)$ is the operation of the sampling network, and $T(P)$ is the simplified point cloud. In our LTA-PCS, the learning process of the sampling network $T(\cdot)$ is not related to any downstream tasks. In other words, we aim to find the best sampling network to generate the simplified point cloud that can perform well on the downstream tasks. Traditional point cloud sampling methods (e.g., FPS) only sample the points based on 3D coordinates, which lacks semantic information. To introduce semantic information in sampling, learning-based sampling methods like S-NET [7] and SampleNet [15] use the task-specific loss to train the sampling network. However, when the task is not given, it is still an open problem on how to introduce such a semantic-aware loss for guiding the sampling network to sample semantically meaningful points. Therefore, we propose to use effective and well-performed pretrained 2D feature extractors to facilitate the training of the sampling network for LTA-PCS in a task-agnostic way. + +# 3.2. Overview + +Fig. 2 shows the overview when training our LTA-PCS framework. Given a dense point cloud, we use the sampling network in LTA-PCS to sample points from this point cloud and generate the simplified point cloud. To utilize the pretrained 2D feature extractors, we first project both the dense and simplified point clouds to multi-view depth maps. Then, we use the pretrained 2D feature extractor to extract semantically meaningful features based on multi-view depth maps. In our implementation, we choose the vision branch of the CLIP model [31] as our pretrained 2D feature extractor, and the parameters of the 2D feature extractor are frozen in the training process. After that, we calculate the proposed semantic loss including both inter- and intraview losses based on the output feature from 2D feature extractors, which introduces the semantic-level supervision to help the sampling network in our LTA-PCS framework and sample more semantically meaningful points. Besides, to fully utilize the 3D coordinate information of point clouds, we also use geometric loss [7] to introduce geometric-level supervision when training our LTA-PCS framework in the training process. We update the sampling network based on both geometric loss and semantic loss. + +Sampling network. Our sampling model follows the architecture of [7, 27]. The input points are processed by a series of $1 \times 1$ convolution layers, which produce a feature vector for each point. Then, a symmetric feature-wise max pooling operation is applied to obtain a global feature vector. Finally, we use several fully-connected layers. The output of the last layer is the set of generated points. + +Projection. Directly using pretrained 2D feature extractor is infeasible because of the heterogeneous problem between 2D images and 3D point clouds. To solve this problem, we need to project point clouds to multi-view + +![](images/57606d2da4553e99f02718743d3b3534e58960f02d043d4dbe5e129bfb8e0c7d.jpg) +Figure 2. Training procedure of our LTA-PCS framework. Given a dense point cloud, we first use the initialized sampling network to sample points from it and generate a simplified point cloud. Then, we project both the dense and simplified point clouds to multi-view depth maps, which will be used for feature extraction by a frozen pretrained 2D feature extractor. We calculate the semantic loss including intra-view and inter-view losses using the extracted features. Geometric loss is also used in training. At the inference stage, we directly use the trained sampling network to sample points from dense point clouds. + +depth maps. Specifically, inspired by [55], we take the bottom view as an example, the point with the coordinate of $(x,y,z)$ will be projected to $(\lceil x / z\rceil ,\lceil y / z\rceil)$ . Also, as the vision branch of CLIP model aims to process images with three channels, we copy the projected depth maps three times and form the input of the CLIP model, which ensures there is no shape mismatch when using the CLIP model. + +Inference. At inference, we use the sampling network in Fig. 2 to sample points from the dense point cloud, which can be used for multiple downstream tasks. + +# 3.3. Loss Function + +Overall loss function. Formally, the overall loss function of our LTA-PCS can be written as follows: + +$$ +L = L _ {g e o} + \alpha L _ {s e m}, \tag {2} +$$ + +where $L_{geo}$ is the geometric loss, and $L_{sem}$ is the newly proposed semantic loss, which includes both the intra- and inter-view losses. $\alpha$ is the coefficient to balance these two terms. In LTA-PCS, the geometric and semantic losses aim to preserve the geometric and semantic consistency between the original point cloud and the simplified point cloud. + +Geometric loss. We denote $Q$ as the simplified point cloud, i.e., $Q = T(P)$ . Inspired by [7], the geometric loss + +is defined as follows: + +$$ +L _ {g e o} = L _ {a} (Q, P) + \beta L _ {w} (Q, P) + \gamma L _ {s} (Q, P), +$$ + +$$ +\text {w h e r e} L _ {a} (Q, P) = \frac {1}{| Q |} \sum_ {q \in Q} \min _ {p \in P} \| q - p \| _ {2} ^ {2}, +$$ + +$$ +L _ {w} (Q, P) = \max _ {q \in Q} \min _ {p \in P} \| q - p \| _ {2} ^ {2}, \tag {3} +$$ + +$$ +L _ {s} (Q, P) = \frac {1}{| P |} \sum_ {p \in P} m i n _ {q \in Q} | | p - q | | _ {2} ^ {2}. +$$ + +Here, $L_{a}(Q,P)$ is to minimize the average distance between $Q$ and $P$ . $L_{w}(Q,P)$ is to minimize the maximum distance between $Q$ and $P$ . $L_{s}(Q,P)$ ensures the points in $Q$ evenly spread in the 3D coordinate space. Following [7], $\beta$ and $\gamma$ are the loss weights. + +Semantic loss. To enforce the sampling network to sample more semantically meaningful points, we also introduce the newly proposed semantic loss $L_{sem}$ in the training process. Formally, let us suppose the dense point cloud as $P$ and the simplified point cloud generated by the initialized sampling network as $Q$ . The semantic loss consists of intraview and inter-view loss, which can be written as follows: + +$$ +L _ {s e m} = L _ {i n t r a} + \lambda L _ {i n t e r}. \tag {4} +$$ + +The intra-view loss aims to minimize the semantic information loss under the same view as follows: + +$$ +L _ {i n t r a} = \sum_ {k = 1} ^ {K} M S E [ E (J _ {k} (P)), E (J _ {k} (Q)) ], \tag {5} +$$ + +# Algorithm 1 Training procedure of our LTA-PCS. + +Input: Dense point cloud $P$ ; Simplified point cloud $Q$ ; Randomly initialized sampling network $T$ ; Number of projected views $K$ ; Pretrained 2D feature extractor $E$ ; + +1: for each iteration in the training process do +2: // Calculate geometric loss $L_{geo}$ +3: Use $T$ to generate the simplified point cloud $Q$ ; +4: Calculate geometric loss $L_{geo}$ based on dense point cloud $P$ and the simplified point cloud $Q$ using Eq. (3); +5: // Calculate semantic loss $L_{sem}$ +6: for view $k$ in $K$ do +7: Project dense point cloud $P$ and simplified point cloud $Q$ to the $k$ -th view depth map $J_{k}(P)$ and $J_{k}(Q)$ , respectively; +8: Extract the features of $J_{k}(P)$ and $J_{k}(Q)$ using the 2D feature extractor $E$ , respectively; +9: end for +10: Calculate intra- and intra-view losses by Eq. (5) and Eq. (6); +11: Calculate semantic loss $L_{sem}$ based on Eq. (4); +12: Calculate overall object function $L$ using Eq. (2); +13: Back propagate the gradients and update the parameters of the sampling network $T$ ; +14: end for + +Output: The optimized sampling network $T$ + +where $J_{k}(\cdot)$ denotes the projection operation for the $k$ -th view, and $E(\cdot)$ is the 2D feature extractor for extracting semantic features. $MSE(\cdot)$ is the mean square error, and $K$ is the total number of projected views. + +As the projection operation inevitably causes information loss and the relationship of different views can also provide essential information [25, 26], we calculate the interview loss for compensation as follows: + +$$ +L _ {i n t e r} = M S E \left(\boldsymbol {\Phi} \boldsymbol {\Phi} ^ {T}, \boldsymbol {\Psi} \boldsymbol {\Psi} ^ {T}\right), +$$ + +$$ +\text {w h e r e} \Phi = \underset {K} {\operatorname {C o n c a t}} \left[ \operatorname {F l a t} \left(E \left(J _ {k} (P)\right)\right) \right], \tag {6} +$$ + +$$ +\Psi = \underset {K} {C o n c a t} \left[ F l a t (E (J _ {k} (Q))) \right]. +$$ + +Here, $Flat(\cdot)$ is the flatten operation. $Concat_{K}$ concatenate the vectors along different views. By using the interview loss, we obtain the relation information in different views from dense point cloud and enforce the simplified point cloud to mimic this information. + +Based on the intra-view loss $L_{intra}$ , inter-view loss $L_{inter}$ , and the geometric loss $L_{geo}$ , we obtain the loss function $L$ in Eq. (2) to train the sampling network $T$ , which is used to generate the simplified point cloud at inference. The training procedure is also illustrated in Alg. 1. + +# 3.4. Inference + +At inference, we use the trained sampling network $T$ to generate the coordinates of the sampled points. However, as we use several fully-connected layers to directly predict the coordinates of the sampled points at the end of the sampling network, we cannot ensure the simplified point cloud $Q$ generated by the sampling network to be a subset of dense point cloud $P$ . Thus, following S-Net [7], we can obtain the points in $Q$ to its nearest neighbors in dense point cloud $P$ . Specifically, for each point $q \in \mathbb{R}^3$ in $Q$ , the matched point $q^*$ is written as follows: + +$$ +q ^ {*} = \underset {p \in P} {\operatorname {a r g m i n}} | | q - p | | _ {2} ^ {2}. \tag {7} +$$ + +After the matching process for $Q$ with $m$ points, we can obtain the simplified point cloud $Q^{*} = \{q_{i}^{*}\}_{i = 1}^{m}$ , which is a subset of $P$ . Note that the matching process is very fast. + +# 3.5. Downstream Tasks Usage + +After training the sampling network in our LTA-PCS, we can use it to effectively generate the simplified point cloud with the information from both 3D coordinates and the semantic meanings. As the downstream tasks are not decided at this stage, we store these simplified point clouds as a new point cloud dataset to simulate this scenario. When the downstream tasks are given, we use the simplified point clouds to train different backbone network architectures (e.g., PointNet [27]) for different downstream tasks as before, and the backbone network is initialized from scratch. + +# 3.6. Discussion + +Necessity of Learnable Task-Agnostic Point Cloud Sampling. We would emphasize that LTA-PCS tries to solve the point cloud sampling under an important scenario where the collected point clouds are given but the downstream tasks are not given. In real-world applications, this scenario often occurs. We often use LiDAR to collect point clouds and store them. When specific tasks are given, we use the stored point cloud data for downstream tasks. Thus, sampling semantically meaningful points from dense collected point clouds becomes critical for storage saving. Existing learning-based methods [15] use task-specific loss to introduce semantic information in sampling, which is infeasible in this scenario as the downstream tasks and the corresponding labels are required. Although we can remove task-specific loss in these learning-based methods for task-agnostic point cloud sampling, semantic information will not be considered in this case. Moreover, the task-specific loss will degrade the generalization ability of these methods on different datasets. LTA-PCS targets at solving the point cloud sampling under a new setting, which cannot be achieved by existing learning-based approaches. + +Discussion on the reason that our task-agnostic point cloud sampling method (i.e., LTA-PCS) yields better results relative to task-specific methods (e.g., S-Net, SampleNet). We hypothesize this is because we introduce the 2D pretrained model in the training process. Therefore, the knowledge in 2D pretrained model can significantly improve the performance of the sampling network. In contrast, due to limited network capacity or dataset size, there is no general and well-performed 3D pretrained model when compared with 2D pretrained models. Therefore, although existing state-of-the-art point cloud sampling methods are task-intensive, they cannot well utilize the prior knowledge in 2D pretrained model, resulting worse performance than our task-agnostic method LTA-PCS. Our LTA-PCS framework can be also combined with task-specific loss, and the results show the performance can be further boosted. + +Discussion on the necessity of 2D feature extractor. We choose the pretrained 2D feature extractor to introduce semantic information when training our LTA-PCS framework as there is no existing well-performed pretrained 3D feature extractor on large-scale datasets in the literature. Current 3D feature extractors like PointNet [27] or PointNet++ [28] are trained based on specific tasks and datasets, which cannot be used in the task-agnostic setting. We hypothesize that this is because of the lack of large-scale high-quality 3D point cloud datasets. Therefore, although we encounter information loss in the projection process, we choose a 2D feature extractor to introduce semantic information when constructing our LTA-PCS. + +# 4. Experiments + +# 4.1. Implementation Details + +We follow [7] to utilize a set of multi-layer perceptrons (MLPs) as the sampling network, and batch normalization and ReLU layers are appended after each MLP. In training, we train the sampling network for 50 epochs with Adam optimizer [14] on both ModelNet40, ShapeNet core55 and ScanObjectNN datasets. The batch size is set as 8 and the learning rate is $5e^{-4}$ . We set the weight decay as $1e^{-4}$ . In the training process, we set the hyperparameter $\alpha$ in Eq. (2) as 1. $\beta$ and $\gamma$ in Eq. (3) are set as 1 and 1, respectively, and $\lambda$ in Eq. (4) is set as 0.01. The number of views (i.e., $K$ ) is set as 10. After training the sampling network, we use it to generate simplified point clouds. Then, we use the generated simplified point clouds to train an initialized backbone network for different downstream tasks. For the point cloud classification task, we use PointNet [27] as our backbone network and train the initialized PointNet for 200 epochs with a batch size of 12. The learning rate is set as $1e^{-3}$ with weight decay of $1e^{-4}$ . Step learning rate decay is used. The learning rate will multiply 0.7 after every 20 epochs. For the point cloud completion and registration tasks, we + +use PCN [54] and PointNetLK [1] as baseline methods, respectively, and the training epochs are set as 400 and 200, respectively. + +# 4.2. Point Cloud Classification + +In Table 2, we compare our LTA-PCS with several existing point cloud sampling methods on the ModelNet40 [47], ShapeNet [3] and ScanObjectNN (hardest perturbed variant (PB_T50_RS)) [36] for point cloud classification, where the instance accuracy is used as the evaluation metric. Specifically, Random and FPS are task-agnostic non-learnable methods, and we directly sample a specific number of points by using the corresponding methods. The baseline approaches S-Net [7], SampleNet [15] are task-specific learnable methods, and we first train the sampling network and use the trained network to downsample the input point cloud to a predefined number of points. However, we can remove the task-specific loss in these learnable sampling approaches to achieve task-agnostic point cloud sampling, which are denoted by S-Net* and SampleNet*, respectively. In contrast, LTA-PCS is a learnable approach without task-specific loss in training. Therefore, it is a task-agnostic point cloud sampling method. In Table 2, we observe that our LTA-PCS achieves better results than other methods. Remarkably, we outperform the state-of-the-art SampleNet* under the task-agnostic setting by $2.1\%$ when sampling to 256 points on ModelNet40. Moreover, LTA-PCS even performs better than the task-specific learnable point cloud sampling methods S-Net and SampleNet. One possible explanation is that S-Net and SampleNet use pretrained 3D feature extractors (e.g., PointNet [27]) to calculate the task loss, and their results are limited by the capacity of 3D feature extractors. + +# 4.3. Point Cloud Retrieval + +We also follow S-Net [7] to report the results of the point cloud retrieval task. For point cloud retrieval, we use ModelNet40 to evaluate our LTA-PCS framework, where the mean average precision (mAP) evaluation metric is used. In Table 3, our LTA-PCS achieves promising performance. Specifically, we first downsample the point cloud to the corresponding number of points, and then extract the features by PointNet. Following [7], the results are computed by $L_{2}$ distance of the shape descriptor, which is the feature of the penultimate layer. + +# 4.4. Point Cloud Completion + +We follow PCN [54] to reconstruct the simplified point clouds into the original dense point clouds with 2,048 points on ShapeNet Core55, and we provide the $L_{1}$ Chamfer Distance (CD) results of different methods in Fig. 3. Note that for the task-specific methods, we need to additionally train the sampling network when training the PCN method. From + +Table 2. Comparison of our LTA-PCS with other methods under the point cloud classification task. * indicates we remove the task-specific loss in training for task-agnostic setting. "Oracle" denotes the accuracy using original dense point clouds. + +
DatasetsModelNet (Oracle: 90.3)ShapeNet (Oracle: 85.0)
#Sampled points6412825651264128256512
Task-AgnosticRandom81.586.286.688.279.380.981.983.2
FPS86.187.988.188.380.581.582.483.3
S-Net* [7]86.187.888.288.580.681.382.283.0
SampleNet* [15]86.688.387.988.480.381.682.482.9
LTA-PCS88.289.190.090.382.083.183.584.4
Task-SpecificS-Net [7]87.788.188.489.080.982.382.783.6
SampleNet [15]87.988.388.088.781.082.582.783.4
+ +Table 3. Comparison of our LTA-PCS with other methods under the point cloud retrieval task. * indicates we remove the task-specific loss in training for task-agnostic setting. + +
#Sampled points64128256512
Task-Agnostic
Random66.169.472.573.3
FPS69.172.573.074.2
S-Net* [7]69.471.773.674.1
SampleNet* [15]70.472.372.674.7
LTA-PCS71.473.274.877.2
Task-Specific
S-Net [7]70.572.774.076.2
SampleNet [15]70.972.573.676.0
+ +![](images/43949d0a2cf7f86b16f404b23e8ed2a520b91e2f61594e1073695ce9b421a7e3.jpg) +Figure 3. Performance on the point cloud completion task. Lower $L_{1}$ Chamber Distance ( $L_{1}$ CD) denotes better performance. + +Fig. 3, we observe that the quantitative results of our LTA-PCS are better than other methods under different numbers of sampled points. The results show our LTA-PCS is effective for different downstream tasks. + +Visualization. We provide the visualization results of dif + +![](images/67249e5907f907dc8eb9a74b3df33bbd0840092a8175338d7ece78cb81ef125d.jpg) + +![](images/2b2339d6ca3e998e3c70e6b9b168daf3849aea1d03a7c0852928e966b373a531.jpg) +(a) Simplified Point Clouds + +![](images/2bc7da72e66a44494ed9853d8d4bb794fb684ec868c21581016f4e59a82be310.jpg) + +![](images/72670124e1ff0375e23e17baeb240a0ae37aa6779b6c9d7f72da1ac08c0b23e8.jpg) +FPS + +![](images/405ca913013e76355f628aba34b52ed1b8ef4943e9c2a0476fde4789cb783d98.jpg) +S-Net +(b) Completed Point Clouds + +![](images/c64edf7b5c65fc1cc71e7aa59d0e4759ee95ae4e2debde5753d3097f3528ce7e.jpg) +LTA-PCS +Figure 4. (a). Simplified point clouds. (The sampled points are enlarged for better visualization.) (b). Completed point clouds by using the corresponding simplified point clouds as input. + +ferent methods in Fig. 4. Specifically, in Fig. 4(a), the simplified point clouds are produced by FPS, S-Net and LTA-PCS, where the number of points in each simplified point cloud are 64, and in Fig. 4(b), we observe that the reconstruction quality of LTA-PCS is better than other methods. + +# 4.5. Ablation Study + +Effect of different losses. In Table 4 we provide three alternative variants of our LTA-PCS: LTA-PCS (w/o inter & w/o intra), LTA-PCS (w/o inter), LTA-PCS (w/o intra). Specifically, for LTA-PCS (w/o inter & w/o intra), we remove the + +Table 4. Ablation on using different losses. + +
#Sampled points128256
LTA-PCS89.190.0
LTA-PCS (w/o inter & w/o intra)87.888.2
LTA-PCS (w/o inter)88.689.4
LTA-PCS (w/o intra)88.389.1
+ +Table 5. Ablation on using different feature extractors. + +
#Sampled points128
LTA-PCS (PointNet)88.0
LTA-PCS (R50 trained on ImageNet)88.3
LTA-PCS (R50 of CLIP vision encoder)89.1
+ +Table 6. Ablation on using different numbers of views. + +
# views2610
LTA-PCS88.288.989.1
+ +semantic losses (i.e., both inter-view and intra-view losses), which is the same as the S-Net [7] without using the task loss. For LTA-PCS (w/o inter), we only use the intra-view loss without using the inter-view loss to train our LTA-PCS framework. For LTA-LTA-PCS (w/o intra), we only use the inter-view loss without using the intra-view loss to train the LTA-PCS framework. In Table 4, our LTA-PCS is better than these three alternative variants, which shows that it is beneficial to utilize both intra-view and inter-view losses. + +Effect of 2D feature extractor. In Table 5, we compare the results of our LTA-PCS by using different pretrained vision backbones [6, 12] to show the effect of 2D feature extractor. Specifically, for LTA-PCS (PointNet), we directly use the PointNet backbone to extract the features of the original dense point cloud and simplified point cloud and calculate the MSE distance of corresponding features to preserve semantic consistency. For LTA-PCS (R50 trained on ImageNet), we use the ResNet-50 [12] model trained on ImageNet as the feature extractor. For LTA-PCS (R50 of CLIP vision encoder) and LTA-PCS (VIT-B/16 of CLIP vision encoder), we use the pretrained CLIP vision encoder with different backbones. In Table 5, first, the results of LTA-PCS (PointNet) are lower than other methods a lot, which indicates that the 3D feature extractor (i.e., PointNet) cannot provide high-quality semantic supervision to train the sampling network. Second, we observe that the results of LTA-PCS (R50 of CLIP vision encoder) are better than the results of LTA-PCS (R50 trained on ImageNet) by a large margin, which shows the capacity of the feature extractor is critical when training the sampling network. Therefore, + +we choose the pre-trained 2D feature extractor to maintain semantic consistency when training the sampling network instead of using the pre-trained 3D feature extractor. + +Effect of the number of views. In Table 6, we report the performance results of using 128 sampled points as input to analyze the effect of the number of projected views, where we report the results on the ModelNet40 dataset based on the PointNet backbone. When the number of views increases from 2 to 10, LTA-PCS can achieve better performance results, which indicates that it is effective to use more views for maintaining the semantic consistency between the original and simplified point clouds. Therefore, we set the projected number of views as 10 to achieve better performance with acceptable memory usage. + +# 4.6. Further Analysis + +Table 7. Inference time and GPU memory usage of our LTA-PCS with different sampled points. + +
#Sampled pointsInference time (ms)GPU usage (MB)
640.28578
1280.37616
2560.58724
5121.05940
20481.772326
+ +Comparison on inference time and GPU memory usage. In Table 7, we take the ModelNet40 dataset as an example and compare the inference time and GPU memory usage under different sampled points when using PointNet as the backbone network. As shown in Table 7, we observe that using fewer points can reduce inference time and GPU memory usage greatly, which further demonstrates the advantages of point cloud sampling. + +# 5. Conclusion + +In this work, we propose the Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework for task-agnostic point cloud sampling task, which aims to preserve the geometric and semantic consistency in point cloud sampling. Specifically, we first propose to use the pre-trained 2D feature extractor for training the sampling network, and use semantic loss including both inter-view and intra-view losses. Comprehensive experimental results on multiple downstream tasks show the effectiveness of LTA-PCS. In the future, we will continue to explore more tasks based on our LTA-PCS method. + +Acknowledgement. This work is supported by National Science and Technology Major Project (2022ZD0116405) and National Natural Science Foundation of China (No. 61932002, No. 62306025, No. 92367204). + +# References + +[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6 +[2] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2 +[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 6 +[4] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. arXiv preprint arXiv:1904.00069, 2019. 2 +[5] Ta-Ying Cheng, Qingyong Hu, Qian Xie, Niki Trigoni, and Andrew Markham. Meta-sampler: Almost-universal yet task-oriented sampling for point clouds. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 694-710, Berlin, Heidelberg, 2022. Springer-Verlag. 3 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8 +[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 3, 4, 5, 6, 7, 8 +[8] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning. PMLR, 2021. 2 +[9] Jinyang Guo, Jiaheng Liu, and Dong Xu. Jointpruning: Pruning networks along multiple dimensions for efficient point cloud processing. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2 +[10] Jinyang Guo, Jiaheng Liu, and Dong Xu. 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8717-8729, 2022. +[11] Jinyang Guo, Dong Xu, and Wanli Ouyang. Multidimensional pruning and its extension: A unified framework for model compression. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8 + +[13] Asako Kanezaki, Yasuyuki Matsushita, and Yoshifumi Nishida. Rotationnet: Joint object categorization and pose estimation using multiviews from unsupervised viewpoints. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2 +[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6 +[15] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6, 7 +[16] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeruIPS, 2018. 2 +[17] Jiaheng Liu and Dong Xu. Geometrhythm-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31 (12):4711-4721, 2021. 2 +[18] Jiaheng Liu, Yudong Wu, Yichao Wu, Chuming Li, Xiaolin Hu, Ding Liang, and Mengyu Wang. Dam: discrepancy alignment metric for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3814-3823, 2021. 2 +[19] Jiaheng Liu, Jinyang Guo, and Dong Xu. Apsnet: Towards adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing, 2022. 2 +[20] Jiaheng Liu, Jinyang Guo, and Dong Xu. Geometriontransformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia, 2022. +[21] Jiaheng Liu, Tong He, Honghui Yang, Rui Su, Jiayi Tian, Junran Wu, Hongcheng Guo, Ke Xu, and Wanli Ouyang. 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375, 2022. 2 +[22] Jiaheng Liu, Haoyu Qin, Yichao Wu, Jinyang Guo, Ding Liang, and Ke Xu. Coupleface: Relation matters for face recognition distillation. In European Conference on Computer Vision, pages 683-700. Springer, 2022. 2 +[23] Jiaheng Liu, Zhipeng Yu, Haoyu Qin, Yichao Wu, Ding Liang, Gangming Zhao, and Ke Xu. Oneface: one threshold for all. In European Conference on Computer Vision, pages 545-561. Springer, 2022. 2 +[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2015. 2 +[25] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 5 +[26] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In ICCV, pages 5007-5016, 2019. 5 +[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification + +and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2, 3, 5, 6 +[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6 +[29] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 2 +[30] Yue Qian, Junhui Hou, Qijian Zhang, Yiming Zeng, Sam Kwong, and Ying He. Mops-net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. arXiv preprint arXiv:2005.00383, 2020. 3 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 3 +[32] Muhammad Sarmad, Hyunjoo Jenny Lee, and Young Min Kim. Rl-gan-net: A reinforcement learning agent controlled gan network for real-time point cloud shape completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2 +[33] Vinit Sarode, Xueqian Li, Hunter Goforth, Yasuhiro Aoki, Rangaprasad Arun Srivatsan, Simon Lucey, and Howie Choset. Pcrnet: Point cloud registration network using point-net encoding. arXiv preprint arXiv:1908.07906, 2019. 2 +[34] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 2 +[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, 2015. 2 +[36] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6 +[37] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2 +[38] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 8565-8574. Computer Vision Foundation / IEEE, 2021. +[39] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2022. 2 +[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the + +IEEE conference on computer vision and pattern recognition, 2018. 2 +[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2 +[42] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. Lightn: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. arXiv preprint arXiv:2202.06263, 2022. 3 +[43] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2 +[44] C. Wen, B. Yu, and D. Tao. Learnable skeleton-aware 3d point cloud sampling. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3 +[45] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2 +[46] Yue Wu, Xidao Hu, Yue Zhang, Maoguo Gong, Wenping Ma, and Qiguang Miao. Sacf-net: Skip-attention based correspondence filtering network for point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3585-3595, 2023. 2 +[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6 +[48] Ge Yang, Chao Zhang, Ling Gao, Yufei Guo, and Jinyang Guo. Domain adaptive channel pruning. *Electronics*, 13(5), 2024. 2 +[49] Honghui Yang, Tong He, Jiaheng Liu, Hua Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wanli Ouyang. Gd-mae: generative decoder for mae pre-training on lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9403-9414, 2023. 2 +[50] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 2 +[51] Yang Ye, Xiulong Yang, and Shihao Ji. Apsnet: Attention based point cloud sampling. arXiv preprint arXiv:2210.05638, 2022. 3 +[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2 +[53] Zhipeng Yu, Jiaheng Liu, Haoyu Qin, Yichao Wu, Kun Hu, Jiayi Tian, and Ding Liang. Icd-face: Intra-class compactness distillation for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21042-21052, 2023. 2 + +[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV). IEEE, 2018. 2,6 +[55] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4 +[56] Lichen Zhao, Jinyang Guo, Dong Xu, and Lu Sheng. Transformer3d-det: Improving 3d object detection by vote refinement. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2 +[57] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2 +[58] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 3 \ No newline at end of file diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/images.zip b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..611a94aa1dcd9ac6508bb852adfaab329453001f --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2814be19017737d2b1c068ec5e9241d9b6bb004ddbd6056411daa4a435dbbaca +size 435797 diff --git a/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/layout.json b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d48b193f18133c2e33227f8172af3252ab9ad75b --- /dev/null +++ b/2024/LTA-PCS_ Learnable Task-Agnostic Point Cloud Sampling/layout.json @@ -0,0 +1,9737 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 479, + 121 + ], + "type": "text", + "content": "LTA-PCS: Learnable Task-Agnostic Point Cloud Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "spans": [ + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": "Jiaheng Liu" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Jianhao Li" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Kaisiyuan Wang" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Hongcheng Guo" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Junran Peng" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Ke Xu" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Xianglong Liu" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "text", + "content": ", Jinyang Guo" + }, + { + "bbox": [ + 108, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "spans": [ + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "text", + "content": "Beihang University " + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "text", + "content": "Institute of Automation, Chinese Academy of Sciences " + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 108, + 171, + 485, + 200 + ], + "type": "text", + "content": "The University of Sydney" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 480 + ], + "type": "text", + "content": "Recently, many approaches directly operate on point clouds for different tasks. These approaches become more computation and storage demanding when point cloud size is large. To reduce the required computation and storage, one possible solution is to sample the point cloud. In this paper, we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework. Existing task-agnostic point cloud sampling strategy (e.g., FPS) does not consider semantic information of point clouds, causing degraded performance on downstream tasks. While learning-based point cloud sampling methods consider semantic information, they are task-specific and require task-oriented ground-truth annotations. So they cannot generalize well on different downstream tasks. Our LTA-PCS achieves task-agnostic point cloud sampling without requiring task-oriented labels, in which both the geometric and semantic information of points is considered in sampling. Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 128, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 633 + ], + "type": "text", + "content": "3D point cloud has been widely used in many areas like autonomous driving and robotics. However, the point clouds collected by LiDAR or other sensors contain millions or even billions of points, making it challenging to process, store, and transmit them efficiently. Therefore, to improve computational efficiency and reduce memory costs, it is critical to develop point cloud sampling methods to reduce the number of points while preserving the essential features of the original point clouds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 288, + 683 + ], + "type": "text", + "content": "Among point cloud sampling methods, the most widely used one is the farthest point sampling (FPS) strategy in Fig. 1(a). It samples the critical points based on their 3D coordinates iteratively and selects a group of points that" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 228, + 544, + 268 + ], + "blocks": [ + { + "bbox": [ + 310, + 228, + 544, + 268 + ], + "lines": [ + { + "bbox": [ + 310, + 228, + 544, + 268 + ], + "spans": [ + { + "bbox": [ + 310, + 228, + 544, + 268 + ], + "type": "image", + "image_path": "a09d4f79be3b936d458399d759655c882de173dd15cd354a980314777899c59c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 278, + 543, + 407 + ], + "blocks": [ + { + "bbox": [ + 413, + 269, + 441, + 278 + ], + "lines": [ + { + "bbox": [ + 413, + 269, + 441, + 278 + ], + "spans": [ + { + "bbox": [ + 413, + 269, + 441, + 278 + ], + "type": "text", + "content": "(a) FPS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 278, + 543, + 407 + ], + "lines": [ + { + "bbox": [ + 309, + 278, + 543, + 407 + ], + "spans": [ + { + "bbox": [ + 309, + 278, + 543, + 407 + ], + "type": "image", + "image_path": "639e8aff6fa3336d3c4766d69905bb6f61f3e8403cc7098d05fcd14f02faa9b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 407, + 500, + 417 + ], + "lines": [ + { + "bbox": [ + 366, + 407, + 500, + 417 + ], + "spans": [ + { + "bbox": [ + 366, + 407, + 500, + 417 + ], + "type": "text", + "content": "(b) Learnable Task-specific methods" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 417, + 543, + 513 + ], + "blocks": [ + { + "bbox": [ + 310, + 417, + 543, + 513 + ], + "lines": [ + { + "bbox": [ + 310, + 417, + 543, + 513 + ], + "spans": [ + { + "bbox": [ + 310, + 417, + 543, + 513 + ], + "type": "image", + "image_path": "91731ea9eb165e108319db947258a4e810b7899e89e1ac1c7044c5d97d857fe8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 514, + 516, + 523 + ], + "lines": [ + { + "bbox": [ + 358, + 514, + 516, + 523 + ], + "spans": [ + { + "bbox": [ + 358, + 514, + 516, + 523 + ], + "type": "text", + "content": "(c) Learnable Task-agnostic method (Ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 533, + 547, + 610 + ], + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 610 + ], + "type": "text", + "content": "Figure 1. Comparison of FPS, learnable task-specific methods, and our LTA-PCS. FPS samples points using 3D coordinates, which ignores semantic meaning. Learnable Task-specific methods need different sampling networks for different tasks. Our LTA-PCS is task-agnostic, where one trained sampling network can be used for different tasks. Note that the pretrained 2D feature extractor is frozen in training, and not used at inference." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "are farthest apart from each other. Although FPS is task-agnostic, it does not utilize the semantic information of point clouds as the sampling process is solely based on the point coordinates, causing degraded performance on downstream tasks. Recently, in Fig. 1 (b), learning-based point cloud sampling methods [7, 15] are proposed to sample the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 189, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 189, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 189, + 703 + ], + "type": "text", + "content": "* First two authors contributed equally." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 61, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 141, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28035" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "points with more semantic meanings for better downstream task performance. They additionally introduce task-specific losses when learning the network for point cloud sampling. However, these methods are task-specific and require task-oriented labels. In real-world applications, the downstream tasks are usually not specified. But we need to sample fewer points in the collected point cloud by LiDAR or other sensors to save the storage. Moreover, even downstream tasks are given, the existing learning-based methods need to sample the points again based on the new downstream task for better performance, which is cumbersome in practice. Besides, the existing learning-based methods use the pretrained 3D feature extractor (e.g., PointNet [27]) to calculate the task loss for training the sampling network, while the capacity of these 3D feature extractors is limited by both model size and the scale of 3D point cloud datasets, which becomes the bottleneck for training the sampling network with semantically meaningful sampling ability. Finally, the task-specific loss makes them prone to overfit on specific datasets and hard to generalize on different datasets, and the requirement of labels increases the training cost. Although we can remove the task-specific loss in these learning-based methods for task-agnostic point cloud sampling, the sampled points still lack semantic meaning in this case." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 363, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 363, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 363, + 289, + 483 + ], + "type": "text", + "content": "So the problem is: How to design a learnable task-agnostic label-free point cloud sampling strategy with the consideration of semantic information? To solve this problem, in Fig. 1(c), we propose the first Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework, in which the semantic meaning of points is considered in the sampling process. It is non-trivial to construct such a framework. Without the specific downstream tasks, the designed sampling loss for learning-based sampling methods is based on the coordinates, which lack semantic meaning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 289, + 715 + ], + "type": "text", + "content": "In LTA-PCS, we observe 2D feature extractors [31] are effective tools to extract task-agnostic but semantically meaningful features and propose to introduce the fixed 2D feature extractors pre-trained on large-scale datasets to help the training of our LTA-PCS for sampling more semantic meaningful points, where the specific tasks and the corresponding labels are both not required. Specifically, to solve the heterogeneous problem of image and point cloud data, we first project 3D point clouds to multi-view depth maps, which can be readily used by 2D feature extractors to extract semantic meaningful features. Then, we calculate the loss based on both the 3D coordinates and the extracted features and train the sampling network, which can enforce it to recognize meaningful points in terms of both geometric and semantic aspects. After training, we directly use the trained sampling network to generate the simplified point clouds for different downstream tasks, in which an inference alignment strategy is used to ensure the simplified point cloud is a subset of the original dense point cloud." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "In the aforementioned training paradigm, another issue is that the projection from point clouds to multi-view depth maps will inevitably cause information loss, causing inaccurate learning of semantic meaningful points. To this end, we propose a new loss function called semantic loss, which includes both intra- and inter-view losses. Specifically, the intra-view loss minimizes the information loss between the dense point cloud and simplified point cloud under the same views, while the inter-view loss is calculated based on the relationship between different views. By introducing the semantic loss, we can consider both the information under the same and cross views, which can help the sampling network to generate more semantic meaningful points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 287 + ], + "type": "text", + "content": "Moreover, our LTA-PCS framework is also general. Although it is designed for task-agnostic point cloud sampling, we can further combine it with the task-specific loss for task-specific point cloud sampling to achieve better performance when the downstream task is given." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 288, + 533, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 288, + 533, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 533, + 299 + ], + "type": "text", + "content": "The contributions of LTA-PCS are shown as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 300, + 545, + 442 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 300, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 545, + 358 + ], + "type": "text", + "content": "- To the best of our knowledge, we propose the first learnable task-agnostic point cloud sampling framework called LTA-PCS, in which both the geometric and semantic information are considered and the task-oriented labels are not required." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 360, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 545, + 418 + ], + "type": "text", + "content": "- We introduce the pre-trained 2D feature extractor to compute the inter- and intra-view learning objective functions to help train the sampling network in our LTA-PCS framework, which can enforce the sampling network to sample more semantically meaningful points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 419, + 545, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 419, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 419, + 545, + 442 + ], + "type": "text", + "content": "- Extensive experiments on multiple downstream tasks demonstrate the effectiveness of our LTA-PCS." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 392, + 467 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "type": "text", + "content": "Deep learning on point clouds. Recently, deep learning methods [18, 22, 23, 37-39, 48, 53] attract many attentions, in which a large number of methods using point clouds are proposed for 3D scene understanding [9-11, 17, 19-21, 46, 49, 50]. For example, PointNet [27] uses multilayer perceptron to process points for different tasks. PointNet++ [28] proposes a hierarchical structure to extract features. Li et al. [16] proposed PointCNN to perform convolution operations on point clouds. Wu et al. [45] proposed PointConv to introduce point density when perform convolution on point clouds. There are also many methods on point cloud segmentation [40, 41, 43] and object detection [29, 34, 56]. In addition, approaches are also proposed to focus on point cloud completion [4, 32, 54], point cloud registration [2, 33]. On the other hand, some methods project point clouds into volumetric [24, 57] or multiview [13, 31, 35, 52] data forms and process them using neural networks. For example, Zhang et al. [55] project point clouds into multi-view depth images and process them using pretrained CLIP model [31]. Goyal et al. [8] propose" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28036" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 102, + 277, + 170 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 1. Comparison between our LTA-PCS framework with existing point cloud sampling methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 102, + 277, + 170 + ], + "lines": [ + { + "bbox": [ + 53, + 102, + 277, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 102, + 277, + 170 + ], + "type": "table", + "html": "
MethodsSemantic informationTask-agnostic
FPS×✓
Learning-based methods [5, 7, 15]✓×
LTA-PCS✓✓
", + "image_path": "e7bc20b5e7992f8fc801702ec9a33b09f61ee8884c344389c6b3539474f7a6f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 191, + 286, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 286, + 287 + ], + "type": "text", + "content": "SimpleView projection methods for point cloud classification. Zhu et al. [58] utilize pretrained CLIP model for 3D open-world learning. Although these methods use CNNs or CLIP models for learning point cloud features, they do not focus on point cloud sampling. Besides, we utilize the 2D feature extractor in a different perspective, in which the 2D feature extractors are used to facilitate the training of our LTA-PCS for more semantically meaningful sampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 289, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 287, + 563 + ], + "type": "text", + "content": "Point cloud sampling methods. Recently, many point cloud sampling methods were proposed. In Table 1, for the task-agnostic point cloud sampling, the FPS method samples the point that is farthest from the previously sampled ones iteratively. However, this method only uses point coordinates, which does not consider the semantic information of points and leads to a sub-optimal solution when using the sampled points for downstream tasks. Besides, many learning-based approaches were proposed [5, 30, 44]. For example, Dovrat et al. [7] proposed S-NET to use networks for point cloud sampling, where a sampling regularization loss is used to train this S-NET. Lang et al. [15] proposed SampleNet to use differentiable relaxation for point cloud sampling. Qian et al. [30] proposed a task-oriented point cloud downsampling approach. Attention operation is also used for point cloud sampling [42, 51]. Although these methods use networks to sample semantic meaningful points from a dense point cloud, they are task-oriented. Therefore, these approaches need specific downstream tasks to train the sampling network, which is often infeasible in real-world applications. In contrast, our LTA-PCS framework is task-agnostic, and the comparison between our LTA-PCS and the existing methods is shown in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 578, + 129, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 129, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 129, + 590 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 598, + 160, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 160, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 160, + 609 + ], + "type": "text", + "content": "3.1. Problem Statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": "Given a dense point cloud with " + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "inline_equation", + "content": "P = \\{p_i \\in \\mathbb{R}^3, i = 1, \\dots, n\\}" + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": " and a target size " + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "inline_equation", + "content": "m \\leq n" + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": ", our goal in this work is to learn a sampling network that can sample " + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": " points from " + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 616, + 286, + 688 + ], + "type": "text", + "content": " and generate a simplified point cloud that can best represent the dense point cloud with task-agnostic semantic meaning as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 700, + 286, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 700, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 102, + 700, + 286, + 714 + ], + "type": "interline_equation", + "content": "T (P) \\subset P, | T (P) | = m \\leq n, \\tag {1}", + "image_path": "6269a993355e617a60d32a0377fea969a8a62f0bbe1be3cabe4de553f5d89631.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": " is the operation of the sampling network, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "T(P)" + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": " is the simplified point cloud. In our LTA-PCS, the learning process of the sampling network " + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": " is not related to any downstream tasks. In other words, we aim to find the best sampling network to generate the simplified point cloud that can perform well on the downstream tasks. Traditional point cloud sampling methods (e.g., FPS) only sample the points based on 3D coordinates, which lacks semantic information. To introduce semantic information in sampling, learning-based sampling methods like S-NET [7] and SampleNet [15] use the task-specific loss to train the sampling network. However, when the task is not given, it is still an open problem on how to introduce such a semantic-aware loss for guiding the sampling network to sample semantically meaningful points. Therefore, we propose to use effective and well-performed pretrained 2D feature extractors to facilitate the training of the sampling network for LTA-PCS in a task-agnostic way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 297, + 373, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 297, + 373, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 373, + 308 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 316, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 545, + 579 + ], + "type": "text", + "content": "Fig. 2 shows the overview when training our LTA-PCS framework. Given a dense point cloud, we use the sampling network in LTA-PCS to sample points from this point cloud and generate the simplified point cloud. To utilize the pretrained 2D feature extractors, we first project both the dense and simplified point clouds to multi-view depth maps. Then, we use the pretrained 2D feature extractor to extract semantically meaningful features based on multi-view depth maps. In our implementation, we choose the vision branch of the CLIP model [31] as our pretrained 2D feature extractor, and the parameters of the 2D feature extractor are frozen in the training process. After that, we calculate the proposed semantic loss including both inter- and intraview losses based on the output feature from 2D feature extractors, which introduces the semantic-level supervision to help the sampling network in our LTA-PCS framework and sample more semantically meaningful points. Besides, to fully utilize the 3D coordinate information of point clouds, we also use geometric loss [7] to introduce geometric-level supervision when training our LTA-PCS framework in the training process. We update the sampling network based on both geometric loss and semantic loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "content": "Sampling network. Our sampling model follows the architecture of [7, 27]. The input points are processed by a series of " + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 580, + 545, + 665 + ], + "type": "text", + "content": " convolution layers, which produce a feature vector for each point. Then, a symmetric feature-wise max pooling operation is applied to obtain a global feature vector. Finally, we use several fully-connected layers. The output of the last layer is the set of generated points." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Projection. Directly using pretrained 2D feature extractor is infeasible because of the heterogeneous problem between 2D images and 3D point clouds. To solve this problem, we need to project point clouds to multi-view" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28037" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 247 + ], + "type": "image", + "image_path": "57606d2da4553e99f02718743d3b3534e58960f02d043d4dbe5e129bfb8e0c7d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 546, + 312 + ], + "type": "text", + "content": "Figure 2. Training procedure of our LTA-PCS framework. Given a dense point cloud, we first use the initialized sampling network to sample points from it and generate a simplified point cloud. Then, we project both the dense and simplified point clouds to multi-view depth maps, which will be used for feature extraction by a frozen pretrained 2D feature extractor. We calculate the semantic loss including intra-view and inter-view losses using the extracted features. Geometric loss is also used in training. At the inference stage, we directly use the trained sampling network to sample points from dense point clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "text", + "content": "depth maps. Specifically, inspired by [55], we take the bottom view as an example, the point with the coordinate of " + }, + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "text", + "content": " will be projected to " + }, + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "inline_equation", + "content": "(\\lceil x / z\\rceil ,\\lceil y / z\\rceil)" + }, + { + "bbox": [ + 46, + 342, + 288, + 426 + ], + "type": "text", + "content": ". Also, as the vision branch of CLIP model aims to process images with three channels, we copy the projected depth maps three times and form the input of the CLIP model, which ensures there is no shape mismatch when using the CLIP model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 433, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 469 + ], + "type": "text", + "content": "Inference. At inference, we use the sampling network in Fig. 2 to sample points from the dense point cloud, which can be used for multiple downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 498, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 135, + 510 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 525, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 548 + ], + "type": "text", + "content": "Overall loss function. Formally, the overall loss function of our LTA-PCS can be written as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 574, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 574, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 124, + 574, + 287, + 587 + ], + "type": "interline_equation", + "content": "L = L _ {g e o} + \\alpha L _ {s e m}, \\tag {2}", + "image_path": "56bc4139a36c7ce3951f32e5162627664b14d48fe34efe4bdfe18b5a3a10c9f2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "inline_equation", + "content": "L_{geo}" + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "text", + "content": " is the geometric loss, and " + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "text", + "content": " is the newly proposed semantic loss, which includes both the intra- and inter-view losses. " + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 610, + 287, + 682 + ], + "type": "text", + "content": " is the coefficient to balance these two terms. In LTA-PCS, the geometric and semantic losses aim to preserve the geometric and semantic consistency between the original point cloud and the simplified point cloud." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Geometric loss. We denote " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " as the simplified point cloud, i.e., " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "Q = T(P)" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": ". Inspired by [7], the geometric loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 342, + 394, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 394, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 394, + 353 + ], + "type": "text", + "content": "is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 362, + 512, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 362, + 512, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 362, + 512, + 376 + ], + "type": "interline_equation", + "content": "L _ {g e o} = L _ {a} (Q, P) + \\beta L _ {w} (Q, P) + \\gamma L _ {s} (Q, P),", + "image_path": "2699844fed035a078227f3cf5f28b24d51a24b0997a40a3252a6b9165ef1fa67.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 378, + 512, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 378, + 512, + 407 + ], + "spans": [ + { + "bbox": [ + 335, + 378, + 512, + 407 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} L _ {a} (Q, P) = \\frac {1}{| Q |} \\sum_ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2},", + "image_path": "15b98ed49de8c96eb6f5e75590e2578f70c98fd96b286a9d379785545a714792.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 406, + 545, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 406, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 375, + 406, + 545, + 429 + ], + "type": "interline_equation", + "content": "L _ {w} (Q, P) = \\max _ {q \\in Q} \\min _ {p \\in P} \\| q - p \\| _ {2} ^ {2}, \\tag {3}", + "image_path": "df877d6c69762e2bd91a54787639880ce8284f6b643fe2bc263e0aba71247861.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 430, + 512, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 430, + 512, + 460 + ], + "spans": [ + { + "bbox": [ + 362, + 430, + 512, + 460 + ], + "type": "interline_equation", + "content": "L _ {s} (Q, P) = \\frac {1}{| P |} \\sum_ {p \\in P} m i n _ {q \\in Q} | | p - q | | _ {2} ^ {2}.", + "image_path": "29ff6c71ff002717e018636d251042aada27fe460b98fcd99977f3eef16717d8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "L_{a}(Q,P)" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " is to minimize the average distance between " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "L_{w}(Q,P)" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " is to minimize the maximum distance between " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "L_{s}(Q,P)" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " ensures the points in " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " evenly spread in the 3D coordinate space. Following [7], " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 471, + 545, + 530 + ], + "type": "text", + "content": " are the loss weights." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": "Semantic loss. To enforce the sampling network to sample more semantically meaningful points, we also introduce the newly proposed semantic loss " + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": " in the training process. Formally, let us suppose the dense point cloud as " + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": " and the simplified point cloud generated by the initialized sampling network as " + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 531, + 545, + 615 + ], + "type": "text", + "content": ". The semantic loss consists of intraview and inter-view loss, which can be written as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 629, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 629, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 370, + 629, + 545, + 643 + ], + "type": "interline_equation", + "content": "L _ {s e m} = L _ {i n t r a} + \\lambda L _ {i n t e r}. \\tag {4}", + "image_path": "58c3e0649221ce7272a4cc45dc2cd5f970f6fc92fe218cf9fa0fd6e4bd12a629.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 651, + 544, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 544, + 673 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 544, + 673 + ], + "type": "text", + "content": "The intra-view loss aims to minimize the semantic information loss under the same view as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 683, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 683, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 335, + 683, + 545, + 716 + ], + "type": "interline_equation", + "content": "L _ {i n t r a} = \\sum_ {k = 1} ^ {K} M S E [ E (J _ {k} (P)), E (J _ {k} (Q)) ], \\tag {5}", + "image_path": "60de9073fe76db92f28ca3da2cdb6d4950a92a5596385cc45bd0fb7796f4714f.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28038" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 251, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 251, + 83 + ], + "type": "text", + "content": "Algorithm 1 Training procedure of our LTA-PCS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": "Input: Dense point cloud " + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": "; Simplified point cloud " + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": "; Randomly initialized sampling network " + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": "; Number of projected views " + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": "; Pretrained 2D feature extractor " + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 49, + 87, + 287, + 123 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 50, + 124, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 124, + 240, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 240, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 240, + 135 + ], + "type": "text", + "content": "1: for each iteration in the training process do" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 136, + 209, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 209, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 209, + 148 + ], + "type": "text", + "content": "2: // Calculate geometric loss " + }, + { + "bbox": [ + 53, + 136, + 209, + 148 + ], + "type": "inline_equation", + "content": "L_{geo}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "text", + "content": "3: Use " + }, + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "text", + "content": " to generate the simplified point cloud " + }, + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 53, + 148, + 272, + 159 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "text", + "content": "4: Calculate geometric loss " + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "inline_equation", + "content": "L_{geo}" + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "text", + "content": " based on dense point cloud " + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "text", + "content": " and the simplified point cloud " + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 53, + 159, + 287, + 183 + ], + "type": "text", + "content": " using Eq. (3);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 184, + 208, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 208, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 208, + 194 + ], + "type": "text", + "content": "5: // Calculate semantic loss " + }, + { + "bbox": [ + 53, + 184, + 208, + 194 + ], + "type": "inline_equation", + "content": "L_{sem}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "text", + "content": "6: for view " + }, + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 53, + 196, + 159, + 206 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": "7: Project dense point cloud " + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": " and simplified point cloud " + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": "-th view depth map " + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "J_{k}(P)" + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "J_{k}(Q)" + }, + { + "bbox": [ + 53, + 207, + 287, + 243 + ], + "type": "text", + "content": ", respectively;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "text", + "content": "8: Extract the features of " + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "inline_equation", + "content": "J_{k}(P)" + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "inline_equation", + "content": "J_{k}(Q)" + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "text", + "content": " using the 2D feature extractor " + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 53, + 243, + 287, + 266 + ], + "type": "text", + "content": ", respectively;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 267, + 113, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 267, + 113, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 267, + 113, + 277 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 50, + 279, + 287, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 279, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 50, + 279, + 287, + 303 + ], + "type": "text", + "content": "10: Calculate intra- and intra-view losses by Eq. (5) and Eq. (6);" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 50, + 303, + 270, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 303, + 270, + 315 + ], + "spans": [ + { + "bbox": [ + 50, + 303, + 270, + 315 + ], + "type": "text", + "content": "11: Calculate semantic loss " + }, + { + "bbox": [ + 50, + 303, + 270, + 315 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 50, + 303, + 270, + 315 + ], + "type": "text", + "content": " based on Eq. (4);" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 50, + 316, + 278, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 316, + 278, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 316, + 278, + 327 + ], + "type": "text", + "content": "12: Calculate overall object function " + }, + { + "bbox": [ + 50, + 316, + 278, + 327 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 50, + 316, + 278, + 327 + ], + "type": "text", + "content": " using Eq. (2);" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 50, + 327, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 327, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 50, + 327, + 286, + 350 + ], + "type": "text", + "content": "13: Back propagate the gradients and update the parameters of the sampling network " + }, + { + "bbox": [ + 50, + 327, + 286, + 350 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 50, + 327, + 286, + 350 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 50, + 351, + 98, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 351, + 98, + 361 + ], + "spans": [ + { + "bbox": [ + 50, + 351, + 98, + 361 + ], + "type": "text", + "content": "14: end for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 362, + 234, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 362, + 234, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 362, + 234, + 374 + ], + "type": "text", + "content": "Output: The optimized sampling network " + }, + { + "bbox": [ + 48, + 362, + 234, + 374 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "inline_equation", + "content": "J_{k}(\\cdot)" + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": " denotes the projection operation for the " + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": "-th view, and " + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": " is the 2D feature extractor for extracting semantic features. " + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "inline_equation", + "content": "MSE(\\cdot)" + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": " is the mean square error, and " + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 404, + 287, + 453 + ], + "type": "text", + "content": " is the total number of projected views." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 503 + ], + "type": "text", + "content": "As the projection operation inevitably causes information loss and the relationship of different views can also provide essential information [25, 26], we calculate the interview loss for compensation as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 523, + 235, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 523, + 235, + 537 + ], + "spans": [ + { + "bbox": [ + 107, + 523, + 235, + 537 + ], + "type": "interline_equation", + "content": "L _ {i n t e r} = M S E \\left(\\boldsymbol {\\Phi} \\boldsymbol {\\Phi} ^ {T}, \\boldsymbol {\\Psi} \\boldsymbol {\\Psi} ^ {T}\\right),", + "image_path": "f5e7abf0d96e3c76d9c736beb6974da2d73d2041607952aff0ed13ca4f9ddb54.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 86, + 540, + 287, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 540, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 86, + 540, + 287, + 557 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\Phi = \\underset {K} {\\operatorname {C o n c a t}} \\left[ \\operatorname {F l a t} \\left(E \\left(J _ {k} (P)\\right)\\right) \\right], \\tag {6}", + "image_path": "514e751cd23b2be5b402a385b34eacba688a3404256fcb83163b4f7bd419f210.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 113, + 559, + 248, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 559, + 248, + 577 + ], + "spans": [ + { + "bbox": [ + 113, + 559, + 248, + 577 + ], + "type": "interline_equation", + "content": "\\Psi = \\underset {K} {C o n c a t} \\left[ F l a t (E (J _ {k} (Q))) \\right].", + "image_path": "6dee0b410d1a3c18369d9a6fa2501f1e11b638cc79e1761d441acc3908e9f3db.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "inline_equation", + "content": "Flat(\\cdot)" + }, + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "text", + "content": " is the flatten operation. " + }, + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "inline_equation", + "content": "Concat_{K}" + }, + { + "bbox": [ + 47, + 590, + 287, + 651 + ], + "type": "text", + "content": " concatenate the vectors along different views. By using the interview loss, we obtain the relation information in different views from dense point cloud and enforce the simplified point cloud to mimic this information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": "Based on the intra-view loss " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{intra}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ", inter-view loss " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{inter}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ", and the geometric loss " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "L_{geo}" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ", we obtain the loss function " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": " in Eq. (2) to train the sampling network " + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 654, + 287, + 713 + ], + "type": "text", + "content": ", which is used to generate the simplified point cloud at inference. The training procedure is also illustrated in Alg. 1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 72, + 373, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 373, + 83 + ], + "type": "text", + "content": "3.4. Inference" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": "At inference, we use the trained sampling network " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": " to generate the coordinates of the sampled points. However, as we use several fully-connected layers to directly predict the coordinates of the sampled points at the end of the sampling network, we cannot ensure the simplified point cloud " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": " generated by the sampling network to be a subset of dense point cloud " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": ". Thus, following S-Net [7], we can obtain the points in " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": " to its nearest neighbors in dense point cloud " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": ". Specifically, for each point " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": ", the matched point " + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "inline_equation", + "content": "q^*" + }, + { + "bbox": [ + 304, + 91, + 545, + 210 + ], + "type": "text", + "content": " is written as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 375, + 220, + 545, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 220, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 375, + 220, + 545, + 242 + ], + "type": "interline_equation", + "content": "q ^ {*} = \\underset {p \\in P} {\\operatorname {a r g m i n}} | | q - p | | _ {2} ^ {2}. \\tag {7}", + "image_path": "647b4d454773419073fbe1e471d975d2b8ec1476b006d4799b46261d55cc42fc.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "content": "After the matching process for " + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "content": " points, we can obtain the simplified point cloud " + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "inline_equation", + "content": "Q^{*} = \\{q_{i}^{*}\\}_{i = 1}^{m}" + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "content": ", which is a subset of " + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 251, + 545, + 287 + ], + "type": "text", + "content": ". Note that the matching process is very fast." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 448, + 308 + ], + "type": "text", + "content": "3.5. Downstream Tasks Usage" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 434 + ], + "type": "text", + "content": "After training the sampling network in our LTA-PCS, we can use it to effectively generate the simplified point cloud with the information from both 3D coordinates and the semantic meanings. As the downstream tasks are not decided at this stage, we store these simplified point clouds as a new point cloud dataset to simulate this scenario. When the downstream tasks are given, we use the simplified point clouds to train different backbone network architectures (e.g., PointNet [27]) for different downstream tasks as before, and the backbone network is initialized from scratch." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 444, + 378, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 378, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 378, + 455 + ], + "type": "text", + "content": "3.6. Discussion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "Necessity of Learnable Task-Agnostic Point Cloud Sampling. We would emphasize that LTA-PCS tries to solve the point cloud sampling under an important scenario where the collected point clouds are given but the downstream tasks are not given. In real-world applications, this scenario often occurs. We often use LiDAR to collect point clouds and store them. When specific tasks are given, we use the stored point cloud data for downstream tasks. Thus, sampling semantically meaningful points from dense collected point clouds becomes critical for storage saving. Existing learning-based methods [15] use task-specific loss to introduce semantic information in sampling, which is infeasible in this scenario as the downstream tasks and the corresponding labels are required. Although we can remove task-specific loss in these learning-based methods for task-agnostic point cloud sampling, semantic information will not be considered in this case. Moreover, the task-specific loss will degrade the generalization ability of these methods on different datasets. LTA-PCS targets at solving the point cloud sampling under a new setting, which cannot be achieved by existing learning-based approaches." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28039" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "Discussion on the reason that our task-agnostic point cloud sampling method (i.e., LTA-PCS) yields better results relative to task-specific methods (e.g., S-Net, SampleNet). We hypothesize this is because we introduce the 2D pretrained model in the training process. Therefore, the knowledge in 2D pretrained model can significantly improve the performance of the sampling network. In contrast, due to limited network capacity or dataset size, there is no general and well-performed 3D pretrained model when compared with 2D pretrained models. Therefore, although existing state-of-the-art point cloud sampling methods are task-intensive, they cannot well utilize the prior knowledge in 2D pretrained model, resulting worse performance than our task-agnostic method LTA-PCS. Our LTA-PCS framework can be also combined with task-specific loss, and the results show the performance can be further boosted." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 289, + 420 + ], + "type": "text", + "content": "Discussion on the necessity of 2D feature extractor. We choose the pretrained 2D feature extractor to introduce semantic information when training our LTA-PCS framework as there is no existing well-performed pretrained 3D feature extractor on large-scale datasets in the literature. Current 3D feature extractors like PointNet [27] or PointNet++ [28] are trained based on specific tasks and datasets, which cannot be used in the task-agnostic setting. We hypothesize that this is because of the lack of large-scale high-quality 3D point cloud datasets. Therefore, although we encounter information loss in the projection process, we choose a 2D feature extractor to introduce semantic information when constructing our LTA-PCS." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 128, + 448 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 180, + 467 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "We follow [7] to utilize a set of multi-layer perceptrons (MLPs) as the sampling network, and batch normalization and ReLU layers are appended after each MLP. In training, we train the sampling network for 50 epochs with Adam optimizer [14] on both ModelNet40, ShapeNet core55 and ScanObjectNN datasets. The batch size is set as 8 and the learning rate is " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": ". We set the weight decay as " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": ". In the training process, we set the hyperparameter " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " in Eq. (2) as 1. " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " in Eq. (3) are set as 1 and 1, respectively, and " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " in Eq. (4) is set as 0.01. The number of views (i.e., " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": ") is set as 10. After training the sampling network, we use it to generate simplified point clouds. Then, we use the generated simplified point clouds to train an initialized backbone network for different downstream tasks. For the point cloud classification task, we use PointNet [27] as our backbone network and train the initialized PointNet for 200 epochs with a batch size of 12. The learning rate is set as " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": " with weight decay of " + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": ". Step learning rate decay is used. The learning rate will multiply 0.7 after every 20 epochs. For the point cloud completion and registration tasks, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "use PCN [54] and PointNetLK [1] as baseline methods, respectively, and the training epochs are set as 400 and 200, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 115, + 451, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 115, + 451, + 128 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 451, + 128 + ], + "type": "text", + "content": "4.2. Point Cloud Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 134, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 547, + 469 + ], + "type": "text", + "content": "In Table 2, we compare our LTA-PCS with several existing point cloud sampling methods on the ModelNet40 [47], ShapeNet [3] and ScanObjectNN (hardest perturbed variant (PB_T50_RS)) [36] for point cloud classification, where the instance accuracy is used as the evaluation metric. Specifically, Random and FPS are task-agnostic non-learnable methods, and we directly sample a specific number of points by using the corresponding methods. The baseline approaches S-Net [7], SampleNet [15] are task-specific learnable methods, and we first train the sampling network and use the trained network to downsample the input point cloud to a predefined number of points. However, we can remove the task-specific loss in these learnable sampling approaches to achieve task-agnostic point cloud sampling, which are denoted by S-Net* and SampleNet*, respectively. In contrast, LTA-PCS is a learnable approach without task-specific loss in training. Therefore, it is a task-agnostic point cloud sampling method. In Table 2, we observe that our LTA-PCS achieves better results than other methods. Remarkably, we outperform the state-of-the-art SampleNet* under the task-agnostic setting by " + }, + { + "bbox": [ + 304, + 134, + 547, + 469 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 134, + 547, + 469 + ], + "type": "text", + "content": " when sampling to 256 points on ModelNet40. Moreover, LTA-PCS even performs better than the task-specific learnable point cloud sampling methods S-Net and SampleNet. One possible explanation is that S-Net and SampleNet use pretrained 3D feature extractors (e.g., PointNet [27]) to calculate the task loss, and their results are limited by the capacity of 3D feature extractors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 477, + 432, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 432, + 489 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 432, + 489 + ], + "type": "text", + "content": "4.3. Point Cloud Retrieval" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "content": "We also follow S-Net [7] to report the results of the point cloud retrieval task. For point cloud retrieval, we use ModelNet40 to evaluate our LTA-PCS framework, where the mean average precision (mAP) evaluation metric is used. In Table 3, our LTA-PCS achieves promising performance. Specifically, we first downsample the point cloud to the corresponding number of points, and then extract the features by PointNet. Following [7], the results are computed by " + }, + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 495, + 547, + 616 + ], + "type": "text", + "content": " distance of the shape descriptor, which is the feature of the penultimate layer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 623, + 443, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 443, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 443, + 635 + ], + "type": "text", + "content": "4.4. Point Cloud Completion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "We follow PCN [54] to reconstruct the simplified point clouds into the original dense point clouds with 2,048 points on ShapeNet Core55, and we provide the " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " Chamfer Distance (CD) results of different methods in Fig. 3. Note that for the task-specific methods, we need to additionally train the sampling network when training the PCN method. From" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28040" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 102, + 492, + 235 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 2. Comparison of our LTA-PCS with other methods under the point cloud classification task. * indicates we remove the task-specific loss in training for task-agnostic setting. \"Oracle\" denotes the accuracy using original dense point clouds." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 102, + 492, + 235 + ], + "lines": [ + { + "bbox": [ + 100, + 102, + 492, + 235 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 492, + 235 + ], + "type": "table", + "html": "
DatasetsModelNet (Oracle: 90.3)ShapeNet (Oracle: 85.0)
#Sampled points6412825651264128256512
Task-AgnosticRandom81.586.286.688.279.380.981.983.2
FPS86.187.988.188.380.581.582.483.3
S-Net* [7]86.187.888.288.580.681.382.283.0
SampleNet* [15]86.688.387.988.480.381.682.482.9
LTA-PCS88.289.190.090.382.083.183.584.4
Task-SpecificS-Net [7]87.788.188.489.080.982.382.783.6
SampleNet [15]87.988.388.088.781.082.582.783.4
", + "image_path": "3e9d3417494ec47559b4264c6e8c54ecb90ad25b77f6913e12ba1e20dbb599bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 293, + 267, + 441 + ], + "blocks": [ + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "type": "text", + "content": "Table 3. Comparison of our LTA-PCS with other methods under the point cloud retrieval task. * indicates we remove the task-specific loss in training for task-agnostic setting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 293, + 267, + 441 + ], + "lines": [ + { + "bbox": [ + 66, + 293, + 267, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 293, + 267, + 441 + ], + "type": "table", + "html": "
#Sampled points64128256512
Task-Agnostic
Random66.169.472.573.3
FPS69.172.573.074.2
S-Net* [7]69.471.773.674.1
SampleNet* [15]70.472.372.674.7
LTA-PCS71.473.274.877.2
Task-Specific
S-Net [7]70.572.774.076.2
SampleNet [15]70.972.573.676.0
", + "image_path": "c0e017188ca4333f0fc28e23bc7f9e7dbe89200b28fc1a9cc639436658b1ae63.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 72, + 452, + 262, + 590 + ], + "blocks": [ + { + "bbox": [ + 72, + 452, + 262, + 590 + ], + "lines": [ + { + "bbox": [ + 72, + 452, + 262, + 590 + ], + "spans": [ + { + "bbox": [ + 72, + 452, + 262, + 590 + ], + "type": "image", + "image_path": "43949d0a2cf7f86b16f404b23e8ed2a520b91e2f61594e1073695ce9b421a7e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "lines": [ + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "type": "text", + "content": "Figure 3. Performance on the point cloud completion task. Lower " + }, + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "type": "text", + "content": " Chamber Distance (" + }, + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 599, + 288, + 623 + ], + "type": "text", + "content": " CD) denotes better performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 700 + ], + "type": "text", + "content": "Fig. 3, we observe that the quantitative results of our LTA-PCS are better than other methods under different numbers of sampled points. The results show our LTA-PCS is effective for different downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 713 + ], + "type": "text", + "content": "Visualization. We provide the visualization results of dif" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 337, + 251, + 379, + 369 + ], + "blocks": [ + { + "bbox": [ + 337, + 251, + 379, + 369 + ], + "lines": [ + { + "bbox": [ + 337, + 251, + 379, + 369 + ], + "spans": [ + { + "bbox": [ + 337, + 251, + 379, + 369 + ], + "type": "image", + "image_path": "67249e5907f907dc8eb9a74b3df33bbd0840092a8175338d7ece78cb81ef125d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 381, + 251, + 447, + 369 + ], + "blocks": [ + { + "bbox": [ + 381, + 251, + 447, + 369 + ], + "lines": [ + { + "bbox": [ + 381, + 251, + 447, + 369 + ], + "spans": [ + { + "bbox": [ + 381, + 251, + 447, + 369 + ], + "type": "image", + "image_path": "2b2339d6ca3e998e3c70e6b9b168daf3849aea1d03a7c0852928e966b373a531.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 370, + 468, + 378 + ], + "lines": [ + { + "bbox": [ + 386, + 370, + 468, + 378 + ], + "spans": [ + { + "bbox": [ + 386, + 370, + 468, + 378 + ], + "type": "text", + "content": "(a) Simplified Point Clouds" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 474, + 252, + 514, + 369 + ], + "blocks": [ + { + "bbox": [ + 474, + 252, + 514, + 369 + ], + "lines": [ + { + "bbox": [ + 474, + 252, + 514, + 369 + ], + "spans": [ + { + "bbox": [ + 474, + 252, + 514, + 369 + ], + "type": "image", + "image_path": "2bc7da72e66a44494ed9853d8d4bb794fb684ec868c21581016f4e59a82be310.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 338, + 380, + 375, + 491 + ], + "blocks": [ + { + "bbox": [ + 338, + 380, + 375, + 491 + ], + "lines": [ + { + "bbox": [ + 338, + 380, + 375, + 491 + ], + "spans": [ + { + "bbox": [ + 338, + 380, + 375, + 491 + ], + "type": "image", + "image_path": "72670124e1ff0375e23e17baeb240a0ae37aa6779b6c9d7f72da1ac08c0b23e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 492, + 365, + 499 + ], + "lines": [ + { + "bbox": [ + 352, + 492, + 365, + 499 + ], + "spans": [ + { + "bbox": [ + 352, + 492, + 365, + 499 + ], + "type": "text", + "content": "FPS" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 408, + 383, + 446, + 492 + ], + "blocks": [ + { + "bbox": [ + 408, + 383, + 446, + 492 + ], + "lines": [ + { + "bbox": [ + 408, + 383, + 446, + 492 + ], + "spans": [ + { + "bbox": [ + 408, + 383, + 446, + 492 + ], + "type": "image", + "image_path": "405ca913013e76355f628aba34b52ed1b8ef4943e9c2a0476fde4789cb783d98.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 493, + 437, + 499 + ], + "lines": [ + { + "bbox": [ + 420, + 493, + 437, + 499 + ], + "spans": [ + { + "bbox": [ + 420, + 493, + 437, + 499 + ], + "type": "text", + "content": "S-Net" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 499, + 474, + 507 + ], + "lines": [ + { + "bbox": [ + 391, + 499, + 474, + 507 + ], + "spans": [ + { + "bbox": [ + 391, + 499, + 474, + 507 + ], + "type": "text", + "content": "(b) Completed Point Clouds" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 380, + 515, + 491 + ], + "blocks": [ + { + "bbox": [ + 476, + 380, + 515, + 491 + ], + "lines": [ + { + "bbox": [ + 476, + 380, + 515, + 491 + ], + "spans": [ + { + "bbox": [ + 476, + 380, + 515, + 491 + ], + "type": "image", + "image_path": "c64edf7b5c65fc1cc71e7aa59d0e4759ee95ae4e2debde5753d3097f3528ce7e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 493, + 511, + 499 + ], + "lines": [ + { + "bbox": [ + 481, + 493, + 511, + 499 + ], + "spans": [ + { + "bbox": [ + 481, + 493, + 511, + 499 + ], + "type": "text", + "content": "LTA-PCS" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 517, + 547, + 551 + ], + "lines": [ + { + "bbox": [ + 305, + 517, + 547, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 517, + 547, + 551 + ], + "type": "text", + "content": "Figure 4. (a). Simplified point clouds. (The sampled points are enlarged for better visualization.) (b). Completed point clouds by using the corresponding simplified point clouds as input." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 547, + 641 + ], + "type": "text", + "content": "ferent methods in Fig. 4. Specifically, in Fig. 4(a), the simplified point clouds are produced by FPS, S-Net and LTA-PCS, where the number of points in each simplified point cloud are 64, and in Fig. 4(b), we observe that the reconstruction quality of LTA-PCS is better than other methods." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 400, + 659 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Effect of different losses. In Table 4 we provide three alternative variants of our LTA-PCS: LTA-PCS (w/o inter & w/o intra), LTA-PCS (w/o inter), LTA-PCS (w/o intra). Specifically, for LTA-PCS (w/o inter & w/o intra), we remove the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28041" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 91, + 269, + 164 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 246, + 82 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 246, + 82 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 246, + 82 + ], + "type": "text", + "content": "Table 4. Ablation on using different losses." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 91, + 269, + 164 + ], + "lines": [ + { + "bbox": [ + 63, + 91, + 269, + 164 + ], + "spans": [ + { + "bbox": [ + 63, + 91, + 269, + 164 + ], + "type": "table", + "html": "
#Sampled points128256
LTA-PCS89.190.0
LTA-PCS (w/o inter & w/o intra)87.888.2
LTA-PCS (w/o inter)88.689.4
LTA-PCS (w/o intra)88.389.1
", + "image_path": "830cdd102554f83472e7d9a75ace05d1653f650ee19122348cb1fa640cfff76f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 198, + 270, + 259 + ], + "blocks": [ + { + "bbox": [ + 67, + 177, + 266, + 189 + ], + "lines": [ + { + "bbox": [ + 67, + 177, + 266, + 189 + ], + "spans": [ + { + "bbox": [ + 67, + 177, + 266, + 189 + ], + "type": "text", + "content": "Table 5. Ablation on using different feature extractors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 198, + 270, + 259 + ], + "lines": [ + { + "bbox": [ + 65, + 198, + 270, + 259 + ], + "spans": [ + { + "bbox": [ + 65, + 198, + 270, + 259 + ], + "type": "table", + "html": "
#Sampled points128
LTA-PCS (PointNet)88.0
LTA-PCS (R50 trained on ImageNet)88.3
LTA-PCS (R50 of CLIP vision encoder)89.1
", + "image_path": "4777b606c138c56074ae186d20a4aae98be20c0c14dc06f48c896d66dde0051b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 95, + 293, + 237, + 330 + ], + "blocks": [ + { + "bbox": [ + 67, + 273, + 266, + 283 + ], + "lines": [ + { + "bbox": [ + 67, + 273, + 266, + 283 + ], + "spans": [ + { + "bbox": [ + 67, + 273, + 266, + 283 + ], + "type": "text", + "content": "Table 6. Ablation on using different numbers of views." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 293, + 237, + 330 + ], + "lines": [ + { + "bbox": [ + 95, + 293, + 237, + 330 + ], + "spans": [ + { + "bbox": [ + 95, + 293, + 237, + 330 + ], + "type": "table", + "html": "
# views2610
LTA-PCS88.288.989.1
", + "image_path": "1f95826a4604520d2d2a255a809fd9db0c2a7a4b19bd4985fba4953f5ac42c59.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 460 + ], + "type": "text", + "content": "semantic losses (i.e., both inter-view and intra-view losses), which is the same as the S-Net [7] without using the task loss. For LTA-PCS (w/o inter), we only use the intra-view loss without using the inter-view loss to train our LTA-PCS framework. For LTA-LTA-PCS (w/o intra), we only use the inter-view loss without using the intra-view loss to train the LTA-PCS framework. In Table 4, our LTA-PCS is better than these three alternative variants, which shows that it is beneficial to utilize both intra-view and inter-view losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 714 + ], + "type": "text", + "content": "Effect of 2D feature extractor. In Table 5, we compare the results of our LTA-PCS by using different pretrained vision backbones [6, 12] to show the effect of 2D feature extractor. Specifically, for LTA-PCS (PointNet), we directly use the PointNet backbone to extract the features of the original dense point cloud and simplified point cloud and calculate the MSE distance of corresponding features to preserve semantic consistency. For LTA-PCS (R50 trained on ImageNet), we use the ResNet-50 [12] model trained on ImageNet as the feature extractor. For LTA-PCS (R50 of CLIP vision encoder) and LTA-PCS (VIT-B/16 of CLIP vision encoder), we use the pretrained CLIP vision encoder with different backbones. In Table 5, first, the results of LTA-PCS (PointNet) are lower than other methods a lot, which indicates that the 3D feature extractor (i.e., PointNet) cannot provide high-quality semantic supervision to train the sampling network. Second, we observe that the results of LTA-PCS (R50 of CLIP vision encoder) are better than the results of LTA-PCS (R50 trained on ImageNet) by a large margin, which shows the capacity of the feature extractor is critical when training the sampling network. Therefore," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "we choose the pre-trained 2D feature extractor to maintain semantic consistency when training the sampling network instead of using the pre-trained 3D feature extractor." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 108, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 240 + ], + "type": "text", + "content": "Effect of the number of views. In Table 6, we report the performance results of using 128 sampled points as input to analyze the effect of the number of projected views, where we report the results on the ModelNet40 dataset based on the PointNet backbone. When the number of views increases from 2 to 10, LTA-PCS can achieve better performance results, which indicates that it is effective to use more views for maintaining the semantic consistency between the original and simplified point clouds. Therefore, we set the projected number of views as 10 to achieve better performance with acceptable memory usage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 247, + 408, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 408, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 408, + 260 + ], + "type": "text", + "content": "4.6. Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 308, + 544, + 387 + ], + "blocks": [ + { + "bbox": [ + 306, + 277, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 277, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 545, + 300 + ], + "type": "text", + "content": "Table 7. Inference time and GPU memory usage of our LTA-PCS with different sampled points." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 308, + 544, + 387 + ], + "lines": [ + { + "bbox": [ + 309, + 308, + 544, + 387 + ], + "spans": [ + { + "bbox": [ + 309, + 308, + 544, + 387 + ], + "type": "table", + "html": "
#Sampled pointsInference time (ms)GPU usage (MB)
640.28578
1280.37616
2560.58724
5121.05940
20481.772326
", + "image_path": "3c98702e118e200faab87238d5556ad2d249fda66e879dfc5c5187816d1a00b8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 399, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 496 + ], + "type": "text", + "content": "Comparison on inference time and GPU memory usage. In Table 7, we take the ModelNet40 dataset as an example and compare the inference time and GPU memory usage under different sampled points when using PointNet as the backbone network. As shown in Table 7, we observe that using fewer points can reduce inference time and GPU memory usage greatly, which further demonstrates the advantages of point cloud sampling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 506, + 378, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 506, + 378, + 518 + ], + "spans": [ + { + "bbox": [ + 306, + 506, + 378, + 518 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 526, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 657 + ], + "type": "text", + "content": "In this work, we propose the Learnable Task-Agnostic Point Cloud Sampling (LTA-PCS) framework for task-agnostic point cloud sampling task, which aims to preserve the geometric and semantic consistency in point cloud sampling. Specifically, we first propose to use the pre-trained 2D feature extractor for training the sampling network, and use semantic loss including both inter-view and intra-view losses. Comprehensive experimental results on multiple downstream tasks show the effectiveness of LTA-PCS. In the future, we will continue to explore more tasks based on our LTA-PCS method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 702 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by National Science and Technology Major Project (2022ZD0116405) and National Natural Science Foundation of China (No. 61932002, No. 62306025, No. 92367204)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28042" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Yasuhiro Aoki, Hunter Goforth, Rangaprasad Arun Srivatsan, and Simon Lucey. Pointnetlk: Robust & efficient point cloud registration using pointnet. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 281 + ], + "type": "text", + "content": "[3] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 316 + ], + "type": "text", + "content": "[4] Xuelin Chen, Baoquan Chen, and Niloy J Mitra. Unpaired point cloud completion on real scans using adversarial training. arXiv preprint arXiv:1904.00069, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "type": "text", + "content": "[5] Ta-Ying Cheng, Qingyong Hu, Qian Xie, Niki Trigoni, and Andrew Markham. Meta-sampler: Almost-universal yet task-oriented sampling for point clouds. In Computer Vision - ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, page 694-710, Berlin, Heidelberg, 2022. Springer-Verlag. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 450 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 450 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 495 + ], + "type": "text", + "content": "[7] Oren Dovrat, Itai Lang, and Shai Avidan. Learning to sample. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 287, + 541 + ], + "type": "text", + "content": "[8] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning. PMLR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 587 + ], + "type": "text", + "content": "[9] Jinyang Guo, Jiaheng Liu, and Dong Xu. Jointpruning: Pruning networks along multiple dimensions for efficient point cloud processing. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 632 + ], + "type": "text", + "content": "[10] Jinyang Guo, Jiaheng Liu, and Dong Xu. 3d-pruning: A model compression framework for efficient 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(12):8717-8729, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[11] Jinyang Guo, Dong Xu, and Wanli Ouyang. Multidimensional pruning and its extension: A unified framework for model compression. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[13] Asako Kanezaki, Yasuyuki Matsushita, and Yoshifumi Nishida. Rotationnet: Joint object categorization and pose estimation using multiviews from unsupervised viewpoints. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[15] Itai Lang, Asaf Manor, and Shai Avidan. Samplenet: Differentiable point cloud sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[16] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeruIPS, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 296 + ], + "type": "text", + "content": "[17] Jiaheng Liu and Dong Xu. Geometrhythm-net: A strong two-stream baseline for 3d action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 31 (12):4711-4721, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 352 + ], + "type": "text", + "content": "[18] Jiaheng Liu, Yudong Wu, Yichao Wu, Chuming Li, Xiaolin Hu, Ding Liang, and Mengyu Wang. Dam: discrepancy alignment metric for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3814-3823, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 387 + ], + "type": "text", + "content": "[19] Jiaheng Liu, Jinyang Guo, and Dong Xu. Apsnet: Towards adaptive point sampling for efficient 3d action recognition. IEEE Transactions on Image Processing, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[20] Jiaheng Liu, Jinyang Guo, and Dong Xu. Geometriontransformer: An end-to-end framework for 3d action recognition. IEEE Transactions on Multimedia, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[21] Jiaheng Liu, Tong He, Honghui Yang, Rui Su, Jiayi Tian, Junran Wu, Hongcheng Guo, Ke Xu, and Wanli Ouyang. 3d-queryis: A query-based framework for 3d instance segmentation. arXiv preprint arXiv:2211.09375, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[22] Jiaheng Liu, Haoyu Qin, Yichao Wu, Jinyang Guo, Ding Liang, and Ke Xu. Coupleface: Relation matters for face recognition distillation. In European Conference on Computer Vision, pages 683-700. Springer, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 555 + ], + "type": "text", + "content": "[23] Jiaheng Liu, Zhipeng Yu, Haoyu Qin, Yichao Wu, Ding Liang, Gangming Zhao, and Ke Xu. Oneface: one threshold for all. In European Conference on Computer Vision, pages 545-561. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[24] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[25] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In ICCV, pages 5007-5016, 2019. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[27] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28043" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2017. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[28] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[29] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 208 + ], + "type": "text", + "content": "[30] Yue Qian, Junhui Hou, Qijian Zhang, Yiming Zeng, Sam Kwong, and Ying He. Mops-net: A matrix optimization-driven network for task-oriented 3d point cloud downsampling. arXiv preprint arXiv:2005.00383, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 274 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 331 + ], + "type": "text", + "content": "[32] Muhammad Sarmad, Hyunjoo Jenny Lee, and Young Min Kim. Rl-gan-net: A reinforcement learning agent controlled gan network for real-time point cloud shape completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 376 + ], + "type": "text", + "content": "[33] Vinit Sarode, Xueqian Li, Hunter Goforth, Yasuhiro Aoki, Rangaprasad Arun Srivatsan, Simon Lucey, and Howie Choset. Pcrnet: Point cloud registration network using point-net encoding. arXiv preprint arXiv:1908.07906, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 410 + ], + "type": "text", + "content": "[34] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 456 + ], + "type": "text", + "content": "[35] Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. In Proceedings of the IEEE international conference on computer vision, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 511 + ], + "type": "text", + "content": "[36] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 288, + 567 + ], + "type": "text", + "content": "[37] Haodi Wang, Kai Dong, Zhilei Zhu, Haotong Qin, Aishan Liu, Xiaolin Fang, Jiakai Wang, and Xianglong Liu. Transferable multimodal attack on vision-language pre-training models. In 2024 IEEE Symposium on Security and Privacy (SP), pages 102-102. IEEE Computer Society, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 634 + ], + "type": "text", + "content": "[38] Jiakai Wang, Aishan Liu, Zixin Yin, Shunchang Liu, Shiyu Tang, and Xianglong Liu. Dual attention suppression attack: Generate adversarial camouflage in physical world. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 8565-8574. Computer Vision Foundation / IEEE, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[39] Jiakai Wang, Aishan Liu, Xiao Bai, and Xianglong Liu. Universal adversarial patch attack for automatic checkout using perceptual and attentional bias. IEEE Trans. Image Process., 31:598-611, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[40] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "IEEE conference on computer vision and pattern recognition, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[41] Xinlong Wang, Shu Liu, Xiaoyong Shen, Chunhua Shen, and Jiaya Jia. Associatively segmenting instances and semantics in point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[42] Xu Wang, Yi Jin, Yigang Cen, Tao Wang, Bowen Tang, and Yidong Li. Lightn: Light-weight transformer network for performance-overhead tradeoff in point cloud downsampling. arXiv preprint arXiv:2202.06263, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 230 + ], + "type": "text", + "content": "[43] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (TOG), 38(5):1-12, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 286 + ], + "type": "text", + "content": "[44] C. Wen, B. Yu, and D. Tao. Learnable skeleton-aware 3d point cloud sampling. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17671-17681, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 319 + ], + "type": "text", + "content": "[45] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 321, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 376 + ], + "type": "text", + "content": "[46] Yue Wu, Xidao Hu, Yue Zhang, Maoguo Gong, Wenping Ma, and Qiguang Miao. Sacf-net: Skip-attention based correspondence filtering network for point cloud registration. IEEE Transactions on Circuits and Systems for Video Technology, 33(8):3585-3595, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 421 + ], + "type": "text", + "content": "[47] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In CVPR, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[48] Ge Yang, Chao Zhang, Ling Gao, Yufei Guo, and Jinyang Guo. Domain adaptive channel pruning. *Electronics*, 13(5), 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 511 + ], + "type": "text", + "content": "[49] Honghui Yang, Tong He, Jiaheng Liu, Hua Chen, Boxi Wu, Binbin Lin, Xiaofei He, and Wanli Ouyang. Gd-mae: generative decoder for mae pre-training on lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9403-9414, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 567 + ], + "type": "text", + "content": "[50] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3323–3332, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[51] Yang Ye, Xiulong Yang, and Shihao Ji. Apsnet: Attention based point cloud sampling. arXiv preprint arXiv:2210.05638, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 658 + ], + "type": "text", + "content": "[52] Licheng Yu, Zhe Lin, Xiaohui Shen, Jimei Yang, Xin Lu, Mohit Bansal, and Tamara L Berg. Mattnet: Modular attention network for referring expression comprehension. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 714 + ], + "type": "text", + "content": "[53] Zhipeng Yu, Jiaheng Liu, Haoyu Qin, Yichao Wu, Kun Hu, Jiayi Tian, and Ding Liang. Icd-face: Intra-class compactness distillation for face recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21042-21052, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28044" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 284 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[54] Wentao Yuan, Tejas Khot, David Held, Christoph Mertz, and Martial Hebert. Pcn: Point completion network. In 2018 international conference on 3D vision (3DV). IEEE, 2018. 2,6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 173 + ], + "type": "text", + "content": "[55] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "text", + "content": "[56] Lichen Zhao, Jinyang Guo, Dong Xu, and Lu Sheng. Transformer3d-det: Improving 3d object detection by vote refinement. IEEE Transactions on Circuits and Systems for Video Technology, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 240 + ], + "type": "text", + "content": "[57] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 287, + 284 + ], + "type": "text", + "content": "[58] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyao Zeng, Shanghang Zhang, and Peng Gao. Pointclip v2: Adapting clip for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28045" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_content_list.json b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d6b7482a856c347aac3e8208c6f3850243824484 --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_content_list.json @@ -0,0 +1,1305 @@ +[ + { + "type": "text", + "text": "LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content", + "text_level": 1, + "bbox": [ + 86, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qihao Zhao $^{1,2,*}$ , Yalun Dai $^{3,*}$ , Hao Li $^{4}$ , Wei Hu $^{1}$ , Fan Zhang $^{1\\dagger}$ , Jun Liu $^{2}$ $^{1}$ Beijing University of Chemical Technology, China \n $^{2}$ Singapore University of Technology and Design, Singapore \n $^{3}$ Nanyang Technological University, Singapore \n $^{4}$ Northwestern Polytechnical University, China", + "bbox": [ + 207, + 179, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long-tail recognition is challenging because it requires the model to learn good representations from tail categories and address imbalances across all categories. In this paper, we propose a novel generative and fine-tuning framework, LTGC, to handle long-tail recognition via leveraging generated content. Firstly, inspired by the rich implicit knowledge in large-scale models (e.g., large language models, LLMs), LTGC leverages the power of these models to parse and reason over the original tail data to produce diverse tail-class content. We then propose several novel designs for LTGC to ensure the quality of the generated data and to efficiently fine-tune the model using both the generated and original data. The visualization demonstrates the effectiveness of the generation module in LTGC, which produces accurate and diverse tail data. Additionally, the experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tailed benchmarks.", + "bbox": [ + 75, + 335, + 473, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 619, + 209, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the real world, data often exhibits a long-tailed distribution, posing significant challenges for computer vision recognition [44, 54]. These challenges include (1) Class Imbalance: within the dataset, some classes (termed \"head\" classes) are abundantly represented, while others (termed \"tail\" classes) have few samples. This imbalanced distribution during training may cause the model to attend more on the head classes, neglecting the tail classes [42]. (2) Tail Data Scarcity: Data scarcity refers to tail classes having extremely limited samples, which lack diversity and are insufficient to effectively train a model [8, 24]. It prevents a model's ability to learn the feature invariant, which is necessary to recognize these categories correctly [35].", + "bbox": [ + 75, + 645, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31f361492f3359d2f385fee6d7a43ccc82031d239f95078b919e7867dc0958ff.jpg", + "image_caption": [ + "Figure 1. Top: Here, our LMMs use ChatGPT. For the Trogon Rufus category, when we asked ChatGPT, \"What species is in the picture?\" we did not get the expected answer due to the complexity of the question. Middle and Down: In contrast, when we asked some easy questions for the ChatGPT, \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus.\" It could accurately answer these questions." + ], + "image_footnote": [], + "bbox": [ + 509, + 301, + 883, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, numerous approaches have emerged within the field of long-tailed recognition, such as re-sampling [5, 19], loss re-weighting [1, 6, 12, 23, 26, 27, 29, 39, 48], ensemble learning [42, 47, 56], decoupling [20], and contrastive learning [11, 25, 40, 59]. Their primary objective is to balance the decision boundaries and parameter weights of the model to improve long-tail recognition.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 851, + 205, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†The Corresponding author is with the College of Information Science and Technology and the Interdisciplinary Research Center for Artificial Intelligence, Beijing University of Chemical Technology, China", + "bbox": [ + 76, + 863, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19510", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, these methods encounter bottlenecks due to the scarcity and limited diversity features of tail-class data [24]. To obtain more diverse representations for tail classes, two types of methods appear: Some methods increase tail-class diversity through data augmentation [9, 24, 52, 55]. Others increase tail-class diversity by transferring features from related classes [8] or large pre-trained models (e.g., CLIP [32]). These transfer learning methods [13, 28, 36] based on CLIP have recently shown great potential in boosting long-tail recognition. The LPT [13] fine-tunes pre-trained models to adapt to target datasets. The VL-LTR [36] leverages text features from CLIP to augment the learning of image features. However, these methods find it hard to obtain the correct and desired knowledge for tail categories. For example, in the real-world data iNauralist 2018 [17], the category Aquilegia Pubescens is characterized only by white or pale yellow colors and spurred shapes. If we transfer similar category features or blindly augment semantic information, such as red color semantics, this category will become more confusing.", + "bbox": [ + 75, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Large Language Models (LLMs) (e.g., ChatGPT) and Large Multimodal Models (LMMs) (e.g., GPT-4V (ChatGPT with Vision [50]) and Minigpt4 [7, 58]) due to their wealth of potential knowledge, have been leveraged for a variety of downstream tasks, such as robot task plans [34], open-set object recognition [30], automated robot learning [43] and various visual reasoning tasks [50, 51, 60]. However, due to the bias of imbalanced training data, LMMs perform poorly on some complex tasks [10]. We also find this dilemma in long-tail recognition as shown in Fig. 1. When we asked LMMs about an image in the category of Trogon rufus: \"What species is in the image?\", we did not get the expected answer. In contrast, when we ask some easy questions about this image: \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus\", we get the desired answer. This suggests that although large models underperform in long-tail recognition (e.g., it is difficult to align complex image features with labels.), they still contain desired potential knowledge (e.g., correctly describing images and providing the knowledge of species). Inspired by this, we aim to leverage the abundant potential knowledge of several large models to address the challenges of long-tailed data scarcity and perform long-tailed recognition well. Nevertheless, effectively leveraging large models to encounter these challenges is not trivial: (1) The potential nature of knowledge within large models makes it difficult to extract the desired knowledge that facilitates long-tail recognition tasks. (2) Due to the domain gap between the generated data and the original data [38], effectively using these hybrid data remains an open issue.", + "bbox": [ + 75, + 397, + 472, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel Long-Tail recognition framework via Generated Content, denoted as LTGC,", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which is illustrated in Fig. 2. Our LTGC aims to leverage the capabilities of large models for generating explicitly diverse content [15] tailored to the long-tail classes and incorporate novel designs to enhance long-tail recognition. Firstly, motivated by the wealth of potential knowledge in large models and the fact that texts are more controllable, we employ them to expand the tail classes. Specifically, to produce more diverse and controllable tail data, we follow the rule of analyzing existing features before generating absent features. To achieve this, LTGC utilizes LMMs to analyze images already in the tail classes to obtain the existing tail-class descriptions list. Then LTGC inputs the existing descriptions list to LLMs to obtain the desired extended tail-class descriptions list for images that are absent from the existing tail classes. Secondly, LTGC utilizes the text-to-image (T2I) model (e.g., DALL-E [33]) to transform these textual descriptions into images. Moreover, inspired by the benefits of chain-of-thought [46], we propose a self-reflection and iterative evaluation module for this process to ensure the diversity and quality of the generated content. Finally, inspired by the advantages of Mixup [52] in merging different image domains, we propose the BalanceMix module to address domain shifts of generated images for the fine-tuning process.", + "bbox": [ + 496, + 90, + 893, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are as follows:", + "bbox": [ + 521, + 454, + 890, + 468 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. In the first time, we propose a novel framework via generated content, LTGC, which leverages the power of large models to address long-tail recognition challenges.", + "2. We design a series of novel modules to tackle the tail-class image scarcity problem and design the BalanceMix module to efficiently fine-tune the model using the generated data and the original data.", + "3. Experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tail benchmarks. Also, the visualization illustrates the diversity and controllability of our generated tail images." + ], + "bbox": [ + 500, + 469, + 890, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 648, + 640, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-Tail recognition. Many approaches have been proposed in recent years to address the long-tail recognition challenge, such as resampling [5, 19], loss rebalancing [1, 6, 12, 23, 26, 27, 29, 39, 48] to increase the model's focus on the tail categories, ensemble learning [42, 53, 56] to improve the ability to recognize tail samples by enhancing model performance, contrast learning [11, 21, 25, 40, 49, 59] and calibration learning [9, 57] to improve the model's decision boundaries, but due to the scarcity of data on the tails, the model still doesn't have enough features to learn class representations for long-tail recognition. Recently, data augmentation-based methods [8, 9, 24] proposed to handle scarcity-tailed data by augmenting from related features. However, these methods find it difficult to obtain the correct knowledge for tail classes. Other approaches", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[13, 36] have achieved good results by transfer learning [37] the feature of the visual-language pre-training model (e.g., CLIP [32]) for long-tailed recognition.", + "bbox": [ + 75, + 90, + 467, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the previous methods, in this paper, from a novel perspective, we explore a generative and fine-tuning framework to handle long-tail recognition by leveraging the rich potential knowledge of large models. Specifically, we design several novel modules to leverage the correct knowledge for generating tail data and perform efficient fine-tuning with the generated data to boost long-tail recognition.", + "bbox": [ + 75, + 136, + 467, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Models. Recently, several large models have emerged, including large language models (e.g., ChatGPT [50]), large multi-modal models (e.g., GPT-4 (ision) [50], MiniGPT [58]) and large generated models (e.g., DALLE [33]). These expansive models are repositories of extensive knowledge [15] and have undergone scrutiny in diverse applications, such as robot task plans [34], open-set object recognition [30], contextual object detection [51], video generation [22], and few-shot segmentation [60]. In this work, we leverage the rich potential knowledge of these large models to handle the long-tail recognition challenge.", + "bbox": [ + 75, + 257, + 467, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 436, + 166, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To perform long-tail recognition well, the challenge is making the model learn diverse representations from tail categories and address imbalances across all categories. In this paper, we propose a novel long-tail recognition framework, LTGC, to handle these challenges. As depicted in Fig. 2, LTGC employ potential knowledge from various off-the-shelf large-scale models to generate and iteratively assess the quality and diversity of tail classes (Sec. 3.1). Furthermore, LTGC proposes the BalanceMix module to facilitating the fine-tuning process with the generated and original datasets (Sec. 3.2). These modules are described in detail below.", + "bbox": [ + 75, + 460, + 467, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diverse Tail Images Generation", + "text_level": 1, + "bbox": [ + 76, + 650, + 354, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To learn diversity representations from tail categories, the previous methods have used data augmentation [9, 24] or transfer learning [13, 36]. However, these works find it difficult to obtain the correct and desired diversity knowledge for tail categories. Inspired by the common-sense knowledge in the LLMs and the fact that textual descriptions are more controllable [33, 50], LTGC takes advantage of these to control the detail and diversity of the generated tail classes. Firstly, to better generate diverse images and control the image detail, LTGC aims to generate images that are absent in the original tail data and represent these images in textual form. Specifically, LTGC employs LMMs to analyze the original tail data and obtain the existing tail-class descriptions list. Then LTGC leverages the common-sense knowledge of LLMs to obtain extended tail-class descrip", + "bbox": [ + 75, + 674, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions based on the existing tail-class descriptions list. Finally, as images are more suitable as training data for long-tail recognition tasks, we utilize the text-to-image module to generate diverse images based on these tail-class descriptions.", + "bbox": [ + 498, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Obtaining Existing Tail-class Descriptions List", + "text_level": 1, + "bbox": [ + 498, + 203, + 875, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to ensure the diversity of tail-class images, we first analyze the feature information of the original tail-class images before generating new ones. It guarantees that the content of the generated images is distinct from that of the existing tail-class images. Moreover, due to textual descriptions being more controllable [33], we utilize the textual descriptions to control the detail and diversity of the generated images. To achieve this, our LTGC employs Language Model Multimodals (LMMs), such as GPT-4V (Vision) [50], to analyze the feature information and extract textual descriptions of existing tail classes. During this process, the textual responses from LMMs could be varied and sometimes redundant, which may impede the generation of the desired images. Therefore, we employ textual templates to constrain the responses of LMMs, aiming to unify the textual description formats. Inspired by an image of an object that can be fully described or generated by its class and a list of features [4, 41], we design the textual template to include the given class and its features. Furthermore, the introduction of variations in the scenes plays a crucial role in enhancing the model's ability to generalize [2], a problem that's particularly acute in long-tailed datasets [35]. Therefore, the textual information of scenes is also important for image generation.", + "bbox": [ + 496, + 233, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we design the LMMs' response Template 1 for a given class $y$ as follows: \"A photo of the class [y], {with distinctive features}{in specific scenes}.\", which include the given class, its distinctive features, and specific scenes or environments. With the response template, for a given $M_y$ number class $y$ , we sequentially feed these tail-class images into LMMs along with an instructing [Prompt 1] to analyze the features of these images: \"Please use the Template 1 to briefly describe the image of the class [y].\" This process is illustrated in Fig. 3, where $y$ is the given class label. As shown, by formulating the description using the template, LMMs would automatically replace the class, features, and background in the response. After performing this process for all images of each class, we compile a list of text descriptions corresponding to each class. Then we intend to extend the tail-class descriptions list with the existing tail-class descriptions list to generate the images lacking in the tail classes.", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19512", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/120199c534a605dc42736c8f510eb18b7a984b55c76a75c0830643deeef4af6d.jpg", + "image_caption": [ + "Figure 2. Overall framework of LTGC. LTGC first employs LMMs to analyze the existing tail data to obtain the existing tail-class descriptions list. Then it inputs the list into LLMs to analyze the absent features of the tail classes and employs the T2I model to generate diverse images. Moreover, our designed self-reflection and iterative evaluation modules ensure the diversity and quality of the tail data. Finally, LTGC employs the BalanceMix module to fine-tune the CLIP's visual encoder with the extended and original data." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 851, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Obtaining Extended Tail-class Descriptions List", + "text_level": 1, + "bbox": [ + 75, + 476, + 464, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we aim to analyze features missing from tail classes and enrich the descriptions of these classes based on the existing tail-class descriptions list. To accomplish this, inspired by the rich common-sense knowledge of Large Language Models (LLMs), such as ChatGPT [50], we extend the tail-class descriptions list through a two-step process: 1) Inputting the existing descriptions list into LLMs, and 2) Designing the Prompt 2 to guide LLMs in generating the desired descriptions for images that are absent in the given tail class $y$ : \"Besides these descriptions mentioned above, please use the Template 1 to list other possible {distinctive features} and {specific scenes} for the class [y],\" which is illustrated in Fig. 4. For each class, we repeat the above two-step process, and then we obtain the tail-class descriptions list for all tail classes.", + "bbox": [ + 73, + 505, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, in order for these generated descriptions to better complement each tail class, we encourage LLMs to generate descriptions of sufficient number and diversity. To meet this goal, we introduce a self-reflection module in this process, aiming to guide LMMs in rethinking if there are any features or scenes that are missed or repeated. It includes two key designs: a number-checking module and a repetition-checking module. (1) The number-checking module aims to guide LMMs in rethinking whether there", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are other missing features or scenarios. To achieve this, after posing the initial [Prompt 2] for class $y$ , we update the extended descriptions list and re-ask LLMs the [Prompt 2] question, incorporating the newly acquired list. For each class $y$ , this iterative process of number-checking continues until a maximum number $K_y$ of the tail class is achieved, where $K_y = M_y + N_y$ and $N_y$ is the number of generated descriptions for class $y$ . (2) The repetition-checking module aims to guide LMMs in rethinking if there are other features or scenes that are repeated at the end of the number-checking iteration. Specifically, we input the extended descriptions list and the following [Prompt 3] of each class $y$ for LLMs' repetition checking: \"Please exclude any repetitive {distinctive features} and {specific scenes} for class $[y]$ in this descriptions list.\" After that, LLM will filter out the ones with repeated features based on this prompt and return a new list of descriptions. Through the implementation of this two-step process and LLMs' self-reflection module, we have obtained a reduction in repeated descriptions and an increase in the diversity of the tail-class descriptions for each class.", + "bbox": [ + 496, + 476, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3 Transform Descriptions to Images", + "text_level": 1, + "bbox": [ + 500, + 830, + 790, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Above, we obtain the extended tail-class descriptions list for each class $y$ , denoted as $L_{y}$ . As the images are better adapted to perform visual recognition tasks, in this section,", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/35a579247c0af3a15735f4b4fac9720d328125cbe449dee2933bab4c55392c03.jpg", + "image_caption": [ + "Figure 3. Example of the instruction for LMMs. When both images from tail classes and textual templates are input into LMMs, textual descriptions corresponding to the images can be obtained. By repeatedly performing this operation on the training data, we convert abstract image descriptions into concrete textual descriptions. Finally, we acquire the current textual descriptions list corresponding to each class." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 460, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we aim to leverage the image-generative ability of the text-to-image (T2I) method to generate the images from the tail-class descriptions list. In detail, we employ T2I to generate images based on the descriptions list, denoted as:", + "bbox": [ + 75, + 401, + 470, + 464 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b4d90bf01494086fb2f59df7ac904efe708fda5edd98f816352bfe738aa1a82.jpg", + "image_caption": [ + "Figure 4. Example of the instruction for LLMs. LTGC inputs the existing textual descriptions list to LLMs, which continually extends it with new distinctive features and scene information. During multiple iterations, LTGC generates a new extended textual descriptions list for each class." + ], + "image_footnote": [], + "bbox": [ + 86, + 474, + 460, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ni _ {n} ^ {y} = \\mathrm {T 2 I} \\left(d _ {n} ^ {y}\\right), \\text {w h e r e} n \\in \\{1, \\dots , N \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 739, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n \\in N$ denotes the $n$ -th generated descriptions for class $y$ , and $i_n^y$ denotes the image generated by T2I model based on $d_n^y$ . However, some generated images may not be of sufficiently high quality to accurately represent their desired classes, as the lower-quality images. As shown in Fig. 7, lower-quality images will output poorly distinctive features, resulting in more confusion for the class. Therefore, using these generated images directly in training may result in disrupting the model's prediction of the tail classes.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6becc86c723c3fcee6af0bc5992fcd5c76c29def6417ebc0a1f4f706c92e7d2d.jpg", + "image_caption": [ + "Figure 5. Illustration of the proposed iterative evaluation module framework. This module detects lower-quality images through the similarity score $S$ computed by images and their corresponding class feature template. Then the textual descriptions corresponding to lower-quality images are re-input into LLMs for refinement. Finally, the refined textural descriptions are fed into the T2I model for regeneration." + ], + "image_footnote": [], + "bbox": [ + 506, + 85, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address the challenge of lower-quality images that fail to represent their desired classes accurately, we aim to automatically detect these images, refine their textual descriptions, and regenerate them accordingly. To achieve this, drawing inspiration from the idea that humans improve their understanding through peer feedback, we investigate the possibility of a large model also enhancing its output by integrating feedback from another fundamental model. Therefore, we propose a iterative evaluation module, as illustrated in Fig. 5. This module utilizes the fundamental model, CLIP, to serve as the 'feedback giver' selected for its proficiency in understanding images and effectively connecting them with textual contexts [32]. It provides feedback about the lower-quality images to LLMs, which assists in refining their corresponding descriptions.", + "bbox": [ + 496, + 383, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, to facilitate CLIP in grasping the represented features in images, a straightforward way is to check the degree of match between the image and its class template: \"A photo of a [y].\" However, due to the fact that CLIP's text encoder may contain less feature knowledge of the tail classes (e.g., as presented on the iNaturalist 2018 dataset with poor zero-shot recognition performance [13]), we aim to design a new textual feature template for each class, highlighting its unique features. Then CLIP employs it to determine the correspondence between the template and the image. Drawing inspiration from the chain-of-thought [30, 46] concept, which suggests that LMMs perform better when provided with additional clues, we further guide LMMs in summarizing the most distinctive features of each class. Specifically, after obtaining the extended tail-class descriptions list, we obtain the class feature template $C_y$ that contains the most distinctive features by the following [Prompt 4] and Template 2 for the given class $y$ : [Prompt 4]: \"Please use", + "bbox": [ + 496, + 613, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19514", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Template 2 to summarize the most distinctive features of class $[y]$ \". \" Template 2: \"A photo of the class [y] with {feature 1} {feature 2} {...}. \" This process could guide the LMMs to rethink and summarize the most distinctive features and produce the class feature template $C_y$ for each class $y$ . Furthermore, since we focus on checking whether the feature information of an image could represent its desired class, we do not include scene information in the class feature template. Then we introduce the details of the iterative evaluation module.", + "bbox": [ + 76, + 90, + 467, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 5, our iterative evaluation module conducts the following three steps iteratively: (1) Detection. To identify lower-quality images, we first employ utilize the strong capability of aligning images with text of CLIP to match the feature of generated images $i_n^y$ and its class feature template $C_y$ by cosine similarity metric as follows:", + "bbox": [ + 76, + 258, + 467, + 349 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS = \\operatorname {E n c o d e r} _ {\\text {v i s}} \\left(i _ {n} ^ {y}\\right) \\cdot \\operatorname {E n c o d e r} _ {\\text {t e x t}} \\left(C _ {y}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 368, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathsf{Encoder}_{\\mathrm{vis}}$ denotes the CLIP's visual encoder, and $\\mathsf{Encoder}_{\\mathrm{text}}$ denotes the text encoder of the CLIP. The image $i_n^y$ is detected as lower quality and filtered out if its similarity score $S$ is below a threshold $\\mu$ . (2) Refinement. Next, also inspired by chain-of-thought [46], we propose the following prompt for refine its descriptions. Specifically, if $i_n^y$ is identified as a lower quality image, we prompt LLMs to refine its corresponding description $d_n^y$ to more accurately represent the intended class $y$ , drawing on feedback provided by CLIP, i.e., [Prompt 5]: \"This description $d_n^y$ doesn't seem to be representative of the class $[y]$ . Could you refine it to enhance the distinctive features of class $[y]?\"\" (3)$", + "bbox": [ + 76, + 393, + 467, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Re-generation. Finally, the image $i_n^y$ is regenerated by the T2I model according to the improved textual description.", + "bbox": [ + 76, + 606, + 467, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By iteratively applying these three stages, we employ CLIP in each cycle to furnish precise feedback to the LLMs, thereby steering the refinement of descriptions to better align with our desired direction. These refined descriptions ensure the production of images that more accurately embody the characteristics of each class.", + "bbox": [ + 76, + 637, + 467, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2. BalanceMix", + "text_level": 1, + "bbox": [ + 76, + 739, + 205, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With generated tailed images, the final problem is how to efficiently use these generated data and original data to perform long-tailed recognition well. Due to the domain gap between the generated data and the original long-tail data [38], we propose a method named BalanceMix to handle this challenge. We first define the original data and generated data as $\\mathcal{D}_o$ and $\\mathcal{D}_g$ . Then BalanceMix balance-sample [5] an image $x_i$ from $\\mathcal{D}_o$ and sample an image $x_j$ from $\\mathcal{D}_g$ . Meanwhile, it mixes the images $x_i$ and $x_j$ and their corre", + "bbox": [ + 76, + 763, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sponding labels, denoted as:", + "bbox": [ + 500, + 90, + 687, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {x} = \\lambda \\odot x _ {i} + (\\mathbf {1} - \\lambda) \\odot x _ {j}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 119, + 890, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {y} = \\lambda \\odot y _ {i} + (\\mathbf {1} - \\lambda) \\odot y _ {j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 162, + 890, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ in Beta (0,1) distribution. Finally, we fine-tune the CLIP's vision encoder with LORA [18] on all mixed data pairs $(\\widetilde{x},\\widetilde{y})$ for efficient long-tail recognition.", + "bbox": [ + 498, + 186, + 890, + 233 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 247, + 630, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present the experimental results on three widely used datasets in long-tailed recognition, including ImageNet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. Moreover, we undertake ablation studies specifically on the ImageNet-LT and iNaturalist 2018 datasets to gain deeper insights into the performance of our method. The experimental results of the comparison methods are taken from their original paper, and our results are averaged over three experiments.", + "bbox": [ + 498, + 273, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details.", + "text_level": 1, + "bbox": [ + 500, + 404, + 718, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Setup. In all experiments, we evaluate and report top-1 accuracy on their corresponding test set. We also report accuracy on three splits of the classes: Many-shot (more than 100 images), Medium-shot (20 to 100 images), and Few-shot (less than 20 images) [20].", + "bbox": [ + 498, + 428, + 890, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Method Implementation. In our LTGC, we incorporate diverse and specialized knowledge from the off-the-shelf large models. Specifically, for LMM, we use the GPT-4V (ision) [50] version of ChatGPT. For LLM, we use the GPT-4 version of ChatGPT. For T2I, we use DAll-E [33]. For the pre-trained CLIP [32], we use ViT-B/32 [14] for its visual encoder and the architecture introduced in [31] as the text encoder. In LLM's self-reflection module, we set the maximum number $K_{y}$ to 100, 300, and 800 for iNaturalist 2018, ImageNet-LT, and Place-LT, respectively. In the iterative evaluation module, the threshold $\\mu$ is set at 0.8 for ImageNet-LT and Place-LT, and at 0.6 for iNaturalist. In Appendix, we provide a detailed ablation and discussion on the choice of LMM and LLM models, as well as parameters used in the self-reflection process.", + "bbox": [ + 498, + 503, + 890, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with SOTA on Benchmarks", + "text_level": 1, + "bbox": [ + 500, + 739, + 851, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our proposed LTGC model with state-of-the-art (SOTA) methods on three benchmarks, including Imagenet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. To ensure a fair comparison, we primarily focus on methods based on CLIP, as they also leverage knowledge pretrained on large-scale datasets. These baselines include CLIP Zero-Shot [32, 36] and CLIP Finetune [36], as well as CLIP-based long-tail recognition approaches such as VL-LTR [36], LPT [13], and RAC [28]. Additionally,", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we report comparisons with traditional methods (without CLIP) on the challenging and fine-grained, large-scale long-tail dataset, iNaturalist 2018. These have proven the effectiveness of our method in different scenarios.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Imagenet-LT. In Tab. 1, we observe that our LTGC models are superior to other CLIP-based LT methods. For example, our method achieves an overall accuracy of $80.6\\%$ , which is $3.4\\%$ higher than the existing SOTA method VT-LTR[36]. Moreover, the overall accuracy of our method marginally surpasses the results on the full ImageNet (i.e., $80\\%$ [36]).", + "bbox": [ + 75, + 151, + 470, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Places-LT. Tab. 1 shows that compared to other CLIP variant methods, LTGC achieves an overall accuracy of $54.1\\%$ and an few-shot accuracy of $52.1\\%$ , surpassing the LPT [13] by $4.0\\%$ and $5.2\\%$ , respectively. Even compared with VL-LTR [36] and RAC [28], which have extra data in training and testing, our LTGC achieves remarkable results.", + "bbox": [ + 75, + 257, + 470, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on iNaturalist 2018. In this part, we employ LTGC on a large and fine-grained dataset, iNaturalist 2018. Tab. 2 presents the quantitative results. LTGC leverages the rich knowledge of LLMs and significantly outperforms traditional deep-learning approaches for long-tail recognition. In addition, LTGC attains an overall accuracy of $82.5\\%$ and a few-shot accuracy of $82.6\\%$ , outperforming all existing SOTA methods based on CLIP. In particular, LTGC also surpasses the retrieval augmented method, RAC [28], by $2.3\\%$ .", + "bbox": [ + 75, + 348, + 470, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Compare with different methods of LMMs.", + "text_level": 1, + "bbox": [ + 76, + 510, + 444, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the experiments, our LTGC employs GPT-4V (ision) to describe a given image to obtain a text-based feature description. A straightforward way for long-tail recognition is by querying LMMs for the category of the given image. To evaluate the ability of these methods to recognize the long tail images, we constructed the following baselines of LMMs for comparison: MiniGPT4 [58], MiniGPT4-v2 [7], LENS [3], and GPT-4V (ision) [50]. To perform a fair evaluation, we provide the label list [Class 1, Class 2, ..., Class Y] for each dataset, where $Y$ is the number of classes.", + "bbox": [ + 75, + 534, + 468, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results presented in Tab. 3, show that our method significantly outperforms the baselines that directly query the LMMs. More implementation details are discussed in Appendix.", + "bbox": [ + 75, + 685, + 468, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis and Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 756, + 333, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the BalanceMix module. To evaluate the effectiveness of the two proposed designs, we conduct experiments on three different variants: 1) w/o BalanceMix: This variant is directly fine-tuned using the generated data and the original training data. 2) Balanced Sample: This variant utilizes generated data to perform balanced sampling [45] for training. 3) Mixup: This variant employs generated data with Mixup [52] without performing balanced", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9279d8df0cdc4f4bda329cfd62531fb7fba979f52c065db279a84268146282bf.jpg", + "table_caption": [ + "Table 1. Comparison with SOTA methods on ImageNet-LT and Places-LT." + ], + "table_footnote": [], + "table_body": "
DatasetImageNet-LTPlaces-LT
FewAllFewAll
CLIP Zero [36]58.659.840.138.0
CLIP Finetune [36]34.560.522.739.7
VL-LTR [36]59.377.242.050.1
RAC [28]--41.847.2
LPT [13]--46.950.1
LTGC(Ours)70.580.652.154.1
", + "bbox": [ + 545, + 128, + 846, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/82b6585ea17f2b89f476fbc5d7941a33bdf9ece5df9c5a5c5e799e40ced73b2b.jpg", + "table_caption": [ + "Table 2. Comparison with SOTA methods on iNaturalist 2018." + ], + "table_footnote": [], + "table_body": "
MethodManyMediumFewAll
Softmax74.766.360.064.7
LADE [16]64.447.734.352.3
RIDE [42]71.570.071.671.8
PaCo [11]69.573.473.073.0
MDCS [56]76.575.575.275.6
CLIP Zero [36]6.13.32.93.4
CLIP Finetune [36]76.674.170.272.6
VL-LTR [36]---76.8
RAC [28]75.980.581.080.2
LPT [13]--79.376.1
LTGC(Ours)77.583.982.682.5
", + "bbox": [ + 534, + 276, + 859, + 429 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ea806ac53459be876b54ed4ea047c73d5519769d4b212e5cce34d3356639770.jpg", + "table_caption": [ + "Table 3. Comparison with different LMMs' methods on ImageNet-LT and iNaturalist2018." + ], + "table_footnote": [], + "table_body": "
MethodImageNet-LTiNaturalist 2018
LENS [3]69.517.4
MiniGPT4 [58]60.420.9
MiniGPT4-v2 [7]68.527.1
GPT-472.164.3
Ours80.682.5
", + "bbox": [ + 540, + 486, + 851, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sampling. The results are shown in Tab. 4, demonstrating that our method significantly outperforms the other variants. Although our self-reflection and iterative evaluation modules already ensure the diversity and high quality of the generated images, there still exists a domain gap between the generated and original data [38]. This domain gap could exacerbate the long-tail effect on the test set (discussed in Appendix). A simple balanced sample [45] approach fails to solve the class imbalance caused by the mixing of original and generated data, while the Mixup [52] alone cannot address inter-class imbalances. To this end, our BalanceMix module combines the strengths of both methods, making the generated data well-suited for our framework.", + "bbox": [ + 496, + 594, + 890, + 790 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d64b6641d6ff3b84e8f994feb887d0bf47939fbd636d2eba5cd866c5a57f553c.jpg", + "table_caption": [ + "Table 4. Evaluation on the effectiveness of the BalanceMix." + ], + "table_footnote": [], + "table_body": "
MethodImageNet-LTiNaturalist
w/o BalanceMix58.369.5
Balanced sample [45]63.973.8
Mixup [52]73.475.2
Ours80.682.5
", + "bbox": [ + 542, + 834, + 846, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98e5a1c89b9bc5f0305aa58c42bcb82eeea9a12cfe7f73b130495d906cc0d726.jpg", + "image_caption": [ + "Figure 6. The visualization of generated images: The template \"A photo of the class [y]\" and LTGC. Each row represents a different class. The four images on the left are generated using the simple template \"A photo of the class [y],\" which results in images with uniform poses and plain backgrounds. The four images on the right are from the proposed LTGC and demonstrate the diversity of classes." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 849, + 305 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4ca9b59496760d42a9cb6e31a6e56464d94df592570f291ef5e0c8367d983df.jpg", + "image_caption": [ + "Figure 7. The visualization of the images generated before and after passing the iterative evaluation module. The top row displays images that were filtered out, while the bottom row shows images regenerated by T2I after refining their corresponding descriptions. More visualizations are in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 81, + 383, + 467, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 693, + 215, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Generated Images: Template \"A photo of the class [y]\" vs our LTGC. Furthermore, we compare the images generated by LTGC with a simple prompt, \"A photo of the class [y]\" for the T2I model. As Fig. 6 illustrates, LTGC generates more accurate and diverse images compared to those generated by the simple prompt. For example, in the absence of control over the category description, the T2I model generates red features that do not belong to class Aquilegia Pubescens. In addition, with rich text for image generation, our approach also generates more diverse and accurate images compared to images generated by a simple template.", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization on iterative evaluation module. We compare images before and after refinement using the iterative evaluation module, and the visualization results are shown in Fig. 7. It shows that images before refinement often possess ambiguous semantic information, and the features of the corresponding classes are not distinct. However, after the refinement process, the quality of the images is substantially enhanced, and the distinctive features of the corresponding classes become more pronounced.", + "bbox": [ + 498, + 387, + 890, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 535, + 617, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel generative and fine-tuning framework, named LTGC, to address the challenge of long-tail recognition. LTGC leverages the abundant potential knowledge embedded in large-scale models to generate diverse data for tail categories. The framework incorporates innovative designs to ensure the quality of the generated data and to fine-tune the model efficiently using both the generated and original data. The experimental results indicate that LTGC surpasses existing state-of-the-art methods on well-known long-tail benchmarks. In the future, we will explore the robustness of large-scale models for application in other areas, such as long-tail semantic segmentation and dissent detection.", + "bbox": [ + 496, + 560, + 890, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 660, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant No.62271034. This project is also supported by the Ministry of Education, Singapore, under the SUTD Kickstarter Initiative Project (SKI 2021_02_06). Additionally, I would like to express my gratitude to Haoxuan Qu and He Huang for their valuable suggestions.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emanuel Sanchez Aimar, Arvi Jonnarth, Michael Felsberg, and Marco Kuhlmann. Balanced product of calibrated experts for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19967-19977, 2023. 1, 2", + "[2] Martin Arjovsky, León Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 3", + "[3] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 7", + "[4] Michel Besserve, Arash Mehrjou, Rémy Sun, and Bernhard Schölkopf. Counterfactuals uncover the modular structure of deep generative models. In International Conference on Learning Representations, 2019. 3", + "[5] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 2, 6", + "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. arXiv preprint arXiv:1906.07413, 2019. 1, 2", + "[7] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 7", + "[8] Xiaohua Chen, Yucan Zhou, Dayan Wu, Wanqian Zhang, Yu Zhou, Bo Li, and Weiping Wang. Imagine by reasoning: A reasoning-based implicit semantic data augmentation for long-tailed classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 356-364, 2022. 1, 2", + "[9] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 95–110. Springer, 2020. 2, 3", + "[10] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in gpt-4v(ison): Bias and interference challenges, 2023. 2", + "[11] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 2, 7", + "[12] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 2", + "[13] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zuo. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022. 2, 3, 5, 6, 7", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[15] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 2, 3", + "[16] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 7", + "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The naturalist challenge 2017 dataset. CoRR, abs/1707.06642, 2017. 2, 6", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[19] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5): 429-449, 2002. 1, 2", + "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 6", + "[21] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 2", + "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3", + "[23] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE transactions on neural networks and learning systems, 29(8):3573-3587, 2017. 1, 2", + "[24] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1, 2, 3", + "[25] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 2", + "[26] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19518", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transactions on pattern analysis and machine intelligence, 2018. 1, 2", + "[27] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 6", + "[28] Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, and Anton van den Hengel. Retrieval augmented classification for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6959-6969, 2022. 2, 6, 7", + "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 1, 2", + "[30] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. arXiv preprint arXiv:2309.12780, 2023. 2, 3, 5", + "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6", + "[33] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 6", + "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530, 2023. 2, 3", + "[35] Kaihua Tang, Mingyuan Tao, Jiaxin Qi, Zhenguang Liu, and Hanwang Zhang. Invariant feature learning for generalized long-tailed classification. In European Conference on Computer Vision, pages 709-726. Springer, 2022. 1, 3", + "[36] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 2, 3, 6, 7", + "[37] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3", + "[38] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 6, 7" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 943-952, 2021. 1, 2", + "[40] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 2", + "[41] Tan Wang, Zhongqi Yue, Jianqiang Huang, Qianru Sun, and Hanwang Zhang. Self-supervised learning disentangled group representation as feature. Advances in Neural Information Processing Systems, 34:18225-18240, 2021. 3", + "[42] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020. 1, 2, 7", + "[43] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation, 2023. 2", + "[44] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 7032-7042, 2017. 1", + "[45] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in neural information processing systems, 30, 2017. 7", + "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2, 5, 6", + "[47] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 247–263. Springer, 2020. 1", + "[48] Cihang Xie and Alan Yuille. Intriguing properties of adversarial training at scale. arXiv preprint arXiv:1906.03787, 2019. 1, 2", + "[49] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in neural information processing systems, 33:19290-19301, 2020. 2", + "[50] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 2, 3, 4, 6, 7", + "[51] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 2, 3", + "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19519", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yifan Zhang, Bryan Hooi, Lanqing Hong, and Jiashi Feng. Test-agnostic long-tailed recognition by test-time aggregating diverse experts with self-supervision. arXiv preprint arXiv:2107.09249, 2021. 2", + "[54] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[55] Qihao Zhao, Yangyu Huang, Wei Hu, Fan Zhang, and Jun Liu. Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[56] Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, and Jun Liu. Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11597-11608, 2023. 1, 2, 7", + "[57] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16489-16498, 2021. 2", + "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 7", + "[59] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 2", + "[60] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation, 2023. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 570 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19520", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_model.json b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72c00f2c0c119a38f6918ba01b6f1bd94427fccd --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_model.json @@ -0,0 +1,2026 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.18, + 0.761, + 0.27 + ], + "angle": 0, + "content": "Qihao Zhao\\(^{1,2,*}\\), Yalun Dai\\(^{3,*}\\), Hao Li\\(^{4}\\), Wei Hu\\(^{1}\\), Fan Zhang\\(^{1\\dagger}\\), Jun Liu\\(^{2}\\) \n\\(^{1}\\)Beijing University of Chemical Technology, China \n\\(^{2}\\)Singapore University of Technology and Design, Singapore \n\\(^{3}\\)Nanyang Technological University, Singapore \n\\(^{4}\\)Northwestern Polytechnical University, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.474, + 0.594 + ], + "angle": 0, + "content": "Long-tail recognition is challenging because it requires the model to learn good representations from tail categories and address imbalances across all categories. In this paper, we propose a novel generative and fine-tuning framework, LTGC, to handle long-tail recognition via leveraging generated content. Firstly, inspired by the rich implicit knowledge in large-scale models (e.g., large language models, LLMs), LTGC leverages the power of these models to parse and reason over the original tail data to produce diverse tail-class content. We then propose several novel designs for LTGC to ensure the quality of the generated data and to efficiently fine-tune the model using both the generated and original data. The visualization demonstrates the effectiveness of the generation module in LTGC, which produces accurate and diverse tail data. Additionally, the experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tailed benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.621, + 0.21, + 0.637 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.47, + 0.844 + ], + "angle": 0, + "content": "In the real world, data often exhibits a long-tailed distribution, posing significant challenges for computer vision recognition [44, 54]. These challenges include (1) Class Imbalance: within the dataset, some classes (termed \"head\" classes) are abundantly represented, while others (termed \"tail\" classes) have few samples. This imbalanced distribution during training may cause the model to attend more on the head classes, neglecting the tail classes [42]. (2) Tail Data Scarcity: Data scarcity refers to tail classes having extremely limited samples, which lack diversity and are insufficient to effectively train a model [8, 24]. It prevents a model's ability to learn the feature invariant, which is necessary to recognize these categories correctly [35]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.207, + 0.864 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.864, + 0.469, + 0.901 + ], + "angle": 0, + "content": "†The Corresponding author is with the College of Information Science and Technology and the Interdisciplinary Research Center for Artificial Intelligence, Beijing University of Chemical Technology, China" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.852, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.302, + 0.884, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.682, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Figure 1. Top: Here, our LMMs use ChatGPT. For the Trogon Rufus category, when we asked ChatGPT, \"What species is in the picture?\" we did not get the expected answer due to the complexity of the question. Middle and Down: In contrast, when we asked some easy questions for the ChatGPT, \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus.\" It could accurately answer these questions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, numerous approaches have emerged within the field of long-tailed recognition, such as re-sampling [5, 19], loss re-weighting [1, 6, 12, 23, 26, 27, 29, 39, 48], ensemble learning [42, 47, 56], decoupling [20], and contrastive learning [11, 25, 40, 59]. Their primary objective is to balance the decision boundaries and parameter weights of the model to improve long-tail recognition." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "However, these methods encounter bottlenecks due to the scarcity and limited diversity features of tail-class data [24]. To obtain more diverse representations for tail classes, two types of methods appear: Some methods increase tail-class diversity through data augmentation [9, 24, 52, 55]. Others increase tail-class diversity by transferring features from related classes [8] or large pre-trained models (e.g., CLIP [32]). These transfer learning methods [13, 28, 36] based on CLIP have recently shown great potential in boosting long-tail recognition. The LPT [13] fine-tunes pre-trained models to adapt to target datasets. The VL-LTR [36] leverages text features from CLIP to augment the learning of image features. However, these methods find it hard to obtain the correct and desired knowledge for tail categories. For example, in the real-world data iNauralist 2018 [17], the category Aquilegia Pubescens is characterized only by white or pale yellow colors and spurred shapes. If we transfer similar category features or blindly augment semantic information, such as red color semantics, this category will become more confusing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.866 + ], + "angle": 0, + "content": "Recently, Large Language Models (LLMs) (e.g., ChatGPT) and Large Multimodal Models (LMMs) (e.g., GPT-4V (ChatGPT with Vision [50]) and Minigpt4 [7, 58]) due to their wealth of potential knowledge, have been leveraged for a variety of downstream tasks, such as robot task plans [34], open-set object recognition [30], automated robot learning [43] and various visual reasoning tasks [50, 51, 60]. However, due to the bias of imbalanced training data, LMMs perform poorly on some complex tasks [10]. We also find this dilemma in long-tail recognition as shown in Fig. 1. When we asked LMMs about an image in the category of Trogon rufus: \"What species is in the image?\", we did not get the expected answer. In contrast, when we ask some easy questions about this image: \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus\", we get the desired answer. This suggests that although large models underperform in long-tail recognition (e.g., it is difficult to align complex image features with labels.), they still contain desired potential knowledge (e.g., correctly describing images and providing the knowledge of species). Inspired by this, we aim to leverage the abundant potential knowledge of several large models to address the challenges of long-tailed data scarcity and perform long-tailed recognition well. Nevertheless, effectively leveraging large models to encounter these challenges is not trivial: (1) The potential nature of knowledge within large models makes it difficult to extract the desired knowledge that facilitates long-tail recognition tasks. (2) Due to the domain gap between the generated data and the original data [38], effectively using these hybrid data remains an open issue." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To this end, we propose a novel Long-Tail recognition framework via Generated Content, denoted as LTGC," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.454 + ], + "angle": 0, + "content": "which is illustrated in Fig. 2. Our LTGC aims to leverage the capabilities of large models for generating explicitly diverse content [15] tailored to the long-tail classes and incorporate novel designs to enhance long-tail recognition. Firstly, motivated by the wealth of potential knowledge in large models and the fact that texts are more controllable, we employ them to expand the tail classes. Specifically, to produce more diverse and controllable tail data, we follow the rule of analyzing existing features before generating absent features. To achieve this, LTGC utilizes LMMs to analyze images already in the tail classes to obtain the existing tail-class descriptions list. Then LTGC inputs the existing descriptions list to LLMs to obtain the desired extended tail-class descriptions list for images that are absent from the existing tail classes. Secondly, LTGC utilizes the text-to-image (T2I) model (e.g., DALL-E [33]) to transform these textual descriptions into images. Moreover, inspired by the benefits of chain-of-thought [46], we propose a self-reflection and iterative evaluation module for this process to ensure the diversity and quality of the generated content. Finally, inspired by the advantages of Mixup [52] in merging different image domains, we propose the BalanceMix module to address domain shifts of generated images for the fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.455, + 0.892, + 0.469 + ], + "angle": 0, + "content": "In summary, the contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.47, + 0.892, + 0.514 + ], + "angle": 0, + "content": "1. In the first time, we propose a novel framework via generated content, LTGC, which leverages the power of large models to address long-tail recognition challenges." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.515, + 0.891, + 0.575 + ], + "angle": 0, + "content": "2. We design a series of novel modules to tackle the tail-class image scarcity problem and design the BalanceMix module to efficiently fine-tune the model using the generated data and the original data." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.575, + 0.891, + 0.635 + ], + "angle": 0, + "content": "3. Experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tail benchmarks. Also, the visualization illustrates the diversity and controllability of our generated tail images." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.47, + 0.892, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.649, + 0.642, + 0.665 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Long-Tail recognition. Many approaches have been proposed in recent years to address the long-tail recognition challenge, such as resampling [5, 19], loss rebalancing [1, 6, 12, 23, 26, 27, 29, 39, 48] to increase the model's focus on the tail categories, ensemble learning [42, 53, 56] to improve the ability to recognize tail samples by enhancing model performance, contrast learning [11, 21, 25, 40, 49, 59] and calibration learning [9, 57] to improve the model's decision boundaries, but due to the scarcity of data on the tails, the model still doesn't have enough features to learn class representations for long-tail recognition. Recently, data augmentation-based methods [8, 9, 24] proposed to handle scarcity-tailed data by augmenting from related features. However, these methods find it difficult to obtain the correct knowledge for tail classes. Other approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19511" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "[13, 36] have achieved good results by transfer learning [37] the feature of the visual-language pre-training model (e.g., CLIP [32]) for long-tailed recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.257 + ], + "angle": 0, + "content": "Different from the previous methods, in this paper, from a novel perspective, we explore a generative and fine-tuning framework to handle long-tail recognition by leveraging the rich potential knowledge of large models. Specifically, we design several novel modules to leverage the correct knowledge for generating tail data and perform efficient fine-tuning with the generated data to boost long-tail recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.424 + ], + "angle": 0, + "content": "Large Models. Recently, several large models have emerged, including large language models (e.g., ChatGPT [50]), large multi-modal models (e.g., GPT-4 (ision) [50], MiniGPT [58]) and large generated models (e.g., DALLE [33]). These expansive models are repositories of extensive knowledge [15] and have undergone scrutiny in diverse applications, such as robot task plans [34], open-set object recognition [30], contextual object detection [51], video generation [22], and few-shot segmentation [60]. In this work, we leverage the rich potential knowledge of these large models to handle the long-tail recognition challenge." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.168, + 0.451 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.468, + 0.642 + ], + "angle": 0, + "content": "To perform long-tail recognition well, the challenge is making the model learn diverse representations from tail categories and address imbalances across all categories. In this paper, we propose a novel long-tail recognition framework, LTGC, to handle these challenges. As depicted in Fig. 2, LTGC employ potential knowledge from various off-the-shelf large-scale models to generate and iteratively assess the quality and diversity of tail classes (Sec. 3.1). Furthermore, LTGC proposes the BalanceMix module to facilitating the fine-tuning process with the generated and original datasets (Sec. 3.2). These modules are described in detail below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.355, + 0.667 + ], + "angle": 0, + "content": "3.1. Diverse Tail Images Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To learn diversity representations from tail categories, the previous methods have used data augmentation [9, 24] or transfer learning [13, 36]. However, these works find it difficult to obtain the correct and desired diversity knowledge for tail categories. Inspired by the common-sense knowledge in the LLMs and the fact that textual descriptions are more controllable [33, 50], LTGC takes advantage of these to control the detail and diversity of the generated tail classes. Firstly, to better generate diverse images and control the image detail, LTGC aims to generate images that are absent in the original tail data and represent these images in textual form. Specifically, LTGC employs LMMs to analyze the original tail data and obtain the existing tail-class descriptions list. Then LTGC leverages the common-sense knowledge of LLMs to obtain extended tail-class descrip" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "tions based on the existing tail-class descriptions list. Finally, as images are more suitable as training data for long-tail recognition tasks, we utilize the text-to-image module to generate diverse images based on these tail-class descriptions." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.204, + 0.877, + 0.219 + ], + "angle": 0, + "content": "3.1.1 Obtaining Existing Tail-class Descriptions List" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.596 + ], + "angle": 0, + "content": "In order to ensure the diversity of tail-class images, we first analyze the feature information of the original tail-class images before generating new ones. It guarantees that the content of the generated images is distinct from that of the existing tail-class images. Moreover, due to textual descriptions being more controllable [33], we utilize the textual descriptions to control the detail and diversity of the generated images. To achieve this, our LTGC employs Language Model Multimodals (LMMs), such as GPT-4V (Vision) [50], to analyze the feature information and extract textual descriptions of existing tail classes. During this process, the textual responses from LMMs could be varied and sometimes redundant, which may impede the generation of the desired images. Therefore, we employ textual templates to constrain the responses of LMMs, aiming to unify the textual description formats. Inspired by an image of an object that can be fully described or generated by its class and a list of features [4, 41], we design the textual template to include the given class and its features. Furthermore, the introduction of variations in the scenes plays a crucial role in enhancing the model's ability to generalize [2], a problem that's particularly acute in long-tailed datasets [35]. Therefore, the textual information of scenes is also important for image generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To this end, we design the LMMs' response Template 1 for a given class \\( y \\) as follows: \"A photo of the class [y], {with distinctive features}{in specific scenes}.\", which include the given class, its distinctive features, and specific scenes or environments. With the response template, for a given \\( M_y \\) number class \\( y \\), we sequentially feed these tail-class images into LMMs along with an instructing [Prompt 1] to analyze the features of these images: \"Please use the Template 1 to briefly describe the image of the class [y].\" This process is illustrated in Fig. 3, where \\( y \\) is the given class label. As shown, by formulating the description using the template, LMMs would automatically replace the class, features, and background in the response. After performing this process for all images of each class, we compile a list of text descriptions corresponding to each class. Then we intend to extend the tail-class descriptions list with the existing tail-class descriptions list to generate the images lacking in the tail classes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19512" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.087, + 0.852, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.407, + 0.892, + 0.464 + ], + "angle": 0, + "content": "Figure 2. Overall framework of LTGC. LTGC first employs LMMs to analyze the existing tail data to obtain the existing tail-class descriptions list. Then it inputs the list into LLMs to analyze the absent features of the tail classes and employs the T2I model to generate diverse images. Moreover, our designed self-reflection and iterative evaluation modules ensure the diversity and quality of the tail data. Finally, LTGC employs the BalanceMix module to fine-tune the CLIP's visual encoder with the extended and original data." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.477, + 0.465, + 0.492 + ], + "angle": 0, + "content": "3.1.2 Obtaining Extended Tail-class Descriptions List" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.506, + 0.47, + 0.763 + ], + "angle": 0, + "content": "In this section, we aim to analyze features missing from tail classes and enrich the descriptions of these classes based on the existing tail-class descriptions list. To accomplish this, inspired by the rich common-sense knowledge of Large Language Models (LLMs), such as ChatGPT [50], we extend the tail-class descriptions list through a two-step process: 1) Inputting the existing descriptions list into LLMs, and 2) Designing the Prompt 2 to guide LLMs in generating the desired descriptions for images that are absent in the given tail class \\( y \\): \"Besides these descriptions mentioned above, please use the Template 1 to list other possible {distinctive features} and {specific scenes} for the class [y],\" which is illustrated in Fig. 4. For each class, we repeat the above two-step process, and then we obtain the tail-class descriptions list for all tail classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In addition, in order for these generated descriptions to better complement each tail class, we encourage LLMs to generate descriptions of sufficient number and diversity. To meet this goal, we introduce a self-reflection module in this process, aiming to guide LMMs in rethinking if there are any features or scenes that are missed or repeated. It includes two key designs: a number-checking module and a repetition-checking module. (1) The number-checking module aims to guide LMMs in rethinking whether there" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.81 + ], + "angle": 0, + "content": "are other missing features or scenarios. To achieve this, after posing the initial [Prompt 2] for class \\( y \\), we update the extended descriptions list and re-ask LLMs the [Prompt 2] question, incorporating the newly acquired list. For each class \\( y \\), this iterative process of number-checking continues until a maximum number \\( K_y \\) of the tail class is achieved, where \\( K_y = M_y + N_y \\) and \\( N_y \\) is the number of generated descriptions for class \\( y \\). (2) The repetition-checking module aims to guide LMMs in rethinking if there are other features or scenes that are repeated at the end of the number-checking iteration. Specifically, we input the extended descriptions list and the following [Prompt 3] of each class \\( y \\) for LLMs' repetition checking: \"Please exclude any repetitive {distinctive features} and {specific scenes} for class \\( [y] \\) in this descriptions list.\" After that, LLM will filter out the ones with repeated features based on this prompt and return a new list of descriptions. Through the implementation of this two-step process and LLMs' self-reflection module, we have obtained a reduction in repeated descriptions and an increase in the diversity of the tail-class descriptions for each class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.791, + 0.847 + ], + "angle": 0, + "content": "3.1.3 Transform Descriptions to Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Above, we obtain the extended tail-class descriptions list for each class \\(y\\), denoted as \\(L_{y}\\). As the images are better adapted to perform visual recognition tasks, in this section," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19513" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.462, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.389 + ], + "angle": 0, + "content": "Figure 3. Example of the instruction for LMMs. When both images from tail classes and textual templates are input into LMMs, textual descriptions corresponding to the images can be obtained. By repeatedly performing this operation on the training data, we convert abstract image descriptions into concrete textual descriptions. Finally, we acquire the current textual descriptions list corresponding to each class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.465 + ], + "angle": 0, + "content": "we aim to leverage the image-generative ability of the text-to-image (T2I) method to generate the images from the tail-class descriptions list. In detail, we employ T2I to generate images based on the descriptions list, denoted as:" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.475, + 0.462, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.649, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Figure 4. Example of the instruction for LLMs. LTGC inputs the existing textual descriptions list to LLMs, which continually extends it with new distinctive features and scene information. During multiple iterations, LTGC generates a new extended textual descriptions list for each class." + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.74, + 0.47, + 0.759 + ], + "angle": 0, + "content": "\\[\ni _ {n} ^ {y} = \\mathrm {T 2 I} \\left(d _ {n} ^ {y}\\right), \\text {w h e r e} n \\in \\{1, \\dots , N \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\( n \\in N \\) denotes the \\( n \\)-th generated descriptions for class \\( y \\), and \\( i_n^y \\) denotes the image generated by T2I model based on \\( d_n^y \\). However, some generated images may not be of sufficiently high quality to accurately represent their desired classes, as the lower-quality images. As shown in Fig. 7, lower-quality images will output poorly distinctive features, resulting in more confusion for the class. Therefore, using these generated images directly in training may result in disrupting the model's prediction of the tail classes." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.087, + 0.891, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.267, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 5. Illustration of the proposed iterative evaluation module framework. This module detects lower-quality images through the similarity score \\( S \\) computed by images and their corresponding class feature template. Then the textual descriptions corresponding to lower-quality images are re-input into LLMs for refinement. Finally, the refined textural descriptions are fed into the T2I model for regeneration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.612 + ], + "angle": 0, + "content": "To address the challenge of lower-quality images that fail to represent their desired classes accurately, we aim to automatically detect these images, refine their textual descriptions, and regenerate them accordingly. To achieve this, drawing inspiration from the idea that humans improve their understanding through peer feedback, we investigate the possibility of a large model also enhancing its output by integrating feedback from another fundamental model. Therefore, we propose a iterative evaluation module, as illustrated in Fig. 5. This module utilizes the fundamental model, CLIP, to serve as the 'feedback giver' selected for its proficiency in understanding images and effectively connecting them with textual contexts [32]. It provides feedback about the lower-quality images to LLMs, which assists in refining their corresponding descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Furthermore, to facilitate CLIP in grasping the represented features in images, a straightforward way is to check the degree of match between the image and its class template: \"A photo of a [y].\" However, due to the fact that CLIP's text encoder may contain less feature knowledge of the tail classes (e.g., as presented on the iNaturalist 2018 dataset with poor zero-shot recognition performance [13]), we aim to design a new textual feature template for each class, highlighting its unique features. Then CLIP employs it to determine the correspondence between the template and the image. Drawing inspiration from the chain-of-thought [30, 46] concept, which suggests that LMMs perform better when provided with additional clues, we further guide LMMs in summarizing the most distinctive features of each class. Specifically, after obtaining the extended tail-class descriptions list, we obtain the class feature template \\( C_y \\) that contains the most distinctive features by the following [Prompt 4] and Template 2 for the given class \\( y \\): [Prompt 4]: \"Please use" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19514" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.257 + ], + "angle": 0, + "content": "Template 2 to summarize the most distinctive features of class \\([y]\\)\". \" Template 2: \"A photo of the class [y] with {feature 1} {feature 2} {...}. \" This process could guide the LMMs to rethink and summarize the most distinctive features and produce the class feature template \\( C_y \\) for each class \\( y \\). Furthermore, since we focus on checking whether the feature information of an image could represent its desired class, we do not include scene information in the class feature template. Then we introduce the details of the iterative evaluation module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.468, + 0.35 + ], + "angle": 0, + "content": "As shown in Fig. 5, our iterative evaluation module conducts the following three steps iteratively: (1) Detection. To identify lower-quality images, we first employ utilize the strong capability of aligning images with text of CLIP to match the feature of generated images \\( i_n^y \\) and its class feature template \\( C_y \\) by cosine similarity metric as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.369, + 0.469, + 0.386 + ], + "angle": 0, + "content": "\\[\nS = \\operatorname {E n c o d e r} _ {\\text {v i s}} \\left(i _ {n} ^ {y}\\right) \\cdot \\operatorname {E n c o d e r} _ {\\text {t e x t}} \\left(C _ {y}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.468, + 0.606 + ], + "angle": 0, + "content": "where \\(\\mathsf{Encoder}_{\\mathrm{vis}}\\) denotes the CLIP's visual encoder, and \\(\\mathsf{Encoder}_{\\mathrm{text}}\\) denotes the text encoder of the CLIP. The image \\(i_n^y\\) is detected as lower quality and filtered out if its similarity score \\(S\\) is below a threshold \\(\\mu\\). (2) Refinement. Next, also inspired by chain-of-thought [46], we propose the following prompt for refine its descriptions. Specifically, if \\(i_n^y\\) is identified as a lower quality image, we prompt LLMs to refine its corresponding description \\(d_n^y\\) to more accurately represent the intended class \\(y\\), drawing on feedback provided by CLIP, i.e., [Prompt 5]: \"This description \\(d_n^y\\) doesn't seem to be representative of the class \\([y]\\). Could you refine it to enhance the distinctive features of class \\([y]?\"\" (3)\\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.607, + 0.468, + 0.636 + ], + "angle": 0, + "content": "Re-generation. Finally, the image \\( i_n^y \\) is regenerated by the T2I model according to the improved textual description." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.468, + 0.727 + ], + "angle": 0, + "content": "By iteratively applying these three stages, we employ CLIP in each cycle to furnish precise feedback to the LLMs, thereby steering the refinement of descriptions to better align with our desired direction. These refined descriptions ensure the production of images that more accurately embody the characteristics of each class." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.206, + 0.755 + ], + "angle": 0, + "content": "3.2. BalanceMix" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "With generated tailed images, the final problem is how to efficiently use these generated data and original data to perform long-tailed recognition well. Due to the domain gap between the generated data and the original long-tail data [38], we propose a method named BalanceMix to handle this challenge. We first define the original data and generated data as \\( \\mathcal{D}_o \\) and \\( \\mathcal{D}_g \\). Then BalanceMix balance-sample [5] an image \\( x_i \\) from \\( \\mathcal{D}_o \\) and sample an image \\( x_j \\) from \\( \\mathcal{D}_g \\). Meanwhile, it mixes the images \\( x_i \\) and \\( x_j \\) and their corre" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.689, + 0.106 + ], + "angle": 0, + "content": "sponding labels, denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.12, + 0.892, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\widetilde {x} = \\lambda \\odot x _ {i} + (\\mathbf {1} - \\lambda) \\odot x _ {j}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.164, + 0.892, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\widetilde {y} = \\lambda \\odot y _ {i} + (\\mathbf {1} - \\lambda) \\odot y _ {j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.188, + 0.892, + 0.234 + ], + "angle": 0, + "content": "where \\(\\lambda\\) in Beta (0,1) distribution. Finally, we fine-tune the CLIP's vision encoder with LORA [18] on all mixed data pairs \\((\\widetilde{x},\\widetilde{y})\\) for efficient long-tail recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.248, + 0.632, + 0.265 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.274, + 0.892, + 0.395 + ], + "angle": 0, + "content": "We present the experimental results on three widely used datasets in long-tailed recognition, including ImageNet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. Moreover, we undertake ablation studies specifically on the ImageNet-LT and iNaturalist 2018 datasets to gain deeper insights into the performance of our method. The experimental results of the comparison methods are taken from their original paper, and our results are averaged over three experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.719, + 0.421 + ], + "angle": 0, + "content": "4.1. Implementation details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Evaluation Setup. In all experiments, we evaluate and report top-1 accuracy on their corresponding test set. We also report accuracy on three splits of the classes: Many-shot (more than 100 images), Medium-shot (20 to 100 images), and Few-shot (less than 20 images) [20]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Method Implementation. In our LTGC, we incorporate diverse and specialized knowledge from the off-the-shelf large models. Specifically, for LMM, we use the GPT-4V (ision) [50] version of ChatGPT. For LLM, we use the GPT-4 version of ChatGPT. For T2I, we use DAll-E [33]. For the pre-trained CLIP [32], we use ViT-B/32 [14] for its visual encoder and the architecture introduced in [31] as the text encoder. In LLM's self-reflection module, we set the maximum number \\( K_{y} \\) to 100, 300, and 800 for iNaturalist 2018, ImageNet-LT, and Place-LT, respectively. In the iterative evaluation module, the threshold \\( \\mu \\) is set at 0.8 for ImageNet-LT and Place-LT, and at 0.6 for iNaturalist. In Appendix, we provide a detailed ablation and discussion on the choice of LMM and LLM models, as well as parameters used in the self-reflection process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.852, + 0.758 + ], + "angle": 0, + "content": "4.2. Comparisons with SOTA on Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare our proposed LTGC model with state-of-the-art (SOTA) methods on three benchmarks, including Imagenet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. To ensure a fair comparison, we primarily focus on methods based on CLIP, as they also leverage knowledge pretrained on large-scale datasets. These baselines include CLIP Zero-Shot [32, 36] and CLIP Finetune [36], as well as CLIP-based long-tail recognition approaches such as VL-LTR [36], LPT [13], and RAC [28]. Additionally," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19515" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "we report comparisons with traditional methods (without CLIP) on the challenging and fine-grained, large-scale long-tail dataset, iNaturalist 2018. These have proven the effectiveness of our method in different scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.471, + 0.258 + ], + "angle": 0, + "content": "Results on Imagenet-LT. In Tab. 1, we observe that our LTGC models are superior to other CLIP-based LT methods. For example, our method achieves an overall accuracy of \\(80.6\\%\\), which is \\(3.4\\%\\) higher than the existing SOTA method VT-LTR[36]. Moreover, the overall accuracy of our method marginally surpasses the results on the full ImageNet (i.e., \\(80\\%\\) [36])." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.471, + 0.349 + ], + "angle": 0, + "content": "Results on Places-LT. Tab. 1 shows that compared to other CLIP variant methods, LTGC achieves an overall accuracy of \\(54.1\\%\\) and an few-shot accuracy of \\(52.1\\%\\), surpassing the LPT [13] by \\(4.0\\%\\) and \\(5.2\\%\\), respectively. Even compared with VL-LTR [36] and RAC [28], which have extra data in training and testing, our LTGC achieves remarkable results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.471, + 0.499 + ], + "angle": 0, + "content": "Results on iNaturalist 2018. In this part, we employ LTGC on a large and fine-grained dataset, iNaturalist 2018. Tab. 2 presents the quantitative results. LTGC leverages the rich knowledge of LLMs and significantly outperforms traditional deep-learning approaches for long-tail recognition. In addition, LTGC attains an overall accuracy of \\(82.5\\%\\) and a few-shot accuracy of \\(82.6\\%\\), outperforming all existing SOTA methods based on CLIP. In particular, LTGC also surpasses the retrieval augmented method, RAC [28], by \\(2.3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.511, + 0.445, + 0.527 + ], + "angle": 0, + "content": "4.3. Compare with different methods of LMMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.685 + ], + "angle": 0, + "content": "In the experiments, our LTGC employs GPT-4V (ision) to describe a given image to obtain a text-based feature description. A straightforward way for long-tail recognition is by querying LMMs for the category of the given image. To evaluate the ability of these methods to recognize the long tail images, we constructed the following baselines of LMMs for comparison: MiniGPT4 [58], MiniGPT4-v2 [7], LENS [3], and GPT-4V (ision) [50]. To perform a fair evaluation, we provide the label list [Class 1, Class 2, ..., Class Y] for each dataset, where \\(Y\\) is the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.747 + ], + "angle": 0, + "content": "The results presented in Tab. 3, show that our method significantly outperforms the baselines that directly query the LMMs. More implementation details are discussed in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.334, + 0.773 + ], + "angle": 0, + "content": "4.4. Analysis and Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of the BalanceMix module. To evaluate the effectiveness of the two proposed designs, we conduct experiments on three different variants: 1) w/o BalanceMix: This variant is directly fine-tuned using the generated data and the original training data. 2) Balanced Sample: This variant utilizes generated data to perform balanced sampling [45] for training. 3) Mixup: This variant employs generated data with Mixup [52] without performing balanced" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.116 + ], + "angle": 0, + "content": "Table 1. Comparison with SOTA methods on ImageNet-LT and Places-LT." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.129, + 0.848, + 0.233 + ], + "angle": 0, + "content": "
DatasetImageNet-LTPlaces-LT
FewAllFewAll
CLIP Zero [36]58.659.840.138.0
CLIP Finetune [36]34.560.522.739.7
VL-LTR [36]59.377.242.050.1
RAC [28]--41.847.2
LPT [13]--46.950.1
LTGC(Ours)70.580.652.154.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.252, + 0.882, + 0.266 + ], + "angle": 0, + "content": "Table 2. Comparison with SOTA methods on iNaturalist 2018." + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.277, + 0.86, + 0.43 + ], + "angle": 0, + "content": "
MethodManyMediumFewAll
Softmax74.766.360.064.7
LADE [16]64.447.734.352.3
RIDE [42]71.570.071.671.8
PaCo [11]69.573.473.073.0
MDCS [56]76.575.575.275.6
CLIP Zero [36]6.13.32.93.4
CLIP Finetune [36]76.674.170.272.6
VL-LTR [36]---76.8
RAC [28]75.980.581.080.2
LPT [13]--79.376.1
LTGC(Ours)77.583.982.682.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.475 + ], + "angle": 0, + "content": "Table 3. Comparison with different LMMs' methods on ImageNet-LT and iNaturalist2018." + }, + { + "type": "table", + "bbox": [ + 0.542, + 0.487, + 0.852, + 0.565 + ], + "angle": 0, + "content": "
MethodImageNet-LTiNaturalist 2018
LENS [3]69.517.4
MiniGPT4 [58]60.420.9
MiniGPT4-v2 [7]68.527.1
GPT-472.164.3
Ours80.682.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.791 + ], + "angle": 0, + "content": "sampling. The results are shown in Tab. 4, demonstrating that our method significantly outperforms the other variants. Although our self-reflection and iterative evaluation modules already ensure the diversity and high quality of the generated images, there still exists a domain gap between the generated and original data [38]. This domain gap could exacerbate the long-tail effect on the test set (discussed in Appendix). A simple balanced sample [45] approach fails to solve the class imbalance caused by the mixing of original and generated data, while the Mixup [52] alone cannot address inter-class imbalances. To this end, our BalanceMix module combines the strengths of both methods, making the generated data well-suited for our framework." + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.81, + 0.872, + 0.823 + ], + "angle": 0, + "content": "Table 4. Evaluation on the effectiveness of the BalanceMix." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.835, + 0.848, + 0.9 + ], + "angle": 0, + "content": "
MethodImageNet-LTiNaturalist
w/o BalanceMix58.369.5
Balanced sample [45]63.973.8
Mixup [52]73.475.2
Ours80.682.5
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19516" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.088, + 0.85, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.895, + 0.364 + ], + "angle": 0, + "content": "Figure 6. The visualization of generated images: The template \"A photo of the class [y]\" and LTGC. Each row represents a different class. The four images on the left are generated using the simple template \"A photo of the class [y],\" which results in images with uniform poses and plain backgrounds. The four images on the right are from the proposed LTGC and demonstrate the diversity of classes." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.385, + 0.468, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.593, + 0.471, + 0.665 + ], + "angle": 0, + "content": "Figure 7. The visualization of the images generated before and after passing the iterative evaluation module. The top row displays images that were filtered out, while the bottom row shows images regenerated by T2I after refining their corresponding descriptions. More visualizations are in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.216, + 0.709 + ], + "angle": 0, + "content": "4.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Visualization of Generated Images: Template \"A photo of the class [y]\" vs our LTGC. Furthermore, we compare the images generated by LTGC with a simple prompt, \"A photo of the class [y]\" for the T2I model. As Fig. 6 illustrates, LTGC generates more accurate and diverse images compared to those generated by the simple prompt. For example, in the absence of control over the category description, the T2I model generates red features that do not belong to class Aquilegia Pubescens. In addition, with rich text for image generation, our approach also generates more diverse and accurate images compared to images generated by a simple template." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Visualization on iterative evaluation module. We compare images before and after refinement using the iterative evaluation module, and the visualization results are shown in Fig. 7. It shows that images before refinement often possess ambiguous semantic information, and the features of the corresponding classes are not distinct. However, after the refinement process, the quality of the images is substantially enhanced, and the distinctive features of the corresponding classes become more pronounced." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.536, + 0.619, + 0.551 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.892, + 0.756 + ], + "angle": 0, + "content": "This paper introduces a novel generative and fine-tuning framework, named LTGC, to address the challenge of long-tail recognition. LTGC leverages the abundant potential knowledge embedded in large-scale models to generate diverse data for tail categories. The framework incorporates innovative designs to ensure the quality of the generated data and to fine-tune the model efficiently using both the generated and original data. The experimental results indicate that LTGC surpasses existing state-of-the-art methods on well-known long-tail benchmarks. In the future, we will explore the robustness of large-scale models for application in other areas, such as long-tail semantic segmentation and dissent detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.661, + 0.788 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.62271034. This project is also supported by the Ministry of Education, Singapore, under the SUTD Kickstarter Initiative Project (SKI 2021_02_06). Additionally, I would like to express my gratitude to Haoxuan Qu and He Huang for their valuable suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19517" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Emanuel Sanchez Aimar, Arvi Jonnarth, Michael Felsberg, and Marco Kuhlmann. Balanced product of calibrated experts for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19967-19977, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Martin Arjovsky, León Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Michel Besserve, Arash Mehrjou, Rémy Sun, and Bernhard Schölkopf. Counterfactuals uncover the modular structure of deep generative models. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.4 + ], + "angle": 0, + "content": "[5] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. arXiv preprint arXiv:1906.07413, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.626 + ], + "angle": 0, + "content": "[8] Xiaohua Chen, Yucan Zhou, Dayan Wu, Wanqian Zhang, Yu Zhou, Bo Li, and Weiping Wang. Imagine by reasoning: A reasoning-based implicit semantic data augmentation for long-tailed classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 356-364, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.471, + 0.698 + ], + "angle": 0, + "content": "[9] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 95–110. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.701, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[10] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in gpt-4v(ison): Bias and interference challenges, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.758, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[11] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.884 + ], + "angle": 0, + "content": "[12] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Zuo. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[15] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[16] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The naturalist challenge 2017 dataset. CoRR, abs/1707.06642, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.894, + 0.427 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.894, + 0.468 + ], + "angle": 0, + "content": "[19] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5): 429-449, 2002. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.894, + 0.525 + ], + "angle": 0, + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.58 + ], + "angle": 0, + "content": "[21] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.894, + 0.65 + ], + "angle": 0, + "content": "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[23] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE transactions on neural networks and learning systems, 29(8):3573-3587, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.789 + ], + "angle": 0, + "content": "[24] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[25] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19518" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "transactions on pattern analysis and machine intelligence, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[27] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[28] Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, and Anton van den Hengel. Retrieval augmented classification for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6959-6969, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[30] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. arXiv preprint arXiv:2309.12780, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.433 + ], + "angle": 0, + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[33] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[35] Kaihua Tang, Mingyuan Tao, Jiaxin Qi, Zhenguang Liu, and Hanwang Zhang. Invariant feature learning for generalized long-tailed classification. In European Conference on Computer Vision, pages 709-726. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[36] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[37] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[39] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 943-952, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[40] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[41] Tan Wang, Zhongqi Yue, Jianqiang Huang, Qianru Sun, and Hanwang Zhang. Self-supervised learning disentangled group representation as feature. Advances in Neural Information Processing Systems, 34:18225-18240, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[42] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.343, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[43] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[44] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 7032-7042, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[45] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[47] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 247–263. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[48] Cihang Xie and Alan Yuille. Intriguing properties of adversarial training at scale. arXiv preprint arXiv:1906.03787, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[49] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in neural information processing systems, 33:19290-19301, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[50] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[51] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19519" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[53] Yifan Zhang, Bryan Hooi, Lanqing Hong, and Jiashi Feng. Test-agnostic long-tailed recognition by test-time aggregating diverse experts with self-supervision. arXiv preprint arXiv:2107.09249, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[54] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[55] Qihao Zhao, Yangyu Huang, Wei Hu, Fan Zhang, and Jun Liu. Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[56] Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, and Jun Liu. Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11597-11608, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.347, + 0.47, + 0.402 + ], + "angle": 0, + "content": "[57] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16489-16498, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.404, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[59] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.531, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[60] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19520" + } + ] +] \ No newline at end of file diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_origin.pdf b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad270103c5f55c23f1cd4bb8244dd3220cab11a2 --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/2e9dbf01-8020-47fa-831f-ee0b1bfd9643_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4162eaf2019ea322a21433f47829a90838e3972f20d4b0a017a2ab1d3ad3bb10 +size 8894912 diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/full.md b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e838e62b7fe1eed3817cced16cb082701c91c1bd --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/full.md @@ -0,0 +1,264 @@ +# LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content + +Qihao Zhao $^{1,2,*}$ , Yalun Dai $^{3,*}$ , Hao Li $^{4}$ , Wei Hu $^{1}$ , Fan Zhang $^{1\dagger}$ , Jun Liu $^{2}$ $^{1}$ Beijing University of Chemical Technology, China + $^{2}$ Singapore University of Technology and Design, Singapore + $^{3}$ Nanyang Technological University, Singapore + $^{4}$ Northwestern Polytechnical University, China + +# Abstract + +Long-tail recognition is challenging because it requires the model to learn good representations from tail categories and address imbalances across all categories. In this paper, we propose a novel generative and fine-tuning framework, LTGC, to handle long-tail recognition via leveraging generated content. Firstly, inspired by the rich implicit knowledge in large-scale models (e.g., large language models, LLMs), LTGC leverages the power of these models to parse and reason over the original tail data to produce diverse tail-class content. We then propose several novel designs for LTGC to ensure the quality of the generated data and to efficiently fine-tune the model using both the generated and original data. The visualization demonstrates the effectiveness of the generation module in LTGC, which produces accurate and diverse tail data. Additionally, the experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tailed benchmarks. + +# 1. Introduction + +In the real world, data often exhibits a long-tailed distribution, posing significant challenges for computer vision recognition [44, 54]. These challenges include (1) Class Imbalance: within the dataset, some classes (termed "head" classes) are abundantly represented, while others (termed "tail" classes) have few samples. This imbalanced distribution during training may cause the model to attend more on the head classes, neglecting the tail classes [42]. (2) Tail Data Scarcity: Data scarcity refers to tail classes having extremely limited samples, which lack diversity and are insufficient to effectively train a model [8, 24]. It prevents a model's ability to learn the feature invariant, which is necessary to recognize these categories correctly [35]. + +![](images/31f361492f3359d2f385fee6d7a43ccc82031d239f95078b919e7867dc0958ff.jpg) +Figure 1. Top: Here, our LMMs use ChatGPT. For the Trogon Rufus category, when we asked ChatGPT, "What species is in the picture?" we did not get the expected answer due to the complexity of the question. Middle and Down: In contrast, when we asked some easy questions for the ChatGPT, "Please describe the image." or "Please describe the distinctive features of Trogon Rufus." It could accurately answer these questions. + +To address these challenges, numerous approaches have emerged within the field of long-tailed recognition, such as re-sampling [5, 19], loss re-weighting [1, 6, 12, 23, 26, 27, 29, 39, 48], ensemble learning [42, 47, 56], decoupling [20], and contrastive learning [11, 25, 40, 59]. Their primary objective is to balance the decision boundaries and parameter weights of the model to improve long-tail recognition. + +However, these methods encounter bottlenecks due to the scarcity and limited diversity features of tail-class data [24]. To obtain more diverse representations for tail classes, two types of methods appear: Some methods increase tail-class diversity through data augmentation [9, 24, 52, 55]. Others increase tail-class diversity by transferring features from related classes [8] or large pre-trained models (e.g., CLIP [32]). These transfer learning methods [13, 28, 36] based on CLIP have recently shown great potential in boosting long-tail recognition. The LPT [13] fine-tunes pre-trained models to adapt to target datasets. The VL-LTR [36] leverages text features from CLIP to augment the learning of image features. However, these methods find it hard to obtain the correct and desired knowledge for tail categories. For example, in the real-world data iNauralist 2018 [17], the category Aquilegia Pubescens is characterized only by white or pale yellow colors and spurred shapes. If we transfer similar category features or blindly augment semantic information, such as red color semantics, this category will become more confusing. + +Recently, Large Language Models (LLMs) (e.g., ChatGPT) and Large Multimodal Models (LMMs) (e.g., GPT-4V (ChatGPT with Vision [50]) and Minigpt4 [7, 58]) due to their wealth of potential knowledge, have been leveraged for a variety of downstream tasks, such as robot task plans [34], open-set object recognition [30], automated robot learning [43] and various visual reasoning tasks [50, 51, 60]. However, due to the bias of imbalanced training data, LMMs perform poorly on some complex tasks [10]. We also find this dilemma in long-tail recognition as shown in Fig. 1. When we asked LMMs about an image in the category of Trogon rufus: "What species is in the image?", we did not get the expected answer. In contrast, when we ask some easy questions about this image: "Please describe the image." or "Please describe the distinctive features of Trogon Rufus", we get the desired answer. This suggests that although large models underperform in long-tail recognition (e.g., it is difficult to align complex image features with labels.), they still contain desired potential knowledge (e.g., correctly describing images and providing the knowledge of species). Inspired by this, we aim to leverage the abundant potential knowledge of several large models to address the challenges of long-tailed data scarcity and perform long-tailed recognition well. Nevertheless, effectively leveraging large models to encounter these challenges is not trivial: (1) The potential nature of knowledge within large models makes it difficult to extract the desired knowledge that facilitates long-tail recognition tasks. (2) Due to the domain gap between the generated data and the original data [38], effectively using these hybrid data remains an open issue. + +To this end, we propose a novel Long-Tail recognition framework via Generated Content, denoted as LTGC, + +which is illustrated in Fig. 2. Our LTGC aims to leverage the capabilities of large models for generating explicitly diverse content [15] tailored to the long-tail classes and incorporate novel designs to enhance long-tail recognition. Firstly, motivated by the wealth of potential knowledge in large models and the fact that texts are more controllable, we employ them to expand the tail classes. Specifically, to produce more diverse and controllable tail data, we follow the rule of analyzing existing features before generating absent features. To achieve this, LTGC utilizes LMMs to analyze images already in the tail classes to obtain the existing tail-class descriptions list. Then LTGC inputs the existing descriptions list to LLMs to obtain the desired extended tail-class descriptions list for images that are absent from the existing tail classes. Secondly, LTGC utilizes the text-to-image (T2I) model (e.g., DALL-E [33]) to transform these textual descriptions into images. Moreover, inspired by the benefits of chain-of-thought [46], we propose a self-reflection and iterative evaluation module for this process to ensure the diversity and quality of the generated content. Finally, inspired by the advantages of Mixup [52] in merging different image domains, we propose the BalanceMix module to address domain shifts of generated images for the fine-tuning process. + +In summary, the contributions of our work are as follows: + +1. In the first time, we propose a novel framework via generated content, LTGC, which leverages the power of large models to address long-tail recognition challenges. +2. We design a series of novel modules to tackle the tail-class image scarcity problem and design the BalanceMix module to efficiently fine-tune the model using the generated data and the original data. +3. Experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tail benchmarks. Also, the visualization illustrates the diversity and controllability of our generated tail images. + +# 2. Related Work + +Long-Tail recognition. Many approaches have been proposed in recent years to address the long-tail recognition challenge, such as resampling [5, 19], loss rebalancing [1, 6, 12, 23, 26, 27, 29, 39, 48] to increase the model's focus on the tail categories, ensemble learning [42, 53, 56] to improve the ability to recognize tail samples by enhancing model performance, contrast learning [11, 21, 25, 40, 49, 59] and calibration learning [9, 57] to improve the model's decision boundaries, but due to the scarcity of data on the tails, the model still doesn't have enough features to learn class representations for long-tail recognition. Recently, data augmentation-based methods [8, 9, 24] proposed to handle scarcity-tailed data by augmenting from related features. However, these methods find it difficult to obtain the correct knowledge for tail classes. Other approaches + +[13, 36] have achieved good results by transfer learning [37] the feature of the visual-language pre-training model (e.g., CLIP [32]) for long-tailed recognition. + +Different from the previous methods, in this paper, from a novel perspective, we explore a generative and fine-tuning framework to handle long-tail recognition by leveraging the rich potential knowledge of large models. Specifically, we design several novel modules to leverage the correct knowledge for generating tail data and perform efficient fine-tuning with the generated data to boost long-tail recognition. + +Large Models. Recently, several large models have emerged, including large language models (e.g., ChatGPT [50]), large multi-modal models (e.g., GPT-4 (ision) [50], MiniGPT [58]) and large generated models (e.g., DALLE [33]). These expansive models are repositories of extensive knowledge [15] and have undergone scrutiny in diverse applications, such as robot task plans [34], open-set object recognition [30], contextual object detection [51], video generation [22], and few-shot segmentation [60]. In this work, we leverage the rich potential knowledge of these large models to handle the long-tail recognition challenge. + +# 3. Method + +To perform long-tail recognition well, the challenge is making the model learn diverse representations from tail categories and address imbalances across all categories. In this paper, we propose a novel long-tail recognition framework, LTGC, to handle these challenges. As depicted in Fig. 2, LTGC employ potential knowledge from various off-the-shelf large-scale models to generate and iteratively assess the quality and diversity of tail classes (Sec. 3.1). Furthermore, LTGC proposes the BalanceMix module to facilitating the fine-tuning process with the generated and original datasets (Sec. 3.2). These modules are described in detail below. + +# 3.1. Diverse Tail Images Generation + +To learn diversity representations from tail categories, the previous methods have used data augmentation [9, 24] or transfer learning [13, 36]. However, these works find it difficult to obtain the correct and desired diversity knowledge for tail categories. Inspired by the common-sense knowledge in the LLMs and the fact that textual descriptions are more controllable [33, 50], LTGC takes advantage of these to control the detail and diversity of the generated tail classes. Firstly, to better generate diverse images and control the image detail, LTGC aims to generate images that are absent in the original tail data and represent these images in textual form. Specifically, LTGC employs LMMs to analyze the original tail data and obtain the existing tail-class descriptions list. Then LTGC leverages the common-sense knowledge of LLMs to obtain extended tail-class descrip + +tions based on the existing tail-class descriptions list. Finally, as images are more suitable as training data for long-tail recognition tasks, we utilize the text-to-image module to generate diverse images based on these tail-class descriptions. + +# 3.1.1 Obtaining Existing Tail-class Descriptions List + +In order to ensure the diversity of tail-class images, we first analyze the feature information of the original tail-class images before generating new ones. It guarantees that the content of the generated images is distinct from that of the existing tail-class images. Moreover, due to textual descriptions being more controllable [33], we utilize the textual descriptions to control the detail and diversity of the generated images. To achieve this, our LTGC employs Language Model Multimodals (LMMs), such as GPT-4V (Vision) [50], to analyze the feature information and extract textual descriptions of existing tail classes. During this process, the textual responses from LMMs could be varied and sometimes redundant, which may impede the generation of the desired images. Therefore, we employ textual templates to constrain the responses of LMMs, aiming to unify the textual description formats. Inspired by an image of an object that can be fully described or generated by its class and a list of features [4, 41], we design the textual template to include the given class and its features. Furthermore, the introduction of variations in the scenes plays a crucial role in enhancing the model's ability to generalize [2], a problem that's particularly acute in long-tailed datasets [35]. Therefore, the textual information of scenes is also important for image generation. + +To this end, we design the LMMs' response Template 1 for a given class $y$ as follows: "A photo of the class [y], {with distinctive features}{in specific scenes}.", which include the given class, its distinctive features, and specific scenes or environments. With the response template, for a given $M_y$ number class $y$ , we sequentially feed these tail-class images into LMMs along with an instructing [Prompt 1] to analyze the features of these images: "Please use the Template 1 to briefly describe the image of the class [y]." This process is illustrated in Fig. 3, where $y$ is the given class label. As shown, by formulating the description using the template, LMMs would automatically replace the class, features, and background in the response. After performing this process for all images of each class, we compile a list of text descriptions corresponding to each class. Then we intend to extend the tail-class descriptions list with the existing tail-class descriptions list to generate the images lacking in the tail classes. + +![](images/120199c534a605dc42736c8f510eb18b7a984b55c76a75c0830643deeef4af6d.jpg) +Figure 2. Overall framework of LTGC. LTGC first employs LMMs to analyze the existing tail data to obtain the existing tail-class descriptions list. Then it inputs the list into LLMs to analyze the absent features of the tail classes and employs the T2I model to generate diverse images. Moreover, our designed self-reflection and iterative evaluation modules ensure the diversity and quality of the tail data. Finally, LTGC employs the BalanceMix module to fine-tune the CLIP's visual encoder with the extended and original data. + +# 3.1.2 Obtaining Extended Tail-class Descriptions List + +In this section, we aim to analyze features missing from tail classes and enrich the descriptions of these classes based on the existing tail-class descriptions list. To accomplish this, inspired by the rich common-sense knowledge of Large Language Models (LLMs), such as ChatGPT [50], we extend the tail-class descriptions list through a two-step process: 1) Inputting the existing descriptions list into LLMs, and 2) Designing the Prompt 2 to guide LLMs in generating the desired descriptions for images that are absent in the given tail class $y$ : "Besides these descriptions mentioned above, please use the Template 1 to list other possible {distinctive features} and {specific scenes} for the class [y]," which is illustrated in Fig. 4. For each class, we repeat the above two-step process, and then we obtain the tail-class descriptions list for all tail classes. + +In addition, in order for these generated descriptions to better complement each tail class, we encourage LLMs to generate descriptions of sufficient number and diversity. To meet this goal, we introduce a self-reflection module in this process, aiming to guide LMMs in rethinking if there are any features or scenes that are missed or repeated. It includes two key designs: a number-checking module and a repetition-checking module. (1) The number-checking module aims to guide LMMs in rethinking whether there + +are other missing features or scenarios. To achieve this, after posing the initial [Prompt 2] for class $y$ , we update the extended descriptions list and re-ask LLMs the [Prompt 2] question, incorporating the newly acquired list. For each class $y$ , this iterative process of number-checking continues until a maximum number $K_y$ of the tail class is achieved, where $K_y = M_y + N_y$ and $N_y$ is the number of generated descriptions for class $y$ . (2) The repetition-checking module aims to guide LMMs in rethinking if there are other features or scenes that are repeated at the end of the number-checking iteration. Specifically, we input the extended descriptions list and the following [Prompt 3] of each class $y$ for LLMs' repetition checking: "Please exclude any repetitive {distinctive features} and {specific scenes} for class $[y]$ in this descriptions list." After that, LLM will filter out the ones with repeated features based on this prompt and return a new list of descriptions. Through the implementation of this two-step process and LLMs' self-reflection module, we have obtained a reduction in repeated descriptions and an increase in the diversity of the tail-class descriptions for each class. + +# 3.1.3 Transform Descriptions to Images + +Above, we obtain the extended tail-class descriptions list for each class $y$ , denoted as $L_{y}$ . As the images are better adapted to perform visual recognition tasks, in this section, + +![](images/35a579247c0af3a15735f4b4fac9720d328125cbe449dee2933bab4c55392c03.jpg) +Figure 3. Example of the instruction for LMMs. When both images from tail classes and textual templates are input into LMMs, textual descriptions corresponding to the images can be obtained. By repeatedly performing this operation on the training data, we convert abstract image descriptions into concrete textual descriptions. Finally, we acquire the current textual descriptions list corresponding to each class. + +we aim to leverage the image-generative ability of the text-to-image (T2I) method to generate the images from the tail-class descriptions list. In detail, we employ T2I to generate images based on the descriptions list, denoted as: + +![](images/7b4d90bf01494086fb2f59df7ac904efe708fda5edd98f816352bfe738aa1a82.jpg) +Figure 4. Example of the instruction for LLMs. LTGC inputs the existing textual descriptions list to LLMs, which continually extends it with new distinctive features and scene information. During multiple iterations, LTGC generates a new extended textual descriptions list for each class. + +$$ +i _ {n} ^ {y} = \mathrm {T 2 I} \left(d _ {n} ^ {y}\right), \text {w h e r e} n \in \{1, \dots , N \} \tag {1} +$$ + +where $n \in N$ denotes the $n$ -th generated descriptions for class $y$ , and $i_n^y$ denotes the image generated by T2I model based on $d_n^y$ . However, some generated images may not be of sufficiently high quality to accurately represent their desired classes, as the lower-quality images. As shown in Fig. 7, lower-quality images will output poorly distinctive features, resulting in more confusion for the class. Therefore, using these generated images directly in training may result in disrupting the model's prediction of the tail classes. + +![](images/6becc86c723c3fcee6af0bc5992fcd5c76c29def6417ebc0a1f4f706c92e7d2d.jpg) +Figure 5. Illustration of the proposed iterative evaluation module framework. This module detects lower-quality images through the similarity score $S$ computed by images and their corresponding class feature template. Then the textual descriptions corresponding to lower-quality images are re-input into LLMs for refinement. Finally, the refined textural descriptions are fed into the T2I model for regeneration. + +To address the challenge of lower-quality images that fail to represent their desired classes accurately, we aim to automatically detect these images, refine their textual descriptions, and regenerate them accordingly. To achieve this, drawing inspiration from the idea that humans improve their understanding through peer feedback, we investigate the possibility of a large model also enhancing its output by integrating feedback from another fundamental model. Therefore, we propose a iterative evaluation module, as illustrated in Fig. 5. This module utilizes the fundamental model, CLIP, to serve as the 'feedback giver' selected for its proficiency in understanding images and effectively connecting them with textual contexts [32]. It provides feedback about the lower-quality images to LLMs, which assists in refining their corresponding descriptions. + +Furthermore, to facilitate CLIP in grasping the represented features in images, a straightforward way is to check the degree of match between the image and its class template: "A photo of a [y]." However, due to the fact that CLIP's text encoder may contain less feature knowledge of the tail classes (e.g., as presented on the iNaturalist 2018 dataset with poor zero-shot recognition performance [13]), we aim to design a new textual feature template for each class, highlighting its unique features. Then CLIP employs it to determine the correspondence between the template and the image. Drawing inspiration from the chain-of-thought [30, 46] concept, which suggests that LMMs perform better when provided with additional clues, we further guide LMMs in summarizing the most distinctive features of each class. Specifically, after obtaining the extended tail-class descriptions list, we obtain the class feature template $C_y$ that contains the most distinctive features by the following [Prompt 4] and Template 2 for the given class $y$ : [Prompt 4]: "Please use + +Template 2 to summarize the most distinctive features of class $[y]$ ". " Template 2: "A photo of the class [y] with {feature 1} {feature 2} {...}. " This process could guide the LMMs to rethink and summarize the most distinctive features and produce the class feature template $C_y$ for each class $y$ . Furthermore, since we focus on checking whether the feature information of an image could represent its desired class, we do not include scene information in the class feature template. Then we introduce the details of the iterative evaluation module. + +As shown in Fig. 5, our iterative evaluation module conducts the following three steps iteratively: (1) Detection. To identify lower-quality images, we first employ utilize the strong capability of aligning images with text of CLIP to match the feature of generated images $i_n^y$ and its class feature template $C_y$ by cosine similarity metric as follows: + +$$ +S = \operatorname {E n c o d e r} _ {\text {v i s}} \left(i _ {n} ^ {y}\right) \cdot \operatorname {E n c o d e r} _ {\text {t e x t}} \left(C _ {y}\right), \tag {2} +$$ + +where $\mathsf{Encoder}_{\mathrm{vis}}$ denotes the CLIP's visual encoder, and $\mathsf{Encoder}_{\mathrm{text}}$ denotes the text encoder of the CLIP. The image $i_n^y$ is detected as lower quality and filtered out if its similarity score $S$ is below a threshold $\mu$ . (2) Refinement. Next, also inspired by chain-of-thought [46], we propose the following prompt for refine its descriptions. Specifically, if $i_n^y$ is identified as a lower quality image, we prompt LLMs to refine its corresponding description $d_n^y$ to more accurately represent the intended class $y$ , drawing on feedback provided by CLIP, i.e., [Prompt 5]: "This description $d_n^y$ doesn't seem to be representative of the class $[y]$ . Could you refine it to enhance the distinctive features of class $[y]?"" (3)$ + +Re-generation. Finally, the image $i_n^y$ is regenerated by the T2I model according to the improved textual description. + +By iteratively applying these three stages, we employ CLIP in each cycle to furnish precise feedback to the LLMs, thereby steering the refinement of descriptions to better align with our desired direction. These refined descriptions ensure the production of images that more accurately embody the characteristics of each class. + +# 3.2. BalanceMix + +With generated tailed images, the final problem is how to efficiently use these generated data and original data to perform long-tailed recognition well. Due to the domain gap between the generated data and the original long-tail data [38], we propose a method named BalanceMix to handle this challenge. We first define the original data and generated data as $\mathcal{D}_o$ and $\mathcal{D}_g$ . Then BalanceMix balance-sample [5] an image $x_i$ from $\mathcal{D}_o$ and sample an image $x_j$ from $\mathcal{D}_g$ . Meanwhile, it mixes the images $x_i$ and $x_j$ and their corre + +sponding labels, denoted as: + +$$ +\widetilde {x} = \lambda \odot x _ {i} + (\mathbf {1} - \lambda) \odot x _ {j}, \tag {3} +$$ + +$$ +\widetilde {y} = \lambda \odot y _ {i} + (\mathbf {1} - \lambda) \odot y _ {j}, \tag {4} +$$ + +where $\lambda$ in Beta (0,1) distribution. Finally, we fine-tune the CLIP's vision encoder with LORA [18] on all mixed data pairs $(\widetilde{x},\widetilde{y})$ for efficient long-tail recognition. + +# 4. Experiments + +We present the experimental results on three widely used datasets in long-tailed recognition, including ImageNet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. Moreover, we undertake ablation studies specifically on the ImageNet-LT and iNaturalist 2018 datasets to gain deeper insights into the performance of our method. The experimental results of the comparison methods are taken from their original paper, and our results are averaged over three experiments. + +# 4.1. Implementation details. + +Evaluation Setup. In all experiments, we evaluate and report top-1 accuracy on their corresponding test set. We also report accuracy on three splits of the classes: Many-shot (more than 100 images), Medium-shot (20 to 100 images), and Few-shot (less than 20 images) [20]. + +Method Implementation. In our LTGC, we incorporate diverse and specialized knowledge from the off-the-shelf large models. Specifically, for LMM, we use the GPT-4V (ision) [50] version of ChatGPT. For LLM, we use the GPT-4 version of ChatGPT. For T2I, we use DAll-E [33]. For the pre-trained CLIP [32], we use ViT-B/32 [14] for its visual encoder and the architecture introduced in [31] as the text encoder. In LLM's self-reflection module, we set the maximum number $K_{y}$ to 100, 300, and 800 for iNaturalist 2018, ImageNet-LT, and Place-LT, respectively. In the iterative evaluation module, the threshold $\mu$ is set at 0.8 for ImageNet-LT and Place-LT, and at 0.6 for iNaturalist. In Appendix, we provide a detailed ablation and discussion on the choice of LMM and LLM models, as well as parameters used in the self-reflection process. + +# 4.2. Comparisons with SOTA on Benchmarks + +In this section, we compare our proposed LTGC model with state-of-the-art (SOTA) methods on three benchmarks, including Imagenet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. To ensure a fair comparison, we primarily focus on methods based on CLIP, as they also leverage knowledge pretrained on large-scale datasets. These baselines include CLIP Zero-Shot [32, 36] and CLIP Finetune [36], as well as CLIP-based long-tail recognition approaches such as VL-LTR [36], LPT [13], and RAC [28]. Additionally, + +we report comparisons with traditional methods (without CLIP) on the challenging and fine-grained, large-scale long-tail dataset, iNaturalist 2018. These have proven the effectiveness of our method in different scenarios. + +Results on Imagenet-LT. In Tab. 1, we observe that our LTGC models are superior to other CLIP-based LT methods. For example, our method achieves an overall accuracy of $80.6\%$ , which is $3.4\%$ higher than the existing SOTA method VT-LTR[36]. Moreover, the overall accuracy of our method marginally surpasses the results on the full ImageNet (i.e., $80\%$ [36]). + +Results on Places-LT. Tab. 1 shows that compared to other CLIP variant methods, LTGC achieves an overall accuracy of $54.1\%$ and an few-shot accuracy of $52.1\%$ , surpassing the LPT [13] by $4.0\%$ and $5.2\%$ , respectively. Even compared with VL-LTR [36] and RAC [28], which have extra data in training and testing, our LTGC achieves remarkable results. + +Results on iNaturalist 2018. In this part, we employ LTGC on a large and fine-grained dataset, iNaturalist 2018. Tab. 2 presents the quantitative results. LTGC leverages the rich knowledge of LLMs and significantly outperforms traditional deep-learning approaches for long-tail recognition. In addition, LTGC attains an overall accuracy of $82.5\%$ and a few-shot accuracy of $82.6\%$ , outperforming all existing SOTA methods based on CLIP. In particular, LTGC also surpasses the retrieval augmented method, RAC [28], by $2.3\%$ . + +# 4.3. Compare with different methods of LMMs. + +In the experiments, our LTGC employs GPT-4V (ision) to describe a given image to obtain a text-based feature description. A straightforward way for long-tail recognition is by querying LMMs for the category of the given image. To evaluate the ability of these methods to recognize the long tail images, we constructed the following baselines of LMMs for comparison: MiniGPT4 [58], MiniGPT4-v2 [7], LENS [3], and GPT-4V (ision) [50]. To perform a fair evaluation, we provide the label list [Class 1, Class 2, ..., Class Y] for each dataset, where $Y$ is the number of classes. + +The results presented in Tab. 3, show that our method significantly outperforms the baselines that directly query the LMMs. More implementation details are discussed in Appendix. + +# 4.4. Analysis and Ablation Study + +Effectiveness of the BalanceMix module. To evaluate the effectiveness of the two proposed designs, we conduct experiments on three different variants: 1) w/o BalanceMix: This variant is directly fine-tuned using the generated data and the original training data. 2) Balanced Sample: This variant utilizes generated data to perform balanced sampling [45] for training. 3) Mixup: This variant employs generated data with Mixup [52] without performing balanced + +Table 1. Comparison with SOTA methods on ImageNet-LT and Places-LT. + +
DatasetImageNet-LTPlaces-LT
FewAllFewAll
CLIP Zero [36]58.659.840.138.0
CLIP Finetune [36]34.560.522.739.7
VL-LTR [36]59.377.242.050.1
RAC [28]--41.847.2
LPT [13]--46.950.1
LTGC(Ours)70.580.652.154.1
+ +Table 2. Comparison with SOTA methods on iNaturalist 2018. + +
MethodManyMediumFewAll
Softmax74.766.360.064.7
LADE [16]64.447.734.352.3
RIDE [42]71.570.071.671.8
PaCo [11]69.573.473.073.0
MDCS [56]76.575.575.275.6
CLIP Zero [36]6.13.32.93.4
CLIP Finetune [36]76.674.170.272.6
VL-LTR [36]---76.8
RAC [28]75.980.581.080.2
LPT [13]--79.376.1
LTGC(Ours)77.583.982.682.5
+ +Table 3. Comparison with different LMMs' methods on ImageNet-LT and iNaturalist2018. + +
MethodImageNet-LTiNaturalist 2018
LENS [3]69.517.4
MiniGPT4 [58]60.420.9
MiniGPT4-v2 [7]68.527.1
GPT-472.164.3
Ours80.682.5
+ +sampling. The results are shown in Tab. 4, demonstrating that our method significantly outperforms the other variants. Although our self-reflection and iterative evaluation modules already ensure the diversity and high quality of the generated images, there still exists a domain gap between the generated and original data [38]. This domain gap could exacerbate the long-tail effect on the test set (discussed in Appendix). A simple balanced sample [45] approach fails to solve the class imbalance caused by the mixing of original and generated data, while the Mixup [52] alone cannot address inter-class imbalances. To this end, our BalanceMix module combines the strengths of both methods, making the generated data well-suited for our framework. + +Table 4. Evaluation on the effectiveness of the BalanceMix. + +
MethodImageNet-LTiNaturalist
w/o BalanceMix58.369.5
Balanced sample [45]63.973.8
Mixup [52]73.475.2
Ours80.682.5
+ +![](images/98e5a1c89b9bc5f0305aa58c42bcb82eeea9a12cfe7f73b130495d906cc0d726.jpg) +Figure 6. The visualization of generated images: The template "A photo of the class [y]" and LTGC. Each row represents a different class. The four images on the left are generated using the simple template "A photo of the class [y]," which results in images with uniform poses and plain backgrounds. The four images on the right are from the proposed LTGC and demonstrate the diversity of classes. + +![](images/a4ca9b59496760d42a9cb6e31a6e56464d94df592570f291ef5e0c8367d983df.jpg) +Figure 7. The visualization of the images generated before and after passing the iterative evaluation module. The top row displays images that were filtered out, while the bottom row shows images regenerated by T2I after refining their corresponding descriptions. More visualizations are in the Appendix. + +# 4.5. Visualization + +Visualization of Generated Images: Template "A photo of the class [y]" vs our LTGC. Furthermore, we compare the images generated by LTGC with a simple prompt, "A photo of the class [y]" for the T2I model. As Fig. 6 illustrates, LTGC generates more accurate and diverse images compared to those generated by the simple prompt. For example, in the absence of control over the category description, the T2I model generates red features that do not belong to class Aquilegia Pubescens. In addition, with rich text for image generation, our approach also generates more diverse and accurate images compared to images generated by a simple template. + +Visualization on iterative evaluation module. We compare images before and after refinement using the iterative evaluation module, and the visualization results are shown in Fig. 7. It shows that images before refinement often possess ambiguous semantic information, and the features of the corresponding classes are not distinct. However, after the refinement process, the quality of the images is substantially enhanced, and the distinctive features of the corresponding classes become more pronounced. + +# 5. Conclusion + +This paper introduces a novel generative and fine-tuning framework, named LTGC, to address the challenge of long-tail recognition. LTGC leverages the abundant potential knowledge embedded in large-scale models to generate diverse data for tail categories. The framework incorporates innovative designs to ensure the quality of the generated data and to fine-tune the model efficiently using both the generated and original data. The experimental results indicate that LTGC surpasses existing state-of-the-art methods on well-known long-tail benchmarks. In the future, we will explore the robustness of large-scale models for application in other areas, such as long-tail semantic segmentation and dissent detection. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China under Grant No.62271034. This project is also supported by the Ministry of Education, Singapore, under the SUTD Kickstarter Initiative Project (SKI 2021_02_06). Additionally, I would like to express my gratitude to Haoxuan Qu and He Huang for their valuable suggestions. + +# References + +[1] Emanuel Sanchez Aimar, Arvi Jonnarth, Michael Felsberg, and Marco Kuhlmann. Balanced product of calibrated experts for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19967-19977, 2023. 1, 2 +[2] Martin Arjovsky, León Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 3 +[3] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 7 +[4] Michel Besserve, Arash Mehrjou, Rémy Sun, and Bernhard Schölkopf. Counterfactuals uncover the modular structure of deep generative models. In International Conference on Learning Representations, 2019. 3 +[5] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 2, 6 +[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. arXiv preprint arXiv:1906.07413, 2019. 1, 2 +[7] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 7 +[8] Xiaohua Chen, Yucan Zhou, Dayan Wu, Wanqian Zhang, Yu Zhou, Bo Li, and Weiping Wang. Imagine by reasoning: A reasoning-based implicit semantic data augmentation for long-tailed classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 356-364, 2022. 1, 2 +[9] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 95–110. Springer, 2020. 2, 3 +[10] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in gpt-4v(ison): Bias and interference challenges, 2023. 2 +[11] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 2, 7 +[12] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 2 +[13] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng + +Zuo. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022. 2, 3, 5, 6, 7 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[15] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 2, 3 +[16] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 7 +[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The naturalist challenge 2017 dataset. CoRR, abs/1707.06642, 2017. 2, 6 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[19] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5): 429-449, 2002. 1, 2 +[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 6 +[21] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 2 +[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3 +[23] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE transactions on neural networks and learning systems, 29(8):3573-3587, 2017. 1, 2 +[24] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1, 2, 3 +[25] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 2 +[26] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE + +transactions on pattern analysis and machine intelligence, 2018. 1, 2 +[27] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 6 +[28] Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, and Anton van den Hengel. Retrieval augmented classification for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6959-6969, 2022. 2, 6, 7 +[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 1, 2 +[30] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. arXiv preprint arXiv:2309.12780, 2023. 2, 3, 5 +[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6 +[33] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 6 +[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530, 2023. 2, 3 +[35] Kaihua Tang, Mingyuan Tao, Jiaxin Qi, Zhenguang Liu, and Hanwang Zhang. Invariant feature learning for generalized long-tailed classification. In European Conference on Computer Vision, pages 709-726. Springer, 2022. 1, 3 +[36] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 2, 3, 6, 7 +[37] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3 +[38] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 6, 7 + +[39] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 943-952, 2021. 1, 2 +[40] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 2 +[41] Tan Wang, Zhongqi Yue, Jianqiang Huang, Qianru Sun, and Hanwang Zhang. Self-supervised learning disentangled group representation as feature. Advances in Neural Information Processing Systems, 34:18225-18240, 2021. 3 +[42] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020. 1, 2, 7 +[43] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation, 2023. 2 +[44] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 7032-7042, 2017. 1 +[45] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in neural information processing systems, 30, 2017. 7 +[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2, 5, 6 +[47] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 247–263. Springer, 2020. 1 +[48] Cihang Xie and Alan Yuille. Intriguing properties of adversarial training at scale. arXiv preprint arXiv:1906.03787, 2019. 1, 2 +[49] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in neural information processing systems, 33:19290-19301, 2020. 2 +[50] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 2, 3, 4, 6, 7 +[51] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 2, 3 +[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 7 + +[53] Yifan Zhang, Bryan Hooi, Lanqing Hong, and Jiashi Feng. Test-agnostic long-tailed recognition by test-time aggregating diverse experts with self-supervision. arXiv preprint arXiv:2107.09249, 2021. 2 +[54] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[55] Qihao Zhao, Yangyu Huang, Wei Hu, Fan Zhang, and Jun Liu. Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In The Eleventh International Conference on Learning Representations, 2022. 2 +[56] Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, and Jun Liu. Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11597-11608, 2023. 1, 2, 7 +[57] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16489-16498, 2021. 2 +[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 7 +[59] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 2 +[60] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation, 2023. 2, 3 \ No newline at end of file diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/images.zip b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0998a4827ee87339b12970f2aff12402159f5d07 --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f948bbb41d63aabe5494728519a0bbda364c19ca8483658a1caa784390dda0dc +size 568520 diff --git a/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/layout.json b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d9553c6951c81565287b98ace0d993631641a1b8 --- /dev/null +++ b/2024/LTGC_ Long-tail Recognition via Leveraging LLMs-driven Generated Content/layout.json @@ -0,0 +1,7566 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 539, + 120 + ], + "type": "text", + "content": "LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": "Qihao Zhao" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": ", Yalun Dai" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": ", Hao Li" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": ", Wei Hu" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": ", Fan Zhang" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": ", Jun Liu" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": "Beijing University of Chemical Technology, China \n" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": "Singapore University of Technology and Design, Singapore \n" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": "Nanyang Technological University, Singapore \n" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 127, + 142, + 465, + 213 + ], + "type": "text", + "content": "Northwestern Polytechnical University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 470 + ], + "type": "text", + "content": "Long-tail recognition is challenging because it requires the model to learn good representations from tail categories and address imbalances across all categories. In this paper, we propose a novel generative and fine-tuning framework, LTGC, to handle long-tail recognition via leveraging generated content. Firstly, inspired by the rich implicit knowledge in large-scale models (e.g., large language models, LLMs), LTGC leverages the power of these models to parse and reason over the original tail data to produce diverse tail-class content. We then propose several novel designs for LTGC to ensure the quality of the generated data and to efficiently fine-tune the model using both the generated and original data. The visualization demonstrates the effectiveness of the generation module in LTGC, which produces accurate and diverse tail data. Additionally, the experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tailed benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 511, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 287, + 668 + ], + "type": "text", + "content": "In the real world, data often exhibits a long-tailed distribution, posing significant challenges for computer vision recognition [44, 54]. These challenges include (1) Class Imbalance: within the dataset, some classes (termed \"head\" classes) are abundantly represented, while others (termed \"tail\" classes) have few samples. This imbalanced distribution during training may cause the model to attend more on the head classes, neglecting the tail classes [42]. (2) Tail Data Scarcity: Data scarcity refers to tail classes having extremely limited samples, which lack diversity and are insufficient to effectively train a model [8, 24]. It prevents a model's ability to learn the feature invariant, which is necessary to recognize these categories correctly [35]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 239, + 541, + 533 + ], + "blocks": [ + { + "bbox": [ + 312, + 239, + 541, + 533 + ], + "lines": [ + { + "bbox": [ + 312, + 239, + 541, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 541, + 533 + ], + "type": "image", + "image_path": "31f361492f3359d2f385fee6d7a43ccc82031d239f95078b919e7867dc0958ff.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 540, + 547, + 617 + ], + "lines": [ + { + "bbox": [ + 304, + 540, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 547, + 617 + ], + "type": "text", + "content": "Figure 1. Top: Here, our LMMs use ChatGPT. For the Trogon Rufus category, when we asked ChatGPT, \"What species is in the picture?\" we did not get the expected answer due to the complexity of the question. Middle and Down: In contrast, when we asked some easy questions for the ChatGPT, \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus.\" It could accurately answer these questions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "To address these challenges, numerous approaches have emerged within the field of long-tailed recognition, such as re-sampling [5, 19], loss re-weighting [1, 6, 12, 23, 26, 27, 29, 39, 48], ensemble learning [42, 47, 56], decoupling [20], and contrastive learning [11, 25, 40, 59]. Their primary objective is to balance the decision boundaries and parameter weights of the model to improve long-tail recognition." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 126, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 126, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 126, + 684 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 684, + 287, + 713 + ], + "type": "text", + "content": "†The Corresponding author is with the College of Information Science and Technology and the Interdisciplinary Research Center for Artificial Intelligence, Beijing University of Chemical Technology, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19510" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 312 + ], + "type": "text", + "content": "However, these methods encounter bottlenecks due to the scarcity and limited diversity features of tail-class data [24]. To obtain more diverse representations for tail classes, two types of methods appear: Some methods increase tail-class diversity through data augmentation [9, 24, 52, 55]. Others increase tail-class diversity by transferring features from related classes [8] or large pre-trained models (e.g., CLIP [32]). These transfer learning methods [13, 28, 36] based on CLIP have recently shown great potential in boosting long-tail recognition. The LPT [13] fine-tunes pre-trained models to adapt to target datasets. The VL-LTR [36] leverages text features from CLIP to augment the learning of image features. However, these methods find it hard to obtain the correct and desired knowledge for tail categories. For example, in the real-world data iNauralist 2018 [17], the category Aquilegia Pubescens is characterized only by white or pale yellow colors and spurred shapes. If we transfer similar category features or blindly augment semantic information, such as red color semantics, this category will become more confusing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 315, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 685 + ], + "type": "text", + "content": "Recently, Large Language Models (LLMs) (e.g., ChatGPT) and Large Multimodal Models (LMMs) (e.g., GPT-4V (ChatGPT with Vision [50]) and Minigpt4 [7, 58]) due to their wealth of potential knowledge, have been leveraged for a variety of downstream tasks, such as robot task plans [34], open-set object recognition [30], automated robot learning [43] and various visual reasoning tasks [50, 51, 60]. However, due to the bias of imbalanced training data, LMMs perform poorly on some complex tasks [10]. We also find this dilemma in long-tail recognition as shown in Fig. 1. When we asked LMMs about an image in the category of Trogon rufus: \"What species is in the image?\", we did not get the expected answer. In contrast, when we ask some easy questions about this image: \"Please describe the image.\" or \"Please describe the distinctive features of Trogon Rufus\", we get the desired answer. This suggests that although large models underperform in long-tail recognition (e.g., it is difficult to align complex image features with labels.), they still contain desired potential knowledge (e.g., correctly describing images and providing the knowledge of species). Inspired by this, we aim to leverage the abundant potential knowledge of several large models to address the challenges of long-tailed data scarcity and perform long-tailed recognition well. Nevertheless, effectively leveraging large models to encounter these challenges is not trivial: (1) The potential nature of knowledge within large models makes it difficult to extract the desired knowledge that facilitates long-tail recognition tasks. (2) Due to the domain gap between the generated data and the original data [38], effectively using these hybrid data remains an open issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "To this end, we propose a novel Long-Tail recognition framework via Generated Content, denoted as LTGC," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 359 + ], + "type": "text", + "content": "which is illustrated in Fig. 2. Our LTGC aims to leverage the capabilities of large models for generating explicitly diverse content [15] tailored to the long-tail classes and incorporate novel designs to enhance long-tail recognition. Firstly, motivated by the wealth of potential knowledge in large models and the fact that texts are more controllable, we employ them to expand the tail classes. Specifically, to produce more diverse and controllable tail data, we follow the rule of analyzing existing features before generating absent features. To achieve this, LTGC utilizes LMMs to analyze images already in the tail classes to obtain the existing tail-class descriptions list. Then LTGC inputs the existing descriptions list to LLMs to obtain the desired extended tail-class descriptions list for images that are absent from the existing tail classes. Secondly, LTGC utilizes the text-to-image (T2I) model (e.g., DALL-E [33]) to transform these textual descriptions into images. Moreover, inspired by the benefits of chain-of-thought [46], we propose a self-reflection and iterative evaluation module for this process to ensure the diversity and quality of the generated content. Finally, inspired by the advantages of Mixup [52] in merging different image domains, we propose the BalanceMix module to address domain shifts of generated images for the fine-tuning process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 319, + 360, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 360, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 319, + 360, + 545, + 371 + ], + "type": "text", + "content": "In summary, the contributions of our work are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 372, + 545, + 502 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 307, + 372, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 407 + ], + "type": "text", + "content": "1. In the first time, we propose a novel framework via generated content, LTGC, which leverages the power of large models to address long-tail recognition challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 545, + 455 + ], + "type": "text", + "content": "2. We design a series of novel modules to tackle the tail-class image scarcity problem and design the BalanceMix module to efficiently fine-tune the model using the generated data and the original data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 455, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 545, + 502 + ], + "type": "text", + "content": "3. Experimental results demonstrate that our LTGC outperforms existing state-of-the-art methods on popular long-tail benchmarks. Also, the visualization illustrates the diversity and controllability of our generated tail images." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 392, + 526 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Long-Tail recognition. Many approaches have been proposed in recent years to address the long-tail recognition challenge, such as resampling [5, 19], loss rebalancing [1, 6, 12, 23, 26, 27, 29, 39, 48] to increase the model's focus on the tail categories, ensemble learning [42, 53, 56] to improve the ability to recognize tail samples by enhancing model performance, contrast learning [11, 21, 25, 40, 49, 59] and calibration learning [9, 57] to improve the model's decision boundaries, but due to the scarcity of data on the tails, the model still doesn't have enough features to learn class representations for long-tail recognition. Recently, data augmentation-based methods [8, 9, 24] proposed to handle scarcity-tailed data by augmenting from related features. However, these methods find it difficult to obtain the correct knowledge for tail classes. Other approaches" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19511" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 108 + ], + "type": "text", + "content": "[13, 36] have achieved good results by transfer learning [37] the feature of the visual-language pre-training model (e.g., CLIP [32]) for long-tailed recognition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 203 + ], + "type": "text", + "content": "Different from the previous methods, in this paper, from a novel perspective, we explore a generative and fine-tuning framework to handle long-tail recognition by leveraging the rich potential knowledge of large models. Specifically, we design several novel modules to leverage the correct knowledge for generating tail data and perform efficient fine-tuning with the generated data to boost long-tail recognition." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 286, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 335 + ], + "type": "text", + "content": "Large Models. Recently, several large models have emerged, including large language models (e.g., ChatGPT [50]), large multi-modal models (e.g., GPT-4 (ision) [50], MiniGPT [58]) and large generated models (e.g., DALLE [33]). These expansive models are repositories of extensive knowledge [15] and have undergone scrutiny in diverse applications, such as robot task plans [34], open-set object recognition [30], contextual object detection [51], video generation [22], and few-shot segmentation [60]. In this work, we leverage the rich potential knowledge of these large models to handle the long-tail recognition challenge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 102, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 102, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 102, + 357 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 286, + 508 + ], + "type": "text", + "content": "To perform long-tail recognition well, the challenge is making the model learn diverse representations from tail categories and address imbalances across all categories. In this paper, we propose a novel long-tail recognition framework, LTGC, to handle these challenges. As depicted in Fig. 2, LTGC employ potential knowledge from various off-the-shelf large-scale models to generate and iteratively assess the quality and diversity of tail classes (Sec. 3.1). Furthermore, LTGC proposes the BalanceMix module to facilitating the fine-tuning process with the generated and original datasets (Sec. 3.2). These modules are described in detail below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 217, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 217, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 217, + 528 + ], + "type": "text", + "content": "3.1. Diverse Tail Images Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 286, + 713 + ], + "type": "text", + "content": "To learn diversity representations from tail categories, the previous methods have used data augmentation [9, 24] or transfer learning [13, 36]. However, these works find it difficult to obtain the correct and desired diversity knowledge for tail categories. Inspired by the common-sense knowledge in the LLMs and the fact that textual descriptions are more controllable [33, 50], LTGC takes advantage of these to control the detail and diversity of the generated tail classes. Firstly, to better generate diverse images and control the image detail, LTGC aims to generate images that are absent in the original tail data and represent these images in textual form. Specifically, LTGC employs LMMs to analyze the original tail data and obtain the existing tail-class descriptions list. Then LTGC leverages the common-sense knowledge of LLMs to obtain extended tail-class descrip" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 131 + ], + "type": "text", + "content": "tions based on the existing tail-class descriptions list. Finally, as images are more suitable as training data for long-tail recognition tasks, we utilize the text-to-image module to generate diverse images based on these tail-class descriptions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 161, + 536, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 161, + 536, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 536, + 173 + ], + "type": "text", + "content": "3.1.1 Obtaining Existing Tail-class Descriptions List" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 185, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 472 + ], + "type": "text", + "content": "In order to ensure the diversity of tail-class images, we first analyze the feature information of the original tail-class images before generating new ones. It guarantees that the content of the generated images is distinct from that of the existing tail-class images. Moreover, due to textual descriptions being more controllable [33], we utilize the textual descriptions to control the detail and diversity of the generated images. To achieve this, our LTGC employs Language Model Multimodals (LMMs), such as GPT-4V (Vision) [50], to analyze the feature information and extract textual descriptions of existing tail classes. During this process, the textual responses from LMMs could be varied and sometimes redundant, which may impede the generation of the desired images. Therefore, we employ textual templates to constrain the responses of LMMs, aiming to unify the textual description formats. Inspired by an image of an object that can be fully described or generated by its class and a list of features [4, 41], we design the textual template to include the given class and its features. Furthermore, the introduction of variations in the scenes plays a crucial role in enhancing the model's ability to generalize [2], a problem that's particularly acute in long-tailed datasets [35]. Therefore, the textual information of scenes is also important for image generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "To this end, we design the LMMs' response Template 1 for a given class " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " as follows: \"A photo of the class [y], {with distinctive features}{in specific scenes}.\", which include the given class, its distinctive features, and specific scenes or environments. With the response template, for a given " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_y" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " number class " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": ", we sequentially feed these tail-class images into LMMs along with an instructing [Prompt 1] to analyze the features of these images: \"Please use the Template 1 to briefly describe the image of the class [y].\" This process is illustrated in Fig. 3, where " + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": " is the given class label. As shown, by formulating the description using the template, LMMs would automatically replace the class, features, and background in the response. After performing this process for all images of each class, we compile a list of text descriptions corresponding to each class. Then we intend to extend the tail-class descriptions list with the existing tail-class descriptions list to generate the images lacking in the tail classes." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "19512" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 521, + 315 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 521, + 315 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 521, + 315 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 521, + 315 + ], + "type": "image", + "image_path": "120199c534a605dc42736c8f510eb18b7a984b55c76a75c0830643deeef4af6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 322, + 545, + 367 + ], + "lines": [ + { + "bbox": [ + 45, + 322, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 45, + 322, + 545, + 367 + ], + "type": "text", + "content": "Figure 2. Overall framework of LTGC. LTGC first employs LMMs to analyze the existing tail data to obtain the existing tail-class descriptions list. Then it inputs the list into LLMs to analyze the absent features of the tail classes and employs the T2I model to generate diverse images. Moreover, our designed self-reflection and iterative evaluation modules ensure the diversity and quality of the tail data. Finally, LTGC employs the BalanceMix module to fine-tune the CLIP's visual encoder with the extended and original data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 377, + 284, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 284, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 284, + 389 + ], + "type": "text", + "content": "3.1.2 Obtaining Extended Tail-class Descriptions List" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 400, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 400, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 45, + 400, + 287, + 604 + ], + "type": "text", + "content": "In this section, we aim to analyze features missing from tail classes and enrich the descriptions of these classes based on the existing tail-class descriptions list. To accomplish this, inspired by the rich common-sense knowledge of Large Language Models (LLMs), such as ChatGPT [50], we extend the tail-class descriptions list through a two-step process: 1) Inputting the existing descriptions list into LLMs, and 2) Designing the Prompt 2 to guide LLMs in generating the desired descriptions for images that are absent in the given tail class " + }, + { + "bbox": [ + 45, + 400, + 287, + 604 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 45, + 400, + 287, + 604 + ], + "type": "text", + "content": ": \"Besides these descriptions mentioned above, please use the Template 1 to list other possible {distinctive features} and {specific scenes} for the class [y],\" which is illustrated in Fig. 4. For each class, we repeat the above two-step process, and then we obtain the tail-class descriptions list for all tail classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "In addition, in order for these generated descriptions to better complement each tail class, we encourage LLMs to generate descriptions of sufficient number and diversity. To meet this goal, we introduce a self-reflection module in this process, aiming to guide LMMs in rethinking if there are any features or scenes that are missed or repeated. It includes two key designs: a number-checking module and a repetition-checking module. (1) The number-checking module aims to guide LMMs in rethinking whether there" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": "are other missing features or scenarios. To achieve this, after posing the initial [Prompt 2] for class " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": ", we update the extended descriptions list and re-ask LLMs the [Prompt 2] question, incorporating the newly acquired list. For each class " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": ", this iterative process of number-checking continues until a maximum number " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "K_y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": " of the tail class is achieved, where " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "K_y = M_y + N_y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "N_y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": " is the number of generated descriptions for class " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": ". (2) The repetition-checking module aims to guide LMMs in rethinking if there are other features or scenes that are repeated at the end of the number-checking iteration. Specifically, we input the extended descriptions list and the following [Prompt 3] of each class " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": " for LLMs' repetition checking: \"Please exclude any repetitive {distinctive features} and {specific scenes} for class " + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "inline_equation", + "content": "[y]" + }, + { + "bbox": [ + 304, + 377, + 545, + 641 + ], + "type": "text", + "content": " in this descriptions list.\" After that, LLM will filter out the ones with repeated features based on this prompt and return a new list of descriptions. Through the implementation of this two-step process and LLMs' self-reflection module, we have obtained a reduction in repeated descriptions and an increase in the diversity of the tail-class descriptions for each class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 658, + 484, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 484, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 484, + 670 + ], + "type": "text", + "content": "3.1.3 Transform Descriptions to Images" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Above, we obtain the extended tail-class descriptions list for each class " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{y}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". As the images are better adapted to perform visual recognition tasks, in this section," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19513" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 282, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 282, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 282, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 282, + 222 + ], + "type": "image", + "image_path": "35a579247c0af3a15735f4b4fac9720d328125cbe449dee2933bab4c55392c03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 308 + ], + "type": "text", + "content": "Figure 3. Example of the instruction for LMMs. When both images from tail classes and textual templates are input into LMMs, textual descriptions corresponding to the images can be obtained. By repeatedly performing this operation on the training data, we convert abstract image descriptions into concrete textual descriptions. Finally, we acquire the current textual descriptions list corresponding to each class." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 318, + 288, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 368 + ], + "type": "text", + "content": "we aim to leverage the image-generative ability of the text-to-image (T2I) method to generate the images from the tail-class descriptions list. In detail, we employ T2I to generate images based on the descriptions list, denoted as:" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 376, + 282, + 507 + ], + "blocks": [ + { + "bbox": [ + 53, + 376, + 282, + 507 + ], + "lines": [ + { + "bbox": [ + 53, + 376, + 282, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 282, + 507 + ], + "type": "image", + "image_path": "7b4d90bf01494086fb2f59df7ac904efe708fda5edd98f816352bfe738aa1a82.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 514, + 287, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 571 + ], + "type": "text", + "content": "Figure 4. Example of the instruction for LLMs. LTGC inputs the existing textual descriptions list to LLMs, which continually extends it with new distinctive features and scene information. During multiple iterations, LTGC generates a new extended textual descriptions list for each class." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 586, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 586, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 89, + 586, + 287, + 601 + ], + "type": "interline_equation", + "content": "i _ {n} ^ {y} = \\mathrm {T 2 I} \\left(d _ {n} ^ {y}\\right), \\text {w h e r e} n \\in \\{1, \\dots , N \\} \\tag {1}", + "image_path": "766a3c636049ca4f2e6925337242fb4e45018e1afb2b07a07c05962029af2bf5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "n \\in N" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "-th generated descriptions for class " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "i_n^y" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " denotes the image generated by T2I model based on " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "d_n^y" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ". However, some generated images may not be of sufficiently high quality to accurately represent their desired classes, as the lower-quality images. As shown in Fig. 7, lower-quality images will output poorly distinctive features, resulting in more confusion for the class. Therefore, using these generated images directly in training may result in disrupting the model's prediction of the tail classes." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 68, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 310, + 68, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 310, + 68, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 310, + 68, + 545, + 205 + ], + "type": "image", + "image_path": "6becc86c723c3fcee6af0bc5992fcd5c76c29def6417ebc0a1f4f706c92e7d2d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 211, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 289 + ], + "type": "text", + "content": "Figure 5. Illustration of the proposed iterative evaluation module framework. This module detects lower-quality images through the similarity score " + }, + { + "bbox": [ + 304, + 211, + 547, + 289 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 211, + 547, + 289 + ], + "type": "text", + "content": " computed by images and their corresponding class feature template. Then the textual descriptions corresponding to lower-quality images are re-input into LLMs for refinement. Finally, the refined textural descriptions are fed into the T2I model for regeneration." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 484 + ], + "type": "text", + "content": "To address the challenge of lower-quality images that fail to represent their desired classes accurately, we aim to automatically detect these images, refine their textual descriptions, and regenerate them accordingly. To achieve this, drawing inspiration from the idea that humans improve their understanding through peer feedback, we investigate the possibility of a large model also enhancing its output by integrating feedback from another fundamental model. Therefore, we propose a iterative evaluation module, as illustrated in Fig. 5. This module utilizes the fundamental model, CLIP, to serve as the 'feedback giver' selected for its proficiency in understanding images and effectively connecting them with textual contexts [32]. It provides feedback about the lower-quality images to LLMs, which assists in refining their corresponding descriptions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "text", + "content": "Furthermore, to facilitate CLIP in grasping the represented features in images, a straightforward way is to check the degree of match between the image and its class template: \"A photo of a [y].\" However, due to the fact that CLIP's text encoder may contain less feature knowledge of the tail classes (e.g., as presented on the iNaturalist 2018 dataset with poor zero-shot recognition performance [13]), we aim to design a new textual feature template for each class, highlighting its unique features. Then CLIP employs it to determine the correspondence between the template and the image. Drawing inspiration from the chain-of-thought [30, 46] concept, which suggests that LMMs perform better when provided with additional clues, we further guide LMMs in summarizing the most distinctive features of each class. Specifically, after obtaining the extended tail-class descriptions list, we obtain the class feature template " + }, + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "inline_equation", + "content": "C_y" + }, + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "text", + "content": " that contains the most distinctive features by the following [Prompt 4] and Template 2 for the given class " + }, + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 486, + 546, + 715 + ], + "type": "text", + "content": ": [Prompt 4]: \"Please use" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "19514" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "text", + "content": "Template 2 to summarize the most distinctive features of class " + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "inline_equation", + "content": "[y]" + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "text", + "content": "\". \" Template 2: \"A photo of the class [y] with {feature 1} {feature 2} {...}. \" This process could guide the LMMs to rethink and summarize the most distinctive features and produce the class feature template " + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "inline_equation", + "content": "C_y" + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 72, + 286, + 203 + ], + "type": "text", + "content": ". Furthermore, since we focus on checking whether the feature information of an image could represent its desired class, we do not include scene information in the class feature template. Then we introduce the details of the iterative evaluation module." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "text", + "content": "As shown in Fig. 5, our iterative evaluation module conducts the following three steps iteratively: (1) Detection. To identify lower-quality images, we first employ utilize the strong capability of aligning images with text of CLIP to match the feature of generated images " + }, + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "inline_equation", + "content": "i_n^y" + }, + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "text", + "content": " and its class feature template " + }, + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "inline_equation", + "content": "C_y" + }, + { + "bbox": [ + 47, + 205, + 286, + 277 + ], + "type": "text", + "content": " by cosine similarity metric as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 292, + 287, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 292, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 88, + 292, + 287, + 305 + ], + "type": "interline_equation", + "content": "S = \\operatorname {E n c o d e r} _ {\\text {v i s}} \\left(i _ {n} ^ {y}\\right) \\cdot \\operatorname {E n c o d e r} _ {\\text {t e x t}} \\left(C _ {y}\\right), \\tag {2}", + "image_path": "bad47ec8dc1e52a8478c510496f3f6843a58f7d0eac54e54e20ebde68d07fa3a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "\\mathsf{Encoder}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " denotes the CLIP's visual encoder, and " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "\\mathsf{Encoder}_{\\mathrm{text}}" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " denotes the text encoder of the CLIP. The image " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "i_n^y" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " is detected as lower quality and filtered out if its similarity score " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " is below a threshold " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": ". (2) Refinement. Next, also inspired by chain-of-thought [46], we propose the following prompt for refine its descriptions. Specifically, if " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "i_n^y" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " is identified as a lower quality image, we prompt LLMs to refine its corresponding description " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "d_n^y" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " to more accurately represent the intended class " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": ", drawing on feedback provided by CLIP, i.e., [Prompt 5]: \"This description " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "d_n^y" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": " doesn't seem to be representative of the class " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "[y]" + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "text", + "content": ". Could you refine it to enhance the distinctive features of class " + }, + { + "bbox": [ + 47, + 312, + 286, + 479 + ], + "type": "inline_equation", + "content": "[y]?\"\" (3)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 480, + 286, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 286, + 503 + ], + "type": "text", + "content": "Re-generation. Finally, the image " + }, + { + "bbox": [ + 47, + 480, + 286, + 503 + ], + "type": "inline_equation", + "content": "i_n^y" + }, + { + "bbox": [ + 47, + 480, + 286, + 503 + ], + "type": "text", + "content": " is regenerated by the T2I model according to the improved textual description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 505, + 286, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 286, + 575 + ], + "type": "text", + "content": "By iteratively applying these three stages, we employ CLIP in each cycle to furnish precise feedback to the LLMs, thereby steering the refinement of descriptions to better align with our desired direction. These refined descriptions ensure the production of images that more accurately embody the characteristics of each class." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 126, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 126, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 126, + 597 + ], + "type": "text", + "content": "3.2. BalanceMix" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": "With generated tailed images, the final problem is how to efficiently use these generated data and original data to perform long-tailed recognition well. Due to the domain gap between the generated data and the original long-tail data [38], we propose a method named BalanceMix to handle this challenge. We first define the original data and generated data as " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_o" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": ". Then BalanceMix balance-sample [5] an image " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_o" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " and sample an image " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "x_j" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_g" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": ". Meanwhile, it mixes the images " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "x_j" + }, + { + "bbox": [ + 47, + 605, + 286, + 713 + ], + "type": "text", + "content": " and their corre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 421, + 83 + ], + "type": "text", + "content": "sponding labels, denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 366, + 95, + 545, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 95, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 366, + 95, + 545, + 107 + ], + "type": "interline_equation", + "content": "\\widetilde {x} = \\lambda \\odot x _ {i} + (\\mathbf {1} - \\lambda) \\odot x _ {j}, \\tag {3}", + "image_path": "3a4f98722170bf8d8c248859ed23372c3ab13edc35e0693c2567bb9e25ee0813.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 129, + 545, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 129, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 367, + 129, + 545, + 143 + ], + "type": "interline_equation", + "content": "\\widetilde {y} = \\lambda \\odot y _ {i} + (\\mathbf {1} - \\lambda) \\odot y _ {j}, \\tag {4}", + "image_path": "8528578533e02505904878609a259b5475c10d427ecad26d99281a3db4358e77.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "text", + "content": " in Beta (0,1) distribution. Finally, we fine-tune the CLIP's vision encoder with LORA [18] on all mixed data pairs " + }, + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "inline_equation", + "content": "(\\widetilde{x},\\widetilde{y})" + }, + { + "bbox": [ + 305, + 148, + 545, + 185 + ], + "type": "text", + "content": " for efficient long-tail recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 196, + 386, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 196, + 386, + 209 + ], + "spans": [ + { + "bbox": [ + 306, + 196, + 386, + 209 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 217, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 312 + ], + "type": "text", + "content": "We present the experimental results on three widely used datasets in long-tailed recognition, including ImageNet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. Moreover, we undertake ablation studies specifically on the ImageNet-LT and iNaturalist 2018 datasets to gain deeper insights into the performance of our method. The experimental results of the comparison methods are taken from their original paper, and our results are averaged over three experiments." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 440, + 333 + ], + "type": "text", + "content": "4.1. Implementation details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 545, + 399 + ], + "type": "text", + "content": "Evaluation Setup. In all experiments, we evaluate and report top-1 accuracy on their corresponding test set. We also report accuracy on three splits of the classes: Many-shot (more than 100 images), Medium-shot (20 to 100 images), and Few-shot (less than 20 images) [20]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "text", + "content": "Method Implementation. In our LTGC, we incorporate diverse and specialized knowledge from the off-the-shelf large models. Specifically, for LMM, we use the GPT-4V (ision) [50] version of ChatGPT. For LLM, we use the GPT-4 version of ChatGPT. For T2I, we use DAll-E [33]. For the pre-trained CLIP [32], we use ViT-B/32 [14] for its visual encoder and the architecture introduced in [31] as the text encoder. In LLM's self-reflection module, we set the maximum number " + }, + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "inline_equation", + "content": "K_{y}" + }, + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "text", + "content": " to 100, 300, and 800 for iNaturalist 2018, ImageNet-LT, and Place-LT, respectively. In the iterative evaluation module, the threshold " + }, + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 399, + 545, + 578 + ], + "type": "text", + "content": " is set at 0.8 for ImageNet-LT and Place-LT, and at 0.6 for iNaturalist. In Appendix, we provide a detailed ablation and discussion on the choice of LMM and LLM models, as well as parameters used in the self-reflection process." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 586, + 521, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 521, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 521, + 600 + ], + "type": "text", + "content": "4.2. Comparisons with SOTA on Benchmarks" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "In this section, we compare our proposed LTGC model with state-of-the-art (SOTA) methods on three benchmarks, including Imagenet-LT [27], Places-LT [27], and iNaturalist 2018 [17]. To ensure a fair comparison, we primarily focus on methods based on CLIP, as they also leverage knowledge pretrained on large-scale datasets. These baselines include CLIP Zero-Shot [32, 36] and CLIP Finetune [36], as well as CLIP-based long-tail recognition approaches such as VL-LTR [36], LPT [13], and RAC [28]. Additionally," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19515" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "we report comparisons with traditional methods (without CLIP) on the challenging and fine-grained, large-scale long-tail dataset, iNaturalist 2018. These have proven the effectiveness of our method in different scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "text", + "content": "Results on Imagenet-LT. In Tab. 1, we observe that our LTGC models are superior to other CLIP-based LT methods. For example, our method achieves an overall accuracy of " + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "inline_equation", + "content": "80.6\\%" + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "text", + "content": " higher than the existing SOTA method VT-LTR[36]. Moreover, the overall accuracy of our method marginally surpasses the results on the full ImageNet (i.e., " + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 120, + 288, + 204 + ], + "type": "text", + "content": " [36])." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "content": "Results on Places-LT. Tab. 1 shows that compared to other CLIP variant methods, LTGC achieves an overall accuracy of " + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "inline_equation", + "content": "54.1\\%" + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "content": " and an few-shot accuracy of " + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "inline_equation", + "content": "52.1\\%" + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "content": ", surpassing the LPT [13] by " + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 46, + 204, + 288, + 276 + ], + "type": "text", + "content": ", respectively. Even compared with VL-LTR [36] and RAC [28], which have extra data in training and testing, our LTGC achieves remarkable results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "text", + "content": "Results on iNaturalist 2018. In this part, we employ LTGC on a large and fine-grained dataset, iNaturalist 2018. Tab. 2 presents the quantitative results. LTGC leverages the rich knowledge of LLMs and significantly outperforms traditional deep-learning approaches for long-tail recognition. In addition, LTGC attains an overall accuracy of " + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "inline_equation", + "content": "82.5\\%" + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "text", + "content": " and a few-shot accuracy of " + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "inline_equation", + "content": "82.6\\%" + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "text", + "content": ", outperforming all existing SOTA methods based on CLIP. In particular, LTGC also surpasses the retrieval augmented method, RAC [28], by " + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 276, + 288, + 395 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 404, + 272, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 272, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 272, + 417 + ], + "type": "text", + "content": "4.3. Compare with different methods of LMMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 423, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 542 + ], + "type": "text", + "content": "In the experiments, our LTGC employs GPT-4V (ision) to describe a given image to obtain a text-based feature description. A straightforward way for long-tail recognition is by querying LMMs for the category of the given image. To evaluate the ability of these methods to recognize the long tail images, we constructed the following baselines of LMMs for comparison: MiniGPT4 [58], MiniGPT4-v2 [7], LENS [3], and GPT-4V (ision) [50]. To perform a fair evaluation, we provide the label list [Class 1, Class 2, ..., Class Y] for each dataset, where " + }, + { + "bbox": [ + 46, + 423, + 287, + 542 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 46, + 423, + 287, + 542 + ], + "type": "text", + "content": " is the number of classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 543, + 287, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 591 + ], + "type": "text", + "content": "The results presented in Tab. 3, show that our method significantly outperforms the baselines that directly query the LMMs. More implementation details are discussed in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 204, + 612 + ], + "type": "text", + "content": "4.4. Analysis and Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Effectiveness of the BalanceMix module. To evaluate the effectiveness of the two proposed designs, we conduct experiments on three different variants: 1) w/o BalanceMix: This variant is directly fine-tuned using the generated data and the original training data. 2) Balanced Sample: This variant utilizes generated data to perform balanced sampling [45] for training. 3) Mixup: This variant employs generated data with Mixup [52] without performing balanced" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 334, + 102, + 518, + 184 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 91 + ], + "type": "text", + "content": "Table 1. Comparison with SOTA methods on ImageNet-LT and Places-LT." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 102, + 518, + 184 + ], + "lines": [ + { + "bbox": [ + 334, + 102, + 518, + 184 + ], + "spans": [ + { + "bbox": [ + 334, + 102, + 518, + 184 + ], + "type": "table", + "html": "
DatasetImageNet-LTPlaces-LT
FewAllFewAll
CLIP Zero [36]58.659.840.138.0
CLIP Finetune [36]34.560.522.739.7
VL-LTR [36]59.377.242.050.1
RAC [28]--41.847.2
LPT [13]--46.950.1
LTGC(Ours)70.580.652.154.1
", + "image_path": "9279d8df0cdc4f4bda329cfd62531fb7fba979f52c065db279a84268146282bf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 327, + 219, + 526, + 340 + ], + "blocks": [ + { + "bbox": [ + 312, + 199, + 539, + 210 + ], + "lines": [ + { + "bbox": [ + 312, + 199, + 539, + 210 + ], + "spans": [ + { + "bbox": [ + 312, + 199, + 539, + 210 + ], + "type": "text", + "content": "Table 2. Comparison with SOTA methods on iNaturalist 2018." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 219, + 526, + 340 + ], + "lines": [ + { + "bbox": [ + 327, + 219, + 526, + 340 + ], + "spans": [ + { + "bbox": [ + 327, + 219, + 526, + 340 + ], + "type": "table", + "html": "
MethodManyMediumFewAll
Softmax74.766.360.064.7
LADE [16]64.447.734.352.3
RIDE [42]71.570.071.671.8
PaCo [11]69.573.473.073.0
MDCS [56]76.575.575.275.6
CLIP Zero [36]6.13.32.93.4
CLIP Finetune [36]76.674.170.272.6
VL-LTR [36]---76.8
RAC [28]75.980.581.080.2
LPT [13]--79.376.1
LTGC(Ours)77.583.982.682.5
", + "image_path": "82b6585ea17f2b89f476fbc5d7941a33bdf9ece5df9c5a5c5e799e40ced73b2b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 331, + 385, + 521, + 447 + ], + "blocks": [ + { + "bbox": [ + 305, + 354, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 376 + ], + "type": "text", + "content": "Table 3. Comparison with different LMMs' methods on ImageNet-LT and iNaturalist2018." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 385, + 521, + 447 + ], + "lines": [ + { + "bbox": [ + 331, + 385, + 521, + 447 + ], + "spans": [ + { + "bbox": [ + 331, + 385, + 521, + 447 + ], + "type": "table", + "html": "
MethodImageNet-LTiNaturalist 2018
LENS [3]69.517.4
MiniGPT4 [58]60.420.9
MiniGPT4-v2 [7]68.527.1
GPT-472.164.3
Ours80.682.5
", + "image_path": "8ea806ac53459be876b54ed4ea047c73d5519769d4b212e5cce34d3356639770.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 626 + ], + "type": "text", + "content": "sampling. The results are shown in Tab. 4, demonstrating that our method significantly outperforms the other variants. Although our self-reflection and iterative evaluation modules already ensure the diversity and high quality of the generated images, there still exists a domain gap between the generated and original data [38]. This domain gap could exacerbate the long-tail effect on the test set (discussed in Appendix). A simple balanced sample [45] approach fails to solve the class imbalance caused by the mixing of original and generated data, while the Mixup [52] alone cannot address inter-class imbalances. To this end, our BalanceMix module combines the strengths of both methods, making the generated data well-suited for our framework." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 332, + 661, + 518, + 712 + ], + "blocks": [ + { + "bbox": [ + 317, + 641, + 533, + 651 + ], + "lines": [ + { + "bbox": [ + 317, + 641, + 533, + 651 + ], + "spans": [ + { + "bbox": [ + 317, + 641, + 533, + 651 + ], + "type": "text", + "content": "Table 4. Evaluation on the effectiveness of the BalanceMix." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 661, + 518, + 712 + ], + "lines": [ + { + "bbox": [ + 332, + 661, + 518, + 712 + ], + "spans": [ + { + "bbox": [ + 332, + 661, + 518, + 712 + ], + "type": "table", + "html": "
MethodImageNet-LTiNaturalist
w/o BalanceMix58.369.5
Balanced sample [45]63.973.8
Mixup [52]73.475.2
Ours80.682.5
", + "image_path": "d64b6641d6ff3b84e8f994feb887d0bf47939fbd636d2eba5cd866c5a57f553c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19516" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 520, + 242 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 520, + 242 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 520, + 242 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 520, + 242 + ], + "type": "image", + "image_path": "98e5a1c89b9bc5f0305aa58c42bcb82eeea9a12cfe7f73b130495d906cc0d726.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 547, + 288 + ], + "type": "text", + "content": "Figure 6. The visualization of generated images: The template \"A photo of the class [y]\" and LTGC. Each row represents a different class. The four images on the left are generated using the simple template \"A photo of the class [y],\" which results in images with uniform poses and plain backgrounds. The four images on the right are from the proposed LTGC and demonstrate the diversity of classes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 304, + 286, + 460 + ], + "blocks": [ + { + "bbox": [ + 50, + 304, + 286, + 460 + ], + "lines": [ + { + "bbox": [ + 50, + 304, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 50, + 304, + 286, + 460 + ], + "type": "image", + "image_path": "a4ca9b59496760d42a9cb6e31a6e56464d94df592570f291ef5e0c8367d983df.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 469, + 288, + 526 + ], + "lines": [ + { + "bbox": [ + 46, + 469, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 288, + 526 + ], + "type": "text", + "content": "Figure 7. The visualization of the images generated before and after passing the iterative evaluation module. The top row displays images that were filtered out, while the bottom row shows images regenerated by T2I after refining their corresponding descriptions. More visualizations are in the Appendix." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 549, + 132, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 132, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 132, + 561 + ], + "type": "text", + "content": "4.5. Visualization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Visualization of Generated Images: Template \"A photo of the class [y]\" vs our LTGC. Furthermore, we compare the images generated by LTGC with a simple prompt, \"A photo of the class [y]\" for the T2I model. As Fig. 6 illustrates, LTGC generates more accurate and diverse images compared to those generated by the simple prompt. For example, in the absence of control over the category description, the T2I model generates red features that do not belong to class Aquilegia Pubescens. In addition, with rich text for image generation, our approach also generates more diverse and accurate images compared to images generated by a simple template." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 307, + 545, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 415 + ], + "type": "text", + "content": "Visualization on iterative evaluation module. We compare images before and after refinement using the iterative evaluation module, and the visualization results are shown in Fig. 7. It shows that images before refinement often possess ambiguous semantic information, and the features of the corresponding classes are not distinct. However, after the refinement process, the quality of the images is substantially enhanced, and the distinctive features of the corresponding classes become more pronounced." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 424, + 378, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 424, + 378, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 378, + 436 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 545, + 598 + ], + "type": "text", + "content": "This paper introduces a novel generative and fine-tuning framework, named LTGC, to address the challenge of long-tail recognition. LTGC leverages the abundant potential knowledge embedded in large-scale models to generate diverse data for tail categories. The framework incorporates innovative designs to ensure the quality of the generated data and to fine-tune the model efficiently using both the generated and original data. The experimental results indicate that LTGC surpasses existing state-of-the-art methods on well-known long-tail benchmarks. In the future, we will explore the robustness of large-scale models for application in other areas, such as long-tail semantic segmentation and dissent detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 404, + 624 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.62271034. This project is also supported by the Ministry of Education, Singapore, under the SUTD Kickstarter Initiative Project (SKI 2021_02_06). Additionally, I would like to express my gratitude to Haoxuan Qu and He Huang for their valuable suggestions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19517" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Emanuel Sanchez Aimar, Arvi Jonnarth, Michael Felsberg, and Marco Kuhlmann. Balanced product of calibrated experts for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19967-19977, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 179 + ], + "type": "text", + "content": "[2] Martin Arjovsky, León Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 226 + ], + "type": "text", + "content": "[3] William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, and Amanpreet Singh. Towards language models that can see: Computer vision through the lens of natural language. arXiv preprint arXiv:2306.16410, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Michel Besserve, Arash Mehrjou, Rémy Sun, and Bernhard Schölkopf. Counterfactuals uncover the modular structure of deep generative models. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 316 + ], + "type": "text", + "content": "[5] Mateusz Buda, Atsuto Maki, and Maciej A Mazurowski. A systematic study of the class imbalance problem in convolutional neural networks. Neural Networks, 106:249-259, 2018. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[6] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. arXiv preprint arXiv:1906.07413, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 430 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu1 Xiaoqian Shen1 Xiang Li, Zechun Liu2 Pengchuan Zhang, Raghuraman Krishnamoorthi2 Vikas Chandra2 Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: Large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 495 + ], + "type": "text", + "content": "[8] Xiaohua Chen, Yucan Zhou, Dayan Wu, Wanqian Zhang, Yu Zhou, Bo Li, and Weiping Wang. Imagine by reasoning: A reasoning-based implicit semantic data augmentation for long-tailed classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 356-364, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 552 + ], + "type": "text", + "content": "[9] Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In Computer Vision-ECCV 2020 Workshops: Glasgow, UK, August 23–28, 2020, Proceedings, Part VI 16, pages 95–110. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 288, + 599 + ], + "type": "text", + "content": "[10] Chenhang Cui, Yiyang Zhou, Xinyu Yang, Shirley Wu, Linjun Zhang, James Zou, and Huaxiu Yao. Holistic analysis of hallucination in gpt-4v(ison): Bias and interference challenges, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "text", + "content": "[11] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715-724, 2021. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 700 + ], + "type": "text", + "content": "[12] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[13] Bowen Dong, Pan Zhou, Shuicheng Yan, and Wangmeng" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Zuo. Lpt: Long-tailed prompt tuning for image classification. arXiv preprint arXiv:2210.01033, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 160 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 194 + ], + "type": "text", + "content": "[15] Lin Geng Foo, Hossein Rahmani, and Jun Liu. Ai-generated content (aigc) for various data modalities: A survey, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 195, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 249 + ], + "type": "text", + "content": "[16] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6626-6636, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 250, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 547, + 293 + ], + "type": "text", + "content": "[17] Grant Van Horn, Oisin Mac Aodha, Yang Song, Alexander Shepard, Hartwig Adam, Pietro Perona, and Serge J. Belongie. The naturalist challenge 2017 dataset. CoRR, abs/1707.06642, 2017. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 547, + 338 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 338, + 547, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 370 + ], + "type": "text", + "content": "[19] Nathalie Japkowicz and Shaju Stephen. The class imbalance problem: A systematic study. Intelligent data analysis, 6(5): 429-449, 2002. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 547, + 415 + ], + "type": "text", + "content": "[20] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. arXiv preprint arXiv:1910.09217, 2019. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 459 + ], + "type": "text", + "content": "[21] Bingyi Kang, Yu Li, Sa Xie, Zehuan Yuan, and Jiashi Feng. Exploring balanced feature spaces for representation learning. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 547, + 514 + ], + "type": "text", + "content": "[22] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "text", + "content": "[23] Salman H Khan, Munawar Hayat, Mohammed Bennamoun, Ferdous A Sohel, and Roberto Togneri. Cost-sensitive learning of deep feature representations from imbalanced data. IEEE transactions on neural networks and learning systems, 29(8):3573-3587, 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 624 + ], + "type": "text", + "content": "[24] Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5212-5221, 2021. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 547, + 689 + ], + "type": "text", + "content": "[25] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Priyal Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. IEEE" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19518" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 286, + 95 + ], + "type": "text", + "content": "transactions on pattern analysis and machine intelligence, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[27] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 218 + ], + "type": "text", + "content": "[28] Alexander Long, Wei Yin, Thalaiyasingam Ajanthan, Vu Nguyen, Pulak Purkait, Ravi Garg, Alan Blair, Chunhua Shen, and Anton van den Hengel. Retrieval augmented classification for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6959-6969, 2022. 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[29] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 308 + ], + "type": "text", + "content": "[30] Haoxuan Qu, Xiaofei Hui, Yujun Cai, and Jun Liu. Lmc: Large model collaboration with cross-assessment for training-free open-set object recognition. arXiv preprint arXiv:2309.12780, 2023. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 342 + ], + "type": "text", + "content": "[31] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 409 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 464 + ], + "type": "text", + "content": "[33] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 533 + ], + "type": "text", + "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 578 + ], + "type": "text", + "content": "[35] Kaihua Tang, Mingyuan Tao, Jiaxin Qi, Zhenguang Liu, and Hanwang Zhang. Invariant feature learning for generalized long-tailed classification. In European Conference on Computer Vision, pages 709-726. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 633 + ], + "type": "text", + "content": "[36] Changyao Tian, Wenhai Wang, Xizhou Zhu, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. In European Conference on Computer Vision, pages 73–91. Springer, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[37] Lisa Torrey and Jude Shavlik. Transfer learning. In Handbook of research on machine learning applications and trends: algorithms, methods, and techniques, pages 242-264. IGI global, 2010. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[38] Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models. arXiv preprint arXiv:2302.07944, 2023. 2, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[39] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 943-952, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[40] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943-952, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 184, + 545, + 227 + ], + "type": "text", + "content": "[41] Tan Wang, Zhongqi Yue, Jianqiang Huang, Qianru Sun, and Hanwang Zhang. Self-supervised learning disentangled group representation as feature. Advances in Neural Information Processing Systems, 34:18225-18240, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[42] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long-tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 271, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 271, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 271, + 545, + 326 + ], + "type": "text", + "content": "[43] Yufei Wang, Zhou Xian, Feng Chen, Tsun-Hsuan Wang, Yian Wang, Katerina Fragkiadaki, Zackory Erickson, David Held, and Chuang Gan. Robogen: Towards unleashing infinite data for automated robot learning via generative simulation, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[44] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 7032-7042, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 403 + ], + "type": "text", + "content": "[45] Yu-Xiong Wang, Deva Ramanan, and Martial Hebert. Learning to model the tail. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "text", + "content": "[46] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824-24837, 2022. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 524 + ], + "type": "text", + "content": "[47] Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V 16, pages 247–263. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 557 + ], + "type": "text", + "content": "[48] Cihang Xie and Alan Yuille. Intriguing properties of adversarial training at scale. arXiv preprint arXiv:1906.03787, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[49] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. Advances in neural information processing systems, 33:19290-19301, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[50] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v (isdiction). arXiv preprint arXiv:2309.17421, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 679 + ], + "type": "text", + "content": "[51] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. arXiv preprint arXiv:2305.18279, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[52] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 2, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19519" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 452 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[53] Yifan Zhang, Bryan Hooi, Lanqing Hong, and Jiashi Feng. Test-agnostic long-tailed recognition by test-time aggregating diverse experts with self-supervision. arXiv preprint arXiv:2107.09249, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[54] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 217 + ], + "type": "text", + "content": "[55] Qihao Zhao, Yangyu Huang, Wei Hu, Fan Zhang, and Jun Liu. Mixpro: Data augmentation with maskmix and progressive attention labeling for vision transformer. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 272 + ], + "type": "text", + "content": "[56] Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, and Jun Liu. Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11597-11608, 2023. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 274, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 318 + ], + "type": "text", + "content": "[57] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16489-16498, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 319, + 287, + 363 + ], + "type": "text", + "content": "[58] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 287, + 419 + ], + "type": "text", + "content": "[59] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 420, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 420, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 49, + 420, + 287, + 452 + ], + "type": "text", + "content": "[60] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large-language models meet few-shot segmentation, 2023. 2, 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19520" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_content_list.json b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..980613e4071c39421689b4da43b4cb2c9d103621 --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_content_list.json @@ -0,0 +1,1835 @@ +[ + { + "type": "text", + "text": "LTM: Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering", + "text_level": 1, + "bbox": [ + 127, + 128, + 844, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaehoon Choi $^{1}$ Rajvi Shah $^{2}$ Qinbo Li $^{2}$ Yipeng Wang $^{2}$ Ayush Saraf $^{2}$ Changil Kim $^{2}$ Jia-Bin Huang $^{1,2}$ Dinesh Manocha $^{1}$ Suhib Alsisan $^{2}$ Johannes Kopf $^{2}$ $^{1}$ University of Maryland $^{2}$ Meta", + "bbox": [ + 116, + 200, + 854, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Advancements in neural signed distance fields (SDFs) have enabled modeling 3D surface geometry from a set of 2D images of real-world scenes. Baking neural SDFs can extract explicit mesh with appearance baked into texture maps as neural features. The baked meshes still have a large memory footprint and require a powerful GPU for real-time rendering. Neural optimization of such large meshes with differentiable rendering pose significant challenges. We propose a method to produce optimized meshes for large unbounded scenes with low triangle budget and high fidelity of geometry and appearance. We achieve this by combining advancements in baking neural SDFs with classical mesh simplification techniques and proposing a joint appearance-geometry refinement step. The visual quality is comparable to or better than state-of-the-art neural meshing and baking methods with high geometric accuracy despite significant reduction in triangle count, making the produced meshes efficient for storage, transmission, and rendering on mobile hardware. We validate the effectiveness of the proposed method on large unbounded scenes from mip-NeRF 360, Tanks & Temples, and Deep Blending datasets, achieving at-par rendering quality with $73 \\times$ reduced triangles and $11 \\times$ reduction in memory footprint.", + "bbox": [ + 75, + 325, + 473, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 722, + 209, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photorealistic reconstruction and rendering of real-world objects and scenes is a longstanding problem of importance with several applications in computer vision, robotics and AR/VR. In recent years, Neural Radiance Fields (NeRFs) [1, 2, 38, 39] have been quite successful at novel view synthesis. However, the scene representation is learned as an implicit volumetric representation that is expensive to evaluate. This is a crucial limitation for many applications, such as Augmented and Virtual Reality (AR/VR), often requiring real-time rendering with low memory and compute power.", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd98c5886290e24ae55c9adf33cbe461f3022bfd3af6e4b2ae7a5697cdb3f88f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 290, + 743, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d35952d2936b517156dbff20f54382b2c4a9178dc369af030dcbb225505e9257.jpg", + "image_caption": [ + "Figure 1. Example results using LTM showing mesh geometry (top-left) and novel-view synthesis (right). Bottom right plot compares performance of different methods on mip-NeRF360 outdoor dataset. Our method achieves comparable rendering quality to BakedSDF while reducing the mesh size considerably." + ], + "image_footnote": [], + "bbox": [ + 524, + 388, + 743, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f0cc1c40195cef09e1d21bbc39ceab91b3debf6fbd7501f48ff2357c658a3a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 290, + 861, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a4186137f44c69c544be406205e412211be0ff128b226383c4654ccb833d0fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 349, + 861, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2112e1b85a6718c98e4b5a798ff1bbf3e253d16c73e9b8879bb350b1d22fa518.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 409, + 861, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since modern graphics engines are highly optimized to rasterize triangle meshes, researchers have started exploring ways to extract explicit mesh representations from volumetric functions.", + "bbox": [ + 496, + 551, + 893, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MobileNeRF [5] uses a polygon mesh with texture maps storing feature vectors and opacity. Although their rendering quality is good, the mesh quality is far from an ideal surface. While NeRF2Mesh [50] and NeRFMeshing [44], based on volumetric density fields, excel at generating highly detailed geometry, they often result in noisy and bumpy surfaces and numerous floaters, especially in planar regions, which are common in large-scale scenes. In contrast, neural Signed Distance Function (SDF) methods [42, 52, 54] typically excel in reconstructing accurate surface geometry. However, they may not be able to capture small or detailed geometry, and their rendering quality is inferior to volumetric density-field methods. BakedSDF [55] overcomes these limitations by applying mip-NeRF 360 techniques to train SDF representation, though the resulting mesh representation has a high memory overhead.", + "bbox": [ + 496, + 612, + 895, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous neural SDF methods that rely on the Marching Cube algorithm [34] result in meshes with high storage overhead. This results in a significant challenge for practi", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5053", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal applications on mobile devices, where low disk storage and fast data transfer are essential. Instead of relying on the Marching Cubes, some methods [11,28,48,49] propose differentiable mesh reconstruction methods. These methods aim to extract an isosurface from an implicit function and can be seamlessly integrated with differentiable rendering approaches [16, 17]. However, most methods primarily concentrate on a single object with ideal implicit surface functions, unsuited for large unbounded real-world scenes [2, 55]. Reconstructing an ideal mesh for the unbounded large-scale scene is still an open problem, often resulting in non-manifold geometry or missing surfaces. Consequently, the renderings can result in visual artifacts due to the incompleteness of the mesh geometry.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main Results: In this paper, our goal is to achieve a realistic rendering with a compact mesh representation that optimally utilizes disk storage for unbounded large-scale scenes, all while preserving geometric quality. Our method initially trains neural fields for geometry and appearance separately. For geometry reconstruction, we adopt the BakedSDF approach [55] to train an SDF in a contracted space [2]. Regarding the appearance field, we utilize hash-based encoding [39] with two shallow MLPs to represent diffuse and specular colors. The SDF and appearance fields are compact representations for reconstructing a high-quality mesh and achieving realistic rendering.", + "bbox": [ + 75, + 305, + 472, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given the initial geometry representations, we perform mesh extraction and decimation. As unbounded large-scale scenes occupy a large amount of disk storage, our method focuses on generating a lightweight mesh without compromising geometric details. We adaptively extract the mesh for the central and background regions in the contracted space, respectively. Subsequently, we modify a mesh decimation [14] designed for large-scale scenes. Our method eliminates redundant triangles corresponding to simple geometric features, such as planar surfaces, while preserving complex geometric features.", + "bbox": [ + 75, + 489, + 472, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the appearance field is trained by volumetric rendering, the initial model is unsuitable for rasterization-based rendering. Moreover, the mesh from the previous step still fails to accurately represent areas with highly-detailed geometry (thin structures) or semi-transparent objects. These mesh defects hurt rendering quality. Thus, we use differentiable rendering [27] and utilize image-space supervision to refine the mesh surface by rectifying these mesh defects while preserving geometric accuracy. Due to the highly simplified mesh and various regularization techniques, our approach can deform the mesh topology without causing severe mesh distortion, a common issue in gradient-based mesh optimization [41, 43]. Compared to mesh-based neural rendering methods [5, 44, 50, 55], our approach results in lower storage overhead while providing comparable rendering quality. To sum up, our contributions include:", + "bbox": [ + 75, + 659, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We effectively utilize classical mesh decimation techniques to extract compact geometry from neural SDFs of large unbounded scenes.", + "- We introduce a new approach for joint optimization of appearance and vertex deformation that preserves geometric details and improves rendering quality.", + "- We achieve a trifecta of precise geometry, efficient storage, and rendering quality comparable to the state of the art on diverse captured environments such as mip-NeRF 360, Tanks & Temples, and the Deep Blending dataset." + ], + "bbox": [ + 500, + 90, + 893, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 271, + 640, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering Neural Fields achieve remarkable photo-realistic rendering quality. One of the most successful approaches is Neural Radiance Fields (NeRFs) [38], which adopts volumetric implicit representations and trains a radiance field through ray marching. Mip-NeRF [1, 2] presents casting a conical frustum into the scene space and extends this integrated positional encoding to unbounded scenes by using a novel contraction function. Instead of learning volumetric density, neural surface reconstruction methods [6, 42, 52, 54] propose learning a Signed Distance Function (SDF) for surface reconstruction via volumetric rendering. Furthermore, several methods have been proposed to accelerate NeRFs by utilizing efficient data structure [3, 10, 39], although achieving real-time rendering for high-resolutions remains a challenge. Alternatively, to enhance rendering speed, \"baking\" based methods [5, 19, 44, 45, 50, 55] typically precompute and store a trained NeRF into a more efficient representation. SNeRG [19] adopts a deferred shading model [8] and bakes trained NeRFs into sparse voxel grid representations with neural features. In particular, the mesh is widely used as geometry representation and bakes neural fields into its texture. MobileNeRF [5] builds on a mesh rasterization pipeline for efficient rendering on mobile architecture. NeRF2Mesh [50] introduces a two-step framework for NeRF training and textured mesh extraction. NeRFMeshing [44] proposes a novel method to produce a truncated SDF from NeRFs and extracts the accurate mesh. BakedSDF [55] enables us to reconstruct high-quality mesh and bake it into a small appearance model based on spherical Gaussians for fast rendering. DNMP [35] and NeuRaS [32] adopt mesh-based geometry and rasterization-based rendering for large driving scenes.", + "bbox": [ + 496, + 296, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mesh Processing Traditional graphics research has developed effective techniques for simplifying meshes, based on vertex decimation [46, 47], edge collapse [9, 12-14, 20] and appearance preservation [7]. Vertex decimation methods prioritize vertices based on shape heuristics and remove the least important ones. Edge collapse methods rank the edges based on the cost and subsequently contract pairs of vertices. QSlim [12, 13] introduces the concept of the", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5054", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quadric Error Metric (QEM), which is defined as the sum of quadrics representing the distance of a point from a set of planes. Following the edge collapse, they utilize this QEM to identify which edges should be removed while preserving geometric topology. Lindstrom et al. [29, 30] use vertex clustering [46] on a uniform grid and accumulate a quadric error matrix to the occupied grid cell. Wang et al. [51] employ RGB-D sequences to reconstruct a lightweight mesh.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differentiable Rendering and Mesh Optimization Recently, many differentiable rendering techniques [23] have emerged, aimed at facilitating end-to-end optimization of the rendering process by employing useful gradients. In particular, several works [4, 33] focus on approximating the rasterization stage of the rendering to compute meaningful gradients. Furthermore, modular differentiable renderers [21, 22, 27] exist, allowing users to customize and modify specific functions. Differentiable mesh reconstruction techniques [11, 28, 48] can build a mesh through gradient-based optimization. In particular, differentiable rendering [16, 17, 49] can offer 2D supervision to refine and enhance the geometric quality of the mesh. Some works [16, 17] adopt an inverse rendering approach, leveraging physics-based rendering to synthesize shapes. Nevertheless, gradient-based mesh optimization [41] is highly sensitive to initialization and prone to significant mesh distortion. Some methods [31, 41, 43] incorporate a novel ADAM optimizer and employ strong regularization techniques. However, this research is currently limited to generating single objects and is challenging to apply to large-scale scenes.", + "bbox": [ + 75, + 213, + 472, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Our Method", + "text_level": 1, + "bbox": [ + 76, + 547, + 207, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given multi-view images and corresponding camera poses, we aim to reconstruct the texture and geometry of unbounded large-scale scenes. Fig. 2 shows the three main stages of our pipeline: Initialization, Mesh Processing, and Optimization. In the first stage, we employ volumetric rendering to train neural implicit representation for geometry $f_{sdf}$ and appearance $f_{color}$ . Subsequently, we leverage the geometry representation $f_{sdf}$ to extract a high-resolution mesh via Marching Cubes [34] and reduce the number of triangles without causing geometry distortion. Next, we utilize differentiable rendering [27] to jointly optimize the explicit 3D mesh $M$ and implicit appearance field $f_{color}$ obtained from the first stage and refine the mesh surface.", + "bbox": [ + 75, + 575, + 468, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Initialization of Geometry and Appearance", + "text_level": 1, + "bbox": [ + 76, + 785, + 442, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this stage, we first learn neural implicit representations for both appearance and geometry. A neural radiance field [38] is an implicit representation of a 3D scene volume, a mapping from a 3D position $x$ and a ray viewing direction $d \\in \\mathbb{R}^3$ to an RGB color $c$ and volume density $\\rho$ . In practice, the networks modeled by MLPs are used to generate this", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0e13cda3153589f6e86fb7db0a0c2f728b647df793244f2e55086eadcaf7c75.jpg", + "image_caption": [ + "Figure 2. Overview of Our Pipeline. It consists of three steps: Initialization (Sec. 3.1), Mesh Processing (Sec. 3.2), and Optimization (Sec. 3.3). Our method focuses on baking the initial appearance and geometry representation, trained by volumetric rendering, into an explicit mesh and a memory-efficient texture conducive to real-time rendering. We employ classical mesh decimation to minimize storage requirements, introduce the optimization stage to facilitate rasterization-based rendering and refine the detailed mesh structure." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 877, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "volume density and view-dependent color:", + "bbox": [ + 500, + 301, + 782, + 316 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\rho = f _ {\\text {d e n s i t y}} (x), \\quad c = f _ {\\text {c o l o r}} (x, d). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 325, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To render the color of a pixel, we first compute the 3D points $t_i$ along the viewing ray associated with each pixel $p$ as $x(t) = o + td$ . Then, estimated volume densities $\\rho_i$ and colors $c_i$ are utilized to approximate a volume rendering equation [37] using numerical quadrature:", + "bbox": [ + 496, + 349, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j < i} \\rho_ {i} \\delta_ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 431, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T_{i}$ is the accumulated transmittance and $\\alpha_{i} = 1 - \\exp(-\\rho_{i}\\delta_{i})$ is the discrete opacity of the point. $\\delta_{i} = t_{i} - t_{i-1}$ is the distance between two neighboring points along the ray. To deal with large unbounded scenes, mip-NeRF 360 [2] introduces a proposal MLP for efficient sampling and space contraction to represent an unbounded scene. The scene contraction function $f$ (in Eq. 3) maps the unbounded points into bounded space:", + "bbox": [ + 496, + 472, + 890, + 593 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = \\left\\{ \\begin{array}{l l} x & | | x | | \\leq 1, \\\\ \\left(2 - \\frac {1}{| | x | |}\\right) \\frac {x}{| | x | |} & | | x | | \\geq 1. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 601, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appearance Similar to previous works [5, 19, 50], we decompose an appearance field $f_{color}$ into shallow MLPs to learn diffuse colors $c_d$ and specular colors $c_s$ separately. Particularly, the view-dependent effects are modeled by a small MLP that can be easily integrated into a fragment Shader. In order to ensure the MLPs are able to model an unbounded scene well, we adopt strategies similar to the ones introduced in mip-NeRF360 work, specifically, scene contraction, proposal MLP, and regularization for the background. We train both MLP networks $f_{color}$ and $f_{density}$ with a multi-resolution hash positional encoding [39] using the volume rendering objective in Eq. 2. The appearance field trained in this stage is further optimized to facilitate rasterization-based rendering, replacing the computationally expensive ray marching. This optimization is described later in Section 3.3.", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5055", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a94c3532d95a8f705e7b0348191d136b71eb0df50b5b675286e07fa10dcb62a7.jpg", + "image_caption": [ + "Figure 3. Our Approach: In Section 3.1, we start training both the appearance $f_{color}$ and geometry $f_{sdf}$ representations for a large-scale scene using volumetric rendering (shown in orange). Then, in Section 3.2, our method extracts the mesh from $f_{sdf}$ and significantly simplifies its structure, resulting in the lightweight mesh $M$ (shown in blue). In Section 3.3, we jointly train vertex deformation $\\Delta x$ and $f_{color}$ computed from appearance modeling." + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 836, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Geometry Initial geometry modeling in our method is similar to BakedSDF [55]. The core mechanism for neural surface reconstruction was introduced in VolSDF [54] for small objects. BakedSDF shows that by leveraging the efficient sampling and contraction ideas introduced in mipNeRF360 [2], VolSDF [54] can be extended to large unbounded scenes. We follow this and jointly train two MLPs using the volumetric rendering objective, (i) $f_{sdf}$ - representing geometry of a scene as a signed distance function, and (ii) $g_{color}$ - a color prediction network, as shown in the Geometry Reconstruction block in Fig.3.", + "bbox": [ + 75, + 311, + 472, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following VolSDF, we define the volume density $\\sigma(x)$ as a function of surface geometry $f_{sdf}$ , namely, $\\sigma(x) = \\alpha \\Psi_{\\beta}(f_{sdf}(x))$ , where $\\Psi_{\\beta}$ is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and $\\beta$ is the learnable scaling parameter. The Eikonal loss [15] to regularize the $f_{sdf}$ is defined as,", + "bbox": [ + 76, + 477, + 470, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e i k} = \\left(\\| \\nabla f _ {s d f} (x) \\| _ {2} - 1\\right) ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 573, + 468, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The color MLP $g_{color}$ , in addition to the position and viewing direction, also takes the predicted surface normal and a 256-dimensional geometric feature as inputs. It is important to note that the sole purpose of color MLP $g_{color}$ is geometry initialization and we discard it after this step. This is not to be confused with the shallow decomposed color MLP $f_{color}$ trained specifically to model appearance that is also jointly refined with mesh vertices in the later stage.", + "bbox": [ + 75, + 598, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mesh Extraction and Decimation", + "text_level": 1, + "bbox": [ + 76, + 726, + 367, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this stage, we first extract a mesh by using Marching Cubes (MC) [34] algorithm on learned SDF ( $f_{sdf}$ evaluated over a uniform grid) and simplify the extracted mesh using a geometry-aware quadric error based mesh decimation. While this approach works pretty well for extracting meshes of well-captured small-scale objects and scenes, it presents unique challenges for large-scale scenes. Meshing neural SDFs of large-scale scenes with marching cubes can often result in spurious surfaces arising from (i) insufficient grid resolution to model complex geometry of nearby", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "objects, and (ii) poorly learned geometry of distant objects due to insufficient view-points observing such distant regions. Furthermore, since our SDF $f_{sdf}$ is trained in the contracted space, the resulting mesh is also defined in the contracted space. When we uncontract, i.e. warp the mesh vertices back into the original Euclidean space, it further exacerbates the geometry distortion in the background region.", + "bbox": [ + 496, + 310, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to preserve the detailed geometry in central (foreground) regions while achieving a coarser but less noisy reconstruction of distant (background) regions, within a small triangle budget. We achieve this by treating foreground and background regions distinctly during both extraction and simplification steps. During MC extraction, we set a finer grid resolution of 2048 and a spatial range of $[-1,1]^3$ for the central region, and a coarser resolution of 1024 and a spatial range of $[-2,2]^3$ for the background region. For decimation, we first partition the central and background regions and employ an aggressive mesh decimation in background regions. This helps us transform bumpy and noisy background areas into smooth surfaces while also reducing redundant triangles.", + "bbox": [ + 496, + 417, + 892, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first apply vertex clustering and then run quadric error based simplification algorithm - QSlim [12, 13] for all clusters. Then, we run edge contraction for each cluster to reach a threshold and repeat this process until convergence. The QSlim algorithm calculates a quadric metric for each vertex to decide whether the two vertices should be merged and aggregates these quadrics during each edge collapse operation. The basic quadric error in QSlim [13] is defined as the point-to-plane distance from vertex $v \\in \\mathbb{R}^3$ to all planes $p \\in \\mathbb{R}^4$ of its adjacent faces $N(v)$ :", + "bbox": [ + 496, + 628, + 893, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE (v) = \\sum_ {p \\in N (v)} \\left(p ^ {T} v\\right) ^ {2} = v ^ {T} \\left(\\sum_ {p \\in N (v)} p p ^ {T}\\right) v = v ^ {T} Q _ {v} v. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 792, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The quadric $Q_{v}$ is the symmetric 4x4 matrix, and each edge computes this quadric by adding the quadrics of the corresponding two vertices. The most commonly used quadric error is defined solely based on 3D vertex positions $v$ .", + "bbox": [ + 496, + 840, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5056", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, the most effective strategy is to represent planar surfaces with a smaller number of large triangles while keeping a consistent number of triangles for complex shapes. Variants that incorporate additional surface attributes, such as normals or textures in calculating the quadric error are proposed in [12]. We experimentally observed that compared to the basic QSlim algorithm, using both the vertex positions $v \\in \\mathbb{R}^3$ and the corresponding vertex normal $n \\in \\mathbb{R}^3$ for quadric error is very effective at prioritizing triangle decimation on flat surfaces while retaining them on complex surfaces. The combination of vertex clustering and geometry-aware QSlim helps us in effectively reducing the number of triangles while preserving the original geometry.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Appearance and Geometry Optimization", + "text_level": 1, + "bbox": [ + 76, + 315, + 426, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this stage, we optimize the appearance model jointly with mesh vertex deformation and refine the mesh topology using a geometric complexity based criterion.", + "bbox": [ + 75, + 339, + 468, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint optimization of appearance and vertex deformation Given the simplified mesh $M$ and the initial appearance field $f_{color}$ , we employ the fast differentiable rasterizer introduced in Nvdiffrast [27] to jointly optimize the mesh geometry and the appearance field. The initial appearance field $f_{color}$ trained using the volume rendering objective (Eq. 1) is agnostic of the surface geometry of the extracted and simplified mesh. To finetune the model to fit the mesh geometry, we rasterize the mesh into each of the training camera poses and obtain surface points $s$ and face normals $N$ interpolated per pixel. Then we apply the contraction function in Eq. 3 to warp the surface points. This is because the appearance field defined in contracted space proves to be more effective in enhancing the rendering quality [2]. The appearance field $f_{color}$ takes these surface points $s$ and viewing direction $d$ as input, rendering an image $\\hat{I}_i$ . Initially, rendering performance is degraded because the appearance field only utilizes the intersection point between the ray and the surface rather than accumulating all sampled points along the ray. We finetune this appearance field to be compatible with mesh rendering, akin to the functionality of a fragment Shader.", + "bbox": [ + 75, + 386, + 472, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the positions $x$ of the mesh vertices, we train per-vertex deformation $\\Delta x$ using differentiable rendering [27]. We explore two approaches to optimize vertex deformation: a standard ADAM optimizer [24] and a rotation-invariant ADAM optimizer [43]. The ADAM optimizer is widely employed [16, 50] for deforming mesh topology to enhance rendering performance. While the ADAM optimizer often results in mesh distortions such as flipped triangles and self-intersections, our method mitigates mesh distortion, especially on smooth surfaces, due to the rigorous mesh decimation performed in Section 3.2.", + "bbox": [ + 75, + 719, + 472, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objective Function To jointly train appearance field and vertex deformation, we minimize the image-space L1 loss $L_{\\text{appearance}}$ between the rendered image $\\hat{I}_i$ of the deformed surface and the groundtruth image $I_i$ for a camera view $P_i$ :", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {a p p e a r a n c e}} = \\sum_ {i} \\left| \\hat {I} _ {i} - I _ {i} \\right|. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 157, + 890, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Naively training vertex deformation is highly likely to result in undesirable mesh topology characterized by a disorganized arrangement of triangles, often referred to as a \"polygon soup.\" To prevent this mesh distortion, we employ various regularization methods. We use the Laplacian mesh smoothness loss and the normal consistency loss inspired by [36].", + "bbox": [ + 496, + 196, + 890, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l a p} = \\sum \\| L V \\| _ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 313, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $V \\in \\mathbb{R}^{N \\times 3}$ ( $N$ is the number of vertices) is the matrix of deformed vertex coordinates, indicating $x + \\Delta x$ , at each training step, and $L \\in \\mathbb{R}^{N \\times N}$ is the graph Laplacian of the mesh [40]. Minimizing the magnitude of a vertex's differential coordinates ( $LV$ ) reduces its distance from its neighbors' average position. The normal consistency loss is", + "bbox": [ + 496, + 339, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {n c} = \\sum_ {i, j \\in F} \\left(1 - n _ {i} \\cdot n _ {j}\\right) ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 438, + 890, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_i$ and $n_j$ denote the surface normal of two adjacent faces and $F$ is the set of faces sharing a common edge. Additionally, in Section 3.2, a high-resolution mesh is rasterized in the first pass before decimation. The output of this rasterization is used to interpolate face normals per pixel, resulting in a normal map $\\hat{N}_i$ for every viewpoint. We calculate the L1 loss between a rendered normal $N_i$ from current mesh $M$ and the $\\hat{N}_i$ :", + "bbox": [ + 496, + 477, + 890, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o r m}} = \\sum_ {i} | \\hat {N} _ {i} - N _ {i} |. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 604, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$L_{norm}$ provides geometry supervision, particularly in planar regions, helping prevent severe mesh distortion when moving vertices. The total loss function $L$ is defined as:", + "bbox": [ + 498, + 642, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {c o l o r}} + \\lambda_ {\\text {l a p}} L _ {\\text {l a p}} + \\lambda_ {\\text {n c}} L _ {\\text {n c}} + \\lambda_ {\\text {n o r m}} L _ {\\text {n o r m}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 695, + 890, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{*}$ is the weight to balance between the loss terms.", + "bbox": [ + 500, + 719, + 879, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh Refinement Inspired by NeRF2Mesh [50], we adopt a classical mesh refinement approach based on mesh subdivision and decimation during training. The mesh refinement strategy is applied differently to the central and background regions. Instead of rendering error proposed by [50], which can arise from many factors including specularity, we employ a criterion: the vertex deformation values for subdividing triangle faces. The large vertex deformation implies a more substantial triangle budget to accurately capture the intricacies of highly detailed geometry associated", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5057", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/14527b1c3e0e24795071a1d22597decaa055bf613881ba4a5a53b2c1e2b43560.jpg", + "image_caption": [ + "Figure 4. Visual comparison of Our Method, NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. Our method and NeRF2Mesh show shading mesh and wireframe mesh extracted without texture, respectively." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 356, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a1e8a9d124af83baaa9a7f230bbfd54f09a1dc282b3ba78fdbd72f736b07eff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 90, + 619, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e0d789746e2598603ed6a323aff4a2c18b8a135ce7eb345d64e6f2ae7605ba9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 90, + 754, + 344 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c06279adae33f8bd1274b8a3ed84261d91db0b072cd0ba778d39907affce1112.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 90, + 890, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with those vertices. We sort the absolute values of vertex deformation and determine a $10\\%$ of faces to increase triangles. Then, we apply $\\sqrt{3}$ subdivision [26], which adds a vertex at the barycenter of a triangle and connects it to all vertices of this triangle. This scheme enables us to augment the triangle budget for complex geometry and mitigate self-intersection during vertex deformation. Furthermore, the mesh decimation presented in Section 3.2 is applied after every mesh subdivision to remove redundant triangles on simple and flat geometry. For background regions, we examine all training views and check faces intersecting with ray casting. We utilize dilation for all invisible masks and employ mesh decimation to remove redundant faces.", + "bbox": [ + 75, + 385, + 472, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rotation-Invariant ADAM The rotation-invariant ADAM formulation [43] modifies the update rule in ADAM, resulting in the update step: $\\Delta x = \\zeta \\cdot v \\cdot l_{ref}$ . Here, $\\zeta$ is the learning rate, and $v$ is the standard ADAM update step, which is the loss gradient with respect to $x$ . $l_{ref}$ is the coefficient initially set to the edge length incident to $x$ , and it gets updated based on the average $v$ .", + "bbox": [ + 75, + 582, + 472, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 76, + 726, + 269, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: To validate the effectiveness of our method, we mainly conduct experiments on seven unbounded, large-scale $360^{\\circ}$ indoor and outdoor scenes from the mip-NeRF 360 [2] dataset. Also, we validate the generalization ability of our approach on two different datasets, including large-scale scenes: two scenes from the Tanks & Temples [25] dataset and two scenes from the Deep Blending [18] dataset. Implementation Details: For all datasets, we subsampled every eight frames for test sets following the mip-NeRF 360 paper [2]. For geometry reconstruction in Section 3.1, we", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "use SDFstudio [56] to implement BakedSDF. In Section 3.2, we reduce $95\\%$ of the triangles in the central region and $99\\%$ of the triangles in the background regions through decimation. In Section 3.3, we run all experiments for 40k iterations, and a learning rate is annealed from 1e-2 to 1e-3. We set the loss weights to $\\lambda_{lap} = 1\\mathrm{e - }3$ , $\\lambda_{nc} = 1\\mathrm{e - }3$ , and $\\lambda_{norm} = 1\\mathrm{e - }4$ . We apply mesh refinement at 8k, 16k, and 24k training steps. Subsequently, we compute diffuse color and specular features using $f_{color}$ to export textures. All these color maps and features are 3-channel .png files. Additional details are provided in the supplementary materials.", + "bbox": [ + 496, + 385, + 893, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 560, + 702, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our proposed method with three different methods: 1) MobileNeRF [5] is the state-of-the-art method, enabling NeRF to achieve real-time rendering of explicit 3D mesh and neural features. 2) NeRF2Mesh [50] employs a volumetric rendering to reconstruct the coarse mesh and refine the mesh by jointly optimizing appearance and geometry. 3) BakedSDF [55] is a recent method for optimizing a neural surface-volume representation for real-time rendering of large unbounded scenes. We also add NeRFMeshing [44], which distills the rendering NeRF into a neural field that represents TSDF (Truncated SDF) and reconstructs the 3D mesh. As BakedSDF and NeRFMeshing have not released public code, we adopt the metrics provided in their original papers. BakedSDF has made its GLB file available. We utilize this file for visualization.", + "bbox": [ + 496, + 583, + 892, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 qualitatively compares our method with NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. We visualize the shading meshes extracted without texture, which is the most effective way to visualize geometric differences. Furthermore, we include a wireframe mesh, distinguishing between our method and NeRF2Mesh. Visu", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5058", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fcbe40fc2a1c6e47281651ebb80208178fb08f7882aae6d1bafa2bf83c1b4652.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMeshOutdoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)Indoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)
BicycleGardenStumpMeanRoomCounterKitchenBonsaiMean
Instant-NGP [39]X22.1 / 0.49 / 0.4924.5 / 0.65 / 0.3123.6 / 0.57 / 0.4523.4 / 0.57 / 0.4229.2 / 0.85 / 0.3026.43 / 0.80 / 0.3428.5 / 0.82 / 0.2530.3 / 0.89 / 0.2328.6 / 0.84 / 0.28
mip-NeRF 360 [2]X24.4 / 0.68 / 0.3026.9 / 0.81 / 0.1726.4 / 0.74 / 0.2625.9 / 0.75 / 0.2431.6 / 0.91 / 0.2129.5 / 0.89 / 0.2032.3 / 0.92 / 0.1333.5 / 0.94 / 0.1831.7 / 0.92 / 0.18
MobileNeRF [5]O21.7 / 0.43 / 0.5118.8 / 0.59 / 0.3623.9 / 0.56 / 0.4321.5 / 0.53 / 0.4328.9 / 0.85 / 0.2825.1 / 0.72 / 0.2926.8 / 0.79 / 0.7923.8 / 0.71 / 0.7226.2 / 0.77 / 0.52
NeRF2Mesh [50]O22.1 / 0.48 / 0.5123.4 / 0.55 / 0.4022.5 / 0.54 / 0.4622.7 / 0.52 / 0.4625.7 / 0.79 / 0.3523.9 / 0.71 / 0.3524.0 / 0.61 / 0.3625.0 / 0.77 / 0.2924.7 / 0.72 / 0.34
NeRFMeshing [44]O21.1 / - / -22.9 / - / -22.6 / - / -22.2 / - / -26.1 / - / -20.0 / - / -23.6 / - / -25.6 / - / -23.8 / - / -
BakedSDF [55]O22.0 / 0.57 / 0.3724.9 / 0.75 / 0.2123.6 / 0.59 / 0.3723.5 / 0.64 / 0.3228.7 / 0.87 / 0.2525.7 / 0.81 / 0.2826.7 / 0.82 / 0.2427.2 / 0.85 / 0.2627.0 / 0.84 / 0.26
OursO22.4 / 0.52 / 0.4423.5 / 0.64 / 0.3023.7 / 0.57 / 0.4223.2 / 0.58 / 0.3929.3 / 0.88 / 0.2325.1 / 0.77 / 0.2726.5 / 0.80 / 0.2127.3 / 0.84 / 0.2227.1 / 0.82 / 0.23
Ours w/o OptO16.1 / 0.30 / 0.5318.8 / 0.45 / 0.3919.4 / 0.39 / 0.5018.1 / 0.38 / 0.4817.2 / 0.54 / 0.4616.2 / 0.47 / 0.4718.1 / 0.50 / 0.3616.4 / 0.54 / 0.4317.0 / 0.51 / 0.43
Ours w/ GOO21.1 / 0.48 / 0.4723.2 / 0.65 / 0.3023.0 / 0.53 / 0.4622.5 / 0.55 / 0.4127.5 / 0.85 / 0.2623.6 / 0.72 / 0.3325.5 / 0.78 / 0.2325.9 / 0.76 / 0.325.4 / 0.78 / 0.28
Ours w. RADAMO21.6 / 0.48 / 0.4623.4 / 0.66 / 0.3023.5 / 0.56 / 0.4522.8 / 0.57 / 0.4027.5 / 0.85 / 0.2624.1 / 0.74 / 0.3125.8 / 0.80 / 0.2226.8 / 0.80 / 0.2626.0 / 0.80 / 0.26
", + "bbox": [ + 99, + 88, + 870, + 204 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/db2286bca39d67e0509af4832514d0c3cf00d9ac33817069a80bafd26bfd949e.jpg", + "table_caption": [ + "Table 1. Quantitative Comparison on mip-NeRF 360 Dataset. “-” denotes a missing implementation. The best and second best-performing algorithms for each metric are bolded and underlined." + ], + "table_footnote": [], + "table_body": "
MethodBicycleGardenStumpRoomCounterKitchenBonsaiMean
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]2011435 / 6638135723 / 3401641069 / 497220903 / 421170864 / 39995552 / 25684606 / 284153879 / 1262
NeRF2Mesh [50]129815 / 858117650 / 779123415 / 82069328 / 46279517 / 38489601 / 43470319 / 63697387 / 667
BakedSDF [55]1087111248 / 4438451310773M / 2107294319667 / 384673857227 / 1411058512212 / 2389265913832 / 2705664113629 / 266796871564M / 27951
Ours83270 / 52950113 / 218120425 / 84745124 / 23858138 / 2684070 / 13642233 / 45962196 / 385
", + "bbox": [ + 81, + 244, + 887, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "alizing the wireframe mesh from BakedSDF lacks meaningful interpretation due to its extremely high face density. MobileNeRF generates a \"triangle soup\" that inaccurately represents the scene's geometry. Also, compared to the mesh reconstructed by NeRF2Mesh, our method shows sharp details of complex geometry and smooth surfaces on the ground or wall. Moreover, in the wireframe mesh, our method employs more triangles in intricate shapes like the bicycle or the garden vase while allocating smaller triangles to planar regions or simpler structures. While the geometry of BakedSDF exhibits superior quality compared to our mesh reconstruction, achieving this high-quality mesh necessitates approximately $73 \\times$ more triangles in Table 2.", + "bbox": [ + 75, + 342, + 468, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method effectively achieves the optimal tradeoff between rendering quality and disk storage cost. For comparison, we report two tables for rendering quality and disk storage. Table 1 reports the rendering quality using the standard PSNR, SSIM [53], and LPIPS [57] on mip-NeRF 360 dataset. Instant-NGP [39], and mip-NeRF 360 [2] are volumetric rendering approaches that do not reconstruct explicit geometry, a topic that lies outside the primary focus of our study. In Table 2, we report the disk storage in megabytes and the number of vertices and faces. For a fair comparison, all the mesh files are saved in .obj and .mtl formats. Texture images are stored as .png files. Other assets, such as view-dependency MLP, are saved as .json files, except BakedSDF, which only provides gLTF format files. Our method demonstrates comparable rendering quality in outdoor scenes (with a 0.3 PSNR drop) and indoor scenes (with a 0.1 PSNR increase) while saving $73 \\times$ triangles and $11 \\times$ storage compared to the state-of-the-art BakedSDF. In comparison to NeRF2Mesh and MobileNeRF, our method shows better-rendering quality (Table 1), demands less disk storage, and fewer triangles (Table 2), and captures superior mesh quality (Fig. 4). In Table 5, we measure the rendering speed and GPU memory by running in-browser on a Mac", + "bbox": [ + 75, + 553, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6fe53221e492fba1e06bfbabe8fe5a87c697f4792b9c16690227c3a35d5269c1.jpg", + "table_caption": [ + "Table 2. Disk Storage (MB), Number of Vertices (V) and Faces (F). The unit for the number of vertices and faces is ${10}^{3}$ . \"M\" denotes the ${10}^{6}$ unit. We save ${73} \\times$ the number of triangles and reduce ${11} \\times$ disk storage." + ], + "table_footnote": [], + "table_body": "
MethodMeshDeep Blending [18]Tanks & Temples [25]
DrJohnsonPlayroomBarnCourthouse
Instant-NGP [39]X27.7 / 0.84 / 0.3819.5 / 0.78 / 0.4623.4 / 0.72 / 0.4320.0 / 0.68 / 0.53
mip-NeRF360 [2]X29.1 / 0.90 / 0.2429.6 / 0.90 / 0.2527.2 / 0.81 / 0.3221.3 / 0.72 / 0.47
MobileNeRF [5]O25.9 / 0.78 / 0.3128.7 / 0.85 / 0.2824.4 / 0.77 / 0.2417.7 / 0.54 / 0.36
NeRF2Mesh [50]O24.0 / 0.75 / 0.5126.6 / 0.83 / 0.4123.4 / 0.72 / 0.4318.2 / 0.63 / 0.56
OursO26.5 / 0.82 / 0.4029.1 / 0.87 / 0.3226.5 / 0.83 / 0.2420.2 / 0.73 / 0.34
", + "bbox": [ + 501, + 340, + 888, + 409 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed85c2771f06179a32e63d57bd468d3f311bd1c6d86053b9edb4c6d4aeb3cc3b.jpg", + "table_caption": [ + "Table 3. Quantitative Comparison. We report PSNR $\\uparrow$ , SSIM $\\uparrow$ , and LPIPS $\\downarrow$ on Deep Blending and Tanks & Temples datasets." + ], + "table_footnote": [], + "table_body": "
MethodDrJohnsonPlayroomBarnCourthouse
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]174.9602 / 285281.81097 / 521289.21080 / 502263.6934 / 417
NeRF2Mesh [50]108.6414 / 815118.7401 / 793131.9397 / 793145.7451 / 843
Ours26.746 / 9260.9203 / 4014086 / 17147.497 / 192
", + "bbox": [ + 503, + 446, + 888, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4ba1fbce16581ada433b0a776d6721c8a27c129f08e084f09c5d9750776a24e.jpg", + "table_caption": [ + "Table 4. Disk Storage (MB), Vertices (V) and Faces (F)." + ], + "table_footnote": [], + "table_body": "
FPS / GPU (MB)MobileNeRFNeRF2MeshOurs
MacBookPro (1920x1080 Resolution)59 / 324205 / 130212 / 115
", + "bbox": [ + 519, + 513, + 870, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. FPS and GPU memory on mip-NeRF 360 dataset", + "bbox": [ + 517, + 547, + 870, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BookPro (2020, M1).", + "bbox": [ + 500, + 568, + 643, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 compares our method to the state-of-the-art baselines that focus on rendering quality: NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. In their first stage, BakedSDF* represents the volumetric rendering results obtained from a surface-based volumetric representation. The visual results of our approach exhibit sharper details than other methods. In the Bicycle and Room scenes, the mesh surface of the armrest for a bench and tree leaves exhibits noticeable jiggling, resulting in distorted images. Unlike BakedSDF*, our method effectively manages the thin structure of the wheel.", + "bbox": [ + 496, + 584, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Results on Deep Blending and Tank & Temples Datasets: We train and evaluate our method on four scenes: Playroom and DrJohnson from the Deep Blending dataset, and Barn and Courthouse from the Tanks & Temples dataset. The former includes bounded indoor scenes, while the latter comprises large unbounded outdoor scenes. Their capture methodology differs from that of the mip-NeRF 360 dataset, and they do not utilize 360-degree scenes. In Tanks & Temples, we exclude sky regions during training and evaluation, as mesh-based rendering tends to", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5059", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68e53eb087640f075447d91d36fa3ffe44c4f15000be93b56aa2d8c9a6702af7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 261, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c37e774ade1e6567bc4c09d3d4ba07c869e3ece3fbde836d99caf0eb6cff20d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 169, + 259, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d4394e26a7d4db72ef6404a38ce3006407e2bd58f69f92fed061946b0922e8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 247, + 259, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ec0cbd3339e840ef314752c05fd88c7e50637f40fa36f7f8e8347fa1dd6d9bf.jpg", + "image_caption": [ + "Scene" + ], + "image_footnote": [], + "bbox": [ + 96, + 327, + 259, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c57afcff8883b762d5559ca7da593a550813c2bbf9aec92c4b9f02afba764775.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 90, + 411, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2220ff23b0d996e39d3cd1c7aa77283b0724e2faa2b51024de198f972a6d7ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 169, + 411, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a85d1b7f3df4a4a87659e07c5b000bb4de585d65159b3f53ff8214d2d6bafaa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 247, + 411, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45e7086bb9bb53aa17e139cdf10ae2c7ea4c4f5e85184c1c25544b1c96b834c6.jpg", + "image_caption": [ + "(a) Ours" + ], + "image_footnote": [], + "bbox": [ + 261, + 327, + 411, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9a23a8610c7b0be0ae0756d74821d9204bdd461bc754843a2ebf8da15dccfd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 90, + 565, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/348dd17a52fee2f3d2d563e6f5fab39e675bd93efe0139e7a2aba90d1357aaf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 169, + 565, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4196afc8ff9354ad2d5d27b67851277f28f7446b08aa7e2b171620aa8ee74e27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 247, + 565, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d99d8a57c37905c7efe16961007c2322468200aa98a8834c931362db66460316.jpg", + "image_caption": [ + "(b) NeRF2Mesh" + ], + "image_footnote": [], + "bbox": [ + 413, + 327, + 565, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/00673299fbc26f822afc3ae151745c998281b7f9b02d583d82238d9bffddcbeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 90, + 718, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a866ad403d194ae5594412e83921d3960e82268ccd6c5c37bc5c7fa52cd6b186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 169, + 717, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f698c865df6068129f7b1696be7df410383f488c0e220424b6691895edcb6f0.jpg", + "image_caption": [ + "(c)BakedSDF\\*" + ], + "image_footnote": [], + "bbox": [ + 566, + 247, + 717, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b38533c70d678c3ec03b6be9c1c29a81861d429f4d2376d33256449331c4ddc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 90, + 870, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ab2eb16d18c82c0847b7795b7d411707f8917be93c33303c06933fd50d3e297.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 169, + 870, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab99f518d41a6e923006684fd031a7e7c87e3c964524db63795418d03e0652b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 247, + 870, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86534fee9dbce4b104e47c63b6b3bd7de8024f7aa5966bd39095ba914e4f739a.jpg", + "image_caption": [ + "(d) MobileNeRF" + ], + "image_footnote": [], + "bbox": [ + 720, + 327, + 872, + 405 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/107430fdf18efb43a11e33f7955d6f1d6e636308269803ff535001c3b91204ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 470, + 174, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6cc9d40c8ee9823b8491755349163a077e2f7a87b31894fe927abd6167af91b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 517, + 174, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c67fafc6044e4b38434f0d460526a49fa29f4be110fe71f22e5544a63ad36cd.jpg", + "image_caption": [ + "(a) Ours w/o Opt.", + "Figure 6. Zoom-in Shading Mesh to see more geometric details. encounter challenges with far-away sky regions." + ], + "image_footnote": [], + "bbox": [ + 81, + 563, + 174, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c292a89b0497768ccea656f80a743fe2d11a23280f63b2064d1b2eb7c6c75af4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 470, + 269, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7d8dd9d1c052a2468a72c587665454adacf0f6db51b06f7ac74544f0bc77e3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 517, + 271, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3b2051cacf858e78dae5891bfd83d9a6cb7eb77b772d8b9dcc3d392596225cd.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 176, + 563, + 271, + 630 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6cb5bcb7d3fc2c017ec92a8c7f8e833a27147eb8e945bb6c47e6424b636ff28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 470, + 367, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4553bd33dc7fb5004c38a0c012cecb47d36eb90ffcb1e90b57f26639e9ce0eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 517, + 367, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/893f2fddaaca3a5105cc1ff06217f8d8fbfd1c53a8d9389b44fef24b4ddbc831.jpg", + "image_caption": [ + "(c) NeRF2Mesh" + ], + "image_footnote": [], + "bbox": [ + 274, + 563, + 367, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c871f7b71b6899b394c1467c9103941aba05c65f4d46b604f240b35c33cf8595.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 470, + 465, + 515 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f465ad3f474a6dbaa440e66b4f8bd8fda04f9d233c3bb32a124bea6509e6215f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 517, + 465, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84f2f4a507bcf6e226531e71edf87725319108c1d885124f6636fa7087602421.jpg", + "image_caption": [ + "(d) BakedSDF" + ], + "image_footnote": [], + "bbox": [ + 370, + 563, + 465, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation: In Fig. 6, we visualize the zoom-in shading mesh extracted without texture. \"Ours w/o Opt\" denotes we did not execute our optimization stage in Section 3.3 after mesh processing. The optimization stage enables us to represent the geometric details of garden vases and bicycle wheels. Our method can represent more detailed structures than BakedSDF and a smoother surface than NeRF2Mesh. Moreover, it is noteworthy that when employing rasterization-based rendering with the initial color model without the optimization stage, the rendering quality is poor, as shown in Table 1. In this table, \"Ours w/o Geo\" only trains appearance model $f_{color}$ without training vertex deformation. Training vertex deformation yields a 1.7 gain in indoor scenes by refining the mesh surface. \"Ours w. RADAM\" denotes that we use the rotation-invariant", + "bbox": [ + 75, + 674, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4d21dcc549a2abcbefd5f0aff8296d2d7125c2fc3a01e1d411bf9d265a33197.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparisons with Existing Methods. We visually compare our proposed technique with NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. The visual results of BakedSDF* generated in its first stage through volumetric rendering. Red and blue dotted lines emphasize subtle differences in rendering quality.", + "Figure 7. Naive QSlim Results" + ], + "image_footnote": [], + "bbox": [ + 500, + 469, + 696, + 512 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fbd3150b7787c1382e9040a09771cdfb6e24a77d07f96c5d90075822678598db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 469, + 890, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ADAM (RADAM) for updating mesh vertices instead of a naive ADAM. However, it provides benefits for preventing mesh distortion. In Fig. 7, we show the visualization results of naive QSlim [12]. The naive QSlim leads to shrinkage and degeneration during the edge collapse process, and the boundary is over-smoothed compared to our method.", + "bbox": [ + 500, + 530, + 890, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion, Limitations, and Future Work", + "text_level": 1, + "bbox": [ + 500, + 633, + 879, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a practical approach for reconstructing geometrically accurate meshes with photorealistic rendering quality. Our method achieves a compact mesh representation in unbounded large-scale scenes while preserving highly detailed geometry. Furthermore, the appearance field corresponding to this lightweight mesh is compatible with the classical rasterization pipeline, enabling efficient rendering. However, our method has limitations. BakedSDF employs a well-designed appearance model based on spherical Gaussians. The potential for enhancing rendering quality exists by incorporating a more sophisticated appearance model. Moreover, the gradient-based mesh optimization in the final stage may still result in mesh distortions, including self-intersections. We plan to employ a carefully designed optimizer to mitigate such distortions in future work.", + "bbox": [ + 496, + 659, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5060", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2", + "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2, 3, 4, 5, 6, 7", + "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[4] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3", + "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 1, 2, 3, 6, 7, 8", + "[6] Jaehoon Choi, Dongki Jung, Taejae Lee, Sangwook Kim, Youngdong Jung, Dinesh Manocha, and Donghwan Lee. Tmo: Textured mesh acquisition of objects with a mobile device by using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16674-16684, 2023. 2", + "[7] Jonathan Cohen, Marc Olano, and Dinesh Manocha. Appearance-preserving simplification. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 115-122, 1998. 2", + "[8] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a vsi system for high performance graphics. Acm siggraph computer graphics, 22(4):21-30, 1988. 2", + "[9] Carl Erikson and Dinesh Manocha. Gaps: General and automatic polygonal simplification. In Proceedings of the 1999 symposium on Interactive 3D graphics, pages 79-88, 1999. 2", + "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[11] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 2, 3", + "[12] Michael Garland. Quadric-based polygonal surface simplification. Carnegie Mellon University, 1999. 2, 4, 5, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 2, 4", + "[14] Michael Garland and Paul S Heckbert. Simplifying surfaces with color and texture using quadric error metrics. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 263-269. IEEE, 1998. 2", + "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4", + "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2, 3, 5", + "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 2, 3", + "[18] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 7", + "[19] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2, 3", + "[20] Hugues Hoppe. View-dependent refinement of progressive meshes. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 189-198, 1997. 2", + "[21] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 3", + "[22] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. 2020. 3", + "[23] Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057, 2020. 3", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 6, 7", + "[26] Leif Kobbelt. *Sqrt(3)-subdivision*. In *Proceedings of the 27th annual conference on Computer graphics and interactive techniques*, pages 103–112, 2000. 6", + "[27] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5061", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5", + "[28] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3", + "[29] Peter Lindstrom. Out-of-core simplification of large polygonal models. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 259-262, 2000. 3", + "[30] Peter Lindstrom and Claudio T Silva. A memory insensitive technique for large model simplification. In Proceedings Visualization, 2001. VIS'01., pages 121-550. IEEE, 2001. 3", + "[31] Selena Zihan Ling, Nicholas Sharp, and Alec Jacobson. Vectoradam for rotation equivariant geometry optimization. Advances in Neural Information Processing Systems, 35:4111-4122, 2022. 3", + "[32] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 2", + "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3", + "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 3, 4", + "[35] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2", + "[36] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, volume 40, pages 101-113. Wiley Online Library, 2021. 5", + "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3", + "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3", + "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 1, 2, 3, 7", + "[40] Andrew Nealen, Takeo Igarashi, Olga Sorkine, and Marc Alexa. Laplacian mesh optimization. In Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia, pages 381-389, 2006. 5", + "[41] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021. 2, 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1, 2", + "[43] Werner Palfinger. Continuous remeshing for inverse rendering. Computer Animation and Virtual Worlds, 33(5):e2101, 2022. 2, 3, 5, 6", + "[44] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. In Proc. of the International Conf. on 3D Vision (3DV), 2023. 1, 2, 6, 7", + "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[46] Jarek Rossignac and Paul Borrel. Multi-resolution 3d approximations for rendering complex scenes. In Modeling in computer graphics: methods and applications, pages 455-465. Springer, 1993. 2, 3", + "[47] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 2", + "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 3", + "[49] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2, 3", + "[50] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2022. 1, 2, 3, 5, 6, 7, 8", + "[51] Chao Wang and Xiaohu Guo. Plane-based optimization of geometry and texture for rgb-d reconstruction of indoor scenes. In 2018 International Conference on 3D Vision (3DV), pages 533-541. IEEE, 2018. 3", + "[52] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2", + "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5062", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 4", + "[55] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv, 2023. 1, 2, 4, 6, 7, 8", + "[56] Zehao Yu, Anpei Chen, Bozidar Antic, Songyou Peng, Apratim Bhattacharyya, Michael Niemeyer, Siyu Tang, Torsten Sattler, and Andreas Geiger. Sdfstudio: A unified framework for surface reconstruction, 2022. 6", + "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + ], + "bbox": [ + 78, + 90, + 467, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5063", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_model.json b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22f60e0840ca6d61802643e628dbf5388bac42f1 --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.845, + 0.177 + ], + "angle": 0, + "content": "LTM: Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.202, + 0.856, + 0.258 + ], + "angle": 0, + "content": "Jaehoon Choi\\(^{1}\\) Rajvi Shah\\(^{2}\\) Qinbo Li\\(^{2}\\) Yipeng Wang\\(^{2}\\) Ayush Saraf\\(^{2}\\) Changil Kim\\(^{2}\\) Jia-Bin Huang\\(^{1,2}\\) Dinesh Manocha\\(^{1}\\) Suhib Alsisan\\(^{2}\\) Johannes Kopf\\(^{2}\\) \\(^{1}\\)University of Maryland \\(^{2}\\)Meta" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.675 + ], + "angle": 0, + "content": "Advancements in neural signed distance fields (SDFs) have enabled modeling 3D surface geometry from a set of 2D images of real-world scenes. Baking neural SDFs can extract explicit mesh with appearance baked into texture maps as neural features. The baked meshes still have a large memory footprint and require a powerful GPU for real-time rendering. Neural optimization of such large meshes with differentiable rendering pose significant challenges. We propose a method to produce optimized meshes for large unbounded scenes with low triangle budget and high fidelity of geometry and appearance. We achieve this by combining advancements in baking neural SDFs with classical mesh simplification techniques and proposing a joint appearance-geometry refinement step. The visual quality is comparable to or better than state-of-the-art neural meshing and baking methods with high geometric accuracy despite significant reduction in triangle count, making the produced meshes efficient for storage, transmission, and rendering on mobile hardware. We validate the effectiveness of the proposed method on large unbounded scenes from mip-NeRF 360, Tanks & Temples, and Deep Blending datasets, achieving at-par rendering quality with \\(73 \\times\\) reduced triangles and \\(11 \\times\\) reduction in memory footprint." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.723, + 0.21, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Photorealistic reconstruction and rendering of real-world objects and scenes is a longstanding problem of importance with several applications in computer vision, robotics and AR/VR. In recent years, Neural Radiance Fields (NeRFs) [1, 2, 38, 39] have been quite successful at novel view synthesis. However, the scene representation is learned as an implicit volumetric representation that is expensive to evaluate. This is a crucial limitation for many applications, such as Augmented and Virtual Reality (AR/VR), often requiring real-time rendering with low memory and compute power." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.291, + 0.744, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.389, + 0.744, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.291, + 0.862, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.35, + 0.862, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.41, + 0.862, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.476, + 0.895, + 0.547 + ], + "angle": 0, + "content": "Figure 1. Example results using LTM showing mesh geometry (top-left) and novel-view synthesis (right). Bottom right plot compares performance of different methods on mip-NeRF360 outdoor dataset. Our method achieves comparable rendering quality to BakedSDF while reducing the mesh size considerably." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Since modern graphics engines are highly optimized to rasterize triangle meshes, researchers have started exploring ways to extract explicit mesh representations from volumetric functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.896, + 0.854 + ], + "angle": 0, + "content": "MobileNeRF [5] uses a polygon mesh with texture maps storing feature vectors and opacity. Although their rendering quality is good, the mesh quality is far from an ideal surface. While NeRF2Mesh [50] and NeRFMeshing [44], based on volumetric density fields, excel at generating highly detailed geometry, they often result in noisy and bumpy surfaces and numerous floaters, especially in planar regions, which are common in large-scale scenes. In contrast, neural Signed Distance Function (SDF) methods [42, 52, 54] typically excel in reconstructing accurate surface geometry. However, they may not be able to capture small or detailed geometry, and their rendering quality is inferior to volumetric density-field methods. BakedSDF [55] overcomes these limitations by applying mip-NeRF 360 techniques to train SDF representation, though the resulting mesh representation has a high memory overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Previous neural SDF methods that rely on the Marching Cube algorithm [34] result in meshes with high storage overhead. This results in a significant challenge for practi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5053" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "cal applications on mobile devices, where low disk storage and fast data transfer are essential. Instead of relying on the Marching Cubes, some methods [11,28,48,49] propose differentiable mesh reconstruction methods. These methods aim to extract an isosurface from an implicit function and can be seamlessly integrated with differentiable rendering approaches [16, 17]. However, most methods primarily concentrate on a single object with ideal implicit surface functions, unsuited for large unbounded real-world scenes [2, 55]. Reconstructing an ideal mesh for the unbounded large-scale scene is still an open problem, often resulting in non-manifold geometry or missing surfaces. Consequently, the renderings can result in visual artifacts due to the incompleteness of the mesh geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.473, + 0.488 + ], + "angle": 0, + "content": "Main Results: In this paper, our goal is to achieve a realistic rendering with a compact mesh representation that optimally utilizes disk storage for unbounded large-scale scenes, all while preserving geometric quality. Our method initially trains neural fields for geometry and appearance separately. For geometry reconstruction, we adopt the BakedSDF approach [55] to train an SDF in a contracted space [2]. Regarding the appearance field, we utilize hash-based encoding [39] with two shallow MLPs to represent diffuse and specular colors. The SDF and appearance fields are compact representations for reconstructing a high-quality mesh and achieving realistic rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.473, + 0.656 + ], + "angle": 0, + "content": "Given the initial geometry representations, we perform mesh extraction and decimation. As unbounded large-scale scenes occupy a large amount of disk storage, our method focuses on generating a lightweight mesh without compromising geometric details. We adaptively extract the mesh for the central and background regions in the contracted space, respectively. Subsequently, we modify a mesh decimation [14] designed for large-scale scenes. Our method eliminates redundant triangles corresponding to simple geometric features, such as planar surfaces, while preserving complex geometric features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Since the appearance field is trained by volumetric rendering, the initial model is unsuitable for rasterization-based rendering. Moreover, the mesh from the previous step still fails to accurately represent areas with highly-detailed geometry (thin structures) or semi-transparent objects. These mesh defects hurt rendering quality. Thus, we use differentiable rendering [27] and utilize image-space supervision to refine the mesh surface by rectifying these mesh defects while preserving geometric accuracy. Due to the highly simplified mesh and various regularization techniques, our approach can deform the mesh topology without causing severe mesh distortion, a common issue in gradient-based mesh optimization [41, 43]. Compared to mesh-based neural rendering methods [5, 44, 50, 55], our approach results in lower storage overhead while providing comparable rendering quality. To sum up, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.138 + ], + "angle": 0, + "content": "- We effectively utilize classical mesh decimation techniques to extract compact geometry from neural SDFs of large unbounded scenes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.146, + 0.894, + 0.192 + ], + "angle": 0, + "content": "- We introduce a new approach for joint optimization of appearance and vertex deformation that preserves geometric details and improves rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.201, + 0.894, + 0.261 + ], + "angle": 0, + "content": "- We achieve a trifecta of precise geometry, efficient storage, and rendering quality comparable to the state of the art on diverse captured environments such as mip-NeRF 360, Tanks & Temples, and the Deep Blending dataset." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.894, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.272, + 0.642, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.297, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Neural Rendering Neural Fields achieve remarkable photo-realistic rendering quality. One of the most successful approaches is Neural Radiance Fields (NeRFs) [38], which adopts volumetric implicit representations and trains a radiance field through ray marching. Mip-NeRF [1, 2] presents casting a conical frustum into the scene space and extends this integrated positional encoding to unbounded scenes by using a novel contraction function. Instead of learning volumetric density, neural surface reconstruction methods [6, 42, 52, 54] propose learning a Signed Distance Function (SDF) for surface reconstruction via volumetric rendering. Furthermore, several methods have been proposed to accelerate NeRFs by utilizing efficient data structure [3, 10, 39], although achieving real-time rendering for high-resolutions remains a challenge. Alternatively, to enhance rendering speed, \"baking\" based methods [5, 19, 44, 45, 50, 55] typically precompute and store a trained NeRF into a more efficient representation. SNeRG [19] adopts a deferred shading model [8] and bakes trained NeRFs into sparse voxel grid representations with neural features. In particular, the mesh is widely used as geometry representation and bakes neural fields into its texture. MobileNeRF [5] builds on a mesh rasterization pipeline for efficient rendering on mobile architecture. NeRF2Mesh [50] introduces a two-step framework for NeRF training and textured mesh extraction. NeRFMeshing [44] proposes a novel method to produce a truncated SDF from NeRFs and extracts the accurate mesh. BakedSDF [55] enables us to reconstruct high-quality mesh and bake it into a small appearance model based on spherical Gaussians for fast rendering. DNMP [35] and NeuRaS [32] adopt mesh-based geometry and rasterization-based rendering for large driving scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Mesh Processing Traditional graphics research has developed effective techniques for simplifying meshes, based on vertex decimation [46, 47], edge collapse [9, 12-14, 20] and appearance preservation [7]. Vertex decimation methods prioritize vertices based on shape heuristics and remove the least important ones. Edge collapse methods rank the edges based on the cost and subsequently contract pairs of vertices. QSlim [12, 13] introduces the concept of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5054" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "Quadric Error Metric (QEM), which is defined as the sum of quadrics representing the distance of a point from a set of planes. Following the edge collapse, they utilize this QEM to identify which edges should be removed while preserving geometric topology. Lindstrom et al. [29, 30] use vertex clustering [46] on a uniform grid and accumulate a quadric error matrix to the occupied grid cell. Wang et al. [51] employ RGB-D sequences to reconstruct a lightweight mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.473, + 0.532 + ], + "angle": 0, + "content": "Differentiable Rendering and Mesh Optimization Recently, many differentiable rendering techniques [23] have emerged, aimed at facilitating end-to-end optimization of the rendering process by employing useful gradients. In particular, several works [4, 33] focus on approximating the rasterization stage of the rendering to compute meaningful gradients. Furthermore, modular differentiable renderers [21, 22, 27] exist, allowing users to customize and modify specific functions. Differentiable mesh reconstruction techniques [11, 28, 48] can build a mesh through gradient-based optimization. In particular, differentiable rendering [16, 17, 49] can offer 2D supervision to refine and enhance the geometric quality of the mesh. Some works [16, 17] adopt an inverse rendering approach, leveraging physics-based rendering to synthesize shapes. Nevertheless, gradient-based mesh optimization [41] is highly sensitive to initialization and prone to significant mesh distortion. Some methods [31, 41, 43] incorporate a novel ADAM optimizer and employ strong regularization techniques. However, this research is currently limited to generating single objects and is challenging to apply to large-scale scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.208, + 0.565 + ], + "angle": 0, + "content": "3. Our Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Given multi-view images and corresponding camera poses, we aim to reconstruct the texture and geometry of unbounded large-scale scenes. Fig. 2 shows the three main stages of our pipeline: Initialization, Mesh Processing, and Optimization. In the first stage, we employ volumetric rendering to train neural implicit representation for geometry \\( f_{sdf} \\) and appearance \\( f_{color} \\). Subsequently, we leverage the geometry representation \\( f_{sdf} \\) to extract a high-resolution mesh via Marching Cubes [34] and reduce the number of triangles without causing geometry distortion. Next, we utilize differentiable rendering [27] to jointly optimize the explicit 3D mesh \\( M \\) and implicit appearance field \\( f_{color} \\) obtained from the first stage and refine the mesh surface." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.444, + 0.803 + ], + "angle": 0, + "content": "3.1. Initialization of Geometry and Appearance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this stage, we first learn neural implicit representations for both appearance and geometry. A neural radiance field [38] is an implicit representation of a 3D scene volume, a mapping from a 3D position \\( x \\) and a ray viewing direction \\( d \\in \\mathbb{R}^3 \\) to an RGB color \\( c \\) and volume density \\( \\rho \\). In practice, the networks modeled by MLPs are used to generate this" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.089, + 0.878, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.168, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Figure 2. Overview of Our Pipeline. It consists of three steps: Initialization (Sec. 3.1), Mesh Processing (Sec. 3.2), and Optimization (Sec. 3.3). Our method focuses on baking the initial appearance and geometry representation, trained by volumetric rendering, into an explicit mesh and a memory-efficient texture conducive to real-time rendering. We employ classical mesh decimation to minimize storage requirements, introduce the optimization stage to facilitate rasterization-based rendering and refine the detailed mesh structure." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.302, + 0.783, + 0.317 + ], + "angle": 0, + "content": "volume density and view-dependent color:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.326, + 0.892, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\rho = f _ {\\text {d e n s i t y}} (x), \\quad c = f _ {\\text {c o l o r}} (x, d). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.426 + ], + "angle": 0, + "content": "To render the color of a pixel, we first compute the 3D points \\( t_i \\) along the viewing ray associated with each pixel \\( p \\) as \\( x(t) = o + td \\). Then, estimated volume densities \\( \\rho_i \\) and colors \\( c_i \\) are utilized to approximate a volume rendering equation [37] using numerical quadrature:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.433, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j < i} \\rho_ {i} \\delta_ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.594 + ], + "angle": 0, + "content": "where \\( T_{i} \\) is the accumulated transmittance and \\( \\alpha_{i} = 1 - \\exp(-\\rho_{i}\\delta_{i}) \\) is the discrete opacity of the point. \\( \\delta_{i} = t_{i} - t_{i-1} \\) is the distance between two neighboring points along the ray. To deal with large unbounded scenes, mip-NeRF 360 [2] introduces a proposal MLP for efficient sampling and space contraction to represent an unbounded scene. The scene contraction function \\( f \\) (in Eq. 3) maps the unbounded points into bounded space:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.602, + 0.892, + 0.655 + ], + "angle": 0, + "content": "\\[\nf (x) = \\left\\{ \\begin{array}{l l} x & | | x | | \\leq 1, \\\\ \\left(2 - \\frac {1}{| | x | |}\\right) \\frac {x}{| | x | |} & | | x | | \\geq 1. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Appearance Similar to previous works [5, 19, 50], we decompose an appearance field \\( f_{color} \\) into shallow MLPs to learn diffuse colors \\( c_d \\) and specular colors \\( c_s \\) separately. Particularly, the view-dependent effects are modeled by a small MLP that can be easily integrated into a fragment Shader. In order to ensure the MLPs are able to model an unbounded scene well, we adopt strategies similar to the ones introduced in mip-NeRF360 work, specifically, scene contraction, proposal MLP, and regularization for the background. We train both MLP networks \\( f_{color} \\) and \\( f_{density} \\) with a multi-resolution hash positional encoding [39] using the volume rendering objective in Eq. 2. The appearance field trained in this stage is further optimized to facilitate rasterization-based rendering, replacing the computationally expensive ray marching. This optimization is described later in Section 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5055" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.097, + 0.837, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Figure 3. Our Approach: In Section 3.1, we start training both the appearance \\( f_{color} \\) and geometry \\( f_{sdf} \\) representations for a large-scale scene using volumetric rendering (shown in orange). Then, in Section 3.2, our method extracts the mesh from \\( f_{sdf} \\) and significantly simplifies its structure, resulting in the lightweight mesh \\( M \\) (shown in blue). In Section 3.3, we jointly train vertex deformation \\( \\Delta x \\) and \\( f_{color} \\) computed from appearance modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.473, + 0.476 + ], + "angle": 0, + "content": "Geometry Initial geometry modeling in our method is similar to BakedSDF [55]. The core mechanism for neural surface reconstruction was introduced in VolSDF [54] for small objects. BakedSDF shows that by leveraging the efficient sampling and contraction ideas introduced in mipNeRF360 [2], VolSDF [54] can be extended to large unbounded scenes. We follow this and jointly train two MLPs using the volumetric rendering objective, (i) \\( f_{sdf} \\) - representing geometry of a scene as a signed distance function, and (ii) \\( g_{color} \\) - a color prediction network, as shown in the Geometry Reconstruction block in Fig.3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Following VolSDF, we define the volume density \\(\\sigma(x)\\) as a function of surface geometry \\(f_{sdf}\\), namely, \\(\\sigma(x) = \\alpha \\Psi_{\\beta}(f_{sdf}(x))\\), where \\(\\Psi_{\\beta}\\) is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and \\(\\beta\\) is the learnable scaling parameter. The Eikonal loss [15] to regularize the \\(f_{sdf}\\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.574, + 0.47, + 0.593 + ], + "angle": 0, + "content": "\\[\nL _ {e i k} = \\left(\\| \\nabla f _ {s d f} (x) \\| _ {2} - 1\\right) ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.72 + ], + "angle": 0, + "content": "The color MLP \\( g_{color} \\), in addition to the position and viewing direction, also takes the predicted surface normal and a 256-dimensional geometric feature as inputs. It is important to note that the sole purpose of color MLP \\( g_{color} \\) is geometry initialization and we discard it after this step. This is not to be confused with the shallow decomposed color MLP \\( f_{color} \\) trained specifically to model appearance that is also jointly refined with mesh vertices in the later stage." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.369, + 0.743 + ], + "angle": 0, + "content": "3.2. Mesh Extraction and Decimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this stage, we first extract a mesh by using Marching Cubes (MC) [34] algorithm on learned SDF (\\(f_{sdf}\\) evaluated over a uniform grid) and simplify the extracted mesh using a geometry-aware quadric error based mesh decimation. While this approach works pretty well for extracting meshes of well-captured small-scale objects and scenes, it presents unique challenges for large-scale scenes. Meshing neural SDFs of large-scale scenes with marching cubes can often result in spurious surfaces arising from (i) insufficient grid resolution to model complex geometry of nearby" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.417 + ], + "angle": 0, + "content": "objects, and (ii) poorly learned geometry of distant objects due to insufficient view-points observing such distant regions. Furthermore, since our SDF \\( f_{sdf} \\) is trained in the contracted space, the resulting mesh is also defined in the contracted space. When we uncontract, i.e. warp the mesh vertices back into the original Euclidean space, it further exacerbates the geometry distortion in the background region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.629 + ], + "angle": 0, + "content": "We aim to preserve the detailed geometry in central (foreground) regions while achieving a coarser but less noisy reconstruction of distant (background) regions, within a small triangle budget. We achieve this by treating foreground and background regions distinctly during both extraction and simplification steps. During MC extraction, we set a finer grid resolution of 2048 and a spatial range of \\([-1,1]^3\\) for the central region, and a coarser resolution of 1024 and a spatial range of \\([-2,2]^3\\) for the background region. For decimation, we first partition the central and background regions and employ an aggressive mesh decimation in background regions. This helps us transform bumpy and noisy background areas into smooth surfaces while also reducing redundant triangles." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.782 + ], + "angle": 0, + "content": "We first apply vertex clustering and then run quadric error based simplification algorithm - QSlim [12, 13] for all clusters. Then, we run edge contraction for each cluster to reach a threshold and repeat this process until convergence. The QSlim algorithm calculates a quadric metric for each vertex to decide whether the two vertices should be merged and aggregates these quadrics during each edge collapse operation. The basic quadric error in QSlim [13] is defined as the point-to-plane distance from vertex \\( v \\in \\mathbb{R}^3 \\) to all planes \\( p \\in \\mathbb{R}^4 \\) of its adjacent faces \\( N(v) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.793, + 0.892, + 0.841 + ], + "angle": 0, + "content": "\\[\nE (v) = \\sum_ {p \\in N (v)} \\left(p ^ {T} v\\right) ^ {2} = v ^ {T} \\left(\\sum_ {p \\in N (v)} p p ^ {T}\\right) v = v ^ {T} Q _ {v} v. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The quadric \\( Q_{v} \\) is the symmetric 4x4 matrix, and each edge computes this quadric by adding the quadrics of the corresponding two vertices. The most commonly used quadric error is defined solely based on 3D vertex positions \\( v \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5056" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "Intuitively, the most effective strategy is to represent planar surfaces with a smaller number of large triangles while keeping a consistent number of triangles for complex shapes. Variants that incorporate additional surface attributes, such as normals or textures in calculating the quadric error are proposed in [12]. We experimentally observed that compared to the basic QSlim algorithm, using both the vertex positions \\( v \\in \\mathbb{R}^3 \\) and the corresponding vertex normal \\( n \\in \\mathbb{R}^3 \\) for quadric error is very effective at prioritizing triangle decimation on flat surfaces while retaining them on complex surfaces. The combination of vertex clustering and geometry-aware QSlim helps us in effectively reducing the number of triangles while preserving the original geometry." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.316, + 0.427, + 0.332 + ], + "angle": 0, + "content": "3.3. Appearance and Geometry Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.386 + ], + "angle": 0, + "content": "In this stage, we optimize the appearance model jointly with mesh vertex deformation and refine the mesh topology using a geometric complexity based criterion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.473, + 0.719 + ], + "angle": 0, + "content": "Joint optimization of appearance and vertex deformation Given the simplified mesh \\(M\\) and the initial appearance field \\(f_{color}\\), we employ the fast differentiable rasterizer introduced in Nvdiffrast [27] to jointly optimize the mesh geometry and the appearance field. The initial appearance field \\(f_{color}\\) trained using the volume rendering objective (Eq. 1) is agnostic of the surface geometry of the extracted and simplified mesh. To finetune the model to fit the mesh geometry, we rasterize the mesh into each of the training camera poses and obtain surface points \\(s\\) and face normals \\(N\\) interpolated per pixel. Then we apply the contraction function in Eq. 3 to warp the surface points. This is because the appearance field defined in contracted space proves to be more effective in enhancing the rendering quality [2]. The appearance field \\(f_{color}\\) takes these surface points \\(s\\) and viewing direction \\(d\\) as input, rendering an image \\(\\hat{I}_i\\). Initially, rendering performance is degraded because the appearance field only utilizes the intersection point between the ray and the surface rather than accumulating all sampled points along the ray. We finetune this appearance field to be compatible with mesh rendering, akin to the functionality of a fragment Shader." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.887 + ], + "angle": 0, + "content": "Given the positions \\( x \\) of the mesh vertices, we train per-vertex deformation \\( \\Delta x \\) using differentiable rendering [27]. We explore two approaches to optimize vertex deformation: a standard ADAM optimizer [24] and a rotation-invariant ADAM optimizer [43]. The ADAM optimizer is widely employed [16, 50] for deforming mesh topology to enhance rendering performance. While the ADAM optimizer often results in mesh distortions such as flipped triangles and self-intersections, our method mitigates mesh distortion, especially on smooth surfaces, due to the rigorous mesh decimation performed in Section 3.2." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Objective Function To jointly train appearance field and vertex deformation, we minimize the image-space L1 loss \\(L_{\\text{appearance}}\\) between the rendered image \\(\\hat{I}_i\\) of the deformed surface and the groundtruth image \\(I_i\\) for a camera view \\(P_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.159, + 0.892, + 0.19 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {a p p e a r a n c e}} = \\sum_ {i} \\left| \\hat {I} _ {i} - I _ {i} \\right|. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.892, + 0.304 + ], + "angle": 0, + "content": "Naively training vertex deformation is highly likely to result in undesirable mesh topology characterized by a disorganized arrangement of triangles, often referred to as a \"polygon soup.\" To prevent this mesh distortion, we employ various regularization methods. We use the Laplacian mesh smoothness loss and the normal consistency loss inspired by [36]." + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.314, + 0.892, + 0.337 + ], + "angle": 0, + "content": "\\[\nL _ {l a p} = \\sum \\| L V \\| _ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.432 + ], + "angle": 0, + "content": "where \\(V \\in \\mathbb{R}^{N \\times 3}\\) (\\(N\\) is the number of vertices) is the matrix of deformed vertex coordinates, indicating \\(x + \\Delta x\\), at each training step, and \\(L \\in \\mathbb{R}^{N \\times N}\\) is the graph Laplacian of the mesh [40]. Minimizing the magnitude of a vertex's differential coordinates (\\(LV\\)) reduces its distance from its neighbors' average position. The normal consistency loss is" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.439, + 0.892, + 0.472 + ], + "angle": 0, + "content": "\\[\nL _ {n c} = \\sum_ {i, j \\in F} \\left(1 - n _ {i} \\cdot n _ {j}\\right) ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.6 + ], + "angle": 0, + "content": "where \\( n_i \\) and \\( n_j \\) denote the surface normal of two adjacent faces and \\( F \\) is the set of faces sharing a common edge. Additionally, in Section 3.2, a high-resolution mesh is rasterized in the first pass before decimation. The output of this rasterization is used to interpolate face normals per pixel, resulting in a normal map \\( \\hat{N}_i \\) for every viewpoint. We calculate the L1 loss between a rendered normal \\( N_i \\) from current mesh \\( M \\) and the \\( \\hat{N}_i \\):" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.606, + 0.892, + 0.637 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o r m}} = \\sum_ {i} | \\hat {N} _ {i} - N _ {i} |. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.643, + 0.892, + 0.688 + ], + "angle": 0, + "content": "\\(L_{norm}\\) provides geometry supervision, particularly in planar regions, helping prevent severe mesh distortion when moving vertices. The total loss function \\(L\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.696, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {c o l o r}} + \\lambda_ {\\text {l a p}} L _ {\\text {l a p}} + \\lambda_ {\\text {n c}} L _ {\\text {n c}} + \\lambda_ {\\text {n o r m}} L _ {\\text {n o r m}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.72, + 0.88, + 0.735 + ], + "angle": 0, + "content": "where \\(\\lambda_{*}\\) is the weight to balance between the loss terms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Mesh Refinement Inspired by NeRF2Mesh [50], we adopt a classical mesh refinement approach based on mesh subdivision and decimation during training. The mesh refinement strategy is applied differently to the central and background regions. Instead of rendering error proposed by [50], which can arise from many factors including specularity, we employ a criterion: the vertex deformation values for subdividing triangle faces. The large vertex deformation implies a more substantial triangle budget to accurately capture the intricacies of highly detailed geometry associated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5057" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.357, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.091, + 0.62, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.091, + 0.755, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.091, + 0.892, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.893, + 0.376 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of Our Method, NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. Our method and NeRF2Mesh show shading mesh and wireframe mesh extracted without texture, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.473, + 0.583 + ], + "angle": 0, + "content": "with those vertices. We sort the absolute values of vertex deformation and determine a \\(10\\%\\) of faces to increase triangles. Then, we apply \\(\\sqrt{3}\\) subdivision [26], which adds a vertex at the barycenter of a triangle and connects it to all vertices of this triangle. This scheme enables us to augment the triangle budget for complex geometry and mitigate self-intersection during vertex deformation. Furthermore, the mesh decimation presented in Section 3.2 is applied after every mesh subdivision to remove redundant triangles on simple and flat geometry. For background regions, we examine all training views and check faces intersecting with ray casting. We utilize dilation for all invisible masks and employ mesh decimation to remove redundant faces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.473, + 0.69 + ], + "angle": 0, + "content": "Rotation-Invariant ADAM The rotation-invariant ADAM formulation [43] modifies the update rule in ADAM, resulting in the update step: \\(\\Delta x = \\zeta \\cdot v \\cdot l_{ref}\\). Here, \\(\\zeta\\) is the learning rate, and \\(v\\) is the standard ADAM update step, which is the loss gradient with respect to \\(x\\). \\(l_{ref}\\) is the coefficient initially set to the edge length incident to \\(x\\), and it gets updated based on the average \\(v\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.702, + 0.21, + 0.719 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.271, + 0.743 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Datasets: To validate the effectiveness of our method, we mainly conduct experiments on seven unbounded, large-scale \\(360^{\\circ}\\) indoor and outdoor scenes from the mip-NeRF 360 [2] dataset. Also, we validate the generalization ability of our approach on two different datasets, including large-scale scenes: two scenes from the Tanks & Temples [25] dataset and two scenes from the Deep Blending [18] dataset. Implementation Details: For all datasets, we subsampled every eight frames for test sets following the mip-NeRF 360 paper [2]. For geometry reconstruction in Section 3.1, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.895, + 0.553 + ], + "angle": 0, + "content": "use SDFstudio [56] to implement BakedSDF. In Section 3.2, we reduce \\(95\\%\\) of the triangles in the central region and \\(99\\%\\) of the triangles in the background regions through decimation. In Section 3.3, we run all experiments for 40k iterations, and a learning rate is annealed from 1e-2 to 1e-3. We set the loss weights to \\(\\lambda_{lap} = 1\\mathrm{e - }3\\), \\(\\lambda_{nc} = 1\\mathrm{e - }3\\), and \\(\\lambda_{norm} = 1\\mathrm{e - }4\\). We apply mesh refinement at 8k, 16k, and 24k training steps. Subsequently, we compute diffuse color and specular features using \\(f_{color}\\) to export textures. All these color maps and features are 3-channel .png files. Additional details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.704, + 0.577 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.809 + ], + "angle": 0, + "content": "We compare our proposed method with three different methods: 1) MobileNeRF [5] is the state-of-the-art method, enabling NeRF to achieve real-time rendering of explicit 3D mesh and neural features. 2) NeRF2Mesh [50] employs a volumetric rendering to reconstruct the coarse mesh and refine the mesh by jointly optimizing appearance and geometry. 3) BakedSDF [55] is a recent method for optimizing a neural surface-volume representation for real-time rendering of large unbounded scenes. We also add NeRFMeshing [44], which distills the rendering NeRF into a neural field that represents TSDF (Truncated SDF) and reconstructs the 3D mesh. As BakedSDF and NeRFMeshing have not released public code, we adopt the metrics provided in their original papers. BakedSDF has made its GLB file available. We utilize this file for visualization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Figure 4 qualitatively compares our method with NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. We visualize the shading meshes extracted without texture, which is the most effective way to visualize geometric differences. Furthermore, we include a wireframe mesh, distinguishing between our method and NeRF2Mesh. Visu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5058" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.872, + 0.205 + ], + "angle": 0, + "content": "
MethodMeshOutdoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)Indoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)
BicycleGardenStumpMeanRoomCounterKitchenBonsaiMean
Instant-NGP [39]X22.1 / 0.49 / 0.4924.5 / 0.65 / 0.3123.6 / 0.57 / 0.4523.4 / 0.57 / 0.4229.2 / 0.85 / 0.3026.43 / 0.80 / 0.3428.5 / 0.82 / 0.2530.3 / 0.89 / 0.2328.6 / 0.84 / 0.28
mip-NeRF 360 [2]X24.4 / 0.68 / 0.3026.9 / 0.81 / 0.1726.4 / 0.74 / 0.2625.9 / 0.75 / 0.2431.6 / 0.91 / 0.2129.5 / 0.89 / 0.2032.3 / 0.92 / 0.1333.5 / 0.94 / 0.1831.7 / 0.92 / 0.18
MobileNeRF [5]O21.7 / 0.43 / 0.5118.8 / 0.59 / 0.3623.9 / 0.56 / 0.4321.5 / 0.53 / 0.4328.9 / 0.85 / 0.2825.1 / 0.72 / 0.2926.8 / 0.79 / 0.7923.8 / 0.71 / 0.7226.2 / 0.77 / 0.52
NeRF2Mesh [50]O22.1 / 0.48 / 0.5123.4 / 0.55 / 0.4022.5 / 0.54 / 0.4622.7 / 0.52 / 0.4625.7 / 0.79 / 0.3523.9 / 0.71 / 0.3524.0 / 0.61 / 0.3625.0 / 0.77 / 0.2924.7 / 0.72 / 0.34
NeRFMeshing [44]O21.1 / - / -22.9 / - / -22.6 / - / -22.2 / - / -26.1 / - / -20.0 / - / -23.6 / - / -25.6 / - / -23.8 / - / -
BakedSDF [55]O22.0 / 0.57 / 0.3724.9 / 0.75 / 0.2123.6 / 0.59 / 0.3723.5 / 0.64 / 0.3228.7 / 0.87 / 0.2525.7 / 0.81 / 0.2826.7 / 0.82 / 0.2427.2 / 0.85 / 0.2627.0 / 0.84 / 0.26
OursO22.4 / 0.52 / 0.4423.5 / 0.64 / 0.3023.7 / 0.57 / 0.4223.2 / 0.58 / 0.3929.3 / 0.88 / 0.2325.1 / 0.77 / 0.2726.5 / 0.80 / 0.2127.3 / 0.84 / 0.2227.1 / 0.82 / 0.23
Ours w/o OptO16.1 / 0.30 / 0.5318.8 / 0.45 / 0.3919.4 / 0.39 / 0.5018.1 / 0.38 / 0.4817.2 / 0.54 / 0.4616.2 / 0.47 / 0.4718.1 / 0.50 / 0.3616.4 / 0.54 / 0.4317.0 / 0.51 / 0.43
Ours w/ GOO21.1 / 0.48 / 0.4723.2 / 0.65 / 0.3023.0 / 0.53 / 0.4622.5 / 0.55 / 0.4127.5 / 0.85 / 0.2623.6 / 0.72 / 0.3325.5 / 0.78 / 0.2325.9 / 0.76 / 0.325.4 / 0.78 / 0.28
Ours w. RADAMO21.6 / 0.48 / 0.4623.4 / 0.66 / 0.3023.5 / 0.56 / 0.4522.8 / 0.57 / 0.4027.5 / 0.85 / 0.2624.1 / 0.74 / 0.3125.8 / 0.80 / 0.2226.8 / 0.80 / 0.2626.0 / 0.80 / 0.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.211, + 0.89, + 0.239 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison on mip-NeRF 360 Dataset. “-” denotes a missing implementation. The best and second best-performing algorithms for each metric are bolded and underlined." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.245, + 0.888, + 0.302 + ], + "angle": 0, + "content": "
MethodBicycleGardenStumpRoomCounterKitchenBonsaiMean
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]2011435 / 6638135723 / 3401641069 / 497220903 / 421170864 / 39995552 / 25684606 / 284153879 / 1262
NeRF2Mesh [50]129815 / 858117650 / 779123415 / 82069328 / 46279517 / 38489601 / 43470319 / 63697387 / 667
BakedSDF [55]1087111248 / 4438451310773M / 2107294319667 / 384673857227 / 1411058512212 / 2389265913832 / 2705664113629 / 266796871564M / 27951
Ours83270 / 52950113 / 218120425 / 84745124 / 23858138 / 2684070 / 13642233 / 45962196 / 385
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.306, + 0.89, + 0.334 + ], + "angle": 0, + "content": "Table 2. Disk Storage (MB), Number of Vertices (V) and Faces (F). The unit for the number of vertices and faces is \\( {10}^{3} \\) . \"M\" denotes the \\( {10}^{6} \\) unit. We save \\( {73} \\times \\) the number of triangles and reduce \\( {11} \\times \\) disk storage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.47, + 0.54 + ], + "angle": 0, + "content": "alizing the wireframe mesh from BakedSDF lacks meaningful interpretation due to its extremely high face density. MobileNeRF generates a \"triangle soup\" that inaccurately represents the scene's geometry. Also, compared to the mesh reconstructed by NeRF2Mesh, our method shows sharp details of complex geometry and smooth surfaces on the ground or wall. Moreover, in the wireframe mesh, our method employs more triangles in intricate shapes like the bicycle or the garden vase while allocating smaller triangles to planar regions or simpler structures. While the geometry of BakedSDF exhibits superior quality compared to our mesh reconstruction, achieving this high-quality mesh necessitates approximately \\(73 \\times\\) more triangles in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Our method effectively achieves the optimal tradeoff between rendering quality and disk storage cost. For comparison, we report two tables for rendering quality and disk storage. Table 1 reports the rendering quality using the standard PSNR, SSIM [53], and LPIPS [57] on mip-NeRF 360 dataset. Instant-NGP [39], and mip-NeRF 360 [2] are volumetric rendering approaches that do not reconstruct explicit geometry, a topic that lies outside the primary focus of our study. In Table 2, we report the disk storage in megabytes and the number of vertices and faces. For a fair comparison, all the mesh files are saved in .obj and .mtl formats. Texture images are stored as .png files. Other assets, such as view-dependency MLP, are saved as .json files, except BakedSDF, which only provides gLTF format files. Our method demonstrates comparable rendering quality in outdoor scenes (with a 0.3 PSNR drop) and indoor scenes (with a 0.1 PSNR increase) while saving \\(73 \\times\\) triangles and \\(11 \\times\\) storage compared to the state-of-the-art BakedSDF. In comparison to NeRF2Mesh and MobileNeRF, our method shows better-rendering quality (Table 1), demands less disk storage, and fewer triangles (Table 2), and captures superior mesh quality (Fig. 4). In Table 5, we measure the rendering speed and GPU memory by running in-browser on a Mac" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.341, + 0.89, + 0.41 + ], + "angle": 0, + "content": "
MethodMeshDeep Blending [18]Tanks & Temples [25]
DrJohnsonPlayroomBarnCourthouse
Instant-NGP [39]X27.7 / 0.84 / 0.3819.5 / 0.78 / 0.4623.4 / 0.72 / 0.4320.0 / 0.68 / 0.53
mip-NeRF360 [2]X29.1 / 0.90 / 0.2429.6 / 0.90 / 0.2527.2 / 0.81 / 0.3221.3 / 0.72 / 0.47
MobileNeRF [5]O25.9 / 0.78 / 0.3128.7 / 0.85 / 0.2824.4 / 0.77 / 0.2417.7 / 0.54 / 0.36
NeRF2Mesh [50]O24.0 / 0.75 / 0.5126.6 / 0.83 / 0.4123.4 / 0.72 / 0.4318.2 / 0.63 / 0.56
OursO26.5 / 0.82 / 0.4029.1 / 0.87 / 0.3226.5 / 0.83 / 0.2420.2 / 0.73 / 0.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.415, + 0.89, + 0.443 + ], + "angle": 0, + "content": "Table 3. Quantitative Comparison. We report PSNR \\(\\uparrow\\), SSIM \\(\\uparrow\\), and LPIPS \\(\\downarrow\\) on Deep Blending and Tanks & Temples datasets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.448, + 0.889, + 0.496 + ], + "angle": 0, + "content": "
MethodDrJohnsonPlayroomBarnCourthouse
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]174.9602 / 285281.81097 / 521289.21080 / 502263.6934 / 417
NeRF2Mesh [50]108.6414 / 815118.7401 / 793131.9397 / 793145.7451 / 843
Ours26.746 / 9260.9203 / 4014086 / 17147.497 / 192
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.502, + 0.869, + 0.514 + ], + "angle": 0, + "content": "Table 4. Disk Storage (MB), Vertices (V) and Faces (F)." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.515, + 0.872, + 0.545 + ], + "angle": 0, + "content": "
FPS / GPU (MB)MobileNeRFNeRF2MeshOurs
MacBookPro (1920x1080 Resolution)59 / 324205 / 130212 / 115
" + }, + { + "type": "table_caption", + "bbox": [ + 0.518, + 0.549, + 0.872, + 0.562 + ], + "angle": 0, + "content": "Table 5. FPS and GPU memory on mip-NeRF 360 dataset" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.569, + 0.645, + 0.583 + ], + "angle": 0, + "content": "BookPro (2020, M1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Figure 5 compares our method to the state-of-the-art baselines that focus on rendering quality: NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. In their first stage, BakedSDF* represents the volumetric rendering results obtained from a surface-based volumetric representation. The visual results of our approach exhibit sharper details than other methods. In the Bicycle and Room scenes, the mesh surface of the armrest for a bench and tree leaves exhibits noticeable jiggling, resulting in distorted images. Unlike BakedSDF*, our method effectively manages the thin structure of the wheel." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Experimental Results on Deep Blending and Tank & Temples Datasets: We train and evaluate our method on four scenes: Playroom and DrJohnson from the Deep Blending dataset, and Barn and Courthouse from the Tanks & Temples dataset. The former includes bounded indoor scenes, while the latter comprises large unbounded outdoor scenes. Their capture methodology differs from that of the mip-NeRF 360 dataset, and they do not utilize 360-degree scenes. In Tanks & Temples, we exclude sky regions during training and evaluation, as mesh-based rendering tends to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5059" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.262, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.17, + 0.261, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.248, + 0.261, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.328, + 0.261, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.409, + 0.201, + 0.419 + ], + "angle": 0, + "content": "Scene" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.091, + 0.413, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.17, + 0.413, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.248, + 0.413, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.328, + 0.413, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.409, + 0.36, + 0.419 + ], + "angle": 0, + "content": "(a) Ours" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.091, + 0.566, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.17, + 0.566, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.248, + 0.566, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.328, + 0.566, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.45, + 0.409, + 0.531, + 0.419 + ], + "angle": 0, + "content": "(b) NeRF2Mesh" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.091, + 0.719, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.17, + 0.718, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.248, + 0.718, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.409, + 0.681, + 0.419 + ], + "angle": 0, + "content": "(c)BakedSDF\\*" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.091, + 0.871, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.17, + 0.871, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.248, + 0.871, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.328, + 0.873, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.409, + 0.831, + 0.419 + ], + "angle": 0, + "content": "(d) MobileNeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.422, + 0.894, + 0.464 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparisons with Existing Methods. We visually compare our proposed technique with NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. The visual results of BakedSDF* generated in its first stage through volumetric rendering. Red and blue dotted lines emphasize subtle differences in rendering quality." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.472, + 0.175, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.518, + 0.175, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.564, + 0.175, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.632, + 0.16, + 0.639 + ], + "angle": 0, + "content": "(a) Ours w/o Opt." + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.472, + 0.271, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.518, + 0.272, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.564, + 0.272, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.632, + 0.24, + 0.638 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.472, + 0.369, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.518, + 0.369, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.564, + 0.369, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.631, + 0.354, + 0.638 + ], + "angle": 0, + "content": "(c) NeRF2Mesh" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.472, + 0.466, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.518, + 0.466, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.564, + 0.466, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.631, + 0.449, + 0.638 + ], + "angle": 0, + "content": "(d) BakedSDF" + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.641, + 0.466, + 0.674 + ], + "angle": 0, + "content": "Figure 6. Zoom-in Shading Mesh to see more geometric details. encounter challenges with far-away sky regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Ablation: In Fig. 6, we visualize the zoom-in shading mesh extracted without texture. \"Ours w/o Opt\" denotes we did not execute our optimization stage in Section 3.3 after mesh processing. The optimization stage enables us to represent the geometric details of garden vases and bicycle wheels. Our method can represent more detailed structures than BakedSDF and a smoother surface than NeRF2Mesh. Moreover, it is noteworthy that when employing rasterization-based rendering with the initial color model without the optimization stage, the rendering quality is poor, as shown in Table 1. In this table, \"Ours w/o Geo\" only trains appearance model \\( f_{color} \\) without training vertex deformation. Training vertex deformation yields a 1.7 gain in indoor scenes by refining the mesh surface. \"Ours w. RADAM\" denotes that we use the rotation-invariant" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.47, + 0.697, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.47, + 0.891, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.519, + 0.793, + 0.531 + ], + "angle": 0, + "content": "Figure 7. Naive QSlim Results" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.531, + 0.892, + 0.621 + ], + "angle": 0, + "content": "ADAM (RADAM) for updating mesh vertices instead of a naive ADAM. However, it provides benefits for preventing mesh distortion. In Fig. 7, we show the visualization results of naive QSlim [12]. The naive QSlim leads to shrinkage and degeneration during the edge collapse process, and the boundary is over-smoothed compared to our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.88, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion, Limitations, and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This paper introduces a practical approach for reconstructing geometrically accurate meshes with photorealistic rendering quality. Our method achieves a compact mesh representation in unbounded large-scale scenes while preserving highly detailed geometry. Furthermore, the appearance field corresponding to this lightweight mesh is compatible with the classical rasterization pipeline, enabling efficient rendering. However, our method has limitations. BakedSDF employs a well-designed appearance model based on spherical Gaussians. The potential for enhancing rendering quality exists by incorporating a more sophisticated appearance model. Moreover, the gradient-based mesh optimization in the final stage may still result in mesh distortions, including self-intersections. We plan to employ a carefully designed optimizer to mitigate such distortions in future work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5060" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.197 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.199, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.323 + ], + "angle": 0, + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[4] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.563 + ], + "angle": 0, + "content": "[6] Jaehoon Choi, Dongki Jung, Taejae Lee, Sangwook Kim, Youngdong Jung, Dinesh Manocha, and Donghwan Lee. Tmo: Textured mesh acquisition of objects with a mobile device by using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16674-16684, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.619 + ], + "angle": 0, + "content": "[7] Jonathan Cohen, Marc Olano, and Dinesh Manocha. Appearance-preserving simplification. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 115-122, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[8] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a vsi system for high performance graphics. Acm siggraph computer graphics, 22(4):21-30, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.677, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[9] Carl Erikson and Dinesh Manocha. Gaps: General and automatic polygonal simplification. In Proceedings of the 1999 symposium on Interactive 3D graphics, pages 79-88, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[11] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Michael Garland. Quadric-based polygonal surface simplification. Carnegie Mellon University, 1999. 2, 4, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[14] Michael Garland and Paul S Heckbert. Simplifying surfaces with color and texture using quadric error metrics. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 263-269. IEEE, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[19] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[20] Hugues Hoppe. View-dependent refinement of progressive meshes. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 189-198, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[22] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[23] Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[26] Leif Kobbelt. *Sqrt(3)-subdivision*. In *Proceedings of the 27th annual conference on Computer graphics and interactive techniques*, pages 103–112, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5061" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[28] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[29] Peter Lindstrom. Out-of-core simplification of large polygonal models. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 259-262, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[30] Peter Lindstrom and Claudio T Silva. A memory insensitive technique for large model simplification. In Proceedings Visualization, 2001. VIS'01., pages 121-550. IEEE, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[31] Selena Zihan Ling, Nicholas Sharp, and Alec Jacobson. Vectoradam for rotation equivariant geometry optimization. Advances in Neural Information Processing Systems, 35:4111-4122, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[32] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.457, + 0.469, + 0.497 + ], + "angle": 0, + "content": "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[35] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[36] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, volume 40, pages 101-113. Wiley Online Library, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[40] Andrew Nealen, Takeo Igarashi, Olga Sorkine, and Marc Alexa. Laplacian mesh optimization. In Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia, pages 381-389, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[42] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[43] Werner Palfinger. Continuous remeshing for inverse rendering. Computer Animation and Virtual Worlds, 33(5):e2101, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. In Proc. of the International Conf. on 3D Vision (3DV), 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[46] Jarek Rossignac and Paul Borrel. Multi-resolution 3d approximations for rendering complex scenes. In Modeling in computer graphics: methods and applications, pages 455-465. Springer, 1993. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[47] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[49] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[50] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[51] Chao Wang and Xiaohu Guo. Plane-based optimization of geometry and texture for rgb-d reconstruction of indoor scenes. In 2018 International Conference on 3D Vision (3DV), pages 533-541. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[52] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5062" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[55] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv, 2023. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[56] Zehao Yu, Anpei Chen, Bozidar Antic, Songyou Peng, Apratim Bhattacharyya, Michael Niemeyer, Siyu Tang, Torsten Sattler, and Andreas Geiger. Sdfstudio: A unified framework for surface reconstruction, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5063" + } + ] +] \ No newline at end of file diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_origin.pdf b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..124f8c8a06dd543f7caca6a06dc6e53da1e8d617 --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/b965bd98-b1d7-4faf-8f20-171bc6260c15_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee71ebb263b63a4f30005b747cb5a9a017bdf5d368473a4a8578512281c87443 +size 9494196 diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/full.md b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..796866cfc2a2f661f9017ebe1ca0fb48b3722f6e --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/full.md @@ -0,0 +1,363 @@ +# LTM: Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering + +Jaehoon Choi $^{1}$ Rajvi Shah $^{2}$ Qinbo Li $^{2}$ Yipeng Wang $^{2}$ Ayush Saraf $^{2}$ Changil Kim $^{2}$ Jia-Bin Huang $^{1,2}$ Dinesh Manocha $^{1}$ Suhib Alsisan $^{2}$ Johannes Kopf $^{2}$ $^{1}$ University of Maryland $^{2}$ Meta + +# Abstract + +Advancements in neural signed distance fields (SDFs) have enabled modeling 3D surface geometry from a set of 2D images of real-world scenes. Baking neural SDFs can extract explicit mesh with appearance baked into texture maps as neural features. The baked meshes still have a large memory footprint and require a powerful GPU for real-time rendering. Neural optimization of such large meshes with differentiable rendering pose significant challenges. We propose a method to produce optimized meshes for large unbounded scenes with low triangle budget and high fidelity of geometry and appearance. We achieve this by combining advancements in baking neural SDFs with classical mesh simplification techniques and proposing a joint appearance-geometry refinement step. The visual quality is comparable to or better than state-of-the-art neural meshing and baking methods with high geometric accuracy despite significant reduction in triangle count, making the produced meshes efficient for storage, transmission, and rendering on mobile hardware. We validate the effectiveness of the proposed method on large unbounded scenes from mip-NeRF 360, Tanks & Temples, and Deep Blending datasets, achieving at-par rendering quality with $73 \times$ reduced triangles and $11 \times$ reduction in memory footprint. + +# 1. Introduction + +Photorealistic reconstruction and rendering of real-world objects and scenes is a longstanding problem of importance with several applications in computer vision, robotics and AR/VR. In recent years, Neural Radiance Fields (NeRFs) [1, 2, 38, 39] have been quite successful at novel view synthesis. However, the scene representation is learned as an implicit volumetric representation that is expensive to evaluate. This is a crucial limitation for many applications, such as Augmented and Virtual Reality (AR/VR), often requiring real-time rendering with low memory and compute power. + +![](images/fd98c5886290e24ae55c9adf33cbe461f3022bfd3af6e4b2ae7a5697cdb3f88f.jpg) + +![](images/d35952d2936b517156dbff20f54382b2c4a9178dc369af030dcbb225505e9257.jpg) +Figure 1. Example results using LTM showing mesh geometry (top-left) and novel-view synthesis (right). Bottom right plot compares performance of different methods on mip-NeRF360 outdoor dataset. Our method achieves comparable rendering quality to BakedSDF while reducing the mesh size considerably. + +![](images/2f0cc1c40195cef09e1d21bbc39ceab91b3debf6fbd7501f48ff2357c658a3a9.jpg) + +![](images/9a4186137f44c69c544be406205e412211be0ff128b226383c4654ccb833d0fe.jpg) + +![](images/2112e1b85a6718c98e4b5a798ff1bbf3e253d16c73e9b8879bb350b1d22fa518.jpg) + +Since modern graphics engines are highly optimized to rasterize triangle meshes, researchers have started exploring ways to extract explicit mesh representations from volumetric functions. + +MobileNeRF [5] uses a polygon mesh with texture maps storing feature vectors and opacity. Although their rendering quality is good, the mesh quality is far from an ideal surface. While NeRF2Mesh [50] and NeRFMeshing [44], based on volumetric density fields, excel at generating highly detailed geometry, they often result in noisy and bumpy surfaces and numerous floaters, especially in planar regions, which are common in large-scale scenes. In contrast, neural Signed Distance Function (SDF) methods [42, 52, 54] typically excel in reconstructing accurate surface geometry. However, they may not be able to capture small or detailed geometry, and their rendering quality is inferior to volumetric density-field methods. BakedSDF [55] overcomes these limitations by applying mip-NeRF 360 techniques to train SDF representation, though the resulting mesh representation has a high memory overhead. + +Previous neural SDF methods that rely on the Marching Cube algorithm [34] result in meshes with high storage overhead. This results in a significant challenge for practi + +cal applications on mobile devices, where low disk storage and fast data transfer are essential. Instead of relying on the Marching Cubes, some methods [11,28,48,49] propose differentiable mesh reconstruction methods. These methods aim to extract an isosurface from an implicit function and can be seamlessly integrated with differentiable rendering approaches [16, 17]. However, most methods primarily concentrate on a single object with ideal implicit surface functions, unsuited for large unbounded real-world scenes [2, 55]. Reconstructing an ideal mesh for the unbounded large-scale scene is still an open problem, often resulting in non-manifold geometry or missing surfaces. Consequently, the renderings can result in visual artifacts due to the incompleteness of the mesh geometry. + +Main Results: In this paper, our goal is to achieve a realistic rendering with a compact mesh representation that optimally utilizes disk storage for unbounded large-scale scenes, all while preserving geometric quality. Our method initially trains neural fields for geometry and appearance separately. For geometry reconstruction, we adopt the BakedSDF approach [55] to train an SDF in a contracted space [2]. Regarding the appearance field, we utilize hash-based encoding [39] with two shallow MLPs to represent diffuse and specular colors. The SDF and appearance fields are compact representations for reconstructing a high-quality mesh and achieving realistic rendering. + +Given the initial geometry representations, we perform mesh extraction and decimation. As unbounded large-scale scenes occupy a large amount of disk storage, our method focuses on generating a lightweight mesh without compromising geometric details. We adaptively extract the mesh for the central and background regions in the contracted space, respectively. Subsequently, we modify a mesh decimation [14] designed for large-scale scenes. Our method eliminates redundant triangles corresponding to simple geometric features, such as planar surfaces, while preserving complex geometric features. + +Since the appearance field is trained by volumetric rendering, the initial model is unsuitable for rasterization-based rendering. Moreover, the mesh from the previous step still fails to accurately represent areas with highly-detailed geometry (thin structures) or semi-transparent objects. These mesh defects hurt rendering quality. Thus, we use differentiable rendering [27] and utilize image-space supervision to refine the mesh surface by rectifying these mesh defects while preserving geometric accuracy. Due to the highly simplified mesh and various regularization techniques, our approach can deform the mesh topology without causing severe mesh distortion, a common issue in gradient-based mesh optimization [41, 43]. Compared to mesh-based neural rendering methods [5, 44, 50, 55], our approach results in lower storage overhead while providing comparable rendering quality. To sum up, our contributions include: + +- We effectively utilize classical mesh decimation techniques to extract compact geometry from neural SDFs of large unbounded scenes. +- We introduce a new approach for joint optimization of appearance and vertex deformation that preserves geometric details and improves rendering quality. +- We achieve a trifecta of precise geometry, efficient storage, and rendering quality comparable to the state of the art on diverse captured environments such as mip-NeRF 360, Tanks & Temples, and the Deep Blending dataset. + +# 2. Related Work + +Neural Rendering Neural Fields achieve remarkable photo-realistic rendering quality. One of the most successful approaches is Neural Radiance Fields (NeRFs) [38], which adopts volumetric implicit representations and trains a radiance field through ray marching. Mip-NeRF [1, 2] presents casting a conical frustum into the scene space and extends this integrated positional encoding to unbounded scenes by using a novel contraction function. Instead of learning volumetric density, neural surface reconstruction methods [6, 42, 52, 54] propose learning a Signed Distance Function (SDF) for surface reconstruction via volumetric rendering. Furthermore, several methods have been proposed to accelerate NeRFs by utilizing efficient data structure [3, 10, 39], although achieving real-time rendering for high-resolutions remains a challenge. Alternatively, to enhance rendering speed, "baking" based methods [5, 19, 44, 45, 50, 55] typically precompute and store a trained NeRF into a more efficient representation. SNeRG [19] adopts a deferred shading model [8] and bakes trained NeRFs into sparse voxel grid representations with neural features. In particular, the mesh is widely used as geometry representation and bakes neural fields into its texture. MobileNeRF [5] builds on a mesh rasterization pipeline for efficient rendering on mobile architecture. NeRF2Mesh [50] introduces a two-step framework for NeRF training and textured mesh extraction. NeRFMeshing [44] proposes a novel method to produce a truncated SDF from NeRFs and extracts the accurate mesh. BakedSDF [55] enables us to reconstruct high-quality mesh and bake it into a small appearance model based on spherical Gaussians for fast rendering. DNMP [35] and NeuRaS [32] adopt mesh-based geometry and rasterization-based rendering for large driving scenes. + +Mesh Processing Traditional graphics research has developed effective techniques for simplifying meshes, based on vertex decimation [46, 47], edge collapse [9, 12-14, 20] and appearance preservation [7]. Vertex decimation methods prioritize vertices based on shape heuristics and remove the least important ones. Edge collapse methods rank the edges based on the cost and subsequently contract pairs of vertices. QSlim [12, 13] introduces the concept of the + +Quadric Error Metric (QEM), which is defined as the sum of quadrics representing the distance of a point from a set of planes. Following the edge collapse, they utilize this QEM to identify which edges should be removed while preserving geometric topology. Lindstrom et al. [29, 30] use vertex clustering [46] on a uniform grid and accumulate a quadric error matrix to the occupied grid cell. Wang et al. [51] employ RGB-D sequences to reconstruct a lightweight mesh. + +Differentiable Rendering and Mesh Optimization Recently, many differentiable rendering techniques [23] have emerged, aimed at facilitating end-to-end optimization of the rendering process by employing useful gradients. In particular, several works [4, 33] focus on approximating the rasterization stage of the rendering to compute meaningful gradients. Furthermore, modular differentiable renderers [21, 22, 27] exist, allowing users to customize and modify specific functions. Differentiable mesh reconstruction techniques [11, 28, 48] can build a mesh through gradient-based optimization. In particular, differentiable rendering [16, 17, 49] can offer 2D supervision to refine and enhance the geometric quality of the mesh. Some works [16, 17] adopt an inverse rendering approach, leveraging physics-based rendering to synthesize shapes. Nevertheless, gradient-based mesh optimization [41] is highly sensitive to initialization and prone to significant mesh distortion. Some methods [31, 41, 43] incorporate a novel ADAM optimizer and employ strong regularization techniques. However, this research is currently limited to generating single objects and is challenging to apply to large-scale scenes. + +# 3. Our Method + +Given multi-view images and corresponding camera poses, we aim to reconstruct the texture and geometry of unbounded large-scale scenes. Fig. 2 shows the three main stages of our pipeline: Initialization, Mesh Processing, and Optimization. In the first stage, we employ volumetric rendering to train neural implicit representation for geometry $f_{sdf}$ and appearance $f_{color}$ . Subsequently, we leverage the geometry representation $f_{sdf}$ to extract a high-resolution mesh via Marching Cubes [34] and reduce the number of triangles without causing geometry distortion. Next, we utilize differentiable rendering [27] to jointly optimize the explicit 3D mesh $M$ and implicit appearance field $f_{color}$ obtained from the first stage and refine the mesh surface. + +# 3.1. Initialization of Geometry and Appearance + +In this stage, we first learn neural implicit representations for both appearance and geometry. A neural radiance field [38] is an implicit representation of a 3D scene volume, a mapping from a 3D position $x$ and a ray viewing direction $d \in \mathbb{R}^3$ to an RGB color $c$ and volume density $\rho$ . In practice, the networks modeled by MLPs are used to generate this + +![](images/f0e13cda3153589f6e86fb7db0a0c2f728b647df793244f2e55086eadcaf7c75.jpg) +Figure 2. Overview of Our Pipeline. It consists of three steps: Initialization (Sec. 3.1), Mesh Processing (Sec. 3.2), and Optimization (Sec. 3.3). Our method focuses on baking the initial appearance and geometry representation, trained by volumetric rendering, into an explicit mesh and a memory-efficient texture conducive to real-time rendering. We employ classical mesh decimation to minimize storage requirements, introduce the optimization stage to facilitate rasterization-based rendering and refine the detailed mesh structure. + +volume density and view-dependent color: + +$$ +\rho = f _ {\text {d e n s i t y}} (x), \quad c = f _ {\text {c o l o r}} (x, d). \tag {1} +$$ + +To render the color of a pixel, we first compute the 3D points $t_i$ along the viewing ray associated with each pixel $p$ as $x(t) = o + td$ . Then, estimated volume densities $\rho_i$ and colors $c_i$ are utilized to approximate a volume rendering equation [37] using numerical quadrature: + +$$ +C = \sum_ {i} T _ {i} \alpha_ {i} c _ {i}, \quad T _ {i} = \exp \left(- \sum_ {j < i} \rho_ {i} \delta_ {i}\right), \tag {2} +$$ + +where $T_{i}$ is the accumulated transmittance and $\alpha_{i} = 1 - \exp(-\rho_{i}\delta_{i})$ is the discrete opacity of the point. $\delta_{i} = t_{i} - t_{i-1}$ is the distance between two neighboring points along the ray. To deal with large unbounded scenes, mip-NeRF 360 [2] introduces a proposal MLP for efficient sampling and space contraction to represent an unbounded scene. The scene contraction function $f$ (in Eq. 3) maps the unbounded points into bounded space: + +$$ +f (x) = \left\{ \begin{array}{l l} x & | | x | | \leq 1, \\ \left(2 - \frac {1}{| | x | |}\right) \frac {x}{| | x | |} & | | x | | \geq 1. \end{array} \right. \tag {3} +$$ + +Appearance Similar to previous works [5, 19, 50], we decompose an appearance field $f_{color}$ into shallow MLPs to learn diffuse colors $c_d$ and specular colors $c_s$ separately. Particularly, the view-dependent effects are modeled by a small MLP that can be easily integrated into a fragment Shader. In order to ensure the MLPs are able to model an unbounded scene well, we adopt strategies similar to the ones introduced in mip-NeRF360 work, specifically, scene contraction, proposal MLP, and regularization for the background. We train both MLP networks $f_{color}$ and $f_{density}$ with a multi-resolution hash positional encoding [39] using the volume rendering objective in Eq. 2. The appearance field trained in this stage is further optimized to facilitate rasterization-based rendering, replacing the computationally expensive ray marching. This optimization is described later in Section 3.3. + +![](images/a94c3532d95a8f705e7b0348191d136b71eb0df50b5b675286e07fa10dcb62a7.jpg) +Figure 3. Our Approach: In Section 3.1, we start training both the appearance $f_{color}$ and geometry $f_{sdf}$ representations for a large-scale scene using volumetric rendering (shown in orange). Then, in Section 3.2, our method extracts the mesh from $f_{sdf}$ and significantly simplifies its structure, resulting in the lightweight mesh $M$ (shown in blue). In Section 3.3, we jointly train vertex deformation $\Delta x$ and $f_{color}$ computed from appearance modeling. + +Geometry Initial geometry modeling in our method is similar to BakedSDF [55]. The core mechanism for neural surface reconstruction was introduced in VolSDF [54] for small objects. BakedSDF shows that by leveraging the efficient sampling and contraction ideas introduced in mipNeRF360 [2], VolSDF [54] can be extended to large unbounded scenes. We follow this and jointly train two MLPs using the volumetric rendering objective, (i) $f_{sdf}$ - representing geometry of a scene as a signed distance function, and (ii) $g_{color}$ - a color prediction network, as shown in the Geometry Reconstruction block in Fig.3. + +Following VolSDF, we define the volume density $\sigma(x)$ as a function of surface geometry $f_{sdf}$ , namely, $\sigma(x) = \alpha \Psi_{\beta}(f_{sdf}(x))$ , where $\Psi_{\beta}$ is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and $\beta$ is the learnable scaling parameter. The Eikonal loss [15] to regularize the $f_{sdf}$ is defined as, + +$$ +L _ {e i k} = \left(\| \nabla f _ {s d f} (x) \| _ {2} - 1\right) ^ {2}. \tag {4} +$$ + +The color MLP $g_{color}$ , in addition to the position and viewing direction, also takes the predicted surface normal and a 256-dimensional geometric feature as inputs. It is important to note that the sole purpose of color MLP $g_{color}$ is geometry initialization and we discard it after this step. This is not to be confused with the shallow decomposed color MLP $f_{color}$ trained specifically to model appearance that is also jointly refined with mesh vertices in the later stage. + +# 3.2. Mesh Extraction and Decimation + +In this stage, we first extract a mesh by using Marching Cubes (MC) [34] algorithm on learned SDF ( $f_{sdf}$ evaluated over a uniform grid) and simplify the extracted mesh using a geometry-aware quadric error based mesh decimation. While this approach works pretty well for extracting meshes of well-captured small-scale objects and scenes, it presents unique challenges for large-scale scenes. Meshing neural SDFs of large-scale scenes with marching cubes can often result in spurious surfaces arising from (i) insufficient grid resolution to model complex geometry of nearby + +objects, and (ii) poorly learned geometry of distant objects due to insufficient view-points observing such distant regions. Furthermore, since our SDF $f_{sdf}$ is trained in the contracted space, the resulting mesh is also defined in the contracted space. When we uncontract, i.e. warp the mesh vertices back into the original Euclidean space, it further exacerbates the geometry distortion in the background region. + +We aim to preserve the detailed geometry in central (foreground) regions while achieving a coarser but less noisy reconstruction of distant (background) regions, within a small triangle budget. We achieve this by treating foreground and background regions distinctly during both extraction and simplification steps. During MC extraction, we set a finer grid resolution of 2048 and a spatial range of $[-1,1]^3$ for the central region, and a coarser resolution of 1024 and a spatial range of $[-2,2]^3$ for the background region. For decimation, we first partition the central and background regions and employ an aggressive mesh decimation in background regions. This helps us transform bumpy and noisy background areas into smooth surfaces while also reducing redundant triangles. + +We first apply vertex clustering and then run quadric error based simplification algorithm - QSlim [12, 13] for all clusters. Then, we run edge contraction for each cluster to reach a threshold and repeat this process until convergence. The QSlim algorithm calculates a quadric metric for each vertex to decide whether the two vertices should be merged and aggregates these quadrics during each edge collapse operation. The basic quadric error in QSlim [13] is defined as the point-to-plane distance from vertex $v \in \mathbb{R}^3$ to all planes $p \in \mathbb{R}^4$ of its adjacent faces $N(v)$ : + +$$ +E (v) = \sum_ {p \in N (v)} \left(p ^ {T} v\right) ^ {2} = v ^ {T} \left(\sum_ {p \in N (v)} p p ^ {T}\right) v = v ^ {T} Q _ {v} v. \tag {5} +$$ + +The quadric $Q_{v}$ is the symmetric 4x4 matrix, and each edge computes this quadric by adding the quadrics of the corresponding two vertices. The most commonly used quadric error is defined solely based on 3D vertex positions $v$ . + +Intuitively, the most effective strategy is to represent planar surfaces with a smaller number of large triangles while keeping a consistent number of triangles for complex shapes. Variants that incorporate additional surface attributes, such as normals or textures in calculating the quadric error are proposed in [12]. We experimentally observed that compared to the basic QSlim algorithm, using both the vertex positions $v \in \mathbb{R}^3$ and the corresponding vertex normal $n \in \mathbb{R}^3$ for quadric error is very effective at prioritizing triangle decimation on flat surfaces while retaining them on complex surfaces. The combination of vertex clustering and geometry-aware QSlim helps us in effectively reducing the number of triangles while preserving the original geometry. + +# 3.3. Appearance and Geometry Optimization + +In this stage, we optimize the appearance model jointly with mesh vertex deformation and refine the mesh topology using a geometric complexity based criterion. + +Joint optimization of appearance and vertex deformation Given the simplified mesh $M$ and the initial appearance field $f_{color}$ , we employ the fast differentiable rasterizer introduced in Nvdiffrast [27] to jointly optimize the mesh geometry and the appearance field. The initial appearance field $f_{color}$ trained using the volume rendering objective (Eq. 1) is agnostic of the surface geometry of the extracted and simplified mesh. To finetune the model to fit the mesh geometry, we rasterize the mesh into each of the training camera poses and obtain surface points $s$ and face normals $N$ interpolated per pixel. Then we apply the contraction function in Eq. 3 to warp the surface points. This is because the appearance field defined in contracted space proves to be more effective in enhancing the rendering quality [2]. The appearance field $f_{color}$ takes these surface points $s$ and viewing direction $d$ as input, rendering an image $\hat{I}_i$ . Initially, rendering performance is degraded because the appearance field only utilizes the intersection point between the ray and the surface rather than accumulating all sampled points along the ray. We finetune this appearance field to be compatible with mesh rendering, akin to the functionality of a fragment Shader. + +Given the positions $x$ of the mesh vertices, we train per-vertex deformation $\Delta x$ using differentiable rendering [27]. We explore two approaches to optimize vertex deformation: a standard ADAM optimizer [24] and a rotation-invariant ADAM optimizer [43]. The ADAM optimizer is widely employed [16, 50] for deforming mesh topology to enhance rendering performance. While the ADAM optimizer often results in mesh distortions such as flipped triangles and self-intersections, our method mitigates mesh distortion, especially on smooth surfaces, due to the rigorous mesh decimation performed in Section 3.2. + +Objective Function To jointly train appearance field and vertex deformation, we minimize the image-space L1 loss $L_{\text{appearance}}$ between the rendered image $\hat{I}_i$ of the deformed surface and the groundtruth image $I_i$ for a camera view $P_i$ : + +$$ +L _ {\text {a p p e a r a n c e}} = \sum_ {i} \left| \hat {I} _ {i} - I _ {i} \right|. \tag {6} +$$ + +Naively training vertex deformation is highly likely to result in undesirable mesh topology characterized by a disorganized arrangement of triangles, often referred to as a "polygon soup." To prevent this mesh distortion, we employ various regularization methods. We use the Laplacian mesh smoothness loss and the normal consistency loss inspired by [36]. + +$$ +L _ {l a p} = \sum \| L V \| _ {2} \tag {7} +$$ + +where $V \in \mathbb{R}^{N \times 3}$ ( $N$ is the number of vertices) is the matrix of deformed vertex coordinates, indicating $x + \Delta x$ , at each training step, and $L \in \mathbb{R}^{N \times N}$ is the graph Laplacian of the mesh [40]. Minimizing the magnitude of a vertex's differential coordinates ( $LV$ ) reduces its distance from its neighbors' average position. The normal consistency loss is + +$$ +L _ {n c} = \sum_ {i, j \in F} \left(1 - n _ {i} \cdot n _ {j}\right) ^ {2}, \tag {8} +$$ + +where $n_i$ and $n_j$ denote the surface normal of two adjacent faces and $F$ is the set of faces sharing a common edge. Additionally, in Section 3.2, a high-resolution mesh is rasterized in the first pass before decimation. The output of this rasterization is used to interpolate face normals per pixel, resulting in a normal map $\hat{N}_i$ for every viewpoint. We calculate the L1 loss between a rendered normal $N_i$ from current mesh $M$ and the $\hat{N}_i$ : + +$$ +L _ {\text {n o r m}} = \sum_ {i} | \hat {N} _ {i} - N _ {i} |. \tag {9} +$$ + +$L_{norm}$ provides geometry supervision, particularly in planar regions, helping prevent severe mesh distortion when moving vertices. The total loss function $L$ is defined as: + +$$ +L = L _ {\text {c o l o r}} + \lambda_ {\text {l a p}} L _ {\text {l a p}} + \lambda_ {\text {n c}} L _ {\text {n c}} + \lambda_ {\text {n o r m}} L _ {\text {n o r m}}, \tag {10} +$$ + +where $\lambda_{*}$ is the weight to balance between the loss terms. + +Mesh Refinement Inspired by NeRF2Mesh [50], we adopt a classical mesh refinement approach based on mesh subdivision and decimation during training. The mesh refinement strategy is applied differently to the central and background regions. Instead of rendering error proposed by [50], which can arise from many factors including specularity, we employ a criterion: the vertex deformation values for subdividing triangle faces. The large vertex deformation implies a more substantial triangle budget to accurately capture the intricacies of highly detailed geometry associated + +![](images/14527b1c3e0e24795071a1d22597decaa055bf613881ba4a5a53b2c1e2b43560.jpg) +Figure 4. Visual comparison of Our Method, NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. Our method and NeRF2Mesh show shading mesh and wireframe mesh extracted without texture, respectively. + +![](images/1a1e8a9d124af83baaa9a7f230bbfd54f09a1dc282b3ba78fdbd72f736b07eff.jpg) + +![](images/0e0d789746e2598603ed6a323aff4a2c18b8a135ce7eb345d64e6f2ae7605ba9.jpg) + +![](images/c06279adae33f8bd1274b8a3ed84261d91db0b072cd0ba778d39907affce1112.jpg) + +with those vertices. We sort the absolute values of vertex deformation and determine a $10\%$ of faces to increase triangles. Then, we apply $\sqrt{3}$ subdivision [26], which adds a vertex at the barycenter of a triangle and connects it to all vertices of this triangle. This scheme enables us to augment the triangle budget for complex geometry and mitigate self-intersection during vertex deformation. Furthermore, the mesh decimation presented in Section 3.2 is applied after every mesh subdivision to remove redundant triangles on simple and flat geometry. For background regions, we examine all training views and check faces intersecting with ray casting. We utilize dilation for all invisible masks and employ mesh decimation to remove redundant faces. + +Rotation-Invariant ADAM The rotation-invariant ADAM formulation [43] modifies the update rule in ADAM, resulting in the update step: $\Delta x = \zeta \cdot v \cdot l_{ref}$ . Here, $\zeta$ is the learning rate, and $v$ is the standard ADAM update step, which is the loss gradient with respect to $x$ . $l_{ref}$ is the coefficient initially set to the edge length incident to $x$ , and it gets updated based on the average $v$ . + +# 4. Experiments + +# 4.1. Experiment Settings + +Datasets: To validate the effectiveness of our method, we mainly conduct experiments on seven unbounded, large-scale $360^{\circ}$ indoor and outdoor scenes from the mip-NeRF 360 [2] dataset. Also, we validate the generalization ability of our approach on two different datasets, including large-scale scenes: two scenes from the Tanks & Temples [25] dataset and two scenes from the Deep Blending [18] dataset. Implementation Details: For all datasets, we subsampled every eight frames for test sets following the mip-NeRF 360 paper [2]. For geometry reconstruction in Section 3.1, we + +use SDFstudio [56] to implement BakedSDF. In Section 3.2, we reduce $95\%$ of the triangles in the central region and $99\%$ of the triangles in the background regions through decimation. In Section 3.3, we run all experiments for 40k iterations, and a learning rate is annealed from 1e-2 to 1e-3. We set the loss weights to $\lambda_{lap} = 1\mathrm{e - }3$ , $\lambda_{nc} = 1\mathrm{e - }3$ , and $\lambda_{norm} = 1\mathrm{e - }4$ . We apply mesh refinement at 8k, 16k, and 24k training steps. Subsequently, we compute diffuse color and specular features using $f_{color}$ to export textures. All these color maps and features are 3-channel .png files. Additional details are provided in the supplementary materials. + +# 4.2. Experimental Results + +We compare our proposed method with three different methods: 1) MobileNeRF [5] is the state-of-the-art method, enabling NeRF to achieve real-time rendering of explicit 3D mesh and neural features. 2) NeRF2Mesh [50] employs a volumetric rendering to reconstruct the coarse mesh and refine the mesh by jointly optimizing appearance and geometry. 3) BakedSDF [55] is a recent method for optimizing a neural surface-volume representation for real-time rendering of large unbounded scenes. We also add NeRFMeshing [44], which distills the rendering NeRF into a neural field that represents TSDF (Truncated SDF) and reconstructs the 3D mesh. As BakedSDF and NeRFMeshing have not released public code, we adopt the metrics provided in their original papers. BakedSDF has made its GLB file available. We utilize this file for visualization. + +Figure 4 qualitatively compares our method with NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. We visualize the shading meshes extracted without texture, which is the most effective way to visualize geometric differences. Furthermore, we include a wireframe mesh, distinguishing between our method and NeRF2Mesh. Visu + +
MethodMeshOutdoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)Indoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)
BicycleGardenStumpMeanRoomCounterKitchenBonsaiMean
Instant-NGP [39]X22.1 / 0.49 / 0.4924.5 / 0.65 / 0.3123.6 / 0.57 / 0.4523.4 / 0.57 / 0.4229.2 / 0.85 / 0.3026.43 / 0.80 / 0.3428.5 / 0.82 / 0.2530.3 / 0.89 / 0.2328.6 / 0.84 / 0.28
mip-NeRF 360 [2]X24.4 / 0.68 / 0.3026.9 / 0.81 / 0.1726.4 / 0.74 / 0.2625.9 / 0.75 / 0.2431.6 / 0.91 / 0.2129.5 / 0.89 / 0.2032.3 / 0.92 / 0.1333.5 / 0.94 / 0.1831.7 / 0.92 / 0.18
MobileNeRF [5]O21.7 / 0.43 / 0.5118.8 / 0.59 / 0.3623.9 / 0.56 / 0.4321.5 / 0.53 / 0.4328.9 / 0.85 / 0.2825.1 / 0.72 / 0.2926.8 / 0.79 / 0.7923.8 / 0.71 / 0.7226.2 / 0.77 / 0.52
NeRF2Mesh [50]O22.1 / 0.48 / 0.5123.4 / 0.55 / 0.4022.5 / 0.54 / 0.4622.7 / 0.52 / 0.4625.7 / 0.79 / 0.3523.9 / 0.71 / 0.3524.0 / 0.61 / 0.3625.0 / 0.77 / 0.2924.7 / 0.72 / 0.34
NeRFMeshing [44]O21.1 / - / -22.9 / - / -22.6 / - / -22.2 / - / -26.1 / - / -20.0 / - / -23.6 / - / -25.6 / - / -23.8 / - / -
BakedSDF [55]O22.0 / 0.57 / 0.3724.9 / 0.75 / 0.2123.6 / 0.59 / 0.3723.5 / 0.64 / 0.3228.7 / 0.87 / 0.2525.7 / 0.81 / 0.2826.7 / 0.82 / 0.2427.2 / 0.85 / 0.2627.0 / 0.84 / 0.26
OursO22.4 / 0.52 / 0.4423.5 / 0.64 / 0.3023.7 / 0.57 / 0.4223.2 / 0.58 / 0.3929.3 / 0.88 / 0.2325.1 / 0.77 / 0.2726.5 / 0.80 / 0.2127.3 / 0.84 / 0.2227.1 / 0.82 / 0.23
Ours w/o OptO16.1 / 0.30 / 0.5318.8 / 0.45 / 0.3919.4 / 0.39 / 0.5018.1 / 0.38 / 0.4817.2 / 0.54 / 0.4616.2 / 0.47 / 0.4718.1 / 0.50 / 0.3616.4 / 0.54 / 0.4317.0 / 0.51 / 0.43
Ours w/ GOO21.1 / 0.48 / 0.4723.2 / 0.65 / 0.3023.0 / 0.53 / 0.4622.5 / 0.55 / 0.4127.5 / 0.85 / 0.2623.6 / 0.72 / 0.3325.5 / 0.78 / 0.2325.9 / 0.76 / 0.325.4 / 0.78 / 0.28
Ours w. RADAMO21.6 / 0.48 / 0.4623.4 / 0.66 / 0.3023.5 / 0.56 / 0.4522.8 / 0.57 / 0.4027.5 / 0.85 / 0.2624.1 / 0.74 / 0.3125.8 / 0.80 / 0.2226.8 / 0.80 / 0.2626.0 / 0.80 / 0.26
+ +Table 1. Quantitative Comparison on mip-NeRF 360 Dataset. “-” denotes a missing implementation. The best and second best-performing algorithms for each metric are bolded and underlined. + +
MethodBicycleGardenStumpRoomCounterKitchenBonsaiMean
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]2011435 / 6638135723 / 3401641069 / 497220903 / 421170864 / 39995552 / 25684606 / 284153879 / 1262
NeRF2Mesh [50]129815 / 858117650 / 779123415 / 82069328 / 46279517 / 38489601 / 43470319 / 63697387 / 667
BakedSDF [55]1087111248 / 4438451310773M / 2107294319667 / 384673857227 / 1411058512212 / 2389265913832 / 2705664113629 / 266796871564M / 27951
Ours83270 / 52950113 / 218120425 / 84745124 / 23858138 / 2684070 / 13642233 / 45962196 / 385
+ +alizing the wireframe mesh from BakedSDF lacks meaningful interpretation due to its extremely high face density. MobileNeRF generates a "triangle soup" that inaccurately represents the scene's geometry. Also, compared to the mesh reconstructed by NeRF2Mesh, our method shows sharp details of complex geometry and smooth surfaces on the ground or wall. Moreover, in the wireframe mesh, our method employs more triangles in intricate shapes like the bicycle or the garden vase while allocating smaller triangles to planar regions or simpler structures. While the geometry of BakedSDF exhibits superior quality compared to our mesh reconstruction, achieving this high-quality mesh necessitates approximately $73 \times$ more triangles in Table 2. + +Our method effectively achieves the optimal tradeoff between rendering quality and disk storage cost. For comparison, we report two tables for rendering quality and disk storage. Table 1 reports the rendering quality using the standard PSNR, SSIM [53], and LPIPS [57] on mip-NeRF 360 dataset. Instant-NGP [39], and mip-NeRF 360 [2] are volumetric rendering approaches that do not reconstruct explicit geometry, a topic that lies outside the primary focus of our study. In Table 2, we report the disk storage in megabytes and the number of vertices and faces. For a fair comparison, all the mesh files are saved in .obj and .mtl formats. Texture images are stored as .png files. Other assets, such as view-dependency MLP, are saved as .json files, except BakedSDF, which only provides gLTF format files. Our method demonstrates comparable rendering quality in outdoor scenes (with a 0.3 PSNR drop) and indoor scenes (with a 0.1 PSNR increase) while saving $73 \times$ triangles and $11 \times$ storage compared to the state-of-the-art BakedSDF. In comparison to NeRF2Mesh and MobileNeRF, our method shows better-rendering quality (Table 1), demands less disk storage, and fewer triangles (Table 2), and captures superior mesh quality (Fig. 4). In Table 5, we measure the rendering speed and GPU memory by running in-browser on a Mac + +Table 2. Disk Storage (MB), Number of Vertices (V) and Faces (F). The unit for the number of vertices and faces is ${10}^{3}$ . "M" denotes the ${10}^{6}$ unit. We save ${73} \times$ the number of triangles and reduce ${11} \times$ disk storage. + +
MethodMeshDeep Blending [18]Tanks & Temples [25]
DrJohnsonPlayroomBarnCourthouse
Instant-NGP [39]X27.7 / 0.84 / 0.3819.5 / 0.78 / 0.4623.4 / 0.72 / 0.4320.0 / 0.68 / 0.53
mip-NeRF360 [2]X29.1 / 0.90 / 0.2429.6 / 0.90 / 0.2527.2 / 0.81 / 0.3221.3 / 0.72 / 0.47
MobileNeRF [5]O25.9 / 0.78 / 0.3128.7 / 0.85 / 0.2824.4 / 0.77 / 0.2417.7 / 0.54 / 0.36
NeRF2Mesh [50]O24.0 / 0.75 / 0.5126.6 / 0.83 / 0.4123.4 / 0.72 / 0.4318.2 / 0.63 / 0.56
OursO26.5 / 0.82 / 0.4029.1 / 0.87 / 0.3226.5 / 0.83 / 0.2420.2 / 0.73 / 0.34
+ +Table 3. Quantitative Comparison. We report PSNR $\uparrow$ , SSIM $\uparrow$ , and LPIPS $\downarrow$ on Deep Blending and Tanks & Temples datasets. + +
MethodDrJohnsonPlayroomBarnCourthouse
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]174.9602 / 285281.81097 / 521289.21080 / 502263.6934 / 417
NeRF2Mesh [50]108.6414 / 815118.7401 / 793131.9397 / 793145.7451 / 843
Ours26.746 / 9260.9203 / 4014086 / 17147.497 / 192
+ +Table 4. Disk Storage (MB), Vertices (V) and Faces (F). + +
FPS / GPU (MB)MobileNeRFNeRF2MeshOurs
MacBookPro (1920x1080 Resolution)59 / 324205 / 130212 / 115
+ +Table 5. FPS and GPU memory on mip-NeRF 360 dataset + +BookPro (2020, M1). + +Figure 5 compares our method to the state-of-the-art baselines that focus on rendering quality: NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. In their first stage, BakedSDF* represents the volumetric rendering results obtained from a surface-based volumetric representation. The visual results of our approach exhibit sharper details than other methods. In the Bicycle and Room scenes, the mesh surface of the armrest for a bench and tree leaves exhibits noticeable jiggling, resulting in distorted images. Unlike BakedSDF*, our method effectively manages the thin structure of the wheel. + +Experimental Results on Deep Blending and Tank & Temples Datasets: We train and evaluate our method on four scenes: Playroom and DrJohnson from the Deep Blending dataset, and Barn and Courthouse from the Tanks & Temples dataset. The former includes bounded indoor scenes, while the latter comprises large unbounded outdoor scenes. Their capture methodology differs from that of the mip-NeRF 360 dataset, and they do not utilize 360-degree scenes. In Tanks & Temples, we exclude sky regions during training and evaluation, as mesh-based rendering tends to + +![](images/68e53eb087640f075447d91d36fa3ffe44c4f15000be93b56aa2d8c9a6702af7.jpg) + +![](images/c37e774ade1e6567bc4c09d3d4ba07c869e3ece3fbde836d99caf0eb6cff20d7.jpg) + +![](images/5d4394e26a7d4db72ef6404a38ce3006407e2bd58f69f92fed061946b0922e8b.jpg) + +![](images/6ec0cbd3339e840ef314752c05fd88c7e50637f40fa36f7f8e8347fa1dd6d9bf.jpg) +Scene + +![](images/c57afcff8883b762d5559ca7da593a550813c2bbf9aec92c4b9f02afba764775.jpg) + +![](images/e2220ff23b0d996e39d3cd1c7aa77283b0724e2faa2b51024de198f972a6d7ca.jpg) + +![](images/a85d1b7f3df4a4a87659e07c5b000bb4de585d65159b3f53ff8214d2d6bafaa5.jpg) + +![](images/45e7086bb9bb53aa17e139cdf10ae2c7ea4c4f5e85184c1c25544b1c96b834c6.jpg) +(a) Ours + +![](images/c9a23a8610c7b0be0ae0756d74821d9204bdd461bc754843a2ebf8da15dccfd1.jpg) + +![](images/348dd17a52fee2f3d2d563e6f5fab39e675bd93efe0139e7a2aba90d1357aaf9.jpg) + +![](images/4196afc8ff9354ad2d5d27b67851277f28f7446b08aa7e2b171620aa8ee74e27.jpg) + +![](images/d99d8a57c37905c7efe16961007c2322468200aa98a8834c931362db66460316.jpg) +(b) NeRF2Mesh + +![](images/00673299fbc26f822afc3ae151745c998281b7f9b02d583d82238d9bffddcbeb.jpg) + +![](images/a866ad403d194ae5594412e83921d3960e82268ccd6c5c37bc5c7fa52cd6b186.jpg) + +![](images/9f698c865df6068129f7b1696be7df410383f488c0e220424b6691895edcb6f0.jpg) +(c)BakedSDF\* + +![](images/6b38533c70d678c3ec03b6be9c1c29a81861d429f4d2376d33256449331c4ddc.jpg) + +![](images/2ab2eb16d18c82c0847b7795b7d411707f8917be93c33303c06933fd50d3e297.jpg) + +![](images/ab99f518d41a6e923006684fd031a7e7c87e3c964524db63795418d03e0652b0.jpg) + +![](images/86534fee9dbce4b104e47c63b6b3bd7de8024f7aa5966bd39095ba914e4f739a.jpg) +(d) MobileNeRF + +![](images/107430fdf18efb43a11e33f7955d6f1d6e636308269803ff535001c3b91204ce.jpg) + +![](images/6cc9d40c8ee9823b8491755349163a077e2f7a87b31894fe927abd6167af91b1.jpg) + +![](images/3c67fafc6044e4b38434f0d460526a49fa29f4be110fe71f22e5544a63ad36cd.jpg) +(a) Ours w/o Opt. +Figure 6. Zoom-in Shading Mesh to see more geometric details. encounter challenges with far-away sky regions. + +![](images/c292a89b0497768ccea656f80a743fe2d11a23280f63b2064d1b2eb7c6c75af4.jpg) + +![](images/c7d8dd9d1c052a2468a72c587665454adacf0f6db51b06f7ac74544f0bc77e3f.jpg) + +![](images/c3b2051cacf858e78dae5891bfd83d9a6cb7eb77b772d8b9dcc3d392596225cd.jpg) +(b) Ours + +![](images/f6cb5bcb7d3fc2c017ec92a8c7f8e833a27147eb8e945bb6c47e6424b636ff28.jpg) + +![](images/4553bd33dc7fb5004c38a0c012cecb47d36eb90ffcb1e90b57f26639e9ce0eea.jpg) + +![](images/893f2fddaaca3a5105cc1ff06217f8d8fbfd1c53a8d9389b44fef24b4ddbc831.jpg) +(c) NeRF2Mesh + +![](images/c871f7b71b6899b394c1467c9103941aba05c65f4d46b604f240b35c33cf8595.jpg) + +![](images/f465ad3f474a6dbaa440e66b4f8bd8fda04f9d233c3bb32a124bea6509e6215f.jpg) + +![](images/84f2f4a507bcf6e226531e71edf87725319108c1d885124f6636fa7087602421.jpg) +(d) BakedSDF + +Ablation: In Fig. 6, we visualize the zoom-in shading mesh extracted without texture. "Ours w/o Opt" denotes we did not execute our optimization stage in Section 3.3 after mesh processing. The optimization stage enables us to represent the geometric details of garden vases and bicycle wheels. Our method can represent more detailed structures than BakedSDF and a smoother surface than NeRF2Mesh. Moreover, it is noteworthy that when employing rasterization-based rendering with the initial color model without the optimization stage, the rendering quality is poor, as shown in Table 1. In this table, "Ours w/o Geo" only trains appearance model $f_{color}$ without training vertex deformation. Training vertex deformation yields a 1.7 gain in indoor scenes by refining the mesh surface. "Ours w. RADAM" denotes that we use the rotation-invariant + +![](images/c4d21dcc549a2abcbefd5f0aff8296d2d7125c2fc3a01e1d411bf9d265a33197.jpg) +Figure 5. Qualitative Comparisons with Existing Methods. We visually compare our proposed technique with NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. The visual results of BakedSDF* generated in its first stage through volumetric rendering. Red and blue dotted lines emphasize subtle differences in rendering quality. +Figure 7. Naive QSlim Results + +![](images/fbd3150b7787c1382e9040a09771cdfb6e24a77d07f96c5d90075822678598db.jpg) + +ADAM (RADAM) for updating mesh vertices instead of a naive ADAM. However, it provides benefits for preventing mesh distortion. In Fig. 7, we show the visualization results of naive QSlim [12]. The naive QSlim leads to shrinkage and degeneration during the edge collapse process, and the boundary is over-smoothed compared to our method. + +# 5. Conclusion, Limitations, and Future Work + +This paper introduces a practical approach for reconstructing geometrically accurate meshes with photorealistic rendering quality. Our method achieves a compact mesh representation in unbounded large-scale scenes while preserving highly detailed geometry. Furthermore, the appearance field corresponding to this lightweight mesh is compatible with the classical rasterization pipeline, enabling efficient rendering. However, our method has limitations. BakedSDF employs a well-designed appearance model based on spherical Gaussians. The potential for enhancing rendering quality exists by incorporating a more sophisticated appearance model. Moreover, the gradient-based mesh optimization in the final stage may still result in mesh distortions, including self-intersections. We plan to employ a carefully designed optimizer to mitigate such distortions in future work. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2, 3, 4, 5, 6, 7 +[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[4] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3 +[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 1, 2, 3, 6, 7, 8 +[6] Jaehoon Choi, Dongki Jung, Taejae Lee, Sangwook Kim, Youngdong Jung, Dinesh Manocha, and Donghwan Lee. Tmo: Textured mesh acquisition of objects with a mobile device by using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16674-16684, 2023. 2 +[7] Jonathan Cohen, Marc Olano, and Dinesh Manocha. Appearance-preserving simplification. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 115-122, 1998. 2 +[8] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a vsi system for high performance graphics. Acm siggraph computer graphics, 22(4):21-30, 1988. 2 +[9] Carl Erikson and Dinesh Manocha. Gaps: General and automatic polygonal simplification. In Proceedings of the 1999 symposium on Interactive 3D graphics, pages 79-88, 1999. 2 +[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[11] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 2, 3 +[12] Michael Garland. Quadric-based polygonal surface simplification. Carnegie Mellon University, 1999. 2, 4, 5, 8 + +[13] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 2, 4 +[14] Michael Garland and Paul S Heckbert. Simplifying surfaces with color and texture using quadric error metrics. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 263-269. IEEE, 1998. 2 +[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4 +[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2, 3, 5 +[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 2, 3 +[18] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 7 +[19] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2, 3 +[20] Hugues Hoppe. View-dependent refinement of progressive meshes. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 189-198, 1997. 2 +[21] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 3 +[22] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. 2020. 3 +[23] Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057, 2020. 3 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 6, 7 +[26] Leif Kobbelt. *Sqrt(3)-subdivision*. In *Proceedings of the 27th annual conference on Computer graphics and interactive techniques*, pages 103–112, 2000. 6 +[27] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for + +high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5 +[28] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3 +[29] Peter Lindstrom. Out-of-core simplification of large polygonal models. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 259-262, 2000. 3 +[30] Peter Lindstrom and Claudio T Silva. A memory insensitive technique for large model simplification. In Proceedings Visualization, 2001. VIS'01., pages 121-550. IEEE, 2001. 3 +[31] Selena Zihan Ling, Nicholas Sharp, and Alec Jacobson. Vectoradam for rotation equivariant geometry optimization. Advances in Neural Information Processing Systems, 35:4111-4122, 2022. 3 +[32] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 2 +[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3 +[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 3, 4 +[35] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2 +[36] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, volume 40, pages 101-113. Wiley Online Library, 2021. 5 +[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3 +[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3 +[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 1, 2, 3, 7 +[40] Andrew Nealen, Takeo Igarashi, Olga Sorkine, and Marc Alexa. Laplacian mesh optimization. In Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia, pages 381-389, 2006. 5 +[41] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021. 2, 3 + +[42] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1, 2 +[43] Werner Palfinger. Continuous remeshing for inverse rendering. Computer Animation and Virtual Worlds, 33(5):e2101, 2022. 2, 3, 5, 6 +[44] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. In Proc. of the International Conf. on 3D Vision (3DV), 2023. 1, 2, 6, 7 +[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[46] Jarek Rossignac and Paul Borrel. Multi-resolution 3d approximations for rendering complex scenes. In Modeling in computer graphics: methods and applications, pages 455-465. Springer, 1993. 2, 3 +[47] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 2 +[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 3 +[49] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2, 3 +[50] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2022. 1, 2, 3, 5, 6, 7, 8 +[51] Chao Wang and Xiaohu Guo. Plane-based optimization of geometry and texture for rgb-d reconstruction of indoor scenes. In 2018 International Conference on 3D Vision (3DV), pages 533-541. IEEE, 2018. 3 +[52] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2 +[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu + +ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 4 +[55] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv, 2023. 1, 2, 4, 6, 7, 8 +[56] Zehao Yu, Anpei Chen, Bozidar Antic, Songyou Peng, Apratim Bhattacharyya, Michael Niemeyer, Siyu Tang, Torsten Sattler, and Andreas Geiger. Sdfstudio: A unified framework for surface reconstruction, 2022. 6 +[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 \ No newline at end of file diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/images.zip b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d4902cf0339bc0bf907d7b12c042409a99cee66e --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e27f828f21e2ae71a4c9f31c6bd73628fd2eebdf747bf337c9e4e307c233b073 +size 788175 diff --git a/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/layout.json b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f44128550bf19ba0a9ba4f0cac703da306a8ba42 --- /dev/null +++ b/2024/LTM_ Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering/layout.json @@ -0,0 +1,10527 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 517, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 517, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 517, + 140 + ], + "type": "text", + "content": "LTM: Lightweight Textured Mesh Extraction and Refinement of Large Unbounded Scenes for Efficient Storage and Real-time Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "spans": [ + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": "Jaehoon Choi" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Rajvi Shah" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Qinbo Li" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Yipeng Wang" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Ayush Saraf" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Changil Kim" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Jia-Bin Huang" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Dinesh Manocha" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Suhib Alsisan" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": " Johannes Kopf" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": "University of Maryland " + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 159, + 523, + 204 + ], + "type": "text", + "content": "Meta" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "text", + "content": "Advancements in neural signed distance fields (SDFs) have enabled modeling 3D surface geometry from a set of 2D images of real-world scenes. Baking neural SDFs can extract explicit mesh with appearance baked into texture maps as neural features. The baked meshes still have a large memory footprint and require a powerful GPU for real-time rendering. Neural optimization of such large meshes with differentiable rendering pose significant challenges. We propose a method to produce optimized meshes for large unbounded scenes with low triangle budget and high fidelity of geometry and appearance. We achieve this by combining advancements in baking neural SDFs with classical mesh simplification techniques and proposing a joint appearance-geometry refinement step. The visual quality is comparable to or better than state-of-the-art neural meshing and baking methods with high geometric accuracy despite significant reduction in triangle count, making the produced meshes efficient for storage, transmission, and rendering on mobile hardware. We validate the effectiveness of the proposed method on large unbounded scenes from mip-NeRF 360, Tanks & Temples, and Deep Blending datasets, achieving at-par rendering quality with " + }, + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "text", + "content": " reduced triangles and " + }, + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "inline_equation", + "content": "11 \\times" + }, + { + "bbox": [ + 46, + 258, + 290, + 534 + ], + "type": "text", + "content": " reduction in memory footprint." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Photorealistic reconstruction and rendering of real-world objects and scenes is a longstanding problem of importance with several applications in computer vision, robotics and AR/VR. In recent years, Neural Radiance Fields (NeRFs) [1, 2, 38, 39] have been quite successful at novel view synthesis. However, the scene representation is learned as an implicit volumetric representation that is expensive to evaluate. This is a crucial limitation for many applications, such as Augmented and Virtual Reality (AR/VR), often requiring real-time rendering with low memory and compute power." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 230, + 455, + 305 + ], + "blocks": [ + { + "bbox": [ + 322, + 230, + 455, + 305 + ], + "lines": [ + { + "bbox": [ + 322, + 230, + 455, + 305 + ], + "spans": [ + { + "bbox": [ + 322, + 230, + 455, + 305 + ], + "type": "image", + "image_path": "fd98c5886290e24ae55c9adf33cbe461f3022bfd3af6e4b2ae7a5697cdb3f88f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 308, + 455, + 373 + ], + "blocks": [ + { + "bbox": [ + 321, + 308, + 455, + 373 + ], + "lines": [ + { + "bbox": [ + 321, + 308, + 455, + 373 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 455, + 373 + ], + "type": "image", + "image_path": "d35952d2936b517156dbff20f54382b2c4a9178dc369af030dcbb225505e9257.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 376, + 547, + 433 + ], + "lines": [ + { + "bbox": [ + 304, + 376, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 547, + 433 + ], + "type": "text", + "content": "Figure 1. Example results using LTM showing mesh geometry (top-left) and novel-view synthesis (right). Bottom right plot compares performance of different methods on mip-NeRF360 outdoor dataset. Our method achieves comparable rendering quality to BakedSDF while reducing the mesh size considerably." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 457, + 230, + 527, + 277 + ], + "blocks": [ + { + "bbox": [ + 457, + 230, + 527, + 277 + ], + "lines": [ + { + "bbox": [ + 457, + 230, + 527, + 277 + ], + "spans": [ + { + "bbox": [ + 457, + 230, + 527, + 277 + ], + "type": "image", + "image_path": "2f0cc1c40195cef09e1d21bbc39ceab91b3debf6fbd7501f48ff2357c658a3a9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 457, + 277, + 527, + 324 + ], + "blocks": [ + { + "bbox": [ + 457, + 277, + 527, + 324 + ], + "lines": [ + { + "bbox": [ + 457, + 277, + 527, + 324 + ], + "spans": [ + { + "bbox": [ + 457, + 277, + 527, + 324 + ], + "type": "image", + "image_path": "9a4186137f44c69c544be406205e412211be0ff128b226383c4654ccb833d0fe.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 457, + 324, + 527, + 372 + ], + "blocks": [ + { + "bbox": [ + 457, + 324, + 527, + 372 + ], + "lines": [ + { + "bbox": [ + 457, + 324, + 527, + 372 + ], + "spans": [ + { + "bbox": [ + 457, + 324, + 527, + 372 + ], + "type": "image", + "image_path": "2112e1b85a6718c98e4b5a798ff1bbf3e253d16c73e9b8879bb350b1d22fa518.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 437, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 547, + 484 + ], + "type": "text", + "content": "Since modern graphics engines are highly optimized to rasterize triangle meshes, researchers have started exploring ways to extract explicit mesh representations from volumetric functions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 485, + 548, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 548, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 548, + 676 + ], + "type": "text", + "content": "MobileNeRF [5] uses a polygon mesh with texture maps storing feature vectors and opacity. Although their rendering quality is good, the mesh quality is far from an ideal surface. While NeRF2Mesh [50] and NeRFMeshing [44], based on volumetric density fields, excel at generating highly detailed geometry, they often result in noisy and bumpy surfaces and numerous floaters, especially in planar regions, which are common in large-scale scenes. In contrast, neural Signed Distance Function (SDF) methods [42, 52, 54] typically excel in reconstructing accurate surface geometry. However, they may not be able to capture small or detailed geometry, and their rendering quality is inferior to volumetric density-field methods. BakedSDF [55] overcomes these limitations by applying mip-NeRF 360 techniques to train SDF representation, though the resulting mesh representation has a high memory overhead." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Previous neural SDF methods that rely on the Marching Cube algorithm [34] result in meshes with high storage overhead. This results in a significant challenge for practi" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5053" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "cal applications on mobile devices, where low disk storage and fast data transfer are essential. Instead of relying on the Marching Cubes, some methods [11,28,48,49] propose differentiable mesh reconstruction methods. These methods aim to extract an isosurface from an implicit function and can be seamlessly integrated with differentiable rendering approaches [16, 17]. However, most methods primarily concentrate on a single object with ideal implicit surface functions, unsuited for large unbounded real-world scenes [2, 55]. Reconstructing an ideal mesh for the unbounded large-scale scene is still an open problem, often resulting in non-manifold geometry or missing surfaces. Consequently, the renderings can result in visual artifacts due to the incompleteness of the mesh geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 289, + 386 + ], + "type": "text", + "content": "Main Results: In this paper, our goal is to achieve a realistic rendering with a compact mesh representation that optimally utilizes disk storage for unbounded large-scale scenes, all while preserving geometric quality. Our method initially trains neural fields for geometry and appearance separately. For geometry reconstruction, we adopt the BakedSDF approach [55] to train an SDF in a contracted space [2]. Regarding the appearance field, we utilize hash-based encoding [39] with two shallow MLPs to represent diffuse and specular colors. The SDF and appearance fields are compact representations for reconstructing a high-quality mesh and achieving realistic rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 388, + 289, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 289, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 289, + 519 + ], + "type": "text", + "content": "Given the initial geometry representations, we perform mesh extraction and decimation. As unbounded large-scale scenes occupy a large amount of disk storage, our method focuses on generating a lightweight mesh without compromising geometric details. We adaptively extract the mesh for the central and background regions in the contracted space, respectively. Subsequently, we modify a mesh decimation [14] designed for large-scale scenes. Our method eliminates redundant triangles corresponding to simple geometric features, such as planar surfaces, while preserving complex geometric features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 714 + ], + "type": "text", + "content": "Since the appearance field is trained by volumetric rendering, the initial model is unsuitable for rasterization-based rendering. Moreover, the mesh from the previous step still fails to accurately represent areas with highly-detailed geometry (thin structures) or semi-transparent objects. These mesh defects hurt rendering quality. Thus, we use differentiable rendering [27] and utilize image-space supervision to refine the mesh surface by rectifying these mesh defects while preserving geometric accuracy. Due to the highly simplified mesh and various regularization techniques, our approach can deform the mesh topology without causing severe mesh distortion, a common issue in gradient-based mesh optimization [41, 43]. Compared to mesh-based neural rendering methods [5, 44, 50, 55], our approach results in lower storage overhead while providing comparable rendering quality. To sum up, our contributions include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 72, + 547, + 206 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 72, + 547, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 109 + ], + "type": "text", + "content": "- We effectively utilize classical mesh decimation techniques to extract compact geometry from neural SDFs of large unbounded scenes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 115, + 547, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 115, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 115, + 547, + 152 + ], + "type": "text", + "content": "- We introduce a new approach for joint optimization of appearance and vertex deformation that preserves geometric details and improves rendering quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 547, + 206 + ], + "type": "text", + "content": "- We achieve a trifecta of precise geometry, efficient storage, and rendering quality comparable to the state of the art on diverse captured environments such as mip-NeRF 360, Tanks & Temples, and the Deep Blending dataset." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 392, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 235, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 235, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 235, + 547, + 617 + ], + "type": "text", + "content": "Neural Rendering Neural Fields achieve remarkable photo-realistic rendering quality. One of the most successful approaches is Neural Radiance Fields (NeRFs) [38], which adopts volumetric implicit representations and trains a radiance field through ray marching. Mip-NeRF [1, 2] presents casting a conical frustum into the scene space and extends this integrated positional encoding to unbounded scenes by using a novel contraction function. Instead of learning volumetric density, neural surface reconstruction methods [6, 42, 52, 54] propose learning a Signed Distance Function (SDF) for surface reconstruction via volumetric rendering. Furthermore, several methods have been proposed to accelerate NeRFs by utilizing efficient data structure [3, 10, 39], although achieving real-time rendering for high-resolutions remains a challenge. Alternatively, to enhance rendering speed, \"baking\" based methods [5, 19, 44, 45, 50, 55] typically precompute and store a trained NeRF into a more efficient representation. SNeRG [19] adopts a deferred shading model [8] and bakes trained NeRFs into sparse voxel grid representations with neural features. In particular, the mesh is widely used as geometry representation and bakes neural fields into its texture. MobileNeRF [5] builds on a mesh rasterization pipeline for efficient rendering on mobile architecture. NeRF2Mesh [50] introduces a two-step framework for NeRF training and textured mesh extraction. NeRFMeshing [44] proposes a novel method to produce a truncated SDF from NeRFs and extracts the accurate mesh. BakedSDF [55] enables us to reconstruct high-quality mesh and bake it into a small appearance model based on spherical Gaussians for fast rendering. DNMP [35] and NeuRaS [32] adopt mesh-based geometry and rasterization-based rendering for large driving scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Mesh Processing Traditional graphics research has developed effective techniques for simplifying meshes, based on vertex decimation [46, 47], edge collapse [9, 12-14, 20] and appearance preservation [7]. Vertex decimation methods prioritize vertices based on shape heuristics and remove the least important ones. Edge collapse methods rank the edges based on the cost and subsequently contract pairs of vertices. QSlim [12, 13] introduces the concept of the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5054" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "Quadric Error Metric (QEM), which is defined as the sum of quadrics representing the distance of a point from a set of planes. Following the edge collapse, they utilize this QEM to identify which edges should be removed while preserving geometric topology. Lindstrom et al. [29, 30] use vertex clustering [46] on a uniform grid and accumulate a quadric error matrix to the occupied grid cell. Wang et al. [51] employ RGB-D sequences to reconstruct a lightweight mesh." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 289, + 421 + ], + "type": "text", + "content": "Differentiable Rendering and Mesh Optimization Recently, many differentiable rendering techniques [23] have emerged, aimed at facilitating end-to-end optimization of the rendering process by employing useful gradients. In particular, several works [4, 33] focus on approximating the rasterization stage of the rendering to compute meaningful gradients. Furthermore, modular differentiable renderers [21, 22, 27] exist, allowing users to customize and modify specific functions. Differentiable mesh reconstruction techniques [11, 28, 48] can build a mesh through gradient-based optimization. In particular, differentiable rendering [16, 17, 49] can offer 2D supervision to refine and enhance the geometric quality of the mesh. Some works [16, 17] adopt an inverse rendering approach, leveraging physics-based rendering to synthesize shapes. Nevertheless, gradient-based mesh optimization [41] is highly sensitive to initialization and prone to significant mesh distortion. Some methods [31, 41, 43] incorporate a novel ADAM optimizer and employ strong regularization techniques. However, this research is currently limited to generating single objects and is challenging to apply to large-scale scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 434, + 127, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 127, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 127, + 447 + ], + "type": "text", + "content": "3. Our Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": "Given multi-view images and corresponding camera poses, we aim to reconstruct the texture and geometry of unbounded large-scale scenes. Fig. 2 shows the three main stages of our pipeline: Initialization, Mesh Processing, and Optimization. In the first stage, we employ volumetric rendering to train neural implicit representation for geometry " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " and appearance " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": ". Subsequently, we leverage the geometry representation " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " to extract a high-resolution mesh via Marching Cubes [34] and reduce the number of triangles without causing geometry distortion. Next, we utilize differentiable rendering [27] to jointly optimize the explicit 3D mesh " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " and implicit appearance field " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " obtained from the first stage and refine the mesh surface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 622, + 271, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 271, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 271, + 635 + ], + "type": "text", + "content": "3.1. Initialization of Geometry and Appearance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "In this stage, we first learn neural implicit representations for both appearance and geometry. A neural radiance field [38] is an implicit representation of a 3D scene volume, a mapping from a 3D position " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and a ray viewing direction " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " to an RGB color " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": " and volume density " + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": ". In practice, the networks modeled by MLPs are used to generate this" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 537, + 129 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 537, + 129 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 537, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 537, + 129 + ], + "type": "image", + "image_path": "f0e13cda3153589f6e86fb7db0a0c2f728b647df793244f2e55086eadcaf7c75.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 133, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 304, + 133, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 545, + 230 + ], + "type": "text", + "content": "Figure 2. Overview of Our Pipeline. It consists of three steps: Initialization (Sec. 3.1), Mesh Processing (Sec. 3.2), and Optimization (Sec. 3.3). Our method focuses on baking the initial appearance and geometry representation, trained by volumetric rendering, into an explicit mesh and a memory-efficient texture conducive to real-time rendering. We employ classical mesh decimation to minimize storage requirements, introduce the optimization stage to facilitate rasterization-based rendering and refine the detailed mesh structure." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 239, + 479, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 479, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 479, + 251 + ], + "type": "text", + "content": "volume density and view-dependent color:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 258, + 545, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 258, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 353, + 258, + 545, + 270 + ], + "type": "interline_equation", + "content": "\\rho = f _ {\\text {d e n s i t y}} (x), \\quad c = f _ {\\text {c o l o r}} (x, d). \\tag {1}", + "image_path": "9a203e3f25dfd00ca50c4c5fc527ad11b74e99051a1f7f2a53fb401c43f7fe7f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": "To render the color of a pixel, we first compute the 3D points " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " along the viewing ray associated with each pixel " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "x(t) = o + td" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": ". Then, estimated volume densities " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "\\rho_i" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " and colors " + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": " are utilized to approximate a volume rendering equation [37] using numerical quadrature:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 342, + 342, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 342, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 342, + 342, + 545, + 369 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i} T _ {i} \\alpha_ {i} c _ {i}, \\quad T _ {i} = \\exp \\left(- \\sum_ {j < i} \\rho_ {i} \\delta_ {i}\\right), \\tag {2}", + "image_path": "c52f2a10b0740324575a8710b1b68a336257121e1a1624673e572d2f0631276c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "content": " is the accumulated transmittance and " + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp(-\\rho_{i}\\delta_{i})" + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "content": " is the discrete opacity of the point. " + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t_{i} - t_{i-1}" + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "content": " is the distance between two neighboring points along the ray. To deal with large unbounded scenes, mip-NeRF 360 [2] introduces a proposal MLP for efficient sampling and space contraction to represent an unbounded scene. The scene contraction function " + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 374, + 545, + 470 + ], + "type": "text", + "content": " (in Eq. 3) maps the unbounded points into bounded space:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 476, + 545, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 476, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 347, + 476, + 545, + 518 + ], + "type": "interline_equation", + "content": "f (x) = \\left\\{ \\begin{array}{l l} x & | | x | | \\leq 1, \\\\ \\left(2 - \\frac {1}{| | x | |}\\right) \\frac {x}{| | x | |} & | | x | | \\geq 1. \\end{array} \\right. \\tag {3}", + "image_path": "a750dd4cba741d7e6b6142c17d0699f4a6bcbdc4d2a7f218689a4f441dbc806c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Appearance Similar to previous works [5, 19, 50], we decompose an appearance field " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " into shallow MLPs to learn diffuse colors " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "c_d" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " and specular colors " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "c_s" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " separately. Particularly, the view-dependent effects are modeled by a small MLP that can be easily integrated into a fragment Shader. In order to ensure the MLPs are able to model an unbounded scene well, we adopt strategies similar to the ones introduced in mip-NeRF360 work, specifically, scene contraction, proposal MLP, and regularization for the background. We train both MLP networks " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{density}" + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": " with a multi-resolution hash positional encoding [39] using the volume rendering objective in Eq. 2. The appearance field trained in this stage is further optimized to facilitate rasterization-based rendering, replacing the computationally expensive ray marching. This optimization is described later in Section 3.3." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5055" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 512, + 197 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 512, + 197 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 512, + 197 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 512, + 197 + ], + "type": "image", + "image_path": "a94c3532d95a8f705e7b0348191d136b71eb0df50b5b675286e07fa10dcb62a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": "Figure 3. Our Approach: In Section 3.1, we start training both the appearance " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " and geometry " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " representations for a large-scale scene using volumetric rendering (shown in orange). Then, in Section 3.2, our method extracts the mesh from " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " and significantly simplifies its structure, resulting in the lightweight mesh " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " (shown in blue). In Section 3.3, we jointly train vertex deformation " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 198, + 547, + 243 + ], + "type": "text", + "content": " computed from appearance modeling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "text", + "content": "Geometry Initial geometry modeling in our method is similar to BakedSDF [55]. The core mechanism for neural surface reconstruction was introduced in VolSDF [54] for small objects. BakedSDF shows that by leveraging the efficient sampling and contraction ideas introduced in mipNeRF360 [2], VolSDF [54] can be extended to large unbounded scenes. We follow this and jointly train two MLPs using the volumetric rendering objective, (i) " + }, + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "text", + "content": " - representing geometry of a scene as a signed distance function, and (ii) " + }, + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "inline_equation", + "content": "g_{color}" + }, + { + "bbox": [ + 46, + 247, + 289, + 376 + ], + "type": "text", + "content": " - a color prediction network, as shown in the Geometry Reconstruction block in Fig.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": "Following VolSDF, we define the volume density " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\sigma(x)" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": " as a function of surface geometry " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": ", namely, " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\sigma(x) = \\alpha \\Psi_{\\beta}(f_{sdf}(x))" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\Psi_{\\beta}" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": " is the Cumulative Distribution Function (CDF) of the Laplace distribution with zero mean and " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": " is the learnable scaling parameter. The Eikonal loss [15] to regularize the " + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 47, + 378, + 288, + 450 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 454, + 287, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 454, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 287, + 469 + ], + "type": "interline_equation", + "content": "L _ {e i k} = \\left(\\| \\nabla f _ {s d f} (x) \\| _ {2} - 1\\right) ^ {2}. \\tag {4}", + "image_path": "be743ec697cf97fbf03e2df141999ae04faa4b0b7019160bae268b8cd98fcd58.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "text", + "content": "The color MLP " + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "inline_equation", + "content": "g_{color}" + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "text", + "content": ", in addition to the position and viewing direction, also takes the predicted surface normal and a 256-dimensional geometric feature as inputs. It is important to note that the sole purpose of color MLP " + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "inline_equation", + "content": "g_{color}" + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "text", + "content": " is geometry initialization and we discard it after this step. This is not to be confused with the shallow decomposed color MLP " + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 474, + 287, + 570 + ], + "type": "text", + "content": " trained specifically to model appearance that is also jointly refined with mesh vertices in the later stage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 575, + 225, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 225, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 225, + 588 + ], + "type": "text", + "content": "3.2. Mesh Extraction and Decimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "In this stage, we first extract a mesh by using Marching Cubes (MC) [34] algorithm on learned SDF (" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " evaluated over a uniform grid) and simplify the extracted mesh using a geometry-aware quadric error based mesh decimation. While this approach works pretty well for extracting meshes of well-captured small-scale objects and scenes, it presents unique challenges for large-scale scenes. Meshing neural SDFs of large-scale scenes with marching cubes can often result in spurious surfaces arising from (i) insufficient grid resolution to model complex geometry of nearby" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 246, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 330 + ], + "type": "text", + "content": "objects, and (ii) poorly learned geometry of distant objects due to insufficient view-points observing such distant regions. Furthermore, since our SDF " + }, + { + "bbox": [ + 304, + 246, + 545, + 330 + ], + "type": "inline_equation", + "content": "f_{sdf}" + }, + { + "bbox": [ + 304, + 246, + 545, + 330 + ], + "type": "text", + "content": " is trained in the contracted space, the resulting mesh is also defined in the contracted space. When we uncontract, i.e. warp the mesh vertices back into the original Euclidean space, it further exacerbates the geometry distortion in the background region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "text", + "content": "We aim to preserve the detailed geometry in central (foreground) regions while achieving a coarser but less noisy reconstruction of distant (background) regions, within a small triangle budget. We achieve this by treating foreground and background regions distinctly during both extraction and simplification steps. During MC extraction, we set a finer grid resolution of 2048 and a spatial range of " + }, + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "inline_equation", + "content": "[-1,1]^3" + }, + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "text", + "content": " for the central region, and a coarser resolution of 1024 and a spatial range of " + }, + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "inline_equation", + "content": "[-2,2]^3" + }, + { + "bbox": [ + 304, + 331, + 546, + 498 + ], + "type": "text", + "content": " for the background region. For decimation, we first partition the central and background regions and employ an aggressive mesh decimation in background regions. This helps us transform bumpy and noisy background areas into smooth surfaces while also reducing redundant triangles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "text", + "content": "We first apply vertex clustering and then run quadric error based simplification algorithm - QSlim [12, 13] for all clusters. Then, we run edge contraction for each cluster to reach a threshold and repeat this process until convergence. The QSlim algorithm calculates a quadric metric for each vertex to decide whether the two vertices should be merged and aggregates these quadrics during each edge collapse operation. The basic quadric error in QSlim [13] is defined as the point-to-plane distance from vertex " + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "text", + "content": " to all planes " + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "text", + "content": " of its adjacent faces " + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "inline_equation", + "content": "N(v)" + }, + { + "bbox": [ + 304, + 498, + 547, + 619 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 628, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 628, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 315, + 628, + 545, + 666 + ], + "type": "interline_equation", + "content": "E (v) = \\sum_ {p \\in N (v)} \\left(p ^ {T} v\\right) ^ {2} = v ^ {T} \\left(\\sum_ {p \\in N (v)} p p ^ {T}\\right) v = v ^ {T} Q _ {v} v. \\tag {5}", + "image_path": "d6b2eb601743d6917a088e00171c8f82fe459c516daa2902b0e9817c062976ff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "The quadric " + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "Q_{v}" + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": " is the symmetric 4x4 matrix, and each edge computes this quadric by adding the quadrics of the corresponding two vertices. The most commonly used quadric error is defined solely based on 3D vertex positions " + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5056" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "Intuitively, the most effective strategy is to represent planar surfaces with a smaller number of large triangles while keeping a consistent number of triangles for complex shapes. Variants that incorporate additional surface attributes, such as normals or textures in calculating the quadric error are proposed in [12]. We experimentally observed that compared to the basic QSlim algorithm, using both the vertex positions " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " and the corresponding vertex normal " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " for quadric error is very effective at prioritizing triangle decimation on flat surfaces while retaining them on complex surfaces. The combination of vertex clustering and geometry-aware QSlim helps us in effectively reducing the number of triangles while preserving the original geometry." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 250, + 261, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 250, + 261, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 261, + 262 + ], + "type": "text", + "content": "3.3. Appearance and Geometry Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "content": "In this stage, we optimize the appearance model jointly with mesh vertex deformation and refine the mesh topology using a geometric complexity based criterion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": "Joint optimization of appearance and vertex deformation Given the simplified mesh " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " and the initial appearance field " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": ", we employ the fast differentiable rasterizer introduced in Nvdiffrast [27] to jointly optimize the mesh geometry and the appearance field. The initial appearance field " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " trained using the volume rendering objective (Eq. 1) is agnostic of the surface geometry of the extracted and simplified mesh. To finetune the model to fit the mesh geometry, we rasterize the mesh into each of the training camera poses and obtain surface points " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " and face normals " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " interpolated per pixel. Then we apply the contraction function in Eq. 3 to warp the surface points. This is because the appearance field defined in contracted space proves to be more effective in enhancing the rendering quality [2]. The appearance field " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " takes these surface points " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": " as input, rendering an image " + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 46, + 306, + 289, + 569 + ], + "type": "text", + "content": ". Initially, rendering performance is degraded because the appearance field only utilizes the intersection point between the ray and the surface rather than accumulating all sampled points along the ray. We finetune this appearance field to be compatible with mesh rendering, akin to the functionality of a fragment Shader." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "text", + "content": "Given the positions " + }, + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "text", + "content": " of the mesh vertices, we train per-vertex deformation " + }, + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 46, + 570, + 289, + 702 + ], + "type": "text", + "content": " using differentiable rendering [27]. We explore two approaches to optimize vertex deformation: a standard ADAM optimizer [24] and a rotation-invariant ADAM optimizer [43]. The ADAM optimizer is widely employed [16, 50] for deforming mesh topology to enhance rendering performance. While the ADAM optimizer often results in mesh distortions such as flipped triangles and self-intersections, our method mitigates mesh distortion, especially on smooth surfaces, due to the rigorous mesh decimation performed in Section 3.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "Objective Function To jointly train appearance field and vertex deformation, we minimize the image-space L1 loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "L_{\\text{appearance}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " between the rendered image " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " of the deformed surface and the groundtruth image " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": " for a camera view " + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 367, + 125, + 545, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 125, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 367, + 125, + 545, + 150 + ], + "type": "interline_equation", + "content": "L _ {\\text {a p p e a r a n c e}} = \\sum_ {i} \\left| \\hat {I} _ {i} - I _ {i} \\right|. \\tag {6}", + "image_path": "d99e1b22b1509f6b417cb373be5766e5c0b60f7185e8cb649edec4b945e1b19e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 156, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 240 + ], + "type": "text", + "content": "Naively training vertex deformation is highly likely to result in undesirable mesh topology characterized by a disorganized arrangement of triangles, often referred to as a \"polygon soup.\" To prevent this mesh distortion, we employ various regularization methods. We use the Laplacian mesh smoothness loss and the normal consistency loss inspired by [36]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 382, + 248, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 248, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 382, + 248, + 545, + 266 + ], + "type": "interline_equation", + "content": "L _ {l a p} = \\sum \\| L V \\| _ {2} \\tag {7}", + "image_path": "7be9a1adcfc675b2494ef616644b5a4bad00ea54762390e1a0b37da9daf63ca1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " is the number of vertices) is the matrix of deformed vertex coordinates, indicating " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "x + \\Delta x" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": ", at each training step, and " + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": " is the graph Laplacian of the mesh [40]. Minimizing the magnitude of a vertex's differential coordinates (" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "inline_equation", + "content": "LV" + }, + { + "bbox": [ + 304, + 269, + 545, + 342 + ], + "type": "text", + "content": ") reduces its distance from its neighbors' average position. The normal consistency loss is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 347, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 347, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 370, + 347, + 545, + 373 + ], + "type": "interline_equation", + "content": "L _ {n c} = \\sum_ {i, j \\in F} \\left(1 - n _ {i} \\cdot n _ {j}\\right) ^ {2}, \\tag {8}", + "image_path": "60bd8a7e4afa8ccb14f62ce5f331996d66e99d13d14210e64c948d7995da0c08.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "n_j" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " denote the surface normal of two adjacent faces and " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " is the set of faces sharing a common edge. Additionally, in Section 3.2, a high-resolution mesh is rasterized in the first pass before decimation. The output of this rasterization is used to interpolate face normals per pixel, resulting in a normal map " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\hat{N}_i" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " for every viewpoint. We calculate the L1 loss between a rendered normal " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " from current mesh " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\hat{N}_i" + }, + { + "bbox": [ + 304, + 378, + 545, + 475 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 374, + 479, + 545, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 479, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 374, + 479, + 545, + 504 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o r m}} = \\sum_ {i} | \\hat {N} _ {i} - N _ {i} |. \\tag {9}", + "image_path": "0542358041c728c484e4e594466b3e1bbf515efdc9fd98ed1a51fbc29adeb2d8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "type": "inline_equation", + "content": "L_{norm}" + }, + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "type": "text", + "content": " provides geometry supervision, particularly in planar regions, helping prevent severe mesh distortion when moving vertices. The total loss function " + }, + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 509, + 545, + 544 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 545, + 564 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {c o l o r}} + \\lambda_ {\\text {l a p}} L _ {\\text {l a p}} + \\lambda_ {\\text {n c}} L _ {\\text {n c}} + \\lambda_ {\\text {n o r m}} L _ {\\text {n o r m}}, \\tag {10}", + "image_path": "19c5ea3172e8b2ad6963d24ffde0c645f61fc27523fdc98bf62b38a703721040.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 570, + 538, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 538, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 538, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 570, + 538, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 306, + 570, + 538, + 582 + ], + "type": "text", + "content": " is the weight to balance between the loss terms." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Mesh Refinement Inspired by NeRF2Mesh [50], we adopt a classical mesh refinement approach based on mesh subdivision and decimation during training. The mesh refinement strategy is applied differently to the central and background regions. Instead of rendering error proposed by [50], which can arise from many factors including specularity, we employ a criterion: the vertex deformation values for subdividing triangle faces. The large vertex deformation implies a more substantial triangle budget to accurately capture the intricacies of highly detailed geometry associated" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5057" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 218, + 273 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 218, + 273 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 218, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 218, + 273 + ], + "type": "image", + "image_path": "14527b1c3e0e24795071a1d22597decaa055bf613881ba4a5a53b2c1e2b43560.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 546, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 546, + 297 + ], + "type": "text", + "content": "Figure 4. Visual comparison of Our Method, NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. Our method and NeRF2Mesh show shading mesh and wireframe mesh extracted without texture, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 379, + 273 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 379, + 273 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 379, + 273 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 379, + 273 + ], + "type": "image", + "image_path": "1a1e8a9d124af83baaa9a7f230bbfd54f09a1dc282b3ba78fdbd72f736b07eff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 381, + 72, + 462, + 273 + ], + "blocks": [ + { + "bbox": [ + 381, + 72, + 462, + 273 + ], + "lines": [ + { + "bbox": [ + 381, + 72, + 462, + 273 + ], + "spans": [ + { + "bbox": [ + 381, + 72, + 462, + 273 + ], + "type": "image", + "image_path": "0e0d789746e2598603ed6a323aff4a2c18b8a135ce7eb345d64e6f2ae7605ba9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 545, + 273 + ], + "type": "image", + "image_path": "c06279adae33f8bd1274b8a3ed84261d91db0b072cd0ba778d39907affce1112.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "text", + "content": "with those vertices. We sort the absolute values of vertex deformation and determine a " + }, + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "text", + "content": " of faces to increase triangles. Then, we apply " + }, + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\sqrt{3}" + }, + { + "bbox": [ + 46, + 305, + 289, + 461 + ], + "type": "text", + "content": " subdivision [26], which adds a vertex at the barycenter of a triangle and connects it to all vertices of this triangle. This scheme enables us to augment the triangle budget for complex geometry and mitigate self-intersection during vertex deformation. Furthermore, the mesh decimation presented in Section 3.2 is applied after every mesh subdivision to remove redundant triangles on simple and flat geometry. For background regions, we examine all training views and check faces intersecting with ray casting. We utilize dilation for all invisible masks and employ mesh decimation to remove redundant faces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": "Rotation-Invariant ADAM The rotation-invariant ADAM formulation [43] modifies the update rule in ADAM, resulting in the update step: " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "\\Delta x = \\zeta \\cdot v \\cdot l_{ref}" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": " is the learning rate, and " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": " is the standard ADAM update step, which is the loss gradient with respect to " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "l_{ref}" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": " is the coefficient initially set to the edge length incident to " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": ", and it gets updated based on the average " + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 461, + 289, + 546 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 569 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 165, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 165, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 165, + 588 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Datasets: To validate the effectiveness of our method, we mainly conduct experiments on seven unbounded, large-scale " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " indoor and outdoor scenes from the mip-NeRF 360 [2] dataset. Also, we validate the generalization ability of our approach on two different datasets, including large-scale scenes: two scenes from the Tanks & Temples [25] dataset and two scenes from the Deep Blending [18] dataset. Implementation Details: For all datasets, we subsampled every eight frames for test sets following the mip-NeRF 360 paper [2]. For geometry reconstruction in Section 3.1, we" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": "use SDFstudio [56] to implement BakedSDF. In Section 3.2, we reduce " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": " of the triangles in the central region and " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": " of the triangles in the background regions through decimation. In Section 3.3, we run all experiments for 40k iterations, and a learning rate is annealed from 1e-2 to 1e-3. We set the loss weights to " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{lap} = 1\\mathrm{e - }3" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{nc} = 1\\mathrm{e - }3" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "\\lambda_{norm} = 1\\mathrm{e - }4" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": ". We apply mesh refinement at 8k, 16k, and 24k training steps. Subsequently, we compute diffuse color and specular features using " + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 304, + 305, + 547, + 437 + ], + "type": "text", + "content": " to export textures. All these color maps and features are 3-channel .png files. Additional details are provided in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 430, + 456 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 640 + ], + "type": "text", + "content": "We compare our proposed method with three different methods: 1) MobileNeRF [5] is the state-of-the-art method, enabling NeRF to achieve real-time rendering of explicit 3D mesh and neural features. 2) NeRF2Mesh [50] employs a volumetric rendering to reconstruct the coarse mesh and refine the mesh by jointly optimizing appearance and geometry. 3) BakedSDF [55] is a recent method for optimizing a neural surface-volume representation for real-time rendering of large unbounded scenes. We also add NeRFMeshing [44], which distills the rendering NeRF into a neural field that represents TSDF (Truncated SDF) and reconstructs the 3D mesh. As BakedSDF and NeRFMeshing have not released public code, we adopt the metrics provided in their original papers. BakedSDF has made its GLB file available. We utilize this file for visualization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Figure 4 qualitatively compares our method with NeRF2Mesh [50], BakedSDF [55], and MobileNeRF [5]. We visualize the shading meshes extracted without texture, which is the most effective way to visualize geometric differences. Furthermore, we include a wireframe mesh, distinguishing between our method and NeRF2Mesh. Visu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5058" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 533, + 162 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 533, + 162 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 533, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 533, + 162 + ], + "type": "table", + "html": "
MethodMeshOutdoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)Indoor (PSNR ↑ / SSIM ↑ / LPIPS ↓)
BicycleGardenStumpMeanRoomCounterKitchenBonsaiMean
Instant-NGP [39]X22.1 / 0.49 / 0.4924.5 / 0.65 / 0.3123.6 / 0.57 / 0.4523.4 / 0.57 / 0.4229.2 / 0.85 / 0.3026.43 / 0.80 / 0.3428.5 / 0.82 / 0.2530.3 / 0.89 / 0.2328.6 / 0.84 / 0.28
mip-NeRF 360 [2]X24.4 / 0.68 / 0.3026.9 / 0.81 / 0.1726.4 / 0.74 / 0.2625.9 / 0.75 / 0.2431.6 / 0.91 / 0.2129.5 / 0.89 / 0.2032.3 / 0.92 / 0.1333.5 / 0.94 / 0.1831.7 / 0.92 / 0.18
MobileNeRF [5]O21.7 / 0.43 / 0.5118.8 / 0.59 / 0.3623.9 / 0.56 / 0.4321.5 / 0.53 / 0.4328.9 / 0.85 / 0.2825.1 / 0.72 / 0.2926.8 / 0.79 / 0.7923.8 / 0.71 / 0.7226.2 / 0.77 / 0.52
NeRF2Mesh [50]O22.1 / 0.48 / 0.5123.4 / 0.55 / 0.4022.5 / 0.54 / 0.4622.7 / 0.52 / 0.4625.7 / 0.79 / 0.3523.9 / 0.71 / 0.3524.0 / 0.61 / 0.3625.0 / 0.77 / 0.2924.7 / 0.72 / 0.34
NeRFMeshing [44]O21.1 / - / -22.9 / - / -22.6 / - / -22.2 / - / -26.1 / - / -20.0 / - / -23.6 / - / -25.6 / - / -23.8 / - / -
BakedSDF [55]O22.0 / 0.57 / 0.3724.9 / 0.75 / 0.2123.6 / 0.59 / 0.3723.5 / 0.64 / 0.3228.7 / 0.87 / 0.2525.7 / 0.81 / 0.2826.7 / 0.82 / 0.2427.2 / 0.85 / 0.2627.0 / 0.84 / 0.26
OursO22.4 / 0.52 / 0.4423.5 / 0.64 / 0.3023.7 / 0.57 / 0.4223.2 / 0.58 / 0.3929.3 / 0.88 / 0.2325.1 / 0.77 / 0.2726.5 / 0.80 / 0.2127.3 / 0.84 / 0.2227.1 / 0.82 / 0.23
Ours w/o OptO16.1 / 0.30 / 0.5318.8 / 0.45 / 0.3919.4 / 0.39 / 0.5018.1 / 0.38 / 0.4817.2 / 0.54 / 0.4616.2 / 0.47 / 0.4718.1 / 0.50 / 0.3616.4 / 0.54 / 0.4317.0 / 0.51 / 0.43
Ours w/ GOO21.1 / 0.48 / 0.4723.2 / 0.65 / 0.3023.0 / 0.53 / 0.4622.5 / 0.55 / 0.4127.5 / 0.85 / 0.2623.6 / 0.72 / 0.3325.5 / 0.78 / 0.2325.9 / 0.76 / 0.325.4 / 0.78 / 0.28
Ours w. RADAMO21.6 / 0.48 / 0.4623.4 / 0.66 / 0.3023.5 / 0.56 / 0.4522.8 / 0.57 / 0.4027.5 / 0.85 / 0.2624.1 / 0.74 / 0.3125.8 / 0.80 / 0.2226.8 / 0.80 / 0.2626.0 / 0.80 / 0.26
", + "image_path": "fcbe40fc2a1c6e47281651ebb80208178fb08f7882aae6d1bafa2bf83c1b4652.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 194, + 543, + 239 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 544, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 544, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 544, + 189 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison on mip-NeRF 360 Dataset. “-” denotes a missing implementation. The best and second best-performing algorithms for each metric are bolded and underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 194, + 543, + 239 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 543, + 239 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 543, + 239 + ], + "type": "table", + "html": "
MethodBicycleGardenStumpRoomCounterKitchenBonsaiMean
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]2011435 / 6638135723 / 3401641069 / 497220903 / 421170864 / 39995552 / 25684606 / 284153879 / 1262
NeRF2Mesh [50]129815 / 858117650 / 779123415 / 82069328 / 46279517 / 38489601 / 43470319 / 63697387 / 667
BakedSDF [55]1087111248 / 4438451310773M / 2107294319667 / 384673857227 / 1411058512212 / 2389265913832 / 2705664113629 / 266796871564M / 27951
Ours83270 / 52950113 / 218120425 / 84745124 / 23858138 / 2684070 / 13642233 / 45962196 / 385
", + "image_path": "db2286bca39d67e0509af4832514d0c3cf00d9ac33817069a80bafd26bfd949e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 271, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 287, + 427 + ], + "type": "text", + "content": "alizing the wireframe mesh from BakedSDF lacks meaningful interpretation due to its extremely high face density. MobileNeRF generates a \"triangle soup\" that inaccurately represents the scene's geometry. Also, compared to the mesh reconstructed by NeRF2Mesh, our method shows sharp details of complex geometry and smooth surfaces on the ground or wall. Moreover, in the wireframe mesh, our method employs more triangles in intricate shapes like the bicycle or the garden vase while allocating smaller triangles to planar regions or simpler structures. While the geometry of BakedSDF exhibits superior quality compared to our mesh reconstruction, achieving this high-quality mesh necessitates approximately " + }, + { + "bbox": [ + 46, + 271, + 287, + 427 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 46, + 271, + 287, + 427 + ], + "type": "text", + "content": " more triangles in Table 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "text", + "content": "Our method effectively achieves the optimal tradeoff between rendering quality and disk storage cost. For comparison, we report two tables for rendering quality and disk storage. Table 1 reports the rendering quality using the standard PSNR, SSIM [53], and LPIPS [57] on mip-NeRF 360 dataset. Instant-NGP [39], and mip-NeRF 360 [2] are volumetric rendering approaches that do not reconstruct explicit geometry, a topic that lies outside the primary focus of our study. In Table 2, we report the disk storage in megabytes and the number of vertices and faces. For a fair comparison, all the mesh files are saved in .obj and .mtl formats. Texture images are stored as .png files. Other assets, such as view-dependency MLP, are saved as .json files, except BakedSDF, which only provides gLTF format files. Our method demonstrates comparable rendering quality in outdoor scenes (with a 0.3 PSNR drop) and indoor scenes (with a 0.1 PSNR increase) while saving " + }, + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "inline_equation", + "content": "73 \\times" + }, + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "text", + "content": " triangles and " + }, + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "inline_equation", + "content": "11 \\times" + }, + { + "bbox": [ + 46, + 438, + 287, + 714 + ], + "type": "text", + "content": " storage compared to the state-of-the-art BakedSDF. In comparison to NeRF2Mesh and MobileNeRF, our method shows better-rendering quality (Table 1), demands less disk storage, and fewer triangles (Table 2), and captures superior mesh quality (Fig. 4). In Table 5, we measure the rendering speed and GPU memory by running in-browser on a Mac" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 270, + 544, + 324 + ], + "blocks": [ + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "text", + "content": "Table 2. Disk Storage (MB), Number of Vertices (V) and Faces (F). The unit for the number of vertices and faces is " + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "inline_equation", + "content": "{10}^{3}" + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "text", + "content": " . \"M\" denotes the " + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "inline_equation", + "content": "{10}^{6}" + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "text", + "content": " unit. We save " + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "inline_equation", + "content": "{73} \\times" + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "text", + "content": " the number of triangles and reduce " + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "inline_equation", + "content": "{11} \\times" + }, + { + "bbox": [ + 46, + 242, + 544, + 264 + ], + "type": "text", + "content": " disk storage." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 270, + 544, + 324 + ], + "lines": [ + { + "bbox": [ + 307, + 270, + 544, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 270, + 544, + 324 + ], + "type": "table", + "html": "
MethodMeshDeep Blending [18]Tanks & Temples [25]
DrJohnsonPlayroomBarnCourthouse
Instant-NGP [39]X27.7 / 0.84 / 0.3819.5 / 0.78 / 0.4623.4 / 0.72 / 0.4320.0 / 0.68 / 0.53
mip-NeRF360 [2]X29.1 / 0.90 / 0.2429.6 / 0.90 / 0.2527.2 / 0.81 / 0.3221.3 / 0.72 / 0.47
MobileNeRF [5]O25.9 / 0.78 / 0.3128.7 / 0.85 / 0.2824.4 / 0.77 / 0.2417.7 / 0.54 / 0.36
NeRF2Mesh [50]O24.0 / 0.75 / 0.5126.6 / 0.83 / 0.4123.4 / 0.72 / 0.4318.2 / 0.63 / 0.56
OursO26.5 / 0.82 / 0.4029.1 / 0.87 / 0.3226.5 / 0.83 / 0.2420.2 / 0.73 / 0.34
", + "image_path": "6fe53221e492fba1e06bfbabe8fe5a87c697f4792b9c16690227c3a35d5269c1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 354, + 544, + 392 + ], + "blocks": [ + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "text", + "content": "Table 3. Quantitative Comparison. We report PSNR " + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "text", + "content": ", SSIM " + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "text", + "content": ", and LPIPS " + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 306, + 328, + 544, + 350 + ], + "type": "text", + "content": " on Deep Blending and Tanks & Temples datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 354, + 544, + 392 + ], + "lines": [ + { + "bbox": [ + 308, + 354, + 544, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 544, + 392 + ], + "type": "table", + "html": "
MethodDrJohnsonPlayroomBarnCourthouse
Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓Disk ↓V ↓ / F ↓
MobileNeRF [5]174.9602 / 285281.81097 / 521289.21080 / 502263.6934 / 417
NeRF2Mesh [50]108.6414 / 815118.7401 / 793131.9397 / 793145.7451 / 843
Ours26.746 / 9260.9203 / 4014086 / 17147.497 / 192
", + "image_path": "ed85c2771f06179a32e63d57bd468d3f311bd1c6d86053b9edb4c6d4aeb3cc3b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 407, + 533, + 431 + ], + "blocks": [ + { + "bbox": [ + 320, + 397, + 531, + 407 + ], + "lines": [ + { + "bbox": [ + 320, + 397, + 531, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 397, + 531, + 407 + ], + "type": "text", + "content": "Table 4. Disk Storage (MB), Vertices (V) and Faces (F)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 407, + 533, + 431 + ], + "lines": [ + { + "bbox": [ + 318, + 407, + 533, + 431 + ], + "spans": [ + { + "bbox": [ + 318, + 407, + 533, + 431 + ], + "type": "table", + "html": "
FPS / GPU (MB)MobileNeRFNeRF2MeshOurs
MacBookPro (1920x1080 Resolution)59 / 324205 / 130212 / 115
", + "image_path": "a4ba1fbce16581ada433b0a776d6721c8a27c129f08e084f09c5d9750776a24e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 434, + 533, + 445 + ], + "lines": [ + { + "bbox": [ + 317, + 434, + 533, + 445 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 533, + 445 + ], + "type": "text", + "content": "Table 5. FPS and GPU memory on mip-NeRF 360 dataset" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 450, + 394, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 450, + 394, + 461 + ], + "spans": [ + { + "bbox": [ + 306, + 450, + 394, + 461 + ], + "type": "text", + "content": "BookPro (2020, M1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 593 + ], + "type": "text", + "content": "Figure 5 compares our method to the state-of-the-art baselines that focus on rendering quality: NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. In their first stage, BakedSDF* represents the volumetric rendering results obtained from a surface-based volumetric representation. The visual results of our approach exhibit sharper details than other methods. In the Bicycle and Room scenes, the mesh surface of the armrest for a bench and tree leaves exhibits noticeable jiggling, resulting in distorted images. Unlike BakedSDF*, our method effectively manages the thin structure of the wheel." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Experimental Results on Deep Blending and Tank & Temples Datasets: We train and evaluate our method on four scenes: Playroom and DrJohnson from the Deep Blending dataset, and Barn and Courthouse from the Tanks & Temples dataset. The former includes bounded indoor scenes, while the latter comprises large unbounded outdoor scenes. Their capture methodology differs from that of the mip-NeRF 360 dataset, and they do not utilize 360-degree scenes. In Tanks & Temples, we exclude sky regions during training and evaluation, as mesh-based rendering tends to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5059" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 160, + 133 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 160, + 133 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 160, + 133 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 160, + 133 + ], + "type": "image", + "image_path": "68e53eb087640f075447d91d36fa3ffe44c4f15000be93b56aa2d8c9a6702af7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 134, + 159, + 194 + ], + "blocks": [ + { + "bbox": [ + 59, + 134, + 159, + 194 + ], + "lines": [ + { + "bbox": [ + 59, + 134, + 159, + 194 + ], + "spans": [ + { + "bbox": [ + 59, + 134, + 159, + 194 + ], + "type": "image", + "image_path": "c37e774ade1e6567bc4c09d3d4ba07c869e3ece3fbde836d99caf0eb6cff20d7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 196, + 159, + 259 + ], + "blocks": [ + { + "bbox": [ + 59, + 196, + 159, + 259 + ], + "lines": [ + { + "bbox": [ + 59, + 196, + 159, + 259 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 159, + 259 + ], + "type": "image", + "image_path": "5d4394e26a7d4db72ef6404a38ce3006407e2bd58f69f92fed061946b0922e8b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 59, + 259, + 159, + 321 + ], + "blocks": [ + { + "bbox": [ + 59, + 259, + 159, + 321 + ], + "lines": [ + { + "bbox": [ + 59, + 259, + 159, + 321 + ], + "spans": [ + { + "bbox": [ + 59, + 259, + 159, + 321 + ], + "type": "image", + "image_path": "6ec0cbd3339e840ef314752c05fd88c7e50637f40fa36f7f8e8347fa1dd6d9bf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 323, + 123, + 331 + ], + "lines": [ + { + "bbox": [ + 102, + 323, + 123, + 331 + ], + "spans": [ + { + "bbox": [ + 102, + 323, + 123, + 331 + ], + "type": "text", + "content": "Scene" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 252, + 133 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 252, + 133 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 252, + 133 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 252, + 133 + ], + "type": "image", + "image_path": "c57afcff8883b762d5559ca7da593a550813c2bbf9aec92c4b9f02afba764775.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 160, + 134, + 252, + 194 + ], + "blocks": [ + { + "bbox": [ + 160, + 134, + 252, + 194 + ], + "lines": [ + { + "bbox": [ + 160, + 134, + 252, + 194 + ], + "spans": [ + { + "bbox": [ + 160, + 134, + 252, + 194 + ], + "type": "image", + "image_path": "e2220ff23b0d996e39d3cd1c7aa77283b0724e2faa2b51024de198f972a6d7ca.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 160, + 196, + 252, + 259 + ], + "blocks": [ + { + "bbox": [ + 160, + 196, + 252, + 259 + ], + "lines": [ + { + "bbox": [ + 160, + 196, + 252, + 259 + ], + "spans": [ + { + "bbox": [ + 160, + 196, + 252, + 259 + ], + "type": "image", + "image_path": "a85d1b7f3df4a4a87659e07c5b000bb4de585d65159b3f53ff8214d2d6bafaa5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 160, + 259, + 252, + 321 + ], + "blocks": [ + { + "bbox": [ + 160, + 259, + 252, + 321 + ], + "lines": [ + { + "bbox": [ + 160, + 259, + 252, + 321 + ], + "spans": [ + { + "bbox": [ + 160, + 259, + 252, + 321 + ], + "type": "image", + "image_path": "45e7086bb9bb53aa17e139cdf10ae2c7ea4c4f5e85184c1c25544b1c96b834c6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 323, + 220, + 331 + ], + "lines": [ + { + "bbox": [ + 193, + 323, + 220, + 331 + ], + "spans": [ + { + "bbox": [ + 193, + 323, + 220, + 331 + ], + "type": "text", + "content": "(a) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 253, + 72, + 346, + 133 + ], + "blocks": [ + { + "bbox": [ + 253, + 72, + 346, + 133 + ], + "lines": [ + { + "bbox": [ + 253, + 72, + 346, + 133 + ], + "spans": [ + { + "bbox": [ + 253, + 72, + 346, + 133 + ], + "type": "image", + "image_path": "c9a23a8610c7b0be0ae0756d74821d9204bdd461bc754843a2ebf8da15dccfd1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 253, + 134, + 346, + 194 + ], + "blocks": [ + { + "bbox": [ + 253, + 134, + 346, + 194 + ], + "lines": [ + { + "bbox": [ + 253, + 134, + 346, + 194 + ], + "spans": [ + { + "bbox": [ + 253, + 134, + 346, + 194 + ], + "type": "image", + "image_path": "348dd17a52fee2f3d2d563e6f5fab39e675bd93efe0139e7a2aba90d1357aaf9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 253, + 196, + 346, + 259 + ], + "blocks": [ + { + "bbox": [ + 253, + 196, + 346, + 259 + ], + "lines": [ + { + "bbox": [ + 253, + 196, + 346, + 259 + ], + "spans": [ + { + "bbox": [ + 253, + 196, + 346, + 259 + ], + "type": "image", + "image_path": "4196afc8ff9354ad2d5d27b67851277f28f7446b08aa7e2b171620aa8ee74e27.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 253, + 259, + 346, + 321 + ], + "blocks": [ + { + "bbox": [ + 253, + 259, + 346, + 321 + ], + "lines": [ + { + "bbox": [ + 253, + 259, + 346, + 321 + ], + "spans": [ + { + "bbox": [ + 253, + 259, + 346, + 321 + ], + "type": "image", + "image_path": "d99d8a57c37905c7efe16961007c2322468200aa98a8834c931362db66460316.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 275, + 323, + 324, + 331 + ], + "lines": [ + { + "bbox": [ + 275, + 323, + 324, + 331 + ], + "spans": [ + { + "bbox": [ + 275, + 323, + 324, + 331 + ], + "type": "text", + "content": "(b) NeRF2Mesh" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 347, + 72, + 440, + 133 + ], + "blocks": [ + { + "bbox": [ + 347, + 72, + 440, + 133 + ], + "lines": [ + { + "bbox": [ + 347, + 72, + 440, + 133 + ], + "spans": [ + { + "bbox": [ + 347, + 72, + 440, + 133 + ], + "type": "image", + "image_path": "00673299fbc26f822afc3ae151745c998281b7f9b02d583d82238d9bffddcbeb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 347, + 134, + 439, + 194 + ], + "blocks": [ + { + "bbox": [ + 347, + 134, + 439, + 194 + ], + "lines": [ + { + "bbox": [ + 347, + 134, + 439, + 194 + ], + "spans": [ + { + "bbox": [ + 347, + 134, + 439, + 194 + ], + "type": "image", + "image_path": "a866ad403d194ae5594412e83921d3960e82268ccd6c5c37bc5c7fa52cd6b186.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 347, + 196, + 439, + 321 + ], + "blocks": [ + { + "bbox": [ + 347, + 196, + 439, + 321 + ], + "lines": [ + { + "bbox": [ + 347, + 196, + 439, + 321 + ], + "spans": [ + { + "bbox": [ + 347, + 196, + 439, + 321 + ], + "type": "image", + "image_path": "9f698c865df6068129f7b1696be7df410383f488c0e220424b6691895edcb6f0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 323, + 416, + 331 + ], + "lines": [ + { + "bbox": [ + 370, + 323, + 416, + 331 + ], + "spans": [ + { + "bbox": [ + 370, + 323, + 416, + 331 + ], + "type": "text", + "content": "(c)BakedSDF\\*" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 72, + 533, + 133 + ], + "blocks": [ + { + "bbox": [ + 441, + 72, + 533, + 133 + ], + "lines": [ + { + "bbox": [ + 441, + 72, + 533, + 133 + ], + "spans": [ + { + "bbox": [ + 441, + 72, + 533, + 133 + ], + "type": "image", + "image_path": "6b38533c70d678c3ec03b6be9c1c29a81861d429f4d2376d33256449331c4ddc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 441, + 134, + 533, + 194 + ], + "blocks": [ + { + "bbox": [ + 441, + 134, + 533, + 194 + ], + "lines": [ + { + "bbox": [ + 441, + 134, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 441, + 134, + 533, + 194 + ], + "type": "image", + "image_path": "2ab2eb16d18c82c0847b7795b7d411707f8917be93c33303c06933fd50d3e297.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 441, + 196, + 533, + 258 + ], + "blocks": [ + { + "bbox": [ + 441, + 196, + 533, + 258 + ], + "lines": [ + { + "bbox": [ + 441, + 196, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 441, + 196, + 533, + 258 + ], + "type": "image", + "image_path": "ab99f518d41a6e923006684fd031a7e7c87e3c964524db63795418d03e0652b0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 441, + 259, + 534, + 321 + ], + "blocks": [ + { + "bbox": [ + 441, + 259, + 534, + 321 + ], + "lines": [ + { + "bbox": [ + 441, + 259, + 534, + 321 + ], + "spans": [ + { + "bbox": [ + 441, + 259, + 534, + 321 + ], + "type": "image", + "image_path": "86534fee9dbce4b104e47c63b6b3bd7de8024f7aa5966bd39095ba914e4f739a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 323, + 508, + 331 + ], + "lines": [ + { + "bbox": [ + 459, + 323, + 508, + 331 + ], + "spans": [ + { + "bbox": [ + 459, + 323, + 508, + 331 + ], + "type": "text", + "content": "(d) MobileNeRF" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 50, + 373, + 107, + 408 + ], + "blocks": [ + { + "bbox": [ + 50, + 373, + 107, + 408 + ], + "lines": [ + { + "bbox": [ + 50, + 373, + 107, + 408 + ], + "spans": [ + { + "bbox": [ + 50, + 373, + 107, + 408 + ], + "type": "image", + "image_path": "107430fdf18efb43a11e33f7955d6f1d6e636308269803ff535001c3b91204ce.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 50, + 410, + 107, + 445 + ], + "blocks": [ + { + "bbox": [ + 50, + 410, + 107, + 445 + ], + "lines": [ + { + "bbox": [ + 50, + 410, + 107, + 445 + ], + "spans": [ + { + "bbox": [ + 50, + 410, + 107, + 445 + ], + "type": "image", + "image_path": "6cc9d40c8ee9823b8491755349163a077e2f7a87b31894fe927abd6167af91b1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 50, + 446, + 107, + 499 + ], + "blocks": [ + { + "bbox": [ + 50, + 446, + 107, + 499 + ], + "lines": [ + { + "bbox": [ + 50, + 446, + 107, + 499 + ], + "spans": [ + { + "bbox": [ + 50, + 446, + 107, + 499 + ], + "type": "image", + "image_path": "3c67fafc6044e4b38434f0d460526a49fa29f4be110fe71f22e5544a63ad36cd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 500, + 97, + 506 + ], + "lines": [ + { + "bbox": [ + 58, + 500, + 97, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 500, + 97, + 506 + ], + "type": "text", + "content": "(a) Ours w/o Opt." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 507, + 285, + 533 + ], + "lines": [ + { + "bbox": [ + 47, + 507, + 285, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 285, + 533 + ], + "type": "text", + "content": "Figure 6. Zoom-in Shading Mesh to see more geometric details. encounter challenges with far-away sky regions." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 108, + 373, + 165, + 408 + ], + "blocks": [ + { + "bbox": [ + 108, + 373, + 165, + 408 + ], + "lines": [ + { + "bbox": [ + 108, + 373, + 165, + 408 + ], + "spans": [ + { + "bbox": [ + 108, + 373, + 165, + 408 + ], + "type": "image", + "image_path": "c292a89b0497768ccea656f80a743fe2d11a23280f63b2064d1b2eb7c6c75af4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 108, + 410, + 166, + 445 + ], + "blocks": [ + { + "bbox": [ + 108, + 410, + 166, + 445 + ], + "lines": [ + { + "bbox": [ + 108, + 410, + 166, + 445 + ], + "spans": [ + { + "bbox": [ + 108, + 410, + 166, + 445 + ], + "type": "image", + "image_path": "c7d8dd9d1c052a2468a72c587665454adacf0f6db51b06f7ac74544f0bc77e3f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 108, + 446, + 166, + 499 + ], + "blocks": [ + { + "bbox": [ + 108, + 446, + 166, + 499 + ], + "lines": [ + { + "bbox": [ + 108, + 446, + 166, + 499 + ], + "spans": [ + { + "bbox": [ + 108, + 446, + 166, + 499 + ], + "type": "image", + "image_path": "c3b2051cacf858e78dae5891bfd83d9a6cb7eb77b772d8b9dcc3d392596225cd.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 500, + 146, + 505 + ], + "lines": [ + { + "bbox": [ + 126, + 500, + 146, + 505 + ], + "spans": [ + { + "bbox": [ + 126, + 500, + 146, + 505 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 168, + 373, + 225, + 408 + ], + "blocks": [ + { + "bbox": [ + 168, + 373, + 225, + 408 + ], + "lines": [ + { + "bbox": [ + 168, + 373, + 225, + 408 + ], + "spans": [ + { + "bbox": [ + 168, + 373, + 225, + 408 + ], + "type": "image", + "image_path": "f6cb5bcb7d3fc2c017ec92a8c7f8e833a27147eb8e945bb6c47e6424b636ff28.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 168, + 410, + 225, + 445 + ], + "blocks": [ + { + "bbox": [ + 168, + 410, + 225, + 445 + ], + "lines": [ + { + "bbox": [ + 168, + 410, + 225, + 445 + ], + "spans": [ + { + "bbox": [ + 168, + 410, + 225, + 445 + ], + "type": "image", + "image_path": "4553bd33dc7fb5004c38a0c012cecb47d36eb90ffcb1e90b57f26639e9ce0eea.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 168, + 446, + 225, + 498 + ], + "blocks": [ + { + "bbox": [ + 168, + 446, + 225, + 498 + ], + "lines": [ + { + "bbox": [ + 168, + 446, + 225, + 498 + ], + "spans": [ + { + "bbox": [ + 168, + 446, + 225, + 498 + ], + "type": "image", + "image_path": "893f2fddaaca3a5105cc1ff06217f8d8fbfd1c53a8d9389b44fef24b4ddbc831.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 499, + 216, + 505 + ], + "lines": [ + { + "bbox": [ + 179, + 499, + 216, + 505 + ], + "spans": [ + { + "bbox": [ + 179, + 499, + 216, + 505 + ], + "type": "text", + "content": "(c) NeRF2Mesh" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 227, + 373, + 285, + 408 + ], + "blocks": [ + { + "bbox": [ + 227, + 373, + 285, + 408 + ], + "lines": [ + { + "bbox": [ + 227, + 373, + 285, + 408 + ], + "spans": [ + { + "bbox": [ + 227, + 373, + 285, + 408 + ], + "type": "image", + "image_path": "c871f7b71b6899b394c1467c9103941aba05c65f4d46b604f240b35c33cf8595.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 227, + 410, + 285, + 445 + ], + "blocks": [ + { + "bbox": [ + 227, + 410, + 285, + 445 + ], + "lines": [ + { + "bbox": [ + 227, + 410, + 285, + 445 + ], + "spans": [ + { + "bbox": [ + 227, + 410, + 285, + 445 + ], + "type": "image", + "image_path": "f465ad3f474a6dbaa440e66b4f8bd8fda04f9d233c3bb32a124bea6509e6215f.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 227, + 446, + 285, + 498 + ], + "blocks": [ + { + "bbox": [ + 227, + 446, + 285, + 498 + ], + "lines": [ + { + "bbox": [ + 227, + 446, + 285, + 498 + ], + "spans": [ + { + "bbox": [ + 227, + 446, + 285, + 498 + ], + "type": "image", + "image_path": "84f2f4a507bcf6e226531e71edf87725319108c1d885124f6636fa7087602421.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 499, + 274, + 505 + ], + "lines": [ + { + "bbox": [ + 240, + 499, + 274, + 505 + ], + "spans": [ + { + "bbox": [ + 240, + 499, + 274, + 505 + ], + "type": "text", + "content": "(d) BakedSDF" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "bbox": [ + 46, + 534, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 712 + ], + "type": "text", + "content": "Ablation: In Fig. 6, we visualize the zoom-in shading mesh extracted without texture. \"Ours w/o Opt\" denotes we did not execute our optimization stage in Section 3.3 after mesh processing. The optimization stage enables us to represent the geometric details of garden vases and bicycle wheels. Our method can represent more detailed structures than BakedSDF and a smoother surface than NeRF2Mesh. Moreover, it is noteworthy that when employing rasterization-based rendering with the initial color model without the optimization stage, the rendering quality is poor, as shown in Table 1. In this table, \"Ours w/o Geo\" only trains appearance model " + }, + { + "bbox": [ + 46, + 534, + 287, + 712 + ], + "type": "inline_equation", + "content": "f_{color}" + }, + { + "bbox": [ + 46, + 534, + 287, + 712 + ], + "type": "text", + "content": " without training vertex deformation. Training vertex deformation yields a 1.7 gain in indoor scenes by refining the mesh surface. \"Ours w. RADAM\" denotes that we use the rotation-invariant" + } + ] + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 306, + 372, + 426, + 406 + ], + "blocks": [ + { + "bbox": [ + 46, + 334, + 547, + 367 + ], + "lines": [ + { + "bbox": [ + 46, + 334, + 547, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 547, + 367 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparisons with Existing Methods. We visually compare our proposed technique with NeRF2Mesh [50], BakedSDF* [55], and MobileNeRF [5]. The visual results of BakedSDF* generated in its first stage through volumetric rendering. Red and blue dotted lines emphasize subtle differences in rendering quality." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 372, + 426, + 406 + ], + "lines": [ + { + "bbox": [ + 306, + 372, + 426, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 426, + 406 + ], + "type": "image", + "image_path": "c4d21dcc549a2abcbefd5f0aff8296d2d7125c2fc3a01e1d411bf9d265a33197.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 411, + 485, + 420 + ], + "lines": [ + { + "bbox": [ + 367, + 411, + 485, + 420 + ], + "spans": [ + { + "bbox": [ + 367, + 411, + 485, + 420 + ], + "type": "text", + "content": "Figure 7. Naive QSlim Results" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 427, + 372, + 545, + 406 + ], + "blocks": [ + { + "bbox": [ + 427, + 372, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 427, + 372, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 427, + 372, + 545, + 406 + ], + "type": "image", + "image_path": "fbd3150b7787c1382e9040a09771cdfb6e24a77d07f96c5d90075822678598db.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 306, + 420, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 545, + 491 + ], + "type": "text", + "content": "ADAM (RADAM) for updating mesh vertices instead of a naive ADAM. However, it provides benefits for preventing mesh distortion. In Fig. 7, we show the visualization results of naive QSlim [12]. The naive QSlim leads to shrinkage and degeneration during the edge collapse process, and the boundary is over-smoothed compared to our method." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 306, + 502, + 538, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 538, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 538, + 514 + ], + "type": "text", + "content": "5. Conclusion, Limitations, and Future Work" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 712 + ], + "type": "text", + "content": "This paper introduces a practical approach for reconstructing geometrically accurate meshes with photorealistic rendering quality. Our method achieves a compact mesh representation in unbounded large-scale scenes while preserving highly detailed geometry. Furthermore, the appearance field corresponding to this lightweight mesh is compatible with the classical rasterization pipeline, enabling efficient rendering. However, our method has limitations. BakedSDF employs a well-designed appearance model based on spherical Gaussians. The potential for enhancing rendering quality exists by incorporating a more sophisticated appearance model. Moreover, the gradient-based mesh optimization in the final stage may still result in mesh distortions, including self-intersections. We plan to employ a carefully designed optimizer to mitigate such distortions in future work." + } + ] + } + ], + "index": 48 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5060" + } + ] + } + ], + "index": 49 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 157, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 157, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 157, + 288, + 213 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 255 + ], + "type": "text", + "content": "[3] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[4] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 378 + ], + "type": "text", + "content": "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 445 + ], + "type": "text", + "content": "[6] Jaehoon Choi, Dongki Jung, Taejae Lee, Sangwook Kim, Youngdong Jung, Dinesh Manocha, and Donghwan Lee. Tmo: Textured mesh acquisition of objects with a mobile device by using differentiable rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16674-16684, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "text", + "content": "[7] Jonathan Cohen, Marc Olano, and Dinesh Manocha. Appearance-preserving simplification. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 115-122, 1998. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 534 + ], + "type": "text", + "content": "[8] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a vsi system for high performance graphics. Acm siggraph computer graphics, 22(4):21-30, 1988. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 287, + 578 + ], + "type": "text", + "content": "[9] Carl Erikson and Dinesh Manocha. Gaps: General and automatic polygonal simplification. In Proceedings of the 1999 symposium on Interactive 3D graphics, pages 79-88, 1999. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 689 + ], + "type": "text", + "content": "[11] Jun Gao, Wenzheng Chen, Tommy Xiang, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Learning deformable tetrahedral meshes for 3d reconstruction. Advances In Neural Information Processing Systems, 33:9936-9947, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Michael Garland. Quadric-based polygonal surface simplification. Carnegie Mellon University, 1999. 2, 4, 5, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Michael Garland and Paul S Heckbert. Surface simplification using quadric error metrics. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 209-216, 1997. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[14] Michael Garland and Paul S Heckbert. Simplifying surfaces with color and texture using quadric error metrics. In Proceedings Visualization'98 (Cat. No. 98CB36276), pages 263-269. IEEE, 1998. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[15] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In Proceedings of the 37th International Conference on Machine Learning, ICML'20. JMLR.org, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[16] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, light, and material decomposition from images using monte carlo rendering and denoising. Advances in Neural Information Processing Systems, 35:22856-22869, 2022. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[17] Jon Hasselgren, Nikolai Hofmann, and Jacob Munkberg. Shape, Light, and Material Decomposition from Images using Monte Carlo Rendering and Denoising. arXiv:2206.03380, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 348 + ], + "type": "text", + "content": "[18] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[19] Peter Hedman, Pratul P Srinivasan, Ben Mildenhall, Jonathan T Barron, and Paul Debevec. Baking neural radiance fields for real-time view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5875-5884, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "text", + "content": "[20] Hugues Hoppe. View-dependent refinement of progressive meshes. In Proceedings of the 24th annual conference on Computer graphics and interactive techniques, pages 189-198, 1997. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[21] Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer, 2022. https://mitsuba-renderer.org. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 536 + ], + "type": "text", + "content": "[22] Justin Johnson, Nikhila Ravi, Jeremy Reizenstein, David Novotny, Shubham Tulsiani, Christoph Lassner, and Steve Branson. Accelerating 3d deep learning with pytorch3d. In SIGGRAPH Asia 2020 Courses, pages 1-1. 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 579 + ], + "type": "text", + "content": "[23] Hiroharu Kato, Deniz Beker, Mihai Morariu, Takahiro Ando, Toru Matsuoka, Wadim Kehl, and Adrien Gaidon. Differentiable rendering: A survey. arXiv preprint arXiv:2006.12057, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 581, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 612 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[25] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[26] Leif Kobbelt. *Sqrt(3)-subdivision*. In *Proceedings of the 27th annual conference on Computer graphics and interactive techniques*, pages 103–112, 2000. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[27] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5061" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "high-performance differentiable rendering. ACM Transactions on Graphics (TOG), 39(6):1-14, 2020. 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[28] Yiyi Liao, Simon Donne, and Andreas Geiger. Deep marching cubes: Learning explicit surface representations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2916-2925, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[29] Peter Lindstrom. Out-of-core simplification of large polygonal models. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 259-262, 2000. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 217 + ], + "type": "text", + "content": "[30] Peter Lindstrom and Claudio T Silva. A memory insensitive technique for large model simplification. In Proceedings Visualization, 2001. VIS'01., pages 121-550. IEEE, 2001. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[31] Selena Zihan Ling, Nicholas Sharp, and Alec Jacobson. Vectoradam for rotation equivariant geometry optimization. Advances in Neural Information Processing Systems, 35:4111-4122, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 316 + ], + "type": "text", + "content": "[32] Jeffrey Yunfan Liu, Yun Chen, Ze Yang, Jingkang Wang, Sivabalan Manivasagam, and Raquel Urtasun. Real-time neural rasterization for large scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8416-8427, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 287, + 360 + ], + "type": "text", + "content": "[33] Shichen Liu, Tianye Li, Weikai Chen, and Hao Li. Soft rasterizer: A differentiable renderer for image-based 3d reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7708-7717, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 393 + ], + "type": "text", + "content": "[34] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163-169, 1987. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 287, + 448 + ], + "type": "text", + "content": "[35] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "text", + "content": "[36] Fujun Luan, Shuang Zhao, Kavita Bala, and Zhao Dong. Unified shape and svbrdf recovery using differentiable monte carlo rendering. In Computer Graphics Forum, volume 40, pages 101-113. Wiley Online Library, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 525 + ], + "type": "text", + "content": "[37] Nelson Max. Optical models for direct volume rendering. IEEE Transactions on Visualization and Computer Graphics, 1(2):99-108, 1995. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "type": "text", + "content": "[38] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[39] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[40] Andrew Nealen, Takeo Igarashi, Olga Sorkine, and Marc Alexa. Laplacian mesh optimization. In Proceedings of the 4th international conference on Computer graphics and interactive techniques in Australasia and Southeast Asia, pages 381-389, 2006. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[41] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large steps in inverse rendering of geometry. ACM Transactions on Graphics (TOG), 40(6):1-13, 2021. 2, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[42] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 162 + ], + "type": "text", + "content": "[43] Werner Palfinger. Continuous remeshing for inverse rendering. Computer Animation and Virtual Worlds, 33(5):e2101, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[44] Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo, Michael Niemeyer, Abhijit Kundu, and Federico Tombari. Nerfmeshing: Distilling neural radiance fields into geometrically-accurate 3d meshes. In Proc. of the International Conf. on 3D Vision (3DV), 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[46] Jarek Rossignac and Paul Borrel. Multi-resolution 3d approximations for rendering complex scenes. In Modeling in computer graphics: methods and applications, pages 455-465. Springer, 1993. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 365 + ], + "type": "text", + "content": "[47] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 419 + ], + "type": "text", + "content": "[48] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep marching tetrahedra: a hybrid representation for high-resolution 3d shape synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 476 + ], + "type": "text", + "content": "[49] Tianchang Shen, Jacob Munkberg, Jon Hasselgren, Kangxue Yin, Zian Wang, Wenzheng Chen, Zan Gojcic, Sanja Fidler, Nicholas Sharp, and Jun Gao. Flexible isosurface extraction for gradient-based mesh optimization. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 531 + ], + "type": "text", + "content": "[50] Jiaxiang Tang, Hang Zhou, Xiaokang Chen, Tianshu Hu, Errui Ding, Jingdong Wang, and Gang Zeng. Delicate textured mesh recovery from nerf via adaptive surface refinement. arXiv preprint arXiv:2303.02091, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 578 + ], + "type": "text", + "content": "[51] Chao Wang and Xiaohu Guo. Plane-based optimization of geometry and texture for rgb-d reconstruction of indoor scenes. In 2018 International Conference on 3D Vision (3DV), pages 533-541. IEEE, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 579, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 644 + ], + "type": "text", + "content": "[52] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[53] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[54] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5062" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 239 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 94 + ], + "type": "text", + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 140 + ], + "type": "text", + "content": "[55] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. arXiv, 2023. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 184 + ], + "type": "text", + "content": "[56] Zehao Yu, Anpei Chen, Bozidar Antic, Songyou Peng, Apratim Bhattacharyya, Michael Niemeyer, Siyu Tang, Torsten Sattler, and Andreas Geiger. Sdfstudio: A unified framework for surface reconstruction, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 239 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5063" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_content_list.json b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..16a04e3488bc11ef4d92d0611267921909b78289 --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_content_list.json @@ -0,0 +1,1851 @@ +[ + { + "type": "text", + "text": "LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images", + "text_level": 1, + "bbox": [ + 104, + 130, + 866, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Zhang $^{1,*}$ , Irving Fang $^{1,*}$ , Hao Wu $^{1}$ , Akshit Kaushik $^{1}$ , Alice Rodriguez $^{1}$ , Hanwen Zhao $^{1}$ , Juexiao Zhang $^{1}$ , Zhuo Zheng $^{2}$ , Radu Iovita $^{1,\\dagger}$ , Chen Feng $^{1,\\dagger}$ , New York University", + "bbox": [ + 122, + 179, + 844, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ai4ce.github.io/LUWA/", + "bbox": [ + 351, + 237, + 611, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/90bf1edf98efabb4f624952e97e9e1986bcba9feb09c9fa497ad88f2b266a309.jpg", + "image_caption": [ + "Figure 1. LUWA poses a unique computer vision challenge due to: its complex wear formation and irregular wear patterns, ambiguous sensing modalities and magnifications in microscopic imaging. Facing these challenges, the LUWA dataset encompasses both texture and heightmap with different magnifications, encouraging the exploration of image classification beyond common objects." + ], + "image_footnote": [], + "bbox": [ + 102, + 260, + 251, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c5d082b4abe8c0afaaaa5096928c76addb5ef8128cad1dfe18ed7d44bf574a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 260, + 434, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24e4e8f23090d87e9f0549de6e4464e21a3c8a9c1946e0a7813110ffa5d1b639.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 257, + 864, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f254019f55b495f1d9712d1e46158be6ca39c86cec115693282a96774ad4ab9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 435, + 864, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 604, + 313, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lithic Use-Wear Analysis (LUWA) using microscopic images is an underexplored vision-for-science research area. It seeks to distinguish the worked material, which is critical for understanding archaeological artifacts, material interactions, tool functionalities, and dental records. However, this challenging task goes beyond the well-studied image classification problem for common objects. It is affected by many confounders owing to the complex wear mechanism and microscopic imaging, which makes it difficult even for human experts to identify the worked material successfully. In this paper, we investigate the following three questions on this unique vision task for the first time:(i) How well can state-of-the-art pre-trained models (like DINOv2) generalize to the rarely seen domain? (ii) How can few-shot", + "bbox": [ + 73, + 638, + 470, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning be exploited for scarce microscopic images? (iii) How do the ambiguous magnification and sensing modality influence the classification accuracy? To study these, we collaborated with archaeologists and built the first open-source and the largest LUWA dataset containing 23,130 microscopic images with different magnifications and sensing modalities. Extensive experiments show that existing pretrained models notably outperform human experts but still leave a large gap for improvements. Most importantly, the LUWA dataset provides an underexplored opportunity for vision and learning communities and complements existing image classification problems on common objects.", + "bbox": [ + 500, + 607, + 893, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 799, + 632, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lithic Use-Wear Analysis (LUWA) is a long-standing scientific problem (see Fig. 1) to identify the functions of stone tools by examining wear traces at the microscopic level on the tool's surface [30, 41]. It seeks to distinguish the worked material (like bone, wood, ivory and antler) us-", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 863, + 202, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author: Radu Iovita (iovita@nyu.edu) and Chen Feng (cfeng@nyu.edu)", + "bbox": [ + 78, + 876, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22563", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing microscopic images, creating a classification problem. Investigating this unanswered vision-for-science problem will provide invaluable insights for uncovering the hidden story of ancient tools [38] and advancing the understanding of material interactions [62, 66]. However, few studies have begun to explore advanced learning-based methods, and this field is still in a nascent stage [8, 45].", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides its potential scientific impact, learning-based LUWA on microscopic images also poses unique computer vision challenges beyond common objects. Unlike common objects with distinct boundaries and typical size, wear traces in microscopic images are irregular and discontinuous (see Fig. 1), making it difficult to define clear visual characteristics [40]. In particular, their absence of clear foreground and background might increase the difficulty of feature extraction. Furthermore, recorded wear traces are a hybrid consequence of the complex wear process and microscopic imaging, which is affected by many factors [41]. For example, the wear duration and the motion type that generates the wear can create significant intra-class variability. Another crucial aspect is the ambiguous magnification and sensing modality of microscopic images. High-sensitivity microscopic imaging allows for capturing complementary sensing modalities and zoom-in details at different magnifications, respectively (see Fig. 1). We found that their modality and magnification differences lead to very different visual features. However, due to the understudied characteristics of wear traces, there has been no conclusive answer to this vision-for-science problem about which magnification and sensing modality is better. In practice, the accessibility of expensive equipment also limits the flexible selection of captured data.", + "bbox": [ + 75, + 199, + 473, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in computer vision provide a good opportunity for this challenging scientific problem, especially the emerging paradigm shift with foundation models pretrained on large-scale data [2, 7]. These foundation models generate task-agnostic visual features and have shown excellent performance on image classification with common datasets like ImageNet- $\\{1k, A\\}$ [43]. But how much can state-of-the-art (SOTA) pre-trained representations benefit rarely seen domains not extensively covered by uncurated data available on the Internet? The lack of diversity in domain-specific datasets leaves uncertainty about this question, particularly when addressing real-world vision challenges. Unfortunately, existing studies on LUWA focus on private data and the research community faces a deficiency in accessible datasets. The complexity of wear forms, microscopic imaging, equipment limitations, and the need for expert interpretations hinders the collection of an appropriate dataset that can represent the variability of this domain.", + "bbox": [ + 75, + 580, + 470, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To explore these unique challenges and opportunities, we collaborated with anthropological archaeologists and built the first open-source and the largest LUWA dataset contain-", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing 23,130 microscopic images. Its major characteristics are summarized as follows: (i) Multi-scale wear patterns. Lower and higher magnifications allow for an observation of pattern distributions and topographical details, respectively, which increases the scale variations. (ii) Complementary sensing modalities. Both grayscale microscopic images and corresponding 3D surface profiles are available to provide complementary texture information and geometry cues, helping to identify discriminative features [21]. (iii) Both machine and human wear processes. Samples of stone tools were collected from both machine and human wear processes. The former is to isolate the effect of worked material by tightly controlling other factors [37] while the latter is to reflect the complexity of real scenes, especially for newly discovered categories without baseline data [44, 52]. We envision that the LUWA dataset will encourage researchers to develop and evaluate algorithms for image classification beyond common objects, as a precursor for downstream tasks like segmentation and detection [3].", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the LUWA dataset, we go further in answering the following three problems facing real-world applications: (i) How well representative classification models can generalize to the rarely seen domain? (ii) How can few-shot learning be exploited when scarce microscopic images are available, especially for newly discovered categories? (iii) How do the ambiguous magnification and sensing modality of microscopic images influence the classification accuracy? Our contributions are summarized as follows:", + "bbox": [ + 496, + 382, + 890, + 518 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce the first open-source and the largest LUWA dataset including 23,130 microscopic images with different magnifications and sensing modalities. We collaborated with archaeologists and set up a data collection pipeline considering the influence of wear formation, microscopic imaging, and expert knowledge. The LUWA dataset allows for reproducible investigations on this rarely seen domain and complements existing image classification problems on common objects.", + "- Facing the image classification problem beyond common objects, we benchmark the generalization capability of state-of-the-art vision models (ResNet, ViT, DINOv2, ConvNeXt) in this specific domain. Experimental results show that DINOv2 has the most stable performance amidst varying levels of granularity, magnification, and sensing modalities in the data. We also observed many trends regarding the impacts that features like magnification and sensing modality have on classification accuracy, and some of them are not consistent with experts' heuristics. In general, state-of-the-art computer vision models display super-human accuracy over domain experts.", + "- Considering the scarcity of microscopic images in real scenarios, we investigate the performance of few-shot image classification and reasoning on the LUWA dataset. Particularly, we collected prompts from archaeologists" + ], + "bbox": [ + 500, + 523, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22564", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and did case studies on whether GPT-4V(ison) can mimic the experts' reasoning process. Further explorations are required to improve the performance.", + "bbox": [ + 89, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 150, + 218, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lithic Use-Wear Analysis. Lithic use-wear analysis was originally developed in the 1950s and aims to answer the scientific problem of how to distinguish the worked material according to wear traces on the stone surface [30, 41, 55]. Low-power and High-power microscopy methods provide useful magnified polishes or micro-fractures via confocal microscopes, tactile profilometers, scanning electron microscopes, and even atomic force microscopes [20, 31, 58]. Existing studies focus on blind tests [4, 19, 39] and quantification methods [4, 5, 20, 28, 59, 60]. However, it remains an insufficiently developed research area due to complex wear patterns and the subjectivity of methods [47]. Existing blind tests demonstrate unreliable identification results on different worked materials. For example, correct identifications of plant and wood are $32.4\\%$ and $49.1\\%$ , respectively [19]. The identification of worked materials can be regarded as a unique vision problem, but learning-based algorithms are rarely explored in this research area [8]. The deficiency in accessible datasets also limits its research progress heavily [19]. To solve these, we present the first open-source LUwA dataset and investigate the capability of SOTA classification algorithms on this unique vision-for-science problem for the first time.", + "bbox": [ + 76, + 176, + 472, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Classification beyond Common Objects. Image classification is a fundamental vision task that categorizes a whole image as a specific label or class based on its visual contents. Outstanding performance can be achieved on the image classification task of common objects with representative frameworks like ConvNeXt using only public training data [67]. However, real-world scenarios, especially scientific applications, involve non-trivial objectives with different physical properties like tiny pollen grains [6], constituent materials [15, 18, 61], texture in the wild [12], and hyperspectral remote sensing images [50]. Facing uncommon objects, many studies focus on specialized models in respective fields. However, we seek to answer whether simpler pipelines can provide comparable performance. Complementing existing datasets, we introduce a new vision challenge to identify irregular and discontinuous wear traces at the microscopic level and build a classification benchmark to explore how far the practical performance of image classification on uncommon objects can be pushed utilizing existing datasets and architectures, especially foundation models.", + "bbox": [ + 76, + 523, + 472, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Microscopic Image Classification. In the context of using microscopic images for classification, a wide variety of applications can be found across different scientific disciplines. This includes the study of biological cells [11, 29],", + "bbox": [ + 76, + 839, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bacteria [65], tissue types [64], and material structures [23]. Each application presents unique challenges, particularly in terms of the high-level detail and focus required in the images. These challenges are often compounded by issues such as ambiguity in magnification and sensing modality [23, 68]. To alleviate these challenges, the LUWA Dataset presents a diverse configuration that covers different magnification levels and sensing modalities.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LUWA Dataset", + "text_level": 1, + "bbox": [ + 500, + 224, + 651, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the LUWA dataset creation process and provide basic statistics. To represent the variability of this domain, in Section 3.1, we present a data collection pipeline, which consists of four key aspects (see Fig. 1). Specifically, ① considering the complexity of wear formation, we introduced both machine and human wear experiments [36] to create stone samples; ② to enrich the dataset diversity and investigate the ambiguous magnification and sensing modality, we utilized an optical 3D profilometer with both $20 \\times$ and $50 \\times$ objective lenses to acquire high-quality texture and heightmap; ③ natural materials were selected according to existing blind tests in the literature [19], particularly including fine-grained categories of wood and plants. ④ domain-specific knowledge is twofold including the identification of wear degrees to increase the dataset diversity and expert interpretations for potential explorations on explainability and the application of vision language models. In Section 3.2, we summarized the LUWA dataset and analyzed its diversity in spatial distributions, magnifications, and sensing modalities.", + "bbox": [ + 496, + 250, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Creation", + "text_level": 1, + "bbox": [ + 500, + 560, + 666, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both Machine and Human Wear Processes. LUWA dataset contains stone samples from both machine and human wear processes. Key factors that affect wear results are material properties, mechanical factors, and environmental conditions [48]. To isolate the effect of worked materials, a tightly controlled protocol was used for machine wear experiments [36, 37]. We utilized a tribometer to simulate cutting actions so as to quantify the load applied to the material (a load of 20N), the type of movement (a straight back and forth motion with the speed of 35 repetitions per minute), and the worked duration (Oh, 1h, 3h, 5h, and 12h) (see Fig. 1). To limit the influence of material properties of stone samples, we chose the same flint (Baltic/morainic flint from Denmark) for all experiments [47, 51]. Considering the low classification accuracy of various plants in blind tests [19], we chose the plant cutting process as the human wear experiment.", + "bbox": [ + 496, + 583, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-Quality Microscopic Imaging. To capture the high-precision wear traces on stone samples, we utilized an optical 3D profilometer (S neox, Sensorfar Metrology) to collect data with a standardized and reproducible process (see", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22565", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0e7f10f1f12cf472b1951873d600cc83e0f9e841f8d403c0dea862c53fdf066c.jpg", + "image_caption": [ + "Figure 2. Image diversity of LUwA dataset and corresponding visual explanations for human and model decision-making processes. (i) LUwA dataset provides diverse microscopic images associated with spatial distributions (e.g. Regions 1 and 2), magnifications (e.g. Regions 2 and 4) and sensing modalities (texture in the first row and heightmap in the second row); (ii) We compared visual explanations in both human (in the third row) and model (in the fourth row) decision-making processes. Human experts labeled the most important region with red and the less important region with yellow when looking at details of microscopic images to distinguish the worked material. Similarly, Grad-CAM [54] heatmaps use red for the highest importance, yellow for lower importance, and blue for the lowest importance. Interestingly, similar areas (e.g. Regions 1, 4 and 6) are labeled with higher importance for both humans and models." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 225, + 363 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fbc0531bbf633635a3888c4310b691159cf89f577853140c8b40389b8015517f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 92, + 885, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 1). To test the influence of magnifications for microscopic image classification, both $20\\times$ and $50\\times$ objective lenses were chosen for measurements. Their spatial resolutions are 0.65 and $0.26\\mu \\mathrm{m}$ pixel, respectively. Furthermore, complementary grayscale images and corresponding 3D surface profiles are acquired via Sensormap. We applied a standard filtering protocol to extract the worn surface and alleviate the effect of natural flints' surface topography [9].", + "bbox": [ + 75, + 481, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Domain-Specific Expert Knowledge. Domain-specific knowledge is twofold: (i) human experts help to identify microscopic traces with different wear degrees, enriching the dataset diversity; (ii) for further investigations on explainability and the application of vision language models, human experts also labeled their attention maps when making decisions on worked material (see Fig 1) and provided classification prompt for GPT-4V [42] (see Fig. 6).", + "bbox": [ + 75, + 603, + 468, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Material Selection and Processing. To benchmark models in this field, we chose representative natural materials (see Fig. 1) according to blind test results in the literature [19] and included fine-grained categories on wood and plants in particular. For the further exploration of wear mechanisms, we analyzed their properties, including the hardness [47] and silicon content [22] in the supplementary material.", + "bbox": [ + 75, + 726, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Analysis", + "text_level": 1, + "bbox": [ + 76, + 845, + 240, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reflect the variability of this scientific domain, we built the first public and largest LUwA dataset containing 23,130", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e56b81e876b6bf122e70f79ba706093f68ae64bbc67a59323fdabfe965365600.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Stone SamplesMotion TypesWorked TimeMaterial Categories
34279
MagnificationsSensing Modalities
22
", + "bbox": [ + 503, + 479, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Key factors considered in the LUWA dataset that can reflect the complex wear formation and microscopic imaging.", + "bbox": [ + 500, + 535, + 890, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "microscopic images. Specifically, key factors of the complex wear formation and microscopic imaging are considered in the LUWA dataset. As shown in Tab. 1, we report (i) the number of microscopic images, (ii) the number of stone samples, motion types, worked time, and material classes which are exploited, (iii) the number of magnifications and sensing modalities LUWA dataset supports.", + "bbox": [ + 496, + 578, + 890, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image diversity of LUWA dataset brings challenges for algorithm robustness. It is associated with the spatial distribution of the region collected, the selection of the magnification, and the sensing modality. Greater distances between sampled areas typically result in more pronounced variations in their surface distributions (see Regions 1 and 2 in Fig. 2). Even collected from the same wear trace, the selection of the magnification also contributes to scale difference, which causes totally different wear patterns (see Regions 2 and 4 in Fig. 2). Moreover, LUWA dataset provides both the texture and heightmap, helping to identify discriminative features. We explore the semantic diversity of LUWA dataset on the magnification and sensing modality (see Fig. 2) [3]. We selected VGG [56], ResNet [26],", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22566", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6524fac1bc150096ee6ef536f5b3306fb3a3eea2077df3cc9c5444b6c24ceaca.jpg", + "image_caption": [ + "Figure 3. Cosine similarity distribution of LUWA dataset on different magnifications and sensing modalities." + ], + "image_footnote": [], + "bbox": [ + 86, + 94, + 455, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ConvNeXt [34], and DINOv2 [10, 14, 43] as feature extractors. Then we compute the mean cosine distance of images with different magnifications and sensing modalities, respectively. Scale difference leads to obvious diversity of semantic information and visual descriptions (see Fig. 3).", + "bbox": [ + 75, + 271, + 468, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Algorithm Benchmarking", + "text_level": 1, + "bbox": [ + 76, + 359, + 318, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By benchmarking a wide range of image classification methods, both classic and state-of-the-art, on this unique vision-for-science dataset, we explore how different features of this dataset affect model performance and hope that we can provide some useful analysis that future work can build on. Specifically, we divide our experiments into two major segments: (1) fully-supervised image classification and (2) few-shot image classification, with specific motivations explained in their corresponding sections below.", + "bbox": [ + 75, + 385, + 468, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Fully-Supervised Image Classification", + "text_level": 1, + "bbox": [ + 76, + 530, + 403, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike datasets crawled from the internet, such as ImageNet-1k [16] or LAION-5B [53], LUWA contains niche microscopic images with irregular and discontinuous wear traces that often lack obvious foreground or background. In this experiment, we investigate how well classic and state-of-the-art image classification algorithms generalize to the LUWA dataset and seek to find compelling patterns affecting different models' performance. We want to see how well these patterns align with domain experts' knowledge. We also aim to position SOTA methods from the computer vision community with respect to classification performance achieved by human experts.", + "bbox": [ + 75, + 551, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Settings. We deploy classic methods such as SIFT+FVs [49], ResNets [24] and cutting-edge models such as ViT [17], ConvNeXts [34] and DINOv2 [13, 43] to our LUWA dataset. We believe the time-tested classic methods can serve as a lower-bound benchmark, while the more recent advancements such as DINOv2, which can often be characterized by intensive scale-up in parameter count, can be used as a reference comparable to human experts' performance on the same task.", + "bbox": [ + 75, + 734, + 468, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another major reason that propels us to deploy this wide range of models is to see if there are consistent trends across", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different model architectures and parameter counts and if these trends align with domain experts' knowledge. Specifically, we study the impact of image granularity, magnification, and sensing modality on image classification performance. We also compare different training strategies. By combining the following configurations, we have 324 total experiment results.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Granularity refers to how many pictures one single use-wear is partitioned into. The use-wear is first captured as an image at $865 \\times 865$ resolution. Because many pre-trained models resize the input, which results in pixelated images and loss of fine-grained details that many experts believe are crucial to such a classification task, we partition the original image into 24 or 6 patches and feed all the patches to the model. Importantly, to make our results comparable to those of human experts, we adopt a voting mechanism during test time. If the majority of the 24 patches are classified as class 1, then all the patches of the same use-wear will be classified as class 1 regardless of their actual test results. We believe this method most resembles how archaeologists perform classification when given an $865 \\times 865$ image, as they do not assign a label to each partition.", + "bbox": [ + 496, + 198, + 892, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Magnification represents the magnification multiplier on our microscopic imaging equipment. Our dataset comprises images after $20 \\times$ and $50 \\times$ magnification. We also mix $20 \\times$ and $50 \\times$ data together without any indicator of magnification added to the data to see if mixing magnification will cause any confusion in our models. Note that the magnification multipliers are fixed when the images are taken, and $50 \\times$ images will not look like $20 \\times$ images, even if we significantly downsize them.", + "bbox": [ + 496, + 426, + 892, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sensing Modality refers to whether the picture of the use-wear is stored as texture scans or heightmaps. Texture scans have no depth information, although depth cues are still present. Meanwhile, heightmaps explicitly store the depth information and only the depth information. We want to see if different ways to represent the use-wear will affect computer vision models.", + "bbox": [ + 496, + 563, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy is also varied in our experiments. Some models are initialized with state-of-the-art initialization methods [25] and then trained from scratch. We also apply full-parameter fine-tuning and linear probing [1] with unfrozen and frozen pre-trained weights, respectively.", + "bbox": [ + 496, + 670, + 890, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We deploy ResNet50 (25.6M parameters), ResNet152 (60.4M), ConvNeXt-tiny (28.6M), ConvNeXt-large (197.7M), ViT-H (632M), and DINOv2-ViT-g/14 with registers [13] (1.1B). All models are trained with Adam optimizer [32] on the default setting in PyTorch. We employ linear warmup with cosine annealing [35] as the learning rate scheduler strategy. No data augmentation technique is applied during pre-processing. We defer more details to our supplementary material.", + "bbox": [ + 496, + 747, + 892, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results and Analysis. The overall experiment results can", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22567", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8eb7baa60da76b56b7a0c4c7dc6c891a0aba7996337165938797ca965ee61f4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 89, + 472, + 246 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d5e90ddcd5378f3e36a8dee7f2db970f949e91e19d3700a24e73a5099f6e7e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 90, + 785, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c69bf421866e2a3196552339535d214f28350352597722f331152e828c8b203e.jpg", + "image_caption": [ + "(a) Training Strategy", + "(c) Magnification" + ], + "image_footnote": [ + "ResNet50 ResNet152 CovNeXt-tiny CovNeXt-large ViT-H DINOv2 SIFT + FVs" + ], + "bbox": [ + 178, + 261, + 468, + 415 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc6dcd081af4e687872a2d44b59c3e0eb2a32440c49cfcdea8caf35e8cab30ea.jpg", + "image_caption": [ + "(b) Granularity (Partition Per Sample)", + "(d) Sensing Modality", + "Figure 4. The impact of the training strategy, granularity, magnification, and sensing modality on top-1 classification accuracy in $\\%$ : (a) Due to their huge parameter counts, the experiments do not include full-parameter fine-tuned DINOv2, and ViT-H and DINOv2 trained from scratch. (b) Larger numbers in granularity mean more detailed information about a use-wear is fed into the model." + ], + "image_footnote": [], + "bbox": [ + 491, + 261, + 784, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "be found in Fig. 4. Fig. 4 (a) demonstrated that linear probing yields the most stable and optimal performance across the broad. These results align with our expectation as fine-tuning on uncommon domain-specific datasets may cause catastrophic forgetting [33]. On the other hand, the LUWA dataset itself is too small to train generalizable models from scratch. From the perspective of granularity (Fig. 4 (b)), the more granular partition of the images tends to result in better outcomes, although a diminishing marginal return can be observed. This aligns with our speculation that keeping the original larger image's information as intact as possible is beneficial. It is also possible that the introduction of a voting mechanism brought about a positive regularization effect. More discussion on the voting mechanism can be found in the supplementary material. Considering the selection of magnification, results in Fig. 4 (c) indicated that a higher magnification multiplier is beneficial, which aligns with some human experts' opinions. Notably, mixing data with different magnifications does not confuse the models, and they are able to reap the benefit of abundant data. The same cannot be said for humans, as images with different magnifications can cause confusion. For the sensing modality, we observed that while the best results are usually trained with heightmaps, larger models tend to favor texture. However, the discrepancy is small in general, and the overall performances of the two modalities are com", + "bbox": [ + 75, + 500, + 472, + 893 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "parable as in Fig. 4 (d). More visualizations and detailed tables that follow the trends described above can be found in the supplementary material.", + "bbox": [ + 496, + 500, + 892, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The best performance of $67.05\\%$ top-1 accuracy is achieved by linear probing on ImageNet-1K pre-trained ResNet152 with the heightmap data of 24 partitions and $20 \\times +50 \\times$ magnification. Overall, DINOv2 excels across all aspects, demonstrating the most stable performance amidst varying levels of granularity, magnification, and sensing modalities. The worst performance of the traditional baseline SIFT+FVs is better than the worst configurations of other deep learning methods, but its better configurations are significantly worse than those of the deep learning methods that achieved upper-echelon performance.", + "bbox": [ + 496, + 547, + 893, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Explainability and Comparison with Human Experts. Currently, archaeologists have about $49.5\\%$ accuracy in a double-blind test with a similar setup [19]. Our in-house testing with two professional archaeologists yields an accuracy of $43.75\\%$ in a few-shot setting (Tab. 2). All tested models, except SIFT+FVs, are able to achieve far better accuracy (over $59.5\\%$ ) under several configurations of the dataset. Notably, DINOv2 with linear probing is able to achieve superior or comparable performance under almost all possible configurations.", + "bbox": [ + 496, + 717, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, feature visualization demonstrated that DINOv2 (with registers for better visualization) sometimes", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22568", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ed8c2cf1119049196c3b7a42bba06ece507ee1c3edb632f3afb17e36da363d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PreTr20X TEX50X TEX20X HM50X HM20X+50X TEX20X+50X HM
9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s
ResNet1854.5461.9754.4362.4831.1938.7935.2742.3442.1149.6026.4331.80
ResNet5054.1359.2055.0862.1832.6738.9736.7143.9545.3751.4628.9134.56
ResNet15252.9259.1457.5964.2630.8338.7434.1241.4044.3251.4026.3931.89
ConvNeXt-tiny46.2752.4452.7459.2332.2539.7236.4343.4642.6449.4327.3333.04
ConvNeXt-base48.0454.4554.7462.4831.6239.7035.2643.4641.9148.5626.3532.12
ConvNeXt-large50.8957.0056.6563.5130.1537.7035.2042.9143.8050.6725.4630.79
ViT-base41.0048.8043.8950.9920.6024.9822.6827.5935.6542.3119.3222.64
DINO-small58.8566.1059.5067.3533.5541.2742.5251.1146.9453.9328.0233.49
DINO-base57.2865.3361.3969.6733.0741.8342.3951.2347.5255.3428.2334.00
DeiT-small47.0055.9952.0860.6429.3636.7035.4544.4339.9347.6826.1832.14
DeiT-base53.7061.4855.1263.8132.7140.6837.6746.8044.2152.5727.2034.13
CLIP-base42.9851.3046.5255.0129.7536.9236.9144.5134.4541.2927.8134.02
GPT-4V37.78-31.11-20.00-20.00-21.11-23.33-
Human Expert35.00-43.75-20.00-18.75-33.33-19.44-
", + "bbox": [ + 99, + 88, + 874, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Few-shot image classification performance on LUWA dataset is associated with the magnification and sensing modality. 'PreTr' denotes pre-trained models we used; '20X' and '50X' denote microscopic images at $20 \\times$ and $50 \\times$ magnifications; 'TEX' and 'HM' denote texture and heightmap; '9w5s' and '9w20s' denote 9-way-5-shot and 9-way-20-shot, respectively.", + "bbox": [ + 75, + 324, + 892, + 367 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "recognizes important polished regions in microscopic images of beechwood (see Fig. 5) as the foreground, but recognizes the same polish of different categories ( sprucewood, bone, and antler, see Fig. 5) as the background. More explorations are needed to explain this unwanted behavior. Interestingly, we found that the regions recognized as highly important for classification are similar for human experts and our best model ResNet152, under the same data configuration (see Regions 1, 4, and 6 in Fig. 2). We visualize the results for ResNet152 using Grad-CAM [54].", + "bbox": [ + 75, + 383, + 470, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Few-Shot Image Classification", + "text_level": 1, + "bbox": [ + 76, + 544, + 341, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In practice, LUWA faces a scarcity of microscopic images due to limited stone tools, expertise requirements, and specialized equipment, especially when discovering new categories. Human experts can identify new classes of wear traces with a few examples. To emulate this, we investigate whether few-shot learning can be utilized and how microscopic image magnifications and sensing modalities influence the model's performance.", + "bbox": [ + 75, + 568, + 468, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Settings. We designed two main experiments: (i) Few-shot image classification with a simple but effective pre-train + ProtoNet pipeline [27]. We evaluate the performance of powerful pre-trained models (including ResNet, ViT, DINO [10], ConvNeXt, CLIP [46], DEIT [63]) and popular meta-learners ProtoNet [57] on LUWA dataset. We simulated 600 episodes/tasks and results are demonstrated under 9-way-5/20-shot settings. (ii) GPT-4V [42] experiments: few-shot image classification and reasoning following instructions from human experts. To explore the potential mode of AI-human collaboration facing scientific domains in the advent of large multi-modal models, we collected prompts from three archaeologists and conducted case studies on whether the latest GPT-4V can", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "follow and mimic the experts' reasoning process when analyzing the samples. Then we summarized key points that matter during the experts' analysis and used that to prompt the GPT-4V for few-shot image classification and reasoning. The experiments are illustrated in Fig. 6. Additionally, we included human experts' test results to reflect the difficulty for humans to distinguish these discovered categories with just a few examples. Results are reported under 9-way-5-shot settings in Tab. 2.", + "bbox": [ + 496, + 383, + 890, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Analysis. Experimental results in Tab. 2 demonstrated that DINO excels at few-shot learning classification. Note that in the case of limited microscopic images, classification results of textures at $50 \\times$ magnification setting yield notably superior results compared to others, which provides valuable guidance for few-shot learning tasks in our domain. Moreover, the number of parameters in pre-trained models has a limited impact on this few-shot learning task. We found that GPT-4V can effectively follow the experts' analysis as highlighted in Fig. 6. It learns to emphasize the same points that the experts pay attention to. However, the analysis doesn't always lead to the correct answer. GPT-4V did poorly on few-shot classification. We hypothesize that our data is very different from the web data that GPT models are trained on, and the vision module in GPT-4V still struggles to efficiently present detailed vision information to the language module, especially in a long context such as our multi-image few-shot classification scenario. This means the vision ability of SOTA multi-modal language models still needs improvement before they can be used in scientific tasks with domain-specific data like ours.", + "bbox": [ + 496, + 518, + 892, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Impact and Limitations of LUWA Dataset", + "text_level": 1, + "bbox": [ + 500, + 843, + 875, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scientific Impacts. AI-expert collaboration can provide invaluable insights for scientific research. To tackle the", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22569", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff195c3b5bb4b5a626062a3f354df79507ccea8bb39caede5f9b45e82d8a10ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 84, + 93, + 181, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0649381b2e18effd5b55a428d2e22c6ef97d1a53246b4335d468a56911053452.jpg", + "image_caption": [ + "(a) Sprucewood" + ], + "image_footnote": [], + "bbox": [ + 84, + 170, + 181, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c816ed627a994dd3c605e9d496e962a84203ee20ad86233048c4878fbf1d21c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 184, + 93, + 282, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/728e739159a84a453005f81e7a4c107455d69a5db5d0887429f9b029f0ad7e5e.jpg", + "image_caption": [ + "Figure 5. Feature visualization of LUWA dataset using frozen pre-trained DINOv2." + ], + "image_footnote": [], + "bbox": [ + 184, + 170, + 282, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48f22d766c9fe098378d65d3b8987f32331694015f1e1ff7a4b4bee4b6c4b9b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 93, + 383, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95d91a33759821101f9a261041d2b1a2c2a2b341b6d28c6bc1a7eccdc06462ea.jpg", + "image_caption": [ + "(b) Beechwood" + ], + "image_footnote": [], + "bbox": [ + 287, + 170, + 383, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1771db2c52c69977a0de6c39ffe6e6ea58c6daa17adb349560101f4b18b4d2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 93, + 482, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d478f3862446b768a0730c084a6daa59433310fae6f3829f6fe4001c506dba5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 170, + 482, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09e8dc111431146e0b6ba110d5620d618f7b9fa039f0a72679ad27790dabdc8f.jpg", + "image_caption": [ + "(c) Bone" + ], + "image_footnote": [], + "bbox": [ + 486, + 93, + 584, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c5f05649a664517010fd672d5ff72194eba550355c494cf8eebfef5389c932e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 93, + 683, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tone", + "bbox": [ + 589, + 250, + 609, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a55f425e2e357b6248e93f9c4a43b3166a2bafc783830a5352602dc31af94fc7.jpg", + "image_caption": [ + "(d) Antler" + ], + "image_footnote": [], + "bbox": [ + 686, + 93, + 782, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0b22da14fecc4345b82fc0c51ed9841bfa72c8f060c755b9db0663df6551a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 93, + 883, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Expert Analysis", + "text_level": 1, + "bbox": [ + 112, + 306, + 243, + 316 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/733040ec0b23e5b21d4d857ad3d457cb451e82a4009d41826651e05f35fc34dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 328, + 174, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2844ee651e5d99c9a1b503489929e9345c2b9f5c660548eed599d1f914d33724.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 328, + 228, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First image is more polished I would argue it was used in softer wood (spruce). ... softer materials create larger polishes region. I see linear traces like striated marks ... does not appear that way for harder materials.", + "bbox": [ + 122, + 364, + 240, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second image …\nbeechwood.", + "bbox": [ + 122, + 489, + 210, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Beechwood is a harder material … smaller region of polish … no linear traces seen in the first image.", + "bbox": [ + 122, + 512, + 243, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-4V Analysis", + "text_level": 1, + "bbox": [ + 310, + 308, + 403, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5d9ecf209b0b29afa3e3d5a726a23d916f1c678c19c6a391129dddca6077991.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 330, + 316, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It appears to have a smaller region of polish with sharp bumps and lacks the linear, striated marks that would indicate a softer material like sprucewood. Following the anthropologist's analysis:", + "bbox": [ + 320, + 325, + 473, + 405 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Sprucewood, being softer, would create a larger, more distributed polish and possibly show linear traces.", + "- Beechwood, being harder, would leave smaller and less extensive polish, without the linear traces." + ], + "bbox": [ + 259, + 407, + 473, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The characteristics of the wear in the image suggest the use of beechwood, given the absence of widespread polished areas or striated marks. This aligns with the anthropologist's explanation that harder materials like beechwood tend to create less polish and more pronounced features on the wear surface.", + "bbox": [ + 256, + 474, + 473, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few shot Image Classification with GPT4-V", + "text_level": 1, + "bbox": [ + 571, + 291, + 803, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Prompt:", + "text_level": 1, + "bbox": [ + 506, + 311, + 576, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "An archaeologist conducting lithic use-wear analysis would examine ...", + "bbox": [ + 511, + 327, + 774, + 349 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Polishing...", + "- Striations", + "- Microflaking" + ], + "bbox": [ + 513, + 349, + 612, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Context", + "bbox": [ + 841, + 325, + 854, + 382 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9-8", + "bbox": [ + 844, + 412, + 854, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "#", + "bbox": [ + 866, + 425, + 872, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This is an image taken with advanced microscopes containing traces created by material $\\{\\mathbf{x}\\}$ , so the use-type is $\\{\\mathbf{x}\\}$ .", + "bbox": [ + 506, + 393, + 815, + 419 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2736326d76845e0b369973c663d564424c408786e30c032f23540829e06a4af7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 430, + 555, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6575579de5ed214b8d2cce11201029a332105dc27204649318a0ac5e5cc9dd87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 430, + 620, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b365a1170bb549cdd3ea4bc94ee133d659f89d087dcbe35b7f90cc82a179f70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 430, + 683, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1d3e1ac4d69edd7a67ba91bac032542cce7e631230629f1c35cccd2a604f909.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 430, + 709, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d032f2fd318cf8ffb66af384fd8c40fe7201b5115425c2ea4ff24acc1e7f6f1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 430, + 767, + 467 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41acc77284b3dfd7fc3f413225373474890f5680bd1c7fd0100cb820db023e3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 430, + 823, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "{Antler}", + "bbox": [ + 511, + 478, + 557, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "{Sprucewood} {Barley}", + "bbox": [ + 560, + 478, + 686, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I", + "bbox": [ + 697, + 478, + 735, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "e} {Ivory}", + "bbox": [ + 754, + 477, + 821, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9553326c07ab650062c19ed4d34be186777fdc154dcf1659f1c84479e8147afc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 508, + 557, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This is another image, do the traces on this image belong to which of the nine use-types as demonstrated above? Answer in exactly one word:", + "bbox": [ + 560, + 508, + 831, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-4V Answer: Sprucewood.", + "bbox": [ + 504, + 551, + 673, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query", + "bbox": [ + 844, + 511, + 856, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "long-standing problem of stone tool use, we make the first attempt to collaborate with archaeologists and utilize advanced learning-based methods for worked material inference. The LUWA dataset allows for further investigations to advance our understanding of ancient tool use and material processing techniques.", + "bbox": [ + 75, + 604, + 468, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Directions. We will enrich the LUWA dataset from the following three aspects: (i) supplement microscopic images with both worn and unworn regions using lower-magnification objectives, allowing wear trace segmentation and detection tasks; (ii) collect images including wear traces caused by different worked material; (iii) increase categories of man-made materials for comprehensive analysis on wear features and material properties.", + "bbox": [ + 75, + 695, + 468, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 819, + 194, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We collaborate with anthropological archaeologists and present the first public and the largest Lithic Use-Wear Analysis (LUWA) dataset benefiting both vision and science domains. The LUWA dataset serves as a benchmark", + "bbox": [ + 75, + 839, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to evaluate the generalization capabilities of advanced models on image classification tasks beyond common objects. Addressing specific challenges of wear formation and microscopic imaging, LUWA offers vital guidance on selecting suitable magnifications and sensing modalities facing different scenarios. Our analysis reveals that SOTA models encounter distinct difficulties when facing these specific challenges. Despite DINOV2's superior performance relative to other methods, it overlooks visual features that archaeologists identify as indicative of wear. We anticipate that the LUWA dataset will stimulate further research into enhancing the adaptability of large-scale models to specialized domains within computer vision.", + "bbox": [ + 496, + 604, + 890, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work is supported by NSF Grant 2152565, and by NYU IT High-Performance Computing resources, services, and staff expertise. We gratefully acknowledge Dr. Rakesh Behera for the tribometer hardware, and thank Sara Borsodi, Felix Devis Kisena, Kat Liu, Eugenia Ochoa, Vita Jackman Kuwabara, Alice Jiang, Meiyu Zhang, and Sriram Koushik for their valuable assistance in collecting the microscopic images.", + "bbox": [ + 496, + 801, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22570", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 118, + 174, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. ArXiv, abs/1610.01644, 2016. 5", + "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 2", + "[3] Reza Akbarian Bafghi and Danna Gurari. A new dataset based on images taken by blind people for testing the robustness of image classification models trained for imagenet categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16261-16270, 2023. 2, 4", + "[4] Douglas B Bamforth. Investigating microwear polishes with blind tests: the institute results in context. Journal of Archaeological Science, 15(1):11-23, 1988. 3", + "[5] Tomasz Bartkowiak and Christopher A Brown. Multiscale 3d curvature analysis of processed surface textures of aluminum alloy 6061 t6. Materials, 12(2):257, 2019. 3", + "[6] Sebastiano Battiato, Alessandro Ortis, Francesca Trenta, Lorenzo Ascari, Mara Politi, and Consolata Siniscalco. Detection and classification of pollen grain microscope images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 980-981, 2020. 3", + "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[8] Wonmin Byeon, Manuel Dominguez-Rodrigo, Georgios Arampatzis, Enrique Baquedano, José Yravedra, Miguel Angel Mate-González, and Petros Koumoutsakos. Automated identification and deep classification of cut marks on bones and its paleoanthropological implications. Journal of Computational Science, 32:36-43, 2019. 2, 3", + "[9] Ivan Calandra, Lisa Schunk, Alice Rodriguez, Walter Gneisinger, Antonella Pedergnana, Eduardo Paixao, Telmo Pereira, Radu Iovita, and Joao Marreiros. Back to the edge: relative coordinate system for use-wear analysis. Archaeological and Anthropological Sciences, 11:5937-5948, 2019. 4", + "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 7", + "[11] Claire Lifan Chen, Ata Mahjoubfar, Li-Chia Tai, Ian K Blaby, Allen Huang, Kayvan Reza Niazi, and Bahram Jalali. Deep learning in label-free cell classification. Scientific reports, 6(1):21471, 2016. 3", + "[12] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + ], + "bbox": [ + 78, + 143, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 3", + "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 5", + "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5", + "[15] Aniket Dashpute, Vishwanath Saragadam, Emma Alexander, Florian Willomitzer, Aggelos Katsaggelos, Ashok Veeraraghavan, and Oliver Cossairt. Thermal spread functions (tsf): Physics-guided material classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1641-1650, 2023. 3", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5", + "[18] Manuel S Drehwald, Sagi Eppel, Jolina Li, Han Hao, and Alan Aspuru-Guzik. One-shot recognition of any material anywhere using contrastive learning with physics-based rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23524-23533, 2023. 3", + "[19] Adrian Anthony Evans. On the importance of blind testing in archaeological science: the example from lithic functional studies. Journal of Archaeological Science, 48:5-14, 2014. 3, 4, 6", + "[20] Adrian A Evans and Randolph E Donahue. Laser scanning confocal microscopy: a potential technique for the study of lithic microwear. Journal of Archaeological Science, 35(8): 2223-2230, 2008. 3", + "[21] Andrea Ferreri, Silvia Bucci, and Tatiana Tommasi. Multi-modal rgb-d scene recognition across domains. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2199-2208, 2021. 2", + "[22] Richard LK Fullagar. The role of silica in polish formation. Journal of Archaeological Science, 18(1):1-24, 1991. 4", + "[23] Eric Hayman, Barbara Caputo, Mario Fritz, and Jan-Olof Eklundh. On the significance of real-world conditions for material classification. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part IV 8, pages 253-266. Springer, 2004. 3", + "[24] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 5", + "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1026-1034, 2015. 5" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22571", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[27] Shell Xu Hu, Da Li, Jan Stühmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068–9077, 2022. 7", + "[28] Juan José Ibanez, Talía Lazuen, and J González-Urquijo. Identifying experimental tool use through confocal microscopy. Journal of Archaeological Method and Theory, 26:1176-1215, 2019. 3", + "[29] Hartland W Jackson, Jana R Fischer, Vito RT Zanotelli, H Raza Ali, Robert Mechera, Savas D Soysal, Holger Moch, Simone Muenst, Zsuzsanna Varga, Walter P Weber, et al. The single-cell pathology landscape of breast cancer. Nature, 578 (7796):615-620, 2020. 3", + "[30] L.H. Keeley. Experimental Determination of Stone Tool Uses: A Microwear Analysis. University of Chicago Press, 1980. 1, 3", + "[31] Larry R Kimball, John F Kimball, and Patricia E Allen. Microwave polishes as viewed through the atomic force microscope. Lithic Technology, pages 6-28, 1995. 3", + "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5", + "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6", + "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5", + "[35] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 5", + "[36] João Marreiros, Ivan Calandra, Walter Gneisinger, Eduardo Paixão, Antonella Pedergnana, and Lisa Schunk. Rethinking use-wear analysis and experimentation as applied to the study of past hominin tool use. Journal of Paleolithic Archaeology, 3:475-502, 2020. 3", + "[37] João Marreiros, Telmo Pereira, and Radu Iovita. Controlled experiments in lithic technology and function, 2020. 2, 3", + "[38] Shannon P McPherron, Zeresenay Alemseged, Curtis W Marean, Jonathan G Wynn, Denné Reed, Denis Geraads, René Bobe, and Hamdallah A Béarat. Evidence for stonetool-assisted consumption of animal tissues before 3.39 million years ago at dikika, ethiopia. Nature, 466(7308):857-860, 2010. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Mark Newcomer, Roger Grace, and Romana Unger-Hamilton. Investigating microwear polishes with blind tests. Journal of Archaeological Science, 13(3):203-217, 1986. 3", + "[40] George H Odell. Stone tool research at the end of the millennium: classification, function, and behavior. Journal of Archaeological Research, 9:45-100, 2001. 2", + "[41] George Hamley Odell and Frieda Odell-Vereecken. Verifying the reliability of lithic use-wear assessments by 'blind tests': the low-power approach. Journal of field Archaeology, 7(1):87-120, 1980. 1, 2, 3", + "[42] OpenAI. Chatgpt can now see, hear, and speak, 2023. Accessed: 2023-11-16. 4, 7", + "[43] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 5", + "[44] Antonella Pedergnana and Andreu Olle. Monitoring and interpreting the use-wear formation processes on quartzite flakes through sequential experiments. *Quaternary International*, 427:35–65, 2017. 2", + "[45] Marcos Pizarro-Monzo, Jordi Rosell, Anna Rufa, Florent Rivals, and Ruth Blasco. A deep learning-based taphonomical approach to distinguish the modifying agent in the late pleistocene site of toll cave (barcelona, spain). *Historical Biology*, pages 1–10, 2023. 2", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[47] Alice Rodriguez, Kaushik Yanamandra, Lukasz Witek, Zhong Wang, Rakesh K Behera, and Radu Iovita. The effect of worked material hardness on stone tool wear. Plos one, 17(10):e0276166, 2022. 3, 4", + "[48] Z Rymuza. Tribology of polymers. Archives of civil and mechanical engineering, 7(4):177-184, 2007. 3", + "[49] Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale image classification. In CVPR 2011, pages 1665-1672, 2011. 5", + "[50] Linus Scheibenreif, Michael Mommert, and Damian Borth. Masked vision transformers for hyperspectral image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2165-2175, 2023. 3", + "[51] Patrick Schmidt, Alice Rodriguez, Kaushik Yanamandra, Rakesh K Behera, and Radu Iovita. The mineralogy and structure of use-wear polish on chert. Scientific reports, 10 (1):21512, 2020. 3", + "[52] Benjamin J Schoville, Kyle S Brown, Jacob A Harris, and Jayne Wilkins. New experiments and a model-driven approach for interpreting middle stone age lithic point function using the edge damage distribution method. PloS one, 11 (10):e0164088, 2016. 2", + "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22572", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294. Curran Associates, Inc., 2022. 5", + "[54] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4, 7", + "[55] S.A. SEMENOV. Prehistoric Technology. 1964. 3", + "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4", + "[57] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 7", + "[58] W James Stemp and Michael Stemp. Ubm laser profilometry and lithic use-wear analysis: a variable length scale investigation of surface topography. Journal of Archaeological Science, 28(1):81-88, 2001. 3", + "[59] W James Stemp, Adam S Watson, and Adrian A Evans. Surface analysis of stone and bone tools. Surface Topography: Metrology and Properties, 4(1):013001, 2015. 3", + "[60] Nathan E Stevens, Douglas R Harro, and Alan Hicklin. Practical quantitative lithic use-wear analysis using multiple classifiers. Journal of Archaeological Science, 37(10):2671-2678, 2010. 3", + "[61] Kenichiro Tanaka, Yasuhiro Mukaigawa, Takuya Funatomi, Hiroyuki Kubo, Yasuyuki Matsushita, and Yasushi Yagi. Material classification using frequency- and depth-dependent time-of-flight distortion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3", + "[62] SH Teoh. Fatigue of biomaterials: a review. International journal of fatigue, 22(10):825-837, 2000. 2", + "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7", + "[64] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, page 53, 2019. 3", + "[65] Hongda Wang, Hatice Ceylan Koydemir, Yunzhe Qiu, Bijie Bai, Yibo Zhang, Yiyin Jin, Sabiha Tok, Enis Cagatay Yilmaz, Esin Gumustekin, Yair Rivenson, et al. Early detection and classification of live bacteria using time-lapse coherent imaging and deep learning. Light: Science & Applications, 9(1):118, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Ran Wang, Yuanjing Zhu, Chengxin Chen, Yu Han, and Hongbo Zhou. Tooth wear and tribological investigations in dentistry. Applied Bionics and Biomechanics, 2022, 2022. 2", + "[67] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 3", + "[68] Samuel J Yang, Marc Berndl, D Michael Ando, Mariya Barch, Arunachalam Narayanaswamy, Eric Christiansen, Stephan Hoyer, Chris Roat, Jane Hung, Curtis T Rueden, et al. Assessing microscope image focus quality with deep learning. BMC bioinformatics, 19:1-9, 2018. 3" + ], + "bbox": [ + 501, + 92, + 892, + 301 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22573", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_model.json b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3994d2cc1d4171bb890717f87f80df8ac21e5d33 --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.867, + 0.155 + ], + "angle": 0, + "content": "LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.846, + 0.235 + ], + "angle": 0, + "content": "Jing Zhang\\(^{1,*}\\), Irving Fang\\(^{1,*}\\), Hao Wu\\(^{1}\\), Akshit Kaushik\\(^{1}\\), Alice Rodriguez\\(^{1}\\), Hanwen Zhao\\(^{1}\\), Juexiao Zhang\\(^{1}\\), Zhuo Zheng\\(^{2}\\), Radu Iovita\\(^{1,\\dagger}\\), Chen Feng\\(^{1,\\dagger}\\), New York University" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.238, + 0.612, + 0.251 + ], + "angle": 0, + "content": "https://ai4ce.github.io/LUWA/" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.261, + 0.252, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.261, + 0.436, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.258, + 0.865, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.436, + 0.865, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.557, + 0.895, + 0.6 + ], + "angle": 0, + "content": "Figure 1. LUWA poses a unique computer vision challenge due to: its complex wear formation and irregular wear patterns, ambiguous sensing modalities and magnifications in microscopic imaging. Facing these challenges, the LUWA dataset encompasses both texture and heightmap with different magnifications, encouraging the exploration of image classification beyond common objects." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.606, + 0.314, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.639, + 0.472, + 0.851 + ], + "angle": 0, + "content": "Lithic Use-Wear Analysis (LUWA) using microscopic images is an underexplored vision-for-science research area. It seeks to distinguish the worked material, which is critical for understanding archaeological artifacts, material interactions, tool functionalities, and dental records. However, this challenging task goes beyond the well-studied image classification problem for common objects. It is affected by many confounders owing to the complex wear mechanism and microscopic imaging, which makes it difficult even for human experts to identify the worked material successfully. In this paper, we investigate the following three questions on this unique vision task for the first time:(i) How well can state-of-the-art pre-trained models (like DINOv2) generalize to the rarely seen domain? (ii) How can few-shot" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.608, + 0.895, + 0.79 + ], + "angle": 0, + "content": "learning be exploited for scarce microscopic images? (iii) How do the ambiguous magnification and sensing modality influence the classification accuracy? To study these, we collaborated with archaeologists and built the first open-source and the largest LUWA dataset containing 23,130 microscopic images with different magnifications and sensing modalities. Extensive experiments show that existing pretrained models notably outperform human experts but still leave a large gap for improvements. Most importantly, the LUWA dataset provides an underexplored opportunity for vision and learning communities and complements existing image classification problems on common objects." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.8, + 0.633, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Lithic Use-Wear Analysis (LUWA) is a long-standing scientific problem (see Fig. 1) to identify the functions of stone tools by examining wear traces at the microscopic level on the tool's surface [30, 41]. It seeks to distinguish the worked material (like bone, wood, ivory and antler) us-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.864, + 0.204, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.877, + 0.47, + 0.901 + ], + "angle": 0, + "content": "†Corresponding author: Radu Iovita (iovita@nyu.edu) and Chen Feng (cfeng@nyu.edu)" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.864, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22563" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "ing microscopic images, creating a classification problem. Investigating this unanswered vision-for-science problem will provide invaluable insights for uncovering the hidden story of ancient tools [38] and advancing the understanding of material interactions [62, 66]. However, few studies have begun to explore advanced learning-based methods, and this field is still in a nascent stage [8, 45]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.474, + 0.579 + ], + "angle": 0, + "content": "Besides its potential scientific impact, learning-based LUWA on microscopic images also poses unique computer vision challenges beyond common objects. Unlike common objects with distinct boundaries and typical size, wear traces in microscopic images are irregular and discontinuous (see Fig. 1), making it difficult to define clear visual characteristics [40]. In particular, their absence of clear foreground and background might increase the difficulty of feature extraction. Furthermore, recorded wear traces are a hybrid consequence of the complex wear process and microscopic imaging, which is affected by many factors [41]. For example, the wear duration and the motion type that generates the wear can create significant intra-class variability. Another crucial aspect is the ambiguous magnification and sensing modality of microscopic images. High-sensitivity microscopic imaging allows for capturing complementary sensing modalities and zoom-in details at different magnifications, respectively (see Fig. 1). We found that their modality and magnification differences lead to very different visual features. However, due to the understudied characteristics of wear traces, there has been no conclusive answer to this vision-for-science problem about which magnification and sensing modality is better. In practice, the accessibility of expensive equipment also limits the flexible selection of captured data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.471, + 0.853 + ], + "angle": 0, + "content": "Recent advances in computer vision provide a good opportunity for this challenging scientific problem, especially the emerging paradigm shift with foundation models pretrained on large-scale data [2, 7]. These foundation models generate task-agnostic visual features and have shown excellent performance on image classification with common datasets like ImageNet-\\(\\{1k, A\\}\\) [43]. But how much can state-of-the-art (SOTA) pre-trained representations benefit rarely seen domains not extensively covered by uncurated data available on the Internet? The lack of diversity in domain-specific datasets leaves uncertainty about this question, particularly when addressing real-world vision challenges. Unfortunately, existing studies on LUWA focus on private data and the research community faces a deficiency in accessible datasets. The complexity of wear forms, microscopic imaging, equipment limitations, and the need for expert interpretations hinders the collection of an appropriate dataset that can represent the variability of this domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "To explore these unique challenges and opportunities, we collaborated with anthropological archaeologists and built the first open-source and the largest LUWA dataset contain-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.379 + ], + "angle": 0, + "content": "ing 23,130 microscopic images. Its major characteristics are summarized as follows: (i) Multi-scale wear patterns. Lower and higher magnifications allow for an observation of pattern distributions and topographical details, respectively, which increases the scale variations. (ii) Complementary sensing modalities. Both grayscale microscopic images and corresponding 3D surface profiles are available to provide complementary texture information and geometry cues, helping to identify discriminative features [21]. (iii) Both machine and human wear processes. Samples of stone tools were collected from both machine and human wear processes. The former is to isolate the effect of worked material by tightly controlling other factors [37] while the latter is to reflect the complexity of real scenes, especially for newly discovered categories without baseline data [44, 52]. We envision that the LUWA dataset will encourage researchers to develop and evaluate algorithms for image classification beyond common objects, as a precursor for downstream tasks like segmentation and detection [3]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.519 + ], + "angle": 0, + "content": "Based on the LUWA dataset, we go further in answering the following three problems facing real-world applications: (i) How well representative classification models can generalize to the rarely seen domain? (ii) How can few-shot learning be exploited when scarce microscopic images are available, especially for newly discovered categories? (iii) How do the ambiguous magnification and sensing modality of microscopic images influence the classification accuracy? Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.524, + 0.892, + 0.659 + ], + "angle": 0, + "content": "- We introduce the first open-source and the largest LUWA dataset including 23,130 microscopic images with different magnifications and sensing modalities. We collaborated with archaeologists and set up a data collection pipeline considering the influence of wear formation, microscopic imaging, and expert knowledge. The LUWA dataset allows for reproducible investigations on this rarely seen domain and complements existing image classification problems on common objects." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.66, + 0.892, + 0.84 + ], + "angle": 0, + "content": "- Facing the image classification problem beyond common objects, we benchmark the generalization capability of state-of-the-art vision models (ResNet, ViT, DINOv2, ConvNeXt) in this specific domain. Experimental results show that DINOv2 has the most stable performance amidst varying levels of granularity, magnification, and sensing modalities in the data. We also observed many trends regarding the impacts that features like magnification and sensing modality have on classification accuracy, and some of them are not consistent with experts' heuristics. In general, state-of-the-art computer vision models display super-human accuracy over domain experts." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Considering the scarcity of microscopic images in real scenarios, we investigate the performance of few-shot image classification and reasoning on the LUWA dataset. Particularly, we collected prompts from archaeologists" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.524, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22564" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "and did case studies on whether GPT-4V(ison) can mimic the experts' reasoning process. Further explorations are required to improve the performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.151, + 0.22, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Lithic Use-Wear Analysis. Lithic use-wear analysis was originally developed in the 1950s and aims to answer the scientific problem of how to distinguish the worked material according to wear traces on the stone surface [30, 41, 55]. Low-power and High-power microscopy methods provide useful magnified polishes or micro-fractures via confocal microscopes, tactile profilometers, scanning electron microscopes, and even atomic force microscopes [20, 31, 58]. Existing studies focus on blind tests [4, 19, 39] and quantification methods [4, 5, 20, 28, 59, 60]. However, it remains an insufficiently developed research area due to complex wear patterns and the subjectivity of methods [47]. Existing blind tests demonstrate unreliable identification results on different worked materials. For example, correct identifications of plant and wood are \\(32.4\\%\\) and \\(49.1\\%\\), respectively [19]. The identification of worked materials can be regarded as a unique vision problem, but learning-based algorithms are rarely explored in this research area [8]. The deficiency in accessible datasets also limits its research progress heavily [19]. To solve these, we present the first open-source LUwA dataset and investigate the capability of SOTA classification algorithms on this unique vision-for-science problem for the first time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.524, + 0.473, + 0.84 + ], + "angle": 0, + "content": "Image Classification beyond Common Objects. Image classification is a fundamental vision task that categorizes a whole image as a specific label or class based on its visual contents. Outstanding performance can be achieved on the image classification task of common objects with representative frameworks like ConvNeXt using only public training data [67]. However, real-world scenarios, especially scientific applications, involve non-trivial objectives with different physical properties like tiny pollen grains [6], constituent materials [15, 18, 61], texture in the wild [12], and hyperspectral remote sensing images [50]. Facing uncommon objects, many studies focus on specialized models in respective fields. However, we seek to answer whether simpler pipelines can provide comparable performance. Complementing existing datasets, we introduce a new vision challenge to identify irregular and discontinuous wear traces at the microscopic level and build a classification benchmark to explore how far the practical performance of image classification on uncommon objects can be pushed utilizing existing datasets and architectures, especially foundation models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Microscopic Image Classification. In the context of using microscopic images for classification, a wide variety of applications can be found across different scientific disciplines. This includes the study of biological cells [11, 29]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "bacteria [65], tissue types [64], and material structures [23]. Each application presents unique challenges, particularly in terms of the high-level detail and focus required in the images. These challenges are often compounded by issues such as ambiguity in magnification and sensing modality [23, 68]. To alleviate these challenges, the LUWA Dataset presents a diverse configuration that covers different magnification levels and sensing modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.225, + 0.652, + 0.24 + ], + "angle": 0, + "content": "3. LUWA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.553 + ], + "angle": 0, + "content": "In this section, we describe the LUWA dataset creation process and provide basic statistics. To represent the variability of this domain, in Section 3.1, we present a data collection pipeline, which consists of four key aspects (see Fig. 1). Specifically, ① considering the complexity of wear formation, we introduced both machine and human wear experiments [36] to create stone samples; ② to enrich the dataset diversity and investigate the ambiguous magnification and sensing modality, we utilized an optical 3D profilometer with both \\(20 \\times\\) and \\(50 \\times\\) objective lenses to acquire high-quality texture and heightmap; ③ natural materials were selected according to existing blind tests in the literature [19], particularly including fine-grained categories of wood and plants. ④ domain-specific knowledge is twofold including the identification of wear degrees to increase the dataset diversity and expert interpretations for potential explorations on explainability and the application of vision language models. In Section 3.2, we summarized the LUWA dataset and analyzed its diversity in spatial distributions, magnifications, and sensing modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.667, + 0.575 + ], + "angle": 0, + "content": "3.1. Dataset Creation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Both Machine and Human Wear Processes. LUWA dataset contains stone samples from both machine and human wear processes. Key factors that affect wear results are material properties, mechanical factors, and environmental conditions [48]. To isolate the effect of worked materials, a tightly controlled protocol was used for machine wear experiments [36, 37]. We utilized a tribometer to simulate cutting actions so as to quantify the load applied to the material (a load of 20N), the type of movement (a straight back and forth motion with the speed of 35 repetitions per minute), and the worked duration (Oh, 1h, 3h, 5h, and 12h) (see Fig. 1). To limit the influence of material properties of stone samples, we chose the same flint (Baltic/morainic flint from Denmark) for all experiments [47, 51]. Considering the low classification accuracy of various plants in blind tests [19], we chose the plant cutting process as the human wear experiment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "High-Quality Microscopic Imaging. To capture the high-precision wear traces on stone samples, we utilized an optical 3D profilometer (S neox, Sensorfar Metrology) to collect data with a standardized and reproducible process (see" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.226, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.093, + 0.887, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.377, + 0.892, + 0.472 + ], + "angle": 0, + "content": "Figure 2. Image diversity of LUwA dataset and corresponding visual explanations for human and model decision-making processes. (i) LUwA dataset provides diverse microscopic images associated with spatial distributions (e.g. Regions 1 and 2), magnifications (e.g. Regions 2 and 4) and sensing modalities (texture in the first row and heightmap in the second row); (ii) We compared visual explanations in both human (in the third row) and model (in the fourth row) decision-making processes. Human experts labeled the most important region with red and the less important region with yellow when looking at details of microscopic images to distinguish the worked material. Similarly, Grad-CAM [54] heatmaps use red for the highest importance, yellow for lower importance, and blue for the lowest importance. Interestingly, similar areas (e.g. Regions 1, 4 and 6) are labeled with higher importance for both humans and models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.469, + 0.603 + ], + "angle": 0, + "content": "Fig. 1). To test the influence of magnifications for microscopic image classification, both \\(20\\times\\) and \\(50\\times\\) objective lenses were chosen for measurements. Their spatial resolutions are 0.65 and \\(0.26\\mu \\mathrm{m}\\) pixel, respectively. Furthermore, complementary grayscale images and corresponding 3D surface profiles are acquired via Sensormap. We applied a standard filtering protocol to extract the worn surface and alleviate the effect of natural flints' surface topography [9]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.469, + 0.725 + ], + "angle": 0, + "content": "Domain-Specific Expert Knowledge. Domain-specific knowledge is twofold: (i) human experts help to identify microscopic traces with different wear degrees, enriching the dataset diversity; (ii) for further investigations on explainability and the application of vision language models, human experts also labeled their attention maps when making decisions on worked material (see Fig 1) and provided classification prompt for GPT-4V [42] (see Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.727, + 0.469, + 0.833 + ], + "angle": 0, + "content": "Material Selection and Processing. To benchmark models in this field, we chose representative natural materials (see Fig. 1) according to blind test results in the literature [19] and included fine-grained categories on wood and plants in particular. For the further exploration of wear mechanisms, we analyzed their properties, including the hardness [47] and silicon content [22] in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.241, + 0.862 + ], + "angle": 0, + "content": "3.2. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To reflect the variability of this scientific domain, we built the first public and largest LUwA dataset containing 23,130" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.481, + 0.892, + 0.534 + ], + "angle": 0, + "content": "
Stone SamplesMotion TypesWorked TimeMaterial Categories
34279
MagnificationsSensing Modalities
22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.536, + 0.892, + 0.565 + ], + "angle": 0, + "content": "Table 1. Key factors considered in the LUWA dataset that can reflect the complex wear formation and microscopic imaging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.685 + ], + "angle": 0, + "content": "microscopic images. Specifically, key factors of the complex wear formation and microscopic imaging are considered in the LUWA dataset. As shown in Tab. 1, we report (i) the number of microscopic images, (ii) the number of stone samples, motion types, worked time, and material classes which are exploited, (iii) the number of magnifications and sensing modalities LUWA dataset supports." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Image diversity of LUWA dataset brings challenges for algorithm robustness. It is associated with the spatial distribution of the region collected, the selection of the magnification, and the sensing modality. Greater distances between sampled areas typically result in more pronounced variations in their surface distributions (see Regions 1 and 2 in Fig. 2). Even collected from the same wear trace, the selection of the magnification also contributes to scale difference, which causes totally different wear patterns (see Regions 2 and 4 in Fig. 2). Moreover, LUWA dataset provides both the texture and heightmap, helping to identify discriminative features. We explore the semantic diversity of LUWA dataset on the magnification and sensing modality (see Fig. 2) [3]. We selected VGG [56], ResNet [26]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22566" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.095, + 0.457, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.47, + 0.266 + ], + "angle": 0, + "content": "Figure 3. Cosine similarity distribution of LUWA dataset on different magnifications and sensing modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.47, + 0.349 + ], + "angle": 0, + "content": "ConvNeXt [34], and DINOv2 [10, 14, 43] as feature extractors. Then we compute the mean cosine distance of images with different magnifications and sensing modalities, respectively. Scale difference leads to obvious diversity of semantic information and visual descriptions (see Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.319, + 0.379 + ], + "angle": 0, + "content": "4. Algorithm Benchmarking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.469, + 0.522 + ], + "angle": 0, + "content": "By benchmarking a wide range of image classification methods, both classic and state-of-the-art, on this unique vision-for-science dataset, we explore how different features of this dataset affect model performance and hope that we can provide some useful analysis that future work can build on. Specifically, we divide our experiments into two major segments: (1) fully-supervised image classification and (2) few-shot image classification, with specific motivations explained in their corresponding sections below." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.404, + 0.547 + ], + "angle": 0, + "content": "4.1. Fully-Supervised Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.734 + ], + "angle": 0, + "content": "Unlike datasets crawled from the internet, such as ImageNet-1k [16] or LAION-5B [53], LUWA contains niche microscopic images with irregular and discontinuous wear traces that often lack obvious foreground or background. In this experiment, we investigate how well classic and state-of-the-art image classification algorithms generalize to the LUWA dataset and seek to find compelling patterns affecting different models' performance. We want to see how well these patterns align with domain experts' knowledge. We also aim to position SOTA methods from the computer vision community with respect to classification performance achieved by human experts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.869 + ], + "angle": 0, + "content": "Experimental Settings. We deploy classic methods such as SIFT+FVs [49], ResNets [24] and cutting-edge models such as ViT [17], ConvNeXts [34] and DINOv2 [13, 43] to our LUWA dataset. We believe the time-tested classic methods can serve as a lower-bound benchmark, while the more recent advancements such as DINOv2, which can often be characterized by intensive scale-up in parameter count, can be used as a reference comparable to human experts' performance on the same task." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Another major reason that propels us to deploy this wide range of models is to see if there are consistent trends across" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "different model architectures and parameter counts and if these trends align with domain experts' knowledge. Specifically, we study the impact of image granularity, magnification, and sensing modality on image classification performance. We also compare different training strategies. By combining the following configurations, we have 324 total experiment results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.893, + 0.426 + ], + "angle": 0, + "content": "Granularity refers to how many pictures one single use-wear is partitioned into. The use-wear is first captured as an image at \\(865 \\times 865\\) resolution. Because many pre-trained models resize the input, which results in pixelated images and loss of fine-grained details that many experts believe are crucial to such a classification task, we partition the original image into 24 or 6 patches and feed all the patches to the model. Importantly, to make our results comparable to those of human experts, we adopt a voting mechanism during test time. If the majority of the 24 patches are classified as class 1, then all the patches of the same use-wear will be classified as class 1 regardless of their actual test results. We believe this method most resembles how archaeologists perform classification when given an \\(865 \\times 865\\) image, as they do not assign a label to each partition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.893, + 0.563 + ], + "angle": 0, + "content": "Magnification represents the magnification multiplier on our microscopic imaging equipment. Our dataset comprises images after \\(20 \\times\\) and \\(50 \\times\\) magnification. We also mix \\(20 \\times\\) and \\(50 \\times\\) data together without any indicator of magnification added to the data to see if mixing magnification will cause any confusion in our models. Note that the magnification multipliers are fixed when the images are taken, and \\(50 \\times\\) images will not look like \\(20 \\times\\) images, even if we significantly downsize them." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Sensing Modality refers to whether the picture of the use-wear is stored as texture scans or heightmaps. Texture scans have no depth information, although depth cues are still present. Meanwhile, heightmaps explicitly store the depth information and only the depth information. We want to see if different ways to represent the use-wear will affect computer vision models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.747 + ], + "angle": 0, + "content": "Training Strategy is also varied in our experiments. Some models are initialized with state-of-the-art initialization methods [25] and then trained from scratch. We also apply full-parameter fine-tuning and linear probing [1] with unfrozen and frozen pre-trained weights, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.748, + 0.893, + 0.884 + ], + "angle": 0, + "content": "Implementation Details. We deploy ResNet50 (25.6M parameters), ResNet152 (60.4M), ConvNeXt-tiny (28.6M), ConvNeXt-large (197.7M), ViT-H (632M), and DINOv2-ViT-g/14 with registers [13] (1.1B). All models are trained with Adam optimizer [32] on the default setting in PyTorch. We employ linear warmup with cosine annealing [35] as the learning rate scheduler strategy. No data augmentation technique is applied during pre-processing. We defer more details to our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results and Analysis. The overall experiment results can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22567" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.09, + 0.473, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.251, + 0.394, + 0.261 + ], + "angle": 0, + "content": "(a) Training Strategy" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.091, + 0.787, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.251, + 0.74, + 0.261 + ], + "angle": 0, + "content": "(b) Granularity (Partition Per Sample)" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.262, + 0.47, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.419, + 0.385, + 0.429 + ], + "angle": 0, + "content": "(c) Magnification" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.262, + 0.785, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.419, + 0.701, + 0.429 + ], + "angle": 0, + "content": "(d) Sensing Modality" + }, + { + "type": "image_footnote", + "bbox": [ + 0.212, + 0.431, + 0.782, + 0.446 + ], + "angle": 0, + "content": "ResNet50 ResNet152 CovNeXt-tiny CovNeXt-large ViT-H DINOv2 SIFT + FVs" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.451, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Figure 4. The impact of the training strategy, granularity, magnification, and sensing modality on top-1 classification accuracy in \\(\\%\\) : (a) Due to their huge parameter counts, the experiments do not include full-parameter fine-tuned DINOv2, and ViT-H and DINOv2 trained from scratch. (b) Larger numbers in granularity mean more detailed information about a use-wear is fed into the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.895 + ], + "angle": 0, + "content": "be found in Fig. 4. Fig. 4 (a) demonstrated that linear probing yields the most stable and optimal performance across the broad. These results align with our expectation as fine-tuning on uncommon domain-specific datasets may cause catastrophic forgetting [33]. On the other hand, the LUWA dataset itself is too small to train generalizable models from scratch. From the perspective of granularity (Fig. 4 (b)), the more granular partition of the images tends to result in better outcomes, although a diminishing marginal return can be observed. This aligns with our speculation that keeping the original larger image's information as intact as possible is beneficial. It is also possible that the introduction of a voting mechanism brought about a positive regularization effect. More discussion on the voting mechanism can be found in the supplementary material. Considering the selection of magnification, results in Fig. 4 (c) indicated that a higher magnification multiplier is beneficial, which aligns with some human experts' opinions. Notably, mixing data with different magnifications does not confuse the models, and they are able to reap the benefit of abundant data. The same cannot be said for humans, as images with different magnifications can cause confusion. For the sensing modality, we observed that while the best results are usually trained with heightmaps, larger models tend to favor texture. However, the discrepancy is small in general, and the overall performances of the two modalities are com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.893, + 0.547 + ], + "angle": 0, + "content": "parable as in Fig. 4 (d). More visualizations and detailed tables that follow the trends described above can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.895, + 0.716 + ], + "angle": 0, + "content": "The best performance of \\(67.05\\%\\) top-1 accuracy is achieved by linear probing on ImageNet-1K pre-trained ResNet152 with the heightmap data of 24 partitions and \\(20 \\times +50 \\times\\) magnification. Overall, DINOv2 excels across all aspects, demonstrating the most stable performance amidst varying levels of granularity, magnification, and sensing modalities. The worst performance of the traditional baseline SIFT+FVs is better than the worst configurations of other deep learning methods, but its better configurations are significantly worse than those of the deep learning methods that achieved upper-echelon performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Explainability and Comparison with Human Experts. Currently, archaeologists have about \\(49.5\\%\\) accuracy in a double-blind test with a similar setup [19]. Our in-house testing with two professional archaeologists yields an accuracy of \\(43.75\\%\\) in a few-shot setting (Tab. 2). All tested models, except SIFT+FVs, are able to achieve far better accuracy (over \\(59.5\\%\\)) under several configurations of the dataset. Notably, DINOv2 with linear probing is able to achieve superior or comparable performance under almost all possible configurations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "However, feature visualization demonstrated that DINOv2 (with registers for better visualization) sometimes" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22568" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.875, + 0.32 + ], + "angle": 0, + "content": "
PreTr20X TEX50X TEX20X HM50X HM20X+50X TEX20X+50X HM
9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s
ResNet1854.5461.9754.4362.4831.1938.7935.2742.3442.1149.6026.4331.80
ResNet5054.1359.2055.0862.1832.6738.9736.7143.9545.3751.4628.9134.56
ResNet15252.9259.1457.5964.2630.8338.7434.1241.4044.3251.4026.3931.89
ConvNeXt-tiny46.2752.4452.7459.2332.2539.7236.4343.4642.6449.4327.3333.04
ConvNeXt-base48.0454.4554.7462.4831.6239.7035.2643.4641.9148.5626.3532.12
ConvNeXt-large50.8957.0056.6563.5130.1537.7035.2042.9143.8050.6725.4630.79
ViT-base41.0048.8043.8950.9920.6024.9822.6827.5935.6542.3119.3222.64
DINO-small58.8566.1059.5067.3533.5541.2742.5251.1146.9453.9328.0233.49
DINO-base57.2865.3361.3969.6733.0741.8342.3951.2347.5255.3428.2334.00
DeiT-small47.0055.9952.0860.6429.3636.7035.4544.4339.9347.6826.1832.14
DeiT-base53.7061.4855.1263.8132.7140.6837.6746.8044.2152.5727.2034.13
CLIP-base42.9851.3046.5255.0129.7536.9236.9144.5134.4541.2927.8134.02
GPT-4V37.78-31.11-20.00-20.00-21.11-23.33-
Human Expert35.00-43.75-20.00-18.75-33.33-19.44-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.325, + 0.893, + 0.368 + ], + "angle": 0, + "content": "Table 2. Few-shot image classification performance on LUWA dataset is associated with the magnification and sensing modality. 'PreTr' denotes pre-trained models we used; '20X' and '50X' denote microscopic images at \\(20 \\times\\) and \\(50 \\times\\) magnifications; 'TEX' and 'HM' denote texture and heightmap; '9w5s' and '9w20s' denote 9-way-5-shot and 9-way-20-shot, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.535 + ], + "angle": 0, + "content": "recognizes important polished regions in microscopic images of beechwood (see Fig. 5) as the foreground, but recognizes the same polish of different categories ( sprucewood, bone, and antler, see Fig. 5) as the background. More explorations are needed to explain this unwanted behavior. Interestingly, we found that the regions recognized as highly important for classification are similar for human experts and our best model ResNet152, under the same data configuration (see Regions 1, 4, and 6 in Fig. 2). We visualize the results for ResNet152 using Grad-CAM [54]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.342, + 0.56 + ], + "angle": 0, + "content": "4.2. Few-Shot Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.688 + ], + "angle": 0, + "content": "In practice, LUWA faces a scarcity of microscopic images due to limited stone tools, expertise requirements, and specialized equipment, especially when discovering new categories. Human experts can identify new classes of wear traces with a few examples. To emulate this, we investigate whether few-shot learning can be utilized and how microscopic image magnifications and sensing modalities influence the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Experimental Settings. We designed two main experiments: (i) Few-shot image classification with a simple but effective pre-train + ProtoNet pipeline [27]. We evaluate the performance of powerful pre-trained models (including ResNet, ViT, DINO [10], ConvNeXt, CLIP [46], DEIT [63]) and popular meta-learners ProtoNet [57] on LUWA dataset. We simulated 600 episodes/tasks and results are demonstrated under 9-way-5/20-shot settings. (ii) GPT-4V [42] experiments: few-shot image classification and reasoning following instructions from human experts. To explore the potential mode of AI-human collaboration facing scientific domains in the advent of large multi-modal models, we collected prompts from three archaeologists and conducted case studies on whether the latest GPT-4V can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.519 + ], + "angle": 0, + "content": "follow and mimic the experts' reasoning process when analyzing the samples. Then we summarized key points that matter during the experts' analysis and used that to prompt the GPT-4V for few-shot image classification and reasoning. The experiments are illustrated in Fig. 6. Additionally, we included human experts' test results to reflect the difficulty for humans to distinguish these discovered categories with just a few examples. Results are reported under 9-way-5-shot settings in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.893, + 0.838 + ], + "angle": 0, + "content": "Results and Analysis. Experimental results in Tab. 2 demonstrated that DINO excels at few-shot learning classification. Note that in the case of limited microscopic images, classification results of textures at \\(50 \\times\\) magnification setting yield notably superior results compared to others, which provides valuable guidance for few-shot learning tasks in our domain. Moreover, the number of parameters in pre-trained models has a limited impact on this few-shot learning task. We found that GPT-4V can effectively follow the experts' analysis as highlighted in Fig. 6. It learns to emphasize the same points that the experts pay attention to. However, the analysis doesn't always lead to the correct answer. GPT-4V did poorly on few-shot classification. We hypothesize that our data is very different from the web data that GPT models are trained on, and the vision module in GPT-4V still struggles to efficiently present detailed vision information to the language module, especially in a long context such as our multi-image few-shot classification scenario. This means the vision ability of SOTA multi-modal language models still needs improvement before they can be used in scientific tasks with domain-specific data like ours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.844, + 0.877, + 0.861 + ], + "angle": 0, + "content": "5. Impact and Limitations of LUWA Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Scientific Impacts. AI-expert collaboration can provide invaluable insights for scientific research. To tackle the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.094, + 0.183, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.171, + 0.182, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.25, + 0.225, + 0.261 + ], + "angle": 0, + "content": "(a) Sprucewood" + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.094, + 0.284, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.186, + 0.171, + 0.283, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.094, + 0.384, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.171, + 0.384, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.25, + 0.423, + 0.261 + ], + "angle": 0, + "content": "(b) Beechwood" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.094, + 0.483, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.171, + 0.483, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.094, + 0.585, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.25, + 0.611, + 0.261 + ], + "angle": 0, + "content": "(c) Bone" + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.094, + 0.684, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.25, + 0.61, + 0.261 + ], + "angle": 0, + "content": "tone" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.094, + 0.784, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.763, + 0.25, + 0.814, + 0.261 + ], + "angle": 0, + "content": "(d) Antler" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.094, + 0.885, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.268, + 0.731, + 0.283 + ], + "angle": 0, + "content": "Figure 5. Feature visualization of LUWA dataset using frozen pre-trained DINOv2." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.307, + 0.245, + 0.318 + ], + "angle": 0, + "content": "Prompt: Expert Analysis" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.329, + 0.175, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.329, + 0.229, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.366, + 0.241, + 0.489 + ], + "angle": 0, + "content": "First image is more polished I would argue it was used in softer wood (spruce). ... softer materials create larger polishes region. I see linear traces like striated marks ... does not appear that way for harder materials." + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.49, + 0.212, + 0.511 + ], + "angle": 0, + "content": "Second image …\nbeechwood." + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.513, + 0.244, + 0.569 + ], + "angle": 0, + "content": "Beechwood is a harder material … smaller region of polish … no linear traces seen in the first image." + }, + { + "type": "title", + "bbox": [ + 0.312, + 0.309, + 0.404, + 0.319 + ], + "angle": 0, + "content": "GPT-4V Analysis" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.332, + 0.318, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.327, + 0.475, + 0.406 + ], + "angle": 0, + "content": "It appears to have a smaller region of polish with sharp bumps and lacks the linear, striated marks that would indicate a softer material like sprucewood. Following the anthropologist's analysis:" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.408, + 0.475, + 0.44 + ], + "angle": 0, + "content": "- Sprucewood, being softer, would create a larger, more distributed polish and possibly show linear traces." + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.441, + 0.47, + 0.473 + ], + "angle": 0, + "content": "- Beechwood, being harder, would leave smaller and less extensive polish, without the linear traces." + }, + { + "type": "list", + "bbox": [ + 0.26, + 0.408, + 0.475, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.475, + 0.474, + 0.566 + ], + "angle": 0, + "content": "The characteristics of the wear in the image suggest the use of beechwood, given the absence of widespread polished areas or striated marks. This aligns with the anthropologist's explanation that harder materials like beechwood tend to create less polish and more pronounced features on the wear surface." + }, + { + "type": "title", + "bbox": [ + 0.572, + 0.292, + 0.805, + 0.304 + ], + "angle": 0, + "content": "Few shot Image Classification with GPT4-V" + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.312, + 0.578, + 0.323 + ], + "angle": 0, + "content": "User Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.776, + 0.35 + ], + "angle": 0, + "content": "An archaeologist conducting lithic use-wear analysis would examine ..." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.351, + 0.594, + 0.361 + ], + "angle": 0, + "content": "Polishing..." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.363, + 0.595, + 0.372 + ], + "angle": 0, + "content": "- Striations" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.373, + 0.613, + 0.385 + ], + "angle": 0, + "content": "- Microflaking" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.351, + 0.613, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.843, + 0.326, + 0.855, + 0.383 + ], + "angle": 0, + "content": "Task Context" + }, + { + "type": "text", + "bbox": [ + 0.845, + 0.414, + 0.855, + 0.471 + ], + "angle": 0, + "content": "9-8" + }, + { + "type": "text", + "bbox": [ + 0.867, + 0.426, + 0.874, + 0.469 + ], + "angle": 0, + "content": "#" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.395, + 0.816, + 0.42 + ], + "angle": 0, + "content": "This is an image taken with advanced microscopes containing traces created by material \\(\\{\\mathbf{x}\\}\\), so the use-type is \\(\\{\\mathbf{x}\\}\\)." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.431, + 0.557, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.431, + 0.621, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.431, + 0.684, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.431, + 0.71, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.431, + 0.769, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.431, + 0.825, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.558, + 0.491 + ], + "angle": 0, + "content": "{Antler}" + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.479, + 0.687, + 0.492 + ], + "angle": 0, + "content": "{Sprucewood} {Barley}" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.479, + 0.736, + 0.49 + ], + "angle": 0, + "content": "I" + }, + { + "type": "text", + "bbox": [ + 0.756, + 0.478, + 0.823, + 0.49 + ], + "angle": 0, + "content": "e} {Ivory}" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.509, + 0.558, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.562, + 0.509, + 0.832, + 0.544 + ], + "angle": 0, + "content": "This is another image, do the traces on this image belong to which of the nine use-types as demonstrated above? Answer in exactly one word:" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.553, + 0.674, + 0.565 + ], + "angle": 0, + "content": "GPT-4V Answer: Sprucewood." + }, + { + "type": "text", + "bbox": [ + 0.845, + 0.512, + 0.857, + 0.539 + ], + "angle": 0, + "content": "Query" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.469, + 0.695 + ], + "angle": 0, + "content": "long-standing problem of stone tool use, we make the first attempt to collaborate with archaeologists and utilize advanced learning-based methods for worked material inference. The LUWA dataset allows for further investigations to advance our understanding of ancient tool use and material processing techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.469, + 0.816 + ], + "angle": 0, + "content": "Limitations and Future Directions. We will enrich the LUWA dataset from the following three aspects: (i) supplement microscopic images with both worn and unworn regions using lower-magnification objectives, allowing wear trace segmentation and detection tasks; (ii) collect images including wear traces caused by different worked material; (iii) increase categories of man-made materials for comprehensive analysis on wear features and material properties." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.82, + 0.195, + 0.834 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.9 + ], + "angle": 0, + "content": "We collaborate with anthropological archaeologists and present the first public and the largest Lithic Use-Wear Analysis (LUWA) dataset benefiting both vision and science domains. The LUWA dataset serves as a benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.801 + ], + "angle": 0, + "content": "to evaluate the generalization capabilities of advanced models on image classification tasks beyond common objects. Addressing specific challenges of wear formation and microscopic imaging, LUWA offers vital guidance on selecting suitable magnifications and sensing modalities facing different scenarios. Our analysis reveals that SOTA models encounter distinct difficulties when facing these specific challenges. Despite DINOV2's superior performance relative to other methods, it overlooks visual features that archaeologists identify as indicative of wear. We anticipate that the LUWA dataset will stimulate further research into enhancing the adaptability of large-scale models to specialized domains within computer vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported by NSF Grant 2152565, and by NYU IT High-Performance Computing resources, services, and staff expertise. We gratefully acknowledge Dr. Rakesh Behera for the tribometer hardware, and thank Sara Borsodi, Felix Devis Kisena, Kat Liu, Eugenia Ochoa, Vita Jackman Kuwabara, Alice Jiang, Meiyu Zhang, and Sriram Koushik for their valuable assistance in collecting the microscopic images." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22570" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.119, + 0.176, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[1] Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. ArXiv, abs/1610.01644, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.472, + 0.34 + ], + "angle": 0, + "content": "[3] Reza Akbarian Bafghi and Danna Gurari. A new dataset based on images taken by blind people for testing the robustness of image classification models trained for imagenet categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16261-16270, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.472, + 0.383 + ], + "angle": 0, + "content": "[4] Douglas B Bamforth. Investigating microwear polishes with blind tests: the institute results in context. Journal of Archaeological Science, 15(1):11-23, 1988. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[5] Tomasz Bartkowiak and Christopher A Brown. Multiscale 3d curvature analysis of processed surface textures of aluminum alloy 6061 t6. Materials, 12(2):257, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.472, + 0.508 + ], + "angle": 0, + "content": "[6] Sebastiano Battiato, Alessandro Ortis, Francesca Trenta, Lorenzo Ascari, Mara Politi, and Consolata Siniscalco. Detection and classification of pollen grain microscope images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 980-981, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.472, + 0.579 + ], + "angle": 0, + "content": "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.472, + 0.662 + ], + "angle": 0, + "content": "[8] Wonmin Byeon, Manuel Dominguez-Rodrigo, Georgios Arampatzis, Enrique Baquedano, José Yravedra, Miguel Angel Mate-González, and Petros Koumoutsakos. Automated identification and deep classification of cut marks on bones and its paleoanthropological implications. Journal of Computational Science, 32:36-43, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.663, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[9] Ivan Calandra, Lisa Schunk, Alice Rodriguez, Walter Gneisinger, Antonella Pedergnana, Eduardo Paixao, Telmo Pereira, Radu Iovita, and Joao Marreiros. Back to the edge: relative coordinate system for use-wear analysis. Archaeological and Anthropological Sciences, 11:5937-5948, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.472, + 0.817 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.818, + 0.472, + 0.873 + ], + "angle": 0, + "content": "[11] Claire Lifan Chen, Ata Mahjoubfar, Li-Chia Tai, Ian K Blaby, Allen Huang, Kayvan Reza Niazi, and Bahram Jalali. Deep learning in label-free cell classification. Scientific reports, 6(1):21471, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.145, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.15 + ], + "angle": 0, + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.193 + ], + "angle": 0, + "content": "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[15] Aniket Dashpute, Vishwanath Saragadam, Emma Alexander, Florian Willomitzer, Aggelos Katsaggelos, Ashok Veeraraghavan, and Oliver Cossairt. Thermal spread functions (tsf): Physics-guided material classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1641-1650, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.334 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.419, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[18] Manuel S Drehwald, Sagi Eppel, Jolina Li, Han Hao, and Alan Aspuru-Guzik. One-shot recognition of any material anywhere using contrastive learning with physics-based rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23524-23533, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[19] Adrian Anthony Evans. On the importance of blind testing in archaeological science: the example from lithic functional studies. Journal of Archaeological Science, 48:5-14, 2014. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[20] Adrian A Evans and Randolph E Donahue. Laser scanning confocal microscopy: a potential technique for the study of lithic microwear. Journal of Archaeological Science, 35(8): 2223-2230, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.66 + ], + "angle": 0, + "content": "[21] Andrea Ferreri, Silvia Bucci, and Tatiana Tommasi. Multi-modal rgb-d scene recognition across domains. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2199-2208, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.689 + ], + "angle": 0, + "content": "[22] Richard LK Fullagar. The role of silica in polish formation. Journal of Archaeological Science, 18(1):1-24, 1991. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Eric Hayman, Barbara Caputo, Mario Fritz, and Jan-Olof Eklundh. On the significance of real-world conditions for material classification. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part IV 8, pages 253-266. Springer, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1026-1034, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22571" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.471, + 0.231 + ], + "angle": 0, + "content": "[27] Shell Xu Hu, Da Li, Jan Stühmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068–9077, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[28] Juan José Ibanez, Talía Lazuen, and J González-Urquijo. Identifying experimental tool use through confocal microscopy. Journal of Archaeological Method and Theory, 26:1176-1215, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[29] Hartland W Jackson, Jana R Fischer, Vito RT Zanotelli, H Raza Ali, Robert Mechera, Savas D Soysal, Holger Moch, Simone Muenst, Zsuzsanna Varga, Walter P Weber, et al. The single-cell pathology landscape of breast cancer. Nature, 578 (7796):615-620, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[30] L.H. Keeley. Experimental Determination of Stone Tool Uses: A Microwear Analysis. University of Chicago Press, 1980. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[31] Larry R Kimball, John F Kimball, and Patricia E Allen. Microwave polishes as viewed through the atomic force microscope. Lithic Technology, pages 6-28, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[36] João Marreiros, Ivan Calandra, Walter Gneisinger, Eduardo Paixão, Antonella Pedergnana, and Lisa Schunk. Rethinking use-wear analysis and experimentation as applied to the study of past hominin tool use. Journal of Paleolithic Archaeology, 3:475-502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[37] João Marreiros, Telmo Pereira, and Radu Iovita. Controlled experiments in lithic technology and function, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[38] Shannon P McPherron, Zeresenay Alemseged, Curtis W Marean, Jonathan G Wynn, Denné Reed, Denis Geraads, René Bobe, and Hamdallah A Béarat. Evidence for stonetool-assisted consumption of animal tissues before 3.39 million years ago at dikika, ethiopia. Nature, 466(7308):857-860, 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[39] Mark Newcomer, Roger Grace, and Romana Unger-Hamilton. Investigating microwear polishes with blind tests. Journal of Archaeological Science, 13(3):203-217, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[40] George H Odell. Stone tool research at the end of the millennium: classification, function, and behavior. Journal of Archaeological Research, 9:45-100, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[41] George Hamley Odell and Frieda Odell-Vereecken. Verifying the reliability of lithic use-wear assessments by 'blind tests': the low-power approach. Journal of field Archaeology, 7(1):87-120, 1980. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[42] OpenAI. Chatgpt can now see, hear, and speak, 2023. Accessed: 2023-11-16. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[43] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[44] Antonella Pedergnana and Andreu Olle. Monitoring and interpreting the use-wear formation processes on quartzite flakes through sequential experiments. *Quaternary International*, 427:35–65, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[45] Marcos Pizarro-Monzo, Jordi Rosell, Anna Rufa, Florent Rivals, and Ruth Blasco. A deep learning-based taphonomical approach to distinguish the modifying agent in the late pleistocene site of toll cave (barcelona, spain). *Historical Biology*, pages 1–10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[47] Alice Rodriguez, Kaushik Yanamandra, Lukasz Witek, Zhong Wang, Rakesh K Behera, and Radu Iovita. The effect of worked material hardness on stone tool wear. Plos one, 17(10):e0276166, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[48] Z Rymuza. Tribology of polymers. Archives of civil and mechanical engineering, 7(4):177-184, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[49] Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale image classification. In CVPR 2011, pages 1665-1672, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[50] Linus Scheibenreif, Michael Mommert, and Damian Borth. Masked vision transformers for hyperspectral image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2165-2175, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[51] Patrick Schmidt, Alice Rodriguez, Kaushik Yanamandra, Rakesh K Behera, and Radu Iovita. The mineralogy and structure of use-wear polish on chert. Scientific reports, 10 (1):21512, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[52] Benjamin J Schoville, Kyle S Brown, Jacob A Harris, and Jayne Wilkins. New experiments and a model-driven approach for interpreting middle stone age lithic point function using the edge damage distribution method. PloS one, 11 (10):e0164088, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22572" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.189 + ], + "angle": 0, + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294. Curran Associates, Inc., 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[54] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.276, + 0.414, + 0.29 + ], + "angle": 0, + "content": "[55] S.A. SEMENOV. Prehistoric Technology. 1964. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.376 + ], + "angle": 0, + "content": "[57] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.47, + 0.433 + ], + "angle": 0, + "content": "[58] W James Stemp and Michael Stemp. Ubm laser profilometry and lithic use-wear analysis: a variable length scale investigation of surface topography. Journal of Archaeological Science, 28(1):81-88, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[59] W James Stemp, Adam S Watson, and Adrian A Evans. Surface analysis of stone and bone tools. Surface Topography: Metrology and Properties, 4(1):013001, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.478, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[60] Nathan E Stevens, Douglas R Harro, and Alan Hicklin. Practical quantitative lithic use-wear analysis using multiple classifiers. Journal of Archaeological Science, 37(10):2671-2678, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.47, + 0.618 + ], + "angle": 0, + "content": "[61] Kenichiro Tanaka, Yasuhiro Mukaigawa, Takuya Funatomi, Hiroyuki Kubo, Yasuyuki Matsushita, and Yasushi Yagi. Material classification using frequency- and depth-dependent time-of-flight distortion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.47, + 0.647 + ], + "angle": 0, + "content": "[62] SH Teoh. Fatigue of biomaterials: a review. International journal of fatigue, 22(10):825-837, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[64] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, page 53, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[65] Hongda Wang, Hatice Ceylan Koydemir, Yunzhe Qiu, Bijie Bai, Yibo Zhang, Yiyin Jin, Sabiha Tok, Enis Cagatay Yilmaz, Esin Gumustekin, Yair Rivenson, et al. Early detection and classification of live bacteria using time-lapse coherent imaging and deep learning. Light: Science & Applications, 9(1):118, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[66] Ran Wang, Yuanjing Zhu, Chengxin Chen, Yu Han, and Hongbo Zhou. Tooth wear and tribological investigations in dentistry. Applied Bionics and Biomechanics, 2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[67] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[68] Samuel J Yang, Marc Berndl, D Michael Ando, Mariya Barch, Arunachalam Narayanaswamy, Eric Christiansen, Stephan Hoyer, Chris Roat, Jane Hung, Curtis T Rueden, et al. Assessing microscope image focus quality with deep learning. BMC bioinformatics, 19:1-9, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22573" + } + ] +] \ No newline at end of file diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_origin.pdf b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bac4e5fcb0efa6b0805e69b8eb56146a84e3af93 --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/962a883b-464b-4830-ab92-8bd978dfedca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee766840104ca2e72e383d147805bdb6824e36815b280488b1e943d4d1d9459b +size 7161424 diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/full.md b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..55443e3bc149c8514fb83b455790f3f14b0202fc --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/full.md @@ -0,0 +1,351 @@ +# LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images + +Jing Zhang $^{1,*}$ , Irving Fang $^{1,*}$ , Hao Wu $^{1}$ , Akshit Kaushik $^{1}$ , Alice Rodriguez $^{1}$ , Hanwen Zhao $^{1}$ , Juexiao Zhang $^{1}$ , Zhuo Zheng $^{2}$ , Radu Iovita $^{1,\dagger}$ , Chen Feng $^{1,\dagger}$ , New York University + +https://ai4ce.github.io/LUWA/ + +![](images/90bf1edf98efabb4f624952e97e9e1986bcba9feb09c9fa497ad88f2b266a309.jpg) +Figure 1. LUWA poses a unique computer vision challenge due to: its complex wear formation and irregular wear patterns, ambiguous sensing modalities and magnifications in microscopic imaging. Facing these challenges, the LUWA dataset encompasses both texture and heightmap with different magnifications, encouraging the exploration of image classification beyond common objects. + +![](images/9c5d082b4abe8c0afaaaa5096928c76addb5ef8128cad1dfe18ed7d44bf574a8.jpg) + +![](images/24e4e8f23090d87e9f0549de6e4464e21a3c8a9c1946e0a7813110ffa5d1b639.jpg) + +![](images/2f254019f55b495f1d9712d1e46158be6ca39c86cec115693282a96774ad4ab9.jpg) + +# Abstract + +Lithic Use-Wear Analysis (LUWA) using microscopic images is an underexplored vision-for-science research area. It seeks to distinguish the worked material, which is critical for understanding archaeological artifacts, material interactions, tool functionalities, and dental records. However, this challenging task goes beyond the well-studied image classification problem for common objects. It is affected by many confounders owing to the complex wear mechanism and microscopic imaging, which makes it difficult even for human experts to identify the worked material successfully. In this paper, we investigate the following three questions on this unique vision task for the first time:(i) How well can state-of-the-art pre-trained models (like DINOv2) generalize to the rarely seen domain? (ii) How can few-shot + +learning be exploited for scarce microscopic images? (iii) How do the ambiguous magnification and sensing modality influence the classification accuracy? To study these, we collaborated with archaeologists and built the first open-source and the largest LUWA dataset containing 23,130 microscopic images with different magnifications and sensing modalities. Extensive experiments show that existing pretrained models notably outperform human experts but still leave a large gap for improvements. Most importantly, the LUWA dataset provides an underexplored opportunity for vision and learning communities and complements existing image classification problems on common objects. + +# 1. Introduction + +Lithic Use-Wear Analysis (LUWA) is a long-standing scientific problem (see Fig. 1) to identify the functions of stone tools by examining wear traces at the microscopic level on the tool's surface [30, 41]. It seeks to distinguish the worked material (like bone, wood, ivory and antler) us- + +ing microscopic images, creating a classification problem. Investigating this unanswered vision-for-science problem will provide invaluable insights for uncovering the hidden story of ancient tools [38] and advancing the understanding of material interactions [62, 66]. However, few studies have begun to explore advanced learning-based methods, and this field is still in a nascent stage [8, 45]. + +Besides its potential scientific impact, learning-based LUWA on microscopic images also poses unique computer vision challenges beyond common objects. Unlike common objects with distinct boundaries and typical size, wear traces in microscopic images are irregular and discontinuous (see Fig. 1), making it difficult to define clear visual characteristics [40]. In particular, their absence of clear foreground and background might increase the difficulty of feature extraction. Furthermore, recorded wear traces are a hybrid consequence of the complex wear process and microscopic imaging, which is affected by many factors [41]. For example, the wear duration and the motion type that generates the wear can create significant intra-class variability. Another crucial aspect is the ambiguous magnification and sensing modality of microscopic images. High-sensitivity microscopic imaging allows for capturing complementary sensing modalities and zoom-in details at different magnifications, respectively (see Fig. 1). We found that their modality and magnification differences lead to very different visual features. However, due to the understudied characteristics of wear traces, there has been no conclusive answer to this vision-for-science problem about which magnification and sensing modality is better. In practice, the accessibility of expensive equipment also limits the flexible selection of captured data. + +Recent advances in computer vision provide a good opportunity for this challenging scientific problem, especially the emerging paradigm shift with foundation models pretrained on large-scale data [2, 7]. These foundation models generate task-agnostic visual features and have shown excellent performance on image classification with common datasets like ImageNet- $\{1k, A\}$ [43]. But how much can state-of-the-art (SOTA) pre-trained representations benefit rarely seen domains not extensively covered by uncurated data available on the Internet? The lack of diversity in domain-specific datasets leaves uncertainty about this question, particularly when addressing real-world vision challenges. Unfortunately, existing studies on LUWA focus on private data and the research community faces a deficiency in accessible datasets. The complexity of wear forms, microscopic imaging, equipment limitations, and the need for expert interpretations hinders the collection of an appropriate dataset that can represent the variability of this domain. + +To explore these unique challenges and opportunities, we collaborated with anthropological archaeologists and built the first open-source and the largest LUWA dataset contain- + +ing 23,130 microscopic images. Its major characteristics are summarized as follows: (i) Multi-scale wear patterns. Lower and higher magnifications allow for an observation of pattern distributions and topographical details, respectively, which increases the scale variations. (ii) Complementary sensing modalities. Both grayscale microscopic images and corresponding 3D surface profiles are available to provide complementary texture information and geometry cues, helping to identify discriminative features [21]. (iii) Both machine and human wear processes. Samples of stone tools were collected from both machine and human wear processes. The former is to isolate the effect of worked material by tightly controlling other factors [37] while the latter is to reflect the complexity of real scenes, especially for newly discovered categories without baseline data [44, 52]. We envision that the LUWA dataset will encourage researchers to develop and evaluate algorithms for image classification beyond common objects, as a precursor for downstream tasks like segmentation and detection [3]. + +Based on the LUWA dataset, we go further in answering the following three problems facing real-world applications: (i) How well representative classification models can generalize to the rarely seen domain? (ii) How can few-shot learning be exploited when scarce microscopic images are available, especially for newly discovered categories? (iii) How do the ambiguous magnification and sensing modality of microscopic images influence the classification accuracy? Our contributions are summarized as follows: + +- We introduce the first open-source and the largest LUWA dataset including 23,130 microscopic images with different magnifications and sensing modalities. We collaborated with archaeologists and set up a data collection pipeline considering the influence of wear formation, microscopic imaging, and expert knowledge. The LUWA dataset allows for reproducible investigations on this rarely seen domain and complements existing image classification problems on common objects. +- Facing the image classification problem beyond common objects, we benchmark the generalization capability of state-of-the-art vision models (ResNet, ViT, DINOv2, ConvNeXt) in this specific domain. Experimental results show that DINOv2 has the most stable performance amidst varying levels of granularity, magnification, and sensing modalities in the data. We also observed many trends regarding the impacts that features like magnification and sensing modality have on classification accuracy, and some of them are not consistent with experts' heuristics. In general, state-of-the-art computer vision models display super-human accuracy over domain experts. +- Considering the scarcity of microscopic images in real scenarios, we investigate the performance of few-shot image classification and reasoning on the LUWA dataset. Particularly, we collected prompts from archaeologists + +and did case studies on whether GPT-4V(ison) can mimic the experts' reasoning process. Further explorations are required to improve the performance. + +# 2. Related Work + +Lithic Use-Wear Analysis. Lithic use-wear analysis was originally developed in the 1950s and aims to answer the scientific problem of how to distinguish the worked material according to wear traces on the stone surface [30, 41, 55]. Low-power and High-power microscopy methods provide useful magnified polishes or micro-fractures via confocal microscopes, tactile profilometers, scanning electron microscopes, and even atomic force microscopes [20, 31, 58]. Existing studies focus on blind tests [4, 19, 39] and quantification methods [4, 5, 20, 28, 59, 60]. However, it remains an insufficiently developed research area due to complex wear patterns and the subjectivity of methods [47]. Existing blind tests demonstrate unreliable identification results on different worked materials. For example, correct identifications of plant and wood are $32.4\%$ and $49.1\%$ , respectively [19]. The identification of worked materials can be regarded as a unique vision problem, but learning-based algorithms are rarely explored in this research area [8]. The deficiency in accessible datasets also limits its research progress heavily [19]. To solve these, we present the first open-source LUwA dataset and investigate the capability of SOTA classification algorithms on this unique vision-for-science problem for the first time. + +Image Classification beyond Common Objects. Image classification is a fundamental vision task that categorizes a whole image as a specific label or class based on its visual contents. Outstanding performance can be achieved on the image classification task of common objects with representative frameworks like ConvNeXt using only public training data [67]. However, real-world scenarios, especially scientific applications, involve non-trivial objectives with different physical properties like tiny pollen grains [6], constituent materials [15, 18, 61], texture in the wild [12], and hyperspectral remote sensing images [50]. Facing uncommon objects, many studies focus on specialized models in respective fields. However, we seek to answer whether simpler pipelines can provide comparable performance. Complementing existing datasets, we introduce a new vision challenge to identify irregular and discontinuous wear traces at the microscopic level and build a classification benchmark to explore how far the practical performance of image classification on uncommon objects can be pushed utilizing existing datasets and architectures, especially foundation models. + +Microscopic Image Classification. In the context of using microscopic images for classification, a wide variety of applications can be found across different scientific disciplines. This includes the study of biological cells [11, 29], + +bacteria [65], tissue types [64], and material structures [23]. Each application presents unique challenges, particularly in terms of the high-level detail and focus required in the images. These challenges are often compounded by issues such as ambiguity in magnification and sensing modality [23, 68]. To alleviate these challenges, the LUWA Dataset presents a diverse configuration that covers different magnification levels and sensing modalities. + +# 3. LUWA Dataset + +In this section, we describe the LUWA dataset creation process and provide basic statistics. To represent the variability of this domain, in Section 3.1, we present a data collection pipeline, which consists of four key aspects (see Fig. 1). Specifically, ① considering the complexity of wear formation, we introduced both machine and human wear experiments [36] to create stone samples; ② to enrich the dataset diversity and investigate the ambiguous magnification and sensing modality, we utilized an optical 3D profilometer with both $20 \times$ and $50 \times$ objective lenses to acquire high-quality texture and heightmap; ③ natural materials were selected according to existing blind tests in the literature [19], particularly including fine-grained categories of wood and plants. ④ domain-specific knowledge is twofold including the identification of wear degrees to increase the dataset diversity and expert interpretations for potential explorations on explainability and the application of vision language models. In Section 3.2, we summarized the LUWA dataset and analyzed its diversity in spatial distributions, magnifications, and sensing modalities. + +# 3.1. Dataset Creation + +Both Machine and Human Wear Processes. LUWA dataset contains stone samples from both machine and human wear processes. Key factors that affect wear results are material properties, mechanical factors, and environmental conditions [48]. To isolate the effect of worked materials, a tightly controlled protocol was used for machine wear experiments [36, 37]. We utilized a tribometer to simulate cutting actions so as to quantify the load applied to the material (a load of 20N), the type of movement (a straight back and forth motion with the speed of 35 repetitions per minute), and the worked duration (Oh, 1h, 3h, 5h, and 12h) (see Fig. 1). To limit the influence of material properties of stone samples, we chose the same flint (Baltic/morainic flint from Denmark) for all experiments [47, 51]. Considering the low classification accuracy of various plants in blind tests [19], we chose the plant cutting process as the human wear experiment. + +High-Quality Microscopic Imaging. To capture the high-precision wear traces on stone samples, we utilized an optical 3D profilometer (S neox, Sensorfar Metrology) to collect data with a standardized and reproducible process (see + +![](images/0e7f10f1f12cf472b1951873d600cc83e0f9e841f8d403c0dea862c53fdf066c.jpg) +Figure 2. Image diversity of LUwA dataset and corresponding visual explanations for human and model decision-making processes. (i) LUwA dataset provides diverse microscopic images associated with spatial distributions (e.g. Regions 1 and 2), magnifications (e.g. Regions 2 and 4) and sensing modalities (texture in the first row and heightmap in the second row); (ii) We compared visual explanations in both human (in the third row) and model (in the fourth row) decision-making processes. Human experts labeled the most important region with red and the less important region with yellow when looking at details of microscopic images to distinguish the worked material. Similarly, Grad-CAM [54] heatmaps use red for the highest importance, yellow for lower importance, and blue for the lowest importance. Interestingly, similar areas (e.g. Regions 1, 4 and 6) are labeled with higher importance for both humans and models. + +![](images/fbc0531bbf633635a3888c4310b691159cf89f577853140c8b40389b8015517f.jpg) + +Fig. 1). To test the influence of magnifications for microscopic image classification, both $20\times$ and $50\times$ objective lenses were chosen for measurements. Their spatial resolutions are 0.65 and $0.26\mu \mathrm{m}$ pixel, respectively. Furthermore, complementary grayscale images and corresponding 3D surface profiles are acquired via Sensormap. We applied a standard filtering protocol to extract the worn surface and alleviate the effect of natural flints' surface topography [9]. + +Domain-Specific Expert Knowledge. Domain-specific knowledge is twofold: (i) human experts help to identify microscopic traces with different wear degrees, enriching the dataset diversity; (ii) for further investigations on explainability and the application of vision language models, human experts also labeled their attention maps when making decisions on worked material (see Fig 1) and provided classification prompt for GPT-4V [42] (see Fig. 6). + +Material Selection and Processing. To benchmark models in this field, we chose representative natural materials (see Fig. 1) according to blind test results in the literature [19] and included fine-grained categories on wood and plants in particular. For the further exploration of wear mechanisms, we analyzed their properties, including the hardness [47] and silicon content [22] in the supplementary material. + +# 3.2. Dataset Analysis + +To reflect the variability of this scientific domain, we built the first public and largest LUwA dataset containing 23,130 + +
Stone SamplesMotion TypesWorked TimeMaterial Categories
34279
MagnificationsSensing Modalities
22
+ +Table 1. Key factors considered in the LUWA dataset that can reflect the complex wear formation and microscopic imaging. + +microscopic images. Specifically, key factors of the complex wear formation and microscopic imaging are considered in the LUWA dataset. As shown in Tab. 1, we report (i) the number of microscopic images, (ii) the number of stone samples, motion types, worked time, and material classes which are exploited, (iii) the number of magnifications and sensing modalities LUWA dataset supports. + +Image diversity of LUWA dataset brings challenges for algorithm robustness. It is associated with the spatial distribution of the region collected, the selection of the magnification, and the sensing modality. Greater distances between sampled areas typically result in more pronounced variations in their surface distributions (see Regions 1 and 2 in Fig. 2). Even collected from the same wear trace, the selection of the magnification also contributes to scale difference, which causes totally different wear patterns (see Regions 2 and 4 in Fig. 2). Moreover, LUWA dataset provides both the texture and heightmap, helping to identify discriminative features. We explore the semantic diversity of LUWA dataset on the magnification and sensing modality (see Fig. 2) [3]. We selected VGG [56], ResNet [26], + +![](images/6524fac1bc150096ee6ef536f5b3306fb3a3eea2077df3cc9c5444b6c24ceaca.jpg) +Figure 3. Cosine similarity distribution of LUWA dataset on different magnifications and sensing modalities. + +ConvNeXt [34], and DINOv2 [10, 14, 43] as feature extractors. Then we compute the mean cosine distance of images with different magnifications and sensing modalities, respectively. Scale difference leads to obvious diversity of semantic information and visual descriptions (see Fig. 3). + +# 4. Algorithm Benchmarking + +By benchmarking a wide range of image classification methods, both classic and state-of-the-art, on this unique vision-for-science dataset, we explore how different features of this dataset affect model performance and hope that we can provide some useful analysis that future work can build on. Specifically, we divide our experiments into two major segments: (1) fully-supervised image classification and (2) few-shot image classification, with specific motivations explained in their corresponding sections below. + +# 4.1. Fully-Supervised Image Classification + +Unlike datasets crawled from the internet, such as ImageNet-1k [16] or LAION-5B [53], LUWA contains niche microscopic images with irregular and discontinuous wear traces that often lack obvious foreground or background. In this experiment, we investigate how well classic and state-of-the-art image classification algorithms generalize to the LUWA dataset and seek to find compelling patterns affecting different models' performance. We want to see how well these patterns align with domain experts' knowledge. We also aim to position SOTA methods from the computer vision community with respect to classification performance achieved by human experts. + +Experimental Settings. We deploy classic methods such as SIFT+FVs [49], ResNets [24] and cutting-edge models such as ViT [17], ConvNeXts [34] and DINOv2 [13, 43] to our LUWA dataset. We believe the time-tested classic methods can serve as a lower-bound benchmark, while the more recent advancements such as DINOv2, which can often be characterized by intensive scale-up in parameter count, can be used as a reference comparable to human experts' performance on the same task. + +Another major reason that propels us to deploy this wide range of models is to see if there are consistent trends across + +different model architectures and parameter counts and if these trends align with domain experts' knowledge. Specifically, we study the impact of image granularity, magnification, and sensing modality on image classification performance. We also compare different training strategies. By combining the following configurations, we have 324 total experiment results. + +Granularity refers to how many pictures one single use-wear is partitioned into. The use-wear is first captured as an image at $865 \times 865$ resolution. Because many pre-trained models resize the input, which results in pixelated images and loss of fine-grained details that many experts believe are crucial to such a classification task, we partition the original image into 24 or 6 patches and feed all the patches to the model. Importantly, to make our results comparable to those of human experts, we adopt a voting mechanism during test time. If the majority of the 24 patches are classified as class 1, then all the patches of the same use-wear will be classified as class 1 regardless of their actual test results. We believe this method most resembles how archaeologists perform classification when given an $865 \times 865$ image, as they do not assign a label to each partition. + +Magnification represents the magnification multiplier on our microscopic imaging equipment. Our dataset comprises images after $20 \times$ and $50 \times$ magnification. We also mix $20 \times$ and $50 \times$ data together without any indicator of magnification added to the data to see if mixing magnification will cause any confusion in our models. Note that the magnification multipliers are fixed when the images are taken, and $50 \times$ images will not look like $20 \times$ images, even if we significantly downsize them. + +Sensing Modality refers to whether the picture of the use-wear is stored as texture scans or heightmaps. Texture scans have no depth information, although depth cues are still present. Meanwhile, heightmaps explicitly store the depth information and only the depth information. We want to see if different ways to represent the use-wear will affect computer vision models. + +Training Strategy is also varied in our experiments. Some models are initialized with state-of-the-art initialization methods [25] and then trained from scratch. We also apply full-parameter fine-tuning and linear probing [1] with unfrozen and frozen pre-trained weights, respectively. + +Implementation Details. We deploy ResNet50 (25.6M parameters), ResNet152 (60.4M), ConvNeXt-tiny (28.6M), ConvNeXt-large (197.7M), ViT-H (632M), and DINOv2-ViT-g/14 with registers [13] (1.1B). All models are trained with Adam optimizer [32] on the default setting in PyTorch. We employ linear warmup with cosine annealing [35] as the learning rate scheduler strategy. No data augmentation technique is applied during pre-processing. We defer more details to our supplementary material. + +Results and Analysis. The overall experiment results can + +![](images/8eb7baa60da76b56b7a0c4c7dc6c891a0aba7996337165938797ca965ee61f4b.jpg) + +![](images/9d5e90ddcd5378f3e36a8dee7f2db970f949e91e19d3700a24e73a5099f6e7e6.jpg) + +(a) Training Strategy +(c) Magnification +![](images/c69bf421866e2a3196552339535d214f28350352597722f331152e828c8b203e.jpg) +ResNet50 ResNet152 CovNeXt-tiny CovNeXt-large ViT-H DINOv2 SIFT + FVs + +![](images/cc6dcd081af4e687872a2d44b59c3e0eb2a32440c49cfcdea8caf35e8cab30ea.jpg) +(b) Granularity (Partition Per Sample) +(d) Sensing Modality +Figure 4. The impact of the training strategy, granularity, magnification, and sensing modality on top-1 classification accuracy in $\%$ : (a) Due to their huge parameter counts, the experiments do not include full-parameter fine-tuned DINOv2, and ViT-H and DINOv2 trained from scratch. (b) Larger numbers in granularity mean more detailed information about a use-wear is fed into the model. + +be found in Fig. 4. Fig. 4 (a) demonstrated that linear probing yields the most stable and optimal performance across the broad. These results align with our expectation as fine-tuning on uncommon domain-specific datasets may cause catastrophic forgetting [33]. On the other hand, the LUWA dataset itself is too small to train generalizable models from scratch. From the perspective of granularity (Fig. 4 (b)), the more granular partition of the images tends to result in better outcomes, although a diminishing marginal return can be observed. This aligns with our speculation that keeping the original larger image's information as intact as possible is beneficial. It is also possible that the introduction of a voting mechanism brought about a positive regularization effect. More discussion on the voting mechanism can be found in the supplementary material. Considering the selection of magnification, results in Fig. 4 (c) indicated that a higher magnification multiplier is beneficial, which aligns with some human experts' opinions. Notably, mixing data with different magnifications does not confuse the models, and they are able to reap the benefit of abundant data. The same cannot be said for humans, as images with different magnifications can cause confusion. For the sensing modality, we observed that while the best results are usually trained with heightmaps, larger models tend to favor texture. However, the discrepancy is small in general, and the overall performances of the two modalities are com + +parable as in Fig. 4 (d). More visualizations and detailed tables that follow the trends described above can be found in the supplementary material. + +The best performance of $67.05\%$ top-1 accuracy is achieved by linear probing on ImageNet-1K pre-trained ResNet152 with the heightmap data of 24 partitions and $20 \times +50 \times$ magnification. Overall, DINOv2 excels across all aspects, demonstrating the most stable performance amidst varying levels of granularity, magnification, and sensing modalities. The worst performance of the traditional baseline SIFT+FVs is better than the worst configurations of other deep learning methods, but its better configurations are significantly worse than those of the deep learning methods that achieved upper-echelon performance. + +Explainability and Comparison with Human Experts. Currently, archaeologists have about $49.5\%$ accuracy in a double-blind test with a similar setup [19]. Our in-house testing with two professional archaeologists yields an accuracy of $43.75\%$ in a few-shot setting (Tab. 2). All tested models, except SIFT+FVs, are able to achieve far better accuracy (over $59.5\%$ ) under several configurations of the dataset. Notably, DINOv2 with linear probing is able to achieve superior or comparable performance under almost all possible configurations. + +However, feature visualization demonstrated that DINOv2 (with registers for better visualization) sometimes + +
PreTr20X TEX50X TEX20X HM50X HM20X+50X TEX20X+50X HM
9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s
ResNet1854.5461.9754.4362.4831.1938.7935.2742.3442.1149.6026.4331.80
ResNet5054.1359.2055.0862.1832.6738.9736.7143.9545.3751.4628.9134.56
ResNet15252.9259.1457.5964.2630.8338.7434.1241.4044.3251.4026.3931.89
ConvNeXt-tiny46.2752.4452.7459.2332.2539.7236.4343.4642.6449.4327.3333.04
ConvNeXt-base48.0454.4554.7462.4831.6239.7035.2643.4641.9148.5626.3532.12
ConvNeXt-large50.8957.0056.6563.5130.1537.7035.2042.9143.8050.6725.4630.79
ViT-base41.0048.8043.8950.9920.6024.9822.6827.5935.6542.3119.3222.64
DINO-small58.8566.1059.5067.3533.5541.2742.5251.1146.9453.9328.0233.49
DINO-base57.2865.3361.3969.6733.0741.8342.3951.2347.5255.3428.2334.00
DeiT-small47.0055.9952.0860.6429.3636.7035.4544.4339.9347.6826.1832.14
DeiT-base53.7061.4855.1263.8132.7140.6837.6746.8044.2152.5727.2034.13
CLIP-base42.9851.3046.5255.0129.7536.9236.9144.5134.4541.2927.8134.02
GPT-4V37.78-31.11-20.00-20.00-21.11-23.33-
Human Expert35.00-43.75-20.00-18.75-33.33-19.44-
+ +Table 2. Few-shot image classification performance on LUWA dataset is associated with the magnification and sensing modality. 'PreTr' denotes pre-trained models we used; '20X' and '50X' denote microscopic images at $20 \times$ and $50 \times$ magnifications; 'TEX' and 'HM' denote texture and heightmap; '9w5s' and '9w20s' denote 9-way-5-shot and 9-way-20-shot, respectively. + +recognizes important polished regions in microscopic images of beechwood (see Fig. 5) as the foreground, but recognizes the same polish of different categories ( sprucewood, bone, and antler, see Fig. 5) as the background. More explorations are needed to explain this unwanted behavior. Interestingly, we found that the regions recognized as highly important for classification are similar for human experts and our best model ResNet152, under the same data configuration (see Regions 1, 4, and 6 in Fig. 2). We visualize the results for ResNet152 using Grad-CAM [54]. + +# 4.2. Few-Shot Image Classification + +In practice, LUWA faces a scarcity of microscopic images due to limited stone tools, expertise requirements, and specialized equipment, especially when discovering new categories. Human experts can identify new classes of wear traces with a few examples. To emulate this, we investigate whether few-shot learning can be utilized and how microscopic image magnifications and sensing modalities influence the model's performance. + +Experimental Settings. We designed two main experiments: (i) Few-shot image classification with a simple but effective pre-train + ProtoNet pipeline [27]. We evaluate the performance of powerful pre-trained models (including ResNet, ViT, DINO [10], ConvNeXt, CLIP [46], DEIT [63]) and popular meta-learners ProtoNet [57] on LUWA dataset. We simulated 600 episodes/tasks and results are demonstrated under 9-way-5/20-shot settings. (ii) GPT-4V [42] experiments: few-shot image classification and reasoning following instructions from human experts. To explore the potential mode of AI-human collaboration facing scientific domains in the advent of large multi-modal models, we collected prompts from three archaeologists and conducted case studies on whether the latest GPT-4V can + +follow and mimic the experts' reasoning process when analyzing the samples. Then we summarized key points that matter during the experts' analysis and used that to prompt the GPT-4V for few-shot image classification and reasoning. The experiments are illustrated in Fig. 6. Additionally, we included human experts' test results to reflect the difficulty for humans to distinguish these discovered categories with just a few examples. Results are reported under 9-way-5-shot settings in Tab. 2. + +Results and Analysis. Experimental results in Tab. 2 demonstrated that DINO excels at few-shot learning classification. Note that in the case of limited microscopic images, classification results of textures at $50 \times$ magnification setting yield notably superior results compared to others, which provides valuable guidance for few-shot learning tasks in our domain. Moreover, the number of parameters in pre-trained models has a limited impact on this few-shot learning task. We found that GPT-4V can effectively follow the experts' analysis as highlighted in Fig. 6. It learns to emphasize the same points that the experts pay attention to. However, the analysis doesn't always lead to the correct answer. GPT-4V did poorly on few-shot classification. We hypothesize that our data is very different from the web data that GPT models are trained on, and the vision module in GPT-4V still struggles to efficiently present detailed vision information to the language module, especially in a long context such as our multi-image few-shot classification scenario. This means the vision ability of SOTA multi-modal language models still needs improvement before they can be used in scientific tasks with domain-specific data like ours. + +# 5. Impact and Limitations of LUWA Dataset + +Scientific Impacts. AI-expert collaboration can provide invaluable insights for scientific research. To tackle the + +![](images/ff195c3b5bb4b5a626062a3f354df79507ccea8bb39caede5f9b45e82d8a10ed.jpg) + +![](images/0649381b2e18effd5b55a428d2e22c6ef97d1a53246b4335d468a56911053452.jpg) +(a) Sprucewood + +![](images/c816ed627a994dd3c605e9d496e962a84203ee20ad86233048c4878fbf1d21c4.jpg) + +![](images/728e739159a84a453005f81e7a4c107455d69a5db5d0887429f9b029f0ad7e5e.jpg) +Figure 5. Feature visualization of LUWA dataset using frozen pre-trained DINOv2. + +![](images/48f22d766c9fe098378d65d3b8987f32331694015f1e1ff7a4b4bee4b6c4b9b2.jpg) + +![](images/95d91a33759821101f9a261041d2b1a2c2a2b341b6d28c6bc1a7eccdc06462ea.jpg) +(b) Beechwood + +![](images/f1771db2c52c69977a0de6c39ffe6e6ea58c6daa17adb349560101f4b18b4d2c.jpg) + +![](images/1d478f3862446b768a0730c084a6daa59433310fae6f3829f6fe4001c506dba5.jpg) + +![](images/09e8dc111431146e0b6ba110d5620d618f7b9fa039f0a72679ad27790dabdc8f.jpg) +(c) Bone + +![](images/7c5f05649a664517010fd672d5ff72194eba550355c494cf8eebfef5389c932e.jpg) + +tone + +![](images/a55f425e2e357b6248e93f9c4a43b3166a2bafc783830a5352602dc31af94fc7.jpg) +(d) Antler + +![](images/d0b22da14fecc4345b82fc0c51ed9841bfa72c8f060c755b9db0663df6551a24.jpg) + +# Prompt: Expert Analysis + +![](images/733040ec0b23e5b21d4d857ad3d457cb451e82a4009d41826651e05f35fc34dd.jpg) + +![](images/2844ee651e5d99c9a1b503489929e9345c2b9f5c660548eed599d1f914d33724.jpg) + +First image is more polished I would argue it was used in softer wood (spruce). ... softer materials create larger polishes region. I see linear traces like striated marks ... does not appear that way for harder materials. + +Second image … +beechwood. + +Beechwood is a harder material … smaller region of polish … no linear traces seen in the first image. + +# GPT-4V Analysis + +![](images/e5d9ecf209b0b29afa3e3d5a726a23d916f1c678c19c6a391129dddca6077991.jpg) + +It appears to have a smaller region of polish with sharp bumps and lacks the linear, striated marks that would indicate a softer material like sprucewood. Following the anthropologist's analysis: + +- Sprucewood, being softer, would create a larger, more distributed polish and possibly show linear traces. +- Beechwood, being harder, would leave smaller and less extensive polish, without the linear traces. + +The characteristics of the wear in the image suggest the use of beechwood, given the absence of widespread polished areas or striated marks. This aligns with the anthropologist's explanation that harder materials like beechwood tend to create less polish and more pronounced features on the wear surface. + +# Few shot Image Classification with GPT4-V + +# User Prompt: + +An archaeologist conducting lithic use-wear analysis would examine ... + +Polishing... +- Striations +- Microflaking + +Task Context + +9-8 + +# + +This is an image taken with advanced microscopes containing traces created by material $\{\mathbf{x}\}$ , so the use-type is $\{\mathbf{x}\}$ . + +![](images/2736326d76845e0b369973c663d564424c408786e30c032f23540829e06a4af7.jpg) + +![](images/6575579de5ed214b8d2cce11201029a332105dc27204649318a0ac5e5cc9dd87.jpg) + +![](images/6b365a1170bb549cdd3ea4bc94ee133d659f89d087dcbe35b7f90cc82a179f70.jpg) + +![](images/a1d3e1ac4d69edd7a67ba91bac032542cce7e631230629f1c35cccd2a604f909.jpg) + +![](images/d032f2fd318cf8ffb66af384fd8c40fe7201b5115425c2ea4ff24acc1e7f6f1c.jpg) + +![](images/41acc77284b3dfd7fc3f413225373474890f5680bd1c7fd0100cb820db023e3b.jpg) + +{Antler} + +{Sprucewood} {Barley} + +I + +e} {Ivory} + +![](images/9553326c07ab650062c19ed4d34be186777fdc154dcf1659f1c84479e8147afc.jpg) + +This is another image, do the traces on this image belong to which of the nine use-types as demonstrated above? Answer in exactly one word: + +GPT-4V Answer: Sprucewood. + +Query + +long-standing problem of stone tool use, we make the first attempt to collaborate with archaeologists and utilize advanced learning-based methods for worked material inference. The LUWA dataset allows for further investigations to advance our understanding of ancient tool use and material processing techniques. + +Limitations and Future Directions. We will enrich the LUWA dataset from the following three aspects: (i) supplement microscopic images with both worn and unworn regions using lower-magnification objectives, allowing wear trace segmentation and detection tasks; (ii) collect images including wear traces caused by different worked material; (iii) increase categories of man-made materials for comprehensive analysis on wear features and material properties. + +# 6. Conclusion + +We collaborate with anthropological archaeologists and present the first public and the largest Lithic Use-Wear Analysis (LUWA) dataset benefiting both vision and science domains. The LUWA dataset serves as a benchmark + +to evaluate the generalization capabilities of advanced models on image classification tasks beyond common objects. Addressing specific challenges of wear formation and microscopic imaging, LUWA offers vital guidance on selecting suitable magnifications and sensing modalities facing different scenarios. Our analysis reveals that SOTA models encounter distinct difficulties when facing these specific challenges. Despite DINOV2's superior performance relative to other methods, it overlooks visual features that archaeologists identify as indicative of wear. We anticipate that the LUWA dataset will stimulate further research into enhancing the adaptability of large-scale models to specialized domains within computer vision. + +Acknowledgment. This work is supported by NSF Grant 2152565, and by NYU IT High-Performance Computing resources, services, and staff expertise. We gratefully acknowledge Dr. Rakesh Behera for the tribometer hardware, and thank Sara Borsodi, Felix Devis Kisena, Kat Liu, Eugenia Ochoa, Vita Jackman Kuwabara, Alice Jiang, Meiyu Zhang, and Sriram Koushik for their valuable assistance in collecting the microscopic images. + +# References + +[1] Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. ArXiv, abs/1610.01644, 2016. 5 +[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 2 +[3] Reza Akbarian Bafghi and Danna Gurari. A new dataset based on images taken by blind people for testing the robustness of image classification models trained for imagenet categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16261-16270, 2023. 2, 4 +[4] Douglas B Bamforth. Investigating microwear polishes with blind tests: the institute results in context. Journal of Archaeological Science, 15(1):11-23, 1988. 3 +[5] Tomasz Bartkowiak and Christopher A Brown. Multiscale 3d curvature analysis of processed surface textures of aluminum alloy 6061 t6. Materials, 12(2):257, 2019. 3 +[6] Sebastiano Battiato, Alessandro Ortis, Francesca Trenta, Lorenzo Ascari, Mara Politi, and Consolata Siniscalco. Detection and classification of pollen grain microscope images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 980-981, 2020. 3 +[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[8] Wonmin Byeon, Manuel Dominguez-Rodrigo, Georgios Arampatzis, Enrique Baquedano, José Yravedra, Miguel Angel Mate-González, and Petros Koumoutsakos. Automated identification and deep classification of cut marks on bones and its paleoanthropological implications. Journal of Computational Science, 32:36-43, 2019. 2, 3 +[9] Ivan Calandra, Lisa Schunk, Alice Rodriguez, Walter Gneisinger, Antonella Pedergnana, Eduardo Paixao, Telmo Pereira, Radu Iovita, and Joao Marreiros. Back to the edge: relative coordinate system for use-wear analysis. Archaeological and Anthropological Sciences, 11:5937-5948, 2019. 4 +[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 7 +[11] Claire Lifan Chen, Ata Mahjoubfar, Li-Chia Tai, Ian K Blaby, Allen Huang, Kayvan Reza Niazi, and Bahram Jalali. Deep learning in label-free cell classification. Scientific reports, 6(1):21471, 2016. 3 +[12] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the + +wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 3 +[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 5 +[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5 +[15] Aniket Dashpute, Vishwanath Saragadam, Emma Alexander, Florian Willomitzer, Aggelos Katsaggelos, Ashok Veeraraghavan, and Oliver Cossairt. Thermal spread functions (tsf): Physics-guided material classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1641-1650, 2023. 3 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5 +[18] Manuel S Drehwald, Sagi Eppel, Jolina Li, Han Hao, and Alan Aspuru-Guzik. One-shot recognition of any material anywhere using contrastive learning with physics-based rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23524-23533, 2023. 3 +[19] Adrian Anthony Evans. On the importance of blind testing in archaeological science: the example from lithic functional studies. Journal of Archaeological Science, 48:5-14, 2014. 3, 4, 6 +[20] Adrian A Evans and Randolph E Donahue. Laser scanning confocal microscopy: a potential technique for the study of lithic microwear. Journal of Archaeological Science, 35(8): 2223-2230, 2008. 3 +[21] Andrea Ferreri, Silvia Bucci, and Tatiana Tommasi. Multi-modal rgb-d scene recognition across domains. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2199-2208, 2021. 2 +[22] Richard LK Fullagar. The role of silica in polish formation. Journal of Archaeological Science, 18(1):1-24, 1991. 4 +[23] Eric Hayman, Barbara Caputo, Mario Fritz, and Jan-Olof Eklundh. On the significance of real-world conditions for material classification. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part IV 8, pages 253-266. Springer, 2004. 3 +[24] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 5 +[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1026-1034, 2015. 5 + +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[27] Shell Xu Hu, Da Li, Jan Stühmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068–9077, 2022. 7 +[28] Juan José Ibanez, Talía Lazuen, and J González-Urquijo. Identifying experimental tool use through confocal microscopy. Journal of Archaeological Method and Theory, 26:1176-1215, 2019. 3 +[29] Hartland W Jackson, Jana R Fischer, Vito RT Zanotelli, H Raza Ali, Robert Mechera, Savas D Soysal, Holger Moch, Simone Muenst, Zsuzsanna Varga, Walter P Weber, et al. The single-cell pathology landscape of breast cancer. Nature, 578 (7796):615-620, 2020. 3 +[30] L.H. Keeley. Experimental Determination of Stone Tool Uses: A Microwear Analysis. University of Chicago Press, 1980. 1, 3 +[31] Larry R Kimball, John F Kimball, and Patricia E Allen. Microwave polishes as viewed through the atomic force microscope. Lithic Technology, pages 6-28, 1995. 3 +[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5 +[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6 +[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5 +[35] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 5 +[36] João Marreiros, Ivan Calandra, Walter Gneisinger, Eduardo Paixão, Antonella Pedergnana, and Lisa Schunk. Rethinking use-wear analysis and experimentation as applied to the study of past hominin tool use. Journal of Paleolithic Archaeology, 3:475-502, 2020. 3 +[37] João Marreiros, Telmo Pereira, and Radu Iovita. Controlled experiments in lithic technology and function, 2020. 2, 3 +[38] Shannon P McPherron, Zeresenay Alemseged, Curtis W Marean, Jonathan G Wynn, Denné Reed, Denis Geraads, René Bobe, and Hamdallah A Béarat. Evidence for stonetool-assisted consumption of animal tissues before 3.39 million years ago at dikika, ethiopia. Nature, 466(7308):857-860, 2010. 2 + +[39] Mark Newcomer, Roger Grace, and Romana Unger-Hamilton. Investigating microwear polishes with blind tests. Journal of Archaeological Science, 13(3):203-217, 1986. 3 +[40] George H Odell. Stone tool research at the end of the millennium: classification, function, and behavior. Journal of Archaeological Research, 9:45-100, 2001. 2 +[41] George Hamley Odell and Frieda Odell-Vereecken. Verifying the reliability of lithic use-wear assessments by 'blind tests': the low-power approach. Journal of field Archaeology, 7(1):87-120, 1980. 1, 2, 3 +[42] OpenAI. Chatgpt can now see, hear, and speak, 2023. Accessed: 2023-11-16. 4, 7 +[43] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 5 +[44] Antonella Pedergnana and Andreu Olle. Monitoring and interpreting the use-wear formation processes on quartzite flakes through sequential experiments. *Quaternary International*, 427:35–65, 2017. 2 +[45] Marcos Pizarro-Monzo, Jordi Rosell, Anna Rufa, Florent Rivals, and Ruth Blasco. A deep learning-based taphonomical approach to distinguish the modifying agent in the late pleistocene site of toll cave (barcelona, spain). *Historical Biology*, pages 1–10, 2023. 2 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[47] Alice Rodriguez, Kaushik Yanamandra, Lukasz Witek, Zhong Wang, Rakesh K Behera, and Radu Iovita. The effect of worked material hardness on stone tool wear. Plos one, 17(10):e0276166, 2022. 3, 4 +[48] Z Rymuza. Tribology of polymers. Archives of civil and mechanical engineering, 7(4):177-184, 2007. 3 +[49] Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale image classification. In CVPR 2011, pages 1665-1672, 2011. 5 +[50] Linus Scheibenreif, Michael Mommert, and Damian Borth. Masked vision transformers for hyperspectral image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2165-2175, 2023. 3 +[51] Patrick Schmidt, Alice Rodriguez, Kaushik Yanamandra, Rakesh K Behera, and Radu Iovita. The mineralogy and structure of use-wear polish on chert. Scientific reports, 10 (1):21512, 2020. 3 +[52] Benjamin J Schoville, Kyle S Brown, Jacob A Harris, and Jayne Wilkins. New experiments and a model-driven approach for interpreting middle stone age lithic point function using the edge damage distribution method. PloS one, 11 (10):e0164088, 2016. 2 +[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294. Curran Associates, Inc., 2022. 5 +[54] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4, 7 +[55] S.A. SEMENOV. Prehistoric Technology. 1964. 3 +[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4 +[57] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 7 +[58] W James Stemp and Michael Stemp. Ubm laser profilometry and lithic use-wear analysis: a variable length scale investigation of surface topography. Journal of Archaeological Science, 28(1):81-88, 2001. 3 +[59] W James Stemp, Adam S Watson, and Adrian A Evans. Surface analysis of stone and bone tools. Surface Topography: Metrology and Properties, 4(1):013001, 2015. 3 +[60] Nathan E Stevens, Douglas R Harro, and Alan Hicklin. Practical quantitative lithic use-wear analysis using multiple classifiers. Journal of Archaeological Science, 37(10):2671-2678, 2010. 3 +[61] Kenichiro Tanaka, Yasuhiro Mukaigawa, Takuya Funatomi, Hiroyuki Kubo, Yasuyuki Matsushita, and Yasushi Yagi. Material classification using frequency- and depth-dependent time-of-flight distortion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3 +[62] SH Teoh. Fatigue of biomaterials: a review. International journal of fatigue, 22(10):825-837, 2000. 2 +[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7 +[64] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, page 53, 2019. 3 +[65] Hongda Wang, Hatice Ceylan Koydemir, Yunzhe Qiu, Bijie Bai, Yibo Zhang, Yiyin Jin, Sabiha Tok, Enis Cagatay Yilmaz, Esin Gumustekin, Yair Rivenson, et al. Early detection and classification of live bacteria using time-lapse coherent imaging and deep learning. Light: Science & Applications, 9(1):118, 2020. 3 + +[66] Ran Wang, Yuanjing Zhu, Chengxin Chen, Yu Han, and Hongbo Zhou. Tooth wear and tribological investigations in dentistry. Applied Bionics and Biomechanics, 2022, 2022. 2 +[67] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 3 +[68] Samuel J Yang, Marc Berndl, D Michael Ando, Mariya Barch, Arunachalam Narayanaswamy, Eric Christiansen, Stephan Hoyer, Chris Roat, Jane Hung, Curtis T Rueden, et al. Assessing microscope image focus quality with deep learning. BMC bioinformatics, 19:1-9, 2018. 3 \ No newline at end of file diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/images.zip b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..091d92587deb5c293b3542e9f52a5cb4b1e9953e --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11881ee44541e61b24bf50f0463bdc90b54fdb808e1069280e18e18d36d8bf72 +size 754678 diff --git a/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/layout.json b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dcf9f2b8612a335463aca58b0b6c05a860b926e7 --- /dev/null +++ b/2024/LUWA Dataset_ Learning Lithic Use-Wear Analysis on Microscopic Images/layout.json @@ -0,0 +1,9071 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 530, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 530, + 122 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 530, + 122 + ], + "type": "text", + "content": "LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": "Jing Zhang" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Irving Fang" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Hao Wu" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Akshit Kaushik" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Alice Rodriguez" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Hanwen Zhao" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Juexiao Zhang" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Zhuo Zheng" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Radu Iovita" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", Chen Feng" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 75, + 142, + 517, + 186 + ], + "type": "text", + "content": ", New York University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 215, + 188, + 374, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 188, + 374, + 198 + ], + "spans": [ + { + "bbox": [ + 215, + 188, + 374, + 198 + ], + "type": "text", + "content": "https://ai4ce.github.io/LUWA/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 206, + 154, + 430 + ], + "blocks": [ + { + "bbox": [ + 63, + 206, + 154, + 430 + ], + "lines": [ + { + "bbox": [ + 63, + 206, + 154, + 430 + ], + "spans": [ + { + "bbox": [ + 63, + 206, + 154, + 430 + ], + "type": "image", + "image_path": "90bf1edf98efabb4f624952e97e9e1986bcba9feb09c9fa497ad88f2b266a309.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 441, + 547, + 475 + ], + "lines": [ + { + "bbox": [ + 45, + 441, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 45, + 441, + 547, + 475 + ], + "type": "text", + "content": "Figure 1. LUWA poses a unique computer vision challenge due to: its complex wear formation and irregular wear patterns, ambiguous sensing modalities and magnifications in microscopic imaging. Facing these challenges, the LUWA dataset encompasses both texture and heightmap with different magnifications, encouraging the exploration of image classification beyond common objects." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 164, + 206, + 266, + 428 + ], + "blocks": [ + { + "bbox": [ + 164, + 206, + 266, + 428 + ], + "lines": [ + { + "bbox": [ + 164, + 206, + 266, + 428 + ], + "spans": [ + { + "bbox": [ + 164, + 206, + 266, + 428 + ], + "type": "image", + "image_path": "9c5d082b4abe8c0afaaaa5096928c76addb5ef8128cad1dfe18ed7d44bf574a8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 271, + 204, + 529, + 340 + ], + "blocks": [ + { + "bbox": [ + 271, + 204, + 529, + 340 + ], + "lines": [ + { + "bbox": [ + 271, + 204, + 529, + 340 + ], + "spans": [ + { + "bbox": [ + 271, + 204, + 529, + 340 + ], + "type": "image", + "image_path": "24e4e8f23090d87e9f0549de6e4464e21a3c8a9c1946e0a7813110ffa5d1b639.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 271, + 345, + 529, + 430 + ], + "blocks": [ + { + "bbox": [ + 271, + 345, + 529, + 430 + ], + "lines": [ + { + "bbox": [ + 271, + 345, + 529, + 430 + ], + "spans": [ + { + "bbox": [ + 271, + 345, + 529, + 430 + ], + "type": "image", + "image_path": "2f254019f55b495f1d9712d1e46158be6ca39c86cec115693282a96774ad4ab9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "spans": [ + { + "bbox": [ + 143, + 479, + 192, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 506, + 288, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 506, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 45, + 506, + 288, + 673 + ], + "type": "text", + "content": "Lithic Use-Wear Analysis (LUWA) using microscopic images is an underexplored vision-for-science research area. It seeks to distinguish the worked material, which is critical for understanding archaeological artifacts, material interactions, tool functionalities, and dental records. However, this challenging task goes beyond the well-studied image classification problem for common objects. It is affected by many confounders owing to the complex wear mechanism and microscopic imaging, which makes it difficult even for human experts to identify the worked material successfully. In this paper, we investigate the following three questions on this unique vision task for the first time:(i) How well can state-of-the-art pre-trained models (like DINOv2) generalize to the rarely seen domain? (ii) How can few-shot" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 481, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 547, + 625 + ], + "type": "text", + "content": "learning be exploited for scarce microscopic images? (iii) How do the ambiguous magnification and sensing modality influence the classification accuracy? To study these, we collaborated with archaeologists and built the first open-source and the largest LUWA dataset containing 23,130 microscopic images with different magnifications and sensing modalities. Extensive experiments show that existing pretrained models notably outperform human experts but still leave a large gap for improvements. Most importantly, the LUWA dataset provides an underexplored opportunity for vision and learning communities and complements existing image classification problems on common objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 633, + 387, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 387, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 387, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Lithic Use-Wear Analysis (LUWA) is a long-standing scientific problem (see Fig. 1) to identify the functions of stone tools by examining wear traces at the microscopic level on the tool's surface [30, 41]. It seeks to distinguish the worked material (like bone, wood, ivory and antler) us-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 684, + 124, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 684, + 124, + 693 + ], + "spans": [ + { + "bbox": [ + 58, + 684, + 124, + 693 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "type": "text", + "content": "†Corresponding author: Radu Iovita (iovita@nyu.edu) and Chen Feng (cfeng@nyu.edu)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22563" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "ing microscopic images, creating a classification problem. Investigating this unanswered vision-for-science problem will provide invaluable insights for uncovering the hidden story of ancient tools [38] and advancing the understanding of material interactions [62, 66]. However, few studies have begun to explore advanced learning-based methods, and this field is still in a nascent stage [8, 45]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 290, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 290, + 458 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 290, + 458 + ], + "type": "text", + "content": "Besides its potential scientific impact, learning-based LUWA on microscopic images also poses unique computer vision challenges beyond common objects. Unlike common objects with distinct boundaries and typical size, wear traces in microscopic images are irregular and discontinuous (see Fig. 1), making it difficult to define clear visual characteristics [40]. In particular, their absence of clear foreground and background might increase the difficulty of feature extraction. Furthermore, recorded wear traces are a hybrid consequence of the complex wear process and microscopic imaging, which is affected by many factors [41]. For example, the wear duration and the motion type that generates the wear can create significant intra-class variability. Another crucial aspect is the ambiguous magnification and sensing modality of microscopic images. High-sensitivity microscopic imaging allows for capturing complementary sensing modalities and zoom-in details at different magnifications, respectively (see Fig. 1). We found that their modality and magnification differences lead to very different visual features. However, due to the understudied characteristics of wear traces, there has been no conclusive answer to this vision-for-science problem about which magnification and sensing modality is better. In practice, the accessibility of expensive equipment also limits the flexible selection of captured data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 288, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 675 + ], + "type": "text", + "content": "Recent advances in computer vision provide a good opportunity for this challenging scientific problem, especially the emerging paradigm shift with foundation models pretrained on large-scale data [2, 7]. These foundation models generate task-agnostic visual features and have shown excellent performance on image classification with common datasets like ImageNet-" + }, + { + "bbox": [ + 46, + 460, + 288, + 675 + ], + "type": "inline_equation", + "content": "\\{1k, A\\}" + }, + { + "bbox": [ + 46, + 460, + 288, + 675 + ], + "type": "text", + "content": " [43]. But how much can state-of-the-art (SOTA) pre-trained representations benefit rarely seen domains not extensively covered by uncurated data available on the Internet? The lack of diversity in domain-specific datasets leaves uncertainty about this question, particularly when addressing real-world vision challenges. Unfortunately, existing studies on LUWA focus on private data and the research community faces a deficiency in accessible datasets. The complexity of wear forms, microscopic imaging, equipment limitations, and the need for expert interpretations hinders the collection of an appropriate dataset that can represent the variability of this domain." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "To explore these unique challenges and opportunities, we collaborated with anthropological archaeologists and built the first open-source and the largest LUWA dataset contain-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "ing 23,130 microscopic images. Its major characteristics are summarized as follows: (i) Multi-scale wear patterns. Lower and higher magnifications allow for an observation of pattern distributions and topographical details, respectively, which increases the scale variations. (ii) Complementary sensing modalities. Both grayscale microscopic images and corresponding 3D surface profiles are available to provide complementary texture information and geometry cues, helping to identify discriminative features [21]. (iii) Both machine and human wear processes. Samples of stone tools were collected from both machine and human wear processes. The former is to isolate the effect of worked material by tightly controlling other factors [37] while the latter is to reflect the complexity of real scenes, especially for newly discovered categories without baseline data [44, 52]. We envision that the LUWA dataset will encourage researchers to develop and evaluate algorithms for image classification beyond common objects, as a precursor for downstream tasks like segmentation and detection [3]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 303, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 411 + ], + "type": "text", + "content": "Based on the LUWA dataset, we go further in answering the following three problems facing real-world applications: (i) How well representative classification models can generalize to the rarely seen domain? (ii) How can few-shot learning be exploited when scarce microscopic images are available, especially for newly discovered categories? (iii) How do the ambiguous magnification and sensing modality of microscopic images influence the classification accuracy? Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 415, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 415, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 545, + 521 + ], + "type": "text", + "content": "- We introduce the first open-source and the largest LUWA dataset including 23,130 microscopic images with different magnifications and sensing modalities. We collaborated with archaeologists and set up a data collection pipeline considering the influence of wear formation, microscopic imaging, and expert knowledge. The LUWA dataset allows for reproducible investigations on this rarely seen domain and complements existing image classification problems on common objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 522, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 545, + 665 + ], + "type": "text", + "content": "- Facing the image classification problem beyond common objects, we benchmark the generalization capability of state-of-the-art vision models (ResNet, ViT, DINOv2, ConvNeXt) in this specific domain. Experimental results show that DINOv2 has the most stable performance amidst varying levels of granularity, magnification, and sensing modalities in the data. We also observed many trends regarding the impacts that features like magnification and sensing modality have on classification accuracy, and some of them are not consistent with experts' heuristics. In general, state-of-the-art computer vision models display super-human accuracy over domain experts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 665, + 545, + 713 + ], + "type": "text", + "content": "- Considering the scarcity of microscopic images in real scenarios, we investigate the performance of few-shot image classification and reasoning on the LUWA dataset. Particularly, we collected prompts from archaeologists" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22564" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "type": "text", + "content": "and did case studies on whether GPT-4V(ison) can mimic the experts' reasoning process. Further explorations are required to improve the performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "text", + "content": "Lithic Use-Wear Analysis. Lithic use-wear analysis was originally developed in the 1950s and aims to answer the scientific problem of how to distinguish the worked material according to wear traces on the stone surface [30, 41, 55]. Low-power and High-power microscopy methods provide useful magnified polishes or micro-fractures via confocal microscopes, tactile profilometers, scanning electron microscopes, and even atomic force microscopes [20, 31, 58]. Existing studies focus on blind tests [4, 19, 39] and quantification methods [4, 5, 20, 28, 59, 60]. However, it remains an insufficiently developed research area due to complex wear patterns and the subjectivity of methods [47]. Existing blind tests demonstrate unreliable identification results on different worked materials. For example, correct identifications of plant and wood are " + }, + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "inline_equation", + "content": "32.4\\%" + }, + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "inline_equation", + "content": "49.1\\%" + }, + { + "bbox": [ + 47, + 140, + 289, + 413 + ], + "type": "text", + "content": ", respectively [19]. The identification of worked materials can be regarded as a unique vision problem, but learning-based algorithms are rarely explored in this research area [8]. The deficiency in accessible datasets also limits its research progress heavily [19]. To solve these, we present the first open-source LUwA dataset and investigate the capability of SOTA classification algorithms on this unique vision-for-science problem for the first time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 415, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 289, + 665 + ], + "type": "text", + "content": "Image Classification beyond Common Objects. Image classification is a fundamental vision task that categorizes a whole image as a specific label or class based on its visual contents. Outstanding performance can be achieved on the image classification task of common objects with representative frameworks like ConvNeXt using only public training data [67]. However, real-world scenarios, especially scientific applications, involve non-trivial objectives with different physical properties like tiny pollen grains [6], constituent materials [15, 18, 61], texture in the wild [12], and hyperspectral remote sensing images [50]. Facing uncommon objects, many studies focus on specialized models in respective fields. However, we seek to answer whether simpler pipelines can provide comparable performance. Complementing existing datasets, we introduce a new vision challenge to identify irregular and discontinuous wear traces at the microscopic level and build a classification benchmark to explore how far the practical performance of image classification on uncommon objects can be pushed utilizing existing datasets and architectures, especially foundation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 715 + ], + "type": "text", + "content": "Microscopic Image Classification. In the context of using microscopic images for classification, a wide variety of applications can be found across different scientific disciplines. This includes the study of biological cells [11, 29]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "bacteria [65], tissue types [64], and material structures [23]. Each application presents unique challenges, particularly in terms of the high-level detail and focus required in the images. These challenges are often compounded by issues such as ambiguity in magnification and sensing modality [23, 68]. To alleviate these challenges, the LUWA Dataset presents a diverse configuration that covers different magnification levels and sensing modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 178, + 399, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 399, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 399, + 190 + ], + "type": "text", + "content": "3. LUWA Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": "In this section, we describe the LUWA dataset creation process and provide basic statistics. To represent the variability of this domain, in Section 3.1, we present a data collection pipeline, which consists of four key aspects (see Fig. 1). Specifically, ① considering the complexity of wear formation, we introduced both machine and human wear experiments [36] to create stone samples; ② to enrich the dataset diversity and investigate the ambiguous magnification and sensing modality, we utilized an optical 3D profilometer with both " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 304, + 198, + 547, + 437 + ], + "type": "text", + "content": " objective lenses to acquire high-quality texture and heightmap; ③ natural materials were selected according to existing blind tests in the literature [19], particularly including fine-grained categories of wood and plants. ④ domain-specific knowledge is twofold including the identification of wear degrees to increase the dataset diversity and expert interpretations for potential explorations on explainability and the application of vision language models. In Section 3.2, we summarized the LUWA dataset and analyzed its diversity in spatial distributions, magnifications, and sensing modalities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 444, + 408, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 408, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 408, + 455 + ], + "type": "text", + "content": "3.1. Dataset Creation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 462, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 665 + ], + "type": "text", + "content": "Both Machine and Human Wear Processes. LUWA dataset contains stone samples from both machine and human wear processes. Key factors that affect wear results are material properties, mechanical factors, and environmental conditions [48]. To isolate the effect of worked materials, a tightly controlled protocol was used for machine wear experiments [36, 37]. We utilized a tribometer to simulate cutting actions so as to quantify the load applied to the material (a load of 20N), the type of movement (a straight back and forth motion with the speed of 35 repetitions per minute), and the worked duration (Oh, 1h, 3h, 5h, and 12h) (see Fig. 1). To limit the influence of material properties of stone samples, we chose the same flint (Baltic/morainic flint from Denmark) for all experiments [47, 51]. Considering the low classification accuracy of various plants in blind tests [19], we chose the plant cutting process as the human wear experiment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "High-Quality Microscopic Imaging. To capture the high-precision wear traces on stone samples, we utilized an optical 3D profilometer (S neox, Sensorfar Metrology) to collect data with a standardized and reproducible process (see" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22565" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 138, + 288 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 138, + 288 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 138, + 288 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 138, + 288 + ], + "type": "image", + "image_path": "0e7f10f1f12cf472b1951873d600cc83e0f9e841f8d403c0dea862c53fdf066c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 298, + 545, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 298, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 545, + 373 + ], + "type": "text", + "content": "Figure 2. Image diversity of LUwA dataset and corresponding visual explanations for human and model decision-making processes. (i) LUwA dataset provides diverse microscopic images associated with spatial distributions (e.g. Regions 1 and 2), magnifications (e.g. Regions 2 and 4) and sensing modalities (texture in the first row and heightmap in the second row); (ii) We compared visual explanations in both human (in the third row) and model (in the fourth row) decision-making processes. Human experts labeled the most important region with red and the less important region with yellow when looking at details of microscopic images to distinguish the worked material. Similarly, Grad-CAM [54] heatmaps use red for the highest importance, yellow for lower importance, and blue for the lowest importance. Interestingly, similar areas (e.g. Regions 1, 4 and 6) are labeled with higher importance for both humans and models." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 73, + 542, + 285 + ], + "blocks": [ + { + "bbox": [ + 140, + 73, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 140, + 73, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 140, + 73, + 542, + 285 + ], + "type": "image", + "image_path": "fbc0531bbf633635a3888c4310b691159cf89f577853140c8b40389b8015517f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": "Fig. 1). To test the influence of magnifications for microscopic image classification, both " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "50\\times" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " objective lenses were chosen for measurements. Their spatial resolutions are 0.65 and " + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "inline_equation", + "content": "0.26\\mu \\mathrm{m}" + }, + { + "bbox": [ + 46, + 381, + 287, + 477 + ], + "type": "text", + "content": " pixel, respectively. Furthermore, complementary grayscale images and corresponding 3D surface profiles are acquired via Sensormap. We applied a standard filtering protocol to extract the worn surface and alleviate the effect of natural flints' surface topography [9]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 478, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 287, + 574 + ], + "type": "text", + "content": "Domain-Specific Expert Knowledge. Domain-specific knowledge is twofold: (i) human experts help to identify microscopic traces with different wear degrees, enriching the dataset diversity; (ii) for further investigations on explainability and the application of vision language models, human experts also labeled their attention maps when making decisions on worked material (see Fig 1) and provided classification prompt for GPT-4V [42] (see Fig. 6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 575, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 575, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 575, + 287, + 659 + ], + "type": "text", + "content": "Material Selection and Processing. To benchmark models in this field, we chose representative natural materials (see Fig. 1) according to blind test results in the literature [19] and included fine-grained categories on wood and plants in particular. For the further exploration of wear mechanisms, we analyzed their properties, including the hardness [47] and silicon content [22] in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 147, + 682 + ], + "type": "text", + "content": "3.2. Dataset Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "To reflect the variability of this scientific domain, we built the first public and largest LUwA dataset containing 23,130" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 380, + 545, + 422 + ], + "blocks": [ + { + "bbox": [ + 308, + 380, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 308, + 380, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 380, + 545, + 422 + ], + "type": "table", + "html": "
Stone SamplesMotion TypesWorked TimeMaterial Categories
34279
MagnificationsSensing Modalities
22
", + "image_path": "e56b81e876b6bf122e70f79ba706093f68ae64bbc67a59323fdabfe965365600.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 424, + 545, + 447 + ], + "lines": [ + { + "bbox": [ + 306, + 424, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 424, + 545, + 447 + ], + "type": "text", + "content": "Table 1. Key factors considered in the LUWA dataset that can reflect the complex wear formation and microscopic imaging." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 542 + ], + "type": "text", + "content": "microscopic images. Specifically, key factors of the complex wear formation and microscopic imaging are considered in the LUWA dataset. As shown in Tab. 1, we report (i) the number of microscopic images, (ii) the number of stone samples, motion types, worked time, and material classes which are exploited, (iii) the number of magnifications and sensing modalities LUWA dataset supports." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "Image diversity of LUWA dataset brings challenges for algorithm robustness. It is associated with the spatial distribution of the region collected, the selection of the magnification, and the sensing modality. Greater distances between sampled areas typically result in more pronounced variations in their surface distributions (see Regions 1 and 2 in Fig. 2). Even collected from the same wear trace, the selection of the magnification also contributes to scale difference, which causes totally different wear patterns (see Regions 2 and 4 in Fig. 2). Moreover, LUWA dataset provides both the texture and heightmap, helping to identify discriminative features. We explore the semantic diversity of LUWA dataset on the magnification and sensing modality (see Fig. 2) [3]. We selected VGG [56], ResNet [26]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22566" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 75, + 279, + 173 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 279, + 173 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 279, + 173 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 279, + 173 + ], + "type": "image", + "image_path": "6524fac1bc150096ee6ef536f5b3306fb3a3eea2077df3cc9c5444b6c24ceaca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 287, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 287, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 287, + 210 + ], + "type": "text", + "content": "Figure 3. Cosine similarity distribution of LUWA dataset on different magnifications and sensing modalities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 287, + 276 + ], + "type": "text", + "content": "ConvNeXt [34], and DINOv2 [10, 14, 43] as feature extractors. Then we compute the mean cosine distance of images with different magnifications and sensing modalities, respectively. Scale difference leads to obvious diversity of semantic information and visual descriptions (see Fig. 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 285, + 195, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 195, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 195, + 300 + ], + "type": "text", + "content": "4. Algorithm Benchmarking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 287, + 413 + ], + "type": "text", + "content": "By benchmarking a wide range of image classification methods, both classic and state-of-the-art, on this unique vision-for-science dataset, we explore how different features of this dataset affect model performance and hope that we can provide some useful analysis that future work can build on. Specifically, we divide our experiments into two major segments: (1) fully-supervised image classification and (2) few-shot image classification, with specific motivations explained in their corresponding sections below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 420, + 247, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 247, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 247, + 433 + ], + "type": "text", + "content": "4.1. Fully-Supervised Image Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 581 + ], + "type": "text", + "content": "Unlike datasets crawled from the internet, such as ImageNet-1k [16] or LAION-5B [53], LUWA contains niche microscopic images with irregular and discontinuous wear traces that often lack obvious foreground or background. In this experiment, we investigate how well classic and state-of-the-art image classification algorithms generalize to the LUWA dataset and seek to find compelling patterns affecting different models' performance. We want to see how well these patterns align with domain experts' knowledge. We also aim to position SOTA methods from the computer vision community with respect to classification performance achieved by human experts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 688 + ], + "type": "text", + "content": "Experimental Settings. We deploy classic methods such as SIFT+FVs [49], ResNets [24] and cutting-edge models such as ViT [17], ConvNeXts [34] and DINOv2 [13, 43] to our LUWA dataset. We believe the time-tested classic methods can serve as a lower-bound benchmark, while the more recent advancements such as DINOv2, which can often be characterized by intensive scale-up in parameter count, can be used as a reference comparable to human experts' performance on the same task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Another major reason that propels us to deploy this wide range of models is to see if there are consistent trends across" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "different model architectures and parameter counts and if these trends align with domain experts' knowledge. Specifically, we study the impact of image granularity, magnification, and sensing modality on image classification performance. We also compare different training strategies. By combining the following configurations, we have 324 total experiment results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "text", + "content": "Granularity refers to how many pictures one single use-wear is partitioned into. The use-wear is first captured as an image at " + }, + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "inline_equation", + "content": "865 \\times 865" + }, + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "text", + "content": " resolution. Because many pre-trained models resize the input, which results in pixelated images and loss of fine-grained details that many experts believe are crucial to such a classification task, we partition the original image into 24 or 6 patches and feed all the patches to the model. Importantly, to make our results comparable to those of human experts, we adopt a voting mechanism during test time. If the majority of the 24 patches are classified as class 1, then all the patches of the same use-wear will be classified as class 1 regardless of their actual test results. We believe this method most resembles how archaeologists perform classification when given an " + }, + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "inline_equation", + "content": "865 \\times 865" + }, + { + "bbox": [ + 304, + 157, + 546, + 337 + ], + "type": "text", + "content": " image, as they do not assign a label to each partition." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": "Magnification represents the magnification multiplier on our microscopic imaging equipment. Our dataset comprises images after " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " magnification. We also mix " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " data together without any indicator of magnification added to the data to see if mixing magnification will cause any confusion in our models. Note that the magnification multipliers are fixed when the images are taken, and " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " images will not look like " + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 338, + 546, + 445 + ], + "type": "text", + "content": " images, even if we significantly downsize them." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 529 + ], + "type": "text", + "content": "Sensing Modality refers to whether the picture of the use-wear is stored as texture scans or heightmaps. Texture scans have no depth information, although depth cues are still present. Meanwhile, heightmaps explicitly store the depth information and only the depth information. We want to see if different ways to represent the use-wear will affect computer vision models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 591 + ], + "type": "text", + "content": "Training Strategy is also varied in our experiments. Some models are initialized with state-of-the-art initialization methods [25] and then trained from scratch. We also apply full-parameter fine-tuning and linear probing [1] with unfrozen and frozen pre-trained weights, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 592, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 592, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 592, + 546, + 700 + ], + "type": "text", + "content": "Implementation Details. We deploy ResNet50 (25.6M parameters), ResNet152 (60.4M), ConvNeXt-tiny (28.6M), ConvNeXt-large (197.7M), ViT-H (632M), and DINOv2-ViT-g/14 with registers [13] (1.1B). All models are trained with Adam optimizer [32] on the default setting in PyTorch. We employ linear warmup with cosine annealing [35] as the learning rate scheduler strategy. No data augmentation technique is applied during pre-processing. We defer more details to our supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Results and Analysis. The overall experiment results can" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22567" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 71, + 289, + 195 + ], + "blocks": [ + { + "bbox": [ + 108, + 71, + 289, + 195 + ], + "lines": [ + { + "bbox": [ + 108, + 71, + 289, + 195 + ], + "spans": [ + { + "bbox": [ + 108, + 71, + 289, + 195 + ], + "type": "image", + "image_path": "8eb7baa60da76b56b7a0c4c7dc6c891a0aba7996337165938797ca965ee61f4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 481, + 194 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 481, + 194 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 481, + 194 + ], + "type": "image", + "image_path": "9d5e90ddcd5378f3e36a8dee7f2db970f949e91e19d3700a24e73a5099f6e7e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 109, + 207, + 287, + 329 + ], + "blocks": [ + { + "bbox": [ + 184, + 198, + 241, + 206 + ], + "lines": [ + { + "bbox": [ + 184, + 198, + 241, + 206 + ], + "spans": [ + { + "bbox": [ + 184, + 198, + 241, + 206 + ], + "type": "text", + "content": "(a) Training Strategy" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 207, + 287, + 329 + ], + "lines": [ + { + "bbox": [ + 109, + 207, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 109, + 207, + 287, + 329 + ], + "type": "image", + "image_path": "c69bf421866e2a3196552339535d214f28350352597722f331152e828c8b203e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 331, + 235, + 339 + ], + "lines": [ + { + "bbox": [ + 188, + 331, + 235, + 339 + ], + "spans": [ + { + "bbox": [ + 188, + 331, + 235, + 339 + ], + "type": "text", + "content": "(c) Magnification" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 341, + 478, + 353 + ], + "lines": [ + { + "bbox": [ + 129, + 341, + 478, + 353 + ], + "spans": [ + { + "bbox": [ + 129, + 341, + 478, + 353 + ], + "type": "text", + "content": "ResNet50 ResNet152 CovNeXt-tiny CovNeXt-large ViT-H DINOv2 SIFT + FVs" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 301, + 207, + 480, + 330 + ], + "blocks": [ + { + "bbox": [ + 348, + 198, + 452, + 206 + ], + "lines": [ + { + "bbox": [ + 348, + 198, + 452, + 206 + ], + "spans": [ + { + "bbox": [ + 348, + 198, + 452, + 206 + ], + "type": "text", + "content": "(b) Granularity (Partition Per Sample)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 207, + 480, + 330 + ], + "lines": [ + { + "bbox": [ + 301, + 207, + 480, + 330 + ], + "spans": [ + { + "bbox": [ + 301, + 207, + 480, + 330 + ], + "type": "image", + "image_path": "cc6dcd081af4e687872a2d44b59c3e0eb2a32440c49cfcdea8caf35e8cab30ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 331, + 429, + 339 + ], + "lines": [ + { + "bbox": [ + 372, + 331, + 429, + 339 + ], + "spans": [ + { + "bbox": [ + 372, + 331, + 429, + 339 + ], + "type": "text", + "content": "(d) Sensing Modality" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 357, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 547, + 390 + ], + "type": "text", + "content": "Figure 4. The impact of the training strategy, granularity, magnification, and sensing modality on top-1 classification accuracy in " + }, + { + "bbox": [ + 46, + 357, + 547, + 390 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 46, + 357, + 547, + 390 + ], + "type": "text", + "content": " : (a) Due to their huge parameter counts, the experiments do not include full-parameter fine-tuned DINOv2, and ViT-H and DINOv2 trained from scratch. (b) Larger numbers in granularity mean more detailed information about a use-wear is fed into the model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 396, + 289, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 708 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 708 + ], + "type": "text", + "content": "be found in Fig. 4. Fig. 4 (a) demonstrated that linear probing yields the most stable and optimal performance across the broad. These results align with our expectation as fine-tuning on uncommon domain-specific datasets may cause catastrophic forgetting [33]. On the other hand, the LUWA dataset itself is too small to train generalizable models from scratch. From the perspective of granularity (Fig. 4 (b)), the more granular partition of the images tends to result in better outcomes, although a diminishing marginal return can be observed. This aligns with our speculation that keeping the original larger image's information as intact as possible is beneficial. It is also possible that the introduction of a voting mechanism brought about a positive regularization effect. More discussion on the voting mechanism can be found in the supplementary material. Considering the selection of magnification, results in Fig. 4 (c) indicated that a higher magnification multiplier is beneficial, which aligns with some human experts' opinions. Notably, mixing data with different magnifications does not confuse the models, and they are able to reap the benefit of abundant data. The same cannot be said for humans, as images with different magnifications can cause confusion. For the sensing modality, we observed that while the best results are usually trained with heightmaps, larger models tend to favor texture. However, the discrepancy is small in general, and the overall performances of the two modalities are com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 546, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 433 + ], + "type": "text", + "content": "parable as in Fig. 4 (d). More visualizations and detailed tables that follow the trends described above can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "content": "The best performance of " + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "inline_equation", + "content": "67.05\\%" + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "content": " top-1 accuracy is achieved by linear probing on ImageNet-1K pre-trained ResNet152 with the heightmap data of 24 partitions and " + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "inline_equation", + "content": "20 \\times +50 \\times" + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "content": " magnification. Overall, DINOv2 excels across all aspects, demonstrating the most stable performance amidst varying levels of granularity, magnification, and sensing modalities. The worst performance of the traditional baseline SIFT+FVs is better than the worst configurations of other deep learning methods, but its better configurations are significantly worse than those of the deep learning methods that achieved upper-echelon performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "text", + "content": "Explainability and Comparison with Human Experts. Currently, archaeologists have about " + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "inline_equation", + "content": "49.5\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "text", + "content": " accuracy in a double-blind test with a similar setup [19]. Our in-house testing with two professional archaeologists yields an accuracy of " + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "inline_equation", + "content": "43.75\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "text", + "content": " in a few-shot setting (Tab. 2). All tested models, except SIFT+FVs, are able to achieve far better accuracy (over " + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "inline_equation", + "content": "59.5\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 687 + ], + "type": "text", + "content": ") under several configurations of the dataset. Notably, DINOv2 with linear probing is able to achieve superior or comparable performance under almost all possible configurations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "However, feature visualization demonstrated that DINOv2 (with registers for better visualization) sometimes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22568" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 535, + 253 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 535, + 253 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 535, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 535, + 253 + ], + "type": "table", + "html": "
PreTr20X TEX50X TEX20X HM50X HM20X+50X TEX20X+50X HM
9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s9w5s9w20s
ResNet1854.5461.9754.4362.4831.1938.7935.2742.3442.1149.6026.4331.80
ResNet5054.1359.2055.0862.1832.6738.9736.7143.9545.3751.4628.9134.56
ResNet15252.9259.1457.5964.2630.8338.7434.1241.4044.3251.4026.3931.89
ConvNeXt-tiny46.2752.4452.7459.2332.2539.7236.4343.4642.6449.4327.3333.04
ConvNeXt-base48.0454.4554.7462.4831.6239.7035.2643.4641.9148.5626.3532.12
ConvNeXt-large50.8957.0056.6563.5130.1537.7035.2042.9143.8050.6725.4630.79
ViT-base41.0048.8043.8950.9920.6024.9822.6827.5935.6542.3119.3222.64
DINO-small58.8566.1059.5067.3533.5541.2742.5251.1146.9453.9328.0233.49
DINO-base57.2865.3361.3969.6733.0741.8342.3951.2347.5255.3428.2334.00
DeiT-small47.0055.9952.0860.6429.3636.7035.4544.4339.9347.6826.1832.14
DeiT-base53.7061.4855.1263.8132.7140.6837.6746.8044.2152.5727.2034.13
CLIP-base42.9851.3046.5255.0129.7536.9236.9144.5134.4541.2927.8134.02
GPT-4V37.78-31.11-20.00-20.00-21.11-23.33-
Human Expert35.00-43.75-20.00-18.75-33.33-19.44-
", + "image_path": "6ed8c2cf1119049196c3b7a42bba06ece507ee1c3edb632f3afb17e36da363d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "type": "text", + "content": "Table 2. Few-shot image classification performance on LUWA dataset is associated with the magnification and sensing modality. 'PreTr' denotes pre-trained models we used; '20X' and '50X' denote microscopic images at " + }, + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 46, + 257, + 546, + 291 + ], + "type": "text", + "content": " magnifications; 'TEX' and 'HM' denote texture and heightmap; '9w5s' and '9w20s' denote 9-way-5-shot and 9-way-20-shot, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 304, + 288, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 423 + ], + "type": "text", + "content": "recognizes important polished regions in microscopic images of beechwood (see Fig. 5) as the foreground, but recognizes the same polish of different categories ( sprucewood, bone, and antler, see Fig. 5) as the background. More explorations are needed to explain this unwanted behavior. Interestingly, we found that the regions recognized as highly important for classification are similar for human experts and our best model ResNet152, under the same data configuration (see Regions 1, 4, and 6 in Fig. 2). We visualize the results for ResNet152 using Grad-CAM [54]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 209, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 209, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 209, + 443 + ], + "type": "text", + "content": "4.2. Few-Shot Image Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 544 + ], + "type": "text", + "content": "In practice, LUWA faces a scarcity of microscopic images due to limited stone tools, expertise requirements, and specialized equipment, especially when discovering new categories. Human experts can identify new classes of wear traces with a few examples. To emulate this, we investigate whether few-shot learning can be utilized and how microscopic image magnifications and sensing modalities influence the model's performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Experimental Settings. We designed two main experiments: (i) Few-shot image classification with a simple but effective pre-train + ProtoNet pipeline [27]. We evaluate the performance of powerful pre-trained models (including ResNet, ViT, DINO [10], ConvNeXt, CLIP [46], DEIT [63]) and popular meta-learners ProtoNet [57] on LUWA dataset. We simulated 600 episodes/tasks and results are demonstrated under 9-way-5/20-shot settings. (ii) GPT-4V [42] experiments: few-shot image classification and reasoning following instructions from human experts. To explore the potential mode of AI-human collaboration facing scientific domains in the advent of large multi-modal models, we collected prompts from three archaeologists and conducted case studies on whether the latest GPT-4V can" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 411 + ], + "type": "text", + "content": "follow and mimic the experts' reasoning process when analyzing the samples. Then we summarized key points that matter during the experts' analysis and used that to prompt the GPT-4V for few-shot image classification and reasoning. The experiments are illustrated in Fig. 6. Additionally, we included human experts' test results to reflect the difficulty for humans to distinguish these discovered categories with just a few examples. Results are reported under 9-way-5-shot settings in Tab. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 411, + 546, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 546, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 546, + 663 + ], + "type": "text", + "content": "Results and Analysis. Experimental results in Tab. 2 demonstrated that DINO excels at few-shot learning classification. Note that in the case of limited microscopic images, classification results of textures at " + }, + { + "bbox": [ + 304, + 411, + 546, + 663 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 304, + 411, + 546, + 663 + ], + "type": "text", + "content": " magnification setting yield notably superior results compared to others, which provides valuable guidance for few-shot learning tasks in our domain. Moreover, the number of parameters in pre-trained models has a limited impact on this few-shot learning task. We found that GPT-4V can effectively follow the experts' analysis as highlighted in Fig. 6. It learns to emphasize the same points that the experts pay attention to. However, the analysis doesn't always lead to the correct answer. GPT-4V did poorly on few-shot classification. We hypothesize that our data is very different from the web data that GPT models are trained on, and the vision module in GPT-4V still struggles to efficiently present detailed vision information to the language module, especially in a long context such as our multi-image few-shot classification scenario. This means the vision ability of SOTA multi-modal language models still needs improvement before they can be used in scientific tasks with domain-specific data like ours." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 668, + 536, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 536, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 536, + 681 + ], + "type": "text", + "content": "5. Impact and Limitations of LUWA Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Scientific Impacts. AI-expert collaboration can provide invaluable insights for scientific research. To tackle the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22569" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 74, + 111, + 133 + ], + "blocks": [ + { + "bbox": [ + 52, + 74, + 111, + 133 + ], + "lines": [ + { + "bbox": [ + 52, + 74, + 111, + 133 + ], + "spans": [ + { + "bbox": [ + 52, + 74, + 111, + 133 + ], + "type": "image", + "image_path": "ff195c3b5bb4b5a626062a3f354df79507ccea8bb39caede5f9b45e82d8a10ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 135, + 111, + 194 + ], + "blocks": [ + { + "bbox": [ + 52, + 135, + 111, + 194 + ], + "lines": [ + { + "bbox": [ + 52, + 135, + 111, + 194 + ], + "spans": [ + { + "bbox": [ + 52, + 135, + 111, + 194 + ], + "type": "image", + "image_path": "0649381b2e18effd5b55a428d2e22c6ef97d1a53246b4335d468a56911053452.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 198, + 137, + 206 + ], + "lines": [ + { + "bbox": [ + 87, + 198, + 137, + 206 + ], + "spans": [ + { + "bbox": [ + 87, + 198, + 137, + 206 + ], + "type": "text", + "content": "(a) Sprucewood" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 113, + 74, + 173, + 133 + ], + "blocks": [ + { + "bbox": [ + 113, + 74, + 173, + 133 + ], + "lines": [ + { + "bbox": [ + 113, + 74, + 173, + 133 + ], + "spans": [ + { + "bbox": [ + 113, + 74, + 173, + 133 + ], + "type": "image", + "image_path": "c816ed627a994dd3c605e9d496e962a84203ee20ad86233048c4878fbf1d21c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 113, + 135, + 173, + 194 + ], + "blocks": [ + { + "bbox": [ + 113, + 135, + 173, + 194 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 173, + 194 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 173, + 194 + ], + "type": "image", + "image_path": "728e739159a84a453005f81e7a4c107455d69a5db5d0887429f9b029f0ad7e5e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 144, + 212, + 447, + 224 + ], + "lines": [ + { + "bbox": [ + 144, + 212, + 447, + 224 + ], + "spans": [ + { + "bbox": [ + 144, + 212, + 447, + 224 + ], + "type": "text", + "content": "Figure 5. Feature visualization of LUWA dataset using frozen pre-trained DINOv2." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 74, + 235, + 133 + ], + "blocks": [ + { + "bbox": [ + 176, + 74, + 235, + 133 + ], + "lines": [ + { + "bbox": [ + 176, + 74, + 235, + 133 + ], + "spans": [ + { + "bbox": [ + 176, + 74, + 235, + 133 + ], + "type": "image", + "image_path": "48f22d766c9fe098378d65d3b8987f32331694015f1e1ff7a4b4bee4b6c4b9b2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 135, + 235, + 194 + ], + "blocks": [ + { + "bbox": [ + 176, + 135, + 235, + 194 + ], + "lines": [ + { + "bbox": [ + 176, + 135, + 235, + 194 + ], + "spans": [ + { + "bbox": [ + 176, + 135, + 235, + 194 + ], + "type": "image", + "image_path": "95d91a33759821101f9a261041d2b1a2c2a2b341b6d28c6bc1a7eccdc06462ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 198, + 258, + 206 + ], + "lines": [ + { + "bbox": [ + 210, + 198, + 258, + 206 + ], + "spans": [ + { + "bbox": [ + 210, + 198, + 258, + 206 + ], + "type": "text", + "content": "(b) Beechwood" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 74, + 295, + 133 + ], + "blocks": [ + { + "bbox": [ + 236, + 74, + 295, + 133 + ], + "lines": [ + { + "bbox": [ + 236, + 74, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 236, + 74, + 295, + 133 + ], + "type": "image", + "image_path": "f1771db2c52c69977a0de6c39ffe6e6ea58c6daa17adb349560101f4b18b4d2c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 135, + 295, + 194 + ], + "blocks": [ + { + "bbox": [ + 236, + 135, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 236, + 135, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 236, + 135, + 295, + 194 + ], + "type": "image", + "image_path": "1d478f3862446b768a0730c084a6daa59433310fae6f3829f6fe4001c506dba5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 298, + 74, + 358, + 194 + ], + "blocks": [ + { + "bbox": [ + 298, + 74, + 358, + 194 + ], + "lines": [ + { + "bbox": [ + 298, + 74, + 358, + 194 + ], + "spans": [ + { + "bbox": [ + 298, + 74, + 358, + 194 + ], + "type": "image", + "image_path": "09e8dc111431146e0b6ba110d5620d618f7b9fa039f0a72679ad27790dabdc8f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 198, + 373, + 206 + ], + "lines": [ + { + "bbox": [ + 345, + 198, + 373, + 206 + ], + "spans": [ + { + "bbox": [ + 345, + 198, + 373, + 206 + ], + "type": "text", + "content": "(c) Bone" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 359, + 74, + 418, + 194 + ], + "blocks": [ + { + "bbox": [ + 359, + 74, + 418, + 194 + ], + "lines": [ + { + "bbox": [ + 359, + 74, + 418, + 194 + ], + "spans": [ + { + "bbox": [ + 359, + 74, + 418, + 194 + ], + "type": "image", + "image_path": "7c5f05649a664517010fd672d5ff72194eba550355c494cf8eebfef5389c932e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 198, + 373, + 206 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 198, + 373, + 206 + ], + "spans": [ + { + "bbox": [ + 361, + 198, + 373, + 206 + ], + "type": "text", + "content": "tone" + } + ] + } + ], + "index": 13, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 420, + 74, + 479, + 194 + ], + "blocks": [ + { + "bbox": [ + 420, + 74, + 479, + 194 + ], + "lines": [ + { + "bbox": [ + 420, + 74, + 479, + 194 + ], + "spans": [ + { + "bbox": [ + 420, + 74, + 479, + 194 + ], + "type": "image", + "image_path": "a55f425e2e357b6248e93f9c4a43b3166a2bafc783830a5352602dc31af94fc7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 198, + 498, + 206 + ], + "lines": [ + { + "bbox": [ + 466, + 198, + 498, + 206 + ], + "spans": [ + { + "bbox": [ + 466, + 198, + 498, + 206 + ], + "type": "text", + "content": "(d) Antler" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 482, + 74, + 541, + 194 + ], + "blocks": [ + { + "bbox": [ + 482, + 74, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 482, + 74, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 482, + 74, + 541, + 194 + ], + "type": "image", + "image_path": "d0b22da14fecc4345b82fc0c51ed9841bfa72c8f060c755b9db0663df6551a24.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 243, + 149, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 243, + 149, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 243, + 149, + 251 + ], + "type": "text", + "content": "Prompt: Expert Analysis" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 78, + 260, + 107, + 289 + ], + "blocks": [ + { + "bbox": [ + 78, + 260, + 107, + 289 + ], + "lines": [ + { + "bbox": [ + 78, + 260, + 107, + 289 + ], + "spans": [ + { + "bbox": [ + 78, + 260, + 107, + 289 + ], + "type": "image", + "image_path": "733040ec0b23e5b21d4d857ad3d457cb451e82a4009d41826651e05f35fc34dd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 112, + 260, + 140, + 289 + ], + "blocks": [ + { + "bbox": [ + 112, + 260, + 140, + 289 + ], + "lines": [ + { + "bbox": [ + 112, + 260, + 140, + 289 + ], + "spans": [ + { + "bbox": [ + 112, + 260, + 140, + 289 + ], + "type": "image", + "image_path": "2844ee651e5d99c9a1b503489929e9345c2b9f5c660548eed599d1f914d33724.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 75, + 289, + 147, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 289, + 147, + 387 + ], + "spans": [ + { + "bbox": [ + 75, + 289, + 147, + 387 + ], + "type": "text", + "content": "First image is more polished I would argue it was used in softer wood (spruce). ... softer materials create larger polishes region. I see linear traces like striated marks ... does not appear that way for harder materials." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 75, + 388, + 129, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 388, + 129, + 404 + ], + "spans": [ + { + "bbox": [ + 75, + 388, + 129, + 404 + ], + "type": "text", + "content": "Second image …\nbeechwood." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 75, + 406, + 149, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 406, + 149, + 450 + ], + "spans": [ + { + "bbox": [ + 75, + 406, + 149, + 450 + ], + "type": "text", + "content": "Beechwood is a harder material … smaller region of polish … no linear traces seen in the first image." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 190, + 244, + 247, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 244, + 247, + 252 + ], + "spans": [ + { + "bbox": [ + 190, + 244, + 247, + 252 + ], + "type": "text", + "content": "GPT-4V Analysis" + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 159, + 262, + 194, + 297 + ], + "blocks": [ + { + "bbox": [ + 159, + 262, + 194, + 297 + ], + "lines": [ + { + "bbox": [ + 159, + 262, + 194, + 297 + ], + "spans": [ + { + "bbox": [ + 159, + 262, + 194, + 297 + ], + "type": "image", + "image_path": "e5d9ecf209b0b29afa3e3d5a726a23d916f1c678c19c6a391129dddca6077991.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 196, + 258, + 290, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 258, + 290, + 321 + ], + "spans": [ + { + "bbox": [ + 196, + 258, + 290, + 321 + ], + "type": "text", + "content": "It appears to have a smaller region of polish with sharp bumps and lacks the linear, striated marks that would indicate a softer material like sprucewood. Following the anthropologist's analysis:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 159, + 323, + 290, + 374 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 159, + 323, + 290, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 323, + 290, + 348 + ], + "spans": [ + { + "bbox": [ + 159, + 323, + 290, + 348 + ], + "type": "text", + "content": "- Sprucewood, being softer, would create a larger, more distributed polish and possibly show linear traces." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 159, + 349, + 287, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 349, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 159, + 349, + 287, + 374 + ], + "type": "text", + "content": "- Beechwood, being harder, would leave smaller and less extensive polish, without the linear traces." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 157, + 376, + 290, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 376, + 290, + 448 + ], + "spans": [ + { + "bbox": [ + 157, + 376, + 290, + 448 + ], + "type": "text", + "content": "The characteristics of the wear in the image suggest the use of beechwood, given the absence of widespread polished areas or striated marks. This aligns with the anthropologist's explanation that harder materials like beechwood tend to create less polish and more pronounced features on the wear surface." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 350, + 231, + 492, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 231, + 492, + 240 + ], + "spans": [ + { + "bbox": [ + 350, + 231, + 492, + 240 + ], + "type": "text", + "content": "Few shot Image Classification with GPT4-V" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 310, + 247, + 353, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 247, + 353, + 255 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 353, + 255 + ], + "type": "text", + "content": "User Prompt:" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 259, + 474, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 474, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 474, + 277 + ], + "type": "text", + "content": "An archaeologist conducting lithic use-wear analysis would examine ..." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 277, + 375, + 304 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 314, + 277, + 363, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 363, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 363, + 285 + ], + "type": "text", + "content": "Polishing..." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 287, + 364, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 287, + 364, + 294 + ], + "spans": [ + { + "bbox": [ + 314, + 287, + 364, + 294 + ], + "type": "text", + "content": "- Striations" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 314, + 295, + 375, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 295, + 375, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 295, + 375, + 304 + ], + "type": "text", + "content": "- Microflaking" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 515, + 258, + 523, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 515, + 258, + 523, + 303 + ], + "spans": [ + { + "bbox": [ + 515, + 258, + 523, + 303 + ], + "type": "text", + "content": "Task Context" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 517, + 327, + 523, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 517, + 327, + 523, + 373 + ], + "spans": [ + { + "bbox": [ + 517, + 327, + 523, + 373 + ], + "type": "text", + "content": "9-8" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 530, + 337, + 534, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 530, + 337, + 534, + 371 + ], + "spans": [ + { + "bbox": [ + 530, + 337, + 534, + 371 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "spans": [ + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "text", + "content": "This is an image taken with advanced microscopes containing traces created by material " + }, + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}\\}" + }, + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "text", + "content": ", so the use-type is " + }, + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}\\}" + }, + { + "bbox": [ + 310, + 312, + 499, + 332 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 311, + 341, + 340, + 370 + ], + "blocks": [ + { + "bbox": [ + 311, + 341, + 340, + 370 + ], + "lines": [ + { + "bbox": [ + 311, + 341, + 340, + 370 + ], + "spans": [ + { + "bbox": [ + 311, + 341, + 340, + 370 + ], + "type": "image", + "image_path": "2736326d76845e0b369973c663d564424c408786e30c032f23540829e06a4af7.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 351, + 341, + 380, + 370 + ], + "blocks": [ + { + "bbox": [ + 351, + 341, + 380, + 370 + ], + "lines": [ + { + "bbox": [ + 351, + 341, + 380, + 370 + ], + "spans": [ + { + "bbox": [ + 351, + 341, + 380, + 370 + ], + "type": "image", + "image_path": "6575579de5ed214b8d2cce11201029a332105dc27204649318a0ac5e5cc9dd87.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 390, + 341, + 418, + 370 + ], + "blocks": [ + { + "bbox": [ + 390, + 341, + 418, + 370 + ], + "lines": [ + { + "bbox": [ + 390, + 341, + 418, + 370 + ], + "spans": [ + { + "bbox": [ + 390, + 341, + 418, + 370 + ], + "type": "image", + "image_path": "6b365a1170bb549cdd3ea4bc94ee133d659f89d087dcbe35b7f90cc82a179f70.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 425, + 341, + 434, + 370 + ], + "blocks": [ + { + "bbox": [ + 425, + 341, + 434, + 370 + ], + "lines": [ + { + "bbox": [ + 425, + 341, + 434, + 370 + ], + "spans": [ + { + "bbox": [ + 425, + 341, + 434, + 370 + ], + "type": "image", + "image_path": "a1d3e1ac4d69edd7a67ba91bac032542cce7e631230629f1c35cccd2a604f909.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 455, + 341, + 470, + 370 + ], + "blocks": [ + { + "bbox": [ + 455, + 341, + 470, + 370 + ], + "lines": [ + { + "bbox": [ + 455, + 341, + 470, + 370 + ], + "spans": [ + { + "bbox": [ + 455, + 341, + 470, + 370 + ], + "type": "image", + "image_path": "d032f2fd318cf8ffb66af384fd8c40fe7201b5115425c2ea4ff24acc1e7f6f1c.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 476, + 341, + 504, + 370 + ], + "blocks": [ + { + "bbox": [ + 476, + 341, + 504, + 370 + ], + "lines": [ + { + "bbox": [ + 476, + 341, + 504, + 370 + ], + "spans": [ + { + "bbox": [ + 476, + 341, + 504, + 370 + ], + "type": "image", + "image_path": "41acc77284b3dfd7fc3f413225373474890f5680bd1c7fd0100cb820db023e3b.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 379, + 341, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 341, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 341, + 388 + ], + "type": "text", + "content": "{Antler}" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 343, + 379, + 420, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 379, + 420, + 389 + ], + "spans": [ + { + "bbox": [ + 343, + 379, + 420, + 389 + ], + "type": "text", + "content": "{Sprucewood} {Barley}" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 427, + 379, + 450, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 379, + 450, + 388 + ], + "spans": [ + { + "bbox": [ + 427, + 379, + 450, + 388 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 462, + 378, + 503, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 378, + 503, + 388 + ], + "spans": [ + { + "bbox": [ + 462, + 378, + 503, + 388 + ], + "type": "text", + "content": "e} {Ivory}" + } + ] + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 312, + 403, + 341, + 432 + ], + "blocks": [ + { + "bbox": [ + 312, + 403, + 341, + 432 + ], + "lines": [ + { + "bbox": [ + 312, + 403, + 341, + 432 + ], + "spans": [ + { + "bbox": [ + 312, + 403, + 341, + 432 + ], + "type": "image", + "image_path": "9553326c07ab650062c19ed4d34be186777fdc154dcf1659f1c84479e8147afc.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + } + ], + "index": 52 + }, + { + "bbox": [ + 343, + 403, + 509, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 403, + 509, + 430 + ], + "spans": [ + { + "bbox": [ + 343, + 403, + 509, + 430 + ], + "type": "text", + "content": "This is another image, do the traces on this image belong to which of the nine use-types as demonstrated above? Answer in exactly one word:" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 309, + 437, + 412, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 437, + 412, + 447 + ], + "spans": [ + { + "bbox": [ + 309, + 437, + 412, + 447 + ], + "type": "text", + "content": "GPT-4V Answer: Sprucewood." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 517, + 405, + 524, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 517, + 405, + 524, + 426 + ], + "spans": [ + { + "bbox": [ + 517, + 405, + 524, + 426 + ], + "type": "text", + "content": "Query" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 479, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 550 + ], + "type": "text", + "content": "long-standing problem of stone tool use, we make the first attempt to collaborate with archaeologists and utilize advanced learning-based methods for worked material inference. The LUWA dataset allows for further investigations to advance our understanding of ancient tool use and material processing techniques." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 46, + 551, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 646 + ], + "type": "text", + "content": "Limitations and Future Directions. We will enrich the LUWA dataset from the following three aspects: (i) supplement microscopic images with both worn and unworn regions using lower-magnification objectives, allowing wear trace segmentation and detection tasks; (ii) collect images including wear traces caused by different worked material; (iii) increase categories of man-made materials for comprehensive analysis on wear features and material properties." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 47, + 649, + 119, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 649, + 119, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 649, + 119, + 660 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 712 + ], + "type": "text", + "content": "We collaborate with anthropological archaeologists and present the first public and the largest Lithic Use-Wear Analysis (LUWA) dataset benefiting both vision and science domains. The LUWA dataset serves as a benchmark" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 304, + 479, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 634 + ], + "type": "text", + "content": "to evaluate the generalization capabilities of advanced models on image classification tasks beyond common objects. Addressing specific challenges of wear formation and microscopic imaging, LUWA offers vital guidance on selecting suitable magnifications and sensing modalities facing different scenarios. Our analysis reveals that SOTA models encounter distinct difficulties when facing these specific challenges. Despite DINOV2's superior performance relative to other methods, it overlooks visual features that archaeologists identify as indicative of wear. We anticipate that the LUWA dataset will stimulate further research into enhancing the adaptability of large-scale models to specialized domains within computer vision." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 304, + 635, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work is supported by NSF Grant 2152565, and by NYU IT High-Performance Computing resources, services, and staff expertise. We gratefully acknowledge Dr. Rakesh Behera for the tribometer hardware, and thank Sara Borsodi, Felix Devis Kisena, Kat Liu, Eugenia Ochoa, Vita Jackman Kuwabara, Alice Jiang, Meiyu Zhang, and Sriram Koushik for their valuable assistance in collecting the microscopic images." + } + ] + } + ], + "index": 61 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22570" + } + ] + } + ], + "index": 62 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 94, + 107, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 114, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[1] Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier probes. ArXiv, abs/1610.01644, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv preprint arXiv:2307.13721, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 269 + ], + "type": "text", + "content": "[3] Reza Akbarian Bafghi and Danna Gurari. A new dataset based on images taken by blind people for testing the robustness of image classification models trained for imagenet categories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16261-16270, 2023. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 288, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 288, + 303 + ], + "type": "text", + "content": "[4] Douglas B Bamforth. Investigating microwear polishes with blind tests: the institute results in context. Journal of Archaeological Science, 15(1):11-23, 1988. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 288, + 335 + ], + "type": "text", + "content": "[5] Tomasz Bartkowiak and Christopher A Brown. Multiscale 3d curvature analysis of processed surface textures of aluminum alloy 6061 t6. Materials, 12(2):257, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 337, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 402 + ], + "type": "text", + "content": "[6] Sebastiano Battiato, Alessandro Ortis, Francesca Trenta, Lorenzo Ascari, Mara Politi, and Consolata Siniscalco. Detection and classification of pollen grain microscope images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 980-981, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 458 + ], + "type": "text", + "content": "[7] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 459, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 288, + 524 + ], + "type": "text", + "content": "[8] Wonmin Byeon, Manuel Dominguez-Rodrigo, Georgios Arampatzis, Enrique Baquedano, José Yravedra, Miguel Angel Mate-González, and Petros Koumoutsakos. Automated identification and deep classification of cut marks on bones and its paleoanthropological implications. Journal of Computational Science, 32:36-43, 2019. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 525, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 288, + 590 + ], + "type": "text", + "content": "[9] Ivan Calandra, Lisa Schunk, Alice Rodriguez, Walter Gneisinger, Antonella Pedergnana, Eduardo Paixao, Telmo Pereira, Radu Iovita, and Joao Marreiros. Back to the edge: relative coordinate system for use-wear analysis. Archaeological and Anthropological Sciences, 11:5937-5948, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 647 + ], + "type": "text", + "content": "[10] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 691 + ], + "type": "text", + "content": "[11] Claire Lifan Chen, Ata Mahjoubfar, Li-Chia Tai, Ian K Blaby, Allen Huang, Kayvan Reza Niazi, and Bahram Jalali. Deep learning in label-free cell classification. Scientific reports, 6(1):21471, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[12] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "text", + "content": "[13] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 152 + ], + "type": "text", + "content": "[14] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 153, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 219 + ], + "type": "text", + "content": "[15] Aniket Dashpute, Vishwanath Saragadam, Emma Alexander, Florian Willomitzer, Aggelos Katsaggelos, Ashok Veeraraghavan, and Oliver Cossairt. Thermal spread functions (tsf): Physics-guided material classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1641-1650, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 264 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 331 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 331, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 331, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 331, + 545, + 388 + ], + "type": "text", + "content": "[18] Manuel S Drehwald, Sagi Eppel, Jolina Li, Han Hao, and Alan Aspuru-Guzik. One-shot recognition of any material anywhere using contrastive learning with physics-based rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 23524-23533, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[19] Adrian Anthony Evans. On the importance of blind testing in archaeological science: the example from lithic functional studies. Journal of Archaeological Science, 48:5-14, 2014. 3, 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 476 + ], + "type": "text", + "content": "[20] Adrian A Evans and Randolph E Donahue. Laser scanning confocal microscopy: a potential technique for the study of lithic microwear. Journal of Archaeological Science, 35(8): 2223-2230, 2008. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 477, + 545, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 522 + ], + "type": "text", + "content": "[21] Andrea Ferreri, Silvia Bucci, and Tatiana Tommasi. Multi-modal rgb-d scene recognition across domains. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2199-2208, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 545 + ], + "type": "text", + "content": "[22] Richard LK Fullagar. The role of silica in polish formation. Journal of Archaeological Science, 18(1):1-24, 1991. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 546, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 612 + ], + "type": "text", + "content": "[23] Eric Hayman, Barbara Caputo, Mario Fritz, and Jan-Olof Eklundh. On the significance of real-world conditions for material classification. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part IV 8, pages 253-266. Springer, 2004. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 658 + ], + "type": "text", + "content": "[24] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2015. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In 2015 IEEE International Conference on Computer Vision (ICCV), pages 1026-1034, 2015. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22571" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 182 + ], + "type": "text", + "content": "[27] Shell Xu Hu, Da Li, Jan Stühmer, Minyoung Kim, and Timothy M Hospedales. Pushing the limits of simple pipelines for few-shot learning: External data and fine-tuning make a difference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9068–9077, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 226 + ], + "type": "text", + "content": "[28] Juan José Ibanez, Talía Lazuen, and J González-Urquijo. Identifying experimental tool use through confocal microscopy. Journal of Archaeological Method and Theory, 26:1176-1215, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 281 + ], + "type": "text", + "content": "[29] Hartland W Jackson, Jana R Fischer, Vito RT Zanotelli, H Raza Ali, Robert Mechera, Savas D Soysal, Holger Moch, Simone Muenst, Zsuzsanna Varga, Walter P Weber, et al. The single-cell pathology landscape of breast cancer. Nature, 578 (7796):615-620, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "type": "text", + "content": "[30] L.H. Keeley. Experimental Determination of Stone Tool Uses: A Microwear Analysis. University of Chicago Press, 1980. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "text", + "content": "[31] Larry R Kimball, John F Kimball, and Patricia E Allen. Microwave polishes as viewed through the atomic force microscope. Lithic Technology, pages 6-28, 1995. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 392 + ], + "type": "text", + "content": "[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 394, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 458 + ], + "type": "text", + "content": "[33] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 513 + ], + "type": "text", + "content": "[34] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[36] João Marreiros, Ivan Calandra, Walter Gneisinger, Eduardo Paixão, Antonella Pedergnana, and Lisa Schunk. Rethinking use-wear analysis and experimentation as applied to the study of past hominin tool use. Journal of Paleolithic Archaeology, 3:475-502, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 647 + ], + "type": "text", + "content": "[37] João Marreiros, Telmo Pereira, and Radu Iovita. Controlled experiments in lithic technology and function, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 712 + ], + "type": "text", + "content": "[38] Shannon P McPherron, Zeresenay Alemseged, Curtis W Marean, Jonathan G Wynn, Denné Reed, Denis Geraads, René Bobe, and Hamdallah A Béarat. Evidence for stonetool-assisted consumption of animal tissues before 3.39 million years ago at dikika, ethiopia. Nature, 466(7308):857-860, 2010. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[39] Mark Newcomer, Roger Grace, and Romana Unger-Hamilton. Investigating microwear polishes with blind tests. Journal of Archaeological Science, 13(3):203-217, 1986. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 140 + ], + "type": "text", + "content": "[40] George H Odell. Stone tool research at the end of the millennium: classification, function, and behavior. Journal of Archaeological Research, 9:45-100, 2001. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[41] George Hamley Odell and Frieda Odell-Vereecken. Verifying the reliability of lithic use-wear assessments by 'blind tests': the low-power approach. Journal of field Archaeology, 7(1):87-120, 1980. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 207 + ], + "type": "text", + "content": "[42] OpenAI. Chatgpt can now see, hear, and speak, 2023. Accessed: 2023-11-16. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "text", + "content": "[43] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 308 + ], + "type": "text", + "content": "[44] Antonella Pedergnana and Andreu Olle. Monitoring and interpreting the use-wear formation processes on quartzite flakes through sequential experiments. *Quaternary International*, 427:35–65, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 365 + ], + "type": "text", + "content": "[45] Marcos Pizarro-Monzo, Jordi Rosell, Anna Rufa, Florent Rivals, and Ruth Blasco. A deep learning-based taphonomical approach to distinguish the modifying agent in the late pleistocene site of toll cave (barcelona, spain). *Historical Biology*, pages 1–10, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 430 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 475 + ], + "type": "text", + "content": "[47] Alice Rodriguez, Kaushik Yanamandra, Lukasz Witek, Zhong Wang, Rakesh K Behera, and Radu Iovita. The effect of worked material hardness on stone tool wear. Plos one, 17(10):e0276166, 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "type": "text", + "content": "[48] Z Rymuza. Tribology of polymers. Archives of civil and mechanical engineering, 7(4):177-184, 2007. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 533 + ], + "type": "text", + "content": "[49] Jorge Sanchez and Florent Perronnin. High-dimensional signature compression for large-scale image classification. In CVPR 2011, pages 1665-1672, 2011. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 588 + ], + "type": "text", + "content": "[50] Linus Scheibenreif, Michael Mommert, and Damian Borth. Masked vision transformers for hyperspectral image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2165-2175, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 633 + ], + "type": "text", + "content": "[51] Patrick Schmidt, Alice Rodriguez, Kaushik Yanamandra, Rakesh K Behera, and Radu Iovita. The mineralogy and structure of use-wear polish on chert. Scientific reports, 10 (1):21512, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 689 + ], + "type": "text", + "content": "[52] Benjamin J Schoville, Kyle S Brown, Jacob A Harris, and Jayne Wilkins. New experiments and a model-driven approach for interpreting middle stone age lithic point function using the edge damage distribution method. PloS one, 11 (10):e0164088, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22572" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 149 + ], + "type": "text", + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models. In Advances in Neural Information Processing Systems, pages 25278-25294. Curran Associates, Inc., 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 217 + ], + "type": "text", + "content": "[54] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 218, + 253, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 218, + 253, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 253, + 229 + ], + "type": "text", + "content": "[55] S.A. SEMENOV. Prehistoric Technology. 1964. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "text", + "content": "[56] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 297 + ], + "type": "text", + "content": "[57] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 342 + ], + "type": "text", + "content": "[58] W James Stemp and Michael Stemp. Ubm laser profilometry and lithic use-wear analysis: a variable length scale investigation of surface topography. Journal of Archaeological Science, 28(1):81-88, 2001. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 287, + 376 + ], + "type": "text", + "content": "[59] W James Stemp, Adam S Watson, and Adrian A Evans. Surface analysis of stone and bone tools. Surface Topography: Metrology and Properties, 4(1):013001, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 287, + 421 + ], + "type": "text", + "content": "[60] Nathan E Stevens, Douglas R Harro, and Alan Hicklin. Practical quantitative lithic use-wear analysis using multiple classifiers. Journal of Archaeological Science, 37(10):2671-2678, 2010. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 489 + ], + "type": "text", + "content": "[61] Kenichiro Tanaka, Yasuhiro Mukaigawa, Takuya Funatomi, Hiroyuki Kubo, Yasuyuki Matsushita, and Yasushi Yagi. Material classification using frequency- and depth-dependent time-of-flight distortion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 512 + ], + "type": "text", + "content": "[62] SH Teoh. Fatigue of biomaterials: a review. International journal of fatigue, 22(10):825-837, 2000. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[63] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 645 + ], + "type": "text", + "content": "[64] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, page 53, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[65] Hongda Wang, Hatice Ceylan Koydemir, Yunzhe Qiu, Bijie Bai, Yibo Zhang, Yiyin Jin, Sabiha Tok, Enis Cagatay Yilmaz, Esin Gumustekin, Yair Rivenson, et al. Early detection and classification of live bacteria using time-lapse coherent imaging and deep learning. Light: Science & Applications, 9(1):118, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 239 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[66] Ran Wang, Yuanjing Zhu, Chengxin Chen, Yu Han, and Hongbo Zhou. Tooth wear and tribological investigations in dentistry. Applied Bionics and Biomechanics, 2022, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 183 + ], + "type": "text", + "content": "[67] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 546, + 239 + ], + "type": "text", + "content": "[68] Samuel J Yang, Marc Berndl, D Michael Ando, Mariya Barch, Arunachalam Narayanaswamy, Eric Christiansen, Stephan Hoyer, Chris Roat, Jane Hung, Curtis T Rueden, et al. Assessing microscope image focus quality with deep learning. BMC bioinformatics, 19:1-9, 2018. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22573" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_content_list.json b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fec93f13fdc092301bdcc499be1e5289673e77bc --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_content_list.json @@ -0,0 +1,1811 @@ +[ + { + "type": "text", + "text": "LaMPilot: An Open Benchmark Dataset for Autonomous Driving with Language Model Programs", + "text_level": 1, + "bbox": [ + 127, + 128, + 841, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunsheng Ma $^{1*}$ , Can Cui $^{1*}$ , Xu Cao $^{2*}$ , Wenqian Ye $^{3}$ , Peiran Liu $^{1}$ , Juanwu Lu $^{1}$ , Amr Abdelraouf $^{4}$ , Rohit Gupta $^{4}$ , Kyungtae Han $^{4}$ , Aniket Bera $^{1}$ , James M. Rehg $^{2}$ , Ziran Wang $^{1}$ $^{1}$ Purdue University $^{2}$ University of Illinois Urbana-Champaign \n $^{3}$ University of Virginia $^{4}$ InfoTech Labs, Toyota Motor North America", + "bbox": [ + 109, + 202, + 857, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yunsheng, ziran}@purdue.edu", + "bbox": [ + 357, + 276, + 607, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving (AD) has made significant strides in recent years. However, existing frameworks struggle to interpret and execute spontaneous user instructions, such as \"overtake the car ahead.\" Large Language Models (LLMs) have demonstrated impressive reasoning capabilities showing potential to bridge this gap. In this paper, we present LaMPilot, a novel framework that integrates LLMs into AD systems, enabling them to follow user instructions by generating code that leverages established functional primitives. We also introduce LaMPilot-Bench, the first benchmark dataset specifically designed to quantitatively evaluate the efficacy of language model programs in AD. Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf LLMs on LaMPilot-Bench. Our results demonstrate the potential of LLMs in handling diverse driving scenarios and following user instructions in driving. To facilitate further research in this area, we release our code and data at GitHub.com/PurdueDigitalTwin/LaMPilot.", + "bbox": [ + 76, + 328, + 472, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 627, + 209, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous driving (AD) has witnessed remarkable progress in recent years, with an increasing number of commercial autonomous vehicles (AVs) being deployed on public roads [30]. State-of-the-art AD systems can be broadly classified into two categories: 1) a modular approach, where standalone models are developed for perception, prediction, and planning independently [15], and 2) an end-to-end approach that directly maps sensor data to control signals via a single neural network [16, 18, 38]. Despite significant breakthroughs, both approaches struggle to handle arbitrary user commands effectively, such as \"overtake the car in front of me.\"", + "bbox": [ + 75, + 652, + 468, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have demonstrated impressive capabilities in language comprehension and reasoning [42, 61], showing potential to improve the safety, ex", + "bbox": [ + 76, + 834, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plainability, and user-friendliness of AVs. Utilizing LLMs to solve AV-related tasks is gaining momentum [9, 13, 24, 60, 65]. However, the integration of LLMs into existing AD frameworks presents several challenges. Firstly, there is a lack of well-established paradigms for incorporating LLMs into the decision-making process of AVs. Secondly, there is a shortage of benchmarks designed to evaluate and compare the performance of LLM-based agents in the context of driving.", + "bbox": [ + 500, + 327, + 890, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel framework called LaMPilot. Inspired by Code as Policy [25], which utilizes code LLMs to write robot policy code, LaMPilot employs Language Model Programs (LMPs) as the action space instead of low-level vehicle control signals. We equip LLMs with APIs that cover various functional primitives, enabling them to connect natural language instructions to executable driving plans through code generation.", + "bbox": [ + 496, + 465, + 892, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We also introduce LaMPilot-Bench, the first benchmark for evaluating LMPs in AD. The primary objective for agents operating within LaMPilot-Bench is to accomplish assigned tasks safely and efficiently. LaMPilot-Bench incorporates an interactive simulator and evaluator, featuring programmatic scoring mechanisms to assess policy performance. The main contributions of this work are summarized as follows:", + "bbox": [ + 496, + 604, + 890, + 726 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose LaMPilot, a novel framework that integrates LLMs into autonomous driving systems, enhancing their ability to interpret and follow user commands.", + "- We introduce LaMPilot-Bench, the first benchmark dataset designed to evaluate the performance of LLM-powered agents in autonomous driving. Each scenario in LaMPilot-Bench consists of a task described in natural language, along with a simulated environment for comprehensive evaluation.", + "- Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf" + ], + "bbox": [ + 500, + 734, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15141", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LLMs on LaMPilot-Bench. Our results demonstrate the great potential of LLMs in handling diverse driving scenarios and following user instructions.", + "bbox": [ + 89, + 90, + 468, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 152, + 227, + 169 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Language-Guided Driving", + "text_level": 1, + "bbox": [ + 76, + 178, + 316, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies highlight the potential role of language in enhancing self-driving technology. Companies like Wayve are integrating natural language to improve the learning and explainability of their driving models. For example, their LINGO-1 system combines vision, language, and action modalities [29, 53].", + "bbox": [ + 76, + 200, + 467, + 291 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers have also been incorporating LLMs into AD systems for various purposes, such as generalization [4, 8, 11, 12, 19, 23, 27, 28, 35, 44, 49, 55, 56, 66], interpretability [10, 31, 36, 57, 63], and human-vehicle interaction [6, 7, 39, 40, 49, 59]. DiLu [56], for instance, instills knowledge-driven capability into AD systems by considering how humans drive. It introduces a new framework that includes an interactive environment, a driver agent, and a memory component. LMDrive [39] proposes an end-to-end, language-based AD framework that processes camera-LiDAR sensor data, comprehends natural language driving instructions, and directly generates vehicle control signals.", + "bbox": [ + 75, + 292, + 467, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most existing works either provide high-level goals, let LLMs generate the ego vehicle's future trajectories, or involve detailed vehicle control signals like throttle and steering angles. In contrast, our work uniquely uses LLMs to generate program code as driving policies, which are then executed with classical planners or controllers. Our approach takes advantage of the strengths of both LLMs and traditional AD components.", + "bbox": [ + 75, + 474, + 467, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Large Language Models for Code Generation", + "text_level": 1, + "bbox": [ + 76, + 604, + 464, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The emergence of powerful LLMs has opened up new possibilities for leveraging programming languages to perform various tasks across multiple domains [58]. Code generation [64] is one of the most prominent applications, where LLMs are used to generate code snippets based on natural language descriptions. LLMs have also been employed for information extraction tasks using code as an intermediate representation [48, 50]. The robotics domain has also benefited from the integration of LLMs and code. Code as Policy [25], ProgPrompt[41], and VoxPoser [17] demonstrate how LLMs can generate code to control robots and manipulate objects in 3D environments. In the vision domain, works such as ViperGPT [43] showcase the potential of using LLMs to generate code for visual processing and reasoning. Moreover, code provides a foundation for logical planning [26, 51], which can be leveraged by LLMs to tackle complex, long-horizon tasks. The combination of LLMs and code has revealed a new paradigm with broad", + "bbox": [ + 75, + 628, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "applicability across numerous engineering and scientific domains.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. LaMPilot Methodology", + "text_level": 1, + "bbox": [ + 500, + 133, + 720, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This section elaborates on the framework of LaMPilot, which integrates the world knowledge and reasoning capabilities of LLMs into an AD system to enable user instruction following. LaMPilot uses LLMs to generate LMPs that serve as the action space. LMPs have the inherent potential to symbolize actions that are both temporally generalizable and hierarchically structured. For instance, a complex maneuver like overtaking can be programmatically split into a series of actions, including lane changes, acceleration, and possibly another lane change, composed with conditional logic.", + "bbox": [ + 496, + 160, + 890, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LaMPilot is designed as an add-on module for existing AD systems rather than as a standalone solution. It synergizes the emerging capabilities of LLMs with the proven efficiency of classical AD algorithms. Instead of directly issuing real-time, low-level control signals, the LLM generates code snippets at lower frequencies. These snippets, formulated as temporary policies, guide the strategic navigation of the ego vehicle based on user instructions.", + "bbox": [ + 496, + 325, + 890, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Prompts", + "text_level": 1, + "bbox": [ + 500, + 455, + 602, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the LaMPilot framework takes inputs including API documentation (A), human instructions (I), and driving context (C). It leverages Chain-of-Thought (CoT) reasoning [54] to write policy code (P). This process can be formulated as:", + "bbox": [ + 496, + 479, + 890, + 554 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {R}, \\mathbf {P} ] = \\ell ([ \\mathbf {A}, \\mathbf {I}, \\mathbf {C} ]) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 566, + 890, + 584 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\ell$ represents the LLM backbone.", + "bbox": [ + 500, + 595, + 759, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An example to demonstrate the code generation process is provided in Tab. 1. The instruction is a spontaneous command from the user. In addition, we also provide environment contexts and API documentation as part of the input prompt for the LLM.", + "bbox": [ + 496, + 611, + 890, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The driving contexts include relevant information about the driving environment. We develop an interface that converts the numerical vector data in simulation into a narrative format using a structured language generator [4]. This narrative provides a natural input interface for the LLM and does not require additional fine-tuning for interpretation.", + "bbox": [ + 496, + 686, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Functional Primitives", + "text_level": 1, + "bbox": [ + 500, + 786, + 702, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Policy code involves calls to functional primitives. They are implemented with classical planning and control algorithms. Due to the autoregressive nature of LLMs, generating long completions can introduce significant latency, making them less suitable for time-critical tasks like collision avoidance. The main idea is to utilize LLMs for", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15142", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a4db3a0e9b548f93abe741be89f22a16b994c5bea1bc2893090b3645d650b94.jpg", + "image_caption": [ + "Figure 1. An overview of the LaMPilot framework. The Large Language Model (LLM) receives a prompt containing human instructions, driving context, and API documentation. It then writes language model programs that serve as driving policies. These policies are executed in the simulator to complete the specified driving task and are subsequently evaluated by the evaluator to assess the effectiveness of the generated policy code." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 883, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strategic planning while intentionally avoiding their direct involvement in low-level control tasks.", + "bbox": [ + 75, + 537, + 470, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the concept of Responsibility-Sensitive Safety [37], which ensures that the autonomous driving system will not issue a command that would lead to an accident, we include safety heuristics in our API implementations to minimize risks. Our API suite is categorized into four main types:", + "bbox": [ + 75, + 573, + 468, + 665 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Ego APIs: They provide information about the status of the ego vehicle.", + "- Perception APIs: They include functions like object and lane detection, which can be called to acquire information about the surrounding environment.", + "- Navigation APIs: They provide route planning functionalities given a target destination.", + "- Control APIs: They translate LLM-generated code into low-level control signals for the vehicle." + ], + "bbox": [ + 76, + 667, + 468, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the event of an exception during policy execution, the ego vehicle automatically switches to a predefined autopilot mode to prevent undefined behaviors. To provide LLMs with essential information about the available APIs and guide their correct usage, we include the API Documentation (A) in the prompt. These documents include", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "not only the input and output specifications but also provide descriptions of their usage and the underlying logic.", + "bbox": [ + 498, + 537, + 890, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Completions", + "text_level": 1, + "bbox": [ + 500, + 580, + 635, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The completion produced by the LLMs is anticipated to be valid functions, written with the provided APIs. These functions could range from straightforward, one-off functions to more complex generator functions. A simple example is demonstrated below, where the target speed of the ego vehicle is altered:", + "bbox": [ + 496, + 604, + 892, + 695 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def decrease_speed_by_5():\n set_target_speed(get_target_speed() - 5)", + "guess_lang": "python", + "bbox": [ + 519, + 700, + 823, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, LaMPilot supports the use of control structures such as if-else and loop statements, enabling the LLM to create dynamic feedback policies. An example of this is shown in Tab. 1. In this sample, the LLM generates a policy function overtake_using_right_lane. The function first checks if there is a right lane available and a vehicle in front of the ego vehicle. If both conditions are met, it enters a loop to wait for a safe opportunity to change into the right lane. Once in the right lane, the function enters another loop to monitor the distance to the target vehicle until it has successfully overtaken the target vehicle.", + "bbox": [ + 496, + 734, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15143", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "API Docs (A)", + "text_level": 1, + "bbox": [ + 112, + 101, + 192, + 113 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "You are a helpful assistant that writes python code to complete any autonomous driving task specified by me.", + "bbox": [ + 112, + 113, + 684, + 125 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here are the APIs you can use: ... [omitted]", + "bbox": [ + 112, + 126, + 346, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instruction (I)", + "text_level": 1, + "bbox": [ + 112, + 138, + 196, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Go around the car in front of you using the right lane.", + "bbox": [ + 112, + 150, + 395, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Driving Context (C)", + "text_level": 1, + "bbox": [ + 112, + 162, + 230, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "My current speed is $31.1\\mathrm{m / s}$ . I am driving on a highway with 2 lanes in my direction, and I am in the 2nd lane from the right. There is a car in front of me in my lane, at a distance of $44.9\\mathrm{m}$ , with a speed of $31.1\\mathrm{m / s}$ .", + "bbox": [ + 112, + 172, + 852, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chain-of-Thought Reasoning (R)", + "text_level": 1, + "bbox": [ + 112, + 208, + 303, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Check if there is a right lane available. 2) Check if there is a vehicle in front of the ego vehicle in the current lane within 100 meters. 3) If so, check if it is safe to enter the right lane. 4) If it is safe to enter the right lane, set the target lane of the ego vehicle to the right lane. 5) Monitor the distance to the target vehicle. 6) If the target vehicle is still in front of the ego vehicle, set the target speed of the ego vehicle to 1.5 times the speed of the target vehicle to overtake. 7) Repeat steps 5-6 until the target vehicle is no longer in front of the ego vehicle.", + "bbox": [ + 112, + 220, + 852, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Code as Policies (P)", + "text_level": 1, + "bbox": [ + 112, + 268, + 228, + 280 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Table 1. This example shows the chain-of-thought reasoning and code generation process of LaMPilot. Contents with blue titles are generated by LLMs." + ], + "code_body": "def overtake_using_right_lane( ): egoVehicle = get_egovehicle( ) current_lane = get_lane_of(egovehicle) right_lane = get_right_lane(egovehicle) targetvehicle $=$ detect_frontvehicle_in(current_lane) if right_lane is None: say(\"There is no right lane to change into.\") return if targetvehicle is None: say(\"There is no vehicle in front of me.\") return # Check if it is safe to enter the right lane while True: if is_safe-enter(right_lane): set_target_lane(right_lane) break yield autopilot( ) # Monitor the distance to the target vehicle while True: distance_to_target $=$ get_distance_between_vehicles(egovehicle, targetvehicle) if distance_to_target $< 0$ : # The target vehicle is still in front of the ego vehicle set_target_speed(get_speed_of(targetvehicle) $\\ast$ 1.5) yield autopilot( ) else: break", + "guess_lang": "python", + "bbox": [ + 112, + 290, + 754, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The yield statements allow the AD system to take control and execute the low-level actions required to follow the high-level policy set by the LLM.", + "bbox": [ + 75, + 628, + 468, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This example demonstrates how LaMPilot leverages the reasoning capabilities of LLMs to generate dynamic, context-aware driving policies. By combining the strategic decision-making of LLMs with the real-time execution of classical AD algorithms, LaMPilot enables more flexible instruction following for AVs.", + "bbox": [ + 75, + 679, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. LaMPilot Benchmark", + "text_level": 1, + "bbox": [ + 76, + 795, + 284, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce LaMPilot-Bench, the first benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. As shown in Fig. 1, LaMPilot-Bench consists of three key components: a simulator, a dataset, and an evaluator.", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Simulator", + "text_level": 1, + "bbox": [ + 500, + 628, + 612, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The LaMPilot-Bench simulator is built upon HighwayEnv [21], a widely used platform for AD research and tactical decision-making. HighwayEnv offers various driving models and simulates realistic multi-vehicle interactions. We extend HighwayEnv with interfaces suitable for LLM-based agents and implement custom intersections to diversify the driving scenarios.", + "bbox": [ + 496, + 652, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Dataset", + "text_level": 1, + "bbox": [ + 500, + 768, + 594, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The LaMPilot dataset consists of 4,900 semi-human-annotated traffic scenes, with a subset of 500 samples split as the test set. Each data sample includes:", + "bbox": [ + 496, + 792, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An instruction I: a high-level task description.", + "- An initial state: used to initialize the simulator.", + "- Goal state criteria: aligned with the instruction I." + ], + "bbox": [ + 500, + 840, + 836, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The dataset covers diverse driving scenarios, as shown", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15144", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fd796f27a55644ee3ffaf03a9f31df5d73ef2da9344a394cbfcbb89c1e7234f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StatisticValue
Total scenarios4,900 (100%)
Distance-related1,200 (24.5%)
Speed-related1,200 (24.5%)
Pulling over200 (4.1%)
Routing1,500 (30.2%)
Lane changing400 (8.2%)
Overtaking400 (8.2%)
Highway3,400 (69.4%)
Intersection1,500 (30.6%)
Average instruction length7.6 words
Maximum instruction length14 words
Minimum instruction length2 words
", + "bbox": [ + 119, + 88, + 426, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Main statistics of the LaMPilot-Bench dataset.", + "bbox": [ + 107, + 289, + 436, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a78d3836bf5c059ac368ac735d266976fad0daee2671418ec4a6c872ae995867.jpg", + "image_caption": [ + "Figure 2. Distribution of instructions in the LaMPilot-Bench dataset." + ], + "image_footnote": [], + "bbox": [ + 98, + 320, + 449, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in Table 2. For each driving scenario, LaMPilot includes various situations. Taking turning scenarios as an example, the diversity is reflected in several variables, such as the ego vehicle's initial position and state, the specific task (turning left/right or going straight), the number of other vehicles, and their positions and states. The driving model parameters for other vehicles are randomly initialized, and each scenario is assigned a random seed.", + "bbox": [ + 75, + 667, + 467, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The dataset also contains a variety of instructions reflecting realistic in-cabin human instructions categorized by maneuver types (e.g., routing, lane changing, overtaking) and scenario types (highway and intersection). Fig. 2 shows the distribution of the first four words in the instructions, highlighting their diversity. Main statistics are provided in Tab. 2.", + "bbox": [ + 75, + 794, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluator", + "text_level": 1, + "bbox": [ + 500, + 90, + 612, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The LaMPilot-Bench evaluator incorporates metrics to assess the safety and efficiency of the agent driving policies.", + "bbox": [ + 500, + 113, + 890, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Safety Metric Time-to-collision (TTC) is used to measure the vehicle's ability to maintain safe distances and avoid collisions. In a scenario with $n + 1$ vehicles (ego vehicle labeled as 0), the TTC $(\\tau_{i})$ with vehicle $i$ is calculated using the ego vehicle's velocity $\\mathbf{v}_0$ and position $\\mathbf{p}_0$ , and vehicle $i$ 's velocity $\\mathbf{v}_i$ and position $\\mathbf{p}_i$ $(1 \\leq i \\leq n)$ :", + "bbox": [ + 498, + 162, + 890, + 253 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {i} = - \\frac {\\left(\\mathbf {p} _ {0} - \\mathbf {p} _ {i}\\right) \\cdot \\left(\\mathbf {v} _ {0} - \\mathbf {v} _ {i}\\right)}{\\left\\| \\mathbf {v} _ {0} - \\mathbf {v} _ {i} \\right\\| ^ {2}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 263, + 890, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The minimum positive TTC value $\\tau_{\\mathrm{min}}$ among all $n$ vehicles over all time steps $t$ (up to task completion time $T$ ) represents the nearest potential collision:", + "bbox": [ + 498, + 300, + 890, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {\\min } = \\min _ {1 \\leq i \\leq n, 1 \\leq t \\leq T} \\tau_ {i} ^ {t} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 359, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TTC scores are empirically assigned based on a 2-second safety margin, with values above 2 seconds considered safe and given a score of 100:", + "bbox": [ + 498, + 388, + 890, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T T C} = \\left\\{ \\begin{array}{l l} 1 0 0 & \\text {i f} \\tau_ {\\min } > 2 \\\\ \\max \\left(0, 1 0 0 - \\frac {1}{\\tau_ {\\min }}\\right) & \\text {i f} 0 < \\tau_ {\\min } \\leq 2 \\\\ 0 & \\text {i f} \\tau_ {\\min } \\leq 0 \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 457, + 890, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Speed variance (SV) is considered as another safety metric [14]. The ego vehicle's speed standard deviation $\\sigma_0$ is calculated as:", + "bbox": [ + 498, + 522, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {0} = \\sqrt {\\frac {\\sum_ {t = 1} ^ {T} \\left(\\left\\| \\mathbf {v} _ {0} ^ {t} \\right\\| - \\mu_ {0}\\right) ^ {2}}{T}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 579, + 890, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_0$ is the mean speed:", + "bbox": [ + 519, + 625, + 709, + 640 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {0} = \\frac {\\sum_ {t = 1} ^ {T} \\| \\mathbf {v} _ {0} ^ {t} \\|}{T}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 648, + 890, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SV score is defined as:", + "bbox": [ + 519, + 686, + 700, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S V} = \\max (0, 1 0 0 \\cdot \\left(1 - \\sigma_ {0} / \\sigma_ {\\text {s a f e}}\\right)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 715, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma_{\\mathrm{safe}}$ is the maximum safe speed deviation, determined empirically.", + "bbox": [ + 500, + 739, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficiency Metric The time efficiency (TE) score evaluates the policy's ability to complete the task within a predefined time limit $T_{\\mathrm{limit}}$ :", + "bbox": [ + 500, + 787, + 890, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T E} = \\max (0, 1 0 0 \\cdot (1 - T / T _ {\\text {l i m i t}})) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 847, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A TE score closer to 100 indicates a more efficient policy.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15145", + "bbox": [ + 478, + 924, + 517, + 936 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1024b41434a6ec671064faa2aa9ab6698b19f8974612e471cb7fce6bad21d21d.jpg", + "image_caption": [ + "Figure 3. Schematic view of the human feedback baseline based on retrieval-augmented generation." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 472, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task Completion Criteria A task is considered successfully completed when the agent achieves the objectives specified in the instruction while maintaining safety (i.e., avoiding collisions) and efficiency (i.e., finishing within the allotted time). For example, a lane change task is completed when the vehicle is in the target lane with its heading aligned with the lane's direction within a specified threshold.", + "bbox": [ + 75, + 257, + 472, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall Scoring The final score aggregates all individual metrics, weighted according to their importance:", + "bbox": [ + 75, + 396, + 468, + 428 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S c o r e} = W _ {\\mathrm {T T C}} \\cdot \\mathrm {T T C} + W _ {\\mathrm {S V}} \\cdot \\mathrm {S V} + W _ {\\mathrm {T E}} \\cdot \\mathrm {T E} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 439, + 468, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 481, + 313, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments are designed to establish baselines for the proposed LaMPilot-Bench. The main objectives are to assess the performance of LLM-based agents in interpreting human instructions within driving contexts and to evaluate the capability of LLMs to generate code for motion planning using provided functional primitives.", + "bbox": [ + 75, + 507, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 607, + 253, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models We conduct benchmarking using various state-of-the-art large language models, which include both open-source and proprietary solutions1:", + "bbox": [ + 75, + 631, + 468, + 676 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Llama 2 [45] (llama-2-70b-chat)", + "PaLM 2 [1](code-bison)", + "- ChatGPT [34] (gpt-3.5-turbo)", + "GPT-4 [33] (gpt-4)", + "GPT-4-Turbo [32] (gpt-4-1106-preview)" + ], + "bbox": [ + 76, + 676, + 406, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Code Execution To execute a LMP, the Python exec function is used. It takes the LMP code as an input string, along with two dictionaries that define the execution scope: (i) apis, which includes all APIs the code may invoke, and (ii) policies, an initially empty dictionary that will eventually contain a policy variable. If the LMP is designed to return a generator, this generator is extracted from", + "bbox": [ + 75, + 771, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the policies dictionary after the execution of the exec function.", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation We set a 60-second time limit for each scenario. If a task is not completed within this time frame, the test case is considered a failure, and the simulation is terminated, resulting in a driving score of 0. For successful cases, the driving score is calculated as follows:", + "bbox": [ + 498, + 145, + 890, + 220 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {D r i v i n g S c o r e} = \\frac {\\alpha}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\operatorname {S c o r e} _ {i} - \\beta \\cdot P _ {\\text {c o l l i s i o n}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 234, + 890, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $N_{s}$ represents the number of successful test cases, $\\alpha$ is the success rate (ranging from 0 to 1), $\\beta$ is the collision rate (also between 0 and 1), and $P_{\\mathrm{collision}}$ is the penalty factor for collisions (set at 500 in our experiments). The Scorei for each individual sample is calculated based on Eq. (9), with weights $W_{\\mathrm{TTC}}$ , $W_{\\mathrm{SV}}$ , and $W_{\\mathrm{TE}}$ set to 0.5, 0.3, and 0.2, respectively. The driving score provides a comprehensive assessment of the agent's performance in LaMPilot-Bench.", + "bbox": [ + 498, + 287, + 890, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 500, + 422, + 607, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1 Heuristic Baselines", + "text_level": 1, + "bbox": [ + 500, + 446, + 686, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In autonomous driving, rule-based models are favored for their deterministic and interpretable nature. In our experiments, we employ two rule-based baseline policies: the Intelligent Driver Model (IDM)[46] and the Minimizing Overall Braking Induced by Lane Changes (MOBIL) principle[20]. IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front. MOBIL is an extension of IDM where a lane change is executed if the prospective new lane offers a more favorable driving scenario and the maneuver can be conducted safely. These baselines can be considered as operating on random chance, as the policy is independent of user instructions but follows predefined rules. Additionally, we include a human performance baseline, where a licensed human driver controls the vehicle in the simulation using arrow keys on the keyboard, following the commands and visuals displayed. This baseline provides a reference for human-level performance on LaMPilot-Bench.", + "bbox": [ + 496, + 472, + 890, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.2 Zero-Shot and Few-Shot Baselines", + "text_level": 1, + "bbox": [ + 500, + 784, + 792, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also include the zero-shot and few-shot chain-of-thought (CoT) prompting [54] as baseline methods. The zero-shot CoT baseline leverages the prompt \"How to complete the task step by step.\" In the few-shot setting, we follow the standard practice [2] by including 3 human-written code exemplars before the test instance. These in-context", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1OpenAI API accessed in October and November 2023.", + "bbox": [ + 94, + 886, + 390, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15146", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ad65bf1729246b0833c3618009730da4b853197d61eba3427f7395802b7d3e9e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLearningCollision (↓)Completion (↑)TTC Score (↑)SV Score (↑)TE Score (↑)Driving Score (↑)
Heuristic Baselines
IDMN/A0.0%20.4%92.880.971.017.3
MOBIL0.0%15.3%82.885.946.711.7
Human Driver0.0%98.0%93.486.381.384.6
LLM-based Agents
Llama 20-Shot0.0%20.4%92.881.468.917.3
PaLM 23.1%35.7%78.676.578.812.8
ChatGPT1.0%40.8%83.975.574.027.8
GPT-44.1%72.4%61.370.574.028.5
GPT-4 Turbo3.1%74.5%73.270.173.939.1
Llama 23-Shot1.0% (+1.0)63.3% (+42.9)68.3 (-24.5)71.4 (-10.0)73.8 (+4.9)39.6 (+22.3)
PaLM 23.1% (0.0)71.4% (+35.7)73.4 (-5.2)69.7 (-6.8)72.0 (-6.8)36.6 (+23.8)
ChatGPT4.1% (+3.1)83.7% (+42.9)70.9 (-13.0)73.7 (-1.8)77.7 (+3.7)41.2 (+13.4)
GPT-41.0% (-3.1)84.7% (+12.3)69.4 (+8.1)72.0 (+1.5)76.7 (+2.7)55.9 (+27.4)
GPT-4 Turbo1.0% (-2.1)80.6% (+6.1)69.1 (-4.1)71.5 (+1.4)74.9 (+1.0)52.6 (+13.5)
Llama 2Human Feedback0.9% (-0.1)32.7% (-30.6)83.9 (+15.6)74.6 (+3.2)74.6 (+0.8)21.4 (-18.2)
PaLM 20.0% (-3.1)45.5% (-25.9)81.1 (+7.7)74.4 (+4.7)77.6 (+5.6)35.6 (-1.0)
ChatGPT0.9% (-3.2)80.9% (-2.8)70.2 (-0.7)71.7 (-2.0)77.5 (-0.2)54.2 (+13.0)
GPT-40.9% (-0.1)92.7% (+8.0)67.6 (-1.8)72.2 (+0.2)76.8 (+0.1)64.0 (+8.1)
GPT-4 Turbo0.9% (-0.1)87.3% (+6.7)74.1 (+5.0)73.4 (+1.9)75.5 (+0.6)60.5 (+7.9)
", + "bbox": [ + 101, + 87, + 867, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of baselines on LaMPilot-Bench under different learning settings. Arrows $(\\uparrow/\\downarrow)$ indicate higher or lower scores are better, respectively. Numbers in parentheses show the performance change compared to the previous setting, with (green/red) indicating performance improvements and degradation, respectively.", + "bbox": [ + 75, + 375, + 893, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "examples help the model adapt to the tasks in LaMPilot-Bench. The in-context examples are created by a programmer proficient in Python. They are provided with API descriptions and are allowed to write and test their code using the non-test scenarios. The same set of three examples are used for all test cases. The API Docs are also the same across all test cases for both zero-shot and few-shot settings.", + "bbox": [ + 75, + 444, + 472, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.3 Human Feedback Baselines", + "text_level": 1, + "bbox": [ + 76, + 571, + 323, + 587 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLMs have demonstrated proficiency in generating coherent solutions across various tasks without requiring additional fine-tuning. However, when it comes to code generation, especially for complex scenarios, they can produce suboptimal results. The autoregressive nature of these models poses a significant challenge, as tokens generated early in a sequence cannot be modified in the same iteration. This constraint limits the models' ability to refine initial responses, potentially impacting the effectiveness of the generated code [47].", + "bbox": [ + 75, + 597, + 468, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address these challenges and enhance the performance of LLMs on tasks in LaMPilot-Bench, we introduce a human-in-the-loop approach, where the LLMs serve not only as planners but also as engines for incorporating human feedback. As shown in Fig. 3, human feedback is infused into the learning process, transforming the LaMPilot framework from an open-loop system into an evolving feedback loop. After an LLM generates a policy program $(\\mathbf{P})$ , it can integrate human feedback $(\\mathbf{F})$ with context-specific guidance, enabling the LLM to refine its output.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This approach does not involve fine-tuning or training a new LLM. Instead, we leverage the RAG approach [22] and introduce a code repository module, which is implemented with the Chroma [5] vector database. This repository allows for the storage and retrieval of effective code snippets to be used in similar situations. Following Voyager [47], we use LLM-generated function descriptions, which are then converted into vectors to serve as database keys. These keys are paired with the corresponding function codes as their values. If the human is satisfied with the policy, the code is committed to the repository for future reference. The human feedback learning process is conducted on the non-test set, involving interaction with 100 random samples for each method to ensure fair comparisons.", + "bbox": [ + 496, + 444, + 893, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 662, + 694, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present the experimental results on the LaMPilot-Bench, summarizing the performance of various methods, including heuristic baselines, 0-shot and 3-shot baselines, and human feedback baselines. These results are shown in Tab. 3.", + "bbox": [ + 496, + 686, + 893, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rule-Based Methods We first evaluate the performance of rule-based approaches, specifically the IDM and MOBIL algorithms. Both methods achieve a zero collision rate in LaMPilot-Bench. However, it is important to note that these methods operate independently of the provided instructions. Without considering human instructions, IDM and MOBIL achieve success rates of $20.4\\%$ and $15.3\\%$ , respectively. These results provide a reference point for as-", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15147", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sessing the effectiveness of LLM-based agents in following human instructions.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LLM-Based Agents We assess the effectiveness of off-the-shelf LLMs to reason in the context of autonomous driving and following human instructions. In the 0-shot setting, GPT models and PaLM 2, given only the API Docs, driving context, and instructions, achieve an obvious performance advantage compared to rule-based methods in terms of task completion. This improvement can be attributed to the LLMs' ability to understand and generate code based on natural language instructions, leveraging their extensive pretraining on diverse data sources. However, the 0-shot setting also results in an increase in the collision rate (1%-4%), indicating that the generated code policies do not fully capture the complexities of driving tasks.", + "bbox": [ + 75, + 140, + 467, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When provided with training examples that include exemplar code (3-shot setting), all the evaluated LLMs exhibit notable improvements in completion rates. For instance, Llama 2's completion rate increases from $20.4\\%$ in the 0-shot setting to $63.3\\%$ in the 3-shot setting. This improvement demonstrates the effectiveness of few-shot learning in adapting LLMs to specific task domains, as the exemplar code provides a clearer understanding of the desired behavior and API usage patterns.", + "bbox": [ + 75, + 338, + 467, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The integration of GPT models with human-in-the-loop learning further enhances their performance in driving tasks. Notably, GPT-4 achieves the highest driving score of 64.0. This improvement highlights the great potential of LLMs in following instructions in the driving context when combined with human feedback. The human feedback loop allows for iterative refinement of the generated policies, enabling the LLMs to learn from their mistakes and continuously incorporate human knowledge to generate more effective driving strategies.", + "bbox": [ + 75, + 474, + 467, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussion", + "text_level": 1, + "bbox": [ + 76, + 635, + 194, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Safety Safety is a critical factor for any AD system [52]. In our context, ensuring that the system does not execute an unsafe action, even if instructed to do so by the human, is essential for responsible autonomy. Our framework includes safety checks at two levels: (1) When the LLM generates code based on human instructions, it first reasons about the safety of the action in the context of the current driving environment. By leveraging the vast knowledge and reasoning capabilities of LLMs, the system can identify potentially unsafe instructions and generate alternative actions that prioritize safety. (2) Even if the LLM fails to catch an unsafe instruction, we have a second layer of safety checks built into the APIs that interface with the vehicle actuators. These APIs encapsulate classical planning and control algorithms with hard safety constraints. This level of safety check serves as a fail-safe mechanism, ensuring that the sys", + "bbox": [ + 75, + 657, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tem does not execute actions that violate safety constraints, regardless of the LLM's output.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instruction Tuning Instruction tuning has been validated as a technique for LLMs to acquire domain-specific knowledge efficiently [3]. Fine-tuning LLMs on driving-related tasks could enhance their performance and reduce collision rates by providing them with more targeted knowledge specific to driving. However, this may also jeopardize their generalization capability when there exist some spurious correlations in the finetuning datasets (e.g. between driving context and policy) [62]. Therefore, we hope that LaMPilot-Bench can provide a platform to encourage future researchers to explore generalizable solutions to enhance LLMs' abilities in driving tasks.", + "bbox": [ + 496, + 146, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal LLMs Exploring the integration of vision modality in multimodal LLMs for AD tasks is another important future direction. Our current framework relies on textual input and output, which limits its ability to perceive and understand the visual world. By incorporating visual perception and understanding capabilities, multimodal LLMs could process and reason about real-world driving scenes more effectively.", + "bbox": [ + 496, + 352, + 890, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 491, + 617, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced LaMPilot, a novel framework that integrates Large Language Models (LLMs) into autonomous driving (AD) systems, enabling human-like interaction in diverse driving contexts. We also created LaMPilot-Bench, a benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. Our experiments demonstrated that off-the-shelf LLMs can generate code policies for driving tasks based on human instructions. However, the notable collision rate indicates the need for further research to fully capture the complexities and safety requirements of real-world driving scenarios. LaMPilot provides a starting point for leveraging LLMs in solving AD-related tasks. As LLMs continue to evolve and integrate with more modalities, we envision a future where they can significantly contribute to the development of safe, efficient, and user-friendly autonomous vehicles.", + "bbox": [ + 496, + 518, + 890, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 777, + 651, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially funded by the Digital Twin Roadmap of InfoTech Labs, Toyota Motor North America. The contents of this paper only reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views of Toyota Motor North America.", + "bbox": [ + 496, + 801, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15148", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Stefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov and Yonghui Wu. PaLM 2 Technical Report. arXiv, 2023. 6", + "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In NeurIPS, 2020. 6", + "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Dataset for Map and Traffic Scene Understanding. In CVPR, 2024. 8", + "[4] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving. In ICRA, 2024. 2", + "[5] Chroma. The AI-native open-source embedding database," + ], + "bbox": [ + 84, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "2023. URL https://github.com/chroma-core/chroma.7", + "[6] Can Cui, Zichong Yang, Yupeng Zhou, Yunsheng Ma, Juanwu Lu, Lingxi Li, Yaobin Chen, Jitesh Panchal, and Ziran Wang. Large Language Models for Autonomous Driving: Real-World Experiments. arXiv, 2023. 2", + "[7] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as You Speak: Enabling Human-Like Interaction with Large Language Models in Autonomous Vehicles. In WACVW, 2024. 2", + "[8] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Receive, Reason, and React: Drive as You Say with Large Language Models in Autonomous Vehicles. IEEE Intelligent Transportation Systems Magazine, 2024. 2", + "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A Survey on Multimodal Large Language Models for Autonomous Driving. In WACVW, 2024. 1", + "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. HiLM-D: Towards High-Resolution Understanding in Multimodal Large Language Models for Autonomous Driving. arXiv, 2023. 2", + "[11] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. LimSim++: A Closed-Loop Platform for Deploying Multimodal LLMs in Autonomous Driving. arXiv, 2024. 2", + "[12] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive Like a Human: Rethinking Autonomous Driving with Large Language Models. In WACvw, 2024. 2", + "[13] Haoxiang Gao, Yaqian Li, Kaiwen Long, Ming Yang, and Yiqing Shen. A Survey for Foundation Models in Autonomous Driving. arXiv, 2024. 1", + "[14] Nicholas J. Garber and Ravi Gadiraju. Factors affecting speed variance and its influence on accidents. Transportation research record, 1989. 5", + "[15] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 2020. 1", + "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1", + "[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. In CoRL, 2023. 2", + "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 1", + "[19] Ye Jin, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, and" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15149", + "bbox": [ + 478, + 924, + 519, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jiangtao Gong. SurrealDriver: Designing Generative Driver Agent Simulation Framework in Urban Contexts based on Large Language Model. arXiv, 2023. 2", + "[20] Arne Kesting, Martin Treiber, and Dirk Helbing. General Lane-Changing Model MOBIL for Car-Following Models. Transportation Research Record, 2007. 6", + "[21] Edouard Leurent. An Environment for Autonomous Driving Decision-Making, 2018. URL https://github.com/Farama-Foundation/HighwayEnv.4", + "[22] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In NeurIPS, 2020. 7", + "[23] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving Everywhere with Large Language Model Policy Adaptation. In CVPR, 2024. 2", + "[24] Xin Li, Yeqi Bai, Pinlong Cai, Licheng Wen, Daocheng Fu, Bo Zhang, Xuemeng Yang, Xinyu Cai, Tao Ma, Jianfei Guo, Xing Gao, Min Dou, Botian Shi, Yong Liu, Liang He, and Yu Qiao. Towards Knowledge-driven Autonomous Driving. arXiv, 2023. 1", + "[25] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control. In ICRA, 2023. 1, 2", + "[26] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models. In NeurIPS, 2023. 2", + "[27] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. GPT-Driver: Learning to Drive with GPT. NeurIPS Workshop on Foundation Models for Decision Making, 2023. 2", + "[28] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A Language Agent for Autonomous Driving. arXiv, 2023. 2", + "[29] Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. LingoQA: Video Question Answering for Autonomous Driving. arXiv, 2023. 2", + "[30] McKinsey Center for Future Mobility. Autonomous driving's future: convenient and connected, 2023. URL https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/autonomous-drivings-future-convenient-and-connected.1", + "[31] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving. arXiv, 2023. 2", + "[32] OpenAI. GPT-4-Turbo with 128K context and lower prices, the new Assistants API, GPT-4 Turbo with Vision, DALL-E 3 API, and more, 2023. URL https://openai.com/" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "blog/new-models-and-developer-products-announced-at-devday.6", + "[33] OpenAI. GPT-4 Technical Report. arXiv, 2023. 6", + "[34] OpenAI. Introducing ChatGPT, 2023. URL https://openai.com/blog/chatgpt.6", + "[35] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G. Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In CVPR, 2024. 2", + "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving. arXiv, 2023. 2", + "[37] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. On a Formal Model of Safe and Scalable Self-driving Cars. arXiv, 2017. 3", + "[38] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. ReasonNet: End-to-End Driving with Temporal and Global Reasoning. In CVPR. arXiv, 2023. 1", + "[39] Hao Shao, Yuxuan Hu, Letian Wang, Steven L. Waslander, Yu Liu, and Hongsheng Li. LMDrive: Closed-Loop End-to-End Driving with Large Language Models. In CVPR, 2024. 2", + "[40] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv, 2023. 2", + "[41] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. In ICRA, 2023. 2", + "[42] Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming Zhang, Pheng Ann Heng, Jifeng Dai, Ping Luo, Jingdong Wang, Ji-Rong Wen, Xipeng Qiu, Yike Guo, Hui Xiong, Qun Liu, and Zhenguo Li. A Survey of Reasoning with Foundation Models. arXiv, 2024. 1", + "[43] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In ICCV, 2023. 2", + "[44] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models. arXiv, 2024. 2", + "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15150", + "bbox": [ + 480, + 924, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and FineTuned Chat Models. arXiv, 2023. 6", + "[46] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 2000. 6", + "[47] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv, 2023. 7", + "[48] Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Zenglin Xu, Shaoliang Nie, Sinong Wang, Madian Khabsa, Hamed Firooz, and Dongfang Liu. MUSTIE: Multimodal Structural Transformer for Web Information Extraction. In ACL, 2023. 2", + "[49] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. DriveMLM: Aligning Multi-Modal Large Language Models with Behavioral Planning States for Autonomous Driving. arXiv, 2023. 2", + "[50] Xingyao Wang, Sha Li, and Heng Ji. Code4Struct: Code Generation for Few-Shot Event Structure Prediction. In ACL, 2023. 2", + "[51] Xingyao Wang, Hao Peng, Reyhaneh Jabbarvand, and Heng Ji. LeTI: Learning to Generate from Textual Interactions. arXiv, 2023. 2", + "[52] Yixuan Wang, Ruochen Jiao, Chengtian Lang, Sinong Simon Zhan, Chao Huang, Zhaoran Wang, Zhuoran Yang, and Qi Zhu. Empowering Autonomous Driving with Large Language Models: A Safety Perspective. In ICLR Workshop on LLM Agents, 2024. 8", + "[53] Wayve Technologies Ltd. LINGO-1: Exploring Natural Language for Autonomous Driving, 2023. URL https://wayve.ai/thinking/lingo-natural-language-autonomous-driving/.2", + "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS, 2022. 2, 6", + "[55] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, Tao Ma, Yingxuan Li, Linran Xu, Dengke Shang, Zheng Zhu, Shaoyan Sun, Yeqi Bai, Xinyu Cai, Min Dou, Shuanglu Hu, Botian Shi, and Yu Qiao. On the Road with GPT-4V(ision): Early Explorations of Visual-Language Model on Autonomous Driving. arXiv, 2023. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models. In ICLR, 2024. 2", + "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhengu Li, and Hengshuang Zhao. DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model. arXiv, 2023. 2", + "[58] Ke Yang, Jiateng Liu, John Wu, Chaoqi Yang, Yi R. Fung, Sha Li, Zixuan Huang, Xu Cao, Xingyao Wang, Yiquan Wang, Heng Ji, and Chengxiang Zhai. If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents. ICLR Workshop on LLM Agents, 2024. 2", + "[59] Yi Yang, Qingwen Zhang, Ci Li, Daniel Simões Marta, Nazre Batool, and John Folkesson. Human-Centric Autonomous Systems With LLMs for User Command Reasoning. In WACVW, 2024. 2", + "[60] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. LLM4Drive: A Survey of Large Language Models for Autonomous Driving. arXiv, 2023. 1", + "[61] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In ICLR, 2023. 1", + "[62] Wenqian Ye, Guangtao Zheng, Xu Cao, Yunsheng Ma, Xia Hu, and Aidong Zhang. Spurious Correlations in Machine Learning: A Survey. arXiv, 2024. 8", + "[63] Jianhao Yuan, Shuyang Sun, Daniel Omeiza, Bo Zhao, Paul Newman, Lars Kunze, and Matthew Gadd. RAGDriver: Generalisable Driving Explanations with Retrieval-Augmented In-Context Learning in Multi-Modal Large Language Model. arXiv, 2024. 2", + "[64] Daoguang Zan, Bei Chen, Fengji Zhang, Dianjie Lu, Bingchao Wu, Bei Guan, Yongji Wang, and Jian-Guang Lou. Large Language Models Meet NL2Code: A Survey. In ACL, 2023. 2", + "[65] Xingcheng Zhou, Mingyu Liu, Bare Luka Zagar, Ekim Yurtsever, and Alois C. Knoll. Vision Language Models in Autonomous Driving and Intelligent Transportation Systems. arXiv, 2023. 1", + "[66] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied Understanding of Driving Scenarios. arXiv, 2024. 2" + ], + "bbox": [ + 501, + 92, + 890, + 726 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15151", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_model.json b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..017f3cfd5cd61f2c5b5387ada6cbf991078176b8 --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.13, + 0.842, + 0.176 + ], + "angle": 0, + "content": "LaMPilot: An Open Benchmark Dataset for Autonomous Driving with Language Model Programs" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.203, + 0.858, + 0.275 + ], + "angle": 0, + "content": "Yunsheng Ma\\(^{1*}\\), Can Cui\\(^{1*}\\), Xu Cao\\(^{2*}\\), Wenqian Ye\\(^{3}\\), Peiran Liu\\(^{1}\\), Juanwu Lu\\(^{1}\\), Amr Abdelraouf\\(^{4}\\), Rohit Gupta\\(^{4}\\), Kyungtae Han\\(^{4}\\), Aniket Bera\\(^{1}\\), James M. Rehg\\(^{2}\\), Ziran Wang\\(^{1}\\) \n\\(^{1}\\)Purdue University \\(^{2}\\)University of Illinois Urbana-Champaign \n\\(^{3}\\)University of Virginia \\(^{4}\\)InfoTech Labs, Toyota Motor North America" + }, + { + "type": "text", + "bbox": [ + 0.358, + 0.277, + 0.608, + 0.293 + ], + "angle": 0, + "content": "{yunsheng, ziran}@purdue.edu" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.329, + 0.473, + 0.616 + ], + "angle": 0, + "content": "Autonomous driving (AD) has made significant strides in recent years. However, existing frameworks struggle to interpret and execute spontaneous user instructions, such as \"overtake the car ahead.\" Large Language Models (LLMs) have demonstrated impressive reasoning capabilities showing potential to bridge this gap. In this paper, we present LaMPilot, a novel framework that integrates LLMs into AD systems, enabling them to follow user instructions by generating code that leverages established functional primitives. We also introduce LaMPilot-Bench, the first benchmark dataset specifically designed to quantitatively evaluate the efficacy of language model programs in AD. Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf LLMs on LaMPilot-Bench. Our results demonstrate the potential of LLMs in handling diverse driving scenarios and following user instructions in driving. To facilitate further research in this area, we release our code and data at GitHub.com/PurdueDigitalTwin/LaMPilot." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.628, + 0.21, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Autonomous driving (AD) has witnessed remarkable progress in recent years, with an increasing number of commercial autonomous vehicles (AVs) being deployed on public roads [30]. State-of-the-art AD systems can be broadly classified into two categories: 1) a modular approach, where standalone models are developed for perception, prediction, and planning independently [15], and 2) an end-to-end approach that directly maps sensor data to control signals via a single neural network [16, 18, 38]. Despite significant breakthroughs, both approaches struggle to handle arbitrary user commands effectively, such as \"overtake the car in front of me.\"" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.835, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have demonstrated impressive capabilities in language comprehension and reasoning [42, 61], showing potential to improve the safety, ex" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.892, + 0.464 + ], + "angle": 0, + "content": "plainability, and user-friendliness of AVs. Utilizing LLMs to solve AV-related tasks is gaining momentum [9, 13, 24, 60, 65]. However, the integration of LLMs into existing AD frameworks presents several challenges. Firstly, there is a lack of well-established paradigms for incorporating LLMs into the decision-making process of AVs. Secondly, there is a shortage of benchmarks designed to evaluate and compare the performance of LLM-based agents in the context of driving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.893, + 0.603 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel framework called LaMPilot. Inspired by Code as Policy [25], which utilizes code LLMs to write robot policy code, LaMPilot employs Language Model Programs (LMPs) as the action space instead of low-level vehicle control signals. We equip LLMs with APIs that cover various functional primitives, enabling them to connect natural language instructions to executable driving plans through code generation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.727 + ], + "angle": 0, + "content": "We also introduce LaMPilot-Bench, the first benchmark for evaluating LMPs in AD. The primary objective for agents operating within LaMPilot-Bench is to accomplish assigned tasks safely and efficiently. LaMPilot-Bench incorporates an interactive simulator and evaluator, featuring programmatic scoring mechanisms to assess policy performance. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.735, + 0.892, + 0.779 + ], + "angle": 0, + "content": "- We propose LaMPilot, a novel framework that integrates LLMs into autonomous driving systems, enhancing their ability to interpret and follow user commands." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.781, + 0.892, + 0.869 + ], + "angle": 0, + "content": "- We introduce LaMPilot-Bench, the first benchmark dataset designed to evaluate the performance of LLM-powered agents in autonomous driving. Each scenario in LaMPilot-Bench consists of a task described in natural language, along with a simulated environment for comprehensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "- Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "15141" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "LLMs on LaMPilot-Bench. Our results demonstrate the great potential of LLMs in handling diverse driving scenarios and following user instructions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.154, + 0.228, + 0.17 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.179, + 0.317, + 0.196 + ], + "angle": 0, + "content": "2.1. Language-Guided Driving" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.202, + 0.468, + 0.292 + ], + "angle": 0, + "content": "Recent studies highlight the potential role of language in enhancing self-driving technology. Companies like Wayve are integrating natural language to improve the learning and explainability of their driving models. For example, their LINGO-1 system combines vision, language, and action modalities [29, 53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.468, + 0.475 + ], + "angle": 0, + "content": "Researchers have also been incorporating LLMs into AD systems for various purposes, such as generalization [4, 8, 11, 12, 19, 23, 27, 28, 35, 44, 49, 55, 56, 66], interpretability [10, 31, 36, 57, 63], and human-vehicle interaction [6, 7, 39, 40, 49, 59]. DiLu [56], for instance, instills knowledge-driven capability into AD systems by considering how humans drive. It introduces a new framework that includes an interactive environment, a driver agent, and a memory component. LMDrive [39] proposes an end-to-end, language-based AD framework that processes camera-LiDAR sensor data, comprehends natural language driving instructions, and directly generates vehicle control signals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.595 + ], + "angle": 0, + "content": "However, most existing works either provide high-level goals, let LLMs generate the ego vehicle's future trajectories, or involve detailed vehicle control signals like throttle and steering angles. In contrast, our work uniquely uses LLMs to generate program code as driving policies, which are then executed with classical planners or controllers. Our approach takes advantage of the strengths of both LLMs and traditional AD components." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.465, + 0.623 + ], + "angle": 0, + "content": "2.2. Large Language Models for Code Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The emergence of powerful LLMs has opened up new possibilities for leveraging programming languages to perform various tasks across multiple domains [58]. Code generation [64] is one of the most prominent applications, where LLMs are used to generate code snippets based on natural language descriptions. LLMs have also been employed for information extraction tasks using code as an intermediate representation [48, 50]. The robotics domain has also benefited from the integration of LLMs and code. Code as Policy [25], ProgPrompt[41], and VoxPoser [17] demonstrate how LLMs can generate code to control robots and manipulate objects in 3D environments. In the vision domain, works such as ViperGPT [43] showcase the potential of using LLMs to generate code for visual processing and reasoning. Moreover, code provides a foundation for logical planning [26, 51], which can be leveraged by LLMs to tackle complex, long-horizon tasks. The combination of LLMs and code has revealed a new paradigm with broad" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "applicability across numerous engineering and scientific domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.135, + 0.722, + 0.153 + ], + "angle": 0, + "content": "3. LaMPilot Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.161, + 0.892, + 0.326 + ], + "angle": 0, + "content": "This section elaborates on the framework of LaMPilot, which integrates the world knowledge and reasoning capabilities of LLMs into an AD system to enable user instruction following. LaMPilot uses LLMs to generate LMPs that serve as the action space. LMPs have the inherent potential to symbolize actions that are both temporally generalizable and hierarchically structured. For instance, a complex maneuver like overtaking can be programmatically split into a series of actions, including lane changes, acceleration, and possibly another lane change, composed with conditional logic." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.448 + ], + "angle": 0, + "content": "LaMPilot is designed as an add-on module for existing AD systems rather than as a standalone solution. It synergizes the emerging capabilities of LLMs with the proven efficiency of classical AD algorithms. Instead of directly issuing real-time, low-level control signals, the LLM generates code snippets at lower frequencies. These snippets, formulated as temporary policies, guide the strategic navigation of the ego vehicle based on user instructions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.457, + 0.604, + 0.473 + ], + "angle": 0, + "content": "3.1. Prompts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.555 + ], + "angle": 0, + "content": "As shown in Fig. 1, the LaMPilot framework takes inputs including API documentation (A), human instructions (I), and driving context (C). It leverages Chain-of-Thought (CoT) reasoning [54] to write policy code (P). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.568, + 0.891, + 0.585 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {R}, \\mathbf {P} ] = \\ell ([ \\mathbf {A}, \\mathbf {I}, \\mathbf {C} ]) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.597, + 0.76, + 0.611 + ], + "angle": 0, + "content": "where \\(\\ell\\) represents the LLM backbone." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.686 + ], + "angle": 0, + "content": "An example to demonstrate the code generation process is provided in Tab. 1. The instruction is a spontaneous command from the user. In addition, we also provide environment contexts and API documentation as part of the input prompt for the LLM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.779 + ], + "angle": 0, + "content": "The driving contexts include relevant information about the driving environment. We develop an interface that converts the numerical vector data in simulation into a narrative format using a structured language generator [4]. This narrative provides a natural input interface for the LLM and does not require additional fine-tuning for interpretation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.704, + 0.802 + ], + "angle": 0, + "content": "3.2. Functional Primitives" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Policy code involves calls to functional primitives. They are implemented with classical planning and control algorithms. Due to the autoregressive nature of LLMs, generating long completions can introduce significant latency, making them less suitable for time-critical tasks like collision avoidance. The main idea is to utilize LLMs for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "15142" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.885, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.463, + 0.893, + 0.52 + ], + "angle": 0, + "content": "Figure 1. An overview of the LaMPilot framework. The Large Language Model (LLM) receives a prompt containing human instructions, driving context, and API documentation. It then writes language model programs that serve as driving policies. These policies are executed in the simulator to complete the specified driving task and are subsequently evaluated by the evaluator to assess the effectiveness of the generated policy code." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.569 + ], + "angle": 0, + "content": "strategic planning while intentionally avoiding their direct involvement in low-level control tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Inspired by the concept of Responsibility-Sensitive Safety [37], which ensures that the autonomous driving system will not issue a command that would lead to an accident, we include safety heuristics in our API implementations to minimize risks. Our API suite is categorized into four main types:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.669, + 0.47, + 0.699 + ], + "angle": 0, + "content": "- Ego APIs: They provide information about the status of the ego vehicle." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.7, + 0.47, + 0.745 + ], + "angle": 0, + "content": "- Perception APIs: They include functions like object and lane detection, which can be called to acquire information about the surrounding environment." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.746, + 0.469, + 0.775 + ], + "angle": 0, + "content": "- Navigation APIs: They provide route planning functionalities given a target destination." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.776, + 0.469, + 0.806 + ], + "angle": 0, + "content": "- Control APIs: They translate LLM-generated code into low-level control signals for the vehicle." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.669, + 0.47, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the event of an exception during policy execution, the ego vehicle automatically switches to a predefined autopilot mode to prevent undefined behaviors. To provide LLMs with essential information about the available APIs and guide their correct usage, we include the API Documentation (A) in the prompt. These documents include" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.539, + 0.892, + 0.57 + ], + "angle": 0, + "content": "not only the input and output specifications but also provide descriptions of their usage and the underlying logic." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.582, + 0.636, + 0.599 + ], + "angle": 0, + "content": "3.3. Completions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.893, + 0.696 + ], + "angle": 0, + "content": "The completion produced by the LLMs is anticipated to be valid functions, written with the provided APIs. These functions could range from straightforward, one-off functions to more complex generator functions. A simple example is demonstrated below, where the target speed of the ego vehicle is altered:" + }, + { + "type": "code", + "bbox": [ + 0.52, + 0.702, + 0.825, + 0.725 + ], + "angle": 0, + "content": "def decrease_speed_by_5():\n set_target_speed(get_target_speed() - 5)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Moreover, LaMPilot supports the use of control structures such as if-else and loop statements, enabling the LLM to create dynamic feedback policies. An example of this is shown in Tab. 1. In this sample, the LLM generates a policy function overtake_using_right_lane. The function first checks if there is a right lane available and a vehicle in front of the ego vehicle. If both conditions are met, it enters a loop to wait for a safe opportunity to change into the right lane. Once in the right lane, the function enters another loop to monitor the distance to the target vehicle until it has successfully overtaken the target vehicle." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "15143" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.102, + 0.194, + 0.114 + ], + "angle": 0, + "content": "API Docs (A)" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.114, + 0.685, + 0.126 + ], + "angle": 0, + "content": "You are a helpful assistant that writes python code to complete any autonomous driving task specified by me." + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.127, + 0.348, + 0.138 + ], + "angle": 0, + "content": "Here are the APIs you can use: ... [omitted]" + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.139, + 0.197, + 0.15 + ], + "angle": 0, + "content": "Instruction (I)" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.151, + 0.396, + 0.162 + ], + "angle": 0, + "content": "Go around the car in front of you using the right lane." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.163, + 0.231, + 0.174 + ], + "angle": 0, + "content": "Driving Context (C)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.174, + 0.854, + 0.198 + ], + "angle": 0, + "content": "My current speed is \\(31.1\\mathrm{m / s}\\). I am driving on a highway with 2 lanes in my direction, and I am in the 2nd lane from the right. There is a car in front of me in my lane, at a distance of \\(44.9\\mathrm{m}\\), with a speed of \\(31.1\\mathrm{m / s}\\)." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.209, + 0.304, + 0.222 + ], + "angle": 0, + "content": "Chain-of-Thought Reasoning (R)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.222, + 0.854, + 0.27 + ], + "angle": 0, + "content": "1) Check if there is a right lane available. 2) Check if there is a vehicle in front of the ego vehicle in the current lane within 100 meters. 3) If so, check if it is safe to enter the right lane. 4) If it is safe to enter the right lane, set the target lane of the ego vehicle to the right lane. 5) Monitor the distance to the target vehicle. 6) If the target vehicle is still in front of the ego vehicle, set the target speed of the ego vehicle to 1.5 times the speed of the target vehicle to overtake. 7) Repeat steps 5-6 until the target vehicle is no longer in front of the ego vehicle." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.27, + 0.229, + 0.281 + ], + "angle": 0, + "content": "Code as Policies (P)" + }, + { + "type": "code", + "bbox": [ + 0.114, + 0.291, + 0.756, + 0.546 + ], + "angle": 0, + "content": "def overtake_using_right_lane( ): egoVehicle = get_egovehicle( ) current_lane = get_lane_of(egovehicle) right_lane = get_right_lane(egovehicle) targetvehicle \\(=\\) detect_frontvehicle_in(current_lane) if right_lane is None: say(\"There is no right lane to change into.\") return if targetvehicle is None: say(\"There is no vehicle in front of me.\") return # Check if it is safe to enter the right lane while True: if is_safe-enter(right_lane): set_target_lane(right_lane) break yield autopilot( ) # Monitor the distance to the target vehicle while True: distance_to_target \\(=\\) get_distance_between_vehicles(egovehicle, targetvehicle) if distance_to_target \\(< 0\\) : # The target vehicle is still in front of the ego vehicle set_target_speed(get_speed_of(targetvehicle) \\(\\ast\\) 1.5) yield autopilot( ) else: break" + }, + { + "type": "code_caption", + "bbox": [ + 0.076, + 0.592, + 0.891, + 0.62 + ], + "angle": 0, + "content": "Table 1. This example shows the chain-of-thought reasoning and code generation process of LaMPilot. Contents with blue titles are generated by LLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.676 + ], + "angle": 0, + "content": "The yield statements allow the AD system to take control and execute the low-level actions required to follow the high-level policy set by the LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.77 + ], + "angle": 0, + "content": "This example demonstrates how LaMPilot leverages the reasoning capabilities of LLMs to generate dynamic, context-aware driving policies. By combining the strategic decision-making of LLMs with the real-time execution of classical AD algorithms, LaMPilot enables more flexible instruction following for AVs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.796, + 0.285, + 0.811 + ], + "angle": 0, + "content": "4. LaMPilot Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.899 + ], + "angle": 0, + "content": "We introduce LaMPilot-Bench, the first benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. As shown in Fig. 1, LaMPilot-Bench consists of three key components: a simulator, a dataset, and an evaluator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.614, + 0.644 + ], + "angle": 0, + "content": "4.1. Simulator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.759 + ], + "angle": 0, + "content": "The LaMPilot-Bench simulator is built upon HighwayEnv [21], a widely used platform for AD research and tactical decision-making. HighwayEnv offers various driving models and simulates realistic multi-vehicle interactions. We extend HighwayEnv with interfaces suitable for LLM-based agents and implement custom intersections to diversify the driving scenarios." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.595, + 0.785 + ], + "angle": 0, + "content": "4.2. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.794, + 0.891, + 0.839 + ], + "angle": 0, + "content": "The LaMPilot dataset consists of 4,900 semi-human-annotated traffic scenes, with a subset of 500 samples split as the test set. Each data sample includes:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.841, + 0.821, + 0.854 + ], + "angle": 0, + "content": "- An instruction I: a high-level task description." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.825, + 0.869 + ], + "angle": 0, + "content": "- An initial state: used to initialize the simulator." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.87, + 0.837, + 0.884 + ], + "angle": 0, + "content": "- Goal state criteria: aligned with the instruction I." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.841, + 0.837, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.891, + 0.901 + ], + "angle": 0, + "content": "The dataset covers diverse driving scenarios, as shown" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15144" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.428, + 0.281 + ], + "angle": 0, + "content": "
StatisticValue
Total scenarios4,900 (100%)
Distance-related1,200 (24.5%)
Speed-related1,200 (24.5%)
Pulling over200 (4.1%)
Routing1,500 (30.2%)
Lane changing400 (8.2%)
Overtaking400 (8.2%)
Highway3,400 (69.4%)
Intersection1,500 (30.6%)
Average instruction length7.6 words
Maximum instruction length14 words
Minimum instruction length2 words
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.29, + 0.437, + 0.304 + ], + "angle": 0, + "content": "Table 2. Main statistics of the LaMPilot-Bench dataset." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.321, + 0.45, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.603, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Figure 2. Distribution of instructions in the LaMPilot-Bench dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.468, + 0.79 + ], + "angle": 0, + "content": "in Table 2. For each driving scenario, LaMPilot includes various situations. Taking turning scenarios as an example, the diversity is reflected in several variables, such as the ego vehicle's initial position and state, the specific task (turning left/right or going straight), the number of other vehicles, and their positions and states. The driving model parameters for other vehicles are randomly initialized, and each scenario is assigned a random seed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.9 + ], + "angle": 0, + "content": "The dataset also contains a variety of instructions reflecting realistic in-cabin human instructions categorized by maneuver types (e.g., routing, lane changing, overtaking) and scenario types (highway and intersection). Fig. 2 shows the distribution of the first four words in the instructions, highlighting their diversity. Main statistics are provided in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.614, + 0.105 + ], + "angle": 0, + "content": "4.3. Evaluator" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.114, + 0.892, + 0.145 + ], + "angle": 0, + "content": "The LaMPilot-Bench evaluator incorporates metrics to assess the safety and efficiency of the agent driving policies." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.163, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Safety Metric Time-to-collision (TTC) is used to measure the vehicle's ability to maintain safe distances and avoid collisions. In a scenario with \\( n + 1 \\) vehicles (ego vehicle labeled as 0), the TTC \\( (\\tau_{i}) \\) with vehicle \\( i \\) is calculated using the ego vehicle's velocity \\( \\mathbf{v}_0 \\) and position \\( \\mathbf{p}_0 \\), and vehicle \\( i \\)'s velocity \\( \\mathbf{v}_i \\) and position \\( \\mathbf{p}_i \\) \\( (1 \\leq i \\leq n) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.264, + 0.891, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\tau_ {i} = - \\frac {\\left(\\mathbf {p} _ {0} - \\mathbf {p} _ {i}\\right) \\cdot \\left(\\mathbf {v} _ {0} - \\mathbf {v} _ {i}\\right)}{\\left\\| \\mathbf {v} _ {0} - \\mathbf {v} _ {i} \\right\\| ^ {2}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.347 + ], + "angle": 0, + "content": "The minimum positive TTC value \\(\\tau_{\\mathrm{min}}\\) among all \\(n\\) vehicles over all time steps \\(t\\) (up to task completion time \\(T\\)) represents the nearest potential collision:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.361, + 0.891, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\tau_ {\\min } = \\min _ {1 \\leq i \\leq n, 1 \\leq t \\leq T} \\tau_ {i} ^ {t} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.39, + 0.892, + 0.434 + ], + "angle": 0, + "content": "TTC scores are empirically assigned based on a 2-second safety margin, with values above 2 seconds considered safe and given a score of 100:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.458, + 0.891, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T T C} = \\left\\{ \\begin{array}{l l} 1 0 0 & \\text {i f} \\tau_ {\\min } > 2 \\\\ \\max \\left(0, 1 0 0 - \\frac {1}{\\tau_ {\\min }}\\right) & \\text {i f} 0 < \\tau_ {\\min } \\leq 2 \\\\ 0 & \\text {i f} \\tau_ {\\min } \\leq 0 \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Speed variance (SV) is considered as another safety metric [14]. The ego vehicle's speed standard deviation \\(\\sigma_0\\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.58, + 0.891, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {0} = \\sqrt {\\frac {\\sum_ {t = 1} ^ {T} \\left(\\left\\| \\mathbf {v} _ {0} ^ {t} \\right\\| - \\mu_ {0}\\right) ^ {2}}{T}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.626, + 0.71, + 0.641 + ], + "angle": 0, + "content": "where \\(\\mu_0\\) is the mean speed:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.65, + 0.891, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mu_ {0} = \\frac {\\sum_ {t = 1} ^ {T} \\| \\mathbf {v} _ {0} ^ {t} \\|}{T}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.688, + 0.702, + 0.702 + ], + "angle": 0, + "content": "The SV score is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.717, + 0.891, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S V} = \\max (0, 1 0 0 \\cdot \\left(1 - \\sigma_ {0} / \\sigma_ {\\text {s a f e}}\\right)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.74, + 0.891, + 0.77 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\mathrm{safe}}\\) is the maximum safe speed deviation, determined empirically." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.788, + 0.891, + 0.834 + ], + "angle": 0, + "content": "Efficiency Metric The time efficiency (TE) score evaluates the policy's ability to complete the task within a predefined time limit \\( T_{\\mathrm{limit}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.848, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T E} = \\max (0, 1 0 0 \\cdot (1 - T / T _ {\\text {l i m i t}})) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "A TE score closer to 100 indicates a more efficient policy." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.473, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.242 + ], + "angle": 0, + "content": "Figure 3. Schematic view of the human feedback baseline based on retrieval-augmented generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.379 + ], + "angle": 0, + "content": "Task Completion Criteria A task is considered successfully completed when the agent achieves the objectives specified in the instruction while maintaining safety (i.e., avoiding collisions) and efficiency (i.e., finishing within the allotted time). For example, a lane change task is completed when the vehicle is in the target lane with its heading aligned with the lane's direction within a specified threshold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.429 + ], + "angle": 0, + "content": "Overall Scoring The final score aggregates all individual metrics, weighted according to their importance:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.44, + 0.47, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S c o r e} = W _ {\\mathrm {T T C}} \\cdot \\mathrm {T T C} + W _ {\\mathrm {S V}} \\cdot \\mathrm {S V} + W _ {\\mathrm {T E}} \\cdot \\mathrm {T E} \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.314, + 0.5 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Our experiments are designed to establish baselines for the proposed LaMPilot-Bench. The main objectives are to assess the performance of LLM-based agents in interpreting human instructions within driving contexts and to evaluate the capability of LLMs to generate code for motion planning using provided functional primitives." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.254, + 0.625 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.469, + 0.678 + ], + "angle": 0, + "content": "Models We conduct benchmarking using various state-of-the-art large language models, which include both open-source and proprietary solutions1:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.678, + 0.354, + 0.692 + ], + "angle": 0, + "content": "- Llama 2 [45] (llama-2-70b-chat)" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.693, + 0.288, + 0.707 + ], + "angle": 0, + "content": "PaLM 2 [1](code-bison)" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.708, + 0.334, + 0.723 + ], + "angle": 0, + "content": "- ChatGPT [34] (gpt-3.5-turbo)" + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.724, + 0.235, + 0.738 + ], + "angle": 0, + "content": "GPT-4 [33] (gpt-4)" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.738, + 0.408, + 0.754 + ], + "angle": 0, + "content": "GPT-4-Turbo [32] (gpt-4-1106-preview)" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.678, + 0.408, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Code Execution To execute a LMP, the Python exec function is used. It takes the LMP code as an input string, along with two dictionaries that define the execution scope: (i) apis, which includes all APIs the code may invoke, and (ii) policies, an initially empty dictionary that will eventually contain a policy variable. If the LMP is designed to return a generator, this generator is extracted from" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "the policies dictionary after the execution of the exec function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.146, + 0.892, + 0.222 + ], + "angle": 0, + "content": "Evaluation We set a 60-second time limit for each scenario. If a task is not completed within this time frame, the test case is considered a failure, and the simulation is terminated, resulting in a driving score of 0. For successful cases, the driving score is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.236, + 0.892, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\text {D r i v i n g S c o r e} = \\frac {\\alpha}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\operatorname {S c o r e} _ {i} - \\beta \\cdot P _ {\\text {c o l l i s i o n}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.411 + ], + "angle": 0, + "content": "Here, \\(N_{s}\\) represents the number of successful test cases, \\(\\alpha\\) is the success rate (ranging from 0 to 1), \\(\\beta\\) is the collision rate (also between 0 and 1), and \\(P_{\\mathrm{collision}}\\) is the penalty factor for collisions (set at 500 in our experiments). The Scorei for each individual sample is calculated based on Eq. (9), with weights \\(W_{\\mathrm{TTC}}\\), \\(W_{\\mathrm{SV}}\\), and \\(W_{\\mathrm{TE}}\\) set to 0.5, 0.3, and 0.2, respectively. The driving score provides a comprehensive assessment of the agent's performance in LaMPilot-Bench." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.423, + 0.608, + 0.437 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.687, + 0.462 + ], + "angle": 0, + "content": "5.2.1 Heuristic Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.892, + 0.76 + ], + "angle": 0, + "content": "In autonomous driving, rule-based models are favored for their deterministic and interpretable nature. In our experiments, we employ two rule-based baseline policies: the Intelligent Driver Model (IDM)[46] and the Minimizing Overall Braking Induced by Lane Changes (MOBIL) principle[20]. IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front. MOBIL is an extension of IDM where a lane change is executed if the prospective new lane offers a more favorable driving scenario and the maneuver can be conducted safely. These baselines can be considered as operating on random chance, as the policy is independent of user instructions but follows predefined rules. Additionally, we include a human performance baseline, where a licensed human driver controls the vehicle in the simulation using arrow keys on the keyboard, following the commands and visuals displayed. This baseline provides a reference for human-level performance on LaMPilot-Bench." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.794, + 0.799 + ], + "angle": 0, + "content": "5.2.2 Zero-Shot and Few-Shot Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also include the zero-shot and few-shot chain-of-thought (CoT) prompting [54] as baseline methods. The zero-shot CoT baseline leverages the prompt \"How to complete the task step by step.\" In the few-shot setting, we follow the standard practice [2] by including 3 human-written code exemplars before the test instance. These in-context" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.392, + 0.9 + ], + "angle": 0, + "content": "1OpenAI API accessed in October and November 2023." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15146" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.088, + 0.868, + 0.365 + ], + "angle": 0, + "content": "
MethodLearningCollision (↓)Completion (↑)TTC Score (↑)SV Score (↑)TE Score (↑)Driving Score (↑)
Heuristic Baselines
IDMN/A0.0%20.4%92.880.971.017.3
MOBIL0.0%15.3%82.885.946.711.7
Human Driver0.0%98.0%93.486.381.384.6
LLM-based Agents
Llama 20-Shot0.0%20.4%92.881.468.917.3
PaLM 23.1%35.7%78.676.578.812.8
ChatGPT1.0%40.8%83.975.574.027.8
GPT-44.1%72.4%61.370.574.028.5
GPT-4 Turbo3.1%74.5%73.270.173.939.1
Llama 23-Shot1.0% (+1.0)63.3% (+42.9)68.3 (-24.5)71.4 (-10.0)73.8 (+4.9)39.6 (+22.3)
PaLM 23.1% (0.0)71.4% (+35.7)73.4 (-5.2)69.7 (-6.8)72.0 (-6.8)36.6 (+23.8)
ChatGPT4.1% (+3.1)83.7% (+42.9)70.9 (-13.0)73.7 (-1.8)77.7 (+3.7)41.2 (+13.4)
GPT-41.0% (-3.1)84.7% (+12.3)69.4 (+8.1)72.0 (+1.5)76.7 (+2.7)55.9 (+27.4)
GPT-4 Turbo1.0% (-2.1)80.6% (+6.1)69.1 (-4.1)71.5 (+1.4)74.9 (+1.0)52.6 (+13.5)
Llama 2Human Feedback0.9% (-0.1)32.7% (-30.6)83.9 (+15.6)74.6 (+3.2)74.6 (+0.8)21.4 (-18.2)
PaLM 20.0% (-3.1)45.5% (-25.9)81.1 (+7.7)74.4 (+4.7)77.6 (+5.6)35.6 (-1.0)
ChatGPT0.9% (-3.2)80.9% (-2.8)70.2 (-0.7)71.7 (-2.0)77.5 (-0.2)54.2 (+13.0)
GPT-40.9% (-0.1)92.7% (+8.0)67.6 (-1.8)72.2 (+0.2)76.8 (+0.1)64.0 (+8.1)
GPT-4 Turbo0.9% (-0.1)87.3% (+6.7)74.1 (+5.0)73.4 (+1.9)75.5 (+0.6)60.5 (+7.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.376, + 0.895, + 0.418 + ], + "angle": 0, + "content": "Table 3. Comparison of baselines on LaMPilot-Bench under different learning settings. Arrows \\((\\uparrow/\\downarrow)\\) indicate higher or lower scores are better, respectively. Numbers in parentheses show the performance change compared to the previous setting, with (green/red) indicating performance improvements and degradation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.473, + 0.552 + ], + "angle": 0, + "content": "examples help the model adapt to the tasks in LaMPilot-Bench. The in-context examples are created by a programmer proficient in Python. They are provided with API descriptions and are allowed to write and test their code using the non-test scenarios. The same set of three examples are used for all test cases. The API Docs are also the same across all test cases for both zero-shot and few-shot settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.325, + 0.588 + ], + "angle": 0, + "content": "5.2.3 Human Feedback Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.749 + ], + "angle": 0, + "content": "LLMs have demonstrated proficiency in generating coherent solutions across various tasks without requiring additional fine-tuning. However, when it comes to code generation, especially for complex scenarios, they can produce suboptimal results. The autoregressive nature of these models poses a significant challenge, as tokens generated early in a sequence cannot be modified in the same iteration. This constraint limits the models' ability to refine initial responses, potentially impacting the effectiveness of the generated code [47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To address these challenges and enhance the performance of LLMs on tasks in LaMPilot-Bench, we introduce a human-in-the-loop approach, where the LLMs serve not only as planners but also as engines for incorporating human feedback. As shown in Fig. 3, human feedback is infused into the learning process, transforming the LaMPilot framework from an open-loop system into an evolving feedback loop. After an LLM generates a policy program \\((\\mathbf{P})\\), it can integrate human feedback \\((\\mathbf{F})\\) with context-specific guidance, enabling the LLM to refine its output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.657 + ], + "angle": 0, + "content": "This approach does not involve fine-tuning or training a new LLM. Instead, we leverage the RAG approach [22] and introduce a code repository module, which is implemented with the Chroma [5] vector database. This repository allows for the storage and retrieval of effective code snippets to be used in similar situations. Following Voyager [47], we use LLM-generated function descriptions, which are then converted into vectors to serve as database keys. These keys are paired with the corresponding function codes as their values. If the human is satisfied with the policy, the code is committed to the repository for future reference. The human feedback learning process is conducted on the non-test set, involving interaction with 100 random samples for each method to ensure fair comparisons." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.664, + 0.695, + 0.68 + ], + "angle": 0, + "content": "5.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.894, + 0.763 + ], + "angle": 0, + "content": "In this section, we present the experimental results on the LaMPilot-Bench, summarizing the performance of various methods, including heuristic baselines, 0-shot and 3-shot baselines, and human feedback baselines. These results are shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Rule-Based Methods We first evaluate the performance of rule-based approaches, specifically the IDM and MOBIL algorithms. Both methods achieve a zero collision rate in LaMPilot-Bench. However, it is important to note that these methods operate independently of the provided instructions. Without considering human instructions, IDM and MOBIL achieve success rates of \\(20.4\\%\\) and \\(15.3\\%\\), respectively. These results provide a reference point for as-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.938 + ], + "angle": 0, + "content": "15147" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "sessing the effectiveness of LLM-based agents in following human instructions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.141, + 0.468, + 0.337 + ], + "angle": 0, + "content": "LLM-Based Agents We assess the effectiveness of off-the-shelf LLMs to reason in the context of autonomous driving and following human instructions. In the 0-shot setting, GPT models and PaLM 2, given only the API Docs, driving context, and instructions, achieve an obvious performance advantage compared to rule-based methods in terms of task completion. This improvement can be attributed to the LLMs' ability to understand and generate code based on natural language instructions, leveraging their extensive pretraining on diverse data sources. However, the 0-shot setting also results in an increase in the collision rate (1%-4%), indicating that the generated code policies do not fully capture the complexities of driving tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.468, + 0.474 + ], + "angle": 0, + "content": "When provided with training examples that include exemplar code (3-shot setting), all the evaluated LLMs exhibit notable improvements in completion rates. For instance, Llama 2's completion rate increases from \\(20.4\\%\\) in the 0-shot setting to \\(63.3\\%\\) in the 3-shot setting. This improvement demonstrates the effectiveness of few-shot learning in adapting LLMs to specific task domains, as the exemplar code provides a clearer understanding of the desired behavior and API usage patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.468, + 0.626 + ], + "angle": 0, + "content": "The integration of GPT models with human-in-the-loop learning further enhances their performance in driving tasks. Notably, GPT-4 achieves the highest driving score of 64.0. This improvement highlights the great potential of LLMs in following instructions in the driving context when combined with human feedback. The human feedback loop allows for iterative refinement of the generated policies, enabling the LLMs to learn from their mistakes and continuously incorporate human knowledge to generate more effective driving strategies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.195, + 0.65 + ], + "angle": 0, + "content": "5.4. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Safety Safety is a critical factor for any AD system [52]. In our context, ensuring that the system does not execute an unsafe action, even if instructed to do so by the human, is essential for responsible autonomy. Our framework includes safety checks at two levels: (1) When the LLM generates code based on human instructions, it first reasons about the safety of the action in the context of the current driving environment. By leveraging the vast knowledge and reasoning capabilities of LLMs, the system can identify potentially unsafe instructions and generate alternative actions that prioritize safety. (2) Even if the LLM fails to catch an unsafe instruction, we have a second layer of safety checks built into the APIs that interface with the vehicle actuators. These APIs encapsulate classical planning and control algorithms with hard safety constraints. This level of safety check serves as a fail-safe mechanism, ensuring that the sys" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "tem does not execute actions that violate safety constraints, regardless of the LLM's output." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.147, + 0.892, + 0.328 + ], + "angle": 0, + "content": "Instruction Tuning Instruction tuning has been validated as a technique for LLMs to acquire domain-specific knowledge efficiently [3]. Fine-tuning LLMs on driving-related tasks could enhance their performance and reduce collision rates by providing them with more targeted knowledge specific to driving. However, this may also jeopardize their generalization capability when there exist some spurious correlations in the finetuning datasets (e.g. between driving context and policy) [62]. Therefore, we hope that LaMPilot-Bench can provide a platform to encourage future researchers to explore generalizable solutions to enhance LLMs' abilities in driving tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Multimodal LLMs Exploring the integration of vision modality in multimodal LLMs for AD tasks is another important future direction. Our current framework relies on textual input and output, which limits its ability to perceive and understand the visual world. By incorporating visual perception and understanding capabilities, multimodal LLMs could process and reason about real-world driving scenes more effectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.492, + 0.619, + 0.508 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.76 + ], + "angle": 0, + "content": "In this paper, we introduced LaMPilot, a novel framework that integrates Large Language Models (LLMs) into autonomous driving (AD) systems, enabling human-like interaction in diverse driving contexts. We also created LaMPilot-Bench, a benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. Our experiments demonstrated that off-the-shelf LLMs can generate code policies for driving tasks based on human instructions. However, the notable collision rate indicates the need for further research to fully capture the complexities and safety requirements of real-world driving scenarios. LaMPilot provides a starting point for leveraging LLMs in solving AD-related tasks. As LLMs continue to evolve and integrate with more modalities, we envision a future where they can significantly contribute to the development of safe, efficient, and user-friendly autonomous vehicles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.652, + 0.794 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is partially funded by the Digital Twin Roadmap of InfoTech Labs, Toyota Motor North America. The contents of this paper only reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views of Toyota Motor North America." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "15148" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.115, + 0.473, + 0.6 + ], + "angle": 0, + "content": "[1] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Stefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov and Yonghui Wu. PaLM 2 Technical Report. arXiv, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.086, + 0.602, + 0.473, + 0.741 + ], + "angle": 0, + "content": "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.743, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Dataset for Map and Traffic Scene Understanding. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.814, + 0.472, + 0.884 + ], + "angle": 0, + "content": "[4] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving. In ICRA, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.085, + 0.887, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[5] Chroma. The AI-native open-source embedding database," + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "2023. URL https://github.com/chroma-core/chroma.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.122, + 0.894, + 0.178 + ], + "angle": 0, + "content": "[6] Can Cui, Zichong Yang, Yupeng Zhou, Yunsheng Ma, Juanwu Lu, Lingxi Li, Yaobin Chen, Jitesh Panchal, and Ziran Wang. Large Language Models for Autonomous Driving: Real-World Experiments. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.179, + 0.895, + 0.233 + ], + "angle": 0, + "content": "[7] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as You Speak: Enabling Human-Like Interaction with Large Language Models in Autonomous Vehicles. In WACVW, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.235, + 0.895, + 0.291 + ], + "angle": 0, + "content": "[8] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Receive, Reason, and React: Drive as You Say with Large Language Models in Autonomous Vehicles. IEEE Intelligent Transportation Systems Magazine, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.293, + 0.895, + 0.388 + ], + "angle": 0, + "content": "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A Survey on Multimodal Large Language Models for Autonomous Driving. In WACVW, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. HiLM-D: Towards High-Resolution Understanding in Multimodal Large Language Models for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.447, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[11] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. LimSim++: A Closed-Loop Platform for Deploying Multimodal LLMs in Autonomous Driving. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[12] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive Like a Human: Rethinking Autonomous Driving with Large Language Models. In WACvw, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.561, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[13] Haoxiang Gao, Yaqian Li, Kaiwen Long, Ming Yang, and Yiqing Shen. A Survey for Foundation Models in Autonomous Driving. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[14] Nicholas J. Garber and Ravi Gadiraju. Factors affecting speed variance and its influence on accidents. Transportation research record, 1989. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.646, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[15] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. In CoRL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.502, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[19] Ye Jin, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, and" + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.521, + 0.937 + ], + "angle": 0, + "content": "15149" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Jiangtao Gong. SurrealDriver: Designing Generative Driver Agent Simulation Framework in Urban Contexts based on Large Language Model. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[20] Arne Kesting, Martin Treiber, and Dirk Helbing. General Lane-Changing Model MOBIL for Car-Following Models. Transportation Research Record, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[21] Edouard Leurent. An Environment for Autonomous Driving Decision-Making, 2018. URL https://github.com/Farama-Foundation/HighwayEnv.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[22] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In NeurIPS, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[23] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving Everywhere with Large Language Model Policy Adaptation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[24] Xin Li, Yeqi Bai, Pinlong Cai, Licheng Wen, Daocheng Fu, Bo Zhang, Xuemeng Yang, Xinyu Cai, Tao Ma, Jianfei Guo, Xing Gao, Min Dou, Botian Shi, Yong Liu, Liang He, and Yu Qiao. Towards Knowledge-driven Autonomous Driving. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[25] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control. In ICRA, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[26] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[27] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. GPT-Driver: Learning to Drive with GPT. NeurIPS Workshop on Foundation Models for Decision Making, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[28] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A Language Agent for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[29] Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. LingoQA: Video Question Answering for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[30] McKinsey Center for Future Mobility. Autonomous driving's future: convenient and connected, 2023. URL https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/autonomous-drivings-future-convenient-and-connected.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[31] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[32] OpenAI. GPT-4-Turbo with 128K context and lower prices, the new Assistants API, GPT-4 Turbo with Vision, DALL-E 3 API, and more, 2023. URL https://openai.com/" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.121 + ], + "angle": 0, + "content": "blog/new-models-and-developer-products-announced-at-devday.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.833, + 0.136 + ], + "angle": 0, + "content": "[33] OpenAI. GPT-4 Technical Report. arXiv, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.138, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[34] OpenAI. Introducing ChatGPT, 2023. URL https://openai.com/blog/chatgpt.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[35] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G. Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.294, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[37] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. On a Formal Model of Safe and Scalable Self-driving Cars. arXiv, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.337, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[38] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. ReasonNet: End-to-End Driving with Temporal and Global Reasoning. In CVPR. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.394, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[39] Hao Shao, Yuxuan Hu, Letian Wang, Steven L. Waslander, Yu Liu, and Hongsheng Li. LMDrive: Closed-Loop End-to-End Driving with Large Language Models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.451, + 0.892, + 0.506 + ], + "angle": 0, + "content": "[40] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.508, + 0.892, + 0.563 + ], + "angle": 0, + "content": "[41] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. In ICRA, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.565, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[42] Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming Zhang, Pheng Ann Heng, Jifeng Dai, Ping Luo, Jingdong Wang, Ji-Rong Wen, Xipeng Qiu, Yike Guo, Hui Xiong, Qun Liu, and Zhenguo Li. A Survey of Reasoning with Foundation Models. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[43] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.734, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[44] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.52, + 0.937 + ], + "angle": 0, + "content": "15150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.272 + ], + "angle": 0, + "content": "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and FineTuned Chat Models. arXiv, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.276, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[46] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.47, + 0.374 + ], + "angle": 0, + "content": "[47] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[48] Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Zenglin Xu, Shaoliang Nie, Sinong Wang, Madian Khabsa, Hamed Firooz, and Dongfang Liu. MUSTIE: Multimodal Structural Transformer for Web Information Extraction. In ACL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[49] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. DriveMLM: Aligning Multi-Modal Large Language Models with Behavioral Planning States for Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[50] Xingyao Wang, Sha Li, and Heng Ji. Code4Struct: Code Generation for Few-Shot Event Structure Prediction. In ACL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[51] Xingyao Wang, Hao Peng, Reyhaneh Jabbarvand, and Heng Ji. LeTI: Learning to Generate from Textual Interactions. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[52] Yixuan Wang, Ruochen Jiao, Chengtian Lang, Sinong Simon Zhan, Chao Huang, Zhaoran Wang, Zhuoran Yang, and Qi Zhu. Empowering Autonomous Driving with Large Language Models: A Safety Perspective. In ICLR Workshop on LLM Agents, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[53] Wayve Technologies Ltd. LINGO-1: Exploring Natural Language for Autonomous Driving, 2023. URL https://wayve.ai/thinking/lingo-natural-language-autonomous-driving/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[55] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, Tao Ma, Yingxuan Li, Linran Xu, Dengke Shang, Zheng Zhu, Shaoyan Sun, Yeqi Bai, Xinyu Cai, Min Dou, Shuanglu Hu, Botian Shi, and Yu Qiao. On the Road with GPT-4V(ision): Early Explorations of Visual-Language Model on Autonomous Driving. arXiv, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[56] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhengu Li, and Hengshuang Zhao. DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model. arXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[58] Ke Yang, Jiateng Liu, John Wu, Chaoqi Yang, Yi R. Fung, Sha Li, Zixuan Huang, Xu Cao, Xingyao Wang, Yiquan Wang, Heng Ji, and Chengxiang Zhai. If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents. ICLR Workshop on LLM Agents, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[59] Yi Yang, Qingwen Zhang, Ci Li, Daniel Simões Marta, Nazre Batool, and John Folkesson. Human-Centric Autonomous Systems With LLMs for User Command Reasoning. In WACVW, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[60] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. LLM4Drive: A Survey of Large Language Models for Autonomous Driving. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[61] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[62] Wenqian Ye, Guangtao Zheng, Xu Cao, Yunsheng Ma, Xia Hu, and Aidong Zhang. Spurious Correlations in Machine Learning: A Survey. arXiv, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[63] Jianhao Yuan, Shuyang Sun, Daniel Omeiza, Bo Zhao, Paul Newman, Lars Kunze, and Matthew Gadd. RAGDriver: Generalisable Driving Explanations with Retrieval-Augmented In-Context Learning in Multi-Modal Large Language Model. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.56, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[64] Daoguang Zan, Bei Chen, Fengji Zhang, Dianjie Lu, Bingchao Wu, Bei Guan, Yongji Wang, and Jian-Guang Lou. Large Language Models Meet NL2Code: A Survey. In ACL, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[65] Xingcheng Zhou, Mingyu Liu, Bare Luka Zagar, Ekim Yurtsever, and Alois C. Knoll. Vision Language Models in Autonomous Driving and Intelligent Transportation Systems. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[66] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied Understanding of Driving Scenarios. arXiv, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.518, + 0.937 + ], + "angle": 0, + "content": "15151" + } + ] +] \ No newline at end of file diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_origin.pdf b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb4d43cf7ed5fe74cfb1a8155a0332e245c8d88b --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/c2cae23c-06f6-48c3-8fe3-d3a2dc677cb0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:138403ae864a566ca9a96f9da7be208fb3077ee2b8170c2bc73f303a60a5fdd0 +size 1835056 diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/full.md b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae8c73d2cc37ee7d2445125439d7f5670b5ec0d9 --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/full.md @@ -0,0 +1,372 @@ +# LaMPilot: An Open Benchmark Dataset for Autonomous Driving with Language Model Programs + +Yunsheng Ma $^{1*}$ , Can Cui $^{1*}$ , Xu Cao $^{2*}$ , Wenqian Ye $^{3}$ , Peiran Liu $^{1}$ , Juanwu Lu $^{1}$ , Amr Abdelraouf $^{4}$ , Rohit Gupta $^{4}$ , Kyungtae Han $^{4}$ , Aniket Bera $^{1}$ , James M. Rehg $^{2}$ , Ziran Wang $^{1}$ $^{1}$ Purdue University $^{2}$ University of Illinois Urbana-Champaign + $^{3}$ University of Virginia $^{4}$ InfoTech Labs, Toyota Motor North America + +{yunsheng, ziran}@purdue.edu + +Autonomous driving (AD) has made significant strides in recent years. However, existing frameworks struggle to interpret and execute spontaneous user instructions, such as "overtake the car ahead." Large Language Models (LLMs) have demonstrated impressive reasoning capabilities showing potential to bridge this gap. In this paper, we present LaMPilot, a novel framework that integrates LLMs into AD systems, enabling them to follow user instructions by generating code that leverages established functional primitives. We also introduce LaMPilot-Bench, the first benchmark dataset specifically designed to quantitatively evaluate the efficacy of language model programs in AD. Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf LLMs on LaMPilot-Bench. Our results demonstrate the potential of LLMs in handling diverse driving scenarios and following user instructions in driving. To facilitate further research in this area, we release our code and data at GitHub.com/PurdueDigitalTwin/LaMPilot. + +# 1. Introduction + +Autonomous driving (AD) has witnessed remarkable progress in recent years, with an increasing number of commercial autonomous vehicles (AVs) being deployed on public roads [30]. State-of-the-art AD systems can be broadly classified into two categories: 1) a modular approach, where standalone models are developed for perception, prediction, and planning independently [15], and 2) an end-to-end approach that directly maps sensor data to control signals via a single neural network [16, 18, 38]. Despite significant breakthroughs, both approaches struggle to handle arbitrary user commands effectively, such as "overtake the car in front of me." + +Large Language Models (LLMs) have demonstrated impressive capabilities in language comprehension and reasoning [42, 61], showing potential to improve the safety, ex + +plainability, and user-friendliness of AVs. Utilizing LLMs to solve AV-related tasks is gaining momentum [9, 13, 24, 60, 65]. However, the integration of LLMs into existing AD frameworks presents several challenges. Firstly, there is a lack of well-established paradigms for incorporating LLMs into the decision-making process of AVs. Secondly, there is a shortage of benchmarks designed to evaluate and compare the performance of LLM-based agents in the context of driving. + +To address these challenges, we propose a novel framework called LaMPilot. Inspired by Code as Policy [25], which utilizes code LLMs to write robot policy code, LaMPilot employs Language Model Programs (LMPs) as the action space instead of low-level vehicle control signals. We equip LLMs with APIs that cover various functional primitives, enabling them to connect natural language instructions to executable driving plans through code generation. + +We also introduce LaMPilot-Bench, the first benchmark for evaluating LMPs in AD. The primary objective for agents operating within LaMPilot-Bench is to accomplish assigned tasks safely and efficiently. LaMPilot-Bench incorporates an interactive simulator and evaluator, featuring programmatic scoring mechanisms to assess policy performance. The main contributions of this work are summarized as follows: + +- We propose LaMPilot, a novel framework that integrates LLMs into autonomous driving systems, enhancing their ability to interpret and follow user commands. +- We introduce LaMPilot-Bench, the first benchmark dataset designed to evaluate the performance of LLM-powered agents in autonomous driving. Each scenario in LaMPilot-Bench consists of a task described in natural language, along with a simulated environment for comprehensive evaluation. +- Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf + +LLMs on LaMPilot-Bench. Our results demonstrate the great potential of LLMs in handling diverse driving scenarios and following user instructions. + +# 2. Related Works + +# 2.1. Language-Guided Driving + +Recent studies highlight the potential role of language in enhancing self-driving technology. Companies like Wayve are integrating natural language to improve the learning and explainability of their driving models. For example, their LINGO-1 system combines vision, language, and action modalities [29, 53]. + +Researchers have also been incorporating LLMs into AD systems for various purposes, such as generalization [4, 8, 11, 12, 19, 23, 27, 28, 35, 44, 49, 55, 56, 66], interpretability [10, 31, 36, 57, 63], and human-vehicle interaction [6, 7, 39, 40, 49, 59]. DiLu [56], for instance, instills knowledge-driven capability into AD systems by considering how humans drive. It introduces a new framework that includes an interactive environment, a driver agent, and a memory component. LMDrive [39] proposes an end-to-end, language-based AD framework that processes camera-LiDAR sensor data, comprehends natural language driving instructions, and directly generates vehicle control signals. + +However, most existing works either provide high-level goals, let LLMs generate the ego vehicle's future trajectories, or involve detailed vehicle control signals like throttle and steering angles. In contrast, our work uniquely uses LLMs to generate program code as driving policies, which are then executed with classical planners or controllers. Our approach takes advantage of the strengths of both LLMs and traditional AD components. + +# 2.2. Large Language Models for Code Generation + +The emergence of powerful LLMs has opened up new possibilities for leveraging programming languages to perform various tasks across multiple domains [58]. Code generation [64] is one of the most prominent applications, where LLMs are used to generate code snippets based on natural language descriptions. LLMs have also been employed for information extraction tasks using code as an intermediate representation [48, 50]. The robotics domain has also benefited from the integration of LLMs and code. Code as Policy [25], ProgPrompt[41], and VoxPoser [17] demonstrate how LLMs can generate code to control robots and manipulate objects in 3D environments. In the vision domain, works such as ViperGPT [43] showcase the potential of using LLMs to generate code for visual processing and reasoning. Moreover, code provides a foundation for logical planning [26, 51], which can be leveraged by LLMs to tackle complex, long-horizon tasks. The combination of LLMs and code has revealed a new paradigm with broad + +applicability across numerous engineering and scientific domains. + +# 3. LaMPilot Methodology + +This section elaborates on the framework of LaMPilot, which integrates the world knowledge and reasoning capabilities of LLMs into an AD system to enable user instruction following. LaMPilot uses LLMs to generate LMPs that serve as the action space. LMPs have the inherent potential to symbolize actions that are both temporally generalizable and hierarchically structured. For instance, a complex maneuver like overtaking can be programmatically split into a series of actions, including lane changes, acceleration, and possibly another lane change, composed with conditional logic. + +LaMPilot is designed as an add-on module for existing AD systems rather than as a standalone solution. It synergizes the emerging capabilities of LLMs with the proven efficiency of classical AD algorithms. Instead of directly issuing real-time, low-level control signals, the LLM generates code snippets at lower frequencies. These snippets, formulated as temporary policies, guide the strategic navigation of the ego vehicle based on user instructions. + +# 3.1. Prompts + +As shown in Fig. 1, the LaMPilot framework takes inputs including API documentation (A), human instructions (I), and driving context (C). It leverages Chain-of-Thought (CoT) reasoning [54] to write policy code (P). This process can be formulated as: + +$$ +[ \mathbf {R}, \mathbf {P} ] = \ell ([ \mathbf {A}, \mathbf {I}, \mathbf {C} ]) \tag {1} +$$ + +where $\ell$ represents the LLM backbone. + +An example to demonstrate the code generation process is provided in Tab. 1. The instruction is a spontaneous command from the user. In addition, we also provide environment contexts and API documentation as part of the input prompt for the LLM. + +The driving contexts include relevant information about the driving environment. We develop an interface that converts the numerical vector data in simulation into a narrative format using a structured language generator [4]. This narrative provides a natural input interface for the LLM and does not require additional fine-tuning for interpretation. + +# 3.2. Functional Primitives + +Policy code involves calls to functional primitives. They are implemented with classical planning and control algorithms. Due to the autoregressive nature of LLMs, generating long completions can introduce significant latency, making them less suitable for time-critical tasks like collision avoidance. The main idea is to utilize LLMs for + +![](images/0a4db3a0e9b548f93abe741be89f22a16b994c5bea1bc2893090b3645d650b94.jpg) +Figure 1. An overview of the LaMPilot framework. The Large Language Model (LLM) receives a prompt containing human instructions, driving context, and API documentation. It then writes language model programs that serve as driving policies. These policies are executed in the simulator to complete the specified driving task and are subsequently evaluated by the evaluator to assess the effectiveness of the generated policy code. + +strategic planning while intentionally avoiding their direct involvement in low-level control tasks. + +Inspired by the concept of Responsibility-Sensitive Safety [37], which ensures that the autonomous driving system will not issue a command that would lead to an accident, we include safety heuristics in our API implementations to minimize risks. Our API suite is categorized into four main types: + +- Ego APIs: They provide information about the status of the ego vehicle. +- Perception APIs: They include functions like object and lane detection, which can be called to acquire information about the surrounding environment. +- Navigation APIs: They provide route planning functionalities given a target destination. +- Control APIs: They translate LLM-generated code into low-level control signals for the vehicle. + +In the event of an exception during policy execution, the ego vehicle automatically switches to a predefined autopilot mode to prevent undefined behaviors. To provide LLMs with essential information about the available APIs and guide their correct usage, we include the API Documentation (A) in the prompt. These documents include + +not only the input and output specifications but also provide descriptions of their usage and the underlying logic. + +# 3.3. Completions + +The completion produced by the LLMs is anticipated to be valid functions, written with the provided APIs. These functions could range from straightforward, one-off functions to more complex generator functions. A simple example is demonstrated below, where the target speed of the ego vehicle is altered: + +```python +def decrease_speed_by_5(): + set_target_speed(get_target_speed() - 5) +``` + +Moreover, LaMPilot supports the use of control structures such as if-else and loop statements, enabling the LLM to create dynamic feedback policies. An example of this is shown in Tab. 1. In this sample, the LLM generates a policy function overtake_using_right_lane. The function first checks if there is a right lane available and a vehicle in front of the ego vehicle. If both conditions are met, it enters a loop to wait for a safe opportunity to change into the right lane. Once in the right lane, the function enters another loop to monitor the distance to the target vehicle until it has successfully overtaken the target vehicle. + +# API Docs (A) + +You are a helpful assistant that writes python code to complete any autonomous driving task specified by me. + +Here are the APIs you can use: ... [omitted] + +# Instruction (I) + +Go around the car in front of you using the right lane. + +# Driving Context (C) + +My current speed is $31.1\mathrm{m / s}$ . I am driving on a highway with 2 lanes in my direction, and I am in the 2nd lane from the right. There is a car in front of me in my lane, at a distance of $44.9\mathrm{m}$ , with a speed of $31.1\mathrm{m / s}$ . + +# Chain-of-Thought Reasoning (R) + +1) Check if there is a right lane available. 2) Check if there is a vehicle in front of the ego vehicle in the current lane within 100 meters. 3) If so, check if it is safe to enter the right lane. 4) If it is safe to enter the right lane, set the target lane of the ego vehicle to the right lane. 5) Monitor the distance to the target vehicle. 6) If the target vehicle is still in front of the ego vehicle, set the target speed of the ego vehicle to 1.5 times the speed of the target vehicle to overtake. 7) Repeat steps 5-6 until the target vehicle is no longer in front of the ego vehicle. + +# Code as Policies (P) + +Table 1. This example shows the chain-of-thought reasoning and code generation process of LaMPilot. Contents with blue titles are generated by LLMs. +```python +def overtake_using_right_lane( ): egoVehicle = get_egovehicle( ) current_lane = get_lane_of(egovehicle) right_lane = get_right_lane(egovehicle) targetvehicle $=$ detect_frontvehicle_in(current_lane) if right_lane is None: say("There is no right lane to change into.") return if targetvehicle is None: say("There is no vehicle in front of me.") return # Check if it is safe to enter the right lane while True: if is_safe-enter(right_lane): set_target_lane(right_lane) break yield autopilot( ) # Monitor the distance to the target vehicle while True: distance_to_target $=$ get_distance_between_vehicles(egovehicle, targetvehicle) if distance_to_target $< 0$ : # The target vehicle is still in front of the ego vehicle set_target_speed(get_speed_of(targetvehicle) $\ast$ 1.5) yield autopilot( ) else: break +``` + +The yield statements allow the AD system to take control and execute the low-level actions required to follow the high-level policy set by the LLM. + +This example demonstrates how LaMPilot leverages the reasoning capabilities of LLMs to generate dynamic, context-aware driving policies. By combining the strategic decision-making of LLMs with the real-time execution of classical AD algorithms, LaMPilot enables more flexible instruction following for AVs. + +# 4. LaMPilot Benchmark + +We introduce LaMPilot-Bench, the first benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. As shown in Fig. 1, LaMPilot-Bench consists of three key components: a simulator, a dataset, and an evaluator. + +# 4.1. Simulator + +The LaMPilot-Bench simulator is built upon HighwayEnv [21], a widely used platform for AD research and tactical decision-making. HighwayEnv offers various driving models and simulates realistic multi-vehicle interactions. We extend HighwayEnv with interfaces suitable for LLM-based agents and implement custom intersections to diversify the driving scenarios. + +# 4.2. Dataset + +The LaMPilot dataset consists of 4,900 semi-human-annotated traffic scenes, with a subset of 500 samples split as the test set. Each data sample includes: + +- An instruction I: a high-level task description. +- An initial state: used to initialize the simulator. +- Goal state criteria: aligned with the instruction I. + +The dataset covers diverse driving scenarios, as shown + +
StatisticValue
Total scenarios4,900 (100%)
Distance-related1,200 (24.5%)
Speed-related1,200 (24.5%)
Pulling over200 (4.1%)
Routing1,500 (30.2%)
Lane changing400 (8.2%)
Overtaking400 (8.2%)
Highway3,400 (69.4%)
Intersection1,500 (30.6%)
Average instruction length7.6 words
Maximum instruction length14 words
Minimum instruction length2 words
+ +Table 2. Main statistics of the LaMPilot-Bench dataset. + +![](images/a78d3836bf5c059ac368ac735d266976fad0daee2671418ec4a6c872ae995867.jpg) +Figure 2. Distribution of instructions in the LaMPilot-Bench dataset. + +in Table 2. For each driving scenario, LaMPilot includes various situations. Taking turning scenarios as an example, the diversity is reflected in several variables, such as the ego vehicle's initial position and state, the specific task (turning left/right or going straight), the number of other vehicles, and their positions and states. The driving model parameters for other vehicles are randomly initialized, and each scenario is assigned a random seed. + +The dataset also contains a variety of instructions reflecting realistic in-cabin human instructions categorized by maneuver types (e.g., routing, lane changing, overtaking) and scenario types (highway and intersection). Fig. 2 shows the distribution of the first four words in the instructions, highlighting their diversity. Main statistics are provided in Tab. 2. + +# 4.3. Evaluator + +The LaMPilot-Bench evaluator incorporates metrics to assess the safety and efficiency of the agent driving policies. + +Safety Metric Time-to-collision (TTC) is used to measure the vehicle's ability to maintain safe distances and avoid collisions. In a scenario with $n + 1$ vehicles (ego vehicle labeled as 0), the TTC $(\tau_{i})$ with vehicle $i$ is calculated using the ego vehicle's velocity $\mathbf{v}_0$ and position $\mathbf{p}_0$ , and vehicle $i$ 's velocity $\mathbf{v}_i$ and position $\mathbf{p}_i$ $(1 \leq i \leq n)$ : + +$$ +\tau_ {i} = - \frac {\left(\mathbf {p} _ {0} - \mathbf {p} _ {i}\right) \cdot \left(\mathbf {v} _ {0} - \mathbf {v} _ {i}\right)}{\left\| \mathbf {v} _ {0} - \mathbf {v} _ {i} \right\| ^ {2}} \tag {2} +$$ + +The minimum positive TTC value $\tau_{\mathrm{min}}$ among all $n$ vehicles over all time steps $t$ (up to task completion time $T$ ) represents the nearest potential collision: + +$$ +\tau_ {\min } = \min _ {1 \leq i \leq n, 1 \leq t \leq T} \tau_ {i} ^ {t} \tag {3} +$$ + +TTC scores are empirically assigned based on a 2-second safety margin, with values above 2 seconds considered safe and given a score of 100: + +$$ +\mathrm {T T C} = \left\{ \begin{array}{l l} 1 0 0 & \text {i f} \tau_ {\min } > 2 \\ \max \left(0, 1 0 0 - \frac {1}{\tau_ {\min }}\right) & \text {i f} 0 < \tau_ {\min } \leq 2 \\ 0 & \text {i f} \tau_ {\min } \leq 0 \end{array} \right. \tag {4} +$$ + +Speed variance (SV) is considered as another safety metric [14]. The ego vehicle's speed standard deviation $\sigma_0$ is calculated as: + +$$ +\sigma_ {0} = \sqrt {\frac {\sum_ {t = 1} ^ {T} \left(\left\| \mathbf {v} _ {0} ^ {t} \right\| - \mu_ {0}\right) ^ {2}}{T}}, \tag {5} +$$ + +where $\mu_0$ is the mean speed: + +$$ +\mu_ {0} = \frac {\sum_ {t = 1} ^ {T} \| \mathbf {v} _ {0} ^ {t} \|}{T}. \tag {6} +$$ + +The SV score is defined as: + +$$ +\mathrm {S V} = \max (0, 1 0 0 \cdot \left(1 - \sigma_ {0} / \sigma_ {\text {s a f e}}\right)), \tag {7} +$$ + +where $\sigma_{\mathrm{safe}}$ is the maximum safe speed deviation, determined empirically. + +Efficiency Metric The time efficiency (TE) score evaluates the policy's ability to complete the task within a predefined time limit $T_{\mathrm{limit}}$ : + +$$ +\mathrm {T E} = \max (0, 1 0 0 \cdot (1 - T / T _ {\text {l i m i t}})) \tag {8} +$$ + +A TE score closer to 100 indicates a more efficient policy. + +![](images/1024b41434a6ec671064faa2aa9ab6698b19f8974612e471cb7fce6bad21d21d.jpg) +Figure 3. Schematic view of the human feedback baseline based on retrieval-augmented generation. + +Task Completion Criteria A task is considered successfully completed when the agent achieves the objectives specified in the instruction while maintaining safety (i.e., avoiding collisions) and efficiency (i.e., finishing within the allotted time). For example, a lane change task is completed when the vehicle is in the target lane with its heading aligned with the lane's direction within a specified threshold. + +Overall Scoring The final score aggregates all individual metrics, weighted according to their importance: + +$$ +\mathrm {S c o r e} = W _ {\mathrm {T T C}} \cdot \mathrm {T T C} + W _ {\mathrm {S V}} \cdot \mathrm {S V} + W _ {\mathrm {T E}} \cdot \mathrm {T E} \tag {9} +$$ + +# 5. Experiments and Results + +Our experiments are designed to establish baselines for the proposed LaMPilot-Bench. The main objectives are to assess the performance of LLM-based agents in interpreting human instructions within driving contexts and to evaluate the capability of LLMs to generate code for motion planning using provided functional primitives. + +# 5.1. Experiment Setup + +Models We conduct benchmarking using various state-of-the-art large language models, which include both open-source and proprietary solutions1: + +- Llama 2 [45] (llama-2-70b-chat) +PaLM 2 [1](code-bison) +- ChatGPT [34] (gpt-3.5-turbo) +GPT-4 [33] (gpt-4) +GPT-4-Turbo [32] (gpt-4-1106-preview) + +Code Execution To execute a LMP, the Python exec function is used. It takes the LMP code as an input string, along with two dictionaries that define the execution scope: (i) apis, which includes all APIs the code may invoke, and (ii) policies, an initially empty dictionary that will eventually contain a policy variable. If the LMP is designed to return a generator, this generator is extracted from + +the policies dictionary after the execution of the exec function. + +Evaluation We set a 60-second time limit for each scenario. If a task is not completed within this time frame, the test case is considered a failure, and the simulation is terminated, resulting in a driving score of 0. For successful cases, the driving score is calculated as follows: + +$$ +\text {D r i v i n g S c o r e} = \frac {\alpha}{N _ {s}} \sum_ {i = 1} ^ {N _ {s}} \operatorname {S c o r e} _ {i} - \beta \cdot P _ {\text {c o l l i s i o n}} \tag {10} +$$ + +Here, $N_{s}$ represents the number of successful test cases, $\alpha$ is the success rate (ranging from 0 to 1), $\beta$ is the collision rate (also between 0 and 1), and $P_{\mathrm{collision}}$ is the penalty factor for collisions (set at 500 in our experiments). The Scorei for each individual sample is calculated based on Eq. (9), with weights $W_{\mathrm{TTC}}$ , $W_{\mathrm{SV}}$ , and $W_{\mathrm{TE}}$ set to 0.5, 0.3, and 0.2, respectively. The driving score provides a comprehensive assessment of the agent's performance in LaMPilot-Bench. + +# 5.2. Baselines + +# 5.2.1 Heuristic Baselines + +In autonomous driving, rule-based models are favored for their deterministic and interpretable nature. In our experiments, we employ two rule-based baseline policies: the Intelligent Driver Model (IDM)[46] and the Minimizing Overall Braking Induced by Lane Changes (MOBIL) principle[20]. IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front. MOBIL is an extension of IDM where a lane change is executed if the prospective new lane offers a more favorable driving scenario and the maneuver can be conducted safely. These baselines can be considered as operating on random chance, as the policy is independent of user instructions but follows predefined rules. Additionally, we include a human performance baseline, where a licensed human driver controls the vehicle in the simulation using arrow keys on the keyboard, following the commands and visuals displayed. This baseline provides a reference for human-level performance on LaMPilot-Bench. + +# 5.2.2 Zero-Shot and Few-Shot Baselines + +We also include the zero-shot and few-shot chain-of-thought (CoT) prompting [54] as baseline methods. The zero-shot CoT baseline leverages the prompt "How to complete the task step by step." In the few-shot setting, we follow the standard practice [2] by including 3 human-written code exemplars before the test instance. These in-context + +
MethodLearningCollision (↓)Completion (↑)TTC Score (↑)SV Score (↑)TE Score (↑)Driving Score (↑)
Heuristic Baselines
IDMN/A0.0%20.4%92.880.971.017.3
MOBIL0.0%15.3%82.885.946.711.7
Human Driver0.0%98.0%93.486.381.384.6
LLM-based Agents
Llama 20-Shot0.0%20.4%92.881.468.917.3
PaLM 23.1%35.7%78.676.578.812.8
ChatGPT1.0%40.8%83.975.574.027.8
GPT-44.1%72.4%61.370.574.028.5
GPT-4 Turbo3.1%74.5%73.270.173.939.1
Llama 23-Shot1.0% (+1.0)63.3% (+42.9)68.3 (-24.5)71.4 (-10.0)73.8 (+4.9)39.6 (+22.3)
PaLM 23.1% (0.0)71.4% (+35.7)73.4 (-5.2)69.7 (-6.8)72.0 (-6.8)36.6 (+23.8)
ChatGPT4.1% (+3.1)83.7% (+42.9)70.9 (-13.0)73.7 (-1.8)77.7 (+3.7)41.2 (+13.4)
GPT-41.0% (-3.1)84.7% (+12.3)69.4 (+8.1)72.0 (+1.5)76.7 (+2.7)55.9 (+27.4)
GPT-4 Turbo1.0% (-2.1)80.6% (+6.1)69.1 (-4.1)71.5 (+1.4)74.9 (+1.0)52.6 (+13.5)
Llama 2Human Feedback0.9% (-0.1)32.7% (-30.6)83.9 (+15.6)74.6 (+3.2)74.6 (+0.8)21.4 (-18.2)
PaLM 20.0% (-3.1)45.5% (-25.9)81.1 (+7.7)74.4 (+4.7)77.6 (+5.6)35.6 (-1.0)
ChatGPT0.9% (-3.2)80.9% (-2.8)70.2 (-0.7)71.7 (-2.0)77.5 (-0.2)54.2 (+13.0)
GPT-40.9% (-0.1)92.7% (+8.0)67.6 (-1.8)72.2 (+0.2)76.8 (+0.1)64.0 (+8.1)
GPT-4 Turbo0.9% (-0.1)87.3% (+6.7)74.1 (+5.0)73.4 (+1.9)75.5 (+0.6)60.5 (+7.9)
+ +Table 3. Comparison of baselines on LaMPilot-Bench under different learning settings. Arrows $(\uparrow/\downarrow)$ indicate higher or lower scores are better, respectively. Numbers in parentheses show the performance change compared to the previous setting, with (green/red) indicating performance improvements and degradation, respectively. + +examples help the model adapt to the tasks in LaMPilot-Bench. The in-context examples are created by a programmer proficient in Python. They are provided with API descriptions and are allowed to write and test their code using the non-test scenarios. The same set of three examples are used for all test cases. The API Docs are also the same across all test cases for both zero-shot and few-shot settings. + +# 5.2.3 Human Feedback Baselines + +LLMs have demonstrated proficiency in generating coherent solutions across various tasks without requiring additional fine-tuning. However, when it comes to code generation, especially for complex scenarios, they can produce suboptimal results. The autoregressive nature of these models poses a significant challenge, as tokens generated early in a sequence cannot be modified in the same iteration. This constraint limits the models' ability to refine initial responses, potentially impacting the effectiveness of the generated code [47]. + +To address these challenges and enhance the performance of LLMs on tasks in LaMPilot-Bench, we introduce a human-in-the-loop approach, where the LLMs serve not only as planners but also as engines for incorporating human feedback. As shown in Fig. 3, human feedback is infused into the learning process, transforming the LaMPilot framework from an open-loop system into an evolving feedback loop. After an LLM generates a policy program $(\mathbf{P})$ , it can integrate human feedback $(\mathbf{F})$ with context-specific guidance, enabling the LLM to refine its output. + +This approach does not involve fine-tuning or training a new LLM. Instead, we leverage the RAG approach [22] and introduce a code repository module, which is implemented with the Chroma [5] vector database. This repository allows for the storage and retrieval of effective code snippets to be used in similar situations. Following Voyager [47], we use LLM-generated function descriptions, which are then converted into vectors to serve as database keys. These keys are paired with the corresponding function codes as their values. If the human is satisfied with the policy, the code is committed to the repository for future reference. The human feedback learning process is conducted on the non-test set, involving interaction with 100 random samples for each method to ensure fair comparisons. + +# 5.3.Quantitative Results + +In this section, we present the experimental results on the LaMPilot-Bench, summarizing the performance of various methods, including heuristic baselines, 0-shot and 3-shot baselines, and human feedback baselines. These results are shown in Tab. 3. + +Rule-Based Methods We first evaluate the performance of rule-based approaches, specifically the IDM and MOBIL algorithms. Both methods achieve a zero collision rate in LaMPilot-Bench. However, it is important to note that these methods operate independently of the provided instructions. Without considering human instructions, IDM and MOBIL achieve success rates of $20.4\%$ and $15.3\%$ , respectively. These results provide a reference point for as- + +sessing the effectiveness of LLM-based agents in following human instructions. + +LLM-Based Agents We assess the effectiveness of off-the-shelf LLMs to reason in the context of autonomous driving and following human instructions. In the 0-shot setting, GPT models and PaLM 2, given only the API Docs, driving context, and instructions, achieve an obvious performance advantage compared to rule-based methods in terms of task completion. This improvement can be attributed to the LLMs' ability to understand and generate code based on natural language instructions, leveraging their extensive pretraining on diverse data sources. However, the 0-shot setting also results in an increase in the collision rate (1%-4%), indicating that the generated code policies do not fully capture the complexities of driving tasks. + +When provided with training examples that include exemplar code (3-shot setting), all the evaluated LLMs exhibit notable improvements in completion rates. For instance, Llama 2's completion rate increases from $20.4\%$ in the 0-shot setting to $63.3\%$ in the 3-shot setting. This improvement demonstrates the effectiveness of few-shot learning in adapting LLMs to specific task domains, as the exemplar code provides a clearer understanding of the desired behavior and API usage patterns. + +The integration of GPT models with human-in-the-loop learning further enhances their performance in driving tasks. Notably, GPT-4 achieves the highest driving score of 64.0. This improvement highlights the great potential of LLMs in following instructions in the driving context when combined with human feedback. The human feedback loop allows for iterative refinement of the generated policies, enabling the LLMs to learn from their mistakes and continuously incorporate human knowledge to generate more effective driving strategies. + +# 5.4. Discussion + +Safety Safety is a critical factor for any AD system [52]. In our context, ensuring that the system does not execute an unsafe action, even if instructed to do so by the human, is essential for responsible autonomy. Our framework includes safety checks at two levels: (1) When the LLM generates code based on human instructions, it first reasons about the safety of the action in the context of the current driving environment. By leveraging the vast knowledge and reasoning capabilities of LLMs, the system can identify potentially unsafe instructions and generate alternative actions that prioritize safety. (2) Even if the LLM fails to catch an unsafe instruction, we have a second layer of safety checks built into the APIs that interface with the vehicle actuators. These APIs encapsulate classical planning and control algorithms with hard safety constraints. This level of safety check serves as a fail-safe mechanism, ensuring that the sys + +tem does not execute actions that violate safety constraints, regardless of the LLM's output. + +Instruction Tuning Instruction tuning has been validated as a technique for LLMs to acquire domain-specific knowledge efficiently [3]. Fine-tuning LLMs on driving-related tasks could enhance their performance and reduce collision rates by providing them with more targeted knowledge specific to driving. However, this may also jeopardize their generalization capability when there exist some spurious correlations in the finetuning datasets (e.g. between driving context and policy) [62]. Therefore, we hope that LaMPilot-Bench can provide a platform to encourage future researchers to explore generalizable solutions to enhance LLMs' abilities in driving tasks. + +Multimodal LLMs Exploring the integration of vision modality in multimodal LLMs for AD tasks is another important future direction. Our current framework relies on textual input and output, which limits its ability to perceive and understand the visual world. By incorporating visual perception and understanding capabilities, multimodal LLMs could process and reason about real-world driving scenes more effectively. + +# 6. Conclusion + +In this paper, we introduced LaMPilot, a novel framework that integrates Large Language Models (LLMs) into autonomous driving (AD) systems, enabling human-like interaction in diverse driving contexts. We also created LaMPilot-Bench, a benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. Our experiments demonstrated that off-the-shelf LLMs can generate code policies for driving tasks based on human instructions. However, the notable collision rate indicates the need for further research to fully capture the complexities and safety requirements of real-world driving scenarios. LaMPilot provides a starting point for leveraging LLMs in solving AD-related tasks. As LLMs continue to evolve and integrate with more modalities, we envision a future where they can significantly contribute to the development of safe, efficient, and user-friendly autonomous vehicles. + +# Acknowledgment + +This work is partially funded by the Digital Twin Roadmap of InfoTech Labs, Toyota Motor North America. The contents of this paper only reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views of Toyota Motor North America. + +# References + +[1] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Stefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov and Yonghui Wu. PaLM 2 Technical Report. arXiv, 2023. 6 +[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In NeurIPS, 2020. 6 +[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Dataset for Map and Traffic Scene Understanding. In CVPR, 2024. 8 +[4] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving. In ICRA, 2024. 2 +[5] Chroma. The AI-native open-source embedding database, + +2023. URL https://github.com/chroma-core/chroma.7 +[6] Can Cui, Zichong Yang, Yupeng Zhou, Yunsheng Ma, Juanwu Lu, Lingxi Li, Yaobin Chen, Jitesh Panchal, and Ziran Wang. Large Language Models for Autonomous Driving: Real-World Experiments. arXiv, 2023. 2 +[7] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as You Speak: Enabling Human-Like Interaction with Large Language Models in Autonomous Vehicles. In WACVW, 2024. 2 +[8] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Receive, Reason, and React: Drive as You Say with Large Language Models in Autonomous Vehicles. IEEE Intelligent Transportation Systems Magazine, 2024. 2 +[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A Survey on Multimodal Large Language Models for Autonomous Driving. In WACVW, 2024. 1 +[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. HiLM-D: Towards High-Resolution Understanding in Multimodal Large Language Models for Autonomous Driving. arXiv, 2023. 2 +[11] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. LimSim++: A Closed-Loop Platform for Deploying Multimodal LLMs in Autonomous Driving. arXiv, 2024. 2 +[12] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive Like a Human: Rethinking Autonomous Driving with Large Language Models. In WACvw, 2024. 2 +[13] Haoxiang Gao, Yaqian Li, Kaiwen Long, Ming Yang, and Yiqing Shen. A Survey for Foundation Models in Autonomous Driving. arXiv, 2024. 1 +[14] Nicholas J. Garber and Ravi Gadiraju. Factors affecting speed variance and its influence on accidents. Transportation research record, 1989. 5 +[15] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 2020. 1 +[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1 +[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. In CoRL, 2023. 2 +[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 1 +[19] Ye Jin, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, and + +Jiangtao Gong. SurrealDriver: Designing Generative Driver Agent Simulation Framework in Urban Contexts based on Large Language Model. arXiv, 2023. 2 +[20] Arne Kesting, Martin Treiber, and Dirk Helbing. General Lane-Changing Model MOBIL for Car-Following Models. Transportation Research Record, 2007. 6 +[21] Edouard Leurent. An Environment for Autonomous Driving Decision-Making, 2018. URL https://github.com/Farama-Foundation/HighwayEnv.4 +[22] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In NeurIPS, 2020. 7 +[23] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving Everywhere with Large Language Model Policy Adaptation. In CVPR, 2024. 2 +[24] Xin Li, Yeqi Bai, Pinlong Cai, Licheng Wen, Daocheng Fu, Bo Zhang, Xuemeng Yang, Xinyu Cai, Tao Ma, Jianfei Guo, Xing Gao, Min Dou, Botian Shi, Yong Liu, Liang He, and Yu Qiao. Towards Knowledge-driven Autonomous Driving. arXiv, 2023. 1 +[25] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control. In ICRA, 2023. 1, 2 +[26] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models. In NeurIPS, 2023. 2 +[27] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. GPT-Driver: Learning to Drive with GPT. NeurIPS Workshop on Foundation Models for Decision Making, 2023. 2 +[28] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A Language Agent for Autonomous Driving. arXiv, 2023. 2 +[29] Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. LingoQA: Video Question Answering for Autonomous Driving. arXiv, 2023. 2 +[30] McKinsey Center for Future Mobility. Autonomous driving's future: convenient and connected, 2023. URL https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/autonomous-drivings-future-convenient-and-connected.1 +[31] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving. arXiv, 2023. 2 +[32] OpenAI. GPT-4-Turbo with 128K context and lower prices, the new Assistants API, GPT-4 Turbo with Vision, DALL-E 3 API, and more, 2023. URL https://openai.com/ + +blog/new-models-and-developer-products-announced-at-devday.6 +[33] OpenAI. GPT-4 Technical Report. arXiv, 2023. 6 +[34] OpenAI. Introducing ChatGPT, 2023. URL https://openai.com/blog/chatgpt.6 +[35] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G. Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In CVPR, 2024. 2 +[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving. arXiv, 2023. 2 +[37] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. On a Formal Model of Safe and Scalable Self-driving Cars. arXiv, 2017. 3 +[38] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. ReasonNet: End-to-End Driving with Temporal and Global Reasoning. In CVPR. arXiv, 2023. 1 +[39] Hao Shao, Yuxuan Hu, Letian Wang, Steven L. Waslander, Yu Liu, and Hongsheng Li. LMDrive: Closed-Loop End-to-End Driving with Large Language Models. In CVPR, 2024. 2 +[40] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv, 2023. 2 +[41] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. In ICRA, 2023. 2 +[42] Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming Zhang, Pheng Ann Heng, Jifeng Dai, Ping Luo, Jingdong Wang, Ji-Rong Wen, Xipeng Qiu, Yike Guo, Hui Xiong, Qun Liu, and Zhenguo Li. A Survey of Reasoning with Foundation Models. arXiv, 2024. 1 +[43] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In ICCV, 2023. 2 +[44] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models. arXiv, 2024. 2 +[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik + +tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and FineTuned Chat Models. arXiv, 2023. 6 +[46] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 2000. 6 +[47] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv, 2023. 7 +[48] Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Zenglin Xu, Shaoliang Nie, Sinong Wang, Madian Khabsa, Hamed Firooz, and Dongfang Liu. MUSTIE: Multimodal Structural Transformer for Web Information Extraction. In ACL, 2023. 2 +[49] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. DriveMLM: Aligning Multi-Modal Large Language Models with Behavioral Planning States for Autonomous Driving. arXiv, 2023. 2 +[50] Xingyao Wang, Sha Li, and Heng Ji. Code4Struct: Code Generation for Few-Shot Event Structure Prediction. In ACL, 2023. 2 +[51] Xingyao Wang, Hao Peng, Reyhaneh Jabbarvand, and Heng Ji. LeTI: Learning to Generate from Textual Interactions. arXiv, 2023. 2 +[52] Yixuan Wang, Ruochen Jiao, Chengtian Lang, Sinong Simon Zhan, Chao Huang, Zhaoran Wang, Zhuoran Yang, and Qi Zhu. Empowering Autonomous Driving with Large Language Models: A Safety Perspective. In ICLR Workshop on LLM Agents, 2024. 8 +[53] Wayve Technologies Ltd. LINGO-1: Exploring Natural Language for Autonomous Driving, 2023. URL https://wayve.ai/thinking/lingo-natural-language-autonomous-driving/.2 +[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS, 2022. 2, 6 +[55] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, Tao Ma, Yingxuan Li, Linran Xu, Dengke Shang, Zheng Zhu, Shaoyan Sun, Yeqi Bai, Xinyu Cai, Min Dou, Shuanglu Hu, Botian Shi, and Yu Qiao. On the Road with GPT-4V(ision): Early Explorations of Visual-Language Model on Autonomous Driving. arXiv, 2023. 2 + +[56] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models. In ICLR, 2024. 2 +[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhengu Li, and Hengshuang Zhao. DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model. arXiv, 2023. 2 +[58] Ke Yang, Jiateng Liu, John Wu, Chaoqi Yang, Yi R. Fung, Sha Li, Zixuan Huang, Xu Cao, Xingyao Wang, Yiquan Wang, Heng Ji, and Chengxiang Zhai. If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents. ICLR Workshop on LLM Agents, 2024. 2 +[59] Yi Yang, Qingwen Zhang, Ci Li, Daniel Simões Marta, Nazre Batool, and John Folkesson. Human-Centric Autonomous Systems With LLMs for User Command Reasoning. In WACVW, 2024. 2 +[60] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. LLM4Drive: A Survey of Large Language Models for Autonomous Driving. arXiv, 2023. 1 +[61] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In ICLR, 2023. 1 +[62] Wenqian Ye, Guangtao Zheng, Xu Cao, Yunsheng Ma, Xia Hu, and Aidong Zhang. Spurious Correlations in Machine Learning: A Survey. arXiv, 2024. 8 +[63] Jianhao Yuan, Shuyang Sun, Daniel Omeiza, Bo Zhao, Paul Newman, Lars Kunze, and Matthew Gadd. RAGDriver: Generalisable Driving Explanations with Retrieval-Augmented In-Context Learning in Multi-Modal Large Language Model. arXiv, 2024. 2 +[64] Daoguang Zan, Bei Chen, Fengji Zhang, Dianjie Lu, Bingchao Wu, Bei Guan, Yongji Wang, and Jian-Guang Lou. Large Language Models Meet NL2Code: A Survey. In ACL, 2023. 2 +[65] Xingcheng Zhou, Mingyu Liu, Bare Luka Zagar, Ekim Yurtsever, and Alois C. Knoll. Vision Language Models in Autonomous Driving and Intelligent Transportation Systems. arXiv, 2023. 1 +[66] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied Understanding of Driving Scenarios. arXiv, 2024. 2 \ No newline at end of file diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/images.zip b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..48c286bbe735b9adfef57f390f26337d9a97e3d9 --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee71b9f5ea21194206394c75a4819904a3b7f6c8be29f4de4ff531f5c60953d6 +size 435293 diff --git a/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/layout.json b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea2c2aaaea7aec867c67db8a0be3acdc9fa714be --- /dev/null +++ b/2024/LaMPilot_ An Open Benchmark Dataset for Autonomous Driving with Language Model Programs/layout.json @@ -0,0 +1,8971 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 515, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 515, + 139 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 515, + 139 + ], + "type": "text", + "content": "LaMPilot: An Open Benchmark Dataset for Autonomous Driving with Language Model Programs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": "Yunsheng Ma" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Can Cui" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Xu Cao" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Wenqian Ye" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Peiran Liu" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Juanwu Lu" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Amr Abdelraouf" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Rohit Gupta" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Kyungtae Han" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Aniket Bera" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", James M. Rehg" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": ", Ziran Wang" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": "Purdue University " + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign \n" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": "University of Virginia " + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 525, + 217 + ], + "type": "text", + "content": "InfoTech Labs, Toyota Motor North America" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 219, + 219, + 372, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 219, + 219, + 372, + 232 + ], + "spans": [ + { + "bbox": [ + 219, + 219, + 372, + 232 + ], + "type": "text", + "content": "{yunsheng, ziran}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 260, + 289, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 260, + 289, + 487 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 289, + 487 + ], + "type": "text", + "content": "Autonomous driving (AD) has made significant strides in recent years. However, existing frameworks struggle to interpret and execute spontaneous user instructions, such as \"overtake the car ahead.\" Large Language Models (LLMs) have demonstrated impressive reasoning capabilities showing potential to bridge this gap. In this paper, we present LaMPilot, a novel framework that integrates LLMs into AD systems, enabling them to follow user instructions by generating code that leverages established functional primitives. We also introduce LaMPilot-Bench, the first benchmark dataset specifically designed to quantitatively evaluate the efficacy of language model programs in AD. Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf LLMs on LaMPilot-Bench. Our results demonstrate the potential of LLMs in handling diverse driving scenarios and following user instructions in driving. To facilitate further research in this area, we release our code and data at GitHub.com/PurdueDigitalTwin/LaMPilot." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 497, + 128, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 128, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 128, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 659 + ], + "type": "text", + "content": "Autonomous driving (AD) has witnessed remarkable progress in recent years, with an increasing number of commercial autonomous vehicles (AVs) being deployed on public roads [30]. State-of-the-art AD systems can be broadly classified into two categories: 1) a modular approach, where standalone models are developed for perception, prediction, and planning independently [15], and 2) an end-to-end approach that directly maps sensor data to control signals via a single neural network [16, 18, 38]. Despite significant breakthroughs, both approaches struggle to handle arbitrary user commands effectively, such as \"overtake the car in front of me.\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 661, + 287, + 696 + ], + "type": "text", + "content": "Large Language Models (LLMs) have demonstrated impressive capabilities in language comprehension and reasoning [42, 61], showing potential to improve the safety, ex" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 259, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 545, + 367 + ], + "type": "text", + "content": "plainability, and user-friendliness of AVs. Utilizing LLMs to solve AV-related tasks is gaining momentum [9, 13, 24, 60, 65]. However, the integration of LLMs into existing AD frameworks presents several challenges. Firstly, there is a lack of well-established paradigms for incorporating LLMs into the decision-making process of AVs. Secondly, there is a shortage of benchmarks designed to evaluate and compare the performance of LLM-based agents in the context of driving." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 369, + 546, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 546, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 546, + 477 + ], + "type": "text", + "content": "To address these challenges, we propose a novel framework called LaMPilot. Inspired by Code as Policy [25], which utilizes code LLMs to write robot policy code, LaMPilot employs Language Model Programs (LMPs) as the action space instead of low-level vehicle control signals. We equip LLMs with APIs that cover various functional primitives, enabling them to connect natural language instructions to executable driving plans through code generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 479, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 575 + ], + "type": "text", + "content": "We also introduce LaMPilot-Bench, the first benchmark for evaluating LMPs in AD. The primary objective for agents operating within LaMPilot-Bench is to accomplish assigned tasks safely and efficiently. LaMPilot-Bench incorporates an interactive simulator and evaluator, featuring programmatic scoring mechanisms to assess policy performance. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 582, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 306, + 582, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 545, + 616 + ], + "type": "text", + "content": "- We propose LaMPilot, a novel framework that integrates LLMs into autonomous driving systems, enhancing their ability to interpret and follow user commands." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 618, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 618, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 618, + 545, + 688 + ], + "type": "text", + "content": "- We introduce LaMPilot-Bench, the first benchmark dataset designed to evaluate the performance of LLM-powered agents in autonomous driving. Each scenario in LaMPilot-Bench consists of a task described in natural language, along with a simulated environment for comprehensive evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "- Adopting the LaMPilot framework, we conduct extensive experiments to assess the performance of off-the-shelf" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "15141" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 287, + 109 + ], + "type": "text", + "content": "LLMs on LaMPilot-Bench. Our results demonstrate the great potential of LLMs in handling diverse driving scenarios and following user instructions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 139, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 121, + 139, + 134 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 139, + 134 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 141, + 194, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 141, + 194, + 155 + ], + "spans": [ + { + "bbox": [ + 47, + 141, + 194, + 155 + ], + "type": "text", + "content": "2.1. Language-Guided Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 159, + 286, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 159, + 286, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 286, + 231 + ], + "type": "text", + "content": "Recent studies highlight the potential role of language in enhancing self-driving technology. Companies like Wayve are integrating natural language to improve the learning and explainability of their driving models. For example, their LINGO-1 system combines vision, language, and action modalities [29, 53]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 232, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 286, + 376 + ], + "type": "text", + "content": "Researchers have also been incorporating LLMs into AD systems for various purposes, such as generalization [4, 8, 11, 12, 19, 23, 27, 28, 35, 44, 49, 55, 56, 66], interpretability [10, 31, 36, 57, 63], and human-vehicle interaction [6, 7, 39, 40, 49, 59]. DiLu [56], for instance, instills knowledge-driven capability into AD systems by considering how humans drive. It introduces a new framework that includes an interactive environment, a driver agent, and a memory component. LMDrive [39] proposes an end-to-end, language-based AD framework that processes camera-LiDAR sensor data, comprehends natural language driving instructions, and directly generates vehicle control signals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 286, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 471 + ], + "type": "text", + "content": "However, most existing works either provide high-level goals, let LLMs generate the ego vehicle's future trajectories, or involve detailed vehicle control signals like throttle and steering angles. In contrast, our work uniquely uses LLMs to generate program code as driving policies, which are then executed with classical planners or controllers. Our approach takes advantage of the strengths of both LLMs and traditional AD components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 284, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 284, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 284, + 493 + ], + "type": "text", + "content": "2.2. Large Language Models for Code Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "content": "The emergence of powerful LLMs has opened up new possibilities for leveraging programming languages to perform various tasks across multiple domains [58]. Code generation [64] is one of the most prominent applications, where LLMs are used to generate code snippets based on natural language descriptions. LLMs have also been employed for information extraction tasks using code as an intermediate representation [48, 50]. The robotics domain has also benefited from the integration of LLMs and code. Code as Policy [25], ProgPrompt[41], and VoxPoser [17] demonstrate how LLMs can generate code to control robots and manipulate objects in 3D environments. In the vision domain, works such as ViperGPT [43] showcase the potential of using LLMs to generate code for visual processing and reasoning. Moreover, code provides a foundation for logical planning [26, 51], which can be leveraged by LLMs to tackle complex, long-horizon tasks. The combination of LLMs and code has revealed a new paradigm with broad" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "applicability across numerous engineering and scientific domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 106, + 441, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 106, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 441, + 121 + ], + "type": "text", + "content": "3. LaMPilot Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 127, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 127, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 127, + 545, + 258 + ], + "type": "text", + "content": "This section elaborates on the framework of LaMPilot, which integrates the world knowledge and reasoning capabilities of LLMs into an AD system to enable user instruction following. LaMPilot uses LLMs to generate LMPs that serve as the action space. LMPs have the inherent potential to symbolize actions that are both temporally generalizable and hierarchically structured. For instance, a complex maneuver like overtaking can be programmatically split into a series of actions, including lane changes, acceleration, and possibly another lane change, composed with conditional logic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 258, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 354 + ], + "type": "text", + "content": "LaMPilot is designed as an add-on module for existing AD systems rather than as a standalone solution. It synergizes the emerging capabilities of LLMs with the proven efficiency of classical AD algorithms. Instead of directly issuing real-time, low-level control signals, the LLM generates code snippets at lower frequencies. These snippets, formulated as temporary policies, guide the strategic navigation of the ego vehicle based on user instructions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 361, + 369, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 369, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 369, + 374 + ], + "type": "text", + "content": "3.1. Prompts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 380, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 439 + ], + "type": "text", + "content": "As shown in Fig. 1, the LaMPilot framework takes inputs including API documentation (A), human instructions (I), and driving context (C). It leverages Chain-of-Thought (CoT) reasoning [54] to write policy code (P). This process can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 449, + 545, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 449, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 380, + 449, + 545, + 463 + ], + "type": "interline_equation", + "content": "[ \\mathbf {R}, \\mathbf {P} ] = \\ell ([ \\mathbf {A}, \\mathbf {I}, \\mathbf {C} ]) \\tag {1}", + "image_path": "62feda4ea7b22d82a5629a899547077dc39bf0f4adcedb2cead92582972240f2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 472, + 465, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 465, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 465, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 472, + 465, + 483 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 306, + 472, + 465, + 483 + ], + "type": "text", + "content": " represents the LLM backbone." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 484, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 543 + ], + "type": "text", + "content": "An example to demonstrate the code generation process is provided in Tab. 1. The instruction is a spontaneous command from the user. In addition, we also provide environment contexts and API documentation as part of the input prompt for the LLM." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": "The driving contexts include relevant information about the driving environment. We develop an interface that converts the numerical vector data in simulation into a narrative format using a structured language generator [4]. This narrative provides a natural input interface for the LLM and does not require additional fine-tuning for interpretation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 430, + 635 + ], + "type": "text", + "content": "3.2. Functional Primitives" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "Policy code involves calls to functional primitives. They are implemented with classical planning and control algorithms. Due to the autoregressive nature of LLMs, generating long completions can introduce significant latency, making them less suitable for time-critical tasks like collision avoidance. The main idea is to utilize LLMs for" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "15142" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 541, + 358 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 541, + 358 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 541, + 358 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 541, + 358 + ], + "type": "image", + "image_path": "0a4db3a0e9b548f93abe741be89f22a16b994c5bea1bc2893090b3645d650b94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 366, + 546, + 411 + ], + "lines": [ + { + "bbox": [ + 46, + 366, + 546, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 546, + 411 + ], + "type": "text", + "content": "Figure 1. An overview of the LaMPilot framework. The Large Language Model (LLM) receives a prompt containing human instructions, driving context, and API documentation. It then writes language model programs that serve as driving policies. These policies are executed in the simulator to complete the specified driving task and are subsequently evaluated by the evaluator to assess the effectiveness of the generated policy code." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 426, + 288, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 450 + ], + "type": "text", + "content": "strategic planning while intentionally avoiding their direct involvement in low-level control tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 454, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 527 + ], + "type": "text", + "content": "Inspired by the concept of Responsibility-Sensitive Safety [37], which ensures that the autonomous driving system will not issue a command that would lead to an accident, we include safety heuristics in our API implementations to minimize risks. Our API suite is categorized into four main types:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 529, + 287, + 638 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 287, + 553 + ], + "type": "text", + "content": "- Ego APIs: They provide information about the status of the ego vehicle." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 287, + 590 + ], + "type": "text", + "content": "- Perception APIs: They include functions like object and lane detection, which can be called to acquire information about the surrounding environment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 590, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 287, + 613 + ], + "type": "text", + "content": "- Navigation APIs: They provide route planning functionalities given a target destination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 287, + 638 + ], + "type": "text", + "content": "- Control APIs: They translate LLM-generated code into low-level control signals for the vehicle." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "In the event of an exception during policy execution, the ego vehicle automatically switches to a predefined autopilot mode to prevent undefined behaviors. To provide LLMs with essential information about the available APIs and guide their correct usage, we include the API Documentation (A) in the prompt. These documents include" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 426, + 545, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 451 + ], + "type": "text", + "content": "not only the input and output specifications but also provide descriptions of their usage and the underlying logic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 460, + 389, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 389, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 389, + 474 + ], + "type": "text", + "content": "3.3. Completions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 479, + 546, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 546, + 551 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 546, + 551 + ], + "type": "text", + "content": "The completion produced by the LLMs is anticipated to be valid functions, written with the provided APIs. These functions could range from straightforward, one-off functions to more complex generator functions. A simple example is demonstrated below, where the target speed of the ego vehicle is altered:" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 318, + 555, + 504, + 574 + ], + "blocks": [ + { + "bbox": [ + 318, + 555, + 504, + 574 + ], + "lines": [ + { + "bbox": [ + 318, + 555, + 504, + 574 + ], + "spans": [ + { + "bbox": [ + 318, + 555, + 504, + 574 + ], + "type": "text", + "content": "def decrease_speed_by_5():\n set_target_speed(get_target_speed() - 5)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 714 + ], + "type": "text", + "content": "Moreover, LaMPilot supports the use of control structures such as if-else and loop statements, enabling the LLM to create dynamic feedback policies. An example of this is shown in Tab. 1. In this sample, the LLM generates a policy function overtake_using_right_lane. The function first checks if there is a right lane available and a vehicle in front of the ego vehicle. If both conditions are met, it enters a loop to wait for a safe opportunity to change into the right lane. Once in the right lane, the function enters another loop to monitor the distance to the target vehicle until it has successfully overtaken the target vehicle." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "15143" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 80, + 118, + 90 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 80, + 118, + 90 + ], + "spans": [ + { + "bbox": [ + 69, + 80, + 118, + 90 + ], + "type": "text", + "content": "API Docs (A)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 90, + 419, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 90, + 419, + 99 + ], + "spans": [ + { + "bbox": [ + 69, + 90, + 419, + 99 + ], + "type": "text", + "content": "You are a helpful assistant that writes python code to complete any autonomous driving task specified by me." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 100, + 212, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 100, + 212, + 109 + ], + "spans": [ + { + "bbox": [ + 69, + 100, + 212, + 109 + ], + "type": "text", + "content": "Here are the APIs you can use: ... [omitted]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 110, + 120, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 110, + 120, + 118 + ], + "spans": [ + { + "bbox": [ + 69, + 110, + 120, + 118 + ], + "type": "text", + "content": "Instruction (I)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 119, + 242, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 119, + 242, + 128 + ], + "spans": [ + { + "bbox": [ + 69, + 119, + 242, + 128 + ], + "type": "text", + "content": "Go around the car in front of you using the right lane." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 129, + 141, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 129, + 141, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 129, + 141, + 137 + ], + "type": "text", + "content": "Driving Context (C)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "text", + "content": "My current speed is " + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "inline_equation", + "content": "31.1\\mathrm{m / s}" + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "text", + "content": ". I am driving on a highway with 2 lanes in my direction, and I am in the 2nd lane from the right. There is a car in front of me in my lane, at a distance of " + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "inline_equation", + "content": "44.9\\mathrm{m}" + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "text", + "content": ", with a speed of " + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "inline_equation", + "content": "31.1\\mathrm{m / s}" + }, + { + "bbox": [ + 69, + 137, + 522, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 165, + 186, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 165, + 186, + 175 + ], + "spans": [ + { + "bbox": [ + 69, + 165, + 186, + 175 + ], + "type": "text", + "content": "Chain-of-Thought Reasoning (R)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 175, + 522, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 175, + 522, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 522, + 213 + ], + "type": "text", + "content": "1) Check if there is a right lane available. 2) Check if there is a vehicle in front of the ego vehicle in the current lane within 100 meters. 3) If so, check if it is safe to enter the right lane. 4) If it is safe to enter the right lane, set the target lane of the ego vehicle to the right lane. 5) Monitor the distance to the target vehicle. 6) If the target vehicle is still in front of the ego vehicle, set the target speed of the ego vehicle to 1.5 times the speed of the target vehicle to overtake. 7) Repeat steps 5-6 until the target vehicle is no longer in front of the ego vehicle." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 213, + 140, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 213, + 140, + 222 + ], + "spans": [ + { + "bbox": [ + 69, + 213, + 140, + 222 + ], + "type": "text", + "content": "Code as Policies (P)" + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 69, + 230, + 462, + 432 + ], + "blocks": [ + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "lines": [ + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "spans": [ + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "text", + "content": "def overtake_using_right_lane( ): egoVehicle = get_egovehicle( ) current_lane = get_lane_of(egovehicle) right_lane = get_right_lane(egovehicle) targetvehicle " + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "text", + "content": " detect_frontvehicle_in(current_lane) if right_lane is None: say(\"There is no right lane to change into.\") return if targetvehicle is None: say(\"There is no vehicle in front of me.\") return # Check if it is safe to enter the right lane while True: if is_safe-enter(right_lane): set_target_lane(right_lane) break yield autopilot( ) # Monitor the distance to the target vehicle while True: distance_to_target " + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "inline_equation", + "content": "=" + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "text", + "content": " get_distance_between_vehicles(egovehicle, targetvehicle) if distance_to_target " + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "inline_equation", + "content": "< 0" + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "text", + "content": " : # The target vehicle is still in front of the ego vehicle set_target_speed(get_speed_of(targetvehicle) " + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 69, + 230, + 462, + 432 + ], + "type": "text", + "content": " 1.5) yield autopilot( ) else: break" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + }, + { + "bbox": [ + 46, + 468, + 545, + 491 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 545, + 491 + ], + "type": "text", + "content": "Table 1. This example shows the chain-of-thought reasoning and code generation process of LaMPilot. Contents with blue titles are generated by LLMs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + } + ], + "index": 10, + "sub_type": "code", + "guess_lang": "python" + }, + { + "bbox": [ + 46, + 498, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 535 + ], + "type": "text", + "content": "The yield statements allow the AD system to take control and execute the low-level actions required to follow the high-level policy set by the LLM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 538, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 609 + ], + "type": "text", + "content": "This example demonstrates how LaMPilot leverages the reasoning capabilities of LLMs to generate dynamic, context-aware driving policies. By combining the strategic decision-making of LLMs with the real-time execution of classical AD algorithms, LaMPilot enables more flexible instruction following for AVs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 630, + 174, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 630, + 174, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 630, + 174, + 642 + ], + "type": "text", + "content": "4. LaMPilot Benchmark" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "We introduce LaMPilot-Bench, the first benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. As shown in Fig. 1, LaMPilot-Bench consists of three key components: a simulator, a dataset, and an evaluator." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 375, + 510 + ], + "type": "text", + "content": "4.1. Simulator" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 517, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 601 + ], + "type": "text", + "content": "The LaMPilot-Bench simulator is built upon HighwayEnv [21], a widely used platform for AD research and tactical decision-making. HighwayEnv offers various driving models and simulates realistic multi-vehicle interactions. We extend HighwayEnv with interfaces suitable for LLM-based agents and implement custom intersections to diversify the driving scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 609, + 364, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 364, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 364, + 621 + ], + "type": "text", + "content": "4.2. Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 664 + ], + "type": "text", + "content": "The LaMPilot dataset consists of 4,900 semi-human-annotated traffic scenes, with a subset of 500 samples split as the test set. Each data sample includes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 666, + 512, + 700 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 306, + 666, + 502, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 502, + 676 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 502, + 676 + ], + "type": "text", + "content": "- An instruction I: a high-level task description." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 677, + 504, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 504, + 688 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 504, + 688 + ], + "type": "text", + "content": "- An initial state: used to initialize the simulator." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 689, + 512, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 512, + 700 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 512, + 700 + ], + "type": "text", + "content": "- Goal state criteria: aligned with the instruction I." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "The dataset covers diverse driving scenarios, as shown" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "15144" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 261, + 222 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 261, + 222 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 261, + 222 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 261, + 222 + ], + "type": "table", + "html": "
StatisticValue
Total scenarios4,900 (100%)
Distance-related1,200 (24.5%)
Speed-related1,200 (24.5%)
Pulling over200 (4.1%)
Routing1,500 (30.2%)
Lane changing400 (8.2%)
Overtaking400 (8.2%)
Highway3,400 (69.4%)
Intersection1,500 (30.6%)
Average instruction length7.6 words
Maximum instruction length14 words
Minimum instruction length2 words
", + "image_path": "fd796f27a55644ee3ffaf03a9f31df5d73ef2da9344a394cbfcbb89c1e7234f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 229, + 267, + 240 + ], + "lines": [ + { + "bbox": [ + 66, + 229, + 267, + 240 + ], + "spans": [ + { + "bbox": [ + 66, + 229, + 267, + 240 + ], + "type": "text", + "content": "Table 2. Main statistics of the LaMPilot-Bench dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 254, + 275, + 470 + ], + "blocks": [ + { + "bbox": [ + 60, + 254, + 275, + 470 + ], + "lines": [ + { + "bbox": [ + 60, + 254, + 275, + 470 + ], + "spans": [ + { + "bbox": [ + 60, + 254, + 275, + 470 + ], + "type": "image", + "image_path": "a78d3836bf5c059ac368ac735d266976fad0daee2671418ec4a6c872ae995867.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 477, + 287, + 499 + ], + "lines": [ + { + "bbox": [ + 46, + 477, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 477, + 287, + 499 + ], + "type": "text", + "content": "Figure 2. Distribution of instructions in the LaMPilot-Bench dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 529, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 625 + ], + "type": "text", + "content": "in Table 2. For each driving scenario, LaMPilot includes various situations. Taking turning scenarios as an example, the diversity is reflected in several variables, such as the ego vehicle's initial position and state, the specific task (turning left/right or going straight), the number of other vehicles, and their positions and states. The driving model parameters for other vehicles are randomly initialized, and each scenario is assigned a random seed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 712 + ], + "type": "text", + "content": "The dataset also contains a variety of instructions reflecting realistic in-cabin human instructions categorized by maneuver types (e.g., routing, lane changing, overtaking) and scenario types (highway and intersection). Fig. 2 shows the distribution of the first four words in the instructions, highlighting their diversity. Main statistics are provided in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 375, + 83 + ], + "type": "text", + "content": "4.3. Evaluator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 90, + 545, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 90, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 306, + 90, + 545, + 114 + ], + "type": "text", + "content": "The LaMPilot-Bench evaluator incorporates metrics to assess the safety and efficiency of the agent driving policies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": "Safety Metric Time-to-collision (TTC) is used to measure the vehicle's ability to maintain safe distances and avoid collisions. In a scenario with " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": " vehicles (ego vehicle labeled as 0), the TTC " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "(\\tau_{i})" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": " with vehicle " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": " is calculated using the ego vehicle's velocity " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": " and position " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_0" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": ", and vehicle " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": "'s velocity " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": " and position " + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "inline_equation", + "content": "(1 \\leq i \\leq n)" + }, + { + "bbox": [ + 305, + 129, + 545, + 201 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 209, + 545, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 209, + 545, + 235 + ], + "spans": [ + { + "bbox": [ + 365, + 209, + 545, + 235 + ], + "type": "interline_equation", + "content": "\\tau_ {i} = - \\frac {\\left(\\mathbf {p} _ {0} - \\mathbf {p} _ {i}\\right) \\cdot \\left(\\mathbf {v} _ {0} - \\mathbf {v} _ {i}\\right)}{\\left\\| \\mathbf {v} _ {0} - \\mathbf {v} _ {i} \\right\\| ^ {2}} \\tag {2}", + "image_path": "9d62693f202042c0823d34c950edf06848088bbb37dbb4121f9be9b9cccae8bb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "content": "The minimum positive TTC value " + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{min}}" + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "content": " among all " + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "content": " vehicles over all time steps " + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "content": " (up to task completion time " + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 305, + 238, + 545, + 274 + ], + "type": "text", + "content": ") represents the nearest potential collision:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 376, + 285, + 545, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 285, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 376, + 285, + 545, + 304 + ], + "type": "interline_equation", + "content": "\\tau_ {\\min } = \\min _ {1 \\leq i \\leq n, 1 \\leq t \\leq T} \\tau_ {i} ^ {t} \\tag {3}", + "image_path": "14bc9c50991ff0e0e4b73b1e0c2363404a0392c7992d76b4bdf3384de08fb11a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 308, + 545, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 343 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 343 + ], + "type": "text", + "content": "TTC scores are empirically assigned based on a 2-second safety margin, with values above 2 seconds considered safe and given a score of 100:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 362, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 362, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 318, + 362, + 545, + 407 + ], + "type": "interline_equation", + "content": "\\mathrm {T T C} = \\left\\{ \\begin{array}{l l} 1 0 0 & \\text {i f} \\tau_ {\\min } > 2 \\\\ \\max \\left(0, 1 0 0 - \\frac {1}{\\tau_ {\\min }}\\right) & \\text {i f} 0 < \\tau_ {\\min } \\leq 2 \\\\ 0 & \\text {i f} \\tau_ {\\min } \\leq 0 \\end{array} \\right. \\tag {4}", + "image_path": "13f150d304a767d6a46f84191476d81a15856776203cfadc00ca77be46e6b3a2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "text", + "content": "Speed variance (SV) is considered as another safety metric [14]. The ego vehicle's speed standard deviation " + }, + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "inline_equation", + "content": "\\sigma_0" + }, + { + "bbox": [ + 305, + 414, + 545, + 449 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 459, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 459, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 364, + 459, + 545, + 491 + ], + "type": "interline_equation", + "content": "\\sigma_ {0} = \\sqrt {\\frac {\\sum_ {t = 1} ^ {T} \\left(\\left\\| \\mathbf {v} _ {0} ^ {t} \\right\\| - \\mu_ {0}\\right) ^ {2}}{T}}, \\tag {5}", + "image_path": "9e009a7e1faf79f6867a7def34536c0d2f4f254c5cafc5afd0e2bc934742769c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 495, + 434, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 495, + 434, + 507 + ], + "spans": [ + { + "bbox": [ + 318, + 495, + 434, + 507 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 318, + 495, + 434, + 507 + ], + "type": "inline_equation", + "content": "\\mu_0" + }, + { + "bbox": [ + 318, + 495, + 434, + 507 + ], + "type": "text", + "content": " is the mean speed:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 514, + 545, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 514, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 386, + 514, + 545, + 541 + ], + "type": "interline_equation", + "content": "\\mu_ {0} = \\frac {\\sum_ {t = 1} ^ {T} \\| \\mathbf {v} _ {0} ^ {t} \\|}{T}. \\tag {6}", + "image_path": "92b54617c670076f0d65d44bbccfa69c6f954826da01eefd2d2ad73f27dc67fd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 318, + 544, + 429, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 544, + 429, + 555 + ], + "spans": [ + { + "bbox": [ + 318, + 544, + 429, + 555 + ], + "type": "text", + "content": "The SV score is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 567, + 545, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 567, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 350, + 567, + 545, + 581 + ], + "type": "interline_equation", + "content": "\\mathrm {S V} = \\max (0, 1 0 0 \\cdot \\left(1 - \\sigma_ {0} / \\sigma_ {\\text {s a f e}}\\right)), \\tag {7}", + "image_path": "e76a64296b1debb3ed863769330e410eec78a9d4797f29689f7679c5c02e9b65.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 586, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 545, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 586, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{safe}}" + }, + { + "bbox": [ + 306, + 586, + 545, + 609 + ], + "type": "text", + "content": " is the maximum safe speed deviation, determined empirically." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 624, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 545, + 660 + ], + "type": "text", + "content": "Efficiency Metric The time efficiency (TE) score evaluates the policy's ability to complete the task within a predefined time limit " + }, + { + "bbox": [ + 306, + 624, + 545, + 660 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{limit}}" + }, + { + "bbox": [ + 306, + 624, + 545, + 660 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 351, + 671, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 671, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 351, + 671, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\mathrm {T E} = \\max (0, 1 0 0 \\cdot (1 - T / T _ {\\text {l i m i t}})) \\tag {8}", + "image_path": "81a64202a5432b2b72fddb244b03b7528d11af5063887022f49c17239e699bd8.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "A TE score closer to 100 indicates a more efficient policy." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 317, + 742 + ], + "type": "text", + "content": "15145" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 289, + 162 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 289, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 289, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 289, + 162 + ], + "type": "image", + "image_path": "1024b41434a6ec671064faa2aa9ab6698b19f8974612e471cb7fce6bad21d21d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 287, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 191 + ], + "type": "text", + "content": "Figure 3. Schematic view of the human feedback baseline based on retrieval-augmented generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 300 + ], + "type": "text", + "content": "Task Completion Criteria A task is considered successfully completed when the agent achieves the objectives specified in the instruction while maintaining safety (i.e., avoiding collisions) and efficiency (i.e., finishing within the allotted time). For example, a lane change task is completed when the vehicle is in the target lane with its heading aligned with the lane's direction within a specified threshold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 339 + ], + "type": "text", + "content": "Overall Scoring The final score aggregates all individual metrics, weighted according to their importance:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 348, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\mathrm {S c o r e} = W _ {\\mathrm {T T C}} \\cdot \\mathrm {T T C} + W _ {\\mathrm {S V}} \\cdot \\mathrm {S V} + W _ {\\mathrm {T E}} \\cdot \\mathrm {T E} \\tag {9}", + "image_path": "7e290e70736b1d498d2cb8ea0baf5397f7866fa9f7a871553f1739af05882387.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 381, + 192, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 192, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 192, + 396 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 402, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 475 + ], + "type": "text", + "content": "Our experiments are designed to establish baselines for the proposed LaMPilot-Bench. The main objectives are to assess the performance of LLM-based agents in interpreting human instructions within driving contexts and to evaluate the capability of LLMs to generate code for motion planning using provided functional primitives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 481, + 155, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 155, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 155, + 495 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "text", + "content": "Models We conduct benchmarking using various state-of-the-art large language models, which include both open-source and proprietary solutions1:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 536, + 249, + 597 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 47, + 536, + 216, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 216, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 216, + 548 + ], + "type": "text", + "content": "- Llama 2 [45] (llama-2-70b-chat)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 176, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 176, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 176, + 559 + ], + "type": "text", + "content": "PaLM 2 [1](code-bison)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 560, + 204, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 560, + 204, + 572 + ], + "spans": [ + { + "bbox": [ + 48, + 560, + 204, + 572 + ], + "type": "text", + "content": "- ChatGPT [34] (gpt-3.5-turbo)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 573, + 143, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 573, + 143, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 573, + 143, + 584 + ], + "type": "text", + "content": "GPT-4 [33] (gpt-4)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 584, + 249, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 249, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 249, + 597 + ], + "type": "text", + "content": "GPT-4-Turbo [32] (gpt-4-1106-preview)" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 696 + ], + "type": "text", + "content": "Code Execution To execute a LMP, the Python exec function is used. It takes the LMP code as an input string, along with two dictionaries that define the execution scope: (i) apis, which includes all APIs the code may invoke, and (ii) policies, an initially empty dictionary that will eventually contain a policy variable. If the LMP is designed to return a generator, this generator is extracted from" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "the policies dictionary after the execution of the exec function." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 115, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 115, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 305, + 115, + 545, + 175 + ], + "type": "text", + "content": "Evaluation We set a 60-second time limit for each scenario. If a task is not completed within this time frame, the test case is considered a failure, and the simulation is terminated, resulting in a driving score of 0. For successful cases, the driving score is calculated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 186, + 545, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 545, + 219 + ], + "type": "interline_equation", + "content": "\\text {D r i v i n g S c o r e} = \\frac {\\alpha}{N _ {s}} \\sum_ {i = 1} ^ {N _ {s}} \\operatorname {S c o r e} _ {i} - \\beta \\cdot P _ {\\text {c o l l i s i o n}} \\tag {10}", + "image_path": "175f226ee594303c4ac4de39a168f755b1d10ab3b76460e2444977c447460963.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": " represents the number of successful test cases, " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": " is the success rate (ranging from 0 to 1), " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": " is the collision rate (also between 0 and 1), and " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{collision}}" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": " is the penalty factor for collisions (set at 500 in our experiments). The Scorei for each individual sample is calculated based on Eq. (9), with weights " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{TTC}}" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{SV}}" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{TE}}" + }, + { + "bbox": [ + 305, + 228, + 545, + 325 + ], + "type": "text", + "content": " set to 0.5, 0.3, and 0.2, respectively. The driving score provides a comprehensive assessment of the agent's performance in LaMPilot-Bench." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 335, + 372, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 372, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 372, + 346 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 354, + 420, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 420, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 420, + 365 + ], + "type": "text", + "content": "5.2.1 Heuristic Baselines" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 374, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 545, + 601 + ], + "type": "text", + "content": "In autonomous driving, rule-based models are favored for their deterministic and interpretable nature. In our experiments, we employ two rule-based baseline policies: the Intelligent Driver Model (IDM)[46] and the Minimizing Overall Braking Induced by Lane Changes (MOBIL) principle[20]. IDM describes a rule for updating the acceleration of a vehicle to avoid collisions based on the proximity and relative velocity of the vehicle to the object directly in front. MOBIL is an extension of IDM where a lane change is executed if the prospective new lane offers a more favorable driving scenario and the maneuver can be conducted safely. These baselines can be considered as operating on random chance, as the policy is independent of user instructions but follows predefined rules. Additionally, we include a human performance baseline, where a licensed human driver controls the vehicle in the simulation using arrow keys on the keyboard, following the commands and visuals displayed. This baseline provides a reference for human-level performance on LaMPilot-Bench." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 621, + 485, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 485, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 485, + 632 + ], + "type": "text", + "content": "5.2.2 Zero-Shot and Few-Shot Baselines" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "We also include the zero-shot and few-shot chain-of-thought (CoT) prompting [54] as baseline methods. The zero-shot CoT baseline leverages the prompt \"How to complete the task step by step.\" In the few-shot setting, we follow the standard practice [2] by including 3 human-written code exemplars before the test instance. These in-context" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 239, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 239, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 239, + 712 + ], + "type": "text", + "content": "1OpenAI API accessed in October and November 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "15146" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 69, + 531, + 289 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 531, + 289 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 531, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 531, + 289 + ], + "type": "table", + "html": "
MethodLearningCollision (↓)Completion (↑)TTC Score (↑)SV Score (↑)TE Score (↑)Driving Score (↑)
Heuristic Baselines
IDMN/A0.0%20.4%92.880.971.017.3
MOBIL0.0%15.3%82.885.946.711.7
Human Driver0.0%98.0%93.486.381.384.6
LLM-based Agents
Llama 20-Shot0.0%20.4%92.881.468.917.3
PaLM 23.1%35.7%78.676.578.812.8
ChatGPT1.0%40.8%83.975.574.027.8
GPT-44.1%72.4%61.370.574.028.5
GPT-4 Turbo3.1%74.5%73.270.173.939.1
Llama 23-Shot1.0% (+1.0)63.3% (+42.9)68.3 (-24.5)71.4 (-10.0)73.8 (+4.9)39.6 (+22.3)
PaLM 23.1% (0.0)71.4% (+35.7)73.4 (-5.2)69.7 (-6.8)72.0 (-6.8)36.6 (+23.8)
ChatGPT4.1% (+3.1)83.7% (+42.9)70.9 (-13.0)73.7 (-1.8)77.7 (+3.7)41.2 (+13.4)
GPT-41.0% (-3.1)84.7% (+12.3)69.4 (+8.1)72.0 (+1.5)76.7 (+2.7)55.9 (+27.4)
GPT-4 Turbo1.0% (-2.1)80.6% (+6.1)69.1 (-4.1)71.5 (+1.4)74.9 (+1.0)52.6 (+13.5)
Llama 2Human Feedback0.9% (-0.1)32.7% (-30.6)83.9 (+15.6)74.6 (+3.2)74.6 (+0.8)21.4 (-18.2)
PaLM 20.0% (-3.1)45.5% (-25.9)81.1 (+7.7)74.4 (+4.7)77.6 (+5.6)35.6 (-1.0)
ChatGPT0.9% (-3.2)80.9% (-2.8)70.2 (-0.7)71.7 (-2.0)77.5 (-0.2)54.2 (+13.0)
GPT-40.9% (-0.1)92.7% (+8.0)67.6 (-1.8)72.2 (+0.2)76.8 (+0.1)64.0 (+8.1)
GPT-4 Turbo0.9% (-0.1)87.3% (+6.7)74.1 (+5.0)73.4 (+1.9)75.5 (+0.6)60.5 (+7.9)
", + "image_path": "ad65bf1729246b0833c3618009730da4b853197d61eba3427f7395802b7d3e9e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 297, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 297, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 547, + 331 + ], + "type": "text", + "content": "Table 3. Comparison of baselines on LaMPilot-Bench under different learning settings. Arrows " + }, + { + "bbox": [ + 46, + 297, + 547, + 331 + ], + "type": "inline_equation", + "content": "(\\uparrow/\\downarrow)" + }, + { + "bbox": [ + 46, + 297, + 547, + 331 + ], + "type": "text", + "content": " indicate higher or lower scores are better, respectively. Numbers in parentheses show the performance change compared to the previous setting, with (green/red) indicating performance improvements and degradation, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 352, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 289, + 437 + ], + "type": "text", + "content": "examples help the model adapt to the tasks in LaMPilot-Bench. The in-context examples are created by a programmer proficient in Python. They are provided with API descriptions and are allowed to write and test their code using the non-test scenarios. The same set of three examples are used for all test cases. The API Docs are also the same across all test cases for both zero-shot and few-shot settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 198, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 198, + 465 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 198, + 465 + ], + "type": "text", + "content": "5.2.3 Human Feedback Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": "LLMs have demonstrated proficiency in generating coherent solutions across various tasks without requiring additional fine-tuning. However, when it comes to code generation, especially for complex scenarios, they can produce suboptimal results. The autoregressive nature of these models poses a significant challenge, as tokens generated early in a sequence cannot be modified in the same iteration. This constraint limits the models' ability to refine initial responses, potentially impacting the effectiveness of the generated code [47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "To address these challenges and enhance the performance of LLMs on tasks in LaMPilot-Bench, we introduce a human-in-the-loop approach, where the LLMs serve not only as planners but also as engines for incorporating human feedback. As shown in Fig. 3, human feedback is infused into the learning process, transforming the LaMPilot framework from an open-loop system into an evolving feedback loop. After an LLM generates a policy program " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{P})" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", it can integrate human feedback " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{F})" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " with context-specific guidance, enabling the LLM to refine its output." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 352, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 520 + ], + "type": "text", + "content": "This approach does not involve fine-tuning or training a new LLM. Instead, we leverage the RAG approach [22] and introduce a code repository module, which is implemented with the Chroma [5] vector database. This repository allows for the storage and retrieval of effective code snippets to be used in similar situations. Following Voyager [47], we use LLM-generated function descriptions, which are then converted into vectors to serve as database keys. These keys are paired with the corresponding function codes as their values. If the human is satisfied with the policy, the code is committed to the repository for future reference. The human feedback learning process is conducted on the non-test set, involving interaction with 100 random samples for each method to ensure fair comparisons." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 525, + 425, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 425, + 538 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 425, + 538 + ], + "type": "text", + "content": "5.3.Quantitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "content": "In this section, we present the experimental results on the LaMPilot-Bench, summarizing the performance of various methods, including heuristic baselines, 0-shot and 3-shot baselines, and human feedback baselines. These results are shown in Tab. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Rule-Based Methods We first evaluate the performance of rule-based approaches, specifically the IDM and MOBIL algorithms. Both methods achieve a zero collision rate in LaMPilot-Bench. However, it is important to note that these methods operate independently of the provided instructions. Without considering human instructions, IDM and MOBIL achieve success rates of " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "20.4\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", respectively. These results provide a reference point for as-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "15147" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "sessing the effectiveness of LLM-based agents in following human instructions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 111, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 111, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 111, + 286, + 266 + ], + "type": "text", + "content": "LLM-Based Agents We assess the effectiveness of off-the-shelf LLMs to reason in the context of autonomous driving and following human instructions. In the 0-shot setting, GPT models and PaLM 2, given only the API Docs, driving context, and instructions, achieve an obvious performance advantage compared to rule-based methods in terms of task completion. This improvement can be attributed to the LLMs' ability to understand and generate code based on natural language instructions, leveraging their extensive pretraining on diverse data sources. However, the 0-shot setting also results in an increase in the collision rate (1%-4%), indicating that the generated code policies do not fully capture the complexities of driving tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "content": "When provided with training examples that include exemplar code (3-shot setting), all the evaluated LLMs exhibit notable improvements in completion rates. For instance, Llama 2's completion rate increases from " + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "inline_equation", + "content": "20.4\\%" + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "content": " in the 0-shot setting to " + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "inline_equation", + "content": "63.3\\%" + }, + { + "bbox": [ + 46, + 268, + 286, + 375 + ], + "type": "text", + "content": " in the 3-shot setting. This improvement demonstrates the effectiveness of few-shot learning in adapting LLMs to specific task domains, as the exemplar code provides a clearer understanding of the desired behavior and API usage patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 495 + ], + "type": "text", + "content": "The integration of GPT models with human-in-the-loop learning further enhances their performance in driving tasks. Notably, GPT-4 achieves the highest driving score of 64.0. This improvement highlights the great potential of LLMs in following instructions in the driving context when combined with human feedback. The human feedback loop allows for iterative refinement of the generated policies, enabling the LLMs to learn from their mistakes and continuously incorporate human knowledge to generate more effective driving strategies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 503, + 119, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 119, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 119, + 514 + ], + "type": "text", + "content": "5.4. Discussion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 286, + 713 + ], + "type": "text", + "content": "Safety Safety is a critical factor for any AD system [52]. In our context, ensuring that the system does not execute an unsafe action, even if instructed to do so by the human, is essential for responsible autonomy. Our framework includes safety checks at two levels: (1) When the LLM generates code based on human instructions, it first reasons about the safety of the action in the context of the current driving environment. By leveraging the vast knowledge and reasoning capabilities of LLMs, the system can identify potentially unsafe instructions and generate alternative actions that prioritize safety. (2) Even if the LLM fails to catch an unsafe instruction, we have a second layer of safety checks built into the APIs that interface with the vehicle actuators. These APIs encapsulate classical planning and control algorithms with hard safety constraints. This level of safety check serves as a fail-safe mechanism, ensuring that the sys" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "tem does not execute actions that violate safety constraints, regardless of the LLM's output." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 116, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 116, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 116, + 545, + 259 + ], + "type": "text", + "content": "Instruction Tuning Instruction tuning has been validated as a technique for LLMs to acquire domain-specific knowledge efficiently [3]. Fine-tuning LLMs on driving-related tasks could enhance their performance and reduce collision rates by providing them with more targeted knowledge specific to driving. However, this may also jeopardize their generalization capability when there exist some spurious correlations in the finetuning datasets (e.g. between driving context and policy) [62]. Therefore, we hope that LaMPilot-Bench can provide a platform to encourage future researchers to explore generalizable solutions to enhance LLMs' abilities in driving tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 279, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 375 + ], + "type": "text", + "content": "Multimodal LLMs Exploring the integration of vision modality in multimodal LLMs for AD tasks is another important future direction. Our current framework relies on textual input and output, which limits its ability to perceive and understand the visual world. By incorporating visual perception and understanding capabilities, multimodal LLMs could process and reason about real-world driving scenes more effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 389, + 378, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 389, + 378, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 389, + 378, + 402 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": "In this paper, we introduced LaMPilot, a novel framework that integrates Large Language Models (LLMs) into autonomous driving (AD) systems, enabling human-like interaction in diverse driving contexts. We also created LaMPilot-Bench, a benchmark for evaluating the instruction-following capabilities of LLM-based agents in AD. Our experiments demonstrated that off-the-shelf LLMs can generate code policies for driving tasks based on human instructions. However, the notable collision rate indicates the need for further research to fully capture the complexities and safety requirements of real-world driving scenarios. LaMPilot provides a starting point for leveraging LLMs in solving AD-related tasks. As LLMs continue to evolve and integrate with more modalities, we envision a future where they can significantly contribute to the development of safe, efficient, and user-friendly autonomous vehicles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 399, + 628 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 712 + ], + "type": "text", + "content": "This work is partially funded by the Digital Twin Roadmap of InfoTech Labs, Toyota Motor North America. The contents of this paper only reflect the views of the authors, who are responsible for the facts and the accuracy of the data presented herein. The contents do not necessarily reflect the official views of Toyota Motor North America." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "15148" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 52, + 91, + 289, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 91, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 289, + 475 + ], + "type": "text", + "content": "[1] Rohan Anil, Andrew M. Dai, Orhan First, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, Yi Tay, Stefan Xiao, Yuanzhong Xu, Yujing Zhang, Gustavo Hernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha, James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng, Colin Cherry, Christopher A. Choquette-Choo, Aakanksha Chowdhery, Clément Crepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Diaz, Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, Markus Freitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari, Steven Hand, Hadi Hashemi, Le Hou, Joshua Howland, Andrea Hu, Jeffrey Hui, Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, Wenhao Jia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, Katherine Lee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, Hyeontaek Lim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, Aroma Mahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, John Nham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek, Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, Parker Riley, Alex Castro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, Renee Shelby, Ambrose Slone, Daniel Smilkov, David R. So, Daniel Sohn, Simon Tokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang, Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, Yunhan Xu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng, Ce Zheng, Weikang Zhou, Denny Zhou, Slav Petrov and Yonghui Wu. PaLM 2 Technical Report. arXiv, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 476, + 289, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 476, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 52, + 476, + 289, + 586 + ], + "type": "text", + "content": "[2] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language Models are Few-Shot Learners. In NeurIPS, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 588, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 588, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 52, + 588, + 288, + 643 + ], + "type": "text", + "content": "[3] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M. Rehg, and Chao Zheng. MAPLM: A Real-World Large-Scale Vision-Language Dataset for Map and Traffic Scene Understanding. In CVPR, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 644, + 288, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 644, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 52, + 644, + 288, + 700 + ], + "type": "text", + "content": "[4] Long Chen, Oleg Sinavski, Jan Hunermann, Alice Karnsund, Andrew James Willmott, Danny Birch, Daniel Maund, and Jamie Shotton. Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving. In ICRA, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 52, + 702, + 288, + 714 + ], + "type": "text", + "content": "[5] Chroma. The AI-native open-source embedding database," + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 94 + ], + "type": "text", + "content": "2023. URL https://github.com/chroma-core/chroma.7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 311, + 96, + 547, + 140 + ], + "type": "text", + "content": "[6] Can Cui, Zichong Yang, Yupeng Zhou, Yunsheng Ma, Juanwu Lu, Lingxi Li, Yaobin Chen, Jitesh Panchal, and Ziran Wang. Large Language Models for Autonomous Driving: Real-World Experiments. arXiv, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 141, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 141, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 311, + 141, + 547, + 184 + ], + "type": "text", + "content": "[7] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as You Speak: Enabling Human-Like Interaction with Large Language Models in Autonomous Vehicles. In WACVW, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 311, + 186, + 547, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 186, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 311, + 186, + 547, + 230 + ], + "type": "text", + "content": "[8] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Receive, Reason, and React: Drive as You Say with Large Language Models in Autonomous Vehicles. IEEE Intelligent Transportation Systems Magazine, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 232, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 232, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 311, + 232, + 547, + 307 + ], + "type": "text", + "content": "[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, Tianren Gao, Erlong Li, Kun Tang, Zhipeng Cao, Tong Zhou, Ao Liu, Xinrui Yan, Shuqi Mei, Jianguo Cao, Ziran Wang, and Chao Zheng. A Survey on Multimodal Large Language Models for Autonomous Driving. In WACVW, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[10] Xinpeng Ding, Jianhua Han, Hang Xu, Wei Zhang, and Xiaomeng Li. HiLM-D: Towards High-Resolution Understanding in Multimodal Large Language Models for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[11] Daocheng Fu, Wenjie Lei, Licheng Wen, Pinlong Cai, Song Mao, Min Dou, Botian Shi, and Yu Qiao. LimSim++: A Closed-Loop Platform for Deploying Multimodal LLMs in Autonomous Driving. arXiv, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 442 + ], + "type": "text", + "content": "[12] Daocheng Fu, Xin Li, Licheng Wen, Min Dou, Pinlong Cai, Botian Shi, and Yu Qiao. Drive Like a Human: Rethinking Autonomous Driving with Large Language Models. In WACvw, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 444, + 545, + 476 + ], + "type": "text", + "content": "[13] Haoxiang Gao, Yaqian Li, Kaiwen Long, Ming Yang, and Yiqing Shen. A Survey for Foundation Models in Autonomous Driving. arXiv, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 510 + ], + "type": "text", + "content": "[14] Nicholas J. Garber and Ravi Gadiraju. Factors affecting speed variance and its influence on accidents. Transportation research record, 1989. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 544 + ], + "type": "text", + "content": "[15] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "text", + "content": "[16] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented Autonomous Driving. In CVPR, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[17] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. In CoRL, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[18] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In ICCV, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[19] Ye Jin, Xiaoxi Shen, Huiling Peng, Xiaoan Liu, Jingli Qin, Jiayang Li, Jintao Xie, Peizhong Gao, Guyue Zhou, and" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 318, + 742 + ], + "type": "text", + "content": "15149" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Jiangtao Gong. SurrealDriver: Designing Generative Driver Agent Simulation Framework in Urban Contexts based on Large Language Model. arXiv, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[20] Arne Kesting, Martin Treiber, and Dirk Helbing. General Lane-Changing Model MOBIL for Car-Following Models. Transportation Research Record, 2007. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[21] Edouard Leurent. An Environment for Autonomous Driving Decision-Making, 2018. URL https://github.com/Farama-Foundation/HighwayEnv.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 239 + ], + "type": "text", + "content": "[22] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Kuttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In NeurIPS, 2020. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[23] Boyi Li, Yue Wang, Jiageng Mao, Boris Ivanovic, Sushant Veer, Karen Leung, and Marco Pavone. Driving Everywhere with Large Language Model Policy Adaptation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 342 + ], + "type": "text", + "content": "[24] Xin Li, Yeqi Bai, Pinlong Cai, Licheng Wen, Daocheng Fu, Bo Zhang, Xuemeng Yang, Xinyu Cai, Tao Ma, Jianfei Guo, Xing Gao, Min Dou, Botian Shi, Yong Liu, Liang He, and Yu Qiao. Towards Knowledge-driven Autonomous Driving. arXiv, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 387 + ], + "type": "text", + "content": "[25] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control. In ICRA, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[26] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-Play Compositional Reasoning with Large Language Models. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "text", + "content": "[27] Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. GPT-Driver: Learning to Drive with GPT. NeurIPS Workshop on Foundation Models for Decision Making, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 510 + ], + "type": "text", + "content": "[28] Jiageng Mao, Junjie Ye, Yuxi Qian, Marco Pavone, and Yue Wang. A Language Agent for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 567 + ], + "type": "text", + "content": "[29] Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, and Oleg Sinavski. LingoQA: Video Question Answering for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 633 + ], + "type": "text", + "content": "[30] McKinsey Center for Future Mobility. Autonomous driving's future: convenient and connected, 2023. URL https://www.mckinsey.com/industries/automotive-and-assembly/our-insights/autonomous-drivings-future-convenient-and-connected.1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[31] Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2Drive: Towards Interpretable and Chain-based Reasoning for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[32] OpenAI. GPT-4-Turbo with 128K context and lower prices, the new Assistants API, GPT-4 Turbo with Vision, DALL-E 3 API, and more, 2023. URL https://openai.com/" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "blog/new-models-and-developer-products-announced-at-devday.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 509, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 509, + 107 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 509, + 107 + ], + "type": "text", + "content": "[33] OpenAI. GPT-4 Technical Report. arXiv, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 109, + 545, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 109, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 307, + 109, + 545, + 130 + ], + "type": "text", + "content": "[34] OpenAI. Introducing ChatGPT, 2023. URL https://openai.com/blog/chatgpt.6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 175 + ], + "type": "text", + "content": "[35] Chenbin Pan, Burhaneddin Yaman, Tommaso Nesti, Abhirup Mallik, Alessandro G. Allievi, Senem Velipasalar, and Liu Ren. VLP: Vision Language Planning for Autonomous Driving. In CVPR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 230 + ], + "type": "text", + "content": "[36] Hao Sha, Yao Mu, Yuxuan Jiang, Li Chen, Chenfeng Xu, Ping Luo, Shengbo Eben Li, Masayoshi Tomizuka, Wei Zhan, and Mingyu Ding. LanguageMPC: Large Language Models as Decision Makers for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 265 + ], + "type": "text", + "content": "[37] Shai Shalev-Shwartz, Shaked Shammah, and Amnon Shashua. On a Formal Model of Safe and Scalable Self-driving Cars. arXiv, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 266, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 545, + 309 + ], + "type": "text", + "content": "[38] Hao Shao, Letian Wang, Ruobing Chen, Steven L. Waslander, Hongsheng Li, and Yu Liu. ReasonNet: End-to-End Driving with Temporal and Global Reasoning. In CVPR. arXiv, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 312, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 354 + ], + "type": "text", + "content": "[39] Hao Shao, Yuxuan Hu, Letian Wang, Steven L. Waslander, Yu Liu, and Hongsheng Li. LMDrive: Closed-Loop End-to-End Driving with Large Language Models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 545, + 400 + ], + "type": "text", + "content": "[40] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Ping Luo, Andreas Geiger, and Hongyang Li. DriveLM: Driving with Graph Visual Question Answering. arXiv, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 402, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 402, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 307, + 402, + 545, + 445 + ], + "type": "text", + "content": "[41] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. In ICRA, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 447, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 447, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 447, + 545, + 544 + ], + "type": "text", + "content": "[42] Jiankai Sun, Chuanyang Zheng, Enze Xie, Zhengying Liu, Ruihang Chu, Jianing Qiu, Jiaqi Xu, Mingyu Ding, Hongyang Li, Mengzhe Geng, Yue Wu, Wenhai Wang, Junsong Chen, Zhangyue Yin, Xiaozhe Ren, Jie Fu, Junxian He, Wu Yuan, Qi Liu, Xihui Liu, Yu Li, Hao Dong, Yu Cheng, Ming Zhang, Pheng Ann Heng, Jifeng Dai, Ping Luo, Jingdong Wang, Ji-Rong Wen, Xipeng Qiu, Yike Guo, Hui Xiong, Qun Liu, and Zhenguo Li. A Survey of Reasoning with Foundation Models. arXiv, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 578 + ], + "type": "text", + "content": "[43] Dídac Surís, Sachit Menon, and Carl Vondrick. ViperGPT: Visual Inference via Python Execution for Reasoning. In ICCV, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 581, + 545, + 624 + ], + "type": "text", + "content": "[44] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models. arXiv, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 625, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 713 + ], + "type": "text", + "content": "[45] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 318, + 742 + ], + "type": "text", + "content": "15150" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 215 + ], + "type": "text", + "content": "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open Foundation and FineTuned Chat Models. arXiv, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 218, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 218, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 287, + 250 + ], + "type": "text", + "content": "[46] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 2000. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 251, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 296 + ], + "type": "text", + "content": "[47] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An Open-Ended Embodied Agent with Large Language Models. arXiv, 2023. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[48] Qifan Wang, Jingang Wang, Xiaojun Quan, Fuli Feng, Zenglin Xu, Shaoliang Nie, Sinong Wang, Madian Khabsa, Hamed Firooz, and Dongfang Liu. MUSTIE: Multimodal Structural Transformer for Web Information Extraction. In ACL, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 419 + ], + "type": "text", + "content": "[49] Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, Hao Tian, Lewei Lu, Xizhou Zhu, Xiaogang Wang, Yu Qiao, and Jifeng Dai. DriveMLM: Aligning Multi-Modal Large Language Models with Behavioral Planning States for Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 453 + ], + "type": "text", + "content": "[50] Xingyao Wang, Sha Li, and Heng Ji. Code4Struct: Code Generation for Few-Shot Event Structure Prediction. In ACL, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[51] Xingyao Wang, Hao Peng, Reyhaneh Jabbarvand, and Heng Ji. LeTI: Learning to Generate from Textual Interactions. arXiv, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[52] Yixuan Wang, Ruochen Jiao, Chengtian Lang, Sinong Simon Zhan, Chao Huang, Zhaoran Wang, Zhuoran Yang, and Qi Zhu. Empowering Autonomous Driving with Large Language Models: A Safety Perspective. In ICLR Workshop on LLM Agents, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 590 + ], + "type": "text", + "content": "[53] Wayve Technologies Ltd. LINGO-1: Exploring Natural Language for Autonomous Driving, 2023. URL https://wayve.ai/thinking/lingo-natural-language-autonomous-driving/.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[54] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 712 + ], + "type": "text", + "content": "[55] Licheng Wen, Xuemeng Yang, Daocheng Fu, Xiaofeng Wang, Pinlong Cai, Xin Li, Tao Ma, Yingxuan Li, Linran Xu, Dengke Shang, Zheng Zhu, Shaoyan Sun, Yeqi Bai, Xinyu Cai, Min Dou, Shuanglu Hu, Botian Shi, and Yu Qiao. On the Road with GPT-4V(ision): Early Explorations of Visual-Language Model on Autonomous Driving. arXiv, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 575 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[56] Licheng Wen, Daocheng Fu, Xin Li, Xinyu Cai, Tao Ma, Pinlong Cai, Min Dou, Botian Shi, Liang He, and Yu Qiao. DiLu: A Knowledge-Driven Approach to Autonomous Driving with Large Language Models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 162 + ], + "type": "text", + "content": "[57] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K. Wong, Zhengu Li, and Hengshuang Zhao. DriveGPT4: Interpretable End-to-end Autonomous Driving via Large Language Model. arXiv, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 228 + ], + "type": "text", + "content": "[58] Ke Yang, Jiateng Liu, John Wu, Chaoqi Yang, Yi R. Fung, Sha Li, Zixuan Huang, Xu Cao, Xingyao Wang, Yiquan Wang, Heng Ji, and Chengxiang Zhai. If LLM Is the Wizard, Then Code Is the Wand: A Survey on How Code Empowers Large Language Models to Serve as Intelligent Agents. ICLR Workshop on LLM Agents, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "text", + "content": "[59] Yi Yang, Qingwen Zhang, Ci Li, Daniel Simões Marta, Nazre Batool, and John Folkesson. Human-Centric Autonomous Systems With LLMs for User Command Reasoning. In WACVW, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 307 + ], + "type": "text", + "content": "[60] Zhenjie Yang, Xiaosong Jia, Hongyang Li, and Junchi Yan. LLM4Drive: A Survey of Large Language Models for Autonomous Driving. arXiv, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 350 + ], + "type": "text", + "content": "[61] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. In ICLR, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "text", + "content": "[62] Wenqian Ye, Guangtao Zheng, Xu Cao, Yunsheng Ma, Xia Hu, and Aidong Zhang. Spurious Correlations in Machine Learning: A Survey. arXiv, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 441 + ], + "type": "text", + "content": "[63] Jianhao Yuan, Shuyang Sun, Daniel Omeiza, Bo Zhao, Paul Newman, Lars Kunze, and Matthew Gadd. RAGDriver: Generalisable Driving Explanations with Retrieval-Augmented In-Context Learning in Multi-Modal Large Language Model. arXiv, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 443, + 545, + 485 + ], + "type": "text", + "content": "[64] Daoguang Zan, Bei Chen, Fengji Zhang, Dianjie Lu, Bingchao Wu, Bei Guan, Yongji Wang, and Jian-Guang Lou. Large Language Models Meet NL2Code: A Survey. In ACL, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 487, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 530 + ], + "type": "text", + "content": "[65] Xingcheng Zhou, Mingyu Liu, Bare Luka Zagar, Ekim Yurtsever, and Alois C. Knoll. Vision Language Models in Autonomous Driving and Intelligent Transportation Systems. arXiv, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 533, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 575 + ], + "type": "text", + "content": "[66] Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied Understanding of Driving Scenarios. arXiv, 2024. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "15151" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_content_list.json b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a76f58292cfe07fe15290c0875ff666529189f7b --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_content_list.json @@ -0,0 +1,1725 @@ +[ + { + "type": "text", + "text": "LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection", + "text_level": 1, + "bbox": [ + 220, + 128, + 750, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunpeng Luo Junlong Du Ke Yan* Shouhong Ding \nTencent YouTu Lab", + "bbox": [ + 256, + 203, + 712, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{petterluo, jeffdu, kerwinyan, ericshding} @tencent.com", + "bbox": [ + 315, + 241, + 650, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The evolution of Diffusion Models has dramatically improved image generation quality, making it increasingly difficult to differentiate between real and generated images. This development, while impressive, also raises significant privacy and security concerns. In response to this, we propose a novel Latent REconstruction error guided feature REfinement method $(\\mathsf{LaRE}^2)$ for detecting the diffusion-generated images. We come up with the Latent Reconstruction Error (LaRE), the first reconstruction-error based feature in the latent space for generated image detection. LaRE surpasses existing methods in terms of feature extraction efficiency while preserving crucial cues required to differentiate between the real and the fake. To exploit LaRE, we propose an Error-Guided feature REfinement module (EGRE), which can refine the image feature guided by LaRE to enhance the discriminativeness of the feature. Our EGRE utilizes an align-then-refine mechanism, which effectively refines the image feature for generated-image detection from both spatial and channel perspectives. Extensive experiments on the large-scale GenImage benchmark demonstrate the superiority of our $\\mathsf{LaRE}^2$ , which surpasses the best SoTA method by up to $11.9\\% / 12.1\\%$ average ACC/AP across 8 different image generators. LaRE also surpasses existing methods in terms of feature extraction cost, delivering an impressive speed enhancement of 8 times.", + "bbox": [ + 75, + 323, + 473, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 729, + 209, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of Diffusion Models has heralded a new era in the domain of image generation. Through concerted efforts in refining model architecture [9, 32, 33], optimizing training strategies [15, 27], and enhancing sampling methods [20, 34], contemporary Diffusion Models are now capable of generating images of unprecedented quality, surpassing the boundaries of human imagination. However, this progress raises significant concerns regarding privacy", + "bbox": [ + 75, + 755, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3575822d4c9245aa536155349939aa459d9f1aa2db807d6ab3adc8b2834c2e72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 291, + 720, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fb5f683c615aea3d9fd4c20c86a79da90c331fe9b02b2ad7b325aeba85c17a0.jpg", + "image_caption": [ + "Figure 1. (a) The comparison of reconstruction-based feature extraction. Existing method [40] chooses to completely reconstruct an image by first gradually adding noise to the image and then denoising it, which involves dozens of sampling steps. Our method can directly calculate noisy images and denoise them with a single sample step. (b) Statistical analysis of the relationship between the single-step reconstruction loss (1000 images are used) and time step. The obvious gap between the two lines indicates that single-step reconstruction can also reflect the differences between real and generated images. (c) Comparison of the cost of per image feature extraction. Our method is $8\\mathbf{x}$ faster than DIRE [40]." + ], + "image_footnote": [], + "bbox": [ + 516, + 407, + 720, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3da2ed5f7618302c7e8f23905f0f75e726882bd8b66ba33ad94465c2c68a4d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 291, + 879, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1661650c2bf7aa79a9dccac4a312655962ab80d0cc48285d3683b2a77e96d3dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 405, + 872, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and security associated with the generated images [16]. The potential for the dissemination of toxic content and misinformation through these images poses a threat to society and could mislead the public. Consequently, there is an urgent need to develop techniques to detect the images generated by these models.", + "bbox": [ + 496, + 702, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A flurry of works [6, 7, 21, 31, 40, 41] has been proposed to detect or study the properties of images generated by diffusion models. Recently, DIRE [40] has been proposed to leverage the reconstruction error as a discriminative feature for diffusion-generated image detection. It is based on the assumption that the diffusion-generated images are more easily to be reconstructed by a diffusion", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17006", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model compared with real images. As a result, DIRE shows great cross-model generalizability towards different diffusion models. Based on DIRE, SeDID [21] is also proposed to harness the inherent distributional disparities between naturally occurring and diffusion-synthesized visuals for diffusion-generated image detection.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, both DIRE and SeDID require multi-step DDIM [34] sampling processes in the feature extraction stage, which results in low efficiency for real-world applications. As shown in Fig 1(c), it takes more than 2 seconds to extract the DIRE feature for an image. Besides, the accumulation loss in the multi-step sampling also introduces uncertainty into the extracted features. Though DDIM [34] provides a deterministic inversion method to transform an image into noise, the reliability of this inversion process is not consistently upheld. In addition, DIRE and SeDID use the reconstruction error as the only feature, ignoring the correspondence between the error and the raw image. We therefore ask: (1) Do we need to completely reconstruct the image to get the discriminative feature? (2) Are there better ways to incorporate reconstruction error into image generation detection?", + "bbox": [ + 75, + 184, + 470, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the two questions mentioned above, we conducted a series of exploratory experiments and analyses. For the first question, we investigated the training process of the diffusion model. We found that given any time step $t$ , the forward Markov process has a closed-form solution, allowing us to directly transform $x_0$ into $x_t$ . Additionally, during the training phase, the model takes $x_t$ and $t$ as inputs and computes the denoising loss, enabling us to obtain the loss through single-step denoising. Therefore, we perform single-step denoising on both 1000 real and 1000 generated images respectively by LDM [32]. The results, as shown in Fig. 1(b), indicate that even with only single-step reconstruction, the loss of real images is consistently greater than that of generated images. This not only demonstrates that the real images are harder to reconstruct than the generated ones but also suggests that the loss from single-step reconstruction can reflect the differences between real and generated images. To address the second question, we visualize the reconstruction loss on the original image, as shown in Fig. 2. Upon observation, we have found that the reconstruction loss is positively correlated with the local information frequency of the original image. In the first image, for instance, the reconstruction loss is relatively lower on the low-frequency background, while it is relatively higher on the high-frequency foreground. The same pattern can be observed in other images as well. The reconstruction loss exhibits spatial correlations with the original image, thereby presenting a potential to serve as a valuable cue for the generated image detection.", + "bbox": [ + 75, + 429, + 468, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the findings above, we propose a novel Latent REconstruction error guided feature REfinement method", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/233c52e08f9df682e8d4a4044191cf611574ac3cde37987589a774542ce4a699.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 598, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eb4aa63b0bc20a78427ae7ce47b86bb592eb3f9de635b0aae5175fd77bb90e95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 154, + 596, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6af487aa9aaa230bcadd9aadf0adc762447c754eca4b8b20d71e4c5e1d83afc7.jpg", + "image_caption": [ + "Raw Image" + ], + "image_footnote": [], + "bbox": [ + 516, + 218, + 596, + 280 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/235058afdf4a1f7864948925f4f15997e2e3fa62c728639724acba63ef45709f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 89, + 691, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/132f43cbbfbea2acf82a4fb83ac8b179b1a4eee885f875f56bc52b8bbd496db2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 609, + 154, + 691, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55f643c0432b10b831758b6665e7a67ac421684d863c702590a3008f1299a844.jpg", + "image_caption": [ + "Added Noise" + ], + "image_footnote": [], + "bbox": [ + 609, + 218, + 691, + 280 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fbc8c680b87283785377240463d8f4a4e2659f7281bbab02c2626da7d883c6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 89, + 784, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b89abf9f7d008f05c0757c6973a537394bc0d323652210b9ebada602887fc9a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 154, + 784, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe99fd27eb74ad90dda188033f002330f04f77bb571b96dc0906f16c1a9cdb4e.jpg", + "image_caption": [ + "Reconstruction Loss" + ], + "image_footnote": [], + "bbox": [ + 702, + 218, + 784, + 280 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8940fa9db14b53fcce3e2f74e58a794c234c50cb0612c24c460d0d51f3098b1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 89, + 877, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b052b26dc04a3531b9cff076889038e9b1d2d7a61b10b8e9ce8b4c8ea37db33f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 154, + 877, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5aa745475a498ebf86bafa85231ed703c8ca3c3186e9d5b35c9068009d818cb1.jpg", + "image_caption": [ + "Image + Loss", + "Figure 2. Visualization of reconstruction loss on raw images (i.e. Image + Loss). Though the randomly sampled noises are added to the whole image, there is a trend that the loss in high-frequency regions is typically greater than that in low-frequency regions." + ], + "image_footnote": [], + "bbox": [ + 797, + 218, + 877, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "$(\\mathrm{LaRE}^2)$ for diffusion-generated image detection. Our $\\mathrm{LaRE}^2$ consists of two parts, the Latent Reconstruction Error (LaRE) and the Error-guided Feature REfinement module (EGRE). LaRE is a more efficient reconstruction feature, which has two improvements compared with existing methods: (1) LaRE is extracted in a single step of the diffusion reverse process, which is significantly more efficient than completely reconstructing the image through dozens of denoising steps. (2) We conduct the reconstruction in the latent space, which further improves the efficiency. Besides, we find that the reconstructed loss is positively correlated with the local information frequency of the original image. Based on that, we come up with the Error-Guided Feature Refinement module (EGRE). In EGRE, we first align LaRE with the image feature map for better correspondings for feature refinement. Then the aligned LaRE is used to refine the image feature in both spatial and channel perspectives. With the guidance of LaRE, EGRE refines the feature map in both spatial and channel perspectives to better reveal the discriminativeness for detecting the generated images.", + "bbox": [ + 496, + 388, + 890, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our method, we conduct extensive experiments on the GenImage [47] benchmark, which comprises 2,681,167 images, segregated into 1,331,167 real and 1,350,000 fake images. The fake images are from 8 different generators. As shown in Fig. 1(c), our LaRE is 8 times faster than existing method [40]. In addition, $\\mathrm{LaRE}^2$ achieves a significant performance gain by up to $11.9\\% / 12.1\\%$ ACC/AP compared with the best SoTA. The results demonstrated the superiority of our method, which is both effective and generalizable.", + "bbox": [ + 496, + 693, + 890, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our work are three-fold:", + "bbox": [ + 517, + 844, + 815, + 857 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Novel feature: We are the first to propose the reconstruction error in latent space for generated-image de", + "bbox": [ + 516, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17007", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tection. Compared with the existing method, we remarkably reduce the cost of feature extraction while preserving the essential information required for the detection of diffusion-generated images.", + "bbox": [ + 107, + 90, + 468, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Novel module: We qualitatively analyze the reason for the effectiveness of the reconstruction loss. Based on that, we come up with a novel module EGRE, which conducts an Error-guided feature refinement to enhance the discriminativeness of image features.", + "- Superior performance: Extensive experiments demonstrate the effectiveness of our method. We achieve $11.9\\% / 12.1\\%$ ACC/AP gain on the large-scale GenImage benchmark, significantly outperforming the SoTA methods." + ], + "bbox": [ + 94, + 161, + 470, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 335, + 225, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 359, + 236, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models [14, 33, 43] have become new state-of-the-art deep generative models. Inspired by nonequilibrium thermodynamics, Denoising Diffusion Probabilistic Models (DDPM) demonstrate promising generative quality, which inspires a flurry of studies on diffusion models [3, 5, 9, 32, 34, 35, 42, 46]. DDIM [34] generalize DDPMs [14] via a class of non-Markovian diffusion processes to accelerate sampling. ADM [9] firstly obtain better generation quality than GANs [11] on ImageNet [8]. Latent Diffusion Models [32] apply the diffusion process in the latent space to improve the efficiency of diffusion models, and enable diffusion models with text conditioning inputs through the cross-attention conditioning mechanism. There are also works that employ pre-trained diffusion models for various downstream tasks [3, 5, 35, 42, 46]. The rapid advancement of deep generative models has given rise to concerns regarding the potential for malicious utilization of the generated images [16]. Consequently, it is urgent to develop robust techniques to detect generated images.", + "bbox": [ + 76, + 383, + 470, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Deep Generated Image Detection", + "text_level": 1, + "bbox": [ + 76, + 681, + 366, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep-generated image detection has achieved significant improvements due to the contributions of previous research. Initially, researchers attempted to use hand-crafted features including color cues [23], saturation cues [24], blending [17] artifacts, co-occurrence features [25]. Then CNN is leveraged to detect the generated images [18, 22, 38]. Some works [10, 44] also find that there are obvious visual artifacts in the GAN-generated images and detect these images from the frequency view [28]. Most of the above works are developed and designed for GAN-generated images. With the advance of diffusion models, many methods are also provided for the study of diffusion-generated image detection. Corvi et. al [7, 31] find that state-of-the-art detec", + "bbox": [ + 76, + 704, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/641a0953ccb1dd5d69ac118439d4bd76084b38f76d41465ef82925fc27e33140.jpg", + "image_caption": [ + "Figure 3. Overview of our method. In the first stage, we extract LaRE in the latent space through single-step reconstruction. In the second stage, to exploit LaRE, we propose the Error-guided Feature Refinement Module, which consists of the Error-guided spatial refinement module and the Error-guided Channel Refinement module. From both spatial and channel perspectives, LaRE is used to enhance the discriminativeness of the image feature for generated image detection." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 898, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tors developed for GAN suffer from a severe performance drop when applied to Diffusion-generated images. Corvi et. al [6,7,31] also find that the spectrum artifacts can also rise up in the Diffusion-generated images. Wu et. al [41] formulate the synthetic image detection as an identification problem and achieve generalizable detection through language-guided contrastive learning. DIRE [40] and SeDIE [21] are proposed to leverage reconstruction error from the diffusion model to achieve generalizable diffusion-generated image detection. Different from these methods, we are the first to show that reconstruction error in the latent space can also benefit diffusion-generated image detection, which is more efficient and generalizable.", + "bbox": [ + 496, + 584, + 890, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 797, + 637, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models has achieved remarkable image generation performance. Typically, it involves two Markov processes. In the forward process, Gaussian noise is gradually added to the raw image $\\mathbf{x}_0$ until the image is asymptotically transformed to pure noise, which is defined as:", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17008", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = N (\\mathbf {x} _ {t}; \\sqrt {\\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}} \\mathbf {x} _ {t - 1}, (1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}} \\mathbf {I})), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 113, + 468, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{t}$ is the noisy image at step $t$ and $\\alpha_{t}$ is predefined noise schedule. According to the properties of Markov process and Gaussian distribution, we can get $\\mathbf{x}_t$ from $\\mathbf{x}_0$ directly by:", + "bbox": [ + 76, + 155, + 468, + 217 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = N \\left(\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I})\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 226, + 468, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the reverse process, the noisy image is gradually denoised to get the raw image, which is defined as:", + "bbox": [ + 76, + 253, + 468, + 285 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = N (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\boldsymbol {\\Sigma} _ {\\theta} (\\mathbf {x} _ {t}, t)), \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 295, + 468, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p_{\\theta}$ is typically parameterized by neural networks. During training, a neural network $\\epsilon_{\\theta}$ is trained to predict the added noise $\\epsilon$ , given the noisy image $x_{t}$ and corresponding time step $t$ :", + "bbox": [ + 76, + 321, + 468, + 381 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 391, + 468, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, $t$ and $\\epsilon$ are randomly sampled.", + "bbox": [ + 76, + 420, + 387, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 76, + 448, + 174, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we elaborate on each component of our $\\mathrm{LaRE}^2$ , as illustrated in Fig. 3.", + "bbox": [ + 76, + 473, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Latent Reconstruction Error", + "text_level": 1, + "bbox": [ + 76, + 512, + 334, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent methods [40] propose to use reconstruction error as the feature for generated-image detection. Through DDIM [34] inversion, an image is first inversed to noise and then re-generated based on the noise. However, this method has several limitations: (1) It is relatively slow to inverse and re-generate an image. For example, it takes more than 2 seconds to extract DIRE for an image on a Tesla V100 GPU as shown in Fig. 1(c), which limits its realistic applications; (2) The reconstruction of the image is not always reliable. Error could accumulate during the Markov forward and reverse process, making it hard to tell whether the bad reconstruction is from the difference between real and fake images, or the reconstruction itself.", + "bbox": [ + 75, + 536, + 468, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we come up with a new reconstruction feature, named Latent Reconstruction Error (LaRE). LaRE is based on the assumption that if the generated images can be completely reconstructed with relative ease [40], then they can also be more easily reconstructed at every single step of the reverse diffusion process.", + "bbox": [ + 75, + 732, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For an image $x$ , we first get its latent code $x_0$ by VAE. Then we calculate the LaRE by:", + "bbox": [ + 76, + 821, + 467, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {L a R E} = L _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\mathbb {E} _ {\\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 861, + 486, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, Our LaRE takes advantage of two special properties of Diffusion Models. (1) The forward process has a close form. According to Eq. 2, we can get $\\mathbf{x}_t$ directly from $\\mathbf{x}_0$ ; (2) Diffusion Model is trained to denoise a noisy image from any time step $t$ , as shown in Eq. 4. Therefore, given $\\mathbf{x}_t$ , $t$ , and $\\epsilon$ , we can directly get the reconstruction error by only single-step denoising, which is much more efficient than completely reconstructing an image. Compared with complete image reconstruction, our LaRE has several advantages: (1) Only single-step denoising is required to get LaRE, which significantly improves its efficiency. (2) Reconstruction is conducted in the latent space, which is also more efficient. (3) No need for complete inversion and reconstruction of the image, which eliminates the potential accumulation of errors that may occur during multiple inversion and reconstruction steps.", + "bbox": [ + 496, + 90, + 890, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With efficiency, the easier reconstruction assumption still holds for our LaRE. Recall that a diffusion-generated image is more easily reconstructed by diffusion models [40]. So the reconstruction error should be relatively small when completely reconstructing an image through the whole reverse process. That implies the reconstruction error is also relatively small during each step of the reverse process, which is also validated in Fig. 1(b). We leverage the Monte Carlo method to estimate LaRE in Eq. 6:", + "bbox": [ + 496, + 333, + 890, + 469 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L a R E} = \\frac {1}{e} \\sum_ {i} ^ {e} \\left[ \\| \\epsilon_ {i} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {i}, t\\right) \\| ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 481, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e, t$ are pre-defined hyperparameters. Specifically, we conduct denoising for $e$ times and average the denoising loss to extract LaRE.", + "bbox": [ + 496, + 534, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Error-Guided Feature Refinement", + "text_level": 1, + "bbox": [ + 500, + 589, + 799, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing methods exploit the reconstruction error as the only feature to detect the generated image, which ignores the relationships between the reconstruction error and the raw image. To better understand how reconstruction error helps generated-image detection, we visualize the extracted LaRE on the raw image, as shown in Fig. 2. We find several interesting conclusions. 1) The loss norm is proportional to the noise norm. It is easy to understand that the stronger the added noise is, the harder to reconstruct; 2) The loss norm is also proportional to the frequency of local image patches. That the low-frequency patches are easier to reconstruct than the high-frequency ones. Therefore, the reconstruction performance on the high-frequency parts of the image plays an important role in the generated image detection. To apply this valuable information, we come up with a novel Error-guided feature REfinement Module (EGRE). Based on the reconstruction error, our EGRE can better refine the image feature to reveal the discriminative parts for generated image detection.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17009", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/edb473f6131e88c39bd4fb0cd4a68a5530dc808aeb76c088802f933669124e31.jpg", + "image_caption": [ + "(a) Comparison on ACC. Left: baseline, right: ours" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 491, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c84a2b4faa7412863231eafb74113f7a2884de04631bd73d2c46db20cec09d98.jpg", + "image_caption": [ + "(b) Comparison on AP. Left: baseline, right: ours", + "Figure 4. Results of cross-validation on different training and testing subsets. For each generator, we train a model and test it on all 8 generators. For both DIRE [40] and our method, accuracy (ACC) and average precision(AP) are reported." + ], + "image_footnote": [], + "bbox": [ + 491, + 90, + 890, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Error-guided Spatial Refinement Module", + "text_level": 1, + "bbox": [ + 76, + 343, + 418, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose the Error-guided Spatial Refinement module (ESR) for feature refinement from the spatial perspective. First, we spatially align LaRE and the image feature map by a simple adaptive average pooling layer, so that they have the same spatial size. Now we have the spatially aligned input feature map $\\mathbf{x} \\in \\mathbb{R}^{HW \\times C_1}$ and LaRE $\\mathbf{e} \\in \\mathbb{R}^{HW \\times C_2}$ . $\\bar{\\mathbf{x}} \\in \\mathbb{R}^{1 \\times C_1}$ is the global feature, which is the average of $\\mathbf{x}$ . To refine the feature from the spatial perspective, we choose an Error-guided Spatial Attention module (ESA):", + "bbox": [ + 75, + 366, + 468, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}, \\mathbf {E}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {E}\\right) \\mathbf {V}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 507, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the feature representation, we adopt a multi-head mechanism from Transformer [37]. Therefore, our Multi-Head Error-guided Spatial Attention module (MHESA) is defined as:", + "bbox": [ + 75, + 545, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathrm{MHESA}(\\mathbf{Q},\\mathbf{K},\\mathbf{V},\\mathbf{E}) = \\mathrm{Concat}(\\mathrm{head}_1,\\dots,\\mathrm{head}_h)\\mathbf{W}^O$", + "bbox": [ + 84, + 609, + 460, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{head}_i = \\mathrm{ESA}(\\mathbf{Q}\\mathbf{W}_i^Q,\\mathbf{K}\\mathbf{W}_i^K,\\mathbf{V}\\mathbf{W}_i^V,\\mathbf{E}\\mathbf{W}_i^E)$ (8)", + "bbox": [ + 101, + 631, + 468, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $h$ is the number of heads, $\\mathbf{W}_i^Q\\in \\mathbb{R}^{C_1\\times d}$ , $\\mathbf{W}_i^K\\in \\mathbb{R}^{C_1\\times d}$ , $\\mathbf{W}_i^V\\in \\mathbb{R}^{C_1\\times d}$ , $\\mathbf{W}_E^Q\\in \\mathbb{R}^{C2\\times d}$ are learnable projection matrices, $d$ is the hyperparameter. The spatial refined feature $x^{\\prime}$ is calculated by:", + "bbox": [ + 75, + 662, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {s} = \\operatorname {M H E S A} (\\bar {\\mathbf {x}}, \\mathbf {x}, \\mathbf {x}, \\mathbf {e}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 732, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{x}$ is the only query attending to the feature map $x$ . In this stage, the aligned LaRE is used to re-weight the attention score of scaled dot-product attention. Since LaRE and the feature map are spatially aligned, LaRE can emphasize important information from the spatial perspective.", + "bbox": [ + 75, + 753, + 468, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Error-guided Channel Refinement Module", + "text_level": 1, + "bbox": [ + 76, + 845, + 426, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For channel refinement, we first squeeze both $\\mathbf{x}$ and $\\mathbf{e}$ to $\\bar{\\mathbf{x}}\\in \\mathbb{R}^{1\\times C_1}$ and $\\bar{\\mathbf{e}}\\in \\mathbb{R}^{1\\times C_2}$ . Then our channel refinement", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is achieved by a gate mechanism:", + "bbox": [ + 500, + 343, + 720, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {c} = \\text {s i g m o i d} (\\bar {\\mathbf {e}} \\mathbf {W}) \\odot \\bar {\\mathbf {x}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 364, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W} \\in \\mathbb{R}^{C1 \\times C2}$ are learnable parameters to align LaRE and the feature map from the channel perspective. Finally, $\\mathbf{x}_s$ , $\\mathbf{x}_c$ and the original global feature $\\mathbf{x}_g$ from the last convolution block are concatenated as the final feature $\\mathbf{x}_{\\mathrm{EGRE}}$ .", + "bbox": [ + 496, + 387, + 890, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {E G R E}} = \\operatorname {C o n c a t} \\left(\\mathbf {x} _ {s}, \\mathbf {x} _ {c}, \\mathbf {x} _ {g}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 455, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbf{x}_{\\mathrm{EGRE}}$ is followed by one FC layer. We train our model with the binary cross entropy loss.", + "bbox": [ + 496, + 478, + 890, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 518, + 632, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 544, + 785, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed method using the GenImage [47] dataset. This dataset comprises a total of 2,681,167 images, divided into 1,331,167 real images and 1,350,000 generated images. The generated images are from 8 different generative models, namely BigGAN [4], GLIDE [26], VQDM [12], Stable Diffusion V1.4&V1.5 [32], ADM [9], Midjourney [1], and Wukong [2]. All of the fake images are generated using the template prompt \"photo of CLS\", where \"CLS\" is replaced by one of the 1000 labels from ImageNet [8]. These images are split into 8 subsets where each subset contains partial real images and all the fake images generated by one of the above generators. We adhere to the official division of the dataset in our research, allocating 2,581,167 images for training and reserving the remaining 100,000 images for validation. Following DIRE [40], we employ Accuracy (ACC) and Average Precision (AP) as our evaluation metrics. More details and results are included in our supplementary materials.", + "bbox": [ + 496, + 568, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 847, + 715, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All experiments are conducted using the GenImage dataset. To obtain the LaRE, we employed Stable Diffu", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17010", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9b293d67910dd97398056ea68f3c713d42033ee886b751f2a98f2d01a36649eb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTesting SubsetAvg ACC.(%)
MidjourneySDV1.4SDV1.5ADMGLIDEWukongVQDMBigGAN
CNNSpot [39]58.270.370.257.057.167.756.756.661.7
Spec [45]56.772.472.357.965.470.361.764.365.1
F3Net [29]55.173.173.166.557.872.362.156.564.6
GramNet [19]58.172.872.758.765.371.357.861.264.7
DIRE [40]†65.073.773.761.969.174.363.456.767.2
Ours66.487.387.166.781.385.584.474.079.1
", + "bbox": [ + 89, + 88, + 880, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparisons on GenImage test set. For each data point, eight models are trained on eight generators respectively. Then eight models are tested on the specified test sets and the accuracy scores are averaged. †indicates our reproduction.", + "bbox": [ + 75, + 232, + 892, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sion V1.5 [32] with a step size of $t = 200$ and a noise ensemble size of $e = 4$ . Images are resized to $256 \\times 256$ for LaRE extraction. We use the prompt \"a photo\" for all images. For both training and testing phases, images are resized to $224 \\times 224$ . Our model utilizes the CLIP [30] pretrained ResNet50 [13] as the backbone architecture. We set the batch size to 48 and the learning rate to $1e^{-4}$ . Training is performed on a single Nvidia Tesla V100 GPU. We train eight models on eight subsets respectively, each corresponding to a different generation method. Model selection is based on performance in the validation set, which shares the same generation method as the training set.", + "bbox": [ + 75, + 291, + 472, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Cross-Generator Image Classification", + "text_level": 1, + "bbox": [ + 76, + 481, + 403, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our model trained by one of the subsets and test on all eight subsets. The results are shown in Fig. 4. Several conclusions can be made from the results. (1) It is easy to detect generated images from seen generators. Both the DIRE and our proposed method demonstrate superior performance when training and testing images share the same generator, as shown by the diagonal of each image. (2) It is challenging to generalize the model to unseen generators. However, it is relatively easier to generalize the model to similar generator structures than to different structures. For example, a model trained on the BigGAN subset achieves bad performance when tested on other subsets, which are all diffusion-based generators. While the model trained on the SDV1.5 subset achieves relatively better performance on other diffusion-based models. (3) Our model is more robust and generalizable than DIRE. Our model trained on diffusion-based generators achieves better generalizability on both gan-based and diffusion-based generators. This indicates that our $\\mathrm{LaRE}^2$ can better capture discriminative features to detect the generated images.", + "bbox": [ + 75, + 505, + 472, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Compare with the State of the Arts", + "text_level": 1, + "bbox": [ + 76, + 832, + 382, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method with several state-of-the-art generated image detection methods, including CNNSpot [39], Spec [45], F3Net [29], GramNet [19]", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and DIRE [40]. In accordance with the GenImage benchmark, eight models are trained on each subset. Subsequently, these models are evaluated on the specified test set, and their accuracy scores are averaged to produce a data point for inclusion in the table. This benchmark presents a formidable challenge, as the overall accuracy is closely tied to the cross-generator generalizability of the proposed techniques. As illustrated in Tab. 1 and Fig. 4, our approach attains state-of-the-art performance across all eight generators. Notably, our overall accuracy exhibits a substantial improvement, surpassing the previous state-of-the-art by $11.9\\% / 12.1\\%$ ACC/AP. This result demonstrates the superior generalizability of our method.", + "bbox": [ + 496, + 291, + 893, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 498, + 498, + 653, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct several ablative studies to quantify the contribution of each design in our model and the influence of the hyperparameters.", + "bbox": [ + 496, + 522, + 893, + 569 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5.1 Influence of EGRE", + "text_level": 1, + "bbox": [ + 498, + 589, + 687, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The contribution of individual components in EGRE to the overall performance of the method is reported in Tab. 2. Starting with the most simple baseline, i.e., training without LaRE (Model A in the table), each component is added building up to the proposed approach. We measure the effect of the Error-guided Spatial Refinement module (ESR) and Error-guided Channel Refinement module (ECR). Specifically, ESR achieves $8.6\\% /5.3\\%$ ACC/AP gain and ECR achieves $5.7\\% /2.4\\%$ ACC/AP gain compared with the baseline. Both ESR and ECR significantly improve the performance, demonstrating the effectiveness of the two modules. Relatively, ESR brings a higher improvement in metrics compared with ECR, indicating a greater benefit from spatial feature refinement. This is also attributed to the better alignment of LaRE with the original image at the spatial scale as shown in Fig. 2. By combining ESR and ECR, our approach achieves a further improvement of $11.4\\% /7.2\\%$ ACC/AP. By leveraging LaRE for feature refinement at both spatial and channel scales, our model", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/db3720d9df913cce0141b5a6e7ad49df0b6e9042982d6d37a8d38df6e2ee2702.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelESRECRCLSMJSDV1.4SDV1.5ADMGLIDEWuKongVQDMBigganAvg
A52.9/94.4100.0/100.0100.0/100.055.0/89.575.5/98.799.7/100.051.8/84.450.4/71.773.1/92.3
B✓61.8/96.9100.0/100.0100.0/100.060.3/95.487.2/99.1100.0/100.086.2/94.658.1/95.081.7/97.6
C✓57.4/95.3100.0/100.0100.0/100.057.2/91.681.2/98.9100.0/100.078.3/89.856.2/82.378.8/94.7
D✓✓64.8/98.2100.0/100.0100.0/100.061.9/98.591.4/99.8100.0/100.094.6/100.063.7/99.784.5/99.5
E✓✓✓66.2/99.3100.0/100.0100.0/100.064.5/99.591.5/99.9100.0/100.097.7/100.067.4/99.885.9/99.8
", + "bbox": [ + 86, + 88, + 883, + 198 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablative studies on using different modules or prompts. 'CLS' indicates to use class-specific prompt for LaRE extraction.", + "bbox": [ + 98, + 210, + 869, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3fe8e64e4dab045164e5fac76e51ffd13806feae5e561e17a3e1a5ec26fe935.jpg", + "image_caption": [ + "Figure 5. Trade-off between detection performance and feature extraction cost. When $e = 4$ , the model achieves the best trade-off." + ], + "image_footnote": [], + "bbox": [ + 89, + 250, + 459, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves more robust and generalizable image generation detection performance.", + "bbox": [ + 75, + 521, + 468, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5.2 Influence of Noise Ensemble", + "text_level": 1, + "bbox": [ + 76, + 571, + 328, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We systematically evaluated the impact of the noise ensemble parameter $e$ on both the accuracy and feature extraction runtime. As shown in Fig. 5. The results demonstrate a consistent increase in accuracy as $e$ increased. However, this improvement in accuracy comes at the cost of increased feature extraction runtime. Upon careful consideration of the trade-off between accuracy and computational cost, we propose $e = 4$ as the optimal choice. This selection of $e$ provides an effective balance, maintaining the performance of our model while ensuring computational efficiency.", + "bbox": [ + 75, + 597, + 470, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5.3 Influence of Prompts", + "text_level": 1, + "bbox": [ + 76, + 768, + 277, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Because Latent diffusion is a text-conditioned model, a prompt is needed to denoise an image. Here we try to use a class-specific prompt instead of an unchanged prompt. We choose the template \"a photo of CLS\" where \"CLS\" is replaced with the ImageNet label of the image. Results are shown in Tab. 2. When using class-specific prompts (Model E), we can achieve further improvements by $1.4\\% /0.3\\%$", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/156bc147d5b4040d8b7ca856d9105748cbab059f271f9c23bfd70324cf885d8d.jpg", + "image_caption": [ + "Figure 6. ACC and AP when choosing different $t$ . The results demonstrate that our model is robust to the choice of $t$ ." + ], + "image_footnote": [], + "bbox": [ + 535, + 250, + 859, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ACC/AP. This suggests that the category information of images can aid in the extraction of LaRE features. Even in the absence of category information, it is interesting that the accuracy does not decline significantly. We speculate this is because we did not fully convert the images into noise. Even with the added noise, the image still retains some of the original important semantic information, which can assist the model in denoising and thus obtaining good LaRE features. This experiment validates that when the textual information of an image is known, the extracted LaRE features can further enhance the performance. However, we still use the generic prompt in our method, as it already achieves satisfactory results. As for accurately describing images to further enhance performance, we leave that for future research work.", + "bbox": [ + 496, + 518, + 893, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5.4 Influence of Sample Step", + "text_level": 1, + "bbox": [ + 500, + 768, + 725, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we conduct the ablative experiment on the sample step $t$ . As illustrated in Fig. 1(b), there is a gap between the reconstruction loss of real images and the generated images. When $t \\in [100,500]$ , there is a more obvious difference in the reconstruction loss between the two types of images. To verify the impact of $t$ selection on the performance, we conducted experiments with different values of $t$ , and the", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17012", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6d1690052959bbffafe8201d40fb1305a5a37f19347e2429c7d281efbd62d96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 92, + 460, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e11dda25136f9bea9c49038ee1aab790164f1db6a86da6fe8cc3be4a804b3275.jpg", + "image_caption": [ + "Figure 7. Visualization of the spatial representation of the baseline model (a) and our model (b). Both models were trained on the SDV1.5 subset. In the case of seen data, both models exhibit good discrimination in the feature space. However, for unseen data, the baseline model shows a significant overlap between real images and fake images (red circle), whereas our model exhibits a smaller overlap (red circle). This further indicates the better generalization capability of our model." + ], + "image_footnote": [], + "bbox": [ + 86, + 316, + 460, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results are shown in Fig. 6. When $t \\in [150, 300]$ , the model exhibits stable performance, indicating the robustness of our model with respect to $t$ . However, when $t$ exceeds this range, the performance of the model tends to decline.", + "bbox": [ + 75, + 691, + 468, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.5 Influence of Input Information", + "text_level": 1, + "bbox": [ + 76, + 771, + 344, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments to train the detection model with different inputs. The results are shown in Tab. 3. Since our LaRE is a compressed latent representation whose size is only $32 \\times 32 \\times 4$ . We train a Resnet-20 with LaRE as the only input (Model B). The experimental results show that both the AP and ACC metrics have decreased significantly. We speculate that there are two reasons. Firstly,", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99ebd29418aea7127d883188572dc1c06fcaea1d64bd7f791868d2a65625c483.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelImageLaREMethodAVG. ACCAVG. AP
A✓None73.192.3
B✓None66.268.3
C✓✓Concat76.893.5
D✓✓EGRE84.599.5
", + "bbox": [ + 513, + 89, + 880, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results of using different input. Models are trained on the SDv1.5 subset and tested on all the subsets. 'AVG' indicates average scores over 8 subsets. 'Concat' indicates concat LaRE with the image feature map after the last convolution block.", + "bbox": [ + 498, + 186, + 890, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LaRE is in the latent space and has more information loss compared to the original image. It is not sufficient to serve as the only information for detecting the generated images. Secondly, as shown in the Fig. 2, LaRE mainly reflects the reconstruction quality of the original image at different spatial positions. It is better to leverage LaRE to enhance the representation of the original image features. Therefore, we conducted further experiments. In model C, we only concatenated LaRE with the feature map of the image, which also improves the performance. This validates our hypothesis. Besides, model D achieves significant performance gain compared with Model C. By using EGRE, we can further improve the performance of our model by a significant margin, which also confirms the advantages of EGRE over simple feature concatenation.", + "bbox": [ + 496, + 275, + 890, + 502 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Qualitative Results and Visualizations", + "text_level": 1, + "bbox": [ + 498, + 515, + 826, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We utilize t-SNE [36] visualization to illustrate feature vectors derived from the final layers of our model and the baseline model, as depicted in Fig. 7. Each model is trained on the SDv1.5 subset of GenImage and evaluated across all eight generative methods. Although both models exhibit good feature discrimination on seen images, they both have limitations when it comes to unseen scenarios. For images from MJ, ADM, and BigGAN, both models show overlaps between real and fake images in the feature space. However, our model demonstrates a smaller proportion of overlap, indicating its better generalization ability.", + "bbox": [ + 496, + 539, + 890, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 723, + 617, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel reconstruction-based diffusion-generated image detection method called LaRE $^2$ . We come up with LaRE, a novel and more efficient reconstruction-based feature by reconstructing the image in the Latent space. Notably, LaRE is 8 times faster compared with existing reconstruction-based methods. By incorporating LaRE with the Error-guided Feature Refinement module (EGRE). Our $\\mathrm{LaRE}^2$ achieves superior performance on diffusion-generated image detection, demonstrating state-of-the-art performance.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Midjourney. https://www.midjourney.com/home/, 2022.5", + "[2] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. 2021. 3", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5", + "[5] Kevin Clark and Priyank Jaini. Text-to-image diffusion models are zero-shot classifiers. 2023. 3", + "[6] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In CVPR, pages 973-982, 2023. 1, 3", + "[7] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, pages 1-5. IEEE, 2023. 1, 3", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 3, 5", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. 34:8780-8794, 2021. 1, 3, 5", + "[10] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. 63(11):139–144, 2020. 3", + "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 5", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 33:6840-6851, 2020. 3", + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. 2022. 1", + "[16] Felix Juefei-Xu, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang Liu. Countering malicious deepfakes: Survey, battleground, and horizon. 130(7):1678-1734, 2022. 1, 3", + "[17] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, pages 5001-5010, 2020. 3", + "[18] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 3" + ], + "bbox": [ + 78, + 116, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 6", + "[20] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. 35:5775-5787, 2022. 1", + "[21] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. 2023. 1, 2, 3", + "[22] Francesco Marra, Diego Gragnaniello, Davide Cozzolino, and Luisa Verdoliva. Detection of gan-generated fake images over social networks. In MIPR, pages 384-389. IEEE, 2018. 3", + "[23] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using color cues. 2018. 3", + "[24] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using saturation cues. In ICIP, pages 4584-4588. IEEE, 2019. 3", + "[25] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, Amit K Roy-Chowdhury, and BS Manjunath. Detecting gan generated fake images using co-occurrence matrices. 2019. 3", + "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. 2021. 5", + "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 1", + "[28] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 3", + "[29] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 6", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6", + "[31] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. 2022. 1, 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3, 5, 6", + "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1, 3", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. 2020. 1, 2, 3, 4" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17014", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. 2023. 3", + "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. 9(11), 2008. 8", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 30, 2017. 5", + "[38] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 3", + "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 6", + "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. 2023. 1, 2, 3, 4, 5, 6", + "[41] Haiwei Wu, Jiantao Zhou, and Shile Zhang. Generalizable synthetic image detection via language-guided contrastive learning. 2023. 1, 3", + "[42] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 3", + "[43] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. 2022. 3", + "[44] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 3", + "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 6", + "[46] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. 2023. 3", + "[47] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. 2023. 2, 5" + ], + "bbox": [ + 78, + 90, + 467, + 702 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17015", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_model.json b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1eedaff639295428ac714c729328e007d2308425 --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.222, + 0.13, + 0.75, + 0.177 + ], + "angle": 0, + "content": "LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.204, + 0.713, + 0.238 + ], + "angle": 0, + "content": "Yunpeng Luo Junlong Du Ke Yan* Shouhong Ding \nTencent YouTu Lab" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.242, + 0.651, + 0.256 + ], + "angle": 0, + "content": "{petterluo, jeffdu, kerwinyan, ericshding} @tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.704 + ], + "angle": 0, + "content": "The evolution of Diffusion Models has dramatically improved image generation quality, making it increasingly difficult to differentiate between real and generated images. This development, while impressive, also raises significant privacy and security concerns. In response to this, we propose a novel Latent REconstruction error guided feature REfinement method \\((\\mathsf{LaRE}^2)\\) for detecting the diffusion-generated images. We come up with the Latent Reconstruction Error (LaRE), the first reconstruction-error based feature in the latent space for generated image detection. LaRE surpasses existing methods in terms of feature extraction efficiency while preserving crucial cues required to differentiate between the real and the fake. To exploit LaRE, we propose an Error-Guided feature REfinement module (EGRE), which can refine the image feature guided by LaRE to enhance the discriminativeness of the feature. Our EGRE utilizes an align-then-refine mechanism, which effectively refines the image feature for generated-image detection from both spatial and channel perspectives. Extensive experiments on the large-scale GenImage benchmark demonstrate the superiority of our \\(\\mathsf{LaRE}^2\\), which surpasses the best SoTA method by up to \\(11.9\\% / 12.1\\%\\) average ACC/AP across 8 different image generators. LaRE also surpasses existing methods in terms of feature extraction cost, delivering an impressive speed enhancement of 8 times." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.731, + 0.21, + 0.746 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.756, + 0.47, + 0.879 + ], + "angle": 0, + "content": "The rapid advancement of Diffusion Models has heralded a new era in the domain of image generation. Through concerted efforts in refining model architecture [9, 32, 33], optimizing training strategies [15, 27], and enhancing sampling methods [20, 34], contemporary Diffusion Models are now capable of generating images of unprecedented quality, surpassing the boundaries of human imagination. However, this progress raises significant concerns regarding privacy" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.292, + 0.721, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.408, + 0.721, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.292, + 0.88, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.406, + 0.874, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Figure 1. (a) The comparison of reconstruction-based feature extraction. Existing method [40] chooses to completely reconstruct an image by first gradually adding noise to the image and then denoising it, which involves dozens of sampling steps. Our method can directly calculate noisy images and denoise them with a single sample step. (b) Statistical analysis of the relationship between the single-step reconstruction loss (1000 images are used) and time step. The obvious gap between the two lines indicates that single-step reconstruction can also reflect the differences between real and generated images. (c) Comparison of the cost of per image feature extraction. Our method is \\(8\\mathbf{x}\\) faster than DIRE [40]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.793 + ], + "angle": 0, + "content": "and security associated with the generated images [16]. The potential for the dissemination of toxic content and misinformation through these images poses a threat to society and could mislead the public. Consequently, there is an urgent need to develop techniques to detect the images generated by these models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "A flurry of works [6, 7, 21, 31, 40, 41] has been proposed to detect or study the properties of images generated by diffusion models. Recently, DIRE [40] has been proposed to leverage the reconstruction error as a discriminative feature for diffusion-generated image detection. It is based on the assumption that the diffusion-generated images are more easily to be reconstructed by a diffusion" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.182 + ], + "angle": 0, + "content": "model compared with real images. As a result, DIRE shows great cross-model generalizability towards different diffusion models. Based on DIRE, SeDID [21] is also proposed to harness the inherent distributional disparities between naturally occurring and diffusion-synthesized visuals for diffusion-generated image detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.185, + 0.471, + 0.426 + ], + "angle": 0, + "content": "However, both DIRE and SeDID require multi-step DDIM [34] sampling processes in the feature extraction stage, which results in low efficiency for real-world applications. As shown in Fig 1(c), it takes more than 2 seconds to extract the DIRE feature for an image. Besides, the accumulation loss in the multi-step sampling also introduces uncertainty into the extracted features. Though DDIM [34] provides a deterministic inversion method to transform an image into noise, the reliability of this inversion process is not consistently upheld. In addition, DIRE and SeDID use the reconstruction error as the only feature, ignoring the correspondence between the error and the raw image. We therefore ask: (1) Do we need to completely reconstruct the image to get the discriminative feature? (2) Are there better ways to incorporate reconstruction error into image generation detection?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.867 + ], + "angle": 0, + "content": "To address the two questions mentioned above, we conducted a series of exploratory experiments and analyses. For the first question, we investigated the training process of the diffusion model. We found that given any time step \\( t \\), the forward Markov process has a closed-form solution, allowing us to directly transform \\( x_0 \\) into \\( x_t \\). Additionally, during the training phase, the model takes \\( x_t \\) and \\( t \\) as inputs and computes the denoising loss, enabling us to obtain the loss through single-step denoising. Therefore, we perform single-step denoising on both 1000 real and 1000 generated images respectively by LDM [32]. The results, as shown in Fig. 1(b), indicate that even with only single-step reconstruction, the loss of real images is consistently greater than that of generated images. This not only demonstrates that the real images are harder to reconstruct than the generated ones but also suggests that the loss from single-step reconstruction can reflect the differences between real and generated images. To address the second question, we visualize the reconstruction loss on the original image, as shown in Fig. 2. Upon observation, we have found that the reconstruction loss is positively correlated with the local information frequency of the original image. In the first image, for instance, the reconstruction loss is relatively lower on the low-frequency background, while it is relatively higher on the high-frequency foreground. The same pattern can be observed in other images as well. The reconstruction loss exhibits spatial correlations with the original image, thereby presenting a potential to serve as a valuable cue for the generated image detection." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Based on the findings above, we propose a novel Latent REconstruction error guided feature REfinement method" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.599, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.155, + 0.598, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.219, + 0.598, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.284, + 0.586, + 0.295 + ], + "angle": 0, + "content": "Raw Image" + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.09, + 0.692, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.155, + 0.692, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.61, + 0.219, + 0.692, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.284, + 0.682, + 0.294 + ], + "angle": 0, + "content": "Added Noise" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.09, + 0.785, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.155, + 0.785, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.219, + 0.785, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.284, + 0.792, + 0.294 + ], + "angle": 0, + "content": "Reconstruction Loss" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.09, + 0.879, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.155, + 0.879, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.219, + 0.879, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.284, + 0.872, + 0.295 + ], + "angle": 0, + "content": "Image + Loss" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Visualization of reconstruction loss on raw images (i.e. Image + Loss). Though the randomly sampled noises are added to the whole image, there is a trend that the loss in high-frequency regions is typically greater than that in low-frequency regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.892, + 0.692 + ], + "angle": 0, + "content": "\\((\\mathrm{LaRE}^2)\\) for diffusion-generated image detection. Our \\(\\mathrm{LaRE}^2\\) consists of two parts, the Latent Reconstruction Error (LaRE) and the Error-guided Feature REfinement module (EGRE). LaRE is a more efficient reconstruction feature, which has two improvements compared with existing methods: (1) LaRE is extracted in a single step of the diffusion reverse process, which is significantly more efficient than completely reconstructing the image through dozens of denoising steps. (2) We conduct the reconstruction in the latent space, which further improves the efficiency. Besides, we find that the reconstructed loss is positively correlated with the local information frequency of the original image. Based on that, we come up with the Error-Guided Feature Refinement module (EGRE). In EGRE, we first align LaRE with the image feature map for better correspondings for feature refinement. Then the aligned LaRE is used to refine the image feature in both spatial and channel perspectives. With the guidance of LaRE, EGRE refines the feature map in both spatial and channel perspectives to better reveal the discriminativeness for detecting the generated images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.892, + 0.844 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our method, we conduct extensive experiments on the GenImage [47] benchmark, which comprises 2,681,167 images, segregated into 1,331,167 real and 1,350,000 fake images. The fake images are from 8 different generators. As shown in Fig. 1(c), our LaRE is 8 times faster than existing method [40]. In addition, \\(\\mathrm{LaRE}^2\\) achieves a significant performance gain by up to \\(11.9\\% / 12.1\\%\\) ACC/AP compared with the best SoTA. The results demonstrated the superiority of our method, which is both effective and generalizable." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.845, + 0.816, + 0.858 + ], + "angle": 0, + "content": "The contributions of our work are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Novel feature: We are the first to propose the reconstruction error in latent space for generated-image de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17007" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "tection. Compared with the existing method, we remarkably reduce the cost of feature extraction while preserving the essential information required for the detection of diffusion-generated images." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.162, + 0.47, + 0.238 + ], + "angle": 0, + "content": "- Novel module: We qualitatively analyze the reason for the effectiveness of the reconstruction loss. Based on that, we come up with a novel module EGRE, which conducts an Error-guided feature refinement to enhance the discriminativeness of image features." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.248, + 0.471, + 0.323 + ], + "angle": 0, + "content": "- Superior performance: Extensive experiments demonstrate the effectiveness of our method. We achieve \\(11.9\\% / 12.1\\%\\) ACC/AP gain on the large-scale GenImage benchmark, significantly outperforming the SoTA methods." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.162, + 0.471, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.337, + 0.227, + 0.352 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.361, + 0.238, + 0.376 + ], + "angle": 0, + "content": "2.1. Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.385, + 0.471, + 0.673 + ], + "angle": 0, + "content": "Diffusion Models [14, 33, 43] have become new state-of-the-art deep generative models. Inspired by nonequilibrium thermodynamics, Denoising Diffusion Probabilistic Models (DDPM) demonstrate promising generative quality, which inspires a flurry of studies on diffusion models [3, 5, 9, 32, 34, 35, 42, 46]. DDIM [34] generalize DDPMs [14] via a class of non-Markovian diffusion processes to accelerate sampling. ADM [9] firstly obtain better generation quality than GANs [11] on ImageNet [8]. Latent Diffusion Models [32] apply the diffusion process in the latent space to improve the efficiency of diffusion models, and enable diffusion models with text conditioning inputs through the cross-attention conditioning mechanism. There are also works that employ pre-trained diffusion models for various downstream tasks [3, 5, 35, 42, 46]. The rapid advancement of deep generative models has given rise to concerns regarding the potential for malicious utilization of the generated images [16]. Consequently, it is urgent to develop robust techniques to detect generated images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.367, + 0.697 + ], + "angle": 0, + "content": "2.2. Deep Generated Image Detection" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Deep-generated image detection has achieved significant improvements due to the contributions of previous research. Initially, researchers attempted to use hand-crafted features including color cues [23], saturation cues [24], blending [17] artifacts, co-occurrence features [25]. Then CNN is leveraged to detect the generated images [18, 22, 38]. Some works [10, 44] also find that there are obvious visual artifacts in the GAN-generated images and detect these images from the frequency view [28]. Most of the above works are developed and designed for GAN-generated images. With the advance of diffusion models, many methods are also provided for the study of diffusion-generated image detection. Corvi et. al [7, 31] find that state-of-the-art detec" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.9, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.557 + ], + "angle": 0, + "content": "Figure 3. Overview of our method. In the first stage, we extract LaRE in the latent space through single-step reconstruction. In the second stage, to exploit LaRE, we propose the Error-guided Feature Refinement Module, which consists of the Error-guided spatial refinement module and the Error-guided Channel Refinement module. From both spatial and channel perspectives, LaRE is used to enhance the discriminativeness of the image feature for generated image detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.782 + ], + "angle": 0, + "content": "tors developed for GAN suffer from a severe performance drop when applied to Diffusion-generated images. Corvi et. al [6,7,31] also find that the spectrum artifacts can also rise up in the Diffusion-generated images. Wu et. al [41] formulate the synthetic image detection as an identification problem and achieve generalizable detection through language-guided contrastive learning. DIRE [40] and SeDIE [21] are proposed to leverage reconstruction error from the diffusion model to achieve generalizable diffusion-generated image detection. Different from these methods, we are the first to show that reconstruction error in the latent space can also benefit diffusion-generated image detection, which is more efficient and generalizable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.799, + 0.638, + 0.815 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diffusion Models has achieved remarkable image generation performance. Typically, it involves two Markov processes. In the forward process, Gaussian noise is gradually added to the raw image \\(\\mathbf{x}_0\\) until the image is asymptotically transformed to pure noise, which is defined as:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17008" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.098, + 0.114, + 0.47, + 0.148 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = N (\\mathbf {x} _ {t}; \\sqrt {\\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}} \\mathbf {x} _ {t - 1}, (1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}} \\mathbf {I})), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.156, + 0.47, + 0.218 + ], + "angle": 0, + "content": "where \\( x_{t} \\) is the noisy image at step \\( t \\) and \\( \\alpha_{t} \\) is predefined noise schedule. According to the properties of Markov process and Gaussian distribution, we can get \\( \\mathbf{x}_t \\) from \\( \\mathbf{x}_0 \\) directly by:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.227, + 0.469, + 0.245 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = N \\left(\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I})\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.254, + 0.469, + 0.286 + ], + "angle": 0, + "content": "In the reverse process, the noisy image is gradually denoised to get the raw image, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.296, + 0.47, + 0.313 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = N (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\boldsymbol {\\Sigma} _ {\\theta} (\\mathbf {x} _ {t}, t)), \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.469, + 0.382 + ], + "angle": 0, + "content": "where \\( p_{\\theta} \\) is typically parameterized by neural networks. During training, a neural network \\( \\epsilon_{\\theta} \\) is trained to predict the added noise \\( \\epsilon \\), given the noisy image \\( x_{t} \\) and corresponding time step \\( t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.392, + 0.469, + 0.41 + ], + "angle": 0, + "content": "\\[\nL _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.421, + 0.388, + 0.436 + ], + "angle": 0, + "content": "During training, \\(t\\) and \\(\\epsilon\\) are randomly sampled." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.449, + 0.176, + 0.464 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.474, + 0.469, + 0.505 + ], + "angle": 0, + "content": "In this section, we elaborate on each component of our \\(\\mathrm{LaRE}^2\\), as illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.513, + 0.336, + 0.528 + ], + "angle": 0, + "content": "4.1. Latent Reconstruction Error" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.732 + ], + "angle": 0, + "content": "Recent methods [40] propose to use reconstruction error as the feature for generated-image detection. Through DDIM [34] inversion, an image is first inversed to noise and then re-generated based on the noise. However, this method has several limitations: (1) It is relatively slow to inverse and re-generate an image. For example, it takes more than 2 seconds to extract DIRE for an image on a Tesla V100 GPU as shown in Fig. 1(c), which limits its realistic applications; (2) The reconstruction of the image is not always reliable. Error could accumulate during the Markov forward and reverse process, making it hard to tell whether the bad reconstruction is from the difference between real and fake images, or the reconstruction itself." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Here we come up with a new reconstruction feature, named Latent Reconstruction Error (LaRE). LaRE is based on the assumption that if the generated images can be completely reconstructed with relative ease [40], then they can also be more easily reconstructed at every single step of the reverse diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.823, + 0.468, + 0.854 + ], + "angle": 0, + "content": "For an image \\( x \\), we first get its latent code \\( x_0 \\) by VAE. Then we calculate the LaRE by:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.862, + 0.488, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\operatorname {L a R E} = L _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\mathbb {E} _ {\\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.333 + ], + "angle": 0, + "content": "As shown in Fig. 3, Our LaRE takes advantage of two special properties of Diffusion Models. (1) The forward process has a close form. According to Eq. 2, we can get \\(\\mathbf{x}_t\\) directly from \\(\\mathbf{x}_0\\); (2) Diffusion Model is trained to denoise a noisy image from any time step \\(t\\), as shown in Eq. 4. Therefore, given \\(\\mathbf{x}_t\\), \\(t\\), and \\(\\epsilon\\), we can directly get the reconstruction error by only single-step denoising, which is much more efficient than completely reconstructing an image. Compared with complete image reconstruction, our LaRE has several advantages: (1) Only single-step denoising is required to get LaRE, which significantly improves its efficiency. (2) Reconstruction is conducted in the latent space, which is also more efficient. (3) No need for complete inversion and reconstruction of the image, which eliminates the potential accumulation of errors that may occur during multiple inversion and reconstruction steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.47 + ], + "angle": 0, + "content": "With efficiency, the easier reconstruction assumption still holds for our LaRE. Recall that a diffusion-generated image is more easily reconstructed by diffusion models [40]. So the reconstruction error should be relatively small when completely reconstructing an image through the whole reverse process. That implies the reconstruction error is also relatively small during each step of the reverse process, which is also validated in Fig. 1(b). We leverage the Monte Carlo method to estimate LaRE in Eq. 6:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.482, + 0.891, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L a R E} = \\frac {1}{e} \\sum_ {i} ^ {e} \\left[ \\| \\epsilon_ {i} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {i}, t\\right) \\| ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.579 + ], + "angle": 0, + "content": "where \\( e, t \\) are pre-defined hyperparameters. Specifically, we conduct denoising for \\( e \\) times and average the denoising loss to extract LaRE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.59, + 0.8, + 0.605 + ], + "angle": 0, + "content": "4.2. Error-Guided Feature Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Existing methods exploit the reconstruction error as the only feature to detect the generated image, which ignores the relationships between the reconstruction error and the raw image. To better understand how reconstruction error helps generated-image detection, we visualize the extracted LaRE on the raw image, as shown in Fig. 2. We find several interesting conclusions. 1) The loss norm is proportional to the noise norm. It is easy to understand that the stronger the added noise is, the harder to reconstruct; 2) The loss norm is also proportional to the frequency of local image patches. That the low-frequency patches are easier to reconstruct than the high-frequency ones. Therefore, the reconstruction performance on the high-frequency parts of the image plays an important role in the generated image detection. To apply this valuable information, we come up with a novel Error-guided feature REfinement Module (EGRE). Based on the reconstruction error, our EGRE can better refine the image feature to reveal the discriminative parts for generated image detection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17009" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.493, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.266, + 0.432, + 0.277 + ], + "angle": 0, + "content": "(a) Comparison on ACC. Left: baseline, right: ours" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.091, + 0.892, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.266, + 0.825, + 0.277 + ], + "angle": 0, + "content": "(b) Comparison on AP. Left: baseline, right: ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.894, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Results of cross-validation on different training and testing subsets. For each generator, we train a model and test it on all 8 generators. For both DIRE [40] and our method, accuracy (ACC) and average precision(AP) are reported." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.344, + 0.419, + 0.359 + ], + "angle": 0, + "content": "4.2.1 Error-guided Spatial Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.503 + ], + "angle": 0, + "content": "We propose the Error-guided Spatial Refinement module (ESR) for feature refinement from the spatial perspective. First, we spatially align LaRE and the image feature map by a simple adaptive average pooling layer, so that they have the same spatial size. Now we have the spatially aligned input feature map \\(\\mathbf{x} \\in \\mathbb{R}^{HW \\times C_1}\\) and LaRE \\(\\mathbf{e} \\in \\mathbb{R}^{HW \\times C_2}\\). \\(\\bar{\\mathbf{x}} \\in \\mathbb{R}^{1 \\times C_1}\\) is the global feature, which is the average of \\(\\mathbf{x}\\). To refine the feature from the spatial perspective, we choose an Error-guided Spatial Attention module (ESA):" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.508, + 0.47, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}, \\mathbf {E}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {E}\\right) \\mathbf {V}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.47, + 0.606 + ], + "angle": 0, + "content": "To further enhance the feature representation, we adopt a multi-head mechanism from Transformer [37]. Therefore, our Multi-Head Error-guided Spatial Attention module (MHESA) is defined as:" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.611, + 0.462, + 0.63 + ], + "angle": 0, + "content": "\\(\\mathrm{MHESA}(\\mathbf{Q},\\mathbf{K},\\mathbf{V},\\mathbf{E}) = \\mathrm{Concat}(\\mathrm{head}_1,\\dots,\\mathrm{head}_h)\\mathbf{W}^O\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.632, + 0.469, + 0.663 + ], + "angle": 0, + "content": "where \\(\\mathrm{head}_i = \\mathrm{ESA}(\\mathbf{Q}\\mathbf{W}_i^Q,\\mathbf{K}\\mathbf{W}_i^K,\\mathbf{V}\\mathbf{W}_i^V,\\mathbf{E}\\mathbf{W}_i^E)\\) (8)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.726 + ], + "angle": 0, + "content": "where \\(h\\) is the number of heads, \\(\\mathbf{W}_i^Q\\in \\mathbb{R}^{C_1\\times d}\\), \\(\\mathbf{W}_i^K\\in \\mathbb{R}^{C_1\\times d}\\), \\(\\mathbf{W}_i^V\\in \\mathbb{R}^{C_1\\times d}\\), \\(\\mathbf{W}_E^Q\\in \\mathbb{R}^{C2\\times d}\\) are learnable projection matrices, \\(d\\) is the hyperparameter. The spatial refined feature \\(x^{\\prime}\\) is calculated by:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.733, + 0.469, + 0.749 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {s} = \\operatorname {M H E S A} (\\bar {\\mathbf {x}}, \\mathbf {x}, \\mathbf {x}, \\mathbf {e}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.755, + 0.469, + 0.83 + ], + "angle": 0, + "content": "where \\(\\bar{x}\\) is the only query attending to the feature map \\(x\\). In this stage, the aligned LaRE is used to re-weight the attention score of scaled dot-product attention. Since LaRE and the feature map are spatially aligned, LaRE can emphasize important information from the spatial perspective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.428, + 0.862 + ], + "angle": 0, + "content": "4.2.2 Error-guided Channel Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.9 + ], + "angle": 0, + "content": "For channel refinement, we first squeeze both \\(\\mathbf{x}\\) and \\(\\mathbf{e}\\) to \\(\\bar{\\mathbf{x}}\\in \\mathbb{R}^{1\\times C_1}\\) and \\(\\bar{\\mathbf{e}}\\in \\mathbb{R}^{1\\times C_2}\\). Then our channel refinement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.344, + 0.722, + 0.358 + ], + "angle": 0, + "content": "is achieved by a gate mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.366, + 0.892, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {c} = \\text {s i g m o i d} (\\bar {\\mathbf {e}} \\mathbf {W}) \\odot \\bar {\\mathbf {x}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.449 + ], + "angle": 0, + "content": "where \\(\\mathbf{W} \\in \\mathbb{R}^{C1 \\times C2}\\) are learnable parameters to align LaRE and the feature map from the channel perspective. Finally, \\(\\mathbf{x}_s\\), \\(\\mathbf{x}_c\\) and the original global feature \\(\\mathbf{x}_g\\) from the last convolution block are concatenated as the final feature \\(\\mathbf{x}_{\\mathrm{EGRE}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.457, + 0.891, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {E G R E}} = \\operatorname {C o n c a t} \\left(\\mathbf {x} _ {s}, \\mathbf {x} _ {c}, \\mathbf {x} _ {g}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.509 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_{\\mathrm{EGRE}}\\) is followed by one FC layer. We train our model with the binary cross entropy loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.633, + 0.538 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.787, + 0.56 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.841 + ], + "angle": 0, + "content": "We evaluate our proposed method using the GenImage [47] dataset. This dataset comprises a total of 2,681,167 images, divided into 1,331,167 real images and 1,350,000 generated images. The generated images are from 8 different generative models, namely BigGAN [4], GLIDE [26], VQDM [12], Stable Diffusion V1.4&V1.5 [32], ADM [9], Midjourney [1], and Wukong [2]. All of the fake images are generated using the template prompt \"photo of CLS\", where \"CLS\" is replaced by one of the 1000 labels from ImageNet [8]. These images are split into 8 subsets where each subset contains partial real images and all the fake images generated by one of the above generators. We adhere to the official division of the dataset in our research, allocating 2,581,167 images for training and reserving the remaining 100,000 images for validation. Following DIRE [40], we employ Accuracy (ACC) and Average Precision (AP) as our evaluation metrics. More details and results are included in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.717, + 0.864 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "All experiments are conducted using the GenImage dataset. To obtain the LaRE, we employed Stable Diffu" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17010" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.089, + 0.882, + 0.221 + ], + "angle": 0, + "content": "
MethodsTesting SubsetAvg ACC.(%)
MidjourneySDV1.4SDV1.5ADMGLIDEWukongVQDMBigGAN
CNNSpot [39]58.270.370.257.057.167.756.756.661.7
Spec [45]56.772.472.357.965.470.361.764.365.1
F3Net [29]55.173.173.166.557.872.362.156.564.6
GramNet [19]58.172.872.758.765.371.357.861.264.7
DIRE [40]†65.073.773.761.969.174.363.456.767.2
Ours66.487.387.166.781.385.584.474.079.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.233, + 0.893, + 0.263 + ], + "angle": 0, + "content": "Table 1. Performance comparisons on GenImage test set. For each data point, eight models are trained on eight generators respectively. Then eight models are tested on the specified test sets and the accuracy scores are averaged. †indicates our reproduction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.475 + ], + "angle": 0, + "content": "sion V1.5 [32] with a step size of \\( t = 200 \\) and a noise ensemble size of \\( e = 4 \\). Images are resized to \\( 256 \\times 256 \\) for LaRE extraction. We use the prompt \"a photo\" for all images. For both training and testing phases, images are resized to \\( 224 \\times 224 \\). Our model utilizes the CLIP [30] pretrained ResNet50 [13] as the backbone architecture. We set the batch size to 48 and the learning rate to \\( 1e^{-4} \\). Training is performed on a single Nvidia Tesla V100 GPU. We train eight models on eight subsets respectively, each corresponding to a different generation method. Model selection is based on performance in the validation set, which shares the same generation method as the training set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.405, + 0.499 + ], + "angle": 0, + "content": "5.3. Cross-Generator Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.473, + 0.825 + ], + "angle": 0, + "content": "We evaluate the performance of our model trained by one of the subsets and test on all eight subsets. The results are shown in Fig. 4. Several conclusions can be made from the results. (1) It is easy to detect generated images from seen generators. Both the DIRE and our proposed method demonstrate superior performance when training and testing images share the same generator, as shown by the diagonal of each image. (2) It is challenging to generalize the model to unseen generators. However, it is relatively easier to generalize the model to similar generator structures than to different structures. For example, a model trained on the BigGAN subset achieves bad performance when tested on other subsets, which are all diffusion-based generators. While the model trained on the SDV1.5 subset achieves relatively better performance on other diffusion-based models. (3) Our model is more robust and generalizable than DIRE. Our model trained on diffusion-based generators achieves better generalizability on both gan-based and diffusion-based generators. This indicates that our \\(\\mathrm{LaRE}^2\\) can better capture discriminative features to detect the generated images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.383, + 0.849 + ], + "angle": 0, + "content": "5.4. Compare with the State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we compare our method with several state-of-the-art generated image detection methods, including CNNSpot [39], Spec [45], F3Net [29], GramNet [19]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.49 + ], + "angle": 0, + "content": "and DIRE [40]. In accordance with the GenImage benchmark, eight models are trained on each subset. Subsequently, these models are evaluated on the specified test set, and their accuracy scores are averaged to produce a data point for inclusion in the table. This benchmark presents a formidable challenge, as the overall accuracy is closely tied to the cross-generator generalizability of the proposed techniques. As illustrated in Tab. 1 and Fig. 4, our approach attains state-of-the-art performance across all eight generators. Notably, our overall accuracy exhibits a substantial improvement, surpassing the previous state-of-the-art by \\(11.9\\% / 12.1\\%\\) ACC/AP. This result demonstrates the superior generalizability of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.655, + 0.517 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.894, + 0.57 + ], + "angle": 0, + "content": "In this section, we conduct several ablative studies to quantify the contribution of each design in our model and the influence of the hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.59, + 0.688, + 0.605 + ], + "angle": 0, + "content": "5.5.1 Influence of EGRE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The contribution of individual components in EGRE to the overall performance of the method is reported in Tab. 2. Starting with the most simple baseline, i.e., training without LaRE (Model A in the table), each component is added building up to the proposed approach. We measure the effect of the Error-guided Spatial Refinement module (ESR) and Error-guided Channel Refinement module (ECR). Specifically, ESR achieves \\(8.6\\% /5.3\\%\\) ACC/AP gain and ECR achieves \\(5.7\\% /2.4\\%\\) ACC/AP gain compared with the baseline. Both ESR and ECR significantly improve the performance, demonstrating the effectiveness of the two modules. Relatively, ESR brings a higher improvement in metrics compared with ECR, indicating a greater benefit from spatial feature refinement. This is also attributed to the better alignment of LaRE with the original image at the spatial scale as shown in Fig. 2. By combining ESR and ECR, our approach achieves a further improvement of \\(11.4\\% /7.2\\%\\) ACC/AP. By leveraging LaRE for feature refinement at both spatial and channel scales, our model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17011" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.885, + 0.199 + ], + "angle": 0, + "content": "
ModelESRECRCLSMJSDV1.4SDV1.5ADMGLIDEWuKongVQDMBigganAvg
A52.9/94.4100.0/100.0100.0/100.055.0/89.575.5/98.799.7/100.051.8/84.450.4/71.773.1/92.3
B✓61.8/96.9100.0/100.0100.0/100.060.3/95.487.2/99.1100.0/100.086.2/94.658.1/95.081.7/97.6
C✓57.4/95.3100.0/100.0100.0/100.057.2/91.681.2/98.9100.0/100.078.3/89.856.2/82.378.8/94.7
D✓✓64.8/98.2100.0/100.0100.0/100.061.9/98.591.4/99.8100.0/100.094.6/100.063.7/99.784.5/99.5
E✓✓✓66.2/99.3100.0/100.0100.0/100.064.5/99.591.5/99.9100.0/100.097.7/100.067.4/99.885.9/99.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.212, + 0.87, + 0.226 + ], + "angle": 0, + "content": "Table 2. Ablative studies on using different modules or prompts. 'CLS' indicates to use class-specific prompt for LaRE extraction." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.25, + 0.46, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.494 + ], + "angle": 0, + "content": "Figure 5. Trade-off between detection performance and feature extraction cost. When \\( e = 4 \\), the model achieves the best trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.552 + ], + "angle": 0, + "content": "achieves more robust and generalizable image generation detection performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.329, + 0.587 + ], + "angle": 0, + "content": "5.5.2 Influence of Noise Ensemble" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.471, + 0.751 + ], + "angle": 0, + "content": "We systematically evaluated the impact of the noise ensemble parameter \\( e \\) on both the accuracy and feature extraction runtime. As shown in Fig. 5. The results demonstrate a consistent increase in accuracy as \\( e \\) increased. However, this improvement in accuracy comes at the cost of increased feature extraction runtime. Upon careful consideration of the trade-off between accuracy and computational cost, we propose \\( e = 4 \\) as the optimal choice. This selection of \\( e \\) provides an effective balance, maintaining the performance of our model while ensuring computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.279, + 0.787 + ], + "angle": 0, + "content": "5.5.3 Influence of Prompts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Because Latent diffusion is a text-conditioned model, a prompt is needed to denoise an image. Here we try to use a class-specific prompt instead of an unchanged prompt. We choose the template \"a photo of CLS\" where \"CLS\" is replaced with the ImageNet label of the image. Results are shown in Tab. 2. When using class-specific prompts (Model E), we can achieve further improvements by \\(1.4\\% /0.3\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.251, + 0.861, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.462, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Figure 6. ACC and AP when choosing different \\( t \\). The results demonstrate that our model is robust to the choice of \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.745 + ], + "angle": 0, + "content": "ACC/AP. This suggests that the category information of images can aid in the extraction of LaRE features. Even in the absence of category information, it is interesting that the accuracy does not decline significantly. We speculate this is because we did not fully convert the images into noise. Even with the added noise, the image still retains some of the original important semantic information, which can assist the model in denoising and thus obtaining good LaRE features. This experiment validates that when the textual information of an image is known, the extracted LaRE features can further enhance the performance. However, we still use the generic prompt in our method, as it already achieves satisfactory results. As for accurately describing images to further enhance performance, we leave that for future research work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.727, + 0.785 + ], + "angle": 0, + "content": "5.5.4 Influence of Sample Step" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Here we conduct the ablative experiment on the sample step \\( t \\). As illustrated in Fig. 1(b), there is a gap between the reconstruction loss of real images and the generated images. When \\( t \\in [100,500] \\), there is a more obvious difference in the reconstruction loss between the two types of images. To verify the impact of \\( t \\) selection on the performance, we conducted experiments with different values of \\( t \\), and the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17012" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.462, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.317, + 0.462, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Figure 7. Visualization of the spatial representation of the baseline model (a) and our model (b). Both models were trained on the SDV1.5 subset. In the case of seen data, both models exhibit good discrimination in the feature space. However, for unseen data, the baseline model shows a significant overlap between real images and fake images (red circle), whereas our model exhibits a smaller overlap (red circle). This further indicates the better generalization capability of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.754 + ], + "angle": 0, + "content": "results are shown in Fig. 6. When \\( t \\in [150, 300] \\), the model exhibits stable performance, indicating the robustness of our model with respect to \\( t \\). However, when \\( t \\) exceeds this range, the performance of the model tends to decline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.345, + 0.787 + ], + "angle": 0, + "content": "5.5.5 Influence of Input Information" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments to train the detection model with different inputs. The results are shown in Tab. 3. Since our LaRE is a compressed latent representation whose size is only \\(32 \\times 32 \\times 4\\). We train a Resnet-20 with LaRE as the only input (Model B). The experimental results show that both the AP and ACC metrics have decreased significantly. We speculate that there are two reasons. Firstly," + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.09, + 0.882, + 0.175 + ], + "angle": 0, + "content": "
ModelImageLaREMethodAVG. ACCAVG. AP
A✓None73.192.3
B✓None66.268.3
C✓✓Concat76.893.5
D✓✓EGRE84.599.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.187, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Table 3. Results of using different input. Models are trained on the SDv1.5 subset and tested on all the subsets. 'AVG' indicates average scores over 8 subsets. 'Concat' indicates concat LaRE with the image feature map after the last convolution block." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.503 + ], + "angle": 0, + "content": "LaRE is in the latent space and has more information loss compared to the original image. It is not sufficient to serve as the only information for detecting the generated images. Secondly, as shown in the Fig. 2, LaRE mainly reflects the reconstruction quality of the original image at different spatial positions. It is better to leverage LaRE to enhance the representation of the original image features. Therefore, we conducted further experiments. In model C, we only concatenated LaRE with the feature map of the image, which also improves the performance. This validates our hypothesis. Besides, model D achieves significant performance gain compared with Model C. By using EGRE, we can further improve the performance of our model by a significant margin, which also confirms the advantages of EGRE over simple feature concatenation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.516, + 0.828, + 0.532 + ], + "angle": 0, + "content": "5.6. Qualitative Results and Visualizations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.707 + ], + "angle": 0, + "content": "We utilize t-SNE [36] visualization to illustrate feature vectors derived from the final layers of our model and the baseline model, as depicted in Fig. 7. Each model is trained on the SDv1.5 subset of GenImage and evaluated across all eight generative methods. Although both models exhibit good feature discrimination on seen images, they both have limitations when it comes to unseen scenarios. For images from MJ, ADM, and BigGAN, both models show overlaps between real and fake images in the feature space. However, our model demonstrates a smaller proportion of overlap, indicating its better generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.619, + 0.739 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel reconstruction-based diffusion-generated image detection method called LaRE\\(^2\\). We come up with LaRE, a novel and more efficient reconstruction-based feature by reconstructing the image in the Latent space. Notably, LaRE is 8 times faster compared with existing reconstruction-based methods. By incorporating LaRE with the Error-guided Feature Refinement module (EGRE). Our \\(\\mathrm{LaRE}^2\\) achieves superior performance on diffusion-generated image detection, demonstrating state-of-the-art performance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17013" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Midjourney. https://www.midjourney.com/home/, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.174 + ], + "angle": 0, + "content": "[2] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.468, + 0.218 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.22, + 0.468, + 0.26 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.263, + 0.468, + 0.29 + ], + "angle": 0, + "content": "[5] Kevin Clark and Priyank Jaini. Text-to-image diffusion models are zero-shot classifiers. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.293, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[6] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In CVPR, pages 973-982, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.351, + 0.468, + 0.406 + ], + "angle": 0, + "content": "[7] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, pages 1-5. IEEE, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.408, + 0.468, + 0.45 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. 34:8780-8794, 2021. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.468, + 0.537 + ], + "angle": 0, + "content": "[10] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. 63(11):139–144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.597, + 0.468, + 0.652 + ], + "angle": 0, + "content": "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.655, + 0.468, + 0.696 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.468, + 0.725 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[16] Felix Juefei-Xu, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang Liu. Countering malicious deepfakes: Survey, battleground, and horizon. 130(7):1678-1734, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[17] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, pages 5001-5010, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[18] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[20] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. 35:5775-5787, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[21] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[22] Francesco Marra, Diego Gragnaniello, Davide Cozzolino, and Luisa Verdoliva. Detection of gan-generated fake images over social networks. In MIPR, pages 384-389. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[23] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using color cues. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[24] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using saturation cues. In ICIP, pages 4584-4588. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[25] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, Amit K Roy-Chowdhury, and BS Manjunath. Detecting gan generated fake images using co-occurrence matrices. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[28] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[29] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[31] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. 2020. 1, 2, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17014" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[35] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.163 + ], + "angle": 0, + "content": "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.468, + 0.207 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[38] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.362 + ], + "angle": 0, + "content": "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.468, + 0.405 + ], + "angle": 0, + "content": "[41] Haiwei Wu, Jiantao Zhou, and Shile Zhang. Generalizable synthetic image detection via language-guided contrastive learning. 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[42] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.468, + 0.518 + ], + "angle": 0, + "content": "[43] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.468, + 0.56 + ], + "angle": 0, + "content": "[44] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.468, + 0.603 + ], + "angle": 0, + "content": "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.468, + 0.646 + ], + "angle": 0, + "content": "[46] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.468, + 0.703 + ], + "angle": 0, + "content": "[47] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17015" + } + ] +] \ No newline at end of file diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_origin.pdf b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..042273824fe8edb0aeae7cca5e6423d601869b19 --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/fbde4cbd-2dff-4663-bad3-50e64c26068e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aaa447f5279b1a5f91f1a46ff14f505aad8e7c970058b768523b09f9fa783666 +size 4063717 diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/full.md b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09fbcb2eec6c0abdb19d1c3bd567fc012c8687bd --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/full.md @@ -0,0 +1,334 @@ +# LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection + +Yunpeng Luo Junlong Du Ke Yan* Shouhong Ding +Tencent YouTu Lab + +{petterluo, jeffdu, kerwinyan, ericshding} @tencent.com + +# Abstract + +The evolution of Diffusion Models has dramatically improved image generation quality, making it increasingly difficult to differentiate between real and generated images. This development, while impressive, also raises significant privacy and security concerns. In response to this, we propose a novel Latent REconstruction error guided feature REfinement method $(\mathsf{LaRE}^2)$ for detecting the diffusion-generated images. We come up with the Latent Reconstruction Error (LaRE), the first reconstruction-error based feature in the latent space for generated image detection. LaRE surpasses existing methods in terms of feature extraction efficiency while preserving crucial cues required to differentiate between the real and the fake. To exploit LaRE, we propose an Error-Guided feature REfinement module (EGRE), which can refine the image feature guided by LaRE to enhance the discriminativeness of the feature. Our EGRE utilizes an align-then-refine mechanism, which effectively refines the image feature for generated-image detection from both spatial and channel perspectives. Extensive experiments on the large-scale GenImage benchmark demonstrate the superiority of our $\mathsf{LaRE}^2$ , which surpasses the best SoTA method by up to $11.9\% / 12.1\%$ average ACC/AP across 8 different image generators. LaRE also surpasses existing methods in terms of feature extraction cost, delivering an impressive speed enhancement of 8 times. + +# 1. Introduction + +The rapid advancement of Diffusion Models has heralded a new era in the domain of image generation. Through concerted efforts in refining model architecture [9, 32, 33], optimizing training strategies [15, 27], and enhancing sampling methods [20, 34], contemporary Diffusion Models are now capable of generating images of unprecedented quality, surpassing the boundaries of human imagination. However, this progress raises significant concerns regarding privacy + +![](images/3575822d4c9245aa536155349939aa459d9f1aa2db807d6ab3adc8b2834c2e72.jpg) + +![](images/9fb5f683c615aea3d9fd4c20c86a79da90c331fe9b02b2ad7b325aeba85c17a0.jpg) +Figure 1. (a) The comparison of reconstruction-based feature extraction. Existing method [40] chooses to completely reconstruct an image by first gradually adding noise to the image and then denoising it, which involves dozens of sampling steps. Our method can directly calculate noisy images and denoise them with a single sample step. (b) Statistical analysis of the relationship between the single-step reconstruction loss (1000 images are used) and time step. The obvious gap between the two lines indicates that single-step reconstruction can also reflect the differences between real and generated images. (c) Comparison of the cost of per image feature extraction. Our method is $8\mathbf{x}$ faster than DIRE [40]. + +![](images/3da2ed5f7618302c7e8f23905f0f75e726882bd8b66ba33ad94465c2c68a4d96.jpg) + +![](images/1661650c2bf7aa79a9dccac4a312655962ab80d0cc48285d3683b2a77e96d3dd.jpg) + +and security associated with the generated images [16]. The potential for the dissemination of toxic content and misinformation through these images poses a threat to society and could mislead the public. Consequently, there is an urgent need to develop techniques to detect the images generated by these models. + +A flurry of works [6, 7, 21, 31, 40, 41] has been proposed to detect or study the properties of images generated by diffusion models. Recently, DIRE [40] has been proposed to leverage the reconstruction error as a discriminative feature for diffusion-generated image detection. It is based on the assumption that the diffusion-generated images are more easily to be reconstructed by a diffusion + +model compared with real images. As a result, DIRE shows great cross-model generalizability towards different diffusion models. Based on DIRE, SeDID [21] is also proposed to harness the inherent distributional disparities between naturally occurring and diffusion-synthesized visuals for diffusion-generated image detection. + +However, both DIRE and SeDID require multi-step DDIM [34] sampling processes in the feature extraction stage, which results in low efficiency for real-world applications. As shown in Fig 1(c), it takes more than 2 seconds to extract the DIRE feature for an image. Besides, the accumulation loss in the multi-step sampling also introduces uncertainty into the extracted features. Though DDIM [34] provides a deterministic inversion method to transform an image into noise, the reliability of this inversion process is not consistently upheld. In addition, DIRE and SeDID use the reconstruction error as the only feature, ignoring the correspondence between the error and the raw image. We therefore ask: (1) Do we need to completely reconstruct the image to get the discriminative feature? (2) Are there better ways to incorporate reconstruction error into image generation detection? + +To address the two questions mentioned above, we conducted a series of exploratory experiments and analyses. For the first question, we investigated the training process of the diffusion model. We found that given any time step $t$ , the forward Markov process has a closed-form solution, allowing us to directly transform $x_0$ into $x_t$ . Additionally, during the training phase, the model takes $x_t$ and $t$ as inputs and computes the denoising loss, enabling us to obtain the loss through single-step denoising. Therefore, we perform single-step denoising on both 1000 real and 1000 generated images respectively by LDM [32]. The results, as shown in Fig. 1(b), indicate that even with only single-step reconstruction, the loss of real images is consistently greater than that of generated images. This not only demonstrates that the real images are harder to reconstruct than the generated ones but also suggests that the loss from single-step reconstruction can reflect the differences between real and generated images. To address the second question, we visualize the reconstruction loss on the original image, as shown in Fig. 2. Upon observation, we have found that the reconstruction loss is positively correlated with the local information frequency of the original image. In the first image, for instance, the reconstruction loss is relatively lower on the low-frequency background, while it is relatively higher on the high-frequency foreground. The same pattern can be observed in other images as well. The reconstruction loss exhibits spatial correlations with the original image, thereby presenting a potential to serve as a valuable cue for the generated image detection. + +Based on the findings above, we propose a novel Latent REconstruction error guided feature REfinement method + +![](images/233c52e08f9df682e8d4a4044191cf611574ac3cde37987589a774542ce4a699.jpg) + +![](images/eb4aa63b0bc20a78427ae7ce47b86bb592eb3f9de635b0aae5175fd77bb90e95.jpg) + +![](images/6af487aa9aaa230bcadd9aadf0adc762447c754eca4b8b20d71e4c5e1d83afc7.jpg) +Raw Image + +![](images/235058afdf4a1f7864948925f4f15997e2e3fa62c728639724acba63ef45709f.jpg) + +![](images/132f43cbbfbea2acf82a4fb83ac8b179b1a4eee885f875f56bc52b8bbd496db2.jpg) + +![](images/55f643c0432b10b831758b6665e7a67ac421684d863c702590a3008f1299a844.jpg) +Added Noise + +![](images/fbc8c680b87283785377240463d8f4a4e2659f7281bbab02c2626da7d883c6fd.jpg) + +![](images/b89abf9f7d008f05c0757c6973a537394bc0d323652210b9ebada602887fc9a9.jpg) + +![](images/fe99fd27eb74ad90dda188033f002330f04f77bb571b96dc0906f16c1a9cdb4e.jpg) +Reconstruction Loss + +![](images/8940fa9db14b53fcce3e2f74e58a794c234c50cb0612c24c460d0d51f3098b1c.jpg) + +![](images/b052b26dc04a3531b9cff076889038e9b1d2d7a61b10b8e9ce8b4c8ea37db33f.jpg) + +![](images/5aa745475a498ebf86bafa85231ed703c8ca3c3186e9d5b35c9068009d818cb1.jpg) +Image + Loss +Figure 2. Visualization of reconstruction loss on raw images (i.e. Image + Loss). Though the randomly sampled noises are added to the whole image, there is a trend that the loss in high-frequency regions is typically greater than that in low-frequency regions. + +$(\mathrm{LaRE}^2)$ for diffusion-generated image detection. Our $\mathrm{LaRE}^2$ consists of two parts, the Latent Reconstruction Error (LaRE) and the Error-guided Feature REfinement module (EGRE). LaRE is a more efficient reconstruction feature, which has two improvements compared with existing methods: (1) LaRE is extracted in a single step of the diffusion reverse process, which is significantly more efficient than completely reconstructing the image through dozens of denoising steps. (2) We conduct the reconstruction in the latent space, which further improves the efficiency. Besides, we find that the reconstructed loss is positively correlated with the local information frequency of the original image. Based on that, we come up with the Error-Guided Feature Refinement module (EGRE). In EGRE, we first align LaRE with the image feature map for better correspondings for feature refinement. Then the aligned LaRE is used to refine the image feature in both spatial and channel perspectives. With the guidance of LaRE, EGRE refines the feature map in both spatial and channel perspectives to better reveal the discriminativeness for detecting the generated images. + +To evaluate the effectiveness of our method, we conduct extensive experiments on the GenImage [47] benchmark, which comprises 2,681,167 images, segregated into 1,331,167 real and 1,350,000 fake images. The fake images are from 8 different generators. As shown in Fig. 1(c), our LaRE is 8 times faster than existing method [40]. In addition, $\mathrm{LaRE}^2$ achieves a significant performance gain by up to $11.9\% / 12.1\%$ ACC/AP compared with the best SoTA. The results demonstrated the superiority of our method, which is both effective and generalizable. + +The contributions of our work are three-fold: + +- Novel feature: We are the first to propose the reconstruction error in latent space for generated-image de + +tection. Compared with the existing method, we remarkably reduce the cost of feature extraction while preserving the essential information required for the detection of diffusion-generated images. + +- Novel module: We qualitatively analyze the reason for the effectiveness of the reconstruction loss. Based on that, we come up with a novel module EGRE, which conducts an Error-guided feature refinement to enhance the discriminativeness of image features. +- Superior performance: Extensive experiments demonstrate the effectiveness of our method. We achieve $11.9\% / 12.1\%$ ACC/AP gain on the large-scale GenImage benchmark, significantly outperforming the SoTA methods. + +# 2. Related Works + +# 2.1. Diffusion Model + +Diffusion Models [14, 33, 43] have become new state-of-the-art deep generative models. Inspired by nonequilibrium thermodynamics, Denoising Diffusion Probabilistic Models (DDPM) demonstrate promising generative quality, which inspires a flurry of studies on diffusion models [3, 5, 9, 32, 34, 35, 42, 46]. DDIM [34] generalize DDPMs [14] via a class of non-Markovian diffusion processes to accelerate sampling. ADM [9] firstly obtain better generation quality than GANs [11] on ImageNet [8]. Latent Diffusion Models [32] apply the diffusion process in the latent space to improve the efficiency of diffusion models, and enable diffusion models with text conditioning inputs through the cross-attention conditioning mechanism. There are also works that employ pre-trained diffusion models for various downstream tasks [3, 5, 35, 42, 46]. The rapid advancement of deep generative models has given rise to concerns regarding the potential for malicious utilization of the generated images [16]. Consequently, it is urgent to develop robust techniques to detect generated images. + +# 2.2. Deep Generated Image Detection + +Deep-generated image detection has achieved significant improvements due to the contributions of previous research. Initially, researchers attempted to use hand-crafted features including color cues [23], saturation cues [24], blending [17] artifacts, co-occurrence features [25]. Then CNN is leveraged to detect the generated images [18, 22, 38]. Some works [10, 44] also find that there are obvious visual artifacts in the GAN-generated images and detect these images from the frequency view [28]. Most of the above works are developed and designed for GAN-generated images. With the advance of diffusion models, many methods are also provided for the study of diffusion-generated image detection. Corvi et. al [7, 31] find that state-of-the-art detec + +![](images/641a0953ccb1dd5d69ac118439d4bd76084b38f76d41465ef82925fc27e33140.jpg) +Figure 3. Overview of our method. In the first stage, we extract LaRE in the latent space through single-step reconstruction. In the second stage, to exploit LaRE, we propose the Error-guided Feature Refinement Module, which consists of the Error-guided spatial refinement module and the Error-guided Channel Refinement module. From both spatial and channel perspectives, LaRE is used to enhance the discriminativeness of the image feature for generated image detection. + +tors developed for GAN suffer from a severe performance drop when applied to Diffusion-generated images. Corvi et. al [6,7,31] also find that the spectrum artifacts can also rise up in the Diffusion-generated images. Wu et. al [41] formulate the synthetic image detection as an identification problem and achieve generalizable detection through language-guided contrastive learning. DIRE [40] and SeDIE [21] are proposed to leverage reconstruction error from the diffusion model to achieve generalizable diffusion-generated image detection. Different from these methods, we are the first to show that reconstruction error in the latent space can also benefit diffusion-generated image detection, which is more efficient and generalizable. + +# 3. Preliminaries + +Diffusion Models has achieved remarkable image generation performance. Typically, it involves two Markov processes. In the forward process, Gaussian noise is gradually added to the raw image $\mathbf{x}_0$ until the image is asymptotically transformed to pure noise, which is defined as: + +$$ +q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1}) = N (\mathbf {x} _ {t}; \sqrt {\frac {\alpha_ {t}}{\alpha_ {t - 1}}} \mathbf {x} _ {t - 1}, (1 - \frac {\alpha_ {t}}{\alpha_ {t - 1}} \mathbf {I})), \quad (1) +$$ + +where $x_{t}$ is the noisy image at step $t$ and $\alpha_{t}$ is predefined noise schedule. According to the properties of Markov process and Gaussian distribution, we can get $\mathbf{x}_t$ from $\mathbf{x}_0$ directly by: + +$$ +q \left(\mathbf {x} _ {t} \mid \mathbf {x} _ {0}\right) = N \left(\mathbf {x} _ {t}; \sqrt {\alpha_ {t}} \mathbf {x} _ {0}, (1 - \alpha_ {t}) \mathbf {I})\right). \tag {2} +$$ + +In the reverse process, the noisy image is gradually denoised to get the raw image, which is defined as: + +$$ +p _ {\theta} (\mathbf {x} _ {t - 1} | \mathbf {x} _ {t}) = N (\mathbf {x} _ {t - 1}; \mu_ {\theta} (\mathbf {x} _ {t}, t), \boldsymbol {\Sigma} _ {\theta} (\mathbf {x} _ {t}, t)), \quad (3) +$$ + +where $p_{\theta}$ is typically parameterized by neural networks. During training, a neural network $\epsilon_{\theta}$ is trained to predict the added noise $\epsilon$ , given the noisy image $x_{t}$ and corresponding time step $t$ : + +$$ +L _ {\theta} \left(\mathbf {x} _ {0}, t\right) = \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha}} _ {t} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \| ^ {2}. \tag {4} +$$ + +During training, $t$ and $\epsilon$ are randomly sampled. + +# 4. Methods + +In this section, we elaborate on each component of our $\mathrm{LaRE}^2$ , as illustrated in Fig. 3. + +# 4.1. Latent Reconstruction Error + +Recent methods [40] propose to use reconstruction error as the feature for generated-image detection. Through DDIM [34] inversion, an image is first inversed to noise and then re-generated based on the noise. However, this method has several limitations: (1) It is relatively slow to inverse and re-generate an image. For example, it takes more than 2 seconds to extract DIRE for an image on a Tesla V100 GPU as shown in Fig. 1(c), which limits its realistic applications; (2) The reconstruction of the image is not always reliable. Error could accumulate during the Markov forward and reverse process, making it hard to tell whether the bad reconstruction is from the difference between real and fake images, or the reconstruction itself. + +Here we come up with a new reconstruction feature, named Latent Reconstruction Error (LaRE). LaRE is based on the assumption that if the generated images can be completely reconstructed with relative ease [40], then they can also be more easily reconstructed at every single step of the reverse diffusion process. + +For an image $x$ , we first get its latent code $x_0$ by VAE. Then we calculate the LaRE by: + +$$ +\operatorname {L a R E} = L _ {\theta} \left(\mathbf {x} _ {0}, t\right) = \mathbb {E} _ {\epsilon} \left[ \| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha}} _ {t} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t\right) \| ^ {2} \right]. \tag {5} +$$ + +As shown in Fig. 3, Our LaRE takes advantage of two special properties of Diffusion Models. (1) The forward process has a close form. According to Eq. 2, we can get $\mathbf{x}_t$ directly from $\mathbf{x}_0$ ; (2) Diffusion Model is trained to denoise a noisy image from any time step $t$ , as shown in Eq. 4. Therefore, given $\mathbf{x}_t$ , $t$ , and $\epsilon$ , we can directly get the reconstruction error by only single-step denoising, which is much more efficient than completely reconstructing an image. Compared with complete image reconstruction, our LaRE has several advantages: (1) Only single-step denoising is required to get LaRE, which significantly improves its efficiency. (2) Reconstruction is conducted in the latent space, which is also more efficient. (3) No need for complete inversion and reconstruction of the image, which eliminates the potential accumulation of errors that may occur during multiple inversion and reconstruction steps. + +With efficiency, the easier reconstruction assumption still holds for our LaRE. Recall that a diffusion-generated image is more easily reconstructed by diffusion models [40]. So the reconstruction error should be relatively small when completely reconstructing an image through the whole reverse process. That implies the reconstruction error is also relatively small during each step of the reverse process, which is also validated in Fig. 1(b). We leverage the Monte Carlo method to estimate LaRE in Eq. 6: + +$$ +\mathrm {L a R E} = \frac {1}{e} \sum_ {i} ^ {e} \left[ \| \epsilon_ {i} - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha}} _ {t} \mathbf {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {i}, t\right) \| ^ {2} \right], \tag {6} +$$ + +where $e, t$ are pre-defined hyperparameters. Specifically, we conduct denoising for $e$ times and average the denoising loss to extract LaRE. + +# 4.2. Error-Guided Feature Refinement + +Existing methods exploit the reconstruction error as the only feature to detect the generated image, which ignores the relationships between the reconstruction error and the raw image. To better understand how reconstruction error helps generated-image detection, we visualize the extracted LaRE on the raw image, as shown in Fig. 2. We find several interesting conclusions. 1) The loss norm is proportional to the noise norm. It is easy to understand that the stronger the added noise is, the harder to reconstruct; 2) The loss norm is also proportional to the frequency of local image patches. That the low-frequency patches are easier to reconstruct than the high-frequency ones. Therefore, the reconstruction performance on the high-frequency parts of the image plays an important role in the generated image detection. To apply this valuable information, we come up with a novel Error-guided feature REfinement Module (EGRE). Based on the reconstruction error, our EGRE can better refine the image feature to reveal the discriminative parts for generated image detection. + +![](images/edb473f6131e88c39bd4fb0cd4a68a5530dc808aeb76c088802f933669124e31.jpg) +(a) Comparison on ACC. Left: baseline, right: ours + +![](images/c84a2b4faa7412863231eafb74113f7a2884de04631bd73d2c46db20cec09d98.jpg) +(b) Comparison on AP. Left: baseline, right: ours +Figure 4. Results of cross-validation on different training and testing subsets. For each generator, we train a model and test it on all 8 generators. For both DIRE [40] and our method, accuracy (ACC) and average precision(AP) are reported. + +# 4.2.1 Error-guided Spatial Refinement Module + +We propose the Error-guided Spatial Refinement module (ESR) for feature refinement from the spatial perspective. First, we spatially align LaRE and the image feature map by a simple adaptive average pooling layer, so that they have the same spatial size. Now we have the spatially aligned input feature map $\mathbf{x} \in \mathbb{R}^{HW \times C_1}$ and LaRE $\mathbf{e} \in \mathbb{R}^{HW \times C_2}$ . $\bar{\mathbf{x}} \in \mathbb{R}^{1 \times C_1}$ is the global feature, which is the average of $\mathbf{x}$ . To refine the feature from the spatial perspective, we choose an Error-guided Spatial Attention module (ESA): + +$$ +\operatorname {E S A} (\mathbf {Q}, \mathbf {K}, \mathbf {V}, \mathbf {E}) = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} \mathbf {K} ^ {T}}{\sqrt {d _ {k}}} + \mathbf {E}\right) \mathbf {V}. \tag {7} +$$ + +To further enhance the feature representation, we adopt a multi-head mechanism from Transformer [37]. Therefore, our Multi-Head Error-guided Spatial Attention module (MHESA) is defined as: + +$\mathrm{MHESA}(\mathbf{Q},\mathbf{K},\mathbf{V},\mathbf{E}) = \mathrm{Concat}(\mathrm{head}_1,\dots,\mathrm{head}_h)\mathbf{W}^O$ + +where $\mathrm{head}_i = \mathrm{ESA}(\mathbf{Q}\mathbf{W}_i^Q,\mathbf{K}\mathbf{W}_i^K,\mathbf{V}\mathbf{W}_i^V,\mathbf{E}\mathbf{W}_i^E)$ (8) + +where $h$ is the number of heads, $\mathbf{W}_i^Q\in \mathbb{R}^{C_1\times d}$ , $\mathbf{W}_i^K\in \mathbb{R}^{C_1\times d}$ , $\mathbf{W}_i^V\in \mathbb{R}^{C_1\times d}$ , $\mathbf{W}_E^Q\in \mathbb{R}^{C2\times d}$ are learnable projection matrices, $d$ is the hyperparameter. The spatial refined feature $x^{\prime}$ is calculated by: + +$$ +\mathbf {x} _ {s} = \operatorname {M H E S A} (\bar {\mathbf {x}}, \mathbf {x}, \mathbf {x}, \mathbf {e}), \tag {9} +$$ + +where $\bar{x}$ is the only query attending to the feature map $x$ . In this stage, the aligned LaRE is used to re-weight the attention score of scaled dot-product attention. Since LaRE and the feature map are spatially aligned, LaRE can emphasize important information from the spatial perspective. + +# 4.2.2 Error-guided Channel Refinement Module + +For channel refinement, we first squeeze both $\mathbf{x}$ and $\mathbf{e}$ to $\bar{\mathbf{x}}\in \mathbb{R}^{1\times C_1}$ and $\bar{\mathbf{e}}\in \mathbb{R}^{1\times C_2}$ . Then our channel refinement + +is achieved by a gate mechanism: + +$$ +\mathbf {x} _ {c} = \text {s i g m o i d} (\bar {\mathbf {e}} \mathbf {W}) \odot \bar {\mathbf {x}}, \tag {10} +$$ + +where $\mathbf{W} \in \mathbb{R}^{C1 \times C2}$ are learnable parameters to align LaRE and the feature map from the channel perspective. Finally, $\mathbf{x}_s$ , $\mathbf{x}_c$ and the original global feature $\mathbf{x}_g$ from the last convolution block are concatenated as the final feature $\mathbf{x}_{\mathrm{EGRE}}$ . + +$$ +\mathbf {x} _ {\text {E G R E}} = \operatorname {C o n c a t} \left(\mathbf {x} _ {s}, \mathbf {x} _ {c}, \mathbf {x} _ {g}\right). \tag {11} +$$ + +$\mathbf{x}_{\mathrm{EGRE}}$ is followed by one FC layer. We train our model with the binary cross entropy loss. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Metrics + +We evaluate our proposed method using the GenImage [47] dataset. This dataset comprises a total of 2,681,167 images, divided into 1,331,167 real images and 1,350,000 generated images. The generated images are from 8 different generative models, namely BigGAN [4], GLIDE [26], VQDM [12], Stable Diffusion V1.4&V1.5 [32], ADM [9], Midjourney [1], and Wukong [2]. All of the fake images are generated using the template prompt "photo of CLS", where "CLS" is replaced by one of the 1000 labels from ImageNet [8]. These images are split into 8 subsets where each subset contains partial real images and all the fake images generated by one of the above generators. We adhere to the official division of the dataset in our research, allocating 2,581,167 images for training and reserving the remaining 100,000 images for validation. Following DIRE [40], we employ Accuracy (ACC) and Average Precision (AP) as our evaluation metrics. More details and results are included in our supplementary materials. + +# 5.2. Implementation Details + +All experiments are conducted using the GenImage dataset. To obtain the LaRE, we employed Stable Diffu + +
MethodsTesting SubsetAvg ACC.(%)
MidjourneySDV1.4SDV1.5ADMGLIDEWukongVQDMBigGAN
CNNSpot [39]58.270.370.257.057.167.756.756.661.7
Spec [45]56.772.472.357.965.470.361.764.365.1
F3Net [29]55.173.173.166.557.872.362.156.564.6
GramNet [19]58.172.872.758.765.371.357.861.264.7
DIRE [40]†65.073.773.761.969.174.363.456.767.2
Ours66.487.387.166.781.385.584.474.079.1
+ +Table 1. Performance comparisons on GenImage test set. For each data point, eight models are trained on eight generators respectively. Then eight models are tested on the specified test sets and the accuracy scores are averaged. †indicates our reproduction. + +sion V1.5 [32] with a step size of $t = 200$ and a noise ensemble size of $e = 4$ . Images are resized to $256 \times 256$ for LaRE extraction. We use the prompt "a photo" for all images. For both training and testing phases, images are resized to $224 \times 224$ . Our model utilizes the CLIP [30] pretrained ResNet50 [13] as the backbone architecture. We set the batch size to 48 and the learning rate to $1e^{-4}$ . Training is performed on a single Nvidia Tesla V100 GPU. We train eight models on eight subsets respectively, each corresponding to a different generation method. Model selection is based on performance in the validation set, which shares the same generation method as the training set. + +# 5.3. Cross-Generator Image Classification + +We evaluate the performance of our model trained by one of the subsets and test on all eight subsets. The results are shown in Fig. 4. Several conclusions can be made from the results. (1) It is easy to detect generated images from seen generators. Both the DIRE and our proposed method demonstrate superior performance when training and testing images share the same generator, as shown by the diagonal of each image. (2) It is challenging to generalize the model to unseen generators. However, it is relatively easier to generalize the model to similar generator structures than to different structures. For example, a model trained on the BigGAN subset achieves bad performance when tested on other subsets, which are all diffusion-based generators. While the model trained on the SDV1.5 subset achieves relatively better performance on other diffusion-based models. (3) Our model is more robust and generalizable than DIRE. Our model trained on diffusion-based generators achieves better generalizability on both gan-based and diffusion-based generators. This indicates that our $\mathrm{LaRE}^2$ can better capture discriminative features to detect the generated images. + +# 5.4. Compare with the State of the Arts + +In this section, we compare our method with several state-of-the-art generated image detection methods, including CNNSpot [39], Spec [45], F3Net [29], GramNet [19] + +and DIRE [40]. In accordance with the GenImage benchmark, eight models are trained on each subset. Subsequently, these models are evaluated on the specified test set, and their accuracy scores are averaged to produce a data point for inclusion in the table. This benchmark presents a formidable challenge, as the overall accuracy is closely tied to the cross-generator generalizability of the proposed techniques. As illustrated in Tab. 1 and Fig. 4, our approach attains state-of-the-art performance across all eight generators. Notably, our overall accuracy exhibits a substantial improvement, surpassing the previous state-of-the-art by $11.9\% / 12.1\%$ ACC/AP. This result demonstrates the superior generalizability of our method. + +# 5.5. Ablation Study + +In this section, we conduct several ablative studies to quantify the contribution of each design in our model and the influence of the hyperparameters. + +# 5.5.1 Influence of EGRE + +The contribution of individual components in EGRE to the overall performance of the method is reported in Tab. 2. Starting with the most simple baseline, i.e., training without LaRE (Model A in the table), each component is added building up to the proposed approach. We measure the effect of the Error-guided Spatial Refinement module (ESR) and Error-guided Channel Refinement module (ECR). Specifically, ESR achieves $8.6\% /5.3\%$ ACC/AP gain and ECR achieves $5.7\% /2.4\%$ ACC/AP gain compared with the baseline. Both ESR and ECR significantly improve the performance, demonstrating the effectiveness of the two modules. Relatively, ESR brings a higher improvement in metrics compared with ECR, indicating a greater benefit from spatial feature refinement. This is also attributed to the better alignment of LaRE with the original image at the spatial scale as shown in Fig. 2. By combining ESR and ECR, our approach achieves a further improvement of $11.4\% /7.2\%$ ACC/AP. By leveraging LaRE for feature refinement at both spatial and channel scales, our model + +
ModelESRECRCLSMJSDV1.4SDV1.5ADMGLIDEWuKongVQDMBigganAvg
A52.9/94.4100.0/100.0100.0/100.055.0/89.575.5/98.799.7/100.051.8/84.450.4/71.773.1/92.3
B✓61.8/96.9100.0/100.0100.0/100.060.3/95.487.2/99.1100.0/100.086.2/94.658.1/95.081.7/97.6
C✓57.4/95.3100.0/100.0100.0/100.057.2/91.681.2/98.9100.0/100.078.3/89.856.2/82.378.8/94.7
D✓✓64.8/98.2100.0/100.0100.0/100.061.9/98.591.4/99.8100.0/100.094.6/100.063.7/99.784.5/99.5
E✓✓✓66.2/99.3100.0/100.0100.0/100.064.5/99.591.5/99.9100.0/100.097.7/100.067.4/99.885.9/99.8
+ +Table 2. Ablative studies on using different modules or prompts. 'CLS' indicates to use class-specific prompt for LaRE extraction. + +![](images/d3fe8e64e4dab045164e5fac76e51ffd13806feae5e561e17a3e1a5ec26fe935.jpg) +Figure 5. Trade-off between detection performance and feature extraction cost. When $e = 4$ , the model achieves the best trade-off. + +achieves more robust and generalizable image generation detection performance. + +# 5.5.2 Influence of Noise Ensemble + +We systematically evaluated the impact of the noise ensemble parameter $e$ on both the accuracy and feature extraction runtime. As shown in Fig. 5. The results demonstrate a consistent increase in accuracy as $e$ increased. However, this improvement in accuracy comes at the cost of increased feature extraction runtime. Upon careful consideration of the trade-off between accuracy and computational cost, we propose $e = 4$ as the optimal choice. This selection of $e$ provides an effective balance, maintaining the performance of our model while ensuring computational efficiency. + +# 5.5.3 Influence of Prompts + +Because Latent diffusion is a text-conditioned model, a prompt is needed to denoise an image. Here we try to use a class-specific prompt instead of an unchanged prompt. We choose the template "a photo of CLS" where "CLS" is replaced with the ImageNet label of the image. Results are shown in Tab. 2. When using class-specific prompts (Model E), we can achieve further improvements by $1.4\% /0.3\%$ + +![](images/156bc147d5b4040d8b7ca856d9105748cbab059f271f9c23bfd70324cf885d8d.jpg) +Figure 6. ACC and AP when choosing different $t$ . The results demonstrate that our model is robust to the choice of $t$ . + +ACC/AP. This suggests that the category information of images can aid in the extraction of LaRE features. Even in the absence of category information, it is interesting that the accuracy does not decline significantly. We speculate this is because we did not fully convert the images into noise. Even with the added noise, the image still retains some of the original important semantic information, which can assist the model in denoising and thus obtaining good LaRE features. This experiment validates that when the textual information of an image is known, the extracted LaRE features can further enhance the performance. However, we still use the generic prompt in our method, as it already achieves satisfactory results. As for accurately describing images to further enhance performance, we leave that for future research work. + +# 5.5.4 Influence of Sample Step + +Here we conduct the ablative experiment on the sample step $t$ . As illustrated in Fig. 1(b), there is a gap between the reconstruction loss of real images and the generated images. When $t \in [100,500]$ , there is a more obvious difference in the reconstruction loss between the two types of images. To verify the impact of $t$ selection on the performance, we conducted experiments with different values of $t$ , and the + +![](images/a6d1690052959bbffafe8201d40fb1305a5a37f19347e2429c7d281efbd62d96.jpg) + +![](images/e11dda25136f9bea9c49038ee1aab790164f1db6a86da6fe8cc3be4a804b3275.jpg) +Figure 7. Visualization of the spatial representation of the baseline model (a) and our model (b). Both models were trained on the SDV1.5 subset. In the case of seen data, both models exhibit good discrimination in the feature space. However, for unseen data, the baseline model shows a significant overlap between real images and fake images (red circle), whereas our model exhibits a smaller overlap (red circle). This further indicates the better generalization capability of our model. + +results are shown in Fig. 6. When $t \in [150, 300]$ , the model exhibits stable performance, indicating the robustness of our model with respect to $t$ . However, when $t$ exceeds this range, the performance of the model tends to decline. + +# 5.5.5 Influence of Input Information + +We conduct experiments to train the detection model with different inputs. The results are shown in Tab. 3. Since our LaRE is a compressed latent representation whose size is only $32 \times 32 \times 4$ . We train a Resnet-20 with LaRE as the only input (Model B). The experimental results show that both the AP and ACC metrics have decreased significantly. We speculate that there are two reasons. Firstly, + +
ModelImageLaREMethodAVG. ACCAVG. AP
A✓None73.192.3
B✓None66.268.3
C✓✓Concat76.893.5
D✓✓EGRE84.599.5
+ +Table 3. Results of using different input. Models are trained on the SDv1.5 subset and tested on all the subsets. 'AVG' indicates average scores over 8 subsets. 'Concat' indicates concat LaRE with the image feature map after the last convolution block. + +LaRE is in the latent space and has more information loss compared to the original image. It is not sufficient to serve as the only information for detecting the generated images. Secondly, as shown in the Fig. 2, LaRE mainly reflects the reconstruction quality of the original image at different spatial positions. It is better to leverage LaRE to enhance the representation of the original image features. Therefore, we conducted further experiments. In model C, we only concatenated LaRE with the feature map of the image, which also improves the performance. This validates our hypothesis. Besides, model D achieves significant performance gain compared with Model C. By using EGRE, we can further improve the performance of our model by a significant margin, which also confirms the advantages of EGRE over simple feature concatenation. + +# 5.6. Qualitative Results and Visualizations + +We utilize t-SNE [36] visualization to illustrate feature vectors derived from the final layers of our model and the baseline model, as depicted in Fig. 7. Each model is trained on the SDv1.5 subset of GenImage and evaluated across all eight generative methods. Although both models exhibit good feature discrimination on seen images, they both have limitations when it comes to unseen scenarios. For images from MJ, ADM, and BigGAN, both models show overlaps between real and fake images in the feature space. However, our model demonstrates a smaller proportion of overlap, indicating its better generalization ability. + +# 6. Conclusion + +In this paper, we propose a novel reconstruction-based diffusion-generated image detection method called LaRE $^2$ . We come up with LaRE, a novel and more efficient reconstruction-based feature by reconstructing the image in the Latent space. Notably, LaRE is 8 times faster compared with existing reconstruction-based methods. By incorporating LaRE with the Error-guided Feature Refinement module (EGRE). Our $\mathrm{LaRE}^2$ achieves superior performance on diffusion-generated image detection, demonstrating state-of-the-art performance. + +# References + +[1] Midjourney. https://www.midjourney.com/home/, 2022.5 +[2] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. 2021. 3 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5 +[5] Kevin Clark and Priyank Jaini. Text-to-image diffusion models are zero-shot classifiers. 2023. 3 +[6] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In CVPR, pages 973-982, 2023. 1, 3 +[7] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, pages 1-5. IEEE, 2023. 1, 3 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 3, 5 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. 34:8780-8794, 2021. 1, 3, 5 +[10] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. 63(11):139–144, 2020. 3 +[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 5 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 33:6840-6851, 2020. 3 +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. 2022. 1 +[16] Felix Juefei-Xu, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang Liu. Countering malicious deepfakes: Survey, battleground, and horizon. 130(7):1678-1734, 2022. 1, 3 +[17] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, pages 5001-5010, 2020. 3 +[18] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 3 + +[19] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 6 +[20] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. 35:5775-5787, 2022. 1 +[21] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. 2023. 1, 2, 3 +[22] Francesco Marra, Diego Gragnaniello, Davide Cozzolino, and Luisa Verdoliva. Detection of gan-generated fake images over social networks. In MIPR, pages 384-389. IEEE, 2018. 3 +[23] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using color cues. 2018. 3 +[24] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using saturation cues. In ICIP, pages 4584-4588. IEEE, 2019. 3 +[25] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, Amit K Roy-Chowdhury, and BS Manjunath. Detecting gan generated fake images using co-occurrence matrices. 2019. 3 +[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. 2021. 5 +[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 1 +[28] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 3 +[29] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6 +[31] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. 2022. 1, 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3, 5, 6 +[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1, 3 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. 2020. 1, 2, 3, 4 + +[35] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. 2023. 3 +[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. 9(11), 2008. 8 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 30, 2017. 5 +[38] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 3 +[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 6 +[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. 2023. 1, 2, 3, 4, 5, 6 +[41] Haiwei Wu, Jiantao Zhou, and Shile Zhang. Generalizable synthetic image detection via language-guided contrastive learning. 2023. 1, 3 +[42] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 3 +[43] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. 2022. 3 +[44] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 3 +[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 6 +[46] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. 2023. 3 +[47] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. 2023. 2, 5 \ No newline at end of file diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/images.zip b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3dbf7b117f00208981cf5fc6bb69c8f4e8d7c7ec --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51f34f81a8301b4de259184390198c25d89cf5e8bb58e172c3b577cd56e0657b +size 631319 diff --git a/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/layout.json b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1c59f8fd7cb3cdcef9e274802ea5dad6f52beac0 --- /dev/null +++ b/2024/LaRE^2_ Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection/layout.json @@ -0,0 +1,8834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 102, + 459, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 102, + 459, + 140 + ], + "spans": [ + { + "bbox": [ + 135, + 102, + 459, + 140 + ], + "type": "text", + "content": "LaRE2: Latent Reconstruction Error Based Method for Diffusion-Generated Image Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 161, + 436, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 161, + 436, + 188 + ], + "spans": [ + { + "bbox": [ + 157, + 161, + 436, + 188 + ], + "type": "text", + "content": "Yunpeng Luo Junlong Du Ke Yan* Shouhong Ding \nTencent YouTu Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 191, + 398, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 191, + 398, + 202 + ], + "spans": [ + { + "bbox": [ + 193, + 191, + 398, + 202 + ], + "type": "text", + "content": "{petterluo, jeffdu, kerwinyan, ericshding} @tencent.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "text", + "content": "The evolution of Diffusion Models has dramatically improved image generation quality, making it increasingly difficult to differentiate between real and generated images. This development, while impressive, also raises significant privacy and security concerns. In response to this, we propose a novel Latent REconstruction error guided feature REfinement method " + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "inline_equation", + "content": "(\\mathsf{LaRE}^2)" + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "text", + "content": " for detecting the diffusion-generated images. We come up with the Latent Reconstruction Error (LaRE), the first reconstruction-error based feature in the latent space for generated image detection. LaRE surpasses existing methods in terms of feature extraction efficiency while preserving crucial cues required to differentiate between the real and the fake. To exploit LaRE, we propose an Error-Guided feature REfinement module (EGRE), which can refine the image feature guided by LaRE to enhance the discriminativeness of the feature. Our EGRE utilizes an align-then-refine mechanism, which effectively refines the image feature for generated-image detection from both spatial and channel perspectives. Extensive experiments on the large-scale GenImage benchmark demonstrate the superiority of our " + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "inline_equation", + "content": "\\mathsf{LaRE}^2" + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "text", + "content": ", which surpasses the best SoTA method by up to " + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "inline_equation", + "content": "11.9\\% / 12.1\\%" + }, + { + "bbox": [ + 46, + 256, + 290, + 557 + ], + "type": "text", + "content": " average ACC/AP across 8 different image generators. LaRE also surpasses existing methods in terms of feature extraction cost, delivering an impressive speed enhancement of 8 times." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 128, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 598, + 287, + 696 + ], + "type": "text", + "content": "The rapid advancement of Diffusion Models has heralded a new era in the domain of image generation. Through concerted efforts in refining model architecture [9, 32, 33], optimizing training strategies [15, 27], and enhancing sampling methods [20, 34], contemporary Diffusion Models are now capable of generating images of unprecedented quality, surpassing the boundaries of human imagination. However, this progress raises significant concerns regarding privacy" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 231, + 441, + 322 + ], + "blocks": [ + { + "bbox": [ + 316, + 231, + 441, + 322 + ], + "lines": [ + { + "bbox": [ + 316, + 231, + 441, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 441, + 322 + ], + "type": "image", + "image_path": "3575822d4c9245aa536155349939aa459d9f1aa2db807d6ab3adc8b2834c2e72.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 323, + 441, + 401 + ], + "blocks": [ + { + "bbox": [ + 316, + 323, + 441, + 401 + ], + "lines": [ + { + "bbox": [ + 316, + 323, + 441, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 441, + 401 + ], + "type": "image", + "image_path": "9fb5f683c615aea3d9fd4c20c86a79da90c331fe9b02b2ad7b325aeba85c17a0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "type": "text", + "content": "Figure 1. (a) The comparison of reconstruction-based feature extraction. Existing method [40] chooses to completely reconstruct an image by first gradually adding noise to the image and then denoising it, which involves dozens of sampling steps. Our method can directly calculate noisy images and denoise them with a single sample step. (b) Statistical analysis of the relationship between the single-step reconstruction loss (1000 images are used) and time step. The obvious gap between the two lines indicates that single-step reconstruction can also reflect the differences between real and generated images. (c) Comparison of the cost of per image feature extraction. Our method is " + }, + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "type": "inline_equation", + "content": "8\\mathbf{x}" + }, + { + "bbox": [ + 304, + 411, + 547, + 533 + ], + "type": "text", + "content": " faster than DIRE [40]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 441, + 231, + 538, + 320 + ], + "blocks": [ + { + "bbox": [ + 441, + 231, + 538, + 320 + ], + "lines": [ + { + "bbox": [ + 441, + 231, + 538, + 320 + ], + "spans": [ + { + "bbox": [ + 441, + 231, + 538, + 320 + ], + "type": "image", + "image_path": "3da2ed5f7618302c7e8f23905f0f75e726882bd8b66ba33ad94465c2c68a4d96.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 441, + 321, + 534, + 401 + ], + "blocks": [ + { + "bbox": [ + 441, + 321, + 534, + 401 + ], + "lines": [ + { + "bbox": [ + 441, + 321, + 534, + 401 + ], + "spans": [ + { + "bbox": [ + 441, + 321, + 534, + 401 + ], + "type": "image", + "image_path": "1661650c2bf7aa79a9dccac4a312655962ab80d0cc48285d3683b2a77e96d3dd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 628 + ], + "type": "text", + "content": "and security associated with the generated images [16]. The potential for the dissemination of toxic content and misinformation through these images poses a threat to society and could mislead the public. Consequently, there is an urgent need to develop techniques to detect the images generated by these models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "A flurry of works [6, 7, 21, 31, 40, 41] has been proposed to detect or study the properties of images generated by diffusion models. Recently, DIRE [40] has been proposed to leverage the reconstruction error as a discriminative feature for diffusion-generated image detection. It is based on the assumption that the diffusion-generated images are more easily to be reconstructed by a diffusion" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17006" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "model compared with real images. As a result, DIRE shows great cross-model generalizability towards different diffusion models. Based on DIRE, SeDID [21] is also proposed to harness the inherent distributional disparities between naturally occurring and diffusion-synthesized visuals for diffusion-generated image detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 146, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 146, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 146, + 288, + 337 + ], + "type": "text", + "content": "However, both DIRE and SeDID require multi-step DDIM [34] sampling processes in the feature extraction stage, which results in low efficiency for real-world applications. As shown in Fig 1(c), it takes more than 2 seconds to extract the DIRE feature for an image. Besides, the accumulation loss in the multi-step sampling also introduces uncertainty into the extracted features. Though DDIM [34] provides a deterministic inversion method to transform an image into noise, the reliability of this inversion process is not consistently upheld. In addition, DIRE and SeDID use the reconstruction error as the only feature, ignoring the correspondence between the error and the raw image. We therefore ask: (1) Do we need to completely reconstruct the image to get the discriminative feature? (2) Are there better ways to incorporate reconstruction error into image generation detection?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": "To address the two questions mentioned above, we conducted a series of exploratory experiments and analyses. For the first question, we investigated the training process of the diffusion model. We found that given any time step " + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": ", the forward Markov process has a closed-form solution, allowing us to directly transform " + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": ". Additionally, during the training phase, the model takes " + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 340, + 287, + 686 + ], + "type": "text", + "content": " as inputs and computes the denoising loss, enabling us to obtain the loss through single-step denoising. Therefore, we perform single-step denoising on both 1000 real and 1000 generated images respectively by LDM [32]. The results, as shown in Fig. 1(b), indicate that even with only single-step reconstruction, the loss of real images is consistently greater than that of generated images. This not only demonstrates that the real images are harder to reconstruct than the generated ones but also suggests that the loss from single-step reconstruction can reflect the differences between real and generated images. To address the second question, we visualize the reconstruction loss on the original image, as shown in Fig. 2. Upon observation, we have found that the reconstruction loss is positively correlated with the local information frequency of the original image. In the first image, for instance, the reconstruction loss is relatively lower on the low-frequency background, while it is relatively higher on the high-frequency foreground. The same pattern can be observed in other images as well. The reconstruction loss exhibits spatial correlations with the original image, thereby presenting a potential to serve as a valuable cue for the generated image detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Based on the findings above, we propose a novel Latent REconstruction error guided feature REfinement method" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 366, + 121 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 366, + 121 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 366, + 121 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 366, + 121 + ], + "type": "image", + "image_path": "233c52e08f9df682e8d4a4044191cf611574ac3cde37987589a774542ce4a699.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 122, + 365, + 171 + ], + "blocks": [ + { + "bbox": [ + 316, + 122, + 365, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 122, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 122, + 365, + 171 + ], + "type": "image", + "image_path": "eb4aa63b0bc20a78427ae7ce47b86bb592eb3f9de635b0aae5175fd77bb90e95.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 173, + 365, + 222 + ], + "blocks": [ + { + "bbox": [ + 316, + 173, + 365, + 222 + ], + "lines": [ + { + "bbox": [ + 316, + 173, + 365, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 365, + 222 + ], + "type": "image", + "image_path": "6af487aa9aaa230bcadd9aadf0adc762447c754eca4b8b20d71e4c5e1d83afc7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 224, + 358, + 233 + ], + "lines": [ + { + "bbox": [ + 327, + 224, + 358, + 233 + ], + "spans": [ + { + "bbox": [ + 327, + 224, + 358, + 233 + ], + "type": "text", + "content": "Raw Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 373, + 71, + 423, + 121 + ], + "blocks": [ + { + "bbox": [ + 373, + 71, + 423, + 121 + ], + "lines": [ + { + "bbox": [ + 373, + 71, + 423, + 121 + ], + "spans": [ + { + "bbox": [ + 373, + 71, + 423, + 121 + ], + "type": "image", + "image_path": "235058afdf4a1f7864948925f4f15997e2e3fa62c728639724acba63ef45709f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 373, + 122, + 423, + 171 + ], + "blocks": [ + { + "bbox": [ + 373, + 122, + 423, + 171 + ], + "lines": [ + { + "bbox": [ + 373, + 122, + 423, + 171 + ], + "spans": [ + { + "bbox": [ + 373, + 122, + 423, + 171 + ], + "type": "image", + "image_path": "132f43cbbfbea2acf82a4fb83ac8b179b1a4eee885f875f56bc52b8bbd496db2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 373, + 173, + 423, + 222 + ], + "blocks": [ + { + "bbox": [ + 373, + 173, + 423, + 222 + ], + "lines": [ + { + "bbox": [ + 373, + 173, + 423, + 222 + ], + "spans": [ + { + "bbox": [ + 373, + 173, + 423, + 222 + ], + "type": "image", + "image_path": "55f643c0432b10b831758b6665e7a67ac421684d863c702590a3008f1299a844.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 224, + 417, + 232 + ], + "lines": [ + { + "bbox": [ + 381, + 224, + 417, + 232 + ], + "spans": [ + { + "bbox": [ + 381, + 224, + 417, + 232 + ], + "type": "text", + "content": "Added Noise" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 480, + 121 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 480, + 121 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 480, + 121 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 480, + 121 + ], + "type": "image", + "image_path": "fbc8c680b87283785377240463d8f4a4e2659f7281bbab02c2626da7d883c6fd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 430, + 122, + 480, + 171 + ], + "blocks": [ + { + "bbox": [ + 430, + 122, + 480, + 171 + ], + "lines": [ + { + "bbox": [ + 430, + 122, + 480, + 171 + ], + "spans": [ + { + "bbox": [ + 430, + 122, + 480, + 171 + ], + "type": "image", + "image_path": "b89abf9f7d008f05c0757c6973a537394bc0d323652210b9ebada602887fc9a9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 430, + 173, + 480, + 222 + ], + "blocks": [ + { + "bbox": [ + 430, + 173, + 480, + 222 + ], + "lines": [ + { + "bbox": [ + 430, + 173, + 480, + 222 + ], + "spans": [ + { + "bbox": [ + 430, + 173, + 480, + 222 + ], + "type": "image", + "image_path": "fe99fd27eb74ad90dda188033f002330f04f77bb571b96dc0906f16c1a9cdb4e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 224, + 484, + 232 + ], + "lines": [ + { + "bbox": [ + 430, + 224, + 484, + 232 + ], + "spans": [ + { + "bbox": [ + 430, + 224, + 484, + 232 + ], + "type": "text", + "content": "Reconstruction Loss" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 488, + 71, + 537, + 121 + ], + "blocks": [ + { + "bbox": [ + 488, + 71, + 537, + 121 + ], + "lines": [ + { + "bbox": [ + 488, + 71, + 537, + 121 + ], + "spans": [ + { + "bbox": [ + 488, + 71, + 537, + 121 + ], + "type": "image", + "image_path": "8940fa9db14b53fcce3e2f74e58a794c234c50cb0612c24c460d0d51f3098b1c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 488, + 122, + 537, + 171 + ], + "blocks": [ + { + "bbox": [ + 488, + 122, + 537, + 171 + ], + "lines": [ + { + "bbox": [ + 488, + 122, + 537, + 171 + ], + "spans": [ + { + "bbox": [ + 488, + 122, + 537, + 171 + ], + "type": "image", + "image_path": "b052b26dc04a3531b9cff076889038e9b1d2d7a61b10b8e9ce8b4c8ea37db33f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 488, + 173, + 537, + 222 + ], + "blocks": [ + { + "bbox": [ + 488, + 173, + 537, + 222 + ], + "lines": [ + { + "bbox": [ + 488, + 173, + 537, + 222 + ], + "spans": [ + { + "bbox": [ + 488, + 173, + 537, + 222 + ], + "type": "image", + "image_path": "5aa745475a498ebf86bafa85231ed703c8ca3c3186e9d5b35c9068009d818cb1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 224, + 533, + 233 + ], + "lines": [ + { + "bbox": [ + 496, + 224, + 533, + 233 + ], + "spans": [ + { + "bbox": [ + 496, + 224, + 533, + 233 + ], + "type": "text", + "content": "Image + Loss" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 243, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 288 + ], + "type": "text", + "content": "Figure 2. Visualization of reconstruction loss on raw images (i.e. Image + Loss). Though the randomly sampled noises are added to the whole image, there is a trend that the loss in high-frequency regions is typically greater than that in low-frequency regions." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "(\\mathrm{LaRE}^2)" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": " for diffusion-generated image detection. Our " + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{LaRE}^2" + }, + { + "bbox": [ + 304, + 308, + 545, + 548 + ], + "type": "text", + "content": " consists of two parts, the Latent Reconstruction Error (LaRE) and the Error-guided Feature REfinement module (EGRE). LaRE is a more efficient reconstruction feature, which has two improvements compared with existing methods: (1) LaRE is extracted in a single step of the diffusion reverse process, which is significantly more efficient than completely reconstructing the image through dozens of denoising steps. (2) We conduct the reconstruction in the latent space, which further improves the efficiency. Besides, we find that the reconstructed loss is positively correlated with the local information frequency of the original image. Based on that, we come up with the Error-Guided Feature Refinement module (EGRE). In EGRE, we first align LaRE with the image feature map for better correspondings for feature refinement. Then the aligned LaRE is used to refine the image feature in both spatial and channel perspectives. With the guidance of LaRE, EGRE refines the feature map in both spatial and channel perspectives to better reveal the discriminativeness for detecting the generated images." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "text", + "content": "To evaluate the effectiveness of our method, we conduct extensive experiments on the GenImage [47] benchmark, which comprises 2,681,167 images, segregated into 1,331,167 real and 1,350,000 fake images. The fake images are from 8 different generators. As shown in Fig. 1(c), our LaRE is 8 times faster than existing method [40]. In addition, " + }, + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{LaRE}^2" + }, + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "text", + "content": " achieves a significant performance gain by up to " + }, + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "inline_equation", + "content": "11.9\\% / 12.1\\%" + }, + { + "bbox": [ + 304, + 549, + 545, + 668 + ], + "type": "text", + "content": " ACC/AP compared with the best SoTA. The results demonstrated the superiority of our method, which is both effective and generalizable." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 669, + 499, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 669, + 499, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 669, + 499, + 679 + ], + "type": "text", + "content": "The contributions of our work are three-fold:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 545, + 713 + ], + "type": "text", + "content": "- Novel feature: We are the first to propose the reconstruction error in latent space for generated-image de" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17007" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 120 + ], + "type": "text", + "content": "tection. Compared with the existing method, we remarkably reduce the cost of feature extraction while preserving the essential information required for the detection of diffusion-generated images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 128, + 288, + 255 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 58, + 128, + 287, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 128, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 58, + 128, + 287, + 188 + ], + "type": "text", + "content": "- Novel module: We qualitatively analyze the reason for the effectiveness of the reconstruction loss. Based on that, we come up with a novel module EGRE, which conducts an Error-guided feature refinement to enhance the discriminativeness of image features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 196, + 288, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 196, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 196, + 288, + 255 + ], + "type": "text", + "content": "- Superior performance: Extensive experiments demonstrate the effectiveness of our method. We achieve " + }, + { + "bbox": [ + 58, + 196, + 288, + 255 + ], + "type": "inline_equation", + "content": "11.9\\% / 12.1\\%" + }, + { + "bbox": [ + 58, + 196, + 288, + 255 + ], + "type": "text", + "content": " ACC/AP gain on the large-scale GenImage benchmark, significantly outperforming the SoTA methods." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 266, + 138, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 138, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 138, + 278 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 285, + 145, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 145, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 145, + 297 + ], + "type": "text", + "content": "2.1. Diffusion Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 304, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 288, + 533 + ], + "type": "text", + "content": "Diffusion Models [14, 33, 43] have become new state-of-the-art deep generative models. Inspired by nonequilibrium thermodynamics, Denoising Diffusion Probabilistic Models (DDPM) demonstrate promising generative quality, which inspires a flurry of studies on diffusion models [3, 5, 9, 32, 34, 35, 42, 46]. DDIM [34] generalize DDPMs [14] via a class of non-Markovian diffusion processes to accelerate sampling. ADM [9] firstly obtain better generation quality than GANs [11] on ImageNet [8]. Latent Diffusion Models [32] apply the diffusion process in the latent space to improve the efficiency of diffusion models, and enable diffusion models with text conditioning inputs through the cross-attention conditioning mechanism. There are also works that employ pre-trained diffusion models for various downstream tasks [3, 5, 35, 42, 46]. The rapid advancement of deep generative models has given rise to concerns regarding the potential for malicious utilization of the generated images [16]. Consequently, it is urgent to develop robust techniques to detect generated images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 540, + 224, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 224, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 224, + 552 + ], + "type": "text", + "content": "2.2. Deep Generated Image Detection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 287, + 713 + ], + "type": "text", + "content": "Deep-generated image detection has achieved significant improvements due to the contributions of previous research. Initially, researchers attempted to use hand-crafted features including color cues [23], saturation cues [24], blending [17] artifacts, co-occurrence features [25]. Then CNN is leveraged to detect the generated images [18, 22, 38]. Some works [10, 44] also find that there are obvious visual artifacts in the GAN-generated images and detect these images from the frequency view [28]. Most of the above works are developed and designed for GAN-generated images. With the advance of diffusion models, many methods are also provided for the study of diffusion-generated image detection. Corvi et. al [7, 31] find that state-of-the-art detec" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 550, + 343 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 550, + 343 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 550, + 343 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 550, + 343 + ], + "type": "image", + "image_path": "641a0953ccb1dd5d69ac118439d4bd76084b38f76d41465ef82925fc27e33140.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 352, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 441 + ], + "type": "text", + "content": "Figure 3. Overview of our method. In the first stage, we extract LaRE in the latent space through single-step reconstruction. In the second stage, to exploit LaRE, we propose the Error-guided Feature Refinement Module, which consists of the Error-guided spatial refinement module and the Error-guided Channel Refinement module. From both spatial and channel perspectives, LaRE is used to enhance the discriminativeness of the image feature for generated image detection." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 463, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 619 + ], + "type": "text", + "content": "tors developed for GAN suffer from a severe performance drop when applied to Diffusion-generated images. Corvi et. al [6,7,31] also find that the spectrum artifacts can also rise up in the Diffusion-generated images. Wu et. al [41] formulate the synthetic image detection as an identification problem and achieve generalizable detection through language-guided contrastive learning. DIRE [40] and SeDIE [21] are proposed to leverage reconstruction error from the diffusion model to achieve generalizable diffusion-generated image detection. Different from these methods, we are the first to show that reconstruction error in the latent space can also benefit diffusion-generated image detection, which is more efficient and generalizable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 632, + 390, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 390, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 390, + 645 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Diffusion Models has achieved remarkable image generation performance. Typically, it involves two Markov processes. In the forward process, Gaussian noise is gradually added to the raw image " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " until the image is asymptotically transformed to pure noise, which is defined as:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17008" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 90, + 287, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 90, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 59, + 90, + 287, + 117 + ], + "type": "interline_equation", + "content": "q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = N (\\mathbf {x} _ {t}; \\sqrt {\\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}} \\mathbf {x} _ {t - 1}, (1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}} \\mathbf {I})), \\quad (1)", + "image_path": "51dfcfaa20f01bf5cff5980371c221dd9a7b445e3fd87b2383996d3c3a143625.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": " is the noisy image at step " + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": " is predefined noise schedule. According to the properties of Markov process and Gaussian distribution, we can get " + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 123, + 287, + 172 + ], + "type": "text", + "content": " directly by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 85, + 179, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 179, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 85, + 179, + 287, + 194 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {x} _ {t} \\mid \\mathbf {x} _ {0}\\right) = N \\left(\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I})\\right). \\tag {2}", + "image_path": "e476f354bc54eea354f42c56c836d6a85abfa0bcbc8a0eefe3dbe4d735ef9248.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 201, + 287, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 226 + ], + "type": "text", + "content": "In the reverse process, the noisy image is gradually denoised to get the raw image, which is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 234, + 287, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 234, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 72, + 234, + 287, + 247 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t}) = N (\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} (\\mathbf {x} _ {t}, t), \\boldsymbol {\\Sigma} _ {\\theta} (\\mathbf {x} _ {t}, t)), \\quad (3)", + "image_path": "760b7f0b294af318a053f308b01c9671c44474c8b39e5433f4b7a9391a7e1fd8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": " is typically parameterized by neural networks. During training, a neural network " + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": " is trained to predict the added noise " + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": ", given the noisy image " + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": " and corresponding time step " + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 255, + 287, + 302 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 310, + 287, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 310, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 70, + 310, + 287, + 324 + ], + "type": "interline_equation", + "content": "L _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2}. \\tag {4}", + "image_path": "b75e09b4d63790aed0a878e05a6f77df4b03918bd065183846c6fe1d5652fe66.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "text", + "content": "During training, " + }, + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 47, + 333, + 237, + 345 + ], + "type": "text", + "content": " are randomly sampled." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 355, + 107, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 107, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 107, + 367 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 375, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 287, + 399 + ], + "type": "text", + "content": "In this section, we elaborate on each component of our " + }, + { + "bbox": [ + 47, + 375, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{LaRE}^2" + }, + { + "bbox": [ + 47, + 375, + 287, + 399 + ], + "type": "text", + "content": ", as illustrated in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 406, + 205, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 205, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 205, + 418 + ], + "type": "text", + "content": "4.1. Latent Reconstruction Error" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 425, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 579 + ], + "type": "text", + "content": "Recent methods [40] propose to use reconstruction error as the feature for generated-image detection. Through DDIM [34] inversion, an image is first inversed to noise and then re-generated based on the noise. However, this method has several limitations: (1) It is relatively slow to inverse and re-generate an image. For example, it takes more than 2 seconds to extract DIRE for an image on a Tesla V100 GPU as shown in Fig. 1(c), which limits its realistic applications; (2) The reconstruction of the image is not always reliable. Error could accumulate during the Markov forward and reverse process, making it hard to tell whether the bad reconstruction is from the difference between real and fake images, or the reconstruction itself." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 580, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 651 + ], + "type": "text", + "content": "Here we come up with a new reconstruction feature, named Latent Reconstruction Error (LaRE). LaRE is based on the assumption that if the generated images can be completely reconstructed with relative ease [40], then they can also be more easily reconstructed at every single step of the reverse diffusion process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "text", + "content": "For an image " + }, + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "text", + "content": ", we first get its latent code " + }, + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 47, + 651, + 286, + 676 + ], + "type": "text", + "content": " by VAE. Then we calculate the LaRE by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 682, + 298, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 682, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 682, + 298, + 713 + ], + "type": "interline_equation", + "content": "\\operatorname {L a R E} = L _ {\\theta} \\left(\\mathbf {x} _ {0}, t\\right) = \\mathbb {E} _ {\\epsilon} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t\\right) \\| ^ {2} \\right]. \\tag {5}", + "image_path": "5ca682e77c6801f89d2ee991388d9c5ad0175465928bc610d3ea7a04d5c174f8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "As shown in Fig. 3, Our LaRE takes advantage of two special properties of Diffusion Models. (1) The forward process has a close form. According to Eq. 2, we can get " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": " directly from " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": "; (2) Diffusion Model is trained to denoise a noisy image from any time step " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": ", as shown in Eq. 4. Therefore, given " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 545, + 263 + ], + "type": "text", + "content": ", we can directly get the reconstruction error by only single-step denoising, which is much more efficient than completely reconstructing an image. Compared with complete image reconstruction, our LaRE has several advantages: (1) Only single-step denoising is required to get LaRE, which significantly improves its efficiency. (2) Reconstruction is conducted in the latent space, which is also more efficient. (3) No need for complete inversion and reconstruction of the image, which eliminates the potential accumulation of errors that may occur during multiple inversion and reconstruction steps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 264, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 372 + ], + "type": "text", + "content": "With efficiency, the easier reconstruction assumption still holds for our LaRE. Recall that a diffusion-generated image is more easily reconstructed by diffusion models [40]. So the reconstruction error should be relatively small when completely reconstructing an image through the whole reverse process. That implies the reconstruction error is also relatively small during each step of the reverse process, which is also validated in Fig. 1(b). We leverage the Monte Carlo method to estimate LaRE in Eq. 6:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 381, + 545, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 545, + 422 + ], + "type": "interline_equation", + "content": "\\mathrm {L a R E} = \\frac {1}{e} \\sum_ {i} ^ {e} \\left[ \\| \\epsilon_ {i} - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha}} _ {t} \\mathbf {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {i}, t\\right) \\| ^ {2} \\right], \\tag {6}", + "image_path": "27bc64b3c79ac9013eae1ad7d1b0b002a966b521494f5bdbf71a5f431707c4f0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "inline_equation", + "content": "e, t" + }, + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "text", + "content": " are pre-defined hyperparameters. Specifically, we conduct denoising for " + }, + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 423, + 545, + 458 + ], + "type": "text", + "content": " times and average the denoising loss to extract LaRE." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 467, + 489, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 489, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 489, + 479 + ], + "type": "text", + "content": "4.2. Error-Guided Feature Refinement" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Existing methods exploit the reconstruction error as the only feature to detect the generated image, which ignores the relationships between the reconstruction error and the raw image. To better understand how reconstruction error helps generated-image detection, we visualize the extracted LaRE on the raw image, as shown in Fig. 2. We find several interesting conclusions. 1) The loss norm is proportional to the noise norm. It is easy to understand that the stronger the added noise is, the harder to reconstruct; 2) The loss norm is also proportional to the frequency of local image patches. That the low-frequency patches are easier to reconstruct than the high-frequency ones. Therefore, the reconstruction performance on the high-frequency parts of the image plays an important role in the generated image detection. To apply this valuable information, we come up with a novel Error-guided feature REfinement Module (EGRE). Based on the reconstruction error, our EGRE can better refine the image feature to reveal the discriminative parts for generated image detection." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17009" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 301, + 207 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 301, + 207 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 301, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 301, + 207 + ], + "type": "image", + "image_path": "edb473f6131e88c39bd4fb0cd4a68a5530dc808aeb76c088802f933669124e31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 210, + 264, + 219 + ], + "lines": [ + { + "bbox": [ + 129, + 210, + 264, + 219 + ], + "spans": [ + { + "bbox": [ + 129, + 210, + 264, + 219 + ], + "type": "text", + "content": "(a) Comparison on ACC. Left: baseline, right: ours" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 545, + 208 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 545, + 208 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 545, + 208 + ], + "type": "image", + "image_path": "c84a2b4faa7412863231eafb74113f7a2884de04631bd73d2c46db20cec09d98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 210, + 504, + 219 + ], + "lines": [ + { + "bbox": [ + 374, + 210, + 504, + 219 + ], + "spans": [ + { + "bbox": [ + 374, + 210, + 504, + 219 + ], + "type": "text", + "content": "(b) Comparison on AP. Left: baseline, right: ours" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 251 + ], + "type": "text", + "content": "Figure 4. Results of cross-validation on different training and testing subsets. For each generator, we train a model and test it on all 8 generators. For both DIRE [40] and our method, accuracy (ACC) and average precision(AP) are reported." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 272, + 256, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 256, + 284 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 256, + 284 + ], + "type": "text", + "content": "4.2.1 Error-guided Spatial Refinement Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "content": "We propose the Error-guided Spatial Refinement module (ESR) for feature refinement from the spatial perspective. First, we spatially align LaRE and the image feature map by a simple adaptive average pooling layer, so that they have the same spatial size. Now we have the spatially aligned input feature map " + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{HW \\times C_1}" + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "content": " and LaRE " + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{e} \\in \\mathbb{R}^{HW \\times C_2}" + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}} \\in \\mathbb{R}^{1 \\times C_1}" + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "content": " is the global feature, which is the average of " + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 290, + 287, + 398 + ], + "type": "text", + "content": ". To refine the feature from the spatial perspective, we choose an Error-guided Spatial Attention module (ESA):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 402, + 287, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 402, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 72, + 402, + 287, + 430 + ], + "type": "interline_equation", + "content": "\\operatorname {E S A} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}, \\mathbf {E}) = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} \\mathbf {K} ^ {T}}{\\sqrt {d _ {k}}} + \\mathbf {E}\\right) \\mathbf {V}. \\tag {7}", + "image_path": "35e20f026ca16fce50bde286547b3ca25bbcc1983ec07ed55f595ed8e1fdf84d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 432, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 287, + 479 + ], + "type": "text", + "content": "To further enhance the feature representation, we adopt a multi-head mechanism from Transformer [37]. Therefore, our Multi-Head Error-guided Spatial Attention module (MHESA) is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 52, + 483, + 282, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 483, + 282, + 498 + ], + "spans": [ + { + "bbox": [ + 52, + 483, + 282, + 498 + ], + "type": "inline_equation", + "content": "\\mathrm{MHESA}(\\mathbf{Q},\\mathbf{K},\\mathbf{V},\\mathbf{E}) = \\mathrm{Concat}(\\mathrm{head}_1,\\dots,\\mathrm{head}_h)\\mathbf{W}^O" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 500, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 287, + 525 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 62, + 500, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathrm{head}_i = \\mathrm{ESA}(\\mathbf{Q}\\mathbf{W}_i^Q,\\mathbf{K}\\mathbf{W}_i^K,\\mathbf{V}\\mathbf{W}_i^V,\\mathbf{E}\\mathbf{W}_i^E)" + }, + { + "bbox": [ + 62, + 500, + 287, + 525 + ], + "type": "text", + "content": " (8)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": " is the number of heads, " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^Q\\in \\mathbb{R}^{C_1\\times d}" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^K\\in \\mathbb{R}^{C_1\\times d}" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_i^V\\in \\mathbb{R}^{C_1\\times d}" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_E^Q\\in \\mathbb{R}^{C2\\times d}" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": " are learnable projection matrices, " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": " is the hyperparameter. The spatial refined feature " + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 46, + 525, + 287, + 574 + ], + "type": "text", + "content": " is calculated by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 111, + 580, + 287, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 580, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 111, + 580, + 287, + 593 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {s} = \\operatorname {M H E S A} (\\bar {\\mathbf {x}}, \\mathbf {x}, \\mathbf {x}, \\mathbf {e}), \\tag {9}", + "image_path": "f86e1c60e9d4fd180d5d233f3ebe5a7c56fe952dacca186dd8caebb23be61fc0.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\bar{x}" + }, + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "text", + "content": " is the only query attending to the feature map " + }, + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 597, + 287, + 657 + ], + "type": "text", + "content": ". In this stage, the aligned LaRE is used to re-weight the attention score of scaled dot-product attention. Since LaRE and the feature map are spatially aligned, LaRE can emphasize important information from the spatial perspective." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 670, + 261, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 261, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 261, + 682 + ], + "type": "text", + "content": "4.2.2 Error-guided Channel Refinement Module" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "For channel refinement, we first squeeze both " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}\\in \\mathbb{R}^{1\\times C_1}" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{e}}\\in \\mathbb{R}^{1\\times C_2}" + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": ". Then our channel refinement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 272, + 441, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 441, + 283 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 441, + 283 + ], + "type": "text", + "content": "is achieved by a gate mechanism:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 289, + 545, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 289, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 372, + 289, + 545, + 302 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {c} = \\text {s i g m o i d} (\\bar {\\mathbf {e}} \\mathbf {W}) \\odot \\bar {\\mathbf {x}}, \\tag {10}", + "image_path": "947f429c52e8e7f3e68fd67edaec5525932ac996e6a6f7cc05878450f6b7b6f7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{C1 \\times C2}" + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": " are learnable parameters to align LaRE and the feature map from the channel perspective. Finally, " + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s" + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_c" + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": " and the original global feature " + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_g" + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": " from the last convolution block are concatenated as the final feature " + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{EGRE}}" + }, + { + "bbox": [ + 304, + 307, + 545, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 361, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 361, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 365, + 361, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {E G R E}} = \\operatorname {C o n c a t} \\left(\\mathbf {x} _ {s}, \\mathbf {x} _ {c}, \\mathbf {x} _ {g}\\right). \\tag {11}", + "image_path": "c6cfcc06d9176104f3c4889fd9e517dc9f5729049dedbd67f3562dba0d5a1575.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 379, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{EGRE}}" + }, + { + "bbox": [ + 304, + 379, + 545, + 403 + ], + "type": "text", + "content": " is followed by one FC layer. We train our model with the binary cross entropy loss." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 411, + 387, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 387, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 387, + 426 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 431, + 481, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 481, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 481, + 443 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 450, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 666 + ], + "type": "text", + "content": "We evaluate our proposed method using the GenImage [47] dataset. This dataset comprises a total of 2,681,167 images, divided into 1,331,167 real images and 1,350,000 generated images. The generated images are from 8 different generative models, namely BigGAN [4], GLIDE [26], VQDM [12], Stable Diffusion V1.4&V1.5 [32], ADM [9], Midjourney [1], and Wukong [2]. All of the fake images are generated using the template prompt \"photo of CLS\", where \"CLS\" is replaced by one of the 1000 labels from ImageNet [8]. These images are split into 8 subsets where each subset contains partial real images and all the fake images generated by one of the above generators. We adhere to the official division of the dataset in our research, allocating 2,581,167 images for training and reserving the remaining 100,000 images for validation. Following DIRE [40], we employ Accuracy (ACC) and Average Precision (AP) as our evaluation metrics. More details and results are included in our supplementary materials." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 671, + 438, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 438, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 438, + 684 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "All experiments are conducted using the GenImage dataset. To obtain the LaRE, we employed Stable Diffu" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17010" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 539, + 175 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 175 + ], + "type": "table", + "html": "
MethodsTesting SubsetAvg ACC.(%)
MidjourneySDV1.4SDV1.5ADMGLIDEWukongVQDMBigGAN
CNNSpot [39]58.270.370.257.057.167.756.756.661.7
Spec [45]56.772.472.357.965.470.361.764.365.1
F3Net [29]55.173.173.166.557.872.362.156.564.6
GramNet [19]58.172.872.758.765.371.357.861.264.7
DIRE [40]†65.073.773.761.969.174.363.456.767.2
Ours66.487.387.166.781.385.584.474.079.1
", + "image_path": "9b293d67910dd97398056ea68f3c713d42033ee886b751f2a98f2d01a36649eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 184, + 546, + 208 + ], + "lines": [ + { + "bbox": [ + 46, + 184, + 546, + 208 + ], + "spans": [ + { + "bbox": [ + 46, + 184, + 546, + 208 + ], + "type": "text", + "content": "Table 1. Performance comparisons on GenImage test set. For each data point, eight models are trained on eight generators respectively. Then eight models are tested on the specified test sets and the accuracy scores are averaged. †indicates our reproduction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": "sion V1.5 [32] with a step size of " + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "inline_equation", + "content": "t = 200" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": " and a noise ensemble size of " + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "inline_equation", + "content": "e = 4" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": ". Images are resized to " + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": " for LaRE extraction. We use the prompt \"a photo\" for all images. For both training and testing phases, images are resized to " + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": ". Our model utilizes the CLIP [30] pretrained ResNet50 [13] as the backbone architecture. We set the batch size to 48 and the learning rate to " + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 231, + 289, + 376 + ], + "type": "text", + "content": ". Training is performed on a single Nvidia Tesla V100 GPU. We train eight models on eight subsets respectively, each corresponding to a different generation method. Model selection is based on performance in the validation set, which shares the same generation method as the training set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 381, + 247, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 247, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 247, + 395 + ], + "type": "text", + "content": "5.3. Cross-Generator Image Classification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 400, + 289, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 289, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 289, + 653 + ], + "type": "text", + "content": "We evaluate the performance of our model trained by one of the subsets and test on all eight subsets. The results are shown in Fig. 4. Several conclusions can be made from the results. (1) It is easy to detect generated images from seen generators. Both the DIRE and our proposed method demonstrate superior performance when training and testing images share the same generator, as shown by the diagonal of each image. (2) It is challenging to generalize the model to unseen generators. However, it is relatively easier to generalize the model to similar generator structures than to different structures. For example, a model trained on the BigGAN subset achieves bad performance when tested on other subsets, which are all diffusion-based generators. While the model trained on the SDV1.5 subset achieves relatively better performance on other diffusion-based models. (3) Our model is more robust and generalizable than DIRE. Our model trained on diffusion-based generators achieves better generalizability on both gan-based and diffusion-based generators. This indicates that our " + }, + { + "bbox": [ + 46, + 400, + 289, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{LaRE}^2" + }, + { + "bbox": [ + 46, + 400, + 289, + 653 + ], + "type": "text", + "content": " can better capture discriminative features to detect the generated images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 234, + 672 + ], + "type": "text", + "content": "5.4. Compare with the State of the Arts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we compare our method with several state-of-the-art generated image detection methods, including CNNSpot [39], Spec [45], F3Net [29], GramNet [19]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 231, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 388 + ], + "type": "text", + "content": "and DIRE [40]. In accordance with the GenImage benchmark, eight models are trained on each subset. Subsequently, these models are evaluated on the specified test set, and their accuracy scores are averaged to produce a data point for inclusion in the table. This benchmark presents a formidable challenge, as the overall accuracy is closely tied to the cross-generator generalizability of the proposed techniques. As illustrated in Tab. 1 and Fig. 4, our approach attains state-of-the-art performance across all eight generators. Notably, our overall accuracy exhibits a substantial improvement, surpassing the previous state-of-the-art by " + }, + { + "bbox": [ + 304, + 231, + 547, + 388 + ], + "type": "inline_equation", + "content": "11.9\\% / 12.1\\%" + }, + { + "bbox": [ + 304, + 231, + 547, + 388 + ], + "type": "text", + "content": " ACC/AP. This result demonstrates the superior generalizability of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 395, + 400, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 400, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 400, + 409 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 414, + 547, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 451 + ], + "type": "text", + "content": "In this section, we conduct several ablative studies to quantify the contribution of each design in our model and the influence of the hyperparameters." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 467, + 421, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 467, + 421, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 467, + 421, + 479 + ], + "type": "text", + "content": "5.5.1 Influence of EGRE" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "The contribution of individual components in EGRE to the overall performance of the method is reported in Tab. 2. Starting with the most simple baseline, i.e., training without LaRE (Model A in the table), each component is added building up to the proposed approach. We measure the effect of the Error-guided Spatial Refinement module (ESR) and Error-guided Channel Refinement module (ECR). Specifically, ESR achieves " + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "inline_equation", + "content": "8.6\\% /5.3\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": " ACC/AP gain and ECR achieves " + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "inline_equation", + "content": "5.7\\% /2.4\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": " ACC/AP gain compared with the baseline. Both ESR and ECR significantly improve the performance, demonstrating the effectiveness of the two modules. Relatively, ESR brings a higher improvement in metrics compared with ECR, indicating a greater benefit from spatial feature refinement. This is also attributed to the better alignment of LaRE with the original image at the spatial scale as shown in Fig. 2. By combining ESR and ECR, our approach achieves a further improvement of " + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "inline_equation", + "content": "11.4\\% /7.2\\%" + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": " ACC/AP. By leveraging LaRE for feature refinement at both spatial and channel scales, our model" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17011" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 541, + 157 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 541, + 157 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 541, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 541, + 157 + ], + "type": "table", + "html": "
ModelESRECRCLSMJSDV1.4SDV1.5ADMGLIDEWuKongVQDMBigganAvg
A52.9/94.4100.0/100.0100.0/100.055.0/89.575.5/98.799.7/100.051.8/84.450.4/71.773.1/92.3
B✓61.8/96.9100.0/100.0100.0/100.060.3/95.487.2/99.1100.0/100.086.2/94.658.1/95.081.7/97.6
C✓57.4/95.3100.0/100.0100.0/100.057.2/91.681.2/98.9100.0/100.078.3/89.856.2/82.378.8/94.7
D✓✓64.8/98.2100.0/100.0100.0/100.061.9/98.591.4/99.8100.0/100.094.6/100.063.7/99.784.5/99.5
E✓✓✓66.2/99.3100.0/100.0100.0/100.064.5/99.591.5/99.9100.0/100.097.7/100.067.4/99.885.9/99.8
", + "image_path": "db3720d9df913cce0141b5a6e7ad49df0b6e9042982d6d37a8d38df6e2ee2702.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 60, + 167, + 532, + 178 + ], + "lines": [ + { + "bbox": [ + 60, + 167, + 532, + 178 + ], + "spans": [ + { + "bbox": [ + 60, + 167, + 532, + 178 + ], + "type": "text", + "content": "Table 2. Ablative studies on using different modules or prompts. 'CLS' indicates to use class-specific prompt for LaRE extraction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 55, + 198, + 281, + 358 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 281, + 358 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 281, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 281, + 358 + ], + "type": "image", + "image_path": "d3fe8e64e4dab045164e5fac76e51ffd13806feae5e561e17a3e1a5ec26fe935.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "type": "text", + "content": "Figure 5. Trade-off between detection performance and feature extraction cost. When " + }, + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "type": "inline_equation", + "content": "e = 4" + }, + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "type": "text", + "content": ", the model achieves the best trade-off." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 437 + ], + "type": "text", + "content": "achieves more robust and generalizable image generation detection performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 201, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 201, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 201, + 464 + ], + "type": "text", + "content": "5.5.2 Influence of Noise Ensemble" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "content": "We systematically evaluated the impact of the noise ensemble parameter " + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "content": " on both the accuracy and feature extraction runtime. As shown in Fig. 5. The results demonstrate a consistent increase in accuracy as " + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "content": " increased. However, this improvement in accuracy comes at the cost of increased feature extraction runtime. Upon careful consideration of the trade-off between accuracy and computational cost, we propose " + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "inline_equation", + "content": "e = 4" + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "content": " as the optimal choice. This selection of " + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 46, + 473, + 288, + 594 + ], + "type": "text", + "content": " provides an effective balance, maintaining the performance of our model while ensuring computational efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 170, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 170, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 170, + 623 + ], + "type": "text", + "content": "5.5.3 Influence of Prompts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Because Latent diffusion is a text-conditioned model, a prompt is needed to denoise an image. Here we try to use a class-specific prompt instead of an unchanged prompt. We choose the template \"a photo of CLS\" where \"CLS\" is replaced with the ImageNet label of the image. Results are shown in Tab. 2. When using class-specific prompts (Model E), we can achieve further improvements by " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "1.4\\% /0.3\\%" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 198, + 526, + 357 + ], + "blocks": [ + { + "bbox": [ + 328, + 198, + 526, + 357 + ], + "lines": [ + { + "bbox": [ + 328, + 198, + 526, + 357 + ], + "spans": [ + { + "bbox": [ + 328, + 198, + 526, + 357 + ], + "type": "image", + "image_path": "156bc147d5b4040d8b7ca856d9105748cbab059f271f9c23bfd70324cf885d8d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": "Figure 6. ACC and AP when choosing different " + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": ". The results demonstrate that our model is robust to the choice of " + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 365, + 545, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 590 + ], + "type": "text", + "content": "ACC/AP. This suggests that the category information of images can aid in the extraction of LaRE features. Even in the absence of category information, it is interesting that the accuracy does not decline significantly. We speculate this is because we did not fully convert the images into noise. Even with the added noise, the image still retains some of the original important semantic information, which can assist the model in denoising and thus obtaining good LaRE features. This experiment validates that when the textual information of an image is known, the extracted LaRE features can further enhance the performance. However, we still use the generic prompt in our method, as it already achieves satisfactory results. As for accurately describing images to further enhance performance, we leave that for future research work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 609, + 444, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 444, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 444, + 621 + ], + "type": "text", + "content": "5.5.4 Influence of Sample Step" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Here we conduct the ablative experiment on the sample step " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ". As illustrated in Fig. 1(b), there is a gap between the reconstruction loss of real images and the generated images. When " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "t \\in [100,500]" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", there is a more obvious difference in the reconstruction loss between the two types of images. To verify the impact of " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": " selection on the performance, we conducted experiments with different values of " + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": ", and the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17012" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 282, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 282, + 249 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 282, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 282, + 249 + ], + "type": "image", + "image_path": "a6d1690052959bbffafe8201d40fb1305a5a37f19347e2429c7d281efbd62d96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 251, + 282, + 430 + ], + "blocks": [ + { + "bbox": [ + 53, + 251, + 282, + 430 + ], + "lines": [ + { + "bbox": [ + 53, + 251, + 282, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 282, + 430 + ], + "type": "image", + "image_path": "e11dda25136f9bea9c49038ee1aab790164f1db6a86da6fe8cc3be4a804b3275.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 440, + 287, + 529 + ], + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 529 + ], + "type": "text", + "content": "Figure 7. Visualization of the spatial representation of the baseline model (a) and our model (b). Both models were trained on the SDV1.5 subset. In the case of seen data, both models exhibit good discrimination in the feature space. However, for unseen data, the baseline model shows a significant overlap between real images and fake images (red circle), whereas our model exhibits a smaller overlap (red circle). This further indicates the better generalization capability of our model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "text", + "content": "results are shown in Fig. 6. When " + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "t \\in [150, 300]" + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "text", + "content": ", the model exhibits stable performance, indicating the robustness of our model with respect to " + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "text", + "content": ". However, when " + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 548, + 287, + 597 + ], + "type": "text", + "content": " exceeds this range, the performance of the model tends to decline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 211, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 211, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 211, + 623 + ], + "type": "text", + "content": "5.5.5 Influence of Input Information" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "We conduct experiments to train the detection model with different inputs. The results are shown in Tab. 3. Since our LaRE is a compressed latent representation whose size is only " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 4" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": ". We train a Resnet-20 with LaRE as the only input (Model B). The experimental results show that both the AP and ACC metrics have decreased significantly. We speculate that there are two reasons. Firstly," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 314, + 71, + 539, + 138 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 539, + 138 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 539, + 138 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 539, + 138 + ], + "type": "table", + "html": "
ModelImageLaREMethodAVG. ACCAVG. AP
A✓None73.192.3
B✓None66.268.3
C✓✓Concat76.893.5
D✓✓EGRE84.599.5
", + "image_path": "99ebd29418aea7127d883188572dc1c06fcaea1d64bd7f791868d2a65625c483.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "lines": [ + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 305, + 148, + 545, + 192 + ], + "type": "text", + "content": "Table 3. Results of using different input. Models are trained on the SDv1.5 subset and tested on all the subsets. 'AVG' indicates average scores over 8 subsets. 'Concat' indicates concat LaRE with the image feature map after the last convolution block." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 218, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 398 + ], + "type": "text", + "content": "LaRE is in the latent space and has more information loss compared to the original image. It is not sufficient to serve as the only information for detecting the generated images. Secondly, as shown in the Fig. 2, LaRE mainly reflects the reconstruction quality of the original image at different spatial positions. It is better to leverage LaRE to enhance the representation of the original image features. Therefore, we conducted further experiments. In model C, we only concatenated LaRE with the feature map of the image, which also improves the performance. This validates our hypothesis. Besides, model D achieves significant performance gain compared with Model C. By using EGRE, we can further improve the performance of our model by a significant margin, which also confirms the advantages of EGRE over simple feature concatenation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 408, + 506, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 408, + 506, + 421 + ], + "type": "text", + "content": "5.6. Qualitative Results and Visualizations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 427, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 559 + ], + "type": "text", + "content": "We utilize t-SNE [36] visualization to illustrate feature vectors derived from the final layers of our model and the baseline model, as depicted in Fig. 7. Each model is trained on the SDv1.5 subset of GenImage and evaluated across all eight generative methods. Although both models exhibit good feature discrimination on seen images, they both have limitations when it comes to unseen scenarios. For images from MJ, ADM, and BigGAN, both models show overlaps between real and fake images in the feature space. However, our model demonstrates a smaller proportion of overlap, indicating its better generalization ability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 573, + 378, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 378, + 585 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 378, + 585 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel reconstruction-based diffusion-generated image detection method called LaRE" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". We come up with LaRE, a novel and more efficient reconstruction-based feature by reconstructing the image in the Latent space. Notably, LaRE is 8 times faster compared with existing reconstruction-based methods. By incorporating LaRE with the Error-guided Feature Refinement module (EGRE). Our " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{LaRE}^2" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " achieves superior performance on diffusion-generated image detection, demonstrating state-of-the-art performance." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17013" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "text", + "content": "[1] Midjourney. https://www.midjourney.com/home/, 2022.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 137 + ], + "type": "text", + "content": "[2] Wukong. https://xihe.mindspore.cn/modelzoo/wukong, 2022.5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 139, + 286, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 286, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 286, + 172 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 174, + 286, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 174, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 174, + 286, + 205 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 208, + 286, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 208, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 208, + 286, + 229 + ], + "type": "text", + "content": "[5] Kevin Clark and Priyank Jaini. Text-to-image diffusion models are zero-shot classifiers. 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 232, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 286, + 275 + ], + "type": "text", + "content": "[6] Riccardo Corvi, Davide Cozzolino, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In CVPR, pages 973-982, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 277, + 286, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 286, + 321 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 286, + 321 + ], + "type": "text", + "content": "[7] Riccardo Corvi, Davide Cozzolino, Giada Zingarini, Giovanni Poggi, Koki Nagano, and Luisa Verdoliva. On the detection of synthetic images generated by diffusion models. In ICASSP, pages 1-5. IEEE, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 323, + 286, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 323, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 323, + 286, + 356 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 357, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 286, + 379 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. 34:8780-8794, 2021. 1, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 381, + 286, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 286, + 425 + ], + "type": "text", + "content": "[10] Joel Frank, Thorsten Eisenhofer, Lea Schonherr, Asja Fischer, Dorothea Kolossa, and Thorsten Holz. Leveraging frequency analysis for deep fake image recognition. In ICML, pages 3247-3258. PMLR, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 427, + 286, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 286, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 286, + 470 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. 63(11):139–144, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 472, + 286, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 286, + 516 + ], + "type": "text", + "content": "[12] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, pages 10696-10706, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 518, + 286, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 518, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 518, + 286, + 551 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 553, + 286, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 286, + 574 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 286, + 597 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 286, + 632 + ], + "type": "text", + "content": "[16] Felix Juefei-Xu, Run Wang, Yihao Huang, Qing Guo, Lei Ma, and Yang Liu. Countering malicious deepfakes: Survey, battleground, and horizon. 130(7):1678-1734, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 677 + ], + "type": "text", + "content": "[17] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In CVPR, pages 5001-5010, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[18] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Zhengzhe Liu, Xiaojuan Qi, and Philip HS Torr. Global texture enhancement for fake face detection in the wild. In CVPR, pages 8060-8069, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[20] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. 35:5775-5787, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[21] Ruipeng Ma, Jinhao Duan, Fei Kong, Xiaoshuang Shi, and Kaidi Xu. Exposing the fake: Effective diffusion-generated images detection. 2023. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 228 + ], + "type": "text", + "content": "[22] Francesco Marra, Diego Gragnaniello, Davide Cozzolino, and Luisa Verdoliva. Detection of gan-generated fake images over social networks. In MIPR, pages 384-389. IEEE, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 252 + ], + "type": "text", + "content": "[23] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using color cues. 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[24] Scott McCloskey and Michael Albright. Detecting gang-generated imagery using saturation cues. In ICIP, pages 4584-4588. IEEE, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 341 + ], + "type": "text", + "content": "[25] Lakshmanan Nataraj, Tajuddin Manhar Mohammed, Shivkumar Chandrasekaran, Arjuna Flenner, Jawadul H Bappy, Amit K Roy-Chowdhury, and BS Manjunath. Detecting gan generated fake images using co-occurrence matrices. 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 387 + ], + "type": "text", + "content": "[26] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[27] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171. PMLR, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[28] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 510 + ], + "type": "text", + "content": "[29] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In ECCV, pages 86-103. Springer, 2020. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 566 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 545, + 599 + ], + "type": "text", + "content": "[31] Jonas Ricker, Simon Damm, Thorsten Holz, and Asja Fischer. Towards the detection of diffusion model deepfakes. 2022. 1, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[33] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256-2265. PMLR, 2015. 1, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. 2020. 1, 2, 3, 4" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17014" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 556 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[35] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "text", + "content": "[36] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. 9(11), 2008. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 286, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 286, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 286, + 163 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. 30, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 286, + 206 + ], + "type": "text", + "content": "[38] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "text", + "content": "[39] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. Cnn-generated images are surprisingly easy to spot... for now. In CVPR, pages 8695-8704, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 286 + ], + "type": "text", + "content": "[40] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. Dire for diffusion-generated image detection. 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 288, + 286, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 286, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 286, + 320 + ], + "type": "text", + "content": "[41] Haiwei Wu, Jiantao Zhou, and Shile Zhang. Generalizable synthetic image detection via language-guided contrastive learning. 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 322, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 286, + 365 + ], + "type": "text", + "content": "[42] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 286, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 286, + 410 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 286, + 410 + ], + "type": "text", + "content": "[43] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 286, + 443 + ], + "type": "text", + "content": "[44] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 286, + 477 + ], + "type": "text", + "content": "[45] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and simulating artifacts in gan fake images. In WIFS, pages 1-6. IEEE, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 286, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 286, + 511 + ], + "type": "text", + "content": "[46] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 513, + 286, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 286, + 556 + ], + "type": "text", + "content": "[47] Mingjian Zhu, Hanting Chen, Qiangyu Yan, Xudong Huang, Guanyu Lin, Wei Li, Zhijun Tu, Hailin Hu, Jie Hu, and Yunhe Wang. Genimage: A million-scale benchmark for detecting ai-generated image. 2023. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17015" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_content_list.json b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2b966b81037beab2c8ebd4f4942db8aedb25434b --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_content_list.json @@ -0,0 +1,1697 @@ +[ + { + "type": "text", + "text": "Label Propagation for Zero-shot Classification with Vision-Language Models", + "text_level": 1, + "bbox": [ + 96, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vladan Stojnic1", + "bbox": [ + 222, + 179, + 351, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yannis Kalantidis²", + "bbox": [ + 410, + 180, + 558, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Giorgos Tolias1", + "bbox": [ + 617, + 180, + 743, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ VRG, FEE, Czech Technical University in Prague", + "bbox": [ + 151, + 214, + 560, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Naver Labs Europe", + "bbox": [ + 612, + 215, + 808, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have demonstrated impressive performance on zero-shot classification, i.e. classification when provided merely with a list of class names. In this paper, we tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. We tailor LP to graphs containing both text and image features and further propose an efficient method for performing inductive inference based on a dual solution and a sparsification step. We perform extensive experiments to evaluate the effectiveness of our method on 14 common datasets and show that ZLaP outperforms the latest related works. Code: https://github.com/vladan-stojnic/ZLaP", + "bbox": [ + 75, + 301, + 473, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 569, + 209, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) have demonstrated impressive performance on a variety of computer vision tasks. They are usually trained on large datasets of image-text pairs and contain visual and textual encoders that map to a common feature space. Visual encoders from such models have been shown to produce strong visual representations for perception tasks [31]. Given labeled data from a downstream dataset, one can fine-tune the model or learn classifiers and achieve really high classification accuracy.", + "bbox": [ + 75, + 595, + 468, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Besides using the visual encoder in isolation, the joint text and visual encoder feature space of VLMs enables us to define text-based \"classifiers\", e.g. using the class names as textual prompts. This means that we only need a list of class names to perform zero-shot classification for a target dataset, i.e. without access to any labeled images. Although utilizing priors or devising better textual prompts can improve zero-shot performance [8, 24, 29, 46], here, we are interested in the case where we further have access to unlabeled data. Our goal is to find the best way of utilizing such data for zero-shot classification.", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/842539b56cbdadf340fdccbf2e30b77c762dfba22a17ecd241d32f03c02e10d8.jpg", + "image_caption": [ + "Figure 1. Zero-shot classification performance over 14 datasets using the proposed ZLaP classifier over CLIP [31], as well as over the (concurrent) InMaP [30] approach. Our method offers performance gains for both transductive (left) and inductive (right) inference. Average accuracy over 14 common datasets is reported." + ], + "image_footnote": [], + "bbox": [ + 513, + 271, + 879, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we leverage the inherent structure of the unlabeled data represented by a proximity graph and apply label propagation (LP) between the text-based classifiers and unlabeled images to derive geodesic distances we then use for classification. We tailor LP to VLMs and graphs containing both text and image features, and show that without proper handling of the bimodality, vanilla application of LP fails dramatically. We introduce ZLaP, a novel classification method based on label propagation that can perform both transductive and inductive inference. We perform the former with the standard (primal) solution of LP for classification and devise a more efficient dual solution for the latter. Our method is not only highly effective but also efficient, making LP a more attractive inductive classifier in terms of complexity.", + "bbox": [ + 496, + 609, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We implement our methods using publicly available VLMs as feature encoders, primarily the ResNet and ViT CLIP [31] models, and perform extensive experiments to evaluate the effectiveness of our method on 14 common", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23209", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "datasets. We show that we are able to achieve top performance on two zero-shot inference setups, i.e. inductive and transductive inference. Figure 1 summarizes our gains over 14 datasets on both setups when applying our LP-powered classifiers on top of CLIP, as well as after incorporating the class proxies from the recent InMaP [30] zero-shot approach.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is worth highlighting that ZLaP is a non-parametric method, i.e. it does not involve a learning step. In fact, our approach does not even require access to the VLM model weights and can therefore be used to improve the zero-shot performance of a black-box model even, e.g. provided only via an API. In summary, our contributions are as follows.", + "bbox": [ + 75, + 196, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tailor label propagation to VLMs and zero-shot classification over bi-modal graphs, proposing per modality neighbor search and balancing of contributions.", + "- We propose an efficient way for performing inductive inference with label propagation via a dual solution and through sparsification. This not only improves the test-time efficiency of our method but also performance.", + "- We complement our method with the class proxies presented in concurrent work [30] and achieve state of the art results for zero-shot on 14 common datasets." + ], + "bbox": [ + 76, + 287, + 467, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 460, + 212, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss related works that improve the already impressive zero-shot classification performance of vision-language models [4, 16, 31] even further. This is achieved by devising better distance metrics, utilizing external knowledge to learn more expressive textual prompts, or by leveraging synthetic and unlabeled data.", + "bbox": [ + 75, + 484, + 467, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Improved distance metrics. Zero-shot classification can be improved by devising a better distance metric between image and text representations [8, 46]. CALIP [8] uses a parameter-free attention mechanism and a local patch representation, i.e. instead of global representations, to improve the estimation of class-to-image similarity. CLIP-DN [46] improves the test-time similarity estimation by alignment with the similarity used during contrastive pre-training of VLMs. To achieve this, the method assumes access to unlabeled data from the target distribution. TPT [35] optimizes textual prompts via a consistency objective across test image augmentations. Our method can be considered a part of this line of work, as label propagation is a similarity measure in a geodesic space instead of the Euclidean space.", + "bbox": [ + 75, + 595, + 467, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Improved textual prompts using language models. CLIP [31] uses hand-crafted prompts that are specialized for each domain. Instead of using hand-crafted prompts, generating them with large language models (LLMs) is shown to be promising [24, 29]. VisDesc [24] and", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CuPL [29] query LLMs to generate diverse descriptions of all classes, while WaffleCLIP [32] operates on top of Vis-Desc to systematically analyze which parts of the generated prompts are the most important. Instead of generating class descriptions, CHiLS [26] targets diversifying the set of classes, by generating sub-classes per class, either through an existing class hierarchy or by querying an LLM. It then performs zero-shot classification using sub-classes and linking them to the parent class. We show that methods improving the textual prompts are complementary to our approach.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Synthetic data. Recent methods [9, 38, 43] demonstrate that the use of synthetic data is beneficial for zero-shot classification. CLIP+SYN [9] uses a stable-diffusion-based model to generate synthetic images using class names and uses them to train a linear classifier, initialized by the VLM class representations. SuS-X [38] considers a similar approach, but relies on a non-parametric classifier. CaFO [43] follows the same path, but additionally includes text prompts generated by LLMs.", + "bbox": [ + 496, + 260, + 890, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "External datasets. Besides the use of synthetic data, SuS-X proposes a variant that operates on an extensive unlabeled image dataset (LAION-5B [34]). This dataset encompasses a distribution that is a super set of the target one. The method generates pseudo-labels using the zero-shot approach which are then incorporated within the non-parametric classifier. NeuralPriming [40] additionally assumes that images have captions, which are used to improve the pseudo-labeling.", + "bbox": [ + 496, + 414, + 890, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlabeled images from the target distribution. Another line of research [11, 12, 27, 30] propose operating on unlabeled datasets from the target distribution. The main ingredient of all these methods is the prediction of pseudo-labels for unlabeled examples, that are later used for further processing. UPL [12] optimizes learnable text prompts based on the pseudo-labels. SVL-Adapter [27] first trains a self-supervised model on unlabeled data, and then an adapter module to align its outputs to the pseudolabels. ReCLIP [11] performs transductive label propagation to obtain the pseudo-labels and uses them to fine-tune the VLM visual and textual encoders. In contrast to that, we do not fine-tune the model, which we may not have access to, and efficiently use label propagation for inductive inference too. InMaP [30] is a concurrent work that uses pseudo-labels to update the class representations such that they are now closer to image representations. We show in the experiments that this approach is complementary to ours. In contrast to all those methods, we do not explicitly require pseudo-label prediction, but rather capture interactions between all unlabeled examples through a proximity graph and label propagation.", + "bbox": [ + 496, + 568, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23210", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 169, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first define the task of zero-shot classification with access to unlabeled examples, present label propagation with our contributions and then present the proposed approach for zero-shot classification using unlabeled examples.", + "bbox": [ + 75, + 114, + 468, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem formulation", + "text_level": 1, + "bbox": [ + 76, + 186, + 274, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language models consist of an image encoder $f: \\mathcal{I} \\to \\mathbb{R}^d$ and a text encoder $g: \\mathcal{T} \\to \\mathbb{R}^d$ , where $\\mathcal{I}$ and $\\mathcal{T}$ represent the space of images and text, respectively. We consider the outputs of these encoders to be $\\ell_2$ -normalized.", + "bbox": [ + 76, + 209, + 468, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{C}$ denote a set of known classes with associated class names $\\{l_1,\\dots,l_C\\}$ and $\\mathcal{P} = \\{p_1,\\dots,p_P\\}$ a set of prompt templates. Each prompt is combined with a class name to produce a textual description of the class, i.e. $p_i(l_c)$ for the $i$ -th template used for class $c$ . Class representations $\\mathbf{w}_c = 1 / P\\sum_{i = 1}^{P}g(p_i(l_c))$ are obtained using the VLM. Then, for a test image $u$ , we extract representation $\\mathbf{u} = f(u)$ , and perform zero-shot classification by $\\mathrm{argmax}_c\\mathbf{u}^T\\mathbf{w}_c$ .", + "bbox": [ + 76, + 270, + 468, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We further assume access to a set $\\mathcal{U}$ of $M$ unlabeled images. Let $\\{\\mathbf{u}_1,\\dots ,\\mathbf{u}_M\\}$ denote the representations of the unlabeled images.", + "bbox": [ + 76, + 391, + 468, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we assume no direct access to the VML model weights, i.e. VLM training is neither possible nor desired. This allows to consider the underlying VLM as a black box, possibly only available through an API that generates features. We consider two inference setups:", + "bbox": [ + 76, + 436, + 468, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inductive inference: We consider an inductive setup, where we need to construct a classifier that can operate on new examples. This classifier should take advantage of the unlabeled examples in $\\mathcal{U}$ .", + "bbox": [ + 76, + 513, + 468, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transductive inference: We consider $\\mathcal{U}$ to be the test set, i.e. all test examples are jointly provided in advance. Prediction for test example $\\mathbf{u}_i$ may therefore depend on the representations and predictions of all other test examples. In this transductive setup the models are not required to provide predictions for any example that is not in $\\mathcal{U}$ .", + "bbox": [ + 76, + 573, + 468, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Label propagation (LP)", + "text_level": 1, + "bbox": [ + 76, + 674, + 294, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\{\\mathbf{x}_1, \\dots, \\mathbf{x}_N\\}$ , with $\\mathbf{x}_i \\in \\mathbb{R}^d$ , be a set of features for $N$ examples. Each feature represents a graph node. We construct an adjacency matrix $S \\in \\mathbb{R}^{N \\times N}$ with zero diagonal, and $s_{ij}$ equal to $\\mathbf{x}_i^\\top \\mathbf{x}_j$ if $\\mathbf{x}_j$ is in the $k$ -nearest neighbors of $\\mathbf{x}_i$ (denoted by $\\mathbf{x}_j \\in \\mathrm{kNN}(\\mathbf{x}_i)$ ), and 0 otherwise. We obtain a symmetric adjacency matrix by $\\bar{S} = S + S^\\top$ , and its symmetrically normalized version by $\\hat{S} = D^{-\\frac{1}{2}}\\bar{S}D^{-\\frac{1}{2}}$ , where $D = \\mathrm{diag}(\\bar{S}\\mathbf{1}_N)$ is the degree matrix, and $\\mathbf{1}_N$ is the all-ones $N$ -dimensional vector. We assume the first $C$ examples, and the corresponding nodes, to be labeled among $C$ classes; each class is assigned to a single node1.", + "bbox": [ + 75, + 698, + 468, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transductive inference. Label propagation [44] is originally proposed for the transductive inference setup; we need to predict labels for the unlabeled nodes of the graph. Given the normalized adjacency matrix $\\hat{S}$ , label propagation is an iterative process given by", + "bbox": [ + 500, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {c} ^ {(t + 1)} = \\alpha \\hat {S} \\hat {\\mathbf {y}} _ {c} ^ {(t)} + (1 - \\alpha) \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 176, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "until convergence. Where $\\alpha \\in (0,1)$ is a propagation hyper-parameter, $\\mathbf{y}_c = \\mathbf{e}_c \\in \\{0,1\\}^N$ is a one-hot vector with the non-zero element at index $c$ , and $t$ is the current iteration. Prediction of the label for an unlabeled node $j \\in \\{C + 1,\\dots,N\\}$ is then given by", + "bbox": [ + 496, + 207, + 890, + 286 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {y}} _ {c} (j), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 297, + 890, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{y}}_c(j) = \\mathbf{e}_j^T\\hat{\\mathbf{y}}_c$ is the $j$ -th element of the vector $\\hat{\\mathbf{y}}_c$ . One can show [44] that this iterative solution is equivalent to solving $C$ linear systems", + "bbox": [ + 496, + 325, + 890, + 372 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL \\hat {\\mathbf {y}} _ {c} = \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 385, + 890, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L = I - \\alpha \\tilde{S}$ is the graph Laplacian. These linear systems have a closed-form solution", + "bbox": [ + 496, + 412, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} _ {c} = L ^ {- 1} \\mathbf {y} _ {c} = L _ {\\text {i n v}} \\mathbf {y} _ {c}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 455, + 890, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, this closed-form solution is not practical for large datasets as the inverse graph Laplacian $L_{\\mathrm{inv}}$ is a non-sparse $\\mathbb{R}^{N\\times N}$ matrix. For this reason it is usual [3, 7, 13, 15] to solve (3) using the conjugate-gradient (CG) method, which is known to be faster than running the iterative solution [13]. Using CG is possible because $L$ is positive-definite.", + "bbox": [ + 496, + 486, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observe that (4) simply picks one of the columns of $L_{\\mathrm{inv}}$ . Matrix element $L_{\\mathrm{inv}}(j, c)$ is the confidence of example $j$ belonging to class $c$ . Its values are similarities, after label propagation, between each node pair. It is a type of geodesic similarity that captures the geometry of the feature space as this is indicated by the graph structure. Focusing on a classification task, we are only interested in similarities between an unlabeled example and a class node.", + "bbox": [ + 496, + 577, + 890, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual solution. Herein, we show that solving $C$ linear systems of the form in (4) to obtain predictions for all unlabeled nodes using (2) is equivalent to solving $N - C$ linear systems of form", + "bbox": [ + 496, + 718, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {j} = L ^ {- 1} \\mathbf {e} _ {j} \\quad \\forall j \\in \\{C + 1, \\dots , N \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 790, + 890, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and obtaining the unlabeled node prediction using", + "bbox": [ + 500, + 820, + 830, + 835 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {z}} _ {j} (c). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 848, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The theoretical part described in this section holds for the case of more labeled examples per class too. We consider this specific case for", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "simplicity of the presentation and because it corresponds to the task of zero-shot classification with unlabeled examples.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23211", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This comes from the fact that", + "bbox": [ + 76, + 90, + 274, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {j} (c) = \\mathbf {e} _ {c} ^ {T} L ^ {- 1} \\mathbf {e} _ {j} = \\mathbf {e} _ {j} ^ {T} L ^ {- 1} \\mathbf {e} _ {c} = \\hat {\\mathbf {y}} _ {c} (j). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 111, + 468, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although we present the dual solution using the closed-form (4), the same holds with the CG solution of (3). Using the dual solution (5) is not practical for transductive learning as usually the unlabeled nodes are many more than the labeled ones. However, we show that this dual solution is efficiently used for inductive inference.", + "bbox": [ + 75, + 136, + 468, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed, we can view $L_{\\mathrm{inv}}$ as a pairwise similarity matrix. The confidence of example $j$ belonging to class $c$ , due to symmetry of $L_{\\mathrm{inv}}$ , is equivalently obtained either by $L_{\\mathrm{inv}}(j, c)$ or $L_{\\mathrm{inv}}(c, j)$ . This constitutes, an additional interpretation of the duality in the solution.", + "bbox": [ + 76, + 227, + 468, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inductive inference. Test examples now come individually and are not known during graph construction. A possible way to perform inductive inference is by adding the new node to the graph, which is expensive for a test-time operation as $\\hat{S}$ would have to be updated for each new test example. Instead, inspired by [13] that uses LP for retrieval, we construct indicator vector $\\mathbf{y}_{\\mathbf{x}} \\in \\mathbb{R}^{N}$ for test example $\\mathbf{x}$ such that", + "bbox": [ + 75, + 320, + 468, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {\\mathbf {x}} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {x} ^ {T} \\mathbf {x} _ {j}, & \\text {i f} \\mathbf {x} _ {j} \\in \\mathrm {k N N} (\\mathbf {x}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 444, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we solve linear system", + "bbox": [ + 76, + 489, + 272, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {\\mathbf {x}} = L ^ {- 1} \\mathbf {y} _ {\\mathbf {x}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 511, + 468, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as in the dual formulation (5) and get a prediction with $\\hat{y}_{\\mathbf{x}} = \\operatorname{argmax}_c \\hat{\\mathbf{z}}_{\\mathbf{x}}(c)$ . With the usual formulation of label propagation in (4), $C$ linear systems need to be solved to get prediction for a single test example. The dual formulation allows us to do it by solving only a single linear system.", + "bbox": [ + 75, + 536, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fast inductive inference with sparsification. We further introduce an additional off-line step, where we solve (4), get $\\hat{\\mathbf{y}}_c$ for all $c\\in \\{1,\\dots,C\\}$ , and store them in a matrix $\\hat{Y} = [\\hat{\\mathbf{y}}_1;\\dots;\\hat{\\mathbf{y}}_c]\\in \\mathbb{R}^{N\\times C}$ . Then, the solution for a test example is equivalent to a weighted sum of rows of $\\hat{Y}$ [1], which is a byproduct of using the indicator vector (8) for representing a test example. Its prediction is given by $\\hat{\\mathbf{z}}_{\\mathbf{x}} = \\mathbf{y}_{\\mathbf{x}}^{T}\\hat{Y}$ , and is equivalent to that obtained via (9). However, storing the whole $\\hat{Y}$ can be expensive for very large values of $N$ and $C$ . We propose to sparsify $\\hat{Y}$ by keeping only the largest values in each row, column, or over the whole matrix.", + "bbox": [ + 75, + 628, + 468, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that [14] proposes a low-rank decomposition of the inverse graph Laplacian for the task of retrieval. Our solution is tailored to zero-shot classification, and we choose to obtain and sparsify the first $C$ rows of $L_{\\mathrm{inv}}$ instead of approximating the whole matrix. Additionally, our solution requires one (sparse) vector to matrix multiplications at test-time instead of two.", + "bbox": [ + 75, + 795, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. LP for zero-shot VLM classification", + "text_level": 1, + "bbox": [ + 500, + 90, + 808, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are given a set of classes $\\mathcal{C}$ with extracted VLM representations $\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C\\}$ and a set of unlabeled images $\\mathcal{U}$ with extracted VLM representations $\\{\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}$ . We use them as nodes $\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C,\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}$ of the graph for label propagation. Nodes of class representations (text nodes) are labeled and image nodes unlabeled.", + "bbox": [ + 496, + 114, + 890, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To construct the adjacency matrix $S$ , we need to perform the $k$ -nearest neighbor search between nodes. However, it is known that there exists a large modality gap between image and text representations coming from VLMs [21, 38, 47]. The respective similarity distributions for CLIP are shown in Figure 3a. This modality gap makes standard kNN search between nodes not useful for label propagation; image nodes mostly get connected to image nodes, and text nodes mostly get connected to text nodes. As a consequence, few edges exist between labeled and unlabeled nodes.", + "bbox": [ + 496, + 208, + 890, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this problem, we perform the kNN search separately for connecting image nodes to image nodes and for connecting image nodes and text nodes. We do not perform the search using text nodes as queries, i.e. text nodes get linked only if they appear in the kNN list of an image. This way we also avoid linking text nodes with each other, which is beneficial as each of them is labeled to a different class. Formally, the values of the adjacency matrix are", + "bbox": [ + 496, + 359, + 890, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {u} _ {i} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u} _ {i}) \\\\ \\mathbf {u} _ {i} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u} _ {i}) \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 494, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{kNN_u}$ and $\\mathrm{kNN_w}$ denote that the search is performed within the image or class features only, respectively.", + "bbox": [ + 500, + 565, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, during inductive inference for image $\\mathbf{u}$ , we perform the kNN search in a similar way to construct indicator vector $\\mathbf{y_u}$ whose elements are given by", + "bbox": [ + 496, + 595, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {\\mathbf {u}} (i) = \\left\\{ \\begin{array}{l l} \\mathbf {u} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u}) \\\\ \\mathbf {u} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 655, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the two types of edges, i.e. image-to-image and image-to-text, we use power function $h(v) = v^{\\gamma}$ to transform the image-to-text (cross-modal) similarities. This way we effectively balance their contribution in the graph and the indicator vector. To that end, we use $h(\\mathbf{u}_i^T\\mathbf{w}_j)$ and $h(\\mathbf{u}^T\\mathbf{w}_j)$ in (10) and (11), respectively, instead of $\\mathbf{u}_i^T\\mathbf{w}_j$ and $\\mathbf{u}^T\\mathbf{w}_j$ .", + "bbox": [ + 496, + 732, + 890, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We refer to the proposed method described above as Zero-shot classification with Label Propagation (ZLaP). We further denote the variant of our method after sparsifying the $\\hat{Y}$ matrix for inductive inference as ZLaP*.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23212", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4bd6274117ce7f7e4af3e2e04bbac872ce76af67fcf387ef6b4f212267476378.jpg", + "image_caption": [ + "Figure 2. t-SNE visualization for the original CLIP features (left) and our geodesic similarity (right). The former is estimated with the features as input, while the latter with the $L_{\\mathrm{inv}}$ used as a pairwise similarity matrix. $\\star$ : class representation, $\\bullet$ : image representation. Figure generated for five random classes from the CUB dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 410, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b55e305adcc9efe1fa3a9873ee053366b8234b3efef992554aa7775906de3fd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 95, + 880, + 266 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1dbfac7a1d4b7eeb79c4225058341baba1894e929048f46f7d06ff05a5eee482.jpg", + "image_caption": [ + "(a) Using text prompts", + "Figure 3. Similarity distributions among features of the same or different modality, using 7 textual templates [38] (left) or the InMaP proxies (right) as class representations." + ], + "image_footnote": [], + "bbox": [ + 88, + 352, + 267, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d905689a541494f1b8041d6ce8d311881458b283578130593244ceea09e925a7.jpg", + "image_caption": [ + "(b) Using proxies from InMaP [30]" + ], + "image_footnote": [], + "bbox": [ + 282, + 352, + 460, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$t$ -SNE visualization of the bi-modal space. In Figure 2 we visualize the bi-modal feature space for CLIP features using $t$ -SNE [39] in two cases, i.e. the Euclidean case and using geodesic similarities obtained by $L_{\\mathrm{inv}}$ , i.e. after label propagation. When using Euclidean affinities (left), we see that due to the large differences in the similarity distributions (text-text, image-to-image, and text-to-image, as shown in Figure 3a) all class representations (stars) are clustered together far from the image nodes. However, using the geodesic affinities from $L_{\\mathrm{inv}}$ (right) we see that class representations are more spread.", + "bbox": [ + 75, + 574, + 468, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first present the datasets we use, our experimental setup and competing methods. We then present component analysis for ZLaP and results for transductive and inductive zero-shot classification on 14 datasets.", + "bbox": [ + 75, + 777, + 468, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 847, + 181, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed method on 14 diverse image classification datasets: ImageNet ILSVRC2012 [33],", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Describable Textures Dataset (DTD) [5], EuroSAT [10], FGVC-Aircraft [23], Oxford Flowers 102 [25], Food101 [2], Oxford-IIIT Pet [28], SUN397 [42], Stanford Cars [17], Caltech101 [6], UCF101 [36], CIFAR10 [18], CIFAR100 [18], CUB-200-2011 [41]. For the first 11 datasets we borrow the train and test splits from CoOp[45]. We use the official training and test splits for CIFAR10, CIFAR100 and CUB-200-2011.", + "bbox": [ + 496, + 351, + 893, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental setup", + "text_level": 1, + "bbox": [ + 498, + 487, + 687, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the transductive (inductive) inference setup, unlabeled nodes in the graph are the test (train) images. We always measure classification accuracy over the test images.", + "bbox": [ + 496, + 512, + 893, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VLMs and textual prompts. We report results using the publicly available ResNet50 and ViT-B/16 CLIP [31] models. We adopt the 7 templates from SuS-X [38] as class prompts for all results apart from Table 3 where we utilize the LLM generated prompts from [29].", + "bbox": [ + 496, + 585, + 892, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared methods. Our baseline is zero-shot recognition with CLIP [31] using text encoder features as class representations. TPT [35] is based on test-time prompt tuning such that different image augmentations produce consistent predictions. The aforementioned methods do not exploit unlabeled data; their performance is therefore unchanged in both inference setups. CLIP-DN [46] normalizes feature distributions during test-time and assumes access to the mean feature vector of the target distribution. In the transductive (inductive) setup the mean vector is estimated on the test (training) set. InMaP [30] is a concurrent work that extracts updated class representations using pseudo-labels on the unlabeled set. In the transductive (inductive) setup the learning is performed on the test (training) images.", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23213", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We reproduce results for $\\mathrm{CLIP}^2$ , $\\mathrm{CLIP - DN^3}$ , and $\\mathrm{InMaP^4}$ using their public implementations. For TPT [35] we report the numbers provided in [30]. We run InMaP using a single set of hyper-parameters for all 14 datasets, i.e. the default values reported in the official implementation4. We also fix the values of $k$ , $\\gamma$ , and $\\alpha$ for ZLaP across all datasets to 5, 5.0, and 0.3, respectively, for CLIP, and 10, 3.0, and 0.3, respectively, for InMaP.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ZLaP variants. We refer to ZLaP using text class representations as $\\mathbf{CLIP} + \\mathbf{ZLaP}$ . Since InMaP is complementary to our work, we further evaluate the performance of ZLaP when $\\{\\mathbf{w}_1,\\dots ,\\mathbf{w}_C\\}$ are the InMaP proxies. We refer to this as InMaP + ZLaP in the results. We refer to ZLaP with a sparse $\\hat{Y}$ for inductive inference as ZLaP*.", + "bbox": [ + 75, + 229, + 468, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Components of ZLaP", + "text_level": 1, + "bbox": [ + 76, + 330, + 277, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bi-modal graph adjustments. In Table 1 we show the importance of two design choices to adapt LP to bi-modal graphs, i.e. separating the nearest neighbor search across modalities using (10) and (11), and transforming cross-modal similarities using a power function $h(\\cdot)$ . We see that separate search is crucial; without it LP is not effective at all. The power function gives an extra boost in both setups, especially in the case of transductive inference. In Table 2 we report the percentage of images that are connected to their groundtruth class nodes within a path of length $n$ , with and without our adjustments. We see that for any such paths to exist for the case without adjustments, $k$ needs to be extremely high. With adjustments, $k = 5$ is enough for $71.4\\%$ of the nodes to be connected to the correct class nodes.", + "bbox": [ + 75, + 353, + 472, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparsifying matrix $\\hat{Y}$ for inductive inference. We explore three ways of approximating $\\hat{Y}$ by sparsification, i.e. wither keeping only the largest $\\xi$ columns per row, the largest $\\xi$ rows per column, or the largest $\\xi$ elements of the whole matrix. In all cases, the rest of the elements are set to zero. In Figure 4 we show the influence that these three variants have on performance. Not only these variants speed-up inference, but we can also see improvements in performance when sparsification percentage is high, i.e. low percentage of non-zero elements. We attribute this to the fact that less confident predictions in $\\hat{Y}$ , many of them erroneous, are now set to zero. Although the best variant to choose seems to vary per dataset, we found that keeping the top element per row performs well across different datasets. We therefore use the $\\xi = 1$ top element per row for our experiments. This amounts to different percentages of sparsity per dataset; we are keeping approximately $2.3\\%$ on average", + "bbox": [ + 75, + 583, + 472, + 843 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/23d092a06b029be3cf6c707f74583305bf2c57edb543ac4dd2b1cace93d9dc96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Eq.(10)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X50.232.341.6
✓✓61.841.952.1
", + "bbox": [ + 504, + 88, + 883, + 174 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/379338db492160396706db8e8fc670427df5f9ed76290333d4986a1cd1d2bb4d.jpg", + "table_caption": [ + "(a) Transductive inference" + ], + "table_footnote": [], + "table_body": "
Eq.(10)-(11)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X60.842.449.6
✓✓62.242.849.7
", + "bbox": [ + 503, + 205, + 883, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a19322ce1549f7758a00a068a84fc8fcee522e4b76b48a1d357b7614458e1b20.jpg", + "table_caption": [ + "(b) Inductive inference", + "Table 1. Adjusting LP to bi-modal graphs. Impact of using separate kNN search for constructing the graph (Eq.(10)) or the indicator vector (Eq.(11)), as well as power function $h(\\cdot)$ for balancing the contributions of two types of edges in the graph." + ], + "table_footnote": [], + "table_body": "
kJoint kNN searchSeparate kNN search
n=1n=2n=3n=1n=2n=3
50.00.00.071.485.1100.0
100.00.00.082.995.4100.0
10040.1100.0100.0100.0100.0100.0
", + "bbox": [ + 513, + 387, + 880, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Impact of the separate kNN search on the shortest paths between image nodes and the text node of their class. We report the percentage of images whose shortest path to the text node of their ground-truth class has length equal to or less than $n$ . Left: the vanilla approach. Right: our separate kNN search using Eq. (10). Analysis on DTD for the transductive setup.", + "bbox": [ + 498, + 500, + 890, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "across all datasets. Regarding the inference speed-up, the primal solution takes $\\sim 2.6$ sec per image, the dual takes $\\sim 4.4$ ms, while the sparsified approach takes $\\sim 0.6$ ms, measured on ImageNet dataset.", + "bbox": [ + 498, + 608, + 890, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the class proxies from InMaP. We observe that many class-to-image similarities (e.g. $\\mathbf{u}_i^\\top \\mathbf{w}_j$ in (10)) become negative on some datasets when using ZLaP with InMaP proxies (see Figure 3b). We therefore perform minmax normalization in range [0,1] after constructing adjacency matrix $S$ or the indicator vector, for the transductive and inductive inference setups.", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results", + "text_level": 1, + "bbox": [ + 500, + 801, + 591, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transductive inference. We present results for transductive zero-shot classification in Figure 5. ZLaP improves the zero-shot performance of CLIP significantly on all datasets. It also outperforms the recent TPT and CLIP-DN approaches on the vast majority of cases, with large gains", + "bbox": [ + 496, + 824, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/OpenAI/CLIP \n $^{3}$ https://github.com/fengyuli-dev/distribution-normalization \n $^{4}$ https://github.com/idstcv/InMaP", + "bbox": [ + 76, + 848, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23214", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46306664df705b1ebca34c6625c2b182e9f4c706be33116713144b5162cdc9a2.jpg", + "image_caption": [ + "Figure 4. Sparcifying matrix $\\hat{Y}$ for inductive CLIP+ZLaP: effect of maintaining only the top elements per row/column/matrix." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 377, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a07735124b0c8ecbe47efd9a854d6781b9a92fd09dcbd7e8cb522ef2b281891e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 89, + 612, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e9157546dbf2f27e9da207d3371dedf55ee66739735207d69dfc74d161e1289f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 879, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be8905d402759a5c56f8b2941c49bca0d5b161c621c1e0a07eacbdccaeab26b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 313, + 465, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0521b9b73f6f88a5d39006d3ab6f9e2ec245249dd2214438c6e81ef88104fdc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 311, + 861, + 486 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bf838dc5c502bdef49613c3110f958819fa478ecdcb8972ce41a1563dc5569c9.jpg", + "image_caption": [ + "Figure 5. Zero-shot classification accuracy averaged over 14 datasets for the transductive (top) and inductive (bottom) setups. Results per dataset are reported in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 96, + 491, + 465, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/201bd18f35e357ce7a37bae4c76f7e75e589d004f4c24c172956dcde0f1aae77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 491, + 861, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in average accuracy. Compared to InMaP, ZLaP offers lower accuracy on average. However, by incorporating InMaP's class representations to our graph, we can improve our results even further and outperform all other methods, for an improvement of approximately $+5\\%$ over CLIP with both backbones.", + "bbox": [ + 75, + 751, + 472, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inductive inference. We report results for the inductive inference setup in Figure 5. ZLaP achieves a noticeable im", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "provements over the CLIP baseline in this setup as well. Gains are more prominent for the case of ZLaP using the InMaP proxies, where gains over CLIP are $+4.4\\%$ and $+4.9\\%$ for the two backbones. We also observe that, although InMaP slightly outperforms ZLaP when used over CLIP, the combination of the two achieves state-of-the-art performance in this case as well. We further see that $\\mathrm{ZLaP^{*}}$ retains the state-of-the-art performance of our method, while sparsifying $\\hat{Y}$ offers significant speed-up at inference time.", + "bbox": [ + 496, + 751, + 893, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23215", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3ca31eab13e5820cbcfe6218b688089d1afc5fe287625f570dbd495fe049a46b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TransductiveInductive
Results with RN50
CLIP63.063.0
+ ZLaP64.664.2
InMaP64.864.6
+ ZLaP65.865.0
Results with ViT-B-16
CLIP71.971.9
+ ZLaP72.673.3
InMaP73.974.0
+ ZLaP74.874.2
", + "bbox": [ + 78, + 89, + 472, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leveraging LLM generated prompts. In Table 3 we report average zero-shot classification accuracy for ZLaP using the prompts recently proposed in CuPL [29]. These are prompts generated by LMMs that are available on the CuPL Github page5 for 12 of the datasets we use (all datasets besides CUB and Eurosat). ZLaP improves zero-shot performance in this case as well, for both the transductive and inductive setups. This verifies that our method is complementary to improved prompt engineering.", + "bbox": [ + 75, + 386, + 470, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-label classification. We apply ZLaP for multi-label classification on the MS-COCO [22] dataset. ZLaP improves the zero-shot performance of CLIP by $+6.0\\%$ mAP (56.8% vs. 50.8%) for inductive inference without any modification of the approach or its hyper-parameters.", + "bbox": [ + 75, + 553, + 468, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Web-crawled unlabeled images. All previous experiments use unlabeled images that come from the target distribution, i.e. they are known to depict one of the classes of interest but their labels are discarded. To see the impact of ZLaP in a more realistic setup using web-crawled images we rely on LAION-400M [34] composed by image-caption pairs. We construct the set of unlabeled images with 10,000 images per class that are chosen either randomly, or based on proximity of their image or text features to the class representation. Random selection fails, but the other two options provide some improvement compared to CLIP, with the caption-based neighbors being a bit better. The complete set of results is presented in the supplementary material.", + "bbox": [ + 75, + 659, + 470, + 869 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b621872e8c517a2962b13bc6e9f8982b1e06670762abf9953627c9ff67516488.jpg", + "table_caption": [ + "Table 3. Zero-shot classification using prompts generated by LLMs [29]. We report average accuracy on 12 datasets using prompts from CuPL [29] together with our 7 standard prompts. Results per dataset are reported in the supplementary material." + ], + "table_footnote": [], + "table_body": "
TransductiveInductive
BLIP [20]54.654.6
+ ZLaP59.657.9
ALBEF [19]36.036.0
+ ZLaP41.246.8
EVA-CLIP-8B [37]83.683.6
+ ZLaP84.684.5
EVA-CLIP-18B [37]83.983.9
+ ZLaP84.884.7
", + "bbox": [ + 535, + 88, + 857, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Accuracy on ImageNet using different VLMs.", + "bbox": [ + 524, + 258, + 867, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different VLMs We use CLIP as the VLM of choice throughout our experiments. In Table 4, we present results when ZLaP is applied on top of four recent VLMs, namely BLIP [20], ALBEF [19], and two versions of EVA-CLIP [37]. We use the implementations of BLIP and ALBEF that are available in the LAVIS library6, while for EVA-CLIP we use implementation from the official Github repository7. ZLaP improves the results of all four different VLMs in both transductive and inductive setups.", + "bbox": [ + 496, + 299, + 893, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 449, + 627, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Label propagation is an intuitive way of encoding the global structure of unlabeled data into geodesic distances over a locally Euclidean space. In this paper, we show that this method can be successfully tailored to both transductive and inductive zero-shot classification with vision-language models, and achieve state-of-the-art performance on both setups. To that end, we show that it is highly important to take proper care of the peculiarities of the bi-modal nature of the task during graph construction. We further carefully design an efficient variant of label propagation for the inductive inference case, that may enable label propagation to be applied to other tasks beyond zero-shot classification.", + "bbox": [ + 496, + 474, + 892, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision-language models trained on billion-scale datasets are redefining computer vision research. The proposed ZLaP is a training-free approach able to improve the generalization performance of black-box VLMs using only unlabeled data, for an annotation-free, text-based and open-world classification paradigm that will inevitably be ubiquitous in the near future.", + "bbox": [ + 496, + 656, + 893, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the Junior Star GACR GM 21-28830M and the Czech Technical University in Prague grant No. SGS23/173/OHK3/3T/13. We thank Ahmet Iscen for many helpful comments.", + "bbox": [ + 496, + 780, + 893, + 842 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "5https://github.com/sarahpratt/CuPL", + "bbox": [ + 99, + 886, + 379, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://github.com/salesforce/LAVIS", + "bbox": [ + 522, + 862, + 808, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "https://github.com/baaivision/EVA/tree/master/EVA-CLIP-18B", + "bbox": [ + 500, + 876, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23216", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. Label propagation and quadratic criterion. In Semi-Supervised Learning, pages 192-216. The MIT Press, 2006. 4", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5", + "[3] Siddhartha Chandra and Iasonas Kokkinos. Fast, exact and multi-scale inference for semantic image segmentation with deep Gaussian CRFs. In ECCV, 2016. 3", + "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2", + "[5] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In CVPR, 2014. 5", + "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, 2004. 5", + "[7] Leo Grady. Random walks for image segmentation. pami, 28(11):1768-1783, 2006. 3", + "[8] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. CALIP: zero-shot enhancement of CLIP with parameter-free attention. In AAAI, 2023. 1, 2", + "[9] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip H. S. Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In ICLR, 2023. 2", + "[10] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 5", + "[11] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, Cheng-Hao Kuo, and Ram Nevatia. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2", + "[12] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2", + "[13] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Teddy Furon, and Ondrej Chum. Efficient diffusion on region manifolds: Recovering small objects with compact CNN representations. In CVPR, 2017. 3, 4", + "[14] Ahmet Iscen, Yannis Avrithis, Giorgos Tolias, Teddy Furon, and Ondrej Chum. Fast spectral ranking for similarity search. In CVPR, 2018. 4", + "[15] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In CVPR, 2019. 3", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2", + "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, 2013. 5", + "[18] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5", + "[19] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 8", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 8", + "[21] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y. Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4", + "[22] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 8", + "[23] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[24] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1, 2", + "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5", + "[26] Zachary Novack, Julian J. McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2", + "[27] Omiros Pantazis, Gabriel J. Brostow, Kate E. Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. In BMVC, 2022. 2", + "[28] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5", + "[29] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 1, 2, 5, 8", + "[30] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In NeurIPS, 2023. 1, 2, 5, 6", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5", + "[32] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23217", + "bbox": [ + 478, + 943, + 517, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5", + "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 8", + "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 5, 6", + "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5", + "[37] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP18B: scaling CLIP to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 8", + "[38] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 4, 5", + "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 5", + "[40] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. Neural priming for sample-efficient adaptation. In NeurIPS, 2023. 2", + "[41] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 5", + "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5", + "[43] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In CVPR, 2023. 2", + "[44] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In NIPS, 2003. 3", + "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 5", + "[46] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Test-time distribution normalization for contrastively learned vision-language models. In NeurlPS, 2023. 1, 2, 5", + "[47] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 4" + ], + "bbox": [ + 78, + 92, + 467, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23218", + "bbox": [ + 478, + 944, + 517, + 954 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_model.json b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97a8dadffe53f56c5542b53646cb7fbf021214ba --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Label Propagation for Zero-shot Classification with Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.18, + 0.352, + 0.2 + ], + "angle": 0, + "content": "Vladan Stojnic1" + }, + { + "type": "text", + "bbox": [ + 0.411, + 0.181, + 0.56, + 0.198 + ], + "angle": 0, + "content": "Yannis Kalantidis²" + }, + { + "type": "text", + "bbox": [ + 0.619, + 0.181, + 0.745, + 0.199 + ], + "angle": 0, + "content": "Giorgos Tolias1" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.215, + 0.562, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) VRG, FEE, Czech Technical University in Prague" + }, + { + "type": "text", + "bbox": [ + 0.614, + 0.216, + 0.81, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\) Naver Labs Europe" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.53 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have demonstrated impressive performance on zero-shot classification, i.e. classification when provided merely with a list of class names. In this paper, we tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. We tailor LP to graphs containing both text and image features and further propose an efficient method for performing inductive inference based on a dual solution and a sparsification step. We perform extensive experiments to evaluate the effectiveness of our method on 14 common datasets and show that ZLaP outperforms the latest related works. Code: https://github.com/vladan-stojnic/ZLaP" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.57, + 0.21, + 0.586 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) have demonstrated impressive performance on a variety of computer vision tasks. They are usually trained on large datasets of image-text pairs and contain visual and textual encoders that map to a common feature space. Visual encoders from such models have been shown to produce strong visual representations for perception tasks [31]. Given labeled data from a downstream dataset, one can fine-tune the model or learn classifiers and achieve really high classification accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Besides using the visual encoder in isolation, the joint text and visual encoder feature space of VLMs enables us to define text-based \"classifiers\", e.g. using the class names as textual prompts. This means that we only need a list of class names to perform zero-shot classification for a target dataset, i.e. without access to any labeled images. Although utilizing priors or devising better textual prompts can improve zero-shot performance [8, 24, 29, 46], here, we are interested in the case where we further have access to unlabeled data. Our goal is to find the best way of utilizing such data for zero-shot classification." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.272, + 0.88, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.503, + 0.895, + 0.574 + ], + "angle": 0, + "content": "Figure 1. Zero-shot classification performance over 14 datasets using the proposed ZLaP classifier over CLIP [31], as well as over the (concurrent) InMaP [30] approach. Our method offers performance gains for both transductive (left) and inductive (right) inference. Average accuracy over 14 common datasets is reported." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.837 + ], + "angle": 0, + "content": "In this paper, we leverage the inherent structure of the unlabeled data represented by a proximity graph and apply label propagation (LP) between the text-based classifiers and unlabeled images to derive geodesic distances we then use for classification. We tailor LP to VLMs and graphs containing both text and image features, and show that without proper handling of the bimodality, vanilla application of LP fails dramatically. We introduce ZLaP, a novel classification method based on label propagation that can perform both transductive and inductive inference. We perform the former with the standard (primal) solution of LP for classification and devise a more efficient dual solution for the latter. Our method is not only highly effective but also efficient, making LP a more attractive inductive classifier in terms of complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We implement our methods using publicly available VLMs as feature encoders, primarily the ResNet and ViT CLIP [31] models, and perform extensive experiments to evaluate the effectiveness of our method on 14 common" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "datasets. We show that we are able to achieve top performance on two zero-shot inference setups, i.e. inductive and transductive inference. Figure 1 summarizes our gains over 14 datasets on both setups when applying our LP-powered classifiers on top of CLIP, as well as after incorporating the class proxies from the recent InMaP [30] zero-shot approach." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.197, + 0.468, + 0.287 + ], + "angle": 0, + "content": "It is worth highlighting that ZLaP is a non-parametric method, i.e. it does not involve a learning step. In fact, our approach does not even require access to the VLM model weights and can therefore be used to improve the zero-shot performance of a black-box model even, e.g. provided only via an API. In summary, our contributions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.288, + 0.468, + 0.332 + ], + "angle": 0, + "content": "- We tailor label propagation to VLMs and zero-shot classification over bi-modal graphs, proposing per modality neighbor search and balancing of contributions." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.338, + 0.468, + 0.397 + ], + "angle": 0, + "content": "- We propose an efficient way for performing inductive inference with label propagation via a dual solution and through sparsification. This not only improves the test-time efficiency of our method but also performance." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.402, + 0.468, + 0.445 + ], + "angle": 0, + "content": "- We complement our method with the class proxies presented in concurrent work [30] and achieve state of the art results for zero-shot on 14 common datasets." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.288, + 0.468, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.461, + 0.214, + 0.475 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.468, + 0.576 + ], + "angle": 0, + "content": "In this section, we discuss related works that improve the already impressive zero-shot classification performance of vision-language models [4, 16, 31] even further. This is achieved by devising better distance metrics, utilizing external knowledge to learn more expressive textual prompts, or by leveraging synthetic and unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.806 + ], + "angle": 0, + "content": "Improved distance metrics. Zero-shot classification can be improved by devising a better distance metric between image and text representations [8, 46]. CALIP [8] uses a parameter-free attention mechanism and a local patch representation, i.e. instead of global representations, to improve the estimation of class-to-image similarity. CLIP-DN [46] improves the test-time similarity estimation by alignment with the similarity used during contrastive pre-training of VLMs. To achieve this, the method assumes access to unlabeled data from the target distribution. TPT [35] optimizes textual prompts via a consistency objective across test image augmentations. Our method can be considered a part of this line of work, as label propagation is a similarity measure in a geodesic space instead of the Euclidean space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Improved textual prompts using language models. CLIP [31] uses hand-crafted prompts that are specialized for each domain. Instead of using hand-crafted prompts, generating them with large language models (LLMs) is shown to be promising [24, 29]. VisDesc [24] and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "CuPL [29] query LLMs to generate diverse descriptions of all classes, while WaffleCLIP [32] operates on top of Vis-Desc to systematically analyze which parts of the generated prompts are the most important. Instead of generating class descriptions, CHiLS [26] targets diversifying the set of classes, by generating sub-classes per class, either through an existing class hierarchy or by querying an LLM. It then performs zero-shot classification using sub-classes and linking them to the parent class. We show that methods improving the textual prompts are complementary to our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.397 + ], + "angle": 0, + "content": "Synthetic data. Recent methods [9, 38, 43] demonstrate that the use of synthetic data is beneficial for zero-shot classification. CLIP+SYN [9] uses a stable-diffusion-based model to generate synthetic images using class names and uses them to train a linear classifier, initialized by the VLM class representations. SuS-X [38] considers a similar approach, but relies on a non-parametric classifier. CaFO [43] follows the same path, but additionally includes text prompts generated by LLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.892, + 0.551 + ], + "angle": 0, + "content": "External datasets. Besides the use of synthetic data, SuS-X proposes a variant that operates on an extensive unlabeled image dataset (LAION-5B [34]). This dataset encompasses a distribution that is a super set of the target one. The method generates pseudo-labels using the zero-shot approach which are then incorporated within the non-parametric classifier. NeuralPriming [40] additionally assumes that images have captions, which are used to improve the pseudo-labeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Unlabeled images from the target distribution. Another line of research [11, 12, 27, 30] propose operating on unlabeled datasets from the target distribution. The main ingredient of all these methods is the prediction of pseudo-labels for unlabeled examples, that are later used for further processing. UPL [12] optimizes learnable text prompts based on the pseudo-labels. SVL-Adapter [27] first trains a self-supervised model on unlabeled data, and then an adapter module to align its outputs to the pseudolabels. ReCLIP [11] performs transductive label propagation to obtain the pseudo-labels and uses them to fine-tune the VLM visual and textual encoders. In contrast to that, we do not fine-tune the model, which we may not have access to, and efficiently use label propagation for inductive inference too. InMaP [30] is a concurrent work that uses pseudo-labels to update the class representations such that they are now closer to image representations. We show in the experiments that this approach is complementary to ours. In contrast to all those methods, we do not explicitly require pseudo-label prediction, but rather capture interactions between all unlabeled examples through a proximity graph and label propagation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23210" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.17, + 0.106 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.178 + ], + "angle": 0, + "content": "We first define the task of zero-shot classification with access to unlabeled examples, present label propagation with our contributions and then present the proposed approach for zero-shot classification using unlabeled examples." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.187, + 0.276, + 0.202 + ], + "angle": 0, + "content": "3.1. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.21, + 0.469, + 0.271 + ], + "angle": 0, + "content": "Vision-language models consist of an image encoder \\( f: \\mathcal{I} \\to \\mathbb{R}^d \\) and a text encoder \\( g: \\mathcal{T} \\to \\mathbb{R}^d \\), where \\( \\mathcal{I} \\) and \\( \\mathcal{T} \\) represent the space of images and text, respectively. We consider the outputs of these encoders to be \\( \\ell_2 \\)-normalized." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.271, + 0.469, + 0.392 + ], + "angle": 0, + "content": "Let \\(\\mathcal{C}\\) denote a set of known classes with associated class names \\(\\{l_1,\\dots,l_C\\}\\) and \\(\\mathcal{P} = \\{p_1,\\dots,p_P\\}\\) a set of prompt templates. Each prompt is combined with a class name to produce a textual description of the class, i.e. \\(p_i(l_c)\\) for the \\(i\\)-th template used for class \\(c\\). Class representations \\(\\mathbf{w}_c = 1 / P\\sum_{i = 1}^{P}g(p_i(l_c))\\) are obtained using the VLM. Then, for a test image \\(u\\), we extract representation \\(\\mathbf{u} = f(u)\\), and perform zero-shot classification by \\(\\mathrm{argmax}_c\\mathbf{u}^T\\mathbf{w}_c\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.392, + 0.469, + 0.437 + ], + "angle": 0, + "content": "We further assume access to a set \\(\\mathcal{U}\\) of \\(M\\) unlabeled images. Let \\(\\{\\mathbf{u}_1,\\dots ,\\mathbf{u}_M\\}\\) denote the representations of the unlabeled images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.438, + 0.469, + 0.514 + ], + "angle": 0, + "content": "In this work, we assume no direct access to the VML model weights, i.e. VLM training is neither possible nor desired. This allows to consider the underlying VLM as a black box, possibly only available through an API that generates features. We consider two inference setups:" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.514, + 0.469, + 0.574 + ], + "angle": 0, + "content": "Inductive inference: We consider an inductive setup, where we need to construct a classifier that can operate on new examples. This classifier should take advantage of the unlabeled examples in \\(\\mathcal{U}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.574, + 0.469, + 0.665 + ], + "angle": 0, + "content": "Transductive inference: We consider \\(\\mathcal{U}\\) to be the test set, i.e. all test examples are jointly provided in advance. Prediction for test example \\(\\mathbf{u}_i\\) may therefore depend on the representations and predictions of all other test examples. In this transductive setup the models are not required to provide predictions for any example that is not in \\(\\mathcal{U}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.675, + 0.295, + 0.692 + ], + "angle": 0, + "content": "3.2. Label propagation (LP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.469, + 0.865 + ], + "angle": 0, + "content": "Let \\(\\{\\mathbf{x}_1, \\dots, \\mathbf{x}_N\\}\\), with \\(\\mathbf{x}_i \\in \\mathbb{R}^d\\), be a set of features for \\(N\\) examples. Each feature represents a graph node. We construct an adjacency matrix \\(S \\in \\mathbb{R}^{N \\times N}\\) with zero diagonal, and \\(s_{ij}\\) equal to \\(\\mathbf{x}_i^\\top \\mathbf{x}_j\\) if \\(\\mathbf{x}_j\\) is in the \\(k\\)-nearest neighbors of \\(\\mathbf{x}_i\\) (denoted by \\(\\mathbf{x}_j \\in \\mathrm{kNN}(\\mathbf{x}_i)\\)), and 0 otherwise. We obtain a symmetric adjacency matrix by \\(\\bar{S} = S + S^\\top\\), and its symmetrically normalized version by \\(\\hat{S} = D^{-\\frac{1}{2}}\\bar{S}D^{-\\frac{1}{2}}\\), where \\(D = \\mathrm{diag}(\\bar{S}\\mathbf{1}_N)\\) is the degree matrix, and \\(\\mathbf{1}_N\\) is the all-ones \\(N\\)-dimensional vector. We assume the first \\(C\\) examples, and the corresponding nodes, to be labeled among \\(C\\) classes; each class is assigned to a single node1." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Transductive inference. Label propagation [44] is originally proposed for the transductive inference setup; we need to predict labels for the unlabeled nodes of the graph. Given the normalized adjacency matrix \\(\\hat{S}\\), label propagation is an iterative process given by" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.178, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {c} ^ {(t + 1)} = \\alpha \\hat {S} \\hat {\\mathbf {y}} _ {c} ^ {(t)} + (1 - \\alpha) \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.208, + 0.892, + 0.287 + ], + "angle": 0, + "content": "until convergence. Where \\(\\alpha \\in (0,1)\\) is a propagation hyper-parameter, \\(\\mathbf{y}_c = \\mathbf{e}_c \\in \\{0,1\\}^N\\) is a one-hot vector with the non-zero element at index \\(c\\), and \\(t\\) is the current iteration. Prediction of the label for an unlabeled node \\(j \\in \\{C + 1,\\dots,N\\}\\) is then given by" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.299, + 0.891, + 0.316 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {y}} _ {c} (j), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.373 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{y}}_c(j) = \\mathbf{e}_j^T\\hat{\\mathbf{y}}_c\\) is the \\(j\\)-th element of the vector \\(\\hat{\\mathbf{y}}_c\\). One can show [44] that this iterative solution is equivalent to solving \\(C\\) linear systems" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.386, + 0.891, + 0.403 + ], + "angle": 0, + "content": "\\[\nL \\hat {\\mathbf {y}} _ {c} = \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.445 + ], + "angle": 0, + "content": "where \\(L = I - \\alpha \\tilde{S}\\) is the graph Laplacian. These linear systems have a closed-form solution" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.457, + 0.891, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} _ {c} = L ^ {- 1} \\mathbf {y} _ {c} = L _ {\\text {i n v}} \\mathbf {y} _ {c}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.892, + 0.577 + ], + "angle": 0, + "content": "However, this closed-form solution is not practical for large datasets as the inverse graph Laplacian \\( L_{\\mathrm{inv}} \\) is a non-sparse \\( \\mathbb{R}^{N\\times N} \\) matrix. For this reason it is usual [3, 7, 13, 15] to solve (3) using the conjugate-gradient (CG) method, which is known to be faster than running the iterative solution [13]. Using CG is possible because \\( L \\) is positive-definite." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.699 + ], + "angle": 0, + "content": "Observe that (4) simply picks one of the columns of \\( L_{\\mathrm{inv}} \\). Matrix element \\( L_{\\mathrm{inv}}(j, c) \\) is the confidence of example \\( j \\) belonging to class \\( c \\). Its values are similarities, after label propagation, between each node pair. It is a type of geodesic similarity that captures the geometry of the feature space as this is indicated by the graph structure. Focusing on a classification task, we are only interested in similarities between an unlabeled example and a class node." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Dual solution. Herein, we show that solving \\( C \\) linear systems of the form in (4) to obtain predictions for all unlabeled nodes using (2) is equivalent to solving \\( N - C \\) linear systems of form" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.791, + 0.891, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {j} = L ^ {- 1} \\mathbf {e} _ {j} \\quad \\forall j \\in \\{C + 1, \\dots , N \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.821, + 0.831, + 0.836 + ], + "angle": 0, + "content": "and obtaining the unlabeled node prediction using" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.849, + 0.891, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {z}} _ {j} (c). \\tag {6}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1The theoretical part described in this section holds for the case of more labeled examples per class too. We consider this specific case for" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "simplicity of the presentation and because it corresponds to the task of zero-shot classification with unlabeled examples." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.275, + 0.105 + ], + "angle": 0, + "content": "This comes from the fact that" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.112, + 0.47, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {j} (c) = \\mathbf {e} _ {c} ^ {T} L ^ {- 1} \\mathbf {e} _ {j} = \\mathbf {e} _ {j} ^ {T} L ^ {- 1} \\mathbf {e} _ {c} = \\hat {\\mathbf {y}} _ {c} (j). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.469, + 0.227 + ], + "angle": 0, + "content": "Although we present the dual solution using the closed-form (4), the same holds with the CG solution of (3). Using the dual solution (5) is not practical for transductive learning as usually the unlabeled nodes are many more than the labeled ones. However, we show that this dual solution is efficiently used for inductive inference." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.228, + 0.469, + 0.304 + ], + "angle": 0, + "content": "As discussed, we can view \\( L_{\\mathrm{inv}} \\) as a pairwise similarity matrix. The confidence of example \\( j \\) belonging to class \\( c \\), due to symmetry of \\( L_{\\mathrm{inv}} \\), is equivalently obtained either by \\( L_{\\mathrm{inv}}(j, c) \\) or \\( L_{\\mathrm{inv}}(c, j) \\). This constitutes, an additional interpretation of the duality in the solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.44 + ], + "angle": 0, + "content": "Inductive inference. Test examples now come individually and are not known during graph construction. A possible way to perform inductive inference is by adding the new node to the graph, which is expensive for a test-time operation as \\(\\hat{S}\\) would have to be updated for each new test example. Instead, inspired by [13] that uses LP for retrieval, we construct indicator vector \\(\\mathbf{y}_{\\mathbf{x}} \\in \\mathbb{R}^{N}\\) for test example \\(\\mathbf{x}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.445, + 0.469, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {\\mathbf {x}} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {x} ^ {T} \\mathbf {x} _ {j}, & \\text {i f} \\mathbf {x} _ {j} \\in \\mathrm {k N N} (\\mathbf {x}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.491, + 0.273, + 0.506 + ], + "angle": 0, + "content": "Then, we solve linear system" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.512, + 0.469, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {\\mathbf {x}} = L ^ {- 1} \\mathbf {y} _ {\\mathbf {x}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.469, + 0.612 + ], + "angle": 0, + "content": "as in the dual formulation (5) and get a prediction with \\(\\hat{y}_{\\mathbf{x}} = \\operatorname{argmax}_c \\hat{\\mathbf{z}}_{\\mathbf{x}}(c)\\). With the usual formulation of label propagation in (4), \\(C\\) linear systems need to be solved to get prediction for a single test example. The dual formulation allows us to do it by solving only a single linear system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.794 + ], + "angle": 0, + "content": "Fast inductive inference with sparsification. We further introduce an additional off-line step, where we solve (4), get \\(\\hat{\\mathbf{y}}_c\\) for all \\(c\\in \\{1,\\dots,C\\}\\), and store them in a matrix \\(\\hat{Y} = [\\hat{\\mathbf{y}}_1;\\dots;\\hat{\\mathbf{y}}_c]\\in \\mathbb{R}^{N\\times C}\\). Then, the solution for a test example is equivalent to a weighted sum of rows of \\(\\hat{Y}\\) [1], which is a byproduct of using the indicator vector (8) for representing a test example. Its prediction is given by \\(\\hat{\\mathbf{z}}_{\\mathbf{x}} = \\mathbf{y}_{\\mathbf{x}}^{T}\\hat{Y}\\), and is equivalent to that obtained via (9). However, storing the whole \\(\\hat{Y}\\) can be expensive for very large values of \\(N\\) and \\(C\\). We propose to sparsify \\(\\hat{Y}\\) by keeping only the largest values in each row, column, or over the whole matrix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Note that [14] proposes a low-rank decomposition of the inverse graph Laplacian for the task of retrieval. Our solution is tailored to zero-shot classification, and we choose to obtain and sparsify the first \\(C\\) rows of \\(L_{\\mathrm{inv}}\\) instead of approximating the whole matrix. Additionally, our solution requires one (sparse) vector to matrix multiplications at test-time instead of two." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.81, + 0.106 + ], + "angle": 0, + "content": "3.3. LP for zero-shot VLM classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.206 + ], + "angle": 0, + "content": "We are given a set of classes \\(\\mathcal{C}\\) with extracted VLM representations \\(\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C\\}\\) and a set of unlabeled images \\(\\mathcal{U}\\) with extracted VLM representations \\(\\{\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}\\). We use them as nodes \\(\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C,\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}\\) of the graph for label propagation. Nodes of class representations (text nodes) are labeled and image nodes unlabeled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.358 + ], + "angle": 0, + "content": "To construct the adjacency matrix \\( S \\), we need to perform the \\( k \\)-nearest neighbor search between nodes. However, it is known that there exists a large modality gap between image and text representations coming from VLMs [21, 38, 47]. The respective similarity distributions for CLIP are shown in Figure 3a. This modality gap makes standard kNN search between nodes not useful for label propagation; image nodes mostly get connected to image nodes, and text nodes mostly get connected to text nodes. As a consequence, few edges exist between labeled and unlabeled nodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.482 + ], + "angle": 0, + "content": "To alleviate this problem, we perform the kNN search separately for connecting image nodes to image nodes and for connecting image nodes and text nodes. We do not perform the search using text nodes as queries, i.e. text nodes get linked only if they appear in the kNN list of an image. This way we also avoid linking text nodes with each other, which is beneficial as each of them is labeled to a different class. Formally, the values of the adjacency matrix are" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.496, + 0.891, + 0.552 + ], + "angle": 0, + "content": "\\[\ns _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {u} _ {i} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u} _ {i}) \\\\ \\mathbf {u} _ {i} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u} _ {i}) \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.566, + 0.891, + 0.595 + ], + "angle": 0, + "content": "where \\(\\mathrm{kNN_u}\\) and \\(\\mathrm{kNN_w}\\) denote that the search is performed within the image or class features only, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.891, + 0.643 + ], + "angle": 0, + "content": "Moreover, during inductive inference for image \\(\\mathbf{u}\\), we perform the kNN search in a similar way to construct indicator vector \\(\\mathbf{y_u}\\) whose elements are given by" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.656, + 0.891, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {\\mathbf {u}} (i) = \\left\\{ \\begin{array}{l l} \\mathbf {u} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u}) \\\\ \\mathbf {u} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Due to the two types of edges, i.e. image-to-image and image-to-text, we use power function \\( h(v) = v^{\\gamma} \\) to transform the image-to-text (cross-modal) similarities. This way we effectively balance their contribution in the graph and the indicator vector. To that end, we use \\( h(\\mathbf{u}_i^T\\mathbf{w}_j) \\) and \\( h(\\mathbf{u}^T\\mathbf{w}_j) \\) in (10) and (11), respectively, instead of \\( \\mathbf{u}_i^T\\mathbf{w}_j \\) and \\( \\mathbf{u}^T\\mathbf{w}_j \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We refer to the proposed method described above as Zero-shot classification with Label Propagation (ZLaP). We further denote the variant of our method after sparsifying the \\(\\hat{Y}\\) matrix for inductive inference as ZLaP*." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "23212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.096, + 0.411, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.096, + 0.882, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.282, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 2. t-SNE visualization for the original CLIP features (left) and our geodesic similarity (right). The former is estimated with the features as input, while the latter with the \\( L_{\\mathrm{inv}} \\) used as a pairwise similarity matrix. \\( \\star \\): class representation, \\( \\bullet \\): image representation. Figure generated for five random classes from the CUB dataset." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.353, + 0.268, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.484, + 0.236, + 0.497 + ], + "angle": 0, + "content": "(a) Using text prompts" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.353, + 0.462, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.483, + 0.465, + 0.497 + ], + "angle": 0, + "content": "(b) Using proxies from InMaP [30]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.508, + 0.472, + 0.551 + ], + "angle": 0, + "content": "Figure 3. Similarity distributions among features of the same or different modality, using 7 textual templates [38] (left) or the InMaP proxies (right) as class representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.47, + 0.742 + ], + "angle": 0, + "content": "\\(t\\)-SNE visualization of the bi-modal space. In Figure 2 we visualize the bi-modal feature space for CLIP features using \\(t\\)-SNE [39] in two cases, i.e. the Euclidean case and using geodesic similarities obtained by \\(L_{\\mathrm{inv}}\\), i.e. after label propagation. When using Euclidean affinities (left), we see that due to the large differences in the similarity distributions (text-text, image-to-image, and text-to-image, as shown in Figure 3a) all class representations (stars) are clustered together far from the image nodes. However, using the geodesic affinities from \\(L_{\\mathrm{inv}}\\) (right) we see that class representations are more spread." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.21, + 0.77 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.84 + ], + "angle": 0, + "content": "In this section, we first present the datasets we use, our experimental setup and competing methods. We then present component analysis for ZLaP and results for transductive and inductive zero-shot classification on 14 datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.182, + 0.863 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We evaluate the proposed method on 14 diverse image classification datasets: ImageNet ILSVRC2012 [33]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Describable Textures Dataset (DTD) [5], EuroSAT [10], FGVC-Aircraft [23], Oxford Flowers 102 [25], Food101 [2], Oxford-IIIT Pet [28], SUN397 [42], Stanford Cars [17], Caltech101 [6], UCF101 [36], CIFAR10 [18], CIFAR100 [18], CUB-200-2011 [41]. For the first 11 datasets we borrow the train and test splits from CoOp[45]. We use the official training and test splits for CIFAR10, CIFAR100 and CUB-200-2011." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.488, + 0.689, + 0.505 + ], + "angle": 0, + "content": "4.2. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.894, + 0.559 + ], + "angle": 0, + "content": "In the transductive (inductive) inference setup, unlabeled nodes in the graph are the test (train) images. We always measure classification accuracy over the test images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.893, + 0.663 + ], + "angle": 0, + "content": "VLMs and textual prompts. We report results using the publicly available ResNet50 and ViT-B/16 CLIP [31] models. We adopt the 7 templates from SuS-X [38] as class prompts for all results apart from Table 3 where we utilize the LLM generated prompts from [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Compared methods. Our baseline is zero-shot recognition with CLIP [31] using text encoder features as class representations. TPT [35] is based on test-time prompt tuning such that different image augmentations produce consistent predictions. The aforementioned methods do not exploit unlabeled data; their performance is therefore unchanged in both inference setups. CLIP-DN [46] normalizes feature distributions during test-time and assumes access to the mean feature vector of the target distribution. In the transductive (inductive) setup the mean vector is estimated on the test (training) set. InMaP [30] is a concurrent work that extracts updated class representations using pseudo-labels on the unlabeled set. In the transductive (inductive) setup the learning is performed on the test (training) images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23213" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.213 + ], + "angle": 0, + "content": "Implementation details. We reproduce results for \\(\\mathrm{CLIP}^2\\), \\(\\mathrm{CLIP - DN^3}\\), and \\(\\mathrm{InMaP^4}\\) using their public implementations. For TPT [35] we report the numbers provided in [30]. We run InMaP using a single set of hyper-parameters for all 14 datasets, i.e. the default values reported in the official implementation4. We also fix the values of \\(k\\), \\(\\gamma\\), and \\(\\alpha\\) for ZLaP across all datasets to 5, 5.0, and 0.3, respectively, for CLIP, and 10, 3.0, and 0.3, respectively, for InMaP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.47, + 0.322 + ], + "angle": 0, + "content": "ZLaP variants. We refer to ZLaP using text class representations as \\(\\mathbf{CLIP} + \\mathbf{ZLaP}\\). Since InMaP is complementary to our work, we further evaluate the performance of ZLaP when \\(\\{\\mathbf{w}_1,\\dots ,\\mathbf{w}_C\\}\\) are the InMaP proxies. We refer to this as InMaP + ZLaP in the results. We refer to ZLaP with a sparse \\(\\hat{Y}\\) for inductive inference as ZLaP*." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.331, + 0.279, + 0.347 + ], + "angle": 0, + "content": "4.3. Components of ZLaP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.473, + 0.566 + ], + "angle": 0, + "content": "Bi-modal graph adjustments. In Table 1 we show the importance of two design choices to adapt LP to bi-modal graphs, i.e. separating the nearest neighbor search across modalities using (10) and (11), and transforming cross-modal similarities using a power function \\( h(\\cdot) \\). We see that separate search is crucial; without it LP is not effective at all. The power function gives an extra boost in both setups, especially in the case of transductive inference. In Table 2 we report the percentage of images that are connected to their groundtruth class nodes within a path of length \\( n \\), with and without our adjustments. We see that for any such paths to exist for the case without adjustments, \\( k \\) needs to be extremely high. With adjustments, \\( k = 5 \\) is enough for \\( 71.4\\% \\) of the nodes to be connected to the correct class nodes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.844 + ], + "angle": 0, + "content": "Sparsifying matrix \\(\\hat{Y}\\) for inductive inference. We explore three ways of approximating \\(\\hat{Y}\\) by sparsification, i.e. wither keeping only the largest \\(\\xi\\) columns per row, the largest \\(\\xi\\) rows per column, or the largest \\(\\xi\\) elements of the whole matrix. In all cases, the rest of the elements are set to zero. In Figure 4 we show the influence that these three variants have on performance. Not only these variants speed-up inference, but we can also see improvements in performance when sparsification percentage is high, i.e. low percentage of non-zero elements. We attribute this to the fact that less confident predictions in \\(\\hat{Y}\\), many of them erroneous, are now set to zero. Although the best variant to choose seems to vary per dataset, we found that keeping the top element per row performs well across different datasets. We therefore use the \\(\\xi = 1\\) top element per row for our experiments. This amounts to different percentages of sparsity per dataset; we are keeping approximately \\(2.3\\%\\) on average" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.885, + 0.175 + ], + "angle": 0, + "content": "
Eq.(10)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X50.232.341.6
✓✓61.841.952.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.625, + 0.177, + 0.764, + 0.189 + ], + "angle": 0, + "content": "(a) Transductive inference" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.207, + 0.885, + 0.293 + ], + "angle": 0, + "content": "
Eq.(10)-(11)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X60.842.449.6
✓✓62.242.849.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.633, + 0.295, + 0.756, + 0.306 + ], + "angle": 0, + "content": "(b) Inductive inference" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.319, + 0.892, + 0.376 + ], + "angle": 0, + "content": "Table 1. Adjusting LP to bi-modal graphs. Impact of using separate kNN search for constructing the graph (Eq.(10)) or the indicator vector (Eq.(11)), as well as power function \\( h(\\cdot) \\) for balancing the contributions of two types of edges in the graph." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.388, + 0.882, + 0.482 + ], + "angle": 0, + "content": "
kJoint kNN searchSeparate kNN search
n=1n=2n=3n=1n=2n=3
50.00.00.071.485.1100.0
100.00.00.082.995.4100.0
10040.1100.0100.0100.0100.0100.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Table 2. Impact of the separate kNN search on the shortest paths between image nodes and the text node of their class. We report the percentage of images whose shortest path to the text node of their ground-truth class has length equal to or less than \\( n \\). Left: the vanilla approach. Right: our separate kNN search using Eq. (10). Analysis on DTD for the transductive setup." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.609, + 0.892, + 0.671 + ], + "angle": 0, + "content": "across all datasets. Regarding the inference speed-up, the primal solution takes \\(\\sim 2.6\\) sec per image, the dual takes \\(\\sim 4.4\\) ms, while the sparsified approach takes \\(\\sim 0.6\\) ms, measured on ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Using the class proxies from InMaP. We observe that many class-to-image similarities (e.g. \\(\\mathbf{u}_i^\\top \\mathbf{w}_j\\) in (10)) become negative on some datasets when using ZLaP with InMaP proxies (see Figure 3b). We therefore perform minmax normalization in range [0,1] after constructing adjacency matrix \\(S\\) or the indicator vector, for the transductive and inductive inference setups." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.593, + 0.817 + ], + "angle": 0, + "content": "4.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Transductive inference. We present results for transductive zero-shot classification in Figure 5. ZLaP improves the zero-shot performance of CLIP significantly on all datasets. It also outperforms the recent TPT and CLIP-DN approaches on the vast majority of cases, with large gains" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.849, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/OpenAI/CLIP \n\\(^{3}\\)https://github.com/fengyuli-dev/distribution-normalization \n\\(^{4}\\)https://github.com/idstcv/InMaP" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.379, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.09, + 0.614, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.88, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.284, + 0.866, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Sparcifying matrix \\(\\hat{Y}\\) for inductive CLIP+ZLaP: effect of maintaining only the top elements per row/column/matrix." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.314, + 0.466, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.313, + 0.862, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.492, + 0.467, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.492, + 0.862, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.697, + 0.895, + 0.727 + ], + "angle": 0, + "content": "Figure 5. Zero-shot classification accuracy averaged over 14 datasets for the transductive (top) and inductive (bottom) setups. Results per dataset are reported in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.473, + 0.844 + ], + "angle": 0, + "content": "in average accuracy. Compared to InMaP, ZLaP offers lower accuracy on average. However, by incorporating InMaP's class representations to our graph, we can improve our results even further and outperform all other methods, for an improvement of approximately \\(+5\\%\\) over CLIP with both backbones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Inductive inference. We report results for the inductive inference setup in Figure 5. ZLaP achieves a noticeable im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.895, + 0.89 + ], + "angle": 0, + "content": "provements over the CLIP baseline in this setup as well. Gains are more prominent for the case of ZLaP using the InMaP proxies, where gains over CLIP are \\(+4.4\\%\\) and \\(+4.9\\%\\) for the two backbones. We also observe that, although InMaP slightly outperforms ZLaP when used over CLIP, the combination of the two achieves state-of-the-art performance in this case as well. We further see that \\(\\mathrm{ZLaP^{*}}\\) retains the state-of-the-art performance of our method, while sparsifying \\(\\hat{Y}\\) offers significant speed-up at inference time." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23215" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.473, + 0.282 + ], + "angle": 0, + "content": "
TransductiveInductive
Results with RN50
CLIP63.063.0
+ ZLaP64.664.2
InMaP64.864.6
+ ZLaP65.865.0
Results with ViT-B-16
CLIP71.971.9
+ ZLaP72.673.3
InMaP73.974.0
+ ZLaP74.874.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.3, + 0.471, + 0.358 + ], + "angle": 0, + "content": "Table 3. Zero-shot classification using prompts generated by LLMs [29]. We report average accuracy on 12 datasets using prompts from CuPL [29] together with our 7 standard prompts. Results per dataset are reported in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.472, + 0.525 + ], + "angle": 0, + "content": "Leveraging LLM generated prompts. In Table 3 we report average zero-shot classification accuracy for ZLaP using the prompts recently proposed in CuPL [29]. These are prompts generated by LMMs that are available on the CuPL Github page5 for 12 of the datasets we use (all datasets besides CUB and Eurosat). ZLaP improves zero-shot performance in this case as well, for both the transductive and inductive setups. This verifies that our method is complementary to improved prompt engineering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.631 + ], + "angle": 0, + "content": "Multi-label classification. We apply ZLaP for multi-label classification on the MS-COCO [22] dataset. ZLaP improves the zero-shot performance of CLIP by \\(+6.0\\%\\) mAP (56.8% vs. 50.8%) for inductive inference without any modification of the approach or its hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.871 + ], + "angle": 0, + "content": "Web-crawled unlabeled images. All previous experiments use unlabeled images that come from the target distribution, i.e. they are known to depict one of the classes of interest but their labels are discarded. To see the impact of ZLaP in a more realistic setup using web-crawled images we rely on LAION-400M [34] composed by image-caption pairs. We construct the set of unlabeled images with 10,000 images per class that are chosen either randomly, or based on proximity of their image or text features to the class representation. Random selection fails, but the other two options provide some improvement compared to CLIP, with the caption-based neighbors being a bit better. The complete set of results is presented in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.089, + 0.858, + 0.25 + ], + "angle": 0, + "content": "
TransductiveInductive
BLIP [20]54.654.6
+ ZLaP59.657.9
ALBEF [19]36.036.0
+ ZLaP41.246.8
EVA-CLIP-8B [37]83.683.6
+ ZLaP84.684.5
EVA-CLIP-18B [37]83.983.9
+ ZLaP84.884.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.259, + 0.868, + 0.275 + ], + "angle": 0, + "content": "Table 4. Accuracy on ImageNet using different VLMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.438 + ], + "angle": 0, + "content": "Different VLMs We use CLIP as the VLM of choice throughout our experiments. In Table 4, we present results when ZLaP is applied on top of four recent VLMs, namely BLIP [20], ALBEF [19], and two versions of EVA-CLIP [37]. We use the implementations of BLIP and ALBEF that are available in the LAVIS library6, while for EVA-CLIP we use implementation from the official Github repository7. ZLaP improves the results of all four different VLMs in both transductive and inductive setups." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.45, + 0.628, + 0.466 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.475, + 0.893, + 0.656 + ], + "angle": 0, + "content": "Label propagation is an intuitive way of encoding the global structure of unlabeled data into geodesic distances over a locally Euclidean space. In this paper, we show that this method can be successfully tailored to both transductive and inductive zero-shot classification with vision-language models, and achieve state-of-the-art performance on both setups. To that end, we show that it is highly important to take proper care of the peculiarities of the bi-modal nature of the task during graph construction. We further carefully design an efficient variant of label propagation for the inductive inference case, that may enable label propagation to be applied to other tasks beyond zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.763 + ], + "angle": 0, + "content": "Vision-language models trained on billion-scale datasets are redefining computer vision research. The proposed ZLaP is a training-free approach able to improve the generalization performance of black-box VLMs using only unlabeled data, for an annotation-free, text-based and open-world classification paradigm that will inevitably be ubiquitous in the near future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.894, + 0.843 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the Junior Star GACR GM 21-28830M and the Czech Technical University in Prague grant No. SGS23/173/OHK3/3T/13. We thank Ahmet Iscen for many helpful comments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.38, + 0.902 + ], + "angle": 0, + "content": "5https://github.com/sarahpratt/CuPL" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.863, + 0.81, + 0.876 + ], + "angle": 0, + "content": "\\(^{6}\\)https://github.com/salesforce/LAVIS" + }, + { + "type": "page_footnote", + "bbox": [ + 0.501, + 0.877, + 0.892, + 0.899 + ], + "angle": 0, + "content": "https://github.com/baaivision/EVA/tree/master/EVA-CLIP-18B" + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.863, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "23216" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. Label propagation and quadratic criterion. In Semi-Supervised Learning, pages 192-216. The MIT Press, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.47, + 0.24 + ], + "angle": 0, + "content": "[3] Siddhartha Chandra and Iasonas Kokkinos. Fast, exact and multi-scale inference for semantic image segmentation with deep Gaussian CRFs. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.469, + 0.309 + ], + "angle": 0, + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[5] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In CVPR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.469, + 0.436 + ], + "angle": 0, + "content": "[7] Leo Grady. Random walks for image segmentation. pami, 28(11):1768-1783, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.492 + ], + "angle": 0, + "content": "[8] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. CALIP: zero-shot enhancement of CLIP with parameter-free attention. In AAAI, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.469, + 0.548 + ], + "angle": 0, + "content": "[9] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip H. S. Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[10] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[11] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, Cheng-Hao Kuo, and Ram Nevatia. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[12] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[13] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Teddy Furon, and Ondrej Chum. Efficient diffusion on region manifolds: Recovering small objects with compact CNN representations. In CVPR, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[14] Ahmet Iscen, Yannis Avrithis, Giorgos Tolias, Teddy Furon, and Ondrej Chum. Fast spectral ranking for similarity search. In CVPR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[18] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.194, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[19] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[21] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y. Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[23] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[24] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[26] Zachary Novack, Julian J. McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[27] Omiros Pantazis, Gabriel J. Brostow, Kate E. Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. In BMVC, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[28] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[29] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[30] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In NeurIPS, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "23217" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.147 + ], + "angle": 0, + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.245 + ], + "angle": 0, + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.248, + 0.468, + 0.302 + ], + "angle": 0, + "content": "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.304, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.361, + 0.468, + 0.414 + ], + "angle": 0, + "content": "[37] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP18B: scaling CLIP to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.418, + 0.468, + 0.458 + ], + "angle": 0, + "content": "[38] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.461, + 0.468, + 0.486 + ], + "angle": 0, + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.543 + ], + "angle": 0, + "content": "[40] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. Neural priming for sample-efficient adaptation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.546, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[41] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.589, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.632, + 0.468, + 0.685 + ], + "angle": 0, + "content": "[43] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[44] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In NIPS, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.774, + 0.468, + 0.827 + ], + "angle": 0, + "content": "[46] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Test-time distribution normalization for contrastively learned vision-language models. In NeurlPS, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.468, + 0.884 + ], + "angle": 0, + "content": "[47] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.955 + ], + "angle": 0, + "content": "23218" + } + ] +] \ No newline at end of file diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_origin.pdf b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7bfc59e61156e5205a6f04b8f6e01e5d53912491 --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/f789bcea-03da-4382-8fd5-69048545fa3d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8286a26ac51c6e7eab5990dbf29f2364a4b8469e1752da7c6b830a2ac46cfdf1 +size 210195 diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/full.md b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eee4b78b0a0e05bf68580f03a50d4ebfaa537f71 --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/full.md @@ -0,0 +1,328 @@ +# Label Propagation for Zero-shot Classification with Vision-Language Models + +Vladan Stojnic1 + +Yannis Kalantidis² + +Giorgos Tolias1 + +$^{1}$ VRG, FEE, Czech Technical University in Prague + +$^{2}$ Naver Labs Europe + +# Abstract + +Vision-Language Models (VLMs) have demonstrated impressive performance on zero-shot classification, i.e. classification when provided merely with a list of class names. In this paper, we tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. We tailor LP to graphs containing both text and image features and further propose an efficient method for performing inductive inference based on a dual solution and a sparsification step. We perform extensive experiments to evaluate the effectiveness of our method on 14 common datasets and show that ZLaP outperforms the latest related works. Code: https://github.com/vladan-stojnic/ZLaP + +# 1. Introduction + +Vision-Language Models (VLMs) have demonstrated impressive performance on a variety of computer vision tasks. They are usually trained on large datasets of image-text pairs and contain visual and textual encoders that map to a common feature space. Visual encoders from such models have been shown to produce strong visual representations for perception tasks [31]. Given labeled data from a downstream dataset, one can fine-tune the model or learn classifiers and achieve really high classification accuracy. + +Besides using the visual encoder in isolation, the joint text and visual encoder feature space of VLMs enables us to define text-based "classifiers", e.g. using the class names as textual prompts. This means that we only need a list of class names to perform zero-shot classification for a target dataset, i.e. without access to any labeled images. Although utilizing priors or devising better textual prompts can improve zero-shot performance [8, 24, 29, 46], here, we are interested in the case where we further have access to unlabeled data. Our goal is to find the best way of utilizing such data for zero-shot classification. + +![](images/842539b56cbdadf340fdccbf2e30b77c762dfba22a17ecd241d32f03c02e10d8.jpg) +Figure 1. Zero-shot classification performance over 14 datasets using the proposed ZLaP classifier over CLIP [31], as well as over the (concurrent) InMaP [30] approach. Our method offers performance gains for both transductive (left) and inductive (right) inference. Average accuracy over 14 common datasets is reported. + +In this paper, we leverage the inherent structure of the unlabeled data represented by a proximity graph and apply label propagation (LP) between the text-based classifiers and unlabeled images to derive geodesic distances we then use for classification. We tailor LP to VLMs and graphs containing both text and image features, and show that without proper handling of the bimodality, vanilla application of LP fails dramatically. We introduce ZLaP, a novel classification method based on label propagation that can perform both transductive and inductive inference. We perform the former with the standard (primal) solution of LP for classification and devise a more efficient dual solution for the latter. Our method is not only highly effective but also efficient, making LP a more attractive inductive classifier in terms of complexity. + +We implement our methods using publicly available VLMs as feature encoders, primarily the ResNet and ViT CLIP [31] models, and perform extensive experiments to evaluate the effectiveness of our method on 14 common + +datasets. We show that we are able to achieve top performance on two zero-shot inference setups, i.e. inductive and transductive inference. Figure 1 summarizes our gains over 14 datasets on both setups when applying our LP-powered classifiers on top of CLIP, as well as after incorporating the class proxies from the recent InMaP [30] zero-shot approach. + +It is worth highlighting that ZLaP is a non-parametric method, i.e. it does not involve a learning step. In fact, our approach does not even require access to the VLM model weights and can therefore be used to improve the zero-shot performance of a black-box model even, e.g. provided only via an API. In summary, our contributions are as follows. + +- We tailor label propagation to VLMs and zero-shot classification over bi-modal graphs, proposing per modality neighbor search and balancing of contributions. +- We propose an efficient way for performing inductive inference with label propagation via a dual solution and through sparsification. This not only improves the test-time efficiency of our method but also performance. +- We complement our method with the class proxies presented in concurrent work [30] and achieve state of the art results for zero-shot on 14 common datasets. + +# 2. Related work + +In this section, we discuss related works that improve the already impressive zero-shot classification performance of vision-language models [4, 16, 31] even further. This is achieved by devising better distance metrics, utilizing external knowledge to learn more expressive textual prompts, or by leveraging synthetic and unlabeled data. + +Improved distance metrics. Zero-shot classification can be improved by devising a better distance metric between image and text representations [8, 46]. CALIP [8] uses a parameter-free attention mechanism and a local patch representation, i.e. instead of global representations, to improve the estimation of class-to-image similarity. CLIP-DN [46] improves the test-time similarity estimation by alignment with the similarity used during contrastive pre-training of VLMs. To achieve this, the method assumes access to unlabeled data from the target distribution. TPT [35] optimizes textual prompts via a consistency objective across test image augmentations. Our method can be considered a part of this line of work, as label propagation is a similarity measure in a geodesic space instead of the Euclidean space. + +Improved textual prompts using language models. CLIP [31] uses hand-crafted prompts that are specialized for each domain. Instead of using hand-crafted prompts, generating them with large language models (LLMs) is shown to be promising [24, 29]. VisDesc [24] and + +CuPL [29] query LLMs to generate diverse descriptions of all classes, while WaffleCLIP [32] operates on top of Vis-Desc to systematically analyze which parts of the generated prompts are the most important. Instead of generating class descriptions, CHiLS [26] targets diversifying the set of classes, by generating sub-classes per class, either through an existing class hierarchy or by querying an LLM. It then performs zero-shot classification using sub-classes and linking them to the parent class. We show that methods improving the textual prompts are complementary to our approach. + +Synthetic data. Recent methods [9, 38, 43] demonstrate that the use of synthetic data is beneficial for zero-shot classification. CLIP+SYN [9] uses a stable-diffusion-based model to generate synthetic images using class names and uses them to train a linear classifier, initialized by the VLM class representations. SuS-X [38] considers a similar approach, but relies on a non-parametric classifier. CaFO [43] follows the same path, but additionally includes text prompts generated by LLMs. + +External datasets. Besides the use of synthetic data, SuS-X proposes a variant that operates on an extensive unlabeled image dataset (LAION-5B [34]). This dataset encompasses a distribution that is a super set of the target one. The method generates pseudo-labels using the zero-shot approach which are then incorporated within the non-parametric classifier. NeuralPriming [40] additionally assumes that images have captions, which are used to improve the pseudo-labeling. + +Unlabeled images from the target distribution. Another line of research [11, 12, 27, 30] propose operating on unlabeled datasets from the target distribution. The main ingredient of all these methods is the prediction of pseudo-labels for unlabeled examples, that are later used for further processing. UPL [12] optimizes learnable text prompts based on the pseudo-labels. SVL-Adapter [27] first trains a self-supervised model on unlabeled data, and then an adapter module to align its outputs to the pseudolabels. ReCLIP [11] performs transductive label propagation to obtain the pseudo-labels and uses them to fine-tune the VLM visual and textual encoders. In contrast to that, we do not fine-tune the model, which we may not have access to, and efficiently use label propagation for inductive inference too. InMaP [30] is a concurrent work that uses pseudo-labels to update the class representations such that they are now closer to image representations. We show in the experiments that this approach is complementary to ours. In contrast to all those methods, we do not explicitly require pseudo-label prediction, but rather capture interactions between all unlabeled examples through a proximity graph and label propagation. + +# 3. Method + +We first define the task of zero-shot classification with access to unlabeled examples, present label propagation with our contributions and then present the proposed approach for zero-shot classification using unlabeled examples. + +# 3.1. Problem formulation + +Vision-language models consist of an image encoder $f: \mathcal{I} \to \mathbb{R}^d$ and a text encoder $g: \mathcal{T} \to \mathbb{R}^d$ , where $\mathcal{I}$ and $\mathcal{T}$ represent the space of images and text, respectively. We consider the outputs of these encoders to be $\ell_2$ -normalized. + +Let $\mathcal{C}$ denote a set of known classes with associated class names $\{l_1,\dots,l_C\}$ and $\mathcal{P} = \{p_1,\dots,p_P\}$ a set of prompt templates. Each prompt is combined with a class name to produce a textual description of the class, i.e. $p_i(l_c)$ for the $i$ -th template used for class $c$ . Class representations $\mathbf{w}_c = 1 / P\sum_{i = 1}^{P}g(p_i(l_c))$ are obtained using the VLM. Then, for a test image $u$ , we extract representation $\mathbf{u} = f(u)$ , and perform zero-shot classification by $\mathrm{argmax}_c\mathbf{u}^T\mathbf{w}_c$ . + +We further assume access to a set $\mathcal{U}$ of $M$ unlabeled images. Let $\{\mathbf{u}_1,\dots ,\mathbf{u}_M\}$ denote the representations of the unlabeled images. + +In this work, we assume no direct access to the VML model weights, i.e. VLM training is neither possible nor desired. This allows to consider the underlying VLM as a black box, possibly only available through an API that generates features. We consider two inference setups: + +Inductive inference: We consider an inductive setup, where we need to construct a classifier that can operate on new examples. This classifier should take advantage of the unlabeled examples in $\mathcal{U}$ . + +Transductive inference: We consider $\mathcal{U}$ to be the test set, i.e. all test examples are jointly provided in advance. Prediction for test example $\mathbf{u}_i$ may therefore depend on the representations and predictions of all other test examples. In this transductive setup the models are not required to provide predictions for any example that is not in $\mathcal{U}$ . + +# 3.2. Label propagation (LP) + +Let $\{\mathbf{x}_1, \dots, \mathbf{x}_N\}$ , with $\mathbf{x}_i \in \mathbb{R}^d$ , be a set of features for $N$ examples. Each feature represents a graph node. We construct an adjacency matrix $S \in \mathbb{R}^{N \times N}$ with zero diagonal, and $s_{ij}$ equal to $\mathbf{x}_i^\top \mathbf{x}_j$ if $\mathbf{x}_j$ is in the $k$ -nearest neighbors of $\mathbf{x}_i$ (denoted by $\mathbf{x}_j \in \mathrm{kNN}(\mathbf{x}_i)$ ), and 0 otherwise. We obtain a symmetric adjacency matrix by $\bar{S} = S + S^\top$ , and its symmetrically normalized version by $\hat{S} = D^{-\frac{1}{2}}\bar{S}D^{-\frac{1}{2}}$ , where $D = \mathrm{diag}(\bar{S}\mathbf{1}_N)$ is the degree matrix, and $\mathbf{1}_N$ is the all-ones $N$ -dimensional vector. We assume the first $C$ examples, and the corresponding nodes, to be labeled among $C$ classes; each class is assigned to a single node1. + +Transductive inference. Label propagation [44] is originally proposed for the transductive inference setup; we need to predict labels for the unlabeled nodes of the graph. Given the normalized adjacency matrix $\hat{S}$ , label propagation is an iterative process given by + +$$ +\hat {\mathbf {y}} _ {c} ^ {(t + 1)} = \alpha \hat {S} \hat {\mathbf {y}} _ {c} ^ {(t)} + (1 - \alpha) \mathbf {y} _ {c} \quad \forall c \in \{1, \dots , C \} \tag {1} +$$ + +until convergence. Where $\alpha \in (0,1)$ is a propagation hyper-parameter, $\mathbf{y}_c = \mathbf{e}_c \in \{0,1\}^N$ is a one-hot vector with the non-zero element at index $c$ , and $t$ is the current iteration. Prediction of the label for an unlabeled node $j \in \{C + 1,\dots,N\}$ is then given by + +$$ +\hat {y} _ {j} = \operatorname {a r g m a x} _ {c} \hat {\mathbf {y}} _ {c} (j), \tag {2} +$$ + +where $\hat{\mathbf{y}}_c(j) = \mathbf{e}_j^T\hat{\mathbf{y}}_c$ is the $j$ -th element of the vector $\hat{\mathbf{y}}_c$ . One can show [44] that this iterative solution is equivalent to solving $C$ linear systems + +$$ +L \hat {\mathbf {y}} _ {c} = \mathbf {y} _ {c} \quad \forall c \in \{1, \dots , C \}, \tag {3} +$$ + +where $L = I - \alpha \tilde{S}$ is the graph Laplacian. These linear systems have a closed-form solution + +$$ +\hat {\mathbf {y}} _ {c} = L ^ {- 1} \mathbf {y} _ {c} = L _ {\text {i n v}} \mathbf {y} _ {c}. \tag {4} +$$ + +However, this closed-form solution is not practical for large datasets as the inverse graph Laplacian $L_{\mathrm{inv}}$ is a non-sparse $\mathbb{R}^{N\times N}$ matrix. For this reason it is usual [3, 7, 13, 15] to solve (3) using the conjugate-gradient (CG) method, which is known to be faster than running the iterative solution [13]. Using CG is possible because $L$ is positive-definite. + +Observe that (4) simply picks one of the columns of $L_{\mathrm{inv}}$ . Matrix element $L_{\mathrm{inv}}(j, c)$ is the confidence of example $j$ belonging to class $c$ . Its values are similarities, after label propagation, between each node pair. It is a type of geodesic similarity that captures the geometry of the feature space as this is indicated by the graph structure. Focusing on a classification task, we are only interested in similarities between an unlabeled example and a class node. + +Dual solution. Herein, we show that solving $C$ linear systems of the form in (4) to obtain predictions for all unlabeled nodes using (2) is equivalent to solving $N - C$ linear systems of form + +$$ +\hat {\mathbf {z}} _ {j} = L ^ {- 1} \mathbf {e} _ {j} \quad \forall j \in \{C + 1, \dots , N \}, \tag {5} +$$ + +and obtaining the unlabeled node prediction using + +$$ +\hat {y} _ {j} = \operatorname {a r g m a x} _ {c} \hat {\mathbf {z}} _ {j} (c). \tag {6} +$$ + +This comes from the fact that + +$$ +\hat {\mathbf {z}} _ {j} (c) = \mathbf {e} _ {c} ^ {T} L ^ {- 1} \mathbf {e} _ {j} = \mathbf {e} _ {j} ^ {T} L ^ {- 1} \mathbf {e} _ {c} = \hat {\mathbf {y}} _ {c} (j). \tag {7} +$$ + +Although we present the dual solution using the closed-form (4), the same holds with the CG solution of (3). Using the dual solution (5) is not practical for transductive learning as usually the unlabeled nodes are many more than the labeled ones. However, we show that this dual solution is efficiently used for inductive inference. + +As discussed, we can view $L_{\mathrm{inv}}$ as a pairwise similarity matrix. The confidence of example $j$ belonging to class $c$ , due to symmetry of $L_{\mathrm{inv}}$ , is equivalently obtained either by $L_{\mathrm{inv}}(j, c)$ or $L_{\mathrm{inv}}(c, j)$ . This constitutes, an additional interpretation of the duality in the solution. + +Inductive inference. Test examples now come individually and are not known during graph construction. A possible way to perform inductive inference is by adding the new node to the graph, which is expensive for a test-time operation as $\hat{S}$ would have to be updated for each new test example. Instead, inspired by [13] that uses LP for retrieval, we construct indicator vector $\mathbf{y}_{\mathbf{x}} \in \mathbb{R}^{N}$ for test example $\mathbf{x}$ such that + +$$ +\mathbf {y} _ {\mathbf {x}} (j) = \left\{ \begin{array}{l l} \mathbf {x} ^ {T} \mathbf {x} _ {j}, & \text {i f} \mathbf {x} _ {j} \in \mathrm {k N N} (\mathbf {x}) \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +Then, we solve linear system + +$$ +\hat {\mathbf {z}} _ {\mathbf {x}} = L ^ {- 1} \mathbf {y} _ {\mathbf {x}}, \tag {9} +$$ + +as in the dual formulation (5) and get a prediction with $\hat{y}_{\mathbf{x}} = \operatorname{argmax}_c \hat{\mathbf{z}}_{\mathbf{x}}(c)$ . With the usual formulation of label propagation in (4), $C$ linear systems need to be solved to get prediction for a single test example. The dual formulation allows us to do it by solving only a single linear system. + +Fast inductive inference with sparsification. We further introduce an additional off-line step, where we solve (4), get $\hat{\mathbf{y}}_c$ for all $c\in \{1,\dots,C\}$ , and store them in a matrix $\hat{Y} = [\hat{\mathbf{y}}_1;\dots;\hat{\mathbf{y}}_c]\in \mathbb{R}^{N\times C}$ . Then, the solution for a test example is equivalent to a weighted sum of rows of $\hat{Y}$ [1], which is a byproduct of using the indicator vector (8) for representing a test example. Its prediction is given by $\hat{\mathbf{z}}_{\mathbf{x}} = \mathbf{y}_{\mathbf{x}}^{T}\hat{Y}$ , and is equivalent to that obtained via (9). However, storing the whole $\hat{Y}$ can be expensive for very large values of $N$ and $C$ . We propose to sparsify $\hat{Y}$ by keeping only the largest values in each row, column, or over the whole matrix. + +Note that [14] proposes a low-rank decomposition of the inverse graph Laplacian for the task of retrieval. Our solution is tailored to zero-shot classification, and we choose to obtain and sparsify the first $C$ rows of $L_{\mathrm{inv}}$ instead of approximating the whole matrix. Additionally, our solution requires one (sparse) vector to matrix multiplications at test-time instead of two. + +# 3.3. LP for zero-shot VLM classification + +We are given a set of classes $\mathcal{C}$ with extracted VLM representations $\{\mathbf{w}_1,\dots,\mathbf{w}_C\}$ and a set of unlabeled images $\mathcal{U}$ with extracted VLM representations $\{\mathbf{u}_1,\dots,\mathbf{u}_M\}$ . We use them as nodes $\{\mathbf{w}_1,\dots,\mathbf{w}_C,\mathbf{u}_1,\dots,\mathbf{u}_M\}$ of the graph for label propagation. Nodes of class representations (text nodes) are labeled and image nodes unlabeled. + +To construct the adjacency matrix $S$ , we need to perform the $k$ -nearest neighbor search between nodes. However, it is known that there exists a large modality gap between image and text representations coming from VLMs [21, 38, 47]. The respective similarity distributions for CLIP are shown in Figure 3a. This modality gap makes standard kNN search between nodes not useful for label propagation; image nodes mostly get connected to image nodes, and text nodes mostly get connected to text nodes. As a consequence, few edges exist between labeled and unlabeled nodes. + +To alleviate this problem, we perform the kNN search separately for connecting image nodes to image nodes and for connecting image nodes and text nodes. We do not perform the search using text nodes as queries, i.e. text nodes get linked only if they appear in the kNN list of an image. This way we also avoid linking text nodes with each other, which is beneficial as each of them is labeled to a different class. Formally, the values of the adjacency matrix are + +$$ +s _ {i j} = \left\{ \begin{array}{l l} \mathbf {u} _ {i} ^ {T} \mathbf {u} _ {j}, & \text {i f} \mathbf {u} _ {j} \in \mathrm {k N N} _ {\mathbf {u}} (\mathbf {u} _ {i}) \\ \mathbf {u} _ {i} ^ {T} \mathbf {w} _ {j}, & \text {i f} \mathbf {w} _ {j} \in \mathrm {k N N} _ {\mathbf {w}} (\mathbf {u} _ {i}) \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {10} +$$ + +where $\mathrm{kNN_u}$ and $\mathrm{kNN_w}$ denote that the search is performed within the image or class features only, respectively. + +Moreover, during inductive inference for image $\mathbf{u}$ , we perform the kNN search in a similar way to construct indicator vector $\mathbf{y_u}$ whose elements are given by + +$$ +\mathbf {y} _ {\mathbf {u}} (i) = \left\{ \begin{array}{l l} \mathbf {u} ^ {T} \mathbf {u} _ {j}, & \text {i f} \mathbf {u} _ {j} \in \mathrm {k N N} _ {\mathbf {u}} (\mathbf {u}) \\ \mathbf {u} ^ {T} \mathbf {w} _ {j}, & \text {i f} \mathbf {w} _ {j} \in \mathrm {k N N} _ {\mathbf {w}} (\mathbf {u}) \\ 0, & \text {o t h e r w i s e .} \end{array} \right. \tag {11} +$$ + +Due to the two types of edges, i.e. image-to-image and image-to-text, we use power function $h(v) = v^{\gamma}$ to transform the image-to-text (cross-modal) similarities. This way we effectively balance their contribution in the graph and the indicator vector. To that end, we use $h(\mathbf{u}_i^T\mathbf{w}_j)$ and $h(\mathbf{u}^T\mathbf{w}_j)$ in (10) and (11), respectively, instead of $\mathbf{u}_i^T\mathbf{w}_j$ and $\mathbf{u}^T\mathbf{w}_j$ . + +We refer to the proposed method described above as Zero-shot classification with Label Propagation (ZLaP). We further denote the variant of our method after sparsifying the $\hat{Y}$ matrix for inductive inference as ZLaP*. + +![](images/4bd6274117ce7f7e4af3e2e04bbac872ce76af67fcf387ef6b4f212267476378.jpg) +Figure 2. t-SNE visualization for the original CLIP features (left) and our geodesic similarity (right). The former is estimated with the features as input, while the latter with the $L_{\mathrm{inv}}$ used as a pairwise similarity matrix. $\star$ : class representation, $\bullet$ : image representation. Figure generated for five random classes from the CUB dataset. + +![](images/b55e305adcc9efe1fa3a9873ee053366b8234b3efef992554aa7775906de3fd3.jpg) + +![](images/1dbfac7a1d4b7eeb79c4225058341baba1894e929048f46f7d06ff05a5eee482.jpg) +(a) Using text prompts +Figure 3. Similarity distributions among features of the same or different modality, using 7 textual templates [38] (left) or the InMaP proxies (right) as class representations. + +![](images/d905689a541494f1b8041d6ce8d311881458b283578130593244ceea09e925a7.jpg) +(b) Using proxies from InMaP [30] + +$t$ -SNE visualization of the bi-modal space. In Figure 2 we visualize the bi-modal feature space for CLIP features using $t$ -SNE [39] in two cases, i.e. the Euclidean case and using geodesic similarities obtained by $L_{\mathrm{inv}}$ , i.e. after label propagation. When using Euclidean affinities (left), we see that due to the large differences in the similarity distributions (text-text, image-to-image, and text-to-image, as shown in Figure 3a) all class representations (stars) are clustered together far from the image nodes. However, using the geodesic affinities from $L_{\mathrm{inv}}$ (right) we see that class representations are more spread. + +# 4. Experiments + +In this section, we first present the datasets we use, our experimental setup and competing methods. We then present component analysis for ZLaP and results for transductive and inductive zero-shot classification on 14 datasets. + +# 4.1. Datasets + +We evaluate the proposed method on 14 diverse image classification datasets: ImageNet ILSVRC2012 [33], + +Describable Textures Dataset (DTD) [5], EuroSAT [10], FGVC-Aircraft [23], Oxford Flowers 102 [25], Food101 [2], Oxford-IIIT Pet [28], SUN397 [42], Stanford Cars [17], Caltech101 [6], UCF101 [36], CIFAR10 [18], CIFAR100 [18], CUB-200-2011 [41]. For the first 11 datasets we borrow the train and test splits from CoOp[45]. We use the official training and test splits for CIFAR10, CIFAR100 and CUB-200-2011. + +# 4.2. Experimental setup + +In the transductive (inductive) inference setup, unlabeled nodes in the graph are the test (train) images. We always measure classification accuracy over the test images. + +VLMs and textual prompts. We report results using the publicly available ResNet50 and ViT-B/16 CLIP [31] models. We adopt the 7 templates from SuS-X [38] as class prompts for all results apart from Table 3 where we utilize the LLM generated prompts from [29]. + +Compared methods. Our baseline is zero-shot recognition with CLIP [31] using text encoder features as class representations. TPT [35] is based on test-time prompt tuning such that different image augmentations produce consistent predictions. The aforementioned methods do not exploit unlabeled data; their performance is therefore unchanged in both inference setups. CLIP-DN [46] normalizes feature distributions during test-time and assumes access to the mean feature vector of the target distribution. In the transductive (inductive) setup the mean vector is estimated on the test (training) set. InMaP [30] is a concurrent work that extracts updated class representations using pseudo-labels on the unlabeled set. In the transductive (inductive) setup the learning is performed on the test (training) images. + +Implementation details. We reproduce results for $\mathrm{CLIP}^2$ , $\mathrm{CLIP - DN^3}$ , and $\mathrm{InMaP^4}$ using their public implementations. For TPT [35] we report the numbers provided in [30]. We run InMaP using a single set of hyper-parameters for all 14 datasets, i.e. the default values reported in the official implementation4. We also fix the values of $k$ , $\gamma$ , and $\alpha$ for ZLaP across all datasets to 5, 5.0, and 0.3, respectively, for CLIP, and 10, 3.0, and 0.3, respectively, for InMaP. + +ZLaP variants. We refer to ZLaP using text class representations as $\mathbf{CLIP} + \mathbf{ZLaP}$ . Since InMaP is complementary to our work, we further evaluate the performance of ZLaP when $\{\mathbf{w}_1,\dots ,\mathbf{w}_C\}$ are the InMaP proxies. We refer to this as InMaP + ZLaP in the results. We refer to ZLaP with a sparse $\hat{Y}$ for inductive inference as ZLaP*. + +# 4.3. Components of ZLaP + +Bi-modal graph adjustments. In Table 1 we show the importance of two design choices to adapt LP to bi-modal graphs, i.e. separating the nearest neighbor search across modalities using (10) and (11), and transforming cross-modal similarities using a power function $h(\cdot)$ . We see that separate search is crucial; without it LP is not effective at all. The power function gives an extra boost in both setups, especially in the case of transductive inference. In Table 2 we report the percentage of images that are connected to their groundtruth class nodes within a path of length $n$ , with and without our adjustments. We see that for any such paths to exist for the case without adjustments, $k$ needs to be extremely high. With adjustments, $k = 5$ is enough for $71.4\%$ of the nodes to be connected to the correct class nodes. + +Sparsifying matrix $\hat{Y}$ for inductive inference. We explore three ways of approximating $\hat{Y}$ by sparsification, i.e. wither keeping only the largest $\xi$ columns per row, the largest $\xi$ rows per column, or the largest $\xi$ elements of the whole matrix. In all cases, the rest of the elements are set to zero. In Figure 4 we show the influence that these three variants have on performance. Not only these variants speed-up inference, but we can also see improvements in performance when sparsification percentage is high, i.e. low percentage of non-zero elements. We attribute this to the fact that less confident predictions in $\hat{Y}$ , many of them erroneous, are now set to zero. Although the best variant to choose seems to vary per dataset, we found that keeping the top element per row performs well across different datasets. We therefore use the $\xi = 1$ top element per row for our experiments. This amounts to different percentages of sparsity per dataset; we are keeping approximately $2.3\%$ on average + +
Eq.(10)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X50.232.341.6
✓✓61.841.952.1
+ +(a) Transductive inference + +
Eq.(10)-(11)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X60.842.449.6
✓✓62.242.849.7
+ +(b) Inductive inference +Table 1. Adjusting LP to bi-modal graphs. Impact of using separate kNN search for constructing the graph (Eq.(10)) or the indicator vector (Eq.(11)), as well as power function $h(\cdot)$ for balancing the contributions of two types of edges in the graph. + +
kJoint kNN searchSeparate kNN search
n=1n=2n=3n=1n=2n=3
50.00.00.071.485.1100.0
100.00.00.082.995.4100.0
10040.1100.0100.0100.0100.0100.0
+ +Table 2. Impact of the separate kNN search on the shortest paths between image nodes and the text node of their class. We report the percentage of images whose shortest path to the text node of their ground-truth class has length equal to or less than $n$ . Left: the vanilla approach. Right: our separate kNN search using Eq. (10). Analysis on DTD for the transductive setup. + +across all datasets. Regarding the inference speed-up, the primal solution takes $\sim 2.6$ sec per image, the dual takes $\sim 4.4$ ms, while the sparsified approach takes $\sim 0.6$ ms, measured on ImageNet dataset. + +Using the class proxies from InMaP. We observe that many class-to-image similarities (e.g. $\mathbf{u}_i^\top \mathbf{w}_j$ in (10)) become negative on some datasets when using ZLaP with InMaP proxies (see Figure 3b). We therefore perform minmax normalization in range [0,1] after constructing adjacency matrix $S$ or the indicator vector, for the transductive and inductive inference setups. + +# 4.4. Results + +Transductive inference. We present results for transductive zero-shot classification in Figure 5. ZLaP improves the zero-shot performance of CLIP significantly on all datasets. It also outperforms the recent TPT and CLIP-DN approaches on the vast majority of cases, with large gains + +![](images/46306664df705b1ebca34c6625c2b182e9f4c706be33116713144b5162cdc9a2.jpg) +Figure 4. Sparcifying matrix $\hat{Y}$ for inductive CLIP+ZLaP: effect of maintaining only the top elements per row/column/matrix. + +![](images/a07735124b0c8ecbe47efd9a854d6781b9a92fd09dcbd7e8cb522ef2b281891e.jpg) + +![](images/e9157546dbf2f27e9da207d3371dedf55ee66739735207d69dfc74d161e1289f.jpg) + +![](images/be8905d402759a5c56f8b2941c49bca0d5b161c621c1e0a07eacbdccaeab26b4.jpg) + +![](images/e0521b9b73f6f88a5d39006d3ab6f9e2ec245249dd2214438c6e81ef88104fdc.jpg) + +![](images/bf838dc5c502bdef49613c3110f958819fa478ecdcb8972ce41a1563dc5569c9.jpg) +Figure 5. Zero-shot classification accuracy averaged over 14 datasets for the transductive (top) and inductive (bottom) setups. Results per dataset are reported in the supplementary material. + +![](images/201bd18f35e357ce7a37bae4c76f7e75e589d004f4c24c172956dcde0f1aae77.jpg) + +in average accuracy. Compared to InMaP, ZLaP offers lower accuracy on average. However, by incorporating InMaP's class representations to our graph, we can improve our results even further and outperform all other methods, for an improvement of approximately $+5\%$ over CLIP with both backbones. + +Inductive inference. We report results for the inductive inference setup in Figure 5. ZLaP achieves a noticeable im + +provements over the CLIP baseline in this setup as well. Gains are more prominent for the case of ZLaP using the InMaP proxies, where gains over CLIP are $+4.4\%$ and $+4.9\%$ for the two backbones. We also observe that, although InMaP slightly outperforms ZLaP when used over CLIP, the combination of the two achieves state-of-the-art performance in this case as well. We further see that $\mathrm{ZLaP^{*}}$ retains the state-of-the-art performance of our method, while sparsifying $\hat{Y}$ offers significant speed-up at inference time. + +
TransductiveInductive
Results with RN50
CLIP63.063.0
+ ZLaP64.664.2
InMaP64.864.6
+ ZLaP65.865.0
Results with ViT-B-16
CLIP71.971.9
+ ZLaP72.673.3
InMaP73.974.0
+ ZLaP74.874.2
+ +Leveraging LLM generated prompts. In Table 3 we report average zero-shot classification accuracy for ZLaP using the prompts recently proposed in CuPL [29]. These are prompts generated by LMMs that are available on the CuPL Github page5 for 12 of the datasets we use (all datasets besides CUB and Eurosat). ZLaP improves zero-shot performance in this case as well, for both the transductive and inductive setups. This verifies that our method is complementary to improved prompt engineering. + +Multi-label classification. We apply ZLaP for multi-label classification on the MS-COCO [22] dataset. ZLaP improves the zero-shot performance of CLIP by $+6.0\%$ mAP (56.8% vs. 50.8%) for inductive inference without any modification of the approach or its hyper-parameters. + +Web-crawled unlabeled images. All previous experiments use unlabeled images that come from the target distribution, i.e. they are known to depict one of the classes of interest but their labels are discarded. To see the impact of ZLaP in a more realistic setup using web-crawled images we rely on LAION-400M [34] composed by image-caption pairs. We construct the set of unlabeled images with 10,000 images per class that are chosen either randomly, or based on proximity of their image or text features to the class representation. Random selection fails, but the other two options provide some improvement compared to CLIP, with the caption-based neighbors being a bit better. The complete set of results is presented in the supplementary material. + +Table 3. Zero-shot classification using prompts generated by LLMs [29]. We report average accuracy on 12 datasets using prompts from CuPL [29] together with our 7 standard prompts. Results per dataset are reported in the supplementary material. + +
TransductiveInductive
BLIP [20]54.654.6
+ ZLaP59.657.9
ALBEF [19]36.036.0
+ ZLaP41.246.8
EVA-CLIP-8B [37]83.683.6
+ ZLaP84.684.5
EVA-CLIP-18B [37]83.983.9
+ ZLaP84.884.7
+ +Table 4. Accuracy on ImageNet using different VLMs. + +Different VLMs We use CLIP as the VLM of choice throughout our experiments. In Table 4, we present results when ZLaP is applied on top of four recent VLMs, namely BLIP [20], ALBEF [19], and two versions of EVA-CLIP [37]. We use the implementations of BLIP and ALBEF that are available in the LAVIS library6, while for EVA-CLIP we use implementation from the official Github repository7. ZLaP improves the results of all four different VLMs in both transductive and inductive setups. + +# 5. Conclusions + +Label propagation is an intuitive way of encoding the global structure of unlabeled data into geodesic distances over a locally Euclidean space. In this paper, we show that this method can be successfully tailored to both transductive and inductive zero-shot classification with vision-language models, and achieve state-of-the-art performance on both setups. To that end, we show that it is highly important to take proper care of the peculiarities of the bi-modal nature of the task during graph construction. We further carefully design an efficient variant of label propagation for the inductive inference case, that may enable label propagation to be applied to other tasks beyond zero-shot classification. + +Vision-language models trained on billion-scale datasets are redefining computer vision research. The proposed ZLaP is a training-free approach able to improve the generalization performance of black-box VLMs using only unlabeled data, for an annotation-free, text-based and open-world classification paradigm that will inevitably be ubiquitous in the near future. + +Acknowledgements. This work was supported by the Junior Star GACR GM 21-28830M and the Czech Technical University in Prague grant No. SGS23/173/OHK3/3T/13. We thank Ahmet Iscen for many helpful comments. + +# References + +[1] Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. Label propagation and quadratic criterion. In Semi-Supervised Learning, pages 192-216. The MIT Press, 2006. 4 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5 +[3] Siddhartha Chandra and Iasonas Kokkinos. Fast, exact and multi-scale inference for semantic image segmentation with deep Gaussian CRFs. In ECCV, 2016. 3 +[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2 +[5] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In CVPR, 2014. 5 +[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, 2004. 5 +[7] Leo Grady. Random walks for image segmentation. pami, 28(11):1768-1783, 2006. 3 +[8] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. CALIP: zero-shot enhancement of CLIP with parameter-free attention. In AAAI, 2023. 1, 2 +[9] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip H. S. Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In ICLR, 2023. 2 +[10] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 5 +[11] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, Cheng-Hao Kuo, and Ram Nevatia. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2 +[12] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2 +[13] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Teddy Furon, and Ondrej Chum. Efficient diffusion on region manifolds: Recovering small objects with compact CNN representations. In CVPR, 2017. 3, 4 +[14] Ahmet Iscen, Yannis Avrithis, Giorgos Tolias, Teddy Furon, and Ondrej Chum. Fast spectral ranking for similarity search. In CVPR, 2018. 4 +[15] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In CVPR, 2019. 3 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom + +Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2 +[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, 2013. 5 +[18] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5 +[19] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 8 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 8 +[21] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y. Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4 +[22] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 8 +[23] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[24] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1, 2 +[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5 +[26] Zachary Novack, Julian J. McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2 +[27] Omiros Pantazis, Gabriel J. Brostow, Kate E. Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. In BMVC, 2022. 2 +[28] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5 +[29] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 1, 2, 5, 8 +[30] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In NeurIPS, 2023. 1, 2, 5, 6 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5 +[32] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2 + +[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5 +[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 8 +[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 5, 6 +[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5 +[37] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP18B: scaling CLIP to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 8 +[38] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 4, 5 +[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 5 +[40] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. Neural priming for sample-efficient adaptation. In NeurIPS, 2023. 2 +[41] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 5 +[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5 +[43] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In CVPR, 2023. 2 +[44] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In NIPS, 2003. 3 +[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 5 +[46] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Test-time distribution normalization for contrastively learned vision-language models. In NeurlPS, 2023. 1, 2, 5 +[47] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 4 \ No newline at end of file diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/images.zip b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d1f8995270e5242328d62540019b3807a6892b85 --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdb10902ab2ecd0d12454cc2a00e9c3ac6c677b51082b243745282c997871dbc +size 359303 diff --git a/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/layout.json b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9840139dbf2db7aa0543dfbb864377024dac81c6 --- /dev/null +++ b/2024/Label Propagation for Zero-shot Classification with Vision-Language Models/layout.json @@ -0,0 +1,9849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 534, + 121 + ], + "type": "text", + "content": "Label Propagation for Zero-shot Classification with Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 142, + 215, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 142, + 215, + 158 + ], + "spans": [ + { + "bbox": [ + 136, + 142, + 215, + 158 + ], + "type": "text", + "content": "Vladan Stojnic1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 251, + 143, + 342, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 251, + 143, + 342, + 156 + ], + "spans": [ + { + "bbox": [ + 251, + 143, + 342, + 156 + ], + "type": "text", + "content": "Yannis Kalantidis²" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 378, + 143, + 455, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 143, + 455, + 157 + ], + "spans": [ + { + "bbox": [ + 378, + 143, + 455, + 157 + ], + "type": "text", + "content": "Giorgos Tolias1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 170, + 343, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 170, + 343, + 186 + ], + "spans": [ + { + "bbox": [ + 93, + 170, + 343, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 93, + 170, + 343, + 186 + ], + "type": "text", + "content": " VRG, FEE, Czech Technical University in Prague" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 375, + 171, + 495, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 171, + 495, + 186 + ], + "spans": [ + { + "bbox": [ + 375, + 171, + 495, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 375, + 171, + 495, + 186 + ], + "type": "text", + "content": " Naver Labs Europe" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 239, + 290, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 419 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have demonstrated impressive performance on zero-shot classification, i.e. classification when provided merely with a list of class names. In this paper, we tackle the case of zero-shot classification in the presence of unlabeled data. We leverage the graph structure of the unlabeled data and introduce ZLaP, a method based on label propagation (LP) that utilizes geodesic distances for classification. We tailor LP to graphs containing both text and image features and further propose an efficient method for performing inductive inference based on a dual solution and a sparsification step. We perform extensive experiments to evaluate the effectiveness of our method on 14 common datasets and show that ZLaP outperforms the latest related works. Code: https://github.com/vladan-stojnic/ZLaP" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 451, + 128, + 464 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 472, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 581 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) have demonstrated impressive performance on a variety of computer vision tasks. They are usually trained on large datasets of image-text pairs and contain visual and textual encoders that map to a common feature space. Visual encoders from such models have been shown to produce strong visual representations for perception tasks [31]. Given labeled data from a downstream dataset, one can fine-tune the model or learn classifiers and achieve really high classification accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Besides using the visual encoder in isolation, the joint text and visual encoder feature space of VLMs enables us to define text-based \"classifiers\", e.g. using the class names as textual prompts. This means that we only need a list of class names to perform zero-shot classification for a target dataset, i.e. without access to any labeled images. Although utilizing priors or devising better textual prompts can improve zero-shot performance [8, 24, 29, 46], here, we are interested in the case where we further have access to unlabeled data. Our goal is to find the best way of utilizing such data for zero-shot classification." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 215, + 538, + 389 + ], + "blocks": [ + { + "bbox": [ + 314, + 215, + 538, + 389 + ], + "lines": [ + { + "bbox": [ + 314, + 215, + 538, + 389 + ], + "spans": [ + { + "bbox": [ + 314, + 215, + 538, + 389 + ], + "type": "image", + "image_path": "842539b56cbdadf340fdccbf2e30b77c762dfba22a17ecd241d32f03c02e10d8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 398, + 547, + 454 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 547, + 454 + ], + "type": "text", + "content": "Figure 1. Zero-shot classification performance over 14 datasets using the proposed ZLaP classifier over CLIP [31], as well as over the (concurrent) InMaP [30] approach. Our method offers performance gains for both transductive (left) and inductive (right) inference. Average accuracy over 14 common datasets is reported." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 662 + ], + "type": "text", + "content": "In this paper, we leverage the inherent structure of the unlabeled data represented by a proximity graph and apply label propagation (LP) between the text-based classifiers and unlabeled images to derive geodesic distances we then use for classification. We tailor LP to VLMs and graphs containing both text and image features, and show that without proper handling of the bimodality, vanilla application of LP fails dramatically. We introduce ZLaP, a novel classification method based on label propagation that can perform both transductive and inductive inference. We perform the former with the standard (primal) solution of LP for classification and devise a more efficient dual solution for the latter. Our method is not only highly effective but also efficient, making LP a more attractive inductive classifier in terms of complexity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We implement our methods using publicly available VLMs as feature encoders, primarily the ResNet and ViT CLIP [31] models, and perform extensive experiments to evaluate the effectiveness of our method on 14 common" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23209" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "datasets. We show that we are able to achieve top performance on two zero-shot inference setups, i.e. inductive and transductive inference. Figure 1 summarizes our gains over 14 datasets on both setups when applying our LP-powered classifiers on top of CLIP, as well as after incorporating the class proxies from the recent InMaP [30] zero-shot approach." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 286, + 227 + ], + "type": "text", + "content": "It is worth highlighting that ZLaP is a non-parametric method, i.e. it does not involve a learning step. In fact, our approach does not even require access to the VLM model weights and can therefore be used to improve the zero-shot performance of a black-box model even, e.g. provided only via an API. In summary, our contributions are as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 286, + 352 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 228, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 262 + ], + "type": "text", + "content": "- We tailor label propagation to VLMs and zero-shot classification over bi-modal graphs, proposing per modality neighbor search and balancing of contributions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 267, + 286, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 267, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 267, + 286, + 314 + ], + "type": "text", + "content": "- We propose an efficient way for performing inductive inference with label propagation via a dual solution and through sparsification. This not only improves the test-time efficiency of our method but also performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 318, + 286, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 286, + 352 + ], + "type": "text", + "content": "- We complement our method with the class proxies presented in concurrent work [30] and achieve state of the art results for zero-shot on 14 common datasets." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 365, + 130, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 130, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 130, + 376 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 456 + ], + "type": "text", + "content": "In this section, we discuss related works that improve the already impressive zero-shot classification performance of vision-language models [4, 16, 31] even further. This is achieved by devising better distance metrics, utilizing external knowledge to learn more expressive textual prompts, or by leveraging synthetic and unlabeled data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 286, + 638 + ], + "type": "text", + "content": "Improved distance metrics. Zero-shot classification can be improved by devising a better distance metric between image and text representations [8, 46]. CALIP [8] uses a parameter-free attention mechanism and a local patch representation, i.e. instead of global representations, to improve the estimation of class-to-image similarity. CLIP-DN [46] improves the test-time similarity estimation by alignment with the similarity used during contrastive pre-training of VLMs. To achieve this, the method assumes access to unlabeled data from the target distribution. TPT [35] optimizes textual prompts via a consistency objective across test image augmentations. Our method can be considered a part of this line of work, as label propagation is a similarity measure in a geodesic space instead of the Euclidean space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "Improved textual prompts using language models. CLIP [31] uses hand-crafted prompts that are specialized for each domain. Instead of using hand-crafted prompts, generating them with large language models (LLMs) is shown to be promising [24, 29]. VisDesc [24] and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "CuPL [29] query LLMs to generate diverse descriptions of all classes, while WaffleCLIP [32] operates on top of Vis-Desc to systematically analyze which parts of the generated prompts are the most important. Instead of generating class descriptions, CHiLS [26] targets diversifying the set of classes, by generating sub-classes per class, either through an existing class hierarchy or by querying an LLM. It then performs zero-shot classification using sub-classes and linking them to the parent class. We show that methods improving the textual prompts are complementary to our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 314 + ], + "type": "text", + "content": "Synthetic data. Recent methods [9, 38, 43] demonstrate that the use of synthetic data is beneficial for zero-shot classification. CLIP+SYN [9] uses a stable-diffusion-based model to generate synthetic images using class names and uses them to train a linear classifier, initialized by the VLM class representations. SuS-X [38] considers a similar approach, but relies on a non-parametric classifier. CaFO [43] follows the same path, but additionally includes text prompts generated by LLMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 545, + 436 + ], + "type": "text", + "content": "External datasets. Besides the use of synthetic data, SuS-X proposes a variant that operates on an extensive unlabeled image dataset (LAION-5B [34]). This dataset encompasses a distribution that is a super set of the target one. The method generates pseudo-labels using the zero-shot approach which are then incorporated within the non-parametric classifier. NeuralPriming [40] additionally assumes that images have captions, which are used to improve the pseudo-labeling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 713 + ], + "type": "text", + "content": "Unlabeled images from the target distribution. Another line of research [11, 12, 27, 30] propose operating on unlabeled datasets from the target distribution. The main ingredient of all these methods is the prediction of pseudo-labels for unlabeled examples, that are later used for further processing. UPL [12] optimizes learnable text prompts based on the pseudo-labels. SVL-Adapter [27] first trains a self-supervised model on unlabeled data, and then an adapter module to align its outputs to the pseudolabels. ReCLIP [11] performs transductive label propagation to obtain the pseudo-labels and uses them to fine-tune the VLM visual and textual encoders. In contrast to that, we do not fine-tune the model, which we may not have access to, and efficiently use label propagation for inductive inference too. InMaP [30] is a concurrent work that uses pseudo-labels to update the class representations such that they are now closer to image representations. We show in the experiments that this approach is complementary to ours. In contrast to all those methods, we do not explicitly require pseudo-label prediction, but rather capture interactions between all unlabeled examples through a proximity graph and label propagation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23210" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 140 + ], + "type": "text", + "content": "We first define the task of zero-shot classification with access to unlabeled examples, present label propagation with our contributions and then present the proposed approach for zero-shot classification using unlabeled examples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 148, + 168, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 148, + 168, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 148, + 168, + 159 + ], + "type": "text", + "content": "3.1. Problem formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": "Vision-language models consist of an image encoder " + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "inline_equation", + "content": "f: \\mathcal{I} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "inline_equation", + "content": "g: \\mathcal{T} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": " represent the space of images and text, respectively. We consider the outputs of these encoders to be " + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 166, + 287, + 214 + ], + "type": "text", + "content": "-normalized." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": " denote a set of known classes with associated class names " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\{l_1,\\dots,l_C\\}" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_1,\\dots,p_P\\}" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": " a set of prompt templates. Each prompt is combined with a class name to produce a textual description of the class, i.e. " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "p_i(l_c)" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": "-th template used for class " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": ". Class representations " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_c = 1 / P\\sum_{i = 1}^{P}g(p_i(l_c))" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": " are obtained using the VLM. Then, for a test image " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": ", we extract representation " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = f(u)" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": ", and perform zero-shot classification by " + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathrm{argmax}_c\\mathbf{u}^T\\mathbf{w}_c" + }, + { + "bbox": [ + 47, + 214, + 287, + 310 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "text", + "content": "We further assume access to a set " + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "text", + "content": " unlabeled images. Let " + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{u}_1,\\dots ,\\mathbf{u}_M\\}" + }, + { + "bbox": [ + 47, + 310, + 287, + 346 + ], + "type": "text", + "content": " denote the representations of the unlabeled images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 346, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 287, + 407 + ], + "type": "text", + "content": "In this work, we assume no direct access to the VML model weights, i.e. VLM training is neither possible nor desired. This allows to consider the underlying VLM as a black box, possibly only available through an API that generates features. We consider two inference setups:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 407, + 287, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 287, + 454 + ], + "type": "text", + "content": "Inductive inference: We consider an inductive setup, where we need to construct a classifier that can operate on new examples. This classifier should take advantage of the unlabeled examples in " + }, + { + "bbox": [ + 47, + 407, + 287, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 47, + 407, + 287, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "text", + "content": "Transductive inference: We consider " + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "text", + "content": " to be the test set, i.e. all test examples are jointly provided in advance. Prediction for test example " + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "text", + "content": " may therefore depend on the representations and predictions of all other test examples. In this transductive setup the models are not required to provide predictions for any example that is not in " + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 47, + 454, + 287, + 526 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 534, + 180, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 534, + 180, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 180, + 548 + ], + "type": "text", + "content": "3.2. Label propagation (LP)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_1, \\dots, \\mathbf{x}_N\\}" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": ", be a set of features for " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " examples. Each feature represents a graph node. We construct an adjacency matrix " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " with zero diagonal, and " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " equal to " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^\\top \\mathbf{x}_j" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " is in the " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": "-nearest neighbors of " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " (denoted by " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j \\in \\mathrm{kNN}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": "), and 0 otherwise. We obtain a symmetric adjacency matrix by " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\bar{S} = S + S^\\top" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": ", and its symmetrically normalized version by " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\hat{S} = D^{-\\frac{1}{2}}\\bar{S}D^{-\\frac{1}{2}}" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "D = \\mathrm{diag}(\\bar{S}\\mathbf{1}_N)" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " is the degree matrix, and " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_N" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " is the all-ones " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": "-dimensional vector. We assume the first " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " examples, and the corresponding nodes, to be labeled among " + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 553, + 287, + 685 + ], + "type": "text", + "content": " classes; each class is assigned to a single node1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 133 + ], + "type": "text", + "content": "Transductive inference. Label propagation [44] is originally proposed for the transductive inference setup; we need to predict labels for the unlabeled nodes of the graph. Given the normalized adjacency matrix " + }, + { + "bbox": [ + 306, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 306, + 72, + 545, + 133 + ], + "type": "text", + "content": ", label propagation is an iterative process given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 140, + 545, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 545, + 156 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {c} ^ {(t + 1)} = \\alpha \\hat {S} \\hat {\\mathbf {y}} _ {c} ^ {(t)} + (1 - \\alpha) \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\} \\tag {1}", + "image_path": "cd5a95c0146ffa5506e81cb857f0a4bd0301dbd895c47e1a065fc66daa0ed4c9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": "until convergence. Where " + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": " is a propagation hyper-parameter, " + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_c = \\mathbf{e}_c \\in \\{0,1\\}^N" + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": " is a one-hot vector with the non-zero element at index " + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": " is the current iteration. Prediction of the label for an unlabeled node " + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "inline_equation", + "content": "j \\in \\{C + 1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 164, + 545, + 227 + ], + "type": "text", + "content": " is then given by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 236, + 545, + 250 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 236, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 380, + 236, + 545, + 250 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {y}} _ {c} (j), \\tag {2}", + "image_path": "12b4a7afc53fdae10a56f4312e296a5db501a8066e875c2a574d1552a097bb63.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_c(j) = \\mathbf{e}_j^T\\hat{\\mathbf{y}}_c" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": "-th element of the vector " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_c" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": ". One can show [44] that this iterative solution is equivalent to solving " + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 258, + 545, + 295 + ], + "type": "text", + "content": " linear systems" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 305, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 305, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 365, + 305, + 545, + 319 + ], + "type": "interline_equation", + "content": "L \\hat {\\mathbf {y}} _ {c} = \\mathbf {y} _ {c} \\quad \\forall c \\in \\{1, \\dots , C \\}, \\tag {3}", + "image_path": "265e4a958432c7f176d4bcaaa6ea7adbb103d98b199165926e7a4c00bc240b57.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 327, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 327, + 545, + 352 + ], + "type": "inline_equation", + "content": "L = I - \\alpha \\tilde{S}" + }, + { + "bbox": [ + 304, + 327, + 545, + 352 + ], + "type": "text", + "content": " is the graph Laplacian. These linear systems have a closed-form solution" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 376, + 361, + 545, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 361, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 376, + 361, + 545, + 375 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} _ {c} = L ^ {- 1} \\mathbf {y} _ {c} = L _ {\\text {i n v}} \\mathbf {y} _ {c}. \\tag {4}", + "image_path": "8a864f169cef34e800a227f783a31da5f07eb8c2cf5cecb3e45f804a351fd412.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "text", + "content": "However, this closed-form solution is not practical for large datasets as the inverse graph Laplacian " + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "text", + "content": " is a non-sparse " + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "text", + "content": " matrix. For this reason it is usual [3, 7, 13, 15] to solve (3) using the conjugate-gradient (CG) method, which is known to be faster than running the iterative solution [13]. Using CG is possible because " + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 385, + 545, + 456 + ], + "type": "text", + "content": " is positive-definite." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "content": "Observe that (4) simply picks one of the columns of " + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "content": ". Matrix element " + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}(j, c)" + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "content": " is the confidence of example " + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 457, + 545, + 553 + ], + "type": "text", + "content": ". Its values are similarities, after label propagation, between each node pair. It is a type of geodesic similarity that captures the geometry of the feature space as this is indicated by the graph structure. Focusing on a classification task, we are only interested in similarities between an unlabeled example and a class node." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "content": "Dual solution. Herein, we show that solving " + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "content": " linear systems of the form in (4) to obtain predictions for all unlabeled nodes using (2) is equivalent to solving " + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "inline_equation", + "content": "N - C" + }, + { + "bbox": [ + 304, + 569, + 545, + 616 + ], + "type": "text", + "content": " linear systems of form" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 349, + 626, + 545, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 626, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 349, + 626, + 545, + 640 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {j} = L ^ {- 1} \\mathbf {e} _ {j} \\quad \\forall j \\in \\{C + 1, \\dots , N \\}, \\tag {5}", + "image_path": "30217dda52527ff220f67989628c88bd0453a1f1d4db735907d8caa87aef6f51.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 650, + 508, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 508, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 508, + 662 + ], + "type": "text", + "content": "and obtaining the unlabeled node prediction using" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 381, + 672, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 672, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 381, + 672, + 545, + 685 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {j} = \\operatorname {a r g m a x} _ {c} \\hat {\\mathbf {z}} _ {j} (c). \\tag {6}", + "image_path": "6e19e8cba5e538900d81a1523b072b024808c51d244d107ab0d99260061a8083.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "1The theoretical part described in this section holds for the case of more labeled examples per class too. We consider this specific case for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "simplicity of the presentation and because it corresponds to the task of zero-shot classification with unlabeled examples." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "23211" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 168, + 83 + ], + "type": "text", + "content": "This comes from the fact that" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 83, + 88, + 287, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 88, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 83, + 88, + 287, + 105 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {j} (c) = \\mathbf {e} _ {c} ^ {T} L ^ {- 1} \\mathbf {e} _ {j} = \\mathbf {e} _ {j} ^ {T} L ^ {- 1} \\mathbf {e} _ {c} = \\hat {\\mathbf {y}} _ {c} (j). \\tag {7}", + "image_path": "dff889d2cc40906f4a5e9e5ca32e6af57c353c008c3d0eeb22e1b66366101986.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 108, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 179 + ], + "type": "text", + "content": "Although we present the dual solution using the closed-form (4), the same holds with the CG solution of (3). Using the dual solution (5) is not practical for transductive learning as usually the unlabeled nodes are many more than the labeled ones. However, we show that this dual solution is efficiently used for inductive inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": "As discussed, we can view " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": " as a pairwise similarity matrix. The confidence of example " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": " belonging to class " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": ", due to symmetry of " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": ", is equivalently obtained either by " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}(j, c)" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}(c, j)" + }, + { + "bbox": [ + 47, + 180, + 287, + 240 + ], + "type": "text", + "content": ". This constitutes, an additional interpretation of the duality in the solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "text", + "content": "Inductive inference. Test examples now come individually and are not known during graph construction. A possible way to perform inductive inference is by adding the new node to the graph, which is expensive for a test-time operation as " + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "text", + "content": " would have to be updated for each new test example. Instead, inspired by [13] that uses LP for retrieval, we construct indicator vector " + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\mathbf{x}} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "text", + "content": " for test example " + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 254, + 287, + 348 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 352, + 287, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 352, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 91, + 352, + 287, + 384 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {\\mathbf {x}} (j) = \\left\\{ \\begin{array}{l l} \\mathbf {x} ^ {T} \\mathbf {x} _ {j}, & \\text {i f} \\mathbf {x} _ {j} \\in \\mathrm {k N N} (\\mathbf {x}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "ed4aa8c05637bf2d9316ed879538109eb3192f0ccc6b2187cd0331744fb6ec2b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 388, + 167, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 167, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 167, + 400 + ], + "type": "text", + "content": "Then, we solve linear system" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 405, + 287, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 405, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 137, + 405, + 287, + 419 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {\\mathbf {x}} = L ^ {- 1} \\mathbf {y} _ {\\mathbf {x}}, \\tag {9}", + "image_path": "dc1e515a82ff14a4163aa8dae0e3348cfe62bc19992d95c68f82ac81ec507c99.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": "as in the dual formulation (5) and get a prediction with " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\hat{y}_{\\mathbf{x}} = \\operatorname{argmax}_c \\hat{\\mathbf{z}}_{\\mathbf{x}}(c)" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": ". With the usual formulation of label propagation in (4), " + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 425, + 287, + 484 + ], + "type": "text", + "content": " linear systems need to be solved to get prediction for a single test example. The dual formulation allows us to do it by solving only a single linear system." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": "Fast inductive inference with sparsification. We further introduce an additional off-line step, where we solve (4), get " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{y}}_c" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "c\\in \\{1,\\dots,C\\}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ", and store them in a matrix " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Y} = [\\hat{\\mathbf{y}}_1;\\dots;\\hat{\\mathbf{y}}_c]\\in \\mathbb{R}^{N\\times C}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ". Then, the solution for a test example is equivalent to a weighted sum of rows of " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " [1], which is a byproduct of using the indicator vector (8) for representing a test example. Its prediction is given by " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathbf{x}} = \\mathbf{y}_{\\mathbf{x}}^{T}\\hat{Y}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ", and is equivalent to that obtained via (9). However, storing the whole " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " can be expensive for very large values of " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": ". We propose to sparsify " + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 498, + 287, + 628 + ], + "type": "text", + "content": " by keeping only the largest values in each row, column, or over the whole matrix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "content": "Note that [14] proposes a low-rank decomposition of the inverse graph Laplacian for the task of retrieval. Our solution is tailored to zero-shot classification, and we choose to obtain and sparsify the first " + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "content": " rows of " + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 630, + 287, + 712 + ], + "type": "text", + "content": " instead of approximating the whole matrix. Additionally, our solution requires one (sparse) vector to matrix multiplications at test-time instead of two." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 495, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 495, + 83 + ], + "type": "text", + "content": "3.3. LP for zero-shot VLM classification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": "We are given a set of classes " + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": " with extracted VLM representations " + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C\\}" + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": " and a set of unlabeled images " + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": " with extracted VLM representations " + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}" + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": ". We use them as nodes " + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_1,\\dots,\\mathbf{w}_C,\\mathbf{u}_1,\\dots,\\mathbf{u}_M\\}" + }, + { + "bbox": [ + 304, + 91, + 545, + 163 + ], + "type": "text", + "content": " of the graph for label propagation. Nodes of class representations (text nodes) are labeled and image nodes unlabeled." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "text", + "content": "To construct the adjacency matrix " + }, + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "text", + "content": ", we need to perform the " + }, + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 165, + 545, + 283 + ], + "type": "text", + "content": "-nearest neighbor search between nodes. However, it is known that there exists a large modality gap between image and text representations coming from VLMs [21, 38, 47]. The respective similarity distributions for CLIP are shown in Figure 3a. This modality gap makes standard kNN search between nodes not useful for label propagation; image nodes mostly get connected to image nodes, and text nodes mostly get connected to text nodes. As a consequence, few edges exist between labeled and unlabeled nodes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 381 + ], + "type": "text", + "content": "To alleviate this problem, we perform the kNN search separately for connecting image nodes to image nodes and for connecting image nodes and text nodes. We do not perform the search using text nodes as queries, i.e. text nodes get linked only if they appear in the kNN list of an image. This way we also avoid linking text nodes with each other, which is beneficial as each of them is labeled to a different class. Formally, the values of the adjacency matrix are" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 349, + 392, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 392, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 349, + 392, + 545, + 437 + ], + "type": "interline_equation", + "content": "s _ {i j} = \\left\\{ \\begin{array}{l l} \\mathbf {u} _ {i} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u} _ {i}) \\\\ \\mathbf {u} _ {i} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u} _ {i}) \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {10}", + "image_path": "7f8046e0283c27b370b377eefe165e10af87aac123c9870b1f5b14e4c6d912c1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{kNN_u}" + }, + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{kNN_w}" + }, + { + "bbox": [ + 306, + 448, + 545, + 471 + ], + "type": "text", + "content": " denote that the search is performed within the image or class features only, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "content": "Moreover, during inductive inference for image " + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "content": ", we perform the kNN search in a similar way to construct indicator vector " + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{y_u}" + }, + { + "bbox": [ + 304, + 472, + 545, + 509 + ], + "type": "text", + "content": " whose elements are given by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 344, + 519, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 519, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 344, + 519, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {\\mathbf {u}} (i) = \\left\\{ \\begin{array}{l l} \\mathbf {u} ^ {T} \\mathbf {u} _ {j}, & \\text {i f} \\mathbf {u} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {u}} (\\mathbf {u}) \\\\ \\mathbf {u} ^ {T} \\mathbf {w} _ {j}, & \\text {i f} \\mathbf {w} _ {j} \\in \\mathrm {k N N} _ {\\mathbf {w}} (\\mathbf {u}) \\\\ 0, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}", + "image_path": "69b5093d09a09e113d335773327345e22a49b3b5e1a5d46cfab5f7b2a97b2300.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "Due to the two types of edges, i.e. image-to-image and image-to-text, we use power function " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "h(v) = v^{\\gamma}" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " to transform the image-to-text (cross-modal) similarities. This way we effectively balance their contribution in the graph and the indicator vector. To that end, we use " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "h(\\mathbf{u}_i^T\\mathbf{w}_j)" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "h(\\mathbf{u}^T\\mathbf{w}_j)" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " in (10) and (11), respectively, instead of " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i^T\\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^T\\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 580, + 545, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We refer to the proposed method described above as Zero-shot classification with Label Propagation (ZLaP). We further denote the variant of our method after sparsifying the " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " matrix for inductive inference as ZLaP*." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "23212" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 251, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 251, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 251, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 251, + 210 + ], + "type": "image", + "image_path": "4bd6274117ce7f7e4af3e2e04bbac872ce76af67fcf387ef6b4f212267476378.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "text", + "content": "Figure 2. t-SNE visualization for the original CLIP features (left) and our geodesic similarity (right). The former is estimated with the features as input, while the latter with the " + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "text", + "content": " used as a pairwise similarity matrix. " + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "text", + "content": ": class representation, " + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 46, + 223, + 547, + 258 + ], + "type": "text", + "content": ": image representation. Figure generated for five random classes from the CUB dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 343, + 76, + 539, + 211 + ], + "blocks": [ + { + "bbox": [ + 343, + 76, + 539, + 211 + ], + "lines": [ + { + "bbox": [ + 343, + 76, + 539, + 211 + ], + "spans": [ + { + "bbox": [ + 343, + 76, + 539, + 211 + ], + "type": "image", + "image_path": "b55e305adcc9efe1fa3a9873ee053366b8234b3efef992554aa7775906de3fd3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 54, + 279, + 164, + 378 + ], + "blocks": [ + { + "bbox": [ + 54, + 279, + 164, + 378 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 164, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 164, + 378 + ], + "type": "image", + "image_path": "1dbfac7a1d4b7eeb79c4225058341baba1894e929048f46f7d06ff05a5eee482.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 383, + 144, + 393 + ], + "lines": [ + { + "bbox": [ + 69, + 383, + 144, + 393 + ], + "spans": [ + { + "bbox": [ + 69, + 383, + 144, + 393 + ], + "type": "text", + "content": "(a) Using text prompts" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 402, + 288, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 436 + ], + "type": "text", + "content": "Figure 3. Similarity distributions among features of the same or different modality, using 7 textual templates [38] (left) or the InMaP proxies (right) as class representations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 173, + 279, + 282, + 378 + ], + "blocks": [ + { + "bbox": [ + 173, + 279, + 282, + 378 + ], + "lines": [ + { + "bbox": [ + 173, + 279, + 282, + 378 + ], + "spans": [ + { + "bbox": [ + 173, + 279, + 282, + 378 + ], + "type": "image", + "image_path": "d905689a541494f1b8041d6ce8d311881458b283578130593244ceea09e925a7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 382, + 284, + 393 + ], + "lines": [ + { + "bbox": [ + 169, + 382, + 284, + 393 + ], + "spans": [ + { + "bbox": [ + 169, + 382, + 284, + 393 + ], + "type": "text", + "content": "(b) Using proxies from InMaP [30]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": "-SNE visualization of the bi-modal space. In Figure 2 we visualize the bi-modal feature space for CLIP features using " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": "-SNE [39] in two cases, i.e. the Euclidean case and using geodesic similarities obtained by " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": ", i.e. after label propagation. When using Euclidean affinities (left), we see that due to the large differences in the similarity distributions (text-text, image-to-image, and text-to-image, as shown in Figure 3a) all class representations (stars) are clustered together far from the image nodes. However, using the geodesic affinities from " + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{inv}}" + }, + { + "bbox": [ + 46, + 455, + 287, + 587 + ], + "type": "text", + "content": " (right) we see that class representations are more spread." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 616, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 665 + ], + "type": "text", + "content": "In this section, we first present the datasets we use, our experimental setup and competing methods. We then present component analysis for ZLaP and results for transductive and inductive zero-shot classification on 14 datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 671, + 111, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 111, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 111, + 683 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "We evaluate the proposed method on 14 diverse image classification datasets: ImageNet ILSVRC2012 [33]," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 278, + 547, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 547, + 374 + ], + "type": "text", + "content": "Describable Textures Dataset (DTD) [5], EuroSAT [10], FGVC-Aircraft [23], Oxford Flowers 102 [25], Food101 [2], Oxford-IIIT Pet [28], SUN397 [42], Stanford Cars [17], Caltech101 [6], UCF101 [36], CIFAR10 [18], CIFAR100 [18], CUB-200-2011 [41]. For the first 11 datasets we borrow the train and test splits from CoOp[45]. We use the official training and test splits for CIFAR10, CIFAR100 and CUB-200-2011." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 386, + 421, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 421, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 421, + 399 + ], + "type": "text", + "content": "4.2. Experimental setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 406, + 547, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 442 + ], + "type": "text", + "content": "In the transductive (inductive) inference setup, unlabeled nodes in the graph are the test (train) images. We always measure classification accuracy over the test images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 546, + 525 + ], + "type": "text", + "content": "VLMs and textual prompts. We report results using the publicly available ResNet50 and ViT-B/16 CLIP [31] models. We adopt the 7 templates from SuS-X [38] as class prompts for all results apart from Table 3 where we utilize the LLM generated prompts from [29]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "Compared methods. Our baseline is zero-shot recognition with CLIP [31] using text encoder features as class representations. TPT [35] is based on test-time prompt tuning such that different image augmentations produce consistent predictions. The aforementioned methods do not exploit unlabeled data; their performance is therefore unchanged in both inference setups. CLIP-DN [46] normalizes feature distributions during test-time and assumes access to the mean feature vector of the target distribution. In the transductive (inductive) setup the mean vector is estimated on the test (training) set. InMaP [30] is a concurrent work that extracts updated class representations using pseudo-labels on the unlabeled set. In the transductive (inductive) setup the learning is performed on the test (training) images." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23213" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "Implementation details. We reproduce results for " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - DN^3}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{InMaP^4}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " using their public implementations. For TPT [35] we report the numbers provided in [30]. We run InMaP using a single set of hyper-parameters for all 14 datasets, i.e. the default values reported in the official implementation4. We also fix the values of " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " for ZLaP across all datasets to 5, 5.0, and 0.3, respectively, for CLIP, and 10, 3.0, and 0.3, respectively, for InMaP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": "ZLaP variants. We refer to ZLaP using text class representations as " + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\mathbf{CLIP} + \\mathbf{ZLaP}" + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": ". Since InMaP is complementary to our work, we further evaluate the performance of ZLaP when " + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{w}_1,\\dots ,\\mathbf{w}_C\\}" + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": " are the InMaP proxies. We refer to this as InMaP + ZLaP in the results. We refer to ZLaP with a sparse " + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 182, + 287, + 255 + ], + "type": "text", + "content": " for inductive inference as ZLaP*." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 262, + 170, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 170, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 170, + 274 + ], + "type": "text", + "content": "4.3. Components of ZLaP" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": "Bi-modal graph adjustments. In Table 1 we show the importance of two design choices to adapt LP to bi-modal graphs, i.e. separating the nearest neighbor search across modalities using (10) and (11), and transforming cross-modal similarities using a power function " + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": ". We see that separate search is crucial; without it LP is not effective at all. The power function gives an extra boost in both setups, especially in the case of transductive inference. In Table 2 we report the percentage of images that are connected to their groundtruth class nodes within a path of length " + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": ", with and without our adjustments. We see that for any such paths to exist for the case without adjustments, " + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": " needs to be extremely high. With adjustments, " + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": " is enough for " + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "inline_equation", + "content": "71.4\\%" + }, + { + "bbox": [ + 46, + 280, + 289, + 448 + ], + "type": "text", + "content": " of the nodes to be connected to the correct class nodes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": "Sparsifying matrix " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " for inductive inference. We explore three ways of approximating " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " by sparsification, i.e. wither keeping only the largest " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " columns per row, the largest " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " rows per column, or the largest " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " elements of the whole matrix. In all cases, the rest of the elements are set to zero. In Figure 4 we show the influence that these three variants have on performance. Not only these variants speed-up inference, but we can also see improvements in performance when sparsification percentage is high, i.e. low percentage of non-zero elements. We attribute this to the fact that less confident predictions in " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": ", many of them erroneous, are now set to zero. Although the best variant to choose seems to vary per dataset, we found that keeping the top element per row performs well across different datasets. We therefore use the " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "\\xi = 1" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " top element per row for our experiments. This amounts to different percentages of sparsity per dataset; we are keeping approximately " + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 46, + 462, + 289, + 668 + ], + "type": "text", + "content": " on average" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 541, + 138 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 541, + 138 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 541, + 138 + ], + "type": "table", + "html": "
Eq.(10)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X50.232.341.6
✓✓61.841.952.1
", + "image_path": "23d092a06b029be3cf6c707f74583305bf2c57edb543ac4dd2b1cace93d9dc96.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 163, + 541, + 232 + ], + "blocks": [ + { + "bbox": [ + 382, + 140, + 467, + 149 + ], + "lines": [ + { + "bbox": [ + 382, + 140, + 467, + 149 + ], + "spans": [ + { + "bbox": [ + 382, + 140, + 467, + 149 + ], + "type": "text", + "content": "(a) Transductive inference" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 163, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 308, + 163, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 541, + 232 + ], + "type": "table", + "html": "
Eq.(10)-(11)h(·)ImageNetDTDCUB
XX0.12.10.5
X✓0.12.10.5
✓X60.842.449.6
✓✓62.242.849.7
", + "image_path": "379338db492160396706db8e8fc670427df5f9ed76290333d4986a1cd1d2bb4d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 314, + 307, + 539, + 381 + ], + "blocks": [ + { + "bbox": [ + 387, + 233, + 462, + 242 + ], + "lines": [ + { + "bbox": [ + 387, + 233, + 462, + 242 + ], + "spans": [ + { + "bbox": [ + 387, + 233, + 462, + 242 + ], + "type": "text", + "content": "(b) Inductive inference" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 305, + 252, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 305, + 252, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 252, + 545, + 297 + ], + "type": "text", + "content": "Table 1. Adjusting LP to bi-modal graphs. Impact of using separate kNN search for constructing the graph (Eq.(10)) or the indicator vector (Eq.(11)), as well as power function " + }, + { + "bbox": [ + 305, + 252, + 545, + 297 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 305, + 252, + 545, + 297 + ], + "type": "text", + "content": " for balancing the contributions of two types of edges in the graph." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 307, + 539, + 381 + ], + "lines": [ + { + "bbox": [ + 314, + 307, + 539, + 381 + ], + "spans": [ + { + "bbox": [ + 314, + 307, + 539, + 381 + ], + "type": "table", + "html": "
kJoint kNN searchSeparate kNN search
n=1n=2n=3n=1n=2n=3
50.00.00.071.485.1100.0
100.00.00.082.995.4100.0
10040.1100.0100.0100.0100.0100.0
", + "image_path": "a19322ce1549f7758a00a068a84fc8fcee522e4b76b48a1d357b7614458e1b20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 396, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 462 + ], + "type": "text", + "content": "Table 2. Impact of the separate kNN search on the shortest paths between image nodes and the text node of their class. We report the percentage of images whose shortest path to the text node of their ground-truth class has length equal to or less than " + }, + { + "bbox": [ + 305, + 396, + 545, + 462 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 396, + 545, + 462 + ], + "type": "text", + "content": ". Left: the vanilla approach. Right: our separate kNN search using Eq. (10). Analysis on DTD for the transductive setup." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "text", + "content": "across all datasets. Regarding the inference speed-up, the primal solution takes " + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\sim 2.6" + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "text", + "content": " sec per image, the dual takes " + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\sim 4.4" + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "text", + "content": " ms, while the sparsified approach takes " + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "inline_equation", + "content": "\\sim 0.6" + }, + { + "bbox": [ + 305, + 482, + 545, + 531 + ], + "type": "text", + "content": " ms, measured on ImageNet dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "Using the class proxies from InMaP. We observe that many class-to-image similarities (e.g. " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i^\\top \\mathbf{w}_j" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " in (10)) become negative on some datasets when using ZLaP with InMaP proxies (see Figure 3b). We therefore perform minmax normalization in range [0,1] after constructing adjacency matrix " + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": " or the indicator vector, for the transductive and inductive inference setups." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 635, + 362, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 362, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 362, + 647 + ], + "type": "text", + "content": "4.4. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 714 + ], + "type": "text", + "content": "Transductive inference. We present results for transductive zero-shot classification in Figure 5. ZLaP improves the zero-shot performance of CLIP significantly on all datasets. It also outperforms the recent TPT and CLIP-DN approaches on the vast majority of cases, with large gains" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "text", + "content": "https://github.com/OpenAI/CLIP \n" + }, + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "text", + "content": "https://github.com/fengyuli-dev/distribution-normalization \n" + }, + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 47, + 672, + 287, + 715 + ], + "type": "text", + "content": "https://github.com/idstcv/InMaP" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23214" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 231, + 215 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 231, + 215 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 231, + 215 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 231, + 215 + ], + "type": "image", + "image_path": "46306664df705b1ebca34c6625c2b182e9f4c706be33116713144b5162cdc9a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 224, + 529, + 239 + ], + "lines": [ + { + "bbox": [ + 59, + 224, + 529, + 239 + ], + "spans": [ + { + "bbox": [ + 59, + 224, + 529, + 239 + ], + "type": "text", + "content": "Figure 4. Sparcifying matrix " + }, + { + "bbox": [ + 59, + 224, + 529, + 239 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 59, + 224, + 529, + 239 + ], + "type": "text", + "content": " for inductive CLIP+ZLaP: effect of maintaining only the top elements per row/column/matrix." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 71, + 375, + 214 + ], + "blocks": [ + { + "bbox": [ + 232, + 71, + 375, + 214 + ], + "lines": [ + { + "bbox": [ + 232, + 71, + 375, + 214 + ], + "spans": [ + { + "bbox": [ + 232, + 71, + 375, + 214 + ], + "type": "image", + "image_path": "a07735124b0c8ecbe47efd9a854d6781b9a92fd09dcbd7e8cb522ef2b281891e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 538, + 215 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 538, + 215 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 538, + 215 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 538, + 215 + ], + "type": "image", + "image_path": "e9157546dbf2f27e9da207d3371dedf55ee66739735207d69dfc74d161e1289f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 248, + 285, + 385 + ], + "blocks": [ + { + "bbox": [ + 58, + 248, + 285, + 385 + ], + "lines": [ + { + "bbox": [ + 58, + 248, + 285, + 385 + ], + "spans": [ + { + "bbox": [ + 58, + 248, + 285, + 385 + ], + "type": "image", + "image_path": "be8905d402759a5c56f8b2941c49bca0d5b161c621c1e0a07eacbdccaeab26b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 247, + 527, + 385 + ], + "blocks": [ + { + "bbox": [ + 299, + 247, + 527, + 385 + ], + "lines": [ + { + "bbox": [ + 299, + 247, + 527, + 385 + ], + "spans": [ + { + "bbox": [ + 299, + 247, + 527, + 385 + ], + "type": "image", + "image_path": "e0521b9b73f6f88a5d39006d3ab6f9e2ec245249dd2214438c6e81ef88104fdc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 389, + 285, + 544 + ], + "blocks": [ + { + "bbox": [ + 59, + 389, + 285, + 544 + ], + "lines": [ + { + "bbox": [ + 59, + 389, + 285, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 389, + 285, + 544 + ], + "type": "image", + "image_path": "bf838dc5c502bdef49613c3110f958819fa478ecdcb8972ce41a1563dc5569c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 552, + 547, + 575 + ], + "lines": [ + { + "bbox": [ + 46, + 552, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 547, + 575 + ], + "type": "text", + "content": "Figure 5. Zero-shot classification accuracy averaged over 14 datasets for the transductive (top) and inductive (bottom) setups. Results per dataset are reported in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 298, + 389, + 527, + 544 + ], + "blocks": [ + { + "bbox": [ + 298, + 389, + 527, + 544 + ], + "lines": [ + { + "bbox": [ + 298, + 389, + 527, + 544 + ], + "spans": [ + { + "bbox": [ + 298, + 389, + 527, + 544 + ], + "type": "image", + "image_path": "201bd18f35e357ce7a37bae4c76f7e75e589d004f4c24c172956dcde0f1aae77.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 595, + 289, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 289, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 289, + 668 + ], + "type": "text", + "content": "in average accuracy. Compared to InMaP, ZLaP offers lower accuracy on average. However, by incorporating InMaP's class representations to our graph, we can improve our results even further and outperform all other methods, for an improvement of approximately " + }, + { + "bbox": [ + 46, + 595, + 289, + 668 + ], + "type": "inline_equation", + "content": "+5\\%" + }, + { + "bbox": [ + 46, + 595, + 289, + 668 + ], + "type": "text", + "content": " over CLIP with both backbones." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Inductive inference. We report results for the inductive inference setup in Figure 5. ZLaP achieves a noticeable im" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "content": "provements over the CLIP baseline in this setup as well. Gains are more prominent for the case of ZLaP using the InMaP proxies, where gains over CLIP are " + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "inline_equation", + "content": "+4.4\\%" + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "inline_equation", + "content": "+4.9\\%" + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "content": " for the two backbones. We also observe that, although InMaP slightly outperforms ZLaP when used over CLIP, the combination of the two achieves state-of-the-art performance in this case as well. We further see that " + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{ZLaP^{*}}" + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "content": " retains the state-of-the-art performance of our method, while sparsifying " + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 304, + 595, + 547, + 704 + ], + "type": "text", + "content": " offers significant speed-up at inference time." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23215" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 289, + 223 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 223 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 223 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 223 + ], + "type": "table", + "html": "
TransductiveInductive
Results with RN50
CLIP63.063.0
+ ZLaP64.664.2
InMaP64.864.6
+ ZLaP65.865.0
Results with ViT-B-16
CLIP71.971.9
+ ZLaP72.673.3
InMaP73.974.0
+ ZLaP74.874.2
", + "image_path": "3ca31eab13e5820cbcfe6218b688089d1afc5fe287625f570dbd495fe049a46b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 306, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 415 + ], + "type": "text", + "content": "Leveraging LLM generated prompts. In Table 3 we report average zero-shot classification accuracy for ZLaP using the prompts recently proposed in CuPL [29]. These are prompts generated by LMMs that are available on the CuPL Github page5 for 12 of the datasets we use (all datasets besides CUB and Eurosat). ZLaP improves zero-shot performance in this case as well, for both the transductive and inductive setups. This verifies that our method is complementary to improved prompt engineering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 499 + ], + "type": "text", + "content": "Multi-label classification. We apply ZLaP for multi-label classification on the MS-COCO [22] dataset. ZLaP improves the zero-shot performance of CLIP by " + }, + { + "bbox": [ + 46, + 438, + 287, + 499 + ], + "type": "inline_equation", + "content": "+6.0\\%" + }, + { + "bbox": [ + 46, + 438, + 287, + 499 + ], + "type": "text", + "content": " mAP (56.8% vs. 50.8%) for inductive inference without any modification of the approach or its hyper-parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 689 + ], + "type": "text", + "content": "Web-crawled unlabeled images. All previous experiments use unlabeled images that come from the target distribution, i.e. they are known to depict one of the classes of interest but their labels are discarded. To see the impact of ZLaP in a more realistic setup using web-crawled images we rely on LAION-400M [34] composed by image-caption pairs. We construct the set of unlabeled images with 10,000 images per class that are chosen either randomly, or based on proximity of their image or text features to the class representation. Random selection fails, but the other two options provide some improvement compared to CLIP, with the caption-based neighbors being a bit better. The complete set of results is presented in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 328, + 70, + 525, + 198 + ], + "blocks": [ + { + "bbox": [ + 46, + 237, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 237, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 288, + 283 + ], + "type": "text", + "content": "Table 3. Zero-shot classification using prompts generated by LLMs [29]. We report average accuracy on 12 datasets using prompts from CuPL [29] together with our 7 standard prompts. Results per dataset are reported in the supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 70, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 525, + 198 + ], + "type": "table", + "html": "
TransductiveInductive
BLIP [20]54.654.6
+ ZLaP59.657.9
ALBEF [19]36.036.0
+ ZLaP41.246.8
EVA-CLIP-8B [37]83.683.6
+ ZLaP84.684.5
EVA-CLIP-18B [37]83.983.9
+ ZLaP84.884.7
", + "image_path": "b621872e8c517a2962b13bc6e9f8982b1e06670762abf9953627c9ff67516488.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 205, + 531, + 217 + ], + "lines": [ + { + "bbox": [ + 321, + 205, + 531, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 205, + 531, + 217 + ], + "type": "text", + "content": "Table 4. Accuracy on ImageNet using different VLMs." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 346 + ], + "type": "text", + "content": "Different VLMs We use CLIP as the VLM of choice throughout our experiments. In Table 4, we present results when ZLaP is applied on top of four recent VLMs, namely BLIP [20], ALBEF [19], and two versions of EVA-CLIP [37]. We use the implementations of BLIP and ALBEF that are available in the LAVIS library6, while for EVA-CLIP we use implementation from the official Github repository7. ZLaP improves the results of all four different VLMs in both transductive and inductive setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 356, + 384, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 384, + 369 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 384, + 369 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 376, + 546, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 519 + ], + "type": "text", + "content": "Label propagation is an intuitive way of encoding the global structure of unlabeled data into geodesic distances over a locally Euclidean space. In this paper, we show that this method can be successfully tailored to both transductive and inductive zero-shot classification with vision-language models, and achieve state-of-the-art performance on both setups. To that end, we show that it is highly important to take proper care of the peculiarities of the bi-modal nature of the task during graph construction. We further carefully design an efficient variant of label propagation for the inductive inference case, that may enable label propagation to be applied to other tasks beyond zero-shot classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 520, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 604 + ], + "type": "text", + "content": "Vision-language models trained on billion-scale datasets are redefining computer vision research. The proposed ZLaP is a training-free approach able to improve the generalization performance of black-box VLMs using only unlabeled data, for an annotation-free, text-based and open-world classification paradigm that will inevitably be ubiquitous in the near future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 667 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 667 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the Junior Star GACR GM 21-28830M and the Czech Technical University in Prague grant No. SGS23/173/OHK3/3T/13. We thank Ahmet Iscen for many helpful comments." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 702, + 232, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 232, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 232, + 714 + ], + "type": "text", + "content": "5https://github.com/sarahpratt/CuPL" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 683, + 495, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 683, + 495, + 693 + ], + "spans": [ + { + "bbox": [ + 320, + 683, + 495, + 693 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 320, + 683, + 495, + 693 + ], + "type": "text", + "content": "https://github.com/salesforce/LAVIS" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 694, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 694, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 694, + 545, + 712 + ], + "type": "text", + "content": "https://github.com/baaivision/EVA/tree/master/EVA-CLIP-18B" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "23216" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Olivier Delalleau, and Nicolas Le Roux. Label propagation and quadratic criterion. In Semi-Supervised Learning, pages 192-216. The MIT Press, 2006. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 - mining discriminative components with random forests. In ECCV, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 190 + ], + "type": "text", + "content": "[3] Siddhartha Chandra and Iasonas Kokkinos. Fast, exact and multi-scale inference for semantic image segmentation with deep Gaussian CRFs. In ECCV, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "type": "text", + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 287, + 278 + ], + "type": "text", + "content": "[5] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In CVPR, 2014. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "text", + "content": "[6] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, 2004. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 345 + ], + "type": "text", + "content": "[7] Leo Grady. Random walks for image segmentation. pami, 28(11):1768-1783, 2006. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 389 + ], + "type": "text", + "content": "[8] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzheng Ma, Xu-peng Miao, Xuming He, and Bin Cui. CALIP: zero-shot enhancement of CLIP with parameter-free attention. In AAAI, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[9] Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip H. S. Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? In ICLR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 490 + ], + "type": "text", + "content": "[10] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[11] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, Cheng-Hao Kuo, and Ram Nevatia. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In WACV, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 578 + ], + "type": "text", + "content": "[12] Tony Huang, Jack Chu, and Fangyun Wei. Unsupervised prompt learning for vision-language models. arXiv preprint arXiv:2204.03649, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 623 + ], + "type": "text", + "content": "[13] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, Teddy Furon, and Ondrej Chum. Efficient diffusion on region manifolds: Recovering small objects with compact CNN representations. In CVPR, 2017. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 656 + ], + "type": "text", + "content": "[14] Ahmet Iscen, Yannis Avrithis, Giorgos Tolias, Teddy Furon, and Ondrej Chum. Fast spectral ranking for similarity search. In CVPR, 2018. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[15] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In CVPR, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 130, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 152 + ], + "type": "text", + "content": "[18] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, 2009. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 153, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 153, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 153, + 545, + 196 + ], + "type": "text", + "content": "[19] Junnan Li, Ramprasaath R. Selvaraju, Akhilesh Gotmare, Shafiq R. Joty, Caiming Xiong, and Steven Chu-Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 240 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven C. H. Hoi. BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[21] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y. Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 330 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 374 + ], + "type": "text", + "content": "[23] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 399 + ], + "type": "text", + "content": "[24] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[25] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 476 + ], + "type": "text", + "content": "[26] Zachary Novack, Julian J. McAuley, Zachary Chase Lipton, and Saurabh Garg. Chils: Zero-shot image classification with hierarchical label sets. In ICML, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "text", + "content": "[27] Omiros Pantazis, Gabriel J. Brostow, Kate E. Jones, and Oisin Mac Aodha. Svl-adapter: Self-supervised adapter for vision-language pretrained models. In BMVC, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 533 + ], + "type": "text", + "content": "[28] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In CVPR, 2012. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "text", + "content": "[29] Sarah M. Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In ICCV, 2023. 1, 2, 5, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 601 + ], + "type": "text", + "content": "[30] Qi Qian, Yuanhong Xu, and Juhua Hu. Intra-modal proxy learning for zero-shot visual categorization with clip. In NeurIPS, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 602, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 667 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[32] Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "text", + "content": "23217" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 700 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 116 + ], + "type": "text", + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 194 + ], + "type": "text", + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 196, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 196, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 196, + 286, + 239 + ], + "type": "text", + "content": "[35] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. In NeurIPS, 2022. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 240, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 240, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 240, + 286, + 282 + ], + "type": "text", + "content": "[36] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 285, + 286, + 327 + ], + "type": "text", + "content": "[37] Quan Sun, Jinsheng Wang, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, and Xinlong Wang. EVA-CLIP18B: scaling CLIP to 18 billion parameters. arXiv preprint arXiv:2402.04252, 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 331, + 286, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 362 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 362 + ], + "type": "text", + "content": "[38] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. Sus-x: Training-free name-only transfer of vision-language models. In ICCV, 2023. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 365, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 286, + 384 + ], + "type": "text", + "content": "[39] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 430 + ], + "type": "text", + "content": "[40] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, Ludwig Schmidt, and Ali Farhadi. Neural priming for sample-efficient adaptation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 432, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 432, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 49, + 432, + 286, + 463 + ], + "type": "text", + "content": "[41] Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. 2010. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 466, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 466, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 466, + 286, + 498 + ], + "type": "text", + "content": "[42] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 500, + 286, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 500, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 500, + 286, + 542 + ], + "type": "text", + "content": "[43] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 544, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 286, + 577 + ], + "type": "text", + "content": "[44] Dengyong Zhou, Olivier Bousquet, Thomas Navin Lal, Jason Weston, and Bernhard Scholkopf. Learning with local and global consistency. In NIPS, 2003. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 578, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 286, + 609 + ], + "type": "text", + "content": "[45] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 613, + 286, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 286, + 654 + ], + "type": "text", + "content": "[46] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser-Nam Lim. Test-time distribution normalization for contrastively learned vision-language models. In NeurlPS, 2023. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 658, + 286, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 286, + 700 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 286, + 700 + ], + "type": "text", + "content": "[47] Xiangyang Zhu, Renrui Zhang, Bowei He, Aojun Zhou, Dong Wang, Bin Zhao, and Peng Gao. Not all features matter: Enhancing few-shot clip with adaptive prior refinement. In ICCV, 2023. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 756 + ], + "type": "text", + "content": "23218" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_content_list.json b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1c60cda488c74d20f3bba72ebeba85b15f159fc --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_content_list.json @@ -0,0 +1,1450 @@ +[ + { + "type": "text", + "text": "Label-Efficient Group Robustness via Out-of-Distribution Concept Curation", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiwei Yang1 Anthony Z. Liu2 Robert Wolfe1 Aylin Caliskan1 Bill Howe1 \n1University of Washington, 2University of Michigan \n{yanyiwei, rwolfe3, aylin, billhowe}@uw.edu anthliu@umich.edu", + "bbox": [ + 125, + 179, + 844, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to $33.1\\%$ on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias.", + "bbox": [ + 75, + 300, + 472, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 799, + 209, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models are prone to capture spurious correlations, patterns that are predictive of the target class in training data but are inherently irrelevant to the classification task. For example, consider an image classification task to distinguish doctors from nurses. Suppose that in", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the training data, $95\\%$ of the doctors present as men, and $95\\%$ of the nurses present as women. A model trained with the standard empirical risk minimization (ERM) tends to classify images of men as doctors and images of women as nurses by learning a spurious correlation between gender-expression features (e.g., hairstyle, clothing) and occupation. While the model may achieve high average accuracy, it performs poorly on the minority groups (in this example, female-presenting doctors and male-presenting nurses), thus exhibiting low worst-group accuracy. Such spurious correlations happen in a wide variety of real-world tasks, including facial recognition, medical imaging, and language tasks [3, 8, 11, 17, 18, 23].", + "bbox": [ + 496, + 268, + 893, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches typically consider the setting in which each training (or validation) sample is explicitly assigned a group, represented as a pair $(y, a)$ where $y$ is a class label and $a$ is a value from the (potentially unknown) spurious attribute. For instance, Group DRO assumes the availability of group labels during training, and directly minimizes the worst-group loss [17]. Recognizing that obtaining labels requires expensive human annotation, several recent approaches avoid relying on group labels during training [7, 9, 19, 24]. These approaches train an ERM model to infer group labels, then a second model is trained to optimize for worst-group loss with the inferred labels. However, these approaches require validation labels for crucial hyperparameter tuning [9], since their effectiveness depends on the accuracy of the group inference. One exception is GEORGE [19], which assumes that the feature space of a model trained for the predictive task is easily separable by groups and learns group assignment via clustering. GEORGE then optimizes for worst-cluster accuracy without requiring labels. However, while GEORGE is effective when the input features are simple and cluster well (e.g. colored digits), its performance drops significantly when the input features are more complex and clusters become less well-defined (e.g., for human faces) [19]", + "bbox": [ + 496, + 472, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose Concept Correction (figure 1), a framework that instead uses a set of example images to represent a concept involved in the spurious correlation. These examples can be out-of-distribution and task-agnostic, as long as", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12426", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6b093bec8798c3637a5d3318f19a10f7f573de96c46fbb7b9f445e8753d83c1.jpg", + "image_caption": [ + "Figure 1. We show our main idea Concept Correction, a framework for correcting spurious attributes in biased models using concepts, set of examples representing some pattern. We show an example from the Waterbirds dataset [17], where (a) directly training an ERM results in a model that spuriously correlates background type with bird type. A practitioner can correct this by (b) finding images with land and water backgrounds and using concept correction. Notably, these example images can be obtained outside of the original Waterbirds dataset, such as using generative AI, and using search engines." + ], + "image_footnote": [], + "bbox": [ + 150, + 78, + 803, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "they contain the features representing the spurious attribute. For example, concept sets designed to counteract a spurious attribute of gender presentation may include a set of images of female-presenting people, but without any reference to the task-specific class label of doctor or nurse. Since the examples need not be selected from the training distribution, they are reusable across models and tasks, and practitioners can retrieve them from various sources such as a search engines, datasets on the Web, or image generation models, all of which require significantly less cognitive and technical effort than manually labeling in-distribution data. In fact, by using images generated with only two lines of prompts to the text-to-image generator Stable Diffusion [16], we show that an implementation of Concept Correction can achieve competitive results to state-of-the-art methods that require manual labeling of tens of thousands of data points [9, 17, 24].", + "bbox": [ + 75, + 406, + 472, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate an instance of Concept Correction via Concept DRO (CDRO), an approach that leverages concepts to infer group labels, then trains with a distributively robust optimization objective using the inferred labels. Similar to the existing two-stage methods, CDRO first trains a neural network with ERM. Then, exploiting the fact that the ERM model learns spurious attribute labels[7], we use the model to extract features from the concept sets, the training data, and the validation data. We then train a classifier to discriminate concepts then use the weights of this classifier to determine the extent to which each training sample exhibits that concept. These concept similarity scores are then used to infer a spurious attribute label for every data point in the training and validation sets, providing the information needed by state-of-the-art methods to improve", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "worst-group accuracy. In our example, CDRO trains a classifier on two sets of images representing the concepts of presenting-female and presenting-male, then uses the coefficients of this classifier to estimate the extent to which each training sample exhibits \"maleness\" vs. \"femaleness.\"", + "bbox": [ + 496, + 406, + 890, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate CDRO on three image classification datasets with spurious correlations: CMNIST [1], Waterbirds [17], and CelebA [10]. Among methods that do not assume spurious attribute labels are provided, CDRO improves worstgroup accuracy by up to $12.6\\%$ on Waterbirds and $33.1\\%$ on CelebA. CDRO also achieves competitive performance to methods that do assume access to spurious attribute labels. Finally, we show that CDRO is robust to the quality of concepts in two aspects: size of concept and distribution distance from training data.", + "bbox": [ + 496, + 482, + 892, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 517, + 632, + 821, + 646 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Concept Correction, a label-efficient framework that uses out-of-distribution sets of images to correct spurious correlations.", + "- We present CDRO, an instance of the Concept Correction framework that uses concepts to infer labels of spurious attributes and then trains with a distributively robust optimization objective.", + "- We show CDRO significantly outperforms ERM and GEORGE across three benchmarks while achieving competitive results to methods that require group labels during validation." + ], + "bbox": [ + 500, + 650, + 890, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 829, + 638, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first present the problem setting, then describe concepts and the Concept Activation Vector (CAV) introduced by Kim et al. [6].", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Problem Setting", + "text_level": 1, + "bbox": [ + 76, + 90, + 238, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a classification setting, where given an input $x \\in \\mathcal{X}$ , our goal is to predict $y \\in \\mathcal{Y}$ . Let $X = \\{x_{1}, \\ldots, x_{n}\\}$ , $Y = \\{y_{1}, \\ldots, y_{n}\\}$ be our training set of size $n$ . Each datapoint $x_{i}$ has a spurious attribute $a_{i}$ , which is unobserved during training. Each $x_{i}$ is assumed to belong to a group $g_{i}$ where $g_{i}$ is a pair $(y_{i}, a_{i})$ , the combination of a class label $y_{i}$ and a spurious attribute label $a_{i}$ . The spurious correlation problem occurs when accuracy across groups varies significantly; i.e., a model trained with ERM may achieve a high average accuracy, yet still performs poorly on some minority groups. We therefore evaluate classifiers on the accuracy of the worst-performing group.", + "bbox": [ + 75, + 113, + 468, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are interested in the setting where group information is not available for either the training set nor the validation set. Instead, we assume that users curate a set of images representing the concept without regard to any particular task. The human effort to curate the concept depends on the user's domain knowledge. For example, in Waterbirds, to curate a concept for 'landbird background,' one can generate images with the prompt 'a photo of land background'. However, if the user knows that land backgrounds of birds tend to be forests, the prompt can be adjusted accordingly.", + "bbox": [ + 75, + 296, + 470, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Concepts and Concept Activation Vectors", + "text_level": 1, + "bbox": [ + 76, + 457, + 431, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A concept is defined using a set of images. For example, a set of ocean and lake images can represent the concept water background. The concept examples need not be drawn from the training data and need not be relevant to the classification task.", + "bbox": [ + 75, + 481, + 468, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a concept $C$ and a layer $l$ of the neural network, a concept activation vector (CAV) $\\mathbf{v}_C^l \\in R^m$ [6] is defined as the vector normal to the hyperplane separating samples with a concept (the concept set $P_C$ ) and examples without a concept (the contrastive set $N$ ) in the model's activations. Given layer activations of both the concept set and the contrastive set, CAV can be obtained by training a linear classifier distinguishing the two sets of activations.", + "bbox": [ + 75, + 556, + 470, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Concept Correction", + "text_level": 1, + "bbox": [ + 76, + 693, + 267, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose Concept Correction, a framework that uses concepts to correct spurious correlations, without relying on manual labeling of in-distribution data. Instead, Concept Correction takes a set of a small number of examples as input and leverages the examples to infer spurious attribute information. Then, a robust model is trained with the inferred group of spurious attribute information using any worst-group loss minimization algorithms.", + "bbox": [ + 75, + 719, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Correction expects the concept examples to represent the spurious attribute. As described in the introduction, these images can be out-of-distribution and irrelevant to the prediction task, de-coupling concept curation from", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the model setting. In fact, using generative models, practitioners can easily generate many examples with just a few lines of prompting. In contrast to existing methods that require manual labeling of either training or validation samples, concept curation does not require access to the training data. Practitioners can reason about potential biases the model may exhibit then curate concept sets accordingly.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept Correction is compatible with any method that requires spurious attribute or group information, such as Group DRO [17], Just-Train-Twice [9], Correct-n-Contrast [24], and more. For example, a simple way of applying Concept Correction to Correct-n-Contrast and JustTrain-Twice is simply replacing the validation group labels with inferred labels using concepts, alleviating the need of manual labeling of a large amount of data.", + "bbox": [ + 496, + 198, + 893, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Concept DRO", + "text_level": 1, + "bbox": [ + 498, + 330, + 645, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present Concept DRO (CDRO), an implementation of Concept Correction on Group DRO. As illustrated in Figure 2, in stage 1, we first train an ERM model and then train a linear classifier on concepts representing the spurious attribute to infer group labels. In stage 2, we use Group DRO to train a robust model with the inferred group labels. We include further details on both stages below, and give pseudocode for CDRO in the Appendix.", + "bbox": [ + 496, + 353, + 893, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Stage 1: Inferring Labels of Spurious Attributes", + "text_level": 1, + "bbox": [ + 498, + 497, + 888, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train a neural network with ERM on the training data. Given a spurious attribute $\\mathcal{A} = \\{a_1,\\dots,a_m\\}$ , we curate two sets of examples for each concept $a_i$ : a concept set $C = \\{c_1,\\dots,c_k\\}$ , which represents the presence of the concept, and a contrastive set $N = \\{n_1,\\dots,n_k\\}$ , which represents the absence of the concept. We then use the ERM model as a feature extractor, by taking the first to the penultimate layer of the model, denoted as $f$ , to extract features from the two sets. We then train a linear classifier on the ERM features of the concept sets, $\\{f(c_1),\\dots,f(c_k)\\}$ and $\\{f(n_1),\\dots,f(n_k)\\}$ . The resulting coefficient of the linear classifier gives us a CAV pointing in the direction of the concept $a_i$ . Then, for both training and validation data, denoted as $X = \\{x_1,\\dots,x_n\\}$ , we use $f$ to extract representations, $\\{f(x_1),\\dots,f(x_n)\\}$ .", + "bbox": [ + 496, + 522, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then compute the cosine similarity between the representations and the CAV. We use these similarity scores to separate the data into groups — examples with the same attributes are likely to activate similar similarity scores. To separate points by similarity scores, we use a naive approach: we train a Gaussian Mixture Model (GMM) on the similarity scores, where the number of mixtures is set to $m$ , the size of the domain of the spurious attribute. Typically (but not exclusively) $m = 2$ , since many formulations of bias posit a binary distinction between minority and major-", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12428", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity populations. The mixture with the highest mean similarity to the concept is labeled as $\\mathcal{A} = a_{i}$ . Intuitively, data with high similarity scores with respect to the concept set tend to exhibit concept $a_{i}$ , and thus belong to groups with $\\mathcal{A} = a_{i}$ (see figure 3). We repeat the above steps for all $a_{i} \\in \\{a_{1}, \\dots, a_{m}\\}$ . Lastly, for the data points with missing labels, we randomly assign one from $\\{a_{1}, \\dots, a_{m}\\}$ .", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the case that the spurious attribute is binary, i.e. $\\mathcal{A} = \\{a_1,a_2\\}$ , we only need to curate concepts for $a_1$ or $a_2$ and train one CAV. Suppose we train a CAV for $a_1$ , and label data with the CAV, yielding $X_{a_1} = \\{x\\in X|\\mathcal{A}(x) = a_1\\}$ . We can then use the negative of the CAV to label $a_2$ , which can be considered the complement, $X_{a_1}^C$ .", + "bbox": [ + 76, + 198, + 467, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since Group DRO takes group labels as input, where a group label $g$ is the pair $(y, a)$ defined by the label $y$ and spurious attribute $a$ . We generate inferred group labels $\\hat{g}$ using ground truth labels $y$ and inferred spurious attribute labels $\\hat{a}$ to use in Stage 2.", + "bbox": [ + 76, + 287, + 467, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Stage 2: Optimize for worst-group loss with inferred labels", + "text_level": 1, + "bbox": [ + 76, + 382, + 467, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use Group DRO to train a classifier with inferred group labels $\\hat{g}$ from stage 1. Instead of overall loss, Group DRO optimizes for worst-group loss.", + "bbox": [ + 76, + 421, + 467, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Group DRO uses ground truth group labels to select the best model checkpoint based on validation worst-group accuracy. We instead use the inferred labels to compute the surrogate worst-group accuracy at validation.", + "bbox": [ + 76, + 467, + 467, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 541, + 281, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we first show that CDRO substantially outperforms ERM and GEORGE, and achieves comparable worst-group accuracy to existing methods that require group labels. To explain the effectiveness of CDRO, we then demonstrate that CDRO separates data into corresponding groups with concepts, and infers labels of spurious attributes with high accuracy. Lastly, we show that CDRO is robust to the quality of concepts in two aspects: concept size and distribution distance from training data.", + "bbox": [ + 75, + 566, + 467, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 712, + 156, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe three image datasets, corresponding concept sets, and baseline methods used in our experiments.", + "bbox": [ + 76, + 736, + 467, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1 Datasets", + "text_level": 1, + "bbox": [ + 76, + 785, + 189, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We briefly describe the three image classification benchmarks used in our evaluation. Full dataset and training details are included in the Appendix.", + "bbox": [ + 76, + 809, + 467, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CMNIST [1]: The task is to classify digits into one of the five classes: $\\mathcal{Y} = \\{(0,1),(2,3),(4,5),(6,7),(8,9)\\}$ . Each class is correlated with one of five colors: $\\mathcal{A}$", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$= \\{$ red, yellow, green, blue, purple\\}, respectively, each represented by its hex code (see figure 3).", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Waterbirds [17]: The task is to classify birds into one of the 2 classes: $\\mathcal{Y} = \\{\\text{waterbird, landbird}\\}$ . Each class is correlated with the backgrounds: $\\mathcal{A} = \\{\\text{water background, land background}\\}$ , respectively.", + "bbox": [ + 498, + 121, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CelebA [10]: The task is to classify celebrities' hair color $\\mathcal{Y} = \\{\\text{blond, not blond}\\}$ , with the spurious attribute $\\mathcal{A} = \\{\\text{female, male}\\}$ . The class blond correlates with female, and the class not blond correlates with male.", + "bbox": [ + 498, + 181, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2 Baselines", + "text_level": 1, + "bbox": [ + 500, + 261, + 617, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider seven baselines that train models under different assumptions regarding the availability of group labels.", + "bbox": [ + 498, + 284, + 888, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Group labels available during training - Group DRO [17] is a state-of-the-art method, which optimizes for the worst-group loss using group labels.", + "bbox": [ + 498, + 315, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Group labels available during validation - Just Train Twice (JTT) [9] trains a model to detect minority group examples, then upweights those examples during training. Correct-n-Contrast (CNC) [24] infers group information similar to JTT, then uses contrastive learning to learn representations robust to spurious correlations. Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. Spread Spurious Attribute (SSA) [14] uses semi-supervised learning to train a predictor to infer pseudo-group label, then trains a robust model with worst-group loss minimization.", + "bbox": [ + 498, + 359, + 890, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Group labels available during neither training nor validation - Empirical Risk Minimization (ERM) is the standard training procedure that minimizes the average loss. Similar to JTT, GEORGE also follows a two-step procedure by first estimating the group labels and then using these estimates to train a robust classifier. Unlike JTT, after training a standard model, the feature space of each class is split into groups via unsupervised clustering. During validation, GEORGE then optimizes for worst-cluster accuracy without requiring group labels.", + "bbox": [ + 498, + 527, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.3 Concepts", + "text_level": 1, + "bbox": [ + 500, + 696, + 617, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments we generated 50 images per concept set. We generated synthetic data as concepts for CMNIST, and used a popular off-the-shelf generative AI model Stable Diffusion [16] to generate concepts for Waterbirds and CelebA.", + "bbox": [ + 498, + 719, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CMNIST - Since the 5 target classes are correlated with 5 colors, we curate concept sets for each color. For each color, we sample 50 images from the original MNIST dataset [2], which consists of white digits 0-9 on black background, and we paint the digits with the corresponding color. We then curate the contrastive set by concatenating the concept sets of the 4 remaining colors, and trained a CAV pointing towards the color.", + "bbox": [ + 498, + 780, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12429", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/384035a350053ee80de40b000ea2154758c35d94ddb37fe8f960958226157ee5.jpg", + "image_caption": [ + "Figure 2. We show Concept DRO (CDRO), a method of training an unbiased model using a concept set. (1) First, we train a neural network model $(h\\circ f)$ using ERM. (2) We train a concept activation vector (CAV) using the given concept set and the model [6]. A CAV is a linear decision boundary between the model representations of the concepts $(f(\\cdot))$ . (3) We use the inferred CAV to infer pseudo-group labels over the training data. (4) Finally, we run Group DRO [17] the training data using the pseudo-labels." + ], + "image_footnote": [], + "bbox": [ + 173, + 78, + 787, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Waterbirds - Since land background is correlated with landbirds, and water background is correlated with waterbirds, we curate sets of examples to represent the concepts \"land background\" and \"water background.\" With the knowledge that land background in the dataset are either bamboo or broadleaf trees, we used prompts \"A photo of bamboo trees\" and \"A photo of broadleaf trees\" to generate a total of 50 (25 each) images to represent the \"land background\" concept. Similarly, knowing that the water background in the dataset are either ocean or lake, we used prompts \"A photo of ocean\" and \"A photo of lake\" to generate a total of 50 (25 each) images to represent the \"water background\" concept. Since the spurious attribute is binary, we trained a CAV in the direction of \"land background,\" using the \"water background\" concept as the contrastive set.", + "bbox": [ + 75, + 428, + 472, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CelebA - Since female is correlated with blond and male is correlated with not blond, we curate sets of examples to represent the concepts female and male. With the knowledge that the dataset consists of celebrities' faces, we used prompts \"A photo of a female celebrity face\" and \"A photo of a male celebrity face\" to generate a total of 100 images (50 each) for the female and male concepts, respectively. We trained a CAV with female as the concept set and male as the contrastive set, such that positive similarity can be interpreted as exhibiting female features.", + "bbox": [ + 75, + 655, + 472, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 816, + 217, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 reports the average and worst-group accuracies of all approaches. CDRO outperforms ERM across all tasks. Comparing to GEORGE, CDRO achieves higher worst-group accuracy on Waterbirds and CelebA. The perfor", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mance of GEORGE relies on how well the learned features can be clustered. In CMNIST, where the features (color and shape) are easily distinguishable, GEORGE easily finds clusters that correspond to groups and therefore yields a high worst-group accuracy. However, in Waterbirds and CelebA, where the features are more complex, clustering is less effective and the worst-group accuracy is lower. Further, with only 100 out-of-distribution images, CDRO achieves similar performance to methods that require validation labels, which may require large-scale manual labeling: CelebA, for example, has 20K validation images. Specifically, CDRO outperforms JTT by $5\\%$ on Waterbirds, and $6.5\\%$ on CelebA and approaches the performance of Group DRO itself. CDRO outperforms CNC by $0.3\\%$ on Waterbirds, and falls short by $0.8\\%$ on CelebA. CDRO outperforms SSA with $5\\%$ validation labels by $1.7\\%$ on Waterbirds, and $1.3\\%$ on CelebA. Lastly, CDRO outperforms AFR with $5\\%$ validation labels by $2.2\\%$ on Waterbirds, and $10.5\\%$ on CelebA.", + "bbox": [ + 496, + 428, + 893, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 498, + 728, + 663, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3.1 CDRO effectively infers group labels", + "text_level": 1, + "bbox": [ + 498, + 753, + 808, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To examine the effectiveness of CDRO in inferring group labels, we first visualize the distribution of both train and validation data by their cosine similarity scores to the corresponding CAV for all three datasets. As shown in Figure 3, for both train and validation data of the three datasets, data representations with high similarity scores tend to exhibit the concept that the CAV directed to. For example, Figure 3 shows that, given a CAV in the direction of the concept red,", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12430", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c73be5748e90d4087658ff64593f485afcdc0a5ec7847f6c72d863799fb53449.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGroup InfoCMNIST [1]Waterbirds [17]CelebA [10]
Worst(%)Mean(%)Worst(%)Mean(%)Worst(%)Mean(%)
Group DRO [17]Train & Val74.7 (1.0)190.6 (0.1)89.9 (0.6)92.0 (0.6)88.9 (1.3)93.9 (0.1)
JTT [9]Val74.5 (2.4)90.2 (0.8)83.8 (1.2)89.3 (0.7)81.5 (1.7)88.1 (0.3)
CNC [24]Val77.4 (3.0)90.9 (0.6)88.5 (0.3)90.9 (0.1)88.8 (0.9)89.9 (0.5)
SSA [14]Val--89.0 (0.6)92.2 (0.9)89.8 (1.3)92.8 (0.1)
5% Val--87.1 (0.7)92.6 (0.2)86.7 (1.1)92.8 (0.3)
AFR [15]Val--90.4 (1.1)94.2 (1.2)82.0 (0.5)91.3 (0.3)
5% Val--86.6 (4.7)-77.5 (1.3)-
GEORGE [19]None76.4 (2.3)89.5 (0.3)76.2 (2.0)95.7 (0.5)54.9 (1.9)94.6 (0.2)
ERMNone0.0 (0.0)20.1 (0.2)62.6 (0.3)97.3 (1.0)47.7 (2.1)94.9 (0.3)
CDRO (ours)Concepts69.3 (2.7)85.0 (1.0)88.8 (0.2)90.5 (0.1)88.0 (0.7)90.8 (0.3)
", + "bbox": [ + 114, + 88, + 857, + 303 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3228c83f99f01725907992037b17de978d8de15bae91f53973784009983457e.jpg", + "table_caption": [ + "Table 1. Average and worst-group accuracies of models trained with CDRO and baselines on the benchmarks. CDRO outperforms approaches that do not use group information (GEORGE, ERM) and approaches that use $5\\%$ of Validation (SSA, AFR). CDRO also has comparable performance to approaches that use validation group labels (JTT, CNC, AFR, SSA)." + ], + "table_footnote": [], + "table_body": "
Precision(%)Recall(%)
CMNIST87.096.9
Waterbirds76.383.9
CelebA78.078.5
", + "bbox": [ + 130, + 359, + 416, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. We show the average precision and recall of the pseudogroup labels inferred by CDRO compared to the ground truth. We show CDRO effectively infers group labels.", + "bbox": [ + 75, + 448, + 470, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "data points with high similarity scores tend to be the red digits. Similarly, given a CAV in the direction of the concept land background, data points with high similarity scores tend to be images with land background. These results justify the decision to label all data points in the highest-mean mixture with the group associated with the concept.", + "bbox": [ + 75, + 505, + 468, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further, Table 2 shows that CDRO estimates group labels with high accuracy in all three datasets.", + "bbox": [ + 76, + 597, + 470, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 CDRO is robust to quality of concepts", + "text_level": 1, + "bbox": [ + 76, + 648, + 395, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From a practical perspective, since CDRO requires practitioners to curate the concepts, the quality of concepts depends on how much time and effort the practitioners spend on curating the concepts. We therefore evaluate the effectiveness of CDRO when the quality of concepts varies on Waterbirds and CelebA. To examine this, we ask the following two questions: 1) Does CDRO require that the concept images be drawn from the training distribution? 2) Does CDRO require a large number of concept images?", + "bbox": [ + 75, + 672, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first fix the size of concepts to be 100 (50 for the concept set and 50 for the contrastive set), and examine the effect of distribution distance between concepts and training data on CDRO. We vary the distance between training distribution and concept distribution by varying the prompts we use for Stable Diffusion.", + "bbox": [ + 75, + 810, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We design the prompts based on how much effort a hypothetical practitioner would spend on curating the concepts. A practitioner may not have specific information about the dataset and decide to use very general prompts, potentially resulting in a low-quality concept set. For the Waterbirds dataset, we use the following prompts, varying the level of effort and domain knowledge: \"A photo of a [water/land] background\" as the \"distant\" distribution, \"A photo of a [water/land]bird habitat\" as the \"somewhat near\" distribution, and \"A photo of a [water/land] background\" as the \"near\" distribution. For the CelebA dataset, we use the following prompts: \"A photo of a [female/male] person\" as the \"distant\" distribution, \"A photo of a [female/male] celebrity\" as the \"somewhat near\" distribution, and \"A photo of a [female/male] celebrity face\" as the \"near\" distribution. We include our reasoning and a discussion regarding these prompts in the Appendix.", + "bbox": [ + 496, + 363, + 890, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 shows that CDRO is effective even if the concept images are out of distribution. In both Waterbirds and CelebA, CDRO achieves similar worst group accuracy to the in-distribution set even when the concept distribution is only somewhat near. This result suggests that practitioners need not meticulously curate the concept sets for CDRO to be effective.", + "bbox": [ + 496, + 625, + 893, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then fix the concept distribution to be near training distribution, and vary the size from the list (4, 8, 40, 100, 200), splitting equally between concept set and contrastive set. For example, if size is 4, then the concept and contrastive sets each have two images. Figure 4 shows that CDRO is robust to the size of the concepts. Even when there are only four images, CDRO achieves higher worst-group accuracy than GEORGE on both Waterbirds and CelebA. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC. A small concept set is sufficient in part since the", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ebdf4d58afa644bfa27870431a34425f2d0845623f803aa1c29383ef401399b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 80, + 267, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f268eb9a50068ec790dbf0fce63c934472a6a3afa0d70742d8ccd88996658fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 80, + 421, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf3b6303c3a42ae07f70167c4baa68de9a71e298664cde0ebede13dc66bd6202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 423, + 80, + 575, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fcef310329df0b558cc28da4ed3f18fb7568fa6859c4bc2920bd8f2faf7f148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 80, + 730, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/483da73f71d471fe04ea562534360b3b33dec50af129c3bb8d365c8b6b4ee5f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 80, + 885, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfb52e8ec05d6454499a01e242c54cdefda5dfa67dea34c92623b88739324739.jpg", + "image_caption": [ + "Figure 3. Distributions of similarity between the CAV trained with out-of-distribution concepts and validation samples. Each peak is colored by the ground truth group label. The peaks are separable, and the peak furthest to the right (highest similarity to the concept) corresponds to the ground truth label in all cases, indicating how we can infer spurious attribute labels. Top row: CMNIST validation samples for each of five color concepts. Bottom left: Waterbird validation samples with land as the concept set and water as the contrastive set. Bottom right: CelebA validation samples with women as the concept set and men as the contrastive set." + ], + "image_footnote": [], + "bbox": [ + 222, + 200, + 482, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1fc5d0e9451bf8baa00f0134cef9b10ac18f73827e738ac22b97253d6b58904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 485, + 200, + 740, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0444e06a7a4bd6bd4d3b231035e164a0c1b5df75a1502717142e746f9dfeb78b.jpg", + "image_caption": [ + "Figure 4. CDRO is robust to size of concepts. Even when there are only 4 images, CDRO achieves higher worst group accuracy than GEORGE on both Waterbirds and CelebA, since the learned embeddings from ERM make it easy to train a classifier. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC." + ], + "image_footnote": [], + "bbox": [ + 102, + 414, + 431, + 585 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf4ef2a02139965ef7c2fec79951745ed4b2415fcaff6f08336a50c64b03d77e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceWorst(%)
In DistributionWaterbirds [17]89.3
NearStable Diffusion [16]88.5
Somewhat NearStable Diffusion [16]80.1
DistantStable Diffusion [16]53.1
", + "bbox": [ + 89, + 676, + 457, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "input to the GMM is ERM-trained embeddings, such that training data features are already represented.", + "bbox": [ + 75, + 830, + 470, + 862 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf211782ca2b3e7117d2e388c6c8b10c07cc0aed03d7f607f99e4336bc1377f8.jpg", + "table_caption": [ + "Table 3. The worst group accuracy given different concept set qualities on Waterbirds. Concept set quality is judged by the level of perceived effort and domain knowledge needed to create." + ], + "table_footnote": [], + "table_body": "
SourceWorst(%)
In DistributionCelebA [10]87.2
NearStable Diffusion [16]87.9
Somewhat NearStable Diffusion [16]87.2
DistantStable Diffusion [16]75.6
", + "bbox": [ + 511, + 417, + 880, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The worst group accuracy given different concept set qualities on CelebA. The concept set qualities are judged similar to Waterbird concepts.", + "bbox": [ + 498, + 521, + 893, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3 Concept Correction is Compatible with CNC", + "text_level": 1, + "bbox": [ + 500, + 573, + 867, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate that Concept Correction is compatible with other robust training methods via CNC. Specifically, CNC consists of two stages: first, CNC trains a model with ERM and uses the predictions of the ERM model to identify majority and minority groups; then CNC uses contrastive learning to maximize the similarity of examples with same class label but different ERM predictions, and minimize the similarity examples with different class labels but same ERM predictions. We replace ERM predictions with inferred labels of spurious attributes, which we generate according to stage 1-3 of Figure 2. Table 5 shows that models trained with concept-based pseudo-labels achieve comparable worst-group accuracy to models trained with ground truth group labels on both train and validation set.", + "bbox": [ + 496, + 599, + 890, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 500, + 828, + 640, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We discuss related work on improving robustness to spurious correlations, both with and without group information available.", + "bbox": [ + 498, + 854, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1Note that we ran Group DRO on CMNIST according to hyperparameters specified in [24], whereas the result reported by [24] was 78.5(4.5)", + "bbox": [ + 76, + 875, + 467, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12432", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0cb647e92f8c219be1856727a6321c6cba58f7835380e71f1a07a84fc76f5144.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGroup InfoWaterbirds Worst(%)CelebA Worst(%)
CNCTrain & Val88.7 (0.4)88.4 (0.1)
Concept CNCConcepts86.2 (0.7)87.2 (0.1)
", + "bbox": [ + 232, + 88, + 736, + 150 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Worst-group accuracy of Concept CNC compared to CNC. We show that we can use concept sets to achieve comparable performance to CNC, which uses both train and validation group labels in our implementation.", + "bbox": [ + 75, + 160, + 890, + 189 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improving Robustness with Group Information \nSagawa et al. proposed Group Distributionally Robust Optimization (DRO) [17], which assumes access to spurious attribute labels and optimizes for worst-group loss instead of average loss. The authors show that Group DRO can achieve state-of-the-art worst-group-accuracy. To demonstrate CDRO, we use the same distributionally robust objective of Group DRO but infer group labels instead of assuming they are provided.", + "bbox": [ + 75, + 203, + 468, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Several proposals to improve worst group accuracy re-sample the training data to balance classes or groups [4, 5, 7, 12]. These approaches show that while a neural network trained via ERM may be biased at the classification layer, it can still learn the core features for the task at the penultimate layer. Therefore it is effective to adjust the weights of the features by fine-tuning on a balanced hold-out set [5, 7], or re-weighting the loss function for majority and minority groups [4], or processing the model outputs post-hoc [12]. However, these methods acknowledge the expense of obtaining spurious attribute labels for a large dataset. Further, multiple spurious correlations can occur in practice [25], complicating reweighting schemes.", + "bbox": [ + 75, + 339, + 470, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Improving Robustness without Group Information: Recent proposals to improve worst-group accuracy without access to group labels typically follow a two-stage paradigm: first train a model with ERM, then use the predictions of the ERM model to identify minority groups, then train another model with emphasis on the minority groups, thereby achieving high accuracy on both the majority and minority groups [5, 9, 14, 15, 19, 20, 22, 24]. For example, JTT [9] uses the incorrect predictions of the ERM classifier as a proxy for the minority group, then trains a model with an upweight on the misclassified examples. Similarly, Correct-N-Contrast (CNC) [24] uses misclassifications as pseudo spurious attribute labels, then uses contrastive learning to align representations for samples within the same class but different spurious attribute. Recognizing that retraining a model from scratch is computationally expensive, Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. However, these methods still require a validation set with group labels for hyperparameter tuning and selecting the best model checkpoint, which turns out to be crucial for achieving their results. CDRO avoids the need of validation group labels by using concepts to infer pseudo group labels.", + "bbox": [ + 75, + 537, + 470, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In [14] and [20], the authors assume access to a small", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of group-labeled data and use semi-supervised learning to exploit the group information for better worst group accuracy. Specifically, Spread Spurious Attribute (SSA) [14] trains a spurious attribute predictor using both group-labeled and group unlabeled data, then uses the predicted labels for worst-group loss minimization. Unlike SSA, CDRO does not require concepts to represent task-specific groups, which encapsulate both the spurious attribute and target class. Instead, concepts can be out-of-distribution examples solely representing the spurious attribute.", + "bbox": [ + 496, + 203, + 890, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GEORGE [19] requires group labels neither during training nor validation, instead inferring group labels by clustering the feature space learned by the ERM model, then running Group DRO. During validation, GEORGE uses worst-cluster accuracy to select the best hyperparameters and model checkpoint, which avoids the need of group labels. However, while GEORGE achieves high worst group accuracy on synthetic datasets where clusters are well-defined, it is less effective on real-world datasets such as CelebA where clusters are harder to distinguish [19].", + "bbox": [ + 496, + 369, + 890, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Some approaches do not follow the two-stage paradigm. Learning from Failure (LfF) [13] trains two models simultaneously, one intended to be biased and one intended to be unbiased. When training the unbiased model, LfF focuses on the examples on which the biased model tends to make mistakes thereby learning to ignore the sources of bias. Further, Taghanaki et al. introduced CIM [21], a contrastive learning approach that learns input-space transformation of the data to preserve task-relevant information.", + "bbox": [ + 496, + 523, + 890, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 676, + 617, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present Concept Correction, a framework that uses out-of-distribution concepts to improve worst-group accuracy in the presence of spurious correlations. We demonstrate an instance of the framework via Concept DRO, which uses concepts to infer group labels, and then uses these to train with worst-group loss minimization. We show that Concept DRO achieves competitive results to existing methods that assume access to group labels.", + "bbox": [ + 496, + 704, + 890, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 843, + 679, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by NIST Grant 60NANB23D194, NSF Award 1934464, Cisco Grant 70618863.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2, 4, 6", + "[2] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 4", + "[3] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1", + "[4] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 8", + "[5] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. Advances in Neural Information Processing Systems, 35:38516-38532, 2022. 8", + "[6] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5", + "[7] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 1, 2, 8", + "[8] Anthony Liu, Santiago Guerra, Isaac Fung, Gabriel Matute, Ece Kamar, and Walter Lasecki. Towards hybrid human-air workflows for unknown unknown detection. In Proceedings of The Web Conference 2020, pages 2432-2442, 2020. 1", + "[9] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 1, 2, 3, 4, 6, 8", + "[10] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 2, 4, 6, 7", + "[11] R Thomas McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007, 2019. 1", + "[12] Aditya Krishna Menon, Ankit Singh Rawat, and Sanjiv Kumar. Overparameterisation and worst-case generalisation: friend or foe? In International Conference on Learning Representations, 2020. 8", + "[13] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020. 8", + "[14] Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accu" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "racy with spurious attribute estimation. arXiv preprint arXiv:2204.02070, 2022. 4, 6, 8", + "[15] Shikai Qiu, Andres Potapczynski, Pavel Izmailov, and Andrew Gordon Wilson. Simple and fast group robustness by automatic feature reweighting. arXiv preprint arXiv:2306.11074, 2023. 4, 6, 8", + "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4, 7", + "[17] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 1, 2, 3, 4, 5, 6, 7, 8", + "[18] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 1", + "[19] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1, 6, 8", + "[20] Nimit S Sohoni, Maziar Sanjabi, Nicolas Ballas, Aditya Grover, Shaoliang Nie, Hamed Firooz, and Christopher Ré. Barack: Partially supervised group robustness with guarantees. arXiv preprint arXiv:2201.00072, 2021. 8", + "[21] Saeid A Taghanaki, Kristy Choi, Amir Hosein Khasahmadi, and Anirudh Goyal. Robust representation learning via perceptual similarity metrics. In International Conference on Machine Learning, pages 10043-10053. PMLR, 2021. 8", + "[22] Haotian Ye, James Zou, and Linjun Zhang. Freeze then train: Towards provable representation learning under spurious correlations and feature noise. In International Conference on Artificial Intelligence and Statistics, pages 8968-8990. PMLR, 2023. 8", + "[23] John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS medicine, 15(11):e1002683, 2018. 1", + "[24] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. 1, 2, 3, 4, 6, 7, 8", + "[25] Dora Zhao, Jerone Andrews, and Alice Xiang. Men also do laundry: Multi-attribute bias amplification. In International Conference on Machine Learning, pages 42000–42017. PMLR, 2023. 8" + ], + "bbox": [ + 503, + 93, + 890, + 810 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12434", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_model.json b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..528223058bc691fe31aa639d7a8647cf26b5939c --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_model.json @@ -0,0 +1,1747 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Label-Efficient Group Robustness via Out-of-Distribution Concept Curation" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.846, + 0.234 + ], + "angle": 0, + "content": "Yiwei Yang1 Anthony Z. Liu2 Robert Wolfe1 Aylin Caliskan1 Bill Howe1 \n1University of Washington, 2University of Michigan \n{yanyiwei, rwolfe3, aylin, billhowe}@uw.edu anthliu@umich.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.771 + ], + "angle": 0, + "content": "Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to \\(33.1\\%\\) on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.8, + 0.21, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Deep learning models are prone to capture spurious correlations, patterns that are predictive of the target class in training data but are inherently irrelevant to the classification task. For example, consider an image classification task to distinguish doctors from nurses. Suppose that in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.468 + ], + "angle": 0, + "content": "the training data, \\(95\\%\\) of the doctors present as men, and \\(95\\%\\) of the nurses present as women. A model trained with the standard empirical risk minimization (ERM) tends to classify images of men as doctors and images of women as nurses by learning a spurious correlation between gender-expression features (e.g., hairstyle, clothing) and occupation. While the model may achieve high average accuracy, it performs poorly on the minority groups (in this example, female-presenting doctors and male-presenting nurses), thus exhibiting low worst-group accuracy. Such spurious correlations happen in a wide variety of real-world tasks, including facial recognition, medical imaging, and language tasks [3, 8, 11, 17, 18, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.837 + ], + "angle": 0, + "content": "Existing approaches typically consider the setting in which each training (or validation) sample is explicitly assigned a group, represented as a pair \\((y, a)\\) where \\(y\\) is a class label and \\(a\\) is a value from the (potentially unknown) spurious attribute. For instance, Group DRO assumes the availability of group labels during training, and directly minimizes the worst-group loss [17]. Recognizing that obtaining labels requires expensive human annotation, several recent approaches avoid relying on group labels during training [7, 9, 19, 24]. These approaches train an ERM model to infer group labels, then a second model is trained to optimize for worst-group loss with the inferred labels. However, these approaches require validation labels for crucial hyperparameter tuning [9], since their effectiveness depends on the accuracy of the group inference. One exception is GEORGE [19], which assumes that the feature space of a model trained for the predictive task is easily separable by groups and learns group assignment via clustering. GEORGE then optimizes for worst-cluster accuracy without requiring labels. However, while GEORGE is effective when the input features are simple and cluster well (e.g. colored digits), its performance drops significantly when the input features are more complex and clusters become less well-defined (e.g., for human faces) [19]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We propose Concept Correction (figure 1), a framework that instead uses a set of example images to represent a concept involved in the spurious correlation. These examples can be out-of-distribution and task-agnostic, as long as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "12426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.151, + 0.079, + 0.805, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 1. We show our main idea Concept Correction, a framework for correcting spurious attributes in biased models using concepts, set of examples representing some pattern. We show an example from the Waterbirds dataset [17], where (a) directly training an ERM results in a model that spuriously correlates background type with bird type. A practitioner can correct this by (b) finding images with land and water backgrounds and using concept correction. Notably, these example images can be obtained outside of the original Waterbirds dataset, such as using generative AI, and using search engines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.473, + 0.664 + ], + "angle": 0, + "content": "they contain the features representing the spurious attribute. For example, concept sets designed to counteract a spurious attribute of gender presentation may include a set of images of female-presenting people, but without any reference to the task-specific class label of doctor or nurse. Since the examples need not be selected from the training distribution, they are reusable across models and tasks, and practitioners can retrieve them from various sources such as a search engines, datasets on the Web, or image generation models, all of which require significantly less cognitive and technical effort than manually labeling in-distribution data. In fact, by using images generated with only two lines of prompts to the text-to-image generator Stable Diffusion [16], we show that an implementation of Concept Correction can achieve competitive results to state-of-the-art methods that require manual labeling of tens of thousands of data points [9, 17, 24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We demonstrate an instance of Concept Correction via Concept DRO (CDRO), an approach that leverages concepts to infer group labels, then trains with a distributively robust optimization objective using the inferred labels. Similar to the existing two-stage methods, CDRO first trains a neural network with ERM. Then, exploiting the fact that the ERM model learns spurious attribute labels[7], we use the model to extract features from the concept sets, the training data, and the validation data. We then train a classifier to discriminate concepts then use the weights of this classifier to determine the extent to which each training sample exhibits that concept. These concept similarity scores are then used to infer a spurious attribute label for every data point in the training and validation sets, providing the information needed by state-of-the-art methods to improve" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.481 + ], + "angle": 0, + "content": "worst-group accuracy. In our example, CDRO trains a classifier on two sets of images representing the concepts of presenting-female and presenting-male, then uses the coefficients of this classifier to estimate the extent to which each training sample exhibits \"maleness\" vs. \"femaleness.\"" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.893, + 0.633 + ], + "angle": 0, + "content": "We evaluate CDRO on three image classification datasets with spurious correlations: CMNIST [1], Waterbirds [17], and CelebA [10]. Among methods that do not assume spurious attribute labels are provided, CDRO improves worstgroup accuracy by up to \\(12.6\\%\\) on Waterbirds and \\(33.1\\%\\) on CelebA. CDRO also achieves competitive performance to methods that do assume access to spurious attribute labels. Finally, we show that CDRO is robust to the quality of concepts in two aspects: size of concept and distribution distance from training data." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.633, + 0.822, + 0.647 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.651, + 0.891, + 0.695 + ], + "angle": 0, + "content": "- We propose Concept Correction, a label-efficient framework that uses out-of-distribution sets of images to correct spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.696, + 0.892, + 0.756 + ], + "angle": 0, + "content": "- We present CDRO, an instance of the Concept Correction framework that uses concepts to infer labels of spurious attributes and then trains with a distributively robust optimization objective." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.757, + 0.892, + 0.817 + ], + "angle": 0, + "content": "- We show CDRO significantly outperforms ERM and GEORGE across three benchmarks while achieving competitive results to methods that require group labels during validation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.651, + 0.892, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.64, + 0.846 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we first present the problem setting, then describe concepts and the Concept Activation Vector (CAV) introduced by Kim et al. [6]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12427" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.24, + 0.108 + ], + "angle": 0, + "content": "2.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.297 + ], + "angle": 0, + "content": "We consider a classification setting, where given an input \\( x \\in \\mathcal{X} \\), our goal is to predict \\( y \\in \\mathcal{Y} \\). Let \\( X = \\{x_{1}, \\ldots, x_{n}\\} \\), \\( Y = \\{y_{1}, \\ldots, y_{n}\\} \\) be our training set of size \\( n \\). Each datapoint \\( x_{i} \\) has a spurious attribute \\( a_{i} \\), which is unobserved during training. Each \\( x_{i} \\) is assumed to belong to a group \\( g_{i} \\) where \\( g_{i} \\) is a pair \\( (y_{i}, a_{i}) \\), the combination of a class label \\( y_{i} \\) and a spurious attribute label \\( a_{i} \\). The spurious correlation problem occurs when accuracy across groups varies significantly; i.e., a model trained with ERM may achieve a high average accuracy, yet still performs poorly on some minority groups. We therefore evaluate classifiers on the accuracy of the worst-performing group." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.297, + 0.471, + 0.449 + ], + "angle": 0, + "content": "We are interested in the setting where group information is not available for either the training set nor the validation set. Instead, we assume that users curate a set of images representing the concept without regard to any particular task. The human effort to curate the concept depends on the user's domain knowledge. For example, in Waterbirds, to curate a concept for 'landbird background,' one can generate images with the prompt 'a photo of land background'. However, if the user knows that land backgrounds of birds tend to be forests, the prompt can be adjusted accordingly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.458, + 0.433, + 0.475 + ], + "angle": 0, + "content": "2.2. Concepts and Concept Activation Vectors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.557 + ], + "angle": 0, + "content": "A concept is defined using a set of images. For example, a set of ocean and lake images can represent the concept water background. The concept examples need not be drawn from the training data and need not be relevant to the classification task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.558, + 0.471, + 0.68 + ], + "angle": 0, + "content": "Given a concept \\( C \\) and a layer \\( l \\) of the neural network, a concept activation vector (CAV) \\( \\mathbf{v}_C^l \\in R^m \\) [6] is defined as the vector normal to the hyperplane separating samples with a concept (the concept set \\( P_C \\)) and examples without a concept (the contrastive set \\( N \\)) in the model's activations. Given layer activations of both the concept set and the contrastive set, CAV can be obtained by training a linear classifier distinguishing the two sets of activations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.268, + 0.71 + ], + "angle": 0, + "content": "3. Concept Correction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.84 + ], + "angle": 0, + "content": "We propose Concept Correction, a framework that uses concepts to correct spurious correlations, without relying on manual labeling of in-distribution data. Instead, Concept Correction takes a set of a small number of examples as input and leverages the examples to infer spurious attribute information. Then, a robust model is trained with the inferred group of spurious attribute information using any worst-group loss minimization algorithms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Concept Correction expects the concept examples to represent the spurious attribute. As described in the introduction, these images can be out-of-distribution and irrelevant to the prediction task, de-coupling concept curation from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "the model setting. In fact, using generative models, practitioners can easily generate many examples with just a few lines of prompting. In contrast to existing methods that require manual labeling of either training or validation samples, concept curation does not require access to the training data. Practitioners can reason about potential biases the model may exhibit then curate concept sets accordingly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Concept Correction is compatible with any method that requires spurious attribute or group information, such as Group DRO [17], Just-Train-Twice [9], Correct-n-Contrast [24], and more. For example, a simple way of applying Concept Correction to Correct-n-Contrast and JustTrain-Twice is simply replacing the validation group labels with inferred labels using concepts, alleviating the need of manual labeling of a large amount of data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.331, + 0.646, + 0.347 + ], + "angle": 0, + "content": "3.1. Concept DRO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.894, + 0.476 + ], + "angle": 0, + "content": "We present Concept DRO (CDRO), an implementation of Concept Correction on Group DRO. As illustrated in Figure 2, in stage 1, we first train an ERM model and then train a linear classifier on concepts representing the spurious attribute to infer group labels. In stage 2, we use Group DRO to train a robust model with the inferred group labels. We include further details on both stages below, and give pseudocode for CDRO in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.498, + 0.89, + 0.514 + ], + "angle": 0, + "content": "3.1.1 Stage 1: Inferring Labels of Spurious Attributes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.75 + ], + "angle": 0, + "content": "We train a neural network with ERM on the training data. Given a spurious attribute \\(\\mathcal{A} = \\{a_1,\\dots,a_m\\}\\), we curate two sets of examples for each concept \\(a_i\\): a concept set \\(C = \\{c_1,\\dots,c_k\\}\\), which represents the presence of the concept, and a contrastive set \\(N = \\{n_1,\\dots,n_k\\}\\), which represents the absence of the concept. We then use the ERM model as a feature extractor, by taking the first to the penultimate layer of the model, denoted as \\(f\\), to extract features from the two sets. We then train a linear classifier on the ERM features of the concept sets, \\(\\{f(c_1),\\dots,f(c_k)\\}\\) and \\(\\{f(n_1),\\dots,f(n_k)\\}\\). The resulting coefficient of the linear classifier gives us a CAV pointing in the direction of the concept \\(a_i\\). Then, for both training and validation data, denoted as \\(X = \\{x_1,\\dots,x_n\\}\\), we use \\(f\\) to extract representations, \\(\\{f(x_1),\\dots,f(x_n)\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We then compute the cosine similarity between the representations and the CAV. We use these similarity scores to separate the data into groups — examples with the same attributes are likely to activate similar similarity scores. To separate points by similarity scores, we use a naive approach: we train a Gaussian Mixture Model (GMM) on the similarity scores, where the number of mixtures is set to \\( m \\), the size of the domain of the spurious attribute. Typically (but not exclusively) \\( m = 2 \\), since many formulations of bias posit a binary distinction between minority and major-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12428" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "ity populations. The mixture with the highest mean similarity to the concept is labeled as \\(\\mathcal{A} = a_{i}\\). Intuitively, data with high similarity scores with respect to the concept set tend to exhibit concept \\(a_{i}\\), and thus belong to groups with \\(\\mathcal{A} = a_{i}\\) (see figure 3). We repeat the above steps for all \\(a_{i} \\in \\{a_{1}, \\dots, a_{m}\\}\\). Lastly, for the data points with missing labels, we randomly assign one from \\(\\{a_{1}, \\dots, a_{m}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.199, + 0.468, + 0.288 + ], + "angle": 0, + "content": "In the case that the spurious attribute is binary, i.e. \\(\\mathcal{A} = \\{a_1,a_2\\}\\), we only need to curate concepts for \\(a_1\\) or \\(a_2\\) and train one CAV. Suppose we train a CAV for \\(a_1\\), and label data with the CAV, yielding \\(X_{a_1} = \\{x\\in X|\\mathcal{A}(x) = a_1\\}\\). We can then use the negative of the CAV to label \\(a_2\\), which can be considered the complement, \\(X_{a_1}^C\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.289, + 0.468, + 0.364 + ], + "angle": 0, + "content": "Since Group DRO takes group labels as input, where a group label \\( g \\) is the pair \\( (y, a) \\) defined by the label \\( y \\) and spurious attribute \\( a \\). We generate inferred group labels \\( \\hat{g} \\) using ground truth labels \\( y \\) and inferred spurious attribute labels \\( \\hat{a} \\) to use in Stage 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.468, + 0.412 + ], + "angle": 0, + "content": "3.1.2 Stage 2: Optimize for worst-group loss with inferred labels" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.468, + 0.467 + ], + "angle": 0, + "content": "We use Group DRO to train a classifier with inferred group labels \\(\\hat{g}\\) from stage 1. Instead of overall loss, Group DRO optimizes for worst-group loss." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.468, + 0.468, + 0.527 + ], + "angle": 0, + "content": "Group DRO uses ground truth group labels to select the best model checkpoint based on validation worst-group accuracy. We instead use the inferred labels to compute the surrogate worst-group accuracy at validation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.282, + 0.559 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.704 + ], + "angle": 0, + "content": "In our experiments, we first show that CDRO substantially outperforms ERM and GEORGE, and achieves comparable worst-group accuracy to existing methods that require group labels. To explain the effectiveness of CDRO, we then demonstrate that CDRO separates data into corresponding groups with concepts, and infers labels of spurious attributes with high accuracy. Lastly, we show that CDRO is robust to the quality of concepts in two aspects: concept size and distribution distance from training data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.158, + 0.729 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.737, + 0.468, + 0.767 + ], + "angle": 0, + "content": "We describe three image datasets, corresponding concept sets, and baseline methods used in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.19, + 0.8 + ], + "angle": 0, + "content": "4.1.1 Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.468, + 0.855 + ], + "angle": 0, + "content": "We briefly describe the three image classification benchmarks used in our evaluation. Full dataset and training details are included in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "CMNIST [1]: The task is to classify digits into one of the five classes: \\(\\mathcal{Y} = \\{(0,1),(2,3),(4,5),(6,7),(8,9)\\}\\). Each class is correlated with one of five colors: \\(\\mathcal{A}\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "\\(= \\{\\) red, yellow, green, blue, purple\\}, respectively, each represented by its hex code (see figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Waterbirds [17]: The task is to classify birds into one of the 2 classes: \\(\\mathcal{Y} = \\{\\text{waterbird, landbird}\\}\\). Each class is correlated with the backgrounds: \\(\\mathcal{A} = \\{\\text{water background, land background}\\}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.242 + ], + "angle": 0, + "content": "CelebA [10]: The task is to classify celebrities' hair color \\(\\mathcal{Y} = \\{\\text{blond, not blond}\\}\\), with the spurious attribute \\(\\mathcal{A} = \\{\\text{female, male}\\}\\). The class blond correlates with female, and the class not blond correlates with male." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.262, + 0.618, + 0.275 + ], + "angle": 0, + "content": "4.1.2 Baselines" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.285, + 0.89, + 0.315 + ], + "angle": 0, + "content": "We consider seven baselines that train models under different assumptions regarding the availability of group labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.316, + 0.891, + 0.36 + ], + "angle": 0, + "content": "Group labels available during training - Group DRO [17] is a state-of-the-art method, which optimizes for the worst-group loss using group labels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.526 + ], + "angle": 0, + "content": "Group labels available during validation - Just Train Twice (JTT) [9] trains a model to detect minority group examples, then upweights those examples during training. Correct-n-Contrast (CNC) [24] infers group information similar to JTT, then uses contrastive learning to learn representations robust to spurious correlations. Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. Spread Spurious Attribute (SSA) [14] uses semi-supervised learning to train a predictor to infer pseudo-group label, then trains a robust model with worst-group loss minimization." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.678 + ], + "angle": 0, + "content": "Group labels available during neither training nor validation - Empirical Risk Minimization (ERM) is the standard training procedure that minimizes the average loss. Similar to JTT, GEORGE also follows a two-step procedure by first estimating the group labels and then using these estimates to train a robust classifier. Unlike JTT, after training a standard model, the feature space of each class is split into groups via unsupervised clustering. During validation, GEORGE then optimizes for worst-cluster accuracy without requiring group labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.618, + 0.711 + ], + "angle": 0, + "content": "4.1.3 Concepts" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.72, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In our experiments we generated 50 images per concept set. We generated synthetic data as concepts for CMNIST, and used a popular off-the-shelf generative AI model Stable Diffusion [16] to generate concepts for Waterbirds and CelebA." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.781, + 0.892, + 0.9 + ], + "angle": 0, + "content": "CMNIST - Since the 5 target classes are correlated with 5 colors, we curate concept sets for each color. For each color, we sample 50 images from the original MNIST dataset [2], which consists of white digits 0-9 on black background, and we paint the digits with the corresponding color. We then curate the contrastive set by concatenating the concept sets of the 4 remaining colors, and trained a CAV pointing towards the color." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12429" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.079, + 0.789, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.896, + 0.416 + ], + "angle": 0, + "content": "Figure 2. We show Concept DRO (CDRO), a method of training an unbiased model using a concept set. (1) First, we train a neural network model \\((h\\circ f)\\) using ERM. (2) We train a concept activation vector (CAV) using the given concept set and the model [6]. A CAV is a linear decision boundary between the model representations of the concepts \\((f(\\cdot))\\). (3) We use the inferred CAV to infer pseudo-group labels over the training data. (4) Finally, we run Group DRO [17] the training data using the pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Waterbirds - Since land background is correlated with landbirds, and water background is correlated with waterbirds, we curate sets of examples to represent the concepts \"land background\" and \"water background.\" With the knowledge that land background in the dataset are either bamboo or broadleaf trees, we used prompts \"A photo of bamboo trees\" and \"A photo of broadleaf trees\" to generate a total of 50 (25 each) images to represent the \"land background\" concept. Similarly, knowing that the water background in the dataset are either ocean or lake, we used prompts \"A photo of ocean\" and \"A photo of lake\" to generate a total of 50 (25 each) images to represent the \"water background\" concept. Since the spurious attribute is binary, we trained a CAV in the direction of \"land background,\" using the \"water background\" concept as the contrastive set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.473, + 0.807 + ], + "angle": 0, + "content": "CelebA - Since female is correlated with blond and male is correlated with not blond, we curate sets of examples to represent the concepts female and male. With the knowledge that the dataset consists of celebrities' faces, we used prompts \"A photo of a female celebrity face\" and \"A photo of a male celebrity face\" to generate a total of 100 images (50 each) for the female and male concepts, respectively. We trained a CAV with female as the concept set and male as the contrastive set, such that positive similarity can be interpreted as exhibiting female features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.218, + 0.832 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Table 1 reports the average and worst-group accuracies of all approaches. CDRO outperforms ERM across all tasks. Comparing to GEORGE, CDRO achieves higher worst-group accuracy on Waterbirds and CelebA. The perfor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.895, + 0.716 + ], + "angle": 0, + "content": "mance of GEORGE relies on how well the learned features can be clustered. In CMNIST, where the features (color and shape) are easily distinguishable, GEORGE easily finds clusters that correspond to groups and therefore yields a high worst-group accuracy. However, in Waterbirds and CelebA, where the features are more complex, clustering is less effective and the worst-group accuracy is lower. Further, with only 100 out-of-distribution images, CDRO achieves similar performance to methods that require validation labels, which may require large-scale manual labeling: CelebA, for example, has 20K validation images. Specifically, CDRO outperforms JTT by \\(5\\%\\) on Waterbirds, and \\(6.5\\%\\) on CelebA and approaches the performance of Group DRO itself. CDRO outperforms CNC by \\(0.3\\%\\) on Waterbirds, and falls short by \\(0.8\\%\\) on CelebA. CDRO outperforms SSA with \\(5\\%\\) validation labels by \\(1.7\\%\\) on Waterbirds, and \\(1.3\\%\\) on CelebA. Lastly, CDRO outperforms AFR with \\(5\\%\\) validation labels by \\(2.2\\%\\) on Waterbirds, and \\(10.5\\%\\) on CelebA." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.729, + 0.665, + 0.744 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.754, + 0.81, + 0.77 + ], + "angle": 0, + "content": "4.3.1 CDRO effectively infers group labels" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.903 + ], + "angle": 0, + "content": "To examine the effectiveness of CDRO in inferring group labels, we first visualize the distribution of both train and validation data by their cosine similarity scores to the corresponding CAV for all three datasets. As shown in Figure 3, for both train and validation data of the three datasets, data representations with high similarity scores tend to exhibit the concept that the CAV directed to. For example, Figure 3 shows that, given a CAV in the direction of the concept red," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.859, + 0.304 + ], + "angle": 0, + "content": "
MethodGroup InfoCMNIST [1]Waterbirds [17]CelebA [10]
Worst(%)Mean(%)Worst(%)Mean(%)Worst(%)Mean(%)
Group DRO [17]Train & Val74.7 (1.0)190.6 (0.1)89.9 (0.6)92.0 (0.6)88.9 (1.3)93.9 (0.1)
JTT [9]Val74.5 (2.4)90.2 (0.8)83.8 (1.2)89.3 (0.7)81.5 (1.7)88.1 (0.3)
CNC [24]Val77.4 (3.0)90.9 (0.6)88.5 (0.3)90.9 (0.1)88.8 (0.9)89.9 (0.5)
SSA [14]Val--89.0 (0.6)92.2 (0.9)89.8 (1.3)92.8 (0.1)
5% Val--87.1 (0.7)92.6 (0.2)86.7 (1.1)92.8 (0.3)
AFR [15]Val--90.4 (1.1)94.2 (1.2)82.0 (0.5)91.3 (0.3)
5% Val--86.6 (4.7)-77.5 (1.3)-
GEORGE [19]None76.4 (2.3)89.5 (0.3)76.2 (2.0)95.7 (0.5)54.9 (1.9)94.6 (0.2)
ERMNone0.0 (0.0)20.1 (0.2)62.6 (0.3)97.3 (1.0)47.7 (2.1)94.9 (0.3)
CDRO (ours)Concepts69.3 (2.7)85.0 (1.0)88.8 (0.2)90.5 (0.1)88.0 (0.7)90.8 (0.3)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.314, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Table 1. Average and worst-group accuracies of models trained with CDRO and baselines on the benchmarks. CDRO outperforms approaches that do not use group information (GEORGE, ERM) and approaches that use \\(5\\%\\) of Validation (SSA, AFR). CDRO also has comparable performance to approaches that use validation group labels (JTT, CNC, AFR, SSA)." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.361, + 0.418, + 0.439 + ], + "angle": 0, + "content": "
Precision(%)Recall(%)
CMNIST87.096.9
Waterbirds76.383.9
CelebA78.078.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.449, + 0.471, + 0.492 + ], + "angle": 0, + "content": "Table 2. We show the average precision and recall of the pseudogroup labels inferred by CDRO compared to the ground truth. We show CDRO effectively infers group labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.597 + ], + "angle": 0, + "content": "data points with high similarity scores tend to be the red digits. Similarly, given a CAV in the direction of the concept land background, data points with high similarity scores tend to be images with land background. These results justify the decision to label all data points in the highest-mean mixture with the group associated with the concept." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.471, + 0.629 + ], + "angle": 0, + "content": "Further, Table 2 shows that CDRO estimates group labels with high accuracy in all three datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.396, + 0.667 + ], + "angle": 0, + "content": "4.3.2 CDRO is robust to quality of concepts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.81 + ], + "angle": 0, + "content": "From a practical perspective, since CDRO requires practitioners to curate the concepts, the quality of concepts depends on how much time and effort the practitioners spend on curating the concepts. We therefore evaluate the effectiveness of CDRO when the quality of concepts varies on Waterbirds and CelebA. To examine this, we ask the following two questions: 1) Does CDRO require that the concept images be drawn from the training distribution? 2) Does CDRO require a large number of concept images?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.9 + ], + "angle": 0, + "content": "We first fix the size of concepts to be 100 (50 for the concept set and 50 for the contrastive set), and examine the effect of distribution distance between concepts and training data on CDRO. We vary the distance between training distribution and concept distribution by varying the prompts we use for Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.622 + ], + "angle": 0, + "content": "We design the prompts based on how much effort a hypothetical practitioner would spend on curating the concepts. A practitioner may not have specific information about the dataset and decide to use very general prompts, potentially resulting in a low-quality concept set. For the Waterbirds dataset, we use the following prompts, varying the level of effort and domain knowledge: \"A photo of a [water/land] background\" as the \"distant\" distribution, \"A photo of a [water/land]bird habitat\" as the \"somewhat near\" distribution, and \"A photo of a [water/land] background\" as the \"near\" distribution. For the CelebA dataset, we use the following prompts: \"A photo of a [female/male] person\" as the \"distant\" distribution, \"A photo of a [female/male] celebrity\" as the \"somewhat near\" distribution, and \"A photo of a [female/male] celebrity face\" as the \"near\" distribution. We include our reasoning and a discussion regarding these prompts in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.894, + 0.731 + ], + "angle": 0, + "content": "Table 3 shows that CDRO is effective even if the concept images are out of distribution. In both Waterbirds and CelebA, CDRO achieves similar worst group accuracy to the in-distribution set even when the concept distribution is only somewhat near. This result suggests that practitioners need not meticulously curate the concept sets for CDRO to be effective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We then fix the concept distribution to be near training distribution, and vary the size from the list (4, 8, 40, 100, 200), splitting equally between concept set and contrastive set. For example, if size is 4, then the concept and contrastive sets each have two images. Figure 4 shows that CDRO is robust to the size of the concepts. Even when there are only four images, CDRO achieves higher worst-group accuracy than GEORGE on both Waterbirds and CelebA. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC. A small concept set is sufficient in part since the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12431" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.082, + 0.268, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.081, + 0.422, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.424, + 0.082, + 0.576, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.082, + 0.732, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.082, + 0.887, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.202, + 0.483, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.202, + 0.741, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Figure 3. Distributions of similarity between the CAV trained with out-of-distribution concepts and validation samples. Each peak is colored by the ground truth group label. The peaks are separable, and the peak furthest to the right (highest similarity to the concept) corresponds to the ground truth label in all cases, indicating how we can infer spurious attribute labels. Top row: CMNIST validation samples for each of five color concepts. Bottom left: Waterbird validation samples with land as the concept set and water as the contrastive set. Bottom right: CelebA validation samples with women as the concept set and men as the contrastive set." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.415, + 0.432, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.59, + 0.471, + 0.674 + ], + "angle": 0, + "content": "Figure 4. CDRO is robust to size of concepts. Even when there are only 4 images, CDRO achieves higher worst group accuracy than GEORGE on both Waterbirds and CelebA, since the learned embeddings from ERM make it easy to train a classifier. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.677, + 0.458, + 0.77 + ], + "angle": 0, + "content": "
SourceWorst(%)
In DistributionWaterbirds [17]89.3
NearStable Diffusion [16]88.5
Somewhat NearStable Diffusion [16]80.1
DistantStable Diffusion [16]53.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.822 + ], + "angle": 0, + "content": "Table 3. The worst group accuracy given different concept set qualities on Waterbirds. Concept set quality is judged by the level of perceived effort and domain knowledge needed to create." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.832, + 0.471, + 0.863 + ], + "angle": 0, + "content": "input to the GMM is ERM-trained embeddings, such that training data features are already represented." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.419, + 0.882, + 0.512 + ], + "angle": 0, + "content": "
SourceWorst(%)
In DistributionCelebA [10]87.2
NearStable Diffusion [16]87.9
Somewhat NearStable Diffusion [16]87.2
DistantStable Diffusion [16]75.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.522, + 0.894, + 0.565 + ], + "angle": 0, + "content": "Table 4. The worst group accuracy given different concept set qualities on CelebA. The concept set qualities are judged similar to Waterbird concepts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.868, + 0.59 + ], + "angle": 0, + "content": "4.3.3 Concept Correction is Compatible with CNC" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.812 + ], + "angle": 0, + "content": "We demonstrate that Concept Correction is compatible with other robust training methods via CNC. Specifically, CNC consists of two stages: first, CNC trains a model with ERM and uses the predictions of the ERM model to identify majority and minority groups; then CNC uses contrastive learning to maximize the similarity of examples with same class label but different ERM predictions, and minimize the similarity examples with different class labels but same ERM predictions. We replace ERM predictions with inferred labels of spurious attributes, which we generate according to stage 1-3 of Figure 2. Table 5 shows that models trained with concept-based pseudo-labels achieve comparable worst-group accuracy to models trained with ground truth group labels on both train and validation set." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.829, + 0.642, + 0.845 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.9 + ], + "angle": 0, + "content": "We discuss related work on improving robustness to spurious correlations, both with and without group information available." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.902 + ], + "angle": 0, + "content": "1Note that we ran Group DRO on CMNIST according to hyperparameters specified in [24], whereas the result reported by [24] was 78.5(4.5)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12432" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.233, + 0.089, + 0.738, + 0.151 + ], + "angle": 0, + "content": "
MethodGroup InfoWaterbirds Worst(%)CelebA Worst(%)
CNCTrain & Val88.7 (0.4)88.4 (0.1)
Concept CNCConcepts86.2 (0.7)87.2 (0.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.161, + 0.892, + 0.19 + ], + "angle": 0, + "content": "Table 5. Worst-group accuracy of Concept CNC compared to CNC. We show that we can use concept sets to achieve comparable performance to CNC, which uses both train and validation group labels in our implementation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.204, + 0.47, + 0.339 + ], + "angle": 0, + "content": "Improving Robustness with Group Information \nSagawa et al. proposed Group Distributionally Robust Optimization (DRO) [17], which assumes access to spurious attribute labels and optimizes for worst-group loss instead of average loss. The authors show that Group DRO can achieve state-of-the-art worst-group-accuracy. To demonstrate CDRO, we use the same distributionally robust objective of Group DRO but infer group labels instead of assuming they are provided." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.471, + 0.537 + ], + "angle": 0, + "content": "Several proposals to improve worst group accuracy re-sample the training data to balance classes or groups [4, 5, 7, 12]. These approaches show that while a neural network trained via ERM may be biased at the classification layer, it can still learn the core features for the task at the penultimate layer. Therefore it is effective to adjust the weights of the features by fine-tuning on a balanced hold-out set [5, 7], or re-weighting the loss function for majority and minority groups [4], or processing the model outputs post-hoc [12]. However, these methods acknowledge the expense of obtaining spurious attribute labels for a large dataset. Further, multiple spurious correlations can occur in practice [25], complicating reweighting schemes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.471, + 0.884 + ], + "angle": 0, + "content": "Improving Robustness without Group Information: Recent proposals to improve worst-group accuracy without access to group labels typically follow a two-stage paradigm: first train a model with ERM, then use the predictions of the ERM model to identify minority groups, then train another model with emphasis on the minority groups, thereby achieving high accuracy on both the majority and minority groups [5, 9, 14, 15, 19, 20, 22, 24]. For example, JTT [9] uses the incorrect predictions of the ERM classifier as a proxy for the minority group, then trains a model with an upweight on the misclassified examples. Similarly, Correct-N-Contrast (CNC) [24] uses misclassifications as pseudo spurious attribute labels, then uses contrastive learning to align representations for samples within the same class but different spurious attribute. Recognizing that retraining a model from scratch is computationally expensive, Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. However, these methods still require a validation set with group labels for hyperparameter tuning and selecting the best model checkpoint, which turns out to be crucial for achieving their results. CDRO avoids the need of validation group labels by using concepts to infer pseudo group labels." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In [14] and [20], the authors assume access to a small" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.368 + ], + "angle": 0, + "content": "number of group-labeled data and use semi-supervised learning to exploit the group information for better worst group accuracy. Specifically, Spread Spurious Attribute (SSA) [14] trains a spurious attribute predictor using both group-labeled and group unlabeled data, then uses the predicted labels for worst-group loss minimization. Unlike SSA, CDRO does not require concepts to represent task-specific groups, which encapsulate both the spurious attribute and target class. Instead, concepts can be out-of-distribution examples solely representing the spurious attribute." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.522 + ], + "angle": 0, + "content": "GEORGE [19] requires group labels neither during training nor validation, instead inferring group labels by clustering the feature space learned by the ERM model, then running Group DRO. During validation, GEORGE uses worst-cluster accuracy to select the best hyperparameters and model checkpoint, which avoids the need of group labels. However, while GEORGE achieves high worst group accuracy on synthetic datasets where clusters are well-defined, it is less effective on real-world datasets such as CelebA where clusters are harder to distinguish [19]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Some approaches do not follow the two-stage paradigm. Learning from Failure (LfF) [13] trains two models simultaneously, one intended to be biased and one intended to be unbiased. When training the unbiased model, LfF focuses on the examples on which the biased model tends to make mistakes thereby learning to ignore the sources of bias. Further, Taghanaki et al. introduced CIM [21], a contrastive learning approach that learns input-space transformation of the data to preserve task-relevant information." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.619, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.826 + ], + "angle": 0, + "content": "In this work, we present Concept Correction, a framework that uses out-of-distribution concepts to improve worst-group accuracy in the presence of spurious correlations. We demonstrate an instance of the framework via Concept DRO, which uses concepts to infer group labels, and then uses these to train with worst-group loss minimization. We show that Concept DRO achieves competitive results to existing methods that assume access to group labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.844, + 0.681, + 0.861 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by NIST Grant 60NANB23D194, NSF Award 1934464, Cisco Grant 70618863." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12433" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.257 + ], + "angle": 0, + "content": "[3] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[4] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[5] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. Advances in Neural Information Processing Systems, 35:38516-38532, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.443 + ], + "angle": 0, + "content": "[6] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[7] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[8] Anthony Liu, Santiago Guerra, Isaac Fung, Gabriel Matute, Ece Kamar, and Walter Lasecki. Towards hybrid human-air workflows for unknown unknown detection. In Proceedings of The Web Conference 2020, pages 2432-2442, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[9] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[10] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[11] R Thomas McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[12] Aditya Krishna Menon, Ankit Singh Rawat, and Sanjiv Kumar. Overparameterisation and worst-case generalisation: friend or foe? In International Conference on Learning Representations, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accu" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "racy with spurious attribute estimation. arXiv preprint arXiv:2204.02070, 2022. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Shikai Qiu, Andres Potapczynski, Pavel Izmailov, and Andrew Gordon Wilson. Simple and fast group robustness by automatic feature reweighting. arXiv preprint arXiv:2306.11074, 2023. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[17] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.358 + ], + "angle": 0, + "content": "[18] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[19] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[20] Nimit S Sohoni, Maziar Sanjabi, Nicolas Ballas, Aditya Grover, Shaoliang Nie, Hamed Firooz, and Christopher Ré. Barack: Partially supervised group robustness with guarantees. arXiv preprint arXiv:2201.00072, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.49, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[21] Saeid A Taghanaki, Kristy Choi, Amir Hosein Khasahmadi, and Anirudh Goyal. Robust representation learning via perceptual similarity metrics. In International Conference on Machine Learning, pages 10043-10053. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[22] Haotian Ye, James Zou, and Linjun Zhang. Freeze then train: Towards provable representation learning under spurious correlations and feature noise. In International Conference on Artificial Intelligence and Statistics, pages 8968-8990. PMLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.616, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[23] John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS medicine, 15(11):e1002683, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[24] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.758, + 0.892, + 0.811 + ], + "angle": 0, + "content": "[25] Dora Zhao, Jerone Andrews, and Alice Xiang. Men also do laundry: Multi-attribute bias amplification. In International Conference on Machine Learning, pages 42000–42017. PMLR, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12434" + } + ] +] \ No newline at end of file diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_origin.pdf b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a1506379caa3667f441c3becfd08ce3b902cd63c --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/71c0abd3-bdae-44a0-ba02-b4e8dc4dac4c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d58bf59f469c7c1f4d824d7848c7c49247dc40876618fa92d4281c411cfd74b +size 5049131 diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/full.md b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f19b0c2e24f27d8353b6a0fc2755555c752e9665 --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/full.md @@ -0,0 +1,256 @@ +# Label-Efficient Group Robustness via Out-of-Distribution Concept Curation + +Yiwei Yang1 Anthony Z. Liu2 Robert Wolfe1 Aylin Caliskan1 Bill Howe1 +1University of Washington, 2University of Michigan +{yanyiwei, rwolfe3, aylin, billhowe}@uw.edu anthliu@umich.edu + +# Abstract + +Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to $33.1\%$ on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias. + +# 1. Introduction + +Deep learning models are prone to capture spurious correlations, patterns that are predictive of the target class in training data but are inherently irrelevant to the classification task. For example, consider an image classification task to distinguish doctors from nurses. Suppose that in + +the training data, $95\%$ of the doctors present as men, and $95\%$ of the nurses present as women. A model trained with the standard empirical risk minimization (ERM) tends to classify images of men as doctors and images of women as nurses by learning a spurious correlation between gender-expression features (e.g., hairstyle, clothing) and occupation. While the model may achieve high average accuracy, it performs poorly on the minority groups (in this example, female-presenting doctors and male-presenting nurses), thus exhibiting low worst-group accuracy. Such spurious correlations happen in a wide variety of real-world tasks, including facial recognition, medical imaging, and language tasks [3, 8, 11, 17, 18, 23]. + +Existing approaches typically consider the setting in which each training (or validation) sample is explicitly assigned a group, represented as a pair $(y, a)$ where $y$ is a class label and $a$ is a value from the (potentially unknown) spurious attribute. For instance, Group DRO assumes the availability of group labels during training, and directly minimizes the worst-group loss [17]. Recognizing that obtaining labels requires expensive human annotation, several recent approaches avoid relying on group labels during training [7, 9, 19, 24]. These approaches train an ERM model to infer group labels, then a second model is trained to optimize for worst-group loss with the inferred labels. However, these approaches require validation labels for crucial hyperparameter tuning [9], since their effectiveness depends on the accuracy of the group inference. One exception is GEORGE [19], which assumes that the feature space of a model trained for the predictive task is easily separable by groups and learns group assignment via clustering. GEORGE then optimizes for worst-cluster accuracy without requiring labels. However, while GEORGE is effective when the input features are simple and cluster well (e.g. colored digits), its performance drops significantly when the input features are more complex and clusters become less well-defined (e.g., for human faces) [19] + +We propose Concept Correction (figure 1), a framework that instead uses a set of example images to represent a concept involved in the spurious correlation. These examples can be out-of-distribution and task-agnostic, as long as + +![](images/e6b093bec8798c3637a5d3318f19a10f7f573de96c46fbb7b9f445e8753d83c1.jpg) +Figure 1. We show our main idea Concept Correction, a framework for correcting spurious attributes in biased models using concepts, set of examples representing some pattern. We show an example from the Waterbirds dataset [17], where (a) directly training an ERM results in a model that spuriously correlates background type with bird type. A practitioner can correct this by (b) finding images with land and water backgrounds and using concept correction. Notably, these example images can be obtained outside of the original Waterbirds dataset, such as using generative AI, and using search engines. + +they contain the features representing the spurious attribute. For example, concept sets designed to counteract a spurious attribute of gender presentation may include a set of images of female-presenting people, but without any reference to the task-specific class label of doctor or nurse. Since the examples need not be selected from the training distribution, they are reusable across models and tasks, and practitioners can retrieve them from various sources such as a search engines, datasets on the Web, or image generation models, all of which require significantly less cognitive and technical effort than manually labeling in-distribution data. In fact, by using images generated with only two lines of prompts to the text-to-image generator Stable Diffusion [16], we show that an implementation of Concept Correction can achieve competitive results to state-of-the-art methods that require manual labeling of tens of thousands of data points [9, 17, 24]. + +We demonstrate an instance of Concept Correction via Concept DRO (CDRO), an approach that leverages concepts to infer group labels, then trains with a distributively robust optimization objective using the inferred labels. Similar to the existing two-stage methods, CDRO first trains a neural network with ERM. Then, exploiting the fact that the ERM model learns spurious attribute labels[7], we use the model to extract features from the concept sets, the training data, and the validation data. We then train a classifier to discriminate concepts then use the weights of this classifier to determine the extent to which each training sample exhibits that concept. These concept similarity scores are then used to infer a spurious attribute label for every data point in the training and validation sets, providing the information needed by state-of-the-art methods to improve + +worst-group accuracy. In our example, CDRO trains a classifier on two sets of images representing the concepts of presenting-female and presenting-male, then uses the coefficients of this classifier to estimate the extent to which each training sample exhibits "maleness" vs. "femaleness." + +We evaluate CDRO on three image classification datasets with spurious correlations: CMNIST [1], Waterbirds [17], and CelebA [10]. Among methods that do not assume spurious attribute labels are provided, CDRO improves worstgroup accuracy by up to $12.6\%$ on Waterbirds and $33.1\%$ on CelebA. CDRO also achieves competitive performance to methods that do assume access to spurious attribute labels. Finally, we show that CDRO is robust to the quality of concepts in two aspects: size of concept and distribution distance from training data. + +Our contributions are summarized as follows: + +- We propose Concept Correction, a label-efficient framework that uses out-of-distribution sets of images to correct spurious correlations. +- We present CDRO, an instance of the Concept Correction framework that uses concepts to infer labels of spurious attributes and then trains with a distributively robust optimization objective. +- We show CDRO significantly outperforms ERM and GEORGE across three benchmarks while achieving competitive results to methods that require group labels during validation. + +# 2. Preliminaries + +In this section, we first present the problem setting, then describe concepts and the Concept Activation Vector (CAV) introduced by Kim et al. [6]. + +# 2.1. Problem Setting + +We consider a classification setting, where given an input $x \in \mathcal{X}$ , our goal is to predict $y \in \mathcal{Y}$ . Let $X = \{x_{1}, \ldots, x_{n}\}$ , $Y = \{y_{1}, \ldots, y_{n}\}$ be our training set of size $n$ . Each datapoint $x_{i}$ has a spurious attribute $a_{i}$ , which is unobserved during training. Each $x_{i}$ is assumed to belong to a group $g_{i}$ where $g_{i}$ is a pair $(y_{i}, a_{i})$ , the combination of a class label $y_{i}$ and a spurious attribute label $a_{i}$ . The spurious correlation problem occurs when accuracy across groups varies significantly; i.e., a model trained with ERM may achieve a high average accuracy, yet still performs poorly on some minority groups. We therefore evaluate classifiers on the accuracy of the worst-performing group. + +We are interested in the setting where group information is not available for either the training set nor the validation set. Instead, we assume that users curate a set of images representing the concept without regard to any particular task. The human effort to curate the concept depends on the user's domain knowledge. For example, in Waterbirds, to curate a concept for 'landbird background,' one can generate images with the prompt 'a photo of land background'. However, if the user knows that land backgrounds of birds tend to be forests, the prompt can be adjusted accordingly. + +# 2.2. Concepts and Concept Activation Vectors + +A concept is defined using a set of images. For example, a set of ocean and lake images can represent the concept water background. The concept examples need not be drawn from the training data and need not be relevant to the classification task. + +Given a concept $C$ and a layer $l$ of the neural network, a concept activation vector (CAV) $\mathbf{v}_C^l \in R^m$ [6] is defined as the vector normal to the hyperplane separating samples with a concept (the concept set $P_C$ ) and examples without a concept (the contrastive set $N$ ) in the model's activations. Given layer activations of both the concept set and the contrastive set, CAV can be obtained by training a linear classifier distinguishing the two sets of activations. + +# 3. Concept Correction + +We propose Concept Correction, a framework that uses concepts to correct spurious correlations, without relying on manual labeling of in-distribution data. Instead, Concept Correction takes a set of a small number of examples as input and leverages the examples to infer spurious attribute information. Then, a robust model is trained with the inferred group of spurious attribute information using any worst-group loss minimization algorithms. + +Concept Correction expects the concept examples to represent the spurious attribute. As described in the introduction, these images can be out-of-distribution and irrelevant to the prediction task, de-coupling concept curation from + +the model setting. In fact, using generative models, practitioners can easily generate many examples with just a few lines of prompting. In contrast to existing methods that require manual labeling of either training or validation samples, concept curation does not require access to the training data. Practitioners can reason about potential biases the model may exhibit then curate concept sets accordingly. + +Concept Correction is compatible with any method that requires spurious attribute or group information, such as Group DRO [17], Just-Train-Twice [9], Correct-n-Contrast [24], and more. For example, a simple way of applying Concept Correction to Correct-n-Contrast and JustTrain-Twice is simply replacing the validation group labels with inferred labels using concepts, alleviating the need of manual labeling of a large amount of data. + +# 3.1. Concept DRO + +We present Concept DRO (CDRO), an implementation of Concept Correction on Group DRO. As illustrated in Figure 2, in stage 1, we first train an ERM model and then train a linear classifier on concepts representing the spurious attribute to infer group labels. In stage 2, we use Group DRO to train a robust model with the inferred group labels. We include further details on both stages below, and give pseudocode for CDRO in the Appendix. + +# 3.1.1 Stage 1: Inferring Labels of Spurious Attributes + +We train a neural network with ERM on the training data. Given a spurious attribute $\mathcal{A} = \{a_1,\dots,a_m\}$ , we curate two sets of examples for each concept $a_i$ : a concept set $C = \{c_1,\dots,c_k\}$ , which represents the presence of the concept, and a contrastive set $N = \{n_1,\dots,n_k\}$ , which represents the absence of the concept. We then use the ERM model as a feature extractor, by taking the first to the penultimate layer of the model, denoted as $f$ , to extract features from the two sets. We then train a linear classifier on the ERM features of the concept sets, $\{f(c_1),\dots,f(c_k)\}$ and $\{f(n_1),\dots,f(n_k)\}$ . The resulting coefficient of the linear classifier gives us a CAV pointing in the direction of the concept $a_i$ . Then, for both training and validation data, denoted as $X = \{x_1,\dots,x_n\}$ , we use $f$ to extract representations, $\{f(x_1),\dots,f(x_n)\}$ . + +We then compute the cosine similarity between the representations and the CAV. We use these similarity scores to separate the data into groups — examples with the same attributes are likely to activate similar similarity scores. To separate points by similarity scores, we use a naive approach: we train a Gaussian Mixture Model (GMM) on the similarity scores, where the number of mixtures is set to $m$ , the size of the domain of the spurious attribute. Typically (but not exclusively) $m = 2$ , since many formulations of bias posit a binary distinction between minority and major- + +ity populations. The mixture with the highest mean similarity to the concept is labeled as $\mathcal{A} = a_{i}$ . Intuitively, data with high similarity scores with respect to the concept set tend to exhibit concept $a_{i}$ , and thus belong to groups with $\mathcal{A} = a_{i}$ (see figure 3). We repeat the above steps for all $a_{i} \in \{a_{1}, \dots, a_{m}\}$ . Lastly, for the data points with missing labels, we randomly assign one from $\{a_{1}, \dots, a_{m}\}$ . + +In the case that the spurious attribute is binary, i.e. $\mathcal{A} = \{a_1,a_2\}$ , we only need to curate concepts for $a_1$ or $a_2$ and train one CAV. Suppose we train a CAV for $a_1$ , and label data with the CAV, yielding $X_{a_1} = \{x\in X|\mathcal{A}(x) = a_1\}$ . We can then use the negative of the CAV to label $a_2$ , which can be considered the complement, $X_{a_1}^C$ . + +Since Group DRO takes group labels as input, where a group label $g$ is the pair $(y, a)$ defined by the label $y$ and spurious attribute $a$ . We generate inferred group labels $\hat{g}$ using ground truth labels $y$ and inferred spurious attribute labels $\hat{a}$ to use in Stage 2. + +# 3.1.2 Stage 2: Optimize for worst-group loss with inferred labels + +We use Group DRO to train a classifier with inferred group labels $\hat{g}$ from stage 1. Instead of overall loss, Group DRO optimizes for worst-group loss. + +Group DRO uses ground truth group labels to select the best model checkpoint based on validation worst-group accuracy. We instead use the inferred labels to compute the surrogate worst-group accuracy at validation. + +# 4. Experimental Results + +In our experiments, we first show that CDRO substantially outperforms ERM and GEORGE, and achieves comparable worst-group accuracy to existing methods that require group labels. To explain the effectiveness of CDRO, we then demonstrate that CDRO separates data into corresponding groups with concepts, and infers labels of spurious attributes with high accuracy. Lastly, we show that CDRO is robust to the quality of concepts in two aspects: concept size and distribution distance from training data. + +# 4.1. Setup + +We describe three image datasets, corresponding concept sets, and baseline methods used in our experiments. + +# 4.1.1 Datasets + +We briefly describe the three image classification benchmarks used in our evaluation. Full dataset and training details are included in the Appendix. + +CMNIST [1]: The task is to classify digits into one of the five classes: $\mathcal{Y} = \{(0,1),(2,3),(4,5),(6,7),(8,9)\}$ . Each class is correlated with one of five colors: $\mathcal{A}$ + +$= \{$ red, yellow, green, blue, purple\}, respectively, each represented by its hex code (see figure 3). + +Waterbirds [17]: The task is to classify birds into one of the 2 classes: $\mathcal{Y} = \{\text{waterbird, landbird}\}$ . Each class is correlated with the backgrounds: $\mathcal{A} = \{\text{water background, land background}\}$ , respectively. + +CelebA [10]: The task is to classify celebrities' hair color $\mathcal{Y} = \{\text{blond, not blond}\}$ , with the spurious attribute $\mathcal{A} = \{\text{female, male}\}$ . The class blond correlates with female, and the class not blond correlates with male. + +# 4.1.2 Baselines + +We consider seven baselines that train models under different assumptions regarding the availability of group labels. + +Group labels available during training - Group DRO [17] is a state-of-the-art method, which optimizes for the worst-group loss using group labels. + +Group labels available during validation - Just Train Twice (JTT) [9] trains a model to detect minority group examples, then upweights those examples during training. Correct-n-Contrast (CNC) [24] infers group information similar to JTT, then uses contrastive learning to learn representations robust to spurious correlations. Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. Spread Spurious Attribute (SSA) [14] uses semi-supervised learning to train a predictor to infer pseudo-group label, then trains a robust model with worst-group loss minimization. + +Group labels available during neither training nor validation - Empirical Risk Minimization (ERM) is the standard training procedure that minimizes the average loss. Similar to JTT, GEORGE also follows a two-step procedure by first estimating the group labels and then using these estimates to train a robust classifier. Unlike JTT, after training a standard model, the feature space of each class is split into groups via unsupervised clustering. During validation, GEORGE then optimizes for worst-cluster accuracy without requiring group labels. + +# 4.1.3 Concepts + +In our experiments we generated 50 images per concept set. We generated synthetic data as concepts for CMNIST, and used a popular off-the-shelf generative AI model Stable Diffusion [16] to generate concepts for Waterbirds and CelebA. + +CMNIST - Since the 5 target classes are correlated with 5 colors, we curate concept sets for each color. For each color, we sample 50 images from the original MNIST dataset [2], which consists of white digits 0-9 on black background, and we paint the digits with the corresponding color. We then curate the contrastive set by concatenating the concept sets of the 4 remaining colors, and trained a CAV pointing towards the color. + +![](images/384035a350053ee80de40b000ea2154758c35d94ddb37fe8f960958226157ee5.jpg) +Figure 2. We show Concept DRO (CDRO), a method of training an unbiased model using a concept set. (1) First, we train a neural network model $(h\circ f)$ using ERM. (2) We train a concept activation vector (CAV) using the given concept set and the model [6]. A CAV is a linear decision boundary between the model representations of the concepts $(f(\cdot))$ . (3) We use the inferred CAV to infer pseudo-group labels over the training data. (4) Finally, we run Group DRO [17] the training data using the pseudo-labels. + +Waterbirds - Since land background is correlated with landbirds, and water background is correlated with waterbirds, we curate sets of examples to represent the concepts "land background" and "water background." With the knowledge that land background in the dataset are either bamboo or broadleaf trees, we used prompts "A photo of bamboo trees" and "A photo of broadleaf trees" to generate a total of 50 (25 each) images to represent the "land background" concept. Similarly, knowing that the water background in the dataset are either ocean or lake, we used prompts "A photo of ocean" and "A photo of lake" to generate a total of 50 (25 each) images to represent the "water background" concept. Since the spurious attribute is binary, we trained a CAV in the direction of "land background," using the "water background" concept as the contrastive set. + +CelebA - Since female is correlated with blond and male is correlated with not blond, we curate sets of examples to represent the concepts female and male. With the knowledge that the dataset consists of celebrities' faces, we used prompts "A photo of a female celebrity face" and "A photo of a male celebrity face" to generate a total of 100 images (50 each) for the female and male concepts, respectively. We trained a CAV with female as the concept set and male as the contrastive set, such that positive similarity can be interpreted as exhibiting female features. + +# 4.2. Main Results + +Table 1 reports the average and worst-group accuracies of all approaches. CDRO outperforms ERM across all tasks. Comparing to GEORGE, CDRO achieves higher worst-group accuracy on Waterbirds and CelebA. The perfor + +mance of GEORGE relies on how well the learned features can be clustered. In CMNIST, where the features (color and shape) are easily distinguishable, GEORGE easily finds clusters that correspond to groups and therefore yields a high worst-group accuracy. However, in Waterbirds and CelebA, where the features are more complex, clustering is less effective and the worst-group accuracy is lower. Further, with only 100 out-of-distribution images, CDRO achieves similar performance to methods that require validation labels, which may require large-scale manual labeling: CelebA, for example, has 20K validation images. Specifically, CDRO outperforms JTT by $5\%$ on Waterbirds, and $6.5\%$ on CelebA and approaches the performance of Group DRO itself. CDRO outperforms CNC by $0.3\%$ on Waterbirds, and falls short by $0.8\%$ on CelebA. CDRO outperforms SSA with $5\%$ validation labels by $1.7\%$ on Waterbirds, and $1.3\%$ on CelebA. Lastly, CDRO outperforms AFR with $5\%$ validation labels by $2.2\%$ on Waterbirds, and $10.5\%$ on CelebA. + +# 4.3. Ablation Studies + +# 4.3.1 CDRO effectively infers group labels + +To examine the effectiveness of CDRO in inferring group labels, we first visualize the distribution of both train and validation data by their cosine similarity scores to the corresponding CAV for all three datasets. As shown in Figure 3, for both train and validation data of the three datasets, data representations with high similarity scores tend to exhibit the concept that the CAV directed to. For example, Figure 3 shows that, given a CAV in the direction of the concept red, + +
MethodGroup InfoCMNIST [1]Waterbirds [17]CelebA [10]
Worst(%)Mean(%)Worst(%)Mean(%)Worst(%)Mean(%)
Group DRO [17]Train & Val74.7 (1.0)190.6 (0.1)89.9 (0.6)92.0 (0.6)88.9 (1.3)93.9 (0.1)
JTT [9]Val74.5 (2.4)90.2 (0.8)83.8 (1.2)89.3 (0.7)81.5 (1.7)88.1 (0.3)
CNC [24]Val77.4 (3.0)90.9 (0.6)88.5 (0.3)90.9 (0.1)88.8 (0.9)89.9 (0.5)
SSA [14]Val--89.0 (0.6)92.2 (0.9)89.8 (1.3)92.8 (0.1)
5% Val--87.1 (0.7)92.6 (0.2)86.7 (1.1)92.8 (0.3)
AFR [15]Val--90.4 (1.1)94.2 (1.2)82.0 (0.5)91.3 (0.3)
5% Val--86.6 (4.7)-77.5 (1.3)-
GEORGE [19]None76.4 (2.3)89.5 (0.3)76.2 (2.0)95.7 (0.5)54.9 (1.9)94.6 (0.2)
ERMNone0.0 (0.0)20.1 (0.2)62.6 (0.3)97.3 (1.0)47.7 (2.1)94.9 (0.3)
CDRO (ours)Concepts69.3 (2.7)85.0 (1.0)88.8 (0.2)90.5 (0.1)88.0 (0.7)90.8 (0.3)
+ +Table 1. Average and worst-group accuracies of models trained with CDRO and baselines on the benchmarks. CDRO outperforms approaches that do not use group information (GEORGE, ERM) and approaches that use $5\%$ of Validation (SSA, AFR). CDRO also has comparable performance to approaches that use validation group labels (JTT, CNC, AFR, SSA). + +
Precision(%)Recall(%)
CMNIST87.096.9
Waterbirds76.383.9
CelebA78.078.5
+ +Table 2. We show the average precision and recall of the pseudogroup labels inferred by CDRO compared to the ground truth. We show CDRO effectively infers group labels. + +data points with high similarity scores tend to be the red digits. Similarly, given a CAV in the direction of the concept land background, data points with high similarity scores tend to be images with land background. These results justify the decision to label all data points in the highest-mean mixture with the group associated with the concept. + +Further, Table 2 shows that CDRO estimates group labels with high accuracy in all three datasets. + +# 4.3.2 CDRO is robust to quality of concepts + +From a practical perspective, since CDRO requires practitioners to curate the concepts, the quality of concepts depends on how much time and effort the practitioners spend on curating the concepts. We therefore evaluate the effectiveness of CDRO when the quality of concepts varies on Waterbirds and CelebA. To examine this, we ask the following two questions: 1) Does CDRO require that the concept images be drawn from the training distribution? 2) Does CDRO require a large number of concept images? + +We first fix the size of concepts to be 100 (50 for the concept set and 50 for the contrastive set), and examine the effect of distribution distance between concepts and training data on CDRO. We vary the distance between training distribution and concept distribution by varying the prompts we use for Stable Diffusion. + +We design the prompts based on how much effort a hypothetical practitioner would spend on curating the concepts. A practitioner may not have specific information about the dataset and decide to use very general prompts, potentially resulting in a low-quality concept set. For the Waterbirds dataset, we use the following prompts, varying the level of effort and domain knowledge: "A photo of a [water/land] background" as the "distant" distribution, "A photo of a [water/land]bird habitat" as the "somewhat near" distribution, and "A photo of a [water/land] background" as the "near" distribution. For the CelebA dataset, we use the following prompts: "A photo of a [female/male] person" as the "distant" distribution, "A photo of a [female/male] celebrity" as the "somewhat near" distribution, and "A photo of a [female/male] celebrity face" as the "near" distribution. We include our reasoning and a discussion regarding these prompts in the Appendix. + +Table 3 shows that CDRO is effective even if the concept images are out of distribution. In both Waterbirds and CelebA, CDRO achieves similar worst group accuracy to the in-distribution set even when the concept distribution is only somewhat near. This result suggests that practitioners need not meticulously curate the concept sets for CDRO to be effective. + +We then fix the concept distribution to be near training distribution, and vary the size from the list (4, 8, 40, 100, 200), splitting equally between concept set and contrastive set. For example, if size is 4, then the concept and contrastive sets each have two images. Figure 4 shows that CDRO is robust to the size of the concepts. Even when there are only four images, CDRO achieves higher worst-group accuracy than GEORGE on both Waterbirds and CelebA. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC. A small concept set is sufficient in part since the + +![](images/ebdf4d58afa644bfa27870431a34425f2d0845623f803aa1c29383ef401399b7.jpg) + +![](images/3f268eb9a50068ec790dbf0fce63c934472a6a3afa0d70742d8ccd88996658fe.jpg) + +![](images/cf3b6303c3a42ae07f70167c4baa68de9a71e298664cde0ebede13dc66bd6202.jpg) + +![](images/5fcef310329df0b558cc28da4ed3f18fb7568fa6859c4bc2920bd8f2faf7f148.jpg) + +![](images/483da73f71d471fe04ea562534360b3b33dec50af129c3bb8d365c8b6b4ee5f9.jpg) + +![](images/cfb52e8ec05d6454499a01e242c54cdefda5dfa67dea34c92623b88739324739.jpg) +Figure 3. Distributions of similarity between the CAV trained with out-of-distribution concepts and validation samples. Each peak is colored by the ground truth group label. The peaks are separable, and the peak furthest to the right (highest similarity to the concept) corresponds to the ground truth label in all cases, indicating how we can infer spurious attribute labels. Top row: CMNIST validation samples for each of five color concepts. Bottom left: Waterbird validation samples with land as the concept set and water as the contrastive set. Bottom right: CelebA validation samples with women as the concept set and men as the contrastive set. + +![](images/f1fc5d0e9451bf8baa00f0134cef9b10ac18f73827e738ac22b97253d6b58904.jpg) + +![](images/0444e06a7a4bd6bd4d3b231035e164a0c1b5df75a1502717142e746f9dfeb78b.jpg) +Figure 4. CDRO is robust to size of concepts. Even when there are only 4 images, CDRO achieves higher worst group accuracy than GEORGE on both Waterbirds and CelebA, since the learned embeddings from ERM make it easy to train a classifier. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC. + +
SourceWorst(%)
In DistributionWaterbirds [17]89.3
NearStable Diffusion [16]88.5
Somewhat NearStable Diffusion [16]80.1
DistantStable Diffusion [16]53.1
+ +input to the GMM is ERM-trained embeddings, such that training data features are already represented. + +Table 3. The worst group accuracy given different concept set qualities on Waterbirds. Concept set quality is judged by the level of perceived effort and domain knowledge needed to create. + +
SourceWorst(%)
In DistributionCelebA [10]87.2
NearStable Diffusion [16]87.9
Somewhat NearStable Diffusion [16]87.2
DistantStable Diffusion [16]75.6
+ +Table 4. The worst group accuracy given different concept set qualities on CelebA. The concept set qualities are judged similar to Waterbird concepts. + +# 4.3.3 Concept Correction is Compatible with CNC + +We demonstrate that Concept Correction is compatible with other robust training methods via CNC. Specifically, CNC consists of two stages: first, CNC trains a model with ERM and uses the predictions of the ERM model to identify majority and minority groups; then CNC uses contrastive learning to maximize the similarity of examples with same class label but different ERM predictions, and minimize the similarity examples with different class labels but same ERM predictions. We replace ERM predictions with inferred labels of spurious attributes, which we generate according to stage 1-3 of Figure 2. Table 5 shows that models trained with concept-based pseudo-labels achieve comparable worst-group accuracy to models trained with ground truth group labels on both train and validation set. + +# 5. Related Work + +We discuss related work on improving robustness to spurious correlations, both with and without group information available. + +
MethodGroup InfoWaterbirds Worst(%)CelebA Worst(%)
CNCTrain & Val88.7 (0.4)88.4 (0.1)
Concept CNCConcepts86.2 (0.7)87.2 (0.1)
+ +Table 5. Worst-group accuracy of Concept CNC compared to CNC. We show that we can use concept sets to achieve comparable performance to CNC, which uses both train and validation group labels in our implementation. + +Improving Robustness with Group Information +Sagawa et al. proposed Group Distributionally Robust Optimization (DRO) [17], which assumes access to spurious attribute labels and optimizes for worst-group loss instead of average loss. The authors show that Group DRO can achieve state-of-the-art worst-group-accuracy. To demonstrate CDRO, we use the same distributionally robust objective of Group DRO but infer group labels instead of assuming they are provided. + +Several proposals to improve worst group accuracy re-sample the training data to balance classes or groups [4, 5, 7, 12]. These approaches show that while a neural network trained via ERM may be biased at the classification layer, it can still learn the core features for the task at the penultimate layer. Therefore it is effective to adjust the weights of the features by fine-tuning on a balanced hold-out set [5, 7], or re-weighting the loss function for majority and minority groups [4], or processing the model outputs post-hoc [12]. However, these methods acknowledge the expense of obtaining spurious attribute labels for a large dataset. Further, multiple spurious correlations can occur in practice [25], complicating reweighting schemes. + +Improving Robustness without Group Information: Recent proposals to improve worst-group accuracy without access to group labels typically follow a two-stage paradigm: first train a model with ERM, then use the predictions of the ERM model to identify minority groups, then train another model with emphasis on the minority groups, thereby achieving high accuracy on both the majority and minority groups [5, 9, 14, 15, 19, 20, 22, 24]. For example, JTT [9] uses the incorrect predictions of the ERM classifier as a proxy for the minority group, then trains a model with an upweight on the misclassified examples. Similarly, Correct-N-Contrast (CNC) [24] uses misclassifications as pseudo spurious attribute labels, then uses contrastive learning to align representations for samples within the same class but different spurious attribute. Recognizing that retraining a model from scratch is computationally expensive, Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. However, these methods still require a validation set with group labels for hyperparameter tuning and selecting the best model checkpoint, which turns out to be crucial for achieving their results. CDRO avoids the need of validation group labels by using concepts to infer pseudo group labels. + +In [14] and [20], the authors assume access to a small + +number of group-labeled data and use semi-supervised learning to exploit the group information for better worst group accuracy. Specifically, Spread Spurious Attribute (SSA) [14] trains a spurious attribute predictor using both group-labeled and group unlabeled data, then uses the predicted labels for worst-group loss minimization. Unlike SSA, CDRO does not require concepts to represent task-specific groups, which encapsulate both the spurious attribute and target class. Instead, concepts can be out-of-distribution examples solely representing the spurious attribute. + +GEORGE [19] requires group labels neither during training nor validation, instead inferring group labels by clustering the feature space learned by the ERM model, then running Group DRO. During validation, GEORGE uses worst-cluster accuracy to select the best hyperparameters and model checkpoint, which avoids the need of group labels. However, while GEORGE achieves high worst group accuracy on synthetic datasets where clusters are well-defined, it is less effective on real-world datasets such as CelebA where clusters are harder to distinguish [19]. + +Some approaches do not follow the two-stage paradigm. Learning from Failure (LfF) [13] trains two models simultaneously, one intended to be biased and one intended to be unbiased. When training the unbiased model, LfF focuses on the examples on which the biased model tends to make mistakes thereby learning to ignore the sources of bias. Further, Taghanaki et al. introduced CIM [21], a contrastive learning approach that learns input-space transformation of the data to preserve task-relevant information. + +# 6. Conclusion + +In this work, we present Concept Correction, a framework that uses out-of-distribution concepts to improve worst-group accuracy in the presence of spurious correlations. We demonstrate an instance of the framework via Concept DRO, which uses concepts to infer group labels, and then uses these to train with worst-group loss minimization. We show that Concept DRO achieves competitive results to existing methods that assume access to group labels. + +# 7. Acknowledgement + +This work was supported by NIST Grant 60NANB23D194, NSF Award 1934464, Cisco Grant 70618863. + +# References + +[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2, 4, 6 +[2] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 4 +[3] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1 +[4] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 8 +[5] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. Advances in Neural Information Processing Systems, 35:38516-38532, 2022. 8 +[6] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5 +[7] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 1, 2, 8 +[8] Anthony Liu, Santiago Guerra, Isaac Fung, Gabriel Matute, Ece Kamar, and Walter Lasecki. Towards hybrid human-air workflows for unknown unknown detection. In Proceedings of The Web Conference 2020, pages 2432-2442, 2020. 1 +[9] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 1, 2, 3, 4, 6, 8 +[10] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 2, 4, 6, 7 +[11] R Thomas McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007, 2019. 1 +[12] Aditya Krishna Menon, Ankit Singh Rawat, and Sanjiv Kumar. Overparameterisation and worst-case generalisation: friend or foe? In International Conference on Learning Representations, 2020. 8 +[13] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020. 8 +[14] Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accu + +racy with spurious attribute estimation. arXiv preprint arXiv:2204.02070, 2022. 4, 6, 8 +[15] Shikai Qiu, Andres Potapczynski, Pavel Izmailov, and Andrew Gordon Wilson. Simple and fast group robustness by automatic feature reweighting. arXiv preprint arXiv:2306.11074, 2023. 4, 6, 8 +[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4, 7 +[17] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 1, 2, 3, 4, 5, 6, 7, 8 +[18] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 1 +[19] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1, 6, 8 +[20] Nimit S Sohoni, Maziar Sanjabi, Nicolas Ballas, Aditya Grover, Shaoliang Nie, Hamed Firooz, and Christopher Ré. Barack: Partially supervised group robustness with guarantees. arXiv preprint arXiv:2201.00072, 2021. 8 +[21] Saeid A Taghanaki, Kristy Choi, Amir Hosein Khasahmadi, and Anirudh Goyal. Robust representation learning via perceptual similarity metrics. In International Conference on Machine Learning, pages 10043-10053. PMLR, 2021. 8 +[22] Haotian Ye, James Zou, and Linjun Zhang. Freeze then train: Towards provable representation learning under spurious correlations and feature noise. In International Conference on Artificial Intelligence and Statistics, pages 8968-8990. PMLR, 2023. 8 +[23] John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS medicine, 15(11):e1002683, 2018. 1 +[24] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. 1, 2, 3, 4, 6, 7, 8 +[25] Dora Zhao, Jerone Andrews, and Alice Xiang. Men also do laundry: Multi-attribute bias amplification. In International Conference on Machine Learning, pages 42000–42017. PMLR, 2023. 8 \ No newline at end of file diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/images.zip b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..af5cc1ee1c713fee9d05cb708c466acf5abe7fda --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:643b7b28baa95731c1d90c81fd879f5c88ca17c6321a40a6486cd3bbfe0fe466 +size 448691 diff --git a/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/layout.json b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f58710140cb540a4c764fc6032ab31a56fb410b0 --- /dev/null +++ b/2024/Label-Efficient Group Robustness via Out-of-Distribution Concept Curation/layout.json @@ -0,0 +1,6974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "Label-Efficient Group Robustness via Out-of-Distribution Concept Curation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 517, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 517, + 185 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 517, + 185 + ], + "type": "text", + "content": "Yiwei Yang1 Anthony Z. Liu2 Robert Wolfe1 Aylin Caliskan1 Bill Howe1 \n1University of Washington, 2University of Michigan \n{yanyiwei, rwolfe3, aylin, billhowe}@uw.edu anthliu@umich.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 289, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 610 + ], + "type": "text", + "content": "Deep neural networks are prone to capture correlations between spurious attributes and class labels, leading to low accuracy on some combinations of class labels and spurious attribute values. When a spurious attribute represents a protected class, these low-accuracy groups can manifest discriminatory bias. Existing methods attempting to improve worst-group accuracy assume the training data, validation data, or both are reliably labeled by the spurious attribute. But a model may be perceived to be biased towards a concept that is not represented by pre-existing labels on the training data. In these situations, the spurious attribute must be defined with external information. We propose Concept Correction, a framework that represents a concept as a curated set of images from any source, then labels each training sample by its similarity to the concept set to control spurious correlations. For example, concept sets representing gender can be used to measure and control gender bias even without explicit labels. We demonstrate and evaluate an instance of the framework as Concept DRO, which uses concept sets to estimate group labels, then uses these labels to train with a state of the art distributively robust optimization objective. We show that Concept DRO outperforms existing methods that do not require labels of spurious attributes by up to " + }, + { + "bbox": [ + 46, + 238, + 289, + 610 + ], + "type": "inline_equation", + "content": "33.1\\%" + }, + { + "bbox": [ + 46, + 238, + 289, + 610 + ], + "type": "text", + "content": " on three image classification datasets and is competitive with the best methods that assume access to labels. We consider how the size and quality of the concept set influences performance and find that even smaller, manually curated sets of noisy AI-generated images are effective at controlling spurious correlations, suggesting that high-quality, reusable concept sets are easy to create and effective in reducing bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 128, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Deep learning models are prone to capture spurious correlations, patterns that are predictive of the target class in training data but are inherently irrelevant to the classification task. For example, consider an image classification task to distinguish doctors from nurses. Suppose that in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "text", + "content": "the training data, " + }, + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "text", + "content": " of the doctors present as men, and " + }, + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 213, + 547, + 370 + ], + "type": "text", + "content": " of the nurses present as women. A model trained with the standard empirical risk minimization (ERM) tends to classify images of men as doctors and images of women as nurses by learning a spurious correlation between gender-expression features (e.g., hairstyle, clothing) and occupation. While the model may achieve high average accuracy, it performs poorly on the minority groups (in this example, female-presenting doctors and male-presenting nurses), thus exhibiting low worst-group accuracy. Such spurious correlations happen in a wide variety of real-world tasks, including facial recognition, medical imaging, and language tasks [3, 8, 11, 17, 18, 23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "text", + "content": "Existing approaches typically consider the setting in which each training (or validation) sample is explicitly assigned a group, represented as a pair " + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "inline_equation", + "content": "(y, a)" + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "text", + "content": " is a class label and " + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 374, + 547, + 662 + ], + "type": "text", + "content": " is a value from the (potentially unknown) spurious attribute. For instance, Group DRO assumes the availability of group labels during training, and directly minimizes the worst-group loss [17]. Recognizing that obtaining labels requires expensive human annotation, several recent approaches avoid relying on group labels during training [7, 9, 19, 24]. These approaches train an ERM model to infer group labels, then a second model is trained to optimize for worst-group loss with the inferred labels. However, these approaches require validation labels for crucial hyperparameter tuning [9], since their effectiveness depends on the accuracy of the group inference. One exception is GEORGE [19], which assumes that the feature space of a model trained for the predictive task is easily separable by groups and learns group assignment via clustering. GEORGE then optimizes for worst-cluster accuracy without requiring labels. However, while GEORGE is effective when the input features are simple and cluster well (e.g. colored digits), its performance drops significantly when the input features are more complex and clusters become less well-defined (e.g., for human faces) [19]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "We propose Concept Correction (figure 1), a framework that instead uses a set of example images to represent a concept involved in the spurious correlation. These examples can be out-of-distribution and task-agnostic, as long as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 92, + 62, + 492, + 258 + ], + "blocks": [ + { + "bbox": [ + 92, + 62, + 492, + 258 + ], + "lines": [ + { + "bbox": [ + 92, + 62, + 492, + 258 + ], + "spans": [ + { + "bbox": [ + 92, + 62, + 492, + 258 + ], + "type": "image", + "image_path": "e6b093bec8798c3637a5d3318f19a10f7f573de96c46fbb7b9f445e8753d83c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 316 + ], + "type": "text", + "content": "Figure 1. We show our main idea Concept Correction, a framework for correcting spurious attributes in biased models using concepts, set of examples representing some pattern. We show an example from the Waterbirds dataset [17], where (a) directly training an ERM results in a model that spuriously correlates background type with bird type. A practitioner can correct this by (b) finding images with land and water backgrounds and using concept correction. Notably, these example images can be obtained outside of the original Waterbirds dataset, such as using generative AI, and using search engines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 322, + 289, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 289, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 289, + 525 + ], + "type": "text", + "content": "they contain the features representing the spurious attribute. For example, concept sets designed to counteract a spurious attribute of gender presentation may include a set of images of female-presenting people, but without any reference to the task-specific class label of doctor or nurse. Since the examples need not be selected from the training distribution, they are reusable across models and tasks, and practitioners can retrieve them from various sources such as a search engines, datasets on the Web, or image generation models, all of which require significantly less cognitive and technical effort than manually labeling in-distribution data. In fact, by using images generated with only two lines of prompts to the text-to-image generator Stable Diffusion [16], we show that an implementation of Concept Correction can achieve competitive results to state-of-the-art methods that require manual labeling of tens of thousands of data points [9, 17, 24]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "We demonstrate an instance of Concept Correction via Concept DRO (CDRO), an approach that leverages concepts to infer group labels, then trains with a distributively robust optimization objective using the inferred labels. Similar to the existing two-stage methods, CDRO first trains a neural network with ERM. Then, exploiting the fact that the ERM model learns spurious attribute labels[7], we use the model to extract features from the concept sets, the training data, and the validation data. We then train a classifier to discriminate concepts then use the weights of this classifier to determine the extent to which each training sample exhibits that concept. These concept similarity scores are then used to infer a spurious attribute label for every data point in the training and validation sets, providing the information needed by state-of-the-art methods to improve" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 322, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 380 + ], + "type": "text", + "content": "worst-group accuracy. In our example, CDRO trains a classifier on two sets of images representing the concepts of presenting-female and presenting-male, then uses the coefficients of this classifier to estimate the extent to which each training sample exhibits \"maleness\" vs. \"femaleness.\"" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "text", + "content": "We evaluate CDRO on three image classification datasets with spurious correlations: CMNIST [1], Waterbirds [17], and CelebA [10]. Among methods that do not assume spurious attribute labels are provided, CDRO improves worstgroup accuracy by up to " + }, + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "inline_equation", + "content": "12.6\\%" + }, + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "text", + "content": " on Waterbirds and " + }, + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "inline_equation", + "content": "33.1\\%" + }, + { + "bbox": [ + 304, + 382, + 546, + 501 + ], + "type": "text", + "content": " on CelebA. CDRO also achieves competitive performance to methods that do assume access to spurious attribute labels. Finally, we show that CDRO is robust to the quality of concepts in two aspects: size of concept and distribution distance from training data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 501, + 503, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 503, + 512 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 503, + 512 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 515, + 545, + 647 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 545, + 550 + ], + "type": "text", + "content": "- We propose Concept Correction, a label-efficient framework that uses out-of-distribution sets of images to correct spurious correlations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 551, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 545, + 598 + ], + "type": "text", + "content": "- We present CDRO, an instance of the Concept Correction framework that uses concepts to infer labels of spurious attributes and then trains with a distributively robust optimization objective." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 599, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 545, + 647 + ], + "type": "text", + "content": "- We show CDRO significantly outperforms ERM and GEORGE across three benchmarks while achieving competitive results to methods that require group labels during validation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 391, + 670 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "In this section, we first present the problem setting, then describe concepts and the Concept Activation Vector (CAV) introduced by Kim et al. [6]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12427" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 146, + 85 + ], + "type": "text", + "content": "2.1. Problem Setting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": "We consider a classification setting, where given an input " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ", our goal is to predict " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, \\ldots, x_{n}\\}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "Y = \\{y_{1}, \\ldots, y_{n}\\}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " be our training set of size " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ". Each datapoint " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " has a spurious attribute " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ", which is unobserved during training. Each " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " is assumed to belong to a group " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " is a pair " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "(y_{i}, a_{i})" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ", the combination of a class label " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": " and a spurious attribute label " + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 46, + 90, + 287, + 235 + ], + "type": "text", + "content": ". The spurious correlation problem occurs when accuracy across groups varies significantly; i.e., a model trained with ERM may achieve a high average accuracy, yet still performs poorly on some minority groups. We therefore evaluate classifiers on the accuracy of the worst-performing group." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 235, + 288, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 235, + 288, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 235, + 288, + 355 + ], + "type": "text", + "content": "We are interested in the setting where group information is not available for either the training set nor the validation set. Instead, we assume that users curate a set of images representing the concept without regard to any particular task. The human effort to curate the concept depends on the user's domain knowledge. For example, in Waterbirds, to curate a concept for 'landbird background,' one can generate images with the prompt 'a photo of land background'. However, if the user knows that land backgrounds of birds tend to be forests, the prompt can be adjusted accordingly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 362, + 264, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 264, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 264, + 376 + ], + "type": "text", + "content": "2.2. Concepts and Concept Activation Vectors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 441 + ], + "type": "text", + "content": "A concept is defined using a set of images. For example, a set of ocean and lake images can represent the concept water background. The concept examples need not be drawn from the training data and need not be relevant to the classification task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": "Given a concept " + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": " and a layer " + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": " of the neural network, a concept activation vector (CAV) " + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_C^l \\in R^m" + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": " [6] is defined as the vector normal to the hyperplane separating samples with a concept (the concept set " + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "inline_equation", + "content": "P_C" + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": ") and examples without a concept (the contrastive set " + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 441, + 288, + 538 + ], + "type": "text", + "content": ") in the model's activations. Given layer activations of both the concept set and the contrastive set, CAV can be obtained by training a linear classifier distinguishing the two sets of activations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 164, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 164, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 164, + 562 + ], + "type": "text", + "content": "3. Concept Correction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 665 + ], + "type": "text", + "content": "We propose Concept Correction, a framework that uses concepts to correct spurious correlations, without relying on manual labeling of in-distribution data. Instead, Concept Correction takes a set of a small number of examples as input and leverages the examples to infer spurious attribute information. Then, a robust model is trained with the inferred group of spurious attribute information using any worst-group loss minimization algorithms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "Concept Correction expects the concept examples to represent the spurious attribute. As described in the introduction, these images can be out-of-distribution and irrelevant to the prediction task, de-coupling concept curation from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "the model setting. In fact, using generative models, practitioners can easily generate many examples with just a few lines of prompting. In contrast to existing methods that require manual labeling of either training or validation samples, concept curation does not require access to the training data. Practitioners can reason about potential biases the model may exhibit then curate concept sets accordingly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 547, + 253 + ], + "type": "text", + "content": "Concept Correction is compatible with any method that requires spurious attribute or group information, such as Group DRO [17], Just-Train-Twice [9], Correct-n-Contrast [24], and more. For example, a simple way of applying Concept Correction to Correct-n-Contrast and JustTrain-Twice is simply replacing the validation group labels with inferred labels using concepts, alleviating the need of manual labeling of a large amount of data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 262, + 395, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 395, + 274 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 395, + 274 + ], + "type": "text", + "content": "3.1. Concept DRO" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 280, + 547, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 547, + 376 + ], + "type": "text", + "content": "We present Concept DRO (CDRO), an implementation of Concept Correction on Group DRO. As illustrated in Figure 2, in stage 1, we first train an ERM model and then train a linear classifier on concepts representing the spurious attribute to infer group labels. In stage 2, we use Group DRO to train a robust model with the inferred group labels. We include further details on both stages below, and give pseudocode for CDRO in the Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 394, + 544, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 394, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 394, + 544, + 407 + ], + "type": "text", + "content": "3.1.1 Stage 1: Inferring Labels of Spurious Attributes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": "We train a neural network with ERM on the training data. Given a spurious attribute " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{a_1,\\dots,a_m\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ", we curate two sets of examples for each concept " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ": a concept set " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "C = \\{c_1,\\dots,c_k\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ", which represents the presence of the concept, and a contrastive set " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "N = \\{n_1,\\dots,n_k\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ", which represents the absence of the concept. We then use the ERM model as a feature extractor, by taking the first to the penultimate layer of the model, denoted as " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ", to extract features from the two sets. We then train a linear classifier on the ERM features of the concept sets, " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{f(c_1),\\dots,f(c_k)\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{f(n_1),\\dots,f(n_k)\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ". The resulting coefficient of the linear classifier gives us a CAV pointing in the direction of the concept " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ". Then, for both training and validation data, denoted as " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "X = \\{x_1,\\dots,x_n\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": ", we use " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": " to extract representations, " + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{f(x_1),\\dots,f(x_n)\\}" + }, + { + "bbox": [ + 304, + 414, + 545, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "We then compute the cosine similarity between the representations and the CAV. We use these similarity scores to separate the data into groups — examples with the same attributes are likely to activate similar similarity scores. To separate points by similarity scores, we use a naive approach: we train a Gaussian Mixture Model (GMM) on the similarity scores, where the number of mixtures is set to " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ", the size of the domain of the spurious attribute. Typically (but not exclusively) " + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "inline_equation", + "content": "m = 2" + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": ", since many formulations of bias posit a binary distinction between minority and major-" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12428" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "ity populations. The mixture with the highest mean similarity to the concept is labeled as " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = a_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". Intuitively, data with high similarity scores with respect to the concept set tend to exhibit concept " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", and thus belong to groups with " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = a_{i}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": " (see figure 3). We repeat the above steps for all " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "a_{i} \\in \\{a_{1}, \\dots, a_{m}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". Lastly, for the data points with missing labels, we randomly assign one from " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\{a_{1}, \\dots, a_{m}\\}" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": "In the case that the spurious attribute is binary, i.e. " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{a_1,a_2\\}" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": ", we only need to curate concepts for " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": " and train one CAV. Suppose we train a CAV for " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "a_1" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": ", and label data with the CAV, yielding " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "X_{a_1} = \\{x\\in X|\\mathcal{A}(x) = a_1\\}" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": ". We can then use the negative of the CAV to label " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "a_2" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": ", which can be considered the complement, " + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "inline_equation", + "content": "X_{a_1}^C" + }, + { + "bbox": [ + 47, + 157, + 286, + 228 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": "Since Group DRO takes group labels as input, where a group label " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " is the pair " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "(y, a)" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " defined by the label " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " and spurious attribute " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": ". We generate inferred group labels " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " using ground truth labels " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " and inferred spurious attribute labels " + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "inline_equation", + "content": "\\hat{a}" + }, + { + "bbox": [ + 47, + 228, + 286, + 288 + ], + "type": "text", + "content": " to use in Stage 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 286, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 286, + 326 + ], + "type": "text", + "content": "3.1.2 Stage 2: Optimize for worst-group loss with inferred labels" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "content": "We use Group DRO to train a classifier with inferred group labels " + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "inline_equation", + "content": "\\hat{g}" + }, + { + "bbox": [ + 47, + 334, + 286, + 369 + ], + "type": "text", + "content": " from stage 1. Instead of overall loss, Group DRO optimizes for worst-group loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 370, + 286, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 370, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 47, + 370, + 286, + 417 + ], + "type": "text", + "content": "Group DRO uses ground truth group labels to select the best model checkpoint based on validation worst-group accuracy. We instead use the inferred labels to compute the surrogate worst-group accuracy at validation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 429, + 172, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 172, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 172, + 442 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 557 + ], + "type": "text", + "content": "In our experiments, we first show that CDRO substantially outperforms ERM and GEORGE, and achieves comparable worst-group accuracy to existing methods that require group labels. To explain the effectiveness of CDRO, we then demonstrate that CDRO separates data into corresponding groups with concepts, and infers labels of spurious attributes with high accuracy. Lastly, we show that CDRO is robust to the quality of concepts in two aspects: concept size and distribution distance from training data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 564, + 96, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 96, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 96, + 577 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 583, + 286, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 286, + 607 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 286, + 607 + ], + "type": "text", + "content": "We describe three image datasets, corresponding concept sets, and baseline methods used in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 622, + 116, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 116, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 116, + 633 + ], + "type": "text", + "content": "4.1.1 Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 641, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 286, + 677 + ], + "type": "text", + "content": "We briefly describe the three image classification benchmarks used in our evaluation. Full dataset and training details are included in the Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "CMNIST [1]: The task is to classify digits into one of the five classes: " + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{(0,1),(2,3),(4,5),(6,7),(8,9)\\}" + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": ". Each class is correlated with one of five colors: " + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "= \\{" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": " red, yellow, green, blue, purple\\}, respectively, each represented by its hex code (see figure 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "content": "Waterbirds [17]: The task is to classify birds into one of the 2 classes: " + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{\\text{waterbird, landbird}\\}" + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "content": ". Each class is correlated with the backgrounds: " + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{\\text{water background, land background}\\}" + }, + { + "bbox": [ + 305, + 96, + 545, + 144 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "text", + "content": "CelebA [10]: The task is to classify celebrities' hair color " + }, + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{\\text{blond, not blond}\\}" + }, + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "text", + "content": ", with the spurious attribute " + }, + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "inline_equation", + "content": "\\mathcal{A} = \\{\\text{female, male}\\}" + }, + { + "bbox": [ + 305, + 144, + 545, + 191 + ], + "type": "text", + "content": ". The class blond correlates with female, and the class not blond correlates with male." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 207, + 378, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 207, + 378, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 378, + 217 + ], + "type": "text", + "content": "4.1.2 Baselines" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 225, + 544, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 225, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 305, + 225, + 544, + 249 + ], + "type": "text", + "content": "We consider seven baselines that train models under different assumptions regarding the availability of group labels." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 250, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 545, + 285 + ], + "type": "text", + "content": "Group labels available during training - Group DRO [17] is a state-of-the-art method, which optimizes for the worst-group loss using group labels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 285, + 545, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 416 + ], + "type": "text", + "content": "Group labels available during validation - Just Train Twice (JTT) [9] trains a model to detect minority group examples, then upweights those examples during training. Correct-n-Contrast (CNC) [24] infers group information similar to JTT, then uses contrastive learning to learn representations robust to spurious correlations. Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. Spread Spurious Attribute (SSA) [14] uses semi-supervised learning to train a predictor to infer pseudo-group label, then trains a robust model with worst-group loss minimization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "type": "text", + "content": "Group labels available during neither training nor validation - Empirical Risk Minimization (ERM) is the standard training procedure that minimizes the average loss. Similar to JTT, GEORGE also follows a two-step procedure by first estimating the group labels and then using these estimates to train a robust classifier. Unlike JTT, after training a standard model, the feature space of each class is split into groups via unsupervised clustering. During validation, GEORGE then optimizes for worst-cluster accuracy without requiring group labels." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 552, + 378, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 378, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 378, + 563 + ], + "type": "text", + "content": "4.1.3 Concepts" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 570, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 570, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 570, + 545, + 617 + ], + "type": "text", + "content": "In our experiments we generated 50 images per concept set. We generated synthetic data as concepts for CMNIST, and used a popular off-the-shelf generative AI model Stable Diffusion [16] to generate concepts for Waterbirds and CelebA." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 618, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 618, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 618, + 545, + 712 + ], + "type": "text", + "content": "CMNIST - Since the 5 target classes are correlated with 5 colors, we curate concept sets for each color. For each color, we sample 50 images from the original MNIST dataset [2], which consists of white digits 0-9 on black background, and we paint the digits with the corresponding color. We then curate the contrastive set by concatenating the concept sets of the 4 remaining colors, and trained a CAV pointing towards the color." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12429" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 62, + 482, + 281 + ], + "blocks": [ + { + "bbox": [ + 106, + 62, + 482, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 62, + 482, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 62, + 482, + 281 + ], + "type": "image", + "image_path": "384035a350053ee80de40b000ea2154758c35d94ddb37fe8f960958226157ee5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "type": "text", + "content": "Figure 2. We show Concept DRO (CDRO), a method of training an unbiased model using a concept set. (1) First, we train a neural network model " + }, + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "type": "inline_equation", + "content": "(h\\circ f)" + }, + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "type": "text", + "content": " using ERM. (2) We train a concept activation vector (CAV) using the given concept set and the model [6]. A CAV is a linear decision boundary between the model representations of the concepts " + }, + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "type": "inline_equation", + "content": "(f(\\cdot))" + }, + { + "bbox": [ + 46, + 284, + 548, + 329 + ], + "type": "text", + "content": ". (3) We use the inferred CAV to infer pseudo-group labels over the training data. (4) Finally, we run Group DRO [17] the training data using the pseudo-labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 339, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 289, + 518 + ], + "type": "text", + "content": "Waterbirds - Since land background is correlated with landbirds, and water background is correlated with waterbirds, we curate sets of examples to represent the concepts \"land background\" and \"water background.\" With the knowledge that land background in the dataset are either bamboo or broadleaf trees, we used prompts \"A photo of bamboo trees\" and \"A photo of broadleaf trees\" to generate a total of 50 (25 each) images to represent the \"land background\" concept. Similarly, knowing that the water background in the dataset are either ocean or lake, we used prompts \"A photo of ocean\" and \"A photo of lake\" to generate a total of 50 (25 each) images to represent the \"water background\" concept. Since the spurious attribute is binary, we trained a CAV in the direction of \"land background,\" using the \"water background\" concept as the contrastive set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 519, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 289, + 639 + ], + "type": "text", + "content": "CelebA - Since female is correlated with blond and male is correlated with not blond, we curate sets of examples to represent the concepts female and male. With the knowledge that the dataset consists of celebrities' faces, we used prompts \"A photo of a female celebrity face\" and \"A photo of a male celebrity face\" to generate a total of 100 images (50 each) for the female and male concepts, respectively. We trained a CAV with female as the concept set and male as the contrastive set, such that positive similarity can be interpreted as exhibiting female features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 647, + 133, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 133, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 133, + 658 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Table 1 reports the average and worst-group accuracies of all approaches. CDRO outperforms ERM across all tasks. Comparing to GEORGE, CDRO achieves higher worst-group accuracy on Waterbirds and CelebA. The perfor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": "mance of GEORGE relies on how well the learned features can be clustered. In CMNIST, where the features (color and shape) are easily distinguishable, GEORGE easily finds clusters that correspond to groups and therefore yields a high worst-group accuracy. However, in Waterbirds and CelebA, where the features are more complex, clustering is less effective and the worst-group accuracy is lower. Further, with only 100 out-of-distribution images, CDRO achieves similar performance to methods that require validation labels, which may require large-scale manual labeling: CelebA, for example, has 20K validation images. Specifically, CDRO outperforms JTT by " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on Waterbirds, and " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on CelebA and approaches the performance of Group DRO itself. CDRO outperforms CNC by " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on Waterbirds, and falls short by " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on CelebA. CDRO outperforms SSA with " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " validation labels by " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on Waterbirds, and " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on CelebA. Lastly, CDRO outperforms AFR with " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " validation labels by " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on Waterbirds, and " + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 304, + 339, + 547, + 567 + ], + "type": "text", + "content": " on CelebA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 577, + 406, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 577, + 406, + 589 + ], + "spans": [ + { + "bbox": [ + 305, + 577, + 406, + 589 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 597, + 495, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 597, + 495, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 597, + 495, + 609 + ], + "type": "text", + "content": "4.3.1 CDRO effectively infers group labels" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "To examine the effectiveness of CDRO in inferring group labels, we first visualize the distribution of both train and validation data by their cosine similarity scores to the corresponding CAV for all three datasets. As shown in Figure 3, for both train and validation data of the three datasets, data representations with high similarity scores tend to exhibit the concept that the CAV directed to. For example, Figure 3 shows that, given a CAV in the direction of the concept red," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12430" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 525, + 240 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 525, + 240 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 525, + 240 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 525, + 240 + ], + "type": "table", + "html": "
MethodGroup InfoCMNIST [1]Waterbirds [17]CelebA [10]
Worst(%)Mean(%)Worst(%)Mean(%)Worst(%)Mean(%)
Group DRO [17]Train & Val74.7 (1.0)190.6 (0.1)89.9 (0.6)92.0 (0.6)88.9 (1.3)93.9 (0.1)
JTT [9]Val74.5 (2.4)90.2 (0.8)83.8 (1.2)89.3 (0.7)81.5 (1.7)88.1 (0.3)
CNC [24]Val77.4 (3.0)90.9 (0.6)88.5 (0.3)90.9 (0.1)88.8 (0.9)89.9 (0.5)
SSA [14]Val--89.0 (0.6)92.2 (0.9)89.8 (1.3)92.8 (0.1)
5% Val--87.1 (0.7)92.6 (0.2)86.7 (1.1)92.8 (0.3)
AFR [15]Val--90.4 (1.1)94.2 (1.2)82.0 (0.5)91.3 (0.3)
5% Val--86.6 (4.7)-77.5 (1.3)-
GEORGE [19]None76.4 (2.3)89.5 (0.3)76.2 (2.0)95.7 (0.5)54.9 (1.9)94.6 (0.2)
ERMNone0.0 (0.0)20.1 (0.2)62.6 (0.3)97.3 (1.0)47.7 (2.1)94.9 (0.3)
CDRO (ours)Concepts69.3 (2.7)85.0 (1.0)88.8 (0.2)90.5 (0.1)88.0 (0.7)90.8 (0.3)
", + "image_path": "c73be5748e90d4087658ff64593f485afcdc0a5ec7847f6c72d863799fb53449.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 80, + 285, + 255, + 347 + ], + "blocks": [ + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "text", + "content": "Table 1. Average and worst-group accuracies of models trained with CDRO and baselines on the benchmarks. CDRO outperforms approaches that do not use group information (GEORGE, ERM) and approaches that use " + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 248, + 547, + 283 + ], + "type": "text", + "content": " of Validation (SSA, AFR). CDRO also has comparable performance to approaches that use validation group labels (JTT, CNC, AFR, SSA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 285, + 255, + 347 + ], + "lines": [ + { + "bbox": [ + 80, + 285, + 255, + 347 + ], + "spans": [ + { + "bbox": [ + 80, + 285, + 255, + 347 + ], + "type": "table", + "html": "
Precision(%)Recall(%)
CMNIST87.096.9
Waterbirds76.383.9
CelebA78.078.5
", + "image_path": "c3228c83f99f01725907992037b17de978d8de15bae91f53973784009983457e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 288, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 355, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 288, + 389 + ], + "type": "text", + "content": "Table 2. We show the average precision and recall of the pseudogroup labels inferred by CDRO compared to the ground truth. We show CDRO effectively infers group labels." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 400, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 472 + ], + "type": "text", + "content": "data points with high similarity scores tend to be the red digits. Similarly, given a CAV in the direction of the concept land background, data points with high similarity scores tend to be images with land background. These results justify the decision to label all data points in the highest-mean mixture with the group associated with the concept." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 473, + 288, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 288, + 498 + ], + "type": "text", + "content": "Further, Table 2 shows that CDRO estimates group labels with high accuracy in all three datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 514, + 242, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 242, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 242, + 528 + ], + "type": "text", + "content": "4.3.2 CDRO is robust to quality of concepts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 641 + ], + "type": "text", + "content": "From a practical perspective, since CDRO requires practitioners to curate the concepts, the quality of concepts depends on how much time and effort the practitioners spend on curating the concepts. We therefore evaluate the effectiveness of CDRO when the quality of concepts varies on Waterbirds and CelebA. To examine this, we ask the following two questions: 1) Does CDRO require that the concept images be drawn from the training distribution? 2) Does CDRO require a large number of concept images?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 712 + ], + "type": "text", + "content": "We first fix the size of concepts to be 100 (50 for the concept set and 50 for the contrastive set), and examine the effect of distribution distance between concepts and training data on CDRO. We vary the distance between training distribution and concept distribution by varying the prompts we use for Stable Diffusion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 288, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 492 + ], + "type": "text", + "content": "We design the prompts based on how much effort a hypothetical practitioner would spend on curating the concepts. A practitioner may not have specific information about the dataset and decide to use very general prompts, potentially resulting in a low-quality concept set. For the Waterbirds dataset, we use the following prompts, varying the level of effort and domain knowledge: \"A photo of a [water/land] background\" as the \"distant\" distribution, \"A photo of a [water/land]bird habitat\" as the \"somewhat near\" distribution, and \"A photo of a [water/land] background\" as the \"near\" distribution. For the CelebA dataset, we use the following prompts: \"A photo of a [female/male] person\" as the \"distant\" distribution, \"A photo of a [female/male] celebrity\" as the \"somewhat near\" distribution, and \"A photo of a [female/male] celebrity face\" as the \"near\" distribution. We include our reasoning and a discussion regarding these prompts in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 495, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 578 + ], + "type": "text", + "content": "Table 3 shows that CDRO is effective even if the concept images are out of distribution. In both Waterbirds and CelebA, CDRO achieves similar worst group accuracy to the in-distribution set even when the concept distribution is only somewhat near. This result suggests that practitioners need not meticulously curate the concept sets for CDRO to be effective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "We then fix the concept distribution to be near training distribution, and vary the size from the list (4, 8, 40, 100, 200), splitting equally between concept set and contrastive set. For example, if size is 4, then the concept and contrastive sets each have two images. Figure 4 shows that CDRO is robust to the size of the concepts. Even when there are only four images, CDRO achieves higher worst-group accuracy than GEORGE on both Waterbirds and CelebA. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC. A small concept set is sufficient in part since the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12431" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 164, + 152 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 164, + 152 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 164, + 152 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 164, + 152 + ], + "type": "image", + "image_path": "ebdf4d58afa644bfa27870431a34425f2d0845623f803aa1c29383ef401399b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 165, + 64, + 258, + 152 + ], + "blocks": [ + { + "bbox": [ + 165, + 64, + 258, + 152 + ], + "lines": [ + { + "bbox": [ + 165, + 64, + 258, + 152 + ], + "spans": [ + { + "bbox": [ + 165, + 64, + 258, + 152 + ], + "type": "image", + "image_path": "3f268eb9a50068ec790dbf0fce63c934472a6a3afa0d70742d8ccd88996658fe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 259, + 64, + 352, + 157 + ], + "blocks": [ + { + "bbox": [ + 259, + 64, + 352, + 157 + ], + "lines": [ + { + "bbox": [ + 259, + 64, + 352, + 157 + ], + "spans": [ + { + "bbox": [ + 259, + 64, + 352, + 157 + ], + "type": "image", + "image_path": "cf3b6303c3a42ae07f70167c4baa68de9a71e298664cde0ebede13dc66bd6202.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 64, + 447, + 152 + ], + "blocks": [ + { + "bbox": [ + 354, + 64, + 447, + 152 + ], + "lines": [ + { + "bbox": [ + 354, + 64, + 447, + 152 + ], + "spans": [ + { + "bbox": [ + 354, + 64, + 447, + 152 + ], + "type": "image", + "image_path": "5fcef310329df0b558cc28da4ed3f18fb7568fa6859c4bc2920bd8f2faf7f148.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 449, + 64, + 542, + 152 + ], + "blocks": [ + { + "bbox": [ + 449, + 64, + 542, + 152 + ], + "lines": [ + { + "bbox": [ + 449, + 64, + 542, + 152 + ], + "spans": [ + { + "bbox": [ + 449, + 64, + 542, + 152 + ], + "type": "image", + "image_path": "483da73f71d471fe04ea562534360b3b33dec50af129c3bb8d365c8b6b4ee5f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 136, + 159, + 295, + 251 + ], + "blocks": [ + { + "bbox": [ + 136, + 159, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 136, + 159, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 136, + 159, + 295, + 251 + ], + "type": "image", + "image_path": "cfb52e8ec05d6454499a01e242c54cdefda5dfa67dea34c92623b88739324739.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 313 + ], + "type": "text", + "content": "Figure 3. Distributions of similarity between the CAV trained with out-of-distribution concepts and validation samples. Each peak is colored by the ground truth group label. The peaks are separable, and the peak furthest to the right (highest similarity to the concept) corresponds to the ground truth label in all cases, indicating how we can infer spurious attribute labels. Top row: CMNIST validation samples for each of five color concepts. Bottom left: Waterbird validation samples with land as the concept set and water as the contrastive set. Bottom right: CelebA validation samples with women as the concept set and men as the contrastive set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 297, + 159, + 453, + 253 + ], + "blocks": [ + { + "bbox": [ + 297, + 159, + 453, + 253 + ], + "lines": [ + { + "bbox": [ + 297, + 159, + 453, + 253 + ], + "spans": [ + { + "bbox": [ + 297, + 159, + 453, + 253 + ], + "type": "image", + "image_path": "f1fc5d0e9451bf8baa00f0134cef9b10ac18f73827e738ac22b97253d6b58904.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 328, + 264, + 464 + ], + "blocks": [ + { + "bbox": [ + 63, + 328, + 264, + 464 + ], + "lines": [ + { + "bbox": [ + 63, + 328, + 264, + 464 + ], + "spans": [ + { + "bbox": [ + 63, + 328, + 264, + 464 + ], + "type": "image", + "image_path": "0444e06a7a4bd6bd4d3b231035e164a0c1b5df75a1502717142e746f9dfeb78b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 467, + 288, + 533 + ], + "lines": [ + { + "bbox": [ + 46, + 467, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 288, + 533 + ], + "type": "text", + "content": "Figure 4. CDRO is robust to size of concepts. Even when there are only 4 images, CDRO achieves higher worst group accuracy than GEORGE on both Waterbirds and CelebA, since the learned embeddings from ERM make it easy to train a classifier. Further, with 40 concept images, CDRO achieves competitive results to methods that require validation group labels, e.g., CNC." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 55, + 536, + 280, + 609 + ], + "blocks": [ + { + "bbox": [ + 55, + 536, + 280, + 609 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 280, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 280, + 609 + ], + "type": "table", + "html": "
SourceWorst(%)
In DistributionWaterbirds [17]89.3
NearStable Diffusion [16]88.5
Somewhat NearStable Diffusion [16]80.1
DistantStable Diffusion [16]53.1
", + "image_path": "cf4ef2a02139965ef7c2fec79951745ed4b2415fcaff6f08336a50c64b03d77e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 658, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 658, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 658, + 288, + 683 + ], + "type": "text", + "content": "input to the GMM is ERM-trained embeddings, such that training data features are already represented." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 313, + 331, + 539, + 405 + ], + "blocks": [ + { + "bbox": [ + 46, + 617, + 288, + 651 + ], + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 651 + ], + "type": "text", + "content": "Table 3. The worst group accuracy given different concept set qualities on Waterbirds. Concept set quality is judged by the level of perceived effort and domain knowledge needed to create." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 331, + 539, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 539, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 539, + 405 + ], + "type": "table", + "html": "
SourceWorst(%)
In DistributionCelebA [10]87.2
NearStable Diffusion [16]87.9
Somewhat NearStable Diffusion [16]87.2
DistantStable Diffusion [16]75.6
", + "image_path": "bf211782ca2b3e7117d2e388c6c8b10c07cc0aed03d7f607f99e4336bc1377f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 413, + 547, + 447 + ], + "lines": [ + { + "bbox": [ + 305, + 413, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 547, + 447 + ], + "type": "text", + "content": "Table 4. The worst group accuracy given different concept set qualities on CelebA. The concept set qualities are judged similar to Waterbird concepts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 454, + 531, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 531, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 531, + 467 + ], + "type": "text", + "content": "4.3.3 Concept Correction is Compatible with CNC" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 475, + 545, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 643 + ], + "type": "text", + "content": "We demonstrate that Concept Correction is compatible with other robust training methods via CNC. Specifically, CNC consists of two stages: first, CNC trains a model with ERM and uses the predictions of the ERM model to identify majority and minority groups; then CNC uses contrastive learning to maximize the similarity of examples with same class label but different ERM predictions, and minimize the similarity examples with different class labels but same ERM predictions. We replace ERM predictions with inferred labels of spurious attributes, which we generate according to stage 1-3 of Figure 2. Table 5 shows that models trained with concept-based pseudo-labels achieve comparable worst-group accuracy to models trained with ground truth group labels on both train and validation set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 392, + 669 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 712 + ], + "type": "text", + "content": "We discuss related work on improving robustness to spurious correlations, both with and without group information available." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 714 + ], + "type": "text", + "content": "1Note that we ran Group DRO on CMNIST according to hyperparameters specified in [24], whereas the result reported by [24] was 78.5(4.5)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12432" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 70, + 451, + 119 + ], + "blocks": [ + { + "bbox": [ + 142, + 70, + 451, + 119 + ], + "lines": [ + { + "bbox": [ + 142, + 70, + 451, + 119 + ], + "spans": [ + { + "bbox": [ + 142, + 70, + 451, + 119 + ], + "type": "table", + "html": "
MethodGroup InfoWaterbirds Worst(%)CelebA Worst(%)
CNCTrain & Val88.7 (0.4)88.4 (0.1)
Concept CNCConcepts86.2 (0.7)87.2 (0.1)
", + "image_path": "0cb647e92f8c219be1856727a6321c6cba58f7835380e71f1a07a84fc76f5144.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 127, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 46, + 127, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 545, + 150 + ], + "type": "text", + "content": "Table 5. Worst-group accuracy of Concept CNC compared to CNC. We show that we can use concept sets to achieve comparable performance to CNC, which uses both train and validation group labels in our implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 161, + 287, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 161, + 287, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 287, + 268 + ], + "type": "text", + "content": "Improving Robustness with Group Information \nSagawa et al. proposed Group Distributionally Robust Optimization (DRO) [17], which assumes access to spurious attribute labels and optimizes for worst-group loss instead of average loss. The authors show that Group DRO can achieve state-of-the-art worst-group-accuracy. To demonstrate CDRO, we use the same distributionally robust objective of Group DRO but infer group labels instead of assuming they are provided." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 288, + 425 + ], + "type": "text", + "content": "Several proposals to improve worst group accuracy re-sample the training data to balance classes or groups [4, 5, 7, 12]. These approaches show that while a neural network trained via ERM may be biased at the classification layer, it can still learn the core features for the task at the penultimate layer. Therefore it is effective to adjust the weights of the features by fine-tuning on a balanced hold-out set [5, 7], or re-weighting the loss function for majority and minority groups [4], or processing the model outputs post-hoc [12]. However, these methods acknowledge the expense of obtaining spurious attribute labels for a large dataset. Further, multiple spurious correlations can occur in practice [25], complicating reweighting schemes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 288, + 700 + ], + "type": "text", + "content": "Improving Robustness without Group Information: Recent proposals to improve worst-group accuracy without access to group labels typically follow a two-stage paradigm: first train a model with ERM, then use the predictions of the ERM model to identify minority groups, then train another model with emphasis on the minority groups, thereby achieving high accuracy on both the majority and minority groups [5, 9, 14, 15, 19, 20, 22, 24]. For example, JTT [9] uses the incorrect predictions of the ERM classifier as a proxy for the minority group, then trains a model with an upweight on the misclassified examples. Similarly, Correct-N-Contrast (CNC) [24] uses misclassifications as pseudo spurious attribute labels, then uses contrastive learning to align representations for samples within the same class but different spurious attribute. Recognizing that retraining a model from scratch is computationally expensive, Automatic Feature Reweighting (AFR) [15] finetunes the last layer of the ERM model on a reweighted training set. However, these methods still require a validation set with group labels for hyperparameter tuning and selecting the best model checkpoint, which turns out to be crucial for achieving their results. CDRO avoids the need of validation group labels by using concepts to infer pseudo group labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "In [14] and [20], the authors assume access to a small" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 161, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 291 + ], + "type": "text", + "content": "number of group-labeled data and use semi-supervised learning to exploit the group information for better worst group accuracy. Specifically, Spread Spurious Attribute (SSA) [14] trains a spurious attribute predictor using both group-labeled and group unlabeled data, then uses the predicted labels for worst-group loss minimization. Unlike SSA, CDRO does not require concepts to represent task-specific groups, which encapsulate both the spurious attribute and target class. Instead, concepts can be out-of-distribution examples solely representing the spurious attribute." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 413 + ], + "type": "text", + "content": "GEORGE [19] requires group labels neither during training nor validation, instead inferring group labels by clustering the feature space learned by the ERM model, then running Group DRO. During validation, GEORGE uses worst-cluster accuracy to select the best hyperparameters and model checkpoint, which avoids the need of group labels. However, while GEORGE achieves high worst group accuracy on synthetic datasets where clusters are well-defined, it is less effective on real-world datasets such as CelebA where clusters are harder to distinguish [19]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 521 + ], + "type": "text", + "content": "Some approaches do not follow the two-stage paradigm. Learning from Failure (LfF) [13] trains two models simultaneously, one intended to be biased and one intended to be unbiased. When training the unbiased model, LfF focuses on the examples on which the biased model tends to make mistakes thereby learning to ignore the sources of bias. Further, Taghanaki et al. introduced CIM [21], a contrastive learning approach that learns input-space transformation of the data to preserve task-relevant information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 378, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 654 + ], + "type": "text", + "content": "In this work, we present Concept Correction, a framework that uses out-of-distribution concepts to improve worst-group accuracy in the presence of spurious correlations. We demonstrate an instance of the framework via Concept DRO, which uses concepts to infer group labels, and then uses these to train with worst-group loss minimization. We show that Concept DRO achieves competitive results to existing methods that assume access to group labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 668, + 416, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 416, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 416, + 681 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by NIST Grant 60NANB23D194, NSF Award 1934464, Cisco Grant 70618863." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12433" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Li Deng. The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine, 29(6):141-142, 2012. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 203 + ], + "type": "text", + "content": "[3] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 249 + ], + "type": "text", + "content": "[4] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "text", + "content": "[5] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. Advances in Neural Information Processing Systems, 35:38516-38532, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "text", + "content": "[6] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 395 + ], + "type": "text", + "content": "[7] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. arXiv preprint arXiv:2204.02937, 2022. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 441 + ], + "type": "text", + "content": "[8] Anthony Liu, Santiago Guerra, Isaac Fung, Gabriel Matute, Ece Kamar, and Walter Lasecki. Towards hybrid human-air workflows for unknown unknown detection. In Proceedings of The Web Conference 2020, pages 2432-2442, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 287, + 507 + ], + "type": "text", + "content": "[9] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 553 + ], + "type": "text", + "content": "[10] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 2, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 597 + ], + "type": "text", + "content": "[11] R Thomas McCoy, Ellie Pavlick, and Tal Linzen. Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference. arXiv preprint arXiv:1902.01007, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 643 + ], + "type": "text", + "content": "[12] Aditya Krishna Menon, Ankit Singh Rawat, and Sanjiv Kumar. Overparameterisation and worst-case generalisation: friend or foe? In International Conference on Learning Representations, 2020. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: De-biasing classifier from biased classifier. Advances in Neural Information Processing Systems, 33:20673-20684, 2020. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accu" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 642 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 95 + ], + "type": "text", + "content": "racy with spurious attribute estimation. arXiv preprint arXiv:2204.02070, 2022. 4, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Shikai Qiu, Andres Potapczynski, Pavel Izmailov, and Andrew Gordon Wilson. Simple and fast group robustness by automatic feature reweighting. arXiv preprint arXiv:2306.11074, 2023. 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[16] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 2, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 228 + ], + "type": "text", + "content": "[17] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 283 + ], + "type": "text", + "content": "[18] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 341 + ], + "type": "text", + "content": "[19] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 386 + ], + "type": "text", + "content": "[20] Nimit S Sohoni, Maziar Sanjabi, Nicolas Ballas, Aditya Grover, Shaoliang Nie, Hamed Firooz, and Christopher Ré. Barack: Partially supervised group robustness with guarantees. arXiv preprint arXiv:2201.00072, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 430 + ], + "type": "text", + "content": "[21] Saeid A Taghanaki, Kristy Choi, Amir Hosein Khasahmadi, and Anirudh Goyal. Robust representation learning via perceptual similarity metrics. In International Conference on Machine Learning, pages 10043-10053. PMLR, 2021. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 485 + ], + "type": "text", + "content": "[22] Haotian Ye, James Zou, and Linjun Zhang. Freeze then train: Towards provable representation learning under spurious correlations and feature noise. In International Conference on Artificial Intelligence and Statistics, pages 8968-8990. PMLR, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 487, + 545, + 542 + ], + "type": "text", + "content": "[23] John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study. PLoS medicine, 15(11):e1002683, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 544, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 597 + ], + "type": "text", + "content": "[24] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. arXiv preprint arXiv:2203.01517, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 600, + 545, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 600, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 308, + 600, + 545, + 642 + ], + "type": "text", + "content": "[25] Dora Zhao, Jerone Andrews, and Alice Xiang. Men also do laundry: Multi-attribute bias amplification. In International Conference on Machine Learning, pages 42000–42017. PMLR, 2023. 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12434" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_content_list.json b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f745bfe11b72cdbfff9f1539f912cc30847a70f0 --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_content_list.json @@ -0,0 +1,1531 @@ +[ + { + "type": "text", + "text": "Lane2Seq: Towards Unified Lane Detection via Sequence Generation", + "text_level": 1, + "bbox": [ + 135, + 130, + 833, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kunyang Zhou Southeast University", + "text_level": 1, + "bbox": [ + 398, + 181, + 570, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kunyangzhou@seu.edu.cn", + "bbox": [ + 385, + 219, + 583, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorporate the task-specific knowledge into Lane2Seq. Experimental results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets $97.95\\%$ and $97.42\\%$ F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks.", + "bbox": [ + 75, + 301, + 473, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 587, + 209, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lane detection is a fundamental task in computer vision [31, 36, 40]. It aims to predict the location of the lane in a given image. Lane detection plays a crucial role in many applications, such as adaptive cruise control and lane keeping. Existing lane detection methods generally adopt a divide-and-conquer strategy, which decomposes the lane detection into multiple subtasks. Each subtask is accomplished by a task-specific head network. For instance, as shown in Fig. 1, segmentation-based methods [28,49] adopt a head network along with a task-specific module, such as the message-passing module [28], to predict per-pixel masks. Anchor-based methods [17, 36] utilize a classification head network for distinguishing lane instances and an anchor refinement network for regressing accurate lanes. Parameter-based methods [13, 23] use a network to predict the parameters of lanes and a vertical offset prediction network to locate the start points of lanes.", + "bbox": [ + 75, + 613, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although divide-and-conquer strategy has been proved an effectively way to address the certain subtask in the ex", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4a37222f67df15e4bb504ee6e2aa97b51868270bd642c8465112acf001dbb588.jpg", + "image_caption": [ + "(a) Segmentation-based lane detection" + ], + "image_footnote": [], + "bbox": [ + 504, + 266, + 890, + 303 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d620d4f44898b74f691c1aeedaee1f466a7f2a97d8c4d7fe101ab9f1459b206.jpg", + "image_caption": [ + "(b) Anchor-based lane detection" + ], + "image_footnote": [], + "bbox": [ + 504, + 324, + 890, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b023bb946345dd46ef4f22d65b7c07a531cbbe7c794f93a21fd611efb8d8244e.jpg", + "image_caption": [ + "(c) Parameter-based lane detection" + ], + "image_footnote": [], + "bbox": [ + 504, + 402, + 890, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66d1487b06eae28eceb22e021472fd336771ba70435b7eef1f4c5b5c0c3b1b40.jpg", + "image_caption": [ + "(d) Our sequence generation-based lane detection", + "Figure 1. Comparison of different lane detection frameworks." + ], + "image_footnote": [], + "bbox": [ + 501, + 484, + 890, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "isting methods, there still exist several limitations. (1) Each subtask needs a customized task-specific head network, resulting in a complicated lane detection model. (2) Each task-specific head requires one or more loss functions, e.g., cross-entropy loss and Line-IOU loss [17], bringing out extra hyper-parameters.", + "bbox": [ + 496, + 625, + 890, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a novel sequence generation-based lane detector (Lane2Seq) framework to tackle the aforementioned issues. By formulating the lane detection as a sequence generation task, Lane2Seq gets rid of customized head networks and task-specific loss function. It is based on the intuition that if the detection model knows where the target lane is, model can be simply taught how to read the location of lane out, instead of designing additional classification head or regression head by the divide-and-conquer strategy.", + "bbox": [ + 496, + 717, + 892, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, we convert the outputs of different lane detection formats, i.e., segmentation-based, anchor-based, and", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16944", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "parameter-based, into a sequence of discrete tokens and the model learns to generate this sequence token-by-token. As shown in Fig. 1 (d), to achieve a specific lane detection format, Lane2Seq uses a prompt to specify the detection format and the generated sequence adapts to the prompt so the model can produce format-specific output. By the format-specific prompt, Lane2Seq unifies different lane detection formats into a model.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While Lane2Seq does not contain task-specific components, task-specific knowledge contained in these components can help the model learn the features of lanes better. We propose a Multi-Format model tuning method based on Reinforcement Learning (MFRL) to incorporate the task-specific knowledge into the model without changing model's architecture. Inspired by Task-Reward [30], MFRL takes the evaluation metrics, which naturally integrates task-specific knowledge, as the reward and tunes Lane2Seq using REINFORCE [41] algorithm. However, evaluation metrics like F1 score cannot be used as the reward directly due to undecomposable as a sum of perexample rewards. In this paper, we propose three new evaluation metric-based rewards for segmentation, anchor, and parameter format, based on their task-specific knowledge.", + "bbox": [ + 75, + 213, + 472, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate that our Lane2Seq achieves competitive performance on three public datasets, Tusimple, CULane, and LLAMAS. For instance, Lane2Seq with ViT-Base encoder obtains $97.95\\%$ and $97.42\\%$ F1 score on Tusimple and LLAMAS, setting a new state-of-the-art result for two datasets. It should be noted that all existing lane detection methods heavily rely on well-designed task-specific head networks and the corresponding complicated loss functions. Instead, our Lane2Seq utilizes a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss.", + "bbox": [ + 75, + 441, + 472, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows.", + "bbox": [ + 96, + 609, + 437, + 625 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a sequence generation-based method for lane detection, which casts the lane detection as a sequence generation task. To the best of our knowledge, we are the first to unify the lane detection through the sequence generation, which offers a new perspective on lane detection.", + "- We present a novel reinforcement learning-based multi-format model tuning, including three new evaluation metric-based reward functions, to incorporate the task-specific knowledge into the model.", + "- Experimental results show that our method achieves competitive performance on lane detection benchmarks. Remarkably, we establish a new state-of-the-art result on Tusimple and LLAMAS." + ], + "bbox": [ + 94, + 638, + 468, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lane detection. Existing lane detection methods can be divided into three categories based on the representation of the lane: segmentation-based method, anchor-based method and parameter-based method. Segmentation-based methods [28, 47, 49] consider lane detection as a semantic segmentation task and performs pixel-wise prediction. SCNN [28] enhances the visual evidence by a message-passing module, which can capture spatial dependency for lanes. Anchor-based methods [20, 31, 36, 51] predict the accurate lanes by refining the predefined lane anchors. UFLD [31] proposes a novel row anchor-based approach to detect lanes. Different from the segmentation-based and anchor-based methods, parameter-based methods [13, 23, 37] regard the lane detection as the parametric modeling and regress the parameters of the lane. PolyLaneNet [37] models a lane curve as a polynomial and regresses the parameters of the polynomial. In this paper, we treat the segmentation-based, anchor-based methods, and parameter-based metohods uniformly as the sequence generation task, getting rid of the complicate structure and task-specific modules, e.g., head network. It only adopts the cross-entropy loss and plain transformer architecture.", + "bbox": [ + 496, + 114, + 890, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sequence generation for vision tasks. Recently, sequence-to-sequence (seq2seq) method used in the natural language processing (NLP) has been applied for vision tasks. Seq2seq utilizes a basic transformer encoder and decoder architecture and accomplishes the task by making sequence predictions, rather than designing a model tailored to the vision task. Pix2seq [8] is the pioneering work to cast the object detection as the sequence generation task. It shows that object can be detected well without any task-specific modules like label assignment. Besides object detection, seq2seq has been extended to other vision tasks such as instance segmentation [9], keypoint detection [9], text spotting [19,29] and object tracking [10]. UniTAB [45] adopts the task prompt [33] to perform multi-task learning. Unified-IO [25] jointly trains various vision tasks by setting the unified input/output formats for all tasks. Besides, seq2seq becomes increasingly popular in the multimodal model. Text-to-image models like DALL-E [34] and vision-language models like Flamingo [2] all use seq2seq to unify the multi-modal tasks. Nevertheless, how to perform the seq2seq in the lane detection to unify different detection formats is still unexplored.", + "bbox": [ + 496, + 448, + 892, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lane2Seq has a similar spirit with the Pix2seq and its successors [9, 10, 19, 29]. All of them view the vision tasks as the sequence generation. The main difference between these methods and Lane2Seq is the sequence format. Previous works use coordinates and category to construct the sequence, while Lane2Seq also adopts the parameter of the lane to construct the sequence.", + "bbox": [ + 496, + 779, + 890, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reinforcement learning in the computer vision. Many", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63e1195f0cba21dbe15638dac9d121f95f6aaf9a247c799564b18d7078d670c5.jpg", + "image_caption": [ + "Figure 2. Inference pipeline of Lane2Seq. The model perceives input image, prompt and generates format-specific tokens, which can be detokenized into required detection format for visualization." + ], + "image_footnote": [], + "bbox": [ + 135, + 85, + 836, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "previously researches introduce the reinforcement learning for the vision tasks, such as object detection [27], object tracking [26], image segmentation [21] and lane detection [48]. They generally focus on learning different parts of an image and refine the outputs iteratively. DQLL [48] localizes the lanes as a group of lanemarks and refine the location of lanemarks with the deep Q-learning. Recently, Task Reward [30] adopts a novel task risk-based reinforcement learning to tune the vision model without changing the model architecture. Reward functions in Task-Reward are designed for object detection, instance segmentation, colorization, and image captioning but not for lane detection. Our MFRL is the first attempt to design effective evaluation metric-based reward functions for different lane detection formats.", + "bbox": [ + 75, + 426, + 472, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 670, + 169, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the proposed sequence generation-based lane detection method, named Lane2Seq, in detail. The overall pipeline of Lane2Seq is illustrated in Fig. 2. Sec 3.1 depicts the unified interface for lane detection, Sec 3.2 details the model architecture and objective function, and Sec 3.3 introduces the proposed multi-format model tuning based on reinforcement learning.", + "bbox": [ + 75, + 696, + 468, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unified Interface for Lane Detection", + "text_level": 1, + "bbox": [ + 76, + 815, + 393, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As presented in Figure 1, lane detection formats in the existing method are diverse and have been formulated quite differently. Considering differences in the form of outputs, customized models with specialized head networks and loss", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "functions are designed for different lane detection formats.", + "bbox": [ + 498, + 426, + 888, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To integrate different lane detection formats into a model, we propose an unified sequence interface for lane detection, where both format transcription like segmentation and outputs are treated as sequences of discrete tokens. As shown in Fig. 2, the generated sequence is composed of four parts, a starting token $\\langle \\text{starting} \\rangle$ , a format transcription token like $\\langle \\text{Segmentation} \\rangle$ , a format-specific sequence, and an ending token $\\langle \\text{end} \\rangle$ . The format-specific sequence for three detection formats can be constructed as follows.", + "bbox": [ + 496, + 443, + 892, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segmentation sequence. Instead of performing pixelwise mask prediction, we predict the polygon [5] corresponding to the mask as a sequence of coordinates conditioned on a given lane instance. Then, we convert the polygon into a sequence by quantizing the coordinates of its points into discrete tokens. Specifically, $x, y$ coordinates of a point are normalized to the width and height of the image, and then quantized to $[1, n_{\\text{bins}}]$ , where $n_{\\text{bins}}$ is the size of the vocabulary. Vocabulary will be described subsequently. A polygon sequence can be expressed as $[x_1, y_1, x_2, y_2, \\ldots, x_{28}, y_{28}, <\\text{Lane}>]$ , where $<\\text{Lane}>$ is the category token. If there are multiple lanes in an image, we concatenate the all polygon sequences. The segmentation sequence consists of a starting point and all polygon sequences, where starting point is the left-top point of the image $x_0 = 0, y_0 = 0$ .", + "bbox": [ + 496, + 597, + 893, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anchor sequence. Since the essence of anchor-based methods is regressing the location of keypoints of lane anchors, we treat the anchor prediction as the keypoint sequence generation. Specifically, the", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16946", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e09fe164266f283dbbe79453cf1e3756568de1f8c1d134a19d0d6fffaa811609.jpg", + "image_caption": [ + "Figure 3. Training pipeline of Lane2Seq. The input sequence can be the segmentation sequence or anchor sequence or parameter sequence." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 460, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "keypoint sequence of a lane can be expressed as \\([x_1, y_1, x_2, y_2, \\ldots, x_{14}, y_{14}, <\\text{Lane}>\\]\\. The normalization and quantization of keypoint coordinates are same to that of polygon point. The anchor sequence contains a starting point and all keypoint sequences, where starting point is still the left-top point of the image.", + "bbox": [ + 75, + 262, + 468, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Parameter sequence. the sequence of the lane includes two parts: the parameter of polynomial function and vertical offsets. We set the polynomial degree as 5th. Specifically, the parameter sequence of a lane can be represented as $[a_1,a_2,a_3,a_4,a_5,s,\\langle \\mathrm{Lane}\\rangle ]$ , where $s$ is the vertical offset. $s$ is normalized to the height of the image and quantized to $[1,n_{\\mathrm{bins}}]$ . We use sigmoid to normalize $a_i$ by $a_{i} = \\operatorname {Sigmoid}(a_{i})$ . Then $a_{i}$ is quantized to $[1,n_{\\mathrm{bins}}]$ . Parameter sequence does not require the starting point.", + "bbox": [ + 75, + 353, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vocabulary. We use a shared vocabulary for all formats and each integer between $[1, n_{bins}]$ can be regarded as a word in the vocabulary. Each word in the vocabulary corresponds to a learnable embedding.", + "bbox": [ + 75, + 489, + 468, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Unified Architecture and Objective Function", + "text_level": 1, + "bbox": [ + 76, + 559, + 455, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow [10] and use a transformer-based encoder-decoder architecture to deal with image input and sequence output flexibly. As shown in Fig. 3, the image encoder takes the pixels as the input and outputs the corresponding image representations. We utilize the Vision Transformer (ViT) [11] to instantiate the image encoder. We adopt a transformer-based sequence decoder, which is widely used in the language modeling [32, 33], to generate the output sequence. The decoder generates one token at a time based on the preceding tokens and the image representations. Sequence decoder removes the well-designed task-specific heads for different detection formats.", + "bbox": [ + 75, + 583, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Pipeline. We first tokenize annotation of each format into the corresponding sequence. Then, we construct a training batch using images and sequences from all detection formats. Finally, we compute loss between generated sequence and target sequence for each format. For a certain detection format, the input sequence is $[<\\text{starting}>$ , format transcription token, format-specific sequence] and the target sequence is [format transcription token, format-specific sequence, $\\langle \\mathrm{end}\\rangle$ ]. For example, in the anchor format, the in", + "bbox": [ + 75, + 763, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "put sequence is $[<\\text{starting}>, <\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]$ and the target sequence is $[<\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]$ . The starting token $<\\text{starting}>$ and ending token $<\\text{end}>$ are learnable embeddings, which tell the decoder when to begin and end the sequence generation. The sequence decoder perceives the image representations and input sequence and reconstructs the target sequence.", + "bbox": [ + 496, + 92, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objective Function. Similar to Pix2seq [8], we train the Lane2Seq with a simple cross-entropy loss. At each time stamp $j$ , Lane2Seq aims to maximize the likelihood of the target tokens conditioned on the image representations $I$ and previously generated tokens $y_{1:j-1}$ ,", + "bbox": [ + 496, + 210, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {o b j} = - \\sum_ {j = 1} ^ {N} w _ {j} \\log P (\\hat {m} _ {j} | I, m _ {1: j - 1}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 292, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ and $\\hat{m}$ denote the input and target sequence, respectively. $N$ is the length of sequence. $w_{j}$ represents the weight for the $j$ -th token, 0 is assigned to the format transcription token and 1 is assigned to other tokens to ensure the model is trained to predict the desired tokens but not the format transcription tokens.", + "bbox": [ + 496, + 339, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inference Pipeline. We take the format transcription to-ken as the prompt to achieve the format-specific detection and the sequence decoder generates the rest of the sequence conditioned on the prompt and image representation. Once the whole sequence is generated, we de-quantize the sequence to obtain the location of lane. More details about de-quantization for each format can be found in the supplementary materials. Inference process is presented in Fig. 2.", + "bbox": [ + 496, + 430, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Format Model Tuning Based on Reinforcement Learning", + "text_level": 1, + "bbox": [ + 498, + 560, + 890, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lane2Seq get rids of the complicated task-specific components, such as task-specific head networks, via the sequence generation. However, this task-agnostic architecture inevitably lacks the task-specific knowledge of the lane detection, making the model ineffective in learning the features of lane. Existing lane detection methods [17, 49] usually design the specific modules to incorporate the task-specific knowledge into the model. But this way makes the model architecture complicate. We propose a novel Multi-Format model tuning method based on Reinforcement Learning (MFRL) to learn task-specific knowledge effectively without changing any components of the model.", + "bbox": [ + 496, + 598, + 890, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by Task-Reward [30], MFRL takes the evaluation metric, which naturally contains the task-specific knowledge, as the reward function and adopts the reinforcement learning to tune the model, which is common practice in the modern language model [4, 12]. Specifically, it is difficult for the model to converge if model is trained with reinforcement learning method from scratch [30]. Hence, MFRL consists of two stages: the pretraining stage and", + "bbox": [ + 496, + 780, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the model tuning stage. In the pretraining stage, model is trained on the lane detection dataset with objective function in Eq. 1 to have a good weight initialization.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the model tuning stage, we use the REINFORCE algorithm [35] to maximize the objective function as below,", + "bbox": [ + 76, + 136, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO b j = E _ {c \\sim D} \\left[ E _ {t \\sim Q (\\cdot | c, u)} R (t, g) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 181, + 468, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c, t,$ and $g$ represent the input image, generated format-specific sequences, and ground truths, respectively. $E$ and $u$ denote the mathematic expectation and the model parameter. $R, D,$ and $Q$ stand for the reward function, data distribution of the dataset, and conditional distribution parameterized by $u$ . REINFORCE algorithm estimates the gradient of the reward function by", + "bbox": [ + 75, + 199, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {u} E _ {t \\sim Q} [ R (t, g) ] = E _ {t \\sim Q} [ R (t, g) \\nabla_ {u} \\log Q (t | c, u) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 316, + 468, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, Eq. 3 is computed as an average of per-sequence gradient and reward function of per-sequence is not required to be differentiable. More details about REINFORCE algorithm can be found in the supplementary materials.", + "bbox": [ + 75, + 344, + 468, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward Function. Evaluation metric F1 score and accuracy are unable to adopt as the reward directly, because they cannot be decomposable as a sum of per-sequence rewards (see Sec 4.2). However, F1 score or accuracy is composed of false positives $(FP)$ , true positives $(TP)$ , and false negatives $(FN)$ . All three indicators can be computed per-sequence. We opt to use $TP$ and $FP$ to design the reward. Specifically, reward for three detection formats are constructed as follows.", + "bbox": [ + 75, + 405, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward for the segmentation format (termed as $R_{seg}$ ). Segmentation-based lane detection includes two task-specific knowledge, i.e., knowledge of the segmentation and lane detection. For the knowledge of lane detection, we adopt a matched Line-IOU (LIOU) [50], which can demonstrate the quality of location prediction and shape prediction of the lane. For the knowledge of the segmentation, we adopt the matched Mean Intersection over Union (mIOU), which is a widely used metric in the semantic segmentation to evaluate the quality of segmentation, as the reward. Finally, we add both of them to compute $r_{seg}$ as below,", + "bbox": [ + 75, + 541, + 468, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {s e g} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + m I O U \\left(p _ {k}, g\\right) \\right] \\tag {4} \\\\ - \\lambda_ {1} F P _ {s e g} (t, g), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 717, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ and $p_k$ represents the number of true positives in $t$ and the $k$ -th true positive. $FP_{seg}$ is the false positive rate of the segmentation format and $\\lambda_1$ is the weight to control the affect of $FP_{seg}$ . We introduce $FP_{seg}$ to penalize for false positives.", + "bbox": [ + 75, + 779, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward for the anchor format (termed as $R_{a}$ ). Anchors-based lane detection contains the knowledge of the keypoint location and lane detection. For the knowledge of keypoint", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "location, we simply adopt the matched Euclidean distance $d(p_k, g)$ between true positives and corresponding ground truths as the reward. Since MFRL needs to maximize the reward, we rescale the reward as $d_r(p_k, g) = 1 - \\frac{d(p_k, g)}{H}$ , where $H$ is the height of image. For the knowledge of lane detection, we also utilize the matched LIOU as the reward. The formulation of $r_a$ is:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {a} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + d _ {r} \\left(p _ {k}, g\\right) \\right] - \\lambda_ {2} F P _ {a} (t, g), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 204, + 890, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $FP_{a}$ and $\\lambda_{2}$ stand for the false positives rate of the anchor format and the weight of $FP_{a}$ .", + "bbox": [ + 496, + 253, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward for the parameter format (termed as $R_p$ ). Parameter-based detection only has the knowledge of lane detection, hence we directly use the matched LIOU as the reward.", + "bbox": [ + 496, + 284, + 890, + 342 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {p} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} L I O U \\left(p _ {k}, g\\right) - \\lambda_ {3} F P _ {p} (t, g), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 345, + 890, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $FP_{p}$ and $\\lambda_{3}$ stand for the false positives rate of the parameter format and the weight of $FP_{p}$ .", + "bbox": [ + 496, + 383, + 890, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since different detection formats have different contributions, we weight the objective function of different formats. The final objective function of MFRL can be computed as,", + "bbox": [ + 496, + 412, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO b j _ {t o t a l} = \\lambda_ {4} O b j _ {s e g} + \\lambda_ {5} O b j _ {a} + \\lambda_ {6} O b j _ {p}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 467, + 890, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Obj_{seg}$ , $Obj_a$ , and $Obj_p$ denote the objective function of segmentation, anchor, and parameter format, respectively. The objective function of three formats is the same as Eq. 2, except the format-specific sequence and reward function. $\\lambda_4$ , $\\lambda_5$ , and $\\lambda_6$ are the scale factor.", + "bbox": [ + 496, + 489, + 890, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 579, + 632, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 604, + 601, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments on three lane detection benchmarks: CULane [28], Tusimple [1], and LLAMAS [3].", + "bbox": [ + 496, + 627, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CULane is a widely used large-scale dataset for lane detection. It contains a lot of challenging scenarios such as crowded roads. The CULane dataset consists of 88.9K images for training, 9.7K images in the validation set, and 34.7K images for the test. Image size is $1640 \\times 590$ .", + "bbox": [ + 496, + 657, + 890, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tusimple is a real highway dataset consisting of 3,626 training images and 2,782 testing images. All images have $1280 \\times 720$ pixels.", + "bbox": [ + 496, + 732, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLAMAS is a recently released large-scale lane detection dataset with over 100K images. All lane markers are annotated with high-accurate maps. Image size is $1280 \\times 717$ .", + "bbox": [ + 496, + 777, + 890, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 847, + 684, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For CULane and LLAMAS dataset, we adopt the F1 score to measure the performance: $F_{1} =$", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0a3e3b1826030828e48d7cd598e27c26c6896ff44e18e04f685781ffddf0cefe.jpg", + "table_caption": [ + "Table 1. Comparison of F1 score and MACs (multiply-accumulate operations) on CULane testing set. We only report the false positives for \"Cross\" category following [50]. FPS is tested on a single 2080Ti." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderFPS↑Normal↑Crowded↑Dazzle↑Shadow↑No line↑Arrow↑Curve↑Night↑Cross↓Total↑MACs(G)↓
Segmentation based
SCNN [28]ResNet501290.6069.7058.5066.9043.4084.1064.4066.10190071.608.4
RESA [49]ResNet5042.192.1073.1069.2072.8047.7088.3070.3069.90150375.37.6
AtrousFormer [44]ResNet34-92.8375.9669.4877.8650.1588.6671.1473.74105478.0814.22
LaneAF [44]DLA34 [46]-91.8075.6171.7879.1251.3886.8872.7073.03136077.4112.73
Lane2Seq (segmentation)ViT-Base1993.3977.2773.4579.6953.9190.5373.3774.96112979.6415.19
Anchor based
LaneATT [36]ResNet1222391.7476.1669.4776.3150.4686.2964.0570.81126477.0270.5
O2SFformer [51]ResNet504493.0976.5772.2576.5652.8089.5069.6073.85311877.8327.51
UFLD [31]ResNet3419290.7070.2059.5069.3044.4085.7069.5066.70203772.303.3
CondLaneNet [22]ResNet3413593.3877.1471.1779.9351.8589.8973.8873.92138778.7419.6
ADNet [43]ResNet34-92.9077.4571.7179.1152.8989.9070.6474.78149978.9418.6
CLRNet [50]ResNet3411293.4978.0674.5779.9254.0190.5972.7775.02121679.7317.3
Lane2Seq (anchor)ViT-Base2393.1177.4373.2579.4653.7490.0272.4475.12117379.2715.19
Parameter based
LSTR [23]ResNet18----------64.005.9
BezierLaneNet [14]ResNet18-90.2271.5562.4970.9145.3084.0958.9868.7099673.674.7
BSNet [6]ResNet34-93.7578.0176.6579.5554.6990.7273.9975.28144579.8917.2
Eigenlanes [18]ResNet50-91.7076.0069.8074.1052.2087.7062.9071.80150977.2018.7
Laneformer [15]ResNet50-91.7775.7470.1775.7548.7387.6566.3371.041977.0626.2
Lane2Seq (parameter)ViT-Base3093.0376.4272.1778.3252.8989.6772.6773.98131978.3915.19
", + "bbox": [ + 80, + 127, + 893, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\frac{2 \\times \\text{Precision} \\times \\text{Recall}}{\\text{Precision} + \\text{Recall}}$ , where Precision = TP/TP+FP and Recall = TP/TP+FN.", + "bbox": [ + 76, + 393, + 455, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Tusimple dataset, we use F1 score, accuracy, false positives, and false negatives to evaluate the model performance. Accuracy is defined as Accuracy = $\\frac{\\sum_{clip} C_{clip}}{\\sum_{clip} S_{clip}}$ , where $C_{clip}$ represents the number of accurately predicted lane points and $S_{clip}$ denotes the total number of lane points of a clip. A lane point is considered as correct if its distance is smaller than the given threshold $t_{pc} = \\frac{20}{\\cos(a_{yl})}$ , where $a_{yl}$ denotes the angle of the corresponding ground truth.", + "bbox": [ + 75, + 431, + 470, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 571, + 294, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model. We adopt the ViT-Base initialized with the MAE [16] pretrained parameters as the image encoder. The patch size is $16 \\times 16$ . The decoder is composed of 2 transformer blocks and its hidden size is 256. The number of attention heads is 8 and the hidden size of feed-forward network is 1024. $n_{bins}$ is set to 1000 and the dimension of word embedding of the vocabulary is 256. Our Lane2Seq is implemented based on MMDetection toolbox [7]. $\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4, \\lambda_5$ , and $\\lambda_6$ are set to 0.3, 0.3, 0.1, 0.2, 1, and 1.5, respectively.", + "bbox": [ + 75, + 597, + 468, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. All input images are resized to $320 \\times 800$ . We utilize the AdamW [24] as the optimizer with initial learning rate 1e-4. The training epochs of the pretraining stage are 5, 20, and 15 for CULane, Tusimple, and LLAMAS. We set the training epochs of the model tuning stage to 15, 30, and 55 for CULane, Tusimple, and LLAMAS, respectively. For the data augmentation, we use the random horizontal flips and random affine transformation including scaling, rotation, and translation. All experiments are conducted on 8 A100 GPUs with total batch size 384.", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b82bf5924c6687a491d276b6bde023c7141a3158a60c51645edb43dc91877f7d.jpg", + "table_caption": [ + "Table 2. Comparison of the performance of different models on LLAMAS." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderF1(%)↑
PolyLaneNet [37]EfficientNet-b0 [38]90.20
BezierLaneNetResNet3496.11
LaneATTResNet3494.96
LaneATTResNet12295.17
LaneAFDLA3496.90
CLRNetResNet1896.96
CLRNetDLA3497.16
Lane2Seq (segmentation)ViT-Base97.42
Lane2Seq (anchor)ViT-Base97.05
Lane2Seq (parameter)ViT-Base96.74
", + "bbox": [ + 503, + 433, + 890, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with the State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 498, + 585, + 890, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on LLAMAS. Our Lane2Seq achieves a new state-of-the-art performance on LLAMAS dataset using segmentation format. In Table 2, Lane2Seq increases F1 score from $97.16\\%$ to $97.42\\%$ compare to the previous state-of-the-art method CLRNet. Compared with parameter-based state-of-the-art method BezierLaneNet, Lane2Seq performs better than BezierLaneNet (96.74% vs. $96.11\\%$ ). The results manifest the detection strengths of Lane2Seq in the multi-lane scenario (The number of lanes $\\geq 5$ ), such as highways. The reason may be that transformer-based architecture has the advantages in capturing long-range dependency.", + "bbox": [ + 496, + 611, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on Tusimple. Table 3 presents the performance comparison with state-of-the-art approaches on Tusimple. Due to small data scale and single scene, the performance gap between different models is small. Our Lane2Seq using segmentation format sets a new state-of-the-art F1 score of $97.95\\%$ . The results also validate the detection strengths of Lane2Seq in multi-lane scenario.", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16949", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09482a1866244daadd92eef0b271f5d628c88ac8d5238f3374acc19fa4d05867.jpg", + "table_caption": [ + "Table 3. Comparison of the performance of different models on Tusimple. Acc denotes accuracy." + ], + "table_footnote": [], + "table_body": "
MethodsEncoderF1(%)↑Acc(%)↑FP(%)↓FN(%)↓
SCNNVGG1695.9796.536.171.80
RESAResNet3496.9396.823.632.48
PolyLaneNetEfficientNet-bo90.6293.369.429.33
UFLDResNet3488.0295.8618.913.75
LaneATTResNet12296.0696.105.642.17
GANet [39]ResNet3497.6895.871.992.64
CondLaneNetResNet3496.9895.372.203.82
CondLaneNetResNet10197.2496.542.013.50
CLRNetResNet3497.8296.872.272.08
CLRNetResNet10197.6296.832.372.38
Lane2Seq (segmentation)ViT-Base97.9596.852.012.03
Lane2Seq (anchor)ViT-Base97.8696.722.212.05
Lane2Seq (parameter)ViT-Base96.5996.002.233.54
", + "bbox": [ + 80, + 127, + 468, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on CULane. We compare Lane2Seq with other state-of-the-art methods on CULane and results are shown in Table 1. For the segmentation-based methods, Lane2Seq improves F1 score from $78.08\\%$ to $79.64\\%$ compared to the previous state-of-the-art methods AtrousFormer. Compared with anchor-based methods, Lane2Seq outperforms most previous methods. For example, Lane2Seq achieves the better performance than existing row anchor-based methods, such as CondLaneNet (79.27% vs. $78.74\\%$ ). Moreover, Lane2Seq achieves competitive performance compared to CLRNet (79.27% vs. $79.73\\%$ ). One possible reason for the performance gap between CLRNet and Lane2Seq is that CLRNet adopts multi-scale detection, which can notably improve the detection performance, while Lane2Seq only uses the plain transformer architecture with single-scale detection. For the comparison of parameter-based methods, Lane2Seq surpasses the existing transformer-based methods. For example, Lane2Seq significantly increases F1 score from $64.00\\%$ to $78.39\\%$ compared with LSTR and outperforms Laneformer by $1.33\\%$ (78.39% vs. $77.06\\%$ ). But Lane2Seq is ranked second after BSNet. As for the inference speed, we can see that Lane2Seq does not have a significant advantage in inference speed. We attribute that token-by-token prediction method slows the inference speed.", + "bbox": [ + 76, + 295, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the above results and analyses, we can conclude that the unified lane detection via the sequence generation without any well-designed task-specific components, combining our reinforcement learning-based multi-format model tuning, can achieve promising performance.", + "bbox": [ + 75, + 672, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We display the qualitative results in Fig. 4. The results show that Lane2Seq can effectively detect lanes in the multi-lane scenario (see Fig 4 (a) and (c)). Even in the case of the night scene, Lane2Seq successfully discriminates the lanes (see Fig 4 (b)).", + "bbox": [ + 75, + 747, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 832, + 232, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct the ablation experiments on CULane dataset to validate the effectiveness of each component. More ablation studies are in the supplementary materials.", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36b5580639691b58cde949a59edf57afac424eb24066be334ade7b64b562fd97.jpg", + "table_caption": [ + "Table 4. Ablation study on the multi-format model tuning based on reinforcement learning." + ], + "table_footnote": [], + "table_body": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
Lane2Seq (segmentation)64.2770.4452.77
Lane2Seq (segmentation) + MFRL79.64 (+15.37)85.26 (+14.82)69.00 (+16.23)
Lane2Seq (anchor)66.2371.3655.75
Lane2Seq (anchor) + MFRL79.27 (+13.04)84.72 (+13.36)67.28 (+11.53)
Lane2Seq (parameter)68.1473.2757.89
Lane2Seq (parameter) + MFRL78.39 (+10.25)83.68 (+10.41)66.94 (+9.05)
", + "bbox": [ + 503, + 127, + 890, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/22096d4e1c2c53f2df10527b56d881e67c650144733f2cb21b773936ce27d584.jpg", + "table_caption": [ + "Table 5. Ablation study on the different reward functions. F1 score is adopted as the indicator." + ], + "table_footnote": [], + "table_body": "
\\( R_{seg} \\)\\( R_a \\)\\( R_p \\)SegmentationAnchorParameter
64.2766.2368.14
✓77.69 (+13.42)66.23 (+0.0)68.14 (+0.0)
✓64.27 (+0.0)76.94 (+10.71)68.14 (+0.0)
✓64.27 (+0.0)66.23 (+0.0)77.02 (+8.88)
✓✓78.72 (+14.45)78.14 (+11.91)68.14 (+0.0)
✓✓78.70 (+14.43)66.23 (+0.0)77.90 (+9.76)
✓✓64.27 (+0.0)78.20 (+11.97)77.93 (+9.79)
✓✓✓79.64 (+15.37)79.27 (+13.04)78.39 (+10.25)
", + "bbox": [ + 504, + 260, + 890, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Format Model Tuning Based on Reinforcement Learning. We first ablate the effectiveness of the proposed Multi-Format Model Tuning Based on Reinforcement Learning (MFRL). As shown in Table 4, MFRL significantly improves the performance of different detection formats. For the segmentation format, MFRL gains $15.37\\%$ F1 score improvement $(79.64\\%$ vs. $64.27\\%)$ . For the anchor and parameter format, MFRL increases $13.04\\%$ $(79.27\\%$ vs. $66.23\\%)$ and $10.25\\%$ $(78.39\\%$ vs. $68.14\\%)$ F1 score, respectively. Reasons of performance improvement can be attributed in two folds: (1) the vanilla Lane2Seq is a task-agnostic architecture and lacks the task-specific knowledge, leading to the ineffective feature learning of lane. MFRL tackles this problem by adopting the evaluation metric, which is designed based on the lane prior information like angle, as the reward to tune the model. (2) MFRL makes the model's predictions aligned with their intended usage, i.e., how to achieve a high performance. Optimizing the loss function is a common practice in the compute vision. However, this way indirectly optimizes for model's intended usage, because the lower loss does not mean the higher performance. Instead, MFRL views the maximizing the evaluation metric as the optimization objective, which is positively correlated with the model performance.", + "bbox": [ + 496, + 402, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Reward Function of Different Formats. We ablate the effectiveness of reward function of different formats and results are shwon in Table 5. It can be observed that our reward function can bring consistent performance improvement. However, we can also see that the performance will not be improved if the corresponding reward is not optimized. Besides, we can find that multi-reward optimization can achieves better results than single-reward optimization. Reason may be that multi-reward optimiza", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16950", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f024bdb0778d6a88ea6226dcb7224b0a91ff2310b956d59788a39416b9595492.jpg", + "image_caption": [ + "Figure 4. Visualization results of LaneATT, CLRNet, and Lane2Seq on Tusimple, CULane, and LLAMAS." + ], + "image_footnote": [], + "bbox": [ + 254, + 88, + 717, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73e73b5965c210a9eef740d96793ec307cfb847c6c8632f8966588948eba6598.jpg", + "table_caption": [ + "Table 6. Comparison between Lane2Seq and YOLOP." + ], + "table_footnote": [], + "table_body": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
YOLOP (segmentation)75.2379.4660.58
Lane2Seq (segmentation)79.64 (+4.41)85.26 (+5.8)69.00 (+8.42)
YOLOP (anchor)76.4480.9962.15
Lane2Seq (anchor)79.27 (+2.83)84.72 (+3.73)67.28 (+5.13)
YOLOP (parameter)69.4374.3858.46
Lane2Seq (parameter)78.39 (+8.96)83.68 (+9.30)66.94 (+8.48)
", + "bbox": [ + 81, + 324, + 468, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion enables model to learn correlations between different detection formats and provides more supervised signals.", + "bbox": [ + 75, + 421, + 468, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with other Multi-Task Methods. An alternative method to unify different detection formats is using different detection heads in a model. YOLOP [42] is a representative work for this method. We replace the heads of YOLOP with anchor-based head, segmentation-based head, and parameter-based head. The segmentation-based loss, anchor-based loss, parameter-based loss for YOLOP are same as that in SCNN [28], LaneATT [36], and PolyLaneNet [37]. From Table 6, we can observe that Lane2Seq surpasses YOLOP for all detection formats. It should be noted that YOLOP contains task-specific modules, such as detection head and loss function, but Lane2Seq does not. Results manifest the sequence generation combining with MFRL is a more simple and effective way to unify lane detection.", + "bbox": [ + 75, + 452, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study on the Scale Factor of Different Objective Functions. For the limitation of the space, we take the segmentation format as an example to present the influence of the scale factor of different objective functions in MFRL. Results of other two formats can be found in the supplementary materials. As shown in Table 7, assigning the same factor to different objective functions causes insuperior performance, indicating different detection formats have different contributions. We set $\\lambda_4$ , $\\lambda_5$ , and $\\lambda_6$ to 0.2, 1, and 1.5 according to the model performance.", + "bbox": [ + 75, + 679, + 468, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 844, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a novel sequence generation-based lane detection framework, i.e., Lane2Seq, to unify lane de", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/372aeb99d3350d5adc31c940bc4e18370f14593ba92733e671b64bf9e628351e.jpg", + "table_caption": [ + "Table 7. Ablation study on different scale factors." + ], + "table_footnote": [], + "table_body": "
λ4λ5λ6F1(%)↑
11173.23
0.71174.96
0.41176.27
0.21177.35
0.11177.02
0.21.1177.03
0.21.2176.75
0.20.9177.09
0.20.8177.00
0.210.976.87
0.211.277.37
0.211.479.46
0.211.579.64
0.211.679.39
", + "bbox": [ + 504, + 324, + 890, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tection, which casts lane detection as a sequence generation task. Lane2Seq gets rid of complicated task-specific modules and adopts a simple transformer-based encoder-decoder architecture. To incorporate the task-specific knowledge, we employ a multi-format model tuning based on reinforcement learning (MFRL). Extensive experiments show Lane2Seq is effective and achieves competitive results compared to the state-of-the-art methods.", + "bbox": [ + 496, + 566, + 890, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. One major limitation of Lane2Seq is that sequence generation model is expensive for long sequence (mainly for inference). The inference speed is low when there are more than 5 lanes in an image despite its detection strength in multi-lane scenario. Therefore, future work is required to make it faster for real-time lane detection applications. Another limitation is that MFRL can only be applied to sequence generation-based model currently. Our ongoing work is applying MFRL to other vision models.", + "bbox": [ + 496, + 686, + 892, + 823 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tusimple dataset. https://github.com/TuSimple/tusimple-benchmark. Accessed on 11th August 2023. 5", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2", + "[3] Karsten Behrendt and Ryan Soussan. Unsupervised labeled lane markers using maps. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 5", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4", + "[5] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 3", + "[6] Haoxin Chen, Mengmeng Wang, and Yong Liu. Bsnet: Lane detection via draw b-spline curves nearby. arXiv preprint arXiv:2301.06910, 2023. 6", + "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6", + "[8] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 2, 4", + "[9] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 2", + "[10] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 2, 4", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4", + "[12] Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. Guiding pretraining in reinforcement learning with" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large language models. arXiv preprint arXiv:2302.06692, 2023.4", + "[13] Ruochen Fan, Xuanrun Wang, Qibin Hou, Hanchao Liu, and Tai-Jiang Mu. Spinnet: Spinning convolutional network for lane boundary detection. Computational Visual Media, 5:417-428, 2019. 1, 2", + "[14] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17062-17070, 2022. 6", + "[15] Jianhua Han, Xiajun Deng, Xinyue Cai, Zhen Yang, Hang Xu, Chunjing Xu, and Xiaodan Liang. Laneformer: Object-aware row-column transformers for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 799-807, 2022. 6", + "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 6", + "[17] Hiroto Honda and Yusuke Uchida. Clrernet: Improving confidence of lane detection with laneiou. arXiv preprint arXiv:2305.08366, 2023. 1, 4", + "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17163-17171, 2022. 6", + "[19] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2", + "[20] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019. 2", + "[21] Xuan Liao, Wenhao Li, Qisen Xu, Xiangfeng Wang, Bo Jin, Xiaoyun Zhang, Yanfeng Wang, and Ya Zhang. Iteratively-refined interactive 3d medical image segmentation with multi-agent reinforcement learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9394-9402, 2020. 3", + "[22] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condlanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021. 6", + "[23] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3694-3702, 2021. 1, 2, 6", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[25] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16952", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022. 2", + "[26] Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. End-to-end active object tracking via reinforcement learning. In International conference on machine learning, pages 3286-3295. PMLR, 2018. 3", + "[27] Stefan Mathe, Aleksis Pirinen, and Cristian Sminchisescu. Reinforcement learning for visual object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2894-2902, 2016. 3", + "[28] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5, 6, 8", + "[29] Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Jing Li, Shenggao Zhu, Dahua Lin, Chunhua Shen, et al. Spts: single-point text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4272-4281, 2022. 2", + "[30] André Susano Pinto, Alexander Kolesnikov, Yuge Shi, Lucas Beyer, and Xiaohua Zhai. Tuning computer vision models with task rewards. arXiv preprint arXiv:2302.08242, 2023. 2, 3, 4", + "[31] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 276-291. Springer, 2020. 1, 2, 6", + "[32] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4", + "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 4", + "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[35] Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12, 1999. 5", + "[36] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021. 1, 2, 6, 8", + "[37] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. 2, 6, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6", + "[39] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1392-1401, 2022. 7", + "[40] Ze Wang, Weiqiang Ren, and Qiang Qiu. Lanenet: Realtime lane detection networks for autonomous driving. arXiv preprint arXiv:1807.01726, 2018. 1", + "[41] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8:229-256, 1992. 2", + "[42] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 8", + "[43] Lingyu Xiao, Xiang Li, Sen Yang, and Wankou Yang. Adnet: Lane shape prediction via anchor decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6404-6413, 2023. 6", + "[44] Jiaxing Yang, Lihe Zhang, and Huchuan Lu. Lane detection with versatile atrousformer and local semantic guidance. Pattern Recognition, 133:109053, 2023. 6", + "[45] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2", + "[46] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 6", + "[47] Jia-Qi Zhang, Hao-Bin Duan, Jun-Long Chen, Ariel Shamir, and Miao Wang. Houghlanenet: Lane detection with deep hough transform and dynamic convolution. arXiv preprint arXiv:2307.03494, 2023. 2", + "[48] Zhiyuan Zhao, Qi Wang, and Xuelong Li. Deep reinforcement learning based lane detection and localization. Neurocomputing, 413:328-338, 2020. 3", + "[49] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3547-3554, 2021. 1, 2, 4, 6", + "[50] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 898-907, 2022. 5, 6", + "[51] Kunyang Zhou and Rui Zhou. End to end lane detection with one-to-several transformer. arXiv preprint arXiv:2305.00675, 2023. 2, 6" + ], + "bbox": [ + 501, + 92, + 890, + 868 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16953", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_model.json b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f20f53b55cdf47ddd760c506b28ede3369d43e6f --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.834, + 0.152 + ], + "angle": 0, + "content": "Lane2Seq: Towards Unified Lane Detection via Sequence Generation" + }, + { + "type": "title", + "bbox": [ + 0.4, + 0.182, + 0.571, + 0.218 + ], + "angle": 0, + "content": "Kunyang Zhou Southeast University" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.22, + 0.584, + 0.233 + ], + "angle": 0, + "content": "kunyangzhou@seu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.558 + ], + "angle": 0, + "content": "In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorporate the task-specific knowledge into Lane2Seq. Experimental results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets \\(97.95\\%\\) and \\(97.42\\%\\) F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.588, + 0.21, + 0.604 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Lane detection is a fundamental task in computer vision [31, 36, 40]. It aims to predict the location of the lane in a given image. Lane detection plays a crucial role in many applications, such as adaptive cruise control and lane keeping. Existing lane detection methods generally adopt a divide-and-conquer strategy, which decomposes the lane detection into multiple subtasks. Each subtask is accomplished by a task-specific head network. For instance, as shown in Fig. 1, segmentation-based methods [28,49] adopt a head network along with a task-specific module, such as the message-passing module [28], to predict per-pixel masks. Anchor-based methods [17, 36] utilize a classification head network for distinguishing lane instances and an anchor refinement network for regressing accurate lanes. Parameter-based methods [13, 23] use a network to predict the parameters of lanes and a vertical offset prediction network to locate the start points of lanes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Although divide-and-conquer strategy has been proved an effectively way to address the certain subtask in the ex" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.267, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.302, + 0.743, + 0.312 + ], + "angle": 0, + "content": "(a) Segmentation-based lane detection" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.325, + 0.891, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.378, + 0.72, + 0.388 + ], + "angle": 0, + "content": "(b) Anchor-based lane detection" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.404, + 0.891, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.461, + 0.733, + 0.47 + ], + "angle": 0, + "content": "(c) Parameter-based lane detection" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.485, + 0.891, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.579, + 0.796, + 0.589 + ], + "angle": 0, + "content": "(d) Our sequence generation-based lane detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.601, + 0.88, + 0.615 + ], + "angle": 0, + "content": "Figure 1. Comparison of different lane detection frameworks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.717 + ], + "angle": 0, + "content": "isting methods, there still exist several limitations. (1) Each subtask needs a customized task-specific head network, resulting in a complicated lane detection model. (2) Each task-specific head requires one or more loss functions, e.g., cross-entropy loss and Line-IOU loss [17], bringing out extra hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.893, + 0.87 + ], + "angle": 0, + "content": "In this paper, we present a novel sequence generation-based lane detector (Lane2Seq) framework to tackle the aforementioned issues. By formulating the lane detection as a sequence generation task, Lane2Seq gets rid of customized head networks and task-specific loss function. It is based on the intuition that if the detection model knows where the target lane is, model can be simply taught how to read the location of lane out, instead of designing additional classification head or regression head by the divide-and-conquer strategy." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Therefore, we convert the outputs of different lane detection formats, i.e., segmentation-based, anchor-based, and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "parameter-based, into a sequence of discrete tokens and the model learns to generate this sequence token-by-token. As shown in Fig. 1 (d), to achieve a specific lane detection format, Lane2Seq uses a prompt to specify the detection format and the generated sequence adapts to the prompt so the model can produce format-specific output. By the format-specific prompt, Lane2Seq unifies different lane detection formats into a model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.473, + 0.442 + ], + "angle": 0, + "content": "While Lane2Seq does not contain task-specific components, task-specific knowledge contained in these components can help the model learn the features of lanes better. We propose a Multi-Format model tuning method based on Reinforcement Learning (MFRL) to incorporate the task-specific knowledge into the model without changing model's architecture. Inspired by Task-Reward [30], MFRL takes the evaluation metrics, which naturally integrates task-specific knowledge, as the reward and tunes Lane2Seq using REINFORCE [41] algorithm. However, evaluation metrics like F1 score cannot be used as the reward directly due to undecomposable as a sum of perexample rewards. In this paper, we propose three new evaluation metric-based rewards for segmentation, anchor, and parameter format, based on their task-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.609 + ], + "angle": 0, + "content": "Experimental results demonstrate that our Lane2Seq achieves competitive performance on three public datasets, Tusimple, CULane, and LLAMAS. For instance, Lane2Seq with ViT-Base encoder obtains \\(97.95\\%\\) and \\(97.42\\%\\) F1 score on Tusimple and LLAMAS, setting a new state-of-the-art result for two datasets. It should be noted that all existing lane detection methods heavily rely on well-designed task-specific head networks and the corresponding complicated loss functions. Instead, our Lane2Seq utilizes a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.61, + 0.438, + 0.626 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.64, + 0.47, + 0.73 + ], + "angle": 0, + "content": "- We propose a sequence generation-based method for lane detection, which casts the lane detection as a sequence generation task. To the best of our knowledge, we are the first to unify the lane detection through the sequence generation, which offers a new perspective on lane detection." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.747, + 0.47, + 0.809 + ], + "angle": 0, + "content": "- We present a novel reinforcement learning-based multi-format model tuning, including three new evaluation metric-based reward functions, to incorporate the task-specific knowledge into the model." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.825, + 0.47, + 0.886 + ], + "angle": 0, + "content": "- Experimental results show that our method achieves competitive performance on lane detection benchmarks. Remarkably, we establish a new state-of-the-art result on Tusimple and LLAMAS." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.64, + 0.47, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Lane detection. Existing lane detection methods can be divided into three categories based on the representation of the lane: segmentation-based method, anchor-based method and parameter-based method. Segmentation-based methods [28, 47, 49] consider lane detection as a semantic segmentation task and performs pixel-wise prediction. SCNN [28] enhances the visual evidence by a message-passing module, which can capture spatial dependency for lanes. Anchor-based methods [20, 31, 36, 51] predict the accurate lanes by refining the predefined lane anchors. UFLD [31] proposes a novel row anchor-based approach to detect lanes. Different from the segmentation-based and anchor-based methods, parameter-based methods [13, 23, 37] regard the lane detection as the parametric modeling and regress the parameters of the lane. PolyLaneNet [37] models a lane curve as a polynomial and regresses the parameters of the polynomial. In this paper, we treat the segmentation-based, anchor-based methods, and parameter-based metohods uniformly as the sequence generation task, getting rid of the complicate structure and task-specific modules, e.g., head network. It only adopts the cross-entropy loss and plain transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Sequence generation for vision tasks. Recently, sequence-to-sequence (seq2seq) method used in the natural language processing (NLP) has been applied for vision tasks. Seq2seq utilizes a basic transformer encoder and decoder architecture and accomplishes the task by making sequence predictions, rather than designing a model tailored to the vision task. Pix2seq [8] is the pioneering work to cast the object detection as the sequence generation task. It shows that object can be detected well without any task-specific modules like label assignment. Besides object detection, seq2seq has been extended to other vision tasks such as instance segmentation [9], keypoint detection [9], text spotting [19,29] and object tracking [10]. UniTAB [45] adopts the task prompt [33] to perform multi-task learning. Unified-IO [25] jointly trains various vision tasks by setting the unified input/output formats for all tasks. Besides, seq2seq becomes increasingly popular in the multimodal model. Text-to-image models like DALL-E [34] and vision-language models like Flamingo [2] all use seq2seq to unify the multi-modal tasks. Nevertheless, how to perform the seq2seq in the lane detection to unify different detection formats is still unexplored." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.886 + ], + "angle": 0, + "content": "Lane2Seq has a similar spirit with the Pix2seq and its successors [9, 10, 19, 29]. All of them view the vision tasks as the sequence generation. The main difference between these methods and Lane2Seq is the sequence format. Previous works use coordinates and category to construct the sequence, while Lane2Seq also adopts the parameter of the lane to construct the sequence." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Reinforcement learning in the computer vision. Many" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.087, + 0.837, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.387, + 0.895, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Inference pipeline of Lane2Seq. The model perceives input image, prompt and generates format-specific tokens, which can be detokenized into required detection format for visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.653 + ], + "angle": 0, + "content": "previously researches introduce the reinforcement learning for the vision tasks, such as object detection [27], object tracking [26], image segmentation [21] and lane detection [48]. They generally focus on learning different parts of an image and refine the outputs iteratively. DQLL [48] localizes the lanes as a group of lanemarks and refine the location of lanemarks with the deep Q-learning. Recently, Task Reward [30] adopts a novel task risk-based reinforcement learning to tune the vision model without changing the model architecture. Reward functions in Task-Reward are designed for object detection, instance segmentation, colorization, and image captioning but not for lane detection. Our MFRL is the first attempt to design effective evaluation metric-based reward functions for different lane detection formats." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.671, + 0.17, + 0.687 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.804 + ], + "angle": 0, + "content": "In this section, we present the proposed sequence generation-based lane detection method, named Lane2Seq, in detail. The overall pipeline of Lane2Seq is illustrated in Fig. 2. Sec 3.1 depicts the unified interface for lane detection, Sec 3.2 details the model architecture and objective function, and Sec 3.3 introduces the proposed multi-format model tuning based on reinforcement learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.394, + 0.832 + ], + "angle": 0, + "content": "3.1. Unified Interface for Lane Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As presented in Figure 1, lane detection formats in the existing method are diverse and have been formulated quite differently. Considering differences in the form of outputs, customized models with specialized head networks and loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.428, + 0.889, + 0.442 + ], + "angle": 0, + "content": "functions are designed for different lane detection formats." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.594 + ], + "angle": 0, + "content": "To integrate different lane detection formats into a model, we propose an unified sequence interface for lane detection, where both format transcription like segmentation and outputs are treated as sequences of discrete tokens. As shown in Fig. 2, the generated sequence is composed of four parts, a starting token \\( \\langle \\text{starting} \\rangle \\), a format transcription token like \\( \\langle \\text{Segmentation} \\rangle \\), a format-specific sequence, and an ending token \\( \\langle \\text{end} \\rangle \\). The format-specific sequence for three detection formats can be constructed as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.839 + ], + "angle": 0, + "content": "Segmentation sequence. Instead of performing pixelwise mask prediction, we predict the polygon [5] corresponding to the mask as a sequence of coordinates conditioned on a given lane instance. Then, we convert the polygon into a sequence by quantizing the coordinates of its points into discrete tokens. Specifically, \\( x, y \\) coordinates of a point are normalized to the width and height of the image, and then quantized to \\( [1, n_{\\text{bins}}] \\), where \\( n_{\\text{bins}} \\) is the size of the vocabulary. Vocabulary will be described subsequently. A polygon sequence can be expressed as \\( [x_1, y_1, x_2, y_2, \\ldots, x_{28}, y_{28}, <\\text{Lane}>] \\), where \\( <\\text{Lane}> \\) is the category token. If there are multiple lanes in an image, we concatenate the all polygon sequences. The segmentation sequence consists of a starting point and all polygon sequences, where starting point is the left-top point of the image \\( x_0 = 0, y_0 = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Anchor sequence. Since the essence of anchor-based methods is regressing the location of keypoints of lane anchors, we treat the anchor prediction as the keypoint sequence generation. Specifically, the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.462, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Training pipeline of Lane2Seq. The input sequence can be the segmentation sequence or anchor sequence or parameter sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.354 + ], + "angle": 0, + "content": "keypoint sequence of a lane can be expressed as \\([x_1, y_1, x_2, y_2, \\ldots, x_{14}, y_{14}, <\\text{Lane}>\\]\\. The normalization and quantization of keypoint coordinates are same to that of polygon point. The anchor sequence contains a starting point and all keypoint sequences, where starting point is still the left-top point of the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.469, + 0.489 + ], + "angle": 0, + "content": "Parameter sequence. the sequence of the lane includes two parts: the parameter of polynomial function and vertical offsets. We set the polynomial degree as 5th. Specifically, the parameter sequence of a lane can be represented as \\([a_1,a_2,a_3,a_4,a_5,s,\\langle \\mathrm{Lane}\\rangle ]\\), where \\(s\\) is the vertical offset. \\(s\\) is normalized to the height of the image and quantized to \\([1,n_{\\mathrm{bins}}]\\). We use sigmoid to normalize \\(a_i\\) by \\(a_{i} = \\operatorname {Sigmoid}(a_{i})\\). Then \\(a_{i}\\) is quantized to \\([1,n_{\\mathrm{bins}}]\\). Parameter sequence does not require the starting point." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.469, + 0.551 + ], + "angle": 0, + "content": "Vocabulary. We use a shared vocabulary for all formats and each integer between \\([1, n_{bins}]\\) can be regarded as a word in the vocabulary. Each word in the vocabulary corresponds to a learnable embedding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.456, + 0.577 + ], + "angle": 0, + "content": "3.2. Unified Architecture and Objective Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.469, + 0.764 + ], + "angle": 0, + "content": "We follow [10] and use a transformer-based encoder-decoder architecture to deal with image input and sequence output flexibly. As shown in Fig. 3, the image encoder takes the pixels as the input and outputs the corresponding image representations. We utilize the Vision Transformer (ViT) [11] to instantiate the image encoder. We adopt a transformer-based sequence decoder, which is widely used in the language modeling [32, 33], to generate the output sequence. The decoder generates one token at a time based on the preceding tokens and the image representations. Sequence decoder removes the well-designed task-specific heads for different detection formats." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Training Pipeline. We first tokenize annotation of each format into the corresponding sequence. Then, we construct a training batch using images and sequences from all detection formats. Finally, we compute loss between generated sequence and target sequence for each format. For a certain detection format, the input sequence is \\([<\\text{starting}>\\), format transcription token, format-specific sequence] and the target sequence is [format transcription token, format-specific sequence, \\(\\langle \\mathrm{end}\\rangle\\)]. For example, in the anchor format, the in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.093, + 0.892, + 0.212 + ], + "angle": 0, + "content": "put sequence is \\([<\\text{starting}>, <\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]\\) and the target sequence is \\([<\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]\\). The starting token \\(<\\text{starting}>\\) and ending token \\(<\\text{end}>\\) are learnable embeddings, which tell the decoder when to begin and end the sequence generation. The sequence decoder perceives the image representations and input sequence and reconstructs the target sequence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Objective Function. Similar to Pix2seq [8], we train the Lane2Seq with a simple cross-entropy loss. At each time stamp \\( j \\), Lane2Seq aims to maximize the likelihood of the target tokens conditioned on the image representations \\( I \\) and previously generated tokens \\( y_{1:j-1} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.293, + 0.892, + 0.336 + ], + "angle": 0, + "content": "\\[\nL o s s _ {o b j} = - \\sum_ {j = 1} ^ {N} w _ {j} \\log P (\\hat {m} _ {j} | I, m _ {1: j - 1}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.43 + ], + "angle": 0, + "content": "where \\( m \\) and \\( \\hat{m} \\) denote the input and target sequence, respectively. \\( N \\) is the length of sequence. \\( w_{j} \\) represents the weight for the \\( j \\)-th token, 0 is assigned to the format transcription token and 1 is assigned to other tokens to ensure the model is trained to predict the desired tokens but not the format transcription tokens." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Inference Pipeline. We take the format transcription to-ken as the prompt to achieve the format-specific detection and the sequence decoder generates the rest of the sequence conditioned on the prompt and image representation. Once the whole sequence is generated, we de-quantize the sequence to obtain the location of lane. More details about de-quantization for each format can be found in the supplementary materials. Inference process is presented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.593 + ], + "angle": 0, + "content": "3.3. Multi-Format Model Tuning Based on Reinforcement Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Lane2Seq get rids of the complicated task-specific components, such as task-specific head networks, via the sequence generation. However, this task-agnostic architecture inevitably lacks the task-specific knowledge of the lane detection, making the model ineffective in learning the features of lane. Existing lane detection methods [17, 49] usually design the specific modules to incorporate the task-specific knowledge into the model. But this way makes the model architecture complicate. We propose a novel Multi-Format model tuning method based on Reinforcement Learning (MFRL) to learn task-specific knowledge effectively without changing any components of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Inspired by Task-Reward [30], MFRL takes the evaluation metric, which naturally contains the task-specific knowledge, as the reward function and adopts the reinforcement learning to tune the model, which is common practice in the modern language model [4, 12]. Specifically, it is difficult for the model to converge if model is trained with reinforcement learning method from scratch [30]. Hence, MFRL consists of two stages: the pretraining stage and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "the model tuning stage. In the pretraining stage, model is trained on the lane detection dataset with objective function in Eq. 1 to have a good weight initialization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.137, + 0.47, + 0.168 + ], + "angle": 0, + "content": "In the model tuning stage, we use the REINFORCE algorithm [35] to maximize the objective function as below," + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.182, + 0.469, + 0.2 + ], + "angle": 0, + "content": "\\[\nO b j = E _ {c \\sim D} \\left[ E _ {t \\sim Q (\\cdot | c, u)} R (t, g) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.469, + 0.304 + ], + "angle": 0, + "content": "where \\(c, t,\\) and \\(g\\) represent the input image, generated format-specific sequences, and ground truths, respectively. \\(E\\) and \\(u\\) denote the mathematic expectation and the model parameter. \\(R, D,\\) and \\(Q\\) stand for the reward function, data distribution of the dataset, and conditional distribution parameterized by \\(u\\). REINFORCE algorithm estimates the gradient of the reward function by" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.317, + 0.469, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {u} E _ {t \\sim Q} [ R (t, g) ] = E _ {t \\sim Q} [ R (t, g) \\nabla_ {u} \\log Q (t | c, u) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.469, + 0.405 + ], + "angle": 0, + "content": "In practice, Eq. 3 is computed as an average of per-sequence gradient and reward function of per-sequence is not required to be differentiable. More details about REINFORCE algorithm can be found in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.469, + 0.54 + ], + "angle": 0, + "content": "Reward Function. Evaluation metric F1 score and accuracy are unable to adopt as the reward directly, because they cannot be decomposable as a sum of per-sequence rewards (see Sec 4.2). However, F1 score or accuracy is composed of false positives \\((FP)\\), true positives \\((TP)\\), and false negatives \\((FN)\\). All three indicators can be computed per-sequence. We opt to use \\(TP\\) and \\(FP\\) to design the reward. Specifically, reward for three detection formats are constructed as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.469, + 0.709 + ], + "angle": 0, + "content": "Reward for the segmentation format (termed as \\( R_{seg} \\)). Segmentation-based lane detection includes two task-specific knowledge, i.e., knowledge of the segmentation and lane detection. For the knowledge of lane detection, we adopt a matched Line-IOU (LIOU) [50], which can demonstrate the quality of location prediction and shape prediction of the lane. For the knowledge of the segmentation, we adopt the matched Mean Intersection over Union (mIOU), which is a widely used metric in the semantic segmentation to evaluate the quality of segmentation, as the reward. Finally, we add both of them to compute \\( r_{seg} \\) as below," + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.718, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {s e g} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + m I O U \\left(p _ {k}, g\\right) \\right] \\tag {4} \\\\ - \\lambda_ {1} F P _ {s e g} (t, g), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.854 + ], + "angle": 0, + "content": "where \\(K\\) and \\(p_k\\) represents the number of true positives in \\(t\\) and the \\(k\\)-th true positive. \\(FP_{seg}\\) is the false positive rate of the segmentation format and \\(\\lambda_1\\) is the weight to control the affect of \\(FP_{seg}\\). We introduce \\(FP_{seg}\\) to penalize for false positives." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Reward for the anchor format (termed as \\( R_{a} \\)). Anchors-based lane detection contains the knowledge of the keypoint location and lane detection. For the knowledge of keypoint" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "location, we simply adopt the matched Euclidean distance \\( d(p_k, g) \\) between true positives and corresponding ground truths as the reward. Since MFRL needs to maximize the reward, we rescale the reward as \\( d_r(p_k, g) = 1 - \\frac{d(p_k, g)}{H} \\), where \\( H \\) is the height of image. For the knowledge of lane detection, we also utilize the matched LIOU as the reward. The formulation of \\( r_a \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.205, + 0.892, + 0.253 + ], + "angle": 0, + "content": "\\[\nR _ {a} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + d _ {r} \\left(p _ {k}, g\\right) \\right] - \\lambda_ {2} F P _ {a} (t, g), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.284 + ], + "angle": 0, + "content": "where \\( FP_{a} \\) and \\( \\lambda_{2} \\) stand for the false positives rate of the anchor format and the weight of \\( FP_{a} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.343 + ], + "angle": 0, + "content": "Reward for the parameter format (termed as \\( R_p \\)). Parameter-based detection only has the knowledge of lane detection, hence we directly use the matched LIOU as the reward." + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.346, + 0.892, + 0.383 + ], + "angle": 0, + "content": "\\[\nR _ {p} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} L I O U \\left(p _ {k}, g\\right) - \\lambda_ {3} F P _ {p} (t, g), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.384, + 0.892, + 0.414 + ], + "angle": 0, + "content": "where \\( FP_{p} \\) and \\( \\lambda_{3} \\) stand for the false positives rate of the parameter format and the weight of \\( FP_{p} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Since different detection formats have different contributions, we weight the objective function of different formats. The final objective function of MFRL can be computed as," + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.468, + 0.892, + 0.483 + ], + "angle": 0, + "content": "\\[\nO b j _ {t o t a l} = \\lambda_ {4} O b j _ {s e g} + \\lambda_ {5} O b j _ {a} + \\lambda_ {6} O b j _ {p}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.566 + ], + "angle": 0, + "content": "where \\( Obj_{seg} \\), \\( Obj_a \\), and \\( Obj_p \\) denote the objective function of segmentation, anchor, and parameter format, respectively. The objective function of three formats is the same as Eq. 2, except the format-specific sequence and reward function. \\( \\lambda_4 \\), \\( \\lambda_5 \\), and \\( \\lambda_6 \\) are the scale factor." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.633, + 0.596 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.602, + 0.619 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.658 + ], + "angle": 0, + "content": "We conduct experiments on three lane detection benchmarks: CULane [28], Tusimple [1], and LLAMAS [3]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.733 + ], + "angle": 0, + "content": "CULane is a widely used large-scale dataset for lane detection. It contains a lot of challenging scenarios such as crowded roads. The CULane dataset consists of 88.9K images for training, 9.7K images in the validation set, and 34.7K images for the test. Image size is \\(1640 \\times 590\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.778 + ], + "angle": 0, + "content": "Tusimple is a real highway dataset consisting of 3,626 training images and 2,782 testing images. All images have \\(1280 \\times 720\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.838 + ], + "angle": 0, + "content": "LLAMAS is a recently released large-scale lane detection dataset with over 100K images. All lane markers are annotated with high-accurate maps. Image size is \\(1280 \\times 717\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.685, + 0.862 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For CULane and LLAMAS dataset, we adopt the F1 score to measure the performance: \\( F_{1} = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16948" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of F1 score and MACs (multiply-accumulate operations) on CULane testing set. We only report the false positives for \"Cross\" category following [50]. FPS is tested on a single 2080Ti." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.894, + 0.373 + ], + "angle": 0, + "content": "
MethodsEncoderFPS↑Normal↑Crowded↑Dazzle↑Shadow↑No line↑Arrow↑Curve↑Night↑Cross↓Total↑MACs(G)↓
Segmentation based
SCNN [28]ResNet501290.6069.7058.5066.9043.4084.1064.4066.10190071.608.4
RESA [49]ResNet5042.192.1073.1069.2072.8047.7088.3070.3069.90150375.37.6
AtrousFormer [44]ResNet34-92.8375.9669.4877.8650.1588.6671.1473.74105478.0814.22
LaneAF [44]DLA34 [46]-91.8075.6171.7879.1251.3886.8872.7073.03136077.4112.73
Lane2Seq (segmentation)ViT-Base1993.3977.2773.4579.6953.9190.5373.3774.96112979.6415.19
Anchor based
LaneATT [36]ResNet1222391.7476.1669.4776.3150.4686.2964.0570.81126477.0270.5
O2SFformer [51]ResNet504493.0976.5772.2576.5652.8089.5069.6073.85311877.8327.51
UFLD [31]ResNet3419290.7070.2059.5069.3044.4085.7069.5066.70203772.303.3
CondLaneNet [22]ResNet3413593.3877.1471.1779.9351.8589.8973.8873.92138778.7419.6
ADNet [43]ResNet34-92.9077.4571.7179.1152.8989.9070.6474.78149978.9418.6
CLRNet [50]ResNet3411293.4978.0674.5779.9254.0190.5972.7775.02121679.7317.3
Lane2Seq (anchor)ViT-Base2393.1177.4373.2579.4653.7490.0272.4475.12117379.2715.19
Parameter based
LSTR [23]ResNet18----------64.005.9
BezierLaneNet [14]ResNet18-90.2271.5562.4970.9145.3084.0958.9868.7099673.674.7
BSNet [6]ResNet34-93.7578.0176.6579.5554.6990.7273.9975.28144579.8917.2
Eigenlanes [18]ResNet50-91.7076.0069.8074.1052.2087.7062.9071.80150977.2018.7
Laneformer [15]ResNet50-91.7775.7470.1775.7548.7387.6566.3371.041977.0626.2
Lane2Seq (parameter)ViT-Base3093.0376.4272.1778.3252.8989.6772.6773.98131978.3915.19
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.457, + 0.434 + ], + "angle": 0, + "content": "\\(\\frac{2 \\times \\text{Precision} \\times \\text{Recall}}{\\text{Precision} + \\text{Recall}}\\), where Precision = TP/TP+FP and Recall = TP/TP+FN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.471, + 0.562 + ], + "angle": 0, + "content": "For Tusimple dataset, we use F1 score, accuracy, false positives, and false negatives to evaluate the model performance. Accuracy is defined as Accuracy = \\(\\frac{\\sum_{clip} C_{clip}}{\\sum_{clip} S_{clip}}\\), where \\(C_{clip}\\) represents the number of accurately predicted lane points and \\(S_{clip}\\) denotes the total number of lane points of a clip. A lane point is considered as correct if its distance is smaller than the given threshold \\(t_{pc} = \\frac{20}{\\cos(a_{yl})}\\), where \\(a_{yl}\\) denotes the angle of the corresponding ground truth." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.295, + 0.59 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Model. We adopt the ViT-Base initialized with the MAE [16] pretrained parameters as the image encoder. The patch size is \\( 16 \\times 16 \\). The decoder is composed of 2 transformer blocks and its hidden size is 256. The number of attention heads is 8 and the hidden size of feed-forward network is 1024. \\( n_{bins} \\) is set to 1000 and the dimension of word embedding of the vocabulary is 256. Our Lane2Seq is implemented based on MMDetection toolbox [7]. \\( \\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4, \\lambda_5 \\), and \\( \\lambda_6 \\) are set to 0.3, 0.3, 0.1, 0.2, 1, and 1.5, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Training. All input images are resized to \\(320 \\times 800\\). We utilize the AdamW [24] as the optimizer with initial learning rate 1e-4. The training epochs of the pretraining stage are 5, 20, and 15 for CULane, Tusimple, and LLAMAS. We set the training epochs of the model tuning stage to 15, 30, and 55 for CULane, Tusimple, and LLAMAS, respectively. For the data augmentation, we use the random horizontal flips and random affine transformation including scaling, rotation, and translation. All experiments are conducted on 8 A100 GPUs with total batch size 384." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.395, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Table 2. Comparison of the performance of different models on LLAMAS." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.573 + ], + "angle": 0, + "content": "
MethodsEncoderF1(%)↑
PolyLaneNet [37]EfficientNet-b0 [38]90.20
BezierLaneNetResNet3496.11
LaneATTResNet3494.96
LaneATTResNet12295.17
LaneAFDLA3496.90
CLRNetResNet1896.96
CLRNetDLA3497.16
Lane2Seq (segmentation)ViT-Base97.42
Lane2Seq (anchor)ViT-Base97.05
Lane2Seq (parameter)ViT-Base96.74
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.587, + 0.892, + 0.604 + ], + "angle": 0, + "content": "4.4. Comparison with the State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Performance on LLAMAS. Our Lane2Seq achieves a new state-of-the-art performance on LLAMAS dataset using segmentation format. In Table 2, Lane2Seq increases F1 score from \\(97.16\\%\\) to \\(97.42\\%\\) compare to the previous state-of-the-art method CLRNet. Compared with parameter-based state-of-the-art method BezierLaneNet, Lane2Seq performs better than BezierLaneNet (96.74% vs. \\(96.11\\%\\)). The results manifest the detection strengths of Lane2Seq in the multi-lane scenario (The number of lanes \\(\\geq 5\\)), such as highways. The reason may be that transformer-based architecture has the advantages in capturing long-range dependency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Performance on Tusimple. Table 3 presents the performance comparison with state-of-the-art approaches on Tusimple. Due to small data scale and single scene, the performance gap between different models is small. Our Lane2Seq using segmentation format sets a new state-of-the-art F1 score of \\(97.95\\%\\). The results also validate the detection strengths of Lane2Seq in multi-lane scenario." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16949" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.09, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 3. Comparison of the performance of different models on Tusimple. Acc denotes accuracy." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.47, + 0.272 + ], + "angle": 0, + "content": "
MethodsEncoderF1(%)↑Acc(%)↑FP(%)↓FN(%)↓
SCNNVGG1695.9796.536.171.80
RESAResNet3496.9396.823.632.48
PolyLaneNetEfficientNet-bo90.6293.369.429.33
UFLDResNet3488.0295.8618.913.75
LaneATTResNet12296.0696.105.642.17
GANet [39]ResNet3497.6895.871.992.64
CondLaneNetResNet3496.9895.372.203.82
CondLaneNetResNet10197.2496.542.013.50
CLRNetResNet3497.8296.872.272.08
CLRNetResNet10197.6296.832.372.38
Lane2Seq (segmentation)ViT-Base97.9596.852.012.03
Lane2Seq (anchor)ViT-Base97.8696.722.212.05
Lane2Seq (parameter)ViT-Base96.5996.002.233.54
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.296, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Performance on CULane. We compare Lane2Seq with other state-of-the-art methods on CULane and results are shown in Table 1. For the segmentation-based methods, Lane2Seq improves F1 score from \\(78.08\\%\\) to \\(79.64\\%\\) compared to the previous state-of-the-art methods AtrousFormer. Compared with anchor-based methods, Lane2Seq outperforms most previous methods. For example, Lane2Seq achieves the better performance than existing row anchor-based methods, such as CondLaneNet (79.27% vs. \\(78.74\\%\\)). Moreover, Lane2Seq achieves competitive performance compared to CLRNet (79.27% vs. \\(79.73\\%\\)). One possible reason for the performance gap between CLRNet and Lane2Seq is that CLRNet adopts multi-scale detection, which can notably improve the detection performance, while Lane2Seq only uses the plain transformer architecture with single-scale detection. For the comparison of parameter-based methods, Lane2Seq surpasses the existing transformer-based methods. For example, Lane2Seq significantly increases F1 score from \\(64.00\\%\\) to \\(78.39\\%\\) compared with LSTR and outperforms Laneformer by \\(1.33\\%\\) (78.39% vs. \\(77.06\\%\\)). But Lane2Seq is ranked second after BSNet. As for the inference speed, we can see that Lane2Seq does not have a significant advantage in inference speed. We attribute that token-by-token prediction method slows the inference speed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Based on the above results and analyses, we can conclude that the unified lane detection via the sequence generation without any well-designed task-specific components, combining our reinforcement learning-based multi-format model tuning, can achieve promising performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Qualitative Results. We display the qualitative results in Fig. 4. The results show that Lane2Seq can effectively detect lanes in the multi-lane scenario (see Fig 4 (a) and (c)). Even in the case of the night scene, Lane2Seq successfully discriminates the lanes (see Fig 4 (b))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.233, + 0.849 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.903 + ], + "angle": 0, + "content": "We conduct the ablation experiments on CULane dataset to validate the effectiveness of each component. More ablation studies are in the supplementary materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation study on the multi-format model tuning based on reinforcement learning." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.128, + 0.892, + 0.206 + ], + "angle": 0, + "content": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
Lane2Seq (segmentation)64.2770.4452.77
Lane2Seq (segmentation) + MFRL79.64 (+15.37)85.26 (+14.82)69.00 (+16.23)
Lane2Seq (anchor)66.2371.3655.75
Lane2Seq (anchor) + MFRL79.27 (+13.04)84.72 (+13.36)67.28 (+11.53)
Lane2Seq (parameter)68.1473.2757.89
Lane2Seq (parameter) + MFRL78.39 (+10.25)83.68 (+10.41)66.94 (+9.05)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.223, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Table 5. Ablation study on the different reward functions. F1 score is adopted as the indicator." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.261, + 0.892, + 0.382 + ], + "angle": 0, + "content": "
\\( R_{seg} \\)\\( R_a \\)\\( R_p \\)SegmentationAnchorParameter
64.2766.2368.14
✓77.69 (+13.42)66.23 (+0.0)68.14 (+0.0)
✓64.27 (+0.0)76.94 (+10.71)68.14 (+0.0)
✓64.27 (+0.0)66.23 (+0.0)77.02 (+8.88)
✓✓78.72 (+14.45)78.14 (+11.91)68.14 (+0.0)
✓✓78.70 (+14.43)66.23 (+0.0)77.90 (+9.76)
✓✓64.27 (+0.0)78.20 (+11.97)77.93 (+9.79)
✓✓✓79.64 (+15.37)79.27 (+13.04)78.39 (+10.25)
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Multi-Format Model Tuning Based on Reinforcement Learning. We first ablate the effectiveness of the proposed Multi-Format Model Tuning Based on Reinforcement Learning (MFRL). As shown in Table 4, MFRL significantly improves the performance of different detection formats. For the segmentation format, MFRL gains \\(15.37\\%\\) F1 score improvement \\((79.64\\%\\) vs. \\(64.27\\%)\\). For the anchor and parameter format, MFRL increases \\(13.04\\%\\) \\((79.27\\%\\) vs. \\(66.23\\%)\\) and \\(10.25\\%\\) \\((78.39\\%\\) vs. \\(68.14\\%)\\) F1 score, respectively. Reasons of performance improvement can be attributed in two folds: (1) the vanilla Lane2Seq is a task-agnostic architecture and lacks the task-specific knowledge, leading to the ineffective feature learning of lane. MFRL tackles this problem by adopting the evaluation metric, which is designed based on the lane prior information like angle, as the reward to tune the model. (2) MFRL makes the model's predictions aligned with their intended usage, i.e., how to achieve a high performance. Optimizing the loss function is a common practice in the compute vision. However, this way indirectly optimizes for model's intended usage, because the lower loss does not mean the higher performance. Instead, MFRL views the maximizing the evaluation metric as the optimization objective, which is positively correlated with the model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Reward Function of Different Formats. We ablate the effectiveness of reward function of different formats and results are shwon in Table 5. It can be observed that our reward function can bring consistent performance improvement. However, we can also see that the performance will not be improved if the corresponding reward is not optimized. Besides, we can find that multi-reward optimization can achieves better results than single-reward optimization. Reason may be that multi-reward optimiza" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.256, + 0.089, + 0.718, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.271, + 0.802, + 0.287 + ], + "angle": 0, + "content": "Figure 4. Visualization results of LaneATT, CLRNet, and Lane2Seq on Tusimple, CULane, and LLAMAS." + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.3, + 0.434, + 0.314 + ], + "angle": 0, + "content": "Table 6. Comparison between Lane2Seq and YOLOP." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.325, + 0.469, + 0.414 + ], + "angle": 0, + "content": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
YOLOP (segmentation)75.2379.4660.58
Lane2Seq (segmentation)79.64 (+4.41)85.26 (+5.8)69.00 (+8.42)
YOLOP (anchor)76.4480.9962.15
Lane2Seq (anchor)79.27 (+2.83)84.72 (+3.73)67.28 (+5.13)
YOLOP (parameter)69.4374.3858.46
Lane2Seq (parameter)78.39 (+8.96)83.68 (+9.30)66.94 (+8.48)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.452 + ], + "angle": 0, + "content": "tion enables model to learn correlations between different detection formats and provides more supervised signals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.453, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Comparison with other Multi-Task Methods. An alternative method to unify different detection formats is using different detection heads in a model. YOLOP [42] is a representative work for this method. We replace the heads of YOLOP with anchor-based head, segmentation-based head, and parameter-based head. The segmentation-based loss, anchor-based loss, parameter-based loss for YOLOP are same as that in SCNN [28], LaneATT [36], and PolyLaneNet [37]. From Table 6, we can observe that Lane2Seq surpasses YOLOP for all detection formats. It should be noted that YOLOP contains task-specific modules, such as detection head and loss function, but Lane2Seq does not. Results manifest the sequence generation combining with MFRL is a more simple and effective way to unify lane detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Ablation Study on the Scale Factor of Different Objective Functions. For the limitation of the space, we take the segmentation format as an example to present the influence of the scale factor of different objective functions in MFRL. Results of other two formats can be found in the supplementary materials. As shown in Table 7, assigning the same factor to different objective functions causes insuperior performance, indicating different detection formats have different contributions. We set \\(\\lambda_4\\), \\(\\lambda_5\\), and \\(\\lambda_6\\) to 0.2, 1, and 1.5 according to the model performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.196, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "This paper presents a novel sequence generation-based lane detection framework, i.e., Lane2Seq, to unify lane de" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.3, + 0.843, + 0.314 + ], + "angle": 0, + "content": "Table 7. Ablation study on different scale factors." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.325, + 0.892, + 0.551 + ], + "angle": 0, + "content": "
λ4λ5λ6F1(%)↑
11173.23
0.71174.96
0.41176.27
0.21177.35
0.11177.02
0.21.1177.03
0.21.2176.75
0.20.9177.09
0.20.8177.00
0.210.976.87
0.211.277.37
0.211.479.46
0.211.579.64
0.211.679.39
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.687 + ], + "angle": 0, + "content": "tection, which casts lane detection as a sequence generation task. Lane2Seq gets rid of complicated task-specific modules and adopts a simple transformer-based encoder-decoder architecture. To incorporate the task-specific knowledge, we employ a multi-format model tuning based on reinforcement learning (MFRL). Extensive experiments show Lane2Seq is effective and achieves competitive results compared to the state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Limitation. One major limitation of Lane2Seq is that sequence generation model is expensive for long sequence (mainly for inference). The inference speed is low when there are more than 5 lanes in an image despite its detection strength in multi-lane scenario. Therefore, future work is required to make it faster for real-time lane detection applications. Another limitation is that MFRL can only be applied to sequence generation-based model currently. Our ongoing work is applying MFRL to other vision models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16951" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Tusimple dataset. https://github.com/TuSimple/tusimple-benchmark. Accessed on 11th August 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[3] Karsten Behrendt and Ryan Soussan. Unsupervised labeled lane markers using maps. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.471, + 0.431 + ], + "angle": 0, + "content": "[5] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[6] Haoxin Chen, Mengmeng Wang, and Yong Liu. Bsnet: Lane detection via draw b-spline curves nearby. arXiv preprint arXiv:2301.06910, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.471, + 0.585 + ], + "angle": 0, + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.471, + 0.642 + ], + "angle": 0, + "content": "[8] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.645, + 0.471, + 0.701 + ], + "angle": 0, + "content": "[9] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[10] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[12] Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. Guiding pretraining in reinforcement learning with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "large language models. arXiv preprint arXiv:2302.06692, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Ruochen Fan, Xuanrun Wang, Qibin Hou, Hanchao Liu, and Tai-Jiang Mu. Spinnet: Spinning convolutional network for lane boundary detection. Computational Visual Media, 5:417-428, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[14] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17062-17070, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[15] Jianhua Han, Xiajun Deng, Xinyue Cai, Zhen Yang, Hang Xu, Chunjing Xu, and Xiaodan Liang. Laneformer: Object-aware row-column transformers for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 799-807, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[17] Hiroto Honda and Yusuke Uchida. Clrernet: Improving confidence of lane detection with laneiou. arXiv preprint arXiv:2305.08366, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17163-17171, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[19] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[20] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[21] Xuan Liao, Wenhao Li, Qisen Xu, Xiangfeng Wang, Bo Jin, Xiaoyun Zhang, Yanfeng Wang, and Ya Zhang. Iteratively-refined interactive 3d medical image segmentation with multi-agent reinforcement learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9394-9402, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[22] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condlanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.787, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[23] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3694-3702, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[26] Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. End-to-end active object tracking via reinforcement learning. In International conference on machine learning, pages 3286-3295. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[27] Stefan Mathe, Aleksis Pirinen, and Cristian Sminchisescu. Reinforcement learning for visual object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2894-2902, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[28] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.362 + ], + "angle": 0, + "content": "[29] Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Jing Li, Shenggao Zhu, Dahua Lin, Chunhua Shen, et al. Spts: single-point text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4272-4281, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[30] André Susano Pinto, Alexander Kolesnikov, Yuge Shi, Lucas Beyer, and Xiaohua Zhai. Tuning computer vision models with task rewards. arXiv preprint arXiv:2302.08242, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[31] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 276-291. Springer, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[32] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.535, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.606, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[35] Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[36] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[37] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[38] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[39] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1392-1401, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[40] Ze Wang, Weiqiang Ren, and Qiang Qiu. Lanenet: Realtime lane detection networks for autonomous driving. arXiv preprint arXiv:1807.01726, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[41] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8:229-256, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[43] Lingyu Xiao, Xiang Li, Sen Yang, and Wankou Yang. Adnet: Lane shape prediction via anchor decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6404-6413, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[44] Jiaxing Yang, Lihe Zhang, and Huchuan Lu. Lane detection with versatile atrousformer and local semantic guidance. Pattern Recognition, 133:109053, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[45] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[46] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[47] Jia-Qi Zhang, Hao-Bin Duan, Jun-Long Chen, Ariel Shamir, and Miao Wang. Houghlanenet: Lane detection with deep hough transform and dynamic convolution. arXiv preprint arXiv:2307.03494, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[48] Zhiyuan Zhao, Qi Wang, and Xuelong Li. Deep reinforcement learning based lane detection and localization. Neurocomputing, 413:328-338, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[49] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3547-3554, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[50] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 898-907, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[51] Kunyang Zhou and Rui Zhou. End to end lane detection with one-to-several transformer. arXiv preprint arXiv:2305.00675, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.869 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16953" + } + ] +] \ No newline at end of file diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_origin.pdf b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f8501500d8279937155fe42d907c4b38f06adb12 --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/0b70847d-c1fa-4821-9d6c-0fc3de9d8952_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bf1716f5a6b7ef659e8c9c74c712efab1c78ca59816463de084e41158f9cb78 +size 1840420 diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/full.md b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28f0183bf81f795190deeaa7890a7e02989282fc --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/full.md @@ -0,0 +1,312 @@ +# Lane2Seq: Towards Unified Lane Detection via Sequence Generation + +# Kunyang Zhou Southeast University + +kunyangzhou@seu.edu.cn + +# Abstract + +In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorporate the task-specific knowledge into Lane2Seq. Experimental results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets $97.95\%$ and $97.42\%$ F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks. + +# 1. Introduction + +Lane detection is a fundamental task in computer vision [31, 36, 40]. It aims to predict the location of the lane in a given image. Lane detection plays a crucial role in many applications, such as adaptive cruise control and lane keeping. Existing lane detection methods generally adopt a divide-and-conquer strategy, which decomposes the lane detection into multiple subtasks. Each subtask is accomplished by a task-specific head network. For instance, as shown in Fig. 1, segmentation-based methods [28,49] adopt a head network along with a task-specific module, such as the message-passing module [28], to predict per-pixel masks. Anchor-based methods [17, 36] utilize a classification head network for distinguishing lane instances and an anchor refinement network for regressing accurate lanes. Parameter-based methods [13, 23] use a network to predict the parameters of lanes and a vertical offset prediction network to locate the start points of lanes. + +Although divide-and-conquer strategy has been proved an effectively way to address the certain subtask in the ex + +![](images/4a37222f67df15e4bb504ee6e2aa97b51868270bd642c8465112acf001dbb588.jpg) +(a) Segmentation-based lane detection + +![](images/2d620d4f44898b74f691c1aeedaee1f466a7f2a97d8c4d7fe101ab9f1459b206.jpg) +(b) Anchor-based lane detection + +![](images/b023bb946345dd46ef4f22d65b7c07a531cbbe7c794f93a21fd611efb8d8244e.jpg) +(c) Parameter-based lane detection + +![](images/66d1487b06eae28eceb22e021472fd336771ba70435b7eef1f4c5b5c0c3b1b40.jpg) +(d) Our sequence generation-based lane detection +Figure 1. Comparison of different lane detection frameworks. + +isting methods, there still exist several limitations. (1) Each subtask needs a customized task-specific head network, resulting in a complicated lane detection model. (2) Each task-specific head requires one or more loss functions, e.g., cross-entropy loss and Line-IOU loss [17], bringing out extra hyper-parameters. + +In this paper, we present a novel sequence generation-based lane detector (Lane2Seq) framework to tackle the aforementioned issues. By formulating the lane detection as a sequence generation task, Lane2Seq gets rid of customized head networks and task-specific loss function. It is based on the intuition that if the detection model knows where the target lane is, model can be simply taught how to read the location of lane out, instead of designing additional classification head or regression head by the divide-and-conquer strategy. + +Therefore, we convert the outputs of different lane detection formats, i.e., segmentation-based, anchor-based, and + +parameter-based, into a sequence of discrete tokens and the model learns to generate this sequence token-by-token. As shown in Fig. 1 (d), to achieve a specific lane detection format, Lane2Seq uses a prompt to specify the detection format and the generated sequence adapts to the prompt so the model can produce format-specific output. By the format-specific prompt, Lane2Seq unifies different lane detection formats into a model. + +While Lane2Seq does not contain task-specific components, task-specific knowledge contained in these components can help the model learn the features of lanes better. We propose a Multi-Format model tuning method based on Reinforcement Learning (MFRL) to incorporate the task-specific knowledge into the model without changing model's architecture. Inspired by Task-Reward [30], MFRL takes the evaluation metrics, which naturally integrates task-specific knowledge, as the reward and tunes Lane2Seq using REINFORCE [41] algorithm. However, evaluation metrics like F1 score cannot be used as the reward directly due to undecomposable as a sum of perexample rewards. In this paper, we propose three new evaluation metric-based rewards for segmentation, anchor, and parameter format, based on their task-specific knowledge. + +Experimental results demonstrate that our Lane2Seq achieves competitive performance on three public datasets, Tusimple, CULane, and LLAMAS. For instance, Lane2Seq with ViT-Base encoder obtains $97.95\%$ and $97.42\%$ F1 score on Tusimple and LLAMAS, setting a new state-of-the-art result for two datasets. It should be noted that all existing lane detection methods heavily rely on well-designed task-specific head networks and the corresponding complicated loss functions. Instead, our Lane2Seq utilizes a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. + +The main contributions of this paper are as follows. + +- We propose a sequence generation-based method for lane detection, which casts the lane detection as a sequence generation task. To the best of our knowledge, we are the first to unify the lane detection through the sequence generation, which offers a new perspective on lane detection. +- We present a novel reinforcement learning-based multi-format model tuning, including three new evaluation metric-based reward functions, to incorporate the task-specific knowledge into the model. +- Experimental results show that our method achieves competitive performance on lane detection benchmarks. Remarkably, we establish a new state-of-the-art result on Tusimple and LLAMAS. + +# 2. Related Work + +Lane detection. Existing lane detection methods can be divided into three categories based on the representation of the lane: segmentation-based method, anchor-based method and parameter-based method. Segmentation-based methods [28, 47, 49] consider lane detection as a semantic segmentation task and performs pixel-wise prediction. SCNN [28] enhances the visual evidence by a message-passing module, which can capture spatial dependency for lanes. Anchor-based methods [20, 31, 36, 51] predict the accurate lanes by refining the predefined lane anchors. UFLD [31] proposes a novel row anchor-based approach to detect lanes. Different from the segmentation-based and anchor-based methods, parameter-based methods [13, 23, 37] regard the lane detection as the parametric modeling and regress the parameters of the lane. PolyLaneNet [37] models a lane curve as a polynomial and regresses the parameters of the polynomial. In this paper, we treat the segmentation-based, anchor-based methods, and parameter-based metohods uniformly as the sequence generation task, getting rid of the complicate structure and task-specific modules, e.g., head network. It only adopts the cross-entropy loss and plain transformer architecture. + +Sequence generation for vision tasks. Recently, sequence-to-sequence (seq2seq) method used in the natural language processing (NLP) has been applied for vision tasks. Seq2seq utilizes a basic transformer encoder and decoder architecture and accomplishes the task by making sequence predictions, rather than designing a model tailored to the vision task. Pix2seq [8] is the pioneering work to cast the object detection as the sequence generation task. It shows that object can be detected well without any task-specific modules like label assignment. Besides object detection, seq2seq has been extended to other vision tasks such as instance segmentation [9], keypoint detection [9], text spotting [19,29] and object tracking [10]. UniTAB [45] adopts the task prompt [33] to perform multi-task learning. Unified-IO [25] jointly trains various vision tasks by setting the unified input/output formats for all tasks. Besides, seq2seq becomes increasingly popular in the multimodal model. Text-to-image models like DALL-E [34] and vision-language models like Flamingo [2] all use seq2seq to unify the multi-modal tasks. Nevertheless, how to perform the seq2seq in the lane detection to unify different detection formats is still unexplored. + +Lane2Seq has a similar spirit with the Pix2seq and its successors [9, 10, 19, 29]. All of them view the vision tasks as the sequence generation. The main difference between these methods and Lane2Seq is the sequence format. Previous works use coordinates and category to construct the sequence, while Lane2Seq also adopts the parameter of the lane to construct the sequence. + +Reinforcement learning in the computer vision. Many + +![](images/63e1195f0cba21dbe15638dac9d121f95f6aaf9a247c799564b18d7078d670c5.jpg) +Figure 2. Inference pipeline of Lane2Seq. The model perceives input image, prompt and generates format-specific tokens, which can be detokenized into required detection format for visualization. + +previously researches introduce the reinforcement learning for the vision tasks, such as object detection [27], object tracking [26], image segmentation [21] and lane detection [48]. They generally focus on learning different parts of an image and refine the outputs iteratively. DQLL [48] localizes the lanes as a group of lanemarks and refine the location of lanemarks with the deep Q-learning. Recently, Task Reward [30] adopts a novel task risk-based reinforcement learning to tune the vision model without changing the model architecture. Reward functions in Task-Reward are designed for object detection, instance segmentation, colorization, and image captioning but not for lane detection. Our MFRL is the first attempt to design effective evaluation metric-based reward functions for different lane detection formats. + +# 3. Method + +In this section, we present the proposed sequence generation-based lane detection method, named Lane2Seq, in detail. The overall pipeline of Lane2Seq is illustrated in Fig. 2. Sec 3.1 depicts the unified interface for lane detection, Sec 3.2 details the model architecture and objective function, and Sec 3.3 introduces the proposed multi-format model tuning based on reinforcement learning. + +# 3.1. Unified Interface for Lane Detection + +As presented in Figure 1, lane detection formats in the existing method are diverse and have been formulated quite differently. Considering differences in the form of outputs, customized models with specialized head networks and loss + +functions are designed for different lane detection formats. + +To integrate different lane detection formats into a model, we propose an unified sequence interface for lane detection, where both format transcription like segmentation and outputs are treated as sequences of discrete tokens. As shown in Fig. 2, the generated sequence is composed of four parts, a starting token $\langle \text{starting} \rangle$ , a format transcription token like $\langle \text{Segmentation} \rangle$ , a format-specific sequence, and an ending token $\langle \text{end} \rangle$ . The format-specific sequence for three detection formats can be constructed as follows. + +Segmentation sequence. Instead of performing pixelwise mask prediction, we predict the polygon [5] corresponding to the mask as a sequence of coordinates conditioned on a given lane instance. Then, we convert the polygon into a sequence by quantizing the coordinates of its points into discrete tokens. Specifically, $x, y$ coordinates of a point are normalized to the width and height of the image, and then quantized to $[1, n_{\text{bins}}]$ , where $n_{\text{bins}}$ is the size of the vocabulary. Vocabulary will be described subsequently. A polygon sequence can be expressed as $[x_1, y_1, x_2, y_2, \ldots, x_{28}, y_{28}, <\text{Lane}>]$ , where $<\text{Lane}>$ is the category token. If there are multiple lanes in an image, we concatenate the all polygon sequences. The segmentation sequence consists of a starting point and all polygon sequences, where starting point is the left-top point of the image $x_0 = 0, y_0 = 0$ . + +Anchor sequence. Since the essence of anchor-based methods is regressing the location of keypoints of lane anchors, we treat the anchor prediction as the keypoint sequence generation. Specifically, the + +![](images/e09fe164266f283dbbe79453cf1e3756568de1f8c1d134a19d0d6fffaa811609.jpg) +Figure 3. Training pipeline of Lane2Seq. The input sequence can be the segmentation sequence or anchor sequence or parameter sequence. + +keypoint sequence of a lane can be expressed as \([x_1, y_1, x_2, y_2, \ldots, x_{14}, y_{14}, <\text{Lane}>\]\. The normalization and quantization of keypoint coordinates are same to that of polygon point. The anchor sequence contains a starting point and all keypoint sequences, where starting point is still the left-top point of the image. + +Parameter sequence. the sequence of the lane includes two parts: the parameter of polynomial function and vertical offsets. We set the polynomial degree as 5th. Specifically, the parameter sequence of a lane can be represented as $[a_1,a_2,a_3,a_4,a_5,s,\langle \mathrm{Lane}\rangle ]$ , where $s$ is the vertical offset. $s$ is normalized to the height of the image and quantized to $[1,n_{\mathrm{bins}}]$ . We use sigmoid to normalize $a_i$ by $a_{i} = \operatorname {Sigmoid}(a_{i})$ . Then $a_{i}$ is quantized to $[1,n_{\mathrm{bins}}]$ . Parameter sequence does not require the starting point. + +Vocabulary. We use a shared vocabulary for all formats and each integer between $[1, n_{bins}]$ can be regarded as a word in the vocabulary. Each word in the vocabulary corresponds to a learnable embedding. + +# 3.2. Unified Architecture and Objective Function + +We follow [10] and use a transformer-based encoder-decoder architecture to deal with image input and sequence output flexibly. As shown in Fig. 3, the image encoder takes the pixels as the input and outputs the corresponding image representations. We utilize the Vision Transformer (ViT) [11] to instantiate the image encoder. We adopt a transformer-based sequence decoder, which is widely used in the language modeling [32, 33], to generate the output sequence. The decoder generates one token at a time based on the preceding tokens and the image representations. Sequence decoder removes the well-designed task-specific heads for different detection formats. + +Training Pipeline. We first tokenize annotation of each format into the corresponding sequence. Then, we construct a training batch using images and sequences from all detection formats. Finally, we compute loss between generated sequence and target sequence for each format. For a certain detection format, the input sequence is $[<\text{starting}>$ , format transcription token, format-specific sequence] and the target sequence is [format transcription token, format-specific sequence, $\langle \mathrm{end}\rangle$ ]. For example, in the anchor format, the in + +put sequence is $[<\text{starting}>, <\text{Anchor}>, x_0, y_0, x_1, y_1, \ldots, <\text{Lane}>, <\text{end}>]$ and the target sequence is $[<\text{Anchor}>, x_0, y_0, x_1, y_1, \ldots, <\text{Lane}>, <\text{end}>]$ . The starting token $<\text{starting}>$ and ending token $<\text{end}>$ are learnable embeddings, which tell the decoder when to begin and end the sequence generation. The sequence decoder perceives the image representations and input sequence and reconstructs the target sequence. + +Objective Function. Similar to Pix2seq [8], we train the Lane2Seq with a simple cross-entropy loss. At each time stamp $j$ , Lane2Seq aims to maximize the likelihood of the target tokens conditioned on the image representations $I$ and previously generated tokens $y_{1:j-1}$ , + +$$ +L o s s _ {o b j} = - \sum_ {j = 1} ^ {N} w _ {j} \log P (\hat {m} _ {j} | I, m _ {1: j - 1}), \tag {1} +$$ + +where $m$ and $\hat{m}$ denote the input and target sequence, respectively. $N$ is the length of sequence. $w_{j}$ represents the weight for the $j$ -th token, 0 is assigned to the format transcription token and 1 is assigned to other tokens to ensure the model is trained to predict the desired tokens but not the format transcription tokens. + +Inference Pipeline. We take the format transcription to-ken as the prompt to achieve the format-specific detection and the sequence decoder generates the rest of the sequence conditioned on the prompt and image representation. Once the whole sequence is generated, we de-quantize the sequence to obtain the location of lane. More details about de-quantization for each format can be found in the supplementary materials. Inference process is presented in Fig. 2. + +# 3.3. Multi-Format Model Tuning Based on Reinforcement Learning + +Lane2Seq get rids of the complicated task-specific components, such as task-specific head networks, via the sequence generation. However, this task-agnostic architecture inevitably lacks the task-specific knowledge of the lane detection, making the model ineffective in learning the features of lane. Existing lane detection methods [17, 49] usually design the specific modules to incorporate the task-specific knowledge into the model. But this way makes the model architecture complicate. We propose a novel Multi-Format model tuning method based on Reinforcement Learning (MFRL) to learn task-specific knowledge effectively without changing any components of the model. + +Inspired by Task-Reward [30], MFRL takes the evaluation metric, which naturally contains the task-specific knowledge, as the reward function and adopts the reinforcement learning to tune the model, which is common practice in the modern language model [4, 12]. Specifically, it is difficult for the model to converge if model is trained with reinforcement learning method from scratch [30]. Hence, MFRL consists of two stages: the pretraining stage and + +the model tuning stage. In the pretraining stage, model is trained on the lane detection dataset with objective function in Eq. 1 to have a good weight initialization. + +In the model tuning stage, we use the REINFORCE algorithm [35] to maximize the objective function as below, + +$$ +O b j = E _ {c \sim D} \left[ E _ {t \sim Q (\cdot | c, u)} R (t, g) \right], \tag {2} +$$ + +where $c, t,$ and $g$ represent the input image, generated format-specific sequences, and ground truths, respectively. $E$ and $u$ denote the mathematic expectation and the model parameter. $R, D,$ and $Q$ stand for the reward function, data distribution of the dataset, and conditional distribution parameterized by $u$ . REINFORCE algorithm estimates the gradient of the reward function by + +$$ +\nabla_ {u} E _ {t \sim Q} [ R (t, g) ] = E _ {t \sim Q} [ R (t, g) \nabla_ {u} \log Q (t | c, u) ], \tag {3} +$$ + +In practice, Eq. 3 is computed as an average of per-sequence gradient and reward function of per-sequence is not required to be differentiable. More details about REINFORCE algorithm can be found in the supplementary materials. + +Reward Function. Evaluation metric F1 score and accuracy are unable to adopt as the reward directly, because they cannot be decomposable as a sum of per-sequence rewards (see Sec 4.2). However, F1 score or accuracy is composed of false positives $(FP)$ , true positives $(TP)$ , and false negatives $(FN)$ . All three indicators can be computed per-sequence. We opt to use $TP$ and $FP$ to design the reward. Specifically, reward for three detection formats are constructed as follows. + +Reward for the segmentation format (termed as $R_{seg}$ ). Segmentation-based lane detection includes two task-specific knowledge, i.e., knowledge of the segmentation and lane detection. For the knowledge of lane detection, we adopt a matched Line-IOU (LIOU) [50], which can demonstrate the quality of location prediction and shape prediction of the lane. For the knowledge of the segmentation, we adopt the matched Mean Intersection over Union (mIOU), which is a widely used metric in the semantic segmentation to evaluate the quality of segmentation, as the reward. Finally, we add both of them to compute $r_{seg}$ as below, + +$$ +\begin{array}{l} R _ {s e g} (t, g) = \frac {1}{K} \sum_ {k = 1} ^ {K} \left[ L I O U \left(p _ {k}, g\right) + m I O U \left(p _ {k}, g\right) \right] \tag {4} \\ - \lambda_ {1} F P _ {s e g} (t, g), \\ \end{array} +$$ + +where $K$ and $p_k$ represents the number of true positives in $t$ and the $k$ -th true positive. $FP_{seg}$ is the false positive rate of the segmentation format and $\lambda_1$ is the weight to control the affect of $FP_{seg}$ . We introduce $FP_{seg}$ to penalize for false positives. + +Reward for the anchor format (termed as $R_{a}$ ). Anchors-based lane detection contains the knowledge of the keypoint location and lane detection. For the knowledge of keypoint + +location, we simply adopt the matched Euclidean distance $d(p_k, g)$ between true positives and corresponding ground truths as the reward. Since MFRL needs to maximize the reward, we rescale the reward as $d_r(p_k, g) = 1 - \frac{d(p_k, g)}{H}$ , where $H$ is the height of image. For the knowledge of lane detection, we also utilize the matched LIOU as the reward. The formulation of $r_a$ is: + +$$ +R _ {a} (t, g) = \frac {1}{K} \sum_ {k = 1} ^ {K} \left[ L I O U \left(p _ {k}, g\right) + d _ {r} \left(p _ {k}, g\right) \right] - \lambda_ {2} F P _ {a} (t, g), \tag {5} +$$ + +where $FP_{a}$ and $\lambda_{2}$ stand for the false positives rate of the anchor format and the weight of $FP_{a}$ . + +Reward for the parameter format (termed as $R_p$ ). Parameter-based detection only has the knowledge of lane detection, hence we directly use the matched LIOU as the reward. + +$$ +R _ {p} (t, g) = \frac {1}{K} \sum_ {k = 1} ^ {K} L I O U \left(p _ {k}, g\right) - \lambda_ {3} F P _ {p} (t, g), \tag {6} +$$ + +where $FP_{p}$ and $\lambda_{3}$ stand for the false positives rate of the parameter format and the weight of $FP_{p}$ . + +Since different detection formats have different contributions, we weight the objective function of different formats. The final objective function of MFRL can be computed as, + +$$ +O b j _ {t o t a l} = \lambda_ {4} O b j _ {s e g} + \lambda_ {5} O b j _ {a} + \lambda_ {6} O b j _ {p}, \tag {7} +$$ + +where $Obj_{seg}$ , $Obj_a$ , and $Obj_p$ denote the objective function of segmentation, anchor, and parameter format, respectively. The objective function of three formats is the same as Eq. 2, except the format-specific sequence and reward function. $\lambda_4$ , $\lambda_5$ , and $\lambda_6$ are the scale factor. + +# 4. Experiments + +# 4.1. Datasets + +We conduct experiments on three lane detection benchmarks: CULane [28], Tusimple [1], and LLAMAS [3]. + +CULane is a widely used large-scale dataset for lane detection. It contains a lot of challenging scenarios such as crowded roads. The CULane dataset consists of 88.9K images for training, 9.7K images in the validation set, and 34.7K images for the test. Image size is $1640 \times 590$ . + +Tusimple is a real highway dataset consisting of 3,626 training images and 2,782 testing images. All images have $1280 \times 720$ pixels. + +LLAMAS is a recently released large-scale lane detection dataset with over 100K images. All lane markers are annotated with high-accurate maps. Image size is $1280 \times 717$ . + +# 4.2. Evaluation Metrics + +For CULane and LLAMAS dataset, we adopt the F1 score to measure the performance: $F_{1} =$ + +Table 1. Comparison of F1 score and MACs (multiply-accumulate operations) on CULane testing set. We only report the false positives for "Cross" category following [50]. FPS is tested on a single 2080Ti. + +
MethodsEncoderFPS↑Normal↑Crowded↑Dazzle↑Shadow↑No line↑Arrow↑Curve↑Night↑Cross↓Total↑MACs(G)↓
Segmentation based
SCNN [28]ResNet501290.6069.7058.5066.9043.4084.1064.4066.10190071.608.4
RESA [49]ResNet5042.192.1073.1069.2072.8047.7088.3070.3069.90150375.37.6
AtrousFormer [44]ResNet34-92.8375.9669.4877.8650.1588.6671.1473.74105478.0814.22
LaneAF [44]DLA34 [46]-91.8075.6171.7879.1251.3886.8872.7073.03136077.4112.73
Lane2Seq (segmentation)ViT-Base1993.3977.2773.4579.6953.9190.5373.3774.96112979.6415.19
Anchor based
LaneATT [36]ResNet1222391.7476.1669.4776.3150.4686.2964.0570.81126477.0270.5
O2SFformer [51]ResNet504493.0976.5772.2576.5652.8089.5069.6073.85311877.8327.51
UFLD [31]ResNet3419290.7070.2059.5069.3044.4085.7069.5066.70203772.303.3
CondLaneNet [22]ResNet3413593.3877.1471.1779.9351.8589.8973.8873.92138778.7419.6
ADNet [43]ResNet34-92.9077.4571.7179.1152.8989.9070.6474.78149978.9418.6
CLRNet [50]ResNet3411293.4978.0674.5779.9254.0190.5972.7775.02121679.7317.3
Lane2Seq (anchor)ViT-Base2393.1177.4373.2579.4653.7490.0272.4475.12117379.2715.19
Parameter based
LSTR [23]ResNet18----------64.005.9
BezierLaneNet [14]ResNet18-90.2271.5562.4970.9145.3084.0958.9868.7099673.674.7
BSNet [6]ResNet34-93.7578.0176.6579.5554.6990.7273.9975.28144579.8917.2
Eigenlanes [18]ResNet50-91.7076.0069.8074.1052.2087.7062.9071.80150977.2018.7
Laneformer [15]ResNet50-91.7775.7470.1775.7548.7387.6566.3371.041977.0626.2
Lane2Seq (parameter)ViT-Base3093.0376.4272.1778.3252.8989.6772.6773.98131978.3915.19
+ +$\frac{2 \times \text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}}$ , where Precision = TP/TP+FP and Recall = TP/TP+FN. + +For Tusimple dataset, we use F1 score, accuracy, false positives, and false negatives to evaluate the model performance. Accuracy is defined as Accuracy = $\frac{\sum_{clip} C_{clip}}{\sum_{clip} S_{clip}}$ , where $C_{clip}$ represents the number of accurately predicted lane points and $S_{clip}$ denotes the total number of lane points of a clip. A lane point is considered as correct if its distance is smaller than the given threshold $t_{pc} = \frac{20}{\cos(a_{yl})}$ , where $a_{yl}$ denotes the angle of the corresponding ground truth. + +# 4.3. Implementation Details + +Model. We adopt the ViT-Base initialized with the MAE [16] pretrained parameters as the image encoder. The patch size is $16 \times 16$ . The decoder is composed of 2 transformer blocks and its hidden size is 256. The number of attention heads is 8 and the hidden size of feed-forward network is 1024. $n_{bins}$ is set to 1000 and the dimension of word embedding of the vocabulary is 256. Our Lane2Seq is implemented based on MMDetection toolbox [7]. $\lambda_1, \lambda_2, \lambda_3, \lambda_4, \lambda_5$ , and $\lambda_6$ are set to 0.3, 0.3, 0.1, 0.2, 1, and 1.5, respectively. + +Training. All input images are resized to $320 \times 800$ . We utilize the AdamW [24] as the optimizer with initial learning rate 1e-4. The training epochs of the pretraining stage are 5, 20, and 15 for CULane, Tusimple, and LLAMAS. We set the training epochs of the model tuning stage to 15, 30, and 55 for CULane, Tusimple, and LLAMAS, respectively. For the data augmentation, we use the random horizontal flips and random affine transformation including scaling, rotation, and translation. All experiments are conducted on 8 A100 GPUs with total batch size 384. + +Table 2. Comparison of the performance of different models on LLAMAS. + +
MethodsEncoderF1(%)↑
PolyLaneNet [37]EfficientNet-b0 [38]90.20
BezierLaneNetResNet3496.11
LaneATTResNet3494.96
LaneATTResNet12295.17
LaneAFDLA3496.90
CLRNetResNet1896.96
CLRNetDLA3497.16
Lane2Seq (segmentation)ViT-Base97.42
Lane2Seq (anchor)ViT-Base97.05
Lane2Seq (parameter)ViT-Base96.74
+ +# 4.4. Comparison with the State-of-the-art Methods + +Performance on LLAMAS. Our Lane2Seq achieves a new state-of-the-art performance on LLAMAS dataset using segmentation format. In Table 2, Lane2Seq increases F1 score from $97.16\%$ to $97.42\%$ compare to the previous state-of-the-art method CLRNet. Compared with parameter-based state-of-the-art method BezierLaneNet, Lane2Seq performs better than BezierLaneNet (96.74% vs. $96.11\%$ ). The results manifest the detection strengths of Lane2Seq in the multi-lane scenario (The number of lanes $\geq 5$ ), such as highways. The reason may be that transformer-based architecture has the advantages in capturing long-range dependency. + +Performance on Tusimple. Table 3 presents the performance comparison with state-of-the-art approaches on Tusimple. Due to small data scale and single scene, the performance gap between different models is small. Our Lane2Seq using segmentation format sets a new state-of-the-art F1 score of $97.95\%$ . The results also validate the detection strengths of Lane2Seq in multi-lane scenario. + +Table 3. Comparison of the performance of different models on Tusimple. Acc denotes accuracy. + +
MethodsEncoderF1(%)↑Acc(%)↑FP(%)↓FN(%)↓
SCNNVGG1695.9796.536.171.80
RESAResNet3496.9396.823.632.48
PolyLaneNetEfficientNet-bo90.6293.369.429.33
UFLDResNet3488.0295.8618.913.75
LaneATTResNet12296.0696.105.642.17
GANet [39]ResNet3497.6895.871.992.64
CondLaneNetResNet3496.9895.372.203.82
CondLaneNetResNet10197.2496.542.013.50
CLRNetResNet3497.8296.872.272.08
CLRNetResNet10197.6296.832.372.38
Lane2Seq (segmentation)ViT-Base97.9596.852.012.03
Lane2Seq (anchor)ViT-Base97.8696.722.212.05
Lane2Seq (parameter)ViT-Base96.5996.002.233.54
+ +Performance on CULane. We compare Lane2Seq with other state-of-the-art methods on CULane and results are shown in Table 1. For the segmentation-based methods, Lane2Seq improves F1 score from $78.08\%$ to $79.64\%$ compared to the previous state-of-the-art methods AtrousFormer. Compared with anchor-based methods, Lane2Seq outperforms most previous methods. For example, Lane2Seq achieves the better performance than existing row anchor-based methods, such as CondLaneNet (79.27% vs. $78.74\%$ ). Moreover, Lane2Seq achieves competitive performance compared to CLRNet (79.27% vs. $79.73\%$ ). One possible reason for the performance gap between CLRNet and Lane2Seq is that CLRNet adopts multi-scale detection, which can notably improve the detection performance, while Lane2Seq only uses the plain transformer architecture with single-scale detection. For the comparison of parameter-based methods, Lane2Seq surpasses the existing transformer-based methods. For example, Lane2Seq significantly increases F1 score from $64.00\%$ to $78.39\%$ compared with LSTR and outperforms Laneformer by $1.33\%$ (78.39% vs. $77.06\%$ ). But Lane2Seq is ranked second after BSNet. As for the inference speed, we can see that Lane2Seq does not have a significant advantage in inference speed. We attribute that token-by-token prediction method slows the inference speed. + +Based on the above results and analyses, we can conclude that the unified lane detection via the sequence generation without any well-designed task-specific components, combining our reinforcement learning-based multi-format model tuning, can achieve promising performance. + +Qualitative Results. We display the qualitative results in Fig. 4. The results show that Lane2Seq can effectively detect lanes in the multi-lane scenario (see Fig 4 (a) and (c)). Even in the case of the night scene, Lane2Seq successfully discriminates the lanes (see Fig 4 (b)). + +# 4.5. Ablation Study + +We conduct the ablation experiments on CULane dataset to validate the effectiveness of each component. More ablation studies are in the supplementary materials. + +Table 4. Ablation study on the multi-format model tuning based on reinforcement learning. + +
MethodsF1(%)↑Precision(%)↑Recall(%)↑
Lane2Seq (segmentation)64.2770.4452.77
Lane2Seq (segmentation) + MFRL79.64 (+15.37)85.26 (+14.82)69.00 (+16.23)
Lane2Seq (anchor)66.2371.3655.75
Lane2Seq (anchor) + MFRL79.27 (+13.04)84.72 (+13.36)67.28 (+11.53)
Lane2Seq (parameter)68.1473.2757.89
Lane2Seq (parameter) + MFRL78.39 (+10.25)83.68 (+10.41)66.94 (+9.05)
+ +Table 5. Ablation study on the different reward functions. F1 score is adopted as the indicator. + +
\( R_{seg} \)\( R_a \)\( R_p \)SegmentationAnchorParameter
64.2766.2368.14
✓77.69 (+13.42)66.23 (+0.0)68.14 (+0.0)
✓64.27 (+0.0)76.94 (+10.71)68.14 (+0.0)
✓64.27 (+0.0)66.23 (+0.0)77.02 (+8.88)
✓✓78.72 (+14.45)78.14 (+11.91)68.14 (+0.0)
✓✓78.70 (+14.43)66.23 (+0.0)77.90 (+9.76)
✓✓64.27 (+0.0)78.20 (+11.97)77.93 (+9.79)
✓✓✓79.64 (+15.37)79.27 (+13.04)78.39 (+10.25)
+ +Multi-Format Model Tuning Based on Reinforcement Learning. We first ablate the effectiveness of the proposed Multi-Format Model Tuning Based on Reinforcement Learning (MFRL). As shown in Table 4, MFRL significantly improves the performance of different detection formats. For the segmentation format, MFRL gains $15.37\%$ F1 score improvement $(79.64\%$ vs. $64.27\%)$ . For the anchor and parameter format, MFRL increases $13.04\%$ $(79.27\%$ vs. $66.23\%)$ and $10.25\%$ $(78.39\%$ vs. $68.14\%)$ F1 score, respectively. Reasons of performance improvement can be attributed in two folds: (1) the vanilla Lane2Seq is a task-agnostic architecture and lacks the task-specific knowledge, leading to the ineffective feature learning of lane. MFRL tackles this problem by adopting the evaluation metric, which is designed based on the lane prior information like angle, as the reward to tune the model. (2) MFRL makes the model's predictions aligned with their intended usage, i.e., how to achieve a high performance. Optimizing the loss function is a common practice in the compute vision. However, this way indirectly optimizes for model's intended usage, because the lower loss does not mean the higher performance. Instead, MFRL views the maximizing the evaluation metric as the optimization objective, which is positively correlated with the model performance. + +Effectiveness of Reward Function of Different Formats. We ablate the effectiveness of reward function of different formats and results are shwon in Table 5. It can be observed that our reward function can bring consistent performance improvement. However, we can also see that the performance will not be improved if the corresponding reward is not optimized. Besides, we can find that multi-reward optimization can achieves better results than single-reward optimization. Reason may be that multi-reward optimiza + +![](images/f024bdb0778d6a88ea6226dcb7224b0a91ff2310b956d59788a39416b9595492.jpg) +Figure 4. Visualization results of LaneATT, CLRNet, and Lane2Seq on Tusimple, CULane, and LLAMAS. + +Table 6. Comparison between Lane2Seq and YOLOP. + +
MethodsF1(%)↑Precision(%)↑Recall(%)↑
YOLOP (segmentation)75.2379.4660.58
Lane2Seq (segmentation)79.64 (+4.41)85.26 (+5.8)69.00 (+8.42)
YOLOP (anchor)76.4480.9962.15
Lane2Seq (anchor)79.27 (+2.83)84.72 (+3.73)67.28 (+5.13)
YOLOP (parameter)69.4374.3858.46
Lane2Seq (parameter)78.39 (+8.96)83.68 (+9.30)66.94 (+8.48)
+ +tion enables model to learn correlations between different detection formats and provides more supervised signals. + +Comparison with other Multi-Task Methods. An alternative method to unify different detection formats is using different detection heads in a model. YOLOP [42] is a representative work for this method. We replace the heads of YOLOP with anchor-based head, segmentation-based head, and parameter-based head. The segmentation-based loss, anchor-based loss, parameter-based loss for YOLOP are same as that in SCNN [28], LaneATT [36], and PolyLaneNet [37]. From Table 6, we can observe that Lane2Seq surpasses YOLOP for all detection formats. It should be noted that YOLOP contains task-specific modules, such as detection head and loss function, but Lane2Seq does not. Results manifest the sequence generation combining with MFRL is a more simple and effective way to unify lane detection. + +Ablation Study on the Scale Factor of Different Objective Functions. For the limitation of the space, we take the segmentation format as an example to present the influence of the scale factor of different objective functions in MFRL. Results of other two formats can be found in the supplementary materials. As shown in Table 7, assigning the same factor to different objective functions causes insuperior performance, indicating different detection formats have different contributions. We set $\lambda_4$ , $\lambda_5$ , and $\lambda_6$ to 0.2, 1, and 1.5 according to the model performance. + +# 5. Conclusion + +This paper presents a novel sequence generation-based lane detection framework, i.e., Lane2Seq, to unify lane de + +Table 7. Ablation study on different scale factors. + +
λ4λ5λ6F1(%)↑
11173.23
0.71174.96
0.41176.27
0.21177.35
0.11177.02
0.21.1177.03
0.21.2176.75
0.20.9177.09
0.20.8177.00
0.210.976.87
0.211.277.37
0.211.479.46
0.211.579.64
0.211.679.39
+ +tection, which casts lane detection as a sequence generation task. Lane2Seq gets rid of complicated task-specific modules and adopts a simple transformer-based encoder-decoder architecture. To incorporate the task-specific knowledge, we employ a multi-format model tuning based on reinforcement learning (MFRL). Extensive experiments show Lane2Seq is effective and achieves competitive results compared to the state-of-the-art methods. + +Limitation. One major limitation of Lane2Seq is that sequence generation model is expensive for long sequence (mainly for inference). The inference speed is low when there are more than 5 lanes in an image despite its detection strength in multi-lane scenario. Therefore, future work is required to make it faster for real-time lane detection applications. Another limitation is that MFRL can only be applied to sequence generation-based model currently. Our ongoing work is applying MFRL to other vision models. + +# References + +[1] Tusimple dataset. https://github.com/TuSimple/tusimple-benchmark. Accessed on 11th August 2023. 5 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2 +[3] Karsten Behrendt and Ryan Soussan. Unsupervised labeled lane markers using maps. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 5 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4 +[5] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 3 +[6] Haoxin Chen, Mengmeng Wang, and Yong Liu. Bsnet: Lane detection via draw b-spline curves nearby. arXiv preprint arXiv:2301.06910, 2023. 6 +[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6 +[8] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 2, 4 +[9] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 2 +[10] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 2, 4 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4 +[12] Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. Guiding pretraining in reinforcement learning with + +large language models. arXiv preprint arXiv:2302.06692, 2023.4 +[13] Ruochen Fan, Xuanrun Wang, Qibin Hou, Hanchao Liu, and Tai-Jiang Mu. Spinnet: Spinning convolutional network for lane boundary detection. Computational Visual Media, 5:417-428, 2019. 1, 2 +[14] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17062-17070, 2022. 6 +[15] Jianhua Han, Xiajun Deng, Xinyue Cai, Zhen Yang, Hang Xu, Chunjing Xu, and Xiaodan Liang. Laneformer: Object-aware row-column transformers for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 799-807, 2022. 6 +[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 6 +[17] Hiroto Honda and Yusuke Uchida. Clrernet: Improving confidence of lane detection with laneiou. arXiv preprint arXiv:2305.08366, 2023. 1, 4 +[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17163-17171, 2022. 6 +[19] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2 +[20] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019. 2 +[21] Xuan Liao, Wenhao Li, Qisen Xu, Xiangfeng Wang, Bo Jin, Xiaoyun Zhang, Yanfeng Wang, and Ya Zhang. Iteratively-refined interactive 3d medical image segmentation with multi-agent reinforcement learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9394-9402, 2020. 3 +[22] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condlanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021. 6 +[23] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3694-3702, 2021. 1, 2, 6 +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[25] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified + +model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022. 2 +[26] Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. End-to-end active object tracking via reinforcement learning. In International conference on machine learning, pages 3286-3295. PMLR, 2018. 3 +[27] Stefan Mathe, Aleksis Pirinen, and Cristian Sminchisescu. Reinforcement learning for visual object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2894-2902, 2016. 3 +[28] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5, 6, 8 +[29] Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Jing Li, Shenggao Zhu, Dahua Lin, Chunhua Shen, et al. Spts: single-point text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4272-4281, 2022. 2 +[30] André Susano Pinto, Alexander Kolesnikov, Yuge Shi, Lucas Beyer, and Xiaohua Zhai. Tuning computer vision models with task rewards. arXiv preprint arXiv:2302.08242, 2023. 2, 3, 4 +[31] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 276-291. Springer, 2020. 1, 2, 6 +[32] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4 +[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 4 +[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[35] Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12, 1999. 5 +[36] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021. 1, 2, 6, 8 +[37] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. 2, 6, 8 + +[38] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6 +[39] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1392-1401, 2022. 7 +[40] Ze Wang, Weiqiang Ren, and Qiang Qiu. Lanenet: Realtime lane detection networks for autonomous driving. arXiv preprint arXiv:1807.01726, 2018. 1 +[41] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8:229-256, 1992. 2 +[42] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 8 +[43] Lingyu Xiao, Xiang Li, Sen Yang, and Wankou Yang. Adnet: Lane shape prediction via anchor decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6404-6413, 2023. 6 +[44] Jiaxing Yang, Lihe Zhang, and Huchuan Lu. Lane detection with versatile atrousformer and local semantic guidance. Pattern Recognition, 133:109053, 2023. 6 +[45] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2 +[46] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 6 +[47] Jia-Qi Zhang, Hao-Bin Duan, Jun-Long Chen, Ariel Shamir, and Miao Wang. Houghlanenet: Lane detection with deep hough transform and dynamic convolution. arXiv preprint arXiv:2307.03494, 2023. 2 +[48] Zhiyuan Zhao, Qi Wang, and Xuelong Li. Deep reinforcement learning based lane detection and localization. Neurocomputing, 413:328-338, 2020. 3 +[49] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3547-3554, 2021. 1, 2, 4, 6 +[50] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 898-907, 2022. 5, 6 +[51] Kunyang Zhou and Rui Zhou. End to end lane detection with one-to-several transformer. arXiv preprint arXiv:2305.00675, 2023. 2, 6 \ No newline at end of file diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/images.zip b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b7baf8dae9d00f412eeed23c7313c541ebe29a42 --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d6a9d0330a125beb1f8849c7ce6f41a8e5a85a79b94b56a0d7b21846830d12d +size 523390 diff --git a/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/layout.json b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5755f25802356d7c9f0b738315c4846839ad2445 --- /dev/null +++ b/2024/Lane2Seq_ Towards Unified Lane Detection via Sequence Generation/layout.json @@ -0,0 +1,8951 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 510, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 510, + 120 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 510, + 120 + ], + "type": "text", + "content": "Lane2Seq: Towards Unified Lane Detection via Sequence Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 244, + 144, + 349, + 172 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 144, + 349, + 172 + ], + "spans": [ + { + "bbox": [ + 244, + 144, + 349, + 172 + ], + "type": "text", + "content": "Kunyang Zhou Southeast University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 236, + 174, + 357, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 174, + 357, + 184 + ], + "spans": [ + { + "bbox": [ + 236, + 174, + 357, + 184 + ], + "type": "text", + "content": "kunyangzhou@seu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": "In this paper, we present a novel sequence generation-based framework for lane detection, called Lane2Seq. It unifies various lane detection formats by casting lane detection as a sequence generation task. This is different from previous lane detection methods, which depend on well-designed task-specific head networks and corresponding loss functions. Lane2Seq only adopts a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss. Additionally, we propose a new multi-format model tuning based on reinforcement learning to incorporate the task-specific knowledge into Lane2Seq. Experimental results demonstrate that such a simple sequence generation paradigm not only unifies lane detection but also achieves competitive performance on benchmarks. For example, Lane2Seq gets " + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "inline_equation", + "content": "97.95\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "inline_equation", + "content": "97.42\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 441 + ], + "type": "text", + "content": " F1 score on Tusimple and LLAMAS datasets, establishing a new state-of-the-art result for two benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "spans": [ + { + "bbox": [ + 47, + 465, + 128, + 478 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 689 + ], + "type": "text", + "content": "Lane detection is a fundamental task in computer vision [31, 36, 40]. It aims to predict the location of the lane in a given image. Lane detection plays a crucial role in many applications, such as adaptive cruise control and lane keeping. Existing lane detection methods generally adopt a divide-and-conquer strategy, which decomposes the lane detection into multiple subtasks. Each subtask is accomplished by a task-specific head network. For instance, as shown in Fig. 1, segmentation-based methods [28,49] adopt a head network along with a task-specific module, such as the message-passing module [28], to predict per-pixel masks. Anchor-based methods [17, 36] utilize a classification head network for distinguishing lane instances and an anchor refinement network for regressing accurate lanes. Parameter-based methods [13, 23] use a network to predict the parameters of lanes and a vertical offset prediction network to locate the start points of lanes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Although divide-and-conquer strategy has been proved an effectively way to address the certain subtask in the ex" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 211, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 309, + 211, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 309, + 211, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 309, + 211, + 545, + 240 + ], + "type": "image", + "image_path": "4a37222f67df15e4bb504ee6e2aa97b51868270bd642c8465112acf001dbb588.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 239, + 454, + 247 + ], + "lines": [ + { + "bbox": [ + 359, + 239, + 454, + 247 + ], + "spans": [ + { + "bbox": [ + 359, + 239, + 454, + 247 + ], + "type": "text", + "content": "(a) Segmentation-based lane detection" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 257, + 545, + 294 + ], + "blocks": [ + { + "bbox": [ + 309, + 257, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 309, + 257, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 309, + 257, + 545, + 294 + ], + "type": "image", + "image_path": "2d620d4f44898b74f691c1aeedaee1f466a7f2a97d8c4d7fe101ab9f1459b206.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 299, + 440, + 307 + ], + "lines": [ + { + "bbox": [ + 360, + 299, + 440, + 307 + ], + "spans": [ + { + "bbox": [ + 360, + 299, + 440, + 307 + ], + "type": "text", + "content": "(b) Anchor-based lane detection" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 319, + 545, + 363 + ], + "blocks": [ + { + "bbox": [ + 309, + 319, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 309, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 309, + 319, + 545, + 363 + ], + "type": "image", + "image_path": "b023bb946345dd46ef4f22d65b7c07a531cbbe7c794f93a21fd611efb8d8244e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 365, + 448, + 372 + ], + "lines": [ + { + "bbox": [ + 362, + 365, + 448, + 372 + ], + "spans": [ + { + "bbox": [ + 362, + 365, + 448, + 372 + ], + "type": "text", + "content": "(c) Parameter-based lane detection" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 384, + 545, + 454 + ], + "blocks": [ + { + "bbox": [ + 307, + 384, + 545, + 454 + ], + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 454 + ], + "type": "image", + "image_path": "66d1487b06eae28eceb22e021472fd336771ba70435b7eef1f4c5b5c0c3b1b40.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 458, + 487, + 466 + ], + "lines": [ + { + "bbox": [ + 364, + 458, + 487, + 466 + ], + "spans": [ + { + "bbox": [ + 364, + 458, + 487, + 466 + ], + "type": "text", + "content": "(d) Our sequence generation-based lane detection" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 475, + 538, + 487 + ], + "lines": [ + { + "bbox": [ + 312, + 475, + 538, + 487 + ], + "spans": [ + { + "bbox": [ + 312, + 475, + 538, + 487 + ], + "type": "text", + "content": "Figure 1. Comparison of different lane detection frameworks." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 567 + ], + "type": "text", + "content": "isting methods, there still exist several limitations. (1) Each subtask needs a customized task-specific head network, resulting in a complicated lane detection model. (2) Each task-specific head requires one or more loss functions, e.g., cross-entropy loss and Line-IOU loss [17], bringing out extra hyper-parameters." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 568, + 546, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 546, + 689 + ], + "type": "text", + "content": "In this paper, we present a novel sequence generation-based lane detector (Lane2Seq) framework to tackle the aforementioned issues. By formulating the lane detection as a sequence generation task, Lane2Seq gets rid of customized head networks and task-specific loss function. It is based on the intuition that if the detection model knows where the target lane is, model can be simply taught how to read the location of lane out, instead of designing additional classification head or regression head by the divide-and-conquer strategy." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "Therefore, we convert the outputs of different lane detection formats, i.e., segmentation-based, anchor-based, and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16944" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "parameter-based, into a sequence of discrete tokens and the model learns to generate this sequence token-by-token. As shown in Fig. 1 (d), to achieve a specific lane detection format, Lane2Seq uses a prompt to specify the detection format and the generated sequence adapts to the prompt so the model can produce format-specific output. By the format-specific prompt, Lane2Seq unifies different lane detection formats into a model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 169, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 289, + 350 + ], + "type": "text", + "content": "While Lane2Seq does not contain task-specific components, task-specific knowledge contained in these components can help the model learn the features of lanes better. We propose a Multi-Format model tuning method based on Reinforcement Learning (MFRL) to incorporate the task-specific knowledge into the model without changing model's architecture. Inspired by Task-Reward [30], MFRL takes the evaluation metrics, which naturally integrates task-specific knowledge, as the reward and tunes Lane2Seq using REINFORCE [41] algorithm. However, evaluation metrics like F1 score cannot be used as the reward directly due to undecomposable as a sum of perexample rewards. In this paper, we propose three new evaluation metric-based rewards for segmentation, anchor, and parameter format, based on their task-specific knowledge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "text", + "content": "Experimental results demonstrate that our Lane2Seq achieves competitive performance on three public datasets, Tusimple, CULane, and LLAMAS. For instance, Lane2Seq with ViT-Base encoder obtains " + }, + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "inline_equation", + "content": "97.95\\%" + }, + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "inline_equation", + "content": "97.42\\%" + }, + { + "bbox": [ + 46, + 350, + 289, + 482 + ], + "type": "text", + "content": " F1 score on Tusimple and LLAMAS, setting a new state-of-the-art result for two datasets. It should be noted that all existing lane detection methods heavily rely on well-designed task-specific head networks and the corresponding complicated loss functions. Instead, our Lane2Seq utilizes a plain transformer-based encoder-decoder architecture with a simple cross-entropy loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 483, + 268, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 483, + 268, + 495 + ], + "spans": [ + { + "bbox": [ + 59, + 483, + 268, + 495 + ], + "type": "text", + "content": "The main contributions of this paper are as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 506, + 287, + 701 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 506, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 506, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 59, + 506, + 287, + 578 + ], + "type": "text", + "content": "- We propose a sequence generation-based method for lane detection, which casts the lane detection as a sequence generation task. To the best of our knowledge, we are the first to unify the lane detection through the sequence generation, which offers a new perspective on lane detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 591, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 591, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 58, + 591, + 287, + 640 + ], + "type": "text", + "content": "- We present a novel reinforcement learning-based multi-format model tuning, including three new evaluation metric-based reward functions, to incorporate the task-specific knowledge into the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 653, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 653, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 58, + 653, + 287, + 701 + ], + "type": "text", + "content": "- Experimental results show that our method achieves competitive performance on lane detection benchmarks. Remarkably, we establish a new state-of-the-art result on Tusimple and LLAMAS." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "type": "text", + "content": "Lane detection. Existing lane detection methods can be divided into three categories based on the representation of the lane: segmentation-based method, anchor-based method and parameter-based method. Segmentation-based methods [28, 47, 49] consider lane detection as a semantic segmentation task and performs pixel-wise prediction. SCNN [28] enhances the visual evidence by a message-passing module, which can capture spatial dependency for lanes. Anchor-based methods [20, 31, 36, 51] predict the accurate lanes by refining the predefined lane anchors. UFLD [31] proposes a novel row anchor-based approach to detect lanes. Different from the segmentation-based and anchor-based methods, parameter-based methods [13, 23, 37] regard the lane detection as the parametric modeling and regress the parameters of the lane. PolyLaneNet [37] models a lane curve as a polynomial and regresses the parameters of the polynomial. In this paper, we treat the segmentation-based, anchor-based methods, and parameter-based metohods uniformly as the sequence generation task, getting rid of the complicate structure and task-specific modules, e.g., head network. It only adopts the cross-entropy loss and plain transformer architecture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 355, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 546, + 617 + ], + "type": "text", + "content": "Sequence generation for vision tasks. Recently, sequence-to-sequence (seq2seq) method used in the natural language processing (NLP) has been applied for vision tasks. Seq2seq utilizes a basic transformer encoder and decoder architecture and accomplishes the task by making sequence predictions, rather than designing a model tailored to the vision task. Pix2seq [8] is the pioneering work to cast the object detection as the sequence generation task. It shows that object can be detected well without any task-specific modules like label assignment. Besides object detection, seq2seq has been extended to other vision tasks such as instance segmentation [9], keypoint detection [9], text spotting [19,29] and object tracking [10]. UniTAB [45] adopts the task prompt [33] to perform multi-task learning. Unified-IO [25] jointly trains various vision tasks by setting the unified input/output formats for all tasks. Besides, seq2seq becomes increasingly popular in the multimodal model. Text-to-image models like DALL-E [34] and vision-language models like Flamingo [2] all use seq2seq to unify the multi-modal tasks. Nevertheless, how to perform the seq2seq in the lane detection to unify different detection formats is still unexplored." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 617, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 701 + ], + "type": "text", + "content": "Lane2Seq has a similar spirit with the Pix2seq and its successors [9, 10, 19, 29]. All of them view the vision tasks as the sequence generation. The main difference between these methods and Lane2Seq is the sequence format. Previous works use coordinates and category to construct the sequence, while Lane2Seq also adopts the parameter of the lane to construct the sequence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "Reinforcement learning in the computer vision. Many" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16945" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 68, + 512, + 299 + ], + "blocks": [ + { + "bbox": [ + 83, + 68, + 512, + 299 + ], + "lines": [ + { + "bbox": [ + 83, + 68, + 512, + 299 + ], + "spans": [ + { + "bbox": [ + 83, + 68, + 512, + 299 + ], + "type": "image", + "image_path": "63e1195f0cba21dbe15638dac9d121f95f6aaf9a247c799564b18d7078d670c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 547, + 329 + ], + "type": "text", + "content": "Figure 2. Inference pipeline of Lane2Seq. The model perceives input image, prompt and generates format-specific tokens, which can be detokenized into required detection format for visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 338, + 289, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 517 + ], + "type": "text", + "content": "previously researches introduce the reinforcement learning for the vision tasks, such as object detection [27], object tracking [26], image segmentation [21] and lane detection [48]. They generally focus on learning different parts of an image and refine the outputs iteratively. DQLL [48] localizes the lanes as a group of lanemarks and refine the location of lanemarks with the deep Q-learning. Recently, Task Reward [30] adopts a novel task risk-based reinforcement learning to tune the vision model without changing the model architecture. Reward functions in Task-Reward are designed for object detection, instance segmentation, colorization, and image captioning but not for lane detection. Our MFRL is the first attempt to design effective evaluation metric-based reward functions for different lane detection formats." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 531, + 104, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 531, + 104, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 531, + 104, + 544 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 636 + ], + "type": "text", + "content": "In this section, we present the proposed sequence generation-based lane detection method, named Lane2Seq, in detail. The overall pipeline of Lane2Seq is illustrated in Fig. 2. Sec 3.1 depicts the unified interface for lane detection, Sec 3.2 details the model architecture and objective function, and Sec 3.3 introduces the proposed multi-format model tuning based on reinforcement learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 646, + 241, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 241, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 241, + 658 + ], + "type": "text", + "content": "3.1. Unified Interface for Lane Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "As presented in Figure 1, lane detection formats in the existing method are diverse and have been formulated quite differently. Considering differences in the form of outputs, customized models with specialized head networks and loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 338, + 544, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 338, + 544, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 338, + 544, + 350 + ], + "type": "text", + "content": "functions are designed for different lane detection formats." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "text", + "content": "To integrate different lane detection formats into a model, we propose an unified sequence interface for lane detection, where both format transcription like segmentation and outputs are treated as sequences of discrete tokens. As shown in Fig. 2, the generated sequence is composed of four parts, a starting token " + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "inline_equation", + "content": "\\langle \\text{starting} \\rangle" + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "text", + "content": ", a format transcription token like " + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "inline_equation", + "content": "\\langle \\text{Segmentation} \\rangle" + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "text", + "content": ", a format-specific sequence, and an ending token " + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "inline_equation", + "content": "\\langle \\text{end} \\rangle" + }, + { + "bbox": [ + 304, + 351, + 546, + 470 + ], + "type": "text", + "content": ". The format-specific sequence for three detection formats can be constructed as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": "Segmentation sequence. Instead of performing pixelwise mask prediction, we predict the polygon [5] corresponding to the mask as a sequence of coordinates conditioned on a given lane instance. Then, we convert the polygon into a sequence by quantizing the coordinates of its points into discrete tokens. Specifically, " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": " coordinates of a point are normalized to the width and height of the image, and then quantized to " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "[1, n_{\\text{bins}}]" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "n_{\\text{bins}}" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": " is the size of the vocabulary. Vocabulary will be described subsequently. A polygon sequence can be expressed as " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "[x_1, y_1, x_2, y_2, \\ldots, x_{28}, y_{28}, <\\text{Lane}>]" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "<\\text{Lane}>" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": " is the category token. If there are multiple lanes in an image, we concatenate the all polygon sequences. The segmentation sequence consists of a starting point and all polygon sequences, where starting point is the left-top point of the image " + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "inline_equation", + "content": "x_0 = 0, y_0 = 0" + }, + { + "bbox": [ + 304, + 473, + 547, + 664 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "Anchor sequence. Since the essence of anchor-based methods is regressing the location of keypoints of lane anchors, we treat the anchor prediction as the keypoint sequence generation. Specifically, the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16946" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 282, + 159 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 282, + 159 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 282, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 282, + 159 + ], + "type": "image", + "image_path": "e09fe164266f283dbbe79453cf1e3756568de1f8c1d134a19d0d6fffaa811609.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 288, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 201 + ], + "type": "text", + "content": "Figure 3. Training pipeline of Lane2Seq. The input sequence can be the segmentation sequence or anchor sequence or parameter sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 280 + ], + "type": "text", + "content": "keypoint sequence of a lane can be expressed as \\([x_1, y_1, x_2, y_2, \\ldots, x_{14}, y_{14}, <\\text{Lane}>\\]\\. The normalization and quantization of keypoint coordinates are same to that of polygon point. The anchor sequence contains a starting point and all keypoint sequences, where starting point is still the left-top point of the image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": "Parameter sequence. the sequence of the lane includes two parts: the parameter of polynomial function and vertical offsets. We set the polynomial degree as 5th. Specifically, the parameter sequence of a lane can be represented as " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "[a_1,a_2,a_3,a_4,a_5,s,\\langle \\mathrm{Lane}\\rangle ]" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": " is the vertical offset. " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": " is normalized to the height of the image and quantized to " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "[1,n_{\\mathrm{bins}}]" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": ". We use sigmoid to normalize " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "a_{i} = \\operatorname {Sigmoid}(a_{i})" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": " is quantized to " + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "inline_equation", + "content": "[1,n_{\\mathrm{bins}}]" + }, + { + "bbox": [ + 46, + 280, + 287, + 387 + ], + "type": "text", + "content": ". Parameter sequence does not require the starting point." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 436 + ], + "type": "text", + "content": "Vocabulary. We use a shared vocabulary for all formats and each integer between " + }, + { + "bbox": [ + 46, + 388, + 287, + 436 + ], + "type": "inline_equation", + "content": "[1, n_{bins}]" + }, + { + "bbox": [ + 46, + 388, + 287, + 436 + ], + "type": "text", + "content": " can be regarded as a word in the vocabulary. Each word in the vocabulary corresponds to a learnable embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 279, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 279, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 279, + 456 + ], + "type": "text", + "content": "3.2. Unified Architecture and Objective Function" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 605 + ], + "type": "text", + "content": "We follow [10] and use a transformer-based encoder-decoder architecture to deal with image input and sequence output flexibly. As shown in Fig. 3, the image encoder takes the pixels as the input and outputs the corresponding image representations. We utilize the Vision Transformer (ViT) [11] to instantiate the image encoder. We adopt a transformer-based sequence decoder, which is widely used in the language modeling [32, 33], to generate the output sequence. The decoder generates one token at a time based on the preceding tokens and the image representations. Sequence decoder removes the well-designed task-specific heads for different detection formats." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "Training Pipeline. We first tokenize annotation of each format into the corresponding sequence. Then, we construct a training batch using images and sequences from all detection formats. Finally, we compute loss between generated sequence and target sequence for each format. For a certain detection format, the input sequence is " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "[<\\text{starting}>" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": ", format transcription token, format-specific sequence] and the target sequence is [format transcription token, format-specific sequence, " + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{end}\\rangle" + }, + { + "bbox": [ + 46, + 605, + 287, + 714 + ], + "type": "text", + "content": "]. For example, in the anchor format, the in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "content": "put sequence is " + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "inline_equation", + "content": "[<\\text{starting}>, <\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]" + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "content": " and the target sequence is " + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "inline_equation", + "content": "[<\\text{Anchor}>, x_0, y_0, x_1, y_1, \\ldots, <\\text{Lane}>, <\\text{end}>]" + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "content": ". The starting token " + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "inline_equation", + "content": "<\\text{starting}>" + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "content": " and ending token " + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "inline_equation", + "content": "<\\text{end}>" + }, + { + "bbox": [ + 304, + 73, + 545, + 167 + ], + "type": "text", + "content": " are learnable embeddings, which tell the decoder when to begin and end the sequence generation. The sequence decoder perceives the image representations and input sequence and reconstructs the target sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "content": "Objective Function. Similar to Pix2seq [8], we train the Lane2Seq with a simple cross-entropy loss. At each time stamp " + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "content": ", Lane2Seq aims to maximize the likelihood of the target tokens conditioned on the image representations " + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "content": " and previously generated tokens " + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "inline_equation", + "content": "y_{1:j-1}" + }, + { + "bbox": [ + 304, + 167, + 545, + 228 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 232, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 232, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 339, + 232, + 545, + 266 + ], + "type": "interline_equation", + "content": "L o s s _ {o b j} = - \\sum_ {j = 1} ^ {N} w _ {j} \\log P (\\hat {m} _ {j} | I, m _ {1: j - 1}), \\tag {1}", + "image_path": "4a44a553a2801b32774238f2c69fed8b5db57d17df8b631e09d12afe94081e09.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "inline_equation", + "content": "\\hat{m}" + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": " denote the input and target sequence, respectively. " + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": " is the length of sequence. " + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "inline_equation", + "content": "w_{j}" + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": " represents the weight for the " + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 269, + 545, + 340 + ], + "type": "text", + "content": "-th token, 0 is assigned to the format transcription token and 1 is assigned to other tokens to ensure the model is trained to predict the desired tokens but not the format transcription tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 437 + ], + "type": "text", + "content": "Inference Pipeline. We take the format transcription to-ken as the prompt to achieve the format-specific detection and the sequence decoder generates the rest of the sequence conditioned on the prompt and image representation. Once the whole sequence is generated, we de-quantize the sequence to obtain the location of lane. More details about de-quantization for each format can be found in the supplementary materials. Inference process is presented in Fig. 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 469 + ], + "type": "text", + "content": "3.3. Multi-Format Model Tuning Based on Reinforcement Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 617 + ], + "type": "text", + "content": "Lane2Seq get rids of the complicated task-specific components, such as task-specific head networks, via the sequence generation. However, this task-agnostic architecture inevitably lacks the task-specific knowledge of the lane detection, making the model ineffective in learning the features of lane. Existing lane detection methods [17, 49] usually design the specific modules to incorporate the task-specific knowledge into the model. But this way makes the model architecture complicate. We propose a novel Multi-Format model tuning method based on Reinforcement Learning (MFRL) to learn task-specific knowledge effectively without changing any components of the model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": "Inspired by Task-Reward [30], MFRL takes the evaluation metric, which naturally contains the task-specific knowledge, as the reward function and adopts the reinforcement learning to tune the model, which is common practice in the modern language model [4, 12]. Specifically, it is difficult for the model to converge if model is trained with reinforcement learning method from scratch [30]. Hence, MFRL consists of two stages: the pretraining stage and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16947" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "the model tuning stage. In the pretraining stage, model is trained on the lane detection dataset with objective function in Eq. 1 to have a good weight initialization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 108, + 287, + 133 + ], + "type": "text", + "content": "In the model tuning stage, we use the REINFORCE algorithm [35] to maximize the objective function as below," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 144, + 287, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 144, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 144, + 287, + 158 + ], + "type": "interline_equation", + "content": "O b j = E _ {c \\sim D} \\left[ E _ {t \\sim Q (\\cdot | c, u)} R (t, g) \\right], \\tag {2}", + "image_path": "70055fb8ac4a72eaf05e37ecf5d36bf9c17a123c3d4c0ecaa5332f0884e6c084.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "c, t," + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " represent the input image, generated format-specific sequences, and ground truths, respectively. " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " denote the mathematic expectation and the model parameter. " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "R, D," + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": " stand for the reward function, data distribution of the dataset, and conditional distribution parameterized by " + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 158, + 287, + 240 + ], + "type": "text", + "content": ". REINFORCE algorithm estimates the gradient of the reward function by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 287, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 287, + 264 + ], + "type": "interline_equation", + "content": "\\nabla_ {u} E _ {t \\sim Q} [ R (t, g) ] = E _ {t \\sim Q} [ R (t, g) \\nabla_ {u} \\log Q (t | c, u) ], \\tag {3}", + "image_path": "f894ec121eddef5e4204c127bc22ded340af45697d88f0731cde86a63bf6534c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 273, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 320 + ], + "type": "text", + "content": "In practice, Eq. 3 is computed as an average of per-sequence gradient and reward function of per-sequence is not required to be differentiable. More details about REINFORCE algorithm can be found in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": "Reward Function. Evaluation metric F1 score and accuracy are unable to adopt as the reward directly, because they cannot be decomposable as a sum of per-sequence rewards (see Sec 4.2). However, F1 score or accuracy is composed of false positives " + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "inline_equation", + "content": "(FP)" + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": ", true positives " + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "inline_equation", + "content": "(TP)" + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": ", and false negatives " + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "inline_equation", + "content": "(FN)" + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": ". All three indicators can be computed per-sequence. We opt to use " + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 46, + 321, + 287, + 427 + ], + "type": "text", + "content": " to design the reward. Specifically, reward for three detection formats are constructed as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "content": "Reward for the segmentation format (termed as " + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "inline_equation", + "content": "R_{seg}" + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "content": "). Segmentation-based lane detection includes two task-specific knowledge, i.e., knowledge of the segmentation and lane detection. For the knowledge of lane detection, we adopt a matched Line-IOU (LIOU) [50], which can demonstrate the quality of location prediction and shape prediction of the lane. For the knowledge of the segmentation, we adopt the matched Mean Intersection over Union (mIOU), which is a widely used metric in the semantic segmentation to evaluate the quality of segmentation, as the reward. Finally, we add both of them to compute " + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "inline_equation", + "content": "r_{seg}" + }, + { + "bbox": [ + 46, + 429, + 287, + 561 + ], + "type": "text", + "content": " as below," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 568, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 568, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 68, + 568, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {s e g} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + m I O U \\left(p _ {k}, g\\right) \\right] \\tag {4} \\\\ - \\lambda_ {1} F P _ {s e g} (t, g), \\\\ \\end{array}", + "image_path": "107d3bd2f3f65013716eaebac844f351f1e788a06880d93b177ad8d776558e8d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " represents the number of true positives in " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": "-th true positive. " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "FP_{seg}" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " is the false positive rate of the segmentation format and " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " is the weight to control the affect of " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "FP_{seg}" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": ". We introduce " + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "inline_equation", + "content": "FP_{seg}" + }, + { + "bbox": [ + 46, + 617, + 287, + 676 + ], + "type": "text", + "content": " to penalize for false positives." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Reward for the anchor format (termed as " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "R_{a}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "). Anchors-based lane detection contains the knowledge of the keypoint location and lane detection. For the knowledge of keypoint" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "location, we simply adopt the matched Euclidean distance " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "d(p_k, g)" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " between true positives and corresponding ground truths as the reward. Since MFRL needs to maximize the reward, we rescale the reward as " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "d_r(p_k, g) = 1 - \\frac{d(p_k, g)}{H}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is the height of image. For the knowledge of lane detection, we also utilize the matched LIOU as the reward. The formulation of " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "r_a" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 162, + 545, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 162, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 309, + 162, + 545, + 200 + ], + "type": "interline_equation", + "content": "R _ {a} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left[ L I O U \\left(p _ {k}, g\\right) + d _ {r} \\left(p _ {k}, g\\right) \\right] - \\lambda_ {2} F P _ {a} (t, g), \\tag {5}", + "image_path": "c18e6c68c27f19959b07a55bdc9301789d610543d921eb9f41f95b6fe842d78e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "inline_equation", + "content": "FP_{a}" + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "text", + "content": " stand for the false positives rate of the anchor format and the weight of " + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "inline_equation", + "content": "FP_{a}" + }, + { + "bbox": [ + 304, + 201, + 545, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 225, + 545, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 271 + ], + "type": "text", + "content": "Reward for the parameter format (termed as " + }, + { + "bbox": [ + 304, + 225, + 545, + 271 + ], + "type": "inline_equation", + "content": "R_p" + }, + { + "bbox": [ + 304, + 225, + 545, + 271 + ], + "type": "text", + "content": "). Parameter-based detection only has the knowledge of lane detection, hence we directly use the matched LIOU as the reward." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 333, + 274, + 545, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 274, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 333, + 274, + 545, + 303 + ], + "type": "interline_equation", + "content": "R _ {p} (t, g) = \\frac {1}{K} \\sum_ {k = 1} ^ {K} L I O U \\left(p _ {k}, g\\right) - \\lambda_ {3} F P _ {p} (t, g), \\tag {6}", + "image_path": "da292e3bb33dae65fa005845c40c1a7cbb91b44c07015345660ac08c83c53311.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "inline_equation", + "content": "FP_{p}" + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "content": " stand for the false positives rate of the parameter format and the weight of " + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "inline_equation", + "content": "FP_{p}" + }, + { + "bbox": [ + 304, + 304, + 545, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 363 + ], + "type": "text", + "content": "Since different detection formats have different contributions, we weight the objective function of different formats. The final objective function of MFRL can be computed as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 370, + 545, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 370, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 342, + 370, + 545, + 382 + ], + "type": "interline_equation", + "content": "O b j _ {t o t a l} = \\lambda_ {4} O b j _ {s e g} + \\lambda_ {5} O b j _ {a} + \\lambda_ {6} O b j _ {p}, \\tag {7}", + "image_path": "2a05d5e1855eb6e1d9c01dd7d23555db640da501a59dcffd310a4402874ec352.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "Obj_{seg}" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "Obj_a" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "Obj_p" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " denote the objective function of segmentation, anchor, and parameter format, respectively. The objective function of three formats is the same as Eq. 2, except the format-specific sequence and reward function. " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_5" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\lambda_6" + }, + { + "bbox": [ + 304, + 388, + 545, + 448 + ], + "type": "text", + "content": " are the scale factor." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 459, + 387, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 387, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 387, + 472 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 479, + 368, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 368, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 368, + 490 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 521 + ], + "type": "text", + "content": "We conduct experiments on three lane detection benchmarks: CULane [28], Tusimple [1], and LLAMAS [3]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 521, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 580 + ], + "type": "text", + "content": "CULane is a widely used large-scale dataset for lane detection. It contains a lot of challenging scenarios such as crowded roads. The CULane dataset consists of 88.9K images for training, 9.7K images in the validation set, and 34.7K images for the test. Image size is " + }, + { + "bbox": [ + 304, + 521, + 545, + 580 + ], + "type": "inline_equation", + "content": "1640 \\times 590" + }, + { + "bbox": [ + 304, + 521, + 545, + 580 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "text", + "content": "Tusimple is a real highway dataset consisting of 3,626 training images and 2,782 testing images. All images have " + }, + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 304, + 580, + 545, + 616 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 616, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 663 + ], + "type": "text", + "content": "LLAMAS is a recently released large-scale lane detection dataset with over 100K images. All lane markers are annotated with high-accurate maps. Image size is " + }, + { + "bbox": [ + 304, + 616, + 545, + 663 + ], + "type": "inline_equation", + "content": "1280 \\times 717" + }, + { + "bbox": [ + 304, + 616, + 545, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 671, + 419, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 419, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 419, + 682 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "For CULane and LLAMAS dataset, we adopt the F1 score to measure the performance: " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "F_{1} =" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16948" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 547, + 295 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of F1 score and MACs (multiply-accumulate operations) on CULane testing set. We only report the false positives for \"Cross\" category following [50]. FPS is tested on a single 2080Ti." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 547, + 295 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 547, + 295 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 547, + 295 + ], + "type": "table", + "html": "
MethodsEncoderFPS↑Normal↑Crowded↑Dazzle↑Shadow↑No line↑Arrow↑Curve↑Night↑Cross↓Total↑MACs(G)↓
Segmentation based
SCNN [28]ResNet501290.6069.7058.5066.9043.4084.1064.4066.10190071.608.4
RESA [49]ResNet5042.192.1073.1069.2072.8047.7088.3070.3069.90150375.37.6
AtrousFormer [44]ResNet34-92.8375.9669.4877.8650.1588.6671.1473.74105478.0814.22
LaneAF [44]DLA34 [46]-91.8075.6171.7879.1251.3886.8872.7073.03136077.4112.73
Lane2Seq (segmentation)ViT-Base1993.3977.2773.4579.6953.9190.5373.3774.96112979.6415.19
Anchor based
LaneATT [36]ResNet1222391.7476.1669.4776.3150.4686.2964.0570.81126477.0270.5
O2SFformer [51]ResNet504493.0976.5772.2576.5652.8089.5069.6073.85311877.8327.51
UFLD [31]ResNet3419290.7070.2059.5069.3044.4085.7069.5066.70203772.303.3
CondLaneNet [22]ResNet3413593.3877.1471.1779.9351.8589.8973.8873.92138778.7419.6
ADNet [43]ResNet34-92.9077.4571.7179.1152.8989.9070.6474.78149978.9418.6
CLRNet [50]ResNet3411293.4978.0674.5779.9254.0190.5972.7775.02121679.7317.3
Lane2Seq (anchor)ViT-Base2393.1177.4373.2579.4653.7490.0272.4475.12117379.2715.19
Parameter based
LSTR [23]ResNet18----------64.005.9
BezierLaneNet [14]ResNet18-90.2271.5562.4970.9145.3084.0958.9868.7099673.674.7
BSNet [6]ResNet34-93.7578.0176.6579.5554.6990.7273.9975.28144579.8917.2
Eigenlanes [18]ResNet50-91.7076.0069.8074.1052.2087.7062.9071.80150977.2018.7
Laneformer [15]ResNet50-91.7775.7470.1775.7548.7387.6566.3371.041977.0626.2
Lane2Seq (parameter)ViT-Base3093.0376.4272.1778.3252.8989.6772.6773.98131978.3915.19
", + "image_path": "0a3e3b1826030828e48d7cd598e27c26c6896ff44e18e04f685781ffddf0cefe.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 312, + 279, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 279, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 279, + 343 + ], + "type": "inline_equation", + "content": "\\frac{2 \\times \\text{Precision} \\times \\text{Recall}}{\\text{Precision} + \\text{Recall}}" + }, + { + "bbox": [ + 47, + 312, + 279, + 343 + ], + "type": "text", + "content": ", where Precision = TP/TP+FP and Recall = TP/TP+FN." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": "For Tusimple dataset, we use F1 score, accuracy, false positives, and false negatives to evaluate the model performance. Accuracy is defined as Accuracy = " + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "inline_equation", + "content": "\\frac{\\sum_{clip} C_{clip}}{\\sum_{clip} S_{clip}}" + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "inline_equation", + "content": "C_{clip}" + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": " represents the number of accurately predicted lane points and " + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "inline_equation", + "content": "S_{clip}" + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": " denotes the total number of lane points of a clip. A lane point is considered as correct if its distance is smaller than the given threshold " + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "inline_equation", + "content": "t_{pc} = \\frac{20}{\\cos(a_{yl})}" + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "inline_equation", + "content": "a_{yl}" + }, + { + "bbox": [ + 46, + 342, + 288, + 445 + ], + "type": "text", + "content": " denotes the angle of the corresponding ground truth." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 453, + 180, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 180, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 180, + 467 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": "Model. We adopt the ViT-Base initialized with the MAE [16] pretrained parameters as the image encoder. The patch size is " + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": ". The decoder is composed of 2 transformer blocks and its hidden size is 256. The number of attention heads is 8 and the hidden size of feed-forward network is 1024. " + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "inline_equation", + "content": "n_{bins}" + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": " is set to 1000 and the dimension of word embedding of the vocabulary is 256. Our Lane2Seq is implemented based on MMDetection toolbox [7]. " + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4, \\lambda_5" + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\lambda_6" + }, + { + "bbox": [ + 46, + 473, + 287, + 593 + ], + "type": "text", + "content": " are set to 0.3, 0.3, 0.1, 0.2, 1, and 1.5, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Training. All input images are resized to " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 800" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ". We utilize the AdamW [24] as the optimizer with initial learning rate 1e-4. The training epochs of the pretraining stage are 5, 20, and 15 for CULane, Tusimple, and LLAMAS. We set the training epochs of the model tuning stage to 15, 30, and 55 for CULane, Tusimple, and LLAMAS, respectively. For the data augmentation, we use the random horizontal flips and random affine transformation including scaling, rotation, and translation. All experiments are conducted on 8 A100 GPUs with total batch size 384." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 343, + 545, + 453 + ], + "blocks": [ + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 545, + 335 + ], + "type": "text", + "content": "Table 2. Comparison of the performance of different models on LLAMAS." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 343, + 545, + 453 + ], + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 453 + ], + "type": "table", + "html": "
MethodsEncoderF1(%)↑
PolyLaneNet [37]EfficientNet-b0 [38]90.20
BezierLaneNetResNet3496.11
LaneATTResNet3494.96
LaneATTResNet12295.17
LaneAFDLA3496.90
CLRNetResNet1896.96
CLRNetDLA3497.16
Lane2Seq (segmentation)ViT-Base97.42
Lane2Seq (anchor)ViT-Base97.05
Lane2Seq (parameter)ViT-Base96.74
", + "image_path": "b82bf5924c6687a491d276b6bde023c7141a3158a60c51645edb43dc91877f7d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 464, + 545, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 464, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 464, + 545, + 478 + ], + "type": "text", + "content": "4.4. Comparison with the State-of-the-art Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "Performance on LLAMAS. Our Lane2Seq achieves a new state-of-the-art performance on LLAMAS dataset using segmentation format. In Table 2, Lane2Seq increases F1 score from " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "97.16\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "97.42\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": " compare to the previous state-of-the-art method CLRNet. Compared with parameter-based state-of-the-art method BezierLaneNet, Lane2Seq performs better than BezierLaneNet (96.74% vs. " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "96.11\\%" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "). The results manifest the detection strengths of Lane2Seq in the multi-lane scenario (The number of lanes " + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "inline_equation", + "content": "\\geq 5" + }, + { + "bbox": [ + 304, + 484, + 545, + 628 + ], + "type": "text", + "content": "), such as highways. The reason may be that transformer-based architecture has the advantages in capturing long-range dependency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Performance on Tusimple. Table 3 presents the performance comparison with state-of-the-art approaches on Tusimple. Due to small data scale and single scene, the performance gap between different models is small. Our Lane2Seq using segmentation format sets a new state-of-the-art F1 score of " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "97.95\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ". The results also validate the detection strengths of Lane2Seq in multi-lane scenario." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16949" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 287, + 215 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 93 + ], + "type": "text", + "content": "Table 3. Comparison of the performance of different models on Tusimple. Acc denotes accuracy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 287, + 215 + ], + "type": "table", + "html": "
MethodsEncoderF1(%)↑Acc(%)↑FP(%)↓FN(%)↓
SCNNVGG1695.9796.536.171.80
RESAResNet3496.9396.823.632.48
PolyLaneNetEfficientNet-bo90.6293.369.429.33
UFLDResNet3488.0295.8618.913.75
LaneATTResNet12296.0696.105.642.17
GANet [39]ResNet3497.6895.871.992.64
CondLaneNetResNet3496.9895.372.203.82
CondLaneNetResNet10197.2496.542.013.50
CLRNetResNet3497.8296.872.272.08
CLRNetResNet10197.6296.832.372.38
Lane2Seq (segmentation)ViT-Base97.9596.852.012.03
Lane2Seq (anchor)ViT-Base97.8696.722.212.05
Lane2Seq (parameter)ViT-Base96.5996.002.233.54
", + "image_path": "09482a1866244daadd92eef0b271f5d628c88ac8d5238f3374acc19fa4d05867.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": "Performance on CULane. We compare Lane2Seq with other state-of-the-art methods on CULane and results are shown in Table 1. For the segmentation-based methods, Lane2Seq improves F1 score from " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "78.08\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "79.64\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": " compared to the previous state-of-the-art methods AtrousFormer. Compared with anchor-based methods, Lane2Seq outperforms most previous methods. For example, Lane2Seq achieves the better performance than existing row anchor-based methods, such as CondLaneNet (79.27% vs. " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "78.74\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": "). Moreover, Lane2Seq achieves competitive performance compared to CLRNet (79.27% vs. " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "79.73\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": "). One possible reason for the performance gap between CLRNet and Lane2Seq is that CLRNet adopts multi-scale detection, which can notably improve the detection performance, while Lane2Seq only uses the plain transformer architecture with single-scale detection. For the comparison of parameter-based methods, Lane2Seq surpasses the existing transformer-based methods. For example, Lane2Seq significantly increases F1 score from " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "64.00\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "78.39\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": " compared with LSTR and outperforms Laneformer by " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "1.33\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": " (78.39% vs. " + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "inline_equation", + "content": "77.06\\%" + }, + { + "bbox": [ + 47, + 234, + 287, + 532 + ], + "type": "text", + "content": "). But Lane2Seq is ranked second after BSNet. As for the inference speed, we can see that Lane2Seq does not have a significant advantage in inference speed. We attribute that token-by-token prediction method slows the inference speed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "type": "text", + "content": "Based on the above results and analyses, we can conclude that the unified lane detection via the sequence generation without any well-designed task-specific components, combining our reinforcement learning-based multi-format model tuning, can achieve promising performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 592, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 653 + ], + "type": "text", + "content": "Qualitative Results. We display the qualitative results in Fig. 4. The results show that Lane2Seq can effectively detect lanes in the multi-lane scenario (see Fig 4 (a) and (c)). Even in the case of the night scene, Lane2Seq successfully discriminates the lanes (see Fig 4 (b))." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 142, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 142, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 142, + 672 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We conduct the ablation experiments on CULane dataset to validate the effectiveness of each component. More ablation studies are in the supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 101, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 4. Ablation study on the multi-format model tuning based on reinforcement learning." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 101, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 308, + 101, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 308, + 101, + 545, + 163 + ], + "type": "table", + "html": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
Lane2Seq (segmentation)64.2770.4452.77
Lane2Seq (segmentation) + MFRL79.64 (+15.37)85.26 (+14.82)69.00 (+16.23)
Lane2Seq (anchor)66.2371.3655.75
Lane2Seq (anchor) + MFRL79.27 (+13.04)84.72 (+13.36)67.28 (+11.53)
Lane2Seq (parameter)68.1473.2757.89
Lane2Seq (parameter) + MFRL78.39 (+10.25)83.68 (+10.41)66.94 (+9.05)
", + "image_path": "36b5580639691b58cde949a59edf57afac424eb24066be334ade7b64b562fd97.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 206, + 545, + 302 + ], + "blocks": [ + { + "bbox": [ + 305, + 176, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 305, + 176, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 545, + 198 + ], + "type": "text", + "content": "Table 5. Ablation study on the different reward functions. F1 score is adopted as the indicator." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 206, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 309, + 206, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 309, + 206, + 545, + 302 + ], + "type": "table", + "html": "
\\( R_{seg} \\)\\( R_a \\)\\( R_p \\)SegmentationAnchorParameter
64.2766.2368.14
✓77.69 (+13.42)66.23 (+0.0)68.14 (+0.0)
✓64.27 (+0.0)76.94 (+10.71)68.14 (+0.0)
✓64.27 (+0.0)66.23 (+0.0)77.02 (+8.88)
✓✓78.72 (+14.45)78.14 (+11.91)68.14 (+0.0)
✓✓78.70 (+14.43)66.23 (+0.0)77.90 (+9.76)
✓✓64.27 (+0.0)78.20 (+11.97)77.93 (+9.79)
✓✓✓79.64 (+15.37)79.27 (+13.04)78.39 (+10.25)
", + "image_path": "22096d4e1c2c53f2df10527b56d881e67c650144733f2cb21b773936ce27d584.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": "Multi-Format Model Tuning Based on Reinforcement Learning. We first ablate the effectiveness of the proposed Multi-Format Model Tuning Based on Reinforcement Learning (MFRL). As shown in Table 4, MFRL significantly improves the performance of different detection formats. For the segmentation format, MFRL gains " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "15.37\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " F1 score improvement " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "(79.64\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "64.27\\%)" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": ". For the anchor and parameter format, MFRL increases " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "13.04\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "(79.27\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "66.23\\%)" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "10.25\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "(78.39\\%" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "inline_equation", + "content": "68.14\\%)" + }, + { + "bbox": [ + 304, + 319, + 545, + 605 + ], + "type": "text", + "content": " F1 score, respectively. Reasons of performance improvement can be attributed in two folds: (1) the vanilla Lane2Seq is a task-agnostic architecture and lacks the task-specific knowledge, leading to the ineffective feature learning of lane. MFRL tackles this problem by adopting the evaluation metric, which is designed based on the lane prior information like angle, as the reward to tune the model. (2) MFRL makes the model's predictions aligned with their intended usage, i.e., how to achieve a high performance. Optimizing the loss function is a common practice in the compute vision. However, this way indirectly optimizes for model's intended usage, because the lower loss does not mean the higher performance. Instead, MFRL views the maximizing the evaluation metric as the optimization objective, which is positively correlated with the model performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Effectiveness of Reward Function of Different Formats. We ablate the effectiveness of reward function of different formats and results are shwon in Table 5. It can be observed that our reward function can bring consistent performance improvement. However, we can also see that the performance will not be improved if the corresponding reward is not optimized. Besides, we can find that multi-reward optimization can achieves better results than single-reward optimization. Reason may be that multi-reward optimiza" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16950" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 70, + 439, + 205 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 439, + 205 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 439, + 205 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 439, + 205 + ], + "type": "image", + "image_path": "f024bdb0778d6a88ea6226dcb7224b0a91ff2310b956d59788a39416b9595492.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 214, + 490, + 227 + ], + "lines": [ + { + "bbox": [ + 101, + 214, + 490, + 227 + ], + "spans": [ + { + "bbox": [ + 101, + 214, + 490, + 227 + ], + "type": "text", + "content": "Figure 4. Visualization results of LaneATT, CLRNet, and Lane2Seq on Tusimple, CULane, and LLAMAS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 257, + 287, + 327 + ], + "blocks": [ + { + "bbox": [ + 68, + 237, + 265, + 248 + ], + "lines": [ + { + "bbox": [ + 68, + 237, + 265, + 248 + ], + "spans": [ + { + "bbox": [ + 68, + 237, + 265, + 248 + ], + "type": "text", + "content": "Table 6. Comparison between Lane2Seq and YOLOP." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 257, + 287, + 327 + ], + "lines": [ + { + "bbox": [ + 50, + 257, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 50, + 257, + 287, + 327 + ], + "type": "table", + "html": "
MethodsF1(%)↑Precision(%)↑Recall(%)↑
YOLOP (segmentation)75.2379.4660.58
Lane2Seq (segmentation)79.64 (+4.41)85.26 (+5.8)69.00 (+8.42)
YOLOP (anchor)76.4480.9962.15
Lane2Seq (anchor)79.27 (+2.83)84.72 (+3.73)67.28 (+5.13)
YOLOP (parameter)69.4374.3858.46
Lane2Seq (parameter)78.39 (+8.96)83.68 (+9.30)66.94 (+8.48)
", + "image_path": "73e73b5965c210a9eef740d96793ec307cfb847c6c8632f8966588948eba6598.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 334, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 357 + ], + "type": "text", + "content": "tion enables model to learn correlations between different detection formats and provides more supervised signals." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 358, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 358, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 287, + 537 + ], + "type": "text", + "content": "Comparison with other Multi-Task Methods. An alternative method to unify different detection formats is using different detection heads in a model. YOLOP [42] is a representative work for this method. We replace the heads of YOLOP with anchor-based head, segmentation-based head, and parameter-based head. The segmentation-based loss, anchor-based loss, parameter-based loss for YOLOP are same as that in SCNN [28], LaneATT [36], and PolyLaneNet [37]. From Table 6, we can observe that Lane2Seq surpasses YOLOP for all detection formats. It should be noted that YOLOP contains task-specific modules, such as detection head and loss function, but Lane2Seq does not. Results manifest the sequence generation combining with MFRL is a more simple and effective way to unify lane detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": "Ablation Study on the Scale Factor of Different Objective Functions. For the limitation of the space, we take the segmentation format as an example to present the influence of the scale factor of different objective functions in MFRL. Results of other two formats can be found in the supplementary materials. As shown in Table 7, assigning the same factor to different objective functions causes insuperior performance, indicating different detection formats have different contributions. We set " + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\lambda_5" + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\lambda_6" + }, + { + "bbox": [ + 46, + 538, + 287, + 658 + ], + "type": "text", + "content": " to 0.2, 1, and 1.5 according to the model performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "This paper presents a novel sequence generation-based lane detection framework, i.e., Lane2Seq, to unify lane de" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 257, + 545, + 436 + ], + "blocks": [ + { + "bbox": [ + 335, + 237, + 515, + 248 + ], + "lines": [ + { + "bbox": [ + 335, + 237, + 515, + 248 + ], + "spans": [ + { + "bbox": [ + 335, + 237, + 515, + 248 + ], + "type": "text", + "content": "Table 7. Ablation study on different scale factors." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 257, + 545, + 436 + ], + "lines": [ + { + "bbox": [ + 309, + 257, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 309, + 257, + 545, + 436 + ], + "type": "table", + "html": "
λ4λ5λ6F1(%)↑
11173.23
0.71174.96
0.41176.27
0.21177.35
0.11177.02
0.21.1177.03
0.21.2176.75
0.20.9177.09
0.20.8177.00
0.210.976.87
0.211.277.37
0.211.479.46
0.211.579.64
0.211.679.39
", + "image_path": "372aeb99d3350d5adc31c940bc4e18370f14593ba92733e671b64bf9e628351e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 544 + ], + "type": "text", + "content": "tection, which casts lane detection as a sequence generation task. Lane2Seq gets rid of complicated task-specific modules and adopts a simple transformer-based encoder-decoder architecture. To incorporate the task-specific knowledge, we employ a multi-format model tuning based on reinforcement learning (MFRL). Extensive experiments show Lane2Seq is effective and achieves competitive results compared to the state-of-the-art methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 544, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 546, + 652 + ], + "type": "text", + "content": "Limitation. One major limitation of Lane2Seq is that sequence generation model is expensive for long sequence (mainly for inference). The inference speed is low when there are more than 5 lanes in an image despite its detection strength in multi-lane scenario. Therefore, future work is required to make it faster for real-time lane detection applications. Another limitation is that MFRL can only be applied to sequence generation-based model currently. Our ongoing work is applying MFRL to other vision models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16951" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Tusimple dataset. https://github.com/TuSimple/tusimple-benchmark. Accessed on 11th August 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 192 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 239 + ], + "type": "text", + "content": "[3] Karsten Behrendt and Ryan Soussan. Unsupervised labeled lane markers using maps. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 295 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 341 + ], + "type": "text", + "content": "[5] Lluis Castrejon, Kaustav Kundu, Raquel Urtasun, and Sanja Fidler. Annotating object instances with a polygon-rnn. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5230-5238, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 373 + ], + "type": "text", + "content": "[6] Haoxin Chen, Mengmeng Wang, and Yong Liu. Bsnet: Lane detection via draw b-spline curves nearby. arXiv preprint arXiv:2301.06910, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 288, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 463 + ], + "type": "text", + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 508 + ], + "type": "text", + "content": "[8] Ting Chen, Saurabh Saxena, Lala Li, David J Fleet, and Geoffrey Hinton. Pix2seq: A language modeling framework for object detection. arXiv preprint arXiv:2109.10852, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 510, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 510, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 510, + 288, + 555 + ], + "type": "text", + "content": "[9] Ting Chen, Saurabh Saxena, Lala Li, Tsung-Yi Lin, David J Fleet, and Geoffrey E Hinton. A unified sequence interface for vision tasks. Advances in Neural Information Processing Systems, 35:31333-31346, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 610 + ], + "type": "text", + "content": "[10] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14572-14581, 2023. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Yuqing Du, Olivia Watkins, Zihan Wang, Cédric Colas, Trevor Darrell, Pieter Abbeel, Abhishek Gupta, and Jacob Andreas. Guiding pretraining in reinforcement learning with" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "large language models. arXiv preprint arXiv:2302.06692, 2023.4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[13] Ruochen Fan, Xuanrun Wang, Qibin Hou, Hanchao Liu, and Tai-Jiang Mu. Spinnet: Spinning convolutional network for lane boundary detection. Computational Visual Media, 5:417-428, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 195 + ], + "type": "text", + "content": "[14] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17062-17070, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 252 + ], + "type": "text", + "content": "[15] Jianhua Han, Xiajun Deng, Xinyue Cai, Zhen Yang, Hang Xu, Chunjing Xu, and Xiaodan Liang. Laneformer: Object-aware row-column transformers for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 799-807, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[16] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 342 + ], + "type": "text", + "content": "[17] Hiroto Honda and Yusuke Uchida. Clrernet: Improving confidence of lane detection with laneiou. arXiv preprint arXiv:2305.08366, 2023. 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "text", + "content": "[18] Dongkwon Jin, Wonhui Park, Seong-Gyun Jeong, Heeyeon Kwon, and Chang-Su Kim. Eigenlanes: Data-driven lane descriptors for structurally diverse lanes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17163-17171, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 453 + ], + "type": "text", + "content": "[19] Taeho Kil, Seonghyeon Kim, Sukmin Seo, Yoonsik Kim, and Daehee Kim. Towards unified scene text spotting based on sequence generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15223-15232, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[20] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 21(1):248-258, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 500, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 566 + ], + "type": "text", + "content": "[21] Xuan Liao, Wenhao Li, Qisen Xu, Xiangfeng Wang, Bo Jin, Xiaoyun Zhang, Yanfeng Wang, and Ya Zhang. Iteratively-refined interactive 3d medical image segmentation with multi-agent reinforcement learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9394-9402, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 621 + ], + "type": "text", + "content": "[22] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condlanenet: a top-to-down lane detection framework based on conditional convolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3773-3782, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 623, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 623, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 623, + 545, + 667 + ], + "type": "text", + "content": "[23] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 3694-3702, 2021. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 690 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mottaghi, and Aniruddha Kembhavi. Unified-io: A unified" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16952" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "model for vision, language, and multi-modal tasks. arXiv preprint arXiv:2206.08916, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[26] Wenhan Luo, Peng Sun, Fangwei Zhong, Wei Liu, Tong Zhang, and Yizhou Wang. End-to-end active object tracking via reinforcement learning. In International conference on machine learning, pages 3286-3295. PMLR, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 185 + ], + "type": "text", + "content": "[27] Stefan Mathe, Aleksis Pirinen, and Cristian Sminchisescu. Reinforcement learning for visual object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2894-2902, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 230 + ], + "type": "text", + "content": "[28] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 286 + ], + "type": "text", + "content": "[29] Dezhi Peng, Xinyu Wang, Yuliang Liu, Jiaxin Zhang, Mingxin Huang, Songxuan Lai, Jing Li, Shenggao Zhu, Dahua Lin, Chunhua Shen, et al. Spts: single-point text spotting. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4272-4281, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 287, + 330 + ], + "type": "text", + "content": "[30] André Susano Pinto, Alexander Kolesnikov, Yuge Shi, Lucas Beyer, and Xiaohua Zhai. Tuning computer vision models with task rewards. arXiv preprint arXiv:2302.08242, 2023. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 387 + ], + "type": "text", + "content": "[31] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIV 16, pages 276-291. Springer, 2020. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 422 + ], + "type": "text", + "content": "[32] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 478 + ], + "type": "text", + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 533 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 579 + ], + "type": "text", + "content": "[35] Richard S Sutton, David McAllester, Satinder Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. Advances in neural information processing systems, 12, 1999. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 645 + ], + "type": "text", + "content": "[36] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 294-302, 2021. 1, 2, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[37] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6150-6156. IEEE, 2021. 2, 6, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 688 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[38] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[39] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1392-1401, 2022. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 206 + ], + "type": "text", + "content": "[40] Ze Wang, Weiqiang Ren, and Qiang Qiu. Lanenet: Realtime lane detection networks for autonomous driving. arXiv preprint arXiv:1807.01726, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 239 + ], + "type": "text", + "content": "[41] Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8:229-256, 1992. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[42] Dong Wu, Man-Wen Liao, Wei-Tian Zhang, Xing-Gang Wang, Xiang Bai, Wen-Qing Cheng, and Wen-Yu Liu. Yolop: You only look once for panoptic driving perception. Machine Intelligence Research, 19(6):550-562, 2022. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[43] Lingyu Xiao, Xiang Li, Sen Yang, and Wankou Yang. Adnet: Lane shape prediction via anchor decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6404-6413, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[44] Jiaxing Yang, Lihe Zhang, and Huchuan Lu. Lane detection with versatile atrousformer and local semantic guidance. Pattern Recognition, 133:109053, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[45] Zhengyuan Yang, Zhe Gan, Jianfeng Wang, Xiaowei Hu, Faisal Ahmed, Zicheng Liu, Yumao Lu, and Lijuan Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision, pages 521-539. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 464 + ], + "type": "text", + "content": "[46] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 509 + ], + "type": "text", + "content": "[47] Jia-Qi Zhang, Hao-Bin Duan, Jun-Long Chen, Ariel Shamir, and Miao Wang. Houghlanenet: Lane detection with deep hough transform and dynamic convolution. arXiv preprint arXiv:2307.03494, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 545, + 543 + ], + "type": "text", + "content": "[48] Zhiyuan Zhao, Qi Wang, and Xuelong Li. Deep reinforcement learning based lane detection and localization. Neurocomputing, 413:328-338, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 598 + ], + "type": "text", + "content": "[49] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3547-3554, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 600, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 545, + 655 + ], + "type": "text", + "content": "[50] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 898-907, 2022. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "type": "text", + "content": "[51] Kunyang Zhou and Rui Zhou. End to end lane detection with one-to-several transformer. arXiv preprint arXiv:2305.00675, 2023. 2, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16953" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_content_list.json b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9f15ac216624f18cddfdf94621a1206fe51482e --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_content_list.json @@ -0,0 +1,1780 @@ +[ + { + "type": "text", + "text": "LaneCPP: Continuous 3D Lane Detection using Physical Priors", + "text_level": 1, + "bbox": [ + 165, + 130, + 803, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Maximilian Pittner $^{1,2}$ , Joel Janai $^{1}$ , Alexandru P. Condurache $^{1,2}$", + "bbox": [ + 233, + 179, + 733, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Bosch Mobility Solutions, Robert Bosch GmbH", + "bbox": [ + 294, + 198, + 678, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Institute of Signal Processing, University of Lübeck", + "bbox": [ + 277, + 215, + 692, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{Maximilian.Pittner, Joel.Janai, AlexandruPaul.Condurache}@de.bosch.com", + "bbox": [ + 169, + 236, + 795, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 310, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors.", + "bbox": [ + 76, + 316, + 472, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 694, + 207, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robust and precise lane detection systems build one of the most essential components in the perception stack of autonomous vehicles. While some approaches utilize LiDAR sensors or multi-sensor setups, the application of monocular cameras has become more popular due to their lower cost and the high-resolution visual representation that provides valuable information to detect lane markings.", + "bbox": [ + 76, + 719, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the past, lane detection was mainly treated as a 2D detection task. Deep learning based methods achieved good results by treating the problem as a segmentation task in pixel space [7, 11, 16, 26, 29, 33, 53], used to classify and regress lanes using anchor-based [19, 41] representations,", + "bbox": [ + 76, + 825, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "or as key-points on a grid structure [13, 15, 34, 45]. However, due to the lack of depth information, these 2D representations fail to model lane markings and road geometry in 3D space, which forms an important prerequisite for later functionalities like trajectory planning. Consequently, approaches for monocular 3D lane detection were introduced, which adapted lane representations for the 3D domain by modeling vertical anchors [6, 9] or local segments on a grid [4] in a Birds-Eye-View (BEV) oriented 3D-frame.", + "bbox": [ + 496, + 287, + 890, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A crucial topic for the application of lane detection algorithms in autonomous systems is safety, which requires predictable and robust behavior in any traffic situation. One risk of learning-based methods is the tendency to show unpredictable behavior in cases of rarely observed scenarios. Since obtaining large amounts of data with high-quality annotations is cumbersome and expensive, publicly available 3D datasets are limited in size and accuracy. Hence, they do not reflect the variability of real-world scenarios sufficiently. This makes learning-based models prone to overfitting, and eventually, diminishes predictability.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One common way to deal with such problems is the integration of prior knowledge. Physics provides us a profound understanding of the 3D world, allowing us to make valid assumptions about the lane structure and road surface geometry. Therefore, we introduce physically motivated priors into the lane detection objective to cope with the limited data problem and achieve robust and predictable behavior.", + "bbox": [ + 496, + 595, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are certain geometric properties that should generally hold for detected lane lines. For instance, we know that most lines progress parallel to each other, reside on a smooth surface and should not exceed certain thresholds in terms of curvature and slope. However, integrating such assumptions into prevailing discrete representations is not straight forward as strong simplifications are necessary. In contrast, continuous 3D lane representations directly provide parametric curves using polynomials [1, 23] or more sophisticated B-Splines [32]. These allow for analytical computations on the curve function, which enables the integration of such priors into the lane representation. By modeling these priors explicitly instead of learning them from", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10639", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "data, the model can focus its full capacity on learning richer features for the lane detection task.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We can further use physical knowledge about the road geometry to support the model in learning an internal transformation from image features to 3D space. While methods based on Inverse Perspective Mapping (IPM) [4, 6, 9, 17, 23, 32] make false flat-ground assumptions, learning based transformations [1, 2, 46] completely ignore road properties. In contrast, integrating prior knowledge about the road surface allows us to model 3D features geometry-aware and helps the network to focus on the 3D region of interest.", + "bbox": [ + 75, + 121, + 467, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, we propose a novel 3D lane detection approach named LaneCPP that leverages valuable prior knowledge to achieve accurate and robust perception behavior. It introduces a new sophisticated continuous curve representation, which enables us to incorporate physical priors. In addition, we present a spatial transformation component for learning a physically inspired mapping from 2D image to 3D space providing meaningful spatial features.", + "bbox": [ + 75, + 257, + 467, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 76, + 378, + 437, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel architecture for 3D lane detection from monocular images using a more sophisticated flexible parametric spline-based lane representation.", + "- We present a way to incorporate priors about lane structure and geometry into our continuous representation.", + "- We introduce a new way to use prior knowledge about the road surface geometry for learning spatial features.", + "- We demonstrate the benefits of our contributions in several ablation studies.", + "- We show state-of-the-art performance of our model." + ], + "bbox": [ + 76, + 393, + 467, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 558, + 212, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different Lane Representations. An important design choice in deep learning based lane detection is the representation that the network uses to model lane line geometry, which can be categorized as follows: 1) Pixelwise representations, which formulate lane detection as a segmentation problem, were used mainly in 2D methods [7, 11, 16, 26, 29, 33, 51, 53] and were adopted in 3D by SALAD [50] combining line segmentation with depth-prediction. These representations come with high computational load since a large amount of parameters is required. 2) Grid-based approaches divide the space into cells and model lanes using local segments [13] or keypoints [15, 34, 45]. 3D-LaneNet+ [4] suggests to use local line-segments and BEV-LaneDet [46] defines key-points on a BEV grid representation. Both depend on the grid resolution and require costly post-processing to obtain lines. 3) Anchor-based representations [19, 40, 41, 52] model lines as straight anchors with positional offsets at predefined locations. They are widely used in 3D detection approaches including 3D-LaneNet [6] and Gen-LaneNet [9], which use vertical anchors in the top-view, and An", + "bbox": [ + 75, + 583, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "chor3DLane [12], introducing anchor projection with iterative regression. Similar to grid-based representations, it requires subsequent curve-fitting to obtain smooth lines. 4) Continuous curve representations [5, 22, 24, 43, 44] instead directly model smooth curves without requiring costly post-processing. While CLGO [23] and CurveFormer [1] use simple polynomials, 3D-SpLineNet [32] proposes B-Splines [3]. Since B-Splines offer local control over curve segments, they are compatible to model complex shapes with low-degree basis functions, while polynomials and Bezier curves show global dependence and thus require higher degrees causing expensive computation. Although 3D-SpLineNet achieves superior detection performance on synthetic data, it unfortunately lacks flexibility as the curve formulation is limited to monotonically progressing lanes, making it hardly applicable to real-world data. To resolve this issue, we propose a more flexible representation based on actual 3D B-Splines. In contrast to discrete grids and anchors, continuous representation even allows us to integrate prior knowledge in an analytical manner.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Geometry Priors. Several approaches suggest to incorporate prior knowledge into learning-based methods, e.g. by integrating invariance into the model architecture [35, 36] or task-specific transformations as for trajectory planning [10, 47, 49]. In the field of lane detection, line parallelism has been formulated as a hard constraint to resolve depth ambiguity and determine camera parameters [27, 48]. Deep declarative networks [8] offer a general framework to incorporate arbitrary properties as constraints, by solving a constrained optimization problem in the forward pass. While such methods are appropriate when hard constraints must be enforced, our goal is rather to guide the network in learning typical geometric lane properties by formulating soft constraints in a regularization objective. Such a regularization only affects training and does not require resolving an optimization problem in the forward pass, and thus, comes without additional computational cost during inference. Following this paradigm, SGNet [24, 40] proposes to penalize the deviation of lateral distance from a constant lane width in the IPM warped top-view, but ignores that the property does not hold for lines deviating from the ground plane. GP [17] presents a parallelism loss that enforces constant distance between nearest neighbors locally, which depends on the number of anchor points. In contrast, our method presents a way to learn parallelism globally and independent of resolutions of discrete lane representations. We propose an elegant way to learn parallelism as well as other geometry priors using analytical formulations of tangents and normals, which are well-defined on our continuous spline representation.", + "bbox": [ + 496, + 397, + 890, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging 3D Features. An important model component consists in the extraction of 3D features, encoding valuable information to detect lanes along the road surface.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10640", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8c794e6cb38623d03b88c4f43ce7126e1c660d926c4bc4c53cce5ce995d69c5e.jpg", + "image_caption": [ + "Figure 1. Our approach: First, front-view image $I$ is propagated through the backbone extracting multi-scale feature maps. These are transformed to 3D using our spatial transformation and then fused to obtain a single 3D feature map. Feature pooling is applied to obtain features for each line proposal that are propagated through fully connected layers to obtain the parameters for our line representation. Finally, prior knowledge is exploited to regularize the lane representation and to produce surface hypotheses for the spatial transformation." + ], + "image_footnote": [], + "bbox": [ + 83, + 87, + 885, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While some works predict 3D lanes directly from the front-view, e.g. by utilizing pixel-wise depth estimation [50] or 3D anchor-projection mechanisms [12], prevalent methods employ an intermediate 3D or BEV feature representation with an internal transformation from the front-view to the 3D space. 3D-LaneNet [6] proposes to utilize IPM [25] to project front-view features to a flat road plane due to the spatial correlation between the warped top-view image and 3D lane geometry and was adopted in several other works [4, 9, 17, 23, 32]. However, IPM causes visual distortions in the top-view representation when the flat road assumption is violated. In related fields like BEV semantic segmentation, BEV transformations are learned via Multi-Layer-Perceptrons (MLPs) [18, 28], depth prediction [31, 37, 38] or transformer-based attention mechanisms [20, 30, 39]. In 3D lane detection, PersFormer [2] utilizes attention between front- and top-view, CurveFormer [1] introduces dynamic 3D anchors that model queries as parametric curves and BEV-LaneDet [46] uses MLPs for the spatial transformations. However, these learned transformations do not necessarily provide a 3D feature representation since they are not guided by valuable priors about the road surface geometry, which potentially results in unforeseen behavior for out-of-distribution data. Our approach instead aims for carefully modeling a geometry-aware feature space using a depth classification method inspired by [31] that exploits knowledge about the distribution of the road surface.", + "bbox": [ + 75, + 373, + 472, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 797, + 212, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following section describes our 3D lane detection approach. An overview of the overall architecture is described and illustrated in Fig. 1. The main focus lies on our continuous 3D lane line representation, our regularization mechanism using physical priors and our prior-based spatial trans", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/988c2dc4d8c350930fdb1f68812b31a63b7e9786bb66de69fea9c2e8a4452c32.jpg", + "image_caption": [ + "Figure 2. Our 3D lane line representation: For each proposal $\\bar{f}$ (purple lines), line geometry is described by 3D B-Splines with control points $c_{k}$ (green dots). Each control point is determined by the offsets $\\alpha_{k},\\beta_{k}$ from the control points of the initial proposal in normal direction (orange vectors). Additionally, visibility $v(t)$ is modeled by splines with 1D control points $\\gamma_{k}$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 369, + 883, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formation module, which we explain in the following.", + "bbox": [ + 500, + 714, + 857, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Lane line representation", + "text_level": 1, + "bbox": [ + 500, + 751, + 723, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by prior work in 3D lane detection [32], we leverage the benefits of continuous representations and employ a parametric model based on B-Splines. However, modeling only lateral $(x-)$ and vertical $(z-)$ components with spline-based functions (as done in previous approaches) is limited to lanes that merely progress along the longitudinal $(y-)$ direction. Instead, we propose the first full 3D lane line representation modeling each component $(x, y, z)$ such that we", + "bbox": [ + 496, + 779, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10641", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "obtain", + "bbox": [ + 76, + 90, + 124, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} (t) = \\left( \\begin{array}{c} x (t) \\\\ y (t) \\\\ z (t) \\end{array} \\right) = \\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {k} \\cdot B _ {k, d} (t) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 112, + 468, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with curve argument $t \\in [0, 1]$ and $K$ control points $c_k = (x_k, y_k, z_k)^T$ . Each control point $c_k$ weights the respective basis function $B_{k,d}(t)$ (recursive polynomials of degree $d$ ) controlling the curve shape.", + "bbox": [ + 76, + 169, + 468, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the ambiguity of curves using 3D B-Splines (the same spline curve can be described by different configurations of its control points), regressing all three dimensions per control point results in strong overfitting during training. We resolve this issue by limiting the degrees of freedom per control point to two and constraining the control points deflection to one direction in the $x - y$ -plane and one direction in the $y - z$ -plane as illustrated in Fig. 2. More precisely, the degrees of freedom per control point are specified by the directions of the normals $\\mathbf{N}_{xy}$ and $\\mathrm{N}_z$ of an initial curve proposal $\\bar{f}$ with control points $\\bar{c}_k = (\\bar{x}_k,\\bar{y}_k,\\bar{z}_k)^T$ . The control points are then defined as", + "bbox": [ + 75, + 233, + 468, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {c} _ {k} = \\left( \\begin{array}{l} x _ {k} \\\\ y _ {k} \\\\ z _ {k} \\end{array} \\right) = \\left( \\begin{array}{l} \\bar {x} _ {k} + \\mathrm {N} _ {x} \\cdot \\alpha_ {k} \\\\ \\bar {y} _ {k} + \\mathrm {N} _ {y} \\cdot \\alpha_ {k} \\\\ \\bar {z} _ {k} + \\mathrm {N} _ {z} \\cdot \\beta_ {k} \\end{array} \\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 426, + 468, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{N}_x, \\mathrm{N}_y$ describe the $x$ - and $y$ -component of the normal vector $\\mathbf{N}_{xy}$ in the $x$ - $y$ -plane. As shown in Eq. (2) and illustrated in Fig. 2, modeling splines as deflections in normal direction of its underlying initial line proposal only requires two parameters $\\alpha_k, \\beta_k$ per control point to describe the 3D shape. We use a wide variety of orientations for the initial proposals $\\bar{\\pmb{f}}$ (see Fig. 2), which allows us to detect any kind of lines with this formulation. More details about the initial proposals are provided in the supplementary.", + "bbox": [ + 75, + 482, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While [32] models the curve range using start- and endpoints that are learned by means of regression, we instead propose to model visibility using a continuous representation $v(t)$ and treat the visibility estimation as a classification problem. We obtain probability values applying sigmoid activation and consider $\\sigma(v(t)) > 0.5$ the visible range. While in theory any kind of function can be utilized, we found that B-Splines with the same configuration as $f(t)$ are well-suited and introduce spline control points $\\gamma_k$ defining the shape of $v(t)$ .", + "bbox": [ + 75, + 619, + 468, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eventually, binary cross-entropy is used as a classification loss to learn visibility", + "bbox": [ + 76, + 768, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v i s} = - \\frac {1}{| \\mathcal {P} _ {G T} |} \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} _ {G T}} \\hat {v} _ {\\boldsymbol {p}} \\cdot \\log (\\sigma (v (t _ {\\boldsymbol {p}}))) + \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 808, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left(1 - \\hat {v} _ {\\boldsymbol {p}}\\right) \\cdot \\log \\left(1 - \\sigma \\left(v \\left(t _ {\\boldsymbol {p}}\\right)\\right)\\right), \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 848, + 468, + 867 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ec2e5bc323da2a922bc627d2bc7fa1a7741f98fc7dc54044906dc1789c137bed.jpg", + "image_caption": [ + "Figure 3. Illustration of different priors expressed by line tangents and surface normals." + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 882, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_{GT}$ denotes the ground truth set of points, $\\hat{v}_p\\in \\{0,1\\}$ the ground truth visibility for point $p$ . $t_p$ represents the respective curve argument obtained by orthogonal projection of $p$ onto the underlying line proposal.", + "bbox": [ + 498, + 292, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Regularization using physical priors", + "text_level": 1, + "bbox": [ + 498, + 363, + 815, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we describe our regularization method to integrate prior knowledge about lane structure and surface geometry into our parametric line representation (see Fig. 3).", + "bbox": [ + 496, + 386, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Line parallelism. In order to reinforce parallel lines, the tangents at point pairs located in opposite normal direction on neighboring lines must be similar (see Fig. 3 left). We realize this by penalizing the cosine distance of the unit tangents $\\mathbf{T}(t)$ on neighboring lines $i$ and $j$ for normal point pairs. More precisely, for each point $\\pmb{p} \\in \\mathcal{P}^{(i)}$ on line $i$ we select the normal pair point $\\pmb{p}^*$ on neighbor line $j$ that minimizes the distance to the normal plane, which is defined by the plane equation $\\mathbf{T}^{(i)}(t)^T \\cdot ((x,y,z)^T - \\pmb{f}^{(i)}(t)) = 0$ . In Fig. 3 the normal planes are visualized in a 2D top-view as lines (orange) for simplicity. Hence the respective curve argument $t_{\\pmb{p}^*}$ for point $\\pmb{p}^*$ on line $j$ is given as", + "bbox": [ + 496, + 431, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\boldsymbol {p} ^ {*}} = \\underset {\\boldsymbol {p} ^ {\\prime} \\in \\mathcal {P} ^ {(j)}} {\\operatorname {a r g m i n}} \\left. \\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right) ^ {T} \\cdot \\left(\\boldsymbol {f} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {\\prime}}\\right) - \\boldsymbol {f} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right), \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 625, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{P}^{(j)}$ denotes the points on line $j$ . While in theory Eq. (5) can be solved analytically, the simpler way is to sample the set of points $\\mathcal{P}^{(j)}$ instead. (Note that our continuous representation allows us to choose high sampling rates without losing precision as no interpolation is required.)", + "bbox": [ + 496, + 666, + 890, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the normal point pairs, we define the parallelism loss for a neighbor line pair based on the cosine distance of their tangents as", + "bbox": [ + 496, + 743, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p a r} ^ {(i j)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {T} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {*}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 799, + 890, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the criterion of line parallelism should not hold for all normal point pairs of neighboring lines (e.g. merging or splitting lines), $\\mathbb{1}_{\\pmb{p}}^{(ij)} \\in \\{0, 1\\}$ represents the indicator", + "bbox": [ + 496, + 852, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "For the concept of visibility, we follow the prevailing definition from the literature [2, 9].", + "bbox": [ + 76, + 875, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10642", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "function determining whether the parallelism loss is applied to the point pair. More precisely, the function ensures that only the overlapping range of neighboring lines is taken into account. Furthermore, it determines whether the line pair should be considered as a parallel pair based on the standard deviation of euclidean distances between normal point pairs, i.e. high deviations indicate that the line pair might belong to a merge or split structure. In our experiments, we achieve state-of-the-art performance on test sets containing merges and splits, proving that our model is also capable of learning non-parallel line pairs using this indicator function.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Surface smoothness. Since the lines reside on a smooth road, the surface normals of neighboring lanes should be similar. Analogously to $\\mathcal{L}_{par}$ , we express this with the cosine distance between surface normals $\\mathbf{N}^{(ih)}$ and $\\mathbf{N}^{(ij)}$ as", + "bbox": [ + 75, + 257, + 470, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s m} ^ {(i)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(h i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {N} ^ {(i h)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {N} ^ {(i j)} \\left(t _ {\\boldsymbol {p}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 334, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with indicator function $\\mathbb{1}_{\\pmb{p}}^{(hij)}$ . The surface normal between line $i$ and left neighbor line $h$ at point $\\pmb{p}$ can be expressed as the cross product of the tangent on line $i$ and the normalized connection vector between lines $i$ and $h$ , hence $\\mathbf{N}^{(ih)}(t_{\\pmb{p}}) = \\mathbf{T}^{(i)}(t_{\\pmb{p}}) \\times \\frac{\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})}{||\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})||}$ . For the normal between line $i$ and right neighbor $j$ the sign is flipped to obtain upwards pointing normal vectors.", + "bbox": [ + 76, + 386, + 468, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Curvature.** We determine lane curvature by computing the second order derivatives as the difference of tangents at consecutive points divided by their euclidean distance as $\\mathbf{T}'(t_p) = \\frac{\\mathbf{T}(t_p) - \\mathbf{T}(t_p - \\Delta t)}{||\\mathbf{f}(t_p) - \\mathbf{f}(t_p - \\Delta t)||}$ . The maximum curvature in $x-y$ -plane (inverse curve radius) and in $z$ (rate of slope change) have very different value ranges and are therefore restricted by different limits. Hence, we define the two thresholds $\\kappa_{xy}$ and $\\kappa_z$ and formulate the curvature loss on line $i$ as", + "bbox": [ + 75, + 502, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {c u r v} ^ {(i)} = \\frac {1}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} \\max \\left(\\mathrm {T} _ {x y} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {x y}\\right) (8) \\\\ + \\max \\left(\\mathrm {T} _ {z} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {z}\\right). (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 643, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the prior regularization loss is given as", + "bbox": [ + 76, + 709, + 387, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r i o r}} = \\sum_ {i = 1} ^ {M} \\lambda_ {s m} \\mathcal {L} _ {s m} ^ {(i)} + \\lambda_ {\\text {c u r v}} \\mathcal {L} _ {\\text {c u r v}} ^ {(i)} + \\sum_ {j = 1} ^ {N} \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} ^ {(i j)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 729, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with individual weights $\\lambda_{par}$ , $\\lambda_{sm}$ , $\\lambda_{curv}$ . Note that all these properties are expressible by means of tangents and normals, which can be computed analytically on our parametric representation in continuous space. Consequently, minimization of the herein introduced prior losses does not depend on numerical approximations as is the case for anchor-, grid- or key-point representations.", + "bbox": [ + 75, + 794, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1de4bcab885767d964f33dfd0955e89c71c85842a7b3cf317c0e6e7a880bf5a.jpg", + "image_caption": [ + "Figure 4. Our proposed spatial transformation module. First, several road surface hypotheses are defined (a) to which front-view features are lifted (b) and weighted according to the predicted depth distribution. Afterwards, point features are aggregated in a weighted manner to obtain the 3D feature map (c)." + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 883, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spatial transformation", + "text_level": 1, + "bbox": [ + 500, + 416, + 710, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe our spatial transformation (shown in Fig. 4) that is leveraging valuable physical knowledge about surface geometry. We know that the road surface typically shows small deviations from the ground level $(z = 0)$ in the near-range and stronger deviations in the far-range. Based on this knowledge, we sample ground surface hypotheses that reflect the distribution of the road surface height profile (Fig. 4a). While in theory different types of surface functions could be utilized as hypotheses, we decide to merely rely on planes, since this facilitates the computation of ray intersections described in the following step.", + "bbox": [ + 496, + 441, + 890, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, the multi-scale front-view feature maps extracted by the backbone are lifted to 3D space (Fig. 4b). Our approach is inspired by [31], where front-view features are spreading along rays throughout the space of the road surface. These rays intersect with the surface hypotheses at different depths spanning a frustum-like point cloud in 3D space, where each point is affiliated with a $C$ -dimensional feature vector and additionally attached with its height value $z$ , hence, each point in the cloud has dimension $(C + 1)$ . The front-view feature map is propagated through a depth branch with a channel-wise softmax applied to obtain a categorical distribution for each ray, resulting in a tensor of size $H\\times W\\times S$ , where $H$ , $W$ denote height and width and channel size $S$ the number of surface hypotheses.", + "bbox": [ + 496, + 609, + 892, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to aggregate the information in 3D space, a BEV grid of size $X \\times Y$ is defined. Features from points mapping to the same grid cell are weighted by the categorical depth distribution for the respective ray and accumulated in terms of a weighted sum (Fig. 4c). Since the $z$ -component", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/94752f31db4f17f6b15cdfcd30c94b725916952e8dd449c2b1457992486f2bbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PriorsF1(%)↑X-near(m)↓X-far(m)↓Z-near(m)↓Z-far(m)↓
None65.00.3160.3840.1060.153
Par.66.20.2910.3730.1030.150
Surf.65.80.3200.3560.1030.144
Curv.66.70.3220.3660.1050.146
Comb.66.70.3010.3590.1030.144
", + "bbox": [ + 81, + 88, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/27bf1a0ffaabed838f141746171d0d04cab441ca857bc9bffbeb6cdf47866f9b.jpg", + "table_caption": [ + "Table 1. Effect of different prior losses on OpenLane300." + ], + "table_footnote": [], + "table_body": "
# Surface Hypotheses1351527
F1-Score(%)↑65.065.966.666.166.0
", + "bbox": [ + 81, + 205, + 467, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7cd9b6523d7260208023e39deca678805e87474317985f9e3e7b7a643c4a039d.jpg", + "table_caption": [ + "Table 2. Effect of the surface hypotheses on OpenLane300." + ], + "table_footnote": [], + "table_body": "
Lane Rep.Prior Reg.Spatial T.F1(%)↑Gain(%)
62.9(baseline)
✓65.0+2.1
✓✓66.7+3.8
✓✓66.6+3.7
✓✓✓66.9+4.0
", + "bbox": [ + 81, + 284, + 467, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Performance gain for different contributions on OpenLane300 using our novel Lane Representation, Prior Regularization and Spatial Transformation instead of IPM.", + "bbox": [ + 75, + 387, + 468, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the points is also combined by a weighted sum, the value $z_{uv}$ can be interpreted as the height value of the surface for grid cell $(u, v)$ . We guide the model in learning the real surface and prevent it from learning an arbitrary mapping by introducing a simple grid-based regression loss as", + "bbox": [ + 75, + 458, + 468, + 535 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s u r f}} = \\frac {1}{X \\cdot Y} \\sum_ {(u, v) \\in X \\times Y} \\mathbb {1} _ {u v} \\cdot \\| z _ {u v} - \\hat {z} _ {u v} \\| _ {1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 545, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with $\\mathbb{1}_{uv}$ indicating whether surface ground truth $\\hat{z}_{uv}$ is available for cell $(u,v)$ . The height ground truth is obtained by interpolation of the 3D lane annotations at cell locations.", + "bbox": [ + 75, + 595, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Loss functions", + "text_level": 1, + "bbox": [ + 76, + 652, + 225, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The overall loss used during training is given as the weighted sum of loss components", + "bbox": [ + 75, + 676, + 468, + 708 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {c a t} \\mathcal {L} _ {c a t} + \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 719, + 468, + 738 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {v i s} \\mathcal {L} _ {v i s} + \\lambda_ {p r i o r} \\mathcal {L} _ {p r i o r} + \\lambda_ {s u r f} \\mathcal {L} _ {s u r f}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 739, + 468, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use focal loss [21] for lane presence $\\mathcal{L}_{pr}$ and category classification $\\mathcal{L}_{cat}$ . For the regression loss $\\mathcal{L}_{reg}$ , we adapt the formulation of [32] to three instead of two dimensions. More details are provided in the supplementary.", + "bbox": [ + 75, + 768, + 468, + 829 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first describe our experimental setup and then analyze our approach on two 3D lane datasets.", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8f50475f56b7aba490beb4d19e4267cd4ec7c906cfdcffb6c0cc88bc6a4a7a5.jpg", + "image_caption": [ + "Front-view" + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 686, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ef57ac9d6f77182779fcca8b7d4ad9ac45485f1d14a7e2bdc9683598ac864d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 175, + 687, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1c6df6aaf53bfd0000d1ffa8f888fe304f8da0a3b5ecdab4015a24ae931f747.jpg", + "image_caption": [ + "Front-view" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 870, + 169 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cea5e6527d9d5251b5eb2a2e3f1699a0379f260c07652769b87f01b4b783d563.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 176, + 870, + 279 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2139bb034d4a0cc9d059fbf2d02f4db9fbf7aae7d5502c6c9657ecdc18d131da.jpg", + "image_caption": [ + "(a)", + "Figure 5. Qualitative comparison of our model trained with prior regularization to the same model without regularization both trained on OpenLane300 with main differences highlighted by arrows. As a reference ground truth lines are visualized dashed." + ], + "image_footnote": [], + "bbox": [ + 522, + 280, + 689, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a4b18d813c32b0f089a8931428c4b75653410ced963493e011883d811c1096a.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 702, + 280, + 870, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 498, + 445, + 687, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on two different datasets: OpenLane and Apollo 3D Synthetic - both containing 3D lane ground truth as well as camera parameters per frame.", + "bbox": [ + 496, + 470, + 890, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OpenLane [2] is a real-world dataset containing 150,000 images in the training and 40,000 in the test set from 1000 different sequences. In order to evaluate different driving scenarios the test set is divided into different situations, namely Up & Down, Curve, Extreme Weather, Night, Intersection and Merge & Split. For ablation studies we use the smaller version OpenLane300 including 300 sequences.", + "bbox": [ + 496, + 517, + 890, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Apollo 3D Synthetic [9] is a small synthetic dataset, consisting of only 10,500 examples from rather simple scenarios of highway, urban and rural environments. The data is split into three subsets, (1) Standard (simple) scenarios, (2) Rare Scenes and (3) Visual Variations.", + "bbox": [ + 496, + 625, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. For the quantitative evaluation both datasets utilize the evaluation scheme proposed in [9]. It evaluates the euclidean distance at uniformly distributed points in the range of $0 - 100\\mathrm{m}$ along the $y$ direction. Based on the mean distance and range, $Fl$ -Score is computed, as well as the mean $x$ - and $z$ -errors in near- $(0 - 40\\mathrm{m})$ and far-range $(40 - 100\\mathrm{m})$ to evaluate geometric accuracy.", + "bbox": [ + 496, + 700, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. Our approach builds up on 3D-SpLineNet. Since it was applied on synthetic data only, it showed poor performance on real data. We applied some straightforward design adaptations - e.g. larger backbone, multiscale features (see supplementary) - and use this modified 3D-SpLineNet as our baseline (first row Table 3).", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10644", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e93d21a97eaa0753b2d0fa89ac87470188d08fbe57f72f2bb10e820facc0e92.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [6]44.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [9]32.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [2]50.50.4850.5530.3640.43142.455.648.646.640.050.7
PersFormer* [2]53.10.3610.3280.1240.12946.858.754.048.441.452.5
CurveFormer [1]50.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [46]58.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane [12]53.70.2760.3110.1070.13846.757.252.547.845.451.2
Anchor3DLane-T [12]54.30.2750.3100.1050.13547.258.052.748.745.851.7
LaneCPP (Ours)60.30.2640.3100.0770.11753.664.456.754.952.058.7
", + "bbox": [ + 81, + 88, + 893, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparison on OpenLane [2]. Best performance and second best are highlighted. The scenario categories are Up and Down (U&D), Curve (C), Extreme Weather (EW), Night (N), Intersection (I), Merge and Split (M&S). PersFormer* denotes the latest performance reported on the official code base, Anchor3DLane-T represents the temporal multi-frame method of [12].", + "bbox": [ + 75, + 255, + 892, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/448e72ae7ae1e35ac542cb83f5290e47082516b13502ec82a11e9167a004a195.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 311, + 235, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1025d877499044a6cc701c3d2d84f82a0c0639438fb8d45a60dde87feff7ba5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 378, + 233, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5cead738cf41e319a1183663c7702360da8a0769fc98d72bc06f6a98060f51c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 245, + 311, + 393, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c005d70086866981a42937a877a761e0a9f9d9e9d4bf15da9d840ce28b4d4bbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 382, + 395, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/448dd77f44973c75ff32a31560e0c84bacf7755cb35e8c408d8da6b6f8dd5a4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 311, + 557, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fe17f78161de68b46601a22002c94452a439b37f0c1816b160f56bc0aefb07e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 380, + 558, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3671b8e28ca4ab8accad2368143da3e9f789f3d40145ed2fe46ba41a2fac6ecf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 311, + 720, + 377 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e09f8a23b45c6c49babfd81ae9f26a97c28460450f73e1400ce75e9bafa98d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 380, + 723, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96d046c2e04dba0541828eec494db8b826bcfb1223f1acdc07c63f6ccaaa3bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 311, + 885, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec75a55116b88d66845331c61fccb2a12d696d00a92d2dcf24ae1c56b35013b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 380, + 887, + 477 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3444e3b48b655990b1f0b25d82dccb6d9bdbf7f92086c3c390b68c09491cc803.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 479, + 238, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fadd5d84a5837a9291f902c12685dd462fd31009a7857bd92afd2defd26bdb7.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 241, + 481, + 398, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c56e09f51d0bd0fbacc6de1527732c23acc442a6b853e08a026b3ccbce670bff.jpg", + "image_caption": [ + "(c)", + "Figure 6. Qualitative comparison on OpenLane. Our method is compared to PersFormer* with ground truth visualized as dashed lines." + ], + "image_footnote": [], + "bbox": [ + 405, + 479, + 560, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ae2c4c7081a6fb245227744324630c8fa59e554a633a2aa96ba4a3df74e7bd6.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 566, + 481, + 725, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0125e786f8849c86237e2b209555ad669774de6694d2f9697a7804128f96aa5.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 730, + 481, + 888, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. We use input size $360 \\times 480$ and adopt the same backbone as in [2] based on a modified EfficientNet [42]. We extract four feature maps of resolutions $\\left[\\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}\\right]$ . The final 3D feature map has size $26 \\times 16$ with 64 channels. We use $M = 64$ initial line proposals and B-Splines of degree $d = 3$ and $K = 10$ control points. We apply Adam optimizer [14] with an initial learning rate of $2 \\times 10^{-4}$ for OpenLane and $10^{-4}$ for Apollo and a dataset specific step-wise scheduler. We train for 30 epochs on OpenLane and 300 epochs on Apollo with batch size 16. For more details we refer to the supplementary.", + "bbox": [ + 75, + 595, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 771, + 236, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 indicates the effect of our proposed prior-based regularization. It is evident that each prior improves the F1-Score as well as geometric errors. While the surface and curvature priors result in better far-range estimates, line parallelism supports X-regression in the near-range. Besides, using surface smoothness loss results in lowest Z-far errors. Finally, a combination of priors yields a good balance of", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "F1-Score and geometric errors. The positive effect of parallelism is confirmed by Fig. 5, where reinforcing parallel lane structure leads to better estimates in the near-range (a) and far-range (b) compared to the unregularized model. Learning parallel lines also is evidently beneficial in cases of poor visibility (b) and occlusions (a). In the latter case, the regularized model even shows better predictions than the noisy ground truth. This emphasizes the high relevance of priors for more robust behavior for real-world datasets, where 3D ground truth often comes with inaccuracies.", + "bbox": [ + 496, + 595, + 890, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the spatial transformation (see Table 2), too low numbers of surface hypotheses result in worse score, presumably as 3D geometry is not captured sufficiently, whereas larger numbers tend to decreasing performance due to the higher complexity. The best F1-Score is obtained with 5 hypotheses, which is chosen for further experiments. While the improvement over IPM is already considerable, we think that with the simplifications of plane hypotheses prevent the component from developing its full potential. We see ways to enhance the 3D transformation even further", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10645", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b158c4f590d85dde71ffbeda325a5cda2faee6f167cff4716673f55fd79306a4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBalanced ScenesRare Scenes
F1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfarF1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfar
3D-LaneNet [6]86.40.0680.4770.0150.20272.00.1660.8550.0390.521
GP [17]91.90.0490.3870.0080.21383.70.1260.9030.0230.625
PersFormer [2]92.90.0540.3560.010.23487.50.1070.7820.0240.602
3D-SpLineNet [32]96.30.0370.3240.0090.21392.90.0770.6990.0210.562
CurveFormer [1]95.80.0780.3260.0180.21995.60.1820.7370.0390.561
BEV-LaneDet [46]96.90.0160.2420.020.21697.60.0310.5940.0400.556
Anchor3DLane [12]95.40.0450.3000.0160.22394.40.0820.6990.0300.580
LaneCPP (Ours)97.40.0300.2770.0110.20696.20.0730.6510.0230.543
", + "bbox": [ + 81, + 88, + 890, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparison of best methods on Apollo 3D Synthetic [9]. Best performance and second best are highlighted.", + "bbox": [ + 96, + 279, + 870, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "using more sophisticated spatial representations in future.", + "bbox": [ + 76, + 321, + 455, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of our different contributions is summarized in Table 3, where the first row shows our baseline (see Sec. 4.1). More than two percent in F1-Score are gained with our novel lane representation compared to the simplified one from [32]. Moreover, it is clear that both, the regularization using combined priors and the spatial transformation using 5 hypotheses result in significant improvement. Eventually, combining all components yields the best model configuration, which we choose for further evaluation.", + "bbox": [ + 75, + 335, + 468, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Evaluation on OpenLane", + "text_level": 1, + "bbox": [ + 76, + 483, + 307, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the real-world OpenLane benchmark our model evidently outperforms all other methods with respect to F1-Score as well as geometric errors as shown in Table 4. Compared to BEV-Lanedet, which achieves a high detection score, our model gains $+1.9\\%$ , while reaching significantly lower geometric errors. In comparison to Anchor3DLane the improvements with respect to X-errors are less substantial, however, our approach surpasses the F1-Score by a large gap of $+6.6\\%$ . Analyzing the detection scores among different scenarios, outstanding performance gain is observed on the up- and down-hill test set $(+5.9\\%)$ that highlights the capability of our approach to capture 3D space proficiently, which is supported by the low Z-errors.", + "bbox": [ + 75, + 507, + 468, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Apart from quantitative results, we show qualitative examples in Fig. 6. In up-hill scenarios like Fig. 6b our model manages to estimate both lateral and height profile accurately, since our assumptions about road surface and line parallelism are satisfied. In contrast, PersFormer lacks spatial features and does not use any kind of physical regularization. Consequently, it fails to estimate the 3D lane geometry and even collapses in Fig. 6c, whereas our surface and curvature priors always prevent such a behavior. Noteworthy is also the top performance on the merges and splits set. This proves that our soft regularization is even capable to handle situations containing non-parallel lines, which is also confirmed by Fig. 6d. However, we rarely observe limi", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tations with our formulation for line pairs with a similar orientation but weakly converging course as shown in Fig. 6e. In such cases the indicator function might erroneously decide for parallelism loss during training. One possible solution for future work would be to consider ground truth for the indicator function to identify such situations.", + "bbox": [ + 496, + 321, + 890, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Evaluation on Apollo 3D Synthetic", + "text_level": 1, + "bbox": [ + 498, + 424, + 802, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Apollo 3D Synthetic dataset is very limited in size and only consists of simple situations in contrast to OpenLane. While we find the results on OpenLane more meaningful, we would like to still provide and discuss the quantitative results on the Apollo dataset. Due to the simplicity of the dataset, our model cannot benefit that significantly from our priors but still achieves competitive results to state of the art with the highest F1-Score on the balanced scenes dataset and comparable error metrics (second best for most errors).", + "bbox": [ + 496, + 448, + 890, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions and future work", + "text_level": 1, + "bbox": [ + 498, + 602, + 767, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present LaneCPP, a novel approach for 3D lane detection that leverages physical prior knowledge about lane structure and road geometry. Our new continuous lane representation overcomes previous deficiencies by allowing arbitrary lane structures and enables us to regularize lane geometry based on analytically formulated priors. We further introduce a novel spatial transformation module that models 3D features carefully considering knowledge about road surface geometry. In our experiments, we demonstrate state-of-the-art performance on real and synthetic benchmarks. The full capability of our approach is revealed on real-world OpenLane, for which we prove the relevance of priors quantitatively and qualitatively. In future, priors could be individualized for different driving scenarios and might support to learn inter-lane relations to achieve better scene understanding in a global context. We also see ways to leverage the full potential of the spatial transformation by using more sophisticated surface representations.", + "bbox": [ + 496, + 628, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10646", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 7, 8", + "[2] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2, 3, 4, 6, 7, 8", + "[3] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2", + "[4] Netalee Efrat, Max Bluvstein, Shaul Oron, Dan Levi, Noa Garnett, and Bat El Shlomo. 3d-lanenet+: Anchor free lane detection using a semi-local representation. arXiv/2011.01535, 2020. 1, 2, 3", + "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 7, 8", + "[7] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1, 2", + "[8] Stephen Gould, Richard Hartley, and Dylan Campbell. Deep declarative networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2021. 2", + "[9] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 4, 6, 7, 8", + "[10] Steffen Hagedorn, Marcel Milich, and Alexandru P. Condurache. Pioneering se (2)-equivariant trajectory planning for automated driving. arXiv:2403.11304, 2024. 2", + "[11] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2", + "[12] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8", + "[13] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1, 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of the International Conf. on Learning Representations (ICLR), 2015. 7", + "[15] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020. 1, 2", + "[16] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1, 2", + "[17] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8", + "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3", + "[19] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 2", + "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 3", + "[21] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 6", + "[22] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2", + "[23] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 1, 2, 3", + "[24] Pingping Lu, Chen Cui, Shaobing Xu, Huei Peng, and Fan Wang. SUPER: A novel lane detection system. IEEE Trans. on Intelligent Vehicles (T-IV), 2021. 2", + "[25] Hanspeter Mallot, Heinrich Blthoff, J.J. Little, and S Bohrer. Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological Cybernetics, 1991. 3", + "[26] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1, 2", + "[27] Marcos Nieto, Luis Salgado, Fernando Jaureguizar, and Jon Arróspide. Robust multiple lane road modeling based on perspective analysis. In Proc. IEEE International Conf. on Image Processing (ICIP), 2008. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10647", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics Autom. Lett., 2020. 3", + "[29] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1, 2", + "[30] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3", + "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 3, 5", + "[32] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3, 4, 6, 8", + "[33] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1, 2", + "[34] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[35] Matthias Rath and Alexandru Paul Condurache. Invariant integration in deep convolutional feature space. In Proc. of European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2020. 2", + "[36] Matthias Rath and Alexandru Paul Condurache. Improving the sample-complexity of deep classification networks with invariant integration. In Proc. of International Joint Conf. on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), 2022. 2", + "[37] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[38] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Proc. of the British Machine Vision Conf. (BMVC), 2019. 3", + "[39] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3", + "[40] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2", + "[41] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2", + "[42] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. of the International Conf. on Machine learning (ICML), 2019. 7", + "[43] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polyanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2", + "[44] Bingke Wang, Zilei Wang, and Yixin Zhang. Polynomial regression network for variable-number lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2", + "[45] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[46] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8", + "[47] Yuping Wang and Jier Chen. Eqdrive: Efficient equivariant motion forecasting with multi-modality for autonomous driving. arXiv:2310.17540, 2023. 2", + "[48] Lu Xiong, Zhenwen Deng, Peizhi Zhang, and Zhiqiang Fu. A 3d estimation of structural road surface based on lane-line information. IFAC Conf. on Engine and Powertrain Control, Simulation and Modeling (E-COSM), 2018. 2", + "[49] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[50] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3", + "[51] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. RESA: recurrent feature-shift aggregator for lane detection. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2021. 2", + "[52] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[53] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 854 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10648", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_model.json b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ad6161b8e8c1f5f90f53de5a42d77b993a8f73c --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_model.json @@ -0,0 +1,2442 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.804, + 0.152 + ], + "angle": 0, + "content": "LaneCPP: Continuous 3D Lane Detection using Physical Priors" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.18, + 0.734, + 0.198 + ], + "angle": 0, + "content": "Maximilian Pittner\\(^{1,2}\\), Joel Janai\\(^{1}\\), Alexandru P. Condurache\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.295, + 0.199, + 0.679, + 0.216 + ], + "angle": 0, + "content": "1Bosch Mobility Solutions, Robert Bosch GmbH" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.217, + 0.694, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Institute of Signal Processing, University of Lübeck" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.237, + 0.796, + 0.251 + ], + "angle": 0, + "content": "{Maximilian.Pittner, Joel.Janai, AlexandruPaul.Condurache}@de.bosch.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.312, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.208, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Robust and precise lane detection systems build one of the most essential components in the perception stack of autonomous vehicles. While some approaches utilize LiDAR sensors or multi-sensor setups, the application of monocular cameras has become more popular due to their lower cost and the high-resolution visual representation that provides valuable information to detect lane markings." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In the past, lane detection was mainly treated as a 2D detection task. Deep learning based methods achieved good results by treating the problem as a segmentation task in pixel space [7, 11, 16, 26, 29, 33, 53], used to classify and regress lanes using anchor-based [19, 41] representations," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.424 + ], + "angle": 0, + "content": "or as key-points on a grid structure [13, 15, 34, 45]. However, due to the lack of depth information, these 2D representations fail to model lane markings and road geometry in 3D space, which forms an important prerequisite for later functionalities like trajectory planning. Consequently, approaches for monocular 3D lane detection were introduced, which adapted lane representations for the 3D domain by modeling vertical anchors [6, 9] or local segments on a grid [4] in a Birds-Eye-View (BEV) oriented 3D-frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.593 + ], + "angle": 0, + "content": "A crucial topic for the application of lane detection algorithms in autonomous systems is safety, which requires predictable and robust behavior in any traffic situation. One risk of learning-based methods is the tendency to show unpredictable behavior in cases of rarely observed scenarios. Since obtaining large amounts of data with high-quality annotations is cumbersome and expensive, publicly available 3D datasets are limited in size and accuracy. Hence, they do not reflect the variability of real-world scenarios sufficiently. This makes learning-based models prone to overfitting, and eventually, diminishes predictability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.702 + ], + "angle": 0, + "content": "One common way to deal with such problems is the integration of prior knowledge. Physics provides us a profound understanding of the 3D world, allowing us to make valid assumptions about the lane structure and road surface geometry. Therefore, we introduce physically motivated priors into the lane detection objective to cope with the limited data problem and achieve robust and predictable behavior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "There are certain geometric properties that should generally hold for detected lane lines. For instance, we know that most lines progress parallel to each other, reside on a smooth surface and should not exceed certain thresholds in terms of curvature and slope. However, integrating such assumptions into prevailing discrete representations is not straight forward as strong simplifications are necessary. In contrast, continuous 3D lane representations directly provide parametric curves using polynomials [1, 23] or more sophisticated B-Splines [32]. These allow for analytical computations on the curve function, which enables the integration of such priors into the lane representation. By modeling these priors explicitly instead of learning them from" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "data, the model can focus its full capacity on learning richer features for the lane detection task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.257 + ], + "angle": 0, + "content": "We can further use physical knowledge about the road geometry to support the model in learning an internal transformation from image features to 3D space. While methods based on Inverse Perspective Mapping (IPM) [4, 6, 9, 17, 23, 32] make false flat-ground assumptions, learning based transformations [1, 2, 46] completely ignore road properties. In contrast, integrating prior knowledge about the road surface allows us to model 3D features geometry-aware and helps the network to focus on the 3D region of interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.468, + 0.379 + ], + "angle": 0, + "content": "Thus, we propose a novel 3D lane detection approach named LaneCPP that leverages valuable prior knowledge to achieve accurate and robust perception behavior. It introduces a new sophisticated continuous curve representation, which enables us to incorporate physical priors. In addition, we present a spatial transformation component for learning a physically inspired mapping from 2D image to 3D space providing meaningful spatial features." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.438, + 0.393 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.394, + 0.468, + 0.439 + ], + "angle": 0, + "content": "- We propose a novel architecture for 3D lane detection from monocular images using a more sophisticated flexible parametric spline-based lane representation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.44, + 0.468, + 0.469 + ], + "angle": 0, + "content": "- We present a way to incorporate priors about lane structure and geometry into our continuous representation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.47, + 0.468, + 0.499 + ], + "angle": 0, + "content": "- We introduce a new way to use prior knowledge about the road surface geometry for learning spatial features." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.5, + 0.468, + 0.528 + ], + "angle": 0, + "content": "- We demonstrate the benefits of our contributions in several ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.53, + 0.434, + 0.545 + ], + "angle": 0, + "content": "- We show state-of-the-art performance of our model." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.394, + 0.468, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.214, + 0.573 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Different Lane Representations. An important design choice in deep learning based lane detection is the representation that the network uses to model lane line geometry, which can be categorized as follows: 1) Pixelwise representations, which formulate lane detection as a segmentation problem, were used mainly in 2D methods [7, 11, 16, 26, 29, 33, 51, 53] and were adopted in 3D by SALAD [50] combining line segmentation with depth-prediction. These representations come with high computational load since a large amount of parameters is required. 2) Grid-based approaches divide the space into cells and model lanes using local segments [13] or keypoints [15, 34, 45]. 3D-LaneNet+ [4] suggests to use local line-segments and BEV-LaneDet [46] defines key-points on a BEV grid representation. Both depend on the grid resolution and require costly post-processing to obtain lines. 3) Anchor-based representations [19, 40, 41, 52] model lines as straight anchors with positional offsets at predefined locations. They are widely used in 3D detection approaches including 3D-LaneNet [6] and Gen-LaneNet [9], which use vertical anchors in the top-view, and An" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "chor3DLane [12], introducing anchor projection with iterative regression. Similar to grid-based representations, it requires subsequent curve-fitting to obtain smooth lines. 4) Continuous curve representations [5, 22, 24, 43, 44] instead directly model smooth curves without requiring costly post-processing. While CLGO [23] and CurveFormer [1] use simple polynomials, 3D-SpLineNet [32] proposes B-Splines [3]. Since B-Splines offer local control over curve segments, they are compatible to model complex shapes with low-degree basis functions, while polynomials and Bezier curves show global dependence and thus require higher degrees causing expensive computation. Although 3D-SpLineNet achieves superior detection performance on synthetic data, it unfortunately lacks flexibility as the curve formulation is limited to monotonically progressing lanes, making it hardly applicable to real-world data. To resolve this issue, we propose a more flexible representation based on actual 3D B-Splines. In contrast to discrete grids and anchors, continuous representation even allows us to integrate prior knowledge in an analytical manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Geometry Priors. Several approaches suggest to incorporate prior knowledge into learning-based methods, e.g. by integrating invariance into the model architecture [35, 36] or task-specific transformations as for trajectory planning [10, 47, 49]. In the field of lane detection, line parallelism has been formulated as a hard constraint to resolve depth ambiguity and determine camera parameters [27, 48]. Deep declarative networks [8] offer a general framework to incorporate arbitrary properties as constraints, by solving a constrained optimization problem in the forward pass. While such methods are appropriate when hard constraints must be enforced, our goal is rather to guide the network in learning typical geometric lane properties by formulating soft constraints in a regularization objective. Such a regularization only affects training and does not require resolving an optimization problem in the forward pass, and thus, comes without additional computational cost during inference. Following this paradigm, SGNet [24, 40] proposes to penalize the deviation of lateral distance from a constant lane width in the IPM warped top-view, but ignores that the property does not hold for lines deviating from the ground plane. GP [17] presents a parallelism loss that enforces constant distance between nearest neighbors locally, which depends on the number of anchor points. In contrast, our method presents a way to learn parallelism globally and independent of resolutions of discrete lane representations. We propose an elegant way to learn parallelism as well as other geometry priors using analytical formulations of tangents and normals, which are well-defined on our continuous spline representation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Leveraging 3D Features. An important model component consists in the extraction of 3D features, encoding valuable information to detect lanes along the road surface." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.088, + 0.887, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.893, + 0.351 + ], + "angle": 0, + "content": "Figure 1. Our approach: First, front-view image \\( I \\) is propagated through the backbone extracting multi-scale feature maps. These are transformed to 3D using our spatial transformation and then fused to obtain a single 3D feature map. Feature pooling is applied to obtain features for each line proposal that are propagated through fully connected layers to obtain the parameters for our line representation. Finally, prior knowledge is exploited to regularize the lane representation and to produce surface hypotheses for the spatial transformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.473, + 0.784 + ], + "angle": 0, + "content": "While some works predict 3D lanes directly from the front-view, e.g. by utilizing pixel-wise depth estimation [50] or 3D anchor-projection mechanisms [12], prevalent methods employ an intermediate 3D or BEV feature representation with an internal transformation from the front-view to the 3D space. 3D-LaneNet [6] proposes to utilize IPM [25] to project front-view features to a flat road plane due to the spatial correlation between the warped top-view image and 3D lane geometry and was adopted in several other works [4, 9, 17, 23, 32]. However, IPM causes visual distortions in the top-view representation when the flat road assumption is violated. In related fields like BEV semantic segmentation, BEV transformations are learned via Multi-Layer-Perceptrons (MLPs) [18, 28], depth prediction [31, 37, 38] or transformer-based attention mechanisms [20, 30, 39]. In 3D lane detection, PersFormer [2] utilizes attention between front- and top-view, CurveFormer [1] introduces dynamic 3D anchors that model queries as parametric curves and BEV-LaneDet [46] uses MLPs for the spatial transformations. However, these learned transformations do not necessarily provide a 3D feature representation since they are not guided by valuable priors about the road surface geometry, which potentially results in unforeseen behavior for out-of-distribution data. Our approach instead aims for carefully modeling a geometry-aware feature space using a depth classification method inspired by [31] that exploits knowledge about the distribution of the road surface." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.799, + 0.214, + 0.817 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The following section describes our 3D lane detection approach. An overview of the overall architecture is described and illustrated in Fig. 1. The main focus lies on our continuous 3D lane line representation, our regularization mechanism using physical priors and our prior-based spatial trans" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.371, + 0.885, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.596, + 0.895, + 0.682 + ], + "angle": 0, + "content": "Figure 2. Our 3D lane line representation: For each proposal \\(\\bar{f}\\) (purple lines), line geometry is described by 3D B-Splines with control points \\(c_{k}\\) (green dots). Each control point is determined by the offsets \\(\\alpha_{k},\\beta_{k}\\) from the control points of the initial proposal in normal direction (orange vectors). Additionally, visibility \\(v(t)\\) is modeled by splines with 1D control points \\(\\gamma_{k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.715, + 0.858, + 0.731 + ], + "angle": 0, + "content": "formation module, which we explain in the following." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.752, + 0.724, + 0.769 + ], + "angle": 0, + "content": "3.1. Lane line representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Inspired by prior work in 3D lane detection [32], we leverage the benefits of continuous representations and employ a parametric model based on B-Splines. However, modeling only lateral \\((x-)\\) and vertical \\((z-)\\) components with spline-based functions (as done in previous approaches) is limited to lanes that merely progress along the longitudinal \\((y-)\\) direction. Instead, we propose the first full 3D lane line representation modeling each component \\((x, y, z)\\) such that we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10641" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.125, + 0.105 + ], + "angle": 0, + "content": "obtain" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.113, + 0.47, + 0.162 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} (t) = \\left( \\begin{array}{c} x (t) \\\\ y (t) \\\\ z (t) \\end{array} \\right) = \\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {k} \\cdot B _ {k, d} (t) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.17, + 0.469, + 0.233 + ], + "angle": 0, + "content": "with curve argument \\( t \\in [0, 1] \\) and \\( K \\) control points \\( c_k = (x_k, y_k, z_k)^T \\). Each control point \\( c_k \\) weights the respective basis function \\( B_{k,d}(t) \\) (recursive polynomials of degree \\( d \\)) controlling the curve shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.419 + ], + "angle": 0, + "content": "Due to the ambiguity of curves using 3D B-Splines (the same spline curve can be described by different configurations of its control points), regressing all three dimensions per control point results in strong overfitting during training. We resolve this issue by limiting the degrees of freedom per control point to two and constraining the control points deflection to one direction in the \\(x - y\\) -plane and one direction in the \\(y - z\\) -plane as illustrated in Fig. 2. More precisely, the degrees of freedom per control point are specified by the directions of the normals \\(\\mathbf{N}_{xy}\\) and \\(\\mathrm{N}_z\\) of an initial curve proposal \\(\\bar{f}\\) with control points \\(\\bar{c}_k = (\\bar{x}_k,\\bar{y}_k,\\bar{z}_k)^T\\). The control points are then defined as" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.427, + 0.47, + 0.477 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {c} _ {k} = \\left( \\begin{array}{l} x _ {k} \\\\ y _ {k} \\\\ z _ {k} \\end{array} \\right) = \\left( \\begin{array}{l} \\bar {x} _ {k} + \\mathrm {N} _ {x} \\cdot \\alpha_ {k} \\\\ \\bar {y} _ {k} + \\mathrm {N} _ {y} \\cdot \\alpha_ {k} \\\\ \\bar {z} _ {k} + \\mathrm {N} _ {z} \\cdot \\beta_ {k} \\end{array} \\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.62 + ], + "angle": 0, + "content": "where \\(\\mathrm{N}_x, \\mathrm{N}_y\\) describe the \\(x\\)- and \\(y\\)-component of the normal vector \\(\\mathbf{N}_{xy}\\) in the \\(x\\)-\\(y\\)-plane. As shown in Eq. (2) and illustrated in Fig. 2, modeling splines as deflections in normal direction of its underlying initial line proposal only requires two parameters \\(\\alpha_k, \\beta_k\\) per control point to describe the 3D shape. We use a wide variety of orientations for the initial proposals \\(\\bar{\\pmb{f}}\\) (see Fig. 2), which allows us to detect any kind of lines with this formulation. More details about the initial proposals are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.469, + 0.77 + ], + "angle": 0, + "content": "While [32] models the curve range using start- and endpoints that are learned by means of regression, we instead propose to model visibility using a continuous representation \\( v(t) \\) and treat the visibility estimation as a classification problem. We obtain probability values applying sigmoid activation and consider \\( \\sigma(v(t)) > 0.5 \\) the visible range. While in theory any kind of function can be utilized, we found that B-Splines with the same configuration as \\( f(t) \\) are well-suited and introduce spline control points \\( \\gamma_k \\) defining the shape of \\( v(t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.77, + 0.469, + 0.801 + ], + "angle": 0, + "content": "Eventually, binary cross-entropy is used as a classification loss to learn visibility" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.809, + 0.47, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v i s} = - \\frac {1}{| \\mathcal {P} _ {G T} |} \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} _ {G T}} \\hat {v} _ {\\boldsymbol {p}} \\cdot \\log (\\sigma (v (t _ {\\boldsymbol {p}}))) + \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.849, + 0.469, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\left. \\left(1 - \\hat {v} _ {\\boldsymbol {p}}\\right) \\cdot \\log \\left(1 - \\sigma \\left(v \\left(t _ {\\boldsymbol {p}}\\right)\\right)\\right), \\right. \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.902 + ], + "angle": 0, + "content": "For the concept of visibility, we follow the prevailing definition from the literature [2, 9]." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.883, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.239, + 0.892, + 0.267 + ], + "angle": 0, + "content": "Figure 3. Illustration of different priors expressed by line tangents and surface normals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.356 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_{GT}\\) denotes the ground truth set of points, \\(\\hat{v}_p\\in \\{0,1\\}\\) the ground truth visibility for point \\(p\\). \\(t_p\\) represents the respective curve argument obtained by orthogonal projection of \\(p\\) onto the underlying line proposal." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.816, + 0.381 + ], + "angle": 0, + "content": "3.2. Regularization using physical priors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.891, + 0.432 + ], + "angle": 0, + "content": "In this section, we describe our regularization method to integrate prior knowledge about lane structure and surface geometry into our parametric line representation (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Line parallelism. In order to reinforce parallel lines, the tangents at point pairs located in opposite normal direction on neighboring lines must be similar (see Fig. 3 left). We realize this by penalizing the cosine distance of the unit tangents \\(\\mathbf{T}(t)\\) on neighboring lines \\(i\\) and \\(j\\) for normal point pairs. More precisely, for each point \\(\\pmb{p} \\in \\mathcal{P}^{(i)}\\) on line \\(i\\) we select the normal pair point \\(\\pmb{p}^*\\) on neighbor line \\(j\\) that minimizes the distance to the normal plane, which is defined by the plane equation \\(\\mathbf{T}^{(i)}(t)^T \\cdot ((x,y,z)^T - \\pmb{f}^{(i)}(t)) = 0\\). In Fig. 3 the normal planes are visualized in a 2D top-view as lines (orange) for simplicity. Hence the respective curve argument \\(t_{\\pmb{p}^*}\\) for point \\(\\pmb{p}^*\\) on line \\(j\\) is given as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.625, + 0.892, + 0.656 + ], + "angle": 0, + "content": "\\[\nt _ {\\boldsymbol {p} ^ {*}} = \\underset {\\boldsymbol {p} ^ {\\prime} \\in \\mathcal {P} ^ {(j)}} {\\operatorname {a r g m i n}} \\left. \\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right) ^ {T} \\cdot \\left(\\boldsymbol {f} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {\\prime}}\\right) - \\boldsymbol {f} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right), \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.744 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}^{(j)}\\) denotes the points on line \\(j\\). While in theory Eq. (5) can be solved analytically, the simpler way is to sample the set of points \\(\\mathcal{P}^{(j)}\\) instead. (Note that our continuous representation allows us to choose high sampling rates without losing precision as no interpolation is required.)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.892, + 0.789 + ], + "angle": 0, + "content": "With the normal point pairs, we define the parallelism loss for a neighbor line pair based on the cosine distance of their tangents as" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.8, + 0.892, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p a r} ^ {(i j)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {T} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {*}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.853, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Since the criterion of line parallelism should not hold for all normal point pairs of neighboring lines (e.g. merging or splitting lines), \\(\\mathbb{1}_{\\pmb{p}}^{(ij)} \\in \\{0, 1\\}\\) represents the indicator" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10642" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "function determining whether the parallelism loss is applied to the point pair. More precisely, the function ensures that only the overlapping range of neighboring lines is taken into account. Furthermore, it determines whether the line pair should be considered as a parallel pair based on the standard deviation of euclidean distances between normal point pairs, i.e. high deviations indicate that the line pair might belong to a merge or split structure. In our experiments, we achieve state-of-the-art performance on test sets containing merges and splits, proving that our model is also capable of learning non-parallel line pairs using this indicator function." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.472, + 0.318 + ], + "angle": 0, + "content": "Surface smoothness. Since the lines reside on a smooth road, the surface normals of neighboring lanes should be similar. Analogously to \\(\\mathcal{L}_{par}\\), we express this with the cosine distance between surface normals \\(\\mathbf{N}^{(ih)}\\) and \\(\\mathbf{N}^{(ij)}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.335, + 0.47, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s m} ^ {(i)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(h i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {N} ^ {(i h)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {N} ^ {(i j)} \\left(t _ {\\boldsymbol {p}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.387, + 0.47, + 0.502 + ], + "angle": 0, + "content": "with indicator function \\(\\mathbb{1}_{\\pmb{p}}^{(hij)}\\). The surface normal between line \\(i\\) and left neighbor line \\(h\\) at point \\(\\pmb{p}\\) can be expressed as the cross product of the tangent on line \\(i\\) and the normalized connection vector between lines \\(i\\) and \\(h\\), hence \\(\\mathbf{N}^{(ih)}(t_{\\pmb{p}}) = \\mathbf{T}^{(i)}(t_{\\pmb{p}}) \\times \\frac{\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})}{||\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})||}\\). For the normal between line \\(i\\) and right neighbor \\(j\\) the sign is flipped to obtain upwards pointing normal vectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.642 + ], + "angle": 0, + "content": "**Curvature.** We determine lane curvature by computing the second order derivatives as the difference of tangents at consecutive points divided by their euclidean distance as \\(\\mathbf{T}'(t_p) = \\frac{\\mathbf{T}(t_p) - \\mathbf{T}(t_p - \\Delta t)}{||\\mathbf{f}(t_p) - \\mathbf{f}(t_p - \\Delta t)||}\\). The maximum curvature in \\(x-y\\)-plane (inverse curve radius) and in \\(z\\) (rate of slope change) have very different value ranges and are therefore restricted by different limits. Hence, we define the two thresholds \\(\\kappa_{xy}\\) and \\(\\kappa_z\\) and formulate the curvature loss on line \\(i\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.644, + 0.47, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {c u r v} ^ {(i)} = \\frac {1}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} \\max \\left(\\mathrm {T} _ {x y} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {x y}\\right) (8) \\\\ + \\max \\left(\\mathrm {T} _ {z} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {z}\\right). (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.71, + 0.388, + 0.726 + ], + "angle": 0, + "content": "Finally, the prior regularization loss is given as" + }, + { + "type": "equation", + "bbox": [ + 0.087, + 0.731, + 0.469, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r i o r}} = \\sum_ {i = 1} ^ {M} \\lambda_ {s m} \\mathcal {L} _ {s m} ^ {(i)} + \\lambda_ {\\text {c u r v}} \\mathcal {L} _ {\\text {c u r v}} ^ {(i)} + \\sum_ {j = 1} ^ {N} \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} ^ {(i j)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.903 + ], + "angle": 0, + "content": "with individual weights \\(\\lambda_{par}\\), \\(\\lambda_{sm}\\), \\(\\lambda_{curv}\\). Note that all these properties are expressible by means of tangents and normals, which can be computed analytically on our parametric representation in continuous space. Consequently, minimization of the herein introduced prior losses does not depend on numerical approximations as is the case for anchor-, grid- or key-point representations." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.089, + 0.885, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.315, + 0.895, + 0.386 + ], + "angle": 0, + "content": "Figure 4. Our proposed spatial transformation module. First, several road surface hypotheses are defined (a) to which front-view features are lifted (b) and weighted according to the predicted depth distribution. Afterwards, point features are aggregated in a weighted manner to obtain the 3D feature map (c)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.712, + 0.433 + ], + "angle": 0, + "content": "3.3. Spatial transformation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.443, + 0.892, + 0.609 + ], + "angle": 0, + "content": "In this section, we describe our spatial transformation (shown in Fig. 4) that is leveraging valuable physical knowledge about surface geometry. We know that the road surface typically shows small deviations from the ground level \\((z = 0)\\) in the near-range and stronger deviations in the far-range. Based on this knowledge, we sample ground surface hypotheses that reflect the distribution of the road surface height profile (Fig. 4a). While in theory different types of surface functions could be utilized as hypotheses, we decide to merely rely on planes, since this facilitates the computation of ray intersections described in the following step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.893, + 0.822 + ], + "angle": 0, + "content": "Next, the multi-scale front-view feature maps extracted by the backbone are lifted to 3D space (Fig. 4b). Our approach is inspired by [31], where front-view features are spreading along rays throughout the space of the road surface. These rays intersect with the surface hypotheses at different depths spanning a frustum-like point cloud in 3D space, where each point is affiliated with a \\(C\\)-dimensional feature vector and additionally attached with its height value \\(z\\), hence, each point in the cloud has dimension \\((C + 1)\\). The front-view feature map is propagated through a depth branch with a channel-wise softmax applied to obtain a categorical distribution for each ray, resulting in a tensor of size \\(H\\times W\\times S\\), where \\(H\\), \\(W\\) denote height and width and channel size \\(S\\) the number of surface hypotheses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In order to aggregate the information in 3D space, a BEV grid of size \\( X \\times Y \\) is defined. Features from points mapping to the same grid cell are weighted by the categorical depth distribution for the respective ray and accumulated in terms of a weighted sum (Fig. 4c). Since the \\( z \\)-component" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10643" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.47, + 0.168 + ], + "angle": 0, + "content": "
PriorsF1(%)↑X-near(m)↓X-far(m)↓Z-near(m)↓Z-far(m)↓
None65.00.3160.3840.1060.153
Par.66.20.2910.3730.1030.150
Surf.65.80.3200.3560.1030.144
Curv.66.70.3220.3660.1050.146
Comb.66.70.3010.3590.1030.144
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.177, + 0.443, + 0.192 + ], + "angle": 0, + "content": "Table 1. Effect of different prior losses on OpenLane300." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.207, + 0.468, + 0.245 + ], + "angle": 0, + "content": "
# Surface Hypotheses1351527
F1-Score(%)↑65.065.966.666.166.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.255, + 0.449, + 0.27 + ], + "angle": 0, + "content": "Table 2. Effect of the surface hypotheses on OpenLane300." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.285, + 0.468, + 0.378 + ], + "angle": 0, + "content": "
Lane Rep.Prior Reg.Spatial T.F1(%)↑Gain(%)
62.9(baseline)
✓65.0+2.1
✓✓66.7+3.8
✓✓66.6+3.7
✓✓✓66.9+4.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Table 3. Performance gain for different contributions on OpenLane300 using our novel Lane Representation, Prior Regularization and Spatial Transformation instead of IPM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.536 + ], + "angle": 0, + "content": "of the points is also combined by a weighted sum, the value \\( z_{uv} \\) can be interpreted as the height value of the surface for grid cell \\( (u, v) \\). We guide the model in learning the real surface and prevent it from learning an arbitrary mapping by introducing a simple grid-based regression loss as" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.546, + 0.47, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s u r f}} = \\frac {1}{X \\cdot Y} \\sum_ {(u, v) \\in X \\times Y} \\mathbb {1} _ {u v} \\cdot \\| z _ {u v} - \\hat {z} _ {u v} \\| _ {1}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.643 + ], + "angle": 0, + "content": "with \\(\\mathbb{1}_{uv}\\) indicating whether surface ground truth \\(\\hat{z}_{uv}\\) is available for cell \\((u,v)\\). The height ground truth is obtained by interpolation of the 3D lane annotations at cell locations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.653, + 0.226, + 0.669 + ], + "angle": 0, + "content": "3.4. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.709 + ], + "angle": 0, + "content": "The overall loss used during training is given as the weighted sum of loss components" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.72, + 0.469, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {c a t} \\mathcal {L} _ {c a t} + \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.74, + 0.469, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {v i s} \\mathcal {L} _ {v i s} + \\lambda_ {p r i o r} \\mathcal {L} _ {p r i o r} + \\lambda_ {s u r f} \\mathcal {L} _ {s u r f}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.47, + 0.83 + ], + "angle": 0, + "content": "We use focal loss [21] for lane presence \\(\\mathcal{L}_{pr}\\) and category classification \\(\\mathcal{L}_{cat}\\). For the regression loss \\(\\mathcal{L}_{reg}\\), we adapt the formulation of [32] to three instead of two dimensions. More details are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.21, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We first describe our experimental setup and then analyze our approach on two 3D lane datasets." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.687, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.169, + 0.629, + 0.175 + ], + "angle": 0, + "content": "Front-view" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.176, + 0.689, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.871, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.169, + 0.811, + 0.176 + ], + "angle": 0, + "content": "Front-view" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.177, + 0.871, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.281, + 0.69, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.336, + 0.621, + 0.348 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.281, + 0.872, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.336, + 0.801, + 0.348 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.36, + 0.892, + 0.416 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of our model trained with prior regularization to the same model without regularization both trained on OpenLane300 with main differences highlighted by arrows. As a reference ground truth lines are visualized dashed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.446, + 0.688, + 0.463 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.892, + 0.517 + ], + "angle": 0, + "content": "We evaluate our method on two different datasets: OpenLane and Apollo 3D Synthetic - both containing 3D lane ground truth as well as camera parameters per frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.518, + 0.892, + 0.624 + ], + "angle": 0, + "content": "OpenLane [2] is a real-world dataset containing 150,000 images in the training and 40,000 in the test set from 1000 different sequences. In order to evaluate different driving scenarios the test set is divided into different situations, namely Up & Down, Curve, Extreme Weather, Night, Intersection and Merge & Split. For ablation studies we use the smaller version OpenLane300 including 300 sequences." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.701 + ], + "angle": 0, + "content": "Apollo 3D Synthetic [9] is a small synthetic dataset, consisting of only 10,500 examples from rather simple scenarios of highway, urban and rural environments. The data is split into three subsets, (1) Standard (simple) scenarios, (2) Rare Scenes and (3) Visual Variations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Evaluation metrics. For the quantitative evaluation both datasets utilize the evaluation scheme proposed in [9]. It evaluates the euclidean distance at uniformly distributed points in the range of \\(0 - 100\\mathrm{m}\\) along the \\(y\\) direction. Based on the mean distance and range, \\(Fl\\)-Score is computed, as well as the mean \\(x\\)- and \\(z\\)-errors in near- \\((0 - 40\\mathrm{m})\\) and far-range \\((40 - 100\\mathrm{m})\\) to evaluate geometric accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baseline. Our approach builds up on 3D-SpLineNet. Since it was applied on synthetic data only, it showed poor performance on real data. We applied some straightforward design adaptations - e.g. larger backbone, multiscale features (see supplementary) - and use this modified 3D-SpLineNet as our baseline (first row Table 3)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10644" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.246 + ], + "angle": 0, + "content": "
MethodF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [6]44.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [9]32.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [2]50.50.4850.5530.3640.43142.455.648.646.640.050.7
PersFormer* [2]53.10.3610.3280.1240.12946.858.754.048.441.452.5
CurveFormer [1]50.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [46]58.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane [12]53.70.2760.3110.1070.13846.757.252.547.845.451.2
Anchor3DLane-T [12]54.30.2750.3100.1050.13547.258.052.748.745.851.7
LaneCPP (Ours)60.30.2640.3100.0770.11753.664.456.754.952.058.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.256, + 0.893, + 0.3 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on OpenLane [2]. Best performance and second best are highlighted. The scenario categories are Up and Down (U&D), Curve (C), Extreme Weather (EW), Night (N), Intersection (I), Merge and Split (M&S). PersFormer* denotes the latest performance reported on the official code base, Anchor3DLane-T represents the temporal multi-frame method of [12]." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.313, + 0.236, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.379, + 0.235, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.246, + 0.313, + 0.395, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.383, + 0.397, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.313, + 0.558, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.381, + 0.559, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.313, + 0.722, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.381, + 0.724, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.313, + 0.886, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.381, + 0.888, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.481, + 0.239, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.53, + 0.169, + 0.541 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.482, + 0.399, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.531, + 0.328, + 0.542 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.481, + 0.562, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.532, + 0.491, + 0.543 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.482, + 0.726, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.531, + 0.655, + 0.543 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.482, + 0.889, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.532, + 0.819, + 0.543 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.556, + 0.882, + 0.571 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on OpenLane. Our method is compared to PersFormer* with ground truth visualized as dashed lines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Implementation details. We use input size \\(360 \\times 480\\) and adopt the same backbone as in [2] based on a modified EfficientNet [42]. We extract four feature maps of resolutions \\(\\left[\\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}\\right]\\). The final 3D feature map has size \\(26 \\times 16\\) with 64 channels. We use \\(M = 64\\) initial line proposals and B-Splines of degree \\(d = 3\\) and \\(K = 10\\) control points. We apply Adam optimizer [14] with an initial learning rate of \\(2 \\times 10^{-4}\\) for OpenLane and \\(10^{-4}\\) for Apollo and a dataset specific step-wise scheduler. We train for 30 epochs on OpenLane and 300 epochs on Apollo with batch size 16. For more details we refer to the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.238, + 0.787 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Table 1 indicates the effect of our proposed prior-based regularization. It is evident that each prior improves the F1-Score as well as geometric errors. While the surface and curvature priors result in better far-range estimates, line parallelism supports X-regression in the near-range. Besides, using surface smoothness loss results in lowest Z-far errors. Finally, a combination of priors yields a good balance of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.748 + ], + "angle": 0, + "content": "F1-Score and geometric errors. The positive effect of parallelism is confirmed by Fig. 5, where reinforcing parallel lane structure leads to better estimates in the near-range (a) and far-range (b) compared to the unregularized model. Learning parallel lines also is evidently beneficial in cases of poor visibility (b) and occlusions (a). In the latter case, the regularized model even shows better predictions than the noisy ground truth. This emphasizes the high relevance of priors for more robust behavior for real-world datasets, where 3D ground truth often comes with inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "For the spatial transformation (see Table 2), too low numbers of surface hypotheses result in worse score, presumably as 3D geometry is not captured sufficiently, whereas larger numbers tend to decreasing performance due to the higher complexity. The best F1-Score is obtained with 5 hypotheses, which is chosen for further experiments. While the improvement over IPM is already considerable, we think that with the simplifications of plane hypotheses prevent the component from developing its full potential. We see ways to enhance the 3D transformation even further" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10645" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.272 + ], + "angle": 0, + "content": "
MethodBalanced ScenesRare Scenes
F1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfarF1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfar
3D-LaneNet [6]86.40.0680.4770.0150.20272.00.1660.8550.0390.521
GP [17]91.90.0490.3870.0080.21383.70.1260.9030.0230.625
PersFormer [2]92.90.0540.3560.010.23487.50.1070.7820.0240.602
3D-SpLineNet [32]96.30.0370.3240.0090.21392.90.0770.6990.0210.562
CurveFormer [1]95.80.0780.3260.0180.21995.60.1820.7370.0390.561
BEV-LaneDet [46]96.90.0160.2420.020.21697.60.0310.5940.0400.556
Anchor3DLane [12]95.40.0450.3000.0160.22394.40.0820.6990.0300.580
LaneCPP (Ours)97.40.0300.2770.0110.20696.20.0730.6510.0230.543
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.28, + 0.872, + 0.296 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison of best methods on Apollo 3D Synthetic [9]. Best performance and second best are highlighted." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.322, + 0.457, + 0.337 + ], + "angle": 0, + "content": "using more sophisticated spatial representations in future." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.47, + 0.474 + ], + "angle": 0, + "content": "The impact of our different contributions is summarized in Table 3, where the first row shows our baseline (see Sec. 4.1). More than two percent in F1-Score are gained with our novel lane representation compared to the simplified one from [32]. Moreover, it is clear that both, the regularization using combined priors and the spatial transformation using 5 hypotheses result in significant improvement. Eventually, combining all components yields the best model configuration, which we choose for further evaluation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.308, + 0.501 + ], + "angle": 0, + "content": "4.3. Evaluation on OpenLane" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.705 + ], + "angle": 0, + "content": "On the real-world OpenLane benchmark our model evidently outperforms all other methods with respect to F1-Score as well as geometric errors as shown in Table 4. Compared to BEV-Lanedet, which achieves a high detection score, our model gains \\(+1.9\\%\\), while reaching significantly lower geometric errors. In comparison to Anchor3DLane the improvements with respect to X-errors are less substantial, however, our approach surpasses the F1-Score by a large gap of \\(+6.6\\%\\). Analyzing the detection scores among different scenarios, outstanding performance gain is observed on the up- and down-hill test set \\((+5.9\\%)\\) that highlights the capability of our approach to capture 3D space proficiently, which is supported by the low Z-errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Apart from quantitative results, we show qualitative examples in Fig. 6. In up-hill scenarios like Fig. 6b our model manages to estimate both lateral and height profile accurately, since our assumptions about road surface and line parallelism are satisfied. In contrast, PersFormer lacks spatial features and does not use any kind of physical regularization. Consequently, it fails to estimate the 3D lane geometry and even collapses in Fig. 6c, whereas our surface and curvature priors always prevent such a behavior. Noteworthy is also the top performance on the merges and splits set. This proves that our soft regularization is even capable to handle situations containing non-parallel lines, which is also confirmed by Fig. 6d. However, we rarely observe limi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.892, + 0.414 + ], + "angle": 0, + "content": "tations with our formulation for line pairs with a similar orientation but weakly converging course as shown in Fig. 6e. In such cases the indicator function might erroneously decide for parallelism loss during training. One possible solution for future work would be to consider ground truth for the indicator function to identify such situations." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.425, + 0.803, + 0.442 + ], + "angle": 0, + "content": "4.4. Evaluation on Apollo 3D Synthetic" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.587 + ], + "angle": 0, + "content": "The Apollo 3D Synthetic dataset is very limited in size and only consists of simple situations in contrast to OpenLane. While we find the results on OpenLane more meaningful, we would like to still provide and discuss the quantitative results on the Apollo dataset. Due to the simplicity of the dataset, our model cannot benefit that significantly from our priors but still achieves competitive results to state of the art with the highest F1-Score on the balanced scenes dataset and comparable error metrics (second best for most errors)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.603, + 0.768, + 0.619 + ], + "angle": 0, + "content": "5. Conclusions and future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this work, we present LaneCPP, a novel approach for 3D lane detection that leverages physical prior knowledge about lane structure and road geometry. Our new continuous lane representation overcomes previous deficiencies by allowing arbitrary lane structures and enables us to regularize lane geometry based on analytically formulated priors. We further introduce a novel spatial transformation module that models 3D features carefully considering knowledge about road surface geometry. In our experiments, we demonstrate state-of-the-art performance on real and synthetic benchmarks. The full capability of our approach is revealed on real-world OpenLane, for which we prove the relevance of priors quantitatively and qualitatively. In future, priors could be individualized for different driving scenarios and might support to learn inter-lane relations to achieve better scene understanding in a global context. We also see ways to leverage the full potential of the spatial transformation by using more sophisticated surface representations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10646" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[3] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.47, + 0.354 + ], + "angle": 0, + "content": "[4] Netalee Efrat, Max Bluvstein, Shaul Oron, Dan Levi, Noa Garnett, and Bat El Shlomo. 3d-lanenet+: Anchor free lane detection using a semi-local representation. arXiv/2011.01535, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.47, + 0.41 + ], + "angle": 0, + "content": "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.47, + 0.466 + ], + "angle": 0, + "content": "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[8] Stephen Gould, Richard Hartley, and Dylan Campbell. Deep declarative networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.58, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[9] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[10] Steffen Hagedorn, Marcel Milich, and Alexandru P. Condurache. Pioneering se (2)-equivariant trajectory planning for automated driving. arXiv:2403.11304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[11] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[12] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of the International Conf. on Learning Representations (ICLR), 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[15] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[19] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.518, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[22] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[23] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[24] Pingping Lu, Chen Cui, Shaobing Xu, Huei Peng, and Fan Wang. SUPER: A novel lane detection system. IEEE Trans. on Intelligent Vehicles (T-IV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[25] Hanspeter Mallot, Heinrich Blthoff, J.J. Little, and S Bohrer. Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological Cybernetics, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[26] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Marcos Nieto, Luis Salgado, Fernando Jaureguizar, and Jon Arróspide. Robust multiple lane road modeling based on perspective analysis. In Proc. IEEE International Conf. on Image Processing (ICIP), 2008. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[28] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics Autom. Lett., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[29] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[30] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.398 + ], + "angle": 0, + "content": "[32] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[33] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.469, + 0.51 + ], + "angle": 0, + "content": "[34] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.469, + 0.579 + ], + "angle": 0, + "content": "[35] Matthias Rath and Alexandru Paul Condurache. Invariant integration in deep convolutional feature space. In Proc. of European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[36] Matthias Rath and Alexandru Paul Condurache. Improving the sample-complexity of deep classification networks with invariant integration. In Proc. of International Joint Conf. on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[37] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[38] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Proc. of the British Machine Vision Conf. (BMVC), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[39] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[40] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[42] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. of the International Conf. on Machine learning (ICML), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[43] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polyanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[44] Bingke Wang, Zilei Wang, and Yixin Zhang. Polynomial regression network for variable-number lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[45] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[46] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[47] Yuping Wang and Jier Chen. Eqdrive: Efficient equivariant motion forecasting with multi-modality for autonomous driving. arXiv:2310.17540, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[48] Lu Xiong, Zhenwen Deng, Peizhi Zhang, and Zhiqiang Fu. A 3d estimation of structural road surface based on lane-line information. IFAC Conf. on Engine and Powertrain Control, Simulation and Modeling (E-COSM), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[49] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.616, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[50] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.687, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[51] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. RESA: recurrent feature-shift aggregator for lane detection. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[52] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.8, + 0.892, + 0.855 + ], + "angle": 0, + "content": "[53] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.855 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10648" + } + ] +] \ No newline at end of file diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_origin.pdf b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..894e7af71339b80d2bed01a2dd073f25faaead2c --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/7a911166-f711-4534-9bf2-2cb8d0604f4d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c9fe3453eef0458a1adacfa6b1640299b81060c1f284240428b2b602cd957e5 +size 3079267 diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/full.md b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7901e515958850aa72b650da8d066d67cd681d0c --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/full.md @@ -0,0 +1,358 @@ +# LaneCPP: Continuous 3D Lane Detection using Physical Priors + +Maximilian Pittner $^{1,2}$ , Joel Janai $^{1}$ , Alexandru P. Condurache $^{1,2}$ + +1Bosch Mobility Solutions, Robert Bosch GmbH + +$^{2}$ Institute of Signal Processing, University of Lübeck + +{Maximilian.Pittner, Joel.Janai, AlexandruPaul.Condurache}@de.bosch.com + +# Abstract + +Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors. + +# 1. Introduction + +Robust and precise lane detection systems build one of the most essential components in the perception stack of autonomous vehicles. While some approaches utilize LiDAR sensors or multi-sensor setups, the application of monocular cameras has become more popular due to their lower cost and the high-resolution visual representation that provides valuable information to detect lane markings. + +In the past, lane detection was mainly treated as a 2D detection task. Deep learning based methods achieved good results by treating the problem as a segmentation task in pixel space [7, 11, 16, 26, 29, 33, 53], used to classify and regress lanes using anchor-based [19, 41] representations, + +or as key-points on a grid structure [13, 15, 34, 45]. However, due to the lack of depth information, these 2D representations fail to model lane markings and road geometry in 3D space, which forms an important prerequisite for later functionalities like trajectory planning. Consequently, approaches for monocular 3D lane detection were introduced, which adapted lane representations for the 3D domain by modeling vertical anchors [6, 9] or local segments on a grid [4] in a Birds-Eye-View (BEV) oriented 3D-frame. + +A crucial topic for the application of lane detection algorithms in autonomous systems is safety, which requires predictable and robust behavior in any traffic situation. One risk of learning-based methods is the tendency to show unpredictable behavior in cases of rarely observed scenarios. Since obtaining large amounts of data with high-quality annotations is cumbersome and expensive, publicly available 3D datasets are limited in size and accuracy. Hence, they do not reflect the variability of real-world scenarios sufficiently. This makes learning-based models prone to overfitting, and eventually, diminishes predictability. + +One common way to deal with such problems is the integration of prior knowledge. Physics provides us a profound understanding of the 3D world, allowing us to make valid assumptions about the lane structure and road surface geometry. Therefore, we introduce physically motivated priors into the lane detection objective to cope with the limited data problem and achieve robust and predictable behavior. + +There are certain geometric properties that should generally hold for detected lane lines. For instance, we know that most lines progress parallel to each other, reside on a smooth surface and should not exceed certain thresholds in terms of curvature and slope. However, integrating such assumptions into prevailing discrete representations is not straight forward as strong simplifications are necessary. In contrast, continuous 3D lane representations directly provide parametric curves using polynomials [1, 23] or more sophisticated B-Splines [32]. These allow for analytical computations on the curve function, which enables the integration of such priors into the lane representation. By modeling these priors explicitly instead of learning them from + +data, the model can focus its full capacity on learning richer features for the lane detection task. + +We can further use physical knowledge about the road geometry to support the model in learning an internal transformation from image features to 3D space. While methods based on Inverse Perspective Mapping (IPM) [4, 6, 9, 17, 23, 32] make false flat-ground assumptions, learning based transformations [1, 2, 46] completely ignore road properties. In contrast, integrating prior knowledge about the road surface allows us to model 3D features geometry-aware and helps the network to focus on the 3D region of interest. + +Thus, we propose a novel 3D lane detection approach named LaneCPP that leverages valuable prior knowledge to achieve accurate and robust perception behavior. It introduces a new sophisticated continuous curve representation, which enables us to incorporate physical priors. In addition, we present a spatial transformation component for learning a physically inspired mapping from 2D image to 3D space providing meaningful spatial features. + +Our main contributions can be summarized as follows: + +- We propose a novel architecture for 3D lane detection from monocular images using a more sophisticated flexible parametric spline-based lane representation. +- We present a way to incorporate priors about lane structure and geometry into our continuous representation. +- We introduce a new way to use prior knowledge about the road surface geometry for learning spatial features. +- We demonstrate the benefits of our contributions in several ablation studies. +- We show state-of-the-art performance of our model. + +# 2. Related work + +Different Lane Representations. An important design choice in deep learning based lane detection is the representation that the network uses to model lane line geometry, which can be categorized as follows: 1) Pixelwise representations, which formulate lane detection as a segmentation problem, were used mainly in 2D methods [7, 11, 16, 26, 29, 33, 51, 53] and were adopted in 3D by SALAD [50] combining line segmentation with depth-prediction. These representations come with high computational load since a large amount of parameters is required. 2) Grid-based approaches divide the space into cells and model lanes using local segments [13] or keypoints [15, 34, 45]. 3D-LaneNet+ [4] suggests to use local line-segments and BEV-LaneDet [46] defines key-points on a BEV grid representation. Both depend on the grid resolution and require costly post-processing to obtain lines. 3) Anchor-based representations [19, 40, 41, 52] model lines as straight anchors with positional offsets at predefined locations. They are widely used in 3D detection approaches including 3D-LaneNet [6] and Gen-LaneNet [9], which use vertical anchors in the top-view, and An + +chor3DLane [12], introducing anchor projection with iterative regression. Similar to grid-based representations, it requires subsequent curve-fitting to obtain smooth lines. 4) Continuous curve representations [5, 22, 24, 43, 44] instead directly model smooth curves without requiring costly post-processing. While CLGO [23] and CurveFormer [1] use simple polynomials, 3D-SpLineNet [32] proposes B-Splines [3]. Since B-Splines offer local control over curve segments, they are compatible to model complex shapes with low-degree basis functions, while polynomials and Bezier curves show global dependence and thus require higher degrees causing expensive computation. Although 3D-SpLineNet achieves superior detection performance on synthetic data, it unfortunately lacks flexibility as the curve formulation is limited to monotonically progressing lanes, making it hardly applicable to real-world data. To resolve this issue, we propose a more flexible representation based on actual 3D B-Splines. In contrast to discrete grids and anchors, continuous representation even allows us to integrate prior knowledge in an analytical manner. + +Geometry Priors. Several approaches suggest to incorporate prior knowledge into learning-based methods, e.g. by integrating invariance into the model architecture [35, 36] or task-specific transformations as for trajectory planning [10, 47, 49]. In the field of lane detection, line parallelism has been formulated as a hard constraint to resolve depth ambiguity and determine camera parameters [27, 48]. Deep declarative networks [8] offer a general framework to incorporate arbitrary properties as constraints, by solving a constrained optimization problem in the forward pass. While such methods are appropriate when hard constraints must be enforced, our goal is rather to guide the network in learning typical geometric lane properties by formulating soft constraints in a regularization objective. Such a regularization only affects training and does not require resolving an optimization problem in the forward pass, and thus, comes without additional computational cost during inference. Following this paradigm, SGNet [24, 40] proposes to penalize the deviation of lateral distance from a constant lane width in the IPM warped top-view, but ignores that the property does not hold for lines deviating from the ground plane. GP [17] presents a parallelism loss that enforces constant distance between nearest neighbors locally, which depends on the number of anchor points. In contrast, our method presents a way to learn parallelism globally and independent of resolutions of discrete lane representations. We propose an elegant way to learn parallelism as well as other geometry priors using analytical formulations of tangents and normals, which are well-defined on our continuous spline representation. + +Leveraging 3D Features. An important model component consists in the extraction of 3D features, encoding valuable information to detect lanes along the road surface. + +![](images/8c794e6cb38623d03b88c4f43ce7126e1c660d926c4bc4c53cce5ce995d69c5e.jpg) +Figure 1. Our approach: First, front-view image $I$ is propagated through the backbone extracting multi-scale feature maps. These are transformed to 3D using our spatial transformation and then fused to obtain a single 3D feature map. Feature pooling is applied to obtain features for each line proposal that are propagated through fully connected layers to obtain the parameters for our line representation. Finally, prior knowledge is exploited to regularize the lane representation and to produce surface hypotheses for the spatial transformation. + +While some works predict 3D lanes directly from the front-view, e.g. by utilizing pixel-wise depth estimation [50] or 3D anchor-projection mechanisms [12], prevalent methods employ an intermediate 3D or BEV feature representation with an internal transformation from the front-view to the 3D space. 3D-LaneNet [6] proposes to utilize IPM [25] to project front-view features to a flat road plane due to the spatial correlation between the warped top-view image and 3D lane geometry and was adopted in several other works [4, 9, 17, 23, 32]. However, IPM causes visual distortions in the top-view representation when the flat road assumption is violated. In related fields like BEV semantic segmentation, BEV transformations are learned via Multi-Layer-Perceptrons (MLPs) [18, 28], depth prediction [31, 37, 38] or transformer-based attention mechanisms [20, 30, 39]. In 3D lane detection, PersFormer [2] utilizes attention between front- and top-view, CurveFormer [1] introduces dynamic 3D anchors that model queries as parametric curves and BEV-LaneDet [46] uses MLPs for the spatial transformations. However, these learned transformations do not necessarily provide a 3D feature representation since they are not guided by valuable priors about the road surface geometry, which potentially results in unforeseen behavior for out-of-distribution data. Our approach instead aims for carefully modeling a geometry-aware feature space using a depth classification method inspired by [31] that exploits knowledge about the distribution of the road surface. + +# 3. Methodology + +The following section describes our 3D lane detection approach. An overview of the overall architecture is described and illustrated in Fig. 1. The main focus lies on our continuous 3D lane line representation, our regularization mechanism using physical priors and our prior-based spatial trans + +![](images/988c2dc4d8c350930fdb1f68812b31a63b7e9786bb66de69fea9c2e8a4452c32.jpg) +Figure 2. Our 3D lane line representation: For each proposal $\bar{f}$ (purple lines), line geometry is described by 3D B-Splines with control points $c_{k}$ (green dots). Each control point is determined by the offsets $\alpha_{k},\beta_{k}$ from the control points of the initial proposal in normal direction (orange vectors). Additionally, visibility $v(t)$ is modeled by splines with 1D control points $\gamma_{k}$ . + +formation module, which we explain in the following. + +# 3.1. Lane line representation + +Inspired by prior work in 3D lane detection [32], we leverage the benefits of continuous representations and employ a parametric model based on B-Splines. However, modeling only lateral $(x-)$ and vertical $(z-)$ components with spline-based functions (as done in previous approaches) is limited to lanes that merely progress along the longitudinal $(y-)$ direction. Instead, we propose the first full 3D lane line representation modeling each component $(x, y, z)$ such that we + +obtain + +$$ +\boldsymbol {f} (t) = \left( \begin{array}{c} x (t) \\ y (t) \\ z (t) \end{array} \right) = \sum_ {k = 1} ^ {K} \boldsymbol {c} _ {k} \cdot B _ {k, d} (t) \tag {1} +$$ + +with curve argument $t \in [0, 1]$ and $K$ control points $c_k = (x_k, y_k, z_k)^T$ . Each control point $c_k$ weights the respective basis function $B_{k,d}(t)$ (recursive polynomials of degree $d$ ) controlling the curve shape. + +Due to the ambiguity of curves using 3D B-Splines (the same spline curve can be described by different configurations of its control points), regressing all three dimensions per control point results in strong overfitting during training. We resolve this issue by limiting the degrees of freedom per control point to two and constraining the control points deflection to one direction in the $x - y$ -plane and one direction in the $y - z$ -plane as illustrated in Fig. 2. More precisely, the degrees of freedom per control point are specified by the directions of the normals $\mathbf{N}_{xy}$ and $\mathrm{N}_z$ of an initial curve proposal $\bar{f}$ with control points $\bar{c}_k = (\bar{x}_k,\bar{y}_k,\bar{z}_k)^T$ . The control points are then defined as + +$$ +\boldsymbol {c} _ {k} = \left( \begin{array}{l} x _ {k} \\ y _ {k} \\ z _ {k} \end{array} \right) = \left( \begin{array}{l} \bar {x} _ {k} + \mathrm {N} _ {x} \cdot \alpha_ {k} \\ \bar {y} _ {k} + \mathrm {N} _ {y} \cdot \alpha_ {k} \\ \bar {z} _ {k} + \mathrm {N} _ {z} \cdot \beta_ {k} \end{array} \right), \tag {2} +$$ + +where $\mathrm{N}_x, \mathrm{N}_y$ describe the $x$ - and $y$ -component of the normal vector $\mathbf{N}_{xy}$ in the $x$ - $y$ -plane. As shown in Eq. (2) and illustrated in Fig. 2, modeling splines as deflections in normal direction of its underlying initial line proposal only requires two parameters $\alpha_k, \beta_k$ per control point to describe the 3D shape. We use a wide variety of orientations for the initial proposals $\bar{\pmb{f}}$ (see Fig. 2), which allows us to detect any kind of lines with this formulation. More details about the initial proposals are provided in the supplementary. + +While [32] models the curve range using start- and endpoints that are learned by means of regression, we instead propose to model visibility using a continuous representation $v(t)$ and treat the visibility estimation as a classification problem. We obtain probability values applying sigmoid activation and consider $\sigma(v(t)) > 0.5$ the visible range. While in theory any kind of function can be utilized, we found that B-Splines with the same configuration as $f(t)$ are well-suited and introduce spline control points $\gamma_k$ defining the shape of $v(t)$ . + +Eventually, binary cross-entropy is used as a classification loss to learn visibility + +$$ +\mathcal {L} _ {v i s} = - \frac {1}{| \mathcal {P} _ {G T} |} \sum_ {\boldsymbol {p} \in \mathcal {P} _ {G T}} \hat {v} _ {\boldsymbol {p}} \cdot \log (\sigma (v (t _ {\boldsymbol {p}}))) + \tag {3} +$$ + +$$ +\left. \left(1 - \hat {v} _ {\boldsymbol {p}}\right) \cdot \log \left(1 - \sigma \left(v \left(t _ {\boldsymbol {p}}\right)\right)\right), \right. \tag {4} +$$ + +![](images/ec2e5bc323da2a922bc627d2bc7fa1a7741f98fc7dc54044906dc1789c137bed.jpg) +Figure 3. Illustration of different priors expressed by line tangents and surface normals. + +where $\mathcal{P}_{GT}$ denotes the ground truth set of points, $\hat{v}_p\in \{0,1\}$ the ground truth visibility for point $p$ . $t_p$ represents the respective curve argument obtained by orthogonal projection of $p$ onto the underlying line proposal. + +# 3.2. Regularization using physical priors + +In this section, we describe our regularization method to integrate prior knowledge about lane structure and surface geometry into our parametric line representation (see Fig. 3). + +Line parallelism. In order to reinforce parallel lines, the tangents at point pairs located in opposite normal direction on neighboring lines must be similar (see Fig. 3 left). We realize this by penalizing the cosine distance of the unit tangents $\mathbf{T}(t)$ on neighboring lines $i$ and $j$ for normal point pairs. More precisely, for each point $\pmb{p} \in \mathcal{P}^{(i)}$ on line $i$ we select the normal pair point $\pmb{p}^*$ on neighbor line $j$ that minimizes the distance to the normal plane, which is defined by the plane equation $\mathbf{T}^{(i)}(t)^T \cdot ((x,y,z)^T - \pmb{f}^{(i)}(t)) = 0$ . In Fig. 3 the normal planes are visualized in a 2D top-view as lines (orange) for simplicity. Hence the respective curve argument $t_{\pmb{p}^*}$ for point $\pmb{p}^*$ on line $j$ is given as + +$$ +t _ {\boldsymbol {p} ^ {*}} = \underset {\boldsymbol {p} ^ {\prime} \in \mathcal {P} ^ {(j)}} {\operatorname {a r g m i n}} \left. \mathbf {T} ^ {(i)} \left(t _ {\boldsymbol {p}}\right) ^ {T} \cdot \left(\boldsymbol {f} ^ {(j)} \left(t _ {\boldsymbol {p} ^ {\prime}}\right) - \boldsymbol {f} ^ {(i)} \left(t _ {\boldsymbol {p}}\right)\right), \right. \tag {5} +$$ + +where $\mathcal{P}^{(j)}$ denotes the points on line $j$ . While in theory Eq. (5) can be solved analytically, the simpler way is to sample the set of points $\mathcal{P}^{(j)}$ instead. (Note that our continuous representation allows us to choose high sampling rates without losing precision as no interpolation is required.) + +With the normal point pairs, we define the parallelism loss for a neighbor line pair based on the cosine distance of their tangents as + +$$ +\mathcal {L} _ {p a r} ^ {(i j)} = \frac {\mathbb {1} _ {\boldsymbol {p}} ^ {(i j)}}{| \mathcal {P} ^ {(i)} |} \cdot \sum_ {\boldsymbol {p} \in \mathcal {P} ^ {(i)}} 1 - \left(\mathbf {T} ^ {(i)} \left(t _ {\boldsymbol {p}}\right)\right) ^ {T} \cdot \mathbf {T} ^ {(j)} \left(t _ {\boldsymbol {p} ^ {*}}\right). \tag {6} +$$ + +Since the criterion of line parallelism should not hold for all normal point pairs of neighboring lines (e.g. merging or splitting lines), $\mathbb{1}_{\pmb{p}}^{(ij)} \in \{0, 1\}$ represents the indicator + +function determining whether the parallelism loss is applied to the point pair. More precisely, the function ensures that only the overlapping range of neighboring lines is taken into account. Furthermore, it determines whether the line pair should be considered as a parallel pair based on the standard deviation of euclidean distances between normal point pairs, i.e. high deviations indicate that the line pair might belong to a merge or split structure. In our experiments, we achieve state-of-the-art performance on test sets containing merges and splits, proving that our model is also capable of learning non-parallel line pairs using this indicator function. + +Surface smoothness. Since the lines reside on a smooth road, the surface normals of neighboring lanes should be similar. Analogously to $\mathcal{L}_{par}$ , we express this with the cosine distance between surface normals $\mathbf{N}^{(ih)}$ and $\mathbf{N}^{(ij)}$ as + +$$ +\mathcal {L} _ {s m} ^ {(i)} = \frac {\mathbb {1} _ {\boldsymbol {p}} ^ {(h i j)}}{| \mathcal {P} ^ {(i)} |} \cdot \sum_ {\boldsymbol {p} \in \mathcal {P} ^ {(i)}} 1 - \left(\mathbf {N} ^ {(i h)} \left(t _ {\boldsymbol {p}}\right)\right) ^ {T} \cdot \mathbf {N} ^ {(i j)} \left(t _ {\boldsymbol {p}}\right), \tag {7} +$$ + +with indicator function $\mathbb{1}_{\pmb{p}}^{(hij)}$ . The surface normal between line $i$ and left neighbor line $h$ at point $\pmb{p}$ can be expressed as the cross product of the tangent on line $i$ and the normalized connection vector between lines $i$ and $h$ , hence $\mathbf{N}^{(ih)}(t_{\pmb{p}}) = \mathbf{T}^{(i)}(t_{\pmb{p}}) \times \frac{\pmb{f}^{(h)}(t_{\pmb{p}^*}) - \pmb{f}^{(i)}(t_{\pmb{p}})}{||\pmb{f}^{(h)}(t_{\pmb{p}^*}) - \pmb{f}^{(i)}(t_{\pmb{p}})||}$ . For the normal between line $i$ and right neighbor $j$ the sign is flipped to obtain upwards pointing normal vectors. + +**Curvature.** We determine lane curvature by computing the second order derivatives as the difference of tangents at consecutive points divided by their euclidean distance as $\mathbf{T}'(t_p) = \frac{\mathbf{T}(t_p) - \mathbf{T}(t_p - \Delta t)}{||\mathbf{f}(t_p) - \mathbf{f}(t_p - \Delta t)||}$ . The maximum curvature in $x-y$ -plane (inverse curve radius) and in $z$ (rate of slope change) have very different value ranges and are therefore restricted by different limits. Hence, we define the two thresholds $\kappa_{xy}$ and $\kappa_z$ and formulate the curvature loss on line $i$ as + +$$ +\begin{array}{l} \mathcal {L} _ {c u r v} ^ {(i)} = \frac {1}{| \mathcal {P} ^ {(i)} |} \cdot \sum_ {\boldsymbol {p} \in \mathcal {P} ^ {(i)}} \max \left(\mathrm {T} _ {x y} ^ {\prime (i)} \left(t _ {\boldsymbol {p}}\right), \kappa_ {x y}\right) (8) \\ + \max \left(\mathrm {T} _ {z} ^ {\prime (i)} \left(t _ {\boldsymbol {p}}\right), \kappa_ {z}\right). (9) \\ \end{array} +$$ + +Finally, the prior regularization loss is given as + +$$ +\mathcal {L} _ {\text {p r i o r}} = \sum_ {i = 1} ^ {M} \lambda_ {s m} \mathcal {L} _ {s m} ^ {(i)} + \lambda_ {\text {c u r v}} \mathcal {L} _ {\text {c u r v}} ^ {(i)} + \sum_ {j = 1} ^ {N} \lambda_ {\text {p a r}} \mathcal {L} _ {\text {p a r}} ^ {(i j)}, \tag {10} +$$ + +with individual weights $\lambda_{par}$ , $\lambda_{sm}$ , $\lambda_{curv}$ . Note that all these properties are expressible by means of tangents and normals, which can be computed analytically on our parametric representation in continuous space. Consequently, minimization of the herein introduced prior losses does not depend on numerical approximations as is the case for anchor-, grid- or key-point representations. + +![](images/d1de4bcab885767d964f33dfd0955e89c71c85842a7b3cf317c0e6e7a880bf5a.jpg) +Figure 4. Our proposed spatial transformation module. First, several road surface hypotheses are defined (a) to which front-view features are lifted (b) and weighted according to the predicted depth distribution. Afterwards, point features are aggregated in a weighted manner to obtain the 3D feature map (c). + +# 3.3. Spatial transformation + +In this section, we describe our spatial transformation (shown in Fig. 4) that is leveraging valuable physical knowledge about surface geometry. We know that the road surface typically shows small deviations from the ground level $(z = 0)$ in the near-range and stronger deviations in the far-range. Based on this knowledge, we sample ground surface hypotheses that reflect the distribution of the road surface height profile (Fig. 4a). While in theory different types of surface functions could be utilized as hypotheses, we decide to merely rely on planes, since this facilitates the computation of ray intersections described in the following step. + +Next, the multi-scale front-view feature maps extracted by the backbone are lifted to 3D space (Fig. 4b). Our approach is inspired by [31], where front-view features are spreading along rays throughout the space of the road surface. These rays intersect with the surface hypotheses at different depths spanning a frustum-like point cloud in 3D space, where each point is affiliated with a $C$ -dimensional feature vector and additionally attached with its height value $z$ , hence, each point in the cloud has dimension $(C + 1)$ . The front-view feature map is propagated through a depth branch with a channel-wise softmax applied to obtain a categorical distribution for each ray, resulting in a tensor of size $H\times W\times S$ , where $H$ , $W$ denote height and width and channel size $S$ the number of surface hypotheses. + +In order to aggregate the information in 3D space, a BEV grid of size $X \times Y$ is defined. Features from points mapping to the same grid cell are weighted by the categorical depth distribution for the respective ray and accumulated in terms of a weighted sum (Fig. 4c). Since the $z$ -component + +
PriorsF1(%)↑X-near(m)↓X-far(m)↓Z-near(m)↓Z-far(m)↓
None65.00.3160.3840.1060.153
Par.66.20.2910.3730.1030.150
Surf.65.80.3200.3560.1030.144
Curv.66.70.3220.3660.1050.146
Comb.66.70.3010.3590.1030.144
+ +Table 1. Effect of different prior losses on OpenLane300. + +
# Surface Hypotheses1351527
F1-Score(%)↑65.065.966.666.166.0
+ +Table 2. Effect of the surface hypotheses on OpenLane300. + +
Lane Rep.Prior Reg.Spatial T.F1(%)↑Gain(%)
62.9(baseline)
✓65.0+2.1
✓✓66.7+3.8
✓✓66.6+3.7
✓✓✓66.9+4.0
+ +Table 3. Performance gain for different contributions on OpenLane300 using our novel Lane Representation, Prior Regularization and Spatial Transformation instead of IPM. + +of the points is also combined by a weighted sum, the value $z_{uv}$ can be interpreted as the height value of the surface for grid cell $(u, v)$ . We guide the model in learning the real surface and prevent it from learning an arbitrary mapping by introducing a simple grid-based regression loss as + +$$ +\mathcal {L} _ {\text {s u r f}} = \frac {1}{X \cdot Y} \sum_ {(u, v) \in X \times Y} \mathbb {1} _ {u v} \cdot \| z _ {u v} - \hat {z} _ {u v} \| _ {1}, \tag {11} +$$ + +with $\mathbb{1}_{uv}$ indicating whether surface ground truth $\hat{z}_{uv}$ is available for cell $(u,v)$ . The height ground truth is obtained by interpolation of the 3D lane annotations at cell locations. + +# 3.4. Loss functions + +The overall loss used during training is given as the weighted sum of loss components + +$$ +\mathcal {L} = \lambda_ {p r} \mathcal {L} _ {p r} + \lambda_ {c a t} \mathcal {L} _ {c a t} + \lambda_ {r e g} \mathcal {L} _ {r e g} + \tag {12} +$$ + +$$ +\lambda_ {v i s} \mathcal {L} _ {v i s} + \lambda_ {p r i o r} \mathcal {L} _ {p r i o r} + \lambda_ {s u r f} \mathcal {L} _ {s u r f}. \tag {13} +$$ + +We use focal loss [21] for lane presence $\mathcal{L}_{pr}$ and category classification $\mathcal{L}_{cat}$ . For the regression loss $\mathcal{L}_{reg}$ , we adapt the formulation of [32] to three instead of two dimensions. More details are provided in the supplementary. + +# 4. Experiments + +We first describe our experimental setup and then analyze our approach on two 3D lane datasets. + +![](images/d8f50475f56b7aba490beb4d19e4267cd4ec7c906cfdcffb6c0cc88bc6a4a7a5.jpg) +Front-view + +![](images/4ef57ac9d6f77182779fcca8b7d4ad9ac45485f1d14a7e2bdc9683598ac864d3.jpg) + +![](images/b1c6df6aaf53bfd0000d1ffa8f888fe304f8da0a3b5ecdab4015a24ae931f747.jpg) +Front-view + +![](images/cea5e6527d9d5251b5eb2a2e3f1699a0379f260c07652769b87f01b4b783d563.jpg) + +![](images/2139bb034d4a0cc9d059fbf2d02f4db9fbf7aae7d5502c6c9657ecdc18d131da.jpg) +(a) +Figure 5. Qualitative comparison of our model trained with prior regularization to the same model without regularization both trained on OpenLane300 with main differences highlighted by arrows. As a reference ground truth lines are visualized dashed. + +![](images/8a4b18d813c32b0f089a8931428c4b75653410ced963493e011883d811c1096a.jpg) +(b) + +# 4.1. Experimental setup + +We evaluate our method on two different datasets: OpenLane and Apollo 3D Synthetic - both containing 3D lane ground truth as well as camera parameters per frame. + +OpenLane [2] is a real-world dataset containing 150,000 images in the training and 40,000 in the test set from 1000 different sequences. In order to evaluate different driving scenarios the test set is divided into different situations, namely Up & Down, Curve, Extreme Weather, Night, Intersection and Merge & Split. For ablation studies we use the smaller version OpenLane300 including 300 sequences. + +Apollo 3D Synthetic [9] is a small synthetic dataset, consisting of only 10,500 examples from rather simple scenarios of highway, urban and rural environments. The data is split into three subsets, (1) Standard (simple) scenarios, (2) Rare Scenes and (3) Visual Variations. + +Evaluation metrics. For the quantitative evaluation both datasets utilize the evaluation scheme proposed in [9]. It evaluates the euclidean distance at uniformly distributed points in the range of $0 - 100\mathrm{m}$ along the $y$ direction. Based on the mean distance and range, $Fl$ -Score is computed, as well as the mean $x$ - and $z$ -errors in near- $(0 - 40\mathrm{m})$ and far-range $(40 - 100\mathrm{m})$ to evaluate geometric accuracy. + +Baseline. Our approach builds up on 3D-SpLineNet. Since it was applied on synthetic data only, it showed poor performance on real data. We applied some straightforward design adaptations - e.g. larger backbone, multiscale features (see supplementary) - and use this modified 3D-SpLineNet as our baseline (first row Table 3). + +
MethodF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [6]44.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [9]32.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [2]50.50.4850.5530.3640.43142.455.648.646.640.050.7
PersFormer* [2]53.10.3610.3280.1240.12946.858.754.048.441.452.5
CurveFormer [1]50.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [46]58.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane [12]53.70.2760.3110.1070.13846.757.252.547.845.451.2
Anchor3DLane-T [12]54.30.2750.3100.1050.13547.258.052.748.745.851.7
LaneCPP (Ours)60.30.2640.3100.0770.11753.664.456.754.952.058.7
+ +Table 4. Quantitative comparison on OpenLane [2]. Best performance and second best are highlighted. The scenario categories are Up and Down (U&D), Curve (C), Extreme Weather (EW), Night (N), Intersection (I), Merge and Split (M&S). PersFormer* denotes the latest performance reported on the official code base, Anchor3DLane-T represents the temporal multi-frame method of [12]. + +![](images/448e72ae7ae1e35ac542cb83f5290e47082516b13502ec82a11e9167a004a195.jpg) + +![](images/b1025d877499044a6cc701c3d2d84f82a0c0639438fb8d45a60dde87feff7ba5.jpg) + +![](images/c5cead738cf41e319a1183663c7702360da8a0769fc98d72bc06f6a98060f51c.jpg) + +![](images/c005d70086866981a42937a877a761e0a9f9d9e9d4bf15da9d840ce28b4d4bbd.jpg) + +![](images/448dd77f44973c75ff32a31560e0c84bacf7755cb35e8c408d8da6b6f8dd5a4d.jpg) + +![](images/8fe17f78161de68b46601a22002c94452a439b37f0c1816b160f56bc0aefb07e.jpg) + +![](images/3671b8e28ca4ab8accad2368143da3e9f789f3d40145ed2fe46ba41a2fac6ecf.jpg) + +![](images/2e09f8a23b45c6c49babfd81ae9f26a97c28460450f73e1400ce75e9bafa98d1.jpg) + +![](images/96d046c2e04dba0541828eec494db8b826bcfb1223f1acdc07c63f6ccaaa3bd7.jpg) + +![](images/ec75a55116b88d66845331c61fccb2a12d696d00a92d2dcf24ae1c56b35013b1.jpg) + +![](images/3444e3b48b655990b1f0b25d82dccb6d9bdbf7f92086c3c390b68c09491cc803.jpg) +(a) + +![](images/7fadd5d84a5837a9291f902c12685dd462fd31009a7857bd92afd2defd26bdb7.jpg) +(b) + +![](images/c56e09f51d0bd0fbacc6de1527732c23acc442a6b853e08a026b3ccbce670bff.jpg) +(c) +Figure 6. Qualitative comparison on OpenLane. Our method is compared to PersFormer* with ground truth visualized as dashed lines. + +![](images/2ae2c4c7081a6fb245227744324630c8fa59e554a633a2aa96ba4a3df74e7bd6.jpg) +(d) + +![](images/d0125e786f8849c86237e2b209555ad669774de6694d2f9697a7804128f96aa5.jpg) +(e) + +Implementation details. We use input size $360 \times 480$ and adopt the same backbone as in [2] based on a modified EfficientNet [42]. We extract four feature maps of resolutions $\left[\frac{1}{2}, \frac{1}{4}, \frac{1}{8}, \frac{1}{16}\right]$ . The final 3D feature map has size $26 \times 16$ with 64 channels. We use $M = 64$ initial line proposals and B-Splines of degree $d = 3$ and $K = 10$ control points. We apply Adam optimizer [14] with an initial learning rate of $2 \times 10^{-4}$ for OpenLane and $10^{-4}$ for Apollo and a dataset specific step-wise scheduler. We train for 30 epochs on OpenLane and 300 epochs on Apollo with batch size 16. For more details we refer to the supplementary. + +# 4.2. Ablation studies + +Table 1 indicates the effect of our proposed prior-based regularization. It is evident that each prior improves the F1-Score as well as geometric errors. While the surface and curvature priors result in better far-range estimates, line parallelism supports X-regression in the near-range. Besides, using surface smoothness loss results in lowest Z-far errors. Finally, a combination of priors yields a good balance of + +F1-Score and geometric errors. The positive effect of parallelism is confirmed by Fig. 5, where reinforcing parallel lane structure leads to better estimates in the near-range (a) and far-range (b) compared to the unregularized model. Learning parallel lines also is evidently beneficial in cases of poor visibility (b) and occlusions (a). In the latter case, the regularized model even shows better predictions than the noisy ground truth. This emphasizes the high relevance of priors for more robust behavior for real-world datasets, where 3D ground truth often comes with inaccuracies. + +For the spatial transformation (see Table 2), too low numbers of surface hypotheses result in worse score, presumably as 3D geometry is not captured sufficiently, whereas larger numbers tend to decreasing performance due to the higher complexity. The best F1-Score is obtained with 5 hypotheses, which is chosen for further experiments. While the improvement over IPM is already considerable, we think that with the simplifications of plane hypotheses prevent the component from developing its full potential. We see ways to enhance the 3D transformation even further + +
MethodBalanced ScenesRare Scenes
F1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfarF1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfar
3D-LaneNet [6]86.40.0680.4770.0150.20272.00.1660.8550.0390.521
GP [17]91.90.0490.3870.0080.21383.70.1260.9030.0230.625
PersFormer [2]92.90.0540.3560.010.23487.50.1070.7820.0240.602
3D-SpLineNet [32]96.30.0370.3240.0090.21392.90.0770.6990.0210.562
CurveFormer [1]95.80.0780.3260.0180.21995.60.1820.7370.0390.561
BEV-LaneDet [46]96.90.0160.2420.020.21697.60.0310.5940.0400.556
Anchor3DLane [12]95.40.0450.3000.0160.22394.40.0820.6990.0300.580
LaneCPP (Ours)97.40.0300.2770.0110.20696.20.0730.6510.0230.543
+ +Table 5. Quantitative comparison of best methods on Apollo 3D Synthetic [9]. Best performance and second best are highlighted. + +using more sophisticated spatial representations in future. + +The impact of our different contributions is summarized in Table 3, where the first row shows our baseline (see Sec. 4.1). More than two percent in F1-Score are gained with our novel lane representation compared to the simplified one from [32]. Moreover, it is clear that both, the regularization using combined priors and the spatial transformation using 5 hypotheses result in significant improvement. Eventually, combining all components yields the best model configuration, which we choose for further evaluation. + +# 4.3. Evaluation on OpenLane + +On the real-world OpenLane benchmark our model evidently outperforms all other methods with respect to F1-Score as well as geometric errors as shown in Table 4. Compared to BEV-Lanedet, which achieves a high detection score, our model gains $+1.9\%$ , while reaching significantly lower geometric errors. In comparison to Anchor3DLane the improvements with respect to X-errors are less substantial, however, our approach surpasses the F1-Score by a large gap of $+6.6\%$ . Analyzing the detection scores among different scenarios, outstanding performance gain is observed on the up- and down-hill test set $(+5.9\%)$ that highlights the capability of our approach to capture 3D space proficiently, which is supported by the low Z-errors. + +Apart from quantitative results, we show qualitative examples in Fig. 6. In up-hill scenarios like Fig. 6b our model manages to estimate both lateral and height profile accurately, since our assumptions about road surface and line parallelism are satisfied. In contrast, PersFormer lacks spatial features and does not use any kind of physical regularization. Consequently, it fails to estimate the 3D lane geometry and even collapses in Fig. 6c, whereas our surface and curvature priors always prevent such a behavior. Noteworthy is also the top performance on the merges and splits set. This proves that our soft regularization is even capable to handle situations containing non-parallel lines, which is also confirmed by Fig. 6d. However, we rarely observe limi + +tations with our formulation for line pairs with a similar orientation but weakly converging course as shown in Fig. 6e. In such cases the indicator function might erroneously decide for parallelism loss during training. One possible solution for future work would be to consider ground truth for the indicator function to identify such situations. + +# 4.4. Evaluation on Apollo 3D Synthetic + +The Apollo 3D Synthetic dataset is very limited in size and only consists of simple situations in contrast to OpenLane. While we find the results on OpenLane more meaningful, we would like to still provide and discuss the quantitative results on the Apollo dataset. Due to the simplicity of the dataset, our model cannot benefit that significantly from our priors but still achieves competitive results to state of the art with the highest F1-Score on the balanced scenes dataset and comparable error metrics (second best for most errors). + +# 5. Conclusions and future work + +In this work, we present LaneCPP, a novel approach for 3D lane detection that leverages physical prior knowledge about lane structure and road geometry. Our new continuous lane representation overcomes previous deficiencies by allowing arbitrary lane structures and enables us to regularize lane geometry based on analytically formulated priors. We further introduce a novel spatial transformation module that models 3D features carefully considering knowledge about road surface geometry. In our experiments, we demonstrate state-of-the-art performance on real and synthetic benchmarks. The full capability of our approach is revealed on real-world OpenLane, for which we prove the relevance of priors quantitatively and qualitatively. In future, priors could be individualized for different driving scenarios and might support to learn inter-lane relations to achieve better scene understanding in a global context. We also see ways to leverage the full potential of the spatial transformation by using more sophisticated surface representations. + +# References + +[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 7, 8 +[2] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2, 3, 4, 6, 7, 8 +[3] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2 +[4] Netalee Efrat, Max Bluvstein, Shaul Oron, Dan Levi, Noa Garnett, and Bat El Shlomo. 3d-lanenet+: Anchor free lane detection using a semi-local representation. arXiv/2011.01535, 2020. 1, 2, 3 +[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 7, 8 +[7] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1, 2 +[8] Stephen Gould, Richard Hartley, and Dylan Campbell. Deep declarative networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2021. 2 +[9] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 4, 6, 7, 8 +[10] Steffen Hagedorn, Marcel Milich, and Alexandru P. Condurache. Pioneering se (2)-equivariant trajectory planning for automated driving. arXiv:2403.11304, 2024. 2 +[11] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2 +[12] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8 +[13] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1, 2 + +[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of the International Conf. on Learning Representations (ICLR), 2015. 7 +[15] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020. 1, 2 +[16] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1, 2 +[17] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8 +[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3 +[19] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 2 +[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 3 +[21] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 6 +[22] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2 +[23] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 1, 2, 3 +[24] Pingping Lu, Chen Cui, Shaobing Xu, Huei Peng, and Fan Wang. SUPER: A novel lane detection system. IEEE Trans. on Intelligent Vehicles (T-IV), 2021. 2 +[25] Hanspeter Mallot, Heinrich Blthoff, J.J. Little, and S Bohrer. Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological Cybernetics, 1991. 3 +[26] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1, 2 +[27] Marcos Nieto, Luis Salgado, Fernando Jaureguizar, and Jon Arróspide. Robust multiple lane road modeling based on perspective analysis. In Proc. IEEE International Conf. on Image Processing (ICIP), 2008. 2 + +[28] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics Autom. Lett., 2020. 3 +[29] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1, 2 +[30] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3 +[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 3, 5 +[32] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3, 4, 6, 8 +[33] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1, 2 +[34] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[35] Matthias Rath and Alexandru Paul Condurache. Invariant integration in deep convolutional feature space. In Proc. of European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2020. 2 +[36] Matthias Rath and Alexandru Paul Condurache. Improving the sample-complexity of deep classification networks with invariant integration. In Proc. of International Joint Conf. on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), 2022. 2 +[37] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[38] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Proc. of the British Machine Vision Conf. (BMVC), 2019. 3 +[39] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3 +[40] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2 +[41] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. + +Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2 +[42] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. of the International Conf. on Machine learning (ICML), 2019. 7 +[43] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polyanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2 +[44] Bingke Wang, Zilei Wang, and Yixin Zhang. Polynomial regression network for variable-number lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2 +[45] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[46] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8 +[47] Yuping Wang and Jier Chen. Eqdrive: Efficient equivariant motion forecasting with multi-modality for autonomous driving. arXiv:2310.17540, 2023. 2 +[48] Lu Xiong, Zhenwen Deng, Peizhi Zhang, and Zhiqiang Fu. A 3d estimation of structural road surface based on lane-line information. IFAC Conf. on Engine and Powertrain Control, Simulation and Modeling (E-COSM), 2018. 2 +[49] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[50] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[51] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. RESA: recurrent feature-shift aggregator for lane detection. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2021. 2 +[52] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[53] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1, 2 \ No newline at end of file diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/images.zip b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e782edb6dc2e43075e4942352ff955269433f637 --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57d0dbed471aea68eaa105cc8c7a4873a525ed31ff607fd30a67bca6349ebe63 +size 636660 diff --git a/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/layout.json b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d182adde1aebd0da77a5a2ca22deb9eab5240ee7 --- /dev/null +++ b/2024/LaneCPP_ Continuous 3D Lane Detection using Physical Priors/layout.json @@ -0,0 +1,9978 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 492, + 120 + ], + "type": "text", + "content": "LaneCPP: Continuous 3D Lane Detection using Physical Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "spans": [ + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "text", + "content": "Maximilian Pittner" + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "text", + "content": ", Joel Janai" + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "text", + "content": ", Alexandru P. Condurache" + }, + { + "bbox": [ + 143, + 142, + 449, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 180, + 157, + 415, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 180, + 157, + 415, + 171 + ], + "spans": [ + { + "bbox": [ + 180, + 157, + 415, + 171 + ], + "type": "text", + "content": "1Bosch Mobility Solutions, Robert Bosch GmbH" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 170, + 171, + 424, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 171, + 424, + 185 + ], + "spans": [ + { + "bbox": [ + 170, + 171, + 424, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 171, + 424, + 185 + ], + "type": "text", + "content": "Institute of Signal Processing, University of Lübeck" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 187, + 487, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 187, + 487, + 198 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 487, + 198 + ], + "type": "text", + "content": "{Maximilian.Pittner, Joel.Janai, AlexandruPaul.Condurache}@de.bosch.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 190, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 251, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 289, + 526 + ], + "type": "text", + "content": "Monocular 3D lane detection has become a fundamental problem in the context of autonomous driving, which comprises the tasks of finding the road surface and locating lane markings. One major challenge lies in a flexible but robust line representation capable of modeling complex lane structures, while still avoiding unpredictable behavior. While previous methods rely on fully data-driven approaches, we instead introduce a novel approach LaneCPP that uses a continuous 3D lane detection model leveraging physical prior knowledge about the lane structure and road geometry. While our sophisticated lane model is capable of modeling complex road structures, it also shows robust behavior since physical constraints are incorporated by means of a regularization scheme that can be analytically applied to our parametric representation. Moreover, we incorporate prior knowledge about the road geometry into the 3D feature space by modeling geometry-aware spatial features, guiding the network to learn an internal road surface representation. In our experiments, we show the benefits of our contributions and prove the meaningfulness of using priors to make 3D lane detection more robust. The results show that LaneCPP achieves state-of-the-art performance in terms of F-Score and geometric errors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 127, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 653 + ], + "type": "text", + "content": "Robust and precise lane detection systems build one of the most essential components in the perception stack of autonomous vehicles. While some approaches utilize LiDAR sensors or multi-sensor setups, the application of monocular cameras has become more popular due to their lower cost and the high-resolution visual representation that provides valuable information to detect lane markings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "text", + "content": "In the past, lane detection was mainly treated as a 2D detection task. Deep learning based methods achieved good results by treating the problem as a segmentation task in pixel space [7, 11, 16, 26, 29, 33, 53], used to classify and regress lanes using anchor-based [19, 41] representations," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 335 + ], + "type": "text", + "content": "or as key-points on a grid structure [13, 15, 34, 45]. However, due to the lack of depth information, these 2D representations fail to model lane markings and road geometry in 3D space, which forms an important prerequisite for later functionalities like trajectory planning. Consequently, approaches for monocular 3D lane detection were introduced, which adapted lane representations for the 3D domain by modeling vertical anchors [6, 9] or local segments on a grid [4] in a Birds-Eye-View (BEV) oriented 3D-frame." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "A crucial topic for the application of lane detection algorithms in autonomous systems is safety, which requires predictable and robust behavior in any traffic situation. One risk of learning-based methods is the tendency to show unpredictable behavior in cases of rarely observed scenarios. Since obtaining large amounts of data with high-quality annotations is cumbersome and expensive, publicly available 3D datasets are limited in size and accuracy. Hence, they do not reflect the variability of real-world scenarios sufficiently. This makes learning-based models prone to overfitting, and eventually, diminishes predictability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 555 + ], + "type": "text", + "content": "One common way to deal with such problems is the integration of prior knowledge. Physics provides us a profound understanding of the 3D world, allowing us to make valid assumptions about the lane structure and road surface geometry. Therefore, we introduce physically motivated priors into the lane detection objective to cope with the limited data problem and achieve robust and predictable behavior." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "There are certain geometric properties that should generally hold for detected lane lines. For instance, we know that most lines progress parallel to each other, reside on a smooth surface and should not exceed certain thresholds in terms of curvature and slope. However, integrating such assumptions into prevailing discrete representations is not straight forward as strong simplifications are necessary. In contrast, continuous 3D lane representations directly provide parametric curves using polynomials [1, 23] or more sophisticated B-Splines [32]. These allow for analytical computations on the curve function, which enables the integration of such priors into the lane representation. By modeling these priors explicitly instead of learning them from" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10639" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "data, the model can focus its full capacity on learning richer features for the lane detection task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 203 + ], + "type": "text", + "content": "We can further use physical knowledge about the road geometry to support the model in learning an internal transformation from image features to 3D space. While methods based on Inverse Perspective Mapping (IPM) [4, 6, 9, 17, 23, 32] make false flat-ground assumptions, learning based transformations [1, 2, 46] completely ignore road properties. In contrast, integrating prior knowledge about the road surface allows us to model 3D features geometry-aware and helps the network to focus on the 3D region of interest." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 286, + 300 + ], + "type": "text", + "content": "Thus, we propose a novel 3D lane detection approach named LaneCPP that leverages valuable prior knowledge to achieve accurate and robust perception behavior. It introduces a new sophisticated continuous curve representation, which enables us to incorporate physical priors. In addition, we present a spatial transformation component for learning a physically inspired mapping from 2D image to 3D space providing meaningful spatial features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 268, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 268, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 268, + 311 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 312, + 286, + 431 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 312, + 286, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 286, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 286, + 347 + ], + "type": "text", + "content": "- We propose a novel architecture for 3D lane detection from monocular images using a more sophisticated flexible parametric spline-based lane representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 348, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 286, + 371 + ], + "type": "text", + "content": "- We present a way to incorporate priors about lane structure and geometry into our continuous representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 372, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 286, + 395 + ], + "type": "text", + "content": "- We introduce a new way to use prior knowledge about the road surface geometry for learning spatial features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 286, + 418 + ], + "type": "text", + "content": "- We demonstrate the benefits of our contributions in several ablation studies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 419, + 265, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 265, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 265, + 431 + ], + "type": "text", + "content": "- We show state-of-the-art performance of our model." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 442, + 130, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 130, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 130, + 453 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 713 + ], + "type": "text", + "content": "Different Lane Representations. An important design choice in deep learning based lane detection is the representation that the network uses to model lane line geometry, which can be categorized as follows: 1) Pixelwise representations, which formulate lane detection as a segmentation problem, were used mainly in 2D methods [7, 11, 16, 26, 29, 33, 51, 53] and were adopted in 3D by SALAD [50] combining line segmentation with depth-prediction. These representations come with high computational load since a large amount of parameters is required. 2) Grid-based approaches divide the space into cells and model lanes using local segments [13] or keypoints [15, 34, 45]. 3D-LaneNet+ [4] suggests to use local line-segments and BEV-LaneDet [46] defines key-points on a BEV grid representation. Both depend on the grid resolution and require costly post-processing to obtain lines. 3) Anchor-based representations [19, 40, 41, 52] model lines as straight anchors with positional offsets at predefined locations. They are widely used in 3D detection approaches including 3D-LaneNet [6] and Gen-LaneNet [9], which use vertical anchors in the top-view, and An" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "chor3DLane [12], introducing anchor projection with iterative regression. Similar to grid-based representations, it requires subsequent curve-fitting to obtain smooth lines. 4) Continuous curve representations [5, 22, 24, 43, 44] instead directly model smooth curves without requiring costly post-processing. While CLGO [23] and CurveFormer [1] use simple polynomials, 3D-SpLineNet [32] proposes B-Splines [3]. Since B-Splines offer local control over curve segments, they are compatible to model complex shapes with low-degree basis functions, while polynomials and Bezier curves show global dependence and thus require higher degrees causing expensive computation. Although 3D-SpLineNet achieves superior detection performance on synthetic data, it unfortunately lacks flexibility as the curve formulation is limited to monotonically progressing lanes, making it hardly applicable to real-world data. To resolve this issue, we propose a more flexible representation based on actual 3D B-Splines. In contrast to discrete grids and anchors, continuous representation even allows us to integrate prior knowledge in an analytical manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 315, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 674 + ], + "type": "text", + "content": "Geometry Priors. Several approaches suggest to incorporate prior knowledge into learning-based methods, e.g. by integrating invariance into the model architecture [35, 36] or task-specific transformations as for trajectory planning [10, 47, 49]. In the field of lane detection, line parallelism has been formulated as a hard constraint to resolve depth ambiguity and determine camera parameters [27, 48]. Deep declarative networks [8] offer a general framework to incorporate arbitrary properties as constraints, by solving a constrained optimization problem in the forward pass. While such methods are appropriate when hard constraints must be enforced, our goal is rather to guide the network in learning typical geometric lane properties by formulating soft constraints in a regularization objective. Such a regularization only affects training and does not require resolving an optimization problem in the forward pass, and thus, comes without additional computational cost during inference. Following this paradigm, SGNet [24, 40] proposes to penalize the deviation of lateral distance from a constant lane width in the IPM warped top-view, but ignores that the property does not hold for lines deviating from the ground plane. GP [17] presents a parallelism loss that enforces constant distance between nearest neighbors locally, which depends on the number of anchor points. In contrast, our method presents a way to learn parallelism globally and independent of resolutions of discrete lane representations. We propose an elegant way to learn parallelism as well as other geometry priors using analytical formulations of tangents and normals, which are well-defined on our continuous spline representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Leveraging 3D Features. An important model component consists in the extraction of 3D features, encoding valuable information to detect lanes along the road surface." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10640" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 69, + 542, + 222 + ], + "blocks": [ + { + "bbox": [ + 51, + 69, + 542, + 222 + ], + "lines": [ + { + "bbox": [ + 51, + 69, + 542, + 222 + ], + "spans": [ + { + "bbox": [ + 51, + 69, + 542, + 222 + ], + "type": "image", + "image_path": "8c794e6cb38623d03b88c4f43ce7126e1c660d926c4bc4c53cce5ce995d69c5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 546, + 277 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 546, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 546, + 277 + ], + "type": "text", + "content": "Figure 1. Our approach: First, front-view image " + }, + { + "bbox": [ + 46, + 231, + 546, + 277 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 231, + 546, + 277 + ], + "type": "text", + "content": " is propagated through the backbone extracting multi-scale feature maps. These are transformed to 3D using our spatial transformation and then fused to obtain a single 3D feature map. Feature pooling is applied to obtain features for each line proposal that are propagated through fully connected layers to obtain the parameters for our line representation. Finally, prior knowledge is exploited to regularize the lane representation and to produce surface hypotheses for the spatial transformation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 296, + 289, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 289, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 289, + 620 + ], + "type": "text", + "content": "While some works predict 3D lanes directly from the front-view, e.g. by utilizing pixel-wise depth estimation [50] or 3D anchor-projection mechanisms [12], prevalent methods employ an intermediate 3D or BEV feature representation with an internal transformation from the front-view to the 3D space. 3D-LaneNet [6] proposes to utilize IPM [25] to project front-view features to a flat road plane due to the spatial correlation between the warped top-view image and 3D lane geometry and was adopted in several other works [4, 9, 17, 23, 32]. However, IPM causes visual distortions in the top-view representation when the flat road assumption is violated. In related fields like BEV semantic segmentation, BEV transformations are learned via Multi-Layer-Perceptrons (MLPs) [18, 28], depth prediction [31, 37, 38] or transformer-based attention mechanisms [20, 30, 39]. In 3D lane detection, PersFormer [2] utilizes attention between front- and top-view, CurveFormer [1] introduces dynamic 3D anchors that model queries as parametric curves and BEV-LaneDet [46] uses MLPs for the spatial transformations. However, these learned transformations do not necessarily provide a 3D feature representation since they are not guided by valuable priors about the road surface geometry, which potentially results in unforeseen behavior for out-of-distribution data. Our approach instead aims for carefully modeling a geometry-aware feature space using a depth classification method inspired by [31] that exploits knowledge about the distribution of the road surface." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 632, + 130, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 130, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 130, + 647 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "The following section describes our 3D lane detection approach. An overview of the overall architecture is described and illustrated in Fig. 1. The main focus lies on our continuous 3D lane line representation, our regularization mechanism using physical priors and our prior-based spatial trans" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 318, + 293, + 541, + 464 + ], + "blocks": [ + { + "bbox": [ + 318, + 293, + 541, + 464 + ], + "lines": [ + { + "bbox": [ + 318, + 293, + 541, + 464 + ], + "spans": [ + { + "bbox": [ + 318, + 293, + 541, + 464 + ], + "type": "image", + "image_path": "988c2dc4d8c350930fdb1f68812b31a63b7e9786bb66de69fea9c2e8a4452c32.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": "Figure 2. Our 3D lane line representation: For each proposal " + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\bar{f}" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": " (purple lines), line geometry is described by 3D B-Splines with control points " + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": " (green dots). Each control point is determined by the offsets " + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\alpha_{k},\\beta_{k}" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": " from the control points of the initial proposal in normal direction (orange vectors). Additionally, visibility " + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "inline_equation", + "content": "v(t)" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": " is modeled by splines with 1D control points " + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "inline_equation", + "content": "\\gamma_{k}" + }, + { + "bbox": [ + 304, + 472, + 547, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 566, + 525, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 525, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 525, + 578 + ], + "type": "text", + "content": "formation module, which we explain in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 595, + 443, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 443, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 443, + 609 + ], + "type": "text", + "content": "3.1. Lane line representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": "Inspired by prior work in 3D lane detection [32], we leverage the benefits of continuous representations and employ a parametric model based on B-Splines. However, modeling only lateral " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(x-)" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " and vertical " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(z-)" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " components with spline-based functions (as done in previous approaches) is limited to lanes that merely progress along the longitudinal " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(y-)" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " direction. Instead, we propose the first full 3D lane line representation modeling each component " + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 617, + 545, + 715 + ], + "type": "text", + "content": " such that we" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10641" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 76, + 83 + ], + "type": "text", + "content": "obtain" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 92, + 89, + 287, + 128 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 89, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 92, + 89, + 287, + 128 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} (t) = \\left( \\begin{array}{c} x (t) \\\\ y (t) \\\\ z (t) \\end{array} \\right) = \\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {k} \\cdot B _ {k, d} (t) \\tag {1}", + "image_path": "f5f4dbfe27de5066c5c9a193d5592af6994283280578db936614b625415ffcf2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": "with curve argument " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "t \\in [0, 1]" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": " control points " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "c_k = (x_k, y_k, z_k)^T" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": ". Each control point " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": " weights the respective basis function " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "B_{k,d}(t)" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": " (recursive polynomials of degree " + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 134, + 287, + 184 + ], + "type": "text", + "content": ") controlling the curve shape." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": "Due to the ambiguity of curves using 3D B-Splines (the same spline curve can be described by different configurations of its control points), regressing all three dimensions per control point results in strong overfitting during training. We resolve this issue by limiting the degrees of freedom per control point to two and constraining the control points deflection to one direction in the " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "x - y" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": " -plane and one direction in the " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "y - z" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": " -plane as illustrated in Fig. 2. More precisely, the degrees of freedom per control point are specified by the directions of the normals " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{xy}" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_z" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": " of an initial curve proposal " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\bar{f}" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": " with control points " + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\bar{c}_k = (\\bar{x}_k,\\bar{y}_k,\\bar{z}_k)^T" + }, + { + "bbox": [ + 46, + 185, + 287, + 331 + ], + "type": "text", + "content": ". The control points are then defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 338, + 287, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 338, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 93, + 338, + 287, + 377 + ], + "type": "interline_equation", + "content": "\\boldsymbol {c} _ {k} = \\left( \\begin{array}{l} x _ {k} \\\\ y _ {k} \\\\ z _ {k} \\end{array} \\right) = \\left( \\begin{array}{l} \\bar {x} _ {k} + \\mathrm {N} _ {x} \\cdot \\alpha_ {k} \\\\ \\bar {y} _ {k} + \\mathrm {N} _ {y} \\cdot \\alpha_ {k} \\\\ \\bar {z} _ {k} + \\mathrm {N} _ {z} \\cdot \\beta_ {k} \\end{array} \\right), \\tag {2}", + "image_path": "ad396f94b08c1c5d0e885f61521fd69bc7e7525c54ef712864b26763dc88d7d8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{N}_x, \\mathrm{N}_y" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": " describe the " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": "-component of the normal vector " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{xy}" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": "-plane. As shown in Eq. (2) and illustrated in Fig. 2, modeling splines as deflections in normal direction of its underlying initial line proposal only requires two parameters " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\alpha_k, \\beta_k" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": " per control point to describe the 3D shape. We use a wide variety of orientations for the initial proposals " + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{f}}" + }, + { + "bbox": [ + 46, + 382, + 287, + 491 + ], + "type": "text", + "content": " (see Fig. 2), which allows us to detect any kind of lines with this formulation. More details about the initial proposals are provided in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": "While [32] models the curve range using start- and endpoints that are learned by means of regression, we instead propose to model visibility using a continuous representation " + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "inline_equation", + "content": "v(t)" + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": " and treat the visibility estimation as a classification problem. We obtain probability values applying sigmoid activation and consider " + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\sigma(v(t)) > 0.5" + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": " the visible range. While in theory any kind of function can be utilized, we found that B-Splines with the same configuration as " + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "inline_equation", + "content": "f(t)" + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": " are well-suited and introduce spline control points " + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\gamma_k" + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": " defining the shape of " + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "inline_equation", + "content": "v(t)" + }, + { + "bbox": [ + 46, + 491, + 287, + 609 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 609, + 287, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 287, + 634 + ], + "type": "text", + "content": "Eventually, binary cross-entropy is used as a classification loss to learn visibility" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 640, + 287, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 640, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 73, + 640, + 287, + 670 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v i s} = - \\frac {1}{| \\mathcal {P} _ {G T} |} \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} _ {G T}} \\hat {v} _ {\\boldsymbol {p}} \\cdot \\log (\\sigma (v (t _ {\\boldsymbol {p}}))) + \\tag {3}", + "image_path": "3ebaa8d5aee3a35c2d644d8ae0ba71e04171fd20b3bf49035c8922945131542f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 672, + 287, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 672, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 102, + 672, + 287, + 687 + ], + "type": "interline_equation", + "content": "\\left. \\left(1 - \\hat {v} _ {\\boldsymbol {p}}\\right) \\cdot \\log \\left(1 - \\sigma \\left(v \\left(t _ {\\boldsymbol {p}}\\right)\\right)\\right), \\right. \\tag {4}", + "image_path": "27ab45d64e3bb32b18ac7a36dcacc44db314ebe8ef9e1d81da5d26d7de0fb1ea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 540, + 182 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 540, + 182 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 540, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 540, + 182 + ], + "type": "image", + "image_path": "ec2e5bc323da2a922bc627d2bc7fa1a7741f98fc7dc54044906dc1789c137bed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 189, + 545, + 211 + ], + "lines": [ + { + "bbox": [ + 305, + 189, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 189, + 545, + 211 + ], + "type": "text", + "content": "Figure 3. Illustration of different priors expressed by line tangents and surface normals." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{GT}" + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": " denotes the ground truth set of points, " + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\hat{v}_p\\in \\{0,1\\}" + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": " the ground truth visibility for point " + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "inline_equation", + "content": "t_p" + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": " represents the respective curve argument obtained by orthogonal projection of " + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 232, + 545, + 281 + ], + "type": "text", + "content": " onto the underlying line proposal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 288, + 499, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 499, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 499, + 301 + ], + "type": "text", + "content": "3.2. Regularization using physical priors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 342 + ], + "type": "text", + "content": "In this section, we describe our regularization method to integrate prior knowledge about lane structure and surface geometry into our parametric line representation (see Fig. 3)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": "Line parallelism. In order to reinforce parallel lines, the tangents at point pairs located in opposite normal direction on neighboring lines must be similar (see Fig. 3 left). We realize this by penalizing the cosine distance of the unit tangents " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t)" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " on neighboring lines " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " for normal point pairs. More precisely, for each point " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p} \\in \\mathcal{P}^{(i)}" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " on line " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " we select the normal pair point " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^*" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " on neighbor line " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " that minimizes the distance to the normal plane, which is defined by the plane equation " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^{(i)}(t)^T \\cdot ((x,y,z)^T - \\pmb{f}^{(i)}(t)) = 0" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": ". In Fig. 3 the normal planes are visualized in a 2D top-view as lines (orange) for simplicity. Hence the respective curve argument " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "t_{\\pmb{p}^*}" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " for point " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "\\pmb{p}^*" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " on line " + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": " is given as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 495, + 545, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 495, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 315, + 495, + 545, + 519 + ], + "type": "interline_equation", + "content": "t _ {\\boldsymbol {p} ^ {*}} = \\underset {\\boldsymbol {p} ^ {\\prime} \\in \\mathcal {P} ^ {(j)}} {\\operatorname {a r g m i n}} \\left. \\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right) ^ {T} \\cdot \\left(\\boldsymbol {f} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {\\prime}}\\right) - \\boldsymbol {f} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right), \\right. \\tag {5}", + "image_path": "fc1b1ca4e80e7410689b99c98cb0b04dfa325ea2b4d690439b01781b3824be9a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(j)}" + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "text", + "content": " denotes the points on line " + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "text", + "content": ". While in theory Eq. (5) can be solved analytically, the simpler way is to sample the set of points " + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{(j)}" + }, + { + "bbox": [ + 304, + 528, + 545, + 589 + ], + "type": "text", + "content": " instead. (Note that our continuous representation allows us to choose high sampling rates without losing precision as no interpolation is required.)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 589, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 545, + 624 + ], + "type": "text", + "content": "With the normal point pairs, we define the parallelism loss for a neighbor line pair based on the cosine distance of their tangents as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 633, + 545, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 545, + 667 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p a r} ^ {(i j)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {T} ^ {(i)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {T} ^ {(j)} \\left(t _ {\\boldsymbol {p} ^ {*}}\\right). \\tag {6}", + "image_path": "d8ef4dc7255249533e9614f5c59c9ecc1fe2bd3ba86f2dc355fdc8be7ac57360.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 675, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 675, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 675, + 545, + 715 + ], + "type": "text", + "content": "Since the criterion of line parallelism should not hold for all normal point pairs of neighboring lines (e.g. merging or splitting lines), " + }, + { + "bbox": [ + 304, + 675, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\pmb{p}}^{(ij)} \\in \\{0, 1\\}" + }, + { + "bbox": [ + 304, + 675, + 545, + 715 + ], + "type": "text", + "content": " represents the indicator" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 714 + ], + "type": "text", + "content": "For the concept of visibility, we follow the prevailing definition from the literature [2, 9]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10642" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "function determining whether the parallelism loss is applied to the point pair. More precisely, the function ensures that only the overlapping range of neighboring lines is taken into account. Furthermore, it determines whether the line pair should be considered as a parallel pair based on the standard deviation of euclidean distances between normal point pairs, i.e. high deviations indicate that the line pair might belong to a merge or split structure. In our experiments, we achieve state-of-the-art performance on test sets containing merges and splits, proving that our model is also capable of learning non-parallel line pairs using this indicator function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": "Surface smoothness. Since the lines reside on a smooth road, the surface normals of neighboring lanes should be similar. Analogously to " + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{par}" + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": ", we express this with the cosine distance between surface normals " + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^{(ih)}" + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^{(ij)}" + }, + { + "bbox": [ + 46, + 204, + 288, + 251 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 265, + 287, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 265, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 52, + 265, + 287, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s m} ^ {(i)} = \\frac {\\mathbb {1} _ {\\boldsymbol {p}} ^ {(h i j)}}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} 1 - \\left(\\mathbf {N} ^ {(i h)} \\left(t _ {\\boldsymbol {p}}\\right)\\right) ^ {T} \\cdot \\mathbf {N} ^ {(i j)} \\left(t _ {\\boldsymbol {p}}\\right), \\tag {7}", + "image_path": "85cfc52de2480fdea712d197f44b06c9285b9fc458fbce73f1d4b894cb9cf275.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": "with indicator function " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\pmb{p}}^{(hij)}" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": ". The surface normal between line " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " and left neighbor line " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " can be expressed as the cross product of the tangent on line " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " and the normalized connection vector between lines " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": ", hence " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^{(ih)}(t_{\\pmb{p}}) = \\mathbf{T}^{(i)}(t_{\\pmb{p}}) \\times \\frac{\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})}{||\\pmb{f}^{(h)}(t_{\\pmb{p}^*}) - \\pmb{f}^{(i)}(t_{\\pmb{p}})||}" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": ". For the normal between line " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " and right neighbor " + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 306, + 287, + 397 + ], + "type": "text", + "content": " the sign is flipped to obtain upwards pointing normal vectors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": "**Curvature.** We determine lane curvature by computing the second order derivatives as the difference of tangents at consecutive points divided by their euclidean distance as " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{T}'(t_p) = \\frac{\\mathbf{T}(t_p) - \\mathbf{T}(t_p - \\Delta t)}{||\\mathbf{f}(t_p) - \\mathbf{f}(t_p - \\Delta t)||}" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": ". The maximum curvature in " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "x-y" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": "-plane (inverse curve radius) and in " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": " (rate of slope change) have very different value ranges and are therefore restricted by different limits. Hence, we define the two thresholds " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\kappa_{xy}" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "\\kappa_z" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": " and formulate the curvature loss on line " + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 398, + 287, + 508 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 510, + 287, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 510, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 75, + 510, + 287, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {c u r v} ^ {(i)} = \\frac {1}{| \\mathcal {P} ^ {(i)} |} \\cdot \\sum_ {\\boldsymbol {p} \\in \\mathcal {P} ^ {(i)}} \\max \\left(\\mathrm {T} _ {x y} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {x y}\\right) (8) \\\\ + \\max \\left(\\mathrm {T} _ {z} ^ {\\prime (i)} \\left(t _ {\\boldsymbol {p}}\\right), \\kappa_ {z}\\right). (9) \\\\ \\end{array}", + "image_path": "936a823ff6e946ae5a46a5852f368a073a3acf2a97cc34f771c3f74297088d27.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 237, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 237, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 237, + 574 + ], + "type": "text", + "content": "Finally, the prior regularization loss is given as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 578, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r i o r}} = \\sum_ {i = 1} ^ {M} \\lambda_ {s m} \\mathcal {L} _ {s m} ^ {(i)} + \\lambda_ {\\text {c u r v}} \\mathcal {L} _ {\\text {c u r v}} ^ {(i)} + \\sum_ {j = 1} ^ {N} \\lambda_ {\\text {p a r}} \\mathcal {L} _ {\\text {p a r}} ^ {(i j)}, \\tag {10}", + "image_path": "4604310a805f3c8168937b909f7ab4cc83acbb492c797b09f4690cd31ad49add.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": "with individual weights " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{par}" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{sm}" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{curv}" + }, + { + "bbox": [ + 46, + 629, + 287, + 715 + ], + "type": "text", + "content": ". Note that all these properties are expressible by means of tangents and normals, which can be computed analytically on our parametric representation in continuous space. Consequently, minimization of the herein introduced prior losses does not depend on numerical approximations as is the case for anchor-, grid- or key-point representations." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 541, + 242 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 541, + 242 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 541, + 242 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 541, + 242 + ], + "type": "image", + "image_path": "d1de4bcab885767d964f33dfd0955e89c71c85842a7b3cf317c0e6e7a880bf5a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 249, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 304, + 249, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 547, + 305 + ], + "type": "text", + "content": "Figure 4. Our proposed spatial transformation module. First, several road surface hypotheses are defined (a) to which front-view features are lifted (b) and weighted according to the predicted depth distribution. Afterwards, point features are aggregated in a weighted manner to obtain the 3D feature map (c)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 330, + 435, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 435, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 435, + 342 + ], + "type": "text", + "content": "3.3. Spatial transformation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 350, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 482 + ], + "type": "text", + "content": "In this section, we describe our spatial transformation (shown in Fig. 4) that is leveraging valuable physical knowledge about surface geometry. We know that the road surface typically shows small deviations from the ground level " + }, + { + "bbox": [ + 304, + 350, + 545, + 482 + ], + "type": "inline_equation", + "content": "(z = 0)" + }, + { + "bbox": [ + 304, + 350, + 545, + 482 + ], + "type": "text", + "content": " in the near-range and stronger deviations in the far-range. Based on this knowledge, we sample ground surface hypotheses that reflect the distribution of the road surface height profile (Fig. 4a). While in theory different types of surface functions could be utilized as hypotheses, we decide to merely rely on planes, since this facilitates the computation of ray intersections described in the following step." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": "Next, the multi-scale front-view feature maps extracted by the backbone are lifted to 3D space (Fig. 4b). Our approach is inspired by [31], where front-view features are spreading along rays throughout the space of the road surface. These rays intersect with the surface hypotheses at different depths spanning a frustum-like point cloud in 3D space, where each point is affiliated with a " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": "-dimensional feature vector and additionally attached with its height value " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": ", hence, each point in the cloud has dimension " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "(C + 1)" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": ". The front-view feature map is propagated through a depth branch with a channel-wise softmax applied to obtain a categorical distribution for each ray, resulting in a tensor of size " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "H\\times W\\times S" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": " denote height and width and channel size " + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": " the number of surface hypotheses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "In order to aggregate the information in 3D space, a BEV grid of size " + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "X \\times Y" + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": " is defined. Features from points mapping to the same grid cell are weighted by the categorical depth distribution for the respective ray and accumulated in terms of a weighted sum (Fig. 4c). Since the " + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "-component" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10643" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 287, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 287, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 287, + 133 + ], + "type": "table", + "html": "
PriorsF1(%)↑X-near(m)↓X-far(m)↓Z-near(m)↓Z-far(m)↓
None65.00.3160.3840.1060.153
Par.66.20.2910.3730.1030.150
Surf.65.80.3200.3560.1030.144
Curv.66.70.3220.3660.1050.146
Comb.66.70.3010.3590.1030.144
", + "image_path": "94752f31db4f17f6b15cdfcd30c94b725916952e8dd449c2b1457992486f2bbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 163, + 286, + 194 + ], + "blocks": [ + { + "bbox": [ + 63, + 140, + 271, + 152 + ], + "lines": [ + { + "bbox": [ + 63, + 140, + 271, + 152 + ], + "spans": [ + { + "bbox": [ + 63, + 140, + 271, + 152 + ], + "type": "text", + "content": "Table 1. Effect of different prior losses on OpenLane300." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 163, + 286, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 163, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 163, + 286, + 194 + ], + "type": "table", + "html": "
# Surface Hypotheses1351527
F1-Score(%)↑65.065.966.666.166.0
", + "image_path": "27bf1a0ffaabed838f141746171d0d04cab441ca857bc9bffbeb6cdf47866f9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 225, + 286, + 299 + ], + "blocks": [ + { + "bbox": [ + 59, + 201, + 274, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 201, + 274, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 201, + 274, + 213 + ], + "type": "text", + "content": "Table 2. Effect of the surface hypotheses on OpenLane300." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 225, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 50, + 225, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 50, + 225, + 286, + 299 + ], + "type": "table", + "html": "
Lane Rep.Prior Reg.Spatial T.F1(%)↑Gain(%)
62.9(baseline)
✓65.0+2.1
✓✓66.7+3.8
✓✓66.6+3.7
✓✓✓66.9+4.0
", + "image_path": "7cd9b6523d7260208023e39deca678805e87474317985f9e3e7b7a643c4a039d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 340 + ], + "type": "text", + "content": "Table 3. Performance gain for different contributions on OpenLane300 using our novel Lane Representation, Prior Regularization and Spatial Transformation instead of IPM." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "content": "of the points is also combined by a weighted sum, the value " + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "inline_equation", + "content": "z_{uv}" + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "content": " can be interpreted as the height value of the surface for grid cell " + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 363, + 287, + 424 + ], + "type": "text", + "content": ". We guide the model in learning the real surface and prevent it from learning an arbitrary mapping by introducing a simple grid-based regression loss as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 432, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 432, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 60, + 432, + 287, + 462 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s u r f}} = \\frac {1}{X \\cdot Y} \\sum_ {(u, v) \\in X \\times Y} \\mathbb {1} _ {u v} \\cdot \\| z _ {u v} - \\hat {z} _ {u v} \\| _ {1}, \\tag {11}", + "image_path": "ae795d2c797b7351918e7fcd0176fccad9e73bb14eed8c75859af5a993aadd80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{uv}" + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": " indicating whether surface ground truth " + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "inline_equation", + "content": "\\hat{z}_{uv}" + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": " is available for cell " + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 46, + 472, + 287, + 509 + ], + "type": "text", + "content": ". The height ground truth is obtained by interpolation of the 3D lane annotations at cell locations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 517, + 138, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 138, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 138, + 529 + ], + "type": "text", + "content": "3.4. Loss functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 536, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 561 + ], + "type": "text", + "content": "The overall loss used during training is given as the weighted sum of loss components" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 570, + 287, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 287, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {p r} \\mathcal {L} _ {p r} + \\lambda_ {c a t} \\mathcal {L} _ {c a t} + \\lambda_ {r e g} \\mathcal {L} _ {r e g} + \\tag {12}", + "image_path": "c86e5aa539555aa1c92d7fc01fd7b511c95533c94f19ac4e0a440f4f5ae4618b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 586, + 287, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 586, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 85, + 586, + 287, + 599 + ], + "type": "interline_equation", + "content": "\\lambda_ {v i s} \\mathcal {L} _ {v i s} + \\lambda_ {p r i o r} \\mathcal {L} _ {p r i o r} + \\lambda_ {s u r f} \\mathcal {L} _ {s u r f}. \\tag {13}", + "image_path": "54d36515bc3bee6d90f78801cbbcbc6cf4f4b693e4854f9a44279e6ecf212165.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": "We use focal loss [21] for lane presence " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pr}" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": " and category classification " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cat}" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": ". For the regression loss " + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 46, + 609, + 287, + 657 + ], + "type": "text", + "content": ", we adapt the formulation of [32] to three instead of two dimensions. More details are provided in the supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We first describe our experimental setup and then analyze our approach on two 3D lane datasets." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 420, + 134 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 420, + 134 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 420, + 134 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 420, + 134 + ], + "type": "image", + "image_path": "d8f50475f56b7aba490beb4d19e4267cd4ec7c906cfdcffb6c0cc88bc6a4a7a5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 133, + 384, + 138 + ], + "lines": [ + { + "bbox": [ + 357, + 133, + 384, + 138 + ], + "spans": [ + { + "bbox": [ + 357, + 133, + 384, + 138 + ], + "type": "text", + "content": "Front-view" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 321, + 139, + 421, + 220 + ], + "blocks": [ + { + "bbox": [ + 321, + 139, + 421, + 220 + ], + "lines": [ + { + "bbox": [ + 321, + 139, + 421, + 220 + ], + "spans": [ + { + "bbox": [ + 321, + 139, + 421, + 220 + ], + "type": "image", + "image_path": "4ef57ac9d6f77182779fcca8b7d4ad9ac45485f1d14a7e2bdc9683598ac864d3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 533, + 134 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 533, + 134 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 533, + 134 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 533, + 134 + ], + "type": "image", + "image_path": "b1c6df6aaf53bfd0000d1ffa8f888fe304f8da0a3b5ecdab4015a24ae931f747.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 133, + 496, + 139 + ], + "lines": [ + { + "bbox": [ + 469, + 133, + 496, + 139 + ], + "spans": [ + { + "bbox": [ + 469, + 133, + 496, + 139 + ], + "type": "text", + "content": "Front-view" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 432, + 140, + 533, + 221 + ], + "blocks": [ + { + "bbox": [ + 432, + 140, + 533, + 221 + ], + "lines": [ + { + "bbox": [ + 432, + 140, + 533, + 221 + ], + "spans": [ + { + "bbox": [ + 432, + 140, + 533, + 221 + ], + "type": "image", + "image_path": "cea5e6527d9d5251b5eb2a2e3f1699a0379f260c07652769b87f01b4b783d563.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 320, + 222, + 422, + 262 + ], + "blocks": [ + { + "bbox": [ + 320, + 222, + 422, + 262 + ], + "lines": [ + { + "bbox": [ + 320, + 222, + 422, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 222, + 422, + 262 + ], + "type": "image", + "image_path": "2139bb034d4a0cc9d059fbf2d02f4db9fbf7aae7d5502c6c9657ecdc18d131da.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 266, + 380, + 275 + ], + "lines": [ + { + "bbox": [ + 369, + 266, + 380, + 275 + ], + "spans": [ + { + "bbox": [ + 369, + 266, + 380, + 275 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 285, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 329 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of our model trained with prior regularization to the same model without regularization both trained on OpenLane300 with main differences highlighted by arrows. As a reference ground truth lines are visualized dashed." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 430, + 222, + 533, + 262 + ], + "blocks": [ + { + "bbox": [ + 430, + 222, + 533, + 262 + ], + "lines": [ + { + "bbox": [ + 430, + 222, + 533, + 262 + ], + "spans": [ + { + "bbox": [ + 430, + 222, + 533, + 262 + ], + "type": "image", + "image_path": "8a4b18d813c32b0f089a8931428c4b75653410ced963493e011883d811c1096a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 266, + 490, + 275 + ], + "lines": [ + { + "bbox": [ + 480, + 266, + 490, + 275 + ], + "spans": [ + { + "bbox": [ + 480, + 266, + 490, + 275 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 353, + 421, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 421, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 421, + 366 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 373, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 409 + ], + "type": "text", + "content": "We evaluate our method on two different datasets: OpenLane and Apollo 3D Synthetic - both containing 3D lane ground truth as well as camera parameters per frame." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 410, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 410, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 410, + 545, + 494 + ], + "type": "text", + "content": "OpenLane [2] is a real-world dataset containing 150,000 images in the training and 40,000 in the test set from 1000 different sequences. In order to evaluate different driving scenarios the test set is divided into different situations, namely Up & Down, Curve, Extreme Weather, Night, Intersection and Merge & Split. For ablation studies we use the smaller version OpenLane300 including 300 sequences." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 555 + ], + "type": "text", + "content": "Apollo 3D Synthetic [9] is a small synthetic dataset, consisting of only 10,500 examples from rather simple scenarios of highway, urban and rural environments. The data is split into three subsets, (1) Standard (simple) scenarios, (2) Rare Scenes and (3) Visual Variations." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "Evaluation metrics. For the quantitative evaluation both datasets utilize the evaluation scheme proposed in [9]. It evaluates the euclidean distance at uniformly distributed points in the range of " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "0 - 100\\mathrm{m}" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": " direction. Based on the mean distance and range, " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "Fl" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "-Score is computed, as well as the mean " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "-errors in near- " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "(0 - 40\\mathrm{m})" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": " and far-range " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "(40 - 100\\mathrm{m})" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": " to evaluate geometric accuracy." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Baseline. Our approach builds up on 3D-SpLineNet. Since it was applied on synthetic data only, it showed poor performance on real data. We applied some straightforward design adaptations - e.g. larger backbone, multiscale features (see supplementary) - and use this modified 3D-SpLineNet as our baseline (first row Table 3)." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10644" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 194 + ], + "type": "table", + "html": "
MethodF1-Score(%)↑X-error near(m)↓X-error far(m)↓Z-error near(m)↓Z-error far(m)↓U&DF1-Score(%) per Scenario ↑
CEWNIM&S
3D-LaneNet [6]44.10.4790.5720.3670.44340.846.547.541.532.141.7
Gen-LaneNet [9]32.30.5910.6840.4110.52125.433.528.118.721.431.0
PersFormer [2]50.50.4850.5530.3640.43142.455.648.646.640.050.7
PersFormer* [2]53.10.3610.3280.1240.12946.858.754.048.441.452.5
CurveFormer [1]50.50.3400.7720.2070.65145.256.649.749.142.945.4
BEV-LaneDet [46]58.40.3090.6590.2440.63148.763.153.453.450.353.7
Anchor3DLane [12]53.70.2760.3110.1070.13846.757.252.547.845.451.2
Anchor3DLane-T [12]54.30.2750.3100.1050.13547.258.052.748.745.851.7
LaneCPP (Ours)60.30.2640.3100.0770.11753.664.456.754.952.058.7
", + "image_path": "8e93d21a97eaa0753b2d0fa89ac87470188d08fbe57f72f2bb10e820facc0e92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 202, + 546, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 546, + 237 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on OpenLane [2]. Best performance and second best are highlighted. The scenario categories are Up and Down (U&D), Curve (C), Extreme Weather (EW), Night (N), Intersection (I), Merge and Split (M&S). PersFormer* denotes the latest performance reported on the official code base, Anchor3DLane-T represents the temporal multi-frame method of [12]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 247, + 144, + 300 + ], + "blocks": [ + { + "bbox": [ + 50, + 247, + 144, + 300 + ], + "lines": [ + { + "bbox": [ + 50, + 247, + 144, + 300 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 144, + 300 + ], + "type": "image", + "image_path": "448e72ae7ae1e35ac542cb83f5290e47082516b13502ec82a11e9167a004a195.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 300, + 143, + 377 + ], + "blocks": [ + { + "bbox": [ + 53, + 300, + 143, + 377 + ], + "lines": [ + { + "bbox": [ + 53, + 300, + 143, + 377 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 143, + 377 + ], + "type": "image", + "image_path": "b1025d877499044a6cc701c3d2d84f82a0c0639438fb8d45a60dde87feff7ba5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 150, + 247, + 241, + 301 + ], + "blocks": [ + { + "bbox": [ + 150, + 247, + 241, + 301 + ], + "lines": [ + { + "bbox": [ + 150, + 247, + 241, + 301 + ], + "spans": [ + { + "bbox": [ + 150, + 247, + 241, + 301 + ], + "type": "image", + "image_path": "c5cead738cf41e319a1183663c7702360da8a0769fc98d72bc06f6a98060f51c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 151, + 303, + 242, + 377 + ], + "blocks": [ + { + "bbox": [ + 151, + 303, + 242, + 377 + ], + "lines": [ + { + "bbox": [ + 151, + 303, + 242, + 377 + ], + "spans": [ + { + "bbox": [ + 151, + 303, + 242, + 377 + ], + "type": "image", + "image_path": "c005d70086866981a42937a877a761e0a9f9d9e9d4bf15da9d840ce28b4d4bbd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 250, + 247, + 341, + 300 + ], + "blocks": [ + { + "bbox": [ + 250, + 247, + 341, + 300 + ], + "lines": [ + { + "bbox": [ + 250, + 247, + 341, + 300 + ], + "spans": [ + { + "bbox": [ + 250, + 247, + 341, + 300 + ], + "type": "image", + "image_path": "448dd77f44973c75ff32a31560e0c84bacf7755cb35e8c408d8da6b6f8dd5a4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 301, + 342, + 378 + ], + "blocks": [ + { + "bbox": [ + 251, + 301, + 342, + 378 + ], + "lines": [ + { + "bbox": [ + 251, + 301, + 342, + 378 + ], + "spans": [ + { + "bbox": [ + 251, + 301, + 342, + 378 + ], + "type": "image", + "image_path": "8fe17f78161de68b46601a22002c94452a439b37f0c1816b160f56bc0aefb07e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 350, + 247, + 441, + 299 + ], + "blocks": [ + { + "bbox": [ + 350, + 247, + 441, + 299 + ], + "lines": [ + { + "bbox": [ + 350, + 247, + 441, + 299 + ], + "spans": [ + { + "bbox": [ + 350, + 247, + 441, + 299 + ], + "type": "image", + "image_path": "3671b8e28ca4ab8accad2368143da3e9f789f3d40145ed2fe46ba41a2fac6ecf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 351, + 301, + 443, + 378 + ], + "blocks": [ + { + "bbox": [ + 351, + 301, + 443, + 378 + ], + "lines": [ + { + "bbox": [ + 351, + 301, + 443, + 378 + ], + "spans": [ + { + "bbox": [ + 351, + 301, + 443, + 378 + ], + "type": "image", + "image_path": "2e09f8a23b45c6c49babfd81ae9f26a97c28460450f73e1400ce75e9bafa98d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 451, + 247, + 542, + 300 + ], + "blocks": [ + { + "bbox": [ + 451, + 247, + 542, + 300 + ], + "lines": [ + { + "bbox": [ + 451, + 247, + 542, + 300 + ], + "spans": [ + { + "bbox": [ + 451, + 247, + 542, + 300 + ], + "type": "image", + "image_path": "96d046c2e04dba0541828eec494db8b826bcfb1223f1acdc07c63f6ccaaa3bd7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 452, + 301, + 543, + 378 + ], + "blocks": [ + { + "bbox": [ + 452, + 301, + 543, + 378 + ], + "lines": [ + { + "bbox": [ + 452, + 301, + 543, + 378 + ], + "spans": [ + { + "bbox": [ + 452, + 301, + 543, + 378 + ], + "type": "image", + "image_path": "ec75a55116b88d66845331c61fccb2a12d696d00a92d2dcf24ae1c56b35013b1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 380, + 146, + 416 + ], + "blocks": [ + { + "bbox": [ + 50, + 380, + 146, + 416 + ], + "lines": [ + { + "bbox": [ + 50, + 380, + 146, + 416 + ], + "spans": [ + { + "bbox": [ + 50, + 380, + 146, + 416 + ], + "type": "image", + "image_path": "3444e3b48b655990b1f0b25d82dccb6d9bdbf7f92086c3c390b68c09491cc803.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 419, + 103, + 428 + ], + "lines": [ + { + "bbox": [ + 93, + 419, + 103, + 428 + ], + "spans": [ + { + "bbox": [ + 93, + 419, + 103, + 428 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 148, + 381, + 244, + 417 + ], + "blocks": [ + { + "bbox": [ + 148, + 381, + 244, + 417 + ], + "lines": [ + { + "bbox": [ + 148, + 381, + 244, + 417 + ], + "spans": [ + { + "bbox": [ + 148, + 381, + 244, + 417 + ], + "type": "image", + "image_path": "7fadd5d84a5837a9291f902c12685dd462fd31009a7857bd92afd2defd26bdb7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 420, + 200, + 429 + ], + "lines": [ + { + "bbox": [ + 190, + 420, + 200, + 429 + ], + "spans": [ + { + "bbox": [ + 190, + 420, + 200, + 429 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 248, + 380, + 343, + 417 + ], + "blocks": [ + { + "bbox": [ + 248, + 380, + 343, + 417 + ], + "lines": [ + { + "bbox": [ + 248, + 380, + 343, + 417 + ], + "spans": [ + { + "bbox": [ + 248, + 380, + 343, + 417 + ], + "type": "image", + "image_path": "c56e09f51d0bd0fbacc6de1527732c23acc442a6b853e08a026b3ccbce670bff.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 421, + 300, + 430 + ], + "lines": [ + { + "bbox": [ + 290, + 421, + 300, + 430 + ], + "spans": [ + { + "bbox": [ + 290, + 421, + 300, + 430 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 440, + 539, + 452 + ], + "lines": [ + { + "bbox": [ + 52, + 440, + 539, + 452 + ], + "spans": [ + { + "bbox": [ + 52, + 440, + 539, + 452 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on OpenLane. Our method is compared to PersFormer* with ground truth visualized as dashed lines." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 347, + 381, + 444, + 417 + ], + "blocks": [ + { + "bbox": [ + 347, + 381, + 444, + 417 + ], + "lines": [ + { + "bbox": [ + 347, + 381, + 444, + 417 + ], + "spans": [ + { + "bbox": [ + 347, + 381, + 444, + 417 + ], + "type": "image", + "image_path": "2ae2c4c7081a6fb245227744324630c8fa59e554a633a2aa96ba4a3df74e7bd6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 420, + 400, + 430 + ], + "lines": [ + { + "bbox": [ + 391, + 420, + 400, + 430 + ], + "spans": [ + { + "bbox": [ + 391, + 420, + 400, + 430 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 447, + 381, + 544, + 417 + ], + "blocks": [ + { + "bbox": [ + 447, + 381, + 544, + 417 + ], + "lines": [ + { + "bbox": [ + 447, + 381, + 544, + 417 + ], + "spans": [ + { + "bbox": [ + 447, + 381, + 544, + 417 + ], + "type": "image", + "image_path": "d0125e786f8849c86237e2b209555ad669774de6694d2f9697a7804128f96aa5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 421, + 501, + 430 + ], + "lines": [ + { + "bbox": [ + 491, + 421, + 501, + 430 + ], + "spans": [ + { + "bbox": [ + 491, + 421, + 501, + 430 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": "Implementation details. We use input size " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "360 \\times 480" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " and adopt the same backbone as in [2] based on a modified EfficientNet [42]. We extract four feature maps of resolutions " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\left[\\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}, \\frac{1}{16}\\right]" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": ". The final 3D feature map has size " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "26 \\times 16" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " with 64 channels. We use " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "M = 64" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " initial line proposals and B-Splines of degree " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "d = 3" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " control points. We apply Adam optimizer [14] with an initial learning rate of " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " for OpenLane and " + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 472, + 287, + 604 + ], + "type": "text", + "content": " for Apollo and a dataset specific step-wise scheduler. We train for 30 epochs on OpenLane and 300 epochs on Apollo with batch size 16. For more details we refer to the supplementary." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 611, + 145, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 145, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 145, + 623 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Table 1 indicates the effect of our proposed prior-based regularization. It is evident that each prior improves the F1-Score as well as geometric errors. While the surface and curvature priors result in better far-range estimates, line parallelism supports X-regression in the near-range. Besides, using surface smoothness loss results in lowest Z-far errors. Finally, a combination of priors yields a good balance of" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 592 + ], + "type": "text", + "content": "F1-Score and geometric errors. The positive effect of parallelism is confirmed by Fig. 5, where reinforcing parallel lane structure leads to better estimates in the near-range (a) and far-range (b) compared to the unregularized model. Learning parallel lines also is evidently beneficial in cases of poor visibility (b) and occlusions (a). In the latter case, the regularized model even shows better predictions than the noisy ground truth. This emphasizes the high relevance of priors for more robust behavior for real-world datasets, where 3D ground truth often comes with inaccuracies." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "For the spatial transformation (see Table 2), too low numbers of surface hypotheses result in worse score, presumably as 3D geometry is not captured sufficiently, whereas larger numbers tend to decreasing performance due to the higher complexity. The best F1-Score is obtained with 5 hypotheses, which is chosen for further experiments. While the improvement over IPM is already considerable, we think that with the simplifications of plane hypotheses prevent the component from developing its full potential. We see ways to enhance the 3D transformation even further" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10645" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 215 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 215 + ], + "type": "table", + "html": "
MethodBalanced ScenesRare Scenes
F1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfarF1(%)↑X-error (m) ↓nearZ-error (m) ↓farnearfar
3D-LaneNet [6]86.40.0680.4770.0150.20272.00.1660.8550.0390.521
GP [17]91.90.0490.3870.0080.21383.70.1260.9030.0230.625
PersFormer [2]92.90.0540.3560.010.23487.50.1070.7820.0240.602
3D-SpLineNet [32]96.30.0370.3240.0090.21392.90.0770.6990.0210.562
CurveFormer [1]95.80.0780.3260.0180.21995.60.1820.7370.0390.561
BEV-LaneDet [46]96.90.0160.2420.020.21697.60.0310.5940.0400.556
Anchor3DLane [12]95.40.0450.3000.0160.22394.40.0820.6990.0300.580
LaneCPP (Ours)97.40.0300.2770.0110.20696.20.0730.6510.0230.543
", + "image_path": "b158c4f590d85dde71ffbeda325a5cda2faee6f167cff4716673f55fd79306a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 221, + 533, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 221, + 533, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 221, + 533, + 234 + ], + "type": "text", + "content": "Table 5. Quantitative comparison of best methods on Apollo 3D Synthetic [9]. Best performance and second best are highlighted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 255, + 279, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 279, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 279, + 266 + ], + "type": "text", + "content": "using more sophisticated spatial representations in future." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 375 + ], + "type": "text", + "content": "The impact of our different contributions is summarized in Table 3, where the first row shows our baseline (see Sec. 4.1). More than two percent in F1-Score are gained with our novel lane representation compared to the simplified one from [32]. Moreover, it is clear that both, the regularization using combined priors and the spatial transformation using 5 hypotheses result in significant improvement. Eventually, combining all components yields the best model configuration, which we choose for further evaluation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 188, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 188, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 188, + 396 + ], + "type": "text", + "content": "4.3. Evaluation on OpenLane" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": "On the real-world OpenLane benchmark our model evidently outperforms all other methods with respect to F1-Score as well as geometric errors as shown in Table 4. Compared to BEV-Lanedet, which achieves a high detection score, our model gains " + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "inline_equation", + "content": "+1.9\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": ", while reaching significantly lower geometric errors. In comparison to Anchor3DLane the improvements with respect to X-errors are less substantial, however, our approach surpasses the F1-Score by a large gap of " + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "inline_equation", + "content": "+6.6\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": ". Analyzing the detection scores among different scenarios, outstanding performance gain is observed on the up- and down-hill test set " + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "inline_equation", + "content": "(+5.9\\%)" + }, + { + "bbox": [ + 46, + 402, + 287, + 558 + ], + "type": "text", + "content": " that highlights the capability of our approach to capture 3D space proficiently, which is supported by the low Z-errors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Apart from quantitative results, we show qualitative examples in Fig. 6. In up-hill scenarios like Fig. 6b our model manages to estimate both lateral and height profile accurately, since our assumptions about road surface and line parallelism are satisfied. In contrast, PersFormer lacks spatial features and does not use any kind of physical regularization. Consequently, it fails to estimate the 3D lane geometry and even collapses in Fig. 6c, whereas our surface and curvature priors always prevent such a behavior. Noteworthy is also the top performance on the merges and splits set. This proves that our soft regularization is even capable to handle situations containing non-parallel lines, which is also confirmed by Fig. 6d. However, we rarely observe limi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 255, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 327 + ], + "type": "text", + "content": "tations with our formulation for line pairs with a similar orientation but weakly converging course as shown in Fig. 6e. In such cases the indicator function might erroneously decide for parallelism loss during training. One possible solution for future work would be to consider ground truth for the indicator function to identify such situations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 336, + 491, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 336, + 491, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 336, + 491, + 350 + ], + "type": "text", + "content": "4.4. Evaluation on Apollo 3D Synthetic" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 355, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 464 + ], + "type": "text", + "content": "The Apollo 3D Synthetic dataset is very limited in size and only consists of simple situations in contrast to OpenLane. While we find the results on OpenLane more meaningful, we would like to still provide and discuss the quantitative results on the Apollo dataset. Due to the simplicity of the dataset, our model cannot benefit that significantly from our priors but still achieves competitive results to state of the art with the highest F1-Score on the balanced scenes dataset and comparable error metrics (second best for most errors)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 477, + 470, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 477, + 470, + 490 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 470, + 490 + ], + "type": "text", + "content": "5. Conclusions and future work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 715 + ], + "type": "text", + "content": "In this work, we present LaneCPP, a novel approach for 3D lane detection that leverages physical prior knowledge about lane structure and road geometry. Our new continuous lane representation overcomes previous deficiencies by allowing arbitrary lane structures and enables us to regularize lane geometry based on analytically formulated priors. We further introduce a novel spatial transformation module that models 3D features carefully considering knowledge about road surface geometry. In our experiments, we demonstrate state-of-the-art performance on real and synthetic benchmarks. The full capability of our approach is revealed on real-world OpenLane, for which we prove the relevance of priors quantitatively and qualitatively. In future, priors could be individualized for different driving scenarios and might support to learn inter-lane relations to achieve better scene understanding in a global context. We also see ways to leverage the full potential of the spatial transformation by using more sophisticated surface representations." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10646" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Yifeng Bai, Zhirong Chen, Zhangjie Fu, Lang Peng, Pengpeng Liang, and Erkang Cheng. Curveformer: 3d lane detection by curve propagation with curve queries and attention. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Li Chen, Chonghao Sima, Yang Li, Zehan Zheng, Jiajie Xu, Xiangwei Geng, Hongyang Li, Conghui He, Jianping Shi, Yu Qiao, et al. Persformer: 3d lane detection via perspective transformer and the openlane benchmark. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "type": "text", + "content": "[3] Carl de Boor. On calculating with b-splines. Journal of Approximation Theory, 1972. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 280 + ], + "type": "text", + "content": "[4] Netalee Efrat, Max Bluvstein, Shaul Oron, Dan Levi, Noa Garnett, and Bat El Shlomo. 3d-lanenet+: Anchor free lane detection using a semi-local representation. arXiv/2011.01535, 2020. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 324 + ], + "type": "text", + "content": "[5] Zhengyang Feng, Shaohua Guo, Xin Tan, Ke Xu, Min Wang, and Lizhuang Ma. Rethinking efficient lane detection via curve modeling. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 369 + ], + "type": "text", + "content": "[6] Noa Garnett, Rafi Cohen, Tomer Pe'er, Roee Lahav, and Dan Levi. 3d-lanenet: End-to-end 3d multiple lane detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 423 + ], + "type": "text", + "content": "[7] Mohsen Ghafoorian, Cedric Nugteren, Nóra Baka, Olaf Booij, and Michael Hofmann. EL-GAN: embedding loss driven generative adversarial networks for lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2018. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 458 + ], + "type": "text", + "content": "[8] Stephen Gould, Richard Hartley, and Dylan Campbell. Deep declarative networks. IEEE Trans. on Pattern Analysis and Machine Intelligence (PAMI), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 459, + 287, + 513 + ], + "type": "text", + "content": "[9] Yuliang Guo, Guang Chen, Peitao Zhao, Weide Zhang, Jinghao Miao, Jingao Wang, and Tae Eun Choe. Gen-lanenet: A generalized and scalable approach for 3d lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 547 + ], + "type": "text", + "content": "[10] Steffen Hagedorn, Marcel Milich, and Alexandru P. Condurache. Pioneering se (2)-equivariant trajectory planning for automated driving. arXiv:2403.11304, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[11] Yuenan Hou, Zheng Ma, Chunxiao Liu, and Chen Change Loy. Learning lightweight lane detection cnns by self attention distillation. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 647 + ], + "type": "text", + "content": "[12] Shaofei Huang, Zhenwei Shen, Zehao Huang, Zi han Ding, Jiao Dai, Jizhong Han, Naiyan Wang, and Si Liu. Anchor3dlane: Learning to regress 3d anchors for monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[13] Brody Huval, Tao Wang, Sameep Tandon, Jeff Kiske, Will Song, Joel Pazhayampallil, Mykhaylo Andriluka, Pranav Rajpurkar, Toki Migimatsu, Royce Cheng-Yue, Fernando A. Mujica, Adam Coates, and Andrew Y. Ng. An empirical evaluation of deep learning on highway driving. arXiv/1504.01716, 2015. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of the International Conf. on Learning Representations (ICLR), 2015. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[15] YeongMin Ko, Jiwon Jun, Donghwuy Ko, and Moongu Jeon. Key points estimation and point instance segmentation approach for lane detection. arXiv/2002.06604, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 141, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 207 + ], + "type": "text", + "content": "[16] Seokju Lee, Junsik Kim, Jae Shin Yoon, Seunghak Shin, Oleksandr Bailo, Namil Kim, Tae-Hee Lee, Hyun Seok Hong, Seung-Hoon Han, and In So Kweon. Vpgnet: Vanishing point guided network for lane and road marking detection and recognition. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 262 + ], + "type": "text", + "content": "[17] Chenguang Li, Jia Shi, Ya Wang, and Guangliang Cheng. Reconstruct from top view: A 3d lane detection approach based on geometry structure prior. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 308 + ], + "type": "text", + "content": "[18] Qi Li, Yue Wang, Yilun Wang, and Hang Zhao. Hdmapnet: An online HD map construction and evaluation framework. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 545, + 353 + ], + "type": "text", + "content": "[19] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Trans. on Intelligent Transportation Systems (T-ITS), 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 409 + ], + "type": "text", + "content": "[20] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In Proc. of the European Conf. on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 410, + 545, + 454 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proc. of the IEEE International Conf. on Computer Vision (ICCV), 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 498 + ], + "type": "text", + "content": "[22] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 544 + ], + "type": "text", + "content": "[23] Ruijin Liu, Dapeng Chen, Tie Liu, Zhiliang Xiong, and Zejian Yuan. Learning to predict 3d lane shape and camera pose from a single image via geometry constraints. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2022. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[24] Pingping Lu, Chen Cui, Shaobing Xu, Huei Peng, and Fan Wang. SUPER: A novel lane detection system. IEEE Trans. on Intelligent Vehicles (T-IV), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 622 + ], + "type": "text", + "content": "[25] Hanspeter Mallot, Heinrich Blthoff, J.J. Little, and S Bohrer. Inverse perspective mapping simplifies optical flow computation and obstacle detection. Biological Cybernetics, 1991. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[26] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In Proc. IEEE Intelligent Vehicles Symposium (IV), 2018. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Marcos Nieto, Luis Salgado, Fernando Jaureguizar, and Jon Arróspide. Robust multiple lane road modeling based on perspective analysis. In Proc. IEEE International Conf. on Image Processing (ICIP), 2008. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10647" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[28] Bowen Pan, Jiankai Sun, Ho Yin Tiga Leung, Alex Andonian, and Bolei Zhou. Cross-view semantic segmentation for sensing surroundings. IEEE Robotics Autom. Lett., 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[29] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial CNN for traffic scene understanding. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2018. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 215 + ], + "type": "text", + "content": "[30] Lang Peng, Zhirong Chen, Zhangjie Fu, Pengpeng Liang, and Erkang Cheng. Bevsegformer: Bird's eye view semantic segmentation from arbitrary camera rigs. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[31] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 315 + ], + "type": "text", + "content": "[32] Maximilian Pittner, Alexandru Condurache, and Joel Janai. 3d-splinenet: 3d traffic line detection using parametric spline representations. In Proc. of the IEEE Winter Conference on Applications of Computer Vision (WACV), 2023. 1, 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 360 + ], + "type": "text", + "content": "[33] Fabio Pizzati, Marco Allodi, Alejandro Barrera, and Fernando García. Lane detection and classification using cascaded cnns. In Proc. of the International Conf. on Computer Aided Systems Theory (EUROCAST), 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "text", + "content": "[34] Zhan Qu, Huan Jin, Yang Zhou, Zhen Yang, and Wei Zhang. Focus on local: Detecting lane marker from bottom up via key point. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 458 + ], + "type": "text", + "content": "[35] Matthias Rath and Alexandru Paul Condurache. Invariant integration in deep convolutional feature space. In Proc. of European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 514 + ], + "type": "text", + "content": "[36] Matthias Rath and Alexandru Paul Condurache. Improving the sample-complexity of deep classification networks with invariant integration. In Proc. of International Joint Conf. on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP), 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 558 + ], + "type": "text", + "content": "[37] Thomas Roddick and Roberto Cipolla. Predicting semantic map representations from images using pyramid occupancy networks. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 602 + ], + "type": "text", + "content": "[38] Thomas Roddick, Alex Kendall, and Roberto Cipolla. Orthographic feature transform for monocular 3d object detection. In Proc. of the British Machine Vision Conf. (BMVC), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 645 + ], + "type": "text", + "content": "[39] Avishkar Saha, Oscar Mendez, Chris Russell, and Richard Bowden. Translating images into maps. In Proc. IEEE International Conf. on Robotics and Automation (ICRA), 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[40] Jinming Su, Chao Chen, Ke Zhang, Junfeng Luo, Xiaoming Wei, and Xiaolin Wei. Structure guided lane detection. In Proc. of the International Joint Conf. on Artificial Intelligence (IJCAI), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 677 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Keep your eyes on the lane: Real-time attention-guided lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[42] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proc. of the International Conf. on Machine learning (ICML), 2019. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[43] Lucas Tabelini Torres, Rodrigo Ferreira Berriel, Thiago M. Paixão, Claudine Badue, Alberto F. De Souza, and Thiago Oliveira-Santos. Polyanenet: Lane estimation via deep polynomial regression. In Proc. of the International Conf. on Pattern Recognition (ICPR), 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[44] Bingke Wang, Zilei Wang, and Yixin Zhang. Polynomial regression network for variable-number lane detection. In Proc. of the European Conf. on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[45] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 352 + ], + "type": "text", + "content": "[46] Ruihao Wang, Jian Qin, Kaiying Li, Yaochen Li, Dong Cao, and Jintao Xu. Bev-lanedet: An efficient 3d lane detection based on virtual camera via key-points. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 353, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 386 + ], + "type": "text", + "content": "[47] Yuping Wang and Jier Chen. Eqdrive: Efficient equivariant motion forecasting with multi-modality for autonomous driving. arXiv:2310.17540, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 545, + 430 + ], + "type": "text", + "content": "[48] Lu Xiong, Zhenwen Deng, Peizhi Zhang, and Zhiqiang Fu. A 3d estimation of structural road surface based on lane-line information. IFAC Conf. on Engine and Powertrain Control, Simulation and Modeling (E-COSM), 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 487 + ], + "type": "text", + "content": "[49] Chenxin Xu, Robby T. Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 542 + ], + "type": "text", + "content": "[50] Fan Yan, Ming Nie, Xinyue Cai, Jianhua Han, Hang Xu, Zhen Yang, Chaoqiang Ye, Yanwei Fu, Michael Bi Mi, and Li Zhang. Once-3dlanes: Building monocular 3d lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 587 + ], + "type": "text", + "content": "[51] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. RESA: recurrent feature-shift aggregator for lane detection. In Proc. of the Conf. on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 632 + ], + "type": "text", + "content": "[52] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 633, + 545, + 677 + ], + "type": "text", + "content": "[53] Qin Zou, Hanwen Jiang, Qiyu Dai, Yuanhao Yue, Long Chen, and Qian Wang. Robust lane detection from continuous driving scenes using deep neural networks. IEEE Trans. on Vehicular Technology (VTC), 2020. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10648" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_content_list.json b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75a0eb32faec2db6e194889e7d5ac11b82a222c6 --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_content_list.json @@ -0,0 +1,1502 @@ +[ + { + "type": "text", + "text": "LangSplat: 3D Language Gaussian Splatting", + "text_level": 1, + "bbox": [ + 258, + 130, + 712, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minghan Qin $^{1,*}$ , Wanhua Li $^{2,*, \\boxtimes}$ , Jiawei Zhou $^{1,*}$ , Haoqian Wang $^{1, \\boxtimes}$ , Hanspeter Pfister $^{2}$ , Tsinghua University $^{2}$ Harvard University", + "bbox": [ + 104, + 186, + 864, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qmh21@mails.tsinghua.edu.cn, wanhua@seas.harvard.edu, zhoujw22@mails.tsinghua.edu.cn", + "bbox": [ + 112, + 228, + 854, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wanghaoqian@tsinghua.edu.cn, pfister@seas.harvard.edu", + "bbox": [ + 246, + 246, + 718, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21fa475a5295e806312cea614041d6ce79c4aa276fbce9c80a008a9c8c0dc2f4.jpg", + "image_caption": [ + "Figure 1. Visualization of learned 3D language features of the previous SOTA method LERF and our LangSplat. While LERF generates imprecise and vague 3D features, our LangSplat accurately captures object boundaries and provides precise 3D language fields. While being effective, our LangSplat is also $199 \\times$ faster than LERF at the resolution of $1440 \\times 1080$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 300, + 361, + 532 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eebeafe6fee027ca32321aee7ad3bf6256d000592a8899ead1a4bf28170911ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 301, + 624, + 529 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b42dcdfeb3e4b5023664ed51b30238f0a4b497487723d096fe8aeed6a18e0e09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 301, + 887, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 595, + 313, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans live in a 3D world and commonly use natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space,", + "bbox": [ + 73, + 628, + 470, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experimental results show that LangSplat significantly outperforms the previous state-of-the-art method LERF by a large margin. Notably, LangSplat is extremely efficient, achieving a $199 \\times$ speedup compared to LERF at the resolution of $1440 \\times 1080$ . We strongly recommend readers to check out our video results at https://langsplat.github.io/.", + "bbox": [ + 500, + 597, + 892, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 843, + 632, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language is the primary means of communication for human beings [3]. Modeling a 3D language field allows users", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. ${}^{ \\ominus }$ Corresponding authors.", + "bbox": [ + 99, + 885, + 349, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to interact with and query 3D worlds using open-ended language, which presents a promising avenue for human-computer interaction and understanding [1, 5, 13]. The field of open-ended language queries in 3D has attracted increasing attention due to its various applications such as robotic navigation [14] and manipulation [39], 3D semantic understanding [9, 54] and editing [20], autonomous driving [16], and augmented/virtual reality [24].", + "bbox": [ + 75, + 90, + 470, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the absence of large-scale and diverse 3D scene data with language annotations, the current prevailing approach like LERF [18] involves feature distillation from off-the-shelf vision-language models such as CLIP into a 3D scene. However, these methods [18, 24] suffer from significant limitations in both speed and accuracy, severely constraining their practical applicability. To address these two issues, we revisit two key aspects of 3D language field modeling: the 3D modeling approach that bridges the gap between 2D and 3D, and the rendering target, which determines what to learn for a 3D point. For the 3D modeling technology, most methods utilize neural radiance fields (NeRFs) to represent a 3D scene, where volume rendering techniques are employed to accumulate 3D points along a ray into a single pixel. While NeRF has been demonstrated for its powerful 3D representation capabilities [7, 32, 33, 37, 38, 52], the volume rendering nature leads to its computationally expensive rendering speed [8, 31, 41], which imposes notable constraints on the potential applications within the NeRF-based language field.", + "bbox": [ + 75, + 215, + 473, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the rendering target, learning a CLIP embedding for a 3D point could be ambiguous as CLIP embeddings are aligned with images rather than pixels. Employing CLIP embeddings from a cropped patch also raises the point ambiguity issue, as the same 3D position can be associated with semantic concepts of varying scales. For instance, a point located on a bear's nose should yield high response values for three distinct textual queries: \"bear's nose\", \"bear's head\", and \"bear\" given that this point contributes to all three hierarchical regions. To deal with this issue, current methods [18, 24] introduce an additional absolute scale input to NeRF, trains with patch-wise CLIP features at different scales, and densely renders 2D maps at multiple scales during querying to select the optimal one. However, this scale-based solution compromises both efficiency and effectiveness. It could increase query time by up to 30 times as it needs to render at multiple different scales. Moreover, most patches with varying scales often fail to accurately encompass objects, either frequently including other objects from the background or omitting portions of the target object. These inaccurate CLIP features lead to the trained 3D language field lacking clear boundaries and containing a significant amount of noise. Therefore, they often simultaneously learn pixel-aligned DINO features to mitigate this issue. However, the performance remains un", + "bbox": [ + 75, + 523, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "satisfactory. As shown in Figure 1, LERF still generates imprecise 3D language features.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the 3D Language Gaussian Splatting (LangSplat) to address the above issues. Instead of using NeRF to build 3D representations, we resort to 3D Gaussian Splatting, which represents a 3D scene as a collection of 3D Gaussians and uses tile-based splatting to achieve efficient rendering at high resolutions. Our LangSplat defines a set of 3D language Gaussians, with each Gaussian being enhanced by a language embedding. These language-enhanced Gaussians are supervised using CLIP embeddings extracted from image patches captured from multiple training views, ensuring multi-view consistency. As an explicit modeling method, directly storing the high-dimensional language embeddings for each 3D language Gaussian is memory-inefficient. To reduce the memory cost and further improve the rendering efficiency, we propose to first learn a scene-wise language autoencoder, which maps CLIP embeddings in a scene to a low-dimensional latent space. In this way, each language Gaussian only contains the low-dimensional latent language features and the final language embeddings are obtained with decoding of the rendered features. To address the point ambiguity issue, we propose to employ the semantic hierarchy defined by the Segment Anything Model (SAM) [19]. Specifically, for each 2D image, we obtain three well-segmented maps at different semantic levels with SAM. Then we extract the CLIP feature for each mask with precise object boundaries and assign this feature to every point on the corresponding mask. Learning with SAM-based masks not only endows each point with precise CLIP embeddings, resulting in higher model accuracy, but also enables direct querying at predefined three semantic scales. This circumvents the need for intensive searches across multiple absolute scales and the auxiliary DINO features, thereby effectively improving efficiency. We summarize the contributions of this paper as follows:", + "bbox": [ + 496, + 121, + 893, + 635 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the LangSplat, which is the first 3D Gaussian Splitting-based method for 3D language fields. A scene-specific autoencoder is further introduced to alleviate the memory cost issue imposed by explicit modeling.", + "- We propose to learn the hierarchical semantics defined by SAM to address the point ambiguity issue for 3D language field modeling.", + "- Experimental results show that our method outperforms the state-of-the-art methods on open-vocabulary 3D object localization and semantic segmentation tasks while being $199 \\times$ faster than LERF at $1440 \\times 1080$ resolution." + ], + "bbox": [ + 500, + 636, + 890, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 814, + 640, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. Real-time rendering has always been a pursued objective for neural rendering. Recently, Kerbl et al. [17] proposed to represent the 3D scene with a set of 3D Gaussians, which attained real-time rendering for", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20052", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1080p resolution while maintaining state-of-the-art visual quality. Encouraged by the success of 3D Gaussian Splatting, many works extend it to other tasks. To achieve real-time dynamic scene rendering, some studies [27, 45, 49] have extended the 3D Gaussian Splatting technique to dynamic scenes. Luiten et al. [27] proposed the Dynamic 3D Gaussians, which extended 3D Gaussians to dynamic scenes by explicitly modeling the 3D Gaussians across different time steps. Yang et al. [48] presented a deformable 3D Gaussians Splatting method, which learned 3D Gaussians in canonical space and modeled the dynamic scenes with a deformation field. Meanwhile, some researchers have combined 3D Gaussian Splatting with diffusion models to achieve efficient text-to-3D generation [10, 42, 51]. For example, Tang et al. [42] introduced DreamGaussian for efficient 3D content generation with a generative 3D Gaussian Splatting model. Unlike these methods, our paper extends each 3D Gaussian with language embeddings for open-vocabulary 3D queries.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM. The Segment Anything Model [19], released by Meta in 2023, has attracted considerable attention [15, 28, 29], which achieved impressive zero-shot performance. It has become the foundational model for image segmentation. SAM has been used for many vision tasks [26, 30, 44, 50] such as image inpainting [53], object tracking [11, 47], image editing [12], and so on. Many efforts have also been made to utilize SAM in the 3D domain [6, 36]. Liu et al. [25] proposed Seal to explore the potential of VFMs including SAM for point cloud segmentation. SA3D [6] generalized SAM to 3D objects by leveraging NeRF to connect 2D images and 3D space. Different from these works, we use SAM to obtain accurate object masks with three well-defined hierarchical semantics to train a 3D language field.", + "bbox": [ + 76, + 382, + 472, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Langue fields. Some early attempts to construct 3D feature fields included Distilled Feature Fields [20] and Neural Feature Fusion Fields [43]. They learned 3D-consistent features by distilling LSeg [21] or DINO [4] features across multiple views into a NeRF. Shen et al. [39] further adopted distilled feature fields for few-shot language-guided robotic manipulation by distilling CLIP feature into a NeRF. There are also some efforts [40, 54] that embed semantic information into NeRFs. For example, Semantic NeRF [54] jointly encoded semantics with appearance and geometry within a NeRF for novel semantic view synthesis. LERF [18] was the first to embed CLIP features into NeRF, enabling open-vocabulary 3D queries leveraging the powerful CLIP representation. DINO features were also used for supervising LERF to improve its performance. Liu et al. [24] also utilized CLIP and DINO features to train a NeRF model for 3D open-vocabulary segmentation. While these methods use NeRF for 3D modeling and suffer from the costly rendering process, we propose the 3D language Gaussian Splatting to obtain efficient 3D language fields.", + "bbox": [ + 76, + 598, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 498, + 89, + 691, + 108 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first revisit the challenges of modeling 3D language fields and then elaborate on how our proposed LangSplat addresses these issues. Figure 2 depicts the framework of our proposed LangSplat.", + "bbox": [ + 496, + 114, + 893, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting the Challenges of Language Fields", + "text_level": 1, + "bbox": [ + 498, + 181, + 879, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote an input image as $I \\in \\mathbb{R}^{3 \\times H \\times W}$ , where $H$ and $W$ represent the height and weight of the image size. We take a set of calibrated images $\\{I_t | t = 1, 2, \\ldots, T\\}$ as input and train a 3D language field $\\Phi$ with these images. Most existing methods [18, 24, 35] employ the CLIP image encoder $V$ to extract image features and utilize the extracted CLIP embeddings to supervise the 3D language field $\\Phi$ , leveraging the well-aligned text-image latent space [22, 34, 55] provided by CLIP, thus facilitating open-vocabulary queries. However, CLIP embeddings are image-aligned rather than pixel-aligned. In other words, simply computing $V(I_t) \\in \\mathbb{R}^D$ only obtains an image-level feature, whereas what we need is a pixel-aligned language embedding $L_t \\in \\mathbb{R}^{D \\times H \\times W}$ , where $D$ represents the CLIP feature dimension. Meanwhile, modeling pixel-aligned language features faces the issue of point ambiguity, as a single point on an object contributes to multiple semantic levels of regions. For instance, a point on a cat's ear simultaneously contributes to the cat's ear, the cat's head, and the entire cat, and should be activated by all three types of textual queries.", + "bbox": [ + 496, + 205, + 892, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these issues, most existing methods [18, 24] extract a hierarchy of CLIP features from cropped image patches. Specifically, for a pixel with coordinates $v \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}$ , the corresponding CLIP features are obtained from image patches centered around $v$ at different absolute physical scales $s$ , with the expectation that at a certain scale $s$ , the patch can fully encompass the object. However, this multi-scale approach has two limitations. Firstly, patch features are imprecise because they often include additional contextual object information, leading to overly smoothed language fields with indistinct object boundaries. To alleviate the patchy issue, most methods [18, 24] leverage additional pixel-aligned DINO features to supervise the network. However, the learned 3D language features are still imprecise, as illustrated in Figure 1. Secondly, it requires simultaneous rendering at multiple scales during inference to find the optimal scale. With the number of scales $s$ potentially reaching as high as 30 [18], this significantly diminishes the inference speed.", + "bbox": [ + 496, + 508, + 893, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides the rendering target, another key design space is the 3D modeling approach. Most existing methods [2, 43] employ NeRFs for 3D representation, where they learn a language feature at each 3D point and subsequently render the language feature onto an image, similar to color rendering. However, NeRF-based methods are constrained by their time-consuming rendering process, even though the", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20053", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6044390a5174794305df198a967eb2f351851440a5dd54a91c6bbe3e252316cd.jpg", + "image_caption": [ + "Figure 2. The framework of our LangSplat. Our LangSplat leverages SAM to learn hierarchical semantics to address the point ambiguity issue. Then segment masks are sent to the CLIP image encoder to extract the corresponding CLIP embeddings. We learn an autoencoder with these obtained CLIP embeddings. Our 3D language Gaussian learn language features on the scene-specific latent space to reduce the memory cost. During querying, the rendered language embeddings are sent to the decoder to recover the features on the CLIP space." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "most advanced NeRF techniques currently available cannot achieve real-time rendering in high-resolution, unrestricted scenes [17]. Meanwhile, there is a high demand for efficient open vocabulary querying in practical applications, especially in fields such as intelligent robotics.", + "bbox": [ + 75, + 329, + 470, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Hierarchical Semantics with SAM", + "text_level": 1, + "bbox": [ + 76, + 419, + 449, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a foundation model for image segmentation, SAM [19] can accurately group a pixel with its surrounding pixels belonging to the same object, thereby segmenting the image into many object masks with clear boundaries. Furthermore, SAM addresses point ambiguity by generating three different masks for a point prompt, namely, whole, part, and subpart, representing three hierarchical levels of semantics. In this paper, we propose leveraging SAM to obtain precise object masks, which are then used to acquire pixel-aligned features. We also explicitly model the semantic hierarchy defined by SAM to address the point ambiguity issue. With SAM, we can capture the semantic hierarchy of objects in 3D scenes, providing accurate and multi-scale segmentation maps for each input image.", + "bbox": [ + 75, + 445, + 468, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we feed a regular grid of $32 \\times 32$ point prompts into SAM to obtain the masks under three different semantic levels: $M_0^s, M_0^p, M_0^w$ , where $M_0^s, M_0^p$ , and $M_0^w$ represent the masks at subpart, part, and whole levels, respectively. Then we remove redundant masks for each of the three mask sets based on the predicted IoU score, stability score, and overlap rate between masks. Each filtered mask set independently performs a comprehensive full-image segmentation based on its respective semantic level, resulting in three segmentation maps: $M^s, M^p, M^w$ . These segmentation maps precisely delineate the boundaries of objects at their hierarchical levels, effectively partitioning the scene into semantically meaningful regions. With the obtained segmentation maps, we proceed to extract CLIP features for each segmented region. These features capture the semantic context of the objects", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "at various levels within the scene. Mathematically, the obtained pixel-aligned language embeddings are:", + "bbox": [ + 498, + 329, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {L} _ {t} ^ {l} (v) = V \\left(\\boldsymbol {I} _ {t} \\odot \\boldsymbol {M} ^ {l} (v)\\right), l \\in \\{s, p, w \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 369, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M^l(v)$ represents the mask region to which pixel $v$ belongs at the semantic level $l$ .", + "bbox": [ + 498, + 397, + 890, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each pixel rendered from the 3D language scene now possesses a CLIP feature that aligns with its precise semantic context. This alignment reduces ambiguity and enhances the accuracy of language-based queries. We can learn an accurate 3D language field even without the commonly used DINO regularization. Another advantage of our SAM-based approach is the predefined semantic scales. Since we have distinct segmentation maps for \"whole,\" \"part,\" and \"subpart\" levels, we can directly query the 3D language field at these predefined scales. This eliminates the need for intensive searches across multiple absolute scales, making the querying process more efficient. By incorporating SAM's semantic hierarchy into our approach, we not only improve the accuracy of our 3D language field but also streamline the querying process, making it more efficient and effective for a wide range of applications.", + "bbox": [ + 496, + 429, + 892, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 3D Gaussian Splitting for Language Fields", + "text_level": 1, + "bbox": [ + 498, + 678, + 864, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Having obtained the language embeddings on a set of 2D images $\\{L_t^l, |t = 1, \\dots, T\\}$ , we can learn a 3D language scene by modeling the relations between 3D points and 2D pixels. Most existing methods [2, 43] suffer from the costly rendering process as they adopt NeRFs for 3D modeling. To address this issue, we present the first 3D Gaussian Splitting-based method for 3D language field modeling.", + "bbox": [ + 496, + 700, + 890, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian Splitting explicitly represents a 3D scene as a collection of anisotropic 3D Gaussians, with each Gaussian $G(x)$ characterized by a mean $\\mu \\in \\mathbb{R}^3$ and a covariance matrix $\\Sigma$ :", + "bbox": [ + 496, + 808, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 875, + 890, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20054", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the parameters of 3D Gaussians, they are rendered into 2D image planes [56], and a tile-based rasterizer is used to improve the rendering efficiency:", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC (v) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 148, + 468, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{i}$ is the color of the $i$ -th Gaussian, $\\mathcal{N}$ denotes the Gaussians in the tile, $C(v)$ is the rendered color at pixel $v$ , and $\\alpha_{i} = o_{i}G_{i}^{2D}(v)$ . Here $o_{i}$ is the opacity of the $i$ -th Gaussian and $G_{i}^{2D}(\\cdot)$ represents the function of the $i$ -th Gaussian projected onto 2D.", + "bbox": [ + 76, + 199, + 468, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this paper, we proposes the 3D language Gaussian Splatting, which augments each 3D Gaussian with three language embeddings $\\{\\pmb{f}^s,\\pmb{f}^p,\\pmb{f}^w\\}$ . These embeddings are derived from CLIP features, which capture the hierarchical semantics provided by SAM. The augmented Gaussians are named as 3D language Gaussians. We also adopt the tile-based rasterizer to retain the rendering efficiency:", + "bbox": [ + 75, + 276, + 468, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {l} (v) = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {f} _ {i} ^ {l} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), l \\in \\{s, p, w \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 393, + 468, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F^l(v)$ represents the language embedding rendered at pixel $v$ with the semantic level $l$ . By incorporating language information directly into the Gaussians, we enable the 3D language field to respond to language-based queries.", + "bbox": [ + 76, + 446, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As an explicit modeling approach, our LangSplat may create millions of 3D points to model a complex 3D scene. As CLIP embeddings are high-dimensional features, directly learning $\\pmb{f}^{l}$ on the CLIP latent space significantly increases memory and time cost. Compared to learning RGB colors without spherical harmonics coefficients, learning 512-dimensional CLIP features increases the memory requirements for storing 3D Gaussians by over 35 times, easily leading to running out of L1 cache memory. To reduce memory cost and improve efficiency, we introduce a scenewise language autoencoder. This autoencoder maps CLIP embeddings in a scene to a lower-dimensional latent space, reducing memory requirements. The CLIP model is trained using 400 million (image, text) pairs, and its $D$ -dimensional latent space could be highly compact, as it needs to align arbitrary text and images in this space. However, the language field $\\Phi$ we train here is scene-specific, meaning we can leverage scene priors to compress CLIP features. In fact, for each input image, we will obtain hundreds of masks segmented by SAM, which is significantly smaller than the number of images used in CLIP training. Therefore, all the segmented regions in a scene are sparsely distributed in the CLIP latent space, allowing us to further compress these CLIP features using a scene-specific autoencoder.", + "bbox": [ + 75, + 508, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we use the collections of CLIP features of SAM segmented masks $\\{L_t^l | l \\in \\{s, p, w\\}, 1 \\leq t \\leq T\\}$ to", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "train a lightweight autoencoder. An encoder $E$ maps the $D$ -dimensional CLIP features $\\pmb{L}_t^l (v)\\in \\mathbb{R}^D$ to $\\pmb{H}_t^l (v) = E(\\pmb{L}_t^l (v))\\in \\mathbb{R}^d$ , where $d\\ll D$ . Then we learn a decoder $\\Psi$ to reconstruct the original CLIP embeddings from the compressed representation. The autoencoder is trained with a reconstruction objective on the CLIP embeddings $\\{\\pmb{L}_t^l\\}$ :", + "bbox": [ + 498, + 90, + 890, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a e} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {a e} \\left(\\Psi \\left(E \\left(\\boldsymbol {L} _ {t} ^ {l} (v)\\right)\\right), \\boldsymbol {L} _ {t} ^ {l} (v)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 191, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{ae}()$ denotes a distance function used for the autoencoder. Here we adopt both $\\mathcal{L}_1$ and a cosine distance loss.", + "bbox": [ + 498, + 244, + 890, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training the autoencoder, we transform all CLIP embeddings $\\{L_t^l\\}$ into scene-specific latent features $\\{H_t^l\\}$ . We let our 3D language Gaussians learn language embeddings in the scene-specific latent space instead of the CLIP latent space. Therefore, we have $f^l \\in \\mathbb{R}^d$ . In practice, we choose $d = 3$ as it yields excellent model efficiency and accuracy. Compared to directly modeling the $D$ -dimensional CLIP embeddings, our method significantly reduced the memory cost by incorporating scene priors. We optimized the language embeddings with the objective:", + "bbox": [ + 498, + 276, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l a n g} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {l a n g} \\left(\\boldsymbol {F} _ {t} ^ {l} (v), \\boldsymbol {H} _ {t} ^ {l} (v)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 436, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_{lang}()$ denotes the distance function used for our 3D Language Gaussians.", + "bbox": [ + 498, + 489, + 890, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we follow the Eq. (4) to render the language embeddings from 3D to 2D, and then we use the trained scene-specific decoder $\\Psi$ to recover the CLIP image embeddings $\\Psi (\\pmb {F}_t^l)\\in \\mathbb{R}^{D\\times H\\times W}$ , which enable open-vocabulary queries with the CLIP text encoder.", + "bbox": [ + 498, + 521, + 890, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By enhancing 3D Gaussians with language embedding and employing a scene-wise language autoencoder, our proposed LangSplat presents a powerful and efficient solution for building 3D language fields. This approach not only preserves the rendering efficiency of Gaussian Splatting but also mitigates the catastrophic memory explosion associated with explicit modeling.", + "bbox": [ + 498, + 597, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Open-vocabulary Querying", + "text_level": 1, + "bbox": [ + 500, + 710, + 746, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the well-aligned latent space between images and text provided by the CLIP model, our learned 3D language field can easily support open-vocabulary 3D queries, including open-vocabulary 3D object localization and open-vocabulary 3D semantic segmentation. Many existing open-vocabulary 3D semantic segmentation methods [24] usually select the category from a category list, which includes the categories present in the images. However, obtaining a comprehensive category list for in-the-wild scenes is challenging. Different from them, our method generates precise object masks given an arbitrary text query.", + "bbox": [ + 496, + 734, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20055", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following LERF [18], we compute the relevancy score for each text query. Specifically, for each rendered language embedding $\\phi_{img}$ and each text query $\\phi_{qry}$ , the relevancy score is defined as $\\min_i\\frac{\\exp(\\phi_{img} \\cdot \\phi_{qry})}{\\exp(\\phi_{img} \\cdot \\phi_{qry}) + \\exp(\\phi_{img} \\cdot \\phi_{canon}^i)}$ , where $\\phi_{canon}^i$ is the CLIP embeddings of a predefined canonical phrase chosen from \"object\", \"things\", \"stuff\", and \"texture\". Hence, for each text query, we will obtain three relevancy maps, each representing results at a specific semantic level. We follow the strategy used in LERF [18] and choose the semantic level that yields the highest relevancy score. For the 3D object localization task, we directly choose the point with the highest relevance score. For the 3D semantic segmentation task, we filter out points with relevancy scores lower than a chosen threshold, and predict the object masks with remaining regions. Please refer to the appendix for additional details.", + "bbox": [ + 75, + 90, + 472, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 353, + 209, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 76, + 378, + 174, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We employ two datasets for evaluation. The LERF dataset [18] is captured using the iPhone App Polycam, which consists of complex in-the-wild scenes. The LERF dataset is designed for 3D object localization tasks, here we extend the LERF dataset by annotating ground truth masks for textual queries, enabling the evaluation of the open-vocabulary 3D semantic segmentation on the LERF dataset. As the original LERF annotations for 3D object localization are relatively simple, performance in some scenarios has already approached saturation. Therefore, we further manually annotated additional challenging localization samples to better evaluate method performance. We report localization accuracy for the 3D object localization task following LERF [18], and report the IoU results for the 3D semantic segmentation task. We also employ the 3D-OVS dataset [24], which comprises a collection of long-tail objects captured in diverse poses and backgrounds. This dataset is developed for open-vocabulary 3D semantic segmentation, where a full list of categories is provided. While other methods use the full list to generate the predicted masks, we only use the query category to generate the corresponding masks. The mIoU metric is used for this dataset.", + "bbox": [ + 75, + 401, + 472, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. To extract the language features of each image, we utilize the OpenCLIP ViT-B/16 model. For SAM, we use the ViT-H model to segment 2D masks. For each scene, we first use 3D Gaussian Splatting to train an RGB scene. We train it for 30,000 iterations, and in the end, each scene comprises around 2,500,000 points. We follow the default parameter setting as in [17] to train the RGB scene. Then we train our 3D language Gaussians by fixing all other parameters of 3D Gaussians such as mean and opacity. Only the language features are learnable during this stage. We train the language features for 30,000", + "bbox": [ + 75, + 734, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2c27469ba1ff54ff355ef44f4c37fa9d60a189cafa66443913bf7a8deecd7e53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen14.162.073.2
figurines8.975.080.4
teatime33.984.888.1
waldo_kitchen27.372.795.5
overall21.173.684.3
", + "bbox": [ + 506, + 88, + 887, + 202 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/da6a82063fe732c3252edff1ec9b9c9a0e00a4ac9f567f88e2405ed62d1b2b55.jpg", + "table_caption": [ + "Table 1. Localization accuracy (%) comparisons on LERF dataset." + ], + "table_footnote": [], + "table_body": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen7.028.251.2
figurines7.638.644.7
teatime21.745.065.1
waldo_kitchen29.937.944.5
overall16.637.451.4
", + "bbox": [ + 506, + 234, + 887, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons of 3D semantic segmentation on the LERF dataset. We report the average IoU scores $(\\%)$", + "bbox": [ + 498, + 359, + 890, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "iterations. Our autoencoder is implemented by an MLP, which compresses the 512-dimensional CLIP features into 3-dimensional latent features. For a $1440 \\times 1080$ resolution scene, our model is trained for $\\sim 25$ minutes on an NVIDIA RTX-3090 GPU and takes roughly 4GB of memory.", + "bbox": [ + 498, + 410, + 890, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on the LERF dataset", + "text_level": 1, + "bbox": [ + 500, + 497, + 756, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. We first compare our method with other methods on the LERF dataset. Table 1 shows the localization results. We observe that our method achieves an overall accuracy of $84.3\\%$ , significantly outperforming LERF. Table 2 further shows the IoU results of 3D semantic segmentation, our method outperforms LERF by $14.0\\%$ , which illustrates the superiority of our proposed LangSplat. Visualization Results. To show the learned 3D language field, we visualize the learned features by computing 3-dimensional PCA components of learned language features following [20]. The results are shown in Figure 1. We see that the LERF learned features fail to generate clear boundaries between objects while our method gives precise object shapes solely using CLIP features. We further show the visualization results of object localization and semantic segmentation in Figure 3 and Figure 4, respectively. We observe that the activation regions generated by LERF are more dispersed, while ours are more concentrated, and our activation regions can better align with the ground truth shape compared to those produced by LERF.", + "bbox": [ + 496, + 521, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study. We conduct ablations on the ramen scene and report the semantic segmentation results in Table 3. We test the query speed on an NVIDIA RTX-3090 GPU. Here AE represents autoencoder and 3D-GS denotes 3D Gaussian Splatting. Without any of our proposed components,", + "bbox": [ + 496, + 825, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20056", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/297a3511d890d5cf3169741f7f9387f8c7ef02c5094bef0cd55e7dc2d3ded6f1.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 210, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dface535705b12611324635b6e3be140a956aac20f0304b1c60f63abbc32543.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 95, + 339, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ca856bb07ca8822c3207a00fa5a719b68d2f9c1408989ad929f5f4a8a3739b6.jpg", + "image_caption": [ + "LERF" + ], + "image_footnote": [], + "bbox": [ + 86, + 176, + 210, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a15befdb603f253347b5bbe0fcf817b1e03173650902d9414e012c18c86959b.jpg", + "image_caption": [ + "LSeg", + "Ours", + "'spoon'" + ], + "image_footnote": [], + "bbox": [ + 215, + 176, + 339, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48501b968ef74b9b4932a69cf60bcf1c69b2338bbee29f24b3ae79ebbe125f9f.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 359, + 94, + 480, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11d888b7f4801b5aab536ef7eb2176c1b78e7f1ed7b77db58c0bdf724ee4d459.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 95, + 611, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/116feb8303a161929e797879fe807c60d476b245bc84d67db8f85f63d8ef0f0c.jpg", + "image_caption": [ + "LERF", + "'paper napkin'" + ], + "image_footnote": [], + "bbox": [ + 359, + 176, + 480, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f52d64e734b1a6e886766b7ed63ec5bda073e5348b188a44a4acf861eaff662b.jpg", + "image_caption": [ + "LSeg", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 488, + 176, + 611, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cbc8586c1493da12a3a04dc50d3658bb19ce1fdbf838ff73144eb40bc658710.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 629, + 94, + 753, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac2d94c40620224d73343a9d05a3d9f2214ee89670afd9c72515a41da401db05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 95, + 880, + 166 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/943809e1a7975c721d13079ccf52caeaf265cff030232e288948c880f67584ac.jpg", + "image_caption": [ + "LERF", + "'spatula'" + ], + "image_footnote": [], + "bbox": [ + 630, + 176, + 751, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/041282c7b0a12dc3d46646cd6d5a37ace264767cf9a903de2bdf8f9fda52bb9e.jpg", + "image_caption": [ + "LSeg", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 759, + 176, + 880, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f6cb00c205fbe5d62bf2238bab3476e5b59dfa7000793a15a4ff1a2a1101a3f.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons of open-vocabulary 3D object localization on the LERF dataset. The red points are the model predictions and the black dashed bounding boxes denote the annotations.", + "Figure 4. Qualitative comparisons of open-vocabulary 3D semantic segmentation on the LERF dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 325, + 890, + 458 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7f59a20cdde0a3da12e2926a97179e967af38e98fea2922a3356f035e3b11440.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ComponentPerformance
AE3D-GSSAMIoU (%)Speed (s/q)
✓28.2030.93
46.747.77
✓✓OOMOOM
✓✓✓51.150.26
", + "bbox": [ + 89, + 489, + 457, + 609 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ea05b997b1b9b1ca545b03744726975a420387de61129d875046a9eacd8e02d.jpg", + "table_caption": [ + "Table 3. Ablations. The results are obtained on the ramen scene." + ], + "table_footnote": [], + "table_body": "
ComponentPerformance
AE3D-GSSAMmIoU (%)Speed (s/q)
✓53.255.7
85.718.4
✓✓OOMOOM
✓✓✓94.20.28
", + "bbox": [ + 83, + 643, + 462, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablations result on the bench scene of the 3D-OVS dataset. The image resolution is $1440 \\times 1080$ .", + "bbox": [ + 75, + 773, + 467, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our baseline equals LERF, which has a speed of 30.93 seconds per text query at the resolution of $988 \\times 731$ . Using SAM to replace the scale-based solution significantly increases the IoU by $18.54\\%$ , showing our SAM-based solution effectively addresses the point ambiguity issue, leading", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to accurate 3D language features. Simply replacing NeRF with 3D Gaussian Splatting leads to running out of L1 cache memory as explicitly modeling CLIP features poses huge memory demands. Incorporating a scene-specific autoencoder effectively addresses this issue and results in further improvements in both accuracy and efficiency. In the end, our LangSplat achieved a $119 \\times$ speedup over LERF while significantly surpassing LERF in terms of accuracy.", + "bbox": [ + 496, + 492, + 890, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further conducted the ablations on the 3D-OVS dataset, which has a higher image resolution of $1440 \\times 1080$ . Table 4 lists the results on the bench scene. We also tested the query speed on an NVIDIA RTX-3090 GPU. We observed that with the increase in image resolution, the speedup over LERF further improved to $199 \\times$ , which demonstrates the huge potential of our method. Our further study shows that most of the computational time is allocated to the decoder rather than the rendering process. We could replace the decoder with a $1 \\times 1$ convolutional layer to attain a higher speedup.", + "bbox": [ + 496, + 616, + 892, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on the 3D-OVS dataset", + "text_level": 1, + "bbox": [ + 498, + 799, + 769, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. We compare LangSplat with other 2D and 3D state-of-the-art methods on the 3D-OVS dataset in Table 5. We observe that LangSplat not only outperforms 2D-based methods such as ODISE [46] and OV-Seg [23], but also achieves better results than 3D-based methods in", + "bbox": [ + 496, + 824, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20057", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/296657d11ae3cf385d956d05e6d7acc4da81d368a54d8df763ff998a37a83707.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons of different methods on the 3D-OVS dataset. We visualize the segmentation results in 2 scenes. We observe that our method gives the most accurate segmentation maps." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 885, + 486 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b5a6aa3d12e2862abfca92cf9f526dec5f4da45abef4f23d6c315d89e2a24f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodbedbenchroomsofalawnoverall
LSeg [21]56.06.019.24.517.520.6
ODISE [46]52.624.152.548.339.843.5
OV-Seg [23]79.888.971.466.181.277.5
FFD [20]56.66.125.13.742.926.9
LERF [18]73.553.246.62773.754.8
3D-OVS [24]89.589.392.87488.286.8
LangSplat92.594.294.190.096.193.4
", + "bbox": [ + 78, + 534, + 468, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative comparisons of 3D semantic segmentation on the 3D-OVS dataset. We report the mIoU scores $(\\%)$", + "bbox": [ + 75, + 695, + 468, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cluding LERF [18] and 3D-OVS [24] by a large margin. Note that in this dataset, we generate object masks only based on the query text while others, such as 3D-OVS, require the complete category list. In the end, our method achieves an overall mIoU of $93.4\\%$ , which shows that LangSplat effectively learns a precise 3D language field.", + "bbox": [ + 75, + 733, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. We present the qualitative results in Figure 5. As LERF suffers from the patchy issue and learns over-smoothed features, it fails to find accurate object boundaries. Among all state-of-the-art methods, our methods give the most accurate segmentation maps, which", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "further demonstrates the effectiveness of our LangSplat.", + "bbox": [ + 500, + 537, + 870, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 570, + 619, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented LangSplat, a method for constructing 3D language fields that enables precise and efficient open-vocabulary querying within 3D spaces. By extending 3D Gaussian Splatting with language features and learning a scene-specific language autoencoder, LangSplat circumvents slow rendering speed associated with NeRF-based methods. Furthermore, we propose to learn the semantic hierarchy defined by SAM, which effectively resolves the point ambiguity problem. The experimental results clearly demonstrate LangSplat's superiority over existing state-of-the-art methods like LERF, particularly in terms of its remarkable $199 \\times$ speed improvement and enhanced performance in open-ended 3D language query tasks.", + "bbox": [ + 496, + 597, + 890, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was funded through National Key Research and Development Program of China (Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (JCYJ20220818101001004, JSGG20210802153150005), and in part by NIH grant R01HD104969, 1U01CA284207, and NSF grant CRCNS-2309041.", + "bbox": [ + 496, + 795, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20058", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 2", + "[2] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633, 2023. 3, 4", + "[3] Nancy Bonvillain. Language, culture, and communication: The meaning of messages. Rowman & Littlefield, 2019. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3", + "[5] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, pages 5056-5066, 2022. 2", + "[6] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. 2023. 3", + "[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 2", + "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCVn, pages 333-350. Springer, 2022. 2", + "[9] Boyuan Chen, Fei Xia, Brian Ichter, Kanishka Rao, Keerthana Gopalakrishnan, Michael S Ryoo, Austin Stone, and Daniel Kappler. Open-vocabulary queryable scene representations for real world planning. In ICRA, pages 11509-11522. IEEE, 2023. 2", + "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 3", + "[11] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3", + "[12] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, MingMing Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In ACM MM, Demo track, 2023. 3", + "[13] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In CVPR, pages 4089-4098, 2018. 2", + "[14] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In ICRA, pages 10608-10615. IEEE, 2023. 2", + "[15] Siyuan Huang, Zhengkai Jiang, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Instruct2act: Mapping multimodality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176, 2023. 3", + "[16] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):1-14, 2023. 2, 4, 6", + "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2, 3, 6, 8", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 4", + "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. NeurlPS, 35:23311-23330, 2022. 2, 3, 6, 8", + "[21] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 3, 6, 8", + "[22] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, 35:35313-35325, 2022. 3", + "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, pages 7061-7070, 2023. 7, 8", + "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2, 3, 5, 6, 8", + "[25] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347, 2023. 3", + "[26] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 3", + "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3", + "[28] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 3", + "[29] Gengchen Mai, Weiming Huang, Jin Sun, Suhang Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, et al. On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798, 2023. 3", + "[30] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 3", + "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20059", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2", + "[33] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318–10327, 2021. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3", + "[35] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. Clip-fields: Weakly supervised semantic fields for robotic memory. arXiv preprint arXiv:2210.05663, 2022. 3", + "[36] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3", + "[37] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In ECCV, pages 666-682. Springer, 2022. 2", + "[38] Shuai Shen, Wanhua Li, Xiaoke Huang, Zheng Zhu, Jie Zhou, and Jiwen Lu. Sd-nerf: Towards lifelike talking head animation via spatially-adaptive dual-driven nerfs. TMM, 2023. 2", + "[39] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. arXiv preprint arXiv:2308.07931, 2023. 2, 3", + "[40] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In CVPR, pages 9043–9052, 2023. 3", + "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2", + "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3", + "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 3DV, pages 443-453. IEEE, 2022. 3, 4", + "[44] Jia Wan, Wanhua Li, Atmadeep Banerjee, Jason Ken Adhinarta, Evelina Sjostedt, Jingpeng Wu, Jeff Lichtman, Hanspeter Pfister, and Donglai Wei. Trisam: Tri-plane sam for zero-shot cortical blood vessel segmentation in vem images. arXiv preprint arXiv:2401.13961, 2024. 3", + "[45] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3", + "[46] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panop" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 7, 8", + "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3", + "[48] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3", + "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3", + "[50] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121, 2023.3", + "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3", + "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2", + "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3", + "[54] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, pages 15838-15847, 2021. 2, 3", + "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 3", + "[56] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In VIS, pages 29-538. IEEE, 2001. 5" + ], + "bbox": [ + 501, + 92, + 890, + 643 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20060", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_model.json b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5afcb1216c14650c0614c7769cc6ddbc41eed522 --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.131, + 0.713, + 0.154 + ], + "angle": 0, + "content": "LangSplat: 3D Language Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.188, + 0.865, + 0.226 + ], + "angle": 0, + "content": "Minghan Qin\\(^{1,*}\\), Wanhua Li\\(^{2,*, \\boxtimes}\\), Jiawei Zhou\\(^{1,*}\\), Haoqian Wang\\(^{1, \\boxtimes}\\), Hanspeter Pfister\\(^{2}\\), Tsinghua University\\(^{2}\\) Harvard University" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.229, + 0.856, + 0.244 + ], + "angle": 0, + "content": "qmh21@mails.tsinghua.edu.cn, wanhua@seas.harvard.edu, zhoujw22@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.247, + 0.719, + 0.261 + ], + "angle": 0, + "content": "wanghaoqian@tsinghua.edu.cn, pfister@seas.harvard.edu" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.301, + 0.362, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.302, + 0.625, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.302, + 0.888, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.549, + 0.893, + 0.591 + ], + "angle": 0, + "content": "Figure 1. Visualization of learned 3D language features of the previous SOTA method LERF and our LangSplat. While LERF generates imprecise and vague 3D features, our LangSplat accurately captures object boundaries and provides precise 3D language fields. While being effective, our LangSplat is also \\(199 \\times\\) faster than LERF at the resolution of \\(1440 \\times 1080\\)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.597, + 0.314, + 0.612 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.471, + 0.871 + ], + "angle": 0, + "content": "Humans live in a 3D world and commonly use natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.598, + 0.893, + 0.809 + ], + "angle": 0, + "content": "thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experimental results show that LangSplat significantly outperforms the previous state-of-the-art method LERF by a large margin. Notably, LangSplat is extremely efficient, achieving a \\(199 \\times\\) speedup compared to LERF at the resolution of \\(1440 \\times 1080\\). We strongly recommend readers to check out our video results at https://langsplat.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.844, + 0.633, + 0.859 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Language is the primary means of communication for human beings [3]. Modeling a 3D language field allows users" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.886, + 0.351, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. \\( {}^{ \\ominus } \\) Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20051" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.213 + ], + "angle": 0, + "content": "to interact with and query 3D worlds using open-ended language, which presents a promising avenue for human-computer interaction and understanding [1, 5, 13]. The field of open-ended language queries in 3D has attracted increasing attention due to its various applications such as robotic navigation [14] and manipulation [39], 3D semantic understanding [9, 54] and editing [20], autonomous driving [16], and augmented/virtual reality [24]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.474, + 0.52 + ], + "angle": 0, + "content": "Due to the absence of large-scale and diverse 3D scene data with language annotations, the current prevailing approach like LERF [18] involves feature distillation from off-the-shelf vision-language models such as CLIP into a 3D scene. However, these methods [18, 24] suffer from significant limitations in both speed and accuracy, severely constraining their practical applicability. To address these two issues, we revisit two key aspects of 3D language field modeling: the 3D modeling approach that bridges the gap between 2D and 3D, and the rendering target, which determines what to learn for a 3D point. For the 3D modeling technology, most methods utilize neural radiance fields (NeRFs) to represent a 3D scene, where volume rendering techniques are employed to accumulate 3D points along a ray into a single pixel. While NeRF has been demonstrated for its powerful 3D representation capabilities [7, 32, 33, 37, 38, 52], the volume rendering nature leads to its computationally expensive rendering speed [8, 31, 41], which imposes notable constraints on the potential applications within the NeRF-based language field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Regarding the rendering target, learning a CLIP embedding for a 3D point could be ambiguous as CLIP embeddings are aligned with images rather than pixels. Employing CLIP embeddings from a cropped patch also raises the point ambiguity issue, as the same 3D position can be associated with semantic concepts of varying scales. For instance, a point located on a bear's nose should yield high response values for three distinct textual queries: \"bear's nose\", \"bear's head\", and \"bear\" given that this point contributes to all three hierarchical regions. To deal with this issue, current methods [18, 24] introduce an additional absolute scale input to NeRF, trains with patch-wise CLIP features at different scales, and densely renders 2D maps at multiple scales during querying to select the optimal one. However, this scale-based solution compromises both efficiency and effectiveness. It could increase query time by up to 30 times as it needs to render at multiple different scales. Moreover, most patches with varying scales often fail to accurately encompass objects, either frequently including other objects from the background or omitting portions of the target object. These inaccurate CLIP features lead to the trained 3D language field lacking clear boundaries and containing a significant amount of noise. Therefore, they often simultaneously learn pixel-aligned DINO features to mitigate this issue. However, the performance remains un" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "satisfactory. As shown in Figure 1, LERF still generates imprecise 3D language features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.636 + ], + "angle": 0, + "content": "In this paper, we propose the 3D Language Gaussian Splatting (LangSplat) to address the above issues. Instead of using NeRF to build 3D representations, we resort to 3D Gaussian Splatting, which represents a 3D scene as a collection of 3D Gaussians and uses tile-based splatting to achieve efficient rendering at high resolutions. Our LangSplat defines a set of 3D language Gaussians, with each Gaussian being enhanced by a language embedding. These language-enhanced Gaussians are supervised using CLIP embeddings extracted from image patches captured from multiple training views, ensuring multi-view consistency. As an explicit modeling method, directly storing the high-dimensional language embeddings for each 3D language Gaussian is memory-inefficient. To reduce the memory cost and further improve the rendering efficiency, we propose to first learn a scene-wise language autoencoder, which maps CLIP embeddings in a scene to a low-dimensional latent space. In this way, each language Gaussian only contains the low-dimensional latent language features and the final language embeddings are obtained with decoding of the rendered features. To address the point ambiguity issue, we propose to employ the semantic hierarchy defined by the Segment Anything Model (SAM) [19]. Specifically, for each 2D image, we obtain three well-segmented maps at different semantic levels with SAM. Then we extract the CLIP feature for each mask with precise object boundaries and assign this feature to every point on the corresponding mask. Learning with SAM-based masks not only endows each point with precise CLIP embeddings, resulting in higher model accuracy, but also enables direct querying at predefined three semantic scales. This circumvents the need for intensive searches across multiple absolute scales and the auxiliary DINO features, thereby effectively improving efficiency. We summarize the contributions of this paper as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.637, + 0.892, + 0.698 + ], + "angle": 0, + "content": "- We propose the LangSplat, which is the first 3D Gaussian Splitting-based method for 3D language fields. A scene-specific autoencoder is further introduced to alleviate the memory cost issue imposed by explicit modeling." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.698, + 0.892, + 0.742 + ], + "angle": 0, + "content": "- We propose to learn the hierarchical semantics defined by SAM to address the point ambiguity issue for 3D language field modeling." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.743, + 0.892, + 0.803 + ], + "angle": 0, + "content": "- Experimental results show that our method outperforms the state-of-the-art methods on open-vocabulary 3D object localization and semantic segmentation tasks while being \\(199 \\times\\) faster than LERF at \\(1440 \\times 1080\\) resolution." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.637, + 0.892, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.642, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D Gaussian Splatting. Real-time rendering has always been a pursued objective for neural rendering. Recently, Kerbl et al. [17] proposed to represent the 3D scene with a set of 3D Gaussians, which attained real-time rendering for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20052" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.38 + ], + "angle": 0, + "content": "1080p resolution while maintaining state-of-the-art visual quality. Encouraged by the success of 3D Gaussian Splatting, many works extend it to other tasks. To achieve real-time dynamic scene rendering, some studies [27, 45, 49] have extended the 3D Gaussian Splatting technique to dynamic scenes. Luiten et al. [27] proposed the Dynamic 3D Gaussians, which extended 3D Gaussians to dynamic scenes by explicitly modeling the 3D Gaussians across different time steps. Yang et al. [48] presented a deformable 3D Gaussians Splatting method, which learned 3D Gaussians in canonical space and modeled the dynamic scenes with a deformation field. Meanwhile, some researchers have combined 3D Gaussian Splatting with diffusion models to achieve efficient text-to-3D generation [10, 42, 51]. For example, Tang et al. [42] introduced DreamGaussian for efficient 3D content generation with a generative 3D Gaussian Splatting model. Unlike these methods, our paper extends each 3D Gaussian with language embeddings for open-vocabulary 3D queries." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.383, + 0.473, + 0.597 + ], + "angle": 0, + "content": "SAM. The Segment Anything Model [19], released by Meta in 2023, has attracted considerable attention [15, 28, 29], which achieved impressive zero-shot performance. It has become the foundational model for image segmentation. SAM has been used for many vision tasks [26, 30, 44, 50] such as image inpainting [53], object tracking [11, 47], image editing [12], and so on. Many efforts have also been made to utilize SAM in the 3D domain [6, 36]. Liu et al. [25] proposed Seal to explore the potential of VFMs including SAM for point cloud segmentation. SA3D [6] generalized SAM to 3D objects by leveraging NeRF to connect 2D images and 3D space. Different from these works, we use SAM to obtain accurate object masks with three well-defined hierarchical semantics to train a 3D language field." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "3D Langue fields. Some early attempts to construct 3D feature fields included Distilled Feature Fields [20] and Neural Feature Fusion Fields [43]. They learned 3D-consistent features by distilling LSeg [21] or DINO [4] features across multiple views into a NeRF. Shen et al. [39] further adopted distilled feature fields for few-shot language-guided robotic manipulation by distilling CLIP feature into a NeRF. There are also some efforts [40, 54] that embed semantic information into NeRFs. For example, Semantic NeRF [54] jointly encoded semantics with appearance and geometry within a NeRF for novel semantic view synthesis. LERF [18] was the first to embed CLIP features into NeRF, enabling open-vocabulary 3D queries leveraging the powerful CLIP representation. DINO features were also used for supervising LERF to improve its performance. Liu et al. [24] also utilized CLIP and DINO features to train a NeRF model for 3D open-vocabulary segmentation. While these methods use NeRF for 3D modeling and suffer from the costly rendering process, we propose the 3D language Gaussian Splatting to obtain efficient 3D language fields." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.692, + 0.109 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.894, + 0.178 + ], + "angle": 0, + "content": "In this section, we first revisit the challenges of modeling 3D language fields and then elaborate on how our proposed LangSplat addresses these issues. Figure 2 depicts the framework of our proposed LangSplat." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.183, + 0.88, + 0.2 + ], + "angle": 0, + "content": "3.1. Revisiting the Challenges of Language Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.893, + 0.508 + ], + "angle": 0, + "content": "We denote an input image as \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\), where \\(H\\) and \\(W\\) represent the height and weight of the image size. We take a set of calibrated images \\(\\{I_t | t = 1, 2, \\ldots, T\\}\\) as input and train a 3D language field \\(\\Phi\\) with these images. Most existing methods [18, 24, 35] employ the CLIP image encoder \\(V\\) to extract image features and utilize the extracted CLIP embeddings to supervise the 3D language field \\(\\Phi\\), leveraging the well-aligned text-image latent space [22, 34, 55] provided by CLIP, thus facilitating open-vocabulary queries. However, CLIP embeddings are image-aligned rather than pixel-aligned. In other words, simply computing \\(V(I_t) \\in \\mathbb{R}^D\\) only obtains an image-level feature, whereas what we need is a pixel-aligned language embedding \\(L_t \\in \\mathbb{R}^{D \\times H \\times W}\\), where \\(D\\) represents the CLIP feature dimension. Meanwhile, modeling pixel-aligned language features faces the issue of point ambiguity, as a single point on an object contributes to multiple semantic levels of regions. For instance, a point on a cat's ear simultaneously contributes to the cat's ear, the cat's head, and the entire cat, and should be activated by all three types of textual queries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.894, + 0.795 + ], + "angle": 0, + "content": "To address these issues, most existing methods [18, 24] extract a hierarchy of CLIP features from cropped image patches. Specifically, for a pixel with coordinates \\( v \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\} \\), the corresponding CLIP features are obtained from image patches centered around \\( v \\) at different absolute physical scales \\( s \\), with the expectation that at a certain scale \\( s \\), the patch can fully encompass the object. However, this multi-scale approach has two limitations. Firstly, patch features are imprecise because they often include additional contextual object information, leading to overly smoothed language fields with indistinct object boundaries. To alleviate the patchy issue, most methods [18, 24] leverage additional pixel-aligned DINO features to supervise the network. However, the learned 3D language features are still imprecise, as illustrated in Figure 1. Secondly, it requires simultaneous rendering at multiple scales during inference to find the optimal scale. With the number of scales \\( s \\) potentially reaching as high as 30 [18], this significantly diminishes the inference speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Besides the rendering target, another key design space is the 3D modeling approach. Most existing methods [2, 43] employ NeRFs for 3D representation, where they learn a language feature at each 3D point and subsequently render the language feature onto an image, similar to color rendering. However, NeRF-based methods are constrained by their time-consuming rendering process, even though the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20053" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 2. The framework of our LangSplat. Our LangSplat leverages SAM to learn hierarchical semantics to address the point ambiguity issue. Then segment masks are sent to the CLIP image encoder to extract the corresponding CLIP embeddings. We learn an autoencoder with these obtained CLIP embeddings. Our 3D language Gaussian learn language features on the scene-specific latent space to reduce the memory cost. During querying, the rendered language embeddings are sent to the decoder to recover the features on the CLIP space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.406 + ], + "angle": 0, + "content": "most advanced NeRF techniques currently available cannot achieve real-time rendering in high-resolution, unrestricted scenes [17]. Meanwhile, there is a high demand for efficient open vocabulary querying in practical applications, especially in fields such as intelligent robotics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.42, + 0.45, + 0.436 + ], + "angle": 0, + "content": "3.2. Learning Hierarchical Semantics with SAM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.657 + ], + "angle": 0, + "content": "As a foundation model for image segmentation, SAM [19] can accurately group a pixel with its surrounding pixels belonging to the same object, thereby segmenting the image into many object masks with clear boundaries. Furthermore, SAM addresses point ambiguity by generating three different masks for a point prompt, namely, whole, part, and subpart, representing three hierarchical levels of semantics. In this paper, we propose leveraging SAM to obtain precise object masks, which are then used to acquire pixel-aligned features. We also explicitly model the semantic hierarchy defined by SAM to address the point ambiguity issue. With SAM, we can capture the semantic hierarchy of objects in 3D scenes, providing accurate and multi-scale segmentation maps for each input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Specifically, we feed a regular grid of \\(32 \\times 32\\) point prompts into SAM to obtain the masks under three different semantic levels: \\(M_0^s, M_0^p, M_0^w\\), where \\(M_0^s, M_0^p\\), and \\(M_0^w\\) represent the masks at subpart, part, and whole levels, respectively. Then we remove redundant masks for each of the three mask sets based on the predicted IoU score, stability score, and overlap rate between masks. Each filtered mask set independently performs a comprehensive full-image segmentation based on its respective semantic level, resulting in three segmentation maps: \\(M^s, M^p, M^w\\). These segmentation maps precisely delineate the boundaries of objects at their hierarchical levels, effectively partitioning the scene into semantically meaningful regions. With the obtained segmentation maps, we proceed to extract CLIP features for each segmented region. These features capture the semantic context of the objects" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.361 + ], + "angle": 0, + "content": "at various levels within the scene. Mathematically, the obtained pixel-aligned language embeddings are:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.37, + 0.892, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {L} _ {t} ^ {l} (v) = V \\left(\\boldsymbol {I} _ {t} \\odot \\boldsymbol {M} ^ {l} (v)\\right), l \\in \\{s, p, w \\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.398, + 0.892, + 0.428 + ], + "angle": 0, + "content": "where \\( M^l(v) \\) represents the mask region to which pixel \\( v \\) belongs at the semantic level \\( l \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.893, + 0.672 + ], + "angle": 0, + "content": "Each pixel rendered from the 3D language scene now possesses a CLIP feature that aligns with its precise semantic context. This alignment reduces ambiguity and enhances the accuracy of language-based queries. We can learn an accurate 3D language field even without the commonly used DINO regularization. Another advantage of our SAM-based approach is the predefined semantic scales. Since we have distinct segmentation maps for \"whole,\" \"part,\" and \"subpart\" levels, we can directly query the 3D language field at these predefined scales. This eliminates the need for intensive searches across multiple absolute scales, making the querying process more efficient. By incorporating SAM's semantic hierarchy into our approach, we not only improve the accuracy of our 3D language field but also streamline the querying process, making it more efficient and effective for a wide range of applications." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.679, + 0.866, + 0.696 + ], + "angle": 0, + "content": "3.3. 3D Gaussian Splitting for Language Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Having obtained the language embeddings on a set of 2D images \\(\\{L_t^l, |t = 1, \\dots, T\\}\\), we can learn a 3D language scene by modeling the relations between 3D points and 2D pixels. Most existing methods [2, 43] suffer from the costly rendering process as they adopt NeRFs for 3D modeling. To address this issue, we present the first 3D Gaussian Splitting-based method for 3D language field modeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.892, + 0.868 + ], + "angle": 0, + "content": "3D Gaussian Splitting explicitly represents a 3D scene as a collection of anisotropic 3D Gaussians, with each Gaussian \\( G(x) \\) characterized by a mean \\( \\mu \\in \\mathbb{R}^3 \\) and a covariance matrix \\( \\Sigma \\):" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.876, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nG (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)\\right). \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20054" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "To optimize the parameters of 3D Gaussians, they are rendered into 2D image planes [56], and a tile-based rasterizer is used to improve the rendering efficiency:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.149, + 0.47, + 0.192 + ], + "angle": 0, + "content": "\\[\nC (v) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.2, + 0.469, + 0.276 + ], + "angle": 0, + "content": "where \\(c_{i}\\) is the color of the \\(i\\)-th Gaussian, \\(\\mathcal{N}\\) denotes the Gaussians in the tile, \\(C(v)\\) is the rendered color at pixel \\(v\\), and \\(\\alpha_{i} = o_{i}G_{i}^{2D}(v)\\). Here \\(o_{i}\\) is the opacity of the \\(i\\)-th Gaussian and \\(G_{i}^{2D}(\\cdot)\\) represents the function of the \\(i\\)-th Gaussian projected onto 2D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.469, + 0.382 + ], + "angle": 0, + "content": "In this paper, we proposes the 3D language Gaussian Splatting, which augments each 3D Gaussian with three language embeddings \\(\\{\\pmb{f}^s,\\pmb{f}^p,\\pmb{f}^w\\}\\). These embeddings are derived from CLIP features, which capture the hierarchical semantics provided by SAM. The augmented Gaussians are named as 3D language Gaussians. We also adopt the tile-based rasterizer to retain the rendering efficiency:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.394, + 0.47, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {l} (v) = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {f} _ {i} ^ {l} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), l \\in \\{s, p, w \\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.469, + 0.509 + ], + "angle": 0, + "content": "where \\( F^l(v) \\) represents the language embedding rendered at pixel \\( v \\) with the semantic level \\( l \\). By incorporating language information directly into the Gaussians, we enable the 3D language field to respond to language-based queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.87 + ], + "angle": 0, + "content": "As an explicit modeling approach, our LangSplat may create millions of 3D points to model a complex 3D scene. As CLIP embeddings are high-dimensional features, directly learning \\( \\pmb{f}^{l} \\) on the CLIP latent space significantly increases memory and time cost. Compared to learning RGB colors without spherical harmonics coefficients, learning 512-dimensional CLIP features increases the memory requirements for storing 3D Gaussians by over 35 times, easily leading to running out of L1 cache memory. To reduce memory cost and improve efficiency, we introduce a scenewise language autoencoder. This autoencoder maps CLIP embeddings in a scene to a lower-dimensional latent space, reducing memory requirements. The CLIP model is trained using 400 million (image, text) pairs, and its \\( D \\)-dimensional latent space could be highly compact, as it needs to align arbitrary text and images in this space. However, the language field \\( \\Phi \\) we train here is scene-specific, meaning we can leverage scene priors to compress CLIP features. In fact, for each input image, we will obtain hundreds of masks segmented by SAM, which is significantly smaller than the number of images used in CLIP training. Therefore, all the segmented regions in a scene are sparsely distributed in the CLIP latent space, allowing us to further compress these CLIP features using a scene-specific autoencoder." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Specifically, we use the collections of CLIP features of SAM segmented masks \\(\\{L_t^l | l \\in \\{s, p, w\\}, 1 \\leq t \\leq T\\}\\) to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "train a lightweight autoencoder. An encoder \\(E\\) maps the \\(D\\)-dimensional CLIP features \\(\\pmb{L}_t^l (v)\\in \\mathbb{R}^D\\) to \\(\\pmb{H}_t^l (v) = E(\\pmb{L}_t^l (v))\\in \\mathbb{R}^d\\), where \\(d\\ll D\\). Then we learn a decoder \\(\\Psi\\) to reconstruct the original CLIP embeddings from the compressed representation. The autoencoder is trained with a reconstruction objective on the CLIP embeddings \\(\\{\\pmb{L}_t^l\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.193, + 0.892, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a e} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {a e} \\left(\\Psi \\left(E \\left(\\boldsymbol {L} _ {t} ^ {l} (v)\\right)\\right), \\boldsymbol {L} _ {t} ^ {l} (v)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.246, + 0.891, + 0.276 + ], + "angle": 0, + "content": "where \\(d_{ae}()\\) denotes a distance function used for the autoencoder. Here we adopt both \\(\\mathcal{L}_1\\) and a cosine distance loss." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.277, + 0.892, + 0.429 + ], + "angle": 0, + "content": "After training the autoencoder, we transform all CLIP embeddings \\(\\{L_t^l\\}\\) into scene-specific latent features \\(\\{H_t^l\\}\\). We let our 3D language Gaussians learn language embeddings in the scene-specific latent space instead of the CLIP latent space. Therefore, we have \\(f^l \\in \\mathbb{R}^d\\). In practice, we choose \\(d = 3\\) as it yields excellent model efficiency and accuracy. Compared to directly modeling the \\(D\\)-dimensional CLIP embeddings, our method significantly reduced the memory cost by incorporating scene priors. We optimized the language embeddings with the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.438, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l a n g} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {l a n g} \\left(\\boldsymbol {F} _ {t} ^ {l} (v), \\boldsymbol {H} _ {t} ^ {l} (v)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.891, + 0.521 + ], + "angle": 0, + "content": "where \\(d_{lang}()\\) denotes the distance function used for our 3D Language Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.522, + 0.892, + 0.597 + ], + "angle": 0, + "content": "During inference, we follow the Eq. (4) to render the language embeddings from 3D to 2D, and then we use the trained scene-specific decoder \\(\\Psi\\) to recover the CLIP image embeddings \\(\\Psi (\\pmb {F}_t^l)\\in \\mathbb{R}^{D\\times H\\times W}\\), which enable open-vocabulary queries with the CLIP text encoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.703 + ], + "angle": 0, + "content": "By enhancing 3D Gaussians with language embedding and employing a scene-wise language autoencoder, our proposed LangSplat presents a powerful and efficient solution for building 3D language fields. This approach not only preserves the rendering efficiency of Gaussian Splatting but also mitigates the catastrophic memory explosion associated with explicit modeling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.748, + 0.729 + ], + "angle": 0, + "content": "3.4. Open-vocabulary Querying" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Due to the well-aligned latent space between images and text provided by the CLIP model, our learned 3D language field can easily support open-vocabulary 3D queries, including open-vocabulary 3D object localization and open-vocabulary 3D semantic segmentation. Many existing open-vocabulary 3D semantic segmentation methods [24] usually select the category from a category list, which includes the categories present in the images. However, obtaining a comprehensive category list for in-the-wild scenes is challenging. Different from them, our method generates precise object masks given an arbitrary text query." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20055" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.34 + ], + "angle": 0, + "content": "Following LERF [18], we compute the relevancy score for each text query. Specifically, for each rendered language embedding \\(\\phi_{img}\\) and each text query \\(\\phi_{qry}\\), the relevancy score is defined as \\(\\min_i\\frac{\\exp(\\phi_{img} \\cdot \\phi_{qry})}{\\exp(\\phi_{img} \\cdot \\phi_{qry}) + \\exp(\\phi_{img} \\cdot \\phi_{canon}^i)}\\), where \\(\\phi_{canon}^i\\) is the CLIP embeddings of a predefined canonical phrase chosen from \"object\", \"things\", \"stuff\", and \"texture\". Hence, for each text query, we will obtain three relevancy maps, each representing results at a specific semantic level. We follow the strategy used in LERF [18] and choose the semantic level that yields the highest relevancy score. For the 3D object localization task, we directly choose the point with the highest relevance score. For the 3D semantic segmentation task, we filter out points with relevancy scores lower than a chosen threshold, and predict the object masks with remaining regions. Please refer to the appendix for additional details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.354, + 0.21, + 0.372 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.379, + 0.176, + 0.395 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.473, + 0.735 + ], + "angle": 0, + "content": "Datasets. We employ two datasets for evaluation. The LERF dataset [18] is captured using the iPhone App Polycam, which consists of complex in-the-wild scenes. The LERF dataset is designed for 3D object localization tasks, here we extend the LERF dataset by annotating ground truth masks for textual queries, enabling the evaluation of the open-vocabulary 3D semantic segmentation on the LERF dataset. As the original LERF annotations for 3D object localization are relatively simple, performance in some scenarios has already approached saturation. Therefore, we further manually annotated additional challenging localization samples to better evaluate method performance. We report localization accuracy for the 3D object localization task following LERF [18], and report the IoU results for the 3D semantic segmentation task. We also employ the 3D-OVS dataset [24], which comprises a collection of long-tail objects captured in diverse poses and backgrounds. This dataset is developed for open-vocabulary 3D semantic segmentation, where a full list of categories is provided. While other methods use the full list to generate the predicted masks, we only use the query category to generate the corresponding masks. The mIoU metric is used for this dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.736, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. To extract the language features of each image, we utilize the OpenCLIP ViT-B/16 model. For SAM, we use the ViT-H model to segment 2D masks. For each scene, we first use 3D Gaussian Splatting to train an RGB scene. We train it for 30,000 iterations, and in the end, each scene comprises around 2,500,000 points. We follow the default parameter setting as in [17] to train the RGB scene. Then we train our 3D language Gaussians by fixing all other parameters of 3D Gaussians such as mean and opacity. Only the language features are learnable during this stage. We train the language features for 30,000" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.089, + 0.888, + 0.203 + ], + "angle": 0, + "content": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen14.162.073.2
figurines8.975.080.4
teatime33.984.888.1
waldo_kitchen27.372.795.5
overall21.173.684.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.213, + 0.892, + 0.227 + ], + "angle": 0, + "content": "Table 1. Localization accuracy (%) comparisons on LERF dataset." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.236, + 0.888, + 0.351 + ], + "angle": 0, + "content": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen7.028.251.2
figurines7.638.644.7
teatime21.745.065.1
waldo_kitchen29.937.944.5
overall16.637.451.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.389 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of 3D semantic segmentation on the LERF dataset. We report the average IoU scores \\((\\%)\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.411, + 0.892, + 0.487 + ], + "angle": 0, + "content": "iterations. Our autoencoder is implemented by an MLP, which compresses the 512-dimensional CLIP features into 3-dimensional latent features. For a \\(1440 \\times 1080\\) resolution scene, our model is trained for \\(\\sim 25\\) minutes on an NVIDIA RTX-3090 GPU and takes roughly 4GB of memory." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.757, + 0.513 + ], + "angle": 0, + "content": "4.2. Results on the LERF dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Quantitative Results. We first compare our method with other methods on the LERF dataset. Table 1 shows the localization results. We observe that our method achieves an overall accuracy of \\(84.3\\%\\), significantly outperforming LERF. Table 2 further shows the IoU results of 3D semantic segmentation, our method outperforms LERF by \\(14.0\\%\\), which illustrates the superiority of our proposed LangSplat. Visualization Results. To show the learned 3D language field, we visualize the learned features by computing 3-dimensional PCA components of learned language features following [20]. The results are shown in Figure 1. We see that the LERF learned features fail to generate clear boundaries between objects while our method gives precise object shapes solely using CLIP features. We further show the visualization results of object localization and semantic segmentation in Figure 3 and Figure 4, respectively. We observe that the activation regions generated by LERF are more dispersed, while ours are more concentrated, and our activation regions can better align with the ground truth shape compared to those produced by LERF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Ablation Study. We conduct ablations on the ramen scene and report the semantic segmentation results in Table 3. We test the query speed on an NVIDIA RTX-3090 GPU. Here AE represents autoencoder and 3D-GS denotes 3D Gaussian Splatting. Without any of our proposed components," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20056" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.212, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.169, + 0.162, + 0.176 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.096, + 0.34, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.169, + 0.29, + 0.177 + ], + "angle": 0, + "content": "LSeg" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.178, + 0.212, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.25, + 0.163, + 0.257 + ], + "angle": 0, + "content": "LERF" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.178, + 0.34, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.25, + 0.291, + 0.257 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.264, + 0.246, + 0.278 + ], + "angle": 0, + "content": "'spoon'" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.095, + 0.482, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.169, + 0.433, + 0.176 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.096, + 0.612, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.169, + 0.561, + 0.177 + ], + "angle": 0, + "content": "LSeg" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.178, + 0.482, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.25, + 0.434, + 0.257 + ], + "angle": 0, + "content": "LERF" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.178, + 0.612, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.25, + 0.561, + 0.257 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.264, + 0.537, + 0.278 + ], + "angle": 0, + "content": "'paper napkin'" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.095, + 0.754, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.169, + 0.704, + 0.176 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.096, + 0.882, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.169, + 0.832, + 0.177 + ], + "angle": 0, + "content": "LSeg" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.178, + 0.753, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.25, + 0.705, + 0.257 + ], + "angle": 0, + "content": "LERF" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.178, + 0.882, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.25, + 0.832, + 0.257 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.264, + 0.788, + 0.278 + ], + "angle": 0, + "content": "'spatula'" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.289, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons of open-vocabulary 3D object localization on the LERF dataset. The red points are the model predictions and the black dashed bounding boxes denote the annotations." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.326, + 0.891, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.468, + 0.788, + 0.483 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons of open-vocabulary 3D semantic segmentation on the LERF dataset." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.491, + 0.458, + 0.611 + ], + "angle": 0, + "content": "
ComponentPerformance
AE3D-GSSAMIoU (%)Speed (s/q)
✓28.2030.93
46.747.77
✓✓OOMOOM
✓✓✓51.150.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.621, + 0.465, + 0.635 + ], + "angle": 0, + "content": "Table 3. Ablations. The results are obtained on the ramen scene." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.645, + 0.464, + 0.764 + ], + "angle": 0, + "content": "
ComponentPerformance
AE3D-GSSAMmIoU (%)Speed (s/q)
✓53.255.7
85.718.4
✓✓OOMOOM
✓✓✓94.20.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.775, + 0.468, + 0.803 + ], + "angle": 0, + "content": "Table 4. Ablations result on the bench scene of the 3D-OVS dataset. The image resolution is \\(1440 \\times 1080\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.903 + ], + "angle": 0, + "content": "our baseline equals LERF, which has a speed of 30.93 seconds per text query at the resolution of \\(988 \\times 731\\). Using SAM to replace the scale-based solution significantly increases the IoU by \\(18.54\\%\\), showing our SAM-based solution effectively addresses the point ambiguity issue, leading" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.614 + ], + "angle": 0, + "content": "to accurate 3D language features. Simply replacing NeRF with 3D Gaussian Splatting leads to running out of L1 cache memory as explicitly modeling CLIP features poses huge memory demands. Incorporating a scene-specific autoencoder effectively addresses this issue and results in further improvements in both accuracy and efficiency. In the end, our LangSplat achieved a \\(119 \\times\\) speedup over LERF while significantly surpassing LERF in terms of accuracy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.893, + 0.783 + ], + "angle": 0, + "content": "We further conducted the ablations on the 3D-OVS dataset, which has a higher image resolution of \\(1440 \\times 1080\\). Table 4 lists the results on the bench scene. We also tested the query speed on an NVIDIA RTX-3090 GPU. We observed that with the increase in image resolution, the speedup over LERF further improved to \\(199 \\times\\), which demonstrates the huge potential of our method. Our further study shows that most of the computational time is allocated to the decoder rather than the rendering process. We could replace the decoder with a \\(1 \\times 1\\) convolutional layer to attain a higher speedup." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.8, + 0.771, + 0.815 + ], + "angle": 0, + "content": "4.3. Results on the 3D-OVS dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Quantitative Results. We compare LangSplat with other 2D and 3D state-of-the-art methods on the 3D-OVS dataset in Table 5. We observe that LangSplat not only outperforms 2D-based methods such as ODISE [46] and OV-Seg [23], but also achieves better results than 3D-based methods in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20057" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.887, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.5, + 0.893, + 0.53 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of different methods on the 3D-OVS dataset. We visualize the segmentation results in 2 scenes. We observe that our method gives the most accurate segmentation maps." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.535, + 0.47, + 0.686 + ], + "angle": 0, + "content": "
Methodbedbenchroomsofalawnoverall
LSeg [21]56.06.019.24.517.520.6
ODISE [46]52.624.152.548.339.843.5
OV-Seg [23]79.888.971.466.181.277.5
FFD [20]56.66.125.13.742.926.9
LERF [18]73.553.246.62773.754.8
3D-OVS [24]89.589.392.87488.286.8
LangSplat92.594.294.190.096.193.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.696, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Table 5. Quantitative comparisons of 3D semantic segmentation on the 3D-OVS dataset. We report the mIoU scores \\((\\%)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.825 + ], + "angle": 0, + "content": "cluding LERF [18] and 3D-OVS [24] by a large margin. Note that in this dataset, we generate object masks only based on the query text while others, such as 3D-OVS, require the complete category list. In the end, our method achieves an overall mIoU of \\(93.4\\%\\), which shows that LangSplat effectively learns a precise 3D language field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. We present the qualitative results in Figure 5. As LERF suffers from the patchy issue and learns over-smoothed features, it fails to find accurate object boundaries. Among all state-of-the-art methods, our methods give the most accurate segmentation maps, which" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.538, + 0.871, + 0.555 + ], + "angle": 0, + "content": "further demonstrates the effectiveness of our LangSplat." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.62, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.794 + ], + "angle": 0, + "content": "In this paper, we have presented LangSplat, a method for constructing 3D language fields that enables precise and efficient open-vocabulary querying within 3D spaces. By extending 3D Gaussian Splatting with language features and learning a scene-specific language autoencoder, LangSplat circumvents slow rendering speed associated with NeRF-based methods. Furthermore, we propose to learn the semantic hierarchy defined by SAM, which effectively resolves the point ambiguity problem. The experimental results clearly demonstrate LangSplat's superiority over existing state-of-the-art methods like LERF, particularly in terms of its remarkable \\(199 \\times\\) speed improvement and enhanced performance in open-ended 3D language query tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was funded through National Key Research and Development Program of China (Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (JCYJ20220818101001004, JSGG20210802153150005), and in part by NIH grant R01HD104969, 1U01CA284207, and NSF grant CRCNS-2309041." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20058" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.213 + ], + "angle": 0, + "content": "[2] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.468, + 0.243 + ], + "angle": 0, + "content": "[3] Nancy Bonvillain. Language, culture, and communication: The meaning of messages. Rowman & Littlefield, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.468, + 0.3 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[5] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, pages 5056-5066, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[6] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCVn, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[9] Boyuan Chen, Fei Xia, Brian Ichter, Kanishka Rao, Keerthana Gopalakrishnan, Michael S Ryoo, Austin Stone, and Daniel Kappler. Open-vocabulary queryable scene representations for real world planning. In ICRA, pages 11509-11522. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.468, + 0.613 + ], + "angle": 0, + "content": "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[11] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[12] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, MingMing Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In ACM MM, Demo track, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.468, + 0.771 + ], + "angle": 0, + "content": "[13] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In CVPR, pages 4089-4098, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[14] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In ICRA, pages 10608-10615. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[15] Siyuan Huang, Zhengkai Jiang, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Instruct2act: Mapping multimodality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[16] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):1-14, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. NeurlPS, 35:23311-23330, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[22] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, 35:35313-35325, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, pages 7061-7070, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[25] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[26] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.703, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[28] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[29] Gengchen Mai, Weiming Huang, Jin Sun, Suhang Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, et al. On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[30] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20059" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[32] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[33] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[35] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. Clip-fields: Weakly supervised semantic fields for robotic memory. arXiv preprint arXiv:2210.05663, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[36] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[37] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In ECCV, pages 666-682. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[38] Shuai Shen, Wanhua Li, Xiaoke Huang, Zheng Zhu, Jie Zhou, and Jiwen Lu. Sd-nerf: Towards lifelike talking head animation via spatially-adaptive dual-driven nerfs. TMM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[39] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. arXiv preprint arXiv:2308.07931, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[40] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In CVPR, pages 9043–9052, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 3DV, pages 443-453. IEEE, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[44] Jia Wan, Wanhua Li, Atmadeep Banerjee, Jason Ken Adhinarta, Evelina Sjostedt, Jingpeng Wu, Jeff Lichtman, Hanspeter Pfister, and Donglai Wei. Trisam: Tri-plane sam for zero-shot cortical blood vessel segmentation in vem images. arXiv preprint arXiv:2401.13961, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[45] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[46] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panop" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[48] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[50] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[54] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, pages 15838-15847, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.561, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.604, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[56] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In VIS, pages 29-538. IEEE, 2001. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20060" + } + ] +] \ No newline at end of file diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_origin.pdf b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..401ef82df663ffe3090efc47e49f4a3627425fa0 --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/591ba264-2503-43cc-8f70-dc10b9821047_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56302bcfd99dc62b4f76e4d40aabbd93e78d59cfc32f76e138caff99cc164b0a +size 4704866 diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/full.md b/2024/LangSplat_ 3D Language Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c077ad2831cbf0a23ea78a15cd936e41628b5af --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/full.md @@ -0,0 +1,305 @@ +# LangSplat: 3D Language Gaussian Splatting + +Minghan Qin $^{1,*}$ , Wanhua Li $^{2,*, \boxtimes}$ , Jiawei Zhou $^{1,*}$ , Haoqian Wang $^{1, \boxtimes}$ , Hanspeter Pfister $^{2}$ , Tsinghua University $^{2}$ Harvard University + +qmh21@mails.tsinghua.edu.cn, wanhua@seas.harvard.edu, zhoujw22@mails.tsinghua.edu.cn + +wanghaoqian@tsinghua.edu.cn, pfister@seas.harvard.edu + +![](images/21fa475a5295e806312cea614041d6ce79c4aa276fbce9c80a008a9c8c0dc2f4.jpg) +Figure 1. Visualization of learned 3D language features of the previous SOTA method LERF and our LangSplat. While LERF generates imprecise and vague 3D features, our LangSplat accurately captures object boundaries and provides precise 3D language fields. While being effective, our LangSplat is also $199 \times$ faster than LERF at the resolution of $1440 \times 1080$ . + +![](images/eebeafe6fee027ca32321aee7ad3bf6256d000592a8899ead1a4bf28170911ac.jpg) + +![](images/b42dcdfeb3e4b5023664ed51b30238f0a4b497487723d096fe8aeed6a18e0e09.jpg) + +# Abstract + +Humans live in a 3D world and commonly use natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space, + +thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experimental results show that LangSplat significantly outperforms the previous state-of-the-art method LERF by a large margin. Notably, LangSplat is extremely efficient, achieving a $199 \times$ speedup compared to LERF at the resolution of $1440 \times 1080$ . We strongly recommend readers to check out our video results at https://langsplat.github.io/. + +# 1. Introduction + +Language is the primary means of communication for human beings [3]. Modeling a 3D language field allows users + +to interact with and query 3D worlds using open-ended language, which presents a promising avenue for human-computer interaction and understanding [1, 5, 13]. The field of open-ended language queries in 3D has attracted increasing attention due to its various applications such as robotic navigation [14] and manipulation [39], 3D semantic understanding [9, 54] and editing [20], autonomous driving [16], and augmented/virtual reality [24]. + +Due to the absence of large-scale and diverse 3D scene data with language annotations, the current prevailing approach like LERF [18] involves feature distillation from off-the-shelf vision-language models such as CLIP into a 3D scene. However, these methods [18, 24] suffer from significant limitations in both speed and accuracy, severely constraining their practical applicability. To address these two issues, we revisit two key aspects of 3D language field modeling: the 3D modeling approach that bridges the gap between 2D and 3D, and the rendering target, which determines what to learn for a 3D point. For the 3D modeling technology, most methods utilize neural radiance fields (NeRFs) to represent a 3D scene, where volume rendering techniques are employed to accumulate 3D points along a ray into a single pixel. While NeRF has been demonstrated for its powerful 3D representation capabilities [7, 32, 33, 37, 38, 52], the volume rendering nature leads to its computationally expensive rendering speed [8, 31, 41], which imposes notable constraints on the potential applications within the NeRF-based language field. + +Regarding the rendering target, learning a CLIP embedding for a 3D point could be ambiguous as CLIP embeddings are aligned with images rather than pixels. Employing CLIP embeddings from a cropped patch also raises the point ambiguity issue, as the same 3D position can be associated with semantic concepts of varying scales. For instance, a point located on a bear's nose should yield high response values for three distinct textual queries: "bear's nose", "bear's head", and "bear" given that this point contributes to all three hierarchical regions. To deal with this issue, current methods [18, 24] introduce an additional absolute scale input to NeRF, trains with patch-wise CLIP features at different scales, and densely renders 2D maps at multiple scales during querying to select the optimal one. However, this scale-based solution compromises both efficiency and effectiveness. It could increase query time by up to 30 times as it needs to render at multiple different scales. Moreover, most patches with varying scales often fail to accurately encompass objects, either frequently including other objects from the background or omitting portions of the target object. These inaccurate CLIP features lead to the trained 3D language field lacking clear boundaries and containing a significant amount of noise. Therefore, they often simultaneously learn pixel-aligned DINO features to mitigate this issue. However, the performance remains un + +satisfactory. As shown in Figure 1, LERF still generates imprecise 3D language features. + +In this paper, we propose the 3D Language Gaussian Splatting (LangSplat) to address the above issues. Instead of using NeRF to build 3D representations, we resort to 3D Gaussian Splatting, which represents a 3D scene as a collection of 3D Gaussians and uses tile-based splatting to achieve efficient rendering at high resolutions. Our LangSplat defines a set of 3D language Gaussians, with each Gaussian being enhanced by a language embedding. These language-enhanced Gaussians are supervised using CLIP embeddings extracted from image patches captured from multiple training views, ensuring multi-view consistency. As an explicit modeling method, directly storing the high-dimensional language embeddings for each 3D language Gaussian is memory-inefficient. To reduce the memory cost and further improve the rendering efficiency, we propose to first learn a scene-wise language autoencoder, which maps CLIP embeddings in a scene to a low-dimensional latent space. In this way, each language Gaussian only contains the low-dimensional latent language features and the final language embeddings are obtained with decoding of the rendered features. To address the point ambiguity issue, we propose to employ the semantic hierarchy defined by the Segment Anything Model (SAM) [19]. Specifically, for each 2D image, we obtain three well-segmented maps at different semantic levels with SAM. Then we extract the CLIP feature for each mask with precise object boundaries and assign this feature to every point on the corresponding mask. Learning with SAM-based masks not only endows each point with precise CLIP embeddings, resulting in higher model accuracy, but also enables direct querying at predefined three semantic scales. This circumvents the need for intensive searches across multiple absolute scales and the auxiliary DINO features, thereby effectively improving efficiency. We summarize the contributions of this paper as follows: + +- We propose the LangSplat, which is the first 3D Gaussian Splitting-based method for 3D language fields. A scene-specific autoencoder is further introduced to alleviate the memory cost issue imposed by explicit modeling. +- We propose to learn the hierarchical semantics defined by SAM to address the point ambiguity issue for 3D language field modeling. +- Experimental results show that our method outperforms the state-of-the-art methods on open-vocabulary 3D object localization and semantic segmentation tasks while being $199 \times$ faster than LERF at $1440 \times 1080$ resolution. + +# 2. Related Work + +3D Gaussian Splatting. Real-time rendering has always been a pursued objective for neural rendering. Recently, Kerbl et al. [17] proposed to represent the 3D scene with a set of 3D Gaussians, which attained real-time rendering for + +1080p resolution while maintaining state-of-the-art visual quality. Encouraged by the success of 3D Gaussian Splatting, many works extend it to other tasks. To achieve real-time dynamic scene rendering, some studies [27, 45, 49] have extended the 3D Gaussian Splatting technique to dynamic scenes. Luiten et al. [27] proposed the Dynamic 3D Gaussians, which extended 3D Gaussians to dynamic scenes by explicitly modeling the 3D Gaussians across different time steps. Yang et al. [48] presented a deformable 3D Gaussians Splatting method, which learned 3D Gaussians in canonical space and modeled the dynamic scenes with a deformation field. Meanwhile, some researchers have combined 3D Gaussian Splatting with diffusion models to achieve efficient text-to-3D generation [10, 42, 51]. For example, Tang et al. [42] introduced DreamGaussian for efficient 3D content generation with a generative 3D Gaussian Splatting model. Unlike these methods, our paper extends each 3D Gaussian with language embeddings for open-vocabulary 3D queries. + +SAM. The Segment Anything Model [19], released by Meta in 2023, has attracted considerable attention [15, 28, 29], which achieved impressive zero-shot performance. It has become the foundational model for image segmentation. SAM has been used for many vision tasks [26, 30, 44, 50] such as image inpainting [53], object tracking [11, 47], image editing [12], and so on. Many efforts have also been made to utilize SAM in the 3D domain [6, 36]. Liu et al. [25] proposed Seal to explore the potential of VFMs including SAM for point cloud segmentation. SA3D [6] generalized SAM to 3D objects by leveraging NeRF to connect 2D images and 3D space. Different from these works, we use SAM to obtain accurate object masks with three well-defined hierarchical semantics to train a 3D language field. + +3D Langue fields. Some early attempts to construct 3D feature fields included Distilled Feature Fields [20] and Neural Feature Fusion Fields [43]. They learned 3D-consistent features by distilling LSeg [21] or DINO [4] features across multiple views into a NeRF. Shen et al. [39] further adopted distilled feature fields for few-shot language-guided robotic manipulation by distilling CLIP feature into a NeRF. There are also some efforts [40, 54] that embed semantic information into NeRFs. For example, Semantic NeRF [54] jointly encoded semantics with appearance and geometry within a NeRF for novel semantic view synthesis. LERF [18] was the first to embed CLIP features into NeRF, enabling open-vocabulary 3D queries leveraging the powerful CLIP representation. DINO features were also used for supervising LERF to improve its performance. Liu et al. [24] also utilized CLIP and DINO features to train a NeRF model for 3D open-vocabulary segmentation. While these methods use NeRF for 3D modeling and suffer from the costly rendering process, we propose the 3D language Gaussian Splatting to obtain efficient 3D language fields. + +# 3. Proposed Approach + +In this section, we first revisit the challenges of modeling 3D language fields and then elaborate on how our proposed LangSplat addresses these issues. Figure 2 depicts the framework of our proposed LangSplat. + +# 3.1. Revisiting the Challenges of Language Fields + +We denote an input image as $I \in \mathbb{R}^{3 \times H \times W}$ , where $H$ and $W$ represent the height and weight of the image size. We take a set of calibrated images $\{I_t | t = 1, 2, \ldots, T\}$ as input and train a 3D language field $\Phi$ with these images. Most existing methods [18, 24, 35] employ the CLIP image encoder $V$ to extract image features and utilize the extracted CLIP embeddings to supervise the 3D language field $\Phi$ , leveraging the well-aligned text-image latent space [22, 34, 55] provided by CLIP, thus facilitating open-vocabulary queries. However, CLIP embeddings are image-aligned rather than pixel-aligned. In other words, simply computing $V(I_t) \in \mathbb{R}^D$ only obtains an image-level feature, whereas what we need is a pixel-aligned language embedding $L_t \in \mathbb{R}^{D \times H \times W}$ , where $D$ represents the CLIP feature dimension. Meanwhile, modeling pixel-aligned language features faces the issue of point ambiguity, as a single point on an object contributes to multiple semantic levels of regions. For instance, a point on a cat's ear simultaneously contributes to the cat's ear, the cat's head, and the entire cat, and should be activated by all three types of textual queries. + +To address these issues, most existing methods [18, 24] extract a hierarchy of CLIP features from cropped image patches. Specifically, for a pixel with coordinates $v \in \{1, \dots, H\} \times \{1, \dots, W\}$ , the corresponding CLIP features are obtained from image patches centered around $v$ at different absolute physical scales $s$ , with the expectation that at a certain scale $s$ , the patch can fully encompass the object. However, this multi-scale approach has two limitations. Firstly, patch features are imprecise because they often include additional contextual object information, leading to overly smoothed language fields with indistinct object boundaries. To alleviate the patchy issue, most methods [18, 24] leverage additional pixel-aligned DINO features to supervise the network. However, the learned 3D language features are still imprecise, as illustrated in Figure 1. Secondly, it requires simultaneous rendering at multiple scales during inference to find the optimal scale. With the number of scales $s$ potentially reaching as high as 30 [18], this significantly diminishes the inference speed. + +Besides the rendering target, another key design space is the 3D modeling approach. Most existing methods [2, 43] employ NeRFs for 3D representation, where they learn a language feature at each 3D point and subsequently render the language feature onto an image, similar to color rendering. However, NeRF-based methods are constrained by their time-consuming rendering process, even though the + +![](images/6044390a5174794305df198a967eb2f351851440a5dd54a91c6bbe3e252316cd.jpg) +Figure 2. The framework of our LangSplat. Our LangSplat leverages SAM to learn hierarchical semantics to address the point ambiguity issue. Then segment masks are sent to the CLIP image encoder to extract the corresponding CLIP embeddings. We learn an autoencoder with these obtained CLIP embeddings. Our 3D language Gaussian learn language features on the scene-specific latent space to reduce the memory cost. During querying, the rendered language embeddings are sent to the decoder to recover the features on the CLIP space. + +most advanced NeRF techniques currently available cannot achieve real-time rendering in high-resolution, unrestricted scenes [17]. Meanwhile, there is a high demand for efficient open vocabulary querying in practical applications, especially in fields such as intelligent robotics. + +# 3.2. Learning Hierarchical Semantics with SAM + +As a foundation model for image segmentation, SAM [19] can accurately group a pixel with its surrounding pixels belonging to the same object, thereby segmenting the image into many object masks with clear boundaries. Furthermore, SAM addresses point ambiguity by generating three different masks for a point prompt, namely, whole, part, and subpart, representing three hierarchical levels of semantics. In this paper, we propose leveraging SAM to obtain precise object masks, which are then used to acquire pixel-aligned features. We also explicitly model the semantic hierarchy defined by SAM to address the point ambiguity issue. With SAM, we can capture the semantic hierarchy of objects in 3D scenes, providing accurate and multi-scale segmentation maps for each input image. + +Specifically, we feed a regular grid of $32 \times 32$ point prompts into SAM to obtain the masks under three different semantic levels: $M_0^s, M_0^p, M_0^w$ , where $M_0^s, M_0^p$ , and $M_0^w$ represent the masks at subpart, part, and whole levels, respectively. Then we remove redundant masks for each of the three mask sets based on the predicted IoU score, stability score, and overlap rate between masks. Each filtered mask set independently performs a comprehensive full-image segmentation based on its respective semantic level, resulting in three segmentation maps: $M^s, M^p, M^w$ . These segmentation maps precisely delineate the boundaries of objects at their hierarchical levels, effectively partitioning the scene into semantically meaningful regions. With the obtained segmentation maps, we proceed to extract CLIP features for each segmented region. These features capture the semantic context of the objects + +at various levels within the scene. Mathematically, the obtained pixel-aligned language embeddings are: + +$$ +\boldsymbol {L} _ {t} ^ {l} (v) = V \left(\boldsymbol {I} _ {t} \odot \boldsymbol {M} ^ {l} (v)\right), l \in \{s, p, w \}, \tag {1} +$$ + +where $M^l(v)$ represents the mask region to which pixel $v$ belongs at the semantic level $l$ . + +Each pixel rendered from the 3D language scene now possesses a CLIP feature that aligns with its precise semantic context. This alignment reduces ambiguity and enhances the accuracy of language-based queries. We can learn an accurate 3D language field even without the commonly used DINO regularization. Another advantage of our SAM-based approach is the predefined semantic scales. Since we have distinct segmentation maps for "whole," "part," and "subpart" levels, we can directly query the 3D language field at these predefined scales. This eliminates the need for intensive searches across multiple absolute scales, making the querying process more efficient. By incorporating SAM's semantic hierarchy into our approach, we not only improve the accuracy of our 3D language field but also streamline the querying process, making it more efficient and effective for a wide range of applications. + +# 3.3. 3D Gaussian Splitting for Language Fields + +Having obtained the language embeddings on a set of 2D images $\{L_t^l, |t = 1, \dots, T\}$ , we can learn a 3D language scene by modeling the relations between 3D points and 2D pixels. Most existing methods [2, 43] suffer from the costly rendering process as they adopt NeRFs for 3D modeling. To address this issue, we present the first 3D Gaussian Splitting-based method for 3D language field modeling. + +3D Gaussian Splitting explicitly represents a 3D scene as a collection of anisotropic 3D Gaussians, with each Gaussian $G(x)$ characterized by a mean $\mu \in \mathbb{R}^3$ and a covariance matrix $\Sigma$ : + +$$ +G (x) = \exp \left(- \frac {1}{2} (x - \mu) ^ {\top} \Sigma^ {- 1} (x - \mu)\right). \tag {2} +$$ + +To optimize the parameters of 3D Gaussians, they are rendered into 2D image planes [56], and a tile-based rasterizer is used to improve the rendering efficiency: + +$$ +C (v) = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {3} +$$ + +where $c_{i}$ is the color of the $i$ -th Gaussian, $\mathcal{N}$ denotes the Gaussians in the tile, $C(v)$ is the rendered color at pixel $v$ , and $\alpha_{i} = o_{i}G_{i}^{2D}(v)$ . Here $o_{i}$ is the opacity of the $i$ -th Gaussian and $G_{i}^{2D}(\cdot)$ represents the function of the $i$ -th Gaussian projected onto 2D. + +In this paper, we proposes the 3D language Gaussian Splatting, which augments each 3D Gaussian with three language embeddings $\{\pmb{f}^s,\pmb{f}^p,\pmb{f}^w\}$ . These embeddings are derived from CLIP features, which capture the hierarchical semantics provided by SAM. The augmented Gaussians are named as 3D language Gaussians. We also adopt the tile-based rasterizer to retain the rendering efficiency: + +$$ +\boldsymbol {F} ^ {l} (v) = \sum_ {i \in \mathcal {N}} \boldsymbol {f} _ {i} ^ {l} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), l \in \{s, p, w \}, \tag {4} +$$ + +where $F^l(v)$ represents the language embedding rendered at pixel $v$ with the semantic level $l$ . By incorporating language information directly into the Gaussians, we enable the 3D language field to respond to language-based queries. + +As an explicit modeling approach, our LangSplat may create millions of 3D points to model a complex 3D scene. As CLIP embeddings are high-dimensional features, directly learning $\pmb{f}^{l}$ on the CLIP latent space significantly increases memory and time cost. Compared to learning RGB colors without spherical harmonics coefficients, learning 512-dimensional CLIP features increases the memory requirements for storing 3D Gaussians by over 35 times, easily leading to running out of L1 cache memory. To reduce memory cost and improve efficiency, we introduce a scenewise language autoencoder. This autoencoder maps CLIP embeddings in a scene to a lower-dimensional latent space, reducing memory requirements. The CLIP model is trained using 400 million (image, text) pairs, and its $D$ -dimensional latent space could be highly compact, as it needs to align arbitrary text and images in this space. However, the language field $\Phi$ we train here is scene-specific, meaning we can leverage scene priors to compress CLIP features. In fact, for each input image, we will obtain hundreds of masks segmented by SAM, which is significantly smaller than the number of images used in CLIP training. Therefore, all the segmented regions in a scene are sparsely distributed in the CLIP latent space, allowing us to further compress these CLIP features using a scene-specific autoencoder. + +Specifically, we use the collections of CLIP features of SAM segmented masks $\{L_t^l | l \in \{s, p, w\}, 1 \leq t \leq T\}$ to + +train a lightweight autoencoder. An encoder $E$ maps the $D$ -dimensional CLIP features $\pmb{L}_t^l (v)\in \mathbb{R}^D$ to $\pmb{H}_t^l (v) = E(\pmb{L}_t^l (v))\in \mathbb{R}^d$ , where $d\ll D$ . Then we learn a decoder $\Psi$ to reconstruct the original CLIP embeddings from the compressed representation. The autoencoder is trained with a reconstruction objective on the CLIP embeddings $\{\pmb{L}_t^l\}$ : + +$$ +\mathcal {L} _ {a e} = \sum_ {l \in \{s, p, w \}} \sum_ {t = 1} ^ {T} d _ {a e} \left(\Psi \left(E \left(\boldsymbol {L} _ {t} ^ {l} (v)\right)\right), \boldsymbol {L} _ {t} ^ {l} (v)\right), \tag {5} +$$ + +where $d_{ae}()$ denotes a distance function used for the autoencoder. Here we adopt both $\mathcal{L}_1$ and a cosine distance loss. + +After training the autoencoder, we transform all CLIP embeddings $\{L_t^l\}$ into scene-specific latent features $\{H_t^l\}$ . We let our 3D language Gaussians learn language embeddings in the scene-specific latent space instead of the CLIP latent space. Therefore, we have $f^l \in \mathbb{R}^d$ . In practice, we choose $d = 3$ as it yields excellent model efficiency and accuracy. Compared to directly modeling the $D$ -dimensional CLIP embeddings, our method significantly reduced the memory cost by incorporating scene priors. We optimized the language embeddings with the objective: + +$$ +\mathcal {L} _ {l a n g} = \sum_ {l \in \{s, p, w \}} \sum_ {t = 1} ^ {T} d _ {l a n g} \left(\boldsymbol {F} _ {t} ^ {l} (v), \boldsymbol {H} _ {t} ^ {l} (v)\right), \tag {6} +$$ + +where $d_{lang}()$ denotes the distance function used for our 3D Language Gaussians. + +During inference, we follow the Eq. (4) to render the language embeddings from 3D to 2D, and then we use the trained scene-specific decoder $\Psi$ to recover the CLIP image embeddings $\Psi (\pmb {F}_t^l)\in \mathbb{R}^{D\times H\times W}$ , which enable open-vocabulary queries with the CLIP text encoder. + +By enhancing 3D Gaussians with language embedding and employing a scene-wise language autoencoder, our proposed LangSplat presents a powerful and efficient solution for building 3D language fields. This approach not only preserves the rendering efficiency of Gaussian Splatting but also mitigates the catastrophic memory explosion associated with explicit modeling. + +# 3.4. Open-vocabulary Querying + +Due to the well-aligned latent space between images and text provided by the CLIP model, our learned 3D language field can easily support open-vocabulary 3D queries, including open-vocabulary 3D object localization and open-vocabulary 3D semantic segmentation. Many existing open-vocabulary 3D semantic segmentation methods [24] usually select the category from a category list, which includes the categories present in the images. However, obtaining a comprehensive category list for in-the-wild scenes is challenging. Different from them, our method generates precise object masks given an arbitrary text query. + +Following LERF [18], we compute the relevancy score for each text query. Specifically, for each rendered language embedding $\phi_{img}$ and each text query $\phi_{qry}$ , the relevancy score is defined as $\min_i\frac{\exp(\phi_{img} \cdot \phi_{qry})}{\exp(\phi_{img} \cdot \phi_{qry}) + \exp(\phi_{img} \cdot \phi_{canon}^i)}$ , where $\phi_{canon}^i$ is the CLIP embeddings of a predefined canonical phrase chosen from "object", "things", "stuff", and "texture". Hence, for each text query, we will obtain three relevancy maps, each representing results at a specific semantic level. We follow the strategy used in LERF [18] and choose the semantic level that yields the highest relevancy score. For the 3D object localization task, we directly choose the point with the highest relevance score. For the 3D semantic segmentation task, we filter out points with relevancy scores lower than a chosen threshold, and predict the object masks with remaining regions. Please refer to the appendix for additional details. + +# 4. Experiments + +# 4.1. Settings + +Datasets. We employ two datasets for evaluation. The LERF dataset [18] is captured using the iPhone App Polycam, which consists of complex in-the-wild scenes. The LERF dataset is designed for 3D object localization tasks, here we extend the LERF dataset by annotating ground truth masks for textual queries, enabling the evaluation of the open-vocabulary 3D semantic segmentation on the LERF dataset. As the original LERF annotations for 3D object localization are relatively simple, performance in some scenarios has already approached saturation. Therefore, we further manually annotated additional challenging localization samples to better evaluate method performance. We report localization accuracy for the 3D object localization task following LERF [18], and report the IoU results for the 3D semantic segmentation task. We also employ the 3D-OVS dataset [24], which comprises a collection of long-tail objects captured in diverse poses and backgrounds. This dataset is developed for open-vocabulary 3D semantic segmentation, where a full list of categories is provided. While other methods use the full list to generate the predicted masks, we only use the query category to generate the corresponding masks. The mIoU metric is used for this dataset. + +Implementation Details. To extract the language features of each image, we utilize the OpenCLIP ViT-B/16 model. For SAM, we use the ViT-H model to segment 2D masks. For each scene, we first use 3D Gaussian Splatting to train an RGB scene. We train it for 30,000 iterations, and in the end, each scene comprises around 2,500,000 points. We follow the default parameter setting as in [17] to train the RGB scene. Then we train our 3D language Gaussians by fixing all other parameters of 3D Gaussians such as mean and opacity. Only the language features are learnable during this stage. We train the language features for 30,000 + +
Test SceneLSeg [21]LERF [18]LangSplat
ramen14.162.073.2
figurines8.975.080.4
teatime33.984.888.1
waldo_kitchen27.372.795.5
overall21.173.684.3
+ +Table 1. Localization accuracy (%) comparisons on LERF dataset. + +
Test SceneLSeg [21]LERF [18]LangSplat
ramen7.028.251.2
figurines7.638.644.7
teatime21.745.065.1
waldo_kitchen29.937.944.5
overall16.637.451.4
+ +Table 2. Quantitative comparisons of 3D semantic segmentation on the LERF dataset. We report the average IoU scores $(\%)$ + +iterations. Our autoencoder is implemented by an MLP, which compresses the 512-dimensional CLIP features into 3-dimensional latent features. For a $1440 \times 1080$ resolution scene, our model is trained for $\sim 25$ minutes on an NVIDIA RTX-3090 GPU and takes roughly 4GB of memory. + +# 4.2. Results on the LERF dataset + +Quantitative Results. We first compare our method with other methods on the LERF dataset. Table 1 shows the localization results. We observe that our method achieves an overall accuracy of $84.3\%$ , significantly outperforming LERF. Table 2 further shows the IoU results of 3D semantic segmentation, our method outperforms LERF by $14.0\%$ , which illustrates the superiority of our proposed LangSplat. Visualization Results. To show the learned 3D language field, we visualize the learned features by computing 3-dimensional PCA components of learned language features following [20]. The results are shown in Figure 1. We see that the LERF learned features fail to generate clear boundaries between objects while our method gives precise object shapes solely using CLIP features. We further show the visualization results of object localization and semantic segmentation in Figure 3 and Figure 4, respectively. We observe that the activation regions generated by LERF are more dispersed, while ours are more concentrated, and our activation regions can better align with the ground truth shape compared to those produced by LERF. + +Ablation Study. We conduct ablations on the ramen scene and report the semantic segmentation results in Table 3. We test the query speed on an NVIDIA RTX-3090 GPU. Here AE represents autoencoder and 3D-GS denotes 3D Gaussian Splatting. Without any of our proposed components, + +![](images/297a3511d890d5cf3169741f7f9387f8c7ef02c5094bef0cd55e7dc2d3ded6f1.jpg) +RGB + +![](images/0dface535705b12611324635b6e3be140a956aac20f0304b1c60f63abbc32543.jpg) + +![](images/0ca856bb07ca8822c3207a00fa5a719b68d2f9c1408989ad929f5f4a8a3739b6.jpg) +LERF + +![](images/1a15befdb603f253347b5bbe0fcf817b1e03173650902d9414e012c18c86959b.jpg) +LSeg +Ours +'spoon' + +![](images/48501b968ef74b9b4932a69cf60bcf1c69b2338bbee29f24b3ae79ebbe125f9f.jpg) +RGB + +![](images/11d888b7f4801b5aab536ef7eb2176c1b78e7f1ed7b77db58c0bdf724ee4d459.jpg) + +![](images/116feb8303a161929e797879fe807c60d476b245bc84d67db8f85f63d8ef0f0c.jpg) +LERF +'paper napkin' + +![](images/f52d64e734b1a6e886766b7ed63ec5bda073e5348b188a44a4acf861eaff662b.jpg) +LSeg +Ours + +![](images/7cbc8586c1493da12a3a04dc50d3658bb19ce1fdbf838ff73144eb40bc658710.jpg) +RGB + +![](images/ac2d94c40620224d73343a9d05a3d9f2214ee89670afd9c72515a41da401db05.jpg) + +![](images/943809e1a7975c721d13079ccf52caeaf265cff030232e288948c880f67584ac.jpg) +LERF +'spatula' + +![](images/041282c7b0a12dc3d46646cd6d5a37ace264767cf9a903de2bdf8f9fda52bb9e.jpg) +LSeg +Ours + +![](images/7f6cb00c205fbe5d62bf2238bab3476e5b59dfa7000793a15a4ff1a2a1101a3f.jpg) +Figure 3. Qualitative comparisons of open-vocabulary 3D object localization on the LERF dataset. The red points are the model predictions and the black dashed bounding boxes denote the annotations. +Figure 4. Qualitative comparisons of open-vocabulary 3D semantic segmentation on the LERF dataset. + +
ComponentPerformance
AE3D-GSSAMIoU (%)Speed (s/q)
✓28.2030.93
46.747.77
✓✓OOMOOM
✓✓✓51.150.26
+ +Table 3. Ablations. The results are obtained on the ramen scene. + +
ComponentPerformance
AE3D-GSSAMmIoU (%)Speed (s/q)
✓53.255.7
85.718.4
✓✓OOMOOM
✓✓✓94.20.28
+ +Table 4. Ablations result on the bench scene of the 3D-OVS dataset. The image resolution is $1440 \times 1080$ . + +our baseline equals LERF, which has a speed of 30.93 seconds per text query at the resolution of $988 \times 731$ . Using SAM to replace the scale-based solution significantly increases the IoU by $18.54\%$ , showing our SAM-based solution effectively addresses the point ambiguity issue, leading + +to accurate 3D language features. Simply replacing NeRF with 3D Gaussian Splatting leads to running out of L1 cache memory as explicitly modeling CLIP features poses huge memory demands. Incorporating a scene-specific autoencoder effectively addresses this issue and results in further improvements in both accuracy and efficiency. In the end, our LangSplat achieved a $119 \times$ speedup over LERF while significantly surpassing LERF in terms of accuracy. + +We further conducted the ablations on the 3D-OVS dataset, which has a higher image resolution of $1440 \times 1080$ . Table 4 lists the results on the bench scene. We also tested the query speed on an NVIDIA RTX-3090 GPU. We observed that with the increase in image resolution, the speedup over LERF further improved to $199 \times$ , which demonstrates the huge potential of our method. Our further study shows that most of the computational time is allocated to the decoder rather than the rendering process. We could replace the decoder with a $1 \times 1$ convolutional layer to attain a higher speedup. + +# 4.3. Results on the 3D-OVS dataset + +Quantitative Results. We compare LangSplat with other 2D and 3D state-of-the-art methods on the 3D-OVS dataset in Table 5. We observe that LangSplat not only outperforms 2D-based methods such as ODISE [46] and OV-Seg [23], but also achieves better results than 3D-based methods in + +![](images/296657d11ae3cf385d956d05e6d7acc4da81d368a54d8df763ff998a37a83707.jpg) +Figure 5. Qualitative comparisons of different methods on the 3D-OVS dataset. We visualize the segmentation results in 2 scenes. We observe that our method gives the most accurate segmentation maps. + +
Methodbedbenchroomsofalawnoverall
LSeg [21]56.06.019.24.517.520.6
ODISE [46]52.624.152.548.339.843.5
OV-Seg [23]79.888.971.466.181.277.5
FFD [20]56.66.125.13.742.926.9
LERF [18]73.553.246.62773.754.8
3D-OVS [24]89.589.392.87488.286.8
LangSplat92.594.294.190.096.193.4
+ +Table 5. Quantitative comparisons of 3D semantic segmentation on the 3D-OVS dataset. We report the mIoU scores $(\%)$ + +cluding LERF [18] and 3D-OVS [24] by a large margin. Note that in this dataset, we generate object masks only based on the query text while others, such as 3D-OVS, require the complete category list. In the end, our method achieves an overall mIoU of $93.4\%$ , which shows that LangSplat effectively learns a precise 3D language field. + +Qualitative Results. We present the qualitative results in Figure 5. As LERF suffers from the patchy issue and learns over-smoothed features, it fails to find accurate object boundaries. Among all state-of-the-art methods, our methods give the most accurate segmentation maps, which + +further demonstrates the effectiveness of our LangSplat. + +# 5. Conclusion + +In this paper, we have presented LangSplat, a method for constructing 3D language fields that enables precise and efficient open-vocabulary querying within 3D spaces. By extending 3D Gaussian Splatting with language features and learning a scene-specific language autoencoder, LangSplat circumvents slow rendering speed associated with NeRF-based methods. Furthermore, we propose to learn the semantic hierarchy defined by SAM, which effectively resolves the point ambiguity problem. The experimental results clearly demonstrate LangSplat's superiority over existing state-of-the-art methods like LERF, particularly in terms of its remarkable $199 \times$ speed improvement and enhanced performance in open-ended 3D language query tasks. + +Acknowledgements This research was funded through National Key Research and Development Program of China (Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (JCYJ20220818101001004, JSGG20210802153150005), and in part by NIH grant R01HD104969, 1U01CA284207, and NSF grant CRCNS-2309041. + +# References + +[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 2 +[2] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633, 2023. 3, 4 +[3] Nancy Bonvillain. Language, culture, and communication: The meaning of messages. Rowman & Littlefield, 2019. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3 +[5] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, pages 5056-5066, 2022. 2 +[6] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. 2023. 3 +[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 2 +[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCVn, pages 333-350. Springer, 2022. 2 +[9] Boyuan Chen, Fei Xia, Brian Ichter, Kanishka Rao, Keerthana Gopalakrishnan, Michael S Ryoo, Austin Stone, and Daniel Kappler. Open-vocabulary queryable scene representations for real world planning. In ICRA, pages 11509-11522. IEEE, 2023. 2 +[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 3 +[11] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3 +[12] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, MingMing Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In ACM MM, Demo track, 2023. 3 +[13] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In CVPR, pages 4089-4098, 2018. 2 +[14] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In ICRA, pages 10608-10615. IEEE, 2023. 2 +[15] Siyuan Huang, Zhengkai Jiang, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Instruct2act: Mapping multimodality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176, 2023. 3 +[16] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh + +Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):1-14, 2023. 2, 4, 6 +[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2, 3, 6, 8 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 4 +[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. NeurlPS, 35:23311-23330, 2022. 2, 3, 6, 8 +[21] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 3, 6, 8 +[22] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, 35:35313-35325, 2022. 3 +[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, pages 7061-7070, 2023. 7, 8 +[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2, 3, 5, 6, 8 +[25] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347, 2023. 3 +[26] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 3 +[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3 +[28] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 3 +[29] Gengchen Mai, Weiming Huang, Jin Sun, Suhang Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, et al. On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798, 2023. 3 +[30] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 3 +[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2 + +[32] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2 +[33] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318–10327, 2021. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3 +[35] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. Clip-fields: Weakly supervised semantic fields for robotic memory. arXiv preprint arXiv:2210.05663, 2022. 3 +[36] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3 +[37] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In ECCV, pages 666-682. Springer, 2022. 2 +[38] Shuai Shen, Wanhua Li, Xiaoke Huang, Zheng Zhu, Jie Zhou, and Jiwen Lu. Sd-nerf: Towards lifelike talking head animation via spatially-adaptive dual-driven nerfs. TMM, 2023. 2 +[39] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. arXiv preprint arXiv:2308.07931, 2023. 2, 3 +[40] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In CVPR, pages 9043–9052, 2023. 3 +[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2 +[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3 +[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 3DV, pages 443-453. IEEE, 2022. 3, 4 +[44] Jia Wan, Wanhua Li, Atmadeep Banerjee, Jason Ken Adhinarta, Evelina Sjostedt, Jingpeng Wu, Jeff Lichtman, Hanspeter Pfister, and Donglai Wei. Trisam: Tri-plane sam for zero-shot cortical blood vessel segmentation in vem images. arXiv preprint arXiv:2401.13961, 2024. 3 +[45] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3 +[46] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panop + +tic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 7, 8 +[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3 +[48] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3 +[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3 +[50] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121, 2023.3 +[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3 +[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2 +[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3 +[54] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, pages 15838-15847, 2021. 2, 3 +[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 3 +[56] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In VIS, pages 29-538. IEEE, 2001. 5 \ No newline at end of file diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/images.zip b/2024/LangSplat_ 3D Language Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b12ab3272398f454020b98edc5774ed488bfe26 --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ccdf8737ce8591ba7de0661845952b6d859828155ed68b17461cc0b472a671c +size 626238 diff --git a/2024/LangSplat_ 3D Language Gaussian Splatting/layout.json b/2024/LangSplat_ 3D Language Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..71ec10ac03e15eb324200d07bbc4eec843bea2be --- /dev/null +++ b/2024/LangSplat_ 3D Language Gaussian Splatting/layout.json @@ -0,0 +1,8867 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "spans": [ + { + "bbox": [ + 158, + 103, + 436, + 121 + ], + "type": "text", + "content": "LangSplat: 3D Language Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": "Minghan Qin" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": ", Wanhua Li" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{2,*, \\boxtimes}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": ", Jiawei Zhou" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": ", Haoqian Wang" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{1, \\boxtimes}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": ", Hanspeter Pfister" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": ", Tsinghua University" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 148, + 529, + 178 + ], + "type": "text", + "content": " Harvard University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 181, + 523, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 181, + 523, + 193 + ], + "spans": [ + { + "bbox": [ + 69, + 181, + 523, + 193 + ], + "type": "text", + "content": "qmh21@mails.tsinghua.edu.cn, wanhua@seas.harvard.edu, zhoujw22@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 195, + 440, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 195, + 440, + 206 + ], + "spans": [ + { + "bbox": [ + 151, + 195, + 440, + 206 + ], + "type": "text", + "content": "wanghaoqian@tsinghua.edu.cn, pfister@seas.harvard.edu" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 238, + 221, + 422 + ], + "blocks": [ + { + "bbox": [ + 50, + 238, + 221, + 422 + ], + "lines": [ + { + "bbox": [ + 50, + 238, + 221, + 422 + ], + "spans": [ + { + "bbox": [ + 50, + 238, + 221, + 422 + ], + "type": "image", + "image_path": "21fa475a5295e806312cea614041d6ce79c4aa276fbce9c80a008a9c8c0dc2f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "lines": [ + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "type": "text", + "content": "Figure 1. Visualization of learned 3D language features of the previous SOTA method LERF and our LangSplat. While LERF generates imprecise and vague 3D features, our LangSplat accurately captures object boundaries and provides precise 3D language fields. While being effective, our LangSplat is also " + }, + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "type": "inline_equation", + "content": "199 \\times" + }, + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "type": "text", + "content": " faster than LERF at the resolution of " + }, + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 45, + 434, + 546, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 224, + 239, + 382, + 419 + ], + "blocks": [ + { + "bbox": [ + 224, + 239, + 382, + 419 + ], + "lines": [ + { + "bbox": [ + 224, + 239, + 382, + 419 + ], + "spans": [ + { + "bbox": [ + 224, + 239, + 382, + 419 + ], + "type": "image", + "image_path": "eebeafe6fee027ca32321aee7ad3bf6256d000592a8899ead1a4bf28170911ac.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 384, + 239, + 543, + 419 + ], + "blocks": [ + { + "bbox": [ + 384, + 239, + 543, + 419 + ], + "lines": [ + { + "bbox": [ + 384, + 239, + 543, + 419 + ], + "spans": [ + { + "bbox": [ + 384, + 239, + 543, + 419 + ], + "type": "image", + "image_path": "b42dcdfeb3e4b5023664ed51b30238f0a4b497487723d096fe8aeed6a18e0e09.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "spans": [ + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 45, + 498, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 288, + 689 + ], + "type": "text", + "content": "Humans live in a 3D world and commonly use natural language to interact with a 3D scene. Modeling a 3D language field to support open-ended language queries in 3D has gained increasing attention recently. This paper introduces LangSplat, which constructs a 3D language field that enables precise and efficient open-vocabulary querying within 3D spaces. Unlike existing methods that ground CLIP language embeddings in a NeRF model, LangSplat advances the field by utilizing a collection of 3D Gaussians, each encoding language features distilled from CLIP, to represent the language field. By employing a tile-based splatting technique for rendering language features, we circumvent the costly rendering process inherent in NeRF. Instead of directly learning CLIP embeddings, LangSplat first trains a scene-wise language autoencoder and then learns language features on the scene-specific latent space," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "text", + "content": "thereby alleviating substantial memory demands imposed by explicit modeling. Existing methods struggle with imprecise and vague 3D language fields, which fail to discern clear boundaries between objects. We delve into this issue and propose to learn hierarchical semantics using SAM, thereby eliminating the need for extensively querying the language field across various scales and the regularization of DINO features. Extensive experimental results show that LangSplat significantly outperforms the previous state-of-the-art method LERF by a large margin. Notably, LangSplat is extremely efficient, achieving a " + }, + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "inline_equation", + "content": "199 \\times" + }, + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "text", + "content": " speedup compared to LERF at the resolution of " + }, + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 306, + 473, + 546, + 640 + ], + "type": "text", + "content": ". We strongly recommend readers to check out our video results at https://langsplat.github.io/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 668, + 387, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 387, + 680 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 387, + 680 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Language is the primary means of communication for human beings [3]. Modeling a 3D language field allows users" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 701, + 214, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 701, + 214, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 701, + 214, + 713 + ], + "type": "text", + "content": "* Equal contribution. " + }, + { + "bbox": [ + 61, + 701, + 214, + 713 + ], + "type": "inline_equation", + "content": "{}^{ \\ominus }" + }, + { + "bbox": [ + 61, + 701, + 214, + 713 + ], + "type": "text", + "content": " Corresponding authors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20051" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 168 + ], + "type": "text", + "content": "to interact with and query 3D worlds using open-ended language, which presents a promising avenue for human-computer interaction and understanding [1, 5, 13]. The field of open-ended language queries in 3D has attracted increasing attention due to its various applications such as robotic navigation [14] and manipulation [39], 3D semantic understanding [9, 54] and editing [20], autonomous driving [16], and augmented/virtual reality [24]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 171, + 290, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 290, + 411 + ], + "type": "text", + "content": "Due to the absence of large-scale and diverse 3D scene data with language annotations, the current prevailing approach like LERF [18] involves feature distillation from off-the-shelf vision-language models such as CLIP into a 3D scene. However, these methods [18, 24] suffer from significant limitations in both speed and accuracy, severely constraining their practical applicability. To address these two issues, we revisit two key aspects of 3D language field modeling: the 3D modeling approach that bridges the gap between 2D and 3D, and the rendering target, which determines what to learn for a 3D point. For the 3D modeling technology, most methods utilize neural radiance fields (NeRFs) to represent a 3D scene, where volume rendering techniques are employed to accumulate 3D points along a ray into a single pixel. While NeRF has been demonstrated for its powerful 3D representation capabilities [7, 32, 33, 37, 38, 52], the volume rendering nature leads to its computationally expensive rendering speed [8, 31, 41], which imposes notable constraints on the potential applications within the NeRF-based language field." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 715 + ], + "type": "text", + "content": "Regarding the rendering target, learning a CLIP embedding for a 3D point could be ambiguous as CLIP embeddings are aligned with images rather than pixels. Employing CLIP embeddings from a cropped patch also raises the point ambiguity issue, as the same 3D position can be associated with semantic concepts of varying scales. For instance, a point located on a bear's nose should yield high response values for three distinct textual queries: \"bear's nose\", \"bear's head\", and \"bear\" given that this point contributes to all three hierarchical regions. To deal with this issue, current methods [18, 24] introduce an additional absolute scale input to NeRF, trains with patch-wise CLIP features at different scales, and densely renders 2D maps at multiple scales during querying to select the optimal one. However, this scale-based solution compromises both efficiency and effectiveness. It could increase query time by up to 30 times as it needs to render at multiple different scales. Moreover, most patches with varying scales often fail to accurately encompass objects, either frequently including other objects from the background or omitting portions of the target object. These inaccurate CLIP features lead to the trained 3D language field lacking clear boundaries and containing a significant amount of noise. Therefore, they often simultaneously learn pixel-aligned DINO features to mitigate this issue. However, the performance remains un" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "satisfactory. As shown in Figure 1, LERF still generates imprecise 3D language features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 96, + 547, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 503 + ], + "type": "text", + "content": "In this paper, we propose the 3D Language Gaussian Splatting (LangSplat) to address the above issues. Instead of using NeRF to build 3D representations, we resort to 3D Gaussian Splatting, which represents a 3D scene as a collection of 3D Gaussians and uses tile-based splatting to achieve efficient rendering at high resolutions. Our LangSplat defines a set of 3D language Gaussians, with each Gaussian being enhanced by a language embedding. These language-enhanced Gaussians are supervised using CLIP embeddings extracted from image patches captured from multiple training views, ensuring multi-view consistency. As an explicit modeling method, directly storing the high-dimensional language embeddings for each 3D language Gaussian is memory-inefficient. To reduce the memory cost and further improve the rendering efficiency, we propose to first learn a scene-wise language autoencoder, which maps CLIP embeddings in a scene to a low-dimensional latent space. In this way, each language Gaussian only contains the low-dimensional latent language features and the final language embeddings are obtained with decoding of the rendered features. To address the point ambiguity issue, we propose to employ the semantic hierarchy defined by the Segment Anything Model (SAM) [19]. Specifically, for each 2D image, we obtain three well-segmented maps at different semantic levels with SAM. Then we extract the CLIP feature for each mask with precise object boundaries and assign this feature to every point on the corresponding mask. Learning with SAM-based masks not only endows each point with precise CLIP embeddings, resulting in higher model accuracy, but also enables direct querying at predefined three semantic scales. This circumvents the need for intensive searches across multiple absolute scales and the auxiliary DINO features, thereby effectively improving efficiency. We summarize the contributions of this paper as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 504, + 545, + 635 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 306, + 504, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 545, + 552 + ], + "type": "text", + "content": "- We propose the LangSplat, which is the first 3D Gaussian Splitting-based method for 3D language fields. A scene-specific autoencoder is further introduced to alleviate the memory cost issue imposed by explicit modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 545, + 587 + ], + "type": "text", + "content": "- We propose to learn the hierarchical semantics defined by SAM to address the point ambiguity issue for 3D language field modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "text", + "content": "- Experimental results show that our method outperforms the state-of-the-art methods on open-vocabulary 3D object localization and semantic segmentation tasks while being " + }, + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "inline_equation", + "content": "199 \\times" + }, + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "text", + "content": " faster than LERF at " + }, + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 306, + 588, + 545, + 635 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 392, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "3D Gaussian Splatting. Real-time rendering has always been a pursued objective for neural rendering. Recently, Kerbl et al. [17] proposed to represent the 3D scene with a set of 3D Gaussians, which attained real-time rendering for" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20052" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "1080p resolution while maintaining state-of-the-art visual quality. Encouraged by the success of 3D Gaussian Splatting, many works extend it to other tasks. To achieve real-time dynamic scene rendering, some studies [27, 45, 49] have extended the 3D Gaussian Splatting technique to dynamic scenes. Luiten et al. [27] proposed the Dynamic 3D Gaussians, which extended 3D Gaussians to dynamic scenes by explicitly modeling the 3D Gaussians across different time steps. Yang et al. [48] presented a deformable 3D Gaussians Splatting method, which learned 3D Gaussians in canonical space and modeled the dynamic scenes with a deformation field. Meanwhile, some researchers have combined 3D Gaussian Splatting with diffusion models to achieve efficient text-to-3D generation [10, 42, 51]. For example, Tang et al. [42] introduced DreamGaussian for efficient 3D content generation with a generative 3D Gaussian Splatting model. Unlike these methods, our paper extends each 3D Gaussian with language embeddings for open-vocabulary 3D queries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 303, + 289, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 289, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 289, + 472 + ], + "type": "text", + "content": "SAM. The Segment Anything Model [19], released by Meta in 2023, has attracted considerable attention [15, 28, 29], which achieved impressive zero-shot performance. It has become the foundational model for image segmentation. SAM has been used for many vision tasks [26, 30, 44, 50] such as image inpainting [53], object tracking [11, 47], image editing [12], and so on. Many efforts have also been made to utilize SAM in the 3D domain [6, 36]. Liu et al. [25] proposed Seal to explore the potential of VFMs including SAM for point cloud segmentation. SA3D [6] generalized SAM to 3D objects by leveraging NeRF to connect 2D images and 3D space. Different from these works, we use SAM to obtain accurate object masks with three well-defined hierarchical semantics to train a 3D language field." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 289, + 715 + ], + "type": "text", + "content": "3D Langue fields. Some early attempts to construct 3D feature fields included Distilled Feature Fields [20] and Neural Feature Fusion Fields [43]. They learned 3D-consistent features by distilling LSeg [21] or DINO [4] features across multiple views into a NeRF. Shen et al. [39] further adopted distilled feature fields for few-shot language-guided robotic manipulation by distilling CLIP feature into a NeRF. There are also some efforts [40, 54] that embed semantic information into NeRFs. For example, Semantic NeRF [54] jointly encoded semantics with appearance and geometry within a NeRF for novel semantic view synthesis. LERF [18] was the first to embed CLIP features into NeRF, enabling open-vocabulary 3D queries leveraging the powerful CLIP representation. DINO features were also used for supervising LERF to improve its performance. Liu et al. [24] also utilized CLIP and DINO features to train a NeRF model for 3D open-vocabulary segmentation. While these methods use NeRF for 3D modeling and suffer from the costly rendering process, we propose the 3D language Gaussian Splatting to obtain efficient 3D language fields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 71, + 423, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 423, + 86 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 423, + 86 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 140 + ], + "type": "text", + "content": "In this section, we first revisit the challenges of modeling 3D language fields and then elaborate on how our proposed LangSplat addresses these issues. Figure 2 depicts the framework of our proposed LangSplat." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 144, + 538, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 538, + 158 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 538, + 158 + ], + "type": "text", + "content": "3.1. Revisiting the Challenges of Language Fields" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": "We denote an input image as " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " represent the height and weight of the image size. We take a set of calibrated images " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "\\{I_t | t = 1, 2, \\ldots, T\\}" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " as input and train a 3D language field " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " with these images. Most existing methods [18, 24, 35] employ the CLIP image encoder " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " to extract image features and utilize the extracted CLIP embeddings to supervise the 3D language field " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": ", leveraging the well-aligned text-image latent space [22, 34, 55] provided by CLIP, thus facilitating open-vocabulary queries. However, CLIP embeddings are image-aligned rather than pixel-aligned. In other words, simply computing " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "V(I_t) \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " only obtains an image-level feature, whereas what we need is a pixel-aligned language embedding " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "L_t \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 163, + 546, + 402 + ], + "type": "text", + "content": " represents the CLIP feature dimension. Meanwhile, modeling pixel-aligned language features faces the issue of point ambiguity, as a single point on an object contributes to multiple semantic levels of regions. For instance, a point on a cat's ear simultaneously contributes to the cat's ear, the cat's head, and the entire cat, and should be activated by all three types of textual queries." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": "To address these issues, most existing methods [18, 24] extract a hierarchy of CLIP features from cropped image patches. Specifically, for a pixel with coordinates " + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "inline_equation", + "content": "v \\in \\{1, \\dots, H\\} \\times \\{1, \\dots, W\\}" + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": ", the corresponding CLIP features are obtained from image patches centered around " + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": " at different absolute physical scales " + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": ", with the expectation that at a certain scale " + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": ", the patch can fully encompass the object. However, this multi-scale approach has two limitations. Firstly, patch features are imprecise because they often include additional contextual object information, leading to overly smoothed language fields with indistinct object boundaries. To alleviate the patchy issue, most methods [18, 24] leverage additional pixel-aligned DINO features to supervise the network. However, the learned 3D language features are still imprecise, as illustrated in Figure 1. Secondly, it requires simultaneous rendering at multiple scales during inference to find the optimal scale. With the number of scales " + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 403, + 547, + 629 + ], + "type": "text", + "content": " potentially reaching as high as 30 [18], this significantly diminishes the inference speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Besides the rendering target, another key design space is the 3D modeling approach. Most existing methods [2, 43] employ NeRFs for 3D representation, where they learn a language feature at each 3D point and subsequently render the language feature onto an image, similar to color rendering. However, NeRF-based methods are constrained by their time-consuming rendering process, even though the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20053" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 200 + ], + "type": "image", + "image_path": "6044390a5174794305df198a967eb2f351851440a5dd54a91c6bbe3e252316cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "type": "text", + "content": "Figure 2. The framework of our LangSplat. Our LangSplat leverages SAM to learn hierarchical semantics to address the point ambiguity issue. Then segment masks are sent to the CLIP image encoder to extract the corresponding CLIP embeddings. We learn an autoencoder with these obtained CLIP embeddings. Our 3D language Gaussian learn language features on the scene-specific latent space to reduce the memory cost. During querying, the rendered language embeddings are sent to the decoder to recover the features on the CLIP space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 288, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 321 + ], + "type": "text", + "content": "most advanced NeRF techniques currently available cannot achieve real-time rendering in high-resolution, unrestricted scenes [17]. Meanwhile, there is a high demand for efficient open vocabulary querying in practical applications, especially in fields such as intelligent robotics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 332, + 275, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 275, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 275, + 345 + ], + "type": "text", + "content": "3.2. Learning Hierarchical Semantics with SAM" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 520 + ], + "type": "text", + "content": "As a foundation model for image segmentation, SAM [19] can accurately group a pixel with its surrounding pixels belonging to the same object, thereby segmenting the image into many object masks with clear boundaries. Furthermore, SAM addresses point ambiguity by generating three different masks for a point prompt, namely, whole, part, and subpart, representing three hierarchical levels of semantics. In this paper, we propose leveraging SAM to obtain precise object masks, which are then used to acquire pixel-aligned features. We also explicitly model the semantic hierarchy defined by SAM to address the point ambiguity issue. With SAM, we can capture the semantic hierarchy of objects in 3D scenes, providing accurate and multi-scale segmentation maps for each input image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "Specifically, we feed a regular grid of " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " point prompts into SAM to obtain the masks under three different semantic levels: " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_0^s, M_0^p, M_0^w" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_0^s, M_0^p" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "M_0^w" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": " represent the masks at subpart, part, and whole levels, respectively. Then we remove redundant masks for each of the three mask sets based on the predicted IoU score, stability score, and overlap rate between masks. Each filtered mask set independently performs a comprehensive full-image segmentation based on its respective semantic level, resulting in three segmentation maps: " + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "inline_equation", + "content": "M^s, M^p, M^w" + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": ". These segmentation maps precisely delineate the boundaries of objects at their hierarchical levels, effectively partitioning the scene into semantically meaningful regions. With the obtained segmentation maps, we proceed to extract CLIP features for each segmented region. These features capture the semantic context of the objects" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 261, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 285 + ], + "type": "text", + "content": "at various levels within the scene. Mathematically, the obtained pixel-aligned language embeddings are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 343, + 293, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 293, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 343, + 293, + 545, + 308 + ], + "type": "interline_equation", + "content": "\\boldsymbol {L} _ {t} ^ {l} (v) = V \\left(\\boldsymbol {I} _ {t} \\odot \\boldsymbol {M} ^ {l} (v)\\right), l \\in \\{s, p, w \\}, \\tag {1}", + "image_path": "7e461e3f8f5cab16b946e197b492db582d69c314c6112ffb31be1c09d5193eb7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "inline_equation", + "content": "M^l(v)" + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "text", + "content": " represents the mask region to which pixel " + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "text", + "content": " belongs at the semantic level " + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 305, + 315, + 545, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 340, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 546, + 532 + ], + "type": "text", + "content": "Each pixel rendered from the 3D language scene now possesses a CLIP feature that aligns with its precise semantic context. This alignment reduces ambiguity and enhances the accuracy of language-based queries. We can learn an accurate 3D language field even without the commonly used DINO regularization. Another advantage of our SAM-based approach is the predefined semantic scales. Since we have distinct segmentation maps for \"whole,\" \"part,\" and \"subpart\" levels, we can directly query the 3D language field at these predefined scales. This eliminates the need for intensive searches across multiple absolute scales, making the querying process more efficient. By incorporating SAM's semantic hierarchy into our approach, we not only improve the accuracy of our 3D language field but also streamline the querying process, making it more efficient and effective for a wide range of applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 537, + 529, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 537, + 529, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 529, + 551 + ], + "type": "text", + "content": "3.3. 3D Gaussian Splitting for Language Fields" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": "Having obtained the language embeddings on a set of 2D images " + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "inline_equation", + "content": "\\{L_t^l, |t = 1, \\dots, T\\}" + }, + { + "bbox": [ + 304, + 555, + 545, + 640 + ], + "type": "text", + "content": ", we can learn a 3D language scene by modeling the relations between 3D points and 2D pixels. Most existing methods [2, 43] suffer from the costly rendering process as they adopt NeRFs for 3D modeling. To address this issue, we present the first 3D Gaussian Splitting-based method for 3D language field modeling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "text", + "content": "3D Gaussian Splitting explicitly represents a 3D scene as a collection of anisotropic 3D Gaussians, with each Gaussian " + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "text", + "content": " characterized by a mean " + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "text", + "content": " and a covariance matrix " + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 304, + 640, + 545, + 687 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 342, + 693, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 342, + 693, + 545, + 717 + ], + "type": "interline_equation", + "content": "G (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)\\right). \\tag {2}", + "image_path": "952b8a84810546ae2256174ddf102680e07f3fcd6773741d3d51334ab80d35c3.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20054" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "To optimize the parameters of 3D Gaussians, they are rendered into 2D image planes [56], and a tile-based rasterizer is used to improve the rendering efficiency:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 118, + 287, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 118, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 104, + 118, + 287, + 152 + ], + "type": "interline_equation", + "content": "C (v) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {3}", + "image_path": "922e055cd0817cefaee18fdeaefa48660c6a889767f175f52c2ca1ae3ac1ee43.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": " is the color of the " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": "-th Gaussian, " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": " denotes the Gaussians in the tile, " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "C(v)" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": " is the rendered color at pixel " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = o_{i}G_{i}^{2D}(v)" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "o_{i}" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": " is the opacity of the " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": "-th Gaussian and " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "G_{i}^{2D}(\\cdot)" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": " represents the function of the " + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 158, + 287, + 218 + ], + "type": "text", + "content": "-th Gaussian projected onto 2D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "content": "In this paper, we proposes the 3D language Gaussian Splatting, which augments each 3D Gaussian with three language embeddings " + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "inline_equation", + "content": "\\{\\pmb{f}^s,\\pmb{f}^p,\\pmb{f}^w\\}" + }, + { + "bbox": [ + 46, + 219, + 287, + 302 + ], + "type": "text", + "content": ". These embeddings are derived from CLIP features, which capture the hierarchical semantics provided by SAM. The augmented Gaussians are named as 3D language Gaussians. We also adopt the tile-based rasterizer to retain the rendering efficiency:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 312, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 312, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 73, + 312, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {l} (v) = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {f} _ {i} ^ {l} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), l \\in \\{s, p, w \\}, \\tag {4}", + "image_path": "5dffc34503a368a7b1ba2f5e4436bdc02ba8164834765b6c09a6373a4d94f1f4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "inline_equation", + "content": "F^l(v)" + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "text", + "content": " represents the language embedding rendered at pixel " + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "text", + "content": " with the semantic level " + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 354, + 287, + 403 + ], + "type": "text", + "content": ". By incorporating language information directly into the Gaussians, we enable the 3D language field to respond to language-based queries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "text", + "content": "As an explicit modeling approach, our LangSplat may create millions of 3D points to model a complex 3D scene. As CLIP embeddings are high-dimensional features, directly learning " + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{l}" + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "text", + "content": " on the CLIP latent space significantly increases memory and time cost. Compared to learning RGB colors without spherical harmonics coefficients, learning 512-dimensional CLIP features increases the memory requirements for storing 3D Gaussians by over 35 times, easily leading to running out of L1 cache memory. To reduce memory cost and improve efficiency, we introduce a scenewise language autoencoder. This autoencoder maps CLIP embeddings in a scene to a lower-dimensional latent space, reducing memory requirements. The CLIP model is trained using 400 million (image, text) pairs, and its " + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "text", + "content": "-dimensional latent space could be highly compact, as it needs to align arbitrary text and images in this space. However, the language field " + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 403, + 287, + 689 + ], + "type": "text", + "content": " we train here is scene-specific, meaning we can leverage scene priors to compress CLIP features. In fact, for each input image, we will obtain hundreds of masks segmented by SAM, which is significantly smaller than the number of images used in CLIP training. Therefore, all the segmented regions in a scene are sparsely distributed in the CLIP latent space, allowing us to further compress these CLIP features using a scene-specific autoencoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Specifically, we use the collections of CLIP features of SAM segmented masks " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{L_t^l | l \\in \\{s, p, w\\}, 1 \\leq t \\leq T\\}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": "train a lightweight autoencoder. An encoder " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " maps the " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": "-dimensional CLIP features " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\pmb{L}_t^l (v)\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\pmb{H}_t^l (v) = E(\\pmb{L}_t^l (v))\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "d\\ll D" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": ". Then we learn a decoder " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": " to reconstruct the original CLIP embeddings from the compressed representation. The autoencoder is trained with a reconstruction objective on the CLIP embeddings " + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\{\\pmb{L}_t^l\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 152, + 545, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 152, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 329, + 152, + 545, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a e} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {a e} \\left(\\Psi \\left(E \\left(\\boldsymbol {L} _ {t} ^ {l} (v)\\right)\\right), \\boldsymbol {L} _ {t} ^ {l} (v)\\right), \\tag {5}", + "image_path": "fdb8e3fbbf825c6ebcd04c101254748b67272464279d22f6d3bb092903be6ae5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "inline_equation", + "content": "d_{ae}()" + }, + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "text", + "content": " denotes a distance function used for the autoencoder. Here we adopt both " + }, + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 305, + 194, + 545, + 218 + ], + "type": "text", + "content": " and a cosine distance loss." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": "After training the autoencoder, we transform all CLIP embeddings " + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{L_t^l\\}" + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": " into scene-specific latent features " + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\{H_t^l\\}" + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": ". We let our 3D language Gaussians learn language embeddings in the scene-specific latent space instead of the CLIP latent space. Therefore, we have " + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "inline_equation", + "content": "f^l \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": ". In practice, we choose " + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "inline_equation", + "content": "d = 3" + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": " as it yields excellent model efficiency and accuracy. Compared to directly modeling the " + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 219, + 545, + 339 + ], + "type": "text", + "content": "-dimensional CLIP embeddings, our method significantly reduced the memory cost by incorporating scene priors. We optimized the language embeddings with the objective:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 346, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 346, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 335, + 346, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l a n g} = \\sum_ {l \\in \\{s, p, w \\}} \\sum_ {t = 1} ^ {T} d _ {l a n g} \\left(\\boldsymbol {F} _ {t} ^ {l} (v), \\boldsymbol {H} _ {t} ^ {l} (v)\\right), \\tag {6}", + "image_path": "556e526bf80be7550de076e5876e187c3424b698406dd8cd113e07656474068e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 388, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 388, + 545, + 412 + ], + "type": "inline_equation", + "content": "d_{lang}()" + }, + { + "bbox": [ + 305, + 388, + 545, + 412 + ], + "type": "text", + "content": " denotes the distance function used for our 3D Language Gaussians." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "text", + "content": "During inference, we follow the Eq. (4) to render the language embeddings from 3D to 2D, and then we use the trained scene-specific decoder " + }, + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "text", + "content": " to recover the CLIP image embeddings " + }, + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\Psi (\\pmb {F}_t^l)\\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 305, + 413, + 545, + 472 + ], + "type": "text", + "content": ", which enable open-vocabulary queries with the CLIP text encoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 473, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 556 + ], + "type": "text", + "content": "By enhancing 3D Gaussians with language embedding and employing a scene-wise language autoencoder, our proposed LangSplat presents a powerful and efficient solution for building 3D language fields. This approach not only preserves the rendering efficiency of Gaussian Splatting but also mitigates the catastrophic memory explosion associated with explicit modeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 563, + 457, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 457, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 457, + 577 + ], + "type": "text", + "content": "3.4. Open-vocabulary Querying" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 714 + ], + "type": "text", + "content": "Due to the well-aligned latent space between images and text provided by the CLIP model, our learned 3D language field can easily support open-vocabulary 3D queries, including open-vocabulary 3D object localization and open-vocabulary 3D semantic segmentation. Many existing open-vocabulary 3D semantic segmentation methods [24] usually select the category from a category list, which includes the categories present in the images. However, obtaining a comprehensive category list for in-the-wild scenes is challenging. Different from them, our method generates precise object masks given an arbitrary text query." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20055" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "content": "Following LERF [18], we compute the relevancy score for each text query. Specifically, for each rendered language embedding " + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\phi_{img}" + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "content": " and each text query " + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\phi_{qry}" + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "content": ", the relevancy score is defined as " + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\min_i\\frac{\\exp(\\phi_{img} \\cdot \\phi_{qry})}{\\exp(\\phi_{img} \\cdot \\phi_{qry}) + \\exp(\\phi_{img} \\cdot \\phi_{canon}^i)}" + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "inline_equation", + "content": "\\phi_{canon}^i" + }, + { + "bbox": [ + 46, + 72, + 289, + 269 + ], + "type": "text", + "content": " is the CLIP embeddings of a predefined canonical phrase chosen from \"object\", \"things\", \"stuff\", and \"texture\". Hence, for each text query, we will obtain three relevancy maps, each representing results at a specific semantic level. We follow the strategy used in LERF [18] and choose the semantic level that yields the highest relevancy score. For the 3D object localization task, we directly choose the point with the highest relevance score. For the 3D semantic segmentation task, we filter out points with relevancy scores lower than a chosen threshold, and predict the object masks with remaining regions. Please refer to the appendix for additional details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 280, + 128, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 280, + 128, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 280, + 128, + 294 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 107, + 312 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 318, + 289, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 289, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 289, + 582 + ], + "type": "text", + "content": "Datasets. We employ two datasets for evaluation. The LERF dataset [18] is captured using the iPhone App Polycam, which consists of complex in-the-wild scenes. The LERF dataset is designed for 3D object localization tasks, here we extend the LERF dataset by annotating ground truth masks for textual queries, enabling the evaluation of the open-vocabulary 3D semantic segmentation on the LERF dataset. As the original LERF annotations for 3D object localization are relatively simple, performance in some scenarios has already approached saturation. Therefore, we further manually annotated additional challenging localization samples to better evaluate method performance. We report localization accuracy for the 3D object localization task following LERF [18], and report the IoU results for the 3D semantic segmentation task. We also employ the 3D-OVS dataset [24], which comprises a collection of long-tail objects captured in diverse poses and backgrounds. This dataset is developed for open-vocabulary 3D semantic segmentation, where a full list of categories is provided. While other methods use the full list to generate the predicted masks, we only use the query category to generate the corresponding masks. The mIoU metric is used for this dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 714 + ], + "type": "text", + "content": "Implementation Details. To extract the language features of each image, we utilize the OpenCLIP ViT-B/16 model. For SAM, we use the ViT-H model to segment 2D masks. For each scene, we first use 3D Gaussian Splatting to train an RGB scene. We train it for 30,000 iterations, and in the end, each scene comprises around 2,500,000 points. We follow the default parameter setting as in [17] to train the RGB scene. Then we train our 3D language Gaussians by fixing all other parameters of 3D Gaussians such as mean and opacity. Only the language features are learnable during this stage. We train the language features for 30,000" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 310, + 70, + 543, + 160 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 543, + 160 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 543, + 160 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 543, + 160 + ], + "type": "table", + "html": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen14.162.073.2
figurines8.975.080.4
teatime33.984.888.1
waldo_kitchen27.372.795.5
overall21.173.684.3
", + "image_path": "2c27469ba1ff54ff355ef44f4c37fa9d60a189cafa66443913bf7a8deecd7e53.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 186, + 543, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 168, + 545, + 179 + ], + "lines": [ + { + "bbox": [ + 306, + 168, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 168, + 545, + 179 + ], + "type": "text", + "content": "Table 1. Localization accuracy (%) comparisons on LERF dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 186, + 543, + 277 + ], + "lines": [ + { + "bbox": [ + 310, + 186, + 543, + 277 + ], + "spans": [ + { + "bbox": [ + 310, + 186, + 543, + 277 + ], + "type": "table", + "html": "
Test SceneLSeg [21]LERF [18]LangSplat
ramen7.028.251.2
figurines7.638.644.7
teatime21.745.065.1
waldo_kitchen29.937.944.5
overall16.637.451.4
", + "image_path": "da6a82063fe732c3252edff1ec9b9c9a0e00a4ac9f567f88e2405ed62d1b2b55.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of 3D semantic segmentation on the LERF dataset. We report the average IoU scores " + }, + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "text", + "content": "iterations. Our autoencoder is implemented by an MLP, which compresses the 512-dimensional CLIP features into 3-dimensional latent features. For a " + }, + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "text", + "content": " resolution scene, our model is trained for " + }, + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "inline_equation", + "content": "\\sim 25" + }, + { + "bbox": [ + 305, + 325, + 545, + 385 + ], + "type": "text", + "content": " minutes on an NVIDIA RTX-3090 GPU and takes roughly 4GB of memory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 394, + 463, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 463, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 463, + 406 + ], + "type": "text", + "content": "4.2. Results on the LERF dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "content": "Quantitative Results. We first compare our method with other methods on the LERF dataset. Table 1 shows the localization results. We observe that our method achieves an overall accuracy of " + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "inline_equation", + "content": "84.3\\%" + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "content": ", significantly outperforming LERF. Table 2 further shows the IoU results of 3D semantic segmentation, our method outperforms LERF by " + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "inline_equation", + "content": "14.0\\%" + }, + { + "bbox": [ + 304, + 413, + 545, + 653 + ], + "type": "text", + "content": ", which illustrates the superiority of our proposed LangSplat. Visualization Results. To show the learned 3D language field, we visualize the learned features by computing 3-dimensional PCA components of learned language features following [20]. The results are shown in Figure 1. We see that the LERF learned features fail to generate clear boundaries between objects while our method gives precise object shapes solely using CLIP features. We further show the visualization results of object localization and semantic segmentation in Figure 3 and Figure 4, respectively. We observe that the activation regions generated by LERF are more dispersed, while ours are more concentrated, and our activation regions can better align with the ground truth shape compared to those produced by LERF." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 714 + ], + "type": "text", + "content": "Ablation Study. We conduct ablations on the ramen scene and report the semantic segmentation results in Table 3. We test the query speed on an NVIDIA RTX-3090 GPU. Here AE represents autoencoder and 3D-GS denotes 3D Gaussian Splatting. Without any of our proposed components," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20056" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 129, + 132 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 129, + 132 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 129, + 132 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 129, + 132 + ], + "type": "image", + "image_path": "297a3511d890d5cf3169741f7f9387f8c7ef02c5094bef0cd55e7dc2d3ded6f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 133, + 99, + 139 + ], + "lines": [ + { + "bbox": [ + 85, + 133, + 99, + 139 + ], + "spans": [ + { + "bbox": [ + 85, + 133, + 99, + 139 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 76, + 208, + 132 + ], + "blocks": [ + { + "bbox": [ + 132, + 76, + 208, + 132 + ], + "lines": [ + { + "bbox": [ + 132, + 76, + 208, + 132 + ], + "spans": [ + { + "bbox": [ + 132, + 76, + 208, + 132 + ], + "type": "image", + "image_path": "0dface535705b12611324635b6e3be140a956aac20f0304b1c60f63abbc32543.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 140, + 129, + 195 + ], + "blocks": [ + { + "bbox": [ + 53, + 140, + 129, + 195 + ], + "lines": [ + { + "bbox": [ + 53, + 140, + 129, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 140, + 129, + 195 + ], + "type": "image", + "image_path": "0ca856bb07ca8822c3207a00fa5a719b68d2f9c1408989ad929f5f4a8a3739b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 198, + 99, + 203 + ], + "lines": [ + { + "bbox": [ + 85, + 198, + 99, + 203 + ], + "spans": [ + { + "bbox": [ + 85, + 198, + 99, + 203 + ], + "type": "text", + "content": "LERF" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 140, + 208, + 196 + ], + "blocks": [ + { + "bbox": [ + 163, + 133, + 177, + 140 + ], + "lines": [ + { + "bbox": [ + 163, + 133, + 177, + 140 + ], + "spans": [ + { + "bbox": [ + 163, + 133, + 177, + 140 + ], + "type": "text", + "content": "LSeg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 132, + 140, + 208, + 196 + ], + "lines": [ + { + "bbox": [ + 132, + 140, + 208, + 196 + ], + "spans": [ + { + "bbox": [ + 132, + 140, + 208, + 196 + ], + "type": "image", + "image_path": "1a15befdb603f253347b5bbe0fcf817b1e03173650902d9414e012c18c86959b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 198, + 178, + 203 + ], + "lines": [ + { + "bbox": [ + 162, + 198, + 178, + 203 + ], + "spans": [ + { + "bbox": [ + 162, + 198, + 178, + 203 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 209, + 150, + 220 + ], + "lines": [ + { + "bbox": [ + 119, + 209, + 150, + 220 + ], + "spans": [ + { + "bbox": [ + 119, + 209, + 150, + 220 + ], + "type": "text", + "content": "'spoon'" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 220, + 75, + 294, + 132 + ], + "blocks": [ + { + "bbox": [ + 220, + 75, + 294, + 132 + ], + "lines": [ + { + "bbox": [ + 220, + 75, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 220, + 75, + 294, + 132 + ], + "type": "image", + "image_path": "48501b968ef74b9b4932a69cf60bcf1c69b2338bbee29f24b3ae79ebbe125f9f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 133, + 264, + 139 + ], + "lines": [ + { + "bbox": [ + 251, + 133, + 264, + 139 + ], + "spans": [ + { + "bbox": [ + 251, + 133, + 264, + 139 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 299, + 76, + 374, + 132 + ], + "blocks": [ + { + "bbox": [ + 299, + 76, + 374, + 132 + ], + "lines": [ + { + "bbox": [ + 299, + 76, + 374, + 132 + ], + "spans": [ + { + "bbox": [ + 299, + 76, + 374, + 132 + ], + "type": "image", + "image_path": "11d888b7f4801b5aab536ef7eb2176c1b78e7f1ed7b77db58c0bdf724ee4d459.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 220, + 140, + 294, + 196 + ], + "blocks": [ + { + "bbox": [ + 220, + 140, + 294, + 196 + ], + "lines": [ + { + "bbox": [ + 220, + 140, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 220, + 140, + 294, + 196 + ], + "type": "image", + "image_path": "116feb8303a161929e797879fe807c60d476b245bc84d67db8f85f63d8ef0f0c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 198, + 265, + 203 + ], + "lines": [ + { + "bbox": [ + 251, + 198, + 265, + 203 + ], + "spans": [ + { + "bbox": [ + 251, + 198, + 265, + 203 + ], + "type": "text", + "content": "LERF" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 209, + 328, + 220 + ], + "lines": [ + { + "bbox": [ + 270, + 209, + 328, + 220 + ], + "spans": [ + { + "bbox": [ + 270, + 209, + 328, + 220 + ], + "type": "text", + "content": "'paper napkin'" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 299, + 140, + 374, + 196 + ], + "blocks": [ + { + "bbox": [ + 329, + 133, + 343, + 140 + ], + "lines": [ + { + "bbox": [ + 329, + 133, + 343, + 140 + ], + "spans": [ + { + "bbox": [ + 329, + 133, + 343, + 140 + ], + "type": "text", + "content": "LSeg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 140, + 374, + 196 + ], + "lines": [ + { + "bbox": [ + 299, + 140, + 374, + 196 + ], + "spans": [ + { + "bbox": [ + 299, + 140, + 374, + 196 + ], + "type": "image", + "image_path": "f52d64e734b1a6e886766b7ed63ec5bda073e5348b188a44a4acf861eaff662b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 198, + 343, + 203 + ], + "lines": [ + { + "bbox": [ + 329, + 198, + 343, + 203 + ], + "spans": [ + { + "bbox": [ + 329, + 198, + 343, + 203 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 385, + 75, + 461, + 132 + ], + "blocks": [ + { + "bbox": [ + 385, + 75, + 461, + 132 + ], + "lines": [ + { + "bbox": [ + 385, + 75, + 461, + 132 + ], + "spans": [ + { + "bbox": [ + 385, + 75, + 461, + 132 + ], + "type": "image", + "image_path": "7cbc8586c1493da12a3a04dc50d3658bb19ce1fdbf838ff73144eb40bc658710.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 133, + 430, + 139 + ], + "lines": [ + { + "bbox": [ + 418, + 133, + 430, + 139 + ], + "spans": [ + { + "bbox": [ + 418, + 133, + 430, + 139 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 465, + 76, + 539, + 132 + ], + "blocks": [ + { + "bbox": [ + 465, + 76, + 539, + 132 + ], + "lines": [ + { + "bbox": [ + 465, + 76, + 539, + 132 + ], + "spans": [ + { + "bbox": [ + 465, + 76, + 539, + 132 + ], + "type": "image", + "image_path": "ac2d94c40620224d73343a9d05a3d9f2214ee89670afd9c72515a41da401db05.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 386, + 140, + 460, + 196 + ], + "blocks": [ + { + "bbox": [ + 386, + 140, + 460, + 196 + ], + "lines": [ + { + "bbox": [ + 386, + 140, + 460, + 196 + ], + "spans": [ + { + "bbox": [ + 386, + 140, + 460, + 196 + ], + "type": "image", + "image_path": "943809e1a7975c721d13079ccf52caeaf265cff030232e288948c880f67584ac.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 198, + 431, + 203 + ], + "lines": [ + { + "bbox": [ + 417, + 198, + 431, + 203 + ], + "spans": [ + { + "bbox": [ + 417, + 198, + 431, + 203 + ], + "type": "text", + "content": "LERF" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 209, + 482, + 220 + ], + "lines": [ + { + "bbox": [ + 446, + 209, + 482, + 220 + ], + "spans": [ + { + "bbox": [ + 446, + 209, + 482, + 220 + ], + "type": "text", + "content": "'spatula'" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 465, + 140, + 539, + 196 + ], + "blocks": [ + { + "bbox": [ + 495, + 133, + 509, + 140 + ], + "lines": [ + { + "bbox": [ + 495, + 133, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 495, + 133, + 509, + 140 + ], + "type": "text", + "content": "LSeg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 465, + 140, + 539, + 196 + ], + "lines": [ + { + "bbox": [ + 465, + 140, + 539, + 196 + ], + "spans": [ + { + "bbox": [ + 465, + 140, + 539, + 196 + ], + "type": "image", + "image_path": "041282c7b0a12dc3d46646cd6d5a37ace264767cf9a903de2bdf8f9fda52bb9e.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 198, + 509, + 203 + ], + "lines": [ + { + "bbox": [ + 495, + 198, + 509, + 203 + ], + "spans": [ + { + "bbox": [ + 495, + 198, + 509, + 203 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 50, + 258, + 545, + 363 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 545, + 251 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons of open-vocabulary 3D object localization on the LERF dataset. The red points are the model predictions and the black dashed bounding boxes denote the annotations." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 258, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 50, + 258, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 50, + 258, + 545, + 363 + ], + "type": "image", + "image_path": "7f6cb00c205fbe5d62bf2238bab3476e5b59dfa7000793a15a4ff1a2a1101a3f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 370, + 482, + 382 + ], + "lines": [ + { + "bbox": [ + 109, + 370, + 482, + 382 + ], + "spans": [ + { + "bbox": [ + 109, + 370, + 482, + 382 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons of open-vocabulary 3D semantic segmentation on the LERF dataset." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 55, + 388, + 280, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 388, + 280, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 280, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 280, + 483 + ], + "type": "table", + "html": "
ComponentPerformance
AE3D-GSSAMIoU (%)Speed (s/q)
✓28.2030.93
46.747.77
✓✓OOMOOM
✓✓✓51.150.26
", + "image_path": "7f59a20cdde0a3da12e2926a97179e967af38e98fea2922a3356f035e3b11440.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_body" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 51, + 510, + 283, + 605 + ], + "blocks": [ + { + "bbox": [ + 50, + 491, + 284, + 502 + ], + "lines": [ + { + "bbox": [ + 50, + 491, + 284, + 502 + ], + "spans": [ + { + "bbox": [ + 50, + 491, + 284, + 502 + ], + "type": "text", + "content": "Table 3. Ablations. The results are obtained on the ramen scene." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 510, + 283, + 605 + ], + "lines": [ + { + "bbox": [ + 51, + 510, + 283, + 605 + ], + "spans": [ + { + "bbox": [ + 51, + 510, + 283, + 605 + ], + "type": "table", + "html": "
ComponentPerformance
AE3D-GSSAMmIoU (%)Speed (s/q)
✓53.255.7
85.718.4
✓✓OOMOOM
✓✓✓94.20.28
", + "image_path": "7ea05b997b1b9b1ca545b03744726975a420387de61129d875046a9eacd8e02d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 46, + 613, + 286, + 635 + ], + "lines": [ + { + "bbox": [ + 46, + 613, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 286, + 635 + ], + "type": "text", + "content": "Table 4. Ablations result on the bench scene of the 3D-OVS dataset. The image resolution is " + }, + { + "bbox": [ + 46, + 613, + 286, + 635 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 46, + 613, + 286, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "our baseline equals LERF, which has a speed of 30.93 seconds per text query at the resolution of " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "988 \\times 731" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". Using SAM to replace the scale-based solution significantly increases the IoU by " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "18.54\\%" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", showing our SAM-based solution effectively addresses the point ambiguity issue, leading" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": "to accurate 3D language features. Simply replacing NeRF with 3D Gaussian Splatting leads to running out of L1 cache memory as explicitly modeling CLIP features poses huge memory demands. Incorporating a scene-specific autoencoder effectively addresses this issue and results in further improvements in both accuracy and efficiency. In the end, our LangSplat achieved a " + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "inline_equation", + "content": "119 \\times" + }, + { + "bbox": [ + 304, + 390, + 545, + 486 + ], + "type": "text", + "content": " speedup over LERF while significantly surpassing LERF in terms of accuracy." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "text", + "content": "We further conducted the ablations on the 3D-OVS dataset, which has a higher image resolution of " + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "text", + "content": ". Table 4 lists the results on the bench scene. We also tested the query speed on an NVIDIA RTX-3090 GPU. We observed that with the increase in image resolution, the speedup over LERF further improved to " + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "inline_equation", + "content": "199 \\times" + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "text", + "content": ", which demonstrates the huge potential of our method. Our further study shows that most of the computational time is allocated to the decoder rather than the rendering process. We could replace the decoder with a " + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 488, + 546, + 620 + ], + "type": "text", + "content": " convolutional layer to attain a higher speedup." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 305, + 633, + 471, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 633, + 471, + 645 + ], + "spans": [ + { + "bbox": [ + 305, + 633, + 471, + 645 + ], + "type": "text", + "content": "4.3. Results on the 3D-OVS dataset" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 712 + ], + "type": "text", + "content": "Quantitative Results. We compare LangSplat with other 2D and 3D state-of-the-art methods on the 3D-OVS dataset in Table 5. We observe that LangSplat not only outperforms 2D-based methods such as ODISE [46] and OV-Seg [23], but also achieves better results than 3D-based methods in" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20057" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 542, + 385 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 542, + 385 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 542, + 385 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 542, + 385 + ], + "type": "image", + "image_path": "296657d11ae3cf385d956d05e6d7acc4da81d368a54d8df763ff998a37a83707.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 396, + 546, + 419 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 546, + 419 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of different methods on the 3D-OVS dataset. We visualize the segmentation results in 2 scenes. We observe that our method gives the most accurate segmentation maps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 423, + 287, + 543 + ], + "blocks": [ + { + "bbox": [ + 48, + 423, + 287, + 543 + ], + "lines": [ + { + "bbox": [ + 48, + 423, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 423, + 287, + 543 + ], + "type": "table", + "html": "
Methodbedbenchroomsofalawnoverall
LSeg [21]56.06.019.24.517.520.6
ODISE [46]52.624.152.548.339.843.5
OV-Seg [23]79.888.971.466.181.277.5
FFD [20]56.66.125.13.742.926.9
LERF [18]73.553.246.62773.754.8
3D-OVS [24]89.589.392.87488.286.8
LangSplat92.594.294.190.096.193.4
", + "image_path": "8b5a6aa3d12e2862abfca92cf9f526dec5f4da45abef4f23d6c315d89e2a24f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "text", + "content": "Table 5. Quantitative comparisons of 3D semantic segmentation on the 3D-OVS dataset. We report the mIoU scores " + }, + { + "bbox": [ + 46, + 551, + 287, + 574 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 581, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 653 + ], + "type": "text", + "content": "cluding LERF [18] and 3D-OVS [24] by a large margin. Note that in this dataset, we generate object masks only based on the query text while others, such as 3D-OVS, require the complete category list. In the end, our method achieves an overall mIoU of " + }, + { + "bbox": [ + 46, + 581, + 287, + 653 + ], + "type": "inline_equation", + "content": "93.4\\%" + }, + { + "bbox": [ + 46, + 581, + 287, + 653 + ], + "type": "text", + "content": ", which shows that LangSplat effectively learns a precise 3D language field." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Qualitative Results. We present the qualitative results in Figure 5. As LERF suffers from the patchy issue and learns over-smoothed features, it fails to find accurate object boundaries. Among all state-of-the-art methods, our methods give the most accurate segmentation maps, which" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 426, + 533, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 533, + 439 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 533, + 439 + ], + "type": "text", + "content": "further demonstrates the effectiveness of our LangSplat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 452, + 379, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 379, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 379, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": "In this paper, we have presented LangSplat, a method for constructing 3D language fields that enables precise and efficient open-vocabulary querying within 3D spaces. By extending 3D Gaussian Splatting with language features and learning a scene-specific language autoencoder, LangSplat circumvents slow rendering speed associated with NeRF-based methods. Furthermore, we propose to learn the semantic hierarchy defined by SAM, which effectively resolves the point ambiguity problem. The experimental results clearly demonstrate LangSplat's superiority over existing state-of-the-art methods like LERF, particularly in terms of its remarkable " + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "inline_equation", + "content": "199 \\times" + }, + { + "bbox": [ + 304, + 473, + 545, + 628 + ], + "type": "text", + "content": " speed improvement and enhanced performance in open-ended 3D language query tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was funded through National Key Research and Development Program of China (Project No. 2022YFB36066), in part by the Shenzhen Science and Technology Project under Grant (JCYJ20220818101001004, JSGG20210802153150005), and in part by NIH grant R01HD104969, 1U01CA284207, and NSF grant CRCNS-2309041." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20058" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In CVPR, pages 19129-19139, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 168 + ], + "type": "text", + "content": "[2] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. Contrastive lift: 3d object instance segmentation by slow-fast contrastive fusion. arXiv preprint arXiv:2306.04633, 2023. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 192 + ], + "type": "text", + "content": "[3] Nancy Bonvillain. Language, culture, and communication: The meaning of messages. Rowman & Littlefield, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 237 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 286, + 282 + ], + "type": "text", + "content": "[5] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, pages 5056-5066, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 317 + ], + "type": "text", + "content": "[6] Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Chen Yang, Wei Shen, Lingxi Xie, Xiaopeng Zhang, and Qi Tian. Segment anything in 3d with nerfs. 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 361 + ], + "type": "text", + "content": "[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 363, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 286, + 396 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCVn, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 286, + 452 + ], + "type": "text", + "content": "[9] Boyuan Chen, Fei Xia, Brian Ichter, Kanishka Rao, Keerthana Gopalakrishnan, Michael S Ryoo, Austin Stone, and Daniel Kappler. Open-vocabulary queryable scene representations for real world planning. In ICRA, pages 11509-11522. IEEE, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 286, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 286, + 485 + ], + "type": "text", + "content": "[10] Zilong Chen, Feng Wang, and Huaping Liu. Text-to-3d using gaussian splatting. arXiv preprint arXiv:2309.16585, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "text", + "content": "[11] Yangming Cheng, Lielei Li, Yuanyou Xu, Xiaodi Li, Zongxin Yang, Wenguan Wang, and Yi Yang. Segment and track anything. arXiv preprint arXiv:2305.06558, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "text", + "content": "[12] Shanghua Gao, Zhijie Lin, Xingyu Xie, Pan Zhou, MingMing Cheng, and Shuicheng Yan. Editanything: Empowering unparalleled flexibility in image editing and generation. In ACM MM, Demo track, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 610 + ], + "type": "text", + "content": "[13] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In CVPR, pages 4089-4098, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "text", + "content": "[14] Chenguang Huang, Oier Mees, Andy Zeng, and Wolfram Burgard. Visual language maps for robot navigation. In ICRA, pages 10608-10615. IEEE, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 286, + 689 + ], + "type": "text", + "content": "[15] Siyuan Huang, Zhengkai Jiang, Hao Dong, Yu Qiao, Peng Gao, and Hongsheng Li. Instruct2act: Mapping multimodality instructions to robotic actions with large language model. arXiv preprint arXiv:2305.11176, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 713 + ], + "type": "text", + "content": "[16] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):1-14, 2023. 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In ICCV, pages 19729-19739, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 229 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 262 + ], + "type": "text", + "content": "[20] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. NeurlPS, 35:23311-23330, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[21] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In ICLR, 2022. 3, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 298, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 342 + ], + "type": "text", + "content": "[22] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. NeurIPS, 35:35313-35325, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 387 + ], + "type": "text", + "content": "[23] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In CVPR, pages 7061-7070, 2023. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 431 + ], + "type": "text", + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 477 + ], + "type": "text", + "content": "[25] Youquan Liu, Lingdong Kong, Jun Cen, Runnan Chen, Wenwei Zhang, Liang Pan, Kai Chen, and Ziwei Liu. Segment any point cloud sequences by distilling vision foundation models. arXiv preprint arXiv:2306.09347, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 511 + ], + "type": "text", + "content": "[26] Zhihe Lu, Zeyu Xiao, Jiawang Bai, Zhiwei Xiong, and Xin chao Wang. Can sam boost video super-resolution? arXiv preprint arXiv:2305.06524, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 555 + ], + "type": "text", + "content": "[27] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 556, + 545, + 578 + ], + "type": "text", + "content": "[28] Jun Ma and Bo Wang. Segment anything in medical images. arXiv preprint arXiv:2304.12306, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 579, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 635 + ], + "type": "text", + "content": "[29] Gengchen Mai, Weiming Huang, Jin Sun, Suhang Song, Deepak Mishra, Ninghao Liu, Song Gao, Tianming Liu, Gao Cong, Yingjie Hu, et al. On the opportunities and challenges of foundation models for geospatial artificial intelligence. arXiv preprint arXiv:2304.06798, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[30] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin Zhang. Segment anything model for medical image analysis: an experimental study. Medical Image Analysis, 89:102918, 2023. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[31] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20059" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[32] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[33] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 206 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[35] Nur Muhammad Mahi Shafiullah, Chris Paxton, Lerrel Pinto, Soumith Chintala, and Arthur Szlam. Clip-fields: Weakly supervised semantic fields for robotic memory. arXiv preprint arXiv:2210.05663, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 285 + ], + "type": "text", + "content": "[36] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[37] Shuai Shen, Wanhua Li, Zheng Zhu, Yueqi Duan, Jie Zhou, and Jiwen Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In ECCV, pages 666-682. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 374 + ], + "type": "text", + "content": "[38] Shuai Shen, Wanhua Li, Xiaoke Huang, Zheng Zhu, Jie Zhou, and Jiwen Lu. Sd-nerf: Towards lifelike talking head animation via spatially-adaptive dual-driven nerfs. TMM, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 420 + ], + "type": "text", + "content": "[39] William Shen, Ge Yang, Alan Yu, Jansen Wong, Leslie Pack Kaelbling, and Phillip Isola. Distilled feature fields enable few-shot language-guided manipulation. arXiv preprint arXiv:2308.07931, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[40] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In CVPR, pages 9043–9052, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 499 + ], + "type": "text", + "content": "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 543 + ], + "type": "text", + "content": "[42] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[43] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 3DV, pages 443-453. IEEE, 2022. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[44] Jia Wan, Wanhua Li, Atmadeep Banerjee, Jason Ken Adhinarta, Evelina Sjostedt, Jingpeng Wu, Jeff Lichtman, Hanspeter Pfister, and Donglai Wei. Trisam: Tri-plane sam for zero-shot cortical blood vessel segmentation in vem images. arXiv preprint arXiv:2401.13961, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[45] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[46] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panop" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 510 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tic segmentation with text-to-image diffusion models. In CVPR, pages 2955-2966, 2023. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[47] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[48] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[49] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[50] Jingfeng Yao, Xinggang Wang, Lang Ye, and Wenyu Liu. Matte anything: Interactive natural image matting with segment anything models. arXiv preprint arXiv:2306.04121, 2023.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 318 + ], + "type": "text", + "content": "[51] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3d gaussian splattering with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 353 + ], + "type": "text", + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "text", + "content": "[53] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 442 + ], + "type": "text", + "content": "[54] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In ICCV, pages 15838-15847, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 444, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 444, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 308, + 444, + 545, + 476 + ], + "type": "text", + "content": "[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 478, + 545, + 510 + ], + "type": "text", + "content": "[56] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In VIS, pages 29-538. IEEE, 2001. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20060" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_content_list.json b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..139db906c9eb673a2f31934c21b64f9d96c2c2c4 --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_content_list.json @@ -0,0 +1,2128 @@ +[ + { + "type": "text", + "text": "Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding", + "text_level": 1, + "bbox": [ + 84, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jin-Chuan Shi $^{1}$ Miao Wang $^{1,2*}$ Hao-Bin Duan $^{1}$ Shao-Hua Guan $^{1}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, SCSE, Beihang University \n $^{2}$ Zhongguancun Laboratory", + "bbox": [ + 114, + 179, + 854, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0570b97684c89522e054a3189fc5c2e4fad8082300910a0b8b2b23ea1c5157e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 268, + 467, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e907397f291ccc9103f688d35ae1a586d808fb4a7e22b05f24db24727adc9887.jpg", + "image_caption": [ + "Figure 1. We present Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary querying. The top row visualizes the original image, novel view synthesis result with query relevancy and PCA of learned semantic features. The bottom row compares our method with other language-embedded representations. The right-side bar maps relevancy values to heatmap colors. Our method achieves better fidelity and query accuracy while rendering at higher frame rates." + ], + "image_footnote": [], + "bbox": [ + 78, + 407, + 274, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2632718df8f54bbf20e2cdb2f02d088daa68ab171c1b2f131a4b0e9577e411c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 407, + 467, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5081425f055e55df3075f169bfd2ffb2897a1cc5cf1ebe8adc0744d44c27f6d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 268, + 857, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e22724e9f82363b7f2140980d0676fe110fc06e28ed4eec560a448044521348.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 407, + 663, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed18a6ac3dc0f2229462af24a1ba84253a61cad9da38099b1b098eb8f27be5fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 407, + 857, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/35382f7f847764445a1edb7da8935a9346fa836ebbc31143619a7b7ee874eaa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 857, + 271, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 603, + 313, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However, their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis, directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work, we introduce Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians, we propose a dedicated quantization scheme that drastically alleviates the memory requirement, and a novel embedding procedure that", + "bbox": [ + 75, + 635, + 472, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "achieves smoother yet high accuracy query, countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations, while maintaining real-time rendering frame rates on a single desktop GPU. Project page: https://buaavrcg.github.io/LEGaussians/.", + "bbox": [ + 500, + 589, + 892, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 753, + 632, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRFs) [2, 3, 32, 57] and 3D Gaussian Splatting [20] has advanced the development of efficient and high-quality 3D scene novel view synthesis from multi-view images. Nevertheless, these representations solely encapsulate geometric and appearance details without any semantic information. To bridge this gap, language-embedded neural representations [21, 22] try to integrate semantic information from multi-view im", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 99, + 887, + 232, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5333", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages into 3D scenes for open-vocabulary querying tasks, which allows intuitive interaction with large language models (LLMs) [53, 54] and human users, and powers broad applications including scene editing, VR/AR, autonomous driving, and robotic navigation. Compared to traditional semantic labeling methods, language features from visual-language models like CLIP [36] offer more comprehensive semantic understanding capability, as they encompass objects with long-tail distribution, enhancing their suitability for real-world applications. However, accurately incorporating language embedding into current 3D scene representations, while maintaining their efficiency and visual quality, presents a significant challenge.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent techniques [21, 22, 27] extract dense language features from multi-view 2D images and incorporate additional output branches in scene representation to predict semantic features. However, the quality of semantic features heavily relies on scene representation, and trivially expanding the output channels poses significant challenges in recovering high-precision and robust semantics of the scenes. Furthermore, while 3D Gaussian [20] are efficient and fast for high-quality 3D scene representation, embedding raw semantic information into a massive number of points can cause prohibitive memory requirements and significantly lower the efficiency of both optimization and rendering.", + "bbox": [ + 75, + 287, + 472, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce Language Embedded 3D Gaussians, a semantic scene representation framework that provides both high precision and efficiency for open-vocabulary query tasks. To avoid the substantial memory requirement that would result from directly increasing the feature channels for raw semantic features in the 3D Gaussians, we employ a novel feature quantization approach that significantly decreases the computational and memory cost. By leveraging the redundancy nature of local semantic features, we construct more streamlined language features that are compactly stored on the 3D Gaussians. Additionally, we address the issue of semantic ambiguity caused by visual inconsistency from multi-view images by implementing a mechanism that lowers the spatial frequency of semantic features, guided by learned uncertainty values. This technique enables language features in the 3D Gaussians to be much smoother yet still precise.", + "bbox": [ + 75, + 470, + 472, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate that our method achieves state-of-the-art quality in both novel view synthesis and open-vocabulary querying tasks, while allowing real-time rendering on consumer-level devices.", + "bbox": [ + 76, + 729, + 468, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include:", + "bbox": [ + 94, + 791, + 356, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel quantization scheme that efficiently compresses and integrates semantic features into dense 3D Gaussians, ensuring efficient optimization and rendering on consumer devices while maintaining accurate semantic embedding.", + "- We propose a mechanism that leverages spatial position" + ], + "bbox": [ + 76, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and semantic uncertainty of 3D Gaussians to address the semantic ambiguity arising from visual inconsistency across views.", + "bbox": [ + 511, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Our method outperforms other language-embedded 3D representations, delivering state-of-the-art results in visual quality, language query precision, and rendering speed at the same time.", + "bbox": [ + 500, + 136, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 640, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering. NeRF [32] has demonstrated superior novel view synthesis quality over traditional methods [6, 10, 16, 17, 40], and various efforts have been made to enhance its performance. Although many methods focus on improving NeRF's rendering quality [2, 3, 31, 38, 43, 52, 57], they still suffer from slow training and rendering speed. On the other hand, explicit and hybrid scene representations [4, 7, 8, 14, 28, 33, 41, 49] typically utilize hash grids [33] and point clouds [49] to reduce computational cost of large neural networks. Recent 3D Gaussian [20] sets a new standard on both rendering quality and training speed, by using fast rasterization of 3D Gaussians to replace differential volume rendering. Nevertheless, directly embedding language embeddings on the dense 3D Gaussians is non-trivial, as the massive amount of 3D points requires prohibitive memory usage and causes a drastic performance drop in both training and rendering.", + "bbox": [ + 496, + 237, + 893, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language Embedded Scene Representation. Incorporating specific semantics into NeRF's implicit MLP-based representation is difficult due to challenges in accurately identifying 3D regions. Various approaches [11, 15, 21, 22, 39, 44, 46, 61] have tried to integrate semantic data into NeRF, often by adding new network branches. For instance, DFF [22] embeds a branch for language predictions and uses a pre-trained encoder for supervision, while LERF [21] assimilates CLIP [36] features from multi-scale image crops into hash grid-represented 3D scenes. 3D-OVS [26] takes a different route, creating dual decomposed 3d tensors for geometry and semantics, but it requires predefined segmented classes and doesn't support arbitrary queries. Other methods [23, 60] decompose scenes into smaller MLPs representing local semantics. However, the quality of semantic embeddings and the efficiency of training and rendering of these methods, are all constrained by the limitations of their scene representation models.", + "bbox": [ + 496, + 501, + 893, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open Vocabulary 3D Scene Understanding. Recent progress in open vocabulary scene understanding has been marked by the integration of 2D Vision-Language Models with 3D point cloud processing [18, 50, 51, 59, 62]. These approaches focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Additionally, advancements in 3D object detection and segmentation [9, 19, 25, 30, 35, 42, 56] have shown the effec", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5334", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tiveness of merging point cloud data with visual features extracted from image for scene analysis [48]. However, these methods mainly address the comprehension and analysis of existing scene representations, like point clouds, rather than optimizing scene representation from multi-view images, which is the focus of our work.", + "bbox": [ + 75, + 90, + 472, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 195, + 168, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our training process of Language Embedded 3D Gaussians, including (1) a recap of 3D Gaussian Splatting [20] (Sec. 3.1), (2) extracting dense language features from multi-view images (Sec. 3.2), (3) a quantization scheme for high-dimensional language features that creates a compact feature space for the embedding process (Sec. 3.3), (4) an embedding mechanism that alleviates semantic ambiguity caused by visual difference across views, by lowering the spatial frequency of semantic features with the learned uncertainty values (Sec. 3.4).", + "bbox": [ + 75, + 220, + 472, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Recap: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 76, + 382, + 346, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splatting [20] renders complex scenes by merging a multitude of colored 3D Gaussians, which are subsequently projected onto camera views through a rasterization process. By employing differentiable rendering and gradient descent, the attributes of these 3D Gaussians, including position $p$ , covariance $\\Sigma$ , color $c$ , and opacity $\\alpha$ , are optimized to represent the 3D scene based on a collection of input images. The result image $I$ is rendered from a specific camera pose $p_{\\mathrm{cam}}$ , using the differentiable rasterization $R$ , represented as:", + "bbox": [ + 75, + 405, + 468, + 556 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI = R (p, \\Sigma , c, \\alpha ; p _ {\\mathrm {c a m}}). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 565, + 468, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dense Language Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 590, + 390, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first extract pixel-level dense language features from visual-language models. While CLIP [36] encodes images into global language features, its direct application is not feasible for our purposes as we require pixel-level targets to learn 3D scene representations from multi-view images. Prior studies [21, 26] overcome this limitation by computing multi-scale dense CLIP features for each image. To obtain dense language embeddings from multi-view images, we employ a slightly different hierarchical random cropping technique to extract CLIP features, similar to 3DOVS [26]. We aggregate features from all layers and normalize them to produce the final language embeddings.", + "bbox": [ + 75, + 613, + 468, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, features extracted from CLIP only provide a rough boundary of different semantic regions, resulting in ambiguities and inaccuracies in the language embedding of 3D scenes. Conversely, DINO [5] exhibits autonomous object decomposition without requiring labeled data [1], as several studies [21, 22, 26, 47] have shown that DINO effectively enhances language feature grouping without reliance", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "on labels or prior knowledge. Therefore, we also extract DINO features as complementary to enhance the details of extracted language features.", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then concatenate the dense CLIP and DINO features extracted from multi-view images to form hybrid language feature maps. We denote the features at the position $(x,y)$ on the CLIP and DINO feature maps of image $I$ as $\\mathbf{F}_{I,x,y}^{\\mathrm{CLIP}}\\in \\mathbb{R}^{d_{\\mathrm{CLIP}}}$ and $\\mathbf{F}_{I,x,y}^{\\mathrm{DINO}}\\in \\mathbb{R}^{d_{\\mathrm{DINO}}}$ , respectively, where $d_{\\mathrm{CLIP}}$ and $d_{\\mathrm{DINO}}$ represent the dimension of each feature. The hybrid language feature $\\mathbf{F}_{I,x,y}\\in \\mathbb{R}^d$ is given by:", + "bbox": [ + 496, + 137, + 893, + 247 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {I, x, y} = \\mathbf {F} _ {I, x, y} ^ {\\mathrm {C L I P}} \\oplus \\mathbf {F} _ {I, x, y} ^ {\\mathrm {D I N O}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 253, + 890, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d$ is the total channels in the hybrid features.", + "bbox": [ + 498, + 282, + 836, + 299 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Quantization of Language Features", + "text_level": 1, + "bbox": [ + 498, + 311, + 807, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Direct integration of our extracted language features as in prior studies [21, 22, 26] would be infeasible for 3D Gaussians, due to the substantial memory demands of recording high-dimensional language embeddings on them, which significantly hinders rendering efficiency and limits the maximum number of 3D Gaussians that can be trained at the same time. To alleviate such storage and computation costs of raw language embeddings, our insight is to leverage the inherent redundancy in language features, as those within a single object share very similar semantics. Additionally, the semantics of a single scene merely cover only a small fraction of the original CLIP feature space, leading to unnecessary data stored in the 3D representation for all the queries we would need. Previous studies [12, 37, 45, 55] have demonstrated vector quantization can effectively compress the feature space while preserving the semantics contained within the feature space accurately and comprehensively. We propose a dedicated quantization scheme to effectively compress the language features extracted from multiple viewpoints, resulting in a more efficient and compact representation of scene-aware language features.", + "bbox": [ + 496, + 335, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, our goal is to transform the hybrid language features $\\mathbf{F} \\in \\mathbb{R}^d$ into a quantized version $\\hat{\\mathbf{F}} \\in \\mathbb{R}^d$ , which approximates $\\mathbf{F}$ . To achieve this, we develop a discrete language feature space $S = \\{\\mathbf{f}_i \\in \\mathbb{R}^d | i = 1,2,\\dots,N\\}$ , and use an integer index $m \\in \\{1,2,\\dots,N\\}$ for retrieving the nearest language feature in $S$ . The quantized language feature is then given by:", + "bbox": [ + 496, + 654, + 890, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {F}} = \\sum_ {i} ^ {N} \\mathbf {f} _ {i} \\cdot \\operatorname {o n e h o t} (m) _ {i}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 763, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{f}_i$ represents the $i$ -th basis of the language features, and onehot $(m) \\in \\mathbb{R}^N$ is the one-hot vector for index $m$ . This quantization scheme effectively removes excessive information in the original language embeddings by compressing the original continuous language feature space into discrete bases, with the compression rate adjustable through the size of space $N$ .", + "bbox": [ + 496, + 806, + 893, + 912 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/72e709372559d5db05a80fd61389fcddd322026ad12added03c5c553b953952e.jpg", + "image_caption": [ + "Figure 2. The training process for Language-embedded 3D Gaussians starts with initializing scenes following 3D Gaussian Splatting [20] and randomly initializing semantic features and setting uncertainty to zero. Dense language features from multi-view CLIP [36] and DINO [5] are quantized to create a discrete feature space and semantic indices. These attributes of the 3D Gaussians are then rendered into 2D maps using a differentiable rasterizer. The optimization is achieved through semantic and adaptive spatial smoothing loss." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 859, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantization. To quantitize high-dimensional language feature $\\mathbf{F}$ , we execute a max-similarity search within space $S$ , utilizing cosine similarity $\\cos \\langle \\cdot \\rangle$ as the distance metric:", + "bbox": [ + 75, + 430, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} (\\mathbf {F}, \\mathbf {f} _ {i}) = \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {C L I P}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle + \\lambda_ {\\mathrm {D I N O}} \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {D I N O}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {D I N O}} \\right\\rangle , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 479, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{DINO}}$ serves as a hyper-parameter that modulates the significance of DINO within the language feature.", + "bbox": [ + 75, + 513, + 468, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following VQ-VAE [45], the selected language feature index from the set $S$ is determined as $m = \\operatorname{argmax}_i(\\mathcal{D}(\\mathbf{F},\\mathbf{f}_i))$ , and the quantization of $\\mathbf{F}$ is computed using Eq. (3). The result for each image after this quantization procedure is a semantic indices map, denoted as $\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}$ .", + "bbox": [ + 75, + 544, + 468, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization. During the quantization of all language features extracted from multi-view images, the optimization of the discrete feature space $S$ is simultaneously accomplished by minimizing the cosine similarity loss between the language features $\\mathbf{F}_i$ and the quantization $\\hat{\\mathbf{F}}_{\\mathrm{i}}$ :", + "bbox": [ + 75, + 638, + 468, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\cos} \\left(\\mathbf {F} _ {i}\\right) = \\left(1 - \\cos \\left\\langle \\mathbf {F} _ {i} ^ {\\mathrm {C L I P}} \\cdot \\hat {\\mathbf {F}} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle\\right) \\tag {5} \\\\ + \\lambda_ {\\mathrm {D I N O}} (1 - \\cos \\langle \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\cdot \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\rangle). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 715, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to prevent quantization collapse and ensure maximal utilization of each feature in the feature space, we have devised a load balancing loss inspired by the design of the Switch Transformer [13]. The load balancing loss is calculated by performing an element-wise multiplication of the utilization ratio $\\mathbf{r} \\in \\mathbb{R}^N$ and the mean selection probability $\\mathbf{p} \\in \\mathbb{R}^N$ of each feature, followed by their summation:", + "bbox": [ + 75, + 762, + 470, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l b} = \\sum^ {N} (\\mathbf {r} \\circ \\mathbf {p}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 872, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\circ$ represents the element-wise product. We provide a detailed description of the load balancing loss in the supplementary material.", + "bbox": [ + 498, + 430, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, when quantizing the multi-view language features, we optimize the discrete feature space $\\mathcal{S}$ as well as the semantic indices maps $\\mathcal{M}$ using the following loss:", + "bbox": [ + 498, + 474, + 890, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {q} = \\lambda_ {c o s} \\mathcal {L} _ {c o s} + \\lambda_ {l b} \\mathcal {L} _ {l b}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 532, + 890, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Language Embedded 3D Gaussians", + "text_level": 1, + "bbox": [ + 500, + 559, + 807, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing the discrete feature space $S$ and index maps $\\mathcal{M}$ in the previous section, we embed the compressed language features into 3D Gaussians for open-vocabulary scene understanding. Concretely, we expand the number of channels of each 3D Gaussian to include a compact feature vector representing the discrete index of language feature bases. To address semantic ambiguity arising from visual disparities across various viewpoints, we introduce a novel mechanism to reduce the spatial frequency of language embeddings through an adaptive learning loss based on the learned uncertainty values on each point.", + "bbox": [ + 496, + 582, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4.1 Compact Semantic Features on 3D Gaussians", + "text_level": 1, + "bbox": [ + 500, + 756, + 870, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The inherent attributes of 3D Gaussians, such as color, are rendered onto the screen through the processes of rasterization and alpha blending (see Sec. 3.1). However, embedding discrete semantic indices $m$ onto 3D Gaussians would lead to erroneous results when optimizing through differential rendering, as these indices are not in a continuous space. Instead of directly embedding indices, we learn another continuous and compact semantic feature vectors, de", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5336", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noted as $\\mathbf{s}_G\\in \\mathbb{R}^{d_s}$ , where $d_{s}$ is a hyper-parameter for controlling the storage capability for semantics. We then render these compact semantic feature vectors into a 2D feature map with rasterization and alpha blending, and decode the 2D feature map into the discrete semantic indices $m$ using a tiny MLP decoder:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {M}} = \\operatorname {s o f t m a x} \\left(D \\left(R _ {s} \\left(\\mathcal {G}; p _ {\\mathrm {c a m}}\\right)\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 190, + 468, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{s}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times d_{s}}$ represents the semantic feature rendering from a set of 3D Gaussians $\\mathcal{G}$ viewed from camera pose $p_{\\mathrm{cam}}$ . $D$ represents the tiny MLP decoder. During training process, a softmax operation is applied to the decoder's output, yielding the language feature index distribution $\\hat{\\mathcal{M}} \\in \\mathbb{R}^{H \\times W \\times N}$ , where $H$ and $W$ denote the height and width of the image, respectively. To optimize the semantic features of the 3D Gaussians and the MLP decoder, we apply the cross-entropy loss:", + "bbox": [ + 75, + 215, + 472, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = \\operatorname {C E} (\\tilde {\\mathcal {M}}, \\mathcal {M}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 362, + 468, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}$ denotes the discrete language indices map of the image extracted during the language feature quantization process (Sec. 3.3).", + "bbox": [ + 75, + 387, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Semantic Feature Smoothing", + "text_level": 1, + "bbox": [ + 76, + 445, + 333, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to viewing angles, illuminations, existence of specular and semi-transparent materials, language features of the same spatial location in multi-view images may exhibit high variance, posing a challenge for precise semantic learning in 3D scenes. The high variance in language features at the same spatial location in multi-view images, due to differences in viewing angles, illumination, and the presence of specular and semi-transparent materials, presents challenges in accurately learning semantics in 3D scenes. Moreover, partially occluded objects may have their language features inaccurately extracted due to the lack of detection of their entirety. When these biased language features undergo quantization, the error may be amplified, resulting in a single 3D position being linked to several distinct language feature bases and, therefore, different discrete feature indices. Consequently, we introduce a smoothing strategy that limits the spatial frequency of semantic features on 3D Gaussians. This is achieved by incorporating an adaptive loss based on a learnable uncertainty value for each 3D Gaussian.", + "bbox": [ + 75, + 470, + 472, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning of uncertainty. To represent the variance associated with the semantic feature for each Gaussian, we record an estimizable semantic uncertainty on each point, denoted as $u \\in [0,1]$ . A higher $u$ suggests that the semantic feature may exhibit instability and undergo frequent changes during the optimization process. The uncertainty values are jointly optimized when training the compact semantic features $\\mathbf{s}_G$ :", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = \\frac {\\sum \\operatorname {C E} (\\hat {\\mathcal {M}} , \\mathcal {M}) \\circ (1 - R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}}))}{H \\times W}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 89, + 892, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{u}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times 1}$ is the rendered 2D uncertainty map from camera pose $p_{\\mathrm{cam}}$ . At the same time, we regularize these uncertainty values to avoid converging to a trivial solution where all 3D Gaussians have the maximum uncertainties:", + "bbox": [ + 496, + 128, + 893, + 203 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {u} = \\frac {\\sum R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}})}{H \\times W}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 209, + 892, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the total semantic loss for optimizing the compact semantic features of the 3D Gaussians and the MLP decoder is defined as:", + "bbox": [ + 496, + 247, + 893, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 305, + 892, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{u}$ is a hyper-parameter for controlling the weight of regularization. We initialize the uncertainty values to zero at the start of training, and inconsistent semantic feature indices will lead to increased $u$ during optimization.", + "bbox": [ + 496, + 332, + 892, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive spatial smoothing loss. Building on the earlier observation that spatially adjacent positions typically display similar semantic features, we utilize this prior knowledge to deliberately decrease the spatial frequency of the embedded compact semantic features in 3D Gaussians, particularly for those with high uncertainty values. We harness the inductive biases of coordinate-based MLPs, known for learning low-frequency representations of target signals, to regularize the semantic features in 3D Gaussians. We calculate the smoothed semantic features $\\mathbf{s}_{\\mathrm{MLP}} \\in \\mathbb{R}^{d_s}$ by inputting the position of each 3D Gaussian into a small MLP:", + "bbox": [ + 496, + 397, + 893, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s} _ {\\mathrm {M L P}} = \\mathrm {M L P} (\\mathrm {P E} (\\mathbf {p})), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 574, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p} \\in \\mathbb{R}^3$ represents the position of Gaussian, and PE represents the positional encoding used in NeRF [32]. We set a low frequency in the positional encoding to encourage spatial smoothness. We then apply the following loss for imposing the spatial smoothness regularization, where the degree of smoothness is adaptively controlled based on the learned uncertainty values:", + "bbox": [ + 496, + 593, + 893, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s m o}} = \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} - \\mathbf {s} _ {G} ^ {*} \\right\\| _ {2} + \\max \\left(\\mathbf {u} _ {G} ^ {*}, w _ {s}\\right) \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} ^ {*} - \\mathbf {s} _ {G} \\right\\| _ {2}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 708, + 890, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $*$ denotes the stop gradient operator, and $w_{s}$ is the minimal weight for the semantic smoothing term.", + "bbox": [ + 496, + 732, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, the combined loss for optimization of semantic features, semantic uncertainties, and MLPs is given by:", + "bbox": [ + 496, + 763, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {\\mathrm {s m o}} \\mathcal {L} _ {\\mathrm {s m o}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 805, + 890, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 829, + 720, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our method using PyTorch [34], and incorporate the CUDA kernel from 3D Gaussian Splitting [20] to speed up the rasterization rendering process. Our method", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5337", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2bd340c2adfd4dff9c49af92b75078b21dc9546fd01792cdce1330d26fdabe0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓mPA↑mP↑mIoU↑mAP↑FPS↑Memory↓Storage↓Training Time↓
DFF [22]25.3780.7120.3120.8170.1240.0910.1990.20242GB+14GB41GB184min
LERF [21]25.7490.8110.3170.8900.4750.4030.6880.0425GB+6GB320MB54min
3DOVS [26]25.7820.7330.2950.9050.5290.4580.5500.1757GB+15GB205GB158min
Ours29.8260.9010.1120.9470.7530.5780.8158911GB+12GB15MB68min
", + "bbox": [ + 81, + 88, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8332156d34483b35f46174b10d45586a24d14747e0cb26e82856265243b5ef57.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of our method with DFF [22], LeRF [21], 3DOVS [26]." + ], + "table_footnote": [], + "table_body": "
mPA↑mP↑mIoU↑mAP↑
Quantizationw/o DINO0.9270.6040.4810.676
w/o Lb0.9390.6660.5440.738
Embeddingw/o uG0.9440.7170.5760.773
w/o MLP0.9450.7150.5800.774
w/o uG & MLP0.9440.7170.5800.774
Ours-0.9470.7530.5780.815
", + "bbox": [ + 81, + 202, + 465, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of ablation experiments.", + "bbox": [ + 114, + 321, + 429, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimizes the scene's geometry and appearance with the same RGB loss following 3D Gaussian Splatting and enables adaptive density control of 3D Gaussians during the reconstruction process. We modify the CUDA kernel to enable the rendering of semantic features on the 3D Gaussians, and ensure that the optimizing of these semantic parameters does not affect original reconstruction quality. We set $d_{s} = 8$ and $w_{S} = 0.1$ in our modified 3D Gaussians, and set $\\lambda_{D} = \\lambda_{lb} = 0.5$ , and all other $\\lambda$ values are set to 1. After the phase of extracting dense semantic features, which takes about 30 minutes, our model can be trained on one RTX3090 GPU for about 1 hour. The training involves 30,000 iterations, utilizing the Adam optimizer [29], with a learning rate set to 0.001 and betas equal to (0.9, 0.999). We leave the inference details of open-vocabulary language querying and semantic relevancy calculation in the supplementary material.", + "bbox": [ + 75, + 348, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 619, + 209, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Basic Setups", + "text_level": 1, + "bbox": [ + 76, + 643, + 209, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. For a simultaneous evaluation of visual and semantic embedding quality, we select six scenes (excluding Stump) from the Mip-NeRF360 dataset [3] and manually annotate segmentation maps for each scene in the evaluation set. Each scene encompasses 180 to 320 images captured from various angles, and the evaluation set is chosen randomly with many novel-view images. Segmentation masks are annotated for primary objects in each scene.", + "bbox": [ + 75, + 667, + 467, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods and Metrics. We conduct a comparative evaluation of our method with DFF [22], LeRF [21], and 3DOVS [26], focusing on visual quality, language-embedded accuracy, rendering speed and model efficiency. To measure the visual quality in novel views, we report the PSNR, SSIM, and LPIPS [58] metrics. For the accuracy of language embedding, we measure the mean intersection", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "over union (mIoU), mean pixel accuracy (mPA), mean precision (mP), and mean average precision (mAP) based on our annotations. The rendering speed (FPS) is measured by rendering the images with language features at a consistent resolution. Additionally, model efficiency is evaluated based on CPU and GPU memory usage during training, as well as data storage requirements and training duration.", + "bbox": [ + 496, + 205, + 890, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons", + "text_level": 1, + "bbox": [ + 498, + 321, + 638, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method both qualitatively and quantitatively with DFF [22], LeRF [21], and 3DOVS [26] on our annotated Mip-NeRF360 dataset using a single RTX3090 GPU, following their default parameters but at the same resolution as our method. In addition, we compare our method with LeRF [21] quantitatively on the LeRF dataset, and the results are provided in the supplementary material.", + "bbox": [ + 496, + 344, + 890, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 9 displays a qualitative comparison of the novel view synthesis and semantic embedding results, demonstrating our method's efficacy in querying challenging objects in both indoor and outdoor scenes. Our approach notably delivers the highest visual rendering quality and query accuracy across all the tested scenes. Specifically, DFF [22] fails to identify \"asphalt ground\" in scene \"bicycle\" and \"flower\" in scene \"garden\". This may be caused by its use of LSeg [24], which is unstable to compute correct features in complex scenes. Moreover, due to predetermined query categories during training, 3DOVS [26] shows poor performance in scenes with complex objects. While LERF [21] can locate queried objects, its grid-based scene representation limits its ability to define clear boundaries. In contrast, our point-based approach supports high-frequency embedded semantic features, allowing for enhanced spatial semantic accuracy simply by incorporating more 3D Gaussians into scenes. This is facilitated by our quantization method, which substantially lowers memory costs. Additionally, our adaptive semantic smoothing technique effectively manages the high variance and ambiguity observed from different viewpoints, selectively applying spatial consistency as required.", + "bbox": [ + 496, + 455, + 890, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. Tab. 1 presents a comparison across various metrics, including novel view synthesis quality, open-vocabulary query accuracy, and computational efficiency. We report both host memory and video memory usage, as well as the disk space used for storing the learned language features. Our approach outperforms others in ren", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3beea16a26d33e09f99730e0d5b41a77ddcdf8c59197b36f7a25b82b35627c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 240, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d748edfd3cd3f59bd045d1a77a650b929453b4dd21daeb0138bc6903c7a10443.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 88, + 405, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bab43e8ab8304a39eadbceb9e36e2d0438f6356c15570d260c6d0c6f3843305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 88, + 566, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11aaf59852aa68692c18124b2bd760406f1c6a65389dee0c956f5652f49b1f8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 88, + 728, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b23130d91e34e39b7989f0534e9f4dab141d1c2b61757408b4206fecb83c5f96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 88, + 890, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e813163128efd3975fc6e18de4b4fa9147484b5c7945fea6ad290b440e9579f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 172, + 240, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53c1b0fc2c2a55285987f5d62e7efc8f62af638d2a74e96bbf8442095e06f983.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 172, + 405, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/037de452f3342c5117320faaab01dedf77847c47c4fcd46d7ab24598fdc48efc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 172, + 566, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bc9562fa882d2f53c8c20ae1020f90cc7c11f63c2c115eab513c2b22d5b0f60.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 172, + 728, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7e0669e02efc1d959a597141497c1d4a145c24f40cc97a389602f33b76ac837.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 172, + 890, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28216add99a92599650e64576905b360ccac7ba2b6958deec5f18e304d1c9391.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 258, + 240, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80544ede6198e2209383531f6712cabd3a62203c7a9f3956b61a52693aa49f48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 258, + 405, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13799b202273aa181ffc8f98293eba2efacdcf970402eca3433548ee0d449034.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 258, + 566, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8fd951d288bc09b025787757ba4b4061f0e46bb00ed294028f79fe2bcfd85086.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 258, + 728, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1876ff650db292cba9bb2f759b11be5b76cb8bd9b5f7d2d88ca7ca4c112b3bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 258, + 890, + 340 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9da1d5541d70b4d70ad675072c9a07942404d2e098992fe033053ac8846758cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 340, + 240, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4841bbc29e9ff3ba924232f06a9ab5fc837236e245628066c9cbe88a80ebadcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 340, + 405, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb20061b2727c5f0aa8242bceda6dd1e45bc87e42558ce27c0a1024922ee5c52.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 340, + 566, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b0ef5f244801cfb7dd89e338437c30a41927bf122775b15a7037c7af5a35d4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 340, + 728, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59c8227f226b756b184cbb03e2ad54bfb647cf13dbd352d5dde0064304245566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 340, + 890, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f97e957663cc5d829c31840aff6bffcddce141b5507cac0b669fd100eda4273.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 422, + 240, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebaa8957bd0a82b73a62ce68a0b6012f87f84be6a67e7246b4c18c7cf73896c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 240, + 422, + 405, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2a23a3c3991b88851ac495810f6497ee65d0b192c5d2d47349a1040851f23790.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 422, + 566, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68a6719d21c2b94a70d85fced18fd402cf7947b6b67e416fbdf22e255a5415b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 422, + 728, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0b3f0c998e79d873897da18f4ffea36eb25b15762a8f63155a556533e265985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 422, + 890, + 506 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4511d76c1f9b16a9f55c8a464f6703eb7a03609287abe4d76eeb2bf556e92cd8.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 80, + 507, + 240, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f59be792d90f005260b7eb64cbff61da4b4b427ab8be9832d3dd48675dc9901.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 240, + 507, + 405, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/289c48aab2956e1745fdba16c45d9fa770d570be5b88f17ddf8bfe0b6e2cf011.jpg", + "image_caption": [ + "DFF" + ], + "image_footnote": [], + "bbox": [ + 405, + 507, + 566, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fbfeb8890e2cb36af4f85eae78013568252e365ee28282210493d34e567ca15c.jpg", + "image_caption": [ + "LERF", + "Figure 3. Comparison of novel view synthesis and query relevance visualization. Left to right: Ground truth novel view synthesis, novel view images with relevance visualization from our method, DFF [22], LeRF [21], and 3DOVS [26]. Top to bottom: Query words \"asphalt ground\", \"bicycle\", \"jar of coconut oil\", \"flower\", \"LEGO Technic 856 Bulldozer\", and \"brown shoes\"." + ], + "image_footnote": [], + "bbox": [ + 566, + 507, + 728, + 590 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1697d935bb8456de43e540f7efab943c1f748a4b1bc64223b314a1acba3d0f90.jpg", + "image_caption": [ + "3DOVS" + ], + "image_footnote": [], + "bbox": [ + 728, + 507, + 890, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dering quality and semantic query accuracy, while also offering lower computational demands and a significant speed increase, nearly 100 times faster in inference. It's noteworthy that methods like 3DOVS and DFF require extensive memory and storage due to their use of raw language features during training. In contrast, our use of the quantization scheme facilitates the incorporation of detailed semantics into complex 3D scenes with numerous 3D Gaussians, and concurrently achieves the most efficient storage utilization among the baseline methods.", + "bbox": [ + 75, + 671, + 472, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Open-vocabulary Query", + "text_level": 1, + "bbox": [ + 76, + 832, + 302, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 illustrates the results of different open-vocabulary queries. We use a diverse range of vocabulary categories to identify objects in scenes, such as visual attribute terms", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "like \"green\", and subjective adjectives like \"cute\". Furthermore, the figure shows our method's effectiveness in identifying both large-scale objects like \"lego\" and specific components of an object, such as a \"engine\". We show more examples in the supplementary material.", + "bbox": [ + 496, + 671, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 756, + 653, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the results of ablation studies in Tab. 2 and Fig. 6. The results show that embedding uncertainty without spatial smoothing of semantic features leads to suboptimal optimization. Conversely, using MLP solely for spatial smoothing, without accounting for semantic variance across multiple views, also impedes the optimization of precise language features. However, our full model, combining uncertainty with MLP smoothing in an adaptive man", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/402888a12bc39c2a4ae6326453c04a48e7efa8264665a201f4a58a01ddd467e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 250, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90e89a48b54af345a5bfc9ce1b4e5c49f3333baa21d7255c1279a88bd5374a45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 88, + 359, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35dc12049086a5880c4f1c8f32ca34fef3badc47fc049b37e6989b6f837637d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 470, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce8d02a2cfe64a9bd217155fce36b8485c63ea44eb522bc09dd5654a9b91f2cf.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 80, + 215, + 250, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38ddaef95e1b73f0ba4894d2323dc97295bb8fde2f5fe1fc87a75aee7e367ff1.jpg", + "image_caption": [ + "Rendering Results" + ], + "image_footnote": [], + "bbox": [ + 253, + 215, + 359, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b951dfa67d089da120267d2dd2882babfb87ab8d9f23d541e8714eb1784367fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 215, + 470, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d263510b576a0f5c6dbcddbb31ded87118de7da935f15ac5cee470b0b1ad704.jpg", + "image_caption": [ + "Figure 4. Visual quality comparison of novel view synthesis results. Our method is able to recover more detailed geometry and appearance compared to other methods." + ], + "image_footnote": [], + "bbox": [ + 80, + 410, + 274, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f887ca98be12d205afd38462acb44ce1686a344cd172ba01725711e8223523f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 410, + 470, + 508 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bcac5c00a47638894c19ce06d04ae398edb665cfe168558e62fff794d42753c2.jpg", + "image_caption": [ + "\"green\"" + ], + "image_footnote": [], + "bbox": [ + 80, + 522, + 274, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18437561d7fd9116ae0df5d05f0285ba4104406e461bc54edb2c0ff9980ba66f.jpg", + "image_caption": [ + "\"grass\"" + ], + "image_footnote": [], + "bbox": [ + 276, + 522, + 470, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09465d65994fa22688be2ca6fbc13c452c79f92ab53284605eecb622b7cb8f6e.jpg", + "image_caption": [ + "\"lego\"", + "\"sheep toy\"", + "Figure 5. Images of various open-vocabulary queries." + ], + "image_footnote": [], + "bbox": [ + 80, + 638, + 274, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b430bddc68b24a48179b9897face2ea8520730ab266e7b404f1ae61a24427a7f.jpg", + "image_caption": [ + "\"engine\"", + "\"cute\"" + ], + "image_footnote": [], + "bbox": [ + 276, + 638, + 470, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ner effectively diminishes ambiguity and enhances the mean average precision (mAP) metric. Furthermore, integrating DINO features significantly improves the definition of object query boundaries. The load balancing loss, introduced during the quantization phase, results in a more utilized discrete feature space, facilitating the distinguish of objects with similar semantics, thereby boosting overall accuracy.", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8c9d528545a599a55a6dcd1bf72cc2ed755f75ce75479d2ef42208e04ad08642.jpg", + "image_caption": [ + "Figure 6. Comparison of ablation experiments." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 349, + 617, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Language Embedded 3D Gaussians, a novel scene representation designed for open-vocabulary query tasks. Our method successfully embeds quantized compact semantics features onto massive 3D Gaussians, while only maintaining minimal memory and storage requirements. To address semantic inconsistencies across different viewpoints, we propose a feature smoothing procedure that adaptively lowers the spatial frequency of embedded semantic features, guided by the uncertainty values learned on the 3D Gaussians. The result is a highly effective scene representation that not only enables high-quality novel view synthesis but also provides high accuracy in open-vocabulary querying, all achieved with modest computational resources.", + "bbox": [ + 496, + 375, + 890, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future works. Although DINO features improve object boundary detection, they fall short in pinpointing fine-grained object geometries at high resolutions when using CLIP-derived semantics. Furthermore, while vector quantization can efficiently compress scene-specific semantic feature space, it may compromise the finer-grained semantic information in the feature space due to its inherent clustering of semantic concepts. In addition, the detection of highly reflective or translucent objects, like televisions and mirrors, poses a challenge. These limitations might be overcome with more advanced visual-language models and native per-pixel semantic features. Nevertheless, our approach can be further adapted for broader open-vocabulary tasks, such as editing and generation of semantic-level scene objects.", + "bbox": [ + 496, + 601, + 892, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Project Number: 62372025 and 61932003) and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5340", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 3", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4", + "[6] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3), 2013. 2", + "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2", + "[8] Zhi Deng, Haoyao Xiao, Yining Lang, Hao Feng, and Juyong Zhang. Multi-scale hash encoding based neural geometry representation. Computational Visual Media, pages 1-18, 2024. 2", + "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[10] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 2", + "[11] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Open-set 3d scene segmentation with rendered novel views. 2023. 2", + "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3", + "[13] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 4, 1", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[15] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 2", + "[16] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 2", + "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2", + "[18] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2", + "[19] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, Joshua B. Tenenbaum, Celso Miguel de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. Robotics: Science and Systems (RSS), 2023. 2", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5", + "[21] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 3, 6, 7, 4", + "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 1, 2, 3, 6, 7, 4", + "[23] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2", + "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 6", + "[25] Gengxin Liu, Oliver van Kaick, Hui Huang, and Ruizhen Hu. Active self-training for weakly supervised 3d scene seman" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5341", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tic segmentation. Computational Visual Media, pages 1-14, 2024. 2", + "[26] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. 3d open-vocabulary segmentation with foundation models. arXiv preprint arXiv:2305.14093, 2023. 2, 3, 6, 7, 4", + "[27] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. arXiv preprint arXiv:2305.14093, 2023. 2", + "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6", + "[30] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. 2023. 2", + "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2", + "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5", + "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2", + "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5", + "[35] Songyou Peng, Kyle Genova, Chiyu \"Max\" Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4", + "[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 3", + "[38] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043–9052, 2023. 2", + "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2", + "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[42] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2", + "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2", + "[44] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 2", + "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 4", + "[46] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 2", + "[47] Yuang Wang, Xingyi He, Sida Peng, Haotong Lin, Hujun Bao, and Xiaowei Zhou. Autorecon: Automated 3d object discovery and reconstruction. In CVPR, 2023. 3", + "[48] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 3", + "[49] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2", + "[50] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171, 2022. 2", + "[51] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding, 2023. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Guo-Wei Yang, Zheng-Ning Liu, Dong-Yang Li, and Hao-Yang Peng. Jnerf: An efficient heterogeneous nerf model zoo based on jitter. Computational Visual Media, 9(2):401-404, 2023. 2", + "[53] Jianing Yang, Xuweiyi Chen, Shengyi Qian, David Fouhey, and Joyce Chai. Chat with nerf: Grounding 3d objects in neural radiance field through dialog, 2023. 2", + "[54] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F. Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent, 2023. 2", + "[55] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3", + "[56] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip, 2023. 2", + "[57] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2", + "[58] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 6", + "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 2", + "[60] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8274-8284, 2023. 2", + "[61] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 2", + "[62] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 699 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_model.json b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44c069ab32e3673beb38bf582998d65a4d1e9f80 --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_model.json @@ -0,0 +1,2840 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.086, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.18, + 0.856, + 0.236 + ], + "angle": 0, + "content": "Jin-Chuan Shi\\(^{1}\\) Miao Wang\\(^{1,2*}\\) Hao-Bin Duan\\(^{1}\\) Shao-Hua Guan\\(^{1}\\) \n\\(^{1}\\)State Key Laboratory of Virtual Reality Technology and Systems, SCSE, Beihang University \n\\(^{2}\\)Zhongguancun Laboratory" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.269, + 0.468, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.408, + 0.275, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.409, + 0.468, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.269, + 0.859, + 0.408 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.409, + 0.664, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.409, + 0.859, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.859, + 0.272, + 0.892, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.52, + 0.893, + 0.577 + ], + "angle": 0, + "content": "Figure 1. We present Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary querying. The top row visualizes the original image, novel view synthesis result with query relevancy and PCA of learned semantic features. The bottom row compares our method with other language-embedded representations. The right-side bar maps relevancy values to heatmap colors. Our method achieves better fidelity and query accuracy while rendering at higher frame rates." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.604, + 0.314, + 0.62 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.473, + 0.878 + ], + "angle": 0, + "content": "Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However, their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis, directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work, we introduce Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians, we propose a dedicated quantization scheme that drastically alleviates the memory requirement, and a novel embedding procedure that" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.59, + 0.893, + 0.726 + ], + "angle": 0, + "content": "achieves smoother yet high accuracy query, countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations, while maintaining real-time rendering frame rates on a single desktop GPU. Project page: https://buaavrcg.github.io/LEGaussians/." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.755, + 0.633, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRFs) [2, 3, 32, 57] and 3D Gaussian Splatting [20] has advanced the development of efficient and high-quality 3D scene novel view synthesis from multi-view images. Nevertheless, these representations solely encapsulate geometric and appearance details without any semantic information. To bridge this gap, language-embedded neural representations [21, 22] try to integrate semantic information from multi-view im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.233, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5333" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "ages into 3D scenes for open-vocabulary querying tasks, which allows intuitive interaction with large language models (LLMs) [53, 54] and human users, and powers broad applications including scene editing, VR/AR, autonomous driving, and robotic navigation. Compared to traditional semantic labeling methods, language features from visual-language models like CLIP [36] offer more comprehensive semantic understanding capability, as they encompass objects with long-tail distribution, enhancing their suitability for real-world applications. However, accurately incorporating language embedding into current 3D scene representations, while maintaining their efficiency and visual quality, presents a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.473, + 0.472 + ], + "angle": 0, + "content": "Recent techniques [21, 22, 27] extract dense language features from multi-view 2D images and incorporate additional output branches in scene representation to predict semantic features. However, the quality of semantic features heavily relies on scene representation, and trivially expanding the output channels poses significant challenges in recovering high-precision and robust semantics of the scenes. Furthermore, while 3D Gaussian [20] are efficient and fast for high-quality 3D scene representation, embedding raw semantic information into a massive number of points can cause prohibitive memory requirements and significantly lower the efficiency of both optimization and rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.473, + 0.729 + ], + "angle": 0, + "content": "In this paper, we introduce Language Embedded 3D Gaussians, a semantic scene representation framework that provides both high precision and efficiency for open-vocabulary query tasks. To avoid the substantial memory requirement that would result from directly increasing the feature channels for raw semantic features in the 3D Gaussians, we employ a novel feature quantization approach that significantly decreases the computational and memory cost. By leveraging the redundancy nature of local semantic features, we construct more streamlined language features that are compactly stored on the 3D Gaussians. Additionally, we address the issue of semantic ambiguity caused by visual inconsistency from multi-view images by implementing a mechanism that lowers the spatial frequency of semantic features, guided by learned uncertainty values. This technique enables language features in the 3D Gaussians to be much smoother yet still precise." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate that our method achieves state-of-the-art quality in both novel view synthesis and open-vocabulary querying tasks, while allowing real-time rendering on consumer-level devices." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.792, + 0.357, + 0.807 + ], + "angle": 0, + "content": "In summary, our contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.886 + ], + "angle": 0, + "content": "- We introduce a novel quantization scheme that efficiently compresses and integrates semantic features into dense 3D Gaussians, ensuring efficient optimization and rendering on consumer devices while maintaining accurate semantic embedding." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose a mechanism that leverages spatial position" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "and semantic uncertainty of 3D Gaussians to address the semantic ambiguity arising from visual inconsistency across views." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.894, + 0.198 + ], + "angle": 0, + "content": "- Our method outperforms other language-embedded 3D representations, delivering state-of-the-art results in visual quality, language query precision, and rendering speed at the same time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.212, + 0.642, + 0.228 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.238, + 0.895, + 0.496 + ], + "angle": 0, + "content": "Neural Rendering. NeRF [32] has demonstrated superior novel view synthesis quality over traditional methods [6, 10, 16, 17, 40], and various efforts have been made to enhance its performance. Although many methods focus on improving NeRF's rendering quality [2, 3, 31, 38, 43, 52, 57], they still suffer from slow training and rendering speed. On the other hand, explicit and hybrid scene representations [4, 7, 8, 14, 28, 33, 41, 49] typically utilize hash grids [33] and point clouds [49] to reduce computational cost of large neural networks. Recent 3D Gaussian [20] sets a new standard on both rendering quality and training speed, by using fast rasterization of 3D Gaussians to replace differential volume rendering. Nevertheless, directly embedding language embeddings on the dense 3D Gaussians is non-trivial, as the massive amount of 3D points requires prohibitive memory usage and causes a drastic performance drop in both training and rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.895, + 0.775 + ], + "angle": 0, + "content": "Language Embedded Scene Representation. Incorporating specific semantics into NeRF's implicit MLP-based representation is difficult due to challenges in accurately identifying 3D regions. Various approaches [11, 15, 21, 22, 39, 44, 46, 61] have tried to integrate semantic data into NeRF, often by adding new network branches. For instance, DFF [22] embeds a branch for language predictions and uses a pre-trained encoder for supervision, while LERF [21] assimilates CLIP [36] features from multi-scale image crops into hash grid-represented 3D scenes. 3D-OVS [26] takes a different route, creating dual decomposed 3d tensors for geometry and semantics, but it requires predefined segmented classes and doesn't support arbitrary queries. Other methods [23, 60] decompose scenes into smaller MLPs representing local semantics. However, the quality of semantic embeddings and the efficiency of training and rendering of these methods, are all constrained by the limitations of their scene representation models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Open Vocabulary 3D Scene Understanding. Recent progress in open vocabulary scene understanding has been marked by the integration of 2D Vision-Language Models with 3D point cloud processing [18, 50, 51, 59, 62]. These approaches focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Additionally, advancements in 3D object detection and segmentation [9, 19, 25, 30, 35, 42, 56] have shown the effec" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.182 + ], + "angle": 0, + "content": "tiveness of merging point cloud data with visual features extracted from image for scene analysis [48]. However, these methods mainly address the comprehension and analysis of existing scene representations, like point clouds, rather than optimizing scene representation from multi-view images, which is the focus of our work." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.169, + 0.212 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.373 + ], + "angle": 0, + "content": "In this section, we introduce our training process of Language Embedded 3D Gaussians, including (1) a recap of 3D Gaussian Splatting [20] (Sec. 3.1), (2) extracting dense language features from multi-view images (Sec. 3.2), (3) a quantization scheme for high-dimensional language features that creates a compact feature space for the embedding process (Sec. 3.3), (4) an embedding mechanism that alleviates semantic ambiguity caused by visual difference across views, by lowering the spatial frequency of semantic features with the learned uncertainty values (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.383, + 0.347, + 0.4 + ], + "angle": 0, + "content": "3.1. Recap: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.47, + 0.557 + ], + "angle": 0, + "content": "3D Gaussian Splatting [20] renders complex scenes by merging a multitude of colored 3D Gaussians, which are subsequently projected onto camera views through a rasterization process. By employing differentiable rendering and gradient descent, the attributes of these 3D Gaussians, including position \\( p \\), covariance \\( \\Sigma \\), color \\( c \\), and opacity \\( \\alpha \\), are optimized to represent the 3D scene based on a collection of input images. The result image \\( I \\) is rendered from a specific camera pose \\( p_{\\mathrm{cam}} \\), using the differentiable rasterization \\( R \\), represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.566, + 0.469, + 0.583 + ], + "angle": 0, + "content": "\\[\nI = R (p, \\Sigma , c, \\alpha ; p _ {\\mathrm {c a m}}). \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.391, + 0.607 + ], + "angle": 0, + "content": "3.2. Dense Language Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.795 + ], + "angle": 0, + "content": "We first extract pixel-level dense language features from visual-language models. While CLIP [36] encodes images into global language features, its direct application is not feasible for our purposes as we require pixel-level targets to learn 3D scene representations from multi-view images. Prior studies [21, 26] overcome this limitation by computing multi-scale dense CLIP features for each image. To obtain dense language embeddings from multi-view images, we employ a slightly different hierarchical random cropping technique to extract CLIP features, similar to 3DOVS [26]. We aggregate features from all layers and normalize them to produce the final language embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Nevertheless, features extracted from CLIP only provide a rough boundary of different semantic regions, resulting in ambiguities and inaccuracies in the language embedding of 3D scenes. Conversely, DINO [5] exhibits autonomous object decomposition without requiring labeled data [1], as several studies [21, 22, 26, 47] have shown that DINO effectively enhances language feature grouping without reliance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "on labels or prior knowledge. Therefore, we also extract DINO features as complementary to enhance the details of extracted language features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.248 + ], + "angle": 0, + "content": "We then concatenate the dense CLIP and DINO features extracted from multi-view images to form hybrid language feature maps. We denote the features at the position \\((x,y)\\) on the CLIP and DINO feature maps of image \\(I\\) as \\(\\mathbf{F}_{I,x,y}^{\\mathrm{CLIP}}\\in \\mathbb{R}^{d_{\\mathrm{CLIP}}}\\) and \\(\\mathbf{F}_{I,x,y}^{\\mathrm{DINO}}\\in \\mathbb{R}^{d_{\\mathrm{DINO}}}\\), respectively, where \\(d_{\\mathrm{CLIP}}\\) and \\(d_{\\mathrm{DINO}}\\) represent the dimension of each feature. The hybrid language feature \\(\\mathbf{F}_{I,x,y}\\in \\mathbb{R}^d\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.255, + 0.892, + 0.275 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {I, x, y} = \\mathbf {F} _ {I, x, y} ^ {\\mathrm {C L I P}} \\oplus \\mathbf {F} _ {I, x, y} ^ {\\mathrm {D I N O}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.837, + 0.3 + ], + "angle": 0, + "content": "where \\(d\\) is the total channels in the hybrid features." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.312, + 0.808, + 0.329 + ], + "angle": 0, + "content": "3.3. Quantization of Language Features" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Direct integration of our extracted language features as in prior studies [21, 22, 26] would be infeasible for 3D Gaussians, due to the substantial memory demands of recording high-dimensional language embeddings on them, which significantly hinders rendering efficiency and limits the maximum number of 3D Gaussians that can be trained at the same time. To alleviate such storage and computation costs of raw language embeddings, our insight is to leverage the inherent redundancy in language features, as those within a single object share very similar semantics. Additionally, the semantics of a single scene merely cover only a small fraction of the original CLIP feature space, leading to unnecessary data stored in the 3D representation for all the queries we would need. Previous studies [12, 37, 45, 55] have demonstrated vector quantization can effectively compress the feature space while preserving the semantics contained within the feature space accurately and comprehensively. We propose a dedicated quantization scheme to effectively compress the language features extracted from multiple viewpoints, resulting in a more efficient and compact representation of scene-aware language features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Specifically, our goal is to transform the hybrid language features \\(\\mathbf{F} \\in \\mathbb{R}^d\\) into a quantized version \\(\\hat{\\mathbf{F}} \\in \\mathbb{R}^d\\), which approximates \\(\\mathbf{F}\\). To achieve this, we develop a discrete language feature space \\(S = \\{\\mathbf{f}_i \\in \\mathbb{R}^d | i = 1,2,\\dots,N\\}\\), and use an integer index \\(m \\in \\{1,2,\\dots,N\\}\\) for retrieving the nearest language feature in \\(S\\). The quantized language feature is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.764, + 0.892, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {F}} = \\sum_ {i} ^ {N} \\mathbf {f} _ {i} \\cdot \\operatorname {o n e h o t} (m) _ {i}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.894, + 0.913 + ], + "angle": 0, + "content": "\\(\\mathbf{f}_i\\) represents the \\(i\\)-th basis of the language features, and onehot \\((m) \\in \\mathbb{R}^N\\) is the one-hot vector for index \\(m\\). This quantization scheme effectively removes excessive information in the original language embeddings by compressing the original continuous language feature space into discrete bases, with the compression rate adjustable through the size of space \\(N\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5335" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.86, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 2. The training process for Language-embedded 3D Gaussians starts with initializing scenes following 3D Gaussian Splatting [20] and randomly initializing semantic features and setting uncertainty to zero. Dense language features from multi-view CLIP [36] and DINO [5] are quantized to create a discrete feature space and semantic indices. These attributes of the 3D Gaussians are then rendered into 2D maps using a differentiable rasterizer. The optimization is achieved through semantic and adaptive spatial smoothing loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.471, + 0.476 + ], + "angle": 0, + "content": "Quantization. To quantitize high-dimensional language feature \\(\\mathbf{F}\\), we execute a max-similarity search within space \\(S\\), utilizing cosine similarity \\(\\cos \\langle \\cdot \\rangle\\) as the distance metric:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.481, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} (\\mathbf {F}, \\mathbf {f} _ {i}) = \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {C L I P}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle + \\lambda_ {\\mathrm {D I N O}} \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {D I N O}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {D I N O}} \\right\\rangle , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{DINO}}\\) serves as a hyper-parameter that modulates the significance of DINO within the language feature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.635 + ], + "angle": 0, + "content": "Following VQ-VAE [45], the selected language feature index from the set \\(S\\) is determined as \\(m = \\operatorname{argmax}_i(\\mathcal{D}(\\mathbf{F},\\mathbf{f}_i))\\), and the quantization of \\(\\mathbf{F}\\) is computed using Eq. (3). The result for each image after this quantization procedure is a semantic indices map, denoted as \\(\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.639, + 0.47, + 0.716 + ], + "angle": 0, + "content": "Optimization. During the quantization of all language features extracted from multi-view images, the optimization of the discrete feature space \\(S\\) is simultaneously accomplished by minimizing the cosine similarity loss between the language features \\(\\mathbf{F}_i\\) and the quantization \\(\\hat{\\mathbf{F}}_{\\mathrm{i}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.717, + 0.47, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\cos} \\left(\\mathbf {F} _ {i}\\right) = \\left(1 - \\cos \\left\\langle \\mathbf {F} _ {i} ^ {\\mathrm {C L I P}} \\cdot \\hat {\\mathbf {F}} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle\\right) \\tag {5} \\\\ + \\lambda_ {\\mathrm {D I N O}} (1 - \\cos \\langle \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\cdot \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\rangle). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Furthermore, to prevent quantization collapse and ensure maximal utilization of each feature in the feature space, we have devised a load balancing loss inspired by the design of the Switch Transformer [13]. The load balancing loss is calculated by performing an element-wise multiplication of the utilization ratio \\(\\mathbf{r} \\in \\mathbb{R}^N\\) and the mean selection probability \\(\\mathbf{p} \\in \\mathbb{R}^N\\) of each feature, followed by their summation:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.873, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l b} = \\sum^ {N} (\\mathbf {r} \\circ \\mathbf {p}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.431, + 0.892, + 0.476 + ], + "angle": 0, + "content": "where \\( \\circ \\) represents the element-wise product. We provide a detailed description of the load balancing loss in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.892, + 0.522 + ], + "angle": 0, + "content": "In summary, when quantizing the multi-view language features, we optimize the discrete feature space \\(\\mathcal{S}\\) as well as the semantic indices maps \\(\\mathcal{M}\\) using the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.533, + 0.892, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {q} = \\lambda_ {c o s} \\mathcal {L} _ {c o s} + \\lambda_ {l b} \\mathcal {L} _ {l b}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.808, + 0.576 + ], + "angle": 0, + "content": "3.4. Language Embedded 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Utilizing the discrete feature space \\( S \\) and index maps \\( \\mathcal{M} \\) in the previous section, we embed the compressed language features into 3D Gaussians for open-vocabulary scene understanding. Concretely, we expand the number of channels of each 3D Gaussian to include a compact feature vector representing the discrete index of language feature bases. To address semantic ambiguity arising from visual disparities across various viewpoints, we introduce a novel mechanism to reduce the spatial frequency of language embeddings through an adaptive learning loss based on the learned uncertainty values on each point." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.871, + 0.772 + ], + "angle": 0, + "content": "3.4.1 Compact Semantic Features on 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The inherent attributes of 3D Gaussians, such as color, are rendered onto the screen through the processes of rasterization and alpha blending (see Sec. 3.1). However, embedding discrete semantic indices \\( m \\) onto 3D Gaussians would lead to erroneous results when optimizing through differential rendering, as these indices are not in a continuous space. Instead of directly embedding indices, we learn another continuous and compact semantic feature vectors, de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5336" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.184 + ], + "angle": 0, + "content": "noted as \\(\\mathbf{s}_G\\in \\mathbb{R}^{d_s}\\), where \\(d_{s}\\) is a hyper-parameter for controlling the storage capability for semantics. We then render these compact semantic feature vectors into a 2D feature map with rasterization and alpha blending, and decode the 2D feature map into the discrete semantic indices \\(m\\) using a tiny MLP decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.191, + 0.47, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {M}} = \\operatorname {s o f t m a x} \\left(D \\left(R _ {s} \\left(\\mathcal {G}; p _ {\\mathrm {c a m}}\\right)\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.217, + 0.473, + 0.355 + ], + "angle": 0, + "content": "where \\( R_{s}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times d_{s}} \\) represents the semantic feature rendering from a set of 3D Gaussians \\( \\mathcal{G} \\) viewed from camera pose \\( p_{\\mathrm{cam}} \\). \\( D \\) represents the tiny MLP decoder. During training process, a softmax operation is applied to the decoder's output, yielding the language feature index distribution \\( \\hat{\\mathcal{M}} \\in \\mathbb{R}^{H \\times W \\times N} \\), where \\( H \\) and \\( W \\) denote the height and width of the image, respectively. To optimize the semantic features of the 3D Gaussians and the MLP decoder, we apply the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.363, + 0.47, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = \\operatorname {C E} (\\tilde {\\mathcal {M}}, \\mathcal {M}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.435 + ], + "angle": 0, + "content": "where \\(\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}\\) denotes the discrete language indices map of the image extracted during the language feature quantization process (Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.446, + 0.334, + 0.461 + ], + "angle": 0, + "content": "3.4.2 Semantic Feature Smoothing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.473, + 0.773 + ], + "angle": 0, + "content": "Due to viewing angles, illuminations, existence of specular and semi-transparent materials, language features of the same spatial location in multi-view images may exhibit high variance, posing a challenge for precise semantic learning in 3D scenes. The high variance in language features at the same spatial location in multi-view images, due to differences in viewing angles, illumination, and the presence of specular and semi-transparent materials, presents challenges in accurately learning semantics in 3D scenes. Moreover, partially occluded objects may have their language features inaccurately extracted due to the lack of detection of their entirety. When these biased language features undergo quantization, the error may be amplified, resulting in a single 3D position being linked to several distinct language feature bases and, therefore, different discrete feature indices. Consequently, we introduce a smoothing strategy that limits the spatial frequency of semantic features on 3D Gaussians. This is achieved by incorporating an adaptive loss based on a learnable uncertainty value for each 3D Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Learning of uncertainty. To represent the variance associated with the semantic feature for each Gaussian, we record an estimizable semantic uncertainty on each point, denoted as \\( u \\in [0,1] \\). A higher \\( u \\) suggests that the semantic feature may exhibit instability and undergo frequent changes during the optimization process. The uncertainty values are jointly optimized when training the compact semantic features \\( \\mathbf{s}_G \\):" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.09, + 0.893, + 0.123 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = \\frac {\\sum \\operatorname {C E} (\\hat {\\mathcal {M}} , \\mathcal {M}) \\circ (1 - R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}}))}{H \\times W}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.129, + 0.894, + 0.204 + ], + "angle": 0, + "content": "where \\( R_{u}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times 1} \\) is the rendered 2D uncertainty map from camera pose \\( p_{\\mathrm{cam}} \\). At the same time, we regularize these uncertainty values to avoid converging to a trivial solution where all 3D Gaussians have the maximum uncertainties:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.21, + 0.893, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {u} = \\frac {\\sum R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}})}{H \\times W}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.894, + 0.292 + ], + "angle": 0, + "content": "In summary, the total semantic loss for optimizing the compact semantic features of the 3D Gaussians and the MLP decoder is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.306, + 0.893, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.893, + 0.394 + ], + "angle": 0, + "content": "where \\(\\lambda_{u}\\) is a hyper-parameter for controlling the weight of regularization. We initialize the uncertainty values to zero at the start of training, and inconsistent semantic feature indices will lead to increased \\(u\\) during optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Adaptive spatial smoothing loss. Building on the earlier observation that spatially adjacent positions typically display similar semantic features, we utilize this prior knowledge to deliberately decrease the spatial frequency of the embedded compact semantic features in 3D Gaussians, particularly for those with high uncertainty values. We harness the inductive biases of coordinate-based MLPs, known for learning low-frequency representations of target signals, to regularize the semantic features in 3D Gaussians. We calculate the smoothed semantic features \\(\\mathbf{s}_{\\mathrm{MLP}} \\in \\mathbb{R}^{d_s}\\) by inputting the position of each 3D Gaussian into a small MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.575, + 0.892, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s} _ {\\mathrm {M L P}} = \\mathrm {M L P} (\\mathrm {P E} (\\mathbf {p})), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.894, + 0.701 + ], + "angle": 0, + "content": "where \\(\\mathbf{p} \\in \\mathbb{R}^3\\) represents the position of Gaussian, and PE represents the positional encoding used in NeRF [32]. We set a low frequency in the positional encoding to encourage spatial smoothness. We then apply the following loss for imposing the spatial smoothness regularization, where the degree of smoothness is adaptively controlled based on the learned uncertainty values:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.709, + 0.892, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s m o}} = \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} - \\mathbf {s} _ {G} ^ {*} \\right\\| _ {2} + \\max \\left(\\mathbf {u} _ {G} ^ {*}, w _ {s}\\right) \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} ^ {*} - \\mathbf {s} _ {G} \\right\\| _ {2}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where the \\(*\\) denotes the stop gradient operator, and \\(w_{s}\\) is the minimal weight for the semantic smoothing term." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.807 + ], + "angle": 0, + "content": "To summarize, the combined loss for optimization of semantic features, semantic uncertainties, and MLPs is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.806, + 0.892, + 0.821 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {\\mathrm {s m o}} \\mathcal {L} _ {\\mathrm {s m o}}. \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.83, + 0.722, + 0.848 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We implement our method using PyTorch [34], and incorporate the CUDA kernel from 3D Gaussian Splitting [20] to speed up the rasterization rendering process. Our method" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5337" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.891, + 0.169 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓mPA↑mP↑mIoU↑mAP↑FPS↑Memory↓Storage↓Training Time↓
DFF [22]25.3780.7120.3120.8170.1240.0910.1990.20242GB+14GB41GB184min
LERF [21]25.7490.8110.3170.8900.4750.4030.6880.0425GB+6GB320MB54min
3DOVS [26]25.7820.7330.2950.9050.5290.4580.5500.1757GB+15GB205GB158min
Ours29.8260.9010.1120.9470.7530.5780.8158911GB+12GB15MB68min
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.179, + 0.748, + 0.194 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of our method with DFF [22], LeRF [21], 3DOVS [26]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.203, + 0.466, + 0.311 + ], + "angle": 0, + "content": "
mPA↑mP↑mIoU↑mAP↑
Quantizationw/o DINO0.9270.6040.4810.676
w/o Lb0.9390.6660.5440.738
Embeddingw/o uG0.9440.7170.5760.773
w/o MLP0.9450.7150.5800.774
w/o uG & MLP0.9440.7170.5800.774
Ours-0.9470.7530.5780.815
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.322, + 0.431, + 0.336 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.47, + 0.606 + ], + "angle": 0, + "content": "optimizes the scene's geometry and appearance with the same RGB loss following 3D Gaussian Splatting and enables adaptive density control of 3D Gaussians during the reconstruction process. We modify the CUDA kernel to enable the rendering of semantic features on the 3D Gaussians, and ensure that the optimizing of these semantic parameters does not affect original reconstruction quality. We set \\( d_{s} = 8 \\) and \\( w_{S} = 0.1 \\) in our modified 3D Gaussians, and set \\( \\lambda_{D} = \\lambda_{lb} = 0.5 \\), and all other \\( \\lambda \\) values are set to 1. After the phase of extracting dense semantic features, which takes about 30 minutes, our model can be trained on one RTX3090 GPU for about 1 hour. The training involves 30,000 iterations, utilizing the Adam optimizer [29], with a learning rate set to 0.001 and betas equal to (0.9, 0.999). We leave the inference details of open-vocabulary language querying and semantic relevancy calculation in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.62, + 0.21, + 0.637 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.645, + 0.21, + 0.661 + ], + "angle": 0, + "content": "5.1. Basic Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.468, + 0.789 + ], + "angle": 0, + "content": "Dataset. For a simultaneous evaluation of visual and semantic embedding quality, we select six scenes (excluding Stump) from the Mip-NeRF360 dataset [3] and manually annotate segmentation maps for each scene in the evaluation set. Each scene encompasses 180 to 320 images captured from various angles, and the evaluation set is chosen randomly with many novel-view images. Segmentation masks are annotated for primary objects in each scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Baseline Methods and Metrics. We conduct a comparative evaluation of our method with DFF [22], LeRF [21], and 3DOVS [26], focusing on visual quality, language-embedded accuracy, rendering speed and model efficiency. To measure the visual quality in novel views, we report the PSNR, SSIM, and LPIPS [58] metrics. For the accuracy of language embedding, we measure the mean intersection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.206, + 0.892, + 0.312 + ], + "angle": 0, + "content": "over union (mIoU), mean pixel accuracy (mPA), mean precision (mP), and mean average precision (mAP) based on our annotations. The rendering speed (FPS) is measured by rendering the images with language features at a consistent resolution. Additionally, model efficiency is evaluated based on CPU and GPU memory usage during training, as well as data storage requirements and training duration." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.322, + 0.64, + 0.338 + ], + "angle": 0, + "content": "5.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.452 + ], + "angle": 0, + "content": "We compare our method both qualitatively and quantitatively with DFF [22], LeRF [21], and 3DOVS [26] on our annotated Mip-NeRF360 dataset using a single RTX3090 GPU, following their default parameters but at the same resolution as our method. In addition, we compare our method with LeRF [21] quantitatively on the LeRF dataset, and the results are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.805 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 9 displays a qualitative comparison of the novel view synthesis and semantic embedding results, demonstrating our method's efficacy in querying challenging objects in both indoor and outdoor scenes. Our approach notably delivers the highest visual rendering quality and query accuracy across all the tested scenes. Specifically, DFF [22] fails to identify \"asphalt ground\" in scene \"bicycle\" and \"flower\" in scene \"garden\". This may be caused by its use of LSeg [24], which is unstable to compute correct features in complex scenes. Moreover, due to predetermined query categories during training, 3DOVS [26] shows poor performance in scenes with complex objects. While LERF [21] can locate queried objects, its grid-based scene representation limits its ability to define clear boundaries. In contrast, our point-based approach supports high-frequency embedded semantic features, allowing for enhanced spatial semantic accuracy simply by incorporating more 3D Gaussians into scenes. This is facilitated by our quantization method, which substantially lowers memory costs. Additionally, our adaptive semantic smoothing technique effectively manages the high variance and ambiguity observed from different viewpoints, selectively applying spatial consistency as required." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. Tab. 1 presents a comparison across various metrics, including novel view synthesis quality, open-vocabulary query accuracy, and computational efficiency. We report both host memory and video memory usage, as well as the disk space used for storing the learned language features. Our approach outperforms others in ren" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5338" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.241, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.089, + 0.406, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.089, + 0.567, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.089, + 0.729, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.089, + 0.892, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.174, + 0.241, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.174, + 0.406, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.174, + 0.567, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.174, + 0.729, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.174, + 0.892, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.259, + 0.241, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.259, + 0.406, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.259, + 0.567, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.259, + 0.729, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.259, + 0.892, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.342, + 0.241, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.342, + 0.406, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.342, + 0.567, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.342, + 0.729, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.342, + 0.892, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.424, + 0.241, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.424, + 0.406, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.424, + 0.567, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.424, + 0.729, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.424, + 0.892, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.508, + 0.241, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.594, + 0.208, + 0.607 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.241, + 0.508, + 0.406, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.594, + 0.342, + 0.606 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.508, + 0.567, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.594, + 0.5, + 0.606 + ], + "angle": 0, + "content": "DFF" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.508, + 0.729, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.594, + 0.665, + 0.606 + ], + "angle": 0, + "content": "LERF" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.508, + 0.892, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.594, + 0.837, + 0.606 + ], + "angle": 0, + "content": "3DOVS" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.618, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Figure 3. Comparison of novel view synthesis and query relevance visualization. Left to right: Ground truth novel view synthesis, novel view images with relevance visualization from our method, DFF [22], LeRF [21], and 3DOVS [26]. Top to bottom: Query words \"asphalt ground\", \"bicycle\", \"jar of coconut oil\", \"flower\", \"LEGO Technic 856 Bulldozer\", and \"brown shoes\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.473, + 0.822 + ], + "angle": 0, + "content": "dering quality and semantic query accuracy, while also offering lower computational demands and a significant speed increase, nearly 100 times faster in inference. It's noteworthy that methods like 3DOVS and DFF require extensive memory and storage due to their use of raw language features during training. In contrast, our use of the quantization scheme facilitates the incorporation of detailed semantics into complex 3D scenes with numerous 3D Gaussians, and concurrently achieves the most efficient storage utilization among the baseline methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.303, + 0.85 + ], + "angle": 0, + "content": "5.3. Open-vocabulary Query" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 illustrates the results of different open-vocabulary queries. We use a diverse range of vocabulary categories to identify objects in scenes, such as visual attribute terms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.749 + ], + "angle": 0, + "content": "like \"green\", and subjective adjectives like \"cute\". Furthermore, the figure shows our method's effectiveness in identifying both large-scale objects like \"lego\" and specific components of an object, such as a \"engine\". We show more examples in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.655, + 0.774 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We demonstrate the results of ablation studies in Tab. 2 and Fig. 6. The results show that embedding uncertainty without spatial smoothing of semantic features leads to suboptimal optimization. Conversely, using MLP solely for spatial smoothing, without accounting for semantic variance across multiple views, also impedes the optimization of precise language features. However, our full model, combining uncertainty with MLP smoothing in an adaptive man" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.251, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.089, + 0.361, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.089, + 0.471, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.217, + 0.251, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.34, + 0.211, + 0.352 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.217, + 0.361, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.217, + 0.471, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.34, + 0.424, + 0.353 + ], + "angle": 0, + "content": "Rendering Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.364, + 0.47, + 0.406 + ], + "angle": 0, + "content": "Figure 4. Visual quality comparison of novel view synthesis results. Our method is able to recover more detailed geometry and appearance compared to other methods." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.411, + 0.276, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.511, + 0.204, + 0.523 + ], + "angle": 0, + "content": "\"green\"" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.411, + 0.471, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.511, + 0.4, + 0.523 + ], + "angle": 0, + "content": "\"grass\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.523, + 0.276, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.627, + 0.198, + 0.639 + ], + "angle": 0, + "content": "\"lego\"" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.523, + 0.471, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.627, + 0.404, + 0.639 + ], + "angle": 0, + "content": "\"engine\"" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.64, + 0.276, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.743, + 0.216, + 0.755 + ], + "angle": 0, + "content": "\"sheep toy\"" + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.64, + 0.471, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.743, + 0.396, + 0.754 + ], + "angle": 0, + "content": "\"cute\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.768, + 0.432, + 0.782 + ], + "angle": 0, + "content": "Figure 5. Images of various open-vocabulary queries." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "ner effectively diminishes ambiguity and enhances the mean average precision (mAP) metric. Furthermore, integrating DINO features significantly improves the definition of object query boundaries. The load balancing loss, introduced during the quantization phase, results in a more utilized discrete feature space, facilitating the distinguish of objects with similar semantics, thereby boosting overall accuracy." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.318, + 0.835, + 0.333 + ], + "angle": 0, + "content": "Figure 6. Comparison of ablation experiments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.351, + 0.619, + 0.366 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.588 + ], + "angle": 0, + "content": "We present Language Embedded 3D Gaussians, a novel scene representation designed for open-vocabulary query tasks. Our method successfully embeds quantized compact semantics features onto massive 3D Gaussians, while only maintaining minimal memory and storage requirements. To address semantic inconsistencies across different viewpoints, we propose a feature smoothing procedure that adaptively lowers the spatial frequency of embedded semantic features, guided by the uncertainty values learned on the 3D Gaussians. The result is a highly effective scene representation that not only enables high-quality novel view synthesis but also provides high accuracy in open-vocabulary querying, all achieved with modest computational resources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.893, + 0.83 + ], + "angle": 0, + "content": "Limitations and Future works. Although DINO features improve object boundary detection, they fall short in pinpointing fine-grained object geometries at high resolutions when using CLIP-derived semantics. Furthermore, while vector quantization can efficiently compress scene-specific semantic feature space, it may compromise the finer-grained semantic information in the feature space due to its inherent clustering of semantic concepts. In addition, the detection of highly reflective or translucent objects, like televisions and mirrors, poses a challenge. These limitations might be overcome with more advanced visual-language models and native per-pixel semantic features. Nevertheless, our approach can be further adapted for broader open-vocabulary tasks, such as editing and generation of semantic-level scene objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Project Number: 62372025 and 61932003) and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5340" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.493 + ], + "angle": 0, + "content": "[6] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.47, + 0.548 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[8] Zhi Deng, Haoyao Xiao, Yining Lang, Hao Feng, and Juyong Zhang. Multi-scale hash encoding based neural geometry representation. Computational Visual Media, pages 1-18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.675 + ], + "angle": 0, + "content": "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[10] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[11] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Open-set 3d scene segmentation with rendered novel views. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[15] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[16] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[18] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[19] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, Joshua B. Tenenbaum, Celso Miguel de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. Robotics: Science and Systems (RSS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[21] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 3, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 1, 2, 3, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Gengxin Liu, Oliver van Kaick, Hui Huang, and Ruizhen Hu. Active self-training for weakly supervised 3d scene seman" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5341" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "tic segmentation. Computational Visual Media, pages 1-14, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[26] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. 3d open-vocabulary segmentation with foundation models. arXiv preprint arXiv:2305.14093, 2023. 2, 3, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.259 + ], + "angle": 0, + "content": "[27] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. arXiv preprint arXiv:2305.14093, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.468, + 0.344 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[30] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[35] Songyou Peng, Kyle Genova, Chiyu \"Max\" Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[39] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043–9052, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[42] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[44] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[46] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[47] Yuang Wang, Xingyi He, Sida Peng, Haotong Lin, Hujun Bao, and Xiaowei Zhou. Autorecon: Automated 3d object discovery and reconstruction. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[48] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[49] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[50] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[51] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[52] Guo-Wei Yang, Zheng-Ning Liu, Dong-Yang Li, and Hao-Yang Peng. Jnerf: An efficient heterogeneous nerf model zoo based on jitter. Computational Visual Media, 9(2):401-404, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[53] Jianing Yang, Xuweiyi Chen, Shengyi Qian, David Fouhey, and Joyce Chai. Chat with nerf: Grounding 3d objects in neural radiance field through dialog, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[54] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F. Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[55] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.347 + ], + "angle": 0, + "content": "[56] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[57] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.432 + ], + "angle": 0, + "content": "[58] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.434, + 0.468, + 0.488 + ], + "angle": 0, + "content": "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.49, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[60] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8274-8284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.468, + 0.643 + ], + "angle": 0, + "content": "[61] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.646, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[62] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5343" + } + ] +] \ No newline at end of file diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_origin.pdf b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae4eef67d90bfa9a32f58e62b1e157047999fe31 --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/887164f8-5ac1-4491-be92-996acbb7aa0b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9aca593b30950534b8be5f2c9152b8da41a542c8090b09e7064f559a37c2aef9 +size 8354600 diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/full.md b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c71ca8884c0b743671f48ecb9b1d88d264a6de0 --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/full.md @@ -0,0 +1,423 @@ +# Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding + +Jin-Chuan Shi $^{1}$ Miao Wang $^{1,2*}$ Hao-Bin Duan $^{1}$ Shao-Hua Guan $^{1}$ $^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, SCSE, Beihang University + $^{2}$ Zhongguancun Laboratory + +![](images/c0570b97684c89522e054a3189fc5c2e4fad8082300910a0b8b2b23ea1c5157e.jpg) + +![](images/e907397f291ccc9103f688d35ae1a586d808fb4a7e22b05f24db24727adc9887.jpg) +Figure 1. We present Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary querying. The top row visualizes the original image, novel view synthesis result with query relevancy and PCA of learned semantic features. The bottom row compares our method with other language-embedded representations. The right-side bar maps relevancy values to heatmap colors. Our method achieves better fidelity and query accuracy while rendering at higher frame rates. + +![](images/2632718df8f54bbf20e2cdb2f02d088daa68ab171c1b2f131a4b0e9577e411c8.jpg) + +![](images/5081425f055e55df3075f169bfd2ffb2897a1cc5cf1ebe8adc0744d44c27f6d2.jpg) + +![](images/5e22724e9f82363b7f2140980d0676fe110fc06e28ed4eec560a448044521348.jpg) + +![](images/ed18a6ac3dc0f2229462af24a1ba84253a61cad9da38099b1b098eb8f27be5fd.jpg) + +![](images/35382f7f847764445a1edb7da8935a9346fa836ebbc31143619a7b7ee874eaa0.jpg) + +# Abstract + +Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However, their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis, directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work, we introduce Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians, we propose a dedicated quantization scheme that drastically alleviates the memory requirement, and a novel embedding procedure that + +achieves smoother yet high accuracy query, countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations, while maintaining real-time rendering frame rates on a single desktop GPU. Project page: https://buaavrcg.github.io/LEGaussians/. + +# 1. Introduction + +Neural Radiance Field (NeRFs) [2, 3, 32, 57] and 3D Gaussian Splatting [20] has advanced the development of efficient and high-quality 3D scene novel view synthesis from multi-view images. Nevertheless, these representations solely encapsulate geometric and appearance details without any semantic information. To bridge this gap, language-embedded neural representations [21, 22] try to integrate semantic information from multi-view im + +ages into 3D scenes for open-vocabulary querying tasks, which allows intuitive interaction with large language models (LLMs) [53, 54] and human users, and powers broad applications including scene editing, VR/AR, autonomous driving, and robotic navigation. Compared to traditional semantic labeling methods, language features from visual-language models like CLIP [36] offer more comprehensive semantic understanding capability, as they encompass objects with long-tail distribution, enhancing their suitability for real-world applications. However, accurately incorporating language embedding into current 3D scene representations, while maintaining their efficiency and visual quality, presents a significant challenge. + +Recent techniques [21, 22, 27] extract dense language features from multi-view 2D images and incorporate additional output branches in scene representation to predict semantic features. However, the quality of semantic features heavily relies on scene representation, and trivially expanding the output channels poses significant challenges in recovering high-precision and robust semantics of the scenes. Furthermore, while 3D Gaussian [20] are efficient and fast for high-quality 3D scene representation, embedding raw semantic information into a massive number of points can cause prohibitive memory requirements and significantly lower the efficiency of both optimization and rendering. + +In this paper, we introduce Language Embedded 3D Gaussians, a semantic scene representation framework that provides both high precision and efficiency for open-vocabulary query tasks. To avoid the substantial memory requirement that would result from directly increasing the feature channels for raw semantic features in the 3D Gaussians, we employ a novel feature quantization approach that significantly decreases the computational and memory cost. By leveraging the redundancy nature of local semantic features, we construct more streamlined language features that are compactly stored on the 3D Gaussians. Additionally, we address the issue of semantic ambiguity caused by visual inconsistency from multi-view images by implementing a mechanism that lowers the spatial frequency of semantic features, guided by learned uncertainty values. This technique enables language features in the 3D Gaussians to be much smoother yet still precise. + +Our extensive experiments demonstrate that our method achieves state-of-the-art quality in both novel view synthesis and open-vocabulary querying tasks, while allowing real-time rendering on consumer-level devices. + +In summary, our contributions include: + +- We introduce a novel quantization scheme that efficiently compresses and integrates semantic features into dense 3D Gaussians, ensuring efficient optimization and rendering on consumer devices while maintaining accurate semantic embedding. +- We propose a mechanism that leverages spatial position + +and semantic uncertainty of 3D Gaussians to address the semantic ambiguity arising from visual inconsistency across views. + +- Our method outperforms other language-embedded 3D representations, delivering state-of-the-art results in visual quality, language query precision, and rendering speed at the same time. + +# 2. Related Work + +Neural Rendering. NeRF [32] has demonstrated superior novel view synthesis quality over traditional methods [6, 10, 16, 17, 40], and various efforts have been made to enhance its performance. Although many methods focus on improving NeRF's rendering quality [2, 3, 31, 38, 43, 52, 57], they still suffer from slow training and rendering speed. On the other hand, explicit and hybrid scene representations [4, 7, 8, 14, 28, 33, 41, 49] typically utilize hash grids [33] and point clouds [49] to reduce computational cost of large neural networks. Recent 3D Gaussian [20] sets a new standard on both rendering quality and training speed, by using fast rasterization of 3D Gaussians to replace differential volume rendering. Nevertheless, directly embedding language embeddings on the dense 3D Gaussians is non-trivial, as the massive amount of 3D points requires prohibitive memory usage and causes a drastic performance drop in both training and rendering. + +Language Embedded Scene Representation. Incorporating specific semantics into NeRF's implicit MLP-based representation is difficult due to challenges in accurately identifying 3D regions. Various approaches [11, 15, 21, 22, 39, 44, 46, 61] have tried to integrate semantic data into NeRF, often by adding new network branches. For instance, DFF [22] embeds a branch for language predictions and uses a pre-trained encoder for supervision, while LERF [21] assimilates CLIP [36] features from multi-scale image crops into hash grid-represented 3D scenes. 3D-OVS [26] takes a different route, creating dual decomposed 3d tensors for geometry and semantics, but it requires predefined segmented classes and doesn't support arbitrary queries. Other methods [23, 60] decompose scenes into smaller MLPs representing local semantics. However, the quality of semantic embeddings and the efficiency of training and rendering of these methods, are all constrained by the limitations of their scene representation models. + +Open Vocabulary 3D Scene Understanding. Recent progress in open vocabulary scene understanding has been marked by the integration of 2D Vision-Language Models with 3D point cloud processing [18, 50, 51, 59, 62]. These approaches focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Additionally, advancements in 3D object detection and segmentation [9, 19, 25, 30, 35, 42, 56] have shown the effec + +tiveness of merging point cloud data with visual features extracted from image for scene analysis [48]. However, these methods mainly address the comprehension and analysis of existing scene representations, like point clouds, rather than optimizing scene representation from multi-view images, which is the focus of our work. + +# 3. Method + +In this section, we introduce our training process of Language Embedded 3D Gaussians, including (1) a recap of 3D Gaussian Splatting [20] (Sec. 3.1), (2) extracting dense language features from multi-view images (Sec. 3.2), (3) a quantization scheme for high-dimensional language features that creates a compact feature space for the embedding process (Sec. 3.3), (4) an embedding mechanism that alleviates semantic ambiguity caused by visual difference across views, by lowering the spatial frequency of semantic features with the learned uncertainty values (Sec. 3.4). + +# 3.1. Recap: 3D Gaussian Splitting + +3D Gaussian Splatting [20] renders complex scenes by merging a multitude of colored 3D Gaussians, which are subsequently projected onto camera views through a rasterization process. By employing differentiable rendering and gradient descent, the attributes of these 3D Gaussians, including position $p$ , covariance $\Sigma$ , color $c$ , and opacity $\alpha$ , are optimized to represent the 3D scene based on a collection of input images. The result image $I$ is rendered from a specific camera pose $p_{\mathrm{cam}}$ , using the differentiable rasterization $R$ , represented as: + +$$ +I = R (p, \Sigma , c, \alpha ; p _ {\mathrm {c a m}}). \tag {1} +$$ + +# 3.2. Dense Language Feature Extraction + +We first extract pixel-level dense language features from visual-language models. While CLIP [36] encodes images into global language features, its direct application is not feasible for our purposes as we require pixel-level targets to learn 3D scene representations from multi-view images. Prior studies [21, 26] overcome this limitation by computing multi-scale dense CLIP features for each image. To obtain dense language embeddings from multi-view images, we employ a slightly different hierarchical random cropping technique to extract CLIP features, similar to 3DOVS [26]. We aggregate features from all layers and normalize them to produce the final language embeddings. + +Nevertheless, features extracted from CLIP only provide a rough boundary of different semantic regions, resulting in ambiguities and inaccuracies in the language embedding of 3D scenes. Conversely, DINO [5] exhibits autonomous object decomposition without requiring labeled data [1], as several studies [21, 22, 26, 47] have shown that DINO effectively enhances language feature grouping without reliance + +on labels or prior knowledge. Therefore, we also extract DINO features as complementary to enhance the details of extracted language features. + +We then concatenate the dense CLIP and DINO features extracted from multi-view images to form hybrid language feature maps. We denote the features at the position $(x,y)$ on the CLIP and DINO feature maps of image $I$ as $\mathbf{F}_{I,x,y}^{\mathrm{CLIP}}\in \mathbb{R}^{d_{\mathrm{CLIP}}}$ and $\mathbf{F}_{I,x,y}^{\mathrm{DINO}}\in \mathbb{R}^{d_{\mathrm{DINO}}}$ , respectively, where $d_{\mathrm{CLIP}}$ and $d_{\mathrm{DINO}}$ represent the dimension of each feature. The hybrid language feature $\mathbf{F}_{I,x,y}\in \mathbb{R}^d$ is given by: + +$$ +\mathbf {F} _ {I, x, y} = \mathbf {F} _ {I, x, y} ^ {\mathrm {C L I P}} \oplus \mathbf {F} _ {I, x, y} ^ {\mathrm {D I N O}}, \tag {2} +$$ + +where $d$ is the total channels in the hybrid features. + +# 3.3. Quantization of Language Features + +Direct integration of our extracted language features as in prior studies [21, 22, 26] would be infeasible for 3D Gaussians, due to the substantial memory demands of recording high-dimensional language embeddings on them, which significantly hinders rendering efficiency and limits the maximum number of 3D Gaussians that can be trained at the same time. To alleviate such storage and computation costs of raw language embeddings, our insight is to leverage the inherent redundancy in language features, as those within a single object share very similar semantics. Additionally, the semantics of a single scene merely cover only a small fraction of the original CLIP feature space, leading to unnecessary data stored in the 3D representation for all the queries we would need. Previous studies [12, 37, 45, 55] have demonstrated vector quantization can effectively compress the feature space while preserving the semantics contained within the feature space accurately and comprehensively. We propose a dedicated quantization scheme to effectively compress the language features extracted from multiple viewpoints, resulting in a more efficient and compact representation of scene-aware language features. + +Specifically, our goal is to transform the hybrid language features $\mathbf{F} \in \mathbb{R}^d$ into a quantized version $\hat{\mathbf{F}} \in \mathbb{R}^d$ , which approximates $\mathbf{F}$ . To achieve this, we develop a discrete language feature space $S = \{\mathbf{f}_i \in \mathbb{R}^d | i = 1,2,\dots,N\}$ , and use an integer index $m \in \{1,2,\dots,N\}$ for retrieving the nearest language feature in $S$ . The quantized language feature is then given by: + +$$ +\hat {\mathbf {F}} = \sum_ {i} ^ {N} \mathbf {f} _ {i} \cdot \operatorname {o n e h o t} (m) _ {i}. \tag {3} +$$ + +$\mathbf{f}_i$ represents the $i$ -th basis of the language features, and onehot $(m) \in \mathbb{R}^N$ is the one-hot vector for index $m$ . This quantization scheme effectively removes excessive information in the original language embeddings by compressing the original continuous language feature space into discrete bases, with the compression rate adjustable through the size of space $N$ . + +![](images/72e709372559d5db05a80fd61389fcddd322026ad12added03c5c553b953952e.jpg) +Figure 2. The training process for Language-embedded 3D Gaussians starts with initializing scenes following 3D Gaussian Splatting [20] and randomly initializing semantic features and setting uncertainty to zero. Dense language features from multi-view CLIP [36] and DINO [5] are quantized to create a discrete feature space and semantic indices. These attributes of the 3D Gaussians are then rendered into 2D maps using a differentiable rasterizer. The optimization is achieved through semantic and adaptive spatial smoothing loss. + +Quantization. To quantitize high-dimensional language feature $\mathbf{F}$ , we execute a max-similarity search within space $S$ , utilizing cosine similarity $\cos \langle \cdot \rangle$ as the distance metric: + +$$ +\mathcal {D} (\mathbf {F}, \mathbf {f} _ {i}) = \cos \left\langle \mathbf {F} ^ {\mathrm {C L I P}} \cdot \mathbf {f} _ {i} ^ {\mathrm {C L I P}} \right\rangle + \lambda_ {\mathrm {D I N O}} \cos \left\langle \mathbf {F} ^ {\mathrm {D I N O}} \cdot \mathbf {f} _ {i} ^ {\mathrm {D I N O}} \right\rangle , \tag {4} +$$ + +where $\lambda_{\mathrm{DINO}}$ serves as a hyper-parameter that modulates the significance of DINO within the language feature. + +Following VQ-VAE [45], the selected language feature index from the set $S$ is determined as $m = \operatorname{argmax}_i(\mathcal{D}(\mathbf{F},\mathbf{f}_i))$ , and the quantization of $\mathbf{F}$ is computed using Eq. (3). The result for each image after this quantization procedure is a semantic indices map, denoted as $\mathcal{M} \in \mathbb{R}^{H \times W \times 1}$ . + +Optimization. During the quantization of all language features extracted from multi-view images, the optimization of the discrete feature space $S$ is simultaneously accomplished by minimizing the cosine similarity loss between the language features $\mathbf{F}_i$ and the quantization $\hat{\mathbf{F}}_{\mathrm{i}}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\cos} \left(\mathbf {F} _ {i}\right) = \left(1 - \cos \left\langle \mathbf {F} _ {i} ^ {\mathrm {C L I P}} \cdot \hat {\mathbf {F}} _ {i} ^ {\mathrm {C L I P}} \right\rangle\right) \tag {5} \\ + \lambda_ {\mathrm {D I N O}} (1 - \cos \langle \mathbf {F} _ {i} ^ {\mathrm {D D N O}} \cdot \mathbf {F} _ {i} ^ {\mathrm {D D N O}} \rangle). \\ \end{array} +$$ + +Furthermore, to prevent quantization collapse and ensure maximal utilization of each feature in the feature space, we have devised a load balancing loss inspired by the design of the Switch Transformer [13]. The load balancing loss is calculated by performing an element-wise multiplication of the utilization ratio $\mathbf{r} \in \mathbb{R}^N$ and the mean selection probability $\mathbf{p} \in \mathbb{R}^N$ of each feature, followed by their summation: + +$$ +\mathcal {L} _ {l b} = \sum^ {N} (\mathbf {r} \circ \mathbf {p}), \tag {6} +$$ + +where $\circ$ represents the element-wise product. We provide a detailed description of the load balancing loss in the supplementary material. + +In summary, when quantizing the multi-view language features, we optimize the discrete feature space $\mathcal{S}$ as well as the semantic indices maps $\mathcal{M}$ using the following loss: + +$$ +\mathcal {L} _ {q} = \lambda_ {c o s} \mathcal {L} _ {c o s} + \lambda_ {l b} \mathcal {L} _ {l b}. \tag {7} +$$ + +# 3.4. Language Embedded 3D Gaussians + +Utilizing the discrete feature space $S$ and index maps $\mathcal{M}$ in the previous section, we embed the compressed language features into 3D Gaussians for open-vocabulary scene understanding. Concretely, we expand the number of channels of each 3D Gaussian to include a compact feature vector representing the discrete index of language feature bases. To address semantic ambiguity arising from visual disparities across various viewpoints, we introduce a novel mechanism to reduce the spatial frequency of language embeddings through an adaptive learning loss based on the learned uncertainty values on each point. + +# 3.4.1 Compact Semantic Features on 3D Gaussians + +The inherent attributes of 3D Gaussians, such as color, are rendered onto the screen through the processes of rasterization and alpha blending (see Sec. 3.1). However, embedding discrete semantic indices $m$ onto 3D Gaussians would lead to erroneous results when optimizing through differential rendering, as these indices are not in a continuous space. Instead of directly embedding indices, we learn another continuous and compact semantic feature vectors, de + +noted as $\mathbf{s}_G\in \mathbb{R}^{d_s}$ , where $d_{s}$ is a hyper-parameter for controlling the storage capability for semantics. We then render these compact semantic feature vectors into a 2D feature map with rasterization and alpha blending, and decode the 2D feature map into the discrete semantic indices $m$ using a tiny MLP decoder: + +$$ +\hat {\mathcal {M}} = \operatorname {s o f t m a x} \left(D \left(R _ {s} \left(\mathcal {G}; p _ {\mathrm {c a m}}\right)\right)\right), \tag {8} +$$ + +where $R_{s}(\mathcal{G};p_{\mathrm{cam}}) \in \mathbb{R}^{H \times W \times d_{s}}$ represents the semantic feature rendering from a set of 3D Gaussians $\mathcal{G}$ viewed from camera pose $p_{\mathrm{cam}}$ . $D$ represents the tiny MLP decoder. During training process, a softmax operation is applied to the decoder's output, yielding the language feature index distribution $\hat{\mathcal{M}} \in \mathbb{R}^{H \times W \times N}$ , where $H$ and $W$ denote the height and width of the image, respectively. To optimize the semantic features of the 3D Gaussians and the MLP decoder, we apply the cross-entropy loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} = \operatorname {C E} (\tilde {\mathcal {M}}, \mathcal {M}), \tag {9} +$$ + +where $\mathcal{M} \in \mathbb{R}^{H \times W \times 1}$ denotes the discrete language indices map of the image extracted during the language feature quantization process (Sec. 3.3). + +# 3.4.2 Semantic Feature Smoothing + +Due to viewing angles, illuminations, existence of specular and semi-transparent materials, language features of the same spatial location in multi-view images may exhibit high variance, posing a challenge for precise semantic learning in 3D scenes. The high variance in language features at the same spatial location in multi-view images, due to differences in viewing angles, illumination, and the presence of specular and semi-transparent materials, presents challenges in accurately learning semantics in 3D scenes. Moreover, partially occluded objects may have their language features inaccurately extracted due to the lack of detection of their entirety. When these biased language features undergo quantization, the error may be amplified, resulting in a single 3D position being linked to several distinct language feature bases and, therefore, different discrete feature indices. Consequently, we introduce a smoothing strategy that limits the spatial frequency of semantic features on 3D Gaussians. This is achieved by incorporating an adaptive loss based on a learnable uncertainty value for each 3D Gaussian. + +Learning of uncertainty. To represent the variance associated with the semantic feature for each Gaussian, we record an estimizable semantic uncertainty on each point, denoted as $u \in [0,1]$ . A higher $u$ suggests that the semantic feature may exhibit instability and undergo frequent changes during the optimization process. The uncertainty values are jointly optimized when training the compact semantic features $\mathbf{s}_G$ : + +$$ +\mathcal {L} _ {\mathrm {C E}} = \frac {\sum \operatorname {C E} (\hat {\mathcal {M}} , \mathcal {M}) \circ (1 - R _ {u} (\mathcal {G} ; p _ {\mathrm {c a m}}))}{H \times W}, \tag {10} +$$ + +where $R_{u}(\mathcal{G};p_{\mathrm{cam}}) \in \mathbb{R}^{H \times W \times 1}$ is the rendered 2D uncertainty map from camera pose $p_{\mathrm{cam}}$ . At the same time, we regularize these uncertainty values to avoid converging to a trivial solution where all 3D Gaussians have the maximum uncertainties: + +$$ +\mathcal {L} _ {u} = \frac {\sum R _ {u} (\mathcal {G} ; p _ {\mathrm {c a m}})}{H \times W}. \tag {11} +$$ + +In summary, the total semantic loss for optimizing the compact semantic features of the 3D Gaussians and the MLP decoder is defined as: + +$$ +\mathcal {L} _ {s} = \lambda_ {\mathrm {C E}} \mathcal {L} _ {\mathrm {C E}} + \lambda_ {u} \mathcal {L} _ {u}, \tag {12} +$$ + +where $\lambda_{u}$ is a hyper-parameter for controlling the weight of regularization. We initialize the uncertainty values to zero at the start of training, and inconsistent semantic feature indices will lead to increased $u$ during optimization. + +Adaptive spatial smoothing loss. Building on the earlier observation that spatially adjacent positions typically display similar semantic features, we utilize this prior knowledge to deliberately decrease the spatial frequency of the embedded compact semantic features in 3D Gaussians, particularly for those with high uncertainty values. We harness the inductive biases of coordinate-based MLPs, known for learning low-frequency representations of target signals, to regularize the semantic features in 3D Gaussians. We calculate the smoothed semantic features $\mathbf{s}_{\mathrm{MLP}} \in \mathbb{R}^{d_s}$ by inputting the position of each 3D Gaussian into a small MLP: + +$$ +\mathrm {s} _ {\mathrm {M L P}} = \mathrm {M L P} (\mathrm {P E} (\mathbf {p})), \tag {13} +$$ + +where $\mathbf{p} \in \mathbb{R}^3$ represents the position of Gaussian, and PE represents the positional encoding used in NeRF [32]. We set a low frequency in the positional encoding to encourage spatial smoothness. We then apply the following loss for imposing the spatial smoothness regularization, where the degree of smoothness is adaptively controlled based on the learned uncertainty values: + +$$ +\mathcal {L} _ {\mathrm {s m o}} = \left\| \mathbf {s} _ {\mathrm {M L P}} - \mathbf {s} _ {G} ^ {*} \right\| _ {2} + \max \left(\mathbf {u} _ {G} ^ {*}, w _ {s}\right) \left\| \mathbf {s} _ {\mathrm {M L P}} ^ {*} - \mathbf {s} _ {G} \right\| _ {2}, \tag {14} +$$ + +where the $*$ denotes the stop gradient operator, and $w_{s}$ is the minimal weight for the semantic smoothing term. + +To summarize, the combined loss for optimization of semantic features, semantic uncertainties, and MLPs is given by: + +$$ +\mathcal {L} = \lambda_ {s} \mathcal {L} _ {s} + \lambda_ {\mathrm {s m o}} \mathcal {L} _ {\mathrm {s m o}}. \tag {15} +$$ + +# 4. Implementation Details + +We implement our method using PyTorch [34], and incorporate the CUDA kernel from 3D Gaussian Splitting [20] to speed up the rasterization rendering process. Our method + +
MethodPSNR↑SSIM↑LPIPS↓mPA↑mP↑mIoU↑mAP↑FPS↑Memory↓Storage↓Training Time↓
DFF [22]25.3780.7120.3120.8170.1240.0910.1990.20242GB+14GB41GB184min
LERF [21]25.7490.8110.3170.8900.4750.4030.6880.0425GB+6GB320MB54min
3DOVS [26]25.7820.7330.2950.9050.5290.4580.5500.1757GB+15GB205GB158min
Ours29.8260.9010.1120.9470.7530.5780.8158911GB+12GB15MB68min
+ +Table 1. Quantitative comparison of our method with DFF [22], LeRF [21], 3DOVS [26]. + +
mPA↑mP↑mIoU↑mAP↑
Quantizationw/o DINO0.9270.6040.4810.676
w/o Lb0.9390.6660.5440.738
Embeddingw/o uG0.9440.7170.5760.773
w/o MLP0.9450.7150.5800.774
w/o uG & MLP0.9440.7170.5800.774
Ours-0.9470.7530.5780.815
+ +Table 2. Quantitative results of ablation experiments. + +optimizes the scene's geometry and appearance with the same RGB loss following 3D Gaussian Splatting and enables adaptive density control of 3D Gaussians during the reconstruction process. We modify the CUDA kernel to enable the rendering of semantic features on the 3D Gaussians, and ensure that the optimizing of these semantic parameters does not affect original reconstruction quality. We set $d_{s} = 8$ and $w_{S} = 0.1$ in our modified 3D Gaussians, and set $\lambda_{D} = \lambda_{lb} = 0.5$ , and all other $\lambda$ values are set to 1. After the phase of extracting dense semantic features, which takes about 30 minutes, our model can be trained on one RTX3090 GPU for about 1 hour. The training involves 30,000 iterations, utilizing the Adam optimizer [29], with a learning rate set to 0.001 and betas equal to (0.9, 0.999). We leave the inference details of open-vocabulary language querying and semantic relevancy calculation in the supplementary material. + +# 5. Experiments + +# 5.1. Basic Setups + +Dataset. For a simultaneous evaluation of visual and semantic embedding quality, we select six scenes (excluding Stump) from the Mip-NeRF360 dataset [3] and manually annotate segmentation maps for each scene in the evaluation set. Each scene encompasses 180 to 320 images captured from various angles, and the evaluation set is chosen randomly with many novel-view images. Segmentation masks are annotated for primary objects in each scene. + +Baseline Methods and Metrics. We conduct a comparative evaluation of our method with DFF [22], LeRF [21], and 3DOVS [26], focusing on visual quality, language-embedded accuracy, rendering speed and model efficiency. To measure the visual quality in novel views, we report the PSNR, SSIM, and LPIPS [58] metrics. For the accuracy of language embedding, we measure the mean intersection + +over union (mIoU), mean pixel accuracy (mPA), mean precision (mP), and mean average precision (mAP) based on our annotations. The rendering speed (FPS) is measured by rendering the images with language features at a consistent resolution. Additionally, model efficiency is evaluated based on CPU and GPU memory usage during training, as well as data storage requirements and training duration. + +# 5.2. Comparisons + +We compare our method both qualitatively and quantitatively with DFF [22], LeRF [21], and 3DOVS [26] on our annotated Mip-NeRF360 dataset using a single RTX3090 GPU, following their default parameters but at the same resolution as our method. In addition, we compare our method with LeRF [21] quantitatively on the LeRF dataset, and the results are provided in the supplementary material. + +Qualitative Results. Fig. 9 displays a qualitative comparison of the novel view synthesis and semantic embedding results, demonstrating our method's efficacy in querying challenging objects in both indoor and outdoor scenes. Our approach notably delivers the highest visual rendering quality and query accuracy across all the tested scenes. Specifically, DFF [22] fails to identify "asphalt ground" in scene "bicycle" and "flower" in scene "garden". This may be caused by its use of LSeg [24], which is unstable to compute correct features in complex scenes. Moreover, due to predetermined query categories during training, 3DOVS [26] shows poor performance in scenes with complex objects. While LERF [21] can locate queried objects, its grid-based scene representation limits its ability to define clear boundaries. In contrast, our point-based approach supports high-frequency embedded semantic features, allowing for enhanced spatial semantic accuracy simply by incorporating more 3D Gaussians into scenes. This is facilitated by our quantization method, which substantially lowers memory costs. Additionally, our adaptive semantic smoothing technique effectively manages the high variance and ambiguity observed from different viewpoints, selectively applying spatial consistency as required. + +Quantitative Results. Tab. 1 presents a comparison across various metrics, including novel view synthesis quality, open-vocabulary query accuracy, and computational efficiency. We report both host memory and video memory usage, as well as the disk space used for storing the learned language features. Our approach outperforms others in ren + +![](images/d3beea16a26d33e09f99730e0d5b41a77ddcdf8c59197b36f7a25b82b35627c4.jpg) + +![](images/d748edfd3cd3f59bd045d1a77a650b929453b4dd21daeb0138bc6903c7a10443.jpg) + +![](images/5bab43e8ab8304a39eadbceb9e36e2d0438f6356c15570d260c6d0c6f3843305.jpg) + +![](images/11aaf59852aa68692c18124b2bd760406f1c6a65389dee0c956f5652f49b1f8c.jpg) + +![](images/b23130d91e34e39b7989f0534e9f4dab141d1c2b61757408b4206fecb83c5f96.jpg) + +![](images/e813163128efd3975fc6e18de4b4fa9147484b5c7945fea6ad290b440e9579f0.jpg) + +![](images/53c1b0fc2c2a55285987f5d62e7efc8f62af638d2a74e96bbf8442095e06f983.jpg) + +![](images/037de452f3342c5117320faaab01dedf77847c47c4fcd46d7ab24598fdc48efc.jpg) + +![](images/2bc9562fa882d2f53c8c20ae1020f90cc7c11f63c2c115eab513c2b22d5b0f60.jpg) + +![](images/e7e0669e02efc1d959a597141497c1d4a145c24f40cc97a389602f33b76ac837.jpg) + +![](images/28216add99a92599650e64576905b360ccac7ba2b6958deec5f18e304d1c9391.jpg) + +![](images/80544ede6198e2209383531f6712cabd3a62203c7a9f3956b61a52693aa49f48.jpg) + +![](images/13799b202273aa181ffc8f98293eba2efacdcf970402eca3433548ee0d449034.jpg) + +![](images/8fd951d288bc09b025787757ba4b4061f0e46bb00ed294028f79fe2bcfd85086.jpg) + +![](images/1876ff650db292cba9bb2f759b11be5b76cb8bd9b5f7d2d88ca7ca4c112b3bcb.jpg) + +![](images/9da1d5541d70b4d70ad675072c9a07942404d2e098992fe033053ac8846758cf.jpg) + +![](images/4841bbc29e9ff3ba924232f06a9ab5fc837236e245628066c9cbe88a80ebadcb.jpg) + +![](images/bb20061b2727c5f0aa8242bceda6dd1e45bc87e42558ce27c0a1024922ee5c52.jpg) + +![](images/5b0ef5f244801cfb7dd89e338437c30a41927bf122775b15a7037c7af5a35d4d.jpg) + +![](images/59c8227f226b756b184cbb03e2ad54bfb647cf13dbd352d5dde0064304245566.jpg) + +![](images/8f97e957663cc5d829c31840aff6bffcddce141b5507cac0b669fd100eda4273.jpg) + +![](images/ebaa8957bd0a82b73a62ce68a0b6012f87f84be6a67e7246b4c18c7cf73896c6.jpg) + +![](images/2a23a3c3991b88851ac495810f6497ee65d0b192c5d2d47349a1040851f23790.jpg) + +![](images/68a6719d21c2b94a70d85fced18fd402cf7947b6b67e416fbdf22e255a5415b7.jpg) + +![](images/d0b3f0c998e79d873897da18f4ffea36eb25b15762a8f63155a556533e265985.jpg) + +![](images/4511d76c1f9b16a9f55c8a464f6703eb7a03609287abe4d76eeb2bf556e92cd8.jpg) +Ground Truth + +![](images/7f59be792d90f005260b7eb64cbff61da4b4b427ab8be9832d3dd48675dc9901.jpg) +Ours + +![](images/289c48aab2956e1745fdba16c45d9fa770d570be5b88f17ddf8bfe0b6e2cf011.jpg) +DFF + +![](images/fbfeb8890e2cb36af4f85eae78013568252e365ee28282210493d34e567ca15c.jpg) +LERF +Figure 3. Comparison of novel view synthesis and query relevance visualization. Left to right: Ground truth novel view synthesis, novel view images with relevance visualization from our method, DFF [22], LeRF [21], and 3DOVS [26]. Top to bottom: Query words "asphalt ground", "bicycle", "jar of coconut oil", "flower", "LEGO Technic 856 Bulldozer", and "brown shoes". + +![](images/1697d935bb8456de43e540f7efab943c1f748a4b1bc64223b314a1acba3d0f90.jpg) +3DOVS + +dering quality and semantic query accuracy, while also offering lower computational demands and a significant speed increase, nearly 100 times faster in inference. It's noteworthy that methods like 3DOVS and DFF require extensive memory and storage due to their use of raw language features during training. In contrast, our use of the quantization scheme facilitates the incorporation of detailed semantics into complex 3D scenes with numerous 3D Gaussians, and concurrently achieves the most efficient storage utilization among the baseline methods. + +# 5.3. Open-vocabulary Query + +Fig. 5 illustrates the results of different open-vocabulary queries. We use a diverse range of vocabulary categories to identify objects in scenes, such as visual attribute terms + +like "green", and subjective adjectives like "cute". Furthermore, the figure shows our method's effectiveness in identifying both large-scale objects like "lego" and specific components of an object, such as a "engine". We show more examples in the supplementary material. + +# 5.4. Ablation Study + +We demonstrate the results of ablation studies in Tab. 2 and Fig. 6. The results show that embedding uncertainty without spatial smoothing of semantic features leads to suboptimal optimization. Conversely, using MLP solely for spatial smoothing, without accounting for semantic variance across multiple views, also impedes the optimization of precise language features. However, our full model, combining uncertainty with MLP smoothing in an adaptive man + +![](images/402888a12bc39c2a4ae6326453c04a48e7efa8264665a201f4a58a01ddd467e9.jpg) + +![](images/90e89a48b54af345a5bfc9ce1b4e5c49f3333baa21d7255c1279a88bd5374a45.jpg) + +![](images/35dc12049086a5880c4f1c8f32ca34fef3badc47fc049b37e6989b6f837637d6.jpg) + +![](images/ce8d02a2cfe64a9bd217155fce36b8485c63ea44eb522bc09dd5654a9b91f2cf.jpg) +Ground Truth + +![](images/38ddaef95e1b73f0ba4894d2323dc97295bb8fde2f5fe1fc87a75aee7e367ff1.jpg) +Rendering Results + +![](images/b951dfa67d089da120267d2dd2882babfb87ab8d9f23d541e8714eb1784367fc.jpg) + +![](images/4d263510b576a0f5c6dbcddbb31ded87118de7da935f15ac5cee470b0b1ad704.jpg) +Figure 4. Visual quality comparison of novel view synthesis results. Our method is able to recover more detailed geometry and appearance compared to other methods. + +![](images/f887ca98be12d205afd38462acb44ce1686a344cd172ba01725711e8223523f2.jpg) + +![](images/bcac5c00a47638894c19ce06d04ae398edb665cfe168558e62fff794d42753c2.jpg) +"green" + +![](images/18437561d7fd9116ae0df5d05f0285ba4104406e461bc54edb2c0ff9980ba66f.jpg) +"grass" + +![](images/09465d65994fa22688be2ca6fbc13c452c79f92ab53284605eecb622b7cb8f6e.jpg) +"lego" +"sheep toy" +Figure 5. Images of various open-vocabulary queries. + +![](images/b430bddc68b24a48179b9897face2ea8520730ab266e7b404f1ae61a24427a7f.jpg) +"engine" +"cute" + +ner effectively diminishes ambiguity and enhances the mean average precision (mAP) metric. Furthermore, integrating DINO features significantly improves the definition of object query boundaries. The load balancing loss, introduced during the quantization phase, results in a more utilized discrete feature space, facilitating the distinguish of objects with similar semantics, thereby boosting overall accuracy. + +![](images/8c9d528545a599a55a6dcd1bf72cc2ed755f75ce75479d2ef42208e04ad08642.jpg) +Figure 6. Comparison of ablation experiments. + +# 6. Conclusion + +We present Language Embedded 3D Gaussians, a novel scene representation designed for open-vocabulary query tasks. Our method successfully embeds quantized compact semantics features onto massive 3D Gaussians, while only maintaining minimal memory and storage requirements. To address semantic inconsistencies across different viewpoints, we propose a feature smoothing procedure that adaptively lowers the spatial frequency of embedded semantic features, guided by the uncertainty values learned on the 3D Gaussians. The result is a highly effective scene representation that not only enables high-quality novel view synthesis but also provides high accuracy in open-vocabulary querying, all achieved with modest computational resources. + +Limitations and Future works. Although DINO features improve object boundary detection, they fall short in pinpointing fine-grained object geometries at high resolutions when using CLIP-derived semantics. Furthermore, while vector quantization can efficiently compress scene-specific semantic feature space, it may compromise the finer-grained semantic information in the feature space due to its inherent clustering of semantic concepts. In addition, the detection of highly reflective or translucent objects, like televisions and mirrors, poses a challenge. These limitations might be overcome with more advanced visual-language models and native per-pixel semantic features. Nevertheless, our approach can be further adapted for broader open-vocabulary tasks, such as editing and generation of semantic-level scene objects. + +Acknowledgment. This work was supported by the National Natural Science Foundation of China (Project Number: 62372025 and 61932003) and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 3 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4 +[6] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3), 2013. 2 +[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2 +[8] Zhi Deng, Haoyao Xiao, Yining Lang, Hao Feng, and Juyong Zhang. Multi-scale hash encoding based neural geometry representation. Computational Visual Media, pages 1-18, 2024. 2 +[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[10] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 2 +[11] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Open-set 3d scene segmentation with rendered novel views. 2023. 2 +[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3 +[13] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim + +ple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 4, 1 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[15] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 2 +[16] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 2 +[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2 +[18] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2 +[19] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, Joshua B. Tenenbaum, Celso Miguel de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. Robotics: Science and Systems (RSS), 2023. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5 +[21] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 3, 6, 7, 4 +[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 1, 2, 3, 6, 7, 4 +[23] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2 +[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 6 +[25] Gengxin Liu, Oliver van Kaick, Hui Huang, and Ruizhen Hu. Active self-training for weakly supervised 3d scene seman + +tic segmentation. Computational Visual Media, pages 1-14, 2024. 2 +[26] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. 3d open-vocabulary segmentation with foundation models. arXiv preprint arXiv:2305.14093, 2023. 2, 3, 6, 7, 4 +[27] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. arXiv preprint arXiv:2305.14093, 2023. 2 +[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6 +[30] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. 2023. 2 +[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2 +[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5 +[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2 +[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5 +[35] Songyou Peng, Kyle Genova, Chiyu "Max" Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4 +[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 3 +[38] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2 + +[39] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043–9052, 2023. 2 +[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2 +[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[42] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 +[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2 +[44] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 2 +[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 4 +[46] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 2 +[47] Yuang Wang, Xingyi He, Sida Peng, Haotong Lin, Hujun Bao, and Xiaowei Zhou. Autorecon: Automated 3d object discovery and reconstruction. In CVPR, 2023. 3 +[48] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 3 +[49] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2 +[50] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171, 2022. 2 +[51] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding, 2023. 2 + +[52] Guo-Wei Yang, Zheng-Ning Liu, Dong-Yang Li, and Hao-Yang Peng. Jnerf: An efficient heterogeneous nerf model zoo based on jitter. Computational Visual Media, 9(2):401-404, 2023. 2 +[53] Jianing Yang, Xuweiyi Chen, Shengyi Qian, David Fouhey, and Joyce Chai. Chat with nerf: Grounding 3d objects in neural radiance field through dialog, 2023. 2 +[54] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F. Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent, 2023. 2 +[55] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3 +[56] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip, 2023. 2 +[57] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2 +[58] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 6 +[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 2 +[60] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8274-8284, 2023. 2 +[61] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 2 +[62] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 2 \ No newline at end of file diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/images.zip b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2287d8662067f26d307a62202fa0035fb8051ff4 --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4671899ba668932a5efff809fc415cf99044c22ad6633918a3bdc20f16eed0a0 +size 1005015 diff --git a/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/layout.json b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfb52c4e68eb73eb452645ea7ebab7dedf020c4c --- /dev/null +++ b/2024/Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding/layout.json @@ -0,0 +1,10634 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 103, + 541, + 121 + ], + "type": "text", + "content": "Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": "Jin-Chuan Shi" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": " Miao Wang" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": " Hao-Bin Duan" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": " Shao-Hua Guan" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, SCSE, Beihang University \n" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 142, + 523, + 186 + ], + "type": "text", + "content": "Zhongguancun Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 213, + 286, + 323 + ], + "blocks": [ + { + "bbox": [ + 48, + 213, + 286, + 323 + ], + "lines": [ + { + "bbox": [ + 48, + 213, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 213, + 286, + 323 + ], + "type": "image", + "image_path": "c0570b97684c89522e054a3189fc5c2e4fad8082300910a0b8b2b23ea1c5157e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 323, + 168, + 403 + ], + "blocks": [ + { + "bbox": [ + 48, + 323, + 168, + 403 + ], + "lines": [ + { + "bbox": [ + 48, + 323, + 168, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 323, + 168, + 403 + ], + "type": "image", + "image_path": "e907397f291ccc9103f688d35ae1a586d808fb4a7e22b05f24db24727adc9887.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 411, + 546, + 456 + ], + "lines": [ + { + "bbox": [ + 45, + 411, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 45, + 411, + 546, + 456 + ], + "type": "text", + "content": "Figure 1. We present Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary querying. The top row visualizes the original image, novel view synthesis result with query relevancy and PCA of learned semantic features. The bottom row compares our method with other language-embedded representations. The right-side bar maps relevancy values to heatmap colors. Our method achieves better fidelity and query accuracy while rendering at higher frame rates." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 323, + 286, + 403 + ], + "blocks": [ + { + "bbox": [ + 168, + 323, + 286, + 403 + ], + "lines": [ + { + "bbox": [ + 168, + 323, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 168, + 323, + 286, + 403 + ], + "type": "image", + "image_path": "2632718df8f54bbf20e2cdb2f02d088daa68ab171c1b2f131a4b0e9577e411c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 286, + 213, + 525, + 323 + ], + "blocks": [ + { + "bbox": [ + 286, + 213, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 286, + 213, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 286, + 213, + 525, + 323 + ], + "type": "image", + "image_path": "5081425f055e55df3075f169bfd2ffb2897a1cc5cf1ebe8adc0744d44c27f6d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 286, + 323, + 406, + 403 + ], + "blocks": [ + { + "bbox": [ + 286, + 323, + 406, + 403 + ], + "lines": [ + { + "bbox": [ + 286, + 323, + 406, + 403 + ], + "spans": [ + { + "bbox": [ + 286, + 323, + 406, + 403 + ], + "type": "image", + "image_path": "5e22724e9f82363b7f2140980d0676fe110fc06e28ed4eec560a448044521348.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 406, + 323, + 525, + 403 + ], + "blocks": [ + { + "bbox": [ + 406, + 323, + 525, + 403 + ], + "lines": [ + { + "bbox": [ + 406, + 323, + 525, + 403 + ], + "spans": [ + { + "bbox": [ + 406, + 323, + 525, + 403 + ], + "type": "image", + "image_path": "ed18a6ac3dc0f2229462af24a1ba84253a61cad9da38099b1b098eb8f27be5fd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 525, + 215, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 525, + 215, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 525, + 215, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 525, + 215, + 545, + 403 + ], + "type": "image", + "image_path": "35382f7f847764445a1edb7da8935a9346fa836ebbc31143619a7b7ee874eaa0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 503, + 289, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 289, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 289, + 695 + ], + "type": "text", + "content": "Open-vocabulary querying in 3D space is challenging but essential for scene understanding tasks such as object localization and segmentation. Language-embedded scene representations have made progress by incorporating language features into 3D spaces. However, their efficacy heavily depends on neural networks that are resource-intensive in training and rendering. Although recent 3D Gaussians offer efficient and high-quality novel view synthesis, directly embedding language features in them leads to prohibitive memory usage and decreased performance. In this work, we introduce Language Embedded 3D Gaussians, a novel scene representation for open-vocabulary query tasks. Instead of embedding high-dimensional raw semantic features on 3D Gaussians, we propose a dedicated quantization scheme that drastically alleviates the memory requirement, and a novel embedding procedure that" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 467, + 546, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 467, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 467, + 546, + 574 + ], + "type": "text", + "content": "achieves smoother yet high accuracy query, countering the multi-view feature inconsistencies and the high-frequency inductive bias in point-based representations. Our comprehensive experiments show that our representation achieves the best visual quality and language querying accuracy across current language-embedded representations, while maintaining real-time rendering frame rates on a single desktop GPU. Project page: https://buaavrcg.github.io/LEGaussians/." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 387, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Neural Radiance Field (NeRFs) [2, 3, 32, 57] and 3D Gaussian Splatting [20] has advanced the development of efficient and high-quality 3D scene novel view synthesis from multi-view images. Nevertheless, these representations solely encapsulate geometric and appearance details without any semantic information. To bridge this gap, language-embedded neural representations [21, 22] try to integrate semantic information from multi-view im" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 142, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 142, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 142, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5333" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "ages into 3D scenes for open-vocabulary querying tasks, which allows intuitive interaction with large language models (LLMs) [53, 54] and human users, and powers broad applications including scene editing, VR/AR, autonomous driving, and robotic navigation. Compared to traditional semantic labeling methods, language features from visual-language models like CLIP [36] offer more comprehensive semantic understanding capability, as they encompass objects with long-tail distribution, enhancing their suitability for real-world applications. However, accurately incorporating language embedding into current 3D scene representations, while maintaining their efficiency and visual quality, presents a significant challenge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 373 + ], + "type": "text", + "content": "Recent techniques [21, 22, 27] extract dense language features from multi-view 2D images and incorporate additional output branches in scene representation to predict semantic features. However, the quality of semantic features heavily relies on scene representation, and trivially expanding the output channels poses significant challenges in recovering high-precision and robust semantics of the scenes. Furthermore, while 3D Gaussian [20] are efficient and fast for high-quality 3D scene representation, embedding raw semantic information into a massive number of points can cause prohibitive memory requirements and significantly lower the efficiency of both optimization and rendering." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 577 + ], + "type": "text", + "content": "In this paper, we introduce Language Embedded 3D Gaussians, a semantic scene representation framework that provides both high precision and efficiency for open-vocabulary query tasks. To avoid the substantial memory requirement that would result from directly increasing the feature channels for raw semantic features in the 3D Gaussians, we employ a novel feature quantization approach that significantly decreases the computational and memory cost. By leveraging the redundancy nature of local semantic features, we construct more streamlined language features that are compactly stored on the 3D Gaussians. Additionally, we address the issue of semantic ambiguity caused by visual inconsistency from multi-view images by implementing a mechanism that lowers the spatial frequency of semantic features, guided by learned uncertainty values. This technique enables language features in the 3D Gaussians to be much smoother yet still precise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 578, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 625 + ], + "type": "text", + "content": "Our extensive experiments demonstrate that our method achieves state-of-the-art quality in both novel view synthesis and open-vocabulary querying tasks, while allowing real-time rendering on consumer-level devices." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 627, + 218, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 627, + 218, + 639 + ], + "spans": [ + { + "bbox": [ + 58, + 627, + 218, + 639 + ], + "type": "text", + "content": "In summary, our contributions include:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 642, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 642, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 701 + ], + "type": "text", + "content": "- We introduce a novel quantization scheme that efficiently compresses and integrates semantic features into dense 3D Gaussians, ensuring efficient optimization and rendering on consumer devices while maintaining accurate semantic embedding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "- We propose a mechanism that leverages spatial position" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 107 + ], + "type": "text", + "content": "and semantic uncertainty of 3D Gaussians to address the semantic ambiguity arising from visual inconsistency across views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 547, + 156 + ], + "type": "text", + "content": "- Our method outperforms other language-embedded 3D representations, delivering state-of-the-art results in visual quality, language query precision, and rendering speed at the same time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 167, + 392, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 392, + 180 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 392, + 180 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 188, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 188, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 188, + 547, + 392 + ], + "type": "text", + "content": "Neural Rendering. NeRF [32] has demonstrated superior novel view synthesis quality over traditional methods [6, 10, 16, 17, 40], and various efforts have been made to enhance its performance. Although many methods focus on improving NeRF's rendering quality [2, 3, 31, 38, 43, 52, 57], they still suffer from slow training and rendering speed. On the other hand, explicit and hybrid scene representations [4, 7, 8, 14, 28, 33, 41, 49] typically utilize hash grids [33] and point clouds [49] to reduce computational cost of large neural networks. Recent 3D Gaussian [20] sets a new standard on both rendering quality and training speed, by using fast rasterization of 3D Gaussians to replace differential volume rendering. Nevertheless, directly embedding language embeddings on the dense 3D Gaussians is non-trivial, as the massive amount of 3D points requires prohibitive memory usage and causes a drastic performance drop in both training and rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 397, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 547, + 613 + ], + "type": "text", + "content": "Language Embedded Scene Representation. Incorporating specific semantics into NeRF's implicit MLP-based representation is difficult due to challenges in accurately identifying 3D regions. Various approaches [11, 15, 21, 22, 39, 44, 46, 61] have tried to integrate semantic data into NeRF, often by adding new network branches. For instance, DFF [22] embeds a branch for language predictions and uses a pre-trained encoder for supervision, while LERF [21] assimilates CLIP [36] features from multi-scale image crops into hash grid-represented 3D scenes. 3D-OVS [26] takes a different route, creating dual decomposed 3d tensors for geometry and semantics, but it requires predefined segmented classes and doesn't support arbitrary queries. Other methods [23, 60] decompose scenes into smaller MLPs representing local semantics. However, the quality of semantic embeddings and the efficiency of training and rendering of these methods, are all constrained by the limitations of their scene representation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Open Vocabulary 3D Scene Understanding. Recent progress in open vocabulary scene understanding has been marked by the integration of 2D Vision-Language Models with 3D point cloud processing [18, 50, 51, 59, 62]. These approaches focus on aligning features and projecting 3D data into 2D to enhance zero-shot learning capabilities. Additionally, advancements in 3D object detection and segmentation [9, 19, 25, 30, 35, 42, 56] have shown the effec" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5334" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 144 + ], + "type": "text", + "content": "tiveness of merging point cloud data with visual features extracted from image for scene analysis [48]. However, these methods mainly address the comprehension and analysis of existing scene representations, like point clouds, rather than optimizing scene representation from multi-view images, which is the focus of our work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 103, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 103, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 103, + 167 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 289, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 295 + ], + "type": "text", + "content": "In this section, we introduce our training process of Language Embedded 3D Gaussians, including (1) a recap of 3D Gaussian Splatting [20] (Sec. 3.1), (2) extracting dense language features from multi-view images (Sec. 3.2), (3) a quantization scheme for high-dimensional language features that creates a compact feature space for the embedding process (Sec. 3.3), (4) an embedding mechanism that alleviates semantic ambiguity caused by visual difference across views, by lowering the spatial frequency of semantic features with the learned uncertainty values (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 212, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 212, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 212, + 316 + ], + "type": "text", + "content": "3.1. Recap: 3D Gaussian Splitting" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": "3D Gaussian Splatting [20] renders complex scenes by merging a multitude of colored 3D Gaussians, which are subsequently projected onto camera views through a rasterization process. By employing differentiable rendering and gradient descent, the attributes of these 3D Gaussians, including position " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", color " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", and opacity " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", are optimized to represent the 3D scene based on a collection of input images. The result image " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": " is rendered from a specific camera pose " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", using the differentiable rasterization " + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 321, + 287, + 441 + ], + "type": "text", + "content": ", represented as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 448, + 287, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 448, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 118, + 448, + 287, + 461 + ], + "type": "interline_equation", + "content": "I = R (p, \\Sigma , c, \\alpha ; p _ {\\mathrm {c a m}}). \\tag {1}", + "image_path": "835f0a0ed95236055148d8f87c698dbaef39db78bccce888ff222bd3548263ed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 468, + 239, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 239, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 239, + 480 + ], + "type": "text", + "content": "3.2. Dense Language Feature Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 629 + ], + "type": "text", + "content": "We first extract pixel-level dense language features from visual-language models. While CLIP [36] encodes images into global language features, its direct application is not feasible for our purposes as we require pixel-level targets to learn 3D scene representations from multi-view images. Prior studies [21, 26] overcome this limitation by computing multi-scale dense CLIP features for each image. To obtain dense language embeddings from multi-view images, we employ a slightly different hierarchical random cropping technique to extract CLIP features, similar to 3DOVS [26]. We aggregate features from all layers and normalize them to produce the final language embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Nevertheless, features extracted from CLIP only provide a rough boundary of different semantic regions, resulting in ambiguities and inaccuracies in the language embedding of 3D scenes. Conversely, DINO [5] exhibits autonomous object decomposition without requiring labeled data [1], as several studies [21, 22, 26, 47] have shown that DINO effectively enhances language feature grouping without reliance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "on labels or prior knowledge. Therefore, we also extract DINO features as complementary to enhance the details of extracted language features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": "We then concatenate the dense CLIP and DINO features extracted from multi-view images to form hybrid language feature maps. We denote the features at the position " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " on the CLIP and DINO feature maps of image " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{I,x,y}^{\\mathrm{CLIP}}\\in \\mathbb{R}^{d_{\\mathrm{CLIP}}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{I,x,y}^{\\mathrm{DINO}}\\in \\mathbb{R}^{d_{\\mathrm{DINO}}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{DINO}}" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " represent the dimension of each feature. The hybrid language feature " + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{I,x,y}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 109, + 547, + 196 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 201, + 545, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 201, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 372, + 201, + 545, + 217 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {I, x, y} = \\mathbf {F} _ {I, x, y} ^ {\\mathrm {C L I P}} \\oplus \\mathbf {F} _ {I, x, y} ^ {\\mathrm {D I N O}}, \\tag {2}", + "image_path": "f5d02f2608699ae7273e0319497e45a29413f925ee804f56d8470c62a13a76e7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 224, + 512, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 512, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 512, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 224, + 512, + 237 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 305, + 224, + 512, + 237 + ], + "type": "text", + "content": " is the total channels in the hybrid features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 247, + 494, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 494, + 260 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 494, + 260 + ], + "type": "text", + "content": "3.3. Quantization of Language Features" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 266, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 518 + ], + "type": "text", + "content": "Direct integration of our extracted language features as in prior studies [21, 22, 26] would be infeasible for 3D Gaussians, due to the substantial memory demands of recording high-dimensional language embeddings on them, which significantly hinders rendering efficiency and limits the maximum number of 3D Gaussians that can be trained at the same time. To alleviate such storage and computation costs of raw language embeddings, our insight is to leverage the inherent redundancy in language features, as those within a single object share very similar semantics. Additionally, the semantics of a single scene merely cover only a small fraction of the original CLIP feature space, leading to unnecessary data stored in the 3D representation for all the queries we would need. Previous studies [12, 37, 45, 55] have demonstrated vector quantization can effectively compress the feature space while preserving the semantics contained within the feature space accurately and comprehensively. We propose a dedicated quantization scheme to effectively compress the language features extracted from multiple viewpoints, resulting in a more efficient and compact representation of scene-aware language features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": "Specifically, our goal is to transform the hybrid language features " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " into a quantized version " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": ", which approximates " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": ". To achieve this, we develop a discrete language feature space " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "S = \\{\\mathbf{f}_i \\in \\mathbb{R}^d | i = 1,2,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": ", and use an integer index " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "m \\in \\{1,2,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " for retrieving the nearest language feature in " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": ". The quantized language feature is then given by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 605, + 545, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 605, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 373, + 605, + 545, + 637 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {F}} = \\sum_ {i} ^ {N} \\mathbf {f} _ {i} \\cdot \\operatorname {o n e h o t} (m) _ {i}. \\tag {3}", + "image_path": "46e428b017f0956f14065ec476a5b476062914cb5744f20eba200fa386db24b8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "content": "-th basis of the language features, and onehot " + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "inline_equation", + "content": "(m) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "content": " is the one-hot vector for index " + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "content": ". This quantization scheme effectively removes excessive information in the original language embeddings by compressing the original continuous language feature space into discrete bases, with the compression rate adjustable through the size of space " + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 639, + 547, + 723 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5335" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 526, + 277 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 526, + 277 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 526, + 277 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 526, + 277 + ], + "type": "image", + "image_path": "72e709372559d5db05a80fd61389fcddd322026ad12added03c5c553b953952e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 331 + ], + "type": "text", + "content": "Figure 2. The training process for Language-embedded 3D Gaussians starts with initializing scenes following 3D Gaussian Splatting [20] and randomly initializing semantic features and setting uncertainty to zero. Dense language features from multi-view CLIP [36] and DINO [5] are quantized to create a discrete feature space and semantic indices. These attributes of the 3D Gaussians are then rendered into 2D maps using a differentiable rasterizer. The optimization is achieved through semantic and adaptive spatial smoothing loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "text", + "content": "Quantization. To quantitize high-dimensional language feature " + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "text", + "content": ", we execute a max-similarity search within space " + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "text", + "content": ", utilizing cosine similarity " + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\cos \\langle \\cdot \\rangle" + }, + { + "bbox": [ + 46, + 341, + 288, + 376 + ], + "type": "text", + "content": " as the distance metric:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 380, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\mathcal {D} (\\mathbf {F}, \\mathbf {f} _ {i}) = \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {C L I P}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle + \\lambda_ {\\mathrm {D I N O}} \\cos \\left\\langle \\mathbf {F} ^ {\\mathrm {D I N O}} \\cdot \\mathbf {f} _ {i} ^ {\\mathrm {D I N O}} \\right\\rangle , \\tag {4}", + "image_path": "b0db789c68b157de19f88b50f5a5fcf3c383dc85d91e144025fb4413f8fdb4d7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{DINO}}" + }, + { + "bbox": [ + 46, + 407, + 287, + 431 + ], + "type": "text", + "content": " serves as a hyper-parameter that modulates the significance of DINO within the language feature." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "content": "Following VQ-VAE [45], the selected language feature index from the set " + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "content": " is determined as " + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "inline_equation", + "content": "m = \\operatorname{argmax}_i(\\mathcal{D}(\\mathbf{F},\\mathbf{f}_i))" + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "content": ", and the quantization of " + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "content": " is computed using Eq. (3). The result for each image after this quantization procedure is a semantic indices map, denoted as " + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 46, + 431, + 287, + 502 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": "Optimization. During the quantization of all language features extracted from multi-view images, the optimization of the discrete feature space " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " is simultaneously accomplished by minimizing the cosine similarity loss between the language features " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": " and the quantization " + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathrm{i}}" + }, + { + "bbox": [ + 46, + 506, + 287, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 567, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 567, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 64, + 567, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\cos} \\left(\\mathbf {F} _ {i}\\right) = \\left(1 - \\cos \\left\\langle \\mathbf {F} _ {i} ^ {\\mathrm {C L I P}} \\cdot \\hat {\\mathbf {F}} _ {i} ^ {\\mathrm {C L I P}} \\right\\rangle\\right) \\tag {5} \\\\ + \\lambda_ {\\mathrm {D I N O}} (1 - \\cos \\langle \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\cdot \\mathbf {F} _ {i} ^ {\\mathrm {D D N O}} \\rangle). \\\\ \\end{array}", + "image_path": "6cc899f3b2f505928a384d330ca5125adef42aa7bbb39db4dd1d31bd9d1c5929.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": "Furthermore, to prevent quantization collapse and ensure maximal utilization of each feature in the feature space, we have devised a load balancing loss inspired by the design of the Switch Transformer [13]. The load balancing loss is calculated by performing an element-wise multiplication of the utilization ratio " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": " and the mean selection probability " + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 604, + 288, + 688 + ], + "type": "text", + "content": " of each feature, followed by their summation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 691, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 691, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 129, + 691, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l b} = \\sum^ {N} (\\mathbf {r} \\circ \\mathbf {p}), \\tag {6}", + "image_path": "cf56c09e9bc24f16d26a89398d53fe13be06eb94c708b60d22ab5c564e51a2b7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 341, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 341, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 341, + 545, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 341, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 305, + 341, + 545, + 376 + ], + "type": "text", + "content": " represents the element-wise product. We provide a detailed description of the load balancing loss in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "content": "In summary, when quantizing the multi-view language features, we optimize the discrete feature space " + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "content": " as well as the semantic indices maps " + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 305, + 376, + 545, + 413 + ], + "type": "text", + "content": " using the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 422, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 422, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 373, + 422, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {q} = \\lambda_ {c o s} \\mathcal {L} _ {c o s} + \\lambda_ {l b} \\mathcal {L} _ {l b}. \\tag {7}", + "image_path": "3c4cdf34a82923a54cfe26d70dcda108a584af406310258f03dbd1b33fe952ff.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 443, + 494, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 494, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 494, + 456 + ], + "type": "text", + "content": "3.4. Language Embedded 3D Gaussians" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "text", + "content": "Utilizing the discrete feature space " + }, + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "text", + "content": " and index maps " + }, + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 461, + 545, + 594 + ], + "type": "text", + "content": " in the previous section, we embed the compressed language features into 3D Gaussians for open-vocabulary scene understanding. Concretely, we expand the number of channels of each 3D Gaussian to include a compact feature vector representing the discrete index of language feature bases. To address semantic ambiguity arising from visual disparities across various viewpoints, we introduce a novel mechanism to reduce the spatial frequency of language embeddings through an adaptive learning loss based on the learned uncertainty values on each point." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 599, + 533, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 533, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 533, + 611 + ], + "type": "text", + "content": "3.4.1 Compact Semantic Features on 3D Gaussians" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "The inherent attributes of 3D Gaussians, such as color, are rendered onto the screen through the processes of rasterization and alpha blending (see Sec. 3.1). However, embedding discrete semantic indices " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " onto 3D Gaussians would lead to erroneous results when optimizing through differential rendering, as these indices are not in a continuous space. Instead of directly embedding indices, we learn another continuous and compact semantic feature vectors, de" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5336" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "noted as " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_G\\in \\mathbb{R}^{d_s}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "d_{s}" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " is a hyper-parameter for controlling the storage capability for semantics. We then render these compact semantic feature vectors into a 2D feature map with rasterization and alpha blending, and decode the 2D feature map into the discrete semantic indices " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " using a tiny MLP decoder:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 151, + 287, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 151, + 287, + 166 + ], + "spans": [ + { + "bbox": [ + 99, + 151, + 287, + 166 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {M}} = \\operatorname {s o f t m a x} \\left(D \\left(R _ {s} \\left(\\mathcal {G}; p _ {\\mathrm {c a m}}\\right)\\right)\\right), \\tag {8}", + "image_path": "6d749ec6e29c4c2c7c9aa3996369637c1cdb8f5676edc34d0b3c65ac3dfdf29a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "R_{s}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times d_{s}}" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": " represents the semantic feature rendering from a set of 3D Gaussians " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": " viewed from camera pose " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{cam}}" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": " represents the tiny MLP decoder. During training process, a softmax operation is applied to the decoder's output, yielding the language feature index distribution " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{M}} \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 171, + 289, + 281 + ], + "type": "text", + "content": " denote the height and width of the image, respectively. To optimize the semantic features of the 3D Gaussians and the MLP decoder, we apply the cross-entropy loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 287, + 287, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 287, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 124, + 287, + 287, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = \\operatorname {C E} (\\tilde {\\mathcal {M}}, \\mathcal {M}), \\tag {9}", + "image_path": "4e0bda5169bbc6ff889a4abc4e709bb9650eb7150e4a072dca7d614cf3db64b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 307, + 287, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 307, + 287, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 46, + 307, + 287, + 344 + ], + "type": "text", + "content": " denotes the discrete language indices map of the image extracted during the language feature quantization process (Sec. 3.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 353, + 204, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 204, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 204, + 365 + ], + "type": "text", + "content": "3.4.2 Semantic Feature Smoothing" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 373, + 289, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 289, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 289, + 612 + ], + "type": "text", + "content": "Due to viewing angles, illuminations, existence of specular and semi-transparent materials, language features of the same spatial location in multi-view images may exhibit high variance, posing a challenge for precise semantic learning in 3D scenes. The high variance in language features at the same spatial location in multi-view images, due to differences in viewing angles, illumination, and the presence of specular and semi-transparent materials, presents challenges in accurately learning semantics in 3D scenes. Moreover, partially occluded objects may have their language features inaccurately extracted due to the lack of detection of their entirety. When these biased language features undergo quantization, the error may be amplified, resulting in a single 3D position being linked to several distinct language feature bases and, therefore, different discrete feature indices. Consequently, we introduce a smoothing strategy that limits the spatial frequency of semantic features on 3D Gaussians. This is achieved by incorporating an adaptive loss based on a learnable uncertainty value for each 3D Gaussian." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "Learning of uncertainty. To represent the variance associated with the semantic feature for each Gaussian, we record an estimizable semantic uncertainty on each point, denoted as " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "u \\in [0,1]" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": ". A higher " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": " suggests that the semantic feature may exhibit instability and undergo frequent changes during the optimization process. The uncertainty values are jointly optimized when training the compact semantic features " + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_G" + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 71, + 546, + 97 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 71, + 546, + 97 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 546, + 97 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = \\frac {\\sum \\operatorname {C E} (\\hat {\\mathcal {M}} , \\mathcal {M}) \\circ (1 - R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}}))}{H \\times W}, \\tag {10}", + "image_path": "ac1cbe4cb6d3483085406eed6dca2dfbc6fe4192334fa78afa35d9d21b8f3922.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "inline_equation", + "content": "R_{u}(\\mathcal{G};p_{\\mathrm{cam}}) \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "text", + "content": " is the rendered 2D uncertainty map from camera pose " + }, + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{cam}}" + }, + { + "bbox": [ + 304, + 102, + 547, + 161 + ], + "type": "text", + "content": ". At the same time, we regularize these uncertainty values to avoid converging to a trivial solution where all 3D Gaussians have the maximum uncertainties:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 166, + 546, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 166, + 546, + 191 + ], + "spans": [ + { + "bbox": [ + 378, + 166, + 546, + 191 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {u} = \\frac {\\sum R _ {u} (\\mathcal {G} ; p _ {\\mathrm {c a m}})}{H \\times W}. \\tag {11}", + "image_path": "bd8d77af73ac15b11a0480a40d27e06cf29cb01374141cd4257d055a595e96c9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 196, + 547, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 231 + ], + "type": "text", + "content": "In summary, the total semantic loss for optimizing the compact semantic features of the 3D Gaussians and the MLP decoder is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 242, + 546, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 242, + 546, + 255 + ], + "spans": [ + { + "bbox": [ + 377, + 242, + 546, + 255 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\lambda_ {\\mathrm {C E}} \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {u} \\mathcal {L} _ {u}, \\tag {12}", + "image_path": "f4ae3c80fa6d7aee588b948a1df3ee45794e35b48226e1d8df201fe232d41257.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\lambda_{u}" + }, + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "text", + "content": " is a hyper-parameter for controlling the weight of regularization. We initialize the uncertainty values to zero at the start of training, and inconsistent semantic feature indices will lead to increased " + }, + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 263, + 546, + 312 + ], + "type": "text", + "content": " during optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 315, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 447 + ], + "type": "text", + "content": "Adaptive spatial smoothing loss. Building on the earlier observation that spatially adjacent positions typically display similar semantic features, we utilize this prior knowledge to deliberately decrease the spatial frequency of the embedded compact semantic features in 3D Gaussians, particularly for those with high uncertainty values. We harness the inductive biases of coordinate-based MLPs, known for learning low-frequency representations of target signals, to regularize the semantic features in 3D Gaussians. We calculate the smoothed semantic features " + }, + { + "bbox": [ + 304, + 315, + 547, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{MLP}} \\in \\mathbb{R}^{d_s}" + }, + { + "bbox": [ + 304, + 315, + 547, + 447 + ], + "type": "text", + "content": " by inputting the position of each 3D Gaussian into a small MLP:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 455, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 455, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 378, + 455, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\mathrm {s} _ {\\mathrm {M L P}} = \\mathrm {M L P} (\\mathrm {P E} (\\mathbf {p})), \\tag {13}", + "image_path": "5a9fa7d1a156a567c22c08b5dc835bf197955ee3a0c3a8fa8df9809d536b5fb8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 470, + 547, + 555 + ], + "type": "text", + "content": " represents the position of Gaussian, and PE represents the positional encoding used in NeRF [32]. We set a low frequency in the positional encoding to encourage spatial smoothness. We then apply the following loss for imposing the spatial smoothness regularization, where the degree of smoothness is adaptively controlled based on the learned uncertainty values:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 561, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 561, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 310, + 561, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s m o}} = \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} - \\mathbf {s} _ {G} ^ {*} \\right\\| _ {2} + \\max \\left(\\mathbf {u} _ {G} ^ {*}, w _ {s}\\right) \\left\\| \\mathbf {s} _ {\\mathrm {M L P}} ^ {*} - \\mathbf {s} _ {G} \\right\\| _ {2}, \\tag {14}", + "image_path": "c481ee205ec37a760118a6598d619f44282c449f5098976827efd94b3693819e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "content": " denotes the stop gradient operator, and " + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "inline_equation", + "content": "w_{s}" + }, + { + "bbox": [ + 304, + 580, + 545, + 604 + ], + "type": "text", + "content": " is the minimal weight for the semantic smoothing term." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 639 + ], + "type": "text", + "content": "To summarize, the combined loss for optimization of semantic features, semantic uncertainties, and MLPs is given by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 377, + 638, + 545, + 650 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 638, + 545, + 650 + ], + "spans": [ + { + "bbox": [ + 377, + 638, + 545, + 650 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {s} \\mathcal {L} _ {s} + \\lambda_ {\\mathrm {s m o}} \\mathcal {L} _ {\\mathrm {s m o}}. \\tag {15}", + "image_path": "31fdf7c278f696cc7375f38cb21f6d9c77a9f7753428fe128f7fc7505d057258.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 657, + 441, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 441, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 441, + 671 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "We implement our method using PyTorch [34], and incorporate the CUDA kernel from 3D Gaussian Splitting [20] to speed up the rasterization rendering process. Our method" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5337" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 545, + 133 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 133 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 133 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓mPA↑mP↑mIoU↑mAP↑FPS↑Memory↓Storage↓Training Time↓
DFF [22]25.3780.7120.3120.8170.1240.0910.1990.20242GB+14GB41GB184min
LERF [21]25.7490.8110.3170.8900.4750.4030.6880.0425GB+6GB320MB54min
3DOVS [26]25.7820.7330.2950.9050.5290.4580.5500.1757GB+15GB205GB158min
Ours29.8260.9010.1120.9470.7530.5780.8158911GB+12GB15MB68min
", + "image_path": "2bd340c2adfd4dff9c49af92b75078b21dc9546fd01792cdce1330d26fdabe0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 160, + 285, + 246 + ], + "blocks": [ + { + "bbox": [ + 133, + 141, + 457, + 153 + ], + "lines": [ + { + "bbox": [ + 133, + 141, + 457, + 153 + ], + "spans": [ + { + "bbox": [ + 133, + 141, + 457, + 153 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of our method with DFF [22], LeRF [21], 3DOVS [26]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 160, + 285, + 246 + ], + "lines": [ + { + "bbox": [ + 50, + 160, + 285, + 246 + ], + "spans": [ + { + "bbox": [ + 50, + 160, + 285, + 246 + ], + "type": "table", + "html": "
mPA↑mP↑mIoU↑mAP↑
Quantizationw/o DINO0.9270.6040.4810.676
w/o Lb0.9390.6660.5440.738
Embeddingw/o uG0.9440.7170.5760.773
w/o MLP0.9450.7150.5800.774
w/o uG & MLP0.9440.7170.5800.774
Ours-0.9470.7530.5780.815
", + "image_path": "8332156d34483b35f46174b10d45586a24d14747e0cb26e82856265243b5ef57.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 255, + 263, + 266 + ], + "lines": [ + { + "bbox": [ + 70, + 255, + 263, + 266 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 263, + 266 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation experiments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "content": "optimizes the scene's geometry and appearance with the same RGB loss following 3D Gaussian Splatting and enables adaptive density control of 3D Gaussians during the reconstruction process. We modify the CUDA kernel to enable the rendering of semantic features on the 3D Gaussians, and ensure that the optimizing of these semantic parameters does not affect original reconstruction quality. We set " + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "inline_equation", + "content": "d_{s} = 8" + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "inline_equation", + "content": "w_{S} = 0.1" + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "content": " in our modified 3D Gaussians, and set " + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\lambda_{D} = \\lambda_{lb} = 0.5" + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "content": ", and all other " + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 276, + 287, + 479 + ], + "type": "text", + "content": " values are set to 1. After the phase of extracting dense semantic features, which takes about 30 minutes, our model can be trained on one RTX3090 GPU for about 1 hour. The training involves 30,000 iterations, utilizing the Adam optimizer [29], with a learning rate set to 0.001 and betas equal to (0.9, 0.999). We leave the inference details of open-vocabulary language querying and semantic relevancy calculation in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 128, + 504 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 128, + 523 + ], + "type": "text", + "content": "5.1. Basic Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 529, + 286, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 286, + 624 + ], + "type": "text", + "content": "Dataset. For a simultaneous evaluation of visual and semantic embedding quality, we select six scenes (excluding Stump) from the Mip-NeRF360 dataset [3] and manually annotate segmentation maps for each scene in the evaluation set. Each scene encompasses 180 to 320 images captured from various angles, and the evaluation set is chosen randomly with many novel-view images. Segmentation masks are annotated for primary objects in each scene." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Baseline Methods and Metrics. We conduct a comparative evaluation of our method with DFF [22], LeRF [21], and 3DOVS [26], focusing on visual quality, language-embedded accuracy, rendering speed and model efficiency. To measure the visual quality in novel views, we report the PSNR, SSIM, and LPIPS [58] metrics. For the accuracy of language embedding, we measure the mean intersection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 247 + ], + "type": "text", + "content": "over union (mIoU), mean pixel accuracy (mPA), mean precision (mP), and mean average precision (mAP) based on our annotations. The rendering speed (FPS) is measured by rendering the images with language features at a consistent resolution. Additionally, model efficiency is evaluated based on CPU and GPU memory usage during training, as well as data storage requirements and training duration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 255, + 391, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 391, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 391, + 267 + ], + "type": "text", + "content": "5.2. Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 273, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 357 + ], + "type": "text", + "content": "We compare our method both qualitatively and quantitatively with DFF [22], LeRF [21], and 3DOVS [26] on our annotated Mip-NeRF360 dataset using a single RTX3090 GPU, following their default parameters but at the same resolution as our method. In addition, we compare our method with LeRF [21] quantitatively on the LeRF dataset, and the results are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 361, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 637 + ], + "type": "text", + "content": "Qualitative Results. Fig. 9 displays a qualitative comparison of the novel view synthesis and semantic embedding results, demonstrating our method's efficacy in querying challenging objects in both indoor and outdoor scenes. Our approach notably delivers the highest visual rendering quality and query accuracy across all the tested scenes. Specifically, DFF [22] fails to identify \"asphalt ground\" in scene \"bicycle\" and \"flower\" in scene \"garden\". This may be caused by its use of LSeg [24], which is unstable to compute correct features in complex scenes. Moreover, due to predetermined query categories during training, 3DOVS [26] shows poor performance in scenes with complex objects. While LERF [21] can locate queried objects, its grid-based scene representation limits its ability to define clear boundaries. In contrast, our point-based approach supports high-frequency embedded semantic features, allowing for enhanced spatial semantic accuracy simply by incorporating more 3D Gaussians into scenes. This is facilitated by our quantization method, which substantially lowers memory costs. Additionally, our adaptive semantic smoothing technique effectively manages the high variance and ambiguity observed from different viewpoints, selectively applying spatial consistency as required." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "Quantitative Results. Tab. 1 presents a comparison across various metrics, including novel view synthesis quality, open-vocabulary query accuracy, and computational efficiency. We report both host memory and video memory usage, as well as the disk space used for storing the learned language features. Our approach outperforms others in ren" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5338" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 147, + 137 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 147, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 147, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 147, + 137 + ], + "type": "image", + "image_path": "d3beea16a26d33e09f99730e0d5b41a77ddcdf8c59197b36f7a25b82b35627c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 147, + 70, + 248, + 137 + ], + "blocks": [ + { + "bbox": [ + 147, + 70, + 248, + 137 + ], + "lines": [ + { + "bbox": [ + 147, + 70, + 248, + 137 + ], + "spans": [ + { + "bbox": [ + 147, + 70, + 248, + 137 + ], + "type": "image", + "image_path": "d748edfd3cd3f59bd045d1a77a650b929453b4dd21daeb0138bc6903c7a10443.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 70, + 347, + 137 + ], + "blocks": [ + { + "bbox": [ + 248, + 70, + 347, + 137 + ], + "lines": [ + { + "bbox": [ + 248, + 70, + 347, + 137 + ], + "spans": [ + { + "bbox": [ + 248, + 70, + 347, + 137 + ], + "type": "image", + "image_path": "5bab43e8ab8304a39eadbceb9e36e2d0438f6356c15570d260c6d0c6f3843305.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 347, + 70, + 446, + 137 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 446, + 137 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 446, + 137 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 446, + 137 + ], + "type": "image", + "image_path": "11aaf59852aa68692c18124b2bd760406f1c6a65389dee0c956f5652f49b1f8c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 446, + 70, + 545, + 137 + ], + "blocks": [ + { + "bbox": [ + 446, + 70, + 545, + 137 + ], + "lines": [ + { + "bbox": [ + 446, + 70, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 446, + 70, + 545, + 137 + ], + "type": "image", + "image_path": "b23130d91e34e39b7989f0534e9f4dab141d1c2b61757408b4206fecb83c5f96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 137, + 147, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 137, + 147, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 147, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 147, + 205 + ], + "type": "image", + "image_path": "e813163128efd3975fc6e18de4b4fa9147484b5c7945fea6ad290b440e9579f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 147, + 137, + 248, + 205 + ], + "blocks": [ + { + "bbox": [ + 147, + 137, + 248, + 205 + ], + "lines": [ + { + "bbox": [ + 147, + 137, + 248, + 205 + ], + "spans": [ + { + "bbox": [ + 147, + 137, + 248, + 205 + ], + "type": "image", + "image_path": "53c1b0fc2c2a55285987f5d62e7efc8f62af638d2a74e96bbf8442095e06f983.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 248, + 137, + 347, + 205 + ], + "blocks": [ + { + "bbox": [ + 248, + 137, + 347, + 205 + ], + "lines": [ + { + "bbox": [ + 248, + 137, + 347, + 205 + ], + "spans": [ + { + "bbox": [ + 248, + 137, + 347, + 205 + ], + "type": "image", + "image_path": "037de452f3342c5117320faaab01dedf77847c47c4fcd46d7ab24598fdc48efc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 347, + 137, + 446, + 205 + ], + "blocks": [ + { + "bbox": [ + 347, + 137, + 446, + 205 + ], + "lines": [ + { + "bbox": [ + 347, + 137, + 446, + 205 + ], + "spans": [ + { + "bbox": [ + 347, + 137, + 446, + 205 + ], + "type": "image", + "image_path": "2bc9562fa882d2f53c8c20ae1020f90cc7c11f63c2c115eab513c2b22d5b0f60.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 446, + 137, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 446, + 137, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 446, + 137, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 446, + 137, + 545, + 205 + ], + "type": "image", + "image_path": "e7e0669e02efc1d959a597141497c1d4a145c24f40cc97a389602f33b76ac837.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 49, + 205, + 147, + 270 + ], + "blocks": [ + { + "bbox": [ + 49, + 205, + 147, + 270 + ], + "lines": [ + { + "bbox": [ + 49, + 205, + 147, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 205, + 147, + 270 + ], + "type": "image", + "image_path": "28216add99a92599650e64576905b360ccac7ba2b6958deec5f18e304d1c9391.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 147, + 205, + 248, + 270 + ], + "blocks": [ + { + "bbox": [ + 147, + 205, + 248, + 270 + ], + "lines": [ + { + "bbox": [ + 147, + 205, + 248, + 270 + ], + "spans": [ + { + "bbox": [ + 147, + 205, + 248, + 270 + ], + "type": "image", + "image_path": "80544ede6198e2209383531f6712cabd3a62203c7a9f3956b61a52693aa49f48.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 248, + 205, + 347, + 270 + ], + "blocks": [ + { + "bbox": [ + 248, + 205, + 347, + 270 + ], + "lines": [ + { + "bbox": [ + 248, + 205, + 347, + 270 + ], + "spans": [ + { + "bbox": [ + 248, + 205, + 347, + 270 + ], + "type": "image", + "image_path": "13799b202273aa181ffc8f98293eba2efacdcf970402eca3433548ee0d449034.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 347, + 205, + 446, + 270 + ], + "blocks": [ + { + "bbox": [ + 347, + 205, + 446, + 270 + ], + "lines": [ + { + "bbox": [ + 347, + 205, + 446, + 270 + ], + "spans": [ + { + "bbox": [ + 347, + 205, + 446, + 270 + ], + "type": "image", + "image_path": "8fd951d288bc09b025787757ba4b4061f0e46bb00ed294028f79fe2bcfd85086.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 446, + 205, + 545, + 270 + ], + "blocks": [ + { + "bbox": [ + 446, + 205, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 446, + 205, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 446, + 205, + 545, + 270 + ], + "type": "image", + "image_path": "1876ff650db292cba9bb2f759b11be5b76cb8bd9b5f7d2d88ca7ca4c112b3bcb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 49, + 270, + 147, + 335 + ], + "blocks": [ + { + "bbox": [ + 49, + 270, + 147, + 335 + ], + "lines": [ + { + "bbox": [ + 49, + 270, + 147, + 335 + ], + "spans": [ + { + "bbox": [ + 49, + 270, + 147, + 335 + ], + "type": "image", + "image_path": "9da1d5541d70b4d70ad675072c9a07942404d2e098992fe033053ac8846758cf.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 147, + 270, + 248, + 335 + ], + "blocks": [ + { + "bbox": [ + 147, + 270, + 248, + 335 + ], + "lines": [ + { + "bbox": [ + 147, + 270, + 248, + 335 + ], + "spans": [ + { + "bbox": [ + 147, + 270, + 248, + 335 + ], + "type": "image", + "image_path": "4841bbc29e9ff3ba924232f06a9ab5fc837236e245628066c9cbe88a80ebadcb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 248, + 270, + 347, + 335 + ], + "blocks": [ + { + "bbox": [ + 248, + 270, + 347, + 335 + ], + "lines": [ + { + "bbox": [ + 248, + 270, + 347, + 335 + ], + "spans": [ + { + "bbox": [ + 248, + 270, + 347, + 335 + ], + "type": "image", + "image_path": "bb20061b2727c5f0aa8242bceda6dd1e45bc87e42558ce27c0a1024922ee5c52.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 347, + 270, + 446, + 335 + ], + "blocks": [ + { + "bbox": [ + 347, + 270, + 446, + 335 + ], + "lines": [ + { + "bbox": [ + 347, + 270, + 446, + 335 + ], + "spans": [ + { + "bbox": [ + 347, + 270, + 446, + 335 + ], + "type": "image", + "image_path": "5b0ef5f244801cfb7dd89e338437c30a41927bf122775b15a7037c7af5a35d4d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 446, + 270, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 446, + 270, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 446, + 270, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 446, + 270, + 545, + 335 + ], + "type": "image", + "image_path": "59c8227f226b756b184cbb03e2ad54bfb647cf13dbd352d5dde0064304245566.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 49, + 335, + 147, + 401 + ], + "blocks": [ + { + "bbox": [ + 49, + 335, + 147, + 401 + ], + "lines": [ + { + "bbox": [ + 49, + 335, + 147, + 401 + ], + "spans": [ + { + "bbox": [ + 49, + 335, + 147, + 401 + ], + "type": "image", + "image_path": "8f97e957663cc5d829c31840aff6bffcddce141b5507cac0b669fd100eda4273.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 147, + 335, + 248, + 401 + ], + "blocks": [ + { + "bbox": [ + 147, + 335, + 248, + 401 + ], + "lines": [ + { + "bbox": [ + 147, + 335, + 248, + 401 + ], + "spans": [ + { + "bbox": [ + 147, + 335, + 248, + 401 + ], + "type": "image", + "image_path": "ebaa8957bd0a82b73a62ce68a0b6012f87f84be6a67e7246b4c18c7cf73896c6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 248, + 335, + 347, + 401 + ], + "blocks": [ + { + "bbox": [ + 248, + 335, + 347, + 401 + ], + "lines": [ + { + "bbox": [ + 248, + 335, + 347, + 401 + ], + "spans": [ + { + "bbox": [ + 248, + 335, + 347, + 401 + ], + "type": "image", + "image_path": "2a23a3c3991b88851ac495810f6497ee65d0b192c5d2d47349a1040851f23790.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 347, + 335, + 446, + 401 + ], + "blocks": [ + { + "bbox": [ + 347, + 335, + 446, + 401 + ], + "lines": [ + { + "bbox": [ + 347, + 335, + 446, + 401 + ], + "spans": [ + { + "bbox": [ + 347, + 335, + 446, + 401 + ], + "type": "image", + "image_path": "68a6719d21c2b94a70d85fced18fd402cf7947b6b67e416fbdf22e255a5415b7.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 446, + 335, + 545, + 401 + ], + "blocks": [ + { + "bbox": [ + 446, + 335, + 545, + 401 + ], + "lines": [ + { + "bbox": [ + 446, + 335, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 446, + 335, + 545, + 401 + ], + "type": "image", + "image_path": "d0b3f0c998e79d873897da18f4ffea36eb25b15762a8f63155a556533e265985.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 49, + 402, + 147, + 468 + ], + "blocks": [ + { + "bbox": [ + 49, + 402, + 147, + 468 + ], + "lines": [ + { + "bbox": [ + 49, + 402, + 147, + 468 + ], + "spans": [ + { + "bbox": [ + 49, + 402, + 147, + 468 + ], + "type": "image", + "image_path": "4511d76c1f9b16a9f55c8a464f6703eb7a03609287abe4d76eeb2bf556e92cd8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 470, + 127, + 480 + ], + "lines": [ + { + "bbox": [ + 70, + 470, + 127, + 480 + ], + "spans": [ + { + "bbox": [ + 70, + 470, + 127, + 480 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 147, + 402, + 248, + 468 + ], + "blocks": [ + { + "bbox": [ + 147, + 402, + 248, + 468 + ], + "lines": [ + { + "bbox": [ + 147, + 402, + 248, + 468 + ], + "spans": [ + { + "bbox": [ + 147, + 402, + 248, + 468 + ], + "type": "image", + "image_path": "7f59be792d90f005260b7eb64cbff61da4b4b427ab8be9832d3dd48675dc9901.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 470, + 209, + 479 + ], + "lines": [ + { + "bbox": [ + 188, + 470, + 209, + 479 + ], + "spans": [ + { + "bbox": [ + 188, + 470, + 209, + 479 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 248, + 402, + 347, + 468 + ], + "blocks": [ + { + "bbox": [ + 248, + 402, + 347, + 468 + ], + "lines": [ + { + "bbox": [ + 248, + 402, + 347, + 468 + ], + "spans": [ + { + "bbox": [ + 248, + 402, + 347, + 468 + ], + "type": "image", + "image_path": "289c48aab2956e1745fdba16c45d9fa770d570be5b88f17ddf8bfe0b6e2cf011.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 470, + 306, + 479 + ], + "lines": [ + { + "bbox": [ + 289, + 470, + 306, + 479 + ], + "spans": [ + { + "bbox": [ + 289, + 470, + 306, + 479 + ], + "type": "text", + "content": "DFF" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 347, + 402, + 446, + 468 + ], + "blocks": [ + { + "bbox": [ + 347, + 402, + 446, + 468 + ], + "lines": [ + { + "bbox": [ + 347, + 402, + 446, + 468 + ], + "spans": [ + { + "bbox": [ + 347, + 402, + 446, + 468 + ], + "type": "image", + "image_path": "fbfeb8890e2cb36af4f85eae78013568252e365ee28282210493d34e567ca15c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 470, + 406, + 479 + ], + "lines": [ + { + "bbox": [ + 387, + 470, + 406, + 479 + ], + "spans": [ + { + "bbox": [ + 387, + 470, + 406, + 479 + ], + "type": "text", + "content": "LERF" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 489, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 46, + 489, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 547, + 522 + ], + "type": "text", + "content": "Figure 3. Comparison of novel view synthesis and query relevance visualization. Left to right: Ground truth novel view synthesis, novel view images with relevance visualization from our method, DFF [22], LeRF [21], and 3DOVS [26]. Top to bottom: Query words \"asphalt ground\", \"bicycle\", \"jar of coconut oil\", \"flower\", \"LEGO Technic 856 Bulldozer\", and \"brown shoes\"." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 446, + 402, + 545, + 468 + ], + "blocks": [ + { + "bbox": [ + 446, + 402, + 545, + 468 + ], + "lines": [ + { + "bbox": [ + 446, + 402, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 446, + 402, + 545, + 468 + ], + "type": "image", + "image_path": "1697d935bb8456de43e540f7efab943c1f748a4b1bc64223b314a1acba3d0f90.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 470, + 512, + 479 + ], + "lines": [ + { + "bbox": [ + 483, + 470, + 512, + 479 + ], + "spans": [ + { + "bbox": [ + 483, + 470, + 512, + 479 + ], + "type": "text", + "content": "3DOVS" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 532, + 289, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 289, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 289, + 651 + ], + "type": "text", + "content": "dering quality and semantic query accuracy, while also offering lower computational demands and a significant speed increase, nearly 100 times faster in inference. It's noteworthy that methods like 3DOVS and DFF require extensive memory and storage due to their use of raw language features during training. In contrast, our use of the quantization scheme facilitates the incorporation of detailed semantics into complex 3D scenes with numerous 3D Gaussians, and concurrently achieves the most efficient storage utilization among the baseline methods." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 47, + 659, + 185, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 185, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 185, + 673 + ], + "type": "text", + "content": "5.3. Open-vocabulary Query" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Fig. 5 illustrates the results of different open-vocabulary queries. We use a diverse range of vocabulary categories to identify objects in scenes, such as visual attribute terms" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 532, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 593 + ], + "type": "text", + "content": "like \"green\", and subjective adjectives like \"cute\". Furthermore, the figure shows our method's effectiveness in identifying both large-scale objects like \"lego\" and specific components of an object, such as a \"engine\". We show more examples in the supplementary material." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 306, + 599, + 400, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 400, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 400, + 613 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We demonstrate the results of ablation studies in Tab. 2 and Fig. 6. The results show that embedding uncertainty without spatial smoothing of semantic features leads to suboptimal optimization. Conversely, using MLP solely for spatial smoothing, without accounting for semantic variance across multiple views, also impedes the optimization of precise language features. However, our full model, combining uncertainty with MLP smoothing in an adaptive man" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5339" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 153, + 171 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 153, + 171 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 153, + 171 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 153, + 171 + ], + "type": "image", + "image_path": "402888a12bc39c2a4ae6326453c04a48e7efa8264665a201f4a58a01ddd467e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 155, + 70, + 220, + 171 + ], + "blocks": [ + { + "bbox": [ + 155, + 70, + 220, + 171 + ], + "lines": [ + { + "bbox": [ + 155, + 70, + 220, + 171 + ], + "spans": [ + { + "bbox": [ + 155, + 70, + 220, + 171 + ], + "type": "image", + "image_path": "90e89a48b54af345a5bfc9ce1b4e5c49f3333baa21d7255c1279a88bd5374a45.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 288, + 171 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 288, + 171 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 288, + 171 + ], + "type": "image", + "image_path": "35dc12049086a5880c4f1c8f32ca34fef3badc47fc049b37e6989b6f837637d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 171, + 153, + 268 + ], + "blocks": [ + { + "bbox": [ + 49, + 171, + 153, + 268 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 153, + 268 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 153, + 268 + ], + "type": "image", + "image_path": "ce8d02a2cfe64a9bd217155fce36b8485c63ea44eb522bc09dd5654a9b91f2cf.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 269, + 129, + 278 + ], + "lines": [ + { + "bbox": [ + 72, + 269, + 129, + 278 + ], + "spans": [ + { + "bbox": [ + 72, + 269, + 129, + 278 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 155, + 171, + 220, + 267 + ], + "blocks": [ + { + "bbox": [ + 155, + 171, + 220, + 267 + ], + "lines": [ + { + "bbox": [ + 155, + 171, + 220, + 267 + ], + "spans": [ + { + "bbox": [ + 155, + 171, + 220, + 267 + ], + "type": "image", + "image_path": "38ddaef95e1b73f0ba4894d2323dc97295bb8fde2f5fe1fc87a75aee7e367ff1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 269, + 259, + 279 + ], + "lines": [ + { + "bbox": [ + 184, + 269, + 259, + 279 + ], + "spans": [ + { + "bbox": [ + 184, + 269, + 259, + 279 + ], + "type": "text", + "content": "Rendering Results" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 222, + 171, + 288, + 267 + ], + "blocks": [ + { + "bbox": [ + 222, + 171, + 288, + 267 + ], + "lines": [ + { + "bbox": [ + 222, + 171, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 222, + 171, + 288, + 267 + ], + "type": "image", + "image_path": "b951dfa67d089da120267d2dd2882babfb87ab8d9f23d541e8714eb1784367fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 325, + 168, + 403 + ], + "blocks": [ + { + "bbox": [ + 47, + 288, + 287, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 288, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 288, + 287, + 321 + ], + "type": "text", + "content": "Figure 4. Visual quality comparison of novel view synthesis results. Our method is able to recover more detailed geometry and appearance compared to other methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 325, + 168, + 403 + ], + "lines": [ + { + "bbox": [ + 49, + 325, + 168, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 325, + 168, + 403 + ], + "type": "image", + "image_path": "4d263510b576a0f5c6dbcddbb31ded87118de7da935f15ac5cee470b0b1ad704.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 170, + 325, + 288, + 403 + ], + "blocks": [ + { + "bbox": [ + 170, + 325, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 170, + 325, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 170, + 325, + 288, + 403 + ], + "type": "image", + "image_path": "f887ca98be12d205afd38462acb44ce1686a344cd172ba01725711e8223523f2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 49, + 414, + 168, + 495 + ], + "blocks": [ + { + "bbox": [ + 93, + 404, + 124, + 414 + ], + "lines": [ + { + "bbox": [ + 93, + 404, + 124, + 414 + ], + "spans": [ + { + "bbox": [ + 93, + 404, + 124, + 414 + ], + "type": "text", + "content": "\"green\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 414, + 168, + 495 + ], + "lines": [ + { + "bbox": [ + 49, + 414, + 168, + 495 + ], + "spans": [ + { + "bbox": [ + 49, + 414, + 168, + 495 + ], + "type": "image", + "image_path": "bcac5c00a47638894c19ce06d04ae398edb665cfe168558e62fff794d42753c2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 169, + 414, + 288, + 495 + ], + "blocks": [ + { + "bbox": [ + 214, + 404, + 244, + 414 + ], + "lines": [ + { + "bbox": [ + 214, + 404, + 244, + 414 + ], + "spans": [ + { + "bbox": [ + 214, + 404, + 244, + 414 + ], + "type": "text", + "content": "\"grass\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 414, + 288, + 495 + ], + "lines": [ + { + "bbox": [ + 169, + 414, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 169, + 414, + 288, + 495 + ], + "type": "image", + "image_path": "18437561d7fd9116ae0df5d05f0285ba4104406e461bc54edb2c0ff9980ba66f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 49, + 506, + 168, + 586 + ], + "blocks": [ + { + "bbox": [ + 94, + 496, + 121, + 506 + ], + "lines": [ + { + "bbox": [ + 94, + 496, + 121, + 506 + ], + "spans": [ + { + "bbox": [ + 94, + 496, + 121, + 506 + ], + "type": "text", + "content": "\"lego\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 506, + 168, + 586 + ], + "lines": [ + { + "bbox": [ + 49, + 506, + 168, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 506, + 168, + 586 + ], + "type": "image", + "image_path": "09465d65994fa22688be2ca6fbc13c452c79f92ab53284605eecb622b7cb8f6e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 588, + 132, + 597 + ], + "lines": [ + { + "bbox": [ + 84, + 588, + 132, + 597 + ], + "spans": [ + { + "bbox": [ + 84, + 588, + 132, + 597 + ], + "type": "text", + "content": "\"sheep toy\"" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 608, + 264, + 619 + ], + "lines": [ + { + "bbox": [ + 69, + 608, + 264, + 619 + ], + "spans": [ + { + "bbox": [ + 69, + 608, + 264, + 619 + ], + "type": "text", + "content": "Figure 5. Images of various open-vocabulary queries." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 169, + 506, + 288, + 586 + ], + "blocks": [ + { + "bbox": [ + 211, + 496, + 247, + 506 + ], + "lines": [ + { + "bbox": [ + 211, + 496, + 247, + 506 + ], + "spans": [ + { + "bbox": [ + 211, + 496, + 247, + 506 + ], + "type": "text", + "content": "\"engine\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 506, + 288, + 586 + ], + "lines": [ + { + "bbox": [ + 169, + 506, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 169, + 506, + 288, + 586 + ], + "type": "image", + "image_path": "b430bddc68b24a48179b9897face2ea8520730ab266e7b404f1ae61a24427a7f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 588, + 242, + 597 + ], + "lines": [ + { + "bbox": [ + 216, + 588, + 242, + 597 + ], + "spans": [ + { + "bbox": [ + 216, + 588, + 242, + 597 + ], + "type": "text", + "content": "\"cute\"" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "ner effectively diminishes ambiguity and enhances the mean average precision (mAP) metric. Furthermore, integrating DINO features significantly improves the definition of object query boundaries. The load balancing loss, introduced during the quantization phase, results in a more utilized discrete feature space, facilitating the distinguish of objects with similar semantics, thereby boosting overall accuracy." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 244 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 244 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 244 + ], + "type": "image", + "image_path": "8c9d528545a599a55a6dcd1bf72cc2ed755f75ce75479d2ef42208e04ad08642.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 251, + 511, + 263 + ], + "lines": [ + { + "bbox": [ + 340, + 251, + 511, + 263 + ], + "spans": [ + { + "bbox": [ + 340, + 251, + 511, + 263 + ], + "type": "text", + "content": "Figure 6. Comparison of ablation experiments." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 277, + 378, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 277, + 378, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 277, + 378, + 289 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 297, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 465 + ], + "type": "text", + "content": "We present Language Embedded 3D Gaussians, a novel scene representation designed for open-vocabulary query tasks. Our method successfully embeds quantized compact semantics features onto massive 3D Gaussians, while only maintaining minimal memory and storage requirements. To address semantic inconsistencies across different viewpoints, we propose a feature smoothing procedure that adaptively lowers the spatial frequency of embedded semantic features, guided by the uncertainty values learned on the 3D Gaussians. The result is a highly effective scene representation that not only enables high-quality novel view synthesis but also provides high accuracy in open-vocabulary querying, all achieved with modest computational resources." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 476, + 546, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 546, + 657 + ], + "type": "text", + "content": "Limitations and Future works. Although DINO features improve object boundary detection, they fall short in pinpointing fine-grained object geometries at high resolutions when using CLIP-derived semantics. Furthermore, while vector quantization can efficiently compress scene-specific semantic feature space, it may compromise the finer-grained semantic information in the feature space due to its inherent clustering of semantic concepts. In addition, the detection of highly reflective or translucent objects, like televisions and mirrors, poses a challenge. These limitations might be overcome with more advanced visual-language models and native per-pixel semantic features. Nevertheless, our approach can be further adapted for broader open-vocabulary tasks, such as editing and generation of semantic-level scene objects." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgment. This work was supported by the National Natural Science Foundation of China (Project Number: 62372025 and 61932003) and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5340" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 190 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 290 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. arXiv preprint arXiv:2304.06706, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), 2021. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "text", + "content": "[6] Gaurav Chaurasia, Sylvain Duchene, Olga Sorkine-Hornung, and George Drettakis. Depth synthesis and local warps for plausible image-based navigation. ACM Transactions on Graphics (TOG), 32(3), 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 287, + 434 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 479 + ], + "type": "text", + "content": "[8] Zhi Deng, Haoyao Xiao, Yining Lang, Hao Feng, and Juyong Zhang. Multi-scale hash encoding based neural geometry representation. Computational Visual Media, pages 1-18, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 534 + ], + "type": "text", + "content": "[9] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[10] Martin Eisemann, Bert De Decker, Marcus Magnor, Philippe Bekaert, Edilson De Aguiar, Naveed Ahmed, Christian Theobalt, and Anita Sellent. Floating textures. In Computer graphics forum, pages 409-418. Wiley Online Library, 2008. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 634 + ], + "type": "text", + "content": "[11] Francis Engelmann, Fabian Manhardt, Michael Niemeyer, Keisuke Tateno, Marc Pollefeys, and Federico Tombari. Open-set 3d scene segmentation with rendered novel views. 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 713 + ], + "type": "text", + "content": "[13] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ple and efficient sparsity. The Journal of Machine Learning Research, 23(1):5232-5270, 2022. 4, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 149 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 205 + ], + "type": "text", + "content": "[15] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 2022 International Conference on 3D Vision (3DV), pages 1–11. IEEE, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 249 + ], + "type": "text", + "content": "[16] Michael Goesele, Noah Snavely, Brian Curless, Hugues Hoppe, and Steven M Seitz. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 293 + ], + "type": "text", + "content": "[17] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[18] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. arXiv preprint arXiv:2210.01055, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 425 + ], + "type": "text", + "content": "[19] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, Joshua B. Tenenbaum, Celso Miguel de Melo, Madhava Krishna, Liam Paull, Florian Shkurti, and Antonio Torralba. Conceptfusion: Open-set multimodal 3d mapping. Robotics: Science and Systems (RSS), 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 469 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (ToG), 42(4):1-14, 2023. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "text", + "content": "[21] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 1, 2, 3, 6, 7, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[22] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nerf for editing via feature field distillation. Advances in Neural Information Processing Systems, 35:23311-23330, 2022. 1, 2, 3, 6, 7, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 570, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 645 + ], + "type": "text", + "content": "[23] Abhijit Kundu, Kyle Genova, Xiaoqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[24] Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl. Language-driven semantic segmentation. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[25] Gengxin Liu, Oliver van Kaick, Hui Huang, and Ruizhen Hu. Active self-training for weakly supervised 3d scene seman" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5341" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "tic segmentation. Computational Visual Media, pages 1-14, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 149 + ], + "type": "text", + "content": "[26] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. 3d open-vocabulary segmentation with foundation models. arXiv preprint arXiv:2305.14093, 2023. 2, 3, 6, 7, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[27] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. arXiv preprint arXiv:2305.14093, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "text", + "content": "[28] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 286, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 286, + 272 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 316 + ], + "type": "text", + "content": "[30] Yuheng Lu, Chenfeng Xu, Xiaobao Wei, Xiaodong Xie, Masayoshi Tomizuka, Kurt Keutzer, and Shanghang Zhang. Open-vocabulary point-cloud object detection without 3d annotation. 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 381 + ], + "type": "text", + "content": "[31] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "text", + "content": "[32] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 480 + ], + "type": "text", + "content": "[33] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 525 + ], + "type": "text", + "content": "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 581 + ], + "type": "text", + "content": "[35] Songyou Peng, Kyle Genova, Chiyu \"Max\" Jiang, Andrea Tagliasacchi, Marc Pollefeys, and Thomas Funkhouser. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 582, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 646 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[37] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[38] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Tom Funkhouser, and Vittorio Ferrari. Urban radiance fields. CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 138 + ], + "type": "text", + "content": "[39] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulò, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3d scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043–9052, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[40] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM siggraph 2006 papers, pages 835-846. 2006. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 226 + ], + "type": "text", + "content": "[41] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 281 + ], + "type": "text", + "content": "[42] Ayca Takmaz, Elisabetta Fedele, Robert W. Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. OpenMask3D: Open-Vocabulary 3D Instance Segmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 545, + 348 + ], + "type": "text", + "content": "[43] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P. Srinivasan, Jonathan T. Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "text", + "content": "[44] Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In 2022 International Conference on 3D Vision (3DV), pages 443-453. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[45] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 470 + ], + "type": "text", + "content": "[46] Bing Wang, Lu Chen, and Bo Yang. Dm-nerf: 3d scene geometry decomposition and manipulation from 2d images. arXiv preprint arXiv:2208.07227, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 471, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 504 + ], + "type": "text", + "content": "[47] Yuang Wang, Xingyi He, Sida Peng, Haotong Lin, Hujun Bao, and Xiaowei Zhou. Autorecon: Automated 3d object discovery and reconstruction. In CVPR, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 505, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 545, + 557 + ], + "type": "text", + "content": "[48] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. arXiv pre-print, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 613 + ], + "type": "text", + "content": "[49] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5438-5448, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 669 + ], + "type": "text", + "content": "[50] Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning unified representation of language, image and point cloud for 3d understanding. arXiv preprint arXiv:2212.05171, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[51] Le Xue, Ning Yu, Shu Zhang, Junnan Li, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip-2: Towards scalable multimodal pre-training for 3d understanding, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5342" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 554 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[52] Guo-Wei Yang, Zheng-Ning Liu, Dong-Yang Li, and Hao-Yang Peng. Jnerf: An efficient heterogeneous nerf model zoo based on jitter. Computational Visual Media, 9(2):401-404, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[53] Jianing Yang, Xuweiyi Chen, Shengyi Qian, David Fouhey, and Joyce Chai. Chat with nerf: Grounding 3d objects in neural radiance field through dialog, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 195 + ], + "type": "text", + "content": "[54] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F. Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "text", + "content": "[55] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 274 + ], + "type": "text", + "content": "[56] Junbo Zhang, Runpei Dong, and Kaisheng Ma. Clip-fo3d: Learning free open-world 3d scene representations from 2d dense clip, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 286, + 308 + ], + "type": "text", + "content": "[57] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 342 + ], + "type": "text", + "content": "[58] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "spans": [ + { + "bbox": [ + 49, + 343, + 286, + 386 + ], + "type": "text", + "content": "[59] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xu-peng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. arXiv preprint arXiv:2112.02413, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 388, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 388, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 388, + 286, + 453 + ], + "type": "text", + "content": "[60] Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser, Leonidas Guibas, Hao Su, and Kyle Genova. Nerflets: Local radiance fields for efficient structure-aware 3d scene representation from 2d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8274-8284, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 455, + 286, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 286, + 509 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 286, + 509 + ], + "type": "text", + "content": "[61] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 49, + 511, + 286, + 554 + ], + "type": "text", + "content": "[62] Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Pointclip v2: Prompting clip and gpt for powerful 3d open-world learning. arXiv preprint arXiv:2211.11682, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5343" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_content_list.json b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b1c1a1dd08e33f6df2a0998bfc18d72051d412e5 --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_content_list.json @@ -0,0 +1,1748 @@ +[ + { + "type": "text", + "text": "Language Model Guided Interpretable Video Action Reasoning", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ning Wang $^{1*}$ , Guangming Zhu $^{1\\dagger,*}$ , HS Li $^{1}$ , Liang Zhang $^{1\\dagger}$ , Syed Afaq Ali Shah $^{2}$ , Mohammed Bennamoun $^{3}$ , Xidian University, $^{2}$ Edith Cowan University, $^{3}$ University of Western Australia", + "bbox": [ + 86, + 179, + 915, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ningwang,hsli}@stu.xidian.edu.cn，{gmzhu,liangzhang}@xidian.edu.cn,", + "bbox": [ + 186, + 218, + 813, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "afaq.shah@ecu.edu.au, mohammed.bennamoun@uwa.edu.au", + "bbox": [ + 272, + 237, + 725, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While neural networks have excelled in video action recognition tasks, their \"black-box\" nature often obscures the understanding of their decision-making processes. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. These models, however, usually fall short in performance compared to their \"black-box\" counterparts. In this work, we present a new framework named Language-guided Interpretable Action Recognition framework (LaIAR). LaIAR leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we redefine the problem of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on two complex video action datasets, Charades & CAD-120, validates the improved performance and interpretability of our LaIAR framework. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR.", + "bbox": [ + 75, + 319, + 473, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building on the advancements of deep learning in image recognition [9, 16, 31], neural network (NN) models have become the leading approach for video-related challenges, including action recognition [18, 25, 30]. Yet, many of the top-tier action recognition techniques [19, 34] deploy NNs in an opaque, black-box fashion. This lack of transparency does not offer clear justification for their decisions, hindering their utility in various real-world contexts [13], especially those with rigorous security demands. These considerations drive us to develop an action reasoning system that pairs exceptional performance with clear interpretability.", + "bbox": [ + 75, + 693, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Action : “Someone is standing up from somewhere” Time", + "text_level": 1, + "bbox": [ + 506, + 285, + 890, + 299 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7301122c9a65b8e52e309af6da4b44743dedf2800b8c7f61892606943ccd6386.jpg", + "image_caption": [ + "(a) An illustration example of relationship transition in the videos." + ], + "image_footnote": [], + "bbox": [ + 504, + 300, + 666, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0e14a0c61638542a95e0520ff39b4ae5c57466ba70388a63799699f7d21c478.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 300, + 784, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ed8c89f66ce624c56a36b8ddeafa9565c41240285e7974ba78fe9d445ef8f57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 301, + 887, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f0578ad47c906b7a5f213f40b53bef711deb37dcbae8835b275818ceee13d81.jpg", + "image_caption": [ + "(b) Traditional two-stage interpretable reasoning model." + ], + "image_footnote": [], + "bbox": [ + 504, + 381, + 665, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff44629ff6dafbdc2ebb43b1840f890d6388078fa6877f7b91eb59663ef0df4b.jpg", + "image_caption": [ + "language model", + "first stage second stage" + ], + "image_footnote": [], + "bbox": [ + 665, + 381, + 803, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c1ada2e923cfbfbe8a8ab54f98fb6d30ca55bafe56d6f4ade6878cc2fb58df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 465, + 606, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8225b8383cd0fbe7a010e04eb985ec0d08ebccac9905d55abfbbb2967953e305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 606, + 465, + 691, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2956b3b4d3c93f4d7f545ded966cdc1da86807fee68ab7d2fb45ca2f97501aa0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 465, + 803, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2984a53c740bfe00f5dcc495001af5bc0ac4119827729c768a7196f5f6e7f61f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 465, + 885, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e23a94b8156380744e980fc765f2fc1e3a337cfcad2fdc8f2411ef9a6eec9b33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 539, + 602, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9349b1ce5232d5a3984b6ff140dd1c92ecdfe1960f8241849304bea75db44cb7.jpg", + "image_caption": [ + "(c) Language model guided video interpretable reasoning model." + ], + "image_footnote": [], + "bbox": [ + 604, + 539, + 691, + 603 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/849f7cc58e06e0e8af6de2ddf321ed5084b5d36310be59c97a78cee60df8895f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 539, + 803, + 590 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d11af26f49783e77a436264323eb94144aead1e7e13ddc205b939d6b75c510a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 539, + 883, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. (a) An example of action that can be decomposed into relationship transitions (i.e., when the transition is 'sitting on' $\\rightarrow$ 'not contacting' between $<\\text{person, bed}>$ pair, it represents the action \"Someone is standing up from somewhere\"). (b) Traditional two-stage methods usually predict the scene graph first, and then use language models to capture the semantic-level relationship transitions. (c) Our method exploits a language model to guide the video model to capture the relationship transition during training. During inference, our method processes videos and directly recognizes actions, providing supportive evidence.", + "bbox": [ + 496, + 630, + 890, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most of the current interpretable action recognition techniques [21, 22, 29] aim to elucidate the decision-making process of NNs using post-hoc explanations, with a particular emphasis on gradient-based and perturbation-based approaches. However, despite notable advancements, these explanations can be problematic because they might not be faithful to what the network computes, as highlighted by [27]. A compelling direction in interpretability re", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Ning Wang and Guangming Zhu are co-first authors.", + "bbox": [ + 94, + 875, + 380, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Liang Zhang and Guangming Zhu are both the corresponding authors.", + "bbox": [ + 94, + 887, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volves around the concept of built-in explanation models [10, 12, 24, 42]. The essence of these models is their inherent interpretability right from the design stage. Recent strategies decompose a complex action into temporal transitions of human-object relationships, drawing inspiration from the event segmentation theory [17]. An illustration in Figure 1 (a) depicts that for a relationship involving a $<$ person, $\\mathrm{bed}>$ pairing, a transition sequence from 'sitting on' to 'not contacting' signifies the action of \"Someone is standing up from somewhere\". This methodology facilitates action recognition by pinpointing semantic transitions through language models, offering a granular insight into action execution. As shown in Figure 1 (b), Jin and Ou et al. [12, 24] extract spatio-temporal scene graphs from video content and apply Markov Logic Network (MLN) based probabilistic logical inference and relation reasoning graphs to create an interpretable representation for a variety of complex actions, respectively. However, it is believed that such models will perform worse than their black-box alternatives [8]. Moreover, these methods divide the process into two stages, namely scene graph prediction and relation modeling. Optimizing these components separately might lead to sub-optimal results. In this paper, we propose to harness the explicit logical inference rules of an interpretable language model to guide the learning process of a video black-box model. Interpretability and strong performance can be attained by focusing solely on the video model during the inference stage. To achieve this, two main challenges arise: 1) Designing a language model that can automatically grasp logical reasoning patterns, sidestepping manual rule creation. 2) Developing a decoupled language-video model architecture that enables the language model to guide the video model's training process.", + "bbox": [ + 76, + 90, + 472, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we have developed a new framework called Language-guided Interpretable Action Recognition framework named LaIAR. As depicted in Figure 1 (c), LaIAR constructs an action recognition model that both implicitly and explicitly exploits fine-grained knowledge of relationship transitions from an interpretable built-in language model. Specifically, we use dynamic token transformers (DT-Former) to both video and language inputs, selectively focusing on important relationships in a data-driven manner, and disregarding non-contributory ones. We aim to redefine the traditional decision interpretation challenge of video models towards a visual-language relation alignment problem. The relationship prioritization determined by the language model then explicitly guides the video model in identifying the most relevant relationships. We propose a learning strategy to facilitate knowledge transfer between language and video to improve the performance of the video model. A key feature of LaIAR is its modular design: during the inference phase, only RGB data serves as input to predict actions, providing", + "bbox": [ + 76, + 599, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a direct and transparent justification.", + "bbox": [ + 500, + 90, + 743, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are three-fold: 1) We propose a novel LaIAR framework that can automatically mine fine-grained relation transitions from data and create interpretable representations for various complex actions. 2) We design a decoupled cross-modal knowledge transfer architecture that leverages useful knowledge from language models to improve the performance and interpretability of the video model at training time, and achieves high-performance interpretable reasoning for videos at test time. 3) Our method achieves state-of-the-art results on two large-scale action recognition benchmarks.", + "bbox": [ + 496, + 107, + 893, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 287, + 640, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Interpretable Video Action Recognition", + "text_level": 1, + "bbox": [ + 500, + 313, + 838, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interpretable video action recognition methods can be categorized into two types: post-hoc method and built-in method. post-hoc techniques generate explanations for the network's decision-making process after the network is trained. [22] introduced an interpretable and easy plug-in spatial-temporal attention mechanism for video action recognition to improve the interpretability of the model for video action recognition. [29] developed an interpretable temporal convolutional network to explain the decision-making process of action recognition through each of the learned filters in a Res-TCN. [21] combined both global dynamics and local details to learn human action, using gradient-weighted class activation mapping (Grad-CAM) to visualize the model's attention to action-critical regions. Although these methods have the advantage of not imposing any model constraints, they may be incomplete or unfaithful to the model's reasoning [27]. In contrast, built-in methods restrict the interpretation to be consistent with the model's inferences. [42] approached action reasoning by modeling semantic-level state transitions between two consecutive frames as defined by domain experts. In [10], a method is proposed to achieve interpretability of action recognition by incorporating qualitative spatial reasoning and extracting salient relation chains. Some recent methods, like [12, 24] decompose complex action into continuous relationship transitions according to the event segmentation theory [17]. These methods model the relationship transitions at the semantic level to recognize actions. In this paper, we propose to construct a high performance interpretable-by-design action classifier by guiding a video model with an interpretable language model.", + "bbox": [ + 496, + 337, + 893, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adaptive Inference in Transformers.", + "text_level": 1, + "bbox": [ + 500, + 816, + 815, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As their popularity soars, adaptive inference for language and vision transformers has caught the attention of researchers. In [37], an adaptive language transformer is proposed to achieve fixed-scale reduction of the input se", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18879", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "quence to improve inference speed by dynamically selecting important tokens and removing the irrelevant ones. In [23], a threshold mechanism is introduced to determine the importance of each token and dynamically select the tokens according to the importance of the input sequence. [14] used the mean of attention matrix column values of the transformers to determine the importance score of each token, facilitating token pruning. [38] developed an adaptive token generation mechanism to determine the required number and size of tokens, thereby reducing the computational and memory overhead of the model on images. [32] devised a token selection framework to dynamically select important tokens across the temporal and spatial dimensions of the video input. In this paper, we propose a lightweight token selection method based on Gumbel-Softmax and apply it to our cross-modal transformers for spatio-temporal token selection.", + "bbox": [ + 75, + 90, + 472, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Cross-modal Knowledge Transferring", + "text_level": 1, + "bbox": [ + 76, + 356, + 405, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The past few years have seen an increasing interest in cross-modal knowledge transfer techniques for detection and segmentation tasks. [20] introduced a method to transfer the motion-related knowledge of unlabeled videos to Human-Object Interactions (HOI) detection to infer rare or unseen HOIs. In [40], reliable domain-invariant sound cues are exploited to help video activity recognition models adapt to video distribution shifts. Lately, knowledge distillation techniques have been extended to transfer knowledge across different modalities. For instance, [18] proposed a decomposed cross-modal distillation framework to improve RGB-based temporal action detection by transferring knowledge from the optical flow modality. Similarly, [39] proposed a modified knowledge distillation method that boosts the performance of single-modal 3D captioning by transferring color and texture-aware information from 2D images into 3D object representations. In contrast to these methods, we propose a well-designed knowledge-guided framework to enable cross-modal learning by decoupling information transfer in video and language.", + "bbox": [ + 75, + 378, + 472, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Approach", + "text_level": 1, + "bbox": [ + 76, + 694, + 305, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method is designed to exploit multimodality by enabling information transfer from language descriptions to videos. This enables the video model to effectively learn from the language model effectively. This is achieved by the video model mimicking the output of the language model, thereby leveraging the intrinsic capabilities of the language model. Specifically, the video frames and the language description (represented as a spatio-temporal scene graph in [11]) are first processed by the encoding network to extract the paired visual and semantic relationship representations. Then, these paired visual and semantic relationship representations are sepa", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rately fed to DT-Former module, which models the key relationship transition for action recognition. Finally, we propose a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning) to improve the performance and interpretability of the video model by facilitating the knowledge transfer from the language model to the video model. An overview of our proposed method is shown in Figure 2 (a). Note that, in our proposed approach, only the video model is employed for inference once training is complete.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 500, + 252, + 633, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Video and Language Encoder", + "text_level": 1, + "bbox": [ + 500, + 275, + 759, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a video consisting of $T$ frames with $N$ entities of either human or object classes, we use Faster R-CNN [26] with ResNet-101 [16] backbone to detect these entities and extract their features from the video. For the frame $I_{t}$ at time step $t$ , the visual features $\\{v_{(t,1)}, v_{(t,2)}, \\dots, v_{(t,N)}\\}$ , bounding boxes $\\{b_{(t,1)}, b_{(t,2)}, \\dots, b_{(t,N)}\\}$ and object category $\\{c_{(t,1)}, c_{(t,2)}, \\dots, c_{(t,N)}\\}$ of the objects proposals are supplied by the detector. Between each $<\\text{human, object}>$ pair in the frame, there is a set of relationships $R_{t} = \\{r_{(t,1)}, r_{(t,2)}, \\dots, r_{(t,K)}\\}$ . Concatenating the visual appearance, spatial information and category embedding between the $i$ -th human and $j$ -th object proposals can represent the visual relation feature $\\mathbf{v}_{(t,k)}$ , as follows:", + "bbox": [ + 496, + 299, + 893, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {(t, k)} = \\left[ W _ {s} v _ {(t, i)}, W _ {o} v _ {(t, j)}, W _ {u} \\varphi \\left(u _ {(t, i j)} \\oplus f _ {b o x} \\left(b _ {(t, i)}, b _ {(t, j)}\\right)\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 508, + 919, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{s}$ , $W_{o}$ and $W_{u}$ represent the parameter matrix of the linear transformation. $[,]$ is concatenation operation, $\\varphi$ is flattening operation and $\\oplus$ is element-wise addition. $u_{(t,ij)}$ the visual feature of the union box of $b_{(t,i)}$ and $b_{(t,j)}$ extracted from the detector. $f_{box}$ maps the 2-channel binary spatial configuration map of bounding boxes $b_{(t,i)}$ and $b_{(t,j)}$ into features of the same dimension as $u_{(t,ij)}$ .", + "bbox": [ + 496, + 541, + 890, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike the visual relation feature, the semantic relation feature $\\mathbf{s}_{(t,k)}$ provides high-level descriptions of the relationship between humans and objects in the videos. The visual relationship categories are either provided as groundtruth or determined by the fine-tuned visual relationship detection network [5]. The features of the semantic relation are obtained by concatenating the three features as follows:", + "bbox": [ + 496, + 646, + 890, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {(t, k)} = \\left[ s _ {(t, i)}, r _ {(t, i j)}, s _ {(t, j)} \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 765, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the $r_{(t,ij)}$ is extracted by embedding the visual relationship category to the semantic feature space. The category embedding vectors $s_{(t,i)}$ and $s_{(t,j)}$ are determined by the categories of human and object, respectively.", + "bbox": [ + 496, + 790, + 890, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the features $\\{\\mathbf{v}_{(t,k)}\\}_{t=1,k=1}^{T,K}$ and the features $\\{\\mathbf{s}_{(t,k)}\\}_{t=1,k=1}^{T,K}$ , we further map the visual relation feature and the semantic relation feature into a joint embedding", + "bbox": [ + 500, + 849, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18880", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6b999cfe46c881921eee6a83a22a16f653604e30c8bd721e3cdbfdc175916575.jpg", + "image_caption": [ + "(a) overview of the proposed method." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 640, + 253 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1d6dc2c76dd0caecf95c4e583ab58afe0f2c03ccb5d77ea7e623a737df79ef71.jpg", + "image_caption": [ + "(b) DT-Former", + "Figure 2. Overview of our LaIAR. The architecture comprises a language model (top) which takes the language description (represented as a spatio-temporal scene graph in [11]) as input and a video model (bottom) which takes the video frames as input. Both models use DT-Former to capture key relational transitions to recognize actions. We transfer knowledge across modalities using a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning), which can help video model benefit from language model during training. For inference, only the video model is considered." + ], + "image_footnote": [], + "bbox": [ + 640, + 98, + 851, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space as follows:", + "bbox": [ + 76, + 364, + 192, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} = f _ {\\mathrm {v}} \\left(\\left\\{\\mathbf {v} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\right), \\quad \\mathcal {S} = f _ {\\mathrm {s}} \\left( \\right.\\left\\{ \\right.\\left\\{\\mathbf {s} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\left. \\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 81, + 387, + 460, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $f_{\\mathrm{V}}$ and $f_{\\mathrm{S}}$ are the visual encoder and the semantic encoder, respectively. In each encoder, each element of the input is first mapped to a local representation via a linear projection. We then apply a Generalized Pooling Operator (GPO) [4] to aggregate the input, creating a global representation. This global representation is combined with each local representation along the channel dimension. This approach helps to use the contextual information from the entire sequence. The visual embedding $\\mathcal{V} \\in \\mathbb{R}^{T \\times K \\times D}$ and semantic embedding $S \\in \\mathbb{R}^{T \\times K \\times D}$ are aligned in the spatiotemporal dimension, where $D$ denotes the dimension in the common space.", + "bbox": [ + 75, + 421, + 468, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2 Dynamic Token Transformers", + "text_level": 1, + "bbox": [ + 76, + 619, + 339, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video understanding shares several high-level similarities with natural language processing (NLP), as they are both fundamentally based on sequential structures [2]. Our intuition is that we can easily model visual and semantic relations simultaneously from joint embedding space (see Sec 3.2.1 for details). Therefore, we introduce a shared dynamic token transformers (DT-Former), which employs the transformer structure to capture key relationship transitions for action reasoning. It mainly consists of the adaptive token selection and the video transformer module. The adaptive token selection module calculates the contribution score of each token to the classification output, and tokens with lower contribution scores will be discarded. The retained tokens, i.e. important relationship representations, are fed to the video transformer module to capture relation transition cues. Figure 2 (b) shows the architecture of DT-Former. Since the video model and the language", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model share the same DT-Former, we denote the input as $X = \\{x_{(t,k)}\\}_{t=1,k=1}^{T,K}$ for simplicity. We add a learnable spatiotemporal positional embedding $e_{(t,k)}^{pos}$ to each vector $x_{(t,k)}$ to obtain the embedding token $x_{(t,k)}^{(0)}$ . The superscript corresponds to the layer of the transformer encoder.", + "bbox": [ + 498, + 364, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Token Selection. Following the ViT approach [6], we concatenate a special learnable vector $(x_{(0,0)}^{(0)} = x_{class})$ representing the embedding of the [class] token in the first position of the sequence. As a large number of relationships between human and objects in a scene are usually redundant, it is essential to reduce these relationships. Inspired by the recent work on token reduction for accelerating transformer inference, we formulate parsing important relations as a token selection problem. To determine whether a token is discarded or retained, we introduce a token selector that consists of an MLP $\\sigma$ and a differentiable discrete-valued estimator using the Gumbel-Softmax (GSM) operator:", + "bbox": [ + 498, + 446, + 892, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {(t, k)} = \\operatorname {G S M} \\left\\{\\sigma \\left(\\left[ W _ {1} x _ {(t, k)} ^ {(0)}, W _ {2} x _ {\\text {c l a s s}} \\right]\\right) \\right\\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 655, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{1}$ and $W_{2}$ represent the linear matrices for dimension compression. We concatenate [class] tokens that represent the global representation with input tokens to exploit contextual information of the entire sequence. The binary output $u_{(t,k)} = 0$ indicates that the $t$ -th token of frame $t$ is to be discarded and $u_{(t,k)} = 1$ is to be retained. Token selection can be represented as: $y_{(t,k)}^{(0)} = u_{(t,k)}x_{(t,k)}^{(0)}$ . Note that this operation is differentiable, allowing for end-to-end training tailored for token selection.", + "bbox": [ + 496, + 685, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure consistent token reduction across consecutive frames, we apply token selector to both the temporal and spatial dimensions. Recent efforts in the field of frame sampling [41] indicate inherent temporal redundancy in frames. Inspired by this, we first focus on salient", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18881", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "frames over the entire time horizon, and then delve into those frames to find key relationships. For the input tokens $X = \\{x_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}$ , we first apply an average-pooling operation to tokens in the spatial dimension to get a sequence of temporal-based tokens $\\{x_{(t)}^{(0)}\\}_{t=1}^{T}$ , and then feed it to the token selector to generate temporal-based selection signal matrix $\\{\\dot{u}_{(t)}\\}_{t=1}^{T}$ . Finally, we repeat it along the spatial dimension to obtain $\\dot{U} = \\{\\dot{u}_{(t,k)}\\}_{t=1,k=1}^{T,K}$ for downstream processing. Similarly, we perform the token selector on each frame separately to generate a selection signal matrix. The spatial-based selection signal matrix can be expressed as $\\dot{U} = \\{\\acute{u}_{(t,k)}\\}_{t=1,k=1}^{T,K}$ . Further, the final selection signal matrix can be expressed as $U = \\dot{U} \\cdot \\dot{U} = \\{u_{(t,k)}\\}_{t=1,k=1}^{T,K}$ . We use matrix multiplication for token selection: $Y = U \\cdot X = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}$ .", + "bbox": [ + 76, + 90, + 472, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer Encoder. In order to model the relationship transition in videos, the token $Y = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}$ are fed to stack of transformer blocks which compute the spatial and temporal self-attention jointly. We convert $Y$ into a set of sequences $\\mathbf{Y}^{(0)} = \\{y_{(p)}^{(0)}\\}_{p=1}^{T \\times K}$ , which are then fed into the transformer encoder to extract a video-level representation:", + "bbox": [ + 76, + 335, + 472, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {\\prime \\ell} = \\operatorname {M S A} \\left(\\operatorname {L N} \\left(\\mathbf {Y} ^ {\\ell - 1}\\right)\\right) + \\mathbf {Y} ^ {\\ell - 1} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 454, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {\\ell} = \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {Y} ^ {\\ell}\\right)\\right) + \\mathbf {Y} ^ {\\prime^ {\\ell}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 477, + 468, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\operatorname {L N} \\left(\\mathbf {Y} _ {(0)} ^ {L}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 500, + 468, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where MSA() and LN() denotes multiheaded self-attention and LayerNorm [1], respectively. $L$ represents the number of transformer blocks. $\\mathbf{z}$ denotes the video-level representation, which can be used to predict the final action classes. Based on the above steps, we can get the selection signal matrix $U^{v} = \\{u_{(t,k)}^{v}\\}_{t=1,k=1}^{T,K}$ , video-level representation $\\mathbf{z}^{\\mathrm{v}}$ of the video model, the selection signal matrix $U^{s} = \\{u_{(t,k)}^{s}\\}_{t=1,k=1}^{T,K}$ and video-level representation $\\mathbf{z}^{\\mathrm{s}}$ of the language model.", + "bbox": [ + 76, + 529, + 472, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1.3 Classification Head", + "text_level": 1, + "bbox": [ + 76, + 685, + 266, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The two classification heads of video and language model predict, $P^v = \\{p_{(c)}^v\\}_{c=1}^C$ and $P^s = \\{p_{(c)}^s\\}_{c=1}^C$ respectively for each branch, where $C$ is the number of classes. We minimize the cross-entropy losses between action scores $P^v$ , and $P^s$ and the ground-truth action labels for each action category, denoted as $\\mathcal{L}_v$ and $\\mathcal{L}_s$ , respectively. The overall loss of two branches can be written as:", + "bbox": [ + 75, + 709, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} = \\mathcal {L} _ {v} + \\mathcal {L} _ {s} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 830, + 468, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The two classification heads only utilize the private information of each modality, in order to allow the video model to benefit from the knowledge of the language model,", + "bbox": [ + 76, + 854, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we propose an additional classification head to estimate the other modality's output: the video model estimates the language model $(P^{v2s} = \\{p_{(c)}^{v2s}\\}_{c=1}^{C})$ . By mimicking not only the class with maximum probability, but also the whole distribution, more information is exchanged, leading to softer labels, which is more beneficial for training our model.", + "bbox": [ + 496, + 90, + 892, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Learning Scheme", + "text_level": 1, + "bbox": [ + 500, + 189, + 671, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The goal of our learning scheme is to transfer information across modalities in a controlled manner thus allowing the video model to learn from the language model. This auxiliary objective can effectively improve the performance of the video modality and does not require additional labels from the datasets. Here, we define the visual-semantic joint embedding learning, our token selection supervision loss and an additional cross-modal learning method.", + "bbox": [ + 496, + 212, + 893, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1 Visual-Semantic Joint Embedding Space", + "text_level": 1, + "bbox": [ + 498, + 351, + 836, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach begins by aligning the visual and semantic relation representations within a shared vector space. In this configuration, each visual embedding $\\hat{\\mathbf{v}}_{(t,k)}\\in \\mathcal{V}$ and $\\hat{\\mathbf{s}}_{(t,k)}\\in S$ pair converge to proximate points. This visual-semantic joint embedding has two main advantages: 1) it helps the video model to improve its generalization since semantic representations are invariant to complex appearance variations. 2) it enables the video model to explicitly represent the relationship transition process, since the visual-semantic joint embedding space can provide semantic labels for each visual relation representation. In this paper, we introduce the contrastive learning of visual-semantic joint embedding and discuss its implementation as following.", + "bbox": [ + 496, + 375, + 892, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a mini-batch $B = \\{(\\hat{\\mathbf{v}}_{(0,0)},\\hat{\\mathbf{s}}_{(0,0)}),\\ldots \\}$ of visual-semantic relationship representation pairs, the contrastive learning objective encourages embeddings of positive pairs $(\\hat{\\mathbf{v}}_{(t,k)},\\hat{\\mathbf{s}}_{(t,k)})$ to align with each other, while pushing embeddings of the negative pairs apart. Formally, the contrastive loss $\\mathcal{L}_{sim}$ is formulated using the symmetric contrastive loss, as follows:", + "bbox": [ + 498, + 570, + 893, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i m} = - \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {v i s u a l} \\rightarrow \\text {s e m a n t i c}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {s e m a n t i c} \\rightarrow \\text {v i s u a l}}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 683, + 906, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i = \\frac{\\mathbf{v}_{(t,k)}}{\\|\\mathbf{v}_{(t,k)}\\|_2}$ and $\\mathbf{y}_i = \\frac{\\mathbf{s}_{(t,k)}}{\\|\\mathbf{s}_{(t,k)}\\|_2}$ . $|\\mathcal{B}|$ is size of the mini-batch $B$ . $t$ is the learnable temperature parameter.", + "bbox": [ + 498, + 763, + 890, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Token Selection Supervision", + "text_level": 1, + "bbox": [ + 500, + 816, + 750, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We aim that the key relations obtained by the language model can guide the video model to perform key relations mining. Therefore, we align the token selection signal matrix of the two models. We minimize the mean-squared loss", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18882", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between the token selection signal matrix $u_{(t,k)}^v$ of the video model and the token selection signal matrix $u_{(t,k)}^s$ of the language model. To maintain the token selection signal matrix's sparsity, we apply L1 norm to allow $u_{(t,k)}^s$ to have a small number of non-zero values. The token selection supervision loss is defined as:", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t s s} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\left\\| u _ {(t, k)} ^ {v} - u _ {(t, k)} ^ {s} \\right\\| + \\left\\| u _ {(t, k)} ^ {s} \\right\\| _ {1}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 189, + 468, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.3 Cross-Modal Learning", + "text_level": 1, + "bbox": [ + 76, + 261, + 290, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given that the visual modality is highly sensitive and the semantic modality more robust to the domain shift, the robust semantic modality can guide the sensitive visual modality to the correct classification. We allow the video model estimate the entire distribution of the language model's prediction. Through cross-modal learning, we aim to transfer knowledge from the language model to the video model. We choose KL divergence for the cross-modal loss $\\mathcal{L}_{\\mathrm{xm}}$ and define it as follows:", + "bbox": [ + 75, + 285, + 472, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {x m}} = D _ {\\mathrm {K L}} \\left(P ^ {s} \\| P ^ {v 2 s}\\right) = - \\sum_ {c = 1} ^ {C} p _ {(c)} ^ {s} \\log \\frac {p _ {(c)} ^ {s}}{p _ {(c)} ^ {v 2 s}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 428, + 468, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 477, + 290, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.1 Training", + "text_level": 1, + "bbox": [ + 76, + 500, + 192, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the training process, we adopt a random sampling strategy to sample fixed $T$ frames for each video. In order to obtain the best possible performance, our framework jointly trains the classification objective and the learning scheme in an end-to-end manner. The final loss is:", + "bbox": [ + 75, + 523, + 468, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\delta \\mathcal {L} _ {s i m} + \\zeta \\mathcal {L} _ {t s s} + \\eta \\mathcal {L} _ {x m} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 608, + 468, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\delta, \\zeta$ and $\\eta$ are hyperparameters that control the importance of the learning scheme. We use $\\delta = 0.1$ , $\\zeta = 1$ and $\\eta = 0.1$ in our experiments.", + "bbox": [ + 75, + 632, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2 Inference", + "text_level": 1, + "bbox": [ + 76, + 695, + 197, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the inference process, a uniform sampling strategy is applied to sample fixed $T$ frames for each video. Only the video model is considered for inference. The explanation of the reasoning process can be explicitly shown by the proximity of the visual representation to the semantic representation in the joint embedding space.", + "bbox": [ + 75, + 719, + 468, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 821, + 209, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 76, + 847, + 274, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments on two extensive video datasets, detailed as follows: (1) Charades [28].", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dd6bb759beed83d255d3b85eddf50a47dda8e3276e90f445e5842d5b33f88e93.jpg", + "table_caption": [ + "Table 1. Ablation study on the Charades and CAD-120 datasets using each proposed module. \"S\" denotes the spatial token selection and \"T\" denotes the temporal token selection." + ], + "table_footnote": [], + "table_body": "
MethodsCharadesCAD-120
mAP (%)↑Num ↓mAR ↑Num ↓
DT-Former w/o S,T61.436.60.7349.8
DT-Former w/o S61.231.10.7217.4
DT-Former w/o T61.230.70.7416.7
DT-Former61.126.00.7514.2
", + "bbox": [ + 504, + 132, + 885, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It contains 157 action classes and consists of about $9.8\\mathrm{k}$ untrimmed videos, among which $7.9\\mathrm{k}$ are used for training and $1.8\\mathrm{k}$ for testing. Each video contains an average of 6.8 distinct action categories and multiple actions can happen at the same time, which makes the recognition extremely challenging. The Action Genome dataset [11] provides fine-grained annotations for the Charades dataset, which provides frame-level relation annotations for videos. Overall, it annotates 476K object bounding boxes and 1.72M relations. (2) CAD-120. Introduced by [15], the CAD-120 dataset is an RGB-D dataset designed for activity understanding. It contains 551 video clips of 4 subjects performing 10 different activities in different environments, such as a kitchen, a living room, and office, etc. To train on our method, we leverage the re-annotated version provided by [42], which provides detailed relationships and attributes for the video frames.", + "bbox": [ + 496, + 253, + 890, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. Following the experimental protocol of [11], We measure multi-label action recognition performance in term of the Mean Average Precision (mAP) on Charades dataset. For CAD-120 dataset, we calculate the Mean Average Recall (mAR) to evaluate whether the model successfully recognizes the performed actions.", + "bbox": [ + 496, + 510, + 890, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 611, + 663, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Effectiveness of Each Module", + "text_level": 1, + "bbox": [ + 500, + 635, + 758, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 reports the effectiveness of each module of the proposed architecture. We evaluate the performance of the DT-Former using different settings, i.e., canceling spatial token selection or canceling temporal token selection or both. Here, the DT-Former corresponds to the performance obtained by the video model. The metric scheme 'Num' refers to the average number of tokens retained per video after token selection. On the CAD-120 dataset, we noted a modest improvement in the accuracy of our architecture, attributable to the spatial-temporal token selection. The CAD-120 dataset typically features videos with a single action, usually characterized by a pair of relational transitions. By eliminating irrelevant features, the model's risk of overfitting is reduced, thereby enhancing its ability to generalize. As expected, more tokens are discarded in the CAD-120 dataset than in the complex Charades dataset. Overall,", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adding either or both of the token selection modules can reduce the number of tokens without significantly affecting that the recognition performance. This demonstrates our architecture's capability in identifying key relational transitions for action recognition and disregarding superfluous/redundant relations.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d464b0c32b2200f353dbaea924a2da810797c0ca0fa992c7ce305d607948a685.jpg", + "table_caption": [ + "Table 2. Ablation study of learning scheme on the Charades and CAD-120 datasets. $\\checkmark$ indicates that the component is applied in the experiments." + ], + "table_footnote": [], + "table_body": "
w/ \nLsimw/ \nLtssw/ \nLxmmAP on \nCharades (%) ↑mAR on \nCAD-120 ↑
61.10.75
✓--62.20.79
-✓-61.70.79
--✓62.90.81
✓-✓63.40.83
✓✓✓63.60.85
", + "bbox": [ + 117, + 236, + 421, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Effectiveness of Learning Scheme", + "text_level": 1, + "bbox": [ + 76, + 375, + 364, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore the effectiveness of the visual-semantic joint embedding $(\\mathcal{L}_{sim})$ , token selection supervision $(\\mathcal{L}_{tss})$ and cross-modal learning $(\\mathcal{L}_{xm})$ in our learning scheme, we conduct related ablation studies using different settings, i.e., cancelling one or any two or all our key modules. As reported in Table 2, the visual-semantic joint embedding, token selection supervision and cross-modal learning components collectively or individually contribute to the final performance improvement. This improvement confirms our hypothesis that language models can help video models to improve performance through knowledge transfer. It can be noted that due to the advantages of the learning scheme, the visual model improves from 61.1 to 63.6 in terms of mAP metric and from 0.75 to 0.85 in terms of mAR metric on Charade and CAD-120, respectively. These results demonstrate that the learning scheme plays an important role in our proposed LaIAR.", + "bbox": [ + 75, + 400, + 470, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Effectiveness of Robustness Against Domain Shift", + "text_level": 1, + "bbox": [ + 76, + 679, + 468, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of RGB-based methods drops drastically when the training and testing data do not share the same distribution caused by change of scene, camera viewpoint or actor [40]. Our proposed model can adapt to video distribution shifts with the aid of semantic modality, which are invariant to complex appearance variations. To demonstrate the robustness of our proposed framework to domain shift, we split the Charades dataset into five subsets with non-overlapping training scenes and test scenes. Table 3 reports the average and variance of five accuracies for these five subsets. The variance of our method is clearly stable and indicates robustness to domain shift.", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ffe2d020e4bf2a7b34338d21ee4c11d9ae44f430ec33289007f036f6f7ab82be.jpg", + "table_caption": [ + "Table 3. Ablation study of the domain shift on the Charades dataset." + ], + "table_footnote": [], + "table_body": "
MethodAccuracy
AverageVariance
STIGPN [34]54.10.30
Ours57.20.11
", + "bbox": [ + 504, + 119, + 715, + 170 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/530293305811a44be96d3512bd0350fb155a9c9fea23690e9de6df105d217661.jpg", + "table_caption": [ + "Table 4. Comparison of accuracy using predicted and annotated relationships." + ], + "table_footnote": [], + "table_body": "
Evaluation ModemAP
Prediction62.4
Label63.6
", + "bbox": [ + 722, + 131, + 874, + 171 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.4 Effectiveness of Using Predicted Relationships.", + "text_level": 1, + "bbox": [ + 500, + 195, + 877, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As previously stated, visual relationship categories can be manually annotated or identified by the visual relationship detection network [5]. To explore the impact of the two modes on accuracy, we compared the proposed method based on the ground truth and the predicted semantic relationships during training. The results, as detailed in Table 4, reveal that using predicted semantic relations can indeed enhance the accuracy of the video model on the Charades dataset (achieving $62.4\\%$ mean Average Precision (mAP) versus $61.1\\%$ mAP shown in Table 2). Moreover, the accuracy does not significantly decrease when using predicted data instead of ground truth. This demonstrates the effectiveness of the proposed method in mining relational transformations from real video data.", + "bbox": [ + 496, + 219, + 890, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to the State-of-the-Art", + "text_level": 1, + "bbox": [ + 500, + 439, + 805, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the action recognition accuracy of the proposed method and the state-of-the-art methods (SoTA) on the Charades and the CAD-120 datasets, respectively. Table 5 summarizes the results on Charades. It can be seen that our proposed method outperforms several previous 3D CNN approaches in terms of mAP, including I3D [3], SlowFast [7] and LFB [36]. This demonstrates that our method can fully capture action cues through the visual relationship transitions, based on the human/objects information detected from a single video frame (rather than using the entire scene like I3D). STRG [35] and SGFB [11] model the action based objects and visual relationships, respectively, and overlook explicit modeling of temporal dynamics of the interaction between objects. Though VideoLN [12] and OR2G [24] takes visual relationship transitions into account, it is difficult for these methods to achieve accurate action inference due to the limitations of scene graph predictors at test time. For comparison in a modality with only RGB video frames, our method achieved the best performance compared with the existing methods. We also evaluated our method in Oracle evaluation mode, which leverages the ground-truth of bounding box and relationships on a frame. As reported in the last row of Table 5, our method still achieves best performance on the Charades dataset. It is important to mention that OR2G [24] used ground-truth scene graphs to enhance its final performance, whereas our network uses only the bounding boxes of humans and objects. Despite this, our method demonstrates strong performance in both evaluation modes, validating the effective", + "bbox": [ + 496, + 463, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18884", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness of our proposed approach.", + "bbox": [ + 76, + 90, + 284, + 106 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5d203cc1335cb3522748942329bf6b88902e6e72e40f470c686715546faeb39c.jpg", + "table_caption": [ + "Table 5. Multi-label action recognition performance comparison on the Charades's validation set in term of mAP. SG: ground truth scene graph. Bbox: Bounding Box. Higher values are better." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneModalitymAP
I3D [3]R101-I3DRGB15.6
VideoMLN [12]R101RGB38.4
STRG [35]R101-I3D-NLRGB39.7
SlowFast [7]R101RGB42.1
LFB [36]R101-I3D-NLRGB42.5
SGFB [11]R101-I3D-NLRGB44.3
OR2G [24]R101-I3D-NLRGB44.9
OursR101-I3D-NLRGB45.1
SGFB Oracle [11]R101-I3D-NLRGB+SG60.3
VideoMLN Oracle [12]R101RGB+SG62.8
OR2G Oracle [24]R101RGB+SG63.3
Ours OracleR101RGB+Bbox63.6
", + "bbox": [ + 80, + 162, + 460, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For CAD-120 dataset, we follow the same experimental protocol as in [42] and divide the long video sequences into small segments based on individual sub-actions and evaluate the average recall metric for each sub-action. Table 6 summarizes the results on CAD-120. Explainable AAR-RAR [42] interpret the action reasoning process through the changes of relationship between objects or the attribute of objects across time. Our method is able to give the same explanation and outperforms these methods, achieving state-of-the-art performance with $0.85\\mathrm{mAR}$ .", + "bbox": [ + 75, + 375, + 468, + 525 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/18f5a7d4c8563f8b720f9621319ca5a666fb21cd5decd328057e2d241387fff9.jpg", + "table_caption": [ + "Table 6. Experimental results on CAD-120 for action recognition." + ], + "table_footnote": [], + "table_body": "
MethodsModalitymAR
Temporal Segment [33]RGB0.42
Flow0.71
RGB + Flow0.77
Explainable AAR-RAR [42]RGB0.80
VideoMLN [12]RGB0.83
OursRGB0.85
", + "bbox": [ + 96, + 551, + 446, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Interpretation and Visualization", + "text_level": 1, + "bbox": [ + 76, + 709, + 370, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To intuitively demonstrate the interpretability effect of our proposed model, we provide interpretable representation and a visualization example. As shown in Figure 3, in the inference stage, we first extract the visual relation representations of human-object pairs in each frame. Then, our proposed DT-Former selects important relations in temporal and spatial dimensions and predicts action by modeling the important relation transition. Finally, the visual representations of important relations are mapped into the joint embedding space to find their nearest neighbor semantic labels, which can provide explicit evidence for the action", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reasoning process. In this example, the relation representations between the person and the box in the second and tenth frames are selected as cues for the action recognition. The nearest semantic labels of these two representations in the joint embedding space are \"holding box\" and \"not holding box\", respectively. Here, the consequences of observations 'holding' $\\rightarrow$ 'not holding' provide a clear sign of the action \"place\".", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb328bafcbf76717fd84a6636a58c004ace07a0b747546f3b9675369602b7074.jpg", + "image_caption": [ + "Figure 3. An example of action recognition performed by the proposed method and its corresponding process of providing explanations. The shaded visual relation representations indicate that it is not selected by DT-Former." + ], + "image_footnote": [], + "bbox": [ + 500, + 213, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 489, + 617, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a new framework, LaIAR, designed to transfer the knowledge from the language model to the video model to improve the recognition performance and interpretability of video models. Specifically, we build a language model and a video model, which take semantic relation and visual relation representations as input, respectively. These two models share the same architecture, namely DT-Former. This architecture is tailored to select the most important relations for action recognition from all the relations in video and to model the fine-grained relation transitions within videos. Our framework also incorporates three novel knowledge transfer strategies in our learning scheme to facilitate the knowledge transfer from the language model to the video model. This not only boosts the performance but also enhances the interpretability of the video model. Ablation experiments verified the effectiveness of the DT-Former, the learning scheme module and the robustness against domain shift. We conducted extensive experiments on Charades and CAD-120 datasets to demonstrate the superior performance of our proposed method.", + "bbox": [ + 496, + 515, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements: This work is partially supported by the National Natural Science Foundation of China under Grant No.62073252 and No.62072358. It was also supported by Natural Science Basic Research Program of Shaanxi under Grant No.2024JC-JCQN-66.", + "bbox": [ + 496, + 816, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 4", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8", + "[4] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4", + "[5] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 3, 7", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2020. 4", + "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7, 8", + "[8] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 2", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[10] Hua Hua, Dongxu Li, Ruiqi Li, Peng Zhang, Jochen Renz, and Anthony Cohn. Towards explainable action recognition by salient qualitative spatial object relation chains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5710-5718, 2022. 2", + "[11] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 3, 4, 6, 7, 8", + "[12] Yang Jin, Linchao Zhu, and Yadong Mu. Complex video action reasoning via learnable markov logic network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3242-3251, 2022. 2, 7, 8", + "[13] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In European Conference on Computer Vision, pages 370-387. Springer, 2022. 1" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 784-794, 2022. 3", + "[15] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 6", + "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3", + "[17] Christopher A Kurby and Jeffrey M Zacks. Segmentation in the perception and memory of events. Trends in cognitive sciences, 12(2):72-79, 2008. 2", + "[18] Pilhyeon Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, and Hyeran Byun. Decomposed cross-modal distillation for rgb-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2373-2383, 2023. 1, 3", + "[19] Yi Li and Nuno Vasconcelos. Improving video model transfer with dynamic representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19280-19291, 2022. 1", + "[20] Xue Lin, Qi Zou, Xixia Xu, Yaping Huang, and Ding Ding. Effects of motion-relevant knowledge from unlabeled video to human-object interaction detection. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3", + "[21] Guan Luo, Shuang Yang, Guodong Tian, Chunfeng Yuan, Weiming Hu, and Stephen J Maybank. Learning human actions by combining global dynamics and local appearance. IEEE transactions on pattern analysis and machine intelligence, 36(12):2466-2482, 2014. 1, 2", + "[22] Lili Meng, Bo Zhao, Bo Chang, Gao Huang, Wei Sun, Frederick Tung, and Leonid Sigal. Interpretable spatio-temporal attention for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1, 2", + "[23] Ali Modarressi, Hosein Mohebbi, and Mohammad Taher Pilehvar. Adapler: Speeding up inference by adaptive length reduction. arXiv preprint arXiv:2203.08991, 2022. 3", + "[24] Yangjun Ou, Li Mi, and Zhenzhong Chen. Object-relation reasoning graph for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20133-20142, 2022. 2, 7, 8", + "[25] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 1", + "[26] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[27] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18886", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models instead. Nature machine intelligence, 1(5):206-215, 2019. 1, 2", + "[28] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6", + "[29] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017. 1, 2", + "[30] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence, 2022. 1", + "[31] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1", + "[32] Junke Wang, Xitong Yang, Hengduo Li, Li Liu, Zuxuan Wu, and Yu-Gang Jiang. Efficient video transformers with spatial-temporal token selection. In European Conference on Computer Vision, pages 69-86. Springer, 2022. 3", + "[33] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 8", + "[34] Ning Wang, Guangming Zhu, Hongsheng Li, Mingtao Feng, Xia Zhao, Lan Ni, Peiyi Shen, Lin Mei, and Liang Zhang. Exploring spatio-temporal graph convolution for video-based human-object interaction recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 7", + "[35] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In Proceedings of the European conference on computer vision (ECCV), pages 399-417, 2018. 7, 8", + "[36] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 7, 8", + "[37] Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618, 2021. 2", + "[38] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3", + "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3", + "[40] Yunhua Zhang, Hazel Doughty, Ling Shao, and Cees GM Snoek. Audio-adaptive activity recognition across video domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13791-13800, 2022. 3, 7", + "[41] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 4", + "[42] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Explainable video action reasoning via prior knowledge and state transitions. In Proceedings of the 27th acm international conference on multimedia, pages 521-529, 2019. 2, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 343 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18887", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_model.json b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cfb7a6a02876db4363fc2fa9ac719a4674424f11 --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Language Model Guided Interpretable Video Action Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.18, + 0.916, + 0.217 + ], + "angle": 0, + "content": "Ning Wang\\(^{1*}\\), Guangming Zhu\\(^{1\\dagger,*}\\), HS Li\\(^{1}\\), Liang Zhang\\(^{1\\dagger}\\), Syed Afaq Ali Shah\\(^{2}\\), Mohammed Bennamoun\\(^{3}\\), Xidian University, \\(^{2}\\)Edith Cowan University, \\(^{3}\\)University of Western Australia" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.219, + 0.815, + 0.235 + ], + "angle": 0, + "content": "{ningwang,hsli}@stu.xidian.edu.cn，{gmzhu,liangzhang}@xidian.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.238, + 0.726, + 0.251 + ], + "angle": 0, + "content": "afaq.shah@ecu.edu.au, mohammed.bennamoun@uwa.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.474, + 0.653 + ], + "angle": 0, + "content": "While neural networks have excelled in video action recognition tasks, their \"black-box\" nature often obscures the understanding of their decision-making processes. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. These models, however, usually fall short in performance compared to their \"black-box\" counterparts. In this work, we present a new framework named Language-guided Interpretable Action Recognition framework (LaIAR). LaIAR leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we redefine the problem of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on two complex video action datasets, Charades & CAD-120, validates the improved performance and interpretability of our LaIAR framework. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.668, + 0.21, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Building on the advancements of deep learning in image recognition [9, 16, 31], neural network (NN) models have become the leading approach for video-related challenges, including action recognition [18, 25, 30]. Yet, many of the top-tier action recognition techniques [19, 34] deploy NNs in an opaque, black-box fashion. This lack of transparency does not offer clear justification for their decisions, hindering their utility in various real-world contexts [13], especially those with rigorous security demands. These considerations drive us to develop an action reasoning system that pairs exceptional performance with clear interpretability." + }, + { + "type": "title", + "bbox": [ + 0.507, + 0.286, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Action : “Someone is standing up from somewhere” Time" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.301, + 0.667, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.301, + 0.785, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.302, + 0.888, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.368, + 0.886, + 0.381 + ], + "angle": 0, + "content": "(a) An illustration example of relationship transition in the videos." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.382, + 0.666, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.382, + 0.805, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.394, + 0.875, + 0.423 + ], + "angle": 0, + "content": "language model" + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.435, + 0.885, + 0.449 + ], + "angle": 0, + "content": "first stage second stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.505, + 0.451, + 0.83, + 0.465 + ], + "angle": 0, + "content": "(b) Traditional two-stage interpretable reasoning model." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.466, + 0.607, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.607, + 0.466, + 0.692, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.466, + 0.805, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.466, + 0.886, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.54, + 0.604, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.54, + 0.692, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.54, + 0.805, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.54, + 0.885, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.605, + 0.88, + 0.62 + ], + "angle": 0, + "content": "(c) Language model guided video interpretable reasoning model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.631, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Figure 1. (a) An example of action that can be decomposed into relationship transitions (i.e., when the transition is 'sitting on' \\(\\rightarrow\\) 'not contacting' between \\(<\\text{person, bed}>\\) pair, it represents the action \"Someone is standing up from somewhere\"). (b) Traditional two-stage methods usually predict the scene graph first, and then use language models to capture the semantic-level relationship transitions. (c) Our method exploits a language model to guide the video model to capture the relationship transition during training. During inference, our method processes videos and directly recognizes actions, providing supportive evidence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Most of the current interpretable action recognition techniques [21, 22, 29] aim to elucidate the decision-making process of NNs using post-hoc explanations, with a particular emphasis on gradient-based and perturbation-based approaches. However, despite notable advancements, these explanations can be problematic because they might not be faithful to what the network computes, as highlighted by [27]. A compelling direction in interpretability re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.381, + 0.888 + ], + "angle": 0, + "content": "* Ning Wang and Guangming Zhu are co-first authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Liang Zhang and Guangming Zhu are both the corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18878" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.591 + ], + "angle": 0, + "content": "volves around the concept of built-in explanation models [10, 12, 24, 42]. The essence of these models is their inherent interpretability right from the design stage. Recent strategies decompose a complex action into temporal transitions of human-object relationships, drawing inspiration from the event segmentation theory [17]. An illustration in Figure 1 (a) depicts that for a relationship involving a \\(<\\) person, \\(\\mathrm{bed}>\\) pairing, a transition sequence from 'sitting on' to 'not contacting' signifies the action of \"Someone is standing up from somewhere\". This methodology facilitates action recognition by pinpointing semantic transitions through language models, offering a granular insight into action execution. As shown in Figure 1 (b), Jin and Ou et al. [12, 24] extract spatio-temporal scene graphs from video content and apply Markov Logic Network (MLN) based probabilistic logical inference and relation reasoning graphs to create an interpretable representation for a variety of complex actions, respectively. However, it is believed that such models will perform worse than their black-box alternatives [8]. Moreover, these methods divide the process into two stages, namely scene graph prediction and relation modeling. Optimizing these components separately might lead to sub-optimal results. In this paper, we propose to harness the explicit logical inference rules of an interpretable language model to guide the learning process of a video black-box model. Interpretability and strong performance can be attained by focusing solely on the video model during the inference stage. To achieve this, two main challenges arise: 1) Designing a language model that can automatically grasp logical reasoning patterns, sidestepping manual rule creation. 2) Developing a decoupled language-video model architecture that enables the language model to guide the video model's training process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.6, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we have developed a new framework called Language-guided Interpretable Action Recognition framework named LaIAR. As depicted in Figure 1 (c), LaIAR constructs an action recognition model that both implicitly and explicitly exploits fine-grained knowledge of relationship transitions from an interpretable built-in language model. Specifically, we use dynamic token transformers (DT-Former) to both video and language inputs, selectively focusing on important relationships in a data-driven manner, and disregarding non-contributory ones. We aim to redefine the traditional decision interpretation challenge of video models towards a visual-language relation alignment problem. The relationship prioritization determined by the language model then explicitly guides the video model in identifying the most relevant relationships. We propose a learning strategy to facilitate knowledge transfer between language and video to improve the performance of the video model. A key feature of LaIAR is its modular design: during the inference phase, only RGB data serves as input to predict actions, providing" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.744, + 0.107 + ], + "angle": 0, + "content": "a direct and transparent justification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.895, + 0.274 + ], + "angle": 0, + "content": "To summarize, our contributions are three-fold: 1) We propose a novel LaIAR framework that can automatically mine fine-grained relation transitions from data and create interpretable representations for various complex actions. 2) We design a decoupled cross-modal knowledge transfer architecture that leverages useful knowledge from language models to improve the performance and interpretability of the video model at training time, and achieves high-performance interpretable reasoning for videos at test time. 3) Our method achieves state-of-the-art results on two large-scale action recognition benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.289, + 0.642, + 0.305 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.839, + 0.331 + ], + "angle": 0, + "content": "2.1. Interpretable Video Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.895, + 0.807 + ], + "angle": 0, + "content": "Interpretable video action recognition methods can be categorized into two types: post-hoc method and built-in method. post-hoc techniques generate explanations for the network's decision-making process after the network is trained. [22] introduced an interpretable and easy plug-in spatial-temporal attention mechanism for video action recognition to improve the interpretability of the model for video action recognition. [29] developed an interpretable temporal convolutional network to explain the decision-making process of action recognition through each of the learned filters in a Res-TCN. [21] combined both global dynamics and local details to learn human action, using gradient-weighted class activation mapping (Grad-CAM) to visualize the model's attention to action-critical regions. Although these methods have the advantage of not imposing any model constraints, they may be incomplete or unfaithful to the model's reasoning [27]. In contrast, built-in methods restrict the interpretation to be consistent with the model's inferences. [42] approached action reasoning by modeling semantic-level state transitions between two consecutive frames as defined by domain experts. In [10], a method is proposed to achieve interpretability of action recognition by incorporating qualitative spatial reasoning and extracting salient relation chains. Some recent methods, like [12, 24] decompose complex action into continuous relationship transitions according to the event segmentation theory [17]. These methods model the relationship transitions at the semantic level to recognize actions. In this paper, we propose to construct a high performance interpretable-by-design action classifier by guiding a video model with an interpretable language model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.816, + 0.833 + ], + "angle": 0, + "content": "2.2. Adaptive Inference in Transformers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As their popularity soars, adaptive inference for language and vision transformers has caught the attention of researchers. In [37], an adaptive language transformer is proposed to achieve fixed-scale reduction of the input se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18879" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.348 + ], + "angle": 0, + "content": "quence to improve inference speed by dynamically selecting important tokens and removing the irrelevant ones. In [23], a threshold mechanism is introduced to determine the importance of each token and dynamically select the tokens according to the importance of the input sequence. [14] used the mean of attention matrix column values of the transformers to determine the importance score of each token, facilitating token pruning. [38] developed an adaptive token generation mechanism to determine the required number and size of tokens, thereby reducing the computational and memory overhead of the model on images. [32] devised a token selection framework to dynamically select important tokens across the temporal and spatial dimensions of the video input. In this paper, we propose a lightweight token selection method based on Gumbel-Softmax and apply it to our cross-modal transformers for spatio-temporal token selection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.357, + 0.406, + 0.373 + ], + "angle": 0, + "content": "2.3. Cross-modal Knowledge Transferring" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.683 + ], + "angle": 0, + "content": "The past few years have seen an increasing interest in cross-modal knowledge transfer techniques for detection and segmentation tasks. [20] introduced a method to transfer the motion-related knowledge of unlabeled videos to Human-Object Interactions (HOI) detection to infer rare or unseen HOIs. In [40], reliable domain-invariant sound cues are exploited to help video activity recognition models adapt to video distribution shifts. Lately, knowledge distillation techniques have been extended to transfer knowledge across different modalities. For instance, [18] proposed a decomposed cross-modal distillation framework to improve RGB-based temporal action detection by transferring knowledge from the optical flow modality. Similarly, [39] proposed a modified knowledge distillation method that boosts the performance of single-modal 3D captioning by transferring color and texture-aware information from 2D images into 3D object representations. In contrast to these methods, we propose a well-designed knowledge-guided framework to enable cross-modal learning by decoupling information transfer in video and language." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.695, + 0.307, + 0.712 + ], + "angle": 0, + "content": "3. The Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Our proposed method is designed to exploit multimodality by enabling information transfer from language descriptions to videos. This enables the video model to effectively learn from the language model effectively. This is achieved by the video model mimicking the output of the language model, thereby leveraging the intrinsic capabilities of the language model. Specifically, the video frames and the language description (represented as a spatio-temporal scene graph in [11]) are first processed by the encoding network to extract the paired visual and semantic relationship representations. Then, these paired visual and semantic relationship representations are sepa" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "rately fed to DT-Former module, which models the key relationship transition for action recognition. Finally, we propose a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning) to improve the performance and interpretability of the video model by facilitating the knowledge transfer from the language model to the video model. An overview of our proposed method is shown in Figure 2 (a). Note that, in our proposed approach, only the video model is employed for inference once training is complete." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.635, + 0.267 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.276, + 0.761, + 0.291 + ], + "angle": 0, + "content": "3.1.1 Video and Language Encoder" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.3, + 0.895, + 0.499 + ], + "angle": 0, + "content": "Given a video consisting of \\(T\\) frames with \\(N\\) entities of either human or object classes, we use Faster R-CNN [26] with ResNet-101 [16] backbone to detect these entities and extract their features from the video. For the frame \\(I_{t}\\) at time step \\(t\\), the visual features \\(\\{v_{(t,1)}, v_{(t,2)}, \\dots, v_{(t,N)}\\}\\), bounding boxes \\(\\{b_{(t,1)}, b_{(t,2)}, \\dots, b_{(t,N)}\\}\\) and object category \\(\\{c_{(t,1)}, c_{(t,2)}, \\dots, c_{(t,N)}\\}\\) of the objects proposals are supplied by the detector. Between each \\(<\\text{human, object}>\\) pair in the frame, there is a set of relationships \\(R_{t} = \\{r_{(t,1)}, r_{(t,2)}, \\dots, r_{(t,K)}\\}\\). Concatenating the visual appearance, spatial information and category embedding between the \\(i\\)-th human and \\(j\\)-th object proposals can represent the visual relation feature \\(\\mathbf{v}_{(t,k)}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.509, + 0.92, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {(t, k)} = \\left[ W _ {s} v _ {(t, i)}, W _ {o} v _ {(t, j)}, W _ {u} \\varphi \\left(u _ {(t, i j)} \\oplus f _ {b o x} \\left(b _ {(t, i)}, b _ {(t, j)}\\right)\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.648 + ], + "angle": 0, + "content": "where \\( W_{s} \\), \\( W_{o} \\) and \\( W_{u} \\) represent the parameter matrix of the linear transformation. \\([,]\\) is concatenation operation, \\( \\varphi \\) is flattening operation and \\( \\oplus \\) is element-wise addition. \\( u_{(t,ij)} \\) the visual feature of the union box of \\( b_{(t,i)} \\) and \\( b_{(t,j)} \\) extracted from the detector. \\( f_{box} \\) maps the 2-channel binary spatial configuration map of bounding boxes \\( b_{(t,i)} \\) and \\( b_{(t,j)} \\) into features of the same dimension as \\( u_{(t,ij)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.647, + 0.892, + 0.753 + ], + "angle": 0, + "content": "Unlike the visual relation feature, the semantic relation feature \\(\\mathbf{s}_{(t,k)}\\) provides high-level descriptions of the relationship between humans and objects in the videos. The visual relationship categories are either provided as groundtruth or determined by the fine-tuned visual relationship detection network [5]. The features of the semantic relation are obtained by concatenating the three features as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.766, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {(t, k)} = \\left[ s _ {(t, i)}, r _ {(t, i j)}, s _ {(t, j)} \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.851 + ], + "angle": 0, + "content": "where the \\(r_{(t,ij)}\\) is extracted by embedding the visual relationship category to the semantic feature space. The category embedding vectors \\(s_{(t,i)}\\) and \\(s_{(t,j)}\\) are determined by the categories of human and object, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.851, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Given the features \\(\\{\\mathbf{v}_{(t,k)}\\}_{t=1,k=1}^{T,K}\\) and the features \\(\\{\\mathbf{s}_{(t,k)}\\}_{t=1,k=1}^{T,K}\\), we further map the visual relation feature and the semantic relation feature into a joint embedding" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18880" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.641, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.257, + 0.493, + 0.272 + ], + "angle": 0, + "content": "(a) overview of the proposed method." + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.099, + 0.852, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.257, + 0.794, + 0.271 + ], + "angle": 0, + "content": "(b) DT-Former" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.283, + 0.892, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Overview of our LaIAR. The architecture comprises a language model (top) which takes the language description (represented as a spatio-temporal scene graph in [11]) as input and a video model (bottom) which takes the video frames as input. Both models use DT-Former to capture key relational transitions to recognize actions. We transfer knowledge across modalities using a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning), which can help video model benefit from language model during training. For inference, only the video model is considered." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.366, + 0.193, + 0.38 + ], + "angle": 0, + "content": "space as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.082, + 0.388, + 0.462, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} = f _ {\\mathrm {v}} \\left(\\left\\{\\mathbf {v} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\right), \\quad \\mathcal {S} = f _ {\\mathrm {s}} \\left( \\right.\\left\\{ \\right.\\left\\{\\mathbf {s} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\left. \\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.603 + ], + "angle": 0, + "content": "where the \\(f_{\\mathrm{V}}\\) and \\(f_{\\mathrm{S}}\\) are the visual encoder and the semantic encoder, respectively. In each encoder, each element of the input is first mapped to a local representation via a linear projection. We then apply a Generalized Pooling Operator (GPO) [4] to aggregate the input, creating a global representation. This global representation is combined with each local representation along the channel dimension. This approach helps to use the contextual information from the entire sequence. The visual embedding \\(\\mathcal{V} \\in \\mathbb{R}^{T \\times K \\times D}\\) and semantic embedding \\(S \\in \\mathbb{R}^{T \\times K \\times D}\\) are aligned in the spatiotemporal dimension, where \\(D\\) denotes the dimension in the common space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.34, + 0.636 + ], + "angle": 0, + "content": "3.1.2 Dynamic Token Transformers" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Video understanding shares several high-level similarities with natural language processing (NLP), as they are both fundamentally based on sequential structures [2]. Our intuition is that we can easily model visual and semantic relations simultaneously from joint embedding space (see Sec 3.2.1 for details). Therefore, we introduce a shared dynamic token transformers (DT-Former), which employs the transformer structure to capture key relationship transitions for action reasoning. It mainly consists of the adaptive token selection and the video transformer module. The adaptive token selection module calculates the contribution score of each token to the classification output, and tokens with lower contribution scores will be discarded. The retained tokens, i.e. important relationship representations, are fed to the video transformer module to capture relation transition cues. Figure 2 (b) shows the architecture of DT-Former. Since the video model and the language" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.447 + ], + "angle": 0, + "content": "model share the same DT-Former, we denote the input as \\( X = \\{x_{(t,k)}\\}_{t=1,k=1}^{T,K} \\) for simplicity. We add a learnable spatiotemporal positional embedding \\( e_{(t,k)}^{pos} \\) to each vector \\( x_{(t,k)} \\) to obtain the embedding token \\( x_{(t,k)}^{(0)} \\). The superscript corresponds to the layer of the transformer encoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.893, + 0.646 + ], + "angle": 0, + "content": "Adaptive Token Selection. Following the ViT approach [6], we concatenate a special learnable vector \\((x_{(0,0)}^{(0)} = x_{class})\\) representing the embedding of the [class] token in the first position of the sequence. As a large number of relationships between human and objects in a scene are usually redundant, it is essential to reduce these relationships. Inspired by the recent work on token reduction for accelerating transformer inference, we formulate parsing important relations as a token selection problem. To determine whether a token is discarded or retained, we introduce a token selector that consists of an MLP \\(\\sigma\\) and a differentiable discrete-valued estimator using the Gumbel-Softmax (GSM) operator:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.656, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\nu _ {(t, k)} = \\operatorname {G S M} \\left\\{\\sigma \\left(\\left[ W _ {1} x _ {(t, k)} ^ {(0)}, W _ {2} x _ {\\text {c l a s s}} \\right]\\right) \\right\\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.825 + ], + "angle": 0, + "content": "where \\( W_{1} \\) and \\( W_{2} \\) represent the linear matrices for dimension compression. We concatenate [class] tokens that represent the global representation with input tokens to exploit contextual information of the entire sequence. The binary output \\( u_{(t,k)} = 0 \\) indicates that the \\( t \\)-th token of frame \\( t \\) is to be discarded and \\( u_{(t,k)} = 1 \\) is to be retained. Token selection can be represented as: \\( y_{(t,k)}^{(0)} = u_{(t,k)}x_{(t,k)}^{(0)} \\). Note that this operation is differentiable, allowing for end-to-end training tailored for token selection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To ensure consistent token reduction across consecutive frames, we apply token selector to both the temporal and spatial dimensions. Recent efforts in the field of frame sampling [41] indicate inherent temporal redundancy in frames. Inspired by this, we first focus on salient" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.337 + ], + "angle": 0, + "content": "frames over the entire time horizon, and then delve into those frames to find key relationships. For the input tokens \\( X = \\{x_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K} \\), we first apply an average-pooling operation to tokens in the spatial dimension to get a sequence of temporal-based tokens \\( \\{x_{(t)}^{(0)}\\}_{t=1}^{T} \\), and then feed it to the token selector to generate temporal-based selection signal matrix \\( \\{\\dot{u}_{(t)}\\}_{t=1}^{T} \\). Finally, we repeat it along the spatial dimension to obtain \\( \\dot{U} = \\{\\dot{u}_{(t,k)}\\}_{t=1,k=1}^{T,K} \\) for downstream processing. Similarly, we perform the token selector on each frame separately to generate a selection signal matrix. The spatial-based selection signal matrix can be expressed as \\( \\dot{U} = \\{\\acute{u}_{(t,k)}\\}_{t=1,k=1}^{T,K} \\). Further, the final selection signal matrix can be expressed as \\( U = \\dot{U} \\cdot \\dot{U} = \\{u_{(t,k)}\\}_{t=1,k=1}^{T,K} \\). We use matrix multiplication for token selection: \\( Y = U \\cdot X = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.473, + 0.446 + ], + "angle": 0, + "content": "Transformer Encoder. In order to model the relationship transition in videos, the token \\( Y = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K} \\) are fed to stack of transformer blocks which compute the spatial and temporal self-attention jointly. We convert \\( Y \\) into a set of sequences \\( \\mathbf{Y}^{(0)} = \\{y_{(p)}^{(0)}\\}_{p=1}^{T \\times K} \\), which are then fed into the transformer encoder to extract a video-level representation:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.455, + 0.47, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {\\prime \\ell} = \\operatorname {M S A} \\left(\\operatorname {L N} \\left(\\mathbf {Y} ^ {\\ell - 1}\\right)\\right) + \\mathbf {Y} ^ {\\ell - 1} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.478, + 0.47, + 0.498 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {\\ell} = \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {Y} ^ {\\ell}\\right)\\right) + \\mathbf {Y} ^ {\\prime^ {\\ell}} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.501, + 0.47, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\operatorname {L N} \\left(\\mathbf {Y} _ {(0)} ^ {L}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.53, + 0.473, + 0.668 + ], + "angle": 0, + "content": "where MSA() and LN() denotes multiheaded self-attention and LayerNorm [1], respectively. \\(L\\) represents the number of transformer blocks. \\(\\mathbf{z}\\) denotes the video-level representation, which can be used to predict the final action classes. Based on the above steps, we can get the selection signal matrix \\(U^{v} = \\{u_{(t,k)}^{v}\\}_{t=1,k=1}^{T,K}\\), video-level representation \\(\\mathbf{z}^{\\mathrm{v}}\\) of the video model, the selection signal matrix \\(U^{s} = \\{u_{(t,k)}^{s}\\}_{t=1,k=1}^{T,K}\\) and video-level representation \\(\\mathbf{z}^{\\mathrm{s}}\\) of the language model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.686, + 0.267, + 0.7 + ], + "angle": 0, + "content": "3.1.3 Classification Head" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.816 + ], + "angle": 0, + "content": "The two classification heads of video and language model predict, \\( P^v = \\{p_{(c)}^v\\}_{c=1}^C \\) and \\( P^s = \\{p_{(c)}^s\\}_{c=1}^C \\) respectively for each branch, where \\( C \\) is the number of classes. We minimize the cross-entropy losses between action scores \\( P^v \\), and \\( P^s \\) and the ground-truth action labels for each action category, denoted as \\( \\mathcal{L}_v \\) and \\( \\mathcal{L}_s \\), respectively. The overall loss of two branches can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.832, + 0.469, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} = \\mathcal {L} _ {v} + \\mathcal {L} _ {s} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The two classification heads only utilize the private information of each modality, in order to allow the video model to benefit from the knowledge of the language model," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.183 + ], + "angle": 0, + "content": "we propose an additional classification head to estimate the other modality's output: the video model estimates the language model \\((P^{v2s} = \\{p_{(c)}^{v2s}\\}_{c=1}^{C})\\). By mimicking not only the class with maximum probability, but also the whole distribution, more information is exchanged, leading to softer labels, which is more beneficial for training our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.19, + 0.673, + 0.207 + ], + "angle": 0, + "content": "3.2. Learning Scheme" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.895, + 0.336 + ], + "angle": 0, + "content": "The goal of our learning scheme is to transfer information across modalities in a controlled manner thus allowing the video model to learn from the language model. This auxiliary objective can effectively improve the performance of the video modality and does not require additional labels from the datasets. Here, we define the visual-semantic joint embedding learning, our token selection supervision loss and an additional cross-modal learning method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.352, + 0.837, + 0.368 + ], + "angle": 0, + "content": "3.2.1 Visual-Semantic Joint Embedding Space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.893, + 0.571 + ], + "angle": 0, + "content": "Our approach begins by aligning the visual and semantic relation representations within a shared vector space. In this configuration, each visual embedding \\(\\hat{\\mathbf{v}}_{(t,k)}\\in \\mathcal{V}\\) and \\(\\hat{\\mathbf{s}}_{(t,k)}\\in S\\) pair converge to proximate points. This visual-semantic joint embedding has two main advantages: 1) it helps the video model to improve its generalization since semantic representations are invariant to complex appearance variations. 2) it enables the video model to explicitly represent the relationship transition process, since the visual-semantic joint embedding space can provide semantic labels for each visual relation representation. In this paper, we introduce the contrastive learning of visual-semantic joint embedding and discuss its implementation as following." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.571, + 0.894, + 0.678 + ], + "angle": 0, + "content": "Given a mini-batch \\(B = \\{(\\hat{\\mathbf{v}}_{(0,0)},\\hat{\\mathbf{s}}_{(0,0)}),\\ldots \\}\\) of visual-semantic relationship representation pairs, the contrastive learning objective encourages embeddings of positive pairs \\((\\hat{\\mathbf{v}}_{(t,k)},\\hat{\\mathbf{s}}_{(t,k)})\\) to align with each other, while pushing embeddings of the negative pairs apart. Formally, the contrastive loss \\(\\mathcal{L}_{sim}\\) is formulated using the symmetric contrastive loss, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.684, + 0.908, + 0.766 + ], + "angle": 0, + "content": "\\[\nL _ {s i m} = - \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {v i s u a l} \\rightarrow \\text {s e m a n t i c}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {s e m a n t i c} \\rightarrow \\text {v i s u a l}}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.8 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i = \\frac{\\mathbf{v}_{(t,k)}}{\\|\\mathbf{v}_{(t,k)}\\|_2}\\) and \\(\\mathbf{y}_i = \\frac{\\mathbf{s}_{(t,k)}}{\\|\\mathbf{s}_{(t,k)}\\|_2}\\). \\(|\\mathcal{B}|\\) is size of the mini-batch \\(B\\). \\(t\\) is the learnable temperature parameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.75, + 0.832 + ], + "angle": 0, + "content": "3.2.2 Token Selection Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We aim that the key relations obtained by the language model can guide the video model to perform key relations mining. Therefore, we align the token selection signal matrix of the two models. We minimize the mean-squared loss" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "between the token selection signal matrix \\( u_{(t,k)}^v \\) of the video model and the token selection signal matrix \\( u_{(t,k)}^s \\) of the language model. To maintain the token selection signal matrix's sparsity, we apply L1 norm to allow \\( u_{(t,k)}^s \\) to have a small number of non-zero values. The token selection supervision loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.19, + 0.47, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t s s} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\left\\| u _ {(t, k)} ^ {v} - u _ {(t, k)} ^ {s} \\right\\| + \\left\\| u _ {(t, k)} ^ {s} \\right\\| _ {1}\\right) \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.262, + 0.292, + 0.278 + ], + "angle": 0, + "content": "3.2.3 Cross-Modal Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.422 + ], + "angle": 0, + "content": "Given that the visual modality is highly sensitive and the semantic modality more robust to the domain shift, the robust semantic modality can guide the sensitive visual modality to the correct classification. We allow the video model estimate the entire distribution of the language model's prediction. Through cross-modal learning, we aim to transfer knowledge from the language model to the video model. We choose KL divergence for the cross-modal loss \\(\\mathcal{L}_{\\mathrm{xm}}\\) and define it as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.429, + 0.47, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {x m}} = D _ {\\mathrm {K L}} \\left(P ^ {s} \\| P ^ {v 2 s}\\right) = - \\sum_ {c = 1} ^ {C} p _ {(c)} ^ {s} \\log \\frac {p _ {(c)} ^ {s}}{p _ {(c)} ^ {v 2 s}} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.291, + 0.493 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.194, + 0.516 + ], + "angle": 0, + "content": "3.3.1 Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.6 + ], + "angle": 0, + "content": "During the training process, we adopt a random sampling strategy to sample fixed \\( T \\) frames for each video. In order to obtain the best possible performance, our framework jointly trains the classification objective and the learning scheme in an end-to-end manner. The final loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.609, + 0.469, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\delta \\mathcal {L} _ {s i m} + \\zeta \\mathcal {L} _ {t s s} + \\eta \\mathcal {L} _ {x m} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.469, + 0.679 + ], + "angle": 0, + "content": "where \\(\\delta, \\zeta\\) and \\(\\eta\\) are hyperparameters that control the importance of the learning scheme. We use \\(\\delta = 0.1\\), \\(\\zeta = 1\\) and \\(\\eta = 0.1\\) in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.198, + 0.71 + ], + "angle": 0, + "content": "3.3.2 Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.811 + ], + "angle": 0, + "content": "During the inference process, a uniform sampling strategy is applied to sample fixed \\( T \\) frames for each video. Only the video model is considered for inference. The explanation of the reasoning process can be explicitly shown by the proximity of the visual representation to the semantic representation in the joint embedding space." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.21, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.276, + 0.862 + ], + "angle": 0, + "content": "4.1. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments on two extensive video datasets, detailed as follows: (1) Charades [28]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.09, + 0.892, + 0.131 + ], + "angle": 0, + "content": "Table 1. Ablation study on the Charades and CAD-120 datasets using each proposed module. \"S\" denotes the spatial token selection and \"T\" denotes the temporal token selection." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.133, + 0.887, + 0.23 + ], + "angle": 0, + "content": "
MethodsCharadesCAD-120
mAP (%)↑Num ↓mAR ↑Num ↓
DT-Former w/o S,T61.436.60.7349.8
DT-Former w/o S61.231.10.7217.4
DT-Former w/o T61.230.70.7416.7
DT-Former61.126.00.7514.2
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.254, + 0.892, + 0.51 + ], + "angle": 0, + "content": "It contains 157 action classes and consists of about \\(9.8\\mathrm{k}\\) untrimmed videos, among which \\(7.9\\mathrm{k}\\) are used for training and \\(1.8\\mathrm{k}\\) for testing. Each video contains an average of 6.8 distinct action categories and multiple actions can happen at the same time, which makes the recognition extremely challenging. The Action Genome dataset [11] provides fine-grained annotations for the Charades dataset, which provides frame-level relation annotations for videos. Overall, it annotates 476K object bounding boxes and 1.72M relations. (2) CAD-120. Introduced by [15], the CAD-120 dataset is an RGB-D dataset designed for activity understanding. It contains 551 video clips of 4 subjects performing 10 different activities in different environments, such as a kitchen, a living room, and office, etc. To train on our method, we leverage the re-annotated version provided by [42], which provides detailed relationships and attributes for the video frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.511, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Evaluation protocol. Following the experimental protocol of [11], We measure multi-label action recognition performance in term of the Mean Average Precision (mAP) on Charades dataset. For CAD-120 dataset, we calculate the Mean Average Recall (mAR) to evaluate whether the model successfully recognizes the performed actions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.612, + 0.665, + 0.627 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.759, + 0.65 + ], + "angle": 0, + "content": "4.2.1 Effectiveness of Each Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 1 reports the effectiveness of each module of the proposed architecture. We evaluate the performance of the DT-Former using different settings, i.e., canceling spatial token selection or canceling temporal token selection or both. Here, the DT-Former corresponds to the performance obtained by the video model. The metric scheme 'Num' refers to the average number of tokens retained per video after token selection. On the CAD-120 dataset, we noted a modest improvement in the accuracy of our architecture, attributable to the spatial-temporal token selection. The CAD-120 dataset typically features videos with a single action, usually characterized by a pair of relational transitions. By eliminating irrelevant features, the model's risk of overfitting is reduced, thereby enhancing its ability to generalize. As expected, more tokens are discarded in the CAD-120 dataset than in the complex Charades dataset. Overall," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "adding either or both of the token selection modules can reduce the number of tokens without significantly affecting that the recognition performance. This demonstrates our architecture's capability in identifying key relational transitions for action recognition and disregarding superfluous/redundant relations." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.194, + 0.471, + 0.235 + ], + "angle": 0, + "content": "Table 2. Ablation study of learning scheme on the Charades and CAD-120 datasets. \\(\\checkmark\\) indicates that the component is applied in the experiments." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.237, + 0.422, + 0.354 + ], + "angle": 0, + "content": "
w/ \nLsimw/ \nLtssw/ \nLxmmAP on \nCharades (%) ↑mAR on \nCAD-120 ↑
61.10.75
✓--62.20.79
-✓-61.70.79
--✓62.90.81
✓-✓63.40.83
✓✓✓63.60.85
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.365, + 0.391 + ], + "angle": 0, + "content": "4.2.2 Effectiveness of Learning Scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.657 + ], + "angle": 0, + "content": "To explore the effectiveness of the visual-semantic joint embedding \\((\\mathcal{L}_{sim})\\), token selection supervision \\((\\mathcal{L}_{tss})\\) and cross-modal learning \\((\\mathcal{L}_{xm})\\) in our learning scheme, we conduct related ablation studies using different settings, i.e., cancelling one or any two or all our key modules. As reported in Table 2, the visual-semantic joint embedding, token selection supervision and cross-modal learning components collectively or individually contribute to the final performance improvement. This improvement confirms our hypothesis that language models can help video models to improve performance through knowledge transfer. It can be noted that due to the advantages of the learning scheme, the visual model improves from 61.1 to 63.6 in terms of mAP metric and from 0.75 to 0.85 in terms of mAR metric on Charade and CAD-120, respectively. These results demonstrate that the learning scheme plays an important role in our proposed LaIAR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.469, + 0.709 + ], + "angle": 0, + "content": "4.2.3 Effectiveness of Robustness Against Domain Shift" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The performance of RGB-based methods drops drastically when the training and testing data do not share the same distribution caused by change of scene, camera viewpoint or actor [40]. Our proposed model can adapt to video distribution shifts with the aid of semantic modality, which are invariant to complex appearance variations. To demonstrate the robustness of our proposed framework to domain shift, we split the Charades dataset into five subsets with non-overlapping training scenes and test scenes. Table 3 reports the average and variance of five accuracies for these five subsets. The variance of our method is clearly stable and indicates robustness to domain shift." + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.712, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation study of the domain shift on the Charades dataset." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.12, + 0.716, + 0.171 + ], + "angle": 0, + "content": "
MethodAccuracy
AverageVariance
STIGPN [34]54.10.30
Ours57.20.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.718, + 0.09, + 0.885, + 0.131 + ], + "angle": 0, + "content": "Table 4. Comparison of accuracy using predicted and annotated relationships." + }, + { + "type": "table", + "bbox": [ + 0.723, + 0.132, + 0.875, + 0.172 + ], + "angle": 0, + "content": "
Evaluation ModemAP
Prediction62.4
Label63.6
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.196, + 0.879, + 0.212 + ], + "angle": 0, + "content": "4.2.4 Effectiveness of Using Predicted Relationships." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.22, + 0.892, + 0.431 + ], + "angle": 0, + "content": "As previously stated, visual relationship categories can be manually annotated or identified by the visual relationship detection network [5]. To explore the impact of the two modes on accuracy, we compared the proposed method based on the ground truth and the predicted semantic relationships during training. The results, as detailed in Table 4, reveal that using predicted semantic relations can indeed enhance the accuracy of the video model on the Charades dataset (achieving \\(62.4\\%\\) mean Average Precision (mAP) versus \\(61.1\\%\\) mAP shown in Table 2). Moreover, the accuracy does not significantly decrease when using predicted data instead of ground truth. This demonstrates the effectiveness of the proposed method in mining relational transformations from real video data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.807, + 0.456 + ], + "angle": 0, + "content": "4.3. Comparison to the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare the action recognition accuracy of the proposed method and the state-of-the-art methods (SoTA) on the Charades and the CAD-120 datasets, respectively. Table 5 summarizes the results on Charades. It can be seen that our proposed method outperforms several previous 3D CNN approaches in terms of mAP, including I3D [3], SlowFast [7] and LFB [36]. This demonstrates that our method can fully capture action cues through the visual relationship transitions, based on the human/objects information detected from a single video frame (rather than using the entire scene like I3D). STRG [35] and SGFB [11] model the action based objects and visual relationships, respectively, and overlook explicit modeling of temporal dynamics of the interaction between objects. Though VideoLN [12] and OR2G [24] takes visual relationship transitions into account, it is difficult for these methods to achieve accurate action inference due to the limitations of scene graph predictors at test time. For comparison in a modality with only RGB video frames, our method achieved the best performance compared with the existing methods. We also evaluated our method in Oracle evaluation mode, which leverages the ground-truth of bounding box and relationships on a frame. As reported in the last row of Table 5, our method still achieves best performance on the Charades dataset. It is important to mention that OR2G [24] used ground-truth scene graphs to enhance its final performance, whereas our network uses only the bounding boxes of humans and objects. Despite this, our method demonstrates strong performance in both evaluation modes, validating the effective" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18884" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.285, + 0.107 + ], + "angle": 0, + "content": "ness of our proposed approach." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.119, + 0.468, + 0.161 + ], + "angle": 0, + "content": "Table 5. Multi-label action recognition performance comparison on the Charades's validation set in term of mAP. SG: ground truth scene graph. Bbox: Bounding Box. Higher values are better." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.164, + 0.462, + 0.359 + ], + "angle": 0, + "content": "
MethodsBackboneModalitymAP
I3D [3]R101-I3DRGB15.6
VideoMLN [12]R101RGB38.4
STRG [35]R101-I3D-NLRGB39.7
SlowFast [7]R101RGB42.1
LFB [36]R101-I3D-NLRGB42.5
SGFB [11]R101-I3D-NLRGB44.3
OR2G [24]R101-I3D-NLRGB44.9
OursR101-I3D-NLRGB45.1
SGFB Oracle [11]R101-I3D-NLRGB+SG60.3
VideoMLN Oracle [12]R101RGB+SG62.8
OR2G Oracle [24]R101RGB+SG63.3
Ours OracleR101RGB+Bbox63.6
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.47, + 0.526 + ], + "angle": 0, + "content": "For CAD-120 dataset, we follow the same experimental protocol as in [42] and divide the long video sequences into small segments based on individual sub-actions and evaluate the average recall metric for each sub-action. Table 6 summarizes the results on CAD-120. Explainable AAR-RAR [42] interpret the action reasoning process through the changes of relationship between objects or the attribute of objects across time. Our method is able to give the same explanation and outperforms these methods, achieving state-of-the-art performance with \\(0.85\\mathrm{mAR}\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.538, + 0.468, + 0.552 + ], + "angle": 0, + "content": "Table 6. Experimental results on CAD-120 for action recognition." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.553, + 0.447, + 0.682 + ], + "angle": 0, + "content": "
MethodsModalitymAR
Temporal Segment [33]RGB0.42
Flow0.71
RGB + Flow0.77
Explainable AAR-RAR [42]RGB0.80
VideoMLN [12]RGB0.83
OursRGB0.85
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.371, + 0.726 + ], + "angle": 0, + "content": "5. Interpretation and Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To intuitively demonstrate the interpretability effect of our proposed model, we provide interpretable representation and a visualization example. As shown in Figure 3, in the inference stage, we first extract the visual relation representations of human-object pairs in each frame. Then, our proposed DT-Former selects important relations in temporal and spatial dimensions and predicts action by modeling the important relation transition. Finally, the visual representations of important relations are mapped into the joint embedding space to find their nearest neighbor semantic labels, which can provide explicit evidence for the action" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "reasoning process. In this example, the relation representations between the person and the box in the second and tenth frames are selected as cues for the action recognition. The nearest semantic labels of these two representations in the joint embedding space are \"holding box\" and \"not holding box\", respectively. Here, the consequences of observations 'holding' \\(\\rightarrow\\) 'not holding' provide a clear sign of the action \"place\"." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.214, + 0.891, + 0.418 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 3. An example of action recognition performed by the proposed method and its corresponding process of providing explanations. The shaded visual relation representations indicate that it is not selected by DT-Former." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.619, + 0.506 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.817 + ], + "angle": 0, + "content": "In this paper, we introduced a new framework, LaIAR, designed to transfer the knowledge from the language model to the video model to improve the recognition performance and interpretability of video models. Specifically, we build a language model and a video model, which take semantic relation and visual relation representations as input, respectively. These two models share the same architecture, namely DT-Former. This architecture is tailored to select the most important relations for action recognition from all the relations in video and to model the fine-grained relation transitions within videos. Our framework also incorporates three novel knowledge transfer strategies in our learning scheme to facilitate the knowledge transfer from the language model to the video model. This not only boosts the performance but also enhances the interpretability of the video model. Ablation experiments verified the effectiveness of the DT-Former, the learning scheme module and the robustness against domain shift. We conducted extensive experiments on Charades and CAD-120 datasets to demonstrate the superior performance of our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "Acknowledgements: This work is partially supported by the National Natural Science Foundation of China under Grant No.62073252 and No.62072358. It was also supported by Natural Science Basic Research Program of Shaanxi under Grant No.2024JC-JCQN-66." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18885" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.261, + 0.47, + 0.316 + ], + "angle": 0, + "content": "[4] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.318, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[5] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.571 + ], + "angle": 0, + "content": "[8] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.575, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.47, + 0.701 + ], + "angle": 0, + "content": "[10] Hua Hua, Dongxu Li, Ruiqi Li, Peng Zhang, Jochen Renz, and Anthony Cohn. Towards explainable action recognition by salient qualitative spatial object relation chains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5710-5718, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[11] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[12] Yang Jin, Linchao Zhu, and Yadong Mu. Complex video action reasoning via learnable markov logic network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3242-3251, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.833, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In European Conference on Computer Vision, pages 370-387. Springer, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.162 + ], + "angle": 0, + "content": "[14] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 784-794, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[15] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[17] Christopher A Kurby and Jeffrey M Zacks. Segmentation in the perception and memory of events. Trends in cognitive sciences, 12(2):72-79, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[18] Pilhyeon Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, and Hyeran Byun. Decomposed cross-modal distillation for rgb-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2373-2383, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.894, + 0.441 + ], + "angle": 0, + "content": "[19] Yi Li and Nuno Vasconcelos. Improving video model transfer with dynamic representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19280-19291, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.443, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[20] Xue Lin, Qi Zou, Xixia Xu, Yaping Huang, and Ding Ding. Effects of motion-relevant knowledge from unlabeled video to human-object interaction detection. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.894, + 0.566 + ], + "angle": 0, + "content": "[21] Guan Luo, Shuang Yang, Guodong Tian, Chunfeng Yuan, Weiming Hu, and Stephen J Maybank. Learning human actions by combining global dynamics and local appearance. IEEE transactions on pattern analysis and machine intelligence, 36(12):2466-2482, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.894, + 0.636 + ], + "angle": 0, + "content": "[22] Lili Meng, Bo Zhao, Bo Chang, Gao Huang, Wei Sun, Frederick Tung, and Leonid Sigal. Interpretable spatio-temporal attention for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.638, + 0.894, + 0.678 + ], + "angle": 0, + "content": "[23] Ali Modarressi, Hosein Mohebbi, and Mohammad Taher Pilehvar. Adapler: Speeding up inference by adaptive length reduction. arXiv preprint arXiv:2203.08991, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.894, + 0.734 + ], + "angle": 0, + "content": "[24] Yangjun Ou, Li Mi, and Zhenzhong Chen. Object-relation reasoning graph for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20133-20142, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[25] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[26] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[27] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18886" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "models instead. Nature machine intelligence, 1(5):206-215, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.216 + ], + "angle": 0, + "content": "[28] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[29] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[30] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.424 + ], + "angle": 0, + "content": "[31] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[32] Junke Wang, Xitong Yang, Hengduo Li, Li Liu, Zuxuan Wu, and Yu-Gang Jiang. Efficient video transformers with spatial-temporal token selection. In European Conference on Computer Vision, pages 69-86. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.55 + ], + "angle": 0, + "content": "[33] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.634 + ], + "angle": 0, + "content": "[34] Ning Wang, Guangming Zhu, Hongsheng Li, Mingtao Feng, Xia Zhao, Lan Ni, Peiyi Shen, Lin Mei, and Liang Zhang. Exploring spatio-temporal graph convolution for video-based human-object interaction recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.469, + 0.677 + ], + "angle": 0, + "content": "[35] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In Proceedings of the European conference on computer vision (ECCV), pages 399-417, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[36] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[37] Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[38] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[40] Yunhua Zhang, Hazel Doughty, Ling Shao, and Cees GM Snoek. Audio-adaptive activity recognition across video domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13791-13800, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[41] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[42] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Explainable video action reasoning via prior knowledge and state transitions. In Proceedings of the 27th acm international conference on multimedia, pages 521-529, 2019. 2, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18887" + } + ] +] \ No newline at end of file diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_origin.pdf b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..be05c527396694a999110a51570a944a0ebb5b52 --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/f807e59b-d0ea-4264-a8c5-fca1bb6bb589_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66aef2b28ca74257e427fbfc27f052701a7ded6713f27e02e1ace4c6e4b7ad10 +size 942884 diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/full.md b/2024/Language Model Guided Interpretable Video Action Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..40c5fb5b6166417293592adcec942d2ac558ee2d --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/full.md @@ -0,0 +1,347 @@ +# Language Model Guided Interpretable Video Action Reasoning + +Ning Wang $^{1*}$ , Guangming Zhu $^{1\dagger,*}$ , HS Li $^{1}$ , Liang Zhang $^{1\dagger}$ , Syed Afaq Ali Shah $^{2}$ , Mohammed Bennamoun $^{3}$ , Xidian University, $^{2}$ Edith Cowan University, $^{3}$ University of Western Australia + +{ningwang,hsli}@stu.xidian.edu.cn，{gmzhu,liangzhang}@xidian.edu.cn, + +afaq.shah@ecu.edu.au, mohammed.bennamoun@uwa.edu.au + +# Abstract + +While neural networks have excelled in video action recognition tasks, their "black-box" nature often obscures the understanding of their decision-making processes. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. These models, however, usually fall short in performance compared to their "black-box" counterparts. In this work, we present a new framework named Language-guided Interpretable Action Recognition framework (LaIAR). LaIAR leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we redefine the problem of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on two complex video action datasets, Charades & CAD-120, validates the improved performance and interpretability of our LaIAR framework. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR. + +# 1. Introduction + +Building on the advancements of deep learning in image recognition [9, 16, 31], neural network (NN) models have become the leading approach for video-related challenges, including action recognition [18, 25, 30]. Yet, many of the top-tier action recognition techniques [19, 34] deploy NNs in an opaque, black-box fashion. This lack of transparency does not offer clear justification for their decisions, hindering their utility in various real-world contexts [13], especially those with rigorous security demands. These considerations drive us to develop an action reasoning system that pairs exceptional performance with clear interpretability. + +# Action : “Someone is standing up from somewhere” Time + +![](images/7301122c9a65b8e52e309af6da4b44743dedf2800b8c7f61892606943ccd6386.jpg) +(a) An illustration example of relationship transition in the videos. + +![](images/d0e14a0c61638542a95e0520ff39b4ae5c57466ba70388a63799699f7d21c478.jpg) + +![](images/6ed8c89f66ce624c56a36b8ddeafa9565c41240285e7974ba78fe9d445ef8f57.jpg) + +![](images/9f0578ad47c906b7a5f213f40b53bef711deb37dcbae8835b275818ceee13d81.jpg) +(b) Traditional two-stage interpretable reasoning model. + +![](images/ff44629ff6dafbdc2ebb43b1840f890d6388078fa6877f7b91eb59663ef0df4b.jpg) +language model +first stage second stage + +![](images/4c1ada2e923cfbfbe8a8ab54f98fb6d30ca55bafe56d6f4ade6878cc2fb58df4.jpg) + +![](images/8225b8383cd0fbe7a010e04eb985ec0d08ebccac9905d55abfbbb2967953e305.jpg) + +![](images/2956b3b4d3c93f4d7f545ded966cdc1da86807fee68ab7d2fb45ca2f97501aa0.jpg) + +![](images/2984a53c740bfe00f5dcc495001af5bc0ac4119827729c768a7196f5f6e7f61f.jpg) + +![](images/e23a94b8156380744e980fc765f2fc1e3a337cfcad2fdc8f2411ef9a6eec9b33.jpg) + +![](images/9349b1ce5232d5a3984b6ff140dd1c92ecdfe1960f8241849304bea75db44cb7.jpg) +(c) Language model guided video interpretable reasoning model. + +![](images/849f7cc58e06e0e8af6de2ddf321ed5084b5d36310be59c97a78cee60df8895f.jpg) + +![](images/d11af26f49783e77a436264323eb94144aead1e7e13ddc205b939d6b75c510a3.jpg) + +Figure 1. (a) An example of action that can be decomposed into relationship transitions (i.e., when the transition is 'sitting on' $\rightarrow$ 'not contacting' between $<\text{person, bed}>$ pair, it represents the action "Someone is standing up from somewhere"). (b) Traditional two-stage methods usually predict the scene graph first, and then use language models to capture the semantic-level relationship transitions. (c) Our method exploits a language model to guide the video model to capture the relationship transition during training. During inference, our method processes videos and directly recognizes actions, providing supportive evidence. + +Most of the current interpretable action recognition techniques [21, 22, 29] aim to elucidate the decision-making process of NNs using post-hoc explanations, with a particular emphasis on gradient-based and perturbation-based approaches. However, despite notable advancements, these explanations can be problematic because they might not be faithful to what the network computes, as highlighted by [27]. A compelling direction in interpretability re + +volves around the concept of built-in explanation models [10, 12, 24, 42]. The essence of these models is their inherent interpretability right from the design stage. Recent strategies decompose a complex action into temporal transitions of human-object relationships, drawing inspiration from the event segmentation theory [17]. An illustration in Figure 1 (a) depicts that for a relationship involving a $<$ person, $\mathrm{bed}>$ pairing, a transition sequence from 'sitting on' to 'not contacting' signifies the action of "Someone is standing up from somewhere". This methodology facilitates action recognition by pinpointing semantic transitions through language models, offering a granular insight into action execution. As shown in Figure 1 (b), Jin and Ou et al. [12, 24] extract spatio-temporal scene graphs from video content and apply Markov Logic Network (MLN) based probabilistic logical inference and relation reasoning graphs to create an interpretable representation for a variety of complex actions, respectively. However, it is believed that such models will perform worse than their black-box alternatives [8]. Moreover, these methods divide the process into two stages, namely scene graph prediction and relation modeling. Optimizing these components separately might lead to sub-optimal results. In this paper, we propose to harness the explicit logical inference rules of an interpretable language model to guide the learning process of a video black-box model. Interpretability and strong performance can be attained by focusing solely on the video model during the inference stage. To achieve this, two main challenges arise: 1) Designing a language model that can automatically grasp logical reasoning patterns, sidestepping manual rule creation. 2) Developing a decoupled language-video model architecture that enables the language model to guide the video model's training process. + +To address the aforementioned challenges, we have developed a new framework called Language-guided Interpretable Action Recognition framework named LaIAR. As depicted in Figure 1 (c), LaIAR constructs an action recognition model that both implicitly and explicitly exploits fine-grained knowledge of relationship transitions from an interpretable built-in language model. Specifically, we use dynamic token transformers (DT-Former) to both video and language inputs, selectively focusing on important relationships in a data-driven manner, and disregarding non-contributory ones. We aim to redefine the traditional decision interpretation challenge of video models towards a visual-language relation alignment problem. The relationship prioritization determined by the language model then explicitly guides the video model in identifying the most relevant relationships. We propose a learning strategy to facilitate knowledge transfer between language and video to improve the performance of the video model. A key feature of LaIAR is its modular design: during the inference phase, only RGB data serves as input to predict actions, providing + +a direct and transparent justification. + +To summarize, our contributions are three-fold: 1) We propose a novel LaIAR framework that can automatically mine fine-grained relation transitions from data and create interpretable representations for various complex actions. 2) We design a decoupled cross-modal knowledge transfer architecture that leverages useful knowledge from language models to improve the performance and interpretability of the video model at training time, and achieves high-performance interpretable reasoning for videos at test time. 3) Our method achieves state-of-the-art results on two large-scale action recognition benchmarks. + +# 2. Related Work + +# 2.1. Interpretable Video Action Recognition + +Interpretable video action recognition methods can be categorized into two types: post-hoc method and built-in method. post-hoc techniques generate explanations for the network's decision-making process after the network is trained. [22] introduced an interpretable and easy plug-in spatial-temporal attention mechanism for video action recognition to improve the interpretability of the model for video action recognition. [29] developed an interpretable temporal convolutional network to explain the decision-making process of action recognition through each of the learned filters in a Res-TCN. [21] combined both global dynamics and local details to learn human action, using gradient-weighted class activation mapping (Grad-CAM) to visualize the model's attention to action-critical regions. Although these methods have the advantage of not imposing any model constraints, they may be incomplete or unfaithful to the model's reasoning [27]. In contrast, built-in methods restrict the interpretation to be consistent with the model's inferences. [42] approached action reasoning by modeling semantic-level state transitions between two consecutive frames as defined by domain experts. In [10], a method is proposed to achieve interpretability of action recognition by incorporating qualitative spatial reasoning and extracting salient relation chains. Some recent methods, like [12, 24] decompose complex action into continuous relationship transitions according to the event segmentation theory [17]. These methods model the relationship transitions at the semantic level to recognize actions. In this paper, we propose to construct a high performance interpretable-by-design action classifier by guiding a video model with an interpretable language model. + +# 2.2. Adaptive Inference in Transformers. + +As their popularity soars, adaptive inference for language and vision transformers has caught the attention of researchers. In [37], an adaptive language transformer is proposed to achieve fixed-scale reduction of the input se + +quence to improve inference speed by dynamically selecting important tokens and removing the irrelevant ones. In [23], a threshold mechanism is introduced to determine the importance of each token and dynamically select the tokens according to the importance of the input sequence. [14] used the mean of attention matrix column values of the transformers to determine the importance score of each token, facilitating token pruning. [38] developed an adaptive token generation mechanism to determine the required number and size of tokens, thereby reducing the computational and memory overhead of the model on images. [32] devised a token selection framework to dynamically select important tokens across the temporal and spatial dimensions of the video input. In this paper, we propose a lightweight token selection method based on Gumbel-Softmax and apply it to our cross-modal transformers for spatio-temporal token selection. + +# 2.3. Cross-modal Knowledge Transferring + +The past few years have seen an increasing interest in cross-modal knowledge transfer techniques for detection and segmentation tasks. [20] introduced a method to transfer the motion-related knowledge of unlabeled videos to Human-Object Interactions (HOI) detection to infer rare or unseen HOIs. In [40], reliable domain-invariant sound cues are exploited to help video activity recognition models adapt to video distribution shifts. Lately, knowledge distillation techniques have been extended to transfer knowledge across different modalities. For instance, [18] proposed a decomposed cross-modal distillation framework to improve RGB-based temporal action detection by transferring knowledge from the optical flow modality. Similarly, [39] proposed a modified knowledge distillation method that boosts the performance of single-modal 3D captioning by transferring color and texture-aware information from 2D images into 3D object representations. In contrast to these methods, we propose a well-designed knowledge-guided framework to enable cross-modal learning by decoupling information transfer in video and language. + +# 3. The Proposed Approach + +Our proposed method is designed to exploit multimodality by enabling information transfer from language descriptions to videos. This enables the video model to effectively learn from the language model effectively. This is achieved by the video model mimicking the output of the language model, thereby leveraging the intrinsic capabilities of the language model. Specifically, the video frames and the language description (represented as a spatio-temporal scene graph in [11]) are first processed by the encoding network to extract the paired visual and semantic relationship representations. Then, these paired visual and semantic relationship representations are sepa + +rately fed to DT-Former module, which models the key relationship transition for action recognition. Finally, we propose a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning) to improve the performance and interpretability of the video model by facilitating the knowledge transfer from the language model to the video model. An overview of our proposed method is shown in Figure 2 (a). Note that, in our proposed approach, only the video model is employed for inference once training is complete. + +# 3.1. Architecture + +# 3.1.1 Video and Language Encoder + +Given a video consisting of $T$ frames with $N$ entities of either human or object classes, we use Faster R-CNN [26] with ResNet-101 [16] backbone to detect these entities and extract their features from the video. For the frame $I_{t}$ at time step $t$ , the visual features $\{v_{(t,1)}, v_{(t,2)}, \dots, v_{(t,N)}\}$ , bounding boxes $\{b_{(t,1)}, b_{(t,2)}, \dots, b_{(t,N)}\}$ and object category $\{c_{(t,1)}, c_{(t,2)}, \dots, c_{(t,N)}\}$ of the objects proposals are supplied by the detector. Between each $<\text{human, object}>$ pair in the frame, there is a set of relationships $R_{t} = \{r_{(t,1)}, r_{(t,2)}, \dots, r_{(t,K)}\}$ . Concatenating the visual appearance, spatial information and category embedding between the $i$ -th human and $j$ -th object proposals can represent the visual relation feature $\mathbf{v}_{(t,k)}$ , as follows: + +$$ +\mathbf {v} _ {(t, k)} = \left[ W _ {s} v _ {(t, i)}, W _ {o} v _ {(t, j)}, W _ {u} \varphi \left(u _ {(t, i j)} \oplus f _ {b o x} \left(b _ {(t, i)}, b _ {(t, j)}\right)\right) \right] \tag {1} +$$ + +where $W_{s}$ , $W_{o}$ and $W_{u}$ represent the parameter matrix of the linear transformation. $[,]$ is concatenation operation, $\varphi$ is flattening operation and $\oplus$ is element-wise addition. $u_{(t,ij)}$ the visual feature of the union box of $b_{(t,i)}$ and $b_{(t,j)}$ extracted from the detector. $f_{box}$ maps the 2-channel binary spatial configuration map of bounding boxes $b_{(t,i)}$ and $b_{(t,j)}$ into features of the same dimension as $u_{(t,ij)}$ . + +Unlike the visual relation feature, the semantic relation feature $\mathbf{s}_{(t,k)}$ provides high-level descriptions of the relationship between humans and objects in the videos. The visual relationship categories are either provided as groundtruth or determined by the fine-tuned visual relationship detection network [5]. The features of the semantic relation are obtained by concatenating the three features as follows: + +$$ +\mathbf {s} _ {(t, k)} = \left[ s _ {(t, i)}, r _ {(t, i j)}, s _ {(t, j)} \right] \tag {2} +$$ + +where the $r_{(t,ij)}$ is extracted by embedding the visual relationship category to the semantic feature space. The category embedding vectors $s_{(t,i)}$ and $s_{(t,j)}$ are determined by the categories of human and object, respectively. + +Given the features $\{\mathbf{v}_{(t,k)}\}_{t=1,k=1}^{T,K}$ and the features $\{\mathbf{s}_{(t,k)}\}_{t=1,k=1}^{T,K}$ , we further map the visual relation feature and the semantic relation feature into a joint embedding + +![](images/6b999cfe46c881921eee6a83a22a16f653604e30c8bd721e3cdbfdc175916575.jpg) +(a) overview of the proposed method. + +![](images/1d6dc2c76dd0caecf95c4e583ab58afe0f2c03ccb5d77ea7e623a737df79ef71.jpg) +(b) DT-Former +Figure 2. Overview of our LaIAR. The architecture comprises a language model (top) which takes the language description (represented as a spatio-temporal scene graph in [11]) as input and a video model (bottom) which takes the video frames as input. Both models use DT-Former to capture key relational transitions to recognize actions. We transfer knowledge across modalities using a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning), which can help video model benefit from language model during training. For inference, only the video model is considered. + +space as follows: + +$$ +\mathcal {V} = f _ {\mathrm {v}} \left(\left\{\mathbf {v} _ {(t, k)} \right\} _ {t = 1, k = 1} ^ {T, K}\right), \quad \mathcal {S} = f _ {\mathrm {s}} \left( \right.\left\{ \right.\left\{\mathbf {s} _ {(t, k)} \right\} _ {t = 1, k = 1} ^ {T, K}\left. \right). +$$ + +where the $f_{\mathrm{V}}$ and $f_{\mathrm{S}}$ are the visual encoder and the semantic encoder, respectively. In each encoder, each element of the input is first mapped to a local representation via a linear projection. We then apply a Generalized Pooling Operator (GPO) [4] to aggregate the input, creating a global representation. This global representation is combined with each local representation along the channel dimension. This approach helps to use the contextual information from the entire sequence. The visual embedding $\mathcal{V} \in \mathbb{R}^{T \times K \times D}$ and semantic embedding $S \in \mathbb{R}^{T \times K \times D}$ are aligned in the spatiotemporal dimension, where $D$ denotes the dimension in the common space. + +# 3.1.2 Dynamic Token Transformers + +Video understanding shares several high-level similarities with natural language processing (NLP), as they are both fundamentally based on sequential structures [2]. Our intuition is that we can easily model visual and semantic relations simultaneously from joint embedding space (see Sec 3.2.1 for details). Therefore, we introduce a shared dynamic token transformers (DT-Former), which employs the transformer structure to capture key relationship transitions for action reasoning. It mainly consists of the adaptive token selection and the video transformer module. The adaptive token selection module calculates the contribution score of each token to the classification output, and tokens with lower contribution scores will be discarded. The retained tokens, i.e. important relationship representations, are fed to the video transformer module to capture relation transition cues. Figure 2 (b) shows the architecture of DT-Former. Since the video model and the language + +model share the same DT-Former, we denote the input as $X = \{x_{(t,k)}\}_{t=1,k=1}^{T,K}$ for simplicity. We add a learnable spatiotemporal positional embedding $e_{(t,k)}^{pos}$ to each vector $x_{(t,k)}$ to obtain the embedding token $x_{(t,k)}^{(0)}$ . The superscript corresponds to the layer of the transformer encoder. + +Adaptive Token Selection. Following the ViT approach [6], we concatenate a special learnable vector $(x_{(0,0)}^{(0)} = x_{class})$ representing the embedding of the [class] token in the first position of the sequence. As a large number of relationships between human and objects in a scene are usually redundant, it is essential to reduce these relationships. Inspired by the recent work on token reduction for accelerating transformer inference, we formulate parsing important relations as a token selection problem. To determine whether a token is discarded or retained, we introduce a token selector that consists of an MLP $\sigma$ and a differentiable discrete-valued estimator using the Gumbel-Softmax (GSM) operator: + +$$ +u _ {(t, k)} = \operatorname {G S M} \left\{\sigma \left(\left[ W _ {1} x _ {(t, k)} ^ {(0)}, W _ {2} x _ {\text {c l a s s}} \right]\right) \right\} \tag {3} +$$ + +where $W_{1}$ and $W_{2}$ represent the linear matrices for dimension compression. We concatenate [class] tokens that represent the global representation with input tokens to exploit contextual information of the entire sequence. The binary output $u_{(t,k)} = 0$ indicates that the $t$ -th token of frame $t$ is to be discarded and $u_{(t,k)} = 1$ is to be retained. Token selection can be represented as: $y_{(t,k)}^{(0)} = u_{(t,k)}x_{(t,k)}^{(0)}$ . Note that this operation is differentiable, allowing for end-to-end training tailored for token selection. + +To ensure consistent token reduction across consecutive frames, we apply token selector to both the temporal and spatial dimensions. Recent efforts in the field of frame sampling [41] indicate inherent temporal redundancy in frames. Inspired by this, we first focus on salient + +frames over the entire time horizon, and then delve into those frames to find key relationships. For the input tokens $X = \{x_{(t,k)}^{(0)}\}_{t=1,k=1}^{T,K}$ , we first apply an average-pooling operation to tokens in the spatial dimension to get a sequence of temporal-based tokens $\{x_{(t)}^{(0)}\}_{t=1}^{T}$ , and then feed it to the token selector to generate temporal-based selection signal matrix $\{\dot{u}_{(t)}\}_{t=1}^{T}$ . Finally, we repeat it along the spatial dimension to obtain $\dot{U} = \{\dot{u}_{(t,k)}\}_{t=1,k=1}^{T,K}$ for downstream processing. Similarly, we perform the token selector on each frame separately to generate a selection signal matrix. The spatial-based selection signal matrix can be expressed as $\dot{U} = \{\acute{u}_{(t,k)}\}_{t=1,k=1}^{T,K}$ . Further, the final selection signal matrix can be expressed as $U = \dot{U} \cdot \dot{U} = \{u_{(t,k)}\}_{t=1,k=1}^{T,K}$ . We use matrix multiplication for token selection: $Y = U \cdot X = \{y_{(t,k)}^{(0)}\}_{t=1,k=1}^{T,K}$ . + +Transformer Encoder. In order to model the relationship transition in videos, the token $Y = \{y_{(t,k)}^{(0)}\}_{t=1,k=1}^{T,K}$ are fed to stack of transformer blocks which compute the spatial and temporal self-attention jointly. We convert $Y$ into a set of sequences $\mathbf{Y}^{(0)} = \{y_{(p)}^{(0)}\}_{p=1}^{T \times K}$ , which are then fed into the transformer encoder to extract a video-level representation: + +$$ +\mathbf {Y} ^ {\prime \ell} = \operatorname {M S A} \left(\operatorname {L N} \left(\mathbf {Y} ^ {\ell - 1}\right)\right) + \mathbf {Y} ^ {\ell - 1} \tag {4} +$$ + +$$ +\mathbf {Y} ^ {\ell} = \operatorname {M L P} \left(\ln \left(\mathbf {Y} ^ {\ell}\right)\right) + \mathbf {Y} ^ {\prime^ {\ell}} \tag {5} +$$ + +$$ +\mathbf {z} = \operatorname {L N} \left(\mathbf {Y} _ {(0)} ^ {L}\right) \tag {6} +$$ + +where MSA() and LN() denotes multiheaded self-attention and LayerNorm [1], respectively. $L$ represents the number of transformer blocks. $\mathbf{z}$ denotes the video-level representation, which can be used to predict the final action classes. Based on the above steps, we can get the selection signal matrix $U^{v} = \{u_{(t,k)}^{v}\}_{t=1,k=1}^{T,K}$ , video-level representation $\mathbf{z}^{\mathrm{v}}$ of the video model, the selection signal matrix $U^{s} = \{u_{(t,k)}^{s}\}_{t=1,k=1}^{T,K}$ and video-level representation $\mathbf{z}^{\mathrm{s}}$ of the language model. + +# 3.1.3 Classification Head + +The two classification heads of video and language model predict, $P^v = \{p_{(c)}^v\}_{c=1}^C$ and $P^s = \{p_{(c)}^s\}_{c=1}^C$ respectively for each branch, where $C$ is the number of classes. We minimize the cross-entropy losses between action scores $P^v$ , and $P^s$ and the ground-truth action labels for each action category, denoted as $\mathcal{L}_v$ and $\mathcal{L}_s$ , respectively. The overall loss of two branches can be written as: + +$$ +\mathcal {L} _ {c l s} = \mathcal {L} _ {v} + \mathcal {L} _ {s} \tag {7} +$$ + +The two classification heads only utilize the private information of each modality, in order to allow the video model to benefit from the knowledge of the language model, + +we propose an additional classification head to estimate the other modality's output: the video model estimates the language model $(P^{v2s} = \{p_{(c)}^{v2s}\}_{c=1}^{C})$ . By mimicking not only the class with maximum probability, but also the whole distribution, more information is exchanged, leading to softer labels, which is more beneficial for training our model. + +# 3.2. Learning Scheme + +The goal of our learning scheme is to transfer information across modalities in a controlled manner thus allowing the video model to learn from the language model. This auxiliary objective can effectively improve the performance of the video modality and does not require additional labels from the datasets. Here, we define the visual-semantic joint embedding learning, our token selection supervision loss and an additional cross-modal learning method. + +# 3.2.1 Visual-Semantic Joint Embedding Space + +Our approach begins by aligning the visual and semantic relation representations within a shared vector space. In this configuration, each visual embedding $\hat{\mathbf{v}}_{(t,k)}\in \mathcal{V}$ and $\hat{\mathbf{s}}_{(t,k)}\in S$ pair converge to proximate points. This visual-semantic joint embedding has two main advantages: 1) it helps the video model to improve its generalization since semantic representations are invariant to complex appearance variations. 2) it enables the video model to explicitly represent the relationship transition process, since the visual-semantic joint embedding space can provide semantic labels for each visual relation representation. In this paper, we introduce the contrastive learning of visual-semantic joint embedding and discuss its implementation as following. + +Given a mini-batch $B = \{(\hat{\mathbf{v}}_{(0,0)},\hat{\mathbf{s}}_{(0,0)}),\ldots \}$ of visual-semantic relationship representation pairs, the contrastive learning objective encourages embeddings of positive pairs $(\hat{\mathbf{v}}_{(t,k)},\hat{\mathbf{s}}_{(t,k)})$ to align with each other, while pushing embeddings of the negative pairs apart. Formally, the contrastive loss $\mathcal{L}_{sim}$ is formulated using the symmetric contrastive loss, as follows: + +$$ +L _ {s i m} = - \frac {1}{2 | \mathcal {B} |} \sum_ {i = 1} ^ {| \mathcal {B} |} \left(\overbrace {\log \frac {e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {B} |} e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {j}}}} ^ {\text {v i s u a l} \rightarrow \text {s e m a n t i c}} + \overbrace {\log \frac {e ^ {t \mathbf {x} _ {i} \cdot \mathbf {y} _ {i}}}{\sum_ {j = 1} ^ {| \mathcal {B} |} e ^ {t \mathbf {x} _ {j} \cdot \mathbf {y} _ {i}}}} ^ {\text {s e m a n t i c} \rightarrow \text {v i s u a l}}\right) \tag {8} +$$ + +where $\mathbf{x}_i = \frac{\mathbf{v}_{(t,k)}}{\|\mathbf{v}_{(t,k)}\|_2}$ and $\mathbf{y}_i = \frac{\mathbf{s}_{(t,k)}}{\|\mathbf{s}_{(t,k)}\|_2}$ . $|\mathcal{B}|$ is size of the mini-batch $B$ . $t$ is the learnable temperature parameter. + +# 3.2.2 Token Selection Supervision + +We aim that the key relations obtained by the language model can guide the video model to perform key relations mining. Therefore, we align the token selection signal matrix of the two models. We minimize the mean-squared loss + +between the token selection signal matrix $u_{(t,k)}^v$ of the video model and the token selection signal matrix $u_{(t,k)}^s$ of the language model. To maintain the token selection signal matrix's sparsity, we apply L1 norm to allow $u_{(t,k)}^s$ to have a small number of non-zero values. The token selection supervision loss is defined as: + +$$ +\mathcal {L} _ {t s s} = \frac {1}{T} \sum_ {t = 1} ^ {T} \frac {1}{K} \sum_ {k = 1} ^ {K} \left(\left\| u _ {(t, k)} ^ {v} - u _ {(t, k)} ^ {s} \right\| + \left\| u _ {(t, k)} ^ {s} \right\| _ {1}\right) \tag {9} +$$ + +# 3.2.3 Cross-Modal Learning + +Given that the visual modality is highly sensitive and the semantic modality more robust to the domain shift, the robust semantic modality can guide the sensitive visual modality to the correct classification. We allow the video model estimate the entire distribution of the language model's prediction. Through cross-modal learning, we aim to transfer knowledge from the language model to the video model. We choose KL divergence for the cross-modal loss $\mathcal{L}_{\mathrm{xm}}$ and define it as follows: + +$$ +\mathcal {L} _ {\mathrm {x m}} = D _ {\mathrm {K L}} \left(P ^ {s} \| P ^ {v 2 s}\right) = - \sum_ {c = 1} ^ {C} p _ {(c)} ^ {s} \log \frac {p _ {(c)} ^ {s}}{p _ {(c)} ^ {v 2 s}} \tag {10} +$$ + +# 3.3. Training and Inference + +# 3.3.1 Training + +During the training process, we adopt a random sampling strategy to sample fixed $T$ frames for each video. In order to obtain the best possible performance, our framework jointly trains the classification objective and the learning scheme in an end-to-end manner. The final loss is: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \delta \mathcal {L} _ {s i m} + \zeta \mathcal {L} _ {t s s} + \eta \mathcal {L} _ {x m} \tag {11} +$$ + +where $\delta, \zeta$ and $\eta$ are hyperparameters that control the importance of the learning scheme. We use $\delta = 0.1$ , $\zeta = 1$ and $\eta = 0.1$ in our experiments. + +# 3.3.2 Inference + +During the inference process, a uniform sampling strategy is applied to sample fixed $T$ frames for each video. Only the video model is considered for inference. The explanation of the reasoning process can be explicitly shown by the proximity of the visual representation to the semantic representation in the joint embedding space. + +# 4. Experiments + +# 4.1. Datasets and Metrics + +Datasets. We conduct our experiments on two extensive video datasets, detailed as follows: (1) Charades [28]. + +Table 1. Ablation study on the Charades and CAD-120 datasets using each proposed module. "S" denotes the spatial token selection and "T" denotes the temporal token selection. + +
MethodsCharadesCAD-120
mAP (%)↑Num ↓mAR ↑Num ↓
DT-Former w/o S,T61.436.60.7349.8
DT-Former w/o S61.231.10.7217.4
DT-Former w/o T61.230.70.7416.7
DT-Former61.126.00.7514.2
+ +It contains 157 action classes and consists of about $9.8\mathrm{k}$ untrimmed videos, among which $7.9\mathrm{k}$ are used for training and $1.8\mathrm{k}$ for testing. Each video contains an average of 6.8 distinct action categories and multiple actions can happen at the same time, which makes the recognition extremely challenging. The Action Genome dataset [11] provides fine-grained annotations for the Charades dataset, which provides frame-level relation annotations for videos. Overall, it annotates 476K object bounding boxes and 1.72M relations. (2) CAD-120. Introduced by [15], the CAD-120 dataset is an RGB-D dataset designed for activity understanding. It contains 551 video clips of 4 subjects performing 10 different activities in different environments, such as a kitchen, a living room, and office, etc. To train on our method, we leverage the re-annotated version provided by [42], which provides detailed relationships and attributes for the video frames. + +Evaluation protocol. Following the experimental protocol of [11], We measure multi-label action recognition performance in term of the Mean Average Precision (mAP) on Charades dataset. For CAD-120 dataset, we calculate the Mean Average Recall (mAR) to evaluate whether the model successfully recognizes the performed actions. + +# 4.2. Ablation Studies + +# 4.2.1 Effectiveness of Each Module + +Table 1 reports the effectiveness of each module of the proposed architecture. We evaluate the performance of the DT-Former using different settings, i.e., canceling spatial token selection or canceling temporal token selection or both. Here, the DT-Former corresponds to the performance obtained by the video model. The metric scheme 'Num' refers to the average number of tokens retained per video after token selection. On the CAD-120 dataset, we noted a modest improvement in the accuracy of our architecture, attributable to the spatial-temporal token selection. The CAD-120 dataset typically features videos with a single action, usually characterized by a pair of relational transitions. By eliminating irrelevant features, the model's risk of overfitting is reduced, thereby enhancing its ability to generalize. As expected, more tokens are discarded in the CAD-120 dataset than in the complex Charades dataset. Overall, + +adding either or both of the token selection modules can reduce the number of tokens without significantly affecting that the recognition performance. This demonstrates our architecture's capability in identifying key relational transitions for action recognition and disregarding superfluous/redundant relations. + +Table 2. Ablation study of learning scheme on the Charades and CAD-120 datasets. $\checkmark$ indicates that the component is applied in the experiments. + +
w/ +Lsimw/ +Ltssw/ +LxmmAP on +Charades (%) ↑mAR on +CAD-120 ↑
61.10.75
✓--62.20.79
-✓-61.70.79
--✓62.90.81
✓-✓63.40.83
✓✓✓63.60.85
+ +# 4.2.2 Effectiveness of Learning Scheme + +To explore the effectiveness of the visual-semantic joint embedding $(\mathcal{L}_{sim})$ , token selection supervision $(\mathcal{L}_{tss})$ and cross-modal learning $(\mathcal{L}_{xm})$ in our learning scheme, we conduct related ablation studies using different settings, i.e., cancelling one or any two or all our key modules. As reported in Table 2, the visual-semantic joint embedding, token selection supervision and cross-modal learning components collectively or individually contribute to the final performance improvement. This improvement confirms our hypothesis that language models can help video models to improve performance through knowledge transfer. It can be noted that due to the advantages of the learning scheme, the visual model improves from 61.1 to 63.6 in terms of mAP metric and from 0.75 to 0.85 in terms of mAR metric on Charade and CAD-120, respectively. These results demonstrate that the learning scheme plays an important role in our proposed LaIAR. + +# 4.2.3 Effectiveness of Robustness Against Domain Shift + +The performance of RGB-based methods drops drastically when the training and testing data do not share the same distribution caused by change of scene, camera viewpoint or actor [40]. Our proposed model can adapt to video distribution shifts with the aid of semantic modality, which are invariant to complex appearance variations. To demonstrate the robustness of our proposed framework to domain shift, we split the Charades dataset into five subsets with non-overlapping training scenes and test scenes. Table 3 reports the average and variance of five accuracies for these five subsets. The variance of our method is clearly stable and indicates robustness to domain shift. + +Table 3. Ablation study of the domain shift on the Charades dataset. + +
MethodAccuracy
AverageVariance
STIGPN [34]54.10.30
Ours57.20.11
+ +Table 4. Comparison of accuracy using predicted and annotated relationships. + +
Evaluation ModemAP
Prediction62.4
Label63.6
+ +# 4.2.4 Effectiveness of Using Predicted Relationships. + +As previously stated, visual relationship categories can be manually annotated or identified by the visual relationship detection network [5]. To explore the impact of the two modes on accuracy, we compared the proposed method based on the ground truth and the predicted semantic relationships during training. The results, as detailed in Table 4, reveal that using predicted semantic relations can indeed enhance the accuracy of the video model on the Charades dataset (achieving $62.4\%$ mean Average Precision (mAP) versus $61.1\%$ mAP shown in Table 2). Moreover, the accuracy does not significantly decrease when using predicted data instead of ground truth. This demonstrates the effectiveness of the proposed method in mining relational transformations from real video data. + +# 4.3. Comparison to the State-of-the-Art + +We compare the action recognition accuracy of the proposed method and the state-of-the-art methods (SoTA) on the Charades and the CAD-120 datasets, respectively. Table 5 summarizes the results on Charades. It can be seen that our proposed method outperforms several previous 3D CNN approaches in terms of mAP, including I3D [3], SlowFast [7] and LFB [36]. This demonstrates that our method can fully capture action cues through the visual relationship transitions, based on the human/objects information detected from a single video frame (rather than using the entire scene like I3D). STRG [35] and SGFB [11] model the action based objects and visual relationships, respectively, and overlook explicit modeling of temporal dynamics of the interaction between objects. Though VideoLN [12] and OR2G [24] takes visual relationship transitions into account, it is difficult for these methods to achieve accurate action inference due to the limitations of scene graph predictors at test time. For comparison in a modality with only RGB video frames, our method achieved the best performance compared with the existing methods. We also evaluated our method in Oracle evaluation mode, which leverages the ground-truth of bounding box and relationships on a frame. As reported in the last row of Table 5, our method still achieves best performance on the Charades dataset. It is important to mention that OR2G [24] used ground-truth scene graphs to enhance its final performance, whereas our network uses only the bounding boxes of humans and objects. Despite this, our method demonstrates strong performance in both evaluation modes, validating the effective + +ness of our proposed approach. + +Table 5. Multi-label action recognition performance comparison on the Charades's validation set in term of mAP. SG: ground truth scene graph. Bbox: Bounding Box. Higher values are better. + +
MethodsBackboneModalitymAP
I3D [3]R101-I3DRGB15.6
VideoMLN [12]R101RGB38.4
STRG [35]R101-I3D-NLRGB39.7
SlowFast [7]R101RGB42.1
LFB [36]R101-I3D-NLRGB42.5
SGFB [11]R101-I3D-NLRGB44.3
OR2G [24]R101-I3D-NLRGB44.9
OursR101-I3D-NLRGB45.1
SGFB Oracle [11]R101-I3D-NLRGB+SG60.3
VideoMLN Oracle [12]R101RGB+SG62.8
OR2G Oracle [24]R101RGB+SG63.3
Ours OracleR101RGB+Bbox63.6
+ +For CAD-120 dataset, we follow the same experimental protocol as in [42] and divide the long video sequences into small segments based on individual sub-actions and evaluate the average recall metric for each sub-action. Table 6 summarizes the results on CAD-120. Explainable AAR-RAR [42] interpret the action reasoning process through the changes of relationship between objects or the attribute of objects across time. Our method is able to give the same explanation and outperforms these methods, achieving state-of-the-art performance with $0.85\mathrm{mAR}$ . + +Table 6. Experimental results on CAD-120 for action recognition. + +
MethodsModalitymAR
Temporal Segment [33]RGB0.42
Flow0.71
RGB + Flow0.77
Explainable AAR-RAR [42]RGB0.80
VideoMLN [12]RGB0.83
OursRGB0.85
+ +# 5. Interpretation and Visualization + +To intuitively demonstrate the interpretability effect of our proposed model, we provide interpretable representation and a visualization example. As shown in Figure 3, in the inference stage, we first extract the visual relation representations of human-object pairs in each frame. Then, our proposed DT-Former selects important relations in temporal and spatial dimensions and predicts action by modeling the important relation transition. Finally, the visual representations of important relations are mapped into the joint embedding space to find their nearest neighbor semantic labels, which can provide explicit evidence for the action + +reasoning process. In this example, the relation representations between the person and the box in the second and tenth frames are selected as cues for the action recognition. The nearest semantic labels of these two representations in the joint embedding space are "holding box" and "not holding box", respectively. Here, the consequences of observations 'holding' $\rightarrow$ 'not holding' provide a clear sign of the action "place". + +![](images/cb328bafcbf76717fd84a6636a58c004ace07a0b747546f3b9675369602b7074.jpg) +Figure 3. An example of action recognition performed by the proposed method and its corresponding process of providing explanations. The shaded visual relation representations indicate that it is not selected by DT-Former. + +# 6. Conclusion + +In this paper, we introduced a new framework, LaIAR, designed to transfer the knowledge from the language model to the video model to improve the recognition performance and interpretability of video models. Specifically, we build a language model and a video model, which take semantic relation and visual relation representations as input, respectively. These two models share the same architecture, namely DT-Former. This architecture is tailored to select the most important relations for action recognition from all the relations in video and to model the fine-grained relation transitions within videos. Our framework also incorporates three novel knowledge transfer strategies in our learning scheme to facilitate the knowledge transfer from the language model to the video model. This not only boosts the performance but also enhances the interpretability of the video model. Ablation experiments verified the effectiveness of the DT-Former, the learning scheme module and the robustness against domain shift. We conducted extensive experiments on Charades and CAD-120 datasets to demonstrate the superior performance of our proposed method. + +Acknowledgements: This work is partially supported by the National Natural Science Foundation of China under Grant No.62073252 and No.62072358. It was also supported by Natural Science Basic Research Program of Shaanxi under Grant No.2024JC-JCQN-66. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 4 +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8 +[4] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4 +[5] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 3, 7 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2020. 4 +[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7, 8 +[8] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 2 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[10] Hua Hua, Dongxu Li, Ruiqi Li, Peng Zhang, Jochen Renz, and Anthony Cohn. Towards explainable action recognition by salient qualitative spatial object relation chains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5710-5718, 2022. 2 +[11] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 3, 4, 6, 7, 8 +[12] Yang Jin, Linchao Zhu, and Yadong Mu. Complex video action reasoning via learnable markov logic network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3242-3251, 2022. 2, 7, 8 +[13] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In European Conference on Computer Vision, pages 370-387. Springer, 2022. 1 + +[14] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 784-794, 2022. 3 +[15] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 6 +[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3 +[17] Christopher A Kurby and Jeffrey M Zacks. Segmentation in the perception and memory of events. Trends in cognitive sciences, 12(2):72-79, 2008. 2 +[18] Pilhyeon Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, and Hyeran Byun. Decomposed cross-modal distillation for rgb-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2373-2383, 2023. 1, 3 +[19] Yi Li and Nuno Vasconcelos. Improving video model transfer with dynamic representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19280-19291, 2022. 1 +[20] Xue Lin, Qi Zou, Xixia Xu, Yaping Huang, and Ding Ding. Effects of motion-relevant knowledge from unlabeled video to human-object interaction detection. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3 +[21] Guan Luo, Shuang Yang, Guodong Tian, Chunfeng Yuan, Weiming Hu, and Stephen J Maybank. Learning human actions by combining global dynamics and local appearance. IEEE transactions on pattern analysis and machine intelligence, 36(12):2466-2482, 2014. 1, 2 +[22] Lili Meng, Bo Zhao, Bo Chang, Gao Huang, Wei Sun, Frederick Tung, and Leonid Sigal. Interpretable spatio-temporal attention for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1, 2 +[23] Ali Modarressi, Hosein Mohebbi, and Mohammad Taher Pilehvar. Adapler: Speeding up inference by adaptive length reduction. arXiv preprint arXiv:2203.08991, 2022. 3 +[24] Yangjun Ou, Li Mi, and Zhenzhong Chen. Object-relation reasoning graph for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20133-20142, 2022. 2, 7, 8 +[25] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 1 +[26] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[27] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable + +models instead. Nature machine intelligence, 1(5):206-215, 2019. 1, 2 +[28] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6 +[29] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017. 1, 2 +[30] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence, 2022. 1 +[31] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1 +[32] Junke Wang, Xitong Yang, Hengduo Li, Li Liu, Zuxuan Wu, and Yu-Gang Jiang. Efficient video transformers with spatial-temporal token selection. In European Conference on Computer Vision, pages 69-86. Springer, 2022. 3 +[33] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 8 +[34] Ning Wang, Guangming Zhu, Hongsheng Li, Mingtao Feng, Xia Zhao, Lan Ni, Peiyi Shen, Lin Mei, and Liang Zhang. Exploring spatio-temporal graph convolution for video-based human-object interaction recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 7 +[35] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In Proceedings of the European conference on computer vision (ECCV), pages 399-417, 2018. 7, 8 +[36] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 7, 8 +[37] Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618, 2021. 2 +[38] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3 +[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense + +captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3 +[40] Yunhua Zhang, Hazel Doughty, Ling Shao, and Cees GM Snoek. Audio-adaptive activity recognition across video domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13791-13800, 2022. 3, 7 +[41] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 4 +[42] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Explainable video action reasoning via prior knowledge and state transitions. In Proceedings of the 27th acm international conference on multimedia, pages 521-529, 2019. 2, 6, 8 \ No newline at end of file diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/images.zip b/2024/Language Model Guided Interpretable Video Action Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a34b2378749657e6f70cf53a648ede8c4ab8cbdb --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f46a86b0e8d58acc52dde908ec1cd9451ec0d354a6d73f54faf446c80c6ba85 +size 446530 diff --git a/2024/Language Model Guided Interpretable Video Action Reasoning/layout.json b/2024/Language Model Guided Interpretable Video Action Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..467a314dc52f59f1a8b8dd0887624472ca735d51 --- /dev/null +++ b/2024/Language Model Guided Interpretable Video Action Reasoning/layout.json @@ -0,0 +1,9338 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "text", + "content": "Language Model Guided Interpretable Video Action Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": "Ning Wang" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", Guangming Zhu" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger,*}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", HS Li" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", Liang Zhang" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", Syed Afaq Ali Shah" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", Mohammed Bennamoun" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": ", Xidian University, " + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": "Edith Cowan University, " + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 560, + 171 + ], + "type": "text", + "content": "University of Western Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 173, + 498, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 173, + 498, + 186 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 498, + 186 + ], + "type": "text", + "content": "{ningwang,hsli}@stu.xidian.edu.cn，{gmzhu,liangzhang}@xidian.edu.cn," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 167, + 188, + 444, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 188, + 444, + 198 + ], + "spans": [ + { + "bbox": [ + 167, + 188, + 444, + 198 + ], + "type": "text", + "content": "afaq.shah@ecu.edu.au, mohammed.bennamoun@uwa.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 253, + 290, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 290, + 517 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 290, + 517 + ], + "type": "text", + "content": "While neural networks have excelled in video action recognition tasks, their \"black-box\" nature often obscures the understanding of their decision-making processes. Recent approaches used inherently interpretable models to analyze video actions in a manner akin to human reasoning. These models, however, usually fall short in performance compared to their \"black-box\" counterparts. In this work, we present a new framework named Language-guided Interpretable Action Recognition framework (LaIAR). LaIAR leverages knowledge from language models to enhance both the recognition capabilities and the interpretability of video models. In essence, we redefine the problem of understanding video model decisions as a task of aligning video and language models. Using the logical reasoning captured by the language model, we steer the training of the video model. This integrated approach not only improves the video model's adaptability to different domains but also boosts its overall performance. Extensive experiments on two complex video action datasets, Charades & CAD-120, validates the improved performance and interpretability of our LaIAR framework. The code of LaIAR is available at https://github.com/NingWang2049/LaIAR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 549, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 682 + ], + "type": "text", + "content": "Building on the advancements of deep learning in image recognition [9, 16, 31], neural network (NN) models have become the leading approach for video-related challenges, including action recognition [18, 25, 30]. Yet, many of the top-tier action recognition techniques [19, 34] deploy NNs in an opaque, black-box fashion. This lack of transparency does not offer clear justification for their decisions, hindering their utility in various real-world contexts [13], especially those with rigorous security demands. These considerations drive us to develop an action reasoning system that pairs exceptional performance with clear interpretability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 226, + 545, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 226, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 310, + 226, + 545, + 237 + ], + "type": "text", + "content": "Action : “Someone is standing up from somewhere” Time" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 238, + 408, + 291 + ], + "blocks": [ + { + "bbox": [ + 309, + 238, + 408, + 291 + ], + "lines": [ + { + "bbox": [ + 309, + 238, + 408, + 291 + ], + "spans": [ + { + "bbox": [ + 309, + 238, + 408, + 291 + ], + "type": "image", + "image_path": "7301122c9a65b8e52e309af6da4b44743dedf2800b8c7f61892606943ccd6386.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 291, + 542, + 301 + ], + "lines": [ + { + "bbox": [ + 309, + 291, + 542, + 301 + ], + "spans": [ + { + "bbox": [ + 309, + 291, + 542, + 301 + ], + "type": "text", + "content": "(a) An illustration example of relationship transition in the videos." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 408, + 238, + 480, + 291 + ], + "blocks": [ + { + "bbox": [ + 408, + 238, + 480, + 291 + ], + "lines": [ + { + "bbox": [ + 408, + 238, + 480, + 291 + ], + "spans": [ + { + "bbox": [ + 408, + 238, + 480, + 291 + ], + "type": "image", + "image_path": "d0e14a0c61638542a95e0520ff39b4ae5c57466ba70388a63799699f7d21c478.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 481, + 239, + 543, + 291 + ], + "blocks": [ + { + "bbox": [ + 481, + 239, + 543, + 291 + ], + "lines": [ + { + "bbox": [ + 481, + 239, + 543, + 291 + ], + "spans": [ + { + "bbox": [ + 481, + 239, + 543, + 291 + ], + "type": "image", + "image_path": "6ed8c89f66ce624c56a36b8ddeafa9565c41240285e7974ba78fe9d445ef8f57.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 302, + 407, + 357 + ], + "blocks": [ + { + "bbox": [ + 309, + 302, + 407, + 357 + ], + "lines": [ + { + "bbox": [ + 309, + 302, + 407, + 357 + ], + "spans": [ + { + "bbox": [ + 309, + 302, + 407, + 357 + ], + "type": "image", + "image_path": "9f0578ad47c906b7a5f213f40b53bef711deb37dcbae8835b275818ceee13d81.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 357, + 507, + 368 + ], + "lines": [ + { + "bbox": [ + 309, + 357, + 507, + 368 + ], + "spans": [ + { + "bbox": [ + 309, + 357, + 507, + 368 + ], + "type": "text", + "content": "(b) Traditional two-stage interpretable reasoning model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 407, + 302, + 492, + 344 + ], + "blocks": [ + { + "bbox": [ + 407, + 302, + 492, + 344 + ], + "lines": [ + { + "bbox": [ + 407, + 302, + 492, + 344 + ], + "spans": [ + { + "bbox": [ + 407, + 302, + 492, + 344 + ], + "type": "image", + "image_path": "ff44629ff6dafbdc2ebb43b1840f890d6388078fa6877f7b91eb59663ef0df4b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 312, + 535, + 335 + ], + "lines": [ + { + "bbox": [ + 495, + 312, + 535, + 335 + ], + "spans": [ + { + "bbox": [ + 495, + 312, + 535, + 335 + ], + "type": "text", + "content": "language model" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 344, + 541, + 355 + ], + "lines": [ + { + "bbox": [ + 427, + 344, + 541, + 355 + ], + "spans": [ + { + "bbox": [ + 427, + 344, + 541, + 355 + ], + "type": "text", + "content": "first stage second stage" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 369, + 371, + 426 + ], + "blocks": [ + { + "bbox": [ + 309, + 369, + 371, + 426 + ], + "lines": [ + { + "bbox": [ + 309, + 369, + 371, + 426 + ], + "spans": [ + { + "bbox": [ + 309, + 369, + 371, + 426 + ], + "type": "image", + "image_path": "4c1ada2e923cfbfbe8a8ab54f98fb6d30ca55bafe56d6f4ade6878cc2fb58df4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 371, + 369, + 423, + 426 + ], + "blocks": [ + { + "bbox": [ + 371, + 369, + 423, + 426 + ], + "lines": [ + { + "bbox": [ + 371, + 369, + 423, + 426 + ], + "spans": [ + { + "bbox": [ + 371, + 369, + 423, + 426 + ], + "type": "image", + "image_path": "8225b8383cd0fbe7a010e04eb985ec0d08ebccac9905d55abfbbb2967953e305.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 369, + 492, + 426 + ], + "blocks": [ + { + "bbox": [ + 427, + 369, + 492, + 426 + ], + "lines": [ + { + "bbox": [ + 427, + 369, + 492, + 426 + ], + "spans": [ + { + "bbox": [ + 427, + 369, + 492, + 426 + ], + "type": "image", + "image_path": "2956b3b4d3c93f4d7f545ded966cdc1da86807fee68ab7d2fb45ca2f97501aa0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 493, + 369, + 542, + 426 + ], + "blocks": [ + { + "bbox": [ + 493, + 369, + 542, + 426 + ], + "lines": [ + { + "bbox": [ + 493, + 369, + 542, + 426 + ], + "spans": [ + { + "bbox": [ + 493, + 369, + 542, + 426 + ], + "type": "image", + "image_path": "2984a53c740bfe00f5dcc495001af5bc0ac4119827729c768a7196f5f6e7f61f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 309, + 427, + 369, + 468 + ], + "blocks": [ + { + "bbox": [ + 309, + 427, + 369, + 468 + ], + "lines": [ + { + "bbox": [ + 309, + 427, + 369, + 468 + ], + "spans": [ + { + "bbox": [ + 309, + 427, + 369, + 468 + ], + "type": "image", + "image_path": "e23a94b8156380744e980fc765f2fc1e3a337cfcad2fdc8f2411ef9a6eec9b33.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 370, + 427, + 423, + 478 + ], + "blocks": [ + { + "bbox": [ + 370, + 427, + 423, + 478 + ], + "lines": [ + { + "bbox": [ + 370, + 427, + 423, + 478 + ], + "spans": [ + { + "bbox": [ + 370, + 427, + 423, + 478 + ], + "type": "image", + "image_path": "9349b1ce5232d5a3984b6ff140dd1c92ecdfe1960f8241849304bea75db44cb7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 479, + 538, + 491 + ], + "lines": [ + { + "bbox": [ + 308, + 479, + 538, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 538, + 491 + ], + "type": "text", + "content": "(c) Language model guided video interpretable reasoning model." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 427, + 427, + 492, + 468 + ], + "blocks": [ + { + "bbox": [ + 427, + 427, + 492, + 468 + ], + "lines": [ + { + "bbox": [ + 427, + 427, + 492, + 468 + ], + "spans": [ + { + "bbox": [ + 427, + 427, + 492, + 468 + ], + "type": "image", + "image_path": "849f7cc58e06e0e8af6de2ddf321ed5084b5d36310be59c97a78cee60df8895f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 493, + 427, + 541, + 478 + ], + "blocks": [ + { + "bbox": [ + 493, + 427, + 541, + 478 + ], + "lines": [ + { + "bbox": [ + 493, + 427, + 541, + 478 + ], + "spans": [ + { + "bbox": [ + 493, + 427, + 541, + 478 + ], + "type": "image", + "image_path": "d11af26f49783e77a436264323eb94144aead1e7e13ddc205b939d6b75c510a3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "text", + "content": "Figure 1. (a) An example of action that can be decomposed into relationship transitions (i.e., when the transition is 'sitting on' " + }, + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "text", + "content": " 'not contacting' between " + }, + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "inline_equation", + "content": "<\\text{person, bed}>" + }, + { + "bbox": [ + 304, + 499, + 545, + 609 + ], + "type": "text", + "content": " pair, it represents the action \"Someone is standing up from somewhere\"). (b) Traditional two-stage methods usually predict the scene graph first, and then use language models to capture the semantic-level relationship transitions. (c) Our method exploits a language model to guide the video model to capture the relationship transition during training. During inference, our method processes videos and directly recognizes actions, providing supportive evidence." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "Most of the current interpretable action recognition techniques [21, 22, 29] aim to elucidate the decision-making process of NNs using post-hoc explanations, with a particular emphasis on gradient-based and perturbation-based approaches. However, despite notable advancements, these explanations can be problematic because they might not be faithful to what the network computes, as highlighted by [27]. A compelling direction in interpretability re" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 233, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 233, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 233, + 703 + ], + "type": "text", + "content": "* Ning Wang and Guangming Zhu are co-first authors." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 58, + 703, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 703, + 287, + 713 + ], + "type": "text", + "content": " Liang Zhang and Guangming Zhu are both the corresponding authors." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18878" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "text", + "content": "volves around the concept of built-in explanation models [10, 12, 24, 42]. The essence of these models is their inherent interpretability right from the design stage. Recent strategies decompose a complex action into temporal transitions of human-object relationships, drawing inspiration from the event segmentation theory [17]. An illustration in Figure 1 (a) depicts that for a relationship involving a " + }, + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "text", + "content": " person, " + }, + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "inline_equation", + "content": "\\mathrm{bed}>" + }, + { + "bbox": [ + 47, + 72, + 289, + 468 + ], + "type": "text", + "content": " pairing, a transition sequence from 'sitting on' to 'not contacting' signifies the action of \"Someone is standing up from somewhere\". This methodology facilitates action recognition by pinpointing semantic transitions through language models, offering a granular insight into action execution. As shown in Figure 1 (b), Jin and Ou et al. [12, 24] extract spatio-temporal scene graphs from video content and apply Markov Logic Network (MLN) based probabilistic logical inference and relation reasoning graphs to create an interpretable representation for a variety of complex actions, respectively. However, it is believed that such models will perform worse than their black-box alternatives [8]. Moreover, these methods divide the process into two stages, namely scene graph prediction and relation modeling. Optimizing these components separately might lead to sub-optimal results. In this paper, we propose to harness the explicit logical inference rules of an interpretable language model to guide the learning process of a video black-box model. Interpretability and strong performance can be attained by focusing solely on the video model during the inference stage. To achieve this, two main challenges arise: 1) Designing a language model that can automatically grasp logical reasoning patterns, sidestepping manual rule creation. 2) Developing a decoupled language-video model architecture that enables the language model to guide the video model's training process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 289, + 715 + ], + "type": "text", + "content": "To address the aforementioned challenges, we have developed a new framework called Language-guided Interpretable Action Recognition framework named LaIAR. As depicted in Figure 1 (c), LaIAR constructs an action recognition model that both implicitly and explicitly exploits fine-grained knowledge of relationship transitions from an interpretable built-in language model. Specifically, we use dynamic token transformers (DT-Former) to both video and language inputs, selectively focusing on important relationships in a data-driven manner, and disregarding non-contributory ones. We aim to redefine the traditional decision interpretation challenge of video models towards a visual-language relation alignment problem. The relationship prioritization determined by the language model then explicitly guides the video model in identifying the most relevant relationships. We propose a learning strategy to facilitate knowledge transfer between language and video to improve the performance of the video model. A key feature of LaIAR is its modular design: during the inference phase, only RGB data serves as input to predict actions, providing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 72, + 455, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 455, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 455, + 84 + ], + "type": "text", + "content": "a direct and transparent justification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 85, + 547, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 547, + 217 + ], + "type": "text", + "content": "To summarize, our contributions are three-fold: 1) We propose a novel LaIAR framework that can automatically mine fine-grained relation transitions from data and create interpretable representations for various complex actions. 2) We design a decoupled cross-modal knowledge transfer architecture that leverages useful knowledge from language models to improve the performance and interpretability of the video model at training time, and achieves high-performance interpretable reasoning for videos at test time. 3) Our method achieves state-of-the-art results on two large-scale action recognition benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 228, + 392, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 392, + 241 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 392, + 241 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 248, + 513, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 513, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 513, + 262 + ], + "type": "text", + "content": "2.1. Interpretable Video Action Recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 267, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 547, + 639 + ], + "type": "text", + "content": "Interpretable video action recognition methods can be categorized into two types: post-hoc method and built-in method. post-hoc techniques generate explanations for the network's decision-making process after the network is trained. [22] introduced an interpretable and easy plug-in spatial-temporal attention mechanism for video action recognition to improve the interpretability of the model for video action recognition. [29] developed an interpretable temporal convolutional network to explain the decision-making process of action recognition through each of the learned filters in a Res-TCN. [21] combined both global dynamics and local details to learn human action, using gradient-weighted class activation mapping (Grad-CAM) to visualize the model's attention to action-critical regions. Although these methods have the advantage of not imposing any model constraints, they may be incomplete or unfaithful to the model's reasoning [27]. In contrast, built-in methods restrict the interpretation to be consistent with the model's inferences. [42] approached action reasoning by modeling semantic-level state transitions between two consecutive frames as defined by domain experts. In [10], a method is proposed to achieve interpretability of action recognition by incorporating qualitative spatial reasoning and extracting salient relation chains. Some recent methods, like [12, 24] decompose complex action into continuous relationship transitions according to the event segmentation theory [17]. These methods model the relationship transitions at the semantic level to recognize actions. In this paper, we propose to construct a high performance interpretable-by-design action classifier by guiding a video model with an interpretable language model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 647, + 499, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 499, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 499, + 659 + ], + "type": "text", + "content": "2.2. Adaptive Inference in Transformers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "As their popularity soars, adaptive inference for language and vision transformers has caught the attention of researchers. In [37], an adaptive language transformer is proposed to achieve fixed-scale reduction of the input se" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18879" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 275 + ], + "type": "text", + "content": "quence to improve inference speed by dynamically selecting important tokens and removing the irrelevant ones. In [23], a threshold mechanism is introduced to determine the importance of each token and dynamically select the tokens according to the importance of the input sequence. [14] used the mean of attention matrix column values of the transformers to determine the importance score of each token, facilitating token pruning. [38] developed an adaptive token generation mechanism to determine the required number and size of tokens, thereby reducing the computational and memory overhead of the model on images. [32] devised a token selection framework to dynamically select important tokens across the temporal and spatial dimensions of the video input. In this paper, we propose a lightweight token selection method based on Gumbel-Softmax and apply it to our cross-modal transformers for spatio-temporal token selection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 282, + 248, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 248, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 248, + 295 + ], + "type": "text", + "content": "2.3. Cross-modal Knowledge Transferring" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 540 + ], + "type": "text", + "content": "The past few years have seen an increasing interest in cross-modal knowledge transfer techniques for detection and segmentation tasks. [20] introduced a method to transfer the motion-related knowledge of unlabeled videos to Human-Object Interactions (HOI) detection to infer rare or unseen HOIs. In [40], reliable domain-invariant sound cues are exploited to help video activity recognition models adapt to video distribution shifts. Lately, knowledge distillation techniques have been extended to transfer knowledge across different modalities. For instance, [18] proposed a decomposed cross-modal distillation framework to improve RGB-based temporal action detection by transferring knowledge from the optical flow modality. Similarly, [39] proposed a modified knowledge distillation method that boosts the performance of single-modal 3D captioning by transferring color and texture-aware information from 2D images into 3D object representations. In contrast to these methods, we propose a well-designed knowledge-guided framework to enable cross-modal learning by decoupling information transfer in video and language." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 187, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 187, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 187, + 563 + ], + "type": "text", + "content": "3. The Proposed Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Our proposed method is designed to exploit multimodality by enabling information transfer from language descriptions to videos. This enables the video model to effectively learn from the language model effectively. This is achieved by the video model mimicking the output of the language model, thereby leveraging the intrinsic capabilities of the language model. Specifically, the video frames and the language description (represented as a spatio-temporal scene graph in [11]) are first processed by the encoding network to extract the paired visual and semantic relationship representations. Then, these paired visual and semantic relationship representations are sepa" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "rately fed to DT-Former module, which models the key relationship transition for action recognition. Finally, we propose a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning) to improve the performance and interpretability of the video model by facilitating the knowledge transfer from the language model to the video model. An overview of our proposed method is shown in Figure 2 (a). Note that, in our proposed approach, only the video model is employed for inference once training is complete." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 200, + 388, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 388, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 388, + 211 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 218, + 465, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 218, + 465, + 230 + ], + "spans": [ + { + "bbox": [ + 306, + 218, + 465, + 230 + ], + "type": "text", + "content": "3.1.1 Video and Language Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": "Given a video consisting of " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " frames with " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " entities of either human or object classes, we use Faster R-CNN [26] with ResNet-101 [16] backbone to detect these entities and extract their features from the video. For the frame " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": ", the visual features " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\{v_{(t,1)}, v_{(t,2)}, \\dots, v_{(t,N)}\\}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": ", bounding boxes " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\{b_{(t,1)}, b_{(t,2)}, \\dots, b_{(t,N)}\\}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " and object category " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\{c_{(t,1)}, c_{(t,2)}, \\dots, c_{(t,N)}\\}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " of the objects proposals are supplied by the detector. Between each " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "<\\text{human, object}>" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": " pair in the frame, there is a set of relationships " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "R_{t} = \\{r_{(t,1)}, r_{(t,2)}, \\dots, r_{(t,K)}\\}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": ". Concatenating the visual appearance, spatial information and category embedding between the " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": "-th human and " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": "-th object proposals can represent the visual relation feature " + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{(t,k)}" + }, + { + "bbox": [ + 304, + 237, + 547, + 395 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 403, + 563, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 563, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 563, + 427 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {(t, k)} = \\left[ W _ {s} v _ {(t, i)}, W _ {o} v _ {(t, j)}, W _ {u} \\varphi \\left(u _ {(t, i j)} \\oplus f _ {b o x} \\left(b _ {(t, i)}, b _ {(t, j)}\\right)\\right) \\right] \\tag {1}", + "image_path": "bbcac07e795d1015a9197b91a84c28e06a974bd48cb7e3c48f6176cdb3ef9e6f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "W_{o}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "W_{u}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " represent the parameter matrix of the linear transformation. " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "[,]" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " is concatenation operation, " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " is flattening operation and " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " is element-wise addition. " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "u_{(t,ij)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " the visual feature of the union box of " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "b_{(t,i)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "b_{(t,j)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " extracted from the detector. " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "f_{box}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " maps the 2-channel binary spatial configuration map of bounding boxes " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "b_{(t,i)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "b_{(t,j)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": " into features of the same dimension as " + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "inline_equation", + "content": "u_{(t,ij)}" + }, + { + "bbox": [ + 304, + 429, + 545, + 513 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 512, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 512, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 512, + 545, + 596 + ], + "type": "text", + "content": "Unlike the visual relation feature, the semantic relation feature " + }, + { + "bbox": [ + 304, + 512, + 545, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{(t,k)}" + }, + { + "bbox": [ + 304, + 512, + 545, + 596 + ], + "type": "text", + "content": " provides high-level descriptions of the relationship between humans and objects in the videos. The visual relationship categories are either provided as groundtruth or determined by the fine-tuned visual relationship detection network [5]. The features of the semantic relation are obtained by concatenating the three features as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 366, + 606, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 606, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 366, + 606, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {(t, k)} = \\left[ s _ {(t, i)}, r _ {(t, i j)}, s _ {(t, j)} \\right] \\tag {2}", + "image_path": "2a8cd64057e74cb82d148fff3c66b4dec884cac24488a4f6491829163c3e4781.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "inline_equation", + "content": "r_{(t,ij)}" + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "text", + "content": " is extracted by embedding the visual relationship category to the semantic feature space. The category embedding vectors " + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "inline_equation", + "content": "s_{(t,i)}" + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "inline_equation", + "content": "s_{(t,j)}" + }, + { + "bbox": [ + 304, + 626, + 545, + 673 + ], + "type": "text", + "content": " are determined by the categories of human and object, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "text", + "content": "Given the features " + }, + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "text", + "content": " and the features " + }, + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 306, + 673, + 545, + 714 + ], + "type": "text", + "content": ", we further map the visual relation feature and the semantic relation feature into a joint embedding" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18880" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 392, + 201 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 392, + 201 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 392, + 201 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 392, + 201 + ], + "type": "image", + "image_path": "6b999cfe46c881921eee6a83a22a16f653604e30c8bd721e3cdbfdc175916575.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 203, + 301, + 215 + ], + "lines": [ + { + "bbox": [ + 156, + 203, + 301, + 215 + ], + "spans": [ + { + "bbox": [ + 156, + 203, + 301, + 215 + ], + "type": "text", + "content": "(a) overview of the proposed method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 392, + 78, + 521, + 200 + ], + "blocks": [ + { + "bbox": [ + 392, + 78, + 521, + 200 + ], + "lines": [ + { + "bbox": [ + 392, + 78, + 521, + 200 + ], + "spans": [ + { + "bbox": [ + 392, + 78, + 521, + 200 + ], + "type": "image", + "image_path": "1d6dc2c76dd0caecf95c4e583ab58afe0f2c03ccb5d77ea7e623a737df79ef71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 426, + 203, + 485, + 214 + ], + "lines": [ + { + "bbox": [ + 426, + 203, + 485, + 214 + ], + "spans": [ + { + "bbox": [ + 426, + 203, + 485, + 214 + ], + "type": "text", + "content": "(b) DT-Former" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 224, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 545, + 278 + ], + "type": "text", + "content": "Figure 2. Overview of our LaIAR. The architecture comprises a language model (top) which takes the language description (represented as a spatio-temporal scene graph in [11]) as input and a video model (bottom) which takes the video frames as input. Both models use DT-Former to capture key relational transitions to recognize actions. We transfer knowledge across modalities using a learning scheme (i.e., Joint Embedding Space, Token Selection Supervision and Cross-Modal Learning), which can help video model benefit from language model during training. For inference, only the video model is considered." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 118, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 118, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 118, + 300 + ], + "type": "text", + "content": "space as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 50, + 307, + 282, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 307, + 282, + 328 + ], + "spans": [ + { + "bbox": [ + 50, + 307, + 282, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {V} = f _ {\\mathrm {v}} \\left(\\left\\{\\mathbf {v} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\right), \\quad \\mathcal {S} = f _ {\\mathrm {s}} \\left( \\right.\\left\\{ \\right.\\left\\{\\mathbf {s} _ {(t, k)} \\right\\} _ {t = 1, k = 1} ^ {T, K}\\left. \\right).", + "image_path": "cbb2d4e19e7a6c37e365d88747343c1bbb5109d1734ccb282816ea2c7d1a248a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{V}}" + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{S}}" + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": " are the visual encoder and the semantic encoder, respectively. In each encoder, each element of the input is first mapped to a local representation via a linear projection. We then apply a Generalized Pooling Operator (GPO) [4] to aggregate the input, creating a global representation. This global representation is combined with each local representation along the channel dimension. This approach helps to use the contextual information from the entire sequence. The visual embedding " + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\in \\mathbb{R}^{T \\times K \\times D}" + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": " and semantic embedding " + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{T \\times K \\times D}" + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": " are aligned in the spatiotemporal dimension, where " + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 334, + 287, + 477 + ], + "type": "text", + "content": " denotes the dimension in the common space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 491, + 208, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 208, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 208, + 503 + ], + "type": "text", + "content": "3.1.2 Dynamic Token Transformers" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Video understanding shares several high-level similarities with natural language processing (NLP), as they are both fundamentally based on sequential structures [2]. Our intuition is that we can easily model visual and semantic relations simultaneously from joint embedding space (see Sec 3.2.1 for details). Therefore, we introduce a shared dynamic token transformers (DT-Former), which employs the transformer structure to capture key relationship transitions for action reasoning. It mainly consists of the adaptive token selection and the video transformer module. The adaptive token selection module calculates the contribution score of each token to the classification output, and tokens with lower contribution scores will be discarded. The retained tokens, i.e. important relationship representations, are fed to the video transformer module to capture relation transition cues. Figure 2 (b) shows the architecture of DT-Former. Since the video model and the language" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "content": "model share the same DT-Former, we denote the input as " + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "inline_equation", + "content": "X = \\{x_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "content": " for simplicity. We add a learnable spatiotemporal positional embedding " + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "inline_equation", + "content": "e_{(t,k)}^{pos}" + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "content": " to each vector " + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "inline_equation", + "content": "x_{(t,k)}" + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "content": " to obtain the embedding token " + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "inline_equation", + "content": "x_{(t,k)}^{(0)}" + }, + { + "bbox": [ + 305, + 289, + 545, + 354 + ], + "type": "text", + "content": ". The superscript corresponds to the layer of the transformer encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "text", + "content": "Adaptive Token Selection. Following the ViT approach [6], we concatenate a special learnable vector " + }, + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "inline_equation", + "content": "(x_{(0,0)}^{(0)} = x_{class})" + }, + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "text", + "content": " representing the embedding of the [class] token in the first position of the sequence. As a large number of relationships between human and objects in a scene are usually redundant, it is essential to reduce these relationships. Inspired by the recent work on token reduction for accelerating transformer inference, we formulate parsing important relations as a token selection problem. To determine whether a token is discarded or retained, we introduce a token selector that consists of an MLP " + }, + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 305, + 354, + 546, + 511 + ], + "type": "text", + "content": " and a differentiable discrete-valued estimator using the Gumbel-Softmax (GSM) operator:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 519, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 519, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 341, + 519, + 545, + 537 + ], + "type": "interline_equation", + "content": "u _ {(t, k)} = \\operatorname {G S M} \\left\\{\\sigma \\left(\\left[ W _ {1} x _ {(t, k)} ^ {(0)}, W _ {2} x _ {\\text {c l a s s}} \\right]\\right) \\right\\} \\tag {3}", + "image_path": "588daac50c9dc817cf73a55a26826b64e6661a9a338402992165bd8de736e461.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "W_{2}" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": " represent the linear matrices for dimension compression. We concatenate [class] tokens that represent the global representation with input tokens to exploit contextual information of the entire sequence. The binary output " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "u_{(t,k)} = 0" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": " indicates that the " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": "-th token of frame " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": " is to be discarded and " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "u_{(t,k)} = 1" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": " is to be retained. Token selection can be represented as: " + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "inline_equation", + "content": "y_{(t,k)}^{(0)} = u_{(t,k)}x_{(t,k)}^{(0)}" + }, + { + "bbox": [ + 304, + 543, + 545, + 653 + ], + "type": "text", + "content": ". Note that this operation is differentiable, allowing for end-to-end training tailored for token selection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "To ensure consistent token reduction across consecutive frames, we apply token selector to both the temporal and spatial dimensions. Recent efforts in the field of frame sampling [41] indicate inherent temporal redundancy in frames. Inspired by this, we first focus on salient" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18881" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": "frames over the entire time horizon, and then delve into those frames to find key relationships. For the input tokens " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "X = \\{x_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": ", we first apply an average-pooling operation to tokens in the spatial dimension to get a sequence of temporal-based tokens " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\{x_{(t)}^{(0)}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": ", and then feed it to the token selector to generate temporal-based selection signal matrix " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\{\\dot{u}_{(t)}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": ". Finally, we repeat it along the spatial dimension to obtain " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\dot{U} = \\{\\dot{u}_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": " for downstream processing. Similarly, we perform the token selector on each frame separately to generate a selection signal matrix. The spatial-based selection signal matrix can be expressed as " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "\\dot{U} = \\{\\acute{u}_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": ". Further, the final selection signal matrix can be expressed as " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "U = \\dot{U} \\cdot \\dot{U} = \\{u_{(t,k)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": ". We use matrix multiplication for token selection: " + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "inline_equation", + "content": "Y = U \\cdot X = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 72, + 289, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "text", + "content": "Transformer Encoder. In order to model the relationship transition in videos, the token " + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "inline_equation", + "content": "Y = \\{y_{(t,k)}^{(0)}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "text", + "content": " are fed to stack of transformer blocks which compute the spatial and temporal self-attention jointly. We convert " + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "text", + "content": " into a set of sequences " + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^{(0)} = \\{y_{(p)}^{(0)}\\}_{p=1}^{T \\times K}" + }, + { + "bbox": [ + 47, + 266, + 289, + 353 + ], + "type": "text", + "content": ", which are then fed into the transformer encoder to extract a video-level representation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 360, + 287, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 360, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 96, + 360, + 287, + 376 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {\\prime \\ell} = \\operatorname {M S A} \\left(\\operatorname {L N} \\left(\\mathbf {Y} ^ {\\ell - 1}\\right)\\right) + \\mathbf {Y} ^ {\\ell - 1} \\tag {4}", + "image_path": "4fd0927caec362d98873273880f44f341e7ca61af8fee9cabfff9722eb097b4e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 378, + 287, + 394 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 378, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 99, + 378, + 287, + 394 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {\\ell} = \\operatorname {M L P} \\left(\\ln \\left(\\mathbf {Y} ^ {\\ell}\\right)\\right) + \\mathbf {Y} ^ {\\prime^ {\\ell}} \\tag {5}", + "image_path": "723beb589aca1b40c239a6446ea558c894c28bb7e298dc41a63843fb30019fa3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 396, + 287, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 396, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 107, + 396, + 287, + 412 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\operatorname {L N} \\left(\\mathbf {Y} _ {(0)} ^ {L}\\right) \\tag {6}", + "image_path": "f18abb401cc8464f3bf1d56cc912e167f48e279d0d7928cfd0033d8daea271ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": "where MSA() and LN() denotes multiheaded self-attention and LayerNorm [1], respectively. " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": " represents the number of transformer blocks. " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": " denotes the video-level representation, which can be used to predict the final action classes. Based on the above steps, we can get the selection signal matrix " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "U^{v} = \\{u_{(t,k)}^{v}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": ", video-level representation " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{v}}" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": " of the video model, the selection signal matrix " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "U^{s} = \\{u_{(t,k)}^{s}\\}_{t=1,k=1}^{T,K}" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": " and video-level representation " + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{s}}" + }, + { + "bbox": [ + 47, + 419, + 289, + 529 + ], + "type": "text", + "content": " of the language model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 543, + 163, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 163, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 163, + 554 + ], + "type": "text", + "content": "3.1.3 Classification Head" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": "The two classification heads of video and language model predict, " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "P^v = \\{p_{(c)}^v\\}_{c=1}^C" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "P^s = \\{p_{(c)}^s\\}_{c=1}^C" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": " respectively for each branch, where " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": " is the number of classes. We minimize the cross-entropy losses between action scores " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "P^v" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "P^s" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": " and the ground-truth action labels for each action category, denoted as " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_v" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 46, + 562, + 287, + 646 + ], + "type": "text", + "content": ", respectively. The overall loss of two branches can be written as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 658, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 658, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 132, + 658, + 287, + 671 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} = \\mathcal {L} _ {v} + \\mathcal {L} _ {s} \\tag {7}", + "image_path": "cd4faffe69580f8af226db39424384957dc2172524a47c4fa2dc8981b4114168.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 289, + 715 + ], + "type": "text", + "content": "The two classification heads only utilize the private information of each modality, in order to allow the video model to benefit from the knowledge of the language model," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "text", + "content": "we propose an additional classification head to estimate the other modality's output: the video model estimates the language model " + }, + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "inline_equation", + "content": "(P^{v2s} = \\{p_{(c)}^{v2s}\\}_{c=1}^{C})" + }, + { + "bbox": [ + 304, + 72, + 546, + 144 + ], + "type": "text", + "content": ". By mimicking not only the class with maximum probability, but also the whole distribution, more information is exchanged, leading to softer labels, which is more beneficial for training our model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 150, + 411, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 411, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 411, + 163 + ], + "type": "text", + "content": "3.2. Learning Scheme" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 168, + 547, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 266 + ], + "type": "text", + "content": "The goal of our learning scheme is to transfer information across modalities in a controlled manner thus allowing the video model to learn from the language model. This auxiliary objective can effectively improve the performance of the video modality and does not require additional labels from the datasets. Here, we define the visual-semantic joint embedding learning, our token selection supervision loss and an additional cross-modal learning method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 278, + 512, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 512, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 512, + 291 + ], + "type": "text", + "content": "3.2.1 Visual-Semantic Joint Embedding Space" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "content": "Our approach begins by aligning the visual and semantic relation representations within a shared vector space. In this configuration, each visual embedding " + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_{(t,k)}\\in \\mathcal{V}" + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{(t,k)}\\in S" + }, + { + "bbox": [ + 304, + 297, + 546, + 452 + ], + "type": "text", + "content": " pair converge to proximate points. This visual-semantic joint embedding has two main advantages: 1) it helps the video model to improve its generalization since semantic representations are invariant to complex appearance variations. 2) it enables the video model to explicitly represent the relationship transition process, since the visual-semantic joint embedding space can provide semantic labels for each visual relation representation. In this paper, we introduce the contrastive learning of visual-semantic joint embedding and discuss its implementation as following." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "text", + "content": "Given a mini-batch " + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "inline_equation", + "content": "B = \\{(\\hat{\\mathbf{v}}_{(0,0)},\\hat{\\mathbf{s}}_{(0,0)}),\\ldots \\}" + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "text", + "content": " of visual-semantic relationship representation pairs, the contrastive learning objective encourages embeddings of positive pairs " + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{v}}_{(t,k)},\\hat{\\mathbf{s}}_{(t,k)})" + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "text", + "content": " to align with each other, while pushing embeddings of the negative pairs apart. Formally, the contrastive loss " + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sim}" + }, + { + "bbox": [ + 305, + 452, + 547, + 536 + ], + "type": "text", + "content": " is formulated using the symmetric contrastive loss, as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 541, + 555, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 541, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 555, + 606 + ], + "type": "interline_equation", + "content": "L _ {s i m} = - \\frac {1}{2 | \\mathcal {B} |} \\sum_ {i = 1} ^ {| \\mathcal {B} |} \\left(\\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {j}}}} ^ {\\text {v i s u a l} \\rightarrow \\text {s e m a n t i c}} + \\overbrace {\\log \\frac {e ^ {t \\mathbf {x} _ {i} \\cdot \\mathbf {y} _ {i}}}{\\sum_ {j = 1} ^ {| \\mathcal {B} |} e ^ {t \\mathbf {x} _ {j} \\cdot \\mathbf {y} _ {i}}}} ^ {\\text {s e m a n t i c} \\rightarrow \\text {v i s u a l}}\\right) \\tag {8}", + "image_path": "7020282b13b51c8a85458fec498ff931291dbb651615273302a3d100fee98147.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = \\frac{\\mathbf{v}_{(t,k)}}{\\|\\mathbf{v}_{(t,k)}\\|_2}" + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i = \\frac{\\mathbf{s}_{(t,k)}}{\\|\\mathbf{s}_{(t,k)}\\|_2}" + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "inline_equation", + "content": "|\\mathcal{B}|" + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": " is size of the mini-batch " + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 605, + 545, + 633 + ], + "type": "text", + "content": " is the learnable temperature parameter." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 459, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 459, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 459, + 658 + ], + "type": "text", + "content": "3.2.2 Token Selection Supervision" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We aim that the key relations obtained by the language model can guide the video model to perform key relations mining. Therefore, we align the token selection signal matrix of the two models. We minimize the mean-squared loss" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18882" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "between the token selection signal matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "u_{(t,k)}^v" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " of the video model and the token selection signal matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "u_{(t,k)}^s" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " of the language model. To maintain the token selection signal matrix's sparsity, we apply L1 norm to allow " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "u_{(t,k)}^s" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " to have a small number of non-zero values. The token selection supervision loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 150, + 287, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 287, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t s s} = \\frac {1}{T} \\sum_ {t = 1} ^ {T} \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\left(\\left\\| u _ {(t, k)} ^ {v} - u _ {(t, k)} ^ {s} \\right\\| + \\left\\| u _ {(t, k)} ^ {s} \\right\\| _ {1}\\right) \\tag {9}", + "image_path": "1e47481c55637aa85be188af8c879ad5473ff6fafa7e237b2978030fc6659655.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 207, + 178, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 178, + 220 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 178, + 220 + ], + "type": "text", + "content": "3.2.3 Cross-Modal Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 226, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 334 + ], + "type": "text", + "content": "Given that the visual modality is highly sensitive and the semantic modality more robust to the domain shift, the robust semantic modality can guide the sensitive visual modality to the correct classification. We allow the video model estimate the entire distribution of the language model's prediction. Through cross-modal learning, we aim to transfer knowledge from the language model to the video model. We choose KL divergence for the cross-modal loss " + }, + { + "bbox": [ + 46, + 226, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{xm}}" + }, + { + "bbox": [ + 46, + 226, + 289, + 334 + ], + "type": "text", + "content": " and define it as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 339, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 339, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 63, + 339, + 287, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {x m}} = D _ {\\mathrm {K L}} \\left(P ^ {s} \\| P ^ {v 2 s}\\right) = - \\sum_ {c = 1} ^ {C} p _ {(c)} ^ {s} \\log \\frac {p _ {(c)} ^ {s}}{p _ {(c)} ^ {v 2 s}} \\tag {10}", + "image_path": "e0878a8f56ffcbb7bce469c44c658515307c8d2634d88c2e1f5892e82feef182.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 378, + 178, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 178, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 178, + 390 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 118, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 118, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 118, + 408 + ], + "type": "text", + "content": "3.3.1 Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": "During the training process, we adopt a random sampling strategy to sample fixed " + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 415, + 287, + 475 + ], + "type": "text", + "content": " frames for each video. In order to obtain the best possible performance, our framework jointly trains the classification objective and the learning scheme in an end-to-end manner. The final loss is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 482, + 287, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 482, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 91, + 482, + 287, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\delta \\mathcal {L} _ {s i m} + \\zeta \\mathcal {L} _ {t s s} + \\eta \\mathcal {L} _ {x m} \\tag {11}", + "image_path": "55699e98fbd445051b4c1bf58f819c1ec59bbaec4d63ab06a63bf3deddb18f07.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\delta, \\zeta" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " are hyperparameters that control the importance of the learning scheme. We use " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\delta = 0.1" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\zeta = 1" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "inline_equation", + "content": "\\eta = 0.1" + }, + { + "bbox": [ + 46, + 501, + 287, + 537 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 551, + 121, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 121, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 121, + 562 + ], + "type": "text", + "content": "3.3.2 Inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "text", + "content": "During the inference process, a uniform sampling strategy is applied to sample fixed " + }, + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 570, + 287, + 642 + ], + "type": "text", + "content": " frames for each video. Only the video model is considered for inference. The explanation of the reasoning process can be explicitly shown by the proximity of the visual representation to the semantic representation in the joint embedding space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 128, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 671, + 168, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 168, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 168, + 682 + ], + "type": "text", + "content": "4.1. Datasets and Metrics" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We conduct our experiments on two extensive video datasets, detailed as follows: (1) Charades [28]." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 309, + 105, + 542, + 182 + ], + "blocks": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "lines": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 545, + 103 + ], + "type": "text", + "content": "Table 1. Ablation study on the Charades and CAD-120 datasets using each proposed module. \"S\" denotes the spatial token selection and \"T\" denotes the temporal token selection." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 105, + 542, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 105, + 542, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 105, + 542, + 182 + ], + "type": "table", + "html": "
MethodsCharadesCAD-120
mAP (%)↑Num ↓mAR ↑Num ↓
DT-Former w/o S,T61.436.60.7349.8
DT-Former w/o S61.231.10.7217.4
DT-Former w/o T61.230.70.7416.7
DT-Former61.126.00.7514.2
", + "image_path": "dd6bb759beed83d255d3b85eddf50a47dda8e3276e90f445e5842d5b33f88e93.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "text", + "content": "It contains 157 action classes and consists of about " + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "inline_equation", + "content": "9.8\\mathrm{k}" + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "text", + "content": " untrimmed videos, among which " + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "inline_equation", + "content": "7.9\\mathrm{k}" + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "text", + "content": " are used for training and " + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "inline_equation", + "content": "1.8\\mathrm{k}" + }, + { + "bbox": [ + 304, + 201, + 545, + 403 + ], + "type": "text", + "content": " for testing. Each video contains an average of 6.8 distinct action categories and multiple actions can happen at the same time, which makes the recognition extremely challenging. The Action Genome dataset [11] provides fine-grained annotations for the Charades dataset, which provides frame-level relation annotations for videos. Overall, it annotates 476K object bounding boxes and 1.72M relations. (2) CAD-120. Introduced by [15], the CAD-120 dataset is an RGB-D dataset designed for activity understanding. It contains 551 video clips of 4 subjects performing 10 different activities in different environments, such as a kitchen, a living room, and office, etc. To train on our method, we leverage the re-annotated version provided by [42], which provides detailed relationships and attributes for the video frames." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 404, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 404, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 404, + 545, + 477 + ], + "type": "text", + "content": "Evaluation protocol. Following the experimental protocol of [11], We measure multi-label action recognition performance in term of the Mean Average Precision (mAP) on Charades dataset. For CAD-120 dataset, we calculate the Mean Average Recall (mAR) to evaluate whether the model successfully recognizes the performed actions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 484, + 406, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 484, + 406, + 496 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 406, + 496 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 503, + 464, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 464, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 464, + 514 + ], + "type": "text", + "content": "4.2.1 Effectiveness of Each Module" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Table 1 reports the effectiveness of each module of the proposed architecture. We evaluate the performance of the DT-Former using different settings, i.e., canceling spatial token selection or canceling temporal token selection or both. Here, the DT-Former corresponds to the performance obtained by the video model. The metric scheme 'Num' refers to the average number of tokens retained per video after token selection. On the CAD-120 dataset, we noted a modest improvement in the accuracy of our architecture, attributable to the spatial-temporal token selection. The CAD-120 dataset typically features videos with a single action, usually characterized by a pair of relational transitions. By eliminating irrelevant features, the model's risk of overfitting is reduced, thereby enhancing its ability to generalize. As expected, more tokens are discarded in the CAD-120 dataset than in the complex Charades dataset. Overall," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18883" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "adding either or both of the token selection modules can reduce the number of tokens without significantly affecting that the recognition performance. This demonstrates our architecture's capability in identifying key relational transitions for action recognition and disregarding superfluous/redundant relations." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 187, + 258, + 280 + ], + "blocks": [ + { + "bbox": [ + 46, + 153, + 288, + 186 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 288, + 186 + ], + "type": "text", + "content": "Table 2. Ablation study of learning scheme on the Charades and CAD-120 datasets. " + }, + { + "bbox": [ + 46, + 153, + 288, + 186 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 153, + 288, + 186 + ], + "type": "text", + "content": " indicates that the component is applied in the experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 187, + 258, + 280 + ], + "lines": [ + { + "bbox": [ + 72, + 187, + 258, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 187, + 258, + 280 + ], + "type": "table", + "html": "
w/ \nLsimw/ \nLtssw/ \nLxmmAP on \nCharades (%) ↑mAR on \nCAD-120 ↑
61.10.75
✓--62.20.79
-✓-61.70.79
--✓62.90.81
✓-✓63.40.83
✓✓✓63.60.85
", + "image_path": "d464b0c32b2200f353dbaea924a2da810797c0ca0fa992c7ce305d607948a685.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 223, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 223, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 223, + 309 + ], + "type": "text", + "content": "4.2.2 Effectiveness of Learning Scheme" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "text", + "content": "To explore the effectiveness of the visual-semantic joint embedding " + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{sim})" + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "text", + "content": ", token selection supervision " + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{tss})" + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "text", + "content": " and cross-modal learning " + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{xm})" + }, + { + "bbox": [ + 46, + 317, + 288, + 520 + ], + "type": "text", + "content": " in our learning scheme, we conduct related ablation studies using different settings, i.e., cancelling one or any two or all our key modules. As reported in Table 2, the visual-semantic joint embedding, token selection supervision and cross-modal learning components collectively or individually contribute to the final performance improvement. This improvement confirms our hypothesis that language models can help video models to improve performance through knowledge transfer. It can be noted that due to the advantages of the learning scheme, the visual model improves from 61.1 to 63.6 in terms of mAP metric and from 0.75 to 0.85 in terms of mAR metric on Charade and CAD-120, respectively. These results demonstrate that the learning scheme plays an important role in our proposed LaIAR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 538, + 287, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 287, + 561 + ], + "type": "text", + "content": "4.2.3 Effectiveness of Robustness Against Domain Shift" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "The performance of RGB-based methods drops drastically when the training and testing data do not share the same distribution caused by change of scene, camera viewpoint or actor [40]. Our proposed model can adapt to video distribution shifts with the aid of semantic modality, which are invariant to complex appearance variations. To demonstrate the robustness of our proposed framework to domain shift, we split the Charades dataset into five subsets with non-overlapping training scenes and test scenes. Table 3 reports the average and variance of five accuracies for these five subsets. The variance of our method is clearly stable and indicates robustness to domain shift." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 309, + 95, + 438, + 135 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 435, + 93 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 435, + 93 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 435, + 93 + ], + "type": "text", + "content": "Table 3. Ablation study of the domain shift on the Charades dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 95, + 438, + 135 + ], + "lines": [ + { + "bbox": [ + 309, + 95, + 438, + 135 + ], + "spans": [ + { + "bbox": [ + 309, + 95, + 438, + 135 + ], + "type": "table", + "html": "
MethodAccuracy
AverageVariance
STIGPN [34]54.10.30
Ours57.20.11
", + "image_path": "ffe2d020e4bf2a7b34338d21ee4c11d9ae44f430ec33289007f036f6f7ab82be.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 442, + 104, + 535, + 136 + ], + "blocks": [ + { + "bbox": [ + 439, + 71, + 541, + 103 + ], + "lines": [ + { + "bbox": [ + 439, + 71, + 541, + 103 + ], + "spans": [ + { + "bbox": [ + 439, + 71, + 541, + 103 + ], + "type": "text", + "content": "Table 4. Comparison of accuracy using predicted and annotated relationships." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 442, + 104, + 535, + 136 + ], + "lines": [ + { + "bbox": [ + 442, + 104, + 535, + 136 + ], + "spans": [ + { + "bbox": [ + 442, + 104, + 535, + 136 + ], + "type": "table", + "html": "
Evaluation ModemAP
Prediction62.4
Label63.6
", + "image_path": "530293305811a44be96d3512bd0350fb155a9c9fea23690e9de6df105d217661.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 155, + 537, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 155, + 537, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 537, + 167 + ], + "type": "text", + "content": "4.2.4 Effectiveness of Using Predicted Relationships." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "text", + "content": "As previously stated, visual relationship categories can be manually annotated or identified by the visual relationship detection network [5]. To explore the impact of the two modes on accuracy, we compared the proposed method based on the ground truth and the predicted semantic relationships during training. The results, as detailed in Table 4, reveal that using predicted semantic relations can indeed enhance the accuracy of the video model on the Charades dataset (achieving " + }, + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "inline_equation", + "content": "62.4\\%" + }, + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "text", + "content": " mean Average Precision (mAP) versus " + }, + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "inline_equation", + "content": "61.1\\%" + }, + { + "bbox": [ + 304, + 174, + 545, + 341 + ], + "type": "text", + "content": " mAP shown in Table 2). Moreover, the accuracy does not significantly decrease when using predicted data instead of ground truth. This demonstrates the effectiveness of the proposed method in mining relational transformations from real video data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 348, + 493, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 493, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 493, + 361 + ], + "type": "text", + "content": "4.3. Comparison to the State-of-the-Art" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 713 + ], + "type": "text", + "content": "We compare the action recognition accuracy of the proposed method and the state-of-the-art methods (SoTA) on the Charades and the CAD-120 datasets, respectively. Table 5 summarizes the results on Charades. It can be seen that our proposed method outperforms several previous 3D CNN approaches in terms of mAP, including I3D [3], SlowFast [7] and LFB [36]. This demonstrates that our method can fully capture action cues through the visual relationship transitions, based on the human/objects information detected from a single video frame (rather than using the entire scene like I3D). STRG [35] and SGFB [11] model the action based objects and visual relationships, respectively, and overlook explicit modeling of temporal dynamics of the interaction between objects. Though VideoLN [12] and OR2G [24] takes visual relationship transitions into account, it is difficult for these methods to achieve accurate action inference due to the limitations of scene graph predictors at test time. For comparison in a modality with only RGB video frames, our method achieved the best performance compared with the existing methods. We also evaluated our method in Oracle evaluation mode, which leverages the ground-truth of bounding box and relationships on a frame. As reported in the last row of Table 5, our method still achieves best performance on the Charades dataset. It is important to mention that OR2G [24] used ground-truth scene graphs to enhance its final performance, whereas our network uses only the bounding boxes of humans and objects. Despite this, our method demonstrates strong performance in both evaluation modes, validating the effective" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18884" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 174, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 174, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 174, + 84 + ], + "type": "text", + "content": "ness of our proposed approach." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 129, + 282, + 284 + ], + "blocks": [ + { + "bbox": [ + 47, + 94, + 286, + 127 + ], + "lines": [ + { + "bbox": [ + 47, + 94, + 286, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 94, + 286, + 127 + ], + "type": "text", + "content": "Table 5. Multi-label action recognition performance comparison on the Charades's validation set in term of mAP. SG: ground truth scene graph. Bbox: Bounding Box. Higher values are better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 129, + 282, + 284 + ], + "lines": [ + { + "bbox": [ + 49, + 129, + 282, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 282, + 284 + ], + "type": "table", + "html": "
MethodsBackboneModalitymAP
I3D [3]R101-I3DRGB15.6
VideoMLN [12]R101RGB38.4
STRG [35]R101-I3D-NLRGB39.7
SlowFast [7]R101RGB42.1
LFB [36]R101-I3D-NLRGB42.5
SGFB [11]R101-I3D-NLRGB44.3
OR2G [24]R101-I3D-NLRGB44.9
OursR101-I3D-NLRGB45.1
SGFB Oracle [11]R101-I3D-NLRGB+SG60.3
VideoMLN Oracle [12]R101RGB+SG62.8
OR2G Oracle [24]R101RGB+SG63.3
Ours OracleR101RGB+Bbox63.6
", + "image_path": "5d203cc1335cb3522748942329bf6b88902e6e72e40f470c686715546faeb39c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 297, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 416 + ], + "type": "text", + "content": "For CAD-120 dataset, we follow the same experimental protocol as in [42] and divide the long video sequences into small segments based on individual sub-actions and evaluate the average recall metric for each sub-action. Table 6 summarizes the results on CAD-120. Explainable AAR-RAR [42] interpret the action reasoning process through the changes of relationship between objects or the attribute of objects across time. Our method is able to give the same explanation and outperforms these methods, achieving state-of-the-art performance with " + }, + { + "bbox": [ + 46, + 297, + 287, + 416 + ], + "type": "inline_equation", + "content": "0.85\\mathrm{mAR}" + }, + { + "bbox": [ + 46, + 297, + 287, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 59, + 437, + 273, + 540 + ], + "blocks": [ + { + "bbox": [ + 47, + 426, + 286, + 437 + ], + "lines": [ + { + "bbox": [ + 47, + 426, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 286, + 437 + ], + "type": "text", + "content": "Table 6. Experimental results on CAD-120 for action recognition." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 437, + 273, + 540 + ], + "lines": [ + { + "bbox": [ + 59, + 437, + 273, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 437, + 273, + 540 + ], + "type": "table", + "html": "
MethodsModalitymAR
Temporal Segment [33]RGB0.42
Flow0.71
RGB + Flow0.77
Explainable AAR-RAR [42]RGB0.80
VideoMLN [12]RGB0.83
OursRGB0.85
", + "image_path": "18f5a7d4c8563f8b720f9621319ca5a666fb21cd5decd328057e2d241387fff9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 562, + 227, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 227, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 227, + 574 + ], + "type": "text", + "content": "5. Interpretation and Visualization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "To intuitively demonstrate the interpretability effect of our proposed model, we provide interpretable representation and a visualization example. As shown in Figure 3, in the inference stage, we first extract the visual relation representations of human-object pairs in each frame. Then, our proposed DT-Former selects important relations in temporal and spatial dimensions and predicts action by modeling the important relation transition. Finally, the visual representations of important relations are mapped into the joint embedding space to find their nearest neighbor semantic labels, which can provide explicit evidence for the action" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "reasoning process. In this example, the relation representations between the person and the box in the second and tenth frames are selected as cues for the action recognition. The nearest semantic labels of these two representations in the joint embedding space are \"holding box\" and \"not holding box\", respectively. Here, the consequences of observations 'holding' " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " 'not holding' provide a clear sign of the action \"place\"." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 169, + 545, + 331 + ], + "blocks": [ + { + "bbox": [ + 306, + 169, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 306, + 169, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 169, + 545, + 331 + ], + "type": "image", + "image_path": "cb328bafcbf76717fd84a6636a58c004ace07a0b747546f3b9675369602b7074.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 338, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 381 + ], + "type": "text", + "content": "Figure 3. An example of action recognition performed by the proposed method and its corresponding process of providing explanations. The shaded visual relation representations indicate that it is not selected by DT-Former." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 378, + 400 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 408, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 647 + ], + "type": "text", + "content": "In this paper, we introduced a new framework, LaIAR, designed to transfer the knowledge from the language model to the video model to improve the recognition performance and interpretability of video models. Specifically, we build a language model and a video model, which take semantic relation and visual relation representations as input, respectively. These two models share the same architecture, namely DT-Former. This architecture is tailored to select the most important relations for action recognition from all the relations in video and to model the fine-grained relation transitions within videos. Our framework also incorporates three novel knowledge transfer strategies in our learning scheme to facilitate the knowledge transfer from the language model to the video model. This not only boosts the performance but also enhances the interpretability of the video model. Ablation experiments verified the effectiveness of the DT-Former, the learning scheme module and the robustness against domain shift. We conducted extensive experiments on Charades and CAD-120 datasets to demonstrate the superior performance of our proposed method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgements: This work is partially supported by the National Natural Science Foundation of China under Grant No.62073252 and No.62072358. It was also supported by Natural Science Basic Research Program of Shaanxi under Grant No.2024JC-JCQN-66." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18885" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 204 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 206, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 287, + 250 + ], + "type": "text", + "content": "[4] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 306 + ], + "type": "text", + "content": "[5] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16372-16382, 2021. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 308, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 287, + 373 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arxiv 2020. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 419 + ], + "type": "text", + "content": "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 452 + ], + "type": "text", + "content": "[8] David Gunning and David Aha. Darpa's explainable artificial intelligence (xai) program. AI magazine, 40(2):44-58, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 287, + 498 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 555 + ], + "type": "text", + "content": "[10] Hua Hua, Dongxu Li, Ruiqi Li, Peng Zhang, Jochen Renz, and Anthony Cohn. Towards explainable action recognition by salient qualitative spatial object relation chains. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5710-5718, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[11] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10236-10247, 2020. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[12] Yang Jin, Linchao Zhu, and Yadong Mu. Complex video action reasoning via learnable markov logic network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3242-3251, 2022. 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 659, + 287, + 713 + ], + "type": "text", + "content": "[13] Taotao Jing, Haifeng Xia, Renran Tian, Haoran Ding, Xiao Luo, Joshua Domeyer, Rini Sherony, and Zhengming Ding. Inaction: Interpretable action decision making for autonomous driving. In European Conference on Computer Vision, pages 370-387. Springer, 2022. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 128 + ], + "type": "text", + "content": "[14] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 784-794, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 171 + ], + "type": "text", + "content": "[15] Hema Swetha Koppula, Rudhir Gupta, and Ashutosh Saxena. Learning human activities and object affordances from rgb-d videos. The International journal of robotics research, 32(8):951-970, 2013. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 215 + ], + "type": "text", + "content": "[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 249 + ], + "type": "text", + "content": "[17] Christopher A Kurby and Jeffrey M Zacks. Segmentation in the perception and memory of events. Trends in cognitive sciences, 12(2):72-79, 2008. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 251, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 547, + 304 + ], + "type": "text", + "content": "[18] Pilhyeon Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, and Hyeran Byun. Decomposed cross-modal distillation for rgb-based temporal action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2373-2383, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 306, + 547, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 547, + 349 + ], + "type": "text", + "content": "[19] Yi Li and Nuno Vasconcelos. Improving video model transfer with dynamic representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19280-19291, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "text", + "content": "[20] Xue Lin, Qi Zou, Xixia Xu, Yaping Huang, and Ding Ding. Effects of motion-relevant knowledge from unlabeled video to human-object interaction detection. IEEE Transactions on Neural Networks and Learning Systems, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 394, + 547, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 547, + 448 + ], + "type": "text", + "content": "[21] Guan Luo, Shuang Yang, Guodong Tian, Chunfeng Yuan, Weiming Hu, and Stephen J Maybank. Learning human actions by combining global dynamics and local appearance. IEEE transactions on pattern analysis and machine intelligence, 36(12):2466-2482, 2014. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "type": "text", + "content": "[22] Lili Meng, Bo Zhao, Bo Chang, Gao Huang, Wei Sun, Frederick Tung, and Leonid Sigal. Interpretable spatio-temporal attention for video action recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 547, + 536 + ], + "type": "text", + "content": "[23] Ali Modarressi, Hosein Mohebbi, and Mohammad Taher Pilehvar. Adapler: Speeding up inference by adaptive length reduction. arXiv preprint arXiv:2203.08991, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 537, + 547, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 547, + 581 + ], + "type": "text", + "content": "[24] Yangjun Ou, Li Mi, and Zhenzhong Chen. Object-relation reasoning graph for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20133-20142, 2022. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 582, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 547, + 646 + ], + "type": "text", + "content": "[25] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 690 + ], + "type": "text", + "content": "[26] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "text", + "content": "[27] Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18886" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "models instead. Nature machine intelligence, 1(5):206-215, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 171 + ], + "type": "text", + "content": "[28] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 510-526. Springer, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[29] Tae Soo Kim and Austin Reiter. Interpretable 3d human action analysis with temporal convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 20-28, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 270 + ], + "type": "text", + "content": "[30] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE transactions on pattern analysis and machine intelligence, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 335 + ], + "type": "text", + "content": "[31] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1-9, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 381 + ], + "type": "text", + "content": "[32] Junke Wang, Xitong Yang, Hengduo Li, Li Liu, Zuxuan Wu, and Yu-Gang Jiang. Efficient video transformers with spatial-temporal token selection. In European Conference on Computer Vision, pages 69-86. Springer, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 435 + ], + "type": "text", + "content": "[33] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks for action recognition in videos. IEEE transactions on pattern analysis and machine intelligence, 41(11):2740-2755, 2018. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 502 + ], + "type": "text", + "content": "[34] Ning Wang, Guangming Zhu, Hongsheng Li, Mingtao Feng, Xia Zhao, Lan Ni, Peiyi Shen, Lin Mei, and Liang Zhang. Exploring spatio-temporal graph convolution for video-based human-object interaction recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[35] Xiaolong Wang and Abhinav Gupta. Videos as space-time region graphs. In Proceedings of the European conference on computer vision (ECCV), pages 399-417, 2018. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 591 + ], + "type": "text", + "content": "[36] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[37] Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[38] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10809-10818, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[39] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 272 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[40] Yunhua Zhang, Hazel Doughty, Ling Shao, and Cees GM Snoek. Audio-adaptive activity recognition across video domains. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13791-13800, 2022. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[41] Yuan Zhi, Zhan Tong, Limin Wang, and Gangshan Wu. Mgsampler: An explainable sampling strategy for video action recognition. In Proceedings of the IEEE/CVF International conference on Computer Vision, pages 1513-1522, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 272 + ], + "type": "text", + "content": "[42] Tao Zhuo, Zhiyong Cheng, Peng Zhang, Yongkang Wong, and Mohan Kankanhalli. Explainable video action reasoning via prior knowledge and state transitions. In Proceedings of the 27th acm international conference on multimedia, pages 521-529, 2019. 2, 6, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18887" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_content_list.json b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ebed40b70df45e02e3c04510459c9302733510b2 --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_content_list.json @@ -0,0 +1,1161 @@ +[ + { + "type": "text", + "text": "Language Models as Black-Box Optimizers for Vision-Language Models", + "text_level": 1, + "bbox": [ + 120, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shihong Liu* Samuel Yu* Zhiqiu Lin* Deepak Pathak Deva Ramanan Carnegie Mellon University", + "bbox": [ + 176, + 180, + 792, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities on downstream tasks when fine-tuned with minimal data. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. As such, we aim to develop a black-box approach to optimize VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or even output logits. We propose employing chat-based LLMs to search for the best text prompt for VLMs. Specifically, we adopt an automatic \"hill-climbing\" procedure that converges to an effective prompt by evaluating the performance of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot image classification setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of $1.5\\%$ across 11 datasets including ImageNet. Our approach also outperforms both human-engineered and LLM-generated prompts. We highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit \"gradient\" direction in textual feedback for a more efficient search. In addition, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different VLM architectures in a black-box manner. Lastly, we demonstrate our framework on a state-of-the-art black-box VLM (DALL-E 3) for text-to-image optimization.", + "bbox": [ + 75, + 300, + 473, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 768, + 209, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models [1, 28, 51, 64] (VLMs) excel at a wide range of classic vision and multimodal [2, 10, 15, 32, 70] tasks, surpassing the performance of their fully-supervised counterparts on downstream tasks even when fine-tuned with minimal data [34, 74]. However, fine-tuning VLMs typically requires transparent white-box access to the model weights, such as gradient-based approaches that rely", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on backpropagation.", + "bbox": [ + 500, + 270, + 637, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VLMs as black-box services. Despite community efforts to collect web-scale public datasets [55, 56] and to replicate proprietary VLMs [3, 24], an increasing number of models [1, 4, 13, 45, 64, 71] are not releasing their weights due to privacy and legal concerns [30, 38]. Therefore, one cannot use popular white-box fine-tuning strategies (such as LoRA [23] and Adapter [22]) that rely on model weights, feature embeddings, and output logits. Given that contemporary black-box VLMs [45, 47] like DALL-E [4, 52] still offer a language-based user interface and may be accessed through APIs that facilitate input and output in natural language, this allows users to customize these models through optimizing textual prompts.", + "bbox": [ + 496, + 287, + 895, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Manual prompting. Manual prompt engineering has been proven successful in adapting black-box LLMs to language tasks [25, 66]. Similarly, carefully crafted prompts can enhance the performance of VLMs. For instance, CLIP has demonstrated improved zero-shot recognition performance using specifically tailored prompts, such as \"a photo of a {class}\" for Internet photos and \"a satellite image of a {class}\" for satellite imagery. Despite its effectiveness, manual prompting can be a laborious process, inspiring efforts to explore automated prompt creation and thereby remove the need for human involvement. These strategies typically leverage an LLM as a knowledge base to create rich visual descriptors that augment the prompts for each class [40, 50] in a zero-shot fashion.", + "bbox": [ + 496, + 487, + 895, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-free prompting with conversational LLMs (our approach). We show how to effectively leverage chat-based LLMs [45] to emulate human-level prompt engineering without any human input. We first address an illustrative low-shot image classification task, aiming to find the best class-agnostic prompt (or \"template\") for image classification with CLIP. We start with a random set of prompts and evaluate the one-shot training accuracy of each. Then, akin to human prompt engineering, our method repeatedly presents ChatGPT with the best and worst prompts, asking it to review the results and suggest an improvement (see Figure 1).", + "bbox": [ + 496, + 700, + 895, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning with implicit \"gradients\" provided through conversational feedback. One of our key findings is that", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12687", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc43833313de44e59df331a20d486ab4abda1512f0f8455752468f4af3832194.jpg", + "image_caption": [ + "Figure 1. Prompting VLMs using chat-based LLMs. Similar to how human prompt engineers iteratively test and refine prompts, we employ ChatGPT [45, 47] to continuously optimize prompts for vision-language models (VLMs). Our iterative approach assesses the performance of ChatGPT-generated prompts on a few-shot dataset (highlighted in blue) and provides feedback (marked in violet) to ChatGPT through simple conversations, as depicted in the illustrative figure. This straightforward method delivers state-of-the-art results for one-shot image classification across 11 datasets using CLIP, operated in a black-box manner without accessing model weights, feature embeddings, or output logits. We show that providing both positive (in green) and negative prompts (in red) enhances efficiency. Remarkably, our approach outperforms both white-box methods such as gradient-based continuous prompting (CoOp [74]) and human-engineered prompts [51] in this extremely low-shot scenario. This figure only shows a typical conversation using ChatGPT's web user interface. Our code implementation follows this pattern using the ChatGPT API. We detail and ablate the prompts in section 8." + ], + "image_footnote": [], + "bbox": [ + 160, + 87, + 816, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLMs can learn the difference between effective and ineffective prompts, and can use this implicit \"gradient\" direction provided through language to perform more efficient searches. Compared to previous automatic prompting methods that only use LLMs as a knowledge base [40, 50] or paraphrasing tool [75], we show a novel use of LLMs as an optimizer that can utilize the patterns hidden in textual feedback. In our experiments, we find that the inclusion of such feedback greatly improves the efficiency and accuracy of our method, sometimes surpassing existing white-box methods [67, 74] on challenging one-shot scenarios.", + "bbox": [ + 75, + 577, + 473, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimizing text-to-image generation with DALL-E 3. We further demonstrate our optimization framework on a state-of-the-art black-box VLM, DALL-E [4], for two illustrative one-shot generative tasks: (1) Text-to-image (T2I) generation (see Figure 3), where we sample challenging text queries from Winoground [63] that involve reasoning over compositions of objects, attributes, and relations. Examples include \"an animal watches a person\" and \"there is less milk than orange juice\", which DALL-E 3 might initially", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fail to generate. (2) Prompt inversion (see Figure 4), which attempts to reverse-engineer the textual prompt to generate a specific image for later customization [53] (see Table 5). To achieve this, we leverage conversational feedback from a multimodal LLM (GPT4-V [45]) to iteratively refine the prompts based on the current generated images. We present qualitative results in Table 4 and conduct a user study to demonstrate that our framework can be more efficient than manual prompting, even for graphical designers experienced with AI content-generation tools.", + "bbox": [ + 496, + 577, + 893, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions. In this work, we introduce a novel prompting method for VLMs, utilizing an LLM as an optimizer. Our black-box approach can surprisingly compete with various white-box methods in a low-shot setting. Additionally, we extensively explore various strategies for conversing with ChatGPT, uncovering several key factors that significantly enhance the efficiency of this tool. We also show that our discovered natural language prompts are not only interpretable but also transfer better across CLIP architectures, eg., from RN50 to ViT/B-16, than continuous prompts discovered by previous white-box prompting", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12688", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method [74]. Finally, we show practical applications of our framework on text-to-image generation using black-box DALL-E 3. We release our code for future research on prompt optimization and AI-driven content creation1.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 171, + 225, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs for multimodal tasks. Cutting-edge LLMs like GPTs [45, 47] have been successfully applied to multimodal tasks, either through zero-shot composition with pre-trained multimodal models [29, 72] or by jointly fine-tuning with modality-specific encoders [1, 28] on large-scale multimodal datasets [56]. LLMs are also utilized as neuro-symbolic reasoners [16, 37, 58, 73], translating natural language instructions into modular programs (like Python code) that invoke APIs of multimodal models. In this work, we show the potential of LLMs as a black-box optimizer for multimodal foundation models with language interfaces, and more specifically vision-language models (VLMs).", + "bbox": [ + 75, + 199, + 470, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt optimization of foundation models. Following the success of in-context learning [6], which appends user-generated natural language instruction and few-shot samples to text inputs, prompting [35] has emerged as the preferred fine-tuning paradigm for LLMs due to its superior performance and parameter-efficiency. However, recent prompt optimization methods, including continuous prefix-tuning [7, 31, 61, 62, 69] and discrete token-searching [11, 12, 59], still operate in a white-box manner, requiring access to either the tokenizer or output logits. Moreover, black-box prompting methods, such as heuristic-based editing [41, 49], are tailored towards language-only tasks and are thus not applicable in VLM settings.", + "bbox": [ + 75, + 383, + 470, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs for prompt optimization. APE [75] leverages an LLM to automatically write prompts using few-shot samples based on instruction induction [21] and paraphrasing [42, 54]. However, it is only designed to address language tasks, while we focus on multimodal tasks using black-box VLMs. LLMs have also proven to be an effective external knowledge base [40, 50, 57] for generating prompts in a zero-shot setting for multimodal models. For example, DCLIP [40] uses GPT3 to come up with rich visual descriptions to improve zero-shot classification with CLIP [51]. We extend this line of work to show that LLMs can iteratively optimize prompts for VLMs in a black-box fashion given few-shot samples. We further illustrate that prompt optimization with LLMs can be made more efficient by leveraging conversational feedback, such as providing ChatGPT with explicit language feedback on how well the most recent prompt performs. Our findings align with the perspective [9] of LLMs as meta-optimizers that can implicitly perform gradient search through in-context learning.", + "bbox": [ + 75, + 580, + 470, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot adaptation of VLMs. Prompting has also been successfully adopted in VLMs [14], as demonstrated by methods like CoOp [74] that fine-tune an ensemble of continuous prefix tokens using cross-entropy loss. [34] achieves state-of-the-art few-shot performance with a cross-modal (image and text) cross-entropy loss. However, these methods all require access to model parameters for gradient backpropagation. We also note that while some concurrent works, such as BlackVIP [44] and LFA [46], claim to operate in a \"black-box\" setting, they still require access to privileged information including output logits and embeddings. In this work, we introduce a truly black-box and gradient-free approach that yields competitive results to white-box approaches in extremely low-shot scenarios.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Prompting VLMs Using Chat-Based LLMs", + "text_level": 1, + "bbox": [ + 498, + 316, + 883, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now present our approach for prompting VLMs using chat-based LLMs as optimizers.", + "bbox": [ + 496, + 340, + 890, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminaries. Motivated by recent proprietary VLMs [4, 45], we adopt a stricter yet practical black-box setting compared to prior works [44, 46], requiring minimal knowledge about the model's inner workings. This is crucial since releasing output logits or embeddings can potentially facilitate unauthorized knowledge extraction through distillation methods [19]. Our objective is to enhance the performance of a VLM equipped with a language interface capable of processing a textual prompt $p \\in T$ . We assume that the targeted task is accompanied by a training dataset denoted as $D_{train} \\subset D$ and its performance can be evaluated with respect to the prompt, represented as a function $F: D \\times T \\to \\mathbb{R}$ . For example, in a classification task, $D_{train} = \\{x, y\\}_n$ where $x$ is an image and $y$ is its class label. The black-box VLM takes the image as input and returns a predicted label. We measure the performance of the textual prompt by calculating the average classification accuracy as $F(D_{train}, p)$ . Our goal in prompt engineering is to search for the optimal prompt $p^*$ without accessing or modifying the black-box VLM.", + "bbox": [ + 496, + 371, + 893, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background: human prompt engineering. Our method draws inspiration from the typical workflow of human prompt engineers. Prompt engineering is often an iterative process that involves: (a) creating an initial prompt $U = \\{p_1\\}$ based on the understanding of a task, (b) evaluating the performance of prompts in $U$ , (c) refining prompts based on the outcomes, (d) repeating the last two steps until convergence, and (e) returning the prompt $p^*$ with the highest $F(D_{train}, p^*)$ . This hands-on approach helps optimize the model's performance, but it can be tedious and labor-intensive. Algorithm 1 formally illustrates this process.", + "bbox": [ + 496, + 659, + 893, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Example: prompting for image classification with CLIP [51]. CLIP is one of the most popular VLM that takes a set of class-specific prompts when performing \"zero-shot\" image classification. [51] details the laborious prompting procedure over the course of a year. Interestingly, they find", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1Project site: llm-can-optimize-vlm.github.io", + "bbox": [ + 94, + 886, + 408, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12689", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 We formalize human prompt engineering with the following algorithm, which motivates our LLM-based algorithm (2)." + ], + "code_body": "Require: $D_{\\mathrm{train}} = \\{x,y\\}_{n}$ training samples, $F:D\\times T\\to$ R: evaluation function \n1: Create initial prompts: $\\mathcal{U}\\gets \\{p_1\\}$ \n2: Evaluate prompts on training set: $S\\gets \\{F(D_{\\mathrm{train}},p_1)\\}$ \n3: while not converged do \n4: Generate a new prompt $p^\\prime$ based on $S$ \n5: Evaluate the new prompt: $s^{\\prime} = F(D_{\\mathrm{train}},p^{\\prime})$ \n6: $\\mathcal{U}\\leftarrow \\mathcal{U}\\cup \\{p^{\\prime}\\}$ \n7: $S\\gets S\\cup \\{s^{\\prime}\\}$ \n8: end while \n9: return optimal prompt $p^*\\gets \\arg \\max_{p\\in \\mathcal{U}}F(D_{\\mathrm{train}},p)$", + "bbox": [ + 76, + 140, + 470, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that a default class-agnostic prompt (or so-called \"template\"), \"a photo of a {class}\" can provide a decent boost in accuracy for most datasets compared to using vanilla class labels. In this scenario, the evaluation function $F$ is the classification accuracy on the test set, and the prompt $p = \\{ \\text{\"a photo of a {class}\" } | c \\in C \\}$ , where $C$ is the set of class names for a given dataset.", + "bbox": [ + 75, + 345, + 472, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompting with chat-based LLMs (our approach). Given the strong in-context reasoning capabilities of LLMs, we envision them as a black-box optimizers that can improve prompts based on their performance outcomes, akin to how human prompt engineers iteratively refine prompts. Specifically, we maintain a pool of prompts $U$ and their corresponding performance outcomes $S$ . In each iteration, we provide the LLM with both positive and negative prompts, such as the highest and lowest-performing candidates. Such textual feedback through in-context prompts offers LLMs an implied \"gradient\" direction [9], making optimization more efficient than taking random local steps. We facilitate this feedback mechanism through conversations with state-of-the-art chat-based LLMs like ChatGPT [47] as illustrated in Figure 1. We note that such a multi-turn conversation is not the only way of conversing with ChatGPT, and ablate different in-context feedback mechanisms in section 8.", + "bbox": [ + 75, + 455, + 472, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Illustrative Few-Shot Classification Task", + "text_level": 1, + "bbox": [ + 76, + 734, + 439, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We illustrate our approach using a few-shot image classification task. Specifically, a prompt $p \\in T$ consists of a set of class-specific prompts – that is, one textual description per class. The evaluation function $F$ takes the prompt $p$ , along with an image dataset $D_{train}$ , and returns the accuracy using the black-box VLM. To prevent overfitting and simplify our search space, we restrict our search to finding a single class-agnostic template, e.g., a photo of a {}, filling in the blank with label names provided with the dataset.", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 2 LLM-based prompt engineering on the illustrative classification task. Our algorithm requires a chat-based LLM and a (black-box) evaluation function, such as accuracy. We highlight mechanisms for “exploration” (restart and reset) in blue and “exploitation” (iter) in red. We mark the key component of “conversational feedback” of our approach in violet. The actual prompts are attached in section 8.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Require: $D_{\\mathrm{train}} = \\{x,y\\}_{n}$ training samples, $F:D\\times T\\to$ R: evaluation function. \nRequire: $n_{\\mathrm{restart}}$ : number of initial sampled prompt sets, $n_{\\mathrm{reset}}$ : number of resets for a prompt set, $n_{\\mathrm{iter}}$ : number of hill-climbing iterations, $m$ size of one initial prompt set, $k$ : number of prompts send to ChatGPT. \n1: $p^* \\gets \\emptyset$ \n2: for 1:: $n_{\\mathrm{restart}}$ do \n3: Sample a new prompt set, $\\mathcal{U}_{\\mathrm{init}} \\leftarrow \\{p_1,\\dots,p_m\\}$ \n4: for 1:: $n_{\\mathrm{reset}}$ do \n5: Reset to initial prompt set: $\\mathcal{U} \\leftarrow \\mathcal{U}_{\\mathrm{init}}$ \n6: for 1:: $n_{\\mathrm{iter}}$ do \n7: Sort $\\mathcal{U}$ by score outcomes $\\{F(D_{\\mathrm{train}},p)\\}_{p\\in U}$ \n8: $\\mathcal{U}_{\\mathrm{top}} \\leftarrow$ top-k prompts in $\\mathcal{U}$ \n9: $\\mathcal{U}_{\\mathrm{bot}} \\leftarrow$ bottom-k prompts in $\\mathcal{U}$ \n10: Get a new prompt $p_{\\mathrm{new}} \\leftarrow$ LLM( $\\mathcal{U}_{\\mathrm{top}},\\mathcal{U}_{\\mathrm{bot}}$ ) \n11: $\\mathcal{U} \\leftarrow \\mathcal{U} \\cup \\{p_{\\mathrm{new}}\\}$ \n12: end for \n13: $p^* \\leftarrow$ arg max $p\\in \\mathcal{U}\\cup \\{p^*\\}$ $F(D_{\\mathrm{train}},p)$ \n14: end for \n15: end for \n16: return prompt with highest score $p^*$", + "bbox": [ + 500, + 200, + 893, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outline of our approach (Alg. 2). To start, we sample entirely random initial prompts from a text corpus such as LAION-COCO [55] captions. Our approach follows the classical stochastic hill-climbing framework with random restart [54], which prevents ChatGPT from being trapped in local optima by balancing \"exploration\" and \"exploitation\". Our restart mechanism is implemented by sampling $n_{\\mathrm{restart}}$ initial prompt sets to encourage exploration. Because ChatGPT performs stochastic top-k sampling for text generation (as we adopt the default temperature of 1.0), we also implement a reset mechanism to foster additional exploration by retrying a given prompt set $n_{\\mathrm{reset}}$ times. For exploitation, we converse with ChatGPT for $n_{\\mathrm{iter}}$ iterations. We find that it is critical to balance exploration and exploitation for optimal performance, and thoroughly examine this trade-off in section 9. Lastly, we present ChatGPT both the top and bottom-performing prompts, denoted as $(\\mathcal{U}_{top}, \\mathcal{U}_{bot})$ . We show that this simple adjustment can improve the efficiency of our approach in Figure 2.", + "bbox": [ + 496, + 566, + 893, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experimental setup. We apply our approach to the few-shot image classification benchmark introduced in CoOp [74], which is the most commonly studied setup for", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12690", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "fine-tuning VLMs. This benchmark involves a collection of 11 datasets covering diverse image domains including ImageNet [10] and more niche datasets such as FGVC-Aircraft [39]. For each dataset, we adhere to the same three-fold k-shot train sets in [34], reporting the average accuracy across all folds. Importantly, our method only utilizes the train set to compute the score and does not require the few-shot validation set. We use CLIP following prior work [34, 74] to emulate a black-box VLM, and we employ ChatGPT (GPT3.5) as the chat-based LLM.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. To start, we sample entirely random 1M initial prompts from a text corpus (LAION-COCO [55] captions). For each caption, we extract all the noun phrases using spaCy part-of-speech tagging [20]. Subsequently, we replace one noun phrase in the caption with $\\{\\}$ (a placeholder where the class name will be inserted) to create a template. Given that each caption contains an average of 2 noun phrases, our initial prompt pool consists of approximately 2M templates. We run our algorithm with $n_{\\mathrm{start}} = 20$ restarts, $n_{\\mathrm{resets}} = 50$ resets, and $n_{\\mathrm{iter}} = 10$ iterations. We opt to sample $m = 100$ prompts per restart and present the top and bottom $k = 15$ prompts to ChatGPT. We ablate different sets of hyperparameters and explain how we balance the tradeoff between exploration and exploitation in section 9. We adopt gpt-3.5-turbo-0301 model for ChatGPT using OpenAI's official API and keep the default sampling temperature of 1.0. We also ablate gpt-4 in Table 10 and find it achieves similar performance. The exact prompts used to converse with ChatGPT are documented in section 8. For a fair comparison, we use CLIP-RN50 for our experiments following prior work [34, 74]. We will open-source our code and release the initial prompt pool (LAIONCOCO-1M) to the public.", + "bbox": [ + 75, + 246, + 472, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Oracle white-box baselines. Our black-box setup substantially differs from, and is more constrained than, the scenarios considered in previous white-box baselines. Specifically, we do not expose the pre-trained weights, model architectures, feature embeddings, or even output logits of VLMs. These constraints render many established gradient-based fine-tuning baselines inapplicable. Among the oracle white-box approaches we later compare to, CoOp [74] performs continuous prompting and requires backpropagation across all layers. WiSE-FT [67] ensembles fine-tuned weights with the original CLIP weights. Cross-Modal Adaptation [34] fine-tunes a linear classifier leveraging both image and text embeddings from CLIP. BlackVIP [44] and LFA [46] are two most recent baselines that apply CLIP logits or embeddings for gradient back-propagation. Finally, while DCLIP [40] queries GPT3 for rich visual descriptors for each class and does not require gradient-based fine-tuning, it performs prompt assembling using 4-6 class-specific prompts, which breaches our black-box assumption for accessing the output logits.", + "bbox": [ + 75, + 598, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Black-box methods. We additionally benchmark our method against truly black-box solutions, including the vanilla class-agnostic templates “{class)” and “a photo of a {class}”. Also, we compare our approach to the best Hand-Engineered templates released by OpenAI, searched using test set performance to represent the theoretical upper bound of human performance, eg., “a centered satellite photo of {class}.” for EuroSAT [17]. Finally, we present two versions of conversational feedback of our approach: (a) using 30 positive (P only) or (b) using 15 positive and 15 negative prompts $(\\mathbf{P} + \\mathbf{N})$ in each iteration. For a fair comparison, both of our approaches start with the same initial sampled prompts, referred to as LAIONCOCO-1M. We also show the performance of the best initial sampled prompt searched using trainset performance.", + "bbox": [ + 496, + 90, + 893, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SOTA one-shot performance against existing methods on 11 datasets. We report the test set performance of our method versus the aforementioned baselines in a challenging 1-shot classification scenario in Table 1. First, compared to the top-performing initial prompts selected from LAIONCOCO-1M based on train set performance, our prompt optimization using ChatGPT notably improves the initial prompts by an average of $5\\%$ (56% to 61%). Remarkably, our black-box approach surpasses the two white-box gradient-based fine-tuning techniques CoOp and WiSE-FT by at least $1.5\\%$ . Given that both CoOp and our method optimize a single class-agnostic template, we attribute this gap in performance to reduced overfitting. More specifically, we posit that our optimization space of natural language effectively acts as a regularizer in extremely low-shot tasks, standing as a more robust alternative to the continuous prompting approach of CoOp. Furthermore, our method benefits from textual feedback and shows improved performance by $1.0\\%$ when using both positive and negative prompts. In section 9, we show that our approach remains effective across different CLIP and ChatGPT variants.", + "bbox": [ + 496, + 332, + 895, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Incorporating negative prompts leads to more efficient optimization. In Figure 2, we demonstrate that incorporating both positive and negative prompts fosters better optimization efficiency, achieving higher accuracy within a much fewer number of resets. Specifically, we hypothesize that LLMs can leverage the implicit \"gradient\" direction suggested in textual feedback to achieve faster convergence. For additional analysis, we ablate different ways of providing conversational feedback to ChatGPT in section 8 and conclude that iteratively updating both positive and negative prompts is the key for efficient optimization.", + "bbox": [ + 496, + 650, + 895, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. More Benefits of Natural Language Prompts", + "text_level": 1, + "bbox": [ + 500, + 830, + 890, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we delve deeper into the advantages of utilizing natural language prompts compared to the continuous prompts [74]. We highlight that the prompts derived through", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12691", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e4ace9614662bebfcb80cc23f784cd43c78f6fcfe26c7e963608a3affdcb1f89.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDatasetAvg
CaltechImageNetAircraftFoodPetsCarsSUNUCFDTDEuroSATFlowers
Oracle white-box approaches
Cross-Modal [34]89.161.620.677.185.759.063.464.749.961.876.364.7
WiSE-FT [67]85.558.318.671.981.755.756.659.444.252.365.859.1
CoOp [74]87.557.29.674.385.955.660.361.944.450.668.159.6
LFA [46]81.652.417.063.175.341.458.456.738.460.774.956.4
BlackVIP [44]85.858.815.376.785.256.457.058.840.130.061.156.8
DCLIP [40]-59.6-76.483.8---41.734.7--
Manual prompting approaches
{}78.555.315.574.078.952.253.455.541.432.157.354.0
a photo of a {}84.557.915.974.083.253.958.056.938.828.660.255.6
Hand-Engineered [51]86.358.217.377.385.855.658.561.542.337.666.158.8
Our black-box approaches
LAIONCOCO-1M81.456.217.476.579.651.354.955.843.138.661.356.0
Ours (P only)89.059.417.977.885.755.760.458.743.646.766.660.1
Ours (P+N)89.159.618.178.388.156.261.060.244.849.067.261.1
", + "bbox": [ + 228, + 88, + 743, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of our method with other baselines on one-shot classification tasks. We report the average accuracy of each method across three folds, optimized using 1-shot training sets. We bold the best black-box result for each dataset, and underline the second best result. First, we note that our approach can effectively improve upon the initial prompts selected from LAIONCOCO-1M from $56\\%$ to $61\\%$ . Our approach is also competitive against the best Human-Engineered prompts released by OpenAI [51] searched using test set performance. Additionally, we show that using both positive and negative prompts improves the overall accuracy by $1\\%$ . For reference, we report oracle white-box approaches in gray. Remarkably, we also surpass white-box solutions such as WiSE-FT [67] and CoOp [74] by $1.5\\%$ . These methods require either gradient-based fine-tuning (CoOp/WiSE-FT/Cross-Modal) or prompt ensembling using output logits (DCLIP). While our approach is less effective than the SOTA white-box method (Cross-Modal Adaptation), we stress that our black-box setup is significantly more challenging, because we restrict the optimization space to natural language and do not access the pre-trained weights, model architectures, feature embeddings, and output logits of VLMs.", + "bbox": [ + 75, + 306, + 893, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/984ce6c2fc4524419f5608bf8d5dcb04a71044111e8c5f49adc33ff86558f74f.jpg", + "image_caption": [ + "Figure 2. Conversational feedback incorporating both positive and negative prompts leads to improved efficiency. We fix the number of restarts to 20 and iterations to 10, and ablate different numbers of resets on all 11 datasets (left) and ImageNet (right). Notably, our approach using “ $\\mathrm{P} + \\mathrm{N}$ ” (both top-15 and bottom-15 prompts) can optimize faster within a much fewer number of resets than using “P-Only” (top-30 prompts), resulting in the highest overall performance." + ], + "image_footnote": [], + "bbox": [ + 81, + 477, + 256, + 584 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d848393515ce207d267624f6e16fe8f7a8264c37a7dc772585090ac716b31b14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 477, + 450, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our method are interpretable; for instance, they often contain descriptions of the targeted image domain. Our prompts can also transfer across CLIP architectures in a black-box manner, such as from RN50 to ViT/B-16.", + "bbox": [ + 75, + 729, + 470, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interpretable natural language prompts. While CoOp [74] concedes that continuous prompts can be difficult to interpret, our method – without explicitly instructing ChatGPT to generate interpretation – often yields interpretable results. Table 2 showcases the templates returned by our algorithm for each dataset, frequently including keywords that reflect the targeted image domain. For example, the template", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for Food101 [5] mentions \"diverse cuisine and ingredients\", and the template for UCF101 [60] (an action recognition dataset) mentions \"in motion\". Likewise, these templates identify general stylistic attributes of the datasets; they refer to \"bright and natural lighting\" for ImageNet [10] and note images that \"emphasize the subject\" for Caltech101 [27]. These prompts are particularly intriguing because we do not provide ChatGPT with any information about the downstream task, yet it manages to generate prompts containing domain-specific keywords that are similar to those engineered by human experts.", + "bbox": [ + 496, + 473, + 893, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Black-box prompt transfer. Our text prompts also maintain consistently high performance across different CLIP backbones. For comparison, since CoOp uses the same tokenizer for all CLIP architectures (including RN50, RN101, ViT/B-32, and ViT/B-16) and optimizes continuous prompts of the same shape (16 x 512), we assess the transferability of these learned continuous prompts from RN50 to other backbones using the official weights on 16-shot ImageNet. Table 3 showcases the results of this experiment, where we also include the baseline prompt \"a photo of a {}\" for reference. We observe a significant decline in accuracy when transferring CoOp's prompts (up to a $40\\%$ decrease despite utilizing more powerful backbones), implying that continuous prompts tend to overfit to the specific CLIP model. In contrast, our natural language prompts maintain their performance and outperform the baseline across all backbones.", + "bbox": [ + 496, + 659, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12692", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea6d9f0e7b1ec8586e72976cdc91803d2b56b30100c76065ea76f4cff64084fa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetExample of Top Templates
Caltech [27]An image of a {} with a blurred background that emphasizes the subject
DTD [8]The essential elements of {} are amplified with visual simplicity
EuroSAT [17]A top-down view of {} arranged in a pattern {}
Aircraft [39]A clear, high-quality image of a single {} with a white background
Food [5]A {} featuring diverse cuisine and ingredients
ImageNet [10]An image of a {} with bright and natural lighting
Flowers [43]A clear and vivid photograph of the {} in its natural setting
Pets [48]A {} with distinct and recognizable characteristics
Cars [26]A {} featuring a wide range of color options for easy selection
SUN [68]A high-resolution photo of a {} with clear background and natural lighting
UCF [60]A black and white photo of a {} in motion
", + "bbox": [ + 248, + 89, + 722, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f57995e32de08a6207cd280e91777439d334a8db733fb1291b2950e6fef7a02.jpg", + "table_caption": [ + "Table 2. Example templates returned by our algorithm on each dataset. Although we do not provide ChatGPT with any information regarding the targeted dataset, we observe that the resulting templates are remarkably similar to human-engineered templates, with many domain-specific details such as \"motion\" and \"cuisine\", and stylistic elements such as \"bright and natural lighting\"." + ], + "table_footnote": [], + "table_body": "
MethodRN50→RN101→ViT-B/32→ViT-B/16
a photo of a {}57.960.661.966.6
CoOp63.020.631.739.5
Ours59.960.762.267.0
", + "bbox": [ + 96, + 275, + 450, + 342 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Black-box prompt transfer from ResNet-50 to other CLIP architectures. We evaluate both our natural language prompts and CoOp's continuous prompts on 16-shot ImageNet, which are trained using the RN50 CLIP backbone. As a reference point, we include the baseline prompt \"a photo of a $\\{\\}$ \", and show that the prompts derived from our method using RN50 consistently surpass it after transferring to different backbones. In contrast, while CoOp achieves better 16-shot ImageNet performance using RN50, its performance plummets during the transfer, e.g., from $63\\%$ to a mere $21\\%$ for RN101.", + "bbox": [ + 75, + 354, + 472, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Application: Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 76, + 503, + 429, + 522 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a25d3c154e797990230cc64699ff124918931546f0a783046bca6c20a42ffac.jpg", + "image_caption": [ + "Figure 3. Improving text-to-image (T2I) generation using chat-based multimodal LLMs. Our framework can use the multimodal GPT4-V [45] to optimize prompts for the state-of-the-art black-box generative VLM, DALL-E 3 [4]. For complex user queries that DALL-E 3 may initially fail to generate, we send the generated image (in violet) along with the current prompt to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue). Our simple framework is surprisingly effective at correcting DALL-E 3 mistakes on some challenging Winoground [63] text queries that involve action, logical, and spatial reasoning. We conduct a human evaluation on the quality of generated images in Table 6 and include the actual prompts in section 8." + ], + "image_footnote": [], + "bbox": [ + 81, + 541, + 477, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present a direct application of our prompt optimization framework to generative tasks using a truly black-box text-to-image (T2I) VLM, DALL-E 3 [4].", + "bbox": [ + 496, + 277, + 890, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimizing T2I using a multimodal LLM. DALL-E 3 can generate high-fidelity images following diverse user queries, but crafting effective prompts is tricky even for designers experienced with AI content generation tools [36]. Therefore, we are motivated to implement our LLM-based optimization framework to assist with creative visual design. Our framework is shown in Figure 3 for the illustrative task of text-to-image generation. In this task, the user specifies a query (topic) in text, such as \"an animal watches a person\", and the goal is to write a prompt that can generate an image reflecting this topic. We adopt a multimodal LLM GPT4-V [45] (gpt-4-1106-preview) to provide feedback on the generated image and optimize the prompt. We find that this framework is surprisingly effective due to GPT4-V's strong visual reasoning capabilities, which can often spot subtle errors in generated images and offer more accurate prompts.", + "bbox": [ + 496, + 324, + 893, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2afe40db62a631c711c0caf5cb07baf88a646951f2afcb4573f046aaed48770d.jpg", + "image_caption": [ + "Figure 4. Prompt inversion using chat-based multimodal LLMs. We apply our framework to reverse engineer the text prompt to generate the same user-queried image. We send the generated image (in violet) along with the original image to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue)." + ], + "image_footnote": [], + "bbox": [ + 501, + 597, + 901, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Task setup. For T2I generation, we experiment with a", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12693", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bd80818b456d5d614873a091fd61a92fa9e36a3893eaabc4d597f51f5c01e13.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
User QueryInit. ImageLLM FeedbackFinal Image
Text-to-image generation
There is less milk than orange juice.Incorrect, the milk bottle appears full, more than orange juice...
A shorter person is covering the eyes of a taller person.Incorrect, the taller person is covering the shorter person's eyes. Instead, ...
Prompt inversion
The scarf should feature red and white stripes, and the fur is fluffy...
The coat should be buttoned and the lighting exhibits a stronger contrast...
", + "bbox": [ + 102, + 88, + 444, + 313 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/db73f4ca7cac10f0b776e0138004c46b18324d6a9bc3cfbc080f1b9841820588.jpg", + "table_caption": [ + "Table 4. Examples of T2I optimization. We show that our framework (Figure 3) can automatically improve the faithfulness of images generated by DALL-E 3, with respect to user-specified textual topics (for T2I generation) or reference images (for prompt inversion). This is achieved through three rounds of prompt optimization, using feedback from the multimodal LLM (GPT4-V). Table 11 and Table 12 shows more examples with actual prompts." + ], + "table_footnote": [], + "table_body": "
User QueryInverted ImageExample 1Example 2Example 3Example 4Example 5
Give the dog a cat friend.Make the dog be in the middle of a jump.Make the dog do a handstand.Make the dog lie down on its side.Make the dog swim in water.
Make the owl fight a hawk.Make the owl flap its wings.Make the owl fully green.Make the owl stand in front of the moon.Make the owl walk in the city.
", + "bbox": [ + 81, + 425, + 478, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "subset of 100 text queries from Winoground [63] that involve complex attribute and relation reasoning, which DALL-E might initially fail to generate. Our framework refines the prompts to capture the user-specified topics using a few (three) iterations. We also attempt a reverse task of prompt inversion: given a user-specified reference (query) image, our framework reverse-engineers the prompt to have DALL-E generate the same object or scene in the query image (see Figure 4). This enables users to easily make customizations [53] (see Table 5), such as having the character in a reference image perform various actions or change scenes. For this task, we sample 100 random queries from DiffusionDB [65]. We provide qualitative results in Table 4, Table 11, and Table 12. We hire two volunteers to assess the faithfulness of the images generated by our method, and to compare these with the images manually prompted by two", + "bbox": [ + 75, + 657, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "designers (each with one year of experience in AI content generation), as shown in Table 6.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarks on limitations. While we show promising results, we note some failure cases in Table 13 and Table 14 due to the inherent limitations of foundation models. For example, GPT4-V might fail to describe abstract and artistic details, and DALL-E 3 often fails to generate the correct number of objects. We believe that our framework can benefit from more capable foundation models in the future.", + "bbox": [ + 498, + 121, + 893, + 226 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66305845930b1860ae78e1a5685d5837a93908c652e1dc2a53d7d62c597bf344.jpg", + "table_caption": [ + "Table 5. Customization via prompt inversion. Users can simply append extra descriptions to the inverted prompts to customize their main characters in queried images." + ], + "table_footnote": [], + "table_body": "
TaskMethodInit. (std)Final (std)Δ
Text-to-ImageHuman2.28 (.45)2.86 (.61)0.58
Ours2.62 (.36)3.56 (.54)0.94
Prompt InversionHuman1.58 (.48)2.76 (.53)1.18
Ours1.94 (.39)3.68 (.47)1.74
", + "bbox": [ + 545, + 238, + 846, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Our method enhances faithfulness in T2I generation. We hire two human annotators to assess the faithfulness of images generated from user queries, e.g., textual topics for Text-to-Image, or reference images for Prompt Inversion. The scores are measured on a 1-to-5 Likert scale, with 1 signifying contradiction and 5 indicating perfect alignment with the user's goal. Our approach benefits from three iterations of prompt optimization and consistently outperforms human-engineered prompts by designers who have one year of experience in AI content generation.", + "bbox": [ + 498, + 337, + 893, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 474, + 751, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. We present the first attempt to leverage LLMs as prompt engineers for VLMs. On one-shot image classification, our method surpasses existing human-engineered prompts and even rivals white-box approaches. Central to the success of our method is the use of conversational feedback, enabling chat-based LLMs to efficiently steer VLMs in the right direction. This process leads to a set of interpretable prompts bearing considerable resemblance to those crafted by humans. Importantly, our natural language prompting setup is a lot more constrained than the assumed scenarios of previous white-box or even some black-box settings [44], because we do not expose the model weights and outputs of VLMs. We finally apply our framework to illustrative generative tasks using a truly black-box text-to-image VLM (DALL-E 3).", + "bbox": [ + 496, + 500, + 893, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. While we try to minimize the overall cost and the total number of API calls, the energy consumption associated with LLMs remains a substantial concern. It is vital to note that we do not intend to compete directly with white-box baselines that can improve visual and text representations with more data. Further details on the higher-shot performance of our method can be found in section 9. Lastly, we are limited to costly human evaluation for T2I generation in this study. Future work may adopt automatic evaluation [18, 33] for large-scale experiments.", + "bbox": [ + 496, + 726, + 893, + 877 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12694", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1", + "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Joyce Zhuang, Juntang and Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiang, and Aditya Ramesh. Improving image generation with better captions. Note on Dalle-3, 2023. 1, 2, 3, 7", + "[5] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6, 7", + "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[7] Yekun Chai, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Clip-tuning: Towards derivative-free prompt learning with a mixture of rewards. arXiv preprint arXiv:2210.12050, 2022. 3", + "[8] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 7", + "[9] Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Zhifang Sui, and Furu Wei. Why can gpt learn in-context? language models secretly perform gradient descent as meta optimizers. ACL, 2023. 3, 4", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 1, 5, 6, 7", + "[11] Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P. Xing, and Zhiting Hu. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548, 2022.3", + "[12] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tong Zhang. Black-box prompt learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022. 3", + "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1", + "[14] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, and Jianfeng Gao. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 3", + "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1", + "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. arXiv preprint arXiv:2211.11559, 2022. 3", + "[17] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification, 2017. 5, 7", + "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8", + "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[20] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 5", + "[21] Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.3", + "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1", + "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 1", + "[25] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1", + "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + ], + "bbox": [ + 503, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12695", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 7", + "[27] Li, Andreeto, Ranzato, and Perona. Caltech 101, 2022. 6, 7", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3", + "[29] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pre-trained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022.3", + "[30] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1", + "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1", + "[33] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 8", + "[34] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramana. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. arXiv preprint arXiv:2301.06267, 2023. 1, 3, 5, 6", + "[35] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 3", + "[36] Vivian Liu. Beyond text-to-image: Multimodal prompts to explore generative ai. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-6, 2023. 7", + "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3", + "[38] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1", + "[39] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5, 7", + "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 1, 2, 3, 5, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830, 2021. 3", + "[42] Melanie Mitchell, John Holland, and Stephanie Forrest. When will a genetic algorithm outperform hill climbing. Advances in neural information processing systems, 6, 1993. 3", + "[43] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7", + "[44] Changdae Oh, Hyeji Hwang, Hee-young Lee, YongTaek Lim, Geunyoung Jung, Jiyoung Jung, Hosik Choi, and Kyungwoo Song. Blackvip: Black-box visual prompting for robust transfer learning. In CVPR, 2023. 3, 5, 6, 8", + "[45] OpenAI. Gpt-4 technical report. 2023. 1, 2, 3, 7", + "[46] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In ICCV, 2023. 3, 5, 6", + "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 1, 2, 3, 4", + "[48] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 7", + "[49] Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. Grips: Gradient-free, edit-based instruction search for prompting large language models. arXiv preprint arXiv:2203.07281, 2022. 3", + "[50] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2023. 1, 2, 3", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6, 13", + "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1", + "[53] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500–22510, 2023. 2, 8", + "[54] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010. 3, 4", + "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 4, 5" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12696", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 3", + "[57] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Zhe Gan, Lijuan Wang, Lu Yuan, Ce Liu, et al. K-lite: Learning transferable visual models with external knowledge. Advances in Neural Information Processing Systems, 35:15558-15573, 2022. 3", + "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3", + "[59] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3", + "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6, 7", + "[61] Tianxiang Sun, Zhengfu He, Hong Qian, Yunhua Zhou, Xuan-Jing Huang, and Xipeng Qiu. Bbtv2: Towards a gradient-free future with large language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3916-3930, 2022. 3", + "[62] Tianxiang Sun, Yunfan Shao, Hong Qian, Xuanjing Huang, and Xipeng Qiu. Black-box tuning for language-model-as-a-service. In International Conference on Machine Learning, pages 20841-20855. PMLR, 2022. 3", + "[63] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 7, 8", + "[64] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1", + "[65] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 8, 15", + "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35: 24824-24837, 2022. 1", + "[67] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. arXiv preprint arXiv:2109.01903, 2021. https://arxiv.org/abs/2109.01903.2,5,6" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. Int. J. Comput. Vision, 119(1): 3-22, 2016. 7", + "[69] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. Gps: Genetic prompt search for efficient few-shot learning. arXiv preprint arXiv:2210.17041, 2022.3", + "[70] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1", + "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1", + "[72] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3", + "[73] Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023. 3", + "[74] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1, 2, 3, 4, 5, 6", + "[75] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. *ICLR*, 2023. 2, 3, 13, 14" + ], + "bbox": [ + 501, + 92, + 893, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12697", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_model.json b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_model.json new file mode 100644 index 0000000000000000000000000000000000000000..01b27807d7f9546475b796860d47c40a453bb45e --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_model.json @@ -0,0 +1,1971 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.85, + 0.154 + ], + "angle": 0, + "content": "Language Models as Black-Box Optimizers for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.181, + 0.794, + 0.235 + ], + "angle": 0, + "content": "Shihong Liu* Samuel Yu* Zhiqiu Lin* Deepak Pathak Deva Ramanan Carnegie Mellon University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.741 + ], + "angle": 0, + "content": "Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities on downstream tasks when fine-tuned with minimal data. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. As such, we aim to develop a black-box approach to optimize VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or even output logits. We propose employing chat-based LLMs to search for the best text prompt for VLMs. Specifically, we adopt an automatic \"hill-climbing\" procedure that converges to an effective prompt by evaluating the performance of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot image classification setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of \\(1.5\\%\\) across 11 datasets including ImageNet. Our approach also outperforms both human-engineered and LLM-generated prompts. We highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit \"gradient\" direction in textual feedback for a more efficient search. In addition, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different VLM architectures in a black-box manner. Lastly, we demonstrate our framework on a state-of-the-art black-box VLM (DALL-E 3) for text-to-image optimization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.769, + 0.21, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Vision-language models [1, 28, 51, 64] (VLMs) excel at a wide range of classic vision and multimodal [2, 10, 15, 32, 70] tasks, surpassing the performance of their fully-supervised counterparts on downstream tasks even when fine-tuned with minimal data [34, 74]. However, fine-tuning VLMs typically requires transparent white-box access to the model weights, such as gradient-based approaches that rely" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.638, + 0.286 + ], + "angle": 0, + "content": "on backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.896, + 0.485 + ], + "angle": 0, + "content": "VLMs as black-box services. Despite community efforts to collect web-scale public datasets [55, 56] and to replicate proprietary VLMs [3, 24], an increasing number of models [1, 4, 13, 45, 64, 71] are not releasing their weights due to privacy and legal concerns [30, 38]. Therefore, one cannot use popular white-box fine-tuning strategies (such as LoRA [23] and Adapter [22]) that rely on model weights, feature embeddings, and output logits. Given that contemporary black-box VLMs [45, 47] like DALL-E [4, 52] still offer a language-based user interface and may be accessed through APIs that facilitate input and output in natural language, this allows users to customize these models through optimizing textual prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.897, + 0.699 + ], + "angle": 0, + "content": "Manual prompting. Manual prompt engineering has been proven successful in adapting black-box LLMs to language tasks [25, 66]. Similarly, carefully crafted prompts can enhance the performance of VLMs. For instance, CLIP has demonstrated improved zero-shot recognition performance using specifically tailored prompts, such as \"a photo of a {class}\" for Internet photos and \"a satellite image of a {class}\" for satellite imagery. Despite its effectiveness, manual prompting can be a laborious process, inspiring efforts to explore automated prompt creation and thereby remove the need for human involvement. These strategies typically leverage an LLM as a knowledge base to create rich visual descriptors that augment the prompts for each class [40, 50] in a zero-shot fashion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.897, + 0.869 + ], + "angle": 0, + "content": "Human-free prompting with conversational LLMs (our approach). We show how to effectively leverage chat-based LLMs [45] to emulate human-level prompt engineering without any human input. We first address an illustrative low-shot image classification task, aiming to find the best class-agnostic prompt (or \"template\") for image classification with CLIP. We start with a random set of prompts and evaluate the one-shot training accuracy of each. Then, akin to human prompt engineering, our method repeatedly presents ChatGPT with the best and worst prompts, asking it to review the results and suggest an improvement (see Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Learning with implicit \"gradients\" provided through conversational feedback. One of our key findings is that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.088, + 0.817, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.44, + 0.896, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Prompting VLMs using chat-based LLMs. Similar to how human prompt engineers iteratively test and refine prompts, we employ ChatGPT [45, 47] to continuously optimize prompts for vision-language models (VLMs). Our iterative approach assesses the performance of ChatGPT-generated prompts on a few-shot dataset (highlighted in blue) and provides feedback (marked in violet) to ChatGPT through simple conversations, as depicted in the illustrative figure. This straightforward method delivers state-of-the-art results for one-shot image classification across 11 datasets using CLIP, operated in a black-box manner without accessing model weights, feature embeddings, or output logits. We show that providing both positive (in green) and negative prompts (in red) enhances efficiency. Remarkably, our approach outperforms both white-box methods such as gradient-based continuous prompting (CoOp [74]) and human-engineered prompts [51] in this extremely low-shot scenario. This figure only shows a typical conversation using ChatGPT's web user interface. Our code implementation follows this pattern using the ChatGPT API. We detail and ablate the prompts in section 8." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.475, + 0.744 + ], + "angle": 0, + "content": "LLMs can learn the difference between effective and ineffective prompts, and can use this implicit \"gradient\" direction provided through language to perform more efficient searches. Compared to previous automatic prompting methods that only use LLMs as a knowledge base [40, 50] or paraphrasing tool [75], we show a novel use of LLMs as an optimizer that can utilize the patterns hidden in textual feedback. In our experiments, we find that the inclusion of such feedback greatly improves the efficiency and accuracy of our method, sometimes surpassing existing white-box methods [67, 74] on challenging one-shot scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Optimizing text-to-image generation with DALL-E 3. We further demonstrate our optimization framework on a state-of-the-art black-box VLM, DALL-E [4], for two illustrative one-shot generative tasks: (1) Text-to-image (T2I) generation (see Figure 3), where we sample challenging text queries from Winoground [63] that involve reasoning over compositions of objects, attributes, and relations. Examples include \"an animal watches a person\" and \"there is less milk than orange juice\", which DALL-E 3 might initially" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.73 + ], + "angle": 0, + "content": "fail to generate. (2) Prompt inversion (see Figure 4), which attempts to reverse-engineer the textual prompt to generate a specific image for later customization [53] (see Table 5). To achieve this, we leverage conversational feedback from a multimodal LLM (GPT4-V [45]) to iteratively refine the prompts based on the current generated images. We present qualitative results in Table 4 and conduct a user study to demonstrate that our framework can be more efficient than manual prompting, even for graphical designers experienced with AI content-generation tools." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Our contributions. In this work, we introduce a novel prompting method for VLMs, utilizing an LLM as an optimizer. Our black-box approach can surprisingly compete with various white-box methods in a low-shot setting. Additionally, we extensively explore various strategies for conversing with ChatGPT, uncovering several key factors that significantly enhance the efficiency of this tool. We also show that our discovered natural language prompts are not only interpretable but also transfer better across CLIP architectures, eg., from RN50 to ViT/B-16, than continuous prompts discovered by previous white-box prompting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12688" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.152 + ], + "angle": 0, + "content": "method [74]. Finally, we show practical applications of our framework on text-to-image generation using black-box DALL-E 3. We release our code for future research on prompt optimization and AI-driven content creation1." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.172, + 0.226, + 0.187 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.382 + ], + "angle": 0, + "content": "LLMs for multimodal tasks. Cutting-edge LLMs like GPTs [45, 47] have been successfully applied to multimodal tasks, either through zero-shot composition with pre-trained multimodal models [29, 72] or by jointly fine-tuning with modality-specific encoders [1, 28] on large-scale multimodal datasets [56]. LLMs are also utilized as neuro-symbolic reasoners [16, 37, 58, 73], translating natural language instructions into modular programs (like Python code) that invoke APIs of multimodal models. In this work, we show the potential of LLMs as a black-box optimizer for multimodal foundation models with language interfaces, and more specifically vision-language models (VLMs)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.581 + ], + "angle": 0, + "content": "Prompt optimization of foundation models. Following the success of in-context learning [6], which appends user-generated natural language instruction and few-shot samples to text inputs, prompting [35] has emerged as the preferred fine-tuning paradigm for LLMs due to its superior performance and parameter-efficiency. However, recent prompt optimization methods, including continuous prefix-tuning [7, 31, 61, 62, 69] and discrete token-searching [11, 12, 59], still operate in a white-box manner, requiring access to either the tokenizer or output logits. Moreover, black-box prompting methods, such as heuristic-based editing [41, 49], are tailored towards language-only tasks and are thus not applicable in VLM settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.471, + 0.87 + ], + "angle": 0, + "content": "LLMs for prompt optimization. APE [75] leverages an LLM to automatically write prompts using few-shot samples based on instruction induction [21] and paraphrasing [42, 54]. However, it is only designed to address language tasks, while we focus on multimodal tasks using black-box VLMs. LLMs have also proven to be an effective external knowledge base [40, 50, 57] for generating prompts in a zero-shot setting for multimodal models. For example, DCLIP [40] uses GPT3 to come up with rich visual descriptions to improve zero-shot classification with CLIP [51]. We extend this line of work to show that LLMs can iteratively optimize prompts for VLMs in a black-box fashion given few-shot samples. We further illustrate that prompt optimization with LLMs can be made more efficient by leveraging conversational feedback, such as providing ChatGPT with explicit language feedback on how well the most recent prompt performs. Our findings align with the perspective [9] of LLMs as meta-optimizers that can implicitly perform gradient search through in-context learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Few-shot adaptation of VLMs. Prompting has also been successfully adopted in VLMs [14], as demonstrated by methods like CoOp [74] that fine-tune an ensemble of continuous prefix tokens using cross-entropy loss. [34] achieves state-of-the-art few-shot performance with a cross-modal (image and text) cross-entropy loss. However, these methods all require access to model parameters for gradient backpropagation. We also note that while some concurrent works, such as BlackVIP [44] and LFA [46], claim to operate in a \"black-box\" setting, they still require access to privileged information including output logits and embeddings. In this work, we introduce a truly black-box and gradient-free approach that yields competitive results to white-box approaches in extremely low-shot scenarios." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.317, + 0.885, + 0.334 + ], + "angle": 0, + "content": "3. Prompting VLMs Using Chat-Based LLMs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.372 + ], + "angle": 0, + "content": "We now present our approach for prompting VLMs using chat-based LLMs as optimizers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Preliminaries. Motivated by recent proprietary VLMs [4, 45], we adopt a stricter yet practical black-box setting compared to prior works [44, 46], requiring minimal knowledge about the model's inner workings. This is crucial since releasing output logits or embeddings can potentially facilitate unauthorized knowledge extraction through distillation methods [19]. Our objective is to enhance the performance of a VLM equipped with a language interface capable of processing a textual prompt \\( p \\in T \\). We assume that the targeted task is accompanied by a training dataset denoted as \\( D_{train} \\subset D \\) and its performance can be evaluated with respect to the prompt, represented as a function \\( F: D \\times T \\to \\mathbb{R} \\). For example, in a classification task, \\( D_{train} = \\{x, y\\}_n \\) where \\( x \\) is an image and \\( y \\) is its class label. The black-box VLM takes the image as input and returns a predicted label. We measure the performance of the textual prompt by calculating the average classification accuracy as \\( F(D_{train}, p) \\). Our goal in prompt engineering is to search for the optimal prompt \\( p^* \\) without accessing or modifying the black-box VLM." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Background: human prompt engineering. Our method draws inspiration from the typical workflow of human prompt engineers. Prompt engineering is often an iterative process that involves: (a) creating an initial prompt \\( U = \\{p_1\\} \\) based on the understanding of a task, (b) evaluating the performance of prompts in \\( U \\), (c) refining prompts based on the outcomes, (d) repeating the last two steps until convergence, and (e) returning the prompt \\( p^* \\) with the highest \\( F(D_{train}, p^*) \\). This hands-on approach helps optimize the model's performance, but it can be tedious and labor-intensive. Algorithm 1 formally illustrates this process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Example: prompting for image classification with CLIP [51]. CLIP is one of the most popular VLM that takes a set of class-specific prompts when performing \"zero-shot\" image classification. [51] details the laborious prompting procedure over the course of a year. Interestingly, they find" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.409, + 0.901 + ], + "angle": 0, + "content": "1Project site: llm-can-optimize-vlm.github.io" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12689" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.078, + 0.091, + 0.473, + 0.136 + ], + "angle": 0, + "content": "Algorithm 1 We formalize human prompt engineering with the following algorithm, which motivates our LLM-based algorithm (2)." + }, + { + "type": "algorithm", + "bbox": [ + 0.078, + 0.141, + 0.472, + 0.314 + ], + "angle": 0, + "content": "Require: \\(D_{\\mathrm{train}} = \\{x,y\\}_{n}\\) training samples, \\(F:D\\times T\\to\\) R: evaluation function \n1: Create initial prompts: \\(\\mathcal{U}\\gets \\{p_1\\}\\) \n2: Evaluate prompts on training set: \\(S\\gets \\{F(D_{\\mathrm{train}},p_1)\\}\\) \n3: while not converged do \n4: Generate a new prompt \\(p^\\prime\\) based on \\(S\\) \n5: Evaluate the new prompt: \\(s^{\\prime} = F(D_{\\mathrm{train}},p^{\\prime})\\) \n6: \\(\\mathcal{U}\\leftarrow \\mathcal{U}\\cup \\{p^{\\prime}\\}\\) \n7: \\(S\\gets S\\cup \\{s^{\\prime}\\}\\) \n8: end while \n9: return optimal prompt \\(p^*\\gets \\arg \\max_{p\\in \\mathcal{U}}F(D_{\\mathrm{train}},p)\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.452 + ], + "angle": 0, + "content": "that a default class-agnostic prompt (or so-called \"template\"), \"a photo of a {class}\" can provide a decent boost in accuracy for most datasets compared to using vanilla class labels. In this scenario, the evaluation function \\( F \\) is the classification accuracy on the test set, and the prompt \\( p = \\{ \\text{\"a photo of a {class}\" } | c \\in C \\} \\), where \\( C \\) is the set of class names for a given dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.713 + ], + "angle": 0, + "content": "Prompting with chat-based LLMs (our approach). Given the strong in-context reasoning capabilities of LLMs, we envision them as a black-box optimizers that can improve prompts based on their performance outcomes, akin to how human prompt engineers iteratively refine prompts. Specifically, we maintain a pool of prompts \\( U \\) and their corresponding performance outcomes \\( S \\). In each iteration, we provide the LLM with both positive and negative prompts, such as the highest and lowest-performing candidates. Such textual feedback through in-context prompts offers LLMs an implied \"gradient\" direction [9], making optimization more efficient than taking random local steps. We facilitate this feedback mechanism through conversations with state-of-the-art chat-based LLMs like ChatGPT [47] as illustrated in Figure 1. We note that such a multi-turn conversation is not the only way of conversing with ChatGPT, and ablate different in-context feedback mechanisms in section 8." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.736, + 0.44, + 0.753 + ], + "angle": 0, + "content": "4. Illustrative Few-Shot Classification Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We illustrate our approach using a few-shot image classification task. Specifically, a prompt \\( p \\in T \\) consists of a set of class-specific prompts – that is, one textual description per class. The evaluation function \\( F \\) takes the prompt \\( p \\), along with an image dataset \\( D_{train} \\), and returns the accuracy using the black-box VLM. To prevent overfitting and simplify our search space, we restrict our search to finding a single class-agnostic template, e.g., a photo of a {}, filling in the blank with label names provided with the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.895, + 0.198 + ], + "angle": 0, + "content": "Algorithm 2 LLM-based prompt engineering on the illustrative classification task. Our algorithm requires a chat-based LLM and a (black-box) evaluation function, such as accuracy. We highlight mechanisms for “exploration” (restart and reset) in blue and “exploitation” (iter) in red. We mark the key component of “conversational feedback” of our approach in violet. The actual prompts are attached in section 8." + }, + { + "type": "algorithm", + "bbox": [ + 0.5, + 0.201, + 0.895, + 0.537 + ], + "angle": 0, + "content": "Require: \\(D_{\\mathrm{train}} = \\{x,y\\}_{n}\\) training samples, \\(F:D\\times T\\to\\) R: evaluation function. \nRequire: \\(n_{\\mathrm{restart}}\\) : number of initial sampled prompt sets, \\(n_{\\mathrm{reset}}\\) : number of resets for a prompt set, \\(n_{\\mathrm{iter}}\\) : number of hill-climbing iterations, \\(m\\) size of one initial prompt set, \\(k\\) : number of prompts send to ChatGPT. \n1: \\(p^* \\gets \\emptyset\\) \n2: for 1::\\(n_{\\mathrm{restart}}\\) do \n3: Sample a new prompt set, \\(\\mathcal{U}_{\\mathrm{init}} \\leftarrow \\{p_1,\\dots,p_m\\}\\) \n4: for 1::\\(n_{\\mathrm{reset}}\\) do \n5: Reset to initial prompt set: \\(\\mathcal{U} \\leftarrow \\mathcal{U}_{\\mathrm{init}}\\) \n6: for 1::\\(n_{\\mathrm{iter}}\\) do \n7: Sort \\(\\mathcal{U}\\) by score outcomes \\(\\{F(D_{\\mathrm{train}},p)\\}_{p\\in U}\\) \n8: \\(\\mathcal{U}_{\\mathrm{top}} \\leftarrow\\) top-k prompts in \\(\\mathcal{U}\\) \n9: \\(\\mathcal{U}_{\\mathrm{bot}} \\leftarrow\\) bottom-k prompts in \\(\\mathcal{U}\\) \n10: Get a new prompt \\(p_{\\mathrm{new}} \\leftarrow\\) LLM(\\(\\mathcal{U}_{\\mathrm{top}},\\mathcal{U}_{\\mathrm{bot}}\\)) \n11: \\(\\mathcal{U} \\leftarrow \\mathcal{U} \\cup \\{p_{\\mathrm{new}}\\}\\) \n12: end for \n13: \\(p^* \\leftarrow\\) arg max \\(p\\in \\mathcal{U}\\cup \\{p^*\\}\\) \\(F(D_{\\mathrm{train}},p)\\) \n14: end for \n15: end for \n16: return prompt with highest score \\(p^*\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Outline of our approach (Alg. 2). To start, we sample entirely random initial prompts from a text corpus such as LAION-COCO [55] captions. Our approach follows the classical stochastic hill-climbing framework with random restart [54], which prevents ChatGPT from being trapped in local optima by balancing \"exploration\" and \"exploitation\". Our restart mechanism is implemented by sampling \\( n_{\\mathrm{restart}} \\) initial prompt sets to encourage exploration. Because ChatGPT performs stochastic top-k sampling for text generation (as we adopt the default temperature of 1.0), we also implement a reset mechanism to foster additional exploration by retrying a given prompt set \\( n_{\\mathrm{reset}} \\) times. For exploitation, we converse with ChatGPT for \\( n_{\\mathrm{iter}} \\) iterations. We find that it is critical to balance exploration and exploitation for optimal performance, and thoroughly examine this trade-off in section 9. Lastly, we present ChatGPT both the top and bottom-performing prompts, denoted as \\( (\\mathcal{U}_{top}, \\mathcal{U}_{bot}) \\). We show that this simple adjustment can improve the efficiency of our approach in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Experimental setup. We apply our approach to the few-shot image classification benchmark introduced in CoOp [74], which is the most commonly studied setup for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "fine-tuning VLMs. This benchmark involves a collection of 11 datasets covering diverse image domains including ImageNet [10] and more niche datasets such as FGVC-Aircraft [39]. For each dataset, we adhere to the same three-fold k-shot train sets in [34], reporting the average accuracy across all folds. Importantly, our method only utilizes the train set to compute the score and does not require the few-shot validation set. We use CLIP following prior work [34, 74] to emulate a black-box VLM, and we employ ChatGPT (GPT3.5) as the chat-based LLM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.247, + 0.473, + 0.595 + ], + "angle": 0, + "content": "Implementation details. To start, we sample entirely random 1M initial prompts from a text corpus (LAION-COCO [55] captions). For each caption, we extract all the noun phrases using spaCy part-of-speech tagging [20]. Subsequently, we replace one noun phrase in the caption with \\(\\{\\}\\) (a placeholder where the class name will be inserted) to create a template. Given that each caption contains an average of 2 noun phrases, our initial prompt pool consists of approximately 2M templates. We run our algorithm with \\(n_{\\mathrm{start}} = 20\\) restarts, \\(n_{\\mathrm{resets}} = 50\\) resets, and \\(n_{\\mathrm{iter}} = 10\\) iterations. We opt to sample \\(m = 100\\) prompts per restart and present the top and bottom \\(k = 15\\) prompts to ChatGPT. We ablate different sets of hyperparameters and explain how we balance the tradeoff between exploration and exploitation in section 9. We adopt gpt-3.5-turbo-0301 model for ChatGPT using OpenAI's official API and keep the default sampling temperature of 1.0. We also ablate gpt-4 in Table 10 and find it achieves similar performance. The exact prompts used to converse with ChatGPT are documented in section 8. For a fair comparison, we use CLIP-RN50 for our experiments following prior work [34, 74]. We will open-source our code and release the initial prompt pool (LAIONCOCO-1M) to the public." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Oracle white-box baselines. Our black-box setup substantially differs from, and is more constrained than, the scenarios considered in previous white-box baselines. Specifically, we do not expose the pre-trained weights, model architectures, feature embeddings, or even output logits of VLMs. These constraints render many established gradient-based fine-tuning baselines inapplicable. Among the oracle white-box approaches we later compare to, CoOp [74] performs continuous prompting and requires backpropagation across all layers. WiSE-FT [67] ensembles fine-tuned weights with the original CLIP weights. Cross-Modal Adaptation [34] fine-tunes a linear classifier leveraging both image and text embeddings from CLIP. BlackVIP [44] and LFA [46] are two most recent baselines that apply CLIP logits or embeddings for gradient back-propagation. Finally, while DCLIP [40] queries GPT3 for rich visual descriptors for each class and does not require gradient-based fine-tuning, it performs prompt assembling using 4-6 class-specific prompts, which breaches our black-box assumption for accessing the output logits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.332 + ], + "angle": 0, + "content": "Black-box methods. We additionally benchmark our method against truly black-box solutions, including the vanilla class-agnostic templates “{class)” and “a photo of a {class}”. Also, we compare our approach to the best Hand-Engineered templates released by OpenAI, searched using test set performance to represent the theoretical upper bound of human performance, eg., “a centered satellite photo of {class}.” for EuroSAT [17]. Finally, we present two versions of conversational feedback of our approach: (a) using 30 positive (P only) or (b) using 15 positive and 15 negative prompts \\((\\mathbf{P} + \\mathbf{N})\\) in each iteration. For a fair comparison, both of our approaches start with the same initial sampled prompts, referred to as LAIONCOCO-1M. We also show the performance of the best initial sampled prompt searched using trainset performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.897, + 0.649 + ], + "angle": 0, + "content": "SOTA one-shot performance against existing methods on 11 datasets. We report the test set performance of our method versus the aforementioned baselines in a challenging 1-shot classification scenario in Table 1. First, compared to the top-performing initial prompts selected from LAIONCOCO-1M based on train set performance, our prompt optimization using ChatGPT notably improves the initial prompts by an average of \\(5\\%\\) (56% to 61%). Remarkably, our black-box approach surpasses the two white-box gradient-based fine-tuning techniques CoOp and WiSE-FT by at least \\(1.5\\%\\). Given that both CoOp and our method optimize a single class-agnostic template, we attribute this gap in performance to reduced overfitting. More specifically, we posit that our optimization space of natural language effectively acts as a regularizer in extremely low-shot tasks, standing as a more robust alternative to the continuous prompting approach of CoOp. Furthermore, our method benefits from textual feedback and shows improved performance by \\(1.0\\%\\) when using both positive and negative prompts. In section 9, we show that our approach remains effective across different CLIP and ChatGPT variants." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.897, + 0.817 + ], + "angle": 0, + "content": "Incorporating negative prompts leads to more efficient optimization. In Figure 2, we demonstrate that incorporating both positive and negative prompts fosters better optimization efficiency, achieving higher accuracy within a much fewer number of resets. Specifically, we hypothesize that LLMs can leverage the implicit \"gradient\" direction suggested in textual feedback to achieve faster convergence. For additional analysis, we ablate different ways of providing conversational feedback to ChatGPT in section 8 and conclude that iteratively updating both positive and negative prompts is the key for efficient optimization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.892, + 0.848 + ], + "angle": 0, + "content": "5. More Benefits of Natural Language Prompts" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we delve deeper into the advantages of utilizing natural language prompts compared to the continuous prompts [74]. We highlight that the prompts derived through" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12691" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.229, + 0.089, + 0.744, + 0.299 + ], + "angle": 0, + "content": "
MethodDatasetAvg
CaltechImageNetAircraftFoodPetsCarsSUNUCFDTDEuroSATFlowers
Oracle white-box approaches
Cross-Modal [34]89.161.620.677.185.759.063.464.749.961.876.364.7
WiSE-FT [67]85.558.318.671.981.755.756.659.444.252.365.859.1
CoOp [74]87.557.29.674.385.955.660.361.944.450.668.159.6
LFA [46]81.652.417.063.175.341.458.456.738.460.774.956.4
BlackVIP [44]85.858.815.376.785.256.457.058.840.130.061.156.8
DCLIP [40]-59.6-76.483.8---41.734.7--
Manual prompting approaches
{}78.555.315.574.078.952.253.455.541.432.157.354.0
a photo of a {}84.557.915.974.083.253.958.056.938.828.660.255.6
Hand-Engineered [51]86.358.217.377.385.855.658.561.542.337.666.158.8
Our black-box approaches
LAIONCOCO-1M81.456.217.476.579.651.354.955.843.138.661.356.0
Ours (P only)89.059.417.977.885.755.760.458.743.646.766.660.1
Ours (P+N)89.159.618.178.388.156.261.060.244.849.067.261.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.308, + 0.895, + 0.449 + ], + "angle": 0, + "content": "Table 1. Comparison of our method with other baselines on one-shot classification tasks. We report the average accuracy of each method across three folds, optimized using 1-shot training sets. We bold the best black-box result for each dataset, and underline the second best result. First, we note that our approach can effectively improve upon the initial prompts selected from LAIONCOCO-1M from \\(56\\%\\) to \\(61\\%\\). Our approach is also competitive against the best Human-Engineered prompts released by OpenAI [51] searched using test set performance. Additionally, we show that using both positive and negative prompts improves the overall accuracy by \\(1\\%\\). For reference, we report oracle white-box approaches in gray. Remarkably, we also surpass white-box solutions such as WiSE-FT [67] and CoOp [74] by \\(1.5\\%\\). These methods require either gradient-based fine-tuning (CoOp/WiSE-FT/Cross-Modal) or prompt ensembling using output logits (DCLIP). While our approach is less effective than the SOTA white-box method (Cross-Modal Adaptation), we stress that our black-box setup is significantly more challenging, because we restrict the optimization space to natural language and do not access the pre-trained weights, model architectures, feature embeddings, and output logits of VLMs." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.478, + 0.258, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.478, + 0.452, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.597, + 0.472, + 0.708 + ], + "angle": 0, + "content": "Figure 2. Conversational feedback incorporating both positive and negative prompts leads to improved efficiency. We fix the number of restarts to 20 and iterations to 10, and ablate different numbers of resets on all 11 datasets (left) and ImageNet (right). Notably, our approach using “\\(\\mathrm{P} + \\mathrm{N}\\)” (both top-15 and bottom-15 prompts) can optimize faster within a much fewer number of resets than using “P-Only” (top-30 prompts), resulting in the highest overall performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.471, + 0.79 + ], + "angle": 0, + "content": "our method are interpretable; for instance, they often contain descriptions of the targeted image domain. Our prompts can also transfer across CLIP architectures in a black-box manner, such as from RN50 to ViT/B-16." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Interpretable natural language prompts. While CoOp [74] concedes that continuous prompts can be difficult to interpret, our method – without explicitly instructing ChatGPT to generate interpretation – often yields interpretable results. Table 2 showcases the templates returned by our algorithm for each dataset, frequently including keywords that reflect the targeted image domain. For example, the template" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.895, + 0.641 + ], + "angle": 0, + "content": "for Food101 [5] mentions \"diverse cuisine and ingredients\", and the template for UCF101 [60] (an action recognition dataset) mentions \"in motion\". Likewise, these templates identify general stylistic attributes of the datasets; they refer to \"bright and natural lighting\" for ImageNet [10] and note images that \"emphasize the subject\" for Caltech101 [27]. These prompts are particularly intriguing because we do not provide ChatGPT with any information about the downstream task, yet it manages to generate prompts containing domain-specific keywords that are similar to those engineered by human experts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Black-box prompt transfer. Our text prompts also maintain consistently high performance across different CLIP backbones. For comparison, since CoOp uses the same tokenizer for all CLIP architectures (including RN50, RN101, ViT/B-32, and ViT/B-16) and optimizes continuous prompts of the same shape (16 x 512), we assess the transferability of these learned continuous prompts from RN50 to other backbones using the official weights on 16-shot ImageNet. Table 3 showcases the results of this experiment, where we also include the baseline prompt \"a photo of a {}\" for reference. We observe a significant decline in accuracy when transferring CoOp's prompts (up to a \\(40\\%\\) decrease despite utilizing more powerful backbones), implying that continuous prompts tend to overfit to the specific CLIP model. In contrast, our natural language prompts maintain their performance and outperform the baseline across all backbones." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.249, + 0.09, + 0.723, + 0.216 + ], + "angle": 0, + "content": "
DatasetExample of Top Templates
Caltech [27]An image of a {} with a blurred background that emphasizes the subject
DTD [8]The essential elements of {} are amplified with visual simplicity
EuroSAT [17]A top-down view of {} arranged in a pattern {}
Aircraft [39]A clear, high-quality image of a single {} with a white background
Food [5]A {} featuring diverse cuisine and ingredients
ImageNet [10]An image of a {} with bright and natural lighting
Flowers [43]A clear and vivid photograph of the {} in its natural setting
Pets [48]A {} with distinct and recognizable characteristics
Cars [26]A {} featuring a wide range of color options for easy selection
SUN [68]A high-resolution photo of a {} with clear background and natural lighting
UCF [60]A black and white photo of a {} in motion
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.227, + 0.895, + 0.27 + ], + "angle": 0, + "content": "Table 2. Example templates returned by our algorithm on each dataset. Although we do not provide ChatGPT with any information regarding the targeted dataset, we observe that the resulting templates are remarkably similar to human-engineered templates, with many domain-specific details such as \"motion\" and \"cuisine\", and stylistic elements such as \"bright and natural lighting\"." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.276, + 0.451, + 0.343 + ], + "angle": 0, + "content": "
MethodRN50→RN101→ViT-B/32→ViT-B/16
a photo of a {}57.960.661.966.6
CoOp63.020.631.739.5
Ours59.960.762.267.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.355, + 0.473, + 0.493 + ], + "angle": 0, + "content": "Table 3. Black-box prompt transfer from ResNet-50 to other CLIP architectures. We evaluate both our natural language prompts and CoOp's continuous prompts on 16-shot ImageNet, which are trained using the RN50 CLIP backbone. As a reference point, we include the baseline prompt \"a photo of a \\(\\{\\}\\)\", and show that the prompts derived from our method using RN50 consistently surpass it after transferring to different backbones. In contrast, while CoOp achieves better 16-shot ImageNet performance using RN50, its performance plummets during the transfer, e.g., from \\(63\\%\\) to a mere \\(21\\%\\) for RN101." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.505, + 0.43, + 0.523 + ], + "angle": 0, + "content": "6. Application: Text-to-Image Generation" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.542, + 0.478, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.742, + 0.473, + 0.907 + ], + "angle": 0, + "content": "Figure 3. Improving text-to-image (T2I) generation using chat-based multimodal LLMs. Our framework can use the multimodal GPT4-V [45] to optimize prompts for the state-of-the-art black-box generative VLM, DALL-E 3 [4]. For complex user queries that DALL-E 3 may initially fail to generate, we send the generated image (in violet) along with the current prompt to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue). Our simple framework is surprisingly effective at correcting DALL-E 3 mistakes on some challenging Winoground [63] text queries that involve action, logical, and spatial reasoning. We conduct a human evaluation on the quality of generated images in Table 6 and include the actual prompts in section 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.323 + ], + "angle": 0, + "content": "In this section, we present a direct application of our prompt optimization framework to generative tasks using a truly black-box text-to-image (T2I) VLM, DALL-E 3 [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.583 + ], + "angle": 0, + "content": "Optimizing T2I using a multimodal LLM. DALL-E 3 can generate high-fidelity images following diverse user queries, but crafting effective prompts is tricky even for designers experienced with AI content generation tools [36]. Therefore, we are motivated to implement our LLM-based optimization framework to assist with creative visual design. Our framework is shown in Figure 3 for the illustrative task of text-to-image generation. In this task, the user specifies a query (topic) in text, such as \"an animal watches a person\", and the goal is to write a prompt that can generate an image reflecting this topic. We adopt a multimodal LLM GPT4-V [45] (gpt-4-1106-preview) to provide feedback on the generated image and optimize the prompt. We find that this framework is surprisingly effective due to GPT4-V's strong visual reasoning capabilities, which can often spot subtle errors in generated images and offer more accurate prompts." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.598, + 0.902, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.797, + 0.895, + 0.88 + ], + "angle": 0, + "content": "Figure 4. Prompt inversion using chat-based multimodal LLMs. We apply our framework to reverse engineer the text prompt to generate the same user-queried image. We send the generated image (in violet) along with the original image to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue)." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Task setup. For T2I generation, we experiment with a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12693" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.446, + 0.314 + ], + "angle": 0, + "content": "
User QueryInit. ImageLLM FeedbackFinal Image
Text-to-image generation
There is less milk than orange juice.Incorrect, the milk bottle appears full, more than orange juice...
A shorter person is covering the eyes of a taller person.Incorrect, the taller person is covering the shorter person's eyes. Instead, ...
Prompt inversion
The scarf should feature red and white stripes, and the fur is fluffy...
The coat should be buttoned and the lighting exhibits a stronger contrast...
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.325, + 0.474, + 0.424 + ], + "angle": 0, + "content": "Table 4. Examples of T2I optimization. We show that our framework (Figure 3) can automatically improve the faithfulness of images generated by DALL-E 3, with respect to user-specified textual topics (for T2I generation) or reference images (for prompt inversion). This is achieved through three rounds of prompt optimization, using feedback from the multimodal LLM (GPT4-V). Table 11 and Table 12 shows more examples with actual prompts." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.426, + 0.479, + 0.584 + ], + "angle": 0, + "content": "
User QueryInverted ImageExample 1Example 2Example 3Example 4Example 5
Give the dog a cat friend.Make the dog be in the middle of a jump.Make the dog do a handstand.Make the dog lie down on its side.Make the dog swim in water.
Make the owl fight a hawk.Make the owl flap its wings.Make the owl fully green.Make the owl stand in front of the moon.Make the owl walk in the city.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.595, + 0.471, + 0.637 + ], + "angle": 0, + "content": "Table 5. Customization via prompt inversion. Users can simply append extra descriptions to the inverted prompts to customize their main characters in queried images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "subset of 100 text queries from Winoground [63] that involve complex attribute and relation reasoning, which DALL-E might initially fail to generate. Our framework refines the prompts to capture the user-specified topics using a few (three) iterations. We also attempt a reverse task of prompt inversion: given a user-specified reference (query) image, our framework reverse-engineers the prompt to have DALL-E generate the same object or scene in the query image (see Figure 4). This enables users to easily make customizations [53] (see Table 5), such as having the character in a reference image perform various actions or change scenes. For this task, we sample 100 random queries from DiffusionDB [65]. We provide qualitative results in Table 4, Table 11, and Table 12. We hire two volunteers to assess the faithfulness of the images generated by our method, and to compare these with the images manually prompted by two" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "designers (each with one year of experience in AI content generation), as shown in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.895, + 0.227 + ], + "angle": 0, + "content": "Remarks on limitations. While we show promising results, we note some failure cases in Table 13 and Table 14 due to the inherent limitations of foundation models. For example, GPT4-V might fail to describe abstract and artistic details, and DALL-E 3 often fails to generate the correct number of objects. We believe that our framework can benefit from more capable foundation models in the future." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.239, + 0.848, + 0.327 + ], + "angle": 0, + "content": "
TaskMethodInit. (std)Final (std)Δ
Text-to-ImageHuman2.28 (.45)2.86 (.61)0.58
Ours2.62 (.36)3.56 (.54)0.94
Prompt InversionHuman1.58 (.48)2.76 (.53)1.18
Ours1.94 (.39)3.68 (.47)1.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.338, + 0.895, + 0.463 + ], + "angle": 0, + "content": "Table 6. Our method enhances faithfulness in T2I generation. We hire two human annotators to assess the faithfulness of images generated from user queries, e.g., textual topics for Text-to-Image, or reference images for Prompt Inversion. The scores are measured on a 1-to-5 Likert scale, with 1 signifying contradiction and 5 indicating perfect alignment with the user's goal. Our approach benefits from three iterations of prompt optimization and consistently outperforms human-engineered prompts by designers who have one year of experience in AI content generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.475, + 0.753, + 0.491 + ], + "angle": 0, + "content": "7. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.725 + ], + "angle": 0, + "content": "Summary. We present the first attempt to leverage LLMs as prompt engineers for VLMs. On one-shot image classification, our method surpasses existing human-engineered prompts and even rivals white-box approaches. Central to the success of our method is the use of conversational feedback, enabling chat-based LLMs to efficiently steer VLMs in the right direction. This process leads to a set of interpretable prompts bearing considerable resemblance to those crafted by humans. Importantly, our natural language prompting setup is a lot more constrained than the assumed scenarios of previous white-box or even some black-box settings [44], because we do not expose the model weights and outputs of VLMs. We finally apply our framework to illustrative generative tasks using a truly black-box text-to-image VLM (DALL-E 3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.894, + 0.878 + ], + "angle": 0, + "content": "Limitations and future work. While we try to minimize the overall cost and the total number of API calls, the energy consumption associated with LLMs remains a substantial concern. It is vital to note that we do not intend to compete directly with white-box baselines that can improve visual and text representations with more data. Further details on the higher-shot performance of our method can be found in section 9. Lastly, we are limited to costly human evaluation for T2I generation in this study. Future work may adopt automatic evaluation [18, 33] for large-scale experiments." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12694" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.472, + 0.369 + ], + "angle": 0, + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.371, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Joyce Zhuang, Juntang and Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiang, and Aditya Ramesh. Improving image generation with better captions. Note on Dalle-3, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[5] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.514, + 0.472, + 0.583 + ], + "angle": 0, + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.585, + 0.472, + 0.64 + ], + "angle": 0, + "content": "[7] Yekun Chai, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Clip-tuning: Towards derivative-free prompt learning with a mixture of rewards. arXiv preprint arXiv:2210.12050, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.472, + 0.698 + ], + "angle": 0, + "content": "[8] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.701, + 0.472, + 0.755 + ], + "angle": 0, + "content": "[9] Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Zhifang Sui, and Furu Wei. Why can gpt learn in-context? language models secretly perform gradient descent as meta optimizers. ACL, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.471, + 0.814 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.883 + ], + "angle": 0, + "content": "[11] Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P. Xing, and Zhiting Hu. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.887, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[12] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Tong Zhang. Black-box prompt learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.188 + ], + "angle": 0, + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[14] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, and Jianfeng Gao. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.894, + 0.315 + ], + "angle": 0, + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. arXiv preprint arXiv:2211.11559, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.894, + 0.412 + ], + "angle": 0, + "content": "[17] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.894, + 0.466 + ], + "angle": 0, + "content": "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.894, + 0.51 + ], + "angle": 0, + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.565 + ], + "angle": 0, + "content": "[20] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.894, + 0.621 + ], + "angle": 0, + "content": "[21] Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.894, + 0.747 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.872 + ], + "angle": 0, + "content": "[25] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12695" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.465, + 0.137 + ], + "angle": 0, + "content": "[27] Li, Andreeto, Ranzato, and Perona. Caltech 101, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.139, + 0.471, + 0.193 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.197, + 0.472, + 0.25 + ], + "angle": 0, + "content": "[29] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pre-trained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.255, + 0.471, + 0.308 + ], + "angle": 0, + "content": "[30] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.313, + 0.471, + 0.353 + ], + "angle": 0, + "content": "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.357, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.471, + 0.495 + ], + "angle": 0, + "content": "[33] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.5, + 0.471, + 0.554 + ], + "angle": 0, + "content": "[34] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramana. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. arXiv preprint arXiv:2301.06267, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.557, + 0.471, + 0.624 + ], + "angle": 0, + "content": "[35] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.629, + 0.471, + 0.683 + ], + "angle": 0, + "content": "[36] Vivian Liu. Beyond text-to-image: Multimodal prompts to explore generative ai. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-6, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.687, + 0.471, + 0.754 + ], + "angle": 0, + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.759, + 0.471, + 0.812 + ], + "angle": 0, + "content": "[38] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[39] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.471, + 0.899 + ], + "angle": 0, + "content": "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[41] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[42] Melanie Mitchell, John Holland, and Stephanie Forrest. When will a genetic algorithm outperform hill climbing. Advances in neural information processing systems, 6, 1993. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[43] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[44] Changdae Oh, Hyeji Hwang, Hee-young Lee, YongTaek Lim, Geunyoung Jung, Jiyoung Jung, Hosik Choi, and Kyungwoo Song. Blackvip: Black-box visual prompting for robust transfer learning. In CVPR, 2023. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.297, + 0.822, + 0.31 + ], + "angle": 0, + "content": "[45] OpenAI. Gpt-4 technical report. 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.313, + 0.892, + 0.353 + ], + "angle": 0, + "content": "[46] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In ICCV, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.486, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[49] Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. Grips: Gradient-free, edit-based instruction search for prompting large language models. arXiv preprint arXiv:2203.07281, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.544, + 0.892, + 0.584 + ], + "angle": 0, + "content": "[50] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.587, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.673, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[53] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500–22510, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[54] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12696" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.174 + ], + "angle": 0, + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[57] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Zhe Gan, Lijuan Wang, Lu Yuan, Ce Liu, et al. K-lite: Learning transferable visual models with external knowledge. Advances in Neural Information Processing Systems, 35:15558-15573, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.303 + ], + "angle": 0, + "content": "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[59] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.472, + 0.404 + ], + "angle": 0, + "content": "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.472, + 0.475 + ], + "angle": 0, + "content": "[61] Tianxiang Sun, Zhengfu He, Hong Qian, Yunhua Zhou, Xuan-Jing Huang, and Xipeng Qiu. Bbtv2: Towards a gradient-free future with large language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3916-3930, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.472, + 0.532 + ], + "angle": 0, + "content": "[62] Tianxiang Sun, Yunfan Shao, Hong Qian, Xuanjing Huang, and Xipeng Qiu. Black-box tuning for language-model-as-a-service. In International Conference on Machine Learning, pages 20841-20855. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.472, + 0.617 + ], + "angle": 0, + "content": "[63] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.472, + 0.674 + ], + "angle": 0, + "content": "[64] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[65] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 8, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.472, + 0.815 + ], + "angle": 0, + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35: 24824-24837, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[67] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. arXiv preprint arXiv:2109.01903, 2021. https://arxiv.org/abs/2109.01903.2,5,6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[68] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. Int. J. Comput. Vision, 119(1): 3-22, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[69] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. Gps: Genetic prompt search for efficient few-shot learning. arXiv preprint arXiv:2210.17041, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[70] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.33 + ], + "angle": 0, + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.894, + 0.402 + ], + "angle": 0, + "content": "[72] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.894, + 0.457 + ], + "angle": 0, + "content": "[73] Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.894, + 0.501 + ], + "angle": 0, + "content": "[74] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.894, + 0.557 + ], + "angle": 0, + "content": "[75] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. *ICLR*, 2023. 2, 3, 13, 14" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12697" + } + ] +] \ No newline at end of file diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_origin.pdf b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..45458b4e8c9c641f4fed08ddf45ca2a14a56c26c --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/40b8d27b-a1d9-46c0-919c-6079601bc412_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:527e4319bae2fcf93457b616405597aa4221d9c21cb0f697a433b6613b0b1290 +size 7951838 diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/full.md b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37005fb7874e4d5249e9efb630db0b4333996f79 --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/full.md @@ -0,0 +1,263 @@ +# Language Models as Black-Box Optimizers for Vision-Language Models + +Shihong Liu* Samuel Yu* Zhiqiu Lin* Deepak Pathak Deva Ramanan Carnegie Mellon University + +# Abstract + +Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities on downstream tasks when fine-tuned with minimal data. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. As such, we aim to develop a black-box approach to optimize VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or even output logits. We propose employing chat-based LLMs to search for the best text prompt for VLMs. Specifically, we adopt an automatic "hill-climbing" procedure that converges to an effective prompt by evaluating the performance of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot image classification setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of $1.5\%$ across 11 datasets including ImageNet. Our approach also outperforms both human-engineered and LLM-generated prompts. We highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit "gradient" direction in textual feedback for a more efficient search. In addition, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different VLM architectures in a black-box manner. Lastly, we demonstrate our framework on a state-of-the-art black-box VLM (DALL-E 3) for text-to-image optimization. + +# 1. Introduction + +Vision-language models [1, 28, 51, 64] (VLMs) excel at a wide range of classic vision and multimodal [2, 10, 15, 32, 70] tasks, surpassing the performance of their fully-supervised counterparts on downstream tasks even when fine-tuned with minimal data [34, 74]. However, fine-tuning VLMs typically requires transparent white-box access to the model weights, such as gradient-based approaches that rely + +on backpropagation. + +VLMs as black-box services. Despite community efforts to collect web-scale public datasets [55, 56] and to replicate proprietary VLMs [3, 24], an increasing number of models [1, 4, 13, 45, 64, 71] are not releasing their weights due to privacy and legal concerns [30, 38]. Therefore, one cannot use popular white-box fine-tuning strategies (such as LoRA [23] and Adapter [22]) that rely on model weights, feature embeddings, and output logits. Given that contemporary black-box VLMs [45, 47] like DALL-E [4, 52] still offer a language-based user interface and may be accessed through APIs that facilitate input and output in natural language, this allows users to customize these models through optimizing textual prompts. + +Manual prompting. Manual prompt engineering has been proven successful in adapting black-box LLMs to language tasks [25, 66]. Similarly, carefully crafted prompts can enhance the performance of VLMs. For instance, CLIP has demonstrated improved zero-shot recognition performance using specifically tailored prompts, such as "a photo of a {class}" for Internet photos and "a satellite image of a {class}" for satellite imagery. Despite its effectiveness, manual prompting can be a laborious process, inspiring efforts to explore automated prompt creation and thereby remove the need for human involvement. These strategies typically leverage an LLM as a knowledge base to create rich visual descriptors that augment the prompts for each class [40, 50] in a zero-shot fashion. + +Human-free prompting with conversational LLMs (our approach). We show how to effectively leverage chat-based LLMs [45] to emulate human-level prompt engineering without any human input. We first address an illustrative low-shot image classification task, aiming to find the best class-agnostic prompt (or "template") for image classification with CLIP. We start with a random set of prompts and evaluate the one-shot training accuracy of each. Then, akin to human prompt engineering, our method repeatedly presents ChatGPT with the best and worst prompts, asking it to review the results and suggest an improvement (see Figure 1). + +Learning with implicit "gradients" provided through conversational feedback. One of our key findings is that + +![](images/cc43833313de44e59df331a20d486ab4abda1512f0f8455752468f4af3832194.jpg) +Figure 1. Prompting VLMs using chat-based LLMs. Similar to how human prompt engineers iteratively test and refine prompts, we employ ChatGPT [45, 47] to continuously optimize prompts for vision-language models (VLMs). Our iterative approach assesses the performance of ChatGPT-generated prompts on a few-shot dataset (highlighted in blue) and provides feedback (marked in violet) to ChatGPT through simple conversations, as depicted in the illustrative figure. This straightforward method delivers state-of-the-art results for one-shot image classification across 11 datasets using CLIP, operated in a black-box manner without accessing model weights, feature embeddings, or output logits. We show that providing both positive (in green) and negative prompts (in red) enhances efficiency. Remarkably, our approach outperforms both white-box methods such as gradient-based continuous prompting (CoOp [74]) and human-engineered prompts [51] in this extremely low-shot scenario. This figure only shows a typical conversation using ChatGPT's web user interface. Our code implementation follows this pattern using the ChatGPT API. We detail and ablate the prompts in section 8. + +LLMs can learn the difference between effective and ineffective prompts, and can use this implicit "gradient" direction provided through language to perform more efficient searches. Compared to previous automatic prompting methods that only use LLMs as a knowledge base [40, 50] or paraphrasing tool [75], we show a novel use of LLMs as an optimizer that can utilize the patterns hidden in textual feedback. In our experiments, we find that the inclusion of such feedback greatly improves the efficiency and accuracy of our method, sometimes surpassing existing white-box methods [67, 74] on challenging one-shot scenarios. + +Optimizing text-to-image generation with DALL-E 3. We further demonstrate our optimization framework on a state-of-the-art black-box VLM, DALL-E [4], for two illustrative one-shot generative tasks: (1) Text-to-image (T2I) generation (see Figure 3), where we sample challenging text queries from Winoground [63] that involve reasoning over compositions of objects, attributes, and relations. Examples include "an animal watches a person" and "there is less milk than orange juice", which DALL-E 3 might initially + +fail to generate. (2) Prompt inversion (see Figure 4), which attempts to reverse-engineer the textual prompt to generate a specific image for later customization [53] (see Table 5). To achieve this, we leverage conversational feedback from a multimodal LLM (GPT4-V [45]) to iteratively refine the prompts based on the current generated images. We present qualitative results in Table 4 and conduct a user study to demonstrate that our framework can be more efficient than manual prompting, even for graphical designers experienced with AI content-generation tools. + +Our contributions. In this work, we introduce a novel prompting method for VLMs, utilizing an LLM as an optimizer. Our black-box approach can surprisingly compete with various white-box methods in a low-shot setting. Additionally, we extensively explore various strategies for conversing with ChatGPT, uncovering several key factors that significantly enhance the efficiency of this tool. We also show that our discovered natural language prompts are not only interpretable but also transfer better across CLIP architectures, eg., from RN50 to ViT/B-16, than continuous prompts discovered by previous white-box prompting + +method [74]. Finally, we show practical applications of our framework on text-to-image generation using black-box DALL-E 3. We release our code for future research on prompt optimization and AI-driven content creation1. + +# 2. Related Works + +LLMs for multimodal tasks. Cutting-edge LLMs like GPTs [45, 47] have been successfully applied to multimodal tasks, either through zero-shot composition with pre-trained multimodal models [29, 72] or by jointly fine-tuning with modality-specific encoders [1, 28] on large-scale multimodal datasets [56]. LLMs are also utilized as neuro-symbolic reasoners [16, 37, 58, 73], translating natural language instructions into modular programs (like Python code) that invoke APIs of multimodal models. In this work, we show the potential of LLMs as a black-box optimizer for multimodal foundation models with language interfaces, and more specifically vision-language models (VLMs). + +Prompt optimization of foundation models. Following the success of in-context learning [6], which appends user-generated natural language instruction and few-shot samples to text inputs, prompting [35] has emerged as the preferred fine-tuning paradigm for LLMs due to its superior performance and parameter-efficiency. However, recent prompt optimization methods, including continuous prefix-tuning [7, 31, 61, 62, 69] and discrete token-searching [11, 12, 59], still operate in a white-box manner, requiring access to either the tokenizer or output logits. Moreover, black-box prompting methods, such as heuristic-based editing [41, 49], are tailored towards language-only tasks and are thus not applicable in VLM settings. + +LLMs for prompt optimization. APE [75] leverages an LLM to automatically write prompts using few-shot samples based on instruction induction [21] and paraphrasing [42, 54]. However, it is only designed to address language tasks, while we focus on multimodal tasks using black-box VLMs. LLMs have also proven to be an effective external knowledge base [40, 50, 57] for generating prompts in a zero-shot setting for multimodal models. For example, DCLIP [40] uses GPT3 to come up with rich visual descriptions to improve zero-shot classification with CLIP [51]. We extend this line of work to show that LLMs can iteratively optimize prompts for VLMs in a black-box fashion given few-shot samples. We further illustrate that prompt optimization with LLMs can be made more efficient by leveraging conversational feedback, such as providing ChatGPT with explicit language feedback on how well the most recent prompt performs. Our findings align with the perspective [9] of LLMs as meta-optimizers that can implicitly perform gradient search through in-context learning. + +Few-shot adaptation of VLMs. Prompting has also been successfully adopted in VLMs [14], as demonstrated by methods like CoOp [74] that fine-tune an ensemble of continuous prefix tokens using cross-entropy loss. [34] achieves state-of-the-art few-shot performance with a cross-modal (image and text) cross-entropy loss. However, these methods all require access to model parameters for gradient backpropagation. We also note that while some concurrent works, such as BlackVIP [44] and LFA [46], claim to operate in a "black-box" setting, they still require access to privileged information including output logits and embeddings. In this work, we introduce a truly black-box and gradient-free approach that yields competitive results to white-box approaches in extremely low-shot scenarios. + +# 3. Prompting VLMs Using Chat-Based LLMs + +We now present our approach for prompting VLMs using chat-based LLMs as optimizers. + +Preliminaries. Motivated by recent proprietary VLMs [4, 45], we adopt a stricter yet practical black-box setting compared to prior works [44, 46], requiring minimal knowledge about the model's inner workings. This is crucial since releasing output logits or embeddings can potentially facilitate unauthorized knowledge extraction through distillation methods [19]. Our objective is to enhance the performance of a VLM equipped with a language interface capable of processing a textual prompt $p \in T$ . We assume that the targeted task is accompanied by a training dataset denoted as $D_{train} \subset D$ and its performance can be evaluated with respect to the prompt, represented as a function $F: D \times T \to \mathbb{R}$ . For example, in a classification task, $D_{train} = \{x, y\}_n$ where $x$ is an image and $y$ is its class label. The black-box VLM takes the image as input and returns a predicted label. We measure the performance of the textual prompt by calculating the average classification accuracy as $F(D_{train}, p)$ . Our goal in prompt engineering is to search for the optimal prompt $p^*$ without accessing or modifying the black-box VLM. + +Background: human prompt engineering. Our method draws inspiration from the typical workflow of human prompt engineers. Prompt engineering is often an iterative process that involves: (a) creating an initial prompt $U = \{p_1\}$ based on the understanding of a task, (b) evaluating the performance of prompts in $U$ , (c) refining prompts based on the outcomes, (d) repeating the last two steps until convergence, and (e) returning the prompt $p^*$ with the highest $F(D_{train}, p^*)$ . This hands-on approach helps optimize the model's performance, but it can be tedious and labor-intensive. Algorithm 1 formally illustrates this process. + +Example: prompting for image classification with CLIP [51]. CLIP is one of the most popular VLM that takes a set of class-specific prompts when performing "zero-shot" image classification. [51] details the laborious prompting procedure over the course of a year. Interestingly, they find + +Algorithm 1 We formalize human prompt engineering with the following algorithm, which motivates our LLM-based algorithm (2). +Require: $D_{\mathrm{train}} = \{x,y\}_{n}$ training samples, $F:D\times T\to$ R: evaluation function +1: Create initial prompts: $\mathcal{U}\gets \{p_1\}$ +2: Evaluate prompts on training set: $S\gets \{F(D_{\mathrm{train}},p_1)\}$ +3: while not converged do +4: Generate a new prompt $p^\prime$ based on $S$ +5: Evaluate the new prompt: $s^{\prime} = F(D_{\mathrm{train}},p^{\prime})$ +6: $\mathcal{U}\leftarrow \mathcal{U}\cup \{p^{\prime}\}$ +7: $S\gets S\cup \{s^{\prime}\}$ +8: end while +9: return optimal prompt $p^*\gets \arg \max_{p\in \mathcal{U}}F(D_{\mathrm{train}},p)$ + +that a default class-agnostic prompt (or so-called "template"), "a photo of a {class}" can provide a decent boost in accuracy for most datasets compared to using vanilla class labels. In this scenario, the evaluation function $F$ is the classification accuracy on the test set, and the prompt $p = \{ \text{"a photo of a {class}" } | c \in C \}$ , where $C$ is the set of class names for a given dataset. + +Prompting with chat-based LLMs (our approach). Given the strong in-context reasoning capabilities of LLMs, we envision them as a black-box optimizers that can improve prompts based on their performance outcomes, akin to how human prompt engineers iteratively refine prompts. Specifically, we maintain a pool of prompts $U$ and their corresponding performance outcomes $S$ . In each iteration, we provide the LLM with both positive and negative prompts, such as the highest and lowest-performing candidates. Such textual feedback through in-context prompts offers LLMs an implied "gradient" direction [9], making optimization more efficient than taking random local steps. We facilitate this feedback mechanism through conversations with state-of-the-art chat-based LLMs like ChatGPT [47] as illustrated in Figure 1. We note that such a multi-turn conversation is not the only way of conversing with ChatGPT, and ablate different in-context feedback mechanisms in section 8. + +# 4. Illustrative Few-Shot Classification Task + +We illustrate our approach using a few-shot image classification task. Specifically, a prompt $p \in T$ consists of a set of class-specific prompts – that is, one textual description per class. The evaluation function $F$ takes the prompt $p$ , along with an image dataset $D_{train}$ , and returns the accuracy using the black-box VLM. To prevent overfitting and simplify our search space, we restrict our search to finding a single class-agnostic template, e.g., a photo of a {}, filling in the blank with label names provided with the dataset. + +Algorithm 2 LLM-based prompt engineering on the illustrative classification task. Our algorithm requires a chat-based LLM and a (black-box) evaluation function, such as accuracy. We highlight mechanisms for “exploration” (restart and reset) in blue and “exploitation” (iter) in red. We mark the key component of “conversational feedback” of our approach in violet. The actual prompts are attached in section 8. + +Require: $D_{\mathrm{train}} = \{x,y\}_{n}$ training samples, $F:D\times T\to$ R: evaluation function. +Require: $n_{\mathrm{restart}}$ : number of initial sampled prompt sets, $n_{\mathrm{reset}}$ : number of resets for a prompt set, $n_{\mathrm{iter}}$ : number of hill-climbing iterations, $m$ size of one initial prompt set, $k$ : number of prompts send to ChatGPT. +1: $p^* \gets \emptyset$ +2: for 1:: $n_{\mathrm{restart}}$ do +3: Sample a new prompt set, $\mathcal{U}_{\mathrm{init}} \leftarrow \{p_1,\dots,p_m\}$ +4: for 1:: $n_{\mathrm{reset}}$ do +5: Reset to initial prompt set: $\mathcal{U} \leftarrow \mathcal{U}_{\mathrm{init}}$ +6: for 1:: $n_{\mathrm{iter}}$ do +7: Sort $\mathcal{U}$ by score outcomes $\{F(D_{\mathrm{train}},p)\}_{p\in U}$ +8: $\mathcal{U}_{\mathrm{top}} \leftarrow$ top-k prompts in $\mathcal{U}$ +9: $\mathcal{U}_{\mathrm{bot}} \leftarrow$ bottom-k prompts in $\mathcal{U}$ +10: Get a new prompt $p_{\mathrm{new}} \leftarrow$ LLM( $\mathcal{U}_{\mathrm{top}},\mathcal{U}_{\mathrm{bot}}$ ) +11: $\mathcal{U} \leftarrow \mathcal{U} \cup \{p_{\mathrm{new}}\}$ +12: end for +13: $p^* \leftarrow$ arg max $p\in \mathcal{U}\cup \{p^*\}$ $F(D_{\mathrm{train}},p)$ +14: end for +15: end for +16: return prompt with highest score $p^*$ + +Outline of our approach (Alg. 2). To start, we sample entirely random initial prompts from a text corpus such as LAION-COCO [55] captions. Our approach follows the classical stochastic hill-climbing framework with random restart [54], which prevents ChatGPT from being trapped in local optima by balancing "exploration" and "exploitation". Our restart mechanism is implemented by sampling $n_{\mathrm{restart}}$ initial prompt sets to encourage exploration. Because ChatGPT performs stochastic top-k sampling for text generation (as we adopt the default temperature of 1.0), we also implement a reset mechanism to foster additional exploration by retrying a given prompt set $n_{\mathrm{reset}}$ times. For exploitation, we converse with ChatGPT for $n_{\mathrm{iter}}$ iterations. We find that it is critical to balance exploration and exploitation for optimal performance, and thoroughly examine this trade-off in section 9. Lastly, we present ChatGPT both the top and bottom-performing prompts, denoted as $(\mathcal{U}_{top}, \mathcal{U}_{bot})$ . We show that this simple adjustment can improve the efficiency of our approach in Figure 2. + +Experimental setup. We apply our approach to the few-shot image classification benchmark introduced in CoOp [74], which is the most commonly studied setup for + +fine-tuning VLMs. This benchmark involves a collection of 11 datasets covering diverse image domains including ImageNet [10] and more niche datasets such as FGVC-Aircraft [39]. For each dataset, we adhere to the same three-fold k-shot train sets in [34], reporting the average accuracy across all folds. Importantly, our method only utilizes the train set to compute the score and does not require the few-shot validation set. We use CLIP following prior work [34, 74] to emulate a black-box VLM, and we employ ChatGPT (GPT3.5) as the chat-based LLM. + +Implementation details. To start, we sample entirely random 1M initial prompts from a text corpus (LAION-COCO [55] captions). For each caption, we extract all the noun phrases using spaCy part-of-speech tagging [20]. Subsequently, we replace one noun phrase in the caption with $\{\}$ (a placeholder where the class name will be inserted) to create a template. Given that each caption contains an average of 2 noun phrases, our initial prompt pool consists of approximately 2M templates. We run our algorithm with $n_{\mathrm{start}} = 20$ restarts, $n_{\mathrm{resets}} = 50$ resets, and $n_{\mathrm{iter}} = 10$ iterations. We opt to sample $m = 100$ prompts per restart and present the top and bottom $k = 15$ prompts to ChatGPT. We ablate different sets of hyperparameters and explain how we balance the tradeoff between exploration and exploitation in section 9. We adopt gpt-3.5-turbo-0301 model for ChatGPT using OpenAI's official API and keep the default sampling temperature of 1.0. We also ablate gpt-4 in Table 10 and find it achieves similar performance. The exact prompts used to converse with ChatGPT are documented in section 8. For a fair comparison, we use CLIP-RN50 for our experiments following prior work [34, 74]. We will open-source our code and release the initial prompt pool (LAIONCOCO-1M) to the public. + +Oracle white-box baselines. Our black-box setup substantially differs from, and is more constrained than, the scenarios considered in previous white-box baselines. Specifically, we do not expose the pre-trained weights, model architectures, feature embeddings, or even output logits of VLMs. These constraints render many established gradient-based fine-tuning baselines inapplicable. Among the oracle white-box approaches we later compare to, CoOp [74] performs continuous prompting and requires backpropagation across all layers. WiSE-FT [67] ensembles fine-tuned weights with the original CLIP weights. Cross-Modal Adaptation [34] fine-tunes a linear classifier leveraging both image and text embeddings from CLIP. BlackVIP [44] and LFA [46] are two most recent baselines that apply CLIP logits or embeddings for gradient back-propagation. Finally, while DCLIP [40] queries GPT3 for rich visual descriptors for each class and does not require gradient-based fine-tuning, it performs prompt assembling using 4-6 class-specific prompts, which breaches our black-box assumption for accessing the output logits. + +Black-box methods. We additionally benchmark our method against truly black-box solutions, including the vanilla class-agnostic templates “{class)” and “a photo of a {class}”. Also, we compare our approach to the best Hand-Engineered templates released by OpenAI, searched using test set performance to represent the theoretical upper bound of human performance, eg., “a centered satellite photo of {class}.” for EuroSAT [17]. Finally, we present two versions of conversational feedback of our approach: (a) using 30 positive (P only) or (b) using 15 positive and 15 negative prompts $(\mathbf{P} + \mathbf{N})$ in each iteration. For a fair comparison, both of our approaches start with the same initial sampled prompts, referred to as LAIONCOCO-1M. We also show the performance of the best initial sampled prompt searched using trainset performance. + +SOTA one-shot performance against existing methods on 11 datasets. We report the test set performance of our method versus the aforementioned baselines in a challenging 1-shot classification scenario in Table 1. First, compared to the top-performing initial prompts selected from LAIONCOCO-1M based on train set performance, our prompt optimization using ChatGPT notably improves the initial prompts by an average of $5\%$ (56% to 61%). Remarkably, our black-box approach surpasses the two white-box gradient-based fine-tuning techniques CoOp and WiSE-FT by at least $1.5\%$ . Given that both CoOp and our method optimize a single class-agnostic template, we attribute this gap in performance to reduced overfitting. More specifically, we posit that our optimization space of natural language effectively acts as a regularizer in extremely low-shot tasks, standing as a more robust alternative to the continuous prompting approach of CoOp. Furthermore, our method benefits from textual feedback and shows improved performance by $1.0\%$ when using both positive and negative prompts. In section 9, we show that our approach remains effective across different CLIP and ChatGPT variants. + +Incorporating negative prompts leads to more efficient optimization. In Figure 2, we demonstrate that incorporating both positive and negative prompts fosters better optimization efficiency, achieving higher accuracy within a much fewer number of resets. Specifically, we hypothesize that LLMs can leverage the implicit "gradient" direction suggested in textual feedback to achieve faster convergence. For additional analysis, we ablate different ways of providing conversational feedback to ChatGPT in section 8 and conclude that iteratively updating both positive and negative prompts is the key for efficient optimization. + +# 5. More Benefits of Natural Language Prompts + +In this section, we delve deeper into the advantages of utilizing natural language prompts compared to the continuous prompts [74]. We highlight that the prompts derived through + +
MethodDatasetAvg
CaltechImageNetAircraftFoodPetsCarsSUNUCFDTDEuroSATFlowers
Oracle white-box approaches
Cross-Modal [34]89.161.620.677.185.759.063.464.749.961.876.364.7
WiSE-FT [67]85.558.318.671.981.755.756.659.444.252.365.859.1
CoOp [74]87.557.29.674.385.955.660.361.944.450.668.159.6
LFA [46]81.652.417.063.175.341.458.456.738.460.774.956.4
BlackVIP [44]85.858.815.376.785.256.457.058.840.130.061.156.8
DCLIP [40]-59.6-76.483.8---41.734.7--
Manual prompting approaches
{}78.555.315.574.078.952.253.455.541.432.157.354.0
a photo of a {}84.557.915.974.083.253.958.056.938.828.660.255.6
Hand-Engineered [51]86.358.217.377.385.855.658.561.542.337.666.158.8
Our black-box approaches
LAIONCOCO-1M81.456.217.476.579.651.354.955.843.138.661.356.0
Ours (P only)89.059.417.977.885.755.760.458.743.646.766.660.1
Ours (P+N)89.159.618.178.388.156.261.060.244.849.067.261.1
+ +Table 1. Comparison of our method with other baselines on one-shot classification tasks. We report the average accuracy of each method across three folds, optimized using 1-shot training sets. We bold the best black-box result for each dataset, and underline the second best result. First, we note that our approach can effectively improve upon the initial prompts selected from LAIONCOCO-1M from $56\%$ to $61\%$ . Our approach is also competitive against the best Human-Engineered prompts released by OpenAI [51] searched using test set performance. Additionally, we show that using both positive and negative prompts improves the overall accuracy by $1\%$ . For reference, we report oracle white-box approaches in gray. Remarkably, we also surpass white-box solutions such as WiSE-FT [67] and CoOp [74] by $1.5\%$ . These methods require either gradient-based fine-tuning (CoOp/WiSE-FT/Cross-Modal) or prompt ensembling using output logits (DCLIP). While our approach is less effective than the SOTA white-box method (Cross-Modal Adaptation), we stress that our black-box setup is significantly more challenging, because we restrict the optimization space to natural language and do not access the pre-trained weights, model architectures, feature embeddings, and output logits of VLMs. + +![](images/984ce6c2fc4524419f5608bf8d5dcb04a71044111e8c5f49adc33ff86558f74f.jpg) +Figure 2. Conversational feedback incorporating both positive and negative prompts leads to improved efficiency. We fix the number of restarts to 20 and iterations to 10, and ablate different numbers of resets on all 11 datasets (left) and ImageNet (right). Notably, our approach using “ $\mathrm{P} + \mathrm{N}$ ” (both top-15 and bottom-15 prompts) can optimize faster within a much fewer number of resets than using “P-Only” (top-30 prompts), resulting in the highest overall performance. + +![](images/d848393515ce207d267624f6e16fe8f7a8264c37a7dc772585090ac716b31b14.jpg) + +our method are interpretable; for instance, they often contain descriptions of the targeted image domain. Our prompts can also transfer across CLIP architectures in a black-box manner, such as from RN50 to ViT/B-16. + +Interpretable natural language prompts. While CoOp [74] concedes that continuous prompts can be difficult to interpret, our method – without explicitly instructing ChatGPT to generate interpretation – often yields interpretable results. Table 2 showcases the templates returned by our algorithm for each dataset, frequently including keywords that reflect the targeted image domain. For example, the template + +for Food101 [5] mentions "diverse cuisine and ingredients", and the template for UCF101 [60] (an action recognition dataset) mentions "in motion". Likewise, these templates identify general stylistic attributes of the datasets; they refer to "bright and natural lighting" for ImageNet [10] and note images that "emphasize the subject" for Caltech101 [27]. These prompts are particularly intriguing because we do not provide ChatGPT with any information about the downstream task, yet it manages to generate prompts containing domain-specific keywords that are similar to those engineered by human experts. + +Black-box prompt transfer. Our text prompts also maintain consistently high performance across different CLIP backbones. For comparison, since CoOp uses the same tokenizer for all CLIP architectures (including RN50, RN101, ViT/B-32, and ViT/B-16) and optimizes continuous prompts of the same shape (16 x 512), we assess the transferability of these learned continuous prompts from RN50 to other backbones using the official weights on 16-shot ImageNet. Table 3 showcases the results of this experiment, where we also include the baseline prompt "a photo of a {}" for reference. We observe a significant decline in accuracy when transferring CoOp's prompts (up to a $40\%$ decrease despite utilizing more powerful backbones), implying that continuous prompts tend to overfit to the specific CLIP model. In contrast, our natural language prompts maintain their performance and outperform the baseline across all backbones. + +
DatasetExample of Top Templates
Caltech [27]An image of a {} with a blurred background that emphasizes the subject
DTD [8]The essential elements of {} are amplified with visual simplicity
EuroSAT [17]A top-down view of {} arranged in a pattern {}
Aircraft [39]A clear, high-quality image of a single {} with a white background
Food [5]A {} featuring diverse cuisine and ingredients
ImageNet [10]An image of a {} with bright and natural lighting
Flowers [43]A clear and vivid photograph of the {} in its natural setting
Pets [48]A {} with distinct and recognizable characteristics
Cars [26]A {} featuring a wide range of color options for easy selection
SUN [68]A high-resolution photo of a {} with clear background and natural lighting
UCF [60]A black and white photo of a {} in motion
+ +Table 2. Example templates returned by our algorithm on each dataset. Although we do not provide ChatGPT with any information regarding the targeted dataset, we observe that the resulting templates are remarkably similar to human-engineered templates, with many domain-specific details such as "motion" and "cuisine", and stylistic elements such as "bright and natural lighting". + +
MethodRN50→RN101→ViT-B/32→ViT-B/16
a photo of a {}57.960.661.966.6
CoOp63.020.631.739.5
Ours59.960.762.267.0
+ +Table 3. Black-box prompt transfer from ResNet-50 to other CLIP architectures. We evaluate both our natural language prompts and CoOp's continuous prompts on 16-shot ImageNet, which are trained using the RN50 CLIP backbone. As a reference point, we include the baseline prompt "a photo of a $\{\}$ ", and show that the prompts derived from our method using RN50 consistently surpass it after transferring to different backbones. In contrast, while CoOp achieves better 16-shot ImageNet performance using RN50, its performance plummets during the transfer, e.g., from $63\%$ to a mere $21\%$ for RN101. + +# 6. Application: Text-to-Image Generation + +![](images/1a25d3c154e797990230cc64699ff124918931546f0a783046bca6c20a42ffac.jpg) +Figure 3. Improving text-to-image (T2I) generation using chat-based multimodal LLMs. Our framework can use the multimodal GPT4-V [45] to optimize prompts for the state-of-the-art black-box generative VLM, DALL-E 3 [4]. For complex user queries that DALL-E 3 may initially fail to generate, we send the generated image (in violet) along with the current prompt to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue). Our simple framework is surprisingly effective at correcting DALL-E 3 mistakes on some challenging Winoground [63] text queries that involve action, logical, and spatial reasoning. We conduct a human evaluation on the quality of generated images in Table 6 and include the actual prompts in section 8. + +In this section, we present a direct application of our prompt optimization framework to generative tasks using a truly black-box text-to-image (T2I) VLM, DALL-E 3 [4]. + +Optimizing T2I using a multimodal LLM. DALL-E 3 can generate high-fidelity images following diverse user queries, but crafting effective prompts is tricky even for designers experienced with AI content generation tools [36]. Therefore, we are motivated to implement our LLM-based optimization framework to assist with creative visual design. Our framework is shown in Figure 3 for the illustrative task of text-to-image generation. In this task, the user specifies a query (topic) in text, such as "an animal watches a person", and the goal is to write a prompt that can generate an image reflecting this topic. We adopt a multimodal LLM GPT4-V [45] (gpt-4-1106-preview) to provide feedback on the generated image and optimize the prompt. We find that this framework is surprisingly effective due to GPT4-V's strong visual reasoning capabilities, which can often spot subtle errors in generated images and offer more accurate prompts. + +![](images/2afe40db62a631c711c0caf5cb07baf88a646951f2afcb4573f046aaed48770d.jpg) +Figure 4. Prompt inversion using chat-based multimodal LLMs. We apply our framework to reverse engineer the text prompt to generate the same user-queried image. We send the generated image (in violet) along with the original image to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue). + +Task setup. For T2I generation, we experiment with a + +
User QueryInit. ImageLLM FeedbackFinal Image
Text-to-image generation
There is less milk than orange juice.Incorrect, the milk bottle appears full, more than orange juice...
A shorter person is covering the eyes of a taller person.Incorrect, the taller person is covering the shorter person's eyes. Instead, ...
Prompt inversion
The scarf should feature red and white stripes, and the fur is fluffy...
The coat should be buttoned and the lighting exhibits a stronger contrast...
+ +Table 4. Examples of T2I optimization. We show that our framework (Figure 3) can automatically improve the faithfulness of images generated by DALL-E 3, with respect to user-specified textual topics (for T2I generation) or reference images (for prompt inversion). This is achieved through three rounds of prompt optimization, using feedback from the multimodal LLM (GPT4-V). Table 11 and Table 12 shows more examples with actual prompts. + +
User QueryInverted ImageExample 1Example 2Example 3Example 4Example 5
Give the dog a cat friend.Make the dog be in the middle of a jump.Make the dog do a handstand.Make the dog lie down on its side.Make the dog swim in water.
Make the owl fight a hawk.Make the owl flap its wings.Make the owl fully green.Make the owl stand in front of the moon.Make the owl walk in the city.
+ +subset of 100 text queries from Winoground [63] that involve complex attribute and relation reasoning, which DALL-E might initially fail to generate. Our framework refines the prompts to capture the user-specified topics using a few (three) iterations. We also attempt a reverse task of prompt inversion: given a user-specified reference (query) image, our framework reverse-engineers the prompt to have DALL-E generate the same object or scene in the query image (see Figure 4). This enables users to easily make customizations [53] (see Table 5), such as having the character in a reference image perform various actions or change scenes. For this task, we sample 100 random queries from DiffusionDB [65]. We provide qualitative results in Table 4, Table 11, and Table 12. We hire two volunteers to assess the faithfulness of the images generated by our method, and to compare these with the images manually prompted by two + +designers (each with one year of experience in AI content generation), as shown in Table 6. + +Remarks on limitations. While we show promising results, we note some failure cases in Table 13 and Table 14 due to the inherent limitations of foundation models. For example, GPT4-V might fail to describe abstract and artistic details, and DALL-E 3 often fails to generate the correct number of objects. We believe that our framework can benefit from more capable foundation models in the future. + +Table 5. Customization via prompt inversion. Users can simply append extra descriptions to the inverted prompts to customize their main characters in queried images. + +
TaskMethodInit. (std)Final (std)Δ
Text-to-ImageHuman2.28 (.45)2.86 (.61)0.58
Ours2.62 (.36)3.56 (.54)0.94
Prompt InversionHuman1.58 (.48)2.76 (.53)1.18
Ours1.94 (.39)3.68 (.47)1.74
+ +Table 6. Our method enhances faithfulness in T2I generation. We hire two human annotators to assess the faithfulness of images generated from user queries, e.g., textual topics for Text-to-Image, or reference images for Prompt Inversion. The scores are measured on a 1-to-5 Likert scale, with 1 signifying contradiction and 5 indicating perfect alignment with the user's goal. Our approach benefits from three iterations of prompt optimization and consistently outperforms human-engineered prompts by designers who have one year of experience in AI content generation. + +# 7. Discussion and Limitations + +Summary. We present the first attempt to leverage LLMs as prompt engineers for VLMs. On one-shot image classification, our method surpasses existing human-engineered prompts and even rivals white-box approaches. Central to the success of our method is the use of conversational feedback, enabling chat-based LLMs to efficiently steer VLMs in the right direction. This process leads to a set of interpretable prompts bearing considerable resemblance to those crafted by humans. Importantly, our natural language prompting setup is a lot more constrained than the assumed scenarios of previous white-box or even some black-box settings [44], because we do not expose the model weights and outputs of VLMs. We finally apply our framework to illustrative generative tasks using a truly black-box text-to-image VLM (DALL-E 3). + +Limitations and future work. While we try to minimize the overall cost and the total number of API calls, the energy consumption associated with LLMs remains a substantial concern. It is vital to note that we do not intend to compete directly with white-box baselines that can improve visual and text representations with more data. Further details on the higher-shot performance of our method can be found in section 9. Lastly, we are limited to costly human evaluation for T2I generation in this study. Future work may adopt automatic evaluation [18, 33] for large-scale experiments. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1 +[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Joyce Zhuang, Juntang and Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiang, and Aditya Ramesh. Improving image generation with better captions. Note on Dalle-3, 2023. 1, 2, 3, 7 +[5] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6, 7 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[7] Yekun Chai, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Clip-tuning: Towards derivative-free prompt learning with a mixture of rewards. arXiv preprint arXiv:2210.12050, 2022. 3 +[8] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 7 +[9] Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Zhifang Sui, and Furu Wei. Why can gpt learn in-context? language models secretly perform gradient descent as meta optimizers. ACL, 2023. 3, 4 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 1, 5, 6, 7 +[11] Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P. Xing, and Zhiting Hu. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548, 2022.3 +[12] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and + +Tong Zhang. Black-box prompt learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022. 3 +[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1 +[14] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, and Jianfeng Gao. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 3 +[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1 +[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. arXiv preprint arXiv:2211.11559, 2022. 3 +[17] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification, 2017. 5, 7 +[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8 +[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[20] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 5 +[21] Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.3 +[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1 +[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 1 +[25] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1 +[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In + +4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 7 +[27] Li, Andreeto, Ranzato, and Perona. Caltech 101, 2022. 6, 7 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3 +[29] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pre-trained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022.3 +[30] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1 +[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 +[33] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 8 +[34] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramana. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. arXiv preprint arXiv:2301.06267, 2023. 1, 3, 5, 6 +[35] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 3 +[36] Vivian Liu. Beyond text-to-image: Multimodal prompts to explore generative ai. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-6, 2023. 7 +[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3 +[38] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1 +[39] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5, 7 +[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 1, 2, 3, 5, 6 + +[41] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830, 2021. 3 +[42] Melanie Mitchell, John Holland, and Stephanie Forrest. When will a genetic algorithm outperform hill climbing. Advances in neural information processing systems, 6, 1993. 3 +[43] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7 +[44] Changdae Oh, Hyeji Hwang, Hee-young Lee, YongTaek Lim, Geunyoung Jung, Jiyoung Jung, Hosik Choi, and Kyungwoo Song. Blackvip: Black-box visual prompting for robust transfer learning. In CVPR, 2023. 3, 5, 6, 8 +[45] OpenAI. Gpt-4 technical report. 2023. 1, 2, 3, 7 +[46] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In ICCV, 2023. 3, 5, 6 +[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 1, 2, 3, 4 +[48] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 7 +[49] Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. Grips: Gradient-free, edit-based instruction search for prompting large language models. arXiv preprint arXiv:2203.07281, 2022. 3 +[50] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2023. 1, 2, 3 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6, 13 +[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1 +[53] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500–22510, 2023. 2, 8 +[54] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010. 3, 4 +[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 4, 5 + +[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 3 +[57] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Zhe Gan, Lijuan Wang, Lu Yuan, Ce Liu, et al. K-lite: Learning transferable visual models with external knowledge. Advances in Neural Information Processing Systems, 35:15558-15573, 2022. 3 +[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3 +[59] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3 +[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6, 7 +[61] Tianxiang Sun, Zhengfu He, Hong Qian, Yunhua Zhou, Xuan-Jing Huang, and Xipeng Qiu. Bbtv2: Towards a gradient-free future with large language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3916-3930, 2022. 3 +[62] Tianxiang Sun, Yunfan Shao, Hong Qian, Xuanjing Huang, and Xipeng Qiu. Black-box tuning for language-model-as-a-service. In International Conference on Machine Learning, pages 20841-20855. PMLR, 2022. 3 +[63] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 7, 8 +[64] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1 +[65] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 8, 15 +[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35: 24824-24837, 2022. 1 +[67] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. arXiv preprint arXiv:2109.01903, 2021. https://arxiv.org/abs/2109.01903.2,5,6 + +[68] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. Int. J. Comput. Vision, 119(1): 3-22, 2016. 7 +[69] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. Gps: Genetic prompt search for efficient few-shot learning. arXiv preprint arXiv:2210.17041, 2022.3 +[70] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1 +[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1 +[72] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3 +[73] Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023. 3 +[74] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1, 2, 3, 4, 5, 6 +[75] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. *ICLR*, 2023. 2, 3, 13, 14 \ No newline at end of file diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/images.zip b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eb66df049088d942cd02bee0306ab0ebe7488952 --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:075c2587a015455ad33502edb8a942158579d6bdd888fc2d7c14eea7ed4d7911 +size 434400 diff --git a/2024/Language Models as Black-Box Optimizers for Vision-Language Models/layout.json b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bcb0b0ee214d0eb4c6aca158147198237a21112f --- /dev/null +++ b/2024/Language Models as Black-Box Optimizers for Vision-Language Models/layout.json @@ -0,0 +1,7608 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 520, + 121 + ], + "type": "text", + "content": "Language Models as Black-Box Optimizers for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 143, + 485, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 143, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 143, + 485, + 186 + ], + "type": "text", + "content": "Shihong Liu* Samuel Yu* Zhiqiu Lin* Deepak Pathak Deva Ramanan Carnegie Mellon University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 586 + ], + "type": "text", + "content": "Vision-language models (VLMs) pre-trained on web-scale datasets have demonstrated remarkable capabilities on downstream tasks when fine-tuned with minimal data. However, many VLMs rely on proprietary data and are not open-source, which restricts the use of white-box approaches for fine-tuning. As such, we aim to develop a black-box approach to optimize VLMs through natural language prompts, thereby avoiding the need to access model parameters, feature embeddings, or even output logits. We propose employing chat-based LLMs to search for the best text prompt for VLMs. Specifically, we adopt an automatic \"hill-climbing\" procedure that converges to an effective prompt by evaluating the performance of current prompts and asking LLMs to refine them based on textual feedback, all within a conversational process without human-in-the-loop. In a challenging 1-shot image classification setup, our simple approach surpasses the white-box continuous prompting method (CoOp) by an average of " + }, + { + "bbox": [ + 46, + 238, + 290, + 586 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 238, + 290, + 586 + ], + "type": "text", + "content": " across 11 datasets including ImageNet. Our approach also outperforms both human-engineered and LLM-generated prompts. We highlight the advantage of conversational feedback that incorporates both positive and negative prompts, suggesting that LLMs can utilize the implicit \"gradient\" direction in textual feedback for a more efficient search. In addition, we find that the text prompts generated through our strategy are not only more interpretable but also transfer well across different VLM architectures in a black-box manner. Lastly, we demonstrate our framework on a state-of-the-art black-box VLM (DALL-E 3) for text-to-image optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 128, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Vision-language models [1, 28, 51, 64] (VLMs) excel at a wide range of classic vision and multimodal [2, 10, 15, 32, 70] tasks, surpassing the performance of their fully-supervised counterparts on downstream tasks even when fine-tuned with minimal data [34, 74]. However, fine-tuning VLMs typically requires transparent white-box access to the model weights, such as gradient-based approaches that rely" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 214, + 390, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 390, + 226 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 390, + 226 + ], + "type": "text", + "content": "on backpropagation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 548, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 548, + 384 + ], + "type": "text", + "content": "VLMs as black-box services. Despite community efforts to collect web-scale public datasets [55, 56] and to replicate proprietary VLMs [3, 24], an increasing number of models [1, 4, 13, 45, 64, 71] are not releasing their weights due to privacy and legal concerns [30, 38]. Therefore, one cannot use popular white-box fine-tuning strategies (such as LoRA [23] and Adapter [22]) that rely on model weights, feature embeddings, and output logits. Given that contemporary black-box VLMs [45, 47] like DALL-E [4, 52] still offer a language-based user interface and may be accessed through APIs that facilitate input and output in natural language, this allows users to customize these models through optimizing textual prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 386, + 548, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 548, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 548, + 553 + ], + "type": "text", + "content": "Manual prompting. Manual prompt engineering has been proven successful in adapting black-box LLMs to language tasks [25, 66]. Similarly, carefully crafted prompts can enhance the performance of VLMs. For instance, CLIP has demonstrated improved zero-shot recognition performance using specifically tailored prompts, such as \"a photo of a {class}\" for Internet photos and \"a satellite image of a {class}\" for satellite imagery. Despite its effectiveness, manual prompting can be a laborious process, inspiring efforts to explore automated prompt creation and thereby remove the need for human involvement. These strategies typically leverage an LLM as a knowledge base to create rich visual descriptors that augment the prompts for each class [40, 50] in a zero-shot fashion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 555, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 548, + 688 + ], + "type": "text", + "content": "Human-free prompting with conversational LLMs (our approach). We show how to effectively leverage chat-based LLMs [45] to emulate human-level prompt engineering without any human input. We first address an illustrative low-shot image classification task, aiming to find the best class-agnostic prompt (or \"template\") for image classification with CLIP. We start with a random set of prompts and evaluate the one-shot training accuracy of each. Then, akin to human prompt engineering, our method repeatedly presents ChatGPT with the best and worst prompts, asking it to review the results and suggest an improvement (see Figure 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "Learning with implicit \"gradients\" provided through conversational feedback. One of our key findings is that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12687" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 69, + 500, + 340 + ], + "blocks": [ + { + "bbox": [ + 98, + 69, + 500, + 340 + ], + "lines": [ + { + "bbox": [ + 98, + 69, + 500, + 340 + ], + "spans": [ + { + "bbox": [ + 98, + 69, + 500, + 340 + ], + "type": "image", + "image_path": "cc43833313de44e59df331a20d486ab4abda1512f0f8455752468f4af3832194.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 348, + 548, + 449 + ], + "lines": [ + { + "bbox": [ + 46, + 348, + 548, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 548, + 449 + ], + "type": "text", + "content": "Figure 1. Prompting VLMs using chat-based LLMs. Similar to how human prompt engineers iteratively test and refine prompts, we employ ChatGPT [45, 47] to continuously optimize prompts for vision-language models (VLMs). Our iterative approach assesses the performance of ChatGPT-generated prompts on a few-shot dataset (highlighted in blue) and provides feedback (marked in violet) to ChatGPT through simple conversations, as depicted in the illustrative figure. This straightforward method delivers state-of-the-art results for one-shot image classification across 11 datasets using CLIP, operated in a black-box manner without accessing model weights, feature embeddings, or output logits. We show that providing both positive (in green) and negative prompts (in red) enhances efficiency. Remarkably, our approach outperforms both white-box methods such as gradient-based continuous prompting (CoOp [74]) and human-engineered prompts [51] in this extremely low-shot scenario. This figure only shows a typical conversation using ChatGPT's web user interface. Our code implementation follows this pattern using the ChatGPT API. We detail and ablate the prompts in section 8." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 457, + 290, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 290, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 290, + 589 + ], + "type": "text", + "content": "LLMs can learn the difference between effective and ineffective prompts, and can use this implicit \"gradient\" direction provided through language to perform more efficient searches. Compared to previous automatic prompting methods that only use LLMs as a knowledge base [40, 50] or paraphrasing tool [75], we show a novel use of LLMs as an optimizer that can utilize the patterns hidden in textual feedback. In our experiments, we find that the inclusion of such feedback greatly improves the efficiency and accuracy of our method, sometimes surpassing existing white-box methods [67, 74] on challenging one-shot scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Optimizing text-to-image generation with DALL-E 3. We further demonstrate our optimization framework on a state-of-the-art black-box VLM, DALL-E [4], for two illustrative one-shot generative tasks: (1) Text-to-image (T2I) generation (see Figure 3), where we sample challenging text queries from Winoground [63] that involve reasoning over compositions of objects, attributes, and relations. Examples include \"an animal watches a person\" and \"there is less milk than orange juice\", which DALL-E 3 might initially" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 457, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 578 + ], + "type": "text", + "content": "fail to generate. (2) Prompt inversion (see Figure 4), which attempts to reverse-engineer the textual prompt to generate a specific image for later customization [53] (see Table 5). To achieve this, we leverage conversational feedback from a multimodal LLM (GPT4-V [45]) to iteratively refine the prompts based on the current generated images. We present qualitative results in Table 4 and conduct a user study to demonstrate that our framework can be more efficient than manual prompting, even for graphical designers experienced with AI content-generation tools." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Our contributions. In this work, we introduce a novel prompting method for VLMs, utilizing an LLM as an optimizer. Our black-box approach can surprisingly compete with various white-box methods in a low-shot setting. Additionally, we extensively explore various strategies for conversing with ChatGPT, uncovering several key factors that significantly enhance the efficiency of this tool. We also show that our discovered natural language prompts are not only interpretable but also transfer better across CLIP architectures, eg., from RN50 to ViT/B-16, than continuous prompts discovered by previous white-box prompting" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12688" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "method [74]. Finally, we show practical applications of our framework on text-to-image generation using black-box DALL-E 3. We release our code for future research on prompt optimization and AI-driven content creation1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 136, + 138, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 136, + 138, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 136, + 138, + 148 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 302 + ], + "type": "text", + "content": "LLMs for multimodal tasks. Cutting-edge LLMs like GPTs [45, 47] have been successfully applied to multimodal tasks, either through zero-shot composition with pre-trained multimodal models [29, 72] or by jointly fine-tuning with modality-specific encoders [1, 28] on large-scale multimodal datasets [56]. LLMs are also utilized as neuro-symbolic reasoners [16, 37, 58, 73], translating natural language instructions into modular programs (like Python code) that invoke APIs of multimodal models. In this work, we show the potential of LLMs as a black-box optimizer for multimodal foundation models with language interfaces, and more specifically vision-language models (VLMs)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 288, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 460 + ], + "type": "text", + "content": "Prompt optimization of foundation models. Following the success of in-context learning [6], which appends user-generated natural language instruction and few-shot samples to text inputs, prompting [35] has emerged as the preferred fine-tuning paradigm for LLMs due to its superior performance and parameter-efficiency. However, recent prompt optimization methods, including continuous prefix-tuning [7, 31, 61, 62, 69] and discrete token-searching [11, 12, 59], still operate in a white-box manner, requiring access to either the tokenizer or output logits. Moreover, black-box prompting methods, such as heuristic-based editing [41, 49], are tailored towards language-only tasks and are thus not applicable in VLM settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 460, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 288, + 689 + ], + "type": "text", + "content": "LLMs for prompt optimization. APE [75] leverages an LLM to automatically write prompts using few-shot samples based on instruction induction [21] and paraphrasing [42, 54]. However, it is only designed to address language tasks, while we focus on multimodal tasks using black-box VLMs. LLMs have also proven to be an effective external knowledge base [40, 50, 57] for generating prompts in a zero-shot setting for multimodal models. For example, DCLIP [40] uses GPT3 to come up with rich visual descriptions to improve zero-shot classification with CLIP [51]. We extend this line of work to show that LLMs can iteratively optimize prompts for VLMs in a black-box fashion given few-shot samples. We further illustrate that prompt optimization with LLMs can be made more efficient by leveraging conversational feedback, such as providing ChatGPT with explicit language feedback on how well the most recent prompt performs. Our findings align with the perspective [9] of LLMs as meta-optimizers that can implicitly perform gradient search through in-context learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "Few-shot adaptation of VLMs. Prompting has also been successfully adopted in VLMs [14], as demonstrated by methods like CoOp [74] that fine-tune an ensemble of continuous prefix tokens using cross-entropy loss. [34] achieves state-of-the-art few-shot performance with a cross-modal (image and text) cross-entropy loss. However, these methods all require access to model parameters for gradient backpropagation. We also note that while some concurrent works, such as BlackVIP [44] and LFA [46], claim to operate in a \"black-box\" setting, they still require access to privileged information including output logits and embeddings. In this work, we introduce a truly black-box and gradient-free approach that yields competitive results to white-box approaches in extremely low-shot scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 251, + 541, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 541, + 264 + ], + "type": "text", + "content": "3. Prompting VLMs Using Chat-Based LLMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 270, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 294 + ], + "type": "text", + "content": "We now present our approach for prompting VLMs using chat-based LLMs as optimizers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": "Preliminaries. Motivated by recent proprietary VLMs [4, 45], we adopt a stricter yet practical black-box setting compared to prior works [44, 46], requiring minimal knowledge about the model's inner workings. This is crucial since releasing output logits or embeddings can potentially facilitate unauthorized knowledge extraction through distillation methods [19]. Our objective is to enhance the performance of a VLM equipped with a language interface capable of processing a textual prompt " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "p \\in T" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": ". We assume that the targeted task is accompanied by a training dataset denoted as " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "D_{train} \\subset D" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": " and its performance can be evaluated with respect to the prompt, represented as a function " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "F: D \\times T \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": ". For example, in a classification task, " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "D_{train} = \\{x, y\\}_n" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": " is an image and " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": " is its class label. The black-box VLM takes the image as input and returns a predicted label. We measure the performance of the textual prompt by calculating the average classification accuracy as " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "F(D_{train}, p)" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": ". Our goal in prompt engineering is to search for the optimal prompt " + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 304, + 294, + 547, + 521 + ], + "type": "text", + "content": " without accessing or modifying the black-box VLM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": "Background: human prompt engineering. Our method draws inspiration from the typical workflow of human prompt engineers. Prompt engineering is often an iterative process that involves: (a) creating an initial prompt " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "U = \\{p_1\\}" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": " based on the understanding of a task, (b) evaluating the performance of prompts in " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": ", (c) refining prompts based on the outcomes, (d) repeating the last two steps until convergence, and (e) returning the prompt " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "p^*" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": " with the highest " + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "inline_equation", + "content": "F(D_{train}, p^*)" + }, + { + "bbox": [ + 304, + 522, + 547, + 653 + ], + "type": "text", + "content": ". This hands-on approach helps optimize the model's performance, but it can be tedious and labor-intensive. Algorithm 1 formally illustrates this process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "Example: prompting for image classification with CLIP [51]. CLIP is one of the most popular VLM that takes a set of class-specific prompts when performing \"zero-shot\" image classification. [51] details the laborious prompting procedure over the course of a year. Interestingly, they find" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 250, + 713 + ], + "type": "text", + "content": "1Project site: llm-can-optimize-vlm.github.io" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12689" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 47, + 111, + 288, + 248 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 107 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 107 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 107 + ], + "type": "text", + "content": "Algorithm 1 We formalize human prompt engineering with the following algorithm, which motivates our LLM-based algorithm (2)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "lines": [ + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}} = \\{x,y\\}_{n}" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "F:D\\times T\\to" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " R: evaluation function \n1: Create initial prompts: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\gets \\{p_1\\}" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n2: Evaluate prompts on training set: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "S\\gets \\{F(D_{\\mathrm{train}},p_1)\\}" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n3: while not converged do \n4: Generate a new prompt " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "p^\\prime" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n5: Evaluate the new prompt: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "s^{\\prime} = F(D_{\\mathrm{train}},p^{\\prime})" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{U}\\leftarrow \\mathcal{U}\\cup \\{p^{\\prime}\\}" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "S\\gets S\\cup \\{s^{\\prime}\\}" + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "text", + "content": " \n8: end while \n9: return optimal prompt " + }, + { + "bbox": [ + 47, + 111, + 288, + 248 + ], + "type": "inline_equation", + "content": "p^*\\gets \\arg \\max_{p\\in \\mathcal{U}}F(D_{\\mathrm{train}},p)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": "that a default class-agnostic prompt (or so-called \"template\"), \"a photo of a {class}\" can provide a decent boost in accuracy for most datasets compared to using vanilla class labels. In this scenario, the evaluation function " + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": " is the classification accuracy on the test set, and the prompt " + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "inline_equation", + "content": "p = \\{ \\text{\"a photo of a {class}\" } | c \\in C \\}" + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 274, + 289, + 357 + ], + "type": "text", + "content": " is the set of class names for a given dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "text", + "content": "Prompting with chat-based LLMs (our approach). Given the strong in-context reasoning capabilities of LLMs, we envision them as a black-box optimizers that can improve prompts based on their performance outcomes, akin to how human prompt engineers iteratively refine prompts. Specifically, we maintain a pool of prompts " + }, + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "text", + "content": " and their corresponding performance outcomes " + }, + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 361, + 289, + 564 + ], + "type": "text", + "content": ". In each iteration, we provide the LLM with both positive and negative prompts, such as the highest and lowest-performing candidates. Such textual feedback through in-context prompts offers LLMs an implied \"gradient\" direction [9], making optimization more efficient than taking random local steps. We facilitate this feedback mechanism through conversations with state-of-the-art chat-based LLMs like ChatGPT [47] as illustrated in Figure 1. We note that such a multi-turn conversation is not the only way of conversing with ChatGPT, and ablate different in-context feedback mechanisms in section 8." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 582, + 269, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 269, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 269, + 596 + ], + "type": "text", + "content": "4. Illustrative Few-Shot Classification Task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "We illustrate our approach using a few-shot image classification task. Specifically, a prompt " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "p \\in T" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " consists of a set of class-specific prompts – that is, one textual description per class. The evaluation function " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " takes the prompt " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", along with an image dataset " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{train}" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": ", and returns the accuracy using the black-box VLM. To prevent overfitting and simplify our search space, we restrict our search to finding a single class-agnostic template, e.g., a photo of a {}, filling in the blank with label names provided with the dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "Algorithm 2 LLM-based prompt engineering on the illustrative classification task. Our algorithm requires a chat-based LLM and a (black-box) evaluation function, such as accuracy. We highlight mechanisms for “exploration” (restart and reset) in blue and “exploitation” (iter) in red. We mark the key component of “conversational feedback” of our approach in violet. The actual prompts are attached in section 8." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 306, + 159, + 547, + 425 + ], + "blocks": [ + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "lines": [ + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{train}} = \\{x,y\\}_{n}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "F:D\\times T\\to" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " R: evaluation function. \nRequire: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{restart}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " : number of initial sampled prompt sets, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{reset}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " : number of resets for a prompt set, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{iter}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " : number of hill-climbing iterations, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " size of one initial prompt set, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " : number of prompts send to ChatGPT. \n1: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "p^* \\gets \\emptyset" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n2: for 1::" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{restart}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " do \n3: Sample a new prompt set, " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{\\mathrm{init}} \\leftarrow \\{p_1,\\dots,p_m\\}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n4: for 1::" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{reset}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " do \n5: Reset to initial prompt set: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\leftarrow \\mathcal{U}_{\\mathrm{init}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n6: for 1::" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{iter}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " do \n7: Sort " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " by score outcomes " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\{F(D_{\\mathrm{train}},p)\\}_{p\\in U}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{\\mathrm{top}} \\leftarrow" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " top-k prompts in " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{\\mathrm{bot}} \\leftarrow" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " bottom-k prompts in " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n10: Get a new prompt " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{new}} \\leftarrow" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " LLM(" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{\\mathrm{top}},\\mathcal{U}_{\\mathrm{bot}}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": ") \n11: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\leftarrow \\mathcal{U} \\cup \\{p_{\\mathrm{new}}\\}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n12: end for \n13: " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "p^* \\leftarrow" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " arg max " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "p\\in \\mathcal{U}\\cup \\{p^*\\}" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "F(D_{\\mathrm{train}},p)" + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "text", + "content": " \n14: end for \n15: end for \n16: return prompt with highest score " + }, + { + "bbox": [ + 306, + 159, + 547, + 425 + ], + "type": "inline_equation", + "content": "p^*" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": "Outline of our approach (Alg. 2). To start, we sample entirely random initial prompts from a text corpus such as LAION-COCO [55] captions. Our approach follows the classical stochastic hill-climbing framework with random restart [54], which prevents ChatGPT from being trapped in local optima by balancing \"exploration\" and \"exploitation\". Our restart mechanism is implemented by sampling " + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{restart}}" + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": " initial prompt sets to encourage exploration. Because ChatGPT performs stochastic top-k sampling for text generation (as we adopt the default temperature of 1.0), we also implement a reset mechanism to foster additional exploration by retrying a given prompt set " + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{reset}}" + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": " times. For exploitation, we converse with ChatGPT for " + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{iter}}" + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": " iterations. We find that it is critical to balance exploration and exploitation for optimal performance, and thoroughly examine this trade-off in section 9. Lastly, we present ChatGPT both the top and bottom-performing prompts, denoted as " + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "inline_equation", + "content": "(\\mathcal{U}_{top}, \\mathcal{U}_{bot})" + }, + { + "bbox": [ + 304, + 449, + 547, + 676 + ], + "type": "text", + "content": ". We show that this simple adjustment can improve the efficiency of our approach in Figure 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Experimental setup. We apply our approach to the few-shot image classification benchmark introduced in CoOp [74], which is the most commonly studied setup for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12690" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "fine-tuning VLMs. This benchmark involves a collection of 11 datasets covering diverse image domains including ImageNet [10] and more niche datasets such as FGVC-Aircraft [39]. For each dataset, we adhere to the same three-fold k-shot train sets in [34], reporting the average accuracy across all folds. Importantly, our method only utilizes the train set to compute the score and does not require the few-shot validation set. We use CLIP following prior work [34, 74] to emulate a black-box VLM, and we employ ChatGPT (GPT3.5) as the chat-based LLM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": "Implementation details. To start, we sample entirely random 1M initial prompts from a text corpus (LAION-COCO [55] captions). For each caption, we extract all the noun phrases using spaCy part-of-speech tagging [20]. Subsequently, we replace one noun phrase in the caption with " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " (a placeholder where the class name will be inserted) to create a template. Given that each caption contains an average of 2 noun phrases, our initial prompt pool consists of approximately 2M templates. We run our algorithm with " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{start}} = 20" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " restarts, " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{resets}} = 50" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " resets, and " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{iter}} = 10" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " iterations. We opt to sample " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "m = 100" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " prompts per restart and present the top and bottom " + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "inline_equation", + "content": "k = 15" + }, + { + "bbox": [ + 46, + 195, + 289, + 471 + ], + "type": "text", + "content": " prompts to ChatGPT. We ablate different sets of hyperparameters and explain how we balance the tradeoff between exploration and exploitation in section 9. We adopt gpt-3.5-turbo-0301 model for ChatGPT using OpenAI's official API and keep the default sampling temperature of 1.0. We also ablate gpt-4 in Table 10 and find it achieves similar performance. The exact prompts used to converse with ChatGPT are documented in section 8. For a fair comparison, we use CLIP-RN50 for our experiments following prior work [34, 74]. We will open-source our code and release the initial prompt pool (LAIONCOCO-1M) to the public." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 289, + 715 + ], + "type": "text", + "content": "Oracle white-box baselines. Our black-box setup substantially differs from, and is more constrained than, the scenarios considered in previous white-box baselines. Specifically, we do not expose the pre-trained weights, model architectures, feature embeddings, or even output logits of VLMs. These constraints render many established gradient-based fine-tuning baselines inapplicable. Among the oracle white-box approaches we later compare to, CoOp [74] performs continuous prompting and requires backpropagation across all layers. WiSE-FT [67] ensembles fine-tuned weights with the original CLIP weights. Cross-Modal Adaptation [34] fine-tunes a linear classifier leveraging both image and text embeddings from CLIP. BlackVIP [44] and LFA [46] are two most recent baselines that apply CLIP logits or embeddings for gradient back-propagation. Finally, while DCLIP [40] queries GPT3 for rich visual descriptors for each class and does not require gradient-based fine-tuning, it performs prompt assembling using 4-6 class-specific prompts, which breaches our black-box assumption for accessing the output logits." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "content": "Black-box methods. We additionally benchmark our method against truly black-box solutions, including the vanilla class-agnostic templates “{class)” and “a photo of a {class}”. Also, we compare our approach to the best Hand-Engineered templates released by OpenAI, searched using test set performance to represent the theoretical upper bound of human performance, eg., “a centered satellite photo of {class}.” for EuroSAT [17]. Finally, we present two versions of conversational feedback of our approach: (a) using 30 positive (P only) or (b) using 15 positive and 15 negative prompts " + }, + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "inline_equation", + "content": "(\\mathbf{P} + \\mathbf{N})" + }, + { + "bbox": [ + 304, + 72, + 547, + 262 + ], + "type": "text", + "content": " in each iteration. For a fair comparison, both of our approaches start with the same initial sampled prompts, referred to as LAIONCOCO-1M. We also show the performance of the best initial sampled prompt searched using trainset performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "text", + "content": "SOTA one-shot performance against existing methods on 11 datasets. We report the test set performance of our method versus the aforementioned baselines in a challenging 1-shot classification scenario in Table 1. First, compared to the top-performing initial prompts selected from LAIONCOCO-1M based on train set performance, our prompt optimization using ChatGPT notably improves the initial prompts by an average of " + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "text", + "content": " (56% to 61%). Remarkably, our black-box approach surpasses the two white-box gradient-based fine-tuning techniques CoOp and WiSE-FT by at least " + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "text", + "content": ". Given that both CoOp and our method optimize a single class-agnostic template, we attribute this gap in performance to reduced overfitting. More specifically, we posit that our optimization space of natural language effectively acts as a regularizer in extremely low-shot tasks, standing as a more robust alternative to the continuous prompting approach of CoOp. Furthermore, our method benefits from textual feedback and shows improved performance by " + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 304, + 263, + 548, + 514 + ], + "type": "text", + "content": " when using both positive and negative prompts. In section 9, we show that our approach remains effective across different CLIP and ChatGPT variants." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 515, + 548, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 548, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 548, + 647 + ], + "type": "text", + "content": "Incorporating negative prompts leads to more efficient optimization. In Figure 2, we demonstrate that incorporating both positive and negative prompts fosters better optimization efficiency, achieving higher accuracy within a much fewer number of resets. Specifically, we hypothesize that LLMs can leverage the implicit \"gradient\" direction suggested in textual feedback to achieve faster convergence. For additional analysis, we ablate different ways of providing conversational feedback to ChatGPT in section 8 and conclude that iteratively updating both positive and negative prompts is the key for efficient optimization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 658, + 545, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 545, + 671 + ], + "type": "text", + "content": "5. More Benefits of Natural Language Prompts" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "In this section, we delve deeper into the advantages of utilizing natural language prompts compared to the continuous prompts [74]. We highlight that the prompts derived through" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12691" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 70, + 455, + 236 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 455, + 236 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 455, + 236 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 455, + 236 + ], + "type": "table", + "html": "
MethodDatasetAvg
CaltechImageNetAircraftFoodPetsCarsSUNUCFDTDEuroSATFlowers
Oracle white-box approaches
Cross-Modal [34]89.161.620.677.185.759.063.464.749.961.876.364.7
WiSE-FT [67]85.558.318.671.981.755.756.659.444.252.365.859.1
CoOp [74]87.557.29.674.385.955.660.361.944.450.668.159.6
LFA [46]81.652.417.063.175.341.458.456.738.460.774.956.4
BlackVIP [44]85.858.815.376.785.256.457.058.840.130.061.156.8
DCLIP [40]-59.6-76.483.8---41.734.7--
Manual prompting approaches
{}78.555.315.574.078.952.253.455.541.432.157.354.0
a photo of a {}84.557.915.974.083.253.958.056.938.828.660.255.6
Hand-Engineered [51]86.358.217.377.385.855.658.561.542.337.666.158.8
Our black-box approaches
LAIONCOCO-1M81.456.217.476.579.651.354.955.843.138.661.356.0
Ours (P only)89.059.417.977.885.755.760.458.743.646.766.660.1
Ours (P+N)89.159.618.178.388.156.261.060.244.849.067.261.1
", + "image_path": "e4ace9614662bebfcb80cc23f784cd43c78f6fcfe26c7e963608a3affdcb1f89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "text", + "content": "Table 1. Comparison of our method with other baselines on one-shot classification tasks. We report the average accuracy of each method across three folds, optimized using 1-shot training sets. We bold the best black-box result for each dataset, and underline the second best result. First, we note that our approach can effectively improve upon the initial prompts selected from LAIONCOCO-1M from " + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "text", + "content": ". Our approach is also competitive against the best Human-Engineered prompts released by OpenAI [51] searched using test set performance. Additionally, we show that using both positive and negative prompts improves the overall accuracy by " + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "text", + "content": ". For reference, we report oracle white-box approaches in gray. Remarkably, we also surpass white-box solutions such as WiSE-FT [67] and CoOp [74] by " + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 243, + 547, + 355 + ], + "type": "text", + "content": ". These methods require either gradient-based fine-tuning (CoOp/WiSE-FT/Cross-Modal) or prompt ensembling using output logits (DCLIP). While our approach is less effective than the SOTA white-box method (Cross-Modal Adaptation), we stress that our black-box setup is significantly more challenging, because we restrict the optimization space to natural language and do not access the pre-trained weights, model architectures, feature embeddings, and output logits of VLMs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 378, + 157, + 463 + ], + "blocks": [ + { + "bbox": [ + 50, + 378, + 157, + 463 + ], + "lines": [ + { + "bbox": [ + 50, + 378, + 157, + 463 + ], + "spans": [ + { + "bbox": [ + 50, + 378, + 157, + 463 + ], + "type": "image", + "image_path": "984ce6c2fc4524419f5608bf8d5dcb04a71044111e8c5f49adc33ff86558f74f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 472, + 288, + 560 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 560 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 560 + ], + "type": "text", + "content": "Figure 2. Conversational feedback incorporating both positive and negative prompts leads to improved efficiency. We fix the number of restarts to 20 and iterations to 10, and ablate different numbers of resets on all 11 datasets (left) and ImageNet (right). Notably, our approach using “" + }, + { + "bbox": [ + 46, + 472, + 288, + 560 + ], + "type": "inline_equation", + "content": "\\mathrm{P} + \\mathrm{N}" + }, + { + "bbox": [ + 46, + 472, + 288, + 560 + ], + "type": "text", + "content": "” (both top-15 and bottom-15 prompts) can optimize faster within a much fewer number of resets than using “P-Only” (top-30 prompts), resulting in the highest overall performance." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 378, + 276, + 462 + ], + "blocks": [ + { + "bbox": [ + 169, + 378, + 276, + 462 + ], + "lines": [ + { + "bbox": [ + 169, + 378, + 276, + 462 + ], + "spans": [ + { + "bbox": [ + 169, + 378, + 276, + 462 + ], + "type": "image", + "image_path": "d848393515ce207d267624f6e16fe8f7a8264c37a7dc772585090ac716b31b14.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 578, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 625 + ], + "type": "text", + "content": "our method are interpretable; for instance, they often contain descriptions of the targeted image domain. Our prompts can also transfer across CLIP architectures in a black-box manner, such as from RN50 to ViT/B-16." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Interpretable natural language prompts. While CoOp [74] concedes that continuous prompts can be difficult to interpret, our method – without explicitly instructing ChatGPT to generate interpretation – often yields interpretable results. Table 2 showcases the templates returned by our algorithm for each dataset, frequently including keywords that reflect the targeted image domain. For example, the template" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 375, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 507 + ], + "type": "text", + "content": "for Food101 [5] mentions \"diverse cuisine and ingredients\", and the template for UCF101 [60] (an action recognition dataset) mentions \"in motion\". Likewise, these templates identify general stylistic attributes of the datasets; they refer to \"bright and natural lighting\" for ImageNet [10] and note images that \"emphasize the subject\" for Caltech101 [27]. These prompts are particularly intriguing because we do not provide ChatGPT with any information about the downstream task, yet it manages to generate prompts containing domain-specific keywords that are similar to those engineered by human experts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": "Black-box prompt transfer. Our text prompts also maintain consistently high performance across different CLIP backbones. For comparison, since CoOp uses the same tokenizer for all CLIP architectures (including RN50, RN101, ViT/B-32, and ViT/B-16) and optimizes continuous prompts of the same shape (16 x 512), we assess the transferability of these learned continuous prompts from RN50 to other backbones using the official weights on 16-shot ImageNet. Table 3 showcases the results of this experiment, where we also include the baseline prompt \"a photo of a {}\" for reference. We observe a significant decline in accuracy when transferring CoOp's prompts (up to a " + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": " decrease despite utilizing more powerful backbones), implying that continuous prompts tend to overfit to the specific CLIP model. In contrast, our natural language prompts maintain their performance and outperform the baseline across all backbones." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12692" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 152, + 71, + 442, + 171 + ], + "blocks": [ + { + "bbox": [ + 152, + 71, + 442, + 171 + ], + "lines": [ + { + "bbox": [ + 152, + 71, + 442, + 171 + ], + "spans": [ + { + "bbox": [ + 152, + 71, + 442, + 171 + ], + "type": "table", + "html": "
DatasetExample of Top Templates
Caltech [27]An image of a {} with a blurred background that emphasizes the subject
DTD [8]The essential elements of {} are amplified with visual simplicity
EuroSAT [17]A top-down view of {} arranged in a pattern {}
Aircraft [39]A clear, high-quality image of a single {} with a white background
Food [5]A {} featuring diverse cuisine and ingredients
ImageNet [10]An image of a {} with bright and natural lighting
Flowers [43]A clear and vivid photograph of the {} in its natural setting
Pets [48]A {} with distinct and recognizable characteristics
Cars [26]A {} featuring a wide range of color options for easy selection
SUN [68]A high-resolution photo of a {} with clear background and natural lighting
UCF [60]A black and white photo of a {} in motion
", + "image_path": "ea6d9f0e7b1ec8586e72976cdc91803d2b56b30100c76065ea76f4cff64084fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 218, + 276, + 271 + ], + "blocks": [ + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 547, + 213 + ], + "type": "text", + "content": "Table 2. Example templates returned by our algorithm on each dataset. Although we do not provide ChatGPT with any information regarding the targeted dataset, we observe that the resulting templates are remarkably similar to human-engineered templates, with many domain-specific details such as \"motion\" and \"cuisine\", and stylistic elements such as \"bright and natural lighting\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 218, + 276, + 271 + ], + "lines": [ + { + "bbox": [ + 59, + 218, + 276, + 271 + ], + "spans": [ + { + "bbox": [ + 59, + 218, + 276, + 271 + ], + "type": "table", + "html": "
MethodRN50→RN101→ViT-B/32→ViT-B/16
a photo of a {}57.960.661.966.6
CoOp63.020.631.739.5
Ours59.960.762.267.0
", + "image_path": "9f57995e32de08a6207cd280e91777439d334a8db733fb1291b2950e6fef7a02.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "text", + "content": "Table 3. Black-box prompt transfer from ResNet-50 to other CLIP architectures. We evaluate both our natural language prompts and CoOp's continuous prompts on 16-shot ImageNet, which are trained using the RN50 CLIP backbone. As a reference point, we include the baseline prompt \"a photo of a " + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "inline_equation", + "content": "\\{\\}" + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "text", + "content": "\", and show that the prompts derived from our method using RN50 consistently surpass it after transferring to different backbones. In contrast, while CoOp achieves better 16-shot ImageNet performance using RN50, its performance plummets during the transfer, e.g., from " + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "text", + "content": " to a mere " + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 281, + 289, + 390 + ], + "type": "text", + "content": " for RN101." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 399, + 263, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 399, + 263, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 399, + 263, + 414 + ], + "type": "text", + "content": "6. Application: Text-to-Image Generation" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 429, + 292, + 578 + ], + "blocks": [ + { + "bbox": [ + 50, + 429, + 292, + 578 + ], + "lines": [ + { + "bbox": [ + 50, + 429, + 292, + 578 + ], + "spans": [ + { + "bbox": [ + 50, + 429, + 292, + 578 + ], + "type": "image", + "image_path": "1a25d3c154e797990230cc64699ff124918931546f0a783046bca6c20a42ffac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 587, + 289, + 718 + ], + "lines": [ + { + "bbox": [ + 46, + 587, + 289, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 289, + 718 + ], + "type": "text", + "content": "Figure 3. Improving text-to-image (T2I) generation using chat-based multimodal LLMs. Our framework can use the multimodal GPT4-V [45] to optimize prompts for the state-of-the-art black-box generative VLM, DALL-E 3 [4]. For complex user queries that DALL-E 3 may initially fail to generate, we send the generated image (in violet) along with the current prompt to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue). Our simple framework is surprisingly effective at correcting DALL-E 3 mistakes on some challenging Winoground [63] text queries that involve action, logical, and spatial reasoning. We conduct a human evaluation on the quality of generated images in Table 6 and include the actual prompts in section 8." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 255 + ], + "type": "text", + "content": "In this section, we present a direct application of our prompt optimization framework to generative tasks using a truly black-box text-to-image (T2I) VLM, DALL-E 3 [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 257, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 461 + ], + "type": "text", + "content": "Optimizing T2I using a multimodal LLM. DALL-E 3 can generate high-fidelity images following diverse user queries, but crafting effective prompts is tricky even for designers experienced with AI content generation tools [36]. Therefore, we are motivated to implement our LLM-based optimization framework to assist with creative visual design. Our framework is shown in Figure 3 for the illustrative task of text-to-image generation. In this task, the user specifies a query (topic) in text, such as \"an animal watches a person\", and the goal is to write a prompt that can generate an image reflecting this topic. We adopt a multimodal LLM GPT4-V [45] (gpt-4-1106-preview) to provide feedback on the generated image and optimize the prompt. We find that this framework is surprisingly effective due to GPT4-V's strong visual reasoning capabilities, which can often spot subtle errors in generated images and offer more accurate prompts." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 473, + 552, + 622 + ], + "blocks": [ + { + "bbox": [ + 307, + 473, + 552, + 622 + ], + "lines": [ + { + "bbox": [ + 307, + 473, + 552, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 473, + 552, + 622 + ], + "type": "image", + "image_path": "2afe40db62a631c711c0caf5cb07baf88a646951f2afcb4573f046aaed48770d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 631, + 547, + 696 + ], + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 696 + ], + "type": "text", + "content": "Figure 4. Prompt inversion using chat-based multimodal LLMs. We apply our framework to reverse engineer the text prompt to generate the same user-queried image. We send the generated image (in violet) along with the original image to GPT4-V to ask for feedback on improvements (in red) and then generate a new prompt (in blue)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "Task setup. For T2I generation, we experiment with a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12693" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 272, + 248 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 272, + 248 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 272, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 272, + 248 + ], + "type": "table", + "html": "
User QueryInit. ImageLLM FeedbackFinal Image
Text-to-image generation
There is less milk than orange juice.Incorrect, the milk bottle appears full, more than orange juice...
A shorter person is covering the eyes of a taller person.Incorrect, the taller person is covering the shorter person's eyes. Instead, ...
Prompt inversion
The scarf should feature red and white stripes, and the fur is fluffy...
The coat should be buttoned and the lighting exhibits a stronger contrast...
", + "image_path": "3bd80818b456d5d614873a091fd61a92fa9e36a3893eaabc4d597f51f5c01e13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 337, + 293, + 462 + ], + "blocks": [ + { + "bbox": [ + 46, + 257, + 290, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 290, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 290, + 335 + ], + "type": "text", + "content": "Table 4. Examples of T2I optimization. We show that our framework (Figure 3) can automatically improve the faithfulness of images generated by DALL-E 3, with respect to user-specified textual topics (for T2I generation) or reference images (for prompt inversion). This is achieved through three rounds of prompt optimization, using feedback from the multimodal LLM (GPT4-V). Table 11 and Table 12 shows more examples with actual prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 337, + 293, + 462 + ], + "lines": [ + { + "bbox": [ + 50, + 337, + 293, + 462 + ], + "spans": [ + { + "bbox": [ + 50, + 337, + 293, + 462 + ], + "type": "table", + "html": "
User QueryInverted ImageExample 1Example 2Example 3Example 4Example 5
Give the dog a cat friend.Make the dog be in the middle of a jump.Make the dog do a handstand.Make the dog lie down on its side.Make the dog swim in water.
Make the owl fight a hawk.Make the owl flap its wings.Make the owl fully green.Make the owl stand in front of the moon.Make the owl walk in the city.
", + "image_path": "db73f4ca7cac10f0b776e0138004c46b18324d6a9bc3cfbc080f1b9841820588.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 715 + ], + "type": "text", + "content": "subset of 100 text queries from Winoground [63] that involve complex attribute and relation reasoning, which DALL-E might initially fail to generate. Our framework refines the prompts to capture the user-specified topics using a few (three) iterations. We also attempt a reverse task of prompt inversion: given a user-specified reference (query) image, our framework reverse-engineers the prompt to have DALL-E generate the same object or scene in the query image (see Figure 4). This enables users to easily make customizations [53] (see Table 5), such as having the character in a reference image perform various actions or change scenes. For this task, we sample 100 random queries from DiffusionDB [65]. We provide qualitative results in Table 4, Table 11, and Table 12. We hire two volunteers to assess the faithfulness of the images generated by our method, and to compare these with the images manually prompted by two" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "designers (each with one year of experience in AI content generation), as shown in Table 6." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 96, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 547, + 179 + ], + "type": "text", + "content": "Remarks on limitations. While we show promising results, we note some failure cases in Table 13 and Table 14 due to the inherent limitations of foundation models. For example, GPT4-V might fail to describe abstract and artistic details, and DALL-E 3 often fails to generate the correct number of objects. We believe that our framework can benefit from more capable foundation models in the future." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 334, + 189, + 518, + 258 + ], + "blocks": [ + { + "bbox": [ + 46, + 471, + 288, + 504 + ], + "lines": [ + { + "bbox": [ + 46, + 471, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 288, + 504 + ], + "type": "text", + "content": "Table 5. Customization via prompt inversion. Users can simply append extra descriptions to the inverted prompts to customize their main characters in queried images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 189, + 518, + 258 + ], + "lines": [ + { + "bbox": [ + 334, + 189, + 518, + 258 + ], + "spans": [ + { + "bbox": [ + 334, + 189, + 518, + 258 + ], + "type": "table", + "html": "
TaskMethodInit. (std)Final (std)Δ
Text-to-ImageHuman2.28 (.45)2.86 (.61)0.58
Ours2.62 (.36)3.56 (.54)0.94
Prompt InversionHuman1.58 (.48)2.76 (.53)1.18
Ours1.94 (.39)3.68 (.47)1.74
", + "image_path": "66305845930b1860ae78e1a5685d5837a93908c652e1dc2a53d7d62c597bf344.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 267, + 547, + 366 + ], + "lines": [ + { + "bbox": [ + 305, + 267, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 547, + 366 + ], + "type": "text", + "content": "Table 6. Our method enhances faithfulness in T2I generation. We hire two human annotators to assess the faithfulness of images generated from user queries, e.g., textual topics for Text-to-Image, or reference images for Prompt Inversion. The scores are measured on a 1-to-5 Likert scale, with 1 signifying contradiction and 5 indicating perfect alignment with the user's goal. Our approach benefits from three iterations of prompt optimization and consistently outperforms human-engineered prompts by designers who have one year of experience in AI content generation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 376, + 460, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 460, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 460, + 388 + ], + "type": "text", + "content": "7. Discussion and Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 574 + ], + "type": "text", + "content": "Summary. We present the first attempt to leverage LLMs as prompt engineers for VLMs. On one-shot image classification, our method surpasses existing human-engineered prompts and even rivals white-box approaches. Central to the success of our method is the use of conversational feedback, enabling chat-based LLMs to efficiently steer VLMs in the right direction. This process leads to a set of interpretable prompts bearing considerable resemblance to those crafted by humans. Importantly, our natural language prompting setup is a lot more constrained than the assumed scenarios of previous white-box or even some black-box settings [44], because we do not expose the model weights and outputs of VLMs. We finally apply our framework to illustrative generative tasks using a truly black-box text-to-image VLM (DALL-E 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 575, + 547, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 547, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 547, + 695 + ], + "type": "text", + "content": "Limitations and future work. While we try to minimize the overall cost and the total number of API calls, the energy consumption associated with LLMs remains a substantial concern. It is vital to note that we do not intend to compete directly with white-box baselines that can improve visual and text representations with more data. Further details on the higher-shot performance of our method can be found in section 9. Lastly, we are limited to costly human evaluation for T2I generation in this study. Future work may adopt automatic evaluation [18, 33] for large-scale experiments." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12694" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 213 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pages 2425-2433, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 292 + ], + "type": "text", + "content": "[3] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, Jenia Jitsev, Simon Kornblith, Pang Wei Koh, Gabriel Ilharco, Mitchell Wortsman, and Ludwig Schmidt. Openflamingo: An open-source framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 293, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 288, + 358 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Joyce Zhuang, Juntang and Lee, Yufei Guo, Wesam Manassra, Prafulla Dhariwal, Casey Chu, Yunxin Jiang, and Aditya Ramesh. Improving image generation with better captions. Note on Dalle-3, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 361, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 288, + 404 + ], + "type": "text", + "content": "[5] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 407, + 288, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 461 + ], + "type": "text", + "content": "[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 463, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 463, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 463, + 288, + 506 + ], + "type": "text", + "content": "[7] Yekun Chai, Shuohuan Wang, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Clip-tuning: Towards derivative-free prompt learning with a mixture of rewards. arXiv preprint arXiv:2210.12050, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 288, + 552 + ], + "type": "text", + "content": "[8] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2014. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 597 + ], + "type": "text", + "content": "[9] Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Zhifang Sui, and Furu Wei. Why can gpt learn in-context? language models secretly perform gradient descent as meta optimizers. ACL, 2023. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 644 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 1, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 288, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 699 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 699 + ], + "type": "text", + "content": "[11] Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh, Yihan Wang, Han Guo, Tianmin Shu, Meng Song, Eric P. Xing, and Zhiting Hu. Rlprompt: Optimizing discrete text prompts with reinforcement learning. arXiv preprint arXiv:2205.12548, 2022.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 714 + ], + "type": "text", + "content": "[12] Shizhe Diao, Xuechun Li, Yong Lin, Zhichao Huang, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 95 + ], + "type": "text", + "content": "Tong Zhang. Black-box prompt learning for pre-trained language models. arXiv preprint arXiv:2201.08531, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 148 + ], + "type": "text", + "content": "[13] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 547, + 194 + ], + "type": "text", + "content": "[14] Zhe Gan, Linjie Li, Chunyuan Li, Lijuan Wang, Zicheng Liu, and Jianfeng Gao. Vision-language pre-training: Basics, recent advances, and future trends. Foundations and Trends® in Computer Graphics and Vision, 14(3-4):163-352, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 547, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 547, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 547, + 249 + ], + "type": "text", + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 250, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 547, + 282 + ], + "type": "text", + "content": "[16] Tanmay Gupta and Aniruddha Kembhavi. Visual programming: Compositional visual reasoning without training. arXiv preprint arXiv:2211.11559, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 547, + 326 + ], + "type": "text", + "content": "[17] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification, 2017. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 547, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 547, + 369 + ], + "type": "text", + "content": "[18] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 371, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 547, + 403 + ], + "type": "text", + "content": "[19] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 447 + ], + "type": "text", + "content": "[20] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. To appear, 7(1): 411-420, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 449, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 547, + 491 + ], + "type": "text", + "content": "[21] Or Honovich, Uri Shaham, Samuel R Bowman, and Omer Levy. Instruction induction: From few examples to natural language task descriptions. arXiv preprint arXiv:2205.10782, 2022.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 547, + 547 + ], + "type": "text", + "content": "[22] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. In International Conference on Machine Learning, pages 2790-2799. PMLR, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 548, + 547, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 547, + 591 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 547, + 645 + ], + "type": "text", + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 690 + ], + "type": "text", + "content": "[25] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199-22213, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 714 + ], + "type": "text", + "content": "[26] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12695" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "4th International IEEE Workshop on 3D Representation and Recognition (3dRR-13), Sydney, Australia, 2013. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 284, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 284, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 284, + 108 + ], + "type": "text", + "content": "[27] Li, Andreeto, Ranzato, and Perona. Caltech 101, 2022. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 110, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 110, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 49, + 110, + 288, + 152 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 156, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 156, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 156, + 288, + 198 + ], + "type": "text", + "content": "[29] Shuang Li, Yilun Du, Joshua B Tenenbaum, Antonio Torralba, and Igor Mordatch. Composing ensembles of pre-trained models via iterative consensus. arXiv preprint arXiv:2210.11522, 2022.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 201, + 288, + 243 + ], + "type": "text", + "content": "[30] Tian Li, Anit Kumar Sahu, Ameet Talwalkar, and Virginia Smith. Federated learning: Challenges, methods, and future directions. IEEE signal processing magazine, 37(3):50-60, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 247, + 288, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 247, + 288, + 279 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 288, + 279 + ], + "type": "text", + "content": "[31] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 282, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 282, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 49, + 282, + 287, + 347 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 288, + 392 + ], + "type": "text", + "content": "[33] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 396, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 396, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 49, + 396, + 288, + 438 + ], + "type": "text", + "content": "[34] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramana. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. arXiv preprint arXiv:2301.06267, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 441, + 288, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 441, + 288, + 494 + ], + "spans": [ + { + "bbox": [ + 49, + 441, + 288, + 494 + ], + "type": "text", + "content": "[35] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1-35, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 498, + 288, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 498, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 49, + 498, + 288, + 540 + ], + "type": "text", + "content": "[36] Vivian Liu. Beyond text-to-image: Multimodal prompts to explore generative ai. In Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1-6, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 544, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 288, + 597 + ], + "type": "text", + "content": "[37] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. arXiv preprint arXiv:2304.09842, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 601, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 601, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 601, + 288, + 643 + ], + "type": "text", + "content": "[38] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 646, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 288, + 677 + ], + "type": "text", + "content": "[39] S. Maji, J. Kannala, E. Rahtu, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. Technical report, 2013. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 681, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 288, + 712 + ], + "type": "text", + "content": "[40] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[41] Swaroop Mishra, Daniel Khashabi, Chitta Baral, Yejin Choi, and Hannaneh Hajishirzi. Reframing instructional prompts to gptk's language. arXiv preprint arXiv:2109.07830, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[42] Melanie Mitchell, John Holland, and Stephanie Forrest. When will a genetic algorithm outperform hill climbing. Advances in neural information processing systems, 6, 1993. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[43] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 545, + 232 + ], + "type": "text", + "content": "[44] Changdae Oh, Hyeji Hwang, Hee-young Lee, YongTaek Lim, Geunyoung Jung, Jiyoung Jung, Hosik Choi, and Kyungwoo Song. Blackvip: Black-box visual prompting for robust transfer learning. In CVPR, 2023. 3, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 235, + 503, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 235, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 308, + 235, + 503, + 245 + ], + "type": "text", + "content": "[45] OpenAI. Gpt-4 technical report. 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 247, + 545, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 247, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 545, + 279 + ], + "type": "text", + "content": "[46] Yassine Ouali, Adrian Bulat, Brais Matinez, and Georgios Tzimiropoulos. Black box few-shot adaptation for vision-language models. In ICCV, 2023. 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 282, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 346 + ], + "type": "text", + "content": "[47] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35:27730-27744, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[48] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 384, + 545, + 426 + ], + "type": "text", + "content": "[49] Archiki Prasad, Peter Hase, Xiang Zhou, and Mohit Bansal. Grips: Gradient-free, edit-based instruction search for prompting large language models. arXiv preprint arXiv:2203.07281, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 430, + 545, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 430, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 308, + 430, + 545, + 462 + ], + "type": "text", + "content": "[50] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. ICCV, 2023. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 545, + 529 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 6, 13" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 564 + ], + "type": "text", + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 567, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 632 + ], + "type": "text", + "content": "[53] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500–22510, 2023. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 656 + ], + "type": "text", + "content": "[54] Stuart J Russell. Artificial intelligence a modern approach. Pearson Education, Inc., 2010. 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 4, 5" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12696" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 137 + ], + "type": "text", + "content": "[56] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[57] Sheng Shen, Chunyuan Li, Xiaowei Hu, Yujia Xie, Jianwei Yang, Pengchuan Zhang, Zhe Gan, Lijuan Wang, Lu Yuan, Ce Liu, et al. K-lite: Learning transferable visual models with external knowledge. Advances in Neural Information Processing Systems, 35:15558-15573, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 239 + ], + "type": "text", + "content": "[58] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. arXiv preprint arXiv:2303.17580, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 242, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 288, + 285 + ], + "type": "text", + "content": "[59] Taylor Shin, Yasaman Razeghi, Robert L Logan IV, Eric Wallace, and Sameer Singh. Autoprompt: Eliciting knowledge from language models with automatically generated prompts. arXiv preprint arXiv:2010.15980, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 288, + 319 + ], + "type": "text", + "content": "[60] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "text", + "content": "[61] Tianxiang Sun, Zhengfu He, Hong Qian, Yunhua Zhou, Xuan-Jing Huang, and Xipeng Qiu. Bbtv2: Towards a gradient-free future with large language models. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 3916-3930, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 421 + ], + "type": "text", + "content": "[62] Tianxiang Sun, Yunfan Shao, Hong Qian, Xuanjing Huang, and Xipeng Qiu. Black-box tuning for language-model-as-a-service. In International Conference on Machine Learning, pages 20841-20855. PMLR, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 488 + ], + "type": "text", + "content": "[63] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 288, + 533 + ], + "type": "text", + "content": "[64] Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, and Lijuan Wang. Git: A generative image-to-text transformer for vision and language. arXiv preprint arXiv:2205.14100, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 588 + ], + "type": "text", + "content": "[65] Zijie J Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models. arXiv preprint arXiv:2210.14896, 2022. 8, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 645 + ], + "type": "text", + "content": "[66] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35: 24824-24837, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 713 + ], + "type": "text", + "content": "[67] Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo-Lopes, Hannaneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, and Ludwig Schmidt. Robust fine-tuning of zero-shot models. arXiv preprint arXiv:2109.01903, 2021. https://arxiv.org/abs/2109.01903.2,5,6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 441 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[68] Jianxiong Xiao, Krista A. Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. Int. J. Comput. Vision, 119(1): 3-22, 2016. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[69] Hanwei Xu, Yujun Chen, Yulun Du, Nan Shao, Yanggang Wang, Haiyu Li, and Zhilin Yang. Gps: Genetic prompt search for efficient few-shot learning. arXiv preprint arXiv:2210.17041, 2022.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[70] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 261 + ], + "type": "text", + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 547, + 318 + ], + "type": "text", + "content": "[72] Andy Zeng, Adrian Wong, Stefan Welker, Krzysztof Choromanski, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 547, + 361 + ], + "type": "text", + "content": "[73] Mingkai Zheng, Xiu Su, Shan You, Fei Wang, Chen Qian, Chang Xu, and Samuel Albanie. Can gpt-4 perform neural architecture search? arXiv preprint arXiv:2304.10970, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 547, + 396 + ], + "type": "text", + "content": "[74] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 398, + 547, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 547, + 441 + ], + "type": "text", + "content": "[75] Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. Large language models are human-level prompt engineers. *ICLR*, 2023. 2, 3, 13, 14" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12697" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_content_list.json b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..48a02b0bff149a216c8bafad61ac64f8d5b5d341 --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "Language-Driven Anchors for Zero-Shot Adversarial Robustness", + "text_level": 1, + "bbox": [ + 156, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Li $^{1}$ Wei Zhang $^{1}$ Yining Liu $^{2}$ Zhanhao Hu $^{1}$ Bo Zhang $^{1}$ Xiaolin Hu $^{1,3*}$ \\\n $^{1}$ Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China\\\n $^{2}$ Harbin Institute of Technology, Weihai, China\\\n $^{3}$ Chinese Institute for Brain Research (CIBR), Beijing, China\\\n{lixiao20, zhang-w19, huzhanha17}@mail.tsinghua.edu.cn\\\n22s030184@stu.hit.edu.cn\\\n{dcszb, xlhu}@mail.tsinghua.edu.cn", + "bbox": [ + 109, + 178, + 859, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 366, + 313, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) are known to be susceptible to adversarial attacks. Previous researches mainly focus on improving adversarial robustness in the fully supervised setting, leaving the challenging domain of zero-shot adversarial robustness an open question. In this work, we investigate this domain by leveraging the recent advances in large vision-language models, such as CLIP, to introduce zero-shot adversarial robustness to DNNs. We propose LAAT, a Language-driven, Anchor-based Adversarial Training strategy. LAAT utilizes the features of a text encoder for each category as fixed anchors (normalized feature embeddings) for each category, which are then employed for adversarial training. By leveraging the semantic consistency of the text encoders, LAAT aims to enhance the adversarial robustness of the image model on novel categories. However, naively using text encoders leads to poor results. Through analysis, we identified the issue to be the high cosine similarity between text encoders. We then design an expansion algorithm and an alignment cross-entropy loss to alleviate the problem. Our experimental results demonstrated that LAAT significantly improves zero-shot adversarial robustness over state-of-the-art methods. LAAT has the potential to enhance adversarial robustness by large-scale multimodal models, especially when labeled data is unavailable during training. Code is available at https://github.com/LixiaoTHU/LAAT.", + "bbox": [ + 76, + 398, + 473, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 820, + 209, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attacks [42], by adding deliberately designed perturbations to inputs, have posed a significant threat to", + "bbox": [ + 76, + 845, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6c7a4984faab35b63942b9e8134d0ddc50121a2c225caf308d2aa28606f0ead.jpg", + "image_caption": [ + "Figure 1. The illustration of zero-shot ability with LAAT. Different colors of the marks indicate different categories. When a model is adversarially trained with the text anchors of table and lion (seen categories), it can recognize adversarial examples of the two categories (grey and green). Then due to the text anchors of chair and tiger (novel categories) being close to those of table and lion, respectively, the model can also recognize the two novel categories." + ], + "image_footnote": [], + "bbox": [ + 545, + 364, + 875, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the application of Deep Neural Networks (DNNs) [16, 57]. In response to the threats, various techniques have been proposed to enhance adversarial robustness of DNNs. Among these, Adversarial Training (AT) and its variants [32, 36, 37, 52] have shown to be highly effective [16, 42, 57]. However, applying AT from scratch to novel datasets or categories can be time-consuming. In real-world scenarios, it is common for models to encounter examples from novel categories, known as zero-shot learning (ZSL) setting [27]. If adversarially trained models possess zero-shot capability, it can significantly enhance the scalability and practical utility of adversarially robust models.", + "bbox": [ + 496, + 641, + 893, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial robustness in the ZSL setting is challenging. Earlier attempts [51, 55] try to combine attribute-based ZSL methods with AT. However, these methods yielded modest improvements and lacked scalability due to the practical difficulties in obtaining attributes. Recently, large-scale", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 223, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24686", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vision-language models such as CLIP [39] have made significant strides and are considered foundational models for various downstream applications. These models, trained on extensive text-image data, demonstrate strong zero-shot capability by using textual descriptions to recognize new visual categories, eliminating the need for explicit reliance on image attributes. In this work, different from previous attribute-based ZSL methods, we explore the utilization of foundation models to introduce zero-shot adversarial robustness to DNNs.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a Language-driven, Anchor-based Adversarial Training strategy, denoted as LAAT, to enhance the zero-shot adversarial robustness. Specifically, we use the popular text encoder from vision-language models, specifically CLIP [39], to obtain $l_{2}$ normalized feature embedding (named anchor in this paper) for the text (label name) corresponding to image categories. The text encoder exhibits a property called semantic consistency, where semantically similar categories are mapped to neighboring anchors in the feature space. Then we utilize the text anchors to supervise the image classification model with AT. After this process, the image model not only obtains adversarial robustness on seen categories but also aligns the image features with the text anchors. During zero-shot inference, by leveraging the semantic consistency between seen and novel categories, the image model can recognize novel categories while maintaining adversarial robustness. Fig. 1 illustrates the idea. We note that one recent work, TeCoA [34], also utilizes the ability of large-scale vision-language models to enhance the zero-shot adversarial robustness. However, TeCoA was only shown effective for extremely tiny adversarial noise (i.e., $\\epsilon = 1 / 255$ ).", + "bbox": [ + 75, + 244, + 472, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As CLIP models are not designed for adversarial robustness, directly applying the obtained anchors from CLIP text encoders to anchor-based AT [36, 37] (refer to Sec. 2.1) results in poor performance. We attribute it to the high Cosine Similarity (CoS) problem, i.e., the average cosine similarity between CLIP anchors are too high to be directly used in AT. See Sec. 3.3 for details. This problem has never been discovered in previous work utilizing CLIP models [22, 28, 34, 45]. To address it, we propose several techniques. We first introduce an expansion algorithm that maps the original CLIP anchors to new anchors with increased distances while maintaining semantic consistency. We then incorporate an Alignment Cross-Entropy (A-CE) loss as the optimization objective. In addition, we use a smoothness loss to enhance the robustness on unseen categories.", + "bbox": [ + 75, + 580, + 472, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The experimental results showed that the models trained with LAAT achieved remarkable zero-shot performance against strong adversarial perturbations, even surpassing the previous state-of-the-art (SOTA) adversarily robust few-shot methods. We also trained several models with LAAT on a large dataset ImageNet-1K [40]. These mod", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els showed substantial adversarial robustness across various downstream datasets in the zero-shot setting, significantly outperforming the recent TeCoA method [34]. The main contributions can be summarized as follows:", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the high CoS problem in CLIP anchors for adversarial robustness and propose effective techniques to alleviate it.", + "- The models trained by LAAT show substantial adversarial robustness across downstream datasets, suggesting that AT in the ZSL setting could be a promising way to improve the practical usefulness of AT." + ], + "bbox": [ + 500, + 151, + 893, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 272, + 640, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anchor-based AT", + "text_level": 1, + "bbox": [ + 500, + 297, + 669, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "AT has become one of the most effective strategies to improve adversarial robustness [3]. Several studies try to improve AT from various aspects [26, 29, 30, 36, 37, 52]. One series of works focuses on refining cross-entropy loss, which is the most commonly used loss in classification, to obtain better robustness with AT [36, 37]. Pang et al. [36] boosts AT by introducing the Max-Mahalanobis center (MMC) loss, a Euclidean loss where there is a fixed feature vector $\\mu_y$ for each category $y$ , and the optimization object is $||\\mathbf{z} - \\mu_y||_2^2$ , where $\\mathbf{z}$ is the output feature. Pang et al. [37] uses Hypersphere Embedding to boost AT. It normalizes both weights $\\mathbf{W}$ and features $\\mathbf{z}$ of the output layer, and proposes two losses in the form of $\\cos \\theta$ and $\\theta$ , where $\\theta$ is the angle between $\\frac{\\mathbf{W}}{||\\mathbf{W}||_2}$ and $\\frac{\\mathbf{z}}{||\\mathbf{z}||_2}$ . We call these AT methods anchor-based methods as they all used $l_{2}$ normalized feature embeddings (anchors) for training. Besides, several metric learning works [33, 56] also used a similar paradigm to anchor-based AT for enhancing robustness.", + "bbox": [ + 496, + 321, + 893, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Language-driven Recognition.", + "text_level": 1, + "bbox": [ + 500, + 604, + 767, + 621 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Earlier ZSL models exploit auxiliary information, commonly in the form of image attributes, to transfer knowledge between seen and novel categories [47, 50]. Recently, language-driven recognition has caught attention in the ZSL field, especially with the recent advance of large-scale pretrained vision-language models [24, 39]. CLIP [39] demonstrated that classic recognition tasks can strongly benefit from millions of raw texts pertaining to images. CLIP uses an independent image encoder and text encoder to perform contrastive learning on 400M image-text pairs. After pretraining, natural language is used to reference visual concepts by computing the CoS between text features and the visual feature, enabling the zero-shot transfer of the model to downstream classification datasets. The basic paradigm of CLIP can be extended to several tasks such as semantic segmentation and object detection [22, 28, 45]. However, recent works show that CLIP models lack adversarial robustness and can be easily attacked [4, 34, 53].", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24687", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b685b0efffcee3d0afa86d067db947d125bf0445ae02e8ba3b1933f71f306736.jpg", + "image_caption": [ + "Figure 2. The pipeline of LAAT. Only the image encoder is trainable in the figure. $\\odot$ indicates computing the CoS. Red arrows indicate the adversarial example generation process and brown arrows indicate the inference." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 885, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Adversarily Robust Zero/few-shot Learning", + "text_level": 1, + "bbox": [ + 76, + 404, + 464, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An early preprint work [55] combines AT with attribute-based ZSL methods but only gets a mild improvement over non-AT models. In addition, the method is hard to generalize without the attribute annotations. Yucel et al. [51] evaluates the adversarial robustness of classic attribute-based zero-shot models and concludes that adversarial robustness for ZSL field is suffering an immature state. Mao et al. [34] proposes several techniques such as visual prompting and fine-tuning to adapt large-scale models, e.g., the image models of CLIP, for zero-shot adversarial robustness, while their method called TeCoA was only shown to be effective under extremely tiny adversarial noise. Different from TeCoA, LAAT only uses the anchors from the CLIP text encoders.", + "bbox": [ + 75, + 428, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A closely related area is adversarially robust few-shot learning, which has caught attention recently, too. We also compare LAAT with recent few-shot methods directly to comprehensively evaluate it. AQ [19], R-MAML [46], and ITS-MAML [15] study meta-learning jointly with AT. GR [13] tries to learn a generalizable robust representation by the combination of several AT techniques.", + "bbox": [ + 75, + 640, + 470, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LAAT", + "text_level": 1, + "bbox": [ + 76, + 761, + 155, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Pipeline", + "text_level": 1, + "bbox": [ + 76, + 785, + 238, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LAAT first obtains fixed text anchors from the text encoder, then uses them to adversarially supervise the training of image classification model. After AT, the image model will obtain adversarial robustness on seen categories and also align the image features with the text anchors to obtain zero-shot ability (see Fig. 1). The overall pipeline of LAAT is", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "illustrated in Fig. 2. Given the original text anchors, LAAT uses an expansion algorithm to obtain fixed Ground-Truth (GT) anchors suitable for AT, then uses an image encoder to extract both benign and adversarial features. The CoS between adversarial features and GT anchors are maximized by minimizing the A-CE loss $(L_{1})$ . Besides, a smoothness loss $(L_{2})$ is applied to the adversarial and benign features. We then describe each module of LAAT.", + "bbox": [ + 496, + 405, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text Anchor Extraction", + "text_level": 1, + "bbox": [ + 498, + 536, + 718, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the names of $N$ categories in the training set, they are first filled into $N$ sentences with a fixed prompt text, such as \"A photo of {}\". After that, these sentences are encoded into $N$ anchors $\\{\\mathbf{a}_i\\}_{i=1}^N$ by a fixed pre-trained text encoder, where $\\mathbf{a}_i \\in \\mathbb{R}^n$ and $||\\mathbf{a}_i||_2 = 1$ . Many architectures are feasible if they have semantic consistency as mentioned before and in this work, we use the large-scale pre-trained CLIP text encoder. However, CLIP anchors cannot be used directly due to the high CoS problem, as analyzed below.", + "bbox": [ + 496, + 559, + 893, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. High Cosine Similarity Problem", + "text_level": 1, + "bbox": [ + 498, + 705, + 782, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first make an intriguing observation that directly using CLIP anchors for anchor-based AT poses significant challenges in achieving convergence. This phenomenon is universal among different CLIP text encoders and different models. To illustrate this, we adversarially trained a ResNet18 [23] on Cifar100 [25] by maximizing the $\\cos \\theta$ objective [37], i.e., the CoS between the model output $\\mathbf{z}$ and the GT anchor $\\mathbf{a}_y$ [37], which is computed as $\\frac{\\mathbf{z}^T\\mathbf{a}_y}{||\\mathbf{z}||_2}$ (note that $||\\mathbf{a}_y||_2 = 1$ ). The experimental details can be found in Appendix A. Fig. 3 shows the learning curves with CLIP anchors (the orange curves). For comparison, here we also", + "bbox": [ + 496, + 729, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24688", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5acee1955a3b9094233ca2dd05ae15db8dd9bad084366207c6d3cbab3f5cf0e9.jpg", + "image_caption": [ + "(a) Clean accuracy." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 264, + 199 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/306e86cb3b75467e3f676cd1e9059bcb21776e635b445500961d6f90110cda9f.jpg", + "image_caption": [ + "(b) Robust accuracy.", + "Figure 3. Learning curves of AT supervised by $\\cos \\theta$ with fixed anchors generated from MMC method, CLIP text encoder, and the expansion algorithm (see Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 282, + 92, + 460, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "show the learning curves with MMC [36] anchors (the blue curves) in the same training setting. MMC anchors are precomputed by maximizing the distance between anchors [36] and thus the average CoS between MMC anchors is close to 0 (e.g., -0.01 on Cifar100). The accuracy with CLIP anchors increased quite slowly. Even after 200 epoch training, the accuracy was far lower than that of MMC anchors (note that zero-shot recognition cannot be achieved with MMC anchors for the lack of semantic consistency).", + "bbox": [ + 75, + 294, + 468, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The stark discrepancy between the two learning curves motivates us to analyze the essential difference between the two sets of anchors. Inspired by the anchor generation process of MMC, we calculated the CoS between CLIP anchors and found them to be considerably larger than that of MMC anchors (-0.01). The first row of Tab. 1 presents the average CoS between anchors from Cifar100 categories generated by different CLIP models. Note that the high CoS between CLIP anchors is irrelevant to the prompt text. We found that even using several random sentences from Wikipedia to encode anchors by CLIP text encoders, the average CoS remained significantly large.", + "bbox": [ + 75, + 430, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although high CoS seems to have no effect on downstream tasks on clean images [22, 28, 45], we discovered that AT was remarkably affected because higher CoS implies that the anchors are closer to the decision boundary, which could leave more room for adversarial examples [36, 52]. This problem may also explain why TeCoA [34], which also used the text anchors during adaptation, failed to convert CLIP image models to adversially robust models against strong perturbations. Importantly, random or MMC anchors cannot mitigate this issue as they lack the semantic consistency provided by CLIP anchors and thus cannot be used in the ZSL setting. Therefore, the micro designs in our LAAT approach primarily aim to address the challenges associated with CLIP anchors.", + "bbox": [ + 75, + 611, + 468, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Expansion Algorithm", + "text_level": 1, + "bbox": [ + 76, + 832, + 279, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first propose an expansion algorithm to increase the distances between anchors (reducing the CoS). The expansion algorithm can remap the anchors to increase the dis", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c3f91c122e6254f387c51fb762a16792769b24771e32276943c4617f8a4d6405.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeRN50x4RN50x16ViT-B/32ViT-B/16ViT-L/14
Original0.7000.7100.7790.7610.746
Expanded0.2380.2530.1990.1950.222
", + "bbox": [ + 504, + 89, + 887, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Average CoS between anchors before and after expansion algorithm (see Sec. 3.4) from Cifar100 categories generated by different CLIP text encoders.", + "bbox": [ + 498, + 156, + 890, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tances between anchors while preserving the semantic consistency between them. High CoS means the normalized anchors $\\{\\mathbf{a}_i\\}_{i=1}^N$ are dispersed over a cluster of the unit hypersphere (see Fig. 4). As we expect the remapped anchors to be also on the unit hyper-sphere, the expansion algorithm should be designed under spherical coordinates. Let us recap the transformation between $n$ -dimensional spherical coordinate system and Cartesian coordinates. Suppose $(x^{(1)}, x^{(2)}, \\ldots, x^{(n)})$ is the Cartesian coordinate, we can compute them by:", + "bbox": [ + 496, + 222, + 890, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {(1)} = r \\cos \\phi^ {(1)}; \\quad x ^ {(2)} = r \\sin \\phi^ {(1)} \\cos \\phi^ {(2)}; \\quad \\dots ;\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 378, + 862, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {(n - 1)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\cos \\phi^ {(n - 1)}; \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 400, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {(n)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\sin \\phi^ {(n - 1)},\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 421, + 790, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r$ is the radial coordinate and $\\phi^{(1)},\\dots ,\\phi^{(n - 1)}$ are angular coordinates [6]. Here $\\phi^{(1)}$ is the polar angle (taking the three-dimensional spherical coordinate system like longitude and latitude system as an example, the polar angle is the angle with respect to the $z$ axis). If the anchors are clustered around the polar, i.e., $\\phi_0\\coloneqq \\max_i\\phi_i^{(1)} < \\frac{\\pi}{2}$ , where $\\phi_i^{(1)}$ denotes the polar angle of anchor $\\mathbf{a}_i$ , a natural expansion method is to enlarge the polar angle $\\phi_i^{(1)}$ to $\\frac{\\pi}{2}\\cdot \\phi_i^{(1)} / \\phi_0$ and keep the other angular coordinates unchanged. Then the cluster can be expanded to the whole hemisphere. Note that we cannot expand them larger than $\\frac{\\pi}{2}$ as anchors may be close to each other from another hemisphere, which may hurt the semantic consistency of anchors.", + "bbox": [ + 496, + 446, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, given the original anchor $\\mathbf{a}_i$ , the first step of the expansion is to calculate a rotation matrix $\\mathbf{R}$ to transform the center $\\mathbf{v}$ of the clustered $\\mathbf{a}_i$ 's ( $\\mathbf{v} := \\sum_{i=1}^{N} \\mathbf{a}_i / || \\sum_{i=1}^{N} \\mathbf{a}_i||_2$ ) to the polar $\\mathbf{p} = [1,0,\\dots,0]^T$ , and the rotated anchor $\\tilde{\\mathbf{a}}_i$ is computed by: $\\tilde{\\mathbf{a}}_i = \\mathbf{Ra}_i$ . Note that computing $\\mathbf{R}$ between two unit vectors has several ways and we use the Aguilera and Pérez-Aguila [1] algorithm. As $\\tilde{\\mathbf{a}}_i$ 's are clustered around the polar $\\mathbf{p}$ , we then enlarge the polar angle of $\\tilde{\\mathbf{a}}_i$ as described above. Denoting the $j$ -th element of $\\mathbf{a}_i$ as $a_i^{(j)}$ , the largest polar angle between $\\tilde{\\mathbf{a}}_i$ 's and $\\mathbf{p}$ can be computed as $\\phi_0 = \\max_{1 \\leq i \\leq N} \\{\\arccos \\tilde{a}_i^{(1)}\\}$ . According to Eq. (1), the Cartesian coordinates of expanded anchor $\\bar{\\mathbf{a}}_i$ is given by:", + "bbox": [ + 496, + 654, + 892, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {a} _ {i} ^ {(1)} = \\cos \\bar {\\phi} _ {i} ^ {(1)}; \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 859, + 890, + 887 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {a} _ {i} ^ {(j)} = \\tilde {a} _ {i} ^ {(j)} \\cdot \\sin \\tilde {\\phi} _ {i} ^ {(1)} / \\sin \\tilde {\\phi} _ {i} ^ {(1)}, \\quad j = 2, \\dots , n,\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 885, + 852, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24689", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/20e83b15f2d4fc4f986b6d231a0ba81a4a42b0a61e5d9903b91a1a70c6d68a19.jpg", + "image_caption": [ + "Figure 4. An illustration of the expansion operation in 3D space." + ], + "image_footnote": [], + "bbox": [ + 119, + 87, + 428, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\phi}_i^{(1)} = \\arccos \\tilde{a}_i^{(1)}$ denotes the polar angle of each $\\tilde{\\mathbf{a}}_i$ and $\\bar{\\phi}_i^{(1)} = \\frac{\\pi}{2}\\cdot \\tilde{\\phi}_i^{(1)} / \\phi_0$ denotes the expanded polar angle. Fig. 4 illustrates this operation and the detailed derivation is shown in Appendix B. In the end, we use the inverse of $\\mathbf{R}$ to map the expanded anchor $\\bar{\\mathbf{a}}_i$ to the original locations and obtain the final anchor: $\\hat{\\mathbf{a}}_i = \\mathbf{R}^{\\mathrm{T}}\\bar{\\mathbf{a}}_i$ . The pseudo-code of the whole expansion algorithm is provided in Appendix C. As shown, the expansion algorithm can increase the distances between different anchors while maintaining the semantic consistency of the original CLIP anchors (close anchors are still close after expansion).", + "bbox": [ + 75, + 255, + 468, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second row of Tab. 1 shows the average CoS between anchors after the expansion algorithm. Compared with the original CLIP anchors, the average CoS between the expanded CLIP anchors is greatly reduced. Fig. 3 shows the learning curves with the expanded CLIP anchors (the green curves). The final robust accuracy (Fig. 3, right) is quite close to that of MMC anchors.", + "bbox": [ + 75, + 429, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Supervision Objective", + "text_level": 1, + "bbox": [ + 76, + 544, + 282, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to improve the anchor-based AT methods further by introducing an A-CE loss into the supervision objective. The A-CE loss is a composition of CE loss and the anchor alignment process (the $\\cos \\theta$ objective), which can also be seen as a special case of InfoNCE [35] with using $l_{2}$ normalized features. This could be confusing considering that previous works have shown that CE is inferior to other anchor-based objectives such as the $\\theta$ and $\\cos \\theta$ [37] in the standard AT setting (see Sec. 2.1). However, we find that with the high CoS problem, A-CE can boost the performance of anchor-based AT methods in that the softmax function of A-CE could make a relaxation on the anchor alignment objective of the original anchor-based AT methods. Assuming that the image encoder $f_{\\Theta}(\\cdot)$ outputs normalized features $\\mathbf{z} = f_{\\Theta}(\\mathbf{x}) \\in \\mathbb{R}^n$ and $||\\mathbf{z}||_2 = 1$ and taking the $\\cos \\theta$ objective as an example, the $\\cos \\theta$ objective is to maximize $\\mathbf{z}^T\\hat{\\mathbf{a}}_y$ , where $\\hat{\\mathbf{a}}_y$ denotes the GT anchor. It encourages the visual feature's CoS to be as close as possible to the GT anchor, while the A-CE loss, by putting $\\cos \\theta$ into the softmax function, only encourages the visual feature's CoS with GT to be larger than those with other anchors. With A-CE supervision, the output feature could be opti", + "bbox": [ + 75, + 568, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mized to be far away from all anchors, but it can still be correctly classified as long as it is closest to the GT anchor. This will be discussed further in Sec. 4.4. Overall, the supervision objective is:", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - \\log \\frac {\\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {y}\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {i}\\right)} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 162, + 890, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ is adversarial perturbation generated by $\\frac{\\partial L_1}{\\partial\\mathbf{x}}$ . Unlike previous works [39] on standard training, we do not use a temperature parameter $\\tau$ to scale the CoS. This is discussed in Appendix D.", + "bbox": [ + 496, + 215, + 890, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Smoothness Loss", + "text_level": 1, + "bbox": [ + 500, + 287, + 666, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since we prefer the adversarial robustness of the model in the zero-shot transfer setting, we additionally introduce a smoothing loss, encouraging the CoS between adversarial features to be similar to the benign features of an image:", + "bbox": [ + 496, + 310, + 890, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {2} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - f _ {\\Theta} (\\mathbf {x}) ^ {T} f _ {\\Theta} (\\mathbf {x} + \\delta) \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 381, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The smoothing loss is independent of the category labels of training examples. It can be expected to improve the adversarial robustness on novel categories. Finally, the whole loss in the training stage is given by $L = L_{1} + \\alpha L_{2}$ , where $\\alpha$ is a hyper-parameter.", + "bbox": [ + 496, + 411, + 890, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.7. Performing Zero-Shot Recognition", + "text_level": 1, + "bbox": [ + 500, + 497, + 802, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When performing zero-shot recognition on novel categories, the image features of a test image are generated by the image encoder, and the anchors for novel categories are generated by the text encoder and the expansion algorithm. Note that the internal parameters of the expansion algorithm, such as $\\phi_0$ , are pre-computed using anchors from the training categories. Finally, the zero-shot prediction is the category whose text anchor exhibits the highest CoS with the image features.", + "bbox": [ + 496, + 521, + 890, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 670, + 633, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 696, + 689, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. We first consider the standard ZSL setting. Due to the lack of adversarially robust methods in this setting, we compared LAAT with several robust few-shot methods directly: AQ [19], R-MAML [46], and GR [13]. ITS-MAML [15] was not compared as it used a different attack setting and no code was released. Note that zero-shot is harder than few-shot as the latter can have access to a few examples from the novel categories.", + "bbox": [ + 496, + 718, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then considered a more pragmatic version of ZSL, which requires recognizing examples from both seen and novel categories. This is called the generalized zero-shot learning (GZSL) setting [8]. We trained several models with", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24690", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/eaf28adb344707102d204cc0e1fd5bb3b953d0d0c5648a9ac7472b8875c40f7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMethodSettingCleanFGSMPGDCWAA
CIFAR-FSAQ [19]1-shot54.31 ± 0.5032.47 ± 0.4730.41 ± 0.4630.14 ± 0.4627.09 ± 2.07
R-MAML [46]1-shot38.53 ± 0.4726.76 ± 0.4426.21 ± 0.4329.76 ± 0.4616.67 ± 5.48
GR [13]1-shot45.27 ± 0.4939.60 ± 0.4638.03 ± 0.4637.00 ± 0.46-
LAATzero-shot55.60 ± 0.4641.17 ± 0.4440.12 ± 0.4437.35 ± 0.4436.45 ± 1.90
LAAT1-shot57.88 ± 0.4543.73 ± 0.4442.74 ± 0.4440.10 ± 0.4539.23 ± 2.12
MiniImageNetAQ [19]1-shot39.17 ± 0.4022.53 ± 0.3520.95 ± 0.3419.33 ± 0.3817.85 ± 1.41
R-MAML [46]1-shot36.18 ± 0.3623.85 ± 0.4621.74 ± 0.3419.68 ± 0.4013.69 ± 3.95
GR [13]1-shot36.14 ± 0.4529.23 ± 0.3327.57 ± 0.3826.61 ± 0.33-
LAATzero-shot46.97 ± 0.3530.69 ± 0.3029.27 ± 0.3127.91 ± 0.2925.24 ± 1.42
LAAT1-shot47.37 ± 0.3531.55 ± 0.3129.93 ± 0.3128.19 ± 0.2925.75 ± 1.45
", + "bbox": [ + 89, + 87, + 883, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ccf0252226909b80f774d0a123e35dc3412f5d1dd161d514b3d9377a18bd5578.jpg", + "table_caption": [ + "Table 2. The comparison of LAAT in 5-way zero-shot setting and previous works in 5-way 1-shot setting under $\\epsilon = 8 / 255$ . The $95\\%$ confidence interval was reported. Here all methods used Conv4-512 as the image model. The accuracy of GR under AA is not shown as it was not reported and no code was released. We also list the performance of LAAT in the 1-shot setting with image-text blended anchors (discussed in Sec. 4.3). The results higher than previous methods in each column are in bold." + ], + "table_footnote": [], + "table_body": "
MethodModelTrainingCifar100 100aPY 32AwA2 50COCO Obj. 80STL10 10OxfordPet 37DTD 47Caltech101 101Caltech256 257SUN 397
TeCoA [34]ViT-B/32*ε = 1/2550.004.390.881.070.980.001.431.661.330.00
ViT-B/32ε = 8/2550.005.271.451.331.460.281.892.972.360.14
ViT-B/16ε = 8/2550.206.451.951.561.660.492.257.033.810.29
LAATViT-B/32ε = 8/2556.6424.3223.3418.5641.1220.314.5934.4321.394.30
ViT-B/16ε = 8/2557.3234.9726.2120.6149.4128.814.7937.8924.025.18
", + "bbox": [ + 86, + 323, + 883, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Adversarial accuracy (\\%) of models trained on ImageNet-1K on ten downstream datasets in GZSL setting. The number of categories is shown below the dataset name. * indicates the results of the originally released checkpoint of TeCoA with the training perturbation $\\epsilon = 1/255$ . The robustness was all evaluated by PGD-20 under $\\epsilon = 8/255$ .", + "bbox": [ + 75, + 435, + 895, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LAAT on ImageNet-1K and evaluated them directly on ten downstream datasets. In this setting, we mainly compare LAAT with TeCoA [34]. We did not compare with Yucel et al. [51] and Zhang et al. [55] as they relied on image attributes and were unsuitable for this setting.", + "bbox": [ + 75, + 503, + 470, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. When comparing with few-shot methods, we used two popular few-shot benchmarks CIFAR-FS [5] and MiniImageNet [43]. CIFAR-FS, a variant of Cifar100 [25], contains 64, 16, and 20 categories for training, validation, and testing, respectively. MiniImageNet, a subset of ImageNet-1k [40], contains 600 images of $84 \\times 84$ size per category. We used the same dataset splitting setting as previous works [13, 19, 46], for training and testing.", + "bbox": [ + 75, + 595, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the GZSL setting, we tested models trained on ImageNet-1K on ten downstream datasets such as AwA2 [48] and aPY [17]. Note that all of these datasets include several novel categories that have never appeared in ImageNet-1K, e.g., statue and ass in aPY [17], blue whale and dolphin in AwA2 [48]. More introduction to the datasets is shown in the Appendix E.", + "bbox": [ + 75, + 717, + 470, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Unless otherwise stated, we used the CLIP ViT-B/16 text encoder and the prompt text was set as \"This is a photo of a {}\". $\\alpha$ was set to 3.", + "bbox": [ + 76, + 838, + 470, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the standard ZSL setting, we followed previous few-", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shot works [13, 46] and used the widely applied model Conv4-512 [54] as the image model. The training recipe strictly followed GR [13]. We performed AT by generating adversarial examples via PGD method with maximum adversarial perturbation $\\epsilon = 8 / 255$ in $l_{\\infty}$ -norm bound, with $T = 7$ iterative steps (step size $s = 2 / 255$ ).", + "bbox": [ + 496, + 503, + 890, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the GZSL setting, we trained two image models ViT-B/32 and ViT-B/16 [14] with LAAT on the large ImageNet-1K. The training basically followed the recipe proposed by Debenedetti et al. [12]. Here we performed AT by generating adversarial examples via PGD method with $\\epsilon = 8 / 255$ in $l_{\\infty}$ -norm bound and $T = 2$ iterative steps ( $s = 8 / 255$ ).", + "bbox": [ + 496, + 595, + 890, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Results", + "text_level": 1, + "bbox": [ + 500, + 695, + 692, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Standard ZSL Setting. We performed experiments in the standard and usual 5-way one/zero-shot setting [13, 19, 46]. Following previous works [19, 46], we report the robustness of our method against common adversarial attacks, including FGSM [20], PGD [32] with 20 iterative steps and step size $s = 2 / 255$ (denoted as PGD-20 in the following text), CW [7] (optimized by PGD for 30 steps with step size $s = 0.8 / 255$ ), and AutoAttack (AA) [11], an effective ensemble attack (several variants of PGD and one query-based black-box Square attack [2]) used to assess adversarial robustness. The maximal perturbation was set to commonly used $\\epsilon = 8 / 255$ in $l_{\\infty}$ setting. The experi", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24691", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4121eefba3a8e2b786e703bc121c0526128dab87eb520566af9221000fc5223a.jpg", + "image_caption": [ + "(a) Benign accuracy." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 267, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10e7f7101445f6788aa6f1609ee5038c08bf84e34e76c4a9734b9686f8d6707d.jpg", + "image_caption": [ + "(b) Robust accuracy.", + "Figure 5. Classification accuracy on both benign and adversarial examples in 5-way few-shot setting on CIFAR-FS, with image feature anchors or with image-text blended anchors. The dashed line denotes 5-way zero-shot accuracy." + ], + "image_footnote": [], + "bbox": [ + 274, + 92, + 460, + 199 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mental results are reported from 2000 randomly sampled 5-way tasks. The results presented in Tab. 2 demonstrate that LAAT in the zero-shot setting achieved competitive performance with several methods in the few-shot setting. Notably, the adversarial robustness of LAAT models even surpassed that of the previous SOTA adversarily robust few-shot method GR [13]. Additionally, our method exhibited significantly better clean accuracy than all three few-shot methods. We note that these results should be interpreted cautiously because LAAT and these few-shot works used different auxiliary information: the former used text information of the test category names and the latter relied on a small number of example images from the test categories.", + "bbox": [ + 75, + 309, + 468, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GZSL Setting across Datasets. In the GZSL setting, the models were evaluated in $N$ -way classification setting, where $N$ is the number of categories in each dataset. All images on downstream datasets are resized to $224 \\times 224$ in inference. Here we compared LAAT with TeCoA. TeCoA originally performed adversarial fine-tuning on ImageNet-1K with maximum perturbation $\\epsilon = 1/255$ . For a fair comparison, here adversarial fine-tuning was also performed with the released code of TeCoA [34]. Tab. 3 shows the zero-shot results on ten downstream datasets under PGD-20. Here the accuracy is reported by sampling 100 examples per category. We can see that TeCoA almost failed to train adversarily robust models in the common $\\epsilon$ setting, while the models trained with LAAT can obtain substantial adversarial robustness on several downstream datasets.", + "bbox": [ + 75, + 518, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that the models trained with LAAT did not perform quite well on Cifar100, DTD [9], and SUN [49]. This could be caused by the original small resolution of Cifar100 $(32 \\times 32)$ [12] and the large inter-class differences between DTD, SUN, and ImageNet-1K.", + "bbox": [ + 75, + 746, + 468, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Extending LAAT to Few-Shot Setting", + "text_level": 1, + "bbox": [ + 76, + 830, + 401, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our zero-shot models supervised by LAAT can be naturally extended to the few-shot setting and the robustness can be further improved. Given the images of novel categories as", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af8bcb23df979fd8e2b254de6d0473c5fc71bfac2ade925131792fe57fa3a9ec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ObjectiveCleanRobust
Eucl39.1728.04
θ37.9227.84
cosθ41.9127.91
", + "bbox": [ + 598, + 88, + 794, + 147 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6dfbd35b776eb5061825e811d09ced29553af95ff3490959b5c20371899eefb1.jpg", + "table_caption": [ + "Table 4. Classification accuracy (%) on CIFAR-FS supervised by the original CLIP anchors with different optimization objectives." + ], + "table_footnote": [], + "table_body": "
EXPA-CESMCleanRobust
41.9127.91
✓46.9832.71
✓✓54.0237.69
✓✓✓55.6040.12
", + "bbox": [ + 566, + 199, + 826, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Classification accuracy $(\\%)$ of Conv4-512 on CIFAR-FS in different ablation experiments. The first two lines were the results obtained with $\\cos \\theta$ as the objective.", + "bbox": [ + 498, + 282, + 890, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the support set, we can follow the previous few-shot SOTA method [13] and use the prototype-based metric learning [41] to build image feature-based anchors. That is, in the $K$ -shot case, the image anchor $\\mathbf{v}_y$ for a novel category $y$ can be computed as $\\mathbf{v}_y = \\mathrm{Norm}_2\\{\\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}$ , where $\\mathbf{x}_i$ denotes the images of $y$ in the support set and $\\mathrm{Norm}_2\\{\\cdot\\}$ denotes the $l_2$ normalization. We can also weight the original text anchor $\\mathbf{a}_y$ from the text encoder and the image feature from the support set to build an image-text blended anchor for $y$ : $\\hat{\\mathbf{a}}_y^* = \\mathrm{Norm}_2\\{\\beta \\cdot \\hat{\\mathbf{a}}_y + \\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}$ . Here the text anchor can be regarded as the prior knowledge of $y$ . We set $\\beta = 2$ in all experiments.", + "bbox": [ + 496, + 352, + 890, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the results of different models with image-text blended anchors in the 1-shot setting. The 1-shot performance of LAAT was better than the zero-shot performance. Fig. 5 further shows the results with two different anchors $\\hat{\\mathbf{a}}^{\\star}$ and $\\mathbf{v}$ in $K$ -shot settings. We can see that the text anchor used to perform zero-shot recognition boosted the performance in various $K$ -shot settings, especially when the support set is small.", + "bbox": [ + 496, + 536, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 667, + 653, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limited by computing resources, the ablation study is not studied on ImageNet-1K. Unless specified, it was performed with Conv4-512 in the 5-way zero-shot setting. The robustness was evaluated by the PGD-20 attack method.", + "bbox": [ + 496, + 691, + 890, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of each design. In Fig. 3, we have shown that if the model is supervised by CoS between the original CLIP anchors and the output feature, say, $\\cos \\theta$ , it is difficult to converge under AT. In Tab. 4, we show the results of other anchor-based AT optimization objectives, such as the angle [37] and the Euclidean distance [36], denoted as $\\theta$ and Eucl respectively, in the ZSL setting. The results indicate that none of these three optimization objectives led to good classification accuracy. Note that random guessing in", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24692", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1df121190250c86ba1adbdb484fb645c1f792b5e99b9b76ec7a470d3f1226e7b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AT MethodHyper-ParameterCleanRobust
EXP + A-CE + TRADES1/λ = 155.2619.22
1/λ = 655.8225.02
EXP + A-CE + SMα = 1.055.6239.95
α = 3.055.6040.12
α = 6.053.0140.29
", + "bbox": [ + 86, + 88, + 459, + 175 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df51eb95c67a0606ea19945275bf52d6cdb3717fa5e0dac94137c479b28427cd.jpg", + "table_caption": [ + "Table 6. Classification accuracy (%) on CIFAR-FS with different AT methods." + ], + "table_footnote": [], + "table_body": "
MethodSettingConv4-64ResNet12
CleanRobustCleanRobust
AQ1-shot43.7427.5443.5831.45
R-MAML1-shot32.4624.6142.5432.54
GR1-shot44.5137.4548.1339.29
LAATzero-shot49.8438.7458.3742.88
", + "bbox": [ + 99, + 227, + 449, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "this setting can obtain $20\\%$ accuracy.", + "bbox": [ + 76, + 380, + 323, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We then conducted an ablation study to show the effectiveness of each design in LAAT: the expansion algorithm, the A-CE loss, and the smoothness loss, denoted as EXP, A-CE, and SM, respectively. The results in Tab. 5 show that all of these designs are helpful in improving the zero-shot adversarial robustness.", + "bbox": [ + 75, + 395, + 468, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As stated in Sec. 3.5, A-CE could make a relaxation on the original anchor-based AT objectives. To validate this, we computed the average CoS between GT anchors and the output visual features of the two models trained: EXP and $\\mathrm{EXP} + \\mathrm{A - CE}$ (EXP was supervised by $\\cos \\theta$ ; see Tab. 5). With A-CE loss, the visual feature's CoS with GT anchors dropped from 0.455 to 0.205 on average, while the accuracies on both benign and adversarial examples were improved (see Tab. 5). These results support our conjecture that some examples not quite close to the GT anchor can still be classified correctly with A-CE supervision.", + "bbox": [ + 75, + 484, + 468, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Smoothness v.s. TRADES. It is seen that SM has similar objectives with TRADES [52], but they are different in many aspects (see Appendix F). Here we also compare SM with TRADES empirically. The experiments were performed based on expanded anchors with A-CE supervision. The results are shown in Tab. 6, together with the most important hyper-parameter of each method. We can clearly see that TRADES performed much worse than our smoothness loss. In addition, the results show that the effectiveness of smoothing loss was not quite sensitive to $\\alpha$ .", + "bbox": [ + 75, + 662, + 468, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different image models. We also performed experiments with other popular image models in the few-shot setting other than Conv4-512, e.g., Conv4-64 [44] and ResNet12 [19]. Tab. 7 shows the results of the two image models on CIFAR-FS. Conv4-64 is smaller than Conv4-512 while", + "bbox": [ + 76, + 825, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f087c58a1eeb5d127d83cd2370a289dccc1e0b75fb9645908a44359e893a6959.jpg", + "table_caption": [ + "Table 7. Classification accuracy (%) on CIFAR-FS with different image models." + ], + "table_footnote": [], + "table_body": "
Text EncoderEmbed-dimCleanRobust
None (MMC, one-shot)51247.5532.79
RN50x464052.2838.65
RN50x1676853.8638.02
ViT-B/3251252.6838.39
ViT-B/1651255.6040.12
ViT-L/1476852.7839.48
", + "bbox": [ + 514, + 88, + 880, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Classification accuracy (\\%) supervised by MMC anchors in one-shot setting and the full LAAT with different CLIP text encoders in zero-shot setting on CIFAR-FS. Embed-dim denotes the dimension of the text anchors.", + "bbox": [ + 498, + 199, + 890, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ResNet12 is slightly larger. The models trained with LAAT showed stronger zero-shot adversarial robustness than those few-shot methods regardless of the model size.", + "bbox": [ + 498, + 281, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different text encoders. Without text encoders, it is impossible for LAAT to achieve zero-shot robustness. In principle, arbitrary text encoders are feasible for LAAT if they have semantic consistency. We show the influence of using different CLIP text encoders in Tab. 8. We also show the results without using any text encoders. Instead, the image model was supervised by the MMC anchors (here zero-shot recognition cannot be performed while one-shot results are reported). We can see that using any text encoders can obtain strong zero-shot robustness, significantly surpassing the few-shot robustness without using text encoders. We also studied the relationship between the distances of different CLIP models' anchors and Cifar100 super-categories. The results in Appendix G further show that semantic consistency of the text encoder is one of the keys to zero-shot adversarial robustness.", + "bbox": [ + 496, + 338, + 890, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 500, + 594, + 748, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel LAAT strategy for adversarially robust image classification in a zero-shot setting, inspired by recent vision-language models. Extensive experiments demonstrate that our method has strong adversarial robustness across different models and several zero-shot settings. We hope our work could encourage more researchers to investigate the robustness in the zero-shot setting, which could be a promising way to improve the scalability and flexibility of adversarially robust models. Limited by the data and computing resources, we cannot perform AT on large-scale image-text pairs directly like CLIP [39]. We believe it could be another promising way to achieve zero-shot adversarial robustness.", + "bbox": [ + 496, + 619, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 660, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (Nos. U2341228 and U19B2034).", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24693", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antonio Aguilera and Ricardo Pérez-Aguila. General n-dimensional rotations. In WSCG, pages 1-8, 2004. 4", + "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: A query-efficient black-box adversarial attack via random search. In Eur. Conf. Comput. Vis. (ECCV), pages 484-501, 2020. 6", + "[3] Anish Athalye et al. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Int. Conf. Mach. Learn. (ICML), pages 274-283, 2018. 2", + "[4] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 2", + "[5] Luca Bertinetto, João F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Int. Conf. Learn. Represent. (ICLR), 2019, 6, 3", + "[6] LE Blumenson. A derivation of n-dimensional spherical coordinates. The American Mathematical Monthly, 67(1):63-66, 1960. 4", + "[7] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy, SP, pages 39-57, 2017. 6", + "[8] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In Eur. Conf. Comput. Vis. (ECCV), pages 52-68, 2016. 5", + "[9] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3606-3613, 2014. 7, 2", + "[10] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, pages 215-223, 2011. 2", + "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Int. Conf. Mach. Learn. (ICML), pages 2206-2216, 2020. 6", + "[12] Edoardo Debenedetti et al. A light recipe to train robust vision transformers. arXiv preprint arXiv:2209.07399, 2022.6, 7", + "[13] Junhao Dong, Yuan Wang, Jianhuang Lai, and Xiaohua Xie. Improving adversarially robust few-shot image classification with generalizable representations. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9015-9024. IEEE, 2022. 3, 5, 6, 7", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent. (ICLR), 2021. 6", + "[15] Xiaoyue Duan, Guoliang Kang, Runqi Wang, Shumin Han, Song Xue, Tian Wang, and Baochang Zhang. Rethinking the number of shots in robust model-agnostic meta-learning. arXiv preprint arXiv:2211.15180, 2022. 3, 5", + "[16] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1625-1634, 2018. 1", + "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David A. Forsyth. Describing objects by their attributes. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1778-1785, 2009. 6, 2", + "[18] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004. 2", + "[19] Micah Goldblum et al. Adversarily robust few-shot learning: A meta-learning approach. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 3, 5, 6, 8", + "[20] Ian J. Goodfellow et al. Explaining and harnessing adversarial examples. In Int. Conf. Learn. Represent. (ICLR), 2015. 6", + "[21] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 2", + "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zero-shot detection via vision and language knowledge distillation. In Int. Conf. Learn. Represent. (ICLR), 2021. 2, 4", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 770-778, 2016. 3", + "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Int. Conf. Mach. Learn. (ICML), pages 4904-4916, 2021. 2", + "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. California Institute of Technology, 2009. 3, 6, 2", + "[26] Huafeng Kuang, Hong Liu, Yongjian Wu, and Rongrong Ji. Semantically consistent visual representation for adversarial robustness. IEEE Transactions on Information Forensics and Security, 2023. 2", + "[27] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 36(3):453-465, 2014. 1", + "[28] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In Int. Conf. Learn. Represent. (ICLR), 2022. 2, 4", + "[29] Xiao Li, Hang Chen, and Xiaolin Hu. On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438, 2023. 2", + "[30] Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, and Xiaolin Hu. Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 2", + "[31] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24694", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2", + "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Int. Conf. Learn. Represent. (ICLR), 2018. 1, 6", + "[33] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. Adv. Neural Inform. Process. Syst. (NeurIPS), 32, 2019. 2", + "[34] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. Int. Conf. Learn. Represent. (ICLR), 2023. 2, 3, 4, 6, 7", + "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5", + "[36] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In Int. Conf. Learn. Represent. (ICLR), 2020. 1, 2, 4, 7", + "[37] Tianyu Pang, Xiao Yang, Yinpeng Dong, Taufik Xu, Jun Zhu, and Hang Su. Boosting adversarial training with hypersphere embedding. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 1, 2, 3, 5, 7", + "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3498-3505, 2012. 2", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 2, 5, 8", + "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. (IJCV), 115(3):211-252, 2015. 2, 6", + "[41] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4077-4087, 2017. 7", + "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In Int. Conf. Learn. Represent. (ICLR), 2014. 1", + "[43] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 6", + "[44] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 8", + "[45] Mengmeng Wang et al. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 3, 5, 6", + "[47] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 69-77, 2016. 2", + "[48] Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - A comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 41(9):2251-2265, 2019. 6, 2", + "[49] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3485-3492, 2010. 7, 2", + "[50] Yunlong Yu, Zhong Ji, Yanwei Fu, Jichang Guo, Yanwei Pang, and Zhongfei (Mark) Zhang. Stacked semantics-guided attention model for fine-grained zero-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6007, 2018. 2", + "[51] Mehmet Kerim Yucel et al. How robust are discriminatively trained zero-shot learning models? Image Vis. Comput., 119: 104392, 2022. 1, 3, 6", + "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In Int. Conf. Mach. Learn. (ICML), pages 7472-7482, 2019. 1, 2, 4, 8", + "[53] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM Int. Conf. Multimedia, pages 5005-5013, 2022. 2", + "[54] Manli Zhang, Jianhong Zhang, Zhiwu Lu, Tao Xiang, Mingyu Ding, and Songfang Huang. IEPT: instance-level and episode-level pretext tasks for few-shot learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 6", + "[55] Xingxing Zhang, Shupeng Gui, Zhenfeng Zhu, Yao Zhao, and Ji Liu. ATZSL: defensive zero-shot recognition in the presence of adversaries. arXiv preprint arXiv:1910.10994, 2019. 1, 3, 6", + "[56] Mo Zhou and Vishal M Patel. Enhancing adversarial robustness for deep metric learning. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15325-15334, 2022. 2", + "[57] Xiaopei Zhu, Xiao Li, Jianmin Li, Zheyao Wang, and Xiaolin Hu. Fooling thermal infrared pedestrian detectors in real world using small bulbs. In AAAI, pages 3616-3624, 2021. 1" + ], + "bbox": [ + 501, + 92, + 890, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24695", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_model.json b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e6499614116ebbbf7d31331989ae4d57a13d090f --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "Language-Driven Anchors for Zero-Shot Adversarial Robustness" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.179, + 0.861, + 0.341 + ], + "angle": 0, + "content": "Xiao Li\\(^{1}\\) Wei Zhang\\(^{1}\\) Yining Liu\\(^{2}\\) Zhanhao Hu\\(^{1}\\) Bo Zhang\\(^{1}\\) Xiaolin Hu\\(^{1,3*}\\)\\\n\\(^{1}\\)Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China\\\n\\(^{2}\\)Harbin Institute of Technology, Weihai, China\\\n\\(^{3}\\)Chinese Institute for Brain Research (CIBR), Beijing, China\\\n{lixiao20, zhang-w19, huzhanha17}@mail.tsinghua.edu.cn\\\n22s030184@stu.hit.edu.cn\\\n{dcszb, xlhu}@mail.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.367, + 0.314, + 0.384 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.399, + 0.474, + 0.794 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) are known to be susceptible to adversarial attacks. Previous researches mainly focus on improving adversarial robustness in the fully supervised setting, leaving the challenging domain of zero-shot adversarial robustness an open question. In this work, we investigate this domain by leveraging the recent advances in large vision-language models, such as CLIP, to introduce zero-shot adversarial robustness to DNNs. We propose LAAT, a Language-driven, Anchor-based Adversarial Training strategy. LAAT utilizes the features of a text encoder for each category as fixed anchors (normalized feature embeddings) for each category, which are then employed for adversarial training. By leveraging the semantic consistency of the text encoders, LAAT aims to enhance the adversarial robustness of the image model on novel categories. However, naively using text encoders leads to poor results. Through analysis, we identified the issue to be the high cosine similarity between text encoders. We then design an expansion algorithm and an alignment cross-entropy loss to alleviate the problem. Our experimental results demonstrated that LAAT significantly improves zero-shot adversarial robustness over state-of-the-art methods. LAAT has the potential to enhance adversarial robustness by large-scale multimodal models, especially when labeled data is unavailable during training. Code is available at https://github.com/LixiaoTHU/LAAT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.821, + 0.21, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Adversarial attacks [42], by adding deliberately designed perturbations to inputs, have posed a significant threat to" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.365, + 0.877, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.515, + 0.895, + 0.614 + ], + "angle": 0, + "content": "Figure 1. The illustration of zero-shot ability with LAAT. Different colors of the marks indicate different categories. When a model is adversarially trained with the text anchors of table and lion (seen categories), it can recognize adversarial examples of the two categories (grey and green). Then due to the text anchors of chair and tiger (novel categories) being close to those of table and lion, respectively, the model can also recognize the two novel categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.824 + ], + "angle": 0, + "content": "the application of Deep Neural Networks (DNNs) [16, 57]. In response to the threats, various techniques have been proposed to enhance adversarial robustness of DNNs. Among these, Adversarial Training (AT) and its variants [32, 36, 37, 52] have shown to be highly effective [16, 42, 57]. However, applying AT from scratch to novel datasets or categories can be time-consuming. In real-world scenarios, it is common for models to encounter examples from novel categories, known as zero-shot learning (ZSL) setting [27]. If adversarially trained models possess zero-shot capability, it can significantly enhance the scalability and practical utility of adversarially robust models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Adversarial robustness in the ZSL setting is challenging. Earlier attempts [51, 55] try to combine attribute-based ZSL methods with AT. However, these methods yielded modest improvements and lacked scalability due to the practical difficulties in obtaining attributes. Recently, large-scale" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.224, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "vision-language models such as CLIP [39] have made significant strides and are considered foundational models for various downstream applications. These models, trained on extensive text-image data, demonstrate strong zero-shot capability by using textual descriptions to recognize new visual categories, eliminating the need for explicit reliance on image attributes. In this work, different from previous attribute-based ZSL methods, we explore the utilization of foundation models to introduce zero-shot adversarial robustness to DNNs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.578 + ], + "angle": 0, + "content": "We propose a Language-driven, Anchor-based Adversarial Training strategy, denoted as LAAT, to enhance the zero-shot adversarial robustness. Specifically, we use the popular text encoder from vision-language models, specifically CLIP [39], to obtain \\( l_{2} \\) normalized feature embedding (named anchor in this paper) for the text (label name) corresponding to image categories. The text encoder exhibits a property called semantic consistency, where semantically similar categories are mapped to neighboring anchors in the feature space. Then we utilize the text anchors to supervise the image classification model with AT. After this process, the image model not only obtains adversarial robustness on seen categories but also aligns the image features with the text anchors. During zero-shot inference, by leveraging the semantic consistency between seen and novel categories, the image model can recognize novel categories while maintaining adversarial robustness. Fig. 1 illustrates the idea. We note that one recent work, TeCoA [34], also utilizes the ability of large-scale vision-language models to enhance the zero-shot adversarial robustness. However, TeCoA was only shown effective for extremely tiny adversarial noise (i.e., \\( \\epsilon = 1 / 255 \\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.473, + 0.808 + ], + "angle": 0, + "content": "As CLIP models are not designed for adversarial robustness, directly applying the obtained anchors from CLIP text encoders to anchor-based AT [36, 37] (refer to Sec. 2.1) results in poor performance. We attribute it to the high Cosine Similarity (CoS) problem, i.e., the average cosine similarity between CLIP anchors are too high to be directly used in AT. See Sec. 3.3 for details. This problem has never been discovered in previous work utilizing CLIP models [22, 28, 34, 45]. To address it, we propose several techniques. We first introduce an expansion algorithm that maps the original CLIP anchors to new anchors with increased distances while maintaining semantic consistency. We then incorporate an Alignment Cross-Entropy (A-CE) loss as the optimization objective. In addition, we use a smoothness loss to enhance the robustness on unseen categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The experimental results showed that the models trained with LAAT achieved remarkable zero-shot performance against strong adversarial perturbations, even surpassing the previous state-of-the-art (SOTA) adversarily robust few-shot methods. We also trained several models with LAAT on a large dataset ImageNet-1K [40]. These mod" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "els showed substantial adversarial robustness across various downstream datasets in the zero-shot setting, significantly outperforming the recent TeCoA method [34]. The main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.152, + 0.894, + 0.196 + ], + "angle": 0, + "content": "- We identify the high CoS problem in CLIP anchors for adversarial robustness and propose effective techniques to alleviate it." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.198, + 0.895, + 0.258 + ], + "angle": 0, + "content": "- The models trained by LAAT show substantial adversarial robustness across downstream datasets, suggesting that AT in the ZSL setting could be a promising way to improve the practical usefulness of AT." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.152, + 0.895, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.273, + 0.642, + 0.289 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.298, + 0.671, + 0.313 + ], + "angle": 0, + "content": "2.1. Anchor-based AT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.596 + ], + "angle": 0, + "content": "AT has become one of the most effective strategies to improve adversarial robustness [3]. Several studies try to improve AT from various aspects [26, 29, 30, 36, 37, 52]. One series of works focuses on refining cross-entropy loss, which is the most commonly used loss in classification, to obtain better robustness with AT [36, 37]. Pang et al. [36] boosts AT by introducing the Max-Mahalanobis center (MMC) loss, a Euclidean loss where there is a fixed feature vector \\(\\mu_y\\) for each category \\(y\\), and the optimization object is \\(||\\mathbf{z} - \\mu_y||_2^2\\), where \\(\\mathbf{z}\\) is the output feature. Pang et al. [37] uses Hypersphere Embedding to boost AT. It normalizes both weights \\(\\mathbf{W}\\) and features \\(\\mathbf{z}\\) of the output layer, and proposes two losses in the form of \\(\\cos \\theta\\) and \\(\\theta\\), where \\(\\theta\\) is the angle between \\(\\frac{\\mathbf{W}}{||\\mathbf{W}||_2}\\) and \\(\\frac{\\mathbf{z}}{||\\mathbf{z}||_2}\\). We call these AT methods anchor-based methods as they all used \\(l_{2}\\) normalized feature embeddings (anchors) for training. Besides, several metric learning works [33, 56] also used a similar paradigm to anchor-based AT for enhancing robustness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.768, + 0.622 + ], + "angle": 0, + "content": "2.2. Language-driven Recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Earlier ZSL models exploit auxiliary information, commonly in the form of image attributes, to transfer knowledge between seen and novel categories [47, 50]. Recently, language-driven recognition has caught attention in the ZSL field, especially with the recent advance of large-scale pretrained vision-language models [24, 39]. CLIP [39] demonstrated that classic recognition tasks can strongly benefit from millions of raw texts pertaining to images. CLIP uses an independent image encoder and text encoder to perform contrastive learning on 400M image-text pairs. After pretraining, natural language is used to reference visual concepts by computing the CoS between text features and the visual feature, enabling the zero-shot transfer of the model to downstream classification datasets. The basic paradigm of CLIP can be extended to several tasks such as semantic segmentation and object detection [22, 28, 45]. However, recent works show that CLIP models lack adversarial robustness and can be easily attacked [4, 34, 53]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24687" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.887, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.893, + 0.38 + ], + "angle": 0, + "content": "Figure 2. The pipeline of LAAT. Only the image encoder is trainable in the figure. \\(\\odot\\) indicates computing the CoS. Red arrows indicate the adversarial example generation process and brown arrows indicate the inference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.405, + 0.465, + 0.423 + ], + "angle": 0, + "content": "2.3. Adversarily Robust Zero/few-shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.639 + ], + "angle": 0, + "content": "An early preprint work [55] combines AT with attribute-based ZSL methods but only gets a mild improvement over non-AT models. In addition, the method is hard to generalize without the attribute annotations. Yucel et al. [51] evaluates the adversarial robustness of classic attribute-based zero-shot models and concludes that adversarial robustness for ZSL field is suffering an immature state. Mao et al. [34] proposes several techniques such as visual prompting and fine-tuning to adapt large-scale models, e.g., the image models of CLIP, for zero-shot adversarial robustness, while their method called TeCoA was only shown to be effective under extremely tiny adversarial noise. Different from TeCoA, LAAT only uses the anchors from the CLIP text encoders." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.471, + 0.748 + ], + "angle": 0, + "content": "A closely related area is adversarially robust few-shot learning, which has caught attention recently, too. We also compare LAAT with recent few-shot methods directly to comprehensively evaluate it. AQ [19], R-MAML [46], and ITS-MAML [15] study meta-learning jointly with AT. GR [13] tries to learn a generalizable robust representation by the combination of several AT techniques." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.156, + 0.777 + ], + "angle": 0, + "content": "3. LAAT" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.239, + 0.803 + ], + "angle": 0, + "content": "3.1. Overall Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "LAAT first obtains fixed text anchors from the text encoder, then uses them to adversarially supervise the training of image classification model. After AT, the image model will obtain adversarial robustness on seen categories and also align the image features with the text anchors to obtain zero-shot ability (see Fig. 1). The overall pipeline of LAAT is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.527 + ], + "angle": 0, + "content": "illustrated in Fig. 2. Given the original text anchors, LAAT uses an expansion algorithm to obtain fixed Ground-Truth (GT) anchors suitable for AT, then uses an image encoder to extract both benign and adversarial features. The CoS between adversarial features and GT anchors are maximized by minimizing the A-CE loss \\((L_{1})\\). Besides, a smoothness loss \\((L_{2})\\) is applied to the adversarial and benign features. We then describe each module of LAAT." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.537, + 0.72, + 0.552 + ], + "angle": 0, + "content": "3.2. Text Anchor Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.895, + 0.697 + ], + "angle": 0, + "content": "Given the names of \\(N\\) categories in the training set, they are first filled into \\(N\\) sentences with a fixed prompt text, such as \"A photo of {}\". After that, these sentences are encoded into \\(N\\) anchors \\(\\{\\mathbf{a}_i\\}_{i=1}^N\\) by a fixed pre-trained text encoder, where \\(\\mathbf{a}_i \\in \\mathbb{R}^n\\) and \\(||\\mathbf{a}_i||_2 = 1\\). Many architectures are feasible if they have semantic consistency as mentioned before and in this work, we use the large-scale pre-trained CLIP text encoder. However, CLIP anchors cannot be used directly due to the high CoS problem, as analyzed below." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.706, + 0.783, + 0.723 + ], + "angle": 0, + "content": "3.3. High Cosine Similarity Problem" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We first make an intriguing observation that directly using CLIP anchors for anchor-based AT poses significant challenges in achieving convergence. This phenomenon is universal among different CLIP text encoders and different models. To illustrate this, we adversarially trained a ResNet18 [23] on Cifar100 [25] by maximizing the \\(\\cos \\theta\\) objective [37], i.e., the CoS between the model output \\(\\mathbf{z}\\) and the GT anchor \\(\\mathbf{a}_y\\) [37], which is computed as \\(\\frac{\\mathbf{z}^T\\mathbf{a}_y}{||\\mathbf{z}||_2}\\) (note that \\(||\\mathbf{a}_y||_2 = 1\\)). The experimental details can be found in Appendix A. Fig. 3 shows the learning curves with CLIP anchors (the orange curves). For comparison, here we also" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.266, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.203, + 0.227, + 0.216 + ], + "angle": 0, + "content": "(a) Clean accuracy." + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.093, + 0.462, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.203, + 0.428, + 0.216 + ], + "angle": 0, + "content": "(b) Robust accuracy." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.227, + 0.47, + 0.27 + ], + "angle": 0, + "content": "Figure 3. Learning curves of AT supervised by \\(\\cos \\theta\\) with fixed anchors generated from MMC method, CLIP text encoder, and the expansion algorithm (see Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.469, + 0.431 + ], + "angle": 0, + "content": "show the learning curves with MMC [36] anchors (the blue curves) in the same training setting. MMC anchors are precomputed by maximizing the distance between anchors [36] and thus the average CoS between MMC anchors is close to 0 (e.g., -0.01 on Cifar100). The accuracy with CLIP anchors increased quite slowly. Even after 200 epoch training, the accuracy was far lower than that of MMC anchors (note that zero-shot recognition cannot be achieved with MMC anchors for the lack of semantic consistency)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.431, + 0.469, + 0.612 + ], + "angle": 0, + "content": "The stark discrepancy between the two learning curves motivates us to analyze the essential difference between the two sets of anchors. Inspired by the anchor generation process of MMC, we calculated the CoS between CLIP anchors and found them to be considerably larger than that of MMC anchors (-0.01). The first row of Tab. 1 presents the average CoS between anchors from Cifar100 categories generated by different CLIP models. Note that the high CoS between CLIP anchors is irrelevant to the prompt text. We found that even using several random sentences from Wikipedia to encode anchors by CLIP text encoders, the average CoS remained significantly large." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.469, + 0.822 + ], + "angle": 0, + "content": "Although high CoS seems to have no effect on downstream tasks on clean images [22, 28, 45], we discovered that AT was remarkably affected because higher CoS implies that the anchors are closer to the decision boundary, which could leave more room for adversarial examples [36, 52]. This problem may also explain why TeCoA [34], which also used the text anchors during adaptation, failed to convert CLIP image models to adversially robust models against strong perturbations. Importantly, random or MMC anchors cannot mitigate this issue as they lack the semantic consistency provided by CLIP anchors and thus cannot be used in the ZSL setting. Therefore, the micro designs in our LAAT approach primarily aim to address the challenges associated with CLIP anchors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.28, + 0.849 + ], + "angle": 0, + "content": "3.4. Expansion Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first propose an expansion algorithm to increase the distances between anchors (reducing the CoS). The expansion algorithm can remap the anchors to increase the dis" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.09, + 0.888, + 0.148 + ], + "angle": 0, + "content": "
TypeRN50x4RN50x16ViT-B/32ViT-B/16ViT-L/14
Original0.7000.7100.7790.7610.746
Expanded0.2380.2530.1990.1950.222
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.157, + 0.892, + 0.199 + ], + "angle": 0, + "content": "Table 1. Average CoS between anchors before and after expansion algorithm (see Sec. 3.4) from Cifar100 categories generated by different CLIP text encoders." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.223, + 0.892, + 0.374 + ], + "angle": 0, + "content": "tances between anchors while preserving the semantic consistency between them. High CoS means the normalized anchors \\(\\{\\mathbf{a}_i\\}_{i=1}^N\\) are dispersed over a cluster of the unit hypersphere (see Fig. 4). As we expect the remapped anchors to be also on the unit hyper-sphere, the expansion algorithm should be designed under spherical coordinates. Let us recap the transformation between \\(n\\)-dimensional spherical coordinate system and Cartesian coordinates. Suppose \\((x^{(1)}, x^{(2)}, \\ldots, x^{(n)})\\) is the Cartesian coordinate, we can compute them by:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.379, + 0.864, + 0.397 + ], + "angle": 0, + "content": "\\[\nx ^ {(1)} = r \\cos \\phi^ {(1)}; \\quad x ^ {(2)} = r \\sin \\phi^ {(1)} \\cos \\phi^ {(2)}; \\quad \\dots ;\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.401, + 0.891, + 0.419 + ], + "angle": 0, + "content": "\\[\nx ^ {(n - 1)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\cos \\phi^ {(n - 1)}; \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.422, + 0.792, + 0.44 + ], + "angle": 0, + "content": "\\[\nx ^ {(n)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\sin \\phi^ {(n - 1)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.653 + ], + "angle": 0, + "content": "where \\(r\\) is the radial coordinate and \\(\\phi^{(1)},\\dots ,\\phi^{(n - 1)}\\) are angular coordinates [6]. Here \\(\\phi^{(1)}\\) is the polar angle (taking the three-dimensional spherical coordinate system like longitude and latitude system as an example, the polar angle is the angle with respect to the \\(z\\) axis). If the anchors are clustered around the polar, i.e., \\(\\phi_0\\coloneqq \\max_i\\phi_i^{(1)} < \\frac{\\pi}{2}\\), where \\(\\phi_i^{(1)}\\) denotes the polar angle of anchor \\(\\mathbf{a}_i\\), a natural expansion method is to enlarge the polar angle \\(\\phi_i^{(1)}\\) to \\(\\frac{\\pi}{2}\\cdot \\phi_i^{(1)} / \\phi_0\\) and keep the other angular coordinates unchanged. Then the cluster can be expanded to the whole hemisphere. Note that we cannot expand them larger than \\(\\frac{\\pi}{2}\\) as anchors may be close to each other from another hemisphere, which may hurt the semantic consistency of anchors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.893, + 0.856 + ], + "angle": 0, + "content": "Thus, given the original anchor \\(\\mathbf{a}_i\\), the first step of the expansion is to calculate a rotation matrix \\(\\mathbf{R}\\) to transform the center \\(\\mathbf{v}\\) of the clustered \\(\\mathbf{a}_i\\)'s (\\(\\mathbf{v} := \\sum_{i=1}^{N} \\mathbf{a}_i / || \\sum_{i=1}^{N} \\mathbf{a}_i||_2\\)) to the polar \\(\\mathbf{p} = [1,0,\\dots,0]^T\\), and the rotated anchor \\(\\tilde{\\mathbf{a}}_i\\) is computed by: \\(\\tilde{\\mathbf{a}}_i = \\mathbf{Ra}_i\\). Note that computing \\(\\mathbf{R}\\) between two unit vectors has several ways and we use the Aguilera and Pérez-Aguila [1] algorithm. As \\(\\tilde{\\mathbf{a}}_i\\)'s are clustered around the polar \\(\\mathbf{p}\\), we then enlarge the polar angle of \\(\\tilde{\\mathbf{a}}_i\\) as described above. Denoting the \\(j\\)-th element of \\(\\mathbf{a}_i\\) as \\(a_i^{(j)}\\), the largest polar angle between \\(\\tilde{\\mathbf{a}}_i\\)'s and \\(\\mathbf{p}\\) can be computed as \\(\\phi_0 = \\max_{1 \\leq i \\leq N} \\{\\arccos \\tilde{a}_i^{(1)}\\}\\). According to Eq. (1), the Cartesian coordinates of expanded anchor \\(\\bar{\\mathbf{a}}_i\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.861, + 0.891, + 0.888 + ], + "angle": 0, + "content": "\\[\n\\bar {a} _ {i} ^ {(1)} = \\cos \\bar {\\phi} _ {i} ^ {(1)}; \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.886, + 0.853, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\bar {a} _ {i} ^ {(j)} = \\tilde {a} _ {i} ^ {(j)} \\cdot \\sin \\tilde {\\phi} _ {i} ^ {(1)} / \\sin \\tilde {\\phi} _ {i} ^ {(1)}, \\quad j = 2, \\dots , n,\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.088, + 0.429, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.081, + 0.217, + 0.466, + 0.232 + ], + "angle": 0, + "content": "Figure 4. An illustration of the expansion operation in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.256, + 0.47, + 0.429 + ], + "angle": 0, + "content": "where \\(\\tilde{\\phi}_i^{(1)} = \\arccos \\tilde{a}_i^{(1)}\\) denotes the polar angle of each \\(\\tilde{\\mathbf{a}}_i\\) and \\(\\bar{\\phi}_i^{(1)} = \\frac{\\pi}{2}\\cdot \\tilde{\\phi}_i^{(1)} / \\phi_0\\) denotes the expanded polar angle. Fig. 4 illustrates this operation and the detailed derivation is shown in Appendix B. In the end, we use the inverse of \\(\\mathbf{R}\\) to map the expanded anchor \\(\\bar{\\mathbf{a}}_i\\) to the original locations and obtain the final anchor: \\(\\hat{\\mathbf{a}}_i = \\mathbf{R}^{\\mathrm{T}}\\bar{\\mathbf{a}}_i\\). The pseudo-code of the whole expansion algorithm is provided in Appendix C. As shown, the expansion algorithm can increase the distances between different anchors while maintaining the semantic consistency of the original CLIP anchors (close anchors are still close after expansion)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.47, + 0.535 + ], + "angle": 0, + "content": "The second row of Tab. 1 shows the average CoS between anchors after the expansion algorithm. Compared with the original CLIP anchors, the average CoS between the expanded CLIP anchors is greatly reduced. Fig. 3 shows the learning curves with the expanded CLIP anchors (the green curves). The final robust accuracy (Fig. 3, right) is quite close to that of MMC anchors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.283, + 0.561 + ], + "angle": 0, + "content": "3.5. Supervision Objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We propose to improve the anchor-based AT methods further by introducing an A-CE loss into the supervision objective. The A-CE loss is a composition of CE loss and the anchor alignment process (the \\(\\cos \\theta\\) objective), which can also be seen as a special case of InfoNCE [35] with using \\(l_{2}\\) normalized features. This could be confusing considering that previous works have shown that CE is inferior to other anchor-based objectives such as the \\(\\theta\\) and \\(\\cos \\theta\\) [37] in the standard AT setting (see Sec. 2.1). However, we find that with the high CoS problem, A-CE can boost the performance of anchor-based AT methods in that the softmax function of A-CE could make a relaxation on the anchor alignment objective of the original anchor-based AT methods. Assuming that the image encoder \\(f_{\\Theta}(\\cdot)\\) outputs normalized features \\(\\mathbf{z} = f_{\\Theta}(\\mathbf{x}) \\in \\mathbb{R}^n\\) and \\(||\\mathbf{z}||_2 = 1\\) and taking the \\(\\cos \\theta\\) objective as an example, the \\(\\cos \\theta\\) objective is to maximize \\(\\mathbf{z}^T\\hat{\\mathbf{a}}_y\\), where \\(\\hat{\\mathbf{a}}_y\\) denotes the GT anchor. It encourages the visual feature's CoS to be as close as possible to the GT anchor, while the A-CE loss, by putting \\(\\cos \\theta\\) into the softmax function, only encourages the visual feature's CoS with GT to be larger than those with other anchors. With A-CE supervision, the output feature could be opti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "mized to be far away from all anchors, but it can still be correctly classified as long as it is closest to the GT anchor. This will be discussed further in Sec. 4.4. Overall, the supervision objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.163, + 0.892, + 0.205 + ], + "angle": 0, + "content": "\\[\nL _ {1} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - \\log \\frac {\\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {y}\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {i}\\right)} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.892, + 0.278 + ], + "angle": 0, + "content": "where \\(\\delta\\) is adversarial perturbation generated by \\(\\frac{\\partial L_1}{\\partial\\mathbf{x}}\\). Unlike previous works [39] on standard training, we do not use a temperature parameter \\(\\tau\\) to scale the CoS. This is discussed in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.288, + 0.667, + 0.302 + ], + "angle": 0, + "content": "3.6. Smoothness Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Since we prefer the adversarial robustness of the model in the zero-shot transfer setting, we additionally introduce a smoothing loss, encouraging the CoS between adversarial features to be similar to the benign features of an image:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.382, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\nL _ {2} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - f _ {\\Theta} (\\mathbf {x}) ^ {T} f _ {\\Theta} (\\mathbf {x} + \\delta) \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.488 + ], + "angle": 0, + "content": "The smoothing loss is independent of the category labels of training examples. It can be expected to improve the adversarial robustness on novel categories. Finally, the whole loss in the training stage is given by \\( L = L_{1} + \\alpha L_{2} \\), where \\( \\alpha \\) is a hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.803, + 0.514 + ], + "angle": 0, + "content": "3.7. Performing Zero-Shot Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.657 + ], + "angle": 0, + "content": "When performing zero-shot recognition on novel categories, the image features of a test image are generated by the image encoder, and the anchors for novel categories are generated by the text encoder and the expansion algorithm. Note that the internal parameters of the expansion algorithm, such as \\(\\phi_0\\), are pre-computed using anchors from the training categories. Finally, the zero-shot prediction is the category whose text anchor exhibits the highest CoS with the image features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.634, + 0.688 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.691, + 0.712 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Baseline. We first consider the standard ZSL setting. Due to the lack of adversarially robust methods in this setting, we compared LAAT with several robust few-shot methods directly: AQ [19], R-MAML [46], and GR [13]. ITS-MAML [15] was not compared as it used a different attack setting and no code was released. Note that zero-shot is harder than few-shot as the latter can have access to a few examples from the novel categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We then considered a more pragmatic version of ZSL, which requires recognizing examples from both seen and novel categories. This is called the generalized zero-shot learning (GZSL) setting [8]. We trained several models with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24690" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.09, + 0.088, + 0.885, + 0.248 + ], + "angle": 0, + "content": "
DatasetMethodSettingCleanFGSMPGDCWAA
CIFAR-FSAQ [19]1-shot54.31 ± 0.5032.47 ± 0.4730.41 ± 0.4630.14 ± 0.4627.09 ± 2.07
R-MAML [46]1-shot38.53 ± 0.4726.76 ± 0.4426.21 ± 0.4329.76 ± 0.4616.67 ± 5.48
GR [13]1-shot45.27 ± 0.4939.60 ± 0.4638.03 ± 0.4637.00 ± 0.46-
LAATzero-shot55.60 ± 0.4641.17 ± 0.4440.12 ± 0.4437.35 ± 0.4436.45 ± 1.90
LAAT1-shot57.88 ± 0.4543.73 ± 0.4442.74 ± 0.4440.10 ± 0.4539.23 ± 2.12
MiniImageNetAQ [19]1-shot39.17 ± 0.4022.53 ± 0.3520.95 ± 0.3419.33 ± 0.3817.85 ± 1.41
R-MAML [46]1-shot36.18 ± 0.3623.85 ± 0.4621.74 ± 0.3419.68 ± 0.4013.69 ± 3.95
GR [13]1-shot36.14 ± 0.4529.23 ± 0.3327.57 ± 0.3826.61 ± 0.33-
LAATzero-shot46.97 ± 0.3530.69 ± 0.3029.27 ± 0.3127.91 ± 0.2925.24 ± 1.42
LAAT1-shot47.37 ± 0.3531.55 ± 0.3129.93 ± 0.3128.19 ± 0.2925.75 ± 1.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Table 2. The comparison of LAAT in 5-way zero-shot setting and previous works in 5-way 1-shot setting under \\(\\epsilon = 8 / 255\\). The \\(95\\%\\) confidence interval was reported. Here all methods used Conv4-512 as the image model. The accuracy of GR under AA is not shown as it was not reported and no code was released. We also list the performance of LAAT in the 1-shot setting with image-text blended anchors (discussed in Sec. 4.3). The results higher than previous methods in each column are in bold." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.324, + 0.885, + 0.429 + ], + "angle": 0, + "content": "
MethodModelTrainingCifar100 100aPY 32AwA2 50COCO Obj. 80STL10 10OxfordPet 37DTD 47Caltech101 101Caltech256 257SUN 397
TeCoA [34]ViT-B/32*ε = 1/2550.004.390.881.070.980.001.431.661.330.00
ViT-B/32ε = 8/2550.005.271.451.331.460.281.892.972.360.14
ViT-B/16ε = 8/2550.206.451.951.561.660.492.257.033.810.29
LAATViT-B/32ε = 8/2556.6424.3223.3418.5641.1220.314.5934.4321.394.30
ViT-B/16ε = 8/2557.3234.9726.2120.6149.4128.814.7937.8924.025.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.436, + 0.897, + 0.482 + ], + "angle": 0, + "content": "Table 3. Adversarial accuracy (\\%) of models trained on ImageNet-1K on ten downstream datasets in GZSL setting. The number of categories is shown below the dataset name. * indicates the results of the originally released checkpoint of TeCoA with the training perturbation \\(\\epsilon = 1/255\\). The robustness was all evaluated by PGD-20 under \\(\\epsilon = 8/255\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.583 + ], + "angle": 0, + "content": "LAAT on ImageNet-1K and evaluated them directly on ten downstream datasets. In this setting, we mainly compare LAAT with TeCoA [34]. We did not compare with Yucel et al. [51] and Zhang et al. [55] as they relied on image attributes and were unsuitable for this setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Datasets. When comparing with few-shot methods, we used two popular few-shot benchmarks CIFAR-FS [5] and MiniImageNet [43]. CIFAR-FS, a variant of Cifar100 [25], contains 64, 16, and 20 categories for training, validation, and testing, respectively. MiniImageNet, a subset of ImageNet-1k [40], contains 600 images of \\(84 \\times 84\\) size per category. We used the same dataset splitting setting as previous works [13, 19, 46], for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.471, + 0.825 + ], + "angle": 0, + "content": "In the GZSL setting, we tested models trained on ImageNet-1K on ten downstream datasets such as AwA2 [48] and aPY [17]. Note that all of these datasets include several novel categories that have never appeared in ImageNet-1K, e.g., statue and ass in aPY [17], blue whale and dolphin in AwA2 [48]. More introduction to the datasets is shown in the Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.471, + 0.885 + ], + "angle": 0, + "content": "Implementation details. Unless otherwise stated, we used the CLIP ViT-B/16 text encoder and the prompt text was set as \"This is a photo of a {}\". \\(\\alpha\\) was set to 3." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In the standard ZSL setting, we followed previous few-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.596 + ], + "angle": 0, + "content": "shot works [13, 46] and used the widely applied model Conv4-512 [54] as the image model. The training recipe strictly followed GR [13]. We performed AT by generating adversarial examples via PGD method with maximum adversarial perturbation \\(\\epsilon = 8 / 255\\) in \\(l_{\\infty}\\)-norm bound, with \\(T = 7\\) iterative steps (step size \\(s = 2 / 255\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.689 + ], + "angle": 0, + "content": "In the GZSL setting, we trained two image models ViT-B/32 and ViT-B/16 [14] with LAAT on the large ImageNet-1K. The training basically followed the recipe proposed by Debenedetti et al. [12]. Here we performed AT by generating adversarial examples via PGD method with \\(\\epsilon = 8 / 255\\) in \\(l_{\\infty}\\)-norm bound and \\(T = 2\\) iterative steps (\\(s = 8 / 255\\))." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.693, + 0.712 + ], + "angle": 0, + "content": "4.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The Standard ZSL Setting. We performed experiments in the standard and usual 5-way one/zero-shot setting [13, 19, 46]. Following previous works [19, 46], we report the robustness of our method against common adversarial attacks, including FGSM [20], PGD [32] with 20 iterative steps and step size \\( s = 2 / 255 \\) (denoted as PGD-20 in the following text), CW [7] (optimized by PGD for 30 steps with step size \\( s = 0.8 / 255 \\)), and AutoAttack (AA) [11], an effective ensemble attack (several variants of PGD and one query-based black-box Square attack [2]) used to assess adversarial robustness. The maximal perturbation was set to commonly used \\( \\epsilon = 8 / 255 \\) in \\( l_{\\infty} \\) setting. The experi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "24691" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.268, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.203, + 0.232, + 0.216 + ], + "angle": 0, + "content": "(a) Benign accuracy." + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.093, + 0.462, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.203, + 0.428, + 0.216 + ], + "angle": 0, + "content": "(b) Robust accuracy." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.471, + 0.284 + ], + "angle": 0, + "content": "Figure 5. Classification accuracy on both benign and adversarial examples in 5-way few-shot setting on CIFAR-FS, with image feature anchors or with image-text blended anchors. The dashed line denotes 5-way zero-shot accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.47, + 0.508 + ], + "angle": 0, + "content": "mental results are reported from 2000 randomly sampled 5-way tasks. The results presented in Tab. 2 demonstrate that LAAT in the zero-shot setting achieved competitive performance with several methods in the few-shot setting. Notably, the adversarial robustness of LAAT models even surpassed that of the previous SOTA adversarily robust few-shot method GR [13]. Additionally, our method exhibited significantly better clean accuracy than all three few-shot methods. We note that these results should be interpreted cautiously because LAAT and these few-shot works used different auxiliary information: the former used text information of the test category names and the latter relied on a small number of example images from the test categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.469, + 0.745 + ], + "angle": 0, + "content": "GZSL Setting across Datasets. In the GZSL setting, the models were evaluated in \\(N\\)-way classification setting, where \\(N\\) is the number of categories in each dataset. All images on downstream datasets are resized to \\(224 \\times 224\\) in inference. Here we compared LAAT with TeCoA. TeCoA originally performed adversarial fine-tuning on ImageNet-1K with maximum perturbation \\(\\epsilon = 1/255\\). For a fair comparison, here adversarial fine-tuning was also performed with the released code of TeCoA [34]. Tab. 3 shows the zero-shot results on ten downstream datasets under PGD-20. Here the accuracy is reported by sampling 100 examples per category. We can see that TeCoA almost failed to train adversarily robust models in the common \\(\\epsilon\\) setting, while the models trained with LAAT can obtain substantial adversarial robustness on several downstream datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Note that the models trained with LAAT did not perform quite well on Cifar100, DTD [9], and SUN [49]. This could be caused by the original small resolution of Cifar100 \\((32 \\times 32)\\) [12] and the large inter-class differences between DTD, SUN, and ImageNet-1K." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.403, + 0.849 + ], + "angle": 0, + "content": "4.3. Extending LAAT to Few-Shot Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our zero-shot models supervised by LAAT can be naturally extended to the few-shot setting and the robustness can be further improved. Given the images of novel categories as" + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.089, + 0.795, + 0.148 + ], + "angle": 0, + "content": "
ObjectiveCleanRobust
Eucl39.1728.04
θ37.9227.84
cosθ41.9127.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.158, + 0.892, + 0.187 + ], + "angle": 0, + "content": "Table 4. Classification accuracy (%) on CIFAR-FS supervised by the original CLIP anchors with different optimization objectives." + }, + { + "type": "table", + "bbox": [ + 0.567, + 0.2, + 0.828, + 0.273 + ], + "angle": 0, + "content": "
EXPA-CESMCleanRobust
41.9127.91
✓46.9832.71
✓✓54.0237.69
✓✓✓55.6040.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.326 + ], + "angle": 0, + "content": "Table 5. Classification accuracy \\((\\%)\\) of Conv4-512 on CIFAR-FS in different ablation experiments. The first two lines were the results obtained with \\(\\cos \\theta\\) as the objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.536 + ], + "angle": 0, + "content": "the support set, we can follow the previous few-shot SOTA method [13] and use the prototype-based metric learning [41] to build image feature-based anchors. That is, in the \\(K\\)-shot case, the image anchor \\(\\mathbf{v}_y\\) for a novel category \\(y\\) can be computed as \\(\\mathbf{v}_y = \\mathrm{Norm}_2\\{\\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}\\), where \\(\\mathbf{x}_i\\) denotes the images of \\(y\\) in the support set and \\(\\mathrm{Norm}_2\\{\\cdot\\}\\) denotes the \\(l_2\\) normalization. We can also weight the original text anchor \\(\\mathbf{a}_y\\) from the text encoder and the image feature from the support set to build an image-text blended anchor for \\(y\\): \\(\\hat{\\mathbf{a}}_y^* = \\mathrm{Norm}_2\\{\\beta \\cdot \\hat{\\mathbf{a}}_y + \\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}\\). Here the text anchor can be regarded as the prior knowledge of \\(y\\). We set \\(\\beta = 2\\) in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Tab. 2 shows the results of different models with image-text blended anchors in the 1-shot setting. The 1-shot performance of LAAT was better than the zero-shot performance. Fig. 5 further shows the results with two different anchors \\(\\hat{\\mathbf{a}}^{\\star}\\) and \\(\\mathbf{v}\\) in \\(K\\)-shot settings. We can see that the text anchor used to perform zero-shot recognition boosted the performance in various \\(K\\)-shot settings, especially when the support set is small." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.668, + 0.655, + 0.685 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.752 + ], + "angle": 0, + "content": "Limited by computing resources, the ablation study is not studied on ImageNet-1K. Unless specified, it was performed with Conv4-512 in the 5-way zero-shot setting. The robustness was evaluated by the PGD-20 attack method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of each design. In Fig. 3, we have shown that if the model is supervised by CoS between the original CLIP anchors and the output feature, say, \\(\\cos \\theta\\), it is difficult to converge under AT. In Tab. 4, we show the results of other anchor-based AT optimization objectives, such as the angle [37] and the Euclidean distance [36], denoted as \\(\\theta\\) and Eucl respectively, in the ZSL setting. The results indicate that none of these three optimization objectives led to good classification accuracy. Note that random guessing in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24692" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.46, + 0.176 + ], + "angle": 0, + "content": "
AT MethodHyper-ParameterCleanRobust
EXP + A-CE + TRADES1/λ = 155.2619.22
1/λ = 655.8225.02
EXP + A-CE + SMα = 1.055.6239.95
α = 3.055.6040.12
α = 6.053.0140.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.186, + 0.47, + 0.214 + ], + "angle": 0, + "content": "Table 6. Classification accuracy (%) on CIFAR-FS with different AT methods." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.228, + 0.45, + 0.315 + ], + "angle": 0, + "content": "
MethodSettingConv4-64ResNet12
CleanRobustCleanRobust
AQ1-shot43.7427.5443.5831.45
R-MAML1-shot32.4624.6142.5432.54
GR1-shot44.5137.4548.1339.29
LAATzero-shot49.8438.7458.3742.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.325, + 0.47, + 0.355 + ], + "angle": 0, + "content": "Table 7. Classification accuracy (%) on CIFAR-FS with different image models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.324, + 0.395 + ], + "angle": 0, + "content": "this setting can obtain \\(20\\%\\) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.469, + 0.485 + ], + "angle": 0, + "content": "We then conducted an ablation study to show the effectiveness of each design in LAAT: the expansion algorithm, the A-CE loss, and the smoothness loss, denoted as EXP, A-CE, and SM, respectively. The results in Tab. 5 show that all of these designs are helpful in improving the zero-shot adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.651 + ], + "angle": 0, + "content": "As stated in Sec. 3.5, A-CE could make a relaxation on the original anchor-based AT objectives. To validate this, we computed the average CoS between GT anchors and the output visual features of the two models trained: EXP and \\(\\mathrm{EXP} + \\mathrm{A - CE}\\) (EXP was supervised by \\(\\cos \\theta\\); see Tab. 5). With A-CE loss, the visual feature's CoS with GT anchors dropped from 0.455 to 0.205 on average, while the accuracies on both benign and adversarial examples were improved (see Tab. 5). These results support our conjecture that some examples not quite close to the GT anchor can still be classified correctly with A-CE supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.814 + ], + "angle": 0, + "content": "Smoothness v.s. TRADES. It is seen that SM has similar objectives with TRADES [52], but they are different in many aspects (see Appendix F). Here we also compare SM with TRADES empirically. The experiments were performed based on expanded anchors with A-CE supervision. The results are shown in Tab. 6, together with the most important hyper-parameter of each method. We can clearly see that TRADES performed much worse than our smoothness loss. In addition, the results show that the effectiveness of smoothing loss was not quite sensitive to \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Different image models. We also performed experiments with other popular image models in the few-shot setting other than Conv4-512, e.g., Conv4-64 [44] and ResNet12 [19]. Tab. 7 shows the results of the two image models on CIFAR-FS. Conv4-64 is smaller than Conv4-512 while" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.882, + 0.189 + ], + "angle": 0, + "content": "
Text EncoderEmbed-dimCleanRobust
None (MMC, one-shot)51247.5532.79
RN50x464052.2838.65
RN50x1676853.8638.02
ViT-B/3251252.6838.39
ViT-B/1651255.6040.12
ViT-L/1476852.7839.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.2, + 0.892, + 0.255 + ], + "angle": 0, + "content": "Table 8. Classification accuracy (\\%) supervised by MMC anchors in one-shot setting and the full LAAT with different CLIP text encoders in zero-shot setting on CIFAR-FS. Embed-dim denotes the dimension of the text anchors." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.282, + 0.892, + 0.328 + ], + "angle": 0, + "content": "ResNet12 is slightly larger. The models trained with LAAT showed stronger zero-shot adversarial robustness than those few-shot methods regardless of the model size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.58 + ], + "angle": 0, + "content": "Different text encoders. Without text encoders, it is impossible for LAAT to achieve zero-shot robustness. In principle, arbitrary text encoders are feasible for LAAT if they have semantic consistency. We show the influence of using different CLIP text encoders in Tab. 8. We also show the results without using any text encoders. Instead, the image model was supervised by the MMC anchors (here zero-shot recognition cannot be performed while one-shot results are reported). We can see that using any text encoders can obtain strong zero-shot robustness, significantly surpassing the few-shot robustness without using text encoders. We also studied the relationship between the distances of different CLIP models' anchors and Cifar100 super-categories. The results in Appendix G further show that semantic consistency of the text encoder is one of the keys to zero-shot adversarial robustness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.749, + 0.611 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.816 + ], + "angle": 0, + "content": "In this work, we propose a novel LAAT strategy for adversarially robust image classification in a zero-shot setting, inspired by recent vision-language models. Extensive experiments demonstrate that our method has strong adversarial robustness across different models and several zero-shot settings. We hope our work could encourage more researchers to investigate the robustness in the zero-shot setting, which could be a promising way to improve the scalability and flexibility of adversarially robust models. Limited by the data and computing resources, we cannot perform AT on large-scale image-text pairs directly like CLIP [39]. We believe it could be another promising way to achieve zero-shot adversarial robustness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.661, + 0.847 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (Nos. U2341228 and U19B2034)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Antonio Aguilera and Ricardo Pérez-Aguila. General n-dimensional rotations. In WSCG, pages 1-8, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.2 + ], + "angle": 0, + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: A query-efficient black-box adversarial attack via random search. In Eur. Conf. Comput. Vis. (ECCV), pages 484-501, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[3] Anish Athalye et al. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Int. Conf. Mach. Learn. (ICML), pages 274-283, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.469, + 0.285 + ], + "angle": 0, + "content": "[4] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[5] Luca Bertinetto, João F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Int. Conf. Learn. Represent. (ICLR), 2019, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[6] LE Blumenson. A derivation of n-dimensional spherical coordinates. The American Mathematical Monthly, 67(1):63-66, 1960. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[7] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy, SP, pages 39-57, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[8] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In Eur. Conf. Comput. Vis. (ECCV), pages 52-68, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.469, + 0.543 + ], + "angle": 0, + "content": "[9] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3606-3613, 2014. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[10] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, pages 215-223, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Int. Conf. Mach. Learn. (ICML), pages 2206-2216, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[12] Edoardo Debenedetti et al. A light recipe to train robust vision transformers. arXiv preprint arXiv:2209.07399, 2022.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[13] Junhao Dong, Yuan Wang, Jianhuang Lai, and Xiaohua Xie. Improving adversarially robust few-shot image classification with generalizable representations. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9015-9024. IEEE, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent. (ICLR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[15] Xiaoyue Duan, Guoliang Kang, Runqi Wang, Shumin Han, Song Xue, Tian Wang, and Baochang Zhang. Rethinking the number of shots in robust model-agnostic meta-learning. arXiv preprint arXiv:2211.15180, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1625-1634, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David A. Forsyth. Describing objects by their attributes. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1778-1785, 2009. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Micah Goldblum et al. Adversarily robust few-shot learning: A meta-learning approach. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[20] Ian J. Goodfellow et al. Explaining and harnessing adversarial examples. In Int. Conf. Learn. Represent. (ICLR), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[21] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zero-shot detection via vision and language knowledge distillation. In Int. Conf. Learn. Represent. (ICLR), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Int. Conf. Mach. Learn. (ICML), pages 4904-4916, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. California Institute of Technology, 2009. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[26] Huafeng Kuang, Hong Liu, Yongjian Wu, and Rongrong Ji. Semantically consistent visual representation for adversarial robustness. IEEE Transactions on Information Forensics and Security, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[27] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 36(3):453-465, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[28] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In Int. Conf. Learn. Represent. (ICLR), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[29] Xiao Li, Hang Chen, and Xiaolin Hu. On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[30] Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, and Xiaolin Hu. Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24694" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Int. Conf. Learn. Represent. (ICLR), 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[33] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. Adv. Neural Inform. Process. Syst. (NeurIPS), 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[34] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. Int. Conf. Learn. Represent. (ICLR), 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.333 + ], + "angle": 0, + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.469, + 0.389 + ], + "angle": 0, + "content": "[36] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In Int. Conf. Learn. Represent. (ICLR), 2020. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.392, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[37] Tianyu Pang, Xiao Yang, Yinpeng Dong, Taufik Xu, Jun Zhu, and Hang Su. Boosting adversarial training with hypersphere embedding. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.49 + ], + "angle": 0, + "content": "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3498-3505, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. (IJCV), 115(3):211-252, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[41] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4077-4087, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In Int. Conf. Learn. Represent. (ICLR), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[43] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[44] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[45] Mengmeng Wang et al. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[46] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[47] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 69-77, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[48] Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - A comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 41(9):2251-2265, 2019. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[49] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3485-3492, 2010. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.387 + ], + "angle": 0, + "content": "[50] Yunlong Yu, Zhong Ji, Yanwei Fu, Jichang Guo, Yanwei Pang, and Zhongfei (Mark) Zhang. Stacked semantics-guided attention model for fine-grained zero-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6007, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[51] Mehmet Kerim Yucel et al. How robust are discriminatively trained zero-shot learning models? Image Vis. Comput., 119: 104392, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.5 + ], + "angle": 0, + "content": "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In Int. Conf. Mach. Learn. (ICML), pages 7472-7482, 2019. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.503, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[53] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM Int. Conf. Multimedia, pages 5005-5013, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.546, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[54] Manli Zhang, Jianhong Zhang, Zhiwu Lu, Tao Xiang, Mingyu Ding, and Songfang Huang. IEPT: instance-level and episode-level pretext tasks for few-shot learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[55] Xingxing Zhang, Shupeng Gui, Zhenfeng Zhu, Yao Zhao, and Ji Liu. ATZSL: defensive zero-shot recognition in the presence of adversaries. arXiv preprint arXiv:1910.10994, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.7 + ], + "angle": 0, + "content": "[56] Mo Zhou and Vishal M Patel. Enhancing adversarial robustness for deep metric learning. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15325-15334, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[57] Xiaopei Zhu, Xiao Li, Jianmin Li, Zheyao Wang, and Xiaolin Hu. Fooling thermal infrared pedestrian detectors in real world using small bulbs. In AAAI, pages 3616-3624, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24695" + } + ] +] \ No newline at end of file diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_origin.pdf b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82fc9e8d308f78295bed34316a511b22d99d9cf2 --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/5bd9567c-d374-4782-9305-9ad2d0f7be37_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11abf24858709afbde6637b3431c033dd0a3056196430308337c07d0ef984098 +size 909340 diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/full.md b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ce3122b1f68c07c58c8111f4815671b777a957a4 --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/full.md @@ -0,0 +1,320 @@ +# Language-Driven Anchors for Zero-Shot Adversarial Robustness + +Xiao Li $^{1}$ Wei Zhang $^{1}$ Yining Liu $^{2}$ Zhanhao Hu $^{1}$ Bo Zhang $^{1}$ Xiaolin Hu $^{1,3*}$ \ + $^{1}$ Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China\ + $^{2}$ Harbin Institute of Technology, Weihai, China\ + $^{3}$ Chinese Institute for Brain Research (CIBR), Beijing, China\ +{lixiao20, zhang-w19, huzhanha17}@mail.tsinghua.edu.cn\ +22s030184@stu.hit.edu.cn\ +{dcszb, xlhu}@mail.tsinghua.edu.cn + +# Abstract + +Deep Neural Networks (DNNs) are known to be susceptible to adversarial attacks. Previous researches mainly focus on improving adversarial robustness in the fully supervised setting, leaving the challenging domain of zero-shot adversarial robustness an open question. In this work, we investigate this domain by leveraging the recent advances in large vision-language models, such as CLIP, to introduce zero-shot adversarial robustness to DNNs. We propose LAAT, a Language-driven, Anchor-based Adversarial Training strategy. LAAT utilizes the features of a text encoder for each category as fixed anchors (normalized feature embeddings) for each category, which are then employed for adversarial training. By leveraging the semantic consistency of the text encoders, LAAT aims to enhance the adversarial robustness of the image model on novel categories. However, naively using text encoders leads to poor results. Through analysis, we identified the issue to be the high cosine similarity between text encoders. We then design an expansion algorithm and an alignment cross-entropy loss to alleviate the problem. Our experimental results demonstrated that LAAT significantly improves zero-shot adversarial robustness over state-of-the-art methods. LAAT has the potential to enhance adversarial robustness by large-scale multimodal models, especially when labeled data is unavailable during training. Code is available at https://github.com/LixiaoTHU/LAAT. + +# 1. Introduction + +Adversarial attacks [42], by adding deliberately designed perturbations to inputs, have posed a significant threat to + +![](images/e6c7a4984faab35b63942b9e8134d0ddc50121a2c225caf308d2aa28606f0ead.jpg) +Figure 1. The illustration of zero-shot ability with LAAT. Different colors of the marks indicate different categories. When a model is adversarially trained with the text anchors of table and lion (seen categories), it can recognize adversarial examples of the two categories (grey and green). Then due to the text anchors of chair and tiger (novel categories) being close to those of table and lion, respectively, the model can also recognize the two novel categories. + +the application of Deep Neural Networks (DNNs) [16, 57]. In response to the threats, various techniques have been proposed to enhance adversarial robustness of DNNs. Among these, Adversarial Training (AT) and its variants [32, 36, 37, 52] have shown to be highly effective [16, 42, 57]. However, applying AT from scratch to novel datasets or categories can be time-consuming. In real-world scenarios, it is common for models to encounter examples from novel categories, known as zero-shot learning (ZSL) setting [27]. If adversarially trained models possess zero-shot capability, it can significantly enhance the scalability and practical utility of adversarially robust models. + +Adversarial robustness in the ZSL setting is challenging. Earlier attempts [51, 55] try to combine attribute-based ZSL methods with AT. However, these methods yielded modest improvements and lacked scalability due to the practical difficulties in obtaining attributes. Recently, large-scale + +vision-language models such as CLIP [39] have made significant strides and are considered foundational models for various downstream applications. These models, trained on extensive text-image data, demonstrate strong zero-shot capability by using textual descriptions to recognize new visual categories, eliminating the need for explicit reliance on image attributes. In this work, different from previous attribute-based ZSL methods, we explore the utilization of foundation models to introduce zero-shot adversarial robustness to DNNs. + +We propose a Language-driven, Anchor-based Adversarial Training strategy, denoted as LAAT, to enhance the zero-shot adversarial robustness. Specifically, we use the popular text encoder from vision-language models, specifically CLIP [39], to obtain $l_{2}$ normalized feature embedding (named anchor in this paper) for the text (label name) corresponding to image categories. The text encoder exhibits a property called semantic consistency, where semantically similar categories are mapped to neighboring anchors in the feature space. Then we utilize the text anchors to supervise the image classification model with AT. After this process, the image model not only obtains adversarial robustness on seen categories but also aligns the image features with the text anchors. During zero-shot inference, by leveraging the semantic consistency between seen and novel categories, the image model can recognize novel categories while maintaining adversarial robustness. Fig. 1 illustrates the idea. We note that one recent work, TeCoA [34], also utilizes the ability of large-scale vision-language models to enhance the zero-shot adversarial robustness. However, TeCoA was only shown effective for extremely tiny adversarial noise (i.e., $\epsilon = 1 / 255$ ). + +As CLIP models are not designed for adversarial robustness, directly applying the obtained anchors from CLIP text encoders to anchor-based AT [36, 37] (refer to Sec. 2.1) results in poor performance. We attribute it to the high Cosine Similarity (CoS) problem, i.e., the average cosine similarity between CLIP anchors are too high to be directly used in AT. See Sec. 3.3 for details. This problem has never been discovered in previous work utilizing CLIP models [22, 28, 34, 45]. To address it, we propose several techniques. We first introduce an expansion algorithm that maps the original CLIP anchors to new anchors with increased distances while maintaining semantic consistency. We then incorporate an Alignment Cross-Entropy (A-CE) loss as the optimization objective. In addition, we use a smoothness loss to enhance the robustness on unseen categories. + +The experimental results showed that the models trained with LAAT achieved remarkable zero-shot performance against strong adversarial perturbations, even surpassing the previous state-of-the-art (SOTA) adversarily robust few-shot methods. We also trained several models with LAAT on a large dataset ImageNet-1K [40]. These mod + +els showed substantial adversarial robustness across various downstream datasets in the zero-shot setting, significantly outperforming the recent TeCoA method [34]. The main contributions can be summarized as follows: + +- We identify the high CoS problem in CLIP anchors for adversarial robustness and propose effective techniques to alleviate it. +- The models trained by LAAT show substantial adversarial robustness across downstream datasets, suggesting that AT in the ZSL setting could be a promising way to improve the practical usefulness of AT. + +# 2. Related Work + +# 2.1. Anchor-based AT + +AT has become one of the most effective strategies to improve adversarial robustness [3]. Several studies try to improve AT from various aspects [26, 29, 30, 36, 37, 52]. One series of works focuses on refining cross-entropy loss, which is the most commonly used loss in classification, to obtain better robustness with AT [36, 37]. Pang et al. [36] boosts AT by introducing the Max-Mahalanobis center (MMC) loss, a Euclidean loss where there is a fixed feature vector $\mu_y$ for each category $y$ , and the optimization object is $||\mathbf{z} - \mu_y||_2^2$ , where $\mathbf{z}$ is the output feature. Pang et al. [37] uses Hypersphere Embedding to boost AT. It normalizes both weights $\mathbf{W}$ and features $\mathbf{z}$ of the output layer, and proposes two losses in the form of $\cos \theta$ and $\theta$ , where $\theta$ is the angle between $\frac{\mathbf{W}}{||\mathbf{W}||_2}$ and $\frac{\mathbf{z}}{||\mathbf{z}||_2}$ . We call these AT methods anchor-based methods as they all used $l_{2}$ normalized feature embeddings (anchors) for training. Besides, several metric learning works [33, 56] also used a similar paradigm to anchor-based AT for enhancing robustness. + +# 2.2. Language-driven Recognition. + +Earlier ZSL models exploit auxiliary information, commonly in the form of image attributes, to transfer knowledge between seen and novel categories [47, 50]. Recently, language-driven recognition has caught attention in the ZSL field, especially with the recent advance of large-scale pretrained vision-language models [24, 39]. CLIP [39] demonstrated that classic recognition tasks can strongly benefit from millions of raw texts pertaining to images. CLIP uses an independent image encoder and text encoder to perform contrastive learning on 400M image-text pairs. After pretraining, natural language is used to reference visual concepts by computing the CoS between text features and the visual feature, enabling the zero-shot transfer of the model to downstream classification datasets. The basic paradigm of CLIP can be extended to several tasks such as semantic segmentation and object detection [22, 28, 45]. However, recent works show that CLIP models lack adversarial robustness and can be easily attacked [4, 34, 53]. + +![](images/b685b0efffcee3d0afa86d067db947d125bf0445ae02e8ba3b1933f71f306736.jpg) +Figure 2. The pipeline of LAAT. Only the image encoder is trainable in the figure. $\odot$ indicates computing the CoS. Red arrows indicate the adversarial example generation process and brown arrows indicate the inference. + +# 2.3. Adversarily Robust Zero/few-shot Learning + +An early preprint work [55] combines AT with attribute-based ZSL methods but only gets a mild improvement over non-AT models. In addition, the method is hard to generalize without the attribute annotations. Yucel et al. [51] evaluates the adversarial robustness of classic attribute-based zero-shot models and concludes that adversarial robustness for ZSL field is suffering an immature state. Mao et al. [34] proposes several techniques such as visual prompting and fine-tuning to adapt large-scale models, e.g., the image models of CLIP, for zero-shot adversarial robustness, while their method called TeCoA was only shown to be effective under extremely tiny adversarial noise. Different from TeCoA, LAAT only uses the anchors from the CLIP text encoders. + +A closely related area is adversarially robust few-shot learning, which has caught attention recently, too. We also compare LAAT with recent few-shot methods directly to comprehensively evaluate it. AQ [19], R-MAML [46], and ITS-MAML [15] study meta-learning jointly with AT. GR [13] tries to learn a generalizable robust representation by the combination of several AT techniques. + +# 3. LAAT + +# 3.1. Overall Pipeline + +LAAT first obtains fixed text anchors from the text encoder, then uses them to adversarially supervise the training of image classification model. After AT, the image model will obtain adversarial robustness on seen categories and also align the image features with the text anchors to obtain zero-shot ability (see Fig. 1). The overall pipeline of LAAT is + +illustrated in Fig. 2. Given the original text anchors, LAAT uses an expansion algorithm to obtain fixed Ground-Truth (GT) anchors suitable for AT, then uses an image encoder to extract both benign and adversarial features. The CoS between adversarial features and GT anchors are maximized by minimizing the A-CE loss $(L_{1})$ . Besides, a smoothness loss $(L_{2})$ is applied to the adversarial and benign features. We then describe each module of LAAT. + +# 3.2. Text Anchor Extraction + +Given the names of $N$ categories in the training set, they are first filled into $N$ sentences with a fixed prompt text, such as "A photo of {}". After that, these sentences are encoded into $N$ anchors $\{\mathbf{a}_i\}_{i=1}^N$ by a fixed pre-trained text encoder, where $\mathbf{a}_i \in \mathbb{R}^n$ and $||\mathbf{a}_i||_2 = 1$ . Many architectures are feasible if they have semantic consistency as mentioned before and in this work, we use the large-scale pre-trained CLIP text encoder. However, CLIP anchors cannot be used directly due to the high CoS problem, as analyzed below. + +# 3.3. High Cosine Similarity Problem + +We first make an intriguing observation that directly using CLIP anchors for anchor-based AT poses significant challenges in achieving convergence. This phenomenon is universal among different CLIP text encoders and different models. To illustrate this, we adversarially trained a ResNet18 [23] on Cifar100 [25] by maximizing the $\cos \theta$ objective [37], i.e., the CoS between the model output $\mathbf{z}$ and the GT anchor $\mathbf{a}_y$ [37], which is computed as $\frac{\mathbf{z}^T\mathbf{a}_y}{||\mathbf{z}||_2}$ (note that $||\mathbf{a}_y||_2 = 1$ ). The experimental details can be found in Appendix A. Fig. 3 shows the learning curves with CLIP anchors (the orange curves). For comparison, here we also + +![](images/5acee1955a3b9094233ca2dd05ae15db8dd9bad084366207c6d3cbab3f5cf0e9.jpg) +(a) Clean accuracy. + +![](images/306e86cb3b75467e3f676cd1e9059bcb21776e635b445500961d6f90110cda9f.jpg) +(b) Robust accuracy. +Figure 3. Learning curves of AT supervised by $\cos \theta$ with fixed anchors generated from MMC method, CLIP text encoder, and the expansion algorithm (see Sec. 3.4). + +show the learning curves with MMC [36] anchors (the blue curves) in the same training setting. MMC anchors are precomputed by maximizing the distance between anchors [36] and thus the average CoS between MMC anchors is close to 0 (e.g., -0.01 on Cifar100). The accuracy with CLIP anchors increased quite slowly. Even after 200 epoch training, the accuracy was far lower than that of MMC anchors (note that zero-shot recognition cannot be achieved with MMC anchors for the lack of semantic consistency). + +The stark discrepancy between the two learning curves motivates us to analyze the essential difference between the two sets of anchors. Inspired by the anchor generation process of MMC, we calculated the CoS between CLIP anchors and found them to be considerably larger than that of MMC anchors (-0.01). The first row of Tab. 1 presents the average CoS between anchors from Cifar100 categories generated by different CLIP models. Note that the high CoS between CLIP anchors is irrelevant to the prompt text. We found that even using several random sentences from Wikipedia to encode anchors by CLIP text encoders, the average CoS remained significantly large. + +Although high CoS seems to have no effect on downstream tasks on clean images [22, 28, 45], we discovered that AT was remarkably affected because higher CoS implies that the anchors are closer to the decision boundary, which could leave more room for adversarial examples [36, 52]. This problem may also explain why TeCoA [34], which also used the text anchors during adaptation, failed to convert CLIP image models to adversially robust models against strong perturbations. Importantly, random or MMC anchors cannot mitigate this issue as they lack the semantic consistency provided by CLIP anchors and thus cannot be used in the ZSL setting. Therefore, the micro designs in our LAAT approach primarily aim to address the challenges associated with CLIP anchors. + +# 3.4. Expansion Algorithm + +We first propose an expansion algorithm to increase the distances between anchors (reducing the CoS). The expansion algorithm can remap the anchors to increase the dis + +
TypeRN50x4RN50x16ViT-B/32ViT-B/16ViT-L/14
Original0.7000.7100.7790.7610.746
Expanded0.2380.2530.1990.1950.222
+ +Table 1. Average CoS between anchors before and after expansion algorithm (see Sec. 3.4) from Cifar100 categories generated by different CLIP text encoders. + +tances between anchors while preserving the semantic consistency between them. High CoS means the normalized anchors $\{\mathbf{a}_i\}_{i=1}^N$ are dispersed over a cluster of the unit hypersphere (see Fig. 4). As we expect the remapped anchors to be also on the unit hyper-sphere, the expansion algorithm should be designed under spherical coordinates. Let us recap the transformation between $n$ -dimensional spherical coordinate system and Cartesian coordinates. Suppose $(x^{(1)}, x^{(2)}, \ldots, x^{(n)})$ is the Cartesian coordinate, we can compute them by: + +$$ +x ^ {(1)} = r \cos \phi^ {(1)}; \quad x ^ {(2)} = r \sin \phi^ {(1)} \cos \phi^ {(2)}; \quad \dots ; +$$ + +$$ +x ^ {(n - 1)} = r \sin \phi^ {(1)} \dots \sin \phi^ {(n - 2)} \cos \phi^ {(n - 1)}; \tag {1} +$$ + +$$ +x ^ {(n)} = r \sin \phi^ {(1)} \dots \sin \phi^ {(n - 2)} \sin \phi^ {(n - 1)}, +$$ + +where $r$ is the radial coordinate and $\phi^{(1)},\dots ,\phi^{(n - 1)}$ are angular coordinates [6]. Here $\phi^{(1)}$ is the polar angle (taking the three-dimensional spherical coordinate system like longitude and latitude system as an example, the polar angle is the angle with respect to the $z$ axis). If the anchors are clustered around the polar, i.e., $\phi_0\coloneqq \max_i\phi_i^{(1)} < \frac{\pi}{2}$ , where $\phi_i^{(1)}$ denotes the polar angle of anchor $\mathbf{a}_i$ , a natural expansion method is to enlarge the polar angle $\phi_i^{(1)}$ to $\frac{\pi}{2}\cdot \phi_i^{(1)} / \phi_0$ and keep the other angular coordinates unchanged. Then the cluster can be expanded to the whole hemisphere. Note that we cannot expand them larger than $\frac{\pi}{2}$ as anchors may be close to each other from another hemisphere, which may hurt the semantic consistency of anchors. + +Thus, given the original anchor $\mathbf{a}_i$ , the first step of the expansion is to calculate a rotation matrix $\mathbf{R}$ to transform the center $\mathbf{v}$ of the clustered $\mathbf{a}_i$ 's ( $\mathbf{v} := \sum_{i=1}^{N} \mathbf{a}_i / || \sum_{i=1}^{N} \mathbf{a}_i||_2$ ) to the polar $\mathbf{p} = [1,0,\dots,0]^T$ , and the rotated anchor $\tilde{\mathbf{a}}_i$ is computed by: $\tilde{\mathbf{a}}_i = \mathbf{Ra}_i$ . Note that computing $\mathbf{R}$ between two unit vectors has several ways and we use the Aguilera and Pérez-Aguila [1] algorithm. As $\tilde{\mathbf{a}}_i$ 's are clustered around the polar $\mathbf{p}$ , we then enlarge the polar angle of $\tilde{\mathbf{a}}_i$ as described above. Denoting the $j$ -th element of $\mathbf{a}_i$ as $a_i^{(j)}$ , the largest polar angle between $\tilde{\mathbf{a}}_i$ 's and $\mathbf{p}$ can be computed as $\phi_0 = \max_{1 \leq i \leq N} \{\arccos \tilde{a}_i^{(1)}\}$ . According to Eq. (1), the Cartesian coordinates of expanded anchor $\bar{\mathbf{a}}_i$ is given by: + +$$ +\bar {a} _ {i} ^ {(1)} = \cos \bar {\phi} _ {i} ^ {(1)}; \tag {2} +$$ + +$$ +\bar {a} _ {i} ^ {(j)} = \tilde {a} _ {i} ^ {(j)} \cdot \sin \tilde {\phi} _ {i} ^ {(1)} / \sin \tilde {\phi} _ {i} ^ {(1)}, \quad j = 2, \dots , n, +$$ + +![](images/20e83b15f2d4fc4f986b6d231a0ba81a4a42b0a61e5d9903b91a1a70c6d68a19.jpg) +Figure 4. An illustration of the expansion operation in 3D space. + +where $\tilde{\phi}_i^{(1)} = \arccos \tilde{a}_i^{(1)}$ denotes the polar angle of each $\tilde{\mathbf{a}}_i$ and $\bar{\phi}_i^{(1)} = \frac{\pi}{2}\cdot \tilde{\phi}_i^{(1)} / \phi_0$ denotes the expanded polar angle. Fig. 4 illustrates this operation and the detailed derivation is shown in Appendix B. In the end, we use the inverse of $\mathbf{R}$ to map the expanded anchor $\bar{\mathbf{a}}_i$ to the original locations and obtain the final anchor: $\hat{\mathbf{a}}_i = \mathbf{R}^{\mathrm{T}}\bar{\mathbf{a}}_i$ . The pseudo-code of the whole expansion algorithm is provided in Appendix C. As shown, the expansion algorithm can increase the distances between different anchors while maintaining the semantic consistency of the original CLIP anchors (close anchors are still close after expansion). + +The second row of Tab. 1 shows the average CoS between anchors after the expansion algorithm. Compared with the original CLIP anchors, the average CoS between the expanded CLIP anchors is greatly reduced. Fig. 3 shows the learning curves with the expanded CLIP anchors (the green curves). The final robust accuracy (Fig. 3, right) is quite close to that of MMC anchors. + +# 3.5. Supervision Objective + +We propose to improve the anchor-based AT methods further by introducing an A-CE loss into the supervision objective. The A-CE loss is a composition of CE loss and the anchor alignment process (the $\cos \theta$ objective), which can also be seen as a special case of InfoNCE [35] with using $l_{2}$ normalized features. This could be confusing considering that previous works have shown that CE is inferior to other anchor-based objectives such as the $\theta$ and $\cos \theta$ [37] in the standard AT setting (see Sec. 2.1). However, we find that with the high CoS problem, A-CE can boost the performance of anchor-based AT methods in that the softmax function of A-CE could make a relaxation on the anchor alignment objective of the original anchor-based AT methods. Assuming that the image encoder $f_{\Theta}(\cdot)$ outputs normalized features $\mathbf{z} = f_{\Theta}(\mathbf{x}) \in \mathbb{R}^n$ and $||\mathbf{z}||_2 = 1$ and taking the $\cos \theta$ objective as an example, the $\cos \theta$ objective is to maximize $\mathbf{z}^T\hat{\mathbf{a}}_y$ , where $\hat{\mathbf{a}}_y$ denotes the GT anchor. It encourages the visual feature's CoS to be as close as possible to the GT anchor, while the A-CE loss, by putting $\cos \theta$ into the softmax function, only encourages the visual feature's CoS with GT to be larger than those with other anchors. With A-CE supervision, the output feature could be opti + +mized to be far away from all anchors, but it can still be correctly classified as long as it is closest to the GT anchor. This will be discussed further in Sec. 4.4. Overall, the supervision objective is: + +$$ +L _ {1} = \mathbb {E} _ {(\mathbf {x}, y)} \left[ - \log \frac {\exp \left(f _ {\Theta} (\mathbf {x} + \delta) ^ {T} \hat {\mathbf {a}} _ {y}\right)}{\sum_ {i = 1} ^ {N} \exp \left(f _ {\Theta} (\mathbf {x} + \delta) ^ {T} \hat {\mathbf {a}} _ {i}\right)} \right], \tag {3} +$$ + +where $\delta$ is adversarial perturbation generated by $\frac{\partial L_1}{\partial\mathbf{x}}$ . Unlike previous works [39] on standard training, we do not use a temperature parameter $\tau$ to scale the CoS. This is discussed in Appendix D. + +# 3.6. Smoothness Loss + +Since we prefer the adversarial robustness of the model in the zero-shot transfer setting, we additionally introduce a smoothing loss, encouraging the CoS between adversarial features to be similar to the benign features of an image: + +$$ +L _ {2} = \mathbb {E} _ {(\mathbf {x}, y)} \left[ - f _ {\Theta} (\mathbf {x}) ^ {T} f _ {\Theta} (\mathbf {x} + \delta) \right]. \tag {4} +$$ + +The smoothing loss is independent of the category labels of training examples. It can be expected to improve the adversarial robustness on novel categories. Finally, the whole loss in the training stage is given by $L = L_{1} + \alpha L_{2}$ , where $\alpha$ is a hyper-parameter. + +# 3.7. Performing Zero-Shot Recognition + +When performing zero-shot recognition on novel categories, the image features of a test image are generated by the image encoder, and the anchors for novel categories are generated by the text encoder and the expansion algorithm. Note that the internal parameters of the expansion algorithm, such as $\phi_0$ , are pre-computed using anchors from the training categories. Finally, the zero-shot prediction is the category whose text anchor exhibits the highest CoS with the image features. + +# 4. Experiments + +# 4.1. Experimental Setup + +Baseline. We first consider the standard ZSL setting. Due to the lack of adversarially robust methods in this setting, we compared LAAT with several robust few-shot methods directly: AQ [19], R-MAML [46], and GR [13]. ITS-MAML [15] was not compared as it used a different attack setting and no code was released. Note that zero-shot is harder than few-shot as the latter can have access to a few examples from the novel categories. + +We then considered a more pragmatic version of ZSL, which requires recognizing examples from both seen and novel categories. This is called the generalized zero-shot learning (GZSL) setting [8]. We trained several models with + +
DatasetMethodSettingCleanFGSMPGDCWAA
CIFAR-FSAQ [19]1-shot54.31 ± 0.5032.47 ± 0.4730.41 ± 0.4630.14 ± 0.4627.09 ± 2.07
R-MAML [46]1-shot38.53 ± 0.4726.76 ± 0.4426.21 ± 0.4329.76 ± 0.4616.67 ± 5.48
GR [13]1-shot45.27 ± 0.4939.60 ± 0.4638.03 ± 0.4637.00 ± 0.46-
LAATzero-shot55.60 ± 0.4641.17 ± 0.4440.12 ± 0.4437.35 ± 0.4436.45 ± 1.90
LAAT1-shot57.88 ± 0.4543.73 ± 0.4442.74 ± 0.4440.10 ± 0.4539.23 ± 2.12
MiniImageNetAQ [19]1-shot39.17 ± 0.4022.53 ± 0.3520.95 ± 0.3419.33 ± 0.3817.85 ± 1.41
R-MAML [46]1-shot36.18 ± 0.3623.85 ± 0.4621.74 ± 0.3419.68 ± 0.4013.69 ± 3.95
GR [13]1-shot36.14 ± 0.4529.23 ± 0.3327.57 ± 0.3826.61 ± 0.33-
LAATzero-shot46.97 ± 0.3530.69 ± 0.3029.27 ± 0.3127.91 ± 0.2925.24 ± 1.42
LAAT1-shot47.37 ± 0.3531.55 ± 0.3129.93 ± 0.3128.19 ± 0.2925.75 ± 1.45
+ +Table 2. The comparison of LAAT in 5-way zero-shot setting and previous works in 5-way 1-shot setting under $\epsilon = 8 / 255$ . The $95\%$ confidence interval was reported. Here all methods used Conv4-512 as the image model. The accuracy of GR under AA is not shown as it was not reported and no code was released. We also list the performance of LAAT in the 1-shot setting with image-text blended anchors (discussed in Sec. 4.3). The results higher than previous methods in each column are in bold. + +
MethodModelTrainingCifar100 100aPY 32AwA2 50COCO Obj. 80STL10 10OxfordPet 37DTD 47Caltech101 101Caltech256 257SUN 397
TeCoA [34]ViT-B/32*ε = 1/2550.004.390.881.070.980.001.431.661.330.00
ViT-B/32ε = 8/2550.005.271.451.331.460.281.892.972.360.14
ViT-B/16ε = 8/2550.206.451.951.561.660.492.257.033.810.29
LAATViT-B/32ε = 8/2556.6424.3223.3418.5641.1220.314.5934.4321.394.30
ViT-B/16ε = 8/2557.3234.9726.2120.6149.4128.814.7937.8924.025.18
+ +Table 3. Adversarial accuracy (\%) of models trained on ImageNet-1K on ten downstream datasets in GZSL setting. The number of categories is shown below the dataset name. * indicates the results of the originally released checkpoint of TeCoA with the training perturbation $\epsilon = 1/255$ . The robustness was all evaluated by PGD-20 under $\epsilon = 8/255$ . + +LAAT on ImageNet-1K and evaluated them directly on ten downstream datasets. In this setting, we mainly compare LAAT with TeCoA [34]. We did not compare with Yucel et al. [51] and Zhang et al. [55] as they relied on image attributes and were unsuitable for this setting. + +Datasets. When comparing with few-shot methods, we used two popular few-shot benchmarks CIFAR-FS [5] and MiniImageNet [43]. CIFAR-FS, a variant of Cifar100 [25], contains 64, 16, and 20 categories for training, validation, and testing, respectively. MiniImageNet, a subset of ImageNet-1k [40], contains 600 images of $84 \times 84$ size per category. We used the same dataset splitting setting as previous works [13, 19, 46], for training and testing. + +In the GZSL setting, we tested models trained on ImageNet-1K on ten downstream datasets such as AwA2 [48] and aPY [17]. Note that all of these datasets include several novel categories that have never appeared in ImageNet-1K, e.g., statue and ass in aPY [17], blue whale and dolphin in AwA2 [48]. More introduction to the datasets is shown in the Appendix E. + +Implementation details. Unless otherwise stated, we used the CLIP ViT-B/16 text encoder and the prompt text was set as "This is a photo of a {}". $\alpha$ was set to 3. + +In the standard ZSL setting, we followed previous few- + +shot works [13, 46] and used the widely applied model Conv4-512 [54] as the image model. The training recipe strictly followed GR [13]. We performed AT by generating adversarial examples via PGD method with maximum adversarial perturbation $\epsilon = 8 / 255$ in $l_{\infty}$ -norm bound, with $T = 7$ iterative steps (step size $s = 2 / 255$ ). + +In the GZSL setting, we trained two image models ViT-B/32 and ViT-B/16 [14] with LAAT on the large ImageNet-1K. The training basically followed the recipe proposed by Debenedetti et al. [12]. Here we performed AT by generating adversarial examples via PGD method with $\epsilon = 8 / 255$ in $l_{\infty}$ -norm bound and $T = 2$ iterative steps ( $s = 8 / 255$ ). + +# 4.2. Comparison Results + +The Standard ZSL Setting. We performed experiments in the standard and usual 5-way one/zero-shot setting [13, 19, 46]. Following previous works [19, 46], we report the robustness of our method against common adversarial attacks, including FGSM [20], PGD [32] with 20 iterative steps and step size $s = 2 / 255$ (denoted as PGD-20 in the following text), CW [7] (optimized by PGD for 30 steps with step size $s = 0.8 / 255$ ), and AutoAttack (AA) [11], an effective ensemble attack (several variants of PGD and one query-based black-box Square attack [2]) used to assess adversarial robustness. The maximal perturbation was set to commonly used $\epsilon = 8 / 255$ in $l_{\infty}$ setting. The experi + +![](images/4121eefba3a8e2b786e703bc121c0526128dab87eb520566af9221000fc5223a.jpg) +(a) Benign accuracy. + +![](images/10e7f7101445f6788aa6f1609ee5038c08bf84e34e76c4a9734b9686f8d6707d.jpg) +(b) Robust accuracy. +Figure 5. Classification accuracy on both benign and adversarial examples in 5-way few-shot setting on CIFAR-FS, with image feature anchors or with image-text blended anchors. The dashed line denotes 5-way zero-shot accuracy. + +mental results are reported from 2000 randomly sampled 5-way tasks. The results presented in Tab. 2 demonstrate that LAAT in the zero-shot setting achieved competitive performance with several methods in the few-shot setting. Notably, the adversarial robustness of LAAT models even surpassed that of the previous SOTA adversarily robust few-shot method GR [13]. Additionally, our method exhibited significantly better clean accuracy than all three few-shot methods. We note that these results should be interpreted cautiously because LAAT and these few-shot works used different auxiliary information: the former used text information of the test category names and the latter relied on a small number of example images from the test categories. + +GZSL Setting across Datasets. In the GZSL setting, the models were evaluated in $N$ -way classification setting, where $N$ is the number of categories in each dataset. All images on downstream datasets are resized to $224 \times 224$ in inference. Here we compared LAAT with TeCoA. TeCoA originally performed adversarial fine-tuning on ImageNet-1K with maximum perturbation $\epsilon = 1/255$ . For a fair comparison, here adversarial fine-tuning was also performed with the released code of TeCoA [34]. Tab. 3 shows the zero-shot results on ten downstream datasets under PGD-20. Here the accuracy is reported by sampling 100 examples per category. We can see that TeCoA almost failed to train adversarily robust models in the common $\epsilon$ setting, while the models trained with LAAT can obtain substantial adversarial robustness on several downstream datasets. + +Note that the models trained with LAAT did not perform quite well on Cifar100, DTD [9], and SUN [49]. This could be caused by the original small resolution of Cifar100 $(32 \times 32)$ [12] and the large inter-class differences between DTD, SUN, and ImageNet-1K. + +# 4.3. Extending LAAT to Few-Shot Setting + +Our zero-shot models supervised by LAAT can be naturally extended to the few-shot setting and the robustness can be further improved. Given the images of novel categories as + +
ObjectiveCleanRobust
Eucl39.1728.04
θ37.9227.84
cosθ41.9127.91
+ +Table 4. Classification accuracy (%) on CIFAR-FS supervised by the original CLIP anchors with different optimization objectives. + +
EXPA-CESMCleanRobust
41.9127.91
✓46.9832.71
✓✓54.0237.69
✓✓✓55.6040.12
+ +Table 5. Classification accuracy $(\%)$ of Conv4-512 on CIFAR-FS in different ablation experiments. The first two lines were the results obtained with $\cos \theta$ as the objective. + +the support set, we can follow the previous few-shot SOTA method [13] and use the prototype-based metric learning [41] to build image feature-based anchors. That is, in the $K$ -shot case, the image anchor $\mathbf{v}_y$ for a novel category $y$ can be computed as $\mathbf{v}_y = \mathrm{Norm}_2\{\sum_{i=1}^{K} f_\Theta(\mathbf{x}_i)\}$ , where $\mathbf{x}_i$ denotes the images of $y$ in the support set and $\mathrm{Norm}_2\{\cdot\}$ denotes the $l_2$ normalization. We can also weight the original text anchor $\mathbf{a}_y$ from the text encoder and the image feature from the support set to build an image-text blended anchor for $y$ : $\hat{\mathbf{a}}_y^* = \mathrm{Norm}_2\{\beta \cdot \hat{\mathbf{a}}_y + \sum_{i=1}^{K} f_\Theta(\mathbf{x}_i)\}$ . Here the text anchor can be regarded as the prior knowledge of $y$ . We set $\beta = 2$ in all experiments. + +Tab. 2 shows the results of different models with image-text blended anchors in the 1-shot setting. The 1-shot performance of LAAT was better than the zero-shot performance. Fig. 5 further shows the results with two different anchors $\hat{\mathbf{a}}^{\star}$ and $\mathbf{v}$ in $K$ -shot settings. We can see that the text anchor used to perform zero-shot recognition boosted the performance in various $K$ -shot settings, especially when the support set is small. + +# 4.4. Ablation Study + +Limited by computing resources, the ablation study is not studied on ImageNet-1K. Unless specified, it was performed with Conv4-512 in the 5-way zero-shot setting. The robustness was evaluated by the PGD-20 attack method. + +Effectiveness of each design. In Fig. 3, we have shown that if the model is supervised by CoS between the original CLIP anchors and the output feature, say, $\cos \theta$ , it is difficult to converge under AT. In Tab. 4, we show the results of other anchor-based AT optimization objectives, such as the angle [37] and the Euclidean distance [36], denoted as $\theta$ and Eucl respectively, in the ZSL setting. The results indicate that none of these three optimization objectives led to good classification accuracy. Note that random guessing in + +
AT MethodHyper-ParameterCleanRobust
EXP + A-CE + TRADES1/λ = 155.2619.22
1/λ = 655.8225.02
EXP + A-CE + SMα = 1.055.6239.95
α = 3.055.6040.12
α = 6.053.0140.29
+ +Table 6. Classification accuracy (%) on CIFAR-FS with different AT methods. + +
MethodSettingConv4-64ResNet12
CleanRobustCleanRobust
AQ1-shot43.7427.5443.5831.45
R-MAML1-shot32.4624.6142.5432.54
GR1-shot44.5137.4548.1339.29
LAATzero-shot49.8438.7458.3742.88
+ +this setting can obtain $20\%$ accuracy. + +We then conducted an ablation study to show the effectiveness of each design in LAAT: the expansion algorithm, the A-CE loss, and the smoothness loss, denoted as EXP, A-CE, and SM, respectively. The results in Tab. 5 show that all of these designs are helpful in improving the zero-shot adversarial robustness. + +As stated in Sec. 3.5, A-CE could make a relaxation on the original anchor-based AT objectives. To validate this, we computed the average CoS between GT anchors and the output visual features of the two models trained: EXP and $\mathrm{EXP} + \mathrm{A - CE}$ (EXP was supervised by $\cos \theta$ ; see Tab. 5). With A-CE loss, the visual feature's CoS with GT anchors dropped from 0.455 to 0.205 on average, while the accuracies on both benign and adversarial examples were improved (see Tab. 5). These results support our conjecture that some examples not quite close to the GT anchor can still be classified correctly with A-CE supervision. + +Smoothness v.s. TRADES. It is seen that SM has similar objectives with TRADES [52], but they are different in many aspects (see Appendix F). Here we also compare SM with TRADES empirically. The experiments were performed based on expanded anchors with A-CE supervision. The results are shown in Tab. 6, together with the most important hyper-parameter of each method. We can clearly see that TRADES performed much worse than our smoothness loss. In addition, the results show that the effectiveness of smoothing loss was not quite sensitive to $\alpha$ . + +Different image models. We also performed experiments with other popular image models in the few-shot setting other than Conv4-512, e.g., Conv4-64 [44] and ResNet12 [19]. Tab. 7 shows the results of the two image models on CIFAR-FS. Conv4-64 is smaller than Conv4-512 while + +Table 7. Classification accuracy (%) on CIFAR-FS with different image models. + +
Text EncoderEmbed-dimCleanRobust
None (MMC, one-shot)51247.5532.79
RN50x464052.2838.65
RN50x1676853.8638.02
ViT-B/3251252.6838.39
ViT-B/1651255.6040.12
ViT-L/1476852.7839.48
+ +Table 8. Classification accuracy (\%) supervised by MMC anchors in one-shot setting and the full LAAT with different CLIP text encoders in zero-shot setting on CIFAR-FS. Embed-dim denotes the dimension of the text anchors. + +ResNet12 is slightly larger. The models trained with LAAT showed stronger zero-shot adversarial robustness than those few-shot methods regardless of the model size. + +Different text encoders. Without text encoders, it is impossible for LAAT to achieve zero-shot robustness. In principle, arbitrary text encoders are feasible for LAAT if they have semantic consistency. We show the influence of using different CLIP text encoders in Tab. 8. We also show the results without using any text encoders. Instead, the image model was supervised by the MMC anchors (here zero-shot recognition cannot be performed while one-shot results are reported). We can see that using any text encoders can obtain strong zero-shot robustness, significantly surpassing the few-shot robustness without using text encoders. We also studied the relationship between the distances of different CLIP models' anchors and Cifar100 super-categories. The results in Appendix G further show that semantic consistency of the text encoder is one of the keys to zero-shot adversarial robustness. + +# 5. Conclusion and Discussion + +In this work, we propose a novel LAAT strategy for adversarially robust image classification in a zero-shot setting, inspired by recent vision-language models. Extensive experiments demonstrate that our method has strong adversarial robustness across different models and several zero-shot settings. We hope our work could encourage more researchers to investigate the robustness in the zero-shot setting, which could be a promising way to improve the scalability and flexibility of adversarially robust models. Limited by the data and computing resources, we cannot perform AT on large-scale image-text pairs directly like CLIP [39]. We believe it could be another promising way to achieve zero-shot adversarial robustness. + +# Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China (Nos. U2341228 and U19B2034). + +# References + +[1] Antonio Aguilera and Ricardo Pérez-Aguila. General n-dimensional rotations. In WSCG, pages 1-8, 2004. 4 +[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: A query-efficient black-box adversarial attack via random search. In Eur. Conf. Comput. Vis. (ECCV), pages 484-501, 2020. 6 +[3] Anish Athalye et al. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Int. Conf. Mach. Learn. (ICML), pages 274-283, 2018. 2 +[4] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 2 +[5] Luca Bertinetto, João F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Int. Conf. Learn. Represent. (ICLR), 2019, 6, 3 +[6] LE Blumenson. A derivation of n-dimensional spherical coordinates. The American Mathematical Monthly, 67(1):63-66, 1960. 4 +[7] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy, SP, pages 39-57, 2017. 6 +[8] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In Eur. Conf. Comput. Vis. (ECCV), pages 52-68, 2016. 5 +[9] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3606-3613, 2014. 7, 2 +[10] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, pages 215-223, 2011. 2 +[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Int. Conf. Mach. Learn. (ICML), pages 2206-2216, 2020. 6 +[12] Edoardo Debenedetti et al. A light recipe to train robust vision transformers. arXiv preprint arXiv:2209.07399, 2022.6, 7 +[13] Junhao Dong, Yuan Wang, Jianhuang Lai, and Xiaohua Xie. Improving adversarially robust few-shot image classification with generalizable representations. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9015-9024. IEEE, 2022. 3, 5, 6, 7 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent. (ICLR), 2021. 6 +[15] Xiaoyue Duan, Guoliang Kang, Runqi Wang, Shumin Han, Song Xue, Tian Wang, and Baochang Zhang. Rethinking the number of shots in robust model-agnostic meta-learning. arXiv preprint arXiv:2211.15180, 2022. 3, 5 +[16] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi + +Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1625-1634, 2018. 1 +[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David A. Forsyth. Describing objects by their attributes. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1778-1785, 2009. 6, 2 +[18] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004. 2 +[19] Micah Goldblum et al. Adversarily robust few-shot learning: A meta-learning approach. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 3, 5, 6, 8 +[20] Ian J. Goodfellow et al. Explaining and harnessing adversarial examples. In Int. Conf. Learn. Represent. (ICLR), 2015. 6 +[21] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 2 +[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zero-shot detection via vision and language knowledge distillation. In Int. Conf. Learn. Represent. (ICLR), 2021. 2, 4 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 770-778, 2016. 3 +[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Int. Conf. Mach. Learn. (ICML), pages 4904-4916, 2021. 2 +[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. California Institute of Technology, 2009. 3, 6, 2 +[26] Huafeng Kuang, Hong Liu, Yongjian Wu, and Rongrong Ji. Semantically consistent visual representation for adversarial robustness. IEEE Transactions on Information Forensics and Security, 2023. 2 +[27] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 36(3):453-465, 2014. 1 +[28] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In Int. Conf. Learn. Represent. (ICLR), 2022. 2, 4 +[29] Xiao Li, Hang Chen, and Xiaolin Hu. On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438, 2023. 2 +[30] Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, and Xiaolin Hu. Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 2 +[31] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in + +context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2 +[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Int. Conf. Learn. Represent. (ICLR), 2018. 1, 6 +[33] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. Adv. Neural Inform. Process. Syst. (NeurIPS), 32, 2019. 2 +[34] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. Int. Conf. Learn. Represent. (ICLR), 2023. 2, 3, 4, 6, 7 +[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5 +[36] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In Int. Conf. Learn. Represent. (ICLR), 2020. 1, 2, 4, 7 +[37] Tianyu Pang, Xiao Yang, Yinpeng Dong, Taufik Xu, Jun Zhu, and Hang Su. Boosting adversarial training with hypersphere embedding. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 1, 2, 3, 5, 7 +[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3498-3505, 2012. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 2, 5, 8 +[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. (IJCV), 115(3):211-252, 2015. 2, 6 +[41] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4077-4087, 2017. 7 +[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In Int. Conf. Learn. Represent. (ICLR), 2014. 1 +[43] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 6 +[44] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 8 +[45] Mengmeng Wang et al. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 4 + +[46] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 3, 5, 6 +[47] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 69-77, 2016. 2 +[48] Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - A comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 41(9):2251-2265, 2019. 6, 2 +[49] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3485-3492, 2010. 7, 2 +[50] Yunlong Yu, Zhong Ji, Yanwei Fu, Jichang Guo, Yanwei Pang, and Zhongfei (Mark) Zhang. Stacked semantics-guided attention model for fine-grained zero-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6007, 2018. 2 +[51] Mehmet Kerim Yucel et al. How robust are discriminatively trained zero-shot learning models? Image Vis. Comput., 119: 104392, 2022. 1, 3, 6 +[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In Int. Conf. Mach. Learn. (ICML), pages 7472-7482, 2019. 1, 2, 4, 8 +[53] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM Int. Conf. Multimedia, pages 5005-5013, 2022. 2 +[54] Manli Zhang, Jianhong Zhang, Zhiwu Lu, Tao Xiang, Mingyu Ding, and Songfang Huang. IEPT: instance-level and episode-level pretext tasks for few-shot learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 6 +[55] Xingxing Zhang, Shupeng Gui, Zhenfeng Zhu, Yao Zhao, and Ji Liu. ATZSL: defensive zero-shot recognition in the presence of adversaries. arXiv preprint arXiv:1910.10994, 2019. 1, 3, 6 +[56] Mo Zhou and Vishal M Patel. Enhancing adversarial robustness for deep metric learning. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15325-15334, 2022. 2 +[57] Xiaopei Zhu, Xiao Li, Jianmin Li, Zheyao Wang, and Xiaolin Hu. Fooling thermal infrared pedestrian detectors in real world using small bulbs. In AAAI, pages 3616-3624, 2021. 1 \ No newline at end of file diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/images.zip b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c66e5f46c5a754335a6aca0b51aaa32a23fd105b --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aaf223f44a9f50e737cb4860dbb2ca42d5449619d1e55b58afc1f26535eb968 +size 457951 diff --git a/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/layout.json b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1eed2968c1ae6b585fafea5269d130953bbbc4d9 --- /dev/null +++ b/2024/Language-Driven Anchors for Zero-Shot Adversarial Robustness/layout.json @@ -0,0 +1,9709 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "text", + "content": "Language-Driven Anchors for Zero-Shot Adversarial Robustness" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": "Xiao Li" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": " Wei Zhang" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": " Yining Liu" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": " Zhanhao Hu" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": " Bo Zhang" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": " Xiaolin Hu" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": "\\\n" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": "Department of Computer Science and Technology, BNRist, Institute for Artificial Intelligence, IDG/McGovern Institute for Brain Research, Tsinghua Laboratory of Brain and Intelligence, Tsinghua University, Beijing, China\\\n" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": "Harbin Institute of Technology, Weihai, China\\\n" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 526, + 270 + ], + "type": "text", + "content": "Chinese Institute for Brain Research (CIBR), Beijing, China\\\n{lixiao20, zhang-w19, huzhanha17}@mail.tsinghua.edu.cn\\\n22s030184@stu.hit.edu.cn\\\n{dcszb, xlhu}@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 290, + 192, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 290, + 192, + 304 + ], + "spans": [ + { + "bbox": [ + 143, + 290, + 192, + 304 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 316, + 290, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 290, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 290, + 628 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) are known to be susceptible to adversarial attacks. Previous researches mainly focus on improving adversarial robustness in the fully supervised setting, leaving the challenging domain of zero-shot adversarial robustness an open question. In this work, we investigate this domain by leveraging the recent advances in large vision-language models, such as CLIP, to introduce zero-shot adversarial robustness to DNNs. We propose LAAT, a Language-driven, Anchor-based Adversarial Training strategy. LAAT utilizes the features of a text encoder for each category as fixed anchors (normalized feature embeddings) for each category, which are then employed for adversarial training. By leveraging the semantic consistency of the text encoders, LAAT aims to enhance the adversarial robustness of the image model on novel categories. However, naively using text encoders leads to poor results. Through analysis, we identified the issue to be the high cosine similarity between text encoders. We then design an expansion algorithm and an alignment cross-entropy loss to alleviate the problem. Our experimental results demonstrated that LAAT significantly improves zero-shot adversarial robustness over state-of-the-art methods. LAAT has the potential to enhance adversarial robustness by large-scale multimodal models, especially when labeled data is unavailable during training. Code is available at https://github.com/LixiaoTHU/LAAT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 650, + 128, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 696 + ], + "type": "text", + "content": "Adversarial attacks [42], by adding deliberately designed perturbations to inputs, have posed a significant threat to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 334, + 289, + 536, + 395 + ], + "blocks": [ + { + "bbox": [ + 334, + 289, + 536, + 395 + ], + "lines": [ + { + "bbox": [ + 334, + 289, + 536, + 395 + ], + "spans": [ + { + "bbox": [ + 334, + 289, + 536, + 395 + ], + "type": "image", + "image_path": "e6c7a4984faab35b63942b9e8134d0ddc50121a2c225caf308d2aa28606f0ead.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 486 + ], + "type": "text", + "content": "Figure 1. The illustration of zero-shot ability with LAAT. Different colors of the marks indicate different categories. When a model is adversarially trained with the text anchors of table and lion (seen categories), it can recognize adversarial examples of the two categories (grey and green). Then due to the text anchors of chair and tiger (novel categories) being close to those of table and lion, respectively, the model can also recognize the two novel categories." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 652 + ], + "type": "text", + "content": "the application of Deep Neural Networks (DNNs) [16, 57]. In response to the threats, various techniques have been proposed to enhance adversarial robustness of DNNs. Among these, Adversarial Training (AT) and its variants [32, 36, 37, 52] have shown to be highly effective [16, 42, 57]. However, applying AT from scratch to novel datasets or categories can be time-consuming. In real-world scenarios, it is common for models to encounter examples from novel categories, known as zero-shot learning (ZSL) setting [27]. If adversarially trained models possess zero-shot capability, it can significantly enhance the scalability and practical utility of adversarially robust models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Adversarial robustness in the ZSL setting is challenging. Earlier attempts [51, 55] try to combine attribute-based ZSL methods with AT. However, these methods yielded modest improvements and lacked scalability due to the practical difficulties in obtaining attributes. Recently, large-scale" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 137, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24686" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "vision-language models such as CLIP [39] have made significant strides and are considered foundational models for various downstream applications. These models, trained on extensive text-image data, demonstrate strong zero-shot capability by using textual descriptions to recognize new visual categories, eliminating the need for explicit reliance on image attributes. In this work, different from previous attribute-based ZSL methods, we explore the utilization of foundation models to introduce zero-shot adversarial robustness to DNNs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "text", + "content": "We propose a Language-driven, Anchor-based Adversarial Training strategy, denoted as LAAT, to enhance the zero-shot adversarial robustness. Specifically, we use the popular text encoder from vision-language models, specifically CLIP [39], to obtain " + }, + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "text", + "content": " normalized feature embedding (named anchor in this paper) for the text (label name) corresponding to image categories. The text encoder exhibits a property called semantic consistency, where semantically similar categories are mapped to neighboring anchors in the feature space. Then we utilize the text anchors to supervise the image classification model with AT. After this process, the image model not only obtains adversarial robustness on seen categories but also aligns the image features with the text anchors. During zero-shot inference, by leveraging the semantic consistency between seen and novel categories, the image model can recognize novel categories while maintaining adversarial robustness. Fig. 1 illustrates the idea. We note that one recent work, TeCoA [34], also utilizes the ability of large-scale vision-language models to enhance the zero-shot adversarial robustness. However, TeCoA was only shown effective for extremely tiny adversarial noise (i.e., " + }, + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "inline_equation", + "content": "\\epsilon = 1 / 255" + }, + { + "bbox": [ + 46, + 194, + 289, + 457 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 289, + 639 + ], + "type": "text", + "content": "As CLIP models are not designed for adversarial robustness, directly applying the obtained anchors from CLIP text encoders to anchor-based AT [36, 37] (refer to Sec. 2.1) results in poor performance. We attribute it to the high Cosine Similarity (CoS) problem, i.e., the average cosine similarity between CLIP anchors are too high to be directly used in AT. See Sec. 3.3 for details. This problem has never been discovered in previous work utilizing CLIP models [22, 28, 34, 45]. To address it, we propose several techniques. We first introduce an expansion algorithm that maps the original CLIP anchors to new anchors with increased distances while maintaining semantic consistency. We then incorporate an Alignment Cross-Entropy (A-CE) loss as the optimization objective. In addition, we use a smoothness loss to enhance the robustness on unseen categories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "The experimental results showed that the models trained with LAAT achieved remarkable zero-shot performance against strong adversarial perturbations, even surpassing the previous state-of-the-art (SOTA) adversarily robust few-shot methods. We also trained several models with LAAT on a large dataset ImageNet-1K [40]. These mod" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "els showed substantial adversarial robustness across various downstream datasets in the zero-shot setting, significantly outperforming the recent TeCoA method [34]. The main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 120, + 547, + 204 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 306, + 120, + 547, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 547, + 155 + ], + "type": "text", + "content": "- We identify the high CoS problem in CLIP anchors for adversarial robustness and propose effective techniques to alleviate it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 547, + 204 + ], + "type": "text", + "content": "- The models trained by LAAT show substantial adversarial robustness across downstream datasets, suggesting that AT in the ZSL setting could be a promising way to improve the practical usefulness of AT." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 216, + 392, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 216, + 392, + 228 + ], + "spans": [ + { + "bbox": [ + 306, + 216, + 392, + 228 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 236, + 410, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 410, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 410, + 247 + ], + "type": "text", + "content": "2.1. Anchor-based AT" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": "AT has become one of the most effective strategies to improve adversarial robustness [3]. Several studies try to improve AT from various aspects [26, 29, 30, 36, 37, 52]. One series of works focuses on refining cross-entropy loss, which is the most commonly used loss in classification, to obtain better robustness with AT [36, 37]. Pang et al. [36] boosts AT by introducing the Max-Mahalanobis center (MMC) loss, a Euclidean loss where there is a fixed feature vector " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mu_y" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " for each category " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": ", and the optimization object is " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "||\\mathbf{z} - \\mu_y||_2^2" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " is the output feature. Pang et al. [37] uses Hypersphere Embedding to boost AT. It normalizes both weights " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " of the output layer, and proposes two losses in the form of " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " is the angle between " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\frac{\\mathbf{W}}{||\\mathbf{W}||_2}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "\\frac{\\mathbf{z}}{||\\mathbf{z}||_2}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": ". We call these AT methods anchor-based methods as they all used " + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 304, + 255, + 547, + 472 + ], + "type": "text", + "content": " normalized feature embeddings (anchors) for training. Besides, several metric learning works [33, 56] also used a similar paradigm to anchor-based AT for enhancing robustness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 479, + 470, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 470, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 470, + 492 + ], + "type": "text", + "content": "2.2. Language-driven Recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "Earlier ZSL models exploit auxiliary information, commonly in the form of image attributes, to transfer knowledge between seen and novel categories [47, 50]. Recently, language-driven recognition has caught attention in the ZSL field, especially with the recent advance of large-scale pretrained vision-language models [24, 39]. CLIP [39] demonstrated that classic recognition tasks can strongly benefit from millions of raw texts pertaining to images. CLIP uses an independent image encoder and text encoder to perform contrastive learning on 400M image-text pairs. After pretraining, natural language is used to reference visual concepts by computing the CoS between text features and the visual feature, enabling the zero-shot transfer of the model to downstream classification datasets. The basic paradigm of CLIP can be extended to several tasks such as semantic segmentation and object detection [22, 28, 45]. However, recent works show that CLIP models lack adversarial robustness and can be easily attacked [4, 34, 53]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24687" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 542, + 268 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 542, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 542, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 542, + 268 + ], + "type": "image", + "image_path": "b685b0efffcee3d0afa86d067db947d125bf0445ae02e8ba3b1933f71f306736.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 546, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 546, + 300 + ], + "type": "text", + "content": "Figure 2. The pipeline of LAAT. Only the image encoder is trainable in the figure. " + }, + { + "bbox": [ + 46, + 277, + 546, + 300 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 277, + 546, + 300 + ], + "type": "text", + "content": " indicates computing the CoS. Red arrows indicate the adversarial example generation process and brown arrows indicate the inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 320, + 284, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 320, + 284, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 284, + 335 + ], + "type": "text", + "content": "2.3. Adversarily Robust Zero/few-shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 339, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 506 + ], + "type": "text", + "content": "An early preprint work [55] combines AT with attribute-based ZSL methods but only gets a mild improvement over non-AT models. In addition, the method is hard to generalize without the attribute annotations. Yucel et al. [51] evaluates the adversarial robustness of classic attribute-based zero-shot models and concludes that adversarial robustness for ZSL field is suffering an immature state. Mao et al. [34] proposes several techniques such as visual prompting and fine-tuning to adapt large-scale models, e.g., the image models of CLIP, for zero-shot adversarial robustness, while their method called TeCoA was only shown to be effective under extremely tiny adversarial noise. Different from TeCoA, LAAT only uses the anchors from the CLIP text encoders." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 507, + 288, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 288, + 592 + ], + "type": "text", + "content": "A closely related area is adversarially robust few-shot learning, which has caught attention recently, too. We also compare LAAT with recent few-shot methods directly to comprehensively evaluate it. AQ [19], R-MAML [46], and ITS-MAML [15] study meta-learning jointly with AT. GR [13] tries to learn a generalizable robust representation by the combination of several AT techniques." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 603, + 95, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 95, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 95, + 615 + ], + "type": "text", + "content": "3. LAAT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 622, + 146, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 146, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 146, + 635 + ], + "type": "text", + "content": "3.1. Overall Pipeline" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "LAAT first obtains fixed text anchors from the text encoder, then uses them to adversarially supervise the training of image classification model. After AT, the image model will obtain adversarial robustness on seen categories and also align the image features with the text anchors to obtain zero-shot ability (see Fig. 1). The overall pipeline of LAAT is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "text", + "content": "illustrated in Fig. 2. Given the original text anchors, LAAT uses an expansion algorithm to obtain fixed Ground-Truth (GT) anchors suitable for AT, then uses an image encoder to extract both benign and adversarial features. The CoS between adversarial features and GT anchors are maximized by minimizing the A-CE loss " + }, + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "inline_equation", + "content": "(L_{1})" + }, + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "text", + "content": ". Besides, a smoothness loss " + }, + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "inline_equation", + "content": "(L_{2})" + }, + { + "bbox": [ + 304, + 321, + 547, + 417 + ], + "type": "text", + "content": " is applied to the adversarial and benign features. We then describe each module of LAAT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 425, + 440, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 440, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 440, + 437 + ], + "type": "text", + "content": "3.2. Text Anchor Extraction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": "Given the names of " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": " categories in the training set, they are first filled into " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": " sentences with a fixed prompt text, such as \"A photo of {}\". After that, these sentences are encoded into " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": " anchors " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": " by a fixed pre-trained text encoder, where " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "inline_equation", + "content": "||\\mathbf{a}_i||_2 = 1" + }, + { + "bbox": [ + 304, + 443, + 547, + 552 + ], + "type": "text", + "content": ". Many architectures are feasible if they have semantic consistency as mentioned before and in this work, we use the large-scale pre-trained CLIP text encoder. However, CLIP anchors cannot be used directly due to the high CoS problem, as analyzed below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 559, + 479, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 559, + 479, + 572 + ], + "spans": [ + { + "bbox": [ + 305, + 559, + 479, + 572 + ], + "type": "text", + "content": "3.3. High Cosine Similarity Problem" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": "We first make an intriguing observation that directly using CLIP anchors for anchor-based AT poses significant challenges in achieving convergence. This phenomenon is universal among different CLIP text encoders and different models. To illustrate this, we adversarially trained a ResNet18 [23] on Cifar100 [25] by maximizing the " + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": " objective [37], i.e., the CoS between the model output " + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": " and the GT anchor " + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_y" + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": " [37], which is computed as " + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\frac{\\mathbf{z}^T\\mathbf{a}_y}{||\\mathbf{z}||_2}" + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": " (note that " + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "inline_equation", + "content": "||\\mathbf{a}_y||_2 = 1" + }, + { + "bbox": [ + 304, + 578, + 547, + 713 + ], + "type": "text", + "content": "). The experimental details can be found in Appendix A. Fig. 3 shows the learning curves with CLIP anchors (the orange curves). For comparison, here we also" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24688" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 162, + 158 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 162, + 158 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 162, + 158 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 162, + 158 + ], + "type": "image", + "image_path": "5acee1955a3b9094233ca2dd05ae15db8dd9bad084366207c6d3cbab3f5cf0e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 160, + 138, + 171 + ], + "lines": [ + { + "bbox": [ + 76, + 160, + 138, + 171 + ], + "spans": [ + { + "bbox": [ + 76, + 160, + 138, + 171 + ], + "type": "text", + "content": "(a) Clean accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 73, + 282, + 158 + ], + "blocks": [ + { + "bbox": [ + 173, + 73, + 282, + 158 + ], + "lines": [ + { + "bbox": [ + 173, + 73, + 282, + 158 + ], + "spans": [ + { + "bbox": [ + 173, + 73, + 282, + 158 + ], + "type": "image", + "image_path": "306e86cb3b75467e3f676cd1e9059bcb21776e635b445500961d6f90110cda9f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "lines": [ + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "type": "text", + "content": "(b) Robust accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 179, + 287, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 179, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 287, + 213 + ], + "type": "text", + "content": "Figure 3. Learning curves of AT supervised by " + }, + { + "bbox": [ + 46, + 179, + 287, + 213 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 179, + 287, + 213 + ], + "type": "text", + "content": " with fixed anchors generated from MMC method, CLIP text encoder, and the expansion algorithm (see Sec. 3.4)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 341 + ], + "type": "text", + "content": "show the learning curves with MMC [36] anchors (the blue curves) in the same training setting. MMC anchors are precomputed by maximizing the distance between anchors [36] and thus the average CoS between MMC anchors is close to 0 (e.g., -0.01 on Cifar100). The accuracy with CLIP anchors increased quite slowly. Even after 200 epoch training, the accuracy was far lower than that of MMC anchors (note that zero-shot recognition cannot be achieved with MMC anchors for the lack of semantic consistency)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 287, + 484 + ], + "type": "text", + "content": "The stark discrepancy between the two learning curves motivates us to analyze the essential difference between the two sets of anchors. Inspired by the anchor generation process of MMC, we calculated the CoS between CLIP anchors and found them to be considerably larger than that of MMC anchors (-0.01). The first row of Tab. 1 presents the average CoS between anchors from Cifar100 categories generated by different CLIP models. Note that the high CoS between CLIP anchors is irrelevant to the prompt text. We found that even using several random sentences from Wikipedia to encode anchors by CLIP text encoders, the average CoS remained significantly large." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 651 + ], + "type": "text", + "content": "Although high CoS seems to have no effect on downstream tasks on clean images [22, 28, 45], we discovered that AT was remarkably affected because higher CoS implies that the anchors are closer to the decision boundary, which could leave more room for adversarial examples [36, 52]. This problem may also explain why TeCoA [34], which also used the text anchors during adaptation, failed to convert CLIP image models to adversially robust models against strong perturbations. Importantly, random or MMC anchors cannot mitigate this issue as they lack the semantic consistency provided by CLIP anchors and thus cannot be used in the ZSL setting. Therefore, the micro designs in our LAAT approach primarily aim to address the challenges associated with CLIP anchors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 171, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 171, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 171, + 672 + ], + "type": "text", + "content": "3.4. Expansion Algorithm" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "We first propose an expansion algorithm to increase the distances between anchors (reducing the CoS). The expansion algorithm can remap the anchors to increase the dis" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 543, + 117 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 543, + 117 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 543, + 117 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 543, + 117 + ], + "type": "table", + "html": "
TypeRN50x4RN50x16ViT-B/32ViT-B/16ViT-L/14
Original0.7000.7100.7790.7610.746
Expanded0.2380.2530.1990.1950.222
", + "image_path": "c3f91c122e6254f387c51fb762a16792769b24771e32276943c4617f8a4d6405.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 124, + 545, + 157 + ], + "lines": [ + { + "bbox": [ + 305, + 124, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 305, + 124, + 545, + 157 + ], + "type": "text", + "content": "Table 1. Average CoS between anchors before and after expansion algorithm (see Sec. 3.4) from Cifar100 categories generated by different CLIP text encoders." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": "tances between anchors while preserving the semantic consistency between them. High CoS means the normalized anchors " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " are dispersed over a cluster of the unit hypersphere (see Fig. 4). As we expect the remapped anchors to be also on the unit hyper-sphere, the expansion algorithm should be designed under spherical coordinates. Let us recap the transformation between " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": "-dimensional spherical coordinate system and Cartesian coordinates. Suppose " + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "inline_equation", + "content": "(x^{(1)}, x^{(2)}, \\ldots, x^{(n)})" + }, + { + "bbox": [ + 304, + 176, + 545, + 296 + ], + "type": "text", + "content": " is the Cartesian coordinate, we can compute them by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 300, + 528, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 300, + 528, + 314 + ], + "spans": [ + { + "bbox": [ + 310, + 300, + 528, + 314 + ], + "type": "interline_equation", + "content": "x ^ {(1)} = r \\cos \\phi^ {(1)}; \\quad x ^ {(2)} = r \\sin \\phi^ {(1)} \\cos \\phi^ {(2)}; \\quad \\dots ;", + "image_path": "69a6995d498453943aa891bc3c2ee9a7c11cc20e0dbbd40b112e3590f5f38aab.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 317, + 545, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 317, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 311, + 317, + 545, + 331 + ], + "type": "interline_equation", + "content": "x ^ {(n - 1)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\cos \\phi^ {(n - 1)}; \\tag {1}", + "image_path": "80df84c0f8ac9bb84a4a90434e58b69701ff1092c667cbc0906d4bd46f328de4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 334, + 484, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 334, + 484, + 348 + ], + "spans": [ + { + "bbox": [ + 311, + 334, + 484, + 348 + ], + "type": "interline_equation", + "content": "x ^ {(n)} = r \\sin \\phi^ {(1)} \\dots \\sin \\phi^ {(n - 2)} \\sin \\phi^ {(n - 1)},", + "image_path": "7098a278e09661ccd71335891202c04df91a29411047e9ff72d36b76eca33252.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " is the radial coordinate and " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi^{(1)},\\dots ,\\phi^{(n - 1)}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " are angular coordinates [6]. Here " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi^{(1)}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " is the polar angle (taking the three-dimensional spherical coordinate system like longitude and latitude system as an example, the polar angle is the angle with respect to the " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " axis). If the anchors are clustered around the polar, i.e., " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi_0\\coloneqq \\max_i\\phi_i^{(1)} < \\frac{\\pi}{2}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi_i^{(1)}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " denotes the polar angle of anchor " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": ", a natural expansion method is to enlarge the polar angle " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi_i^{(1)}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\frac{\\pi}{2}\\cdot \\phi_i^{(1)} / \\phi_0" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " and keep the other angular coordinates unchanged. Then the cluster can be expanded to the whole hemisphere. Note that we cannot expand them larger than " + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\frac{\\pi}{2}" + }, + { + "bbox": [ + 304, + 354, + 545, + 517 + ], + "type": "text", + "content": " as anchors may be close to each other from another hemisphere, which may hurt the semantic consistency of anchors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": "Thus, given the original anchor " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ", the first step of the expansion is to calculate a rotation matrix " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " to transform the center " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " of the clustered " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": "'s (" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{v} := \\sum_{i=1}^{N} \\mathbf{a}_i / || \\sum_{i=1}^{N} \\mathbf{a}_i||_2" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ") to the polar " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [1,0,\\dots,0]^T" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ", and the rotated anchor " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " is computed by: " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i = \\mathbf{Ra}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ". Note that computing " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " between two unit vectors has several ways and we use the Aguilera and Pérez-Aguila [1] algorithm. As " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": "'s are clustered around the polar " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ", we then enlarge the polar angle of " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " as described above. Denoting the " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": "-th element of " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "a_i^{(j)}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ", the largest polar angle between " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": "'s and " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " can be computed as " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\phi_0 = \\max_{1 \\leq i \\leq N} \\{\\arccos \\tilde{a}_i^{(1)}\\}" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": ". According to Eq. (1), the Cartesian coordinates of expanded anchor " + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_i" + }, + { + "bbox": [ + 304, + 518, + 546, + 677 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 681, + 545, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 681, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 329, + 681, + 545, + 703 + ], + "type": "interline_equation", + "content": "\\bar {a} _ {i} ^ {(1)} = \\cos \\bar {\\phi} _ {i} ^ {(1)}; \\tag {2}", + "image_path": "609c0e0934bdae2096177bc7732093b2c35a8baa694f781b5af409561d5d2eb7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 701, + 522, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 701, + 522, + 715 + ], + "spans": [ + { + "bbox": [ + 329, + 701, + 522, + 715 + ], + "type": "interline_equation", + "content": "\\bar {a} _ {i} ^ {(j)} = \\tilde {a} _ {i} ^ {(j)} \\cdot \\sin \\tilde {\\phi} _ {i} ^ {(1)} / \\sin \\tilde {\\phi} _ {i} ^ {(1)}, \\quad j = 2, \\dots , n,", + "image_path": "680ed940642325634e53a16c0c527c5cc8f262407aa887a356beeaad18e83a25.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24689" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 69, + 262, + 160 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 262, + 160 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 262, + 160 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 262, + 160 + ], + "type": "image", + "image_path": "20e83b15f2d4fc4f986b6d231a0ba81a4a42b0a61e5d9903b91a1a70c6d68a19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 49, + 171, + 285, + 183 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 285, + 183 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 285, + 183 + ], + "type": "text", + "content": "Figure 4. An illustration of the expansion operation in 3D space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\tilde{\\phi}_i^{(1)} = \\arccos \\tilde{a}_i^{(1)}" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": " denotes the polar angle of each " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{a}}_i" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\bar{\\phi}_i^{(1)} = \\frac{\\pi}{2}\\cdot \\tilde{\\phi}_i^{(1)} / \\phi_0" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": " denotes the expanded polar angle. Fig. 4 illustrates this operation and the detailed derivation is shown in Appendix B. In the end, we use the inverse of " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": " to map the expanded anchor " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{a}}_i" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": " to the original locations and obtain the final anchor: " + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_i = \\mathbf{R}^{\\mathrm{T}}\\bar{\\mathbf{a}}_i" + }, + { + "bbox": [ + 46, + 202, + 287, + 339 + ], + "type": "text", + "content": ". The pseudo-code of the whole expansion algorithm is provided in Appendix C. As shown, the expansion algorithm can increase the distances between different anchors while maintaining the semantic consistency of the original CLIP anchors (close anchors are still close after expansion)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 340, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 287, + 423 + ], + "type": "text", + "content": "The second row of Tab. 1 shows the average CoS between anchors after the expansion algorithm. Compared with the original CLIP anchors, the average CoS between the expanded CLIP anchors is greatly reduced. Fig. 3 shows the learning curves with the expanded CLIP anchors (the green curves). The final robust accuracy (Fig. 3, right) is quite close to that of MMC anchors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 173, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 173, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 173, + 444 + ], + "type": "text", + "content": "3.5. Supervision Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": "We propose to improve the anchor-based AT methods further by introducing an A-CE loss into the supervision objective. The A-CE loss is a composition of CE loss and the anchor alignment process (the " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " objective), which can also be seen as a special case of InfoNCE [35] with using " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " normalized features. This could be confusing considering that previous works have shown that CE is inferior to other anchor-based objectives such as the " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " [37] in the standard AT setting (see Sec. 2.1). However, we find that with the high CoS problem, A-CE can boost the performance of anchor-based AT methods in that the softmax function of A-CE could make a relaxation on the anchor alignment objective of the original anchor-based AT methods. Assuming that the image encoder " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_{\\Theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " outputs normalized features " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = f_{\\Theta}(\\mathbf{x}) \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "||\\mathbf{z}||_2 = 1" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " and taking the " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " objective as an example, the " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " objective is to maximize " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^T\\hat{\\mathbf{a}}_y" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_y" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " denotes the GT anchor. It encourages the visual feature's CoS to be as close as possible to the GT anchor, while the A-CE loss, by putting " + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 450, + 289, + 715 + ], + "type": "text", + "content": " into the softmax function, only encourages the visual feature's CoS with GT to be larger than those with other anchors. With A-CE supervision, the output feature could be opti" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "mized to be far away from all anchors, but it can still be correctly classified as long as it is closest to the GT anchor. This will be discussed further in Sec. 4.4. Overall, the supervision objective is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 129, + 545, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 129, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 318, + 129, + 545, + 162 + ], + "type": "interline_equation", + "content": "L _ {1} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - \\log \\frac {\\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {y}\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(f _ {\\Theta} (\\mathbf {x} + \\delta) ^ {T} \\hat {\\mathbf {a}} _ {i}\\right)} \\right], \\tag {3}", + "image_path": "ca881728c77508e1ae6ca105879e3ceda4030673eb316a7ffecb5e73b6064fb5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "text", + "content": " is adversarial perturbation generated by " + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\frac{\\partial L_1}{\\partial\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "text", + "content": ". Unlike previous works [39] on standard training, we do not use a temperature parameter " + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 171, + 545, + 220 + ], + "type": "text", + "content": " to scale the CoS. This is discussed in Appendix D." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 228, + 408, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 408, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 408, + 239 + ], + "type": "text", + "content": "3.6. Smoothness Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 545, + 294 + ], + "type": "text", + "content": "Since we prefer the adversarial robustness of the model in the zero-shot transfer setting, we additionally introduce a smoothing loss, encouraging the CoS between adversarial features to be similar to the benign features of an image:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 302, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 302, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 351, + 302, + 545, + 319 + ], + "type": "interline_equation", + "content": "L _ {2} = \\mathbb {E} _ {(\\mathbf {x}, y)} \\left[ - f _ {\\Theta} (\\mathbf {x}) ^ {T} f _ {\\Theta} (\\mathbf {x} + \\delta) \\right]. \\tag {4}", + "image_path": "da370c7683c12c8018fa0dbe6e7223b63aee42d376213d65313def75dc85bc24.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": "The smoothing loss is independent of the category labels of training examples. It can be expected to improve the adversarial robustness on novel categories. Finally, the whole loss in the training stage is given by " + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "inline_equation", + "content": "L = L_{1} + \\alpha L_{2}" + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 326, + 545, + 386 + ], + "type": "text", + "content": " is a hyper-parameter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 394, + 491, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 491, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 491, + 407 + ], + "type": "text", + "content": "3.7. Performing Zero-Shot Recognition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": "When performing zero-shot recognition on novel categories, the image features of a test image are generated by the image encoder, and the anchors for novel categories are generated by the text encoder and the expansion algorithm. Note that the internal parameters of the expansion algorithm, such as " + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\phi_0" + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": ", are pre-computed using anchors from the training categories. Finally, the zero-shot prediction is the category whose text anchor exhibits the highest CoS with the image features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 388, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 422, + 563 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 665 + ], + "type": "text", + "content": "Baseline. We first consider the standard ZSL setting. Due to the lack of adversarially robust methods in this setting, we compared LAAT with several robust few-shot methods directly: AQ [19], R-MAML [46], and GR [13]. ITS-MAML [15] was not compared as it used a different attack setting and no code was released. Note that zero-shot is harder than few-shot as the latter can have access to a few examples from the novel categories." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "We then considered a more pragmatic version of ZSL, which requires recognizing examples from both seen and novel categories. This is called the generalized zero-shot learning (GZSL) setting [8]. We trained several models with" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24690" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 541, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 541, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 541, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 541, + 196 + ], + "type": "table", + "html": "
DatasetMethodSettingCleanFGSMPGDCWAA
CIFAR-FSAQ [19]1-shot54.31 ± 0.5032.47 ± 0.4730.41 ± 0.4630.14 ± 0.4627.09 ± 2.07
R-MAML [46]1-shot38.53 ± 0.4726.76 ± 0.4426.21 ± 0.4329.76 ± 0.4616.67 ± 5.48
GR [13]1-shot45.27 ± 0.4939.60 ± 0.4638.03 ± 0.4637.00 ± 0.46-
LAATzero-shot55.60 ± 0.4641.17 ± 0.4440.12 ± 0.4437.35 ± 0.4436.45 ± 1.90
LAAT1-shot57.88 ± 0.4543.73 ± 0.4442.74 ± 0.4440.10 ± 0.4539.23 ± 2.12
MiniImageNetAQ [19]1-shot39.17 ± 0.4022.53 ± 0.3520.95 ± 0.3419.33 ± 0.3817.85 ± 1.41
R-MAML [46]1-shot36.18 ± 0.3623.85 ± 0.4621.74 ± 0.3419.68 ± 0.4013.69 ± 3.95
GR [13]1-shot36.14 ± 0.4529.23 ± 0.3327.57 ± 0.3826.61 ± 0.33-
LAATzero-shot46.97 ± 0.3530.69 ± 0.3029.27 ± 0.3127.91 ± 0.2925.24 ± 1.42
LAAT1-shot47.37 ± 0.3531.55 ± 0.3129.93 ± 0.3128.19 ± 0.2925.75 ± 1.45
", + "image_path": "eaf28adb344707102d204cc0e1fd5bb3b953d0d0c5648a9ac7472b8875c40f7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 256, + 541, + 339 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "type": "text", + "content": "Table 2. The comparison of LAAT in 5-way zero-shot setting and previous works in 5-way 1-shot setting under " + }, + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 201, + 547, + 248 + ], + "type": "text", + "content": " confidence interval was reported. Here all methods used Conv4-512 as the image model. The accuracy of GR under AA is not shown as it was not reported and no code was released. We also list the performance of LAAT in the 1-shot setting with image-text blended anchors (discussed in Sec. 4.3). The results higher than previous methods in each column are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 256, + 541, + 339 + ], + "lines": [ + { + "bbox": [ + 53, + 256, + 541, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 256, + 541, + 339 + ], + "type": "table", + "html": "
MethodModelTrainingCifar100 100aPY 32AwA2 50COCO Obj. 80STL10 10OxfordPet 37DTD 47Caltech101 101Caltech256 257SUN 397
TeCoA [34]ViT-B/32*ε = 1/2550.004.390.881.070.980.001.431.661.330.00
ViT-B/32ε = 8/2550.005.271.451.331.460.281.892.972.360.14
ViT-B/16ε = 8/2550.206.451.951.561.660.492.257.033.810.29
LAATViT-B/32ε = 8/2556.6424.3223.3418.5641.1220.314.5934.4321.394.30
ViT-B/16ε = 8/2557.3234.9726.2120.6149.4128.814.7937.8924.025.18
", + "image_path": "ccf0252226909b80f774d0a123e35dc3412f5d1dd161d514b3d9377a18bd5578.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "type": "text", + "content": "Table 3. Adversarial accuracy (\\%) of models trained on ImageNet-1K on ten downstream datasets in GZSL setting. The number of categories is shown below the dataset name. * indicates the results of the originally released checkpoint of TeCoA with the training perturbation " + }, + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\epsilon = 1/255" + }, + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "type": "text", + "content": ". The robustness was all evaluated by PGD-20 under " + }, + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 46, + 345, + 548, + 381 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 399, + 288, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 461 + ], + "type": "text", + "content": "LAAT on ImageNet-1K and evaluated them directly on ten downstream datasets. In this setting, we mainly compare LAAT with TeCoA [34]. We did not compare with Yucel et al. [51] and Zhang et al. [55] as they relied on image attributes and were unsuitable for this setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": "Datasets. When comparing with few-shot methods, we used two popular few-shot benchmarks CIFAR-FS [5] and MiniImageNet [43]. CIFAR-FS, a variant of Cifar100 [25], contains 64, 16, and 20 categories for training, validation, and testing, respectively. MiniImageNet, a subset of ImageNet-1k [40], contains 600 images of " + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "inline_equation", + "content": "84 \\times 84" + }, + { + "bbox": [ + 46, + 472, + 287, + 567 + ], + "type": "text", + "content": " size per category. We used the same dataset splitting setting as previous works [13, 19, 46], for training and testing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 288, + 653 + ], + "type": "text", + "content": "In the GZSL setting, we tested models trained on ImageNet-1K on ten downstream datasets such as AwA2 [48] and aPY [17]. Note that all of these datasets include several novel categories that have never appeared in ImageNet-1K, e.g., statue and ass in aPY [17], blue whale and dolphin in AwA2 [48]. More introduction to the datasets is shown in the Appendix E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 664, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 288, + 700 + ], + "type": "text", + "content": "Implementation details. Unless otherwise stated, we used the CLIP ViT-B/16 text encoder and the prompt text was set as \"This is a photo of a {}\". " + }, + { + "bbox": [ + 47, + 664, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 664, + 288, + 700 + ], + "type": "text", + "content": " was set to 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "In the standard ZSL setting, we followed previous few-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "shot works [13, 46] and used the widely applied model Conv4-512 [54] as the image model. The training recipe strictly followed GR [13]. We performed AT by generating adversarial examples via PGD method with maximum adversarial perturbation " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": "-norm bound, with " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "T = 7" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": " iterative steps (step size " + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "inline_equation", + "content": "s = 2 / 255" + }, + { + "bbox": [ + 304, + 399, + 545, + 472 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": "In the GZSL setting, we trained two image models ViT-B/32 and ViT-B/16 [14] with LAAT on the large ImageNet-1K. The training basically followed the recipe proposed by Debenedetti et al. [12]. Here we performed AT by generating adversarial examples via PGD method with " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": "-norm bound and " + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "T = 2" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": " iterative steps (" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "inline_equation", + "content": "s = 8 / 255" + }, + { + "bbox": [ + 304, + 472, + 545, + 545 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 551, + 424, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 424, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 424, + 563 + ], + "type": "text", + "content": "4.2. Comparison Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "The Standard ZSL Setting. We performed experiments in the standard and usual 5-way one/zero-shot setting [13, 19, 46]. Following previous works [19, 46], we report the robustness of our method against common adversarial attacks, including FGSM [20], PGD [32] with 20 iterative steps and step size " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "s = 2 / 255" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " (denoted as PGD-20 in the following text), CW [7] (optimized by PGD for 30 steps with step size " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "s = 0.8 / 255" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "), and AutoAttack (AA) [11], an effective ensemble attack (several variants of PGD and one query-based black-box Square attack [2]) used to assess adversarial robustness. The maximal perturbation was set to commonly used " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "l_{\\infty}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " setting. The experi" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "24691" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 164, + 158 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 164, + 158 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 164, + 158 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 164, + 158 + ], + "type": "image", + "image_path": "4121eefba3a8e2b786e703bc121c0526128dab87eb520566af9221000fc5223a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 160, + 141, + 171 + ], + "lines": [ + { + "bbox": [ + 73, + 160, + 141, + 171 + ], + "spans": [ + { + "bbox": [ + 73, + 160, + 141, + 171 + ], + "type": "text", + "content": "(a) Benign accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 73, + 282, + 158 + ], + "blocks": [ + { + "bbox": [ + 168, + 73, + 282, + 158 + ], + "lines": [ + { + "bbox": [ + 168, + 73, + 282, + 158 + ], + "spans": [ + { + "bbox": [ + 168, + 73, + 282, + 158 + ], + "type": "image", + "image_path": "10e7f7101445f6788aa6f1609ee5038c08bf84e34e76c4a9734b9686f8d6707d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "lines": [ + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "spans": [ + { + "bbox": [ + 194, + 160, + 261, + 171 + ], + "type": "text", + "content": "(b) Robust accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 180, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 288, + 224 + ], + "type": "text", + "content": "Figure 5. Classification accuracy on both benign and adversarial examples in 5-way few-shot setting on CIFAR-FS, with image feature anchors or with image-text blended anchors. The dashed line denotes 5-way zero-shot accuracy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 287, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 402 + ], + "type": "text", + "content": "mental results are reported from 2000 randomly sampled 5-way tasks. The results presented in Tab. 2 demonstrate that LAAT in the zero-shot setting achieved competitive performance with several methods in the few-shot setting. Notably, the adversarial robustness of LAAT models even surpassed that of the previous SOTA adversarily robust few-shot method GR [13]. Additionally, our method exhibited significantly better clean accuracy than all three few-shot methods. We note that these results should be interpreted cautiously because LAAT and these few-shot works used different auxiliary information: the former used text information of the test category names and the latter relied on a small number of example images from the test categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": "GZSL Setting across Datasets. In the GZSL setting, the models were evaluated in " + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": "-way classification setting, where " + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": " is the number of categories in each dataset. All images on downstream datasets are resized to " + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": " in inference. Here we compared LAAT with TeCoA. TeCoA originally performed adversarial fine-tuning on ImageNet-1K with maximum perturbation " + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon = 1/255" + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": ". For a fair comparison, here adversarial fine-tuning was also performed with the released code of TeCoA [34]. Tab. 3 shows the zero-shot results on ten downstream datasets under PGD-20. Here the accuracy is reported by sampling 100 examples per category. We can see that TeCoA almost failed to train adversarily robust models in the common " + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 411, + 287, + 590 + ], + "type": "text", + "content": " setting, while the models trained with LAAT can obtain substantial adversarial robustness on several downstream datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": "Note that the models trained with LAAT did not perform quite well on Cifar100, DTD [9], and SUN [49]. This could be caused by the original small resolution of Cifar100 " + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "inline_equation", + "content": "(32 \\times 32)" + }, + { + "bbox": [ + 46, + 591, + 287, + 651 + ], + "type": "text", + "content": " [12] and the large inter-class differences between DTD, SUN, and ImageNet-1K." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 246, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 246, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 246, + 672 + ], + "type": "text", + "content": "4.3. Extending LAAT to Few-Shot Setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Our zero-shot models supervised by LAAT can be naturally extended to the few-shot setting and the robustness can be further improved. Given the images of novel categories as" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 366, + 70, + 486, + 117 + ], + "blocks": [ + { + "bbox": [ + 366, + 70, + 486, + 117 + ], + "lines": [ + { + "bbox": [ + 366, + 70, + 486, + 117 + ], + "spans": [ + { + "bbox": [ + 366, + 70, + 486, + 117 + ], + "type": "table", + "html": "
ObjectiveCleanRobust
Eucl39.1728.04
θ37.9227.84
cosθ41.9127.91
", + "image_path": "af8bcb23df979fd8e2b254de6d0473c5fc71bfac2ade925131792fe57fa3a9ec.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 347, + 158, + 506, + 216 + ], + "blocks": [ + { + "bbox": [ + 305, + 125, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 305, + 125, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 305, + 125, + 545, + 148 + ], + "type": "text", + "content": "Table 4. Classification accuracy (%) on CIFAR-FS supervised by the original CLIP anchors with different optimization objectives." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 158, + 506, + 216 + ], + "lines": [ + { + "bbox": [ + 347, + 158, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 347, + 158, + 506, + 216 + ], + "type": "table", + "html": "
EXPA-CESMCleanRobust
41.9127.91
✓46.9832.71
✓✓54.0237.69
✓✓✓55.6040.12
", + "image_path": "6dfbd35b776eb5061825e811d09ced29553af95ff3490959b5c20371899eefb1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": "Table 5. Classification accuracy " + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": " of Conv4-512 on CIFAR-FS in different ablation experiments. The first two lines were the results obtained with " + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": " as the objective." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": "the support set, we can follow the previous few-shot SOTA method [13] and use the prototype-based metric learning [41] to build image feature-based anchors. That is, in the " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": "-shot case, the image anchor " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " for a novel category " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " can be computed as " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_y = \\mathrm{Norm}_2\\{\\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " denotes the images of " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " in the support set and " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathrm{Norm}_2\\{\\cdot\\}" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " normalization. We can also weight the original text anchor " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " from the text encoder and the image feature from the support set to build an image-text blended anchor for " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}_y^* = \\mathrm{Norm}_2\\{\\beta \\cdot \\hat{\\mathbf{a}}_y + \\sum_{i=1}^{K} f_\\Theta(\\mathbf{x}_i)\\}" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": ". Here the text anchor can be regarded as the prior knowledge of " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\beta = 2" + }, + { + "bbox": [ + 304, + 279, + 545, + 424 + ], + "type": "text", + "content": " in all experiments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": "Tab. 2 shows the results of different models with image-text blended anchors in the 1-shot setting. The 1-shot performance of LAAT was better than the zero-shot performance. Fig. 5 further shows the results with two different anchors " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}^{\\star}" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": "-shot settings. We can see that the text anchor used to perform zero-shot recognition boosted the performance in various " + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 425, + 545, + 520 + ], + "type": "text", + "content": "-shot settings, especially when the support set is small." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 400, + 542 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 548, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 595 + ], + "type": "text", + "content": "Limited by computing resources, the ablation study is not studied on ImageNet-1K. Unless specified, it was performed with Conv4-512 in the 5-way zero-shot setting. The robustness was evaluated by the PGD-20 attack method." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "Effectiveness of each design. In Fig. 3, we have shown that if the model is supervised by CoS between the original CLIP anchors and the output feature, say, " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": ", it is difficult to converge under AT. In Tab. 4, we show the results of other anchor-based AT optimization objectives, such as the angle [37] and the Euclidean distance [36], denoted as " + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": " and Eucl respectively, in the ZSL setting. The results indicate that none of these three optimization objectives led to good classification accuracy. Note that random guessing in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24692" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 281, + 139 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 281, + 139 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 281, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 281, + 139 + ], + "type": "table", + "html": "
AT MethodHyper-ParameterCleanRobust
EXP + A-CE + TRADES1/λ = 155.2619.22
1/λ = 655.8225.02
EXP + A-CE + SMα = 1.055.6239.95
α = 3.055.6040.12
α = 6.053.0140.29
", + "image_path": "1df121190250c86ba1adbdb484fb645c1f792b5e99b9b76ec7a470d3f1226e7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 61, + 180, + 275, + 249 + ], + "blocks": [ + { + "bbox": [ + 47, + 147, + 287, + 169 + ], + "lines": [ + { + "bbox": [ + 47, + 147, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 287, + 169 + ], + "type": "text", + "content": "Table 6. Classification accuracy (%) on CIFAR-FS with different AT methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 180, + 275, + 249 + ], + "lines": [ + { + "bbox": [ + 61, + 180, + 275, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 275, + 249 + ], + "type": "table", + "html": "
MethodSettingConv4-64ResNet12
CleanRobustCleanRobust
AQ1-shot43.7427.5443.5831.45
R-MAML1-shot32.4624.6142.5432.54
GR1-shot44.5137.4548.1339.29
LAATzero-shot49.8438.7458.3742.88
", + "image_path": "df51eb95c67a0606ea19945275bf52d6cdb3717fa5e0dac94137c479b28427cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 301, + 198, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 198, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 198, + 312 + ], + "type": "text", + "content": "this setting can obtain " + }, + { + "bbox": [ + 47, + 301, + 198, + 312 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 47, + 301, + 198, + 312 + ], + "type": "text", + "content": " accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 313, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 287, + 384 + ], + "type": "text", + "content": "We then conducted an ablation study to show the effectiveness of each design in LAAT: the expansion algorithm, the A-CE loss, and the smoothness loss, denoted as EXP, A-CE, and SM, respectively. The results in Tab. 5 show that all of these designs are helpful in improving the zero-shot adversarial robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "content": "As stated in Sec. 3.5, A-CE could make a relaxation on the original anchor-based AT objectives. To validate this, we computed the average CoS between GT anchors and the output visual features of the two models trained: EXP and " + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{EXP} + \\mathrm{A - CE}" + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "content": " (EXP was supervised by " + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "inline_equation", + "content": "\\cos \\theta" + }, + { + "bbox": [ + 46, + 384, + 287, + 515 + ], + "type": "text", + "content": "; see Tab. 5). With A-CE loss, the visual feature's CoS with GT anchors dropped from 0.455 to 0.205 on average, while the accuracies on both benign and adversarial examples were improved (see Tab. 5). These results support our conjecture that some examples not quite close to the GT anchor can still be classified correctly with A-CE supervision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 644 + ], + "type": "text", + "content": "Smoothness v.s. TRADES. It is seen that SM has similar objectives with TRADES [52], but they are different in many aspects (see Appendix F). Here we also compare SM with TRADES empirically. The experiments were performed based on expanded anchors with A-CE supervision. The results are shown in Tab. 6, together with the most important hyper-parameter of each method. We can clearly see that TRADES performed much worse than our smoothness loss. In addition, the results show that the effectiveness of smoothing loss was not quite sensitive to " + }, + { + "bbox": [ + 46, + 525, + 287, + 644 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 525, + 287, + 644 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": "Different image models. We also performed experiments with other popular image models in the few-shot setting other than Conv4-512, e.g., Conv4-64 [44] and ResNet12 [19]. Tab. 7 shows the results of the two image models on CIFAR-FS. Conv4-64 is smaller than Conv4-512 while" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 539, + 149 + ], + "blocks": [ + { + "bbox": [ + 47, + 257, + 287, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 257, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 257, + 287, + 281 + ], + "type": "text", + "content": "Table 7. Classification accuracy (%) on CIFAR-FS with different image models." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 539, + 149 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 539, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 539, + 149 + ], + "type": "table", + "html": "
Text EncoderEmbed-dimCleanRobust
None (MMC, one-shot)51247.5532.79
RN50x464052.2838.65
RN50x1676853.8638.02
ViT-B/3251252.6838.39
ViT-B/1651255.6040.12
ViT-L/1476852.7839.48
", + "image_path": "f087c58a1eeb5d127d83cd2370a289dccc1e0b75fb9645908a44359e893a6959.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 158, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 545, + 201 + ], + "type": "text", + "content": "Table 8. Classification accuracy (\\%) supervised by MMC anchors in one-shot setting and the full LAAT with different CLIP text encoders in zero-shot setting on CIFAR-FS. Embed-dim denotes the dimension of the text anchors." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 223, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 223, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 223, + 545, + 259 + ], + "type": "text", + "content": "ResNet12 is slightly larger. The models trained with LAAT showed stronger zero-shot adversarial robustness than those few-shot methods regardless of the model size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 268, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 459 + ], + "type": "text", + "content": "Different text encoders. Without text encoders, it is impossible for LAAT to achieve zero-shot robustness. In principle, arbitrary text encoders are feasible for LAAT if they have semantic consistency. We show the influence of using different CLIP text encoders in Tab. 8. We also show the results without using any text encoders. Instead, the image model was supervised by the MMC anchors (here zero-shot recognition cannot be performed while one-shot results are reported). We can see that using any text encoders can obtain strong zero-shot robustness, significantly surpassing the few-shot robustness without using text encoders. We also studied the relationship between the distances of different CLIP models' anchors and Cifar100 super-categories. The results in Appendix G further show that semantic consistency of the text encoder is one of the keys to zero-shot adversarial robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 458, + 483 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 491, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 646 + ], + "type": "text", + "content": "In this work, we propose a novel LAAT strategy for adversarially robust image classification in a zero-shot setting, inspired by recent vision-language models. Extensive experiments demonstrate that our method has strong adversarial robustness across different models and several zero-shot settings. We hope our work could encourage more researchers to investigate the robustness in the zero-shot setting, which could be a promising way to improve the scalability and flexibility of adversarially robust models. Limited by the data and computing resources, we cannot perform AT on large-scale image-text pairs directly like CLIP [39]. We believe it could be another promising way to achieve zero-shot adversarial robustness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 404, + 670 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (Nos. U2341228 and U19B2034)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24693" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Antonio Aguilera and Ricardo Pérez-Aguila. General n-dimensional rotations. In WSCG, pages 1-8, 2004. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 158 + ], + "type": "text", + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: A query-efficient black-box adversarial attack via random search. In Eur. Conf. Comput. Vis. (ECCV), pages 484-501, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 287, + 192 + ], + "type": "text", + "content": "[3] Anish Athalye et al. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Int. Conf. Mach. Learn. (ICML), pages 274-283, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 225 + ], + "type": "text", + "content": "[4] Yuanhao Ban and Yinpeng Dong. Pre-trained adversarial perturbations. Adv. Neural Inform. Process. Syst. (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 270 + ], + "type": "text", + "content": "[5] Luca Bertinetto, João F. Henriques, Philip H. S. Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In Int. Conf. Learn. Represent. (ICLR), 2019, 6, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 304 + ], + "type": "text", + "content": "[6] LE Blumenson. A derivation of n-dimensional spherical coordinates. The American Mathematical Monthly, 67(1):63-66, 1960. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 307, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 287, + 340 + ], + "type": "text", + "content": "[7] Nicholas Carlini and David A. Wagner. Towards evaluating the robustness of neural networks. In IEEE Symposium on Security and Privacy, SP, pages 39-57, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 287, + 385 + ], + "type": "text", + "content": "[8] Wei-Lun Chao, Soravit Changpinyo, Boqing Gong, and Fei Sha. An empirical study and analysis of generalized zero-shot learning for object recognition in the wild. In Eur. Conf. Comput. Vis. (ECCV), pages 52-68, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 387, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 430 + ], + "type": "text", + "content": "[9] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3606-3613, 2014. 7, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 464 + ], + "type": "text", + "content": "[10] Adam Coates, Andrew Ng, and Honglak Lee. An analysis of single-layer networks in unsupervised feature learning. In AISTATS, pages 215-223, 2011. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 509 + ], + "type": "text", + "content": "[11] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In Int. Conf. Mach. Learn. (ICML), pages 2206-2216, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 542 + ], + "type": "text", + "content": "[12] Edoardo Debenedetti et al. A light recipe to train robust vision transformers. arXiv preprint arXiv:2209.07399, 2022.6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[13] Junhao Dong, Yuan Wang, Jianhuang Lai, and Xiaohua Xie. Improving adversarially robust few-shot image classification with generalizable representations. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 9015-9024. IEEE, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 644 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In Int. Conf. Learn. Represent. (ICLR), 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Xiaoyue Duan, Guoliang Kang, Runqi Wang, Shumin Han, Song Xue, Tian Wang, and Baochang Zhang. Rethinking the number of shots in robust model-agnostic meta-learning. arXiv preprint arXiv:2211.15180, 2022. 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[16] Kevin Eykholt, Ivan Evtimov, Earlence Fernandes, Bo Li, Amir Rahmati, Chaowei Xiao, Atul Prakash, Tadayoshi" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Kohno, and Dawn Song. Robust physical-world attacks on deep learning visual classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1625-1634, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Ali Farhadi, Ian Endres, Derek Hoiem, and David A. Forsyth. Describing objects by their attributes. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 1778-1785, 2009. 6, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[19] Micah Goldblum et al. Adversarily robust few-shot learning: A meta-learning approach. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 3, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 262 + ], + "type": "text", + "content": "[20] Ian J. Goodfellow et al. Explaining and harnessing adversarial examples. In Int. Conf. Learn. Represent. (ICLR), 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 286 + ], + "type": "text", + "content": "[21] Gregory Griffin, Alex Holub, and Pietro Perona. Caltech-256 object category dataset. 2007. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 320 + ], + "type": "text", + "content": "[22] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Zero-shot detection via vision and language knowledge distillation. In Int. Conf. Learn. Represent. (ICLR), 2021. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 364 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 770-778, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 421 + ], + "type": "text", + "content": "[24] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Int. Conf. Mach. Learn. (ICML), pages 4904-4916, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 455 + ], + "type": "text", + "content": "[25] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. California Institute of Technology, 2009. 3, 6, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[26] Huafeng Kuang, Hong Liu, Yongjian Wu, and Rongrong Ji. Semantically consistent visual representation for adversarial robustness. IEEE Transactions on Information Forensics and Security, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[27] Christoph H. Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 36(3):453-465, 2014. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 588 + ], + "type": "text", + "content": "[28] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In Int. Conf. Learn. Represent. (ICLR), 2022. 2, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[29] Xiao Li, Hang Chen, and Xiaolin Hu. On the importance of backbone to the adversarial robustness of object detectors. arXiv preprint arXiv:2305.17438, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[30] Xiao Li, Ziqi Wang, Bo Zhang, Fuchun Sun, and Xiaolin Hu. Recognizing object by components with human prior knowledge enhances adversarial robustness of deep neural networks. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: common objects in" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24694" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "context. In Eur. Conf. Comput. Vis. (ECCV), pages 740-755, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[32] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Int. Conf. Learn. Represent. (ICLR), 2018. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[33] Chengzhi Mao, Ziyuan Zhong, Junfeng Yang, Carl Vondrick, and Baishakhi Ray. Metric learning for adversarial robustness. Adv. Neural Inform. Process. Syst. (NeurIPS), 32, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 229 + ], + "type": "text", + "content": "[34] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. Int. Conf. Learn. Represent. (ICLR), 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 263 + ], + "type": "text", + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 308 + ], + "type": "text", + "content": "[36] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In Int. Conf. Learn. Represent. (ICLR), 2020. 1, 2, 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 287, + 354 + ], + "type": "text", + "content": "[37] Tianyu Pang, Xiao Yang, Yinpeng Dong, Taufik Xu, Jun Zhu, and Hang Su. Boosting adversarial training with hypersphere embedding. In Adv. Neural Inform. Process. Syst. (NeurIPS), 2020. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 388 + ], + "type": "text", + "content": "[38] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3498-3505, 2012. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 389, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 442 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, et al. Learning transferable visual models from natural language supervision. In Int. Conf. Mach. Learn. (ICML), pages 8748-8763, 2021. 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "text", + "content": "[40] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. (IJCV), 115(3):211-252, 2015. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 544 + ], + "type": "text", + "content": "[41] Jake Snell, Kevin Swersky, and Richard S. Zemel. Prototypical networks for few-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 4077-4087, 2017. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[42] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In Int. Conf. Learn. Represent. (ICLR), 2014. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[43] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[44] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching networks for one shot learning. In Adv. Neural Inform. Process. Syst. (NeurlPS), pages 3630-3638, 2016. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[45] Mengmeng Wang et al. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2, 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 597 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[46] Ren Wang, Kaidi Xu, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Chuang Gan, and Meng Wang. On fast adversarial robustness adaptation in model-agnostic meta-learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 3, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[47] Yongqin Xian, Zeynep Akata, Gaurav Sharma, Quynh Nguyen, Matthias Hein, and Bernt Schiele. Latent embeddings for zero-shot classification. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 69-77, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[48] Yongqin Xian, Christoph H. Lampert, Bernt Schiele, and Zeynep Akata. Zero-shot learning - A comprehensive evaluation of the good, the bad and the ugly. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 41(9):2251-2265, 2019. 6, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[49] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 3485-3492, 2010. 7, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 306 + ], + "type": "text", + "content": "[50] Yunlong Yu, Zhong Ji, Yanwei Fu, Jichang Guo, Yanwei Pang, and Zhongfei (Mark) Zhang. Stacked semantics-guided attention model for fine-grained zero-shot learning. In Adv. Neural Inform. Process. Syst. (NeurIPS), pages 5998-6007, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 341 + ], + "type": "text", + "content": "[51] Mehmet Kerim Yucel et al. How robust are discriminatively trained zero-shot learning models? Image Vis. Comput., 119: 104392, 2022. 1, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 396 + ], + "type": "text", + "content": "[52] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric P. Xing, Laurent El Ghaoui, and Michael I. Jordan. Theoretically principled trade-off between robustness and accuracy. In Int. Conf. Mach. Learn. (ICML), pages 7472-7482, 2019. 1, 2, 4, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 398, + 545, + 430 + ], + "type": "text", + "content": "[53] Jiaming Zhang, Qi Yi, and Jitao Sang. Towards adversarial attack on vision-language pre-training models. In ACM Int. Conf. Multimedia, pages 5005-5013, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 432, + 545, + 475 + ], + "type": "text", + "content": "[54] Manli Zhang, Jianhong Zhang, Zhiwu Lu, Tao Xiang, Mingyu Ding, and Songfang Huang. IEPT: instance-level and episode-level pretext tasks for few-shot learning. In Int. Conf. Learn. Represent. (ICLR), 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 477, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 520 + ], + "type": "text", + "content": "[55] Xingxing Zhang, Shupeng Gui, Zhenfeng Zhu, Yao Zhao, and Ji Liu. ATZSL: defensive zero-shot recognition in the presence of adversaries. arXiv preprint arXiv:1910.10994, 2019. 1, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 554 + ], + "type": "text", + "content": "[56] Mo Zhou and Vishal M Patel. Enhancing adversarial robustness for deep metric learning. In IEEE Conf. Comput. Vis. Pattern Recog. (CVPR), pages 15325-15334, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 597 + ], + "type": "text", + "content": "[57] Xiaopei Zhu, Xiao Li, Jianmin Li, Zheyao Wang, and Xiaolin Hu. Fooling thermal infrared pedestrian detectors in real world using small bulbs. In AAAI, pages 3616-3624, 2021. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24695" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_content_list.json b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e6b326c4d4ce37af1728eea8bf585ace45a9d56f --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_content_list.json @@ -0,0 +1,1851 @@ +[ + { + "type": "text", + "text": "Language-aware Visual Semantic Distillation for Video Question Answering", + "text_level": 1, + "bbox": [ + 101, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bo Zou *", + "bbox": [ + 209, + 156, + 289, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University Beijing, China", + "bbox": [ + 168, + 174, + 331, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zoub21@mails.tsinghua.edu.cn", + "bbox": [ + 125, + 212, + 375, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Yang *", + "bbox": [ + 459, + 156, + 566, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai AI Laboratory Shanghai, China", + "bbox": [ + 416, + 174, + 609, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yangchao@pjlab.org.cn", + "bbox": [ + 418, + 212, + 607, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Qiao", + "bbox": [ + 715, + 156, + 782, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanghai AI Laboratory Shanghai, China", + "bbox": [ + 651, + 174, + 844, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qiaoyu@pjlab.org.cn", + "bbox": [ + 663, + 212, + 833, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengbin Quan", + "bbox": [ + 267, + 234, + 395, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University", + "bbox": [ + 251, + 253, + 411, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing, China", + "bbox": [ + 272, + 270, + 390, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quancb@tsinghua.edu.cn", + "bbox": [ + 233, + 290, + 429, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youjian Zhao†", + "bbox": [ + 550, + 234, + 674, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tsinghua University", + "bbox": [ + 532, + 253, + 694, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongguancun Laboratory", + "bbox": [ + 506, + 270, + 718, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beijing, China", + "bbox": [ + 555, + 287, + 671, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhaoyoujian@tsinghua.edu.cn", + "bbox": [ + 491, + 308, + 730, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 361, + 312, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Significant progress in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although image-language models can efficiently represent both video and language branches, they typically employ goal-free vision perception and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure extracting multi-grained visual semantics associated with the questions. We conduct evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset.", + "bbox": [ + 76, + 393, + 473, + 848 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d90d72df9833dde436ea53e7778a2f58cb4e42c69df533c70fbf3be776f4b073.jpg", + "image_caption": [ + "(a)", + "Answer: Car radio" + ], + "image_footnote": [], + "bbox": [ + 524, + 354, + 869, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ae326244236e0ac116abb45449c8ee4044bac881b2673645ff8ebcd0f9a77a9.jpg", + "image_caption": [ + "Question: The man is singing along with what object?", + "Question: What object is taken out from the cabinet?", + "Answer: Cup" + ], + "image_footnote": [], + "bbox": [ + 522, + 405, + 869, + 454 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ff03425bcc0e91988c86e4cca1be3106d4dff9a6ce32a827e02fd01f04f3165.jpg", + "image_caption": [ + "(b)", + "(c)", + "Question: The person put the fish in the bowl into what object on the table? Answer: Tank" + ], + "image_footnote": [], + "bbox": [ + 522, + 464, + 867, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c73333246d7aaaccab8845a932cced907171b2f90b6dc4913234b4565c4e9ce.jpg", + "image_caption": [ + "(d)", + "Question: What color is the banana in the photo?", + "Figure 1. Challenges of goal-free VideoQA models. They can not efficiently handle (a) Long-term dependencies, (b) Multi-events, and (c) Multi-scale semantics in the videos. They also suffer from (d) language prior phenomenon in training question-answer pairs." + ], + "image_footnote": [], + "bbox": [ + 524, + 516, + 867, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 689, + 632, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, large-scale video-and-language pretraining has seen remarkable progress. Most modern video-language understanding models [4, 15, 26, 31, 35, 36, 50] independently encode uni-modality and then fuse them. Human motion perception studies [5, 17, 53] suggest that humans perceive motion and environments as goal-driven behavior. This discrepancy leads to several issues, especially in long-form video understanding.", + "bbox": [ + 496, + 715, + 895, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "First, goal-free video representation struggles with long-term dependencies and multi-event reasoning. While enabled by progress in cross-modal pretraining[29, 43, 46] and datasets[39, 40], current methods excel at question an", + "bbox": [ + 496, + 839, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 20 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution, † Corresponding author", + "bbox": [ + 99, + 875, + 344, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was done during an internship at Shanghai AI Lab.", + "bbox": [ + 104, + 888, + 428, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27113", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "swering for images and short videos, but long-form videos contain many clips irrelevant or redundant to the questions, interfering with overall understanding. Encoding long videos also brings immense computational costs. Second, accurate semantic reasoning usually relies on multi-scale perception from local-spatial regions to global temporal dynamics. Goal-free methods for multi-scale visual modeling require custom submodels[9, 52, 56] for each scale or extra modalities like bounding boxes and OCR features[23, 66] but large-scale pretraining makes these approaches inefficient or infeasible. Third, goal-free video embedding requires a multimodal fusion module to synthesize questions and visual embeddings to predict answers. Incorporating questions directly can lead to shortcut solutions [1, 2, 8, 12, 19, 25, 41, 65], which means utilizing obvious clues in questions (mainly in data distribution and the relation between keywords) that exhibit more reliability in answer prediction than complicated visual reasoning, especially in the early training phases. This phenomenon is also known as language bias, which often causes major performance gaps in out-of-distribution tests. The above challenges are visualized in Figure 1.", + "bbox": [ + 76, + 90, + 472, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a language-aware (goal-driven) visual semantic distillation framework called VideoDistill. Semantic distillation means the visual encoder must embed relevant frames of questions and focus on question-related multi-scale visual semantics. Closely resembles human behavior, semantic distillation functions like keeping the goal (question) in mind and conducting meaningful and precise visual reasoning. Differing from previous VideoQA frameworks [3, 31, 32, 42, 45, 59, 60], VideoDistill can generate answers only from question-related visual embeddings (without additional textual embeddings) since the interaction of text and vision is adequate during semantic distillation. This feature also enhances the significance of visual reasoning when generating answers.", + "bbox": [ + 76, + 411, + 472, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To realize semantic distillation, we first introduce Language-Aware Gate (LA-Gate), a multi-head cross-gating mechanism for cross-model interaction, which is inspired by self-gating and gated attention[14, 22, 37, 51] and is integrable into any existing transformer-based models like[4, 50]. LA-Gates compute questions' dependencies on video patch embeddings and depress or excite corresponding patches in subsequent attention layers. The proposed LA-Gate is not only the key to semantic distillation but also a new efficient modality fusion method, which is a powerful competitor of predominant Cross-Attention in VideoQA. We discuss the differences between LA-Gate and attention mechanisms in Section 3.2 and summarize three merits of LA-Gate: (1) It can alleviate the influence of language bias by avoiding the direct involvement of text embeddings. (2) It can better maintain local diversity within the video embeddings. (3) It enhances the interpretability of the modality fusion process.", + "bbox": [ + 76, + 626, + 472, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VideoDistill has two LA-Gate-based modules. The first", + "bbox": [ + 96, + 885, + 470, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is a differentiable sparse sampling module, which uses pretrained image-language models like CLIP[43] to encode frames, then performs goal-driven frame sampling to remarkably reduce subsequent spatial-temporal attention overhead and naturally avoid long-term dependencies and multi-event reasoning by retaining only question-related frames. It also provides our framework with a good nature that is insensitive to the number of sampled frames (see section 4.3).", + "bbox": [ + 496, + 90, + 893, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second is a vision refinement module eliminating unrelated visual semantics at different perceptual levels and enhancing related multi-scale semantics to support multilevel refinement. It encodes sparsely sampled frames into question-related global embeddings for generating answers. This module brings obvious performance boosts, especially on object-related questions (see section 4.5). Our contributions are summarized as follows:", + "bbox": [ + 496, + 212, + 893, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the Language-Aware Gate (LA-Gate) that enables interacting vision with language meanwhile not directly introducing language into visual representations. It can alleviate the influence of language bias, better maintain local diversity within the video embeddings, and is more interpretable compared with predominant Cross-Attention.", + "- Based on LA-Gate, we propose a differentiable sparse sampling module to capture question-related frames and a vision refinement module to emphasize multi-scale question-related semantics. They can benefit VideoQA models on the stability under various numbers of sampled frames and the capacity to understand multi-scale objects.", + "- Our model achieves new state-of-the-art performance on a wide range of downstream VideoQA tasks and text-to-video retrieval tasks." + ], + "bbox": [ + 500, + 335, + 893, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 566, + 650, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Long-form Video modeling", + "text_level": 1, + "bbox": [ + 500, + 590, + 743, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-form understanding has recently become a new research hot spot [11, 16, 31, 34, 47, 48, 54, 55, 59, 64]. Current methods develop in two directions to overcome unique challenges not encountered by previous short-term video understanding works. The first direction is to encode long-term dependencies better. LF-VILA [48] proposes hierarchical temporal window attention that begins with a small window learning the attention between adjacent frames, then gradually expands the window size to learn high-level representation. HERO [34] and MERLOT [64] predict the order of shuffled frames to understand sequential characteristics. The second direction aims to reduce the increasing computation cost of encoding longer videos. Since the common practice of randomly selecting 3 or 4 frames per video regardless of length is not suited for long videos, ClipBert [31] randomly samples a sequence of segments from a video and then aggregates their predictions. MIST [16] improves ClipBert by selecting both video segments and frame regions to be encoded and reduces computation costs further.", + "bbox": [ + 496, + 613, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27114", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/626fbdfe04d7fdff432f14ca2aa8decdc8648fb266b97ed7d1b7e129e299200e.jpg", + "image_caption": [ + "Figure 2. Overview of VideoDistill. VideoDistill first densely samples video frames and utilizes a pre-trained image-language encoder to extract features, then sparsely samples a small number of question-related frames by a differentiable sparse sampling module. Finally, VideoDistill uses a vision refinement module to emphasize necessary multi-scale visual semantics in selected frames." + ], + "image_footnote": [], + "bbox": [ + 138, + 87, + 831, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TALLFormer [11] proposes short-term feature extraction and long-term memory mechanisms that avoid repetitive calculations during training.", + "bbox": [ + 75, + 446, + 470, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to the second category, we reduce the workload by decreasing inputs, but our goal-driven differentiable sparse sampling works directly on frames, the minimum composition of the video, rather than rough video segments. Kim et al. [27] propose to use a self-gating mechanism to realize frame sampling, but it is a goal-free sampling method. Also, unlike random selection [31] or simple selector based on similarity[16], our learnable sampler excavates the semantic relationships between visual candidates. In addition, sparsely sampled frames from our sampler can be considered a concise summary of the long video sequence. Therefore, the following inference can perform similarly to models on short-form inputs. Thus, VideoDistill does not make efforts in the first direction.", + "bbox": [ + 75, + 494, + 473, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Video-Language Fusion in VideoQA", + "text_level": 1, + "bbox": [ + 76, + 723, + 393, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The answer generation of VideoQA can be formulated as the interaction of the two modalities. Some methods propose fusing language and vision in the feature extraction stage for a more interactive combination. PSAC [36] combines regular visual self-attention with visual-linguistic co-attention. HCRN [30] supports conditioning video features with linguistic clues as a context factor in every inference stage. PMT [42] adopts a pyramidal video-language interaction. MCAN[63] utilizes Guided-Attention (V2T cross-attention in this paper) to fuse language into visual feature extraction.", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike these techniques, our VideoDistill only keeps the question \"in mind.\" We do not directly fuse the question information into the video embedding, thus generating \"purer\" visual embeddings. Since the answer prediction is only based on visual embeddings, VideoDistill can alleviate just assuming the answer from the question and follow the \"look and answer\" criterion [2].", + "bbox": [ + 496, + 446, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides VideoDistill, some existing methods can also be considered as indirect fusion. Contrastive learning methods like [10, 18, 20, 43, 57, 67] pull the matched pairs closer to interact vision with language. However, they do not demonstrate the same level of performance in tasks other than retrieval [28] due to ignoring the misalignment of vision and language. The method of fine-grained contrast to mitigate the misalignment has become a recent research hot spot.", + "bbox": [ + 496, + 553, + 893, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 689, + 635, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce VideoDistill, a pretraining framework that enables goal-driven VideoQA relying on question-related frames and their multi-scale semantics. Figure 2 gives an overview of VideoDistill. VideoDistill consists of two sub-modules: Differentiable Sparse Sampling and Vision Refinement. Both sub-modules are built upon our proposed language-aware gate (LA-Gate) to perform in a goal-driven manner.", + "bbox": [ + 496, + 715, + 893, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Vision-Language Representations", + "text_level": 1, + "bbox": [ + 500, + 845, + 794, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VideoDistill begins with densely sampling $N$ uniformly distributed frames and embedding them into frame represent", + "bbox": [ + 496, + 869, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27115", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ca32b2d5e3347226dfc993cccb89c381cf492851e87fe4e9cf0b582db00a2fa8.jpg", + "image_caption": [ + "Figure 3. Illustrations of Self-Attention, Cross Attention, and our LA-Gate mechanisms" + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 357, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/659c9c9cce9b2386df7b74c2cb763293a40f6774e57ab5fc58391923618fc584.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 612, + 237 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d94d1263b62a166d54a368a8b7123f1b16aeaf63cf6659859fd4922d8510da66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 89, + 851, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tations. We adopt the pretrained image-language encoder CLIP [43] with frozen parameters in the visual branch to reduce the computational overhead. CLIP divides $N$ frames into $N \\times n \\times n$ patches, and extracts patch embeddings $v_{\\mathrm{patch}} \\in \\mathbb{R}^{N \\times n^2 \\times D}$ and CLS tokens $v_{\\mathrm{cls}} \\in \\mathbb{R}^{N \\times D}$ that represent the global understanding of frames, where $D$ is the dimension of representations.", + "bbox": [ + 75, + 276, + 470, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the Language branch, A text is first tokenized and then fed into the pretrained text encoder of CLIP to generate word-level embeddings $t_{\\mathrm{word}} \\in \\mathbb{R}^{M \\times D}$ and a sentence-level token $t_{\\mathrm{cls}} \\in \\mathbb{R}^D$ , where M is the length of embeddings. Note that parameters in the text encoder are updated during pretraining and downstream finetuning to mitigate the linguistic gaps between datasets.", + "bbox": [ + 75, + 383, + 468, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Language-Aware Gate", + "text_level": 1, + "bbox": [ + 76, + 500, + 285, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a reusable unit, termed language-aware gate (LA-Gate), which takes an arbitrary visual representation $v$ and a sentence-level language representation $t_{\\mathrm{cls}}$ as input. LA-Gate generates $v_{\\mathrm{distill}}$ , an output sequence of language-aware visual representation with the same dimension as $v$ , by exciting or depressing the components of $v$ according to their similarities with $t_{\\mathrm{cls}}$ . For a better understanding of LA-Gate, we illustrate self-attention, cross-attention, and LA-Gate in Figure 3. We first discuss why commonly used cross-attention is not suitable for our purposes, then point out the design concept and benefits of LA-Gate.", + "bbox": [ + 75, + 522, + 468, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3(b) illustrates two types of cross-attention: Text-to-Video (T2V, text as Query) and Video-to-Text (V2T, video as Query). Since outputs of the attention have the same length as Query and carry information sourced from Value, applying cross-attention in VideoQA will have two problems. Firstly, T2V is unsuitable for the modality interaction in the middle layers as it alters the shape of visual representations. Although V2T can maintain the shape, its output will carry the info directly from the text (like reconstructing text representations into the shape of visual input).", + "bbox": [ + 75, + 688, + 470, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The structure of LA-Gate is shown in Figure 3(c). There are two divergences between V2T and LA-Gate. (1) When text input length $L > 1$ , we sum the similarity matrix to form a vector of importance for rows of the video input.", + "bbox": [ + 75, + 840, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L$ functions like the number of multi-heads in the attention mechanism, and each text representation will partially dictate the importance of each row in the video input. (Note that in VideoDistill, we always have $L = 1$ since the text input is $t_{\\mathrm{cls}}$ .) (2) We expand the importance vector (repeat D times) to match the shape of the video input. Finally, we apply an element-wise product rather than a dot product on the importance matrix and the Value.", + "bbox": [ + 496, + 276, + 890, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first merit of LA-Gate is the text input only controls the importance of each visual representation, and the output does not directly involve text. Thus, the answer decoder can make decisions based on \"purer\" visual semantics and alleviate shortcut solutions hidden in texts.", + "bbox": [ + 496, + 398, + 892, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A more significant characteristic of LA-Gate in general modality fusion usage: it can better maintain local diversity within the visual input. In Figure 3 (a) and (b), the output displays rows in mixed colors, representing the weighted sum of rows in the Value. Cross-attention will gather information from all rows in the Value for each output token and present more global attributes than LA-Gate since LA-Gate does not blend all features. In Figure 4(c), we maintain the uniqueness of colors for LA-Gate in each output row because each row is produced by multiplying the corresponding row in the Value and a scalar of importance. This characteristic makes LA-Gate fit more into our idea of information distillation because the language-aware distillation of a local area should be irrelevant to other regions.", + "bbox": [ + 496, + 477, + 893, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The last benefit of LA-Gate is we can enhance the interpretability of modality fusion by incorporating LA-Gate with self-attention layers. Due to the skip connection between the input and the output, commonly used V2T cross-attention brutally adds a reconstructed text (maybe linearly projected) into the visual input to fuse vision and language. The physical meaning of the structure is difficult to explain. By contrast, LA-Gate first distills the visual input by emphasizing language-related visual semantics, and then the interaction within visual representations is performed by self-attention.", + "bbox": [ + 496, + 691, + 893, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The implementation of LA-Gate is visualized in Figure 4 (we demonstrate how LA-Gate works in a vision refinement block, where $v$ consists of frame patches from $K$ selected frames and a video-level CLS token). Assume LA-Gate", + "bbox": [ + 496, + 844, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27116", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "receives $v \\in \\mathbb{R}^{m \\times D}$ , where $m$ is the sequence length, and $t_{\\mathrm{cls}} \\in \\mathbb{R}^D$ . We first produce the key from the language representation $t_{\\mathrm{cls}}$ and queries and values from the visual representation $v$ . Then, we calculate cosine distances between each query and the key:", + "bbox": [ + 75, + 90, + 470, + 161 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {k e y} = \\mathrm {w} _ {\\mathrm {k}} \\left(t _ {\\mathrm {c l s}}\\right), \\text {q u e r y} _ {i} = \\mathrm {w} _ {\\mathrm {q}} \\left(v _ {i}\\right), \\text {v a l u e} _ {i} = \\mathrm {w} _ {\\mathrm {v}} \\left(v _ {i}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 169, + 470, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {d i s t} _ {i} = \\operatorname {r e p e a t} \\left(\\frac {\\operatorname {q u e r y} _ {i} \\cdot \\operatorname {k e y}}{\\| \\operatorname {q u e r y} _ {i} \\| \\times \\| \\operatorname {k e y} \\|}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 202, + 470, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i} \\in \\mathbb{R}^{D}, i \\in [1, m]$ is a feature vector in $v$ . $\\mathrm{w_k}$ , $\\mathrm{w_q}$ and $\\mathrm{w_v}$ are trainable linear projection layers, and repeat means expanding a scalar into a vector with the dimension of $D$ . Since $\\mathrm{dist}_i$ reflects the correlation between vision and language, we treat it as an importance coefficient of $\\mathrm{value}_i$ . Then, we generate distilled $v_{i}$ as formulated:", + "bbox": [ + 75, + 236, + 470, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {\\text {d i s t i l l}} ^ {i} = \\mathrm {w} _ {\\mathrm {o}} \\left(\\operatorname {d i s t} _ {i} \\odot \\operatorname {v a l u e} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 328, + 470, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise production and $\\mathrm{w_o}$ is the linear output layer. In practice, we perform a standard multihead attention setting [49] on $\\mathrm{w_k}$ , $\\mathrm{w_q}$ , $\\mathrm{w_v}$ and $\\mathrm{W_o}$ , and concatenate outputs of each head to form $v_{\\mathrm{distill}}^i$ . Although, we only illustrate a single attention head in Figure 2 for simplicity. Also, we apply an input skip connection on $v_{\\mathrm{distill}}^i$ for better training stability and convergence. Finally, LA-Gate outputs $v_{\\mathrm{distill}} = \\left\\{v_{\\mathrm{distill}}^i | i \\in [1, m]\\right\\}$ .", + "bbox": [ + 75, + 352, + 470, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Differentiable Sparse Sampling", + "text_level": 1, + "bbox": [ + 76, + 496, + 354, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $N$ densely sampled frames, VideoDistill further adaptively picks out $K(K < N)$ language-related frames. To this end, we utilize stacked frame sampling blocks (FS-Blocks, $L$ layers) which take as input $v_{\\mathrm{cls}}$ and $t_\\mathrm{cls}$ to perform a top- $K$ selection. Since $v_{\\mathrm{cls}}$ of frames are separately extracted, to indicate their temporal positions in the whole video, we add temporal embedding $\\mathrm{T_f} \\in \\{\\phi_{\\mathrm{T}}(\\mathrm{f}) | \\mathrm{f} \\in [1,\\mathrm{N}]\\}$ for each of them according to their frame index. After adding temporal information, each FS-Block first distills $v_{\\mathrm{cls}}$ conditioned on $t_\\mathrm{cls}$ by LA-Gate to emphasize question-related frames, then performs an interframe interaction through standard multi-head attention. Figure 4 shows the architecture of FS-Blocks. We borrow the form of skip connections from the divided block in [4].", + "bbox": [ + 75, + 520, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To realize the differentiable selection, we first project $v_{\\mathrm{cls}}^{L}$ , the output of the $L$ -th FS-Block with the dimension identical to $v_{\\mathrm{cls}}$ , onto a feature space with the dimension of $K$ by a linear layer $W_{s}$ , where $K$ is the number of frames to be selected. Then, we conduct Gumbel-Softmax sampling [24] on each $x^{k} \\in \\mathbb{R}^{N}, k \\in [1, K]$ , the row of $\\mathrm{W_s}\\left(v_{\\mathrm{cls}}^L\\right)$ . The procedures are formulated as follows:", + "bbox": [ + 75, + 729, + 470, + 828 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {\\mathrm {s o f t}} ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {x ^ {k} + g u m b e l s}{\\tau}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 835, + 470, + 867 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3ab6d664f307c3dff80b4e8eb181ad725bb754d880c0e5f60f1f54ee3b229a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 852, + 231 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f3cdc518f4e623ccb44883fecb4fac28123a40d223674c4f08d7be1361939db6.jpg", + "image_caption": [ + "Figure 4. Architectures of Language-Aware Gate, Frame Sampling Block, and Vision Refinement Block." + ], + "image_footnote": [], + "bbox": [ + 540, + 232, + 691, + 371 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0f2f9654508f02079edec81448f0484474b342476edab5d6188e734956e5484.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 232, + 851, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {\\text {h a r d}} ^ {k} = \\operatorname {o n e h o t} \\left(\\arg \\max \\left(y _ {\\text {s o f t}} ^ {k}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 419, + 890, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm a s k ^ {k} = y _ {\\text {h a r d}} ^ {k} + y _ {\\text {s o f t}} ^ {k} - \\operatorname {s t o p g r a d} \\left(y _ {\\text {s o f t}} ^ {k}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 448, + 890, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where gumbels is a noise sampled from the Gumbel distribution with $\\mu = 0$ and $\\beta = 1$ , $y_{\\mathrm{soft}}^k$ is the possibility of frames to be selected as the $k$ -th language-related frame and $y_{\\mathrm{hard}}^k$ reflects the index of the chosen frame. Since the argmax operation has no gradient, we adopt a code-level trick in equation 6 to generate mask $= \\{mask^{k}|k\\in [1,K]\\}$ which can properly pass the gradient. Finally, we apply mask on $v_{\\mathrm{patch}}$ and generate $v_{\\mathrm{patch}}^K\\in \\mathbb{R}^{K\\times n^2\\times D}$ for the following vision refinement.", + "bbox": [ + 496, + 473, + 892, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Vision Refinement", + "text_level": 1, + "bbox": [ + 500, + 619, + 676, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the selected patch embeddings $v_{\\mathrm{patch}}^{K}$ and the language representation $t_{\\mathrm{cls}}$ , this module generates video-level representation $v_{\\mathrm{cls}}^{*}$ , which synthesizes multi-scale language-related visual semantics. $v_{\\mathrm{cls}}^{*}$ is set as a learnable token with the dimension of $D$ and is concatenated to $v_{\\mathrm{patch}}^{K}$ as the visual input $v_{*}$ . We combine the LA-Gate and any existing spatial-temporal self-attention layer [4, 6, 50] to form a vision refinement block as shown in Figure 4. When iteratively applying vision refinement blocks on $v_{*}$ , LA-Gates spontaneously distill language-related visual semantics in different perceptive fields. Since the effective perceptive field gradually expands in stacked blocks [44], our module can consider multi-scale objects.", + "bbox": [ + 496, + 642, + 893, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike predominant pyramidal models [33, 42, 61] for multi-scale reasoning that gathers the intermediate results of each encoding stage as final representation, we empirically adopt only the $v_{\\mathrm{cls}}^*$ from the last vision refinement block.", + "bbox": [ + 496, + 839, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "The code will be available at https://zoubo9034.github.io/VideoDistill/", + "bbox": [ + 76, + 875, + 470, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27117", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Pretraining Tasks", + "text_level": 1, + "bbox": [ + 76, + 90, + 250, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the two pretraining tasks to facilitate cross-modal interaction: Video-Text Matching (VTM) and Vision-Guided Masked Language Modeling (VG-MLM).", + "bbox": [ + 75, + 114, + 468, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video-Text Matching. To facilitate cross-modal interaction, we use VTM as the first pre-training task. VTM predicts whether input video-language pairs are matched. In practice, we randomly exchange the annotations of two input pairs (matched) in a mini-batch with a probability of 0.5. Then, we use a linear projection head on the top of $v_{\\mathrm{cls}}^*$ to predict two-class matching logits $y$ , and formulate the VTM loss as negative log-likelihood:", + "bbox": [ + 75, + 159, + 470, + 270 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V T M}} = - \\mathbb {E} _ {\\left(v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p (y \\mid v _ {\\mathrm {c l s}} ^ {*}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 282, + 470, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We do not adopt predominant contrastive pretraining because this paradigm requires the contrast between all possible combinations of vision and language. However, for VideoDistill, the number of possible visual embeddings of each video is proportional to the number of annotations. This leads to a quadratic growth in computational overhead (see Appendix A for more details). Nevertheless, we still apply a contrastive constraint on matched pairs (w/o exchange) to stabilize the training:", + "bbox": [ + 75, + 311, + 472, + 438 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C L}} = - \\sum_ {i \\in B} \\log \\frac {\\mathbb {I} (\\text {m a t c h e d}) \\exp (s (v _ {i} , t _ {i}) / \\tau)}{\\sum_ {j \\in B} \\exp (s (v _ {i} , t _ {j}) / \\tau)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 446, + 470, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where B is the batch size, $v_{i}$ and $t_i$ are $v_{\\mathrm{cls}}^*$ and $t_\\mathrm{cls}$ of $i$ -th video-language pair, $\\mathbb{I}$ is a indicate function of whether the video and the annotation in the $i$ -th pair are matched, $s$ denotes a cosine similarity function, $\\tau$ is a temperature coefficient equals 0.07 in this paper.", + "bbox": [ + 75, + 491, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision-Guided Masked Language Modeling. The commonly used MLM [13] task aims to predict the ground truth label of the masked token according to textual context. To better map the visual and the language representations in a fine-grained manner, we propose VG-MLM, which is applied only on the matched pairs (w/o exchange) and encourages predicting the masked tokens from the visual context. In particular, we first provide VideoDistill the CLS tokens $t_{\\mathrm{cls}}$ of unmasked annotations to sample frames and extract video representation $v_{\\mathrm{cls}}^*$ . Then, we mask $I$ words in the annotations and encode the masked sentences. The outcome $w_i^{\\mathrm{M}}, i \\in [1, I]$ , which denotes the token of $i$ -th masked word, and $v_{\\mathrm{cls}}^*$ are used to predict the $i$ -th masked word. The objective of VG-MLM is formulated as follows:", + "bbox": [ + 75, + 566, + 470, + 762 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V G - M L M}} = - \\mathbb {E} _ {\\left(t _ {\\mathrm {c l s}}, v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p \\left(w _ {i} | \\mathrm {S G} \\left(w _ {i} ^ {\\mathrm {M}}\\right), v _ {\\mathrm {c l s}} ^ {*}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 768, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where SG means stop gradient, $w_{i}$ denote the logit of $i$ -th masked word. We stop the gradient of $w_{i}^{\\mathrm{M}}$ to enhance the importance of visual clues in language modeling. Following the heuristics of Bert [13], we adopt the same masking strategy. Then we adopt a two-layer MLP on top of the combination of $w_{i}^{\\mathrm{M}}$ and $v_{\\mathrm{cls}}^{*}$ to generate the probability over the vocabulary, which is calculated as the negative log-likelihood loss", + "bbox": [ + 75, + 801, + 472, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the masked word. The overall loss of pretraining is the combination of VTM, VG-MLM, and CL:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {V T M}} + \\mathcal {L} _ {\\mathrm {V G - M L M}} + \\mathcal {L} _ {\\mathrm {C L}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 130, + 890, + 143 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 159, + 633, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, conduct extensive experiments utilizing the pre-trained model on downstream various types of VideoQA tasks to verify the effectiveness of our proposed VideoDistill. We also transfer VideoDistill for video-text retrieval tasks to show the generalization power of the pre-trained model in Appendix C. Our pretraining set consists of three parts: (1) 3M video-caption pairs randomly sampled from generic dataset WebVid10M [4]. (2) 4.2M video-caption pairs randomly sampled from YouTube video dataset HD-VILA [59]. (3) 3.8M video-caption pairs from 1-st person view dataset EgoCLIP. We report results with $N = 100$ , $K = 16$ , and $L = 3$ . For open-ended datasets, train a MLP classification head on the top of $v_{\\mathrm{cls}}^*$ . For multiple-choice datasets, we choose the answer with the maximum logit over the VTM head. More details are in Appendix B.", + "bbox": [ + 496, + 184, + 893, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Generic VideoQA:", + "text_level": 1, + "bbox": [ + 498, + 421, + 679, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate VideoDistill on the four commonly used VideoQA datasets: MSRVTT-QA [58], MSVD-QA [58], EgoMCQ [38] and MSRVTT-multiple-choice test [62]. Details of each dataset and results are in the supplementary material (Appendix C).", + "bbox": [ + 496, + 444, + 893, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Long-Form VideoQA:", + "text_level": 1, + "bbox": [ + 498, + 530, + 705, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our model on two recently proposed challenging datasets for the long-form VideoQA, namely EgoTaskQA[25] and AGQA[21]. We adopt the same finetuning setting with MSRVTT-QA and MSVD-QA. For fairness comparison, we report the results of VideoDistill w/o large-scale pertaining. Details of each dataset are in the supplementary material.", + "bbox": [ + 496, + 553, + 893, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. In the Table 2, our VideoDistill overwhelm previous SOTA method and bring $2.82\\%$ overall performance gains. Remarkably, in the categories of multi-agent and descriptive, we outperform previous work with $5.59\\%$ and $3.72\\%$ improvement. They show VideoDistill can better understand the storyline and temporal relations in long videos. In the category of object, we achieve $9.33\\%$ gains, which indicates VideoDistill can capture multi-scale objects. The limited performance in the category of action is due to the strong correlations between action, object, and their change. Most methods tend to over-fit this strong language bias without thorough task understanding. Although the design of VideoDistill avoids utilizing the shortcut solutions in language. In Table 1, The results in the gray color (AIO, Temp, and MIST-CLIP) are reported by [16]. If they follow the same evaluation metrics [21] as us, one of the reasons for the", + "bbox": [ + 496, + 659, + 895, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27118", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b48d074f635ae7ac0307215976dba48f381e64f7b8330ca5f5cb6a2d25602e02.jpg", + "table_caption": [ + "Table 1. QA accuracies of state-of-the-art methods on AGQA v2 test set." + ], + "table_footnote": [], + "table_body": "
Question TypesMost LikelyPSACHMEHCRNAIO [50]Temp[ATP] [7]MIST-CLIPVideoDistill†
Object-relation9.3937.8437.4240.3348.3450.1551.6856.32
Relation-action50.0049.9549.9049.8648.9949.7667.1855.09
Object-action50.0050.0049.9749.8549.6646.2568.9955.12
Superlative21.0133.2033.2133.5537.5339.7842.0543.30
Sequencing49.7849.7849.7749.7049.6148.2567.2454.49
Exists50.0049.9449.9650.0150.8151.7960.3355.74
Duration comparison24.2745.2147.0343.8445.3649.5954.6249.08
Activity recognition5.524.145.435.5218.9718.9619.6910.16
All10.9940.1839.8942.1148.5949.7954.3955.80
", + "bbox": [ + 156, + 102, + 808, + 223 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5719547475562860f1dc3147fec3cdc8a7f01d08aecb1a79c78796848bab638d.jpg", + "table_caption": [ + "Table 2. Performances on EgoTaskQA normal split. †denotes training from scratch." + ], + "table_footnote": [], + "table_body": "
CategoryVisualBERT [35]PSAC [36]HME [15]HGA [26]HCRN [30]ClipBERT [31]VideoDistill†
Scopeworld39.7340.7641.9138.8244.2742.1547.32
intent44.5146.1948.9242.1249.7740.9452.53
multi-agent26.2930.5927.9823.4331.3627.6336.95
Typedescriptive41.9940.6341.4538.0443.4838.4547.20
predictive30.3731.9835.8825.5736.5631.5040.43
counterfactual41.9941.8944.1341.9448.0046.7549.64
explanatory37.4237.9938.8535.9740.6042.3942.53
Semanticaction15.0214.7514.9915.0814.9222.9116.35
object23.2636.5336.0519.0945.3121.8054.64
state59.2061.8963.4455.6568.2854.3672.37
change68.2765.0568.8768.3867.3866.5871.47
all37.9338.9040.1636.7742.2039.8745.02
", + "bbox": [ + 156, + 262, + 808, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance gaps in some categories is they adopt frames with higher resolution $(448 \\times 448)$ and more powerful encoder (CLIP-ViTB/32). Nevertheless, we still outperform current SOTA methods in Object-relation $(4.64\\%)$ and Superlative $(1.25\\%)$ , which require multi-scale and multi-event reasoning. We also improve the overall performance with $1.41\\%$ gains.", + "bbox": [ + 75, + 460, + 470, + 565 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9301845e814910e72d00c2a5c28328a919818c6893a3a9ea0c016d87bb86b32.jpg", + "image_caption": [ + "Figure 5. The impact of the number of frames." + ], + "image_footnote": [], + "bbox": [ + 107, + 578, + 421, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. The Impact of Differentiable Sparse Sampling", + "text_level": 1, + "bbox": [ + 75, + 801, + 467, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Insensitive to the number of frames. We evaluate the performances on MSRVTT-QA with variable input frames. In Figure 5, we compare the performance change of VideoDistill (w/ & w/o sparse sampling module) with PMT [42] (other methods only report results on 16 frames, which are the best", + "bbox": [ + 75, + 824, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results), and there are two findings. First, the differentiable sparse sampling module makes our 4-frame variation on par with other methods with many more input frames. It is intuitive because other competitors require more frames to improve the change to capture necessary information for reasoning (a disadvantage of goal-free perception). This property can dramatically reduce the computational overhead required to achieve the same performance as previous methods. Second, we notice that too many frames will damage the answering accuracy since the emergence of unrelated information disturbs the reasoning process. However, our sparse sampling module can mitigate the dropping with increasing frames and make the model more robust.", + "bbox": [ + 496, + 460, + 893, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We visualize the results of our sparse sampling module on EgoTaskQA in Appendix E.", + "bbox": [ + 498, + 657, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. The Impact of LA-Gate:", + "text_level": 1, + "bbox": [ + 500, + 696, + 723, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prevent the use of language shortcuts. We come up with two methods to quantitatively analyze this characteristic. First, we test the language-only QA performances on EgoTaskQA normal split in Table 3, which means replacing the visual inputs with a static video (We also report the performances under a Gaussian noise input in Appendix E). The more declines, the better. When compared with the language-based model Bert [13], and the previous SOTA method HCRN [30], the decline of our VideoDistill is the most significant (especially in the object category). It demonstrates that VideoDistill relies more on vision and can avoid using language bias. Second, we test VideoDistill on Ego-", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27119", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f49882e5be8a49beb3a153594d32bd20b3dab83c0eb220686dffcfbf77790dfb.jpg", + "table_caption": [ + "Table 3. Language-only QA results on the EgoTaskQA normal split. The more performance drops, the better." + ], + "table_footnote": [], + "table_body": "
CategoryBERTHCRN (w/o vision)VideoDistill (w/o vision)
Acc.ChangeAcc.ChangeAcc.Change
world36.28-8.7%30.17-31.9%18.78-60.3%
intent35.02-21.3%35.54-28.6%22.57-57.0%
multi-agent20.58-21.7%19.9-36.5%11.50-68.7%
descriptive34.55-17.7%29.97-31.1%21.19-55.1%
predictive24.75-18.5%18.32-49.9%7.42-81.6%
counterfactual41.3-1.6%41.1-14.4%39.70-20.0%
explanatory31.78-15.1%32.41-20.2%13.69-67.8%
action15.72+4.6%17.31+16.0%11.81-27.8%
object7.43-68%8.85-80.5%0-100.0%
state45.03-23.9%35.51-48.0%20.21-72.1%
change69.87+2.3%70.47+4.6%55.38-22.5%
", + "bbox": [ + 160, + 104, + 808, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6e185ffa0cd567c3d7c22bf3974d2235f5e5cb3fe35e1a9a7bb2fcf1fea2591.jpg", + "table_caption": [ + "Table 4. Analysis of the effectiveness of each module." + ], + "table_footnote": [], + "table_body": "
MethodModuleEgoTaskQAMSRVVTT-QA
SSVRCross-ATTLA-Gate
(a) BaselineXX--29.1023.92
(b) Uniform SamplingX✓X✓41.7040.92
(c) Soft SamplingX✓X✓38.9538.16
(d) w/o VR✓XX✓33.4426.51
(e) Cross-Attention✓✓✓X40.9441.17
(f) VideoDistill✓✓X✓45.0244.20
", + "bbox": [ + 161, + 273, + 808, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TaskQA indirect split, which is motivated by the fact [25] that during task execution, actions, objects, and their changes are often strongly correlated. It leaves the chance for the model to perform well by simply over-fitting these strong correlations (language bias) without thorough task understanding. The indirect references can avoid these correlations. In Appendix E, we show that our VideoDistill has the least absolute performance change. It indicates that VideoDistill barely utilizes language bias in questions.", + "bbox": [ + 75, + 385, + 470, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A more effective fusion method than Cross-Attention. In Table 4 (e), we replace LA-Gates in our frameworks with Cross-Attention. When comparing (e) with (f), we can find that LA-Gate brings significant performance boosts on both test sets (4.08% and and 3.03%)", + "bbox": [ + 75, + 521, + 468, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. The Impact of Vision Refinement:", + "text_level": 1, + "bbox": [ + 76, + 604, + 372, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance boost in multi-scale object-related questions. The stacked video distillation blocks containing LA-Gate can obtain multi-scale question-related semantics. We observe huge improvements in Table 2 object category $(9.33\\%)$ and Table 1 object-relation category $(4.64\\%)$ .", + "bbox": [ + 75, + 628, + 470, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 710, + 230, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies to verify the effectiveness of components and the rationality of parameter selection.", + "bbox": [ + 75, + 734, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of each component. We design four variants and report their performances in Table 4. (a) we remove the two modules and apply evenly sampled 16 frames as input and a 3-layer transformer to fuse the vision and language. (b) we remove differentiable sparse sampling and apply evenly sampled 16 frames as input. (c) we replace the hard sampling in SS with soft sampling, which means using the weighted sum of dense frames rather than top-k selection. (d) we remove vision refinement and apply a 3-layer transformer", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to fuse the vision and language. (e) we replace LA-Gates in both modules with cross-attention layers.", + "bbox": [ + 496, + 385, + 890, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Ablations We validate the influence of pretraining tasks, the best number of densely sampled frames, and the reasonable number of stacked layers $L$ in Appendix E.", + "bbox": [ + 496, + 416, + 890, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 470, + 624, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we study language-aware VideoQA to overcome the difficulties of long-term dependencies, multi-events, multi-scale semantics, and shortcut solutions in video understanding. In particular, we introduce a differentiable sparse sampling module that naturally avoids complicated long-term dependencies and multi-event reasoning since it only retains question-related frames and a vision refinement module that eliminates unrelated visual semantics in different perception fields and enhances multi-scale related semantics to support a multi-level refinement process. Both modules are based on our LA-Gate, which realizes cross-modal interaction meanwhile avoiding language's direct fusion into visual representations. LA-Gate is a powerful competitor of predominant Cross-Attention in VideoQA and fits our idea of semantic distillation. Therefore, we can realize high-performance VideoQA by just relying on question-related visual representations and reducing language shortcuts. Experiments on 8 VideoQA and text-to-video retrieval tasks demonstrate the effectiveness of our model.", + "bbox": [ + 496, + 493, + 893, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 500, + 794, + 617, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Post-doctoral Excellent Program (Grant No. 2022234).", + "bbox": [ + 496, + 819, + 893, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27120", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2", + "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2, 3", + "[3] Anonymous. LLaMA-excitor: General instruction tuning via indirect feature interaction. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2", + "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6", + "[5] Dare A Baldwin, Jodie A Baird, Megan M Saylor, and M Angela Clark. Infants parse dynamic action. *Child development*, 72(3):708-717, 2001. 1", + "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5", + "[7] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 7", + "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 2", + "[9] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 2", + "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3", + "[11] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with a long-memory transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 503-521. Springer, 2022. 2, 3", + "[12] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1574–1583, 2021. 2", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6, 7", + "[14] Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text comprehension. arXiv preprint arXiv:1606.01549, 2016.2", + "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 1, 7", + "[16] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. arXiv preprint arXiv:2212.09522, 2022. 2, 3, 6", + "[17] György Gergely, Harold Bekkering, and Ildiko Király. Rational imitation in preverbal infants. Nature, 415(6873):755-755, 2002. 1", + "[18] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5006–5015, 2022. 3", + "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2", + "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3", + "[21] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 6", + "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2", + "[23] Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. Iterative answer prediction with pointeraugmented multimodal transformers for textvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9992-10002, 2020. 2", + "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5", + "[25] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. arXiv preprint arXiv:2210.03929, 2022. 2, 6, 8", + "[26] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 1, 7", + "[27] Hyounghun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27121", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.3", + "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 3", + "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1", + "[30] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 3, 7", + "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 7", + "[32] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2", + "[33] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1902-1910, 2021. 5", + "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 2", + "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 1, 7", + "[36] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8658–8665, 2019. 1, 3, 7", + "[37] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 510-519, 2019. 2", + "[38] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 6", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1", + "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1", + "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2", + "[42] Min Peng, Chongyang Wang, Yu Shi, and Xiang-Dong Zhou. Efficient end-to-end video question answering with pyramidal multimodal transformer. arXiv preprint arXiv:2302.02136, 2023. 2, 3, 5, 7", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4", + "[44] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 5", + "[45] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2", + "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1", + "[47] Mykhailo Shvets, Wei Liu, and Alexander C Berg. Leveraging long-range temporal relationships between proposals for video object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9756-9764, 2019. 2", + "[48] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. arXiv preprint arXiv:2210.06031, 2022. 2", + "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[50] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified videolanguage pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 5, 7", + "[51] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wang-meng Zuo, and Qinghua Hu. Eca-net: Efficient channel" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27122", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 2", + "[52] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale microexpression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2", + "[53] Amanda L Woodward. Infants selectively encode the goal object of an actor's reach. Cognition, 69(1):1-34, 1998. 1", + "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 2", + "[55] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 2", + "[56] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2", + "[57] Xing Wu, Chaochen Gao, Zijia Lin, Zhongyuan Wang, Jizhong Han, and Songlin Hu. Rap: Redundancy-aware video-language pre-training for text-video retrieval. arXiv preprint arXiv:2210.06881, 2022.3", + "[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6", + "[59] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2, 6", + "[60] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2", + "[61] Ceyuan Yang, Yinghao Xu, Jianping Shi, Bo Dai, and Bolei Zhou. Temporal pyramid network for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 591-600, 2020. 5", + "[62] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6", + "[63] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[64] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot:" + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 2", + "[65] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and yang: Balancing and answering binary visual questions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5014-5022, 2016. 2", + "[66] Xinsong Zhang, Yan Zeng, Jipeng Zhang, and Hang Li. Toward building general foundation models for language, vision, and vision-language understanding tasks. arXiv preprint arXiv:2301.05065, 2023. 2", + "[67] Bo Zou, Chao Yang, Chengbin Quan, and Youjian Zhao. Spaceclip: A vision-language pretraining framework with spatial reconstruction on text. In Proceedings of the 31st ACM International Conference on Multimedia, pages 519-528, 2023. 3" + ], + "bbox": [ + 501, + 92, + 893, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27123", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_model.json b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a201627696433fc123b82cf0910b2871d607fa70 --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_model.json @@ -0,0 +1,2730 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.021 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.131, + 0.87, + 0.153 + ], + "angle": 0, + "content": "Language-aware Visual Semantic Distillation for Video Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.157, + 0.29, + 0.173 + ], + "angle": 0, + "content": "Bo Zou *" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.175, + 0.332, + 0.21 + ], + "angle": 0, + "content": "Tsinghua University Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.213, + 0.376, + 0.227 + ], + "angle": 0, + "content": "zoub21@mails.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.157, + 0.567, + 0.174 + ], + "angle": 0, + "content": "Chao Yang *" + }, + { + "type": "text", + "bbox": [ + 0.417, + 0.175, + 0.61, + 0.21 + ], + "angle": 0, + "content": "Shanghai AI Laboratory Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.419, + 0.213, + 0.608, + 0.227 + ], + "angle": 0, + "content": "yangchao@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.157, + 0.783, + 0.174 + ], + "angle": 0, + "content": "Yu Qiao" + }, + { + "type": "text", + "bbox": [ + 0.652, + 0.175, + 0.845, + 0.21 + ], + "angle": 0, + "content": "Shanghai AI Laboratory Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.664, + 0.213, + 0.834, + 0.227 + ], + "angle": 0, + "content": "qiaoyu@pjlab.org.cn" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.235, + 0.396, + 0.252 + ], + "angle": 0, + "content": "Chengbin Quan" + }, + { + "type": "text", + "bbox": [ + 0.252, + 0.254, + 0.413, + 0.27 + ], + "angle": 0, + "content": "Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.271, + 0.392, + 0.288 + ], + "angle": 0, + "content": "Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.291, + 0.431, + 0.305 + ], + "angle": 0, + "content": "quancb@tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.552, + 0.235, + 0.675, + 0.253 + ], + "angle": 0, + "content": "Youjian Zhao†" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.254, + 0.695, + 0.27 + ], + "angle": 0, + "content": "Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.271, + 0.72, + 0.288 + ], + "angle": 0, + "content": "Zhongguancun Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.289, + 0.673, + 0.305 + ], + "angle": 0, + "content": "Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.309, + 0.732, + 0.322 + ], + "angle": 0, + "content": "zhaoyoujian@tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.362, + 0.313, + 0.378 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.474, + 0.849 + ], + "angle": 0, + "content": "Significant progress in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although image-language models can efficiently represent both video and language branches, they typically employ goal-free vision perception and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure extracting multi-grained visual semantics associated with the questions. We conduct evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset." + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.355, + 0.87, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.397, + 0.547, + 0.404 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.397, + 0.743, + 0.405 + ], + "angle": 0, + "content": "Question: The man is singing along with what object?" + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.397, + 0.82, + 0.405 + ], + "angle": 0, + "content": "Answer: Car radio" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.406, + 0.87, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.457, + 0.547, + 0.464 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.457, + 0.739, + 0.464 + ], + "angle": 0, + "content": "Question: What object is taken out from the cabinet?" + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.457, + 0.8, + 0.464 + ], + "angle": 0, + "content": "Answer: Cup" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.465, + 0.869, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.507, + 0.545, + 0.514 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.507, + 0.868, + 0.515 + ], + "angle": 0, + "content": "Question: The person put the fish in the bowl into what object on the table? Answer: Tank" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.517, + 0.868, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.598, + 0.546, + 0.606 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.598, + 0.731, + 0.606 + ], + "angle": 0, + "content": "Question: What color is the banana in the photo?" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.618, + 0.895, + 0.675 + ], + "angle": 0, + "content": "Figure 1. Challenges of goal-free VideoQA models. They can not efficiently handle (a) Long-term dependencies, (b) Multi-events, and (c) Multi-scale semantics in the videos. They also suffer from (d) language prior phenomenon in training question-answer pairs." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.69, + 0.633, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.896, + 0.838 + ], + "angle": 0, + "content": "In recent years, large-scale video-and-language pretraining has seen remarkable progress. Most modern video-language understanding models [4, 15, 26, 31, 35, 36, 50] independently encode uni-modality and then fuse them. Human motion perception studies [5, 17, 53] suggest that humans perceive motion and environments as goal-driven behavior. This discrepancy leads to several issues, especially in long-form video understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.897, + 0.902 + ], + "angle": 0, + "content": "First, goal-free video representation struggles with long-term dependencies and multi-event reasoning. While enabled by progress in cross-modal pretraining[29, 43, 46] and datasets[39, 40], current methods excel at question an" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.875, + 0.346, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution, † Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.105, + 0.889, + 0.429, + 0.9 + ], + "angle": 0, + "content": "This work was done during an internship at Shanghai AI Lab." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.875, + 0.429, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27113" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.41 + ], + "angle": 0, + "content": "swering for images and short videos, but long-form videos contain many clips irrelevant or redundant to the questions, interfering with overall understanding. Encoding long videos also brings immense computational costs. Second, accurate semantic reasoning usually relies on multi-scale perception from local-spatial regions to global temporal dynamics. Goal-free methods for multi-scale visual modeling require custom submodels[9, 52, 56] for each scale or extra modalities like bounding boxes and OCR features[23, 66] but large-scale pretraining makes these approaches inefficient or infeasible. Third, goal-free video embedding requires a multimodal fusion module to synthesize questions and visual embeddings to predict answers. Incorporating questions directly can lead to shortcut solutions [1, 2, 8, 12, 19, 25, 41, 65], which means utilizing obvious clues in questions (mainly in data distribution and the relation between keywords) that exhibit more reliability in answer prediction than complicated visual reasoning, especially in the early training phases. This phenomenon is also known as language bias, which often causes major performance gaps in out-of-distribution tests. The above challenges are visualized in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.412, + 0.473, + 0.624 + ], + "angle": 0, + "content": "To address these issues, we propose a language-aware (goal-driven) visual semantic distillation framework called VideoDistill. Semantic distillation means the visual encoder must embed relevant frames of questions and focus on question-related multi-scale visual semantics. Closely resembles human behavior, semantic distillation functions like keeping the goal (question) in mind and conducting meaningful and precise visual reasoning. Differing from previous VideoQA frameworks [3, 31, 32, 42, 45, 59, 60], VideoDistill can generate answers only from question-related visual embeddings (without additional textual embeddings) since the interaction of text and vision is adequate during semantic distillation. This feature also enhances the significance of visual reasoning when generating answers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.627, + 0.473, + 0.884 + ], + "angle": 0, + "content": "To realize semantic distillation, we first introduce Language-Aware Gate (LA-Gate), a multi-head cross-gating mechanism for cross-model interaction, which is inspired by self-gating and gated attention[14, 22, 37, 51] and is integrable into any existing transformer-based models like[4, 50]. LA-Gates compute questions' dependencies on video patch embeddings and depress or excite corresponding patches in subsequent attention layers. The proposed LA-Gate is not only the key to semantic distillation but also a new efficient modality fusion method, which is a powerful competitor of predominant Cross-Attention in VideoQA. We discuss the differences between LA-Gate and attention mechanisms in Section 3.2 and summarize three merits of LA-Gate: (1) It can alleviate the influence of language bias by avoiding the direct involvement of text embeddings. (2) It can better maintain local diversity within the video embeddings. (3) It enhances the interpretability of the modality fusion process." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.886, + 0.472, + 0.9 + ], + "angle": 0, + "content": "VideoDistill has two LA-Gate-based modules. The first" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.212 + ], + "angle": 0, + "content": "is a differentiable sparse sampling module, which uses pretrained image-language models like CLIP[43] to encode frames, then performs goal-driven frame sampling to remarkably reduce subsequent spatial-temporal attention overhead and naturally avoid long-term dependencies and multi-event reasoning by retaining only question-related frames. It also provides our framework with a good nature that is insensitive to the number of sampled frames (see section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.895, + 0.332 + ], + "angle": 0, + "content": "The second is a vision refinement module eliminating unrelated visual semantics at different perceptual levels and enhancing related multi-scale semantics to support multilevel refinement. It encodes sparsely sampled frames into question-related global embeddings for generating answers. This module brings obvious performance boosts, especially on object-related questions (see section 4.5). Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.336, + 0.894, + 0.426 + ], + "angle": 0, + "content": "- We propose the Language-Aware Gate (LA-Gate) that enables interacting vision with language meanwhile not directly introducing language into visual representations. It can alleviate the influence of language bias, better maintain local diversity within the video embeddings, and is more interpretable compared with predominant Cross-Attention." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.427, + 0.895, + 0.517 + ], + "angle": 0, + "content": "- Based on LA-Gate, we propose a differentiable sparse sampling module to capture question-related frames and a vision refinement module to emphasize multi-scale question-related semantics. They can benefit VideoQA models on the stability under various numbers of sampled frames and the capacity to understand multi-scale objects." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.518, + 0.894, + 0.561 + ], + "angle": 0, + "content": "- Our model achieves new state-of-the-art performance on a wide range of downstream VideoQA tasks and text-to-video retrieval tasks." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.336, + 0.895, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.651, + 0.582 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.745, + 0.607 + ], + "angle": 0, + "content": "2.1. Long-form Video modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Long-form understanding has recently become a new research hot spot [11, 16, 31, 34, 47, 48, 54, 55, 59, 64]. Current methods develop in two directions to overcome unique challenges not encountered by previous short-term video understanding works. The first direction is to encode long-term dependencies better. LF-VILA [48] proposes hierarchical temporal window attention that begins with a small window learning the attention between adjacent frames, then gradually expands the window size to learn high-level representation. HERO [34] and MERLOT [64] predict the order of shuffled frames to understand sequential characteristics. The second direction aims to reduce the increasing computation cost of encoding longer videos. Since the common practice of randomly selecting 3 or 4 frames per video regardless of length is not suited for long videos, ClipBert [31] randomly samples a sequence of segments from a video and then aggregates their predictions. MIST [16] improves ClipBert by selecting both video segments and frame regions to be encoded and reduces computation costs further." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.088, + 0.833, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.401, + 0.895, + 0.444 + ], + "angle": 0, + "content": "Figure 2. Overview of VideoDistill. VideoDistill first densely samples video frames and utilizes a pre-trained image-language encoder to extract features, then sparsely samples a small number of question-related frames by a differentiable sparse sampling module. Finally, VideoDistill uses a vision refinement module to emphasize necessary multi-scale visual semantics in selected frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.495 + ], + "angle": 0, + "content": "TALLFormer [11] proposes short-term feature extraction and long-term memory mechanisms that avoid repetitive calculations during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.496, + 0.474, + 0.707 + ], + "angle": 0, + "content": "Similar to the second category, we reduce the workload by decreasing inputs, but our goal-driven differentiable sparse sampling works directly on frames, the minimum composition of the video, rather than rough video segments. Kim et al. [27] propose to use a self-gating mechanism to realize frame sampling, but it is a goal-free sampling method. Also, unlike random selection [31] or simple selector based on similarity[16], our learnable sampler excavates the semantic relationships between visual candidates. In addition, sparsely sampled frames from our sampler can be considered a concise summary of the long video sequence. Therefore, the following inference can perform similarly to models on short-form inputs. Thus, VideoDistill does not make efforts in the first direction." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.394, + 0.741 + ], + "angle": 0, + "content": "2.2. Video-Language Fusion in VideoQA" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The answer generation of VideoQA can be formulated as the interaction of the two modalities. Some methods propose fusing language and vision in the feature extraction stage for a more interactive combination. PSAC [36] combines regular visual self-attention with visual-linguistic co-attention. HCRN [30] supports conditioning video features with linguistic clues as a context factor in every inference stage. PMT [42] adopts a pyramidal video-language interaction. MCAN[63] utilizes Guided-Attention (V2T cross-attention in this paper) to fuse language into visual feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.895, + 0.553 + ], + "angle": 0, + "content": "Unlike these techniques, our VideoDistill only keeps the question \"in mind.\" We do not directly fuse the question information into the video embedding, thus generating \"purer\" visual embeddings. Since the answer prediction is only based on visual embeddings, VideoDistill can alleviate just assuming the answer from the question and follow the \"look and answer\" criterion [2]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.676 + ], + "angle": 0, + "content": "Besides VideoDistill, some existing methods can also be considered as indirect fusion. Contrastive learning methods like [10, 18, 20, 43, 57, 67] pull the matched pairs closer to interact vision with language. However, they do not demonstrate the same level of performance in tasks other than retrieval [28] due to ignoring the misalignment of vision and language. The method of fine-grained contrast to mitigate the misalignment has become a recent research hot spot." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.69, + 0.637, + 0.708 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.716, + 0.895, + 0.837 + ], + "angle": 0, + "content": "We introduce VideoDistill, a pretraining framework that enables goal-driven VideoQA relying on question-related frames and their multi-scale semantics. Figure 2 gives an overview of VideoDistill. VideoDistill consists of two sub-modules: Differentiable Sparse Sampling and Vision Refinement. Both sub-modules are built upon our proposed language-aware gate (LA-Gate) to perform in a goal-driven manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.795, + 0.865 + ], + "angle": 0, + "content": "3.1. Vision-Language Representations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.897, + 0.903 + ], + "angle": 0, + "content": "VideoDistill begins with densely sampling \\(N\\) uniformly distributed frames and embedding them into frame represent" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27115" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.359, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.614, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.09, + 0.852, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.25, + 0.746, + 0.264 + ], + "angle": 0, + "content": "Figure 3. Illustrations of Self-Attention, Cross Attention, and our LA-Gate mechanisms" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.383 + ], + "angle": 0, + "content": "tations. We adopt the pretrained image-language encoder CLIP [43] with frozen parameters in the visual branch to reduce the computational overhead. CLIP divides \\(N\\) frames into \\(N \\times n \\times n\\) patches, and extracts patch embeddings \\(v_{\\mathrm{patch}} \\in \\mathbb{R}^{N \\times n^2 \\times D}\\) and CLS tokens \\(v_{\\mathrm{cls}} \\in \\mathbb{R}^{N \\times D}\\) that represent the global understanding of frames, where \\(D\\) is the dimension of representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.489 + ], + "angle": 0, + "content": "For the Language branch, A text is first tokenized and then fed into the pretrained text encoder of CLIP to generate word-level embeddings \\( t_{\\mathrm{word}} \\in \\mathbb{R}^{M \\times D} \\) and a sentence-level token \\( t_{\\mathrm{cls}} \\in \\mathbb{R}^D \\), where M is the length of embeddings. Note that parameters in the text encoder are updated during pretraining and downstream finetuning to mitigate the linguistic gaps between datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.5, + 0.287, + 0.515 + ], + "angle": 0, + "content": "3.2. Language-Aware Gate" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.689 + ], + "angle": 0, + "content": "We introduce a reusable unit, termed language-aware gate (LA-Gate), which takes an arbitrary visual representation \\( v \\) and a sentence-level language representation \\( t_{\\mathrm{cls}} \\) as input. LA-Gate generates \\( v_{\\mathrm{distill}} \\), an output sequence of language-aware visual representation with the same dimension as \\( v \\), by exciting or depressing the components of \\( v \\) according to their similarities with \\( t_{\\mathrm{cls}} \\). For a better understanding of LA-Gate, we illustrate self-attention, cross-attention, and LA-Gate in Figure 3. We first discuss why commonly used cross-attention is not suitable for our purposes, then point out the design concept and benefits of LA-Gate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.841 + ], + "angle": 0, + "content": "Figure 3(b) illustrates two types of cross-attention: Text-to-Video (T2V, text as Query) and Video-to-Text (V2T, video as Query). Since outputs of the attention have the same length as Query and carry information sourced from Value, applying cross-attention in VideoQA will have two problems. Firstly, T2V is unsuitable for the modality interaction in the middle layers as it alters the shape of visual representations. Although V2T can maintain the shape, its output will carry the info directly from the text (like reconstructing text representations into the shape of visual input)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.472, + 0.902 + ], + "angle": 0, + "content": "The structure of LA-Gate is shown in Figure 3(c). There are two divergences between V2T and LA-Gate. (1) When text input length \\( L > 1 \\), we sum the similarity matrix to form a vector of importance for rows of the video input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.397 + ], + "angle": 0, + "content": "\\(L\\) functions like the number of multi-heads in the attention mechanism, and each text representation will partially dictate the importance of each row in the video input. (Note that in VideoDistill, we always have \\(L = 1\\) since the text input is \\(t_{\\mathrm{cls}}\\).) (2) We expand the importance vector (repeat D times) to match the shape of the video input. Finally, we apply an element-wise product rather than a dot product on the importance matrix and the Value." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.893, + 0.474 + ], + "angle": 0, + "content": "The first merit of LA-Gate is the text input only controls the importance of each visual representation, and the output does not directly involve text. Thus, the answer decoder can make decisions based on \"purer\" visual semantics and alleviate shortcut solutions hidden in texts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.69 + ], + "angle": 0, + "content": "A more significant characteristic of LA-Gate in general modality fusion usage: it can better maintain local diversity within the visual input. In Figure 3 (a) and (b), the output displays rows in mixed colors, representing the weighted sum of rows in the Value. Cross-attention will gather information from all rows in the Value for each output token and present more global attributes than LA-Gate since LA-Gate does not blend all features. In Figure 4(c), we maintain the uniqueness of colors for LA-Gate in each output row because each row is produced by multiplying the corresponding row in the Value and a scalar of importance. This characteristic makes LA-Gate fit more into our idea of information distillation because the language-aware distillation of a local area should be irrelevant to other regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.895, + 0.844 + ], + "angle": 0, + "content": "The last benefit of LA-Gate is we can enhance the interpretability of modality fusion by incorporating LA-Gate with self-attention layers. Due to the skip connection between the input and the output, commonly used V2T cross-attention brutally adds a reconstructed text (maybe linearly projected) into the visual input to fuse vision and language. The physical meaning of the structure is difficult to explain. By contrast, LA-Gate first distills the visual input by emphasizing language-related visual semantics, and then the interaction within visual representations is performed by self-attention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The implementation of LA-Gate is visualized in Figure 4 (we demonstrate how LA-Gate works in a vision refinement block, where \\( v \\) consists of frame patches from \\( K \\) selected frames and a video-level CLS token). Assume LA-Gate" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.471, + 0.162 + ], + "angle": 0, + "content": "receives \\( v \\in \\mathbb{R}^{m \\times D} \\), where \\( m \\) is the sequence length, and \\( t_{\\mathrm{cls}} \\in \\mathbb{R}^D \\). We first produce the key from the language representation \\( t_{\\mathrm{cls}} \\) and queries and values from the visual representation \\( v \\). Then, we calculate cosine distances between each query and the key:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.17, + 0.471, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\mathrm {k e y} = \\mathrm {w} _ {\\mathrm {k}} \\left(t _ {\\mathrm {c l s}}\\right), \\text {q u e r y} _ {i} = \\mathrm {w} _ {\\mathrm {q}} \\left(v _ {i}\\right), \\text {v a l u e} _ {i} = \\mathrm {w} _ {\\mathrm {v}} \\left(v _ {i}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.203, + 0.471, + 0.234 + ], + "angle": 0, + "content": "\\[\n\\operatorname {d i s t} _ {i} = \\operatorname {r e p e a t} \\left(\\frac {\\operatorname {q u e r y} _ {i} \\cdot \\operatorname {k e y}}{\\| \\operatorname {q u e r y} _ {i} \\| \\times \\| \\operatorname {k e y} \\|}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.321 + ], + "angle": 0, + "content": "where \\( v_{i} \\in \\mathbb{R}^{D}, i \\in [1, m] \\) is a feature vector in \\( v \\). \\( \\mathrm{w_k} \\), \\( \\mathrm{w_q} \\) and \\( \\mathrm{w_v} \\) are trainable linear projection layers, and repeat means expanding a scalar into a vector with the dimension of \\( D \\). Since \\( \\mathrm{dist}_i \\) reflects the correlation between vision and language, we treat it as an importance coefficient of \\( \\mathrm{value}_i \\). Then, we generate distilled \\( v_{i} \\) as formulated:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.329, + 0.471, + 0.345 + ], + "angle": 0, + "content": "\\[\nv _ {\\text {d i s t i l l}} ^ {i} = \\mathrm {w} _ {\\mathrm {o}} \\left(\\operatorname {d i s t} _ {i} \\odot \\operatorname {v a l u e} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.471, + 0.476 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise production and \\(\\mathrm{w_o}\\) is the linear output layer. In practice, we perform a standard multihead attention setting [49] on \\(\\mathrm{w_k}\\), \\(\\mathrm{w_q}\\), \\(\\mathrm{w_v}\\) and \\(\\mathrm{W_o}\\), and concatenate outputs of each head to form \\(v_{\\mathrm{distill}}^i\\). Although, we only illustrate a single attention head in Figure 2 for simplicity. Also, we apply an input skip connection on \\(v_{\\mathrm{distill}}^i\\) for better training stability and convergence. Finally, LA-Gate outputs \\(v_{\\mathrm{distill}} = \\left\\{v_{\\mathrm{distill}}^i | i \\in [1, m]\\right\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.497, + 0.355, + 0.514 + ], + "angle": 0, + "content": "3.3. Differentiable Sparse Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.731 + ], + "angle": 0, + "content": "Given \\(N\\) densely sampled frames, VideoDistill further adaptively picks out \\(K(K < N)\\) language-related frames. To this end, we utilize stacked frame sampling blocks (FS-Blocks, \\(L\\) layers) which take as input \\(v_{\\mathrm{cls}}\\) and \\(t_\\mathrm{cls}\\) to perform a top-\\(K\\) selection. Since \\(v_{\\mathrm{cls}}\\) of frames are separately extracted, to indicate their temporal positions in the whole video, we add temporal embedding \\(\\mathrm{T_f} \\in \\{\\phi_{\\mathrm{T}}(\\mathrm{f}) | \\mathrm{f} \\in [1,\\mathrm{N}]\\}\\) for each of them according to their frame index. After adding temporal information, each FS-Block first distills \\(v_{\\mathrm{cls}}\\) conditioned on \\(t_\\mathrm{cls}\\) by LA-Gate to emphasize question-related frames, then performs an interframe interaction through standard multi-head attention. Figure 4 shows the architecture of FS-Blocks. We borrow the form of skip connections from the divided block in [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.829 + ], + "angle": 0, + "content": "To realize the differentiable selection, we first project \\( v_{\\mathrm{cls}}^{L} \\), the output of the \\( L \\)-th FS-Block with the dimension identical to \\( v_{\\mathrm{cls}} \\), onto a feature space with the dimension of \\( K \\) by a linear layer \\( W_{s} \\), where \\( K \\) is the number of frames to be selected. Then, we conduct Gumbel-Softmax sampling [24] on each \\( x^{k} \\in \\mathbb{R}^{N}, k \\in [1, K] \\), the row of \\( \\mathrm{W_s}\\left(v_{\\mathrm{cls}}^L\\right) \\). The procedures are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.837, + 0.471, + 0.868 + ], + "angle": 0, + "content": "\\[\ny _ {\\mathrm {s o f t}} ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {x ^ {k} + g u m b e l s}{\\tau}\\right), \\tag {4}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.876, + 0.471, + 0.899 + ], + "angle": 0, + "content": "The code will be available at https://zoubo9034.github.io/VideoDistill/" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.853, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.233, + 0.692, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.233, + 0.852, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.383, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 4. Architectures of Language-Aware Gate, Frame Sampling Block, and Vision Refinement Block." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.42, + 0.892, + 0.444 + ], + "angle": 0, + "content": "\\[\ny _ {\\text {h a r d}} ^ {k} = \\operatorname {o n e h o t} \\left(\\arg \\max \\left(y _ {\\text {s o f t}} ^ {k}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.449, + 0.892, + 0.47 + ], + "angle": 0, + "content": "\\[\nm a s k ^ {k} = y _ {\\text {h a r d}} ^ {k} + y _ {\\text {s o f t}} ^ {k} - \\operatorname {s t o p g r a d} \\left(y _ {\\text {s o f t}} ^ {k}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.893, + 0.61 + ], + "angle": 0, + "content": "where gumbels is a noise sampled from the Gumbel distribution with \\(\\mu = 0\\) and \\(\\beta = 1\\), \\(y_{\\mathrm{soft}}^k\\) is the possibility of frames to be selected as the \\(k\\)-th language-related frame and \\(y_{\\mathrm{hard}}^k\\) reflects the index of the chosen frame. Since the argmax operation has no gradient, we adopt a code-level trick in equation 6 to generate mask \\(= \\{mask^{k}|k\\in [1,K]\\}\\) which can properly pass the gradient. Finally, we apply mask on \\(v_{\\mathrm{patch}}\\) and generate \\(v_{\\mathrm{patch}}^K\\in \\mathbb{R}^{K\\times n^2\\times D}\\) for the following vision refinement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.677, + 0.635 + ], + "angle": 0, + "content": "3.4. Vision Refinement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Given the selected patch embeddings \\( v_{\\mathrm{patch}}^{K} \\) and the language representation \\( t_{\\mathrm{cls}} \\), this module generates video-level representation \\( v_{\\mathrm{cls}}^{*} \\), which synthesizes multi-scale language-related visual semantics. \\( v_{\\mathrm{cls}}^{*} \\) is set as a learnable token with the dimension of \\( D \\) and is concatenated to \\( v_{\\mathrm{patch}}^{K} \\) as the visual input \\( v_{*} \\). We combine the LA-Gate and any existing spatial-temporal self-attention layer [4, 6, 50] to form a vision refinement block as shown in Figure 4. When iteratively applying vision refinement blocks on \\( v_{*} \\), LA-Gates spontaneously distill language-related visual semantics in different perceptive fields. Since the effective perceptive field gradually expands in stacked blocks [44], our module can consider multi-scale objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Unlike predominant pyramidal models [33, 42, 61] for multi-scale reasoning that gathers the intermediate results of each encoding stage as final representation, we empirically adopt only the \\( v_{\\mathrm{cls}}^* \\) from the last vision refinement block." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27117" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "3.5. Pretraining Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.16 + ], + "angle": 0, + "content": "We adopt the two pretraining tasks to facilitate cross-modal interaction: Video-Text Matching (VTM) and Vision-Guided Masked Language Modeling (VG-MLM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.16, + 0.471, + 0.271 + ], + "angle": 0, + "content": "Video-Text Matching. To facilitate cross-modal interaction, we use VTM as the first pre-training task. VTM predicts whether input video-language pairs are matched. In practice, we randomly exchange the annotations of two input pairs (matched) in a mini-batch with a probability of 0.5. Then, we use a linear projection head on the top of \\( v_{\\mathrm{cls}}^* \\) to predict two-class matching logits \\( y \\), and formulate the VTM loss as negative log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.283, + 0.471, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V T M}} = - \\mathbb {E} _ {\\left(v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p (y \\mid v _ {\\mathrm {c l s}} ^ {*}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.473, + 0.439 + ], + "angle": 0, + "content": "We do not adopt predominant contrastive pretraining because this paradigm requires the contrast between all possible combinations of vision and language. However, for VideoDistill, the number of possible visual embeddings of each video is proportional to the number of annotations. This leads to a quadratic growth in computational overhead (see Appendix A for more details). Nevertheless, we still apply a contrastive constraint on matched pairs (w/o exchange) to stabilize the training:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.448, + 0.471, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C L}} = - \\sum_ {i \\in B} \\log \\frac {\\mathbb {I} (\\text {m a t c h e d}) \\exp (s (v _ {i} , t _ {i}) / \\tau)}{\\sum_ {j \\in B} \\exp (s (v _ {i} , t _ {j}) / \\tau)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.568 + ], + "angle": 0, + "content": "where B is the batch size, \\( v_{i} \\) and \\( t_i \\) are \\( v_{\\mathrm{cls}}^* \\) and \\( t_\\mathrm{cls} \\) of \\( i \\)-th video-language pair, \\( \\mathbb{I} \\) is a indicate function of whether the video and the annotation in the \\( i \\)-th pair are matched, \\( s \\) denotes a cosine similarity function, \\( \\tau \\) is a temperature coefficient equals 0.07 in this paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.763 + ], + "angle": 0, + "content": "Vision-Guided Masked Language Modeling. The commonly used MLM [13] task aims to predict the ground truth label of the masked token according to textual context. To better map the visual and the language representations in a fine-grained manner, we propose VG-MLM, which is applied only on the matched pairs (w/o exchange) and encourages predicting the masked tokens from the visual context. In particular, we first provide VideoDistill the CLS tokens \\( t_{\\mathrm{cls}} \\) of unmasked annotations to sample frames and extract video representation \\( v_{\\mathrm{cls}}^* \\). Then, we mask \\( I \\) words in the annotations and encode the masked sentences. The outcome \\( w_i^{\\mathrm{M}}, i \\in [1, I] \\), which denotes the token of \\( i \\)-th masked word, and \\( v_{\\mathrm{cls}}^* \\) are used to predict the \\( i \\)-th masked word. The objective of VG-MLM is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.77, + 0.47, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V G - M L M}} = - \\mathbb {E} _ {\\left(t _ {\\mathrm {c l s}}, v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p \\left(w _ {i} | \\mathrm {S G} \\left(w _ {i} ^ {\\mathrm {M}}\\right), v _ {\\mathrm {c l s}} ^ {*}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where SG means stop gradient, \\(w_{i}\\) denote the logit of \\(i\\)-th masked word. We stop the gradient of \\(w_{i}^{\\mathrm{M}}\\) to enhance the importance of visual clues in language modeling. Following the heuristics of Bert [13], we adopt the same masking strategy. Then we adopt a two-layer MLP on top of the combination of \\(w_{i}^{\\mathrm{M}}\\) and \\(v_{\\mathrm{cls}}^{*}\\) to generate the probability over the vocabulary, which is calculated as the negative log-likelihood loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for the masked word. The overall loss of pretraining is the combination of VTM, VG-MLM, and CL:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.131, + 0.892, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {V T M}} + \\mathcal {L} _ {\\mathrm {V G - M L M}} + \\mathcal {L} _ {\\mathrm {C L}}, \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.16, + 0.634, + 0.177 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.895, + 0.412 + ], + "angle": 0, + "content": "In this section, conduct extensive experiments utilizing the pre-trained model on downstream various types of VideoQA tasks to verify the effectiveness of our proposed VideoDistill. We also transfer VideoDistill for video-text retrieval tasks to show the generalization power of the pre-trained model in Appendix C. Our pretraining set consists of three parts: (1) 3M video-caption pairs randomly sampled from generic dataset WebVid10M [4]. (2) 4.2M video-caption pairs randomly sampled from YouTube video dataset HD-VILA [59]. (3) 3.8M video-caption pairs from 1-st person view dataset EgoCLIP. We report results with \\( N = 100 \\), \\( K = 16 \\), and \\( L = 3 \\). For open-ended datasets, train a MLP classification head on the top of \\( v_{\\mathrm{cls}}^* \\). For multiple-choice datasets, we choose the answer with the maximum logit over the VTM head. More details are in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.68, + 0.438 + ], + "angle": 0, + "content": "4.1. Generic VideoQA:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.895, + 0.522 + ], + "angle": 0, + "content": "We first evaluate VideoDistill on the four commonly used VideoQA datasets: MSRVTT-QA [58], MSVD-QA [58], EgoMCQ [38] and MSRVTT-multiple-choice test [62]. Details of each dataset and results are in the supplementary material (Appendix C)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.531, + 0.707, + 0.547 + ], + "angle": 0, + "content": "4.2. Long-Form VideoQA:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.895, + 0.659 + ], + "angle": 0, + "content": "We evaluate our model on two recently proposed challenging datasets for the long-form VideoQA, namely EgoTaskQA[25] and AGQA[21]. We adopt the same finetuning setting with MSRVTT-QA and MSVD-QA. For fairness comparison, we report the results of VideoDistill w/o large-scale pertaining. Details of each dataset are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Results. In the Table 2, our VideoDistill overwhelm previous SOTA method and bring \\(2.82\\%\\) overall performance gains. Remarkably, in the categories of multi-agent and descriptive, we outperform previous work with \\(5.59\\%\\) and \\(3.72\\%\\) improvement. They show VideoDistill can better understand the storyline and temporal relations in long videos. In the category of object, we achieve \\(9.33\\%\\) gains, which indicates VideoDistill can capture multi-scale objects. The limited performance in the category of action is due to the strong correlations between action, object, and their change. Most methods tend to over-fit this strong language bias without thorough task understanding. Although the design of VideoDistill avoids utilizing the shortcut solutions in language. In Table 1, The results in the gray color (AIO, Temp, and MIST-CLIP) are reported by [16]. If they follow the same evaluation metrics [21] as us, one of the reasons for the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27118" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.268, + 0.09, + 0.7, + 0.103 + ], + "angle": 0, + "content": "Table 1. QA accuracies of state-of-the-art methods on AGQA v2 test set." + }, + { + "type": "table", + "bbox": [ + 0.158, + 0.103, + 0.809, + 0.224 + ], + "angle": 0, + "content": "
Question TypesMost LikelyPSACHMEHCRNAIO [50]Temp[ATP] [7]MIST-CLIPVideoDistill†
Object-relation9.3937.8437.4240.3348.3450.1551.6856.32
Relation-action50.0049.9549.9049.8648.9949.7667.1855.09
Object-action50.0050.0049.9749.8549.6646.2568.9955.12
Superlative21.0133.2033.2133.5537.5339.7842.0543.30
Sequencing49.7849.7849.7749.7049.6148.2567.2454.49
Exists50.0049.9449.9650.0150.8151.7960.3355.74
Duration comparison24.2745.2147.0343.8445.3649.5954.6249.08
Activity recognition5.524.145.435.5218.9718.9619.6910.16
All10.9940.1839.8942.1148.5949.7954.3955.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.236, + 0.237, + 0.733, + 0.251 + ], + "angle": 0, + "content": "Table 2. Performances on EgoTaskQA normal split. †denotes training from scratch." + }, + { + "type": "table", + "bbox": [ + 0.157, + 0.263, + 0.81, + 0.436 + ], + "angle": 0, + "content": "
CategoryVisualBERT [35]PSAC [36]HME [15]HGA [26]HCRN [30]ClipBERT [31]VideoDistill†
Scopeworld39.7340.7641.9138.8244.2742.1547.32
intent44.5146.1948.9242.1249.7740.9452.53
multi-agent26.2930.5927.9823.4331.3627.6336.95
Typedescriptive41.9940.6341.4538.0443.4838.4547.20
predictive30.3731.9835.8825.5736.5631.5040.43
counterfactual41.9941.8944.1341.9448.0046.7549.64
explanatory37.4237.9938.8535.9740.6042.3942.53
Semanticaction15.0214.7514.9915.0814.9222.9116.35
object23.2636.5336.0519.0945.3121.8054.64
state59.2061.8963.4455.6568.2854.3672.37
change68.2765.0568.8768.3867.3866.5871.47
all37.9338.9040.1636.7742.2039.8745.02
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.471, + 0.566 + ], + "angle": 0, + "content": "performance gaps in some categories is they adopt frames with higher resolution \\((448 \\times 448)\\) and more powerful encoder (CLIP-ViTB/32). Nevertheless, we still outperform current SOTA methods in Object-relation \\((4.64\\%)\\) and Superlative \\((1.25\\%)\\), which require multi-scale and multi-event reasoning. We also improve the overall performance with \\(1.41\\%\\) gains." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.579, + 0.422, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.774, + 0.411, + 0.787 + ], + "angle": 0, + "content": "Figure 5. The impact of the number of frames." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.802, + 0.468, + 0.819 + ], + "angle": 0, + "content": "4.3. The Impact of Differentiable Sparse Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Insensitive to the number of frames. We evaluate the performances on MSRVTT-QA with variable input frames. In Figure 5, we compare the performance change of VideoDistill (w/ & w/o sparse sampling module) with PMT [42] (other methods only report results on 16 frames, which are the best" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.656 + ], + "angle": 0, + "content": "results), and there are two findings. First, the differentiable sparse sampling module makes our 4-frame variation on par with other methods with many more input frames. It is intuitive because other competitors require more frames to improve the change to capture necessary information for reasoning (a disadvantage of goal-free perception). This property can dramatically reduce the computational overhead required to achieve the same performance as previous methods. Second, we notice that too many frames will damage the answering accuracy since the emergence of unrelated information disturbs the reasoning process. However, our sparse sampling module can mitigate the dropping with increasing frames and make the model more robust." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.658, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Qualitative Results. We visualize the results of our sparse sampling module on EgoTaskQA in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.724, + 0.713 + ], + "angle": 0, + "content": "4.4. The Impact of LA-Gate:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Prevent the use of language shortcuts. We come up with two methods to quantitatively analyze this characteristic. First, we test the language-only QA performances on EgoTaskQA normal split in Table 3, which means replacing the visual inputs with a static video (We also report the performances under a Gaussian noise input in Appendix E). The more declines, the better. When compared with the language-based model Bert [13], and the previous SOTA method HCRN [30], the decline of our VideoDistill is the most significant (especially in the object category). It demonstrates that VideoDistill relies more on vision and can avoid using language bias. Second, we test VideoDistill on Ego-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27119" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.09, + 0.816, + 0.104 + ], + "angle": 0, + "content": "Table 3. Language-only QA results on the EgoTaskQA normal split. The more performance drops, the better." + }, + { + "type": "table", + "bbox": [ + 0.161, + 0.105, + 0.81, + 0.258 + ], + "angle": 0, + "content": "
CategoryBERTHCRN (w/o vision)VideoDistill (w/o vision)
Acc.ChangeAcc.ChangeAcc.Change
world36.28-8.7%30.17-31.9%18.78-60.3%
intent35.02-21.3%35.54-28.6%22.57-57.0%
multi-agent20.58-21.7%19.9-36.5%11.50-68.7%
descriptive34.55-17.7%29.97-31.1%21.19-55.1%
predictive24.75-18.5%18.32-49.9%7.42-81.6%
counterfactual41.3-1.6%41.1-14.4%39.70-20.0%
explanatory31.78-15.1%32.41-20.2%13.69-67.8%
action15.72+4.6%17.31+16.0%11.81-27.8%
object7.43-68%8.85-80.5%0-100.0%
state45.03-23.9%35.51-48.0%20.21-72.1%
change69.87+2.3%70.47+4.6%55.38-22.5%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.325, + 0.26, + 0.645, + 0.272 + ], + "angle": 0, + "content": "Table 4. Analysis of the effectiveness of each module." + }, + { + "type": "table", + "bbox": [ + 0.162, + 0.275, + 0.81, + 0.361 + ], + "angle": 0, + "content": "
MethodModuleEgoTaskQAMSRVVTT-QA
SSVRCross-ATTLA-Gate
(a) BaselineXX--29.1023.92
(b) Uniform SamplingX✓X✓41.7040.92
(c) Soft SamplingX✓X✓38.9538.16
(d) w/o VR✓XX✓33.4426.51
(e) Cross-Attention✓✓✓X40.9441.17
(f) VideoDistill✓✓X✓45.0244.20
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.386, + 0.471, + 0.522 + ], + "angle": 0, + "content": "TaskQA indirect split, which is motivated by the fact [25] that during task execution, actions, objects, and their changes are often strongly correlated. It leaves the chance for the model to perform well by simply over-fitting these strong correlations (language bias) without thorough task understanding. The indirect references can avoid these correlations. In Appendix E, we show that our VideoDistill has the least absolute performance change. It indicates that VideoDistill barely utilizes language bias in questions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.598 + ], + "angle": 0, + "content": "A more effective fusion method than Cross-Attention. In Table 4 (e), we replace LA-Gates in our frameworks with Cross-Attention. When comparing (e) with (f), we can find that LA-Gate brings significant performance boosts on both test sets (4.08% and and 3.03%)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.373, + 0.621 + ], + "angle": 0, + "content": "4.5. The Impact of Vision Refinement:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.705 + ], + "angle": 0, + "content": "Performance boost in multi-scale object-related questions. The stacked video distillation blocks containing LA-Gate can obtain multi-scale question-related semantics. We observe huge improvements in Table 2 object category \\((9.33\\%)\\) and Table 1 object-relation category \\((4.64\\%)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.231, + 0.728 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.765 + ], + "angle": 0, + "content": "We conduct ablation studies to verify the effectiveness of components and the rationality of parameter selection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of each component. We design four variants and report their performances in Table 4. (a) we remove the two modules and apply evenly sampled 16 frames as input and a 3-layer transformer to fuse the vision and language. (b) we remove differentiable sparse sampling and apply evenly sampled 16 frames as input. (c) we replace the hard sampling in SS with soft sampling, which means using the weighted sum of dense frames rather than top-k selection. (d) we remove vision refinement and apply a 3-layer transformer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.386, + 0.892, + 0.416 + ], + "angle": 0, + "content": "to fuse the vision and language. (e) we replace LA-Gates in both modules with cross-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.462 + ], + "angle": 0, + "content": "More Ablations We validate the influence of pretraining tasks, the best number of densely sampled frames, and the reasonable number of stacked layers \\( L \\) in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.625, + 0.485 + ], + "angle": 0, + "content": "4.7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.78 + ], + "angle": 0, + "content": "In this paper, we study language-aware VideoQA to overcome the difficulties of long-term dependencies, multi-events, multi-scale semantics, and shortcut solutions in video understanding. In particular, we introduce a differentiable sparse sampling module that naturally avoids complicated long-term dependencies and multi-event reasoning since it only retains question-related frames and a vision refinement module that eliminates unrelated visual semantics in different perception fields and enhances multi-scale related semantics to support a multi-level refinement process. Both modules are based on our LA-Gate, which realizes cross-modal interaction meanwhile avoiding language's direct fusion into visual representations. LA-Gate is a powerful competitor of predominant Cross-Attention in VideoQA and fits our idea of semantic distillation. Therefore, we can realize high-performance VideoQA by just relying on question-related visual representations and reducing language shortcuts. Experiments on 8 VideoQA and text-to-video retrieval tasks demonstrate the effectiveness of our model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.795, + 0.618, + 0.812 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.82, + 0.895, + 0.896 + ], + "angle": 0, + "content": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Post-doctoral Excellent Program (Grant No. 2022234)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27120" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.157 + ], + "angle": 0, + "content": "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Anonymous. LLaMA-excitor: General instruction tuning via indirect feature interaction. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[5] Dare A Baldwin, Jodie A Baird, Megan M Saylor, and M Angela Clark. Infants parse dynamic action. *Child development*, 72(3):708-717, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.471, + 0.429 + ], + "angle": 0, + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[7] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[9] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[11] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with a long-memory transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 503-521. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[12] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1574–1583, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.119 + ], + "angle": 0, + "content": "text comprehension. arXiv preprint arXiv:1606.01549, 2016.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.895, + 0.204 + ], + "angle": 0, + "content": "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.895, + 0.261 + ], + "angle": 0, + "content": "[16] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. arXiv preprint arXiv:2212.09522, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.895, + 0.304 + ], + "angle": 0, + "content": "[17] György Gergely, Harold Bekkering, and Ildiko Király. Rational imitation in preverbal infants. Nature, 415(6873):755-755, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.895, + 0.388 + ], + "angle": 0, + "content": "[18] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5006–5015, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.391, + 0.894, + 0.46 + ], + "angle": 0, + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[21] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.66, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[23] Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. Iterative answer prediction with pointeraugmented multimodal transformers for textvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9992-10002, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.894, + 0.771 + ], + "angle": 0, + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.894, + 0.814 + ], + "angle": 0, + "content": "[25] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. arXiv preprint arXiv:2210.03929, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.87 + ], + "angle": 0, + "content": "[26] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[27] Hyounghun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.895, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27121" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.471, + 0.119 + ], + "angle": 0, + "content": "localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[30] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.404 + ], + "angle": 0, + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.472, + 0.475 + ], + "angle": 0, + "content": "[32] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.472, + 0.531 + ], + "angle": 0, + "content": "[33] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1902-1910, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.645 + ], + "angle": 0, + "content": "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.717 + ], + "angle": 0, + "content": "[36] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8658–8665, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[37] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 510-519, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[38] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.895, + 0.203 + ], + "angle": 0, + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.895, + 0.273 + ], + "angle": 0, + "content": "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.895, + 0.328 + ], + "angle": 0, + "content": "[42] Min Peng, Chongyang Wang, Yu Shi, and Xiang-Dong Zhou. Efficient end-to-end video question answering with pyramidal multimodal transformer. arXiv preprint arXiv:2302.02136, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.895, + 0.412 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.895, + 0.481 + ], + "angle": 0, + "content": "[44] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.895, + 0.538 + ], + "angle": 0, + "content": "[45] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.895, + 0.621 + ], + "angle": 0, + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.895, + 0.69 + ], + "angle": 0, + "content": "[47] Mykhailo Shvets, Wei Liu, and Alexander C Berg. Leveraging long-range temporal relationships between proposals for video object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9756-9764, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.895, + 0.748 + ], + "angle": 0, + "content": "[48] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. arXiv preprint arXiv:2210.06031, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.895, + 0.803 + ], + "angle": 0, + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.895, + 0.872 + ], + "angle": 0, + "content": "[50] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified videolanguage pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.895, + 0.902 + ], + "angle": 0, + "content": "[51] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wang-meng Zuo, and Qinghua Hu. Eca-net: Efficient channel" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27122" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.471, + 0.134 + ], + "angle": 0, + "content": "attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.189 + ], + "angle": 0, + "content": "[52] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale microexpression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[53] Amanda L Woodward. Infants selectively encode the goal object of an actor's reach. Cognition, 69(1):1-34, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.275, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[55] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[56] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[57] Xing Wu, Chaochen Gao, Zijia Lin, Zhongyuan Wang, Jizhong Han, and Songlin Hu. Rap: Redundancy-aware video-language pre-training for text-video retrieval. arXiv preprint arXiv:2210.06881, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.471, + 0.552 + ], + "angle": 0, + "content": "[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.471, + 0.635 + ], + "angle": 0, + "content": "[59] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[60] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.471, + 0.761 + ], + "angle": 0, + "content": "[61] Ceyuan Yang, Yinghao Xu, Jianping Shi, Bo Dai, and Bolei Zhou. Temporal pyramid network for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 591-600, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[62] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[63] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[64] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[65] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and yang: Balancing and answering binary visual questions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5014-5022, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[66] Xinsong Zhang, Yan Zeng, Jipeng Zhang, and Hang Li. Toward building general foundation models for language, vision, and vision-language understanding tasks. arXiv preprint arXiv:2301.05065, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[67] Bo Zou, Chao Yang, Chengbin Quan, and Youjian Zhao. Spaceclip: A vision-language pretraining framework with spatial reconstruction on text. In Proceedings of the 31st ACM International Conference on Multimedia, pages 519-528, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27123" + } + ] +] \ No newline at end of file diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_origin.pdf b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cc46c0e840285d5d2c83a2cd70dd86b367f6530f --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/f0849498-5020-4fb0-b9e6-adb27b80af64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e6e049c22f7e2231cbafa67dc13e4d52b7f8edc6bfb854c083cd2bc6d073054 +size 5947706 diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/full.md b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6f954d7eed237134f3a6cb9c8d0e1e2418c66854 --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/full.md @@ -0,0 +1,376 @@ +# Language-aware Visual Semantic Distillation for Video Question Answering + +Bo Zou * + +Tsinghua University Beijing, China + +zoub21@mails.tsinghua.edu.cn + +Chao Yang * + +Shanghai AI Laboratory Shanghai, China + +yangchao@pjlab.org.cn + +Yu Qiao + +Shanghai AI Laboratory Shanghai, China + +qiaoyu@pjlab.org.cn + +Chengbin Quan + +Tsinghua University + +Beijing, China + +quancb@tsinghua.edu.cn + +Youjian Zhao† + +Tsinghua University + +Zhongguancun Laboratory + +Beijing, China + +zhaoyoujian@tsinghua.edu.cn + +# Abstract + +Significant progress in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although image-language models can efficiently represent both video and language branches, they typically employ goal-free vision perception and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure extracting multi-grained visual semantics associated with the questions. We conduct evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset. + +![](images/d90d72df9833dde436ea53e7778a2f58cb4e42c69df533c70fbf3be776f4b073.jpg) +(a) +Answer: Car radio + +![](images/8ae326244236e0ac116abb45449c8ee4044bac881b2673645ff8ebcd0f9a77a9.jpg) +Question: The man is singing along with what object? +Question: What object is taken out from the cabinet? +Answer: Cup + +![](images/3ff03425bcc0e91988c86e4cca1be3106d4dff9a6ce32a827e02fd01f04f3165.jpg) +(b) +(c) +Question: The person put the fish in the bowl into what object on the table? Answer: Tank + +![](images/7c73333246d7aaaccab8845a932cced907171b2f90b6dc4913234b4565c4e9ce.jpg) +(d) +Question: What color is the banana in the photo? +Figure 1. Challenges of goal-free VideoQA models. They can not efficiently handle (a) Long-term dependencies, (b) Multi-events, and (c) Multi-scale semantics in the videos. They also suffer from (d) language prior phenomenon in training question-answer pairs. + +# 1. Introduction + +In recent years, large-scale video-and-language pretraining has seen remarkable progress. Most modern video-language understanding models [4, 15, 26, 31, 35, 36, 50] independently encode uni-modality and then fuse them. Human motion perception studies [5, 17, 53] suggest that humans perceive motion and environments as goal-driven behavior. This discrepancy leads to several issues, especially in long-form video understanding. + +First, goal-free video representation struggles with long-term dependencies and multi-event reasoning. While enabled by progress in cross-modal pretraining[29, 43, 46] and datasets[39, 40], current methods excel at question an + +swering for images and short videos, but long-form videos contain many clips irrelevant or redundant to the questions, interfering with overall understanding. Encoding long videos also brings immense computational costs. Second, accurate semantic reasoning usually relies on multi-scale perception from local-spatial regions to global temporal dynamics. Goal-free methods for multi-scale visual modeling require custom submodels[9, 52, 56] for each scale or extra modalities like bounding boxes and OCR features[23, 66] but large-scale pretraining makes these approaches inefficient or infeasible. Third, goal-free video embedding requires a multimodal fusion module to synthesize questions and visual embeddings to predict answers. Incorporating questions directly can lead to shortcut solutions [1, 2, 8, 12, 19, 25, 41, 65], which means utilizing obvious clues in questions (mainly in data distribution and the relation between keywords) that exhibit more reliability in answer prediction than complicated visual reasoning, especially in the early training phases. This phenomenon is also known as language bias, which often causes major performance gaps in out-of-distribution tests. The above challenges are visualized in Figure 1. + +To address these issues, we propose a language-aware (goal-driven) visual semantic distillation framework called VideoDistill. Semantic distillation means the visual encoder must embed relevant frames of questions and focus on question-related multi-scale visual semantics. Closely resembles human behavior, semantic distillation functions like keeping the goal (question) in mind and conducting meaningful and precise visual reasoning. Differing from previous VideoQA frameworks [3, 31, 32, 42, 45, 59, 60], VideoDistill can generate answers only from question-related visual embeddings (without additional textual embeddings) since the interaction of text and vision is adequate during semantic distillation. This feature also enhances the significance of visual reasoning when generating answers. + +To realize semantic distillation, we first introduce Language-Aware Gate (LA-Gate), a multi-head cross-gating mechanism for cross-model interaction, which is inspired by self-gating and gated attention[14, 22, 37, 51] and is integrable into any existing transformer-based models like[4, 50]. LA-Gates compute questions' dependencies on video patch embeddings and depress or excite corresponding patches in subsequent attention layers. The proposed LA-Gate is not only the key to semantic distillation but also a new efficient modality fusion method, which is a powerful competitor of predominant Cross-Attention in VideoQA. We discuss the differences between LA-Gate and attention mechanisms in Section 3.2 and summarize three merits of LA-Gate: (1) It can alleviate the influence of language bias by avoiding the direct involvement of text embeddings. (2) It can better maintain local diversity within the video embeddings. (3) It enhances the interpretability of the modality fusion process. + +VideoDistill has two LA-Gate-based modules. The first + +is a differentiable sparse sampling module, which uses pretrained image-language models like CLIP[43] to encode frames, then performs goal-driven frame sampling to remarkably reduce subsequent spatial-temporal attention overhead and naturally avoid long-term dependencies and multi-event reasoning by retaining only question-related frames. It also provides our framework with a good nature that is insensitive to the number of sampled frames (see section 4.3). + +The second is a vision refinement module eliminating unrelated visual semantics at different perceptual levels and enhancing related multi-scale semantics to support multilevel refinement. It encodes sparsely sampled frames into question-related global embeddings for generating answers. This module brings obvious performance boosts, especially on object-related questions (see section 4.5). Our contributions are summarized as follows: + +- We propose the Language-Aware Gate (LA-Gate) that enables interacting vision with language meanwhile not directly introducing language into visual representations. It can alleviate the influence of language bias, better maintain local diversity within the video embeddings, and is more interpretable compared with predominant Cross-Attention. +- Based on LA-Gate, we propose a differentiable sparse sampling module to capture question-related frames and a vision refinement module to emphasize multi-scale question-related semantics. They can benefit VideoQA models on the stability under various numbers of sampled frames and the capacity to understand multi-scale objects. +- Our model achieves new state-of-the-art performance on a wide range of downstream VideoQA tasks and text-to-video retrieval tasks. + +# 2. Related Works + +# 2.1. Long-form Video modeling + +Long-form understanding has recently become a new research hot spot [11, 16, 31, 34, 47, 48, 54, 55, 59, 64]. Current methods develop in two directions to overcome unique challenges not encountered by previous short-term video understanding works. The first direction is to encode long-term dependencies better. LF-VILA [48] proposes hierarchical temporal window attention that begins with a small window learning the attention between adjacent frames, then gradually expands the window size to learn high-level representation. HERO [34] and MERLOT [64] predict the order of shuffled frames to understand sequential characteristics. The second direction aims to reduce the increasing computation cost of encoding longer videos. Since the common practice of randomly selecting 3 or 4 frames per video regardless of length is not suited for long videos, ClipBert [31] randomly samples a sequence of segments from a video and then aggregates their predictions. MIST [16] improves ClipBert by selecting both video segments and frame regions to be encoded and reduces computation costs further. + +![](images/626fbdfe04d7fdff432f14ca2aa8decdc8648fb266b97ed7d1b7e129e299200e.jpg) +Figure 2. Overview of VideoDistill. VideoDistill first densely samples video frames and utilizes a pre-trained image-language encoder to extract features, then sparsely samples a small number of question-related frames by a differentiable sparse sampling module. Finally, VideoDistill uses a vision refinement module to emphasize necessary multi-scale visual semantics in selected frames. + +TALLFormer [11] proposes short-term feature extraction and long-term memory mechanisms that avoid repetitive calculations during training. + +Similar to the second category, we reduce the workload by decreasing inputs, but our goal-driven differentiable sparse sampling works directly on frames, the minimum composition of the video, rather than rough video segments. Kim et al. [27] propose to use a self-gating mechanism to realize frame sampling, but it is a goal-free sampling method. Also, unlike random selection [31] or simple selector based on similarity[16], our learnable sampler excavates the semantic relationships between visual candidates. In addition, sparsely sampled frames from our sampler can be considered a concise summary of the long video sequence. Therefore, the following inference can perform similarly to models on short-form inputs. Thus, VideoDistill does not make efforts in the first direction. + +# 2.2. Video-Language Fusion in VideoQA + +The answer generation of VideoQA can be formulated as the interaction of the two modalities. Some methods propose fusing language and vision in the feature extraction stage for a more interactive combination. PSAC [36] combines regular visual self-attention with visual-linguistic co-attention. HCRN [30] supports conditioning video features with linguistic clues as a context factor in every inference stage. PMT [42] adopts a pyramidal video-language interaction. MCAN[63] utilizes Guided-Attention (V2T cross-attention in this paper) to fuse language into visual feature extraction. + +Unlike these techniques, our VideoDistill only keeps the question "in mind." We do not directly fuse the question information into the video embedding, thus generating "purer" visual embeddings. Since the answer prediction is only based on visual embeddings, VideoDistill can alleviate just assuming the answer from the question and follow the "look and answer" criterion [2]. + +Besides VideoDistill, some existing methods can also be considered as indirect fusion. Contrastive learning methods like [10, 18, 20, 43, 57, 67] pull the matched pairs closer to interact vision with language. However, they do not demonstrate the same level of performance in tasks other than retrieval [28] due to ignoring the misalignment of vision and language. The method of fine-grained contrast to mitigate the misalignment has become a recent research hot spot. + +# 3. Methodology + +We introduce VideoDistill, a pretraining framework that enables goal-driven VideoQA relying on question-related frames and their multi-scale semantics. Figure 2 gives an overview of VideoDistill. VideoDistill consists of two sub-modules: Differentiable Sparse Sampling and Vision Refinement. Both sub-modules are built upon our proposed language-aware gate (LA-Gate) to perform in a goal-driven manner. + +# 3.1. Vision-Language Representations + +VideoDistill begins with densely sampling $N$ uniformly distributed frames and embedding them into frame represent + +![](images/ca32b2d5e3347226dfc993cccb89c381cf492851e87fe4e9cf0b582db00a2fa8.jpg) +Figure 3. Illustrations of Self-Attention, Cross Attention, and our LA-Gate mechanisms + +![](images/659c9c9cce9b2386df7b74c2cb763293a40f6774e57ab5fc58391923618fc584.jpg) + +![](images/d94d1263b62a166d54a368a8b7123f1b16aeaf63cf6659859fd4922d8510da66.jpg) + +tations. We adopt the pretrained image-language encoder CLIP [43] with frozen parameters in the visual branch to reduce the computational overhead. CLIP divides $N$ frames into $N \times n \times n$ patches, and extracts patch embeddings $v_{\mathrm{patch}} \in \mathbb{R}^{N \times n^2 \times D}$ and CLS tokens $v_{\mathrm{cls}} \in \mathbb{R}^{N \times D}$ that represent the global understanding of frames, where $D$ is the dimension of representations. + +For the Language branch, A text is first tokenized and then fed into the pretrained text encoder of CLIP to generate word-level embeddings $t_{\mathrm{word}} \in \mathbb{R}^{M \times D}$ and a sentence-level token $t_{\mathrm{cls}} \in \mathbb{R}^D$ , where M is the length of embeddings. Note that parameters in the text encoder are updated during pretraining and downstream finetuning to mitigate the linguistic gaps between datasets. + +# 3.2. Language-Aware Gate + +We introduce a reusable unit, termed language-aware gate (LA-Gate), which takes an arbitrary visual representation $v$ and a sentence-level language representation $t_{\mathrm{cls}}$ as input. LA-Gate generates $v_{\mathrm{distill}}$ , an output sequence of language-aware visual representation with the same dimension as $v$ , by exciting or depressing the components of $v$ according to their similarities with $t_{\mathrm{cls}}$ . For a better understanding of LA-Gate, we illustrate self-attention, cross-attention, and LA-Gate in Figure 3. We first discuss why commonly used cross-attention is not suitable for our purposes, then point out the design concept and benefits of LA-Gate. + +Figure 3(b) illustrates two types of cross-attention: Text-to-Video (T2V, text as Query) and Video-to-Text (V2T, video as Query). Since outputs of the attention have the same length as Query and carry information sourced from Value, applying cross-attention in VideoQA will have two problems. Firstly, T2V is unsuitable for the modality interaction in the middle layers as it alters the shape of visual representations. Although V2T can maintain the shape, its output will carry the info directly from the text (like reconstructing text representations into the shape of visual input). + +The structure of LA-Gate is shown in Figure 3(c). There are two divergences between V2T and LA-Gate. (1) When text input length $L > 1$ , we sum the similarity matrix to form a vector of importance for rows of the video input. + +$L$ functions like the number of multi-heads in the attention mechanism, and each text representation will partially dictate the importance of each row in the video input. (Note that in VideoDistill, we always have $L = 1$ since the text input is $t_{\mathrm{cls}}$ .) (2) We expand the importance vector (repeat D times) to match the shape of the video input. Finally, we apply an element-wise product rather than a dot product on the importance matrix and the Value. + +The first merit of LA-Gate is the text input only controls the importance of each visual representation, and the output does not directly involve text. Thus, the answer decoder can make decisions based on "purer" visual semantics and alleviate shortcut solutions hidden in texts. + +A more significant characteristic of LA-Gate in general modality fusion usage: it can better maintain local diversity within the visual input. In Figure 3 (a) and (b), the output displays rows in mixed colors, representing the weighted sum of rows in the Value. Cross-attention will gather information from all rows in the Value for each output token and present more global attributes than LA-Gate since LA-Gate does not blend all features. In Figure 4(c), we maintain the uniqueness of colors for LA-Gate in each output row because each row is produced by multiplying the corresponding row in the Value and a scalar of importance. This characteristic makes LA-Gate fit more into our idea of information distillation because the language-aware distillation of a local area should be irrelevant to other regions. + +The last benefit of LA-Gate is we can enhance the interpretability of modality fusion by incorporating LA-Gate with self-attention layers. Due to the skip connection between the input and the output, commonly used V2T cross-attention brutally adds a reconstructed text (maybe linearly projected) into the visual input to fuse vision and language. The physical meaning of the structure is difficult to explain. By contrast, LA-Gate first distills the visual input by emphasizing language-related visual semantics, and then the interaction within visual representations is performed by self-attention. + +The implementation of LA-Gate is visualized in Figure 4 (we demonstrate how LA-Gate works in a vision refinement block, where $v$ consists of frame patches from $K$ selected frames and a video-level CLS token). Assume LA-Gate + +receives $v \in \mathbb{R}^{m \times D}$ , where $m$ is the sequence length, and $t_{\mathrm{cls}} \in \mathbb{R}^D$ . We first produce the key from the language representation $t_{\mathrm{cls}}$ and queries and values from the visual representation $v$ . Then, we calculate cosine distances between each query and the key: + +$$ +\mathrm {k e y} = \mathrm {w} _ {\mathrm {k}} \left(t _ {\mathrm {c l s}}\right), \text {q u e r y} _ {i} = \mathrm {w} _ {\mathrm {q}} \left(v _ {i}\right), \text {v a l u e} _ {i} = \mathrm {w} _ {\mathrm {v}} \left(v _ {i}\right) \tag {1} +$$ + +$$ +\operatorname {d i s t} _ {i} = \operatorname {r e p e a t} \left(\frac {\operatorname {q u e r y} _ {i} \cdot \operatorname {k e y}}{\| \operatorname {q u e r y} _ {i} \| \times \| \operatorname {k e y} \|}\right), \tag {2} +$$ + +where $v_{i} \in \mathbb{R}^{D}, i \in [1, m]$ is a feature vector in $v$ . $\mathrm{w_k}$ , $\mathrm{w_q}$ and $\mathrm{w_v}$ are trainable linear projection layers, and repeat means expanding a scalar into a vector with the dimension of $D$ . Since $\mathrm{dist}_i$ reflects the correlation between vision and language, we treat it as an importance coefficient of $\mathrm{value}_i$ . Then, we generate distilled $v_{i}$ as formulated: + +$$ +v _ {\text {d i s t i l l}} ^ {i} = \mathrm {w} _ {\mathrm {o}} \left(\operatorname {d i s t} _ {i} \odot \operatorname {v a l u e} _ {i}\right), \tag {3} +$$ + +where $\odot$ denotes element-wise production and $\mathrm{w_o}$ is the linear output layer. In practice, we perform a standard multihead attention setting [49] on $\mathrm{w_k}$ , $\mathrm{w_q}$ , $\mathrm{w_v}$ and $\mathrm{W_o}$ , and concatenate outputs of each head to form $v_{\mathrm{distill}}^i$ . Although, we only illustrate a single attention head in Figure 2 for simplicity. Also, we apply an input skip connection on $v_{\mathrm{distill}}^i$ for better training stability and convergence. Finally, LA-Gate outputs $v_{\mathrm{distill}} = \left\{v_{\mathrm{distill}}^i | i \in [1, m]\right\}$ . + +# 3.3. Differentiable Sparse Sampling + +Given $N$ densely sampled frames, VideoDistill further adaptively picks out $K(K < N)$ language-related frames. To this end, we utilize stacked frame sampling blocks (FS-Blocks, $L$ layers) which take as input $v_{\mathrm{cls}}$ and $t_\mathrm{cls}$ to perform a top- $K$ selection. Since $v_{\mathrm{cls}}$ of frames are separately extracted, to indicate their temporal positions in the whole video, we add temporal embedding $\mathrm{T_f} \in \{\phi_{\mathrm{T}}(\mathrm{f}) | \mathrm{f} \in [1,\mathrm{N}]\}$ for each of them according to their frame index. After adding temporal information, each FS-Block first distills $v_{\mathrm{cls}}$ conditioned on $t_\mathrm{cls}$ by LA-Gate to emphasize question-related frames, then performs an interframe interaction through standard multi-head attention. Figure 4 shows the architecture of FS-Blocks. We borrow the form of skip connections from the divided block in [4]. + +To realize the differentiable selection, we first project $v_{\mathrm{cls}}^{L}$ , the output of the $L$ -th FS-Block with the dimension identical to $v_{\mathrm{cls}}$ , onto a feature space with the dimension of $K$ by a linear layer $W_{s}$ , where $K$ is the number of frames to be selected. Then, we conduct Gumbel-Softmax sampling [24] on each $x^{k} \in \mathbb{R}^{N}, k \in [1, K]$ , the row of $\mathrm{W_s}\left(v_{\mathrm{cls}}^L\right)$ . The procedures are formulated as follows: + +$$ +y _ {\mathrm {s o f t}} ^ {k} = \operatorname {s o f t m a x} \left(\frac {x ^ {k} + g u m b e l s}{\tau}\right), \tag {4} +$$ + +![](images/d3ab6d664f307c3dff80b4e8eb181ad725bb754d880c0e5f60f1f54ee3b229a1.jpg) + +![](images/f3cdc518f4e623ccb44883fecb4fac28123a40d223674c4f08d7be1361939db6.jpg) +Figure 4. Architectures of Language-Aware Gate, Frame Sampling Block, and Vision Refinement Block. + +![](images/c0f2f9654508f02079edec81448f0484474b342476edab5d6188e734956e5484.jpg) + +$$ +y _ {\text {h a r d}} ^ {k} = \operatorname {o n e h o t} \left(\arg \max \left(y _ {\text {s o f t}} ^ {k}\right)\right), \tag {5} +$$ + +$$ +m a s k ^ {k} = y _ {\text {h a r d}} ^ {k} + y _ {\text {s o f t}} ^ {k} - \operatorname {s t o p g r a d} \left(y _ {\text {s o f t}} ^ {k}\right), \tag {6} +$$ + +where gumbels is a noise sampled from the Gumbel distribution with $\mu = 0$ and $\beta = 1$ , $y_{\mathrm{soft}}^k$ is the possibility of frames to be selected as the $k$ -th language-related frame and $y_{\mathrm{hard}}^k$ reflects the index of the chosen frame. Since the argmax operation has no gradient, we adopt a code-level trick in equation 6 to generate mask $= \{mask^{k}|k\in [1,K]\}$ which can properly pass the gradient. Finally, we apply mask on $v_{\mathrm{patch}}$ and generate $v_{\mathrm{patch}}^K\in \mathbb{R}^{K\times n^2\times D}$ for the following vision refinement. + +# 3.4. Vision Refinement + +Given the selected patch embeddings $v_{\mathrm{patch}}^{K}$ and the language representation $t_{\mathrm{cls}}$ , this module generates video-level representation $v_{\mathrm{cls}}^{*}$ , which synthesizes multi-scale language-related visual semantics. $v_{\mathrm{cls}}^{*}$ is set as a learnable token with the dimension of $D$ and is concatenated to $v_{\mathrm{patch}}^{K}$ as the visual input $v_{*}$ . We combine the LA-Gate and any existing spatial-temporal self-attention layer [4, 6, 50] to form a vision refinement block as shown in Figure 4. When iteratively applying vision refinement blocks on $v_{*}$ , LA-Gates spontaneously distill language-related visual semantics in different perceptive fields. Since the effective perceptive field gradually expands in stacked blocks [44], our module can consider multi-scale objects. + +Unlike predominant pyramidal models [33, 42, 61] for multi-scale reasoning that gathers the intermediate results of each encoding stage as final representation, we empirically adopt only the $v_{\mathrm{cls}}^*$ from the last vision refinement block. + +# 3.5. Pretraining Tasks + +We adopt the two pretraining tasks to facilitate cross-modal interaction: Video-Text Matching (VTM) and Vision-Guided Masked Language Modeling (VG-MLM). + +Video-Text Matching. To facilitate cross-modal interaction, we use VTM as the first pre-training task. VTM predicts whether input video-language pairs are matched. In practice, we randomly exchange the annotations of two input pairs (matched) in a mini-batch with a probability of 0.5. Then, we use a linear projection head on the top of $v_{\mathrm{cls}}^*$ to predict two-class matching logits $y$ , and formulate the VTM loss as negative log-likelihood: + +$$ +\mathcal {L} _ {\mathrm {V T M}} = - \mathbb {E} _ {\left(v _ {\mathrm {c l s}} ^ {*}\right)} \log p (y \mid v _ {\mathrm {c l s}} ^ {*}), \tag {7} +$$ + +We do not adopt predominant contrastive pretraining because this paradigm requires the contrast between all possible combinations of vision and language. However, for VideoDistill, the number of possible visual embeddings of each video is proportional to the number of annotations. This leads to a quadratic growth in computational overhead (see Appendix A for more details). Nevertheless, we still apply a contrastive constraint on matched pairs (w/o exchange) to stabilize the training: + +$$ +\mathcal {L} _ {\mathrm {C L}} = - \sum_ {i \in B} \log \frac {\mathbb {I} (\text {m a t c h e d}) \exp (s (v _ {i} , t _ {i}) / \tau)}{\sum_ {j \in B} \exp (s (v _ {i} , t _ {j}) / \tau)}, \tag {8} +$$ + +where B is the batch size, $v_{i}$ and $t_i$ are $v_{\mathrm{cls}}^*$ and $t_\mathrm{cls}$ of $i$ -th video-language pair, $\mathbb{I}$ is a indicate function of whether the video and the annotation in the $i$ -th pair are matched, $s$ denotes a cosine similarity function, $\tau$ is a temperature coefficient equals 0.07 in this paper. + +Vision-Guided Masked Language Modeling. The commonly used MLM [13] task aims to predict the ground truth label of the masked token according to textual context. To better map the visual and the language representations in a fine-grained manner, we propose VG-MLM, which is applied only on the matched pairs (w/o exchange) and encourages predicting the masked tokens from the visual context. In particular, we first provide VideoDistill the CLS tokens $t_{\mathrm{cls}}$ of unmasked annotations to sample frames and extract video representation $v_{\mathrm{cls}}^*$ . Then, we mask $I$ words in the annotations and encode the masked sentences. The outcome $w_i^{\mathrm{M}}, i \in [1, I]$ , which denotes the token of $i$ -th masked word, and $v_{\mathrm{cls}}^*$ are used to predict the $i$ -th masked word. The objective of VG-MLM is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {V G - M L M}} = - \mathbb {E} _ {\left(t _ {\mathrm {c l s}}, v _ {\mathrm {c l s}} ^ {*}\right)} \log p \left(w _ {i} | \mathrm {S G} \left(w _ {i} ^ {\mathrm {M}}\right), v _ {\mathrm {c l s}} ^ {*}\right), \tag {9} +$$ + +where SG means stop gradient, $w_{i}$ denote the logit of $i$ -th masked word. We stop the gradient of $w_{i}^{\mathrm{M}}$ to enhance the importance of visual clues in language modeling. Following the heuristics of Bert [13], we adopt the same masking strategy. Then we adopt a two-layer MLP on top of the combination of $w_{i}^{\mathrm{M}}$ and $v_{\mathrm{cls}}^{*}$ to generate the probability over the vocabulary, which is calculated as the negative log-likelihood loss + +for the masked word. The overall loss of pretraining is the combination of VTM, VG-MLM, and CL: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {V T M}} + \mathcal {L} _ {\mathrm {V G - M L M}} + \mathcal {L} _ {\mathrm {C L}}, \tag {10} +$$ + +# 4. Experiments + +In this section, conduct extensive experiments utilizing the pre-trained model on downstream various types of VideoQA tasks to verify the effectiveness of our proposed VideoDistill. We also transfer VideoDistill for video-text retrieval tasks to show the generalization power of the pre-trained model in Appendix C. Our pretraining set consists of three parts: (1) 3M video-caption pairs randomly sampled from generic dataset WebVid10M [4]. (2) 4.2M video-caption pairs randomly sampled from YouTube video dataset HD-VILA [59]. (3) 3.8M video-caption pairs from 1-st person view dataset EgoCLIP. We report results with $N = 100$ , $K = 16$ , and $L = 3$ . For open-ended datasets, train a MLP classification head on the top of $v_{\mathrm{cls}}^*$ . For multiple-choice datasets, we choose the answer with the maximum logit over the VTM head. More details are in Appendix B. + +# 4.1. Generic VideoQA: + +We first evaluate VideoDistill on the four commonly used VideoQA datasets: MSRVTT-QA [58], MSVD-QA [58], EgoMCQ [38] and MSRVTT-multiple-choice test [62]. Details of each dataset and results are in the supplementary material (Appendix C). + +# 4.2. Long-Form VideoQA: + +We evaluate our model on two recently proposed challenging datasets for the long-form VideoQA, namely EgoTaskQA[25] and AGQA[21]. We adopt the same finetuning setting with MSRVTT-QA and MSVD-QA. For fairness comparison, we report the results of VideoDistill w/o large-scale pertaining. Details of each dataset are in the supplementary material. + +Results. In the Table 2, our VideoDistill overwhelm previous SOTA method and bring $2.82\%$ overall performance gains. Remarkably, in the categories of multi-agent and descriptive, we outperform previous work with $5.59\%$ and $3.72\%$ improvement. They show VideoDistill can better understand the storyline and temporal relations in long videos. In the category of object, we achieve $9.33\%$ gains, which indicates VideoDistill can capture multi-scale objects. The limited performance in the category of action is due to the strong correlations between action, object, and their change. Most methods tend to over-fit this strong language bias without thorough task understanding. Although the design of VideoDistill avoids utilizing the shortcut solutions in language. In Table 1, The results in the gray color (AIO, Temp, and MIST-CLIP) are reported by [16]. If they follow the same evaluation metrics [21] as us, one of the reasons for the + +Table 1. QA accuracies of state-of-the-art methods on AGQA v2 test set. + +
Question TypesMost LikelyPSACHMEHCRNAIO [50]Temp[ATP] [7]MIST-CLIPVideoDistill†
Object-relation9.3937.8437.4240.3348.3450.1551.6856.32
Relation-action50.0049.9549.9049.8648.9949.7667.1855.09
Object-action50.0050.0049.9749.8549.6646.2568.9955.12
Superlative21.0133.2033.2133.5537.5339.7842.0543.30
Sequencing49.7849.7849.7749.7049.6148.2567.2454.49
Exists50.0049.9449.9650.0150.8151.7960.3355.74
Duration comparison24.2745.2147.0343.8445.3649.5954.6249.08
Activity recognition5.524.145.435.5218.9718.9619.6910.16
All10.9940.1839.8942.1148.5949.7954.3955.80
+ +Table 2. Performances on EgoTaskQA normal split. †denotes training from scratch. + +
CategoryVisualBERT [35]PSAC [36]HME [15]HGA [26]HCRN [30]ClipBERT [31]VideoDistill†
Scopeworld39.7340.7641.9138.8244.2742.1547.32
intent44.5146.1948.9242.1249.7740.9452.53
multi-agent26.2930.5927.9823.4331.3627.6336.95
Typedescriptive41.9940.6341.4538.0443.4838.4547.20
predictive30.3731.9835.8825.5736.5631.5040.43
counterfactual41.9941.8944.1341.9448.0046.7549.64
explanatory37.4237.9938.8535.9740.6042.3942.53
Semanticaction15.0214.7514.9915.0814.9222.9116.35
object23.2636.5336.0519.0945.3121.8054.64
state59.2061.8963.4455.6568.2854.3672.37
change68.2765.0568.8768.3867.3866.5871.47
all37.9338.9040.1636.7742.2039.8745.02
+ +performance gaps in some categories is they adopt frames with higher resolution $(448 \times 448)$ and more powerful encoder (CLIP-ViTB/32). Nevertheless, we still outperform current SOTA methods in Object-relation $(4.64\%)$ and Superlative $(1.25\%)$ , which require multi-scale and multi-event reasoning. We also improve the overall performance with $1.41\%$ gains. + +![](images/d9301845e814910e72d00c2a5c28328a919818c6893a3a9ea0c016d87bb86b32.jpg) +Figure 5. The impact of the number of frames. + +# 4.3. The Impact of Differentiable Sparse Sampling + +Insensitive to the number of frames. We evaluate the performances on MSRVTT-QA with variable input frames. In Figure 5, we compare the performance change of VideoDistill (w/ & w/o sparse sampling module) with PMT [42] (other methods only report results on 16 frames, which are the best + +results), and there are two findings. First, the differentiable sparse sampling module makes our 4-frame variation on par with other methods with many more input frames. It is intuitive because other competitors require more frames to improve the change to capture necessary information for reasoning (a disadvantage of goal-free perception). This property can dramatically reduce the computational overhead required to achieve the same performance as previous methods. Second, we notice that too many frames will damage the answering accuracy since the emergence of unrelated information disturbs the reasoning process. However, our sparse sampling module can mitigate the dropping with increasing frames and make the model more robust. + +Qualitative Results. We visualize the results of our sparse sampling module on EgoTaskQA in Appendix E. + +# 4.4. The Impact of LA-Gate: + +Prevent the use of language shortcuts. We come up with two methods to quantitatively analyze this characteristic. First, we test the language-only QA performances on EgoTaskQA normal split in Table 3, which means replacing the visual inputs with a static video (We also report the performances under a Gaussian noise input in Appendix E). The more declines, the better. When compared with the language-based model Bert [13], and the previous SOTA method HCRN [30], the decline of our VideoDistill is the most significant (especially in the object category). It demonstrates that VideoDistill relies more on vision and can avoid using language bias. Second, we test VideoDistill on Ego- + +Table 3. Language-only QA results on the EgoTaskQA normal split. The more performance drops, the better. + +
CategoryBERTHCRN (w/o vision)VideoDistill (w/o vision)
Acc.ChangeAcc.ChangeAcc.Change
world36.28-8.7%30.17-31.9%18.78-60.3%
intent35.02-21.3%35.54-28.6%22.57-57.0%
multi-agent20.58-21.7%19.9-36.5%11.50-68.7%
descriptive34.55-17.7%29.97-31.1%21.19-55.1%
predictive24.75-18.5%18.32-49.9%7.42-81.6%
counterfactual41.3-1.6%41.1-14.4%39.70-20.0%
explanatory31.78-15.1%32.41-20.2%13.69-67.8%
action15.72+4.6%17.31+16.0%11.81-27.8%
object7.43-68%8.85-80.5%0-100.0%
state45.03-23.9%35.51-48.0%20.21-72.1%
change69.87+2.3%70.47+4.6%55.38-22.5%
+ +Table 4. Analysis of the effectiveness of each module. + +
MethodModuleEgoTaskQAMSRVVTT-QA
SSVRCross-ATTLA-Gate
(a) BaselineXX--29.1023.92
(b) Uniform SamplingX✓X✓41.7040.92
(c) Soft SamplingX✓X✓38.9538.16
(d) w/o VR✓XX✓33.4426.51
(e) Cross-Attention✓✓✓X40.9441.17
(f) VideoDistill✓✓X✓45.0244.20
+ +TaskQA indirect split, which is motivated by the fact [25] that during task execution, actions, objects, and their changes are often strongly correlated. It leaves the chance for the model to perform well by simply over-fitting these strong correlations (language bias) without thorough task understanding. The indirect references can avoid these correlations. In Appendix E, we show that our VideoDistill has the least absolute performance change. It indicates that VideoDistill barely utilizes language bias in questions. + +A more effective fusion method than Cross-Attention. In Table 4 (e), we replace LA-Gates in our frameworks with Cross-Attention. When comparing (e) with (f), we can find that LA-Gate brings significant performance boosts on both test sets (4.08% and and 3.03%) + +# 4.5. The Impact of Vision Refinement: + +Performance boost in multi-scale object-related questions. The stacked video distillation blocks containing LA-Gate can obtain multi-scale question-related semantics. We observe huge improvements in Table 2 object category $(9.33\%)$ and Table 1 object-relation category $(4.64\%)$ . + +# 4.6. Ablation Study + +We conduct ablation studies to verify the effectiveness of components and the rationality of parameter selection. + +Effectiveness of each component. We design four variants and report their performances in Table 4. (a) we remove the two modules and apply evenly sampled 16 frames as input and a 3-layer transformer to fuse the vision and language. (b) we remove differentiable sparse sampling and apply evenly sampled 16 frames as input. (c) we replace the hard sampling in SS with soft sampling, which means using the weighted sum of dense frames rather than top-k selection. (d) we remove vision refinement and apply a 3-layer transformer + +to fuse the vision and language. (e) we replace LA-Gates in both modules with cross-attention layers. + +More Ablations We validate the influence of pretraining tasks, the best number of densely sampled frames, and the reasonable number of stacked layers $L$ in Appendix E. + +# 4.7. Conclusion + +In this paper, we study language-aware VideoQA to overcome the difficulties of long-term dependencies, multi-events, multi-scale semantics, and shortcut solutions in video understanding. In particular, we introduce a differentiable sparse sampling module that naturally avoids complicated long-term dependencies and multi-event reasoning since it only retains question-related frames and a vision refinement module that eliminates unrelated visual semantics in different perception fields and enhances multi-scale related semantics to support a multi-level refinement process. Both modules are based on our LA-Gate, which realizes cross-modal interaction meanwhile avoiding language's direct fusion into visual representations. LA-Gate is a powerful competitor of predominant Cross-Attention in VideoQA and fits our idea of semantic distillation. Therefore, we can realize high-performance VideoQA by just relying on question-related visual representations and reducing language shortcuts. Experiments on 8 VideoQA and text-to-video retrieval tasks demonstrate the effectiveness of our model. + +# Acknowledge + +This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Post-doctoral Excellent Program (Grant No. 2022234). + +# References + +[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2 +[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2, 3 +[3] Anonymous. LLaMA-excitor: General instruction tuning via indirect feature interaction. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2 +[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6 +[5] Dare A Baldwin, Jodie A Baird, Megan M Saylor, and M Angela Clark. Infants parse dynamic action. *Child development*, 72(3):708-717, 2001. 1 +[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5 +[7] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the" video" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 7 +[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 2 +[9] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 2 +[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3 +[11] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with a long-memory transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 503-521. Springer, 2022. 2, 3 +[12] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1574–1583, 2021. 2 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6, 7 +[14] Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for + +text comprehension. arXiv preprint arXiv:1606.01549, 2016.2 +[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 1, 7 +[16] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. arXiv preprint arXiv:2212.09522, 2022. 2, 3, 6 +[17] György Gergely, Harold Bekkering, and Ildiko Király. Rational imitation in preverbal infants. Nature, 415(6873):755-755, 2002. 1 +[18] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5006–5015, 2022. 3 +[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2 +[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3 +[21] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 6 +[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2 +[23] Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. Iterative answer prediction with pointeraugmented multimodal transformers for textvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9992-10002, 2020. 2 +[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5 +[25] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. arXiv preprint arXiv:2210.03929, 2022. 2, 6, 8 +[26] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 1, 7 +[27] Hyounghun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal + +localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.3 +[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 3 +[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1 +[30] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 3, 7 +[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 7 +[32] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2 +[33] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1902-1910, 2021. 5 +[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 2 +[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 1, 7 +[36] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8658–8665, 2019. 1, 3, 7 +[37] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 510-519, 2019. 2 +[38] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 6 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, + +Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1 +[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1 +[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2 +[42] Min Peng, Chongyang Wang, Yu Shi, and Xiang-Dong Zhou. Efficient end-to-end video question answering with pyramidal multimodal transformer. arXiv preprint arXiv:2302.02136, 2023. 2, 3, 5, 7 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4 +[44] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 5 +[45] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2 +[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1 +[47] Mykhailo Shvets, Wei Liu, and Alexander C Berg. Leveraging long-range temporal relationships between proposals for video object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9756-9764, 2019. 2 +[48] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. arXiv preprint arXiv:2210.06031, 2022. 2 +[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[50] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified videolanguage pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 5, 7 +[51] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wang-meng Zuo, and Qinghua Hu. Eca-net: Efficient channel + +attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 2 +[52] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale microexpression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2 +[53] Amanda L Woodward. Infants selectively encode the goal object of an actor's reach. Cognition, 69(1):1-34, 1998. 1 +[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 2 +[55] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 2 +[56] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2 +[57] Xing Wu, Chaochen Gao, Zijia Lin, Zhongyuan Wang, Jizhong Han, and Songlin Hu. Rap: Redundancy-aware video-language pre-training for text-video retrieval. arXiv preprint arXiv:2210.06881, 2022.3 +[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6 +[59] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2, 6 +[60] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2 +[61] Ceyuan Yang, Yinghao Xu, Jianping Shi, Bo Dai, and Bolei Zhou. Temporal pyramid network for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 591-600, 2020. 5 +[62] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6 +[63] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[64] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot: + +Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 2 +[65] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and yang: Balancing and answering binary visual questions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5014-5022, 2016. 2 +[66] Xinsong Zhang, Yan Zeng, Jipeng Zhang, and Hang Li. Toward building general foundation models for language, vision, and vision-language understanding tasks. arXiv preprint arXiv:2301.05065, 2023. 2 +[67] Bo Zou, Chao Yang, Chengbin Quan, and Youjian Zhao. Spaceclip: A vision-language pretraining framework with spatial reconstruction on text. In Proceedings of the 31st ACM International Conference on Multimedia, pages 519-528, 2023. 3 \ No newline at end of file diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/images.zip b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f9012811026600d8221000f10c97bf0ffe36209b --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b18e14b13aaa0c890ffc9946a0eca478259e2fa985272d5077015399366cde3 +size 544175 diff --git a/2024/Language-aware Visual Semantic Distillation for Video Question Answering/layout.json b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6a6813047498e94ddf5ff96a939533b664c86dcf --- /dev/null +++ b/2024/Language-aware Visual Semantic Distillation for Video Question Answering/layout.json @@ -0,0 +1,10595 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 62, + 103, + 532, + 121 + ], + "type": "text", + "content": "Language-aware Visual Semantic Distillation for Video Question Answering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 128, + 124, + 177, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 124, + 177, + 137 + ], + "spans": [ + { + "bbox": [ + 128, + 124, + 177, + 137 + ], + "type": "text", + "content": "Bo Zou *" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 138, + 203, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 138, + 203, + 166 + ], + "spans": [ + { + "bbox": [ + 103, + 138, + 203, + 166 + ], + "type": "text", + "content": "Tsinghua University Beijing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 168, + 230, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 168, + 230, + 179 + ], + "spans": [ + { + "bbox": [ + 77, + 168, + 230, + 179 + ], + "type": "text", + "content": "zoub21@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 281, + 124, + 347, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 124, + 347, + 137 + ], + "spans": [ + { + "bbox": [ + 281, + 124, + 347, + 137 + ], + "type": "text", + "content": "Chao Yang *" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 255, + 138, + 373, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 138, + 373, + 166 + ], + "spans": [ + { + "bbox": [ + 255, + 138, + 373, + 166 + ], + "type": "text", + "content": "Shanghai AI Laboratory Shanghai, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 256, + 168, + 372, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 256, + 168, + 372, + 179 + ], + "spans": [ + { + "bbox": [ + 256, + 168, + 372, + 179 + ], + "type": "text", + "content": "yangchao@pjlab.org.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 438, + 124, + 479, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 124, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 438, + 124, + 479, + 137 + ], + "type": "text", + "content": "Yu Qiao" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 138, + 517, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 138, + 517, + 166 + ], + "spans": [ + { + "bbox": [ + 399, + 138, + 517, + 166 + ], + "type": "text", + "content": "Shanghai AI Laboratory Shanghai, China" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 406, + 168, + 510, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 168, + 510, + 179 + ], + "spans": [ + { + "bbox": [ + 406, + 168, + 510, + 179 + ], + "type": "text", + "content": "qiaoyu@pjlab.org.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 164, + 186, + 242, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 186, + 242, + 199 + ], + "spans": [ + { + "bbox": [ + 164, + 186, + 242, + 199 + ], + "type": "text", + "content": "Chengbin Quan" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 154, + 201, + 252, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 201, + 252, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 252, + 213 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 167, + 214, + 239, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 214, + 239, + 228 + ], + "spans": [ + { + "bbox": [ + 167, + 214, + 239, + 228 + ], + "type": "text", + "content": "Beijing, China" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 230, + 263, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 230, + 263, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 230, + 263, + 241 + ], + "type": "text", + "content": "quancb@tsinghua.edu.cn" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 337, + 186, + 413, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 186, + 413, + 200 + ], + "spans": [ + { + "bbox": [ + 337, + 186, + 413, + 200 + ], + "type": "text", + "content": "Youjian Zhao†" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 201, + 425, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 201, + 425, + 213 + ], + "spans": [ + { + "bbox": [ + 326, + 201, + 425, + 213 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 214, + 440, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 214, + 440, + 228 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 440, + 228 + ], + "type": "text", + "content": "Zhongguancun Laboratory" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 228, + 411, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 228, + 411, + 241 + ], + "spans": [ + { + "bbox": [ + 340, + 228, + 411, + 241 + ], + "type": "text", + "content": "Beijing, China" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 301, + 244, + 447, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 244, + 447, + 255 + ], + "spans": [ + { + "bbox": [ + 301, + 244, + 447, + 255 + ], + "type": "text", + "content": "zhaoyoujian@tsinghua.edu.cn" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 143, + 286, + 191, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 286, + 191, + 299 + ], + "spans": [ + { + "bbox": [ + 143, + 286, + 191, + 299 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 312, + 290, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 290, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 290, + 672 + ], + "type": "text", + "content": "Significant progress in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although image-language models can efficiently represent both video and language branches, they typically employ goal-free vision perception and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure extracting multi-grained visual semantics associated with the questions. We conduct evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset." + } + ] + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 321, + 281, + 532, + 313 + ], + "blocks": [ + { + "bbox": [ + 321, + 281, + 532, + 313 + ], + "lines": [ + { + "bbox": [ + 321, + 281, + 532, + 313 + ], + "spans": [ + { + "bbox": [ + 321, + 281, + 532, + 313 + ], + "type": "image", + "image_path": "d90d72df9833dde436ea53e7778a2f58cb4e42c69df533c70fbf3be776f4b073.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 314, + 334, + 319 + ], + "lines": [ + { + "bbox": [ + 327, + 314, + 334, + 319 + ], + "spans": [ + { + "bbox": [ + 327, + 314, + 334, + 319 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 314, + 501, + 320 + ], + "lines": [ + { + "bbox": [ + 460, + 314, + 501, + 320 + ], + "spans": [ + { + "bbox": [ + 460, + 314, + 501, + 320 + ], + "type": "text", + "content": "Answer: Car radio" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 320, + 321, + 532, + 360 + ], + "blocks": [ + { + "bbox": [ + 340, + 314, + 454, + 320 + ], + "lines": [ + { + "bbox": [ + 340, + 314, + 454, + 320 + ], + "spans": [ + { + "bbox": [ + 340, + 314, + 454, + 320 + ], + "type": "text", + "content": "Question: The man is singing along with what object?" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 321, + 532, + 360 + ], + "lines": [ + { + "bbox": [ + 320, + 321, + 532, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 321, + 532, + 360 + ], + "type": "image", + "image_path": "8ae326244236e0ac116abb45449c8ee4044bac881b2673645ff8ebcd0f9a77a9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 361, + 452, + 367 + ], + "lines": [ + { + "bbox": [ + 340, + 361, + 452, + 367 + ], + "spans": [ + { + "bbox": [ + 340, + 361, + 452, + 367 + ], + "type": "text", + "content": "Question: What object is taken out from the cabinet?" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 460, + 361, + 489, + 367 + ], + "lines": [ + { + "bbox": [ + 460, + 361, + 489, + 367 + ], + "spans": [ + { + "bbox": [ + 460, + 361, + 489, + 367 + ], + "type": "text", + "content": "Answer: Cup" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 320, + 368, + 531, + 399 + ], + "blocks": [ + { + "bbox": [ + 326, + 361, + 334, + 367 + ], + "lines": [ + { + "bbox": [ + 326, + 361, + 334, + 367 + ], + "spans": [ + { + "bbox": [ + 326, + 361, + 334, + 367 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 368, + 531, + 399 + ], + "lines": [ + { + "bbox": [ + 320, + 368, + 531, + 399 + ], + "spans": [ + { + "bbox": [ + 320, + 368, + 531, + 399 + ], + "type": "image", + "image_path": "3ff03425bcc0e91988c86e4cca1be3106d4dff9a6ce32a827e02fd01f04f3165.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 401, + 333, + 407 + ], + "lines": [ + { + "bbox": [ + 326, + 401, + 333, + 407 + ], + "spans": [ + { + "bbox": [ + 326, + 401, + 333, + 407 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 401, + 531, + 407 + ], + "lines": [ + { + "bbox": [ + 340, + 401, + 531, + 407 + ], + "spans": [ + { + "bbox": [ + 340, + 401, + 531, + 407 + ], + "type": "text", + "content": "Question: The person put the fish in the bowl into what object on the table? Answer: Tank" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 321, + 409, + 531, + 472 + ], + "blocks": [ + { + "bbox": [ + 321, + 409, + 531, + 472 + ], + "lines": [ + { + "bbox": [ + 321, + 409, + 531, + 472 + ], + "spans": [ + { + "bbox": [ + 321, + 409, + 531, + 472 + ], + "type": "image", + "image_path": "7c73333246d7aaaccab8845a932cced907171b2f90b6dc4913234b4565c4e9ce.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 473, + 334, + 479 + ], + "lines": [ + { + "bbox": [ + 325, + 473, + 334, + 479 + ], + "spans": [ + { + "bbox": [ + 325, + 473, + 334, + 479 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 473, + 447, + 479 + ], + "lines": [ + { + "bbox": [ + 339, + 473, + 447, + 479 + ], + "spans": [ + { + "bbox": [ + 339, + 473, + 447, + 479 + ], + "type": "text", + "content": "Question: What color is the banana in the photo?" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 489, + 547, + 534 + ], + "lines": [ + { + "bbox": [ + 304, + 489, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 547, + 534 + ], + "type": "text", + "content": "Figure 1. Challenges of goal-free VideoQA models. They can not efficiently handle (a) Long-term dependencies, (b) Multi-events, and (c) Multi-scale semantics in the videos. They also suffer from (d) language prior phenomenon in training question-answer pairs." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 546, + 387, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 387, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 387, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 304, + 567, + 548, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 548, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 548, + 663 + ], + "type": "text", + "content": "In recent years, large-scale video-and-language pretraining has seen remarkable progress. Most modern video-language understanding models [4, 15, 26, 31, 35, 36, 50] independently encode uni-modality and then fuse them. Human motion perception studies [5, 17, 53] suggest that humans perceive motion and environments as goal-driven behavior. This discrepancy leads to several issues, especially in long-form video understanding." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 304, + 665, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 548, + 714 + ], + "type": "text", + "content": "First, goal-free video representation struggles with long-term dependencies and multi-event reasoning. While enabled by progress in cross-modal pretraining[29, 43, 46] and datasets[39, 40], current methods excel at question an" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 16 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 693, + 211, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 211, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 211, + 703 + ], + "type": "text", + "content": "* Equal contribution, † Corresponding author" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 64, + 704, + 262, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 704, + 262, + 712 + ], + "spans": [ + { + "bbox": [ + 64, + 704, + 262, + 712 + ], + "type": "text", + "content": "This work was done during an internship at Shanghai AI Lab." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27113" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 324 + ], + "type": "text", + "content": "swering for images and short videos, but long-form videos contain many clips irrelevant or redundant to the questions, interfering with overall understanding. Encoding long videos also brings immense computational costs. Second, accurate semantic reasoning usually relies on multi-scale perception from local-spatial regions to global temporal dynamics. Goal-free methods for multi-scale visual modeling require custom submodels[9, 52, 56] for each scale or extra modalities like bounding boxes and OCR features[23, 66] but large-scale pretraining makes these approaches inefficient or infeasible. Third, goal-free video embedding requires a multimodal fusion module to synthesize questions and visual embeddings to predict answers. Incorporating questions directly can lead to shortcut solutions [1, 2, 8, 12, 19, 25, 41, 65], which means utilizing obvious clues in questions (mainly in data distribution and the relation between keywords) that exhibit more reliability in answer prediction than complicated visual reasoning, especially in the early training phases. This phenomenon is also known as language bias, which often causes major performance gaps in out-of-distribution tests. The above challenges are visualized in Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 326, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 326, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 326, + 289, + 494 + ], + "type": "text", + "content": "To address these issues, we propose a language-aware (goal-driven) visual semantic distillation framework called VideoDistill. Semantic distillation means the visual encoder must embed relevant frames of questions and focus on question-related multi-scale visual semantics. Closely resembles human behavior, semantic distillation functions like keeping the goal (question) in mind and conducting meaningful and precise visual reasoning. Differing from previous VideoQA frameworks [3, 31, 32, 42, 45, 59, 60], VideoDistill can generate answers only from question-related visual embeddings (without additional textual embeddings) since the interaction of text and vision is adequate during semantic distillation. This feature also enhances the significance of visual reasoning when generating answers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 496, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 496, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 289, + 700 + ], + "type": "text", + "content": "To realize semantic distillation, we first introduce Language-Aware Gate (LA-Gate), a multi-head cross-gating mechanism for cross-model interaction, which is inspired by self-gating and gated attention[14, 22, 37, 51] and is integrable into any existing transformer-based models like[4, 50]. LA-Gates compute questions' dependencies on video patch embeddings and depress or excite corresponding patches in subsequent attention layers. The proposed LA-Gate is not only the key to semantic distillation but also a new efficient modality fusion method, which is a powerful competitor of predominant Cross-Attention in VideoQA. We discuss the differences between LA-Gate and attention mechanisms in Section 3.2 and summarize three merits of LA-Gate: (1) It can alleviate the influence of language bias by avoiding the direct involvement of text embeddings. (2) It can better maintain local diversity within the video embeddings. (3) It enhances the interpretability of the modality fusion process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 701, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 712 + ], + "type": "text", + "content": "VideoDistill has two LA-Gate-based modules. The first" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 167 + ], + "type": "text", + "content": "is a differentiable sparse sampling module, which uses pretrained image-language models like CLIP[43] to encode frames, then performs goal-driven frame sampling to remarkably reduce subsequent spatial-temporal attention overhead and naturally avoid long-term dependencies and multi-event reasoning by retaining only question-related frames. It also provides our framework with a good nature that is insensitive to the number of sampled frames (see section 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 168, + 547, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 262 + ], + "type": "text", + "content": "The second is a vision refinement module eliminating unrelated visual semantics at different perceptual levels and enhancing related multi-scale semantics to support multilevel refinement. It encodes sparsely sampled frames into question-related global embeddings for generating answers. This module brings obvious performance boosts, especially on object-related questions (see section 4.5). Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 266, + 547, + 444 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 266, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 266, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 547, + 337 + ], + "type": "text", + "content": "- We propose the Language-Aware Gate (LA-Gate) that enables interacting vision with language meanwhile not directly introducing language into visual representations. It can alleviate the influence of language bias, better maintain local diversity within the video embeddings, and is more interpretable compared with predominant Cross-Attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 338, + 547, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 547, + 409 + ], + "type": "text", + "content": "- Based on LA-Gate, we propose a differentiable sparse sampling module to capture question-related frames and a vision refinement module to emphasize multi-scale question-related semantics. They can benefit VideoQA models on the stability under various numbers of sampled frames and the capacity to understand multi-scale objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 410, + 547, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 410, + 547, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 547, + 444 + ], + "type": "text", + "content": "- Our model achieves new state-of-the-art performance on a wide range of downstream VideoQA tasks and text-to-video retrieval tasks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 449, + 398, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 398, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 398, + 460 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 468, + 455, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 455, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 455, + 480 + ], + "type": "text", + "content": "2.1. Long-form Video modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 548, + 715 + ], + "type": "text", + "content": "Long-form understanding has recently become a new research hot spot [11, 16, 31, 34, 47, 48, 54, 55, 59, 64]. Current methods develop in two directions to overcome unique challenges not encountered by previous short-term video understanding works. The first direction is to encode long-term dependencies better. LF-VILA [48] proposes hierarchical temporal window attention that begins with a small window learning the attention between adjacent frames, then gradually expands the window size to learn high-level representation. HERO [34] and MERLOT [64] predict the order of shuffled frames to understand sequential characteristics. The second direction aims to reduce the increasing computation cost of encoding longer videos. Since the common practice of randomly selecting 3 or 4 frames per video regardless of length is not suited for long videos, ClipBert [31] randomly samples a sequence of segments from a video and then aggregates their predictions. MIST [16] improves ClipBert by selecting both video segments and frame regions to be encoded and reduces computation costs further." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27114" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 69, + 509, + 310 + ], + "blocks": [ + { + "bbox": [ + 85, + 69, + 509, + 310 + ], + "lines": [ + { + "bbox": [ + 85, + 69, + 509, + 310 + ], + "spans": [ + { + "bbox": [ + 85, + 69, + 509, + 310 + ], + "type": "image", + "image_path": "626fbdfe04d7fdff432f14ca2aa8decdc8648fb266b97ed7d1b7e129e299200e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 547, + 351 + ], + "type": "text", + "content": "Figure 2. Overview of VideoDistill. VideoDistill first densely samples video frames and utilizes a pre-trained image-language encoder to extract features, then sparsely samples a small number of question-related frames by a differentiable sparse sampling module. Finally, VideoDistill uses a vision refinement module to emphasize necessary multi-scale visual semantics in selected frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 392 + ], + "type": "text", + "content": "TALLFormer [11] proposes short-term feature extraction and long-term memory mechanisms that avoid repetitive calculations during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 392, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 290, + 559 + ], + "type": "text", + "content": "Similar to the second category, we reduce the workload by decreasing inputs, but our goal-driven differentiable sparse sampling works directly on frames, the minimum composition of the video, rather than rough video segments. Kim et al. [27] propose to use a self-gating mechanism to realize frame sampling, but it is a goal-free sampling method. Also, unlike random selection [31] or simple selector based on similarity[16], our learnable sampler excavates the semantic relationships between visual candidates. In addition, sparsely sampled frames from our sampler can be considered a concise summary of the long video sequence. Therefore, the following inference can perform similarly to models on short-form inputs. Thus, VideoDistill does not make efforts in the first direction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 573, + 241, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 241, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 241, + 586 + ], + "type": "text", + "content": "2.2. Video-Language Fusion in VideoQA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "The answer generation of VideoQA can be formulated as the interaction of the two modalities. Some methods propose fusing language and vision in the feature extraction stage for a more interactive combination. PSAC [36] combines regular visual self-attention with visual-linguistic co-attention. HCRN [30] supports conditioning video features with linguistic clues as a context factor in every inference stage. PMT [42] adopts a pyramidal video-language interaction. MCAN[63] utilizes Guided-Attention (V2T cross-attention in this paper) to fuse language into visual feature extraction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 437 + ], + "type": "text", + "content": "Unlike these techniques, our VideoDistill only keeps the question \"in mind.\" We do not directly fuse the question information into the video embedding, thus generating \"purer\" visual embeddings. Since the answer prediction is only based on visual embeddings, VideoDistill can alleviate just assuming the answer from the question and follow the \"look and answer\" criterion [2]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 438, + 547, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 535 + ], + "type": "text", + "content": "Besides VideoDistill, some existing methods can also be considered as indirect fusion. Contrastive learning methods like [10, 18, 20, 43, 57, 67] pull the matched pairs closer to interact vision with language. However, they do not demonstrate the same level of performance in tasks other than retrieval [28] due to ignoring the misalignment of vision and language. The method of fine-grained contrast to mitigate the misalignment has become a recent research hot spot." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 546, + 389, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 389, + 560 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 389, + 560 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 567, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 662 + ], + "type": "text", + "content": "We introduce VideoDistill, a pretraining framework that enables goal-driven VideoQA relying on question-related frames and their multi-scale semantics. Figure 2 gives an overview of VideoDistill. VideoDistill consists of two sub-modules: Differentiable Sparse Sampling and Vision Refinement. Both sub-modules are built upon our proposed language-aware gate (LA-Gate) to perform in a goal-driven manner." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 670, + 486, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 486, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 486, + 685 + ], + "type": "text", + "content": "3.1. Vision-Language Representations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 689, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 548, + 715 + ], + "type": "text", + "content": "VideoDistill begins with densely sampling " + }, + { + "bbox": [ + 304, + 689, + 548, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 689, + 548, + 715 + ], + "type": "text", + "content": " uniformly distributed frames and embedding them into frame represent" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27115" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 219, + 188 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 219, + 188 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 219, + 188 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 219, + 188 + ], + "type": "image", + "image_path": "ca32b2d5e3347226dfc993cccb89c381cf492851e87fe4e9cf0b582db00a2fa8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 198, + 456, + 209 + ], + "lines": [ + { + "bbox": [ + 136, + 198, + 456, + 209 + ], + "spans": [ + { + "bbox": [ + 136, + 198, + 456, + 209 + ], + "type": "text", + "content": "Figure 3. Illustrations of Self-Attention, Cross Attention, and our LA-Gate mechanisms" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 375, + 188 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 375, + 188 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 375, + 188 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 375, + 188 + ], + "type": "image", + "image_path": "659c9c9cce9b2386df7b74c2cb763293a40f6774e57ab5fc58391923618fc584.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 521, + 187 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 521, + 187 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 521, + 187 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 521, + 187 + ], + "type": "image", + "image_path": "d94d1263b62a166d54a368a8b7123f1b16aeaf63cf6659859fd4922d8510da66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": "tations. We adopt the pretrained image-language encoder CLIP [43] with frozen parameters in the visual branch to reduce the computational overhead. CLIP divides " + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": " frames into " + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "inline_equation", + "content": "N \\times n \\times n" + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": " patches, and extracts patch embeddings " + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{patch}} \\in \\mathbb{R}^{N \\times n^2 \\times D}" + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": " and CLS tokens " + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": " that represent the global understanding of frames, where " + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 219, + 288, + 303 + ], + "type": "text", + "content": " is the dimension of representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "text", + "content": "For the Language branch, A text is first tokenized and then fed into the pretrained text encoder of CLIP to generate word-level embeddings " + }, + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{word}} \\in \\mathbb{R}^{M \\times D}" + }, + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "text", + "content": " and a sentence-level token " + }, + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 304, + 287, + 387 + ], + "type": "text", + "content": ", where M is the length of embeddings. Note that parameters in the text encoder are updated during pretraining and downstream finetuning to mitigate the linguistic gaps between datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 396, + 175, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 396, + 175, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 396, + 175, + 407 + ], + "type": "text", + "content": "3.2. Language-Aware Gate" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": "We introduce a reusable unit, termed language-aware gate (LA-Gate), which takes an arbitrary visual representation " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": " and a sentence-level language representation " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": " as input. LA-Gate generates " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{distill}}" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": ", an output sequence of language-aware visual representation with the same dimension as " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": ", by exciting or depressing the components of " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": " according to their similarities with " + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 414, + 287, + 545 + ], + "type": "text", + "content": ". For a better understanding of LA-Gate, we illustrate self-attention, cross-attention, and LA-Gate in Figure 3. We first discuss why commonly used cross-attention is not suitable for our purposes, then point out the design concept and benefits of LA-Gate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 666 + ], + "type": "text", + "content": "Figure 3(b) illustrates two types of cross-attention: Text-to-Video (T2V, text as Query) and Video-to-Text (V2T, video as Query). Since outputs of the attention have the same length as Query and carry information sourced from Value, applying cross-attention in VideoQA will have two problems. Firstly, T2V is unsuitable for the modality interaction in the middle layers as it alters the shape of visual representations. Although V2T can maintain the shape, its output will carry the info directly from the text (like reconstructing text representations into the shape of visual input)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": "The structure of LA-Gate is shown in Figure 3(c). There are two divergences between V2T and LA-Gate. (1) When text input length " + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "inline_equation", + "content": "L > 1" + }, + { + "bbox": [ + 46, + 666, + 288, + 714 + ], + "type": "text", + "content": ", we sum the similarity matrix to form a vector of importance for rows of the video input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "text", + "content": " functions like the number of multi-heads in the attention mechanism, and each text representation will partially dictate the importance of each row in the video input. (Note that in VideoDistill, we always have " + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "inline_equation", + "content": "L = 1" + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "text", + "content": " since the text input is " + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 219, + 545, + 314 + ], + "type": "text", + "content": ".) (2) We expand the importance vector (repeat D times) to match the shape of the video input. Finally, we apply an element-wise product rather than a dot product on the importance matrix and the Value." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 546, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 546, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 546, + 375 + ], + "type": "text", + "content": "The first merit of LA-Gate is the text input only controls the importance of each visual representation, and the output does not directly involve text. Thus, the answer decoder can make decisions based on \"purer\" visual semantics and alleviate shortcut solutions hidden in texts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 378, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 546 + ], + "type": "text", + "content": "A more significant characteristic of LA-Gate in general modality fusion usage: it can better maintain local diversity within the visual input. In Figure 3 (a) and (b), the output displays rows in mixed colors, representing the weighted sum of rows in the Value. Cross-attention will gather information from all rows in the Value for each output token and present more global attributes than LA-Gate since LA-Gate does not blend all features. In Figure 4(c), we maintain the uniqueness of colors for LA-Gate in each output row because each row is produced by multiplying the corresponding row in the Value and a scalar of importance. This characteristic makes LA-Gate fit more into our idea of information distillation because the language-aware distillation of a local area should be irrelevant to other regions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 547, + 668 + ], + "type": "text", + "content": "The last benefit of LA-Gate is we can enhance the interpretability of modality fusion by incorporating LA-Gate with self-attention layers. Due to the skip connection between the input and the output, commonly used V2T cross-attention brutally adds a reconstructed text (maybe linearly projected) into the visual input to fuse vision and language. The physical meaning of the structure is difficult to explain. By contrast, LA-Gate first distills the visual input by emphasizing language-related visual semantics, and then the interaction within visual representations is performed by self-attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": "The implementation of LA-Gate is visualized in Figure 4 (we demonstrate how LA-Gate works in a vision refinement block, where " + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": " consists of frame patches from " + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": " selected frames and a video-level CLS token). Assume LA-Gate" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27116" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": "receives " + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^{m \\times D}" + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": " is the sequence length, and " + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": ". We first produce the key from the language representation " + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": " and queries and values from the visual representation " + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 72, + 288, + 128 + ], + "type": "text", + "content": ". Then, we calculate cosine distances between each query and the key:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 134, + 288, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 134, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 61, + 134, + 288, + 147 + ], + "type": "interline_equation", + "content": "\\mathrm {k e y} = \\mathrm {w} _ {\\mathrm {k}} \\left(t _ {\\mathrm {c l s}}\\right), \\text {q u e r y} _ {i} = \\mathrm {w} _ {\\mathrm {q}} \\left(v _ {i}\\right), \\text {v a l u e} _ {i} = \\mathrm {w} _ {\\mathrm {v}} \\left(v _ {i}\\right) \\tag {1}", + "image_path": "2521050e583fd6ea473f21ef3048227c26c674dda5a605faf4e750103db8651d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 160, + 288, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 288, + 185 + ], + "type": "interline_equation", + "content": "\\operatorname {d i s t} _ {i} = \\operatorname {r e p e a t} \\left(\\frac {\\operatorname {q u e r y} _ {i} \\cdot \\operatorname {k e y}}{\\| \\operatorname {q u e r y} _ {i} \\| \\times \\| \\operatorname {k e y} \\|}\\right), \\tag {2}", + "image_path": "5e4e80518c5f73b62db993d2f8a1c6e0587a3bbe0f901cf2482db241adf99fde.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "v_{i} \\in \\mathbb{R}^{D}, i \\in [1, m]" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": " is a feature vector in " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{w_k}" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{w_q}" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{w_v}" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": " are trainable linear projection layers, and repeat means expanding a scalar into a vector with the dimension of " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}_i" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": " reflects the correlation between vision and language, we treat it as an importance coefficient of " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "\\mathrm{value}_i" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": ". Then, we generate distilled " + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 187, + 288, + 254 + ], + "type": "text", + "content": " as formulated:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 260, + 288, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 260, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 109, + 260, + 288, + 273 + ], + "type": "interline_equation", + "content": "v _ {\\text {d i s t i l l}} ^ {i} = \\mathrm {w} _ {\\mathrm {o}} \\left(\\operatorname {d i s t} _ {i} \\odot \\operatorname {v a l u e} _ {i}\\right), \\tag {3}", + "image_path": "1d4a8d4f70f0cfb4fd9b5e28bece03bbeb594682b83c4cd98b939cd8b8168a16.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": " denotes element-wise production and " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{w_o}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": " is the linear output layer. In practice, we perform a standard multihead attention setting [49] on " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{w_k}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{w_q}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{w_v}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{W_o}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": ", and concatenate outputs of each head to form " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{distill}}^i" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": ". Although, we only illustrate a single attention head in Figure 2 for simplicity. Also, we apply an input skip connection on " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{distill}}^i" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": " for better training stability and convergence. Finally, LA-Gate outputs " + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{distill}} = \\left\\{v_{\\mathrm{distill}}^i | i \\in [1, m]\\right\\}" + }, + { + "bbox": [ + 46, + 279, + 288, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 393, + 217, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 217, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 217, + 407 + ], + "type": "text", + "content": "3.3. Differentiable Sparse Sampling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " densely sampled frames, VideoDistill further adaptively picks out " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "K(K < N)" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " language-related frames. To this end, we utilize stacked frame sampling blocks (FS-Blocks, " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " layers) which take as input " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "t_\\mathrm{cls}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " to perform a top-" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " selection. Since " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " of frames are separately extracted, to indicate their temporal positions in the whole video, we add temporal embedding " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "\\mathrm{T_f} \\in \\{\\phi_{\\mathrm{T}}(\\mathrm{f}) | \\mathrm{f} \\in [1,\\mathrm{N}]\\}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " for each of them according to their frame index. After adding temporal information, each FS-Block first distills " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "inline_equation", + "content": "t_\\mathrm{cls}" + }, + { + "bbox": [ + 46, + 412, + 287, + 578 + ], + "type": "text", + "content": " by LA-Gate to emphasize question-related frames, then performs an interframe interaction through standard multi-head attention. Figure 4 shows the architecture of FS-Blocks. We borrow the form of skip connections from the divided block in [4]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": "To realize the differentiable selection, we first project " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^{L}" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": ", the output of the " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": "-th FS-Block with the dimension identical to " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": ", onto a feature space with the dimension of " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": " by a linear layer " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": " is the number of frames to be selected. Then, we conduct Gumbel-Softmax sampling [24] on each " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "x^{k} \\in \\mathbb{R}^{N}, k \\in [1, K]" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": ", the row of " + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "inline_equation", + "content": "\\mathrm{W_s}\\left(v_{\\mathrm{cls}}^L\\right)" + }, + { + "bbox": [ + 46, + 578, + 288, + 656 + ], + "type": "text", + "content": ". The procedures are formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 96, + 662, + 288, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 662, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 96, + 662, + 288, + 687 + ], + "type": "interline_equation", + "content": "y _ {\\mathrm {s o f t}} ^ {k} = \\operatorname {s o f t m a x} \\left(\\frac {x ^ {k} + g u m b e l s}{\\tau}\\right), \\tag {4}", + "image_path": "efaeb49154b3dfed4735e3f242f8c4a188c9f238a8774be86735388fb975e42f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 522, + 183 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 522, + 183 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 522, + 183 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 522, + 183 + ], + "type": "image", + "image_path": "d3ab6d664f307c3dff80b4e8eb181ad725bb754d880c0e5f60f1f54ee3b229a1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 184, + 423, + 294 + ], + "blocks": [ + { + "bbox": [ + 331, + 184, + 423, + 294 + ], + "lines": [ + { + "bbox": [ + 331, + 184, + 423, + 294 + ], + "spans": [ + { + "bbox": [ + 331, + 184, + 423, + 294 + ], + "type": "image", + "image_path": "f3cdc518f4e623ccb44883fecb4fac28123a40d223674c4f08d7be1361939db6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 303, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 303, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 303, + 545, + 324 + ], + "type": "text", + "content": "Figure 4. Architectures of Language-Aware Gate, Frame Sampling Block, and Vision Refinement Block." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 184, + 521, + 294 + ], + "blocks": [ + { + "bbox": [ + 425, + 184, + 521, + 294 + ], + "lines": [ + { + "bbox": [ + 425, + 184, + 521, + 294 + ], + "spans": [ + { + "bbox": [ + 425, + 184, + 521, + 294 + ], + "type": "image", + "image_path": "c0f2f9654508f02079edec81448f0484474b342476edab5d6188e734956e5484.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 332, + 545, + 351 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 332, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 353, + 332, + 545, + 351 + ], + "type": "interline_equation", + "content": "y _ {\\text {h a r d}} ^ {k} = \\operatorname {o n e h o t} \\left(\\arg \\max \\left(y _ {\\text {s o f t}} ^ {k}\\right)\\right), \\tag {5}", + "image_path": "96bf0e1bd33b2a1c3b1329649274c45afa05da4cd63682f05a4e834164105195.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 355, + 545, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 355, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 341, + 355, + 545, + 372 + ], + "type": "interline_equation", + "content": "m a s k ^ {k} = y _ {\\text {h a r d}} ^ {k} + y _ {\\text {s o f t}} ^ {k} - \\operatorname {s t o p g r a d} \\left(y _ {\\text {s o f t}} ^ {k}\\right), \\tag {6}", + "image_path": "15208acc593dec22ff63f1c3a4a45d79a6702c297b12f4b68c2a1651f24720c0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": "where gumbels is a noise sampled from the Gumbel distribution with " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\mu = 0" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{soft}}^k" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " is the possibility of frames to be selected as the " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": "-th language-related frame and " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{hard}}^k" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " reflects the index of the chosen frame. Since the argmax operation has no gradient, we adopt a code-level trick in equation 6 to generate mask " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "= \\{mask^{k}|k\\in [1,K]\\}" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " which can properly pass the gradient. Finally, we apply mask on " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " and generate " + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{patch}}^K\\in \\mathbb{R}^{K\\times n^2\\times D}" + }, + { + "bbox": [ + 304, + 375, + 546, + 483 + ], + "type": "text", + "content": " for the following vision refinement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 491, + 414, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 414, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 414, + 502 + ], + "type": "text", + "content": "3.4. Vision Refinement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": "Given the selected patch embeddings " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{patch}}^{K}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": " and the language representation " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": ", this module generates video-level representation " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^{*}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": ", which synthesizes multi-scale language-related visual semantics. " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^{*}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": " is set as a learnable token with the dimension of " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": " and is concatenated to " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{patch}}^{K}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": " as the visual input " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{*}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": ". We combine the LA-Gate and any existing spatial-temporal self-attention layer [4, 6, 50] to form a vision refinement block as shown in Figure 4. When iteratively applying vision refinement blocks on " + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "inline_equation", + "content": "v_{*}" + }, + { + "bbox": [ + 304, + 509, + 547, + 665 + ], + "type": "text", + "content": ", LA-Gates spontaneously distill language-related visual semantics in different perceptive fields. Since the effective perceptive field gradually expands in stacked blocks [44], our module can consider multi-scale objects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": "Unlike predominant pyramidal models [33, 42, 61] for multi-scale reasoning that gathers the intermediate results of each encoding stage as final representation, we empirically adopt only the " + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 304, + 665, + 547, + 714 + ], + "type": "text", + "content": " from the last vision refinement block." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 288, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 288, + 712 + ], + "type": "text", + "content": "The code will be available at https://zoubo9034.github.io/VideoDistill/" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27117" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 153, + 85 + ], + "type": "text", + "content": "3.5. Pretraining Tasks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 126 + ], + "type": "text", + "content": "We adopt the two pretraining tasks to facilitate cross-modal interaction: Video-Text Matching (VTM) and Vision-Guided Masked Language Modeling (VG-MLM)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "text", + "content": "Video-Text Matching. To facilitate cross-modal interaction, we use VTM as the first pre-training task. VTM predicts whether input video-language pairs are matched. In practice, we randomly exchange the annotations of two input pairs (matched) in a mini-batch with a probability of 0.5. Then, we use a linear projection head on the top of " + }, + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "text", + "content": " to predict two-class matching logits " + }, + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 126, + 288, + 214 + ], + "type": "text", + "content": ", and formulate the VTM loss as negative log-likelihood:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 224, + 288, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 224, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 102, + 224, + 288, + 240 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V T M}} = - \\mathbb {E} _ {\\left(v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p (y \\mid v _ {\\mathrm {c l s}} ^ {*}), \\tag {7}", + "image_path": "896c2fc51f03c8d473abe2197478f50ea30c6cec0091468935cef50c3ccd7880.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 289, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 347 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 347 + ], + "type": "text", + "content": "We do not adopt predominant contrastive pretraining because this paradigm requires the contrast between all possible combinations of vision and language. However, for VideoDistill, the number of possible visual embeddings of each video is proportional to the number of annotations. This leads to a quadratic growth in computational overhead (see Appendix A for more details). Nevertheless, we still apply a contrastive constraint on matched pairs (w/o exchange) to stabilize the training:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 354, + 288, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 354, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 73, + 354, + 288, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C L}} = - \\sum_ {i \\in B} \\log \\frac {\\mathbb {I} (\\text {m a t c h e d}) \\exp (s (v _ {i} , t _ {i}) / \\tau)}{\\sum_ {j \\in B} \\exp (s (v _ {i} , t _ {j}) / \\tau)}, \\tag {8}", + "image_path": "ca121ee8d7a094ee26c28cb841c8c56e82b2586d96e2e5fa803da972b349fea8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": "where B is the batch size, " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "t_\\mathrm{cls}" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": "-th video-language pair, " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " is a indicate function of whether the video and the annotation in the " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": "-th pair are matched, " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " denotes a cosine similarity function, " + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 389, + 287, + 449 + ], + "type": "text", + "content": " is a temperature coefficient equals 0.07 in this paper." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": "Vision-Guided Masked Language Modeling. The commonly used MLM [13] task aims to predict the ground truth label of the masked token according to textual context. To better map the visual and the language representations in a fine-grained manner, we propose VG-MLM, which is applied only on the matched pairs (w/o exchange) and encourages predicting the masked tokens from the visual context. In particular, we first provide VideoDistill the CLS tokens " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{cls}}" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": " of unmasked annotations to sample frames and extract video representation " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": ". Then, we mask " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": " words in the annotations and encode the masked sentences. The outcome " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "w_i^{\\mathrm{M}}, i \\in [1, I]" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": ", which denotes the token of " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": "-th masked word, and " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": " are used to predict the " + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 449, + 288, + 604 + ], + "type": "text", + "content": "-th masked word. The objective of VG-MLM is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 609, + 287, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 609, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 61, + 609, + 287, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V G - M L M}} = - \\mathbb {E} _ {\\left(t _ {\\mathrm {c l s}}, v _ {\\mathrm {c l s}} ^ {*}\\right)} \\log p \\left(w _ {i} | \\mathrm {S G} \\left(w _ {i} ^ {\\mathrm {M}}\\right), v _ {\\mathrm {c l s}} ^ {*}\\right), \\tag {9}", + "image_path": "c2b4198af1332db58f0307041e0340467c59194d2e201337bdcedb742ffeebd0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": "where SG means stop gradient, " + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": " denote the logit of " + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": "-th masked word. We stop the gradient of " + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "inline_equation", + "content": "w_{i}^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": " to enhance the importance of visual clues in language modeling. Following the heuristics of Bert [13], we adopt the same masking strategy. Then we adopt a two-layer MLP on top of the combination of " + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "inline_equation", + "content": "w_{i}^{\\mathrm{M}}" + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^{*}" + }, + { + "bbox": [ + 46, + 635, + 289, + 714 + ], + "type": "text", + "content": " to generate the probability over the vocabulary, which is calculated as the negative log-likelihood loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "for the masked word. The overall loss of pretraining is the combination of VTM, VG-MLM, and CL:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 355, + 103, + 545, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 103, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 355, + 103, + 545, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {V T M}} + \\mathcal {L} _ {\\mathrm {V G - M L M}} + \\mathcal {L} _ {\\mathrm {C L}}, \\tag {10}", + "image_path": "d9038a8358d794094cba14bddcdc53d02c6405beef1fd3d1dfbe5d42e58e7d95.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 126, + 388, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 126, + 388, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 126, + 388, + 140 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "content": "In this section, conduct extensive experiments utilizing the pre-trained model on downstream various types of VideoQA tasks to verify the effectiveness of our proposed VideoDistill. We also transfer VideoDistill for video-text retrieval tasks to show the generalization power of the pre-trained model in Appendix C. Our pretraining set consists of three parts: (1) 3M video-caption pairs randomly sampled from generic dataset WebVid10M [4]. (2) 4.2M video-caption pairs randomly sampled from YouTube video dataset HD-VILA [59]. (3) 3.8M video-caption pairs from 1-st person view dataset EgoCLIP. We report results with " + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "inline_equation", + "content": "K = 16" + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "content": ". For open-ended datasets, train a MLP classification head on the top of " + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "inline_equation", + "content": "v_{\\mathrm{cls}}^*" + }, + { + "bbox": [ + 304, + 146, + 547, + 326 + ], + "type": "text", + "content": ". For multiple-choice datasets, we choose the answer with the maximum logit over the VTM head. More details are in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 334, + 416, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 416, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 416, + 346 + ], + "type": "text", + "content": "4.1. Generic VideoQA:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 352, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 547, + 413 + ], + "type": "text", + "content": "We first evaluate VideoDistill on the four commonly used VideoQA datasets: MSRVTT-QA [58], MSVD-QA [58], EgoMCQ [38] and MSRVTT-multiple-choice test [62]. Details of each dataset and results are in the supplementary material (Appendix C)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 420, + 432, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 420, + 432, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 432, + 433 + ], + "type": "text", + "content": "4.2. Long-Form VideoQA:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 438, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 547, + 521 + ], + "type": "text", + "content": "We evaluate our model on two recently proposed challenging datasets for the long-form VideoQA, namely EgoTaskQA[25] and AGQA[21]. We adopt the same finetuning setting with MSRVTT-QA and MSVD-QA. For fairness comparison, we report the results of VideoDistill w/o large-scale pertaining. Details of each dataset are in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": "Results. In the Table 2, our VideoDistill overwhelm previous SOTA method and bring " + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "inline_equation", + "content": "2.82\\%" + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": " overall performance gains. Remarkably, in the categories of multi-agent and descriptive, we outperform previous work with " + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "inline_equation", + "content": "5.59\\%" + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "inline_equation", + "content": "3.72\\%" + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": " improvement. They show VideoDistill can better understand the storyline and temporal relations in long videos. In the category of object, we achieve " + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "inline_equation", + "content": "9.33\\%" + }, + { + "bbox": [ + 304, + 522, + 548, + 713 + ], + "type": "text", + "content": " gains, which indicates VideoDistill can capture multi-scale objects. The limited performance in the category of action is due to the strong correlations between action, object, and their change. Most methods tend to over-fit this strong language bias without thorough task understanding. Although the design of VideoDistill avoids utilizing the shortcut solutions in language. In Table 1, The results in the gray color (AIO, Temp, and MIST-CLIP) are reported by [16]. If they follow the same evaluation metrics [21] as us, one of the reasons for the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27118" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 96, + 81, + 495, + 177 + ], + "blocks": [ + { + "bbox": [ + 164, + 71, + 428, + 81 + ], + "lines": [ + { + "bbox": [ + 164, + 71, + 428, + 81 + ], + "spans": [ + { + "bbox": [ + 164, + 71, + 428, + 81 + ], + "type": "text", + "content": "Table 1. QA accuracies of state-of-the-art methods on AGQA v2 test set." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 81, + 495, + 177 + ], + "lines": [ + { + "bbox": [ + 96, + 81, + 495, + 177 + ], + "spans": [ + { + "bbox": [ + 96, + 81, + 495, + 177 + ], + "type": "table", + "html": "
Question TypesMost LikelyPSACHMEHCRNAIO [50]Temp[ATP] [7]MIST-CLIPVideoDistill†
Object-relation9.3937.8437.4240.3348.3450.1551.6856.32
Relation-action50.0049.9549.9049.8648.9949.7667.1855.09
Object-action50.0050.0049.9749.8549.6646.2568.9955.12
Superlative21.0133.2033.2133.5537.5339.7842.0543.30
Sequencing49.7849.7849.7749.7049.6148.2567.2454.49
Exists50.0049.9449.9650.0150.8151.7960.3355.74
Duration comparison24.2745.2147.0343.8445.3649.5954.6249.08
Activity recognition5.524.145.435.5218.9718.9619.6910.16
All10.9940.1839.8942.1148.5949.7954.3955.80
", + "image_path": "b48d074f635ae7ac0307215976dba48f381e64f7b8330ca5f5cb6a2d25602e02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 96, + 208, + 495, + 345 + ], + "blocks": [ + { + "bbox": [ + 144, + 187, + 448, + 198 + ], + "lines": [ + { + "bbox": [ + 144, + 187, + 448, + 198 + ], + "spans": [ + { + "bbox": [ + 144, + 187, + 448, + 198 + ], + "type": "text", + "content": "Table 2. Performances on EgoTaskQA normal split. †denotes training from scratch." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 96, + 208, + 495, + 345 + ], + "lines": [ + { + "bbox": [ + 96, + 208, + 495, + 345 + ], + "spans": [ + { + "bbox": [ + 96, + 208, + 495, + 345 + ], + "type": "table", + "html": "
CategoryVisualBERT [35]PSAC [36]HME [15]HGA [26]HCRN [30]ClipBERT [31]VideoDistill†
Scopeworld39.7340.7641.9138.8244.2742.1547.32
intent44.5146.1948.9242.1249.7740.9452.53
multi-agent26.2930.5927.9823.4331.3627.6336.95
Typedescriptive41.9940.6341.4538.0443.4838.4547.20
predictive30.3731.9835.8825.5736.5631.5040.43
counterfactual41.9941.8944.1341.9448.0046.7549.64
explanatory37.4237.9938.8535.9740.6042.3942.53
Semanticaction15.0214.7514.9915.0814.9222.9116.35
object23.2636.5336.0519.0945.3121.8054.64
state59.2061.8963.4455.6568.2854.3672.37
change68.2765.0568.8768.3867.3866.5871.47
all37.9338.9040.1636.7742.2039.8745.02
", + "image_path": "5719547475562860f1dc3147fec3cdc8a7f01d08aecb1a79c78796848bab638d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "content": "performance gaps in some categories is they adopt frames with higher resolution " + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "inline_equation", + "content": "(448 \\times 448)" + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "content": " and more powerful encoder (CLIP-ViTB/32). Nevertheless, we still outperform current SOTA methods in Object-relation " + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "inline_equation", + "content": "(4.64\\%)" + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "content": " and Superlative " + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "inline_equation", + "content": "(1.25\\%)" + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "content": ", which require multi-scale and multi-event reasoning. We also improve the overall performance with " + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "inline_equation", + "content": "1.41\\%" + }, + { + "bbox": [ + 46, + 365, + 288, + 448 + ], + "type": "text", + "content": " gains." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 458, + 258, + 610 + ], + "blocks": [ + { + "bbox": [ + 66, + 458, + 258, + 610 + ], + "lines": [ + { + "bbox": [ + 66, + 458, + 258, + 610 + ], + "spans": [ + { + "bbox": [ + 66, + 458, + 258, + 610 + ], + "type": "image", + "image_path": "d9301845e814910e72d00c2a5c28328a919818c6893a3a9ea0c016d87bb86b32.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 613, + 251, + 623 + ], + "lines": [ + { + "bbox": [ + 82, + 613, + 251, + 623 + ], + "spans": [ + { + "bbox": [ + 82, + 613, + 251, + 623 + ], + "type": "text", + "content": "Figure 5. The impact of the number of frames." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 635, + 286, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 286, + 648 + ], + "type": "text", + "content": "4.3. The Impact of Differentiable Sparse Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 713 + ], + "type": "text", + "content": "Insensitive to the number of frames. We evaluate the performances on MSRVTT-QA with variable input frames. In Figure 5, we compare the performance change of VideoDistill (w/ & w/o sparse sampling module) with PMT [42] (other methods only report results on 16 frames, which are the best" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 519 + ], + "type": "text", + "content": "results), and there are two findings. First, the differentiable sparse sampling module makes our 4-frame variation on par with other methods with many more input frames. It is intuitive because other competitors require more frames to improve the change to capture necessary information for reasoning (a disadvantage of goal-free perception). This property can dramatically reduce the computational overhead required to achieve the same performance as previous methods. Second, we notice that too many frames will damage the answering accuracy since the emergence of unrelated information disturbs the reasoning process. However, our sparse sampling module can mitigate the dropping with increasing frames and make the model more robust." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 544 + ], + "type": "text", + "content": "Qualitative Results. We visualize the results of our sparse sampling module on EgoTaskQA in Appendix E." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 552, + 443, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 443, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 443, + 564 + ], + "type": "text", + "content": "4.4. The Impact of LA-Gate:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Prevent the use of language shortcuts. We come up with two methods to quantitatively analyze this characteristic. First, we test the language-only QA performances on EgoTaskQA normal split in Table 3, which means replacing the visual inputs with a static video (We also report the performances under a Gaussian noise input in Appendix E). The more declines, the better. When compared with the language-based model Bert [13], and the previous SOTA method HCRN [30], the decline of our VideoDistill is the most significant (especially in the object category). It demonstrates that VideoDistill relies more on vision and can avoid using language bias. Second, we test VideoDistill on Ego-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27119" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 98, + 83, + 495, + 204 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 499, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 499, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 499, + 82 + ], + "type": "text", + "content": "Table 3. Language-only QA results on the EgoTaskQA normal split. The more performance drops, the better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 98, + 83, + 495, + 204 + ], + "lines": [ + { + "bbox": [ + 98, + 83, + 495, + 204 + ], + "spans": [ + { + "bbox": [ + 98, + 83, + 495, + 204 + ], + "type": "table", + "html": "
CategoryBERTHCRN (w/o vision)VideoDistill (w/o vision)
Acc.ChangeAcc.ChangeAcc.Change
world36.28-8.7%30.17-31.9%18.78-60.3%
intent35.02-21.3%35.54-28.6%22.57-57.0%
multi-agent20.58-21.7%19.9-36.5%11.50-68.7%
descriptive34.55-17.7%29.97-31.1%21.19-55.1%
predictive24.75-18.5%18.32-49.9%7.42-81.6%
counterfactual41.3-1.6%41.1-14.4%39.70-20.0%
explanatory31.78-15.1%32.41-20.2%13.69-67.8%
action15.72+4.6%17.31+16.0%11.81-27.8%
object7.43-68%8.85-80.5%0-100.0%
state45.03-23.9%35.51-48.0%20.21-72.1%
change69.87+2.3%70.47+4.6%55.38-22.5%
", + "image_path": "f49882e5be8a49beb3a153594d32bd20b3dab83c0eb220686dffcfbf77790dfb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 99, + 217, + 495, + 285 + ], + "blocks": [ + { + "bbox": [ + 198, + 205, + 394, + 215 + ], + "lines": [ + { + "bbox": [ + 198, + 205, + 394, + 215 + ], + "spans": [ + { + "bbox": [ + 198, + 205, + 394, + 215 + ], + "type": "text", + "content": "Table 4. Analysis of the effectiveness of each module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 217, + 495, + 285 + ], + "lines": [ + { + "bbox": [ + 99, + 217, + 495, + 285 + ], + "spans": [ + { + "bbox": [ + 99, + 217, + 495, + 285 + ], + "type": "table", + "html": "
MethodModuleEgoTaskQAMSRVVTT-QA
SSVRCross-ATTLA-Gate
(a) BaselineXX--29.1023.92
(b) Uniform SamplingX✓X✓41.7040.92
(c) Soft SamplingX✓X✓38.9538.16
(d) w/o VR✓XX✓33.4426.51
(e) Cross-Attention✓✓✓X40.9441.17
(f) VideoDistill✓✓X✓45.0244.20
", + "image_path": "c6e185ffa0cd567c3d7c22bf3974d2235f5e5cb3fe35e1a9a7bb2fcf1fea2591.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 305, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 305, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 288, + 413 + ], + "type": "text", + "content": "TaskQA indirect split, which is motivated by the fact [25] that during task execution, actions, objects, and their changes are often strongly correlated. It leaves the chance for the model to perform well by simply over-fitting these strong correlations (language bias) without thorough task understanding. The indirect references can avoid these correlations. In Appendix E, we show that our VideoDistill has the least absolute performance change. It indicates that VideoDistill barely utilizes language bias in questions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 413, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 473 + ], + "type": "text", + "content": "A more effective fusion method than Cross-Attention. In Table 4 (e), we replace LA-Gates in our frameworks with Cross-Attention. When comparing (e) with (f), we can find that LA-Gate brings significant performance boosts on both test sets (4.08% and and 3.03%)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 479, + 228, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 228, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 228, + 491 + ], + "type": "text", + "content": "4.5. The Impact of Vision Refinement:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "content": "Performance boost in multi-scale object-related questions. The stacked video distillation blocks containing LA-Gate can obtain multi-scale question-related semantics. We observe huge improvements in Table 2 object category " + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "inline_equation", + "content": "(9.33\\%)" + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "content": " and Table 1 object-relation category " + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "inline_equation", + "content": "(4.64\\%)" + }, + { + "bbox": [ + 46, + 498, + 288, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 141, + 576 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": "We conduct ablation studies to verify the effectiveness of components and the rationality of parameter selection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Effectiveness of each component. We design four variants and report their performances in Table 4. (a) we remove the two modules and apply evenly sampled 16 frames as input and a 3-layer transformer to fuse the vision and language. (b) we remove differentiable sparse sampling and apply evenly sampled 16 frames as input. (c) we replace the hard sampling in SS with soft sampling, which means using the weighted sum of dense frames rather than top-k selection. (d) we remove vision refinement and apply a 3-layer transformer" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 305, + 545, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 305, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 304, + 305, + 545, + 329 + ], + "type": "text", + "content": "to fuse the vision and language. (e) we replace LA-Gates in both modules with cross-attention layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "text", + "content": "More Ablations We validate the influence of pretraining tasks, the best number of densely sampled frames, and the reasonable number of stacked layers " + }, + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 330, + 545, + 365 + ], + "type": "text", + "content": " in Appendix E." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 373, + 382, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 382, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 382, + 384 + ], + "type": "text", + "content": "4.7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 391, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 617 + ], + "type": "text", + "content": "In this paper, we study language-aware VideoQA to overcome the difficulties of long-term dependencies, multi-events, multi-scale semantics, and shortcut solutions in video understanding. In particular, we introduce a differentiable sparse sampling module that naturally avoids complicated long-term dependencies and multi-event reasoning since it only retains question-related frames and a vision refinement module that eliminates unrelated visual semantics in different perception fields and enhances multi-scale related semantics to support a multi-level refinement process. Both modules are based on our LA-Gate, which realizes cross-modal interaction meanwhile avoiding language's direct fusion into visual representations. LA-Gate is a powerful competitor of predominant Cross-Attention in VideoQA and fits our idea of semantic distillation. Therefore, we can realize high-performance VideoQA by just relying on question-related visual representations and reducing language shortcuts. Experiments on 8 VideoQA and text-to-video retrieval tasks demonstrate the effectiveness of our model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 629, + 378, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 378, + 643 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 378, + 643 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 649, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 649, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 649, + 547, + 709 + ], + "type": "text", + "content": "This work is supported in part by the Beijing Natural Science Foundation (No. L222024), and the National Natural Science Foundation of China (No. 62394322). One of the authors, Chao Yang, is supported by the Shanghai Post-doctoral Excellent Program (Grant No. 2022234)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27120" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 124 + ], + "type": "text", + "content": "[1] Aishwarya Agrawal, Dhruv Batra, and Devi Parikh. Analyzing the behavior of visual question answering models. arXiv preprint arXiv:1606.07356, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "text", + "content": "[2] Aishwarya Agrawal, Dhruv Batra, Devi Parikh, and Aniruddha Kembhavi. Don't just assume; look and answer: Overcoming priors for visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4971-4980, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Anonymous. LLaMA-excitor: General instruction tuning via indirect feature interaction. In Conference on Computer Vision and Pattern Recognition 2024, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "text", + "content": "[4] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 304 + ], + "type": "text", + "content": "[5] Dare A Baldwin, Jodie A Baird, Megan M Saylor, and M Angela Clark. Infants parse dynamic action. *Child development*, 72(3):708-717, 2001. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 307, + 288, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 339 + ], + "type": "text", + "content": "[6] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 288, + 396 + ], + "type": "text", + "content": "[7] Shyamal Buch, Cristóbal Eyzaguirre, Adrien Gaidon, Jiajun Wu, Li Fei-Fei, and Juan Carlos Niebles. Revisiting the\" video\" in video-language understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2917-2927, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 441 + ], + "type": "text", + "content": "[8] Wei-Lun Chao, Hexiang Hu, and Fei Sha. Cross-dataset adaptation for visual question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5716-5725, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 443, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 288, + 487 + ], + "type": "text", + "content": "[9] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 531 + ], + "type": "text", + "content": "[10] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "type": "text", + "content": "[11] Feng Cheng and Gedas Bertasius. Tallformer: Temporal action localization with a long-memory transformer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXIV, pages 503-521. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[12] Corentin Dancette, Rémi Cadène, Damien Teney, and Matthieu Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1574–1583, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[14] Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gated-attention readers for" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 94 + ], + "type": "text", + "content": "text comprehension. arXiv preprint arXiv:1606.01549, 2016.2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 161 + ], + "type": "text", + "content": "[15] Chenyou Fan, Xiaofan Zhang, Shu Zhang, Wensheng Wang, Chi Zhang, and Heng Huang. Heterogeneous memory enhanced multimodal attention model for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1999–2007, 2019. 1, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 547, + 206 + ], + "type": "text", + "content": "[16] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. arXiv preprint arXiv:2212.09522, 2022. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 240 + ], + "type": "text", + "content": "[17] György Gergely, Harold Bekkering, and Ildiko Király. Rational imitation in preverbal infants. Nature, 415(6873):755-755, 2002. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 243, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 547, + 307 + ], + "type": "text", + "content": "[18] Satya Krishna Gorti, Noel Vouitsis, Junwei Ma, Keyvan Golestan, Maksims Volkovs, Animesh Garg, and Guang-wei Yu. X-pool: Cross-modal language-video attention for text-video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5006–5015, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 309, + 547, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 547, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 547, + 364 + ], + "type": "text", + "content": "[19] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 547, + 430 + ], + "type": "text", + "content": "[20] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 432, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 547, + 487 + ], + "type": "text", + "content": "[21] Madeleine Grunde-McLaughlin, Ranjay Krishna, and Maneesh Agrawala. Agqa: A benchmark for compositional spatio-temporal reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11287-11297, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 488, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 547, + 520 + ], + "type": "text", + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 522, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 522, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 308, + 522, + 547, + 577 + ], + "type": "text", + "content": "[23] Ronghang Hu, Amanpreet Singh, Trevor Darrell, and Marcus Rohrbach. Iterative answer prediction with pointeraugmented multimodal transformers for textvqa. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9992-10002, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 578, + 547, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 547, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 547, + 610 + ], + "type": "text", + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 613, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 547, + 644 + ], + "type": "text", + "content": "[25] Baoxiong Jia, Ting Lei, Song-Chun Zhu, and Siyuan Huang. Egotaskqa: Understanding human tasks in egocentric videos. arXiv preprint arXiv:2210.03929, 2022. 2, 6, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "type": "text", + "content": "[26] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 1, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 691, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 712 + ], + "type": "text", + "content": "[27] Hyounghun Kim, Zineng Tang, and Mohit Bansal. Dense-ecaption matching and frame-selection gating for temporal" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27121" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 94 + ], + "type": "text", + "content": "localization in videoqa. arXiv preprint arXiv:2005.06409, 2020.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[28] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International Conference on Machine Learning, pages 5583-5594. PMLR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 288, + 207 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[30] Thao Minh Le, Vuong Le, Svetha Venkatesh, and Truyen Tran. Hierarchical conditional relation networks for video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9972–9981, 2020. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 1, 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 288, + 376 + ], + "type": "text", + "content": "[32] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 420 + ], + "type": "text", + "content": "[33] Kun Li, Dan Guo, and Meng Wang. Proposal-free video grounding with contextual pyramid network. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1902-1910, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[34] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. arXiv preprint arXiv:2005.00200, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "text", + "content": "[35] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 1, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 567 + ], + "type": "text", + "content": "[36] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8658–8665, 2019. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[37] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 510-519, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[38] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 714 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 160 + ], + "type": "text", + "content": "[40] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 216 + ], + "type": "text", + "content": "[41] Yulei Niu, Kaihua Tang, Hanwang Zhang, Zhiwu Lu, Xian-Sheng Hua, and Ji-Rong Wen. Counterfactual vqa: A cause-effect look at language bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12700-12710, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 259 + ], + "type": "text", + "content": "[42] Min Peng, Chongyang Wang, Yu Shi, and Xiang-Dong Zhou. Efficient end-to-end video question answering with pyramidal multimodal transformer. arXiv preprint arXiv:2302.02136, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 261, + 547, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 326 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 327, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 547, + 380 + ], + "type": "text", + "content": "[44] Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision transformers see like convolutional neural networks? Advances in Neural Information Processing Systems, 34:12116-12128, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 382, + 547, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 547, + 426 + ], + "type": "text", + "content": "[45] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 426, + 547, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 491 + ], + "type": "text", + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 493, + 547, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 547, + 546 + ], + "type": "text", + "content": "[47] Mykhailo Shvets, Wei Liu, and Alexander C Berg. Leveraging long-range temporal relationships between proposals for video object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9756-9764, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 592 + ], + "type": "text", + "content": "[48] Yuchong Sun, Hongwei Xue, Ruihua Song, Bei Liu, Huan Yang, and Jianlong Fu. Long-form video-language pretraining with multimodal temporal contrastive learning. arXiv preprint arXiv:2210.06031, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 547, + 635 + ], + "type": "text", + "content": "[49] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 690 + ], + "type": "text", + "content": "[50] Alex Jinpeng Wang, Yixiao Ge, Rui Yan, Yuying Ge, Xudong Lin, Guanyu Cai, Jianping Wu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. All in one: Exploring unified videolanguage pre-training. arXiv preprint arXiv:2203.07303, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 714 + ], + "type": "text", + "content": "[51] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wang-meng Zuo, and Qinghua Hu. Eca-net: Efficient channel" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27122" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 288, + 106 + ], + "type": "text", + "content": "attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 149 + ], + "type": "text", + "content": "[52] Su-Jing Wang, Ying He, Jingting Li, and Xiaolan Fu. Mesnet: A convolutional neural network for spotting multi-scale microexpression intervals in long videos. IEEE Transactions on Image Processing, 30:3956-3969, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 288, + 172 + ], + "type": "text", + "content": "[53] Amanda L Woodward. Infants selectively encode the goal object of an actor's reach. Cognition, 69(1):1-34, 1998. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "text", + "content": "[54] Chao-Yuan Wu and Philipp Krahenbuhl. Towards long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1884-1894, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 271 + ], + "type": "text", + "content": "[55] Chao-Yuan Wu, Christoph Feichtenhofer, Haoqi Fan, Kaiming He, Philipp Krahenbuhl, and Ross Girshick. Long-term feature banks for detailed video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 284-293, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 338 + ], + "type": "text", + "content": "[56] Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13587-13597, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 288, + 381 + ], + "type": "text", + "content": "[57] Xing Wu, Chaochen Gao, Zijia Lin, Zhongyuan Wang, Jizhong Han, and Songlin Hu. Rap: Redundancy-aware video-language pre-training for text-video retrieval. arXiv preprint arXiv:2210.06881, 2022.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 437 + ], + "type": "text", + "content": "[58] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645-1653, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 502 + ], + "type": "text", + "content": "[59] Hongwei Xue, Tiankai Hang, Yanhong Zeng, Yuchong Sun, Bei Liu, Huan Yang, Jianlong Fu, and Baining Guo. Advancing high-resolution video-language representation with large-scale video transcriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5036-5045, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 557 + ], + "type": "text", + "content": "[60] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Just ask: Learning to answer questions from millions of narrated videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1686-1697, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 288, + 602 + ], + "type": "text", + "content": "[61] Ceyuan Yang, Yinghao Xu, Jianping Shi, Bo Dai, and Bolei Zhou. Temporal pyramid network for action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 591-600, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 647 + ], + "type": "text", + "content": "[62] Youngjae Yu, Jongseok Kim, and Gunhee Kim. A joint sequence fusion model for video question answering and retrieval. In Proceedings of the European Conference on Computer Vision (ECCV), pages 471-487, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "text", + "content": "[63] Zhou Yu, Jun Yu, Yuhao Cui, Dacheng Tao, and Qi Tian. Deep modular co-attention networks for visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[64] Rowan Zellers, Ximing Lu, Jack Hessel, Youngjae Yu, Jae Sung Park, Jize Cao, Ali Farhadi, and Yejin Choi. Merlot:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 262 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "text", + "content": "Multimodal neural script knowledge models. Advances in Neural Information Processing Systems, 34:23634-23651, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "text", + "content": "[65] Peng Zhang, Yash Goyal, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Yin and yang: Balancing and answering binary visual questions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5014-5022, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[66] Xinsong Zhang, Yan Zeng, Jipeng Zhang, and Hang Li. Toward building general foundation models for language, vision, and vision-language understanding tasks. arXiv preprint arXiv:2301.05065, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 547, + 262 + ], + "type": "text", + "content": "[67] Bo Zou, Chao Yang, Chengbin Quan, and Youjian Zhao. Spaceclip: A vision-language pretraining framework with spatial reconstruction on text. In Proceedings of the 31st ACM International Conference on Multimedia, pages 519-528, 2023. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27123" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_content_list.json b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac2c73bcbb5d68aa75d2409d8b84704fe33fc040 --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_content_list.json @@ -0,0 +1,1360 @@ +[ + { + "type": "text", + "text": "Language-conditioned Detection Transformer", + "text_level": 1, + "bbox": [ + 253, + 130, + 718, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jang Hyun Cho UT Austin", + "bbox": [ + 297, + 181, + 423, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "janghyuncho7@utexas.edu", + "bbox": [ + 258, + 219, + 464, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Philipp Krahenbuhl UT Austin", + "bbox": [ + 540, + 181, + 697, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "philkr@cs.utexas.edu", + "bbox": [ + 527, + 219, + 707, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a new open-vocabulary detection framework. Our framework uses both image-level labels and detailed detection annotations when available. Our framework proceeds in three steps. We first train a language-conditioned object detector on fully-supervised detection data. This detector gets to see the presence or absence of ground truth classes during training, and conditions prediction on the set of present classes. We use this detector to pseudolabel images with image-level labels. Our detector provides much more accurate pseudo-labels than prior approaches with its conditioning mechanism. Finally, we train an unconditioned open-vocabulary detector on the pseudo-annotated images. The resulting detector, named DECOLA, shows strong zero-shot performance in open-vocabulary LVIS benchmark as well as direct zero-shot transfer benchmarks on LVIS, COCO, Object365, and OpenImages. DECOLA outperforms the prior arts by 17.1 AP_rare and 9.4 mAP on zero-shot LVIS benchmark. DECOLA achieves state-of-the-art results in various model sizes, architectures, and datasets by only training on open-sourced data and academic-scale computing. Code is available at https://github.com/janghyuncho/DECOLA.", + "bbox": [ + 75, + 300, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object detection has seen immense progress over the past decade. Classical object detectors reason over datasets of fixed predefined classes. This simplifies the design, training, and evaluation of new methods, and allows for rapid prototyping [2-5, 17, 18, 25, 37, 42, 60, 70, 72, 75]. However, it complicates deployment to downstream applications too. A classical detector requires a new dataset to further finetune for every new concept it encounters. Collecting sufficient data for every new concept is not scalable [20]. Open-vocabulary detection offers an alternative [19, 43, 61, 65, 66, 73]. Open-vocabulary detectors reason about any arbitrary concept with free-form text, using the generalization ability of vision-language models. Yet, common open-vocabulary detectors reuse classical detectors and either replace the last classification layer", + "bbox": [ + 75, + 674, + 472, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eab37f5762d533fbfcf4ee6003dca2f7f131377e0e31f25de59d1d33b5eec667.jpg", + "image_caption": [ + "Figure 1. An illustration of how standard open-vocabulary detectors and DECOLA generate pseudo-labels using image-level data. Standard detectors use image-level information later in the pipeline after initial box proposals, which may result in low coverage of unseen classes (e.g., \"mentos\" and \"cola\"). DECOLA adjusts the prediction to the information and ensures sufficient coverage." + ], + "image_footnote": [], + "bbox": [ + 511, + 268, + 890, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "with [19, 43, 66, 73], or fuse box feature with [31, 65] text representation from pretrained vision-language model. The inner workings of the detector remain unchanged.", + "bbox": [ + 496, + 669, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a transformer-based object detector that adjusts its inner workings to any arbitrary set of concepts represented in language. The detector considers only the queried set of concepts as foreground and disregards any other objects as background. It learns to adapt detection to the language embedding of queried concepts at run-time. Specifically, the detector conditions proposal generation with respect to the text embedding of each queried concept and refines the conditioned proposals into predictions. Our detection transformer conditioned on language (DECOLA) offers a powerful alternative to classical architectures in open-vocabulary detection. Adapting the detector to", + "bbox": [ + 496, + 719, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16593", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "language leads to stronger generalization to unseen concepts, and largely enhances self-training on weakly-labeled data.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DECOLA's ability to readily adapt to any queried concepts makes it particularly suitable for pseudo-labeling weakly-labeled data. Internet data, specifically images with paired text, is highly abundant and semantically rich [50, 51, 53]. The best vision models today build on this massive amount of weakly labeled data [10, 13, 24, 33, 34, 46, 67]. DECOLA leverages the same data to produce high-quality object detection labels from image-level annotations alone. DECOLA takes the image-level tags or text descriptions from the weakly labeled data and generates conditioned predictions as pseudo-labels. It efficiently processes multiple texts in parallel and only adds minimal computational overhead compared to the standard pseudo-labeling process. We finetune DECOLA on this rich detection dataset of pseudo-annotations and achieve the state-of-the-art open-vocabulary detector.", + "bbox": [ + 75, + 121, + 470, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our detector on popular open-vocabulary detection benchmarks on the LVIS dataset [19, 20, 66]. The final model improves the state-of-the-art methods by 4.4 and $4.9\\mathrm{AP}_{\\mathrm{novel}}$ on open-vocabulary LVIS [19] benchmark, and 5.9 and $5.4\\mathrm{AP}_{\\mathrm{rare}}$ on standard LVIS [20] benchmark, with ResNet-50 [21] and Swin-B [39] backbones, respectively. Our largest model with Swin-L achieves $10.4\\mathrm{AP}_{\\mathrm{rare}}$ and $3.6\\mathrm{mAP}$ improvement. Furthermore, DECOLA largely outperforms the state-of-the-art for direct zero-shot transfer benchmark on LVIS, by 12.0 and $17.1\\mathrm{AP}_{\\mathrm{rare}}$ on LVIS minimal and LVIS v1.0, respectively. DECOLA consistently improves frequent, common, and rare classes altogether for different backbones and detection frameworks. Much of this improvement is driven by stronger pseudo-labeling capabilities. All our models are trained using open-sourced datasets with academic-scale computing. We open-source our code, pseudo-annotations, and checkpoints of all the model scales.", + "bbox": [ + 75, + 348, + 470, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 618, + 218, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary detection aims to detect objects of categories beyond the vocabulary of the training classes. A common solution is to inject language embeddings of class names in the last classification layer. OVR-CNN [66] pretrains a detector on image-caption data using BERT model [11] as language embedding. ViLD [19] trains a detector with CLIP text encoder [46] as language embedding with additional knowledge distillation [23] between predicted box features and the image encoder of CLIP. Detic [73] improves the above approaches through weakly-supervised learning on image-level annotations. RegionCLIP [71] introduces an intermediate pretraining step to better align CLIP to box features. BARON [61] improves the alignment between text and image encoders by extracting bag of regions as additional supervision. F-VLM [31] simplifies the training pipeline of open-vocabulary detection and explores the limit of the frozen vision-language model. All of the models above take", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the design of the object detector as granted, and inject language in the last classification layer of the network. We take a different approach and design a detector that adapts predictions to particular categories of interest.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-vocabulary DETR integrates DETR architecture into open-vocabulary detection. OWL [43] introduces a simple ViT architecture using pretrained CLIP and finetune with the DETR objective. OWLv2 uses self-training to further improve the performance [41]. OV-DETR [65] fuses features of a pretrained CLIP model with DETR object queries. Archi-tecturally, OV-DETR is closest to DECOLA. Both OV-DETR and DECOLA condition predictions on the text representation of classes. OV-DETR uses the original DETR queries and expands each with a CLIP feature per class for all classes. This leads to a quadratic number of queries, growing with the original DETR queries and with the classes considered. On the other hand, DECOLA explicitly controls the first-stage predictions (proposals) by formulating the scoring function to respect to the text embedding of each queried class at run-time. We visualize this difference in Figure ?? in the supplementary. The advantage is that we entirely remove inter-class competition and process a manageable amount of queries each focusing on a specific class, and running as fast as the vanilla Deformable DETR. This ability to freely adjust inner workings deviates DECOLA from prior works; it expands detection data through high-quality pseudo-labeling and achieves state-of-the-art results.", + "bbox": [ + 496, + 152, + 893, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-vocabulary object detection shares similar goals with open-vocabulary detection. Both learn from naturally long-tail data over large vocabularies. Vanilla large-vocabulary detectors are often ill-calibrated: The detector's final classification layer favors frequently seen objects over rare ones. This imbalance is usually addressed through a change in loss [7, 55-57, 72], or leveraging additional weakly labeled data for self-training [8, 16, 68, 76]. In large-vocabulary detection, R-CNN-based frameworks [2, 4, 17, 18, 72] dominate despite DETR-style architectures [5, 25, 44, 70] having long surpassed them on standard benchmarks [36]. DETR automatically assigns object queries to output classes, and thus it learns to more heavily focus queries on common classes. We show that language-conditioning helps address this calibration issue. Specifically, it removes inter-class competition in the training objective as queries are no longer shared across categories. As a result, DECOLA equally focuses on as many rare classes as frequent ones whenever they are present in an image. This yields a DETR-style detector that is competitive with the best R-CNN-based large-vocabulary detectors.", + "bbox": [ + 496, + 500, + 895, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 830, + 638, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection transformers (DETR) [3] build an object detection pipeline as a single feed-forward network. The network transforms object queries, arbitrary feature vectors, into la", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16594", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ffeaad042b3f61a7d65e0720779527f587aaa5d31460ffff12f09d84dc4235c3.jpg", + "image_caption": [ + "Figure 2. Overview of DECOLA Phase 1 for conditioned prediction (top) and Phase 2 for open-vocabulary detection (bottom). For language-conditioned detection, each text embedding directly parameterizes the objectness function for each class. In Phase 2, the language-condition reads \"an object\" instead of particular class names, and predicts multi-class scores over all classes after decoding layers." + ], + "image_footnote": [], + "bbox": [ + 148, + 90, + 807, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "beled bounding boxes through a series of cross-attention layers in a decoder architecture. Vanilla DETR [3] learns object queries as free-form parameters, while modern DETRs architectures [25, 37, 60, 70, 75] adopt a two-stage paradigm similar to RCNNs [48]. This query mechanism controls much of the inner workings of the detector. Queries determine what image regions the detector focuses on, and what object classes are prioritized.", + "bbox": [ + 75, + 411, + 470, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Query selection. Modern DETR architectures use image-dependent query selection, analogous to R-CNN's proposal generation [48]. An objectness function $s$ scores each grid location $(i,j)$ in the image using a feature $x_{i,j}$ extracted from the transformer encoder. The top- $k$ scored regions proceed to the second stage as object queries $Q$ :", + "bbox": [ + 75, + 532, + 470, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(x _ {i, j}\\right) = \\left\\langle x _ {i, j}, w \\right\\rangle , \\quad Q = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s \\left(x _ {i, j}\\right)\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 633, + 468, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $w$ are the parameters of the objectness predictor. Each selected query produces a series of predictions that are refined over multiple iterations, similar to Cascade R-CNN [2]. The final prediction $\\vec{p}$ contains scores over all classes $C$ and an associated box. At a high level, DETR and R-CNNs share the same motivation: first, localize all objects in a scene, then refine their predictions.", + "bbox": [ + 75, + 659, + 470, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training objective. During training, DETR assigns each object query to an object or marks it as background. This allows DETR to learn non-overlapping object queries without post-processing such as non-maximum-suppression. The Hungarian matching algorithm finds the optimal assignment between all predictions $P$ and all ground truth $G$ , minimizing the loss function as matching cost:", + "bbox": [ + 75, + 765, + 470, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S}} \\ell (P, G | \\sigma) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 881, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathfrak{S}$ captures all possible assignments from $P$ to $G$ . For each assigned prediction, the loss $\\ell$ maximizes its class log-likelihood and fits its bounding box. For unassigned predictions, the loss $\\ell$ reduces both the objectness score $s$ and class log-likelihood for all classes.", + "bbox": [ + 496, + 411, + 893, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. DECOLA", + "text_level": 1, + "bbox": [ + 498, + 498, + 612, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our detection transformer conditioned on language, DECOLA, changes the DETR architecture in one remarkable way: Object queries are conditioned on a language embedding. Figure 2 illustrates this change. This simple change has a few important implications: First, it allows the language embedding to control and focus queries to localize on the concepts at hand. Second, it removes any contention between different object classes. Each class present in the image uses the same amount of queries. Third, it generalizes to unseen classes by leveraging semantic knowledge encoded in language embedding throughout the detection pipeline. In the remainder of this section, we highlight the changes in the architecture and training objective for conditioning (DECOLA Phase 1), and self-training on image-level data for open-vocabulary detection (DECOLA Phase 2).", + "bbox": [ + 496, + 523, + 893, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-conditioned query selection. DECOLA conditions queries to a specific object category by modeling the objectness function as a similarity score between a region feature $x_{i,j}$ and a text representation of a category name $t(y)$ using their cosine similarity:", + "bbox": [ + 496, + 751, + 893, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ns _ {y} \\left(x _ {i, j}\\right) = \\frac {\\left\\langle x _ {i , j} , t (y) \\right\\rangle}{\\left\\| x _ {i , j} \\right\\| \\left\\| t (y) \\right\\|}, Q _ {y} = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s _ {y} \\left(x _ {i, j}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 832, + 892, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above objective avoids any inter-class calibration issue common in imbalanced data [7, 55, 57]. Queries do not", + "bbox": [ + 498, + 869, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "compete, as DECOLA independently selects top- $k$ scoring regions $Q_{y}$ for each class $y$ . All queries proceed to the second stage in parallel. A memory-efficient attention mechanism isolates interaction within each class. After a series of decoding layers, each language-conditioned query predicts a single scalar score, corresponding to the likelihood of the class $y$ , and the associated box. The overall architecture mirrors the two-stage deformable DETR [75] with two modifications: a language-conditioned query, and a binary output classifier.", + "bbox": [ + 75, + 90, + 472, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory-efficient modeling. DECOLA uses $n = |Q_y|$ queries per class for $K$ classes. Generally, $n$ is smaller than the total number of queries $|Q|$ of a standard DETR model. However, since we produce $n$ queries per class, the total number of queries in DECOLA is much larger $|Q| \\ll nK$ . A naive implementation of the DETR decoder is unable to cope with the $O(n^2 K^2)$ memory requirements of the self-attention layers in the transformer decoder. We thus modify the self-attention formulation to isolate it within each class, reducing the memory cost to $O(n^2 K)$ . The actual implementation uses standard self-attention with a reshaping operation. See Figure 2 (right) for the illustration. DECOLA Phase 1: Train to condition on given concepts. Our goal is to design DECOLA to take a set of class names in an image (or a batch of images) and predict objects of the corresponding classes or backgrounds. For each class $y$ , each conditioned query $q_y \\in Q_y$ therefore only predicts a single presence score for class $y$ and the box location. All predictions from the conditioned queries $P_y$ are matched with $G_y$ , the subset of ground truth with class $y$ :", + "bbox": [ + 75, + 226, + 472, + 530 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {y} ^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S} _ {y}} \\ell \\left(P _ {y}, G _ {y} \\mid \\sigma\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 537, + 468, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathfrak{S}_y$ is the set of possible matches between $P_y$ and $G_y$ , and $\\ell$ is the binary cross-entropy loss. Unlike the original DETR objective in Eqn. 2, Eqn. 4 matches within the conditioned class. It avoids inter-class competition during training and simplifies the training objective. Instead, it learns to adapt its predictions to $y$ ; the set of conditioned query $q_y$ considers any objects other than $y$ as background.", + "bbox": [ + 75, + 568, + 468, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pseudo-labeling weakly-labeled data. DECOLA produces highly accurate predictions when conditioned on the exact categories of a scene, as shown in Section 5.4. This makes DECOLA a strong pseudo-labeler for weakly-labeled data with either image tags or captions. We expand a large amount of such data with pseudo-bounding boxes of DECOLA Phase 1 and self-train altogether to scale up open-vocabulary object detection. Unlike other forms of weakly supervised learning such as knowledge distillation [19] and online pseudolabeling [61, 73], we simply generate labels for all images offline and jointly train over all pseudo-labeled data using the regular detection losses without any additional complication or slowdown. For each image and class $y$ , DECOLA encodes the class' language feature and predicts a set of detections $P_{y}$ . We simply choose the most confident prediction.", + "bbox": [ + 75, + 674, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4d shows our simple offline pseudo-labeling works better than online pseudo-labeling.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The advantage of DECOLA comes from adaptability to specified class names on a per-image basis. However, in open-vocabulary detection, the set of test classes is neither known a priori nor available per image. Hence, we convert DECOLA into a general-purpose detector to detect all objects. We condition DECOLA with \"an object\" as the text input, and inject the class information in the second-stage classifier. Figure 2 highlights this conversion. Since DECOLA is trained to align image features to text embedding in both the first and second stages, this change only introduces inter-class calibration for multi-class object detection. We train DECOLA with pseudo-labeled and human-labeled data as a standard supervised detection training, using the standard matching algorithm of DETR (Section 3). We do not introduce any additional hyper-parameter specifically for the weakly supervised learning [73] or design choices [31, 71], extra loss functions such as alignment loss [61], nor a large teacher model for knowledge distillation [19, 65]. Generating pseudo-labels with DECOLA Phase 1 runs as fast as a regular detector and training Phase 2 is as easy as standard detection training on a supervised dataset. At a high level, DECOLA Phase 1 training objective optimizes for a strong pseudo-labeler instead of a multi-class detector, which differentiates DECOLA from prior works. Leveraging DECOLA Phase 1 to expand weakly-labeled data is the key contribution to scaling up the final open-vocabulary object detection.", + "bbox": [ + 496, + 121, + 893, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 541, + 633, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of DECOLA in two aspects: (1) pseudo-labeling quality of DECOLA Phase 1 (Section 5.4) and (2) benchmark evaluation (Section 5.3). We consider three primary benchmarks to evaluate our final model (DECOLA Phase 2): open-vocabulary LVIS [19], standard LVIS [20], and direct zero-shot evaluation to LVIS, COCO [36], Object365 [52], and OpenImages [32].", + "bbox": [ + 496, + 566, + 893, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 681, + 691, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets and benchmarks. We mainly evaluate our method on the LVIS dataset [20], a large-vocabulary instance segmentation and object detection dataset with 1203 naturally distributed object categories. LVIS splits categories into frequent, common, and rare. For open-vocabulary LVIS, we combine frequent and common categories into LVIS-base and consider the rare categories as novel concepts used for testing only [19]. For standard LVIS, we train and evaluate all classes. Direct zero-shot transfer evaluates models trained on different detection data (e.g., Object365) and other weakly-labeled data without any prior knowledge about the target dataset such as the set of classes or object frequency. In this benchmark, we test DECOLA's generalization to dif", + "bbox": [ + 496, + 704, + 895, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4b05f749b7793e547ba3fd1be801166453298dd417c8c783d360752fe83aee9a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methoddataAPbox novelAPcAPfmAPbox
ResNet-50 (1K)
OV-DETR† [65]LVIS-base, IN-21K18.024.831.826.4
baselineLVIS-base10.230.938.030.1
baseline + self-trainLVIS-base, IN-21K19.231.737.131.7
DECOLA Phase 2LVIS-base, IN-21K23.8 (+4.6)34.4 (+2.7)38.3 (+1.2)34.1 (+2.4)
ResNet-50
baselineLVIS-base9.433.840.432.2
baseline + self-trainLVIS-base, IN-21K23.236.541.636.2
DECOLA Phase 2LVIS-base, IN-21K27.6 (+4.4)38.3 (+1.8)42.9 (+1.3)38.3 (+2.1)
Swin-B
baselineLVIS-base16.243.849.141.1
baseline + self-trainLVIS-base, IN-21K30.843.645.942.3
DECOLA Phase 2LVIS-base, IN-21K35.7 (+4.9)47.5 (+3.9)49.7 (+3.8)46.3 (+4.0)
Swin-L
DITO* [28]O365, LVIS-base, DataComp-1B45.8--44.2
OWLv2‡ [41]O365, LVIS-base, VG, WebLI45.9--50.4
baselineO365, LVIS-base21.953.357.749.6
baseline + self-trainO365, LVIS-base, IN-21K36.553.556.551.8
DECOLA Phase 2O365, LVIS-base, IN-21K46.9 (+10.4)56.0 (+2.5)58.0 (+1.5)55.2 (+3.6)
", + "bbox": [ + 191, + 87, + 777, + 344 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/35cf24c708fd444d521de86984b107f692a062a9ed5f3f14db0de67abc6f408b.jpg", + "table_caption": [ + "Table 1. Open-vocabulary LVIS using DETR architectures. $\\dagger$ : we further finetune the official OV-DETR model trained on LVIS-base by self-training on ImageNet-21K data similar to \"baseline + self-train\" and DECOLA. $\\ddagger$ : uses CLIP L/14, which is comparable to Swin-L backbone. $\\star$ : is based on Mask R-CNN with ViT L/16. We report the improvement between \"baseline + self-train\" to DECOLA in green. Last 4 rows compare DECOLA to Swin-L or equivalent scale of models that use additional detection data (e.g., Objects365, VG [30]) and billion-scale weakly-labeled data (DataComp-1B [15], WebLI [6]) for training." + ], + "table_footnote": [ + "(a) Comparison with ResNet-50 backbone." + ], + "table_body": "
methodframework\\(AP^{box}\\)novel\\(mAP^{box}\\)novel\\(AP^{mask}\\)novel\\(mAP^{mask}\\)
ViLD [19]Mask R-CNN16.727.816.625.5
RegionCLIP [71]Mask R-CNN--17.122.5
DetPro [12]Mask R-CNN20.828.419.825.9
PromptDet [14]Mask R-CNN21.425.3--
F-VLM [31]Mask R-CNN--18.624.2
BARON [61]Mask R-CNN23.229.522.627.6
OADP [58]Mask R-CNN21.928.721.726.6
EdaDet [54]Mask R-CNN--23.727.5
VLDet [35]CenterNet2--21.730.1
\\(\\mathsf{CORA}^{+}\\) [62]CenterNet228.1---
Rasheed et al. [47]CenterNet2--25.232.9
Detic-base [73]CenterNet217.633.816.430.2
Detic [73]CenterNet226.736.324.632.4
DECOLA labelsCenterNet229.537.727.033.7
", + "bbox": [ + 96, + 430, + 480, + 623 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4050687fb3698defcf31652fce07636378dd5f5a564794f0027e9be6862e1641.jpg", + "table_caption": [], + "table_footnote": [ + "(b) System-level comparison." + ], + "table_body": "
methodbackbone\\(AP_{\\text{novel}}^{box}\\)\\(mAP_{\\text{novel}}^{box}\\)\\(AP_{\\text{novel}}^{mask}\\)\\(mAP_{\\text{novel}}^{mask}\\)
RegionCLIP [71]R50×4--22.032.3
CondHead [59]R50×424.133.724.432.0
ViLD [19]EN-B7--26.329.3
OWL-ViT [43]ViT-L/1425.634.7--
F-VLM [31]R50×64--32.834.9
VLDet [35]Swin-B--26.338.1
3Ways [1]NFNet-F630.144.6--
RO-VIT [29]ViT-L/1632.134.0--
CFM-ViT [27]ViT-L/1635.638.533.936.6
DITO [28]ViT-B/1634.936.932.534.0
CoDet [40]Swin-B--29.439.2
Detic-base [73]Swin-B24.643.021.938.4
Detic [73]Swin-B36.645.733.840.7
DECOLA labelsSwin-B38.446.735.342.0
", + "bbox": [ + 504, + 430, + 898, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Open-vocabulary LVIS using Mask R-CNN and CenterNet2 detectors. Methods in both tables use LVIS-base as the only human-labeled data for fair comparison. For system-level comparison (right), we include methods with non R-CNN architectures such as OWL-ViT [43]. The results show the impact of high-quality pseudo-labels generated by DECOLA Phase 1.", + "bbox": [ + 75, + 642, + 892, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ferent domains. We evaluate DECOLA on LVIS, COCO [36], Object365 [52], and OpenImages [32] in a fully zero-shot manner. All our models use the ImageNet-21K [50] dataset as weakly labeled data, which contains 14M of object-centric images annotated with a single class.", + "bbox": [ + 75, + 696, + 470, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate DECOLA on $\\mathrm{AP}_{\\mathrm{novel/rare}}$ , $\\mathrm{AP_c}$ , $\\mathrm{AP_f}$ , and mAP following the LVIS evaluation metric [20]. We highlight the results in all three groups since we believe open-vocabulary detectors should not compensate for the performance of common/frequent classes for novel/rare classes. We evaluate both $\\mathrm{AP^{box}}$ and $\\mathrm{AP^{mask}}$ for object detection and instance segmentation. For zero-shot transfer benchmark with COCO and Object365, we", + "bbox": [ + 75, + 780, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "use AP, $\\mathrm{AP}_{50}$ , and $\\mathrm{AP}_{75}$ following prior work [19, 73]. For OpenImages, we report $\\mathrm{AP}_{50}^{\\mathrm{flat}}$ on the expanded label space [73, 74]. For zero-shot transfer to LVIS, we consider LVIS minival [26] and standard LVIS v1.0 validation set and report $\\mathrm{AP}^{\\mathrm{fixed}}$ [9] following the prior works [26, 34, 38]. In addition, we pursue a more direct measurement of the generated pseudo-labeling quality. Hence, we define conditioned mAP/AR (c-mAP/AR) and compare it to baseline open-vocabulary detectors. c-mAP measures the detection performance in mAP when the detector is provided the set of ground truth classes in each image. For example in Figure 1, both detectors use \"cat\", \"mentos\" and \"cola\" as given. This extra information is used to select scores to rank", + "bbox": [ + 496, + 696, + 893, + 893 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d6fb5e4224f302780ec060d726045c08107f82aa557bfb54e6140e19d18d6750.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methoddataAPboxraremAPbox
ResNet-50
baselineLVIS26.335.6
baseline + self-trainLVIS, IN-21K30.036.6
DECOLA Phase 2LVIS, IN-21K35.9 (+5.9)39.4 (+2.8)
Swin-B
baselineLVIS38.344.5
baseline + self-trainLVIS, IN-21K42.045.2
DECOLA Phase 2LVIS, IN-21K47.4 (+5.4)48.3 (+3.1)
Swin-L
baselineO365, LVIS49.354.4
baseline + self-trainO365, LVIS, IN-21K48.753.4
DECOLA Phase 2O365, LVIS, IN-21K54.9 (+6.2)56.4 (+3.0)
", + "bbox": [ + 89, + 88, + 472, + 256 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/68559292312b3194c1009dfab04fb8ea457c7ab17d317e642e3c4822e69cef2e.jpg", + "table_caption": [ + "(a) Standard LVIS with DETR." + ], + "table_footnote": [], + "table_body": "
methodAPboxraremAPboxAPmaskraremAPmask
ResNet-50
Detic-base [73]28.235.325.631.4
Detic [73]31.436.829.733.2
DECOLA labels35.6 (+4.2)38.6 (+1.8)32.1 (+2.4)34.4 (+1.2)
Swin-B
Detic-base [73]39.945.435.940.7
Detic [73]45.846.941.741.7
DECOLA labels47.6 (+1.8)48.5 (+1.6)43.7 (+2.0)43.6 (+1.9)
", + "bbox": [ + 86, + 270, + 472, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(b) Standard LVIS with CenterNet2.", + "bbox": [ + 179, + 388, + 370, + 401 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Standard LVIS benchmark. DECOLA shows consistent improvement over different model scales and architectures.", + "bbox": [ + 75, + 406, + 468, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the final predictions (belines), or directly condition the detector (DECOLA). We analyze the model's behavior and label quality in Section 5.4 and Section ?? in supplementary.", + "bbox": [ + 75, + 441, + 470, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Models", + "text_level": 1, + "bbox": [ + 76, + 497, + 169, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DECOLA is based on two-stage Deformable DETR [75]. As described in Section 4, the first-stage objectness function for query selection is replaced by a similarity score between the image feature and the CLIP text embedding of each class name. We train the detector with the improved DETR training recipe [44, 70]: look-forward-twice, larger MLP hidden dimension, no dropout, etc. We consider four backbones: a ResNet-50 [21], Swin-B and L for all LVIS benchmarks, and Swin-T and L for the direct zero-shot transfer. Unless otherwise mentioned, all backbones are pretrained on the ImageNet-21K dataset [49]. Next, we describe our key baseline models to directly compare to DECOLA.", + "bbox": [ + 75, + 521, + 468, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. We design a baseline open-vocabulary detector to closely compare to DECOLA. Inspired by Detic [73], baseline replaces classification layers with the class embedding of the pretrained CLIP text encoder and is trained using Federated Loss [72]. DECOLA Phase 1 and baseline are trained using human-labeled data (e.g., LVIS-base). All other settings (training dataset, number of training iterations, etc.) are kept the same between DECOLA and baseline.", + "bbox": [ + 75, + 703, + 468, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline + self-train. Similar to DECOLA Phase 2, we self-train baseline on weakly-labeled data. For the self-training algorithm, we use online self-training with max-size loss from Detic [73] as baseline comparison (baseline + self-train) to DECOLA Phase 2. We tested max-size and max", + "bbox": [ + 75, + 824, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "score losses from Detic [73] (online pseudo-labeling) as well as offline pseudo-labeling similar to DECOLA, and max-size loss consistently performed the best.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DECOLA labels. We train a two-stage detector for broader comparison: CenterNet2 [72]. Specifically, we use Detric's baseline model (\"Detric-base\"), a CenterNet2 trained on LVIS-base with CLIP embedding, and finetune on pseudolabeled ImageNet-21K data using DECOLA Phase 1 of the same backbone size. We denote this as \"DECOLA labels\".", + "bbox": [ + 496, + 137, + 893, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficient modeling. For DECOLA Phase 1, we use $n = 300$ queries per class. One memory and time bottleneck during DECOLA training is the first-stage loss computation. The original Deformable DETR computes Hungarian matching with all pixels to all objects in a class-agnostic manner, which is $\\sum_{l \\in L} H_l \\cdot W_l$ predictions. To reduce the memory and time cost, we only consider the top $K = 10,000$ confident pixels for each class $y$ during the first-stage matching and loss computation. Together with memory-efficient self-attention (Sec. 4), the training time and memory cost of DECOLA increases by less than $20\\%$ over the baselines (See Table 6).", + "bbox": [ + 496, + 229, + 893, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. Following Detic [73], we train DECOLA Phase 1 and baseline $4 \\times$ on LVIS-base, and further fine-tune for another $4 \\times$ on ImageNet-21K data with pseudolabeling. For training CenterNet2 with DECOLA labels, we combine pseudo-labels from different image resolutions for $h \\in \\{240, 280, 320, 360, 400\\}$ , where $h$ is the shorter side of the image. Note that this mimics the random image resizing data augmentation during standard detection training. We use Detector2 [63] based on PyTorch [45] in all of the experiments. More details are in Section ?? of supplementary.", + "bbox": [ + 496, + 396, + 893, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Results", + "text_level": 1, + "bbox": [ + 500, + 559, + 637, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open-vocabulary LVIS. Table 1 compares DECOLA to baseline as well as state-of-the-art DETR-based open-vocabulary detectors; OV-DETR [65], OWLv2 [41], and baseline + self-train. For a fair comparison, we further fine-tune the official OV-DETR model checkpoint on ImageNet21K for $4 \\times$ schedule same as DECOLA Phase 2 and baseline + self-train. For all backbone scales, we show consistent improvement over other methods. Notably, baseline + self-train exhibits degradation in frequent classes as a trade-off with improved novel classes, which is commonly observed behavior in other open-vocabulary detection methods, too. DECOLA improves all categories consistently, which highlights the quality of our pseudo-labels. In the last rows with the Swin-L backbone, we report the result of two concurrent works, DITO [28] (Mask R-CNN-based) and OWLv2, to compare to the method that uses additional detection data (Object365) and billion-scale web data (DataComp1B [15], WebLI [6]). DECOLA demonstrates large improvement over the state-of-the-arts despite using orders of magnitude smaller training data and compute resources. To further examine DECOLA's scalability, we test the pseudo", + "bbox": [ + 496, + 583, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/88facdde53a3e545c6b64ab5665a0214e7bf03c1e119b45c2f610a89994e0c0f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
methoddataLVIS minivalLVIS v1.0 val
APboxrareAPcAPfmAPboxAPboxrareAPcAPfmAPbox
Swin-T
MDETR* [26]LVIS, GoldG, RefC20.924.924.324.27.422.725.022.5
GLIP [34]O365, GoldG, Cap4M20.821.431.026.010.112.525.517.2
GroundingDINO [38]O365, GoldG, Cap4M18.123.332.727.4----
GLIPv2 [69]O365, GoldG, Cap4M---29.0----
MQ-GroundingDINO† [64]O365, GoldG, Cap4M, LVIS-5VQ21.726.235.230.212.917.431.422.1
MQ-GLIP† [64]O365, GoldG, Cap4M, LVIS-5VQ21.027.534.630.415.418.430.422.6
DECOLA Phase 2O365, IN-21K‡32.832.031.832.027.224.928.026.6
Δ(+12.0)(+8.7)-(+3.0)(+17.1)(+12.4)(+2.5)(+9.4)
Swin-L
GLIP [34]FourODs, GoldG, Cap24M28.234.341.537.317.123.335.426.9
GroundingDINO [38]O365, OI, GoldG, Cap4M, COCO, RefC22.230.738.833.9----
MQ-GLIP† [64]FourODs, GoldG, Cap24M, LVIS-5VQ34.541.246.943.426.932.041.334.7
OWLv2 [41]O365, VG, WebLI39.0--38.134.9--33.5
DECOLA Phase 2O365, OI, IN-21K‡41.538.034.936.832.929.130.330.2
Δ(+2.5)(+3.7)---(+5.8)--
", + "bbox": [ + 94, + 88, + 874, + 318 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5132d0af4e0378f011ac9b720469b4f6cc9fe273839ad595fcaed4249f232b9.jpg", + "table_caption": [ + "Table 4. Direct zero-shot transfer to LVIS. †: methods use 5 per-class vision queries of LVIS dataset (denoted as “LVIS-5VQ”), which use images and annotations to extract instance-level features. ☆: MDETR uses ResNet-101 backbone and trained fully-supervised on LVIS. Results that use LVIS data are in gray. ‡: Full ImageNet-21K. No LVIS information is used in DECOLA." + ], + "table_footnote": [], + "table_body": "
methodCOCOO365OI
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{50}^{flat} \\)
R-CNNs
ViLD [19]36.655.639.811.818.212.6-
F-VLM [31]32.553.134.611.919.212.6-
DetPro [12]34.953.837.412.118.812.9-
BARON [61]36.356.139.313.621.014.5-
Detic [73]39.156.342.214.220.715.242.9
DETRs
OV-DETR [65]38.158.441.1----
Detic [73]39.856.643.314.521.415.541.6
DECOLA Phase 240.357.043.715.022.016.043.3
", + "bbox": [ + 78, + 369, + 475, + 523 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a6e8ad1b08545e5985f6ba98570ee374b988bdc61475ecf225e0f578ed49ff68.jpg", + "table_caption": [ + "Table 5. Cross-dataset generalization benchmark on COCO, Object365, and OpenImages. All models use a ResNet-50 backbone and train on LVIS-base and weakly-labeled data." + ], + "table_footnote": [], + "table_body": "
methodtraintest
timemem.timemem.
baseline44 h8.9 G0.07 s/img2.5 G
+ self-train45 h10.2 G0.07 s/img2.5 G
OV-DETR73 h22.0 G6.4 s/img3.4 G
DECOLA Phase 149 h12.6 G--
DECOLA Phase 245 h8.9 G0.07 s/img2.8 G
", + "bbox": [ + 119, + 568, + 424, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Efficiency. Training time is measured with 8 DGX V100 on ResNet-50, 2 images per-GPU. float16 is used in both training and testing. OV-DETR uses the original optimized inference.", + "bbox": [ + 75, + 662, + 470, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labeling capability of DECOLA on R-CNN-based detectors with DECOLA labels (Detic-base finetuned with our pseudolabels). In Table 2, we compare DECOLA labels to a broad range of literature based on CenterNet2 [72] and Mask R-CNN [22]. Table 2a compares methods with ResNet-50 backbone and Table 2b compares larger scale backbones for system-level comparison. In both tables, DECOLA clearly improve upon the state-of-the-art by large margins without additional complication in training and bells-and-whistles.", + "bbox": [ + 75, + 715, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Standard LVIS. Tables 3 evaluate DECOLA and baseline on the standard LVIS benchmark, where all object categories are used to fully supervise the detectors. Similar to the", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "open-vocabulary LVIS, we compare DETR architectures in Table 3a and R-CNN architectures in Table 3b. Table 3a shows that DECOLA remarkably improves baseline by 9.5, 9.1, and 5.6 points on $\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}$ , outperforming baseline + self-train by 5.9, 5.4, and 6.2 $\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}$ for ResNet-50, Swin-B, and Swin-L backbones, respectively. Similarly in Table 3b, DECOLA labels further improves the baseline of Detic by 7.4 and 7.7 $\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}$ , and outperforms Detic [73] by 4.2 and 1.8 $\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}$ with ResNet-50 and Swin-B backbones, respectively. Direct zero-shot evaluation. For direct zero-shot evaluation, we train DECOLA with Swin-T [39] and use Object365 data for Phase 1, and ImageNet-21K for Phase 2 (full dataset and classes). We compare to MDETR [26], GLIP [34], GroundingDINO [38], and MQ-Det [64] finetuned from GLIP and GroundingDINO. Table 4 shows the results. DECOLA outperforms the previous state-of-the-arts, by $12.0 / 17.1\\mathrm{AP}_{\\mathrm{rare}}$ and $3.0 / 9.4\\mathrm{mAP}$ on LVIS minival and LVIS v1.0 val, respectively. Note that all other methods use much richer detection labels from GoldG data [26], a collection of grounding data (box and text expression pairs) curated by MDETR. Furthermore, other benchmark methods show highly imbalanced $\\mathrm{AP}_{\\mathrm{rare}}$ and $\\mathrm{AP_f}$ in both LVIS minival and LVIS v1.0 val (10-20 points gap). We hypothesize that the large collection of training data coincides with LVIS vocabulary, as all data follows a natural distribution of common objects. Also, DECOLA enjoys significantly faster run-time compared to all other models that undergo BERT encoding for grounding, which requires more than 50 forward passes per image in order to predict all LVIS categories. Similarly, our Swin-L model outperforms GroundingDINO and GLIP by 19.3 and $13.3\\mathrm{AP}_{\\mathrm{rare}}$ , respectively, despite much smaller training data compared to FourODs [34] and match to OWLv2 with 10-B private WebLI data [6]. Table 5 further examines the generality of DECOLA on different domains. DECOLA Phase 2 with ResNet-50 outperforms all other competitive baselines.", + "bbox": [ + 496, + 371, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16599", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14d1dd4ea0cc274c81c008bde984323767dcd651e32a292f669c9fc0c03b26f8.jpg", + "image_caption": [ + "Figure 3. Examples of prediction on unseen categories. Images from ImageNet-21K dataset. Boxes are the most confident prediction from DECOLA and baseline. DECOLA conditions on the ground truth label, and baseline selects the max-score box of the ground truth class. Images are all from unseen categories, which neither model was trained on. Green: DECOLA Phase 1 trained on LVIS-base. Red: baseline trained on LVIS-base. Both models use a Deformable DETR detector with a ResNet-50 backbone. More in Fig. ?? and ?? of supplementary." + ], + "image_footnote": [], + "bbox": [ + 165, + 90, + 807, + 356 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bd77d56d18c866bb0b2263450d1459c0dc6abbee2c76f4199e508cf6f9f7261.jpg", + "image_caption": [ + "(a) c-AP for ResNet-50 (IN-1K)" + ], + "image_footnote": [], + "bbox": [ + 81, + 426, + 277, + 545 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ed8e239b3cf9a89151578fbf69b0216a65323826667a60ced142f0d5c08d527.jpg", + "image_caption": [ + "(b) c-AP for Swin-B", + "Figure 4. Analyzing DECOLA. All plots show the conditioned AP/AR for unseen classes. We compare DECOLA Phase 1 and baselines. We highlight more detailed analyses about c-AP and c-AR in Section ??, Table ?? and ?? of the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 279, + 426, + 472, + 544 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b6e4c6ff3f4bbd78d357c45926e0fa2b19cc757fa2a03f4e19e31ca2b78f61f.jpg", + "image_caption": [ + "(c) c-AR for 1st and 2nd stage boxes" + ], + "image_footnote": [], + "bbox": [ + 473, + 426, + 686, + 536 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96c45373ba0c1f741ab0eb7382b39c0c6425e02cb0388b2f7e88658524e408c3.jpg", + "image_caption": [ + "(d) online vs offline: c-AP@20" + ], + "image_footnote": [], + "bbox": [ + 686, + 426, + 887, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Analyses", + "text_level": 1, + "bbox": [ + 76, + 603, + 181, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we analyze the model's behavior with conditioned mAP/AR (c-AP/AR) (defined in Sec. 5.1).", + "bbox": [ + 75, + 627, + 470, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo-labeling quality. Figure 4a and 4b compare DECOLA Phase 1, OV-DETR, and baseline on c-AP for unseen classes. Compared to baseline and OV-DETR, DECOLA Phase 1 generates much higher quality pseudo-labels, especially in low-shot regimes. See examples in Figure 3.", + "bbox": [ + 75, + 657, + 470, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of conditioning. In Figure 4c, we compare c-AR of unseen classes. This measures the detector's ability to localize objects of interest when pseudo-labeling. We observe significant improvement in c-AR on both first-stage (proposals) and second-stage (predictions) due to our conditioning mechanism. This result demonstrates the key difference between DECOLA and other open-vocabulary detectors.", + "bbox": [ + 75, + 734, + 470, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo-labeling algorithms. Figure 4d shows the c-AP of baseline + self-train and DECOLA Phase 1 for unseen classes with 20 predictions per-image. Each red bar indicates the percent of training iteration during self-training. Online", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "self-labeling suffers a sharp drop during the early iterations, and c-AP after full iterations still underperforms compared to DECOLA Phase 1. DECOLA's simple approach of offline self-training is more stable and effective.", + "bbox": [ + 496, + 604, + 890, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 686, + 617, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore a new open-vocabulary detection framework, DECOLA. It adjusts its inner workings to the concepts that the user asks to reason over by conditioning on a language embedding. Our detector generates high-quality pseudo-labels on weakly labeled data through the conditioning mechanism. We finetune it with the pseudolabels to build the state-of-the-art open-vocabulary detector.", + "bbox": [ + 496, + 715, + 893, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank Xingyi Zhou, Yue Zhao, Jeffrey Ouyang-Zhang, and Nayeon Lee for valuable feedback. This material is in part based upon work supported by the National Science Foundation under Grant No. IIS-1845485 and IIS-2006820.", + "bbox": [ + 496, + 824, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16600", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Hénaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection. arXiv, 2023. 5", + "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 1, 2, 3", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3", + "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, 2019. 2", + "[5] Qiang Chen, Xiaokang Chen, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. arXiv preprint arXiv:2207.13085, 2022. 1, 2", + "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 5, 6, 7", + "[7] Jang Hyun Cho and Philipp Krahenbuhl. Long-tail detection with effective class-margins. In ECCV, 2022. 2, 3", + "[8] Jang Hyun Cho, Philipp Krahenbuhl, and Vignesh Ramanathan. Partdistillation: Learning parts from instance segmentation. In CVPR, 2023. 2", + "[9] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv, 2021. 5", + "[10] Mostafa et al. Dehghani. Scaling vision transformers to 22 billion parameters. ICML, 2023. 2", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 2", + "[12] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. 2022. 5, 7", + "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2", + "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 5", + "[15] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M Pratt, Vivek" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Neurips (Datasets and Benchmarks Track), 2023. 5, 6", + "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2", + "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2", + "[18] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. 1, 2", + "[19] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 2, 4, 5, 7", + "[20] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019, 1, 2, 4, 5", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6", + "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7", + "[23] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In Neurips, 2015. 2", + "[24] Gabriel et al Ilharco. Openclip, 2021. 2", + "[25] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. 2023. 1, 2, 3", + "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 5, 7", + "[27] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In ICCV, 2023. 5", + "[28] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Detection-oriented image-text pretraining for open-vocabulary detection. arXiv, 2023. 5, 6", + "[29] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, 2023. 5", + "[30] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 5", + "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2, 4, 5, 7", + "[32] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov," + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16601", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 4, 5", + "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2", + "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 5, 7", + "[35] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 5", + "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5", + "[37] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 1, 3", + "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5, 7", + "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 7", + "[40] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In Nuerips, 2023. 5", + "[41] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 2, 5, 6, 7", + "[42] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 1", + "[43] Matthias Minderer et al. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 2, 5", + "[44] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 2, 6", + "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Neurips. 2019. 6", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammed Uzair Khattak, Salman Khan, and Fahad Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Neurips, 2022. 5", + "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Neurips, 2015. 3", + "[49] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik. Imagenet-21k pretraining for the masses. In Neurips, 2021. 6", + "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 2, 5", + "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Neurips, 2022. 2", + "[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019. 4, 5", + "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2", + "[54] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 5", + "[55] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3", + "[56] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In CVPR, 2021.", + "[57] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In CVPR, 2021. 2, 3", + "[58] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, 2023. 5", + "[59] Tao Wang. Learning to detect and segment for open vocabulary object detection. In CVPR, 2023. 5", + "[60] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 3", + "[61] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 1, 2, 4, 5, 7", + "[62] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 5", + "[63] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. *Detectron2*, 2019. 6" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16602", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. In Neurips, 2023. 7", + "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, 2022. 1, 2, 4, 5, 6, 7", + "[66] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 1, 2", + "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022. 2", + "[68] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. MosaicOS: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2", + "[69] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Neurips, 2022. 7", + "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. 2023. 1, 2, 3, 6", + "[71] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2, 4, 5", + "[72] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 1, 2, 6, 7", + "[73] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 4, 5, 6, 7", + "[74] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 5", + "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\\{\\mathrm{detr}\\}$ : Deformable transformers for end-to-end object detection. In ICLR, 2021. 1, 3, 4, 6", + "[76] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. Neurips, 2020. 2" + ], + "bbox": [ + 78, + 90, + 470, + 729 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_model.json b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a5f14d390a3faecfdb260267a52c4b9c980898 --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.809, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.254, + 0.131, + 0.719, + 0.154 + ], + "angle": 0, + "content": "Language-conditioned Detection Transformer" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.182, + 0.424, + 0.215 + ], + "angle": 0, + "content": "Jang Hyun Cho UT Austin" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.22, + 0.465, + 0.234 + ], + "angle": 0, + "content": "janghyuncho7@utexas.edu" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.182, + 0.699, + 0.216 + ], + "angle": 0, + "content": "Philipp Krahenbuhl UT Austin" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.22, + 0.708, + 0.234 + ], + "angle": 0, + "content": "philkr@cs.utexas.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.633 + ], + "angle": 0, + "content": "We present a new open-vocabulary detection framework. Our framework uses both image-level labels and detailed detection annotations when available. Our framework proceeds in three steps. We first train a language-conditioned object detector on fully-supervised detection data. This detector gets to see the presence or absence of ground truth classes during training, and conditions prediction on the set of present classes. We use this detector to pseudolabel images with image-level labels. Our detector provides much more accurate pseudo-labels than prior approaches with its conditioning mechanism. Finally, we train an unconditioned open-vocabulary detector on the pseudo-annotated images. The resulting detector, named DECOLA, shows strong zero-shot performance in open-vocabulary LVIS benchmark as well as direct zero-shot transfer benchmarks on LVIS, COCO, Object365, and OpenImages. DECOLA outperforms the prior arts by 17.1 AP_rare and 9.4 mAP on zero-shot LVIS benchmark. DECOLA achieves state-of-the-art results in various model sizes, architectures, and datasets by only training on open-sourced data and academic-scale computing. Code is available at https://github.com/janghyuncho/DECOLA." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.21, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Object detection has seen immense progress over the past decade. Classical object detectors reason over datasets of fixed predefined classes. This simplifies the design, training, and evaluation of new methods, and allows for rapid prototyping [2-5, 17, 18, 25, 37, 42, 60, 70, 72, 75]. However, it complicates deployment to downstream applications too. A classical detector requires a new dataset to further finetune for every new concept it encounters. Collecting sufficient data for every new concept is not scalable [20]. Open-vocabulary detection offers an alternative [19, 43, 61, 65, 66, 73]. Open-vocabulary detectors reason about any arbitrary concept with free-form text, using the generalization ability of vision-language models. Yet, common open-vocabulary detectors reuse classical detectors and either replace the last classification layer" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.269, + 0.891, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Figure 1. An illustration of how standard open-vocabulary detectors and DECOLA generate pseudo-labels using image-level data. Standard detectors use image-level information later in the pipeline after initial box proposals, which may result in low coverage of unseen classes (e.g., \"mentos\" and \"cola\"). DECOLA adjusts the prediction to the information and ensures sufficient coverage." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.67, + 0.895, + 0.716 + ], + "angle": 0, + "content": "with [19, 43, 66, 73], or fuse box feature with [31, 65] text representation from pretrained vision-language model. The inner workings of the detector remain unchanged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a transformer-based object detector that adjusts its inner workings to any arbitrary set of concepts represented in language. The detector considers only the queried set of concepts as foreground and disregards any other objects as background. It learns to adapt detection to the language embedding of queried concepts at run-time. Specifically, the detector conditions proposal generation with respect to the text embedding of each queried concept and refines the conditioned proposals into predictions. Our detection transformer conditioned on language (DECOLA) offers a powerful alternative to classical architectures in open-vocabulary detection. Adapting the detector to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "language leads to stronger generalization to unseen concepts, and largely enhances self-training on weakly-labeled data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.471, + 0.348 + ], + "angle": 0, + "content": "DECOLA's ability to readily adapt to any queried concepts makes it particularly suitable for pseudo-labeling weakly-labeled data. Internet data, specifically images with paired text, is highly abundant and semantically rich [50, 51, 53]. The best vision models today build on this massive amount of weakly labeled data [10, 13, 24, 33, 34, 46, 67]. DECOLA leverages the same data to produce high-quality object detection labels from image-level annotations alone. DECOLA takes the image-level tags or text descriptions from the weakly labeled data and generates conditioned predictions as pseudo-labels. It efficiently processes multiple texts in parallel and only adds minimal computational overhead compared to the standard pseudo-labeling process. We finetune DECOLA on this rich detection dataset of pseudo-annotations and achieve the state-of-the-art open-vocabulary detector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.349, + 0.471, + 0.605 + ], + "angle": 0, + "content": "We evaluate our detector on popular open-vocabulary detection benchmarks on the LVIS dataset [19, 20, 66]. The final model improves the state-of-the-art methods by 4.4 and \\(4.9\\mathrm{AP}_{\\mathrm{novel}}\\) on open-vocabulary LVIS [19] benchmark, and 5.9 and \\(5.4\\mathrm{AP}_{\\mathrm{rare}}\\) on standard LVIS [20] benchmark, with ResNet-50 [21] and Swin-B [39] backbones, respectively. Our largest model with Swin-L achieves \\(10.4\\mathrm{AP}_{\\mathrm{rare}}\\) and \\(3.6\\mathrm{mAP}\\) improvement. Furthermore, DECOLA largely outperforms the state-of-the-art for direct zero-shot transfer benchmark on LVIS, by 12.0 and \\(17.1\\mathrm{AP}_{\\mathrm{rare}}\\) on LVIS minimal and LVIS v1.0, respectively. DECOLA consistently improves frequent, common, and rare classes altogether for different backbones and detection frameworks. Much of this improvement is driven by stronger pseudo-labeling capabilities. All our models are trained using open-sourced datasets with academic-scale computing. We open-source our code, pseudo-annotations, and checkpoints of all the model scales." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.619, + 0.22, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Open-vocabulary detection aims to detect objects of categories beyond the vocabulary of the training classes. A common solution is to inject language embeddings of class names in the last classification layer. OVR-CNN [66] pretrains a detector on image-caption data using BERT model [11] as language embedding. ViLD [19] trains a detector with CLIP text encoder [46] as language embedding with additional knowledge distillation [23] between predicted box features and the image encoder of CLIP. Detic [73] improves the above approaches through weakly-supervised learning on image-level annotations. RegionCLIP [71] introduces an intermediate pretraining step to better align CLIP to box features. BARON [61] improves the alignment between text and image encoders by extracting bag of regions as additional supervision. F-VLM [31] simplifies the training pipeline of open-vocabulary detection and explores the limit of the frozen vision-language model. All of the models above take" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "the design of the object detector as granted, and inject language in the last classification layer of the network. We take a different approach and design a detector that adapts predictions to particular categories of interest." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.895, + 0.497 + ], + "angle": 0, + "content": "Open-vocabulary DETR integrates DETR architecture into open-vocabulary detection. OWL [43] introduces a simple ViT architecture using pretrained CLIP and finetune with the DETR objective. OWLv2 uses self-training to further improve the performance [41]. OV-DETR [65] fuses features of a pretrained CLIP model with DETR object queries. Archi-tecturally, OV-DETR is closest to DECOLA. Both OV-DETR and DECOLA condition predictions on the text representation of classes. OV-DETR uses the original DETR queries and expands each with a CLIP feature per class for all classes. This leads to a quadratic number of queries, growing with the original DETR queries and with the classes considered. On the other hand, DECOLA explicitly controls the first-stage predictions (proposals) by formulating the scoring function to respect to the text embedding of each queried class at run-time. We visualize this difference in Figure ?? in the supplementary. The advantage is that we entirely remove inter-class competition and process a manageable amount of queries each focusing on a specific class, and running as fast as the vanilla Deformable DETR. This ability to freely adjust inner workings deviates DECOLA from prior works; it expands detection data through high-quality pseudo-labeling and achieves state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.897, + 0.817 + ], + "angle": 0, + "content": "Large-vocabulary object detection shares similar goals with open-vocabulary detection. Both learn from naturally long-tail data over large vocabularies. Vanilla large-vocabulary detectors are often ill-calibrated: The detector's final classification layer favors frequently seen objects over rare ones. This imbalance is usually addressed through a change in loss [7, 55-57, 72], or leveraging additional weakly labeled data for self-training [8, 16, 68, 76]. In large-vocabulary detection, R-CNN-based frameworks [2, 4, 17, 18, 72] dominate despite DETR-style architectures [5, 25, 44, 70] having long surpassed them on standard benchmarks [36]. DETR automatically assigns object queries to output classes, and thus it learns to more heavily focus queries on common classes. We show that language-conditioning helps address this calibration issue. Specifically, it removes inter-class competition in the training objective as queries are no longer shared across categories. As a result, DECOLA equally focuses on as many rare classes as frequent ones whenever they are present in an image. This yields a DETR-style detector that is competitive with the best R-CNN-based large-vocabulary detectors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.64, + 0.846 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Detection transformers (DETR) [3] build an object detection pipeline as a single feed-forward network. The network transforms object queries, arbitrary feature vectors, into la" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.091, + 0.808, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.894, + 0.404 + ], + "angle": 0, + "content": "Figure 2. Overview of DECOLA Phase 1 for conditioned prediction (top) and Phase 2 for open-vocabulary detection (bottom). For language-conditioned detection, each text embedding directly parameterizes the objectness function for each class. In Phase 2, the language-condition reads \"an object\" instead of particular class names, and predicts multi-class scores over all classes after decoding layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.532 + ], + "angle": 0, + "content": "beled bounding boxes through a series of cross-attention layers in a decoder architecture. Vanilla DETR [3] learns object queries as free-form parameters, while modern DETRs architectures [25, 37, 60, 70, 75] adopt a two-stage paradigm similar to RCNNs [48]. This query mechanism controls much of the inner workings of the detector. Queries determine what image regions the detector focuses on, and what object classes are prioritized." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.472, + 0.625 + ], + "angle": 0, + "content": "Query selection. Modern DETR architectures use image-dependent query selection, analogous to R-CNN's proposal generation [48]. An objectness function \\( s \\) scores each grid location \\( (i,j) \\) in the image using a feature \\( x_{i,j} \\) extracted from the transformer encoder. The top- \\( k \\) scored regions proceed to the second stage as object queries \\( Q \\):" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.634, + 0.47, + 0.653 + ], + "angle": 0, + "content": "\\[\ns \\left(x _ {i, j}\\right) = \\left\\langle x _ {i, j}, w \\right\\rangle , \\quad Q = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s \\left(x _ {i, j}\\right)\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.472, + 0.765 + ], + "angle": 0, + "content": "Here, \\( w \\) are the parameters of the objectness predictor. Each selected query produces a series of predictions that are refined over multiple iterations, similar to Cascade R-CNN [2]. The final prediction \\( \\vec{p} \\) contains scores over all classes \\( C \\) and an associated box. At a high level, DETR and R-CNNs share the same motivation: first, localize all objects in a scene, then refine their predictions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.766, + 0.472, + 0.872 + ], + "angle": 0, + "content": "Training objective. During training, DETR assigns each object query to an object or marks it as background. This allows DETR to learn non-overlapping object queries without post-processing such as non-maximum-suppression. The Hungarian matching algorithm finds the optimal assignment between all predictions \\( P \\) and all ground truth \\( G \\), minimizing the loss function as matching cost:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.882, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S}} \\ell (P, G | \\sigma) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.895, + 0.489 + ], + "angle": 0, + "content": "where \\(\\mathfrak{S}\\) captures all possible assignments from \\(P\\) to \\(G\\). For each assigned prediction, the loss \\(\\ell\\) maximizes its class log-likelihood and fits its bounding box. For unassigned predictions, the loss \\(\\ell\\) reduces both the objectness score \\(s\\) and class log-likelihood for all classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.499, + 0.614, + 0.515 + ], + "angle": 0, + "content": "4. DECOLA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Our detection transformer conditioned on language, DECOLA, changes the DETR architecture in one remarkable way: Object queries are conditioned on a language embedding. Figure 2 illustrates this change. This simple change has a few important implications: First, it allows the language embedding to control and focus queries to localize on the concepts at hand. Second, it removes any contention between different object classes. Each class present in the image uses the same amount of queries. Third, it generalizes to unseen classes by leveraging semantic knowledge encoded in language embedding throughout the detection pipeline. In the remainder of this section, we highlight the changes in the architecture and training objective for conditioning (DECOLA Phase 1), and self-training on image-level data for open-vocabulary detection (DECOLA Phase 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.894, + 0.827 + ], + "angle": 0, + "content": "Language-conditioned query selection. DECOLA conditions queries to a specific object category by modeling the objectness function as a similarity score between a region feature \\( x_{i,j} \\) and a text representation of a category name \\( t(y) \\) using their cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.833, + 0.893, + 0.867 + ], + "angle": 0, + "content": "\\[\ns _ {y} \\left(x _ {i, j}\\right) = \\frac {\\left\\langle x _ {i , j} , t (y) \\right\\rangle}{\\left\\| x _ {i , j} \\right\\| \\left\\| t (y) \\right\\|}, Q _ {y} = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s _ {y} \\left(x _ {i, j}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The above objective avoids any inter-class calibration issue common in imbalanced data [7, 55, 57]. Queries do not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16595" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.226 + ], + "angle": 0, + "content": "compete, as DECOLA independently selects top-\\(k\\) scoring regions \\(Q_{y}\\) for each class \\(y\\). All queries proceed to the second stage in parallel. A memory-efficient attention mechanism isolates interaction within each class. After a series of decoding layers, each language-conditioned query predicts a single scalar score, corresponding to the likelihood of the class \\(y\\), and the associated box. The overall architecture mirrors the two-stage deformable DETR [75] with two modifications: a language-conditioned query, and a binary output classifier." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.227, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Memory-efficient modeling. DECOLA uses \\( n = |Q_y| \\) queries per class for \\( K \\) classes. Generally, \\( n \\) is smaller than the total number of queries \\( |Q| \\) of a standard DETR model. However, since we produce \\( n \\) queries per class, the total number of queries in DECOLA is much larger \\( |Q| \\ll nK \\). A naive implementation of the DETR decoder is unable to cope with the \\( O(n^2 K^2) \\) memory requirements of the self-attention layers in the transformer decoder. We thus modify the self-attention formulation to isolate it within each class, reducing the memory cost to \\( O(n^2 K) \\). The actual implementation uses standard self-attention with a reshaping operation. See Figure 2 (right) for the illustration. DECOLA Phase 1: Train to condition on given concepts. Our goal is to design DECOLA to take a set of class names in an image (or a batch of images) and predict objects of the corresponding classes or backgrounds. For each class \\( y \\), each conditioned query \\( q_y \\in Q_y \\) therefore only predicts a single presence score for class \\( y \\) and the box location. All predictions from the conditioned queries \\( P_y \\) are matched with \\( G_y \\), the subset of ground truth with class \\( y \\):" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.538, + 0.47, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {y} ^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S} _ {y}} \\ell \\left(P _ {y}, G _ {y} \\mid \\sigma\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.675 + ], + "angle": 0, + "content": "where \\(\\mathfrak{S}_y\\) is the set of possible matches between \\(P_y\\) and \\(G_y\\), and \\(\\ell\\) is the binary cross-entropy loss. Unlike the original DETR objective in Eqn. 2, Eqn. 4 matches within the conditioned class. It avoids inter-class competition during training and simplifies the training objective. Instead, it learns to adapt its predictions to \\(y\\); the set of conditioned query \\(q_y\\) considers any objects other than \\(y\\) as background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Pseudo-labeling weakly-labeled data. DECOLA produces highly accurate predictions when conditioned on the exact categories of a scene, as shown in Section 5.4. This makes DECOLA a strong pseudo-labeler for weakly-labeled data with either image tags or captions. We expand a large amount of such data with pseudo-bounding boxes of DECOLA Phase 1 and self-train altogether to scale up open-vocabulary object detection. Unlike other forms of weakly supervised learning such as knowledge distillation [19] and online pseudolabeling [61, 73], we simply generate labels for all images offline and jointly train over all pseudo-labeled data using the regular detection losses without any additional complication or slowdown. For each image and class \\( y \\), DECOLA encodes the class' language feature and predicts a set of detections \\( P_{y} \\). We simply choose the most confident prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Figure 4d shows our simple offline pseudo-labeling works better than online pseudo-labeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.531 + ], + "angle": 0, + "content": "The advantage of DECOLA comes from adaptability to specified class names on a per-image basis. However, in open-vocabulary detection, the set of test classes is neither known a priori nor available per image. Hence, we convert DECOLA into a general-purpose detector to detect all objects. We condition DECOLA with \"an object\" as the text input, and inject the class information in the second-stage classifier. Figure 2 highlights this conversion. Since DECOLA is trained to align image features to text embedding in both the first and second stages, this change only introduces inter-class calibration for multi-class object detection. We train DECOLA with pseudo-labeled and human-labeled data as a standard supervised detection training, using the standard matching algorithm of DETR (Section 3). We do not introduce any additional hyper-parameter specifically for the weakly supervised learning [73] or design choices [31, 71], extra loss functions such as alignment loss [61], nor a large teacher model for knowledge distillation [19, 65]. Generating pseudo-labels with DECOLA Phase 1 runs as fast as a regular detector and training Phase 2 is as easy as standard detection training on a supervised dataset. At a high level, DECOLA Phase 1 training objective optimizes for a strong pseudo-labeler instead of a multi-class detector, which differentiates DECOLA from prior works. Leveraging DECOLA Phase 1 to expand weakly-labeled data is the key contribution to scaling up the final open-vocabulary object detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.634, + 0.559 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.895, + 0.675 + ], + "angle": 0, + "content": "We evaluate the effectiveness of DECOLA in two aspects: (1) pseudo-labeling quality of DECOLA Phase 1 (Section 5.4) and (2) benchmark evaluation (Section 5.3). We consider three primary benchmarks to evaluate our final model (DECOLA Phase 2): open-vocabulary LVIS [19], standard LVIS [20], and direct zero-shot evaluation to LVIS, COCO [36], Object365 [52], and OpenImages [32]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.692, + 0.699 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Datasets and benchmarks. We mainly evaluate our method on the LVIS dataset [20], a large-vocabulary instance segmentation and object detection dataset with 1203 naturally distributed object categories. LVIS splits categories into frequent, common, and rare. For open-vocabulary LVIS, we combine frequent and common categories into LVIS-base and consider the rare categories as novel concepts used for testing only [19]. For standard LVIS, we train and evaluate all classes. Direct zero-shot transfer evaluates models trained on different detection data (e.g., Object365) and other weakly-labeled data without any prior knowledge about the target dataset such as the set of classes or object frequency. In this benchmark, we test DECOLA's generalization to dif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16596" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.192, + 0.088, + 0.778, + 0.345 + ], + "angle": 0, + "content": "
methoddataAPbox novelAPcAPfmAPbox
ResNet-50 (1K)
OV-DETR† [65]LVIS-base, IN-21K18.024.831.826.4
baselineLVIS-base10.230.938.030.1
baseline + self-trainLVIS-base, IN-21K19.231.737.131.7
DECOLA Phase 2LVIS-base, IN-21K23.8 (+4.6)34.4 (+2.7)38.3 (+1.2)34.1 (+2.4)
ResNet-50
baselineLVIS-base9.433.840.432.2
baseline + self-trainLVIS-base, IN-21K23.236.541.636.2
DECOLA Phase 2LVIS-base, IN-21K27.6 (+4.4)38.3 (+1.8)42.9 (+1.3)38.3 (+2.1)
Swin-B
baselineLVIS-base16.243.849.141.1
baseline + self-trainLVIS-base, IN-21K30.843.645.942.3
DECOLA Phase 2LVIS-base, IN-21K35.7 (+4.9)47.5 (+3.9)49.7 (+3.8)46.3 (+4.0)
Swin-L
DITO* [28]O365, LVIS-base, DataComp-1B45.8--44.2
OWLv2‡ [41]O365, LVIS-base, VG, WebLI45.9--50.4
baselineO365, LVIS-base21.953.357.749.6
baseline + self-trainO365, LVIS-base, IN-21K36.553.556.551.8
DECOLA Phase 2O365, LVIS-base, IN-21K46.9 (+10.4)56.0 (+2.5)58.0 (+1.5)55.2 (+3.6)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.348, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Table 1. Open-vocabulary LVIS using DETR architectures. \\(\\dagger\\): we further finetune the official OV-DETR model trained on LVIS-base by self-training on ImageNet-21K data similar to \"baseline + self-train\" and DECOLA. \\(\\ddagger\\): uses CLIP L/14, which is comparable to Swin-L backbone. \\(\\star\\): is based on Mask R-CNN with ViT L/16. We report the improvement between \"baseline + self-train\" to DECOLA in green. Last 4 rows compare DECOLA to Swin-L or equivalent scale of models that use additional detection data (e.g., Objects365, VG [30]) and billion-scale weakly-labeled data (DataComp-1B [15], WebLI [6]) for training." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.431, + 0.481, + 0.624 + ], + "angle": 0, + "content": "
methodframework\\(AP^{box}\\)novel\\(mAP^{box}\\)novel\\(AP^{mask}\\)novel\\(mAP^{mask}\\)
ViLD [19]Mask R-CNN16.727.816.625.5
RegionCLIP [71]Mask R-CNN--17.122.5
DetPro [12]Mask R-CNN20.828.419.825.9
PromptDet [14]Mask R-CNN21.425.3--
F-VLM [31]Mask R-CNN--18.624.2
BARON [61]Mask R-CNN23.229.522.627.6
OADP [58]Mask R-CNN21.928.721.726.6
EdaDet [54]Mask R-CNN--23.727.5
VLDet [35]CenterNet2--21.730.1
\\(\\mathsf{CORA}^{+}\\) [62]CenterNet228.1---
Rasheed et al. [47]CenterNet2--25.232.9
Detic-base [73]CenterNet217.633.816.430.2
Detic [73]CenterNet226.736.324.632.4
DECOLA labelsCenterNet229.537.727.033.7
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.171, + 0.626, + 0.398, + 0.638 + ], + "angle": 0, + "content": "(a) Comparison with ResNet-50 backbone." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.431, + 0.9, + 0.624 + ], + "angle": 0, + "content": "
methodbackbone\\(AP_{\\text{novel}}^{box}\\)\\(mAP_{\\text{novel}}^{box}\\)\\(AP_{\\text{novel}}^{mask}\\)\\(mAP_{\\text{novel}}^{mask}\\)
RegionCLIP [71]R50×4--22.032.3
CondHead [59]R50×424.133.724.432.0
ViLD [19]EN-B7--26.329.3
OWL-ViT [43]ViT-L/1425.634.7--
F-VLM [31]R50×64--32.834.9
VLDet [35]Swin-B--26.338.1
3Ways [1]NFNet-F630.144.6--
RO-VIT [29]ViT-L/1632.134.0--
CFM-ViT [27]ViT-L/1635.638.533.936.6
DITO [28]ViT-B/1634.936.932.534.0
CoDet [40]Swin-B--29.439.2
Detic-base [73]Swin-B24.643.021.938.4
Detic [73]Swin-B36.645.733.840.7
DECOLA labelsSwin-B38.446.735.342.0
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.619, + 0.625, + 0.774, + 0.638 + ], + "angle": 0, + "content": "(b) System-level comparison." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.643, + 0.893, + 0.686 + ], + "angle": 0, + "content": "Table 2. Open-vocabulary LVIS using Mask R-CNN and CenterNet2 detectors. Methods in both tables use LVIS-base as the only human-labeled data for fair comparison. For system-level comparison (right), we include methods with non R-CNN architectures such as OWL-ViT [43]. The results show the impact of high-quality pseudo-labels generated by DECOLA Phase 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.697, + 0.471, + 0.775 + ], + "angle": 0, + "content": "ferent domains. We evaluate DECOLA on LVIS, COCO [36], Object365 [52], and OpenImages [32] in a fully zero-shot manner. All our models use the ImageNet-21K [50] dataset as weakly labeled data, which contains 14M of object-centric images annotated with a single class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate DECOLA on \\(\\mathrm{AP}_{\\mathrm{novel/rare}}\\), \\(\\mathrm{AP_c}\\), \\(\\mathrm{AP_f}\\), and mAP following the LVIS evaluation metric [20]. We highlight the results in all three groups since we believe open-vocabulary detectors should not compensate for the performance of common/frequent classes for novel/rare classes. We evaluate both \\(\\mathrm{AP^{box}}\\) and \\(\\mathrm{AP^{mask}}\\) for object detection and instance segmentation. For zero-shot transfer benchmark with COCO and Object365, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.698, + 0.895, + 0.895 + ], + "angle": 0, + "content": "use AP, \\(\\mathrm{AP}_{50}\\), and \\(\\mathrm{AP}_{75}\\) following prior work [19, 73]. For OpenImages, we report \\(\\mathrm{AP}_{50}^{\\mathrm{flat}}\\) on the expanded label space [73, 74]. For zero-shot transfer to LVIS, we consider LVIS minival [26] and standard LVIS v1.0 validation set and report \\(\\mathrm{AP}^{\\mathrm{fixed}}\\) [9] following the prior works [26, 34, 38]. In addition, we pursue a more direct measurement of the generated pseudo-labeling quality. Hence, we define conditioned mAP/AR (c-mAP/AR) and compare it to baseline open-vocabulary detectors. c-mAP measures the detection performance in mAP when the detector is provided the set of ground truth classes in each image. For example in Figure 1, both detectors use \"cat\", \"mentos\" and \"cola\" as given. This extra information is used to select scores to rank" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16597" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.473, + 0.257 + ], + "angle": 0, + "content": "
methoddataAPboxraremAPbox
ResNet-50
baselineLVIS26.335.6
baseline + self-trainLVIS, IN-21K30.036.6
DECOLA Phase 2LVIS, IN-21K35.9 (+5.9)39.4 (+2.8)
Swin-B
baselineLVIS38.344.5
baseline + self-trainLVIS, IN-21K42.045.2
DECOLA Phase 2LVIS, IN-21K47.4 (+5.4)48.3 (+3.1)
Swin-L
baselineO365, LVIS49.354.4
baseline + self-trainO365, LVIS, IN-21K48.753.4
DECOLA Phase 2O365, LVIS, IN-21K54.9 (+6.2)56.4 (+3.0)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.259, + 0.358, + 0.269 + ], + "angle": 0, + "content": "(a) Standard LVIS with DETR." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.271, + 0.473, + 0.389 + ], + "angle": 0, + "content": "
methodAPboxraremAPboxAPmaskraremAPmask
ResNet-50
Detic-base [73]28.235.325.631.4
Detic [73]31.436.829.733.2
DECOLA labels35.6 (+4.2)38.6 (+1.8)32.1 (+2.4)34.4 (+1.2)
Swin-B
Detic-base [73]39.945.435.940.7
Detic [73]45.846.941.741.7
DECOLA labels47.6 (+1.8)48.5 (+1.6)43.7 (+2.0)43.6 (+1.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.39, + 0.371, + 0.402 + ], + "angle": 0, + "content": "(b) Standard LVIS with CenterNet2." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.435 + ], + "angle": 0, + "content": "Table 3. Standard LVIS benchmark. DECOLA shows consistent improvement over different model scales and architectures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.472, + 0.488 + ], + "angle": 0, + "content": "the final predictions (belines), or directly condition the detector (DECOLA). We analyze the model's behavior and label quality in Section 5.4 and Section ?? in supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.17, + 0.513 + ], + "angle": 0, + "content": "5.2. Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.703 + ], + "angle": 0, + "content": "DECOLA is based on two-stage Deformable DETR [75]. As described in Section 4, the first-stage objectness function for query selection is replaced by a similarity score between the image feature and the CLIP text embedding of each class name. We train the detector with the improved DETR training recipe [44, 70]: look-forward-twice, larger MLP hidden dimension, no dropout, etc. We consider four backbones: a ResNet-50 [21], Swin-B and L for all LVIS benchmarks, and Swin-T and L for the direct zero-shot transfer. Unless otherwise mentioned, all backbones are pretrained on the ImageNet-21K dataset [49]. Next, we describe our key baseline models to directly compare to DECOLA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Baseline. We design a baseline open-vocabulary detector to closely compare to DECOLA. Inspired by Detic [73], baseline replaces classification layers with the class embedding of the pretrained CLIP text encoder and is trained using Federated Loss [72]. DECOLA Phase 1 and baseline are trained using human-labeled data (e.g., LVIS-base). All other settings (training dataset, number of training iterations, etc.) are kept the same between DECOLA and baseline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Baseline + self-train. Similar to DECOLA Phase 2, we self-train baseline on weakly-labeled data. For the self-training algorithm, we use online self-training with max-size loss from Detic [73] as baseline comparison (baseline + self-train) to DECOLA Phase 2. We tested max-size and max" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "score losses from Detic [73] (online pseudo-labeling) as well as offline pseudo-labeling similar to DECOLA, and max-size loss consistently performed the best." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.894, + 0.228 + ], + "angle": 0, + "content": "DECOLA labels. We train a two-stage detector for broader comparison: CenterNet2 [72]. Specifically, we use Detric's baseline model (\"Detric-base\"), a CenterNet2 trained on LVIS-base with CLIP embedding, and finetune on pseudolabeled ImageNet-21K data using DECOLA Phase 1 of the same backbone size. We denote this as \"DECOLA labels\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.23, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Efficient modeling. For DECOLA Phase 1, we use \\( n = 300 \\) queries per class. One memory and time bottleneck during DECOLA training is the first-stage loss computation. The original Deformable DETR computes Hungarian matching with all pixels to all objects in a class-agnostic manner, which is \\( \\sum_{l \\in L} H_l \\cdot W_l \\) predictions. To reduce the memory and time cost, we only consider the top \\( K = 10,000 \\) confident pixels for each class \\( y \\) during the first-stage matching and loss computation. Together with memory-efficient self-attention (Sec. 4), the training time and memory cost of DECOLA increases by less than \\( 20\\% \\) over the baselines (See Table 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.895, + 0.548 + ], + "angle": 0, + "content": "Training details. Following Detic [73], we train DECOLA Phase 1 and baseline \\(4 \\times\\) on LVIS-base, and further fine-tune for another \\(4 \\times\\) on ImageNet-21K data with pseudolabeling. For training CenterNet2 with DECOLA labels, we combine pseudo-labels from different image resolutions for \\(h \\in \\{240, 280, 320, 360, 400\\}\\), where \\(h\\) is the shorter side of the image. Note that this mimics the random image resizing data augmentation during standard detection training. We use Detector2 [63] based on PyTorch [45] in all of the experiments. More details are in Section ?? of supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.638, + 0.575 + ], + "angle": 0, + "content": "5.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Open-vocabulary LVIS. Table 1 compares DECOLA to baseline as well as state-of-the-art DETR-based open-vocabulary detectors; OV-DETR [65], OWLv2 [41], and baseline + self-train. For a fair comparison, we further fine-tune the official OV-DETR model checkpoint on ImageNet21K for \\(4 \\times\\) schedule same as DECOLA Phase 2 and baseline + self-train. For all backbone scales, we show consistent improvement over other methods. Notably, baseline + self-train exhibits degradation in frequent classes as a trade-off with improved novel classes, which is commonly observed behavior in other open-vocabulary detection methods, too. DECOLA improves all categories consistently, which highlights the quality of our pseudo-labels. In the last rows with the Swin-L backbone, we report the result of two concurrent works, DITO [28] (Mask R-CNN-based) and OWLv2, to compare to the method that uses additional detection data (Object365) and billion-scale web data (DataComp1B [15], WebLI [6]). DECOLA demonstrates large improvement over the state-of-the-arts despite using orders of magnitude smaller training data and compute resources. To further examine DECOLA's scalability, we test the pseudo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16598" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.875, + 0.319 + ], + "angle": 0, + "content": "
methoddataLVIS minivalLVIS v1.0 val
APboxrareAPcAPfmAPboxAPboxrareAPcAPfmAPbox
Swin-T
MDETR* [26]LVIS, GoldG, RefC20.924.924.324.27.422.725.022.5
GLIP [34]O365, GoldG, Cap4M20.821.431.026.010.112.525.517.2
GroundingDINO [38]O365, GoldG, Cap4M18.123.332.727.4----
GLIPv2 [69]O365, GoldG, Cap4M---29.0----
MQ-GroundingDINO† [64]O365, GoldG, Cap4M, LVIS-5VQ21.726.235.230.212.917.431.422.1
MQ-GLIP† [64]O365, GoldG, Cap4M, LVIS-5VQ21.027.534.630.415.418.430.422.6
DECOLA Phase 2O365, IN-21K‡32.832.031.832.027.224.928.026.6
Δ(+12.0)(+8.7)-(+3.0)(+17.1)(+12.4)(+2.5)(+9.4)
Swin-L
GLIP [34]FourODs, GoldG, Cap24M28.234.341.537.317.123.335.426.9
GroundingDINO [38]O365, OI, GoldG, Cap4M, COCO, RefC22.230.738.833.9----
MQ-GLIP† [64]FourODs, GoldG, Cap24M, LVIS-5VQ34.541.246.943.426.932.041.334.7
OWLv2 [41]O365, VG, WebLI39.0--38.134.9--33.5
DECOLA Phase 2O365, OI, IN-21K‡41.538.034.936.832.929.130.330.2
Δ(+2.5)(+3.7)---(+5.8)--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.323, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Table 4. Direct zero-shot transfer to LVIS. †: methods use 5 per-class vision queries of LVIS dataset (denoted as “LVIS-5VQ”), which use images and annotations to extract instance-level features. ☆: MDETR uses ResNet-101 backbone and trained fully-supervised on LVIS. Results that use LVIS data are in gray. ‡: Full ImageNet-21K. No LVIS information is used in DECOLA." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.37, + 0.477, + 0.524 + ], + "angle": 0, + "content": "
methodCOCOO365OI
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{50}^{flat} \\)
R-CNNs
ViLD [19]36.655.639.811.818.212.6-
F-VLM [31]32.553.134.611.919.212.6-
DetPro [12]34.953.837.412.118.812.9-
BARON [61]36.356.139.313.621.014.5-
Detic [73]39.156.342.214.220.715.242.9
DETRs
OV-DETR [65]38.158.441.1----
Detic [73]39.856.643.314.521.415.541.6
DECOLA Phase 240.357.043.715.022.016.043.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.529, + 0.471, + 0.569 + ], + "angle": 0, + "content": "Table 5. Cross-dataset generalization benchmark on COCO, Object365, and OpenImages. All models use a ResNet-50 backbone and train on LVIS-base and weakly-labeled data." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.569, + 0.425, + 0.658 + ], + "angle": 0, + "content": "
methodtraintest
timemem.timemem.
baseline44 h8.9 G0.07 s/img2.5 G
+ self-train45 h10.2 G0.07 s/img2.5 G
OV-DETR73 h22.0 G6.4 s/img3.4 G
DECOLA Phase 149 h12.6 G--
DECOLA Phase 245 h8.9 G0.07 s/img2.8 G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.663, + 0.471, + 0.706 + ], + "angle": 0, + "content": "Table 6. Efficiency. Training time is measured with 8 DGX V100 on ResNet-50, 2 images per-GPU. float16 is used in both training and testing. OV-DETR uses the original optimized inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.47, + 0.854 + ], + "angle": 0, + "content": "labeling capability of DECOLA on R-CNN-based detectors with DECOLA labels (Detic-base finetuned with our pseudolabels). In Table 2, we compare DECOLA labels to a broad range of literature based on CenterNet2 [72] and Mask R-CNN [22]. Table 2a compares methods with ResNet-50 backbone and Table 2b compares larger scale backbones for system-level comparison. In both tables, DECOLA clearly improve upon the state-of-the-art by large margins without additional complication in training and bells-and-whistles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Standard LVIS. Tables 3 evaluate DECOLA and baseline on the standard LVIS benchmark, where all object categories are used to fully supervise the detectors. Similar to the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.901 + ], + "angle": 0, + "content": "open-vocabulary LVIS, we compare DETR architectures in Table 3a and R-CNN architectures in Table 3b. Table 3a shows that DECOLA remarkably improves baseline by 9.5, 9.1, and 5.6 points on \\(\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}\\), outperforming baseline + self-train by 5.9, 5.4, and 6.2 \\(\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}\\) for ResNet-50, Swin-B, and Swin-L backbones, respectively. Similarly in Table 3b, DECOLA labels further improves the baseline of Detic by 7.4 and 7.7 \\(\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}\\), and outperforms Detic [73] by 4.2 and 1.8 \\(\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}\\) with ResNet-50 and Swin-B backbones, respectively. Direct zero-shot evaluation. For direct zero-shot evaluation, we train DECOLA with Swin-T [39] and use Object365 data for Phase 1, and ImageNet-21K for Phase 2 (full dataset and classes). We compare to MDETR [26], GLIP [34], GroundingDINO [38], and MQ-Det [64] finetuned from GLIP and GroundingDINO. Table 4 shows the results. DECOLA outperforms the previous state-of-the-arts, by \\(12.0 / 17.1\\mathrm{AP}_{\\mathrm{rare}}\\) and \\(3.0 / 9.4\\mathrm{mAP}\\) on LVIS minival and LVIS v1.0 val, respectively. Note that all other methods use much richer detection labels from GoldG data [26], a collection of grounding data (box and text expression pairs) curated by MDETR. Furthermore, other benchmark methods show highly imbalanced \\(\\mathrm{AP}_{\\mathrm{rare}}\\) and \\(\\mathrm{AP_f}\\) in both LVIS minival and LVIS v1.0 val (10-20 points gap). We hypothesize that the large collection of training data coincides with LVIS vocabulary, as all data follows a natural distribution of common objects. Also, DECOLA enjoys significantly faster run-time compared to all other models that undergo BERT encoding for grounding, which requires more than 50 forward passes per image in order to predict all LVIS categories. Similarly, our Swin-L model outperforms GroundingDINO and GLIP by 19.3 and \\(13.3\\mathrm{AP}_{\\mathrm{rare}}\\), respectively, despite much smaller training data compared to FourODs [34] and match to OWLv2 with 10-B private WebLI data [6]. Table 5 further examines the generality of DECOLA on different domains. DECOLA Phase 2 with ResNet-50 outperforms all other competitive baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.166, + 0.092, + 0.808, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.364, + 0.892, + 0.421 + ], + "angle": 0, + "content": "Figure 3. Examples of prediction on unseen categories. Images from ImageNet-21K dataset. Boxes are the most confident prediction from DECOLA and baseline. DECOLA conditions on the ground truth label, and baseline selects the max-score box of the ground truth class. Images are all from unseen categories, which neither model was trained on. Green: DECOLA Phase 1 trained on LVIS-base. Red: baseline trained on LVIS-base. Both models use a Deformable DETR detector with a ResNet-50 backbone. More in Fig. ?? and ?? of supplementary." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.428, + 0.278, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.551, + 0.267, + 0.562 + ], + "angle": 0, + "content": "(a) c-AP for ResNet-50 (IN-1K)" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.428, + 0.473, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.551, + 0.435, + 0.561 + ], + "angle": 0, + "content": "(b) c-AP for Swin-B" + }, + { + "type": "image", + "bbox": [ + 0.475, + 0.428, + 0.687, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.494, + 0.551, + 0.687, + 0.562 + ], + "angle": 0, + "content": "(c) c-AR for 1st and 2nd stage boxes" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.428, + 0.888, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.551, + 0.869, + 0.561 + ], + "angle": 0, + "content": "(d) online vs offline: c-AP@20" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.568, + 0.89, + 0.596 + ], + "angle": 0, + "content": "Figure 4. Analyzing DECOLA. All plots show the conditioned AP/AR for unseen classes. We compare DECOLA Phase 1 and baselines. We highlight more detailed analyses about c-AP and c-AR in Section ??, Table ?? and ?? of the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.604, + 0.182, + 0.619 + ], + "angle": 0, + "content": "5.4. Analyses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.471, + 0.657 + ], + "angle": 0, + "content": "In this section, we analyze the model's behavior with conditioned mAP/AR (c-AP/AR) (defined in Sec. 5.1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.734 + ], + "angle": 0, + "content": "Pseudo-labeling quality. Figure 4a and 4b compare DECOLA Phase 1, OV-DETR, and baseline on c-AP for unseen classes. Compared to baseline and OV-DETR, DECOLA Phase 1 generates much higher quality pseudo-labels, especially in low-shot regimes. See examples in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.84 + ], + "angle": 0, + "content": "Impact of conditioning. In Figure 4c, we compare c-AR of unseen classes. This measures the detector's ability to localize objects of interest when pseudo-labeling. We observe significant improvement in c-AR on both first-stage (proposals) and second-stage (predictions) due to our conditioning mechanism. This result demonstrates the key difference between DECOLA and other open-vocabulary detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Pseudo-labeling algorithms. Figure 4d shows the c-AP of baseline + self-train and DECOLA Phase 1 for unseen classes with 20 predictions per-image. Each red bar indicates the percent of training iteration during self-training. Online" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.605, + 0.892, + 0.665 + ], + "angle": 0, + "content": "self-labeling suffers a sharp drop during the early iterations, and c-AP after full iterations still underperforms compared to DECOLA Phase 1. DECOLA's simple approach of offline self-training is more stable and effective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.688, + 0.619, + 0.704 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.894, + 0.822 + ], + "angle": 0, + "content": "In this paper, we explore a new open-vocabulary detection framework, DECOLA. It adjusts its inner workings to the concepts that the user asks to reason over by conditioning on a language embedding. Our detector generates high-quality pseudo-labels on weakly labeled data through the conditioning mechanism. We finetune it with the pseudolabels to build the state-of-the-art open-vocabulary detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.899 + ], + "angle": 0, + "content": "Acknowledgement. We thank Xingyi Zhou, Yue Zhao, Jeffrey Ouyang-Zhang, and Nayeon Lee for valuable feedback. This material is in part based upon work supported by the National Science Foundation under Grant No. IIS-1845485 and IIS-2006820." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16600" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.169 + ], + "angle": 0, + "content": "[1] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Hénaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection. arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.201, + 0.471, + 0.253 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.256, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.312, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[5] Qiang Chen, Xiaokang Chen, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. arXiv preprint arXiv:2207.13085, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.369, + 0.471, + 0.506 + ], + "angle": 0, + "content": "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.508, + 0.468, + 0.534 + ], + "angle": 0, + "content": "[7] Jang Hyun Cho and Philipp Krahenbuhl. Long-tail detection with effective class-margins. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.536, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[8] Jang Hyun Cho, Philipp Krahenbuhl, and Vignesh Ramanathan. Partdistillation: Learning parts from instance segmentation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.579, + 0.471, + 0.619 + ], + "angle": 0, + "content": "[9] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[10] Mostafa et al. Dehghani. Scaling vision transformers to 22 billion parameters. ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[12] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M Pratt, Vivek" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.894, + 0.203 + ], + "angle": 0, + "content": "Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Neurips (Datasets and Benchmarks Track), 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.814, + 0.274 + ], + "angle": 0, + "content": "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.276, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[19] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.894, + 0.402 + ], + "angle": 0, + "content": "[20] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019, 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[23] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In Neurips, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.771, + 0.518 + ], + "angle": 0, + "content": "[24] Gabriel et al Ilharco. Openclip, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[25] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.894, + 0.658 + ], + "angle": 0, + "content": "[27] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.702 + ], + "angle": 0, + "content": "[28] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Detection-oriented image-text pretraining for open-vocabulary detection. arXiv, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[29] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[30] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.894, + 0.87 + ], + "angle": 0, + "content": "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[32] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16601" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.471, + 0.134 + ], + "angle": 0, + "content": "Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.472, + 0.287 + ], + "angle": 0, + "content": "[35] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.343 + ], + "angle": 0, + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.472, + 0.397 + ], + "angle": 0, + "content": "[37] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.472, + 0.468 + ], + "angle": 0, + "content": "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.472, + 0.523 + ], + "angle": 0, + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.567 + ], + "angle": 0, + "content": "[40] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In Nuerips, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.472, + 0.595 + ], + "angle": 0, + "content": "[41] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[42] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.472, + 0.678 + ], + "angle": 0, + "content": "[43] Matthias Minderer et al. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.472, + 0.72 + ], + "angle": 0, + "content": "[44] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Neurips. 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammed Uzair Khattak, Salman Khan, and Fahad Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Neurips, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Neurips, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[49] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik. Imagenet-21k pretraining for the masses. In Neurips, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.894, + 0.37 + ], + "angle": 0, + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Neurips, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.894, + 0.425 + ], + "angle": 0, + "content": "[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.427, + 0.894, + 0.48 + ], + "angle": 0, + "content": "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.482, + 0.894, + 0.51 + ], + "angle": 0, + "content": "[54] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[55] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.894, + 0.594 + ], + "angle": 0, + "content": "[56] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.894, + 0.649 + ], + "angle": 0, + "content": "[57] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.894, + 0.704 + ], + "angle": 0, + "content": "[58] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[59] Tao Wang. Learning to detect and segment for open vocabulary object detection. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.894, + 0.774 + ], + "angle": 0, + "content": "[60] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.894, + 0.829 + ], + "angle": 0, + "content": "[61] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.894, + 0.873 + ], + "angle": 0, + "content": "[62] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[63] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. *Detectron2*, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16602" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[64] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. In Neurips, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.471, + 0.219 + ], + "angle": 0, + "content": "[66] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[68] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. MosaicOS: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.471, + 0.401 + ], + "angle": 0, + "content": "[69] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Neurips, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.405, + 0.47, + 0.459 + ], + "angle": 0, + "content": "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.462, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[71] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[72] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.562, + 0.471, + 0.614 + ], + "angle": 0, + "content": "[73] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.471, + 0.644 + ], + "angle": 0, + "content": "[74] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable \\(\\{\\mathrm{detr}\\}\\): Deformable transformers for end-to-end object detection. In ICLR, 2021. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.689, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[76] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. Neurips, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16603" + } + ] +] \ No newline at end of file diff --git a/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_origin.pdf b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..64b8cace9d502e42c95ae022c6c66cbd768152ae --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/4052d6d0-e8f9-4912-885e-f4d1033b20c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d43c008f02f8f23fcbf1d84462ac4e4eeec8c3c8e7117447c1c8e3a2afbfe39c +size 2803316 diff --git a/2024/Language-conditioned Detection Transformer/full.md b/2024/Language-conditioned Detection Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..90448a3a2878b9af4087b0e811d5f0f46a890768 --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/full.md @@ -0,0 +1,290 @@ +# Language-conditioned Detection Transformer + +Jang Hyun Cho UT Austin + +janghyuncho7@utexas.edu + +Philipp Krahenbuhl UT Austin + +philkr@cs.utexas.edu + +# Abstract + +We present a new open-vocabulary detection framework. Our framework uses both image-level labels and detailed detection annotations when available. Our framework proceeds in three steps. We first train a language-conditioned object detector on fully-supervised detection data. This detector gets to see the presence or absence of ground truth classes during training, and conditions prediction on the set of present classes. We use this detector to pseudolabel images with image-level labels. Our detector provides much more accurate pseudo-labels than prior approaches with its conditioning mechanism. Finally, we train an unconditioned open-vocabulary detector on the pseudo-annotated images. The resulting detector, named DECOLA, shows strong zero-shot performance in open-vocabulary LVIS benchmark as well as direct zero-shot transfer benchmarks on LVIS, COCO, Object365, and OpenImages. DECOLA outperforms the prior arts by 17.1 AP_rare and 9.4 mAP on zero-shot LVIS benchmark. DECOLA achieves state-of-the-art results in various model sizes, architectures, and datasets by only training on open-sourced data and academic-scale computing. Code is available at https://github.com/janghyuncho/DECOLA. + +# 1. Introduction + +Object detection has seen immense progress over the past decade. Classical object detectors reason over datasets of fixed predefined classes. This simplifies the design, training, and evaluation of new methods, and allows for rapid prototyping [2-5, 17, 18, 25, 37, 42, 60, 70, 72, 75]. However, it complicates deployment to downstream applications too. A classical detector requires a new dataset to further finetune for every new concept it encounters. Collecting sufficient data for every new concept is not scalable [20]. Open-vocabulary detection offers an alternative [19, 43, 61, 65, 66, 73]. Open-vocabulary detectors reason about any arbitrary concept with free-form text, using the generalization ability of vision-language models. Yet, common open-vocabulary detectors reuse classical detectors and either replace the last classification layer + +![](images/eab37f5762d533fbfcf4ee6003dca2f7f131377e0e31f25de59d1d33b5eec667.jpg) +Figure 1. An illustration of how standard open-vocabulary detectors and DECOLA generate pseudo-labels using image-level data. Standard detectors use image-level information later in the pipeline after initial box proposals, which may result in low coverage of unseen classes (e.g., "mentos" and "cola"). DECOLA adjusts the prediction to the information and ensures sufficient coverage. + +with [19, 43, 66, 73], or fuse box feature with [31, 65] text representation from pretrained vision-language model. The inner workings of the detector remain unchanged. + +In this paper, we introduce a transformer-based object detector that adjusts its inner workings to any arbitrary set of concepts represented in language. The detector considers only the queried set of concepts as foreground and disregards any other objects as background. It learns to adapt detection to the language embedding of queried concepts at run-time. Specifically, the detector conditions proposal generation with respect to the text embedding of each queried concept and refines the conditioned proposals into predictions. Our detection transformer conditioned on language (DECOLA) offers a powerful alternative to classical architectures in open-vocabulary detection. Adapting the detector to + +language leads to stronger generalization to unseen concepts, and largely enhances self-training on weakly-labeled data. + +DECOLA's ability to readily adapt to any queried concepts makes it particularly suitable for pseudo-labeling weakly-labeled data. Internet data, specifically images with paired text, is highly abundant and semantically rich [50, 51, 53]. The best vision models today build on this massive amount of weakly labeled data [10, 13, 24, 33, 34, 46, 67]. DECOLA leverages the same data to produce high-quality object detection labels from image-level annotations alone. DECOLA takes the image-level tags or text descriptions from the weakly labeled data and generates conditioned predictions as pseudo-labels. It efficiently processes multiple texts in parallel and only adds minimal computational overhead compared to the standard pseudo-labeling process. We finetune DECOLA on this rich detection dataset of pseudo-annotations and achieve the state-of-the-art open-vocabulary detector. + +We evaluate our detector on popular open-vocabulary detection benchmarks on the LVIS dataset [19, 20, 66]. The final model improves the state-of-the-art methods by 4.4 and $4.9\mathrm{AP}_{\mathrm{novel}}$ on open-vocabulary LVIS [19] benchmark, and 5.9 and $5.4\mathrm{AP}_{\mathrm{rare}}$ on standard LVIS [20] benchmark, with ResNet-50 [21] and Swin-B [39] backbones, respectively. Our largest model with Swin-L achieves $10.4\mathrm{AP}_{\mathrm{rare}}$ and $3.6\mathrm{mAP}$ improvement. Furthermore, DECOLA largely outperforms the state-of-the-art for direct zero-shot transfer benchmark on LVIS, by 12.0 and $17.1\mathrm{AP}_{\mathrm{rare}}$ on LVIS minimal and LVIS v1.0, respectively. DECOLA consistently improves frequent, common, and rare classes altogether for different backbones and detection frameworks. Much of this improvement is driven by stronger pseudo-labeling capabilities. All our models are trained using open-sourced datasets with academic-scale computing. We open-source our code, pseudo-annotations, and checkpoints of all the model scales. + +# 2. Related Work + +Open-vocabulary detection aims to detect objects of categories beyond the vocabulary of the training classes. A common solution is to inject language embeddings of class names in the last classification layer. OVR-CNN [66] pretrains a detector on image-caption data using BERT model [11] as language embedding. ViLD [19] trains a detector with CLIP text encoder [46] as language embedding with additional knowledge distillation [23] between predicted box features and the image encoder of CLIP. Detic [73] improves the above approaches through weakly-supervised learning on image-level annotations. RegionCLIP [71] introduces an intermediate pretraining step to better align CLIP to box features. BARON [61] improves the alignment between text and image encoders by extracting bag of regions as additional supervision. F-VLM [31] simplifies the training pipeline of open-vocabulary detection and explores the limit of the frozen vision-language model. All of the models above take + +the design of the object detector as granted, and inject language in the last classification layer of the network. We take a different approach and design a detector that adapts predictions to particular categories of interest. + +Open-vocabulary DETR integrates DETR architecture into open-vocabulary detection. OWL [43] introduces a simple ViT architecture using pretrained CLIP and finetune with the DETR objective. OWLv2 uses self-training to further improve the performance [41]. OV-DETR [65] fuses features of a pretrained CLIP model with DETR object queries. Archi-tecturally, OV-DETR is closest to DECOLA. Both OV-DETR and DECOLA condition predictions on the text representation of classes. OV-DETR uses the original DETR queries and expands each with a CLIP feature per class for all classes. This leads to a quadratic number of queries, growing with the original DETR queries and with the classes considered. On the other hand, DECOLA explicitly controls the first-stage predictions (proposals) by formulating the scoring function to respect to the text embedding of each queried class at run-time. We visualize this difference in Figure ?? in the supplementary. The advantage is that we entirely remove inter-class competition and process a manageable amount of queries each focusing on a specific class, and running as fast as the vanilla Deformable DETR. This ability to freely adjust inner workings deviates DECOLA from prior works; it expands detection data through high-quality pseudo-labeling and achieves state-of-the-art results. + +Large-vocabulary object detection shares similar goals with open-vocabulary detection. Both learn from naturally long-tail data over large vocabularies. Vanilla large-vocabulary detectors are often ill-calibrated: The detector's final classification layer favors frequently seen objects over rare ones. This imbalance is usually addressed through a change in loss [7, 55-57, 72], or leveraging additional weakly labeled data for self-training [8, 16, 68, 76]. In large-vocabulary detection, R-CNN-based frameworks [2, 4, 17, 18, 72] dominate despite DETR-style architectures [5, 25, 44, 70] having long surpassed them on standard benchmarks [36]. DETR automatically assigns object queries to output classes, and thus it learns to more heavily focus queries on common classes. We show that language-conditioning helps address this calibration issue. Specifically, it removes inter-class competition in the training objective as queries are no longer shared across categories. As a result, DECOLA equally focuses on as many rare classes as frequent ones whenever they are present in an image. This yields a DETR-style detector that is competitive with the best R-CNN-based large-vocabulary detectors. + +# 3. Preliminaries + +Detection transformers (DETR) [3] build an object detection pipeline as a single feed-forward network. The network transforms object queries, arbitrary feature vectors, into la + +![](images/ffeaad042b3f61a7d65e0720779527f587aaa5d31460ffff12f09d84dc4235c3.jpg) +Figure 2. Overview of DECOLA Phase 1 for conditioned prediction (top) and Phase 2 for open-vocabulary detection (bottom). For language-conditioned detection, each text embedding directly parameterizes the objectness function for each class. In Phase 2, the language-condition reads "an object" instead of particular class names, and predicts multi-class scores over all classes after decoding layers. + +beled bounding boxes through a series of cross-attention layers in a decoder architecture. Vanilla DETR [3] learns object queries as free-form parameters, while modern DETRs architectures [25, 37, 60, 70, 75] adopt a two-stage paradigm similar to RCNNs [48]. This query mechanism controls much of the inner workings of the detector. Queries determine what image regions the detector focuses on, and what object classes are prioritized. + +Query selection. Modern DETR architectures use image-dependent query selection, analogous to R-CNN's proposal generation [48]. An objectness function $s$ scores each grid location $(i,j)$ in the image using a feature $x_{i,j}$ extracted from the transformer encoder. The top- $k$ scored regions proceed to the second stage as object queries $Q$ : + +$$ +s \left(x _ {i, j}\right) = \left\langle x _ {i, j}, w \right\rangle , \quad Q = \operatorname {t o p k} _ {x _ {i, j}} \left(s \left(x _ {i, j}\right)\right). \tag {1} +$$ + +Here, $w$ are the parameters of the objectness predictor. Each selected query produces a series of predictions that are refined over multiple iterations, similar to Cascade R-CNN [2]. The final prediction $\vec{p}$ contains scores over all classes $C$ and an associated box. At a high level, DETR and R-CNNs share the same motivation: first, localize all objects in a scene, then refine their predictions. + +Training objective. During training, DETR assigns each object query to an object or marks it as background. This allows DETR to learn non-overlapping object queries without post-processing such as non-maximum-suppression. The Hungarian matching algorithm finds the optimal assignment between all predictions $P$ and all ground truth $G$ , minimizing the loss function as matching cost: + +$$ +\sigma^ {*} = \arg \min _ {\sigma \in \mathfrak {S}} \ell (P, G | \sigma) \tag {2} +$$ + +where $\mathfrak{S}$ captures all possible assignments from $P$ to $G$ . For each assigned prediction, the loss $\ell$ maximizes its class log-likelihood and fits its bounding box. For unassigned predictions, the loss $\ell$ reduces both the objectness score $s$ and class log-likelihood for all classes. + +# 4. DECOLA + +Our detection transformer conditioned on language, DECOLA, changes the DETR architecture in one remarkable way: Object queries are conditioned on a language embedding. Figure 2 illustrates this change. This simple change has a few important implications: First, it allows the language embedding to control and focus queries to localize on the concepts at hand. Second, it removes any contention between different object classes. Each class present in the image uses the same amount of queries. Third, it generalizes to unseen classes by leveraging semantic knowledge encoded in language embedding throughout the detection pipeline. In the remainder of this section, we highlight the changes in the architecture and training objective for conditioning (DECOLA Phase 1), and self-training on image-level data for open-vocabulary detection (DECOLA Phase 2). + +Language-conditioned query selection. DECOLA conditions queries to a specific object category by modeling the objectness function as a similarity score between a region feature $x_{i,j}$ and a text representation of a category name $t(y)$ using their cosine similarity: + +$$ +s _ {y} \left(x _ {i, j}\right) = \frac {\left\langle x _ {i , j} , t (y) \right\rangle}{\left\| x _ {i , j} \right\| \left\| t (y) \right\|}, Q _ {y} = \operatorname {t o p k} _ {x _ {i, j}} \left(s _ {y} \left(x _ {i, j}\right)\right) \tag {3} +$$ + +The above objective avoids any inter-class calibration issue common in imbalanced data [7, 55, 57]. Queries do not + +compete, as DECOLA independently selects top- $k$ scoring regions $Q_{y}$ for each class $y$ . All queries proceed to the second stage in parallel. A memory-efficient attention mechanism isolates interaction within each class. After a series of decoding layers, each language-conditioned query predicts a single scalar score, corresponding to the likelihood of the class $y$ , and the associated box. The overall architecture mirrors the two-stage deformable DETR [75] with two modifications: a language-conditioned query, and a binary output classifier. + +Memory-efficient modeling. DECOLA uses $n = |Q_y|$ queries per class for $K$ classes. Generally, $n$ is smaller than the total number of queries $|Q|$ of a standard DETR model. However, since we produce $n$ queries per class, the total number of queries in DECOLA is much larger $|Q| \ll nK$ . A naive implementation of the DETR decoder is unable to cope with the $O(n^2 K^2)$ memory requirements of the self-attention layers in the transformer decoder. We thus modify the self-attention formulation to isolate it within each class, reducing the memory cost to $O(n^2 K)$ . The actual implementation uses standard self-attention with a reshaping operation. See Figure 2 (right) for the illustration. DECOLA Phase 1: Train to condition on given concepts. Our goal is to design DECOLA to take a set of class names in an image (or a batch of images) and predict objects of the corresponding classes or backgrounds. For each class $y$ , each conditioned query $q_y \in Q_y$ therefore only predicts a single presence score for class $y$ and the box location. All predictions from the conditioned queries $P_y$ are matched with $G_y$ , the subset of ground truth with class $y$ : + +$$ +\sigma_ {y} ^ {*} = \arg \min _ {\sigma \in \mathfrak {S} _ {y}} \ell \left(P _ {y}, G _ {y} \mid \sigma\right) \tag {4} +$$ + +where $\mathfrak{S}_y$ is the set of possible matches between $P_y$ and $G_y$ , and $\ell$ is the binary cross-entropy loss. Unlike the original DETR objective in Eqn. 2, Eqn. 4 matches within the conditioned class. It avoids inter-class competition during training and simplifies the training objective. Instead, it learns to adapt its predictions to $y$ ; the set of conditioned query $q_y$ considers any objects other than $y$ as background. + +Pseudo-labeling weakly-labeled data. DECOLA produces highly accurate predictions when conditioned on the exact categories of a scene, as shown in Section 5.4. This makes DECOLA a strong pseudo-labeler for weakly-labeled data with either image tags or captions. We expand a large amount of such data with pseudo-bounding boxes of DECOLA Phase 1 and self-train altogether to scale up open-vocabulary object detection. Unlike other forms of weakly supervised learning such as knowledge distillation [19] and online pseudolabeling [61, 73], we simply generate labels for all images offline and jointly train over all pseudo-labeled data using the regular detection losses without any additional complication or slowdown. For each image and class $y$ , DECOLA encodes the class' language feature and predicts a set of detections $P_{y}$ . We simply choose the most confident prediction. + +Figure 4d shows our simple offline pseudo-labeling works better than online pseudo-labeling. + +The advantage of DECOLA comes from adaptability to specified class names on a per-image basis. However, in open-vocabulary detection, the set of test classes is neither known a priori nor available per image. Hence, we convert DECOLA into a general-purpose detector to detect all objects. We condition DECOLA with "an object" as the text input, and inject the class information in the second-stage classifier. Figure 2 highlights this conversion. Since DECOLA is trained to align image features to text embedding in both the first and second stages, this change only introduces inter-class calibration for multi-class object detection. We train DECOLA with pseudo-labeled and human-labeled data as a standard supervised detection training, using the standard matching algorithm of DETR (Section 3). We do not introduce any additional hyper-parameter specifically for the weakly supervised learning [73] or design choices [31, 71], extra loss functions such as alignment loss [61], nor a large teacher model for knowledge distillation [19, 65]. Generating pseudo-labels with DECOLA Phase 1 runs as fast as a regular detector and training Phase 2 is as easy as standard detection training on a supervised dataset. At a high level, DECOLA Phase 1 training objective optimizes for a strong pseudo-labeler instead of a multi-class detector, which differentiates DECOLA from prior works. Leveraging DECOLA Phase 1 to expand weakly-labeled data is the key contribution to scaling up the final open-vocabulary object detection. + +# 5. Experiments + +We evaluate the effectiveness of DECOLA in two aspects: (1) pseudo-labeling quality of DECOLA Phase 1 (Section 5.4) and (2) benchmark evaluation (Section 5.3). We consider three primary benchmarks to evaluate our final model (DECOLA Phase 2): open-vocabulary LVIS [19], standard LVIS [20], and direct zero-shot evaluation to LVIS, COCO [36], Object365 [52], and OpenImages [32]. + +# 5.1. Experimental Setup + +Datasets and benchmarks. We mainly evaluate our method on the LVIS dataset [20], a large-vocabulary instance segmentation and object detection dataset with 1203 naturally distributed object categories. LVIS splits categories into frequent, common, and rare. For open-vocabulary LVIS, we combine frequent and common categories into LVIS-base and consider the rare categories as novel concepts used for testing only [19]. For standard LVIS, we train and evaluate all classes. Direct zero-shot transfer evaluates models trained on different detection data (e.g., Object365) and other weakly-labeled data without any prior knowledge about the target dataset such as the set of classes or object frequency. In this benchmark, we test DECOLA's generalization to dif + +
methoddataAPbox novelAPcAPfmAPbox
ResNet-50 (1K)
OV-DETR† [65]LVIS-base, IN-21K18.024.831.826.4
baselineLVIS-base10.230.938.030.1
baseline + self-trainLVIS-base, IN-21K19.231.737.131.7
DECOLA Phase 2LVIS-base, IN-21K23.8 (+4.6)34.4 (+2.7)38.3 (+1.2)34.1 (+2.4)
ResNet-50
baselineLVIS-base9.433.840.432.2
baseline + self-trainLVIS-base, IN-21K23.236.541.636.2
DECOLA Phase 2LVIS-base, IN-21K27.6 (+4.4)38.3 (+1.8)42.9 (+1.3)38.3 (+2.1)
Swin-B
baselineLVIS-base16.243.849.141.1
baseline + self-trainLVIS-base, IN-21K30.843.645.942.3
DECOLA Phase 2LVIS-base, IN-21K35.7 (+4.9)47.5 (+3.9)49.7 (+3.8)46.3 (+4.0)
Swin-L
DITO* [28]O365, LVIS-base, DataComp-1B45.8--44.2
OWLv2‡ [41]O365, LVIS-base, VG, WebLI45.9--50.4
baselineO365, LVIS-base21.953.357.749.6
baseline + self-trainO365, LVIS-base, IN-21K36.553.556.551.8
DECOLA Phase 2O365, LVIS-base, IN-21K46.9 (+10.4)56.0 (+2.5)58.0 (+1.5)55.2 (+3.6)
+ +Table 1. Open-vocabulary LVIS using DETR architectures. $\dagger$ : we further finetune the official OV-DETR model trained on LVIS-base by self-training on ImageNet-21K data similar to "baseline + self-train" and DECOLA. $\ddagger$ : uses CLIP L/14, which is comparable to Swin-L backbone. $\star$ : is based on Mask R-CNN with ViT L/16. We report the improvement between "baseline + self-train" to DECOLA in green. Last 4 rows compare DECOLA to Swin-L or equivalent scale of models that use additional detection data (e.g., Objects365, VG [30]) and billion-scale weakly-labeled data (DataComp-1B [15], WebLI [6]) for training. + +
methodframework\(AP^{box}\)novel\(mAP^{box}\)novel\(AP^{mask}\)novel\(mAP^{mask}\)
ViLD [19]Mask R-CNN16.727.816.625.5
RegionCLIP [71]Mask R-CNN--17.122.5
DetPro [12]Mask R-CNN20.828.419.825.9
PromptDet [14]Mask R-CNN21.425.3--
F-VLM [31]Mask R-CNN--18.624.2
BARON [61]Mask R-CNN23.229.522.627.6
OADP [58]Mask R-CNN21.928.721.726.6
EdaDet [54]Mask R-CNN--23.727.5
VLDet [35]CenterNet2--21.730.1
\(\mathsf{CORA}^{+}\) [62]CenterNet228.1---
Rasheed et al. [47]CenterNet2--25.232.9
Detic-base [73]CenterNet217.633.816.430.2
Detic [73]CenterNet226.736.324.632.4
DECOLA labelsCenterNet229.537.727.033.7
+ +(a) Comparison with ResNet-50 backbone. + +
methodbackbone\(AP_{\text{novel}}^{box}\)\(mAP_{\text{novel}}^{box}\)\(AP_{\text{novel}}^{mask}\)\(mAP_{\text{novel}}^{mask}\)
RegionCLIP [71]R50×4--22.032.3
CondHead [59]R50×424.133.724.432.0
ViLD [19]EN-B7--26.329.3
OWL-ViT [43]ViT-L/1425.634.7--
F-VLM [31]R50×64--32.834.9
VLDet [35]Swin-B--26.338.1
3Ways [1]NFNet-F630.144.6--
RO-VIT [29]ViT-L/1632.134.0--
CFM-ViT [27]ViT-L/1635.638.533.936.6
DITO [28]ViT-B/1634.936.932.534.0
CoDet [40]Swin-B--29.439.2
Detic-base [73]Swin-B24.643.021.938.4
Detic [73]Swin-B36.645.733.840.7
DECOLA labelsSwin-B38.446.735.342.0
+ +(b) System-level comparison. + +Table 2. Open-vocabulary LVIS using Mask R-CNN and CenterNet2 detectors. Methods in both tables use LVIS-base as the only human-labeled data for fair comparison. For system-level comparison (right), we include methods with non R-CNN architectures such as OWL-ViT [43]. The results show the impact of high-quality pseudo-labels generated by DECOLA Phase 1. + +ferent domains. We evaluate DECOLA on LVIS, COCO [36], Object365 [52], and OpenImages [32] in a fully zero-shot manner. All our models use the ImageNet-21K [50] dataset as weakly labeled data, which contains 14M of object-centric images annotated with a single class. + +Evaluation metrics. We evaluate DECOLA on $\mathrm{AP}_{\mathrm{novel/rare}}$ , $\mathrm{AP_c}$ , $\mathrm{AP_f}$ , and mAP following the LVIS evaluation metric [20]. We highlight the results in all three groups since we believe open-vocabulary detectors should not compensate for the performance of common/frequent classes for novel/rare classes. We evaluate both $\mathrm{AP^{box}}$ and $\mathrm{AP^{mask}}$ for object detection and instance segmentation. For zero-shot transfer benchmark with COCO and Object365, we + +use AP, $\mathrm{AP}_{50}$ , and $\mathrm{AP}_{75}$ following prior work [19, 73]. For OpenImages, we report $\mathrm{AP}_{50}^{\mathrm{flat}}$ on the expanded label space [73, 74]. For zero-shot transfer to LVIS, we consider LVIS minival [26] and standard LVIS v1.0 validation set and report $\mathrm{AP}^{\mathrm{fixed}}$ [9] following the prior works [26, 34, 38]. In addition, we pursue a more direct measurement of the generated pseudo-labeling quality. Hence, we define conditioned mAP/AR (c-mAP/AR) and compare it to baseline open-vocabulary detectors. c-mAP measures the detection performance in mAP when the detector is provided the set of ground truth classes in each image. For example in Figure 1, both detectors use "cat", "mentos" and "cola" as given. This extra information is used to select scores to rank + +
methoddataAPboxraremAPbox
ResNet-50
baselineLVIS26.335.6
baseline + self-trainLVIS, IN-21K30.036.6
DECOLA Phase 2LVIS, IN-21K35.9 (+5.9)39.4 (+2.8)
Swin-B
baselineLVIS38.344.5
baseline + self-trainLVIS, IN-21K42.045.2
DECOLA Phase 2LVIS, IN-21K47.4 (+5.4)48.3 (+3.1)
Swin-L
baselineO365, LVIS49.354.4
baseline + self-trainO365, LVIS, IN-21K48.753.4
DECOLA Phase 2O365, LVIS, IN-21K54.9 (+6.2)56.4 (+3.0)
+ +(a) Standard LVIS with DETR. + +
methodAPboxraremAPboxAPmaskraremAPmask
ResNet-50
Detic-base [73]28.235.325.631.4
Detic [73]31.436.829.733.2
DECOLA labels35.6 (+4.2)38.6 (+1.8)32.1 (+2.4)34.4 (+1.2)
Swin-B
Detic-base [73]39.945.435.940.7
Detic [73]45.846.941.741.7
DECOLA labels47.6 (+1.8)48.5 (+1.6)43.7 (+2.0)43.6 (+1.9)
+ +(b) Standard LVIS with CenterNet2. + +Table 3. Standard LVIS benchmark. DECOLA shows consistent improvement over different model scales and architectures. + +the final predictions (belines), or directly condition the detector (DECOLA). We analyze the model's behavior and label quality in Section 5.4 and Section ?? in supplementary. + +# 5.2. Models + +DECOLA is based on two-stage Deformable DETR [75]. As described in Section 4, the first-stage objectness function for query selection is replaced by a similarity score between the image feature and the CLIP text embedding of each class name. We train the detector with the improved DETR training recipe [44, 70]: look-forward-twice, larger MLP hidden dimension, no dropout, etc. We consider four backbones: a ResNet-50 [21], Swin-B and L for all LVIS benchmarks, and Swin-T and L for the direct zero-shot transfer. Unless otherwise mentioned, all backbones are pretrained on the ImageNet-21K dataset [49]. Next, we describe our key baseline models to directly compare to DECOLA. + +Baseline. We design a baseline open-vocabulary detector to closely compare to DECOLA. Inspired by Detic [73], baseline replaces classification layers with the class embedding of the pretrained CLIP text encoder and is trained using Federated Loss [72]. DECOLA Phase 1 and baseline are trained using human-labeled data (e.g., LVIS-base). All other settings (training dataset, number of training iterations, etc.) are kept the same between DECOLA and baseline. + +Baseline + self-train. Similar to DECOLA Phase 2, we self-train baseline on weakly-labeled data. For the self-training algorithm, we use online self-training with max-size loss from Detic [73] as baseline comparison (baseline + self-train) to DECOLA Phase 2. We tested max-size and max + +score losses from Detic [73] (online pseudo-labeling) as well as offline pseudo-labeling similar to DECOLA, and max-size loss consistently performed the best. + +DECOLA labels. We train a two-stage detector for broader comparison: CenterNet2 [72]. Specifically, we use Detric's baseline model ("Detric-base"), a CenterNet2 trained on LVIS-base with CLIP embedding, and finetune on pseudolabeled ImageNet-21K data using DECOLA Phase 1 of the same backbone size. We denote this as "DECOLA labels". + +Efficient modeling. For DECOLA Phase 1, we use $n = 300$ queries per class. One memory and time bottleneck during DECOLA training is the first-stage loss computation. The original Deformable DETR computes Hungarian matching with all pixels to all objects in a class-agnostic manner, which is $\sum_{l \in L} H_l \cdot W_l$ predictions. To reduce the memory and time cost, we only consider the top $K = 10,000$ confident pixels for each class $y$ during the first-stage matching and loss computation. Together with memory-efficient self-attention (Sec. 4), the training time and memory cost of DECOLA increases by less than $20\%$ over the baselines (See Table 6). + +Training details. Following Detic [73], we train DECOLA Phase 1 and baseline $4 \times$ on LVIS-base, and further fine-tune for another $4 \times$ on ImageNet-21K data with pseudolabeling. For training CenterNet2 with DECOLA labels, we combine pseudo-labels from different image resolutions for $h \in \{240, 280, 320, 360, 400\}$ , where $h$ is the shorter side of the image. Note that this mimics the random image resizing data augmentation during standard detection training. We use Detector2 [63] based on PyTorch [45] in all of the experiments. More details are in Section ?? of supplementary. + +# 5.3. Main Results + +Open-vocabulary LVIS. Table 1 compares DECOLA to baseline as well as state-of-the-art DETR-based open-vocabulary detectors; OV-DETR [65], OWLv2 [41], and baseline + self-train. For a fair comparison, we further fine-tune the official OV-DETR model checkpoint on ImageNet21K for $4 \times$ schedule same as DECOLA Phase 2 and baseline + self-train. For all backbone scales, we show consistent improvement over other methods. Notably, baseline + self-train exhibits degradation in frequent classes as a trade-off with improved novel classes, which is commonly observed behavior in other open-vocabulary detection methods, too. DECOLA improves all categories consistently, which highlights the quality of our pseudo-labels. In the last rows with the Swin-L backbone, we report the result of two concurrent works, DITO [28] (Mask R-CNN-based) and OWLv2, to compare to the method that uses additional detection data (Object365) and billion-scale web data (DataComp1B [15], WebLI [6]). DECOLA demonstrates large improvement over the state-of-the-arts despite using orders of magnitude smaller training data and compute resources. To further examine DECOLA's scalability, we test the pseudo + +
methoddataLVIS minivalLVIS v1.0 val
APboxrareAPcAPfmAPboxAPboxrareAPcAPfmAPbox
Swin-T
MDETR* [26]LVIS, GoldG, RefC20.924.924.324.27.422.725.022.5
GLIP [34]O365, GoldG, Cap4M20.821.431.026.010.112.525.517.2
GroundingDINO [38]O365, GoldG, Cap4M18.123.332.727.4----
GLIPv2 [69]O365, GoldG, Cap4M---29.0----
MQ-GroundingDINO† [64]O365, GoldG, Cap4M, LVIS-5VQ21.726.235.230.212.917.431.422.1
MQ-GLIP† [64]O365, GoldG, Cap4M, LVIS-5VQ21.027.534.630.415.418.430.422.6
DECOLA Phase 2O365, IN-21K‡32.832.031.832.027.224.928.026.6
Δ(+12.0)(+8.7)-(+3.0)(+17.1)(+12.4)(+2.5)(+9.4)
Swin-L
GLIP [34]FourODs, GoldG, Cap24M28.234.341.537.317.123.335.426.9
GroundingDINO [38]O365, OI, GoldG, Cap4M, COCO, RefC22.230.738.833.9----
MQ-GLIP† [64]FourODs, GoldG, Cap24M, LVIS-5VQ34.541.246.943.426.932.041.334.7
OWLv2 [41]O365, VG, WebLI39.0--38.134.9--33.5
DECOLA Phase 2O365, OI, IN-21K‡41.538.034.936.832.929.130.330.2
Δ(+2.5)(+3.7)---(+5.8)--
+ +Table 4. Direct zero-shot transfer to LVIS. †: methods use 5 per-class vision queries of LVIS dataset (denoted as “LVIS-5VQ”), which use images and annotations to extract instance-level features. ☆: MDETR uses ResNet-101 backbone and trained fully-supervised on LVIS. Results that use LVIS data are in gray. ‡: Full ImageNet-21K. No LVIS information is used in DECOLA. + +
methodCOCOO365OI
AP\( AP_{50} \)\( AP_{75} \)AP\( AP_{50} \)\( AP_{75} \)\( AP_{50}^{flat} \)
R-CNNs
ViLD [19]36.655.639.811.818.212.6-
F-VLM [31]32.553.134.611.919.212.6-
DetPro [12]34.953.837.412.118.812.9-
BARON [61]36.356.139.313.621.014.5-
Detic [73]39.156.342.214.220.715.242.9
DETRs
OV-DETR [65]38.158.441.1----
Detic [73]39.856.643.314.521.415.541.6
DECOLA Phase 240.357.043.715.022.016.043.3
+ +Table 5. Cross-dataset generalization benchmark on COCO, Object365, and OpenImages. All models use a ResNet-50 backbone and train on LVIS-base and weakly-labeled data. + +
methodtraintest
timemem.timemem.
baseline44 h8.9 G0.07 s/img2.5 G
+ self-train45 h10.2 G0.07 s/img2.5 G
OV-DETR73 h22.0 G6.4 s/img3.4 G
DECOLA Phase 149 h12.6 G--
DECOLA Phase 245 h8.9 G0.07 s/img2.8 G
+ +Table 6. Efficiency. Training time is measured with 8 DGX V100 on ResNet-50, 2 images per-GPU. float16 is used in both training and testing. OV-DETR uses the original optimized inference. + +labeling capability of DECOLA on R-CNN-based detectors with DECOLA labels (Detic-base finetuned with our pseudolabels). In Table 2, we compare DECOLA labels to a broad range of literature based on CenterNet2 [72] and Mask R-CNN [22]. Table 2a compares methods with ResNet-50 backbone and Table 2b compares larger scale backbones for system-level comparison. In both tables, DECOLA clearly improve upon the state-of-the-art by large margins without additional complication in training and bells-and-whistles. + +Standard LVIS. Tables 3 evaluate DECOLA and baseline on the standard LVIS benchmark, where all object categories are used to fully supervise the detectors. Similar to the + +open-vocabulary LVIS, we compare DETR architectures in Table 3a and R-CNN architectures in Table 3b. Table 3a shows that DECOLA remarkably improves baseline by 9.5, 9.1, and 5.6 points on $\mathrm{AP}_{\mathrm{rare}}^{\mathrm{box}}$ , outperforming baseline + self-train by 5.9, 5.4, and 6.2 $\mathrm{AP}_{\mathrm{rare}}^{\mathrm{box}}$ for ResNet-50, Swin-B, and Swin-L backbones, respectively. Similarly in Table 3b, DECOLA labels further improves the baseline of Detic by 7.4 and 7.7 $\mathrm{AP}_{\mathrm{rare}}^{\mathrm{box}}$ , and outperforms Detic [73] by 4.2 and 1.8 $\mathrm{AP}_{\mathrm{rare}}^{\mathrm{box}}$ with ResNet-50 and Swin-B backbones, respectively. Direct zero-shot evaluation. For direct zero-shot evaluation, we train DECOLA with Swin-T [39] and use Object365 data for Phase 1, and ImageNet-21K for Phase 2 (full dataset and classes). We compare to MDETR [26], GLIP [34], GroundingDINO [38], and MQ-Det [64] finetuned from GLIP and GroundingDINO. Table 4 shows the results. DECOLA outperforms the previous state-of-the-arts, by $12.0 / 17.1\mathrm{AP}_{\mathrm{rare}}$ and $3.0 / 9.4\mathrm{mAP}$ on LVIS minival and LVIS v1.0 val, respectively. Note that all other methods use much richer detection labels from GoldG data [26], a collection of grounding data (box and text expression pairs) curated by MDETR. Furthermore, other benchmark methods show highly imbalanced $\mathrm{AP}_{\mathrm{rare}}$ and $\mathrm{AP_f}$ in both LVIS minival and LVIS v1.0 val (10-20 points gap). We hypothesize that the large collection of training data coincides with LVIS vocabulary, as all data follows a natural distribution of common objects. Also, DECOLA enjoys significantly faster run-time compared to all other models that undergo BERT encoding for grounding, which requires more than 50 forward passes per image in order to predict all LVIS categories. Similarly, our Swin-L model outperforms GroundingDINO and GLIP by 19.3 and $13.3\mathrm{AP}_{\mathrm{rare}}$ , respectively, despite much smaller training data compared to FourODs [34] and match to OWLv2 with 10-B private WebLI data [6]. Table 5 further examines the generality of DECOLA on different domains. DECOLA Phase 2 with ResNet-50 outperforms all other competitive baselines. + +![](images/14d1dd4ea0cc274c81c008bde984323767dcd651e32a292f669c9fc0c03b26f8.jpg) +Figure 3. Examples of prediction on unseen categories. Images from ImageNet-21K dataset. Boxes are the most confident prediction from DECOLA and baseline. DECOLA conditions on the ground truth label, and baseline selects the max-score box of the ground truth class. Images are all from unseen categories, which neither model was trained on. Green: DECOLA Phase 1 trained on LVIS-base. Red: baseline trained on LVIS-base. Both models use a Deformable DETR detector with a ResNet-50 backbone. More in Fig. ?? and ?? of supplementary. + +![](images/2bd77d56d18c866bb0b2263450d1459c0dc6abbee2c76f4199e508cf6f9f7261.jpg) +(a) c-AP for ResNet-50 (IN-1K) + +![](images/1ed8e239b3cf9a89151578fbf69b0216a65323826667a60ced142f0d5c08d527.jpg) +(b) c-AP for Swin-B +Figure 4. Analyzing DECOLA. All plots show the conditioned AP/AR for unseen classes. We compare DECOLA Phase 1 and baselines. We highlight more detailed analyses about c-AP and c-AR in Section ??, Table ?? and ?? of the supplementary materials. + +![](images/8b6e4c6ff3f4bbd78d357c45926e0fa2b19cc757fa2a03f4e19e31ca2b78f61f.jpg) +(c) c-AR for 1st and 2nd stage boxes + +![](images/96c45373ba0c1f741ab0eb7382b39c0c6425e02cb0388b2f7e88658524e408c3.jpg) +(d) online vs offline: c-AP@20 + +# 5.4. Analyses + +In this section, we analyze the model's behavior with conditioned mAP/AR (c-AP/AR) (defined in Sec. 5.1). + +Pseudo-labeling quality. Figure 4a and 4b compare DECOLA Phase 1, OV-DETR, and baseline on c-AP for unseen classes. Compared to baseline and OV-DETR, DECOLA Phase 1 generates much higher quality pseudo-labels, especially in low-shot regimes. See examples in Figure 3. + +Impact of conditioning. In Figure 4c, we compare c-AR of unseen classes. This measures the detector's ability to localize objects of interest when pseudo-labeling. We observe significant improvement in c-AR on both first-stage (proposals) and second-stage (predictions) due to our conditioning mechanism. This result demonstrates the key difference between DECOLA and other open-vocabulary detectors. + +Pseudo-labeling algorithms. Figure 4d shows the c-AP of baseline + self-train and DECOLA Phase 1 for unseen classes with 20 predictions per-image. Each red bar indicates the percent of training iteration during self-training. Online + +self-labeling suffers a sharp drop during the early iterations, and c-AP after full iterations still underperforms compared to DECOLA Phase 1. DECOLA's simple approach of offline self-training is more stable and effective. + +# 6. Conclusion + +In this paper, we explore a new open-vocabulary detection framework, DECOLA. It adjusts its inner workings to the concepts that the user asks to reason over by conditioning on a language embedding. Our detector generates high-quality pseudo-labels on weakly labeled data through the conditioning mechanism. We finetune it with the pseudolabels to build the state-of-the-art open-vocabulary detector. + +Acknowledgement. We thank Xingyi Zhou, Yue Zhao, Jeffrey Ouyang-Zhang, and Nayeon Lee for valuable feedback. This material is in part based upon work supported by the National Science Foundation under Grant No. IIS-1845485 and IIS-2006820. + +# References + +[1] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Hénaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection. arXiv, 2023. 5 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 1, 2, 3 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3 +[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, 2019. 2 +[5] Qiang Chen, Xiaokang Chen, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. arXiv preprint arXiv:2207.13085, 2022. 1, 2 +[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 5, 6, 7 +[7] Jang Hyun Cho and Philipp Krahenbuhl. Long-tail detection with effective class-margins. In ECCV, 2022. 2, 3 +[8] Jang Hyun Cho, Philipp Krahenbuhl, and Vignesh Ramanathan. Partdistillation: Learning parts from instance segmentation. In CVPR, 2023. 2 +[9] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv, 2021. 5 +[10] Mostafa et al. Dehghani. Scaling vision transformers to 22 billion parameters. ICML, 2023. 2 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 2 +[12] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. 2022. 5, 7 +[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2 +[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 5 +[15] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M Pratt, Vivek + +Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Neurips (Datasets and Benchmarks Track), 2023. 5, 6 +[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2 +[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2 +[18] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. 1, 2 +[19] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 2, 4, 5, 7 +[20] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019, 1, 2, 4, 5 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6 +[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7 +[23] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In Neurips, 2015. 2 +[24] Gabriel et al Ilharco. Openclip, 2021. 2 +[25] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. 2023. 1, 2, 3 +[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 5, 7 +[27] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In ICCV, 2023. 5 +[28] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Detection-oriented image-text pretraining for open-vocabulary detection. arXiv, 2023. 5, 6 +[29] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, 2023. 5 +[30] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 5 +[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2, 4, 5, 7 +[32] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, + +Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 4, 5 +[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2 +[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 5, 7 +[35] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 5 +[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5 +[37] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 1, 3 +[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5, 7 +[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 7 +[40] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In Nuerips, 2023. 5 +[41] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 2, 5, 6, 7 +[42] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 1 +[43] Matthias Minderer et al. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 2, 5 +[44] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 2, 6 +[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Neurips. 2019. 6 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2 + +[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammed Uzair Khattak, Salman Khan, and Fahad Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Neurips, 2022. 5 +[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Neurips, 2015. 3 +[49] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik. Imagenet-21k pretraining for the masses. In Neurips, 2021. 6 +[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 2, 5 +[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Neurips, 2022. 2 +[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019. 4, 5 +[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2 +[54] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 5 +[55] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3 +[56] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In CVPR, 2021. +[57] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In CVPR, 2021. 2, 3 +[58] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, 2023. 5 +[59] Tao Wang. Learning to detect and segment for open vocabulary object detection. In CVPR, 2023. 5 +[60] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 3 +[61] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 1, 2, 4, 5, 7 +[62] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 5 +[63] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. *Detectron2*, 2019. 6 + +[64] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. In Neurips, 2023. 7 +[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, 2022. 1, 2, 4, 5, 6, 7 +[66] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 1, 2 +[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022. 2 +[68] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. MosaicOS: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2 +[69] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Neurips, 2022. 7 +[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. 2023. 1, 2, 3, 6 +[71] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2, 4, 5 +[72] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 1, 2, 6, 7 +[73] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 4, 5, 6, 7 +[74] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 5 +[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable $\{\mathrm{detr}\}$ : Deformable transformers for end-to-end object detection. In ICLR, 2021. 1, 3, 4, 6 +[76] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. Neurips, 2020. 2 \ No newline at end of file diff --git a/2024/Language-conditioned Detection Transformer/images.zip b/2024/Language-conditioned Detection Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..787cb872143b354c97224c277531446c58aafd88 --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0c5d00211103a3735ee2bd74132495150def797da1056eb2beec24e166aff6f +size 790107 diff --git a/2024/Language-conditioned Detection Transformer/layout.json b/2024/Language-conditioned Detection Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..61fb89d0310aab2a4eb930b03d5c5fc9b71fdde7 --- /dev/null +++ b/2024/Language-conditioned Detection Transformer/layout.json @@ -0,0 +1,8464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 155, + 103, + 440, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 103, + 440, + 121 + ], + "spans": [ + { + "bbox": [ + 155, + 103, + 440, + 121 + ], + "type": "text", + "content": "Language-conditioned Detection Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 144, + 259, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 144, + 259, + 170 + ], + "spans": [ + { + "bbox": [ + 182, + 144, + 259, + 170 + ], + "type": "text", + "content": "Jang Hyun Cho UT Austin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 158, + 174, + 284, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 174, + 284, + 185 + ], + "spans": [ + { + "bbox": [ + 158, + 174, + 284, + 185 + ], + "type": "text", + "content": "janghyuncho7@utexas.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 331, + 144, + 427, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 144, + 427, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 144, + 427, + 171 + ], + "type": "text", + "content": "Philipp Krahenbuhl UT Austin" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 174, + 433, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 174, + 433, + 185 + ], + "spans": [ + { + "bbox": [ + 323, + 174, + 433, + 185 + ], + "type": "text", + "content": "philkr@cs.utexas.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 501 + ], + "type": "text", + "content": "We present a new open-vocabulary detection framework. Our framework uses both image-level labels and detailed detection annotations when available. Our framework proceeds in three steps. We first train a language-conditioned object detector on fully-supervised detection data. This detector gets to see the presence or absence of ground truth classes during training, and conditions prediction on the set of present classes. We use this detector to pseudolabel images with image-level labels. Our detector provides much more accurate pseudo-labels than prior approaches with its conditioning mechanism. Finally, we train an unconditioned open-vocabulary detector on the pseudo-annotated images. The resulting detector, named DECOLA, shows strong zero-shot performance in open-vocabulary LVIS benchmark as well as direct zero-shot transfer benchmarks on LVIS, COCO, Object365, and OpenImages. DECOLA outperforms the prior arts by 17.1 AP_rare and 9.4 mAP on zero-shot LVIS benchmark. DECOLA achieves state-of-the-art results in various model sizes, architectures, and datasets by only training on open-sourced data and academic-scale computing. Code is available at https://github.com/janghyuncho/DECOLA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": "Object detection has seen immense progress over the past decade. Classical object detectors reason over datasets of fixed predefined classes. This simplifies the design, training, and evaluation of new methods, and allows for rapid prototyping [2-5, 17, 18, 25, 37, 42, 60, 70, 72, 75]. However, it complicates deployment to downstream applications too. A classical detector requires a new dataset to further finetune for every new concept it encounters. Collecting sufficient data for every new concept is not scalable [20]. Open-vocabulary detection offers an alternative [19, 43, 61, 65, 66, 73]. Open-vocabulary detectors reason about any arbitrary concept with free-form text, using the generalization ability of vision-language models. Yet, common open-vocabulary detectors reuse classical detectors and either replace the last classification layer" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 213, + 545, + 450 + ], + "blocks": [ + { + "bbox": [ + 313, + 213, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 545, + 450 + ], + "type": "image", + "image_path": "eab37f5762d533fbfcf4ee6003dca2f7f131377e0e31f25de59d1d33b5eec667.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 456, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 522 + ], + "type": "text", + "content": "Figure 1. An illustration of how standard open-vocabulary detectors and DECOLA generate pseudo-labels using image-level data. Standard detectors use image-level information later in the pipeline after initial box proposals, which may result in low coverage of unseen classes (e.g., \"mentos\" and \"cola\"). DECOLA adjusts the prediction to the information and ensures sufficient coverage." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 530, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 530, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 530, + 547, + 567 + ], + "type": "text", + "content": "with [19, 43, 66, 73], or fuse box feature with [31, 65] text representation from pretrained vision-language model. The inner workings of the detector remain unchanged." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a transformer-based object detector that adjusts its inner workings to any arbitrary set of concepts represented in language. The detector considers only the queried set of concepts as foreground and disregards any other objects as background. It learns to adapt detection to the language embedding of queried concepts at run-time. Specifically, the detector conditions proposal generation with respect to the text embedding of each queried concept and refines the conditioned proposals into predictions. Our detection transformer conditioned on language (DECOLA) offers a powerful alternative to classical architectures in open-vocabulary detection. Adapting the detector to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16593" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "language leads to stronger generalization to unseen concepts, and largely enhances self-training on weakly-labeled data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 288, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 288, + 275 + ], + "type": "text", + "content": "DECOLA's ability to readily adapt to any queried concepts makes it particularly suitable for pseudo-labeling weakly-labeled data. Internet data, specifically images with paired text, is highly abundant and semantically rich [50, 51, 53]. The best vision models today build on this massive amount of weakly labeled data [10, 13, 24, 33, 34, 46, 67]. DECOLA leverages the same data to produce high-quality object detection labels from image-level annotations alone. DECOLA takes the image-level tags or text descriptions from the weakly labeled data and generates conditioned predictions as pseudo-labels. It efficiently processes multiple texts in parallel and only adds minimal computational overhead compared to the standard pseudo-labeling process. We finetune DECOLA on this rich detection dataset of pseudo-annotations and achieve the state-of-the-art open-vocabulary detector." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": "We evaluate our detector on popular open-vocabulary detection benchmarks on the LVIS dataset [19, 20, 66]. The final model improves the state-of-the-art methods by 4.4 and " + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "inline_equation", + "content": "4.9\\mathrm{AP}_{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": " on open-vocabulary LVIS [19] benchmark, and 5.9 and " + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "inline_equation", + "content": "5.4\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": " on standard LVIS [20] benchmark, with ResNet-50 [21] and Swin-B [39] backbones, respectively. Our largest model with Swin-L achieves " + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "inline_equation", + "content": "10.4\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "inline_equation", + "content": "3.6\\mathrm{mAP}" + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": " improvement. Furthermore, DECOLA largely outperforms the state-of-the-art for direct zero-shot transfer benchmark on LVIS, by 12.0 and " + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "inline_equation", + "content": "17.1\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 46, + 276, + 288, + 479 + ], + "type": "text", + "content": " on LVIS minimal and LVIS v1.0, respectively. DECOLA consistently improves frequent, common, and rare classes altogether for different backbones and detection frameworks. Much of this improvement is driven by stronger pseudo-labeling capabilities. All our models are trained using open-sourced datasets with academic-scale computing. We open-source our code, pseudo-annotations, and checkpoints of all the model scales." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 134, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Open-vocabulary detection aims to detect objects of categories beyond the vocabulary of the training classes. A common solution is to inject language embeddings of class names in the last classification layer. OVR-CNN [66] pretrains a detector on image-caption data using BERT model [11] as language embedding. ViLD [19] trains a detector with CLIP text encoder [46] as language embedding with additional knowledge distillation [23] between predicted box features and the image encoder of CLIP. Detic [73] improves the above approaches through weakly-supervised learning on image-level annotations. RegionCLIP [71] introduces an intermediate pretraining step to better align CLIP to box features. BARON [61] improves the alignment between text and image encoders by extracting bag of regions as additional supervision. F-VLM [31] simplifies the training pipeline of open-vocabulary detection and explores the limit of the frozen vision-language model. All of the models above take" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "the design of the object detector as granted, and inject language in the last classification layer of the network. We take a different approach and design a detector that adapts predictions to particular categories of interest." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 547, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 547, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 547, + 393 + ], + "type": "text", + "content": "Open-vocabulary DETR integrates DETR architecture into open-vocabulary detection. OWL [43] introduces a simple ViT architecture using pretrained CLIP and finetune with the DETR objective. OWLv2 uses self-training to further improve the performance [41]. OV-DETR [65] fuses features of a pretrained CLIP model with DETR object queries. Archi-tecturally, OV-DETR is closest to DECOLA. Both OV-DETR and DECOLA condition predictions on the text representation of classes. OV-DETR uses the original DETR queries and expands each with a CLIP feature per class for all classes. This leads to a quadratic number of queries, growing with the original DETR queries and with the classes considered. On the other hand, DECOLA explicitly controls the first-stage predictions (proposals) by formulating the scoring function to respect to the text embedding of each queried class at run-time. We visualize this difference in Figure ?? in the supplementary. The advantage is that we entirely remove inter-class competition and process a manageable amount of queries each focusing on a specific class, and running as fast as the vanilla Deformable DETR. This ability to freely adjust inner workings deviates DECOLA from prior works; it expands detection data through high-quality pseudo-labeling and achieves state-of-the-art results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 548, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 548, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 548, + 647 + ], + "type": "text", + "content": "Large-vocabulary object detection shares similar goals with open-vocabulary detection. Both learn from naturally long-tail data over large vocabularies. Vanilla large-vocabulary detectors are often ill-calibrated: The detector's final classification layer favors frequently seen objects over rare ones. This imbalance is usually addressed through a change in loss [7, 55-57, 72], or leveraging additional weakly labeled data for self-training [8, 16, 68, 76]. In large-vocabulary detection, R-CNN-based frameworks [2, 4, 17, 18, 72] dominate despite DETR-style architectures [5, 25, 44, 70] having long surpassed them on standard benchmarks [36]. DETR automatically assigns object queries to output classes, and thus it learns to more heavily focus queries on common classes. We show that language-conditioning helps address this calibration issue. Specifically, it removes inter-class competition in the training objective as queries are no longer shared across categories. As a result, DECOLA equally focuses on as many rare classes as frequent ones whenever they are present in an image. This yields a DETR-style detector that is competitive with the best R-CNN-based large-vocabulary detectors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 391, + 670 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Detection transformers (DETR) [3] build an object detection pipeline as a single feed-forward network. The network transforms object queries, arbitrary feature vectors, into la" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16594" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 494, + 284 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 494, + 284 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 494, + 284 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 494, + 284 + ], + "type": "image", + "image_path": "ffeaad042b3f61a7d65e0720779527f587aaa5d31460ffff12f09d84dc4235c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 547, + 319 + ], + "type": "text", + "content": "Figure 2. Overview of DECOLA Phase 1 for conditioned prediction (top) and Phase 2 for open-vocabulary detection (bottom). For language-conditioned detection, each text embedding directly parameterizes the objectness function for each class. In Phase 2, the language-condition reads \"an object\" instead of particular class names, and predicts multi-class scores over all classes after decoding layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 288, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 421 + ], + "type": "text", + "content": "beled bounding boxes through a series of cross-attention layers in a decoder architecture. Vanilla DETR [3] learns object queries as free-form parameters, while modern DETRs architectures [25, 37, 60, 70, 75] adopt a two-stage paradigm similar to RCNNs [48]. This query mechanism controls much of the inner workings of the detector. Queries determine what image regions the detector focuses on, and what object classes are prioritized." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": "Query selection. Modern DETR architectures use image-dependent query selection, analogous to R-CNN's proposal generation [48]. An objectness function " + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": " scores each grid location " + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": " in the image using a feature " + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "inline_equation", + "content": "x_{i,j}" + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": " extracted from the transformer encoder. The top- " + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": " scored regions proceed to the second stage as object queries " + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 422, + 288, + 495 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 502, + 287, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 502, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 71, + 502, + 287, + 517 + ], + "type": "interline_equation", + "content": "s \\left(x _ {i, j}\\right) = \\left\\langle x _ {i, j}, w \\right\\rangle , \\quad Q = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s \\left(x _ {i, j}\\right)\\right). \\tag {1}", + "image_path": "f7225c450cd004c071962b4cf28c4eea415c825dd3f393773396effacb94cb8d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "text", + "content": " are the parameters of the objectness predictor. Each selected query produces a series of predictions that are refined over multiple iterations, similar to Cascade R-CNN [2]. The final prediction " + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "inline_equation", + "content": "\\vec{p}" + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "text", + "content": " contains scores over all classes " + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 522, + 288, + 605 + ], + "type": "text", + "content": " and an associated box. At a high level, DETR and R-CNNs share the same motivation: first, localize all objects in a scene, then refine their predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "text", + "content": "Training objective. During training, DETR assigns each object query to an object or marks it as background. This allows DETR to learn non-overlapping object queries without post-processing such as non-maximum-suppression. The Hungarian matching algorithm finds the optimal assignment between all predictions " + }, + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "text", + "content": " and all ground truth " + }, + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 606, + 288, + 690 + ], + "type": "text", + "content": ", minimizing the loss function as matching cost:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 698, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 698, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 116, + 698, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\sigma^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S}} \\ell (P, G | \\sigma) \\tag {2}", + "image_path": "2c7ed1f76b17309a1bf491b9ec69edaf073d162cc5a718381bcd5a1d30b1db94.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": " captures all possible assignments from " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": ". For each assigned prediction, the loss " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": " maximizes its class log-likelihood and fits its bounding box. For unassigned predictions, the loss " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": " reduces both the objectness score " + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 326, + 547, + 387 + ], + "type": "text", + "content": " and class log-likelihood for all classes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 395, + 375, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 395, + 375, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 375, + 407 + ], + "type": "text", + "content": "4. DECOLA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 594 + ], + "type": "text", + "content": "Our detection transformer conditioned on language, DECOLA, changes the DETR architecture in one remarkable way: Object queries are conditioned on a language embedding. Figure 2 illustrates this change. This simple change has a few important implications: First, it allows the language embedding to control and focus queries to localize on the concepts at hand. Second, it removes any contention between different object classes. Each class present in the image uses the same amount of queries. Third, it generalizes to unseen classes by leveraging semantic knowledge encoded in language embedding throughout the detection pipeline. In the remainder of this section, we highlight the changes in the architecture and training objective for conditioning (DECOLA Phase 1), and self-training on image-level data for open-vocabulary detection (DECOLA Phase 2)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "text", + "content": "Language-conditioned query selection. DECOLA conditions queries to a specific object category by modeling the objectness function as a similarity score between a region feature " + }, + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "inline_equation", + "content": "x_{i,j}" + }, + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "text", + "content": " and a text representation of a category name " + }, + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "inline_equation", + "content": "t(y)" + }, + { + "bbox": [ + 304, + 595, + 547, + 654 + ], + "type": "text", + "content": " using their cosine similarity:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 659, + 546, + 686 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 659, + 546, + 686 + ], + "spans": [ + { + "bbox": [ + 312, + 659, + 546, + 686 + ], + "type": "interline_equation", + "content": "s _ {y} \\left(x _ {i, j}\\right) = \\frac {\\left\\langle x _ {i , j} , t (y) \\right\\rangle}{\\left\\| x _ {i , j} \\right\\| \\left\\| t (y) \\right\\|}, Q _ {y} = \\operatorname {t o p k} _ {x _ {i, j}} \\left(s _ {y} \\left(x _ {i, j}\\right)\\right) \\tag {3}", + "image_path": "9db5f806c04a13804f7d255321404216b6a3c1d15d6cfccd37d706079c340b40.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 713 + ], + "type": "text", + "content": "The above objective avoids any inter-class calibration issue common in imbalanced data [7, 55, 57]. Queries do not" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16595" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": "compete, as DECOLA independently selects top-" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " scoring regions " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "Q_{y}" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ". All queries proceed to the second stage in parallel. A memory-efficient attention mechanism isolates interaction within each class. After a series of decoding layers, each language-conditioned query predicts a single scalar score, corresponding to the likelihood of the class " + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 72, + 289, + 178 + ], + "type": "text", + "content": ", and the associated box. The overall architecture mirrors the two-stage deformable DETR [75] with two modifications: a language-conditioned query, and a binary output classifier." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": "Memory-efficient modeling. DECOLA uses " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "n = |Q_y|" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " queries per class for " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " classes. Generally, " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " is smaller than the total number of queries " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "|Q|" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " of a standard DETR model. However, since we produce " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " queries per class, the total number of queries in DECOLA is much larger " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "|Q| \\ll nK" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": ". A naive implementation of the DETR decoder is unable to cope with the " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "O(n^2 K^2)" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " memory requirements of the self-attention layers in the transformer decoder. We thus modify the self-attention formulation to isolate it within each class, reducing the memory cost to " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "O(n^2 K)" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": ". The actual implementation uses standard self-attention with a reshaping operation. See Figure 2 (right) for the illustration. DECOLA Phase 1: Train to condition on given concepts. Our goal is to design DECOLA to take a set of class names in an image (or a batch of images) and predict objects of the corresponding classes or backgrounds. For each class " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": ", each conditioned query " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "q_y \\in Q_y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " therefore only predicts a single presence score for class " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " and the box location. All predictions from the conditioned queries " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "P_y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": " are matched with " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "G_y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": ", the subset of ground truth with class " + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 179, + 289, + 420 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 109, + 426, + 287, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 426, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 109, + 426, + 287, + 444 + ], + "type": "interline_equation", + "content": "\\sigma_ {y} ^ {*} = \\arg \\min _ {\\sigma \\in \\mathfrak {S} _ {y}} \\ell \\left(P _ {y}, G _ {y} \\mid \\sigma\\right) \\tag {4}", + "image_path": "128b7a6ed7dcd645897046aa0c501f6f2f09a342c7d0ec6556e65602ddc6e240.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\mathfrak{S}_y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": " is the set of possible matches between " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "P_y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "G_y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": " is the binary cross-entropy loss. Unlike the original DETR objective in Eqn. 2, Eqn. 4 matches within the conditioned class. It avoids inter-class competition during training and simplifies the training objective. Instead, it learns to adapt its predictions to " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": "; the set of conditioned query " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "q_y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": " considers any objects other than " + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 450, + 287, + 534 + ], + "type": "text", + "content": " as background." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": "Pseudo-labeling weakly-labeled data. DECOLA produces highly accurate predictions when conditioned on the exact categories of a scene, as shown in Section 5.4. This makes DECOLA a strong pseudo-labeler for weakly-labeled data with either image tags or captions. We expand a large amount of such data with pseudo-bounding boxes of DECOLA Phase 1 and self-train altogether to scale up open-vocabulary object detection. Unlike other forms of weakly supervised learning such as knowledge distillation [19] and online pseudolabeling [61, 73], we simply generate labels for all images offline and jointly train over all pseudo-labeled data using the regular detection losses without any additional complication or slowdown. For each image and class " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": ", DECOLA encodes the class' language feature and predicts a set of detections " + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "inline_equation", + "content": "P_{y}" + }, + { + "bbox": [ + 46, + 534, + 289, + 714 + ], + "type": "text", + "content": ". We simply choose the most confident prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "Figure 4d shows our simple offline pseudo-labeling works better than online pseudo-labeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 420 + ], + "type": "text", + "content": "The advantage of DECOLA comes from adaptability to specified class names on a per-image basis. However, in open-vocabulary detection, the set of test classes is neither known a priori nor available per image. Hence, we convert DECOLA into a general-purpose detector to detect all objects. We condition DECOLA with \"an object\" as the text input, and inject the class information in the second-stage classifier. Figure 2 highlights this conversion. Since DECOLA is trained to align image features to text embedding in both the first and second stages, this change only introduces inter-class calibration for multi-class object detection. We train DECOLA with pseudo-labeled and human-labeled data as a standard supervised detection training, using the standard matching algorithm of DETR (Section 3). We do not introduce any additional hyper-parameter specifically for the weakly supervised learning [73] or design choices [31, 71], extra loss functions such as alignment loss [61], nor a large teacher model for knowledge distillation [19, 65]. Generating pseudo-labels with DECOLA Phase 1 runs as fast as a regular detector and training Phase 2 is as easy as standard detection training on a supervised dataset. At a high level, DECOLA Phase 1 training objective optimizes for a strong pseudo-labeler instead of a multi-class detector, which differentiates DECOLA from prior works. Leveraging DECOLA Phase 1 to expand weakly-labeled data is the key contribution to scaling up the final open-vocabulary object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 429, + 388, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 388, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 388, + 442 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 449, + 547, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 534 + ], + "type": "text", + "content": "We evaluate the effectiveness of DECOLA in two aspects: (1) pseudo-labeling quality of DECOLA Phase 1 (Section 5.4) and (2) benchmark evaluation (Section 5.3). We consider three primary benchmarks to evaluate our final model (DECOLA Phase 2): open-vocabulary LVIS [19], standard LVIS [20], and direct zero-shot evaluation to LVIS, COCO [36], Object365 [52], and OpenImages [32]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 540, + 423, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 423, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 423, + 553 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 713 + ], + "type": "text", + "content": "Datasets and benchmarks. We mainly evaluate our method on the LVIS dataset [20], a large-vocabulary instance segmentation and object detection dataset with 1203 naturally distributed object categories. LVIS splits categories into frequent, common, and rare. For open-vocabulary LVIS, we combine frequent and common categories into LVIS-base and consider the rare categories as novel concepts used for testing only [19]. For standard LVIS, we train and evaluate all classes. Direct zero-shot transfer evaluates models trained on different detection data (e.g., Object365) and other weakly-labeled data without any prior knowledge about the target dataset such as the set of classes or object frequency. In this benchmark, we test DECOLA's generalization to dif" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16596" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 69, + 476, + 273 + ], + "blocks": [ + { + "bbox": [ + 117, + 69, + 476, + 273 + ], + "lines": [ + { + "bbox": [ + 117, + 69, + 476, + 273 + ], + "spans": [ + { + "bbox": [ + 117, + 69, + 476, + 273 + ], + "type": "table", + "html": "
methoddataAPbox novelAPcAPfmAPbox
ResNet-50 (1K)
OV-DETR† [65]LVIS-base, IN-21K18.024.831.826.4
baselineLVIS-base10.230.938.030.1
baseline + self-trainLVIS-base, IN-21K19.231.737.131.7
DECOLA Phase 2LVIS-base, IN-21K23.8 (+4.6)34.4 (+2.7)38.3 (+1.2)34.1 (+2.4)
ResNet-50
baselineLVIS-base9.433.840.432.2
baseline + self-trainLVIS-base, IN-21K23.236.541.636.2
DECOLA Phase 2LVIS-base, IN-21K27.6 (+4.4)38.3 (+1.8)42.9 (+1.3)38.3 (+2.1)
Swin-B
baselineLVIS-base16.243.849.141.1
baseline + self-trainLVIS-base, IN-21K30.843.645.942.3
DECOLA Phase 2LVIS-base, IN-21K35.7 (+4.9)47.5 (+3.9)49.7 (+3.8)46.3 (+4.0)
Swin-L
DITO* [28]O365, LVIS-base, DataComp-1B45.8--44.2
OWLv2‡ [41]O365, LVIS-base, VG, WebLI45.9--50.4
baselineO365, LVIS-base21.953.357.749.6
baseline + self-trainO365, LVIS-base, IN-21K36.553.556.551.8
DECOLA Phase 2O365, LVIS-base, IN-21K46.9 (+10.4)56.0 (+2.5)58.0 (+1.5)55.2 (+3.6)
", + "image_path": "4b05f749b7793e547ba3fd1be801166453298dd417c8c783d360752fe83aee9a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 341, + 294, + 494 + ], + "blocks": [ + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "text", + "content": "Table 1. Open-vocabulary LVIS using DETR architectures. " + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "text", + "content": ": we further finetune the official OV-DETR model trained on LVIS-base by self-training on ImageNet-21K data similar to \"baseline + self-train\" and DECOLA. " + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "text", + "content": ": uses CLIP L/14, which is comparable to Swin-L backbone. " + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 275, + 547, + 331 + ], + "type": "text", + "content": ": is based on Mask R-CNN with ViT L/16. We report the improvement between \"baseline + self-train\" to DECOLA in green. Last 4 rows compare DECOLA to Swin-L or equivalent scale of models that use additional detection data (e.g., Objects365, VG [30]) and billion-scale weakly-labeled data (DataComp-1B [15], WebLI [6]) for training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 341, + 294, + 494 + ], + "lines": [ + { + "bbox": [ + 59, + 341, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 59, + 341, + 294, + 494 + ], + "type": "table", + "html": "
methodframework\\(AP^{box}\\)novel\\(mAP^{box}\\)novel\\(AP^{mask}\\)novel\\(mAP^{mask}\\)
ViLD [19]Mask R-CNN16.727.816.625.5
RegionCLIP [71]Mask R-CNN--17.122.5
DetPro [12]Mask R-CNN20.828.419.825.9
PromptDet [14]Mask R-CNN21.425.3--
F-VLM [31]Mask R-CNN--18.624.2
BARON [61]Mask R-CNN23.229.522.627.6
OADP [58]Mask R-CNN21.928.721.726.6
EdaDet [54]Mask R-CNN--23.727.5
VLDet [35]CenterNet2--21.730.1
\\(\\mathsf{CORA}^{+}\\) [62]CenterNet228.1---
Rasheed et al. [47]CenterNet2--25.232.9
Detic-base [73]CenterNet217.633.816.430.2
Detic [73]CenterNet226.736.324.632.4
DECOLA labelsCenterNet229.537.727.033.7
", + "image_path": "35cf24c708fd444d521de86984b107f692a062a9ed5f3f14db0de67abc6f408b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 104, + 495, + 243, + 505 + ], + "lines": [ + { + "bbox": [ + 104, + 495, + 243, + 505 + ], + "spans": [ + { + "bbox": [ + 104, + 495, + 243, + 505 + ], + "type": "text", + "content": "(a) Comparison with ResNet-50 backbone." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 309, + 341, + 550, + 494 + ], + "blocks": [ + { + "bbox": [ + 309, + 341, + 550, + 494 + ], + "lines": [ + { + "bbox": [ + 309, + 341, + 550, + 494 + ], + "spans": [ + { + "bbox": [ + 309, + 341, + 550, + 494 + ], + "type": "table", + "html": "
methodbackbone\\(AP_{\\text{novel}}^{box}\\)\\(mAP_{\\text{novel}}^{box}\\)\\(AP_{\\text{novel}}^{mask}\\)\\(mAP_{\\text{novel}}^{mask}\\)
RegionCLIP [71]R50×4--22.032.3
CondHead [59]R50×424.133.724.432.0
ViLD [19]EN-B7--26.329.3
OWL-ViT [43]ViT-L/1425.634.7--
F-VLM [31]R50×64--32.834.9
VLDet [35]Swin-B--26.338.1
3Ways [1]NFNet-F630.144.6--
RO-VIT [29]ViT-L/1632.134.0--
CFM-ViT [27]ViT-L/1635.638.533.936.6
DITO [28]ViT-B/1634.936.932.534.0
CoDet [40]Swin-B--29.439.2
Detic-base [73]Swin-B24.643.021.938.4
Detic [73]Swin-B36.645.733.840.7
DECOLA labelsSwin-B38.446.735.342.0
", + "image_path": "4050687fb3698defcf31652fce07636378dd5f5a564794f0027e9be6862e1641.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 378, + 495, + 473, + 505 + ], + "lines": [ + { + "bbox": [ + 378, + 495, + 473, + 505 + ], + "spans": [ + { + "bbox": [ + 378, + 495, + 473, + 505 + ], + "type": "text", + "content": "(b) System-level comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 509, + 546, + 543 + ], + "lines": [ + { + "bbox": [ + 46, + 509, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 546, + 543 + ], + "type": "text", + "content": "Table 2. Open-vocabulary LVIS using Mask R-CNN and CenterNet2 detectors. Methods in both tables use LVIS-base as the only human-labeled data for fair comparison. For system-level comparison (right), we include methods with non R-CNN architectures such as OWL-ViT [43]. The results show the impact of high-quality pseudo-labels generated by DECOLA Phase 1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 552, + 288, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 288, + 613 + ], + "type": "text", + "content": "ferent domains. We evaluate DECOLA on LVIS, COCO [36], Object365 [52], and OpenImages [32] in a fully zero-shot manner. All our models use the ImageNet-21K [50] dataset as weakly labeled data, which contains 14M of object-centric images annotated with a single class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate DECOLA on " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{novel/rare}}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_c}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_f}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": ", and mAP following the LVIS evaluation metric [20]. We highlight the results in all three groups since we believe open-vocabulary detectors should not compensate for the performance of common/frequent classes for novel/rare classes. We evaluate both " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{box}}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AP^{mask}}" + }, + { + "bbox": [ + 46, + 618, + 289, + 714 + ], + "type": "text", + "content": " for object detection and instance segmentation. For zero-shot transfer benchmark with COCO and Object365, we" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": "use AP, " + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": " following prior work [19, 73]. For OpenImages, we report " + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}^{\\mathrm{flat}}" + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": " on the expanded label space [73, 74]. For zero-shot transfer to LVIS, we consider LVIS minival [26] and standard LVIS v1.0 validation set and report " + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{\\mathrm{fixed}}" + }, + { + "bbox": [ + 304, + 552, + 547, + 708 + ], + "type": "text", + "content": " [9] following the prior works [26, 34, 38]. In addition, we pursue a more direct measurement of the generated pseudo-labeling quality. Hence, we define conditioned mAP/AR (c-mAP/AR) and compare it to baseline open-vocabulary detectors. c-mAP measures the detection performance in mAP when the detector is provided the set of ground truth classes in each image. For example in Figure 1, both detectors use \"cat\", \"mentos\" and \"cola\" as given. This extra information is used to select scores to rank" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16597" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 289, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 289, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 289, + 203 + ], + "type": "table", + "html": "
methoddataAPboxraremAPbox
ResNet-50
baselineLVIS26.335.6
baseline + self-trainLVIS, IN-21K30.036.6
DECOLA Phase 2LVIS, IN-21K35.9 (+5.9)39.4 (+2.8)
Swin-B
baselineLVIS38.344.5
baseline + self-trainLVIS, IN-21K42.045.2
DECOLA Phase 2LVIS, IN-21K47.4 (+5.4)48.3 (+3.1)
Swin-L
baselineO365, LVIS49.354.4
baseline + self-trainO365, LVIS, IN-21K48.753.4
DECOLA Phase 2O365, LVIS, IN-21K54.9 (+6.2)56.4 (+3.0)
", + "image_path": "d6fb5e4224f302780ec060d726045c08107f82aa557bfb54e6140e19d18d6750.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 214, + 289, + 308 + ], + "blocks": [ + { + "bbox": [ + 118, + 205, + 219, + 213 + ], + "lines": [ + { + "bbox": [ + 118, + 205, + 219, + 213 + ], + "spans": [ + { + "bbox": [ + 118, + 205, + 219, + 213 + ], + "type": "text", + "content": "(a) Standard LVIS with DETR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 214, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 53, + 214, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 289, + 308 + ], + "type": "table", + "html": "
methodAPboxraremAPboxAPmaskraremAPmask
ResNet-50
Detic-base [73]28.235.325.631.4
Detic [73]31.436.829.733.2
DECOLA labels35.6 (+4.2)38.6 (+1.8)32.1 (+2.4)34.4 (+1.2)
Swin-B
Detic-base [73]39.945.435.940.7
Detic [73]45.846.941.741.7
DECOLA labels47.6 (+1.8)48.5 (+1.6)43.7 (+2.0)43.6 (+1.9)
", + "image_path": "68559292312b3194c1009dfab04fb8ea457c7ab17d317e642e3c4822e69cef2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 308, + 227, + 318 + ], + "lines": [ + { + "bbox": [ + 110, + 308, + 227, + 318 + ], + "spans": [ + { + "bbox": [ + 110, + 308, + 227, + 318 + ], + "type": "text", + "content": "(b) Standard LVIS with CenterNet2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 322, + 287, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 344 + ], + "type": "text", + "content": "Table 3. Standard LVIS benchmark. DECOLA shows consistent improvement over different model scales and architectures." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 350, + 288, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 288, + 386 + ], + "type": "text", + "content": "the final predictions (belines), or directly condition the detector (DECOLA). We analyze the model's behavior and label quality in Section 5.4 and Section ?? in supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 394, + 104, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 104, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 104, + 406 + ], + "type": "text", + "content": "5.2. Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 413, + 287, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 556 + ], + "type": "text", + "content": "DECOLA is based on two-stage Deformable DETR [75]. As described in Section 4, the first-stage objectness function for query selection is replaced by a similarity score between the image feature and the CLIP text embedding of each class name. We train the detector with the improved DETR training recipe [44, 70]: look-forward-twice, larger MLP hidden dimension, no dropout, etc. We consider four backbones: a ResNet-50 [21], Swin-B and L for all LVIS benchmarks, and Swin-T and L for the direct zero-shot transfer. Unless otherwise mentioned, all backbones are pretrained on the ImageNet-21K dataset [49]. Next, we describe our key baseline models to directly compare to DECOLA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 652 + ], + "type": "text", + "content": "Baseline. We design a baseline open-vocabulary detector to closely compare to DECOLA. Inspired by Detic [73], baseline replaces classification layers with the class embedding of the pretrained CLIP text encoder and is trained using Federated Loss [72]. DECOLA Phase 1 and baseline are trained using human-labeled data (e.g., LVIS-base). All other settings (training dataset, number of training iterations, etc.) are kept the same between DECOLA and baseline." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 712 + ], + "type": "text", + "content": "Baseline + self-train. Similar to DECOLA Phase 2, we self-train baseline on weakly-labeled data. For the self-training algorithm, we use online self-training with max-size loss from Detic [73] as baseline comparison (baseline + self-train) to DECOLA Phase 2. We tested max-size and max" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "score losses from Detic [73] (online pseudo-labeling) as well as offline pseudo-labeling similar to DECOLA, and max-size loss consistently performed the best." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 109, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 547, + 180 + ], + "type": "text", + "content": "DECOLA labels. We train a two-stage detector for broader comparison: CenterNet2 [72]. Specifically, we use Detric's baseline model (\"Detric-base\"), a CenterNet2 trained on LVIS-base with CLIP embedding, and finetune on pseudolabeled ImageNet-21K data using DECOLA Phase 1 of the same backbone size. We denote this as \"DECOLA labels\"." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": "Efficient modeling. For DECOLA Phase 1, we use " + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "inline_equation", + "content": "n = 300" + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": " queries per class. One memory and time bottleneck during DECOLA training is the first-stage loss computation. The original Deformable DETR computes Hungarian matching with all pixels to all objects in a class-agnostic manner, which is " + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "inline_equation", + "content": "\\sum_{l \\in L} H_l \\cdot W_l" + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": " predictions. To reduce the memory and time cost, we only consider the top " + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "inline_equation", + "content": "K = 10,000" + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": " confident pixels for each class " + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": " during the first-stage matching and loss computation. Together with memory-efficient self-attention (Sec. 4), the training time and memory cost of DECOLA increases by less than " + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 182, + 547, + 312 + ], + "type": "text", + "content": " over the baselines (See Table 6)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "content": "Training details. Following Detic [73], we train DECOLA Phase 1 and baseline " + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "content": " on LVIS-base, and further fine-tune for another " + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "content": " on ImageNet-21K data with pseudolabeling. For training CenterNet2 with DECOLA labels, we combine pseudo-labels from different image resolutions for " + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "inline_equation", + "content": "h \\in \\{240, 280, 320, 360, 400\\}" + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 314, + 547, + 434 + ], + "type": "text", + "content": " is the shorter side of the image. Note that this mimics the random image resizing data augmentation during standard detection training. We use Detector2 [63] based on PyTorch [45] in all of the experiments. More details are in Section ?? of supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 443, + 390, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 390, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 390, + 455 + ], + "type": "text", + "content": "5.3. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "text", + "content": "Open-vocabulary LVIS. Table 1 compares DECOLA to baseline as well as state-of-the-art DETR-based open-vocabulary detectors; OV-DETR [65], OWLv2 [41], and baseline + self-train. For a fair comparison, we further fine-tune the official OV-DETR model checkpoint on ImageNet21K for " + }, + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 462, + 547, + 713 + ], + "type": "text", + "content": " schedule same as DECOLA Phase 2 and baseline + self-train. For all backbone scales, we show consistent improvement over other methods. Notably, baseline + self-train exhibits degradation in frequent classes as a trade-off with improved novel classes, which is commonly observed behavior in other open-vocabulary detection methods, too. DECOLA improves all categories consistently, which highlights the quality of our pseudo-labels. In the last rows with the Swin-L backbone, we report the result of two concurrent works, DITO [28] (Mask R-CNN-based) and OWLv2, to compare to the method that uses additional detection data (Object365) and billion-scale web data (DataComp1B [15], WebLI [6]). DECOLA demonstrates large improvement over the state-of-the-arts despite using orders of magnitude smaller training data and compute resources. To further examine DECOLA's scalability, we test the pseudo" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16598" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 535, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 535, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 535, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 535, + 252 + ], + "type": "table", + "html": "
methoddataLVIS minivalLVIS v1.0 val
APboxrareAPcAPfmAPboxAPboxrareAPcAPfmAPbox
Swin-T
MDETR* [26]LVIS, GoldG, RefC20.924.924.324.27.422.725.022.5
GLIP [34]O365, GoldG, Cap4M20.821.431.026.010.112.525.517.2
GroundingDINO [38]O365, GoldG, Cap4M18.123.332.727.4----
GLIPv2 [69]O365, GoldG, Cap4M---29.0----
MQ-GroundingDINO† [64]O365, GoldG, Cap4M, LVIS-5VQ21.726.235.230.212.917.431.422.1
MQ-GLIP† [64]O365, GoldG, Cap4M, LVIS-5VQ21.027.534.630.415.418.430.422.6
DECOLA Phase 2O365, IN-21K‡32.832.031.832.027.224.928.026.6
Δ(+12.0)(+8.7)-(+3.0)(+17.1)(+12.4)(+2.5)(+9.4)
Swin-L
GLIP [34]FourODs, GoldG, Cap24M28.234.341.537.317.123.335.426.9
GroundingDINO [38]O365, OI, GoldG, Cap4M, COCO, RefC22.230.738.833.9----
MQ-GLIP† [64]FourODs, GoldG, Cap24M, LVIS-5VQ34.541.246.943.426.932.041.334.7
OWLv2 [41]O365, VG, WebLI39.0--38.134.9--33.5
DECOLA Phase 2O365, OI, IN-21K‡41.538.034.936.832.929.130.330.2
Δ(+2.5)(+3.7)---(+5.8)--
", + "image_path": "88facdde53a3e545c6b64ab5665a0214e7bf03c1e119b45c2f610a89994e0c0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 293, + 291, + 415 + ], + "blocks": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "type": "text", + "content": "Table 4. Direct zero-shot transfer to LVIS. †: methods use 5 per-class vision queries of LVIS dataset (denoted as “LVIS-5VQ”), which use images and annotations to extract instance-level features. ☆: MDETR uses ResNet-101 backbone and trained fully-supervised on LVIS. Results that use LVIS data are in gray. ‡: Full ImageNet-21K. No LVIS information is used in DECOLA." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 293, + 291, + 415 + ], + "lines": [ + { + "bbox": [ + 48, + 293, + 291, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 291, + 415 + ], + "type": "table", + "html": "
methodCOCOO365OI
AP\\( AP_{50} \\)\\( AP_{75} \\)AP\\( AP_{50} \\)\\( AP_{75} \\)\\( AP_{50}^{flat} \\)
R-CNNs
ViLD [19]36.655.639.811.818.212.6-
F-VLM [31]32.553.134.611.919.212.6-
DetPro [12]34.953.837.412.118.812.9-
BARON [61]36.356.139.313.621.014.5-
Detic [73]39.156.342.214.220.715.242.9
DETRs
OV-DETR [65]38.158.441.1----
Detic [73]39.856.643.314.521.415.541.6
DECOLA Phase 240.357.043.715.022.016.043.3
", + "image_path": "d5132d0af4e0378f011ac9b720469b4f6cc9fe273839ad595fcaed4249f232b9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 450, + 260, + 521 + ], + "blocks": [ + { + "bbox": [ + 46, + 418, + 288, + 450 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 288, + 450 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 288, + 450 + ], + "type": "text", + "content": "Table 5. Cross-dataset generalization benchmark on COCO, Object365, and OpenImages. All models use a ResNet-50 backbone and train on LVIS-base and weakly-labeled data." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 450, + 260, + 521 + ], + "lines": [ + { + "bbox": [ + 73, + 450, + 260, + 521 + ], + "spans": [ + { + "bbox": [ + 73, + 450, + 260, + 521 + ], + "type": "table", + "html": "
methodtraintest
timemem.timemem.
baseline44 h8.9 G0.07 s/img2.5 G
+ self-train45 h10.2 G0.07 s/img2.5 G
OV-DETR73 h22.0 G6.4 s/img3.4 G
DECOLA Phase 149 h12.6 G--
DECOLA Phase 245 h8.9 G0.07 s/img2.8 G
", + "image_path": "a6e8ad1b08545e5985f6ba98570ee374b988bdc61475ecf225e0f578ed49ff68.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 525, + 288, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 288, + 559 + ], + "type": "text", + "content": "Table 6. Efficiency. Training time is measured with 8 DGX V100 on ResNet-50, 2 images per-GPU. float16 is used in both training and testing. OV-DETR uses the original optimized inference." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 567, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 676 + ], + "type": "text", + "content": "labeling capability of DECOLA on R-CNN-based detectors with DECOLA labels (Detic-base finetuned with our pseudolabels). In Table 2, we compare DECOLA labels to a broad range of literature based on CenterNet2 [72] and Mask R-CNN [22]. Table 2a compares methods with ResNet-50 backbone and Table 2b compares larger scale backbones for system-level comparison. In both tables, DECOLA clearly improve upon the state-of-the-art by large margins without additional complication in training and bells-and-whistles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Standard LVIS. Tables 3 evaluate DECOLA and baseline on the standard LVIS benchmark, where all object categories are used to fully supervise the detectors. Similar to the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": "open-vocabulary LVIS, we compare DETR architectures in Table 3a and R-CNN architectures in Table 3b. Table 3a shows that DECOLA remarkably improves baseline by 9.5, 9.1, and 5.6 points on " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": ", outperforming baseline + self-train by 5.9, 5.4, and 6.2 " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " for ResNet-50, Swin-B, and Swin-L backbones, respectively. Similarly in Table 3b, DECOLA labels further improves the baseline of Detic by 7.4 and 7.7 " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": ", and outperforms Detic [73] by 4.2 and 1.8 " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{rare}}^{\\mathrm{box}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " with ResNet-50 and Swin-B backbones, respectively. Direct zero-shot evaluation. For direct zero-shot evaluation, we train DECOLA with Swin-T [39] and use Object365 data for Phase 1, and ImageNet-21K for Phase 2 (full dataset and classes). We compare to MDETR [26], GLIP [34], GroundingDINO [38], and MQ-Det [64] finetuned from GLIP and GroundingDINO. Table 4 shows the results. DECOLA outperforms the previous state-of-the-arts, by " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "12.0 / 17.1\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "3.0 / 9.4\\mathrm{mAP}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " on LVIS minival and LVIS v1.0 val, respectively. Note that all other methods use much richer detection labels from GoldG data [26], a collection of grounding data (box and text expression pairs) curated by MDETR. Furthermore, other benchmark methods show highly imbalanced " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_f}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": " in both LVIS minival and LVIS v1.0 val (10-20 points gap). We hypothesize that the large collection of training data coincides with LVIS vocabulary, as all data follows a natural distribution of common objects. Also, DECOLA enjoys significantly faster run-time compared to all other models that undergo BERT encoding for grounding, which requires more than 50 forward passes per image in order to predict all LVIS categories. Similarly, our Swin-L model outperforms GroundingDINO and GLIP by 19.3 and " + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "inline_equation", + "content": "13.3\\mathrm{AP}_{\\mathrm{rare}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 713 + ], + "type": "text", + "content": ", respectively, despite much smaller training data compared to FourODs [34] and match to OWLv2 with 10-B private WebLI data [6]. Table 5 further examines the generality of DECOLA on different domains. DECOLA Phase 2 with ResNet-50 outperforms all other competitive baselines." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16599" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 101, + 72, + 494, + 282 + ], + "blocks": [ + { + "bbox": [ + 101, + 72, + 494, + 282 + ], + "lines": [ + { + "bbox": [ + 101, + 72, + 494, + 282 + ], + "spans": [ + { + "bbox": [ + 101, + 72, + 494, + 282 + ], + "type": "image", + "image_path": "14d1dd4ea0cc274c81c008bde984323767dcd651e32a292f669c9fc0c03b26f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 288, + 545, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 288, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 545, + 333 + ], + "type": "text", + "content": "Figure 3. Examples of prediction on unseen categories. Images from ImageNet-21K dataset. Boxes are the most confident prediction from DECOLA and baseline. DECOLA conditions on the ground truth label, and baseline selects the max-score box of the ground truth class. Images are all from unseen categories, which neither model was trained on. Green: DECOLA Phase 1 trained on LVIS-base. Red: baseline trained on LVIS-base. Both models use a Deformable DETR detector with a ResNet-50 backbone. More in Fig. ?? and ?? of supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 338, + 170, + 432 + ], + "blocks": [ + { + "bbox": [ + 50, + 338, + 170, + 432 + ], + "lines": [ + { + "bbox": [ + 50, + 338, + 170, + 432 + ], + "spans": [ + { + "bbox": [ + 50, + 338, + 170, + 432 + ], + "type": "image", + "image_path": "2bd77d56d18c866bb0b2263450d1459c0dc6abbee2c76f4199e508cf6f9f7261.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 436, + 163, + 445 + ], + "lines": [ + { + "bbox": [ + 59, + 436, + 163, + 445 + ], + "spans": [ + { + "bbox": [ + 59, + 436, + 163, + 445 + ], + "type": "text", + "content": "(a) c-AP for ResNet-50 (IN-1K)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 338, + 289, + 431 + ], + "blocks": [ + { + "bbox": [ + 171, + 338, + 289, + 431 + ], + "lines": [ + { + "bbox": [ + 171, + 338, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 171, + 338, + 289, + 431 + ], + "type": "image", + "image_path": "1ed8e239b3cf9a89151578fbf69b0216a65323826667a60ced142f0d5c08d527.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 199, + 436, + 266, + 444 + ], + "lines": [ + { + "bbox": [ + 199, + 436, + 266, + 444 + ], + "spans": [ + { + "bbox": [ + 199, + 436, + 266, + 444 + ], + "type": "text", + "content": "(b) c-AP for Swin-B" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 544, + 472 + ], + "type": "text", + "content": "Figure 4. Analyzing DECOLA. All plots show the conditioned AP/AR for unseen classes. We compare DECOLA Phase 1 and baselines. We highlight more detailed analyses about c-AP and c-AR in Section ??, Table ?? and ?? of the supplementary materials." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 290, + 338, + 420, + 425 + ], + "blocks": [ + { + "bbox": [ + 290, + 338, + 420, + 425 + ], + "lines": [ + { + "bbox": [ + 290, + 338, + 420, + 425 + ], + "spans": [ + { + "bbox": [ + 290, + 338, + 420, + 425 + ], + "type": "image", + "image_path": "8b6e4c6ff3f4bbd78d357c45926e0fa2b19cc757fa2a03f4e19e31ca2b78f61f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 436, + 420, + 445 + ], + "lines": [ + { + "bbox": [ + 302, + 436, + 420, + 445 + ], + "spans": [ + { + "bbox": [ + 302, + 436, + 420, + 445 + ], + "type": "text", + "content": "(c) c-AR for 1st and 2nd stage boxes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 420, + 338, + 543, + 429 + ], + "blocks": [ + { + "bbox": [ + 420, + 338, + 543, + 429 + ], + "lines": [ + { + "bbox": [ + 420, + 338, + 543, + 429 + ], + "spans": [ + { + "bbox": [ + 420, + 338, + 543, + 429 + ], + "type": "image", + "image_path": "96c45373ba0c1f741ab0eb7382b39c0c6425e02cb0388b2f7e88658524e408c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 436, + 531, + 444 + ], + "lines": [ + { + "bbox": [ + 432, + 436, + 531, + 444 + ], + "spans": [ + { + "bbox": [ + 432, + 436, + 531, + 444 + ], + "type": "text", + "content": "(d) online vs offline: c-AP@20" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 478, + 111, + 490 + ], + "type": "text", + "content": "5.4. Analyses" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 497, + 288, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 288, + 520 + ], + "type": "text", + "content": "In this section, we analyze the model's behavior with conditioned mAP/AR (c-AP/AR) (defined in Sec. 5.1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 521, + 288, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 581 + ], + "type": "text", + "content": "Pseudo-labeling quality. Figure 4a and 4b compare DECOLA Phase 1, OV-DETR, and baseline on c-AP for unseen classes. Compared to baseline and OV-DETR, DECOLA Phase 1 generates much higher quality pseudo-labels, especially in low-shot regimes. See examples in Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 665 + ], + "type": "text", + "content": "Impact of conditioning. In Figure 4c, we compare c-AR of unseen classes. This measures the detector's ability to localize objects of interest when pseudo-labeling. We observe significant improvement in c-AR on both first-stage (proposals) and second-stage (predictions) due to our conditioning mechanism. This result demonstrates the key difference between DECOLA and other open-vocabulary detectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Pseudo-labeling algorithms. Figure 4d shows the c-AP of baseline + self-train and DECOLA Phase 1 for unseen classes with 20 predictions per-image. Each red bar indicates the percent of training iteration during self-training. Online" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 479, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 526 + ], + "type": "text", + "content": "self-labeling suffers a sharp drop during the early iterations, and c-AP after full iterations still underperforms compared to DECOLA Phase 1. DECOLA's simple approach of offline self-training is more stable and effective." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 544, + 378, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 378, + 557 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 378, + 557 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 547, + 651 + ], + "type": "text", + "content": "In this paper, we explore a new open-vocabulary detection framework, DECOLA. It adjusts its inner workings to the concepts that the user asks to reason over by conditioning on a language embedding. Our detector generates high-quality pseudo-labels on weakly labeled data through the conditioning mechanism. We finetune it with the pseudolabels to build the state-of-the-art open-vocabulary detector." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement. We thank Xingyi Zhou, Yue Zhao, Jeffrey Ouyang-Zhang, and Nayeon Lee for valuable feedback. This material is in part based upon work supported by the National Science Foundation under Grant No. IIS-1845485 and IIS-2006820." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16600" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 133 + ], + "type": "text", + "content": "[1] Relja Arandjelović, Alex Andonian, Arthur Mensch, Olivier J Hénaff, Jean-Baptiste Alayrac, and Andrew Zisserman. Three ways to improve feature alignment for open vocabulary detection. arXiv, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "type": "text", + "content": "[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In CVPR, 2018. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 200 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 245 + ], + "type": "text", + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 291 + ], + "type": "text", + "content": "[5] Qiang Chen, Xiaokang Chen, Jian Wang, Haocheng Feng, Junyu Han, Errui Ding, Gang Zeng, and Jingdong Wang. Group detr: Fast detr training with group-wise one-to-many assignment. arXiv preprint arXiv:2207.13085, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 288, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 400 + ], + "type": "text", + "content": "[6] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme Ruiz, Andreas Peter Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. PaLI: A jointly-scaled multilingual language-image model. In ICLR, 2023. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 402, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 286, + 422 + ], + "type": "text", + "content": "[7] Jang Hyun Cho and Philipp Krahenbuhl. Long-tail detection with effective class-margins. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 456 + ], + "type": "text", + "content": "[8] Jang Hyun Cho, Philipp Krahenbuhl, and Vignesh Ramanathan. Partdistillation: Learning parts from instance segmentation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 490 + ], + "type": "text", + "content": "[9] Achal Dave, Piotr Dólar, Deva Ramanan, Alexander Kirillov, and Ross Girshick. Evaluating large-vocabulary object detectors: The devil is in the details. arXiv, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 512 + ], + "type": "text", + "content": "[10] Mostafa et al. Dehghani. Scaling vision transformers to 22 billion parameters. ICML, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 546 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[12] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. 2022. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[13] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 668 + ], + "type": "text", + "content": "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In ECCV, 2022. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[15] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, Eyal Orgad, Rahim Entezari, Giannis Daras, Sarah M Pratt, Vivek" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 547, + 160 + ], + "type": "text", + "content": "Ramanujan, Yonatan Bitton, Kalyani Marathe, Stephen Mussmann, Richard Vencu, Mehdi Cherti, Ranjay Krishna, Pang Wei Koh, Olga Saukh, Alexander Ratner, Shuran Song, Hannaneh Hajishirzi, Ali Farhadi, Romain Beaumont, Sewoong Oh, Alex Dimakis, Jenia Jitsev, Yair Carmon, Vaishaal Shankar, and Ludwig Schmidt. Datacomp: In search of the next generation of multimodal datasets. In Neurips (Datasets and Benchmarks Track), 2023. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 547, + 205 + ], + "type": "text", + "content": "[16] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 498, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 498, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 498, + 217 + ], + "type": "text", + "content": "[17] Ross Girshick. Fast r-cnn. In ICCV, 2015. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 218, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 218, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 218, + 545, + 251 + ], + "type": "text", + "content": "[18] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, 2014. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 252, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 547, + 285 + ], + "type": "text", + "content": "[19] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2022. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 286, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 547, + 318 + ], + "type": "text", + "content": "[20] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019, 1, 2, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 319, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 547, + 352 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 354, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 547, + 374 + ], + "type": "text", + "content": "[22] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 398 + ], + "type": "text", + "content": "[23] Geoffrey Hinton, Oriol Vinyls, and Jeffrey Dean. Distilling the knowledge in a neural network. In Neurips, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 399, + 471, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 471, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 471, + 410 + ], + "type": "text", + "content": "[24] Gabriel et al Ilharco. Openclip, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 443 + ], + "type": "text", + "content": "[25] Ding Jia, Yuhui Yuan, Haodi He, Xiaopei Wu, Haojun Yu, Weihong Lin, Lei Sun, Chao Zhang, and Han Hu. Detrs with hybrid matching. 2023. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 445, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 547, + 487 + ], + "type": "text", + "content": "[26] Aishwarya Kamath, Mannat Singh, Yann LeCun, Ishan Misra, Gabriel Synnaeve, and Nicolas Carion. Mdetr - modulated detection for end-to-end multi-modal understanding. ICCV, 2021. 5, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 489, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 547, + 521 + ], + "type": "text", + "content": "[27] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Contrastive feature masking open-vocabulary vision transformer. In ICCV, 2023. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 555 + ], + "type": "text", + "content": "[28] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Detection-oriented image-text pretraining for open-vocabulary detection. arXiv, 2023. 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 557, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 547, + 589 + ], + "type": "text", + "content": "[29] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, 2023. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 547, + 645 + ], + "type": "text", + "content": "[30] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 689 + ], + "type": "text", + "content": "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[32] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov," + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16601" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 106 + ], + "type": "text", + "content": "Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 2020. 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 139 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[34] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 2, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 288, + 227 + ], + "type": "text", + "content": "[35] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In ICLR, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[36] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 314 + ], + "type": "text", + "content": "[37] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. DAB-DETR: Dynamic anchor boxes are better queries for DETR. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 370 + ], + "type": "text", + "content": "[38] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 414 + ], + "type": "text", + "content": "[39] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 449 + ], + "type": "text", + "content": "[40] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In Nuerips, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 471 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 471 + ], + "type": "text", + "content": "[41] Neil Houlsby Matthias Minderer, Alexey Gritsenko. Scaling open-vocabulary object detection. NeurIPS, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 288, + 514 + ], + "type": "text", + "content": "[42] Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, and Jingdong Wang. Conditional detr for fast training convergence. In ICCV, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 515, + 288, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 536 + ], + "type": "text", + "content": "[43] Matthias Minderer et al. Simple open-vocabulary object detection with vision transformers. ECCV, 2022. 1, 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 537, + 288, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 288, + 570 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 288, + 570 + ], + "type": "text", + "content": "[44] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbuhl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 571, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 288, + 657 + ], + "type": "text", + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Neurips. 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammed Uzair Khattak, Salman Khan, and Fahad Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In Neurips, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 161 + ], + "type": "text", + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Neurips, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 183 + ], + "type": "text", + "content": "[49] Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, and Lihi Zelnik. Imagenet-21k pretraining for the masses. In Neurips, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 547, + 237 + ], + "type": "text", + "content": "[50] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 547, + 293 + ], + "type": "text", + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Neurips, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 293, + 547, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 547, + 336 + ], + "type": "text", + "content": "[52] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, 2019. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 547, + 380 + ], + "type": "text", + "content": "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 381, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 381, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 381, + 547, + 403 + ], + "type": "text", + "content": "[54] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. In ICCV, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 547, + 437 + ], + "type": "text", + "content": "[55] Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 437, + 547, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 470 + ], + "type": "text", + "content": "[56] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In CVPR, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 547, + 514 + ], + "type": "text", + "content": "[57] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 547, + 557 + ], + "type": "text", + "content": "[58] Luting Wang, Yi Liu, Penghui Du, Zihan Ding, Yue Liao, Qiaosong Qi, Biaolong Chen, and Si Liu. Object-aware distillation pyramid for open-vocabulary object detection. In CVPR, 2023. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 580 + ], + "type": "text", + "content": "[59] Tao Wang. Learning to detect and segment for open vocabulary object detection. In CVPR, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 582, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 547, + 613 + ], + "type": "text", + "content": "[60] Yingming Wang, Xiangyu Zhang, Tong Yang, and Jian Sun. Anchor detr: Query design for transformer-based detector. In AAAI, 2022. 1, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 614, + 547, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 547, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 547, + 656 + ], + "type": "text", + "content": "[61] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 658, + 547, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 547, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 547, + 691 + ], + "type": "text", + "content": "[62] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, 2023. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 692, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 547, + 712 + ], + "type": "text", + "content": "[63] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. *Detectron2*, 2019. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16602" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 578 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[64] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. In Neurips, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[65] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 173 + ], + "type": "text", + "content": "[66] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 288, + 217 + ], + "type": "text", + "content": "[67] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 288, + 262 + ], + "type": "text", + "content": "[68] Cheng Zhang, Tai-Yu Pan, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. MosaicOS: A simple and effective use of object-centric images for long-tailed object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 264, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 288, + 317 + ], + "type": "text", + "content": "[69] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Neurips, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 320, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 320, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 320, + 287, + 363 + ], + "type": "text", + "content": "[70] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 365, + 288, + 408 + ], + "type": "text", + "content": "[71] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 2, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 411, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 288, + 441 + ], + "type": "text", + "content": "[72] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Probabilistic two-stage detection. arXiv preprint arXiv:2103.07461, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 445, + 288, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 288, + 486 + ], + "type": "text", + "content": "[73] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 288, + 510 + ], + "type": "text", + "content": "[74] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "text", + "content": "[75] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable " + }, + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{detr}\\}" + }, + { + "bbox": [ + 49, + 512, + 287, + 544 + ], + "type": "text", + "content": ": Deformable transformers for end-to-end object detection. In ICLR, 2021. 1, 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 49, + 545, + 288, + 578 + ], + "type": "text", + "content": "[76] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. Neurips, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16603" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_content_list.json b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da52c43dbbede0006a59ceaa0a83a2283fee160d --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_content_list.json @@ -0,0 +1,2707 @@ +[ + { + "type": "text", + "text": "Language-driven All-in-one Adverse Weather Removal", + "text_level": 1, + "bbox": [ + 205, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Yang $^{1}$ , Liyuan Pan $^{1\\dagger}$ , Yan Yang $^{2,3}$ , and Wei Liang $^{1}$ $^{1}$ Beijing Institute of Technology $^{2}$ Australian National University $^{3}$ A&F, CSIRO \n{hao.yang, liyuan.pan, liangwei}@bit.edu.cn, {yan.yang}@anu.edu.au", + "bbox": [ + 148, + 179, + 821, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 282, + 313, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "All-in-one (AiO) frameworks restore various adverse weather degradations with a single set of networks jointly. To handle various weather conditions, an AiO framework is expected to adaptively learn weather-specific knowledge for different degradations and shared knowledge for common patterns. However, existing methods: 1) rely on extra supervision signals, which are usually unknown in real-world applications; 2) employ fixed network structures, which restrict the diversity of weather-specific knowledge. In this paper, we propose a Language-driven Restoration framework (LDR) to alleviate the aforementioned issues. First, we leverage the power of pre-trained vision-language (PVL) models to enrich the diversity of weather-specific knowledge by reasoning about the occurrence, type, and severity of degradation, generating description-based degradation priors. Then, with the guidance of degradation prior, we sparsely select restoration experts from a candidate list dynamically based on a Mixture-of-Experts (MoE) structure. This enables us to adaptively learn the weather-specific and shared knowledge to handle various weather conditions (e.g., unknown or mixed weather). Experiments on extensive restoration scenarios show our superior performance.", + "bbox": [ + 75, + 315, + 473, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 672, + 209, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imaging under unpleasant degradation conditions poses challenges for vision-based systems, such as self-driving cars [15, 33, 34, 46] and outdoor surveillance systems [9, 12], which require 24/7 service regardless of adverse weather conditions like rain [6, 13, 47], haze [14, 38, 41], and snow [4, 26, 29]. To meet the safety demand, compared to tackling each type of weather condition with independent models, the joint task process, i.e., all-in-one framework, has a broader application scenario [17, 35].", + "bbox": [ + 75, + 696, + 468, + 833 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Formally, degraded images can be modeled as masked additive combination of clean images and degradation residuals [35]. Consequently, several works [5, 21, 42] use", + "bbox": [ + 75, + 833, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67f9b721831be68e375c31c389918ce754b9caeed68a7911c7624435bb03c5c9.jpg", + "image_caption": [ + "Figure 1. Score (PSNR and SSIM) comparisons. We compare our model (red) with the top 2 (blue and green) baselines on benchmark datasets with various weather scenarios. Superscripts besides evaluation metrics are used to differentiate datasets." + ], + "image_footnote": [], + "bbox": [ + 553, + 267, + 849, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a single network for all degradation types. Though shared knowledge is learned for restoration, they neglect that different degradations still hold different mathematical formulations, e.g., transmission map produced by scattering effect in haze model [41] is unnecessary for raindrop model [37].", + "bbox": [ + 496, + 565, + 893, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hence, several works [17, 35, 36, 50, 57] use different sub-networks for weather-specific knowledge learning. However, auxiliary supervisions are required to assign the sub-networks, e.g., degradation types [35] or depth maps [57]. Furthermore, the fixed sub-network architecture of existing methods restricts the diversity of learned weather-specific knowledge and their ability to handle images with various weather conditions [43, 55], such as images degraded by weather severity or conditions that have not been encountered before [1, 37], or mixed weather conditions like snow with haze in real-world scenarios.", + "bbox": [ + 496, + 642, + 895, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we question - can we restore the image degraded by various weather conditions by adaptive learning of diverse weather-specific knowledge and shared knowledge, without requiring ground truth weather types and severity data? We answer it by our language-driven all-in-one restoration (LDR) framework in two aspects.", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 101, + 886, + 232, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24902", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50aa2018c212c214be77ad863870163f9cec1a419a19d98e7652354a1aa2e1c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 124, + 88, + 209, + 140 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c734cbb34f9ad4a54233e3b575fd7f3116abc3a59f32166881e5c7402556592.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 142, + 207, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0b9a9c7f677cb834bb966921fe4a2fd811c2546718bb7a791a6da2d5bb33352.jpg", + "image_caption": [ + "(a) 7-th" + ], + "image_footnote": [], + "bbox": [ + 125, + 194, + 207, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/748723c14dc7c7702fd4c127ca11dc510965482ee47cffe92c958dfa4b1aec5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 295, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a2626627053a55e9160fba2dcd7024f4b34055165cdea5f32007169c90c4f80.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 141, + 295, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae952ad470c2933b85efac49a11dd212f781a6f6bd1fa5099db6864a4d72cb76.jpg", + "image_caption": [ + "(b) 10-th" + ], + "image_footnote": [], + "bbox": [ + 212, + 194, + 295, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b09b8e2abea57923d3950697b301fad2851e703c2df24d04acf06d48bd0b4a63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 88, + 385, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/092201e40044eb18b472aff2e3d45ee923fa93e6f2442686fb0cfd4e6da75ffe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 141, + 383, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/327a90d47932adb9d915ddc7eb3a3ad1bc0b05b56c0420d4103c0ab617a4aec8.jpg", + "image_caption": [ + "(c) 54-th" + ], + "image_footnote": [], + "bbox": [ + 302, + 194, + 383, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae711ce1ed555676d531eb23e300e1b02ecd2431428108753f089e1892e1b992.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 387, + 88, + 473, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8b4f9b714e28ca1922da243efb997a6e3051bf93c12b7ba9d2a32d6e39127d1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 141, + 473, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cf9a67bf073c1ded610eeb5a22466e8900ecd7cc57cfd505c8e1fd09c294776d.jpg", + "image_caption": [ + "(d) Input" + ], + "image_footnote": [], + "bbox": [ + 388, + 194, + 473, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bce8aaa65c6a711d0d4ddaee6d0c74b15ad66b515074070b7373b7a77ed80899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 88, + 562, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc31a9180b6675e0d7ce46b41415060d73f4098d1b87926f0675e98754f5bbc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 141, + 562, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ca35d23aa2af05ebfd10917394ad2b4bc5f4e8754c611e31ef0045d3a30fb186.jpg", + "image_caption": [ + "(e) Transweather (f) Zero-out" + ], + "image_footnote": [], + "bbox": [ + 475, + 194, + 562, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e68a7fd0f76b00c65f7c70edd7079a8b4f82f1b0025e1add05a0243fbcb59575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 88, + 651, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d4a08e5b5cf243f0d1f35b597ae6919de46668c3bdab2906ddd44afc2c5db0b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 141, + 651, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/039335b1ea8d96969a6c0aed78049623686b69b442d3c5a2e8a314696fd99f31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 194, + 651, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(f) Zero-out", + "bbox": [ + 568, + 247, + 643, + 260 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dadd85a57d3f5069af34ac9be41c970f73af73540b12034a152708f953edd466.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 655, + 90, + 785, + 135 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e34f96cbf2cd6213d6f3e77f985627acd519882d1f4c6661f2e43e331bc4a336.jpg", + "image_caption": [ + "", + "", + "", + "" + ], + "image_footnote": [], + "bbox": [ + 656, + 137, + 785, + 183 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a23680f075d1303654c8dd674959a887c0c172c5dd3c22c7e33e70ad4e46b73d.jpg", + "image_caption": [ + "(g) Distribution", + "", + "", + "", + "", + "#", + "", + "Figure 2. Ineffective computations of Transweather [42]. We find the activity of some parameters in Transweather [42] are degradation dependent, and zero-outing computations for those inactivate parameters barely affect image restoration. (a) 7-th channel, (b) 10-th channel, and (c) 54-th channel of feature maps from Transweather, the brighter, the larger activation value, are activated for degradation of rain, raindrop, and snow. We show the (d) degraded images, (e) restorations from Transweather, and (f) restorations by zero-outing inactive channels. (g) From top to bottom, distribution of $\\mathcal{L}_2$ norm for the three channels across all images in the All-weather dataset [21]." + ], + "image_footnote": [], + "bbox": [ + 656, + 185, + 790, + 247 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/09e99b2af473f8159c811dd518168b10d5ff7d0b143fc5634ca03e03c03fc11e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 90, + 843, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) The knowledge within the feature space of a pretrained vision-language (PVL) model can benefit various tasks, while its potential in our task is still under exploration. A straightforward way is to use the PVL model as the image degradation classifier. In contrast, we go one step further by reasoning diverse weather-specific knowledge from the feature space of the PVL model beyond the type of weather conditions.", + "bbox": [ + 75, + 340, + 468, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by formulating a question prompt to query the occurrence, type, and severity of degradation in a degraded image. The obtained degradation prior describes what, where, and severity of degradations in high-level textual semantics. Then, we translate the high-level degradation prior to a 2D degradation map by aligning the prior with the degraded image. This degradation map provides a pixel-wise representation of the diverse knowledge of image degradation from the PVL model.", + "bbox": [ + 75, + 462, + 468, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) We then unleash the potential for various weather removal with the guidance of degradation maps. Observing model parameters are weather-specific, e.g., the rain-related parameters are usually inactive for unrelated degradations, and zero-outing the computations of unrelated parameters also barely affects the restoration quality. Take TUM [5] and Transweather [42] as examples, the PSNR and SSIM w/wo zero-outing on the All-weather dataset are $27.93 / 26.52$ (dB) and $0.883 / 0.864$ , and $27.98 / 26.65$ (dB) and $0.884 / 0.866$ . This observation, illustrated in Fig. 2, inspired us to bypass computations for parameters unrelated to the specific weather type and severity during image restoration. With the assistance of MoE structure [28, 40], our LDR framework selects experts dynamically for restoration, therefore, ensuring adaptive learning of weather-specific knowledge which is not limited to a fixed network architecture.", + "bbox": [ + 75, + 598, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we maintain a candidate list of restoration experts and utilize the degradation map to sparsely select the most related restoration experts for each degra", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dition. By applying the selected expert pixel-wisely to restore weather-specific features, we create flexible and degradation-adaptive expert combinations/model architectures for restoration.", + "bbox": [ + 496, + 340, + 890, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Though we can have a preliminary restoration from experts restored feature, considering image regions with similar values from the degradation map tend to benefit restoration features of each other, we re-use the degradation map to aggregate the restoration features, and improve the locality of the obtained restoration by a simple convolutional feedforward network. Our main contributions are:", + "bbox": [ + 496, + 401, + 890, + 506 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present an LDR framework to adaptively remove various adverse weather conditions in an all-in-one solution;", + "- We propose a degradation map measurement module for extracting diverse weather-specific knowledge from a pre-trained vision-language model;", + "- We propose a Top-K expert restoration module, sparsely and adaptively computing pixel-wise restoration features." + ], + "bbox": [ + 500, + 511, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The overall comparison in Fig. 1 shows the superiority of our framework for handling various degradations compared to state-of-the-art methods.", + "bbox": [ + 496, + 618, + 890, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 678, + 640, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adverse Weather Restoration. The field of adverse weather restoration encompasses two distinct approaches: task-specific methods [2, 3, 23, 51, 52] train the model independently for each individual degradation and all-in-one frameworks [5, 17, 21, 32, 35, 42, 44, 54, 57] develop a single unified model capable of handling various adverse weather degradations. A prevalent strategy for the all-in-one framework is to build upon task-specific models by employing multi-task learning techniques [44, 54, 57] or knowledge distillation [5] that consolidates multiple task-specific networks into a single network. Nevertheless, these methods typically rely on fixed computations, and the computation effectiveness is discounted as model parameters are", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e554615ba8772bfb9d1856ab3bef8f3b2dd8afe4b8f6430ef2653db0cf5cbbf4.jpg", + "image_caption": [ + "Figure 3. The pipeline of our method. Given the input degraded image $\\mathbf{I}^d$ , we aim to recover the clean image $\\hat{\\mathbf{I}}^c$ . To tackle various adverse weather conditions, we format a question prompt $\\mathbf{T}$ , and query a pre-trained vision-language (PVL) model with $\\mathbf{I}^d$ and $\\mathbf{T}$ , to estimate the degradation type and severity. The generated descriptions $\\mathbf{P}^{txt}$ are transformed by a multilayer perceptron (MLP) to get the degradation prior $\\mathbf{P}^{emb}$ . Meanwhile, we extract a feature map $\\mathbf{X}$ from the input degraded image $\\mathbf{I}^d$ . A degradation map $\\mathbf{M}$ is computed by cross-attending $\\mathbf{X}$ with the prior $\\mathbf{P}^{emb}$ , describing pixel-wise degradation pattern. We maintain a trainable candidate list of convolution filters, $\\{\\mathbf{F}_n|n = 1,\\dots,N\\}$ , and denote them as experts. We first parse the degradation map $\\mathbf{M}$ into a score map $\\mathbf{S}$ via a feedforward network $FFN^m$ and a softmax layer, and then use $\\mathbf{S}$ to find the best $K$ experts describing the degradation of $\\mathbf{X}$ . Best experts are convolved with $\\mathbf{X}$ to generate an intermediate feature map $\\hat{\\mathbf{X}}^{int}$ . Finally, a cross-attention layer is used to aggregate $\\hat{\\mathbf{X}}^{int}$ with the guidance of the degradation map $\\mathbf{M}$ . We improve the feature locality with a feedforward network $FFN^{int}$ , and the output $\\hat{\\mathbf{X}}$ is decoded to the clean image $\\hat{\\mathbf{I}}^c$ ." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 867, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "separately and degradation-dependently activated.", + "bbox": [ + 76, + 455, + 406, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, several methods [17, 35, 57] have been developed to overcome these limitations by focusing on learning weather-specific knowledge. They utilize indicators like degradation types [27, 35] or depth maps [57] during training to categorize and direct images to appropriate sub-networks for restoration. Though AirNet [17] learns weather-specific knowledge with contrastive learning. Its fixed sub-network designs can only restore images with certain degradation types, and does not generalize to unseen weather degradations, e.g., rain mixed with haze. Furthermore, these mentioned works overlook the fact that varying levels of degradation severity also warrant adaptively tailored computational processes, for a specific degradation type.", + "bbox": [ + 75, + 470, + 468, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diverging from previous methods, we harness PVL models to reason both the type and severity of degradations, enabling adaptive restoration with dynamic sub-networks informed by this rich, weather-specific knowledge.", + "bbox": [ + 75, + 684, + 468, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse Mixture of Expert. The concept of expert models [28] is defined as a subset of model parameters/computation, where each expert is specialized in handling distinct aspects of the input data. The sparse mixture of experts usually employs a routing mechanism [40] to dynamically and adaptively forward input to a subset of these experts, bypassing unnecessary and irrelevant computations. Given fixed computation costs, due to the sparsity, this framework can readily scale the number of experts to improve the model capability that has been widely verified", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the domain of natural language processing [10, 40, 58] and computer vision [7, 11, 22, 30]. We study a prior derived from a PVL model to dynamically select experts, and adaptively apply expert to restore degraded images.", + "bbox": [ + 496, + 455, + 890, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language Model. With the release of ChatGPT, remarkable reasoning abilities of large language model (LLM) [16, 18, 25, 31, 39, 56] have been shown. It motivates the vision-language community to transfer the reasoning abilities to visual data. A common pipeline is to project an image to a joint space of LLM [19, 24, 25, 48, 49], and feed the projected image alongside user-provided text for conditional text generation. This paper proposes to leverage the PVL (e.g., [16]) to generate prior of an adverse weather degraded image for adaptive image restoration.", + "bbox": [ + 496, + 515, + 892, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 679, + 589, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. Given an input image $\\mathbf{I}^{\\mathrm{d}}$ degraded by adverse weather, we aim to recover the corresponding clean image $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ . Our method adaptively recovers $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ with degradation prior $\\mathbf{P}^{\\mathrm{emb}}$ obtained from a pre-trained vision-language (PVL) model. The overall architecture is shown in Fig. 3.", + "bbox": [ + 496, + 703, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To derive degradation prior $\\mathbf{P}^{\\mathrm{emb}}$ , we query a PVL model with a question prompt $\\mathbf{T}$ to reason about the occurrence, type, and severity of degradation in the input image $\\mathbf{I}^{\\mathrm{d}}$ , outputting $\\mathbf{P}^{\\mathrm{txt}}$ . Then a mapping network projects $\\mathbf{P}^{\\mathrm{txt}}$ to $\\mathbf{P}^{\\mathrm{emb}}$ . Meanwhile, we encode $\\mathbf{I}^{\\mathrm{d}}$ to $\\mathbf{X}$ by using an encoder.", + "bbox": [ + 496, + 779, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adaptively restore $\\mathbf{X}$ into $\\hat{\\mathbf{X}}$ in the embedding space, and decode $\\hat{\\mathbf{X}}$ to a restored image $\\hat{\\mathbf{l}}^{\\mathrm{c}}$ , with three steps: i) degradation map measurement, measuring degradation for", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24904", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/29e84c1c9975a12dcb97d88fd3be8ddacf2c2f37f104412b483ee2c619870020.jpg", + "image_caption": [ + "Figure 4. Examples text descriptions from LISA [16] for reasoning weather degraded images." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 455, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each pixel of $\\mathbf{I}^{\\mathrm{d}}$ by using the degradation prior $\\mathbf{P}^{\\mathrm{emb}}$ ; ii) top-K expert restoration, selecting expert/parameters from a trainable candidate list of convolution filters according to the degradation map $\\mathbf{M}$ , and convolving with $\\mathbf{X}$ to get intermediate restoration features $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ ; iii) restoration feature aggregation, deriving $\\hat{\\mathbf{X}}$ by pixel-wisely aggregating $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ with respect to the degradation map $\\mathbf{M}$ , and improving the feature locality with a feedforward network. Finally, $\\hat{\\mathbf{X}}$ is decoded to get the restored image $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ .", + "bbox": [ + 75, + 271, + 470, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Degradation Prior", + "text_level": 1, + "bbox": [ + 76, + 420, + 256, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Degradation Prior Generation. We leverage the context learning capability of the PVL model (e.g., [16]) to reason diverse degradation knowledge of the degraded image $\\mathbf{I}^{\\mathrm{d}}$ with a question prompt $\\mathbf{T}$ . We format the question prompt $\\mathbf{T}$ inspired by the chain-of-thought reasoning that makes the model to identify the occurrence, type, and severity of degradation. In Fig. 4, we provide the text description examples obtained by using different prompts. The generated descriptions $\\mathbf{P}^{\\mathrm{txt}}$ are defined as", + "bbox": [ + 75, + 445, + 472, + 582 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {\\mathrm {t x t}} = \\operatorname {V L} \\left(\\mathbf {I} ^ {\\mathrm {c}}, \\mathbf {T}\\right), \\quad \\mathbf {P} ^ {\\mathrm {t x t}} \\in \\mathbb {R} ^ {L \\times C ^ {\\mathrm {v l}}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 595, + 468, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{VL}(\\cdot ,\\cdot)$ is the PVL model, $L$ is the description length, and $C^{\\mathrm{vl}}$ is the channel dimension. To preserve the representation capabilities of PVL, we prune out the text description output layer of PVL, and use the embedding before the output layer as $\\mathbf{P}^{\\mathrm{xt}}$ .", + "bbox": [ + 75, + 628, + 468, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Degradation Prior Embedding. We align $\\mathbf{P}^{\\mathrm{txt}}$ to the embedding space with size $C$ of our restoration model by using a multilayer perceptron network MLP( $\\cdot$ ). We have", + "bbox": [ + 75, + 710, + 468, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} ^ {\\mathrm {e m b}} = \\operatorname {M L P} \\left(\\mathbf {P} ^ {\\mathrm {t x t}}\\right), \\quad \\mathbf {P} ^ {\\mathrm {e m b}} \\in \\mathbb {R} ^ {L \\times C}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 768, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathbf{P}^{\\mathrm{emb}}$ is then used as the degradation prior in our restoration model to adaptively recover $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ .", + "bbox": [ + 75, + 801, + 468, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Language-driven Restoration Model", + "text_level": 1, + "bbox": [ + 76, + 845, + 393, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use an encoder-decoder architecture [8] as a backbone, and adaptively recover $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ in the embedding space with $\\mathbf{P}^{\\mathrm{emb}}$ and $\\mathbf{X}$ (the encoded embedding of $\\mathbf{I}^{\\mathrm{d}}$ ) as inputs.", + "bbox": [ + 75, + 869, + 468, + 916 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98708d37090c1b36d206cd8a414ffca3b0bdd2d88fb505e013bcb077b684a235.jpg", + "image_caption": [ + "Figure 5. Visualization of $\\mathbf{X}$ , $\\mathbf{M}$ , $\\hat{\\mathbf{X}}^{int}$ and $\\hat{\\mathbf{X}}$ . Without cherry-picking, we show the channel with maximum gradients. From (b) $\\mathbf{X}$ to (c) $\\mathbf{M}$ , degraded regions are highlighted, and regions with similar degradation severity have similar activation values on the feature map $\\mathbf{M}$ . From (b) $\\mathbf{X}$ to (d) $\\hat{\\mathbf{X}}^{int}$ and (e) $\\hat{\\mathbf{X}}$ , degradations are removed step by step." + ], + "image_footnote": [], + "bbox": [ + 509, + 89, + 880, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Degradation Map Measurement. The degradation prior $\\mathbf{P}^{\\mathrm{emb}}$ is text description related with high-level semantics. However, for $\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times C}$ with height $H$ , width $W$ , and channel $C$ , the degradation is usually pixel-wise and distinct in severity. To allow pixel-wise computation of adverse weather removal, we measure the 2D degradation map $\\mathbf{M}$ for $\\mathbf{X}$ by the cross-attention mechanism,", + "bbox": [ + 496, + 338, + 892, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {\\mathrm {q} _ {1}}, \\quad \\mathbf {K} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {k} _ {1}}, \\quad \\mathbf {V} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {v} _ {1}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 455, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top}\\right) \\mathbf {V}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 476, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}^{\\mathrm{q}_1}$ , $\\mathbf{W}^{\\mathrm{k}_1}$ , and $\\mathbf{W}^{\\mathrm{v}_1} \\in \\mathbb{R}^{C \\times C}$ are linear projection matrices for obtaining the query, key, and value. Our cross-attention mechanism first computes the semantic alignments between $\\mathbf{X}$ and $\\mathbf{P}^{\\mathrm{emb}}$ via Softmax $(\\mathbf{QK}^\\top)$ , and then transforms $\\mathbf{P}^{\\mathrm{emb}}$ into a 2D degradation map $\\mathbf{M}$ that describes pixel-wise degradation of $\\mathbf{X}$ . We visualize $\\mathbf{X}$ and $\\mathbf{M}$ in Fig. 5. The regions of similar degradation severity have been individually highlighted, by applying Eq. (4).", + "bbox": [ + 496, + 506, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Top-K Expert Restoration. Not all pixels degrade equally, and we adaptively restore each of them by using pixel-wise degradation knowledge within $\\mathbf{M}$ . We have $N$ candidate experts (convolution filters), $\\{\\mathbf{F}_n|n = 1,\\dots,N\\}$ , for diverse adverse weather conditions, where each candidate $\\mathbf{F}_n$ is expertise at generating restoration features for specific degradation types or severities (see 4.2). With the degradation map $\\mathbf{M}$ , we select experts for each pixel of $\\mathbf{X}$ .", + "bbox": [ + 496, + 628, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the degradation map $\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times C}$ is fed to a feedforward network $\\mathrm{FFN}^{\\mathrm{m}}$ , followed by a Softmax layer along the last dimension, to output a normalized pixel-wise selection score $\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times N}$ , i.e., $\\mathbf{S} = \\operatorname{Softmax}(\\mathrm{FFN}^{\\mathrm{m}}(\\mathbf{M}))$ . For a pixel with location $(i,j), i \\in [1,H], j \\in [1,W]$ , $\\mathbf{S}(i,j) \\in \\mathbb{R}^N$ and $\\mathbf{S}(i,j,n)$ measures how likely the $n$ -th expert $\\mathbf{F}_n$ can be used to describe the degradation of $\\mathbf{X}(i,j)$ .", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To find the best experts describing the degradation of $\\mathbf{X}(i,j)$ , we compute the Top-K scores of $\\mathbf{S}(i,j)$ and record", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24905", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2a7f44d54aba1aaa8e8485e844850426bacf18577e8529aca9fcf86ebccbb85e.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on the All-weather dataset. We respectively color the best and the second-best methods in red and blue." + ], + "table_footnote": [], + "table_body": "
TypeMethodRainSnowRaindropAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
BestT + VL24.330.86028.470.87229.230.89527.340.876
BestT + GT27.040.91330.610.90031.630.93629.760.916
GeneralMPRNet [51]23.080.83927.690.84928.750.87926.510.856
NAFNet [3]23.210.84027.680.84728.900.89026.600.859
Uformer [45]22.930.83527.500.83828.510.87126.310.848
Restormer [52]23.370.84527.810.85029.100.89026.760.862
GRL [23]23.310.84227.790.84929.050.88826.720.860
All-in-OneAll-in-One [21]24.710.89828.330.88231.120.92728.050.902
AirNet [17]23.120.83727.920.85828.230.89226.420.862
TUM [5]23.920.85529.270.88430.750.91227.980.884
Transweather [42]23.180.84127.800.85428.980.90226.650.866
WDiff [32]26.180.90729.690.89329.710.91128.530.904
WGWS [57]25.310.90129.710.89431.310.93228.780.909
Ours26.920.91230.790.90531.540.93329.750.916
", + "bbox": [ + 78, + 103, + 928, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding indices as the best experts. The best experts restored feature for location $(i,j)$ is given by,", + "bbox": [ + 75, + 356, + 468, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} ^ {\\text {i n t}} (i, j) = \\sum_ {k = 1} ^ {K} \\mathbf {S} (i, j, \\rho (k)) \\cdot \\mathcal {E} (i, j, \\rho (k)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 393, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} (i, j, \\rho (k)) = \\sum_ {\\Delta i, \\Delta j} \\mathbf {X} (i + \\Delta i, j + \\Delta j) \\cdot \\mathbf {F} _ {\\rho (k)} (\\Delta i, \\Delta j),\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 439, + 467, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho (k)$ is the index of the selected $k$ -th expert. $\\mathcal{E}(i,j,\\rho (k))$ is the convolution result of $\\mathbf{X}(i,j)$ and the $k$ -th expert $\\mathbf{F}_{\\rho (k)}$ , and $(\\Delta i,\\Delta j)$ iterates over the convolution kernel size. $\\mathbf{S}(i,j,\\rho (k))$ are weights for prioritizing different experts. With our pixel-wise Top-K experts, the intermediate restoration feature $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ for each pixel of $\\mathbf{X}$ is adaptively generated. Practically, we find the best performance by region-wisely smoothing the expert selection score $\\mathbf{S}$ , as pixels of a region frequently suffer the same degradation.", + "bbox": [ + 75, + 479, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Restoration Feature Aggregation. Pixels with similar degradation information in $\\mathbf{M}$ potentially benefit each other in deriving the restoration $\\hat{\\mathbf{X}}$ . We compute the compatibility between degradation prior $\\mathbf{M}$ and restoration features $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ , and aggregate weighted restoration features pixel-wisely by using the cross-attention mechanism,", + "bbox": [ + 75, + 616, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {M} \\mathbf {W} ^ {\\mathrm {q} _ {2}}, \\quad \\mathbf {K} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {k} _ {2}}, \\quad \\mathbf {V} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {v} _ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 713, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} = \\operatorname {F F N} ^ {\\operatorname {i n t}} (\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V})) ， \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 734, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{FFN}^{\\mathrm{int}}$ is a feedforward convolutional network to improve the feature locality. Finally, $\\hat{\\mathbf{X}}$ is decoded to the restoration $\\hat{\\mathbf{I}}^{\\mathrm{c}}$ . In Fig. 5, we compare $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ and $\\hat{\\mathbf{X}}$ , and find that $\\hat{\\mathbf{X}}$ is cleaner than $\\hat{\\mathbf{X}}^{\\mathrm{int}}$ , as pixels with similar degradation measurements benefit each other in the restorations.", + "bbox": [ + 75, + 763, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 76, + 847, + 220, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our network with Charbonnier loss $\\mathcal{L}_{\\mathrm{char}}$ and gradient-level edge loss $\\mathcal{L}_{\\mathrm{edge}}$ , to penalize the deviation of", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "restoration from ground truth clean image, and encourage the consistent image gradients. We have $\\mathcal{L}_{\\mathrm{char}}$ as", + "bbox": [ + 498, + 356, + 890, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\left\\| \\mathbf {I} ^ {\\mathrm {c}} - \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 398, + 890, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varepsilon = 10^{-4}$ is a constant in all experiments. The $\\mathcal{L}_{\\mathrm{edge}}$ is", + "bbox": [ + 498, + 428, + 888, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e d g e}} = \\sqrt {\\| \\nabla \\mathbf {I} ^ {\\mathrm {c}} - \\nabla \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 455, + 890, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla$ is the laplacian gradient operator. With a balance parameter $\\lambda$ , the total loss is given by", + "bbox": [ + 498, + 487, + 888, + 517 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a r}} + \\lambda \\mathcal {L} _ {\\text {e d g e}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 531, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments and Analysis", + "text_level": 1, + "bbox": [ + 498, + 566, + 741, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our model is implemented using the PyTorch framework and all experiments are conducted on an RTX A6000 GPU. We train our network with batch size 4 and $4 \\times 10^{6}$ iterations using the ADAM optimizer. The learning rate is decayed from $2 \\times 10^{-4}$ to $1 \\times 10^{-6}$ by cosine annealing strategy. In training, images are randomly cropped to size $256 \\times 256$ , and $\\lambda = 0.05$ . Code and more results are in our project page.", + "bbox": [ + 496, + 592, + 890, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our experiments are conducted on both synthetic dataset and real dataset, i.e., All-weather dataset [21] and WeatherStream [53] dataset. All-weather dataset comprises 18,609 training images and 17,609 testing images, and is composed of subsets from Outdoor-rain [20], Snow100K-L [26], and Raindrop [37], corresponding to rain, snow, and raindrops weather conditions respectively. WeatherStream dataset is a real-world dataset that has three weather conditions, i.e., rain, snow, and fog, with a total of 176, 100 training images and 11, 400 testing images.", + "bbox": [ + 496, + 715, + 892, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use averages of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation metrics (RGB-channel). The higher, the better.", + "bbox": [ + 496, + 869, + 890, + 915 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24906", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/329c0e3798bdeff94f76b7819b5e17286b09bfa6c8352bbcb1d3e4bc5862775c.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the WeatherStream dataset. We color the best and the second-best methods in red and blue." + ], + "table_footnote": [], + "table_body": "
TypeMethodRainHazeSnowAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
GeneralBestT + VL21.200.78121.600.75520.320.77221.040.769
BestT + GT23.950.81022.970.80422.700.82823.210.814
MPRNet [51]21.500.79121.730.76320.740.80121.320.785
NAFNet [3]23.010.80322.200.80322.110.82622.440.811
Uformer [45]22.250.79118.810.76320.940.80120.670.785
Restormer [52]23.670.80422.900.80322.510.82822.860.812
GRL [23]23.750.80522.880.80222.590.82923.070.812
All-in-OneAll-in-One [21]--------
AirNet [17]22.520.79721.560.77021.440.81221.840.793
TUM [5]23.220.79522.380.80522.250.82722.620.809
Transweather [42]22.210.77222.550.77421.790.79222.180.779
WGWS [57]23.800.80722.780.80022.720.83123.100.813
Ours24.420.81823.110.80923.120.83823.550.822
", + "bbox": [ + 78, + 103, + 897, + 321 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d9bc51965978840ef6d5582de78dbdfd68821b87cb081524193373a2138b83b.jpg", + "image_caption": [ + "Rain" + ], + "image_footnote": [], + "bbox": [ + 173, + 324, + 377, + 430 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bef62892f913d7239b3cb214139f16b73a560daaf49d6bf5def2cd37d60e6754.jpg", + "image_caption": [ + "Snow" + ], + "image_footnote": [], + "bbox": [ + 173, + 440, + 375, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e18c2860f6e47e99f493a0340db23e531ae53cc74a923512b5046ac21a1a223b.jpg", + "image_caption": [ + "Raindrop", + "Figure 6. Qualitative comparison on the All-weather dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + ], + "image_footnote": [], + "bbox": [ + 173, + 556, + 375, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3fd8796e1dc1d017400df17e2d48566b0c508e2ce3d0acf0f47f0ead1595ac10.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 388, + 325, + 480, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf982d708a9dba1011556f472890c0df3e7d3f92c55724f4e09b52e517f09347.jpg", + "image_caption": [ + "TUM [5]" + ], + "image_footnote": [], + "bbox": [ + 388, + 383, + 480, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bdcd9c91803fdd4afae70a3adc334e19270d035e86aedc912e0575983cbab8ca.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 493, + 325, + 584, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75cc649564af012dfe80bf2e1d2da3af3779010ca756477e6f2839f660467a38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 382, + 584, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/450cf566e3cfed04bd508527fee1e39efe2373c25fa380d80346860aea569b92.jpg", + "image_caption": [ + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 596, + 325, + 687, + 371 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf36925a2f00b9bd48e83404f1a227cd5bf7fd93427fc75499caf650f7458df7.jpg", + "image_caption": [ + "WGWS [57]" + ], + "image_footnote": [], + "bbox": [ + 596, + 382, + 687, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25b68cd19a746c4e08df840fe538f500d5c1b0e9826ab271c3bf389b1103f0ff.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 388, + 441, + 480, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1000766bd9a8590ed0282a55c4fdcd7b5fc369d43415bedaa9c01135ca9ca5dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 500, + 480, + 544 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccaebe707ed23737ab05d91c3b7e6a34a80c3b5942fc088264f2a418c3877ee2.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 493, + 441, + 584, + 488 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3d79301cebac642b483332aeb383ea71b14080ac7bd598ef6e891ad6dcf204d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 500, + 584, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da9847f09628e6d8346b52afeae7798b8a183a6ef55e277982a62c5a0c5fea2d.jpg", + "image_caption": [ + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 596, + 441, + 687, + 487 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/00f47c18f4b63113f207ae7feb128b414123411276deca765b2138d3d88c1a6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 500, + 687, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4478250649f099786679deb6dfdb9dc02f61aff497d06ff8d11c45092a59a876.jpg", + "image_caption": [ + "TUM [5]", + "Input" + ], + "image_footnote": [], + "bbox": [ + 388, + 559, + 480, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8edc7f0add0e9455b57aa25d4642b4865743d83b82d3434ffed475e56fe9e990.jpg", + "image_caption": [ + "TUM [5]" + ], + "image_footnote": [], + "bbox": [ + 388, + 616, + 480, + 661 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4853ebd2bc6f24f0bb634480920e6ffc484ac7a1ffb6720a5960ba6f9b8266c.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 493, + 559, + 584, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a56d06adbb5f28e51f7e092408bd41cfcbe1738ac9da525f2a8ad64f85b5ee4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 616, + 584, + 661 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f5c895c271e6f9bf64865cc63f5d952f3ebdf99994e7eeacfc6892e9ae1b5f4.jpg", + "image_caption": [ + "WGWS [57]", + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 596, + 559, + 687, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8e8553e5c9759111372d02df100f4d133f0e24da9f15e6eb655a8477cb12a273.jpg", + "image_caption": [ + "WGWS [57]" + ], + "image_footnote": [], + "bbox": [ + 596, + 616, + 687, + 661 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e91399c81ad3a95d8f04937690cb4b62344a9365acce3e0af01aa256ccc3069.jpg", + "image_caption": [ + "AirNet [17]" + ], + "image_footnote": [], + "bbox": [ + 699, + 325, + 792, + 369 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/746421e067a8deb140344d20efab908d875e66d1b9fb89e35b09a830861cdf0a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 383, + 792, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e75a300df17f845fae0c9a91d5e942f752f9a60c9743d58fa69d1af3337803c7.jpg", + "image_caption": [ + "Ours", + "AirNet [17]" + ], + "image_footnote": [], + "bbox": [ + 699, + 441, + 792, + 487 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86f487730909b5543e1da0353cd708a37362ec2dd9dfb63b7c85bb757bebd880.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 699, + 500, + 792, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4052a061315f926ed695ff6dc611c1e89d1a907e90d706c9087197f6846a6015.jpg", + "image_caption": [ + "Ours", + "AirNet [17]" + ], + "image_footnote": [], + "bbox": [ + 699, + 556, + 792, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ff53dd5f6d2f745eb867376cdd8908c8a38ddc6a01f17b59a7b4bffdd975abb.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 699, + 616, + 792, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. We compare with the state-of-the-art (SOTA) general and all-in-one methods. For general methods, task-specific methods are trained with multi-task learning of different weather conditions, i.e., MPRNet [51], NAFNet [3], Uformer[45], Restormer [52], and GRL [23]. All-in-one methods are All-in-One [21], AirNet[17], TUM [5], Transweather [42], WDiff [32] and WGWS [57].", + "bbox": [ + 75, + 728, + 468, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 277, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We compare with the SOTA general and all-in-one methods on the All-weather and", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- WeatherStream datasets in Tab. 1 and Tab. 2, respectively.\n- Our method achieves the best performance.", + "bbox": [ + 498, + 728, + 888, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we create two strong baselines: i) BestT + VL. We use a PVL model for zero-shot adverse weather classification, and select the best task-specific methods for restoring degraded images; ii) BestT + GT. The ground truth adverse weather type is used to select the best task-specific method for image restoration. We find that i) Our method significantly outperforms 'BestT + VL', though it uses the same PVL as our model and uses multiple models for different adverse weather conditions. This shows", + "bbox": [ + 496, + 763, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24907", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d16cb43e6bccc291ac61bb5942426b177579885b973c424d584a6c6f7aa08a42.jpg", + "image_caption": [ + "Haze" + ], + "image_footnote": [], + "bbox": [ + 169, + 89, + 372, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/117e8ef003f55cd5ed0710c3026996cbce48fecef239179c657f42e708570a45.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 477, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1f1a0967a242331f1bfb152d6738d4b6476e7de88c13a477e2752bd608f2653.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 90, + 581, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66e9655765fce5aace00d58e23c59b0b7069d561c1d9edb7bd011ee5eb83f77c.jpg", + "image_caption": [ + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 591, + 90, + 684, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b574b23598f722843db316a46317228e281c65e103f4adef20f0eef5d4069e5.jpg", + "image_caption": [ + "AirNet 17" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 789, + 137 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e1f8b717c1bb58edbfc0f078a414bccbc04a35d323065307500989f79644c4c.jpg", + "image_caption": [ + "Rain" + ], + "image_footnote": [], + "bbox": [ + 169, + 208, + 372, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc4e924bce113b90da8e98f486fd9873d4848c4d5a24ce2bcb34b182346f5290.jpg", + "image_caption": [ + "TUM [5]" + ], + "image_footnote": [], + "bbox": [ + 385, + 148, + 477, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/296523e3f89a68c56fc1bff2d6f1e1e246a85d9706eb3e873c1b510cc76820e3.jpg", + "image_caption": [ + "GT", + "transweather", + "[42] WGWS [57]" + ], + "image_footnote": [], + "bbox": [ + 485, + 148, + 581, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c79966778d844b3c3d61e0458725134e979be1a5f021c399db8facd3c38fd8d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 148, + 684, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e656e7b8c5b173f97b82aefc7a6af339ce182c16479540294d4dbb1a5110a73.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 148, + 789, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/825729df2f225578c745747c97d0e4c146a41a8ff19e94f1a02b7a0e33398154.jpg", + "image_caption": [ + "Snow", + "Figure 7. Qualitative comparison on the WeatherStream dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + ], + "image_footnote": [], + "bbox": [ + 169, + 324, + 372, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff7d85d49408d93081c2356ebdd41ed8752765f4340cd1b4e1708e92b32f9a99.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 385, + 208, + 477, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82fcea633d4de1827e5d2e2e975514d96acad55762e73947c7ad70e0664a44b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 208, + 789, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AirNet 117", + "bbox": [ + 712, + 253, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/223690d0cd8d9afcaef9ab4c9a7354ee9ca743f022a1c5806b7cbb95578a8b5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 263, + 477, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96937a15dfac9072b01a2b68630f51fb229260583fb992ab30e92f3461f8b84e.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 488, + 209, + 581, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89cff17d97eabdab1cc0cf6b352129ce062332911e58775e271664ecb92ac474.jpg", + "image_caption": [ + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 593, + 209, + 684, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 697, + 209, + 789, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AirNet [17]", + "bbox": [ + 712, + 253, + 785, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/044a9834d14a234dab52255373b6f849874d9e4387debfa9243731b198e10cfb.jpg", + "image_caption": [ + "TUM [5]", + "Input" + ], + "image_footnote": [], + "bbox": [ + 385, + 324, + 477, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/042240e7dfc82474d53059be6ad0a5fdcbb627c00b4cdd2871f14d8a579bfe36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 265, + 581, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c3c9b6392469622c85812006190f80acdc6de6c1d12779ae2cdb87f964a67ce.jpg", + "image_caption": [ + "] WGWS [57]" + ], + "image_footnote": [], + "bbox": [ + 593, + 265, + 684, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60a7727666d5dfdec14fde0859b741c9f8717fed0cf90d3afb8635bd922e2d62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 265, + 789, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 385, + 382, + 477, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TUM [5]", + "bbox": [ + 401, + 430, + 462, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddb873dbd66972312aabf83a658c35bd06a80e6dc637cd81696a6702fcaf6df8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 488, + 324, + 581, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ecd56a84ef379a9e40e1f98acc661a41ded2951aaaff466e045e3b4ef852a35.jpg", + "image_caption": [ + "GRL [23]" + ], + "image_footnote": [], + "bbox": [ + 593, + 324, + 684, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/631df5213e5f17da325f08671401fd4a5c3c3adb9dc39e4ef203ae2e40fd6917.jpg", + "image_caption": [ + "Ours", + "AirNet [17]" + ], + "image_footnote": [], + "bbox": [ + 697, + 324, + 789, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 385, + 382, + 477, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TUM [5]", + "bbox": [ + 401, + 430, + 462, + 441 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fb2c4524ac119102ad155ecb9172af9d454965ddf05f27de3656ecbd4bf7290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 382, + 581, + 429 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/53b169759a1bef939633ee0e16174aa93dcac9561002e2ae8911593b5d7672b7.jpg", + "image_caption": [ + "] WGWS [57]" + ], + "image_footnote": [], + "bbox": [ + 591, + 382, + 684, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51119fccec0d7a82626f46083a1398739ff3386cb10ec091afead246a011884c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 382, + 789, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that non-trivial designs are required for effectively leveraging the PVL model in all-in-one adverse weather removal; ii) Although 'BestT + GT' outperforms most SOTA general and all-in-one methods, our method still achieves competitive performance compared to it. This indicates: 1) the benefit of accessing prior knowledge of adverse weather conditions, and 2) our method has learned the shared and weather-specific knowledge adaptively by selecting sparse experts dynamically.", + "bbox": [ + 75, + 493, + 467, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. We compare with the SOTA methods on restoring images degraded by different adverse weather conditions. The results are given in Fig. 6 and Fig. 7. Our method consistently restores clearer images than SOTA methods under different weather conditions.", + "bbox": [ + 75, + 633, + 467, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies and Discussions", + "text_level": 1, + "bbox": [ + 76, + 719, + 362, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the effectiveness and components of our framework on the All-weather dataset.", + "bbox": [ + 75, + 743, + 467, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Architecture. We study the effectiveness of the degradation map measurement (DMM), top-K expert restoration (TER), and restoration feature aggregation (RFA) modules. The results are given in Tab. 3. The best performance is achieved by using all modules.", + "bbox": [ + 75, + 776, + 467, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Degradation Severity. We use the PVL model to partition the All-weather dataset [21] into three subsets based on degradation severity: slight, moderate, and heavy. We", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7b47ae697cb6bb9483d04fe98d9016042a8c6192d05c7dc43cb57af609e6e56.jpg", + "table_caption": [ + "Table 3. The effectiveness of our model components." + ], + "table_footnote": [], + "table_body": "
DMMTERRFAPSNRSSIM
X✓X27.930.882
XX✓28.370.889
✓✓X28.250.890
✓X✓29.110.902
X✓✓28.550.895
✓✓✓29.750.916
", + "bbox": [ + 504, + 506, + 888, + 609 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/89a71fb760d52f7ff9454655c41dadc4539f504557a5b0f5642be3427b3bdb65.jpg", + "table_caption": [ + "Table 4. Comparisons on the All-weather dataset with degradation severity of slight, moderate, and heavy." + ], + "table_footnote": [], + "table_body": "
MethodSlightModerateHeavy
PSNRSSIMPSNRSSIMPSNRSSIM
AirNet27.590.89526.490.86524.500.818
WGWS29.740.92128.770.91027.140.886
Ours30.460.92529.780.91828.380.899
", + "bbox": [ + 501, + 652, + 890, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "then compare our results with those of AirNet and WGWS in Tab. 4. Our method shows significant improvement over AirNet [17] and WGWS [57] in the heavily degraded subset. This indicates the effectiveness of reasoning with diverse weather-specific knowledge, such as severity.", + "bbox": [ + 496, + 750, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pixel-wise Expert. Different regions of the same degraded image often exhibit varying degrees of degradation severity, and should be adaptively and pixel-wisely restored. We use the degradation prior provided by the PVL model to select the same expert for the degraded image. It shows a", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24908", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a066f61434721720a99a5df10ea3fce47225b5925c783fc1282b549410b910e.jpg", + "image_caption": [ + "Figure 8. The usage frequencies of experts for different degradation types. For example, Expert 8, 11, and 15 focuses on raindrop, rain, and snow, respectively. Expert 12 handles all types of degradation. Please refer to Fig. 9 for visualization." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 890, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f6a36680882ed0072f97a1328e6badd79ca2ffe41cf1878709f4de02cb3de923.jpg", + "image_caption": [ + "(a) Expert 8" + ], + "image_footnote": [], + "bbox": [ + 81, + 246, + 250, + 376 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97439ccd3be4d9464249385dcc176b90590c4284faf519b0290333e19af9a208.jpg", + "image_caption": [ + "(b) Expert 11" + ], + "image_footnote": [], + "bbox": [ + 295, + 246, + 460, + 376 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d893fe87b443140c9cd2a7f5fcf598724031af3f844b7ab110efb7775dc138d2.jpg", + "image_caption": [ + "(c) Expert 12" + ], + "image_footnote": [], + "bbox": [ + 508, + 246, + 674, + 376 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e06a8ab188f3963f7618fb7b9a0b492a5655346cc8e9c218a345824d6e16d62a.jpg", + "image_caption": [ + "(d) Expert 15" + ], + "image_footnote": [], + "bbox": [ + 720, + 246, + 888, + 376 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/929219944645be22c84fe08df0039f92b9cac4ec21d2cd17fdad3ab2efafb640.jpg", + "image_caption": [ + "Figure 9. Sample image regions, activating different experts. We show regions with highest selection scores for experts 8, 11, 12, and 15 on the All-weather dataset.", + "(a) Degraded", + "(b) AirNet", + "(c) WGWS", + "(d) Ours", + "Figure 10. From the first to last rows, the degradation is rain, raindrop, snow, and haze, respectively. From left to right, (a) degraded images are restored by (b) AirNet, (c) WGWS, and (d) our method." + ], + "image_footnote": [], + "bbox": [ + 88, + 443, + 455, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 272, + 443, + 361, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 362, + 443, + 454, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "0.33 dB/0.07 decrease in PSNR/SSIM, indicating the necessity of selecting experts pixel-wisely.", + "bbox": [ + 75, + 712, + 468, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Adverse Weather Removal. We study the model response to different adverse weather degradation by measuring the usage of our experts. Fig. 8 shows the usage for rain, snow, and raindrop. We find that there are experts for weather-specific knowledge and shared knowledge among degradation, such as Expert 8, Expert 11, Expert 12, and Expert 15. We visualize the image patches selecting the four experts in Fig. 9, where Expert 8, Expert 11, and Expert 15 are selected by image regions with raindrop, rain, and snow degradation, and Expert 12 is selected by image", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "regions with sky in all types of degradation.", + "bbox": [ + 500, + 445, + 790, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model Ability Under Complex Weather Condition. We test our model trained on the All-weather dataset with real images degraded by rain, raindrop, snow, and haze, as shown in Fig. 10. We compare our results with the two most competitive methods, AirNet and WGWS. Our model successfully disentangles weather-specific knowledge and generalizes to restore images degraded by haze.", + "bbox": [ + 498, + 469, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Broader Impact", + "text_level": 1, + "bbox": [ + 500, + 595, + 794, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed an LDR framework that adaptively removes various adverse weather conditions in an all-in-one solution. Our key insight is to leverage a pre-trained vision-language model to reason diverse weather-specific knowledge in a degraded image. We then use this knowledge to restore a clean image with three modules: degradation map measurement, Top-K expert restoration, and restoration feature aggregation. Experiments on standard benchmark datasets demonstrate that our method outperforms past works by a large margin.", + "bbox": [ + 496, + 622, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact. Our method is promising to be developed as an image restoration foundation model, prompting by degradation prior generated by a vision-language model.", + "bbox": [ + 496, + 782, + 893, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045.", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24909", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1", + "[2] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 2", + "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 2, 5, 6", + "[4] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1", + "[5] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 5, 6, 7", + "[6] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1", + "[7] Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, and Yuanzhi Li. Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems, 35:23049-23062, 2022. 3", + "[8] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 4", + "[9] Keval Doshi and Yasin Yilmaz. Multi-task learning for video surveillance with limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3889-3899, 2022. 1", + "[10] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 3", + "[11] Markus Enzweiler and Dariu M Gavrila. A multilevel mixture-of-experts framework for pedestrian classification. IEEE Transactions on Image Processing, 20(10):2967-2979, 2011. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Sergio Escalera, Hugo Jair Escalante, Zhen Lei, Hao Fang, Ajian Liu, and Jun Wan. Surveillance face presentation attack detection challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6360-6370, 2023. 1", + "[13] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[14] Trung Hoang, Haichuan Zhang, Amirsaeed Yazdani, and Vishal Monga. Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1670-1679, 2023. 1", + "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[16] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3, 4", + "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17452-17462, 2022. 1, 2, 3, 5, 6, 7", + "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3", + "[20] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 5", + "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 5, 6, 7", + "[22] Xinjie Li and Huijuan Xu. Meid: mixture-of-experts with internal distillation for long-tailed video recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1451-1459, 2023. 3", + "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24910", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 3", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3", + "[26] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 1, 5", + "[27] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 3", + "[28] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2, 3", + "[29] Hamam Mokayed, Amirhossein Nayebiastaneh, Kanjar De, Stergios Sozos, Olle Hagner, and Bjorn Backe. Nordic vehicle dataset (nvd): Performance of vehicle detectors using newly captured nvd from uav in different snowy weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5313-5321, 2023. 1", + "[30] Lynnette Hui Xian Ng and Kathleen M Carley. Botbuster: Multi-platform bot detection using a mixture of experts. In Proceedings of the International AAAI Conference on Web and Social Media, pages 686-697, 2023. 3", + "[31] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24, 2011. 3", + "[32] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5, 6", + "[33] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1", + "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1", + "[35] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 1, 2, 3", + "[36] Prashant W Patil, Sunil Gupta, Santu Rana, Svetha Venkatesh, and Subrahmanyam Murala. Multi-weather image restoration via domain translation. In Proceedings of" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF International Conference on Computer Vision, pages 21696-21705, 2023. 1", + "[37] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 1, 5", + "[38] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence, pages 11908-11915, 2020. 1", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[40] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 3", + "[41] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 1", + "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 5, 6, 7", + "[43] Yecong Wan, Yuanshuo Cheng, Mingwen Shao, and Jordi González. Image rain removal and illumination enhancement done in one go. Knowledge-Based Systems, 252:109244, 2022. 1", + "[44] Yinglong Wang, Chao Ma, and Jianzhuang Liu. Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2023. 2", + "[45] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5, 6", + "[46] Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17535-17544, 2023. 1", + "[47] Hao Yang, Dongming Zhou, Jinde Cao, and Qian Zhao. Dpnet: Detail-preserving image deraining via learning frequency domain knowledge. Digital Signal Processing, 130: 103740, 2022. 1", + "[48] Hao Yang, Liyuan Pan, Yan Yang, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. arXiv preprint arXiv:2307.09815, 2023. 3", + "[49] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24911", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 3", + "[50] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 1", + "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 5, 6", + "[52] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6", + "[53] Howard Zhang, Yunhao Ba, Ethan Yang, Varan Mehra, Blake Gella, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Alex Wong, and Achuta Kadambi. Weatherstream: Light transport automation of single image deweathering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13499-13509, 2023. 5", + "[54] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2", + "[55] Zhihong Zhang, Yuxiao Cheng, Jinli Suo, Liheng Bian, and Qionghai Dai. Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing, 32:1390-1402, 2023. 1", + "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[57] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 1, 2, 3, 5, 6, 7", + "[58] Simiao Zuo, Qingru Zhang, Chen Liang, Pengcheng He, Tuo Zhao, and Weizhu Chen. Moebert: from bert to mixture-of-experts via importance-guided adaptation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1610–1623, 2022. 3" + ], + "bbox": [ + 78, + 90, + 468, + 823 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24912", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_model.json b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd90a58b0616d8d8e162374c0030c11f4c8fd186 --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_model.json @@ -0,0 +1,3940 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.207, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "Language-driven All-in-one Adverse Weather Removal" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.18, + 0.823, + 0.235 + ], + "angle": 0, + "content": "Hao Yang\\(^{1}\\), Liyuan Pan\\(^{1\\dagger}\\), Yan Yang\\(^{2,3}\\), and Wei Liang\\(^{1}\\) \n\\(^{1}\\)Beijing Institute of Technology \\(^{2}\\)Australian National University \\(^{3}\\)A&F, CSIRO \n{hao.yang, liyuan.pan, liangwei}@bit.edu.cn, {yan.yang}@anu.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.284, + 0.314, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.474, + 0.651 + ], + "angle": 0, + "content": "All-in-one (AiO) frameworks restore various adverse weather degradations with a single set of networks jointly. To handle various weather conditions, an AiO framework is expected to adaptively learn weather-specific knowledge for different degradations and shared knowledge for common patterns. However, existing methods: 1) rely on extra supervision signals, which are usually unknown in real-world applications; 2) employ fixed network structures, which restrict the diversity of weather-specific knowledge. In this paper, we propose a Language-driven Restoration framework (LDR) to alleviate the aforementioned issues. First, we leverage the power of pre-trained vision-language (PVL) models to enrich the diversity of weather-specific knowledge by reasoning about the occurrence, type, and severity of degradation, generating description-based degradation priors. Then, with the guidance of degradation prior, we sparsely select restoration experts from a candidate list dynamically based on a Mixture-of-Experts (MoE) structure. This enables us to adaptively learn the weather-specific and shared knowledge to handle various weather conditions (e.g., unknown or mixed weather). Experiments on extensive restoration scenarios show our superior performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.673, + 0.21, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.47, + 0.834 + ], + "angle": 0, + "content": "Imaging under unpleasant degradation conditions poses challenges for vision-based systems, such as self-driving cars [15, 33, 34, 46] and outdoor surveillance systems [9, 12], which require 24/7 service regardless of adverse weather conditions like rain [6, 13, 47], haze [14, 38, 41], and snow [4, 26, 29]. To meet the safety demand, compared to tackling each type of weather condition with independent models, the joint task process, i.e., all-in-one framework, has a broader application scenario [17, 35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.834, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Formally, degraded images can be modeled as masked additive combination of clean images and degradation residuals [35]. Consequently, several works [5, 21, 42] use" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.268, + 0.85, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Figure 1. Score (PSNR and SSIM) comparisons. We compare our model (red) with the top 2 (blue and green) baselines on benchmark datasets with various weather scenarios. Superscripts besides evaluation metrics are used to differentiate datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.895, + 0.642 + ], + "angle": 0, + "content": "a single network for all degradation types. Though shared knowledge is learned for restoration, they neglect that different degradations still hold different mathematical formulations, e.g., transmission map produced by scattering effect in haze model [41] is unnecessary for raindrop model [37]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.896, + 0.809 + ], + "angle": 0, + "content": "Hence, several works [17, 35, 36, 50, 57] use different sub-networks for weather-specific knowledge learning. However, auxiliary supervisions are required to assign the sub-networks, e.g., degradation types [35] or depth maps [57]. Furthermore, the fixed sub-network architecture of existing methods restricts the diversity of learned weather-specific knowledge and their ability to handle images with various weather conditions [43, 55], such as images degraded by weather severity or conditions that have not been encountered before [1, 37], or mixed weather conditions like snow with haze in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this paper, we question - can we restore the image degraded by various weather conditions by adaptive learning of diverse weather-specific knowledge and shared knowledge, without requiring ground truth weather types and severity data? We answer it by our language-driven all-in-one restoration (LDR) framework in two aspects." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.887, + 0.233, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.21, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.143, + 0.209, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.195, + 0.209, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.248, + 0.185, + 0.261 + ], + "angle": 0, + "content": "(a) 7-th" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.297, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.142, + 0.297, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.195, + 0.297, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.248, + 0.279, + 0.261 + ], + "angle": 0, + "content": "(b) 10-th" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.386, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.142, + 0.385, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.195, + 0.385, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.248, + 0.368, + 0.26 + ], + "angle": 0, + "content": "(c) 54-th" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.089, + 0.475, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.142, + 0.474, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.195, + 0.474, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.248, + 0.456, + 0.26 + ], + "angle": 0, + "content": "(d) Input" + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.089, + 0.563, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.142, + 0.563, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.477, + 0.195, + 0.563, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.47, + 0.248, + 0.645, + 0.261 + ], + "angle": 0, + "content": "(e) Transweather (f) Zero-out" + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.089, + 0.652, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.142, + 0.652, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.195, + 0.652, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.248, + 0.645, + 0.261 + ], + "angle": 0, + "content": "(f) Zero-out" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.091, + 0.787, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.138, + 0.787, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.186, + 0.791, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.248, + 0.777, + 0.261 + ], + "angle": 0, + "content": "(g) Distribution" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.091, + 0.844, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.136, + 0.789, + 0.146 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.147, + 0.789, + 0.157 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.158, + 0.789, + 0.169 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.17, + 0.789, + 0.179 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.18, + 0.789, + 0.189 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.19, + 0.789, + 0.2 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.201, + 0.789, + 0.21 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.211, + 0.789, + 0.22 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.221, + 0.789, + 0.23 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.231, + 0.789, + 0.24 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.241, + 0.789, + 0.25 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.262, + 0.894, + 0.332 + ], + "angle": 0, + "content": "Figure 2. Ineffective computations of Transweather [42]. We find the activity of some parameters in Transweather [42] are degradation dependent, and zero-outing computations for those inactivate parameters barely affect image restoration. (a) 7-th channel, (b) 10-th channel, and (c) 54-th channel of feature maps from Transweather, the brighter, the larger activation value, are activated for degradation of rain, raindrop, and snow. We show the (d) degraded images, (e) restorations from Transweather, and (f) restorations by zero-outing inactive channels. (g) From top to bottom, distribution of \\(\\mathcal{L}_2\\) norm for the three channels across all images in the All-weather dataset [21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.461 + ], + "angle": 0, + "content": "1) The knowledge within the feature space of a pretrained vision-language (PVL) model can benefit various tasks, while its potential in our task is still under exploration. A straightforward way is to use the PVL model as the image degradation classifier. In contrast, we go one step further by reasoning diverse weather-specific knowledge from the feature space of the PVL model beyond the type of weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.469, + 0.597 + ], + "angle": 0, + "content": "We start by formulating a question prompt to query the occurrence, type, and severity of degradation in a degraded image. The obtained degradation prior describes what, where, and severity of degradations in high-level textual semantics. Then, we translate the high-level degradation prior to a 2D degradation map by aligning the prior with the degraded image. This degradation map provides a pixel-wise representation of the diverse knowledge of image degradation from the PVL model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.469, + 0.853 + ], + "angle": 0, + "content": "2) We then unleash the potential for various weather removal with the guidance of degradation maps. Observing model parameters are weather-specific, e.g., the rain-related parameters are usually inactive for unrelated degradations, and zero-outing the computations of unrelated parameters also barely affects the restoration quality. Take TUM [5] and Transweather [42] as examples, the PSNR and SSIM w/wo zero-outing on the All-weather dataset are \\(27.93 / 26.52\\) (dB) and \\(0.883 / 0.864\\), and \\(27.98 / 26.65\\) (dB) and \\(0.884 / 0.866\\). This observation, illustrated in Fig. 2, inspired us to bypass computations for parameters unrelated to the specific weather type and severity during image restoration. With the assistance of MoE structure [28, 40], our LDR framework selects experts dynamically for restoration, therefore, ensuring adaptive learning of weather-specific knowledge which is not limited to a fixed network architecture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Specifically, we maintain a candidate list of restoration experts and utilize the degradation map to sparsely select the most related restoration experts for each degra" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.4 + ], + "angle": 0, + "content": "dition. By applying the selected expert pixel-wisely to restore weather-specific features, we create flexible and degradation-adaptive expert combinations/model architectures for restoration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.507 + ], + "angle": 0, + "content": "Though we can have a preliminary restoration from experts restored feature, considering image regions with similar values from the degradation map tend to benefit restoration features of each other, we re-use the degradation map to aggregate the restoration features, and improve the locality of the obtained restoration by a simple convolutional feedforward network. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.512, + 0.892, + 0.54 + ], + "angle": 0, + "content": "- We present an LDR framework to adaptively remove various adverse weather conditions in an all-in-one solution;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.54, + 0.892, + 0.584 + ], + "angle": 0, + "content": "- We propose a degradation map measurement module for extracting diverse weather-specific knowledge from a pre-trained vision-language model;" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.584, + 0.892, + 0.614 + ], + "angle": 0, + "content": "- We propose a Top-K expert restoration module, sparsely and adaptively computing pixel-wise restoration features." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.512, + 0.892, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.619, + 0.892, + 0.662 + ], + "angle": 0, + "content": "The overall comparison in Fig. 1 shows the superiority of our framework for handling various degradations compared to state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.641, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Adverse Weather Restoration. The field of adverse weather restoration encompasses two distinct approaches: task-specific methods [2, 3, 23, 51, 52] train the model independently for each individual degradation and all-in-one frameworks [5, 17, 21, 32, 35, 42, 44, 54, 57] develop a single unified model capable of handling various adverse weather degradations. A prevalent strategy for the all-in-one framework is to build upon task-specific models by employing multi-task learning techniques [44, 54, 57] or knowledge distillation [5] that consolidates multiple task-specific networks into a single network. Nevertheless, these methods typically rely on fixed computations, and the computation effectiveness is discounted as model parameters are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24903" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.868, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Figure 3. The pipeline of our method. Given the input degraded image \\(\\mathbf{I}^d\\), we aim to recover the clean image \\(\\hat{\\mathbf{I}}^c\\). To tackle various adverse weather conditions, we format a question prompt \\(\\mathbf{T}\\), and query a pre-trained vision-language (PVL) model with \\(\\mathbf{I}^d\\) and \\(\\mathbf{T}\\), to estimate the degradation type and severity. The generated descriptions \\(\\mathbf{P}^{txt}\\) are transformed by a multilayer perceptron (MLP) to get the degradation prior \\(\\mathbf{P}^{emb}\\). Meanwhile, we extract a feature map \\(\\mathbf{X}\\) from the input degraded image \\(\\mathbf{I}^d\\). A degradation map \\(\\mathbf{M}\\) is computed by cross-attending \\(\\mathbf{X}\\) with the prior \\(\\mathbf{P}^{emb}\\), describing pixel-wise degradation pattern. We maintain a trainable candidate list of convolution filters, \\(\\{\\mathbf{F}_n|n = 1,\\dots,N\\}\\), and denote them as experts. We first parse the degradation map \\(\\mathbf{M}\\) into a score map \\(\\mathbf{S}\\) via a feedforward network \\(FFN^m\\) and a softmax layer, and then use \\(\\mathbf{S}\\) to find the best \\(K\\) experts describing the degradation of \\(\\mathbf{X}\\). Best experts are convolved with \\(\\mathbf{X}\\) to generate an intermediate feature map \\(\\hat{\\mathbf{X}}^{int}\\). Finally, a cross-attention layer is used to aggregate \\(\\hat{\\mathbf{X}}^{int}\\) with the guidance of the degradation map \\(\\mathbf{M}\\). We improve the feature locality with a feedforward network \\(FFN^{int}\\), and the output \\(\\hat{\\mathbf{X}}\\) is decoded to the clean image \\(\\hat{\\mathbf{I}}^c\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.456, + 0.408, + 0.471 + ], + "angle": 0, + "content": "separately and degradation-dependently activated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.683 + ], + "angle": 0, + "content": "Conversely, several methods [17, 35, 57] have been developed to overcome these limitations by focusing on learning weather-specific knowledge. They utilize indicators like degradation types [27, 35] or depth maps [57] during training to categorize and direct images to appropriate sub-networks for restoration. Though AirNet [17] learns weather-specific knowledge with contrastive learning. Its fixed sub-network designs can only restore images with certain degradation types, and does not generalize to unseen weather degradations, e.g., rain mixed with haze. Furthermore, these mentioned works overlook the fact that varying levels of degradation severity also warrant adaptively tailored computational processes, for a specific degradation type." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.746 + ], + "angle": 0, + "content": "Diverging from previous methods, we harness PVL models to reason both the type and severity of degradations, enabling adaptive restoration with dynamic sub-networks informed by this rich, weather-specific knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Sparse Mixture of Expert. The concept of expert models [28] is defined as a subset of model parameters/computation, where each expert is specialized in handling distinct aspects of the input data. The sparse mixture of experts usually employs a routing mechanism [40] to dynamically and adaptively forward input to a subset of these experts, bypassing unnecessary and irrelevant computations. Given fixed computation costs, due to the sparsity, this framework can readily scale the number of experts to improve the model capability that has been widely verified" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.892, + 0.516 + ], + "angle": 0, + "content": "in the domain of natural language processing [10, 40, 58] and computer vision [7, 11, 22, 30]. We study a prior derived from a PVL model to dynamically select experts, and adaptively apply expert to restore degraded images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.668 + ], + "angle": 0, + "content": "Vision-language Model. With the release of ChatGPT, remarkable reasoning abilities of large language model (LLM) [16, 18, 25, 31, 39, 56] have been shown. It motivates the vision-language community to transfer the reasoning abilities to visual data. A common pipeline is to project an image to a joint space of LLM [19, 24, 25, 48, 49], and feed the projected image alongside user-provided text for conditional text generation. This paper proposes to leverage the PVL (e.g., [16]) to generate prior of an adverse weather degraded image for adaptive image restoration." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.591, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Overview. Given an input image \\(\\mathbf{I}^{\\mathrm{d}}\\) degraded by adverse weather, we aim to recover the corresponding clean image \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\). Our method adaptively recovers \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\) with degradation prior \\(\\mathbf{P}^{\\mathrm{emb}}\\) obtained from a pre-trained vision-language (PVL) model. The overall architecture is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.856 + ], + "angle": 0, + "content": "To derive degradation prior \\(\\mathbf{P}^{\\mathrm{emb}}\\), we query a PVL model with a question prompt \\(\\mathbf{T}\\) to reason about the occurrence, type, and severity of degradation in the input image \\(\\mathbf{I}^{\\mathrm{d}}\\), outputting \\(\\mathbf{P}^{\\mathrm{txt}}\\). Then a mapping network projects \\(\\mathbf{P}^{\\mathrm{txt}}\\) to \\(\\mathbf{P}^{\\mathrm{emb}}\\). Meanwhile, we encode \\(\\mathbf{I}^{\\mathrm{d}}\\) to \\(\\mathbf{X}\\) by using an encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We adaptively restore \\(\\mathbf{X}\\) into \\(\\hat{\\mathbf{X}}\\) in the embedding space, and decode \\(\\hat{\\mathbf{X}}\\) to a restored image \\(\\hat{\\mathbf{l}}^{\\mathrm{c}}\\), with three steps: i) degradation map measurement, measuring degradation for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24904" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.457, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.227, + 0.47, + 0.255 + ], + "angle": 0, + "content": "Figure 4. Examples text descriptions from LISA [16] for reasoning weather degraded images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.272, + 0.471, + 0.41 + ], + "angle": 0, + "content": "each pixel of \\(\\mathbf{I}^{\\mathrm{d}}\\) by using the degradation prior \\(\\mathbf{P}^{\\mathrm{emb}}\\); ii) top-K expert restoration, selecting expert/parameters from a trainable candidate list of convolution filters according to the degradation map \\(\\mathbf{M}\\), and convolving with \\(\\mathbf{X}\\) to get intermediate restoration features \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\); iii) restoration feature aggregation, deriving \\(\\hat{\\mathbf{X}}\\) by pixel-wisely aggregating \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\) with respect to the degradation map \\(\\mathbf{M}\\), and improving the feature locality with a feedforward network. Finally, \\(\\hat{\\mathbf{X}}\\) is decoded to get the restored image \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.421, + 0.258, + 0.438 + ], + "angle": 0, + "content": "3.1. Degradation Prior" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.583 + ], + "angle": 0, + "content": "Degradation Prior Generation. We leverage the context learning capability of the PVL model (e.g., [16]) to reason diverse degradation knowledge of the degraded image \\(\\mathbf{I}^{\\mathrm{d}}\\) with a question prompt \\(\\mathbf{T}\\). We format the question prompt \\(\\mathbf{T}\\) inspired by the chain-of-thought reasoning that makes the model to identify the occurrence, type, and severity of degradation. In Fig. 4, we provide the text description examples obtained by using different prompts. The generated descriptions \\(\\mathbf{P}^{\\mathrm{txt}}\\) are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.596, + 0.47, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {\\mathrm {t x t}} = \\operatorname {V L} \\left(\\mathbf {I} ^ {\\mathrm {c}}, \\mathbf {T}\\right), \\quad \\mathbf {P} ^ {\\mathrm {t x t}} \\in \\mathbb {R} ^ {L \\times C ^ {\\mathrm {v l}}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.705 + ], + "angle": 0, + "content": "where \\(\\mathrm{VL}(\\cdot ,\\cdot)\\) is the PVL model, \\(L\\) is the description length, and \\(C^{\\mathrm{vl}}\\) is the channel dimension. To preserve the representation capabilities of PVL, we prune out the text description output layer of PVL, and use the embedding before the output layer as \\(\\mathbf{P}^{\\mathrm{xt}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.47, + 0.758 + ], + "angle": 0, + "content": "Degradation Prior Embedding. We align \\(\\mathbf{P}^{\\mathrm{txt}}\\) to the embedding space with size \\(C\\) of our restoration model by using a multilayer perceptron network MLP(\\(\\cdot\\)). We have" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.77, + 0.469, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} ^ {\\mathrm {e m b}} = \\operatorname {M L P} \\left(\\mathbf {P} ^ {\\mathrm {t x t}}\\right), \\quad \\mathbf {P} ^ {\\mathrm {e m b}} \\in \\mathbb {R} ^ {L \\times C}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.803, + 0.47, + 0.834 + ], + "angle": 0, + "content": "The \\(\\mathbf{P}^{\\mathrm{emb}}\\) is then used as the degradation prior in our restoration model to adaptively recover \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.394, + 0.863 + ], + "angle": 0, + "content": "3.2. Language-driven Restoration Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.917 + ], + "angle": 0, + "content": "We use an encoder-decoder architecture [8] as a backbone, and adaptively recover \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\) in the embedding space with \\(\\mathbf{P}^{\\mathrm{emb}}\\) and \\(\\mathbf{X}\\) (the encoded embedding of \\(\\mathbf{I}^{\\mathrm{d}}\\)) as inputs." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.09, + 0.881, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.24, + 0.894, + 0.325 + ], + "angle": 0, + "content": "Figure 5. Visualization of \\(\\mathbf{X}\\), \\(\\mathbf{M}\\), \\(\\hat{\\mathbf{X}}^{int}\\) and \\(\\hat{\\mathbf{X}}\\). Without cherry-picking, we show the channel with maximum gradients. From (b) \\(\\mathbf{X}\\) to (c) \\(\\mathbf{M}\\), degraded regions are highlighted, and regions with similar degradation severity have similar activation values on the feature map \\(\\mathbf{M}\\). From (b) \\(\\mathbf{X}\\) to (d) \\(\\hat{\\mathbf{X}}^{int}\\) and (e) \\(\\hat{\\mathbf{X}}\\), degradations are removed step by step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.893, + 0.446 + ], + "angle": 0, + "content": "Degradation Map Measurement. The degradation prior \\(\\mathbf{P}^{\\mathrm{emb}}\\) is text description related with high-level semantics. However, for \\(\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times C}\\) with height \\(H\\), width \\(W\\), and channel \\(C\\), the degradation is usually pixel-wise and distinct in severity. To allow pixel-wise computation of adverse weather removal, we measure the 2D degradation map \\(\\mathbf{M}\\) for \\(\\mathbf{X}\\) by the cross-attention mechanism," + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.456, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {\\mathrm {q} _ {1}}, \\quad \\mathbf {K} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {k} _ {1}}, \\quad \\mathbf {V} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {v} _ {1}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.477, + 0.892, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top}\\right) \\mathbf {V}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}^{\\mathrm{q}_1}\\), \\(\\mathbf{W}^{\\mathrm{k}_1}\\), and \\(\\mathbf{W}^{\\mathrm{v}_1} \\in \\mathbb{R}^{C \\times C}\\) are linear projection matrices for obtaining the query, key, and value. Our cross-attention mechanism first computes the semantic alignments between \\(\\mathbf{X}\\) and \\(\\mathbf{P}^{\\mathrm{emb}}\\) via Softmax \\((\\mathbf{QK}^\\top)\\), and then transforms \\(\\mathbf{P}^{\\mathrm{emb}}\\) into a 2D degradation map \\(\\mathbf{M}\\) that describes pixel-wise degradation of \\(\\mathbf{X}\\). We visualize \\(\\mathbf{X}\\) and \\(\\mathbf{M}\\) in Fig. 5. The regions of similar degradation severity have been individually highlighted, by applying Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Top-K Expert Restoration. Not all pixels degrade equally, and we adaptively restore each of them by using pixel-wise degradation knowledge within \\(\\mathbf{M}\\). We have \\(N\\) candidate experts (convolution filters), \\(\\{\\mathbf{F}_n|n = 1,\\dots,N\\}\\), for diverse adverse weather conditions, where each candidate \\(\\mathbf{F}_n\\) is expertise at generating restoration features for specific degradation types or severities (see 4.2). With the degradation map \\(\\mathbf{M}\\), we select experts for each pixel of \\(\\mathbf{X}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Specifically, the degradation map \\(\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times C}\\) is fed to a feedforward network \\(\\mathrm{FFN}^{\\mathrm{m}}\\), followed by a Softmax layer along the last dimension, to output a normalized pixel-wise selection score \\(\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times N}\\), i.e., \\(\\mathbf{S} = \\operatorname{Softmax}(\\mathrm{FFN}^{\\mathrm{m}}(\\mathbf{M}))\\). For a pixel with location \\((i,j), i \\in [1,H], j \\in [1,W]\\), \\(\\mathbf{S}(i,j) \\in \\mathbb{R}^N\\) and \\(\\mathbf{S}(i,j,n)\\) measures how likely the \\(n\\)-th expert \\(\\mathbf{F}_n\\) can be used to describe the degradation of \\(\\mathbf{X}(i,j)\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "To find the best experts describing the degradation of \\(\\mathbf{X}(i,j)\\), we compute the Top-K scores of \\(\\mathbf{S}(i,j)\\) and record" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24905" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.079, + 0.09, + 0.89, + 0.103 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the All-weather dataset. We respectively color the best and the second-best methods in red and blue." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.104, + 0.929, + 0.336 + ], + "angle": 0, + "content": "
TypeMethodRainSnowRaindropAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
BestT + VL24.330.86028.470.87229.230.89527.340.876
BestT + GT27.040.91330.610.90031.630.93629.760.916
GeneralMPRNet [51]23.080.83927.690.84928.750.87926.510.856
NAFNet [3]23.210.84027.680.84728.900.89026.600.859
Uformer [45]22.930.83527.500.83828.510.87126.310.848
Restormer [52]23.370.84527.810.85029.100.89026.760.862
GRL [23]23.310.84227.790.84929.050.88826.720.860
All-in-OneAll-in-One [21]24.710.89828.330.88231.120.92728.050.902
AirNet [17]23.120.83727.920.85828.230.89226.420.862
TUM [5]23.920.85529.270.88430.750.91227.980.884
Transweather [42]23.180.84127.800.85428.980.90226.650.866
WDiff [32]26.180.90729.690.89329.710.91128.530.904
WGWS [57]25.310.90129.710.89431.310.93228.780.909
Ours26.920.91230.790.90531.540.93329.750.916
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.388 + ], + "angle": 0, + "content": "corresponding indices as the best experts. The best experts restored feature for location \\((i,j)\\) is given by," + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.395, + 0.47, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} ^ {\\text {i n t}} (i, j) = \\sum_ {k = 1} ^ {K} \\mathbf {S} (i, j, \\rho (k)) \\cdot \\mathcal {E} (i, j, \\rho (k)), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.44, + 0.468, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} (i, j, \\rho (k)) = \\sum_ {\\Delta i, \\Delta j} \\mathbf {X} (i + \\Delta i, j + \\Delta j) \\cdot \\mathbf {F} _ {\\rho (k)} (\\Delta i, \\Delta j),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.617 + ], + "angle": 0, + "content": "where \\(\\rho (k)\\) is the index of the selected \\(k\\)-th expert. \\(\\mathcal{E}(i,j,\\rho (k))\\) is the convolution result of \\(\\mathbf{X}(i,j)\\) and the \\(k\\)-th expert \\(\\mathbf{F}_{\\rho (k)}\\), and \\((\\Delta i,\\Delta j)\\) iterates over the convolution kernel size. \\(\\mathbf{S}(i,j,\\rho (k))\\) are weights for prioritizing different experts. With our pixel-wise Top-K experts, the intermediate restoration feature \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\) for each pixel of \\(\\mathbf{X}\\) is adaptively generated. Practically, we find the best performance by region-wisely smoothing the expert selection score \\(\\mathbf{S}\\), as pixels of a region frequently suffer the same degradation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.469, + 0.706 + ], + "angle": 0, + "content": "Restoration Feature Aggregation. Pixels with similar degradation information in \\(\\mathbf{M}\\) potentially benefit each other in deriving the restoration \\(\\hat{\\mathbf{X}}\\). We compute the compatibility between degradation prior \\(\\mathbf{M}\\) and restoration features \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\), and aggregate weighted restoration features pixel-wisely by using the cross-attention mechanism," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.714, + 0.469, + 0.732 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {M} \\mathbf {W} ^ {\\mathrm {q} _ {2}}, \\quad \\mathbf {K} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {k} _ {2}}, \\quad \\mathbf {V} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {v} _ {2}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.736, + 0.469, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} = \\operatorname {F F N} ^ {\\operatorname {i n t}} (\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V})) ， \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.839 + ], + "angle": 0, + "content": "where \\(\\mathrm{FFN}^{\\mathrm{int}}\\) is a feedforward convolutional network to improve the feature locality. Finally, \\(\\hat{\\mathbf{X}}\\) is decoded to the restoration \\(\\hat{\\mathbf{I}}^{\\mathrm{c}}\\). In Fig. 5, we compare \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\) and \\(\\hat{\\mathbf{X}}\\), and find that \\(\\hat{\\mathbf{X}}\\) is cleaner than \\(\\hat{\\mathbf{X}}^{\\mathrm{int}}\\), as pixels with similar degradation measurements benefit each other in the restorations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.222, + 0.862 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We train our network with Charbonnier loss \\(\\mathcal{L}_{\\mathrm{char}}\\) and gradient-level edge loss \\(\\mathcal{L}_{\\mathrm{edge}}\\), to penalize the deviation of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.892, + 0.388 + ], + "angle": 0, + "content": "restoration from ground truth clean image, and encourage the consistent image gradients. We have \\(\\mathcal{L}_{\\mathrm{char}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.4, + 0.891, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\left\\| \\mathbf {I} ^ {\\mathrm {c}} - \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.429, + 0.89, + 0.456 + ], + "angle": 0, + "content": "where \\(\\varepsilon = 10^{-4}\\) is a constant in all experiments. The \\(\\mathcal{L}_{\\mathrm{edge}}\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.457, + 0.891, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e d g e}} = \\sqrt {\\| \\nabla \\mathbf {I} ^ {\\mathrm {c}} - \\nabla \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.488, + 0.89, + 0.518 + ], + "angle": 0, + "content": "where \\(\\nabla\\) is the laplacian gradient operator. With a balance parameter \\(\\lambda\\), the total loss is given by" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.532, + 0.891, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a r}} + \\lambda \\mathcal {L} _ {\\text {e d g e}}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.567, + 0.743, + 0.584 + ], + "angle": 0, + "content": "4. Experiments and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Implementation Details. Our model is implemented using the PyTorch framework and all experiments are conducted on an RTX A6000 GPU. We train our network with batch size 4 and \\(4 \\times 10^{6}\\) iterations using the ADAM optimizer. The learning rate is decayed from \\(2 \\times 10^{-4}\\) to \\(1 \\times 10^{-6}\\) by cosine annealing strategy. In training, images are randomly cropped to size \\(256 \\times 256\\), and \\(\\lambda = 0.05\\). Code and more results are in our project page." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.717, + 0.893, + 0.868 + ], + "angle": 0, + "content": "Datasets. Our experiments are conducted on both synthetic dataset and real dataset, i.e., All-weather dataset [21] and WeatherStream [53] dataset. All-weather dataset comprises 18,609 training images and 17,609 testing images, and is composed of subsets from Outdoor-rain [20], Snow100K-L [26], and Raindrop [37], corresponding to rain, snow, and raindrops weather conditions respectively. WeatherStream dataset is a real-world dataset that has three weather conditions, i.e., rain, snow, and fog, with a total of 176, 100 training images and 11, 400 testing images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.916 + ], + "angle": 0, + "content": "Evaluation Metrics. We use averages of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation metrics (RGB-channel). The higher, the better." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24906" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.09, + 0.864, + 0.103 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the WeatherStream dataset. We color the best and the second-best methods in red and blue." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.104, + 0.898, + 0.322 + ], + "angle": 0, + "content": "
TypeMethodRainHazeSnowAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
GeneralBestT + VL21.200.78121.600.75520.320.77221.040.769
BestT + GT23.950.81022.970.80422.700.82823.210.814
MPRNet [51]21.500.79121.730.76320.740.80121.320.785
NAFNet [3]23.010.80322.200.80322.110.82622.440.811
Uformer [45]22.250.79118.810.76320.940.80120.670.785
Restormer [52]23.670.80422.900.80322.510.82822.860.812
GRL [23]23.750.80522.880.80222.590.82923.070.812
All-in-OneAll-in-One [21]--------
AirNet [17]22.520.79721.560.77021.440.81221.840.793
TUM [5]23.220.79522.380.80522.250.82722.620.809
Transweather [42]22.210.77222.550.77421.790.79222.180.779
WGWS [57]23.800.80722.780.80022.720.83123.100.813
Ours24.420.81823.110.80923.120.83823.550.822
" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.325, + 0.379, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.43, + 0.295, + 0.441 + ], + "angle": 0, + "content": "Rain" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.441, + 0.377, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.546, + 0.298, + 0.557 + ], + "angle": 0, + "content": "Snow" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.558, + 0.377, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.664, + 0.311, + 0.678 + ], + "angle": 0, + "content": "Raindrop" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.326, + 0.482, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.372, + 0.452, + 0.383 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.384, + 0.482, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.429, + 0.464, + 0.443 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.326, + 0.586, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.372, + 0.551, + 0.383 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.383, + 0.585, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.326, + 0.688, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.611, + 0.372, + 0.676, + 0.384 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.383, + 0.688, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.429, + 0.687, + 0.442 + ], + "angle": 0, + "content": "WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.442, + 0.481, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.488, + 0.451, + 0.5 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.5, + 0.481, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.547, + 0.464, + 0.56 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.442, + 0.585, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.488, + 0.551, + 0.5 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.5, + 0.585, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.442, + 0.688, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.488, + 0.675, + 0.5 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.5, + 0.689, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.547, + 0.687, + 0.56 + ], + "angle": 0, + "content": "WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.56, + 0.481, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.605, + 0.451, + 0.617 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.617, + 0.481, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.663, + 0.464, + 0.678 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.56, + 0.585, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.605, + 0.551, + 0.617 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.617, + 0.585, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.56, + 0.688, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.605, + 0.676, + 0.617 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.617, + 0.689, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.663, + 0.687, + 0.677 + ], + "angle": 0, + "content": "WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.326, + 0.793, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.371, + 0.79, + 0.384 + ], + "angle": 0, + "content": "AirNet [17]" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.384, + 0.793, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.431, + 0.768, + 0.442 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.442, + 0.793, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.488, + 0.789, + 0.5 + ], + "angle": 0, + "content": "AirNet [17]" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.5, + 0.793, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.548, + 0.768, + 0.559 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.558, + 0.793, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.605, + 0.789, + 0.617 + ], + "angle": 0, + "content": "AirNet [17]" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.617, + 0.793, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.665, + 0.768, + 0.676 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.681, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on the All-weather dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Baseline Methods. We compare with the state-of-the-art (SOTA) general and all-in-one methods. For general methods, task-specific methods are trained with multi-task learning of different weather conditions, i.e., MPRNet [51], NAFNet [3], Uformer[45], Restormer [52], and GRL [23]. All-in-one methods are All-in-One [21], AirNet[17], TUM [5], Transweather [42], WDiff [32] and WGWS [57]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.279, + 0.862 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. We compare with the SOTA general and all-in-one methods on the All-weather and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.729, + 0.89, + 0.76 + ], + "angle": 0, + "content": "- WeatherStream datasets in Tab. 1 and Tab. 2, respectively.\n- Our method achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Furthermore, we create two strong baselines: i) BestT + VL. We use a PVL model for zero-shot adverse weather classification, and select the best task-specific methods for restoring degraded images; ii) BestT + GT. The ground truth adverse weather type is used to select the best task-specific method for image restoration. We find that i) Our method significantly outperforms 'BestT + VL', though it uses the same PVL as our model and uses multiple models for different adverse weather conditions. This shows" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.09, + 0.374, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.196, + 0.297, + 0.206 + ], + "angle": 0, + "content": "Haze" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.091, + 0.478, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.138, + 0.452, + 0.15 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.091, + 0.582, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.139, + 0.551, + 0.149 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.091, + 0.686, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.138, + 0.675, + 0.15 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.091, + 0.79, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.138, + 0.787, + 0.149 + ], + "angle": 0, + "content": "AirNet 17" + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.209, + 0.374, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.313, + 0.295, + 0.323 + ], + "angle": 0, + "content": "Rain" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.15, + 0.478, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.196, + 0.464, + 0.207 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.486, + 0.15, + 0.582, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.15, + 0.686, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.15, + 0.79, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.197, + 0.768, + 0.207 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.17, + 0.325, + 0.374, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.432, + 0.298, + 0.443 + ], + "angle": 0, + "content": "Snow" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.209, + 0.478, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.254, + 0.451, + 0.265 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.197, + 0.57, + 0.207 + ], + "angle": 0, + "content": "transweather" + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.197, + 0.685, + 0.208 + ], + "angle": 0, + "content": "[42] WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.209, + 0.79, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.254, + 0.786, + 0.265 + ], + "angle": 0, + "content": "AirNet 117" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.265, + 0.478, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.314, + 0.462, + 0.325 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.21, + 0.582, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.255, + 0.55, + 0.265 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.21, + 0.686, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.254, + 0.676, + 0.266 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.21, + 0.79, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.255, + 0.786, + 0.265 + ], + "angle": 0, + "content": "AirNet [17]" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.325, + 0.478, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.372, + 0.451, + 0.382 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.266, + 0.582, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.266, + 0.686, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.313, + 0.685, + 0.325 + ], + "angle": 0, + "content": "] WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.266, + 0.79, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.315, + 0.768, + 0.325 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.383, + 0.478, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.431, + 0.463, + 0.443 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.325, + 0.582, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.372, + 0.55, + 0.382 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.325, + 0.686, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.372, + 0.676, + 0.383 + ], + "angle": 0, + "content": "GRL [23]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.325, + 0.79, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.372, + 0.786, + 0.382 + ], + "angle": 0, + "content": "AirNet [17]" + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.383, + 0.478, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.431, + 0.463, + 0.443 + ], + "angle": 0, + "content": "TUM [5]" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.383, + 0.582, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.383, + 0.686, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.429, + 0.686, + 0.443 + ], + "angle": 0, + "content": "] WGWS [57]" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.383, + 0.79, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.432, + 0.768, + 0.443 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.446, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on the WeatherStream dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.468, + 0.631 + ], + "angle": 0, + "content": "that non-trivial designs are required for effectively leveraging the PVL model in all-in-one adverse weather removal; ii) Although 'BestT + GT' outperforms most SOTA general and all-in-one methods, our method still achieves competitive performance compared to it. This indicates: 1) the benefit of accessing prior knowledge of adverse weather conditions, and 2) our method has learned the shared and weather-specific knowledge adaptively by selecting sparse experts dynamically." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.468, + 0.709 + ], + "angle": 0, + "content": "Qualitative Comparison. We compare with the SOTA methods on restoring images degraded by different adverse weather conditions. The results are given in Fig. 6 and Fig. 7. Our method consistently restores clearer images than SOTA methods under different weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.364, + 0.735 + ], + "angle": 0, + "content": "4.2. Ablation studies and Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.468, + 0.772 + ], + "angle": 0, + "content": "We validate the effectiveness and components of our framework on the All-weather dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.468, + 0.852 + ], + "angle": 0, + "content": "Model Architecture. We study the effectiveness of the degradation map measurement (DMM), top-K expert restoration (TER), and restoration feature aggregation (RFA) modules. The results are given in Tab. 3. The best performance is achieved by using all modules." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Degradation Severity. We use the PVL model to partition the All-weather dataset [21] into three subsets based on degradation severity: slight, moderate, and heavy. We" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.493, + 0.851, + 0.506 + ], + "angle": 0, + "content": "Table 3. The effectiveness of our model components." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.507, + 0.889, + 0.611 + ], + "angle": 0, + "content": "
DMMTERRFAPSNRSSIM
X✓X27.930.882
XX✓28.370.889
✓✓X28.250.890
✓X✓29.110.902
X✓✓28.550.895
✓✓✓29.750.916
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.625, + 0.891, + 0.652 + ], + "angle": 0, + "content": "Table 4. Comparisons on the All-weather dataset with degradation severity of slight, moderate, and heavy." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.653, + 0.892, + 0.73 + ], + "angle": 0, + "content": "
MethodSlightModerateHeavy
PSNRSSIMPSNRSSIMPSNRSSIM
AirNet27.590.89526.490.86524.500.818
WGWS29.740.92128.770.91027.140.886
Ours30.460.92529.780.91828.380.899
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.892, + 0.825 + ], + "angle": 0, + "content": "then compare our results with those of AirNet and WGWS in Tab. 4. Our method shows significant improvement over AirNet [17] and WGWS [57] in the heavily degraded subset. This indicates the effectiveness of reasoning with diverse weather-specific knowledge, such as severity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pixel-wise Expert. Different regions of the same degraded image often exhibit varying degrees of degradation severity, and should be adaptively and pixel-wisely restored. We use the degradation prior provided by the PVL model to select the same expert for the degraded image. It shows a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24908" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.892, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.24 + ], + "angle": 0, + "content": "Figure 8. The usage frequencies of experts for different degradation types. For example, Expert 8, 11, and 15 focuses on raindrop, rain, and snow, respectively. Expert 12 handles all types of degradation. Please refer to Fig. 9 for visualization." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.247, + 0.25, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.379, + 0.215, + 0.392 + ], + "angle": 0, + "content": "(a) Expert 8" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.247, + 0.462, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.379, + 0.42, + 0.392 + ], + "angle": 0, + "content": "(b) Expert 11" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.247, + 0.676, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.379, + 0.631, + 0.392 + ], + "angle": 0, + "content": "(c) Expert 12" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.247, + 0.89, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.378, + 0.842, + 0.392 + ], + "angle": 0, + "content": "(d) Expert 15" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.397, + 0.895, + 0.425 + ], + "angle": 0, + "content": "Figure 9. Sample image regions, activating different experts. We show regions with highest selection scores for experts 8, 11, 12, and 15 on the All-weather dataset." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.444, + 0.457, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.635, + 0.168, + 0.646 + ], + "angle": 0, + "content": "(a) Degraded" + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.635, + 0.254, + 0.645 + ], + "angle": 0, + "content": "(b) AirNet" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.444, + 0.362, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.635, + 0.347, + 0.645 + ], + "angle": 0, + "content": "(c) WGWS" + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.444, + 0.455, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.635, + 0.431, + 0.645 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.649, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Figure 10. From the first to last rows, the degradation is rain, raindrop, snow, and haze, respectively. From left to right, (a) degraded images are restored by (b) AirNet, (c) WGWS, and (d) our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.47, + 0.744 + ], + "angle": 0, + "content": "0.33 dB/0.07 decrease in PSNR/SSIM, indicating the necessity of selecting experts pixel-wisely." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Adaptive Adverse Weather Removal. We study the model response to different adverse weather degradation by measuring the usage of our experts. Fig. 8 shows the usage for rain, snow, and raindrop. We find that there are experts for weather-specific knowledge and shared knowledge among degradation, such as Expert 8, Expert 11, Expert 12, and Expert 15. We visualize the image patches selecting the four experts in Fig. 9, where Expert 8, Expert 11, and Expert 15 are selected by image regions with raindrop, rain, and snow degradation, and Expert 12 is selected by image" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.446, + 0.792, + 0.462 + ], + "angle": 0, + "content": "regions with sky in all types of degradation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.47, + 0.892, + 0.577 + ], + "angle": 0, + "content": "Model Ability Under Complex Weather Condition. We test our model trained on the All-weather dataset with real images degraded by rain, raindrop, snow, and haze, as shown in Fig. 10. We compare our results with the two most competitive methods, AirNet and WGWS. Our model successfully disentangles weather-specific knowledge and generalizes to restore images degraded by haze." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.795, + 0.614 + ], + "angle": 0, + "content": "5. Conclusion and Broader Impact" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.623, + 0.892, + 0.776 + ], + "angle": 0, + "content": "We have proposed an LDR framework that adaptively removes various adverse weather conditions in an all-in-one solution. Our key insight is to leverage a pre-trained vision-language model to reason diverse weather-specific knowledge in a degraded image. We then use this knowledge to restore a clean image with three modules: degradation map measurement, Top-K expert restoration, and restoration feature aggregation. Experiments on standard benchmark datasets demonstrate that our method outperforms past works by a large margin." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.784, + 0.894, + 0.83 + ], + "angle": 0, + "content": "Broader Impact. Our method is promising to be developed as an image restoration foundation model, prompting by degradation prior generated by a vision-language model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24909" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.472, + 0.4 + ], + "angle": 0, + "content": "[4] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.471, + 0.498 + ], + "angle": 0, + "content": "[5] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.471, + 0.57 + ], + "angle": 0, + "content": "[6] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.47, + 0.628 + ], + "angle": 0, + "content": "[7] Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, and Yuanzhi Li. Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems, 35:23049-23062, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.698 + ], + "angle": 0, + "content": "[8] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.471, + 0.758 + ], + "angle": 0, + "content": "[9] Keval Doshi and Yasin Yilmaz. Multi-task learning for video surveillance with limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3889-3899, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[11] Markus Enzweiler and Dariu M Gavrila. A multilevel mixture-of-experts framework for pedestrian classification. IEEE Transactions on Image Processing, 20(10):2967-2979, 2011. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[12] Sergio Escalera, Hugo Jair Escalante, Zhen Lei, Hao Fang, Ajian Liu, and Jun Wan. Surveillance face presentation attack detection challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6360-6370, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[13] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[14] Trung Hoang, Haichuan Zhang, Amirsaeed Yazdani, and Vishal Monga. Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1670-1679, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[16] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17452-17462, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.559 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[20] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[22] Xinjie Li and Huijuan Xu. Meid: mixture-of-experts with internal distillation for long-tailed video recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1451-1459, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[24] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.163, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[26] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[27] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.334, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[28] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[29] Hamam Mokayed, Amirhossein Nayebiastaneh, Kanjar De, Stergios Sozos, Olle Hagner, and Bjorn Backe. Nordic vehicle dataset (nvd): Performance of vehicle detectors using newly captured nvd from uav in different snowy weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5313-5321, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[30] Lynnette Hui Xian Ng and Kathleen M Carley. Botbuster: Multi-platform bot detection using a mixture of experts. In Proceedings of the International AAAI Conference on Web and Social Media, pages 686-697, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.533, + 0.469, + 0.587 + ], + "angle": 0, + "content": "[31] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.59, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[32] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[33] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.775, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[35] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[36] Prashant W Patil, Sunil Gupta, Santu Rana, Svetha Venkatesh, and Subrahmanyam Murala. Multi-weather image restoration via domain translation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF International Conference on Computer Vision, pages 21696-21705, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[37] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[38] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence, pages 11908-11915, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[40] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[41] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.564 + ], + "angle": 0, + "content": "[43] Yecong Wan, Yuanshuo Cheng, Mingwen Shao, and Jordi González. Image rain removal and illumination enhancement done in one go. Knowledge-Based Systems, 252:109244, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[44] Yinglong Wang, Chao Ma, and Jianzhuang Liu. Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[45] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[46] Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17535-17544, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[47] Hao Yang, Dongming Zhou, Jinde Cao, and Qian Zhao. Dpnet: Detail-preserving image deraining via learning frequency domain knowledge. Digital Signal Processing, 130: 103740, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[48] Hao Yang, Liyuan Pan, Yan Yang, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. arXiv preprint arXiv:2307.09815, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[50] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[52] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[53] Howard Zhang, Yunhao Ba, Ethan Yang, Varan Mehra, Blake Gella, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Alex Wong, and Achuta Kadambi. Weatherstream: Light transport automation of single image deweathering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13499-13509, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.47, + 0.514 + ], + "angle": 0, + "content": "[54] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[55] Zhihong Zhang, Yuxiao Cheng, Jinli Suo, Liheng Bian, and Qionghai Dai. Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing, 32:1390-1402, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.586, + 0.469, + 0.641 + ], + "angle": 0, + "content": "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.738 + ], + "angle": 0, + "content": "[57] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.741, + 0.469, + 0.824 + ], + "angle": 0, + "content": "[58] Simiao Zuo, Qingru Zhang, Chen Liang, Pengcheng He, Tuo Zhao, and Weizhu Chen. Moebert: from bert to mixture-of-experts via importance-guided adaptation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1610–1623, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24912" + } + ] +] \ No newline at end of file diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_origin.pdf b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a9a648a88fc04c80f94f71749ff7e698ec03e71d --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/f89b79b0-6a22-4bde-912e-e3cada02d463_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:644d65f6520b94f4eb78014cbebd642b25cba51b239bcfa1d0884141af97d52b +size 5614313 diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/full.md b/2024/Language-driven All-in-one Adverse Weather Removal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dec32a1ee2d123bab27adbe5d83af26654a4e8dc --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/full.md @@ -0,0 +1,541 @@ +# Language-driven All-in-one Adverse Weather Removal + +Hao Yang $^{1}$ , Liyuan Pan $^{1\dagger}$ , Yan Yang $^{2,3}$ , and Wei Liang $^{1}$ $^{1}$ Beijing Institute of Technology $^{2}$ Australian National University $^{3}$ A&F, CSIRO +{hao.yang, liyuan.pan, liangwei}@bit.edu.cn, {yan.yang}@anu.edu.au + +# Abstract + +All-in-one (AiO) frameworks restore various adverse weather degradations with a single set of networks jointly. To handle various weather conditions, an AiO framework is expected to adaptively learn weather-specific knowledge for different degradations and shared knowledge for common patterns. However, existing methods: 1) rely on extra supervision signals, which are usually unknown in real-world applications; 2) employ fixed network structures, which restrict the diversity of weather-specific knowledge. In this paper, we propose a Language-driven Restoration framework (LDR) to alleviate the aforementioned issues. First, we leverage the power of pre-trained vision-language (PVL) models to enrich the diversity of weather-specific knowledge by reasoning about the occurrence, type, and severity of degradation, generating description-based degradation priors. Then, with the guidance of degradation prior, we sparsely select restoration experts from a candidate list dynamically based on a Mixture-of-Experts (MoE) structure. This enables us to adaptively learn the weather-specific and shared knowledge to handle various weather conditions (e.g., unknown or mixed weather). Experiments on extensive restoration scenarios show our superior performance. + +# 1. Introduction + +Imaging under unpleasant degradation conditions poses challenges for vision-based systems, such as self-driving cars [15, 33, 34, 46] and outdoor surveillance systems [9, 12], which require 24/7 service regardless of adverse weather conditions like rain [6, 13, 47], haze [14, 38, 41], and snow [4, 26, 29]. To meet the safety demand, compared to tackling each type of weather condition with independent models, the joint task process, i.e., all-in-one framework, has a broader application scenario [17, 35]. + +Formally, degraded images can be modeled as masked additive combination of clean images and degradation residuals [35]. Consequently, several works [5, 21, 42] use + +![](images/67f9b721831be68e375c31c389918ce754b9caeed68a7911c7624435bb03c5c9.jpg) +Figure 1. Score (PSNR and SSIM) comparisons. We compare our model (red) with the top 2 (blue and green) baselines on benchmark datasets with various weather scenarios. Superscripts besides evaluation metrics are used to differentiate datasets. + +a single network for all degradation types. Though shared knowledge is learned for restoration, they neglect that different degradations still hold different mathematical formulations, e.g., transmission map produced by scattering effect in haze model [41] is unnecessary for raindrop model [37]. + +Hence, several works [17, 35, 36, 50, 57] use different sub-networks for weather-specific knowledge learning. However, auxiliary supervisions are required to assign the sub-networks, e.g., degradation types [35] or depth maps [57]. Furthermore, the fixed sub-network architecture of existing methods restricts the diversity of learned weather-specific knowledge and their ability to handle images with various weather conditions [43, 55], such as images degraded by weather severity or conditions that have not been encountered before [1, 37], or mixed weather conditions like snow with haze in real-world scenarios. + +In this paper, we question - can we restore the image degraded by various weather conditions by adaptive learning of diverse weather-specific knowledge and shared knowledge, without requiring ground truth weather types and severity data? We answer it by our language-driven all-in-one restoration (LDR) framework in two aspects. + +![](images/50aa2018c212c214be77ad863870163f9cec1a419a19d98e7652354a1aa2e1c5.jpg) + +![](images/0c734cbb34f9ad4a54233e3b575fd7f3116abc3a59f32166881e5c7402556592.jpg) + +![](images/e0b9a9c7f677cb834bb966921fe4a2fd811c2546718bb7a791a6da2d5bb33352.jpg) +(a) 7-th + +![](images/748723c14dc7c7702fd4c127ca11dc510965482ee47cffe92c958dfa4b1aec5a.jpg) + +![](images/6a2626627053a55e9160fba2dcd7024f4b34055165cdea5f32007169c90c4f80.jpg) + +![](images/ae952ad470c2933b85efac49a11dd212f781a6f6bd1fa5099db6864a4d72cb76.jpg) +(b) 10-th + +![](images/b09b8e2abea57923d3950697b301fad2851e703c2df24d04acf06d48bd0b4a63.jpg) + +![](images/092201e40044eb18b472aff2e3d45ee923fa93e6f2442686fb0cfd4e6da75ffe.jpg) + +![](images/327a90d47932adb9d915ddc7eb3a3ad1bc0b05b56c0420d4103c0ab617a4aec8.jpg) +(c) 54-th + +![](images/ae711ce1ed555676d531eb23e300e1b02ecd2431428108753f089e1892e1b992.jpg) + +![](images/8b4f9b714e28ca1922da243efb997a6e3051bf93c12b7ba9d2a32d6e39127d1d.jpg) + +![](images/cf9a67bf073c1ded610eeb5a22466e8900ecd7cc57cfd505c8e1fd09c294776d.jpg) +(d) Input + +![](images/bce8aaa65c6a711d0d4ddaee6d0c74b15ad66b515074070b7373b7a77ed80899.jpg) + +![](images/dc31a9180b6675e0d7ce46b41415060d73f4098d1b87926f0675e98754f5bbc7.jpg) + +![](images/ca35d23aa2af05ebfd10917394ad2b4bc5f4e8754c611e31ef0045d3a30fb186.jpg) +(e) Transweather (f) Zero-out + +![](images/e68a7fd0f76b00c65f7c70edd7079a8b4f82f1b0025e1add05a0243fbcb59575.jpg) + +![](images/d4a08e5b5cf243f0d1f35b597ae6919de46668c3bdab2906ddd44afc2c5db0b6.jpg) + +![](images/039335b1ea8d96969a6c0aed78049623686b69b442d3c5a2e8a314696fd99f31.jpg) + +(f) Zero-out + +![](images/dadd85a57d3f5069af34ac9be41c970f73af73540b12034a152708f953edd466.jpg) + +![](images/e34f96cbf2cd6213d6f3e77f985627acd519882d1f4c6661f2e43e331bc4a336.jpg) + +![](images/a23680f075d1303654c8dd674959a887c0c172c5dd3c22c7e33e70ad4e46b73d.jpg) +(g) Distribution + + + + +# + +Figure 2. Ineffective computations of Transweather [42]. We find the activity of some parameters in Transweather [42] are degradation dependent, and zero-outing computations for those inactivate parameters barely affect image restoration. (a) 7-th channel, (b) 10-th channel, and (c) 54-th channel of feature maps from Transweather, the brighter, the larger activation value, are activated for degradation of rain, raindrop, and snow. We show the (d) degraded images, (e) restorations from Transweather, and (f) restorations by zero-outing inactive channels. (g) From top to bottom, distribution of $\mathcal{L}_2$ norm for the three channels across all images in the All-weather dataset [21]. + +![](images/09e99b2af473f8159c811dd518168b10d5ff7d0b143fc5634ca03e03c03fc11e.jpg) + +1) The knowledge within the feature space of a pretrained vision-language (PVL) model can benefit various tasks, while its potential in our task is still under exploration. A straightforward way is to use the PVL model as the image degradation classifier. In contrast, we go one step further by reasoning diverse weather-specific knowledge from the feature space of the PVL model beyond the type of weather conditions. + +We start by formulating a question prompt to query the occurrence, type, and severity of degradation in a degraded image. The obtained degradation prior describes what, where, and severity of degradations in high-level textual semantics. Then, we translate the high-level degradation prior to a 2D degradation map by aligning the prior with the degraded image. This degradation map provides a pixel-wise representation of the diverse knowledge of image degradation from the PVL model. + +2) We then unleash the potential for various weather removal with the guidance of degradation maps. Observing model parameters are weather-specific, e.g., the rain-related parameters are usually inactive for unrelated degradations, and zero-outing the computations of unrelated parameters also barely affects the restoration quality. Take TUM [5] and Transweather [42] as examples, the PSNR and SSIM w/wo zero-outing on the All-weather dataset are $27.93 / 26.52$ (dB) and $0.883 / 0.864$ , and $27.98 / 26.65$ (dB) and $0.884 / 0.866$ . This observation, illustrated in Fig. 2, inspired us to bypass computations for parameters unrelated to the specific weather type and severity during image restoration. With the assistance of MoE structure [28, 40], our LDR framework selects experts dynamically for restoration, therefore, ensuring adaptive learning of weather-specific knowledge which is not limited to a fixed network architecture. + +Specifically, we maintain a candidate list of restoration experts and utilize the degradation map to sparsely select the most related restoration experts for each degra + +dition. By applying the selected expert pixel-wisely to restore weather-specific features, we create flexible and degradation-adaptive expert combinations/model architectures for restoration. + +Though we can have a preliminary restoration from experts restored feature, considering image regions with similar values from the degradation map tend to benefit restoration features of each other, we re-use the degradation map to aggregate the restoration features, and improve the locality of the obtained restoration by a simple convolutional feedforward network. Our main contributions are: + +- We present an LDR framework to adaptively remove various adverse weather conditions in an all-in-one solution; +- We propose a degradation map measurement module for extracting diverse weather-specific knowledge from a pre-trained vision-language model; +- We propose a Top-K expert restoration module, sparsely and adaptively computing pixel-wise restoration features. + +The overall comparison in Fig. 1 shows the superiority of our framework for handling various degradations compared to state-of-the-art methods. + +# 2. Related Work + +Adverse Weather Restoration. The field of adverse weather restoration encompasses two distinct approaches: task-specific methods [2, 3, 23, 51, 52] train the model independently for each individual degradation and all-in-one frameworks [5, 17, 21, 32, 35, 42, 44, 54, 57] develop a single unified model capable of handling various adverse weather degradations. A prevalent strategy for the all-in-one framework is to build upon task-specific models by employing multi-task learning techniques [44, 54, 57] or knowledge distillation [5] that consolidates multiple task-specific networks into a single network. Nevertheless, these methods typically rely on fixed computations, and the computation effectiveness is discounted as model parameters are + +![](images/e554615ba8772bfb9d1856ab3bef8f3b2dd8afe4b8f6430ef2653db0cf5cbbf4.jpg) +Figure 3. The pipeline of our method. Given the input degraded image $\mathbf{I}^d$ , we aim to recover the clean image $\hat{\mathbf{I}}^c$ . To tackle various adverse weather conditions, we format a question prompt $\mathbf{T}$ , and query a pre-trained vision-language (PVL) model with $\mathbf{I}^d$ and $\mathbf{T}$ , to estimate the degradation type and severity. The generated descriptions $\mathbf{P}^{txt}$ are transformed by a multilayer perceptron (MLP) to get the degradation prior $\mathbf{P}^{emb}$ . Meanwhile, we extract a feature map $\mathbf{X}$ from the input degraded image $\mathbf{I}^d$ . A degradation map $\mathbf{M}$ is computed by cross-attending $\mathbf{X}$ with the prior $\mathbf{P}^{emb}$ , describing pixel-wise degradation pattern. We maintain a trainable candidate list of convolution filters, $\{\mathbf{F}_n|n = 1,\dots,N\}$ , and denote them as experts. We first parse the degradation map $\mathbf{M}$ into a score map $\mathbf{S}$ via a feedforward network $FFN^m$ and a softmax layer, and then use $\mathbf{S}$ to find the best $K$ experts describing the degradation of $\mathbf{X}$ . Best experts are convolved with $\mathbf{X}$ to generate an intermediate feature map $\hat{\mathbf{X}}^{int}$ . Finally, a cross-attention layer is used to aggregate $\hat{\mathbf{X}}^{int}$ with the guidance of the degradation map $\mathbf{M}$ . We improve the feature locality with a feedforward network $FFN^{int}$ , and the output $\hat{\mathbf{X}}$ is decoded to the clean image $\hat{\mathbf{I}}^c$ . + +separately and degradation-dependently activated. + +Conversely, several methods [17, 35, 57] have been developed to overcome these limitations by focusing on learning weather-specific knowledge. They utilize indicators like degradation types [27, 35] or depth maps [57] during training to categorize and direct images to appropriate sub-networks for restoration. Though AirNet [17] learns weather-specific knowledge with contrastive learning. Its fixed sub-network designs can only restore images with certain degradation types, and does not generalize to unseen weather degradations, e.g., rain mixed with haze. Furthermore, these mentioned works overlook the fact that varying levels of degradation severity also warrant adaptively tailored computational processes, for a specific degradation type. + +Diverging from previous methods, we harness PVL models to reason both the type and severity of degradations, enabling adaptive restoration with dynamic sub-networks informed by this rich, weather-specific knowledge. + +Sparse Mixture of Expert. The concept of expert models [28] is defined as a subset of model parameters/computation, where each expert is specialized in handling distinct aspects of the input data. The sparse mixture of experts usually employs a routing mechanism [40] to dynamically and adaptively forward input to a subset of these experts, bypassing unnecessary and irrelevant computations. Given fixed computation costs, due to the sparsity, this framework can readily scale the number of experts to improve the model capability that has been widely verified + +in the domain of natural language processing [10, 40, 58] and computer vision [7, 11, 22, 30]. We study a prior derived from a PVL model to dynamically select experts, and adaptively apply expert to restore degraded images. + +Vision-language Model. With the release of ChatGPT, remarkable reasoning abilities of large language model (LLM) [16, 18, 25, 31, 39, 56] have been shown. It motivates the vision-language community to transfer the reasoning abilities to visual data. A common pipeline is to project an image to a joint space of LLM [19, 24, 25, 48, 49], and feed the projected image alongside user-provided text for conditional text generation. This paper proposes to leverage the PVL (e.g., [16]) to generate prior of an adverse weather degraded image for adaptive image restoration. + +# 3. Method + +Overview. Given an input image $\mathbf{I}^{\mathrm{d}}$ degraded by adverse weather, we aim to recover the corresponding clean image $\hat{\mathbf{I}}^{\mathrm{c}}$ . Our method adaptively recovers $\hat{\mathbf{I}}^{\mathrm{c}}$ with degradation prior $\mathbf{P}^{\mathrm{emb}}$ obtained from a pre-trained vision-language (PVL) model. The overall architecture is shown in Fig. 3. + +To derive degradation prior $\mathbf{P}^{\mathrm{emb}}$ , we query a PVL model with a question prompt $\mathbf{T}$ to reason about the occurrence, type, and severity of degradation in the input image $\mathbf{I}^{\mathrm{d}}$ , outputting $\mathbf{P}^{\mathrm{txt}}$ . Then a mapping network projects $\mathbf{P}^{\mathrm{txt}}$ to $\mathbf{P}^{\mathrm{emb}}$ . Meanwhile, we encode $\mathbf{I}^{\mathrm{d}}$ to $\mathbf{X}$ by using an encoder. + +We adaptively restore $\mathbf{X}$ into $\hat{\mathbf{X}}$ in the embedding space, and decode $\hat{\mathbf{X}}$ to a restored image $\hat{\mathbf{l}}^{\mathrm{c}}$ , with three steps: i) degradation map measurement, measuring degradation for + +![](images/29e84c1c9975a12dcb97d88fd3be8ddacf2c2f37f104412b483ee2c619870020.jpg) +Figure 4. Examples text descriptions from LISA [16] for reasoning weather degraded images. + +each pixel of $\mathbf{I}^{\mathrm{d}}$ by using the degradation prior $\mathbf{P}^{\mathrm{emb}}$ ; ii) top-K expert restoration, selecting expert/parameters from a trainable candidate list of convolution filters according to the degradation map $\mathbf{M}$ , and convolving with $\mathbf{X}$ to get intermediate restoration features $\hat{\mathbf{X}}^{\mathrm{int}}$ ; iii) restoration feature aggregation, deriving $\hat{\mathbf{X}}$ by pixel-wisely aggregating $\hat{\mathbf{X}}^{\mathrm{int}}$ with respect to the degradation map $\mathbf{M}$ , and improving the feature locality with a feedforward network. Finally, $\hat{\mathbf{X}}$ is decoded to get the restored image $\hat{\mathbf{I}}^{\mathrm{c}}$ . + +# 3.1. Degradation Prior + +Degradation Prior Generation. We leverage the context learning capability of the PVL model (e.g., [16]) to reason diverse degradation knowledge of the degraded image $\mathbf{I}^{\mathrm{d}}$ with a question prompt $\mathbf{T}$ . We format the question prompt $\mathbf{T}$ inspired by the chain-of-thought reasoning that makes the model to identify the occurrence, type, and severity of degradation. In Fig. 4, we provide the text description examples obtained by using different prompts. The generated descriptions $\mathbf{P}^{\mathrm{txt}}$ are defined as + +$$ +\mathbf {P} ^ {\mathrm {t x t}} = \operatorname {V L} \left(\mathbf {I} ^ {\mathrm {c}}, \mathbf {T}\right), \quad \mathbf {P} ^ {\mathrm {t x t}} \in \mathbb {R} ^ {L \times C ^ {\mathrm {v l}}}, \tag {1} +$$ + +where $\mathrm{VL}(\cdot ,\cdot)$ is the PVL model, $L$ is the description length, and $C^{\mathrm{vl}}$ is the channel dimension. To preserve the representation capabilities of PVL, we prune out the text description output layer of PVL, and use the embedding before the output layer as $\mathbf{P}^{\mathrm{xt}}$ . + +Degradation Prior Embedding. We align $\mathbf{P}^{\mathrm{txt}}$ to the embedding space with size $C$ of our restoration model by using a multilayer perceptron network MLP( $\cdot$ ). We have + +$$ +\mathbf {P} ^ {\mathrm {e m b}} = \operatorname {M L P} \left(\mathbf {P} ^ {\mathrm {t x t}}\right), \quad \mathbf {P} ^ {\mathrm {e m b}} \in \mathbb {R} ^ {L \times C}. \tag {2} +$$ + +The $\mathbf{P}^{\mathrm{emb}}$ is then used as the degradation prior in our restoration model to adaptively recover $\hat{\mathbf{I}}^{\mathrm{c}}$ . + +# 3.2. Language-driven Restoration Model + +We use an encoder-decoder architecture [8] as a backbone, and adaptively recover $\hat{\mathbf{I}}^{\mathrm{c}}$ in the embedding space with $\mathbf{P}^{\mathrm{emb}}$ and $\mathbf{X}$ (the encoded embedding of $\mathbf{I}^{\mathrm{d}}$ ) as inputs. + +![](images/98708d37090c1b36d206cd8a414ffca3b0bdd2d88fb505e013bcb077b684a235.jpg) +Figure 5. Visualization of $\mathbf{X}$ , $\mathbf{M}$ , $\hat{\mathbf{X}}^{int}$ and $\hat{\mathbf{X}}$ . Without cherry-picking, we show the channel with maximum gradients. From (b) $\mathbf{X}$ to (c) $\mathbf{M}$ , degraded regions are highlighted, and regions with similar degradation severity have similar activation values on the feature map $\mathbf{M}$ . From (b) $\mathbf{X}$ to (d) $\hat{\mathbf{X}}^{int}$ and (e) $\hat{\mathbf{X}}$ , degradations are removed step by step. + +Degradation Map Measurement. The degradation prior $\mathbf{P}^{\mathrm{emb}}$ is text description related with high-level semantics. However, for $\mathbf{X} \in \mathbb{R}^{H \times W \times C}$ with height $H$ , width $W$ , and channel $C$ , the degradation is usually pixel-wise and distinct in severity. To allow pixel-wise computation of adverse weather removal, we measure the 2D degradation map $\mathbf{M}$ for $\mathbf{X}$ by the cross-attention mechanism, + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} ^ {\mathrm {q} _ {1}}, \quad \mathbf {K} = \mathbf {P} ^ {\text {e m b}} \mathbf {W} ^ {\mathrm {k} _ {1}}, \quad \mathbf {V} = \mathbf {P} ^ {\text {e m b}} \mathbf {W} ^ {\mathrm {v} _ {1}}, \tag {3} +$$ + +$$ +\mathbf {M} = \text {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {S o f t m a x} \left(\mathbf {Q} \mathbf {K} ^ {\top}\right) \mathbf {V}, \tag {4} +$$ + +where $\mathbf{W}^{\mathrm{q}_1}$ , $\mathbf{W}^{\mathrm{k}_1}$ , and $\mathbf{W}^{\mathrm{v}_1} \in \mathbb{R}^{C \times C}$ are linear projection matrices for obtaining the query, key, and value. Our cross-attention mechanism first computes the semantic alignments between $\mathbf{X}$ and $\mathbf{P}^{\mathrm{emb}}$ via Softmax $(\mathbf{QK}^\top)$ , and then transforms $\mathbf{P}^{\mathrm{emb}}$ into a 2D degradation map $\mathbf{M}$ that describes pixel-wise degradation of $\mathbf{X}$ . We visualize $\mathbf{X}$ and $\mathbf{M}$ in Fig. 5. The regions of similar degradation severity have been individually highlighted, by applying Eq. (4). + +Top-K Expert Restoration. Not all pixels degrade equally, and we adaptively restore each of them by using pixel-wise degradation knowledge within $\mathbf{M}$ . We have $N$ candidate experts (convolution filters), $\{\mathbf{F}_n|n = 1,\dots,N\}$ , for diverse adverse weather conditions, where each candidate $\mathbf{F}_n$ is expertise at generating restoration features for specific degradation types or severities (see 4.2). With the degradation map $\mathbf{M}$ , we select experts for each pixel of $\mathbf{X}$ . + +Specifically, the degradation map $\mathbf{M} \in \mathbb{R}^{H \times W \times C}$ is fed to a feedforward network $\mathrm{FFN}^{\mathrm{m}}$ , followed by a Softmax layer along the last dimension, to output a normalized pixel-wise selection score $\mathbf{S} \in \mathbb{R}^{H \times W \times N}$ , i.e., $\mathbf{S} = \operatorname{Softmax}(\mathrm{FFN}^{\mathrm{m}}(\mathbf{M}))$ . For a pixel with location $(i,j), i \in [1,H], j \in [1,W]$ , $\mathbf{S}(i,j) \in \mathbb{R}^N$ and $\mathbf{S}(i,j,n)$ measures how likely the $n$ -th expert $\mathbf{F}_n$ can be used to describe the degradation of $\mathbf{X}(i,j)$ . + +To find the best experts describing the degradation of $\mathbf{X}(i,j)$ , we compute the Top-K scores of $\mathbf{S}(i,j)$ and record + +Table 1. Quantitative comparison on the All-weather dataset. We respectively color the best and the second-best methods in red and blue. + +
TypeMethodRainSnowRaindropAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
BestT + VL24.330.86028.470.87229.230.89527.340.876
BestT + GT27.040.91330.610.90031.630.93629.760.916
GeneralMPRNet [51]23.080.83927.690.84928.750.87926.510.856
NAFNet [3]23.210.84027.680.84728.900.89026.600.859
Uformer [45]22.930.83527.500.83828.510.87126.310.848
Restormer [52]23.370.84527.810.85029.100.89026.760.862
GRL [23]23.310.84227.790.84929.050.88826.720.860
All-in-OneAll-in-One [21]24.710.89828.330.88231.120.92728.050.902
AirNet [17]23.120.83727.920.85828.230.89226.420.862
TUM [5]23.920.85529.270.88430.750.91227.980.884
Transweather [42]23.180.84127.800.85428.980.90226.650.866
WDiff [32]26.180.90729.690.89329.710.91128.530.904
WGWS [57]25.310.90129.710.89431.310.93228.780.909
Ours26.920.91230.790.90531.540.93329.750.916
+ +corresponding indices as the best experts. The best experts restored feature for location $(i,j)$ is given by, + +$$ +\hat {\mathbf {X}} ^ {\text {i n t}} (i, j) = \sum_ {k = 1} ^ {K} \mathbf {S} (i, j, \rho (k)) \cdot \mathcal {E} (i, j, \rho (k)), \tag {5} +$$ + +$$ +\mathcal {E} (i, j, \rho (k)) = \sum_ {\Delta i, \Delta j} \mathbf {X} (i + \Delta i, j + \Delta j) \cdot \mathbf {F} _ {\rho (k)} (\Delta i, \Delta j), +$$ + +where $\rho (k)$ is the index of the selected $k$ -th expert. $\mathcal{E}(i,j,\rho (k))$ is the convolution result of $\mathbf{X}(i,j)$ and the $k$ -th expert $\mathbf{F}_{\rho (k)}$ , and $(\Delta i,\Delta j)$ iterates over the convolution kernel size. $\mathbf{S}(i,j,\rho (k))$ are weights for prioritizing different experts. With our pixel-wise Top-K experts, the intermediate restoration feature $\hat{\mathbf{X}}^{\mathrm{int}}$ for each pixel of $\mathbf{X}$ is adaptively generated. Practically, we find the best performance by region-wisely smoothing the expert selection score $\mathbf{S}$ , as pixels of a region frequently suffer the same degradation. + +Restoration Feature Aggregation. Pixels with similar degradation information in $\mathbf{M}$ potentially benefit each other in deriving the restoration $\hat{\mathbf{X}}$ . We compute the compatibility between degradation prior $\mathbf{M}$ and restoration features $\hat{\mathbf{X}}^{\mathrm{int}}$ , and aggregate weighted restoration features pixel-wisely by using the cross-attention mechanism, + +$$ +\mathbf {Q} = \mathbf {M} \mathbf {W} ^ {\mathrm {q} _ {2}}, \quad \mathbf {K} = \hat {\mathbf {X}} ^ {\mathrm {i n t}} \mathbf {W} ^ {\mathrm {k} _ {2}}, \quad \mathbf {V} = \hat {\mathbf {X}} ^ {\mathrm {i n t}} \mathbf {W} ^ {\mathrm {v} _ {2}}, \tag {6} +$$ + +$$ +\hat {\mathbf {X}} = \operatorname {F F N} ^ {\operatorname {i n t}} (\operatorname {A t t e n t i o n} (\mathbf {Q}, \mathbf {K}, \mathbf {V})) ， \tag {7} +$$ + +where $\mathrm{FFN}^{\mathrm{int}}$ is a feedforward convolutional network to improve the feature locality. Finally, $\hat{\mathbf{X}}$ is decoded to the restoration $\hat{\mathbf{I}}^{\mathrm{c}}$ . In Fig. 5, we compare $\hat{\mathbf{X}}^{\mathrm{int}}$ and $\hat{\mathbf{X}}$ , and find that $\hat{\mathbf{X}}$ is cleaner than $\hat{\mathbf{X}}^{\mathrm{int}}$ , as pixels with similar degradation measurements benefit each other in the restorations. + +# 3.3. Loss Function + +We train our network with Charbonnier loss $\mathcal{L}_{\mathrm{char}}$ and gradient-level edge loss $\mathcal{L}_{\mathrm{edge}}$ , to penalize the deviation of + +restoration from ground truth clean image, and encourage the consistent image gradients. We have $\mathcal{L}_{\mathrm{char}}$ as + +$$ +\mathcal {L} _ {\text {c h a r}} = \sqrt {\left\| \mathbf {I} ^ {\mathrm {c}} - \hat {\mathbf {I}} ^ {\mathrm {c}} \right\| ^ {2} + \varepsilon^ {2}}, \tag {8} +$$ + +where $\varepsilon = 10^{-4}$ is a constant in all experiments. The $\mathcal{L}_{\mathrm{edge}}$ is + +$$ +\mathcal {L} _ {\text {e d g e}} = \sqrt {\| \nabla \mathbf {I} ^ {\mathrm {c}} - \nabla \hat {\mathbf {I}} ^ {\mathrm {c}} \| ^ {2} + \varepsilon^ {2}}, \tag {9} +$$ + +where $\nabla$ is the laplacian gradient operator. With a balance parameter $\lambda$ , the total loss is given by + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c h a r}} + \lambda \mathcal {L} _ {\text {e d g e}}. \tag {10} +$$ + +# 4. Experiments and Analysis + +Implementation Details. Our model is implemented using the PyTorch framework and all experiments are conducted on an RTX A6000 GPU. We train our network with batch size 4 and $4 \times 10^{6}$ iterations using the ADAM optimizer. The learning rate is decayed from $2 \times 10^{-4}$ to $1 \times 10^{-6}$ by cosine annealing strategy. In training, images are randomly cropped to size $256 \times 256$ , and $\lambda = 0.05$ . Code and more results are in our project page. + +Datasets. Our experiments are conducted on both synthetic dataset and real dataset, i.e., All-weather dataset [21] and WeatherStream [53] dataset. All-weather dataset comprises 18,609 training images and 17,609 testing images, and is composed of subsets from Outdoor-rain [20], Snow100K-L [26], and Raindrop [37], corresponding to rain, snow, and raindrops weather conditions respectively. WeatherStream dataset is a real-world dataset that has three weather conditions, i.e., rain, snow, and fog, with a total of 176, 100 training images and 11, 400 testing images. + +Evaluation Metrics. We use averages of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation metrics (RGB-channel). The higher, the better. + +Table 2. Quantitative comparison on the WeatherStream dataset. We color the best and the second-best methods in red and blue. + +
TypeMethodRainHazeSnowAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
GeneralBestT + VL21.200.78121.600.75520.320.77221.040.769
BestT + GT23.950.81022.970.80422.700.82823.210.814
MPRNet [51]21.500.79121.730.76320.740.80121.320.785
NAFNet [3]23.010.80322.200.80322.110.82622.440.811
Uformer [45]22.250.79118.810.76320.940.80120.670.785
Restormer [52]23.670.80422.900.80322.510.82822.860.812
GRL [23]23.750.80522.880.80222.590.82923.070.812
All-in-OneAll-in-One [21]--------
AirNet [17]22.520.79721.560.77021.440.81221.840.793
TUM [5]23.220.79522.380.80522.250.82722.620.809
Transweather [42]22.210.77222.550.77421.790.79222.180.779
WGWS [57]23.800.80722.780.80022.720.83123.100.813
Ours24.420.81823.110.80923.120.83823.550.822
+ +![](images/1d9bc51965978840ef6d5582de78dbdfd68821b87cb081524193373a2138b83b.jpg) +Rain + +![](images/bef62892f913d7239b3cb214139f16b73a560daaf49d6bf5def2cd37d60e6754.jpg) +Snow + +![](images/e18c2860f6e47e99f493a0340db23e531ae53cc74a923512b5046ac21a1a223b.jpg) +Raindrop +Figure 6. Qualitative comparison on the All-weather dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns. + +![](images/3fd8796e1dc1d017400df17e2d48566b0c508e2ce3d0acf0f47f0ead1595ac10.jpg) +Input + +![](images/bf982d708a9dba1011556f472890c0df3e7d3f92c55724f4e09b52e517f09347.jpg) +TUM [5] + +![](images/bdcd9c91803fdd4afae70a3adc334e19270d035e86aedc912e0575983cbab8ca.jpg) +GT + +![](images/75cc649564af012dfe80bf2e1d2da3af3779010ca756477e6f2839f660467a38.jpg) + +![](images/450cf566e3cfed04bd508527fee1e39efe2373c25fa380d80346860aea569b92.jpg) +GRL [23] + +![](images/cf36925a2f00b9bd48e83404f1a227cd5bf7fd93427fc75499caf650f7458df7.jpg) +WGWS [57] + +![](images/25b68cd19a746c4e08df840fe538f500d5c1b0e9826ab271c3bf389b1103f0ff.jpg) +Input + +![](images/1000766bd9a8590ed0282a55c4fdcd7b5fc369d43415bedaa9c01135ca9ca5dc.jpg) + +![](images/ccaebe707ed23737ab05d91c3b7e6a34a80c3b5942fc088264f2a418c3877ee2.jpg) +GT + +![](images/e3d79301cebac642b483332aeb383ea71b14080ac7bd598ef6e891ad6dcf204d.jpg) + +![](images/da9847f09628e6d8346b52afeae7798b8a183a6ef55e277982a62c5a0c5fea2d.jpg) +GRL [23] + +![](images/00f47c18f4b63113f207ae7feb128b414123411276deca765b2138d3d88c1a6f.jpg) + +![](images/4478250649f099786679deb6dfdb9dc02f61aff497d06ff8d11c45092a59a876.jpg) +TUM [5] +Input + +![](images/8edc7f0add0e9455b57aa25d4642b4865743d83b82d3434ffed475e56fe9e990.jpg) +TUM [5] + +![](images/a4853ebd2bc6f24f0bb634480920e6ffc484ac7a1ffb6720a5960ba6f9b8266c.jpg) +GT + +![](images/6a56d06adbb5f28e51f7e092408bd41cfcbe1738ac9da525f2a8ad64f85b5ee4.jpg) + +![](images/1f5c895c271e6f9bf64865cc63f5d952f3ebdf99994e7eeacfc6892e9ae1b5f4.jpg) +WGWS [57] +GRL [23] + +![](images/8e8553e5c9759111372d02df100f4d133f0e24da9f15e6eb655a8477cb12a273.jpg) +WGWS [57] + +![](images/2e91399c81ad3a95d8f04937690cb4b62344a9365acce3e0af01aa256ccc3069.jpg) +AirNet [17] + +![](images/746421e067a8deb140344d20efab908d875e66d1b9fb89e35b09a830861cdf0a.jpg) + +![](images/e75a300df17f845fae0c9a91d5e942f752f9a60c9743d58fa69d1af3337803c7.jpg) +Ours +AirNet [17] + +![](images/86f487730909b5543e1da0353cd708a37362ec2dd9dfb63b7c85bb757bebd880.jpg) + +![](images/4052a061315f926ed695ff6dc611c1e89d1a907e90d706c9087197f6846a6015.jpg) +Ours +AirNet [17] + +![](images/8ff53dd5f6d2f745eb867376cdd8908c8a38ddc6a01f17b59a7b4bffdd975abb.jpg) +Ours + +Baseline Methods. We compare with the state-of-the-art (SOTA) general and all-in-one methods. For general methods, task-specific methods are trained with multi-task learning of different weather conditions, i.e., MPRNet [51], NAFNet [3], Uformer[45], Restormer [52], and GRL [23]. All-in-one methods are All-in-One [21], AirNet[17], TUM [5], Transweather [42], WDiff [32] and WGWS [57]. + +# 4.1. Experimental Results + +Quantitative Comparison. We compare with the SOTA general and all-in-one methods on the All-weather and + +- WeatherStream datasets in Tab. 1 and Tab. 2, respectively. +- Our method achieves the best performance. + +Furthermore, we create two strong baselines: i) BestT + VL. We use a PVL model for zero-shot adverse weather classification, and select the best task-specific methods for restoring degraded images; ii) BestT + GT. The ground truth adverse weather type is used to select the best task-specific method for image restoration. We find that i) Our method significantly outperforms 'BestT + VL', though it uses the same PVL as our model and uses multiple models for different adverse weather conditions. This shows + +![](images/d16cb43e6bccc291ac61bb5942426b177579885b973c424d584a6c6f7aa08a42.jpg) +Haze + +![](images/117e8ef003f55cd5ed0710c3026996cbce48fecef239179c657f42e708570a45.jpg) +Input + +![](images/c1f1a0967a242331f1bfb152d6738d4b6476e7de88c13a477e2752bd608f2653.jpg) + +![](images/66e9655765fce5aace00d58e23c59b0b7069d561c1d9edb7bd011ee5eb83f77c.jpg) +GRL [23] + +![](images/4b574b23598f722843db316a46317228e281c65e103f4adef20f0eef5d4069e5.jpg) +AirNet 17 + +![](images/1e1f8b717c1bb58edbfc0f078a414bccbc04a35d323065307500989f79644c4c.jpg) +Rain + +![](images/dc4e924bce113b90da8e98f486fd9873d4848c4d5a24ce2bcb34b182346f5290.jpg) +TUM [5] + +![](images/296523e3f89a68c56fc1bff2d6f1e1e246a85d9706eb3e873c1b510cc76820e3.jpg) +GT +transweather +[42] WGWS [57] + +![](images/c79966778d844b3c3d61e0458725134e979be1a5f021c399db8facd3c38fd8d3.jpg) + +![](images/6e656e7b8c5b173f97b82aefc7a6af339ce182c16479540294d4dbb1a5110a73.jpg) +Ours + +![](images/825729df2f225578c745747c97d0e4c146a41a8ff19e94f1a02b7a0e33398154.jpg) +Snow +Figure 7. Qualitative comparison on the WeatherStream dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns. + +![](images/ff7d85d49408d93081c2356ebdd41ed8752765f4340cd1b4e1708e92b32f9a99.jpg) +Input + +![](images/82fcea633d4de1827e5d2e2e975514d96acad55762e73947c7ad70e0664a44b5.jpg) + +AirNet 117 + +![](images/223690d0cd8d9afcaef9ab4c9a7354ee9ca743f022a1c5806b7cbb95578a8b5c.jpg) + +![](images/96937a15dfac9072b01a2b68630f51fb229260583fb992ab30e92f3461f8b84e.jpg) +GT + +![](images/89cff17d97eabdab1cc0cf6b352129ce062332911e58775e271664ecb92ac474.jpg) +GRL [23] + +AirNet [17] + +![](images/044a9834d14a234dab52255373b6f849874d9e4387debfa9243731b198e10cfb.jpg) +TUM [5] +Input + +![](images/042240e7dfc82474d53059be6ad0a5fdcbb627c00b4cdd2871f14d8a579bfe36.jpg) + +![](images/0c3c9b6392469622c85812006190f80acdc6de6c1d12779ae2cdb87f964a67ce.jpg) +] WGWS [57] + +![](images/60a7727666d5dfdec14fde0859b741c9f8717fed0cf90d3afb8635bd922e2d62.jpg) + +TUM [5] + +![](images/ddb873dbd66972312aabf83a658c35bd06a80e6dc637cd81696a6702fcaf6df8.jpg) +GT + +![](images/3ecd56a84ef379a9e40e1f98acc661a41ded2951aaaff466e045e3b4ef852a35.jpg) +GRL [23] + +![](images/631df5213e5f17da325f08671401fd4a5c3c3adb9dc39e4ef203ae2e40fd6917.jpg) +Ours +AirNet [17] + +TUM [5] + +![](images/1fb2c4524ac119102ad155ecb9172af9d454965ddf05f27de3656ecbd4bf7290.jpg) + +![](images/53b169759a1bef939633ee0e16174aa93dcac9561002e2ae8911593b5d7672b7.jpg) +] WGWS [57] + +![](images/51119fccec0d7a82626f46083a1398739ff3386cb10ec091afead246a011884c.jpg) +Ours + +that non-trivial designs are required for effectively leveraging the PVL model in all-in-one adverse weather removal; ii) Although 'BestT + GT' outperforms most SOTA general and all-in-one methods, our method still achieves competitive performance compared to it. This indicates: 1) the benefit of accessing prior knowledge of adverse weather conditions, and 2) our method has learned the shared and weather-specific knowledge adaptively by selecting sparse experts dynamically. + +Qualitative Comparison. We compare with the SOTA methods on restoring images degraded by different adverse weather conditions. The results are given in Fig. 6 and Fig. 7. Our method consistently restores clearer images than SOTA methods under different weather conditions. + +# 4.2. Ablation studies and Discussions + +We validate the effectiveness and components of our framework on the All-weather dataset. + +Model Architecture. We study the effectiveness of the degradation map measurement (DMM), top-K expert restoration (TER), and restoration feature aggregation (RFA) modules. The results are given in Tab. 3. The best performance is achieved by using all modules. + +Degradation Severity. We use the PVL model to partition the All-weather dataset [21] into three subsets based on degradation severity: slight, moderate, and heavy. We + +Table 3. The effectiveness of our model components. + +
DMMTERRFAPSNRSSIM
X✓X27.930.882
XX✓28.370.889
✓✓X28.250.890
✓X✓29.110.902
X✓✓28.550.895
✓✓✓29.750.916
+ +Table 4. Comparisons on the All-weather dataset with degradation severity of slight, moderate, and heavy. + +
MethodSlightModerateHeavy
PSNRSSIMPSNRSSIMPSNRSSIM
AirNet27.590.89526.490.86524.500.818
WGWS29.740.92128.770.91027.140.886
Ours30.460.92529.780.91828.380.899
+ +then compare our results with those of AirNet and WGWS in Tab. 4. Our method shows significant improvement over AirNet [17] and WGWS [57] in the heavily degraded subset. This indicates the effectiveness of reasoning with diverse weather-specific knowledge, such as severity. + +Pixel-wise Expert. Different regions of the same degraded image often exhibit varying degrees of degradation severity, and should be adaptively and pixel-wisely restored. We use the degradation prior provided by the PVL model to select the same expert for the degraded image. It shows a + +![](images/5a066f61434721720a99a5df10ea3fce47225b5925c783fc1282b549410b910e.jpg) +Figure 8. The usage frequencies of experts for different degradation types. For example, Expert 8, 11, and 15 focuses on raindrop, rain, and snow, respectively. Expert 12 handles all types of degradation. Please refer to Fig. 9 for visualization. + +![](images/f6a36680882ed0072f97a1328e6badd79ca2ffe41cf1878709f4de02cb3de923.jpg) +(a) Expert 8 + +![](images/97439ccd3be4d9464249385dcc176b90590c4284faf519b0290333e19af9a208.jpg) +(b) Expert 11 + +![](images/d893fe87b443140c9cd2a7f5fcf598724031af3f844b7ab110efb7775dc138d2.jpg) +(c) Expert 12 + +![](images/e06a8ab188f3963f7618fb7b9a0b492a5655346cc8e9c218a345824d6e16d62a.jpg) +(d) Expert 15 + +![](images/929219944645be22c84fe08df0039f92b9cac4ec21d2cd17fdad3ab2efafb640.jpg) +Figure 9. Sample image regions, activating different experts. We show regions with highest selection scores for experts 8, 11, 12, and 15 on the All-weather dataset. +(a) Degraded +(b) AirNet +(c) WGWS +(d) Ours +Figure 10. From the first to last rows, the degradation is rain, raindrop, snow, and haze, respectively. From left to right, (a) degraded images are restored by (b) AirNet, (c) WGWS, and (d) our method. + +0.33 dB/0.07 decrease in PSNR/SSIM, indicating the necessity of selecting experts pixel-wisely. + +Adaptive Adverse Weather Removal. We study the model response to different adverse weather degradation by measuring the usage of our experts. Fig. 8 shows the usage for rain, snow, and raindrop. We find that there are experts for weather-specific knowledge and shared knowledge among degradation, such as Expert 8, Expert 11, Expert 12, and Expert 15. We visualize the image patches selecting the four experts in Fig. 9, where Expert 8, Expert 11, and Expert 15 are selected by image regions with raindrop, rain, and snow degradation, and Expert 12 is selected by image + +regions with sky in all types of degradation. + +Model Ability Under Complex Weather Condition. We test our model trained on the All-weather dataset with real images degraded by rain, raindrop, snow, and haze, as shown in Fig. 10. We compare our results with the two most competitive methods, AirNet and WGWS. Our model successfully disentangles weather-specific knowledge and generalizes to restore images degraded by haze. + +# 5. Conclusion and Broader Impact + +We have proposed an LDR framework that adaptively removes various adverse weather conditions in an all-in-one solution. Our key insight is to leverage a pre-trained vision-language model to reason diverse weather-specific knowledge in a degraded image. We then use this knowledge to restore a clean image with three modules: degradation map measurement, Top-K expert restoration, and restoration feature aggregation. Experiments on standard benchmark datasets demonstrate that our method outperforms past works by a large margin. + +Broader Impact. Our method is promising to be developed as an image restoration foundation model, prompting by degradation prior generated by a vision-language model. + +Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045. + +# References + +[1] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1 +[2] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 2 +[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 2, 5, 6 +[4] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1 +[5] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 5, 6, 7 +[6] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1 +[7] Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, and Yuanzhi Li. Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems, 35:23049-23062, 2022. 3 +[8] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 4 +[9] Keval Doshi and Yasin Yilmaz. Multi-task learning for video surveillance with limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3889-3899, 2022. 1 +[10] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 3 +[11] Markus Enzweiler and Dariu M Gavrila. A multilevel mixture-of-experts framework for pedestrian classification. IEEE Transactions on Image Processing, 20(10):2967-2979, 2011. 3 + +[12] Sergio Escalera, Hugo Jair Escalante, Zhen Lei, Hao Fang, Ajian Liu, and Jun Wan. Surveillance face presentation attack detection challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6360-6370, 2023. 1 +[13] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[14] Trung Hoang, Haichuan Zhang, Amirsaeed Yazdani, and Vishal Monga. Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1670-1679, 2023. 1 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[16] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3, 4 +[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17452-17462, 2022. 1, 2, 3, 5, 6, 7 +[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3 +[20] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 5 +[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 5, 6, 7 +[22] Xinjie Li and Huijuan Xu. Meid: mixture-of-experts with internal distillation for long-tailed video recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1451-1459, 2023. 3 +[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 5, 6, 7 + +[24] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 3 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3 +[26] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 1, 5 +[27] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 3 +[28] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2, 3 +[29] Hamam Mokayed, Amirhossein Nayebiastaneh, Kanjar De, Stergios Sozos, Olle Hagner, and Bjorn Backe. Nordic vehicle dataset (nvd): Performance of vehicle detectors using newly captured nvd from uav in different snowy weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5313-5321, 2023. 1 +[30] Lynnette Hui Xian Ng and Kathleen M Carley. Botbuster: Multi-platform bot detection using a mixture of experts. In Proceedings of the International AAAI Conference on Web and Social Media, pages 686-697, 2023. 3 +[31] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24, 2011. 3 +[32] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5, 6 +[33] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1 +[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1 +[35] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 1, 2, 3 +[36] Prashant W Patil, Sunil Gupta, Santu Rana, Svetha Venkatesh, and Subrahmanyam Murala. Multi-weather image restoration via domain translation. In Proceedings of + +the IEEE/CVF International Conference on Computer Vision, pages 21696-21705, 2023. 1 +[37] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 1, 5 +[38] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence, pages 11908-11915, 2020. 1 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[40] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 3 +[41] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 1 +[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 5, 6, 7 +[43] Yecong Wan, Yuanshuo Cheng, Mingwen Shao, and Jordi González. Image rain removal and illumination enhancement done in one go. Knowledge-Based Systems, 252:109244, 2022. 1 +[44] Yinglong Wang, Chao Ma, and Jianzhuang Liu. Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2023. 2 +[45] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5, 6 +[46] Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17535-17544, 2023. 1 +[47] Hao Yang, Dongming Zhou, Jinde Cao, and Qian Zhao. Dpnet: Detail-preserving image deraining via learning frequency domain knowledge. Digital Signal Processing, 130: 103740, 2022. 1 +[48] Hao Yang, Liyuan Pan, Yan Yang, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. arXiv preprint arXiv:2307.09815, 2023. 3 +[49] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 3 +[50] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 1 +[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 5, 6 +[52] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6 +[53] Howard Zhang, Yunhao Ba, Ethan Yang, Varan Mehra, Blake Gella, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Alex Wong, and Achuta Kadambi. Weatherstream: Light transport automation of single image deweathering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13499-13509, 2023. 5 +[54] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2 +[55] Zhihong Zhang, Yuxiao Cheng, Jinli Suo, Liheng Bian, and Qionghai Dai. Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing, 32:1390-1402, 2023. 1 +[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[57] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 1, 2, 3, 5, 6, 7 +[58] Simiao Zuo, Qingru Zhang, Chen Liang, Pengcheng He, Tuo Zhao, and Weizhu Chen. Moebert: from bert to mixture-of-experts via importance-guided adaptation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1610–1623, 2022. 3 \ No newline at end of file diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/images.zip b/2024/Language-driven All-in-one Adverse Weather Removal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..57e4d9302ee8fb4e632a2b112ef92b11c86c61c5 --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9015d1b14c06082394a8bf306428b13488fd4abeb3ae40f6c0e38a0c2ba2ca73 +size 955948 diff --git a/2024/Language-driven All-in-one Adverse Weather Removal/layout.json b/2024/Language-driven All-in-one Adverse Weather Removal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6d93ace70cfef752dd99a7fc2da333a9533d62f --- /dev/null +++ b/2024/Language-driven All-in-one Adverse Weather Removal/layout.json @@ -0,0 +1,16074 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 126, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 126, + 103, + 467, + 121 + ], + "type": "text", + "content": "Language-driven All-in-one Adverse Weather Removal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": "Hao Yang" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": ", Liyuan Pan" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": ", Yan Yang" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": ", and Wei Liang" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": "Beijing Institute of Technology " + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": "Australian National University " + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 142, + 503, + 186 + ], + "type": "text", + "content": "A&F, CSIRO \n{hao.yang, liyuan.pan, liangwei}@bit.edu.cn, {yan.yang}@anu.edu.au" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "spans": [ + { + "bbox": [ + 143, + 224, + 192, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 250, + 290, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 290, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 290, + 515 + ], + "type": "text", + "content": "All-in-one (AiO) frameworks restore various adverse weather degradations with a single set of networks jointly. To handle various weather conditions, an AiO framework is expected to adaptively learn weather-specific knowledge for different degradations and shared knowledge for common patterns. However, existing methods: 1) rely on extra supervision signals, which are usually unknown in real-world applications; 2) employ fixed network structures, which restrict the diversity of weather-specific knowledge. In this paper, we propose a Language-driven Restoration framework (LDR) to alleviate the aforementioned issues. First, we leverage the power of pre-trained vision-language (PVL) models to enrich the diversity of weather-specific knowledge by reasoning about the occurrence, type, and severity of degradation, generating description-based degradation priors. Then, with the guidance of degradation prior, we sparsely select restoration experts from a candidate list dynamically based on a Mixture-of-Experts (MoE) structure. This enables us to adaptively learn the weather-specific and shared knowledge to handle various weather conditions (e.g., unknown or mixed weather). Experiments on extensive restoration scenarios show our superior performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 128, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 660 + ], + "type": "text", + "content": "Imaging under unpleasant degradation conditions poses challenges for vision-based systems, such as self-driving cars [15, 33, 34, 46] and outdoor surveillance systems [9, 12], which require 24/7 service regardless of adverse weather conditions like rain [6, 13, 47], haze [14, 38, 41], and snow [4, 26, 29]. To meet the safety demand, compared to tackling each type of weather condition with independent models, the joint task process, i.e., all-in-one framework, has a broader application scenario [17, 35]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 660, + 287, + 696 + ], + "type": "text", + "content": "Formally, degraded images can be modeled as masked additive combination of clean images and degradation residuals [35]. Consequently, several works [5, 21, 42] use" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 339, + 212, + 520, + 392 + ], + "blocks": [ + { + "bbox": [ + 339, + 212, + 520, + 392 + ], + "lines": [ + { + "bbox": [ + 339, + 212, + 520, + 392 + ], + "spans": [ + { + "bbox": [ + 339, + 212, + 520, + 392 + ], + "type": "image", + "image_path": "67f9b721831be68e375c31c389918ce754b9caeed68a7911c7624435bb03c5c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 394, + 547, + 440 + ], + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 440 + ], + "type": "text", + "content": "Figure 1. Score (PSNR and SSIM) comparisons. We compare our model (red) with the top 2 (blue and green) baselines on benchmark datasets with various weather scenarios. Superscripts besides evaluation metrics are used to differentiate datasets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 547, + 508 + ], + "type": "text", + "content": "a single network for all degradation types. Though shared knowledge is learned for restoration, they neglect that different degradations still hold different mathematical formulations, e.g., transmission map produced by scattering effect in haze model [41] is unnecessary for raindrop model [37]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 548, + 640 + ], + "type": "text", + "content": "Hence, several works [17, 35, 36, 50, 57] use different sub-networks for weather-specific knowledge learning. However, auxiliary supervisions are required to assign the sub-networks, e.g., degradation types [35] or depth maps [57]. Furthermore, the fixed sub-network architecture of existing methods restricts the diversity of learned weather-specific knowledge and their ability to handle images with various weather conditions [43, 55], such as images degraded by weather severity or conditions that have not been encountered before [1, 37], or mixed weather conditions like snow with haze in real-world scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "In this paper, we question - can we restore the image degraded by various weather conditions by adaptive learning of diverse weather-specific knowledge and shared knowledge, without requiring ground truth weather types and severity data? We answer it by our language-driven all-in-one restoration (LDR) framework in two aspects." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 702, + 142, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 702, + 142, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 702, + 142, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24902" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 128, + 111 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 128, + 111 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 128, + 111 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 128, + 111 + ], + "type": "image", + "image_path": "50aa2018c212c214be77ad863870163f9cec1a419a19d98e7652354a1aa2e1c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 113, + 127, + 152 + ], + "blocks": [ + { + "bbox": [ + 77, + 113, + 127, + 152 + ], + "lines": [ + { + "bbox": [ + 77, + 113, + 127, + 152 + ], + "spans": [ + { + "bbox": [ + 77, + 113, + 127, + 152 + ], + "type": "image", + "image_path": "0c734cbb34f9ad4a54233e3b575fd7f3116abc3a59f32166881e5c7402556592.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 77, + 154, + 127, + 194 + ], + "blocks": [ + { + "bbox": [ + 77, + 154, + 127, + 194 + ], + "lines": [ + { + "bbox": [ + 77, + 154, + 127, + 194 + ], + "spans": [ + { + "bbox": [ + 77, + 154, + 127, + 194 + ], + "type": "image", + "image_path": "e0b9a9c7f677cb834bb966921fe4a2fd811c2546718bb7a791a6da2d5bb33352.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 196, + 113, + 206 + ], + "lines": [ + { + "bbox": [ + 85, + 196, + 113, + 206 + ], + "spans": [ + { + "bbox": [ + 85, + 196, + 113, + 206 + ], + "type": "text", + "content": "(a) 7-th" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 181, + 110 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 181, + 110 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 181, + 110 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 181, + 110 + ], + "type": "image", + "image_path": "748723c14dc7c7702fd4c127ca11dc510965482ee47cffe92c958dfa4b1aec5a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 130, + 112, + 181, + 152 + ], + "blocks": [ + { + "bbox": [ + 130, + 112, + 181, + 152 + ], + "lines": [ + { + "bbox": [ + 130, + 112, + 181, + 152 + ], + "spans": [ + { + "bbox": [ + 130, + 112, + 181, + 152 + ], + "type": "image", + "image_path": "6a2626627053a55e9160fba2dcd7024f4b34055165cdea5f32007169c90c4f80.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 130, + 154, + 181, + 194 + ], + "blocks": [ + { + "bbox": [ + 130, + 154, + 181, + 194 + ], + "lines": [ + { + "bbox": [ + 130, + 154, + 181, + 194 + ], + "spans": [ + { + "bbox": [ + 130, + 154, + 181, + 194 + ], + "type": "image", + "image_path": "ae952ad470c2933b85efac49a11dd212f781a6f6bd1fa5099db6864a4d72cb76.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 196, + 170, + 206 + ], + "lines": [ + { + "bbox": [ + 137, + 196, + 170, + 206 + ], + "spans": [ + { + "bbox": [ + 137, + 196, + 170, + 206 + ], + "type": "text", + "content": "(b) 10-th" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 236, + 110 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 236, + 110 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 236, + 110 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 236, + 110 + ], + "type": "image", + "image_path": "b09b8e2abea57923d3950697b301fad2851e703c2df24d04acf06d48bd0b4a63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 186, + 112, + 235, + 152 + ], + "blocks": [ + { + "bbox": [ + 186, + 112, + 235, + 152 + ], + "lines": [ + { + "bbox": [ + 186, + 112, + 235, + 152 + ], + "spans": [ + { + "bbox": [ + 186, + 112, + 235, + 152 + ], + "type": "image", + "image_path": "092201e40044eb18b472aff2e3d45ee923fa93e6f2442686fb0cfd4e6da75ffe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 185, + 154, + 235, + 194 + ], + "blocks": [ + { + "bbox": [ + 185, + 154, + 235, + 194 + ], + "lines": [ + { + "bbox": [ + 185, + 154, + 235, + 194 + ], + "spans": [ + { + "bbox": [ + 185, + 154, + 235, + 194 + ], + "type": "image", + "image_path": "327a90d47932adb9d915ddc7eb3a3ad1bc0b05b56c0420d4103c0ab617a4aec8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 196, + 225, + 205 + ], + "lines": [ + { + "bbox": [ + 192, + 196, + 225, + 205 + ], + "spans": [ + { + "bbox": [ + 192, + 196, + 225, + 205 + ], + "type": "text", + "content": "(c) 54-th" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 237, + 70, + 290, + 110 + ], + "blocks": [ + { + "bbox": [ + 237, + 70, + 290, + 110 + ], + "lines": [ + { + "bbox": [ + 237, + 70, + 290, + 110 + ], + "spans": [ + { + "bbox": [ + 237, + 70, + 290, + 110 + ], + "type": "image", + "image_path": "ae711ce1ed555676d531eb23e300e1b02ecd2431428108753f089e1892e1b992.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 238, + 112, + 290, + 152 + ], + "blocks": [ + { + "bbox": [ + 238, + 112, + 290, + 152 + ], + "lines": [ + { + "bbox": [ + 238, + 112, + 290, + 152 + ], + "spans": [ + { + "bbox": [ + 238, + 112, + 290, + 152 + ], + "type": "image", + "image_path": "8b4f9b714e28ca1922da243efb997a6e3051bf93c12b7ba9d2a32d6e39127d1d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 238, + 154, + 290, + 194 + ], + "blocks": [ + { + "bbox": [ + 238, + 154, + 290, + 194 + ], + "lines": [ + { + "bbox": [ + 238, + 154, + 290, + 194 + ], + "spans": [ + { + "bbox": [ + 238, + 154, + 290, + 194 + ], + "type": "image", + "image_path": "cf9a67bf073c1ded610eeb5a22466e8900ecd7cc57cfd505c8e1fd09c294776d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 196, + 279, + 205 + ], + "lines": [ + { + "bbox": [ + 247, + 196, + 279, + 205 + ], + "spans": [ + { + "bbox": [ + 247, + 196, + 279, + 205 + ], + "type": "text", + "content": "(d) Input" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 291, + 70, + 344, + 110 + ], + "blocks": [ + { + "bbox": [ + 291, + 70, + 344, + 110 + ], + "lines": [ + { + "bbox": [ + 291, + 70, + 344, + 110 + ], + "spans": [ + { + "bbox": [ + 291, + 70, + 344, + 110 + ], + "type": "image", + "image_path": "bce8aaa65c6a711d0d4ddaee6d0c74b15ad66b515074070b7373b7a77ed80899.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 291, + 112, + 344, + 152 + ], + "blocks": [ + { + "bbox": [ + 291, + 112, + 344, + 152 + ], + "lines": [ + { + "bbox": [ + 291, + 112, + 344, + 152 + ], + "spans": [ + { + "bbox": [ + 291, + 112, + 344, + 152 + ], + "type": "image", + "image_path": "dc31a9180b6675e0d7ce46b41415060d73f4098d1b87926f0675e98754f5bbc7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 291, + 154, + 344, + 194 + ], + "blocks": [ + { + "bbox": [ + 291, + 154, + 344, + 194 + ], + "lines": [ + { + "bbox": [ + 291, + 154, + 344, + 194 + ], + "spans": [ + { + "bbox": [ + 291, + 154, + 344, + 194 + ], + "type": "image", + "image_path": "ca35d23aa2af05ebfd10917394ad2b4bc5f4e8754c611e31ef0045d3a30fb186.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 287, + 196, + 394, + 206 + ], + "lines": [ + { + "bbox": [ + 287, + 196, + 394, + 206 + ], + "spans": [ + { + "bbox": [ + 287, + 196, + 394, + 206 + ], + "type": "text", + "content": "(e) Transweather (f) Zero-out" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 346, + 70, + 399, + 110 + ], + "blocks": [ + { + "bbox": [ + 346, + 70, + 399, + 110 + ], + "lines": [ + { + "bbox": [ + 346, + 70, + 399, + 110 + ], + "spans": [ + { + "bbox": [ + 346, + 70, + 399, + 110 + ], + "type": "image", + "image_path": "e68a7fd0f76b00c65f7c70edd7079a8b4f82f1b0025e1add05a0243fbcb59575.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 347, + 112, + 399, + 152 + ], + "blocks": [ + { + "bbox": [ + 347, + 112, + 399, + 152 + ], + "lines": [ + { + "bbox": [ + 347, + 112, + 399, + 152 + ], + "spans": [ + { + "bbox": [ + 347, + 112, + 399, + 152 + ], + "type": "image", + "image_path": "d4a08e5b5cf243f0d1f35b597ae6919de46668c3bdab2906ddd44afc2c5db0b6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 347, + 154, + 399, + 194 + ], + "blocks": [ + { + "bbox": [ + 347, + 154, + 399, + 194 + ], + "lines": [ + { + "bbox": [ + 347, + 154, + 399, + 194 + ], + "spans": [ + { + "bbox": [ + 347, + 154, + 399, + 194 + ], + "type": "image", + "image_path": "039335b1ea8d96969a6c0aed78049623686b69b442d3c5a2e8a314696fd99f31.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 196, + 394, + 206 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 196, + 394, + 206 + ], + "spans": [ + { + "bbox": [ + 348, + 196, + 394, + 206 + ], + "type": "text", + "content": "(f) Zero-out" + } + ] + } + ], + "index": 23, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 401, + 72, + 481, + 107 + ], + "blocks": [ + { + "bbox": [ + 401, + 72, + 481, + 107 + ], + "lines": [ + { + "bbox": [ + 401, + 72, + 481, + 107 + ], + "spans": [ + { + "bbox": [ + 401, + 72, + 481, + 107 + ], + "type": "image", + "image_path": "dadd85a57d3f5069af34ac9be41c970f73af73540b12034a152708f953edd466.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 107, + 482, + 115 + ], + "lines": [ + { + "bbox": [ + 481, + 107, + 482, + 115 + ], + "spans": [ + { + "bbox": [ + 481, + 107, + 482, + 115 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 402, + 109, + 481, + 145 + ], + "blocks": [ + { + "bbox": [ + 402, + 109, + 481, + 145 + ], + "lines": [ + { + "bbox": [ + 402, + 109, + 481, + 145 + ], + "spans": [ + { + "bbox": [ + 402, + 109, + 481, + 145 + ], + "type": "image", + "image_path": "e34f96cbf2cd6213d6f3e77f985627acd519882d1f4c6661f2e43e331bc4a336.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 116, + 482, + 124 + ], + "lines": [ + { + "bbox": [ + 481, + 116, + 482, + 124 + ], + "spans": [ + { + "bbox": [ + 481, + 116, + 482, + 124 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 125, + 482, + 133 + ], + "lines": [ + { + "bbox": [ + 481, + 125, + 482, + 133 + ], + "spans": [ + { + "bbox": [ + 481, + 125, + 482, + 133 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 134, + 482, + 141 + ], + "lines": [ + { + "bbox": [ + 481, + 134, + 482, + 141 + ], + "spans": [ + { + "bbox": [ + 481, + 134, + 482, + 141 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 142, + 482, + 149 + ], + "lines": [ + { + "bbox": [ + 481, + 142, + 482, + 149 + ], + "spans": [ + { + "bbox": [ + 481, + 142, + 482, + 149 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 402, + 147, + 484, + 196 + ], + "blocks": [ + { + "bbox": [ + 402, + 147, + 484, + 196 + ], + "lines": [ + { + "bbox": [ + 402, + 147, + 484, + 196 + ], + "spans": [ + { + "bbox": [ + 402, + 147, + 484, + 196 + ], + "type": "image", + "image_path": "a23680f075d1303654c8dd674959a887c0c172c5dd3c22c7e33e70ad4e46b73d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 196, + 475, + 206 + ], + "lines": [ + { + "bbox": [ + 419, + 196, + 475, + 206 + ], + "spans": [ + { + "bbox": [ + 419, + 196, + 475, + 206 + ], + "type": "text", + "content": "(g) Distribution" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 150, + 482, + 158 + ], + "lines": [ + { + "bbox": [ + 481, + 150, + 482, + 158 + ], + "spans": [ + { + "bbox": [ + 481, + 150, + 482, + 158 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 159, + 482, + 166 + ], + "lines": [ + { + "bbox": [ + 481, + 159, + 482, + 166 + ], + "spans": [ + { + "bbox": [ + 481, + 159, + 482, + 166 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 167, + 482, + 174 + ], + "lines": [ + { + "bbox": [ + 481, + 167, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 481, + 167, + 482, + 174 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 175, + 482, + 182 + ], + "lines": [ + { + "bbox": [ + 481, + 175, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 481, + 175, + 482, + 182 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 182, + 482, + 190 + ], + "lines": [ + { + "bbox": [ + 481, + 182, + 482, + 190 + ], + "spans": [ + { + "bbox": [ + 481, + 182, + 482, + 190 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 190, + 482, + 198 + ], + "lines": [ + { + "bbox": [ + 481, + 190, + 482, + 198 + ], + "spans": [ + { + "bbox": [ + 481, + 190, + 482, + 198 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 207, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 207, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 547, + 262 + ], + "type": "text", + "content": "Figure 2. Ineffective computations of Transweather [42]. We find the activity of some parameters in Transweather [42] are degradation dependent, and zero-outing computations for those inactivate parameters barely affect image restoration. (a) 7-th channel, (b) 10-th channel, and (c) 54-th channel of feature maps from Transweather, the brighter, the larger activation value, are activated for degradation of rain, raindrop, and snow. We show the (d) degraded images, (e) restorations from Transweather, and (f) restorations by zero-outing inactive channels. (g) From top to bottom, distribution of " + }, + { + "bbox": [ + 46, + 207, + 547, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 207, + 547, + 262 + ], + "type": "text", + "content": " norm for the three channels across all images in the All-weather dataset [21]." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 481, + 72, + 516, + 106 + ], + "blocks": [ + { + "bbox": [ + 481, + 72, + 516, + 106 + ], + "lines": [ + { + "bbox": [ + 481, + 72, + 516, + 106 + ], + "spans": [ + { + "bbox": [ + 481, + 72, + 516, + 106 + ], + "type": "image", + "image_path": "09e99b2af473f8159c811dd518168b10d5ff7d0b143fc5634ca03e03c03fc11e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 365 + ], + "type": "text", + "content": "1) The knowledge within the feature space of a pretrained vision-language (PVL) model can benefit various tasks, while its potential in our task is still under exploration. A straightforward way is to use the PVL model as the image degradation classifier. In contrast, we go one step further by reasoning diverse weather-specific knowledge from the feature space of the PVL model beyond the type of weather conditions." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 46, + 366, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 472 + ], + "type": "text", + "content": "We start by formulating a question prompt to query the occurrence, type, and severity of degradation in a degraded image. The obtained degradation prior describes what, where, and severity of degradations in high-level textual semantics. Then, we translate the high-level degradation prior to a 2D degradation map by aligning the prior with the degraded image. This degradation map provides a pixel-wise representation of the diverse knowledge of image degradation from the PVL model." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "content": "2) We then unleash the potential for various weather removal with the guidance of degradation maps. Observing model parameters are weather-specific, e.g., the rain-related parameters are usually inactive for unrelated degradations, and zero-outing the computations of unrelated parameters also barely affects the restoration quality. Take TUM [5] and Transweather [42] as examples, the PSNR and SSIM w/wo zero-outing on the All-weather dataset are " + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "inline_equation", + "content": "27.93 / 26.52" + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "content": " (dB) and " + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "inline_equation", + "content": "0.883 / 0.864" + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "inline_equation", + "content": "27.98 / 26.65" + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "content": " (dB) and " + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "inline_equation", + "content": "0.884 / 0.866" + }, + { + "bbox": [ + 46, + 474, + 287, + 675 + ], + "type": "text", + "content": ". This observation, illustrated in Fig. 2, inspired us to bypass computations for parameters unrelated to the specific weather type and severity during image restoration. With the assistance of MoE structure [28, 40], our LDR framework selects experts dynamically for restoration, therefore, ensuring adaptive learning of weather-specific knowledge which is not limited to a fixed network architecture." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Specifically, we maintain a candidate list of restoration experts and utilize the degradation map to sparsely select the most related restoration experts for each degra" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 270, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 316 + ], + "type": "text", + "content": "dition. By applying the selected expert pixel-wisely to restore weather-specific features, we create flexible and degradation-adaptive expert combinations/model architectures for restoration." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 304, + 318, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 401 + ], + "type": "text", + "content": "Though we can have a preliminary restoration from experts restored feature, considering image regions with similar values from the degradation map tend to benefit restoration features of each other, we re-use the degradation map to aggregate the restoration features, and improve the locality of the obtained restoration by a simple convolutional feedforward network. Our main contributions are:" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 306, + 405, + 545, + 486 + ], + "type": "list", + "angle": 0, + "index": 50, + "blocks": [ + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 545, + 427 + ], + "type": "text", + "content": "- We present an LDR framework to adaptively remove various adverse weather conditions in an all-in-one solution;" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 306, + 427, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 427, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 545, + 462 + ], + "type": "text", + "content": "- We propose a degradation map measurement module for extracting diverse weather-specific knowledge from a pre-trained vision-language model;" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 306, + 462, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 462, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 462, + 545, + 486 + ], + "type": "text", + "content": "- We propose a Top-K expert restoration module, sparsely and adaptively computing pixel-wise restoration features." + } + ] + } + ], + "index": 49 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 490, + 545, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 490, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 490, + 545, + 524 + ], + "type": "text", + "content": "The overall comparison in Fig. 1 shows the superiority of our framework for handling various degradations compared to state-of-the-art methods." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 306, + 537, + 392, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 392, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 392, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Adverse Weather Restoration. The field of adverse weather restoration encompasses two distinct approaches: task-specific methods [2, 3, 23, 51, 52] train the model independently for each individual degradation and all-in-one frameworks [5, 17, 21, 32, 35, 42, 44, 54, 57] develop a single unified model capable of handling various adverse weather degradations. A prevalent strategy for the all-in-one framework is to build upon task-specific models by employing multi-task learning techniques [44, 54, 57] or knowledge distillation [5] that consolidates multiple task-specific networks into a single network. Nevertheless, these methods typically rely on fixed computations, and the computation effectiveness is discounted as model parameters are" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24903" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 531, + 249 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 531, + 249 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 531, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 531, + 249 + ], + "type": "image", + "image_path": "e554615ba8772bfb9d1856ab3bef8f3b2dd8afe4b8f6430ef2653db0cf5cbbf4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": "Figure 3. The pipeline of our method. Given the input degraded image " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^d" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", we aim to recover the clean image " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^c" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". To tackle various adverse weather conditions, we format a question prompt " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", and query a pre-trained vision-language (PVL) model with " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^d" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", to estimate the degradation type and severity. The generated descriptions " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{txt}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " are transformed by a multilayer perceptron (MLP) to get the degradation prior " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{emb}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". Meanwhile, we extract a feature map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " from the input degraded image " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^d" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". A degradation map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " is computed by cross-attending " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " with the prior " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{emb}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", describing pixel-wise degradation pattern. We maintain a trainable candidate list of convolution filters, " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_n|n = 1,\\dots,N\\}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", and denote them as experts. We first parse the degradation map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " into a score map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " via a feedforward network " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "FFN^m" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " and a softmax layer, and then use " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " to find the best " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " experts describing the degradation of " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". Best experts are convolved with " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " to generate an intermediate feature map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{int}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". Finally, a cross-attention layer is used to aggregate " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{int}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " with the guidance of the degradation map " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ". We improve the feature locality with a feedforward network " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "FFN^{int}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": ", and the output " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": " is decoded to the clean image " + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^c" + }, + { + "bbox": [ + 46, + 250, + 545, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 249, + 373 + ], + "type": "text", + "content": "separately and degradation-dependently activated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 540 + ], + "type": "text", + "content": "Conversely, several methods [17, 35, 57] have been developed to overcome these limitations by focusing on learning weather-specific knowledge. They utilize indicators like degradation types [27, 35] or depth maps [57] during training to categorize and direct images to appropriate sub-networks for restoration. Though AirNet [17] learns weather-specific knowledge with contrastive learning. Its fixed sub-network designs can only restore images with certain degradation types, and does not generalize to unseen weather degradations, e.g., rain mixed with haze. Furthermore, these mentioned works overlook the fact that varying levels of degradation severity also warrant adaptively tailored computational processes, for a specific degradation type." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 590 + ], + "type": "text", + "content": "Diverging from previous methods, we harness PVL models to reason both the type and severity of degradations, enabling adaptive restoration with dynamic sub-networks informed by this rich, weather-specific knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Sparse Mixture of Expert. The concept of expert models [28] is defined as a subset of model parameters/computation, where each expert is specialized in handling distinct aspects of the input data. The sparse mixture of experts usually employs a routing mechanism [40] to dynamically and adaptively forward input to a subset of these experts, bypassing unnecessary and irrelevant computations. Given fixed computation costs, due to the sparsity, this framework can readily scale the number of experts to improve the model capability that has been widely verified" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 361, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 408 + ], + "type": "text", + "content": "in the domain of natural language processing [10, 40, 58] and computer vision [7, 11, 22, 30]. We study a prior derived from a PVL model to dynamically select experts, and adaptively apply expert to restore degraded images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 408, + 546, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 529 + ], + "type": "text", + "content": "Vision-language Model. With the release of ChatGPT, remarkable reasoning abilities of large language model (LLM) [16, 18, 25, 31, 39, 56] have been shown. It motivates the vision-language community to transfer the reasoning abilities to visual data. A common pipeline is to project an image to a joint space of LLM [19, 24, 25, 48, 49], and feed the projected image alongside user-provided text for conditional text generation. This paper proposes to leverage the PVL (e.g., [16]) to generate prior of an adverse weather degraded image for adaptive image restoration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 361, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": "Overview. Given an input image " + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": " degraded by adverse weather, we aim to recover the corresponding clean image " + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": ". Our method adaptively recovers " + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": " with degradation prior " + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 557, + 545, + 617 + ], + "type": "text", + "content": " obtained from a pre-trained vision-language (PVL) model. The overall architecture is shown in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": "To derive degradation prior " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": ", we query a PVL model with a question prompt " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " to reason about the occurrence, type, and severity of degradation in the input image " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": ", outputting " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{txt}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": ". Then a mapping network projects " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{txt}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": ". Meanwhile, we encode " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 617, + 545, + 677 + ], + "type": "text", + "content": " by using an encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We adaptively restore " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " in the embedding space, and decode " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " to a restored image " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{l}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ", with three steps: i) degradation map measurement, measuring degradation for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24904" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 279, + 178 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 279, + 178 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 279, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 279, + 178 + ], + "type": "image", + "image_path": "29e84c1c9975a12dcb97d88fd3be8ddacf2c2f37f104412b483ee2c619870020.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 179, + 287, + 201 + ], + "lines": [ + { + "bbox": [ + 47, + 179, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 47, + 179, + 287, + 201 + ], + "type": "text", + "content": "Figure 4. Examples text descriptions from LISA [16] for reasoning weather degraded images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": "each pixel of " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": " by using the degradation prior " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": "; ii) top-K expert restoration, selecting expert/parameters from a trainable candidate list of convolution filters according to the degradation map " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": ", and convolving with " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": " to get intermediate restoration features " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": "; iii) restoration feature aggregation, deriving " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": " by pixel-wisely aggregating " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": " with respect to the degradation map " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": ", and improving the feature locality with a feedforward network. Finally, " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": " is decoded to get the restored image " + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 215, + 288, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 333, + 157, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 333, + 157, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 333, + 157, + 346 + ], + "type": "text", + "content": "3.1. Degradation Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "content": "Degradation Prior Generation. We leverage the context learning capability of the PVL model (e.g., [16]) to reason diverse degradation knowledge of the degraded image " + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "content": " with a question prompt " + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "content": ". We format the question prompt " + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "content": " inspired by the chain-of-thought reasoning that makes the model to identify the occurrence, type, and severity of degradation. In Fig. 4, we provide the text description examples obtained by using different prompts. The generated descriptions " + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{txt}}" + }, + { + "bbox": [ + 46, + 353, + 289, + 461 + ], + "type": "text", + "content": " are defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 472, + 287, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 472, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 89, + 472, + 287, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {\\mathrm {t x t}} = \\operatorname {V L} \\left(\\mathbf {I} ^ {\\mathrm {c}}, \\mathbf {T}\\right), \\quad \\mathbf {P} ^ {\\mathrm {t x t}} \\in \\mathbb {R} ^ {L \\times C ^ {\\mathrm {v l}}}, \\tag {1}", + "image_path": "9dce510ccfb1d68f5af3ac2eddba4d6b319729215b632d3cd9f17f346cd9aff8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{VL}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "content": " is the PVL model, " + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "content": " is the description length, and " + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{vl}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "content": " is the channel dimension. To preserve the representation capabilities of PVL, we prune out the text description output layer of PVL, and use the embedding before the output layer as " + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{xt}}" + }, + { + "bbox": [ + 46, + 498, + 287, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "content": "Degradation Prior Embedding. We align " + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{txt}}" + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "content": " to the embedding space with size " + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "content": " of our restoration model by using a multilayer perceptron network MLP(" + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 46, + 563, + 287, + 600 + ], + "type": "text", + "content": "). We have" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 609, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 609, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 86, + 609, + 287, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {P} ^ {\\mathrm {e m b}} = \\operatorname {M L P} \\left(\\mathbf {P} ^ {\\mathrm {t x t}}\\right), \\quad \\mathbf {P} ^ {\\mathrm {e m b}} \\in \\mathbb {R} ^ {L \\times C}. \\tag {2}", + "image_path": "007bce3ec5e4d26f35e5d5bfc7c1edcff135de0fbdce1676c6d900c0b4c6fed6.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "text", + "content": " is then used as the degradation prior in our restoration model to adaptively recover " + }, + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 635, + 287, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 241, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 241, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 241, + 683 + ], + "type": "text", + "content": "3.2. Language-driven Restoration Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "content": "We use an encoder-decoder architecture [8] as a backbone, and adaptively recover " + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "content": " in the embedding space with " + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "content": " (the encoded embedding of " + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathrm{d}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 726 + ], + "type": "text", + "content": ") as inputs." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 71, + 539, + 188 + ], + "blocks": [ + { + "bbox": [ + 312, + 71, + 539, + 188 + ], + "lines": [ + { + "bbox": [ + 312, + 71, + 539, + 188 + ], + "spans": [ + { + "bbox": [ + 312, + 71, + 539, + 188 + ], + "type": "image", + "image_path": "98708d37090c1b36d206cd8a414ffca3b0bdd2d88fb505e013bcb077b684a235.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": "Figure 5. Visualization of " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{int}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ". Without cherry-picking, we show the channel with maximum gradients. From (b) " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": " to (c) " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ", degraded regions are highlighted, and regions with similar degradation severity have similar activation values on the feature map " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ". From (b) " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": " to (d) " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{int}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": " and (e) " + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 304, + 190, + 547, + 257 + ], + "type": "text", + "content": ", degradations are removed step by step." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": "Degradation Map Measurement. The degradation prior " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": " is text description related with high-level semantics. However, for " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": " with height " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": ", and channel " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": ", the degradation is usually pixel-wise and distinct in severity. To allow pixel-wise computation of adverse weather removal, we measure the 2D degradation map " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 268, + 546, + 353 + ], + "type": "text", + "content": " by the cross-attention mechanism," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 361, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 361, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 310, + 361, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} ^ {\\mathrm {q} _ {1}}, \\quad \\mathbf {K} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {k} _ {1}}, \\quad \\mathbf {V} = \\mathbf {P} ^ {\\text {e m b}} \\mathbf {W} ^ {\\mathrm {v} _ {1}}, \\tag {3}", + "image_path": "42e16ebeaacbe3ba1aff4ad96d11edd3c517a658ea022a8fa15e0d25581fdceb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 377, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 377, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 311, + 377, + 545, + 392 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\text {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {S o f t m a x} \\left(\\mathbf {Q} \\mathbf {K} ^ {\\top}\\right) \\mathbf {V}, \\tag {4}", + "image_path": "86deb9e9cf40c1f5c0ed485ac7b6d1044eb0eb3b5ddd35a21591c788981b2e5b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{q}_1}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{k}_1}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{v}_1} \\in \\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " are linear projection matrices for obtaining the query, key, and value. Our cross-attention mechanism first computes the semantic alignments between " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " via Softmax " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "(\\mathbf{QK}^\\top)" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": ", and then transforms " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\mathrm{emb}}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " into a 2D degradation map " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " that describes pixel-wise degradation of " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": ". We visualize " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 401, + 545, + 498 + ], + "type": "text", + "content": " in Fig. 5. The regions of similar degradation severity have been individually highlighted, by applying Eq. (4)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "Top-K Expert Restoration. Not all pixels degrade equally, and we adaptively restore each of them by using pixel-wise degradation knowledge within " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " candidate experts (convolution filters), " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_n|n = 1,\\dots,N\\}" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": ", for diverse adverse weather conditions, where each candidate " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": " is expertise at generating restoration features for specific degradation types or severities (see 4.2). With the degradation map " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": ", we select experts for each pixel of " + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Specifically, the degradation map " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " is fed to a feedforward network " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}^{\\mathrm{m}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", followed by a Softmax layer along the last dimension, to output a normalized pixel-wise selection score " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\operatorname{Softmax}(\\mathrm{FFN}^{\\mathrm{m}}(\\mathbf{M}))" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ". For a pixel with location " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "(i,j), i \\in [1,H], j \\in [1,W]" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(i,j) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(i,j,n)" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " measures how likely the " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "-th expert " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": " can be used to describe the degradation of " + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}(i,j)" + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "To find the best experts describing the degradation of " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}(i,j)" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": ", we compute the Top-K scores of " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(i,j)" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " and record" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24905" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 82, + 568, + 266 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 544, + 81 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 544, + 81 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 544, + 81 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the All-weather dataset. We respectively color the best and the second-best methods in red and blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 82, + 568, + 266 + ], + "lines": [ + { + "bbox": [ + 48, + 82, + 568, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 82, + 568, + 266 + ], + "type": "table", + "html": "
TypeMethodRainSnowRaindropAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
BestT + VL24.330.86028.470.87229.230.89527.340.876
BestT + GT27.040.91330.610.90031.630.93629.760.916
GeneralMPRNet [51]23.080.83927.690.84928.750.87926.510.856
NAFNet [3]23.210.84027.680.84728.900.89026.600.859
Uformer [45]22.930.83527.500.83828.510.87126.310.848
Restormer [52]23.370.84527.810.85029.100.89026.760.862
GRL [23]23.310.84227.790.84929.050.88826.720.860
All-in-OneAll-in-One [21]24.710.89828.330.88231.120.92728.050.902
AirNet [17]23.120.83727.920.85828.230.89226.420.862
TUM [5]23.920.85529.270.88430.750.91227.980.884
Transweather [42]23.180.84127.800.85428.980.90226.650.866
WDiff [32]26.180.90729.690.89329.710.91128.530.904
WGWS [57]25.310.90129.710.89431.310.93228.780.909
Ours26.920.91230.790.90531.540.93329.750.916
", + "image_path": "2a7f44d54aba1aaa8e8485e844850426bacf18577e8529aca9fcf86ebccbb85e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "content": "corresponding indices as the best experts. The best experts restored feature for location " + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 282, + 287, + 307 + ], + "type": "text", + "content": " is given by," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 312, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 287, + 346 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} ^ {\\text {i n t}} (i, j) = \\sum_ {k = 1} ^ {K} \\mathbf {S} (i, j, \\rho (k)) \\cdot \\mathcal {E} (i, j, \\rho (k)), \\tag {5}", + "image_path": "f660d48f4ebae6c0df1b8702188e1c0c4b2faaac4ded9fb3c33db429512fbfe1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 286, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 286, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 286, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {E} (i, j, \\rho (k)) = \\sum_ {\\Delta i, \\Delta j} \\mathbf {X} (i + \\Delta i, j + \\Delta j) \\cdot \\mathbf {F} _ {\\rho (k)} (\\Delta i, \\Delta j),", + "image_path": "4f4ea296ab0aff9b5b9677de39d5467c3936da01feae0d54fff1c7f390196448.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\rho (k)" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " is the index of the selected " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": "-th expert. " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(i,j,\\rho (k))" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " is the convolution result of " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{X}(i,j)" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": "-th expert " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\rho (k)}" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "(\\Delta i,\\Delta j)" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " iterates over the convolution kernel size. " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(i,j,\\rho (k))" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " are weights for prioritizing different experts. With our pixel-wise Top-K experts, the intermediate restoration feature " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " for each pixel of " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": " is adaptively generated. Practically, we find the best performance by region-wisely smoothing the expert selection score " + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": ", as pixels of a region frequently suffer the same degradation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "content": "Restoration Feature Aggregation. Pixels with similar degradation information in " + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "content": " potentially benefit each other in deriving the restoration " + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "content": ". We compute the compatibility between degradation prior " + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "content": " and restoration features " + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 488, + 287, + 559 + ], + "type": "text", + "content": ", and aggregate weighted restoration features pixel-wisely by using the cross-attention mechanism," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 287, + 579 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 287, + 579 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {M} \\mathbf {W} ^ {\\mathrm {q} _ {2}}, \\quad \\mathbf {K} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {k} _ {2}}, \\quad \\mathbf {V} = \\hat {\\mathbf {X}} ^ {\\mathrm {i n t}} \\mathbf {W} ^ {\\mathrm {v} _ {2}}, \\tag {6}", + "image_path": "d802a29773de3b6b8bd81cda33b5e24003853bf46b55e44ab47d870c96ed86ad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} = \\operatorname {F F N} ^ {\\operatorname {i n t}} (\\operatorname {A t t e n t i o n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V})) ， \\tag {7}", + "image_path": "7c06367a9871c2b1308c4c60bbc19bcd6d8d9188aa10e696b438cde1e8f8b667.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": " is a feedforward convolutional network to improve the feature locality. Finally, " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": " is decoded to the restoration " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": ". In Fig. 5, we compare " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": ", and find that " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": " is cleaner than " + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\mathrm{int}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 664 + ], + "type": "text", + "content": ", as pixels with similar degradation measurements benefit each other in the restorations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 135, + 682 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "We train our network with Charbonnier loss " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{char}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": " and gradient-level edge loss " + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{edge}}" + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": ", to penalize the deviation of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 282, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 282, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 545, + 307 + ], + "type": "text", + "content": "restoration from ground truth clean image, and encourage the consistent image gradients. We have " + }, + { + "bbox": [ + 305, + 282, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{char}}" + }, + { + "bbox": [ + 305, + 282, + 545, + 307 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 316, + 545, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 316, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 368, + 316, + 545, + 336 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c h a r}} = \\sqrt {\\left\\| \\mathbf {I} ^ {\\mathrm {c}} - \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\right\\| ^ {2} + \\varepsilon^ {2}}, \\tag {8}", + "image_path": "282bbb454276511a4e552d954fb6f77b9b807ab4f81531485783a01efd818eaa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "inline_equation", + "content": "\\varepsilon = 10^{-4}" + }, + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "text", + "content": " is a constant in all experiments. The " + }, + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{edge}}" + }, + { + "bbox": [ + 305, + 339, + 544, + 361 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 361, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 361, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 359, + 361, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e d g e}} = \\sqrt {\\| \\nabla \\mathbf {I} ^ {\\mathrm {c}} - \\nabla \\hat {\\mathbf {I}} ^ {\\mathrm {c}} \\| ^ {2} + \\varepsilon^ {2}}, \\tag {9}", + "image_path": "6e8e1e2c91ea572f40d9961add9b6a03fbb3ff11f6e4a1fe6eb8f3645350b86d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "spans": [ + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "text", + "content": " is the laplacian gradient operator. With a balance parameter " + }, + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 386, + 544, + 410 + ], + "type": "text", + "content": ", the total loss is given by" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 376, + 421, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 421, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 376, + 421, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c h a r}} + \\lambda \\mathcal {L} _ {\\text {e d g e}}. \\tag {10}", + "image_path": "aeaa90494537b4357711d188dcb539884a51a43b373b32bad8e6b6fdd4d8b1de.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 449, + 454, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 454, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 454, + 462 + ], + "type": "text", + "content": "4. Experiments and Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": "Implementation Details. Our model is implemented using the PyTorch framework and all experiments are conducted on an RTX A6000 GPU. We train our network with batch size 4 and " + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "inline_equation", + "content": "4 \\times 10^{6}" + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": " iterations using the ADAM optimizer. The learning rate is decayed from " + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": " by cosine annealing strategy. In training, images are randomly cropped to size " + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "inline_equation", + "content": "\\lambda = 0.05" + }, + { + "bbox": [ + 304, + 469, + 545, + 565 + ], + "type": "text", + "content": ". Code and more results are in our project page." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 567, + 546, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 567, + 546, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 567, + 546, + 687 + ], + "type": "text", + "content": "Datasets. Our experiments are conducted on both synthetic dataset and real dataset, i.e., All-weather dataset [21] and WeatherStream [53] dataset. All-weather dataset comprises 18,609 training images and 17,609 testing images, and is composed of subsets from Outdoor-rain [20], Snow100K-L [26], and Raindrop [37], corresponding to rain, snow, and raindrops weather conditions respectively. WeatherStream dataset is a real-world dataset that has three weather conditions, i.e., rain, snow, and fog, with a total of 176, 100 training images and 11, 400 testing images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 689, + 545, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 725 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 725 + ], + "type": "text", + "content": "Evaluation Metrics. We use averages of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation metrics (RGB-channel). The higher, the better." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24906" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 82, + 549, + 255 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 528, + 81 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 528, + 81 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 528, + 81 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the WeatherStream dataset. We color the best and the second-best methods in red and blue." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 82, + 549, + 255 + ], + "lines": [ + { + "bbox": [ + 48, + 82, + 549, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 82, + 549, + 255 + ], + "type": "table", + "html": "
TypeMethodRainHazeSnowAverage
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
GeneralBestT + VL21.200.78121.600.75520.320.77221.040.769
BestT + GT23.950.81022.970.80422.700.82823.210.814
MPRNet [51]21.500.79121.730.76320.740.80121.320.785
NAFNet [3]23.010.80322.200.80322.110.82622.440.811
Uformer [45]22.250.79118.810.76320.940.80120.670.785
Restormer [52]23.670.80422.900.80322.510.82822.860.812
GRL [23]23.750.80522.880.80222.590.82923.070.812
All-in-OneAll-in-One [21]--------
AirNet [17]22.520.79721.560.77021.440.81221.840.793
TUM [5]23.220.79522.380.80522.250.82722.620.809
Transweather [42]22.210.77222.550.77421.790.79222.180.779
WGWS [57]23.800.80722.780.80022.720.83123.100.813
Ours24.420.81823.110.80923.120.83823.550.822
", + "image_path": "329c0e3798bdeff94f76b7819b5e17286b09bfa6c8352bbcb1d3e4bc5862775c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 106, + 257, + 231, + 341 + ], + "blocks": [ + { + "bbox": [ + 106, + 257, + 231, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 257, + 231, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 257, + 231, + 341 + ], + "type": "image", + "image_path": "1d9bc51965978840ef6d5582de78dbdfd68821b87cb081524193373a2138b83b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 340, + 180, + 349 + ], + "lines": [ + { + "bbox": [ + 160, + 340, + 180, + 349 + ], + "spans": [ + { + "bbox": [ + 160, + 340, + 180, + 349 + ], + "type": "text", + "content": "Rain" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 106, + 349, + 230, + 432 + ], + "blocks": [ + { + "bbox": [ + 106, + 349, + 230, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 349, + 230, + 432 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 230, + 432 + ], + "type": "image", + "image_path": "bef62892f913d7239b3cb214139f16b73a560daaf49d6bf5def2cd37d60e6754.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 432, + 182, + 441 + ], + "lines": [ + { + "bbox": [ + 158, + 432, + 182, + 441 + ], + "spans": [ + { + "bbox": [ + 158, + 432, + 182, + 441 + ], + "type": "text", + "content": "Snow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 106, + 441, + 230, + 525 + ], + "blocks": [ + { + "bbox": [ + 106, + 441, + 230, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 441, + 230, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 230, + 525 + ], + "type": "image", + "image_path": "e18c2860f6e47e99f493a0340db23e531ae53cc74a923512b5046ac21a1a223b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 525, + 190, + 536 + ], + "lines": [ + { + "bbox": [ + 151, + 525, + 190, + 536 + ], + "spans": [ + { + "bbox": [ + 151, + 525, + 190, + 536 + ], + "type": "text", + "content": "Raindrop" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 539, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 46, + 539, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 545, + 561 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on the All-weather dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 238, + 258, + 294, + 294 + ], + "blocks": [ + { + "bbox": [ + 238, + 258, + 294, + 294 + ], + "lines": [ + { + "bbox": [ + 238, + 258, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 238, + 258, + 294, + 294 + ], + "type": "image", + "image_path": "3fd8796e1dc1d017400df17e2d48566b0c508e2ce3d0acf0f47f0ead1595ac10.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 294, + 276, + 303 + ], + "lines": [ + { + "bbox": [ + 253, + 294, + 276, + 303 + ], + "spans": [ + { + "bbox": [ + 253, + 294, + 276, + 303 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 238, + 304, + 294, + 339 + ], + "blocks": [ + { + "bbox": [ + 238, + 304, + 294, + 339 + ], + "lines": [ + { + "bbox": [ + 238, + 304, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 238, + 304, + 294, + 339 + ], + "type": "image", + "image_path": "bf982d708a9dba1011556f472890c0df3e7d3f92c55724f4e09b52e517f09347.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 339, + 283, + 350 + ], + "lines": [ + { + "bbox": [ + 246, + 339, + 283, + 350 + ], + "spans": [ + { + "bbox": [ + 246, + 339, + 283, + 350 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 302, + 258, + 358, + 294 + ], + "blocks": [ + { + "bbox": [ + 302, + 258, + 358, + 294 + ], + "lines": [ + { + "bbox": [ + 302, + 258, + 358, + 294 + ], + "spans": [ + { + "bbox": [ + 302, + 258, + 358, + 294 + ], + "type": "image", + "image_path": "bdcd9c91803fdd4afae70a3adc334e19270d035e86aedc912e0575983cbab8ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 294, + 337, + 303 + ], + "lines": [ + { + "bbox": [ + 322, + 294, + 337, + 303 + ], + "spans": [ + { + "bbox": [ + 322, + 294, + 337, + 303 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 302, + 303, + 358, + 339 + ], + "blocks": [ + { + "bbox": [ + 302, + 303, + 358, + 339 + ], + "lines": [ + { + "bbox": [ + 302, + 303, + 358, + 339 + ], + "spans": [ + { + "bbox": [ + 302, + 303, + 358, + 339 + ], + "type": "image", + "image_path": "75cc649564af012dfe80bf2e1d2da3af3779010ca756477e6f2839f660467a38.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 258, + 421, + 294 + ], + "blocks": [ + { + "bbox": [ + 365, + 258, + 421, + 294 + ], + "lines": [ + { + "bbox": [ + 365, + 258, + 421, + 294 + ], + "spans": [ + { + "bbox": [ + 365, + 258, + 421, + 294 + ], + "type": "image", + "image_path": "450cf566e3cfed04bd508527fee1e39efe2373c25fa380d80346860aea569b92.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 294, + 413, + 304 + ], + "lines": [ + { + "bbox": [ + 373, + 294, + 413, + 304 + ], + "spans": [ + { + "bbox": [ + 373, + 294, + 413, + 304 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 365, + 303, + 421, + 339 + ], + "blocks": [ + { + "bbox": [ + 365, + 303, + 421, + 339 + ], + "lines": [ + { + "bbox": [ + 365, + 303, + 421, + 339 + ], + "spans": [ + { + "bbox": [ + 365, + 303, + 421, + 339 + ], + "type": "image", + "image_path": "cf36925a2f00b9bd48e83404f1a227cd5bf7fd93427fc75499caf650f7458df7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 339, + 420, + 350 + ], + "lines": [ + { + "bbox": [ + 367, + 339, + 420, + 350 + ], + "spans": [ + { + "bbox": [ + 367, + 339, + 420, + 350 + ], + "type": "text", + "content": "WGWS [57]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 238, + 350, + 294, + 387 + ], + "blocks": [ + { + "bbox": [ + 238, + 350, + 294, + 387 + ], + "lines": [ + { + "bbox": [ + 238, + 350, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 238, + 350, + 294, + 387 + ], + "type": "image", + "image_path": "25b68cd19a746c4e08df840fe538f500d5c1b0e9826ab271c3bf389b1103f0ff.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 386, + 276, + 396 + ], + "lines": [ + { + "bbox": [ + 253, + 386, + 276, + 396 + ], + "spans": [ + { + "bbox": [ + 253, + 386, + 276, + 396 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 238, + 396, + 294, + 431 + ], + "blocks": [ + { + "bbox": [ + 238, + 396, + 294, + 431 + ], + "lines": [ + { + "bbox": [ + 238, + 396, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 238, + 396, + 294, + 431 + ], + "type": "image", + "image_path": "1000766bd9a8590ed0282a55c4fdcd7b5fc369d43415bedaa9c01135ca9ca5dc.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 302, + 350, + 358, + 387 + ], + "blocks": [ + { + "bbox": [ + 302, + 350, + 358, + 387 + ], + "lines": [ + { + "bbox": [ + 302, + 350, + 358, + 387 + ], + "spans": [ + { + "bbox": [ + 302, + 350, + 358, + 387 + ], + "type": "image", + "image_path": "ccaebe707ed23737ab05d91c3b7e6a34a80c3b5942fc088264f2a418c3877ee2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 386, + 337, + 396 + ], + "lines": [ + { + "bbox": [ + 323, + 386, + 337, + 396 + ], + "spans": [ + { + "bbox": [ + 323, + 386, + 337, + 396 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 302, + 396, + 358, + 432 + ], + "blocks": [ + { + "bbox": [ + 302, + 396, + 358, + 432 + ], + "lines": [ + { + "bbox": [ + 302, + 396, + 358, + 432 + ], + "spans": [ + { + "bbox": [ + 302, + 396, + 358, + 432 + ], + "type": "image", + "image_path": "e3d79301cebac642b483332aeb383ea71b14080ac7bd598ef6e891ad6dcf204d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 365, + 350, + 421, + 386 + ], + "blocks": [ + { + "bbox": [ + 365, + 350, + 421, + 386 + ], + "lines": [ + { + "bbox": [ + 365, + 350, + 421, + 386 + ], + "spans": [ + { + "bbox": [ + 365, + 350, + 421, + 386 + ], + "type": "image", + "image_path": "da9847f09628e6d8346b52afeae7798b8a183a6ef55e277982a62c5a0c5fea2d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 386, + 413, + 396 + ], + "lines": [ + { + "bbox": [ + 375, + 386, + 413, + 396 + ], + "spans": [ + { + "bbox": [ + 375, + 386, + 413, + 396 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 365, + 396, + 421, + 432 + ], + "blocks": [ + { + "bbox": [ + 365, + 396, + 421, + 432 + ], + "lines": [ + { + "bbox": [ + 365, + 396, + 421, + 432 + ], + "spans": [ + { + "bbox": [ + 365, + 396, + 421, + 432 + ], + "type": "image", + "image_path": "00f47c18f4b63113f207ae7feb128b414123411276deca765b2138d3d88c1a6f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 238, + 443, + 294, + 479 + ], + "blocks": [ + { + "bbox": [ + 246, + 433, + 283, + 443 + ], + "lines": [ + { + "bbox": [ + 246, + 433, + 283, + 443 + ], + "spans": [ + { + "bbox": [ + 246, + 433, + 283, + 443 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 443, + 294, + 479 + ], + "lines": [ + { + "bbox": [ + 238, + 443, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 238, + 443, + 294, + 479 + ], + "type": "image", + "image_path": "4478250649f099786679deb6dfdb9dc02f61aff497d06ff8d11c45092a59a876.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 479, + 276, + 488 + ], + "lines": [ + { + "bbox": [ + 253, + 479, + 276, + 488 + ], + "spans": [ + { + "bbox": [ + 253, + 479, + 276, + 488 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 238, + 488, + 294, + 524 + ], + "blocks": [ + { + "bbox": [ + 238, + 488, + 294, + 524 + ], + "lines": [ + { + "bbox": [ + 238, + 488, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 238, + 488, + 294, + 524 + ], + "type": "image", + "image_path": "8edc7f0add0e9455b57aa25d4642b4865743d83b82d3434ffed475e56fe9e990.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 525, + 283, + 536 + ], + "lines": [ + { + "bbox": [ + 246, + 525, + 283, + 536 + ], + "spans": [ + { + "bbox": [ + 246, + 525, + 283, + 536 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 302, + 443, + 358, + 479 + ], + "blocks": [ + { + "bbox": [ + 302, + 443, + 358, + 479 + ], + "lines": [ + { + "bbox": [ + 302, + 443, + 358, + 479 + ], + "spans": [ + { + "bbox": [ + 302, + 443, + 358, + 479 + ], + "type": "image", + "image_path": "a4853ebd2bc6f24f0bb634480920e6ffc484ac7a1ffb6720a5960ba6f9b8266c.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 479, + 337, + 488 + ], + "lines": [ + { + "bbox": [ + 323, + 479, + 337, + 488 + ], + "spans": [ + { + "bbox": [ + 323, + 479, + 337, + 488 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 302, + 488, + 358, + 524 + ], + "blocks": [ + { + "bbox": [ + 302, + 488, + 358, + 524 + ], + "lines": [ + { + "bbox": [ + 302, + 488, + 358, + 524 + ], + "spans": [ + { + "bbox": [ + 302, + 488, + 358, + 524 + ], + "type": "image", + "image_path": "6a56d06adbb5f28e51f7e092408bd41cfcbe1738ac9da525f2a8ad64f85b5ee4.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 365, + 443, + 421, + 479 + ], + "blocks": [ + { + "bbox": [ + 367, + 433, + 420, + 443 + ], + "lines": [ + { + "bbox": [ + 367, + 433, + 420, + 443 + ], + "spans": [ + { + "bbox": [ + 367, + 433, + 420, + 443 + ], + "type": "text", + "content": "WGWS [57]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 443, + 421, + 479 + ], + "lines": [ + { + "bbox": [ + 365, + 443, + 421, + 479 + ], + "spans": [ + { + "bbox": [ + 365, + 443, + 421, + 479 + ], + "type": "image", + "image_path": "1f5c895c271e6f9bf64865cc63f5d952f3ebdf99994e7eeacfc6892e9ae1b5f4.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 479, + 413, + 488 + ], + "lines": [ + { + "bbox": [ + 375, + 479, + 413, + 488 + ], + "spans": [ + { + "bbox": [ + 375, + 479, + 413, + 488 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 365, + 488, + 421, + 524 + ], + "blocks": [ + { + "bbox": [ + 365, + 488, + 421, + 524 + ], + "lines": [ + { + "bbox": [ + 365, + 488, + 421, + 524 + ], + "spans": [ + { + "bbox": [ + 365, + 488, + 421, + 524 + ], + "type": "image", + "image_path": "8e8553e5c9759111372d02df100f4d133f0e24da9f15e6eb655a8477cb12a273.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 525, + 420, + 536 + ], + "lines": [ + { + "bbox": [ + 367, + 525, + 420, + 536 + ], + "spans": [ + { + "bbox": [ + 367, + 525, + 420, + 536 + ], + "type": "text", + "content": "WGWS [57]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 428, + 258, + 485, + 293 + ], + "blocks": [ + { + "bbox": [ + 428, + 258, + 485, + 293 + ], + "lines": [ + { + "bbox": [ + 428, + 258, + 485, + 293 + ], + "spans": [ + { + "bbox": [ + 428, + 258, + 485, + 293 + ], + "type": "image", + "image_path": "2e91399c81ad3a95d8f04937690cb4b62344a9365acce3e0af01aa256ccc3069.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 293, + 483, + 304 + ], + "lines": [ + { + "bbox": [ + 436, + 293, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 436, + 293, + 483, + 304 + ], + "type": "text", + "content": "AirNet [17]" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 428, + 304, + 485, + 339 + ], + "blocks": [ + { + "bbox": [ + 428, + 304, + 485, + 339 + ], + "lines": [ + { + "bbox": [ + 428, + 304, + 485, + 339 + ], + "spans": [ + { + "bbox": [ + 428, + 304, + 485, + 339 + ], + "type": "image", + "image_path": "746421e067a8deb140344d20efab908d875e66d1b9fb89e35b09a830861cdf0a.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 428, + 350, + 485, + 386 + ], + "blocks": [ + { + "bbox": [ + 449, + 341, + 470, + 350 + ], + "lines": [ + { + "bbox": [ + 449, + 341, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 449, + 341, + 470, + 350 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 350, + 485, + 386 + ], + "lines": [ + { + "bbox": [ + 428, + 350, + 485, + 386 + ], + "spans": [ + { + "bbox": [ + 428, + 350, + 485, + 386 + ], + "type": "image", + "image_path": "e75a300df17f845fae0c9a91d5e942f752f9a60c9743d58fa69d1af3337803c7.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 386, + 482, + 396 + ], + "lines": [ + { + "bbox": [ + 436, + 386, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 436, + 386, + 482, + 396 + ], + "type": "text", + "content": "AirNet [17]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 428, + 396, + 485, + 432 + ], + "blocks": [ + { + "bbox": [ + 428, + 396, + 485, + 432 + ], + "lines": [ + { + "bbox": [ + 428, + 396, + 485, + 432 + ], + "spans": [ + { + "bbox": [ + 428, + 396, + 485, + 432 + ], + "type": "image", + "image_path": "86f487730909b5543e1da0353cd708a37362ec2dd9dfb63b7c85bb757bebd880.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 428, + 441, + 485, + 479 + ], + "blocks": [ + { + "bbox": [ + 449, + 434, + 470, + 442 + ], + "lines": [ + { + "bbox": [ + 449, + 434, + 470, + 442 + ], + "spans": [ + { + "bbox": [ + 449, + 434, + 470, + 442 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 441, + 485, + 479 + ], + "lines": [ + { + "bbox": [ + 428, + 441, + 485, + 479 + ], + "spans": [ + { + "bbox": [ + 428, + 441, + 485, + 479 + ], + "type": "image", + "image_path": "4052a061315f926ed695ff6dc611c1e89d1a907e90d706c9087197f6846a6015.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 479, + 482, + 488 + ], + "lines": [ + { + "bbox": [ + 436, + 479, + 482, + 488 + ], + "spans": [ + { + "bbox": [ + 436, + 479, + 482, + 488 + ], + "type": "text", + "content": "AirNet [17]" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 428, + 488, + 485, + 525 + ], + "blocks": [ + { + "bbox": [ + 428, + 488, + 485, + 525 + ], + "lines": [ + { + "bbox": [ + 428, + 488, + 485, + 525 + ], + "spans": [ + { + "bbox": [ + 428, + 488, + 485, + 525 + ], + "type": "image", + "image_path": "8ff53dd5f6d2f745eb867376cdd8908c8a38ddc6a01f17b59a7b4bffdd975abb.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 526, + 470, + 535 + ], + "lines": [ + { + "bbox": [ + 449, + 526, + 470, + 535 + ], + "spans": [ + { + "bbox": [ + 449, + 526, + 470, + 535 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 287, + 661 + ], + "type": "text", + "content": "Baseline Methods. We compare with the state-of-the-art (SOTA) general and all-in-one methods. For general methods, task-specific methods are trained with multi-task learning of different weather conditions, i.e., MPRNet [51], NAFNet [3], Uformer[45], Restormer [52], and GRL [23]. All-in-one methods are All-in-One [21], AirNet[17], TUM [5], Transweather [42], WDiff [32] and WGWS [57]." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 170, + 682 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. We compare with the SOTA general and all-in-one methods on the All-weather and" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 305, + 577, + 544, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 577, + 544, + 601 + ], + "spans": [ + { + "bbox": [ + 305, + 577, + 544, + 601 + ], + "type": "text", + "content": "- WeatherStream datasets in Tab. 1 and Tab. 2, respectively.\n- Our method achieves the best performance." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 712 + ], + "type": "text", + "content": "Furthermore, we create two strong baselines: i) BestT + VL. We use a PVL model for zero-shot adverse weather classification, and select the best task-specific methods for restoring degraded images; ii) BestT + GT. The ground truth adverse weather type is used to select the best task-specific method for image restoration. We find that i) Our method significantly outperforms 'BestT + VL', though it uses the same PVL as our model and uses multiple models for different adverse weather conditions. This shows" + } + ] + } + ], + "index": 58 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24907" + } + ] + } + ], + "index": 59 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 71, + 228, + 155 + ], + "blocks": [ + { + "bbox": [ + 104, + 71, + 228, + 155 + ], + "lines": [ + { + "bbox": [ + 104, + 71, + 228, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 71, + 228, + 155 + ], + "type": "image", + "image_path": "d16cb43e6bccc291ac61bb5942426b177579885b973c424d584a6c6f7aa08a42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 155, + 181, + 163 + ], + "lines": [ + { + "bbox": [ + 160, + 155, + 181, + 163 + ], + "spans": [ + { + "bbox": [ + 160, + 155, + 181, + 163 + ], + "type": "text", + "content": "Haze" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 292, + 109 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 292, + 109 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 292, + 109 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 292, + 109 + ], + "type": "image", + "image_path": "117e8ef003f55cd5ed0710c3026996cbce48fecef239179c657f42e708570a45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 109, + 276, + 118 + ], + "lines": [ + { + "bbox": [ + 253, + 109, + 276, + 118 + ], + "spans": [ + { + "bbox": [ + 253, + 109, + 276, + 118 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 72, + 356, + 109 + ], + "blocks": [ + { + "bbox": [ + 299, + 72, + 356, + 109 + ], + "lines": [ + { + "bbox": [ + 299, + 72, + 356, + 109 + ], + "spans": [ + { + "bbox": [ + 299, + 72, + 356, + 109 + ], + "type": "image", + "image_path": "c1f1a0967a242331f1bfb152d6738d4b6476e7de88c13a477e2752bd608f2653.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 362, + 72, + 419, + 109 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 419, + 109 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 419, + 109 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 419, + 109 + ], + "type": "image", + "image_path": "66e9655765fce5aace00d58e23c59b0b7069d561c1d9edb7bd011ee5eb83f77c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 109, + 413, + 118 + ], + "lines": [ + { + "bbox": [ + 375, + 109, + 413, + 118 + ], + "spans": [ + { + "bbox": [ + 375, + 109, + 413, + 118 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 483, + 109 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 483, + 109 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 483, + 109 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 483, + 109 + ], + "type": "image", + "image_path": "4b574b23598f722843db316a46317228e281c65e103f4adef20f0eef5d4069e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 109, + 481, + 118 + ], + "lines": [ + { + "bbox": [ + 436, + 109, + 481, + 118 + ], + "spans": [ + { + "bbox": [ + 436, + 109, + 481, + 118 + ], + "type": "text", + "content": "AirNet 17" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 104, + 165, + 228, + 247 + ], + "blocks": [ + { + "bbox": [ + 104, + 165, + 228, + 247 + ], + "lines": [ + { + "bbox": [ + 104, + 165, + 228, + 247 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 228, + 247 + ], + "type": "image", + "image_path": "1e1f8b717c1bb58edbfc0f078a414bccbc04a35d323065307500989f79644c4c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 247, + 180, + 255 + ], + "lines": [ + { + "bbox": [ + 160, + 247, + 180, + 255 + ], + "spans": [ + { + "bbox": [ + 160, + 247, + 180, + 255 + ], + "type": "text", + "content": "Rain" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 236, + 118, + 292, + 154 + ], + "blocks": [ + { + "bbox": [ + 236, + 118, + 292, + 154 + ], + "lines": [ + { + "bbox": [ + 236, + 118, + 292, + 154 + ], + "spans": [ + { + "bbox": [ + 236, + 118, + 292, + 154 + ], + "type": "image", + "image_path": "dc4e924bce113b90da8e98f486fd9873d4848c4d5a24ce2bcb34b182346f5290.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 155, + 283, + 163 + ], + "lines": [ + { + "bbox": [ + 246, + 155, + 283, + 163 + ], + "spans": [ + { + "bbox": [ + 246, + 155, + 283, + 163 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 297, + 118, + 356, + 154 + ], + "blocks": [ + { + "bbox": [ + 323, + 110, + 337, + 118 + ], + "lines": [ + { + "bbox": [ + 323, + 110, + 337, + 118 + ], + "spans": [ + { + "bbox": [ + 323, + 110, + 337, + 118 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 118, + 356, + 154 + ], + "lines": [ + { + "bbox": [ + 297, + 118, + 356, + 154 + ], + "spans": [ + { + "bbox": [ + 297, + 118, + 356, + 154 + ], + "type": "image", + "image_path": "296523e3f89a68c56fc1bff2d6f1e1e246a85d9706eb3e873c1b510cc76820e3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 156, + 348, + 163 + ], + "lines": [ + { + "bbox": [ + 299, + 156, + 348, + 163 + ], + "spans": [ + { + "bbox": [ + 299, + 156, + 348, + 163 + ], + "type": "text", + "content": "transweather" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 156, + 419, + 164 + ], + "lines": [ + { + "bbox": [ + 349, + 156, + 419, + 164 + ], + "spans": [ + { + "bbox": [ + 349, + 156, + 419, + 164 + ], + "type": "text", + "content": "[42] WGWS [57]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 363, + 118, + 419, + 154 + ], + "blocks": [ + { + "bbox": [ + 363, + 118, + 419, + 154 + ], + "lines": [ + { + "bbox": [ + 363, + 118, + 419, + 154 + ], + "spans": [ + { + "bbox": [ + 363, + 118, + 419, + 154 + ], + "type": "image", + "image_path": "c79966778d844b3c3d61e0458725134e979be1a5f021c399db8facd3c38fd8d3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 427, + 118, + 483, + 155 + ], + "blocks": [ + { + "bbox": [ + 427, + 118, + 483, + 155 + ], + "lines": [ + { + "bbox": [ + 427, + 118, + 483, + 155 + ], + "spans": [ + { + "bbox": [ + 427, + 118, + 483, + 155 + ], + "type": "image", + "image_path": "6e656e7b8c5b173f97b82aefc7a6af339ce182c16479540294d4dbb1a5110a73.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 156, + 470, + 163 + ], + "lines": [ + { + "bbox": [ + 450, + 156, + 470, + 163 + ], + "spans": [ + { + "bbox": [ + 450, + 156, + 470, + 163 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 104, + 257, + 228, + 340 + ], + "blocks": [ + { + "bbox": [ + 104, + 257, + 228, + 340 + ], + "lines": [ + { + "bbox": [ + 104, + 257, + 228, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 257, + 228, + 340 + ], + "type": "image", + "image_path": "825729df2f225578c745747c97d0e4c146a41a8ff19e94f1a02b7a0e33398154.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 342, + 182, + 350 + ], + "lines": [ + { + "bbox": [ + 158, + 342, + 182, + 350 + ], + "spans": [ + { + "bbox": [ + 158, + 342, + 182, + 350 + ], + "type": "text", + "content": "Snow" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 353, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 353, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 545, + 374 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on the WeatherStream dataset. The first column shows degraded images, while the crops for the bounding box regions of degraded images, ground truth, restoration from SOTA methods and our method are shown in the subsequent columns." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 236, + 165, + 292, + 202 + ], + "blocks": [ + { + "bbox": [ + 236, + 165, + 292, + 202 + ], + "lines": [ + { + "bbox": [ + 236, + 165, + 292, + 202 + ], + "spans": [ + { + "bbox": [ + 236, + 165, + 292, + 202 + ], + "type": "image", + "image_path": "ff7d85d49408d93081c2356ebdd41ed8752765f4340cd1b4e1708e92b32f9a99.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 201, + 276, + 209 + ], + "lines": [ + { + "bbox": [ + 254, + 201, + 276, + 209 + ], + "spans": [ + { + "bbox": [ + 254, + 201, + 276, + 209 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 427, + 165, + 483, + 201 + ], + "blocks": [ + { + "bbox": [ + 427, + 165, + 483, + 201 + ], + "lines": [ + { + "bbox": [ + 427, + 165, + 483, + 201 + ], + "spans": [ + { + "bbox": [ + 427, + 165, + 483, + 201 + ], + "type": "image", + "image_path": "82fcea633d4de1827e5d2e2e975514d96acad55762e73947c7ad70e0664a44b5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "type": "text", + "content": "AirNet 117" + } + ] + } + ], + "index": 25, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 236, + 209, + 292, + 247 + ], + "blocks": [ + { + "bbox": [ + 236, + 209, + 292, + 247 + ], + "lines": [ + { + "bbox": [ + 236, + 209, + 292, + 247 + ], + "spans": [ + { + "bbox": [ + 236, + 209, + 292, + 247 + ], + "type": "image", + "image_path": "223690d0cd8d9afcaef9ab4c9a7354ee9ca743f022a1c5806b7cbb95578a8b5c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 299, + 166, + 356, + 202 + ], + "blocks": [ + { + "bbox": [ + 299, + 166, + 356, + 202 + ], + "lines": [ + { + "bbox": [ + 299, + 166, + 356, + 202 + ], + "spans": [ + { + "bbox": [ + 299, + 166, + 356, + 202 + ], + "type": "image", + "image_path": "96937a15dfac9072b01a2b68630f51fb229260583fb992ab30e92f3461f8b84e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 201, + 336, + 209 + ], + "lines": [ + { + "bbox": [ + 323, + 201, + 336, + 209 + ], + "spans": [ + { + "bbox": [ + 323, + 201, + 336, + 209 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 363, + 166, + 419, + 202 + ], + "blocks": [ + { + "bbox": [ + 363, + 166, + 419, + 202 + ], + "lines": [ + { + "bbox": [ + 363, + 166, + 419, + 202 + ], + "spans": [ + { + "bbox": [ + 363, + 166, + 419, + 202 + ], + "type": "image", + "image_path": "89cff17d97eabdab1cc0cf6b352129ce062332911e58775e271664ecb92ac474.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 201, + 413, + 210 + ], + "lines": [ + { + "bbox": [ + 375, + 201, + 413, + 210 + ], + "spans": [ + { + "bbox": [ + 375, + 201, + 413, + 210 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 427, + 166, + 483, + 201 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 166, + 483, + 201 + ], + "spans": [ + { + "bbox": [ + 427, + 166, + 483, + 201 + ], + "type": "image", + "image_path": "9a911179e757235641ad551d56e66a140dd5d4f48ff7fc3224bc2b86aa87b006.jpg" + } + ] + } + ], + "index": 32, + "type": "text" + }, + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "spans": [ + { + "bbox": [ + 436, + 201, + 481, + 209 + ], + "type": "text", + "content": "AirNet [17]" + } + ] + } + ], + "index": 33, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 236, + 257, + 292, + 294 + ], + "blocks": [ + { + "bbox": [ + 246, + 248, + 282, + 257 + ], + "lines": [ + { + "bbox": [ + 246, + 248, + 282, + 257 + ], + "spans": [ + { + "bbox": [ + 246, + 248, + 282, + 257 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 257, + 292, + 294 + ], + "lines": [ + { + "bbox": [ + 236, + 257, + 292, + 294 + ], + "spans": [ + { + "bbox": [ + 236, + 257, + 292, + 294 + ], + "type": "image", + "image_path": "044a9834d14a234dab52255373b6f849874d9e4387debfa9243731b198e10cfb.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 254, + 294, + 276, + 302 + ], + "lines": [ + { + "bbox": [ + 254, + 294, + 276, + 302 + ], + "spans": [ + { + "bbox": [ + 254, + 294, + 276, + 302 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 299, + 210, + 356, + 247 + ], + "blocks": [ + { + "bbox": [ + 299, + 210, + 356, + 247 + ], + "lines": [ + { + "bbox": [ + 299, + 210, + 356, + 247 + ], + "spans": [ + { + "bbox": [ + 299, + 210, + 356, + 247 + ], + "type": "image", + "image_path": "042240e7dfc82474d53059be6ad0a5fdcbb627c00b4cdd2871f14d8a579bfe36.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 363, + 210, + 419, + 247 + ], + "blocks": [ + { + "bbox": [ + 363, + 210, + 419, + 247 + ], + "lines": [ + { + "bbox": [ + 363, + 210, + 419, + 247 + ], + "spans": [ + { + "bbox": [ + 363, + 210, + 419, + 247 + ], + "type": "image", + "image_path": "0c3c9b6392469622c85812006190f80acdc6de6c1d12779ae2cdb87f964a67ce.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 247, + 419, + 257 + ], + "lines": [ + { + "bbox": [ + 363, + 247, + 419, + 257 + ], + "spans": [ + { + "bbox": [ + 363, + 247, + 419, + 257 + ], + "type": "text", + "content": "] WGWS [57]" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 427, + 210, + 483, + 247 + ], + "blocks": [ + { + "bbox": [ + 427, + 210, + 483, + 247 + ], + "lines": [ + { + "bbox": [ + 427, + 210, + 483, + 247 + ], + "spans": [ + { + "bbox": [ + 427, + 210, + 483, + 247 + ], + "type": "image", + "image_path": "60a7727666d5dfdec14fde0859b741c9f8717fed0cf90d3afb8635bd922e2d62.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "spans": [ + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "type": "image", + "image_path": "22e39955c224010e104e718de0fd68409653a020d9e55c6198fc6fc5f28c3487.jpg" + } + ] + } + ], + "index": 41, + "type": "text" + }, + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "spans": [ + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 42, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 299, + 257, + 356, + 294 + ], + "blocks": [ + { + "bbox": [ + 299, + 257, + 356, + 294 + ], + "lines": [ + { + "bbox": [ + 299, + 257, + 356, + 294 + ], + "spans": [ + { + "bbox": [ + 299, + 257, + 356, + 294 + ], + "type": "image", + "image_path": "ddb873dbd66972312aabf83a658c35bd06a80e6dc637cd81696a6702fcaf6df8.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 294, + 336, + 302 + ], + "lines": [ + { + "bbox": [ + 322, + 294, + 336, + 302 + ], + "spans": [ + { + "bbox": [ + 322, + 294, + 336, + 302 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 363, + 257, + 419, + 294 + ], + "blocks": [ + { + "bbox": [ + 363, + 257, + 419, + 294 + ], + "lines": [ + { + "bbox": [ + 363, + 257, + 419, + 294 + ], + "spans": [ + { + "bbox": [ + 363, + 257, + 419, + 294 + ], + "type": "image", + "image_path": "3ecd56a84ef379a9e40e1f98acc661a41ded2951aaaff466e045e3b4ef852a35.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 294, + 413, + 303 + ], + "lines": [ + { + "bbox": [ + 375, + 294, + 413, + 303 + ], + "spans": [ + { + "bbox": [ + 375, + 294, + 413, + 303 + ], + "type": "text", + "content": "GRL [23]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 427, + 257, + 483, + 294 + ], + "blocks": [ + { + "bbox": [ + 449, + 249, + 470, + 257 + ], + "lines": [ + { + "bbox": [ + 449, + 249, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 449, + 249, + 470, + 257 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 257, + 483, + 294 + ], + "lines": [ + { + "bbox": [ + 427, + 257, + 483, + 294 + ], + "spans": [ + { + "bbox": [ + 427, + 257, + 483, + 294 + ], + "type": "image", + "image_path": "631df5213e5f17da325f08671401fd4a5c3c3adb9dc39e4ef203ae2e40fd6917.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 294, + 481, + 302 + ], + "lines": [ + { + "bbox": [ + 436, + 294, + 481, + 302 + ], + "spans": [ + { + "bbox": [ + 436, + 294, + 481, + 302 + ], + "type": "text", + "content": "AirNet [17]" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "spans": [ + { + "bbox": [ + 236, + 303, + 292, + 340 + ], + "type": "image", + "image_path": "22e39955c224010e104e718de0fd68409653a020d9e55c6198fc6fc5f28c3487.jpg" + } + ] + } + ], + "index": 49, + "type": "text" + }, + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "spans": [ + { + "bbox": [ + 246, + 341, + 283, + 350 + ], + "type": "text", + "content": "TUM [5]" + } + ] + } + ], + "index": 50, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 299, + 303, + 356, + 340 + ], + "blocks": [ + { + "bbox": [ + 299, + 303, + 356, + 340 + ], + "lines": [ + { + "bbox": [ + 299, + 303, + 356, + 340 + ], + "spans": [ + { + "bbox": [ + 299, + 303, + 356, + 340 + ], + "type": "image", + "image_path": "1fb2c4524ac119102ad155ecb9172af9d454965ddf05f27de3656ecbd4bf7290.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 362, + 303, + 419, + 339 + ], + "blocks": [ + { + "bbox": [ + 362, + 303, + 419, + 339 + ], + "lines": [ + { + "bbox": [ + 362, + 303, + 419, + 339 + ], + "spans": [ + { + "bbox": [ + 362, + 303, + 419, + 339 + ], + "type": "image", + "image_path": "53b169759a1bef939633ee0e16174aa93dcac9561002e2ae8911593b5d7672b7.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 339, + 419, + 350 + ], + "lines": [ + { + "bbox": [ + 363, + 339, + 419, + 350 + ], + "spans": [ + { + "bbox": [ + 363, + 339, + 419, + 350 + ], + "type": "text", + "content": "] WGWS [57]" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 427, + 303, + 483, + 340 + ], + "blocks": [ + { + "bbox": [ + 427, + 303, + 483, + 340 + ], + "lines": [ + { + "bbox": [ + 427, + 303, + 483, + 340 + ], + "spans": [ + { + "bbox": [ + 427, + 303, + 483, + 340 + ], + "type": "image", + "image_path": "51119fccec0d7a82626f46083a1398739ff3386cb10ec091afead246a011884c.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 342, + 470, + 350 + ], + "lines": [ + { + "bbox": [ + 449, + 342, + 470, + 350 + ], + "spans": [ + { + "bbox": [ + 449, + 342, + 470, + 350 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "bbox": [ + 46, + 391, + 286, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 286, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 286, + 499 + ], + "type": "text", + "content": "that non-trivial designs are required for effectively leveraging the PVL model in all-in-one adverse weather removal; ii) Although 'BestT + GT' outperforms most SOTA general and all-in-one methods, our method still achieves competitive performance compared to it. This indicates: 1) the benefit of accessing prior knowledge of adverse weather conditions, and 2) our method has learned the shared and weather-specific knowledge adaptively by selecting sparse experts dynamically." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 46, + 502, + 286, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 286, + 561 + ], + "type": "text", + "content": "Qualitative Comparison. We compare with the SOTA methods on restoring images degraded by different adverse weather conditions. The results are given in Fig. 6 and Fig. 7. Our method consistently restores clearer images than SOTA methods under different weather conditions." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 47, + 570, + 222, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 222, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 222, + 582 + ], + "type": "text", + "content": "4.2. Ablation studies and Discussions" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 46, + 589, + 286, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 286, + 611 + ], + "type": "text", + "content": "We validate the effectiveness and components of our framework on the All-weather dataset." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 46, + 615, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 286, + 674 + ], + "type": "text", + "content": "Model Architecture. We study the effectiveness of the degradation map measurement (DMM), top-K expert restoration (TER), and restoration feature aggregation (RFA) modules. The results are given in Tab. 3. The best performance is achieved by using all modules." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Degradation Severity. We use the PVL model to partition the All-weather dataset [21] into three subsets based on degradation severity: slight, moderate, and heavy. We" + } + ] + } + ], + "index": 62 + }, + { + "type": "table", + "bbox": [ + 309, + 401, + 544, + 483 + ], + "blocks": [ + { + "bbox": [ + 330, + 390, + 520, + 400 + ], + "lines": [ + { + "bbox": [ + 330, + 390, + 520, + 400 + ], + "spans": [ + { + "bbox": [ + 330, + 390, + 520, + 400 + ], + "type": "text", + "content": "Table 3. The effectiveness of our model components." + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 401, + 544, + 483 + ], + "lines": [ + { + "bbox": [ + 309, + 401, + 544, + 483 + ], + "spans": [ + { + "bbox": [ + 309, + 401, + 544, + 483 + ], + "type": "table", + "html": "
DMMTERRFAPSNRSSIM
X✓X27.930.882
XX✓28.370.889
✓✓X28.250.890
✓X✓29.110.902
X✓✓28.550.895
✓✓✓29.750.916
", + "image_path": "a7b47ae697cb6bb9483d04fe98d9016042a8c6192d05c7dc43cb57af609e6e56.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "table_body" + } + ], + "index": 64 + }, + { + "type": "table", + "bbox": [ + 307, + 517, + 545, + 578 + ], + "blocks": [ + { + "bbox": [ + 306, + 495, + 545, + 516 + ], + "lines": [ + { + "bbox": [ + 306, + 495, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 495, + 545, + 516 + ], + "type": "text", + "content": "Table 4. Comparisons on the All-weather dataset with degradation severity of slight, moderate, and heavy." + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 517, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 307, + 517, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 517, + 545, + 578 + ], + "type": "table", + "html": "
MethodSlightModerateHeavy
PSNRSSIMPSNRSSIMPSNRSSIM
AirNet27.590.89526.490.86524.500.818
WGWS29.740.92128.770.91027.140.886
Ours30.460.92529.780.91828.380.899
", + "image_path": "89a71fb760d52f7ff9454655c41dadc4539f504557a5b0f5642be3427b3bdb65.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "table_body" + } + ], + "index": 66 + }, + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 653 + ], + "type": "text", + "content": "then compare our results with those of AirNet and WGWS in Tab. 4. Our method shows significant improvement over AirNet [17] and WGWS [57] in the heavily degraded subset. This indicates the effectiveness of reasoning with diverse weather-specific knowledge, such as severity." + } + ] + } + ], + "index": 67 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Pixel-wise Expert. Different regions of the same degraded image often exhibit varying degrees of degradation severity, and should be adaptively and pixel-wisely restored. We use the degradation prior provided by the PVL model to select the same expert for the degraded image. It shows a" + } + ] + } + ], + "index": 68 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24908" + } + ] + } + ], + "index": 69 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 545, + 164 + ], + "type": "image", + "image_path": "5a066f61434721720a99a5df10ea3fce47225b5925c783fc1282b549410b910e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "type": "text", + "content": "Figure 8. The usage frequencies of experts for different degradation types. For example, Expert 8, 11, and 15 focuses on raindrop, rain, and snow, respectively. Expert 12 handles all types of degradation. Please refer to Fig. 9 for visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 195, + 153, + 298 + ], + "blocks": [ + { + "bbox": [ + 50, + 195, + 153, + 298 + ], + "lines": [ + { + "bbox": [ + 50, + 195, + 153, + 298 + ], + "spans": [ + { + "bbox": [ + 50, + 195, + 153, + 298 + ], + "type": "image", + "image_path": "f6a36680882ed0072f97a1328e6badd79ca2ffe41cf1878709f4de02cb3de923.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 300, + 131, + 310 + ], + "lines": [ + { + "bbox": [ + 86, + 300, + 131, + 310 + ], + "spans": [ + { + "bbox": [ + 86, + 300, + 131, + 310 + ], + "type": "text", + "content": "(a) Expert 8" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 195, + 282, + 298 + ], + "blocks": [ + { + "bbox": [ + 181, + 195, + 282, + 298 + ], + "lines": [ + { + "bbox": [ + 181, + 195, + 282, + 298 + ], + "spans": [ + { + "bbox": [ + 181, + 195, + 282, + 298 + ], + "type": "image", + "image_path": "97439ccd3be4d9464249385dcc176b90590c4284faf519b0290333e19af9a208.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 300, + 257, + 310 + ], + "lines": [ + { + "bbox": [ + 208, + 300, + 257, + 310 + ], + "spans": [ + { + "bbox": [ + 208, + 300, + 257, + 310 + ], + "type": "text", + "content": "(b) Expert 11" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 311, + 195, + 413, + 298 + ], + "blocks": [ + { + "bbox": [ + 311, + 195, + 413, + 298 + ], + "lines": [ + { + "bbox": [ + 311, + 195, + 413, + 298 + ], + "spans": [ + { + "bbox": [ + 311, + 195, + 413, + 298 + ], + "type": "image", + "image_path": "d893fe87b443140c9cd2a7f5fcf598724031af3f844b7ab110efb7775dc138d2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 300, + 386, + 310 + ], + "lines": [ + { + "bbox": [ + 336, + 300, + 386, + 310 + ], + "spans": [ + { + "bbox": [ + 336, + 300, + 386, + 310 + ], + "type": "text", + "content": "(c) Expert 12" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 441, + 195, + 544, + 298 + ], + "blocks": [ + { + "bbox": [ + 441, + 195, + 544, + 298 + ], + "lines": [ + { + "bbox": [ + 441, + 195, + 544, + 298 + ], + "spans": [ + { + "bbox": [ + 441, + 195, + 544, + 298 + ], + "type": "image", + "image_path": "e06a8ab188f3963f7618fb7b9a0b492a5655346cc8e9c218a345824d6e16d62a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 299, + 515, + 310 + ], + "lines": [ + { + "bbox": [ + 465, + 299, + 515, + 310 + ], + "spans": [ + { + "bbox": [ + 465, + 299, + 515, + 310 + ], + "type": "text", + "content": "(d) Expert 15" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 54, + 351, + 279, + 502 + ], + "blocks": [ + { + "bbox": [ + 46, + 314, + 547, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 314, + 547, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 547, + 336 + ], + "type": "text", + "content": "Figure 9. Sample image regions, activating different experts. We show regions with highest selection scores for experts 8, 11, 12, and 15 on the All-weather dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 351, + 279, + 502 + ], + "lines": [ + { + "bbox": [ + 54, + 351, + 279, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 279, + 502 + ], + "type": "image", + "image_path": "929219944645be22c84fe08df0039f92b9cac4ec21d2cd17fdad3ab2efafb640.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 502, + 102, + 511 + ], + "lines": [ + { + "bbox": [ + 59, + 502, + 102, + 511 + ], + "spans": [ + { + "bbox": [ + 59, + 502, + 102, + 511 + ], + "type": "text", + "content": "(a) Degraded" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 120, + 502, + 155, + 510 + ], + "lines": [ + { + "bbox": [ + 120, + 502, + 155, + 510 + ], + "spans": [ + { + "bbox": [ + 120, + 502, + 155, + 510 + ], + "type": "text", + "content": "(b) AirNet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 174, + 502, + 212, + 510 + ], + "lines": [ + { + "bbox": [ + 174, + 502, + 212, + 510 + ], + "spans": [ + { + "bbox": [ + 174, + 502, + 212, + 510 + ], + "type": "text", + "content": "(c) WGWS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 502, + 263, + 510 + ], + "lines": [ + { + "bbox": [ + 235, + 502, + 263, + 510 + ], + "spans": [ + { + "bbox": [ + 235, + 502, + 263, + 510 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 514, + 287, + 547 + ], + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 547 + ], + "type": "text", + "content": "Figure 10. From the first to last rows, the degradation is rain, raindrop, snow, and haze, respectively. From left to right, (a) degraded images are restored by (b) AirNet, (c) WGWS, and (d) our method." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 167, + 351, + 221, + 502 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 351, + 221, + 502 + ], + "spans": [ + { + "bbox": [ + 167, + 351, + 221, + 502 + ], + "type": "image", + "image_path": "adeb9b0728c899683ffdbee8460888db895db4696f5be0f262fb5d8d1eab75db.jpg" + } + ] + } + ], + "index": 14, + "type": "text" + }, + { + "bbox": [ + 222, + 351, + 278, + 502 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 351, + 278, + 502 + ], + "spans": [ + { + "bbox": [ + 222, + 351, + 278, + 502 + ], + "type": "image", + "image_path": "a9f68b20dad090947302ddd1052b00fdc81912f123fa0ac4dd8b1c308ceca7a6.jpg" + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "bbox": [ + 46, + 564, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 589 + ], + "type": "text", + "content": "0.33 dB/0.07 decrease in PSNR/SSIM, indicating the necessity of selecting experts pixel-wisely." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Adaptive Adverse Weather Removal. We study the model response to different adverse weather degradation by measuring the usage of our experts. Fig. 8 shows the usage for rain, snow, and raindrop. We find that there are experts for weather-specific knowledge and shared knowledge among degradation, such as Expert 8, Expert 11, Expert 12, and Expert 15. We visualize the image patches selecting the four experts in Fig. 9, where Expert 8, Expert 11, and Expert 15 are selected by image regions with raindrop, rain, and snow degradation, and Expert 12 is selected by image" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "spans": [ + { + "bbox": [ + 306, + 353, + 484, + 365 + ], + "type": "text", + "content": "regions with sky in all types of degradation." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 372, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 372, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 372, + 545, + 456 + ], + "type": "text", + "content": "Model Ability Under Complex Weather Condition. We test our model trained on the All-weather dataset with real images degraded by rain, raindrop, snow, and haze, as shown in Fig. 10. We compare our results with the two most competitive methods, AirNet and WGWS. Our model successfully disentangles weather-specific knowledge and generalizes to restore images degraded by haze." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 472, + 486, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 486, + 486 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 486, + 486 + ], + "type": "text", + "content": "5. Conclusion and Broader Impact" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 493, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 493, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 493, + 545, + 614 + ], + "type": "text", + "content": "We have proposed an LDR framework that adaptively removes various adverse weather conditions in an all-in-one solution. Our key insight is to leverage a pre-trained vision-language model to reason diverse weather-specific knowledge in a degraded image. We then use this knowledge to restore a clean image with three modules: degradation map measurement, Top-K expert restoration, and restoration feature aggregation. Experiments on standard benchmark datasets demonstrate that our method outperforms past works by a large margin." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 620, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 620, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 620, + 547, + 657 + ], + "type": "text", + "content": "Broader Impact. Our method is promising to be developed as an image restoration foundation model, prompting by degradation prior generated by a vision-language model." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part by the Beijing Institute of Technology Research Fund Program for Young Scholars, BIT Special-Zone, and National Natural Science Foundation of China 62302045." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24909" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12299-12310, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 203 + ], + "type": "text", + "content": "[2] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 182-192, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[3] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European Conference on Computer Vision, pages 17-33. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 316 + ], + "type": "text", + "content": "[4] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 394 + ], + "type": "text", + "content": "[5] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 451 + ], + "type": "text", + "content": "[6] Xiang Chen, Hao Li, Mingqiang Li, and Jinshan Pan. Learning a sparse transformer network for effective image deraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5896-5905, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 497 + ], + "type": "text", + "content": "[7] Zixiang Chen, Yihe Deng, Yue Wu, Quanquan Gu, and Yuanzhi Li. Towards understanding the mixture-of-experts layer in deep learning. Advances in neural information processing systems, 35:23049-23062, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 499, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 552 + ], + "type": "text", + "content": "[8] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 288, + 600 + ], + "type": "text", + "content": "[9] Keval Doshi and Yasin Yilmaz. Multi-task learning for video surveillance with limited data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3889-3899, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 667 + ], + "type": "text", + "content": "[10] Nan Du, Yanping Huang, Andrew M Dai, Simon Tong, Dmitry Lepikhin, Yuanzhong Xu, Maxim Krikun, Yanqi Zhou, Adams Wei Yu, Orhan First, et al. Glam: Efficient scaling of language models with mixture-of-experts. In International Conference on Machine Learning, pages 5547-5569. PMLR, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[11] Markus Enzweiler and Dariu M Gavrila. A multilevel mixture-of-experts framework for pedestrian classification. IEEE Transactions on Image Processing, 20(10):2967-2979, 2011. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[12] Sergio Escalera, Hugo Jair Escalante, Zhen Lei, Hao Fang, Ajian Liu, and Jun Wan. Surveillance face presentation attack detection challenge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6360-6370, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[13] Xueyang Fu, Jie Xiao, Yurui Zhu, Aiping Liu, Feng Wu, and Zheng-Jun Zha. Continual image deraining with hypergraph convolutional networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 229 + ], + "type": "text", + "content": "[14] Trung Hoang, Haichuan Zhang, Amirsaeed Yazdani, and Vishal Monga. Transer: Hybrid model and ensemble-based sequential learning for non-homogenous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1670-1679, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 285 + ], + "type": "text", + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 330 + ], + "type": "text", + "content": "[16] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. arXiv preprint arXiv:2308.00692, 2023. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "text", + "content": "[17] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17452-17462, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 442 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 488 + ], + "type": "text", + "content": "[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 544 + ], + "type": "text", + "content": "[20] Ruoteng Li, Loong-Fah Cheong, and Robby T Tan. Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1633-1642, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 600 + ], + "type": "text", + "content": "[21] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3175-3185, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 646 + ], + "type": "text", + "content": "[22] Xinjie Li and Huijuan Xu. Meid: mixture-of-experts with internal distillation for long-tailed video recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1451-1459, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[23] Yawei Li, Yuchen Fan, Xiaoyu Xiang, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18278-18289, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24910" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[24] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 161 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 206 + ], + "type": "text", + "content": "[26] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 287, + 262 + ], + "type": "text", + "content": "[27] Yulin Luo, Rui Zhao, Xiaobao Wei, Jinwei Chen, Yijie Lu, Shenghao Xie, Tianyu Wang, Ruiqin Xiong, Ming Lu, and Shanghang Zhang. Mowe: mixture of weather experts for multiple adverse weather removal. arXiv preprint arXiv:2303.13739, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 287, + 297 + ], + "type": "text", + "content": "[28] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 299, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 287, + 374 + ], + "type": "text", + "content": "[29] Hamam Mokayed, Amirhossein Nayebiastaneh, Kanjar De, Stergios Sozos, Olle Hagner, and Bjorn Backe. Nordic vehicle dataset (nvd): Performance of vehicle detectors using newly captured nvd from uav in different snowy weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5313-5321, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 420 + ], + "type": "text", + "content": "[30] Lynnette Hui Xian Ng and Kathleen M Carley. Botbuster: Multi-platform bot detection using a mixture of experts. In Proceedings of the International AAAI Conference on Web and Social Media, pages 686-697, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 49, + 422, + 287, + 464 + ], + "type": "text", + "content": "[31] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. Advances in neural information processing systems, 24, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 467, + 287, + 510 + ], + "type": "text", + "content": "[32] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 555 + ], + "type": "text", + "content": "[33] Liyuan Pan, Yuchao Dai, Miaomiao Liu, Fatih Porkikli, and Quan Pan. Joint stereo video deblurring, scene flow estimation and moving object segmentation. IEEE Transactions on Image Processing, 29:1748-1761, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 287, + 612 + ], + "type": "text", + "content": "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 613, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 613, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 49, + 613, + 287, + 677 + ], + "type": "text", + "content": "[35] Dongwon Park, Byung Hyun Lee, and Se Young Chun. All-in-one image restoration for unknown degradations using adaptive discriminative filters for specific degradations. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5815-5824. IEEE, 2023. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 680, + 287, + 713 + ], + "type": "text", + "content": "[36] Prashant W Patil, Sunil Gupta, Santu Rana, Svetha Venkatesh, and Subrahmanyam Murala. Multi-weather image restoration via domain translation. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF International Conference on Computer Vision, pages 21696-21705, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 149 + ], + "type": "text", + "content": "[37] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2482-2491, 2018. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 194 + ], + "type": "text", + "content": "[38] Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. Ffa-net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence, pages 11908-11915, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[40] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In International Conference on Learning Representations, 2016. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[41] Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927-1941, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 446 + ], + "type": "text", + "content": "[43] Yecong Wan, Yuanshuo Cheng, Mingwen Shao, and Jordi González. Image rain removal and illumination enhancement done in one go. Knowledge-Based Systems, 252:109244, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 502 + ], + "type": "text", + "content": "[44] Yinglong Wang, Chao Ma, and Jianzhuang Liu. Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3677-3686, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[45] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 612 + ], + "type": "text", + "content": "[46] Xuan Xiong, Yicheng Liu, Tianyuan Yuan, Yue Wang, Yilun Wang, and Hang Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17535-17544, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 656 + ], + "type": "text", + "content": "[47] Hao Yang, Dongming Zhou, Jinde Cao, and Qian Zhao. Dpnet: Detail-preserving image deraining via learning frequency domain knowledge. Digital Signal Processing, 130: 103740, 2022. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[48] Hao Yang, Liyuan Pan, Yan Yang, and Miaomiao Liu. Ldp: Language-driven dual-pixel image defocus deblurring network. arXiv preprint arXiv:2307.09815, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[49] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data pre-training. In Proceedings of the IEEE/CVF International" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24911" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 652 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 287, + 94 + ], + "type": "text", + "content": "Conference on Computer Vision (ICCV), pages 10699-10709, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[50] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[51] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 272 + ], + "type": "text", + "content": "[52] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 350 + ], + "type": "text", + "content": "[53] Howard Zhang, Yunhao Ba, Ethan Yang, Varan Mehra, Blake Gella, Akira Suzuki, Arnold Pfahnl, Chethan Chinder Chandrappa, Alex Wong, and Achuta Kadambi. Weatherstream: Light transport automation of single image deweathering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13499-13509, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 407 + ], + "type": "text", + "content": "[54] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 287, + 461 + ], + "type": "text", + "content": "[55] Zhihong Zhang, Yuxiao Cheng, Jinli Suo, Liheng Bian, and Qionghai Dai. Infwide: Image and feature space wiener deconvolution network for non-blind image deblurring in low-light conditions. IEEE Transactions on Image Processing, 32:1390-1402, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 507 + ], + "type": "text", + "content": "[56] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 509, + 287, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 584 + ], + "type": "text", + "content": "[57] Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21747-21758, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 586, + 287, + 652 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 586, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 48, + 586, + 287, + 652 + ], + "type": "text", + "content": "[58] Simiao Zuo, Qingru Zhang, Chen Liang, Pengcheng He, Tuo Zhao, and Weizhu Chen. Moebert: from bert to mixture-of-experts via importance-guided adaptation. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1610–1623, 2022. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24912" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_content_list.json b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..733a768c5750e2c44b9ff6535d70313b9689a860 --- /dev/null +++ b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_content_list.json @@ -0,0 +1,1940 @@ +[ + { + "type": "text", + "text": "Language-driven Grasp Detection", + "text_level": 1, + "bbox": [ + 310, + 130, + 658, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An Dinh Vuong $^{1}$ , Minh Nhat Vu $^{2,*}$ , Baoru Huang $^{3,*}$ , Nghia Nguyen $^{1}$ , Hieu Le $^{1}$ , Thieu Vo $^{4}$ , Anh Nguyen $^{5}$", + "bbox": [ + 86, + 179, + 919, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FPT Software AI Center, Vietnam $^{2}$ Automation & Control Institute, TU Wien, Austria $^{3}$ Imperial College London, UK", + "bbox": [ + 148, + 200, + 856, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Ton Duc Thang University, Vietnam $^{5}$ University of Liverpool, UK *Co-Corresponding authors", + "bbox": [ + 217, + 218, + 787, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://airvlab.github.io/grasp-anything", + "bbox": [ + 385, + 236, + 619, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e25ba918d52692bb9f2a2f4cef694e7122226554dfece7ff4a35af51422c547.jpg", + "image_caption": [ + "A coffee cup, a grape biscuit, a black oven plate on a white dining table", + "Grasp the coffee cup." + ], + "image_footnote": [], + "bbox": [ + 86, + 309, + 245, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7cfffa3db9198cd347875d81b81bd2d3792ab6bd0989dd95035c7de9dfa5b2a5.jpg", + "image_caption": [ + "A black pen, a steel marble and a digital calculator on an office desk", + "Grasp the calculator at its keypad.", + "Figure 1. We present a new dataset and method for language-driven grasp task." + ], + "image_footnote": [], + "bbox": [ + 246, + 309, + 405, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1f93e7891e5c35da4bbcb99c8af49cc675eee568a3a9ca6cd60f62d9ea6d614.jpg", + "image_caption": [ + "A black pot, a green bottle, a bread container arranged on a kitchen counter", + "Grab the neck of the green bottle." + ], + "image_footnote": [], + "bbox": [ + 408, + 309, + 563, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9bafff9a75f75acdec949e96af0d9469c76a55c459bdcda095c0afb8aa904730.jpg", + "image_caption": [ + "A brown-band wristwatch, a sleek black pen and an office clip placed on a table", + "Give me the brown-band wristwatch." + ], + "image_footnote": [], + "bbox": [ + 566, + 309, + 723, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3b83d90a97a9e2785ee86fd3954c919f23f96693945dc7bc25ba5b86f32d7e3.jpg", + "image_caption": [ + "Pick the fountain pen at its cap." + ], + "image_footnote": [], + "bbox": [ + 727, + 273, + 885, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 503, + 313, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Grasp detection is a persistent and intricate challenge with various industrial applications. Recently, many methods and datasets have been proposed to tackle the grasp detection problem. However, most of them do not consider using natural language as a condition to detect the grasp poses. In this paper, we introduce Grasp-Anything++, a new language-driven grasp detection dataset featuring 1M samples, over 3M objects, and upwards of 10M grasping instructions. We utilize foundation models to create a large-scale scene corpus with corresponding images and grasp prompts. We approach the language-driven grasp detection task as a conditional generation problem. Drawing on the success of diffusion models in generative tasks and given that language plays a vital role in this task, we propose a new language-driven grasp detection method based on diffusion models. Our key contribution is the contrastive training objective, which explicitly contributes to the denoising process to detect the grasp pose given the language instructions. We illustrate that our approach is theoretically supportive. The intensive experiments show that our method outperforms state-of-the-art approaches and allows real-world robotic grasping. Finally, we demonstrate our large-scale dataset enables zero-short grasp detection and is a challenging benchmark for future work.", + "bbox": [ + 75, + 521, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 503, + 630, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine we want an assistant robot to grasp a cup among a clutter of daily objects such as a knife, a fork, a cup, and a pair of scissors. Conventionally, to convey the idea of grasping this specific object, humans use the natural language command, \"give me the cup\", for instance. Although humans intuitively know how to grasp the cup given the linguistic command, determining specific grasp actions for objects based on natural language instructions or language-driven grasp detection remains challenging for robots [64]. First, natural language is usually overlooked in existing grasp datasets [56] while training vision-and-language neural networks necessitates an excessive number of labeled examples [66]. Second, recent works usually focus on particular manipulation tasks with limited objects [25], imposing a bottleneck for in-the-wild robot execution [63]. Finally, despite recent developments, bridging the gap between language, vision, and control for real-world robotic experiments remains a challenging task [83].", + "bbox": [ + 496, + 532, + 890, + 805 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, language-driven robotic frameworks are gaining traction, offering the potential for robots to process natural language, and bridging the gap between robotic manipulations and real-world human-robot interaction [51]. PaLM-E [21], EgoCOT [51], and ConceptFusion [32] are some notable embodied robots with the ability to compre", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17902", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hend natural language by harnessing the power of large foundation models such as ChatGPT [54]. However, most works assume the high-level actions of robots and ignore the fundamental grasping actions, restricting the structure for generalization across robotic domains, tasks, and skills [50]. In this paper, we explore training a language-driven agent to implement low-level actions, focusing on the task of object grasping via image observations. Specifically, our hypothesis is centered around the establishment of a robotic system that can execute grasping actions following a given language instruction for any universal object.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first present Grasp-Anything++ to serve as a large-scale dataset for language-driven grasp detection. Our dataset is based on the Grasp-Anything [74], and is synthesized from foundation models. Compared to the original Grasp-Anything dataset, we provide more than 10M grasp prompts and 3M associated object masks, 6M ground truth poses at the object part level. Our dataset showcases the ability to facilitate grasp detection using language instructions. We label the ground truth at both the object level and part level, providing a comprehensive understanding of real-world scenarios. For example, our ground truth includes both general instructions \"give me the knife\" and detail ones such as \"grasp the handle of the steak knife\". We empirically show that our large-scale dataset successfully facilitates zero-shot grasp detection on both vision-based tasks and real-world robotic experiments.", + "bbox": [ + 75, + 257, + 472, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenging language-driven grasp detection task, we propose a new diffusion model-based method. Our selection of diffusion models is motivated by their proven efficacy in conditional generation tasks [28]. These models have shown efficiency beyond image synthesis, including other image-based tasks such as image segmentation [77], and visual grounding [40]. Despite achieving notable success, integrating visual and text features effectively remains a challenge [13] as the majority of existing literature employs latent strategies to combine visual and text features [36]. We address this challenge by employing a new training strategy for learning text and image features, focusing on the use of feature maps as guidance information for grasp pose generation. Our main contribution is a new training objective that incorporates the feature maps and explicitly contributes to the denoising process. In summary, our contributions are three-fold:", + "bbox": [ + 75, + 500, + 472, + 755 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Grasp-Anything++, a large-scale language-driven dataset for grasp detection tasks.", + "- We propose a diffusion model with a training objective that explicitly contributes to the denoising process to detect the grasp poses.", + "- We demonstrate that our Grasp-Anything++ dataset and the proposed method outperform other approaches and enable successful robotic applications." + ], + "bbox": [ + 76, + 763, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Grasp Detection. Grasp detection is a popular task in both computer vision and robotic community [2, 17, 24, 64]. Recently, establishing robotic systems with the ability to follow natural commands has been actively researched [64, 80, 83]. The prevalent solution to the language-driven grasp detection task is to split into two stages: one for grounding the target object, and the other is to synthesize grasp poses from the grounding visual-text correlations [1, 80]. Training in two stages may result in longer inference time [41]. In addition, several works [82, 83, 90] adopt foundation models, such as GroundDINO [42] and GPT-3 [7]. Accessing such commercial foundation models is not always available [71], especially on robotic systems with limited resources or unstable internet connection [37]. In our work, we directly train the model on the large-scale Grasp-Anything++ dataset to inherit the power of a foundation-based dataset, while ensuring a straightforward inference process for the downstream robotic applications.", + "bbox": [ + 496, + 114, + 893, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-driven Grasp Detection Datasets. While there are many grasp datasets have been introduced [9, 17, 22-24, 33, 38, 45, 48, 49, 55, 78], the majority of them overlook the text modality. Therefore, the grasping of objects out of a clutter typically experiences ambiguities in what object to grasp [88]. DailyGrasp [83] is one of the first grasp datasets employing natural language for scene descriptions; however, the scene description corpus in this dataset is relatively small and does not specify which part of the object should be grasped. In our work, we present Grasp Anything++, which is a large-scale language-driven grasping dataset. Grasp-Anything++ provides object instance description with a natural word phrase, for example, \"a blue ceramic vase\". Furthermore, Grasp-Anything++ describes the grasping object at both the part level and object level, providing more information for the robot to execute the grasping [86].", + "bbox": [ + 496, + 387, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models for Robotic Applications. Diffusion models [28] have emerged as the new state-of-the-art method of generative tasks [84]. Recent have witnessed growing attention for utilizing diffusion models in robotic applications [61]. Liu et al. [43] propose a diffusion model to handle the language-guided object rearrangement task. Diffusion models are also applied to other robotic tasks such as motion planning [10], and trajectory optimization [31]. The authors in [70] present a diffusion model to determine grasp poses by minimizing a SDF loss. Overall, the diffusion models employed in previous works often combine visual and text features in a latent mechanism [6], which may cause interpretability problems [83] for robotic systems that require low-level controls [20]. To tackle this challenge, we propose a training objective that explicitly contributes to the denoising process. We demonstrate that our proposed strategy is more effective than the latent strategy.", + "bbox": [ + 496, + 643, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17903", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/c6b0f0d9dc25bce9ff0e4aba3c82a426edbb52f89db8874df2fc57e3eb3cef19.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StepDescription
Scene GenerationUserPlease help me generate scene descriptions for natural arrangements of daily objects. Each description has the following form: <Object_1><Object_2>...<Verb><Container Object>. Please also ensure the incorporation of a rich and varied lexicon in the scene descriptions.
SampleA steel knife, a polished fork and a pristine ceramic plate on a wooden table.
Text-to-ImageWe use Stable Diffusion [60] to proceed text-to-image generation.
Object MaskingUserFor object part-level description, given an input list {<Object_1>, <Object_2>, ...}, the output will be a list that describes the parts of objects as: {<Object_1>: [(Part_1.1>, <Part_1.2>, ..., <Object_2>]: [(Part_2.1>, <Part_2.2>, ...]}.
Sample{knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}
Post ProcessWe use OFA [75] and SAM [34] to locate the region describing the objects.
Part MaskingUserGiven the object list and part lists of each scene description, you will generate for me all prompts with the following format: {<Manipulation Action><Object_ID><Part_ID)}. The part that is more suitable for human grasping is positioned at the start of the list to represent the grasping actions.
SampleGive me the steel knife; Grasp the knife at its handle.
Post ProcessWe leverage VLPart [67] to locate the region describing the parts of objects.
Grasp GenerationUserGenerate for me a scene description with grasp instructions following the templates.
SampleScene description: A steel knife, a polished fork and a pristine ceramic plate on a wooden table. Object list: {knife, fork, plate}. Part lists: {knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}. Prompts: Give me the steel knife; Grasp the knife at its handle.
Grasp LabellingWe utilize a pretrained RAGT-3/3 [9] to generate grasp poses corresponding to the located region.
", + "bbox": [ + 76, + 88, + 890, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Grasp-Anything++ creation pipeline. We utilize ChatGPT to generate scene descriptions and grasp instructions from the user input. We generate images given scene descriptions and automatically synthesize the grasp poses.", + "bbox": [ + 76, + 491, + 892, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Grasp-Anything++ Dataset", + "text_level": 1, + "bbox": [ + 76, + 526, + 366, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We utilize large-scale foundation models to create the Grasp-Anything++. Our dataset offers open-vocabulary grasping commands and images with associated groundtruth. There are three key steps in establishing our dataset: i) prompting procedure, ii) image synthesis and grasp poses annotation, and iii) post-processing.", + "bbox": [ + 75, + 551, + 468, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Prompting Procedure", + "text_level": 1, + "bbox": [ + 76, + 650, + 279, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first establish prompt-based procedures to generate a large-scale scene description corpus as well as grasp prompt instructions. In particular, we utilize ChatGPT to generate the prompts for two tasks: $i$ ) Scene descriptions: Sentences capturing the scene arrangement, including the extracted object and part lists, and $ii$ ) Grasp instructions: Prompts directing the robot to grasp specific objects or parts.", + "bbox": [ + 75, + 674, + 468, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow a procedure in Table 1 to implement ChatGPT's output templates. The reference target in the grasp instruction may be either an object or an object's part. When the reference is an object's part, that part is directly selected as the reference in the grasp instruction sentence. If the reference is an object, we determine the grasping region on the part of the object that is likely to be grasped in everyday scenarios as described in affordance theory [53].", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Image Synthesis and Grasp Annotation", + "text_level": 1, + "bbox": [ + 500, + 527, + 836, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Synthesis. Given the scene description corpus, we first utilize a large-scale pretrained text-to-image model, namely, Stable Diffusion [60] to generate images from scene descriptions. Next, we perform a series of visual grounding and image segmentation using OFA [75], Segment-Anything [34], and VLPart [67] to locate the referenced object or part to the grasp instruction.", + "bbox": [ + 496, + 551, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Grasp Annotation. Grasp poses are represented as 2D rectangles, consistent with prior research and practical compatibility with real-world parallel grippers [17, 33]. Utilizing a pretrained network [9], we annotate grasp poses based on part segmentation masks. Since potential inaccuracies in these candidate poses could occur, we follow the procedure as defined in [74] to evaluate the quality of generated grasp poses to discard unreasonable grasp poses.", + "bbox": [ + 496, + 657, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, grasp quality is evaluated through net torque $\\mathcal{T} = (\\tau_{1} + \\tau_{2}) - RMg$ , where resistance at contact points is $\\tau_{i} = K\\mu_{s}F\\cos \\alpha_{i}$ . With constants such as $M$ (mass), $g$ (gravitational acceleration), $K$ (geometrical characteristics), $\\mu_{s}$ (static friction coefficient), and $F$ (applied force), accurately determining $\\mathcal{T}$ directly is challenging due to the physical difficulties in precisely measuring $M$ , $K$ , and $\\mu_{s}$ . Thus, we employ a surrogate measure,", + "bbox": [ + 496, + 780, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17904", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6dd8292cf455040829c3b39186f0b5587aa1a90bc9ba6d2451d5fef7ba4d0903.jpg", + "image_caption": [ + "(a) Number of categories" + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 330, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5a413908e008f1647111a74eddf433d39bffa972ced42a05b9d3dea412a1315.jpg", + "image_caption": [ + "(b) Scene description length", + "Figure 2. Data statistics. We analyze object categories and the use of natural language in the Grasp-Anything++ dataset." + ], + "image_footnote": [], + "bbox": [ + 334, + 95, + 584, + 200 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a122adec72ea413f0b6ecbafc6da97578f717f46af82b0f44385d364baa1b675.jpg", + "image_caption": [ + "(c) Number of object/part references." + ], + "image_footnote": [], + "bbox": [ + 588, + 90, + 893, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\tilde{\\mathcal{T}} = \\frac{\\cos\\alpha_1 + \\cos\\alpha_2}{R}$ , as an alternative. As a result in [12], antipodal grasps score higher on $\\tilde{\\mathcal{T}}$ , indicating better quality. Consequently, grasps are evaluated based on $\\tilde{\\mathcal{T}}$ , with positive values indicating positive grasps and others considered as negative.", + "bbox": [ + 75, + 236, + 467, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post Processing. Despite training on extensive datasets, Stable Diffusion [60] may produce subpar content, commonly termed as hallucination [30] when generating images from the text prompts. To address this, we perform manual reviews to filter out such images, with qualitative examples in our figures. Our process includes checks at every stage to prevent duplicate or hallucinated content. However, manual inspection introduces biases, which we counter with guidelines focusing on abnormal structures or implausible gravity (Fig. 3), aligns with approach in the literature [62].", + "bbox": [ + 75, + 324, + 467, + 474 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/00cfd26594998ed4dacfb405fc228105fffad00c13fc6abddd77333389d6a20f.jpg", + "image_caption": [ + "Figure 3. Failure image generation cases. Images generated by Stable Diffusion [60] may exhibit hallucinatory artifacts such as sunglasses lacking a lens, scissors with an anomalous structure, and a spoon not resting properly on a table." + ], + "image_footnote": [], + "bbox": [ + 78, + 478, + 202, + 575 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7efe25e835b004bb6cabb821598425187b0795efc3c7b6460d93172f8dfd0bce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 478, + 334, + 574 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/911907bda31ab4c17981e2704b4571e25edea51096c28024977411aed9caed9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 478, + 467, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, ChatGPT-generated scene prompts often duplicate [44]. To address this, we use duplication checking, filtering out identical prompts with BERTScore [89], which assesses sentence similarity through cosine similarities of token embeddings. We remove sentences with a BERTScore above 0.85 as in prior study [76].", + "bbox": [ + 75, + 654, + 467, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Data Statistics", + "text_level": 1, + "bbox": [ + 76, + 755, + 223, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Number of Categories. To evaluate the object category diversity, we apply a methodology akin to that in [16]. Utilizing 300 categories from LVIS dataset [26], we employ a pretrained model [91] to identify 300 candidate objects from our dataset for each category. We then curate a subset comprising 90,000 objects, refining it by excluding items that do not align semantically with their designated categories. A category is considered significant if it has more", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "than 40 objects. Fig. 2a shows the results. Overall, our dataset spans over 236 categories from LVIS dataset, indicating a notable degree of object diversity in our dataset.", + "bbox": [ + 496, + 239, + 890, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Descriptions. Fig. 2b shows the distribution of scene descriptions based on sentence length. The analysis reveals a wide range of sentence lengths, spanning from 10 words to 100 words per sentence. On average, each scene description consists of approximately 54 words, indicative of detailed and descriptive sentences. These scene descriptions correspond to sets of grasp instructions. Fig. 2c further shows the objects and object parts in scene descriptions.", + "bbox": [ + 496, + 284, + 890, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diversity Analysis. We assess the diversity of occlusion and lighting conditions in the dataset. Regarding occlusion, we use a pretrained YOLOv5 model to identify objects within images. The results indicate that $93.8\\%$ of images have a substantial overlap of five or more bounding boxes, which suggests a diverse range of occlusion within the Grasp-Anything++ dataset. Regarding lighting conditions, we convert images to YCbCr to analyze Y channel (luminance) and find that GraspL1M has the most diverse lighting conditions, identifying by the lowest Gini coefficient (a metric to measure the inequality of a distribution) of 0.26, compared to VMRD [87] (0.31), OCID-grasp [2] (0.32), Cornell [33] (0.62), Jacquard [17] (0.91).", + "bbox": [ + 496, + 405, + 890, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Language-driven Grasp Detection", + "text_level": 1, + "bbox": [ + 498, + 614, + 808, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivation. The use of diffusion model for language-driven grasp detection is motivated by its efficiency in various generative tasks [15,28,43,84]. Conditional generation, such as our language-driven grasp detection task, aligns seamlessly with diffusion models' capabilities [29]. Moreover, language-driven grasp detection represents a fine-grained problem in which the outputs strongly depend on the text input [4]. For example, \"grasp the steak knife\" and \"grasp the kraft knife\" refer to two different objects on the image. To this end, we propose using contrastive loss with diffusion model to tackle this task, as contrastive learning is a popular solution for fine-grained tasks [8, 14, 85].", + "bbox": [ + 496, + 640, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Constrastive Loss for Diffusion Model", + "text_level": 1, + "bbox": [ + 498, + 832, + 826, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We represent the target grasp pose as $\\mathbf{x}_0$ in the diffusion model. The objective of our diffusion process of language-driven grasp detection involves denoising from a noisy state", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{x}_T$ to the original grasp pose $\\mathbf{x}_0$ , conditioned on the input image and grasp instruction represented by $y$ .", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a diffusion process [28], assume that $q(\\mathbf{x}_{1:T}|\\mathbf{x}_0)$ is the forward process and we parameterize the reverse process by $p_{\\theta}(\\mathbf{x}_{0:T})$ . The conditional diffusion process [19] assumes $\\hat{q}$ is the forward process but with the inclusion of a condition $y$ . The goal of the reverse process is to optimize the variational bound on negative log likelihood [28]", + "bbox": [ + 75, + 121, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {T}) - \\sum_ {t \\geq 1} \\log \\frac {p _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t})}{q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1})} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 229, + 468, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We prove in the Appendix that", + "bbox": [ + 96, + 285, + 302, + 300 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} \\left[ \\underbrace {C - \\log p _ {\\theta} (\\mathbf {x} _ {T})} _ {\\text {C o n s t a n t}} + \\underbrace {\\log \\hat {q} (\\mathbf {x} _ {0} | \\mathbf {x} _ {T} , y)} _ {\\text {C o n t r a s t i v e}} + \\right. \\\\ \\underbrace {\\sum_ {t > 1} D _ {\\mathrm {K L}} \\left(\\hat {q} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t} , \\mathbf {x} _ {0} , y\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)\\right) - \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1} , y\\right)} _ {\\text {D e n o i s i n g s c o r e}}. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 318, + 482, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The terms $D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t))$ and $\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)$ in Equation 2 are similar to the concept of denoising score used in [28]. Thus, we can represent the quantity $\\sum_{t > 1}D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t))-$ $\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)$ by the loss utilized in [68, 69]", + "bbox": [ + 75, + 434, + 468, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0} | y), t \\sim [ 1, T ]} \\left[ \\mathbf {x} _ {0} - f \\left(\\mathbf {x} _ {t + 1}, t + 1, \\tilde {\\mathbf {x}} _ {0}\\right) \\right] ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 529, + 473, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $q(\\mathbf{x}_T|\\cdot)$ is equivalent to an isotropic Gaussian distribution as $T \\to +\\infty$ , the estimation quantity $\\log p_{\\theta}(\\mathbf{x}_T)$ converges to a constant when $\\theta \\to \\theta^*$ . Therefore, we can ignore the first term of Equation 2.", + "bbox": [ + 75, + 564, + 468, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the term $\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)$ of Equation 2 provides more information about the relation between $\\mathbf{x}_T$ and $\\mathbf{x}_0$ . As this term is intractable [65], we parameterize $\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)$ as $p_{\\psi ,y}(\\mathbf{x}_0,\\mathbf{x}_T)$ . This estimation resembles the noise-contrastive estimation [27], where the $\\mathbf{x}_T$ , $\\mathbf{x}_0$ can be considered as a pair of contrastive estimation and $\\psi$ can be estimated by a contrastive loss.", + "bbox": [ + 75, + 625, + 468, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In [19, 68, 73], the authors indicate that predicting $\\mathbf{x}_0$ is often infeasible but predicting an estimation $\\tilde{\\mathbf{x}}_0$ is tractable and can be used as a 'pseudo' estimation of $\\mathbf{x}_0$ . We denote $\\tilde{\\mathbf{x}}_0$ as an estimation of $\\mathbf{x}_0$ . The loss term $\\hat{q}(\\mathbf{x}_0 | \\mathbf{x}_T, y)$ can be approximated by using the following contrastive loss", + "bbox": [ + 75, + 729, + 468, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n t r a s t i v e}} = \\max \\left(0, \\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} - M\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 823, + 468, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M$ is the number of dimension of $\\mathbf{x}_0$ , and $\\alpha_{t}$ is the variance schedule at timestep $t$ ( $t = \\overline{1;T}$ ).", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. Suppose that $\\tilde{\\mathbf{x}}_0$ , $\\mathbf{x}_0$ and $\\epsilon$ are independent, and that", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} \\geq M.\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 119, + 779, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then there exists $C > 0$ such that: for arbitrary $\\delta > 0$ , if $\\mathcal{L}_{\\mathrm{contrastive}} < \\delta$ , then", + "bbox": [ + 498, + 162, + 893, + 193 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\| \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\| _ {2} ^ {2} \\right] < C \\delta .\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 205, + 774, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. See Supplementary Material.", + "bbox": [ + 500, + 237, + 741, + 252 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2eef1af9873495c5885bb62b027cc1295ddaef5beef475a9a959387855b8c009.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 237, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1.1. Proposition 1 suggests that if the contrastive loss $\\mathcal{L}_{\\mathrm{contrastive}}$ tends to zero, then the prediction $\\tilde{\\mathbf{x}}_0$ will approach the ground truth $\\mathbf{x}_0$ .", + "bbox": [ + 500, + 268, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1.2. The condition $\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 \\geq M$ is suitable for our language-driven grasp detection task as $\\tilde{\\mathbf{x}}_0$ and $\\mathbf{x}_T$ are two contrastive quantities, therefore, we can assume there is a minimum distance between $\\tilde{\\mathbf{x}}_0$ and $\\mathbf{x}_T$ . In addition, in the proof of Proposition 1, we see that $\\mathbb{E}\\left[\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 - M\\right] = \\beta^2\\mathbb{E}\\left[\\| \\tilde{\\mathbf{x}}_0 - \\mathbf{x}_0\\|_2^2\\right]$ , which is always nonnegative. Therefore, it is both theoretically and experimentally reasonable to add this assumption.", + "bbox": [ + 498, + 318, + 890, + 464 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5f8cf0da96fc9cb60067e238944d2fab949583d467d2ea986898423b76b7ced.jpg", + "image_caption": [ + "Figure 4. Language-drive Grasp Detection (LGD) network. We present the network architecture (left) and the proposed training objectives of the denoising process (right)." + ], + "image_footnote": [], + "bbox": [ + 506, + 479, + 908, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Language-driven Grasp Detection Network", + "text_level": 1, + "bbox": [ + 498, + 755, + 867, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network. Our network operates on two conditions: an image denoted as $\\pmb{I}$ and a corresponding text prompt represented as $e$ . To process these conditions, we employ a vision encoder to extract visual features from $\\pmb{I}$ and a text encoder to derive textual embeddings from $e$ . The resulting feature vectors, denoted as $I'$ and $e'$ , are subsequently subjected to a fusion module, ALBEF [39]. We leverage the attention mask generated by the ALBEF fusion module", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17906", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/31af37597999a812ea03d30352742866a3875271953175cf6a4ffff0652e4476.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 102, + 207, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/853a17736d910c97ba413f075bf351c062d98c253545ebe034163a21153f2cc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 102, + 344, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ef3066826567afd9937574af595edf403a795ffbf49d6fe73a2884db31ab644.jpg", + "image_caption": [ + "Give me the bottle" + ], + "image_footnote": [], + "bbox": [ + 351, + 102, + 480, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d040c5158a765ad1a6044d085d67f7decc812517f74e1524a65a4f6f74cba34f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 102, + 616, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c175ad61339879b6e2afa8d1769ec697bffea22e87d70e9611a8752a03d954b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 102, + 753, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14bec2ca7857c1b62b3ceaea57a18780bc18796d6d727500f5815f31d4d1799b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 102, + 890, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0b133780a36ae1595fe6d8c5715704ebd2fbdd2a08ca50223b25be31d333f6e.jpg", + "image_caption": [ + "(a) Ours" + ], + "image_footnote": [], + "bbox": [ + 78, + 219, + 207, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/989181ff313a07de2bfb6e230168a17bc2ccab2c0b3677013935364ead3ea5cb.jpg", + "image_caption": [ + "(b) GR-ConvNet", + "Figure 5. Language-driven grasp detection results visualization." + ], + "image_footnote": [], + "bbox": [ + 215, + 219, + 344, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e2945012a302f4787aa6e85ec085254518d9e7de32d80a98e9c1f9c0a09dbc6.jpg", + "image_caption": [ + "Grasp the wristwatch at its dial", + "(c) Det-Seg-Refine" + ], + "image_footnote": [], + "bbox": [ + 352, + 219, + 480, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f21faf38cd302e13e459d2bece98cbb230586d4b6bb03c29d368ec85a62c9a3.jpg", + "image_caption": [ + "(d) GG-CNN" + ], + "image_footnote": [], + "bbox": [ + 488, + 219, + 616, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/60222a5d53b79fdec6ca8b5bcbbefffec84e1853bb0ecfe3c95933e1258ff630.jpg", + "image_caption": [ + "(e) CLIPORT" + ], + "image_footnote": [], + "bbox": [ + 625, + 219, + 753, + 318 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/468e895b375c80b4d2740a82d90feffd28d1d99d02bcbc9449a60dd580697f48.jpg", + "image_caption": [ + "(f) CLIP-Fusion" + ], + "image_footnote": [], + "bbox": [ + 763, + 219, + 890, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as the estimation $\\tilde{\\mathbf{x}}_0$ of $\\mathbf{x}_0$ . Next, we aggregate three elements: the estimation region $\\tilde{\\mathbf{x}}_0$ , the grasp pose at the current timestep $\\mathbf{x}_{t + 1}$ , and the timestep $t + 1$ . These inputs are combined using MLP layers, similar to the approach outlined in [68]. Specifically, the output operation can be expressed as: $\\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0)$ , where function $f$ encompasses a composition of multiple MLP layers. Additional specifics regarding these universal MLP layers are provided in the Supplementary Material.", + "bbox": [ + 75, + 353, + 468, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Objective. In our context, conditioned grasp detection models the distribution $p(\\mathbf{x}_0|y)$ as the reversed diffusion process of gradually cleaning $\\mathbf{x}_{t + 1}$ . Instead of predicting $\\mathbf{x}_t$ as formulated by [28], we follow Ramesh et al. [58] and predict the signal itself, i.e., $\\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0)$ with the simple objective [28]. To this end, we utilize the contrastive loss as in Equation 4 to explicitly improve the learning objective of the denoising process:", + "bbox": [ + 75, + 489, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n t r a s t i v e}} + \\mathcal {L} _ {\\text {d i f f u s i o n}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 628, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments to evaluate our proposed method and Grasp-Anything++ dataset using both the vision-based metrics and real robot experiments. We then demonstrate zero-shot grasp results and discuss the challenges and open questions for future works.", + "bbox": [ + 75, + 684, + 468, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Language-driven Grasp Detection Results", + "text_level": 1, + "bbox": [ + 76, + 770, + 434, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our language-driven grasp detection method (LGD) with the linguistically supported versions of GR-CNN [35], Det-Seg-Refine [2], GG-CNN [47], CLIPORT [64] and CLIP-Fusion [80]. In all cases, we employ a pretrained CLIP [57] or BERT [18] as the text embedding. The implementation details of all baselines can be found in our Supplementary Material.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/38361574347e58d03965137c19d3fab300af94d36904fa32e6c47cb747e3173b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineSeenUnseenH
GR-ConvNet [35] + CLIP [57]0.370.180.24
Det-Seg-Refine [2] + CLIP [57]0.300.150.20
GG-CNN [47] + CLIP [57]0.120.080.10
CLIPORT [64]0.360.260.29
CLIP-Fusion [80]0.400.290.33
LGD (ours) + BERT [18]0.440.380.41
LGD (ours) + CLIP [57]0.480.420.45
", + "bbox": [ + 517, + 349, + 883, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. To assess the generalization of all methods trained on Grasp-Anything++, we utilize the concept of base and new labels [92] in zero-shot learning. We categorize LVIS labels from Section 3.3 to form labels for our experiment. In particular, we select $70\\%$ of these labels by frequency for 'Base' and assign the remaining $30\\%$ to 'New'. We also use the harmonic mean ('H') to measure the overall success rates [92]. Our primary evaluation metric is the success rate, defined similarly to [35], necessitating an IoU score of the predicted grasp exceeding $25\\%$ with the ground truth grasp and an offset angle less than $30^{\\circ}$ .", + "bbox": [ + 496, + 523, + 890, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. Table 2 shows the results of language-driven grasp detection on the Grasp-Anything++ dataset. The findings indicate a notable performance advantage of our LGD over other baseline approaches, with LGD outperforming the subsequent best-performing baselines (CLIP-Fusion) by margins of 0.14 on Grasp-Anything++ dataset.", + "bbox": [ + 496, + 691, + 890, + 781 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47352291292a3c0213f5b552b9967c29ec62d218fb646a89b1c718626a560a2a.jpg", + "table_caption": [ + "Table 2. Language-driven grasp detection results." + ], + "table_footnote": [], + "table_body": "
BaselineSeenUnseenH
LGD w/o predicting x̂00.150.080.10
LGD w/o contrastive loss0.450.400.42
LGD w contrastive loss0.480.420.45
", + "bbox": [ + 537, + 795, + 862, + 873 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Contrastive loss analysis.", + "bbox": [ + 589, + 877, + 800, + 891 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c307034979a56be59006f6bf22980e82842e7b9f5c980ae459ba3e08fd0bf9c.jpg", + "image_caption": [ + "Figure 6. Loss visualization." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 464, + 215 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c68367085b7a413e767eb39182d92b6b2fad4e5e1a8b4884bb40ecec9166f4c.jpg", + "image_caption": [ + "Figure 7. t-SNE visualization. We apply t-SNE to cluster the vision-and-language features with and without contrastive loss." + ], + "image_footnote": [], + "bbox": [ + 81, + 243, + 274, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34186e8bce2d7cb6da9de2e05cece17a842747e303536e459d5f00a33de8d4db.jpg", + "image_caption": [ + "Figure 8. Attention map visualization. We compare the attention map when utilizing our proposed contrastive loss and when not." + ], + "image_footnote": [], + "bbox": [ + 276, + 243, + 467, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contrastive Loss Analysis. Table 3 presents the performance of LGD under varied configurations. The outcomes emphasize the substantial influence of the training objective (contrastive loss) and the importance of language instructions in enhancing LGD performance on both seen and unseen classes in the Grasp-Anything++ dataset.", + "bbox": [ + 75, + 441, + 468, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 shows the contrastive loss $\\mathcal{L}_{\\mathrm{contrastive}}$ approaching towards 0 during training, indicating the grasp pose estimation $\\tilde{\\mathbf{x}}_0$ aligns with the ground truth $\\mathbf{x}_0$ , as anticipated by Proposition 1. The subsequent attention maps visualization in Fig. 8 shows the attention region is meaningful and improves the results when employing our proposed contrastive loss compared to its absence. Moreover, we employ t-SNE for vision-and-language embedding visualization, as in [46], by processing 2,000 samples from the Grasp-Anything++ dataset through the ALBEF module. The outcomes reveal that our contrastive loss facilitates better object classification, as evidenced in Fig.7 by clearer segregation of pixel embeddings across various semantic classes, underscoring contrastive loss's role in refining embeddings' differentiation for improved class distinctions.", + "bbox": [ + 75, + 535, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 5 presents qualitative results of the language-driven grasp detection task, suggesting that our LGD method generates more semantically plausible than other baselines. Despite satisfactory performance, LGD occasionally predicts incorrect results, with a detailed analysis of these cases available in our Appendix.", + "bbox": [ + 75, + 763, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robotic Validation. We provide quantitative results by integrating our language-driven grasp detection pipeline for a robotic grasping application with a KUKA LBR iwa", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0dcadfe26b1b66087528e78d00518a1f905d590b7151ce2af416a6b221d0d2b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4eb7ccd31a7cfffba546a11c2584d1794f4fd7d5989136f9a19e70a4f6f48bd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineSingleCluttered
GR-ConvNet [35] + CLIP [57]0.330.30
Det-Seg-Refine [2] + CLIP [57]0.300.23
GG-CNN [47] + CLIP [57]0.100.07
CLIPORT [64]0.270.30
CLIP-Fusion [80]0.400.40
LGD (ours)0.430.42
", + "bbox": [ + 531, + 324, + 874, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Robotic language-driven grasp detection results.", + "bbox": [ + 514, + 458, + 875, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "R820 robot. Using the RealSense D435i camera, the grasp pose inferred from approaches in Table 4 is transformed into the 6DoF grasp pose, similar to [35]. The optimization-based trajectory planner in [5,72] is employed to execute the grasps. Experiments are conducted for two scenarios, i.e., the single object scenario and the cluttered scene scenario, of a set of 20 real-world daily objects. In each scenario, we run 30 experiments using baselines listed in Table 4 and a predefined grasping prompt corpus. The results exhibit that our LGD outperforms other baselines. Furthermore, although LGD is trained on our Grasp-Anything++ which is a solely synthesis dataset created by foundation models, it still shows reasonable results on real-world objects.", + "bbox": [ + 496, + 488, + 890, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Zero-shot Grasp Detection", + "text_level": 1, + "bbox": [ + 500, + 695, + 740, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed Grasp-Anything++ is a large-scale dataset. Apart from the language-driven grasp detection task, we believe it can be used for other purposes. In this experiment, we seek to answer the question: Can Grasp-Anything++ be useful in the traditional grasp detection task without text? Consequently, we verify our Grasp-Anything++ and LGD (no text) with other existing datasets and grasping methods.", + "bbox": [ + 496, + 719, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Setup. We setup an LGD (no text) version, and other state-of-the-art grasp detection methods GR-ConvNet [35], Det-Seg-Refine [2], GG-CNN [47]. We use five datasets: our Grasp-Anything++, Jacquard [17], Cornell [33], VMRD [87], and OCID-grasp [2] in this experiment.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e1593434cf1da0bb562034e28ceb91b09db38ce79f5c2d26943e8137b15ae853.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BaselineGrasp-Anything++ (ours)Jacquard [17]Cornell [33]VMRD [87]OCID-grasp [2]
BaseNewHBaseNewHBaseNewHBaseNewHBaseNewH
GR-ConvNet [35]0.710.590.640.880.660.750.980.740.840.770.640.700.860.670.75
Det-Seg-Refine [2]0.620.570.590.860.600.710.990.760.860.750.600.660.800.620.70
GG-CNN [47]0.680.570.620.780.560.650.960.750.840.690.530.590.710.630.67
LGD (no text) (ours)0.740.630.680.890.690.770.970.760.850.790.660.720.880.680.76
", + "bbox": [ + 81, + 99, + 890, + 202 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot Results. Table 5 summarizes the base-to-new grasp detection results on five datasets. Overall, the performance of LGD even without the language branch is better than other baselines across all datasets. Furthermore, this table also shows that our Grasp-Anything++ dataset is more challenging to train as the detection results are lower than related datasets using the same approaches due to the greater coverage of unseen objects in the testing phase.", + "bbox": [ + 76, + 233, + 470, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0372a56ad7adbd16e878c5b14ffa8a808afcdc2d21496ae75bb199206f0bed3a.jpg", + "table_caption": [ + "Table 5. Base-to-new zero-shot grasp detection results." + ], + "table_footnote": [], + "table_body": "
Test TrainJacquardCornellVMRDOCID-graspGrasp-Anything++
Jacquard [17]0.870.510.130.210.17
Cornell [33]0.070.980.200.120.13
VMRD [87]0.060.210.790.110.10
OCID-grasp [2]0.090.120.200.740.11
Grasp-Anything++ (ours)0.410.630.300.390.65
", + "bbox": [ + 81, + 367, + 468, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Cross-dataset grasp detection results.", + "bbox": [ + 133, + 458, + 410, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-dataset Evaluation. To further verify the usefulness of our Grasp-Anything++ dataset, we conduct the cross-dataset validation in Table 6. We use the GRConvNet [35] to reuse its results on existing grasp datasets. GR-ConvNet is trained on a dataset (row) and evaluated on another dataset (column). For example, training on Jacquard and testing on Cornell yields an accuracy of 0.51. Notably, training with our dataset improves performance by approximately $10 - 33\\%$ compared to other datasets.", + "bbox": [ + 75, + 477, + 468, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the wild grasp detection. Fig. 9 shows visualization results using LGD (no text) trained on our Grasp-Anything++ dataset on random internet images and other datasets images. We can see that the detected grasp poses are adequate in quality and quantity. This demonstrates that although our Grasp-Anything++ is fully created by foundation models without having any real images, models trained on our Grasp-Anything++ dataset still generalize well on real-world images.", + "bbox": [ + 75, + 613, + 468, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Discussion", + "text_level": 1, + "bbox": [ + 76, + 756, + 194, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments indicate that Grasp-Anything++ can serve as a foundation dataset for both language-driven and traditional grasp detection tasks. However, there are certain limitations. First, our dataset lacks depth images for directly being applied to robotic applications [52]. Second, we remark that the creation of our dataset is time-consuming and relies on access to the ChatGPT API. Fortunately, future research can reuse our provided assets (images, prompts,", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d45184855b9574769a59b2cc5d6a5a523aadcc2d15ecca86d9317af75a06c019.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 229, + 625, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a10c88bfdc345c472ae706e4ff528b3568882724b860abf6dcfae55a39089ad7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 229, + 754, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74ab2e0f15cced209ade5f6546ff3d4275012a2e365f1bb48382ea7f66ef018d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 229, + 885, + 325 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb740b29edd411199dd92af3a1a8f467e814b76806cffefa8bf4bc433575e067.jpg", + "image_caption": [ + "Figure 9. In the wild grasp detection. Top row images are from GraspNet [24], YCB-Video [78], NBMOD [9] datasets; bottom row shows internet images." + ], + "image_footnote": [], + "bbox": [ + 501, + 330, + 625, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dab9c132ea433c097788cc4bda274d6644c4c48ab2609b7975562fb7471f7f57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 330, + 754, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dd68889eeac9faeb8996a249d69ba68fa871e77f0b86937ff65f3e1ca7351ec5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 330, + 885, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "etc.) without starting from scratch. Furthermore, our experiments show that adding language to the grasp detection task (Table 2) poses a more challenging problem compared to standard grasp detection task (Table 5).", + "bbox": [ + 496, + 484, + 890, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We see several interesting future research directions. First, future work could investigate the use of text or imaged-to-3D models [79] or image-to-depth [59] and reuse our dataset's prompts and images to construct 3D language-driven grasp datasets. Additionally, beyond linguistic grasp instruction adherence, our dataset holds potential for varied applications, including scene understanding [81] and scene generation [3], hallucination analysis [30], and humanrobot interaction [11].", + "bbox": [ + 496, + 547, + 890, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 705, + 617, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Grasp-Anything++, a large-scale dataset with 1M images and 10M grasp prompts for language-driven grasp detection tasks. We propose LGD, a diffusion-based method to tackle the language-driven grasp detection task. Our diffusion model employs a contrastive training objective, which explicitly contributes to the denoising process. Empirically, we have shown that Grasp-Anything++ serves as a foundation grasp detection dataset. Finally, our LGD improves the performance of other baselines, and the real-world robotic experiments further validate the effectiveness of our dataset and approach.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17909", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2", + "[2] Stefan Ainetter and Friedrich Fraundorfer. End-to-end trainable deep neural network for robotic grasp detection and semantic segmentation from rgb. In ICRA, 2021. 2, 4, 6, 7, 8", + "[3] Dor Arad Hudson and Larry Zitnick. Compositional transformers for scene generation. NeurIPS, 2021. 8", + "[4] Umar Asif, Jianbin Tang, and Stefan Harrer. Grasynet: An efficient convolutional neural network for real-time grasp detection for low-powered devices. In IJCAI, 2018. 4", + "[5] Florian Beck, Minh Nhat Vu, Christian Hartl-Nesic, and Andreas Kugi. Singularity avoidance with application to online trajectory optimization for serial manipulators. IFAC-PapersOnLine, 56(2):284-291, 2023. 7", + "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33, 2020. 2", + "[8] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In CVPR, 2021. 4", + "[9] Boyuan Cao, Xinyu Zhou, Congmin Guo, Baohua Zhang, Yuchen Liu, and Qianqiu Tan. Nbmod: Find it and grasp it in noisy background. arXiv preprint arXiv:2306.10265, 2023. 2, 3, 8", + "[10] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 2", + "[11] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, 2022. 8", + "[12] I-Ming Chen and Joel W Burdick. Finding antipodal point grasps on irregularly shaped objects. T-RA, 1993. 4", + "[13] Sijia Chen and Baochun Li. Language-guided diffusion model for visual grounding. arXiv preprint arXiv:2308.09599, 2023. 2", + "[14] Hongjun Choi, Anirudh Som, and Pavan Turaga. Amc-loss: Angular margin contrastive loss for improved explainability in image classification. In CVPR workshops, 2020. 4", + "[15] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. TPAMI, 2023. 4", + "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 4", + "[17] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IROS, 2018. 2, 3, 4, 7, 8", + "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6", + "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34, 2021. 5", + "[20] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 2", + "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1", + "[22] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. A billion ways to grasp: An evaluation of grasp sampling schemes on a dense, physics-based grasp data set. In ISRR, 2019. 2", + "[23] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In ICRA, 2021. 2", + "[24] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. 2, 8", + "[25] Maximilian Gilles, Yuhao Chen, Tim Robin Winter, E Zhixuan Zeng, and Alexander Wong. Metagraspnet: A large-scale benchmark dataset for scene-aware ambidextrous bin picking via physics-based metaverse synthesis. In CASE, 2022. 1", + "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4", + "[27] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTAT, 2010. 5", + "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 4, 5, 6", + "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 4", + "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232, 2023. 4, 8", + "[31] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, 2022. 2", + "[32] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 1" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Yun Jiang, Stephen Moseson, and Ashutosh Saxena. Efficient grasping from rgbd images: Learning using a new rectangle representation. In ICRA, 2011. 2, 3, 4, 7, 8", + "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3", + "[35] Sulabh Kumra, Shirin Joshi, and Ferat Sahin. Antipodal robotic grasping using generative residual convolutional neural network. In IROS, 2020. 6, 7, 8", + "[36] Jaewoong Lee, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, Yunji Kim, Jin-Hwa Kim, Jung-Woo Ha, and Sung Ju Hwang. Text-conditioned sampling framework for text-to-image generation with masked generative models. arXiv preprint arXiv:2304.01515, 2023. 2", + "[37] Meng-Lun Lee, Sara Behdad, Xiao Liang, and Minghui Zheng. Task allocation and planning for product disassembly with human-robot collaboration. Robotics and Computer-Integrated Manufacturing, 76, 2022. 2", + "[38] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. *IJRR*, 2018. 2", + "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 5", + "[40] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2", + "[41] Jirong Liu, Ruo Zhang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Chenxi Wang, Sheng Xu, Hengxu Yan, and Cewu Lu. Target-referenced reactive grasping for dynamic objects. In CVPR, 2023. 2", + "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2", + "[43] Weiyu Liu, Tucker Hermans, Sonia Chernova, and Chris Paxton. Structdiffusion: Object-centric diffusion for semantic rearrangement of novel objects. arXiv preprint arXiv:2211.04604, 2022. 2, 4", + "[44] Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-Radiology, page 100017, 2023. 4", + "[45] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312, 2017. 2", + "[46] Jiaxu Miao, Zongxin Yang, Leilei Fan, and Yi Yang. Fedseg: Class-heterogeneous federated learning for semantic segmentation. In CVPR, 2023. 7" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Douglas Morrison, Peter Corke, and Jürgen Leitner. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. arXiv preprint arXiv:1804.05172, 2018. 6, 7, 8", + "[48] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. RA-L, 2020. 2", + "[49] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In ICCV, 2019. 2", + "[50] Yao Mu, Shunyu Yao, Mingyu Ding, Ping Luo, and Chuang Gan. Ec2: Emergent communication for embodied control. In CVPR, 2023. 2", + "[51] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 1", + "[52] Rhys Newbury, Morris Gu, Lachlan Chumbley, Arsalan Mousavian, Clemens Eppner, Jürgen Leitner, Jeannette Bohg, Antonio Morales, Tamim Asfour, Danica Kragic, et al. Deep learning approaches to grasp synthesis: A review. T-RO, 2023. 8", + "[53] Toan Nguyen, Minh Nhat Vu, An Vuong, Dzung Nguyen, Thieu Vo, Ngan Le, and Anh Nguyen. Open-vocabulary affordance detection in 3d point clouds. In IROS, 2023. 3", + "[54] OpenAI. Introducing ChatGPT. Software. Accessed: July 6th 2023. 2", + "[55] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. 2", + "[56] Robert Platt. Grasp learning: Models, methods, and performance. Annual Review of Control, Robotics, and Autonomous Systems, 2023. 1", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7", + "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6", + "[59] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In CVP4, 2021. 8", + "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4", + "[61] Kallol Saha, Vishal Mandadi, Jayaram Reddy, Ajit Srikanth, Aditya Agarwal, Bipasha Sen, Arun Singh, and Madhava Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. arXiv preprint arXiv:2309.11414, 2023. 2", + "[62] Schuhmann et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 4", + "[63] Dhruv Shah, Błajej Osiński, brian richter, and Sergey Levine. Lm-nav: Robotic navigation with large pre-trained models of" + ], + "bbox": [ + 501, + 90, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "language, vision, and action. In Karen Liu, Dana Kulic, and Jeff Ichnowski, editors, Proceedings of The 6th Conference on Robot Learning, volume 205 of PMLR. PMLR, 2023. 1", + "[64] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1, 2, 6, 7", + "[65] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 5", + "[66] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023. 1", + "[67] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3", + "[68] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2022. 5, 6", + "[69] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 5", + "[70] Julien Urain, Niklas Funk, Jan Peters, and Georgia Chalvatzaki. Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In ICRA, 2023. 2", + "[71] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2023. 2", + "[72] Minh Nhat Vu, Florian Beck, Michael Schwegel, Christian Hartl-Nesic, Anh Nguyen, and Andreas Kugi. Machine learning-based framework for optimally solving the analytical inverse kinematics for redundant manipulators. Mechatronics, 2023. 7", + "[73] An Vuong, Minh Nhat Vu, Toan Tien Nguyen, Baoru Huang, Dzung Nguyen, Thieu Vo, and Anh Nguyen. Language-driven scene synthesis using multi-conditional diffusion model. arXiv preprint arXiv:2310.15948, 2023. 5", + "[74] An Dinh Vuong, Minh Nhat Vu, Hieu Le, Baoru Huang, Binh Huynh, Thieu Vo, Andreas Kugi, and Anh Nguyen. Grasp-anything: Large-scale grasp dataset from foundation models. arXiv preprint arXiv:2309.09818, 2023. 2, 3", + "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 3", + "[76] Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania, and Ernesto W De Luca. Legal norm retrieval with variations of the bert model combined with tfidf vectorization. In Proceedings of the eighteenth international conference on artificial intelligence and law, pages 285-294, 2021. 4", + "[77] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In MIDL, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 2, 8", + "[79] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 8", + "[80] Kechun Xu, Shuqi Zhao, Zhongxiang Zhou, Zizhang Li, Huajin Pi, Yifeng Zhu, Yue Wang, and Rong Xiong. A joint modeling of vision-language-action for target-oriented grasping in clutter. arXiv preprint arXiv:2302.12610, 2023. 2, 6, 7", + "[81] Yiteng Xu, Peishan Cong, Yichen Yao, Runnan Chen, Yuenan Hou, Xinge Zhu, Xuming He, Jingyi Yu, and Yuexin Ma. Human-centric scene understanding for 3d large-scale scenarios. In ICCV, 2023. 8", + "[82] Zhixuan Xu, Kechun Xu, Rong Xiong, and Yue Wang. Object-centric inference for language conditioned placement: A foundation model based approach. In ICARM, 2023. 2", + "[83] Jange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pick-place robots. arXiv preprint arXiv:2306.05716, 2023. 1, 2", + "[84] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 2022, 2, 4", + "[85] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. ICCV, 2023. 4", + "[86] Yang Yang, Xibai Lou, and Changyun Choi. Interactive robotic grasping with attribute-guided disambiguation. In ICRA, 2022. 2", + "[87] Hanbo Zhang, Xuguang Lan, Site Bai, Xinwen Zhou, Zhiqiang Tian, and Nanning Zheng. Roi-based robotic grasp detection for object overlapping scenes. In IROS, 2019. 4, 7, 8", + "[88] Hanbo Zhang, Yunfan Lu, Cunjun Yu, David Hsu, Xuguang La, and Nanning Zheng. Invigorate: Interactive visual grounding and grasping in clutter. arXiv preprint arXiv:2108.11092, 2021. 2", + "[89] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 4", + "[90] Kaizhi Zheng, Xiaotong Chen, Odest Chadwicke Jenkins, and Xin Wang. Vlmbench: A compositional benchmark for vision-and-language manipulation. NeurIPS, 35, 2022. 2", + "[91] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 4", + "[92] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 6" + ], + "bbox": [ + 501, + 92, + 890, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17912", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_model.json b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6342d4d46123f902acfc88666fbb2fb3719d780d --- /dev/null +++ b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_model.json @@ -0,0 +1,3093 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.312, + 0.131, + 0.66, + 0.154 + ], + "angle": 0, + "content": "Language-driven Grasp Detection" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.92, + 0.201 + ], + "angle": 0, + "content": "An Dinh Vuong\\(^{1}\\), Minh Nhat Vu\\(^{2,*}\\), Baoru Huang\\(^{3,*}\\), Nghia Nguyen\\(^{1}\\), Hieu Le\\(^{1}\\), Thieu Vo\\(^{4}\\), Anh Nguyen\\(^{5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.201, + 0.857, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)FPT Software AI Center, Vietnam \\(^{2}\\)Automation & Control Institute, TU Wien, Austria \\(^{3}\\)Imperial College London, UK" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.219, + 0.789, + 0.234 + ], + "angle": 0, + "content": "\\(^{4}\\)Ton Duc Thang University, Vietnam \\(^{5}\\)University of Liverpool, UK *Co-Corresponding authors" + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.237, + 0.62, + 0.252 + ], + "angle": 0, + "content": "https://airvlab.github.io/grasp-anything" + }, + { + "type": "image_caption", + "bbox": [ + 0.088, + 0.274, + 0.245, + 0.308 + ], + "angle": 0, + "content": "A coffee cup, a grape biscuit, a black oven plate on a white dining table" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.31, + 0.246, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.445, + 0.225, + 0.456 + ], + "angle": 0, + "content": "Grasp the coffee cup." + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.275, + 0.401, + 0.308 + ], + "angle": 0, + "content": "A black pen, a steel marble and a digital calculator on an office desk" + }, + { + "type": "image", + "bbox": [ + 0.248, + 0.31, + 0.406, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.445, + 0.38, + 0.468 + ], + "angle": 0, + "content": "Grasp the calculator at its keypad." + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.275, + 0.565, + 0.308 + ], + "angle": 0, + "content": "A black pot, a green bottle, a bread container arranged on a kitchen counter" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.31, + 0.565, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.445, + 0.537, + 0.468 + ], + "angle": 0, + "content": "Grab the neck of the green bottle." + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.275, + 0.724, + 0.308 + ], + "angle": 0, + "content": "A brown-band wristwatch, a sleek black pen and an office clip placed on a table" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.31, + 0.725, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.445, + 0.713, + 0.467 + ], + "angle": 0, + "content": "Give me the brown-band wristwatch." + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.275, + 0.886, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.445, + 0.852, + 0.468 + ], + "angle": 0, + "content": "Pick the fountain pen at its cap." + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.477, + 0.72, + 0.492 + ], + "angle": 0, + "content": "Figure 1. We present a new dataset and method for language-driven grasp task." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.505, + 0.314, + 0.518 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Grasp detection is a persistent and intricate challenge with various industrial applications. Recently, many methods and datasets have been proposed to tackle the grasp detection problem. However, most of them do not consider using natural language as a condition to detect the grasp poses. In this paper, we introduce Grasp-Anything++, a new language-driven grasp detection dataset featuring 1M samples, over 3M objects, and upwards of 10M grasping instructions. We utilize foundation models to create a large-scale scene corpus with corresponding images and grasp prompts. We approach the language-driven grasp detection task as a conditional generation problem. Drawing on the success of diffusion models in generative tasks and given that language plays a vital role in this task, we propose a new language-driven grasp detection method based on diffusion models. Our key contribution is the contrastive training objective, which explicitly contributes to the denoising process to detect the grasp pose given the language instructions. We illustrate that our approach is theoretically supportive. The intensive experiments show that our method outperforms state-of-the-art approaches and allows real-world robotic grasping. Finally, we demonstrate our large-scale dataset enables zero-short grasp detection and is a challenging benchmark for future work." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.504, + 0.631, + 0.519 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Imagine we want an assistant robot to grasp a cup among a clutter of daily objects such as a knife, a fork, a cup, and a pair of scissors. Conventionally, to convey the idea of grasping this specific object, humans use the natural language command, \"give me the cup\", for instance. Although humans intuitively know how to grasp the cup given the linguistic command, determining specific grasp actions for objects based on natural language instructions or language-driven grasp detection remains challenging for robots [64]. First, natural language is usually overlooked in existing grasp datasets [56] while training vision-and-language neural networks necessitates an excessive number of labeled examples [66]. Second, recent works usually focus on particular manipulation tasks with limited objects [25], imposing a bottleneck for in-the-wild robot execution [63]. Finally, despite recent developments, bridging the gap between language, vision, and control for real-world robotic experiments remains a challenging task [83]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, language-driven robotic frameworks are gaining traction, offering the potential for robots to process natural language, and bridging the gap between robotic manipulations and real-world human-robot interaction [51]. PaLM-E [21], EgoCOT [51], and ConceptFusion [32] are some notable embodied robots with the ability to compre" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "hend natural language by harnessing the power of large foundation models such as ChatGPT [54]. However, most works assume the high-level actions of robots and ignore the fundamental grasping actions, restricting the structure for generalization across robotic domains, tasks, and skills [50]. In this paper, we explore training a language-driven agent to implement low-level actions, focusing on the task of object grasping via image observations. Specifically, our hypothesis is centered around the establishment of a robotic system that can execute grasping actions following a given language instruction for any universal object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.473, + 0.5 + ], + "angle": 0, + "content": "We first present Grasp-Anything++ to serve as a large-scale dataset for language-driven grasp detection. Our dataset is based on the Grasp-Anything [74], and is synthesized from foundation models. Compared to the original Grasp-Anything dataset, we provide more than 10M grasp prompts and 3M associated object masks, 6M ground truth poses at the object part level. Our dataset showcases the ability to facilitate grasp detection using language instructions. We label the ground truth at both the object level and part level, providing a comprehensive understanding of real-world scenarios. For example, our ground truth includes both general instructions \"give me the knife\" and detail ones such as \"grasp the handle of the steak knife\". We empirically show that our large-scale dataset successfully facilitates zero-shot grasp detection on both vision-based tasks and real-world robotic experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.473, + 0.756 + ], + "angle": 0, + "content": "To tackle the challenging language-driven grasp detection task, we propose a new diffusion model-based method. Our selection of diffusion models is motivated by their proven efficacy in conditional generation tasks [28]. These models have shown efficiency beyond image synthesis, including other image-based tasks such as image segmentation [77], and visual grounding [40]. Despite achieving notable success, integrating visual and text features effectively remains a challenge [13] as the majority of existing literature employs latent strategies to combine visual and text features [36]. We address this challenge by employing a new training strategy for learning text and image features, focusing on the use of feature maps as guidance information for grasp pose generation. Our main contribution is a new training objective that incorporates the feature maps and explicitly contributes to the denoising process. In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.764, + 0.47, + 0.794 + ], + "angle": 0, + "content": "- We propose Grasp-Anything++, a large-scale language-driven dataset for grasp detection tasks." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.802, + 0.47, + 0.849 + ], + "angle": 0, + "content": "- We propose a diffusion model with a training objective that explicitly contributes to the denoising process to detect the grasp poses." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We demonstrate that our Grasp-Anything++ dataset and the proposed method outperform other approaches and enable successful robotic applications." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.764, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Grasp Detection. Grasp detection is a popular task in both computer vision and robotic community [2, 17, 24, 64]. Recently, establishing robotic systems with the ability to follow natural commands has been actively researched [64, 80, 83]. The prevalent solution to the language-driven grasp detection task is to split into two stages: one for grounding the target object, and the other is to synthesize grasp poses from the grounding visual-text correlations [1, 80]. Training in two stages may result in longer inference time [41]. In addition, several works [82, 83, 90] adopt foundation models, such as GroundDINO [42] and GPT-3 [7]. Accessing such commercial foundation models is not always available [71], especially on robotic systems with limited resources or unstable internet connection [37]. In our work, we directly train the model on the large-scale Grasp-Anything++ dataset to inherit the power of a foundation-based dataset, while ensuring a straightforward inference process for the downstream robotic applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.645 + ], + "angle": 0, + "content": "Language-driven Grasp Detection Datasets. While there are many grasp datasets have been introduced [9, 17, 22-24, 33, 38, 45, 48, 49, 55, 78], the majority of them overlook the text modality. Therefore, the grasping of objects out of a clutter typically experiences ambiguities in what object to grasp [88]. DailyGrasp [83] is one of the first grasp datasets employing natural language for scene descriptions; however, the scene description corpus in this dataset is relatively small and does not specify which part of the object should be grasped. In our work, we present Grasp Anything++, which is a large-scale language-driven grasping dataset. Grasp-Anything++ provides object instance description with a natural word phrase, for example, \"a blue ceramic vase\". Furthermore, Grasp-Anything++ describes the grasping object at both the part level and object level, providing more information for the robot to execute the grasping [86]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.896, + 0.903 + ], + "angle": 0, + "content": "Diffusion Models for Robotic Applications. Diffusion models [28] have emerged as the new state-of-the-art method of generative tasks [84]. Recent have witnessed growing attention for utilizing diffusion models in robotic applications [61]. Liu et al. [43] propose a diffusion model to handle the language-guided object rearrangement task. Diffusion models are also applied to other robotic tasks such as motion planning [10], and trajectory optimization [31]. The authors in [70] present a diffusion model to determine grasp poses by minimizing a SDF loss. Overall, the diffusion models employed in previous works often combine visual and text features in a latent mechanism [6], which may cause interpretability problems [83] for robotic systems that require low-level controls [20]. To tackle this challenge, we propose a training objective that explicitly contributes to the denoising process. We demonstrate that our proposed strategy is more effective than the latent strategy." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17903" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.891, + 0.489 + ], + "angle": 0, + "content": "
StepDescription
Scene GenerationUserPlease help me generate scene descriptions for natural arrangements of daily objects. Each description has the following form: <Object_1><Object_2>...<Verb><Container Object>. Please also ensure the incorporation of a rich and varied lexicon in the scene descriptions.
SampleA steel knife, a polished fork and a pristine ceramic plate on a wooden table.
Text-to-ImageWe use Stable Diffusion [60] to proceed text-to-image generation.
Object MaskingUserFor object part-level description, given an input list {<Object_1>, <Object_2>, ...}, the output will be a list that describes the parts of objects as: {<Object_1>: [(Part_1.1>, <Part_1.2>, ..., <Object_2>]: [(Part_2.1>, <Part_2.2>, ...]}.
Sample{knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}
Post ProcessWe use OFA [75] and SAM [34] to locate the region describing the objects.
Part MaskingUserGiven the object list and part lists of each scene description, you will generate for me all prompts with the following format: {<Manipulation Action><Object_ID><Part_ID)}. The part that is more suitable for human grasping is positioned at the start of the list to represent the grasping actions.
SampleGive me the steel knife; Grasp the knife at its handle.
Post ProcessWe leverage VLPart [67] to locate the region describing the parts of objects.
Grasp GenerationUserGenerate for me a scene description with grasp instructions following the templates.
SampleScene description: A steel knife, a polished fork and a pristine ceramic plate on a wooden table. Object list: {knife, fork, plate}. Part lists: {knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}. Prompts: Give me the steel knife; Grasp the knife at its handle.
Grasp LabellingWe utilize a pretrained RAGT-3/3 [9] to generate grasp poses corresponding to the located region.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.492, + 0.893, + 0.521 + ], + "angle": 0, + "content": "Table 1. Grasp-Anything++ creation pipeline. We utilize ChatGPT to generate scene descriptions and grasp instructions from the user input. We generate images given scene descriptions and automatically synthesize the grasp poses." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.367, + 0.544 + ], + "angle": 0, + "content": "3. The Grasp-Anything++ Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.47, + 0.645 + ], + "angle": 0, + "content": "We utilize large-scale foundation models to create the Grasp-Anything++. Our dataset offers open-vocabulary grasping commands and images with associated groundtruth. There are three key steps in establishing our dataset: i) prompting procedure, ii) image synthesis and grasp poses annotation, and iii) post-processing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.281, + 0.668 + ], + "angle": 0, + "content": "3.1. Prompting Procedure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.78 + ], + "angle": 0, + "content": "We first establish prompt-based procedures to generate a large-scale scene description corpus as well as grasp prompt instructions. In particular, we utilize ChatGPT to generate the prompts for two tasks: \\(i\\) ) Scene descriptions: Sentences capturing the scene arrangement, including the extracted object and part lists, and \\(ii\\) ) Grasp instructions: Prompts directing the robot to grasp specific objects or parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We follow a procedure in Table 1 to implement ChatGPT's output templates. The reference target in the grasp instruction may be either an object or an object's part. When the reference is an object's part, that part is directly selected as the reference in the grasp instruction sentence. If the reference is an object, we determine the grasping region on the part of the object that is likely to be grasped in everyday scenarios as described in affordance theory [53]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.838, + 0.544 + ], + "angle": 0, + "content": "3.2. Image Synthesis and Grasp Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Image Synthesis. Given the scene description corpus, we first utilize a large-scale pretrained text-to-image model, namely, Stable Diffusion [60] to generate images from scene descriptions. Next, we perform a series of visual grounding and image segmentation using OFA [75], Segment-Anything [34], and VLPart [67] to locate the referenced object or part to the grasp instruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Grasp Annotation. Grasp poses are represented as 2D rectangles, consistent with prior research and practical compatibility with real-world parallel grippers [17, 33]. Utilizing a pretrained network [9], we annotate grasp poses based on part segmentation masks. Since potential inaccuracies in these candidate poses could occur, we follow the procedure as defined in [74] to evaluate the quality of generated grasp poses to discard unreasonable grasp poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Specifically, grasp quality is evaluated through net torque \\(\\mathcal{T} = (\\tau_{1} + \\tau_{2}) - RMg\\), where resistance at contact points is \\(\\tau_{i} = K\\mu_{s}F\\cos \\alpha_{i}\\). With constants such as \\(M\\) (mass), \\(g\\) (gravitational acceleration), \\(K\\) (geometrical characteristics), \\(\\mu_{s}\\) (static friction coefficient), and \\(F\\) (applied force), accurately determining \\(\\mathcal{T}\\) directly is challenging due to the physical difficulties in precisely measuring \\(M\\), \\(K\\), and \\(\\mu_{s}\\). Thus, we employ a surrogate measure," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17904" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.331, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.205, + 0.273, + 0.217 + ], + "angle": 0, + "content": "(a) Number of categories" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.096, + 0.585, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.204, + 0.534, + 0.217 + ], + "angle": 0, + "content": "(b) Scene description length" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.091, + 0.895, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.204, + 0.839, + 0.217 + ], + "angle": 0, + "content": "(c) Number of object/part references." + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.221, + 0.842, + 0.235 + ], + "angle": 0, + "content": "Figure 2. Data statistics. We analyze object categories and the use of natural language in the Grasp-Anything++ dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.468, + 0.324 + ], + "angle": 0, + "content": "\\(\\tilde{\\mathcal{T}} = \\frac{\\cos\\alpha_1 + \\cos\\alpha_2}{R}\\), as an alternative. As a result in [12], antipodal grasps score higher on \\(\\tilde{\\mathcal{T}}\\), indicating better quality. Consequently, grasps are evaluated based on \\(\\tilde{\\mathcal{T}}\\), with positive values indicating positive grasps and others considered as negative." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Post Processing. Despite training on extensive datasets, Stable Diffusion [60] may produce subpar content, commonly termed as hallucination [30] when generating images from the text prompts. To address this, we perform manual reviews to filter out such images, with qualitative examples in our figures. Our process includes checks at every stage to prevent duplicate or hallucinated content. However, manual inspection introduces biases, which we counter with guidelines focusing on abnormal structures or implausible gravity (Fig. 3), aligns with approach in the literature [62]." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.479, + 0.203, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.479, + 0.335, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.479, + 0.468, + 0.575 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.592, + 0.468, + 0.648 + ], + "angle": 0, + "content": "Figure 3. Failure image generation cases. Images generated by Stable Diffusion [60] may exhibit hallucinatory artifacts such as sunglasses lacking a lens, scissors with an anomalous structure, and a spoon not resting properly on a table." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.468, + 0.746 + ], + "angle": 0, + "content": "Additionally, ChatGPT-generated scene prompts often duplicate [44]. To address this, we use duplication checking, filtering out identical prompts with BERTScore [89], which assesses sentence similarity through cosine similarities of token embeddings. We remove sentences with a BERTScore above 0.85 as in prior study [76]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.224, + 0.77 + ], + "angle": 0, + "content": "3.3. Data Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Number of Categories. To evaluate the object category diversity, we apply a methodology akin to that in [16]. Utilizing 300 categories from LVIS dataset [26], we employ a pretrained model [91] to identify 300 candidate objects from our dataset for each category. We then curate a subset comprising 90,000 objects, refining it by excluding items that do not align semantically with their designated categories. A category is considered significant if it has more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.285 + ], + "angle": 0, + "content": "than 40 objects. Fig. 2a shows the results. Overall, our dataset spans over 236 categories from LVIS dataset, indicating a notable degree of object diversity in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.406 + ], + "angle": 0, + "content": "Scene Descriptions. Fig. 2b shows the distribution of scene descriptions based on sentence length. The analysis reveals a wide range of sentence lengths, spanning from 10 words to 100 words per sentence. On average, each scene description consists of approximately 54 words, indicative of detailed and descriptive sentences. These scene descriptions correspond to sets of grasp instructions. Fig. 2c further shows the objects and object parts in scene descriptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Diversity Analysis. We assess the diversity of occlusion and lighting conditions in the dataset. Regarding occlusion, we use a pretrained YOLOv5 model to identify objects within images. The results indicate that \\(93.8\\%\\) of images have a substantial overlap of five or more bounding boxes, which suggests a diverse range of occlusion within the Grasp-Anything++ dataset. Regarding lighting conditions, we convert images to YCbCr to analyze Y channel (luminance) and find that GraspL1M has the most diverse lighting conditions, identifying by the lowest Gini coefficient (a metric to measure the inequality of a distribution) of 0.26, compared to VMRD [87] (0.31), OCID-grasp [2] (0.32), Cornell [33] (0.62), Jacquard [17] (0.91)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.616, + 0.809, + 0.634 + ], + "angle": 0, + "content": "4. Language-driven Grasp Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.823 + ], + "angle": 0, + "content": "Motivation. The use of diffusion model for language-driven grasp detection is motivated by its efficiency in various generative tasks [15,28,43,84]. Conditional generation, such as our language-driven grasp detection task, aligns seamlessly with diffusion models' capabilities [29]. Moreover, language-driven grasp detection represents a fine-grained problem in which the outputs strongly depend on the text input [4]. For example, \"grasp the steak knife\" and \"grasp the kraft knife\" refer to two different objects on the image. To this end, we propose using contrastive loss with diffusion model to tackle this task, as contrastive learning is a popular solution for fine-grained tasks [8, 14, 85]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.833, + 0.827, + 0.848 + ], + "angle": 0, + "content": "4.1. Constrastive Loss for Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We represent the target grasp pose as \\(\\mathbf{x}_0\\) in the diffusion model. The objective of our diffusion process of language-driven grasp detection involves denoising from a noisy state" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17905" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_T\\) to the original grasp pose \\(\\mathbf{x}_0\\), conditioned on the input image and grasp instruction represented by \\(y\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.47, + 0.213 + ], + "angle": 0, + "content": "In a diffusion process [28], assume that \\( q(\\mathbf{x}_{1:T}|\\mathbf{x}_0) \\) is the forward process and we parameterize the reverse process by \\( p_{\\theta}(\\mathbf{x}_{0:T}) \\). The conditional diffusion process [19] assumes \\( \\hat{q} \\) is the forward process but with the inclusion of a condition \\( y \\). The goal of the reverse process is to optimize the variational bound on negative log likelihood [28]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.231, + 0.47, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {T}) - \\sum_ {t \\geq 1} \\log \\frac {p _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t})}{q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1})} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.286, + 0.303, + 0.301 + ], + "angle": 0, + "content": "We prove in the Appendix that" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.319, + 0.483, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathbb {E} \\left[ \\underbrace {C - \\log p _ {\\theta} (\\mathbf {x} _ {T})} _ {\\text {C o n s t a n t}} + \\underbrace {\\log \\hat {q} (\\mathbf {x} _ {0} | \\mathbf {x} _ {T} , y)} _ {\\text {C o n t r a s t i v e}} + \\right. \\\\ \\underbrace {\\sum_ {t > 1} D _ {\\mathrm {K L}} \\left(\\hat {q} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t} , \\mathbf {x} _ {0} , y\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)\\right) - \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1} , y\\right)} _ {\\text {D e n o i s i n g s c o r e}}. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.435, + 0.47, + 0.511 + ], + "angle": 0, + "content": "The terms \\(D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t))\\) and \\(\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)\\) in Equation 2 are similar to the concept of denoising score used in [28]. Thus, we can represent the quantity \\(\\sum_{t > 1}D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t))-\\) \\(\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)\\) by the loss utilized in [68, 69]" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.53, + 0.475, + 0.564 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0} | y), t \\sim [ 1, T ]} \\left[ \\mathbf {x} _ {0} - f \\left(\\mathbf {x} _ {t + 1}, t + 1, \\tilde {\\mathbf {x}} _ {0}\\right) \\right] ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.469, + 0.624 + ], + "angle": 0, + "content": "Since \\( q(\\mathbf{x}_T|\\cdot) \\) is equivalent to an isotropic Gaussian distribution as \\( T \\to +\\infty \\), the estimation quantity \\( \\log p_{\\theta}(\\mathbf{x}_T) \\) converges to a constant when \\( \\theta \\to \\theta^* \\). Therefore, we can ignore the first term of Equation 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.469, + 0.729 + ], + "angle": 0, + "content": "Finally, the term \\(\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)\\) of Equation 2 provides more information about the relation between \\(\\mathbf{x}_T\\) and \\(\\mathbf{x}_0\\). As this term is intractable [65], we parameterize \\(\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)\\) as \\(p_{\\psi ,y}(\\mathbf{x}_0,\\mathbf{x}_T)\\). This estimation resembles the noise-contrastive estimation [27], where the \\(\\mathbf{x}_T\\), \\(\\mathbf{x}_0\\) can be considered as a pair of contrastive estimation and \\(\\psi\\) can be estimated by a contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.469, + 0.806 + ], + "angle": 0, + "content": "In [19, 68, 73], the authors indicate that predicting \\(\\mathbf{x}_0\\) is often infeasible but predicting an estimation \\(\\tilde{\\mathbf{x}}_0\\) is tractable and can be used as a 'pseudo' estimation of \\(\\mathbf{x}_0\\). We denote \\(\\tilde{\\mathbf{x}}_0\\) as an estimation of \\(\\mathbf{x}_0\\). The loss term \\(\\hat{q}(\\mathbf{x}_0 | \\mathbf{x}_T, y)\\) can be approximated by using the following contrastive loss" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.824, + 0.47, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n t r a s t i v e}} = \\max \\left(0, \\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} - M\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(M\\) is the number of dimension of \\(\\mathbf{x}_0\\), and \\(\\alpha_{t}\\) is the variance schedule at timestep \\(t\\) (\\(t = \\overline{1;T}\\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.12 + ], + "angle": 0, + "content": "Proposition 1. Suppose that \\(\\tilde{\\mathbf{x}}_0\\), \\(\\mathbf{x}_0\\) and \\(\\epsilon\\) are independent, and that" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.12, + 0.781, + 0.158 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} \\geq M.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.164, + 0.895, + 0.194 + ], + "angle": 0, + "content": "Then there exists \\(C > 0\\) such that: for arbitrary \\(\\delta > 0\\), if \\(\\mathcal{L}_{\\mathrm{contrastive}} < \\delta\\), then" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.206, + 0.775, + 0.225 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\| \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\| _ {2} ^ {2} \\right] < C \\delta .\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.238, + 0.743, + 0.253 + ], + "angle": 0, + "content": "Proof. See Supplementary Material." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.238, + 0.892, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.269, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Remark 1.1. Proposition 1 suggests that if the contrastive loss \\(\\mathcal{L}_{\\mathrm{contrastive}}\\) tends to zero, then the prediction \\(\\tilde{\\mathbf{x}}_0\\) will approach the ground truth \\(\\mathbf{x}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.319, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Remark 1.2. The condition \\(\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 \\geq M\\) is suitable for our language-driven grasp detection task as \\(\\tilde{\\mathbf{x}}_0\\) and \\(\\mathbf{x}_T\\) are two contrastive quantities, therefore, we can assume there is a minimum distance between \\(\\tilde{\\mathbf{x}}_0\\) and \\(\\mathbf{x}_T\\). In addition, in the proof of Proposition 1, we see that \\(\\mathbb{E}\\left[\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 - M\\right] = \\beta^2\\mathbb{E}\\left[\\| \\tilde{\\mathbf{x}}_0 - \\mathbf{x}_0\\|_2^2\\right]\\), which is always nonnegative. Therefore, it is both theoretically and experimentally reasonable to add this assumption." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.48, + 0.909, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.697, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Figure 4. Language-drive Grasp Detection (LGD) network. We present the network architecture (left) and the proposed training objectives of the denoising process (right)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.756, + 0.868, + 0.772 + ], + "angle": 0, + "content": "4.2. Language-driven Grasp Detection Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Network. Our network operates on two conditions: an image denoted as \\( \\pmb{I} \\) and a corresponding text prompt represented as \\( e \\). To process these conditions, we employ a vision encoder to extract visual features from \\( \\pmb{I} \\) and a text encoder to derive textual embeddings from \\( e \\). The resulting feature vectors, denoted as \\( I' \\) and \\( e' \\), are subsequently subjected to a fusion module, ALBEF [39]. We leverage the attention mask generated by the ALBEF fusion module" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17906" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.092, + 0.531, + 0.102 + ], + "angle": 0, + "content": "Give me the bottle" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.103, + 0.208, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.103, + 0.345, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.103, + 0.481, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.103, + 0.617, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.103, + 0.754, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.103, + 0.891, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.209, + 0.566, + 0.22 + ], + "angle": 0, + "content": "Grasp the wristwatch at its dial" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.22, + 0.208, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.32, + 0.168, + 0.332 + ], + "angle": 0, + "content": "(a) Ours" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.22, + 0.345, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.32, + 0.327, + 0.331 + ], + "angle": 0, + "content": "(b) GR-ConvNet" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.22, + 0.481, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.32, + 0.47, + 0.332 + ], + "angle": 0, + "content": "(c) Det-Seg-Refine" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.22, + 0.617, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.32, + 0.59, + 0.332 + ], + "angle": 0, + "content": "(d) GG-CNN" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.22, + 0.754, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.32, + 0.727, + 0.332 + ], + "angle": 0, + "content": "(e) CLIPORT" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.22, + 0.891, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.32, + 0.871, + 0.331 + ], + "angle": 0, + "content": "(f) CLIP-Fusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.335, + 0.686, + 0.349 + ], + "angle": 0, + "content": "Figure 5. Language-driven grasp detection results visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.47, + 0.49 + ], + "angle": 0, + "content": "as the estimation \\(\\tilde{\\mathbf{x}}_0\\) of \\(\\mathbf{x}_0\\). Next, we aggregate three elements: the estimation region \\(\\tilde{\\mathbf{x}}_0\\), the grasp pose at the current timestep \\(\\mathbf{x}_{t + 1}\\), and the timestep \\(t + 1\\). These inputs are combined using MLP layers, similar to the approach outlined in [68]. Specifically, the output operation can be expressed as: \\(\\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0)\\), where function \\(f\\) encompasses a composition of multiple MLP layers. Additional specifics regarding these universal MLP layers are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Training Objective. In our context, conditioned grasp detection models the distribution \\( p(\\mathbf{x}_0|y) \\) as the reversed diffusion process of gradually cleaning \\( \\mathbf{x}_{t + 1} \\). Instead of predicting \\( \\mathbf{x}_t \\) as formulated by [28], we follow Ramesh et al. [58] and predict the signal itself, i.e., \\( \\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0) \\) with the simple objective [28]. To this end, we utilize the contrastive loss as in Equation 4 to explicitly improve the learning objective of the denoising process:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.629, + 0.469, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n t r a s t i v e}} + \\mathcal {L} _ {\\text {d i f f u s i o n}}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.659, + 0.21, + 0.676 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.469, + 0.76 + ], + "angle": 0, + "content": "We conduct experiments to evaluate our proposed method and Grasp-Anything++ dataset using both the vision-based metrics and real robot experiments. We then demonstrate zero-shot grasp results and discuss the challenges and open questions for future works." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.436, + 0.789 + ], + "angle": 0, + "content": "5.1. Language-driven Grasp Detection Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our language-driven grasp detection method (LGD) with the linguistically supported versions of GR-CNN [35], Det-Seg-Refine [2], GG-CNN [47], CLIPORT [64] and CLIP-Fusion [80]. In all cases, we employ a pretrained CLIP [57] or BERT [18] as the text embedding. The implementation details of all baselines can be found in our Supplementary Material." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.351, + 0.885, + 0.495 + ], + "angle": 0, + "content": "
BaselineSeenUnseenH
GR-ConvNet [35] + CLIP [57]0.370.180.24
Det-Seg-Refine [2] + CLIP [57]0.300.150.20
GG-CNN [47] + CLIP [57]0.120.080.10
CLIPORT [64]0.360.260.29
CLIP-Fusion [80]0.400.290.33
LGD (ours) + BERT [18]0.440.380.41
LGD (ours) + CLIP [57]0.480.420.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.54, + 0.5, + 0.852, + 0.514 + ], + "angle": 0, + "content": "Table 2. Language-driven grasp detection results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.525, + 0.892, + 0.691 + ], + "angle": 0, + "content": "Setup. To assess the generalization of all methods trained on Grasp-Anything++, we utilize the concept of base and new labels [92] in zero-shot learning. We categorize LVIS labels from Section 3.3 to form labels for our experiment. In particular, we select \\(70\\%\\) of these labels by frequency for 'Base' and assign the remaining \\(30\\%\\) to 'New'. We also use the harmonic mean ('H') to measure the overall success rates [92]. Our primary evaluation metric is the success rate, defined similarly to [35], necessitating an IoU score of the predicted grasp exceeding \\(25\\%\\) with the ground truth grasp and an offset angle less than \\(30^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.692, + 0.892, + 0.782 + ], + "angle": 0, + "content": "Main Results. Table 2 shows the results of language-driven grasp detection on the Grasp-Anything++ dataset. The findings indicate a notable performance advantage of our LGD over other baseline approaches, with LGD outperforming the subsequent best-performing baselines (CLIP-Fusion) by margins of 0.14 on Grasp-Anything++ dataset." + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.796, + 0.863, + 0.874 + ], + "angle": 0, + "content": "
BaselineSeenUnseenH
LGD w/o predicting x̂00.150.080.10
LGD w/o contrastive loss0.450.400.42
LGD w contrastive loss0.480.420.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.59, + 0.878, + 0.802, + 0.892 + ], + "angle": 0, + "content": "Table 3. Contrastive loss analysis." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17907" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.465, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.223, + 0.362, + 0.236 + ], + "angle": 0, + "content": "Figure 6. Loss visualization." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.244, + 0.275, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.244, + 0.468, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.43 + ], + "angle": 0, + "content": "Figure 7. t-SNE visualization. We apply t-SNE to cluster the vision-and-language features with and without contrastive loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.443, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Contrastive Loss Analysis. Table 3 presents the performance of LGD under varied configurations. The outcomes emphasize the substantial influence of the training objective (contrastive loss) and the importance of language instructions in enhancing LGD performance on both seen and unseen classes in the Grasp-Anything++ dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Fig. 6 shows the contrastive loss \\(\\mathcal{L}_{\\mathrm{contrastive}}\\) approaching towards 0 during training, indicating the grasp pose estimation \\(\\tilde{\\mathbf{x}}_0\\) aligns with the ground truth \\(\\mathbf{x}_0\\), as anticipated by Proposition 1. The subsequent attention maps visualization in Fig. 8 shows the attention region is meaningful and improves the results when employing our proposed contrastive loss compared to its absence. Moreover, we employ t-SNE for vision-and-language embedding visualization, as in [46], by processing 2,000 samples from the Grasp-Anything++ dataset through the ALBEF module. The outcomes reveal that our contrastive loss facilitates better object classification, as evidenced in Fig.7 by clearer segregation of pixel embeddings across various semantic classes, underscoring contrastive loss's role in refining embeddings' differentiation for improved class distinctions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 5 presents qualitative results of the language-driven grasp detection task, suggesting that our LGD method generates more semantically plausible than other baselines. Despite satisfactory performance, LGD occasionally predicts incorrect results, with a detailed analysis of these cases available in our Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Robotic Validation. We provide quantitative results by integrating our language-driven grasp detection pipeline for a robotic grasping application with a KUKA LBR iwa" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.295, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 8. Attention map visualization. We compare the attention map when utilizing our proposed contrastive loss and when not." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.325, + 0.875, + 0.449 + ], + "angle": 0, + "content": "
BaselineSingleCluttered
GR-ConvNet [35] + CLIP [57]0.330.30
Det-Seg-Refine [2] + CLIP [57]0.300.23
GG-CNN [47] + CLIP [57]0.100.07
CLIPORT [64]0.270.30
CLIP-Fusion [80]0.400.40
LGD (ours)0.430.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.459, + 0.877, + 0.473 + ], + "angle": 0, + "content": "Table 4. Robotic language-driven grasp detection results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.686 + ], + "angle": 0, + "content": "R820 robot. Using the RealSense D435i camera, the grasp pose inferred from approaches in Table 4 is transformed into the 6DoF grasp pose, similar to [35]. The optimization-based trajectory planner in [5,72] is employed to execute the grasps. Experiments are conducted for two scenarios, i.e., the single object scenario and the cluttered scene scenario, of a set of 20 real-world daily objects. In each scenario, we run 30 experiments using baselines listed in Table 4 and a predefined grasping prompt corpus. The results exhibit that our LGD outperforms other baselines. Furthermore, although LGD is trained on our Grasp-Anything++ which is a solely synthesis dataset created by foundation models, it still shows reasonable results on real-world objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.741, + 0.712 + ], + "angle": 0, + "content": "5.2. Zero-shot Grasp Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Our proposed Grasp-Anything++ is a large-scale dataset. Apart from the language-driven grasp detection task, we believe it can be used for other purposes. In this experiment, we seek to answer the question: Can Grasp-Anything++ be useful in the traditional grasp detection task without text? Consequently, we verify our Grasp-Anything++ and LGD (no text) with other existing datasets and grasping methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Setup. We setup an LGD (no text) version, and other state-of-the-art grasp detection methods GR-ConvNet [35], Det-Seg-Refine [2], GG-CNN [47]. We use five datasets: our Grasp-Anything++, Jacquard [17], Cornell [33], VMRD [87], and OCID-grasp [2] in this experiment." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17908" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.1, + 0.891, + 0.203 + ], + "angle": 0, + "content": "
BaselineGrasp-Anything++ (ours)Jacquard [17]Cornell [33]VMRD [87]OCID-grasp [2]
BaseNewHBaseNewHBaseNewHBaseNewHBaseNewH
GR-ConvNet [35]0.710.590.640.880.660.750.980.740.840.770.640.700.860.670.75
Det-Seg-Refine [2]0.620.570.590.860.600.710.990.760.860.750.600.660.800.620.70
GG-CNN [47]0.680.570.620.780.560.650.960.750.840.690.530.590.710.630.67
LGD (no text) (ours)0.740.630.680.890.690.770.970.760.850.790.660.720.880.680.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.21, + 0.658, + 0.224 + ], + "angle": 0, + "content": "Table 5. Base-to-new zero-shot grasp detection results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.234, + 0.471, + 0.356 + ], + "angle": 0, + "content": "Zero-shot Results. Table 5 summarizes the base-to-new grasp detection results on five datasets. Overall, the performance of LGD even without the language branch is better than other baselines across all datasets. Furthermore, this table also shows that our Grasp-Anything++ dataset is more challenging to train as the detection results are lower than related datasets using the same approaches due to the greater coverage of unseen objects in the testing phase." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.368, + 0.47, + 0.448 + ], + "angle": 0, + "content": "
Test TrainJacquardCornellVMRDOCID-graspGrasp-Anything++
Jacquard [17]0.870.510.130.210.17
Cornell [33]0.070.980.200.120.13
VMRD [87]0.060.210.790.110.10
OCID-grasp [2]0.090.120.200.740.11
Grasp-Anything++ (ours)0.410.630.300.390.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.459, + 0.411, + 0.473 + ], + "angle": 0, + "content": "Table 6. Cross-dataset grasp detection results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Cross-dataset Evaluation. To further verify the usefulness of our Grasp-Anything++ dataset, we conduct the cross-dataset validation in Table 6. We use the GRConvNet [35] to reuse its results on existing grasp datasets. GR-ConvNet is trained on a dataset (row) and evaluated on another dataset (column). For example, training on Jacquard and testing on Cornell yields an accuracy of 0.51. Notably, training with our dataset improves performance by approximately \\(10 - 33\\%\\) compared to other datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.749 + ], + "angle": 0, + "content": "In the wild grasp detection. Fig. 9 shows visualization results using LGD (no text) trained on our Grasp-Anything++ dataset on random internet images and other datasets images. We can see that the detected grasp poses are adequate in quality and quantity. This demonstrates that although our Grasp-Anything++ is fully created by foundation models without having any real images, models trained on our Grasp-Anything++ dataset still generalize well on real-world images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.195, + 0.771 + ], + "angle": 0, + "content": "5.3. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our experiments indicate that Grasp-Anything++ can serve as a foundation dataset for both language-driven and traditional grasp detection tasks. However, there are certain limitations. First, our dataset lacks depth images for directly being applied to robotic applications [52]. Second, we remark that the creation of our dataset is time-consuming and relies on access to the ChatGPT API. Fortunately, future research can reuse our provided assets (images, prompts," + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.231, + 0.626, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.231, + 0.756, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.231, + 0.887, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.332, + 0.626, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.332, + 0.756, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.332, + 0.887, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Figure 9. In the wild grasp detection. Top row images are from GraspNet [24], YCB-Video [78], NBMOD [9] datasets; bottom row shows internet images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.545 + ], + "angle": 0, + "content": "etc.) without starting from scratch. Furthermore, our experiments show that adding language to the grasp detection task (Table 2) poses a more challenging problem compared to standard grasp detection task (Table 5)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.684 + ], + "angle": 0, + "content": "We see several interesting future research directions. First, future work could investigate the use of text or imaged-to-3D models [79] or image-to-depth [59] and reuse our dataset's prompts and images to construct 3D language-driven grasp datasets. Additionally, beyond linguistic grasp instruction adherence, our dataset holds potential for varied applications, including scene understanding [81] and scene generation [3], hallucination analysis [30], and humanrobot interaction [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.707, + 0.619, + 0.723 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce Grasp-Anything++, a large-scale dataset with 1M images and 10M grasp prompts for language-driven grasp detection tasks. We propose LGD, a diffusion-based method to tackle the language-driven grasp detection task. Our diffusion model employs a contrastive training objective, which explicitly contributes to the denoising process. Empirically, we have shown that Grasp-Anything++ serves as a foundation grasp detection dataset. Finally, our LGD improves the performance of other baselines, and the real-world robotic experiments further validate the effectiveness of our dataset and approach." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17909" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Stefan Ainetter and Friedrich Fraundorfer. End-to-end trainable deep neural network for robotic grasp detection and semantic segmentation from rgb. In ICRA, 2021. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.471, + 0.271 + ], + "angle": 0, + "content": "[3] Dor Arad Hudson and Larry Zitnick. Compositional transformers for scene generation. NeurIPS, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[4] Umar Asif, Jianbin Tang, and Stefan Harrer. Grasynet: An efficient convolutional neural network for real-time grasp detection for low-powered devices. In IJCAI, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.371 + ], + "angle": 0, + "content": "[5] Florian Beck, Minh Nhat Vu, Christian Hartl-Nesic, and Andreas Kugi. Singularity avoidance with application to online trajectory optimization for serial manipulators. IFAC-PapersOnLine, 56(2):284-291, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.469, + 0.541 + ], + "angle": 0, + "content": "[8] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[9] Boyuan Cao, Xinyu Zhou, Congmin Guo, Baohua Zhang, Yuchen Liu, and Qianqiu Tan. Nbmod: Find it and grasp it in noisy background. arXiv preprint arXiv:2306.10265, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[10] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.711 + ], + "angle": 0, + "content": "[11] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[12] I-Ming Chen and Joel W Burdick. Finding antipodal point grasps on irregularly shaped objects. T-RA, 1993. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[13] Sijia Chen and Baochun Li. Language-guided diffusion model for visual grounding. arXiv preprint arXiv:2308.09599, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[14] Hongjun Choi, Anirudh Som, and Pavan Turaga. Amc-loss: Angular margin contrastive loss for improved explainability in image classification. In CVPR workshops, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[15] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. TPAMI, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[17] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IROS, 2018. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[20] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[22] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. A billion ways to grasp: An evaluation of grasp sampling schemes on a dense, physics-based grasp data set. In ISRR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[23] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In ICRA, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[24] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[25] Maximilian Gilles, Yuhao Chen, Tim Robin Winter, E Zhixuan Zeng, and Alexander Wong. Metagraspnet: A large-scale benchmark dataset for scene-aware ambidextrous bin picking via physics-based metaverse synthesis. In CASE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[27] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTAT, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.706 + ], + "angle": 0, + "content": "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[31] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[33] Yun Jiang, Stephen Moseson, and Ashutosh Saxena. Efficient grasping from rgbd images: Learning using a new rectangle representation. In ICRA, 2011. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.192 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.47, + 0.234 + ], + "angle": 0, + "content": "[35] Sulabh Kumra, Shirin Joshi, and Ferat Sahin. Antipodal robotic grasping using generative residual convolutional neural network. In IROS, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.47, + 0.305 + ], + "angle": 0, + "content": "[36] Jaewoong Lee, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, Yunji Kim, Jin-Hwa Kim, Jung-Woo Ha, and Sung Ju Hwang. Text-conditioned sampling framework for text-to-image generation with masked generative models. arXiv preprint arXiv:2304.01515, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.306, + 0.47, + 0.362 + ], + "angle": 0, + "content": "[37] Meng-Lun Lee, Sara Behdad, Xiao Liang, and Minghui Zheng. Task allocation and planning for product disassembly with human-robot collaboration. Robotics and Computer-Integrated Manufacturing, 76, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.47, + 0.418 + ], + "angle": 0, + "content": "[38] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. *IJRR*, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[40] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[41] Jirong Liu, Ruo Zhang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Chenxi Wang, Sheng Xu, Hengxu Yan, and Cewu Lu. Target-referenced reactive grasping for dynamic objects. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[43] Weiyu Liu, Tucker Hermans, Sonia Chernova, and Chris Paxton. Structdiffusion: Object-centric diffusion for semantic rearrangement of novel objects. arXiv preprint arXiv:2211.04604, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[44] Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-Radiology, page 100017, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[45] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[46] Jiaxu Miao, Zongxin Yang, Leilei Fan, and Yi Yang. Fedseg: Class-heterogeneous federated learning for semantic segmentation. In CVPR, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[47] Douglas Morrison, Peter Corke, and Jürgen Leitner. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. arXiv preprint arXiv:1804.05172, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.192 + ], + "angle": 0, + "content": "[48] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. RA-L, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[49] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[50] Yao Mu, Shunyu Yao, Mingyu Ding, Ping Luo, and Chuang Gan. Ec2: Emergent communication for embodied control. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.346 + ], + "angle": 0, + "content": "[51] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.348, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[52] Rhys Newbury, Morris Gu, Lachlan Chumbley, Arsalan Mousavian, Clemens Eppner, Jürgen Leitner, Jeannette Bohg, Antonio Morales, Tamim Asfour, Danica Kragic, et al. Deep learning approaches to grasp synthesis: A review. T-RO, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.893, + 0.46 + ], + "angle": 0, + "content": "[53] Toan Nguyen, Minh Nhat Vu, An Vuong, Dzung Nguyen, Thieu Vo, Ngan Le, and Anh Nguyen. Open-vocabulary affordance detection in 3d point clouds. In IROS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.893, + 0.487 + ], + "angle": 0, + "content": "[54] OpenAI. Introducing ChatGPT. Software. Accessed: July 6th 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.489, + 0.893, + 0.531 + ], + "angle": 0, + "content": "[55] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[56] Robert Platt. Grasp learning: Models, methods, and performance. Annual Review of Control, Robotics, and Autonomous Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.893, + 0.644 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.645, + 0.893, + 0.7 + ], + "angle": 0, + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[59] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In CVP4, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.829 + ], + "angle": 0, + "content": "[61] Kallol Saha, Vishal Mandadi, Jayaram Reddy, Ajit Srikanth, Aditya Agarwal, Bipasha Sen, Arun Singh, and Madhava Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. arXiv preprint arXiv:2309.11414, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.83, + 0.893, + 0.87 + ], + "angle": 0, + "content": "[62] Schuhmann et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[63] Dhruv Shah, Błajej Osiński, brian richter, and Sergey Levine. Lm-nav: Robotic navigation with large pre-trained models of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "language, vision, and action. In Karen Liu, Dana Kulic, and Jeff Ichnowski, editors, Proceedings of The 6th Conference on Robot Learning, volume 205 of PMLR. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[64] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.47, + 0.223 + ], + "angle": 0, + "content": "[65] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.224, + 0.47, + 0.279 + ], + "angle": 0, + "content": "[66] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.282, + 0.47, + 0.337 + ], + "angle": 0, + "content": "[67] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.34, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[68] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.383, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[69] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[70] Julien Urain, Niklas Funk, Jan Peters, and Georgia Chalvatzaki. Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In ICRA, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.513 + ], + "angle": 0, + "content": "[71] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.515, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[72] Minh Nhat Vu, Florian Beck, Michael Schwegel, Christian Hartl-Nesic, Anh Nguyen, and Andreas Kugi. Machine learning-based framework for optimally solving the analytical inverse kinematics for redundant manipulators. Mechatronics, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[73] An Vuong, Minh Nhat Vu, Toan Tien Nguyen, Baoru Huang, Dzung Nguyen, Thieu Vo, and Anh Nguyen. Language-driven scene synthesis using multi-conditional diffusion model. arXiv preprint arXiv:2310.15948, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[74] An Dinh Vuong, Minh Nhat Vu, Hieu Le, Baoru Huang, Binh Huynh, Thieu Vo, Andreas Kugi, and Anh Nguyen. Grasp-anything: Large-scale grasp dataset from foundation models. arXiv preprint arXiv:2309.09818, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.77 + ], + "angle": 0, + "content": "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[76] Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania, and Ernesto W De Luca. Legal norm retrieval with variations of the bert model combined with tfidf vectorization. In Proceedings of the eighteenth international conference on artificial intelligence and law, pages 285-294, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[77] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In MIDL, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[78] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[79] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[80] Kechun Xu, Shuqi Zhao, Zhongxiang Zhou, Zizhang Li, Huajin Pi, Yifeng Zhu, Yue Wang, and Rong Xiong. A joint modeling of vision-language-action for target-oriented grasping in clutter. arXiv preprint arXiv:2302.12610, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[81] Yiteng Xu, Peishan Cong, Yichen Yao, Runnan Chen, Yuenan Hou, Xinge Zhu, Xuming He, Jingyi Yu, and Yuexin Ma. Human-centric scene understanding for 3d large-scale scenarios. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.332, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[82] Zhixuan Xu, Kechun Xu, Rong Xiong, and Yue Wang. Object-centric inference for language conditioned placement: A foundation model based approach. In ICARM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.388, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[83] Jange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pick-place robots. arXiv preprint arXiv:2306.05716, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.445, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[84] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 2022, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.515, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[85] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.557, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[86] Yang Yang, Xibai Lou, and Changyun Choi. Interactive robotic grasping with attribute-guided disambiguation. In ICRA, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[87] Hanbo Zhang, Xuguang Lan, Site Bai, Xinwen Zhou, Zhiqiang Tian, and Nanning Zheng. Roi-based robotic grasp detection for object overlapping scenes. In IROS, 2019. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.656, + 0.892, + 0.71 + ], + "angle": 0, + "content": "[88] Hanbo Zhang, Yunfan Lu, Cunjun Yu, David Hsu, Xuguang La, and Nanning Zheng. Invigorate: Interactive visual grounding and grasping in clutter. arXiv preprint arXiv:2108.11092, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.712, + 0.892, + 0.753 + ], + "angle": 0, + "content": "[89] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.755, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[90] Kaizhi Zheng, Xiaotong Chen, Odest Chadwicke Jenkins, and Xin Wang. Vlmbench: A compositional benchmark for vision-and-language manipulation. NeurIPS, 35, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.798, + 0.892, + 0.852 + ], + "angle": 0, + "content": "[91] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.853, + 0.892, + 0.894 + ], + "angle": 0, + "content": "[92] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.894 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17912" + } + ] +] \ No newline at end of file diff --git a/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_origin.pdf b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ea9fb35900fba489109663157aa07e42959da0d7 --- /dev/null +++ b/2024/Language-driven Grasp Detection/4da4eafd-acd2-41ed-99a7-76c7e2b26e40_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:769fc7b61acd762fe90a8862936ae9e17c73acac905427b12d1cdc9fa5e13907 +size 10077104 diff --git a/2024/Language-driven Grasp Detection/full.md b/2024/Language-driven Grasp Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c21d92594f182e626c2afc8ce2c37ff0a0e019d --- /dev/null +++ b/2024/Language-driven Grasp Detection/full.md @@ -0,0 +1,412 @@ +# Language-driven Grasp Detection + +An Dinh Vuong $^{1}$ , Minh Nhat Vu $^{2,*}$ , Baoru Huang $^{3,*}$ , Nghia Nguyen $^{1}$ , Hieu Le $^{1}$ , Thieu Vo $^{4}$ , Anh Nguyen $^{5}$ + +$^{1}$ FPT Software AI Center, Vietnam $^{2}$ Automation & Control Institute, TU Wien, Austria $^{3}$ Imperial College London, UK + +$^{4}$ Ton Duc Thang University, Vietnam $^{5}$ University of Liverpool, UK *Co-Corresponding authors + +https://airvlab.github.io/grasp-anything + +![](images/7e25ba918d52692bb9f2a2f4cef694e7122226554dfece7ff4a35af51422c547.jpg) +A coffee cup, a grape biscuit, a black oven plate on a white dining table +Grasp the coffee cup. + +![](images/7cfffa3db9198cd347875d81b81bd2d3792ab6bd0989dd95035c7de9dfa5b2a5.jpg) +A black pen, a steel marble and a digital calculator on an office desk +Grasp the calculator at its keypad. +Figure 1. We present a new dataset and method for language-driven grasp task. + +![](images/f1f93e7891e5c35da4bbcb99c8af49cc675eee568a3a9ca6cd60f62d9ea6d614.jpg) +A black pot, a green bottle, a bread container arranged on a kitchen counter +Grab the neck of the green bottle. + +![](images/9bafff9a75f75acdec949e96af0d9469c76a55c459bdcda095c0afb8aa904730.jpg) +A brown-band wristwatch, a sleek black pen and an office clip placed on a table +Give me the brown-band wristwatch. + +![](images/a3b83d90a97a9e2785ee86fd3954c919f23f96693945dc7bc25ba5b86f32d7e3.jpg) +Pick the fountain pen at its cap. + +# Abstract + +Grasp detection is a persistent and intricate challenge with various industrial applications. Recently, many methods and datasets have been proposed to tackle the grasp detection problem. However, most of them do not consider using natural language as a condition to detect the grasp poses. In this paper, we introduce Grasp-Anything++, a new language-driven grasp detection dataset featuring 1M samples, over 3M objects, and upwards of 10M grasping instructions. We utilize foundation models to create a large-scale scene corpus with corresponding images and grasp prompts. We approach the language-driven grasp detection task as a conditional generation problem. Drawing on the success of diffusion models in generative tasks and given that language plays a vital role in this task, we propose a new language-driven grasp detection method based on diffusion models. Our key contribution is the contrastive training objective, which explicitly contributes to the denoising process to detect the grasp pose given the language instructions. We illustrate that our approach is theoretically supportive. The intensive experiments show that our method outperforms state-of-the-art approaches and allows real-world robotic grasping. Finally, we demonstrate our large-scale dataset enables zero-short grasp detection and is a challenging benchmark for future work. + +# 1. Introduction + +Imagine we want an assistant robot to grasp a cup among a clutter of daily objects such as a knife, a fork, a cup, and a pair of scissors. Conventionally, to convey the idea of grasping this specific object, humans use the natural language command, "give me the cup", for instance. Although humans intuitively know how to grasp the cup given the linguistic command, determining specific grasp actions for objects based on natural language instructions or language-driven grasp detection remains challenging for robots [64]. First, natural language is usually overlooked in existing grasp datasets [56] while training vision-and-language neural networks necessitates an excessive number of labeled examples [66]. Second, recent works usually focus on particular manipulation tasks with limited objects [25], imposing a bottleneck for in-the-wild robot execution [63]. Finally, despite recent developments, bridging the gap between language, vision, and control for real-world robotic experiments remains a challenging task [83]. + +Recently, language-driven robotic frameworks are gaining traction, offering the potential for robots to process natural language, and bridging the gap between robotic manipulations and real-world human-robot interaction [51]. PaLM-E [21], EgoCOT [51], and ConceptFusion [32] are some notable embodied robots with the ability to compre + +hend natural language by harnessing the power of large foundation models such as ChatGPT [54]. However, most works assume the high-level actions of robots and ignore the fundamental grasping actions, restricting the structure for generalization across robotic domains, tasks, and skills [50]. In this paper, we explore training a language-driven agent to implement low-level actions, focusing on the task of object grasping via image observations. Specifically, our hypothesis is centered around the establishment of a robotic system that can execute grasping actions following a given language instruction for any universal object. + +We first present Grasp-Anything++ to serve as a large-scale dataset for language-driven grasp detection. Our dataset is based on the Grasp-Anything [74], and is synthesized from foundation models. Compared to the original Grasp-Anything dataset, we provide more than 10M grasp prompts and 3M associated object masks, 6M ground truth poses at the object part level. Our dataset showcases the ability to facilitate grasp detection using language instructions. We label the ground truth at both the object level and part level, providing a comprehensive understanding of real-world scenarios. For example, our ground truth includes both general instructions "give me the knife" and detail ones such as "grasp the handle of the steak knife". We empirically show that our large-scale dataset successfully facilitates zero-shot grasp detection on both vision-based tasks and real-world robotic experiments. + +To tackle the challenging language-driven grasp detection task, we propose a new diffusion model-based method. Our selection of diffusion models is motivated by their proven efficacy in conditional generation tasks [28]. These models have shown efficiency beyond image synthesis, including other image-based tasks such as image segmentation [77], and visual grounding [40]. Despite achieving notable success, integrating visual and text features effectively remains a challenge [13] as the majority of existing literature employs latent strategies to combine visual and text features [36]. We address this challenge by employing a new training strategy for learning text and image features, focusing on the use of feature maps as guidance information for grasp pose generation. Our main contribution is a new training objective that incorporates the feature maps and explicitly contributes to the denoising process. In summary, our contributions are three-fold: + +- We propose Grasp-Anything++, a large-scale language-driven dataset for grasp detection tasks. +- We propose a diffusion model with a training objective that explicitly contributes to the denoising process to detect the grasp poses. +- We demonstrate that our Grasp-Anything++ dataset and the proposed method outperform other approaches and enable successful robotic applications. + +# 2. Related Work + +Grasp Detection. Grasp detection is a popular task in both computer vision and robotic community [2, 17, 24, 64]. Recently, establishing robotic systems with the ability to follow natural commands has been actively researched [64, 80, 83]. The prevalent solution to the language-driven grasp detection task is to split into two stages: one for grounding the target object, and the other is to synthesize grasp poses from the grounding visual-text correlations [1, 80]. Training in two stages may result in longer inference time [41]. In addition, several works [82, 83, 90] adopt foundation models, such as GroundDINO [42] and GPT-3 [7]. Accessing such commercial foundation models is not always available [71], especially on robotic systems with limited resources or unstable internet connection [37]. In our work, we directly train the model on the large-scale Grasp-Anything++ dataset to inherit the power of a foundation-based dataset, while ensuring a straightforward inference process for the downstream robotic applications. + +Language-driven Grasp Detection Datasets. While there are many grasp datasets have been introduced [9, 17, 22-24, 33, 38, 45, 48, 49, 55, 78], the majority of them overlook the text modality. Therefore, the grasping of objects out of a clutter typically experiences ambiguities in what object to grasp [88]. DailyGrasp [83] is one of the first grasp datasets employing natural language for scene descriptions; however, the scene description corpus in this dataset is relatively small and does not specify which part of the object should be grasped. In our work, we present Grasp Anything++, which is a large-scale language-driven grasping dataset. Grasp-Anything++ provides object instance description with a natural word phrase, for example, "a blue ceramic vase". Furthermore, Grasp-Anything++ describes the grasping object at both the part level and object level, providing more information for the robot to execute the grasping [86]. + +Diffusion Models for Robotic Applications. Diffusion models [28] have emerged as the new state-of-the-art method of generative tasks [84]. Recent have witnessed growing attention for utilizing diffusion models in robotic applications [61]. Liu et al. [43] propose a diffusion model to handle the language-guided object rearrangement task. Diffusion models are also applied to other robotic tasks such as motion planning [10], and trajectory optimization [31]. The authors in [70] present a diffusion model to determine grasp poses by minimizing a SDF loss. Overall, the diffusion models employed in previous works often combine visual and text features in a latent mechanism [6], which may cause interpretability problems [83] for robotic systems that require low-level controls [20]. To tackle this challenge, we propose a training objective that explicitly contributes to the denoising process. We demonstrate that our proposed strategy is more effective than the latent strategy. + +
StepDescription
Scene GenerationUserPlease help me generate scene descriptions for natural arrangements of daily objects. Each description has the following form: <Object_1><Object_2>...<Verb><Container Object>. Please also ensure the incorporation of a rich and varied lexicon in the scene descriptions.
SampleA steel knife, a polished fork and a pristine ceramic plate on a wooden table.
Text-to-ImageWe use Stable Diffusion [60] to proceed text-to-image generation.
Object MaskingUserFor object part-level description, given an input list {<Object_1>, <Object_2>, ...}, the output will be a list that describes the parts of objects as: {<Object_1>: [(Part_1.1>, <Part_1.2>, ..., <Object_2>]: [(Part_2.1>, <Part_2.2>, ...]}.
Sample{knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}
Post ProcessWe use OFA [75] and SAM [34] to locate the region describing the objects.
Part MaskingUserGiven the object list and part lists of each scene description, you will generate for me all prompts with the following format: {<Manipulation Action><Object_ID><Part_ID)}. The part that is more suitable for human grasping is positioned at the start of the list to represent the grasping actions.
SampleGive me the steel knife; Grasp the knife at its handle.
Post ProcessWe leverage VLPart [67] to locate the region describing the parts of objects.
Grasp GenerationUserGenerate for me a scene description with grasp instructions following the templates.
SampleScene description: A steel knife, a polished fork and a pristine ceramic plate on a wooden table. Object list: {knife, fork, plate}. Part lists: {knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}. Prompts: Give me the steel knife; Grasp the knife at its handle.
Grasp LabellingWe utilize a pretrained RAGT-3/3 [9] to generate grasp poses corresponding to the located region.
+ +Table 1. Grasp-Anything++ creation pipeline. We utilize ChatGPT to generate scene descriptions and grasp instructions from the user input. We generate images given scene descriptions and automatically synthesize the grasp poses. + +# 3. The Grasp-Anything++ Dataset + +We utilize large-scale foundation models to create the Grasp-Anything++. Our dataset offers open-vocabulary grasping commands and images with associated groundtruth. There are three key steps in establishing our dataset: i) prompting procedure, ii) image synthesis and grasp poses annotation, and iii) post-processing. + +# 3.1. Prompting Procedure + +We first establish prompt-based procedures to generate a large-scale scene description corpus as well as grasp prompt instructions. In particular, we utilize ChatGPT to generate the prompts for two tasks: $i$ ) Scene descriptions: Sentences capturing the scene arrangement, including the extracted object and part lists, and $ii$ ) Grasp instructions: Prompts directing the robot to grasp specific objects or parts. + +We follow a procedure in Table 1 to implement ChatGPT's output templates. The reference target in the grasp instruction may be either an object or an object's part. When the reference is an object's part, that part is directly selected as the reference in the grasp instruction sentence. If the reference is an object, we determine the grasping region on the part of the object that is likely to be grasped in everyday scenarios as described in affordance theory [53]. + +# 3.2. Image Synthesis and Grasp Annotation + +Image Synthesis. Given the scene description corpus, we first utilize a large-scale pretrained text-to-image model, namely, Stable Diffusion [60] to generate images from scene descriptions. Next, we perform a series of visual grounding and image segmentation using OFA [75], Segment-Anything [34], and VLPart [67] to locate the referenced object or part to the grasp instruction. + +Grasp Annotation. Grasp poses are represented as 2D rectangles, consistent with prior research and practical compatibility with real-world parallel grippers [17, 33]. Utilizing a pretrained network [9], we annotate grasp poses based on part segmentation masks. Since potential inaccuracies in these candidate poses could occur, we follow the procedure as defined in [74] to evaluate the quality of generated grasp poses to discard unreasonable grasp poses. + +Specifically, grasp quality is evaluated through net torque $\mathcal{T} = (\tau_{1} + \tau_{2}) - RMg$ , where resistance at contact points is $\tau_{i} = K\mu_{s}F\cos \alpha_{i}$ . With constants such as $M$ (mass), $g$ (gravitational acceleration), $K$ (geometrical characteristics), $\mu_{s}$ (static friction coefficient), and $F$ (applied force), accurately determining $\mathcal{T}$ directly is challenging due to the physical difficulties in precisely measuring $M$ , $K$ , and $\mu_{s}$ . Thus, we employ a surrogate measure, + +![](images/6dd8292cf455040829c3b39186f0b5587aa1a90bc9ba6d2451d5fef7ba4d0903.jpg) +(a) Number of categories + +![](images/b5a413908e008f1647111a74eddf433d39bffa972ced42a05b9d3dea412a1315.jpg) +(b) Scene description length +Figure 2. Data statistics. We analyze object categories and the use of natural language in the Grasp-Anything++ dataset. + +![](images/a122adec72ea413f0b6ecbafc6da97578f717f46af82b0f44385d364baa1b675.jpg) +(c) Number of object/part references. + +$\tilde{\mathcal{T}} = \frac{\cos\alpha_1 + \cos\alpha_2}{R}$ , as an alternative. As a result in [12], antipodal grasps score higher on $\tilde{\mathcal{T}}$ , indicating better quality. Consequently, grasps are evaluated based on $\tilde{\mathcal{T}}$ , with positive values indicating positive grasps and others considered as negative. + +Post Processing. Despite training on extensive datasets, Stable Diffusion [60] may produce subpar content, commonly termed as hallucination [30] when generating images from the text prompts. To address this, we perform manual reviews to filter out such images, with qualitative examples in our figures. Our process includes checks at every stage to prevent duplicate or hallucinated content. However, manual inspection introduces biases, which we counter with guidelines focusing on abnormal structures or implausible gravity (Fig. 3), aligns with approach in the literature [62]. + +![](images/00cfd26594998ed4dacfb405fc228105fffad00c13fc6abddd77333389d6a20f.jpg) +Figure 3. Failure image generation cases. Images generated by Stable Diffusion [60] may exhibit hallucinatory artifacts such as sunglasses lacking a lens, scissors with an anomalous structure, and a spoon not resting properly on a table. + +![](images/7efe25e835b004bb6cabb821598425187b0795efc3c7b6460d93172f8dfd0bce.jpg) + +![](images/911907bda31ab4c17981e2704b4571e25edea51096c28024977411aed9caed9a.jpg) + +Additionally, ChatGPT-generated scene prompts often duplicate [44]. To address this, we use duplication checking, filtering out identical prompts with BERTScore [89], which assesses sentence similarity through cosine similarities of token embeddings. We remove sentences with a BERTScore above 0.85 as in prior study [76]. + +# 3.3. Data Statistics + +Number of Categories. To evaluate the object category diversity, we apply a methodology akin to that in [16]. Utilizing 300 categories from LVIS dataset [26], we employ a pretrained model [91] to identify 300 candidate objects from our dataset for each category. We then curate a subset comprising 90,000 objects, refining it by excluding items that do not align semantically with their designated categories. A category is considered significant if it has more + +than 40 objects. Fig. 2a shows the results. Overall, our dataset spans over 236 categories from LVIS dataset, indicating a notable degree of object diversity in our dataset. + +Scene Descriptions. Fig. 2b shows the distribution of scene descriptions based on sentence length. The analysis reveals a wide range of sentence lengths, spanning from 10 words to 100 words per sentence. On average, each scene description consists of approximately 54 words, indicative of detailed and descriptive sentences. These scene descriptions correspond to sets of grasp instructions. Fig. 2c further shows the objects and object parts in scene descriptions. + +Diversity Analysis. We assess the diversity of occlusion and lighting conditions in the dataset. Regarding occlusion, we use a pretrained YOLOv5 model to identify objects within images. The results indicate that $93.8\%$ of images have a substantial overlap of five or more bounding boxes, which suggests a diverse range of occlusion within the Grasp-Anything++ dataset. Regarding lighting conditions, we convert images to YCbCr to analyze Y channel (luminance) and find that GraspL1M has the most diverse lighting conditions, identifying by the lowest Gini coefficient (a metric to measure the inequality of a distribution) of 0.26, compared to VMRD [87] (0.31), OCID-grasp [2] (0.32), Cornell [33] (0.62), Jacquard [17] (0.91). + +# 4. Language-driven Grasp Detection + +Motivation. The use of diffusion model for language-driven grasp detection is motivated by its efficiency in various generative tasks [15,28,43,84]. Conditional generation, such as our language-driven grasp detection task, aligns seamlessly with diffusion models' capabilities [29]. Moreover, language-driven grasp detection represents a fine-grained problem in which the outputs strongly depend on the text input [4]. For example, "grasp the steak knife" and "grasp the kraft knife" refer to two different objects on the image. To this end, we propose using contrastive loss with diffusion model to tackle this task, as contrastive learning is a popular solution for fine-grained tasks [8, 14, 85]. + +# 4.1. Constrastive Loss for Diffusion Model + +We represent the target grasp pose as $\mathbf{x}_0$ in the diffusion model. The objective of our diffusion process of language-driven grasp detection involves denoising from a noisy state + +$\mathbf{x}_T$ to the original grasp pose $\mathbf{x}_0$ , conditioned on the input image and grasp instruction represented by $y$ . + +In a diffusion process [28], assume that $q(\mathbf{x}_{1:T}|\mathbf{x}_0)$ is the forward process and we parameterize the reverse process by $p_{\theta}(\mathbf{x}_{0:T})$ . The conditional diffusion process [19] assumes $\hat{q}$ is the forward process but with the inclusion of a condition $y$ . The goal of the reverse process is to optimize the variational bound on negative log likelihood [28] + +$$ +\mathcal {L} = \mathbb {E} \left[ - \log p _ {\theta} (\mathbf {x} _ {T}) - \sum_ {t \geq 1} \log \frac {p _ {\theta} (\mathbf {x} _ {t - 1} | \mathbf {x} _ {t})}{q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1})} \right]. \tag {1} +$$ + +We prove in the Appendix that + +$$ +\begin{array}{l} \mathcal {L} = \mathbb {E} \left[ \underbrace {C - \log p _ {\theta} (\mathbf {x} _ {T})} _ {\text {C o n s t a n t}} + \underbrace {\log \hat {q} (\mathbf {x} _ {0} | \mathbf {x} _ {T} , y)} _ {\text {C o n t r a s t i v e}} + \right. \\ \underbrace {\sum_ {t > 1} D _ {\mathrm {K L}} \left(\hat {q} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t} , \mathbf {x} _ {0} , y\right) \| p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right)\right) - \log p _ {\theta} \left(\mathbf {x} _ {0} \mid \mathbf {x} _ {1} , y\right)} _ {\text {D e n o i s i n g s c o r e}}. \tag {2} \\ \end{array} +$$ + +The terms $D_{\mathrm{KL}}(\hat{q} (\mathbf{x}_{t - 1}|\mathbf{x}_t,\mathbf{x}_0,y)\| p_\theta (\mathbf{x}_{t - 1}|\mathbf{x}_t))$ and $\log p_{\theta}(\mathbf{x}_0|\mathbf{x}_1,y)$ in Equation 2 are similar to the concept of denoising score used in [28]. Thus, we can represent the quantity $\sum_{t > 1}D_{\mathrm{KL}}(\hat{q} (\mathbf{x}_{t - 1}|\mathbf{x}_t,\mathbf{x}_0,y)\| p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t))-$ $\log p_{\theta}(\mathbf{x}_0|\mathbf{x}_1,y)$ by the loss utilized in [68, 69] + +$$ +\mathcal {L} _ {\text {d i f f u s i o n}} = \mathbb {E} _ {\mathbf {x} _ {0} \sim q (\mathbf {x} _ {0} | y), t \sim [ 1, T ]} \left[ \mathbf {x} _ {0} - f \left(\mathbf {x} _ {t + 1}, t + 1, \tilde {\mathbf {x}} _ {0}\right) \right] ^ {2}. \tag {3} +$$ + +Since $q(\mathbf{x}_T|\cdot)$ is equivalent to an isotropic Gaussian distribution as $T \to +\infty$ , the estimation quantity $\log p_{\theta}(\mathbf{x}_T)$ converges to a constant when $\theta \to \theta^*$ . Therefore, we can ignore the first term of Equation 2. + +Finally, the term $\hat{q} (\mathbf{x}_0|\mathbf{x}_T,y)$ of Equation 2 provides more information about the relation between $\mathbf{x}_T$ and $\mathbf{x}_0$ . As this term is intractable [65], we parameterize $\hat{q} (\mathbf{x}_0|\mathbf{x}_T,y)$ as $p_{\psi ,y}(\mathbf{x}_0,\mathbf{x}_T)$ . This estimation resembles the noise-contrastive estimation [27], where the $\mathbf{x}_T$ , $\mathbf{x}_0$ can be considered as a pair of contrastive estimation and $\psi$ can be estimated by a contrastive loss. + +In [19, 68, 73], the authors indicate that predicting $\mathbf{x}_0$ is often infeasible but predicting an estimation $\tilde{\mathbf{x}}_0$ is tractable and can be used as a 'pseudo' estimation of $\mathbf{x}_0$ . We denote $\tilde{\mathbf{x}}_0$ as an estimation of $\mathbf{x}_0$ . The loss term $\hat{q}(\mathbf{x}_0 | \mathbf{x}_T, y)$ can be approximated by using the following contrastive loss + +$$ +\mathcal {L} _ {\mathrm {c o n t r a s t i v e}} = \max \left(0, \left\| \frac {\sqrt {\overline {{\alpha_ {T}}}} \tilde {\mathbf {x}} _ {0} - \mathbf {x} _ {T}}{\sqrt {1 - \overline {{\alpha_ {T}}}}} \right\| _ {2} ^ {2} - M\right), \tag {4} +$$ + +where $M$ is the number of dimension of $\mathbf{x}_0$ , and $\alpha_{t}$ is the variance schedule at timestep $t$ ( $t = \overline{1;T}$ ). + +Proposition 1. Suppose that $\tilde{\mathbf{x}}_0$ , $\mathbf{x}_0$ and $\epsilon$ are independent, and that + +$$ +\left\| \frac {\sqrt {\overline {{\alpha_ {T}}}} \tilde {\mathbf {x}} _ {0} - \mathbf {x} _ {T}}{\sqrt {1 - \overline {{\alpha_ {T}}}}} \right\| _ {2} ^ {2} \geq M. +$$ + +Then there exists $C > 0$ such that: for arbitrary $\delta > 0$ , if $\mathcal{L}_{\mathrm{contrastive}} < \delta$ , then + +$$ +\mathbb {E} \left[ \| \tilde {\mathbf {x}} _ {0} - \mathbf {x} _ {0} \| _ {2} ^ {2} \right] < C \delta . +$$ + +Proof. See Supplementary Material. + +![](images/2eef1af9873495c5885bb62b027cc1295ddaef5beef475a9a959387855b8c009.jpg) + +Remark 1.1. Proposition 1 suggests that if the contrastive loss $\mathcal{L}_{\mathrm{contrastive}}$ tends to zero, then the prediction $\tilde{\mathbf{x}}_0$ will approach the ground truth $\mathbf{x}_0$ . + +Remark 1.2. The condition $\left\| \frac{\sqrt{\overline{\alpha_T}}\tilde{\mathbf{x}}_0 - \mathbf{x}_T}{\sqrt{1 - \overline{\alpha_T}}}\right\|_2^2 \geq M$ is suitable for our language-driven grasp detection task as $\tilde{\mathbf{x}}_0$ and $\mathbf{x}_T$ are two contrastive quantities, therefore, we can assume there is a minimum distance between $\tilde{\mathbf{x}}_0$ and $\mathbf{x}_T$ . In addition, in the proof of Proposition 1, we see that $\mathbb{E}\left[\left\| \frac{\sqrt{\overline{\alpha_T}}\tilde{\mathbf{x}}_0 - \mathbf{x}_T}{\sqrt{1 - \overline{\alpha_T}}}\right\|_2^2 - M\right] = \beta^2\mathbb{E}\left[\| \tilde{\mathbf{x}}_0 - \mathbf{x}_0\|_2^2\right]$ , which is always nonnegative. Therefore, it is both theoretically and experimentally reasonable to add this assumption. + +![](images/c5f8cf0da96fc9cb60067e238944d2fab949583d467d2ea986898423b76b7ced.jpg) +Figure 4. Language-drive Grasp Detection (LGD) network. We present the network architecture (left) and the proposed training objectives of the denoising process (right). + +# 4.2. Language-driven Grasp Detection Network + +Network. Our network operates on two conditions: an image denoted as $\pmb{I}$ and a corresponding text prompt represented as $e$ . To process these conditions, we employ a vision encoder to extract visual features from $\pmb{I}$ and a text encoder to derive textual embeddings from $e$ . The resulting feature vectors, denoted as $I'$ and $e'$ , are subsequently subjected to a fusion module, ALBEF [39]. We leverage the attention mask generated by the ALBEF fusion module + +![](images/31af37597999a812ea03d30352742866a3875271953175cf6a4ffff0652e4476.jpg) + +![](images/853a17736d910c97ba413f075bf351c062d98c253545ebe034163a21153f2cc6.jpg) + +![](images/5ef3066826567afd9937574af595edf403a795ffbf49d6fe73a2884db31ab644.jpg) +Give me the bottle + +![](images/d040c5158a765ad1a6044d085d67f7decc812517f74e1524a65a4f6f74cba34f.jpg) + +![](images/c175ad61339879b6e2afa8d1769ec697bffea22e87d70e9611a8752a03d954b0.jpg) + +![](images/14bec2ca7857c1b62b3ceaea57a18780bc18796d6d727500f5815f31d4d1799b.jpg) + +![](images/d0b133780a36ae1595fe6d8c5715704ebd2fbdd2a08ca50223b25be31d333f6e.jpg) +(a) Ours + +![](images/989181ff313a07de2bfb6e230168a17bc2ccab2c0b3677013935364ead3ea5cb.jpg) +(b) GR-ConvNet +Figure 5. Language-driven grasp detection results visualization. + +![](images/6e2945012a302f4787aa6e85ec085254518d9e7de32d80a98e9c1f9c0a09dbc6.jpg) +Grasp the wristwatch at its dial +(c) Det-Seg-Refine + +![](images/2f21faf38cd302e13e459d2bece98cbb230586d4b6bb03c29d368ec85a62c9a3.jpg) +(d) GG-CNN + +![](images/60222a5d53b79fdec6ca8b5bcbbefffec84e1853bb0ecfe3c95933e1258ff630.jpg) +(e) CLIPORT + +![](images/468e895b375c80b4d2740a82d90feffd28d1d99d02bcbc9449a60dd580697f48.jpg) +(f) CLIP-Fusion + +as the estimation $\tilde{\mathbf{x}}_0$ of $\mathbf{x}_0$ . Next, we aggregate three elements: the estimation region $\tilde{\mathbf{x}}_0$ , the grasp pose at the current timestep $\mathbf{x}_{t + 1}$ , and the timestep $t + 1$ . These inputs are combined using MLP layers, similar to the approach outlined in [68]. Specifically, the output operation can be expressed as: $\mathbf{x}_t = f(\mathbf{x}_{t + 1}, t + 1, \tilde{\mathbf{x}}_0)$ , where function $f$ encompasses a composition of multiple MLP layers. Additional specifics regarding these universal MLP layers are provided in the Supplementary Material. + +Training Objective. In our context, conditioned grasp detection models the distribution $p(\mathbf{x}_0|y)$ as the reversed diffusion process of gradually cleaning $\mathbf{x}_{t + 1}$ . Instead of predicting $\mathbf{x}_t$ as formulated by [28], we follow Ramesh et al. [58] and predict the signal itself, i.e., $\mathbf{x}_t = f(\mathbf{x}_{t + 1}, t + 1, \tilde{\mathbf{x}}_0)$ with the simple objective [28]. To this end, we utilize the contrastive loss as in Equation 4 to explicitly improve the learning objective of the denoising process: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {c o n t r a s t i v e}} + \mathcal {L} _ {\text {d i f f u s i o n}}. \tag {5} +$$ + +# 5. Experiments + +We conduct experiments to evaluate our proposed method and Grasp-Anything++ dataset using both the vision-based metrics and real robot experiments. We then demonstrate zero-shot grasp results and discuss the challenges and open questions for future works. + +# 5.1. Language-driven Grasp Detection Results + +Baselines. We compare our language-driven grasp detection method (LGD) with the linguistically supported versions of GR-CNN [35], Det-Seg-Refine [2], GG-CNN [47], CLIPORT [64] and CLIP-Fusion [80]. In all cases, we employ a pretrained CLIP [57] or BERT [18] as the text embedding. The implementation details of all baselines can be found in our Supplementary Material. + +
BaselineSeenUnseenH
GR-ConvNet [35] + CLIP [57]0.370.180.24
Det-Seg-Refine [2] + CLIP [57]0.300.150.20
GG-CNN [47] + CLIP [57]0.120.080.10
CLIPORT [64]0.360.260.29
CLIP-Fusion [80]0.400.290.33
LGD (ours) + BERT [18]0.440.380.41
LGD (ours) + CLIP [57]0.480.420.45
+ +Setup. To assess the generalization of all methods trained on Grasp-Anything++, we utilize the concept of base and new labels [92] in zero-shot learning. We categorize LVIS labels from Section 3.3 to form labels for our experiment. In particular, we select $70\%$ of these labels by frequency for 'Base' and assign the remaining $30\%$ to 'New'. We also use the harmonic mean ('H') to measure the overall success rates [92]. Our primary evaluation metric is the success rate, defined similarly to [35], necessitating an IoU score of the predicted grasp exceeding $25\%$ with the ground truth grasp and an offset angle less than $30^{\circ}$ . + +Main Results. Table 2 shows the results of language-driven grasp detection on the Grasp-Anything++ dataset. The findings indicate a notable performance advantage of our LGD over other baseline approaches, with LGD outperforming the subsequent best-performing baselines (CLIP-Fusion) by margins of 0.14 on Grasp-Anything++ dataset. + +Table 2. Language-driven grasp detection results. + +
BaselineSeenUnseenH
LGD w/o predicting x̂00.150.080.10
LGD w/o contrastive loss0.450.400.42
LGD w contrastive loss0.480.420.45
+ +Table 3. Contrastive loss analysis. + +![](images/1c307034979a56be59006f6bf22980e82842e7b9f5c980ae459ba3e08fd0bf9c.jpg) +Figure 6. Loss visualization. + +![](images/8c68367085b7a413e767eb39182d92b6b2fad4e5e1a8b4884bb40ecec9166f4c.jpg) +Figure 7. t-SNE visualization. We apply t-SNE to cluster the vision-and-language features with and without contrastive loss. + +![](images/34186e8bce2d7cb6da9de2e05cece17a842747e303536e459d5f00a33de8d4db.jpg) +Figure 8. Attention map visualization. We compare the attention map when utilizing our proposed contrastive loss and when not. + +Contrastive Loss Analysis. Table 3 presents the performance of LGD under varied configurations. The outcomes emphasize the substantial influence of the training objective (contrastive loss) and the importance of language instructions in enhancing LGD performance on both seen and unseen classes in the Grasp-Anything++ dataset. + +Fig. 6 shows the contrastive loss $\mathcal{L}_{\mathrm{contrastive}}$ approaching towards 0 during training, indicating the grasp pose estimation $\tilde{\mathbf{x}}_0$ aligns with the ground truth $\mathbf{x}_0$ , as anticipated by Proposition 1. The subsequent attention maps visualization in Fig. 8 shows the attention region is meaningful and improves the results when employing our proposed contrastive loss compared to its absence. Moreover, we employ t-SNE for vision-and-language embedding visualization, as in [46], by processing 2,000 samples from the Grasp-Anything++ dataset through the ALBEF module. The outcomes reveal that our contrastive loss facilitates better object classification, as evidenced in Fig.7 by clearer segregation of pixel embeddings across various semantic classes, underscoring contrastive loss's role in refining embeddings' differentiation for improved class distinctions. + +Qualitative Results. Fig. 5 presents qualitative results of the language-driven grasp detection task, suggesting that our LGD method generates more semantically plausible than other baselines. Despite satisfactory performance, LGD occasionally predicts incorrect results, with a detailed analysis of these cases available in our Appendix. + +Robotic Validation. We provide quantitative results by integrating our language-driven grasp detection pipeline for a robotic grasping application with a KUKA LBR iwa + +![](images/0dcadfe26b1b66087528e78d00518a1f905d590b7151ce2af416a6b221d0d2b4.jpg) + +
BaselineSingleCluttered
GR-ConvNet [35] + CLIP [57]0.330.30
Det-Seg-Refine [2] + CLIP [57]0.300.23
GG-CNN [47] + CLIP [57]0.100.07
CLIPORT [64]0.270.30
CLIP-Fusion [80]0.400.40
LGD (ours)0.430.42
+ +Table 4. Robotic language-driven grasp detection results. + +R820 robot. Using the RealSense D435i camera, the grasp pose inferred from approaches in Table 4 is transformed into the 6DoF grasp pose, similar to [35]. The optimization-based trajectory planner in [5,72] is employed to execute the grasps. Experiments are conducted for two scenarios, i.e., the single object scenario and the cluttered scene scenario, of a set of 20 real-world daily objects. In each scenario, we run 30 experiments using baselines listed in Table 4 and a predefined grasping prompt corpus. The results exhibit that our LGD outperforms other baselines. Furthermore, although LGD is trained on our Grasp-Anything++ which is a solely synthesis dataset created by foundation models, it still shows reasonable results on real-world objects. + +# 5.2. Zero-shot Grasp Detection + +Our proposed Grasp-Anything++ is a large-scale dataset. Apart from the language-driven grasp detection task, we believe it can be used for other purposes. In this experiment, we seek to answer the question: Can Grasp-Anything++ be useful in the traditional grasp detection task without text? Consequently, we verify our Grasp-Anything++ and LGD (no text) with other existing datasets and grasping methods. + +Setup. We setup an LGD (no text) version, and other state-of-the-art grasp detection methods GR-ConvNet [35], Det-Seg-Refine [2], GG-CNN [47]. We use five datasets: our Grasp-Anything++, Jacquard [17], Cornell [33], VMRD [87], and OCID-grasp [2] in this experiment. + +
BaselineGrasp-Anything++ (ours)Jacquard [17]Cornell [33]VMRD [87]OCID-grasp [2]
BaseNewHBaseNewHBaseNewHBaseNewHBaseNewH
GR-ConvNet [35]0.710.590.640.880.660.750.980.740.840.770.640.700.860.670.75
Det-Seg-Refine [2]0.620.570.590.860.600.710.990.760.860.750.600.660.800.620.70
GG-CNN [47]0.680.570.620.780.560.650.960.750.840.690.530.590.710.630.67
LGD (no text) (ours)0.740.630.680.890.690.770.970.760.850.790.660.720.880.680.76
+ +Zero-shot Results. Table 5 summarizes the base-to-new grasp detection results on five datasets. Overall, the performance of LGD even without the language branch is better than other baselines across all datasets. Furthermore, this table also shows that our Grasp-Anything++ dataset is more challenging to train as the detection results are lower than related datasets using the same approaches due to the greater coverage of unseen objects in the testing phase. + +Table 5. Base-to-new zero-shot grasp detection results. + +
Test TrainJacquardCornellVMRDOCID-graspGrasp-Anything++
Jacquard [17]0.870.510.130.210.17
Cornell [33]0.070.980.200.120.13
VMRD [87]0.060.210.790.110.10
OCID-grasp [2]0.090.120.200.740.11
Grasp-Anything++ (ours)0.410.630.300.390.65
+ +Table 6. Cross-dataset grasp detection results. + +Cross-dataset Evaluation. To further verify the usefulness of our Grasp-Anything++ dataset, we conduct the cross-dataset validation in Table 6. We use the GRConvNet [35] to reuse its results on existing grasp datasets. GR-ConvNet is trained on a dataset (row) and evaluated on another dataset (column). For example, training on Jacquard and testing on Cornell yields an accuracy of 0.51. Notably, training with our dataset improves performance by approximately $10 - 33\%$ compared to other datasets. + +In the wild grasp detection. Fig. 9 shows visualization results using LGD (no text) trained on our Grasp-Anything++ dataset on random internet images and other datasets images. We can see that the detected grasp poses are adequate in quality and quantity. This demonstrates that although our Grasp-Anything++ is fully created by foundation models without having any real images, models trained on our Grasp-Anything++ dataset still generalize well on real-world images. + +# 5.3. Discussion + +Our experiments indicate that Grasp-Anything++ can serve as a foundation dataset for both language-driven and traditional grasp detection tasks. However, there are certain limitations. First, our dataset lacks depth images for directly being applied to robotic applications [52]. Second, we remark that the creation of our dataset is time-consuming and relies on access to the ChatGPT API. Fortunately, future research can reuse our provided assets (images, prompts, + +![](images/d45184855b9574769a59b2cc5d6a5a523aadcc2d15ecca86d9317af75a06c019.jpg) + +![](images/a10c88bfdc345c472ae706e4ff528b3568882724b860abf6dcfae55a39089ad7.jpg) + +![](images/74ab2e0f15cced209ade5f6546ff3d4275012a2e365f1bb48382ea7f66ef018d.jpg) + +![](images/cb740b29edd411199dd92af3a1a8f467e814b76806cffefa8bf4bc433575e067.jpg) +Figure 9. In the wild grasp detection. Top row images are from GraspNet [24], YCB-Video [78], NBMOD [9] datasets; bottom row shows internet images. + +![](images/dab9c132ea433c097788cc4bda274d6644c4c48ab2609b7975562fb7471f7f57.jpg) + +![](images/dd68889eeac9faeb8996a249d69ba68fa871e77f0b86937ff65f3e1ca7351ec5.jpg) + +etc.) without starting from scratch. Furthermore, our experiments show that adding language to the grasp detection task (Table 2) poses a more challenging problem compared to standard grasp detection task (Table 5). + +We see several interesting future research directions. First, future work could investigate the use of text or imaged-to-3D models [79] or image-to-depth [59] and reuse our dataset's prompts and images to construct 3D language-driven grasp datasets. Additionally, beyond linguistic grasp instruction adherence, our dataset holds potential for varied applications, including scene understanding [81] and scene generation [3], hallucination analysis [30], and humanrobot interaction [11]. + +# 6. Conclusion + +We introduce Grasp-Anything++, a large-scale dataset with 1M images and 10M grasp prompts for language-driven grasp detection tasks. We propose LGD, a diffusion-based method to tackle the language-driven grasp detection task. Our diffusion model employs a contrastive training objective, which explicitly contributes to the denoising process. Empirically, we have shown that Grasp-Anything++ serves as a foundation grasp detection dataset. Finally, our LGD improves the performance of other baselines, and the real-world robotic experiments further validate the effectiveness of our dataset and approach. + +# References + +[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2 +[2] Stefan Ainetter and Friedrich Fraundorfer. End-to-end trainable deep neural network for robotic grasp detection and semantic segmentation from rgb. In ICRA, 2021. 2, 4, 6, 7, 8 +[3] Dor Arad Hudson and Larry Zitnick. Compositional transformers for scene generation. NeurIPS, 2021. 8 +[4] Umar Asif, Jianbin Tang, and Stefan Harrer. Grasynet: An efficient convolutional neural network for real-time grasp detection for low-powered devices. In IJCAI, 2018. 4 +[5] Florian Beck, Minh Nhat Vu, Christian Hartl-Nesic, and Andreas Kugi. Singularity avoidance with application to online trajectory optimization for serial manipulators. IFAC-PapersOnLine, 56(2):284-291, 2023. 7 +[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33, 2020. 2 +[8] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In CVPR, 2021. 4 +[9] Boyuan Cao, Xinyu Zhou, Congmin Guo, Baohua Zhang, Yuchen Liu, and Qianqiu Tan. Nbmod: Find it and grasp it in noisy background. arXiv preprint arXiv:2306.10265, 2023. 2, 3, 8 +[10] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 2 +[11] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, 2022. 8 +[12] I-Ming Chen and Joel W Burdick. Finding antipodal point grasps on irregularly shaped objects. T-RA, 1993. 4 +[13] Sijia Chen and Baochun Li. Language-guided diffusion model for visual grounding. arXiv preprint arXiv:2308.09599, 2023. 2 +[14] Hongjun Choi, Anirudh Som, and Pavan Turaga. Amc-loss: Angular margin contrastive loss for improved explainability in image classification. In CVPR workshops, 2020. 4 +[15] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. TPAMI, 2023. 4 +[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana + +Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 4 +[17] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IROS, 2018. 2, 3, 4, 7, 8 +[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6 +[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34, 2021. 5 +[20] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 2 +[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1 +[22] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. A billion ways to grasp: An evaluation of grasp sampling schemes on a dense, physics-based grasp data set. In ISRR, 2019. 2 +[23] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In ICRA, 2021. 2 +[24] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. 2, 8 +[25] Maximilian Gilles, Yuhao Chen, Tim Robin Winter, E Zhixuan Zeng, and Alexander Wong. Metagraspnet: A large-scale benchmark dataset for scene-aware ambidextrous bin picking via physics-based metaverse synthesis. In CASE, 2022. 1 +[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4 +[27] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTAT, 2010. 5 +[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 4, 5, 6 +[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 4 +[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232, 2023. 4, 8 +[31] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, 2022. 2 +[32] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 1 + +[33] Yun Jiang, Stephen Moseson, and Ashutosh Saxena. Efficient grasping from rgbd images: Learning using a new rectangle representation. In ICRA, 2011. 2, 3, 4, 7, 8 +[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3 +[35] Sulabh Kumra, Shirin Joshi, and Ferat Sahin. Antipodal robotic grasping using generative residual convolutional neural network. In IROS, 2020. 6, 7, 8 +[36] Jaewoong Lee, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, Yunji Kim, Jin-Hwa Kim, Jung-Woo Ha, and Sung Ju Hwang. Text-conditioned sampling framework for text-to-image generation with masked generative models. arXiv preprint arXiv:2304.01515, 2023. 2 +[37] Meng-Lun Lee, Sara Behdad, Xiao Liang, and Minghui Zheng. Task allocation and planning for product disassembly with human-robot collaboration. Robotics and Computer-Integrated Manufacturing, 76, 2022. 2 +[38] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. *IJRR*, 2018. 2 +[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 5 +[40] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2 +[41] Jirong Liu, Ruo Zhang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Chenxi Wang, Sheng Xu, Hengxu Yan, and Cewu Lu. Target-referenced reactive grasping for dynamic objects. In CVPR, 2023. 2 +[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2 +[43] Weiyu Liu, Tucker Hermans, Sonia Chernova, and Chris Paxton. Structdiffusion: Object-centric diffusion for semantic rearrangement of novel objects. arXiv preprint arXiv:2211.04604, 2022. 2, 4 +[44] Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-Radiology, page 100017, 2023. 4 +[45] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312, 2017. 2 +[46] Jiaxu Miao, Zongxin Yang, Leilei Fan, and Yi Yang. Fedseg: Class-heterogeneous federated learning for semantic segmentation. In CVPR, 2023. 7 + +[47] Douglas Morrison, Peter Corke, and Jürgen Leitner. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. arXiv preprint arXiv:1804.05172, 2018. 6, 7, 8 +[48] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. RA-L, 2020. 2 +[49] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In ICCV, 2019. 2 +[50] Yao Mu, Shunyu Yao, Mingyu Ding, Ping Luo, and Chuang Gan. Ec2: Emergent communication for embodied control. In CVPR, 2023. 2 +[51] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 1 +[52] Rhys Newbury, Morris Gu, Lachlan Chumbley, Arsalan Mousavian, Clemens Eppner, Jürgen Leitner, Jeannette Bohg, Antonio Morales, Tamim Asfour, Danica Kragic, et al. Deep learning approaches to grasp synthesis: A review. T-RO, 2023. 8 +[53] Toan Nguyen, Minh Nhat Vu, An Vuong, Dzung Nguyen, Thieu Vo, Ngan Le, and Anh Nguyen. Open-vocabulary affordance detection in 3d point clouds. In IROS, 2023. 3 +[54] OpenAI. Introducing ChatGPT. Software. Accessed: July 6th 2023. 2 +[55] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. 2 +[56] Robert Platt. Grasp learning: Models, methods, and performance. Annual Review of Control, Robotics, and Autonomous Systems, 2023. 1 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7 +[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6 +[59] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In CVP4, 2021. 8 +[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4 +[61] Kallol Saha, Vishal Mandadi, Jayaram Reddy, Ajit Srikanth, Aditya Agarwal, Bipasha Sen, Arun Singh, and Madhava Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. arXiv preprint arXiv:2309.11414, 2023. 2 +[62] Schuhmann et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 4 +[63] Dhruv Shah, Błajej Osiński, brian richter, and Sergey Levine. Lm-nav: Robotic navigation with large pre-trained models of + +language, vision, and action. In Karen Liu, Dana Kulic, and Jeff Ichnowski, editors, Proceedings of The 6th Conference on Robot Learning, volume 205 of PMLR. PMLR, 2023. 1 +[64] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1, 2, 6, 7 +[65] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 5 +[66] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023. 1 +[67] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3 +[68] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2022. 5, 6 +[69] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 5 +[70] Julien Urain, Niklas Funk, Jan Peters, and Georgia Chalvatzaki. Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In ICRA, 2023. 2 +[71] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2023. 2 +[72] Minh Nhat Vu, Florian Beck, Michael Schwegel, Christian Hartl-Nesic, Anh Nguyen, and Andreas Kugi. Machine learning-based framework for optimally solving the analytical inverse kinematics for redundant manipulators. Mechatronics, 2023. 7 +[73] An Vuong, Minh Nhat Vu, Toan Tien Nguyen, Baoru Huang, Dzung Nguyen, Thieu Vo, and Anh Nguyen. Language-driven scene synthesis using multi-conditional diffusion model. arXiv preprint arXiv:2310.15948, 2023. 5 +[74] An Dinh Vuong, Minh Nhat Vu, Hieu Le, Baoru Huang, Binh Huynh, Thieu Vo, Andreas Kugi, and Anh Nguyen. Grasp-anything: Large-scale grasp dataset from foundation models. arXiv preprint arXiv:2309.09818, 2023. 2, 3 +[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 3 +[76] Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania, and Ernesto W De Luca. Legal norm retrieval with variations of the bert model combined with tfidf vectorization. In Proceedings of the eighteenth international conference on artificial intelligence and law, pages 285-294, 2021. 4 +[77] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In MIDL, 2022. 2 + +[78] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 2, 8 +[79] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 8 +[80] Kechun Xu, Shuqi Zhao, Zhongxiang Zhou, Zizhang Li, Huajin Pi, Yifeng Zhu, Yue Wang, and Rong Xiong. A joint modeling of vision-language-action for target-oriented grasping in clutter. arXiv preprint arXiv:2302.12610, 2023. 2, 6, 7 +[81] Yiteng Xu, Peishan Cong, Yichen Yao, Runnan Chen, Yuenan Hou, Xinge Zhu, Xuming He, Jingyi Yu, and Yuexin Ma. Human-centric scene understanding for 3d large-scale scenarios. In ICCV, 2023. 8 +[82] Zhixuan Xu, Kechun Xu, Rong Xiong, and Yue Wang. Object-centric inference for language conditioned placement: A foundation model based approach. In ICARM, 2023. 2 +[83] Jange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pick-place robots. arXiv preprint arXiv:2306.05716, 2023. 1, 2 +[84] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 2022, 2, 4 +[85] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. ICCV, 2023. 4 +[86] Yang Yang, Xibai Lou, and Changyun Choi. Interactive robotic grasping with attribute-guided disambiguation. In ICRA, 2022. 2 +[87] Hanbo Zhang, Xuguang Lan, Site Bai, Xinwen Zhou, Zhiqiang Tian, and Nanning Zheng. Roi-based robotic grasp detection for object overlapping scenes. In IROS, 2019. 4, 7, 8 +[88] Hanbo Zhang, Yunfan Lu, Cunjun Yu, David Hsu, Xuguang La, and Nanning Zheng. Invigorate: Interactive visual grounding and grasping in clutter. arXiv preprint arXiv:2108.11092, 2021. 2 +[89] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 4 +[90] Kaizhi Zheng, Xiaotong Chen, Odest Chadwicke Jenkins, and Xin Wang. Vlmbench: A compositional benchmark for vision-and-language manipulation. NeurIPS, 35, 2022. 2 +[91] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 4 +[92] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 6 \ No newline at end of file diff --git a/2024/Language-driven Grasp Detection/images.zip b/2024/Language-driven Grasp Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5760cfba484a299d2bc3c1f495448aff35f18018 --- /dev/null +++ b/2024/Language-driven Grasp Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a337a573e06a50c7d0ccf6f3d68dc2b23a91e0fb4e31f442b27aaf7c143a74b1 +size 897876 diff --git a/2024/Language-driven Grasp Detection/layout.json b/2024/Language-driven Grasp Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cab5b9255bfe18de67bd97838c4b05c48497d375 --- /dev/null +++ b/2024/Language-driven Grasp Detection/layout.json @@ -0,0 +1,11726 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "spans": [ + { + "bbox": [ + 190, + 103, + 403, + 121 + ], + "type": "text", + "content": "Language-driven Grasp Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": "An Dinh Vuong" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Minh Nhat Vu" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Baoru Huang" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{3,*}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Nghia Nguyen" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Hieu Le" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Thieu Vo" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "text", + "content": ", Anh Nguyen" + }, + { + "bbox": [ + 53, + 142, + 563, + 159 + ], + "type": "inline_equation", + "content": "^{5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "text", + "content": "FPT Software AI Center, Vietnam " + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "text", + "content": "Automation & Control Institute, TU Wien, Austria " + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 159, + 524, + 171 + ], + "type": "text", + "content": "Imperial College London, UK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "spans": [ + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "type": "text", + "content": "Ton Duc Thang University, Vietnam " + }, + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 173, + 482, + 185 + ], + "type": "text", + "content": "University of Liverpool, UK *Co-Corresponding authors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 236, + 187, + 379, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 187, + 379, + 199 + ], + "spans": [ + { + "bbox": [ + 236, + 187, + 379, + 199 + ], + "type": "text", + "content": "https://airvlab.github.io/grasp-anything" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 53, + 245, + 150, + 350 + ], + "blocks": [ + { + "bbox": [ + 53, + 217, + 149, + 243 + ], + "lines": [ + { + "bbox": [ + 53, + 217, + 149, + 243 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 149, + 243 + ], + "type": "text", + "content": "A coffee cup, a grape biscuit, a black oven plate on a white dining table" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 245, + 150, + 350 + ], + "lines": [ + { + "bbox": [ + 53, + 245, + 150, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 245, + 150, + 350 + ], + "type": "image", + "image_path": "7e25ba918d52692bb9f2a2f4cef694e7122226554dfece7ff4a35af51422c547.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 352, + 137, + 361 + ], + "lines": [ + { + "bbox": [ + 67, + 352, + 137, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 352, + 137, + 361 + ], + "type": "text", + "content": "Grasp the coffee cup." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 151, + 245, + 248, + 350 + ], + "blocks": [ + { + "bbox": [ + 154, + 217, + 245, + 243 + ], + "lines": [ + { + "bbox": [ + 154, + 217, + 245, + 243 + ], + "spans": [ + { + "bbox": [ + 154, + 217, + 245, + 243 + ], + "type": "text", + "content": "A black pen, a steel marble and a digital calculator on an office desk" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 245, + 248, + 350 + ], + "lines": [ + { + "bbox": [ + 151, + 245, + 248, + 350 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 248, + 350 + ], + "type": "image", + "image_path": "7cfffa3db9198cd347875d81b81bd2d3792ab6bd0989dd95035c7de9dfa5b2a5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 352, + 232, + 370 + ], + "lines": [ + { + "bbox": [ + 165, + 352, + 232, + 370 + ], + "spans": [ + { + "bbox": [ + 165, + 352, + 232, + 370 + ], + "type": "text", + "content": "Grasp the calculator at its keypad." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 151, + 377, + 440, + 389 + ], + "lines": [ + { + "bbox": [ + 151, + 377, + 440, + 389 + ], + "spans": [ + { + "bbox": [ + 151, + 377, + 440, + 389 + ], + "type": "text", + "content": "Figure 1. We present a new dataset and method for language-driven grasp task." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 250, + 245, + 345, + 350 + ], + "blocks": [ + { + "bbox": [ + 250, + 217, + 345, + 243 + ], + "lines": [ + { + "bbox": [ + 250, + 217, + 345, + 243 + ], + "spans": [ + { + "bbox": [ + 250, + 217, + 345, + 243 + ], + "type": "text", + "content": "A black pot, a green bottle, a bread container arranged on a kitchen counter" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 250, + 245, + 345, + 350 + ], + "lines": [ + { + "bbox": [ + 250, + 245, + 345, + 350 + ], + "spans": [ + { + "bbox": [ + 250, + 245, + 345, + 350 + ], + "type": "image", + "image_path": "f1f93e7891e5c35da4bbcb99c8af49cc675eee568a3a9ca6cd60f62d9ea6d614.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 352, + 328, + 370 + ], + "lines": [ + { + "bbox": [ + 266, + 352, + 328, + 370 + ], + "spans": [ + { + "bbox": [ + 266, + 352, + 328, + 370 + ], + "type": "text", + "content": "Grab the neck of the green bottle." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 347, + 245, + 443, + 350 + ], + "blocks": [ + { + "bbox": [ + 347, + 217, + 443, + 243 + ], + "lines": [ + { + "bbox": [ + 347, + 217, + 443, + 243 + ], + "spans": [ + { + "bbox": [ + 347, + 217, + 443, + 243 + ], + "type": "text", + "content": "A brown-band wristwatch, a sleek black pen and an office clip placed on a table" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 245, + 443, + 350 + ], + "lines": [ + { + "bbox": [ + 347, + 245, + 443, + 350 + ], + "spans": [ + { + "bbox": [ + 347, + 245, + 443, + 350 + ], + "type": "image", + "image_path": "9bafff9a75f75acdec949e96af0d9469c76a55c459bdcda095c0afb8aa904730.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 352, + 436, + 369 + ], + "lines": [ + { + "bbox": [ + 355, + 352, + 436, + 369 + ], + "spans": [ + { + "bbox": [ + 355, + 352, + 436, + 369 + ], + "type": "text", + "content": "Give me the brown-band wristwatch." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 445, + 217, + 542, + 350 + ], + "blocks": [ + { + "bbox": [ + 445, + 217, + 542, + 350 + ], + "lines": [ + { + "bbox": [ + 445, + 217, + 542, + 350 + ], + "spans": [ + { + "bbox": [ + 445, + 217, + 542, + 350 + ], + "type": "image", + "image_path": "a3b83d90a97a9e2785ee86fd3954c919f23f96693945dc7bc25ba5b86f32d7e3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 352, + 521, + 370 + ], + "lines": [ + { + "bbox": [ + 466, + 352, + 521, + 370 + ], + "spans": [ + { + "bbox": [ + 466, + 352, + 521, + 370 + ], + "type": "text", + "content": "Pick the fountain pen at its cap." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 143, + 399, + 192, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 399, + 192, + 410 + ], + "spans": [ + { + "bbox": [ + 143, + 399, + 192, + 410 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 413, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 289, + 700 + ], + "type": "text", + "content": "Grasp detection is a persistent and intricate challenge with various industrial applications. Recently, many methods and datasets have been proposed to tackle the grasp detection problem. However, most of them do not consider using natural language as a condition to detect the grasp poses. In this paper, we introduce Grasp-Anything++, a new language-driven grasp detection dataset featuring 1M samples, over 3M objects, and upwards of 10M grasping instructions. We utilize foundation models to create a large-scale scene corpus with corresponding images and grasp prompts. We approach the language-driven grasp detection task as a conditional generation problem. Drawing on the success of diffusion models in generative tasks and given that language plays a vital role in this task, we propose a new language-driven grasp detection method based on diffusion models. Our key contribution is the contrastive training objective, which explicitly contributes to the denoising process to detect the grasp pose given the language instructions. We illustrate that our approach is theoretically supportive. The intensive experiments show that our method outperforms state-of-the-art approaches and allows real-world robotic grasping. Finally, we demonstrate our large-scale dataset enables zero-short grasp detection and is a challenging benchmark for future work." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 386, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 386, + 411 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 386, + 411 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 638 + ], + "type": "text", + "content": "Imagine we want an assistant robot to grasp a cup among a clutter of daily objects such as a knife, a fork, a cup, and a pair of scissors. Conventionally, to convey the idea of grasping this specific object, humans use the natural language command, \"give me the cup\", for instance. Although humans intuitively know how to grasp the cup given the linguistic command, determining specific grasp actions for objects based on natural language instructions or language-driven grasp detection remains challenging for robots [64]. First, natural language is usually overlooked in existing grasp datasets [56] while training vision-and-language neural networks necessitates an excessive number of labeled examples [66]. Second, recent works usually focus on particular manipulation tasks with limited objects [25], imposing a bottleneck for in-the-wild robot execution [63]. Finally, despite recent developments, bridging the gap between language, vision, and control for real-world robotic experiments remains a challenging task [83]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Recently, language-driven robotic frameworks are gaining traction, offering the potential for robots to process natural language, and bridging the gap between robotic manipulations and real-world human-robot interaction [51]. PaLM-E [21], EgoCOT [51], and ConceptFusion [32] are some notable embodied robots with the ability to compre" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17902" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "hend natural language by harnessing the power of large foundation models such as ChatGPT [54]. However, most works assume the high-level actions of robots and ignore the fundamental grasping actions, restricting the structure for generalization across robotic domains, tasks, and skills [50]. In this paper, we explore training a language-driven agent to implement low-level actions, focusing on the task of object grasping via image observations. Specifically, our hypothesis is centered around the establishment of a robotic system that can execute grasping actions following a given language instruction for any universal object." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 289, + 396 + ], + "type": "text", + "content": "We first present Grasp-Anything++ to serve as a large-scale dataset for language-driven grasp detection. Our dataset is based on the Grasp-Anything [74], and is synthesized from foundation models. Compared to the original Grasp-Anything dataset, we provide more than 10M grasp prompts and 3M associated object masks, 6M ground truth poses at the object part level. Our dataset showcases the ability to facilitate grasp detection using language instructions. We label the ground truth at both the object level and part level, providing a comprehensive understanding of real-world scenarios. For example, our ground truth includes both general instructions \"give me the knife\" and detail ones such as \"grasp the handle of the steak knife\". We empirically show that our large-scale dataset successfully facilitates zero-shot grasp detection on both vision-based tasks and real-world robotic experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 289, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 598 + ], + "type": "text", + "content": "To tackle the challenging language-driven grasp detection task, we propose a new diffusion model-based method. Our selection of diffusion models is motivated by their proven efficacy in conditional generation tasks [28]. These models have shown efficiency beyond image synthesis, including other image-based tasks such as image segmentation [77], and visual grounding [40]. Despite achieving notable success, integrating visual and text features effectively remains a challenge [13] as the majority of existing literature employs latent strategies to combine visual and text features [36]. We address this challenge by employing a new training strategy for learning text and image features, focusing on the use of feature maps as guidance information for grasp pose generation. Our main contribution is a new training objective that incorporates the feature maps and explicitly contributes to the denoising process. In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 605, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 628 + ], + "type": "text", + "content": "- We propose Grasp-Anything++, a large-scale language-driven dataset for grasp detection tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 287, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 672 + ], + "type": "text", + "content": "- We propose a diffusion model with a training objective that explicitly contributes to the denoising process to detect the grasp poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "- We demonstrate that our Grasp-Anything++ dataset and the proposed method outperform other approaches and enable successful robotic applications." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 547, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 307 + ], + "type": "text", + "content": "Grasp Detection. Grasp detection is a popular task in both computer vision and robotic community [2, 17, 24, 64]. Recently, establishing robotic systems with the ability to follow natural commands has been actively researched [64, 80, 83]. The prevalent solution to the language-driven grasp detection task is to split into two stages: one for grounding the target object, and the other is to synthesize grasp poses from the grounding visual-text correlations [1, 80]. Training in two stages may result in longer inference time [41]. In addition, several works [82, 83, 90] adopt foundation models, such as GroundDINO [42] and GPT-3 [7]. Accessing such commercial foundation models is not always available [71], especially on robotic systems with limited resources or unstable internet connection [37]. In our work, we directly train the model on the large-scale Grasp-Anything++ dataset to inherit the power of a foundation-based dataset, while ensuring a straightforward inference process for the downstream robotic applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 510 + ], + "type": "text", + "content": "Language-driven Grasp Detection Datasets. While there are many grasp datasets have been introduced [9, 17, 22-24, 33, 38, 45, 48, 49, 55, 78], the majority of them overlook the text modality. Therefore, the grasping of objects out of a clutter typically experiences ambiguities in what object to grasp [88]. DailyGrasp [83] is one of the first grasp datasets employing natural language for scene descriptions; however, the scene description corpus in this dataset is relatively small and does not specify which part of the object should be grasped. In our work, we present Grasp Anything++, which is a large-scale language-driven grasping dataset. Grasp-Anything++ provides object instance description with a natural word phrase, for example, \"a blue ceramic vase\". Furthermore, Grasp-Anything++ describes the grasping object at both the part level and object level, providing more information for the robot to execute the grasping [86]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "content": "Diffusion Models for Robotic Applications. Diffusion models [28] have emerged as the new state-of-the-art method of generative tasks [84]. Recent have witnessed growing attention for utilizing diffusion models in robotic applications [61]. Liu et al. [43] propose a diffusion model to handle the language-guided object rearrangement task. Diffusion models are also applied to other robotic tasks such as motion planning [10], and trajectory optimization [31]. The authors in [70] present a diffusion model to determine grasp poses by minimizing a SDF loss. Overall, the diffusion models employed in previous works often combine visual and text features in a latent mechanism [6], which may cause interpretability problems [83] for robotic systems that require low-level controls [20]. To tackle this challenge, we propose a training objective that explicitly contributes to the denoising process. We demonstrate that our proposed strategy is more effective than the latent strategy." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17903" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 545, + 387 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 387 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 387 + ], + "type": "table", + "html": "
StepDescription
Scene GenerationUserPlease help me generate scene descriptions for natural arrangements of daily objects. Each description has the following form: <Object_1><Object_2>...<Verb><Container Object>. Please also ensure the incorporation of a rich and varied lexicon in the scene descriptions.
SampleA steel knife, a polished fork and a pristine ceramic plate on a wooden table.
Text-to-ImageWe use Stable Diffusion [60] to proceed text-to-image generation.
Object MaskingUserFor object part-level description, given an input list {<Object_1>, <Object_2>, ...}, the output will be a list that describes the parts of objects as: {<Object_1>: [(Part_1.1>, <Part_1.2>, ..., <Object_2>]: [(Part_2.1>, <Part_2.2>, ...]}.
Sample{knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}
Post ProcessWe use OFA [75] and SAM [34] to locate the region describing the objects.
Part MaskingUserGiven the object list and part lists of each scene description, you will generate for me all prompts with the following format: {<Manipulation Action><Object_ID><Part_ID)}. The part that is more suitable for human grasping is positioned at the start of the list to represent the grasping actions.
SampleGive me the steel knife; Grasp the knife at its handle.
Post ProcessWe leverage VLPart [67] to locate the region describing the parts of objects.
Grasp GenerationUserGenerate for me a scene description with grasp instructions following the templates.
SampleScene description: A steel knife, a polished fork and a pristine ceramic plate on a wooden table. Object list: {knife, fork, plate}. Part lists: {knife: [handle, blade], fork: [handle, neck, stem, tines], plate: [rim, base]}. Prompts: Give me the steel knife; Grasp the knife at its handle.
Grasp LabellingWe utilize a pretrained RAGT-3/3 [9] to generate grasp poses corresponding to the located region.
", + "image_path": "c6b0f0d9dc25bce9ff0e4aba3c82a426edbb52f89db8874df2fc57e3eb3cef19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 389, + 546, + 412 + ], + "lines": [ + { + "bbox": [ + 47, + 389, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 546, + 412 + ], + "type": "text", + "content": "Table 1. Grasp-Anything++ creation pipeline. We utilize ChatGPT to generate scene descriptions and grasp instructions from the user input. We generate images given scene descriptions and automatically synthesize the grasp poses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 417, + 224, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 224, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 224, + 430 + ], + "type": "text", + "content": "3. The Grasp-Anything++ Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": "We utilize large-scale foundation models to create the Grasp-Anything++. Our dataset offers open-vocabulary grasping commands and images with associated groundtruth. There are three key steps in establishing our dataset: i) prompting procedure, ii) image synthesis and grasp poses annotation, and iii) post-processing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 515, + 171, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 171, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 171, + 529 + ], + "type": "text", + "content": "3.1. Prompting Procedure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": "We first establish prompt-based procedures to generate a large-scale scene description corpus as well as grasp prompt instructions. In particular, we utilize ChatGPT to generate the prompts for two tasks: " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " ) Scene descriptions: Sentences capturing the scene arrangement, including the extracted object and part lists, and " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " ) Grasp instructions: Prompts directing the robot to grasp specific objects or parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "We follow a procedure in Table 1 to implement ChatGPT's output templates. The reference target in the grasp instruction may be either an object or an object's part. When the reference is an object's part, that part is directly selected as the reference in the grasp instruction sentence. If the reference is an object, we determine the grasping region on the part of the object that is likely to be grasped in everyday scenarios as described in affordance theory [53]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 418, + 512, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 512, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 512, + 430 + ], + "type": "text", + "content": "3.2. Image Synthesis and Grasp Annotation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 521 + ], + "type": "text", + "content": "Image Synthesis. Given the scene description corpus, we first utilize a large-scale pretrained text-to-image model, namely, Stable Diffusion [60] to generate images from scene descriptions. Next, we perform a series of visual grounding and image segmentation using OFA [75], Segment-Anything [34], and VLPart [67] to locate the referenced object or part to the grasp instruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 617 + ], + "type": "text", + "content": "Grasp Annotation. Grasp poses are represented as 2D rectangles, consistent with prior research and practical compatibility with real-world parallel grippers [17, 33]. Utilizing a pretrained network [9], we annotate grasp poses based on part segmentation masks. Since potential inaccuracies in these candidate poses could occur, we follow the procedure as defined in [74] to evaluate the quality of generated grasp poses to discard unreasonable grasp poses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": "Specifically, grasp quality is evaluated through net torque " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = (\\tau_{1} + \\tau_{2}) - RMg" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": ", where resistance at contact points is " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{i} = K\\mu_{s}F\\cos \\alpha_{i}" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": ". With constants such as " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " (mass), " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " (gravitational acceleration), " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " (geometrical characteristics), " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{s}" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " (static friction coefficient), and " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " (applied force), accurately determining " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": " directly is challenging due to the physical difficulties in precisely measuring " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{s}" + }, + { + "bbox": [ + 304, + 618, + 546, + 714 + ], + "type": "text", + "content": ". Thus, we employ a surrogate measure," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17904" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 202, + 159 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 202, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 202, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 202, + 159 + ], + "type": "image", + "image_path": "6dd8292cf455040829c3b39186f0b5587aa1a90bc9ba6d2451d5fef7ba4d0903.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 162, + 167, + 171 + ], + "lines": [ + { + "bbox": [ + 85, + 162, + 167, + 171 + ], + "spans": [ + { + "bbox": [ + 85, + 162, + 167, + 171 + ], + "type": "text", + "content": "(a) Number of categories" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 205, + 76, + 358, + 159 + ], + "blocks": [ + { + "bbox": [ + 205, + 76, + 358, + 159 + ], + "lines": [ + { + "bbox": [ + 205, + 76, + 358, + 159 + ], + "spans": [ + { + "bbox": [ + 205, + 76, + 358, + 159 + ], + "type": "image", + "image_path": "b5a413908e008f1647111a74eddf433d39bffa972ced42a05b9d3dea412a1315.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 161, + 326, + 171 + ], + "lines": [ + { + "bbox": [ + 236, + 161, + 326, + 171 + ], + "spans": [ + { + "bbox": [ + 236, + 161, + 326, + 171 + ], + "type": "text", + "content": "(b) Scene description length" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 175, + 515, + 186 + ], + "lines": [ + { + "bbox": [ + 77, + 175, + 515, + 186 + ], + "spans": [ + { + "bbox": [ + 77, + 175, + 515, + 186 + ], + "type": "text", + "content": "Figure 2. Data statistics. We analyze object categories and the use of natural language in the Grasp-Anything++ dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 360, + 72, + 547, + 158 + ], + "blocks": [ + { + "bbox": [ + 360, + 72, + 547, + 158 + ], + "lines": [ + { + "bbox": [ + 360, + 72, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 360, + 72, + 547, + 158 + ], + "type": "image", + "image_path": "a122adec72ea413f0b6ecbafc6da97578f717f46af82b0f44385d364baa1b675.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 161, + 513, + 171 + ], + "lines": [ + { + "bbox": [ + 394, + 161, + 513, + 171 + ], + "spans": [ + { + "bbox": [ + 394, + 161, + 513, + 171 + ], + "type": "text", + "content": "(c) Number of object/part references." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}} = \\frac{\\cos\\alpha_1 + \\cos\\alpha_2}{R}" + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "text", + "content": ", as an alternative. As a result in [12], antipodal grasps score higher on " + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "text", + "content": ", indicating better quality. Consequently, grasps are evaluated based on " + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{T}}" + }, + { + "bbox": [ + 46, + 187, + 286, + 256 + ], + "type": "text", + "content": ", with positive values indicating positive grasps and others considered as negative." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 286, + 376 + ], + "type": "text", + "content": "Post Processing. Despite training on extensive datasets, Stable Diffusion [60] may produce subpar content, commonly termed as hallucination [30] when generating images from the text prompts. To address this, we perform manual reviews to filter out such images, with qualitative examples in our figures. Our process includes checks at every stage to prevent duplicate or hallucinated content. However, manual inspection introduces biases, which we counter with guidelines focusing on abnormal structures or implausible gravity (Fig. 3), aligns with approach in the literature [62]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 48, + 379, + 124, + 456 + ], + "blocks": [ + { + "bbox": [ + 48, + 379, + 124, + 456 + ], + "lines": [ + { + "bbox": [ + 48, + 379, + 124, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 379, + 124, + 456 + ], + "type": "image", + "image_path": "00cfd26594998ed4dacfb405fc228105fffad00c13fc6abddd77333389d6a20f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 468, + 286, + 513 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 286, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 286, + 513 + ], + "type": "text", + "content": "Figure 3. Failure image generation cases. Images generated by Stable Diffusion [60] may exhibit hallucinatory artifacts such as sunglasses lacking a lens, scissors with an anomalous structure, and a spoon not resting properly on a table." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 379, + 205, + 455 + ], + "blocks": [ + { + "bbox": [ + 129, + 379, + 205, + 455 + ], + "lines": [ + { + "bbox": [ + 129, + 379, + 205, + 455 + ], + "spans": [ + { + "bbox": [ + 129, + 379, + 205, + 455 + ], + "type": "image", + "image_path": "7efe25e835b004bb6cabb821598425187b0795efc3c7b6460d93172f8dfd0bce.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 210, + 379, + 286, + 455 + ], + "blocks": [ + { + "bbox": [ + 210, + 379, + 286, + 455 + ], + "lines": [ + { + "bbox": [ + 210, + 379, + 286, + 455 + ], + "spans": [ + { + "bbox": [ + 210, + 379, + 286, + 455 + ], + "type": "image", + "image_path": "911907bda31ab4c17981e2704b4571e25edea51096c28024977411aed9caed9a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 518, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 286, + 590 + ], + "type": "text", + "content": "Additionally, ChatGPT-generated scene prompts often duplicate [44]. To address this, we use duplication checking, filtering out identical prompts with BERTScore [89], which assesses sentence similarity through cosine similarities of token embeddings. We remove sentences with a BERTScore above 0.85 as in prior study [76]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 598, + 137, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 137, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 137, + 609 + ], + "type": "text", + "content": "3.3. Data Statistics" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Number of Categories. To evaluate the object category diversity, we apply a methodology akin to that in [16]. Utilizing 300 categories from LVIS dataset [26], we employ a pretrained model [91] to identify 300 candidate objects from our dataset for each category. We then curate a subset comprising 90,000 objects, refining it by excluding items that do not align semantically with their designated categories. A category is considered significant if it has more" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 190, + 545, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 225 + ], + "type": "text", + "content": "than 40 objects. Fig. 2a shows the results. Overall, our dataset spans over 236 categories from LVIS dataset, indicating a notable degree of object diversity in our dataset." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 225, + 545, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 321 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 321 + ], + "type": "text", + "content": "Scene Descriptions. Fig. 2b shows the distribution of scene descriptions based on sentence length. The analysis reveals a wide range of sentence lengths, spanning from 10 words to 100 words per sentence. On average, each scene description consists of approximately 54 words, indicative of detailed and descriptive sentences. These scene descriptions correspond to sets of grasp instructions. Fig. 2c further shows the objects and object parts in scene descriptions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 321, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 477 + ], + "type": "text", + "content": "Diversity Analysis. We assess the diversity of occlusion and lighting conditions in the dataset. Regarding occlusion, we use a pretrained YOLOv5 model to identify objects within images. The results indicate that " + }, + { + "bbox": [ + 304, + 321, + 545, + 477 + ], + "type": "inline_equation", + "content": "93.8\\%" + }, + { + "bbox": [ + 304, + 321, + 545, + 477 + ], + "type": "text", + "content": " of images have a substantial overlap of five or more bounding boxes, which suggests a diverse range of occlusion within the Grasp-Anything++ dataset. Regarding lighting conditions, we convert images to YCbCr to analyze Y channel (luminance) and find that GraspL1M has the most diverse lighting conditions, identifying by the lowest Gini coefficient (a metric to measure the inequality of a distribution) of 0.26, compared to VMRD [87] (0.31), OCID-grasp [2] (0.32), Cornell [33] (0.62), Jacquard [17] (0.91)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 487, + 495, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 495, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 495, + 502 + ], + "type": "text", + "content": "4. Language-driven Grasp Detection" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 507, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 651 + ], + "type": "text", + "content": "Motivation. The use of diffusion model for language-driven grasp detection is motivated by its efficiency in various generative tasks [15,28,43,84]. Conditional generation, such as our language-driven grasp detection task, aligns seamlessly with diffusion models' capabilities [29]. Moreover, language-driven grasp detection represents a fine-grained problem in which the outputs strongly depend on the text input [4]. For example, \"grasp the steak knife\" and \"grasp the kraft knife\" refer to two different objects on the image. To this end, we propose using contrastive loss with diffusion model to tackle this task, as contrastive learning is a popular solution for fine-grained tasks [8, 14, 85]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 659, + 506, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 506, + 671 + ], + "type": "text", + "content": "4.1. Constrastive Loss for Diffusion Model" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We represent the target grasp pose as " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " in the diffusion model. The objective of our diffusion process of language-driven grasp detection involves denoising from a noisy state" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17905" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": " to the original grasp pose " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": ", conditioned on the input image and grasp instruction represented by " + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "content": "In a diffusion process [28], assume that " + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_{1:T}|\\mathbf{x}_0)" + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "content": " is the forward process and we parameterize the reverse process by " + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{0:T})" + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "content": ". The conditional diffusion process [19] assumes " + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "inline_equation", + "content": "\\hat{q}" + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "content": " is the forward process but with the inclusion of a condition " + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 96, + 287, + 168 + ], + "type": "text", + "content": ". The goal of the reverse process is to optimize the variational bound on negative log likelihood [28]" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 182, + 287, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 287, + 222 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} \\left[ - \\log p _ {\\theta} (\\mathbf {x} _ {T}) - \\sum_ {t \\geq 1} \\log \\frac {p _ {\\theta} (\\mathbf {x} _ {t - 1} | \\mathbf {x} _ {t})}{q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1})} \\right]. \\tag {1}", + "image_path": "8590cebc2a365d019851eaddf6e44e87af67b3d6f80a3a6e356907ac764ed9bf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 226, + 185, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 226, + 185, + 238 + ], + "spans": [ + { + "bbox": [ + 59, + 226, + 185, + 238 + ], + "type": "text", + "content": "We prove in the Appendix that" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 252, + 295, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 295, + 343 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathbb {E} \\left[ \\underbrace {C - \\log p _ {\\theta} (\\mathbf {x} _ {T})} _ {\\text {C o n s t a n t}} + \\underbrace {\\log \\hat {q} (\\mathbf {x} _ {0} | \\mathbf {x} _ {T} , y)} _ {\\text {C o n t r a s t i v e}} + \\right. \\\\ \\underbrace {\\sum_ {t > 1} D _ {\\mathrm {K L}} \\left(\\hat {q} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t} , \\mathbf {x} _ {0} , y\\right) \\| p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right)\\right) - \\log p _ {\\theta} \\left(\\mathbf {x} _ {0} \\mid \\mathbf {x} _ {1} , y\\right)} _ {\\text {D e n o i s i n g s c o r e}}. \\tag {2} \\\\ \\end{array}", + "image_path": "45db0ca29f6235727ffdacedd65eb071d87404ace35c78d8bb08d3e138168585.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "text", + "content": "The terms " + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_\\theta (\\mathbf{x}_{t - 1}|\\mathbf{x}_t))" + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)" + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "text", + "content": " in Equation 2 are similar to the concept of denoising score used in [28]. Thus, we can represent the quantity " + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "inline_equation", + "content": "\\sum_{t > 1}D_{\\mathrm{KL}}(\\hat{q} (\\mathbf{x}_{t - 1}|\\mathbf{x}_t,\\mathbf{x}_0,y)\\| p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t))-" + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\mathbf{x}_0|\\mathbf{x}_1,y)" + }, + { + "bbox": [ + 46, + 344, + 287, + 404 + ], + "type": "text", + "content": " by the loss utilized in [68, 69]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 419, + 290, + 446 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 290, + 446 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 290, + 446 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f u s i o n}} = \\mathbb {E} _ {\\mathbf {x} _ {0} \\sim q (\\mathbf {x} _ {0} | y), t \\sim [ 1, T ]} \\left[ \\mathbf {x} _ {0} - f \\left(\\mathbf {x} _ {t + 1}, t + 1, \\tilde {\\mathbf {x}} _ {0}\\right) \\right] ^ {2}. \\tag {3}", + "image_path": "5f74116164bf82f1c4c0dfb297468a391888c212aa3f62999050c4e8d487b06e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_T|\\cdot)" + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "content": " is equivalent to an isotropic Gaussian distribution as " + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "inline_equation", + "content": "T \\to +\\infty" + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "content": ", the estimation quantity " + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\mathbf{x}_T)" + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "content": " converges to a constant when " + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "inline_equation", + "content": "\\theta \\to \\theta^*" + }, + { + "bbox": [ + 46, + 447, + 287, + 494 + ], + "type": "text", + "content": ". Therefore, we can ignore the first term of Equation 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": "Finally, the term " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": " of Equation 2 provides more information about the relation between " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": ". As this term is intractable [65], we parameterize " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\hat{q} (\\mathbf{x}_0|\\mathbf{x}_T,y)" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "p_{\\psi ,y}(\\mathbf{x}_0,\\mathbf{x}_T)" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": ". This estimation resembles the noise-contrastive estimation [27], where the " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": " can be considered as a pair of contrastive estimation and " + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 495, + 287, + 577 + ], + "type": "text", + "content": " can be estimated by a contrastive loss." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": "In [19, 68, 73], the authors indicate that predicting " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": " is often infeasible but predicting an estimation " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": " is tractable and can be used as a 'pseudo' estimation of " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": ". We denote " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": " as an estimation of " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": ". The loss term " + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "inline_equation", + "content": "\\hat{q}(\\mathbf{x}_0 | \\mathbf{x}_T, y)" + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": " can be approximated by using the following contrastive loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 652, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n t r a s t i v e}} = \\max \\left(0, \\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} - M\\right), \\tag {4}", + "image_path": "9a356b9ae1196b5fea02d7670f147ad94c64a8627a46d140a07d722cba8c85e0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the number of dimension of " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the variance schedule at timestep " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "t = \\overline{1;T}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "Proposition 1. Suppose that " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": " are independent, and that" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 95, + 477, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 95, + 477, + 125 + ], + "spans": [ + { + "bbox": [ + 373, + 95, + 477, + 125 + ], + "type": "interline_equation", + "content": "\\left\\| \\frac {\\sqrt {\\overline {{\\alpha_ {T}}}} \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {T}}{\\sqrt {1 - \\overline {{\\alpha_ {T}}}}} \\right\\| _ {2} ^ {2} \\geq M.", + "image_path": "b208a03c91d18326f3567cd49cfaa98c247feaf0ba1613748445f69ecf0d7a89.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "text", + "content": "Then there exists " + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "inline_equation", + "content": "C > 0" + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "text", + "content": " such that: for arbitrary " + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrastive}} < \\delta" + }, + { + "bbox": [ + 305, + 129, + 547, + 153 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 163, + 474, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 163, + 474, + 178 + ], + "spans": [ + { + "bbox": [ + 376, + 163, + 474, + 178 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\| \\tilde {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\| _ {2} ^ {2} \\right] < C \\delta .", + "image_path": "0fcd6736e401e1b629abdf8e4ca09f171ebe1aa094b5db2be679acbc593cb8bd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 188, + 454, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 188, + 454, + 200 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 454, + 200 + ], + "type": "text", + "content": "Proof. See Supplementary Material." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 535, + 188, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 535, + 188, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 535, + 188, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 535, + 188, + 545, + 198 + ], + "type": "image", + "image_path": "2eef1af9873495c5885bb62b027cc1295ddaef5beef475a9a959387855b8c009.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": "Remark 1.1. Proposition 1 suggests that if the contrastive loss " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrastive}}" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": " tends to zero, then the prediction " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": " will approach the ground truth " + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 306, + 213, + 545, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": "Remark 1.2. The condition " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 \\geq M" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": " is suitable for our language-driven grasp detection task as " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": " are two contrastive quantities, therefore, we can assume there is a minimum distance between " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": ". In addition, in the proof of Proposition 1, we see that " + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\left[\\left\\| \\frac{\\sqrt{\\overline{\\alpha_T}}\\tilde{\\mathbf{x}}_0 - \\mathbf{x}_T}{\\sqrt{1 - \\overline{\\alpha_T}}}\\right\\|_2^2 - M\\right] = \\beta^2\\mathbb{E}\\left[\\| \\tilde{\\mathbf{x}}_0 - \\mathbf{x}_0\\|_2^2\\right]" + }, + { + "bbox": [ + 305, + 252, + 545, + 368 + ], + "type": "text", + "content": ", which is always nonnegative. Therefore, it is both theoretically and experimentally reasonable to add this assumption." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 310, + 380, + 556, + 544 + ], + "blocks": [ + { + "bbox": [ + 310, + 380, + 556, + 544 + ], + "lines": [ + { + "bbox": [ + 310, + 380, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 310, + 380, + 556, + 544 + ], + "type": "image", + "image_path": "c5f8cf0da96fc9cb60067e238944d2fab949583d467d2ea986898423b76b7ced.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 552, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 586 + ], + "type": "text", + "content": "Figure 4. Language-drive Grasp Detection (LGD) network. We present the network architecture (left) and the proposed training objectives of the denoising process (right)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 598, + 531, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 598, + 531, + 611 + ], + "spans": [ + { + "bbox": [ + 305, + 598, + 531, + 611 + ], + "type": "text", + "content": "4.2. Language-driven Grasp Detection Network" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Network. Our network operates on two conditions: an image denoted as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and a corresponding text prompt represented as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". To process these conditions, we employ a vision encoder to extract visual features from " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and a text encoder to derive textual embeddings from " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". The resulting feature vectors, denoted as " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "e'" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ", are subsequently subjected to a fusion module, ALBEF [39]. We leverage the attention mask generated by the ALBEF fusion module" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17906" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 81, + 127, + 159 + ], + "blocks": [ + { + "bbox": [ + 48, + 81, + 127, + 159 + ], + "lines": [ + { + "bbox": [ + 48, + 81, + 127, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 81, + 127, + 159 + ], + "type": "image", + "image_path": "31af37597999a812ea03d30352742866a3875271953175cf6a4ffff0652e4476.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 132, + 81, + 211, + 159 + ], + "blocks": [ + { + "bbox": [ + 132, + 81, + 211, + 159 + ], + "lines": [ + { + "bbox": [ + 132, + 81, + 211, + 159 + ], + "spans": [ + { + "bbox": [ + 132, + 81, + 211, + 159 + ], + "type": "image", + "image_path": "853a17736d910c97ba413f075bf351c062d98c253545ebe034163a21153f2cc6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 81, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 261, + 72, + 324, + 80 + ], + "lines": [ + { + "bbox": [ + 261, + 72, + 324, + 80 + ], + "spans": [ + { + "bbox": [ + 261, + 72, + 324, + 80 + ], + "type": "text", + "content": "Give me the bottle" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 81, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 215, + 81, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 215, + 81, + 294, + 159 + ], + "type": "image", + "image_path": "5ef3066826567afd9937574af595edf403a795ffbf49d6fe73a2884db31ab644.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 81, + 377, + 159 + ], + "blocks": [ + { + "bbox": [ + 299, + 81, + 377, + 159 + ], + "lines": [ + { + "bbox": [ + 299, + 81, + 377, + 159 + ], + "spans": [ + { + "bbox": [ + 299, + 81, + 377, + 159 + ], + "type": "image", + "image_path": "d040c5158a765ad1a6044d085d67f7decc812517f74e1524a65a4f6f74cba34f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 81, + 461, + 159 + ], + "blocks": [ + { + "bbox": [ + 383, + 81, + 461, + 159 + ], + "lines": [ + { + "bbox": [ + 383, + 81, + 461, + 159 + ], + "spans": [ + { + "bbox": [ + 383, + 81, + 461, + 159 + ], + "type": "image", + "image_path": "c175ad61339879b6e2afa8d1769ec697bffea22e87d70e9611a8752a03d954b0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 467, + 81, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 467, + 81, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 467, + 81, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 467, + 81, + 545, + 159 + ], + "type": "image", + "image_path": "14bec2ca7857c1b62b3ceaea57a18780bc18796d6d727500f5815f31d4d1799b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 48, + 174, + 127, + 252 + ], + "blocks": [ + { + "bbox": [ + 48, + 174, + 127, + 252 + ], + "lines": [ + { + "bbox": [ + 48, + 174, + 127, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 127, + 252 + ], + "type": "image", + "image_path": "d0b133780a36ae1595fe6d8c5715704ebd2fbdd2a08ca50223b25be31d333f6e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 253, + 102, + 262 + ], + "lines": [ + { + "bbox": [ + 73, + 253, + 102, + 262 + ], + "spans": [ + { + "bbox": [ + 73, + 253, + 102, + 262 + ], + "type": "text", + "content": "(a) Ours" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 132, + 174, + 211, + 252 + ], + "blocks": [ + { + "bbox": [ + 132, + 174, + 211, + 252 + ], + "lines": [ + { + "bbox": [ + 132, + 174, + 211, + 252 + ], + "spans": [ + { + "bbox": [ + 132, + 174, + 211, + 252 + ], + "type": "image", + "image_path": "989181ff313a07de2bfb6e230168a17bc2ccab2c0b3677013935364ead3ea5cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 253, + 200, + 262 + ], + "lines": [ + { + "bbox": [ + 143, + 253, + 200, + 262 + ], + "spans": [ + { + "bbox": [ + 143, + 253, + 200, + 262 + ], + "type": "text", + "content": "(b) GR-ConvNet" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 265, + 419, + 276 + ], + "lines": [ + { + "bbox": [ + 172, + 265, + 419, + 276 + ], + "spans": [ + { + "bbox": [ + 172, + 265, + 419, + 276 + ], + "type": "text", + "content": "Figure 5. Language-driven grasp detection results visualization." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 216, + 174, + 294, + 252 + ], + "blocks": [ + { + "bbox": [ + 241, + 165, + 346, + 174 + ], + "lines": [ + { + "bbox": [ + 241, + 165, + 346, + 174 + ], + "spans": [ + { + "bbox": [ + 241, + 165, + 346, + 174 + ], + "type": "text", + "content": "Grasp the wristwatch at its dial" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 174, + 294, + 252 + ], + "lines": [ + { + "bbox": [ + 216, + 174, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 216, + 174, + 294, + 252 + ], + "type": "image", + "image_path": "6e2945012a302f4787aa6e85ec085254518d9e7de32d80a98e9c1f9c0a09dbc6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 253, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 223, + 253, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 223, + 253, + 287, + 262 + ], + "type": "text", + "content": "(c) Det-Seg-Refine" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 299, + 174, + 377, + 252 + ], + "blocks": [ + { + "bbox": [ + 299, + 174, + 377, + 252 + ], + "lines": [ + { + "bbox": [ + 299, + 174, + 377, + 252 + ], + "spans": [ + { + "bbox": [ + 299, + 174, + 377, + 252 + ], + "type": "image", + "image_path": "2f21faf38cd302e13e459d2bece98cbb230586d4b6bb03c29d368ec85a62c9a3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 253, + 361, + 262 + ], + "lines": [ + { + "bbox": [ + 317, + 253, + 361, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 361, + 262 + ], + "type": "text", + "content": "(d) GG-CNN" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 174, + 461, + 252 + ], + "blocks": [ + { + "bbox": [ + 383, + 174, + 461, + 252 + ], + "lines": [ + { + "bbox": [ + 383, + 174, + 461, + 252 + ], + "spans": [ + { + "bbox": [ + 383, + 174, + 461, + 252 + ], + "type": "image", + "image_path": "60222a5d53b79fdec6ca8b5bcbbefffec84e1853bb0ecfe3c95933e1258ff630.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 253, + 444, + 262 + ], + "lines": [ + { + "bbox": [ + 400, + 253, + 444, + 262 + ], + "spans": [ + { + "bbox": [ + 400, + 253, + 444, + 262 + ], + "type": "text", + "content": "(e) CLIPORT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 467, + 174, + 545, + 252 + ], + "blocks": [ + { + "bbox": [ + 467, + 174, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 467, + 174, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 467, + 174, + 545, + 252 + ], + "type": "image", + "image_path": "468e895b375c80b4d2740a82d90feffd28d1d99d02bcbc9449a60dd580697f48.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 253, + 533, + 262 + ], + "lines": [ + { + "bbox": [ + 480, + 253, + 533, + 262 + ], + "spans": [ + { + "bbox": [ + 480, + 253, + 533, + 262 + ], + "type": "text", + "content": "(f) CLIP-Fusion" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": "as the estimation " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": ". Next, we aggregate three elements: the estimation region " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": ", the grasp pose at the current timestep " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + 1}" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": ", and the timestep " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": ". These inputs are combined using MLP layers, similar to the approach outlined in [68]. Specifically, the output operation can be expressed as: " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0)" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": ", where function " + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 280, + 287, + 388 + ], + "type": "text", + "content": " encompasses a composition of multiple MLP layers. Additional specifics regarding these universal MLP layers are provided in the Supplementary Material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": "Training Objective. In our context, conditioned grasp detection models the distribution " + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0|y)" + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": " as the reversed diffusion process of gradually cleaning " + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t + 1}" + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": ". Instead of predicting " + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": " as formulated by [28], we follow Ramesh et al. [58] and predict the signal itself, i.e., " + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = f(\\mathbf{x}_{t + 1}, t + 1, \\tilde{\\mathbf{x}}_0)" + }, + { + "bbox": [ + 46, + 388, + 287, + 483 + ], + "type": "text", + "content": " with the simple objective [28]. To this end, we utilize the contrastive loss as in Equation 4 to explicitly improve the learning objective of the denoising process:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 97, + 498, + 287, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 498, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 97, + 498, + 287, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {c o n t r a s t i v e}} + \\mathcal {L} _ {\\text {d i f f u s i o n}}. \\tag {5}", + "image_path": "5b882273c41637bda93456e72d7e606405994a10ca6e435a9f6f0b1adac5c7dd.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 521, + 128, + 535 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 535 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 535 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 542, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 601 + ], + "type": "text", + "content": "We conduct experiments to evaluate our proposed method and Grasp-Anything++ dataset using both the vision-based metrics and real robot experiments. We then demonstrate zero-shot grasp results and discuss the challenges and open questions for future works." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 610, + 266, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 266, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 266, + 624 + ], + "type": "text", + "content": "5.1. Language-driven Grasp Detection Results" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Baselines. We compare our language-driven grasp detection method (LGD) with the linguistically supported versions of GR-CNN [35], Det-Seg-Refine [2], GG-CNN [47], CLIPORT [64] and CLIP-Fusion [80]. In all cases, we employ a pretrained CLIP [57] or BERT [18] as the text embedding. The implementation details of all baselines can be found in our Supplementary Material." + } + ] + } + ], + "index": 27 + }, + { + "type": "table", + "bbox": [ + 317, + 277, + 541, + 392 + ], + "blocks": [ + { + "bbox": [ + 317, + 277, + 541, + 392 + ], + "lines": [ + { + "bbox": [ + 317, + 277, + 541, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 541, + 392 + ], + "type": "table", + "html": "
BaselineSeenUnseenH
GR-ConvNet [35] + CLIP [57]0.370.180.24
Det-Seg-Refine [2] + CLIP [57]0.300.150.20
GG-CNN [47] + CLIP [57]0.120.080.10
CLIPORT [64]0.360.260.29
CLIP-Fusion [80]0.400.290.33
LGD (ours) + BERT [18]0.440.380.41
LGD (ours) + CLIP [57]0.480.420.45
", + "image_path": "38361574347e58d03965137c19d3fab300af94d36904fa32e6c47cb747e3173b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": "Setup. To assess the generalization of all methods trained on Grasp-Anything++, we utilize the concept of base and new labels [92] in zero-shot learning. We categorize LVIS labels from Section 3.3 to form labels for our experiment. In particular, we select " + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": " of these labels by frequency for 'Base' and assign the remaining " + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": " to 'New'. We also use the harmonic mean ('H') to measure the overall success rates [92]. Our primary evaluation metric is the success rate, defined similarly to [35], necessitating an IoU score of the predicted grasp exceeding " + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": " with the ground truth grasp and an offset angle less than " + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 304, + 415, + 545, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "spans": [ + { + "bbox": [ + 304, + 548, + 545, + 619 + ], + "type": "text", + "content": "Main Results. Table 2 shows the results of language-driven grasp detection on the Grasp-Anything++ dataset. The findings indicate a notable performance advantage of our LGD over other baseline approaches, with LGD outperforming the subsequent best-performing baselines (CLIP-Fusion) by margins of 0.14 on Grasp-Anything++ dataset." + } + ] + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 329, + 630, + 528, + 692 + ], + "blocks": [ + { + "bbox": [ + 330, + 396, + 521, + 407 + ], + "lines": [ + { + "bbox": [ + 330, + 396, + 521, + 407 + ], + "spans": [ + { + "bbox": [ + 330, + 396, + 521, + 407 + ], + "type": "text", + "content": "Table 2. Language-driven grasp detection results." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 630, + 528, + 692 + ], + "lines": [ + { + "bbox": [ + 329, + 630, + 528, + 692 + ], + "spans": [ + { + "bbox": [ + 329, + 630, + 528, + 692 + ], + "type": "table", + "html": "
BaselineSeenUnseenH
LGD w/o predicting x̂00.150.080.10
LGD w/o contrastive loss0.450.400.42
LGD w contrastive loss0.480.420.45
", + "image_path": "47352291292a3c0213f5b552b9967c29ec62d218fb646a89b1c718626a560a2a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 361, + 695, + 490, + 706 + ], + "lines": [ + { + "bbox": [ + 361, + 695, + 490, + 706 + ], + "spans": [ + { + "bbox": [ + 361, + 695, + 490, + 706 + ], + "type": "text", + "content": "Table 3. Contrastive loss analysis." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17907" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 284, + 171 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 284, + 171 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 284, + 171 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 284, + 171 + ], + "type": "image", + "image_path": "1c307034979a56be59006f6bf22980e82842e7b9f5c980ae459ba3e08fd0bf9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 176, + 221, + 186 + ], + "lines": [ + { + "bbox": [ + 113, + 176, + 221, + 186 + ], + "spans": [ + { + "bbox": [ + 113, + 176, + 221, + 186 + ], + "type": "text", + "content": "Figure 6. Loss visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 193, + 168, + 310 + ], + "blocks": [ + { + "bbox": [ + 50, + 193, + 168, + 310 + ], + "lines": [ + { + "bbox": [ + 50, + 193, + 168, + 310 + ], + "spans": [ + { + "bbox": [ + 50, + 193, + 168, + 310 + ], + "type": "image", + "image_path": "8c68367085b7a413e767eb39182d92b6b2fad4e5e1a8b4884bb40ecec9166f4c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 318, + 287, + 340 + ], + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 340 + ], + "type": "text", + "content": "Figure 7. t-SNE visualization. We apply t-SNE to cluster the vision-and-language features with and without contrastive loss." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 169, + 193, + 286, + 310 + ], + "blocks": [ + { + "bbox": [ + 169, + 193, + 286, + 310 + ], + "lines": [ + { + "bbox": [ + 169, + 193, + 286, + 310 + ], + "spans": [ + { + "bbox": [ + 169, + 193, + 286, + 310 + ], + "type": "image", + "image_path": "34186e8bce2d7cb6da9de2e05cece17a842747e303536e459d5f00a33de8d4db.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 233, + 545, + 255 + ], + "type": "text", + "content": "Figure 8. Attention map visualization. We compare the attention map when utilizing our proposed contrastive loss and when not." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 287, + 422 + ], + "type": "text", + "content": "Contrastive Loss Analysis. Table 3 presents the performance of LGD under varied configurations. The outcomes emphasize the substantial influence of the training objective (contrastive loss) and the importance of language instructions in enhancing LGD performance on both seen and unseen classes in the Grasp-Anything++ dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "text", + "content": "Fig. 6 shows the contrastive loss " + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contrastive}}" + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "text", + "content": " approaching towards 0 during training, indicating the grasp pose estimation " + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_0" + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "text", + "content": " aligns with the ground truth " + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 46, + 424, + 287, + 604 + ], + "type": "text", + "content": ", as anticipated by Proposition 1. The subsequent attention maps visualization in Fig. 8 shows the attention region is meaningful and improves the results when employing our proposed contrastive loss compared to its absence. Moreover, we employ t-SNE for vision-and-language embedding visualization, as in [46], by processing 2,000 samples from the Grasp-Anything++ dataset through the ALBEF module. The outcomes reveal that our contrastive loss facilitates better object classification, as evidenced in Fig.7 by clearer segregation of pixel embeddings across various semantic classes, underscoring contrastive loss's role in refining embeddings' differentiation for improved class distinctions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 676 + ], + "type": "text", + "content": "Qualitative Results. Fig. 5 presents qualitative results of the language-driven grasp detection task, suggesting that our LGD method generates more semantically plausible than other baselines. Despite satisfactory performance, LGD occasionally predicts incorrect results, with a detailed analysis of these cases available in our Appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Robotic Validation. We provide quantitative results by integrating our language-driven grasp detection pipeline for a robotic grasping application with a KUKA LBR iwa" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 224 + ], + "type": "image", + "image_path": "0dcadfe26b1b66087528e78d00518a1f905d590b7151ce2af416a6b221d0d2b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 325, + 257, + 535, + 355 + ], + "blocks": [ + { + "bbox": [ + 325, + 257, + 535, + 355 + ], + "lines": [ + { + "bbox": [ + 325, + 257, + 535, + 355 + ], + "spans": [ + { + "bbox": [ + 325, + 257, + 535, + 355 + ], + "type": "table", + "html": "
BaselineSingleCluttered
GR-ConvNet [35] + CLIP [57]0.330.30
Det-Seg-Refine [2] + CLIP [57]0.300.23
GG-CNN [47] + CLIP [57]0.100.07
CLIPORT [64]0.270.30
CLIP-Fusion [80]0.400.40
LGD (ours)0.430.42
", + "image_path": "4eb7ccd31a7cfffba546a11c2584d1794f4fd7d5989136f9a19e70a4f6f48bd8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 363, + 536, + 374 + ], + "lines": [ + { + "bbox": [ + 315, + 363, + 536, + 374 + ], + "spans": [ + { + "bbox": [ + 315, + 363, + 536, + 374 + ], + "type": "text", + "content": "Table 4. Robotic language-driven grasp detection results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 387, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 543 + ], + "type": "text", + "content": "R820 robot. Using the RealSense D435i camera, the grasp pose inferred from approaches in Table 4 is transformed into the 6DoF grasp pose, similar to [35]. The optimization-based trajectory planner in [5,72] is employed to execute the grasps. Experiments are conducted for two scenarios, i.e., the single object scenario and the cluttered scene scenario, of a set of 20 real-world daily objects. In each scenario, we run 30 experiments using baselines listed in Table 4 and a predefined grasping prompt corpus. The results exhibit that our LGD outperforms other baselines. Furthermore, although LGD is trained on our Grasp-Anything++ which is a solely synthesis dataset created by foundation models, it still shows reasonable results on real-world objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 453, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 453, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 453, + 563 + ], + "type": "text", + "content": "5.2. Zero-shot Grasp Detection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 653 + ], + "type": "text", + "content": "Our proposed Grasp-Anything++ is a large-scale dataset. Apart from the language-driven grasp detection task, we believe it can be used for other purposes. In this experiment, we seek to answer the question: Can Grasp-Anything++ be useful in the traditional grasp detection task without text? Consequently, we verify our Grasp-Anything++ and LGD (no text) with other existing datasets and grasping methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Setup. We setup an LGD (no text) version, and other state-of-the-art grasp detection methods GR-ConvNet [35], Det-Seg-Refine [2], GG-CNN [47]. We use five datasets: our Grasp-Anything++, Jacquard [17], Cornell [33], VMRD [87], and OCID-grasp [2] in this experiment." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17908" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 79, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 545, + 160 + ], + "type": "table", + "html": "
BaselineGrasp-Anything++ (ours)Jacquard [17]Cornell [33]VMRD [87]OCID-grasp [2]
BaseNewHBaseNewHBaseNewHBaseNewHBaseNewH
GR-ConvNet [35]0.710.590.640.880.660.750.980.740.840.770.640.700.860.670.75
Det-Seg-Refine [2]0.620.570.590.860.600.710.990.760.860.750.600.660.800.620.70
GG-CNN [47]0.680.570.620.780.560.650.960.750.840.690.530.590.710.630.67
LGD (no text) (ours)0.740.630.680.890.690.770.970.760.850.790.660.720.880.680.76
", + "image_path": "e1593434cf1da0bb562034e28ceb91b09db38ce79f5c2d26943e8137b15ae853.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 185, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 288, + 281 + ], + "type": "text", + "content": "Zero-shot Results. Table 5 summarizes the base-to-new grasp detection results on five datasets. Overall, the performance of LGD even without the language branch is better than other baselines across all datasets. Furthermore, this table also shows that our Grasp-Anything++ dataset is more challenging to train as the detection results are lower than related datasets using the same approaches due to the greater coverage of unseen objects in the testing phase." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 291, + 287, + 354 + ], + "blocks": [ + { + "bbox": [ + 190, + 166, + 402, + 177 + ], + "lines": [ + { + "bbox": [ + 190, + 166, + 402, + 177 + ], + "spans": [ + { + "bbox": [ + 190, + 166, + 402, + 177 + ], + "type": "text", + "content": "Table 5. Base-to-new zero-shot grasp detection results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 291, + 287, + 354 + ], + "lines": [ + { + "bbox": [ + 50, + 291, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 50, + 291, + 287, + 354 + ], + "type": "table", + "html": "
Test TrainJacquardCornellVMRDOCID-graspGrasp-Anything++
Jacquard [17]0.870.510.130.210.17
Cornell [33]0.070.980.200.120.13
VMRD [87]0.060.210.790.110.10
OCID-grasp [2]0.090.120.200.740.11
Grasp-Anything++ (ours)0.410.630.300.390.65
", + "image_path": "0372a56ad7adbd16e878c5b14ffa8a808afcdc2d21496ae75bb199206f0bed3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 363, + 251, + 374 + ], + "lines": [ + { + "bbox": [ + 82, + 363, + 251, + 374 + ], + "spans": [ + { + "bbox": [ + 82, + 363, + 251, + 374 + ], + "type": "text", + "content": "Table 6. Cross-dataset grasp detection results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": "Cross-dataset Evaluation. To further verify the usefulness of our Grasp-Anything++ dataset, we conduct the cross-dataset validation in Table 6. We use the GRConvNet [35] to reuse its results on existing grasp datasets. GR-ConvNet is trained on a dataset (row) and evaluated on another dataset (column). For example, training on Jacquard and testing on Cornell yields an accuracy of 0.51. Notably, training with our dataset improves performance by approximately " + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "10 - 33\\%" + }, + { + "bbox": [ + 46, + 378, + 287, + 484 + ], + "type": "text", + "content": " compared to other datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 593 + ], + "type": "text", + "content": "In the wild grasp detection. Fig. 9 shows visualization results using LGD (no text) trained on our Grasp-Anything++ dataset on random internet images and other datasets images. We can see that the detected grasp poses are adequate in quality and quantity. This demonstrates that although our Grasp-Anything++ is fully created by foundation models without having any real images, models trained on our Grasp-Anything++ dataset still generalize well on real-world images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 599, + 119, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 119, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 119, + 610 + ], + "type": "text", + "content": "5.3. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Our experiments indicate that Grasp-Anything++ can serve as a foundation dataset for both language-driven and traditional grasp detection tasks. However, there are certain limitations. First, our dataset lacks depth images for directly being applied to robotic applications [52]. Second, we remark that the creation of our dataset is time-consuming and relies on access to the ChatGPT API. Fortunately, future research can reuse our provided assets (images, prompts," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 182, + 383, + 258 + ], + "blocks": [ + { + "bbox": [ + 307, + 182, + 383, + 258 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 383, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 383, + 258 + ], + "type": "image", + "image_path": "d45184855b9574769a59b2cc5d6a5a523aadcc2d15ecca86d9317af75a06c019.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 387, + 182, + 462, + 258 + ], + "blocks": [ + { + "bbox": [ + 387, + 182, + 462, + 258 + ], + "lines": [ + { + "bbox": [ + 387, + 182, + 462, + 258 + ], + "spans": [ + { + "bbox": [ + 387, + 182, + 462, + 258 + ], + "type": "image", + "image_path": "a10c88bfdc345c472ae706e4ff528b3568882724b860abf6dcfae55a39089ad7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 467, + 182, + 542, + 258 + ], + "blocks": [ + { + "bbox": [ + 467, + 182, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 467, + 182, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 467, + 182, + 542, + 258 + ], + "type": "image", + "image_path": "74ab2e0f15cced209ade5f6546ff3d4275012a2e365f1bb48382ea7f66ef018d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 262, + 383, + 338 + ], + "blocks": [ + { + "bbox": [ + 307, + 262, + 383, + 338 + ], + "lines": [ + { + "bbox": [ + 307, + 262, + 383, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 262, + 383, + 338 + ], + "type": "image", + "image_path": "cb740b29edd411199dd92af3a1a8f467e814b76806cffefa8bf4bc433575e067.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 342, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 375 + ], + "type": "text", + "content": "Figure 9. In the wild grasp detection. Top row images are from GraspNet [24], YCB-Video [78], NBMOD [9] datasets; bottom row shows internet images." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 387, + 262, + 462, + 338 + ], + "blocks": [ + { + "bbox": [ + 387, + 262, + 462, + 338 + ], + "lines": [ + { + "bbox": [ + 387, + 262, + 462, + 338 + ], + "spans": [ + { + "bbox": [ + 387, + 262, + 462, + 338 + ], + "type": "image", + "image_path": "dab9c132ea433c097788cc4bda274d6644c4c48ab2609b7975562fb7471f7f57.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 467, + 262, + 542, + 338 + ], + "blocks": [ + { + "bbox": [ + 467, + 262, + 542, + 338 + ], + "lines": [ + { + "bbox": [ + 467, + 262, + 542, + 338 + ], + "spans": [ + { + "bbox": [ + 467, + 262, + 542, + 338 + ], + "type": "image", + "image_path": "dd68889eeac9faeb8996a249d69ba68fa871e77f0b86937ff65f3e1ca7351ec5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 431 + ], + "type": "text", + "content": "etc.) without starting from scratch. Furthermore, our experiments show that adding language to the grasp detection task (Table 2) poses a more challenging problem compared to standard grasp detection task (Table 5)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 541 + ], + "type": "text", + "content": "We see several interesting future research directions. First, future work could investigate the use of text or imaged-to-3D models [79] or image-to-depth [59] and reuse our dataset's prompts and images to construct 3D language-driven grasp datasets. Additionally, beyond linguistic grasp instruction adherence, our dataset holds potential for varied applications, including scene understanding [81] and scene generation [3], hallucination analysis [30], and humanrobot interaction [11]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 559, + 378, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 378, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 378, + 572 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We introduce Grasp-Anything++, a large-scale dataset with 1M images and 10M grasp prompts for language-driven grasp detection tasks. We propose LGD, a diffusion-based method to tackle the language-driven grasp detection task. Our diffusion model employs a contrastive training objective, which explicitly contributes to the denoising process. Empirically, we have shown that Grasp-Anything++ serves as a foundation grasp detection dataset. Finally, our LGD improves the performance of other baselines, and the real-world robotic experiments further validate the effectiveness of our dataset and approach." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17909" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 190 + ], + "type": "text", + "content": "[2] Stefan Ainetter and Friedrich Fraundorfer. End-to-end trainable deep neural network for robotic grasp detection and semantic segmentation from rgb. In ICRA, 2021. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 288, + 214 + ], + "type": "text", + "content": "[3] Dor Arad Hudson and Larry Zitnick. Compositional transformers for scene generation. NeurIPS, 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 287, + 248 + ], + "type": "text", + "content": "[4] Umar Asif, Jianbin Tang, and Stefan Harrer. Grasynet: An efficient convolutional neural network for real-time grasp detection for low-powered devices. In IJCAI, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 293 + ], + "type": "text", + "content": "[5] Florian Beck, Minh Nhat Vu, Christian Hartl-Nesic, and Andreas Kugi. Singularity avoidance with application to online trajectory optimization for serial manipulators. IFAC-PapersOnLine, 56(2):284-291, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 338 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 287, + 384 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 33, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 428 + ], + "type": "text", + "content": "[8] Guy Bukchin, Eli Schwartz, Kate Saenko, Ori Shahar, Rogerio Feris, Raja Giryes, and Leonid Karlinsky. Fine-grained angular contrastive learning with coarse labels. In CVPR, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "text", + "content": "[9] Boyuan Cao, Xinyu Zhou, Congmin Guo, Baohua Zhang, Yuchen Liu, and Qianqiu Tan. Nbmod: Find it and grasp it in noisy background. arXiv preprint arXiv:2306.10265, 2023. 2, 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 518 + ], + "type": "text", + "content": "[10] Joao Carvalho, An T Le, Mark Baierl, Dorothea Koert, and Jan Peters. Motion planning diffusion: Learning and planning of robot motions with diffusion models. arXiv preprint arXiv:2308.01557, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 563 + ], + "type": "text", + "content": "[11] Paola Cascante-Bonilla, Hui Wu, Letao Wang, Rogerio S Feris, and Vicente Ordonez. Simvqa: Exploring simulated environments for visual question answering. In CVPR, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 566, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 287, + 587 + ], + "type": "text", + "content": "[12] I-Ming Chen and Joel W Burdick. Finding antipodal point grasps on irregularly shaped objects. T-RA, 1993. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 620 + ], + "type": "text", + "content": "[13] Sijia Chen and Baochun Li. Language-guided diffusion model for visual grounding. arXiv preprint arXiv:2308.09599, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 655 + ], + "type": "text", + "content": "[14] Hongjun Choi, Anirudh Som, and Pavan Turaga. Amc-loss: Angular margin contrastive loss for improved explainability in image classification. In CVPR workshops, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[15] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. TPAMI, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 127 + ], + "type": "text", + "content": "[17] Amaury Depierre, Emmanuel Dellandrea, and Liming Chen. Jacquard: A large scale dataset for robotic grasp detection. In IROS, 2018. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 171 + ], + "type": "text", + "content": "[18] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 194 + ], + "type": "text", + "content": "[19] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 226 + ], + "type": "text", + "content": "[20] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[21] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palme: An embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 316 + ], + "type": "text", + "content": "[22] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. A billion ways to grasp: An evaluation of grasp sampling schemes on a dense, physics-based grasp data set. In ISRR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "text", + "content": "[23] Clemens Eppner, Arsalan Mousavian, and Dieter Fox. Acronym: A large-scale grasp dataset based on simulation. In ICRA, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 382 + ], + "type": "text", + "content": "[24] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In CVPR, 2020. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 436 + ], + "type": "text", + "content": "[25] Maximilian Gilles, Yuhao Chen, Tim Robin Winter, E Zhixuan Zeng, and Alexander Wong. Metagraspnet: A large-scale benchmark dataset for scene-aware ambidextrous bin picking via physics-based metaverse synthesis. In CASE, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 469 + ], + "type": "text", + "content": "[26] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[27] Michael Gutmann and Aapo Hyvarinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In AISTAT, 2010. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 525 + ], + "type": "text", + "content": "[28] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 4, 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 559 + ], + "type": "text", + "content": "[29] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv preprint arXiv:2204.03458, 2022. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[30] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232, 2023. 4, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 657 + ], + "type": "text", + "content": "[31] Michael Janner, Yilun Du, Joshua Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. In ICML, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[32] Krishna Murthy Jatavallabhula, Alihusein Kuwajerwala, Qiao Gu, Mohd Omama, Tao Chen, Shuang Li, Ganesh Iyer, Soroush Saryazdi, Nikhil Keetha, Ayush Tewari, et al. Conceptfusion: Open-set multimodal 3d mapping. arXiv preprint arXiv:2302.07241, 2023. 1" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17910" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Yun Jiang, Stephen Moseson, and Ashutosh Saxena. Efficient grasping from rgbd images: Learning using a new rectangle representation. In ICRA, 2011. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 152 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 185 + ], + "type": "text", + "content": "[35] Sulabh Kumra, Shirin Joshi, and Ferat Sahin. Antipodal robotic grasping using generative residual convolutional neural network. In IROS, 2020. 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 241 + ], + "type": "text", + "content": "[36] Jaewoong Lee, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, Yunji Kim, Jin-Hwa Kim, Jung-Woo Ha, and Sung Ju Hwang. Text-conditioned sampling framework for text-to-image generation with masked generative models. arXiv preprint arXiv:2304.01515, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 286 + ], + "type": "text", + "content": "[37] Meng-Lun Lee, Sara Behdad, Xiao Liang, and Minghui Zheng. Task allocation and planning for product disassembly with human-robot collaboration. Robotics and Computer-Integrated Manufacturing, 76, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 331 + ], + "type": "text", + "content": "[38] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. *IJRR*, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[39] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. NeurIPS, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 421 + ], + "type": "text", + "content": "[40] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 466 + ], + "type": "text", + "content": "[41] Jirong Liu, Ruo Zhang, Hao-Shu Fang, Minghao Gou, Hongjie Fang, Chenxi Wang, Sheng Xu, Hengxu Yan, and Cewu Lu. Target-referenced reactive grasping for dynamic objects. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 521 + ], + "type": "text", + "content": "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[43] Weiyu Liu, Tucker Hermans, Sonia Chernova, and Chris Paxton. Structdiffusion: Object-centric diffusion for semantic rearrangement of novel objects. arXiv preprint arXiv:2211.04604, 2022. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 624 + ], + "type": "text", + "content": "[44] Yiheng Liu, Tianle Han, Siyuan Ma, Jiayue Zhang, Yuanyuan Yang, Jiaming Tian, Hao He, Antong Li, Mengshen He, Zhengliang Liu, et al. Summary of chatgpt-related research and perspective towards the future of large language models. Meta-Radiology, page 100017, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 679 + ], + "type": "text", + "content": "[45] Jeffrey Mahler, Jacky Liang, Sherdil Niyaz, Michael Laskey, Richard Doan, Xinyu Liu, Juan Aparicio Ojea, and Ken Goldberg. Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[46] Jiaxu Miao, Zongxin Yang, Leilei Fan, and Yi Yang. Fedseg: Class-heterogeneous federated learning for semantic segmentation. In CVPR, 2023. 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[47] Douglas Morrison, Peter Corke, and Jürgen Leitner. Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach. arXiv preprint arXiv:1804.05172, 2018. 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 152 + ], + "type": "text", + "content": "[48] Douglas Morrison, Peter Corke, and Jürgen Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. RA-L, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "text", + "content": "[49] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 217 + ], + "type": "text", + "content": "[50] Yao Mu, Shunyu Yao, Mingyu Ding, Ping Luo, and Chuang Gan. Ec2: Emergent communication for embodied control. In CVPR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 274 + ], + "type": "text", + "content": "[51] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pretraining via embodied chain of thought. arXiv preprint arXiv:2305.15021, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 275, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 275, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 275, + 546, + 329 + ], + "type": "text", + "content": "[52] Rhys Newbury, Morris Gu, Lachlan Chumbley, Arsalan Mousavian, Clemens Eppner, Jürgen Leitner, Jeannette Bohg, Antonio Morales, Tamim Asfour, Danica Kragic, et al. Deep learning approaches to grasp synthesis: A review. T-RO, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 546, + 364 + ], + "type": "text", + "content": "[53] Toan Nguyen, Minh Nhat Vu, An Vuong, Dzung Nguyen, Thieu Vo, Ngan Le, and Anh Nguyen. Open-vocabulary affordance detection in 3d point clouds. In IROS, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 365, + 546, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 385 + ], + "type": "text", + "content": "[54] OpenAI. Introducing ChatGPT. Software. Accessed: July 6th 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 387, + 546, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 387, + 546, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 387, + 546, + 420 + ], + "type": "text", + "content": "[55] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 546, + 454 + ], + "type": "text", + "content": "[56] Robert Platt. Grasp learning: Models, methods, and performance. Annual Review of Control, Robotics, and Autonomous Systems, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 455, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 546, + 510 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 307, + 510, + 546, + 554 + ], + "type": "text", + "content": "[58] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 555, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 578 + ], + "type": "text", + "content": "[59] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In CVP4, 2021. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 546, + 611 + ], + "type": "text", + "content": "[60] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 656 + ], + "type": "text", + "content": "[61] Kallol Saha, Vishal Mandadi, Jayaram Reddy, Ajit Srikanth, Aditya Agarwal, Bipasha Sen, Arun Singh, and Madhava Krishna. Edmp: Ensemble-of-costs-guided diffusion for motion planning. arXiv preprint arXiv:2309.11414, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 657, + 546, + 689 + ], + "type": "text", + "content": "[62] Schuhmann et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 546, + 713 + ], + "type": "text", + "content": "[63] Dhruv Shah, Błajej Osiński, brian richter, and Sergey Levine. Lm-nav: Robotic navigation with large pre-trained models of" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17911" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "language, vision, and action. In Karen Liu, Dana Kulic, and Jeff Ichnowski, editors, Proceedings of The 6th Conference on Robot Learning, volume 205 of PMLR. PMLR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[64] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. CIport: What and where pathways for robotic manipulation. In CoRL, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 176 + ], + "type": "text", + "content": "[65] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 287, + 220 + ], + "type": "text", + "content": "[66] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In ICCV, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 223, + 287, + 266 + ], + "type": "text", + "content": "[67] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 287, + 301 + ], + "type": "text", + "content": "[68] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2022. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 303, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 303, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 303, + 287, + 326 + ], + "type": "text", + "content": "[69] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In CVPR, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[70] Julien Urain, Niklas Funk, Jan Peters, and Georgia Chalvatzaki. Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion. In ICRA, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 406 + ], + "type": "text", + "content": "[71] Sai Vemprala, Rogerio Bonatti, Arthur Bucker, and Ashish Kapoor. Chatgpt for robotics: Design principles and model abilities. Microsoft Auton. Syst. Robot. Res, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 407, + 287, + 461 + ], + "type": "text", + "content": "[72] Minh Nhat Vu, Florian Beck, Michael Schwegel, Christian Hartl-Nesic, Anh Nguyen, and Andreas Kugi. Machine learning-based framework for optimally solving the analytical inverse kinematics for redundant manipulators. Mechatronics, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 287, + 508 + ], + "type": "text", + "content": "[73] An Vuong, Minh Nhat Vu, Toan Tien Nguyen, Baoru Huang, Dzung Nguyen, Thieu Vo, and Anh Nguyen. Language-driven scene synthesis using multi-conditional diffusion model. arXiv preprint arXiv:2310.15948, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 554 + ], + "type": "text", + "content": "[74] An Dinh Vuong, Minh Nhat Vu, Hieu Le, Baoru Huang, Binh Huynh, Thieu Vo, Andreas Kugi, and Anh Nguyen. Grasp-anything: Large-scale grasp dataset from foundation models. arXiv preprint arXiv:2309.09818, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 609 + ], + "type": "text", + "content": "[75] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In ICML, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 677 + ], + "type": "text", + "content": "[76] Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania, and Ernesto W De Luca. Legal norm retrieval with variations of the bert model combined with tfidf vectorization. In Proceedings of the eighteenth international conference on artificial intelligence and law, pages 285-294, 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[77] Julia Wolleb, Robin Sandkuhler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In MIDL, 2022. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 708 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[78] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[79] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, and Zhangyang Wang. Neurallift-360: Lifting an in-the-wild 2d photo to a 3d object with 360deg views. In CVPR, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 216 + ], + "type": "text", + "content": "[80] Kechun Xu, Shuqi Zhao, Zhongxiang Zhou, Zizhang Li, Huajin Pi, Yifeng Zhu, Yue Wang, and Rong Xiong. A joint modeling of vision-language-action for target-oriented grasping in clutter. arXiv preprint arXiv:2302.12610, 2023. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 261 + ], + "type": "text", + "content": "[81] Yiteng Xu, Peishan Cong, Yichen Yao, Runnan Chen, Yuenan Hou, Xinge Zhu, Xuming He, Jingyi Yu, and Yuexin Ma. Human-centric scene understanding for 3d large-scale scenarios. In ICCV, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 262, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 262, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 262, + 545, + 304 + ], + "type": "text", + "content": "[82] Zhixuan Xu, Kechun Xu, Rong Xiong, and Yue Wang. Object-centric inference for language conditioned placement: A foundation model based approach. In ICARM, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 307, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 307, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 308, + 307, + 545, + 350 + ], + "type": "text", + "content": "[83] Jange Yang, Wenhui Tan, Chuhao Jin, Bei Liu, Jianlong Fu, Ruihua Song, and Limin Wang. Pave the way to grasp anything: Transferring foundation models for universal pick-place robots. arXiv preprint arXiv:2306.05716, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 352, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 352, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 308, + 352, + 545, + 406 + ], + "type": "text", + "content": "[84] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 2022, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 407, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 407, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 308, + 407, + 545, + 439 + ], + "type": "text", + "content": "[85] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. ICCV, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 441, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 473 + ], + "type": "text", + "content": "[86] Yang Yang, Xibai Lou, and Changyun Choi. Interactive robotic grasping with attribute-guided disambiguation. In ICRA, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 517 + ], + "type": "text", + "content": "[87] Hanbo Zhang, Xuguang Lan, Site Bai, Xinwen Zhou, Zhiqiang Tian, and Nanning Zheng. Roi-based robotic grasp detection for object overlapping scenes. In IROS, 2019. 4, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 308, + 519, + 545, + 562 + ], + "type": "text", + "content": "[88] Hanbo Zhang, Yunfan Lu, Cunjun Yu, David Hsu, Xuguang La, and Nanning Zheng. Invigorate: Interactive visual grounding and grasping in clutter. arXiv preprint arXiv:2108.11092, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 563, + 545, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 563, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 308, + 563, + 545, + 596 + ], + "type": "text", + "content": "[89] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 597, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 597, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 308, + 597, + 545, + 630 + ], + "type": "text", + "content": "[90] Kaizhi Zheng, Xiaotong Chen, Odest Chadwicke Jenkins, and Xin Wang. Vlmbench: A compositional benchmark for vision-and-language manipulation. NeurIPS, 35, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 632, + 545, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 632, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 308, + 632, + 545, + 674 + ], + "type": "text", + "content": "[91] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 308, + 675, + 545, + 708 + ], + "type": "text", + "content": "[92] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, 2022. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17912" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_content_list.json b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5dd551e52905d57402f8acd2ea2b15b56bbf115e --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_content_list.json @@ -0,0 +1,1772 @@ +[ + { + "type": "text", + "text": "Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates", + "text_level": 1, + "bbox": [ + 181, + 130, + 789, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ka Chun Shum¹ Jaeyeon Kim¹ Binh-Son Hua²,⁴ Duc Thanh Nguyen³ Sai-Kit Yeung¹", + "bbox": [ + 116, + 214, + 851, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology $^{2}$ VinAI $^{3}$ Deakin University $^{4}$ Trinity College Dublin", + "bbox": [ + 76, + 244, + 890, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing.", + "bbox": [ + 75, + 342, + 473, + 614 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 646, + 207, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Editing of 3D scenes by insertion or removal of objects has been a fundamental task in computer graphics and computer vision. The task has often been performed using traditional 3D scene authoring tools [5, 9]. For example, to insert an object into a 3D scene, traditional approach requires user to manually select the object and position it into the scene. This manual pipeline has been adopted in a wide range of applications, such as furniture arrangement in interior design [15] and asset creation in game development [10, 13].", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in deep learning have opened new directions to scene modeling. Neural radiance field (NeRF) [40] is a pioneer to render view-consistent photo-realistic images using neural networks. Generative models such as generative adversarial networks (GANs) [12] and diffusion models [17] learn to output photo-realistic images from un", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b7ec97ea103525c7798b80878297a59f637096c4f2bdaa606d4b0000184f75d.jpg", + "image_caption": [ + "Input: multi-view object images & object text prompt" + ], + "image_footnote": [], + "bbox": [ + 511, + 325, + 877, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78346f513bd46aa69d4cc1355d1cf5da1407a424aa2e5bdbc06b4a7bfa420998.jpg", + "image_caption": [ + "Input: multi-view background images & background text prompt" + ], + "image_footnote": [], + "bbox": [ + 511, + 441, + 875, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/896e69843f5988da44b044e266c7b7a9fce35bd973a0e30a35200e08444ac447.jpg", + "image_caption": [ + "Output: multi-view object-blended images & novel view synthesis", + "Figure 1. Object insertion. We propose a language-driven method for view-consistent 3D object insertion into a background NeRF scene. Given an object defined by a set of multi-view images, our method generates plausible text-guided insertion results that require geometry manipulation of the original background NeRF." + ], + "image_footnote": [], + "bbox": [ + 513, + 537, + 879, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "constrained image collections. Recent text-guided diffusion models [50, 51, 53] have shown great promise in generating high-quality and diverse images from a single text prompt.", + "bbox": [ + 496, + 762, + 890, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Such successes inspire us to revisit 3D scene editing via language-driven image synthesis techniques. Particularly, for object insertion (see Fig. 1), our method performs view-consistent edits by taking as input a set of multi-view object images, multi-view background images, and a target text prompt. We utilize a text-to-image diffusion model to con", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5176", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tinueously synthesize multi-view images containing both a target object and a background during NeRF training. These synthesized images can be iteratively used to refine the NeRF of the background to learn geometric and appearance features of the object. This approach to refining a radiance field is also known as dataset update [14]. However, a particular challenge is that the refinement process may result in unstable NeRF training, degrading its rendering quality due to inconsistent views synthesized. To address this issue, we propose a pose-conditioned dataset update strategy that gradually engages the target object into the background, beginning at a randomly selected pose (view), then views close to the already-used views before propagating to views further away. We observe that this pose-conditioned strategy significantly improves the NeRF learning, reducing rendering artifacts and maintaining view-consistent rendering. In summary, we make the following contributions in our work.", + "bbox": [ + 76, + 90, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new framework for object manipulation in NeRFs via text-to-image diffusion. Our work can create high-quality 3D scenes from simple inputs (text prompts and multi-view images). It, therefore, has the potential for conveniently building a variant of 3D object libraries that is applicable for the scene editing task.", + "- We propose a pose-conditioned dataset update strategy that stabilizes the fusion of objects into a background NeRF, enabling view-consistent rendering. Our method requires no priors on the geometry or texture of the objects, unlike existing works relying on accurate meshes [7, 20, 31, 81], depth maps [42, 67], fine masks [58, 63], semantics [1], or lighting assumption [69, 78].", + "- We showcase our method in two case studies: object insertion and object removal via a user-friendly manner. Specifically, we use a 3D bounding box to define the location for object insertion/removal. Box-based location requires only a rough orientation, which is easy to be adjusted and visualized using real-time NeRF GUIs [43, 59].", + "- We conduct extensive experiments to validate key techniques of our method and demonstrate its state-of-the-art performance in NeRF editing." + ], + "bbox": [ + 76, + 349, + 470, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 694, + 215, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image synthesis. Early attempts in deep learning-based image synthesis have utilized generative models such as GANs [12]. Several methods combine supervised losses with adversarial losses to learn class-conditional GANs [41] and image-to-image translations [21, 65, 80]. StyleGAN [22] and its variants [23, 24] perform image editing directly on latent representations, but such editorings are not intuitive and thus difficult to control the output.", + "bbox": [ + 76, + 719, + 468, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models [17] have made significant progress in image synthesis with high-quality data samples constructed from sophisticated forward and denoising diffusion steps. In addition, vision-language models such as", + "bbox": [ + 76, + 839, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CLIP [49] have made text prompts an intuitive condition to generate and edit images. These developments have led to text-guided diffusion models [50, 51, 53] which offer excellent synthesis quality. Downstream applications then can be built by fine-tuning these text-guided models to fit with the application domains [4, 25, 36, 52, 70, 77]. Our method follows this streamline where we use a text-to-image diffusion model [52] to guide the fusion of objects into a background.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural 3D modeling. Image-based 3D modeling methods use convolutional neural networks (CNNs) [29] to learn 3D structures from multiple images [11, 18, 33, 60, 79]. However, CNNs often struggle to deal with complex shapes, texture, and lighting captured in the images. Follow-up works integrate differentiable rendering and represent 3D structures as neural surfaces [39] or shapes [8]. 3D-aware GANs integrates volume rendering into their generators to synthesize novel views from a single image [6, 45, 56, 57]. NeRFs [40] apply the same neural rendering approach, but are optimized on a ray rendering loss on multi-view input images. There are methods addressing the limitations of NeRFs in various aspects such as visual artifacts [2], data complexity [37], camera poses [74], and computational efficiency [43, 76].", + "bbox": [ + 496, + 231, + 895, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF editing. NeRF editing has often been carried out by parameter tuning [35], layer feature fusion [61], or deformable rays [75]. An alternative to editing a NeRF is to directly amend the multi-view images used to learn it. For example, NeRF stylization methods [19, 44, 46, 62] freeze the geometry branch and optimize the color branch in a NeRF to stylize multi-view images. Several methods attempt to decompose an existing NeRF, which in turn holds color information and separates voxels considering multi-view masks [30] or semantics [27, 28]. NeRF inpainting fills simple unseen background geometry and colors with help of depth priors [42] or filtering inpainted multi-views [67].", + "bbox": [ + 496, + 462, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Creation of complex geometry and vivid colors for a NeRF is challenging due to higher-level requirements of consistency. DiscoScene [72] fuses background and object NeRFs, thus is unable to condition customized contents. FocalDreamer [31] and DreamEditor [81] rely on fine meshes to function, disregarding the advantage of multi-view representation of NeRFs. Our method is perhaps most similar to Instruct-NeRF2NeRF [14] which shares a related data updating schema. However, Instruct-NeRF2NeRF [14] always generates geometry aligned with the original geometry (e.g., transfer a sneaker to a sneaker-shape apple), thus fails in most cases of object insertion or removals. Besides, it requires heavy retraining of a diffusion model on large-scale self-constructed datasets.", + "bbox": [ + 496, + 643, + 895, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, text-to-image diffusion has been applied to generate 3D contents. DreamFusion [47] introduced a score distillation sampling (SDS) loss that progressively consolidates", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/383d36ec2e8eba0016311fdae1eedf40f9a416e5a7ee6cbc6234feeab3164338.jpg", + "image_caption": [ + "Figure 2. Overview of our pipeline. We customize and fine-tune a text-to-image diffusion model for view synthesis in an inpainting manner (left). We then apply the model to progressively fuse an object into background views to update a background NeRF (right). The process of view synthesis and NeRF updating is performed repeatedly. Views generated by the diffusion model are added to an on-going dataset to strengthen the NeRF. In return, the NeRF renders color hints for the diffusion model to create new views." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 315 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "view information from a diffusion model into NeRFs. This loss has been applied with additional treatments to image resolution [32], conditional images [38], photo-realism [66], or scene geometry [71]. We also adopt the SDS loss but develop a novel training schedule to address challenges in multi-view object and background fusion for NeRF editing.", + "bbox": [ + 75, + 414, + 473, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 518, + 168, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 544, + 187, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the ease of presentation, we first describe our method for object insertion. We then extend it to object removal. An overview of our method for object insertion is illustrated in Fig. 2. Here we aim to insert an object into a background NeRF in two steps. In the first step, we synthesize training views for the NeRF with the target object embedded in the background. In the second step, the NeRF is updated with the synthesized views to learn the geometry and appearance of the object. There are two key challenges in this approach. First, background preserving is required in the synthesized views for the NeRF updating. Second, image synthesis may generate view-inconsistent images, causing artifacts in the resulting NeRF.", + "bbox": [ + 75, + 568, + 470, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the first challenge (i.e., object-blended image synthesis with background preserving), we leverage a state-of-the-art diffusion model to image synthesis, and opt to customize the model with text prompts for object blending (see Sec. 3.2). We formulate this task as image inpainting where we place a binary mask on each background image to indicate the location where the object is inserted in, and adjust the text prompts with both the object and background.", + "bbox": [ + 75, + 763, + 470, + 885 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve view-consistent renderings, we propose a new", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "strategy, namely pose-conditioned dataset updates, to schedule the data used in the NeRF updating (see Sec. 3.3). Our strategy is inspired by an important observation about the nature of NeRF: a view rendered by a NeRF maintains an extent of pose-aware color information from nearby already-used views, the nearer the more noticeable. Therefore, if we pass nearby renderings to the diffusion model with properly controlled noise, view-consistent results can be generated based on the learned color hints. From such an observation, we design a novel data scheduler for our NeRF updating. We initially train the NeRF using regular method on a dataset of multi-view background images. We then progressively fuse the object into the NeRF by iteratively updating the dataset with object-blended background images in a pose-ordered manner, i.e., views are sorted in such a way that new views are acquired nearby already-used views. Our method can also be adapted to implement object removal (see Sec. 3.4).", + "bbox": [ + 496, + 414, + 893, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Object-blended image synthesis", + "text_level": 1, + "bbox": [ + 498, + 680, + 779, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our target object and background by a set of multi-view object images $\\mathbf{I}_o$ and background images $\\mathbf{I}_b$ . Our goal is to build an image synthesis model that can blend the object (from $\\mathbf{I}_o$ ) into the background (from $\\mathbf{I}_b$ ) at custom locations specified by binary masks $M$ .", + "bbox": [ + 496, + 704, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $D_{\\theta}$ (with parameters $\\theta$ ) be such an image synthesis model. Here we adopt a pre-trained Stable Diffusion [51] to implement $D_{\\theta}$ . The model $D_{\\theta}$ includes a denoising model $\\epsilon_{\\theta}$ that learns the noise component $\\epsilon_{\\alpha} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{1})$ added in the diffusion process, where $\\alpha \\in [0, 1]$ is a denoising strength. We customize $D_{\\theta}$ with our background and object images in an inpainting fashion. Specifically, for each image $I \\in \\mathbf{I}_b \\cup \\mathbf{I}_o$ , we make a masked-out (background-preserved)", + "bbox": [ + 496, + 780, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5178", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "version $I\\odot M$ where $M\\sim \\mathcal{U}$ is a random square binary mask whose coordinates are sampled from a uniform distribution $\\mathcal{U}$ and $\\odot$ is a pixel-wise product. The image $I$ is associated with a text prompt $p$ where we use $\" ^ { \\text{丰} }$ to indicate our object/background identifiers [52], e.g., $\" *$ sneaker\" means that our target object is a sneaker and $\" *$ table\" is our target background. We fine-tune $D_{\\theta}$ with the following loss:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s y n t h e s i s}} (\\theta) = \\mathbb {E} _ {I, M \\sim \\mathcal {U}, p, \\epsilon_ {\\alpha} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} \\left[ \\| \\epsilon_ {\\alpha} - \\epsilon_ {\\theta} (I) \\| _ {2} ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 223, + 470, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. 1 represents a general form for the fine-tuning process. Intermediate steps, e.g., input formation, latent encoding-decoding, are presented in our supplementary material.", + "bbox": [ + 75, + 253, + 470, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We fine-tune the diffusion model $D_{\\theta}$ with all the images in $\\mathbf{I}_b$ and $\\mathbf{I}_o$ for $n_{bg}$ and $n_{obj}$ times. After fine-tuning, we can use $D_{\\theta}$ to synthesize images $\\hat{I}_b$ that blend the object presented in $\\mathbf{I}_o$ into the background images $I_b \\in \\mathbf{I}_b$ via a combined text prompt $\\tilde{p}$ , e.g., \"a * sneaker on a * table\",", + "bbox": [ + 75, + 299, + 468, + 376 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\theta} \\left(I _ {b} \\odot M, \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 387, + 468, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is set manually to control how much object content in the masked area to be inpainted by $D_{\\theta}$ (we demonstrate how we set $\\alpha$ for different purposes in Sec. 3.3).", + "bbox": [ + 75, + 417, + 468, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pose-conditioned dataset updates", + "text_level": 1, + "bbox": [ + 76, + 472, + 369, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $R_{\\phi}$ (with parameters $\\phi$ ) be our NeRF, which is initially trained with the multi-view background images in $\\mathbf{I}_b$ using an existing approach [43]. We also retrieve a set of camera poses $\\{\\pi(I_b)\\}$ ( $4 \\times 4$ matrices) for the images $I_b \\in \\mathbf{I}_b$ using the pose estimation method COLMAP [54, 55].", + "bbox": [ + 75, + 496, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose to update the NeRF $R_{\\phi}$ progressively with object-blended background images generated by the finetuned diffusion model $D_{\\theta}$ . Our idea is to build a progressive multi-view and pose-conditioned object-blended image dataset $\\hat{\\mathbf{I}}_b$ from $\\mathbf{I}_b$ to be used to update $R_{\\phi}$ . We initialize $\\hat{\\mathbf{I}}_b^{(0)} = \\emptyset$ and $\\mathbf{I}_b^{(0)} = \\mathbf{I}_b$ . We start with a random background image $I_b \\in \\mathbf{I}_b$ . Let $B$ be a 3D bounding box, and $B(I_b)$ be the projection mask of $B$ onto $I_b$ using the camera pose $\\pi (I_b)$ . We generate an object-blended image $\\hat{I}_b$ as,", + "bbox": [ + 75, + 571, + 470, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\theta} \\left(I _ {b} \\odot B \\left(I _ {b}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 722, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $\\alpha = 1$ (maximum strength) to blend the object into the masked area by using only the knowledge that $D_{\\theta}$ learns from $\\tilde{p}$ (during the fine-tuning), and colors from the background image $I_{b}$ as there is no existing color clue in the masked area $I_{b} \\odot B(I_{b})$ .", + "bbox": [ + 75, + 752, + 468, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We update $\\hat{\\mathbf{I}}_b^{(1)} = \\hat{\\mathbf{I}}_b^{(0)}\\cup \\{\\hat{I}_b\\}$ and $\\mathbf{I}_b^{(1)} = \\mathbf{I}_b^{(0)}\\setminus \\{I_b\\}$ . We then update $R_{\\phi}$ using $\\hat{\\mathbf{I}}_b^{(1)}$ , and keep doing so by adding images into $\\hat{\\mathbf{I}}_b$ sequentially. In particular, let $\\hat{\\mathbf{I}}_b^{(n - 1)}$ be the dataset at the $(n - 1)$ -th step. The next image $I_b^{(n)}\\in \\mathbf{I}_b$ for", + "bbox": [ + 75, + 827, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step $n$ is selected so as it is closest (in terms of the camera poses) to already-used images,", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {b} ^ {(n)} = \\underset {I _ {b, k} \\in \\mathbf {I} _ {b} ^ {(n)}} {\\arg \\min } \\underset {I _ {b, j} \\in \\mathbf {I} _ {b} ^ {(n - 1)}} {\\min } \\| \\pi_ {T} (I _ {b, k}) - \\pi_ {T} (I _ {b, j}) \\| ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 128, + 893, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_T$ is the translation component of the pose $\\pi$ . This view selection reflects the standard multi-view reconstruction pipeline where the multi-view data are collected with a smoothly connected, inward-surrounding camera trajectory. This requirement is also fulfilled by most NeRF datasets.", + "bbox": [ + 496, + 170, + 893, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a background image $I_b^{(n)}$ in Eq. 4, to utilize the nearby color hints learned by $R_{\\phi}$ , we create a background-preserved foreground-rendered image $\\tilde{I}_b^{(n)}$ as,", + "bbox": [ + 496, + 247, + 893, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {I} _ {b} ^ {(n)} = \\left(I _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right)\\right) \\oplus \\left(R _ {\\phi} \\left(\\pi \\left(I _ {b} ^ {(n)}\\right)\\right) \\odot \\bar {B} \\left(I _ {b} ^ {(n)}\\right)\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 304, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R_{\\phi}(\\pi(I_b^{(n)}))$ is the rendering result of $R_{\\phi}$ in the pose $\\pi(I_b^{(n)})$ (for the background image $I_b^{(n)}$ ), $\\bar{B}$ is the complement of $B$ (as we want to maintain the foreground rendered by $R_{\\phi}$ ), and $\\oplus$ is a pixel-wise addition.", + "bbox": [ + 496, + 335, + 893, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then generate a view-consistent image $\\hat{I}_b^{(n)}$ by applying $D_{\\theta}$ to $\\tilde{I}_b^{(n)}$ defined in Eq. 5 as,", + "bbox": [ + 496, + 401, + 893, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\theta} \\left(\\tilde {I} _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} ^ {(n)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 445, + 890, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we set $\\alpha$ to a low value to allow $D_{\\theta}$ to utilize color hints from previous nearby views, provided from the rendering results $R_{\\phi}(\\pi (I_b^{(n)}))$ . We empirically found that $\\alpha \\in [0.3, 0.4]$ gives best view-consistent rendering.", + "bbox": [ + 496, + 478, + 890, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Again, we update $\\hat{\\mathbf{I}}_b^{(n)}$ with $\\hat{I}_b^{(n)}$ , and update $R_{\\phi}$ accordingly by minimizing the loss:", + "bbox": [ + 496, + 541, + 893, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N e R F}} (\\phi) = \\mathbb {E} _ {\\hat {I} _ {b} \\in \\hat {\\mathbf {I}} _ {b} ^ {(n)}} \\left[ \\| R _ {\\phi} \\left(\\pi \\left(\\hat {I} _ {b}\\right)\\right) - \\hat {I} _ {b} \\| ^ {2} \\right]. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 580, + 890, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the updating process, we include $n_{near}$ views into the ongoing dataset for every $n_{new}$ NeRF updating steps. In addition, we periodically replace each old view by a new one using Eqs. 5 and 6 for every $n_{old}$ NeRF updating steps. The updating procedure is completed once all the background images in $\\mathbf{I}_b$ have been processed, i.e., $\\mathbf{I}_b^{(n)} = \\emptyset$ .", + "bbox": [ + 496, + 609, + 890, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Adapting to object removal", + "text_level": 1, + "bbox": [ + 500, + 710, + 745, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our framework can be adapted to object removal. In particular, we also fine-tune the diffusion model $D_{\\theta}$ as in Sec. 3.2. However, we do not fine-tune $D_{\\theta}$ with object images as we want to remove objects. Instead, we first individually inpaint all background images on projection masks using background prompts without identifier (e.g., \"a table\"). These inpainted images are then treated as pseudo ground-truth and used to customize $D_{\\theta}$ with the background text prompts containing identifiers (e.g., \"a * table\"). The pseudo ground-truth backgrounds are visually pleasing but still remain cross-view inconsistencies. To circumvent this issue, we perform", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF updating using dataset updates as in Sec. 3.3. However, only background prompts are used during the NeRF updating. We found that this joint 2D-3D interaction gradually transforms disruptive inpainted image regions into view-consistent background images. We show the necessity of the pseudo ground-truth and the NeRF updating in making view-consistent object removal in our experiments.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 210, + 209, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 234, + 179, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For object insertion, we propose a dataset comprising multiview images of 8 backgrounds and 10 objects. The number of images for each background and object ranges from 60 to 100 and 40 to 80, respectively. We collected the images using an iPhone camera and resized the images to $512 \\times 512$ resolution for the ease of training. Except for our self-captured data, we also test on synthetically rendered multiview images [40, 47].", + "bbox": [ + 75, + 258, + 472, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For object removal, we run experiments on the commonly used inpainting datasets from Mip-NeRF-360 [3] and IBR-Net [64]. We centrally cropped images in those datasets to make $512 \\times 512$ images to fit with our pipeline.", + "bbox": [ + 76, + 380, + 470, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 76, + 450, + 184, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For comparisons, we select SOTA baselines from different branches of work that can perform object insertion. To ensure the fairness, we barely modify the baselines, only when necessary to fit them with our settings.", + "bbox": [ + 75, + 473, + 468, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Traditional 3D editing pipeline. This pipeline creates a 3D model from multi-view images. Scene editing is then performed directly on the 3D model. To simulate this traditional pipeline, we use COLMAP [54, 55] to reconstruct the mesh for the background and object, and then manually crop and place the object mesh into the background mesh.", + "bbox": [ + 75, + 534, + 470, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image inpainting. We adopt the inpainting variant of Stable Diffusion [51] as a baseline. However, since image inpainting treats each view independently, for a fair comparison, we perform single-view inpainting on each background image.", + "bbox": [ + 75, + 625, + 470, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-image-to-3D NeRF. NeRF-based view synthesis partially fulfills scene editing. Here we select Zero123 [34], a SOTA that uses a distillation prior [47] from a 2D diffusion model to synthesize novel views for comparison.", + "bbox": [ + 75, + 686, + 470, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instruct-NeRF2NeRF [14]. This work also applies dataset updates for NeRF training. However, it fails to add/remove objects with noticeable non-uniform appearance. To test this method, we re-format text prompts, e.g., we replace \"a * sneaker on a * table\" by \"add/make a sneaker on a table\".", + "bbox": [ + 75, + 746, + 470, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 830, + 290, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We fine-tune the diffusion model $D_{\\theta}$ with object images $(n_{obj} = 5,000$ times) more than background images $(n_{bg} = 500$ times) as we found the model needs extra training to", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "learn objects with complex geometry and texture. For NeRF training, we empirically found $n_{near} = 3$ , $n_{new} = 500$ , and $n_{old} = 10$ balance well the quality and efficiency.", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We run all experiments on a Nvidia RTX 3090 GPU. Diffusion fine-tuning takes around 30 minutes. NeRF updating speed relies on the number of background images while each backpropagation takes 0.5 seconds. Inference speed of the diffusion model is 8 seconds/image. Our NeRF training costs around 2.7 times more than the standard NeRF training.", + "bbox": [ + 496, + 136, + 893, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Qualitative results", + "text_level": 1, + "bbox": [ + 498, + 234, + 676, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object insertion. We present qualitative results of object insertion in Fig. 3. As shown, our method (Fig. 3-f) can generate plausible contents with view consistency. Moverover, the outputs of our method also match well the text prompts and the generated objects are precisely located.", + "bbox": [ + 496, + 258, + 893, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, the traditional 3D pipeline (Fig. 3-b) suffers from seamed object-background boundaries and unrealistic lighting. The geometry in obscure regions, e.g., corners, is not accurately reconstructed. The image inpainting baseline (Fig. 3-c) produces reasonable results on individual views but generates view-inconsistency. Single-image-to-3D NeRF (Fig. 3-d) fails on complex scenes. A larger camera pose shift can result in background mismatch or even content collapse. Instruct-NeRF2NeRF (Fig. 3-e) is known to be strong at stylizing existing geometry but weak at generating new geometry. Under our setting, it performs random view-consistent editing, but fails to accomplish object insertion.", + "bbox": [ + 496, + 334, + 893, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object removal. We illustrate several results of object removal in Fig. 4, which shows that our method can generate view-consistent backgrounds. We also observed that without using pseudo ground-truth background, the removal can cause gradual background collapse, which we discuss further in our ablation study in Sec. 4.6.", + "bbox": [ + 496, + 515, + 893, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Quantitative results", + "text_level": 1, + "bbox": [ + 498, + 614, + 689, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since there is no real-world ground-truth for scene editing, direct quantitative evaluations are not possible. Instead, we adopt the CLIP Score [16], a well-known metric to measure how well an image correlates to a target prompt in the CLIP space. Let $E(I)$ and $E(p)$ be the CLIP embeddings of an image $I$ and a text prompt $p$ . The CLIP Score is defined as,", + "bbox": [ + 496, + 637, + 893, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C L I P S c o r e} (I, p) = \\cos^ {+} (E (I), E (p)) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 737, + 890, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos^{+}(a,b) = \\max (0,\\cos (a,b))$ . We average the CLIPScore $(\\hat{I}_b,\\tilde{p})$ for all the edited images $\\hat{I}_b\\in \\hat{\\mathbf{I}}_b$ paired with their target prompts $\\tilde{p}$ , and use this average value as a performance metric for scene editing; higher score means better performance.", + "bbox": [ + 496, + 763, + 893, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also use the CLIP Directional Consistency (CLIPDC) in [14] to measure the editing quality and consistency across views in the CLIP space. This metric relies on an assumption that the difference in the CLIP space between a background", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7a469988be365c00da5df2342c1bea6c66583c45b8d8940871a4f7260d82837e.jpg", + "image_caption": [ + "Figure 3. Qualitative results of object insertion. Inputs include multi-view object/background images, a 3D bounding box where the object is inserted in, and a text prompt. Note that some baselines use parts of the inputs due to the nature of their techniques." + ], + "image_footnote": [], + "bbox": [ + 80, + 85, + 893, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image $I_{b}$ and its edited version $\\hat{I}_{b}$ should match with the difference between the original prompt $p$ (paired with $I_{b}$ ) and the customized prompt $\\tilde{p}$ (paired with $\\hat{I}_b$ ). Moreover, a good editing should make consistent differences between every pair of $I_{b}$ and $\\hat{I}_b$ , especially for adjacent views. We follow [14] to calculate the CLIPDC between two adjacent", + "bbox": [ + 75, + 773, + 472, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "edited views $\\hat{I}_b^{(i)}$ and $\\hat{I}_b^{(i + 1)}$ as,", + "bbox": [ + 500, + 772, + 709, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {C L I P D C} \\left(\\hat {I} _ {b} ^ {(i)}, \\hat {I} _ {b} ^ {(i + 1)}\\right) = \\\\ \\cos^ {+} (E (\\tilde {p}) - E (p), E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)})) \\times \\tag {9} \\\\ \\cos^ {+} \\left(E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)}), E (\\hat {I} _ {b} ^ {(i + 1)}) - E (I _ {b} ^ {(i + 1)})\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 800, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measure the CLIPDC of an edited scene by averaging", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b5be5d30166ec4c05c787446d7a26e3f009183097c893855cd6f3e1f5e9ea4e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 205, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6890a1aea494ff605eac3c30bbeadbf2bb41848d5ed6e9d18e359022a0aaf9d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 88, + 336, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/721697224708c1135dd2f92bd36eabd8b0f2fab0bdc698c6780dd16fe247148f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 88, + 468, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17a93be35b9bea6ed344c8fa0029092af4a60b57ad719be707571140a4a0f7a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 188, + 205, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9c628322d00cfa69b9153551f1aaf37555778ef9b008c09a670373f22c205c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 188, + 336, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c30d90367c851029730c85d89ed9e5c35b42b25f4106b465dac5a18202928f78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 188, + 468, + 286 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/746b293bc1723199285da4cdf9baa3de98bfe0fa8598f57f62711b07a152b6a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 292, + 205, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/faa48321ebbe27ee602177623a2a2e1fd533191585dac98cd8e06bb6a35893fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 292, + 336, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae6aafa184f599ed0e73e019d6a6ab82988fed8f3d5c61337a2963919fe48bb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 292, + 468, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb01b77239be06ed66a255a408b84c063d4077750c46b86ae16dcbf066e2539a.jpg", + "image_caption": [ + "(a) Original BG with object bounding box" + ], + "image_footnote": [], + "bbox": [ + 80, + 392, + 205, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6ccf694c33f8ea3b986f7b91ce8338dd465bc4c3a6c32be2dc8cdff0a2d9234.jpg", + "image_caption": [ + "(b) Pseudo ground-truth BG" + ], + "image_footnote": [], + "bbox": [ + 210, + 392, + 336, + 489 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4747e311bac08754a3a78caa193274b3524594a8b95950634bb317b38495b81a.jpg", + "image_caption": [ + "(c) View-consistent editing result", + "Figure 4. Qualitative results of object removal where we show the importance of pseudo ground-truth background (BG) in generating view-consistent editing results." + ], + "image_footnote": [], + "bbox": [ + 341, + 392, + 468, + 489 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f893b151b2dd4cc8ff6f372a558cf434788234d82aab8145f684ecad35285cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCLIPScore ↑CLIPDC ↑
Traditional 3D [54, 55]0.26480.1169
Inpainting [51]0.26830.1325
Zero123 [34]0.21970.0361
Instruct-NeRF2NeRF [14]0.23470.0263
Ours0.27430.1642
", + "bbox": [ + 78, + 598, + 468, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results. Higher CLIP metrics scores indicate higher image editing quality or consistency.", + "bbox": [ + 75, + 722, + 468, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the CLIPDC $(\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})$ for all the adjacent image pairs $(\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})$ in the scene. We report both the CLIPScore and CLIPDC of our method and other baselines on 35 unique edits (cross editing on the more generalizable 5 scenes $\\times$ 7 objects) in Tab. 1. Experimental show that our method outperforms all the baselines on both the CLIPScore and CLIPDC metrics.", + "bbox": [ + 75, + 787, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 90, + 653, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this ablation study, we validate the effectiveness of technical components in our method. We suggest readers observe the results in Fig. 5, which shows two views of an edited scene. We provide detailed analyses on these results below. Diffusion model fine-tuning. Recall that we fine-tune the diffusion model $D_{\\theta}$ on both object and background images (in Sec. 3.2). Here we prove that such a fine-tuning is necessary. Fine-tuning on object images makes the diffusion model aware of the same object during the dataset updates (in Sec. 3.3). We verify this in Fig. 5-a, where we skip the fine-tuning of $D_{\\theta}$ on object images. As shown, without using object images, the diffusion model cannot generate the same object across views.", + "bbox": [ + 496, + 113, + 893, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Likewise, fine-tuning the diffusion model with background images helps to preserve the background at inpainted borders (areas in between mask boundary and inner-mask object boundary). As shown in Fig. 5-b, without fine-tuning on background images, the model cannot inpaint the background within the mask region properly. The inpainted background gets darker at latter NeRF updating steps, and finally ends up with a noticeable error. We hypothesize this collapse that the model, without being fine-tuned on relevant background images, applies the bias from pre-learned knowledge which does not align with the surrounding background.", + "bbox": [ + 496, + 310, + 893, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pose-conditioned dataset updates. To prove the effectiveness of the pose-conditioned dataset update strategy, we experiment with another scheme in which the dataset used to refine the NeRF is updated with random views. We present the results of this experiment in Fig. 5-c. We observe that, with view-random updates, objects in different views can converge into inconsistent poses (e.g., the hats in the left and right view are generated in different poses).", + "bbox": [ + 496, + 477, + 893, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The above inconsistency commonly happens with many text-to-3D NeRF generation or editing methods. We hypothesize this phenomenon as follows. Initially the diffusion model can generate the object in inconsistent poses across views due to little 3D-aware hint available. This inconsistency is continually introduced to the NeRF updating and then, in return, passed partially as input to the diffusion model (as in Eq. 6), making further pose divergence eventually. Without considering nearby-views, the diffusion model cannot fix this inconsistency as objects generated on individual views match well their given text prompts.", + "bbox": [ + 496, + 598, + 893, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed pose-conditioned dataset update strategy simulates the nature of NeRF construction, in which a view rendered closed to already-used views should contain similar but slightly blurry and distorted object content. Pose-conditioned view arrangement thus gives the diffusion model enough hints to generate objects with view-consistent appearance and poses, specified in nearby views. View-consistency is thus achieved progressively in the training dataset and is integrated into the constructed NeRF (see Fig. 5-e).", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5182", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89ffbaa2f17dac027e3910a281c3160d38f308d311280cdf06985953c9dde8d8.jpg", + "image_caption": [ + "(a) Fine-tuning of the diffusion model without using object images" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b52ea12b2c4b0590e356c944e1ddf15f2ab193f2255d38c8e4ebcd84efa6ab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 183, + 893, + 263 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aadfa0bfdb4792f546191664180902d62e6a2a9c7f9613efeb8923906798521.jpg", + "image_caption": [ + "(b) Fine-tuning of the diffusion model without using background images" + ], + "image_footnote": [], + "bbox": [ + 80, + 277, + 893, + 358 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c488f96364b847254e7408a119a8342ea5908ec5ffd4fab9e633e81864f72ff9.jpg", + "image_caption": [ + "(c) Dataset updates with random poses (without pose-conditioned updates)", + "(d) Without periodic dataset updates" + ], + "image_footnote": [], + "bbox": [ + 80, + 372, + 893, + 454 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95cd1e08b02da26e3393e9d9b06ac1405b1d31e15d6171b0a56bbd513feebaad.jpg", + "image_caption": [ + "(e) Full pipeline", + "Figure 5. Ablation study results. Each row shows the results of a variant of our pipeline. The left and right columns include images of two different views of an edited scene. For each column, from left to right are the results of increasing training steps, where the most left image in each column is an early-stage result and the most right is the final output. The left column is a view near the starting view, which converges faster than the right column from a farther view (except for the variant in (c) where both views converge equally fast)." + ], + "image_footnote": [], + "bbox": [ + 80, + 467, + 893, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Periodic dataset updates. We observe that periodically updating of training views during the NeRF updating is important to achieve high-quality renderings. We verify this by keeping all the training views from the pose-conditioned dataset updates fixed during the NeRF updating. We found that although objects are rendered at fairly accurate orientations, their texture and geometry still suffer from inconsistencies (see Fig. 5-d). Having periodically updates on the training data fixes these minor defects and facilitates a more consistent convergence of the NeRF updating (see Fig. 5-e).", + "bbox": [ + 75, + 656, + 472, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 827, + 194, + 842 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a new language-driven method for object manipulation in NeRFs. Our method is built upon a novel idea of joint 2D-3D interaction, keeping both 2D image synthesis", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and 3D NeRF reconstruction in a loop. This idea is enabled by an advanced text-to-image diffusion technique that generates object-blended background images, and a novel pose-conditioned dataset update strategy that learns a NeRF from the multi-view images in a progressive manner.", + "bbox": [ + 496, + 656, + 893, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method is not without limitations. Since our 2D views are synthesized by a diffusion model, we may share the flickering problem with diffusion-based video editing methods [26, 48, 68]. We leave this for future work, which can potentially be addressed by robust video translation methods [73]. It is also of great interest to postulate a theoretical foundation for pose-conditioned dataset update to better understand the convergence of NeRF training in scene editing. Acknowledgment. This work was partially supported by a grant from the RGC of HKSAR, China (Project No. HKUST 16202323) and an internal grant from HKUST (R9429).", + "bbox": [ + 496, + 733, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5183", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 90, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 5", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2", + "[5] Ronald A Castellino. Computer aided detection (cad): an overview. Cancer Imaging, 5(1):17, 2005. 1", + "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2", + "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[9] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 1", + "[10] Epic Games. Unreal engine. 1", + "[11] SM Ali Eslami, Danilo Jimenez Rezende, Frederic Besse, Fabio Viola, Ari S Morcos, Marta Garnelo, Avraham Ruderman, Andrei A Rusu, Ivo Danihelka, Karol Gregor, et al. Neural scene representation and rendering. Science, 360 (6394):1204-1210, 2018. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2", + "[13] John K Haas. A history of the unity game engine. 2014. 1", + "[14] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 5, 6, 7", + "[15] Jeffrey Harper. Mastering Autodesk 3ds Max 2013. John Wiley & Sons, 2012. 1", + "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[18] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multiview stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2", + "[19] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2", + "[20] Yi-Hua Huang, Yan-Pei Cao, Yu-Kun Lai, Ying Shan, and Lin Gao. Nerf-texture: Texture synthesis with neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2", + "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[23] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 2", + "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2", + "[26] Gyeongman Kim, Hajin Shim, Hyunsu Kim, Yunjey Choi, Junho Kim, and Eunho Yang. Diffusion video autoencoders: Toward temporally consistent face video editing via disentangled video encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6091-6100, 2023. 8", + "[27] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. arXiv preprint arXiv:2302.06608, 2023. 2", + "[28] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation." + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5184", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Advances in Neural Information Processing Systems, 35: 23311-23330, 2022. 2", + "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 2", + "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2", + "[31] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 2", + "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3", + "[33] Miaomiao Liu, Xuming He, and Mathieu Salzmann. Geometry-aware deep network for single-image novel view synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4616-4624, 2018. 2", + "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5,6,7", + "[35] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 2", + "[36] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2", + "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2", + "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8446-8455, 2023. 3", + "[39] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2", + "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5", + "[41] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2", + "[42] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 2", + "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 4", + "[44] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2", + "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2", + "[46] Hong-Wing Pang, Binh-Son Hua, and Sai-Kit Yeung. Locally stylized neural radiance fields. In ICCV, 2023. 2", + "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 5", + "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 8", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 5, 6, 7", + "[52] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5185", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2", + "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 6, 7", + "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4, 5, 6, 7", + "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 2", + "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2", + "[58] Hyeonseop Song, Seokhun Choi, Hoseok Do, Chul Lee, and Taehyeong Kim. Blending-nerf: Text-driven localized editing in neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14383–14393, 2023. 2", + "[59] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 2", + "[60] Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Multiview consistency as supervisory signal for learning shape and pose prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2897-2905, 2018. 2", + "[61] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2", + "[62] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[63] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094, 2023. 2", + "[64] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 5", + "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2", + "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3", + "[67] Silvan Weder, Guillermo Garcia-Hernando, Aron Monszpart, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023. 2", + "[68] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 8", + "[69] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 2", + "[70] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2", + "[71] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20908–20918, 2023. 3", + "[72] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4402-4412, 2023. 2", + "[73] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. *Rerender a video: Zero-shot text-guided video-to-video translation.* arXiv preprint arXiv:2306.07954, 2023. 8", + "[74] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2", + "[75] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2", + "[76] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[77] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5186", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2", + "[79] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2", + "[80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2", + "[81] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2" + ], + "bbox": [ + 78, + 90, + 470, + 359 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5187", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_model.json b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7952f9808d028ac63388ee7e156802addcda7e06 --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_model.json @@ -0,0 +1,2754 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.79, + 0.177 + ], + "angle": 0, + "content": "Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.215, + 0.852, + 0.235 + ], + "angle": 0, + "content": "Ka Chun Shum¹ Jaeyeon Kim¹ Binh-Son Hua²,⁴ Duc Thanh Nguyen³ Sai-Kit Yeung¹" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.246, + 0.892, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\)Hong Kong University of Science and Technology \\(^{2}\\)VinAI \\(^{3}\\)Deakin University \\(^{4}\\)Trinity College Dublin" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.616 + ], + "angle": 0, + "content": "Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.647, + 0.208, + 0.66 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Editing of 3D scenes by insertion or removal of objects has been a fundamental task in computer graphics and computer vision. The task has often been performed using traditional 3D scene authoring tools [5, 9]. For example, to insert an object into a 3D scene, traditional approach requires user to manually select the object and position it into the scene. This manual pipeline has been adopted in a wide range of applications, such as furniture arrangement in interior design [15] and asset creation in game development [10, 13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Recent advances in deep learning have opened new directions to scene modeling. Neural radiance field (NeRF) [40] is a pioneer to render view-consistent photo-realistic images using neural networks. Generative models such as generative adversarial networks (GANs) [12] and diffusion models [17] learn to output photo-realistic images from un" + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.315, + 0.82, + 0.326 + ], + "angle": 0, + "content": "Input: multi-view object images & object text prompt" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.327, + 0.879, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.431, + 0.848, + 0.442 + ], + "angle": 0, + "content": "Input: multi-view background images & background text prompt" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.443, + 0.877, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.527, + 0.845, + 0.538 + ], + "angle": 0, + "content": "Output: multi-view object-blended images & novel view synthesis" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.539, + 0.88, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.661, + 0.894, + 0.732 + ], + "angle": 0, + "content": "Figure 1. Object insertion. We propose a language-driven method for view-consistent 3D object insertion into a background NeRF scene. Given an object defined by a set of multi-view images, our method generates plausible text-guided insertion results that require geometry manipulation of the original background NeRF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.809 + ], + "angle": 0, + "content": "constrained image collections. Recent text-guided diffusion models [50, 51, 53] have shown great promise in generating high-quality and diverse images from a single text prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Such successes inspire us to revisit 3D scene editing via language-driven image synthesis techniques. Particularly, for object insertion (see Fig. 1), our method performs view-consistent edits by taking as input a set of multi-view object images, multi-view background images, and a target text prompt. We utilize a text-to-image diffusion model to con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.349 + ], + "angle": 0, + "content": "tinueously synthesize multi-view images containing both a target object and a background during NeRF training. These synthesized images can be iteratively used to refine the NeRF of the background to learn geometric and appearance features of the object. This approach to refining a radiance field is also known as dataset update [14]. However, a particular challenge is that the refinement process may result in unstable NeRF training, degrading its rendering quality due to inconsistent views synthesized. To address this issue, we propose a pose-conditioned dataset update strategy that gradually engages the target object into the background, beginning at a randomly selected pose (view), then views close to the already-used views before propagating to views further away. We observe that this pose-conditioned strategy significantly improves the NeRF learning, reducing rendering artifacts and maintaining view-consistent rendering. In summary, we make the following contributions in our work." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.351, + 0.472, + 0.44 + ], + "angle": 0, + "content": "- We propose a new framework for object manipulation in NeRFs via text-to-image diffusion. Our work can create high-quality 3D scenes from simple inputs (text prompts and multi-view images). It, therefore, has the potential for conveniently building a variant of 3D object libraries that is applicable for the scene editing task." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.442, + 0.472, + 0.547 + ], + "angle": 0, + "content": "- We propose a pose-conditioned dataset update strategy that stabilizes the fusion of objects into a background NeRF, enabling view-consistent rendering. Our method requires no priors on the geometry or texture of the objects, unlike existing works relying on accurate meshes [7, 20, 31, 81], depth maps [42, 67], fine masks [58, 63], semantics [1], or lighting assumption [69, 78]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.548, + 0.472, + 0.637 + ], + "angle": 0, + "content": "- We showcase our method in two case studies: object insertion and object removal via a user-friendly manner. Specifically, we use a 3D bounding box to define the location for object insertion/removal. Box-based location requires only a rough orientation, which is easy to be adjusted and visualized using real-time NeRF GUIs [43, 59]." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.638, + 0.472, + 0.682 + ], + "angle": 0, + "content": "- We conduct extensive experiments to validate key techniques of our method and demonstrate its state-of-the-art performance in NeRF editing." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.351, + 0.472, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.216, + 0.71 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.72, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Image synthesis. Early attempts in deep learning-based image synthesis have utilized generative models such as GANs [12]. Several methods combine supervised losses with adversarial losses to learn class-conditional GANs [41] and image-to-image translations [21, 65, 80]. StyleGAN [22] and its variants [23, 24] perform image editing directly on latent representations, but such editorings are not intuitive and thus difficult to control the output." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Recently, diffusion models [17] have made significant progress in image synthesis with high-quality data samples constructed from sophisticated forward and denoising diffusion steps. In addition, vision-language models such as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.213 + ], + "angle": 0, + "content": "CLIP [49] have made text prompts an intuitive condition to generate and edit images. These developments have led to text-guided diffusion models [50, 51, 53] which offer excellent synthesis quality. Downstream applications then can be built by fine-tuning these text-guided models to fit with the application domains [4, 25, 36, 52, 70, 77]. Our method follows this streamline where we use a text-to-image diffusion model [52] to guide the fusion of objects into a background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.897, + 0.445 + ], + "angle": 0, + "content": "Neural 3D modeling. Image-based 3D modeling methods use convolutional neural networks (CNNs) [29] to learn 3D structures from multiple images [11, 18, 33, 60, 79]. However, CNNs often struggle to deal with complex shapes, texture, and lighting captured in the images. Follow-up works integrate differentiable rendering and represent 3D structures as neural surfaces [39] or shapes [8]. 3D-aware GANs integrates volume rendering into their generators to synthesize novel views from a single image [6, 45, 56, 57]. NeRFs [40] apply the same neural rendering approach, but are optimized on a ray rendering loss on multi-view input images. There are methods addressing the limitations of NeRFs in various aspects such as visual artifacts [2], data complexity [37], camera poses [74], and computational efficiency [43, 76]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.644 + ], + "angle": 0, + "content": "NeRF editing. NeRF editing has often been carried out by parameter tuning [35], layer feature fusion [61], or deformable rays [75]. An alternative to editing a NeRF is to directly amend the multi-view images used to learn it. For example, NeRF stylization methods [19, 44, 46, 62] freeze the geometry branch and optimize the color branch in a NeRF to stylize multi-view images. Several methods attempt to decompose an existing NeRF, which in turn holds color information and separates voxels considering multi-view masks [30] or semantics [27, 28]. NeRF inpainting fills simple unseen background geometry and colors with help of depth priors [42] or filtering inpainted multi-views [67]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.896, + 0.854 + ], + "angle": 0, + "content": "Creation of complex geometry and vivid colors for a NeRF is challenging due to higher-level requirements of consistency. DiscoScene [72] fuses background and object NeRFs, thus is unable to condition customized contents. FocalDreamer [31] and DreamEditor [81] rely on fine meshes to function, disregarding the advantage of multi-view representation of NeRFs. Our method is perhaps most similar to Instruct-NeRF2NeRF [14] which shares a related data updating schema. However, Instruct-NeRF2NeRF [14] always generates geometry aligned with the original geometry (e.g., transfer a sneaker to a sneaker-shape apple), thus fails in most cases of object insertion or removals. Besides, it requires heavy retraining of a diffusion model on large-scale self-constructed datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Recently, text-to-image diffusion has been applied to generate 3D contents. DreamFusion [47] introduced a score distillation sampling (SDS) loss that progressively consolidates" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5177" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Figure 2. Overview of our pipeline. We customize and fine-tune a text-to-image diffusion model for view synthesis in an inpainting manner (left). We then apply the model to progressively fuse an object into background views to update a background NeRF (right). The process of view synthesis and NeRF updating is performed repeatedly. Views generated by the diffusion model are added to an on-going dataset to strengthen the NeRF. In return, the NeRF renders color hints for the diffusion model to create new views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.475, + 0.507 + ], + "angle": 0, + "content": "view information from a diffusion model into NeRFs. This loss has been applied with additional treatments to image resolution [32], conditional images [38], photo-realism [66], or scene geometry [71]. We also adopt the SDS loss but develop a novel training schedule to address challenges in multi-view object and background fusion for NeRF editing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.169, + 0.535 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.188, + 0.56 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.764 + ], + "angle": 0, + "content": "For the ease of presentation, we first describe our method for object insertion. We then extend it to object removal. An overview of our method for object insertion is illustrated in Fig. 2. Here we aim to insert an object into a background NeRF in two steps. In the first step, we synthesize training views for the NeRF with the target object embedded in the background. In the second step, the NeRF is updated with the synthesized views to learn the geometry and appearance of the object. There are two key challenges in this approach. First, background preserving is required in the synthesized views for the NeRF updating. Second, image synthesis may generate view-inconsistent images, causing artifacts in the resulting NeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.886 + ], + "angle": 0, + "content": "To address the first challenge (i.e., object-blended image synthesis with background preserving), we leverage a state-of-the-art diffusion model to image synthesis, and opt to customize the model with text prompts for object blending (see Sec. 3.2). We formulate this task as image inpainting where we place a binary mask on each background image to indicate the location where the object is inserted in, and adjust the text prompts with both the object and background." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "To achieve view-consistent renderings, we propose a new" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.895, + 0.673 + ], + "angle": 0, + "content": "strategy, namely pose-conditioned dataset updates, to schedule the data used in the NeRF updating (see Sec. 3.3). Our strategy is inspired by an important observation about the nature of NeRF: a view rendered by a NeRF maintains an extent of pose-aware color information from nearby already-used views, the nearer the more noticeable. Therefore, if we pass nearby renderings to the diffusion model with properly controlled noise, view-consistent results can be generated based on the learned color hints. From such an observation, we design a novel data scheduler for our NeRF updating. We initially train the NeRF using regular method on a dataset of multi-view background images. We then progressively fuse the object into the NeRF by iteratively updating the dataset with object-blended background images in a pose-ordered manner, i.e., views are sorted in such a way that new views are acquired nearby already-used views. Our method can also be adapted to implement object removal (see Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.681, + 0.78, + 0.699 + ], + "angle": 0, + "content": "3.2. Object-blended image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.78 + ], + "angle": 0, + "content": "We present our target object and background by a set of multi-view object images \\(\\mathbf{I}_o\\) and background images \\(\\mathbf{I}_b\\). Our goal is to build an image synthesis model that can blend the object (from \\(\\mathbf{I}_o\\)) into the background (from \\(\\mathbf{I}_b\\)) at custom locations specified by binary masks \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Let \\(D_{\\theta}\\) (with parameters \\(\\theta\\)) be such an image synthesis model. Here we adopt a pre-trained Stable Diffusion [51] to implement \\(D_{\\theta}\\). The model \\(D_{\\theta}\\) includes a denoising model \\(\\epsilon_{\\theta}\\) that learns the noise component \\(\\epsilon_{\\alpha} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{1})\\) added in the diffusion process, where \\(\\alpha \\in [0, 1]\\) is a denoising strength. We customize \\(D_{\\theta}\\) with our background and object images in an inpainting fashion. Specifically, for each image \\(I \\in \\mathbf{I}_b \\cup \\mathbf{I}_o\\), we make a masked-out (background-preserved)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5178" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "version \\(I\\odot M\\) where \\(M\\sim \\mathcal{U}\\) is a random square binary mask whose coordinates are sampled from a uniform distribution \\(\\mathcal{U}\\) and \\(\\odot\\) is a pixel-wise product. The image \\(I\\) is associated with a text prompt \\(p\\) where we use \\(\" ^ { \\text{丰} }\\) to indicate our object/background identifiers [52], e.g., \\(\" *\\) sneaker\" means that our target object is a sneaker and \\(\" *\\) table\" is our target background. We fine-tune \\(D_{\\theta}\\) with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.224, + 0.471, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s y n t h e s i s}} (\\theta) = \\mathbb {E} _ {I, M \\sim \\mathcal {U}, p, \\epsilon_ {\\alpha} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} \\left[ \\| \\epsilon_ {\\alpha} - \\epsilon_ {\\theta} (I) \\| _ {2} ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.471, + 0.299 + ], + "angle": 0, + "content": "Eq. 1 represents a general form for the fine-tuning process. Intermediate steps, e.g., input formation, latent encoding-decoding, are presented in our supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.47, + 0.377 + ], + "angle": 0, + "content": "We fine-tune the diffusion model \\( D_{\\theta} \\) with all the images in \\( \\mathbf{I}_b \\) and \\( \\mathbf{I}_o \\) for \\( n_{bg} \\) and \\( n_{obj} \\) times. After fine-tuning, we can use \\( D_{\\theta} \\) to synthesize images \\( \\hat{I}_b \\) that blend the object presented in \\( \\mathbf{I}_o \\) into the background images \\( I_b \\in \\mathbf{I}_b \\) via a combined text prompt \\( \\tilde{p} \\), e.g., \"a * sneaker on a * table\"," + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.388, + 0.47, + 0.406 + ], + "angle": 0, + "content": "\\[\nD _ {\\theta} \\left(I _ {b} \\odot M, \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.47, + 0.464 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is set manually to control how much object content in the masked area to be inpainted by \\(D_{\\theta}\\) (we demonstrate how we set \\(\\alpha\\) for different purposes in Sec. 3.3)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.473, + 0.37, + 0.489 + ], + "angle": 0, + "content": "3.3. Pose-conditioned dataset updates" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.47, + 0.572 + ], + "angle": 0, + "content": "Let \\( R_{\\phi} \\) (with parameters \\( \\phi \\)) be our NeRF, which is initially trained with the multi-view background images in \\( \\mathbf{I}_b \\) using an existing approach [43]. We also retrieve a set of camera poses \\( \\{\\pi(I_b)\\} \\) (\\( 4 \\times 4 \\) matrices) for the images \\( I_b \\in \\mathbf{I}_b \\) using the pose estimation method COLMAP [54, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.471, + 0.713 + ], + "angle": 0, + "content": "We propose to update the NeRF \\( R_{\\phi} \\) progressively with object-blended background images generated by the finetuned diffusion model \\( D_{\\theta} \\). Our idea is to build a progressive multi-view and pose-conditioned object-blended image dataset \\( \\hat{\\mathbf{I}}_b \\) from \\( \\mathbf{I}_b \\) to be used to update \\( R_{\\phi} \\). We initialize \\( \\hat{\\mathbf{I}}_b^{(0)} = \\emptyset \\) and \\( \\mathbf{I}_b^{(0)} = \\mathbf{I}_b \\). We start with a random background image \\( I_b \\in \\mathbf{I}_b \\). Let \\( B \\) be a 3D bounding box, and \\( B(I_b) \\) be the projection mask of \\( B \\) onto \\( I_b \\) using the camera pose \\( \\pi (I_b) \\). We generate an object-blended image \\( \\hat{I}_b \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.723, + 0.47, + 0.741 + ], + "angle": 0, + "content": "\\[\nD _ {\\theta} \\left(I _ {b} \\odot B \\left(I _ {b}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.753, + 0.47, + 0.828 + ], + "angle": 0, + "content": "where we set \\(\\alpha = 1\\) (maximum strength) to blend the object into the masked area by using only the knowledge that \\(D_{\\theta}\\) learns from \\(\\tilde{p}\\) (during the fine-tuning), and colors from the background image \\(I_{b}\\) as there is no existing color clue in the masked area \\(I_{b} \\odot B(I_{b})\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We update \\(\\hat{\\mathbf{I}}_b^{(1)} = \\hat{\\mathbf{I}}_b^{(0)}\\cup \\{\\hat{I}_b\\}\\) and \\(\\mathbf{I}_b^{(1)} = \\mathbf{I}_b^{(0)}\\setminus \\{I_b\\}\\). We then update \\(R_{\\phi}\\) using \\(\\hat{\\mathbf{I}}_b^{(1)}\\), and keep doing so by adding images into \\(\\hat{\\mathbf{I}}_b\\) sequentially. In particular, let \\(\\hat{\\mathbf{I}}_b^{(n - 1)}\\) be the dataset at the \\((n - 1)\\)-th step. The next image \\(I_b^{(n)}\\in \\mathbf{I}_b\\) for" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "step \\( n \\) is selected so as it is closest (in terms of the camera poses) to already-used images," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.13, + 0.894, + 0.164 + ], + "angle": 0, + "content": "\\[\nI _ {b} ^ {(n)} = \\underset {I _ {b, k} \\in \\mathbf {I} _ {b} ^ {(n)}} {\\arg \\min } \\underset {I _ {b, j} \\in \\mathbf {I} _ {b} ^ {(n - 1)}} {\\min } \\| \\pi_ {T} (I _ {b, k}) - \\pi_ {T} (I _ {b, j}) \\| ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.895, + 0.246 + ], + "angle": 0, + "content": "where \\(\\pi_T\\) is the translation component of the pose \\(\\pi\\). This view selection reflects the standard multi-view reconstruction pipeline where the multi-view data are collected with a smoothly connected, inward-surrounding camera trajectory. This requirement is also fulfilled by most NeRF datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.248, + 0.894, + 0.299 + ], + "angle": 0, + "content": "Given a background image \\( I_b^{(n)} \\) in Eq. 4, to utilize the nearby color hints learned by \\( R_{\\phi} \\), we create a background-preserved foreground-rendered image \\( \\tilde{I}_b^{(n)} \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.305, + 0.892, + 0.327 + ], + "angle": 0, + "content": "\\[\n\\tilde {I} _ {b} ^ {(n)} = \\left(I _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right)\\right) \\oplus \\left(R _ {\\phi} \\left(\\pi \\left(I _ {b} ^ {(n)}\\right)\\right) \\odot \\bar {B} \\left(I _ {b} ^ {(n)}\\right)\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.894, + 0.401 + ], + "angle": 0, + "content": "where \\( R_{\\phi}(\\pi(I_b^{(n)})) \\) is the rendering result of \\( R_{\\phi} \\) in the pose \\( \\pi(I_b^{(n)}) \\) (for the background image \\( I_b^{(n)} \\)), \\( \\bar{B} \\) is the complement of \\( B \\) (as we want to maintain the foreground rendered by \\( R_{\\phi} \\)), and \\( \\oplus \\) is a pixel-wise addition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.894, + 0.438 + ], + "angle": 0, + "content": "We then generate a view-consistent image \\(\\hat{I}_b^{(n)}\\) by applying \\(D_{\\theta}\\) to \\(\\tilde{I}_b^{(n)}\\) defined in Eq. 5 as," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.446, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\nD _ {\\theta} \\left(\\tilde {I} _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} ^ {(n)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.892, + 0.542 + ], + "angle": 0, + "content": "where we set \\(\\alpha\\) to a low value to allow \\(D_{\\theta}\\) to utilize color hints from previous nearby views, provided from the rendering results \\(R_{\\phi}(\\pi (I_b^{(n)}))\\). We empirically found that \\(\\alpha \\in [0.3, 0.4]\\) gives best view-consistent rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.894, + 0.574 + ], + "angle": 0, + "content": "Again, we update \\(\\hat{\\mathbf{I}}_b^{(n)}\\) with \\(\\hat{I}_b^{(n)}\\), and update \\(R_{\\phi}\\) accordingly by minimizing the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.581, + 0.892, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N e R F}} (\\phi) = \\mathbb {E} _ {\\hat {I} _ {b} \\in \\hat {\\mathbf {I}} _ {b} ^ {(n)}} \\left[ \\| R _ {\\phi} \\left(\\pi \\left(\\hat {I} _ {b}\\right)\\right) - \\hat {I} _ {b} \\| ^ {2} \\right]. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.611, + 0.892, + 0.705 + ], + "angle": 0, + "content": "During the updating process, we include \\( n_{near} \\) views into the ongoing dataset for every \\( n_{new} \\) NeRF updating steps. In addition, we periodically replace each old view by a new one using Eqs. 5 and 6 for every \\( n_{old} \\) NeRF updating steps. The updating procedure is completed once all the background images in \\( \\mathbf{I}_b \\) have been processed, i.e., \\( \\mathbf{I}_b^{(n)} = \\emptyset \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.746, + 0.728 + ], + "angle": 0, + "content": "3.4. Adapting to object removal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our framework can be adapted to object removal. In particular, we also fine-tune the diffusion model \\( D_{\\theta} \\) as in Sec. 3.2. However, we do not fine-tune \\( D_{\\theta} \\) with object images as we want to remove objects. Instead, we first individually inpaint all background images on projection masks using background prompts without identifier (e.g., \"a table\"). These inpainted images are then treated as pseudo ground-truth and used to customize \\( D_{\\theta} \\) with the background text prompts containing identifiers (e.g., \"a * table\"). The pseudo ground-truth backgrounds are visually pleasing but still remain cross-view inconsistencies. To circumvent this issue, we perform" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5179" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "NeRF updating using dataset updates as in Sec. 3.3. However, only background prompts are used during the NeRF updating. We found that this joint 2D-3D interaction gradually transforms disruptive inpainted image regions into view-consistent background images. We show the necessity of the pseudo ground-truth and the NeRF updating in making view-consistent object removal in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.211, + 0.21, + 0.228 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.236, + 0.18, + 0.251 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.473, + 0.38 + ], + "angle": 0, + "content": "For object insertion, we propose a dataset comprising multiview images of 8 backgrounds and 10 objects. The number of images for each background and object ranges from 60 to 100 and 40 to 80, respectively. We collected the images using an iPhone camera and resized the images to \\(512 \\times 512\\) resolution for the ease of training. Except for our self-captured data, we also test on synthetically rendered multiview images [40, 47]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.381, + 0.472, + 0.442 + ], + "angle": 0, + "content": "For object removal, we run experiments on the commonly used inpainting datasets from Mip-NeRF-360 [3] and IBR-Net [64]. We centrally cropped images in those datasets to make \\(512 \\times 512\\) images to fit with our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.186, + 0.466 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.535 + ], + "angle": 0, + "content": "For comparisons, we select SOTA baselines from different branches of work that can perform object insertion. To ensure the fairness, we barely modify the baselines, only when necessary to fit them with our settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.472, + 0.626 + ], + "angle": 0, + "content": "Traditional 3D editing pipeline. This pipeline creates a 3D model from multi-view images. Scene editing is then performed directly on the 3D model. To simulate this traditional pipeline, we use COLMAP [54, 55] to reconstruct the mesh for the background and object, and then manually crop and place the object mesh into the background mesh." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.472, + 0.687 + ], + "angle": 0, + "content": "Image inpainting. We adopt the inpainting variant of Stable Diffusion [51] as a baseline. However, since image inpainting treats each view independently, for a fair comparison, we perform single-view inpainting on each background image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.472, + 0.747 + ], + "angle": 0, + "content": "Single-image-to-3D NeRF. NeRF-based view synthesis partially fulfills scene editing. Here we select Zero123 [34], a SOTA that uses a distillation prior [47] from a 2D diffusion model to synthesize novel views for comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.472, + 0.822 + ], + "angle": 0, + "content": "Instruct-NeRF2NeRF [14]. This work also applies dataset updates for NeRF training. However, it fails to add/remove objects with noticeable non-uniform appearance. To test this method, we re-format text prompts, e.g., we replace \"a * sneaker on a * table\" by \"add/make a sneaker on a table\"." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.292, + 0.848 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We fine-tune the diffusion model \\( D_{\\theta} \\) with object images \\( (n_{obj} = 5,000 \\) times) more than background images \\( (n_{bg} = 500 \\) times) as we found the model needs extra training to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "learn objects with complex geometry and texture. For NeRF training, we empirically found \\( n_{near} = 3 \\), \\( n_{new} = 500 \\), and \\( n_{old} = 10 \\) balance well the quality and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.228 + ], + "angle": 0, + "content": "We run all experiments on a Nvidia RTX 3090 GPU. Diffusion fine-tuning takes around 30 minutes. NeRF updating speed relies on the number of background images while each backpropagation takes 0.5 seconds. Inference speed of the diffusion model is 8 seconds/image. Our NeRF training costs around 2.7 times more than the standard NeRF training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.236, + 0.678, + 0.251 + ], + "angle": 0, + "content": "4.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.895, + 0.334 + ], + "angle": 0, + "content": "Object insertion. We present qualitative results of object insertion in Fig. 3. As shown, our method (Fig. 3-f) can generate plausible contents with view consistency. Moverover, the outputs of our method also match well the text prompts and the generated objects are precisely located." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.895, + 0.516 + ], + "angle": 0, + "content": "In contrast, the traditional 3D pipeline (Fig. 3-b) suffers from seamed object-background boundaries and unrealistic lighting. The geometry in obscure regions, e.g., corners, is not accurately reconstructed. The image inpainting baseline (Fig. 3-c) produces reasonable results on individual views but generates view-inconsistency. Single-image-to-3D NeRF (Fig. 3-d) fails on complex scenes. A larger camera pose shift can result in background mismatch or even content collapse. Instruct-NeRF2NeRF (Fig. 3-e) is known to be strong at stylizing existing geometry but weak at generating new geometry. Under our setting, it performs random view-consistent editing, but fails to accomplish object insertion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.894, + 0.607 + ], + "angle": 0, + "content": "Object removal. We illustrate several results of object removal in Fig. 4, which shows that our method can generate view-consistent backgrounds. We also observed that without using pseudo ground-truth background, the removal can cause gradual background collapse, which we discuss further in our ablation study in Sec. 4.6." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.615, + 0.691, + 0.631 + ], + "angle": 0, + "content": "4.5. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.73 + ], + "angle": 0, + "content": "Since there is no real-world ground-truth for scene editing, direct quantitative evaluations are not possible. Instead, we adopt the CLIP Score [16], a well-known metric to measure how well an image correlates to a target prompt in the CLIP space. Let \\( E(I) \\) and \\( E(p) \\) be the CLIP embeddings of an image \\( I \\) and a text prompt \\( p \\). The CLIP Score is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.738, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C L I P S c o r e} (I, p) = \\cos^ {+} (E (I), E (p)) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.894, + 0.84 + ], + "angle": 0, + "content": "where \\(\\cos^{+}(a,b) = \\max (0,\\cos (a,b))\\). We average the CLIPScore \\((\\hat{I}_b,\\tilde{p})\\) for all the edited images \\(\\hat{I}_b\\in \\hat{\\mathbf{I}}_b\\) paired with their target prompts \\(\\tilde{p}\\), and use this average value as a performance metric for scene editing; higher score means better performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We also use the CLIP Directional Consistency (CLIPDC) in [14] to measure the editing quality and consistency across views in the CLIP space. This metric relies on an assumption that the difference in the CLIP space between a background" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5180" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.087, + 0.895, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.722, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of object insertion. Inputs include multi-view object/background images, a 3D bounding box where the object is inserted in, and a text prompt. Note that some baselines use parts of the inputs due to the nature of their techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.775, + 0.473, + 0.868 + ], + "angle": 0, + "content": "image \\( I_{b} \\) and its edited version \\( \\hat{I}_{b} \\) should match with the difference between the original prompt \\( p \\) (paired with \\( I_{b} \\)) and the customized prompt \\( \\tilde{p} \\) (paired with \\( \\hat{I}_b \\)). Moreover, a good editing should make consistent differences between every pair of \\( I_{b} \\) and \\( \\hat{I}_b \\), especially for adjacent views. We follow [14] to calculate the CLIPDC between two adjacent" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.773, + 0.71, + 0.794 + ], + "angle": 0, + "content": "edited views \\(\\hat{I}_b^{(i)}\\) and \\(\\hat{I}_b^{(i + 1)}\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.801, + 0.892, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {C L I P D C} \\left(\\hat {I} _ {b} ^ {(i)}, \\hat {I} _ {b} ^ {(i + 1)}\\right) = \\\\ \\cos^ {+} (E (\\tilde {p}) - E (p), E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)})) \\times \\tag {9} \\\\ \\cos^ {+} \\left(E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)}), E (\\hat {I} _ {b} ^ {(i + 1)}) - E (I _ {b} ^ {(i + 1)})\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We measure the CLIPDC of an edited scene by averaging" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.207, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.089, + 0.337, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.469, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.189, + 0.206, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.189, + 0.337, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.189, + 0.469, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.293, + 0.206, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.293, + 0.337, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.293, + 0.469, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.393, + 0.206, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.393, + 0.337, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.393, + 0.469, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.493, + 0.206, + 0.517 + ], + "angle": 0, + "content": "(a) Original BG with object bounding box" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.493, + 0.337, + 0.517 + ], + "angle": 0, + "content": "(b) Pseudo ground-truth BG" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.493, + 0.47, + 0.517 + ], + "angle": 0, + "content": "(c) View-consistent editing result" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of object removal where we show the importance of pseudo ground-truth background (BG) in generating view-consistent editing results." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.599, + 0.47, + 0.713 + ], + "angle": 0, + "content": "
MethodCLIPScore ↑CLIPDC ↑
Traditional 3D [54, 55]0.26480.1169
Inpainting [51]0.26830.1325
Zero123 [34]0.21970.0361
Instruct-NeRF2NeRF [14]0.23470.0263
Ours0.27430.1642
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.723, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Table 1. Quantitative results. Higher CLIP metrics scores indicate higher image editing quality or consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.47, + 0.901 + ], + "angle": 0, + "content": "the CLIPDC \\((\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})\\) for all the adjacent image pairs \\((\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})\\) in the scene. We report both the CLIPScore and CLIPDC of our method and other baselines on 35 unique edits (cross editing on the more generalizable 5 scenes \\(\\times\\) 7 objects) in Tab. 1. Experimental show that our method outperforms all the baselines on both the CLIPScore and CLIPDC metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.655, + 0.108 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.311 + ], + "angle": 0, + "content": "In this ablation study, we validate the effectiveness of technical components in our method. We suggest readers observe the results in Fig. 5, which shows two views of an edited scene. We provide detailed analyses on these results below. Diffusion model fine-tuning. Recall that we fine-tune the diffusion model \\( D_{\\theta} \\) on both object and background images (in Sec. 3.2). Here we prove that such a fine-tuning is necessary. Fine-tuning on object images makes the diffusion model aware of the same object during the dataset updates (in Sec. 3.3). We verify this in Fig. 5-a, where we skip the fine-tuning of \\( D_{\\theta} \\) on object images. As shown, without using object images, the diffusion model cannot generate the same object across views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Likewise, fine-tuning the diffusion model with background images helps to preserve the background at inpainted borders (areas in between mask boundary and inner-mask object boundary). As shown in Fig. 5-b, without fine-tuning on background images, the model cannot inpaint the background within the mask region properly. The inpainted background gets darker at latter NeRF updating steps, and finally ends up with a noticeable error. We hypothesize this collapse that the model, without being fine-tuned on relevant background images, applies the bias from pre-learned knowledge which does not align with the surrounding background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.895, + 0.599 + ], + "angle": 0, + "content": "Pose-conditioned dataset updates. To prove the effectiveness of the pose-conditioned dataset update strategy, we experiment with another scheme in which the dataset used to refine the NeRF is updated with random views. We present the results of this experiment in Fig. 5-c. We observe that, with view-random updates, objects in different views can converge into inconsistent poses (e.g., the hats in the left and right view are generated in different poses)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.765 + ], + "angle": 0, + "content": "The above inconsistency commonly happens with many text-to-3D NeRF generation or editing methods. We hypothesize this phenomenon as follows. Initially the diffusion model can generate the object in inconsistent poses across views due to little 3D-aware hint available. This inconsistency is continually introduced to the NeRF updating and then, in return, passed partially as input to the diffusion model (as in Eq. 6), making further pose divergence eventually. Without considering nearby-views, the diffusion model cannot fix this inconsistency as objects generated on individual views match well their given text prompts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Our proposed pose-conditioned dataset update strategy simulates the nature of NeRF construction, in which a view rendered closed to already-used views should contain similar but slightly blurry and distorted object content. Pose-conditioned view arrangement thus gives the diffusion model enough hints to generate objects with view-consistent appearance and poses, specified in nearby views. View-consistency is thus achieved progressively in the training dataset and is integrated into the constructed NeRF (see Fig. 5-e)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.171, + 0.66, + 0.183 + ], + "angle": 0, + "content": "(a) Fine-tuning of the diffusion model without using object images" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.184, + 0.894, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.266, + 0.676, + 0.278 + ], + "angle": 0, + "content": "(b) Fine-tuning of the diffusion model without using background images" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.278, + 0.894, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.361, + 0.681, + 0.373 + ], + "angle": 0, + "content": "(c) Dataset updates with random poses (without pose-conditioned updates)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.373, + 0.894, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.456, + 0.581, + 0.467 + ], + "angle": 0, + "content": "(d) Without periodic dataset updates" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.468, + 0.894, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.551, + 0.528, + 0.563 + ], + "angle": 0, + "content": "(e) Full pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.574, + 0.895, + 0.631 + ], + "angle": 0, + "content": "Figure 5. Ablation study results. Each row shows the results of a variant of our pipeline. The left and right columns include images of two different views of an edited scene. For each column, from left to right are the results of increasing training steps, where the most left image in each column is an early-stage result and the most right is the final output. The left column is a view near the starting view, which converges faster than the right column from a farther view (except for the variant in (c) where both views converge equally fast)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Periodic dataset updates. We observe that periodically updating of training views during the NeRF updating is important to achieve high-quality renderings. We verify this by keeping all the training views from the pose-conditioned dataset updates fixed during the NeRF updating. We found that although objects are rendered at fairly accurate orientations, their texture and geometry still suffer from inconsistencies (see Fig. 5-d). Having periodically updates on the training data fixes these minor defects and facilitates a more consistent convergence of the NeRF updating (see Fig. 5-e)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.828, + 0.196, + 0.843 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "We propose a new language-driven method for object manipulation in NeRFs. Our method is built upon a novel idea of joint 2D-3D interaction, keeping both 2D image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.734 + ], + "angle": 0, + "content": "and 3D NeRF reconstruction in a loop. This idea is enabled by an advanced text-to-image diffusion technique that generates object-blended background images, and a novel pose-conditioned dataset update strategy that learns a NeRF from the multi-view images in a progressive manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.734, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our method is not without limitations. Since our 2D views are synthesized by a diffusion model, we may share the flickering problem with diffusion-based video editing methods [26, 48, 68]. We leave this for future work, which can potentially be addressed by robust video translation methods [73]. It is also of great interest to postulate a theoretical foundation for pose-conditioned dataset update to better understand the convergence of NeRF training in scene editing. Acknowledgment. This work was partially supported by a grant from the RGC of HKSAR, China (Project No. HKUST 16202323) and an internal grant from HKUST (R9429)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5183" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.091, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.198 + ], + "angle": 0, + "content": "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.281 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.357, + 0.471, + 0.423 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.428, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[5] Ronald A Castellino. Computer aided detection (cad): an overview. Cancer Imaging, 5(1):17, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.457, + 0.471, + 0.551 + ], + "angle": 0, + "content": "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.556, + 0.47, + 0.61 + ], + "angle": 0, + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.613, + 0.471, + 0.668 + ], + "angle": 0, + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.671, + 0.47, + 0.711 + ], + "angle": 0, + "content": "[9] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.294, + 0.727 + ], + "angle": 0, + "content": "[10] Epic Games. Unreal engine. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.471, + 0.797 + ], + "angle": 0, + "content": "[11] SM Ali Eslami, Danilo Jimenez Rezende, Frederic Besse, Fabio Viola, Ari S Morcos, Marta Garnelo, Avraham Ruderman, Andrei A Rusu, Ivo Danihelka, Karol Gregor, et al. Neural scene representation and rendering. Science, 360 (6394):1204-1210, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.458, + 0.871 + ], + "angle": 0, + "content": "[13] John K Haas. A history of the unity game engine. 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.893, + 0.119 + ], + "angle": 0, + "content": "3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Jeffrey Harper. Mastering Autodesk 3ds Max 2013. John Wiley & Sons, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.248, + 0.894, + 0.314 + ], + "angle": 0, + "content": "[18] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multiview stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.893, + 0.385 + ], + "angle": 0, + "content": "[19] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.893, + 0.441 + ], + "angle": 0, + "content": "[20] Yi-Hua Huang, Yan-Pei Cao, Yu-Kun Lai, Ying Shan, and Lin Gao. Nerf-texture: Texture synthesis with neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.893, + 0.552 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.893, + 0.608 + ], + "angle": 0, + "content": "[23] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.893, + 0.678 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.893, + 0.748 + ], + "angle": 0, + "content": "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.893, + 0.83 + ], + "angle": 0, + "content": "[26] Gyeongman Kim, Hajin Shim, Hyunsu Kim, Yunjey Choi, Junho Kim, and Eunho Yang. Diffusion video autoencoders: Toward temporally consistent face video editing via disentangled video encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6091-6100, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[27] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. arXiv preprint arXiv:2302.06608, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[28] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.471, + 0.12 + ], + "angle": 0, + "content": "Advances in Neural Information Processing Systems, 35: 23311-23330, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.472, + 0.235 + ], + "angle": 0, + "content": "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.472, + 0.292 + ], + "angle": 0, + "content": "[31] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.472, + 0.376 + ], + "angle": 0, + "content": "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[33] Miaomiao Liu, Xuming He, and Mathieu Salzmann. Geometry-aware deep network for single-image novel view synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4616-4624, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.472, + 0.503 + ], + "angle": 0, + "content": "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.472, + 0.561 + ], + "angle": 0, + "content": "[35] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[36] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8446-8455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[39] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[41] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[42] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.894, + 0.318 + ], + "angle": 0, + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[44] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.447, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[46] Hong-Wing Pang, Binh-Son Hua, and Sai-Kit Yeung. Locally stylized neural radiance fields. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.52, + 0.892, + 0.576 + ], + "angle": 0, + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.894, + 0.716 + ], + "angle": 0, + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.894, + 0.788 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[52] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[58] Hyeonseop Song, Seokhun Choi, Hoseok Do, Chul Lee, and Taehyeong Kim. Blending-nerf: Text-driven localized editing in neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14383–14393, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.435, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[59] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.472, + 0.587 + ], + "angle": 0, + "content": "[60] Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Multiview consistency as supervisory signal for learning shape and pose prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2897-2905, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.472, + 0.659 + ], + "angle": 0, + "content": "[61] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[62] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.772 + ], + "angle": 0, + "content": "[63] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[64] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.257 + ], + "angle": 0, + "content": "[67] Silvan Weder, Guillermo Garcia-Hernando, Aron Monszpart, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[68] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.894, + 0.397 + ], + "angle": 0, + "content": "[69] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.894, + 0.468 + ], + "angle": 0, + "content": "[70] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.894, + 0.551 + ], + "angle": 0, + "content": "[71] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20908–20918, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.894, + 0.649 + ], + "angle": 0, + "content": "[72] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4402-4412, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.894, + 0.691 + ], + "angle": 0, + "content": "[73] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. *Rerender a video: Zero-shot text-guided video-to-video translation.* arXiv preprint arXiv:2306.07954, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.894, + 0.748 + ], + "angle": 0, + "content": "[74] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[75] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.894, + 0.859 + ], + "angle": 0, + "content": "[76] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[77] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5186" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[78] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[79] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.248, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.472, + 0.36 + ], + "angle": 0, + "content": "[81] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5187" + } + ] +] \ No newline at end of file diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_origin.pdf b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a8f1324355d4cccd777c48dfd20d83a24eadc6f --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/af532a3c-a441-40c6-a819-aabe1e6f5939_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5722a35beca9a4152aef8b33c1165296fdcc6b4e28ebc20048da37b2f7c51fc5 +size 9971171 diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/full.md b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c4ecdda919d91b889264c5cdbe66ce629185722 --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/full.md @@ -0,0 +1,371 @@ +# Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates + +Ka Chun Shum¹ Jaeyeon Kim¹ Binh-Son Hua²,⁴ Duc Thanh Nguyen³ Sai-Kit Yeung¹ + +$^{1}$ Hong Kong University of Science and Technology $^{2}$ VinAI $^{3}$ Deakin University $^{4}$ Trinity College Dublin + +# Abstract + +Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing. + +# 1. Introduction + +Editing of 3D scenes by insertion or removal of objects has been a fundamental task in computer graphics and computer vision. The task has often been performed using traditional 3D scene authoring tools [5, 9]. For example, to insert an object into a 3D scene, traditional approach requires user to manually select the object and position it into the scene. This manual pipeline has been adopted in a wide range of applications, such as furniture arrangement in interior design [15] and asset creation in game development [10, 13]. + +Recent advances in deep learning have opened new directions to scene modeling. Neural radiance field (NeRF) [40] is a pioneer to render view-consistent photo-realistic images using neural networks. Generative models such as generative adversarial networks (GANs) [12] and diffusion models [17] learn to output photo-realistic images from un + +![](images/8b7ec97ea103525c7798b80878297a59f637096c4f2bdaa606d4b0000184f75d.jpg) +Input: multi-view object images & object text prompt + +![](images/78346f513bd46aa69d4cc1355d1cf5da1407a424aa2e5bdbc06b4a7bfa420998.jpg) +Input: multi-view background images & background text prompt + +![](images/896e69843f5988da44b044e266c7b7a9fce35bd973a0e30a35200e08444ac447.jpg) +Output: multi-view object-blended images & novel view synthesis +Figure 1. Object insertion. We propose a language-driven method for view-consistent 3D object insertion into a background NeRF scene. Given an object defined by a set of multi-view images, our method generates plausible text-guided insertion results that require geometry manipulation of the original background NeRF. + +constrained image collections. Recent text-guided diffusion models [50, 51, 53] have shown great promise in generating high-quality and diverse images from a single text prompt. + +Such successes inspire us to revisit 3D scene editing via language-driven image synthesis techniques. Particularly, for object insertion (see Fig. 1), our method performs view-consistent edits by taking as input a set of multi-view object images, multi-view background images, and a target text prompt. We utilize a text-to-image diffusion model to con + +tinueously synthesize multi-view images containing both a target object and a background during NeRF training. These synthesized images can be iteratively used to refine the NeRF of the background to learn geometric and appearance features of the object. This approach to refining a radiance field is also known as dataset update [14]. However, a particular challenge is that the refinement process may result in unstable NeRF training, degrading its rendering quality due to inconsistent views synthesized. To address this issue, we propose a pose-conditioned dataset update strategy that gradually engages the target object into the background, beginning at a randomly selected pose (view), then views close to the already-used views before propagating to views further away. We observe that this pose-conditioned strategy significantly improves the NeRF learning, reducing rendering artifacts and maintaining view-consistent rendering. In summary, we make the following contributions in our work. + +- We propose a new framework for object manipulation in NeRFs via text-to-image diffusion. Our work can create high-quality 3D scenes from simple inputs (text prompts and multi-view images). It, therefore, has the potential for conveniently building a variant of 3D object libraries that is applicable for the scene editing task. +- We propose a pose-conditioned dataset update strategy that stabilizes the fusion of objects into a background NeRF, enabling view-consistent rendering. Our method requires no priors on the geometry or texture of the objects, unlike existing works relying on accurate meshes [7, 20, 31, 81], depth maps [42, 67], fine masks [58, 63], semantics [1], or lighting assumption [69, 78]. +- We showcase our method in two case studies: object insertion and object removal via a user-friendly manner. Specifically, we use a 3D bounding box to define the location for object insertion/removal. Box-based location requires only a rough orientation, which is easy to be adjusted and visualized using real-time NeRF GUIs [43, 59]. +- We conduct extensive experiments to validate key techniques of our method and demonstrate its state-of-the-art performance in NeRF editing. + +# 2. Related work + +Image synthesis. Early attempts in deep learning-based image synthesis have utilized generative models such as GANs [12]. Several methods combine supervised losses with adversarial losses to learn class-conditional GANs [41] and image-to-image translations [21, 65, 80]. StyleGAN [22] and its variants [23, 24] perform image editing directly on latent representations, but such editorings are not intuitive and thus difficult to control the output. + +Recently, diffusion models [17] have made significant progress in image synthesis with high-quality data samples constructed from sophisticated forward and denoising diffusion steps. In addition, vision-language models such as + +CLIP [49] have made text prompts an intuitive condition to generate and edit images. These developments have led to text-guided diffusion models [50, 51, 53] which offer excellent synthesis quality. Downstream applications then can be built by fine-tuning these text-guided models to fit with the application domains [4, 25, 36, 52, 70, 77]. Our method follows this streamline where we use a text-to-image diffusion model [52] to guide the fusion of objects into a background. + +Neural 3D modeling. Image-based 3D modeling methods use convolutional neural networks (CNNs) [29] to learn 3D structures from multiple images [11, 18, 33, 60, 79]. However, CNNs often struggle to deal with complex shapes, texture, and lighting captured in the images. Follow-up works integrate differentiable rendering and represent 3D structures as neural surfaces [39] or shapes [8]. 3D-aware GANs integrates volume rendering into their generators to synthesize novel views from a single image [6, 45, 56, 57]. NeRFs [40] apply the same neural rendering approach, but are optimized on a ray rendering loss on multi-view input images. There are methods addressing the limitations of NeRFs in various aspects such as visual artifacts [2], data complexity [37], camera poses [74], and computational efficiency [43, 76]. + +NeRF editing. NeRF editing has often been carried out by parameter tuning [35], layer feature fusion [61], or deformable rays [75]. An alternative to editing a NeRF is to directly amend the multi-view images used to learn it. For example, NeRF stylization methods [19, 44, 46, 62] freeze the geometry branch and optimize the color branch in a NeRF to stylize multi-view images. Several methods attempt to decompose an existing NeRF, which in turn holds color information and separates voxels considering multi-view masks [30] or semantics [27, 28]. NeRF inpainting fills simple unseen background geometry and colors with help of depth priors [42] or filtering inpainted multi-views [67]. + +Creation of complex geometry and vivid colors for a NeRF is challenging due to higher-level requirements of consistency. DiscoScene [72] fuses background and object NeRFs, thus is unable to condition customized contents. FocalDreamer [31] and DreamEditor [81] rely on fine meshes to function, disregarding the advantage of multi-view representation of NeRFs. Our method is perhaps most similar to Instruct-NeRF2NeRF [14] which shares a related data updating schema. However, Instruct-NeRF2NeRF [14] always generates geometry aligned with the original geometry (e.g., transfer a sneaker to a sneaker-shape apple), thus fails in most cases of object insertion or removals. Besides, it requires heavy retraining of a diffusion model on large-scale self-constructed datasets. + +Recently, text-to-image diffusion has been applied to generate 3D contents. DreamFusion [47] introduced a score distillation sampling (SDS) loss that progressively consolidates + +![](images/383d36ec2e8eba0016311fdae1eedf40f9a416e5a7ee6cbc6234feeab3164338.jpg) +Figure 2. Overview of our pipeline. We customize and fine-tune a text-to-image diffusion model for view synthesis in an inpainting manner (left). We then apply the model to progressively fuse an object into background views to update a background NeRF (right). The process of view synthesis and NeRF updating is performed repeatedly. Views generated by the diffusion model are added to an on-going dataset to strengthen the NeRF. In return, the NeRF renders color hints for the diffusion model to create new views. + +view information from a diffusion model into NeRFs. This loss has been applied with additional treatments to image resolution [32], conditional images [38], photo-realism [66], or scene geometry [71]. We also adopt the SDS loss but develop a novel training schedule to address challenges in multi-view object and background fusion for NeRF editing. + +# 3. Method + +# 3.1. Overview + +For the ease of presentation, we first describe our method for object insertion. We then extend it to object removal. An overview of our method for object insertion is illustrated in Fig. 2. Here we aim to insert an object into a background NeRF in two steps. In the first step, we synthesize training views for the NeRF with the target object embedded in the background. In the second step, the NeRF is updated with the synthesized views to learn the geometry and appearance of the object. There are two key challenges in this approach. First, background preserving is required in the synthesized views for the NeRF updating. Second, image synthesis may generate view-inconsistent images, causing artifacts in the resulting NeRF. + +To address the first challenge (i.e., object-blended image synthesis with background preserving), we leverage a state-of-the-art diffusion model to image synthesis, and opt to customize the model with text prompts for object blending (see Sec. 3.2). We formulate this task as image inpainting where we place a binary mask on each background image to indicate the location where the object is inserted in, and adjust the text prompts with both the object and background. + +To achieve view-consistent renderings, we propose a new + +strategy, namely pose-conditioned dataset updates, to schedule the data used in the NeRF updating (see Sec. 3.3). Our strategy is inspired by an important observation about the nature of NeRF: a view rendered by a NeRF maintains an extent of pose-aware color information from nearby already-used views, the nearer the more noticeable. Therefore, if we pass nearby renderings to the diffusion model with properly controlled noise, view-consistent results can be generated based on the learned color hints. From such an observation, we design a novel data scheduler for our NeRF updating. We initially train the NeRF using regular method on a dataset of multi-view background images. We then progressively fuse the object into the NeRF by iteratively updating the dataset with object-blended background images in a pose-ordered manner, i.e., views are sorted in such a way that new views are acquired nearby already-used views. Our method can also be adapted to implement object removal (see Sec. 3.4). + +# 3.2. Object-blended image synthesis + +We present our target object and background by a set of multi-view object images $\mathbf{I}_o$ and background images $\mathbf{I}_b$ . Our goal is to build an image synthesis model that can blend the object (from $\mathbf{I}_o$ ) into the background (from $\mathbf{I}_b$ ) at custom locations specified by binary masks $M$ . + +Let $D_{\theta}$ (with parameters $\theta$ ) be such an image synthesis model. Here we adopt a pre-trained Stable Diffusion [51] to implement $D_{\theta}$ . The model $D_{\theta}$ includes a denoising model $\epsilon_{\theta}$ that learns the noise component $\epsilon_{\alpha} \sim \mathcal{N}(\mathbf{0}, \mathbf{1})$ added in the diffusion process, where $\alpha \in [0, 1]$ is a denoising strength. We customize $D_{\theta}$ with our background and object images in an inpainting fashion. Specifically, for each image $I \in \mathbf{I}_b \cup \mathbf{I}_o$ , we make a masked-out (background-preserved) + +version $I\odot M$ where $M\sim \mathcal{U}$ is a random square binary mask whose coordinates are sampled from a uniform distribution $\mathcal{U}$ and $\odot$ is a pixel-wise product. The image $I$ is associated with a text prompt $p$ where we use $" ^ { \text{丰} }$ to indicate our object/background identifiers [52], e.g., $" *$ sneaker" means that our target object is a sneaker and $" *$ table" is our target background. We fine-tune $D_{\theta}$ with the following loss: + +$$ +\mathcal {L} _ {\text {s y n t h e s i s}} (\theta) = \mathbb {E} _ {I, M \sim \mathcal {U}, p, \epsilon_ {\alpha} \sim \mathcal {N} (\mathbf {0}, \mathbf {1})} \left[ \| \epsilon_ {\alpha} - \epsilon_ {\theta} (I) \| _ {2} ^ {2} \right] \tag {1} +$$ + +Eq. 1 represents a general form for the fine-tuning process. Intermediate steps, e.g., input formation, latent encoding-decoding, are presented in our supplementary material. + +We fine-tune the diffusion model $D_{\theta}$ with all the images in $\mathbf{I}_b$ and $\mathbf{I}_o$ for $n_{bg}$ and $n_{obj}$ times. After fine-tuning, we can use $D_{\theta}$ to synthesize images $\hat{I}_b$ that blend the object presented in $\mathbf{I}_o$ into the background images $I_b \in \mathbf{I}_b$ via a combined text prompt $\tilde{p}$ , e.g., "a * sneaker on a * table", + +$$ +D _ {\theta} \left(I _ {b} \odot M, \tilde {p}, \alpha\right)\rightarrow \hat {I} _ {b} \tag {2} +$$ + +where $\alpha$ is set manually to control how much object content in the masked area to be inpainted by $D_{\theta}$ (we demonstrate how we set $\alpha$ for different purposes in Sec. 3.3). + +# 3.3. Pose-conditioned dataset updates + +Let $R_{\phi}$ (with parameters $\phi$ ) be our NeRF, which is initially trained with the multi-view background images in $\mathbf{I}_b$ using an existing approach [43]. We also retrieve a set of camera poses $\{\pi(I_b)\}$ ( $4 \times 4$ matrices) for the images $I_b \in \mathbf{I}_b$ using the pose estimation method COLMAP [54, 55]. + +We propose to update the NeRF $R_{\phi}$ progressively with object-blended background images generated by the finetuned diffusion model $D_{\theta}$ . Our idea is to build a progressive multi-view and pose-conditioned object-blended image dataset $\hat{\mathbf{I}}_b$ from $\mathbf{I}_b$ to be used to update $R_{\phi}$ . We initialize $\hat{\mathbf{I}}_b^{(0)} = \emptyset$ and $\mathbf{I}_b^{(0)} = \mathbf{I}_b$ . We start with a random background image $I_b \in \mathbf{I}_b$ . Let $B$ be a 3D bounding box, and $B(I_b)$ be the projection mask of $B$ onto $I_b$ using the camera pose $\pi (I_b)$ . We generate an object-blended image $\hat{I}_b$ as, + +$$ +D _ {\theta} \left(I _ {b} \odot B \left(I _ {b}\right), \tilde {p}, \alpha\right)\rightarrow \hat {I} _ {b} \tag {3} +$$ + +where we set $\alpha = 1$ (maximum strength) to blend the object into the masked area by using only the knowledge that $D_{\theta}$ learns from $\tilde{p}$ (during the fine-tuning), and colors from the background image $I_{b}$ as there is no existing color clue in the masked area $I_{b} \odot B(I_{b})$ . + +We update $\hat{\mathbf{I}}_b^{(1)} = \hat{\mathbf{I}}_b^{(0)}\cup \{\hat{I}_b\}$ and $\mathbf{I}_b^{(1)} = \mathbf{I}_b^{(0)}\setminus \{I_b\}$ . We then update $R_{\phi}$ using $\hat{\mathbf{I}}_b^{(1)}$ , and keep doing so by adding images into $\hat{\mathbf{I}}_b$ sequentially. In particular, let $\hat{\mathbf{I}}_b^{(n - 1)}$ be the dataset at the $(n - 1)$ -th step. The next image $I_b^{(n)}\in \mathbf{I}_b$ for + +step $n$ is selected so as it is closest (in terms of the camera poses) to already-used images, + +$$ +I _ {b} ^ {(n)} = \underset {I _ {b, k} \in \mathbf {I} _ {b} ^ {(n)}} {\arg \min } \underset {I _ {b, j} \in \mathbf {I} _ {b} ^ {(n - 1)}} {\min } \| \pi_ {T} (I _ {b, k}) - \pi_ {T} (I _ {b, j}) \| ^ {2} \tag {4} +$$ + +where $\pi_T$ is the translation component of the pose $\pi$ . This view selection reflects the standard multi-view reconstruction pipeline where the multi-view data are collected with a smoothly connected, inward-surrounding camera trajectory. This requirement is also fulfilled by most NeRF datasets. + +Given a background image $I_b^{(n)}$ in Eq. 4, to utilize the nearby color hints learned by $R_{\phi}$ , we create a background-preserved foreground-rendered image $\tilde{I}_b^{(n)}$ as, + +$$ +\tilde {I} _ {b} ^ {(n)} = \left(I _ {b} ^ {(n)} \odot B \left(I _ {b} ^ {(n)}\right)\right) \oplus \left(R _ {\phi} \left(\pi \left(I _ {b} ^ {(n)}\right)\right) \odot \bar {B} \left(I _ {b} ^ {(n)}\right)\right) \tag {5} +$$ + +where $R_{\phi}(\pi(I_b^{(n)}))$ is the rendering result of $R_{\phi}$ in the pose $\pi(I_b^{(n)})$ (for the background image $I_b^{(n)}$ ), $\bar{B}$ is the complement of $B$ (as we want to maintain the foreground rendered by $R_{\phi}$ ), and $\oplus$ is a pixel-wise addition. + +We then generate a view-consistent image $\hat{I}_b^{(n)}$ by applying $D_{\theta}$ to $\tilde{I}_b^{(n)}$ defined in Eq. 5 as, + +$$ +D _ {\theta} \left(\tilde {I} _ {b} ^ {(n)} \odot B \left(I _ {b} ^ {(n)}\right), \tilde {p}, \alpha\right)\rightarrow \hat {I} _ {b} ^ {(n)} \tag {6} +$$ + +where we set $\alpha$ to a low value to allow $D_{\theta}$ to utilize color hints from previous nearby views, provided from the rendering results $R_{\phi}(\pi (I_b^{(n)}))$ . We empirically found that $\alpha \in [0.3, 0.4]$ gives best view-consistent rendering. + +Again, we update $\hat{\mathbf{I}}_b^{(n)}$ with $\hat{I}_b^{(n)}$ , and update $R_{\phi}$ accordingly by minimizing the loss: + +$$ +\mathcal {L} _ {\mathrm {N e R F}} (\phi) = \mathbb {E} _ {\hat {I} _ {b} \in \hat {\mathbf {I}} _ {b} ^ {(n)}} \left[ \| R _ {\phi} \left(\pi \left(\hat {I} _ {b}\right)\right) - \hat {I} _ {b} \| ^ {2} \right]. \tag {7} +$$ + +During the updating process, we include $n_{near}$ views into the ongoing dataset for every $n_{new}$ NeRF updating steps. In addition, we periodically replace each old view by a new one using Eqs. 5 and 6 for every $n_{old}$ NeRF updating steps. The updating procedure is completed once all the background images in $\mathbf{I}_b$ have been processed, i.e., $\mathbf{I}_b^{(n)} = \emptyset$ . + +# 3.4. Adapting to object removal + +Our framework can be adapted to object removal. In particular, we also fine-tune the diffusion model $D_{\theta}$ as in Sec. 3.2. However, we do not fine-tune $D_{\theta}$ with object images as we want to remove objects. Instead, we first individually inpaint all background images on projection masks using background prompts without identifier (e.g., "a table"). These inpainted images are then treated as pseudo ground-truth and used to customize $D_{\theta}$ with the background text prompts containing identifiers (e.g., "a * table"). The pseudo ground-truth backgrounds are visually pleasing but still remain cross-view inconsistencies. To circumvent this issue, we perform + +NeRF updating using dataset updates as in Sec. 3.3. However, only background prompts are used during the NeRF updating. We found that this joint 2D-3D interaction gradually transforms disruptive inpainted image regions into view-consistent background images. We show the necessity of the pseudo ground-truth and the NeRF updating in making view-consistent object removal in our experiments. + +# 4. Experiments + +# 4.1. Datasets + +For object insertion, we propose a dataset comprising multiview images of 8 backgrounds and 10 objects. The number of images for each background and object ranges from 60 to 100 and 40 to 80, respectively. We collected the images using an iPhone camera and resized the images to $512 \times 512$ resolution for the ease of training. Except for our self-captured data, we also test on synthetically rendered multiview images [40, 47]. + +For object removal, we run experiments on the commonly used inpainting datasets from Mip-NeRF-360 [3] and IBR-Net [64]. We centrally cropped images in those datasets to make $512 \times 512$ images to fit with our pipeline. + +# 4.2. Baselines + +For comparisons, we select SOTA baselines from different branches of work that can perform object insertion. To ensure the fairness, we barely modify the baselines, only when necessary to fit them with our settings. + +Traditional 3D editing pipeline. This pipeline creates a 3D model from multi-view images. Scene editing is then performed directly on the 3D model. To simulate this traditional pipeline, we use COLMAP [54, 55] to reconstruct the mesh for the background and object, and then manually crop and place the object mesh into the background mesh. + +Image inpainting. We adopt the inpainting variant of Stable Diffusion [51] as a baseline. However, since image inpainting treats each view independently, for a fair comparison, we perform single-view inpainting on each background image. + +Single-image-to-3D NeRF. NeRF-based view synthesis partially fulfills scene editing. Here we select Zero123 [34], a SOTA that uses a distillation prior [47] from a 2D diffusion model to synthesize novel views for comparison. + +Instruct-NeRF2NeRF [14]. This work also applies dataset updates for NeRF training. However, it fails to add/remove objects with noticeable non-uniform appearance. To test this method, we re-format text prompts, e.g., we replace "a * sneaker on a * table" by "add/make a sneaker on a table". + +# 4.3. Implementation details + +We fine-tune the diffusion model $D_{\theta}$ with object images $(n_{obj} = 5,000$ times) more than background images $(n_{bg} = 500$ times) as we found the model needs extra training to + +learn objects with complex geometry and texture. For NeRF training, we empirically found $n_{near} = 3$ , $n_{new} = 500$ , and $n_{old} = 10$ balance well the quality and efficiency. + +We run all experiments on a Nvidia RTX 3090 GPU. Diffusion fine-tuning takes around 30 minutes. NeRF updating speed relies on the number of background images while each backpropagation takes 0.5 seconds. Inference speed of the diffusion model is 8 seconds/image. Our NeRF training costs around 2.7 times more than the standard NeRF training. + +# 4.4. Qualitative results + +Object insertion. We present qualitative results of object insertion in Fig. 3. As shown, our method (Fig. 3-f) can generate plausible contents with view consistency. Moverover, the outputs of our method also match well the text prompts and the generated objects are precisely located. + +In contrast, the traditional 3D pipeline (Fig. 3-b) suffers from seamed object-background boundaries and unrealistic lighting. The geometry in obscure regions, e.g., corners, is not accurately reconstructed. The image inpainting baseline (Fig. 3-c) produces reasonable results on individual views but generates view-inconsistency. Single-image-to-3D NeRF (Fig. 3-d) fails on complex scenes. A larger camera pose shift can result in background mismatch or even content collapse. Instruct-NeRF2NeRF (Fig. 3-e) is known to be strong at stylizing existing geometry but weak at generating new geometry. Under our setting, it performs random view-consistent editing, but fails to accomplish object insertion. + +Object removal. We illustrate several results of object removal in Fig. 4, which shows that our method can generate view-consistent backgrounds. We also observed that without using pseudo ground-truth background, the removal can cause gradual background collapse, which we discuss further in our ablation study in Sec. 4.6. + +# 4.5. Quantitative results + +Since there is no real-world ground-truth for scene editing, direct quantitative evaluations are not possible. Instead, we adopt the CLIP Score [16], a well-known metric to measure how well an image correlates to a target prompt in the CLIP space. Let $E(I)$ and $E(p)$ be the CLIP embeddings of an image $I$ and a text prompt $p$ . The CLIP Score is defined as, + +$$ +\operatorname {C L I P S c o r e} (I, p) = \cos^ {+} (E (I), E (p)) \tag {8} +$$ + +where $\cos^{+}(a,b) = \max (0,\cos (a,b))$ . We average the CLIPScore $(\hat{I}_b,\tilde{p})$ for all the edited images $\hat{I}_b\in \hat{\mathbf{I}}_b$ paired with their target prompts $\tilde{p}$ , and use this average value as a performance metric for scene editing; higher score means better performance. + +We also use the CLIP Directional Consistency (CLIPDC) in [14] to measure the editing quality and consistency across views in the CLIP space. This metric relies on an assumption that the difference in the CLIP space between a background + +![](images/7a469988be365c00da5df2342c1bea6c66583c45b8d8940871a4f7260d82837e.jpg) +Figure 3. Qualitative results of object insertion. Inputs include multi-view object/background images, a 3D bounding box where the object is inserted in, and a text prompt. Note that some baselines use parts of the inputs due to the nature of their techniques. + +image $I_{b}$ and its edited version $\hat{I}_{b}$ should match with the difference between the original prompt $p$ (paired with $I_{b}$ ) and the customized prompt $\tilde{p}$ (paired with $\hat{I}_b$ ). Moreover, a good editing should make consistent differences between every pair of $I_{b}$ and $\hat{I}_b$ , especially for adjacent views. We follow [14] to calculate the CLIPDC between two adjacent + +edited views $\hat{I}_b^{(i)}$ and $\hat{I}_b^{(i + 1)}$ as, + +$$ +\begin{array}{l} \operatorname {C L I P D C} \left(\hat {I} _ {b} ^ {(i)}, \hat {I} _ {b} ^ {(i + 1)}\right) = \\ \cos^ {+} (E (\tilde {p}) - E (p), E (\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)})) \times \tag {9} \\ \cos^ {+} \left(E (\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)}), E (\hat {I} _ {b} ^ {(i + 1)}) - E (I _ {b} ^ {(i + 1)})\right) \\ \end{array} +$$ + +We measure the CLIPDC of an edited scene by averaging + +![](images/b5be5d30166ec4c05c787446d7a26e3f009183097c893855cd6f3e1f5e9ea4e1.jpg) + +![](images/6890a1aea494ff605eac3c30bbeadbf2bb41848d5ed6e9d18e359022a0aaf9d5.jpg) + +![](images/721697224708c1135dd2f92bd36eabd8b0f2fab0bdc698c6780dd16fe247148f.jpg) + +![](images/17a93be35b9bea6ed344c8fa0029092af4a60b57ad719be707571140a4a0f7a5.jpg) + +![](images/f9c628322d00cfa69b9153551f1aaf37555778ef9b008c09a670373f22c205c9.jpg) + +![](images/c30d90367c851029730c85d89ed9e5c35b42b25f4106b465dac5a18202928f78.jpg) + +![](images/746b293bc1723199285da4cdf9baa3de98bfe0fa8598f57f62711b07a152b6a6.jpg) + +![](images/faa48321ebbe27ee602177623a2a2e1fd533191585dac98cd8e06bb6a35893fd.jpg) + +![](images/ae6aafa184f599ed0e73e019d6a6ab82988fed8f3d5c61337a2963919fe48bb9.jpg) + +![](images/fb01b77239be06ed66a255a408b84c063d4077750c46b86ae16dcbf066e2539a.jpg) +(a) Original BG with object bounding box + +![](images/f6ccf694c33f8ea3b986f7b91ce8338dd465bc4c3a6c32be2dc8cdff0a2d9234.jpg) +(b) Pseudo ground-truth BG + +![](images/4747e311bac08754a3a78caa193274b3524594a8b95950634bb317b38495b81a.jpg) +(c) View-consistent editing result +Figure 4. Qualitative results of object removal where we show the importance of pseudo ground-truth background (BG) in generating view-consistent editing results. + +
MethodCLIPScore ↑CLIPDC ↑
Traditional 3D [54, 55]0.26480.1169
Inpainting [51]0.26830.1325
Zero123 [34]0.21970.0361
Instruct-NeRF2NeRF [14]0.23470.0263
Ours0.27430.1642
+ +Table 1. Quantitative results. Higher CLIP metrics scores indicate higher image editing quality or consistency. + +the CLIPDC $(\hat{I}_b^{(i)},\hat{I}_b^{(i + 1)})$ for all the adjacent image pairs $(\hat{I}_b^{(i)},\hat{I}_b^{(i + 1)})$ in the scene. We report both the CLIPScore and CLIPDC of our method and other baselines on 35 unique edits (cross editing on the more generalizable 5 scenes $\times$ 7 objects) in Tab. 1. Experimental show that our method outperforms all the baselines on both the CLIPScore and CLIPDC metrics. + +# 4.6. Ablation Study + +In this ablation study, we validate the effectiveness of technical components in our method. We suggest readers observe the results in Fig. 5, which shows two views of an edited scene. We provide detailed analyses on these results below. Diffusion model fine-tuning. Recall that we fine-tune the diffusion model $D_{\theta}$ on both object and background images (in Sec. 3.2). Here we prove that such a fine-tuning is necessary. Fine-tuning on object images makes the diffusion model aware of the same object during the dataset updates (in Sec. 3.3). We verify this in Fig. 5-a, where we skip the fine-tuning of $D_{\theta}$ on object images. As shown, without using object images, the diffusion model cannot generate the same object across views. + +Likewise, fine-tuning the diffusion model with background images helps to preserve the background at inpainted borders (areas in between mask boundary and inner-mask object boundary). As shown in Fig. 5-b, without fine-tuning on background images, the model cannot inpaint the background within the mask region properly. The inpainted background gets darker at latter NeRF updating steps, and finally ends up with a noticeable error. We hypothesize this collapse that the model, without being fine-tuned on relevant background images, applies the bias from pre-learned knowledge which does not align with the surrounding background. + +Pose-conditioned dataset updates. To prove the effectiveness of the pose-conditioned dataset update strategy, we experiment with another scheme in which the dataset used to refine the NeRF is updated with random views. We present the results of this experiment in Fig. 5-c. We observe that, with view-random updates, objects in different views can converge into inconsistent poses (e.g., the hats in the left and right view are generated in different poses). + +The above inconsistency commonly happens with many text-to-3D NeRF generation or editing methods. We hypothesize this phenomenon as follows. Initially the diffusion model can generate the object in inconsistent poses across views due to little 3D-aware hint available. This inconsistency is continually introduced to the NeRF updating and then, in return, passed partially as input to the diffusion model (as in Eq. 6), making further pose divergence eventually. Without considering nearby-views, the diffusion model cannot fix this inconsistency as objects generated on individual views match well their given text prompts. + +Our proposed pose-conditioned dataset update strategy simulates the nature of NeRF construction, in which a view rendered closed to already-used views should contain similar but slightly blurry and distorted object content. Pose-conditioned view arrangement thus gives the diffusion model enough hints to generate objects with view-consistent appearance and poses, specified in nearby views. View-consistency is thus achieved progressively in the training dataset and is integrated into the constructed NeRF (see Fig. 5-e). + +![](images/89ffbaa2f17dac027e3910a281c3160d38f308d311280cdf06985953c9dde8d8.jpg) +(a) Fine-tuning of the diffusion model without using object images + +![](images/0b52ea12b2c4b0590e356c944e1ddf15f2ab193f2255d38c8e4ebcd84efa6ab6.jpg) + +![](images/2aadfa0bfdb4792f546191664180902d62e6a2a9c7f9613efeb8923906798521.jpg) +(b) Fine-tuning of the diffusion model without using background images + +![](images/c488f96364b847254e7408a119a8342ea5908ec5ffd4fab9e633e81864f72ff9.jpg) +(c) Dataset updates with random poses (without pose-conditioned updates) +(d) Without periodic dataset updates + +![](images/95cd1e08b02da26e3393e9d9b06ac1405b1d31e15d6171b0a56bbd513feebaad.jpg) +(e) Full pipeline +Figure 5. Ablation study results. Each row shows the results of a variant of our pipeline. The left and right columns include images of two different views of an edited scene. For each column, from left to right are the results of increasing training steps, where the most left image in each column is an early-stage result and the most right is the final output. The left column is a view near the starting view, which converges faster than the right column from a farther view (except for the variant in (c) where both views converge equally fast). + +Periodic dataset updates. We observe that periodically updating of training views during the NeRF updating is important to achieve high-quality renderings. We verify this by keeping all the training views from the pose-conditioned dataset updates fixed during the NeRF updating. We found that although objects are rendered at fairly accurate orientations, their texture and geometry still suffer from inconsistencies (see Fig. 5-d). Having periodically updates on the training data fixes these minor defects and facilitates a more consistent convergence of the NeRF updating (see Fig. 5-e). + +# 5. Conclusion + +We propose a new language-driven method for object manipulation in NeRFs. Our method is built upon a novel idea of joint 2D-3D interaction, keeping both 2D image synthesis + +and 3D NeRF reconstruction in a loop. This idea is enabled by an advanced text-to-image diffusion technique that generates object-blended background images, and a novel pose-conditioned dataset update strategy that learns a NeRF from the multi-view images in a progressive manner. + +Our method is not without limitations. Since our 2D views are synthesized by a diffusion model, we may share the flickering problem with diffusion-based video editing methods [26, 48, 68]. We leave this for future work, which can potentially be addressed by robust video translation methods [73]. It is also of great interest to postulate a theoretical foundation for pose-conditioned dataset update to better understand the convergence of NeRF training in scene editing. Acknowledgment. This work was partially supported by a grant from the RGC of HKSAR, China (Project No. HKUST 16202323) and an internal grant from HKUST (R9429). + +# References + +[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 5 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2 +[5] Ronald A Castellino. Computer aided detection (cad): an overview. Cancer Imaging, 5(1):17, 2005. 1 +[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2 +[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[9] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 1 +[10] Epic Games. Unreal engine. 1 +[11] SM Ali Eslami, Danilo Jimenez Rezende, Frederic Besse, Fabio Viola, Ari S Morcos, Marta Garnelo, Avraham Ruderman, Andrei A Rusu, Ivo Danihelka, Karol Gregor, et al. Neural scene representation and rendering. Science, 360 (6394):1204-1210, 2018. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2 +[13] John K Haas. A history of the unity game engine. 2014. 1 +[14] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing + +3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 5, 6, 7 +[15] Jeffrey Harper. Mastering Autodesk 3ds Max 2013. John Wiley & Sons, 2012. 1 +[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[18] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multiview stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2 +[19] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2 +[20] Yi-Hua Huang, Yan-Pei Cao, Yu-Kun Lai, Ying Shan, and Lin Gao. Nerf-texture: Texture synthesis with neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2 +[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[23] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 2 +[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2 +[26] Gyeongman Kim, Hajin Shim, Hyunsu Kim, Yunjey Choi, Junho Kim, and Eunho Yang. Diffusion video autoencoders: Toward temporally consistent face video editing via disentangled video encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6091-6100, 2023. 8 +[27] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. arXiv preprint arXiv:2302.06608, 2023. 2 +[28] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation. + +Advances in Neural Information Processing Systems, 35: 23311-23330, 2022. 2 +[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 2 +[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2 +[31] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 2 +[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3 +[33] Miaomiao Liu, Xuming He, and Mathieu Salzmann. Geometry-aware deep network for single-image novel view synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4616-4624, 2018. 2 +[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5,6,7 +[35] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 2 +[36] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2 +[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2 +[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8446-8455, 2023. 3 +[39] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2 +[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5 +[41] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[42] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 2 +[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 4 +[44] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2 +[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2 +[46] Hong-Wing Pang, Binh-Son Hua, and Sai-Kit Yeung. Locally stylized neural radiance fields. In ICCV, 2023. 2 +[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 5 +[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 8 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 5, 6, 7 +[52] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael + +Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2 +[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 6, 7 +[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4, 5, 6, 7 +[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 2 +[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2 +[58] Hyeonseop Song, Seokhun Choi, Hoseok Do, Chul Lee, and Taehyeong Kim. Blending-nerf: Text-driven localized editing in neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14383–14393, 2023. 2 +[59] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 2 +[60] Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Multiview consistency as supervisory signal for learning shape and pose prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2897-2905, 2018. 2 +[61] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2 +[62] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[63] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094, 2023. 2 +[64] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 5 +[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. + +In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2 +[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3 +[67] Silvan Weder, Guillermo Garcia-Hernando, Aron Monszpart, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023. 2 +[68] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 8 +[69] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 2 +[70] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2 +[71] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20908–20918, 2023. 3 +[72] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4402-4412, 2023. 2 +[73] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. *Rerender a video: Zero-shot text-guided video-to-video translation.* arXiv preprint arXiv:2306.07954, 2023. 8 +[74] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2 +[75] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2 +[76] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[77] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2 + +[78] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2 +[79] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2 +[80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 +[81] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2 \ No newline at end of file diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/images.zip b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e33398715bfdb45e362240d470e68d5051f0503e --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5b75b254806778e092b9cfab770ec3da4d9cf14f5d7117ab3def0dc1021405e +size 894177 diff --git a/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/layout.json b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..824619ac37e8cd20be048a41703f9061ec5d3d4e --- /dev/null +++ b/2024/Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates/layout.json @@ -0,0 +1,11186 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 483, + 140 + ], + "type": "text", + "content": "Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 170, + 521, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 170, + 521, + 186 + ], + "spans": [ + { + "bbox": [ + 71, + 170, + 521, + 186 + ], + "type": "text", + "content": "Ka Chun Shum¹ Jaeyeon Kim¹ Binh-Son Hua²,⁴ Duc Thanh Nguyen³ Sai-Kit Yeung¹" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "text", + "content": "VinAI " + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "text", + "content": "Deakin University " + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 47, + 194, + 545, + 210 + ], + "type": "text", + "content": "Trinity College Dublin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 271, + 290, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 487 + ], + "type": "text", + "content": "Neural radiance field (NeRF) is an emerging technique for 3D scene reconstruction and modeling. However, current NeRF-based methods are limited in the capabilities of adding or removing objects. This paper fills the aforementioned gap by proposing a new language-driven method for object manipulation in NeRFs through dataset updates. Specifically, to insert an object represented by a set of multi-view images into a background NeRF, we use a text-to-image diffusion model to blend the object into the given background across views. The generated images are then used to update the NeRF so that we can render view-consistent images of the object within the background. To ensure view consistency, we propose a dataset update strategy that prioritizes the radiance field training based on camera poses in a pose-ordered manner. We validate our method in two case studies: object insertion and object removal. Experimental results show that our method can generate photo-realistic results and achieves state-of-the-art performance in NeRF editing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 512, + 127, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 127, + 522 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 127, + 522 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "Editing of 3D scenes by insertion or removal of objects has been a fundamental task in computer graphics and computer vision. The task has often been performed using traditional 3D scene authoring tools [5, 9]. For example, to insert an object into a 3D scene, traditional approach requires user to manually select the object and position it into the scene. This manual pipeline has been adopted in a wide range of applications, such as furniture arrangement in interior design [15] and asset creation in game development [10, 13]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "Recent advances in deep learning have opened new directions to scene modeling. Neural radiance field (NeRF) [40] is a pioneer to render view-consistent photo-realistic images using neural networks. Generative models such as generative adversarial networks (GANs) [12] and diffusion models [17] learn to output photo-realistic images from un" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 258, + 537, + 338 + ], + "blocks": [ + { + "bbox": [ + 350, + 249, + 501, + 258 + ], + "lines": [ + { + "bbox": [ + 350, + 249, + 501, + 258 + ], + "spans": [ + { + "bbox": [ + 350, + 249, + 501, + 258 + ], + "type": "text", + "content": "Input: multi-view object images & object text prompt" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 258, + 537, + 338 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 537, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 537, + 338 + ], + "type": "image", + "image_path": "8b7ec97ea103525c7798b80878297a59f637096c4f2bdaa606d4b0000184f75d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 350, + 536, + 407 + ], + "blocks": [ + { + "bbox": [ + 334, + 341, + 518, + 350 + ], + "lines": [ + { + "bbox": [ + 334, + 341, + 518, + 350 + ], + "spans": [ + { + "bbox": [ + 334, + 341, + 518, + 350 + ], + "type": "text", + "content": "Input: multi-view background images & background text prompt" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 350, + 536, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 536, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 536, + 407 + ], + "type": "image", + "image_path": "78346f513bd46aa69d4cc1355d1cf5da1407a424aa2e5bdbc06b4a7bfa420998.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 314, + 426, + 538, + 507 + ], + "blocks": [ + { + "bbox": [ + 332, + 417, + 517, + 426 + ], + "lines": [ + { + "bbox": [ + 332, + 417, + 517, + 426 + ], + "spans": [ + { + "bbox": [ + 332, + 417, + 517, + 426 + ], + "type": "text", + "content": "Output: multi-view object-blended images & novel view synthesis" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 426, + 538, + 507 + ], + "lines": [ + { + "bbox": [ + 314, + 426, + 538, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 538, + 507 + ], + "type": "image", + "image_path": "896e69843f5988da44b044e266c7b7a9fce35bd973a0e30a35200e08444ac447.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 523, + 547, + 579 + ], + "lines": [ + { + "bbox": [ + 305, + 523, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 523, + 547, + 579 + ], + "type": "text", + "content": "Figure 1. Object insertion. We propose a language-driven method for view-consistent 3D object insertion into a background NeRF scene. Given an object defined by a set of multi-view images, our method generates plausible text-guided insertion results that require geometry manipulation of the original background NeRF." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 604, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 640 + ], + "type": "text", + "content": "constrained image collections. Recent text-guided diffusion models [50, 51, 53] have shown great promise in generating high-quality and diverse images from a single text prompt." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "Such successes inspire us to revisit 3D scene editing via language-driven image synthesis techniques. Particularly, for object insertion (see Fig. 1), our method performs view-consistent edits by taking as input a set of multi-view object images, multi-view background images, and a target text prompt. We utilize a text-to-image diffusion model to con" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5176" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 276 + ], + "type": "text", + "content": "tinueously synthesize multi-view images containing both a target object and a background during NeRF training. These synthesized images can be iteratively used to refine the NeRF of the background to learn geometric and appearance features of the object. This approach to refining a radiance field is also known as dataset update [14]. However, a particular challenge is that the refinement process may result in unstable NeRF training, degrading its rendering quality due to inconsistent views synthesized. To address this issue, we propose a pose-conditioned dataset update strategy that gradually engages the target object into the background, beginning at a randomly selected pose (view), then views close to the already-used views before propagating to views further away. We observe that this pose-conditioned strategy significantly improves the NeRF learning, reducing rendering artifacts and maintaining view-consistent rendering. In summary, we make the following contributions in our work." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 277, + 288, + 540 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 277, + 288, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 288, + 348 + ], + "type": "text", + "content": "- We propose a new framework for object manipulation in NeRFs via text-to-image diffusion. Our work can create high-quality 3D scenes from simple inputs (text prompts and multi-view images). It, therefore, has the potential for conveniently building a variant of 3D object libraries that is applicable for the scene editing task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 350, + 288, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 288, + 433 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 288, + 433 + ], + "type": "text", + "content": "- We propose a pose-conditioned dataset update strategy that stabilizes the fusion of objects into a background NeRF, enabling view-consistent rendering. Our method requires no priors on the geometry or texture of the objects, unlike existing works relying on accurate meshes [7, 20, 31, 81], depth maps [42, 67], fine masks [58, 63], semantics [1], or lighting assumption [69, 78]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 434, + 288, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 288, + 504 + ], + "type": "text", + "content": "- We showcase our method in two case studies: object insertion and object removal via a user-friendly manner. Specifically, we use a 3D bounding box to define the location for object insertion/removal. Box-based location requires only a rough orientation, which is easy to be adjusted and visualized using real-time NeRF GUIs [43, 59]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 505, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 288, + 540 + ], + "type": "text", + "content": "- We conduct extensive experiments to validate key techniques of our method and demonstrate its state-of-the-art performance in NeRF editing." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 570, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 287, + 665 + ], + "type": "text", + "content": "Image synthesis. Early attempts in deep learning-based image synthesis have utilized generative models such as GANs [12]. Several methods combine supervised losses with adversarial losses to learn class-conditional GANs [41] and image-to-image translations [21, 65, 80]. StyleGAN [22] and its variants [23, 24] perform image editing directly on latent representations, but such editorings are not intuitive and thus difficult to control the output." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 288, + 713 + ], + "type": "text", + "content": "Recently, diffusion models [17] have made significant progress in image synthesis with high-quality data samples constructed from sophisticated forward and denoising diffusion steps. In addition, vision-language models such as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "CLIP [49] have made text prompts an intuitive condition to generate and edit images. These developments have led to text-guided diffusion models [50, 51, 53] which offer excellent synthesis quality. Downstream applications then can be built by fine-tuning these text-guided models to fit with the application domains [4, 25, 36, 52, 70, 77]. Our method follows this streamline where we use a text-to-image diffusion model [52] to guide the fusion of objects into a background." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 183, + 548, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 548, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 548, + 352 + ], + "type": "text", + "content": "Neural 3D modeling. Image-based 3D modeling methods use convolutional neural networks (CNNs) [29] to learn 3D structures from multiple images [11, 18, 33, 60, 79]. However, CNNs often struggle to deal with complex shapes, texture, and lighting captured in the images. Follow-up works integrate differentiable rendering and represent 3D structures as neural surfaces [39] or shapes [8]. 3D-aware GANs integrates volume rendering into their generators to synthesize novel views from a single image [6, 45, 56, 57]. NeRFs [40] apply the same neural rendering approach, but are optimized on a ray rendering loss on multi-view input images. There are methods addressing the limitations of NeRFs in various aspects such as visual artifacts [2], data complexity [37], camera poses [74], and computational efficiency [43, 76]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 510 + ], + "type": "text", + "content": "NeRF editing. NeRF editing has often been carried out by parameter tuning [35], layer feature fusion [61], or deformable rays [75]. An alternative to editing a NeRF is to directly amend the multi-view images used to learn it. For example, NeRF stylization methods [19, 44, 46, 62] freeze the geometry branch and optimize the color branch in a NeRF to stylize multi-view images. Several methods attempt to decompose an existing NeRF, which in turn holds color information and separates voxels considering multi-view masks [30] or semantics [27, 28]. NeRF inpainting fills simple unseen background geometry and colors with help of depth priors [42] or filtering inpainted multi-views [67]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 548, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 676 + ], + "type": "text", + "content": "Creation of complex geometry and vivid colors for a NeRF is challenging due to higher-level requirements of consistency. DiscoScene [72] fuses background and object NeRFs, thus is unable to condition customized contents. FocalDreamer [31] and DreamEditor [81] rely on fine meshes to function, disregarding the advantage of multi-view representation of NeRFs. Our method is perhaps most similar to Instruct-NeRF2NeRF [14] which shares a related data updating schema. However, Instruct-NeRF2NeRF [14] always generates geometry aligned with the original geometry (e.g., transfer a sneaker to a sneaker-shape apple), thus fails in most cases of object insertion or removals. Besides, it requires heavy retraining of a diffusion model on large-scale self-constructed datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Recently, text-to-image diffusion has been applied to generate 3D contents. DreamFusion [47] introduced a score distillation sampling (SDS) loss that progressively consolidates" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5177" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 250 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 250 + ], + "type": "image", + "image_path": "383d36ec2e8eba0016311fdae1eedf40f9a416e5a7ee6cbc6234feeab3164338.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "type": "text", + "content": "Figure 2. Overview of our pipeline. We customize and fine-tune a text-to-image diffusion model for view synthesis in an inpainting manner (left). We then apply the model to progressively fuse an object into background views to update a background NeRF (right). The process of view synthesis and NeRF updating is performed repeatedly. Views generated by the diffusion model are added to an on-going dataset to strengthen the NeRF. In return, the NeRF renders color hints for the diffusion model to create new views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 328, + 290, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 290, + 401 + ], + "type": "text", + "content": "view information from a diffusion model into NeRFs. This loss has been applied with additional treatments to image resolution [32], conditional images [38], photo-realism [66], or scene geometry [71]. We also adopt the SDS loss but develop a novel training schedule to address challenges in multi-view object and background fusion for NeRF editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 411, + 103, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 103, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 103, + 423 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 115, + 443 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 605 + ], + "type": "text", + "content": "For the ease of presentation, we first describe our method for object insertion. We then extend it to object removal. An overview of our method for object insertion is illustrated in Fig. 2. Here we aim to insert an object into a background NeRF in two steps. In the first step, we synthesize training views for the NeRF with the target object embedded in the background. In the second step, the NeRF is updated with the synthesized views to learn the geometry and appearance of the object. There are two key challenges in this approach. First, background preserving is required in the synthesized views for the NeRF updating. Second, image synthesis may generate view-inconsistent images, causing artifacts in the resulting NeRF." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 701 + ], + "type": "text", + "content": "To address the first challenge (i.e., object-blended image synthesis with background preserving), we leverage a state-of-the-art diffusion model to image synthesis, and opt to customize the model with text prompts for object blending (see Sec. 3.2). We formulate this task as image inpainting where we place a binary mask on each background image to indicate the location where the object is inserted in, and adjust the text prompts with both the object and background." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "To achieve view-consistent renderings, we propose a new" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 328, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 547, + 533 + ], + "type": "text", + "content": "strategy, namely pose-conditioned dataset updates, to schedule the data used in the NeRF updating (see Sec. 3.3). Our strategy is inspired by an important observation about the nature of NeRF: a view rendered by a NeRF maintains an extent of pose-aware color information from nearby already-used views, the nearer the more noticeable. Therefore, if we pass nearby renderings to the diffusion model with properly controlled noise, view-consistent results can be generated based on the learned color hints. From such an observation, we design a novel data scheduler for our NeRF updating. We initially train the NeRF using regular method on a dataset of multi-view background images. We then progressively fuse the object into the NeRF by iteratively updating the dataset with object-blended background images in a pose-ordered manner, i.e., views are sorted in such a way that new views are acquired nearby already-used views. Our method can also be adapted to implement object removal (see Sec. 3.4)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 539, + 477, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 477, + 553 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 477, + 553 + ], + "type": "text", + "content": "3.2. Object-blended image synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": "We present our target object and background by a set of multi-view object images " + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_o" + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": " and background images " + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": ". Our goal is to build an image synthesis model that can blend the object (from " + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_o" + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": ") into the background (from " + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": ") at custom locations specified by binary masks " + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 558, + 547, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " (with parameters " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": ") be such an image synthesis model. Here we adopt a pre-trained Stable Diffusion [51] to implement " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": ". The model " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " includes a denoising model " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " that learns the noise component " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\alpha} \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{1})" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " added in the diffusion process, where " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0, 1]" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " is a denoising strength. We customize " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": " with our background and object images in an inpainting fashion. Specifically, for each image " + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "inline_equation", + "content": "I \\in \\mathbf{I}_b \\cup \\mathbf{I}_o" + }, + { + "bbox": [ + 304, + 618, + 547, + 714 + ], + "type": "text", + "content": ", we make a masked-out (background-preserved)" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5178" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "version " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "I\\odot M" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "M\\sim \\mathcal{U}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is a random square binary mask whose coordinates are sampled from a uniform distribution " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is a pixel-wise product. The image " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is associated with a text prompt " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " where we use " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\" ^ { \\text{丰} }" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to indicate our object/background identifiers [52], e.g., " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\" *" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " sneaker\" means that our target object is a sneaker and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\" *" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " table\" is our target background. We fine-tune " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " with the following loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 177, + 288, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 177, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 177, + 288, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s y n t h e s i s}} (\\theta) = \\mathbb {E} _ {I, M \\sim \\mathcal {U}, p, \\epsilon_ {\\alpha} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {1})} \\left[ \\| \\epsilon_ {\\alpha} - \\epsilon_ {\\theta} (I) \\| _ {2} ^ {2} \\right] \\tag {1}", + "image_path": "9a10225cadfdeb9aad558703cf8889a4e6a9f0738f9e0cd99af3fcc5d1a4c76a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 201, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 236 + ], + "type": "text", + "content": "Eq. 1 represents a general form for the fine-tuning process. Intermediate steps, e.g., input formation, latent encoding-decoding, are presented in our supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": "We fine-tune the diffusion model " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " with all the images in " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_o" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "n_{bg}" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "n_{obj}" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " times. After fine-tuning, we can use " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " to synthesize images " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\hat{I}_b" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " that blend the object presented in " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_o" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " into the background images " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "I_b \\in \\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": " via a combined text prompt " + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 46, + 237, + 287, + 298 + ], + "type": "text", + "content": ", e.g., \"a * sneaker on a * table\"," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 307, + 287, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 307, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 117, + 307, + 287, + 321 + ], + "type": "interline_equation", + "content": "D _ {\\theta} \\left(I _ {b} \\odot M, \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {2}", + "image_path": "865bc29273005bac960c3c6bdf024604c6a1205742e6daef3f99ff5a385d4ff9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " is set manually to control how much object content in the masked area to be inpainted by " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " (we demonstrate how we set " + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 331, + 287, + 367 + ], + "type": "text", + "content": " for different purposes in Sec. 3.3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 374, + 226, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 226, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 226, + 387 + ], + "type": "text", + "content": "3.3. Pose-conditioned dataset updates" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " (with parameters " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": ") be our NeRF, which is initially trained with the multi-view background images in " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " using an existing approach [43]. We also retrieve a set of camera poses " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "\\{\\pi(I_b)\\}" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " matrices) for the images " + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "inline_equation", + "content": "I_b \\in \\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 393, + 287, + 453 + ], + "type": "text", + "content": " using the pose estimation method COLMAP [54, 55]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": "We propose to update the NeRF " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " progressively with object-blended background images generated by the finetuned diffusion model " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": ". Our idea is to build a progressive multi-view and pose-conditioned object-blended image dataset " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " to be used to update " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": ". We initialize " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b^{(0)} = \\emptyset" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b^{(0)} = \\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": ". We start with a random background image " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "I_b \\in \\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " be a 3D bounding box, and " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "B(I_b)" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " be the projection mask of " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "I_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " using the camera pose " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\pi (I_b)" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": ". We generate an object-blended image " + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "inline_equation", + "content": "\\hat{I}_b" + }, + { + "bbox": [ + 46, + 453, + 288, + 564 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 572, + 287, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 572, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 110, + 572, + 287, + 586 + ], + "type": "interline_equation", + "content": "D _ {\\theta} \\left(I _ {b} \\odot B \\left(I _ {b}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} \\tag {3}", + "image_path": "110a9aa68f9c24ebb186833977ee5f877a0e25e26c371b460b869258ce78ec24.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": " (maximum strength) to blend the object into the masked area by using only the knowledge that " + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": " learns from " + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": " (during the fine-tuning), and colors from the background image " + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": " as there is no existing color clue in the masked area " + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "inline_equation", + "content": "I_{b} \\odot B(I_{b})" + }, + { + "bbox": [ + 46, + 596, + 287, + 655 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": "We update " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b^{(1)} = \\hat{\\mathbf{I}}_b^{(0)}\\cup \\{\\hat{I}_b\\}" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b^{(1)} = \\mathbf{I}_b^{(0)}\\setminus \\{I_b\\}" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": ". We then update " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b^{(1)}" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": ", and keep doing so by adding images into " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": " sequentially. In particular, let " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b^{(n - 1)}" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": " be the dataset at the " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "(n - 1)" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": "-th step. The next image " + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "inline_equation", + "content": "I_b^{(n)}\\in \\mathbf{I}_b" + }, + { + "bbox": [ + 46, + 655, + 288, + 714 + ], + "type": "text", + "content": " for" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "step " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " is selected so as it is closest (in terms of the camera poses) to already-used images," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 102, + 547, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 102, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 318, + 102, + 547, + 129 + ], + "type": "interline_equation", + "content": "I _ {b} ^ {(n)} = \\underset {I _ {b, k} \\in \\mathbf {I} _ {b} ^ {(n)}} {\\arg \\min } \\underset {I _ {b, j} \\in \\mathbf {I} _ {b} ^ {(n - 1)}} {\\min } \\| \\pi_ {T} (I _ {b, k}) - \\pi_ {T} (I _ {b, j}) \\| ^ {2} \\tag {4}", + "image_path": "a54a6aea8995544faff7b05c122fd976a19ab478236905ff061cca66409acbb7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\pi_T" + }, + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "text", + "content": " is the translation component of the pose " + }, + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 135, + 547, + 194 + ], + "type": "text", + "content": ". This view selection reflects the standard multi-view reconstruction pipeline where the multi-view data are collected with a smoothly connected, inward-surrounding camera trajectory. This requirement is also fulfilled by most NeRF datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "text", + "content": "Given a background image " + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "inline_equation", + "content": "I_b^{(n)}" + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "text", + "content": " in Eq. 4, to utilize the nearby color hints learned by " + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "text", + "content": ", we create a background-preserved foreground-rendered image " + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "inline_equation", + "content": "\\tilde{I}_b^{(n)}" + }, + { + "bbox": [ + 304, + 196, + 547, + 236 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 241, + 545, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 241, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 311, + 241, + 545, + 258 + ], + "type": "interline_equation", + "content": "\\tilde {I} _ {b} ^ {(n)} = \\left(I _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right)\\right) \\oplus \\left(R _ {\\phi} \\left(\\pi \\left(I _ {b} ^ {(n)}\\right)\\right) \\odot \\bar {B} \\left(I _ {b} ^ {(n)}\\right)\\right) \\tag {5}", + "image_path": "749d21b286d72872cca7d22853160487f2da7e7d0f02bb1e8dad63f48f53b8ae.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "R_{\\phi}(\\pi(I_b^{(n)}))" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " is the rendering result of " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " in the pose " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\pi(I_b^{(n)})" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " (for the background image " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "I_b^{(n)}" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\bar{B}" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " is the complement of " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " (as we want to maintain the foreground rendered by " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 266, + 547, + 317 + ], + "type": "text", + "content": " is a pixel-wise addition." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "text", + "content": "We then generate a view-consistent image " + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\hat{I}_b^{(n)}" + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "text", + "content": " by applying " + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "inline_equation", + "content": "\\tilde{I}_b^{(n)}" + }, + { + "bbox": [ + 304, + 318, + 547, + 346 + ], + "type": "text", + "content": " defined in Eq. 5 as," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 353, + 353, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 353, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 353, + 353, + 545, + 374 + ], + "type": "interline_equation", + "content": "D _ {\\theta} \\left(\\tilde {I} _ {b} ^ {(n)} \\odot B \\left(I _ {b} ^ {(n)}\\right), \\tilde {p}, \\alpha\\right)\\rightarrow \\hat {I} _ {b} ^ {(n)} \\tag {6}", + "image_path": "7a8b168ecdb42283ae900faca83911f351605e828e315859815b7df04f6f226e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "content": "where we set " + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "content": " to a low value to allow " + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "content": " to utilize color hints from previous nearby views, provided from the rendering results " + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "inline_equation", + "content": "R_{\\phi}(\\pi (I_b^{(n)}))" + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "content": ". We empirically found that " + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0.3, 0.4]" + }, + { + "bbox": [ + 304, + 379, + 545, + 429 + ], + "type": "text", + "content": " gives best view-consistent rendering." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "text", + "content": "Again, we update " + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_b^{(n)}" + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "inline_equation", + "content": "\\hat{I}_b^{(n)}" + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "text", + "content": ", and update " + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "inline_equation", + "content": "R_{\\phi}" + }, + { + "bbox": [ + 304, + 429, + 547, + 454 + ], + "type": "text", + "content": " accordingly by minimizing the loss:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 335, + 460, + 545, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 460, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 335, + 460, + 545, + 478 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N e R F}} (\\phi) = \\mathbb {E} _ {\\hat {I} _ {b} \\in \\hat {\\mathbf {I}} _ {b} ^ {(n)}} \\left[ \\| R _ {\\phi} \\left(\\pi \\left(\\hat {I} _ {b}\\right)\\right) - \\hat {I} _ {b} \\| ^ {2} \\right]. \\tag {7}", + "image_path": "101e83ef8b1c98ccff78550490f6cd84fddef981f5eaa434a469aec9480a6e47.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": "During the updating process, we include " + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "inline_equation", + "content": "n_{near}" + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": " views into the ongoing dataset for every " + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "inline_equation", + "content": "n_{new}" + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": " NeRF updating steps. In addition, we periodically replace each old view by a new one using Eqs. 5 and 6 for every " + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "inline_equation", + "content": "n_{old}" + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": " NeRF updating steps. The updating procedure is completed once all the background images in " + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b" + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": " have been processed, i.e., " + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_b^{(n)} = \\emptyset" + }, + { + "bbox": [ + 304, + 483, + 545, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 563, + 456, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 456, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 456, + 576 + ], + "type": "text", + "content": "3.4. Adapting to object removal" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Our framework can be adapted to object removal. In particular, we also fine-tune the diffusion model " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " as in Sec. 3.2. However, we do not fine-tune " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " with object images as we want to remove objects. Instead, we first individually inpaint all background images on projection masks using background prompts without identifier (e.g., \"a table\"). These inpainted images are then treated as pseudo ground-truth and used to customize " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": " with the background text prompts containing identifiers (e.g., \"a * table\"). The pseudo ground-truth backgrounds are visually pleasing but still remain cross-view inconsistencies. To circumvent this issue, we perform" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5179" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "NeRF updating using dataset updates as in Sec. 3.3. However, only background prompts are used during the NeRF updating. We found that this joint 2D-3D interaction gradually transforms disruptive inpainted image regions into view-consistent background images. We show the necessity of the pseudo ground-truth and the NeRF updating in making view-consistent object removal in our experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 167, + 128, + 180 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 110, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 110, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 110, + 198 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 300 + ], + "type": "text", + "content": "For object insertion, we propose a dataset comprising multiview images of 8 backgrounds and 10 objects. The number of images for each background and object ranges from 60 to 100 and 40 to 80, respectively. We collected the images using an iPhone camera and resized the images to " + }, + { + "bbox": [ + 46, + 205, + 289, + 300 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 205, + 289, + 300 + ], + "type": "text", + "content": " resolution for the ease of training. Except for our self-captured data, we also test on synthetically rendered multiview images [40, 47]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 301, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 288, + 350 + ], + "type": "text", + "content": "For object removal, we run experiments on the commonly used inpainting datasets from Mip-NeRF-360 [3] and IBR-Net [64]. We centrally cropped images in those datasets to make " + }, + { + "bbox": [ + 47, + 301, + 288, + 350 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 301, + 288, + 350 + ], + "type": "text", + "content": " images to fit with our pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 357, + 113, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 113, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 113, + 369 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 423 + ], + "type": "text", + "content": "For comparisons, we select SOTA baselines from different branches of work that can perform object insertion. To ensure the fairness, we barely modify the baselines, only when necessary to fit them with our settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 288, + 495 + ], + "type": "text", + "content": "Traditional 3D editing pipeline. This pipeline creates a 3D model from multi-view images. Scene editing is then performed directly on the 3D model. To simulate this traditional pipeline, we use COLMAP [54, 55] to reconstruct the mesh for the background and object, and then manually crop and place the object mesh into the background mesh." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 495, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 288, + 544 + ], + "type": "text", + "content": "Image inpainting. We adopt the inpainting variant of Stable Diffusion [51] as a baseline. However, since image inpainting treats each view independently, for a fair comparison, we perform single-view inpainting on each background image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 288, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 591 + ], + "type": "text", + "content": "Single-image-to-3D NeRF. NeRF-based view synthesis partially fulfills scene editing. Here we select Zero123 [34], a SOTA that uses a distillation prior [47] from a 2D diffusion model to synthesize novel views for comparison." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 591, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 288, + 651 + ], + "type": "text", + "content": "Instruct-NeRF2NeRF [14]. This work also applies dataset updates for NeRF training. However, it fails to add/remove objects with noticeable non-uniform appearance. To test this method, we re-format text prompts, e.g., we replace \"a * sneaker on a * table\" by \"add/make a sneaker on a table\"." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 178, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 178, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 178, + 671 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We fine-tune the diffusion model " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " with object images " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "(n_{obj} = 5,000" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " times) more than background images " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "(n_{bg} = 500" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " times) as we found the model needs extra training to" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "learn objects with complex geometry and texture. For NeRF training, we empirically found " + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "inline_equation", + "content": "n_{near} = 3" + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "inline_equation", + "content": "n_{new} = 500" + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "inline_equation", + "content": "n_{old} = 10" + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": " balance well the quality and efficiency." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 108, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 180 + ], + "type": "text", + "content": "We run all experiments on a Nvidia RTX 3090 GPU. Diffusion fine-tuning takes around 30 minutes. NeRF updating speed relies on the number of background images while each backpropagation takes 0.5 seconds. Inference speed of the diffusion model is 8 seconds/image. Our NeRF training costs around 2.7 times more than the standard NeRF training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 186, + 414, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 414, + 198 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 414, + 198 + ], + "type": "text", + "content": "4.4. Qualitative results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 205, + 547, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 547, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 547, + 264 + ], + "type": "text", + "content": "Object insertion. We present qualitative results of object insertion in Fig. 3. As shown, our method (Fig. 3-f) can generate plausible contents with view consistency. Moverover, the outputs of our method also match well the text prompts and the generated objects are precisely located." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 265, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 547, + 408 + ], + "type": "text", + "content": "In contrast, the traditional 3D pipeline (Fig. 3-b) suffers from seamed object-background boundaries and unrealistic lighting. The geometry in obscure regions, e.g., corners, is not accurately reconstructed. The image inpainting baseline (Fig. 3-c) produces reasonable results on individual views but generates view-inconsistency. Single-image-to-3D NeRF (Fig. 3-d) fails on complex scenes. A larger camera pose shift can result in background mismatch or even content collapse. Instruct-NeRF2NeRF (Fig. 3-e) is known to be strong at stylizing existing geometry but weak at generating new geometry. Under our setting, it performs random view-consistent editing, but fails to accomplish object insertion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 408, + 547, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 480 + ], + "type": "text", + "content": "Object removal. We illustrate several results of object removal in Fig. 4, which shows that our method can generate view-consistent backgrounds. We also observed that without using pseudo ground-truth background, the removal can cause gradual background collapse, which we discuss further in our ablation study in Sec. 4.6." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 487, + 422, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 422, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 422, + 499 + ], + "type": "text", + "content": "4.5. Quantitative results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "content": "Since there is no real-world ground-truth for scene editing, direct quantitative evaluations are not possible. Instead, we adopt the CLIP Score [16], a well-known metric to measure how well an image correlates to a target prompt in the CLIP space. Let " + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "inline_equation", + "content": "E(I)" + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "inline_equation", + "content": "E(p)" + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "content": " be the CLIP embeddings of an image " + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "content": " and a text prompt " + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 505, + 547, + 578 + ], + "type": "text", + "content": ". The CLIP Score is defined as," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 584, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 584, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 345, + 584, + 545, + 597 + ], + "type": "interline_equation", + "content": "\\operatorname {C L I P S c o r e} (I, p) = \\cos^ {+} (E (I), E (p)) \\tag {8}", + "image_path": "ad104a69bbcf6c69fd3df3be7e3871af88b2dd91efe4731f7c2046c1ec4edbf3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\cos^{+}(a,b) = \\max (0,\\cos (a,b))" + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": ". We average the CLIPScore " + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "inline_equation", + "content": "(\\hat{I}_b,\\tilde{p})" + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": " for all the edited images " + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\hat{I}_b\\in \\hat{\\mathbf{I}}_b" + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": " paired with their target prompts " + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": ", and use this average value as a performance metric for scene editing; higher score means better performance." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We also use the CLIP Directional Consistency (CLIPDC) in [14] to measure the editing quality and consistency across views in the CLIP space. This metric relies on an assumption that the difference in the CLIP space between a background" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5180" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 68, + 547, + 562 + ], + "blocks": [ + { + "bbox": [ + 49, + 68, + 547, + 562 + ], + "lines": [ + { + "bbox": [ + 49, + 68, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 49, + 68, + 547, + 562 + ], + "type": "image", + "image_path": "7a469988be365c00da5df2342c1bea6c66583c45b8d8940871a4f7260d82837e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 571, + 547, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 571, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 547, + 594 + ], + "type": "text", + "content": "Figure 3. Qualitative results of object insertion. Inputs include multi-view object/background images, a 3D bounding box where the object is inserted in, and a text prompt. Note that some baselines use parts of the inputs due to the nature of their techniques." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": " and its edited version " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\hat{I}_{b}" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": " should match with the difference between the original prompt " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": " (paired with " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": ") and the customized prompt " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": " (paired with " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\hat{I}_b" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": "). Moreover, a good editing should make consistent differences between every pair of " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "inline_equation", + "content": "\\hat{I}_b" + }, + { + "bbox": [ + 46, + 613, + 289, + 687 + ], + "type": "text", + "content": ", especially for adjacent views. We follow [14] to calculate the CLIPDC between two adjacent" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "text", + "content": "edited views " + }, + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "inline_equation", + "content": "\\hat{I}_b^{(i)}" + }, + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "inline_equation", + "content": "\\hat{I}_b^{(i + 1)}" + }, + { + "bbox": [ + 306, + 612, + 434, + 628 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 311, + 634, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 634, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 311, + 634, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {C L I P D C} \\left(\\hat {I} _ {b} ^ {(i)}, \\hat {I} _ {b} ^ {(i + 1)}\\right) = \\\\ \\cos^ {+} (E (\\tilde {p}) - E (p), E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)})) \\times \\tag {9} \\\\ \\cos^ {+} \\left(E (\\hat {I} _ {b} ^ {(i)}) - E (I _ {b} ^ {(i)}), E (\\hat {I} _ {b} ^ {(i + 1)}) - E (I _ {b} ^ {(i + 1)})\\right) \\\\ \\end{array}", + "image_path": "186bc699252af5e7638bb37b14ec01efff680e5fab92cffb75e1c5450cc57af4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "We measure the CLIPDC of an edited scene by averaging" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5181" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 126, + 148 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 126, + 148 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 126, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 126, + 148 + ], + "type": "image", + "image_path": "b5be5d30166ec4c05c787446d7a26e3f009183097c893855cd6f3e1f5e9ea4e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 206, + 148 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 206, + 148 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 206, + 148 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 206, + 148 + ], + "type": "image", + "image_path": "6890a1aea494ff605eac3c30bbeadbf2bb41848d5ed6e9d18e359022a0aaf9d5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 287, + 148 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 287, + 148 + ], + "type": "image", + "image_path": "721697224708c1135dd2f92bd36eabd8b0f2fab0bdc698c6780dd16fe247148f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 149, + 126, + 227 + ], + "blocks": [ + { + "bbox": [ + 49, + 149, + 126, + 227 + ], + "lines": [ + { + "bbox": [ + 49, + 149, + 126, + 227 + ], + "spans": [ + { + "bbox": [ + 49, + 149, + 126, + 227 + ], + "type": "image", + "image_path": "17a93be35b9bea6ed344c8fa0029092af4a60b57ad719be707571140a4a0f7a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 129, + 149, + 206, + 227 + ], + "blocks": [ + { + "bbox": [ + 129, + 149, + 206, + 227 + ], + "lines": [ + { + "bbox": [ + 129, + 149, + 206, + 227 + ], + "spans": [ + { + "bbox": [ + 129, + 149, + 206, + 227 + ], + "type": "image", + "image_path": "f9c628322d00cfa69b9153551f1aaf37555778ef9b008c09a670373f22c205c9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 149, + 287, + 227 + ], + "blocks": [ + { + "bbox": [ + 209, + 149, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 209, + 149, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 209, + 149, + 287, + 227 + ], + "type": "image", + "image_path": "c30d90367c851029730c85d89ed9e5c35b42b25f4106b465dac5a18202928f78.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 232, + 126, + 309 + ], + "blocks": [ + { + "bbox": [ + 49, + 232, + 126, + 309 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 126, + 309 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 126, + 309 + ], + "type": "image", + "image_path": "746b293bc1723199285da4cdf9baa3de98bfe0fa8598f57f62711b07a152b6a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 129, + 232, + 206, + 309 + ], + "blocks": [ + { + "bbox": [ + 129, + 232, + 206, + 309 + ], + "lines": [ + { + "bbox": [ + 129, + 232, + 206, + 309 + ], + "spans": [ + { + "bbox": [ + 129, + 232, + 206, + 309 + ], + "type": "image", + "image_path": "faa48321ebbe27ee602177623a2a2e1fd533191585dac98cd8e06bb6a35893fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 209, + 232, + 287, + 309 + ], + "blocks": [ + { + "bbox": [ + 209, + 232, + 287, + 309 + ], + "lines": [ + { + "bbox": [ + 209, + 232, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 209, + 232, + 287, + 309 + ], + "type": "image", + "image_path": "ae6aafa184f599ed0e73e019d6a6ab82988fed8f3d5c61337a2963919fe48bb9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 311, + 126, + 388 + ], + "blocks": [ + { + "bbox": [ + 49, + 311, + 126, + 388 + ], + "lines": [ + { + "bbox": [ + 49, + 311, + 126, + 388 + ], + "spans": [ + { + "bbox": [ + 49, + 311, + 126, + 388 + ], + "type": "image", + "image_path": "fb01b77239be06ed66a255a408b84c063d4077750c46b86ae16dcbf066e2539a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 390, + 126, + 409 + ], + "lines": [ + { + "bbox": [ + 47, + 390, + 126, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 126, + 409 + ], + "type": "text", + "content": "(a) Original BG with object bounding box" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 129, + 311, + 206, + 388 + ], + "blocks": [ + { + "bbox": [ + 129, + 311, + 206, + 388 + ], + "lines": [ + { + "bbox": [ + 129, + 311, + 206, + 388 + ], + "spans": [ + { + "bbox": [ + 129, + 311, + 206, + 388 + ], + "type": "image", + "image_path": "f6ccf694c33f8ea3b986f7b91ce8338dd465bc4c3a6c32be2dc8cdff0a2d9234.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 390, + 206, + 409 + ], + "lines": [ + { + "bbox": [ + 129, + 390, + 206, + 409 + ], + "spans": [ + { + "bbox": [ + 129, + 390, + 206, + 409 + ], + "type": "text", + "content": "(b) Pseudo ground-truth BG" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 209, + 311, + 287, + 388 + ], + "blocks": [ + { + "bbox": [ + 209, + 311, + 287, + 388 + ], + "lines": [ + { + "bbox": [ + 209, + 311, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 209, + 311, + 287, + 388 + ], + "type": "image", + "image_path": "4747e311bac08754a3a78caa193274b3524594a8b95950634bb317b38495b81a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 390, + 287, + 409 + ], + "lines": [ + { + "bbox": [ + 209, + 390, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 209, + 390, + 287, + 409 + ], + "type": "text", + "content": "(c) View-consistent editing result" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 418, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 452 + ], + "type": "text", + "content": "Figure 4. Qualitative results of object removal where we show the importance of pseudo ground-truth background (BG) in generating view-consistent editing results." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 48, + 474, + 287, + 564 + ], + "blocks": [ + { + "bbox": [ + 48, + 474, + 287, + 564 + ], + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 564 + ], + "type": "table", + "html": "
MethodCLIPScore ↑CLIPDC ↑
Traditional 3D [54, 55]0.26480.1169
Inpainting [51]0.26830.1325
Zero123 [34]0.21970.0361
Instruct-NeRF2NeRF [14]0.23470.0263
Ours0.27430.1642
", + "image_path": "0f893b151b2dd4cc8ff6f372a558cf434788234d82aab8145f684ecad35285cd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "lines": [ + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 572, + 287, + 595 + ], + "type": "text", + "content": "Table 1. Quantitative results. Higher CLIP metrics scores indicate higher image editing quality or consistency." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "text", + "content": "the CLIPDC " + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})" + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "text", + "content": " for all the adjacent image pairs " + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\hat{I}_b^{(i)},\\hat{I}_b^{(i + 1)})" + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "text", + "content": " in the scene. We report both the CLIPScore and CLIPDC of our method and other baselines on 35 unique edits (cross editing on the more generalizable 5 scenes " + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 624, + 287, + 713 + ], + "type": "text", + "content": " 7 objects) in Tab. 1. Experimental show that our method outperforms all the baselines on both the CLIPScore and CLIPDC metrics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 400, + 85 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "text", + "content": "In this ablation study, we validate the effectiveness of technical components in our method. We suggest readers observe the results in Fig. 5, which shows two views of an edited scene. We provide detailed analyses on these results below. Diffusion model fine-tuning. Recall that we fine-tune the diffusion model " + }, + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "text", + "content": " on both object and background images (in Sec. 3.2). Here we prove that such a fine-tuning is necessary. Fine-tuning on object images makes the diffusion model aware of the same object during the dataset updates (in Sec. 3.3). We verify this in Fig. 5-a, where we skip the fine-tuning of " + }, + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 304, + 90, + 547, + 246 + ], + "type": "text", + "content": " on object images. As shown, without using object images, the diffusion model cannot generate the same object across views." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 378 + ], + "type": "text", + "content": "Likewise, fine-tuning the diffusion model with background images helps to preserve the background at inpainted borders (areas in between mask boundary and inner-mask object boundary). As shown in Fig. 5-b, without fine-tuning on background images, the model cannot inpaint the background within the mask region properly. The inpainted background gets darker at latter NeRF updating steps, and finally ends up with a noticeable error. We hypothesize this collapse that the model, without being fine-tuned on relevant background images, applies the bias from pre-learned knowledge which does not align with the surrounding background." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 378, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 547, + 474 + ], + "type": "text", + "content": "Pose-conditioned dataset updates. To prove the effectiveness of the pose-conditioned dataset update strategy, we experiment with another scheme in which the dataset used to refine the NeRF is updated with random views. We present the results of this experiment in Fig. 5-c. We observe that, with view-random updates, objects in different views can converge into inconsistent poses (e.g., the hats in the left and right view are generated in different poses)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": "The above inconsistency commonly happens with many text-to-3D NeRF generation or editing methods. We hypothesize this phenomenon as follows. Initially the diffusion model can generate the object in inconsistent poses across views due to little 3D-aware hint available. This inconsistency is continually introduced to the NeRF updating and then, in return, passed partially as input to the diffusion model (as in Eq. 6), making further pose divergence eventually. Without considering nearby-views, the diffusion model cannot fix this inconsistency as objects generated on individual views match well their given text prompts." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "Our proposed pose-conditioned dataset update strategy simulates the nature of NeRF construction, in which a view rendered closed to already-used views should contain similar but slightly blurry and distorted object content. Pose-conditioned view arrangement thus gives the diffusion model enough hints to generate objects with view-consistent appearance and poses, specified in nearby views. View-consistency is thus achieved progressively in the training dataset and is integrated into the constructed NeRF (see Fig. 5-e)." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5182" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 134 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 134 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 134 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 134 + ], + "type": "image", + "image_path": "89ffbaa2f17dac027e3910a281c3160d38f308d311280cdf06985953c9dde8d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 135, + 403, + 144 + ], + "lines": [ + { + "bbox": [ + 189, + 135, + 403, + 144 + ], + "spans": [ + { + "bbox": [ + 189, + 135, + 403, + 144 + ], + "type": "text", + "content": "(a) Fine-tuning of the diffusion model without using object images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 145, + 547, + 209 + ], + "blocks": [ + { + "bbox": [ + 49, + 145, + 547, + 209 + ], + "lines": [ + { + "bbox": [ + 49, + 145, + 547, + 209 + ], + "spans": [ + { + "bbox": [ + 49, + 145, + 547, + 209 + ], + "type": "image", + "image_path": "0b52ea12b2c4b0590e356c944e1ddf15f2ab193f2255d38c8e4ebcd84efa6ab6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 220, + 547, + 284 + ], + "blocks": [ + { + "bbox": [ + 180, + 210, + 413, + 220 + ], + "lines": [ + { + "bbox": [ + 180, + 210, + 413, + 220 + ], + "spans": [ + { + "bbox": [ + 180, + 210, + 413, + 220 + ], + "type": "text", + "content": "(b) Fine-tuning of the diffusion model without using background images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 220, + 547, + 284 + ], + "lines": [ + { + "bbox": [ + 49, + 220, + 547, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 547, + 284 + ], + "type": "image", + "image_path": "2aadfa0bfdb4792f546191664180902d62e6a2a9c7f9613efeb8923906798521.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 295, + 547, + 360 + ], + "blocks": [ + { + "bbox": [ + 177, + 285, + 416, + 295 + ], + "lines": [ + { + "bbox": [ + 177, + 285, + 416, + 295 + ], + "spans": [ + { + "bbox": [ + 177, + 285, + 416, + 295 + ], + "type": "text", + "content": "(c) Dataset updates with random poses (without pose-conditioned updates)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 295, + 547, + 360 + ], + "lines": [ + { + "bbox": [ + 49, + 295, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 547, + 360 + ], + "type": "image", + "image_path": "c488f96364b847254e7408a119a8342ea5908ec5ffd4fab9e633e81864f72ff9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 361, + 355, + 369 + ], + "lines": [ + { + "bbox": [ + 238, + 361, + 355, + 369 + ], + "spans": [ + { + "bbox": [ + 238, + 361, + 355, + 369 + ], + "type": "text", + "content": "(d) Without periodic dataset updates" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 370, + 547, + 434 + ], + "blocks": [ + { + "bbox": [ + 49, + 370, + 547, + 434 + ], + "lines": [ + { + "bbox": [ + 49, + 370, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 49, + 370, + 547, + 434 + ], + "type": "image", + "image_path": "95cd1e08b02da26e3393e9d9b06ac1405b1d31e15d6171b0a56bbd513feebaad.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 436, + 323, + 445 + ], + "lines": [ + { + "bbox": [ + 271, + 436, + 323, + 445 + ], + "spans": [ + { + "bbox": [ + 271, + 436, + 323, + 445 + ], + "type": "text", + "content": "(e) Full pipeline" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 454, + 547, + 499 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 547, + 499 + ], + "type": "text", + "content": "Figure 5. Ablation study results. Each row shows the results of a variant of our pipeline. The left and right columns include images of two different views of an edited scene. For each column, from left to right are the results of increasing training steps, where the most left image in each column is an early-stage result and the most right is the final output. The left column is a view near the starting view, which converges faster than the right column from a farther view (except for the variant in (c) where both views converge equally fast)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 520, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 289, + 640 + ], + "type": "text", + "content": "Periodic dataset updates. We observe that periodically updating of training views during the NeRF updating is important to achieve high-quality renderings. We verify this by keeping all the training views from the pose-conditioned dataset updates fixed during the NeRF updating. We found that although objects are rendered at fairly accurate orientations, their texture and geometry still suffer from inconsistencies (see Fig. 5-d). Having periodically updates on the training data fixes these minor defects and facilitates a more consistent convergence of the NeRF updating (see Fig. 5-e)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 655, + 119, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 655, + 119, + 667 + ], + "spans": [ + { + "bbox": [ + 47, + 655, + 119, + 667 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "We propose a new language-driven method for object manipulation in NeRFs. Our method is built upon a novel idea of joint 2D-3D interaction, keeping both 2D image synthesis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 581 + ], + "type": "text", + "content": "and 3D NeRF reconstruction in a loop. This idea is enabled by an advanced text-to-image diffusion technique that generates object-blended background images, and a novel pose-conditioned dataset update strategy that learns a NeRF from the multi-view images in a progressive manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 581, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 581, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 581, + 547, + 713 + ], + "type": "text", + "content": "Our method is not without limitations. Since our 2D views are synthesized by a diffusion model, we may share the flickering problem with diffusion-based video editing methods [26, 48, 68]. We leave this for future work, which can potentially be addressed by robust video translation methods [73]. It is also of great interest to postulate a theoretical foundation for pose-conditioned dataset update to better understand the convergence of NeRF training in scene editing. Acknowledgment. This work was partially supported by a grant from the RGC of HKSAR, China (Project No. HKUST 16202323) and an internal grant from HKUST (R9429)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5183" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 156 + ], + "type": "text", + "content": "[1] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 222 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 222 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 280 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5479, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 282, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 288, + 335 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 338, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 338, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 338, + 287, + 360 + ], + "type": "text", + "content": "[5] Ronald A Castellino. Computer aided detection (cad): an overview. Cancer Imaging, 5(1):17, 2005. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 361, + 288, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 361, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 288, + 436 + ], + "type": "text", + "content": "[6] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 53, + 440, + 287, + 483 + ], + "type": "text", + "content": "[7] Dave Zhenyu Chen, Yawar Siddiqui, Hsin-Ying Lee, Sergey Tulyakov, and Matthias Nießner. Text2tex: Text-driven texture synthesis via diffusion models. arXiv preprint arXiv:2303.11396, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 485, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 485, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 485, + 288, + 529 + ], + "type": "text", + "content": "[8] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 531, + 287, + 563 + ], + "type": "text", + "content": "[9] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 565, + 179, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 179, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 179, + 575 + ], + "type": "text", + "content": "[10] Epic Games. Unreal engine. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 631 + ], + "type": "text", + "content": "[11] SM Ali Eslami, Danilo Jimenez Rezende, Frederic Besse, Fabio Viola, Ari S Morcos, Marta Garnelo, Avraham Ruderman, Andrei A Rusu, Ivo Danihelka, Karol Gregor, et al. Neural scene representation and rendering. Science, 360 (6394):1204-1210, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 679, + 280, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 280, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 280, + 689 + ], + "type": "text", + "content": "[13] John K Haas. A history of the unity game engine. 2014. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 546, + 94 + ], + "type": "text", + "content": "3d scenes with instructions. arXiv preprint arXiv:2303.12789, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 116 + ], + "type": "text", + "content": "[15] Jeffrey Harper. Mastering Autodesk 3ds Max 2013. John Wiley & Sons, 2012. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 118, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 160 + ], + "type": "text", + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 162, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 547, + 195 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 196, + 547, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 196, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 547, + 248 + ], + "type": "text", + "content": "[18] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multiview stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 250, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 546, + 304 + ], + "type": "text", + "content": "[19] Yi-Hua Huang, Yue He, Yu-Jie Yuan, Yu-Kun Lai, and Lin Gao. Stylizednerf: consistent 3d scene stylization as stylized nerf via 2d-3d mutual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18342–18352, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 305, + 546, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 546, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 546, + 349 + ], + "type": "text", + "content": "[20] Yi-Hua Huang, Yan-Pei Cao, Yu-Kun Lai, Ying Shan, and Lin Gao. Nerf-texture: Texture synthesis with neural radiance fields. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 547, + 392 + ], + "type": "text", + "content": "[21] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 394, + 546, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 546, + 437 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 546, + 481 + ], + "type": "text", + "content": "[23] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. Advances in neural information processing systems, 33:12104-12114, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 482, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 546, + 536 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 546, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 546, + 592 + ], + "type": "text", + "content": "[25] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imag: Text-based real image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6007-6017, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 593, + 546, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 546, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 546, + 657 + ], + "type": "text", + "content": "[26] Gyeongman Kim, Hajin Shim, Hyunsu Kim, Yunjey Choi, Junho Kim, and Eunho Yang. Diffusion video autoencoders: Toward temporally consistent face video editing via disentangled video encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6091-6100, 2023. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 546, + 690 + ], + "type": "text", + "content": "[27] Hyunsu Kim, Gayoung Lee, Yunjay Choi, Jin-Hwa Kim, and Jun-Yan Zhu. 3d-aware blending with generative nerfs. arXiv preprint arXiv:2302.06608, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 546, + 713 + ], + "type": "text", + "content": "[28] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing nef for editing via feature field distillation." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5184" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 288, + 95 + ], + "type": "text", + "content": "Advances in Neural Information Processing Systems, 35: 23311-23330, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 139 + ], + "type": "text", + "content": "[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 186 + ], + "type": "text", + "content": "[30] Gang Li, Heliang Zheng, Chaoyue Wang, Chang Li, Changwen Zheng, and Dacheng Tao. 3ddesigner: Towards photorealistic 3d object generation and editing with text-guided diffusion models. arXiv preprint arXiv:2211.14108, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 231 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 231 + ], + "type": "text", + "content": "[31] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. Focaldreamer: Text-driven 3d editing via focal-fusion assembly. arXiv preprint arXiv:2308.10608, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 288, + 297 + ], + "type": "text", + "content": "[32] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300–309, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 353 + ], + "type": "text", + "content": "[33] Miaomiao Liu, Xuming He, and Mathieu Salzmann. Geometry-aware deep network for single-image novel view synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4616-4624, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 288, + 398 + ], + "type": "text", + "content": "[34] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. arXiv preprint arXiv:2303.11328, 2023.5,6,7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 444 + ], + "type": "text", + "content": "[35] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Jun-Yan Zhu, and Bryan Russell. Editing conditional radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5773-5783, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 445, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 509 + ], + "type": "text", + "content": "[36] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 578 + ], + "type": "text", + "content": "[37] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7210-7219, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "text", + "content": "[38] Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8446-8455, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[39] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "type": "text", + "content": "[41] Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 130, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 547, + 206 + ], + "type": "text", + "content": "[42] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein. Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20669–20679, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 547, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 547, + 251 + ], + "type": "text", + "content": "[43] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics (ToG), 41(4):1-15, 2022. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 285 + ], + "type": "text", + "content": "[44] Thu Nguyen-Phuoc, Feng Liu, and Lei Xiao. Snerf: stylized neural implicit representations for 3d scenes. arXiv preprint arXiv:2207.02363, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 287, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 547, + 352 + ], + "type": "text", + "content": "[45] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13503–13513, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 354, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 376 + ], + "type": "text", + "content": "[46] Hong-Wing Pang, Binh-Son Hua, and Sai-Kit Yeung. Locally stylized neural radiance fields. In ICCV, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 410 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 411, + 545, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 456 + ], + "type": "text", + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 547, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 547, + 567 + ], + "type": "text", + "content": "[50] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 568, + 547, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 547, + 624 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 547, + 689 + ], + "type": "text", + "content": "[52] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 547, + 713 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5185" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 151 + ], + "type": "text", + "content": "[54] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[55] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 241 + ], + "type": "text", + "content": "[56] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. Advances in Neural Information Processing Systems, 35:33999-34011, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 288, + 285 + ], + "type": "text", + "content": "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. Advances in Neural Information Processing Systems, 35:24487-24501, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 342 + ], + "type": "text", + "content": "[58] Hyeonseop Song, Seokhun Choi, Hoseok Do, Chul Lee, and Taehyeong Kim. Blending-nerf: Text-driven localized editing in neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14383–14393, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 344, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 344, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 344, + 288, + 408 + ], + "type": "text", + "content": "[59] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, et al. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-12, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "type": "text", + "content": "[60] Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Multiview consistency as supervisory signal for learning shape and pose prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2897-2905, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 521 + ], + "type": "text", + "content": "[61] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 567 + ], + "type": "text", + "content": "[62] Can Wang, Ruixiang Jiang, Mengei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neural radiance fields stylization. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 611 + ], + "type": "text", + "content": "[63] Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine Susstrunk. Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields. arXiv preprint arXiv:2305.15094, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 678 + ], + "type": "text", + "content": "[64] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8798-8807, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 139 + ], + "type": "text", + "content": "[66] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 140, + 547, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 203 + ], + "type": "text", + "content": "[67] Silvan Weder, Guillermo Garcia-Hernando, Aron Monszpart, Marc Pollefeys, Gabriel J Brostow, Michael Firman, and Sara Vicente. Removing objects from neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16528-16538, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 547, + 259 + ], + "type": "text", + "content": "[68] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Weixian Lei, Yuchao Gu, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. arXiv preprint arXiv:2212.11565, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 547, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 547, + 314 + ], + "type": "text", + "content": "[69] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, and Lin Gao. De-nerf: Decoupled neural radiance fields for view-consistent appearance editing and high-frequency environmental relighting. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1–11, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 547, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 547, + 370 + ], + "type": "text", + "content": "[70] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 372, + 547, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 547, + 436 + ], + "type": "text", + "content": "[71] Jiale Xu, Xintao Wang, Weihao Cheng, Yan-Pei Cao, Ying Shan, Xiaohu Qie, and Shenghua Gao. Dream3d: Zero-shot text-to-3d synthesis using 3d shape prior and text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20908–20918, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 437, + 547, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 547, + 514 + ], + "type": "text", + "content": "[72] Yinghao Xu, Menglei Chai, Zifan Shi, Sida Peng, Ivan Skorokhodov, Aliaksandr Siarohin, Ceyuan Yang, Yujun Shen, Hsin-Ying Lee, Bolei Zhou, et al. Discoscope: Spatially disentangled generative radiance fields for controllable 3d-aware scene synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4402-4412, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 515, + 547, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 547, + 547 + ], + "type": "text", + "content": "[73] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. *Rerender a video: Zero-shot text-guided video-to-video translation.* arXiv preprint arXiv:2306.07954, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 547, + 592 + ], + "type": "text", + "content": "[74] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 593, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 547, + 646 + ], + "type": "text", + "content": "[75] Yu-Jie Yuan, Yang-Tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. Nerf-editing: geometry editing of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18353-18364, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 547, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 547, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 547, + 680 + ], + "type": "text", + "content": "[76] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 547, + 713 + ], + "type": "text", + "content": "[77] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5186" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 285 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[78] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (ToG), 40(6):1-18, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 194 + ], + "type": "text", + "content": "[79] Tinghui Zhou, Shubham Tulsiani, Weilun Sun, Jitendra Malik, and Alexei A Efros. View synthesis by appearance flow. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 286-301. Springer, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 196, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 196, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 196, + 288, + 250 + ], + "type": "text", + "content": "[80] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 285 + ], + "type": "text", + "content": "[81] Jingyu Zhuang, Chen Wang, Lingjie Liu, Liang Lin, and Guanbin Li. Dreameditor: Text-driven 3d scene editing with neural fields. arXiv preprint arXiv:2306.13455, 2023. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5187" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_content_list.json b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1b83c853c96e243879e8208e00585b4f64255e05 --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_content_list.json @@ -0,0 +1,1505 @@ +[ + { + "type": "text", + "text": "Language-guided Image Reflection Separation", + "text_level": 1, + "bbox": [ + 250, + 131, + 720, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haofeng Zhong $^{1,2,3\\#}$ Yuchen Hong $^{1,2\\#}$ Shuchen Weng $^{1,2}$ Jinxiu Liang $^{1,2}$ Boxin Shi $^{1,2,3*}$", + "bbox": [ + 114, + 166, + 852, + 186 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 143, + 186, + 826, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com", + "bbox": [ + 117, + 217, + 846, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 267, + 313, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper studies the problem of language-guided reflection separation, which aims at addressing the ill-posed reflection separation problem by introducing language descriptions to provide layer content. We propose a unified framework to solve this problem, which leverages the crossattention mechanism with contrastive learning strategies to construct the correspondence between language descriptions and image layers. A gated network design and a randomized training strategy are employed to tackle the recognizable layer ambiguity. The effectiveness of the proposed method is validated by the significant performance advantage over existing reflection separation methods on both quantitative and qualitative comparisons.", + "bbox": [ + 73, + 300, + 473, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 527, + 209, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When photographing through transparent materials like glass windows or showcases, the presence of reflections can significantly degrade the image quality of captured images and disrupt downstream computer vision tasks like face recognition [53] or depth estimation [2]. As an attractive topic in computational photography, reflection separation aims at decomposing the contaminated mixture image (denoted as $\\mathbf{M}$ ) into two components that correspond to scenes located at different sides of the glass, i.e., the reflection layer (denoted as $\\mathbf{R}$ ) and the transmission layer (denoted as $\\mathbf{T}$ ). Since reflection separation is a severely ill-posed problem, it is imperative to exploit effective priors or auxiliary information for distinguishing the two components.", + "bbox": [ + 75, + 553, + 468, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary challenge of solving the reflection separation problem lies in the exploration of distinct clues for distinguishing transmission and reflection layers. Multi-image methods handle this problem by introducing additional constraints. Some of them acquire a series of mixture images in different viewpoints [32, 34, 45] to harness the distinct motions of the two layers, while others adopt specialized capturing setups to obtain complementary scene informa", + "bbox": [ + 75, + 751, + 468, + 872 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f45362c64c63d2eacf3472b97a8ae2ef172e097b9bbc0b323aafe6087713a14b.jpg", + "image_caption": [ + "Figure 1. The recognizable layer ambiguity problem causes uncertain quantities of input language descriptions for language-guided image reflection separation. Given language descriptions of either (a) one layer or (b) both two layers, the proposed method achieves robust reflection separation compared with an existing reflection separation method [10]." + ], + "image_footnote": [], + "bbox": [ + 503, + 267, + 890, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [19, 27, 28, 36]. However, the specialized data capture requirements limit the application scope of these methods, especially for images downloaded from the Internet. Single-image methods attempt to tackle reflection separation by utilizing handcrafted priors derived from natural image statistics [30, 33, 44] or leveraging the modeling capacity of neural networks to learn content priors about reflections from a large scale of training data [10, 22, 31]. However, they are prone to fail due to the lack of auxiliary content information about transmission or reflection scenes for solving such a highly ill-posed problem. Recently, language descriptions have shown their effectiveness in providing content information for various vision tasks such as image editing [46, 47, 60, 61], semantic segmentation [57, 65], and image colorization [4-6, 62], which inspires us to think about: Can we leverage the auxiliary content information brought by language descriptions to facilitate the reflection separation problem?", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\# Equal contributions. *Corresponding author.", + "bbox": [ + 91, + 883, + 338, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24913", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Since language can effectively convey humans' prior knowledge about the real world [8] and provide auxiliary information of image semantics [65], introducing language descriptions to guide the separation of reflection and transmission layers from mixture images merits exploration. However, leveraging language descriptions for reflection separation is non-trivial in three aspects: 1) Language-image modality inconsistency. Language and images belong to different modalities, thus it is challenging to establish a cross-modality correspondence between the scene content information provided in language descriptions and the complex blended content present in mixture images. 2) Recognizable layer ambiguity. Since the image content and brightness of reflection and transmission layers are different, the recognizable extents of them in mixture images are also uncertain. Specifically, as shown in Figure 1, it is possible that only one layer's content is recognizable (clothes in Figure 1(a)), or both two layers exhibit recognizable content (the trashbin and cars in Figure 1(b)), which leads to the difficulty of using uncertain quantities of language descriptions for separating mixture images in practice. 3) Language annotation deficiency. All existing datasets for the reflection separation task only contain image data but no correlated language description is provided, raising the challenge for network training and evaluation.", + "bbox": [ + 76, + 90, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce the concept of language-guided image reflection separation for the first time, which leverages flexible natural language to specify the content of one or two layers within a mixture image, relieving the ill-posedness of the reflection separation problem and maintaining a wide applicability for both live captured or online downloaded mixture images. We propose an end-to-end framework that employs adaptive global interaction modules to explore holistic language-image content coherence and utilizes specifically designed loss functions to constrain the correspondence between language descriptions and recovered image layers. A language gate mechanism and a randomized training strategy are designed to deal with the recognizable layer ambiguity problem. To address the language annotation deficiency, we synthesize the training dataset from paired image-language datasets [7, 66] and expand prevailing real reflection separation datasets [31, 49, 67] by manually adding language descriptions. Besides, we further construct a new dataset for visual quality evaluation by collecting mixture images from the Internet and captioning them with language descriptions for recognizable layers. Our contributions are summarized as follows:", + "bbox": [ + 76, + 473, + 472, + 805 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first work that introduces language descriptions to guide the reflection separation task.", + "- We propose adaptive global interaction modules and language-image loss functions to tackle modality inconsistency.", + "- We design a language gate mechanism and a randomized" + ], + "bbox": [ + 76, + 810, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training strategy to handle recognizable layer ambiguity.", + "bbox": [ + 511, + 90, + 883, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We build a dataset with language descriptions to facilitate language-guided image reflection separation.", + "bbox": [ + 501, + 107, + 888, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 150, + 637, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single-image reflection separation methods try to distinguish reflection and transmission layers using a single mixture image, which mainly relies on the assumption that the two layers have different distributions, i.e., reflection layers are more likely to be blurry and appear with lower intensity compared with transmission layers. Conventional methods adopt handcrafted priors derived from natural image statistics in their optimization process, e.g., the gradient sparsity [30], relative smoothness [33], ghosting cues [44], content prior [50], and penalty on the gradient of restored transmission layers [64].", + "bbox": [ + 496, + 176, + 890, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the tremendous progress in the field of deep learning, a series of single-image reflection separation methods concentrate on the improvement of learning strategies or network design, e.g., predicting edges and images with a two-stage [11] or concurrent framework [51, 52], training with the perceptual loss [67], employing generative adversarial network [12] based models [38, 58], adopting iterative refinement strategies [10, 31, 63, 68], and leveraging the complementary two-stream architecture [22]. Meanwhile, research on data synthesis and image models is also ongoing to satisfy the data-driven needs of learning-based methods. Ma et al. [38] utilize generative adversarial networks for data generation while Wen et al. [59] synthesize mixture images with learned non-linear blending masks. Hu et al. [23] introduce a learnable residue term in the mixture image formation model to mitigate the non-linearity caused by the complex camera pipeline. Zheng et al. consider physical factors such as reflective amplitude coefficient maps [69] and the absorption effect [70] in the image formation process of mixture images. To facilitate network training and evaluation, researchers [31, 49, 67] also collect real data by using portable glass. Moreover, as a special form of images, panoramic images are introduced to relieve the content ambiguity in mixture images [13, 19, 21, 40]. We refer readers to [54] for a comprehensive and up-to-date survey on single-image reflection separation.", + "bbox": [ + 496, + 342, + 892, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-image reflection separation methods usually leverage the auxiliary information introduced by additional images and achieve more robust performance than single-image methods. Polarization-based methods [9, 26, 28, 36, 37, 39, 43] distinguish reflection and transmission layers by using images captured with different angles of polarizers or special polarization cameras. Flash-based methods [3, 18, 20, 27, 29] adopt active light sources to illuminate transmission scenes for obtaining reflection-free guidance. Motion-based methods [32, 34, 35, 45] utilize multiple images captured from different viewpoints to harness", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24914", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f8d20ba08f8af659fe038800ae98c653d32eb32f2488da12dd6d017a5ad91a67.jpg", + "image_caption": [ + "Figure 2. The pipeline of the proposed language-guided image reflection separation framework, which extracts features from mixture images and available language descriptions (the description $L_{2}$ with dashed lines is possible to be set to null due to the recognizable layer ambiguity) by image and language encoders (in Sec. 3.1), aggregates global visual information by adaptive global aggregation modules (AGAM) and conducts progressive interactions to exploit distinctive image features with gated language guidance by adaptive global interaction modules (AGIM) (in Sec. 3.2), and recovers image layers by image decoders (in Sec. 3.3)." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 334 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distinct motions of reflection and transmission layers. However, special data capture requirements significantly limit the application scope of these methods, especially for mobile devices or images downloaded from the Internet.", + "bbox": [ + 75, + 431, + 468, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed method", + "text_level": 1, + "bbox": [ + 76, + 508, + 246, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of the proposed framework is illustrated in Figure 2. In this section, we start by introducing the feature extraction stage (in Sec. 3.1) that obtains multi-scale image features and global language features, the gated language interaction stage (in Sec. 3.2) that conducts progressive image-language global interactions to exploit distinctive image features and prevents unavailable language interactions with switchable gates, and the layer recovery stage (in Sec. 3.3) that reprojects features into the image space with a light-weight image decoder. Then we explain loss functions (in Sec. 3.4) employed for network optimization, especially a contrastive correspondence loss and a layer correspondence loss that constrains the network to construct correspondences between the language description and the corresponding layer under the disturbance of the other layer in a superimposed mixture image. Finally, we present our training strategy (in Sec. 3.5) which enables the network to be applicable for varying quantities of input language descriptions and jointly tackles the recognizable layer ambiguity problem with the gated network design.", + "bbox": [ + 75, + 532, + 470, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image and language feature extraction", + "text_level": 1, + "bbox": [ + 76, + 847, + 408, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The inputs of the proposed method consist of a mixture image $\\mathbf{M}$ with two language descriptions $\\{\\mathbf{L}_i|i = 1,2\\}$", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which corresponds to the two image layers. We specify that layer $\\mathbf{I}_i$ corresponds to the description $\\mathbf{L}_i$ . However, due to the recognizable layer ambiguity that in certain cases only one layer of the mixture image is recognizable (usually the transmission layer), for such cases, we set $\\mathbf{L}_1$ to be the available language description (for the recognizable transmission layer) and $\\mathbf{L}_2$ to null (for the unrecognizable reflection layer) to ensure a unified input setting. Then given the input image and language descriptions, the feature extraction stage aims at obtaining the image feature $\\mathbf{F}_{\\mathrm{M}}$ with the image encoder and the multi-scale feature fusion process and extracting the global language feature $\\mathbf{F}_{\\mathbf{L}_i}$ for each description $\\mathbf{L}_i$ via the language encoder for the subsequent interaction procedure.", + "bbox": [ + 496, + 431, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image encoder. Given a mixture image $\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we employ a commonly-used vision backbone ResNet-50 [16] as our image encoder, whose last three layers (i.e., an average pooling layer, a fully connected layer, and a softmax layer) are removed to fit our task. We utilize image features from the first five blocks of the image encoder to form a multi-scale feature pyramid $\\{\\mathbf{F}_{\\mathbf{M}_i}\\}_{i=1}^5$ , where $\\mathbf{F}_{\\mathbf{M}_i} \\in \\mathbb{R}^{h_i \\times w_i \\times C_i}$ , $h_i = H/2^i$ and $w_i = W/2^i$ , $H$ and $W$ is the height and width of the mixture image, respectively, and $C_i$ is the dimension of the $i$ -th extracted feature.", + "bbox": [ + 496, + 648, + 890, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-scale feature fusion. Obtaining the extracted feature pyramid, we first transform it into a hypercolumn feature $\\mathbf{F}_{\\mathbf{M}}^{\\mathrm{hyp}} \\in \\mathbb{R}^{h \\times w \\times C^{\\mathrm{hyp}}} [14]$ (where $h = H/2$ , $w = W/2$ , and $C^{\\mathrm{hyp}} = \\sum_{i=1}^{5} C_i$ ), which has been proved to be effective in fusing multi-scale contextual information for reflection separation [58, 67]. Considering the computational cost,", + "bbox": [ + 496, + 806, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24915", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75de14e1787dca95ffbbbd42d5c9259507ba41a253e09c356f3e49918a8fefe8.jpg", + "image_caption": [ + "(a) AGAM: Adaptive global aggregation module" + ], + "image_footnote": [], + "bbox": [ + 83, + 113, + 464, + 210 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a30e18b3b14960b596b66f902c24e2e5b54bbab10688658b825283817f37c6d.jpg", + "image_caption": [ + "(b) AGIM: Adaptive global interaction module", + "Figure 3. The architecture of the (a) adaptive global aggregation module (AGAM) and (b) adaptive global interaction module (AGIM), which aggregates global contextual information of visual features and achieves feature channel rearrangement with gated language guidance, respectively." + ], + "image_footnote": [ + "$\\otimes$ Matrix multiplication $\\mathbb{R}$ Reshape $\\odot$ Element-wise multiplication", + "Language gate $\\odot$ Concatenation $\\oplus$ Element-wise addition" + ], + "bbox": [ + 83, + 234, + 464, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we condense and refine the hypercolumn feature by a $1 \\times 1$ convolutional layer with a GELU activation [17] followed by a locally-enhanced feed-forward (LeFF) block [56]. The final fused feature of the mixture image is denoted as $\\mathbf{F}_{\\mathbf{M}} \\in \\mathbb{R}^{h \\times w \\times C}$ , which serves as the basis for the subsequent interaction and separation process.", + "bbox": [ + 75, + 483, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Language encoder. Motivated by the rapid development of pre-trained large-scale vision-language models, we employ the language encoder from CLIP [42], which adopts a Transformer architecture [48] to extract language features and obtains a global contextual feature in the multi-modal embedding space by using layer normalization and linear projection layers. Given a language description $\\mathbf{L}_i\\in \\mathbb{R}^L$ we obtain its corresponding global feature $\\mathbf{F}_{\\mathbf{L}_i}\\in \\mathbb{R}^C$ by leveraging the modeling capacity of the language encoder to encode the description, thus extracting the holistic contextual information of the corresponding image layer. Here $L$ denotes the length of the language description and $C$ is the feature dimension as the image feature.", + "bbox": [ + 75, + 575, + 468, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Gated language interaction", + "text_level": 1, + "bbox": [ + 76, + 785, + 321, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The gated language interaction stage aims at leveraging the contextual information from available language descriptions to guide the separation of the corresponding layer feature $\\mathbf{F}_{\\mathbf{I}_i}$ , which is composed of $2N$ cascaded interaction groups to separate image layer features successively. As shown in Figure 2, each group consists of an adaptive global", + "bbox": [ + 75, + 810, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aggregation module (AGAM) to gather global information of input visual features, a language gate to prevent detrimental guidance from unavailable descriptions, an adaptive global interaction module (AGIM) to conduct interactions using global features for exploring holistic image-language content correspondence, and $K$ normalization layers with LeFF blocks [56] for feature refinement. The former $N$ interaction groups are utilized for separating $\\mathbf{F}_{\\mathbf{I}_1}$ from $\\mathbf{F}_{\\mathbf{M}}$ and $\\mathbf{F}_{\\mathbf{L}_1}$ , and the latter $N$ groups for separating $\\mathbf{F}_{\\mathbf{I}_2}$ from $\\mathbf{F}_{\\mathbf{I}_2}^{\\mathrm{ini}}$ (obtained by feeding the concatenation of $\\mathbf{F}_{\\mathbf{M}}$ and $\\mathbf{F}_{\\mathbf{I}_1}$ into a $1\\times 1$ convolutional layer) and $\\mathbf{F}_{\\mathbf{L}_2}$ (if $\\mathbf{L}_2$ is available). We set $N = 4$ and $K = 2$ in practice. Details of the gated language interaction are described as follows.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive global aggregation module (AGAM). As the network structure shown in Figure 3(a), given an input visual feature $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} \\in \\mathbb{R}^{h \\times w \\times C}$ with the spatial resolution of $h \\times w$ , AGAM is designed for adaptively obtaining a global feature $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} \\in \\mathbb{R}^{C}$ that aggregates the contextual information for the subsequent interaction. The visual feature $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}$ is firstly averaged by an average pooling layer to obtain $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}} \\in \\mathbb{R}^{C}$ . Then the aggregation process is accomplished via a cross-attention mechanism which contains three linear projection layers with layer normalization [1] to conduct query, key, and value projections: $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} = \\operatorname{Softmax}(\\mathbf{QK}^{\\top} / \\tau)\\mathbf{V}$ , where the query $\\mathbf{Q} \\in \\mathbb{R}^{C}$ is projected from $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}$ , and the key $\\mathbf{K} \\in \\mathbb{R}^{(hw + 1) \\times C}$ and value $\\mathbf{V} \\in \\mathbb{R}^{(hw + 1) \\times C}$ are from the concatenation of $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}$ and $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}$ , and $\\tau$ is a learnable scaling factor to control the magnitude of the dot product of $\\mathbf{Q}$ and $\\mathbf{K}$ before applying the softmax function.", + "bbox": [ + 496, + 290, + 892, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive global interaction module (AGIM). Inspired by existing reflection separation approaches [23, 58] which attempt to distinguish transmission and reflection layers in the feature space by feature channel rearrangement (i.e., allocating distinct channels from mixture image features to the two layers), we propose to integrate language interactions at the feature channel level. To achieve this, as illustrated in Figure 3(b), we propose the adaptive global interaction module (AGIM), which employs the channel-wise cross-attention mechanism to conduct interactions between global language features and image features for channel rearrangement. The query, key, and value projections using linear projection layers with layer normalization [1] are conducted on the global language feature $\\mathbf{F}_{\\mathbf{L}_i}$ , the global visual feature $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{gl}}$ , and the visual feature $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}$ to generate $\\mathbf{Q} \\in \\mathbb{R}^{C}$ , $\\mathbf{K} \\in \\mathbb{R}^{C}$ , and $\\mathbf{V} \\in \\mathbb{R}^{C \\times h w}$ , respectively. Before the interaction, a language gate is designed to prevent impacts of unavailable language descriptions caused by the recognizable layer ambiguity problem that users may only input one description for the layer recognizable in the mixture image. Specifically, if the description $\\mathbf{L}_i$ corresponding to the current global language feature $\\mathbf{F}_{\\mathbf{L}_i}$ is available (not set to null), the gate will feed $\\mathbf{F}_{\\mathbf{L}_i}$ into the following interaction", + "bbox": [ + 496, + 551, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24916", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "process, otherwise the gate will feed $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}}$ , which turns the interaction process to be a channel-wise self-attention.", + "bbox": [ + 75, + 89, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After filtering by the language gate, an interacted feature can be obtained through the channel-wise cross-attention: $\\mathbf{F}^{\\mathrm{itr}} = \\operatorname{Softmax}(\\mathbf{Q}^\\top \\mathbf{K} / \\eta)\\mathbf{V}$ , where $\\eta$ is a learnable scaling factor to control the magnitude of the attention map. Besides, to adaptively adjust the influence of language guidance based on the correspondence between image and language features, we define another scaling factor $s$ valued as the cosine distance between $\\mathbf{Q}$ and $\\mathbf{K}$ to multiply with the interacted feature $\\mathbf{F}^{\\mathrm{itr}}$ . Finally, the output of AGIM is obtained by a residual structure: $\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{out}} = \\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} + s\\mathbf{F}^{\\mathrm{itr}}$ , which integrates the contextual information from the language description and adjusts features for layer separation.", + "bbox": [ + 75, + 121, + 470, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Layer recovery", + "text_level": 1, + "bbox": [ + 76, + 313, + 230, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After global language-image interaction and progressive refinement by the gated language interaction stage, we obtain layer features $\\mathbf{F}_{\\mathbf{I}_i}$ which integrates holistic contextual information from the corresponding language description $\\mathbf{L}_i$ (if available). The layer recovery stage is dedicated to reconstructing each image layer $\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}$ from its corresponding layer features $\\mathbf{F}_{\\mathbf{I}_i}$ , which is achieved by using individual image decoders. In the image decoder, the layer feature is firstly upsampled by a transposed convolutional layer with a GELU activation [17], then refined by a residual block, and finally projected from the feature domain back into the image domain by a $1 \\times 1$ convolutional layer with a sigmoid activation, thus accomplishing a precise layer recovery.", + "bbox": [ + 75, + 337, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss functions", + "text_level": 1, + "bbox": [ + 76, + 559, + 225, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce the contrastive correspondence loss $\\mathcal{L}_{\\mathrm{ctr}}$ and the layer correspondence loss $\\mathcal{L}_{\\mathrm{lcr}}$ for constraining the proposed method to establish the cross-modality correspondence between language descriptions $\\mathbf{L}_i$ and corresponding image layers under the content disturbance from counter layers in mixture images. We also briefly describe the image layer loss $\\mathcal{L}_{\\mathrm{img}}$ which consists of image- and feature-level loss functions for layer recovery. We denote the estimated image layers as $\\tilde{\\mathbf{I}}_i$ and ground truths as $\\mathbf{I}_i$ . Details of the loss functions are as follows.", + "bbox": [ + 75, + 583, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contrastive correspondence loss. CLIP [42] conducts contrastive language-image pre-training within batches which successes in establishing cross-modality correspondence between language descriptions and clean images, while our language-guided image reflection separation task requires finding the correspondence between the given language description and a certain layer in the mixture image with the interference of extraneous contents. To tackle the above issue, we propose a contrastive correspondence loss that conducts contrastive learning between image layers to establish correct cross-modality correspondence. Specifi-", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cally, given a language description $\\mathbf{L}_i$ , the goal of the contrastive correspondence loss is to force the network to learn the relation that the corresponding image layer $\\mathbf{I}_i$ is more relevant to $\\mathbf{L}_i$ than the counter layer $\\mathbf{I}_j$ ( $j \\neq i$ ), thus constraining the estimated image layer $\\tilde{\\mathbf{I}}_i$ to conform to the associated language description $\\mathbf{L}_i$ .", + "bbox": [ + 496, + 90, + 890, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To measure the relevance between a language description $\\mathbf{L}$ and an image layer $\\mathbf{I}$ , we define a feature-level similarity function $\\mathcal{D}(\\cdot)$ as:", + "bbox": [ + 496, + 183, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} (\\mathbf {L}, \\mathbf {I}) = \\sigma \\left(\\Psi \\left(\\mathbf {F} _ {\\mathbf {L}} ^ {\\mathrm {g l o}}, \\mathbf {F} _ {\\mathbf {I}} ^ {\\mathrm {g l o}}\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 239, + 890, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ represents the sigmoid function, $\\Psi$ represents the cosine distance, $\\mathbf{F}_{\\mathbf{L}}^{\\mathrm{glo}}$ is the global language feature obtained by the language encoder (in Sec. 3.1), and $\\mathbf{F}_{\\mathbf{I}}^{\\mathrm{glo}}$ is the global image feature produced by the AGAM (in Sec. 3.2). For an available language description $\\mathbf{L}_i$ , we calculate its contrastive correspondence loss as:", + "bbox": [ + 496, + 271, + 890, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {j}\\right) = - \\log \\left(\\frac {\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right)}{\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right) + \\mathcal {D} \\left(\\mathbf {L} _ {i} , \\mathbf {I} _ {j}\\right)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 376, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and we sum up the contrastive correspondence loss of each available language description as the final one.", + "bbox": [ + 496, + 422, + 890, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layer correspondence loss. We further define a layer correspondence loss to encourage the relevance between the language description $\\mathbf{L}_i$ and the estimated image layer $\\tilde{\\mathbf{I}}_i$ to approach the relevance between $\\mathbf{L}_i$ and the ground truth image layer $\\mathbf{I}_i$ :", + "bbox": [ + 496, + 454, + 890, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {l c r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {i}\\right) = \\left\\| \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}\\right) - \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\mathbf {I} _ {i}\\right) \\right\\| _ {1}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 540, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}(\\cdot)$ is the same feature-level similarity function as in the contrastive correspondence loss. We also sum up the layer correspondence loss of each available language description for final supervision.", + "bbox": [ + 496, + 577, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image layer loss. To achieve high-fidelity recovery of image layers (i.e., transmission and reflection layers), the proposed method is also optimized with loss functions following previous reflection separation methods [23, 52, 67]. Specifically, we utilize loss functions that conduct constraints on the visual quality of estimated images (i.e., the pixel $\\mathcal{L}_{\\mathrm{pix}}$ , structural similarity $\\mathcal{L}_{\\mathrm{ssim}}$ , and perceptual loss $\\mathcal{L}_{\\mathrm{per}}$ ) or exploit the inherent relationship between two layers (i.e., the exclusion $\\mathcal{L}_{\\mathrm{exc}}$ and reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ ), and we denote the combination of the above image- or feature-level loss functions as the image layer loss $\\mathcal{L}_{\\mathrm{img}}^1$ .", + "bbox": [ + 496, + 638, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, the total loss function is then formulated as:", + "bbox": [ + 517, + 805, + 867, + 820 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {c t r}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {l c r}} + \\mathcal {L} _ {\\mathrm {i m g}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 833, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where coefficients are set as $\\gamma_{1} = \\gamma_{2} = 0.5$", + "bbox": [ + 500, + 862, + 790, + 877 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "Details of $\\mathcal{L}_{\\mathrm{img}}$ are provided in the supplementary material.", + "bbox": [ + 516, + 886, + 841, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24917", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training strategy", + "text_level": 1, + "bbox": [ + 76, + 90, + 246, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the recognizable layer ambiguity that sometimes only one layer in the mixture image is recognizable, we propose a randomized training strategy to synergize with the gated language interaction mechanism (in Sec. 3.2). Since our training data simulates the recognizable layer ambiguity that some image layers do not have corresponding language descriptions (introduced in the next section), we only feed the available language description corresponding to the other image layer to guide the separation. For data with descriptions of both layers available, which indicates that both layers are recognizable, we also randomly drop one language description and feed the remaining one into the network to improve the generalization capacity for the proposed method. In practice, we set the ratio of dropping language descriptions to $30\\%$ .", + "bbox": [ + 75, + 114, + 472, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement the proposed method with PyTorch [41] with a batch size of 16 on two Nvidia GeForce RTX 3090 GPUs. The model is trained for 40 epochs with Adam optimizer [25] to update learnable parameters. Weights are initialized as in [15]. The learning rate is set to $10^{-4}$ initially and decreases to $10^{-5}$ at epoch 30.", + "bbox": [ + 75, + 342, + 470, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Data preparation", + "text_level": 1, + "bbox": [ + 76, + 450, + 246, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Though existing works have constructed several datasets for single-image reflection separation [31, 49, 67], they are unavailable for the proposed language-guided reflection separation framework due to the lack of corresponding language descriptions. Therefore, we build a dataset containing both synthetic and real data to overcome the data deficiency and facilitate network training and evaluation. Each group of data is composed of a mixture image, a transmission layer, a reflection layer, and two language descriptions. Details of synthetic and real data are as follows.", + "bbox": [ + 75, + 477, + 472, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Synthetic data", + "text_level": 1, + "bbox": [ + 76, + 641, + 225, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The synthetic dataset is generated for network training to satisfy the data-driven need of the proposed method. Due to the demand for paired image-language data, we utilize two prevailing image captioning datasets (i.e., Flickr30k [66] and COCO Captions [7]) for data generation, which contain 31,000 and 330,000 images respectively, and each image has 5 independent human-generated language descriptions. We randomly select images from the above two datasets as transmission $\\mathbf{T}_{\\mathrm{S}}$ and reflection scene images $\\mathbf{R}_{\\mathrm{S}}$ and conduct an image synthesis process with linear blending [23]:", + "bbox": [ + 75, + 666, + 468, + 818 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} = \\hat {\\mathbf {T}} + \\hat {\\mathbf {R}} = \\alpha \\hat {\\mathbf {T}} _ {\\mathrm {S}} + \\beta \\hat {\\mathbf {R}} _ {\\mathrm {S}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 825, + 468, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{T}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{T}_{\\mathrm{S}})$ and $\\hat{\\mathbf{R}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{R}_{\\mathrm{S}})$ , $g_{\\mathrm{inv}}$ represents the inverse gamma correction, and $\\alpha \\in [0.8,1]$ and $\\beta \\in [0.4,1]$ are the blending attenuation coefficients as in [23].", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering the recognizable layer ambiguity that sometimes only one layer is recognizable in a mixture image, we assign a language description only when the corresponding layer is obvious enough in the synthesized mixture image, i.e., $\\frac{\\mathrm{mean}(\\hat{\\mathbf{V}}_l)}{\\mathrm{mean}(\\hat{\\mathbf{V}}_{\\mathbf{M}})}\\geqslant \\mu$ , where $\\hat{\\mathbf{V}}_l$ represents the brightness image of $\\hat{\\mathbf{T}}$ or $\\hat{\\mathbf{R}}$ in the HSV color space and $\\hat{\\mathbf{V}}_{\\mathbf{M}}$ represents the brightness image of $\\hat{\\mathbf{M}}$ , and we set $\\mu = 0.3$ in the data generation process. Finally, gamma correction is applied to image triplets $\\{\\hat{\\mathbf{T}},\\hat{\\mathbf{R}},\\hat{\\mathbf{M}}\\}$ to obtain $\\{\\mathbf{T},\\mathbf{R},\\mathbf{M}\\}$ , and we generate 50000 triplets of data in total for network training.", + "bbox": [ + 496, + 90, + 890, + 252 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Real data", + "text_level": 1, + "bbox": [ + 500, + 265, + 609, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Existing real datasets collected for the single-image reflection separation task usually contain mixture images with ground truth of transmission layers (e.g., Zhang et al. [67] and Nature [31]), and $\\mathrm{SIR}^2$ [49, 54] further captures ground truths of reflection layers. For these off-the-shelf real datasets, we augment them by manually adding language descriptions for each group of data to satisfy the input setting of the proposed language-guided reflection separation task. Specifically, following the annotation principle of COCO Captions [7], we first describe the content of transmission layers in mixture images with entities, attributes (e.g., colors or materials), and relative positions between different entities. If the content of reflection layers is recognizable in mixture images, we also give language descriptions for reflection layers in the same way. To further evaluate the generalization capacity of the proposed method, we collect a real dataset (denoted as REFOL dataset) containing 100 mixture images from the Internet that are captured in different scenes and with different cameras. We annotate these mixture images with language descriptions in the same manner as mentioned above. Following the training strategy of previous methods [10, 31], we utilize 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [67] for training, and the rest of real data are used for quantitative and qualitative evaluation.", + "bbox": [ + 496, + 287, + 893, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 684, + 633, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 500, + 709, + 859, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of the proposed method, we conduct quantitative and qualitative experiments on existing real datasets [31, 49, 67] (with our manually annotated language descriptions) and our newly collected dataset REFOL. We compare with state-of-the-art single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], IBCLN [31], CoRRN [52], and Zhang et al. [67]. For fair comparisons, we finetune the above methods on our training data if their training codes are provided. We report better results between the original pre-trained model and the finetuned version.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24918", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6d294cb49880cb8af0cec613e2bd28c4062a0dc2d2ccd20bd29eb7cbc343b06c.jpg", + "table_caption": [ + "Table 1. Comparison of quantitative results in terms of PSNR [24] and SSIM [55] on real datasets for evaluating the recovery of transmission layers. $\\uparrow (\\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + ], + "table_footnote": [], + "table_body": "
Dataset (size)MetricsMethods
Zhang et al. [67]CoRRN [52]IBCLN [31]Dong et al. [10]YTMT [22]DSRNet [23]Ours
Postcard (199)PSNR↑20.8522.0423.4123.7222.8224.8825.02
SSIM↑0.8720.8700.8720.9030.8850.9100.915
Object (200)PSNR↑23.8425.1324.5224.3624.6826.4426.51
SSIM↑0.8720.9120.8910.8980.8920.9210.927
Wild (101)PSNR↑24.9725.1724.7825.7525.7025.8626.23
SSIM↑0.8750.8890.8840.9030.8970.9080.925
Real20 (20)PSNR↑22.3421.4321.4723.3423.2323.8824.05
SSIM↑0.7950.8010.7620.8120.8020.8160.824
Nature (20)PSNR↑20.6220.7523.7223.4521.5322.2623.87
SSIM↑0.7530.7830.8060.8080.7780.8010.812
Average (540)PSNR↑22.7723.7024.0224.3124.0125.5125.72
SSIM↑0.8650.8830.8750.8940.8830.9060.914
", + "bbox": [ + 116, + 121, + 849, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31fd5974efffb740d30321de4e2dc25e963bae97e2aaf4c24cacd2e7740935d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 321, + 349, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17f7b10bc75071d9448dfd2b7eca0c844c71f26f1e39d7117335d3c522fd51b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 321, + 617, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/059f90c0b7d490404328ff5ae74a15d53a34e038f14c2df3ee82211ceb17ff2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 321, + 887, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a4ab7a730efd3cbfc4092ee474c0b8b2d33ed415a98d21b0ed345b9c65a3e67.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 441, + 349, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5485608babf34c448d34d210adc4a6a3ead931fa5a414fc6e64b3a070f2e6dcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 441, + 617, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9d60594cfba8391e7c2d2b5a80d2c6733a02a6be7da9bcecd91bde3255b50fcc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 441, + 887, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d67069550ba292446096683f0bc1ad631394316de46bb16afec1baceaacdd538.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 565, + 349, + 683 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d983f8002a4d9565fb15bf90455a7786d881941321a17cfbfc46af44816c0228.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 565, + 617, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5e58dd925ae40e1a3efcb9d41aef3aa2b55f2a0198b18c3f64bf564399e3836.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 565, + 887, + 681 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/740aa01ce9ffaec58414e5334f1279279cc01321a5e03f938f0465f648d35a74.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison of estimated transmission and reflection layers on real data, compared with the state-of-the-art methods including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Please zoom in for details." + ], + "image_footnote": [], + "bbox": [ + 83, + 686, + 349, + 800 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64d42f4f91c4c08337d7600a88145badafad6673c031b19c7b91a535505974d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 686, + 617, + 800 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7218db79820066185abdbc9e2cbb6e86699ac06643f166da2a840f6817eb7d63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 686, + 887, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection separation,", + "bbox": [ + 76, + 858, + 470, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "i.e., Nature [31], Real20 [67], and three subsets of $\\mathrm{SIR}^2$ [49] dataset. Following the setting of existing reflection separa", + "bbox": [ + 500, + 858, + 890, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24919", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c3ad27059d69f21233bb823cc22740c1eabe7fad97c112435aa6e70ec93aba12.jpg", + "table_caption": [ + "Table 2. Quantitative results of ablation studies." + ], + "table_footnote": [], + "table_body": "
Metricsw/o languagew/o AGIMLimg onlyOurs
PSNR↑24.3124.6924.5225.72
SSIM↑0.8850.9010.8930.914
", + "bbox": [ + 81, + 108, + 465, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion methods [10, 36], we utilize PSNR [24] and SSIM [55] as error metrics for evaluating the recovery of transmission layers2. As quantitative results shown in Table 1, the proposed method achieves the best performance of both PSNR and SSIM, which validates its generalization capacity and the effectiveness of language descriptions.", + "bbox": [ + 75, + 181, + 467, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparison. To evaluate the visual quality of reflection separation results, we compare the proposed method with four single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Qualitative results on recovering both transmission and reflection layers are shown in Figure 4. As can be observed in Figure 4, IBCLN [31] can only remove parts of reflections, while Dong et al. [10] have trouble in dealing with complex semantic images like the reflections of a man in a room (the second example). YTMT [22] separates layers incorrectly and thus brings the content of transmission layers into reflections. DSRNet [23] fails to recover transmission and reflection layers for these challenging cases. Contributing to the language guidance, the proposed method generates better visual results and recovers both transmission and reflection layers neatly.", + "bbox": [ + 75, + 273, + 468, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 523, + 225, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct several ablation studies with quantitative results shown in Table 2 to investigate the influence of the additional input of language descriptions (denoted as 'w/o language'), the network design of AGIM (denoted as 'w/o AGIM'), and the loss functions for cross-modality correspondence (denoted as $\\mathcal{L}_{\\mathrm{img}}$ only). The performance of the variant 'w/o language' suffers from an obvious degradation as we remove language descriptions, which shows the effectiveness of the additional contextual information for layer separation. The variant 'w/o AGIM' replaces AGIMs with simple feature fusion blocks which directly concatenate language and image features and feed them into self-attention blocks, and the decline in performance validates the necessity of our gated interaction mechanism. The variant $\\mathcal{L}_{\\mathrm{img}}$ only is trained with $\\mathcal{L}_{\\mathrm{img}}$ , which obtains results slightly better than the variant 'w/o language', indicating the significance of establishing the cross-modality correspondence.", + "bbox": [ + 75, + 547, + 468, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the effectiveness of the language interaction, we conduct an ablation study by gradually increasing the number of input descriptions. As shown in yellow", + "bbox": [ + 75, + 819, + 467, + 864 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72e62436b20e32a8267b24dcfeec431c63c50b730e8166a45c80428798e75f44.jpg", + "image_caption": [ + "Figure 5. Results with different numbers of input language descriptions." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 887, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "boxes of Figure 5, by utilizing more input language descriptions, the separation of transmission (clothes) and reflection layers (a man with a camera) becomes more thorough, which also demonstrates the efficacy and robustness of the proposed method.", + "bbox": [ + 498, + 378, + 890, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 476, + 617, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces natural language to provide contextual information about image layers for relieving the ill-posed reflection separation problem. We develop an end-to-end framework with adaptive global interaction modules and language-image loss functions to effectively manage the modality inconsistency, and we adopt a language gate mechanism with randomized training strategies to handle the recognizable layer ambiguity. To address data deficiency, a specially built dataset with language annotations significantly aids in training and evaluating the proposed language-guided image reflection separation framework. Quantitative and qualitative experiments on real data demonstrate the effectiveness of introducing language descriptions for reflection separation.", + "bbox": [ + 496, + 503, + 890, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The proposed method may not distinguish transmission and reflection layers accurately when their contents are similar. For such ambiguous cases, a more flexible language-guided mechanism is needed. This might be solved by exploring a better interaction approach, which is left as our future work.", + "bbox": [ + 496, + 718, + 890, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 830, + 658, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102.", + "bbox": [ + 500, + 857, + 890, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "2Quantitative evaluations on the recovery of reflection layers are provided in the supplementary material.", + "bbox": [ + 76, + 875, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24920", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4", + "[2] Yakun Chang, Cheolkon Jung, and Jun Sun. Joint reflection removal and depth estimation from a single image. IEEE TCYB, 2020. 1", + "[3] Yakun Chang, Cheolkon Jung, Jun Sun, and Fengqiao Wang. Siamese dense network for reflection removal with flash and no-flash image pairs. IJCV, 2020. 2", + "[4] Zheng Chang, Shuchen Weng, Yu Li, Si Li, and Boxin Shi. L-CoDer: Language-based colorization with color-object decoupling transformer. In Proc. of ECCV, 2022. 1", + "[5] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CoIns: Language-based colorization with instance awareness. In Proc. of CVPR, 2023.", + "[6] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Proc. of NeurIPS, 2023. 1", + "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 6", + "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. NeRDi: Single-view NeRF synthesis with language-guided diffusion as general image priors. In Proc. of CVPR, 2023. 2", + "[9] Yaron Diamant and Yoav Y Schechner. Overcoming visual reverberations. In Proc. of CVPR, 2008. 2", + "[10] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson WH Lau. Location-aware single image reflection removal. In Proc. of ICCV, 2021. 1, 2, 6, 7, 8", + "[11] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf. A generic deep architecture for single image reflection removal and image smoothing. In Proc. of ICCV, 2017. 2", + "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. of NeurIPS, 2014. 2", + "[13] Byeong-Ju Han and Jae-Young Sim. Zero-shot learning for reflection removal of single 360-degree image. In Proc. of ECCV, 2022. 2", + "[14] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In Proc. of CVPR, 2015. 3", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance onImagenet classification. In Proc. of ICCV, 2015. 6", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of CVPR, 2016. 3", + "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 4, 5" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Yuchen Hong, Youwei Lyu, Si Li, and Boxin Shi. Near-infrared image guided reflection removal. In Proc. of ICME, 2020. 2", + "[19] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. Panoramic image reflection removal. In Proc. of CVPR, 2021. 1, 2", + "[20] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with NIR and RGB image feature fusion. IEEE TMM, 2022. 2", + "[21] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. PAR $^2$ Net: End-to-end panoramic image reflection removal. IEEE TPAMI, 2023. 2", + "[22] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation.* In *Proc. of NeurIPS*, 2021. 1, 2, 6, 7, 8", + "[23] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proc. of ICCV, 2023. 2, 4, 5, 6, 7, 8", + "[24] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics letters*, 2008, 7, 8", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[26] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. A physically-based approach to reflection separation. In Proc. of CVPR, 2012. 2", + "[27] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proc. of CVPR, 2021. 1, 2", + "[28] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proc. of CVPR, 2020. 1, 2", + "[29] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE TPAMI, 2023. 2", + "[30] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE TPAMI, 2007. 1, 2", + "[31] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proc. of CVPR, 2020. 1, 2, 6, 7, 8", + "[32] Yu Li and Michael S Brown. Exploiting reflection change for automatic reflection removal. In Proc. of ICCV, 2013. 1, 2", + "[33] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proc. of CVPR, 2014. 1, 2", + "[34] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In Proc. of CVPR, 2020. 1, 2", + "[35] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. IEEE TPAMI, 2021. 2", + "[36] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolar" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24921", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ized and polarized images. In Proc. of NeurIPS, 2019. 1, 2, 8", + "[37] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE TPAMI, 2022. 2", + "[38] Daiqian Ma, Renjie Wan, Boxin Shi, Alex C Kot, and Ling-Yu Duan. Learning to jointly generate and separate reflections. In Proc. of ICCV, 2019. 2", + "[39] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. IJCV, 1997. 2", + "[40] Jonghyuk Park, Hyeona Kim, Eunpil Park, and Jae-Young Sim. Fully-automatic reflection removal for 360-degree images. In Proc. of WACV, 2024. 2", + "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Proc. of NeurIPS, 2019. 6", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. of ICML, 2021. 4, 5", + "[43] Yoav Y Schechner, Nahum Kiryati, and Ronen Basri. Separation of transparent layers using focus. IJCV, 2000. 2", + "[44] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proc. of CVPR, 2015. 1, 2", + "[45] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In Proc. of CVPR, 2015. 1, 2", + "[46] Jimeng Sun, Shuchen Weng, Zheng Chang, Si Li, and Boxin Shi. UniCoRN: A unified conditional image inpainting network. In Proc. of CVPR, 2022. 1", + "[47] Jiajun Tang, Haofeng Zhong, Shuchen Weng, and Boxin Shi. LuminAIRe: Illumination-aware conditional image inpainting for lighting-realistic generation. In Proc. of NeurIPS, 2023. 1", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, 2017. 4", + "[49] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In Proc. of ICCV, 2017. 2, 6, 7", + "[50] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, Wen Gao, and Alex C. Kot. Region-aware reflection removal with unified content and gradient priors. IEEE TIP, 2018. 2", + "[51] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CRRN: Multi-scale guided concurrent reflection removal network. In Proc. of CVPR, 2018. 2", + "[52] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CoRRN: Cooperative reflection removal network. IEEE TPAMI, 2019. 2, 5, 6, 7", + "[53] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. IJCV, 2021. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Renjie Wan, Boxin Shi, Haoliang Li, Yuchen Hong, Ling-Yu Duan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In IEEE TPAMI, 2022. 2, 6", + "[55] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In Proc. of ACSSC, 2003. 7, 8", + "[56] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proc. of CVPR, 2022. 4", + "[57] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: CLIP-driven referring image segmentation. In Proc. of CVPR, 2022. 1", + "[58] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proc. of CVPR, 2019. 2, 3, 4", + "[59] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In Proc. of CVPR, 2019. 2", + "[60] Shuchen Weng and Boxin Shi. Conditional image repainting. IEEE TPAMI, 2023. 1", + "[61] Shuchen Weng, Wenbo Li, Dawei Li, Hongxia Jin, and Boxin Shi. MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In Proc. of CVPR, 2020. 1", + "[62] Shuchen Weng, Hao Wu, Zheng Chang, Jiajun Tang, Si Li, and Boxin Shi. L-CoDe: Language-based colorization using color-object decoupled conditions. In Proc. of AAAI, 2022. 1", + "[63] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In Proc. of ECCV, 2018. 2", + "[64] Yang Yang, Wenye Ma, Yin Zheng, Jian-Feng Cai, and Weiyu Xu. Fast single image reflection suppression via convex optimization. In Proc. of CVPR, 2019. 2", + "[65] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-aware vision transformer for referring image segmentation. In Proc. of CVPR, 2022. 1, 2", + "[66] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2, 6", + "[67] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proc. of CVPR, 2018. 2, 3, 5, 6, 7", + "[68] Ya-Nan Zhang, Linlin Shen, and Qiufu Li. Content and gradient model-driven deep network for single image reflection removal. In Proc. of ACM MM, 2022. 2", + "[69] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In Proc. of CVPR, 2020. 2", + "[70] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In Proc. of CVPR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24922", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_model.json b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22cbb8e10f6ae24909be601e5e9822862a5b4db1 --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.132, + 0.722, + 0.154 + ], + "angle": 0, + "content": "Language-guided Image Reflection Separation" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.167, + 0.854, + 0.187 + ], + "angle": 0, + "content": "Haofeng Zhong\\(^{1,2,3\\#}\\) Yuchen Hong\\(^{1,2\\#}\\) Shuchen Weng\\(^{1,2}\\) Jinxiu Liang\\(^{1,2}\\) Boxin Shi\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.188, + 0.828, + 0.204 + ], + "angle": 0, + "content": "\\(^{1}\\) National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.204, + 0.815, + 0.218 + ], + "angle": 0, + "content": "\\(^{2}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.144, + 0.188, + 0.828, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.218, + 0.848, + 0.246 + ], + "angle": 0, + "content": "\\(^{3}\\) AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.268, + 0.314, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.301, + 0.474, + 0.498 + ], + "angle": 0, + "content": "This paper studies the problem of language-guided reflection separation, which aims at addressing the ill-posed reflection separation problem by introducing language descriptions to provide layer content. We propose a unified framework to solve this problem, which leverages the crossattention mechanism with contrastive learning strategies to construct the correspondence between language descriptions and image layers. A gated network design and a randomized training strategy are employed to tackle the recognizable layer ambiguity. The effectiveness of the proposed method is validated by the significant performance advantage over existing reflection separation methods on both quantitative and qualitative comparisons." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.528, + 0.21, + 0.543 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.75 + ], + "angle": 0, + "content": "When photographing through transparent materials like glass windows or showcases, the presence of reflections can significantly degrade the image quality of captured images and disrupt downstream computer vision tasks like face recognition [53] or depth estimation [2]. As an attractive topic in computational photography, reflection separation aims at decomposing the contaminated mixture image (denoted as \\(\\mathbf{M}\\)) into two components that correspond to scenes located at different sides of the glass, i.e., the reflection layer (denoted as \\(\\mathbf{R}\\)) and the transmission layer (denoted as \\(\\mathbf{T}\\)). Since reflection separation is a severely ill-posed problem, it is imperative to exploit effective priors or auxiliary information for distinguishing the two components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.752, + 0.47, + 0.873 + ], + "angle": 0, + "content": "The primary challenge of solving the reflection separation problem lies in the exploration of distinct clues for distinguishing transmission and reflection layers. Multi-image methods handle this problem by introducing additional constraints. Some of them acquire a series of mixture images in different viewpoints [32, 34, 45] to harness the distinct motions of the two layers, while others adopt specialized capturing setups to obtain complementary scene informa" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.268, + 0.891, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.519, + 0.894, + 0.603 + ], + "angle": 0, + "content": "Figure 1. The recognizable layer ambiguity problem causes uncertain quantities of input language descriptions for language-guided image reflection separation. Given language descriptions of either (a) one layer or (b) both two layers, the proposed method achieves robust reflection separation compared with an existing reflection separation method [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "tion [19, 27, 28, 36]. However, the specialized data capture requirements limit the application scope of these methods, especially for images downloaded from the Internet. Single-image methods attempt to tackle reflection separation by utilizing handcrafted priors derived from natural image statistics [30, 33, 44] or leveraging the modeling capacity of neural networks to learn content priors about reflections from a large scale of training data [10, 22, 31]. However, they are prone to fail due to the lack of auxiliary content information about transmission or reflection scenes for solving such a highly ill-posed problem. Recently, language descriptions have shown their effectiveness in providing content information for various vision tasks such as image editing [46, 47, 60, 61], semantic segmentation [57, 65], and image colorization [4-6, 62], which inspires us to think about: Can we leverage the auxiliary content information brought by language descriptions to facilitate the reflection separation problem?" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.885, + 0.339, + 0.901 + ], + "angle": 0, + "content": "\\# Equal contributions. *Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24913" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Since language can effectively convey humans' prior knowledge about the real world [8] and provide auxiliary information of image semantics [65], introducing language descriptions to guide the separation of reflection and transmission layers from mixture images merits exploration. However, leveraging language descriptions for reflection separation is non-trivial in three aspects: 1) Language-image modality inconsistency. Language and images belong to different modalities, thus it is challenging to establish a cross-modality correspondence between the scene content information provided in language descriptions and the complex blended content present in mixture images. 2) Recognizable layer ambiguity. Since the image content and brightness of reflection and transmission layers are different, the recognizable extents of them in mixture images are also uncertain. Specifically, as shown in Figure 1, it is possible that only one layer's content is recognizable (clothes in Figure 1(a)), or both two layers exhibit recognizable content (the trashbin and cars in Figure 1(b)), which leads to the difficulty of using uncertain quantities of language descriptions for separating mixture images in practice. 3) Language annotation deficiency. All existing datasets for the reflection separation task only contain image data but no correlated language description is provided, raising the challenge for network training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.474, + 0.473, + 0.806 + ], + "angle": 0, + "content": "In this paper, we introduce the concept of language-guided image reflection separation for the first time, which leverages flexible natural language to specify the content of one or two layers within a mixture image, relieving the ill-posedness of the reflection separation problem and maintaining a wide applicability for both live captured or online downloaded mixture images. We propose an end-to-end framework that employs adaptive global interaction modules to explore holistic language-image content coherence and utilizes specifically designed loss functions to constrain the correspondence between language descriptions and recovered image layers. A language gate mechanism and a randomized training strategy are designed to deal with the recognizable layer ambiguity problem. To address the language annotation deficiency, we synthesize the training dataset from paired image-language datasets [7, 66] and expand prevailing real reflection separation datasets [31, 49, 67] by manually adding language descriptions. Besides, we further construct a new dataset for visual quality evaluation by collecting mixture images from the Internet and captioning them with language descriptions for recognizable layers. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.841 + ], + "angle": 0, + "content": "- We present the first work that introduces language descriptions to guide the reflection separation task." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.841, + 0.468, + 0.886 + ], + "angle": 0, + "content": "- We propose adaptive global interaction modules and language-image loss functions to tackle modality inconsistency." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- We design a language gate mechanism and a randomized" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.811, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.108 + ], + "angle": 0, + "content": "training strategy to handle recognizable layer ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.108, + 0.89, + 0.137 + ], + "angle": 0, + "content": "- We build a dataset with language descriptions to facilitate language-guided image reflection separation." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.151, + 0.638, + 0.166 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.177, + 0.892, + 0.342 + ], + "angle": 0, + "content": "Single-image reflection separation methods try to distinguish reflection and transmission layers using a single mixture image, which mainly relies on the assumption that the two layers have different distributions, i.e., reflection layers are more likely to be blurry and appear with lower intensity compared with transmission layers. Conventional methods adopt handcrafted priors derived from natural image statistics in their optimization process, e.g., the gradient sparsity [30], relative smoothness [33], ghosting cues [44], content prior [50], and penalty on the gradient of restored transmission layers [64]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.893, + 0.735 + ], + "angle": 0, + "content": "Due to the tremendous progress in the field of deep learning, a series of single-image reflection separation methods concentrate on the improvement of learning strategies or network design, e.g., predicting edges and images with a two-stage [11] or concurrent framework [51, 52], training with the perceptual loss [67], employing generative adversarial network [12] based models [38, 58], adopting iterative refinement strategies [10, 31, 63, 68], and leveraging the complementary two-stream architecture [22]. Meanwhile, research on data synthesis and image models is also ongoing to satisfy the data-driven needs of learning-based methods. Ma et al. [38] utilize generative adversarial networks for data generation while Wen et al. [59] synthesize mixture images with learned non-linear blending masks. Hu et al. [23] introduce a learnable residue term in the mixture image formation model to mitigate the non-linearity caused by the complex camera pipeline. Zheng et al. consider physical factors such as reflective amplitude coefficient maps [69] and the absorption effect [70] in the image formation process of mixture images. To facilitate network training and evaluation, researchers [31, 49, 67] also collect real data by using portable glass. Moreover, as a special form of images, panoramic images are introduced to relieve the content ambiguity in mixture images [13, 19, 21, 40]. We refer readers to [54] for a comprehensive and up-to-date survey on single-image reflection separation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-image reflection separation methods usually leverage the auxiliary information introduced by additional images and achieve more robust performance than single-image methods. Polarization-based methods [9, 26, 28, 36, 37, 39, 43] distinguish reflection and transmission layers by using images captured with different angles of polarizers or special polarization cameras. Flash-based methods [3, 18, 20, 27, 29] adopt active light sources to illuminate transmission scenes for obtaining reflection-free guidance. Motion-based methods [32, 34, 35, 45] utilize multiple images captured from different viewpoints to harness" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.893, + 0.407 + ], + "angle": 0, + "content": "Figure 2. The pipeline of the proposed language-guided image reflection separation framework, which extracts features from mixture images and available language descriptions (the description \\( L_{2} \\) with dashed lines is possible to be set to null due to the recognizable layer ambiguity) by image and language encoders (in Sec. 3.1), aggregates global visual information by adaptive global aggregation modules (AGAM) and conducts progressive interactions to exploit distinctive image features with gated language guidance by adaptive global interaction modules (AGIM) (in Sec. 3.2), and recovers image layers by image decoders (in Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.493 + ], + "angle": 0, + "content": "distinct motions of reflection and transmission layers. However, special data capture requirements significantly limit the application scope of these methods, especially for mobile devices or images downloaded from the Internet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.509, + 0.248, + 0.525 + ], + "angle": 0, + "content": "3. Proposed method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.471, + 0.837 + ], + "angle": 0, + "content": "The pipeline of the proposed framework is illustrated in Figure 2. In this section, we start by introducing the feature extraction stage (in Sec. 3.1) that obtains multi-scale image features and global language features, the gated language interaction stage (in Sec. 3.2) that conducts progressive image-language global interactions to exploit distinctive image features and prevents unavailable language interactions with switchable gates, and the layer recovery stage (in Sec. 3.3) that reprojects features into the image space with a light-weight image decoder. Then we explain loss functions (in Sec. 3.4) employed for network optimization, especially a contrastive correspondence loss and a layer correspondence loss that constrains the network to construct correspondences between the language description and the corresponding layer under the disturbance of the other layer in a superimposed mixture image. Finally, we present our training strategy (in Sec. 3.5) which enables the network to be applicable for varying quantities of input language descriptions and jointly tackles the recognizable layer ambiguity problem with the gated network design." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.41, + 0.863 + ], + "angle": 0, + "content": "3.1. Image and language feature extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The inputs of the proposed method consist of a mixture image \\(\\mathbf{M}\\) with two language descriptions \\(\\{\\mathbf{L}_i|i = 1,2\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.644 + ], + "angle": 0, + "content": "which corresponds to the two image layers. We specify that layer \\(\\mathbf{I}_i\\) corresponds to the description \\(\\mathbf{L}_i\\). However, due to the recognizable layer ambiguity that in certain cases only one layer of the mixture image is recognizable (usually the transmission layer), for such cases, we set \\(\\mathbf{L}_1\\) to be the available language description (for the recognizable transmission layer) and \\(\\mathbf{L}_2\\) to null (for the unrecognizable reflection layer) to ensure a unified input setting. Then given the input image and language descriptions, the feature extraction stage aims at obtaining the image feature \\(\\mathbf{F}_{\\mathrm{M}}\\) with the image encoder and the multi-scale feature fusion process and extracting the global language feature \\(\\mathbf{F}_{\\mathbf{L}_i}\\) for each description \\(\\mathbf{L}_i\\) via the language encoder for the subsequent interaction procedure." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.801 + ], + "angle": 0, + "content": "Image encoder. Given a mixture image \\(\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}\\), we employ a commonly-used vision backbone ResNet-50 [16] as our image encoder, whose last three layers (i.e., an average pooling layer, a fully connected layer, and a softmax layer) are removed to fit our task. We utilize image features from the first five blocks of the image encoder to form a multi-scale feature pyramid \\(\\{\\mathbf{F}_{\\mathbf{M}_i}\\}_{i=1}^5\\), where \\(\\mathbf{F}_{\\mathbf{M}_i} \\in \\mathbb{R}^{h_i \\times w_i \\times C_i}\\), \\(h_i = H/2^i\\) and \\(w_i = W/2^i\\), \\(H\\) and \\(W\\) is the height and width of the mixture image, respectively, and \\(C_i\\) is the dimension of the \\(i\\)-th extracted feature." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.807, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-scale feature fusion. Obtaining the extracted feature pyramid, we first transform it into a hypercolumn feature \\(\\mathbf{F}_{\\mathbf{M}}^{\\mathrm{hyp}} \\in \\mathbb{R}^{h \\times w \\times C^{\\mathrm{hyp}}} [14]\\) (where \\(h = H/2\\), \\(w = W/2\\), and \\(C^{\\mathrm{hyp}} = \\sum_{i=1}^{5} C_i\\)), which has been proved to be effective in fusing multi-scale contextual information for reflection separation [58, 67]. Considering the computational cost," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24915" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.095, + 0.411, + 0.109 + ], + "angle": 0, + "content": "(a) AGAM: Adaptive global aggregation module" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.114, + 0.465, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.216, + 0.403, + 0.231 + ], + "angle": 0, + "content": "(b) AGIM: Adaptive global interaction module" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.236, + 0.465, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.085, + 0.351, + 0.465, + 0.365 + ], + "angle": 0, + "content": "\\(\\otimes\\) Matrix multiplication \\(\\mathbb{R}\\) Reshape \\(\\odot\\) Element-wise multiplication" + }, + { + "type": "image_footnote", + "bbox": [ + 0.085, + 0.369, + 0.465, + 0.381 + ], + "angle": 0, + "content": "Language gate \\(\\odot\\) Concatenation \\(\\oplus\\) Element-wise addition" + }, + { + "type": "list", + "bbox": [ + 0.085, + 0.351, + 0.465, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Figure 3. The architecture of the (a) adaptive global aggregation module (AGAM) and (b) adaptive global interaction module (AGIM), which aggregates global contextual information of visual features and achieves feature channel rearrangement with gated language guidance, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.469, + 0.576 + ], + "angle": 0, + "content": "we condense and refine the hypercolumn feature by a \\(1 \\times 1\\) convolutional layer with a GELU activation [17] followed by a locally-enhanced feed-forward (LeFF) block [56]. The final fused feature of the mixture image is denoted as \\(\\mathbf{F}_{\\mathbf{M}} \\in \\mathbb{R}^{h \\times w \\times C}\\), which serves as the basis for the subsequent interaction and separation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Language encoder. Motivated by the rapid development of pre-trained large-scale vision-language models, we employ the language encoder from CLIP [42], which adopts a Transformer architecture [48] to extract language features and obtains a global contextual feature in the multi-modal embedding space by using layer normalization and linear projection layers. Given a language description \\(\\mathbf{L}_i\\in \\mathbb{R}^L\\) we obtain its corresponding global feature \\(\\mathbf{F}_{\\mathbf{L}_i}\\in \\mathbb{R}^C\\) by leveraging the modeling capacity of the language encoder to encode the description, thus extracting the holistic contextual information of the corresponding image layer. Here \\(L\\) denotes the length of the language description and \\(C\\) is the feature dimension as the image feature." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.322, + 0.802 + ], + "angle": 0, + "content": "3.2. Gated language interaction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.903 + ], + "angle": 0, + "content": "The gated language interaction stage aims at leveraging the contextual information from available language descriptions to guide the separation of the corresponding layer feature \\(\\mathbf{F}_{\\mathbf{I}_i}\\), which is composed of \\(2N\\) cascaded interaction groups to separate image layer features successively. As shown in Figure 2, each group consists of an adaptive global" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "aggregation module (AGAM) to gather global information of input visual features, a language gate to prevent detrimental guidance from unavailable descriptions, an adaptive global interaction module (AGIM) to conduct interactions using global features for exploring holistic image-language content correspondence, and \\(K\\) normalization layers with LeFF blocks [56] for feature refinement. The former \\(N\\) interaction groups are utilized for separating \\(\\mathbf{F}_{\\mathbf{I}_1}\\) from \\(\\mathbf{F}_{\\mathbf{M}}\\) and \\(\\mathbf{F}_{\\mathbf{L}_1}\\), and the latter \\(N\\) groups for separating \\(\\mathbf{F}_{\\mathbf{I}_2}\\) from \\(\\mathbf{F}_{\\mathbf{I}_2}^{\\mathrm{ini}}\\) (obtained by feeding the concatenation of \\(\\mathbf{F}_{\\mathbf{M}}\\) and \\(\\mathbf{F}_{\\mathbf{I}_1}\\) into a \\(1\\times 1\\) convolutional layer) and \\(\\mathbf{F}_{\\mathbf{L}_2}\\) (if \\(\\mathbf{L}_2\\) is available). We set \\(N = 4\\) and \\(K = 2\\) in practice. Details of the gated language interaction are described as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.548 + ], + "angle": 0, + "content": "Adaptive global aggregation module (AGAM). As the network structure shown in Figure 3(a), given an input visual feature \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} \\in \\mathbb{R}^{h \\times w \\times C}\\) with the spatial resolution of \\(h \\times w\\), AGAM is designed for adaptively obtaining a global feature \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} \\in \\mathbb{R}^{C}\\) that aggregates the contextual information for the subsequent interaction. The visual feature \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}\\) is firstly averaged by an average pooling layer to obtain \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}} \\in \\mathbb{R}^{C}\\). Then the aggregation process is accomplished via a cross-attention mechanism which contains three linear projection layers with layer normalization [1] to conduct query, key, and value projections: \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} = \\operatorname{Softmax}(\\mathbf{QK}^{\\top} / \\tau)\\mathbf{V}\\), where the query \\(\\mathbf{Q} \\in \\mathbb{R}^{C}\\) is projected from \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}\\), and the key \\(\\mathbf{K} \\in \\mathbb{R}^{(hw + 1) \\times C}\\) and value \\(\\mathbf{V} \\in \\mathbb{R}^{(hw + 1) \\times C}\\) are from the concatenation of \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}\\) and \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}\\), and \\(\\tau\\) is a learnable scaling factor to control the magnitude of the dot product of \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) before applying the softmax function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Adaptive global interaction module (AGIM). Inspired by existing reflection separation approaches [23, 58] which attempt to distinguish transmission and reflection layers in the feature space by feature channel rearrangement (i.e., allocating distinct channels from mixture image features to the two layers), we propose to integrate language interactions at the feature channel level. To achieve this, as illustrated in Figure 3(b), we propose the adaptive global interaction module (AGIM), which employs the channel-wise cross-attention mechanism to conduct interactions between global language features and image features for channel rearrangement. The query, key, and value projections using linear projection layers with layer normalization [1] are conducted on the global language feature \\(\\mathbf{F}_{\\mathbf{L}_i}\\), the global visual feature \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{gl}}\\), and the visual feature \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}\\) to generate \\(\\mathbf{Q} \\in \\mathbb{R}^{C}\\), \\(\\mathbf{K} \\in \\mathbb{R}^{C}\\), and \\(\\mathbf{V} \\in \\mathbb{R}^{C \\times h w}\\), respectively. Before the interaction, a language gate is designed to prevent impacts of unavailable language descriptions caused by the recognizable layer ambiguity problem that users may only input one description for the layer recognizable in the mixture image. Specifically, if the description \\(\\mathbf{L}_i\\) corresponding to the current global language feature \\(\\mathbf{F}_{\\mathbf{L}_i}\\) is available (not set to null), the gate will feed \\(\\mathbf{F}_{\\mathbf{L}_i}\\) into the following interaction" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24916" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.09, + 0.47, + 0.121 + ], + "angle": 0, + "content": "process, otherwise the gate will feed \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}}\\) , which turns the interaction process to be a channel-wise self-attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.471, + 0.304 + ], + "angle": 0, + "content": "After filtering by the language gate, an interacted feature can be obtained through the channel-wise cross-attention: \\(\\mathbf{F}^{\\mathrm{itr}} = \\operatorname{Softmax}(\\mathbf{Q}^\\top \\mathbf{K} / \\eta)\\mathbf{V}\\), where \\(\\eta\\) is a learnable scaling factor to control the magnitude of the attention map. Besides, to adaptively adjust the influence of language guidance based on the correspondence between image and language features, we define another scaling factor \\(s\\) valued as the cosine distance between \\(\\mathbf{Q}\\) and \\(\\mathbf{K}\\) to multiply with the interacted feature \\(\\mathbf{F}^{\\mathrm{itr}}\\). Finally, the output of AGIM is obtained by a residual structure: \\(\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{out}} = \\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} + s\\mathbf{F}^{\\mathrm{itr}}\\), which integrates the contextual information from the language description and adjusts features for layer separation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.314, + 0.232, + 0.33 + ], + "angle": 0, + "content": "3.3. Layer recovery" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.47, + 0.549 + ], + "angle": 0, + "content": "After global language-image interaction and progressive refinement by the gated language interaction stage, we obtain layer features \\(\\mathbf{F}_{\\mathbf{I}_i}\\) which integrates holistic contextual information from the corresponding language description \\(\\mathbf{L}_i\\) (if available). The layer recovery stage is dedicated to reconstructing each image layer \\(\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}\\) from its corresponding layer features \\(\\mathbf{F}_{\\mathbf{I}_i}\\), which is achieved by using individual image decoders. In the image decoder, the layer feature is firstly upsampled by a transposed convolutional layer with a GELU activation [17], then refined by a residual block, and finally projected from the feature domain back into the image domain by a \\(1 \\times 1\\) convolutional layer with a sigmoid activation, thus accomplishing a precise layer recovery." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.226, + 0.574 + ], + "angle": 0, + "content": "3.4. Loss functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.734 + ], + "angle": 0, + "content": "In this section, we introduce the contrastive correspondence loss \\(\\mathcal{L}_{\\mathrm{ctr}}\\) and the layer correspondence loss \\(\\mathcal{L}_{\\mathrm{lcr}}\\) for constraining the proposed method to establish the cross-modality correspondence between language descriptions \\(\\mathbf{L}_i\\) and corresponding image layers under the content disturbance from counter layers in mixture images. We also briefly describe the image layer loss \\(\\mathcal{L}_{\\mathrm{img}}\\) which consists of image- and feature-level loss functions for layer recovery. We denote the estimated image layers as \\(\\tilde{\\mathbf{I}}_i\\) and ground truths as \\(\\mathbf{I}_i\\). Details of the loss functions are as follows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Contrastive correspondence loss. CLIP [42] conducts contrastive language-image pre-training within batches which successes in establishing cross-modality correspondence between language descriptions and clean images, while our language-guided image reflection separation task requires finding the correspondence between the given language description and a certain layer in the mixture image with the interference of extraneous contents. To tackle the above issue, we propose a contrastive correspondence loss that conducts contrastive learning between image layers to establish correct cross-modality correspondence. Specifi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "cally, given a language description \\(\\mathbf{L}_i\\), the goal of the contrastive correspondence loss is to force the network to learn the relation that the corresponding image layer \\(\\mathbf{I}_i\\) is more relevant to \\(\\mathbf{L}_i\\) than the counter layer \\(\\mathbf{I}_j\\) (\\(j \\neq i\\)), thus constraining the estimated image layer \\(\\tilde{\\mathbf{I}}_i\\) to conform to the associated language description \\(\\mathbf{L}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.23 + ], + "angle": 0, + "content": "To measure the relevance between a language description \\(\\mathbf{L}\\) and an image layer \\(\\mathbf{I}\\), we define a feature-level similarity function \\(\\mathcal{D}(\\cdot)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.241, + 0.892, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} (\\mathbf {L}, \\mathbf {I}) = \\sigma \\left(\\Psi \\left(\\mathbf {F} _ {\\mathbf {L}} ^ {\\mathrm {g l o}}, \\mathbf {F} _ {\\mathbf {I}} ^ {\\mathrm {g l o}}\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.272, + 0.892, + 0.364 + ], + "angle": 0, + "content": "where \\(\\sigma\\) represents the sigmoid function, \\(\\Psi\\) represents the cosine distance, \\(\\mathbf{F}_{\\mathbf{L}}^{\\mathrm{glo}}\\) is the global language feature obtained by the language encoder (in Sec. 3.1), and \\(\\mathbf{F}_{\\mathbf{I}}^{\\mathrm{glo}}\\) is the global image feature produced by the AGAM (in Sec. 3.2). For an available language description \\(\\mathbf{L}_i\\), we calculate its contrastive correspondence loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.377, + 0.892, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {j}\\right) = - \\log \\left(\\frac {\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right)}{\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right) + \\mathcal {D} \\left(\\mathbf {L} _ {i} , \\mathbf {I} _ {j}\\right)}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.453 + ], + "angle": 0, + "content": "and we sum up the contrastive correspondence loss of each available language description as the final one." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.53 + ], + "angle": 0, + "content": "Layer correspondence loss. We further define a layer correspondence loss to encourage the relevance between the language description \\(\\mathbf{L}_i\\) and the estimated image layer \\(\\tilde{\\mathbf{I}}_i\\) to approach the relevance between \\(\\mathbf{L}_i\\) and the ground truth image layer \\(\\mathbf{I}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.541, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {l c r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {i}\\right) = \\left\\| \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}\\right) - \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\mathbf {I} _ {i}\\right) \\right\\| _ {1}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.639 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}(\\cdot)\\) is the same feature-level similarity function as in the contrastive correspondence loss. We also sum up the layer correspondence loss of each available language description for final supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.806 + ], + "angle": 0, + "content": "Image layer loss. To achieve high-fidelity recovery of image layers (i.e., transmission and reflection layers), the proposed method is also optimized with loss functions following previous reflection separation methods [23, 52, 67]. Specifically, we utilize loss functions that conduct constraints on the visual quality of estimated images (i.e., the pixel \\(\\mathcal{L}_{\\mathrm{pix}}\\), structural similarity \\(\\mathcal{L}_{\\mathrm{ssim}}\\), and perceptual loss \\(\\mathcal{L}_{\\mathrm{per}}\\)) or exploit the inherent relationship between two layers (i.e., the exclusion \\(\\mathcal{L}_{\\mathrm{exc}}\\) and reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\)), and we denote the combination of the above image- or feature-level loss functions as the image layer loss \\(\\mathcal{L}_{\\mathrm{img}}^1\\)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.806, + 0.868, + 0.821 + ], + "angle": 0, + "content": "Overall, the total loss function is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.834, + 0.892, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {c t r}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {l c r}} + \\mathcal {L} _ {\\mathrm {i m g}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.863, + 0.792, + 0.878 + ], + "angle": 0, + "content": "where coefficients are set as \\(\\gamma_{1} = \\gamma_{2} = 0.5\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.887, + 0.843, + 0.902 + ], + "angle": 0, + "content": "Details of \\(\\mathcal{L}_{\\mathrm{img}}\\) are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24917" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.248, + 0.108 + ], + "angle": 0, + "content": "3.5. Training strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.342 + ], + "angle": 0, + "content": "Due to the recognizable layer ambiguity that sometimes only one layer in the mixture image is recognizable, we propose a randomized training strategy to synergize with the gated language interaction mechanism (in Sec. 3.2). Since our training data simulates the recognizable layer ambiguity that some image layers do not have corresponding language descriptions (introduced in the next section), we only feed the available language description corresponding to the other image layer to guide the separation. For data with descriptions of both layers available, which indicates that both layers are recognizable, we also randomly drop one language description and feed the remaining one into the network to improve the generalization capacity for the proposed method. In practice, we set the ratio of dropping language descriptions to \\(30\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.472, + 0.435 + ], + "angle": 0, + "content": "We implement the proposed method with PyTorch [41] with a batch size of 16 on two Nvidia GeForce RTX 3090 GPUs. The model is trained for 40 epochs with Adam optimizer [25] to update learnable parameters. Weights are initialized as in [15]. The learning rate is set to \\(10^{-4}\\) initially and decreases to \\(10^{-5}\\) at epoch 30." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.452, + 0.248, + 0.469 + ], + "angle": 0, + "content": "4. Data preparation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.473, + 0.63 + ], + "angle": 0, + "content": "Though existing works have constructed several datasets for single-image reflection separation [31, 49, 67], they are unavailable for the proposed language-guided reflection separation framework due to the lack of corresponding language descriptions. Therefore, we build a dataset containing both synthetic and real data to overcome the data deficiency and facilitate network training and evaluation. Each group of data is composed of a mixture image, a transmission layer, a reflection layer, and two language descriptions. Details of synthetic and real data are as follows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.226, + 0.658 + ], + "angle": 0, + "content": "4.1. Synthetic data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.819 + ], + "angle": 0, + "content": "The synthetic dataset is generated for network training to satisfy the data-driven need of the proposed method. Due to the demand for paired image-language data, we utilize two prevailing image captioning datasets (i.e., Flickr30k [66] and COCO Captions [7]) for data generation, which contain 31,000 and 330,000 images respectively, and each image has 5 independent human-generated language descriptions. We randomly select images from the above two datasets as transmission \\(\\mathbf{T}_{\\mathrm{S}}\\) and reflection scene images \\(\\mathbf{R}_{\\mathrm{S}}\\) and conduct an image synthesis process with linear blending [23]:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.827, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} = \\hat {\\mathbf {T}} + \\hat {\\mathbf {R}} = \\alpha \\hat {\\mathbf {T}} _ {\\mathrm {S}} + \\beta \\hat {\\mathbf {R}} _ {\\mathrm {S}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.855, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{T}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{T}_{\\mathrm{S}})\\) and \\(\\hat{\\mathbf{R}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{R}_{\\mathrm{S}})\\), \\(g_{\\mathrm{inv}}\\) represents the inverse gamma correction, and \\(\\alpha \\in [0.8,1]\\) and \\(\\beta \\in [0.4,1]\\) are the blending attenuation coefficients as in [23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Considering the recognizable layer ambiguity that sometimes only one layer is recognizable in a mixture image, we assign a language description only when the corresponding layer is obvious enough in the synthesized mixture image, i.e., \\(\\frac{\\mathrm{mean}(\\hat{\\mathbf{V}}_l)}{\\mathrm{mean}(\\hat{\\mathbf{V}}_{\\mathbf{M}})}\\geqslant \\mu\\), where \\(\\hat{\\mathbf{V}}_l\\) represents the brightness image of \\(\\hat{\\mathbf{T}}\\) or \\(\\hat{\\mathbf{R}}\\) in the HSV color space and \\(\\hat{\\mathbf{V}}_{\\mathbf{M}}\\) represents the brightness image of \\(\\hat{\\mathbf{M}}\\), and we set \\(\\mu = 0.3\\) in the data generation process. Finally, gamma correction is applied to image triplets \\(\\{\\hat{\\mathbf{T}},\\hat{\\mathbf{R}},\\hat{\\mathbf{M}}\\}\\) to obtain \\(\\{\\mathbf{T},\\mathbf{R},\\mathbf{M}\\}\\), and we generate 50000 triplets of data in total for network training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.266, + 0.611, + 0.28 + ], + "angle": 0, + "content": "4.2. Real data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.668 + ], + "angle": 0, + "content": "Existing real datasets collected for the single-image reflection separation task usually contain mixture images with ground truth of transmission layers (e.g., Zhang et al. [67] and Nature [31]), and \\(\\mathrm{SIR}^2\\) [49, 54] further captures ground truths of reflection layers. For these off-the-shelf real datasets, we augment them by manually adding language descriptions for each group of data to satisfy the input setting of the proposed language-guided reflection separation task. Specifically, following the annotation principle of COCO Captions [7], we first describe the content of transmission layers in mixture images with entities, attributes (e.g., colors or materials), and relative positions between different entities. If the content of reflection layers is recognizable in mixture images, we also give language descriptions for reflection layers in the same way. To further evaluate the generalization capacity of the proposed method, we collect a real dataset (denoted as REFOL dataset) containing 100 mixture images from the Internet that are captured in different scenes and with different cameras. We annotate these mixture images with language descriptions in the same manner as mentioned above. Following the training strategy of previous methods [10, 31], we utilize 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [67] for training, and the rest of real data are used for quantitative and qualitative evaluation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.634, + 0.701 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.86, + 0.727 + ], + "angle": 0, + "content": "5.1. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To evaluate the performance of the proposed method, we conduct quantitative and qualitative experiments on existing real datasets [31, 49, 67] (with our manually annotated language descriptions) and our newly collected dataset REFOL. We compare with state-of-the-art single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], IBCLN [31], CoRRN [52], and Zhang et al. [67]. For fair comparisons, we finetune the above methods on our training data if their training codes are provided. We report better results between the original pre-trained model and the finetuned version." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24918" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison of quantitative results in terms of PSNR [24] and SSIM [55] on real datasets for evaluating the recovery of transmission layers. \\( \\uparrow (\\downarrow) \\) indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.122, + 0.85, + 0.307 + ], + "angle": 0, + "content": "
Dataset (size)MetricsMethods
Zhang et al. [67]CoRRN [52]IBCLN [31]Dong et al. [10]YTMT [22]DSRNet [23]Ours
Postcard (199)PSNR↑20.8522.0423.4123.7222.8224.8825.02
SSIM↑0.8720.8700.8720.9030.8850.9100.915
Object (200)PSNR↑23.8425.1324.5224.3624.6826.4426.51
SSIM↑0.8720.9120.8910.8980.8920.9210.927
Wild (101)PSNR↑24.9725.1724.7825.7525.7025.8626.23
SSIM↑0.8750.8890.8840.9030.8970.9080.925
Real20 (20)PSNR↑22.3421.4321.4723.3423.2323.8824.05
SSIM↑0.7950.8010.7620.8120.8020.8160.824
Nature (20)PSNR↑20.6220.7523.7223.4521.5322.2623.87
SSIM↑0.7530.7830.8060.8080.7780.8010.812
Average (540)PSNR↑22.7723.7024.0224.3124.0125.5125.72
SSIM↑0.8650.8830.8750.8940.8830.9060.914
" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.323, + 0.351, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.323, + 0.619, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.323, + 0.888, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.442, + 0.351, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.442, + 0.618, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.442, + 0.888, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.566, + 0.351, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.566, + 0.619, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.566, + 0.888, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.687, + 0.351, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.687, + 0.619, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.687, + 0.888, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.805, + 0.892, + 0.833 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of estimated transmission and reflection layers on real data, compared with the state-of-the-art methods including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Please zoom in for details." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.859, + 0.471, + 0.89 + ], + "angle": 0, + "content": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection separation," + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.859, + 0.892, + 0.89 + ], + "angle": 0, + "content": "i.e., Nature [31], Real20 [67], and three subsets of \\(\\mathrm{SIR}^2\\) [49] dataset. Following the setting of existing reflection separa" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24919" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.09, + 0.414, + 0.103 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation studies." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.109, + 0.466, + 0.164 + ], + "angle": 0, + "content": "
Metricsw/o languagew/o AGIMLimg onlyOurs
PSNR↑24.3124.6924.5225.72
SSIM↑0.8850.9010.8930.914
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.468, + 0.273 + ], + "angle": 0, + "content": "tion methods [10, 36], we utilize PSNR [24] and SSIM [55] as error metrics for evaluating the recovery of transmission layers2. As quantitative results shown in Table 1, the proposed method achieves the best performance of both PSNR and SSIM, which validates its generalization capacity and the effectiveness of language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.469, + 0.515 + ], + "angle": 0, + "content": "Qualitative comparison. To evaluate the visual quality of reflection separation results, we compare the proposed method with four single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Qualitative results on recovering both transmission and reflection layers are shown in Figure 4. As can be observed in Figure 4, IBCLN [31] can only remove parts of reflections, while Dong et al. [10] have trouble in dealing with complex semantic images like the reflections of a man in a room (the second example). YTMT [22] separates layers incorrectly and thus brings the content of transmission layers into reflections. DSRNet [23] fails to recover transmission and reflection layers for these challenging cases. Contributing to the language guidance, the proposed method generates better visual results and recovers both transmission and reflection layers neatly." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.525, + 0.227, + 0.541 + ], + "angle": 0, + "content": "5.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.469, + 0.819 + ], + "angle": 0, + "content": "In this section, we conduct several ablation studies with quantitative results shown in Table 2 to investigate the influence of the additional input of language descriptions (denoted as 'w/o language'), the network design of AGIM (denoted as 'w/o AGIM'), and the loss functions for cross-modality correspondence (denoted as \\( \\mathcal{L}_{\\mathrm{img}} \\) only). The performance of the variant 'w/o language' suffers from an obvious degradation as we remove language descriptions, which shows the effectiveness of the additional contextual information for layer separation. The variant 'w/o AGIM' replaces AGIMs with simple feature fusion blocks which directly concatenate language and image features and feed them into self-attention blocks, and the decline in performance validates the necessity of our gated interaction mechanism. The variant \\( \\mathcal{L}_{\\mathrm{img}} \\) only is trained with \\( \\mathcal{L}_{\\mathrm{img}} \\), which obtains results slightly better than the variant 'w/o language', indicating the significance of establishing the cross-modality correspondence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.82, + 0.468, + 0.866 + ], + "angle": 0, + "content": "To further verify the effectiveness of the language interaction, we conduct an ablation study by gradually increasing the number of input descriptions. As shown in yellow" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.888, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.337, + 0.892, + 0.365 + ], + "angle": 0, + "content": "Figure 5. Results with different numbers of input language descriptions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.379, + 0.892, + 0.456 + ], + "angle": 0, + "content": "boxes of Figure 5, by utilizing more input language descriptions, the separation of transmission (clothes) and reflection layers (a man with a camera) becomes more thorough, which also demonstrates the efficacy and robustness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.477, + 0.619, + 0.492 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.716 + ], + "angle": 0, + "content": "This paper introduces natural language to provide contextual information about image layers for relieving the ill-posed reflection separation problem. We develop an end-to-end framework with adaptive global interaction modules and language-image loss functions to effectively manage the modality inconsistency, and we adopt a language gate mechanism with randomized training strategies to handle the recognizable layer ambiguity. To address data deficiency, a specially built dataset with language annotations significantly aids in training and evaluating the proposed language-guided image reflection separation framework. Quantitative and qualitative experiments on real data demonstrate the effectiveness of introducing language descriptions for reflection separation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Limitations. The proposed method may not distinguish transmission and reflection layers accurately when their contents are similar. For such ambiguous cases, a more flexible language-guided mechanism is needed. This might be solved by exploring a better interaction approach, which is left as our future work." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.832, + 0.66, + 0.848 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.858, + 0.892, + 0.887 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.468, + 0.9 + ], + "angle": 0, + "content": "2Quantitative evaluations on the recovery of reflection layers are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24920" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Yakun Chang, Cheolkon Jung, and Jun Sun. Joint reflection removal and depth estimation from a single image. IEEE TCYB, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.202, + 0.471, + 0.243 + ], + "angle": 0, + "content": "[3] Yakun Chang, Cheolkon Jung, Jun Sun, and Fengqiao Wang. Siamese dense network for reflection removal with flash and no-flash image pairs. IJCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.286 + ], + "angle": 0, + "content": "[4] Zheng Chang, Shuchen Weng, Yu Li, Si Li, and Boxin Shi. L-CoDer: Language-based colorization with color-object decoupling transformer. In Proc. of ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[5] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CoIns: Language-based colorization with instance awareness. In Proc. of CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.471, + 0.386 + ], + "angle": 0, + "content": "[6] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Proc. of NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. NeRDi: Single-view NeRF synthesis with language-guided diffusion as general image priors. In Proc. of CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[9] Yaron Diamant and Yoav Y Schechner. Overcoming visual reverberations. In Proc. of CVPR, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[10] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson WH Lau. Location-aware single image reflection removal. In Proc. of ICCV, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.574, + 0.47, + 0.627 + ], + "angle": 0, + "content": "[11] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf. A generic deep architecture for single image reflection removal and image smoothing. In Proc. of ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. of NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[13] Byeong-Ju Han and Jae-Young Sim. Zero-shot learning for reflection removal of single 360-degree image. In Proc. of ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[14] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In Proc. of CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.828 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance onImagenet classification. In Proc. of ICCV, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[18] Yuchen Hong, Youwei Lyu, Si Li, and Boxin Shi. Near-infrared image guided reflection removal. In Proc. of ICME, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[19] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. Panoramic image reflection removal. In Proc. of CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[20] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with NIR and RGB image feature fusion. IEEE TMM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[21] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. PAR\\(^2\\)Net: End-to-end panoramic image reflection removal. IEEE TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[22] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation.* In *Proc. of NeurIPS*, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[23] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proc. of ICCV, 2023. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[24] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics letters*, 2008, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[26] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. A physically-based approach to reflection separation. In Proc. of CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[27] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proc. of CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[28] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proc. of CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[29] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[30] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE TPAMI, 2007. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[31] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proc. of CVPR, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[32] Yu Li and Michael S Brown. Exploiting reflection change for automatic reflection removal. In Proc. of ICCV, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[33] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proc. of CVPR, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[34] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In Proc. of CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[35] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. IEEE TPAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[36] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolar" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24921" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "ized and polarized images. In Proc. of NeurIPS, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.163 + ], + "angle": 0, + "content": "[37] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[38] Daiqian Ma, Renjie Wan, Boxin Shi, Alex C Kot, and Ling-Yu Duan. Learning to jointly generate and separate reflections. In Proc. of ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[39] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. IJCV, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.291 + ], + "angle": 0, + "content": "[40] Jonghyuk Park, Hyeona Kim, Eunpil Park, and Jae-Young Sim. Fully-automatic reflection removal for 360-degree images. In Proc. of WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.469, + 0.361 + ], + "angle": 0, + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Proc. of NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.432 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. of ICML, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[43] Yoav Y Schechner, Nahum Kiryati, and Ronen Basri. Separation of transparent layers using focus. IJCV, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[44] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proc. of CVPR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[45] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In Proc. of CVPR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[46] Jimeng Sun, Shuchen Weng, Zheng Chang, Si Li, and Boxin Shi. UniCoRN: A unified conditional image inpainting network. In Proc. of CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[47] Jiajun Tang, Haofeng Zhong, Shuchen Weng, and Boxin Shi. LuminAIRe: Illumination-aware conditional image inpainting for lighting-realistic generation. In Proc. of NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[49] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In Proc. of ICCV, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[50] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, Wen Gao, and Alex C. Kot. Region-aware reflection removal with unified content and gradient priors. IEEE TIP, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[51] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CRRN: Multi-scale guided concurrent reflection removal network. In Proc. of CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[52] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CoRRN: Cooperative reflection removal network. IEEE TPAMI, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[53] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. IJCV, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[54] Renjie Wan, Boxin Shi, Haoliang Li, Yuchen Hong, Ling-Yu Duan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In IEEE TPAMI, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[55] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In Proc. of ACSSC, 2003. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.231 + ], + "angle": 0, + "content": "[56] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proc. of CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[57] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: CLIP-driven referring image segmentation. In Proc. of CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[58] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proc. of CVPR, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[59] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In Proc. of CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.891, + 0.412 + ], + "angle": 0, + "content": "[60] Shuchen Weng and Boxin Shi. Conditional image repainting. IEEE TPAMI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[61] Shuchen Weng, Wenbo Li, Dawei Li, Hongxia Jin, and Boxin Shi. MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In Proc. of CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[62] Shuchen Weng, Hao Wu, Zheng Chang, Jiajun Tang, Si Li, and Boxin Shi. L-CoDe: Language-based colorization using color-object decoupled conditions. In Proc. of AAAI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[63] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In Proc. of ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[64] Yang Yang, Wenye Ma, Yin Zheng, Jian-Feng Cai, and Weiyu Xu. Fast single image reflection suppression via convex optimization. In Proc. of CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[65] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-aware vision transformer for referring image segmentation. In Proc. of CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[66] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[67] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proc. of CVPR, 2018. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[68] Ya-Nan Zhang, Linlin Shen, and Qiufu Li. Content and gradient model-driven deep network for single image reflection removal. In Proc. of ACM MM, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[69] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In Proc. of CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[70] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In Proc. of CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24922" + } + ] +] \ No newline at end of file diff --git a/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_origin.pdf b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6c7f6505b34dd4fe1327d63d55f7d6e6c2ddfd8e --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/2783b2be-5a5b-4ab1-8b98-8dc1d3acc65d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2a2c38de08fd21b399ac5257789599b0c4e71afa8294451d4843af7634c8dff +size 1493574 diff --git a/2024/Language-guided Image Reflection Separation/full.md b/2024/Language-guided Image Reflection Separation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8a9d3c17c5c9e8816882b8323f2bf24a52c6fd01 --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/full.md @@ -0,0 +1,304 @@ +# Language-guided Image Reflection Separation + +Haofeng Zhong $^{1,2,3\#}$ Yuchen Hong $^{1,2\#}$ Shuchen Weng $^{1,2}$ Jinxiu Liang $^{1,2}$ Boxin Shi $^{1,2,3*}$ + +$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com + +# Abstract + +This paper studies the problem of language-guided reflection separation, which aims at addressing the ill-posed reflection separation problem by introducing language descriptions to provide layer content. We propose a unified framework to solve this problem, which leverages the crossattention mechanism with contrastive learning strategies to construct the correspondence between language descriptions and image layers. A gated network design and a randomized training strategy are employed to tackle the recognizable layer ambiguity. The effectiveness of the proposed method is validated by the significant performance advantage over existing reflection separation methods on both quantitative and qualitative comparisons. + +# 1. Introduction + +When photographing through transparent materials like glass windows or showcases, the presence of reflections can significantly degrade the image quality of captured images and disrupt downstream computer vision tasks like face recognition [53] or depth estimation [2]. As an attractive topic in computational photography, reflection separation aims at decomposing the contaminated mixture image (denoted as $\mathbf{M}$ ) into two components that correspond to scenes located at different sides of the glass, i.e., the reflection layer (denoted as $\mathbf{R}$ ) and the transmission layer (denoted as $\mathbf{T}$ ). Since reflection separation is a severely ill-posed problem, it is imperative to exploit effective priors or auxiliary information for distinguishing the two components. + +The primary challenge of solving the reflection separation problem lies in the exploration of distinct clues for distinguishing transmission and reflection layers. Multi-image methods handle this problem by introducing additional constraints. Some of them acquire a series of mixture images in different viewpoints [32, 34, 45] to harness the distinct motions of the two layers, while others adopt specialized capturing setups to obtain complementary scene informa + +![](images/f45362c64c63d2eacf3472b97a8ae2ef172e097b9bbc0b323aafe6087713a14b.jpg) +Figure 1. The recognizable layer ambiguity problem causes uncertain quantities of input language descriptions for language-guided image reflection separation. Given language descriptions of either (a) one layer or (b) both two layers, the proposed method achieves robust reflection separation compared with an existing reflection separation method [10]. + +tion [19, 27, 28, 36]. However, the specialized data capture requirements limit the application scope of these methods, especially for images downloaded from the Internet. Single-image methods attempt to tackle reflection separation by utilizing handcrafted priors derived from natural image statistics [30, 33, 44] or leveraging the modeling capacity of neural networks to learn content priors about reflections from a large scale of training data [10, 22, 31]. However, they are prone to fail due to the lack of auxiliary content information about transmission or reflection scenes for solving such a highly ill-posed problem. Recently, language descriptions have shown their effectiveness in providing content information for various vision tasks such as image editing [46, 47, 60, 61], semantic segmentation [57, 65], and image colorization [4-6, 62], which inspires us to think about: Can we leverage the auxiliary content information brought by language descriptions to facilitate the reflection separation problem? + +Since language can effectively convey humans' prior knowledge about the real world [8] and provide auxiliary information of image semantics [65], introducing language descriptions to guide the separation of reflection and transmission layers from mixture images merits exploration. However, leveraging language descriptions for reflection separation is non-trivial in three aspects: 1) Language-image modality inconsistency. Language and images belong to different modalities, thus it is challenging to establish a cross-modality correspondence between the scene content information provided in language descriptions and the complex blended content present in mixture images. 2) Recognizable layer ambiguity. Since the image content and brightness of reflection and transmission layers are different, the recognizable extents of them in mixture images are also uncertain. Specifically, as shown in Figure 1, it is possible that only one layer's content is recognizable (clothes in Figure 1(a)), or both two layers exhibit recognizable content (the trashbin and cars in Figure 1(b)), which leads to the difficulty of using uncertain quantities of language descriptions for separating mixture images in practice. 3) Language annotation deficiency. All existing datasets for the reflection separation task only contain image data but no correlated language description is provided, raising the challenge for network training and evaluation. + +In this paper, we introduce the concept of language-guided image reflection separation for the first time, which leverages flexible natural language to specify the content of one or two layers within a mixture image, relieving the ill-posedness of the reflection separation problem and maintaining a wide applicability for both live captured or online downloaded mixture images. We propose an end-to-end framework that employs adaptive global interaction modules to explore holistic language-image content coherence and utilizes specifically designed loss functions to constrain the correspondence between language descriptions and recovered image layers. A language gate mechanism and a randomized training strategy are designed to deal with the recognizable layer ambiguity problem. To address the language annotation deficiency, we synthesize the training dataset from paired image-language datasets [7, 66] and expand prevailing real reflection separation datasets [31, 49, 67] by manually adding language descriptions. Besides, we further construct a new dataset for visual quality evaluation by collecting mixture images from the Internet and captioning them with language descriptions for recognizable layers. Our contributions are summarized as follows: + +- We present the first work that introduces language descriptions to guide the reflection separation task. +- We propose adaptive global interaction modules and language-image loss functions to tackle modality inconsistency. +- We design a language gate mechanism and a randomized + +training strategy to handle recognizable layer ambiguity. + +- We build a dataset with language descriptions to facilitate language-guided image reflection separation. + +# 2. Related work + +Single-image reflection separation methods try to distinguish reflection and transmission layers using a single mixture image, which mainly relies on the assumption that the two layers have different distributions, i.e., reflection layers are more likely to be blurry and appear with lower intensity compared with transmission layers. Conventional methods adopt handcrafted priors derived from natural image statistics in their optimization process, e.g., the gradient sparsity [30], relative smoothness [33], ghosting cues [44], content prior [50], and penalty on the gradient of restored transmission layers [64]. + +Due to the tremendous progress in the field of deep learning, a series of single-image reflection separation methods concentrate on the improvement of learning strategies or network design, e.g., predicting edges and images with a two-stage [11] or concurrent framework [51, 52], training with the perceptual loss [67], employing generative adversarial network [12] based models [38, 58], adopting iterative refinement strategies [10, 31, 63, 68], and leveraging the complementary two-stream architecture [22]. Meanwhile, research on data synthesis and image models is also ongoing to satisfy the data-driven needs of learning-based methods. Ma et al. [38] utilize generative adversarial networks for data generation while Wen et al. [59] synthesize mixture images with learned non-linear blending masks. Hu et al. [23] introduce a learnable residue term in the mixture image formation model to mitigate the non-linearity caused by the complex camera pipeline. Zheng et al. consider physical factors such as reflective amplitude coefficient maps [69] and the absorption effect [70] in the image formation process of mixture images. To facilitate network training and evaluation, researchers [31, 49, 67] also collect real data by using portable glass. Moreover, as a special form of images, panoramic images are introduced to relieve the content ambiguity in mixture images [13, 19, 21, 40]. We refer readers to [54] for a comprehensive and up-to-date survey on single-image reflection separation. + +Multi-image reflection separation methods usually leverage the auxiliary information introduced by additional images and achieve more robust performance than single-image methods. Polarization-based methods [9, 26, 28, 36, 37, 39, 43] distinguish reflection and transmission layers by using images captured with different angles of polarizers or special polarization cameras. Flash-based methods [3, 18, 20, 27, 29] adopt active light sources to illuminate transmission scenes for obtaining reflection-free guidance. Motion-based methods [32, 34, 35, 45] utilize multiple images captured from different viewpoints to harness + +![](images/f8d20ba08f8af659fe038800ae98c653d32eb32f2488da12dd6d017a5ad91a67.jpg) +Figure 2. The pipeline of the proposed language-guided image reflection separation framework, which extracts features from mixture images and available language descriptions (the description $L_{2}$ with dashed lines is possible to be set to null due to the recognizable layer ambiguity) by image and language encoders (in Sec. 3.1), aggregates global visual information by adaptive global aggregation modules (AGAM) and conducts progressive interactions to exploit distinctive image features with gated language guidance by adaptive global interaction modules (AGIM) (in Sec. 3.2), and recovers image layers by image decoders (in Sec. 3.3). + +distinct motions of reflection and transmission layers. However, special data capture requirements significantly limit the application scope of these methods, especially for mobile devices or images downloaded from the Internet. + +# 3. Proposed method + +The pipeline of the proposed framework is illustrated in Figure 2. In this section, we start by introducing the feature extraction stage (in Sec. 3.1) that obtains multi-scale image features and global language features, the gated language interaction stage (in Sec. 3.2) that conducts progressive image-language global interactions to exploit distinctive image features and prevents unavailable language interactions with switchable gates, and the layer recovery stage (in Sec. 3.3) that reprojects features into the image space with a light-weight image decoder. Then we explain loss functions (in Sec. 3.4) employed for network optimization, especially a contrastive correspondence loss and a layer correspondence loss that constrains the network to construct correspondences between the language description and the corresponding layer under the disturbance of the other layer in a superimposed mixture image. Finally, we present our training strategy (in Sec. 3.5) which enables the network to be applicable for varying quantities of input language descriptions and jointly tackles the recognizable layer ambiguity problem with the gated network design. + +# 3.1. Image and language feature extraction + +The inputs of the proposed method consist of a mixture image $\mathbf{M}$ with two language descriptions $\{\mathbf{L}_i|i = 1,2\}$ + +which corresponds to the two image layers. We specify that layer $\mathbf{I}_i$ corresponds to the description $\mathbf{L}_i$ . However, due to the recognizable layer ambiguity that in certain cases only one layer of the mixture image is recognizable (usually the transmission layer), for such cases, we set $\mathbf{L}_1$ to be the available language description (for the recognizable transmission layer) and $\mathbf{L}_2$ to null (for the unrecognizable reflection layer) to ensure a unified input setting. Then given the input image and language descriptions, the feature extraction stage aims at obtaining the image feature $\mathbf{F}_{\mathrm{M}}$ with the image encoder and the multi-scale feature fusion process and extracting the global language feature $\mathbf{F}_{\mathbf{L}_i}$ for each description $\mathbf{L}_i$ via the language encoder for the subsequent interaction procedure. + +Image encoder. Given a mixture image $\mathbf{M} \in \mathbb{R}^{H \times W \times 3}$ , we employ a commonly-used vision backbone ResNet-50 [16] as our image encoder, whose last three layers (i.e., an average pooling layer, a fully connected layer, and a softmax layer) are removed to fit our task. We utilize image features from the first five blocks of the image encoder to form a multi-scale feature pyramid $\{\mathbf{F}_{\mathbf{M}_i}\}_{i=1}^5$ , where $\mathbf{F}_{\mathbf{M}_i} \in \mathbb{R}^{h_i \times w_i \times C_i}$ , $h_i = H/2^i$ and $w_i = W/2^i$ , $H$ and $W$ is the height and width of the mixture image, respectively, and $C_i$ is the dimension of the $i$ -th extracted feature. + +Multi-scale feature fusion. Obtaining the extracted feature pyramid, we first transform it into a hypercolumn feature $\mathbf{F}_{\mathbf{M}}^{\mathrm{hyp}} \in \mathbb{R}^{h \times w \times C^{\mathrm{hyp}}} [14]$ (where $h = H/2$ , $w = W/2$ , and $C^{\mathrm{hyp}} = \sum_{i=1}^{5} C_i$ ), which has been proved to be effective in fusing multi-scale contextual information for reflection separation [58, 67]. Considering the computational cost, + +![](images/75de14e1787dca95ffbbbd42d5c9259507ba41a253e09c356f3e49918a8fefe8.jpg) +(a) AGAM: Adaptive global aggregation module + +(b) AGIM: Adaptive global interaction module +Figure 3. The architecture of the (a) adaptive global aggregation module (AGAM) and (b) adaptive global interaction module (AGIM), which aggregates global contextual information of visual features and achieves feature channel rearrangement with gated language guidance, respectively. +![](images/9a30e18b3b14960b596b66f902c24e2e5b54bbab10688658b825283817f37c6d.jpg) +$\otimes$ Matrix multiplication $\mathbb{R}$ Reshape $\odot$ Element-wise multiplication +Language gate $\odot$ Concatenation $\oplus$ Element-wise addition + +we condense and refine the hypercolumn feature by a $1 \times 1$ convolutional layer with a GELU activation [17] followed by a locally-enhanced feed-forward (LeFF) block [56]. The final fused feature of the mixture image is denoted as $\mathbf{F}_{\mathbf{M}} \in \mathbb{R}^{h \times w \times C}$ , which serves as the basis for the subsequent interaction and separation process. + +Language encoder. Motivated by the rapid development of pre-trained large-scale vision-language models, we employ the language encoder from CLIP [42], which adopts a Transformer architecture [48] to extract language features and obtains a global contextual feature in the multi-modal embedding space by using layer normalization and linear projection layers. Given a language description $\mathbf{L}_i\in \mathbb{R}^L$ we obtain its corresponding global feature $\mathbf{F}_{\mathbf{L}_i}\in \mathbb{R}^C$ by leveraging the modeling capacity of the language encoder to encode the description, thus extracting the holistic contextual information of the corresponding image layer. Here $L$ denotes the length of the language description and $C$ is the feature dimension as the image feature. + +# 3.2. Gated language interaction + +The gated language interaction stage aims at leveraging the contextual information from available language descriptions to guide the separation of the corresponding layer feature $\mathbf{F}_{\mathbf{I}_i}$ , which is composed of $2N$ cascaded interaction groups to separate image layer features successively. As shown in Figure 2, each group consists of an adaptive global + +aggregation module (AGAM) to gather global information of input visual features, a language gate to prevent detrimental guidance from unavailable descriptions, an adaptive global interaction module (AGIM) to conduct interactions using global features for exploring holistic image-language content correspondence, and $K$ normalization layers with LeFF blocks [56] for feature refinement. The former $N$ interaction groups are utilized for separating $\mathbf{F}_{\mathbf{I}_1}$ from $\mathbf{F}_{\mathbf{M}}$ and $\mathbf{F}_{\mathbf{L}_1}$ , and the latter $N$ groups for separating $\mathbf{F}_{\mathbf{I}_2}$ from $\mathbf{F}_{\mathbf{I}_2}^{\mathrm{ini}}$ (obtained by feeding the concatenation of $\mathbf{F}_{\mathbf{M}}$ and $\mathbf{F}_{\mathbf{I}_1}$ into a $1\times 1$ convolutional layer) and $\mathbf{F}_{\mathbf{L}_2}$ (if $\mathbf{L}_2$ is available). We set $N = 4$ and $K = 2$ in practice. Details of the gated language interaction are described as follows. + +Adaptive global aggregation module (AGAM). As the network structure shown in Figure 3(a), given an input visual feature $\mathbf{F}_{\mathbf{V}}^{\mathrm{in}} \in \mathbb{R}^{h \times w \times C}$ with the spatial resolution of $h \times w$ , AGAM is designed for adaptively obtaining a global feature $\mathbf{F}_{\mathbf{V}}^{\mathrm{glo}} \in \mathbb{R}^{C}$ that aggregates the contextual information for the subsequent interaction. The visual feature $\mathbf{F}_{\mathbf{V}}^{\mathrm{in}}$ is firstly averaged by an average pooling layer to obtain $\mathbf{F}_{\mathbf{V}}^{\mathrm{avg}} \in \mathbb{R}^{C}$ . Then the aggregation process is accomplished via a cross-attention mechanism which contains three linear projection layers with layer normalization [1] to conduct query, key, and value projections: $\mathbf{F}_{\mathbf{V}}^{\mathrm{glo}} = \operatorname{Softmax}(\mathbf{QK}^{\top} / \tau)\mathbf{V}$ , where the query $\mathbf{Q} \in \mathbb{R}^{C}$ is projected from $\mathbf{F}_{\mathbf{V}}^{\mathrm{avg}}$ , and the key $\mathbf{K} \in \mathbb{R}^{(hw + 1) \times C}$ and value $\mathbf{V} \in \mathbb{R}^{(hw + 1) \times C}$ are from the concatenation of $\mathbf{F}_{\mathbf{V}}^{\mathrm{in}}$ and $\mathbf{F}_{\mathbf{V}}^{\mathrm{avg}}$ , and $\tau$ is a learnable scaling factor to control the magnitude of the dot product of $\mathbf{Q}$ and $\mathbf{K}$ before applying the softmax function. + +Adaptive global interaction module (AGIM). Inspired by existing reflection separation approaches [23, 58] which attempt to distinguish transmission and reflection layers in the feature space by feature channel rearrangement (i.e., allocating distinct channels from mixture image features to the two layers), we propose to integrate language interactions at the feature channel level. To achieve this, as illustrated in Figure 3(b), we propose the adaptive global interaction module (AGIM), which employs the channel-wise cross-attention mechanism to conduct interactions between global language features and image features for channel rearrangement. The query, key, and value projections using linear projection layers with layer normalization [1] are conducted on the global language feature $\mathbf{F}_{\mathbf{L}_i}$ , the global visual feature $\mathbf{F}_{\mathbf{V}}^{\mathrm{gl}}$ , and the visual feature $\mathbf{F}_{\mathbf{V}}^{\mathrm{in}}$ to generate $\mathbf{Q} \in \mathbb{R}^{C}$ , $\mathbf{K} \in \mathbb{R}^{C}$ , and $\mathbf{V} \in \mathbb{R}^{C \times h w}$ , respectively. Before the interaction, a language gate is designed to prevent impacts of unavailable language descriptions caused by the recognizable layer ambiguity problem that users may only input one description for the layer recognizable in the mixture image. Specifically, if the description $\mathbf{L}_i$ corresponding to the current global language feature $\mathbf{F}_{\mathbf{L}_i}$ is available (not set to null), the gate will feed $\mathbf{F}_{\mathbf{L}_i}$ into the following interaction + +process, otherwise the gate will feed $\mathbf{F}_{\mathbf{V}}^{\mathrm{glo}}$ , which turns the interaction process to be a channel-wise self-attention. + +After filtering by the language gate, an interacted feature can be obtained through the channel-wise cross-attention: $\mathbf{F}^{\mathrm{itr}} = \operatorname{Softmax}(\mathbf{Q}^\top \mathbf{K} / \eta)\mathbf{V}$ , where $\eta$ is a learnable scaling factor to control the magnitude of the attention map. Besides, to adaptively adjust the influence of language guidance based on the correspondence between image and language features, we define another scaling factor $s$ valued as the cosine distance between $\mathbf{Q}$ and $\mathbf{K}$ to multiply with the interacted feature $\mathbf{F}^{\mathrm{itr}}$ . Finally, the output of AGIM is obtained by a residual structure: $\mathbf{F}_{\mathbf{V}}^{\mathrm{out}} = \mathbf{F}_{\mathbf{V}}^{\mathrm{in}} + s\mathbf{F}^{\mathrm{itr}}$ , which integrates the contextual information from the language description and adjusts features for layer separation. + +# 3.3. Layer recovery + +After global language-image interaction and progressive refinement by the gated language interaction stage, we obtain layer features $\mathbf{F}_{\mathbf{I}_i}$ which integrates holistic contextual information from the corresponding language description $\mathbf{L}_i$ (if available). The layer recovery stage is dedicated to reconstructing each image layer $\mathbf{I}_i \in \mathbb{R}^{H \times W \times 3}$ from its corresponding layer features $\mathbf{F}_{\mathbf{I}_i}$ , which is achieved by using individual image decoders. In the image decoder, the layer feature is firstly upsampled by a transposed convolutional layer with a GELU activation [17], then refined by a residual block, and finally projected from the feature domain back into the image domain by a $1 \times 1$ convolutional layer with a sigmoid activation, thus accomplishing a precise layer recovery. + +# 3.4. Loss functions + +In this section, we introduce the contrastive correspondence loss $\mathcal{L}_{\mathrm{ctr}}$ and the layer correspondence loss $\mathcal{L}_{\mathrm{lcr}}$ for constraining the proposed method to establish the cross-modality correspondence between language descriptions $\mathbf{L}_i$ and corresponding image layers under the content disturbance from counter layers in mixture images. We also briefly describe the image layer loss $\mathcal{L}_{\mathrm{img}}$ which consists of image- and feature-level loss functions for layer recovery. We denote the estimated image layers as $\tilde{\mathbf{I}}_i$ and ground truths as $\mathbf{I}_i$ . Details of the loss functions are as follows. + +Contrastive correspondence loss. CLIP [42] conducts contrastive language-image pre-training within batches which successes in establishing cross-modality correspondence between language descriptions and clean images, while our language-guided image reflection separation task requires finding the correspondence between the given language description and a certain layer in the mixture image with the interference of extraneous contents. To tackle the above issue, we propose a contrastive correspondence loss that conducts contrastive learning between image layers to establish correct cross-modality correspondence. Specifi- + +cally, given a language description $\mathbf{L}_i$ , the goal of the contrastive correspondence loss is to force the network to learn the relation that the corresponding image layer $\mathbf{I}_i$ is more relevant to $\mathbf{L}_i$ than the counter layer $\mathbf{I}_j$ ( $j \neq i$ ), thus constraining the estimated image layer $\tilde{\mathbf{I}}_i$ to conform to the associated language description $\mathbf{L}_i$ . + +To measure the relevance between a language description $\mathbf{L}$ and an image layer $\mathbf{I}$ , we define a feature-level similarity function $\mathcal{D}(\cdot)$ as: + +$$ +\mathcal {D} (\mathbf {L}, \mathbf {I}) = \sigma \left(\Psi \left(\mathbf {F} _ {\mathbf {L}} ^ {\mathrm {g l o}}, \mathbf {F} _ {\mathbf {I}} ^ {\mathrm {g l o}}\right)\right), \tag {1} +$$ + +where $\sigma$ represents the sigmoid function, $\Psi$ represents the cosine distance, $\mathbf{F}_{\mathbf{L}}^{\mathrm{glo}}$ is the global language feature obtained by the language encoder (in Sec. 3.1), and $\mathbf{F}_{\mathbf{I}}^{\mathrm{glo}}$ is the global image feature produced by the AGAM (in Sec. 3.2). For an available language description $\mathbf{L}_i$ , we calculate its contrastive correspondence loss as: + +$$ +\mathcal {L} _ {\mathrm {c t r}} \left(\mathbf {L} _ {i}, \tilde {\mathbf {I}} _ {i}, \mathbf {I} _ {j}\right) = - \log \left(\frac {\mathcal {D} \left(\mathbf {L} _ {i} , \tilde {\mathbf {I}} _ {i}\right)}{\mathcal {D} \left(\mathbf {L} _ {i} , \tilde {\mathbf {I}} _ {i}\right) + \mathcal {D} \left(\mathbf {L} _ {i} , \mathbf {I} _ {j}\right)}\right), \tag {2} +$$ + +and we sum up the contrastive correspondence loss of each available language description as the final one. + +Layer correspondence loss. We further define a layer correspondence loss to encourage the relevance between the language description $\mathbf{L}_i$ and the estimated image layer $\tilde{\mathbf{I}}_i$ to approach the relevance between $\mathbf{L}_i$ and the ground truth image layer $\mathbf{I}_i$ : + +$$ +\mathcal {L} _ {\mathrm {l c r}} \left(\mathbf {L} _ {i}, \tilde {\mathbf {I}} _ {i}, \mathbf {I} _ {i}\right) = \left\| \mathcal {D} \left(\mathbf {L} _ {i}, \tilde {\mathbf {I}} _ {i}\right) - \mathcal {D} \left(\mathbf {L} _ {i}, \mathbf {I} _ {i}\right) \right\| _ {1}, \tag {3} +$$ + +where $\mathcal{D}(\cdot)$ is the same feature-level similarity function as in the contrastive correspondence loss. We also sum up the layer correspondence loss of each available language description for final supervision. + +Image layer loss. To achieve high-fidelity recovery of image layers (i.e., transmission and reflection layers), the proposed method is also optimized with loss functions following previous reflection separation methods [23, 52, 67]. Specifically, we utilize loss functions that conduct constraints on the visual quality of estimated images (i.e., the pixel $\mathcal{L}_{\mathrm{pix}}$ , structural similarity $\mathcal{L}_{\mathrm{ssim}}$ , and perceptual loss $\mathcal{L}_{\mathrm{per}}$ ) or exploit the inherent relationship between two layers (i.e., the exclusion $\mathcal{L}_{\mathrm{exc}}$ and reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ ), and we denote the combination of the above image- or feature-level loss functions as the image layer loss $\mathcal{L}_{\mathrm{img}}^1$ . + +Overall, the total loss function is then formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \gamma_ {1} \mathcal {L} _ {\mathrm {c t r}} + \gamma_ {2} \mathcal {L} _ {\mathrm {l c r}} + \mathcal {L} _ {\mathrm {i m g}}, \tag {4} +$$ + +where coefficients are set as $\gamma_{1} = \gamma_{2} = 0.5$ + +# 3.5. Training strategy + +Due to the recognizable layer ambiguity that sometimes only one layer in the mixture image is recognizable, we propose a randomized training strategy to synergize with the gated language interaction mechanism (in Sec. 3.2). Since our training data simulates the recognizable layer ambiguity that some image layers do not have corresponding language descriptions (introduced in the next section), we only feed the available language description corresponding to the other image layer to guide the separation. For data with descriptions of both layers available, which indicates that both layers are recognizable, we also randomly drop one language description and feed the remaining one into the network to improve the generalization capacity for the proposed method. In practice, we set the ratio of dropping language descriptions to $30\%$ . + +We implement the proposed method with PyTorch [41] with a batch size of 16 on two Nvidia GeForce RTX 3090 GPUs. The model is trained for 40 epochs with Adam optimizer [25] to update learnable parameters. Weights are initialized as in [15]. The learning rate is set to $10^{-4}$ initially and decreases to $10^{-5}$ at epoch 30. + +# 4. Data preparation + +Though existing works have constructed several datasets for single-image reflection separation [31, 49, 67], they are unavailable for the proposed language-guided reflection separation framework due to the lack of corresponding language descriptions. Therefore, we build a dataset containing both synthetic and real data to overcome the data deficiency and facilitate network training and evaluation. Each group of data is composed of a mixture image, a transmission layer, a reflection layer, and two language descriptions. Details of synthetic and real data are as follows. + +# 4.1. Synthetic data + +The synthetic dataset is generated for network training to satisfy the data-driven need of the proposed method. Due to the demand for paired image-language data, we utilize two prevailing image captioning datasets (i.e., Flickr30k [66] and COCO Captions [7]) for data generation, which contain 31,000 and 330,000 images respectively, and each image has 5 independent human-generated language descriptions. We randomly select images from the above two datasets as transmission $\mathbf{T}_{\mathrm{S}}$ and reflection scene images $\mathbf{R}_{\mathrm{S}}$ and conduct an image synthesis process with linear blending [23]: + +$$ +\hat {\mathbf {M}} = \hat {\mathbf {T}} + \hat {\mathbf {R}} = \alpha \hat {\mathbf {T}} _ {\mathrm {S}} + \beta \hat {\mathbf {R}} _ {\mathrm {S}}, \tag {5} +$$ + +where $\hat{\mathbf{T}}_{\mathrm{S}} = g_{\mathrm{inv}}(\mathbf{T}_{\mathrm{S}})$ and $\hat{\mathbf{R}}_{\mathrm{S}} = g_{\mathrm{inv}}(\mathbf{R}_{\mathrm{S}})$ , $g_{\mathrm{inv}}$ represents the inverse gamma correction, and $\alpha \in [0.8,1]$ and $\beta \in [0.4,1]$ are the blending attenuation coefficients as in [23]. + +Considering the recognizable layer ambiguity that sometimes only one layer is recognizable in a mixture image, we assign a language description only when the corresponding layer is obvious enough in the synthesized mixture image, i.e., $\frac{\mathrm{mean}(\hat{\mathbf{V}}_l)}{\mathrm{mean}(\hat{\mathbf{V}}_{\mathbf{M}})}\geqslant \mu$ , where $\hat{\mathbf{V}}_l$ represents the brightness image of $\hat{\mathbf{T}}$ or $\hat{\mathbf{R}}$ in the HSV color space and $\hat{\mathbf{V}}_{\mathbf{M}}$ represents the brightness image of $\hat{\mathbf{M}}$ , and we set $\mu = 0.3$ in the data generation process. Finally, gamma correction is applied to image triplets $\{\hat{\mathbf{T}},\hat{\mathbf{R}},\hat{\mathbf{M}}\}$ to obtain $\{\mathbf{T},\mathbf{R},\mathbf{M}\}$ , and we generate 50000 triplets of data in total for network training. + +# 4.2. Real data + +Existing real datasets collected for the single-image reflection separation task usually contain mixture images with ground truth of transmission layers (e.g., Zhang et al. [67] and Nature [31]), and $\mathrm{SIR}^2$ [49, 54] further captures ground truths of reflection layers. For these off-the-shelf real datasets, we augment them by manually adding language descriptions for each group of data to satisfy the input setting of the proposed language-guided reflection separation task. Specifically, following the annotation principle of COCO Captions [7], we first describe the content of transmission layers in mixture images with entities, attributes (e.g., colors or materials), and relative positions between different entities. If the content of reflection layers is recognizable in mixture images, we also give language descriptions for reflection layers in the same way. To further evaluate the generalization capacity of the proposed method, we collect a real dataset (denoted as REFOL dataset) containing 100 mixture images from the Internet that are captured in different scenes and with different cameras. We annotate these mixture images with language descriptions in the same manner as mentioned above. Following the training strategy of previous methods [10, 31], we utilize 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [67] for training, and the rest of real data are used for quantitative and qualitative evaluation. + +# 5. Experiments + +# 5.1. Comparison with state-of-the-art methods + +To evaluate the performance of the proposed method, we conduct quantitative and qualitative experiments on existing real datasets [31, 49, 67] (with our manually annotated language descriptions) and our newly collected dataset REFOL. We compare with state-of-the-art single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], IBCLN [31], CoRRN [52], and Zhang et al. [67]. For fair comparisons, we finetune the above methods on our training data if their training codes are provided. We report better results between the original pre-trained model and the finetuned version. + +Table 1. Comparison of quantitative results in terms of PSNR [24] and SSIM [55] on real datasets for evaluating the recovery of transmission layers. $\uparrow (\downarrow)$ indicates larger (smaller) values are better. Bold numbers indicate the best-performing results. + +
Dataset (size)MetricsMethods
Zhang et al. [67]CoRRN [52]IBCLN [31]Dong et al. [10]YTMT [22]DSRNet [23]Ours
Postcard (199)PSNR↑20.8522.0423.4123.7222.8224.8825.02
SSIM↑0.8720.8700.8720.9030.8850.9100.915
Object (200)PSNR↑23.8425.1324.5224.3624.6826.4426.51
SSIM↑0.8720.9120.8910.8980.8920.9210.927
Wild (101)PSNR↑24.9725.1724.7825.7525.7025.8626.23
SSIM↑0.8750.8890.8840.9030.8970.9080.925
Real20 (20)PSNR↑22.3421.4321.4723.3423.2323.8824.05
SSIM↑0.7950.8010.7620.8120.8020.8160.824
Nature (20)PSNR↑20.6220.7523.7223.4521.5322.2623.87
SSIM↑0.7530.7830.8060.8080.7780.8010.812
Average (540)PSNR↑22.7723.7024.0224.3124.0125.5125.72
SSIM↑0.8650.8830.8750.8940.8830.9060.914
+ +![](images/31fd5974efffb740d30321de4e2dc25e963bae97e2aaf4c24cacd2e7740935d7.jpg) + +![](images/17f7b10bc75071d9448dfd2b7eca0c844c71f26f1e39d7117335d3c522fd51b1.jpg) + +![](images/059f90c0b7d490404328ff5ae74a15d53a34e038f14c2df3ee82211ceb17ff2e.jpg) + +![](images/0a4ab7a730efd3cbfc4092ee474c0b8b2d33ed415a98d21b0ed345b9c65a3e67.jpg) + +![](images/5485608babf34c448d34d210adc4a6a3ead931fa5a414fc6e64b3a070f2e6dcb.jpg) + +![](images/9d60594cfba8391e7c2d2b5a80d2c6733a02a6be7da9bcecd91bde3255b50fcc.jpg) + +![](images/d67069550ba292446096683f0bc1ad631394316de46bb16afec1baceaacdd538.jpg) + +![](images/d983f8002a4d9565fb15bf90455a7786d881941321a17cfbfc46af44816c0228.jpg) + +![](images/a5e58dd925ae40e1a3efcb9d41aef3aa2b55f2a0198b18c3f64bf564399e3836.jpg) + +![](images/740aa01ce9ffaec58414e5334f1279279cc01321a5e03f938f0465f648d35a74.jpg) +Figure 4. Qualitative comparison of estimated transmission and reflection layers on real data, compared with the state-of-the-art methods including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Please zoom in for details. + +![](images/64d42f4f91c4c08337d7600a88145badafad6673c031b19c7b91a535505974d2.jpg) + +![](images/7218db79820066185abdbc9e2cbb6e86699ac06643f166da2a840f6817eb7d63.jpg) + +Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection separation, + +i.e., Nature [31], Real20 [67], and three subsets of $\mathrm{SIR}^2$ [49] dataset. Following the setting of existing reflection separa + +Table 2. Quantitative results of ablation studies. + +
Metricsw/o languagew/o AGIMLimg onlyOurs
PSNR↑24.3124.6924.5225.72
SSIM↑0.8850.9010.8930.914
+ +tion methods [10, 36], we utilize PSNR [24] and SSIM [55] as error metrics for evaluating the recovery of transmission layers2. As quantitative results shown in Table 1, the proposed method achieves the best performance of both PSNR and SSIM, which validates its generalization capacity and the effectiveness of language descriptions. + +Qualitative comparison. To evaluate the visual quality of reflection separation results, we compare the proposed method with four single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Qualitative results on recovering both transmission and reflection layers are shown in Figure 4. As can be observed in Figure 4, IBCLN [31] can only remove parts of reflections, while Dong et al. [10] have trouble in dealing with complex semantic images like the reflections of a man in a room (the second example). YTMT [22] separates layers incorrectly and thus brings the content of transmission layers into reflections. DSRNet [23] fails to recover transmission and reflection layers for these challenging cases. Contributing to the language guidance, the proposed method generates better visual results and recovers both transmission and reflection layers neatly. + +# 5.2. Ablation study + +In this section, we conduct several ablation studies with quantitative results shown in Table 2 to investigate the influence of the additional input of language descriptions (denoted as 'w/o language'), the network design of AGIM (denoted as 'w/o AGIM'), and the loss functions for cross-modality correspondence (denoted as $\mathcal{L}_{\mathrm{img}}$ only). The performance of the variant 'w/o language' suffers from an obvious degradation as we remove language descriptions, which shows the effectiveness of the additional contextual information for layer separation. The variant 'w/o AGIM' replaces AGIMs with simple feature fusion blocks which directly concatenate language and image features and feed them into self-attention blocks, and the decline in performance validates the necessity of our gated interaction mechanism. The variant $\mathcal{L}_{\mathrm{img}}$ only is trained with $\mathcal{L}_{\mathrm{img}}$ , which obtains results slightly better than the variant 'w/o language', indicating the significance of establishing the cross-modality correspondence. + +To further verify the effectiveness of the language interaction, we conduct an ablation study by gradually increasing the number of input descriptions. As shown in yellow + +![](images/72e62436b20e32a8267b24dcfeec431c63c50b730e8166a45c80428798e75f44.jpg) +Figure 5. Results with different numbers of input language descriptions. + +boxes of Figure 5, by utilizing more input language descriptions, the separation of transmission (clothes) and reflection layers (a man with a camera) becomes more thorough, which also demonstrates the efficacy and robustness of the proposed method. + +# 6. Conclusion + +This paper introduces natural language to provide contextual information about image layers for relieving the ill-posed reflection separation problem. We develop an end-to-end framework with adaptive global interaction modules and language-image loss functions to effectively manage the modality inconsistency, and we adopt a language gate mechanism with randomized training strategies to handle the recognizable layer ambiguity. To address data deficiency, a specially built dataset with language annotations significantly aids in training and evaluating the proposed language-guided image reflection separation framework. Quantitative and qualitative experiments on real data demonstrate the effectiveness of introducing language descriptions for reflection separation. + +Limitations. The proposed method may not distinguish transmission and reflection layers accurately when their contents are similar. For such ambiguous cases, a more flexible language-guided mechanism is needed. This might be solved by exploring a better interaction approach, which is left as our future work. + +# Acknowledgement + +This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4 +[2] Yakun Chang, Cheolkon Jung, and Jun Sun. Joint reflection removal and depth estimation from a single image. IEEE TCYB, 2020. 1 +[3] Yakun Chang, Cheolkon Jung, Jun Sun, and Fengqiao Wang. Siamese dense network for reflection removal with flash and no-flash image pairs. IJCV, 2020. 2 +[4] Zheng Chang, Shuchen Weng, Yu Li, Si Li, and Boxin Shi. L-CoDer: Language-based colorization with color-object decoupling transformer. In Proc. of ECCV, 2022. 1 +[5] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CoIns: Language-based colorization with instance awareness. In Proc. of CVPR, 2023. +[6] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Proc. of NeurIPS, 2023. 1 +[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 6 +[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. NeRDi: Single-view NeRF synthesis with language-guided diffusion as general image priors. In Proc. of CVPR, 2023. 2 +[9] Yaron Diamant and Yoav Y Schechner. Overcoming visual reverberations. In Proc. of CVPR, 2008. 2 +[10] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson WH Lau. Location-aware single image reflection removal. In Proc. of ICCV, 2021. 1, 2, 6, 7, 8 +[11] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf. A generic deep architecture for single image reflection removal and image smoothing. In Proc. of ICCV, 2017. 2 +[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. of NeurIPS, 2014. 2 +[13] Byeong-Ju Han and Jae-Young Sim. Zero-shot learning for reflection removal of single 360-degree image. In Proc. of ECCV, 2022. 2 +[14] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In Proc. of CVPR, 2015. 3 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance onImagenet classification. In Proc. of ICCV, 2015. 6 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of CVPR, 2016. 3 +[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 4, 5 + +[18] Yuchen Hong, Youwei Lyu, Si Li, and Boxin Shi. Near-infrared image guided reflection removal. In Proc. of ICME, 2020. 2 +[19] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. Panoramic image reflection removal. In Proc. of CVPR, 2021. 1, 2 +[20] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with NIR and RGB image feature fusion. IEEE TMM, 2022. 2 +[21] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. PAR $^2$ Net: End-to-end panoramic image reflection removal. IEEE TPAMI, 2023. 2 +[22] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation.* In *Proc. of NeurIPS*, 2021. 1, 2, 6, 7, 8 +[23] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proc. of ICCV, 2023. 2, 4, 5, 6, 7, 8 +[24] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics letters*, 2008, 7, 8 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[26] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. A physically-based approach to reflection separation. In Proc. of CVPR, 2012. 2 +[27] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proc. of CVPR, 2021. 1, 2 +[28] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proc. of CVPR, 2020. 1, 2 +[29] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE TPAMI, 2023. 2 +[30] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE TPAMI, 2007. 1, 2 +[31] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proc. of CVPR, 2020. 1, 2, 6, 7, 8 +[32] Yu Li and Michael S Brown. Exploiting reflection change for automatic reflection removal. In Proc. of ICCV, 2013. 1, 2 +[33] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proc. of CVPR, 2014. 1, 2 +[34] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In Proc. of CVPR, 2020. 1, 2 +[35] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. IEEE TPAMI, 2021. 2 +[36] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolar + +ized and polarized images. In Proc. of NeurIPS, 2019. 1, 2, 8 +[37] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE TPAMI, 2022. 2 +[38] Daiqian Ma, Renjie Wan, Boxin Shi, Alex C Kot, and Ling-Yu Duan. Learning to jointly generate and separate reflections. In Proc. of ICCV, 2019. 2 +[39] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. IJCV, 1997. 2 +[40] Jonghyuk Park, Hyeona Kim, Eunpil Park, and Jae-Young Sim. Fully-automatic reflection removal for 360-degree images. In Proc. of WACV, 2024. 2 +[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Proc. of NeurIPS, 2019. 6 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. of ICML, 2021. 4, 5 +[43] Yoav Y Schechner, Nahum Kiryati, and Ronen Basri. Separation of transparent layers using focus. IJCV, 2000. 2 +[44] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proc. of CVPR, 2015. 1, 2 +[45] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In Proc. of CVPR, 2015. 1, 2 +[46] Jimeng Sun, Shuchen Weng, Zheng Chang, Si Li, and Boxin Shi. UniCoRN: A unified conditional image inpainting network. In Proc. of CVPR, 2022. 1 +[47] Jiajun Tang, Haofeng Zhong, Shuchen Weng, and Boxin Shi. LuminAIRe: Illumination-aware conditional image inpainting for lighting-realistic generation. In Proc. of NeurIPS, 2023. 1 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, 2017. 4 +[49] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In Proc. of ICCV, 2017. 2, 6, 7 +[50] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, Wen Gao, and Alex C. Kot. Region-aware reflection removal with unified content and gradient priors. IEEE TIP, 2018. 2 +[51] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CRRN: Multi-scale guided concurrent reflection removal network. In Proc. of CVPR, 2018. 2 +[52] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CoRRN: Cooperative reflection removal network. IEEE TPAMI, 2019. 2, 5, 6, 7 +[53] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. IJCV, 2021. 1 + +[54] Renjie Wan, Boxin Shi, Haoliang Li, Yuchen Hong, Ling-Yu Duan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In IEEE TPAMI, 2022. 2, 6 +[55] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In Proc. of ACSSC, 2003. 7, 8 +[56] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proc. of CVPR, 2022. 4 +[57] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: CLIP-driven referring image segmentation. In Proc. of CVPR, 2022. 1 +[58] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proc. of CVPR, 2019. 2, 3, 4 +[59] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In Proc. of CVPR, 2019. 2 +[60] Shuchen Weng and Boxin Shi. Conditional image repainting. IEEE TPAMI, 2023. 1 +[61] Shuchen Weng, Wenbo Li, Dawei Li, Hongxia Jin, and Boxin Shi. MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In Proc. of CVPR, 2020. 1 +[62] Shuchen Weng, Hao Wu, Zheng Chang, Jiajun Tang, Si Li, and Boxin Shi. L-CoDe: Language-based colorization using color-object decoupled conditions. In Proc. of AAAI, 2022. 1 +[63] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In Proc. of ECCV, 2018. 2 +[64] Yang Yang, Wenye Ma, Yin Zheng, Jian-Feng Cai, and Weiyu Xu. Fast single image reflection suppression via convex optimization. In Proc. of CVPR, 2019. 2 +[65] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-aware vision transformer for referring image segmentation. In Proc. of CVPR, 2022. 1, 2 +[66] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2, 6 +[67] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proc. of CVPR, 2018. 2, 3, 5, 6, 7 +[68] Ya-Nan Zhang, Linlin Shen, and Qiufu Li. Content and gradient model-driven deep network for single image reflection removal. In Proc. of ACM MM, 2022. 2 +[69] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In Proc. of CVPR, 2020. 2 +[70] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In Proc. of CVPR, 2021. 2 \ No newline at end of file diff --git a/2024/Language-guided Image Reflection Separation/images.zip b/2024/Language-guided Image Reflection Separation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cc41924c32ebc3cc21e281e5c30f7cf6d56a0a96 --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c871a693c8c690fff5223254a7cb238d665fd3020155dc04e64b3717e1a570a +size 562661 diff --git a/2024/Language-guided Image Reflection Separation/layout.json b/2024/Language-guided Image Reflection Separation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a4d8611795c548217c1fe6768f210c577fc8486a --- /dev/null +++ b/2024/Language-guided Image Reflection Separation/layout.json @@ -0,0 +1,10211 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 104, + 441, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 104, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 153, + 104, + 441, + 121 + ], + "type": "text", + "content": "Language-guided Image Reflection Separation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "spans": [ + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "content": "Haofeng Zhong" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "inline_equation", + "content": "^{1,2,3\\#}" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "content": " Yuchen Hong" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "inline_equation", + "content": "^{1,2\\#}" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "content": " Shuchen Weng" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "content": " Jinxiu Liang" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 70, + 132, + 522, + 148 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 148, + 506, + 172 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 88, + 148, + 506, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 148, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 88, + 148, + 506, + 161 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 148, + 506, + 161 + ], + "type": "text", + "content": " National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 161, + 498, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 161, + 498, + 172 + ], + "spans": [ + { + "bbox": [ + 88, + 161, + 498, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 161, + 498, + 172 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 72, + 172, + 518, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 172, + 518, + 194 + ], + "spans": [ + { + "bbox": [ + 72, + 172, + 518, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 72, + 172, + 518, + 194 + ], + "type": "text", + "content": " AI Innovation Center, School of Computer Science, Peking University {hfzhong, shuchenweng, cssherryliang, shiboxin}@pku.edu.cn, yuchenhong.cn@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 212, + 192, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 212, + 192, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 212, + 192, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 238, + 290, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 238, + 290, + 394 + ], + "spans": [ + { + "bbox": [ + 45, + 238, + 290, + 394 + ], + "type": "text", + "content": "This paper studies the problem of language-guided reflection separation, which aims at addressing the ill-posed reflection separation problem by introducing language descriptions to provide layer content. We propose a unified framework to solve this problem, which leverages the crossattention mechanism with contrastive learning strategies to construct the correspondence between language descriptions and image layers. A gated network design and a randomized training strategy are employed to tackle the recognizable layer ambiguity. The effectiveness of the proposed method is validated by the significant performance advantage over existing reflection separation methods on both quantitative and qualitative comparisons." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 128, + 430 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": "When photographing through transparent materials like glass windows or showcases, the presence of reflections can significantly degrade the image quality of captured images and disrupt downstream computer vision tasks like face recognition [53] or depth estimation [2]. As an attractive topic in computational photography, reflection separation aims at decomposing the contaminated mixture image (denoted as " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": ") into two components that correspond to scenes located at different sides of the glass, i.e., the reflection layer (denoted as " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": ") and the transmission layer (denoted as " + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 46, + 438, + 287, + 594 + ], + "type": "text", + "content": "). Since reflection separation is a severely ill-posed problem, it is imperative to exploit effective priors or auxiliary information for distinguishing the two components." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 595, + 287, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 595, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 46, + 595, + 287, + 691 + ], + "type": "text", + "content": "The primary challenge of solving the reflection separation problem lies in the exploration of distinct clues for distinguishing transmission and reflection layers. Multi-image methods handle this problem by introducing additional constraints. Some of them acquire a series of mixture images in different viewpoints [32, 34, 45] to harness the distinct motions of the two layers, while others adopt specialized capturing setups to obtain complementary scene informa" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 212, + 545, + 408 + ], + "blocks": [ + { + "bbox": [ + 308, + 212, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 308, + 212, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 212, + 545, + 408 + ], + "type": "image", + "image_path": "f45362c64c63d2eacf3472b97a8ae2ef172e097b9bbc0b323aafe6087713a14b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 411, + 547, + 477 + ], + "lines": [ + { + "bbox": [ + 305, + 411, + 547, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 547, + 477 + ], + "type": "text", + "content": "Figure 1. The recognizable layer ambiguity problem causes uncertain quantities of input language descriptions for language-guided image reflection separation. Given language descriptions of either (a) one layer or (b) both two layers, the proposed method achieves robust reflection separation compared with an existing reflection separation method [10]." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "tion [19, 27, 28, 36]. However, the specialized data capture requirements limit the application scope of these methods, especially for images downloaded from the Internet. Single-image methods attempt to tackle reflection separation by utilizing handcrafted priors derived from natural image statistics [30, 33, 44] or leveraging the modeling capacity of neural networks to learn content priors about reflections from a large scale of training data [10, 22, 31]. However, they are prone to fail due to the lack of auxiliary content information about transmission or reflection scenes for solving such a highly ill-posed problem. Recently, language descriptions have shown their effectiveness in providing content information for various vision tasks such as image editing [46, 47, 60, 61], semantic segmentation [57, 65], and image colorization [4-6, 62], which inspires us to think about: Can we leverage the auxiliary content information brought by language descriptions to facilitate the reflection separation problem?" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 700, + 207, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 700, + 207, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 700, + 207, + 713 + ], + "type": "text", + "content": "\\# Equal contributions. *Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24913" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 371 + ], + "type": "text", + "content": "Since language can effectively convey humans' prior knowledge about the real world [8] and provide auxiliary information of image semantics [65], introducing language descriptions to guide the separation of reflection and transmission layers from mixture images merits exploration. However, leveraging language descriptions for reflection separation is non-trivial in three aspects: 1) Language-image modality inconsistency. Language and images belong to different modalities, thus it is challenging to establish a cross-modality correspondence between the scene content information provided in language descriptions and the complex blended content present in mixture images. 2) Recognizable layer ambiguity. Since the image content and brightness of reflection and transmission layers are different, the recognizable extents of them in mixture images are also uncertain. Specifically, as shown in Figure 1, it is possible that only one layer's content is recognizable (clothes in Figure 1(a)), or both two layers exhibit recognizable content (the trashbin and cars in Figure 1(b)), which leads to the difficulty of using uncertain quantities of language descriptions for separating mixture images in practice. 3) Language annotation deficiency. All existing datasets for the reflection separation task only contain image data but no correlated language description is provided, raising the challenge for network training and evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 375, + 289, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 289, + 638 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 289, + 638 + ], + "type": "text", + "content": "In this paper, we introduce the concept of language-guided image reflection separation for the first time, which leverages flexible natural language to specify the content of one or two layers within a mixture image, relieving the ill-posedness of the reflection separation problem and maintaining a wide applicability for both live captured or online downloaded mixture images. We propose an end-to-end framework that employs adaptive global interaction modules to explore holistic language-image content coherence and utilizes specifically designed loss functions to constrain the correspondence between language descriptions and recovered image layers. A language gate mechanism and a randomized training strategy are designed to deal with the recognizable layer ambiguity problem. To address the language annotation deficiency, we synthesize the training dataset from paired image-language datasets [7, 66] and expand prevailing real reflection separation datasets [31, 49, 67] by manually adding language descriptions. Besides, we further construct a new dataset for visual quality evaluation by collecting mixture images from the Internet and captioning them with language descriptions for recognizable layers. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 642, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 666 + ], + "type": "text", + "content": "- We present the first work that introduces language descriptions to guide the reflection separation task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 666, + 286, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 286, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 286, + 701 + ], + "type": "text", + "content": "- We propose adaptive global interaction modules and language-image loss functions to tackle modality inconsistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 713 + ], + "type": "text", + "content": "- We design a language gate mechanism and a randomized" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 541, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 85 + ], + "type": "text", + "content": "training strategy to handle recognizable layer ambiguity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 85, + 544, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 85, + 544, + 108 + ], + "spans": [ + { + "bbox": [ + 307, + 85, + 544, + 108 + ], + "type": "text", + "content": "- We build a dataset with language descriptions to facilitate language-guided image reflection separation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 119, + 390, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 119, + 390, + 131 + ], + "spans": [ + { + "bbox": [ + 306, + 119, + 390, + 131 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 140, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 545, + 270 + ], + "type": "text", + "content": "Single-image reflection separation methods try to distinguish reflection and transmission layers using a single mixture image, which mainly relies on the assumption that the two layers have different distributions, i.e., reflection layers are more likely to be blurry and appear with lower intensity compared with transmission layers. Conventional methods adopt handcrafted priors derived from natural image statistics in their optimization process, e.g., the gradient sparsity [30], relative smoothness [33], ghosting cues [44], content prior [50], and penalty on the gradient of restored transmission layers [64]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 271, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 546, + 582 + ], + "type": "text", + "content": "Due to the tremendous progress in the field of deep learning, a series of single-image reflection separation methods concentrate on the improvement of learning strategies or network design, e.g., predicting edges and images with a two-stage [11] or concurrent framework [51, 52], training with the perceptual loss [67], employing generative adversarial network [12] based models [38, 58], adopting iterative refinement strategies [10, 31, 63, 68], and leveraging the complementary two-stream architecture [22]. Meanwhile, research on data synthesis and image models is also ongoing to satisfy the data-driven needs of learning-based methods. Ma et al. [38] utilize generative adversarial networks for data generation while Wen et al. [59] synthesize mixture images with learned non-linear blending masks. Hu et al. [23] introduce a learnable residue term in the mixture image formation model to mitigate the non-linearity caused by the complex camera pipeline. Zheng et al. consider physical factors such as reflective amplitude coefficient maps [69] and the absorption effect [70] in the image formation process of mixture images. To facilitate network training and evaluation, researchers [31, 49, 67] also collect real data by using portable glass. Moreover, as a special form of images, panoramic images are introduced to relieve the content ambiguity in mixture images [13, 19, 21, 40]. We refer readers to [54] for a comprehensive and up-to-date survey on single-image reflection separation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Multi-image reflection separation methods usually leverage the auxiliary information introduced by additional images and achieve more robust performance than single-image methods. Polarization-based methods [9, 26, 28, 36, 37, 39, 43] distinguish reflection and transmission layers by using images captured with different angles of polarizers or special polarization cameras. Flash-based methods [3, 18, 20, 27, 29] adopt active light sources to illuminate transmission scenes for obtaining reflection-free guidance. Motion-based methods [32, 34, 35, 45] utilize multiple images captured from different viewpoints to harness" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24914" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 265 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 265 + ], + "type": "image", + "image_path": "f8d20ba08f8af659fe038800ae98c653d32eb32f2488da12dd6d017a5ad91a67.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 546, + 322 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 546, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 546, + 322 + ], + "type": "text", + "content": "Figure 2. The pipeline of the proposed language-guided image reflection separation framework, which extracts features from mixture images and available language descriptions (the description " + }, + { + "bbox": [ + 46, + 266, + 546, + 322 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 266, + 546, + 322 + ], + "type": "text", + "content": " with dashed lines is possible to be set to null due to the recognizable layer ambiguity) by image and language encoders (in Sec. 3.1), aggregates global visual information by adaptive global aggregation modules (AGAM) and conducts progressive interactions to exploit distinctive image features with gated language guidance by adaptive global interaction modules (AGIM) (in Sec. 3.2), and recovers image layers by image decoders (in Sec. 3.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 390 + ], + "type": "text", + "content": "distinct motions of reflection and transmission layers. However, special data capture requirements significantly limit the application scope of these methods, especially for mobile devices or images downloaded from the Internet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 403, + 151, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 403, + 151, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 403, + 151, + 415 + ], + "type": "text", + "content": "3. Proposed method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 422, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 662 + ], + "type": "text", + "content": "The pipeline of the proposed framework is illustrated in Figure 2. In this section, we start by introducing the feature extraction stage (in Sec. 3.1) that obtains multi-scale image features and global language features, the gated language interaction stage (in Sec. 3.2) that conducts progressive image-language global interactions to exploit distinctive image features and prevents unavailable language interactions with switchable gates, and the layer recovery stage (in Sec. 3.3) that reprojects features into the image space with a light-weight image decoder. Then we explain loss functions (in Sec. 3.4) employed for network optimization, especially a contrastive correspondence loss and a layer correspondence loss that constrains the network to construct correspondences between the language description and the corresponding layer under the disturbance of the other layer in a superimposed mixture image. Finally, we present our training strategy (in Sec. 3.5) which enables the network to be applicable for varying quantities of input language descriptions and jointly tackles the recognizable layer ambiguity problem with the gated network design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 250, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 250, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 250, + 683 + ], + "type": "text", + "content": "3.1. Image and language feature extraction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The inputs of the proposed method consist of a mixture image " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " with two language descriptions " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{L}_i|i = 1,2\\}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": "which corresponds to the two image layers. We specify that layer " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " corresponds to the description " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": ". However, due to the recognizable layer ambiguity that in certain cases only one layer of the mixture image is recognizable (usually the transmission layer), for such cases, we set " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " to be the available language description (for the recognizable transmission layer) and " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " to null (for the unrecognizable reflection layer) to ensure a unified input setting. Then given the input image and language descriptions, the feature extraction stage aims at obtaining the image feature " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{M}}" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " with the image encoder and the multi-scale feature fusion process and extracting the global language feature " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_i}" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " for each description " + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 342, + 545, + 510 + ], + "type": "text", + "content": " via the language encoder for the subsequent interaction procedure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": "Image encoder. Given a mixture image " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": ", we employ a commonly-used vision backbone ResNet-50 [16] as our image encoder, whose last three layers (i.e., an average pooling layer, a fully connected layer, and a softmax layer) are removed to fit our task. We utilize image features from the first five blocks of the image encoder to form a multi-scale feature pyramid " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_{\\mathbf{M}_i}\\}_{i=1}^5" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{M}_i} \\in \\mathbb{R}^{h_i \\times w_i \\times C_i}" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "h_i = H/2^i" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "w_i = W/2^i" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": " is the height and width of the mixture image, respectively, and " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": " is the dimension of the " + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 514, + 545, + 634 + ], + "type": "text", + "content": "-th extracted feature." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": "Multi-scale feature fusion. Obtaining the extracted feature pyramid, we first transform it into a hypercolumn feature " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{M}}^{\\mathrm{hyp}} \\in \\mathbb{R}^{h \\times w \\times C^{\\mathrm{hyp}}} [14]" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "h = H/2" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "w = W/2" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{hyp}} = \\sum_{i=1}^{5} C_i" + }, + { + "bbox": [ + 304, + 639, + 545, + 713 + ], + "type": "text", + "content": "), which has been proved to be effective in fusing multi-scale contextual information for reflection separation [58, 67]. Considering the computational cost," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24915" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 90, + 284, + 167 + ], + "blocks": [ + { + "bbox": [ + 51, + 75, + 251, + 86 + ], + "lines": [ + { + "bbox": [ + 51, + 75, + 251, + 86 + ], + "spans": [ + { + "bbox": [ + 51, + 75, + 251, + 86 + ], + "type": "text", + "content": "(a) AGAM: Adaptive global aggregation module" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 90, + 284, + 167 + ], + "lines": [ + { + "bbox": [ + 51, + 90, + 284, + 167 + ], + "spans": [ + { + "bbox": [ + 51, + 90, + 284, + 167 + ], + "type": "image", + "image_path": "75de14e1787dca95ffbbbd42d5c9259507ba41a253e09c356f3e49918a8fefe8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 51, + 186, + 284, + 271 + ], + "blocks": [ + { + "bbox": [ + 52, + 171, + 246, + 182 + ], + "lines": [ + { + "bbox": [ + 52, + 171, + 246, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 171, + 246, + 182 + ], + "type": "text", + "content": "(b) AGIM: Adaptive global interaction module" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 186, + 284, + 271 + ], + "lines": [ + { + "bbox": [ + 51, + 186, + 284, + 271 + ], + "spans": [ + { + "bbox": [ + 51, + 186, + 284, + 271 + ], + "type": "image", + "image_path": "9a30e18b3b14960b596b66f902c24e2e5b54bbab10688658b825283817f37c6d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "lines": [ + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "spans": [ + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "text", + "content": " Matrix multiplication " + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "text", + "content": " Reshape " + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 52, + 277, + 284, + 289 + ], + "type": "text", + "content": " Element-wise multiplication" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "lines": [ + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "spans": [ + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "type": "text", + "content": "Language gate " + }, + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "type": "text", + "content": " Concatenation " + }, + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 52, + 292, + 284, + 301 + ], + "type": "text", + "content": " Element-wise addition" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 304, + 287, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 361 + ], + "type": "text", + "content": "Figure 3. The architecture of the (a) adaptive global aggregation module (AGAM) and (b) adaptive global interaction module (AGIM), which aggregates global contextual information of visual features and achieves feature channel rearrangement with gated language guidance, respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "text", + "content": "we condense and refine the hypercolumn feature by a " + }, + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "text", + "content": " convolutional layer with a GELU activation [17] followed by a locally-enhanced feed-forward (LeFF) block [56]. The final fused feature of the mixture image is denoted as " + }, + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{M}} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 46, + 383, + 287, + 456 + ], + "type": "text", + "content": ", which serves as the basis for the subsequent interaction and separation process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": "Language encoder. Motivated by the rapid development of pre-trained large-scale vision-language models, we employ the language encoder from CLIP [42], which adopts a Transformer architecture [48] to extract language features and obtains a global contextual feature in the multi-modal embedding space by using layer normalization and linear projection layers. Given a language description " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i\\in \\mathbb{R}^L" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " we obtain its corresponding global feature " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_i}\\in \\mathbb{R}^C" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " by leveraging the modeling capacity of the language encoder to encode the description, thus extracting the holistic contextual information of the corresponding image layer. Here " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " denotes the length of the language description and " + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 456, + 287, + 612 + ], + "type": "text", + "content": " is the feature dimension as the image feature." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 622, + 197, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 197, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 197, + 635 + ], + "type": "text", + "content": "3.2. Gated language interaction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": "The gated language interaction stage aims at leveraging the contextual information from available language descriptions to guide the separation of the corresponding layer feature " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_i}" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": ", which is composed of " + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 46, + 642, + 287, + 715 + ], + "type": "text", + "content": " cascaded interaction groups to separate image layer features successively. As shown in Figure 2, each group consists of an adaptive global" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "aggregation module (AGAM) to gather global information of input visual features, a language gate to prevent detrimental guidance from unavailable descriptions, an adaptive global interaction module (AGIM) to conduct interactions using global features for exploring holistic image-language content correspondence, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " normalization layers with LeFF blocks [56] for feature refinement. The former " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " interaction groups are utilized for separating " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": ", and the latter " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " groups for separating " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_2}^{\\mathrm{ini}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " (obtained by feeding the concatenation of " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_1}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " into a " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " convolutional layer) and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " (if " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " is available). We set " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " in practice. Details of the gated language interaction are described as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": "Adaptive global aggregation module (AGAM). As the network structure shown in Figure 3(a), given an input visual feature " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " with the spatial resolution of " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", AGAM is designed for adaptively obtaining a global feature " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " that aggregates the contextual information for the subsequent interaction. The visual feature " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " is firstly averaged by an average pooling layer to obtain " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ". Then the aggregation process is accomplished via a cross-attention mechanism which contains three linear projection layers with layer normalization [1] to conduct query, key, and value projections: " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}} = \\operatorname{Softmax}(\\mathbf{QK}^{\\top} / \\tau)\\mathbf{V}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", where the query " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " is projected from " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", and the key " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{(hw + 1) \\times C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " and value " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{(hw + 1) \\times C}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " are from the concatenation of " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{avg}}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " is a learnable scaling factor to control the magnitude of the dot product of " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 304, + 230, + 546, + 434 + ], + "type": "text", + "content": " before applying the softmax function." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": "Adaptive global interaction module (AGIM). Inspired by existing reflection separation approaches [23, 58] which attempt to distinguish transmission and reflection layers in the feature space by feature channel rearrangement (i.e., allocating distinct channels from mixture image features to the two layers), we propose to integrate language interactions at the feature channel level. To achieve this, as illustrated in Figure 3(b), we propose the adaptive global interaction module (AGIM), which employs the channel-wise cross-attention mechanism to conduct interactions between global language features and image features for channel rearrangement. The query, key, and value projections using linear projection layers with layer normalization [1] are conducted on the global language feature " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_i}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", the global visual feature " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{gl}}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", and the visual feature " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{C \\times h w}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": ", respectively. Before the interaction, a language gate is designed to prevent impacts of unavailable language descriptions caused by the recognizable layer ambiguity problem that users may only input one description for the layer recognizable in the mixture image. Specifically, if the description " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " corresponding to the current global language feature " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_i}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " is available (not set to null), the gate will feed " + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}_i}" + }, + { + "bbox": [ + 304, + 437, + 546, + 715 + ], + "type": "text", + "content": " into the following interaction" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24916" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 71, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 71, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 71, + 287, + 95 + ], + "type": "text", + "content": "process, otherwise the gate will feed " + }, + { + "bbox": [ + 46, + 71, + 287, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{glo}}" + }, + { + "bbox": [ + 46, + 71, + 287, + 95 + ], + "type": "text", + "content": " , which turns the interaction process to be a channel-wise self-attention." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": "After filtering by the language gate, an interacted feature can be obtained through the channel-wise cross-attention: " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{itr}} = \\operatorname{Softmax}(\\mathbf{Q}^\\top \\mathbf{K} / \\eta)\\mathbf{V}" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": " is a learnable scaling factor to control the magnitude of the attention map. Besides, to adaptively adjust the influence of language guidance based on the correspondence between image and language features, we define another scaling factor " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": " valued as the cosine distance between " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": " to multiply with the interacted feature " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{itr}}" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": ". Finally, the output of AGIM is obtained by a residual structure: " + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{V}}^{\\mathrm{out}} = \\mathbf{F}_{\\mathbf{V}}^{\\mathrm{in}} + s\\mathbf{F}^{\\mathrm{itr}}" + }, + { + "bbox": [ + 46, + 96, + 288, + 240 + ], + "type": "text", + "content": ", which integrates the contextual information from the language description and adjusts features for layer separation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 248, + 141, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 248, + 141, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 248, + 141, + 261 + ], + "type": "text", + "content": "3.3. Layer recovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": "After global language-image interaction and progressive refinement by the gated language interaction stage, we obtain layer features " + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_i}" + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": " which integrates holistic contextual information from the corresponding language description " + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": " (if available). The layer recovery stage is dedicated to reconstructing each image layer " + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": " from its corresponding layer features " + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}_i}" + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": ", which is achieved by using individual image decoders. In the image decoder, the layer feature is firstly upsampled by a transposed convolutional layer with a GELU activation [17], then refined by a residual block, and finally projected from the feature domain back into the image domain by a " + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 267, + 287, + 434 + ], + "type": "text", + "content": " convolutional layer with a sigmoid activation, thus accomplishing a precise layer recovery." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 443, + 138, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 138, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 138, + 454 + ], + "type": "text", + "content": "3.4. Loss functions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": "In this section, we introduce the contrastive correspondence loss " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ctr}}" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": " and the layer correspondence loss " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lcr}}" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": " for constraining the proposed method to establish the cross-modality correspondence between language descriptions " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": " and corresponding image layers under the content disturbance from counter layers in mixture images. We also briefly describe the image layer loss " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": " which consists of image- and feature-level loss functions for layer recovery. We denote the estimated image layers as " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_i" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": " and ground truths as " + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 46, + 462, + 287, + 581 + ], + "type": "text", + "content": ". Details of the loss functions are as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Contrastive correspondence loss. CLIP [42] conducts contrastive language-image pre-training within batches which successes in establishing cross-modality correspondence between language descriptions and clean images, while our language-guided image reflection separation task requires finding the correspondence between the given language description and a certain layer in the mixture image with the interference of extraneous contents. To tackle the above issue, we propose a contrastive correspondence loss that conducts contrastive learning between image layers to establish correct cross-modality correspondence. Specifi-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "cally, given a language description " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": ", the goal of the contrastive correspondence loss is to force the network to learn the relation that the corresponding image layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " is more relevant to " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " than the counter layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "j \\neq i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "), thus constraining the estimated image layer " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": " to conform to the associated language description " + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 72, + 545, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "text", + "content": "To measure the relevance between a language description " + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "text", + "content": " and an image layer " + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "text", + "content": ", we define a feature-level similarity function " + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 145, + 545, + 182 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 355, + 190, + 545, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 190, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 355, + 190, + 545, + 206 + ], + "type": "interline_equation", + "content": "\\mathcal {D} (\\mathbf {L}, \\mathbf {I}) = \\sigma \\left(\\Psi \\left(\\mathbf {F} _ {\\mathbf {L}} ^ {\\mathrm {g l o}}, \\mathbf {F} _ {\\mathbf {I}} ^ {\\mathrm {g l o}}\\right)\\right), \\tag {1}", + "image_path": "9858ca5c2ab35dd966ca8112d05dc67a110f268593806ef76ea95a98c304910b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": " represents the sigmoid function, " + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": " represents the cosine distance, " + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{L}}^{\\mathrm{glo}}" + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": " is the global language feature obtained by the language encoder (in Sec. 3.1), and " + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}}^{\\mathrm{glo}}" + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": " is the global image feature produced by the AGAM (in Sec. 3.2). For an available language description " + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 215, + 545, + 288 + ], + "type": "text", + "content": ", we calculate its contrastive correspondence loss as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 298, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 298, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 315, + 298, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {j}\\right) = - \\log \\left(\\frac {\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right)}{\\mathcal {D} \\left(\\mathbf {L} _ {i} , \\tilde {\\mathbf {I}} _ {i}\\right) + \\mathcal {D} \\left(\\mathbf {L} _ {i} , \\mathbf {I} _ {j}\\right)}\\right), \\tag {2}", + "image_path": "d3459d7e9f2ce86aa38a4b20180b50d2a1d6ec49b27794e94ed32c5a77ab5f9d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 335, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 358 + ], + "type": "text", + "content": "and we sum up the contrastive correspondence loss of each available language description as the final one." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "content": "Layer correspondence loss. We further define a layer correspondence loss to encourage the relevance between the language description " + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "content": " and the estimated image layer " + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{I}}_i" + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "content": " to approach the relevance between " + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i" + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "content": " and the ground truth image layer " + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 304, + 360, + 545, + 419 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 428, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 428, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 324, + 428, + 545, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {l c r}} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}, \\mathbf {I} _ {i}\\right) = \\left\\| \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\tilde {\\mathbf {I}} _ {i}\\right) - \\mathcal {D} \\left(\\mathbf {L} _ {i}, \\mathbf {I} _ {i}\\right) \\right\\| _ {1}, \\tag {3}", + "image_path": "10fc6a1e8aa2f578ef62ab5c4ea67b06e0d589f01c29fe95fab908a74f9e1e57.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 457, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 457, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}(\\cdot)" + }, + { + "bbox": [ + 304, + 457, + 545, + 506 + ], + "type": "text", + "content": " is the same feature-level similarity function as in the contrastive correspondence loss. We also sum up the layer correspondence loss of each available language description for final supervision." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": "Image layer loss. To achieve high-fidelity recovery of image layers (i.e., transmission and reflection layers), the proposed method is also optimized with loss functions following previous reflection separation methods [23, 52, 67]. Specifically, we utilize loss functions that conduct constraints on the visual quality of estimated images (i.e., the pixel " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pix}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": ", structural similarity " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ssim}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": ", and perceptual loss " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{per}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": ") or exploit the inherent relationship between two layers (i.e., the exclusion " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{exc}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": " and reconstruction loss " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": "), and we denote the combination of the above image- or feature-level loss functions as the image layer loss " + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}^1" + }, + { + "bbox": [ + 304, + 506, + 545, + 638 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 638, + 531, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 638, + 531, + 650 + ], + "spans": [ + { + "bbox": [ + 317, + 638, + 531, + 650 + ], + "type": "text", + "content": "Overall, the total loss function is then formulated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 660, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 660, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 345, + 660, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\gamma_ {1} \\mathcal {L} _ {\\mathrm {c t r}} + \\gamma_ {2} \\mathcal {L} _ {\\mathrm {l c r}} + \\mathcal {L} _ {\\mathrm {i m g}}, \\tag {4}", + "image_path": "c0f9f1150a040335883532e75c51b31bbab28e7bf47baff43ef7259925a17fe8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 683, + 484, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 683, + 484, + 695 + ], + "spans": [ + { + "bbox": [ + 306, + 683, + 484, + 695 + ], + "type": "text", + "content": "where coefficients are set as " + }, + { + "bbox": [ + 306, + 683, + 484, + 695 + ], + "type": "inline_equation", + "content": "\\gamma_{1} = \\gamma_{2} = 0.5" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 702, + 515, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 515, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 515, + 714 + ], + "type": "text", + "content": "Details of " + }, + { + "bbox": [ + 316, + 702, + 515, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}" + }, + { + "bbox": [ + 316, + 702, + 515, + 714 + ], + "type": "text", + "content": " are provided in the supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24917" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "type": "text", + "content": "3.5. Training strategy" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 270 + ], + "type": "text", + "content": "Due to the recognizable layer ambiguity that sometimes only one layer in the mixture image is recognizable, we propose a randomized training strategy to synergize with the gated language interaction mechanism (in Sec. 3.2). Since our training data simulates the recognizable layer ambiguity that some image layers do not have corresponding language descriptions (introduced in the next section), we only feed the available language description corresponding to the other image layer to guide the separation. For data with descriptions of both layers available, which indicates that both layers are recognizable, we also randomly drop one language description and feed the remaining one into the network to improve the generalization capacity for the proposed method. In practice, we set the ratio of dropping language descriptions to " + }, + { + "bbox": [ + 46, + 91, + 289, + 270 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 46, + 91, + 289, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "text", + "content": "We implement the proposed method with PyTorch [41] with a batch size of 16 on two Nvidia GeForce RTX 3090 GPUs. The model is trained for 40 epochs with Adam optimizer [25] to update learnable parameters. Weights are initialized as in [15]. The learning rate is set to " + }, + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "text", + "content": " initially and decreases to " + }, + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 271, + 288, + 344 + ], + "type": "text", + "content": " at epoch 30." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 357, + 151, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 151, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 151, + 371 + ], + "type": "text", + "content": "4. Data preparation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 289, + 498 + ], + "type": "text", + "content": "Though existing works have constructed several datasets for single-image reflection separation [31, 49, 67], they are unavailable for the proposed language-guided reflection separation framework due to the lack of corresponding language descriptions. Therefore, we build a dataset containing both synthetic and real data to overcome the data deficiency and facilitate network training and evaluation. Each group of data is composed of a mixture image, a transmission layer, a reflection layer, and two language descriptions. Details of synthetic and real data are as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 508, + 138, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 138, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 138, + 521 + ], + "type": "text", + "content": "4.1. Synthetic data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "text", + "content": "The synthetic dataset is generated for network training to satisfy the data-driven need of the proposed method. Due to the demand for paired image-language data, we utilize two prevailing image captioning datasets (i.e., Flickr30k [66] and COCO Captions [7]) for data generation, which contain 31,000 and 330,000 images respectively, and each image has 5 independent human-generated language descriptions. We randomly select images from the above two datasets as transmission " + }, + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{S}}" + }, + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "text", + "content": " and reflection scene images " + }, + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{S}}" + }, + { + "bbox": [ + 46, + 528, + 287, + 648 + ], + "type": "text", + "content": " and conduct an image synthesis process with linear blending [23]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 654, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 654, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 104, + 654, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} = \\hat {\\mathbf {T}} + \\hat {\\mathbf {R}} = \\alpha \\hat {\\mathbf {T}} _ {\\mathrm {S}} + \\beta \\hat {\\mathbf {R}} _ {\\mathrm {S}}, \\tag {5}", + "image_path": "3e4d7794b11a8a6c38049a5fa2013c9f1dd0278ae4e34bf9ffbf599ec78e0a2c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{T}_{\\mathrm{S}})" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}_{\\mathrm{S}} = g_{\\mathrm{inv}}(\\mathbf{R}_{\\mathrm{S}})" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{inv}}" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " represents the inverse gamma correction, and " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0.8,1]" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\beta \\in [0.4,1]" + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": " are the blending attenuation coefficients as in [23]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": "Considering the recognizable layer ambiguity that sometimes only one layer is recognizable in a mixture image, we assign a language description only when the corresponding layer is obvious enough in the synthesized mixture image, i.e., " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{mean}(\\hat{\\mathbf{V}}_l)}{\\mathrm{mean}(\\hat{\\mathbf{V}}_{\\mathbf{M}})}\\geqslant \\mu" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_l" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " represents the brightness image of " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " in the HSV color space and " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{V}}_{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " represents the brightness image of " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": ", and we set " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\mu = 0.3" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " in the data generation process. Finally, gamma correction is applied to image triplets " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{T}},\\hat{\\mathbf{R}},\\hat{\\mathbf{M}}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T},\\mathbf{R},\\mathbf{M}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 200 + ], + "type": "text", + "content": ", and we generate 50000 triplets of data in total for network training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 210, + 373, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 210, + 373, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 210, + 373, + 221 + ], + "type": "text", + "content": "4.2. Real data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 228, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 529 + ], + "type": "text", + "content": "Existing real datasets collected for the single-image reflection separation task usually contain mixture images with ground truth of transmission layers (e.g., Zhang et al. [67] and Nature [31]), and " + }, + { + "bbox": [ + 304, + 228, + 547, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{SIR}^2" + }, + { + "bbox": [ + 304, + 228, + 547, + 529 + ], + "type": "text", + "content": " [49, 54] further captures ground truths of reflection layers. For these off-the-shelf real datasets, we augment them by manually adding language descriptions for each group of data to satisfy the input setting of the proposed language-guided reflection separation task. Specifically, following the annotation principle of COCO Captions [7], we first describe the content of transmission layers in mixture images with entities, attributes (e.g., colors or materials), and relative positions between different entities. If the content of reflection layers is recognizable in mixture images, we also give language descriptions for reflection layers in the same way. To further evaluate the generalization capacity of the proposed method, we collect a real dataset (denoted as REFOL dataset) containing 100 mixture images from the Internet that are captured in different scenes and with different cameras. We annotate these mixture images with language descriptions in the same manner as mentioned above. Following the training strategy of previous methods [10, 31], we utilize 200 image pairs from Nature dataset [31] and 90 pairs from Zhang et al. [67] for training, and the rest of real data are used for quantitative and qualitative evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 542, + 388, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 388, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 388, + 555 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 562, + 526, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 526, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 526, + 575 + ], + "type": "text", + "content": "5.1. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "To evaluate the performance of the proposed method, we conduct quantitative and qualitative experiments on existing real datasets [31, 49, 67] (with our manually annotated language descriptions) and our newly collected dataset REFOL. We compare with state-of-the-art single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], IBCLN [31], CoRRN [52], and Zhang et al. [67]. For fair comparisons, we finetune the above methods on our training data if their training codes are provided. We report better results between the original pre-trained model and the finetuned version." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24918" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 96, + 520, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 1. Comparison of quantitative results in terms of PSNR [24] and SSIM [55] on real datasets for evaluating the recovery of transmission layers. " + }, + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "type": "text", + "content": " indicates larger (smaller) values are better. Bold numbers indicate the best-performing results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 96, + 520, + 243 + ], + "lines": [ + { + "bbox": [ + 71, + 96, + 520, + 243 + ], + "spans": [ + { + "bbox": [ + 71, + 96, + 520, + 243 + ], + "type": "table", + "html": "
Dataset (size)MetricsMethods
Zhang et al. [67]CoRRN [52]IBCLN [31]Dong et al. [10]YTMT [22]DSRNet [23]Ours
Postcard (199)PSNR↑20.8522.0423.4123.7222.8224.8825.02
SSIM↑0.8720.8700.8720.9030.8850.9100.915
Object (200)PSNR↑23.8425.1324.5224.3624.6826.4426.51
SSIM↑0.8720.9120.8910.8980.8920.9210.927
Wild (101)PSNR↑24.9725.1724.7825.7525.7025.8626.23
SSIM↑0.8750.8890.8840.9030.8970.9080.925
Real20 (20)PSNR↑22.3421.4321.4723.3423.2323.8824.05
SSIM↑0.7950.8010.7620.8120.8020.8160.824
Nature (20)PSNR↑20.6220.7523.7223.4521.5322.2623.87
SSIM↑0.7530.7830.8060.8080.7780.8010.812
Average (540)PSNR↑22.7723.7024.0224.3124.0125.5125.72
SSIM↑0.8650.8830.8750.8940.8830.9060.914
", + "image_path": "6d294cb49880cb8af0cec613e2bd28c4062a0dc2d2ccd20bd29eb7cbc343b06c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 255, + 214, + 348 + ], + "blocks": [ + { + "bbox": [ + 50, + 255, + 214, + 348 + ], + "lines": [ + { + "bbox": [ + 50, + 255, + 214, + 348 + ], + "spans": [ + { + "bbox": [ + 50, + 255, + 214, + 348 + ], + "type": "image", + "image_path": "31fd5974efffb740d30321de4e2dc25e963bae97e2aaf4c24cacd2e7740935d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 215, + 255, + 378, + 348 + ], + "blocks": [ + { + "bbox": [ + 215, + 255, + 378, + 348 + ], + "lines": [ + { + "bbox": [ + 215, + 255, + 378, + 348 + ], + "spans": [ + { + "bbox": [ + 215, + 255, + 378, + 348 + ], + "type": "image", + "image_path": "17f7b10bc75071d9448dfd2b7eca0c844c71f26f1e39d7117335d3c522fd51b1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 379, + 255, + 543, + 348 + ], + "blocks": [ + { + "bbox": [ + 379, + 255, + 543, + 348 + ], + "lines": [ + { + "bbox": [ + 379, + 255, + 543, + 348 + ], + "spans": [ + { + "bbox": [ + 379, + 255, + 543, + 348 + ], + "type": "image", + "image_path": "059f90c0b7d490404328ff5ae74a15d53a34e038f14c2df3ee82211ceb17ff2e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 350, + 214, + 441 + ], + "blocks": [ + { + "bbox": [ + 51, + 350, + 214, + 441 + ], + "lines": [ + { + "bbox": [ + 51, + 350, + 214, + 441 + ], + "spans": [ + { + "bbox": [ + 51, + 350, + 214, + 441 + ], + "type": "image", + "image_path": "0a4ab7a730efd3cbfc4092ee474c0b8b2d33ed415a98d21b0ed345b9c65a3e67.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 215, + 350, + 378, + 441 + ], + "blocks": [ + { + "bbox": [ + 215, + 350, + 378, + 441 + ], + "lines": [ + { + "bbox": [ + 215, + 350, + 378, + 441 + ], + "spans": [ + { + "bbox": [ + 215, + 350, + 378, + 441 + ], + "type": "image", + "image_path": "5485608babf34c448d34d210adc4a6a3ead931fa5a414fc6e64b3a070f2e6dcb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 379, + 350, + 543, + 441 + ], + "blocks": [ + { + "bbox": [ + 379, + 350, + 543, + 441 + ], + "lines": [ + { + "bbox": [ + 379, + 350, + 543, + 441 + ], + "spans": [ + { + "bbox": [ + 379, + 350, + 543, + 441 + ], + "type": "image", + "image_path": "9d60594cfba8391e7c2d2b5a80d2c6733a02a6be7da9bcecd91bde3255b50fcc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 51, + 448, + 214, + 541 + ], + "blocks": [ + { + "bbox": [ + 51, + 448, + 214, + 541 + ], + "lines": [ + { + "bbox": [ + 51, + 448, + 214, + 541 + ], + "spans": [ + { + "bbox": [ + 51, + 448, + 214, + 541 + ], + "type": "image", + "image_path": "d67069550ba292446096683f0bc1ad631394316de46bb16afec1baceaacdd538.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 215, + 448, + 378, + 540 + ], + "blocks": [ + { + "bbox": [ + 215, + 448, + 378, + 540 + ], + "lines": [ + { + "bbox": [ + 215, + 448, + 378, + 540 + ], + "spans": [ + { + "bbox": [ + 215, + 448, + 378, + 540 + ], + "type": "image", + "image_path": "d983f8002a4d9565fb15bf90455a7786d881941321a17cfbfc46af44816c0228.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 379, + 448, + 543, + 540 + ], + "blocks": [ + { + "bbox": [ + 379, + 448, + 543, + 540 + ], + "lines": [ + { + "bbox": [ + 379, + 448, + 543, + 540 + ], + "spans": [ + { + "bbox": [ + 379, + 448, + 543, + 540 + ], + "type": "image", + "image_path": "a5e58dd925ae40e1a3efcb9d41aef3aa2b55f2a0198b18c3f64bf564399e3836.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 51, + 544, + 214, + 634 + ], + "blocks": [ + { + "bbox": [ + 51, + 544, + 214, + 634 + ], + "lines": [ + { + "bbox": [ + 51, + 544, + 214, + 634 + ], + "spans": [ + { + "bbox": [ + 51, + 544, + 214, + 634 + ], + "type": "image", + "image_path": "740aa01ce9ffaec58414e5334f1279279cc01321a5e03f938f0465f648d35a74.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 637, + 545, + 659 + ], + "lines": [ + { + "bbox": [ + 47, + 637, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 545, + 659 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of estimated transmission and reflection layers on real data, compared with the state-of-the-art methods including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Please zoom in for details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 215, + 544, + 378, + 634 + ], + "blocks": [ + { + "bbox": [ + 215, + 544, + 378, + 634 + ], + "lines": [ + { + "bbox": [ + 215, + 544, + 378, + 634 + ], + "spans": [ + { + "bbox": [ + 215, + 544, + 378, + 634 + ], + "type": "image", + "image_path": "64d42f4f91c4c08337d7600a88145badafad6673c031b19c7b91a535505974d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 379, + 544, + 543, + 634 + ], + "blocks": [ + { + "bbox": [ + 379, + 544, + 543, + 634 + ], + "lines": [ + { + "bbox": [ + 379, + 544, + 543, + 634 + ], + "spans": [ + { + "bbox": [ + 379, + 544, + 543, + 634 + ], + "type": "image", + "image_path": "7218db79820066185abdbc9e2cbb6e86699ac06643f166da2a840f6817eb7d63.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 680, + 288, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 704 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 704 + ], + "type": "text", + "content": "Quantitative comparison. Quantitative experiments are conducted on three real datasets for reflection separation," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 680, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 680, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 306, + 680, + 545, + 704 + ], + "type": "text", + "content": "i.e., Nature [31], Real20 [67], and three subsets of " + }, + { + "bbox": [ + 306, + 680, + 545, + 704 + ], + "type": "inline_equation", + "content": "\\mathrm{SIR}^2" + }, + { + "bbox": [ + 306, + 680, + 545, + 704 + ], + "type": "text", + "content": " [49] dataset. Following the setting of existing reflection separa" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24919" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 86, + 285, + 129 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 253, + 81 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation studies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 86, + 285, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 86, + 285, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 86, + 285, + 129 + ], + "type": "table", + "html": "
Metricsw/o languagew/o AGIMLimg onlyOurs
PSNR↑24.3124.6924.5225.72
SSIM↑0.8850.9010.8930.914
", + "image_path": "c3ad27059d69f21233bb823cc22740c1eabe7fad97c112435aa6e70ec93aba12.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 286, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 286, + 216 + ], + "type": "text", + "content": "tion methods [10, 36], we utilize PSNR [24] and SSIM [55] as error metrics for evaluating the recovery of transmission layers2. As quantitative results shown in Table 1, the proposed method achieves the best performance of both PSNR and SSIM, which validates its generalization capacity and the effectiveness of language descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 217, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 287, + 407 + ], + "type": "text", + "content": "Qualitative comparison. To evaluate the visual quality of reflection separation results, we compare the proposed method with four single-image reflection separation methods, including DSRNet [23], YTMT [22], Dong et al. [10], and IBCLN [31]. Qualitative results on recovering both transmission and reflection layers are shown in Figure 4. As can be observed in Figure 4, IBCLN [31] can only remove parts of reflections, while Dong et al. [10] have trouble in dealing with complex semantic images like the reflections of a man in a room (the second example). YTMT [22] separates layers incorrectly and thus brings the content of transmission layers into reflections. DSRNet [23] fails to recover transmission and reflection layers for these challenging cases. Contributing to the language guidance, the proposed method generates better visual results and recovers both transmission and reflection layers neatly." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 415, + 138, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 415, + 138, + 428 + ], + "spans": [ + { + "bbox": [ + 47, + 415, + 138, + 428 + ], + "type": "text", + "content": "5.2. Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "text", + "content": "In this section, we conduct several ablation studies with quantitative results shown in Table 2 to investigate the influence of the additional input of language descriptions (denoted as 'w/o language'), the network design of AGIM (denoted as 'w/o AGIM'), and the loss functions for cross-modality correspondence (denoted as " + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "text", + "content": " only). The performance of the variant 'w/o language' suffers from an obvious degradation as we remove language descriptions, which shows the effectiveness of the additional contextual information for layer separation. The variant 'w/o AGIM' replaces AGIMs with simple feature fusion blocks which directly concatenate language and image features and feed them into self-attention blocks, and the decline in performance validates the necessity of our gated interaction mechanism. The variant " + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "text", + "content": " only is trained with " + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{img}}" + }, + { + "bbox": [ + 46, + 434, + 287, + 648 + ], + "type": "text", + "content": ", which obtains results slightly better than the variant 'w/o language', indicating the significance of establishing the cross-modality correspondence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 649, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 649, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 649, + 286, + 685 + ], + "type": "text", + "content": "To further verify the effectiveness of the language interaction, we conduct an ablation study by gradually increasing the number of input descriptions. As shown in yellow" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 543, + 264 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 543, + 264 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 543, + 264 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 543, + 264 + ], + "type": "image", + "image_path": "72e62436b20e32a8267b24dcfeec431c63c50b730e8166a45c80428798e75f44.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 266, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 306, + 266, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 545, + 289 + ], + "type": "text", + "content": "Figure 5. Results with different numbers of input language descriptions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 300, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 361 + ], + "type": "text", + "content": "boxes of Figure 5, by utilizing more input language descriptions, the separation of transmission (clothes) and reflection layers (a man with a camera) becomes more thorough, which also demonstrates the efficacy and robustness of the proposed method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 567 + ], + "type": "text", + "content": "This paper introduces natural language to provide contextual information about image layers for relieving the ill-posed reflection separation problem. We develop an end-to-end framework with adaptive global interaction modules and language-image loss functions to effectively manage the modality inconsistency, and we adopt a language gate mechanism with randomized training strategies to handle the recognizable layer ambiguity. To address data deficiency, a specially built dataset with language annotations significantly aids in training and evaluating the proposed language-guided image reflection separation framework. Quantitative and qualitative experiments on real data demonstrate the effectiveness of introducing language descriptions for reflection separation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 640 + ], + "type": "text", + "content": "Limitations. The proposed method may not distinguish transmission and reflection layers accurately when their contents are similar. For such ambiguous cases, a more flexible language-guided mechanism is needed. This might be solved by exploring a better interaction approach, which is left as our future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 403, + 671 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 679, + 545, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 679, + 545, + 702 + ], + "spans": [ + { + "bbox": [ + 306, + 679, + 545, + 702 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 286, + 712 + ], + "type": "text", + "content": "2Quantitative evaluations on the recovery of reflection layers are provided in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24920" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Yakun Chang, Cheolkon Jung, and Jun Sun. Joint reflection removal and depth estimation from a single image. IEEE TCYB, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 192 + ], + "type": "text", + "content": "[3] Yakun Chang, Cheolkon Jung, Jun Sun, and Fengqiao Wang. Siamese dense network for reflection removal with flash and no-flash image pairs. IJCV, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[4] Zheng Chang, Shuchen Weng, Yu Li, Si Li, and Boxin Shi. L-CoDer: Language-based colorization with color-object decoupling transformer. In Proc. of ECCV, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 259 + ], + "type": "text", + "content": "[5] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CoIns: Language-based colorization with instance awareness. In Proc. of CVPR, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 305 + ], + "type": "text", + "content": "[6] Zheng Chang, Shuchen Weng, Peixuan Zhang, Yu Li, Si Li, and Boxin Shi. L-CAD: Language-based colorization with any-level descriptions using diffusion priors. In Proc. of NeurIPS, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 288, + 350 + ], + "type": "text", + "content": "[7] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 288, + 396 + ], + "type": "text", + "content": "[8] Congyue Deng, Chiyu Jiang, Charles R Qi, Xinchen Yan, Yin Zhou, Leonidas Guibas, Dragomir Anguelov, et al. NeRDi: Single-view NeRF synthesis with language-guided diffusion as general image priors. In Proc. of CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "type": "text", + "content": "[9] Yaron Diamant and Yoav Y Schechner. Overcoming visual reverberations. In Proc. of CVPR, 2008. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 287, + 453 + ], + "type": "text", + "content": "[10] Zheng Dong, Ke Xu, Yin Yang, Hujun Bao, Weiwei Xu, and Rynson WH Lau. Location-aware single image reflection removal. In Proc. of ICCV, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 454, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 454, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 454, + 287, + 496 + ], + "type": "text", + "content": "[11] Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf. A generic deep architecture for single image reflection removal and image smoothing. In Proc. of ICCV, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "type": "text", + "content": "[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proc. of NeurIPS, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 576 + ], + "type": "text", + "content": "[13] Byeong-Ju Han and Jae-Young Sim. Zero-shot learning for reflection removal of single 360-degree image. In Proc. of ECCV, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 611 + ], + "type": "text", + "content": "[14] Bharath Hariharan, Pablo Arbeláez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In Proc. of CVPR, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance onImagenet classification. In Proc. of ICCV, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. of CVPR, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (GeLUs). arXiv preprint arXiv:1606.08415, 2016. 4, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[18] Yuchen Hong, Youwei Lyu, Si Li, and Boxin Shi. Near-infrared image guided reflection removal. In Proc. of ICME, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[19] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. Panoramic image reflection removal. In Proc. of CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 172 + ], + "type": "text", + "content": "[20] Yuchen Hong, Youwei Lyu, Si Li, Gang Cao, and Boxin Shi. Reflection removal with NIR and RGB image feature fusion. IEEE TMM, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[21] Yuchen Hong, Qian Zheng, Lingran Zhao, Xudong Jiang, Alex C. Kot, and Boxin Shi. PAR" + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "Net: End-to-end panoramic image reflection removal. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 240 + ], + "type": "text", + "content": "[22] Qiming Hu and Xiaojie Guo. *Trash or treasure? an interactive dual-stream strategy for single image reflection separation.* In *Proc. of NeurIPS*, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 274 + ], + "type": "text", + "content": "[23] Qiming Hu and Xiaojie Guo. Single image reflection separation via component synergy. In Proc. of ICCV, 2023. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[24] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of PSNR in image/video quality assessment. *Electronics letters*, 2008, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 310, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 545, + 341 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 375 + ], + "type": "text", + "content": "[26] Naejin Kong, Yu-Wing Tai, and Sung Yong Shin. A physically-based approach to reflection separation. In Proc. of CVPR, 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 408 + ], + "type": "text", + "content": "[27] Chenyang Lei and Qifeng Chen. Robust reflection removal with reflection-free flash-only cues. In Proc. of CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 453 + ], + "type": "text", + "content": "[28] Chenyang Lei, Xuhua Huang, Mengdi Zhang, Qiong Yan, Wenxiu Sun, and Qifeng Chen. Polarized reflection removal with perfect alignment in the wild. In Proc. of CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "text", + "content": "[29] Chenyang Lei, Xudong Jiang, and Qifeng Chen. Robust reflection removal with flash-only cues in the wild. IEEE TPAMI, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 520 + ], + "type": "text", + "content": "[30] Anat Levin and Yair Weiss. User assisted separation of reflections from a single image using a sparsity prior. IEEE TPAMI, 2007. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 555 + ], + "type": "text", + "content": "[31] Chao Li, Yixiao Yang, Kun He, Stephen Lin, and John E Hopcroft. Single image reflection removal through cascaded refinement. In Proc. of CVPR, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 588 + ], + "type": "text", + "content": "[32] Yu Li and Michael S Brown. Exploiting reflection change for automatic reflection removal. In Proc. of ICCV, 2013. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 612 + ], + "type": "text", + "content": "[33] Yu Li and Michael S Brown. Single image layer separation using relative smoothness. In Proc. of CVPR, 2014. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 645 + ], + "type": "text", + "content": "[34] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In Proc. of CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[35] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions with layered decomposition. IEEE TPAMI, 2021. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[36] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Reflection separation using a pair of unpolar" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24921" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 94 + ], + "type": "text", + "content": "ized and polarized images. In Proc. of NeurIPS, 2019. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[37] Youwei Lyu, Zhaopeng Cui, Si Li, Marc Pollefeys, and Boxin Shi. Physics-guided reflection separation from a pair of unpolarized and polarized images. IEEE TPAMI, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 163 + ], + "type": "text", + "content": "[38] Daiqian Ma, Renjie Wan, Boxin Shi, Alex C Kot, and Ling-Yu Duan. Learning to jointly generate and separate reflections. In Proc. of ICCV, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 196 + ], + "type": "text", + "content": "[39] Shree K Nayar, Xi-Sheng Fang, and Terrance Boult. Separation of reflection components using color and polarization. IJCV, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 230 + ], + "type": "text", + "content": "[40] Jonghyuk Park, Hyeona Kim, Eunpil Park, and Jae-Young Sim. Fully-automatic reflection removal for 360-degree images. In Proc. of WACV, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 285 + ], + "type": "text", + "content": "[41] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Proc. of NeurIPS, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 342 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. of ICML, 2021. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 343, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 287, + 365 + ], + "type": "text", + "content": "[43] Yoav Y Schechner, Nahum Kiryati, and Ronen Basri. Separation of transparent layers using focus. IJCV, 2000. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 398 + ], + "type": "text", + "content": "[44] YiChang Shih, Dilip Krishnan, Fredo Durand, and William T Freeman. Reflection removal using ghosting cues. In Proc. of CVPR, 2015. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 421 + ], + "type": "text", + "content": "[45] Christian Simon and In Kyu Park. Reflection removal for in-vehicle black box videos. In Proc. of CVPR, 2015. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 455 + ], + "type": "text", + "content": "[46] Jimeng Sun, Shuchen Weng, Zheng Chang, Si Li, and Boxin Shi. UniCoRN: A unified conditional image inpainting network. In Proc. of CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 498 + ], + "type": "text", + "content": "[47] Jiajun Tang, Haofeng Zhong, Shuchen Weng, and Boxin Shi. LuminAIRe: Illumination-aware conditional image inpainting for lighting-realistic generation. In Proc. of NeurIPS, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 544 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proc. of NeurIPS, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 578 + ], + "type": "text", + "content": "[49] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In Proc. of ICCV, 2017. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "text", + "content": "[50] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, Wen Gao, and Alex C. Kot. Region-aware reflection removal with unified content and gradient priors. IEEE TIP, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[51] Renjie Wan, Boxin Shi, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CRRN: Multi-scale guided concurrent reflection removal network. In Proc. of CVPR, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 679 + ], + "type": "text", + "content": "[52] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, Ah-Hwee Tan, and Alex C. Kot. CoRRN: Cooperative reflection removal network. IEEE TPAMI, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[53] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. IJCV, 2021. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[54] Renjie Wan, Boxin Shi, Haoliang Li, Yuchen Hong, Ling-Yu Duan, and Alex C Kot. Benchmarking single-image reflection removal algorithms. In IEEE TPAMI, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[55] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In Proc. of ACSSC, 2003. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 140, + 547, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 182 + ], + "type": "text", + "content": "[56] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proc. of CVPR, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 226 + ], + "type": "text", + "content": "[57] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. CRIS: CLIP-driven referring image segmentation. In Proc. of CVPR, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 270 + ], + "type": "text", + "content": "[58] Kaixuan Wei, Jiaolong Yang, Ying Fu, David Wipf, and Hua Huang. Single image reflection removal exploiting misaligned training data and network enhancements. In Proc. of CVPR, 2019. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[59] Qiang Wen, Yinjie Tan, Jing Qin, Wenxi Liu, Guoqiang Han, and Shengfeng He. Single image reflection removal beyond linearity. In Proc. of CVPR, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 305, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 326 + ], + "type": "text", + "content": "[60] Shuchen Weng and Boxin Shi. Conditional image repainting. IEEE TPAMI, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 370 + ], + "type": "text", + "content": "[61] Shuchen Weng, Wenbo Li, Dawei Li, Hongxia Jin, and Boxin Shi. MISC: Multi-condition injection and spatially-adaptive compositing for conditional person image synthesis. In Proc. of CVPR, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 414 + ], + "type": "text", + "content": "[62] Shuchen Weng, Hao Wu, Zheng Chang, Jiajun Tang, Si Li, and Boxin Shi. L-CoDe: Language-based colorization using color-object decoupled conditions. In Proc. of AAAI, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "text", + "content": "[63] Jie Yang, Dong Gong, Lingqiao Liu, and Qinfeng Shi. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. In Proc. of ECCV, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[64] Yang Yang, Wenye Ma, Yin Zheng, Jian-Feng Cai, and Weiyu Xu. Fast single image reflection suppression via convex optimization. In Proc. of CVPR, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[65] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-aware vision transformer for referring image segmentation. In Proc. of CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[66] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 602 + ], + "type": "text", + "content": "[67] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In Proc. of CVPR, 2018. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "text", + "content": "[68] Ya-Nan Zhang, Linlin Shen, and Qiufu Li. Content and gradient model-driven deep network for single image reflection removal. In Proc. of ACM MM, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 637, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 679 + ], + "type": "text", + "content": "[69] Qian Zheng, Jinnan Chen, Zhan Lu, Boxin Shi, Xudong Jiang, Kim-Hui Yap, Ling-Yu Duan, and Alex C Kot. What does plate glass reveal about camera calibration? In Proc. of CVPR, 2020. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[70] Qian Zheng, Boxin Shi, Jinnan Chen, Xudong Jiang, LingYu Duan, and Alex C. Kot. Single image reflection removal with absorption effect. In Proc. of CVPR, 2021. 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24922" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_content_list.json b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d3d803c1a4e409510f8e09dbe149fd9962461b6 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_content_list.json @@ -0,0 +1,1329 @@ +[ + { + "type": "text", + "text": "Language-only Efficient Training of Zero-shot Composed Image Retrieval", + "text_level": 1, + "bbox": [ + 112, + 130, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geonmo Gu*,1 Sanghyuk Chun*,2 Wonjae Kim² Yoohoon Kang¹ Sangwoo Yun² \n1NAVER Vision 2NAVER AI Lab * Equal contribution", + "bbox": [ + 127, + 179, + 841, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LinCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LinCIR is surprisingly efficient and highly effective; LinCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ. Code is available at github.com/navervision/lincir", + "bbox": [ + 75, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed image retrieval (CIR) is a challenging vision-language (VL) task that takes a composed query of image and text, aiming to search relative images for both conditions [29]. As language serves as the most natural method for encoding human interaction, CIR provides a higher degree of freedom and a better user experience for image-based search engine applications, such as web commerce.", + "bbox": [ + 75, + 718, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One of the main challenges of CIR is the expensive dataset collection pipeline. CIR datasets consist of triplets $\\langle x_{i_R},x_c,x_i\\rangle$ , where $x_{i_R}$ is an image query, $x_{c}$ is a text query, and $x_{i}$ is the target image. Unlike image-text paired datasets, such as CC3M [39] or LAION [38], such triplets", + "bbox": [ + 76, + 824, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e7e4cb41b2d7b9c6a250bd8548ab831d1dd012bd08b6834836418b9351d41c1.jpg", + "image_caption": [ + "Figure 1. Training time (hours) vs. Zero-shot Composed Image Retrieval (ZS-CIR) performance. Thanks to our efficient language-only training strategy, our LinCIR outperforms the previous ZS-CIR methods in both training time and CIR performance. The training time is measured on 8 A100 GPUs. We compare the models on the CIRCO mAP@5 [3] score for a more comprehensive evaluation of CIR models (more results are in Fig. 4). Notably, when we scale up the backbone CLIP [19, 35] model size by ViT-L, ViT-H and ViT-G, LinCIR shows a promising performance boost with surprisingly short training time (48 mins for ViT-G). On the other hand, Pic2Word [37] and SEARLE [3] cannot be scaled up to CLIP ViT-G due to their limitation on restricted textual expressions and the lack of diversity of input texts." + ], + "image_footnote": [], + "bbox": [ + 522, + 275, + 857, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are almost impossible to collect by web crawling and require expensive human labor to create each triplet. For example, the datasets are constructed by gathering candidates of $\\langle x_{i_R},x_i\\rangle$ and manually annotating $x_{c}$ by human annotators [29, 47], which is hard to scalable. Therefore, the size of the training triplets is usually small (e.g., 46.6k [47], 28.8k [29]), and the existing CIR methods trained on such triplets [2, 6-8, 11, 12, 20-22, 24, 30, 40, 42, 46, 48] suffer from the lack of generalizability to diverse unseen domains.", + "bbox": [ + 496, + 642, + 893, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome the drawback, recent studies have explored zero-shot CIR (ZS-CIR), a scalable direction, by eliminating the dependency on the pre-collected triplet datasets. For example, Saito et al. [37] and Baldrati et al. [3] propose projection-based ZS-CIR methods without using triplet datasets. Based on the pre-trained CLIP [35], they train a lightweight projection module $\\phi$ that projects the CLIP image latent embedding $z_{i}$ to the CLIP text token", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13225", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da7a488300c42f1f41157206c5c13ad3e4c7084ef8f0b852ef775a9cb0dcb844.jpg", + "image_caption": [ + "Figure 2. Overview of ZS-CIR with a projection to the token embedding space. The mainstream ZS-CIR methods, such as Pic2Word [37], SEARLE [3] and LinCIR (ours), train a projection module $\\phi$ that projects the image latent embedding $z_{i}$ into the token embedding space $e_c$ with a custom prompt (e.g., a photo of $[\\mathbb{S}]$ that [cond]). The textual encoder output is used for CIR." + ], + "image_footnote": [], + "bbox": [ + 163, + 87, + 803, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "embedding space $e_{c}$ (see Fig. 2). These approaches have shown promising generalizability to unseen datasets. However, these methods struggle to handle diverse text conditions because they rely on pre-defined naïve text prompts during training (e.g., a photo of [§]). Moreover, their training frameworks need sequential forward operations of the visual and textual encoders (see Fig. 3(a)), resulting in inefficient training and limited scalability to a larger backbone.", + "bbox": [ + 73, + 273, + 468, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a new paradigm of ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). As shown in Fig. 3 (b), instead of projecting the image latent embedding $z_{i}$ , we propose to project the text latent embedding $z_{c}$ to the token embedding space $e_{c}$ . We introduce a novel self-supervision, named Self-Masking Projection (SMP), for language-only training. We replace all the \"keywords\" of the original text with the projected text embedding of the original text to produce an embedding $\\widehat{z}_{c}$ and apply MSE loss between $z_{c}$ and $\\widehat{z}_{c}$ . Here, we define \"keyword\" as consecutive adjectives and nouns. For example, the keywords of \"gray cat sleeps on a pillow\" are \"gray cat\" and \"a pillow\"; therefore, it becomes \"[\\$] sleeps on [\\$]\". The purpose of SMP is to make [\\$] token interpreted as the \"one-word summarization\" of the input by extracting the essential information of the input. During inference, we simply perform text-to-image retrieval by projecting image embedding to the token embedding space using the projection module $\\phi$ as shown in Fig. 2. However, this strategy can suffer from the modality gap between textual and visual modalities [27], i.e., even though our $\\phi$ module works perfectly for text latent embeddings, it can underperform for the target visual embeddings. We mitigate the issue by employing a random noise addition strategy [33], carefully choosing a probability distribution that ensures the diversity of the noise-augmented textual embeddings.", + "bbox": [ + 73, + 396, + 473, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our paradigm has three advantages over the previous approaches. First, while Pic2Word and SEARLE projection modules are trained with a restricted text prompt, i.e., a photo of $\\left[\\S\\right]$ , the projection module of LinCIR is trained with the diverse text inputs from the actual texts. Due to this reason, Pic2Word and SEARLE show degenerated", + "bbox": [ + 75, + 787, + 470, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performances when the backbone size becomes larger (see Fig. 1). On the other hand, LinCIR is more generalizable to complex and diverse text conditions, showing superior ZS-CIR performances than others, especially for larger backbones. Second, as LinCIR only utilizes the textual encoder, our training process is highly efficient and scalable than the methods incorporated with the visual encoder; our language-only training strategy is $\\times 6.0$ faster than Pic2Word [37] and $\\times 8.4$ faster than SEARLE [3] with CLIP ViT-L backbone. When we scale up the backbone size to CLIP ViT-G, the gap becomes $\\times 16.4$ and $\\times 17.6$ , respectively. Even ViT-G training for LinCIR only takes 48 minutes using 8 A100 and less than 2 hours using 1 V100. Third, our method is storage-efficient; for example, the CC3M dataset [39] images occupy about 430GB storage size, while its captions only need 125MB. We train LinCIR only with 571MB storage size for storing the 5.5M training captions. In summary, LinCIR shows the best training time and ZS-CIR performance as shown in Fig. 1 and 4.", + "bbox": [ + 496, + 273, + 893, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution can be summarized as follows: (1) We propose LinCIR, a novel and efficient language-only training framework for ZS-CIR. (2) We introduce a new self-supervision for language-only training, named SelfMasking Projection (SMP). (3) We employ a better random noise addition strategy than naive Gaussian noise to mitigate the modality gap. (4) LinCIR achieves the best training time and the ZS-CIR performances on four ZS-CIR benchmarks (CIRCO [3], GeneCIS [44], FashionIQ [47] and CIRR [29]). Notably, LinCIR even outperforms the state-of-the-art supervised method [2] on FashionIQ.", + "bbox": [ + 496, + 561, + 893, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 739, + 638, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models (VLM). As with previous ZS-CIR methods, LinCIR utilizes a pre-trained VLM, such as CLIP [19, 35] or BLIP [25]. We use VLMs that map an image input $x_{i}$ and a text input $x_{c}$ to the $d$ -dimensional joint embedding space. Here, a given caption $x_{c}$ is tokenized by the pre-defined tokenizer as $t_c = \\{t_c^k\\mid k = 1\\dots K\\}$ , where $K$ is the number of tokens, and mapped to token embeddings $e_{c} = E_{w}(t_{c}) = \\{E_{w}(t_{c}^{k})\\mid k = 1\\dots K\\}$ , where $E_{w}$ is the embedding layer parameterized by $w$ . Afterwords, a", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Pronounced as \"linker\", meaning for linking the two modalities.", + "bbox": [ + 94, + 886, + 441, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13226", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bffd32666c233df65638f0fc875db44563415e25a58ae0809933479b61b140fb.jpg", + "image_caption": [ + "(a) Pic2Word training procedure" + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 421, + 218 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aaf2b51b98802a017066951ce6ddb7b73fa1fa632d3af887a59e9b99845b1df9.jpg", + "image_caption": [ + "(b) Our language-only training procedure", + "Figure 3. Comparison of Pic2Word [37] and LinCIR training procedures. (a) Pic2Word [37] and SEARLE [3] training procedure requires both the visual encoder and the textual encoder. They only need images for training, while the text prompt is pre-defined [37] or automatically generated [3]. (b) LinCIR is trained solely on texts with the frozen textual encoder. First, a projection module $\\phi$ projects a textual latent embedding of a sentence $z_{t}$ into the token embedding space. Before the projection, a random noise $n$ is added to $z_{t}$ to reduce the modality gap between text and image. We introduce a new self-supervision, named Self-Masking Projection (SMP), by replacing all keywords of the given caption with the projected embedding by $\\phi$ and extracting a modified text embedding $\\widehat{z}_{t}$ . Finally, the projection module $\\phi$ is trained by the MSE loss between $z_{t}$ and $\\widehat{z}_{t}$ . Note that both (a) and (b) use the same inference strategy shown in Fig. 2." + ], + "image_footnote": [], + "bbox": [ + 464, + 88, + 848, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "textual encoder $\\psi_C$ encodes the token embeddings to extract textual latent embeddings $z_{c} = \\psi_{C}(x_{c}) = \\psi_{C}(E_{w}(t_{c}))$ . The visual latent embeddings are extracted by the visual encoder $\\psi_I$ : $z_{i} = \\psi_{I}(x_{i})$ . We will use the terminology token embeddings to represent $e_c$ , and use textual (or visual) latent embeddings to represent $z_{c}$ (or $z_{i}$ ).", + "bbox": [ + 75, + 356, + 468, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VLM modality gap. The VLM joint embedding space suffers from the semantic gap between each modality. Namely, the visual and the textual latent embeddings are not exactly aligned with each other, but they are located in completely separate regions of the embedding space [27].", + "bbox": [ + 75, + 460, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Such a modality gap hinders the harmonization of visual and textual latents in the joint embedding space. To mitigate the gap, language-only training has been recently introduced for image captioning tasks [16, 26, 33]. They train a text decoder from the CLIP textual embeddings to generate image captions from visual embeddings. During training the decoder, the modality gap is bridged by injecting Gaussian noises [16, 33] or projection-based method [26].", + "bbox": [ + 75, + 537, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we utilize language-only training and a noise-addition strategy to mitigate the modality gap. We carefully studied the impact of noise and found a better distribution rather than simple Gaussian noise. We believe our findings can be transferred to other language-only training methods. Moreover, our target task, CIR, needs to model a relationship between triplets of $\\langle x_{i_R},x_c,x_i\\rangle$ , whereas previous language-only training methods only consider the pair-wise relationship (i.e., image captioning). We tackle this problem by proposing a novel self-supervision, named Self-Masking Projection (SMP), which injects the projected textual embeddings into the original token embeddings.", + "bbox": [ + 75, + 659, + 470, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CIR by projection to token embeddings. The mainstream ZS-CIR methods, such as Pic2Word [37] and SEARLE [3], employ a projection-based method. Namely,", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "they learn a projection module \\(\\phi\\) from the image latent embedding space to the token embedding space. For inference, they project the image latent embedding \\(z_{i}\\) to the token embedding (\\$), then perform text-to-image retrieval with the prompt \"a photo of [\\\\(] that [cond]\", where [cond] is a text condition. Fig. 2 shows an overview of how textual projection-based ZS-CIR works. The main research point of this field is how to train the projection module \\(\\phi\\) to capture the visual information into the token embedding space.", + "bbox": [ + 496, + 356, + 890, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pic2Word [37] trains the projection module $\\phi$ by minimizing contrastive loss between image latent embedding and the textual latent embedding of \"a photo of [\\$]\" (see Fig. 3 (a)). SEARLE [3] employs a similar approach to Pic2Word. First, they employ optimization-based textual inversion to generate pre-defined special tokens for an image and train $\\phi$ to predict the token embedding. SEARLE employs CLIP zero-shot classification to predict the \"concept\" of the given image and refine the prompt by letting GPT [4] continue the phrase. Both methods only use image inputs $x_{i}$ for training without accessing the CIR triplets.", + "bbox": [ + 496, + 492, + 892, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although Pic2Word and SEARLE achieve reasonable ZS-CIR performances, they have two significant problems. First, they heavily rely on the initial prompt, \"a photo of $[\\mathbb{S}]$ \", limiting the diversity of the textual encoder input. As Fig. 1, we argue that diversifying the input texts during training the projection module $\\phi$ is critical to train with larger backbones (e.g., CLIP ViT-G), while using the naive prompt is failed to scale up the backbone size. Second, they need image inputs, which are less compact and redundant than text datasets. Moreover, the visual encoder usually needs more computation resources than the textual encoder because the visual encoder takes the fixed length token (e.g., 256). In contrast, the textual encoder takes shorter token lengths (e.g., average token length of CC3M ≈ 10). We tackle the problems by introducing a language-only training method, showing remarkable efficiency and scalability.", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Language-only Training of Zero-shot CIR", + "text_level": 1, + "bbox": [ + 76, + 89, + 452, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces a new paradigm for ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). We first introduce a novel language-only self-supervision, Self-Masking Projection (SMP), that enables a language-only training for CIR ( $\\S 3.1$ ). Then, we explain the modality gap problem and empirically show that adding a carefully chosen random noise can mitigate the problem ( $\\S 3.2$ ). Finally, we describe the advantage of LinCIR in terms of efficiency and scalability ( $\\S 3.3$ ).", + "bbox": [ + 75, + 114, + 468, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Self-Masking Projection (SMP)", + "text_level": 1, + "bbox": [ + 76, + 258, + 354, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We aim to learn a projection module $\\phi$ that captures and retains the original visual information after the projection and the textual encoder. While the previous methods focus on directly mapping visual information to the token space with a naive prompt (i.e., a photo of $[\\S])$ , we argue that focusing on the textual encoder is more important. Our zero-shot CIR is based on text-to-image retrieval (see Fig. 2). It means that the quality of the textual latent embedding is more critical to the final ZS-CIR performances. Hence, rather than focusing on minimizing the gap between visual information and the naive prompt, we aim to achieve a projection module $\\phi$ that captures the semantics of the keywords in the given text.", + "bbox": [ + 75, + 281, + 468, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve our goal, we introduce a novel language-only self-supervision named Self-Masking Projection (SMP). First, we project the textual embedding $z_{c}$ of a given text input $x_{c}$ with the projection module $\\phi$ to the token embedding space, i.e., $\\widehat{e}_{c} = \\phi(z_{c})$ , where $\\widehat{e}_{c}$ is the projected textual embedding. Then, we replace the token embeddings of all the \"keywords\" of $x_{c}$ with the projected token embedding $\\widehat{e}_{c}$ . We define the keywords of the sentence as consecutive nouns and adjectives. For example, the keywords of \"A Russian Blue cat is gray and cute\" will be \"A Russian Blue cat\", \"gray\" and \"cute\"; hence it will be converted to \"[$] is [\\$] and [\\$]\", where [\\$] is a special token to represent the projected token embedding $\\widehat{e}_{c}$ . By treating all the main concepts (keywords) in the caption as the same [\\$], we intend [\\$] to represent the overall essential information of the inputs. Note that similar to [MASK] tokens of masked modeling, [\\$] with different positions will be encoded in different features due to the positional embeddings. Using the converted caption, we extract a converted textual latent feature $\\widehat{z}_{c}$ and minimize MSE loss between the original textual embedding $z_{c}$ and $\\widehat{z}_{c}$ . Note that we only train the $\\phi$ module while keeping the textual encoder frozen.", + "bbox": [ + 75, + 463, + 470, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intuition behind SMP is that semantic information is not balanced across the tokens, but concentrated on the specific keywords. We assume that it is more common that adjectives and nouns in the sentence are more important than other part-of-speech (POS), such as verbs or adverbs. We empirically observe that our design choice (replacing all keyword token embeddings with $\\widehat{e}_c$ ) is the best among", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/f65242bbaa7da0c2bfb79f44b7ba62ff792873a750db3cec601c6e4971666dd9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
No noiseStudent-tExpx²N(0,1)Unif(-1,1)Ours
L0.81 (19.8)0.76 (23.1)0.76 (23.5)0.78 (23.5)0.77 (23.7)0.74 (25.1)0.71 (25.5)
H0.63 (31.8)0.59 (32.4)0.67 (28.3)0.64 (27.0)0.60 (32.8)0.59 (33.9)0.53 (34.8)
G0.51 (33.3)0.55 (36.1)0.63 (30.8)0.56 (30.7)0.55 (35.9)0.58 (35.3)0.48 (36.9)
", + "bbox": [ + 501, + 89, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Modality gap vs. distributions. Modality gap [27] (lower denotes less gap) on CC3M and CIRR dev R@1 (higher denotes better performance – in the parentheses) for different noises with different backbone sizes (from ViT-L/14 to ViT-G/14).", + "bbox": [ + 498, + 161, + 890, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the other variants, such as randomly replacing $n$ keyword tokens ( $n = 1,3,5$ ), replacing a random token, replacing all non-keyword tokens, or replacing all noun tokens (See Tab. 8).", + "bbox": [ + 496, + 244, + 890, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SMP has two benefits over the previous image-based ZS-CIR supervision [3, 37]. First, SMP allows the textual encoder to accept more diverse captions rather than \"a photo of $\\left[\\mathbb{S}\\right]$ \". While previous methods risk being sensitive to natural sentence variations, potentially affecting performances, our approach replaces tokens in natural sentences, maintaining robust performance across a more diverse set of sentence constructions. Second, SMP only requires language inputs; therefore, the overall training procedure is efficient regarding the training time and the storage size. It means that LinCIR can easily scale up in terms of the backbone size and the dataset scale. We will discuss the efficiency and the scalability of LinCIR in the Sec. 3.3.", + "bbox": [ + 496, + 306, + 892, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Searching for a better noise distribution for reducing the modality gap.", + "text_level": 1, + "bbox": [ + 498, + 513, + 890, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although SMP enables the language-only training, we still suffer from the modality gap between textual and visual embedding space [27]. Namely, even if the $\\phi$ module works perfectly for language inputs, it can fail to be generalized to visual inputs. To tackle the problem, we employ a simple noise addition strategy following Nukrai et al. [33]: we add a random noise before the projection during training. Nukrai et al. [33] employed a simple Gaussian noise, but we empirically observe that Gaussian noise is not effective in mitigating the gap. Tab. 1 shows the modality gap [27] measured by various CLIP backbones by adding different probability distributions to textual embeddings. In the table, we observe that the careful choice of distribution greatly affects the modality gap and the final performances.", + "bbox": [ + 496, + 551, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also observe that the generally used probabilistic distributions can suffer from a curse of dimensionality in the CLIP embedding space dimensions (e.g., 768-dim). The norm histogram of each probabilistic distribution (Fig. B.2) shows that the samples drawn from a Gaussian distribution have almost identical norm sizes. From this observation, we employ a probability distribution enforcing the diverse norm sizes instead of the Gaussian distribution. We multiply a random scalar value by a random vector drawn from a", + "bbox": [ + 496, + 763, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13228", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da5ecb7babba5693fc13c491934022d9e9799b27316692068159b601a2be93d4.jpg", + "image_caption": [ + "Figure 4. Training time vs. CIR performances. We evaluate three CIR methods with three backbone sizes: ViT-L, ViT-H, and ViT-G. To avoid an unreliable assessment due to the nature of R@1, CIR performances are measured in CIRCO mAP@5 [3], GeneCIS average R@3 [44], FashionIQ Average R@50 [47], and CIRR average R@10 [29]. In all evaluation results, LinCIR achieves the best training time-performance trade-off. Moreover, Pic2Word and SEARLE show degenerated performances when scaling up the backbone size." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 277, + 241 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d8a4550b62f7a70f138cdd6fbd0ada6c648bb72f94f29036c979fbe90202fdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 90, + 475, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/add62770dd1af3bd8c5bbf933849176017c50b0a92d974cf271a37e385a79037.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 676, + 242 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c13d056df8434d878dd904daaa86e118d8bc48d0a8c67b24c20015868d0666ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 683, + 93, + 883, + 242 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d7318f3333f44b366773b3cb79654151c82a12d71d41131e0dbd9356290aa3cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Visual encoderTextual encoder
# layersd# tokensTP# layersd# tokensTP
ViT-L24102425618.51276812.5665.5
ViT-H32128025616.324102412.5635.5
ViT-G48166425611.832128012.5626.5
", + "bbox": [ + 78, + 316, + 468, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Configuration of CLIP visual and textual encoders. Every visual encoder uses the patch size 14 and the input resolution $224 \\times 224$ . The text token length is the average token length of the CC3M [39] captions by the CLIP tokenizer. $d$ denotes the hidden dimension of Transformer blocks, and TP denotes throughput per second (higher means faster). TPs are measured by 1 V100 on CC3M [39] captions and COCO [28] images with FP16 weights.", + "bbox": [ + 75, + 414, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian distribution, i.e., $n \\sim \\mathrm{Unif}(0,1) \\times \\mathcal{N}(0,1)$ . Conceptually, our distribution is a randomized Gaussian distribution with varying variances. In the Appendix, we illustrate that our design choice shows a more diverse norm distribution than the other distributions (See Fig. B.2).", + "bbox": [ + 75, + 541, + 468, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we empirically observe that the noise addition strategy improves the overall CIR performances with a large gap by bridging the modality gap. We also empirically observe that our design choice outperforms other probability distributions with a large gap by diversifying the impact of the random noise (see Tab. 9).", + "bbox": [ + 75, + 619, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Efficiency and scalability", + "text_level": 1, + "bbox": [ + 76, + 724, + 305, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The most remarkable advantages of LinCIR beyond its generalizability are the training efficiency and scalability. First, a text dataset is storage-efficient; the caption storage size of CC3M dataset [39] is only 125 MB, while its image storage size is about 430GB, about 3,400 times larger. Second, the forward complexity of the textual encoder is notably lower than that of the visual encoder. For example, as shown in Tab. 2, the textual encoder has fewer depth, dimension size, and input token size than the visual encoder taking $224 \\times 224$ resolution images. As a result, the average inference", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "time of the CLIP ViT-L visual encoder is $\\times 3.5$ times slower than that of the textual encoder (Tab. 2). Furthermore, the average throughput of the ViT-G textual encoder is even $\\times 1.4$ times faster than the ViT-L visual encoder.", + "bbox": [ + 496, + 320, + 890, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All these advantages make LinCIR easily scalable. Even though we increase the backbone size, the overhead of the textual encoder is not significantly increased. We can train LinCIR with the CLIP ViT-G backbone in 48 minutes using 8 A100 GPUs and 2 hours using a single V100.", + "bbox": [ + 496, + 380, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 470, + 633, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 494, + 714, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use three-layered MLP for the $\\phi$ model: LN [1] - Linear - GeLU [18] - Linear - GeLU - Linear - LN. The intermediate hidden dimension is set to $4d$ ( $d$ for each architecture is shown in Tab. 2). We do not apply the $\\ell_2$ -normalization to the textual encoder outputs during training because, as shown in Fig. B.2, the added random noises have larger norm sizes than 1. If we apply the $\\ell_2$ -normalization, we observe that the $\\phi$ module is not converged. Keywords of the given text are extracted by the POS tagger of spacy library. We use the AdamW optimizer [31] with a fixed learning rate of 0.0001, weight decay of 0.01, and mini-batch size of 512. Dropout with probability $50\\%$ is applied for the regularization. We use CC3M [39] captions and 2.47M number of the curated StableDiffusion prompts² for the training dataset (i.e., there are 5.5M training captions).", + "bbox": [ + 496, + 518, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison between models, we select the model showing the best zero-shot CIRR [29] dev R@1 score for the model selection. We employ an early stopping strategy by monitoring the validation score. We evaluate the CIR performances of the selected model in a zero-shot manner, i.e., one model is evaluated on four benchmarks. We employ the visual and textual encoders of the official", + "bbox": [ + 496, + 760, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://huggingface.co/datasets/FredZhang7/stable-diffusion-prompts-2.47M", + "bbox": [ + 500, + 875, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13229", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/511aa3655e50c8c63e5f072b1b52cff5d495c17bb3a7e3daeaefc2034fcc465b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
mAP@5mAP@10mAP@25mAP@50
ViT-LPic2Word†8.729.5110.6411.29
SEARLE†11.6812.7314.3315.12
LinCIR12.5913.5815.0015.85
ViT-HPic2Word11.6512.3313.7114.43
SEARLE16.0816.9218.8119.69
LinCIR17.6018.5220.4621.39
ViT-GPic2Word5.545.596.687.12
SEARLE13.2013.8515.3216.04
LinCIR19.7121.0123.1324.18
", + "bbox": [ + 78, + 88, + 470, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CLIP ViT-L [35], and OpenCLIP ViT-H and ViT-G [19]. In the Appendix, we show that LinCIR can be easily extended to the other VLMs, such as BLIP [25].", + "bbox": [ + 75, + 339, + 468, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental protocols", + "text_level": 1, + "bbox": [ + 76, + 398, + 294, + 415 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation benchmarks and metrics. As pointed out by Baldrati et al. [3], the existing CIR benchmarks only have a single positive, which can cause an unreliable evaluation. A similar phenomenon is also reported in the image-text cross-modal retrieval problem by Chun et al. [10]; such benchmarks can lead to a wrong model comparison result. For this reason, we use CIRCO as the main benchmark, which has multiple positives and measures a more reliable ranking-based metric, mAP@K [32]. We also report the R@K evaluation results on three additional datasets, GeneCIS [44], FashionIQ [47], and CIRR [29]. We describe the details of each dataset in the Appendix.", + "bbox": [ + 75, + 424, + 468, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this paper, we argue that $\\mathbf{R}@\\mathbf{1}$ results can be somewhat noisy due to the false negatives in the dataset. Note that these benchmarks only have a unique positive triplet for each query $\\langle x_{i_R},x_c,x_i\\rangle$ , i.e., if other plausible images (i.e. false negatives) exist in the gallery set, the $\\mathbf{R}@\\mathbf{1}$ score in these benchmarks cannot correctly measure the actual retrieval performance. Due to this reason, we will focus on the mAP score if it is available. Otherwise, we will concentrate on $\\mathbf{R}@\\mathbf{K}$ with a larger K (e.g., 10) rather than $\\mathbf{R}@\\mathbf{1}$ .", + "bbox": [ + 75, + 607, + 468, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We compare LinCIR with the recent ZS-CIR methods: Pic2Word [37] and SEARLE [3]. For a fair comparison, we train all methods with the same backbone architecture as LinCIR, namely ViT-H and ViT-G CLIP backbones. ViT-L results are measured using the official checkpoints. We did not directly compare our method with recent methods that require massive external triplet datasets or take a long training time [15, 45]. More comparisons with these methods can be found in the Appendix.", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac90ba2290087db6f597b384b31df1092571ba03ac8890378a1ba0ec172da068.jpg", + "table_caption": [ + "Table 3. CIRCO results. Results of Pic2Word [37], SEARLE [3], LinCIR by using different CLIP backbones are shown. $\\dagger$ denotes that the numbers are measured by the official checkpoint." + ], + "table_footnote": [], + "table_body": "
R@1R@2R@3
ViT-LPic2Word†11.1621.4730.38
SEARLE†12.2622.1131.30
LinCIR12.1922.7632.38
ViT-HPic2Word11.8922.1731.94
SEARLE13.3423.7232.72
LinCIR13.7623.8733.16
ViT-GPic2Word10.6720.7029.50
SEARLE12.8722.6132.46
LinCIR13.6624.6433.54
", + "bbox": [ + 545, + 88, + 848, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. GeneCIS results. The average R@1, R@2, R@3 for \"Focus Attribute\", \"Change Attribute\", \"Focus Object\", and \"Change Object\" are shown. The full table is in the Appendix.", + "bbox": [ + 498, + 267, + 890, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main results", + "text_level": 1, + "bbox": [ + 500, + 330, + 633, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The experimental results are summarized in Fig. 4: LinCIR outperforms the comparison methods in training time and retrieval performances. In all benchmarks, we observe that while the performance of LinCIR is enhanced by enlarging the backbone size, Pic2Word and SEARLE show inferior performances with the ViT-G backbone. We presume that it is because Pic2Word and SEARLE have a limited understanding of complex text queries because their $\\phi$ module are trained on texts not diverse enough (i.e., a photo of $[\\S])$ . On the other hand, our $\\phi$ module shows a better understanding of complex texts as LinCIR is trained on diverse real-world texts from the caption datasets.", + "bbox": [ + 496, + 354, + 890, + 535 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also provide the full evaluation results on the four benchmarks below. Tab. 3 shows the evaluation results on the CIRCO dataset. In all experiments, LinCIR outperforms others with a significant gap. We can observe a similar finding in the GeneCIS average R@K results for four different subtasks (Tab. 4), especially for R@K with a larger K. As shown in the Appendix, LinCIR outperforms other models, especially on \"Focus Attribute\" and \"Change Attribute\" tasks. We presume that it is because the Pic2Word and SEARLE training prompts are more specialized to objects (a photo of $[\\\\])$ , where LinCIR can handle more detailed concepts in the image by altering all the nouns in the sentence (e.g., \"gray $[\\\\] sleeps on a \\[\\\\]$ \").", + "bbox": [ + 496, + 537, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the FashionIQ benchmark (Tab. 5), LinCIR even outperforms the state-of-the-art supervised method [2] with a large gap (38.32 vs. 45.11 in the average R@10). Our work is the first ZS-CIR method that outperforms the supervised CIR method despite its generalizability to the other CIR benchmarks and flexibility for handling various conditions.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We observe somewhat mixed results on CIRR (Tab. 6): LinCIR achieves the best R@10 score, but not in some other metrics. We argue that this is because of two reasons. First, R@K with a small K cannot fully reflect the authentic performance. As shown in the Appendix, the retrieval results", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13230", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5341db20759a60ad69d0a3ff7145e2a7497b58fc2abe4dcd73f77ea05b531aeb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ShirtDressTopteeAverage
R@10R@50R@10R@50R@10R@50R@10R@50
ViT-LPic2Word†26.2043.6020.0040.2027.9047.4024.7043.70
SEARLE†26.8945.5820.4843.1329.3249.9725.5646.23
LinCIR29.1046.8120.9242.4428.8150.1826.2846.49
ViT-HPic2Word36.9055.9928.0151.5140.1862.0135.0356.50
SEARLE36.4655.4528.4651.0738.8160.8934.5755.80
LinCIR36.9057.7529.8052.1142.0762.5236.2657.46
ViT-GPic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
Combiner (supervised) [2]†39.9960.4533.8159.4041.4165.3738.3261.74
", + "bbox": [ + 133, + 88, + 839, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1f2ef51d387bcaff6050774d721625f7a21518b618c1496d5f80b4ecc2a749a1.jpg", + "table_caption": [ + "Table 5. FashionIQ results. ${}^{ \\dagger }$ denotes that the numbers are from the original paper. LinCIR ViT-G even outperforms the previous state- of-the-art supervised CIR method [2] with a large gap although LinCIR is not directly trained on the FashionIQ dataset." + ], + "table_footnote": [], + "table_body": "
FullSubset
R@1R@5R@10R@1R@2R@3
ViT-LPic2Word†23.9051.7065.3053.7674.4687.08
SEARLE†24.2452.4866.2953.7675.0188.19
LinCIR25.0453.2566.6857.1177.3788.89
ViT-HPic2Word32.9463.1173.8662.2281.3591.23
SEARLE34.0063.9875.2564.6383.2192.77
LinCIR33.8363.5275.3562.4381.4792.12
ViT-GPic2Word30.4158.1269.2368.9285.4593.04
SEARLE34.8064.0775.1168.7284.7093.23
LinCIR35.2564.7276.0563.3582.2291.98
", + "bbox": [ + 78, + 354, + 475, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of LinCIR are plausible to humans, but because the dataset has incomplete positives (i.e., there are many false negatives), the R@1 score cannot correctly evaluate the model performance. According to Chun et al. [10], the similarity between the rankings measured by R@K on a partially annotated benchmark and those measured by mAP on a fully annotated benchmark becomes lower when we use a small K (e.g., 1). Second, as observed by previous works [3, 37], the quality of the CIRR benchmark as a ZS-CIR benchmark is somewhat doubtful. The CIRR text relative captions are often not truly relative (i.e., there exist false positives), and reference images can even be harmful to retrieval. As pointed out by Baldrati et al. [3], this problem becomes more severe when we use a small subset of images, i.e., the subset R@Ks. In summary, due to the noisy nature of CIRR full and subset R@1s, we propose to focus on the CIRR full R@10 scores rather than other metrics. In CIRR full R@10, LinCIR shows the same trends as the other benchmarks.", + "bbox": [ + 75, + 628, + 472, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e88e9311f7bba67c3bd3bc0262ed56018e55e26e98be6fec0a1fd793ad02e98.jpg", + "table_caption": [ + "Table 6. CIRR results. Due to the noisy nature of CIRR as a ZS-CIR benchmark, we only highlight the R@10 score for the full CIRR set. The detailed discussion can be found in Sec. 4.3" + ], + "table_footnote": [], + "table_body": "
Supervision designCIRR dev R@1Fashion IQ R@10R@50
a photo of [$][37]21.6524.9344.35
Ours, but [$] extracted by ψI22.6322.0139.87
Our SMP design choice25.6626.2846.49
", + "bbox": [ + 500, + 354, + 895, + 440 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4beb56c253f1d10efdb859386335db10e22b65b20c1167e232074f0cf9d1108c.jpg", + "table_caption": [ + "Table 7. Impact of the supervision design. Different target text designs (e.g., \\([\\\\) ]$ sleeps on \\([\\\\( ]\\) in Fig. 3 (b)) affect the performances. \"Ours, but \\([\\\\) ]$ extracted by \\( \\psi_I \\)\" denotes that the textual encoder before the \\( \\phi \\) module is replaced with the visual encoder." + ], + "table_footnote": [], + "table_body": "
SMP Masking strategyCIRR devFashion IQ
R@1R@10R@50
All non-keyword tokens20.5919.6136.97
Random token22.9822.4440.92
All noun tokens24.9525.1645.19
1 keyword token23.9226.4245.91
3 keyword tokens24.6626.6946.54
5 keyword tokens25.1426.2846.29
All keyword tokens25.6626.2846.49
", + "bbox": [ + 516, + 518, + 877, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Impact of the SMP masking strategy. The CIRR retrieval performances by varying the masking strategy for SMP (i.e., \"Keyword masking\" in Fig. 3 (b)) are shown. We define \"keyword\" as consecutive adjectives and nouns, except for \"All noun tokens\". \"All noun tokens\" defines keywords as nouns.", + "bbox": [ + 498, + 684, + 893, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 500, + 773, + 602, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we provide detailed analyses of our design choices. If not specified, we compare the models on the CIRR dev split and FashionIQ test split.", + "bbox": [ + 498, + 797, + 892, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Self-Masking Projection (SMP) supervision. We compare two variants of SMP in Tab. 7. First, instead of employing our keyword masking strategy, we use “a photo", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13231", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d921390bdce3ee57c1dc7757157728a7562b15dfb66572c9be9bdb7fdc2b9e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Noise typeCIRR devFashion IQ
R@1R@10R@50
No noise19.7620.4238.31
Student-t23.0822.8141.00
Exponential23.5125.7845.26
\\( \\chi^2 \\)23.5423.1141.89
N(0,1)23.7023.3141.89
Unif(-1,1)25.1425.8845.78
N(0,1)×Unif(0,1)25.4726.0546.29
", + "bbox": [ + 102, + 88, + 442, + 242 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95489e1d55d5f5beecd0df999d920b24ab654ba2bccebd7ef4e48e5f0ff61d24.jpg", + "table_caption": [ + "Table 9. Impact of the choice of the random noise. Adding noise to the textual latent space helps to mitigate the inferior generalizability due to the modality gap. Moreover, using better random noise can significantly boost the overall performances." + ], + "table_footnote": [], + "table_body": "
CC3MSDPCOYOOWTCIRCO mAP@5GeneCIS R@3FashionIQ R@10CIRR R@10Avg
✓XXX13.7232.8025.1164.9534.15
X✓XX9.5232.3823.6362.5832.03
XX✓X11.3631.4826.1865.3333.59
XXX✓9.6730.9024.4164.0532.26
✓✓XX12.5932.3826.2866.6834.48
✓✓X✓10.0633.3621.1163.1631.92
✓✓✓X11.5432.0826.9766.8034.35
", + "bbox": [ + 81, + 327, + 465, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Impact of the training corpus. OpenWebText (OWT) [14] and SD prompts (SDP) are text-only datasets and CC3M [39] and COYO-700M (COYO) [5] are image-text aligned datasets. In our experiments, we use CC3M + SDP for the training corpus, considering the dataset scale and overall CIR performances.", + "bbox": [ + 75, + 458, + 467, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of \\([\\\\) ]\\( as Pic2Word and SEARLE to train the \\(\\phi\\) module. Second, we replace the textual encoder before \\(\\phi\\) module with the visual encoder using the corresponding image of the caption. In the table, \"a photo of \\([\\\\) ]\\( that [cond]\" variant performs worse than our design choice due to the limited diversity of the input texts. Interestingly, we observe that using both image and text pairs for LinCIR performs worse than our design choice. Note that the second model is trained on CC3M image-text pairs without using 2.47M SD prompts as our design choice. We presume it makes the \\(\\phi\\) model overfitted to CC3M image-text relationships, which significantly differ from our target CIR datasets.", + "bbox": [ + 75, + 559, + 467, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of masking design choice. We compare other masking design choices, such as random tokens or non-keyword tokens, with our design choice in Tab. 8. Tab. 8 shows that (1) masking the keywords performs better than masking the others. (2) Our keyword design – i.e., consecutive adjectives and nouns – is better than defining keywords as nouns. (3) The overall performances are enhanced by increasing the masked keywords. As the differences are not significant, we mask all the keywords for simplicity.", + "bbox": [ + 75, + 763, + 467, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the random noise addition. As we discussed in Sec. 3.2, the random noise addition is critical to mitigating the modality gap. Tab. 9 supports this claim: when we do not add any noise, the performance becomes the worst. Our design choice shows the best performance among the other noise designs due to the diverse norm size of our distribution as shown in Fig. B.2.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The impact of the training corpus. We evaluate the impact of the training corpus in Tab. 10. We use four corpora, CC3M [39] (3M captions), StableDiffusion Prompts (SDP) (2.47M text prompts), COYO-700M (700M captions) [5] and OpenText (OWT) [14] (8M web texts). CC3M and COYO are image-text paired datasets, and OWT and SDP are text-only datasets. The example training samples of each dataset are shown in the Appendix. In Tab. 10, we observe that models trained with image descriptions (CC3M and COYO) perform better than models trained with general texts because general web texts are often irrelevant to visual information. In addition, using multiple corpora improves the overall performance, e.g., CC3M $(34.15)\\rightarrow$ CC3M $^+$ SDP (34.48). Although we observe that using more massive captions (e.g., $3\\mathrm{M} + 2.47\\mathrm{M} + 700\\mathrm{M})$ can be helpful for some CIR tasks, such as FashionIQ and CIRR, we use CC3M and SDP for our training set considering the corpus size $(3\\mathrm{M} + 2.47\\mathrm{M})$ , and the overall performances.", + "bbox": [ + 496, + 208, + 890, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results. We provide the additional qualitative retrieval results in the Appendix. In summary, LinCIR shows qualitatively plausible retrieval performance even in a large-scale image database, e.g. LAION-2B. Also, we observe that the retrieved results by LinCIR often suffer from false negatives in benchmark datasets, showing the limitation of the R@1 evaluation on the existing CIR benchmarks.", + "bbox": [ + 496, + 500, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 617, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel zero-shot composed image retrieval (ZS-CIR) framework named Language Only training for Composed Image Retrieval (LinCIR). LinCIR presents a breakthrough in addressing the challenges associated with the previous ZS-CIR methods. By leveraging a novel self-supervision technique, Self-Masking Projection (SMP), LinCIR eliminates the dependency on expensive CIR triplets, opting for a training process solely based on text inputs. This innovative approach significantly enhances scalability and generalizability, overcoming limitations observed in existing ZS-CIR methods. Notably, our LinCIR model achieves remarkable efficiency and ZS-CIR performances compared to other methods on multiple CIR benchmarks, including CIRCO, GeneCIS, FashionIQ, and CIRR. We underscore the effectiveness of our language-only training framework, offering a potent solution with wide-ranging implications for image retrieval tasks.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13232", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5", + "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, pages 4959-4968, 2022. 1, 2, 6, 7", + "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 11, 13", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3", + "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.8, 11", + "[6] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. In ECCV, pages 136-152. Springer, 2020. 1", + "[7] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In CVPR, pages 3001-3011, 2020.", + "[8] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and TatSeng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization. arXiv preprint arXiv:2211.07394, 2022. 1", + "[9] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 11", + "[10] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image_caption associations for ms-coco. In ECCV, 2022. 6, 7, 11", + "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1", + "[12] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback. arXiv preprint arXiv:2007.00145, 2020. 1", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 13", + "[14] Aaron Gokaslan, Vanya Cohen, Ellie Pavlick, and Stefanie Tellex. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019. 8, 11", + "[15] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 6, 13", + "[16] Sophia Gu, Christopher Clark, and Aniruddha Kembhavi. I can't believe there's no images! learning visual tasks us" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing only language supervision. In ICCV, pages 2672-2683, 2023. 3", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 11", + "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 6", + "[20] Surgan Jandial, Ayush Chopra, Pinkesh Badjatiya, Pranit Chawla, Mausoom Sarkar, and Balaji Krishnamurthy. Trace: Transform aggregate and compose visiolinguistic representations for image search with text feedback. arXiv preprint arXiv:2009.01485, 7, 2020. 1", + "[21] Surgan Jandial, Pinkesh Badjatiya, Pranit Chawla, Ayush Chopra, Mausoom Sarkar, and Balaji Krishnamurthy. Sac: Semantic attention composition for text-conditioned image retrieval. In WACV, pages 4021-4030, 2022.", + "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 1", + "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 11", + "[24] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, pages 802-812, 2021. 1", + "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 2, 6, 13, 14", + "[26] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding CLIP latents for zero-shot captioning via text-only training. In ICLR, 2023. 3", + "[27] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 35:17612-17625, 2022. 2, 3, 4", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 11", + "[29] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1, 2, 5, 6, 11", + "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13233", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5", + "[32] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6, 11", + "[33] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022. 2, 3, 4", + "[34] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, pages 13018-13028, 2021. 11", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6", + "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 11", + "[37] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 4, 6, 7, 11, 13", + "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 1", + "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 2, 5, 8, 11", + "[40] Minchul Shin, Yoonjae Cho, Byungsoo Ko, and Geonmo Gu. Rtic: Residual learning for text and image composition using graph convolutional network. arXiv preprint arXiv:2104.03015, 2021. 1", + "[41] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 11", + "[42] Yuxin Tian, Shawn Newsam, and Kofi Boakye. Image search with text feedback by additive attention compositional learning. arXiv preprint arXiv:2203.03809, 2022. 1", + "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 13", + "[44] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5, 6, 11" + ], + "bbox": [ + 78, + 90, + 467, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 6, 13", + "[46] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1", + "[47] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 2, 5, 6, 11, 13", + "[48] Youngjae Yu, Seunghwan Lee, Yuncheol Choi, and Gunhee Kim. Curlingnet: Compositional learning between images and text for fashion iq data. arXiv preprint arXiv:2003.12299, 2020. 1" + ], + "bbox": [ + 501, + 92, + 890, + 316 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13234", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_model.json b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c307eadc11f7ce724ae2c0d97c97f5913e717b28 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.131, + 0.857, + 0.153 + ], + "angle": 0, + "content": "Language-only Efficient Training of Zero-shot Composed Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.18, + 0.842, + 0.234 + ], + "angle": 0, + "content": "Geonmo Gu*,1 Sanghyuk Chun*,2 Wonjae Kim² Yoohoon Kang¹ Sangwoo Yun² \n1NAVER Vision 2NAVER AI Lab * Equal contribution" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LinCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LinCIR is surprisingly efficient and highly effective; LinCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ. Code is available at github.com/navervision/lincir" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Composed image retrieval (CIR) is a challenging vision-language (VL) task that takes a composed query of image and text, aiming to search relative images for both conditions [29]. As language serves as the most natural method for encoding human interaction, CIR provides a higher degree of freedom and a better user experience for image-based search engine applications, such as web commerce." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "One of the main challenges of CIR is the expensive dataset collection pipeline. CIR datasets consist of triplets \\(\\langle x_{i_R},x_c,x_i\\rangle\\), where \\(x_{i_R}\\) is an image query, \\(x_{c}\\) is a text query, and \\(x_{i}\\) is the target image. Unlike image-text paired datasets, such as CC3M [39] or LAION [38], such triplets" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.276, + 0.859, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.627 + ], + "angle": 0, + "content": "Figure 1. Training time (hours) vs. Zero-shot Composed Image Retrieval (ZS-CIR) performance. Thanks to our efficient language-only training strategy, our LinCIR outperforms the previous ZS-CIR methods in both training time and CIR performance. The training time is measured on 8 A100 GPUs. We compare the models on the CIRCO mAP@5 [3] score for a more comprehensive evaluation of CIR models (more results are in Fig. 4). Notably, when we scale up the backbone CLIP [19, 35] model size by ViT-L, ViT-H and ViT-G, LinCIR shows a promising performance boost with surprisingly short training time (48 mins for ViT-G). On the other hand, Pic2Word [37] and SEARLE [3] cannot be scaled up to CLIP ViT-G due to their limitation on restricted textual expressions and the lack of diversity of input texts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.78 + ], + "angle": 0, + "content": "are almost impossible to collect by web crawling and require expensive human labor to create each triplet. For example, the datasets are constructed by gathering candidates of \\(\\langle x_{i_R},x_i\\rangle\\) and manually annotating \\(x_{c}\\) by human annotators [29, 47], which is hard to scalable. Therefore, the size of the training triplets is usually small (e.g., 46.6k [47], 28.8k [29]), and the existing CIR methods trained on such triplets [2, 6-8, 11, 12, 20-22, 24, 30, 40, 42, 46, 48] suffer from the lack of generalizability to diverse unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To overcome the drawback, recent studies have explored zero-shot CIR (ZS-CIR), a scalable direction, by eliminating the dependency on the pre-collected triplet datasets. For example, Saito et al. [37] and Baldrati et al. [3] propose projection-based ZS-CIR methods without using triplet datasets. Based on the pre-trained CLIP [35], they train a lightweight projection module \\(\\phi\\) that projects the CLIP image latent embedding \\(z_{i}\\) to the CLIP text token" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13225" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.088, + 0.805, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.213, + 0.895, + 0.258 + ], + "angle": 0, + "content": "Figure 2. Overview of ZS-CIR with a projection to the token embedding space. The mainstream ZS-CIR methods, such as Pic2Word [37], SEARLE [3] and LinCIR (ours), train a projection module \\(\\phi\\) that projects the image latent embedding \\(z_{i}\\) into the token embedding space \\(e_c\\) with a custom prompt (e.g., a photo of \\([\\mathbb{S}]\\) that [cond]). The textual encoder output is used for CIR." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.275, + 0.47, + 0.396 + ], + "angle": 0, + "content": "embedding space \\( e_{c} \\) (see Fig. 2). These approaches have shown promising generalizability to unseen datasets. However, these methods struggle to handle diverse text conditions because they rely on pre-defined naïve text prompts during training (e.g., a photo of [§]). Moreover, their training frameworks need sequential forward operations of the visual and textual encoders (see Fig. 3(a)), resulting in inefficient training and limited scalability to a larger backbone." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.397, + 0.474, + 0.789 + ], + "angle": 0, + "content": "In this paper, we introduce a new paradigm of ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). As shown in Fig. 3 (b), instead of projecting the image latent embedding \\( z_{i} \\), we propose to project the text latent embedding \\( z_{c} \\) to the token embedding space \\( e_{c} \\). We introduce a novel self-supervision, named Self-Masking Projection (SMP), for language-only training. We replace all the \"keywords\" of the original text with the projected text embedding of the original text to produce an embedding \\( \\widehat{z}_{c} \\) and apply MSE loss between \\( z_{c} \\) and \\( \\widehat{z}_{c} \\). Here, we define \"keyword\" as consecutive adjectives and nouns. For example, the keywords of \"gray cat sleeps on a pillow\" are \"gray cat\" and \"a pillow\"; therefore, it becomes \"[\\$] sleeps on [\\$]\". The purpose of SMP is to make [\\$] token interpreted as the \"one-word summarization\" of the input by extracting the essential information of the input. During inference, we simply perform text-to-image retrieval by projecting image embedding to the token embedding space using the projection module \\( \\phi \\) as shown in Fig. 2. However, this strategy can suffer from the modality gap between textual and visual modalities [27], i.e., even though our \\( \\phi \\) module works perfectly for text latent embeddings, it can underperform for the target visual embeddings. We mitigate the issue by employing a random noise addition strategy [33], carefully choosing a probability distribution that ensures the diversity of the noise-augmented textual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.789, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Our paradigm has three advantages over the previous approaches. First, while Pic2Word and SEARLE projection modules are trained with a restricted text prompt, i.e., a photo of \\(\\left[\\S\\right]\\), the projection module of LinCIR is trained with the diverse text inputs from the actual texts. Due to this reason, Pic2Word and SEARLE show degenerated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.895, + 0.562 + ], + "angle": 0, + "content": "performances when the backbone size becomes larger (see Fig. 1). On the other hand, LinCIR is more generalizable to complex and diverse text conditions, showing superior ZS-CIR performances than others, especially for larger backbones. Second, as LinCIR only utilizes the textual encoder, our training process is highly efficient and scalable than the methods incorporated with the visual encoder; our language-only training strategy is \\(\\times 6.0\\) faster than Pic2Word [37] and \\(\\times 8.4\\) faster than SEARLE [3] with CLIP ViT-L backbone. When we scale up the backbone size to CLIP ViT-G, the gap becomes \\(\\times 16.4\\) and \\(\\times 17.6\\), respectively. Even ViT-G training for LinCIR only takes 48 minutes using 8 A100 and less than 2 hours using 1 V100. Third, our method is storage-efficient; for example, the CC3M dataset [39] images occupy about 430GB storage size, while its captions only need 125MB. We train LinCIR only with 571MB storage size for storing the 5.5M training captions. In summary, LinCIR shows the best training time and ZS-CIR performance as shown in Fig. 1 and 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.895, + 0.73 + ], + "angle": 0, + "content": "Our contribution can be summarized as follows: (1) We propose LinCIR, a novel and efficient language-only training framework for ZS-CIR. (2) We introduce a new self-supervision for language-only training, named SelfMasking Projection (SMP). (3) We employ a better random noise addition strategy than naive Gaussian noise to mitigate the modality gap. (4) LinCIR achieves the best training time and the ZS-CIR performances on four ZS-CIR benchmarks (CIRCO [3], GeneCIS [44], FashionIQ [47] and CIRR [29]). Notably, LinCIR even outperforms the state-of-the-art supervised method [2] on FashionIQ." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.64, + 0.756 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Vision-language models (VLM). As with previous ZS-CIR methods, LinCIR utilizes a pre-trained VLM, such as CLIP [19, 35] or BLIP [25]. We use VLMs that map an image input \\( x_{i} \\) and a text input \\( x_{c} \\) to the \\( d \\)-dimensional joint embedding space. Here, a given caption \\( x_{c} \\) is tokenized by the pre-defined tokenizer as \\( t_c = \\{t_c^k\\mid k = 1\\dots K\\} \\), where \\( K \\) is the number of tokens, and mapped to token embeddings \\( e_{c} = E_{w}(t_{c}) = \\{E_{w}(t_{c}^{k})\\mid k = 1\\dots K\\} \\), where \\( E_{w} \\) is the embedding layer parameterized by \\( w \\). Afterwords, a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.442, + 0.902 + ], + "angle": 0, + "content": "1 Pronounced as \"linker\", meaning for linking the two modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13226" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.092, + 0.422, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.22, + 0.356, + 0.233 + ], + "angle": 0, + "content": "(a) Pic2Word training procedure" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.849, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.22, + 0.792, + 0.233 + ], + "angle": 0, + "content": "(b) Our language-only training procedure" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.894, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Comparison of Pic2Word [37] and LinCIR training procedures. (a) Pic2Word [37] and SEARLE [3] training procedure requires both the visual encoder and the textual encoder. They only need images for training, while the text prompt is pre-defined [37] or automatically generated [3]. (b) LinCIR is trained solely on texts with the frozen textual encoder. First, a projection module \\(\\phi\\) projects a textual latent embedding of a sentence \\(z_{t}\\) into the token embedding space. Before the projection, a random noise \\(n\\) is added to \\(z_{t}\\) to reduce the modality gap between text and image. We introduce a new self-supervision, named Self-Masking Projection (SMP), by replacing all keywords of the given caption with the projected embedding by \\(\\phi\\) and extracting a modified text embedding \\(\\widehat{z}_{t}\\). Finally, the projection module \\(\\phi\\) is trained by the MSE loss between \\(z_{t}\\) and \\(\\widehat{z}_{t}\\). Note that both (a) and (b) use the same inference strategy shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.448 + ], + "angle": 0, + "content": "textual encoder \\(\\psi_C\\) encodes the token embeddings to extract textual latent embeddings \\(z_{c} = \\psi_{C}(x_{c}) = \\psi_{C}(E_{w}(t_{c}))\\). The visual latent embeddings are extracted by the visual encoder \\(\\psi_I\\): \\(z_{i} = \\psi_{I}(x_{i})\\). We will use the terminology token embeddings to represent \\(e_c\\), and use textual (or visual) latent embeddings to represent \\(z_{c}\\) (or \\(z_{i}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.538 + ], + "angle": 0, + "content": "VLM modality gap. The VLM joint embedding space suffers from the semantic gap between each modality. Namely, the visual and the textual latent embeddings are not exactly aligned with each other, but they are located in completely separate regions of the embedding space [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Such a modality gap hinders the harmonization of visual and textual latents in the joint embedding space. To mitigate the gap, language-only training has been recently introduced for image captioning tasks [16, 26, 33]. They train a text decoder from the CLIP textual embeddings to generate image captions from visual embeddings. During training the decoder, the modality gap is bridged by injecting Gaussian noises [16, 33] or projection-based method [26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.841 + ], + "angle": 0, + "content": "In this paper, we utilize language-only training and a noise-addition strategy to mitigate the modality gap. We carefully studied the impact of noise and found a better distribution rather than simple Gaussian noise. We believe our findings can be transferred to other language-only training methods. Moreover, our target task, CIR, needs to model a relationship between triplets of \\(\\langle x_{i_R},x_c,x_i\\rangle\\), whereas previous language-only training methods only consider the pair-wise relationship (i.e., image captioning). We tackle this problem by proposing a novel self-supervision, named Self-Masking Projection (SMP), which injects the projected textual embeddings into the original token embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "CIR by projection to token embeddings. The mainstream ZS-CIR methods, such as Pic2Word [37] and SEARLE [3], employ a projection-based method. Namely," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.493 + ], + "angle": 0, + "content": "they learn a projection module \\(\\phi\\) from the image latent embedding space to the token embedding space. For inference, they project the image latent embedding \\(z_{i}\\) to the token embedding (\\$), then perform text-to-image retrieval with the prompt \"a photo of [\\\\(] that [cond]\", where [cond] is a text condition. Fig. 2 shows an overview of how textual projection-based ZS-CIR works. The main research point of this field is how to train the projection module \\(\\phi\\) to capture the visual information into the token embedding space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Pic2Word [37] trains the projection module \\(\\phi\\) by minimizing contrastive loss between image latent embedding and the textual latent embedding of \"a photo of [\\$]\" (see Fig. 3 (a)). SEARLE [3] employs a similar approach to Pic2Word. First, they employ optimization-based textual inversion to generate pre-defined special tokens for an image and train \\(\\phi\\) to predict the token embedding. SEARLE employs CLIP zero-shot classification to predict the \"concept\" of the given image and refine the prompt by letting GPT [4] continue the phrase. Both methods only use image inputs \\(x_{i}\\) for training without accessing the CIR triplets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Although Pic2Word and SEARLE achieve reasonable ZS-CIR performances, they have two significant problems. First, they heavily rely on the initial prompt, \"a photo of \\([\\mathbb{S}]\\)\", limiting the diversity of the textual encoder input. As Fig. 1, we argue that diversifying the input texts during training the projection module \\( \\phi \\) is critical to train with larger backbones (e.g., CLIP ViT-G), while using the naive prompt is failed to scale up the backbone size. Second, they need image inputs, which are less compact and redundant than text datasets. Moreover, the visual encoder usually needs more computation resources than the textual encoder because the visual encoder takes the fixed length token (e.g., 256). In contrast, the textual encoder takes shorter token lengths (e.g., average token length of CC3M ≈ 10). We tackle the problems by introducing a language-only training method, showing remarkable efficiency and scalability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13227" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.453, + 0.108 + ], + "angle": 0, + "content": "3. Language-only Training of Zero-shot CIR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.252 + ], + "angle": 0, + "content": "This section introduces a new paradigm for ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). We first introduce a novel language-only self-supervision, Self-Masking Projection (SMP), that enables a language-only training for CIR (\\(\\S 3.1\\)). Then, we explain the modality gap problem and empirically show that adding a carefully chosen random noise can mitigate the problem (\\(\\S 3.2\\)). Finally, we describe the advantage of LinCIR in terms of efficiency and scalability (\\(\\S 3.3\\))." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.259, + 0.355, + 0.275 + ], + "angle": 0, + "content": "3.1. Self-Masking Projection (SMP)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.282, + 0.47, + 0.463 + ], + "angle": 0, + "content": "We aim to learn a projection module \\(\\phi\\) that captures and retains the original visual information after the projection and the textual encoder. While the previous methods focus on directly mapping visual information to the token space with a naive prompt (i.e., a photo of \\([\\S])\\), we argue that focusing on the textual encoder is more important. Our zero-shot CIR is based on text-to-image retrieval (see Fig. 2). It means that the quality of the textual latent embedding is more critical to the final ZS-CIR performances. Hence, rather than focusing on minimizing the gap between visual information and the naive prompt, we aim to achieve a projection module \\(\\phi\\) that captures the semantics of the keywords in the given text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.464, + 0.471, + 0.795 + ], + "angle": 0, + "content": "To achieve our goal, we introduce a novel language-only self-supervision named Self-Masking Projection (SMP). First, we project the textual embedding \\( z_{c} \\) of a given text input \\( x_{c} \\) with the projection module \\( \\phi \\) to the token embedding space, i.e., \\( \\widehat{e}_{c} = \\phi(z_{c}) \\), where \\( \\widehat{e}_{c} \\) is the projected textual embedding. Then, we replace the token embeddings of all the \"keywords\" of \\( x_{c} \\) with the projected token embedding \\( \\widehat{e}_{c} \\). We define the keywords of the sentence as consecutive nouns and adjectives. For example, the keywords of \"A Russian Blue cat is gray and cute\" will be \"A Russian Blue cat\", \"gray\" and \"cute\"; hence it will be converted to \"[$] is [\\$] and [\\$]\", where [\\$] is a special token to represent the projected token embedding \\( \\widehat{e}_{c} \\). By treating all the main concepts (keywords) in the caption as the same [\\$], we intend [\\$] to represent the overall essential information of the inputs. Note that similar to [MASK] tokens of masked modeling, [\\$] with different positions will be encoded in different features due to the positional embeddings. Using the converted caption, we extract a converted textual latent feature \\( \\widehat{z}_{c} \\) and minimize MSE loss between the original textual embedding \\( z_{c} \\) and \\( \\widehat{z}_{c} \\). Note that we only train the \\( \\phi \\) module while keeping the textual encoder frozen." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The intuition behind SMP is that semantic information is not balanced across the tokens, but concentrated on the specific keywords. We assume that it is more common that adjectives and nouns in the sentence are more important than other part-of-speech (POS), such as verbs or adverbs. We empirically observe that our design choice (replacing all keyword token embeddings with \\(\\widehat{e}_c\\)) is the best among" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.09, + 0.891, + 0.152 + ], + "angle": 0, + "content": "
No noiseStudent-tExpx²N(0,1)Unif(-1,1)Ours
L0.81 (19.8)0.76 (23.1)0.76 (23.5)0.78 (23.5)0.77 (23.7)0.74 (25.1)0.71 (25.5)
H0.63 (31.8)0.59 (32.4)0.67 (28.3)0.64 (27.0)0.60 (32.8)0.59 (33.9)0.53 (34.8)
G0.51 (33.3)0.55 (36.1)0.63 (30.8)0.56 (30.7)0.55 (35.9)0.58 (35.3)0.48 (36.9)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.162, + 0.892, + 0.218 + ], + "angle": 0, + "content": "Table 1. Modality gap vs. distributions. Modality gap [27] (lower denotes less gap) on CC3M and CIRR dev R@1 (higher denotes better performance – in the parentheses) for different noises with different backbone sizes (from ViT-L/14 to ViT-G/14)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.305 + ], + "angle": 0, + "content": "the other variants, such as randomly replacing \\( n \\) keyword tokens (\\( n = 1,3,5 \\)), replacing a random token, replacing all non-keyword tokens, or replacing all noun tokens (See Tab. 8)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.893, + 0.503 + ], + "angle": 0, + "content": "SMP has two benefits over the previous image-based ZS-CIR supervision [3, 37]. First, SMP allows the textual encoder to accept more diverse captions rather than \"a photo of \\(\\left[\\mathbb{S}\\right]\\)\". While previous methods risk being sensitive to natural sentence variations, potentially affecting performances, our approach replaces tokens in natural sentences, maintaining robust performance across a more diverse set of sentence constructions. Second, SMP only requires language inputs; therefore, the overall training procedure is efficient regarding the training time and the storage size. It means that LinCIR can easily scale up in terms of the backbone size and the dataset scale. We will discuss the efficiency and the scalability of LinCIR in the Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.892, + 0.546 + ], + "angle": 0, + "content": "3.2. Searching for a better noise distribution for reducing the modality gap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Although SMP enables the language-only training, we still suffer from the modality gap between textual and visual embedding space [27]. Namely, even if the \\(\\phi\\) module works perfectly for language inputs, it can fail to be generalized to visual inputs. To tackle the problem, we employ a simple noise addition strategy following Nukrai et al. [33]: we add a random noise before the projection during training. Nukrai et al. [33] employed a simple Gaussian noise, but we empirically observe that Gaussian noise is not effective in mitigating the gap. Tab. 1 shows the modality gap [27] measured by various CLIP backbones by adding different probability distributions to textual embeddings. In the table, we observe that the careful choice of distribution greatly affects the modality gap and the final performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We also observe that the generally used probabilistic distributions can suffer from a curse of dimensionality in the CLIP embedding space dimensions (e.g., 768-dim). The norm histogram of each probabilistic distribution (Fig. B.2) shows that the samples drawn from a Gaussian distribution have almost identical norm sizes. From this observation, we employ a probability distribution enforcing the diverse norm sizes instead of the Gaussian distribution. We multiply a random scalar value by a random vector drawn from a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13228" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.278, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.091, + 0.477, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.092, + 0.677, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.684, + 0.094, + 0.885, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Training time vs. CIR performances. We evaluate three CIR methods with three backbone sizes: ViT-L, ViT-H, and ViT-G. To avoid an unreliable assessment due to the nature of R@1, CIR performances are measured in CIRCO mAP@5 [3], GeneCIS average R@3 [44], FashionIQ Average R@50 [47], and CIRR average R@10 [29]. In all evaluation results, LinCIR achieves the best training time-performance trade-off. Moreover, Pic2Word and SEARLE show degenerated performances when scaling up the backbone size." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.318, + 0.469, + 0.404 + ], + "angle": 0, + "content": "
Visual encoderTextual encoder
# layersd# tokensTP# layersd# tokensTP
ViT-L24102425618.51276812.5665.5
ViT-H32128025616.324102412.5635.5
ViT-G48166425611.832128012.5626.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.415, + 0.47, + 0.513 + ], + "angle": 0, + "content": "Table 2. Configuration of CLIP visual and textual encoders. Every visual encoder uses the patch size 14 and the input resolution \\(224 \\times 224\\). The text token length is the average token length of the CC3M [39] captions by the CLIP tokenizer. \\(d\\) denotes the hidden dimension of Transformer blocks, and TP denotes throughput per second (higher means faster). TPs are measured by 1 V100 on CC3M [39] captions and COCO [28] images with FP16 weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Gaussian distribution, i.e., \\( n \\sim \\mathrm{Unif}(0,1) \\times \\mathcal{N}(0,1) \\). Conceptually, our distribution is a randomized Gaussian distribution with varying variances. In the Appendix, we illustrate that our design choice shows a more diverse norm distribution than the other distributions (See Fig. B.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.71 + ], + "angle": 0, + "content": "In our experiments, we empirically observe that the noise addition strategy improves the overall CIR performances with a large gap by bridging the modality gap. We also empirically observe that our design choice outperforms other probability distributions with a large gap by diversifying the impact of the random noise (see Tab. 9)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.307, + 0.74 + ], + "angle": 0, + "content": "3.3. Efficiency and scalability" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The most remarkable advantages of LinCIR beyond its generalizability are the training efficiency and scalability. First, a text dataset is storage-efficient; the caption storage size of CC3M dataset [39] is only 125 MB, while its image storage size is about 430GB, about 3,400 times larger. Second, the forward complexity of the textual encoder is notably lower than that of the visual encoder. For example, as shown in Tab. 2, the textual encoder has fewer depth, dimension size, and input token size than the visual encoder taking \\(224 \\times 224\\) resolution images. As a result, the average inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.38 + ], + "angle": 0, + "content": "time of the CLIP ViT-L visual encoder is \\(\\times 3.5\\) times slower than that of the textual encoder (Tab. 2). Furthermore, the average throughput of the ViT-G textual encoder is even \\(\\times 1.4\\) times faster than the ViT-L visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.457 + ], + "angle": 0, + "content": "All these advantages make LinCIR easily scalable. Even though we increase the backbone size, the overhead of the textual encoder is not significantly increased. We can train LinCIR with the CLIP ViT-G backbone in 48 minutes using 8 A100 GPUs and 2 hours using a single V100." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.471, + 0.634, + 0.487 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.496, + 0.715, + 0.511 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.76 + ], + "angle": 0, + "content": "We use three-layered MLP for the \\(\\phi\\) model: LN [1] - Linear - GeLU [18] - Linear - GeLU - Linear - LN. The intermediate hidden dimension is set to \\(4d\\) (\\(d\\) for each architecture is shown in Tab. 2). We do not apply the \\(\\ell_2\\)-normalization to the textual encoder outputs during training because, as shown in Fig. B.2, the added random noises have larger norm sizes than 1. If we apply the \\(\\ell_2\\)-normalization, we observe that the \\(\\phi\\) module is not converged. Keywords of the given text are extracted by the POS tagger of spacy library. We use the AdamW optimizer [31] with a fixed learning rate of 0.0001, weight decay of 0.01, and mini-batch size of 512. Dropout with probability \\(50\\%\\) is applied for the regularization. We use CC3M [39] captions and 2.47M number of the curated StableDiffusion prompts² for the training dataset (i.e., there are 5.5M training captions)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.866 + ], + "angle": 0, + "content": "For a fair comparison between models, we select the model showing the best zero-shot CIRR [29] dev R@1 score for the model selection. We employ an early stopping strategy by monitoring the validation score. We evaluate the CIR performances of the selected model in a zero-shot manner, i.e., one model is evaluated on four benchmarks. We employ the visual and textual encoders of the official" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://huggingface.co/datasets/FredZhang7/stable-diffusion-prompts-2.47M" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13229" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.257 + ], + "angle": 0, + "content": "
mAP@5mAP@10mAP@25mAP@50
ViT-LPic2Word†8.729.5110.6411.29
SEARLE†11.6812.7314.3315.12
LinCIR12.5913.5815.0015.85
ViT-HPic2Word11.6512.3313.7114.43
SEARLE16.0816.9218.8119.69
LinCIR17.6018.5220.4621.39
ViT-GPic2Word5.545.596.687.12
SEARLE13.2013.8515.3216.04
LinCIR19.7121.0123.1324.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.268, + 0.47, + 0.311 + ], + "angle": 0, + "content": "Table 3. CIRCO results. Results of Pic2Word [37], SEARLE [3], LinCIR by using different CLIP backbones are shown. \\(\\dagger\\) denotes that the numbers are measured by the official checkpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.386 + ], + "angle": 0, + "content": "CLIP ViT-L [35], and OpenCLIP ViT-H and ViT-G [19]. In the Appendix, we show that LinCIR can be easily extended to the other VLMs, such as BLIP [25]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.295, + 0.416 + ], + "angle": 0, + "content": "4.2. Experimental protocols" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.47, + 0.606 + ], + "angle": 0, + "content": "Evaluation benchmarks and metrics. As pointed out by Baldrati et al. [3], the existing CIR benchmarks only have a single positive, which can cause an unreliable evaluation. A similar phenomenon is also reported in the image-text cross-modal retrieval problem by Chun et al. [10]; such benchmarks can lead to a wrong model comparison result. For this reason, we use CIRCO as the main benchmark, which has multiple positives and measures a more reliable ranking-based metric, mAP@K [32]. We also report the R@K evaluation results on three additional datasets, GeneCIS [44], FashionIQ [47], and CIRR [29]. We describe the details of each dataset in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.744 + ], + "angle": 0, + "content": "In this paper, we argue that \\(\\mathbf{R}@\\mathbf{1}\\) results can be somewhat noisy due to the false negatives in the dataset. Note that these benchmarks only have a unique positive triplet for each query \\(\\langle x_{i_R},x_c,x_i\\rangle\\), i.e., if other plausible images (i.e. false negatives) exist in the gallery set, the \\(\\mathbf{R}@\\mathbf{1}\\) score in these benchmarks cannot correctly measure the actual retrieval performance. Due to this reason, we will focus on the mAP score if it is available. Otherwise, we will concentrate on \\(\\mathbf{R}@\\mathbf{K}\\) with a larger K (e.g., 10) rather than \\(\\mathbf{R}@\\mathbf{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Comparison methods. We compare LinCIR with the recent ZS-CIR methods: Pic2Word [37] and SEARLE [3]. For a fair comparison, we train all methods with the same backbone architecture as LinCIR, namely ViT-H and ViT-G CLIP backbones. ViT-L results are measured using the official checkpoints. We did not directly compare our method with recent methods that require massive external triplet datasets or take a long training time [15, 45]. More comparisons with these methods can be found in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.089, + 0.849, + 0.257 + ], + "angle": 0, + "content": "
R@1R@2R@3
ViT-LPic2Word†11.1621.4730.38
SEARLE†12.2622.1131.30
LinCIR12.1922.7632.38
ViT-HPic2Word11.8922.1731.94
SEARLE13.3423.7232.72
LinCIR13.7623.8733.16
ViT-GPic2Word10.6720.7029.50
SEARLE12.8722.6132.46
LinCIR13.6624.6433.54
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.311 + ], + "angle": 0, + "content": "Table 4. GeneCIS results. The average R@1, R@2, R@3 for \"Focus Attribute\", \"Change Attribute\", \"Focus Object\", and \"Change Object\" are shown. The full table is in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.332, + 0.634, + 0.346 + ], + "angle": 0, + "content": "4.3. Main results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.536 + ], + "angle": 0, + "content": "The experimental results are summarized in Fig. 4: LinCIR outperforms the comparison methods in training time and retrieval performances. In all benchmarks, we observe that while the performance of LinCIR is enhanced by enlarging the backbone size, Pic2Word and SEARLE show inferior performances with the ViT-G backbone. We presume that it is because Pic2Word and SEARLE have a limited understanding of complex text queries because their \\(\\phi\\) module are trained on texts not diverse enough (i.e., a photo of \\([\\S])\\). On the other hand, our \\(\\phi\\) module shows a better understanding of complex texts as LinCIR is trained on diverse real-world texts from the caption datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.734 + ], + "angle": 0, + "content": "We also provide the full evaluation results on the four benchmarks below. Tab. 3 shows the evaluation results on the CIRCO dataset. In all experiments, LinCIR outperforms others with a significant gap. We can observe a similar finding in the GeneCIS average R@K results for four different subtasks (Tab. 4), especially for R@K with a larger K. As shown in the Appendix, LinCIR outperforms other models, especially on \"Focus Attribute\" and \"Change Attribute\" tasks. We presume that it is because the Pic2Word and SEARLE training prompts are more specialized to objects (a photo of \\([\\\\])\\), where LinCIR can handle more detailed concepts in the image by altering all the nouns in the sentence (e.g., \"gray \\([\\\\] sleeps on a \\[\\\\]\\)\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "In the FashionIQ benchmark (Tab. 5), LinCIR even outperforms the state-of-the-art supervised method [2] with a large gap (38.32 vs. 45.11 in the average R@10). Our work is the first ZS-CIR method that outperforms the supervised CIR method despite its generalizability to the other CIR benchmarks and flexibility for handling various conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We observe somewhat mixed results on CIRR (Tab. 6): LinCIR achieves the best R@10 score, but not in some other metrics. We argue that this is because of two reasons. First, R@K with a small K cannot fully reflect the authentic performance. As shown in the Appendix, the retrieval results" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13230" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.841, + 0.293 + ], + "angle": 0, + "content": "
ShirtDressTopteeAverage
R@10R@50R@10R@50R@10R@50R@10R@50
ViT-LPic2Word†26.2043.6020.0040.2027.9047.4024.7043.70
SEARLE†26.8945.5820.4843.1329.3249.9725.5646.23
LinCIR29.1046.8120.9242.4428.8150.1826.2846.49
ViT-HPic2Word36.9055.9928.0151.5140.1862.0135.0356.50
SEARLE36.4655.4528.4651.0738.8160.8934.5755.80
LinCIR36.9057.7529.8052.1142.0762.5236.2657.46
ViT-GPic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
Combiner (supervised) [2]†39.9960.4533.8159.4041.4165.3738.3261.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.302, + 0.893, + 0.334 + ], + "angle": 0, + "content": "Table 5. FashionIQ results. \\( {}^{ \\dagger } \\) denotes that the numbers are from the original paper. LinCIR ViT-G even outperforms the previous state- of-the-art supervised CIR method [2] with a large gap although LinCIR is not directly trained on the FashionIQ dataset." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.355, + 0.477, + 0.539 + ], + "angle": 0, + "content": "
FullSubset
R@1R@5R@10R@1R@2R@3
ViT-LPic2Word†23.9051.7065.3053.7674.4687.08
SEARLE†24.2452.4866.2953.7675.0188.19
LinCIR25.0453.2566.6857.1177.3788.89
ViT-HPic2Word32.9463.1173.8662.2281.3591.23
SEARLE34.0063.9875.2564.6383.2192.77
LinCIR33.8363.5275.3562.4381.4792.12
ViT-GPic2Word30.4158.1269.2368.9285.4593.04
SEARLE34.8064.0775.1168.7284.7093.23
LinCIR35.2564.7276.0563.3582.2291.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.547, + 0.472, + 0.59 + ], + "angle": 0, + "content": "Table 6. CIRR results. Due to the noisy nature of CIRR as a ZS-CIR benchmark, we only highlight the R@10 score for the full CIRR set. The detailed discussion can be found in Sec. 4.3" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.901 + ], + "angle": 0, + "content": "of LinCIR are plausible to humans, but because the dataset has incomplete positives (i.e., there are many false negatives), the R@1 score cannot correctly evaluate the model performance. According to Chun et al. [10], the similarity between the rankings measured by R@K on a partially annotated benchmark and those measured by mAP on a fully annotated benchmark becomes lower when we use a small K (e.g., 1). Second, as observed by previous works [3, 37], the quality of the CIRR benchmark as a ZS-CIR benchmark is somewhat doubtful. The CIRR text relative captions are often not truly relative (i.e., there exist false positives), and reference images can even be harmful to retrieval. As pointed out by Baldrati et al. [3], this problem becomes more severe when we use a small subset of images, i.e., the subset R@Ks. In summary, due to the noisy nature of CIRR full and subset R@1s, we propose to focus on the CIRR full R@10 scores rather than other metrics. In CIRR full R@10, LinCIR shows the same trends as the other benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.355, + 0.897, + 0.441 + ], + "angle": 0, + "content": "
Supervision designCIRR dev R@1Fashion IQ R@10R@50
a photo of [$][37]21.6524.9344.35
Ours, but [$] extracted by ψI22.6322.0139.87
Our SMP design choice25.6626.2846.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.451, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Table 7. Impact of the supervision design. Different target text designs (e.g., \\([\\\\) ]$ sleeps on \\([\\\\( ]\\) in Fig. 3 (b)) affect the performances. \"Ours, but \\([\\\\) ]$ extracted by \\( \\psi_I \\)\" denotes that the textual encoder before the \\( \\phi \\) module is replaced with the visual encoder." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.52, + 0.878, + 0.676 + ], + "angle": 0, + "content": "
SMP Masking strategyCIRR devFashion IQ
R@1R@10R@50
All non-keyword tokens20.5919.6136.97
Random token22.9822.4440.92
All noun tokens24.9525.1645.19
1 keyword token23.9226.4245.91
3 keyword tokens24.6626.6946.54
5 keyword tokens25.1426.2846.29
All keyword tokens25.6626.2846.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.685, + 0.895, + 0.756 + ], + "angle": 0, + "content": "Table 8. Impact of the SMP masking strategy. The CIRR retrieval performances by varying the masking strategy for SMP (i.e., \"Keyword masking\" in Fig. 3 (b)) are shown. We define \"keyword\" as consecutive adjectives and nouns, except for \"All noun tokens\". \"All noun tokens\" defines keywords as nouns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.775, + 0.603, + 0.792 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.798, + 0.893, + 0.845 + ], + "angle": 0, + "content": "In this subsection, we provide detailed analyses of our design choices. If not specified, we compare the models on the CIRR dev split and FashionIQ test split." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Impact of Self-Masking Projection (SMP) supervision. We compare two variants of SMP in Tab. 7. First, instead of employing our keyword masking strategy, we use “a photo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13231" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.444, + 0.243 + ], + "angle": 0, + "content": "
Noise typeCIRR devFashion IQ
R@1R@10R@50
No noise19.7620.4238.31
Student-t23.0822.8141.00
Exponential23.5125.7845.26
\\( \\chi^2 \\)23.5423.1141.89
N(0,1)23.7023.3141.89
Unif(-1,1)25.1425.8845.78
N(0,1)×Unif(0,1)25.4726.0546.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.309 + ], + "angle": 0, + "content": "Table 9. Impact of the choice of the random noise. Adding noise to the textual latent space helps to mitigate the inferior generalizability due to the modality gap. Moreover, using better random noise can significantly boost the overall performances." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.328, + 0.467, + 0.449 + ], + "angle": 0, + "content": "
CC3MSDPCOYOOWTCIRCO mAP@5GeneCIS R@3FashionIQ R@10CIRR R@10Avg
✓XXX13.7232.8025.1164.9534.15
X✓XX9.5232.3823.6362.5832.03
XX✓X11.3631.4826.1865.3333.59
XXX✓9.6730.9024.4164.0532.26
✓✓XX12.5932.3826.2866.6834.48
✓✓X✓10.0633.3621.1163.1631.92
✓✓✓X11.5432.0826.9766.8034.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.459, + 0.468, + 0.529 + ], + "angle": 0, + "content": "Table 10. Impact of the training corpus. OpenWebText (OWT) [14] and SD prompts (SDP) are text-only datasets and CC3M [39] and COYO-700M (COYO) [5] are image-text aligned datasets. In our experiments, we use CC3M + SDP for the training corpus, considering the dataset scale and overall CIR performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.468, + 0.742 + ], + "angle": 0, + "content": "of \\([\\\\) ]\\( as Pic2Word and SEARLE to train the \\(\\phi\\) module. Second, we replace the textual encoder before \\(\\phi\\) module with the visual encoder using the corresponding image of the caption. In the table, \"a photo of \\([\\\\) ]\\( that [cond]\" variant performs worse than our design choice due to the limited diversity of the input texts. Interestingly, we observe that using both image and text pairs for LinCIR performs worse than our design choice. Note that the second model is trained on CC3M image-text pairs without using 2.47M SD prompts as our design choice. We presume it makes the \\(\\phi\\) model overfitted to CC3M image-text relationships, which significantly differ from our target CIR datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.903 + ], + "angle": 0, + "content": "Impact of masking design choice. We compare other masking design choices, such as random tokens or non-keyword tokens, with our design choice in Tab. 8. Tab. 8 shows that (1) masking the keywords performs better than masking the others. (2) Our keyword design – i.e., consecutive adjectives and nouns – is better than defining keywords as nouns. (3) The overall performances are enhanced by increasing the masked keywords. As the differences are not significant, we mask all the keywords for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Impact of the random noise addition. As we discussed in Sec. 3.2, the random noise addition is critical to mitigating the modality gap. Tab. 9 supports this claim: when we do not add any noise, the performance becomes the worst. Our design choice shows the best performance among the other noise designs due to the diverse norm size of our distribution as shown in Fig. B.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.892, + 0.481 + ], + "angle": 0, + "content": "The impact of the training corpus. We evaluate the impact of the training corpus in Tab. 10. We use four corpora, CC3M [39] (3M captions), StableDiffusion Prompts (SDP) (2.47M text prompts), COYO-700M (700M captions) [5] and OpenText (OWT) [14] (8M web texts). CC3M and COYO are image-text paired datasets, and OWT and SDP are text-only datasets. The example training samples of each dataset are shown in the Appendix. In Tab. 10, we observe that models trained with image descriptions (CC3M and COYO) perform better than models trained with general texts because general web texts are often irrelevant to visual information. In addition, using multiple corpora improves the overall performance, e.g., CC3M \\((34.15)\\rightarrow\\) CC3M \\(^+\\) SDP (34.48). Although we observe that using more massive captions (e.g., \\(3\\mathrm{M} + 2.47\\mathrm{M} + 700\\mathrm{M})\\) can be helpful for some CIR tasks, such as FashionIQ and CIRR, we use CC3M and SDP for our training set considering the corpus size \\((3\\mathrm{M} + 2.47\\mathrm{M})\\), and the overall performances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Qualitative results. We provide the additional qualitative retrieval results in the Appendix. In summary, LinCIR shows qualitatively plausible retrieval performance even in a large-scale image database, e.g. LAION-2B. Also, we observe that the retrieved results by LinCIR often suffer from false negatives in benchmark datasets, showing the limitation of the R@1 evaluation on the existing CIR benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.618, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose a novel zero-shot composed image retrieval (ZS-CIR) framework named Language Only training for Composed Image Retrieval (LinCIR). LinCIR presents a breakthrough in addressing the challenges associated with the previous ZS-CIR methods. By leveraging a novel self-supervision technique, Self-Masking Projection (SMP), LinCIR eliminates the dependency on expensive CIR triplets, opting for a training process solely based on text inputs. This innovative approach significantly enhances scalability and generalizability, overcoming limitations observed in existing ZS-CIR methods. Notably, our LinCIR model achieves remarkable efficiency and ZS-CIR performances compared to other methods on multiple CIR benchmarks, including CIRCO, GeneCIS, FashionIQ, and CIRR. We underscore the effectiveness of our language-only training framework, offering a potent solution with wide-ranging implications for image retrieval tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13232" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.155 + ], + "angle": 0, + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.212 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, pages 4959-4968, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.214, + 0.471, + 0.255 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[6] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. In ECCV, pages 136-152. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.465 + ], + "angle": 0, + "content": "[7] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In CVPR, pages 3001-3011, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.52 + ], + "angle": 0, + "content": "[8] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and TatSeng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization. arXiv preprint arXiv:2211.07394, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.522, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[9] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[10] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image_caption associations for ms-coco. In ECCV, 2022. 6, 7, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.703 + ], + "angle": 0, + "content": "[12] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback. arXiv preprint arXiv:2007.00145, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[14] Aaron Gokaslan, Vanya Cohen, Ellie Pavlick, and Stefanie Tellex. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019. 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[15] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[16] Sophia Gu, Christopher Clark, and Aniruddha Kembhavi. I can't believe there's no images! learning visual tasks us" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ing only language supervision. In ICCV, pages 2672-2683, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[20] Surgan Jandial, Ayush Chopra, Pinkesh Badjatiya, Pranit Chawla, Mausoom Sarkar, and Balaji Krishnamurthy. Trace: Transform aggregate and compose visiolinguistic representations for image search with text feedback. arXiv preprint arXiv:2009.01485, 7, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[21] Surgan Jandial, Pinkesh Badjatiya, Pranit Chawla, Ayush Chopra, Mausoom Sarkar, and Balaji Krishnamurthy. Sac: Semantic attention composition for text-conditioned image retrieval. In WACV, pages 4021-4030, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[24] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, pages 802-812, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 2, 6, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[26] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding CLIP latents for zero-shot captioning via text-only training. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[27] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 35:17612-17625, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[29] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1, 2, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13233" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[32] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[33] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.468, + 0.289 + ], + "angle": 0, + "content": "[34] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, pages 13018-13028, 2021. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.468, + 0.361 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.434, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[37] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 4, 6, 7, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.468, + 0.6 + ], + "angle": 0, + "content": "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 2, 5, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[40] Minchul Shin, Yoonjae Cho, Byungsoo Ko, and Geonmo Gu. Rtic: Residual learning for text and image composition using graph convolutional network. arXiv preprint arXiv:2104.03015, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[41] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[42] Yuxin Tian, Shawn Newsam, and Kofi Boakye. Image search with text feedback by additive attention compositional learning. arXiv preprint arXiv:2203.03809, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[44] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5, 6, 11" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[45] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[46] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[47] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 2, 5, 6, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[48] Youngjae Yu, Seunghwan Lee, Yuncheol Choi, and Gunhee Kim. Curlingnet: Compositional learning between images and text for fashion iq data. arXiv preprint arXiv:2003.12299, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13234" + } + ] +] \ No newline at end of file diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_origin.pdf b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef3c2e6a7c09ed80e62fa039bbcbcb79c1a26098 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/35e5e309-9782-4a25-8b32-4eac66825f83_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:850a1023e91e9a7274130ba80c9f25dbc2102bc7d253b0c96bbc28d3403c882c +size 876588 diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/full.md b/2024/Language-only Training of Zero-shot Composed Image Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b7341625c292b0fe9da5bb16072eaa7b6083536 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/full.md @@ -0,0 +1,249 @@ +# Language-only Efficient Training of Zero-shot Composed Image Retrieval + +Geonmo Gu*,1 Sanghyuk Chun*,2 Wonjae Kim² Yoohoon Kang¹ Sangwoo Yun² +1NAVER Vision 2NAVER AI Lab * Equal contribution + +# Abstract + +Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LinCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LinCIR is surprisingly efficient and highly effective; LinCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ. Code is available at github.com/navervision/lincir + +# 1. Introduction + +Composed image retrieval (CIR) is a challenging vision-language (VL) task that takes a composed query of image and text, aiming to search relative images for both conditions [29]. As language serves as the most natural method for encoding human interaction, CIR provides a higher degree of freedom and a better user experience for image-based search engine applications, such as web commerce. + +One of the main challenges of CIR is the expensive dataset collection pipeline. CIR datasets consist of triplets $\langle x_{i_R},x_c,x_i\rangle$ , where $x_{i_R}$ is an image query, $x_{c}$ is a text query, and $x_{i}$ is the target image. Unlike image-text paired datasets, such as CC3M [39] or LAION [38], such triplets + +![](images/6e7e4cb41b2d7b9c6a250bd8548ab831d1dd012bd08b6834836418b9351d41c1.jpg) +Figure 1. Training time (hours) vs. Zero-shot Composed Image Retrieval (ZS-CIR) performance. Thanks to our efficient language-only training strategy, our LinCIR outperforms the previous ZS-CIR methods in both training time and CIR performance. The training time is measured on 8 A100 GPUs. We compare the models on the CIRCO mAP@5 [3] score for a more comprehensive evaluation of CIR models (more results are in Fig. 4). Notably, when we scale up the backbone CLIP [19, 35] model size by ViT-L, ViT-H and ViT-G, LinCIR shows a promising performance boost with surprisingly short training time (48 mins for ViT-G). On the other hand, Pic2Word [37] and SEARLE [3] cannot be scaled up to CLIP ViT-G due to their limitation on restricted textual expressions and the lack of diversity of input texts. + +are almost impossible to collect by web crawling and require expensive human labor to create each triplet. For example, the datasets are constructed by gathering candidates of $\langle x_{i_R},x_i\rangle$ and manually annotating $x_{c}$ by human annotators [29, 47], which is hard to scalable. Therefore, the size of the training triplets is usually small (e.g., 46.6k [47], 28.8k [29]), and the existing CIR methods trained on such triplets [2, 6-8, 11, 12, 20-22, 24, 30, 40, 42, 46, 48] suffer from the lack of generalizability to diverse unseen domains. + +To overcome the drawback, recent studies have explored zero-shot CIR (ZS-CIR), a scalable direction, by eliminating the dependency on the pre-collected triplet datasets. For example, Saito et al. [37] and Baldrati et al. [3] propose projection-based ZS-CIR methods without using triplet datasets. Based on the pre-trained CLIP [35], they train a lightweight projection module $\phi$ that projects the CLIP image latent embedding $z_{i}$ to the CLIP text token + +![](images/da7a488300c42f1f41157206c5c13ad3e4c7084ef8f0b852ef775a9cb0dcb844.jpg) +Figure 2. Overview of ZS-CIR with a projection to the token embedding space. The mainstream ZS-CIR methods, such as Pic2Word [37], SEARLE [3] and LinCIR (ours), train a projection module $\phi$ that projects the image latent embedding $z_{i}$ into the token embedding space $e_c$ with a custom prompt (e.g., a photo of $[\mathbb{S}]$ that [cond]). The textual encoder output is used for CIR. + +embedding space $e_{c}$ (see Fig. 2). These approaches have shown promising generalizability to unseen datasets. However, these methods struggle to handle diverse text conditions because they rely on pre-defined naïve text prompts during training (e.g., a photo of [§]). Moreover, their training frameworks need sequential forward operations of the visual and textual encoders (see Fig. 3(a)), resulting in inefficient training and limited scalability to a larger backbone. + +In this paper, we introduce a new paradigm of ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). As shown in Fig. 3 (b), instead of projecting the image latent embedding $z_{i}$ , we propose to project the text latent embedding $z_{c}$ to the token embedding space $e_{c}$ . We introduce a novel self-supervision, named Self-Masking Projection (SMP), for language-only training. We replace all the "keywords" of the original text with the projected text embedding of the original text to produce an embedding $\widehat{z}_{c}$ and apply MSE loss between $z_{c}$ and $\widehat{z}_{c}$ . Here, we define "keyword" as consecutive adjectives and nouns. For example, the keywords of "gray cat sleeps on a pillow" are "gray cat" and "a pillow"; therefore, it becomes "[\$] sleeps on [\$]". The purpose of SMP is to make [\$] token interpreted as the "one-word summarization" of the input by extracting the essential information of the input. During inference, we simply perform text-to-image retrieval by projecting image embedding to the token embedding space using the projection module $\phi$ as shown in Fig. 2. However, this strategy can suffer from the modality gap between textual and visual modalities [27], i.e., even though our $\phi$ module works perfectly for text latent embeddings, it can underperform for the target visual embeddings. We mitigate the issue by employing a random noise addition strategy [33], carefully choosing a probability distribution that ensures the diversity of the noise-augmented textual embeddings. + +Our paradigm has three advantages over the previous approaches. First, while Pic2Word and SEARLE projection modules are trained with a restricted text prompt, i.e., a photo of $\left[\S\right]$ , the projection module of LinCIR is trained with the diverse text inputs from the actual texts. Due to this reason, Pic2Word and SEARLE show degenerated + +performances when the backbone size becomes larger (see Fig. 1). On the other hand, LinCIR is more generalizable to complex and diverse text conditions, showing superior ZS-CIR performances than others, especially for larger backbones. Second, as LinCIR only utilizes the textual encoder, our training process is highly efficient and scalable than the methods incorporated with the visual encoder; our language-only training strategy is $\times 6.0$ faster than Pic2Word [37] and $\times 8.4$ faster than SEARLE [3] with CLIP ViT-L backbone. When we scale up the backbone size to CLIP ViT-G, the gap becomes $\times 16.4$ and $\times 17.6$ , respectively. Even ViT-G training for LinCIR only takes 48 minutes using 8 A100 and less than 2 hours using 1 V100. Third, our method is storage-efficient; for example, the CC3M dataset [39] images occupy about 430GB storage size, while its captions only need 125MB. We train LinCIR only with 571MB storage size for storing the 5.5M training captions. In summary, LinCIR shows the best training time and ZS-CIR performance as shown in Fig. 1 and 4. + +Our contribution can be summarized as follows: (1) We propose LinCIR, a novel and efficient language-only training framework for ZS-CIR. (2) We introduce a new self-supervision for language-only training, named SelfMasking Projection (SMP). (3) We employ a better random noise addition strategy than naive Gaussian noise to mitigate the modality gap. (4) LinCIR achieves the best training time and the ZS-CIR performances on four ZS-CIR benchmarks (CIRCO [3], GeneCIS [44], FashionIQ [47] and CIRR [29]). Notably, LinCIR even outperforms the state-of-the-art supervised method [2] on FashionIQ. + +# 2. Preliminaries + +Vision-language models (VLM). As with previous ZS-CIR methods, LinCIR utilizes a pre-trained VLM, such as CLIP [19, 35] or BLIP [25]. We use VLMs that map an image input $x_{i}$ and a text input $x_{c}$ to the $d$ -dimensional joint embedding space. Here, a given caption $x_{c}$ is tokenized by the pre-defined tokenizer as $t_c = \{t_c^k\mid k = 1\dots K\}$ , where $K$ is the number of tokens, and mapped to token embeddings $e_{c} = E_{w}(t_{c}) = \{E_{w}(t_{c}^{k})\mid k = 1\dots K\}$ , where $E_{w}$ is the embedding layer parameterized by $w$ . Afterwords, a + +![](images/bffd32666c233df65638f0fc875db44563415e25a58ae0809933479b61b140fb.jpg) +(a) Pic2Word training procedure + +![](images/aaf2b51b98802a017066951ce6ddb7b73fa1fa632d3af887a59e9b99845b1df9.jpg) +(b) Our language-only training procedure +Figure 3. Comparison of Pic2Word [37] and LinCIR training procedures. (a) Pic2Word [37] and SEARLE [3] training procedure requires both the visual encoder and the textual encoder. They only need images for training, while the text prompt is pre-defined [37] or automatically generated [3]. (b) LinCIR is trained solely on texts with the frozen textual encoder. First, a projection module $\phi$ projects a textual latent embedding of a sentence $z_{t}$ into the token embedding space. Before the projection, a random noise $n$ is added to $z_{t}$ to reduce the modality gap between text and image. We introduce a new self-supervision, named Self-Masking Projection (SMP), by replacing all keywords of the given caption with the projected embedding by $\phi$ and extracting a modified text embedding $\widehat{z}_{t}$ . Finally, the projection module $\phi$ is trained by the MSE loss between $z_{t}$ and $\widehat{z}_{t}$ . Note that both (a) and (b) use the same inference strategy shown in Fig. 2. + +textual encoder $\psi_C$ encodes the token embeddings to extract textual latent embeddings $z_{c} = \psi_{C}(x_{c}) = \psi_{C}(E_{w}(t_{c}))$ . The visual latent embeddings are extracted by the visual encoder $\psi_I$ : $z_{i} = \psi_{I}(x_{i})$ . We will use the terminology token embeddings to represent $e_c$ , and use textual (or visual) latent embeddings to represent $z_{c}$ (or $z_{i}$ ). + +VLM modality gap. The VLM joint embedding space suffers from the semantic gap between each modality. Namely, the visual and the textual latent embeddings are not exactly aligned with each other, but they are located in completely separate regions of the embedding space [27]. + +Such a modality gap hinders the harmonization of visual and textual latents in the joint embedding space. To mitigate the gap, language-only training has been recently introduced for image captioning tasks [16, 26, 33]. They train a text decoder from the CLIP textual embeddings to generate image captions from visual embeddings. During training the decoder, the modality gap is bridged by injecting Gaussian noises [16, 33] or projection-based method [26]. + +In this paper, we utilize language-only training and a noise-addition strategy to mitigate the modality gap. We carefully studied the impact of noise and found a better distribution rather than simple Gaussian noise. We believe our findings can be transferred to other language-only training methods. Moreover, our target task, CIR, needs to model a relationship between triplets of $\langle x_{i_R},x_c,x_i\rangle$ , whereas previous language-only training methods only consider the pair-wise relationship (i.e., image captioning). We tackle this problem by proposing a novel self-supervision, named Self-Masking Projection (SMP), which injects the projected textual embeddings into the original token embeddings. + +CIR by projection to token embeddings. The mainstream ZS-CIR methods, such as Pic2Word [37] and SEARLE [3], employ a projection-based method. Namely, + +they learn a projection module \(\phi\) from the image latent embedding space to the token embedding space. For inference, they project the image latent embedding \(z_{i}\) to the token embedding (\$), then perform text-to-image retrieval with the prompt "a photo of [\\(] that [cond]", where [cond] is a text condition. Fig. 2 shows an overview of how textual projection-based ZS-CIR works. The main research point of this field is how to train the projection module \(\phi\) to capture the visual information into the token embedding space. + +Pic2Word [37] trains the projection module $\phi$ by minimizing contrastive loss between image latent embedding and the textual latent embedding of "a photo of [\$]" (see Fig. 3 (a)). SEARLE [3] employs a similar approach to Pic2Word. First, they employ optimization-based textual inversion to generate pre-defined special tokens for an image and train $\phi$ to predict the token embedding. SEARLE employs CLIP zero-shot classification to predict the "concept" of the given image and refine the prompt by letting GPT [4] continue the phrase. Both methods only use image inputs $x_{i}$ for training without accessing the CIR triplets. + +Although Pic2Word and SEARLE achieve reasonable ZS-CIR performances, they have two significant problems. First, they heavily rely on the initial prompt, "a photo of $[\mathbb{S}]$ ", limiting the diversity of the textual encoder input. As Fig. 1, we argue that diversifying the input texts during training the projection module $\phi$ is critical to train with larger backbones (e.g., CLIP ViT-G), while using the naive prompt is failed to scale up the backbone size. Second, they need image inputs, which are less compact and redundant than text datasets. Moreover, the visual encoder usually needs more computation resources than the textual encoder because the visual encoder takes the fixed length token (e.g., 256). In contrast, the textual encoder takes shorter token lengths (e.g., average token length of CC3M ≈ 10). We tackle the problems by introducing a language-only training method, showing remarkable efficiency and scalability. + +# 3. Language-only Training of Zero-shot CIR + +This section introduces a new paradigm for ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). We first introduce a novel language-only self-supervision, Self-Masking Projection (SMP), that enables a language-only training for CIR ( $\S 3.1$ ). Then, we explain the modality gap problem and empirically show that adding a carefully chosen random noise can mitigate the problem ( $\S 3.2$ ). Finally, we describe the advantage of LinCIR in terms of efficiency and scalability ( $\S 3.3$ ). + +# 3.1. Self-Masking Projection (SMP) + +We aim to learn a projection module $\phi$ that captures and retains the original visual information after the projection and the textual encoder. While the previous methods focus on directly mapping visual information to the token space with a naive prompt (i.e., a photo of $[\S])$ , we argue that focusing on the textual encoder is more important. Our zero-shot CIR is based on text-to-image retrieval (see Fig. 2). It means that the quality of the textual latent embedding is more critical to the final ZS-CIR performances. Hence, rather than focusing on minimizing the gap between visual information and the naive prompt, we aim to achieve a projection module $\phi$ that captures the semantics of the keywords in the given text. + +To achieve our goal, we introduce a novel language-only self-supervision named Self-Masking Projection (SMP). First, we project the textual embedding $z_{c}$ of a given text input $x_{c}$ with the projection module $\phi$ to the token embedding space, i.e., $\widehat{e}_{c} = \phi(z_{c})$ , where $\widehat{e}_{c}$ is the projected textual embedding. Then, we replace the token embeddings of all the "keywords" of $x_{c}$ with the projected token embedding $\widehat{e}_{c}$ . We define the keywords of the sentence as consecutive nouns and adjectives. For example, the keywords of "A Russian Blue cat is gray and cute" will be "A Russian Blue cat", "gray" and "cute"; hence it will be converted to "[$] is [\$] and [\$]", where [\$] is a special token to represent the projected token embedding $\widehat{e}_{c}$ . By treating all the main concepts (keywords) in the caption as the same [\$], we intend [\$] to represent the overall essential information of the inputs. Note that similar to [MASK] tokens of masked modeling, [\$] with different positions will be encoded in different features due to the positional embeddings. Using the converted caption, we extract a converted textual latent feature $\widehat{z}_{c}$ and minimize MSE loss between the original textual embedding $z_{c}$ and $\widehat{z}_{c}$ . Note that we only train the $\phi$ module while keeping the textual encoder frozen. + +The intuition behind SMP is that semantic information is not balanced across the tokens, but concentrated on the specific keywords. We assume that it is more common that adjectives and nouns in the sentence are more important than other part-of-speech (POS), such as verbs or adverbs. We empirically observe that our design choice (replacing all keyword token embeddings with $\widehat{e}_c$ ) is the best among + +
No noiseStudent-tExpx²N(0,1)Unif(-1,1)Ours
L0.81 (19.8)0.76 (23.1)0.76 (23.5)0.78 (23.5)0.77 (23.7)0.74 (25.1)0.71 (25.5)
H0.63 (31.8)0.59 (32.4)0.67 (28.3)0.64 (27.0)0.60 (32.8)0.59 (33.9)0.53 (34.8)
G0.51 (33.3)0.55 (36.1)0.63 (30.8)0.56 (30.7)0.55 (35.9)0.58 (35.3)0.48 (36.9)
+ +Table 1. Modality gap vs. distributions. Modality gap [27] (lower denotes less gap) on CC3M and CIRR dev R@1 (higher denotes better performance – in the parentheses) for different noises with different backbone sizes (from ViT-L/14 to ViT-G/14). + +the other variants, such as randomly replacing $n$ keyword tokens ( $n = 1,3,5$ ), replacing a random token, replacing all non-keyword tokens, or replacing all noun tokens (See Tab. 8). + +SMP has two benefits over the previous image-based ZS-CIR supervision [3, 37]. First, SMP allows the textual encoder to accept more diverse captions rather than "a photo of $\left[\mathbb{S}\right]$ ". While previous methods risk being sensitive to natural sentence variations, potentially affecting performances, our approach replaces tokens in natural sentences, maintaining robust performance across a more diverse set of sentence constructions. Second, SMP only requires language inputs; therefore, the overall training procedure is efficient regarding the training time and the storage size. It means that LinCIR can easily scale up in terms of the backbone size and the dataset scale. We will discuss the efficiency and the scalability of LinCIR in the Sec. 3.3. + +# 3.2. Searching for a better noise distribution for reducing the modality gap. + +Although SMP enables the language-only training, we still suffer from the modality gap between textual and visual embedding space [27]. Namely, even if the $\phi$ module works perfectly for language inputs, it can fail to be generalized to visual inputs. To tackle the problem, we employ a simple noise addition strategy following Nukrai et al. [33]: we add a random noise before the projection during training. Nukrai et al. [33] employed a simple Gaussian noise, but we empirically observe that Gaussian noise is not effective in mitigating the gap. Tab. 1 shows the modality gap [27] measured by various CLIP backbones by adding different probability distributions to textual embeddings. In the table, we observe that the careful choice of distribution greatly affects the modality gap and the final performances. + +We also observe that the generally used probabilistic distributions can suffer from a curse of dimensionality in the CLIP embedding space dimensions (e.g., 768-dim). The norm histogram of each probabilistic distribution (Fig. B.2) shows that the samples drawn from a Gaussian distribution have almost identical norm sizes. From this observation, we employ a probability distribution enforcing the diverse norm sizes instead of the Gaussian distribution. We multiply a random scalar value by a random vector drawn from a + +![](images/da5ecb7babba5693fc13c491934022d9e9799b27316692068159b601a2be93d4.jpg) +Figure 4. Training time vs. CIR performances. We evaluate three CIR methods with three backbone sizes: ViT-L, ViT-H, and ViT-G. To avoid an unreliable assessment due to the nature of R@1, CIR performances are measured in CIRCO mAP@5 [3], GeneCIS average R@3 [44], FashionIQ Average R@50 [47], and CIRR average R@10 [29]. In all evaluation results, LinCIR achieves the best training time-performance trade-off. Moreover, Pic2Word and SEARLE show degenerated performances when scaling up the backbone size. + +![](images/6d8a4550b62f7a70f138cdd6fbd0ada6c648bb72f94f29036c979fbe90202fdf.jpg) + +![](images/add62770dd1af3bd8c5bbf933849176017c50b0a92d974cf271a37e385a79037.jpg) + +![](images/c13d056df8434d878dd904daaa86e118d8bc48d0a8c67b24c20015868d0666ae.jpg) + +
Visual encoderTextual encoder
# layersd# tokensTP# layersd# tokensTP
ViT-L24102425618.51276812.5665.5
ViT-H32128025616.324102412.5635.5
ViT-G48166425611.832128012.5626.5
+ +Table 2. Configuration of CLIP visual and textual encoders. Every visual encoder uses the patch size 14 and the input resolution $224 \times 224$ . The text token length is the average token length of the CC3M [39] captions by the CLIP tokenizer. $d$ denotes the hidden dimension of Transformer blocks, and TP denotes throughput per second (higher means faster). TPs are measured by 1 V100 on CC3M [39] captions and COCO [28] images with FP16 weights. + +Gaussian distribution, i.e., $n \sim \mathrm{Unif}(0,1) \times \mathcal{N}(0,1)$ . Conceptually, our distribution is a randomized Gaussian distribution with varying variances. In the Appendix, we illustrate that our design choice shows a more diverse norm distribution than the other distributions (See Fig. B.2). + +In our experiments, we empirically observe that the noise addition strategy improves the overall CIR performances with a large gap by bridging the modality gap. We also empirically observe that our design choice outperforms other probability distributions with a large gap by diversifying the impact of the random noise (see Tab. 9). + +# 3.3. Efficiency and scalability + +The most remarkable advantages of LinCIR beyond its generalizability are the training efficiency and scalability. First, a text dataset is storage-efficient; the caption storage size of CC3M dataset [39] is only 125 MB, while its image storage size is about 430GB, about 3,400 times larger. Second, the forward complexity of the textual encoder is notably lower than that of the visual encoder. For example, as shown in Tab. 2, the textual encoder has fewer depth, dimension size, and input token size than the visual encoder taking $224 \times 224$ resolution images. As a result, the average inference + +time of the CLIP ViT-L visual encoder is $\times 3.5$ times slower than that of the textual encoder (Tab. 2). Furthermore, the average throughput of the ViT-G textual encoder is even $\times 1.4$ times faster than the ViT-L visual encoder. + +All these advantages make LinCIR easily scalable. Even though we increase the backbone size, the overhead of the textual encoder is not significantly increased. We can train LinCIR with the CLIP ViT-G backbone in 48 minutes using 8 A100 GPUs and 2 hours using a single V100. + +# 4. Experiments + +# 4.1. Implementation details + +We use three-layered MLP for the $\phi$ model: LN [1] - Linear - GeLU [18] - Linear - GeLU - Linear - LN. The intermediate hidden dimension is set to $4d$ ( $d$ for each architecture is shown in Tab. 2). We do not apply the $\ell_2$ -normalization to the textual encoder outputs during training because, as shown in Fig. B.2, the added random noises have larger norm sizes than 1. If we apply the $\ell_2$ -normalization, we observe that the $\phi$ module is not converged. Keywords of the given text are extracted by the POS tagger of spacy library. We use the AdamW optimizer [31] with a fixed learning rate of 0.0001, weight decay of 0.01, and mini-batch size of 512. Dropout with probability $50\%$ is applied for the regularization. We use CC3M [39] captions and 2.47M number of the curated StableDiffusion prompts² for the training dataset (i.e., there are 5.5M training captions). + +For a fair comparison between models, we select the model showing the best zero-shot CIRR [29] dev R@1 score for the model selection. We employ an early stopping strategy by monitoring the validation score. We evaluate the CIR performances of the selected model in a zero-shot manner, i.e., one model is evaluated on four benchmarks. We employ the visual and textual encoders of the official + +
mAP@5mAP@10mAP@25mAP@50
ViT-LPic2Word†8.729.5110.6411.29
SEARLE†11.6812.7314.3315.12
LinCIR12.5913.5815.0015.85
ViT-HPic2Word11.6512.3313.7114.43
SEARLE16.0816.9218.8119.69
LinCIR17.6018.5220.4621.39
ViT-GPic2Word5.545.596.687.12
SEARLE13.2013.8515.3216.04
LinCIR19.7121.0123.1324.18
+ +CLIP ViT-L [35], and OpenCLIP ViT-H and ViT-G [19]. In the Appendix, we show that LinCIR can be easily extended to the other VLMs, such as BLIP [25]. + +# 4.2. Experimental protocols + +Evaluation benchmarks and metrics. As pointed out by Baldrati et al. [3], the existing CIR benchmarks only have a single positive, which can cause an unreliable evaluation. A similar phenomenon is also reported in the image-text cross-modal retrieval problem by Chun et al. [10]; such benchmarks can lead to a wrong model comparison result. For this reason, we use CIRCO as the main benchmark, which has multiple positives and measures a more reliable ranking-based metric, mAP@K [32]. We also report the R@K evaluation results on three additional datasets, GeneCIS [44], FashionIQ [47], and CIRR [29]. We describe the details of each dataset in the Appendix. + +In this paper, we argue that $\mathbf{R}@\mathbf{1}$ results can be somewhat noisy due to the false negatives in the dataset. Note that these benchmarks only have a unique positive triplet for each query $\langle x_{i_R},x_c,x_i\rangle$ , i.e., if other plausible images (i.e. false negatives) exist in the gallery set, the $\mathbf{R}@\mathbf{1}$ score in these benchmarks cannot correctly measure the actual retrieval performance. Due to this reason, we will focus on the mAP score if it is available. Otherwise, we will concentrate on $\mathbf{R}@\mathbf{K}$ with a larger K (e.g., 10) rather than $\mathbf{R}@\mathbf{1}$ . + +Comparison methods. We compare LinCIR with the recent ZS-CIR methods: Pic2Word [37] and SEARLE [3]. For a fair comparison, we train all methods with the same backbone architecture as LinCIR, namely ViT-H and ViT-G CLIP backbones. ViT-L results are measured using the official checkpoints. We did not directly compare our method with recent methods that require massive external triplet datasets or take a long training time [15, 45]. More comparisons with these methods can be found in the Appendix. + +Table 3. CIRCO results. Results of Pic2Word [37], SEARLE [3], LinCIR by using different CLIP backbones are shown. $\dagger$ denotes that the numbers are measured by the official checkpoint. + +
R@1R@2R@3
ViT-LPic2Word†11.1621.4730.38
SEARLE†12.2622.1131.30
LinCIR12.1922.7632.38
ViT-HPic2Word11.8922.1731.94
SEARLE13.3423.7232.72
LinCIR13.7623.8733.16
ViT-GPic2Word10.6720.7029.50
SEARLE12.8722.6132.46
LinCIR13.6624.6433.54
+ +Table 4. GeneCIS results. The average R@1, R@2, R@3 for "Focus Attribute", "Change Attribute", "Focus Object", and "Change Object" are shown. The full table is in the Appendix. + +# 4.3. Main results + +The experimental results are summarized in Fig. 4: LinCIR outperforms the comparison methods in training time and retrieval performances. In all benchmarks, we observe that while the performance of LinCIR is enhanced by enlarging the backbone size, Pic2Word and SEARLE show inferior performances with the ViT-G backbone. We presume that it is because Pic2Word and SEARLE have a limited understanding of complex text queries because their $\phi$ module are trained on texts not diverse enough (i.e., a photo of $[\S])$ . On the other hand, our $\phi$ module shows a better understanding of complex texts as LinCIR is trained on diverse real-world texts from the caption datasets. + +We also provide the full evaluation results on the four benchmarks below. Tab. 3 shows the evaluation results on the CIRCO dataset. In all experiments, LinCIR outperforms others with a significant gap. We can observe a similar finding in the GeneCIS average R@K results for four different subtasks (Tab. 4), especially for R@K with a larger K. As shown in the Appendix, LinCIR outperforms other models, especially on "Focus Attribute" and "Change Attribute" tasks. We presume that it is because the Pic2Word and SEARLE training prompts are more specialized to objects (a photo of $[\\])$ , where LinCIR can handle more detailed concepts in the image by altering all the nouns in the sentence (e.g., "gray $[\\] sleeps on a \[\\]$ "). + +In the FashionIQ benchmark (Tab. 5), LinCIR even outperforms the state-of-the-art supervised method [2] with a large gap (38.32 vs. 45.11 in the average R@10). Our work is the first ZS-CIR method that outperforms the supervised CIR method despite its generalizability to the other CIR benchmarks and flexibility for handling various conditions. + +We observe somewhat mixed results on CIRR (Tab. 6): LinCIR achieves the best R@10 score, but not in some other metrics. We argue that this is because of two reasons. First, R@K with a small K cannot fully reflect the authentic performance. As shown in the Appendix, the retrieval results + +
ShirtDressTopteeAverage
R@10R@50R@10R@50R@10R@50R@10R@50
ViT-LPic2Word†26.2043.6020.0040.2027.9047.4024.7043.70
SEARLE†26.8945.5820.4843.1329.3249.9725.5646.23
LinCIR29.1046.8120.9242.4428.8150.1826.2846.49
ViT-HPic2Word36.9055.9928.0151.5140.1862.0135.0356.50
SEARLE36.4655.4528.4651.0738.8160.8934.5755.80
LinCIR36.9057.7529.8052.1142.0762.5236.2657.46
ViT-GPic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
Combiner (supervised) [2]†39.9960.4533.8159.4041.4165.3738.3261.74
+ +Table 5. FashionIQ results. ${}^{ \dagger }$ denotes that the numbers are from the original paper. LinCIR ViT-G even outperforms the previous state- of-the-art supervised CIR method [2] with a large gap although LinCIR is not directly trained on the FashionIQ dataset. + +
FullSubset
R@1R@5R@10R@1R@2R@3
ViT-LPic2Word†23.9051.7065.3053.7674.4687.08
SEARLE†24.2452.4866.2953.7675.0188.19
LinCIR25.0453.2566.6857.1177.3788.89
ViT-HPic2Word32.9463.1173.8662.2281.3591.23
SEARLE34.0063.9875.2564.6383.2192.77
LinCIR33.8363.5275.3562.4381.4792.12
ViT-GPic2Word30.4158.1269.2368.9285.4593.04
SEARLE34.8064.0775.1168.7284.7093.23
LinCIR35.2564.7276.0563.3582.2291.98
+ +of LinCIR are plausible to humans, but because the dataset has incomplete positives (i.e., there are many false negatives), the R@1 score cannot correctly evaluate the model performance. According to Chun et al. [10], the similarity between the rankings measured by R@K on a partially annotated benchmark and those measured by mAP on a fully annotated benchmark becomes lower when we use a small K (e.g., 1). Second, as observed by previous works [3, 37], the quality of the CIRR benchmark as a ZS-CIR benchmark is somewhat doubtful. The CIRR text relative captions are often not truly relative (i.e., there exist false positives), and reference images can even be harmful to retrieval. As pointed out by Baldrati et al. [3], this problem becomes more severe when we use a small subset of images, i.e., the subset R@Ks. In summary, due to the noisy nature of CIRR full and subset R@1s, we propose to focus on the CIRR full R@10 scores rather than other metrics. In CIRR full R@10, LinCIR shows the same trends as the other benchmarks. + +Table 6. CIRR results. Due to the noisy nature of CIRR as a ZS-CIR benchmark, we only highlight the R@10 score for the full CIRR set. The detailed discussion can be found in Sec. 4.3 + +
Supervision designCIRR dev R@1Fashion IQ R@10R@50
a photo of [$][37]21.6524.9344.35
Ours, but [$] extracted by ψI22.6322.0139.87
Our SMP design choice25.6626.2846.49
+ +Table 7. Impact of the supervision design. Different target text designs (e.g., \([\\) ]$ sleeps on \([\\( ]\) in Fig. 3 (b)) affect the performances. "Ours, but \([\\) ]$ extracted by \( \psi_I \)" denotes that the textual encoder before the \( \phi \) module is replaced with the visual encoder. + +
SMP Masking strategyCIRR devFashion IQ
R@1R@10R@50
All non-keyword tokens20.5919.6136.97
Random token22.9822.4440.92
All noun tokens24.9525.1645.19
1 keyword token23.9226.4245.91
3 keyword tokens24.6626.6946.54
5 keyword tokens25.1426.2846.29
All keyword tokens25.6626.2846.49
+ +Table 8. Impact of the SMP masking strategy. The CIRR retrieval performances by varying the masking strategy for SMP (i.e., "Keyword masking" in Fig. 3 (b)) are shown. We define "keyword" as consecutive adjectives and nouns, except for "All noun tokens". "All noun tokens" defines keywords as nouns. + +# 4.4. Analysis + +In this subsection, we provide detailed analyses of our design choices. If not specified, we compare the models on the CIRR dev split and FashionIQ test split. + +Impact of Self-Masking Projection (SMP) supervision. We compare two variants of SMP in Tab. 7. First, instead of employing our keyword masking strategy, we use “a photo + +
Noise typeCIRR devFashion IQ
R@1R@10R@50
No noise19.7620.4238.31
Student-t23.0822.8141.00
Exponential23.5125.7845.26
\( \chi^2 \)23.5423.1141.89
N(0,1)23.7023.3141.89
Unif(-1,1)25.1425.8845.78
N(0,1)×Unif(0,1)25.4726.0546.29
+ +Table 9. Impact of the choice of the random noise. Adding noise to the textual latent space helps to mitigate the inferior generalizability due to the modality gap. Moreover, using better random noise can significantly boost the overall performances. + +
CC3MSDPCOYOOWTCIRCO mAP@5GeneCIS R@3FashionIQ R@10CIRR R@10Avg
✓XXX13.7232.8025.1164.9534.15
X✓XX9.5232.3823.6362.5832.03
XX✓X11.3631.4826.1865.3333.59
XXX✓9.6730.9024.4164.0532.26
✓✓XX12.5932.3826.2866.6834.48
✓✓X✓10.0633.3621.1163.1631.92
✓✓✓X11.5432.0826.9766.8034.35
+ +Table 10. Impact of the training corpus. OpenWebText (OWT) [14] and SD prompts (SDP) are text-only datasets and CC3M [39] and COYO-700M (COYO) [5] are image-text aligned datasets. In our experiments, we use CC3M + SDP for the training corpus, considering the dataset scale and overall CIR performances. + +of \([\\) ]\( as Pic2Word and SEARLE to train the \(\phi\) module. Second, we replace the textual encoder before \(\phi\) module with the visual encoder using the corresponding image of the caption. In the table, "a photo of \([\\) ]\( that [cond]" variant performs worse than our design choice due to the limited diversity of the input texts. Interestingly, we observe that using both image and text pairs for LinCIR performs worse than our design choice. Note that the second model is trained on CC3M image-text pairs without using 2.47M SD prompts as our design choice. We presume it makes the \(\phi\) model overfitted to CC3M image-text relationships, which significantly differ from our target CIR datasets. + +Impact of masking design choice. We compare other masking design choices, such as random tokens or non-keyword tokens, with our design choice in Tab. 8. Tab. 8 shows that (1) masking the keywords performs better than masking the others. (2) Our keyword design – i.e., consecutive adjectives and nouns – is better than defining keywords as nouns. (3) The overall performances are enhanced by increasing the masked keywords. As the differences are not significant, we mask all the keywords for simplicity. + +Impact of the random noise addition. As we discussed in Sec. 3.2, the random noise addition is critical to mitigating the modality gap. Tab. 9 supports this claim: when we do not add any noise, the performance becomes the worst. Our design choice shows the best performance among the other noise designs due to the diverse norm size of our distribution as shown in Fig. B.2. + +The impact of the training corpus. We evaluate the impact of the training corpus in Tab. 10. We use four corpora, CC3M [39] (3M captions), StableDiffusion Prompts (SDP) (2.47M text prompts), COYO-700M (700M captions) [5] and OpenText (OWT) [14] (8M web texts). CC3M and COYO are image-text paired datasets, and OWT and SDP are text-only datasets. The example training samples of each dataset are shown in the Appendix. In Tab. 10, we observe that models trained with image descriptions (CC3M and COYO) perform better than models trained with general texts because general web texts are often irrelevant to visual information. In addition, using multiple corpora improves the overall performance, e.g., CC3M $(34.15)\rightarrow$ CC3M $^+$ SDP (34.48). Although we observe that using more massive captions (e.g., $3\mathrm{M} + 2.47\mathrm{M} + 700\mathrm{M})$ can be helpful for some CIR tasks, such as FashionIQ and CIRR, we use CC3M and SDP for our training set considering the corpus size $(3\mathrm{M} + 2.47\mathrm{M})$ , and the overall performances. + +Qualitative results. We provide the additional qualitative retrieval results in the Appendix. In summary, LinCIR shows qualitatively plausible retrieval performance even in a large-scale image database, e.g. LAION-2B. Also, we observe that the retrieved results by LinCIR often suffer from false negatives in benchmark datasets, showing the limitation of the R@1 evaluation on the existing CIR benchmarks. + +# 5. Conclusion + +We propose a novel zero-shot composed image retrieval (ZS-CIR) framework named Language Only training for Composed Image Retrieval (LinCIR). LinCIR presents a breakthrough in addressing the challenges associated with the previous ZS-CIR methods. By leveraging a novel self-supervision technique, Self-Masking Projection (SMP), LinCIR eliminates the dependency on expensive CIR triplets, opting for a training process solely based on text inputs. This innovative approach significantly enhances scalability and generalizability, overcoming limitations observed in existing ZS-CIR methods. Notably, our LinCIR model achieves remarkable efficiency and ZS-CIR performances compared to other methods on multiple CIR benchmarks, including CIRCO, GeneCIS, FashionIQ, and CIRR. We underscore the effectiveness of our language-only training framework, offering a potent solution with wide-ranging implications for image retrieval tasks. + +# References + +[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5 +[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, pages 4959-4968, 2022. 1, 2, 6, 7 +[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 11, 13 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3 +[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.8, 11 +[6] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. In ECCV, pages 136-152. Springer, 2020. 1 +[7] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In CVPR, pages 3001-3011, 2020. +[8] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and TatSeng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization. arXiv preprint arXiv:2211.07394, 2022. 1 +[9] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 11 +[10] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image_caption associations for ms-coco. In ECCV, 2022. 6, 7, 11 +[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1 +[12] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback. arXiv preprint arXiv:2007.00145, 2020. 1 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 13 +[14] Aaron Gokaslan, Vanya Cohen, Ellie Pavlick, and Stefanie Tellex. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019. 8, 11 +[15] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 6, 13 +[16] Sophia Gu, Christopher Clark, and Aniruddha Kembhavi. I can't believe there's no images! learning visual tasks us + +ing only language supervision. In ICCV, pages 2672-2683, 2023. 3 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 11 +[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 6 +[20] Surgan Jandial, Ayush Chopra, Pinkesh Badjatiya, Pranit Chawla, Mausoom Sarkar, and Balaji Krishnamurthy. Trace: Transform aggregate and compose visiolinguistic representations for image search with text feedback. arXiv preprint arXiv:2009.01485, 7, 2020. 1 +[21] Surgan Jandial, Pinkesh Badjatiya, Pranit Chawla, Ayush Chopra, Mausoom Sarkar, and Balaji Krishnamurthy. Sac: Semantic attention composition for text-conditioned image retrieval. In WACV, pages 4021-4030, 2022. +[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 1 +[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 11 +[24] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, pages 802-812, 2021. 1 +[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 2, 6, 13, 14 +[26] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding CLIP latents for zero-shot captioning via text-only training. In ICLR, 2023. 3 +[27] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 35:17612-17625, 2022. 2, 3, 4 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 11 +[29] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1, 2, 5, 6, 11 +[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1 + +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5 +[32] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6, 11 +[33] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022. 2, 3, 4 +[34] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, pages 13018-13028, 2021. 11 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6 +[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 11 +[37] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 4, 6, 7, 11, 13 +[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 1 +[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 2, 5, 8, 11 +[40] Minchul Shin, Yoonjae Cho, Byungsoo Ko, and Geonmo Gu. Rtic: Residual learning for text and image composition using graph convolutional network. arXiv preprint arXiv:2104.03015, 2021. 1 +[41] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 11 +[42] Yuxin Tian, Shawn Newsam, and Kofi Boakye. Image search with text feedback by additive attention compositional learning. arXiv preprint arXiv:2203.03809, 2022. 1 +[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 13 +[44] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5, 6, 11 + +[45] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 6, 13 +[46] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1 +[47] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 2, 5, 6, 11, 13 +[48] Youngjae Yu, Seunghwan Lee, Yuncheol Choi, and Gunhee Kim. Curlingnet: Compositional learning between images and text for fashion iq data. arXiv preprint arXiv:2003.12299, 2020. 1 \ No newline at end of file diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/images.zip b/2024/Language-only Training of Zero-shot Composed Image Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b6ffeb59c4efcf3350d98dc45d85c1ccf9b01b30 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c1decdaa66c6cf33df28deeb85130e3dc493c496659b25adbb222857208d0d4 +size 524900 diff --git a/2024/Language-only Training of Zero-shot Composed Image Retrieval/layout.json b/2024/Language-only Training of Zero-shot Composed Image Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8a8a5a53d0a8c52eac6a19044a5799aecf23ac18 --- /dev/null +++ b/2024/Language-only Training of Zero-shot Composed Image Retrieval/layout.json @@ -0,0 +1,8000 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 524, + 121 + ], + "type": "text", + "content": "Language-only Efficient Training of Zero-shot Composed Image Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 515, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 515, + 185 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 515, + 185 + ], + "type": "text", + "content": "Geonmo Gu*,1 Sanghyuk Chun*,2 Wonjae Kim² Yoohoon Kang¹ Sangwoo Yun² \n1NAVER Vision 2NAVER AI Lab * Equal contribution" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": "Composed image retrieval (CIR) task takes a composed query of image and text, aiming to search relative images for both conditions. Conventional CIR approaches need a training dataset composed of triplets of query image, query text, and target image, which is very expensive to collect. Several recent works have worked on the zero-shot (ZS) CIR paradigm to tackle the issue without using pre-collected triplets. However, the existing ZS-CIR methods show limited backbone scalability and generalizability due to the lack of diversity of the input texts during training. We propose a novel CIR framework, only using language for its training. Our LinCIR (Language-only training for CIR) can be trained only with text datasets by a novel self-supervision named self-masking projection (SMP). We project the text latent embedding to the token embedding space and construct a new text by replacing the keyword tokens of the original text. Then, we let the new and original texts have the same latent embedding vector. With this simple strategy, LinCIR is surprisingly efficient and highly effective; LinCIR with CLIP ViT-G backbone is trained in 48 minutes and shows the best ZS-CIR performances on four different CIR benchmarks, CIRCO, GeneCIS, FashionIQ, and CIRR, even outperforming supervised method on FashionIQ. Code is available at github.com/navervision/lincir" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 653 + ], + "type": "text", + "content": "Composed image retrieval (CIR) is a challenging vision-language (VL) task that takes a composed query of image and text, aiming to search relative images for both conditions [29]. As language serves as the most natural method for encoding human interaction, CIR provides a higher degree of freedom and a better user experience for image-based search engine applications, such as web commerce." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": "One of the main challenges of CIR is the expensive dataset collection pipeline. CIR datasets consist of triplets " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\langle x_{i_R},x_c,x_i\\rangle" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{i_R}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is an image query, " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is a text query, and " + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 47, + 653, + 288, + 715 + ], + "type": "text", + "content": " is the target image. Unlike image-text paired datasets, such as CC3M [39] or LAION [38], such triplets" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 218, + 525, + 353 + ], + "blocks": [ + { + "bbox": [ + 320, + 218, + 525, + 353 + ], + "lines": [ + { + "bbox": [ + 320, + 218, + 525, + 353 + ], + "spans": [ + { + "bbox": [ + 320, + 218, + 525, + 353 + ], + "type": "image", + "image_path": "6e7e4cb41b2d7b9c6a250bd8548ab831d1dd012bd08b6834836418b9351d41c1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 547, + 496 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 496 + ], + "type": "text", + "content": "Figure 1. Training time (hours) vs. Zero-shot Composed Image Retrieval (ZS-CIR) performance. Thanks to our efficient language-only training strategy, our LinCIR outperforms the previous ZS-CIR methods in both training time and CIR performance. The training time is measured on 8 A100 GPUs. We compare the models on the CIRCO mAP@5 [3] score for a more comprehensive evaluation of CIR models (more results are in Fig. 4). Notably, when we scale up the backbone CLIP [19, 35] model size by ViT-L, ViT-H and ViT-G, LinCIR shows a promising performance boost with surprisingly short training time (48 mins for ViT-G). On the other hand, Pic2Word [37] and SEARLE [3] cannot be scaled up to CLIP ViT-G due to their limitation on restricted textual expressions and the lack of diversity of input texts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "text", + "content": "are almost impossible to collect by web crawling and require expensive human labor to create each triplet. For example, the datasets are constructed by gathering candidates of " + }, + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\langle x_{i_R},x_i\\rangle" + }, + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "text", + "content": " and manually annotating " + }, + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 304, + 509, + 547, + 617 + ], + "type": "text", + "content": " by human annotators [29, 47], which is hard to scalable. Therefore, the size of the training triplets is usually small (e.g., 46.6k [47], 28.8k [29]), and the existing CIR methods trained on such triplets [2, 6-8, 11, 12, 20-22, 24, 30, 40, 42, 46, 48] suffer from the lack of generalizability to diverse unseen domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "To overcome the drawback, recent studies have explored zero-shot CIR (ZS-CIR), a scalable direction, by eliminating the dependency on the pre-collected triplet datasets. For example, Saito et al. [37] and Baldrati et al. [3] propose projection-based ZS-CIR methods without using triplet datasets. Based on the pre-trained CLIP [35], they train a lightweight projection module " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " that projects the CLIP image latent embedding " + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": " to the CLIP text token" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13225" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 69, + 492, + 166 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 492, + 166 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 492, + 166 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 492, + 166 + ], + "type": "image", + "image_path": "da7a488300c42f1f41157206c5c13ad3e4c7084ef8f0b852ef775a9cb0dcb844.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "text", + "content": "Figure 2. Overview of ZS-CIR with a projection to the token embedding space. The mainstream ZS-CIR methods, such as Pic2Word [37], SEARLE [3] and LinCIR (ours), train a projection module " + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "text", + "content": " that projects the image latent embedding " + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "text", + "content": " into the token embedding space " + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "inline_equation", + "content": "e_c" + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "text", + "content": " with a custom prompt (e.g., a photo of " + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "inline_equation", + "content": "[\\mathbb{S}]" + }, + { + "bbox": [ + 46, + 168, + 547, + 204 + ], + "type": "text", + "content": " that [cond]). The textual encoder output is used for CIR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 217, + 287, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 217, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 45, + 217, + 287, + 313 + ], + "type": "text", + "content": "embedding space " + }, + { + "bbox": [ + 45, + 217, + 287, + 313 + ], + "type": "inline_equation", + "content": "e_{c}" + }, + { + "bbox": [ + 45, + 217, + 287, + 313 + ], + "type": "text", + "content": " (see Fig. 2). These approaches have shown promising generalizability to unseen datasets. However, these methods struggle to handle diverse text conditions because they rely on pre-defined naïve text prompts during training (e.g., a photo of [§]). Moreover, their training frameworks need sequential forward operations of the visual and textual encoders (see Fig. 3(a)), resulting in inefficient training and limited scalability to a larger backbone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": "In this paper, we introduce a new paradigm of ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). As shown in Fig. 3 (b), instead of projecting the image latent embedding " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": ", we propose to project the text latent embedding " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": " to the token embedding space " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "e_{c}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": ". We introduce a novel self-supervision, named Self-Masking Projection (SMP), for language-only training. We replace all the \"keywords\" of the original text with the projected text embedding of the original text to produce an embedding " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{c}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": " and apply MSE loss between " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{c}" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": ". Here, we define \"keyword\" as consecutive adjectives and nouns. For example, the keywords of \"gray cat sleeps on a pillow\" are \"gray cat\" and \"a pillow\"; therefore, it becomes \"[\\$] sleeps on [\\$]\". The purpose of SMP is to make [\\$] token interpreted as the \"one-word summarization\" of the input by extracting the essential information of the input. During inference, we simply perform text-to-image retrieval by projecting image embedding to the token embedding space using the projection module " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": " as shown in Fig. 2. However, this strategy can suffer from the modality gap between textual and visual modalities [27], i.e., even though our " + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 45, + 314, + 290, + 624 + ], + "type": "text", + "content": " module works perfectly for text latent embeddings, it can underperform for the target visual embeddings. We mitigate the issue by employing a random noise addition strategy [33], carefully choosing a probability distribution that ensures the diversity of the noise-augmented textual embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "content": "Our paradigm has three advantages over the previous approaches. First, while Pic2Word and SEARLE projection modules are trained with a restricted text prompt, i.e., a photo of " + }, + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\left[\\S\\right]" + }, + { + "bbox": [ + 46, + 624, + 288, + 696 + ], + "type": "text", + "content": ", the projection module of LinCIR is trained with the diverse text inputs from the actual texts. Due to this reason, Pic2Word and SEARLE show degenerated" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "content": "performances when the backbone size becomes larger (see Fig. 1). On the other hand, LinCIR is more generalizable to complex and diverse text conditions, showing superior ZS-CIR performances than others, especially for larger backbones. Second, as LinCIR only utilizes the textual encoder, our training process is highly efficient and scalable than the methods incorporated with the visual encoder; our language-only training strategy is " + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\times 6.0" + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "content": " faster than Pic2Word [37] and " + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\times 8.4" + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "content": " faster than SEARLE [3] with CLIP ViT-L backbone. When we scale up the backbone size to CLIP ViT-G, the gap becomes " + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\times 16.4" + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "inline_equation", + "content": "\\times 17.6" + }, + { + "bbox": [ + 304, + 217, + 547, + 445 + ], + "type": "text", + "content": ", respectively. Even ViT-G training for LinCIR only takes 48 minutes using 8 A100 and less than 2 hours using 1 V100. Third, our method is storage-efficient; for example, the CC3M dataset [39] images occupy about 430GB storage size, while its captions only need 125MB. We train LinCIR only with 571MB storage size for storing the 5.5M training captions. In summary, LinCIR shows the best training time and ZS-CIR performance as shown in Fig. 1 and 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 578 + ], + "type": "text", + "content": "Our contribution can be summarized as follows: (1) We propose LinCIR, a novel and efficient language-only training framework for ZS-CIR. (2) We introduce a new self-supervision for language-only training, named SelfMasking Projection (SMP). (3) We employ a better random noise addition strategy than naive Gaussian noise to mitigate the modality gap. (4) LinCIR achieves the best training time and the ZS-CIR performances on four ZS-CIR benchmarks (CIRCO [3], GeneCIS [44], FashionIQ [47] and CIRR [29]). Notably, LinCIR even outperforms the state-of-the-art supervised method [2] on FashionIQ." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 586, + 391, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 391, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 391, + 598 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Vision-language models (VLM). As with previous ZS-CIR methods, LinCIR utilizes a pre-trained VLM, such as CLIP [19, 35] or BLIP [25]. We use VLMs that map an image input " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and a text input " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "-dimensional joint embedding space. Here, a given caption " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is tokenized by the pre-defined tokenizer as " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "t_c = \\{t_c^k\\mid k = 1\\dots K\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the number of tokens, and mapped to token embeddings " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "e_{c} = E_{w}(t_{c}) = \\{E_{w}(t_{c}^{k})\\mid k = 1\\dots K\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "E_{w}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the embedding layer parameterized by " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". Afterwords, a" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 270, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 270, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 270, + 714 + ], + "type": "text", + "content": "1 Pronounced as \"linker\", meaning for linking the two modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13226" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 258, + 173 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 258, + 173 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 258, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 258, + 173 + ], + "type": "image", + "image_path": "bffd32666c233df65638f0fc875db44563415e25a58ae0809933479b61b140fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 174, + 217, + 184 + ], + "lines": [ + { + "bbox": [ + 94, + 174, + 217, + 184 + ], + "spans": [ + { + "bbox": [ + 94, + 174, + 217, + 184 + ], + "type": "text", + "content": "(a) Pic2Word training procedure" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 519, + 173 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 519, + 173 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 519, + 173 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 519, + 173 + ], + "type": "image", + "image_path": "aaf2b51b98802a017066951ce6ddb7b73fa1fa632d3af887a59e9b99845b1df9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 174, + 484, + 184 + ], + "lines": [ + { + "bbox": [ + 328, + 174, + 484, + 184 + ], + "spans": [ + { + "bbox": [ + 328, + 174, + 484, + 184 + ], + "type": "text", + "content": "(b) Our language-only training procedure" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": "Figure 3. Comparison of Pic2Word [37] and LinCIR training procedures. (a) Pic2Word [37] and SEARLE [3] training procedure requires both the visual encoder and the textual encoder. They only need images for training, while the text prompt is pre-defined [37] or automatically generated [3]. (b) LinCIR is trained solely on texts with the frozen textual encoder. First, a projection module " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " projects a textual latent embedding of a sentence " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " into the token embedding space. Before the projection, a random noise " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " to reduce the modality gap between text and image. We introduce a new self-supervision, named Self-Masking Projection (SMP), by replacing all keywords of the given caption with the projected embedding by " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " and extracting a modified text embedding " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{t}" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": ". Finally, the projection module " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " is trained by the MSE loss between " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{t}" + }, + { + "bbox": [ + 46, + 188, + 547, + 266 + ], + "type": "text", + "content": ". Note that both (a) and (b) use the same inference strategy shown in Fig. 2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": "textual encoder " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": " encodes the token embeddings to extract textual latent embeddings " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{c} = \\psi_{C}(x_{c}) = \\psi_{C}(E_{w}(t_{c}))" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": ". The visual latent embeddings are extracted by the visual encoder " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "\\psi_I" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{i} = \\psi_{I}(x_{i})" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": ". We will use the terminology token embeddings to represent " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "e_c" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": ", and use textual (or visual) latent embeddings to represent " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 46, + 282, + 287, + 354 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": "VLM modality gap. The VLM joint embedding space suffers from the semantic gap between each modality. Namely, the visual and the textual latent embeddings are not exactly aligned with each other, but they are located in completely separate regions of the embedding space [27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "content": "Such a modality gap hinders the harmonization of visual and textual latents in the joint embedding space. To mitigate the gap, language-only training has been recently introduced for image captioning tasks [16, 26, 33]. They train a text decoder from the CLIP textual embeddings to generate image captions from visual embeddings. During training the decoder, the modality gap is bridged by injecting Gaussian noises [16, 33] or projection-based method [26]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "text", + "content": "In this paper, we utilize language-only training and a noise-addition strategy to mitigate the modality gap. We carefully studied the impact of noise and found a better distribution rather than simple Gaussian noise. We believe our findings can be transferred to other language-only training methods. Moreover, our target task, CIR, needs to model a relationship between triplets of " + }, + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "inline_equation", + "content": "\\langle x_{i_R},x_c,x_i\\rangle" + }, + { + "bbox": [ + 46, + 522, + 288, + 666 + ], + "type": "text", + "content": ", whereas previous language-only training methods only consider the pair-wise relationship (i.e., image captioning). We tackle this problem by proposing a novel self-supervision, named Self-Masking Projection (SMP), which injects the projected textual embeddings into the original token embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "CIR by projection to token embeddings. The mainstream ZS-CIR methods, such as Pic2Word [37] and SEARLE [3], employ a projection-based method. Namely," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 390 + ], + "type": "text", + "content": "they learn a projection module \\(\\phi\\) from the image latent embedding space to the token embedding space. For inference, they project the image latent embedding \\(z_{i}\\) to the token embedding (\\$), then perform text-to-image retrieval with the prompt \"a photo of [\\\\(] that [cond]\", where [cond] is a text condition. Fig. 2 shows an overview of how textual projection-based ZS-CIR works. The main research point of this field is how to train the projection module \\(\\phi\\) to capture the visual information into the token embedding space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "content": "Pic2Word [37] trains the projection module " + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "content": " by minimizing contrastive loss between image latent embedding and the textual latent embedding of \"a photo of [\\$]\" (see Fig. 3 (a)). SEARLE [3] employs a similar approach to Pic2Word. First, they employ optimization-based textual inversion to generate pre-defined special tokens for an image and train " + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "content": " to predict the token embedding. SEARLE employs CLIP zero-shot classification to predict the \"concept\" of the given image and refine the prompt by letting GPT [4] continue the phrase. Both methods only use image inputs " + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 390, + 546, + 521 + ], + "type": "text", + "content": " for training without accessing the CIR triplets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "Although Pic2Word and SEARLE achieve reasonable ZS-CIR performances, they have two significant problems. First, they heavily rely on the initial prompt, \"a photo of " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbb{S}]" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "\", limiting the diversity of the textual encoder input. As Fig. 1, we argue that diversifying the input texts during training the projection module " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " is critical to train with larger backbones (e.g., CLIP ViT-G), while using the naive prompt is failed to scale up the backbone size. Second, they need image inputs, which are less compact and redundant than text datasets. Moreover, the visual encoder usually needs more computation resources than the textual encoder because the visual encoder takes the fixed length token (e.g., 256). In contrast, the textual encoder takes shorter token lengths (e.g., average token length of CC3M ≈ 10). We tackle the problems by introducing a language-only training method, showing remarkable efficiency and scalability." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13227" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 277, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 277, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 277, + 85 + ], + "type": "text", + "content": "3. Language-only Training of Zero-shot CIR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "content": "This section introduces a new paradigm for ZS-CIR, named Language Only training for Composed Image Retrieval (LinCIR). We first introduce a novel language-only self-supervision, Self-Masking Projection (SMP), that enables a language-only training for CIR (" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "inline_equation", + "content": "\\S 3.1" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "content": "). Then, we explain the modality gap problem and empirically show that adding a carefully chosen random noise can mitigate the problem (" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "inline_equation", + "content": "\\S 3.2" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "content": "). Finally, we describe the advantage of LinCIR in terms of efficiency and scalability (" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 46, + 91, + 287, + 199 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 205, + 217, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 217, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 217, + 217 + ], + "type": "text", + "content": "3.1. Self-Masking Projection (SMP)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "text", + "content": "We aim to learn a projection module " + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "text", + "content": " that captures and retains the original visual information after the projection and the textual encoder. While the previous methods focus on directly mapping visual information to the token space with a naive prompt (i.e., a photo of " + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "inline_equation", + "content": "[\\S])" + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "text", + "content": ", we argue that focusing on the textual encoder is more important. Our zero-shot CIR is based on text-to-image retrieval (see Fig. 2). It means that the quality of the textual latent embedding is more critical to the final ZS-CIR performances. Hence, rather than focusing on minimizing the gap between visual information and the naive prompt, we aim to achieve a projection module " + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 223, + 287, + 366 + ], + "type": "text", + "content": " that captures the semantics of the keywords in the given text." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": "To achieve our goal, we introduce a novel language-only self-supervision named Self-Masking Projection (SMP). First, we project the textual embedding " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " of a given text input " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " with the projection module " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " to the token embedding space, i.e., " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{e}_{c} = \\phi(z_{c})" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{e}_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " is the projected textual embedding. Then, we replace the token embeddings of all the \"keywords\" of " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " with the projected token embedding " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{e}_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": ". We define the keywords of the sentence as consecutive nouns and adjectives. For example, the keywords of \"A Russian Blue cat is gray and cute\" will be \"A Russian Blue cat\", \"gray\" and \"cute\"; hence it will be converted to \"[$] is [\\$] and [\\$]\", where [\\$] is a special token to represent the projected token embedding " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{e}_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": ". By treating all the main concepts (keywords) in the caption as the same [\\$], we intend [\\$] to represent the overall essential information of the inputs. Note that similar to [MASK] tokens of masked modeling, [\\$] with different positions will be encoded in different features due to the positional embeddings. Using the converted caption, we extract a converted textual latent feature " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " and minimize MSE loss between the original textual embedding " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "z_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\widehat{z}_{c}" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": ". Note that we only train the " + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 367, + 288, + 629 + ], + "type": "text", + "content": " module while keeping the textual encoder frozen." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "The intuition behind SMP is that semantic information is not balanced across the tokens, but concentrated on the specific keywords. We assume that it is more common that adjectives and nouns in the sentence are more important than other part-of-speech (POS), such as verbs or adverbs. We empirically observe that our design choice (replacing all keyword token embeddings with " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\widehat{e}_c" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ") is the best among" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 545, + 120 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 545, + 120 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 545, + 120 + ], + "type": "table", + "html": "
No noiseStudent-tExpx²N(0,1)Unif(-1,1)Ours
L0.81 (19.8)0.76 (23.1)0.76 (23.5)0.78 (23.5)0.77 (23.7)0.74 (25.1)0.71 (25.5)
H0.63 (31.8)0.59 (32.4)0.67 (28.3)0.64 (27.0)0.60 (32.8)0.59 (33.9)0.53 (34.8)
G0.51 (33.3)0.55 (36.1)0.63 (30.8)0.56 (30.7)0.55 (35.9)0.58 (35.3)0.48 (36.9)
", + "image_path": "f65242bbaa7da0c2bfb79f44b7ba62ff792873a750db3cec601c6e4971666dd9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 128, + 545, + 172 + ], + "lines": [ + { + "bbox": [ + 305, + 128, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 305, + 128, + 545, + 172 + ], + "type": "text", + "content": "Table 1. Modality gap vs. distributions. Modality gap [27] (lower denotes less gap) on CC3M and CIRR dev R@1 (higher denotes better performance – in the parentheses) for different noises with different backbone sizes (from ViT-L/14 to ViT-G/14)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "text", + "content": "the other variants, such as randomly replacing " + }, + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "text", + "content": " keyword tokens (" + }, + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "inline_equation", + "content": "n = 1,3,5" + }, + { + "bbox": [ + 304, + 194, + 545, + 241 + ], + "type": "text", + "content": "), replacing a random token, replacing all non-keyword tokens, or replacing all noun tokens (See Tab. 8)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 243, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 546, + 398 + ], + "type": "text", + "content": "SMP has two benefits over the previous image-based ZS-CIR supervision [3, 37]. First, SMP allows the textual encoder to accept more diverse captions rather than \"a photo of " + }, + { + "bbox": [ + 304, + 243, + 546, + 398 + ], + "type": "inline_equation", + "content": "\\left[\\mathbb{S}\\right]" + }, + { + "bbox": [ + 304, + 243, + 546, + 398 + ], + "type": "text", + "content": "\". While previous methods risk being sensitive to natural sentence variations, potentially affecting performances, our approach replaces tokens in natural sentences, maintaining robust performance across a more diverse set of sentence constructions. Second, SMP only requires language inputs; therefore, the overall training procedure is efficient regarding the training time and the storage size. It means that LinCIR can easily scale up in terms of the backbone size and the dataset scale. We will discuss the efficiency and the scalability of LinCIR in the Sec. 3.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 407, + 545, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 545, + 432 + ], + "type": "text", + "content": "3.2. Searching for a better noise distribution for reducing the modality gap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": "Although SMP enables the language-only training, we still suffer from the modality gap between textual and visual embedding space [27]. Namely, even if the " + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 437, + 545, + 605 + ], + "type": "text", + "content": " module works perfectly for language inputs, it can fail to be generalized to visual inputs. To tackle the problem, we employ a simple noise addition strategy following Nukrai et al. [33]: we add a random noise before the projection during training. Nukrai et al. [33] employed a simple Gaussian noise, but we empirically observe that Gaussian noise is not effective in mitigating the gap. Tab. 1 shows the modality gap [27] measured by various CLIP backbones by adding different probability distributions to textual embeddings. In the table, we observe that the careful choice of distribution greatly affects the modality gap and the final performances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 714 + ], + "type": "text", + "content": "We also observe that the generally used probabilistic distributions can suffer from a curse of dimensionality in the CLIP embedding space dimensions (e.g., 768-dim). The norm histogram of each probabilistic distribution (Fig. B.2) shows that the samples drawn from a Gaussian distribution have almost identical norm sizes. From this observation, we employ a probability distribution enforcing the diverse norm sizes instead of the Gaussian distribution. We multiply a random scalar value by a random vector drawn from a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13228" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 170, + 191 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 170, + 191 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 170, + 191 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 170, + 191 + ], + "type": "image", + "image_path": "da5ecb7babba5693fc13c491934022d9e9799b27316692068159b601a2be93d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 239 + ], + "type": "text", + "content": "Figure 4. Training time vs. CIR performances. We evaluate three CIR methods with three backbone sizes: ViT-L, ViT-H, and ViT-G. To avoid an unreliable assessment due to the nature of R@1, CIR performances are measured in CIRCO mAP@5 [3], GeneCIS average R@3 [44], FashionIQ Average R@50 [47], and CIRR average R@10 [29]. In all evaluation results, LinCIR achieves the best training time-performance trade-off. Moreover, Pic2Word and SEARLE show degenerated performances when scaling up the backbone size." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 72, + 291, + 192 + ], + "blocks": [ + { + "bbox": [ + 173, + 72, + 291, + 192 + ], + "lines": [ + { + "bbox": [ + 173, + 72, + 291, + 192 + ], + "spans": [ + { + "bbox": [ + 173, + 72, + 291, + 192 + ], + "type": "image", + "image_path": "6d8a4550b62f7a70f138cdd6fbd0ada6c648bb72f94f29036c979fbe90202fdf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 414, + 192 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 414, + 192 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 414, + 192 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 414, + 192 + ], + "type": "image", + "image_path": "add62770dd1af3bd8c5bbf933849176017c50b0a92d974cf271a37e385a79037.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 418, + 74, + 541, + 192 + ], + "blocks": [ + { + "bbox": [ + 418, + 74, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 418, + 74, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 418, + 74, + 541, + 192 + ], + "type": "image", + "image_path": "c13d056df8434d878dd904daaa86e118d8bc48d0a8c67b24c20015868d0666ae.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 251, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 48, + 251, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 48, + 251, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 287, + 319 + ], + "type": "table", + "html": "
Visual encoderTextual encoder
# layersd# tokensTP# layersd# tokensTP
ViT-L24102425618.51276812.5665.5
ViT-H32128025616.324102412.5635.5
ViT-G48166425611.832128012.5626.5
", + "image_path": "d7318f3333f44b366773b3cb79654151c82a12d71d41131e0dbd9356290aa3cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "type": "text", + "content": "Table 2. Configuration of CLIP visual and textual encoders. Every visual encoder uses the patch size 14 and the input resolution " + }, + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "type": "text", + "content": ". The text token length is the average token length of the CC3M [39] captions by the CLIP tokenizer. " + }, + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 328, + 287, + 406 + ], + "type": "text", + "content": " denotes the hidden dimension of Transformer blocks, and TP denotes throughput per second (higher means faster). TPs are measured by 1 V100 on CC3M [39] captions and COCO [28] images with FP16 weights." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 429, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 489 + ], + "type": "text", + "content": "Gaussian distribution, i.e., " + }, + { + "bbox": [ + 46, + 429, + 287, + 489 + ], + "type": "inline_equation", + "content": "n \\sim \\mathrm{Unif}(0,1) \\times \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 46, + 429, + 287, + 489 + ], + "type": "text", + "content": ". Conceptually, our distribution is a randomized Gaussian distribution with varying variances. In the Appendix, we illustrate that our design choice shows a more diverse norm distribution than the other distributions (See Fig. B.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 562 + ], + "type": "text", + "content": "In our experiments, we empirically observe that the noise addition strategy improves the overall CIR performances with a large gap by bridging the modality gap. We also empirically observe that our design choice outperforms other probability distributions with a large gap by diversifying the impact of the random noise (see Tab. 9)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 187, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 187, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 187, + 586 + ], + "type": "text", + "content": "3.3. Efficiency and scalability" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "The most remarkable advantages of LinCIR beyond its generalizability are the training efficiency and scalability. First, a text dataset is storage-efficient; the caption storage size of CC3M dataset [39] is only 125 MB, while its image storage size is about 430GB, about 3,400 times larger. Second, the forward complexity of the textual encoder is notably lower than that of the visual encoder. For example, as shown in Tab. 2, the textual encoder has fewer depth, dimension size, and input token size than the visual encoder taking " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " resolution images. As a result, the average inference" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "text", + "content": "time of the CLIP ViT-L visual encoder is " + }, + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\times 3.5" + }, + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "text", + "content": " times slower than that of the textual encoder (Tab. 2). Furthermore, the average throughput of the ViT-G textual encoder is even " + }, + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\times 1.4" + }, + { + "bbox": [ + 304, + 254, + 545, + 300 + ], + "type": "text", + "content": " times faster than the ViT-L visual encoder." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 361 + ], + "type": "text", + "content": "All these advantages make LinCIR easily scalable. Even though we increase the backbone size, the overhead of the textual encoder is not significantly increased. We can train LinCIR with the CLIP ViT-G backbone in 48 minutes using 8 A100 GPUs and 2 hours using a single V100." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 373, + 388, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 388, + 385 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 388, + 385 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 392, + 437, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 392, + 437, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 392, + 437, + 404 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": "We use three-layered MLP for the " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": " model: LN [1] - Linear - GeLU [18] - Linear - GeLU - Linear - LN. The intermediate hidden dimension is set to " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "4d" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": " for each architecture is shown in Tab. 2). We do not apply the " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": "-normalization to the textual encoder outputs during training because, as shown in Fig. B.2, the added random noises have larger norm sizes than 1. If we apply the " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": "-normalization, we observe that the " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": " module is not converged. Keywords of the given text are extracted by the POS tagger of spacy library. We use the AdamW optimizer [31] with a fixed learning rate of 0.0001, weight decay of 0.01, and mini-batch size of 512. Dropout with probability " + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 601 + ], + "type": "text", + "content": " is applied for the regularization. We use CC3M [39] captions and 2.47M number of the curated StableDiffusion prompts² for the training dataset (i.e., there are 5.5M training captions)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 685 + ], + "type": "text", + "content": "For a fair comparison between models, we select the model showing the best zero-shot CIRR [29] dev R@1 score for the model selection. We employ an early stopping strategy by monitoring the validation score. We evaluate the CIR performances of the selected model in a zero-shot manner, i.e., one model is evaluated on four benchmarks. We employ the visual and textual encoders of the official" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 306, + 693, + 545, + 712 + ], + "type": "text", + "content": "https://huggingface.co/datasets/FredZhang7/stable-diffusion-prompts-2.47M" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13229" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 203 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 203 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 203 + ], + "type": "table", + "html": "
mAP@5mAP@10mAP@25mAP@50
ViT-LPic2Word†8.729.5110.6411.29
SEARLE†11.6812.7314.3315.12
LinCIR12.5913.5815.0015.85
ViT-HPic2Word11.6512.3313.7114.43
SEARLE16.0816.9218.8119.69
LinCIR17.6018.5220.4621.39
ViT-GPic2Word5.545.596.687.12
SEARLE13.2013.8515.3216.04
LinCIR19.7121.0123.1324.18
", + "image_path": "511aa3655e50c8c63e5f072b1b52cff5d495c17bb3a7e3daeaefc2034fcc465b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 305 + ], + "type": "text", + "content": "CLIP ViT-L [35], and OpenCLIP ViT-H and ViT-G [19]. In the Appendix, we show that LinCIR can be easily extended to the other VLMs, such as BLIP [25]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 316, + 180, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 180, + 329 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 180, + 329 + ], + "type": "text", + "content": "4.2. Experimental protocols" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 287, + 479 + ], + "type": "text", + "content": "Evaluation benchmarks and metrics. As pointed out by Baldrati et al. [3], the existing CIR benchmarks only have a single positive, which can cause an unreliable evaluation. A similar phenomenon is also reported in the image-text cross-modal retrieval problem by Chun et al. [10]; such benchmarks can lead to a wrong model comparison result. For this reason, we use CIRCO as the main benchmark, which has multiple positives and measures a more reliable ranking-based metric, mAP@K [32]. We also report the R@K evaluation results on three additional datasets, GeneCIS [44], FashionIQ [47], and CIRR [29]. We describe the details of each dataset in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": "In this paper, we argue that " + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{1}" + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": " results can be somewhat noisy due to the false negatives in the dataset. Note that these benchmarks only have a unique positive triplet for each query " + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\langle x_{i_R},x_c,x_i\\rangle" + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": ", i.e., if other plausible images (i.e. false negatives) exist in the gallery set, the " + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{1}" + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": " score in these benchmarks cannot correctly measure the actual retrieval performance. Due to this reason, we will focus on the mAP score if it is available. Otherwise, we will concentrate on " + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{K}" + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": " with a larger K (e.g., 10) rather than " + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{1}" + }, + { + "bbox": [ + 46, + 481, + 287, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Comparison methods. We compare LinCIR with the recent ZS-CIR methods: Pic2Word [37] and SEARLE [3]. For a fair comparison, we train all methods with the same backbone architecture as LinCIR, namely ViT-H and ViT-G CLIP backbones. ViT-L results are measured using the official checkpoints. We did not directly compare our method with recent methods that require massive external triplet datasets or take a long training time [15, 45]. More comparisons with these methods can be found in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 519, + 203 + ], + "blocks": [ + { + "bbox": [ + 46, + 212, + 287, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 212, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 212, + 287, + 246 + ], + "type": "text", + "content": "Table 3. CIRCO results. Results of Pic2Word [37], SEARLE [3], LinCIR by using different CLIP backbones are shown. " + }, + { + "bbox": [ + 46, + 212, + 287, + 246 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 212, + 287, + 246 + ], + "type": "text", + "content": " denotes that the numbers are measured by the official checkpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 70, + 519, + 203 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 519, + 203 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 519, + 203 + ], + "type": "table", + "html": "
R@1R@2R@3
ViT-LPic2Word†11.1621.4730.38
SEARLE†12.2622.1131.30
LinCIR12.1922.7632.38
ViT-HPic2Word11.8922.1731.94
SEARLE13.3423.7232.72
LinCIR13.7623.8733.16
ViT-GPic2Word10.6720.7029.50
SEARLE12.8722.6132.46
LinCIR13.6624.6433.54
", + "image_path": "ac90ba2290087db6f597b384b31df1092571ba03ac8890378a1ba0ec172da068.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 212, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 246 + ], + "type": "text", + "content": "Table 4. GeneCIS results. The average R@1, R@2, R@3 for \"Focus Attribute\", \"Change Attribute\", \"Focus Object\", and \"Change Object\" are shown. The full table is in the Appendix." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 262, + 388, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 262, + 388, + 274 + ], + "spans": [ + { + "bbox": [ + 306, + 262, + 388, + 274 + ], + "type": "text", + "content": "4.3. Main results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": "The experimental results are summarized in Fig. 4: LinCIR outperforms the comparison methods in training time and retrieval performances. In all benchmarks, we observe that while the performance of LinCIR is enhanced by enlarging the backbone size, Pic2Word and SEARLE show inferior performances with the ViT-G backbone. We presume that it is because Pic2Word and SEARLE have a limited understanding of complex text queries because their " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " module are trained on texts not diverse enough (i.e., a photo of " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "[\\S])" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": ". On the other hand, our " + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 281, + 545, + 424 + ], + "type": "text", + "content": " module shows a better understanding of complex texts as LinCIR is trained on diverse real-world texts from the caption datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "We also provide the full evaluation results on the four benchmarks below. Tab. 3 shows the evaluation results on the CIRCO dataset. In all experiments, LinCIR outperforms others with a significant gap. We can observe a similar finding in the GeneCIS average R@K results for four different subtasks (Tab. 4), especially for R@K with a larger K. As shown in the Appendix, LinCIR outperforms other models, especially on \"Focus Attribute\" and \"Change Attribute\" tasks. We presume that it is because the Pic2Word and SEARLE training prompts are more specialized to objects (a photo of " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "[\\\\])" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": ", where LinCIR can handle more detailed concepts in the image by altering all the nouns in the sentence (e.g., \"gray " + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "inline_equation", + "content": "[\\\\] sleeps on a \\[\\\\]" + }, + { + "bbox": [ + 304, + 426, + 545, + 581 + ], + "type": "text", + "content": "\")." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "In the FashionIQ benchmark (Tab. 5), LinCIR even outperforms the state-of-the-art supervised method [2] with a large gap (38.32 vs. 45.11 in the average R@10). Our work is the first ZS-CIR method that outperforms the supervised CIR method despite its generalizability to the other CIR benchmarks and flexibility for handling various conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "We observe somewhat mixed results on CIRR (Tab. 6): LinCIR achieves the best R@10 score, but not in some other metrics. We argue that this is because of two reasons. First, R@K with a small K cannot fully reflect the authentic performance. As shown in the Appendix, the retrieval results" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13230" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 514, + 232 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 514, + 232 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 514, + 232 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 514, + 232 + ], + "type": "table", + "html": "
ShirtDressTopteeAverage
R@10R@50R@10R@50R@10R@50R@10R@50
ViT-LPic2Word†26.2043.6020.0040.2027.9047.4024.7043.70
SEARLE†26.8945.5820.4843.1329.3249.9725.5646.23
LinCIR29.1046.8120.9242.4428.8150.1826.2846.49
ViT-HPic2Word36.9055.9928.0151.5140.1862.0135.0356.50
SEARLE36.4655.4528.4651.0738.8160.8934.5755.80
LinCIR36.9057.7529.8052.1142.0762.5236.2657.46
ViT-GPic2Word33.1750.3925.4347.6535.2457.6231.2851.89
SEARLE36.4655.3528.1650.3239.8361.4534.8155.71
LinCIR46.7665.1138.0860.8850.4871.0945.1165.69
Combiner (supervised) [2]†39.9960.4533.8159.4041.4165.3738.3261.74
", + "image_path": "5341db20759a60ad69d0a3ff7145e2a7497b58fc2abe4dcd73f77ea05b531aeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 281, + 291, + 426 + ], + "blocks": [ + { + "bbox": [ + 46, + 239, + 546, + 264 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 546, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 546, + 264 + ], + "type": "text", + "content": "Table 5. FashionIQ results. " + }, + { + "bbox": [ + 46, + 239, + 546, + 264 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 46, + 239, + 546, + 264 + ], + "type": "text", + "content": " denotes that the numbers are from the original paper. LinCIR ViT-G even outperforms the previous state- of-the-art supervised CIR method [2] with a large gap although LinCIR is not directly trained on the FashionIQ dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 281, + 291, + 426 + ], + "lines": [ + { + "bbox": [ + 48, + 281, + 291, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 291, + 426 + ], + "type": "table", + "html": "
FullSubset
R@1R@5R@10R@1R@2R@3
ViT-LPic2Word†23.9051.7065.3053.7674.4687.08
SEARLE†24.2452.4866.2953.7675.0188.19
LinCIR25.0453.2566.6857.1177.3788.89
ViT-HPic2Word32.9463.1173.8662.2281.3591.23
SEARLE34.0063.9875.2564.6383.2192.77
LinCIR33.8363.5275.3562.4381.4792.12
ViT-GPic2Word30.4158.1269.2368.9285.4593.04
SEARLE34.8064.0775.1168.7284.7093.23
LinCIR35.2564.7276.0563.3582.2291.98
", + "image_path": "1f2ef51d387bcaff6050774d721625f7a21518b618c1496d5f80b4ecc2a749a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 713 + ], + "type": "text", + "content": "of LinCIR are plausible to humans, but because the dataset has incomplete positives (i.e., there are many false negatives), the R@1 score cannot correctly evaluate the model performance. According to Chun et al. [10], the similarity between the rankings measured by R@K on a partially annotated benchmark and those measured by mAP on a fully annotated benchmark becomes lower when we use a small K (e.g., 1). Second, as observed by previous works [3, 37], the quality of the CIRR benchmark as a ZS-CIR benchmark is somewhat doubtful. The CIRR text relative captions are often not truly relative (i.e., there exist false positives), and reference images can even be harmful to retrieval. As pointed out by Baldrati et al. [3], this problem becomes more severe when we use a small subset of images, i.e., the subset R@Ks. In summary, due to the noisy nature of CIRR full and subset R@1s, we propose to focus on the CIRR full R@10 scores rather than other metrics. In CIRR full R@10, LinCIR shows the same trends as the other benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 306, + 281, + 548, + 349 + ], + "blocks": [ + { + "bbox": [ + 46, + 433, + 288, + 467 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 288, + 467 + ], + "type": "text", + "content": "Table 6. CIRR results. Due to the noisy nature of CIRR as a ZS-CIR benchmark, we only highlight the R@10 score for the full CIRR set. The detailed discussion can be found in Sec. 4.3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 281, + 548, + 349 + ], + "lines": [ + { + "bbox": [ + 306, + 281, + 548, + 349 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 548, + 349 + ], + "type": "table", + "html": "
Supervision designCIRR dev R@1Fashion IQ R@10R@50
a photo of [$][37]21.6524.9344.35
Ours, but [$] extracted by ψI22.6322.0139.87
Our SMP design choice25.6626.2846.49
", + "image_path": "1e88e9311f7bba67c3bd3bc0262ed56018e55e26e98be6fec0a1fd793ad02e98.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 411, + 537, + 535 + ], + "blocks": [ + { + "bbox": [ + 305, + 357, + 547, + 402 + ], + "lines": [ + { + "bbox": [ + 305, + 357, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 547, + 402 + ], + "type": "text", + "content": "Table 7. Impact of the supervision design. Different target text designs (e.g., \\([\\\\) ]$ sleeps on \\([\\\\( ]\\) in Fig. 3 (b)) affect the performances. \"Ours, but \\([\\\\) ]$ extracted by \\( \\psi_I \\)\" denotes that the textual encoder before the \\( \\phi \\) module is replaced with the visual encoder." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 411, + 537, + 535 + ], + "lines": [ + { + "bbox": [ + 316, + 411, + 537, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 537, + 535 + ], + "type": "table", + "html": "
SMP Masking strategyCIRR devFashion IQ
R@1R@10R@50
All non-keyword tokens20.5919.6136.97
Random token22.9822.4440.92
All noun tokens24.9525.1645.19
1 keyword token23.9226.4245.91
3 keyword tokens24.6626.6946.54
5 keyword tokens25.1426.2846.29
All keyword tokens25.6626.2846.49
", + "image_path": "4beb56c253f1d10efdb859386335db10e22b65b20c1167e232074f0cf9d1108c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 542, + 547, + 598 + ], + "lines": [ + { + "bbox": [ + 305, + 542, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 547, + 598 + ], + "type": "text", + "content": "Table 8. Impact of the SMP masking strategy. The CIRR retrieval performances by varying the masking strategy for SMP (i.e., \"Keyword masking\" in Fig. 3 (b)) are shown. We define \"keyword\" as consecutive adjectives and nouns, except for \"All noun tokens\". \"All noun tokens\" defines keywords as nouns." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 613, + 369, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 613, + 369, + 627 + ], + "spans": [ + { + "bbox": [ + 306, + 613, + 369, + 627 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 632, + 546, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 632, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 632, + 546, + 669 + ], + "type": "text", + "content": "In this subsection, we provide detailed analyses of our design choices. If not specified, we compare the models on the CIRR dev split and FashionIQ test split." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Impact of Self-Masking Projection (SMP) supervision. We compare two variants of SMP in Tab. 7. First, instead of employing our keyword masking strategy, we use “a photo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13231" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 271, + 192 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 271, + 192 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 271, + 192 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 271, + 192 + ], + "type": "table", + "html": "
Noise typeCIRR devFashion IQ
R@1R@10R@50
No noise19.7620.4238.31
Student-t23.0822.8141.00
Exponential23.5125.7845.26
\\( \\chi^2 \\)23.5423.1141.89
N(0,1)23.7023.3141.89
Unif(-1,1)25.1425.8845.78
N(0,1)×Unif(0,1)25.4726.0546.29
", + "image_path": "7d921390bdce3ee57c1dc7757157728a7562b15dfb66572c9be9bdb7fdc2b9e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 259, + 285, + 355 + ], + "blocks": [ + { + "bbox": [ + 46, + 200, + 287, + 244 + ], + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 244 + ], + "type": "text", + "content": "Table 9. Impact of the choice of the random noise. Adding noise to the textual latent space helps to mitigate the inferior generalizability due to the modality gap. Moreover, using better random noise can significantly boost the overall performances." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 259, + 285, + 355 + ], + "lines": [ + { + "bbox": [ + 50, + 259, + 285, + 355 + ], + "spans": [ + { + "bbox": [ + 50, + 259, + 285, + 355 + ], + "type": "table", + "html": "
CC3MSDPCOYOOWTCIRCO mAP@5GeneCIS R@3FashionIQ R@10CIRR R@10Avg
✓XXX13.7232.8025.1164.9534.15
X✓XX9.5232.3823.6362.5832.03
XX✓X11.3631.4826.1865.3333.59
XXX✓9.6730.9024.4164.0532.26
✓✓XX12.5932.3826.2866.6834.48
✓✓X✓10.0633.3621.1163.1631.92
✓✓✓X11.5432.0826.9766.8034.35
", + "image_path": "95489e1d55d5f5beecd0df999d920b24ab654ba2bccebd7ef4e48e5f0ff61d24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 286, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 286, + 418 + ], + "type": "text", + "content": "Table 10. Impact of the training corpus. OpenWebText (OWT) [14] and SD prompts (SDP) are text-only datasets and CC3M [39] and COYO-700M (COYO) [5] are image-text aligned datasets. In our experiments, we use CC3M + SDP for the training corpus, considering the dataset scale and overall CIR performances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 443, + 286, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 286, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 286, + 587 + ], + "type": "text", + "content": "of \\([\\\\) ]\\( as Pic2Word and SEARLE to train the \\(\\phi\\) module. Second, we replace the textual encoder before \\(\\phi\\) module with the visual encoder using the corresponding image of the caption. In the table, \"a photo of \\([\\\\) ]\\( that [cond]\" variant performs worse than our design choice due to the limited diversity of the input texts. Interestingly, we observe that using both image and text pairs for LinCIR performs worse than our design choice. Note that the second model is trained on CC3M image-text pairs without using 2.47M SD prompts as our design choice. We presume it makes the \\(\\phi\\) model overfitted to CC3M image-text relationships, which significantly differ from our target CIR datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 715 + ], + "type": "text", + "content": "Impact of masking design choice. We compare other masking design choices, such as random tokens or non-keyword tokens, with our design choice in Tab. 8. Tab. 8 shows that (1) masking the keywords performs better than masking the others. (2) Our keyword design – i.e., consecutive adjectives and nouns – is better than defining keywords as nouns. (3) The overall performances are enhanced by increasing the masked keywords. As the differences are not significant, we mask all the keywords for simplicity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Impact of the random noise addition. As we discussed in Sec. 3.2, the random noise addition is critical to mitigating the modality gap. Tab. 9 supports this claim: when we do not add any noise, the performance becomes the worst. Our design choice shows the best performance among the other noise designs due to the diverse norm size of our distribution as shown in Fig. B.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "content": "The impact of the training corpus. We evaluate the impact of the training corpus in Tab. 10. We use four corpora, CC3M [39] (3M captions), StableDiffusion Prompts (SDP) (2.47M text prompts), COYO-700M (700M captions) [5] and OpenText (OWT) [14] (8M web texts). CC3M and COYO are image-text paired datasets, and OWT and SDP are text-only datasets. The example training samples of each dataset are shown in the Appendix. In Tab. 10, we observe that models trained with image descriptions (CC3M and COYO) perform better than models trained with general texts because general web texts are often irrelevant to visual information. In addition, using multiple corpora improves the overall performance, e.g., CC3M " + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "inline_equation", + "content": "(34.15)\\rightarrow" + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "content": " CC3M " + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "inline_equation", + "content": "^+" + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "content": " SDP (34.48). Although we observe that using more massive captions (e.g., " + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "inline_equation", + "content": "3\\mathrm{M} + 2.47\\mathrm{M} + 700\\mathrm{M})" + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "content": " can be helpful for some CIR tasks, such as FashionIQ and CIRR, we use CC3M and SDP for our training set considering the corpus size " + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "inline_equation", + "content": "(3\\mathrm{M} + 2.47\\mathrm{M})" + }, + { + "bbox": [ + 304, + 165, + 545, + 380 + ], + "type": "text", + "content": ", and the overall performances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 479 + ], + "type": "text", + "content": "Qualitative results. We provide the additional qualitative retrieval results in the Appendix. In summary, LinCIR shows qualitatively plausible retrieval performance even in a large-scale image database, e.g. LAION-2B. Also, we observe that the retrieved results by LinCIR often suffer from false negatives in benchmark datasets, showing the limitation of the R@1 evaluation on the existing CIR benchmarks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 378, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "We propose a novel zero-shot composed image retrieval (ZS-CIR) framework named Language Only training for Composed Image Retrieval (LinCIR). LinCIR presents a breakthrough in addressing the challenges associated with the previous ZS-CIR methods. By leveraging a novel self-supervision technique, Self-Masking Projection (SMP), LinCIR eliminates the dependency on expensive CIR triplets, opting for a training process solely based on text inputs. This innovative approach significantly enhances scalability and generalizability, overcoming limitations observed in existing ZS-CIR methods. Notably, our LinCIR model achieves remarkable efficiency and ZS-CIR performances compared to other methods on multiple CIR benchmarks, including CIRCO, GeneCIS, FashionIQ, and CIRR. We underscore the effectiveness of our language-only training framework, offering a potent solution with wide-ranging implications for image retrieval tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13232" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 122 + ], + "type": "text", + "content": "[1] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, pages 4959-4968, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 169, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 169, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 169, + 288, + 201 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 4, 5, 6, 7, 11, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 203, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 288, + 255 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 300 + ], + "type": "text", + "content": "[5] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.8, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "type": "text", + "content": "[6] Yanbei Chen and Loris Bazzani. Learning joint visual semantic matching embeddings for language-guided retrieval. In ECCV, pages 136-152. Springer, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 368 + ], + "type": "text", + "content": "[7] Yanbei Chen, Shaogang Gong, and Loris Bazzani. Image search with text feedback by visiolinguistic attention learning. In CVPR, pages 3001-3011, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 411 + ], + "type": "text", + "content": "[8] Yiyang Chen, Zhedong Zheng, Wei Ji, Leigang Qu, and TatSeng Chua. Composed image retrieval with text feedback via multi-grained uncertainty regularization. arXiv preprint arXiv:2211.07394, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 413, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 413, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 413, + 288, + 435 + ], + "type": "text", + "content": "[9] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 287, + 479 + ], + "type": "text", + "content": "[10] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image_caption associations for ms-coco. In ECCV, 2022. 6, 7, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 513 + ], + "type": "text", + "content": "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[12] Eric Dodds, Jack Culpepper, Simao Herdade, Yang Zhang, and Kofi Boakye. Modality-agnostic attention fusion for visual search with text feedback. arXiv preprint arXiv:2007.00145, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 612 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 13" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[14] Aaron Gokaslan, Vanya Cohen, Ellie Pavlick, and Stefanie Tellex. Openwebtext corpus. http://Skylion007.github.io/OpenWebTextCorpus, 2019. 8, 11" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[15] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. arXiv preprint arXiv:2303.11916, 2023. 6, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Sophia Gu, Christopher Clark, and Aniruddha Kembhavi. I can't believe there's no images! learning visual tasks us" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "ing only language supervision. In ICCV, pages 2672-2683, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 150 + ], + "type": "text", + "content": "[18] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 215 + ], + "type": "text", + "content": "[19] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 1, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 271 + ], + "type": "text", + "content": "[20] Surgan Jandial, Ayush Chopra, Pinkesh Badjatiya, Pranit Chawla, Mausoom Sarkar, and Balaji Krishnamurthy. Trace: Transform aggregate and compose visiolinguistic representations for image search with text feedback. arXiv preprint arXiv:2009.01485, 7, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 316 + ], + "type": "text", + "content": "[21] Surgan Jandial, Pinkesh Badjatiya, Pranit Chawla, Ayush Chopra, Mausoom Sarkar, and Balaji Krishnamurthy. Sac: Semantic attention composition for text-conditioned image retrieval. In WACV, pages 4021-4030, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[22] Jongseok Kim, Youngjae Yu, Hoeseong Kim, and Gunhee Kim. Dual compositional learning in interactive image retrieval. In AAAI, pages 1771-1779, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[23] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 123:32-73, 2017. 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 437 + ], + "type": "text", + "content": "[24] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, pages 802-812, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 481 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 2, 6, 13, 14" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[26] Wei Li, Linchao Zhu, Longyin Wen, and Yi Yang. Decap: Decoding CLIP latents for zero-shot captioning via text-only training. In ICLR, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[27] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, 35:17612-17625, 2022. 2, 3, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 624 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 11" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 668 + ], + "type": "text", + "content": "[29] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1, 2, 5, 6, 11" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[30] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, pages 2125–2134, 2021. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13233" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 150 + ], + "type": "text", + "content": "[32] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 681-699. Springer, 2020. 6, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 286, + 184 + ], + "type": "text", + "content": "[33] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 286, + 228 + ], + "type": "text", + "content": "[34] Khoi Pham, Kushal Kafle, Zhe Lin, Zhihong Ding, Scott Cohen, Quan Tran, and Abhinav Shrivastava. Learning to predict visual attributes in the wild. In CVPR, pages 13018-13028, 2021. 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 286, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 286, + 285 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 286, + 341 + ], + "type": "text", + "content": "[36] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 343, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 343, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 343, + 286, + 387 + ], + "type": "text", + "content": "[37] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 4, 6, 7, 11, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 286, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 286, + 475 + ], + "type": "text", + "content": "[38] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Datasets and Benchmarks Track, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 286, + 520 + ], + "type": "text", + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 2, 5, 8, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 565 + ], + "type": "text", + "content": "[40] Minchul Shin, Yoonjae Cho, Byungsoo Ko, and Geonmo Gu. Rtic: Residual learning for text and image composition using graph convolutional network. arXiv preprint arXiv:2104.03015, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 567, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 609 + ], + "type": "text", + "content": "[41] Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Hua-jun Bai, and Yoav Artzi. A corpus for reasoning about natural language grounded in photographs. arXiv preprint arXiv:1811.00491, 2018. 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 286, + 644 + ], + "type": "text", + "content": "[42] Yuxin Tian, Shawn Newsam, and Kofi Boakye. Image search with text feedback by additive attention compositional learning. arXiv preprint arXiv:2203.03809, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "text", + "content": "[43] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 13" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[44] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5, 6, 11" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 251 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[45] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. arXiv:2308.14746, 2023. 6, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[46] Nam Vo, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval—an empirical odyssey. In CVPR, pages 6439–6448, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[47] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, pages 11307-11317, 2021. 1, 2, 5, 6, 11, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[48] Youngjae Yu, Seunghwan Lee, Yuncheol Choi, and Gunhee Kim. Curlingnet: Compositional learning between images and text for fashion iq data. arXiv preprint arXiv:2003.12299, 2020. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13234" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_content_list.json b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7dd1c9724a0b69c4a937ae50307f45248f6c2558 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_content_list.json @@ -0,0 +1,1696 @@ +[ + { + "type": "text", + "text": "Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Atefeh Khoshkhahtinat, Ali Zafari, Piyush M. Mehta, Nasser M. Nasrabadi \nWest Virginia University, WV, USA", + "bbox": [ + 186, + 180, + 781, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ak00043, az00004}@mix.wvu.edu, {piyush.mehta, nasser.nasrabadi}@mail.wvu.edu", + "bbox": [ + 158, + 218, + 810, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While replacing Gaussian decoders with a conditional diffusion model enhances the perceptual quality of reconstructions in neural image compression, their lack of inductive bias for image data restricts their ability to achieve state-of-the-art perceptual levels. To address this limitation, we adopt a non-isotropic diffusion model at the decoder side. This model imposes an inductive bias aimed at distinguishing between frequency contents, thereby facilitating the generation of high-quality images. Moreover, our framework is equipped with a novel entropy model that accurately models the probability distribution of latent representation by exploiting spatio-channel correlations in latent space, while accelerating the entropy decoding step. This channel-wise entropy model leverages both local and global spatial contexts within each channel chunk. The global spatial context is built upon the Transformer, which is specifically designed for image compression tasks. The designed Transformer employs a Laplacian-shaped positional encoding, the learnable parameters of which are adaptively adjusted for each channel cluster. Our experiments demonstrate that our proposed framework yields better perceptual quality compared to cutting-edge generative-based CODECs, and the proposed entropy model contributes to notable bitrate savings. The code is available at https://github.com/Atefeh-Khoshtinat/Blurdissipated-compression.", + "bbox": [ + 75, + 300, + 473, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 723, + 209, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image compression is a crucial tack in image processing which aims to decrease the amount of data required for storing or transmitting without significant loss of visual content. Recently, learning-based image compression methods [13, 25, 42, 45] have demonstrated the potential to surpass classical hand-engineered CODECs in terms of rate-distortion performance. Learned image compression commonly comprises three steps: transformation, quantization, and lossless entropy coding, which resemble the components found in the traditional transform coding paradigm [44].", + "bbox": [ + 75, + 750, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6f6058ebfc2b1442af93af3be3744a15ea45ea0738d498262b3ab1647638a44.jpg", + "image_caption": [ + "Figure 1. Overview of our proposed neural codec. The quantized semantic latent variable $\\hat{\\pmb{y}}$ is utilized by a diffusion-based decoder to generate realistically reconstructed image." + ], + "image_footnote": [], + "bbox": [ + 526, + 268, + 864, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural image compression methods are typically trained with two primary objectives: minimizing distortion between the original input image and its reconstructed version and reducing the bitrate for efficient data transmission or storage [38]. However, solely optimizing for rate-distortion may lead to blurry reconstructions. To address this issue, various generative-based CODECs [4, 11, 28, 42] have emerged, aiming to generate reconstructions that are not only faithful to the original data but also visually realistic to human observers.[28] introduced HiFiC, an image compression network based on GANs that aims to achieve a balance between fidelity and perceptual quality. The network demonstrated substantial advantages for human observers, surpassing BPG even when using half the bits. However, HiFiC faces challenges due to the instability of GAN training and necessitates various design adjustments. Inspired by the achievements of Denoising Diffusion Probabilistic Models (DDPMs)[17] in generative modeling, authors in [42] replaced the Gaussian decoder of VAEs with a conditional diffusion model to mitigate the issue of blurriness in decoded images. Despite the impressive performance of diffusion-based network, its diffusion and denoising processes do not explicitly incorporate the inductive biases inherent to natural images. The imposing inductive biases,", + "bbox": [ + 496, + 537, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as the multi-scale nature of images, enable the generation of high-quality samples [20].", + "bbox": [ + 78, + 90, + 465, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to enhancing the quality of reconstructed images, accurate entropy estimation of latent representations is vital for boosting compression efficiency. The more the estimated entropy gets closer to the actual entropy of the latent, the lower the bitrate required for generating bitstream file. In this respect, several works have been proposed. Balle et al.[5] introduced an entropy model that is conditioned on an additional latent variable, named the hyperprior, to capture the existing spatial redundancies within the latent space. Studies [24, 30] introduced a sequential autoregressive block into the entropy model. This block leverages causally adjacent latent elements to estimate the distribution of the current latent, which results in slowing down of the decoding step. He et al.[13] utilized a two-way parallel context model as a substitute for the sequential autoregressive context model, aiming to accelerate the decoding of latent code. In another realm of studies, to mitigate the slow decoding times, authors in [29] introduced channel-wise context modeling as an alternative to serial spatial context. To expedite the extracting of the channel-conditional context model, He et al. [14] exploited an unevenly segmented channel-wise autoregressive model, wherein the channels are divided into varying sizes. Strengthening the entropy model through self-attention was initially proposed in [32]. This approach employed a global reference model to identify the most suitable spatial context, which doesn't have to be confined to the local scope. The idea of integrating global context into the entropy model was subsequently refined in a later work [33], where a specially crafted transformer-based entropy model was employed. This entropy model is equipped with the diamond-relative position encoding to effectively model long-distance dependencies. The diamond-shaped relative positional encoding is devised based on the clip function, which relies on the pre-defined threshold to characterize an effective receptive field. This approach is effective in computing context based on the positions of latent elements; yet, it cannot provide an optimal receptive field due to its reliance on the clip function. Recent works[21, 23] focused on developing adaptive context models to extract both channel and spatial correlations within the latent space while expediting the entropy decoding step. Authors in [21] proposed an entropy model named MEM, aiming to capture both local and global spatial correlations in each channel segment by using a parallel bidirectional context approach. However, despite MEM's attention to spatio-channel correlations, it appears to overlook the crucial role of positional encoding in computing long-range spatial dependencies.", + "bbox": [ + 80, + 123, + 467, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose a conditional diffusion-based decoder in the neural compression pipeline to enhance the realism of reconstructed", + "bbox": [ + 78, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images. The traditional diffusion models are generalized into a conditional non-isotropic diffusion model to incorporate an inductive bias, considering the relative importance of each frequency component of images. As a result, each frequency component of an image undergoes diffusion at distinct rates, leading to the generation of decoded images in a coarse-to-fine manner. Furthermore, a novel channel-conditional autoregressive entropy model is introduced to efficiently take into account both local and global spatial dependencies within each channel chunk of the latent space. To leverage global spatial context, a Transformer block is designed which incorporates a Laplacian-shaped positional encoding within checkerboard-shaped self-attention module. The effective Laplacian-shaped receptive field for each channel chunk is dynamically determined by learning the positional encoding parameters during the optimization of the entire neural codec.", + "bbox": [ + 501, + 90, + 890, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 503, + 378, + 645, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Learned Image Compression", + "text_level": 1, + "bbox": [ + 503, + 411, + 754, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural image compression methods adopt a non-linear transform coding paradigm which is founded on the variational autoencoders (VAEs) [12]. Within this scheme, the encoder initially maps the input image $\\pmb{x}$ to a compact latent representation $\\pmb{y}$ . Subsequently, quantization is applied to this latent representation, leading to discrete symbols that minimize the bit requirement. Afterward, entropy coding generates bit strings using an entropy model. Ultimately, the reconstructed image $\\hat{\\pmb{x}}$ is generated by applying the decoder to the quantized latent representation $\\hat{\\pmb{y}}$ . This framework is commonly trained with a trade-off between the rate and distortion, such as Mean Squared Error (MSE) [38].", + "bbox": [ + 501, + 441, + 890, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies have attempted to enhance the realism of reconstructed images by optimizing neural CODEs using a triple rate-distortion-perception loss [4, 28, 43]. Blau and Michaeli [7] explored the essential balance between distortion and realism, demonstrating that within a given rate, improving distortion comes at the cost of decreasing the perceptual quality of a reconstructed image. In line with this premise where perceptual quality is measured as the difference between the image distribution and the distribution of reconstructions, Mentzer et al. [28] employed a conditional GAN within an autoencoder-based compression model to enhance perceptual quality. In a similar vein, [15] introduced a model that aimed to improve the realism by enriching the loss function with an adversarial perceptual loss using LPIPS [46], along with a patch-based style loss [10]. Recently, Agustsson et al. [4] developed a Multi-Realism model that merges the ELIC codec [14] and PatchGAN [8], showing improved performance compared to HiFiC.", + "bbox": [ + 501, + 628, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3046", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Models", + "text_level": 1, + "bbox": [ + 76, + 90, + 246, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [17, 22, 37], which belong to the family of score-based generative models, acquire the data distribution through a gradual iterative denoising process, starting from a Gaussian distribution and progressing towards the actual data distribution. Recently, they have received considerable attention owing to their training stability and high quality image generation compared to GANs. Diffusion models are adopted in various domains, including image and video generation [18, 19, 34], super-resolution [36], inpainting [27], deblurring [41], compression [11, 39, 42].", + "bbox": [ + 75, + 112, + 468, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the domain of image compression, several diffusion-based CODEs have been proposed. The research conducted by Theis et al. [39] involved employing a general unconditional diffusion model for transmitting Gaussian samples in a lossy manner. Their approach leverages a reverse channel coding concept, which stands apart from the conventional transform coding scheme. Although their methodology performs well, it suffers from high computational cost which makes it unfeasible for dealing with high-resolution images. Yang and Mandt [42] introduced a neural codec, called CDC, in which the decoder takes the form of a DDPM, conditioned on a quantized latent representation. Ghouse et al. [11] began by optimizing an autoencoder network using a rate-distortion loss. Subsequently, they train a conditional diffusion model on the output of the decoder, with the goal of enhancing its perceptual quality.", + "bbox": [ + 75, + 265, + 472, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Neural Entropy Model", + "text_level": 1, + "bbox": [ + 76, + 513, + 290, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main purpose of the entropy model is to estimate the joint probability distribution over the quantized latent representation. When the learned entropy model precisely matches the true distribution of the latent representation, a lower bit-rate is required to generate a compressed file. Entropy estimation can take advantage of two key principles: backward and forward adaptation. Forward adaptation uses an extra latent variable, called hyper-prior [5], as a side information to capture spatial dependencies between elements of the latent representation. Backward adaptation, on the other hand, employs an autoregressive-based context model which leverages the previously decoded elements of the latent to predict the probability of the current element. The context model of backward methods can extract relationships between the current symbol and previously decoded symbols in various dimensions, including spatial and channel axes [13, 14, 21, 23, 29, 30, 32, 33, 47].", + "bbox": [ + 75, + 537, + 468, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Minnen et al.[30] incorporated a local spatial context with a hyper-prior network to precisely predict distribution of the latent. This type of context model can exploit local correlations between the current latent and its neighboring causal elements using masked convolutions, which consequently necessitates the adoption of serial decoding. To expedite the decoding process, He et al.[13] evenly parti", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tioned the latent into two groups: anchors and non-anchors. The anchors are exclusively decoded using the hyper-prior, while the non-anchors utilize both the hyper-prior and the local context model. This approach allows for the parallel decoding of both anchors and non-anchors.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative strategy for parallelizing the decoding process is to extract inter-channel dependencies. In [29] the latent code is divided into evenly sized chunks along the channel dimension, with each of these chunks being decoded using information from previously decoded ones. To speed up the channel-wise context extraction, He et al.[14] partitioned channels of the latent into an uneven set of groups, allocating fewer channels to the initial groups and more channels to the subsequent ones. This uneven division is motivated by the observation that the earlier channels possess higher entropy compared to the remaining channels.", + "bbox": [ + 496, + 167, + 892, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some works introduced adaptive context models which focus on capturing long-range spatial correlations. Qian et al.[32] devised a global reference context model which assesses throughout previously decoded elements to detect the most similar one to the current latent. The authors in [33] employed a Transformer-based entropy model, called Entroformer, for capturing long-range contexts. The Entroformer benefits from diamond-relative position encoding and a top-k self-attention mechanism, both of which contribute to providing an efficient receptive field.", + "bbox": [ + 496, + 333, + 890, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 498, + 599, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we develop our decoder based on the blurring diffusion model [20], which is conditioned on the discrete latent representation and enables the differentiation among the frequency components of an image throughout the diffusion and denoising processes. The architecture of decoder will be explained in Appendix. Moreover, a novel entropy model is employed that efficiently encodes the quantized latent representation into a binary stream.", + "bbox": [ + 496, + 523, + 890, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Blurring Diffusion Model", + "text_level": 1, + "bbox": [ + 500, + 654, + 732, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can define the blurring diffusion model as a traditional diffusion model in the frequency space, but with distinct schedules for each dimension of data [20]. The specifics of the schedules for vectors $\\alpha_{t}$ and $\\sigma_{t}$ are crucial in this type of model, which will be discussed.", + "bbox": [ + 496, + 676, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Process: The diffusion process [17] progressively degrades the clean image, transforming it into nearly pure Gaussian noise over the course of T time steps. Each step of the diffusion process in frequency space can be formulated as:", + "bbox": [ + 496, + 753, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {f} _ {t} \\mid \\boldsymbol {f} _ {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t}; \\alpha_ {t} \\boldsymbol {f} _ {\\boldsymbol {x}}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 840, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{t}$ and $f_{x}$ are defined as $\\mathbf{V}^T\\mathbf{z}_t$ and $\\mathbf{V}^T\\mathbf{x}$ , respectively. In such a model, each frequency component could", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "undergo diffusion at varying rates, with the modulation of these rates being determined by the values present in the vectors $\\alpha_{t}$ and $\\sigma_{t}$ .", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Denoising Process: The true denoising process in the frequency space, which is equal to the deblurring operation in the time domain, can be expressed as:", + "bbox": [ + 75, + 136, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\boldsymbol {f} _ {x}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1}, \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 188, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{t\\rightarrow t - 1} = \\sigma_{t|t - 1}\\sigma_{t - 1} / \\sigma_t$ and $\\mu_{t\\to t - 1} = (\\alpha_{t|t - 1}\\sigma_{t - 1}^2 /\\sigma_t^2)f_t + (\\alpha_{t - 1}\\sigma_{t|t - 1}^2 /\\sigma_t^2)f_x.$ The actual denoising process can be approximated using a learned denoising distribution to generate data. Thus, by considering the epsilon reparametrization of Eq. 2, we can derive the expression for the learned denoising process in frequency space:", + "bbox": [ + 76, + 212, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1} \\left(\\hat {\\boldsymbol {f}} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {t}\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 328, + 468, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Re-parametrization: $\\hat{f}_{\\pmb{x}} = (1 / \\alpha_{t})(f_{t} - \\sigma_{t}\\hat{f}_{\\epsilon})$ (3)", + "bbox": [ + 116, + 349, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By substituting the approximation of $\\hat{f}_x$ into the true denoising distribution, the mean of the denoising network can be derived by $(\\alpha_{t|t-1}\\sigma_{t-1}^2/\\sigma_t^2)\\pmb{f}_t + (\\alpha_{t-1}\\sigma_{t|t-1}^2/\\sigma_t^2\\alpha_t)(\\pmb{f}_t - \\sigma_t\\hat{f}_\\epsilon)$ .", + "bbox": [ + 76, + 375, + 468, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization: According to [17], the loss function is simplified as below:", + "bbox": [ + 75, + 454, + 468, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\boldsymbol {f} _ {\\epsilon} - \\hat {\\boldsymbol {f}} _ {\\epsilon} \\| ^ {2} ] = \\boldsymbol {E} [ \\| \\boldsymbol {V} ^ {\\boldsymbol {T}} (\\epsilon - \\hat {\\epsilon}) \\| ^ {2} ] \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\epsilon - \\hat {\\epsilon} \\| ^ {2} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 489, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)$ and $z_t = V(\\alpha_t V^T x + \\sigma_t V^T \\epsilon)$ . The $\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)$ signifies that the neural network $\\phi_{\\theta}$ takes $z_t$ as input and predicts $\\hat{\\epsilon}$ . This formulation satisfies the requirement for neural networks to perform effectively in a standard pixel space. After prediction, it is sufficient to transition back and forth between frequency space and pixel space using the DCT matrix $V^T$ and its inverse $V$ .", + "bbox": [ + 75, + 521, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Schedules of Blurring Diffusion Model", + "text_level": 1, + "bbox": [ + 76, + 635, + 406, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The schedules for the blurring diffusion model, denoted as $\\alpha_{t}$ and $\\sigma_{t}$ , are obtained by combining a standard Gaussian noise diffusion schedule (with scalar values $\\sigma_{t}$ and $\\alpha_{t}$ ) along with a blurring schedule $d_{t}$ . Each element of the vector $\\sigma_{t}$ shares the same value, as identical noise is added to all frequency components. Therefore, it becomes adequate to present a schedule for a scalar value $\\sigma_{t}$ [20]. The noise schedule is chosen based on a variance-preserving cosine [31], specifically $\\sigma_{t}^{2} = 1 - \\alpha_{t}^{2}$ , where $\\alpha_{t} = \\cos(t\\pi / 2T)$ for $t \\in [0, T]$ . Following [35], the blurring schedule is then defined as:", + "bbox": [ + 75, + 657, + 468, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {B, t} = \\sigma_ {B, m a x} \\sin (t \\pi / 2 T) ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 830, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{B,max}$ represents a hyperparameter equal to the maximum level of blur applied to the image. This schedule, in turn, corresponds to the dissipation time through", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\tau_{t} = \\sigma_{B,t}^{2} / 2$ . Based on the formulation discussed in the Appendix, the blurring schedule $d_{t}$ , which is employed for $\\alpha_{t}$ , is defined as follows:", + "bbox": [ + 500, + 90, + 890, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} _ {\\boldsymbol {t}} = \\exp (- \\lambda \\tau_ {t}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 148, + 890, + 164 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ represents the vector containing squared frequencies, and $\\tau_{t}$ corresponds to the dissipation time. To achieve a more gradual amplification of high frequencies during the denoising process, the blurring schedule $d_{t}$ is adjusted to $(1 - d_{min})\\exp (-\\lambda \\tau_{t}) + d_{min}$ , where $d_{min}$ is set to 0.001. Finally, by combining the Gaussian noise schedule with the blurring schedule, resultant schedule is as follows:", + "bbox": [ + 498, + 176, + 890, + 281 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\alpha} _ {t} = \\alpha_ {t}. \\boldsymbol {d} _ {t}, \\boldsymbol {\\sigma} _ {t} = \\mathbf {1} \\sigma_ {t}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 294, + 890, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Blurring Diffusion Model for Compression", + "text_level": 1, + "bbox": [ + 500, + 320, + 862, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The rate-distortion objective in end-to-end compression resembles the loss function of a $\\beta$ -VAE, where a hyperparameter $\\lambda$ is utilized to balance the trade-off between the bit-rate $(R)$ and distortion $(D)$ :", + "bbox": [ + 500, + 343, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} = D + \\lambda R = \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 415, + 890, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While many neural codecs commonly employ Gaussian or Laplacian decoders, we introduce a novel approach using a conditional blurring diffusion model as the decoder. This approach aims to yield new distortion metrics that deviate from those based on Mean Squared Error (MSE) or Mean Absolute Error (MAE). In this framework, our proposed neural codec leverages two distinct types of latent variables: a semantic latent variable $\\mathbf{y}$ and texture latent variables $\\mathbf{z}_{1:T}$ . The semantic latent variable captures and encodes the overall content and meaning of the image. In contrast, the texture latent variables are tailored to carry finer details and intricate patterns that might not be fully represented by the semantic variable alone. Notably, unlike the semantic latent variable, the texture latent variables are not compressed but synthesized during the decoding phase.", + "bbox": [ + 500, + 443, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The forward and backward processes of the conditional blurring diffusion model can be expressed as:", + "bbox": [ + 500, + 670, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq (\\pmb {f} _ {t} | \\pmb {f} _ {t - 1}) = \\mathcal {N} (\\pmb {f} _ {t}; \\alpha_ {t | t - 1} \\pmb {f} _ {t - 1}, \\sigma_ {t | t - 1} ^ {2} I),\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 723, + 803, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {\\boldsymbol {1}: T} | \\tilde {\\boldsymbol {y}}\\right) = p \\left(\\boldsymbol {f} _ {\\boldsymbol {T}}\\right) p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}} \\mid \\boldsymbol {f} _ {\\boldsymbol {1}}, \\tilde {\\boldsymbol {y}}\\right) \\prod_ {t = 2} ^ {T} p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {t} - \\boldsymbol {1}} \\mid \\boldsymbol {f} _ {\\boldsymbol {t}}, \\tilde {\\boldsymbol {y}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 746, + 887, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {f} _ {T}) = \\mathcal {N} (\\boldsymbol {f} _ {T}; \\mathbf {0}, \\boldsymbol {I})\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 789, + 651, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {\\theta} \\left(\\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}, t\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 809, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{t} = V^{T}z_{t}$ and $f_{x} = V^{T}x = V^{T}z_{o}$ . As shown in Eq. 8, distortion is equivalent to the negative marginal likelihood of input data $-\\log p_{\\boldsymbol{x}|\\tilde{\\boldsymbol{y}}}(\\boldsymbol{x}|\\tilde{\\boldsymbol{y}})$ , and minimizing it", + "bbox": [ + 500, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is analogous to minimizing the negative marginal likelihood of the frequency representation of the data $p_{\\pmb{f}_x|\\tilde{\\pmb{y}}}(\\pmb{f}_x|\\tilde{\\pmb{y}}) = \\int p(\\pmb{f}_x, \\pmb{f}_{1:T}|\\tilde{\\pmb{y}}) d\\pmb{f}_{1:T}$ . Since computing the marginal likelihood is intractable, we employ its ELBO with the specified diffusion and de-blurring distributions. Following this substitution, the rate-distortion objective, through the application of Jensen's inequality, can be formulated as follows:", + "bbox": [ + 76, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right] \\\\ \\leq \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} [ - \\log_ {P} \\boldsymbol {x} | \\boldsymbol {y} (\\omega | \\boldsymbol {y}) - \\lambda \\log_ {P} \\boldsymbol {y} (\\boldsymbol {y}) ], \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 219, + 468, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{ELBO} = E_q[\\frac{p_\\theta(f_x,f_{1:T}|\\tilde{y})}{q(f_t|f_{t - 1})} ]$ . As explained in Appendix, the optimizing of ELBO can be simplified to:", + "bbox": [ + 76, + 266, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E L B O} \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} ], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 311, + 468, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{t} = V\\alpha_{t}V^{T}x + \\sigma_{t}\\epsilon$ . Similar to [28], an LPIPS loss [46] is incorporated into the rate-distortion loss to enhance the perceptual quality of the reconstructed image. As the initial image at any time step can be decoded as a function of the texture latent variable $z_{t}$ , the semantic latent variable $y$ , and the time step $t$ , i.e., $\\hat{x}_t = V(1 / \\alpha_t)(V^T z_t - \\sigma_t V^T \\hat{\\epsilon})$ , the total loss becomes as:", + "bbox": [ + 76, + 339, + 472, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ (1 - \\beta) \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} \\tag {12} \\\\ \\left. - \\lambda \\log p _ {\\tilde {\\mathbf {y}}} (\\tilde {\\mathbf {y}}) + \\beta d _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {t}) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 455, + 468, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\lambda$ and $\\beta$ represent hyperparameters that control the trade-off between rate, distortion, and perception.", + "bbox": [ + 76, + 502, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoding Process: After the training, we utilize entropy decoding on $\\hat{\\pmb{y}}$ through the entropy model which estimates the distribution $p_{\\hat{\\pmb{y}}}(\\hat{\\pmb{y}})$ . We then employ ancestral sampling to conditionally decode the image $\\pmb{x}$ , which is equal to $\\pmb{z_0}$ and starts from pure Gaussian noise $\\pmb{z}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})$ :", + "bbox": [ + 76, + 532, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t - 1} \\gets \\boldsymbol {V} (\\hat {\\boldsymbol {\\mu}} _ {t \\rightarrow t - 1} + \\boldsymbol {\\sigma} _ {t \\rightarrow t - 1} \\boldsymbol {V} ^ {T} \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\hat {\\boldsymbol {y}})) \\quad (1 3)\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 618, + 468, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mu}_{t\\rightarrow t - 1} = \\frac{\\alpha_{t|t - 1}\\sigma_{t - 1}^2}{\\sigma_t^2} V^T z_t + \\frac{\\sigma_{t|t - 1}^2}{\\alpha_{t|t - 1}\\sigma_t^2} (V^T z_t - \\sigma_t V^T\\phi_\\theta (z_t,t,\\hat{y}))$ . The latent variables $z_{1:T}$ are not stored but produced during the decoding.", + "bbox": [ + 76, + 646, + 468, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Proposed Entropy Model", + "text_level": 1, + "bbox": [ + 76, + 710, + 307, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The main goal of the proposed entropy model' context is to exploit both channel-wise and spatial correlations, while expediting the decoding process. Inspired by the ELIC [14], we adopt an uneven grouping of latent channels, allowing most low entropy channels to depend on high entropy channels. The latent representation $\\hat{\\pmb{y}}$ , with $M$ channels, is clustered into five chunks along the channel dimension: 16, 16, 32, 64, and M - 128 channels, respectively. In this setup, each chunk depends on all its previous decoded chunks.", + "bbox": [ + 76, + 734, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Fig. 2(a) illustrates our proposed entropy model as applied in the $j$ -th chunk. Within each chunk, we use spatial", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1dc4e0096d8a53b6a62039eb526558c3fb106b94c6036b388ac4daf4f273c896.jpg", + "image_caption": [ + "Figure 2. Diagram illustrating the application of the proposed entropy model for decoding the $j$ -th chunk $\\hat{\\pmb{y}}^{(j)}$ . (b) Global Spatial Context Block. (c) An example of a checkerboard-shaped mask." + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 885, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "context in conjunction with channel context to model correlations along both the channel and spatial dimensions. To expedite the decoding step, we employ a parallel bidirectional spatial context model which is capable of capturing both local and global spatial relationships. So, the anchor part is decoded in parallel by using solely the hyperprior and channel context, while the decoding of the non-anchor part relies on the hyperprior, as well as both the spatial and channel contexts.", + "bbox": [ + 498, + 354, + 890, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Spatial Context", + "text_level": 1, + "bbox": [ + 500, + 512, + 663, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The spatial context design captures both local and global spatial correlations within the latent representation $\\hat{\\pmb{y}}$ . Following the decoding of the anchor group, a checkerboard-shaped convolution is applied to this group, generating local context for all non-anchor elements in a parallel manner. Furthermore, the acquired local contexts of the non-anchor part are subsequently fed into a Transformer-based block to efficiently extract the global spatial context. This Transformer block takes advantage of positional encoding that is customized specifically for the compression task, along with a checkerboard-based attention mechanism.", + "bbox": [ + 498, + 537, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Transformer-based Spatial Context", + "text_level": 1, + "bbox": [ + 500, + 726, + 800, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the Swin Transformer [26] blocks, including a masked window-based multi-head self-attention (W-MSA) and a masked shifted-window-based multi-head self-attention (SW-MSA), as our global spatial context model which shown in Fig. 2.(b). This selection enables us to strike a balance between computational efficiency and modeling capacity. The checkerboard-based self-attention in W-MAS and SW-MSA can be expressed:", + "bbox": [ + 498, + 750, + 890, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA t t e n (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} (\\boldsymbol {Q} \\boldsymbol {K} ^ {T} \\odot \\boldsymbol {M}) \\boldsymbol {V}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3049", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bc3daba084e255fffb60e34138a9a2450ed7c96c5029372656da8fb4e68d630.jpg", + "image_caption": [ + "Figure 3. The procedure for acquiring Laplacian relative position encoding for a window with a size of $2 \\times 2$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 89, + 472, + 164 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $Q, K, V \\in R^{N^2 \\times d}$ are the query, key, and value matrices, respectively. $N^2$ denotes the number of patches in a window and $d$ is the dimension of query/key/value. $M \\in R^{N^2 \\times N^2}$ represents a checkerboard-shaped mask.", + "bbox": [ + 76, + 229, + 470, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4.3 Relative Positional Encoding", + "text_level": 1, + "bbox": [ + 76, + 311, + 331, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The authors in [13, 33] have shown that the bitrate is impacted by the distance of neighboring latents within the spatial context modeling of current latent. Based on this insight, we introduce a receptive field-aware self-attention mechanism that employs a learnable Laplacian-shaped positional encoding for calculating spatial context. In our proposed approach, each channel chunk possesses its own receptive field during self-attention computation, which dynamically adjusts in response to changes in entropy of chunks. In the $j$ -th chunk, the checkerboard-based attention of W-MSA and SW-MSA is refined to receptive field-aware self-attention as follows:", + "bbox": [ + 75, + 335, + 468, + 515 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA t t e n \\left(\\boldsymbol {Q} _ {j}, \\boldsymbol {K} _ {j}, \\boldsymbol {V} _ {j}\\right) = \\text {s o f t m a x} \\left(\\boldsymbol {Q} _ {j} \\boldsymbol {k} _ {j} ^ {T} \\odot \\boldsymbol {M} + \\boldsymbol {P} _ {\\text {L a p}, j}\\right) \\boldsymbol {V} _ {j}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 526, + 478, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{Lap,j}$ represents the learnable Laplacian relative position encoding for chunk $j$ . To create the Laplacian relative position encoding for the $j$ -th chunk, which comprises spatial windows with a size of $N \\times N$ (i.e. including $N^2$ patches), three steps need to be taken. Firstly, We generate a 2D relative position table, where each coordinate of the relative position lies in the range $[-N + 1, N - 1]$ . Subsequently, we derive the relative position matrix for each patch. As shown in Fig. 3, the first patch's relative distance coordinate of $(0,0)$ (relative distance with its position) is located at the top-left of the orange box, while the last patch's relative distance of $(0,0)$ is positioned at the bottom-right of the green box. Afterward, each relative position matrix is flattened and stacked together. Finally, we apply the 2D Laplacian function to each element of the resulting matrix to generate a Laplacian relative position encoding $P_{Lap,j}$ with learnable parameters $A_j$ and $\\sigma_j$ , whose size is $R^{N^2 \\times N^2}$ . The 2D Laplacian function is defined as:", + "bbox": [ + 75, + 558, + 468, + 832 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {L a p} (x, y) = A _ {j} ^ {2} \\exp \\left(\\left(- 1 / 2 \\sigma_ {j} ^ {2}\\right) \\left(| x | + | y |\\right)\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 842, + 468, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x\\in \\{-N + 1,\\dots,N - 1\\}$ $y\\in \\{-N + 1,\\dots,N - 1\\}$ $A_{j}$ and $\\sigma_{j}$ are learnable parameters which are determined", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for each chunk through optimization. The value of them is associated with the wideness of the effective receptive field.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 135, + 630, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 160, + 717, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To train our learned image compression network, we utilize a merged dataset that includes high-resolution images from the DIV2K, Flickr2K, and CLIC [1] datasets. These images are randomly cropped to a size of $256 \\times 256$ during the training phase. The model parameters of all architectures were optimized using the Adam optimizer for a total of 2.4 million steps, with a batch size of 8. The initial learning rate was configured to be $1 \\times 10^{-4}$ and was progressively reduced until the conclusion of training, reaching $1 \\times 10^{-7}$ . In order to cover a broad spectrum of bitrates, we selected a hyperparameter $\\lambda$ from the set $\\{0.0004, 0.005, 0.01, 0.02, 0.04, 0.016\\}$ . The hyperparameter $\\beta$ which specifies the contribution of perceptual loss is considered 0.9 for all models. The parameter $T$ , represents the required time steps for diffusion-based decoder, is consistently set to 500. To evaluate the performance of our compression approach, we select the Kodak dataset [2], which comprises 24 high-quality images with a resolution of $768 \\times 512$ , and the CLIC2020 test set (428 images) with varying resolutions.", + "bbox": [ + 496, + 183, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the SOTA Methods", + "text_level": 1, + "bbox": [ + 500, + 496, + 818, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our model with state-of-the-art generative based image compression networks, including both GAN-based and diffusion-based CODECs, as well as a state-of-the-art hand-crafted codec such as VVC-Intra (VTM) [3]. Additionally, we choose two VAE CODECs for comparison: DGML [9] and NSC [40]. Specifically, within the domain of GAN-based image compression approaches, we select HiFiC [28] and Multi-Realism [4] frameworks for comparison. Furthermore, we assess our method against diffusion-based models, namely DIRAC [11] and CDC [42]. This comparison is performed with respect to both rate-distortion and rate-perception tradeoff. The distortion is measured by the Peak Signal-to-Noise Ratio (PSNR) metric and perceptual quality are quantified via Frechet Inception Distance (FID) [16] and Learned Perceptual Image Patch Similarity (LPIPS). As shown in Fig. 4, our model shows superior performance compared to all other CODECs in terms of rate-perception. However, this heightened realism comes at the cost of distortion where other methods, except for the CDC model, either exhibit better performance or remain competitive.", + "bbox": [ + 496, + 518, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Visual Quality", + "text_level": 1, + "bbox": [ + 500, + 845, + 648, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 illustrates reconstructed images (kodim20.png, kodim07.png) generated by our proposed model, as well as", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3050", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b16130257453f3e6bd81f3c6bf6fb5783500bca1d0bbf3ee1e6c92c937eea3f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 359, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fc7343380d5f1c99e4da77b0fe572a6cfc938c3c8f28e81d984a416d71387e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 89, + 602, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7978237301d02da3664285e7c4354a1fd70900e1911819e114cfa06631546f77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 90, + 852, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5327daa86c66bf863571b6ee4e7eb7e5d8ec87bd1955df156ac8255918736bf1.jpg", + "image_caption": [ + "Figure 4. Comparison of our method with other codescs in terms of rate/distortion [bpp $\\downarrow$ / PSNR $\\uparrow$ ] and rate-perception, including [bpp $\\downarrow$ / FID $\\downarrow$ ] and [bpp $\\downarrow$ / LPIPS $\\downarrow$ ], for both the CLIC2020 test set and the Kodak dataset." + ], + "image_footnote": [], + "bbox": [ + 112, + 248, + 359, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0253b50e4b9e5ee61a10785eaa8e5ff199ff6660421dbac8c917b81a4f8c5a96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 250, + 601, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ab8a502bce3b38a86f10bd620f2c809d014fc5486ade6a3ace9009b94899bca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 250, + 849, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f568de16a342910a3ca170440f73065a3194f08167543fb960aa67f4e4eb491c.jpg", + "image_caption": [ + "Figure 5. Visual comparison of our method to HiFiC and CDC models shows that our model achieves superior reconstruction quality, particularly at lower bit-rates. In addition, our model displays fewer artifacts compared to both the HiFiC and CDC models." + ], + "image_footnote": [], + "bbox": [ + 78, + 468, + 890, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the CDC [42] and HiFiC [28] frameworks. Perceptually-oriented neural codecs, such as HiFiC and CDC, are criticized when their decoding process encounters semantic content, such as text within the compressed image. Our model effectively addresses this challenge, as depicted in Fig. 5. It enhances perceptual quality through a diffusion-based decoder and imposes a strict inductive bias via the", + "bbox": [ + 75, + 790, + 470, + 896 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "de-blurring diffusion process, resulting in both high perceptual quality and better text-preserving reconstructions. Additional qualitative comparisons are provided in the Appendix.", + "bbox": [ + 498, + 790, + 890, + 852 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3051", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 90, + 230, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Maximum Blurring: To explore the impact of the blurring schedule, we vary the values of $\\sigma_{B,max}$ and compare the performance of the resulting models. It is evident that the model with $\\sigma_{B,max} = 0$ is equivalent to a standard denoising diffusion model. As indicated in Table 1a, blurring diffusion models with higher maximum blur levels $\\sigma_{B,max} = 25$ generate higher-quality reconstructed images compared to other variant models in terms of FID score.", + "bbox": [ + 75, + 113, + 468, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ced95dbeffe01e700869746a390eeb2c4b58982c09a47244b1d64265cab499e.jpg", + "table_caption": [ + "Table 1. Ablation Study: all models are optimized with $\\lambda = {0.01}$ . (a) Investigating the impact of maximum blurring ${\\sigma }_{B,\\max }$ . (b) Exploring the effects of different types of positional encoding" + ], + "table_footnote": [], + "table_body": "
σB,maxFID
03.94
53.78
153.56
253.45
", + "bbox": [ + 80, + 296, + 187, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66799f3b750691748aabb58882e07f90a9e37565d4347b2997a4527f14454683.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Position Enc.bpp
-0.2811
Relative Pos.0.2643
Diamond Relative Pos.0.2512
Laplacian-shaped Pos.0.2347
", + "bbox": [ + 241, + 297, + 444, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 344, + 369, + 362, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positional Encoding: We investigate the influence of various types of positional encoding in the Transformer-based entropy model. As shown in Table 1b, the 2D diamond relative positional encoding, which is implemented using a clip function, demonstrates better performance than the relative position encoding. However, adopting the Laplacian-shaped positional encoding results in even more significant bitrate savings compared to the 2D diamond relative positional encoding. As depicted in Fig. 6, distinct values of $A$ and $\\sigma$ are obtained for each channel chunk. As we progress towards the final chunks, the receptive field becomes narrower. As observed in [14], the last chunk of channels, conditioned on all previous chunks, exhibits lower entropy. Their entropy can be estimated by considering only a small neighborhood, corresponding to a narrower receptive field in our proposed relative positioning. Conversely, to provide a reasonable estimate of entropy of the first channel chunks, a wider receptive field gathers more information over a large context to compared to mentioned last chunks.", + "bbox": [ + 75, + 401, + 468, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Context Blocks: We conducted a comparison of the inference latency for entropy parameters during entropy decoding, as well as the bitrate savings of our proposed entropy model compared to other backward adaptation-based entropy models. For a fair comparison, all the models are equipped with a same encoder comprising of ResNet blocks and convolution layers to transform the input image $\\boldsymbol{x} \\in R^{H \\times C \\times 3}$ to a latent representation $\\boldsymbol{y} \\in R^{H / 16 \\times W / 16 \\times 256}$ . As reported in Tabel 2, our entropy model's speed is notably improved by employing unevenly grouped channels and implementing a bidirectional context model, as compared to the sequential spatial context modeling. Moreover, our findings clearly indicate that incorporating spatial global context results in superior per", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aaf6816fced1ece463fb61406af7d9768360c96f9bc9213330d21eabdd709ec8.jpg", + "image_caption": [ + "Figure 6. Receptive fields are extracted for each channel chunk." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 576, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/abb886c94d26bcdbf75da767c9a7eaea83d0acec01962fbddac0fdcda68b6e0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 88, + 658, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdf0e772579027d8120da043af53a415ef7dd84496b21091b32a752c72fb7b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 88, + 738, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2fb3e32f6b9ebffe34f49c74d036f07cd2de45435ef2b613fcfafe1b0b548124.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 88, + 818, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c6685010111f4b5a2a57390062d0260ed436826ffa5ea886628f939a534b246.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 88, + 897, + 154 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79c2e4df442c1384f12b22a3bb7564c315f9411986f2f165cd531582889022ee.jpg", + "table_caption": [ + "Table 2. The Bjøntegaard-delta-rate (BD-rate) [6] and inference latency of entropy parameter estimation during entropy decoding (Dec.) for different context-based models on the Kodak dataset using a GPU (RTX A6000). The BD-Rate is computed relative to VVC. ([P]:parallel, [S]:Serial)" + ], + "table_footnote": [], + "table_body": "
MethodContext ModelDec.(ms)BD-Rate
ChannelSpatial
Minnen et. al.[30]-Local[S]>103-3.24
Minnen et. al.[29]Even-67-3.96
He et. al.[13]-Local[P]28.2-2.91
Qian et. al.[33]-Global[S]>103-4.89
He et. al.[14]Uneven-37.2-3.12
He et. al.[14]UnevenLocal[P]156.9-5.71
Jiang et. al.[21]UnevenLocal[P]+Global[P]207.4-7.89
OursUnevenLocal[S]+Global[S]>103-9.38
OursUnevenLocal[P]+Global[P]196.3-8.25
VVC---0.00
", + "bbox": [ + 504, + 281, + 877, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "formance when contrasted with utilizing only local context. The evaluation emphasizes that the integration of both global and local context enhances the precision of capturing spatial correlations, ultimately leading to a more accurate entropy model. In addition, our results verify that incorporating Laplacian-shaped positional encoding enhances the compression efficiency compared to MEM, which does not consider positional encoding in computing global spatial context.", + "bbox": [ + 496, + 445, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 619, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We developed a neural image compression model which improves perceptual image quality using a non-isotropic diffusion decoder. This decoder's inductive bias effectively separates frequency components, leading to the creation of high-quality images. Moreover, we introduce an innovative entropy model that optimizes the trade-off between compression efficiency and decoding speed. This entropy model, founded on the Transformer architecture with Laplacian-shaped positional encoding, establishes a strong global spatial context. Our results underscore the efficacy of leveraging diffusion models and advanced entropy modeling to achieve outstanding image compression performance.", + "bbox": [ + 496, + 628, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This research is based upon work supported by the National Aeronautics and Space Administration (NASA), via award number 80NSSC21M0322 under the title of Adaptive and Scalable Data Compression for Deep Space Data Transfer Applications using Deep Learning.", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3052", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] CLIC·challenge on learned image compression., Available at http://compression.cc/tasks/, 2022.6", + "[2] Kodak image dataset., Available at https://r0k.us/ graphics/kodak/, 2022.6", + "[3] Versatile Video Coding Reference Software. Available at https://vcgi.hhi.fraunhofer.de/jvet/VVCSoftware_VTM, 2022.6", + "[4] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 1, 2, 6", + "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2, 3", + "[6] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 8", + "[7] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2", + "[8] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9066-9075, 2019. 2", + "[9] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 6", + "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2", + "[11] Noor Fathima Ghouse, Jens Petersen, Auke Wiggers, Tianlin Xu, and Guillaume Sautiere. A residual diffusion model for high perceptual quality codec augmentation. arXiv preprint arXiv:2301.05489, 2023. 1, 3, 6", + "[12] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2", + "[13] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 1, 2, 3, 6, 8", + "[14] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2, 3, 5, 8", + "[15] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision and Pattern Recognition, pages 1764-1769, 2022. 2", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 4", + "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3", + "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 3", + "[20] Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. arXiv preprint arXiv:2209.05557, 2022. 2, 3, 4", + "[21] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 8", + "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3", + "[23] A Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach. Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In European Conference on Computer Vision, pages 447-463. Springer, 2022. 2, 3", + "[24] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beach. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 2", + "[25] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5", + "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3", + "[28] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 1, 2, 5, 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3053", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 8", + "[30] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3, 8", + "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4", + "[32] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Li Hao, and Rong Jin. Learning accurate entropy model with global reference for image compression. In International Conference on Learning Representations, 2020. 2, 3", + "[33] Yichen Qian, Xiuyu Sun, Ming Lin, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. In International Conference on Learning Representations, 2021. 2, 3, 6, 8", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3", + "[35] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 4", + "[36] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 3", + "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[38] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International Conference on Learning Representations, 2016. 1, 2", + "[39] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 3", + "[40] Dezhao Wang, Wenhan Yang, Yueyu Hu, and Jiaying Liu. Neural data-dependent transform for learned image compression. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 17379-17388, 2022. 6", + "[41] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 3", + "[42] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950, 2022. 1, 3, 6, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2", + "[44] Yibo Yang, Stephan Mandt, Lucas Theis, et al. An introduction to neural data compression. Foundations and Trends® in Computer Graphics and Vision, 15(2):113-200, 2023. 1", + "[45] Ali Zafari, Atefeh Khoshkhahtinat, Piyush Mehta, Mohammad Saeed Ebrahimi Saadabadi, Mohammad Akyash, and Nasser M Nasrabadi. Frequency disentangled features in neural image compression. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2815-2819. IEEE, 2023. 1", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 5", + "[47] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3054", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_model.json b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce377d0e3c8ed3e4cbad49bc5f549dd847a14325 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.181, + 0.782, + 0.217 + ], + "angle": 0, + "content": "Atefeh Khoshkhahtinat, Ali Zafari, Piyush M. Mehta, Nasser M. Nasrabadi \nWest Virginia University, WV, USA" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.219, + 0.812, + 0.235 + ], + "angle": 0, + "content": "{ak00043, az00004}@mix.wvu.edu, {piyush.mehta, nasser.nasrabadi}@mail.wvu.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.695 + ], + "angle": 0, + "content": "While replacing Gaussian decoders with a conditional diffusion model enhances the perceptual quality of reconstructions in neural image compression, their lack of inductive bias for image data restricts their ability to achieve state-of-the-art perceptual levels. To address this limitation, we adopt a non-isotropic diffusion model at the decoder side. This model imposes an inductive bias aimed at distinguishing between frequency contents, thereby facilitating the generation of high-quality images. Moreover, our framework is equipped with a novel entropy model that accurately models the probability distribution of latent representation by exploiting spatio-channel correlations in latent space, while accelerating the entropy decoding step. This channel-wise entropy model leverages both local and global spatial contexts within each channel chunk. The global spatial context is built upon the Transformer, which is specifically designed for image compression tasks. The designed Transformer employs a Laplacian-shaped positional encoding, the learnable parameters of which are adaptively adjusted for each channel cluster. Our experiments demonstrate that our proposed framework yields better perceptual quality compared to cutting-edge generative-based CODECs, and the proposed entropy model contributes to notable bitrate savings. The code is available at https://github.com/Atefeh-Khoshtinat/Blurdissipated-compression." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.21, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Image compression is a crucial tack in image processing which aims to decrease the amount of data required for storing or transmitting without significant loss of visual content. Recently, learning-based image compression methods [13, 25, 42, 45] have demonstrated the potential to surpass classical hand-engineered CODECs in terms of rate-distortion performance. Learned image compression commonly comprises three steps: transformation, quantization, and lossless entropy coding, which resemble the components found in the traditional transform coding paradigm [44]." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.269, + 0.865, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.462, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Overview of our proposed neural codec. The quantized semantic latent variable \\(\\hat{\\pmb{y}}\\) is utilized by a diffusion-based decoder to generate realistically reconstructed image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Neural image compression methods are typically trained with two primary objectives: minimizing distortion between the original input image and its reconstructed version and reducing the bitrate for efficient data transmission or storage [38]. However, solely optimizing for rate-distortion may lead to blurry reconstructions. To address this issue, various generative-based CODECs [4, 11, 28, 42] have emerged, aiming to generate reconstructions that are not only faithful to the original data but also visually realistic to human observers.[28] introduced HiFiC, an image compression network based on GANs that aims to achieve a balance between fidelity and perceptual quality. The network demonstrated substantial advantages for human observers, surpassing BPG even when using half the bits. However, HiFiC faces challenges due to the instability of GAN training and necessitates various design adjustments. Inspired by the achievements of Denoising Diffusion Probabilistic Models (DDPMs)[17] in generative modeling, authors in [42] replaced the Gaussian decoder of VAEs with a conditional diffusion model to mitigate the issue of blurriness in decoded images. Despite the impressive performance of diffusion-based network, its diffusion and denoising processes do not explicitly incorporate the inductive biases inherent to natural images. The imposing inductive biases," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.467, + 0.122 + ], + "angle": 0, + "content": "such as the multi-scale nature of images, enable the generation of high-quality samples [20]." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.124, + 0.468, + 0.852 + ], + "angle": 0, + "content": "In addition to enhancing the quality of reconstructed images, accurate entropy estimation of latent representations is vital for boosting compression efficiency. The more the estimated entropy gets closer to the actual entropy of the latent, the lower the bitrate required for generating bitstream file. In this respect, several works have been proposed. Balle et al.[5] introduced an entropy model that is conditioned on an additional latent variable, named the hyperprior, to capture the existing spatial redundancies within the latent space. Studies [24, 30] introduced a sequential autoregressive block into the entropy model. This block leverages causally adjacent latent elements to estimate the distribution of the current latent, which results in slowing down of the decoding step. He et al.[13] utilized a two-way parallel context model as a substitute for the sequential autoregressive context model, aiming to accelerate the decoding of latent code. In another realm of studies, to mitigate the slow decoding times, authors in [29] introduced channel-wise context modeling as an alternative to serial spatial context. To expedite the extracting of the channel-conditional context model, He et al. [14] exploited an unevenly segmented channel-wise autoregressive model, wherein the channels are divided into varying sizes. Strengthening the entropy model through self-attention was initially proposed in [32]. This approach employed a global reference model to identify the most suitable spatial context, which doesn't have to be confined to the local scope. The idea of integrating global context into the entropy model was subsequently refined in a later work [33], where a specially crafted transformer-based entropy model was employed. This entropy model is equipped with the diamond-relative position encoding to effectively model long-distance dependencies. The diamond-shaped relative positional encoding is devised based on the clip function, which relies on the pre-defined threshold to characterize an effective receptive field. This approach is effective in computing context based on the positions of latent elements; yet, it cannot provide an optimal receptive field due to its reliance on the clip function. Recent works[21, 23] focused on developing adaptive context models to extract both channel and spatial correlations within the latent space while expediting the entropy decoding step. Authors in [21] proposed an entropy model named MEM, aiming to capture both local and global spatial correlations in each channel segment by using a parallel bidirectional context approach. However, despite MEM's attention to spatio-channel correlations, it appears to overlook the crucial role of positional encoding in computing long-range spatial dependencies." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.855, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose a conditional diffusion-based decoder in the neural compression pipeline to enhance the realism of reconstructed" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.347 + ], + "angle": 0, + "content": "images. The traditional diffusion models are generalized into a conditional non-isotropic diffusion model to incorporate an inductive bias, considering the relative importance of each frequency component of images. As a result, each frequency component of an image undergoes diffusion at distinct rates, leading to the generation of decoded images in a coarse-to-fine manner. Furthermore, a novel channel-conditional autoregressive entropy model is introduced to efficiently take into account both local and global spatial dependencies within each channel chunk of the latent space. To leverage global spatial context, a Transformer block is designed which incorporates a Laplacian-shaped positional encoding within checkerboard-shaped self-attention module. The effective Laplacian-shaped receptive field for each channel chunk is dynamically determined by learning the positional encoding parameters during the optimization of the entire neural codec." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.38, + 0.647, + 0.396 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.412, + 0.756, + 0.428 + ], + "angle": 0, + "content": "2.1. Learned Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.623 + ], + "angle": 0, + "content": "Neural image compression methods adopt a non-linear transform coding paradigm which is founded on the variational autoencoders (VAEs) [12]. Within this scheme, the encoder initially maps the input image \\( \\pmb{x} \\) to a compact latent representation \\( \\pmb{y} \\). Subsequently, quantization is applied to this latent representation, leading to discrete symbols that minimize the bit requirement. Afterward, entropy coding generates bit strings using an entropy model. Ultimately, the reconstructed image \\( \\hat{\\pmb{x}} \\) is generated by applying the decoder to the quantized latent representation \\( \\hat{\\pmb{y}} \\). This framework is commonly trained with a trade-off between the rate and distortion, such as Mean Squared Error (MSE) [38]." + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recent studies have attempted to enhance the realism of reconstructed images by optimizing neural CODEs using a triple rate-distortion-perception loss [4, 28, 43]. Blau and Michaeli [7] explored the essential balance between distortion and realism, demonstrating that within a given rate, improving distortion comes at the cost of decreasing the perceptual quality of a reconstructed image. In line with this premise where perceptual quality is measured as the difference between the image distribution and the distribution of reconstructions, Mentzer et al. [28] employed a conditional GAN within an autoencoder-based compression model to enhance perceptual quality. In a similar vein, [15] introduced a model that aimed to improve the realism by enriching the loss function with an adversarial perceptual loss using LPIPS [46], along with a patch-based style loss [10]. Recently, Agustsson et al. [4] developed a Multi-Realism model that merges the ELIC codec [14] and PatchGAN [8], showing improved performance compared to HiFiC." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3046" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.248, + 0.106 + ], + "angle": 0, + "content": "2.2. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.113, + 0.47, + 0.265 + ], + "angle": 0, + "content": "Diffusion models [17, 22, 37], which belong to the family of score-based generative models, acquire the data distribution through a gradual iterative denoising process, starting from a Gaussian distribution and progressing towards the actual data distribution. Recently, they have received considerable attention owing to their training stability and high quality image generation compared to GANs. Diffusion models are adopted in various domains, including image and video generation [18, 19, 34], super-resolution [36], inpainting [27], deblurring [41], compression [11, 39, 42]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.473, + 0.508 + ], + "angle": 0, + "content": "In the domain of image compression, several diffusion-based CODEs have been proposed. The research conducted by Theis et al. [39] involved employing a general unconditional diffusion model for transmitting Gaussian samples in a lossy manner. Their approach leverages a reverse channel coding concept, which stands apart from the conventional transform coding scheme. Although their methodology performs well, it suffers from high computational cost which makes it unfeasible for dealing with high-resolution images. Yang and Mandt [42] introduced a neural codec, called CDC, in which the decoder takes the form of a DDPM, conditioned on a quantized latent representation. Ghouse et al. [11] began by optimizing an autoencoder network using a rate-distortion loss. Subsequently, they train a conditional diffusion model on the output of the decoder, with the goal of enhancing its perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.291, + 0.532 + ], + "angle": 0, + "content": "2.3. Neural Entropy Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.795 + ], + "angle": 0, + "content": "The main purpose of the entropy model is to estimate the joint probability distribution over the quantized latent representation. When the learned entropy model precisely matches the true distribution of the latent representation, a lower bit-rate is required to generate a compressed file. Entropy estimation can take advantage of two key principles: backward and forward adaptation. Forward adaptation uses an extra latent variable, called hyper-prior [5], as a side information to capture spatial dependencies between elements of the latent representation. Backward adaptation, on the other hand, employs an autoregressive-based context model which leverages the previously decoded elements of the latent to predict the probability of the current element. The context model of backward methods can extract relationships between the current symbol and previously decoded symbols in various dimensions, including spatial and channel axes [13, 14, 21, 23, 29, 30, 32, 33, 47]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Minnen et al.[30] incorporated a local spatial context with a hyper-prior network to precisely predict distribution of the latent. This type of context model can exploit local correlations between the current latent and its neighboring causal elements using masked convolutions, which consequently necessitates the adoption of serial decoding. To expedite the decoding process, He et al.[13] evenly parti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "tioned the latent into two groups: anchors and non-anchors. The anchors are exclusively decoded using the hyper-prior, while the non-anchors utilize both the hyper-prior and the local context model. This approach allows for the parallel decoding of both anchors and non-anchors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.893, + 0.333 + ], + "angle": 0, + "content": "An alternative strategy for parallelizing the decoding process is to extract inter-channel dependencies. In [29] the latent code is divided into evenly sized chunks along the channel dimension, with each of these chunks being decoded using information from previously decoded ones. To speed up the channel-wise context extraction, He et al.[14] partitioned channels of the latent into an uneven set of groups, allocating fewer channels to the initial groups and more channels to the subsequent ones. This uneven division is motivated by the observation that the earlier channels possess higher entropy compared to the remaining channels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.892, + 0.485 + ], + "angle": 0, + "content": "Some works introduced adaptive context models which focus on capturing long-range spatial correlations. Qian et al.[32] devised a global reference context model which assesses throughout previously decoded elements to detect the most similar one to the current latent. The authors in [33] employed a Transformer-based entropy model, called Entroformer, for capturing long-range contexts. The Entroformer benefits from diamond-relative position encoding and a top-k self-attention mechanism, both of which contribute to providing an efficient receptive field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.499, + 0.6, + 0.515 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.646 + ], + "angle": 0, + "content": "In this work, we develop our decoder based on the blurring diffusion model [20], which is conditioned on the discrete latent representation and enables the differentiation among the frequency components of an image throughout the diffusion and denoising processes. The architecture of decoder will be explained in Appendix. Moreover, a novel entropy model is employed that efficiently encodes the quantized latent representation into a binary stream." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.733, + 0.672 + ], + "angle": 0, + "content": "3.1. Blurring Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.754 + ], + "angle": 0, + "content": "We can define the blurring diffusion model as a traditional diffusion model in the frequency space, but with distinct schedules for each dimension of data [20]. The specifics of the schedules for vectors \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are crucial in this type of model, which will be discussed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.829 + ], + "angle": 0, + "content": "Diffusion Process: The diffusion process [17] progressively degrades the clean image, transforming it into nearly pure Gaussian noise over the course of T time steps. Each step of the diffusion process in frequency space can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.841, + 0.892, + 0.86 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {f} _ {t} \\mid \\boldsymbol {f} _ {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t}; \\alpha_ {t} \\boldsymbol {f} _ {\\boldsymbol {x}}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(f_{t}\\) and \\(f_{x}\\) are defined as \\(\\mathbf{V}^T\\mathbf{z}_t\\) and \\(\\mathbf{V}^T\\mathbf{x}\\), respectively. In such a model, each frequency component could" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "undergo diffusion at varying rates, with the modulation of these rates being determined by the values present in the vectors \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.469, + 0.182 + ], + "angle": 0, + "content": "Denoising Process: The true denoising process in the frequency space, which is equal to the deblurring operation in the time domain, can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.189, + 0.469, + 0.206 + ], + "angle": 0, + "content": "\\[\nq \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\boldsymbol {f} _ {x}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1}, \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.469, + 0.318 + ], + "angle": 0, + "content": "where \\(\\sigma_{t\\rightarrow t - 1} = \\sigma_{t|t - 1}\\sigma_{t - 1} / \\sigma_t\\) and \\(\\mu_{t\\to t - 1} = (\\alpha_{t|t - 1}\\sigma_{t - 1}^2 /\\sigma_t^2)f_t + (\\alpha_{t - 1}\\sigma_{t|t - 1}^2 /\\sigma_t^2)f_x.\\) The actual denoising process can be approximated using a learned denoising distribution to generate data. Thus, by considering the epsilon reparametrization of Eq. 2, we can derive the expression for the learned denoising process in frequency space:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.329, + 0.469, + 0.346 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1} \\left(\\hat {\\boldsymbol {f}} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {t}\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.35, + 0.469, + 0.367 + ], + "angle": 0, + "content": "Re-parametrization: \\(\\hat{f}_{\\pmb{x}} = (1 / \\alpha_{t})(f_{t} - \\sigma_{t}\\hat{f}_{\\epsilon})\\) (3)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.376, + 0.469, + 0.443 + ], + "angle": 0, + "content": "By substituting the approximation of \\(\\hat{f}_x\\) into the true denoising distribution, the mean of the denoising network can be derived by \\((\\alpha_{t|t-1}\\sigma_{t-1}^2/\\sigma_t^2)\\pmb{f}_t + (\\alpha_{t-1}\\sigma_{t|t-1}^2/\\sigma_t^2\\alpha_t)(\\pmb{f}_t - \\sigma_t\\hat{f}_\\epsilon)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.469, + 0.484 + ], + "angle": 0, + "content": "Optimization: According to [17], the loss function is simplified as below:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.49, + 0.47, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\boldsymbol {f} _ {\\epsilon} - \\hat {\\boldsymbol {f}} _ {\\epsilon} \\| ^ {2} ] = \\boldsymbol {E} [ \\| \\boldsymbol {V} ^ {\\boldsymbol {T}} (\\epsilon - \\hat {\\epsilon}) \\| ^ {2} ] \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\epsilon - \\hat {\\epsilon} \\| ^ {2} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.469, + 0.629 + ], + "angle": 0, + "content": "where \\(\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)\\) and \\(z_t = V(\\alpha_t V^T x + \\sigma_t V^T \\epsilon)\\). The \\(\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)\\) signifies that the neural network \\(\\phi_{\\theta}\\) takes \\(z_t\\) as input and predicts \\(\\hat{\\epsilon}\\). This formulation satisfies the requirement for neural networks to perform effectively in a standard pixel space. After prediction, it is sufficient to transition back and forth between frequency space and pixel space using the DCT matrix \\(V^T\\) and its inverse \\(V\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.408, + 0.652 + ], + "angle": 0, + "content": "3.2. Schedules of Blurring Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.469, + 0.824 + ], + "angle": 0, + "content": "The schedules for the blurring diffusion model, denoted as \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\), are obtained by combining a standard Gaussian noise diffusion schedule (with scalar values \\(\\sigma_{t}\\) and \\(\\alpha_{t}\\)) along with a blurring schedule \\(d_{t}\\). Each element of the vector \\(\\sigma_{t}\\) shares the same value, as identical noise is added to all frequency components. Therefore, it becomes adequate to present a schedule for a scalar value \\(\\sigma_{t}\\) [20]. The noise schedule is chosen based on a variance-preserving cosine [31], specifically \\(\\sigma_{t}^{2} = 1 - \\alpha_{t}^{2}\\), where \\(\\alpha_{t} = \\cos(t\\pi / 2T)\\) for \\(t \\in [0, T]\\). Following [35], the blurring schedule is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.832, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {B, t} = \\sigma_ {B, m a x} \\sin (t \\pi / 2 T) ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma_{B,max}\\) represents a hyperparameter equal to the maximum level of blur applied to the image. This schedule, in turn, corresponds to the dissipation time through" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "\\(\\tau_{t} = \\sigma_{B,t}^{2} / 2\\). Based on the formulation discussed in the Appendix, the blurring schedule \\(d_{t}\\), which is employed for \\(\\alpha_{t}\\), is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.15, + 0.891, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} _ {\\boldsymbol {t}} = \\exp (- \\lambda \\tau_ {t}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.177, + 0.892, + 0.282 + ], + "angle": 0, + "content": "where \\(\\lambda\\) represents the vector containing squared frequencies, and \\(\\tau_{t}\\) corresponds to the dissipation time. To achieve a more gradual amplification of high frequencies during the denoising process, the blurring schedule \\(d_{t}\\) is adjusted to \\((1 - d_{min})\\exp (-\\lambda \\tau_{t}) + d_{min}\\), where \\(d_{min}\\) is set to 0.001. Finally, by combining the Gaussian noise schedule with the blurring schedule, resultant schedule is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.295, + 0.891, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\alpha} _ {t} = \\alpha_ {t}. \\boldsymbol {d} _ {t}, \\boldsymbol {\\sigma} _ {t} = \\mathbf {1} \\sigma_ {t}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.321, + 0.864, + 0.337 + ], + "angle": 0, + "content": "3.3. Blurring Diffusion Model for Compression" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.344, + 0.892, + 0.405 + ], + "angle": 0, + "content": "The rate-distortion objective in end-to-end compression resembles the loss function of a \\(\\beta\\)-VAE, where a hyperparameter \\(\\lambda\\) is utilized to balance the trade-off between the bit-rate \\((R)\\) and distortion \\((D)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.416, + 0.891, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} = D + \\lambda R = \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.444, + 0.892, + 0.67 + ], + "angle": 0, + "content": "While many neural codecs commonly employ Gaussian or Laplacian decoders, we introduce a novel approach using a conditional blurring diffusion model as the decoder. This approach aims to yield new distortion metrics that deviate from those based on Mean Squared Error (MSE) or Mean Absolute Error (MAE). In this framework, our proposed neural codec leverages two distinct types of latent variables: a semantic latent variable \\( \\mathbf{y} \\) and texture latent variables \\( \\mathbf{z}_{1:T} \\). The semantic latent variable captures and encodes the overall content and meaning of the image. In contrast, the texture latent variables are tailored to carry finer details and intricate patterns that might not be fully represented by the semantic variable alone. Notably, unlike the semantic latent variable, the texture latent variables are not compressed but synthesized during the decoding phase." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.671, + 0.891, + 0.701 + ], + "angle": 0, + "content": "The forward and backward processes of the conditional blurring diffusion model can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.505, + 0.724, + 0.805, + 0.743 + ], + "angle": 0, + "content": "\\[\nq (\\pmb {f} _ {t} | \\pmb {f} _ {t - 1}) = \\mathcal {N} (\\pmb {f} _ {t}; \\alpha_ {t | t - 1} \\pmb {f} _ {t - 1}, \\sigma_ {t | t - 1} ^ {2} I),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.747, + 0.888, + 0.786 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {\\boldsymbol {1}: T} | \\tilde {\\boldsymbol {y}}\\right) = p \\left(\\boldsymbol {f} _ {\\boldsymbol {T}}\\right) p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}} \\mid \\boldsymbol {f} _ {\\boldsymbol {1}}, \\tilde {\\boldsymbol {y}}\\right) \\prod_ {t = 2} ^ {T} p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {t} - \\boldsymbol {1}} \\mid \\boldsymbol {f} _ {\\boldsymbol {t}}, \\tilde {\\boldsymbol {y}}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.79, + 0.652, + 0.807 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {f} _ {T}) = \\mathcal {N} (\\boldsymbol {f} _ {T}; \\mathbf {0}, \\boldsymbol {I})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.81, + 0.891, + 0.842 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {\\theta} \\left(\\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}, t\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.855, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\( f_{t} = V^{T}z_{t} \\) and \\( f_{x} = V^{T}x = V^{T}z_{o} \\). As shown in Eq. 8, distortion is equivalent to the negative marginal likelihood of input data \\( -\\log p_{\\boldsymbol{x}|\\tilde{\\boldsymbol{y}}}(\\boldsymbol{x}|\\tilde{\\boldsymbol{y}}) \\), and minimizing it" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3048" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "is analogous to minimizing the negative marginal likelihood of the frequency representation of the data \\( p_{\\pmb{f}_x|\\tilde{\\pmb{y}}}(\\pmb{f}_x|\\tilde{\\pmb{y}}) = \\int p(\\pmb{f}_x, \\pmb{f}_{1:T}|\\tilde{\\pmb{y}}) d\\pmb{f}_{1:T} \\). Since computing the marginal likelihood is intractable, we employ its ELBO with the specified diffusion and de-blurring distributions. Following this substitution, the rate-distortion objective, through the application of Jensen's inequality, can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.22, + 0.469, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right] \\\\ \\leq \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} [ - \\log_ {P} \\boldsymbol {x} | \\boldsymbol {y} (\\omega | \\boldsymbol {y}) - \\lambda \\log_ {P} \\boldsymbol {y} (\\boldsymbol {y}) ], \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.267, + 0.47, + 0.303 + ], + "angle": 0, + "content": "where \\(\\mathrm{ELBO} = E_q[\\frac{p_\\theta(f_x,f_{1:T}|\\tilde{y})}{q(f_t|f_{t - 1})} ]\\). As explained in Appendix, the optimizing of ELBO can be simplified to:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.312, + 0.47, + 0.331 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E L B O} \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} ], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.34, + 0.473, + 0.448 + ], + "angle": 0, + "content": "where \\( z_{t} = V\\alpha_{t}V^{T}x + \\sigma_{t}\\epsilon \\). Similar to [28], an LPIPS loss [46] is incorporated into the rate-distortion loss to enhance the perceptual quality of the reconstructed image. As the initial image at any time step can be decoded as a function of the texture latent variable \\( z_{t} \\), the semantic latent variable \\( y \\), and the time step \\( t \\), i.e., \\( \\hat{x}_t = V(1 / \\alpha_t)(V^T z_t - \\sigma_t V^T \\hat{\\epsilon}) \\), the total loss becomes as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.456, + 0.47, + 0.494 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ (1 - \\beta) \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} \\tag {12} \\\\ \\left. - \\lambda \\log p _ {\\tilde {\\mathbf {y}}} (\\tilde {\\mathbf {y}}) + \\beta d _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {t}) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.503, + 0.47, + 0.533 + ], + "angle": 0, + "content": "where the \\(\\lambda\\) and \\(\\beta\\) represent hyperparameters that control the trade-off between rate, distortion, and perception." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.534, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Decoding Process: After the training, we utilize entropy decoding on \\(\\hat{\\pmb{y}}\\) through the entropy model which estimates the distribution \\(p_{\\hat{\\pmb{y}}}(\\hat{\\pmb{y}})\\). We then employ ancestral sampling to conditionally decode the image \\(\\pmb{x}\\), which is equal to \\(\\pmb{z_0}\\) and starts from pure Gaussian noise \\(\\pmb{z}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.619, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t - 1} \\gets \\boldsymbol {V} (\\hat {\\boldsymbol {\\mu}} _ {t \\rightarrow t - 1} + \\boldsymbol {\\sigma} _ {t \\rightarrow t - 1} \\boldsymbol {V} ^ {T} \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\hat {\\boldsymbol {y}})) \\quad (1 3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.647, + 0.469, + 0.704 + ], + "angle": 0, + "content": "where \\(\\hat{\\mu}_{t\\rightarrow t - 1} = \\frac{\\alpha_{t|t - 1}\\sigma_{t - 1}^2}{\\sigma_t^2} V^T z_t + \\frac{\\sigma_{t|t - 1}^2}{\\alpha_{t|t - 1}\\sigma_t^2} (V^T z_t - \\sigma_t V^T\\phi_\\theta (z_t,t,\\hat{y}))\\) . The latent variables \\(z_{1:T}\\) are not stored but produced during the decoding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.308, + 0.728 + ], + "angle": 0, + "content": "3.4. Proposed Entropy Model" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.469, + 0.87 + ], + "angle": 0, + "content": "The main goal of the proposed entropy model' context is to exploit both channel-wise and spatial correlations, while expediting the decoding process. Inspired by the ELIC [14], we adopt an uneven grouping of latent channels, allowing most low entropy channels to depend on high entropy channels. The latent representation \\(\\hat{\\pmb{y}}\\), with \\(M\\) channels, is clustered into five chunks along the channel dimension: 16, 16, 32, 64, and M - 128 channels, respectively. In this setup, each chunk depends on all its previous decoded chunks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The Fig. 2(a) illustrates our proposed entropy model as applied in the \\(j\\)-th chunk. Within each chunk, we use spatial" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.088, + 0.887, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.284, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Diagram illustrating the application of the proposed entropy model for decoding the \\(j\\)-th chunk \\(\\hat{\\pmb{y}}^{(j)}\\). (b) Global Spatial Context Block. (c) An example of a checkerboard-shaped mask." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.49 + ], + "angle": 0, + "content": "context in conjunction with channel context to model correlations along both the channel and spatial dimensions. To expedite the decoding step, we employ a parallel bidirectional spatial context model which is capable of capturing both local and global spatial relationships. So, the anchor part is decoded in parallel by using solely the hyperprior and channel context, while the decoding of the non-anchor part relies on the hyperprior, as well as both the spatial and channel contexts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.664, + 0.528 + ], + "angle": 0, + "content": "3.4.1 Spatial Context" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.892, + 0.704 + ], + "angle": 0, + "content": "The spatial context design captures both local and global spatial correlations within the latent representation \\(\\hat{\\pmb{y}}\\). Following the decoding of the anchor group, a checkerboard-shaped convolution is applied to this group, generating local context for all non-anchor elements in a parallel manner. Furthermore, the acquired local contexts of the non-anchor part are subsequently fed into a Transformer-based block to efficiently extract the global spatial context. This Transformer block takes advantage of positional encoding that is customized specifically for the compression task, along with a checkerboard-based attention mechanism." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.802, + 0.742 + ], + "angle": 0, + "content": "3.4.2 Transformer-based Spatial Context" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.751, + 0.892, + 0.872 + ], + "angle": 0, + "content": "We adopt the Swin Transformer [26] blocks, including a masked window-based multi-head self-attention (W-MSA) and a masked shifted-window-based multi-head self-attention (SW-MSA), as our global spatial context model which shown in Fig. 2.(b). This selection enables us to strike a balance between computational efficiency and modeling capacity. The checkerboard-based self-attention in W-MAS and SW-MSA can be expressed:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nA t t e n (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} (\\boldsymbol {Q} \\boldsymbol {K} ^ {T} \\odot \\boldsymbol {M}) \\boldsymbol {V}, \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3049" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.09, + 0.473, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.179, + 0.47, + 0.207 + ], + "angle": 0, + "content": "Figure 3. The procedure for acquiring Laplacian relative position encoding for a window with a size of \\(2 \\times 2\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.231, + 0.471, + 0.295 + ], + "angle": 0, + "content": "where \\(Q, K, V \\in R^{N^2 \\times d}\\) are the query, key, and value matrices, respectively. \\(N^2\\) denotes the number of patches in a window and \\(d\\) is the dimension of query/key/value. \\(M \\in R^{N^2 \\times N^2}\\) represents a checkerboard-shaped mask." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.313, + 0.332, + 0.327 + ], + "angle": 0, + "content": "3.4.3 Relative Positional Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.47, + 0.516 + ], + "angle": 0, + "content": "The authors in [13, 33] have shown that the bitrate is impacted by the distance of neighboring latents within the spatial context modeling of current latent. Based on this insight, we introduce a receptive field-aware self-attention mechanism that employs a learnable Laplacian-shaped positional encoding for calculating spatial context. In our proposed approach, each channel chunk possesses its own receptive field during self-attention computation, which dynamically adjusts in response to changes in entropy of chunks. In the \\(j\\)-th chunk, the checkerboard-based attention of W-MSA and SW-MSA is refined to receptive field-aware self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.527, + 0.479, + 0.558 + ], + "angle": 0, + "content": "\\[\nA t t e n \\left(\\boldsymbol {Q} _ {j}, \\boldsymbol {K} _ {j}, \\boldsymbol {V} _ {j}\\right) = \\text {s o f t m a x} \\left(\\boldsymbol {Q} _ {j} \\boldsymbol {k} _ {j} ^ {T} \\odot \\boldsymbol {M} + \\boldsymbol {P} _ {\\text {L a p}, j}\\right) \\boldsymbol {V} _ {j}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.833 + ], + "angle": 0, + "content": "where \\( P_{Lap,j} \\) represents the learnable Laplacian relative position encoding for chunk \\( j \\). To create the Laplacian relative position encoding for the \\( j \\)-th chunk, which comprises spatial windows with a size of \\( N \\times N \\) (i.e. including \\( N^2 \\) patches), three steps need to be taken. Firstly, We generate a 2D relative position table, where each coordinate of the relative position lies in the range \\( [-N + 1, N - 1] \\). Subsequently, we derive the relative position matrix for each patch. As shown in Fig. 3, the first patch's relative distance coordinate of \\( (0,0) \\) (relative distance with its position) is located at the top-left of the orange box, while the last patch's relative distance of \\( (0,0) \\) is positioned at the bottom-right of the green box. Afterward, each relative position matrix is flattened and stacked together. Finally, we apply the 2D Laplacian function to each element of the resulting matrix to generate a Laplacian relative position encoding \\( P_{Lap,j} \\) with learnable parameters \\( A_j \\) and \\( \\sigma_j \\), whose size is \\( R^{N^2 \\times N^2} \\). The 2D Laplacian function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.843, + 0.47, + 0.862 + ], + "angle": 0, + "content": "\\[\nf _ {L a p} (x, y) = A _ {j} ^ {2} \\exp \\left(\\left(- 1 / 2 \\sigma_ {j} ^ {2}\\right) \\left(| x | + | y |\\right)\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.469, + 0.903 + ], + "angle": 0, + "content": "where \\(x\\in \\{-N + 1,\\dots,N - 1\\}\\) \\(y\\in \\{-N + 1,\\dots,N - 1\\}\\) \\(A_{j}\\) and \\(\\sigma_{j}\\) are learnable parameters which are determined" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "for each chunk through optimization. The value of them is associated with the wideness of the effective receptive field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.136, + 0.632, + 0.153 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.161, + 0.718, + 0.178 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.486 + ], + "angle": 0, + "content": "To train our learned image compression network, we utilize a merged dataset that includes high-resolution images from the DIV2K, Flickr2K, and CLIC [1] datasets. These images are randomly cropped to a size of \\(256 \\times 256\\) during the training phase. The model parameters of all architectures were optimized using the Adam optimizer for a total of 2.4 million steps, with a batch size of 8. The initial learning rate was configured to be \\(1 \\times 10^{-4}\\) and was progressively reduced until the conclusion of training, reaching \\(1 \\times 10^{-7}\\). In order to cover a broad spectrum of bitrates, we selected a hyperparameter \\(\\lambda\\) from the set \\(\\{0.0004, 0.005, 0.01, 0.02, 0.04, 0.016\\}\\). The hyperparameter \\(\\beta\\) which specifies the contribution of perceptual loss is considered 0.9 for all models. The parameter \\(T\\), represents the required time steps for diffusion-based decoder, is consistently set to 500. To evaluate the performance of our compression approach, we select the Kodak dataset [2], which comprises 24 high-quality images with a resolution of \\(768 \\times 512\\), and the CLIC2020 test set (428 images) with varying resolutions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.82, + 0.513 + ], + "angle": 0, + "content": "4.2. Comparison with the SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.836 + ], + "angle": 0, + "content": "We compare our model with state-of-the-art generative based image compression networks, including both GAN-based and diffusion-based CODECs, as well as a state-of-the-art hand-crafted codec such as VVC-Intra (VTM) [3]. Additionally, we choose two VAE CODECs for comparison: DGML [9] and NSC [40]. Specifically, within the domain of GAN-based image compression approaches, we select HiFiC [28] and Multi-Realism [4] frameworks for comparison. Furthermore, we assess our method against diffusion-based models, namely DIRAC [11] and CDC [42]. This comparison is performed with respect to both rate-distortion and rate-perception tradeoff. The distortion is measured by the Peak Signal-to-Noise Ratio (PSNR) metric and perceptual quality are quantified via Frechet Inception Distance (FID) [16] and Learned Perceptual Image Patch Similarity (LPIPS). As shown in Fig. 4, our model shows superior performance compared to all other CODECs in terms of rate-perception. However, this heightened realism comes at the cost of distortion where other methods, except for the CDC model, either exhibit better performance or remain competitive." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.649, + 0.864 + ], + "angle": 0, + "content": "4.3. Visual Quality" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 illustrates reconstructed images (kodim20.png, kodim07.png) generated by our proposed model, as well as" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3050" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.36, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.09, + 0.603, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.091, + 0.854, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.249, + 0.36, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.25, + 0.602, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.251, + 0.851, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.423, + 0.894, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Comparison of our method with other codescs in terms of rate/distortion [bpp \\(\\downarrow\\) / PSNR \\(\\uparrow\\) ] and rate-perception, including [bpp \\(\\downarrow\\) / FID \\(\\downarrow\\) ] and [bpp \\(\\downarrow\\) / LPIPS \\(\\downarrow\\)], for both the CLIC2020 test set and the Kodak dataset." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.469, + 0.891, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.737, + 0.893, + 0.767 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of our method to HiFiC and CDC models shows that our model achieves superior reconstruction quality, particularly at lower bit-rates. In addition, our model displays fewer artifacts compared to both the HiFiC and CDC models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.471, + 0.897 + ], + "angle": 0, + "content": "the CDC [42] and HiFiC [28] frameworks. Perceptually-oriented neural codecs, such as HiFiC and CDC, are criticized when their decoding process encounters semantic content, such as text within the compressed image. Our model effectively addresses this challenge, as depicted in Fig. 5. It enhances perceptual quality through a diffusion-based decoder and imposes a strict inductive bias via the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.791, + 0.892, + 0.853 + ], + "angle": 0, + "content": "de-blurring diffusion process, resulting in both high perceptual quality and better text-preserving reconstructions. Additional qualitative comparisons are provided in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3051" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.232, + 0.108 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.114, + 0.47, + 0.234 + ], + "angle": 0, + "content": "Maximum Blurring: To explore the impact of the blurring schedule, we vary the values of \\(\\sigma_{B,max}\\) and compare the performance of the resulting models. It is evident that the model with \\(\\sigma_{B,max} = 0\\) is equivalent to a standard denoising diffusion model. As indicated in Table 1a, blurring diffusion models with higher maximum blur levels \\(\\sigma_{B,max} = 25\\) generate higher-quality reconstructed images compared to other variant models in terms of FID score." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.244, + 0.469, + 0.287 + ], + "angle": 0, + "content": "Table 1. Ablation Study: all models are optimized with \\( \\lambda = {0.01} \\) . (a) Investigating the impact of maximum blurring \\( {\\sigma }_{B,\\max } \\) . (b) Exploring the effects of different types of positional encoding" + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.297, + 0.189, + 0.37 + ], + "angle": 0, + "content": "
σB,maxFID
03.94
53.78
153.56
253.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.129, + 0.371, + 0.147, + 0.381 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.242, + 0.298, + 0.445, + 0.37 + ], + "angle": 0, + "content": "
Position Enc.bpp
-0.2811
Relative Pos.0.2643
Diamond Relative Pos.0.2512
Laplacian-shaped Pos.0.2347
" + }, + { + "type": "table_caption", + "bbox": [ + 0.345, + 0.371, + 0.363, + 0.381 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Positional Encoding: We investigate the influence of various types of positional encoding in the Transformer-based entropy model. As shown in Table 1b, the 2D diamond relative positional encoding, which is implemented using a clip function, demonstrates better performance than the relative position encoding. However, adopting the Laplacian-shaped positional encoding results in even more significant bitrate savings compared to the 2D diamond relative positional encoding. As depicted in Fig. 6, distinct values of \\( A \\) and \\( \\sigma \\) are obtained for each channel chunk. As we progress towards the final chunks, the receptive field becomes narrower. As observed in [14], the last chunk of channels, conditioned on all previous chunks, exhibits lower entropy. Their entropy can be estimated by considering only a small neighborhood, corresponding to a narrower receptive field in our proposed relative positioning. Conversely, to provide a reasonable estimate of entropy of the first channel chunks, a wider receptive field gathers more information over a large context to compared to mentioned last chunks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Analysis of Context Blocks: We conducted a comparison of the inference latency for entropy parameters during entropy decoding, as well as the bitrate savings of our proposed entropy model compared to other backward adaptation-based entropy models. For a fair comparison, all the models are equipped with a same encoder comprising of ResNet blocks and convolution layers to transform the input image \\( \\boldsymbol{x} \\in R^{H \\times C \\times 3} \\) to a latent representation \\( \\boldsymbol{y} \\in R^{H / 16 \\times W / 16 \\times 256} \\). As reported in Tabel 2, our entropy model's speed is notably improved by employing unevenly grouped channels and implementing a bidirectional context model, as compared to the sequential spatial context modeling. Moreover, our findings clearly indicate that incorporating spatial global context results in superior per" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.578, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.089, + 0.659, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.089, + 0.739, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.089, + 0.819, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.089, + 0.898, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.167, + 0.885, + 0.181 + ], + "angle": 0, + "content": "Figure 6. Receptive fields are extracted for each channel chunk." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.2, + 0.894, + 0.27 + ], + "angle": 0, + "content": "Table 2. The Bjøntegaard-delta-rate (BD-rate) [6] and inference latency of entropy parameter estimation during entropy decoding (Dec.) for different context-based models on the Kodak dataset using a GPU (RTX A6000). The BD-Rate is computed relative to VVC. ([P]:parallel, [S]:Serial)" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.282, + 0.878, + 0.416 + ], + "angle": 0, + "content": "
MethodContext ModelDec.(ms)BD-Rate
ChannelSpatial
Minnen et. al.[30]-Local[S]>103-3.24
Minnen et. al.[29]Even-67-3.96
He et. al.[13]-Local[P]28.2-2.91
Qian et. al.[33]-Global[S]>103-4.89
He et. al.[14]Uneven-37.2-3.12
He et. al.[14]UnevenLocal[P]156.9-5.71
Jiang et. al.[21]UnevenLocal[P]+Global[P]207.4-7.89
OursUnevenLocal[S]+Global[S]>103-9.38
OursUnevenLocal[P]+Global[P]196.3-8.25
VVC---0.00
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.581 + ], + "angle": 0, + "content": "formance when contrasted with utilizing only local context. The evaluation emphasizes that the integration of both global and local context enhances the precision of capturing spatial correlations, ultimately leading to a more accurate entropy model. In addition, our results verify that incorporating Laplacian-shaped positional encoding enhances the compression efficiency compared to MEM, which does not consider positional encoding in computing global spatial context." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.62, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We developed a neural image compression model which improves perceptual image quality using a non-isotropic diffusion decoder. This decoder's inductive bias effectively separates frequency components, leading to the creation of high-quality images. Moreover, we introduce an innovative entropy model that optimizes the trade-off between compression efficiency and decoding speed. This entropy model, founded on the Transformer architecture with Laplacian-shaped positional encoding, establishes a strong global spatial context. Our results underscore the efficacy of leveraging diffusion models and advanced entropy modeling to achieve outstanding image compression performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement: This research is based upon work supported by the National Aeronautics and Space Administration (NASA), via award number 80NSSC21M0322 under the title of Adaptive and Scalable Data Compression for Deep Space Data Transfer Applications using Deep Learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3052" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] CLIC·challenge on learned image compression., Available at http://compression.cc/tasks/, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Kodak image dataset., Available at https://r0k.us/ graphics/kodak/, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[3] Versatile Video Coding Reference Software. Available at https://vcgi.hhi.fraunhofer.de/jvet/VVCSoftware_VTM, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.282 + ], + "angle": 0, + "content": "[4] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.34 + ], + "angle": 0, + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[6] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.37, + 0.468, + 0.423 + ], + "angle": 0, + "content": "[7] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[8] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9066-9075, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.549 + ], + "angle": 0, + "content": "[9] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[11] Noor Fathima Ghouse, Jens Petersen, Auke Wiggers, Tianlin Xu, and Guillaume Sautiere. A residual diffusion model for high perceptual quality codec augmentation. arXiv preprint arXiv:2301.05489, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.468, + 0.69 + ], + "angle": 0, + "content": "[12] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[13] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.844 + ], + "angle": 0, + "content": "[14] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[15] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision and Pattern Recognition, pages 1764-1769, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.391 + ], + "angle": 0, + "content": "[20] Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. arXiv preprint arXiv:2209.05557, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.393, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[21] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.506, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[23] A Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach. Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In European Conference on Computer Vision, pages 447-463. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[24] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beach. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.648, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[25] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3053" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[30] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.258 + ], + "angle": 0, + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[32] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Li Hao, and Rong Jin. Learning accurate entropy model with global reference for image compression. In International Conference on Learning Representations, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[33] Yichen Qian, Xiuyu Sun, Ming Lin, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. In International Conference on Learning Representations, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.439 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.47, + 0.482 + ], + "angle": 0, + "content": "[35] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.47, + 0.551 + ], + "angle": 0, + "content": "[36] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[38] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International Conference on Learning Representations, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.719 + ], + "angle": 0, + "content": "[39] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[40] Dezhao Wang, Wenhan Yang, Yueyu Hu, and Jiaying Liu. Neural data-dependent transform for learned image compression. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 17379-17388, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[41] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[42] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950, 2022. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[43] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[44] Yibo Yang, Stephan Mandt, Lucas Theis, et al. An introduction to neural data compression. Foundations and Trends® in Computer Graphics and Vision, 15(2):113-200, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[45] Ali Zafari, Atefeh Khoshkhahtinat, Piyush Mehta, Mohammad Saeed Ebrahimi Saadabadi, Mohammad Akyash, and Nasser M Nasrabadi. Frequency disentangled features in neural image compression. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2815-2819. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[47] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3054" + } + ] +] \ No newline at end of file diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_origin.pdf b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a376ebad03670ab3c4f4991c4badebb75913bdc3 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/3c1b4fc3-944d-4df5-a8e9-c0d68cc7d2d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2984663ccfdb5fbd9868d1be4cad222e90bc76509afd1a383b610372b60ffdde +size 4015423 diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/full.md b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c01e7a3bdbe80f9c904a689f98d39aa09f37d69 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/full.md @@ -0,0 +1,350 @@ +# Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis + +Atefeh Khoshkhahtinat, Ali Zafari, Piyush M. Mehta, Nasser M. Nasrabadi +West Virginia University, WV, USA + +{ak00043, az00004}@mix.wvu.edu, {piyush.mehta, nasser.nasrabadi}@mail.wvu.edu + +# Abstract + +While replacing Gaussian decoders with a conditional diffusion model enhances the perceptual quality of reconstructions in neural image compression, their lack of inductive bias for image data restricts their ability to achieve state-of-the-art perceptual levels. To address this limitation, we adopt a non-isotropic diffusion model at the decoder side. This model imposes an inductive bias aimed at distinguishing between frequency contents, thereby facilitating the generation of high-quality images. Moreover, our framework is equipped with a novel entropy model that accurately models the probability distribution of latent representation by exploiting spatio-channel correlations in latent space, while accelerating the entropy decoding step. This channel-wise entropy model leverages both local and global spatial contexts within each channel chunk. The global spatial context is built upon the Transformer, which is specifically designed for image compression tasks. The designed Transformer employs a Laplacian-shaped positional encoding, the learnable parameters of which are adaptively adjusted for each channel cluster. Our experiments demonstrate that our proposed framework yields better perceptual quality compared to cutting-edge generative-based CODECs, and the proposed entropy model contributes to notable bitrate savings. The code is available at https://github.com/Atefeh-Khoshtinat/Blurdissipated-compression. + +# 1. Introduction + +Image compression is a crucial tack in image processing which aims to decrease the amount of data required for storing or transmitting without significant loss of visual content. Recently, learning-based image compression methods [13, 25, 42, 45] have demonstrated the potential to surpass classical hand-engineered CODECs in terms of rate-distortion performance. Learned image compression commonly comprises three steps: transformation, quantization, and lossless entropy coding, which resemble the components found in the traditional transform coding paradigm [44]. + +![](images/f6f6058ebfc2b1442af93af3be3744a15ea45ea0738d498262b3ab1647638a44.jpg) +Figure 1. Overview of our proposed neural codec. The quantized semantic latent variable $\hat{\pmb{y}}$ is utilized by a diffusion-based decoder to generate realistically reconstructed image. + +Neural image compression methods are typically trained with two primary objectives: minimizing distortion between the original input image and its reconstructed version and reducing the bitrate for efficient data transmission or storage [38]. However, solely optimizing for rate-distortion may lead to blurry reconstructions. To address this issue, various generative-based CODECs [4, 11, 28, 42] have emerged, aiming to generate reconstructions that are not only faithful to the original data but also visually realistic to human observers.[28] introduced HiFiC, an image compression network based on GANs that aims to achieve a balance between fidelity and perceptual quality. The network demonstrated substantial advantages for human observers, surpassing BPG even when using half the bits. However, HiFiC faces challenges due to the instability of GAN training and necessitates various design adjustments. Inspired by the achievements of Denoising Diffusion Probabilistic Models (DDPMs)[17] in generative modeling, authors in [42] replaced the Gaussian decoder of VAEs with a conditional diffusion model to mitigate the issue of blurriness in decoded images. Despite the impressive performance of diffusion-based network, its diffusion and denoising processes do not explicitly incorporate the inductive biases inherent to natural images. The imposing inductive biases, + +such as the multi-scale nature of images, enable the generation of high-quality samples [20]. + +In addition to enhancing the quality of reconstructed images, accurate entropy estimation of latent representations is vital for boosting compression efficiency. The more the estimated entropy gets closer to the actual entropy of the latent, the lower the bitrate required for generating bitstream file. In this respect, several works have been proposed. Balle et al.[5] introduced an entropy model that is conditioned on an additional latent variable, named the hyperprior, to capture the existing spatial redundancies within the latent space. Studies [24, 30] introduced a sequential autoregressive block into the entropy model. This block leverages causally adjacent latent elements to estimate the distribution of the current latent, which results in slowing down of the decoding step. He et al.[13] utilized a two-way parallel context model as a substitute for the sequential autoregressive context model, aiming to accelerate the decoding of latent code. In another realm of studies, to mitigate the slow decoding times, authors in [29] introduced channel-wise context modeling as an alternative to serial spatial context. To expedite the extracting of the channel-conditional context model, He et al. [14] exploited an unevenly segmented channel-wise autoregressive model, wherein the channels are divided into varying sizes. Strengthening the entropy model through self-attention was initially proposed in [32]. This approach employed a global reference model to identify the most suitable spatial context, which doesn't have to be confined to the local scope. The idea of integrating global context into the entropy model was subsequently refined in a later work [33], where a specially crafted transformer-based entropy model was employed. This entropy model is equipped with the diamond-relative position encoding to effectively model long-distance dependencies. The diamond-shaped relative positional encoding is devised based on the clip function, which relies on the pre-defined threshold to characterize an effective receptive field. This approach is effective in computing context based on the positions of latent elements; yet, it cannot provide an optimal receptive field due to its reliance on the clip function. Recent works[21, 23] focused on developing adaptive context models to extract both channel and spatial correlations within the latent space while expediting the entropy decoding step. Authors in [21] proposed an entropy model named MEM, aiming to capture both local and global spatial correlations in each channel segment by using a parallel bidirectional context approach. However, despite MEM's attention to spatio-channel correlations, it appears to overlook the crucial role of positional encoding in computing long-range spatial dependencies. + +To address the aforementioned challenges, we propose a conditional diffusion-based decoder in the neural compression pipeline to enhance the realism of reconstructed + +images. The traditional diffusion models are generalized into a conditional non-isotropic diffusion model to incorporate an inductive bias, considering the relative importance of each frequency component of images. As a result, each frequency component of an image undergoes diffusion at distinct rates, leading to the generation of decoded images in a coarse-to-fine manner. Furthermore, a novel channel-conditional autoregressive entropy model is introduced to efficiently take into account both local and global spatial dependencies within each channel chunk of the latent space. To leverage global spatial context, a Transformer block is designed which incorporates a Laplacian-shaped positional encoding within checkerboard-shaped self-attention module. The effective Laplacian-shaped receptive field for each channel chunk is dynamically determined by learning the positional encoding parameters during the optimization of the entire neural codec. + +# 2. Related Works + +# 2.1. Learned Image Compression + +Neural image compression methods adopt a non-linear transform coding paradigm which is founded on the variational autoencoders (VAEs) [12]. Within this scheme, the encoder initially maps the input image $\pmb{x}$ to a compact latent representation $\pmb{y}$ . Subsequently, quantization is applied to this latent representation, leading to discrete symbols that minimize the bit requirement. Afterward, entropy coding generates bit strings using an entropy model. Ultimately, the reconstructed image $\hat{\pmb{x}}$ is generated by applying the decoder to the quantized latent representation $\hat{\pmb{y}}$ . This framework is commonly trained with a trade-off between the rate and distortion, such as Mean Squared Error (MSE) [38]. + +Recent studies have attempted to enhance the realism of reconstructed images by optimizing neural CODEs using a triple rate-distortion-perception loss [4, 28, 43]. Blau and Michaeli [7] explored the essential balance between distortion and realism, demonstrating that within a given rate, improving distortion comes at the cost of decreasing the perceptual quality of a reconstructed image. In line with this premise where perceptual quality is measured as the difference between the image distribution and the distribution of reconstructions, Mentzer et al. [28] employed a conditional GAN within an autoencoder-based compression model to enhance perceptual quality. In a similar vein, [15] introduced a model that aimed to improve the realism by enriching the loss function with an adversarial perceptual loss using LPIPS [46], along with a patch-based style loss [10]. Recently, Agustsson et al. [4] developed a Multi-Realism model that merges the ELIC codec [14] and PatchGAN [8], showing improved performance compared to HiFiC. + +# 2.2. Diffusion Models + +Diffusion models [17, 22, 37], which belong to the family of score-based generative models, acquire the data distribution through a gradual iterative denoising process, starting from a Gaussian distribution and progressing towards the actual data distribution. Recently, they have received considerable attention owing to their training stability and high quality image generation compared to GANs. Diffusion models are adopted in various domains, including image and video generation [18, 19, 34], super-resolution [36], inpainting [27], deblurring [41], compression [11, 39, 42]. + +In the domain of image compression, several diffusion-based CODEs have been proposed. The research conducted by Theis et al. [39] involved employing a general unconditional diffusion model for transmitting Gaussian samples in a lossy manner. Their approach leverages a reverse channel coding concept, which stands apart from the conventional transform coding scheme. Although their methodology performs well, it suffers from high computational cost which makes it unfeasible for dealing with high-resolution images. Yang and Mandt [42] introduced a neural codec, called CDC, in which the decoder takes the form of a DDPM, conditioned on a quantized latent representation. Ghouse et al. [11] began by optimizing an autoencoder network using a rate-distortion loss. Subsequently, they train a conditional diffusion model on the output of the decoder, with the goal of enhancing its perceptual quality. + +# 2.3. Neural Entropy Model + +The main purpose of the entropy model is to estimate the joint probability distribution over the quantized latent representation. When the learned entropy model precisely matches the true distribution of the latent representation, a lower bit-rate is required to generate a compressed file. Entropy estimation can take advantage of two key principles: backward and forward adaptation. Forward adaptation uses an extra latent variable, called hyper-prior [5], as a side information to capture spatial dependencies between elements of the latent representation. Backward adaptation, on the other hand, employs an autoregressive-based context model which leverages the previously decoded elements of the latent to predict the probability of the current element. The context model of backward methods can extract relationships between the current symbol and previously decoded symbols in various dimensions, including spatial and channel axes [13, 14, 21, 23, 29, 30, 32, 33, 47]. + +Minnen et al.[30] incorporated a local spatial context with a hyper-prior network to precisely predict distribution of the latent. This type of context model can exploit local correlations between the current latent and its neighboring causal elements using masked convolutions, which consequently necessitates the adoption of serial decoding. To expedite the decoding process, He et al.[13] evenly parti + +tioned the latent into two groups: anchors and non-anchors. The anchors are exclusively decoded using the hyper-prior, while the non-anchors utilize both the hyper-prior and the local context model. This approach allows for the parallel decoding of both anchors and non-anchors. + +An alternative strategy for parallelizing the decoding process is to extract inter-channel dependencies. In [29] the latent code is divided into evenly sized chunks along the channel dimension, with each of these chunks being decoded using information from previously decoded ones. To speed up the channel-wise context extraction, He et al.[14] partitioned channels of the latent into an uneven set of groups, allocating fewer channels to the initial groups and more channels to the subsequent ones. This uneven division is motivated by the observation that the earlier channels possess higher entropy compared to the remaining channels. + +Some works introduced adaptive context models which focus on capturing long-range spatial correlations. Qian et al.[32] devised a global reference context model which assesses throughout previously decoded elements to detect the most similar one to the current latent. The authors in [33] employed a Transformer-based entropy model, called Entroformer, for capturing long-range contexts. The Entroformer benefits from diamond-relative position encoding and a top-k self-attention mechanism, both of which contribute to providing an efficient receptive field. + +# 3. Methods + +In this work, we develop our decoder based on the blurring diffusion model [20], which is conditioned on the discrete latent representation and enables the differentiation among the frequency components of an image throughout the diffusion and denoising processes. The architecture of decoder will be explained in Appendix. Moreover, a novel entropy model is employed that efficiently encodes the quantized latent representation into a binary stream. + +# 3.1. Blurring Diffusion Model + +We can define the blurring diffusion model as a traditional diffusion model in the frequency space, but with distinct schedules for each dimension of data [20]. The specifics of the schedules for vectors $\alpha_{t}$ and $\sigma_{t}$ are crucial in this type of model, which will be discussed. + +Diffusion Process: The diffusion process [17] progressively degrades the clean image, transforming it into nearly pure Gaussian noise over the course of T time steps. Each step of the diffusion process in frequency space can be formulated as: + +$$ +q \left(\boldsymbol {f} _ {t} \mid \boldsymbol {f} _ {\boldsymbol {x}}\right) = \mathcal {N} \left(\boldsymbol {f} _ {t}; \alpha_ {t} \boldsymbol {f} _ {\boldsymbol {x}}, \sigma_ {t} ^ {2} \boldsymbol {I}\right), \tag {1} +$$ + +where $f_{t}$ and $f_{x}$ are defined as $\mathbf{V}^T\mathbf{z}_t$ and $\mathbf{V}^T\mathbf{x}$ , respectively. In such a model, each frequency component could + +undergo diffusion at varying rates, with the modulation of these rates being determined by the values present in the vectors $\alpha_{t}$ and $\sigma_{t}$ . + +Denoising Process: The true denoising process in the frequency space, which is equal to the deblurring operation in the time domain, can be expressed as: + +$$ +q \left(\boldsymbol {f} _ {t - 1} \mid \boldsymbol {f} _ {t}, \boldsymbol {f} _ {x}\right) = \mathcal {N} \left(\boldsymbol {f} _ {t - 1}; \mu_ {t \rightarrow t - 1}, \sigma^ {2} _ {t \rightarrow t - 1} \boldsymbol {I}\right), \tag {2} +$$ + +where $\sigma_{t\rightarrow t - 1} = \sigma_{t|t - 1}\sigma_{t - 1} / \sigma_t$ and $\mu_{t\to t - 1} = (\alpha_{t|t - 1}\sigma_{t - 1}^2 /\sigma_t^2)f_t + (\alpha_{t - 1}\sigma_{t|t - 1}^2 /\sigma_t^2)f_x.$ The actual denoising process can be approximated using a learned denoising distribution to generate data. Thus, by considering the epsilon reparametrization of Eq. 2, we can derive the expression for the learned denoising process in frequency space: + +$$ +p _ {\theta} \left(\boldsymbol {f} _ {t - 1} \mid \boldsymbol {f} _ {t}\right) = \mathcal {N} \left(\boldsymbol {f} _ {t - 1}; \mu_ {t \rightarrow t - 1} \left(\hat {\boldsymbol {f}} _ {\boldsymbol {x}}, \boldsymbol {f} _ {t}\right), \sigma^ {2} _ {t \rightarrow t - 1} \boldsymbol {I}\right) +$$ + +Re-parametrization: $\hat{f}_{\pmb{x}} = (1 / \alpha_{t})(f_{t} - \sigma_{t}\hat{f}_{\epsilon})$ (3) + +By substituting the approximation of $\hat{f}_x$ into the true denoising distribution, the mean of the denoising network can be derived by $(\alpha_{t|t-1}\sigma_{t-1}^2/\sigma_t^2)\pmb{f}_t + (\alpha_{t-1}\sigma_{t|t-1}^2/\sigma_t^2\alpha_t)(\pmb{f}_t - \sigma_t\hat{f}_\epsilon)$ . + +Optimization: According to [17], the loss function is simplified as below: + +$$ +\boldsymbol {E} _ {t, \boldsymbol {x}, \epsilon} [ \| \boldsymbol {f} _ {\epsilon} - \hat {\boldsymbol {f}} _ {\epsilon} \| ^ {2} ] = \boldsymbol {E} [ \| \boldsymbol {V} ^ {\boldsymbol {T}} (\epsilon - \hat {\epsilon}) \| ^ {2} ] \approx \boldsymbol {E} _ {t, \boldsymbol {x}, \epsilon} [ \| \epsilon - \hat {\epsilon} \| ^ {2} ], \tag {4} +$$ + +where $\hat{\epsilon} = \phi_{\theta}(z_t, t)$ and $z_t = V(\alpha_t V^T x + \sigma_t V^T \epsilon)$ . The $\hat{\epsilon} = \phi_{\theta}(z_t, t)$ signifies that the neural network $\phi_{\theta}$ takes $z_t$ as input and predicts $\hat{\epsilon}$ . This formulation satisfies the requirement for neural networks to perform effectively in a standard pixel space. After prediction, it is sufficient to transition back and forth between frequency space and pixel space using the DCT matrix $V^T$ and its inverse $V$ . + +# 3.2. Schedules of Blurring Diffusion Model + +The schedules for the blurring diffusion model, denoted as $\alpha_{t}$ and $\sigma_{t}$ , are obtained by combining a standard Gaussian noise diffusion schedule (with scalar values $\sigma_{t}$ and $\alpha_{t}$ ) along with a blurring schedule $d_{t}$ . Each element of the vector $\sigma_{t}$ shares the same value, as identical noise is added to all frequency components. Therefore, it becomes adequate to present a schedule for a scalar value $\sigma_{t}$ [20]. The noise schedule is chosen based on a variance-preserving cosine [31], specifically $\sigma_{t}^{2} = 1 - \alpha_{t}^{2}$ , where $\alpha_{t} = \cos(t\pi / 2T)$ for $t \in [0, T]$ . Following [35], the blurring schedule is then defined as: + +$$ +\sigma_ {B, t} = \sigma_ {B, m a x} \sin (t \pi / 2 T) ^ {2}, \tag {5} +$$ + +where $\sigma_{B,max}$ represents a hyperparameter equal to the maximum level of blur applied to the image. This schedule, in turn, corresponds to the dissipation time through + +$\tau_{t} = \sigma_{B,t}^{2} / 2$ . Based on the formulation discussed in the Appendix, the blurring schedule $d_{t}$ , which is employed for $\alpha_{t}$ , is defined as follows: + +$$ +\boldsymbol {d} _ {\boldsymbol {t}} = \exp (- \lambda \tau_ {t}), \tag {6} +$$ + +where $\lambda$ represents the vector containing squared frequencies, and $\tau_{t}$ corresponds to the dissipation time. To achieve a more gradual amplification of high frequencies during the denoising process, the blurring schedule $d_{t}$ is adjusted to $(1 - d_{min})\exp (-\lambda \tau_{t}) + d_{min}$ , where $d_{min}$ is set to 0.001. Finally, by combining the Gaussian noise schedule with the blurring schedule, resultant schedule is as follows: + +$$ +\boldsymbol {\alpha} _ {t} = \alpha_ {t}. \boldsymbol {d} _ {t}, \boldsymbol {\sigma} _ {t} = \mathbf {1} \sigma_ {t}. \tag {7} +$$ + +# 3.3. Blurring Diffusion Model for Compression + +The rate-distortion objective in end-to-end compression resembles the loss function of a $\beta$ -VAE, where a hyperparameter $\lambda$ is utilized to balance the trade-off between the bit-rate $(R)$ and distortion $(D)$ : + +$$ +\mathrm {L} = D + \lambda R = \boldsymbol {E} _ {\tilde {\boldsymbol {y}}} \left[ - \log p _ {\boldsymbol {x} | \tilde {\boldsymbol {y}}} (\boldsymbol {x} | \tilde {\boldsymbol {y}}) - \lambda \log p _ {\tilde {\boldsymbol {y}}} (\tilde {\boldsymbol {y}}) \right]. \tag {8} +$$ + +While many neural codecs commonly employ Gaussian or Laplacian decoders, we introduce a novel approach using a conditional blurring diffusion model as the decoder. This approach aims to yield new distortion metrics that deviate from those based on Mean Squared Error (MSE) or Mean Absolute Error (MAE). In this framework, our proposed neural codec leverages two distinct types of latent variables: a semantic latent variable $\mathbf{y}$ and texture latent variables $\mathbf{z}_{1:T}$ . The semantic latent variable captures and encodes the overall content and meaning of the image. In contrast, the texture latent variables are tailored to carry finer details and intricate patterns that might not be fully represented by the semantic variable alone. Notably, unlike the semantic latent variable, the texture latent variables are not compressed but synthesized during the decoding phase. + +The forward and backward processes of the conditional blurring diffusion model can be expressed as: + +$$ +q (\pmb {f} _ {t} | \pmb {f} _ {t - 1}) = \mathcal {N} (\pmb {f} _ {t}; \alpha_ {t | t - 1} \pmb {f} _ {t - 1}, \sigma_ {t | t - 1} ^ {2} I), +$$ + +$$ +p _ {\theta} \left(\boldsymbol {f} _ {\boldsymbol {x}}, \boldsymbol {f} _ {\boldsymbol {1}: T} | \tilde {\boldsymbol {y}}\right) = p \left(\boldsymbol {f} _ {\boldsymbol {T}}\right) p _ {\theta} \left(\boldsymbol {f} _ {\boldsymbol {x}} \mid \boldsymbol {f} _ {\boldsymbol {1}}, \tilde {\boldsymbol {y}}\right) \prod_ {t = 2} ^ {T} p _ {\theta} \left(\boldsymbol {f} _ {\boldsymbol {t} - \boldsymbol {1}} \mid \boldsymbol {f} _ {\boldsymbol {t}}, \tilde {\boldsymbol {y}}\right), +$$ + +$$ +p (\boldsymbol {f} _ {T}) = \mathcal {N} (\boldsymbol {f} _ {T}; \mathbf {0}, \boldsymbol {I}) +$$ + +$$ +p _ {\theta} \left(\boldsymbol {f} _ {t - 1} \mid \boldsymbol {f} _ {t}, \tilde {\boldsymbol {y}}\right) = \mathcal {N} \left(\boldsymbol {f} _ {t - 1}; \mu_ {\theta} \left(\boldsymbol {f} _ {t}, \tilde {\boldsymbol {y}}, t\right), \sigma^ {2} _ {t \rightarrow t - 1} \boldsymbol {I}\right) \tag {9} +$$ + +where $f_{t} = V^{T}z_{t}$ and $f_{x} = V^{T}x = V^{T}z_{o}$ . As shown in Eq. 8, distortion is equivalent to the negative marginal likelihood of input data $-\log p_{\boldsymbol{x}|\tilde{\boldsymbol{y}}}(\boldsymbol{x}|\tilde{\boldsymbol{y}})$ , and minimizing it + +is analogous to minimizing the negative marginal likelihood of the frequency representation of the data $p_{\pmb{f}_x|\tilde{\pmb{y}}}(\pmb{f}_x|\tilde{\pmb{y}}) = \int p(\pmb{f}_x, \pmb{f}_{1:T}|\tilde{\pmb{y}}) d\pmb{f}_{1:T}$ . Since computing the marginal likelihood is intractable, we employ its ELBO with the specified diffusion and de-blurring distributions. Following this substitution, the rate-distortion objective, through the application of Jensen's inequality, can be formulated as follows: + +$$ +\begin{array}{l} \boldsymbol {E} _ {\tilde {\boldsymbol {y}}} \left[ - \log p _ {\boldsymbol {x} | \tilde {\boldsymbol {y}}} (\boldsymbol {x} | \tilde {\boldsymbol {y}}) - \lambda \log p _ {\tilde {\boldsymbol {y}}} (\tilde {\boldsymbol {y}}) \right] \\ \leq \boldsymbol {E} _ {\tilde {\boldsymbol {y}}} [ - \log_ {P} \boldsymbol {x} | \boldsymbol {y} (\omega | \boldsymbol {y}) - \lambda \log_ {P} \boldsymbol {y} (\boldsymbol {y}) ], \tag {10} \\ \end{array} +$$ + +where $\mathrm{ELBO} = E_q[\frac{p_\theta(f_x,f_{1:T}|\tilde{y})}{q(f_t|f_{t - 1})} ]$ . As explained in Appendix, the optimizing of ELBO can be simplified to: + +$$ +\operatorname {E L B O} \approx \boldsymbol {E} _ {t, \boldsymbol {x}, \epsilon , \tilde {\boldsymbol {y}}} [ \| \epsilon - \phi_ {\theta} (\boldsymbol {z} _ {t}, t, \tilde {\boldsymbol {y}}) \| ^ {2} ], \tag {11} +$$ + +where $z_{t} = V\alpha_{t}V^{T}x + \sigma_{t}\epsilon$ . Similar to [28], an LPIPS loss [46] is incorporated into the rate-distortion loss to enhance the perceptual quality of the reconstructed image. As the initial image at any time step can be decoded as a function of the texture latent variable $z_{t}$ , the semantic latent variable $y$ , and the time step $t$ , i.e., $\hat{x}_t = V(1 / \alpha_t)(V^T z_t - \sigma_t V^T \hat{\epsilon})$ , the total loss becomes as: + +$$ +\begin{array}{l} \boldsymbol {E} _ {t, \boldsymbol {x}, \epsilon , \tilde {\boldsymbol {y}}} [ (1 - \beta) \| \epsilon - \phi_ {\theta} (\boldsymbol {z} _ {t}, t, \tilde {\boldsymbol {y}}) \| ^ {2} \tag {12} \\ \left. - \lambda \log p _ {\tilde {\mathbf {y}}} (\tilde {\mathbf {y}}) + \beta d _ {\mathrm {L P I P S}} (\boldsymbol {x}, \hat {\boldsymbol {x}} _ {t}) \right], \\ \end{array} +$$ + +where the $\lambda$ and $\beta$ represent hyperparameters that control the trade-off between rate, distortion, and perception. + +Decoding Process: After the training, we utilize entropy decoding on $\hat{\pmb{y}}$ through the entropy model which estimates the distribution $p_{\hat{\pmb{y}}}(\hat{\pmb{y}})$ . We then employ ancestral sampling to conditionally decode the image $\pmb{x}$ , which is equal to $\pmb{z_0}$ and starts from pure Gaussian noise $\pmb{z}_T\sim \mathcal{N}(\pmb {0},\pmb {I})$ : + +$$ +\boldsymbol {z} _ {t - 1} \gets \boldsymbol {V} (\hat {\boldsymbol {\mu}} _ {t \rightarrow t - 1} + \boldsymbol {\sigma} _ {t \rightarrow t - 1} \boldsymbol {V} ^ {T} \phi_ {\theta} (\boldsymbol {z} _ {t}, t, \hat {\boldsymbol {y}})) \quad (1 3) +$$ + +where $\hat{\mu}_{t\rightarrow t - 1} = \frac{\alpha_{t|t - 1}\sigma_{t - 1}^2}{\sigma_t^2} V^T z_t + \frac{\sigma_{t|t - 1}^2}{\alpha_{t|t - 1}\sigma_t^2} (V^T z_t - \sigma_t V^T\phi_\theta (z_t,t,\hat{y}))$ . The latent variables $z_{1:T}$ are not stored but produced during the decoding. + +# 3.4. Proposed Entropy Model + +The main goal of the proposed entropy model' context is to exploit both channel-wise and spatial correlations, while expediting the decoding process. Inspired by the ELIC [14], we adopt an uneven grouping of latent channels, allowing most low entropy channels to depend on high entropy channels. The latent representation $\hat{\pmb{y}}$ , with $M$ channels, is clustered into five chunks along the channel dimension: 16, 16, 32, 64, and M - 128 channels, respectively. In this setup, each chunk depends on all its previous decoded chunks. + +The Fig. 2(a) illustrates our proposed entropy model as applied in the $j$ -th chunk. Within each chunk, we use spatial + +![](images/1dc4e0096d8a53b6a62039eb526558c3fb106b94c6036b388ac4daf4f273c896.jpg) +Figure 2. Diagram illustrating the application of the proposed entropy model for decoding the $j$ -th chunk $\hat{\pmb{y}}^{(j)}$ . (b) Global Spatial Context Block. (c) An example of a checkerboard-shaped mask. + +context in conjunction with channel context to model correlations along both the channel and spatial dimensions. To expedite the decoding step, we employ a parallel bidirectional spatial context model which is capable of capturing both local and global spatial relationships. So, the anchor part is decoded in parallel by using solely the hyperprior and channel context, while the decoding of the non-anchor part relies on the hyperprior, as well as both the spatial and channel contexts. + +# 3.4.1 Spatial Context + +The spatial context design captures both local and global spatial correlations within the latent representation $\hat{\pmb{y}}$ . Following the decoding of the anchor group, a checkerboard-shaped convolution is applied to this group, generating local context for all non-anchor elements in a parallel manner. Furthermore, the acquired local contexts of the non-anchor part are subsequently fed into a Transformer-based block to efficiently extract the global spatial context. This Transformer block takes advantage of positional encoding that is customized specifically for the compression task, along with a checkerboard-based attention mechanism. + +# 3.4.2 Transformer-based Spatial Context + +We adopt the Swin Transformer [26] blocks, including a masked window-based multi-head self-attention (W-MSA) and a masked shifted-window-based multi-head self-attention (SW-MSA), as our global spatial context model which shown in Fig. 2.(b). This selection enables us to strike a balance between computational efficiency and modeling capacity. The checkerboard-based self-attention in W-MAS and SW-MSA can be expressed: + +$$ +A t t e n (\boldsymbol {Q}, \boldsymbol {K}, \boldsymbol {V}) = \operatorname {s o f t m a x} (\boldsymbol {Q} \boldsymbol {K} ^ {T} \odot \boldsymbol {M}) \boldsymbol {V}, \tag {14} +$$ + +![](images/8bc3daba084e255fffb60e34138a9a2450ed7c96c5029372656da8fb4e68d630.jpg) +Figure 3. The procedure for acquiring Laplacian relative position encoding for a window with a size of $2 \times 2$ . + +where $Q, K, V \in R^{N^2 \times d}$ are the query, key, and value matrices, respectively. $N^2$ denotes the number of patches in a window and $d$ is the dimension of query/key/value. $M \in R^{N^2 \times N^2}$ represents a checkerboard-shaped mask. + +# 3.4.3 Relative Positional Encoding + +The authors in [13, 33] have shown that the bitrate is impacted by the distance of neighboring latents within the spatial context modeling of current latent. Based on this insight, we introduce a receptive field-aware self-attention mechanism that employs a learnable Laplacian-shaped positional encoding for calculating spatial context. In our proposed approach, each channel chunk possesses its own receptive field during self-attention computation, which dynamically adjusts in response to changes in entropy of chunks. In the $j$ -th chunk, the checkerboard-based attention of W-MSA and SW-MSA is refined to receptive field-aware self-attention as follows: + +$$ +A t t e n \left(\boldsymbol {Q} _ {j}, \boldsymbol {K} _ {j}, \boldsymbol {V} _ {j}\right) = \text {s o f t m a x} \left(\boldsymbol {Q} _ {j} \boldsymbol {k} _ {j} ^ {T} \odot \boldsymbol {M} + \boldsymbol {P} _ {\text {L a p}, j}\right) \boldsymbol {V} _ {j}, \tag {15} +$$ + +where $P_{Lap,j}$ represents the learnable Laplacian relative position encoding for chunk $j$ . To create the Laplacian relative position encoding for the $j$ -th chunk, which comprises spatial windows with a size of $N \times N$ (i.e. including $N^2$ patches), three steps need to be taken. Firstly, We generate a 2D relative position table, where each coordinate of the relative position lies in the range $[-N + 1, N - 1]$ . Subsequently, we derive the relative position matrix for each patch. As shown in Fig. 3, the first patch's relative distance coordinate of $(0,0)$ (relative distance with its position) is located at the top-left of the orange box, while the last patch's relative distance of $(0,0)$ is positioned at the bottom-right of the green box. Afterward, each relative position matrix is flattened and stacked together. Finally, we apply the 2D Laplacian function to each element of the resulting matrix to generate a Laplacian relative position encoding $P_{Lap,j}$ with learnable parameters $A_j$ and $\sigma_j$ , whose size is $R^{N^2 \times N^2}$ . The 2D Laplacian function is defined as: + +$$ +f _ {L a p} (x, y) = A _ {j} ^ {2} \exp \left(\left(- 1 / 2 \sigma_ {j} ^ {2}\right) \left(| x | + | y |\right)\right), \tag {16} +$$ + +where $x\in \{-N + 1,\dots,N - 1\}$ $y\in \{-N + 1,\dots,N - 1\}$ $A_{j}$ and $\sigma_{j}$ are learnable parameters which are determined + +for each chunk through optimization. The value of them is associated with the wideness of the effective receptive field. + +# 4. Experiments + +# 4.1. Implementation Details + +To train our learned image compression network, we utilize a merged dataset that includes high-resolution images from the DIV2K, Flickr2K, and CLIC [1] datasets. These images are randomly cropped to a size of $256 \times 256$ during the training phase. The model parameters of all architectures were optimized using the Adam optimizer for a total of 2.4 million steps, with a batch size of 8. The initial learning rate was configured to be $1 \times 10^{-4}$ and was progressively reduced until the conclusion of training, reaching $1 \times 10^{-7}$ . In order to cover a broad spectrum of bitrates, we selected a hyperparameter $\lambda$ from the set $\{0.0004, 0.005, 0.01, 0.02, 0.04, 0.016\}$ . The hyperparameter $\beta$ which specifies the contribution of perceptual loss is considered 0.9 for all models. The parameter $T$ , represents the required time steps for diffusion-based decoder, is consistently set to 500. To evaluate the performance of our compression approach, we select the Kodak dataset [2], which comprises 24 high-quality images with a resolution of $768 \times 512$ , and the CLIC2020 test set (428 images) with varying resolutions. + +# 4.2. Comparison with the SOTA Methods + +We compare our model with state-of-the-art generative based image compression networks, including both GAN-based and diffusion-based CODECs, as well as a state-of-the-art hand-crafted codec such as VVC-Intra (VTM) [3]. Additionally, we choose two VAE CODECs for comparison: DGML [9] and NSC [40]. Specifically, within the domain of GAN-based image compression approaches, we select HiFiC [28] and Multi-Realism [4] frameworks for comparison. Furthermore, we assess our method against diffusion-based models, namely DIRAC [11] and CDC [42]. This comparison is performed with respect to both rate-distortion and rate-perception tradeoff. The distortion is measured by the Peak Signal-to-Noise Ratio (PSNR) metric and perceptual quality are quantified via Frechet Inception Distance (FID) [16] and Learned Perceptual Image Patch Similarity (LPIPS). As shown in Fig. 4, our model shows superior performance compared to all other CODECs in terms of rate-perception. However, this heightened realism comes at the cost of distortion where other methods, except for the CDC model, either exhibit better performance or remain competitive. + +# 4.3. Visual Quality + +Fig. 5 illustrates reconstructed images (kodim20.png, kodim07.png) generated by our proposed model, as well as + +![](images/b16130257453f3e6bd81f3c6bf6fb5783500bca1d0bbf3ee1e6c92c937eea3f1.jpg) + +![](images/1fc7343380d5f1c99e4da77b0fe572a6cfc938c3c8f28e81d984a416d71387e0.jpg) + +![](images/7978237301d02da3664285e7c4354a1fd70900e1911819e114cfa06631546f77.jpg) + +![](images/5327daa86c66bf863571b6ee4e7eb7e5d8ec87bd1955df156ac8255918736bf1.jpg) +Figure 4. Comparison of our method with other codescs in terms of rate/distortion [bpp $\downarrow$ / PSNR $\uparrow$ ] and rate-perception, including [bpp $\downarrow$ / FID $\downarrow$ ] and [bpp $\downarrow$ / LPIPS $\downarrow$ ], for both the CLIC2020 test set and the Kodak dataset. + +![](images/0253b50e4b9e5ee61a10785eaa8e5ff199ff6660421dbac8c917b81a4f8c5a96.jpg) + +![](images/4ab8a502bce3b38a86f10bd620f2c809d014fc5486ade6a3ace9009b94899bca.jpg) + +![](images/f568de16a342910a3ca170440f73065a3194f08167543fb960aa67f4e4eb491c.jpg) +Figure 5. Visual comparison of our method to HiFiC and CDC models shows that our model achieves superior reconstruction quality, particularly at lower bit-rates. In addition, our model displays fewer artifacts compared to both the HiFiC and CDC models. + +the CDC [42] and HiFiC [28] frameworks. Perceptually-oriented neural codecs, such as HiFiC and CDC, are criticized when their decoding process encounters semantic content, such as text within the compressed image. Our model effectively addresses this challenge, as depicted in Fig. 5. It enhances perceptual quality through a diffusion-based decoder and imposes a strict inductive bias via the + +de-blurring diffusion process, resulting in both high perceptual quality and better text-preserving reconstructions. Additional qualitative comparisons are provided in the Appendix. + +# 4.4. Ablation Study + +Maximum Blurring: To explore the impact of the blurring schedule, we vary the values of $\sigma_{B,max}$ and compare the performance of the resulting models. It is evident that the model with $\sigma_{B,max} = 0$ is equivalent to a standard denoising diffusion model. As indicated in Table 1a, blurring diffusion models with higher maximum blur levels $\sigma_{B,max} = 25$ generate higher-quality reconstructed images compared to other variant models in terms of FID score. + +Table 1. Ablation Study: all models are optimized with $\lambda = {0.01}$ . (a) Investigating the impact of maximum blurring ${\sigma }_{B,\max }$ . (b) Exploring the effects of different types of positional encoding + +
σB,maxFID
03.94
53.78
153.56
253.45
+ +(a) + +
Position Enc.bpp
-0.2811
Relative Pos.0.2643
Diamond Relative Pos.0.2512
Laplacian-shaped Pos.0.2347
+ +(b) + +Positional Encoding: We investigate the influence of various types of positional encoding in the Transformer-based entropy model. As shown in Table 1b, the 2D diamond relative positional encoding, which is implemented using a clip function, demonstrates better performance than the relative position encoding. However, adopting the Laplacian-shaped positional encoding results in even more significant bitrate savings compared to the 2D diamond relative positional encoding. As depicted in Fig. 6, distinct values of $A$ and $\sigma$ are obtained for each channel chunk. As we progress towards the final chunks, the receptive field becomes narrower. As observed in [14], the last chunk of channels, conditioned on all previous chunks, exhibits lower entropy. Their entropy can be estimated by considering only a small neighborhood, corresponding to a narrower receptive field in our proposed relative positioning. Conversely, to provide a reasonable estimate of entropy of the first channel chunks, a wider receptive field gathers more information over a large context to compared to mentioned last chunks. + +Analysis of Context Blocks: We conducted a comparison of the inference latency for entropy parameters during entropy decoding, as well as the bitrate savings of our proposed entropy model compared to other backward adaptation-based entropy models. For a fair comparison, all the models are equipped with a same encoder comprising of ResNet blocks and convolution layers to transform the input image $\boldsymbol{x} \in R^{H \times C \times 3}$ to a latent representation $\boldsymbol{y} \in R^{H / 16 \times W / 16 \times 256}$ . As reported in Tabel 2, our entropy model's speed is notably improved by employing unevenly grouped channels and implementing a bidirectional context model, as compared to the sequential spatial context modeling. Moreover, our findings clearly indicate that incorporating spatial global context results in superior per + +![](images/aaf6816fced1ece463fb61406af7d9768360c96f9bc9213330d21eabdd709ec8.jpg) +Figure 6. Receptive fields are extracted for each channel chunk. + +![](images/abb886c94d26bcdbf75da767c9a7eaea83d0acec01962fbddac0fdcda68b6e0b.jpg) + +![](images/fdf0e772579027d8120da043af53a415ef7dd84496b21091b32a752c72fb7b2f.jpg) + +![](images/2fb3e32f6b9ebffe34f49c74d036f07cd2de45435ef2b613fcfafe1b0b548124.jpg) + +![](images/2c6685010111f4b5a2a57390062d0260ed436826ffa5ea886628f939a534b246.jpg) + +Table 2. The Bjøntegaard-delta-rate (BD-rate) [6] and inference latency of entropy parameter estimation during entropy decoding (Dec.) for different context-based models on the Kodak dataset using a GPU (RTX A6000). The BD-Rate is computed relative to VVC. ([P]:parallel, [S]:Serial) + +
MethodContext ModelDec.(ms)BD-Rate
ChannelSpatial
Minnen et. al.[30]-Local[S]>103-3.24
Minnen et. al.[29]Even-67-3.96
He et. al.[13]-Local[P]28.2-2.91
Qian et. al.[33]-Global[S]>103-4.89
He et. al.[14]Uneven-37.2-3.12
He et. al.[14]UnevenLocal[P]156.9-5.71
Jiang et. al.[21]UnevenLocal[P]+Global[P]207.4-7.89
OursUnevenLocal[S]+Global[S]>103-9.38
OursUnevenLocal[P]+Global[P]196.3-8.25
VVC---0.00
+ +formance when contrasted with utilizing only local context. The evaluation emphasizes that the integration of both global and local context enhances the precision of capturing spatial correlations, ultimately leading to a more accurate entropy model. In addition, our results verify that incorporating Laplacian-shaped positional encoding enhances the compression efficiency compared to MEM, which does not consider positional encoding in computing global spatial context. + +# 5. Conclusion + +We developed a neural image compression model which improves perceptual image quality using a non-isotropic diffusion decoder. This decoder's inductive bias effectively separates frequency components, leading to the creation of high-quality images. Moreover, we introduce an innovative entropy model that optimizes the trade-off between compression efficiency and decoding speed. This entropy model, founded on the Transformer architecture with Laplacian-shaped positional encoding, establishes a strong global spatial context. Our results underscore the efficacy of leveraging diffusion models and advanced entropy modeling to achieve outstanding image compression performance. + +Acknowledgement: This research is based upon work supported by the National Aeronautics and Space Administration (NASA), via award number 80NSSC21M0322 under the title of Adaptive and Scalable Data Compression for Deep Space Data Transfer Applications using Deep Learning. + +# References + +[1] CLIC·challenge on learned image compression., Available at http://compression.cc/tasks/, 2022.6 +[2] Kodak image dataset., Available at https://r0k.us/ graphics/kodak/, 2022.6 +[3] Versatile Video Coding Reference Software. Available at https://vcgi.hhi.fraunhofer.de/jvet/VVCSoftware_VTM, 2022.6 +[4] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 1, 2, 6 +[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2, 3 +[6] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 8 +[7] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2 +[8] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9066-9075, 2019. 2 +[9] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 6 +[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2 +[11] Noor Fathima Ghouse, Jens Petersen, Auke Wiggers, Tianlin Xu, and Guillaume Sautiere. A residual diffusion model for high perceptual quality codec augmentation. arXiv preprint arXiv:2301.05489, 2023. 1, 3, 6 +[12] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2 +[13] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 1, 2, 3, 6, 8 +[14] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2, 3, 5, 8 +[15] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 1764-1769, 2022. 2 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 4 +[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3 +[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 3 +[20] Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. arXiv preprint arXiv:2209.05557, 2022. 2, 3, 4 +[21] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 8 +[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3 +[23] A Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach. Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In European Conference on Computer Vision, pages 447-463. Springer, 2022. 2, 3 +[24] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beach. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 2 +[25] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5 +[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3 +[28] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 1, 2, 5, 6, 7 + +[29] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 8 +[30] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3, 8 +[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4 +[32] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Li Hao, and Rong Jin. Learning accurate entropy model with global reference for image compression. In International Conference on Learning Representations, 2020. 2, 3 +[33] Yichen Qian, Xiuyu Sun, Ming Lin, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. In International Conference on Learning Representations, 2021. 2, 3, 6, 8 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3 +[35] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 4 +[36] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 3 +[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[38] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International Conference on Learning Representations, 2016. 1, 2 +[39] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 3 +[40] Dezhao Wang, Wenhan Yang, Yueyu Hu, and Jiaying Liu. Neural data-dependent transform for learned image compression. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 17379-17388, 2022. 6 +[41] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 3 +[42] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950, 2022. 1, 3, 6, 7 + +[43] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2 +[44] Yibo Yang, Stephan Mandt, Lucas Theis, et al. An introduction to neural data compression. Foundations and Trends® in Computer Graphics and Vision, 15(2):113-200, 2023. 1 +[45] Ali Zafari, Atefeh Khoshkhahtinat, Piyush Mehta, Mohammad Saeed Ebrahimi Saadabadi, Mohammad Akyash, and Nasser M Nasrabadi. Frequency disentangled features in neural image compression. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2815-2819. IEEE, 2023. 1 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 5 +[47] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2021. 3 \ No newline at end of file diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/images.zip b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b74508a1337338a7252b6888bb2dd555b73dacd4 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da08d844bd6e712380c19d4eff2ec23e54dc72bab978d63308e4330472b8b62a +size 564998 diff --git a/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/layout.json b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..21bdcdf8d89567e3c3dfd4bb77b15c3d0f05cb81 --- /dev/null +++ b/2024/Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis/layout.json @@ -0,0 +1,9169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 122 + ], + "type": "text", + "content": "Laplacian-guided Entropy Model in Neural Codec with Blur-dissipated Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 143, + 478, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 143, + 478, + 171 + ], + "spans": [ + { + "bbox": [ + 114, + 143, + 478, + 171 + ], + "type": "text", + "content": "Atefeh Khoshkhahtinat, Ali Zafari, Piyush M. Mehta, Nasser M. Nasrabadi \nWest Virginia University, WV, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 173, + 496, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 173, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 97, + 173, + 496, + 186 + ], + "type": "text", + "content": "{ak00043, az00004}@mix.wvu.edu, {piyush.mehta, nasser.nasrabadi}@mail.wvu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 550 + ], + "type": "text", + "content": "While replacing Gaussian decoders with a conditional diffusion model enhances the perceptual quality of reconstructions in neural image compression, their lack of inductive bias for image data restricts their ability to achieve state-of-the-art perceptual levels. To address this limitation, we adopt a non-isotropic diffusion model at the decoder side. This model imposes an inductive bias aimed at distinguishing between frequency contents, thereby facilitating the generation of high-quality images. Moreover, our framework is equipped with a novel entropy model that accurately models the probability distribution of latent representation by exploiting spatio-channel correlations in latent space, while accelerating the entropy decoding step. This channel-wise entropy model leverages both local and global spatial contexts within each channel chunk. The global spatial context is built upon the Transformer, which is specifically designed for image compression tasks. The designed Transformer employs a Laplacian-shaped positional encoding, the learnable parameters of which are adaptively adjusted for each channel cluster. Our experiments demonstrate that our proposed framework yields better perceptual quality compared to cutting-edge generative-based CODECs, and the proposed entropy model contributes to notable bitrate savings. The code is available at https://github.com/Atefeh-Khoshtinat/Blurdissipated-compression." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 128, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 715 + ], + "type": "text", + "content": "Image compression is a crucial tack in image processing which aims to decrease the amount of data required for storing or transmitting without significant loss of visual content. Recently, learning-based image compression methods [13, 25, 42, 45] have demonstrated the potential to surpass classical hand-engineered CODECs in terms of rate-distortion performance. Learned image compression commonly comprises three steps: transformation, quantization, and lossless entropy coding, which resemble the components found in the traditional transform coding paradigm [44]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 213, + 529, + 357 + ], + "blocks": [ + { + "bbox": [ + 322, + 213, + 529, + 357 + ], + "lines": [ + { + "bbox": [ + 322, + 213, + 529, + 357 + ], + "spans": [ + { + "bbox": [ + 322, + 213, + 529, + 357 + ], + "type": "image", + "image_path": "f6f6058ebfc2b1442af93af3be3744a15ea45ea0738d498262b3ab1647638a44.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 547, + 399 + ], + "type": "text", + "content": "Figure 1. Overview of our proposed neural codec. The quantized semantic latent variable " + }, + { + "bbox": [ + 305, + 365, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 305, + 365, + 547, + 399 + ], + "type": "text", + "content": " is utilized by a diffusion-based decoder to generate realistically reconstructed image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 715 + ], + "type": "text", + "content": "Neural image compression methods are typically trained with two primary objectives: minimizing distortion between the original input image and its reconstructed version and reducing the bitrate for efficient data transmission or storage [38]. However, solely optimizing for rate-distortion may lead to blurry reconstructions. To address this issue, various generative-based CODECs [4, 11, 28, 42] have emerged, aiming to generate reconstructions that are not only faithful to the original data but also visually realistic to human observers.[28] introduced HiFiC, an image compression network based on GANs that aims to achieve a balance between fidelity and perceptual quality. The network demonstrated substantial advantages for human observers, surpassing BPG even when using half the bits. However, HiFiC faces challenges due to the instability of GAN training and necessitates various design adjustments. Inspired by the achievements of Denoising Diffusion Probabilistic Models (DDPMs)[17] in generative modeling, authors in [42] replaced the Gaussian decoder of VAEs with a conditional diffusion model to mitigate the issue of blurriness in decoded images. Despite the impressive performance of diffusion-based network, its diffusion and denoising processes do not explicitly incorporate the inductive biases inherent to natural images. The imposing inductive biases," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3045" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 285, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 285, + 96 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 285, + 96 + ], + "type": "text", + "content": "such as the multi-scale nature of images, enable the generation of high-quality samples [20]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 98, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 98, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 49, + 98, + 286, + 674 + ], + "type": "text", + "content": "In addition to enhancing the quality of reconstructed images, accurate entropy estimation of latent representations is vital for boosting compression efficiency. The more the estimated entropy gets closer to the actual entropy of the latent, the lower the bitrate required for generating bitstream file. In this respect, several works have been proposed. Balle et al.[5] introduced an entropy model that is conditioned on an additional latent variable, named the hyperprior, to capture the existing spatial redundancies within the latent space. Studies [24, 30] introduced a sequential autoregressive block into the entropy model. This block leverages causally adjacent latent elements to estimate the distribution of the current latent, which results in slowing down of the decoding step. He et al.[13] utilized a two-way parallel context model as a substitute for the sequential autoregressive context model, aiming to accelerate the decoding of latent code. In another realm of studies, to mitigate the slow decoding times, authors in [29] introduced channel-wise context modeling as an alternative to serial spatial context. To expedite the extracting of the channel-conditional context model, He et al. [14] exploited an unevenly segmented channel-wise autoregressive model, wherein the channels are divided into varying sizes. Strengthening the entropy model through self-attention was initially proposed in [32]. This approach employed a global reference model to identify the most suitable spatial context, which doesn't have to be confined to the local scope. The idea of integrating global context into the entropy model was subsequently refined in a later work [33], where a specially crafted transformer-based entropy model was employed. This entropy model is equipped with the diamond-relative position encoding to effectively model long-distance dependencies. The diamond-shaped relative positional encoding is devised based on the clip function, which relies on the pre-defined threshold to characterize an effective receptive field. This approach is effective in computing context based on the positions of latent elements; yet, it cannot provide an optimal receptive field due to its reliance on the clip function. Recent works[21, 23] focused on developing adaptive context models to extract both channel and spatial correlations within the latent space while expediting the entropy decoding step. Authors in [21] proposed an entropy model named MEM, aiming to capture both local and global spatial correlations in each channel segment by using a parallel bidirectional context approach. However, despite MEM's attention to spatio-channel correlations, it appears to overlook the crucial role of positional encoding in computing long-range spatial dependencies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose a conditional diffusion-based decoder in the neural compression pipeline to enhance the realism of reconstructed" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 307, + 72, + 545, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 274 + ], + "type": "text", + "content": "images. The traditional diffusion models are generalized into a conditional non-isotropic diffusion model to incorporate an inductive bias, considering the relative importance of each frequency component of images. As a result, each frequency component of an image undergoes diffusion at distinct rates, leading to the generation of decoded images in a coarse-to-fine manner. Furthermore, a novel channel-conditional autoregressive entropy model is introduced to efficiently take into account both local and global spatial dependencies within each channel chunk of the latent space. To leverage global spatial context, a Transformer block is designed which incorporates a Laplacian-shaped positional encoding within checkerboard-shaped self-attention module. The effective Laplacian-shaped receptive field for each channel chunk is dynamically determined by learning the positional encoding parameters during the optimization of the entire neural codec." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 300, + 395, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 395, + 313 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 395, + 313 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 326, + 462, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 462, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 462, + 338 + ], + "type": "text", + "content": "2.1. Learned Image Compression" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "content": "Neural image compression methods adopt a non-linear transform coding paradigm which is founded on the variational autoencoders (VAEs) [12]. Within this scheme, the encoder initially maps the input image " + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "content": " to a compact latent representation " + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "content": ". Subsequently, quantization is applied to this latent representation, leading to discrete symbols that minimize the bit requirement. Afterward, entropy coding generates bit strings using an entropy model. Ultimately, the reconstructed image " + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "content": " is generated by applying the decoder to the quantized latent representation " + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 307, + 350, + 545, + 493 + ], + "type": "text", + "content": ". This framework is commonly trained with a trade-off between the rate and distortion, such as Mean Squared Error (MSE) [38]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 713 + ], + "type": "text", + "content": "Recent studies have attempted to enhance the realism of reconstructed images by optimizing neural CODEs using a triple rate-distortion-perception loss [4, 28, 43]. Blau and Michaeli [7] explored the essential balance between distortion and realism, demonstrating that within a given rate, improving distortion comes at the cost of decreasing the perceptual quality of a reconstructed image. In line with this premise where perceptual quality is measured as the difference between the image distribution and the distribution of reconstructions, Mentzer et al. [28] employed a conditional GAN within an autoencoder-based compression model to enhance perceptual quality. In a similar vein, [15] introduced a model that aimed to improve the realism by enriching the loss function with an adversarial perceptual loss using LPIPS [46], along with a patch-based style loss [10]. Recently, Agustsson et al. [4] developed a Multi-Realism model that merges the ELIC codec [14] and PatchGAN [8], showing improved performance compared to HiFiC." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "3046" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 151, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 151, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 151, + 83 + ], + "type": "text", + "content": "2.2. Diffusion Models" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 89, + 287, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 89, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 89, + 287, + 209 + ], + "type": "text", + "content": "Diffusion models [17, 22, 37], which belong to the family of score-based generative models, acquire the data distribution through a gradual iterative denoising process, starting from a Gaussian distribution and progressing towards the actual data distribution. Recently, they have received considerable attention owing to their training stability and high quality image generation compared to GANs. Diffusion models are adopted in various domains, including image and video generation [18, 19, 34], super-resolution [36], inpainting [27], deblurring [41], compression [11, 39, 42]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 210, + 289, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 289, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 289, + 402 + ], + "type": "text", + "content": "In the domain of image compression, several diffusion-based CODEs have been proposed. The research conducted by Theis et al. [39] involved employing a general unconditional diffusion model for transmitting Gaussian samples in a lossy manner. Their approach leverages a reverse channel coding concept, which stands apart from the conventional transform coding scheme. Although their methodology performs well, it suffers from high computational cost which makes it unfeasible for dealing with high-resolution images. Yang and Mandt [42] introduced a neural codec, called CDC, in which the decoder takes the form of a DDPM, conditioned on a quantized latent representation. Ghouse et al. [11] began by optimizing an autoencoder network using a rate-distortion loss. Subsequently, they train a conditional diffusion model on the output of the decoder, with the goal of enhancing its perceptual quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 407, + 178, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 178, + 421 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 178, + 421 + ], + "type": "text", + "content": "2.3. Neural Entropy Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 629 + ], + "type": "text", + "content": "The main purpose of the entropy model is to estimate the joint probability distribution over the quantized latent representation. When the learned entropy model precisely matches the true distribution of the latent representation, a lower bit-rate is required to generate a compressed file. Entropy estimation can take advantage of two key principles: backward and forward adaptation. Forward adaptation uses an extra latent variable, called hyper-prior [5], as a side information to capture spatial dependencies between elements of the latent representation. Backward adaptation, on the other hand, employs an autoregressive-based context model which leverages the previously decoded elements of the latent to predict the probability of the current element. The context model of backward methods can extract relationships between the current symbol and previously decoded symbols in various dimensions, including spatial and channel axes [13, 14, 21, 23, 29, 30, 32, 33, 47]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Minnen et al.[30] incorporated a local spatial context with a hyper-prior network to precisely predict distribution of the latent. This type of context model can exploit local correlations between the current latent and its neighboring causal elements using masked convolutions, which consequently necessitates the adoption of serial decoding. To expedite the decoding process, He et al.[13] evenly parti" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "tioned the latent into two groups: anchors and non-anchors. The anchors are exclusively decoded using the hyper-prior, while the non-anchors utilize both the hyper-prior and the local context model. This approach allows for the parallel decoding of both anchors and non-anchors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 263 + ], + "type": "text", + "content": "An alternative strategy for parallelizing the decoding process is to extract inter-channel dependencies. In [29] the latent code is divided into evenly sized chunks along the channel dimension, with each of these chunks being decoded using information from previously decoded ones. To speed up the channel-wise context extraction, He et al.[14] partitioned channels of the latent into an uneven set of groups, allocating fewer channels to the initial groups and more channels to the subsequent ones. This uneven division is motivated by the observation that the earlier channels possess higher entropy compared to the remaining channels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 545, + 384 + ], + "type": "text", + "content": "Some works introduced adaptive context models which focus on capturing long-range spatial correlations. Qian et al.[32] devised a global reference context model which assesses throughout previously decoded elements to detect the most similar one to the current latent. The authors in [33] employed a Transformer-based entropy model, called Entroformer, for capturing long-range contexts. The Entroformer benefits from diamond-relative position encoding and a top-k self-attention mechanism, both of which contribute to providing an efficient receptive field." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 395, + 367, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 395, + 367, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 395, + 367, + 407 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 511 + ], + "type": "text", + "content": "In this work, we develop our decoder based on the blurring diffusion model [20], which is conditioned on the discrete latent representation and enables the differentiation among the frequency components of an image throughout the diffusion and denoising processes. The architecture of decoder will be explained in Appendix. Moreover, a novel entropy model is employed that efficiently encodes the quantized latent representation into a binary stream." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 518, + 448, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 448, + 532 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 448, + 532 + ], + "type": "text", + "content": "3.1. Blurring Diffusion Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "content": "We can define the blurring diffusion model as a traditional diffusion model in the frequency space, but with distinct schedules for each dimension of data [20]. The specifics of the schedules for vectors " + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "content": " are crucial in this type of model, which will be discussed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 656 + ], + "type": "text", + "content": "Diffusion Process: The diffusion process [17] progressively degrades the clean image, transforming it into nearly pure Gaussian noise over the course of T time steps. Each step of the diffusion process in frequency space can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 666, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 666, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 358, + 666, + 545, + 681 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {f} _ {t} \\mid \\boldsymbol {f} _ {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t}; \\alpha_ {t} \\boldsymbol {f} _ {\\boldsymbol {x}}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}", + "image_path": "c8804a0f49ad0aa0c28a8a98e8f9bfa9b9e48855fdcd48dc0448cafed3692932.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " are defined as " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^T\\mathbf{z}_t" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^T\\mathbf{x}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": ", respectively. In such a model, each frequency component could" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3047" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "undergo diffusion at varying rates, with the modulation of these rates being determined by the values present in the vectors " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 144 + ], + "type": "text", + "content": "Denoising Process: The true denoising process in the frequency space, which is equal to the deblurring operation in the time domain, can be expressed as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 149, + 287, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 149, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 58, + 149, + 287, + 163 + ], + "type": "interline_equation", + "content": "q \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\boldsymbol {f} _ {x}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1}, \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right), \\tag {2}", + "image_path": "8653aac7d6330968426678447224d3f0b6a1ba213d9e15142c8cdba7512c906e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\sigma_{t\\rightarrow t - 1} = \\sigma_{t|t - 1}\\sigma_{t - 1} / \\sigma_t" + }, + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "inline_equation", + "content": "\\mu_{t\\to t - 1} = (\\alpha_{t|t - 1}\\sigma_{t - 1}^2 /\\sigma_t^2)f_t + (\\alpha_{t - 1}\\sigma_{t|t - 1}^2 /\\sigma_t^2)f_x." + }, + { + "bbox": [ + 47, + 168, + 287, + 251 + ], + "type": "text", + "content": " The actual denoising process can be approximated using a learned denoising distribution to generate data. Thus, by considering the epsilon reparametrization of Eq. 2, we can derive the expression for the learned denoising process in frequency space:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 260, + 287, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 260, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 59, + 260, + 287, + 274 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {t \\rightarrow t - 1} \\left(\\hat {\\boldsymbol {f}} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {t}\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right)", + "image_path": "3abf0b5ef8cfddfd31bfe5d7b5d781a1df49fd45dbec940862c79169d6424c88.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 277, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 277, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 71, + 277, + 287, + 290 + ], + "type": "text", + "content": "Re-parametrization: " + }, + { + "bbox": [ + 71, + 277, + 287, + 290 + ], + "type": "inline_equation", + "content": "\\hat{f}_{\\pmb{x}} = (1 / \\alpha_{t})(f_{t} - \\sigma_{t}\\hat{f}_{\\epsilon})" + }, + { + "bbox": [ + 71, + 277, + 287, + 290 + ], + "type": "text", + "content": " (3)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "text", + "content": "By substituting the approximation of " + }, + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "inline_equation", + "content": "\\hat{f}_x" + }, + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "text", + "content": " into the true denoising distribution, the mean of the denoising network can be derived by " + }, + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "inline_equation", + "content": "(\\alpha_{t|t-1}\\sigma_{t-1}^2/\\sigma_t^2)\\pmb{f}_t + (\\alpha_{t-1}\\sigma_{t|t-1}^2/\\sigma_t^2\\alpha_t)(\\pmb{f}_t - \\sigma_t\\hat{f}_\\epsilon)" + }, + { + "bbox": [ + 47, + 297, + 287, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 360, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 287, + 383 + ], + "type": "text", + "content": "Optimization: According to [17], the loss function is simplified as below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "type": "interline_equation", + "content": "\\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\boldsymbol {f} _ {\\epsilon} - \\hat {\\boldsymbol {f}} _ {\\epsilon} \\| ^ {2} ] = \\boldsymbol {E} [ \\| \\boldsymbol {V} ^ {\\boldsymbol {T}} (\\epsilon - \\hat {\\epsilon}) \\| ^ {2} ] \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon} [ \\| \\epsilon - \\hat {\\epsilon} \\| ^ {2} ], \\tag {4}", + "image_path": "9ed16c060db3680d2fbe7702c4dc73bd25d79060a4cbac0b355fc5f1f08cdf52.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "z_t = V(\\alpha_t V^T x + \\sigma_t V^T \\epsilon)" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon} = \\phi_{\\theta}(z_t, t)" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": " signifies that the neural network " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\phi_{\\theta}" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": " as input and predicts " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": ". This formulation satisfies the requirement for neural networks to perform effectively in a standard pixel space. After prediction, it is sufficient to transition back and forth between frequency space and pixel space using the DCT matrix " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "V^T" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": " and its inverse " + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 413, + 287, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 503, + 249, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 249, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 249, + 516 + ], + "type": "text", + "content": "3.2. Schedules of Blurring Diffusion Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": "The schedules for the blurring diffusion model, denoted as " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ", are obtained by combining a standard Gaussian noise diffusion schedule (with scalar values " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ") along with a blurring schedule " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ". Each element of the vector " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " shares the same value, as identical noise is added to all frequency components. Therefore, it becomes adequate to present a schedule for a scalar value " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " [20]. The noise schedule is chosen based on a variance-preserving cosine [31], specifically " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\sigma_{t}^{2} = 1 - \\alpha_{t}^{2}" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\cos(t\\pi / 2T)" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 46, + 521, + 287, + 652 + ], + "type": "text", + "content": ". Following [35], the blurring schedule is then defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 658, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 658, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 107, + 658, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\sigma_ {B, t} = \\sigma_ {B, m a x} \\sin (t \\pi / 2 T) ^ {2}, \\tag {5}", + "image_path": "f49ad4c4f18c61e68c941e0216c19dee44b1e157d3b44ee8527be62f42b8bc2e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{B,max}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " represents a hyperparameter equal to the maximum level of blur applied to the image. This schedule, in turn, corresponds to the dissipation time through" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\tau_{t} = \\sigma_{B,t}^{2} / 2" + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": ". Based on the formulation discussed in the Appendix, the blurring schedule " + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": ", which is employed for " + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": ", is defined as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 390, + 118, + 545, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 118, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 390, + 118, + 545, + 130 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} _ {\\boldsymbol {t}} = \\exp (- \\lambda \\tau_ {t}), \\tag {6}", + "image_path": "077c600c9abf8c051b218e84f45ae57297bf6c3d6e2b61e16332f771dd1be0a0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": " represents the vector containing squared frequencies, and " + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "inline_equation", + "content": "\\tau_{t}" + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": " corresponds to the dissipation time. To achieve a more gradual amplification of high frequencies during the denoising process, the blurring schedule " + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "inline_equation", + "content": "d_{t}" + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": " is adjusted to " + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "inline_equation", + "content": "(1 - d_{min})\\exp (-\\lambda \\tau_{t}) + d_{min}" + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "inline_equation", + "content": "d_{min}" + }, + { + "bbox": [ + 305, + 140, + 545, + 223 + ], + "type": "text", + "content": " is set to 0.001. Finally, by combining the Gaussian noise schedule with the blurring schedule, resultant schedule is as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 378, + 233, + 545, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 233, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 378, + 233, + 545, + 245 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\alpha} _ {t} = \\alpha_ {t}. \\boldsymbol {d} _ {t}, \\boldsymbol {\\sigma} _ {t} = \\mathbf {1} \\sigma_ {t}. \\tag {7}", + "image_path": "5b325af2c1ca76aca972bf29e602a9348539fd0c2099a91427bc525d0325c48d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 254, + 528, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 254, + 528, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 254, + 528, + 266 + ], + "type": "text", + "content": "3.3. Blurring Diffusion Model for Compression" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "content": "The rate-distortion objective in end-to-end compression resembles the loss function of a " + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "content": "-VAE, where a hyperparameter " + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "content": " is utilized to balance the trade-off between the bit-rate " + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "inline_equation", + "content": "(R)" + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "content": " and distortion " + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 306, + 272, + 545, + 320 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 329, + 545, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 545, + 342 + ], + "type": "interline_equation", + "content": "\\mathrm {L} = D + \\lambda R = \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right]. \\tag {8}", + "image_path": "10b0b48e9e3aaf873d9b11ae297c0e2f9744eefae5d2a9f9b18a512fe2e56aca.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "text", + "content": "While many neural codecs commonly employ Gaussian or Laplacian decoders, we introduce a novel approach using a conditional blurring diffusion model as the decoder. This approach aims to yield new distortion metrics that deviate from those based on Mean Squared Error (MSE) or Mean Absolute Error (MAE). In this framework, our proposed neural codec leverages two distinct types of latent variables: a semantic latent variable " + }, + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "text", + "content": " and texture latent variables " + }, + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{1:T}" + }, + { + "bbox": [ + 306, + 351, + 545, + 530 + ], + "type": "text", + "content": ". The semantic latent variable captures and encodes the overall content and meaning of the image. In contrast, the texture latent variables are tailored to carry finer details and intricate patterns that might not be fully represented by the semantic variable alone. Notably, unlike the semantic latent variable, the texture latent variables are not compressed but synthesized during the decoding phase." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 531, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 545, + 555 + ], + "type": "text", + "content": "The forward and backward processes of the conditional blurring diffusion model can be expressed as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 573, + 492, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 573, + 492, + 588 + ], + "spans": [ + { + "bbox": [ + 309, + 573, + 492, + 588 + ], + "type": "interline_equation", + "content": "q (\\pmb {f} _ {t} | \\pmb {f} _ {t - 1}) = \\mathcal {N} (\\pmb {f} _ {t}; \\alpha_ {t | t - 1} \\pmb {f} _ {t - 1}, \\sigma_ {t | t - 1} ^ {2} I),", + "image_path": "43d04498f61e68ab1a235e30157c2b6029f8262b1905b1771c5c005049a73bb2.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 591, + 543, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 543, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 543, + 622 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}}, \\boldsymbol {f} _ {\\boldsymbol {1}: T} | \\tilde {\\boldsymbol {y}}\\right) = p \\left(\\boldsymbol {f} _ {\\boldsymbol {T}}\\right) p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {x}} \\mid \\boldsymbol {f} _ {\\boldsymbol {1}}, \\tilde {\\boldsymbol {y}}\\right) \\prod_ {t = 2} ^ {T} p _ {\\theta} \\left(\\boldsymbol {f} _ {\\boldsymbol {t} - \\boldsymbol {1}} \\mid \\boldsymbol {f} _ {\\boldsymbol {t}}, \\tilde {\\boldsymbol {y}}\\right),", + "image_path": "139efa87407db548768f40999ac3265a0f88e75ea8940786c03b776d58f63863.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 625, + 399, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 399, + 639 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 399, + 639 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {f} _ {T}) = \\mathcal {N} (\\boldsymbol {f} _ {T}; \\mathbf {0}, \\boldsymbol {I})", + "image_path": "3b487ba1fb1489de7ec48a63e508e6a6f05b04a2f5299dfe7fcb32143e03e7c6.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 641, + 545, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 641, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 641, + 545, + 666 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(\\boldsymbol {f} _ {t - 1} \\mid \\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}\\right) = \\mathcal {N} \\left(\\boldsymbol {f} _ {t - 1}; \\mu_ {\\theta} \\left(\\boldsymbol {f} _ {t}, \\tilde {\\boldsymbol {y}}, t\\right), \\sigma^ {2} _ {t \\rightarrow t - 1} \\boldsymbol {I}\\right) \\tag {9}", + "image_path": "465c53977d2d5a6517afe5f09fb9c03e17a2969dd51391d89ba25734be348aef.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{t} = V^{T}z_{t}" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{x} = V^{T}x = V^{T}z_{o}" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": ". As shown in Eq. 8, distortion is equivalent to the negative marginal likelihood of input data " + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "-\\log p_{\\boldsymbol{x}|\\tilde{\\boldsymbol{y}}}(\\boldsymbol{x}|\\tilde{\\boldsymbol{y}})" + }, + { + "bbox": [ + 306, + 677, + 545, + 714 + ], + "type": "text", + "content": ", and minimizing it" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3048" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": "is analogous to minimizing the negative marginal likelihood of the frequency representation of the data " + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "p_{\\pmb{f}_x|\\tilde{\\pmb{y}}}(\\pmb{f}_x|\\tilde{\\pmb{y}}) = \\int p(\\pmb{f}_x, \\pmb{f}_{1:T}|\\tilde{\\pmb{y}}) d\\pmb{f}_{1:T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": ". Since computing the marginal likelihood is intractable, we employ its ELBO with the specified diffusion and de-blurring distributions. Following this substitution, the rate-distortion objective, through the application of Jensen's inequality, can be formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 174, + 287, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 174, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 91, + 174, + 287, + 202 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} \\left[ - \\log p _ {\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}} (\\boldsymbol {x} | \\tilde {\\boldsymbol {y}}) - \\lambda \\log p _ {\\tilde {\\boldsymbol {y}}} (\\tilde {\\boldsymbol {y}}) \\right] \\\\ \\leq \\boldsymbol {E} _ {\\tilde {\\boldsymbol {y}}} [ - \\log_ {P} \\boldsymbol {x} | \\boldsymbol {y} (\\omega | \\boldsymbol {y}) - \\lambda \\log_ {P} \\boldsymbol {y} (\\boldsymbol {y}) ], \\tag {10} \\\\ \\end{array}", + "image_path": "92cd212a11515713d2e9ab951b5aaab99ea00cfaf754672486424648d22fe53b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 211, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 287, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 211, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{ELBO} = E_q[\\frac{p_\\theta(f_x,f_{1:T}|\\tilde{y})}{q(f_t|f_{t - 1})} ]" + }, + { + "bbox": [ + 47, + 211, + 287, + 239 + ], + "type": "text", + "content": ". As explained in Appendix, the optimizing of ELBO can be simplified to:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 247, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 247, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 85, + 247, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\operatorname {E L B O} \\approx \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} ], \\tag {11}", + "image_path": "11ff1af468b9b5c779e2aab506e6b0143f408f30c53ea7495085614e5d4f0f5c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "inline_equation", + "content": "z_{t} = V\\alpha_{t}V^{T}x + \\sigma_{t}\\epsilon" + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": ". Similar to [28], an LPIPS loss [46] is incorporated into the rate-distortion loss to enhance the perceptual quality of the reconstructed image. As the initial image at any time step can be decoded as a function of the texture latent variable " + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": ", the semantic latent variable " + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": ", and the time step " + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "inline_equation", + "content": "\\hat{x}_t = V(1 / \\alpha_t)(V^T z_t - \\sigma_t V^T \\hat{\\epsilon})" + }, + { + "bbox": [ + 47, + 269, + 289, + 354 + ], + "type": "text", + "content": ", the total loss becomes as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 361, + 287, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 361, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 92, + 361, + 287, + 391 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {E} _ {t, \\boldsymbol {x}, \\epsilon , \\tilde {\\boldsymbol {y}}} [ (1 - \\beta) \\| \\epsilon - \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\tilde {\\boldsymbol {y}}) \\| ^ {2} \\tag {12} \\\\ \\left. - \\lambda \\log p _ {\\tilde {\\mathbf {y}}} (\\tilde {\\mathbf {y}}) + \\beta d _ {\\mathrm {L P I P S}} (\\boldsymbol {x}, \\hat {\\boldsymbol {x}} _ {t}) \\right], \\\\ \\end{array}", + "image_path": "53c15008524a61ceeb73935dbc15ec2800400662371f25811d47a3b40d814526.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 398, + 287, + 422 + ], + "type": "text", + "content": " represent hyperparameters that control the trade-off between rate, distortion, and perception." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": "Decoding Process: After the training, we utilize entropy decoding on " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " through the entropy model which estimates the distribution " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "p_{\\hat{\\pmb{y}}}(\\hat{\\pmb{y}})" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": ". We then employ ancestral sampling to conditionally decode the image " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": ", which is equal to " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{z_0}" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": " and starts from pure Gaussian noise " + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{z}_T\\sim \\mathcal{N}(\\pmb {0},\\pmb {I})" + }, + { + "bbox": [ + 47, + 422, + 287, + 483 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 490, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 490, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 58, + 490, + 287, + 506 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t - 1} \\gets \\boldsymbol {V} (\\hat {\\boldsymbol {\\mu}} _ {t \\rightarrow t - 1} + \\boldsymbol {\\sigma} _ {t \\rightarrow t - 1} \\boldsymbol {V} ^ {T} \\phi_ {\\theta} (\\boldsymbol {z} _ {t}, t, \\hat {\\boldsymbol {y}})) \\quad (1 3)", + "image_path": "6d5e3eb721368217bc35fc28abc2fed9b47818234adbdda4c46dbe771991eaf9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_{t\\rightarrow t - 1} = \\frac{\\alpha_{t|t - 1}\\sigma_{t - 1}^2}{\\sigma_t^2} V^T z_t + \\frac{\\sigma_{t|t - 1}^2}{\\alpha_{t|t - 1}\\sigma_t^2} (V^T z_t - \\sigma_t V^T\\phi_\\theta (z_t,t,\\hat{y}))" + }, + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "text", + "content": " . The latent variables " + }, + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "inline_equation", + "content": "z_{1:T}" + }, + { + "bbox": [ + 47, + 512, + 287, + 557 + ], + "type": "text", + "content": " are not stored but produced during the decoding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 563, + 188, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 188, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 188, + 576 + ], + "type": "text", + "content": "3.4. Proposed Entropy Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": "The main goal of the proposed entropy model' context is to exploit both channel-wise and spatial correlations, while expediting the decoding process. Inspired by the ELIC [14], we adopt an uneven grouping of latent channels, allowing most low entropy channels to depend on high entropy channels. The latent representation " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 582, + 287, + 689 + ], + "type": "text", + "content": " channels, is clustered into five chunks along the channel dimension: 16, 16, 32, 64, and M - 128 channels, respectively. In this setup, each chunk depends on all its previous decoded chunks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "The Fig. 2(a) illustrates our proposed entropy model as applied in the " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "-th chunk. Within each chunk, we use spatial" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 542, + 217 + ], + "type": "image", + "image_path": "1dc4e0096d8a53b6a62039eb526558c3fb106b94c6036b388ac4daf4f273c896.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": "Figure 2. Diagram illustrating the application of the proposed entropy model for decoding the " + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": "-th chunk " + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}^{(j)}" + }, + { + "bbox": [ + 305, + 224, + 545, + 258 + ], + "type": "text", + "content": ". (b) Global Spatial Context Block. (c) An example of a checkerboard-shaped mask." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 281, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 388 + ], + "type": "text", + "content": "context in conjunction with channel context to model correlations along both the channel and spatial dimensions. To expedite the decoding step, we employ a parallel bidirectional spatial context model which is capable of capturing both local and global spatial relationships. So, the anchor part is decoded in parallel by using solely the hyperprior and channel context, while the decoding of the non-anchor part relies on the hyperprior, as well as both the spatial and channel contexts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 406, + 406, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 406, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 406, + 418 + ], + "type": "text", + "content": "3.4.1 Spatial Context" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 426, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 545, + 557 + ], + "type": "text", + "content": "The spatial context design captures both local and global spatial correlations within the latent representation " + }, + { + "bbox": [ + 305, + 426, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}" + }, + { + "bbox": [ + 305, + 426, + 545, + 557 + ], + "type": "text", + "content": ". Following the decoding of the anchor group, a checkerboard-shaped convolution is applied to this group, generating local context for all non-anchor elements in a parallel manner. Furthermore, the acquired local contexts of the non-anchor part are subsequently fed into a Transformer-based block to efficiently extract the global spatial context. This Transformer block takes advantage of positional encoding that is customized specifically for the compression task, along with a checkerboard-based attention mechanism." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 575, + 490, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 490, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 490, + 587 + ], + "type": "text", + "content": "3.4.2 Transformer-based Spatial Context" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 594, + 545, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 690 + ], + "type": "text", + "content": "We adopt the Swin Transformer [26] blocks, including a masked window-based multi-head self-attention (W-MSA) and a masked shifted-window-based multi-head self-attention (SW-MSA), as our global spatial context model which shown in Fig. 2.(b). This selection enables us to strike a balance between computational efficiency and modeling capacity. The checkerboard-based self-attention in W-MAS and SW-MSA can be expressed:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 319, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "A t t e n (\\boldsymbol {Q}, \\boldsymbol {K}, \\boldsymbol {V}) = \\operatorname {s o f t m a x} (\\boldsymbol {Q} \\boldsymbol {K} ^ {T} \\odot \\boldsymbol {M}) \\boldsymbol {V}, \\tag {14}", + "image_path": "5ba175f5db53f140f0f4516ffd6e871f71e47becd290fa72e0c95662aeb0d795.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3049" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 71, + 289, + 130 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 289, + 130 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 289, + 130 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 289, + 130 + ], + "type": "image", + "image_path": "8bc3daba084e255fffb60e34138a9a2450ed7c96c5029372656da8fb4e68d630.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 141, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 141, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 141, + 287, + 163 + ], + "type": "text", + "content": "Figure 3. The procedure for acquiring Laplacian relative position encoding for a window with a size of " + }, + { + "bbox": [ + 46, + 141, + 287, + 163 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 46, + 141, + 287, + 163 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "inline_equation", + "content": "Q, K, V \\in R^{N^2 \\times d}" + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "content": " are the query, key, and value matrices, respectively. " + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "inline_equation", + "content": "N^2" + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "content": " denotes the number of patches in a window and " + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "content": " is the dimension of query/key/value. " + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "inline_equation", + "content": "M \\in R^{N^2 \\times N^2}" + }, + { + "bbox": [ + 47, + 182, + 288, + 233 + ], + "type": "text", + "content": " represents a checkerboard-shaped mask." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 247, + 203, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 203, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 203, + 258 + ], + "type": "text", + "content": "3.4.3 Relative Positional Encoding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 266, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 408 + ], + "type": "text", + "content": "The authors in [13, 33] have shown that the bitrate is impacted by the distance of neighboring latents within the spatial context modeling of current latent. Based on this insight, we introduce a receptive field-aware self-attention mechanism that employs a learnable Laplacian-shaped positional encoding for calculating spatial context. In our proposed approach, each channel chunk possesses its own receptive field during self-attention computation, which dynamically adjusts in response to changes in entropy of chunks. In the " + }, + { + "bbox": [ + 46, + 266, + 287, + 408 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 266, + 287, + 408 + ], + "type": "text", + "content": "-th chunk, the checkerboard-based attention of W-MSA and SW-MSA is refined to receptive field-aware self-attention as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 417, + 293, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 293, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 293, + 441 + ], + "type": "interline_equation", + "content": "A t t e n \\left(\\boldsymbol {Q} _ {j}, \\boldsymbol {K} _ {j}, \\boldsymbol {V} _ {j}\\right) = \\text {s o f t m a x} \\left(\\boldsymbol {Q} _ {j} \\boldsymbol {k} _ {j} ^ {T} \\odot \\boldsymbol {M} + \\boldsymbol {P} _ {\\text {L a p}, j}\\right) \\boldsymbol {V} _ {j}, \\tag {15}", + "image_path": "802f2f0c816c222f5e3291d6527c6f1b02ac93e92f2294f816f629fdf3dc34bc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "P_{Lap,j}" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " represents the learnable Laplacian relative position encoding for chunk " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": ". To create the Laplacian relative position encoding for the " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": "-th chunk, which comprises spatial windows with a size of " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " (i.e. including " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "N^2" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " patches), three steps need to be taken. Firstly, We generate a 2D relative position table, where each coordinate of the relative position lies in the range " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "[-N + 1, N - 1]" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": ". Subsequently, we derive the relative position matrix for each patch. As shown in Fig. 3, the first patch's relative distance coordinate of " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " (relative distance with its position) is located at the top-left of the orange box, while the last patch's relative distance of " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "(0,0)" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " is positioned at the bottom-right of the green box. Afterward, each relative position matrix is flattened and stacked together. Finally, we apply the 2D Laplacian function to each element of the resulting matrix to generate a Laplacian relative position encoding " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "P_{Lap,j}" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "A_j" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\sigma_j" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": ", whose size is " + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "inline_equation", + "content": "R^{N^2 \\times N^2}" + }, + { + "bbox": [ + 46, + 442, + 287, + 659 + ], + "type": "text", + "content": ". The 2D Laplacian function is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 667, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 667, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 64, + 667, + 287, + 682 + ], + "type": "interline_equation", + "content": "f _ {L a p} (x, y) = A _ {j} ^ {2} \\exp \\left(\\left(- 1 / 2 \\sigma_ {j} ^ {2}\\right) \\left(| x | + | y |\\right)\\right), \\tag {16}", + "image_path": "80a85081bed7c6ab0283eb9d889a8b16a813a67b03c97cafe743aff5a8c59b24.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "x\\in \\{-N + 1,\\dots,N - 1\\}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "y\\in \\{-N + 1,\\dots,N - 1\\}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\sigma_{j}" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " are learnable parameters which are determined" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "for each chunk through optimization. The value of them is associated with the wideness of the effective receptive field." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 107, + 386, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 107, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 107, + 386, + 121 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 127, + 439, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 439, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 439, + 140 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": "To train our learned image compression network, we utilize a merged dataset that includes high-resolution images from the DIV2K, Flickr2K, and CLIC [1] datasets. These images are randomly cropped to a size of " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": " during the training phase. The model parameters of all architectures were optimized using the Adam optimizer for a total of 2.4 million steps, with a batch size of 8. The initial learning rate was configured to be " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": " and was progressively reduced until the conclusion of training, reaching " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-7}" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": ". In order to cover a broad spectrum of bitrates, we selected a hyperparameter " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\{0.0004, 0.005, 0.01, 0.02, 0.04, 0.016\\}" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": ". The hyperparameter " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": " which specifies the contribution of perceptual loss is considered 0.9 for all models. The parameter " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": ", represents the required time steps for diffusion-based decoder, is consistently set to 500. To evaluate the performance of our compression approach, we select the Kodak dataset [2], which comprises 24 high-quality images with a resolution of " + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 304, + 145, + 545, + 384 + ], + "type": "text", + "content": ", and the CLIC2020 test set (428 images) with varying resolutions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "type": "text", + "content": "4.2. Comparison with the SOTA Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 662 + ], + "type": "text", + "content": "We compare our model with state-of-the-art generative based image compression networks, including both GAN-based and diffusion-based CODECs, as well as a state-of-the-art hand-crafted codec such as VVC-Intra (VTM) [3]. Additionally, we choose two VAE CODECs for comparison: DGML [9] and NSC [40]. Specifically, within the domain of GAN-based image compression approaches, we select HiFiC [28] and Multi-Realism [4] frameworks for comparison. Furthermore, we assess our method against diffusion-based models, namely DIRAC [11] and CDC [42]. This comparison is performed with respect to both rate-distortion and rate-perception tradeoff. The distortion is measured by the Peak Signal-to-Noise Ratio (PSNR) metric and perceptual quality are quantified via Frechet Inception Distance (FID) [16] and Learned Perceptual Image Patch Similarity (LPIPS). As shown in Fig. 4, our model shows superior performance compared to all other CODECs in terms of rate-perception. However, this heightened realism comes at the cost of distortion where other methods, except for the CDC model, either exhibit better performance or remain competitive." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 397, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 397, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 397, + 684 + ], + "type": "text", + "content": "4.3. Visual Quality" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Fig. 5 illustrates reconstructed images (kodim20.png, kodim07.png) generated by our proposed model, as well as" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3050" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 220, + 192 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 220, + 192 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 220, + 192 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 220, + 192 + ], + "type": "image", + "image_path": "b16130257453f3e6bd81f3c6bf6fb5783500bca1d0bbf3ee1e6c92c937eea3f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 221, + 71, + 369, + 193 + ], + "blocks": [ + { + "bbox": [ + 221, + 71, + 369, + 193 + ], + "lines": [ + { + "bbox": [ + 221, + 71, + 369, + 193 + ], + "spans": [ + { + "bbox": [ + 221, + 71, + 369, + 193 + ], + "type": "image", + "image_path": "1fc7343380d5f1c99e4da77b0fe572a6cfc938c3c8f28e81d984a416d71387e0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 522, + 194 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 522, + 194 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 522, + 194 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 522, + 194 + ], + "type": "image", + "image_path": "7978237301d02da3664285e7c4354a1fd70900e1911819e114cfa06631546f77.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 197, + 220, + 324 + ], + "blocks": [ + { + "bbox": [ + 69, + 197, + 220, + 324 + ], + "lines": [ + { + "bbox": [ + 69, + 197, + 220, + 324 + ], + "spans": [ + { + "bbox": [ + 69, + 197, + 220, + 324 + ], + "type": "image", + "image_path": "5327daa86c66bf863571b6ee4e7eb7e5d8ec87bd1955df156ac8255918736bf1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": "Figure 4. Comparison of our method with other codescs in terms of rate/distortion [bpp " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": " / PSNR " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": " ] and rate-perception, including [bpp " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": " / FID " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": " ] and [bpp " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": " / LPIPS " + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 46, + 335, + 547, + 357 + ], + "type": "text", + "content": "], for both the CLIC2020 test set and the Kodak dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 198, + 368, + 324 + ], + "blocks": [ + { + "bbox": [ + 221, + 198, + 368, + 324 + ], + "lines": [ + { + "bbox": [ + 221, + 198, + 368, + 324 + ], + "spans": [ + { + "bbox": [ + 221, + 198, + 368, + 324 + ], + "type": "image", + "image_path": "0253b50e4b9e5ee61a10785eaa8e5ff199ff6660421dbac8c917b81a4f8c5a96.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 369, + 198, + 520, + 324 + ], + "blocks": [ + { + "bbox": [ + 369, + 198, + 520, + 324 + ], + "lines": [ + { + "bbox": [ + 369, + 198, + 520, + 324 + ], + "spans": [ + { + "bbox": [ + 369, + 198, + 520, + 324 + ], + "type": "image", + "image_path": "4ab8a502bce3b38a86f10bd620f2c809d014fc5486ade6a3ace9009b94899bca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 48, + 371, + 545, + 575 + ], + "blocks": [ + { + "bbox": [ + 48, + 371, + 545, + 575 + ], + "lines": [ + { + "bbox": [ + 48, + 371, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 545, + 575 + ], + "type": "image", + "image_path": "f568de16a342910a3ca170440f73065a3194f08167543fb960aa67f4e4eb491c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 583, + 546, + 607 + ], + "lines": [ + { + "bbox": [ + 46, + 583, + 546, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 546, + 607 + ], + "type": "text", + "content": "Figure 5. Visual comparison of our method to HiFiC and CDC models shows that our model achieves superior reconstruction quality, particularly at lower bit-rates. In addition, our model displays fewer artifacts compared to both the HiFiC and CDC models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 626, + 288, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 288, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 288, + 710 + ], + "type": "text", + "content": "the CDC [42] and HiFiC [28] frameworks. Perceptually-oriented neural codecs, such as HiFiC and CDC, are criticized when their decoding process encounters semantic content, such as text within the compressed image. Our model effectively addresses this challenge, as depicted in Fig. 5. It enhances perceptual quality through a diffusion-based decoder and imposes a strict inductive bias via the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 626, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 626, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 305, + 626, + 545, + 675 + ], + "type": "text", + "content": "de-blurring diffusion process, resulting in both high perceptual quality and better text-preserving reconstructions. Additional qualitative comparisons are provided in the Appendix." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3051" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "text", + "content": "Maximum Blurring: To explore the impact of the blurring schedule, we vary the values of " + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\sigma_{B,max}" + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "text", + "content": " and compare the performance of the resulting models. It is evident that the model with " + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\sigma_{B,max} = 0" + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "text", + "content": " is equivalent to a standard denoising diffusion model. As indicated in Table 1a, blurring diffusion models with higher maximum blur levels " + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "inline_equation", + "content": "\\sigma_{B,max} = 25" + }, + { + "bbox": [ + 46, + 90, + 287, + 185 + ], + "type": "text", + "content": " generate higher-quality reconstructed images compared to other variant models in terms of FID score." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 235, + 115, + 293 + ], + "blocks": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "text", + "content": "Table 1. Ablation Study: all models are optimized with " + }, + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "inline_equation", + "content": "\\lambda = {0.01}" + }, + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "text", + "content": " . (a) Investigating the impact of maximum blurring " + }, + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "inline_equation", + "content": "{\\sigma }_{B,\\max }" + }, + { + "bbox": [ + 47, + 193, + 287, + 227 + ], + "type": "text", + "content": " . (b) Exploring the effects of different types of positional encoding" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 235, + 115, + 293 + ], + "lines": [ + { + "bbox": [ + 49, + 235, + 115, + 293 + ], + "spans": [ + { + "bbox": [ + 49, + 235, + 115, + 293 + ], + "type": "table", + "html": "
σB,maxFID
03.94
53.78
153.56
253.45
", + "image_path": "0ced95dbeffe01e700869746a390eeb2c4b58982c09a47244b1d64265cab499e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 148, + 236, + 272, + 293 + ], + "blocks": [ + { + "bbox": [ + 78, + 293, + 89, + 301 + ], + "lines": [ + { + "bbox": [ + 78, + 293, + 89, + 301 + ], + "spans": [ + { + "bbox": [ + 78, + 293, + 89, + 301 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 148, + 236, + 272, + 293 + ], + "lines": [ + { + "bbox": [ + 148, + 236, + 272, + 293 + ], + "spans": [ + { + "bbox": [ + 148, + 236, + 272, + 293 + ], + "type": "table", + "html": "
Position Enc.bpp
-0.2811
Relative Pos.0.2643
Diamond Relative Pos.0.2512
Laplacian-shaped Pos.0.2347
", + "image_path": "66799f3b750691748aabb58882e07f90a9e37565d4347b2997a4527f14454683.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 211, + 293, + 222, + 301 + ], + "lines": [ + { + "bbox": [ + 211, + 293, + 222, + 301 + ], + "spans": [ + { + "bbox": [ + 211, + 293, + 222, + 301 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "text", + "content": "Positional Encoding: We investigate the influence of various types of positional encoding in the Transformer-based entropy model. As shown in Table 1b, the 2D diamond relative positional encoding, which is implemented using a clip function, demonstrates better performance than the relative position encoding. However, adopting the Laplacian-shaped positional encoding results in even more significant bitrate savings compared to the 2D diamond relative positional encoding. As depicted in Fig. 6, distinct values of " + }, + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 318, + 287, + 544 + ], + "type": "text", + "content": " are obtained for each channel chunk. As we progress towards the final chunks, the receptive field becomes narrower. As observed in [14], the last chunk of channels, conditioned on all previous chunks, exhibits lower entropy. Their entropy can be estimated by considering only a small neighborhood, corresponding to a narrower receptive field in our proposed relative positioning. Conversely, to provide a reasonable estimate of entropy of the first channel chunks, a wider receptive field gathers more information over a large context to compared to mentioned last chunks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Analysis of Context Blocks: We conducted a comparison of the inference latency for entropy parameters during entropy decoding, as well as the bitrate savings of our proposed entropy model compared to other backward adaptation-based entropy models. For a fair comparison, all the models are equipped with a same encoder comprising of ResNet blocks and convolution layers to transform the input image " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in R^{H \\times C \\times 3}" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": " to a latent representation " + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y} \\in R^{H / 16 \\times W / 16 \\times 256}" + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": ". As reported in Tabel 2, our entropy model's speed is notably improved by employing unevenly grouped channels and implementing a bidirectional context model, as compared to the sequential spatial context modeling. Moreover, our findings clearly indicate that incorporating spatial global context results in superior per" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 353, + 122 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 353, + 122 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 353, + 122 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 353, + 122 + ], + "type": "image", + "image_path": "aaf6816fced1ece463fb61406af7d9768360c96f9bc9213330d21eabdd709ec8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 132, + 541, + 143 + ], + "lines": [ + { + "bbox": [ + 309, + 132, + 541, + 143 + ], + "spans": [ + { + "bbox": [ + 309, + 132, + 541, + 143 + ], + "type": "text", + "content": "Figure 6. Receptive fields are extracted for each channel chunk." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 356, + 70, + 403, + 122 + ], + "blocks": [ + { + "bbox": [ + 356, + 70, + 403, + 122 + ], + "lines": [ + { + "bbox": [ + 356, + 70, + 403, + 122 + ], + "spans": [ + { + "bbox": [ + 356, + 70, + 403, + 122 + ], + "type": "image", + "image_path": "abb886c94d26bcdbf75da767c9a7eaea83d0acec01962fbddac0fdcda68b6e0b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 405, + 70, + 452, + 122 + ], + "blocks": [ + { + "bbox": [ + 405, + 70, + 452, + 122 + ], + "lines": [ + { + "bbox": [ + 405, + 70, + 452, + 122 + ], + "spans": [ + { + "bbox": [ + 405, + 70, + 452, + 122 + ], + "type": "image", + "image_path": "fdf0e772579027d8120da043af53a415ef7dd84496b21091b32a752c72fb7b2f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 454, + 70, + 501, + 122 + ], + "blocks": [ + { + "bbox": [ + 454, + 70, + 501, + 122 + ], + "lines": [ + { + "bbox": [ + 454, + 70, + 501, + 122 + ], + "spans": [ + { + "bbox": [ + 454, + 70, + 501, + 122 + ], + "type": "image", + "image_path": "2fb3e32f6b9ebffe34f49c74d036f07cd2de45435ef2b613fcfafe1b0b548124.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 503, + 70, + 549, + 122 + ], + "blocks": [ + { + "bbox": [ + 503, + 70, + 549, + 122 + ], + "lines": [ + { + "bbox": [ + 503, + 70, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 503, + 70, + 549, + 122 + ], + "type": "image", + "image_path": "2c6685010111f4b5a2a57390062d0260ed436826ffa5ea886628f939a534b246.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 223, + 537, + 329 + ], + "blocks": [ + { + "bbox": [ + 305, + 158, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 305, + 158, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 158, + 547, + 213 + ], + "type": "text", + "content": "Table 2. The Bjøntegaard-delta-rate (BD-rate) [6] and inference latency of entropy parameter estimation during entropy decoding (Dec.) for different context-based models on the Kodak dataset using a GPU (RTX A6000). The BD-Rate is computed relative to VVC. ([P]:parallel, [S]:Serial)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 223, + 537, + 329 + ], + "lines": [ + { + "bbox": [ + 309, + 223, + 537, + 329 + ], + "spans": [ + { + "bbox": [ + 309, + 223, + 537, + 329 + ], + "type": "table", + "html": "
MethodContext ModelDec.(ms)BD-Rate
ChannelSpatial
Minnen et. al.[30]-Local[S]>103-3.24
Minnen et. al.[29]Even-67-3.96
He et. al.[13]-Local[P]28.2-2.91
Qian et. al.[33]-Global[S]>103-4.89
He et. al.[14]Uneven-37.2-3.12
He et. al.[14]UnevenLocal[P]156.9-5.71
Jiang et. al.[21]UnevenLocal[P]+Global[P]207.4-7.89
OursUnevenLocal[S]+Global[S]>103-9.38
OursUnevenLocal[P]+Global[P]196.3-8.25
VVC---0.00
", + "image_path": "79c2e4df442c1384f12b22a3bb7564c315f9411986f2f165cd531582889022ee.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 353, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 460 + ], + "type": "text", + "content": "formance when contrasted with utilizing only local context. The evaluation emphasizes that the integration of both global and local context enhances the precision of capturing spatial correlations, ultimately leading to a more accurate entropy model. In addition, our results verify that incorporating Laplacian-shaped positional encoding enhances the compression efficiency compared to MEM, which does not consider positional encoding in computing global spatial context." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 476, + 379, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 379, + 488 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 379, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": "We developed a neural image compression model which improves perceptual image quality using a non-isotropic diffusion decoder. This decoder's inductive bias effectively separates frequency components, leading to the creation of high-quality images. Moreover, we introduce an innovative entropy model that optimizes the trade-off between compression efficiency and decoding speed. This entropy model, founded on the Transformer architecture with Laplacian-shaped positional encoding, establishes a strong global spatial context. Our results underscore the efficacy of leveraging diffusion models and advanced entropy modeling to achieve outstanding image compression performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement: This research is based upon work supported by the National Aeronautics and Space Administration (NASA), via award number 80NSSC21M0322 under the title of Adaptive and Scalable Data Compression for Deep Space Data Transfer Applications using Deep Learning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3052" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] CLIC·challenge on learned image compression., Available at http://compression.cc/tasks/, 2022.6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Kodak image dataset., Available at https://r0k.us/ graphics/kodak/, 2022.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "type": "text", + "content": "[3] Versatile Video Coding Reference Software. Available at https://vcgi.hhi.fraunhofer.de/jvet/VVCSoftware_VTM, 2022.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 286, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 223 + ], + "type": "text", + "content": "[4] Eirikur Agustsson, David Minnen, George Toderici, and Fabian Mentzer. Multi-realism image compression with a conditional generator. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22324-22333, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 286, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 269 + ], + "type": "text", + "content": "[5] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. In International Conference on Learning Representations, 2018. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 291 + ], + "type": "text", + "content": "[6] Gisle Bjontegaard. Calculation of average psnr differences between rd-curves. ITU SG16 Doc. VCEG-M33, 2001. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "type": "text", + "content": "[7] Yochai Blau and Tomer Michaeli. Rethinking lossy compression: The rate-distortion-perception tradeoff. In International Conference on Machine Learning, pages 675-685. PMLR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "text", + "content": "[8] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9066-9075, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "text", + "content": "[9] Zhengxue Cheng, Heming Sun, Masaru Takeuchi, and Jiro Katto. Learned image compression with discretized gaussian mixture likelihoods and attention modules. In CVPR, pages 7939-7948, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 479 + ], + "type": "text", + "content": "[10] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2414-2423, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 286, + 523 + ], + "type": "text", + "content": "[11] Noor Fathima Ghouse, Jens Petersen, Auke Wiggers, Tianlin Xu, and Guillaume Sautiere. A residual diffusion model for high perceptual quality codec augmentation. arXiv preprint arXiv:2301.05489, 2023. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 286, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 546 + ], + "type": "text", + "content": "[12] Vivek K Goyal. Theoretical foundations of transform coding. IEEE Signal Processing Magazine, 18(5):9-21, 2001. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 286, + 601 + ], + "type": "text", + "content": "[13] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 603, + 286, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 668 + ], + "type": "text", + "content": "[14] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2, 3, 5, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[15] Dailan He, Ziming Yang, Hongjiu Yu, Tongda Xu, Jixiang Luo, Yuan Chen, Chenjian Gao, Xinjie Shi, Hongwei Qin, and Yan Wang. Po-elic: Perception-oriented efficient learned image coding. In Proceedings of the IEEE/CVF Conference" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "on Computer Vision and Pattern Recognition, pages 1764-1769, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 241 + ], + "type": "text", + "content": "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsanko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[19] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 287, + 545, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 309 + ], + "type": "text", + "content": "[20] Emiel Hoogeboom and Tim Salimans. Blurring diffusion models. arXiv preprint arXiv:2209.05557, 2022. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 311, + 545, + 364 + ], + "type": "text", + "content": "[21] Wei Jiang, Jiayu Yang, Yongqi Zhai, Peirong Ning, Feng Gao, and Ronggang Wang. Mlic: Multi-reference entropy model for learned image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7618-7627, 2023. 2, 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 399 + ], + "type": "text", + "content": "[22] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 545, + 466 + ], + "type": "text", + "content": "[23] A Burakhan Koyuncu, Han Gao, Atanas Boev, Georgii Gaikov, Elena Alshina, and Eckehard Steinbach. Contextformer: A transformer with spatio-channel attention for context modeling in learned image compression. In European Conference on Computer Vision, pages 447-463. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[24] Jooyoung Lee, Seunghyun Cho, and Seung-Kwon Beach. Context-adaptive entropy model for end-to-end optimized image compression. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 308, + 513, + 545, + 555 + ], + "type": "text", + "content": "[25] Jinming Liu, Heming Sun, and Jiro Katto. Learned image compression with mixed transformer-cnn architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14388-14397, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 612 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 668 + ], + "type": "text", + "content": "[27] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Fabian Mentzer, George D Toderici, Michael Tschannen, and Eirikur Agustsson. High-fidelity generative image compression. Advances in Neural Information Processing Systems, 33:11913-11924, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3053" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing (ICIP), pages 3339-3343. IEEE, 2020. 2, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 160 + ], + "type": "text", + "content": "[30] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 204 + ], + "type": "text", + "content": "[31] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[32] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Li Hao, and Rong Jin. Learning accurate entropy model with global reference for image compression. In International Conference on Learning Representations, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[33] Yichen Qian, Xiuyu Sun, Ming Lin, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. In International Conference on Learning Representations, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 347 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 287, + 381 + ], + "type": "text", + "content": "[35] Severi Rissanen, Markus Heinonen, and Arno Solin. Generative modelling with inverse heat dissipation. arXiv preprint arXiv:2206.13397, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 436 + ], + "type": "text", + "content": "[36] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 536 + ], + "type": "text", + "content": "[38] Lucas Theis, Wenzhe Shi, Andrew Cunningham, and Ferenc Huszár. Lossy image compression with compressive autoencoders. In International Conference on Learning Representations, 2016. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 569 + ], + "type": "text", + "content": "[39] Lucas Theis, Tim Salimans, Matthew D Hoffman, and Fabian Mentzer. Lossy compression with gaussian diffusion. arXiv preprint arXiv:2206.08889, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[40] Dezhao Wang, Wenhan Yang, Yueyu Hu, and Jiaying Liu. Neural data-dependent transform for learned image compression. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 17379-17388, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 679 + ], + "type": "text", + "content": "[41] Jay Whang, Maurizio Delbracio, Hossein Talebi, Chitwan Sahara, Alexandros G Dimakis, and Peyman Milanfar. Deblurring via stochastic refinement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16293-16303, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[42] Ruihan Yang and Stephan Mandt. Lossy image compression with conditional diffusion models. arXiv preprint arXiv:2209.06950, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 285 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[43] Ren Yang, Luc Van Gool, and Radu Timofte. Perceptual learned video compression with recurrent conditional gan. arXiv preprint arXiv:2109.03082, 1, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[44] Yibo Yang, Stephan Mandt, Lucas Theis, et al. An introduction to neural data compression. Foundations and Trends® in Computer Graphics and Vision, 15(2):113-200, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 205 + ], + "type": "text", + "content": "[45] Ali Zafari, Atefeh Khoshkhahtinat, Piyush Mehta, Mohammad Saeed Ebrahimi Saadabadi, Mohammad Akyash, and Nasser M Nasrabadi. Frequency disentangled features in neural image compression. In 2023 IEEE International Conference on Image Processing (ICIP), pages 2815-2819. IEEE, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 262 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 263, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 285 + ], + "type": "text", + "content": "[47] Yinhao Zhu, Yang Yang, and Taco Cohen. Transformer-based transform coding. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3054" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_content_list.json b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f53809f1393cb939b61f02c4aa800429264e8068 --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "Large Language Models are Good Prompt Learners for Low-Shot Image Classification", + "text_level": 1, + "bbox": [ + 222, + 130, + 750, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoheng Zheng Jingmin Wei Xuefeng Hu Haidong Zhu Ram Nevatia Viterbi School of Engineering University of Southern California", + "bbox": [ + 165, + 203, + 803, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhaoheng.zheng, jingminw, xuefengh, haidongz, nevatia}@usc.edu", + "bbox": [ + 204, + 258, + 759, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 309, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-shot image classification, where training images are limited or inaccessible, has benefited from recent progress on pre-trained vision-language (VL) models with strong generalizability, e.g. CLIP. Prompt learning methods built with VL models generate text features from the class names that only have confined class-specific information. Large Language Models (LLMs), with their vast encyclopedic knowledge, emerge as the complement. Thus, in this paper, we discuss the integration of LLMs to enhance pre-trained VL models, specifically on low-shot classification. However, the domain gap between language and vision blocks the direct application of LLMs. Thus, we propose LLaMP, Large Language Models as Prompt learners, that produces adaptive prompts for the CLIP text encoder, establishing it as the connecting bridge. Experiments show that, compared with other state-of-the-art prompt learning methods, LLaMP yields better performance on both zero-shot generalization and few-shot image classification, over a spectrum of 11 datasets. Code will be made available at: https://github.com/zhaohengz/LLaMP.", + "bbox": [ + 75, + 342, + 473, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-shot image classification tasks, including few-shot and zero-shot variants, are to learn from a set of class names along with a limited or null set of images. Such capacities are crucial for the extension and generalization of vision systems. Vision-Language (VL) models trained on large-scale web data, such as CLIP [32] and ALIGN [14], provide a new paradigm due to their generalization capabilities that includes zero-shot classification, and have been used in recent work [17-19, 23, 24, 48, 49]. Due to the scarcity of images for training, methods built for both tasks rely heavily on merely category names as the source of class-specific knowledge, resulting in a shortage of distinguishable descriptions. Meanwhile, Large Language Models (LLMs),", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13ba31720a95e4f52c8efd8e4be9287dc1acb43ea3910e2996b90dc8748222a6.jpg", + "image_caption": [ + "(a) Visual Descriptions from an LLM." + ], + "image_footnote": [], + "bbox": [ + 506, + 309, + 888, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44ef9588350c8f3b111c503448edbde37cb3f6386a248e4c5a54c4b4c7fc6ba8.jpg", + "image_caption": [ + "(b) LLMs' Knowledge Boosts the Performance.", + "Figure 1. Demonstration of LLaMP: (a) LLMs can provide visual descriptions for fine-grained object categories; (b) Zero-shot baseto-novel generalization benefits from the LLM knowledge." + ], + "image_footnote": [], + "bbox": [ + 506, + 421, + 887, + 549 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "e.g. GPT-4 [28] and LLaMA [38, 39], have demonstrated their encyclopedic knowledge and thus can provide linguistic visual descriptions for objects. Here, we investigate how to leverage LLMs for low-shot image classification.", + "bbox": [ + 496, + 640, + 893, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of prompt learning has provided an efficient way to adapt large pre-trained models. Previous work has explored various strategies to prompt vision-language (VL) models, including vision-conditioned text prompt learning [48], joint VL prompt learning [18] and self-regulated VL prompts[19]. On the text side, regardless of the learning strategy, learned prompt vectors are shared across all categories. The only difference among text inputs is the class name. In low-shot scenarios where visual data is limited, the extraction of class-specific knowledge from textual inputs becomes essential. However, the current paradigm, which relies on the CLIP text encoder to distinguish between class names, faces challenges, partic", + "bbox": [ + 496, + 703, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28453", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ularly with fine-grained target categories. For example, in FGVCAircraft [25], the class name \"Yak-40\", can barely provide any information for recognizing the object.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models, trained with large text corpora, are good candidates to serve as the complement. As in Fig. 1a, being queried about \"Yak-40\", the LLM generates a sentence detailing the visual appearance of the Yak-40 that can be further parsed into noun phrases and integrated into text prompts, providing richer information, compared with the ordinary prompt. We also show in Fig. 1b that by simply incorporating noun phrases extracted from a LLM's response, the performance of the ordinary CLIP models is improved by more than $1\\%$ without any training. Although recent prompt-learning based methods have shown notable improvements, it is non-trivial to apply them on textual visual descriptions generated by LLMs. Thus, instead of directly taking LLM generations as the textual input, we aim at producing class-specific representations by adapting LLMs to low-shot image classification.", + "bbox": [ + 75, + 138, + 468, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One challenge of the adaption is the domain gap between vision and language. When trained exclusively with textual corpora, the latent feature space of a LLM significantly diverges from that of its visual counterpart. Even worse, the data scarcity under the low-shot scenario make it virtually impossible to align two spaces through plain contrastive loss. We argue that, the CLIP text encoder, which is trained to project features from the language domain into the joint VL domain, can serve as the bridge. Thus, we propose the LLaMP framework, Large Language Models as Prompt learners, which leverages LLMs to learn informative prompts for CLIP models. In LLaMP, we treat LLMs as the prompt learner of the CLIP text encoder. More specifically, for each object category, LLaMP extracts corresponding knowledge from the LLM and yields class-specific prompt vectors, which are further combined with class-agnostic prompt embeddings (as in previous approaches), and encoded by the CLIP text encoder. We design an efficient tuning pipeline to avoid fully fine-tuning the language model while performing effective adaptation.", + "bbox": [ + 75, + 382, + 468, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Following the protocol in [48, 49], we evaluate LLaMP with two typical scenarios: zero-shot base-to-novel generalization [49] and few-shot image classification. For each scenario, we run LLaMP with 11 datasets covering a spectrum of tasks. On average, LLaMP achieves a $1.3\\%$ boost on the harmonic mean against the state-of-the-art PSRC [19], and $9.6\\%$ over the ordinary CLIP [32], on base-to-novel generalization. We also observe an average improvement of $0.94\\%$ on 16-shot image classification.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our approach makes use of Large Language Models to improve performance in low-shot image classification scenarios. The main contributions are: i) To the best of our knowledge, we are the first to investigate how to use the encyclopedic knowledge inherent in Large Lan", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "guage Models (LLMs) to enhance low-shot image classification; ii) We design a framework, LLaMP, to effectively adapt LLMs for image classification, without training the entire language model, and achieve state-of-the-art in both few-shot and zero-shot settings; iii) We conduct extensive analysis investigating the effectiveness of each components of LLaMP, and discuss the optimal setup for LLM-aided image classification.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 224, + 640, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large Language Models (LLMs). Recent years have witnessed remarkable progress in scaling up the size and capabilities of LLMs. Zhang et al. [47] first introduced a suite of transformers pre-trained at scale, followed by PaLM [6]. ChatGPT/GPT-4 [27, 28] emerged as a milestone conversational model, demonstrated impressive conversational abilities as a generalist model. Vicuna [5] further advanced by learning from ChatGPT, while LLaMA [38] demonstrates that larger scale training yields stronger foundation models. The subsequent LLaMA-2 [39] and PaLM-2 [2] achieved further gains in scale, efficiency and reasoning. Most recently, Almazrouei et al. [1] released Falcon, a 400B model.", + "bbox": [ + 496, + 251, + 890, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Zero-Shot Learning (ZSL). ZSL stands in contrast to traditional fully-supervised paradigms. Instead of relying on direct visual training samples, it leverages side information that can be drawn from a multitude of non-visual domains, including attributes [22], word embeddings [36, 40], and descriptive texts [34]. Zhang et al. [46] designed an embedding model to bridge the gap between seen and unseen categories. Concurrently, studies like [4, 44, 50] have spotlighted that generative models can produce features for unseen categories. Moreover, Graph Convolution Networks (GCN) [20] has been explored in research such as [16, 40] for further generalization.", + "bbox": [ + 496, + 431, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Learning. With the progress in large-scale vision-language models, such as CLIP [32] and ALIGN [14], which reveal their capacity in zero-shot transferability, prompt learning has emerged as an efficient learning scheme, where learnable prompts are appended to the input to fine-tune models. For low-shot image classification, CoOp [49] and CoCoOp [48], which modeled context words as learnable vectors to automate prompt engineering, have shown significant improvements over regular CLIP. MaPLe [18] further employed a hierarchical multi-modal prompting strategy across transformer blocks for progressive feature modeling. Kan et al. [17] incorporated external knowledge by designing knowledge-aware prompts and adaptation head for better generalization. Lee et al. [23] used masked attention to prevent internal representation shift for better generalization. Khattak et al. [19] further improved prompt learning by guiding prompts to balance task-specific and task-agnostic knowledge via mutual agreement maximization and prompt ensemble.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28454", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 90, + 184, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 113, + 217, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to previous CLIP-based learning approaches, we consider the classification problem as an image-text matching problem. We denote the image encoder and the text encoder, in CLIP-like models, as $\\mathcal{F}$ and $\\mathcal{G}$ , parameterized by $\\theta_{\\mathcal{F}}$ and $\\theta_{\\mathcal{G}}$ , respectively. An input image $\\boldsymbol{x} \\in \\mathbb{R}^{C \\times H \\times W}$ is split into $M$ equal-sized patches which are converted into a sequence of embeddings $\\tilde{\\boldsymbol{x}} = \\{\\boldsymbol{e}_{cls}, \\boldsymbol{e}_1, \\boldsymbol{e}_2, \\dots, \\boldsymbol{e}_M\\}$ . The visual input sequence $\\tilde{\\boldsymbol{x}}$ is encoded by the image encoder, producing the image feature $\\boldsymbol{f} = \\mathcal{F}(\\tilde{\\boldsymbol{x}})$ . On the text side, the text label $y$ and the associated name is formatted as \"A photo of [STH]\" and tokenized into a sequence of tokens $\\tilde{\\boldsymbol{y}} = \\{\\boldsymbol{t}_{bos}, \\boldsymbol{t}_1, \\boldsymbol{t}_2, \\dots, \\boldsymbol{t}_L, \\boldsymbol{t}_{eos}\\}$ , where $L$ is the length of input tokens. The input sequence is then encoded into $\\boldsymbol{g} = \\mathcal{G}(\\tilde{\\boldsymbol{y}})$ . For image classification, target class labels $\\{1, 2, \\dots, C\\}$ are encoded into text features $g_i$ . The classification is done by picking the class that has the highest similarity with the vision feature: $\\hat{y} = \\mathrm{argmax}_i \\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}_i)$ , where $\\mathcal{C}$ is the softmax cosine-similarity function $\\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}) = \\frac{\\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g} / \\tau)}{\\sum_{i=1}^{C} \\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g}_j / \\tau)}$ with temperature $\\tau$ .", + "bbox": [ + 76, + 137, + 472, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Prompting Learning. Given the size of the CLIP model, fine-tuning the entire model becomes infeasible. As both image and text encoders are built with standard transformer architecture, prompt learning, which tunes the model by combining trainable prompts with hidden states has been applied on the text encoder [48, 49], the image encoder [15, 41, 42], or both [18, 19, 33]. Similar to [19, 33], we build our method following the vision-language prompting paradigm, with deep prompting [15, 19], which not only insert prompts to the input layer, but to later encoder layers.", + "bbox": [ + 75, + 431, + 468, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, for each transformer layer that takes prompts, we define $V$ learnable visual prompts $\\pmb{p}_v = \\{p_v^1, p_v^2, \\dots, p_v^V\\}$ and $T$ learnable language prompts $\\pmb{p}_t = \\{p_t^1, p_t^2, \\dots, p_t^T\\}$ . For the $i$ -th vision encoder layer, visual prompts $\\pmb{p}_v^i$ are appended to input embeddings: $\\tilde{\\pmb{x}}_p^i = \\{e_{cls}^i, e_1^i, e_2^i, \\dots, e_M^i, p_v^i\\}$ . The prompt-augmented vision feature, $\\pmb{f}_p = \\mathcal{F}(\\tilde{\\pmb{x}}_p)$ , is produced by jointly encoding prompts and the image. As the ViT [9] architecture in CLIP adopts the bi-directional attention mechanism, the placement of $\\pmb{p}_v$ has no effect on $\\pmb{f}_p$ . On the language side, prompts are concatenated with the input of the $i$ -th text encoder: $\\tilde{\\pmb{y}}_p^i = \\{\\pmb{t}_{bos}^i, \\pmb{p}_t^i, \\pmb{t}_1^i, \\pmb{t}_2^i, \\dots, \\pmb{t}_L^i, \\pmb{t}_{eos}^i\\}$ . $\\tilde{\\pmb{y}}_p$ is further processed by the text encoder, resulting in the prompt-augmented language feature $\\pmb{g}_p = \\mathcal{G}(\\tilde{\\pmb{y}}_p)$ . More specifically, prompts to the first layer $\\pmb{p}_t^1$ are initialized with the embeddings of \"A photo of a\".", + "bbox": [ + 76, + 597, + 472, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-Rank Adaptation [13] (LoRA). As a parameter-efficient tuning technique, LoRA is designed to adapt large transformer model without updating original model weights. The LoRA technique is, in particular, applied to", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "linear projection layers. More specifically, for a linear layer with weight $W_0 \\in \\mathbb{R}^{d \\times k}$ , LoRA creates $\\Delta W$ by learning two low rank matrices $B \\in \\mathbb{R}^{d \\times r}$ and $A \\in \\mathbb{R}^{r \\times k}$ :", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} = \\left(W _ {0} + \\Delta W\\right) \\boldsymbol {x} = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 147, + 890, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We adopt a hybrid tuning scheme on the vision encoder, which performs prompt learning on the first few layers and applies LoRA on the rest.", + "bbox": [ + 498, + 172, + 890, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adaptive Prompt Learning with LLMs", + "text_level": 1, + "bbox": [ + 498, + 227, + 834, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of prompt tuning is to find a set of optimal prompts $\\pmb{p} = \\{\\pmb{p}_v, \\pmb{p}_t\\}$ which maximizes the log likelihood of $P(x, y | \\theta_{\\mathcal{F}}, \\theta_{\\mathcal{G}})$ over target downstream distribution $(x, y) \\sim (X, Y)$ :", + "bbox": [ + 498, + 250, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} = \\underset {\\boldsymbol {p}} {\\operatorname {a r g m a x}} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim (\\boldsymbol {X}, \\boldsymbol {Y})} \\log \\mathcal {C} \\left(\\mathcal {F} \\left(\\boldsymbol {x}; \\boldsymbol {p} _ {v}\\right), \\mathcal {G} \\left(\\boldsymbol {y}; \\boldsymbol {p} _ {t}\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 321, + 890, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, the $p$ optimized through Eqn. 2 has two issues. First, $p$ is shared for all categories for the downstream task, while the optimal prompt for each category might be different. Second, in low-shot scenarios, $p$ are usually empirically estimated from a limited training-set $X^{train}$ with limited categories $\\{1, 2, \\dots, C^{base}\\}$ , and therefore such $p$ can often be over-fitted to the small training-set $X^{train}$ and fail to generalize to novel categories outside $\\{1, 2, \\dots, C^{base}\\}$ .", + "bbox": [ + 496, + 359, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome these problems, we propose to learn a meta function on the language side $\\pmb{p}_t = \\Theta(y)$ which can adaptively estimate the optimal prompt for each category. An intuitive way to estimate proper prompts $\\pmb{p}$ for category name $y$ is to take advantage of the knowledge of the pre-trained Large Language Models (LLM) $\\mathcal{D}$ and extract discriminative descriptions of category $y$ . For example, given the input text $z: \\text{\"Describe {y}\"}$ ,", + "bbox": [ + 496, + 481, + 890, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} = \\left\\{p _ {1}, p _ {2}, \\dots , p _ {k} \\right\\} = \\mathcal {D} (\\boldsymbol {z}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 612, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while $p_i$ being sequentially generated by $\\mathcal{D}$ such that", + "bbox": [ + 498, + 638, + 849, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\mathcal {D} (\\boldsymbol {z}, t _ {1}, \\dots , t _ {i - 1}) = \\mathcal {D} ^ {(i)} (\\boldsymbol {z}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 662, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} = \\mathcal {M} (p _ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 686, + 671, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}^{(i)}$ is the $i$ -th forward iteration of $\\mathcal{D}$ , and $\\mathcal{M}$ maps continuous hidden states into discrete language tokens. To accelerate the process and to obtain $p$ in one pass, we approximate the above process with $K$ learnable prompts $\\pmb{p}_l = \\{\\theta_1,\\dots,\\theta_K\\}$ so that", + "bbox": [ + 496, + 710, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {t} = \\Theta (y) = \\mathcal {D} \\left(\\left\\{\\theta_ {1}, \\dots , \\theta_ {K} \\right\\} | z\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 797, + 890, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussion. While Large Language Models (LLMs) possess robust foundational knowledge within the linguistic domain, it is not feasible to directly substitute the text encoder of CLIP with an LLM. The reason lies in the inherent divergence between the LLM's latent space, which", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28455", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/283906f7dd2e1c395cde2299a97a11fd24d0a477ca2588799710fb4f0a87cefd.jpg", + "image_caption": [ + "Figure 2. An Overview of the LLaMP Framework: We first generate the knowledge cache by passing the query prompt through the LLM $\\mathcal{D}$ and use the knowledge cache to encode $\\pmb{p}_l$ , resulting in the adaptive prompts $\\tilde{\\pmb{h}}_l^i = W\\pmb{h}_l^i + b^i$ for the CLIP text encoder. $\\tilde{\\pmb{h}}_l$ is combined with regular learnable prompts of $\\mathcal{G}$ to generate the final text feature vector $\\pmb{g}_p$ . The image feature vector $\\pmb{f}_p$ is obtained through a hybrid-tuning strategy combining prompt learning and low-rank adaptation (LoRA)." + ], + "image_footnote": [], + "bbox": [ + 86, + 89, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is purely language-oriented, and the image-focused latent space of vision encoders. Attempting a direct alignment via contrastive learning would require an extensive dataset that is typically beyond the scope of low-shot learning. To bridge this gap, we introduce LLaMP—an adaptive prompt learning framework that leverages the LLM to craft class-specific prompt vectors, to reinforce the text encoder for low-shot image classification.", + "bbox": [ + 75, + 378, + 470, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. The LLaMP Framework", + "text_level": 1, + "bbox": [ + 76, + 507, + 303, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 shows an overview of the LLaMP framework. For convenience, we denote the decoder-only LLM as $\\mathcal{D}$ . The input to the decoder $\\mathcal{D}$ consists of two components: textual prompts $\\pmb{y}$ in the form of sentences, tokenized as $\\tilde{\\pmb{y}}$ , and learnable prompts $\\pmb{p}_l$ . We append prompt embeddings to the end of the input sequence and obtain the last hidden states of $\\mathcal{D}$ as the feature $\\pmb{h}_l$ :", + "bbox": [ + 75, + 530, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {h} _ {l} = \\mathcal {D} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {p} _ {l}) [ L + 1: L + K ], L = \\operatorname {L e n g t h} (\\tilde {\\boldsymbol {y}}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 645, + 468, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hidden states of $\\mathcal{D}$ are then mapped to the input space of the CLIP text encoder by the projection matrix $W\\in \\mathbb{R}^{d_1\\times d_2}$ , where $d_{1}$ and $d_{2}$ are respectively the hidden sizes of the LLM $\\mathcal{D}$ and the CLIP text encoder $\\mathcal{G}$ . A set of prompt-specific biases $b\\in \\mathbb{R}^{K\\times d_2}$ are added:", + "bbox": [ + 75, + 671, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {h}} _ {l} = W \\boldsymbol {h} _ {l} + b \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 755, + 468, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We combine $\\tilde{h}_l$ from LLM with regular learnable prompts, as in previous approaches [19], to construct the input for CLIP text encoder. Similar to deep prompting [15, 19], we create layer-specific prompts through different $W$ matrices and $b$ vectors. For the $i$ -th layer, we let $\\tilde{\\pmb{h}}_l^i = W^i\\pmb{h}_l + b^i$ and the entire sequence is constructed as", + "bbox": [ + 75, + 781, + 468, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {y}} _ {l} ^ {i} = \\left\\{\\boldsymbol {t} _ {b o s} ^ {i}, \\boldsymbol {p} _ {t} ^ {i}, \\boldsymbol {t} _ {1} ^ {i}, \\boldsymbol {t} _ {2} ^ {i}, \\dots , \\boldsymbol {t} _ {L} ^ {i}, \\tilde {\\boldsymbol {h}} _ {l} ^ {i}, \\boldsymbol {t} _ {e o s} ^ {i} \\right\\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LLM Knowledge Cache. A Large Language Model (LLM), as implied by its name, typically comprises billions of parameters. For example, the most compact LLaMA [38, 39] model has 7B parameters. Thus, even performing prompt learning on a LLM become impractical. The memory consumption to store gradients for back propagation can go beyond the limit of mainstream GPUs. Instead, the causal attention mechanism inherent in decoder-only LLMs, where the embedding of an input token only depends on the preceding tokens, facilitates a feasible workaround.", + "bbox": [ + 496, + 377, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previously mentioned, the prompt embeddings $\\pmb{p}_l$ are appended to the end of text tokens $\\tilde{\\pmb{y}}$ . According to the causal attention mechanism, $\\tilde{\\pmb{y}}$ is encoded independently of $\\pmb{p}_l$ . Thus, we design a two-stage process, where we create the LLM knowledge cache by passing $\\tilde{\\pmb{y}}$ through $\\mathcal{D}$ and leverage the cache to convert $\\pmb{p}_l$ into class-specific embeddings for the CLIP text encoder $\\mathcal{G}$ .", + "bbox": [ + 496, + 532, + 892, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To compute the attention of a token, the only dependency is the Key and Value vectors from the preceding tokens. Thus, we adopt the KV-cache [31, 43], a technique used in inference acceleration of LLMs, to create the knowledge cache. At the first stage, we pass text tokens $\\tilde{\\pmb{y}}$ through the language model $\\mathcal{D}$ and save the Keys and Values as the knowledge cache for the second stage. Once computed, the knowledge cache remains fixed throughout the entire training process and bears the information that is needed for further computation. Thus, in LLaMP, we leverage the knowledge cache obtained at the first stage to generate class-specific prompt embeddings.", + "bbox": [ + 496, + 641, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the second stage, we create class-specific prompt embeddings from the pre-computed knowledge cache. As $\\pmb{p}_l$ is not initialized in the natural language domain, it need not pass through the entire LLM; instead, we insert those prompts $\\pmb{p}_l$ to the last layer of the LLM $\\mathcal{D}_N$ . It is achieved", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28456", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by encoding them alongside the cache from $\\tilde{\\pmb{y}}$ , as in", + "bbox": [ + 76, + 90, + 421, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {H} _ {l} = \\mathcal {D} _ {N} \\left(\\boldsymbol {K} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {V} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {p} _ {l}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 118, + 468, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{\\tilde{y}}, V_{\\tilde{y}}$ represent the knowledge cache. This design enables LLaMP to efficiently learn informative prompt embeddings for the CLIP encoder $\\mathcal{G}$ . It accomplishes this by incurring modest training costs, compared with training the entire LLM. Simultaneously, it maintains the essential knowledge inherent in the LLM decoder $\\mathcal{D}$ .", + "bbox": [ + 75, + 145, + 468, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Targets of LLaMP. Although the training strategy in Eqn. 9 has reduced the number of learnable parameters, a full decoder inside a LLM still consists of an enormous number of parameters. For example, one layer in LLaMA-7B bears 200M parameters, making training of the entire layer costly. As the goal is to leverage the knowledge from LLM, altering a full layer can lead to the loss of knowledge. Thus, as shown in Fig. 2, a typical decoder layer has two major components: the self attention module, consisting Query, Key, Value and Output projection layers, and the Feed-Forward Network (FFN). LLaMP targets the Query and Output projection layer inside the self-attention module. By updating the Query layer, LLM prompts $p_l$ are learned to distill pertinent information from the knowledge cache and the Output layer projects it to the latent space. We keep the Key and Value layers frozen to ensure the alignment between $p_l$ and knowledge cache. We leave the FFN unchanged to preserve the knowledge. Further discussions regarding these choices are made in Sec. 4.3.", + "bbox": [ + 75, + 236, + 470, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Priors from Pre-Generated Responses. We extend the initial prompt, \"In one sentence, describe the distinctive appearance of [STH]\", by incorporating the response generated by the language model into the input sequence. This approach enriches the base content: the generated text provides a clear and explicit description of the object's appearance, acting as a valuable informative prior for language model adaptation. However, it is common for responses from an LLM to include filler words like \"sure\" for sentence structure coherence. To refine the input, we parse the noun phrases from the LLM's response through spaCy [12], an NLP engine, and merge them with the initial prompt, forming a more focused and informative language prior.", + "bbox": [ + 75, + 523, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Augmentations. Following the insights of Khattak et al. [19], which highlight the performance benefits of diverse textual inputs, we aim to further augment the text inputs used in the CLIP text encoder. Our approach, building upon the methods in [19, 48], incorporates handcrafted templates and expands their diversity through a two-step process: i) We introduce noun phrases into the existing templates for CLIP, for example, transforming \"A photo of [STH]\" to \"A photo of [STH] with [NP]\", thereby enriching the descriptive content; ii) We create a variety of new prompt templates for the LLM", + "bbox": [ + 75, + 734, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similar to “In one sentence, describe the distinctive appearance of [STH]” through GPT-4 [28], to further diversify the text input.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 498, + 143, + 714, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to PSRC [19], our objective function consists of three components: The main cross-entropy loss $\\mathcal{L}_{CE}$ , feature-level L1 regularization $\\mathcal{L}_{l1}$ , and soft distillation loss $\\mathcal{L}_{dist}$ . Given $\\mathcal{C}$ training categories and $\\mathcal{N}$ training samples, $\\mathcal{L}_{CE}$ is defined as", + "bbox": [ + 496, + 167, + 893, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} = - \\frac {1}{\\mathcal {N}} \\sum_ {i} \\log \\frac {\\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} / \\tau\\right)}{\\sum \\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {j} / \\tau\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 251, + 890, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The L1 regularization is computed between learned features $\\pmb{f}_p, \\pmb{g}_p$ and pre-trained CLIP features $\\hat{\\pmb{f}}, \\hat{\\pmb{g}}$ :", + "bbox": [ + 498, + 297, + 890, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l 1} = \\frac {1}{\\mathcal {N}} \\sum_ {i} \\lambda_ {v} \\left| \\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {f}} ^ {i} \\right| + \\frac {1}{C} \\sum_ {i} \\lambda_ {t} \\left| \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {g}} ^ {i} \\right|, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 337, + 890, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_v$ and $\\lambda_t$ are coefficients. The prediction of LLaMP is further bound by the KL-Divergence between predicted distributions of LLaMP and vanilla CLIP:", + "bbox": [ + 496, + 380, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\lambda_ {\\text {d i s t}} D _ {K L} \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}}, \\hat {\\boldsymbol {f}} \\cdot \\hat {\\boldsymbol {g}}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 435, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We sum all three losses up as the final objective function: $\\mathcal{L} = \\mathcal{L}_{CE} + \\mathcal{L}_{l1} + \\mathcal{L}_{dist}$ .", + "bbox": [ + 498, + 462, + 890, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we randomly sample one LLM template as the input of LLaMP for each batch. For inference, we compute the probability distribution predicted from each input template and average them.", + "bbox": [ + 496, + 492, + 890, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 565, + 633, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 500, + 590, + 674, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Similar to previous work [18, 19, 48], in our study, we evaluate LLaMP performance over a spectrum of classification tasks with 11 datasets, including ImageNet [8] and Caltech101 [10] for generic image classification, OxfordPets [29], StanfordCars [21], Flowers102 [26], Food101 [3], and FGVCAircraft [25] for fine-grained classification, SUN397 [45] for scene recognition, UCF101 [37] for action recognition, DTD [7] for texture classification, and EuroSAT [11] for satellite image recognition.", + "bbox": [ + 496, + 613, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scenarios & Metrics. We evaluate LLaMP on two typical low-shot scenarios: zero-shot base-to-novel generalization and few-shot image classification. In zero-shot base-to-novel generalization, the base classes are seen during training, while the novel classes are unseen. We measure models performance through accuracies of base and novel classes, and the harmonic mean of the two. For few-shot classification, we assess the accuracy with 16 shots per class.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We build LLaMP through the PyTorch [30] framework. All models are trained with", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28457", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e31197af8adaa8d8b7eae93e47dfe3fd81a52d8a78cffd09ae0dc33bb1edb96a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAverageImageNet [8]Caltech101 [10]OxfordPets [29]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3474.2271.7072.4368.1470.2296.8494.0095.4091.1797.2694.12
CoOp [49]82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp [48]80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
KAPT* [17]78.4170.5274.2671.1065.2068.0297.1093.5395.2893.1396.5394.80
ProDA [24]81.5672.3076.6576.6670.5473.4797.7494.3696.0295.4397.7696.58
MaPLe [18]82.2875.1478.5575.4070.3272.7298.2793.2395.6895.4397.8396.62
RPO [23]81.1375.0077.7876.6071.5774.0097.9794.3796.0394.6397.5096.05
PSRC [19]84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
LLaMP85.1677.7181.2777.9971.2774.4898.4595.8597.1396.3197.7497.02
Δ w.r.t. PSRC+0.90+1.61+1.30+0.39+0.54+0.47+0.35+1.82+1.11+0.98+0.44+0.72
", + "bbox": [ + 106, + 88, + 866, + 282 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72616f794bb299e77101e81471b3aa570be7d7de907b66acd6dd735a05110dc0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStanfordCars [21]Flowers102[26]Food101 [3]FGVCAircraft [25]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]63.3774.8968.6572.0877.8074.8390.1091.2290.6627.1936.2931.09
CoOp [49]78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp [48]70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
KAPT* [17]69.4766.2067.7995.0071.2081.4086.1387.0686.5929.6728.7329.19
ProDA [24]72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
MaPLe [18]74.7071.2072.9197.7068.6880.6690.3088.5789.4336.9034.1335.46
RPO [23]73.8775.5374.6994.1376.6784.5090.3390.8390.5837.3334.2035.70
PSRC [19]78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
LLaMP81.5674.5477.8997.8277.4086.4291.0591.9391.4947.3037.6141.90
Δ w.r.t. PSRC+3.29-0.43+1.31-0.25+0.90+0.47+0.38+0.40+0.39+4.57-0.26+1.75
", + "bbox": [ + 106, + 287, + 864, + 483 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/49bb5878a17491d4a0b157d7e8880559049eb07afca1468b5b03940175ab0b6d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3675.3572.2353.2459.9056.3756.4864.0560.0370.5377.5073.85
CoOp [49]80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp [48]79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
KAPT* [17]79.4074.3376.7875.9758.3065.9784.8067.5775.2180.8367.1073.33
ProDA [24]80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
MaPLe [18]78.4776.9377.7980.6756.4866.4483.9066.0073.8885.2371.9778.04
RPO [23]80.6077.8079.1876.7062.1368.6186.6368.9776.7983.6775.4379.34
PSRC [19]82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
LLaMP83.4179.9081.6283.4964.4972.7791.9383.6687.6087.1380.6683.77
Δ w.r.t. PSRC+0.74+1.43+1.10+0.12+1.52+1.02-0.97+9.76+5.28+0.03+1.86+1.03
", + "bbox": [ + 106, + 487, + 864, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with state-of-the-art methods on base-to-novel generalization. LLaMP shows strong generalization results over previous approaches on 11 image classification tasks. Absolute gains over PSRC are indicated in blue. *KAPT is trained with ViT-B/32 image encoder instead of ViT-B/16.", + "bbox": [ + 75, + 695, + 893, + 737 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2 NVIDIA A100 40GB GPUs. For LLaMP, we adopt LLaMA2-7B [39] as the language model $\\mathcal{D}$ , and ViT-B/16 [9] as the image encoder, following [18, 19, 48, 49]. On the text side, we set prompt learning depth to 9. To tune the vision encoder, we adopt the hybrid tuning scheme which performs deep prompt learning on the first 6 layers and LoRA on the rest. Similar to [13], LoRA is applied to the Query and Value projection layers inside attention modules. The number of $p_l$ prompts, $K$ , is set to 16. We set a global learning rate of 2E-4 with a batch size of 8. The learning", + "bbox": [ + 75, + 744, + 472, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rate of LoRA modules is set to 2E-5. $\\lambda_{t},\\lambda_{v}$ and $\\lambda_{dist}$ are set to 25, 10 and 2.5, respectively.", + "bbox": [ + 496, + 746, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 498, + 786, + 722, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-Shot Base-to-Novel Generalization. LLaMP outperforms existing state-of-the-art prompt learning methods on most metrics of 11 classification datasets in the baseto-novel generalization benchmark. As shown in Tab. 1, compared to the latest model PSRC [19], LLaMP achieves average gains of $0.90\\%$ in base accuracy, $1.61\\%$ in novel", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28458", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/908bf87a600e47c599f21fe166edd5ef55e1b62e24f4faa081e57d22fd4b317b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
16-Shot Classification
AverageImageNet [8]Collect [10]Pets [29]Cars [21]Flowers [26]Food [3]Aircraft [25]SUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
CLIP [32]78.79 (65.02)67.3195.4385.3480.4497.3782.9045.3673.2869.9687.2182.11
CoOp [49]79.89 (73.82)71.8795.5791.8783.0797.0784.2043.4074.6769.8784.9382.23
CoCoOp [48]74.90 (70.70)70.8395.1693.3471.5787.8487.2531.2172.1563.0473.3278.14
MaPLe [18]81.79 (75.58)72.3396.0092.8383.5797.0085.3348.4075.5371.3392.3385.03
PSRC [19]82.87 (77.90)73.1796.0793.6783.8397.6087.5050.8377.2372.7392.4386.47
LLaMP83.81 (78.50)73.4997.0894.2186.0798.0687.6256.0777.0274.1791.3186.84
", + "bbox": [ + 130, + 88, + 841, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accuracy, and $1.30\\%$ in harmonic mean on average. Moreover, LLaMP consistently achieves higher harmonic means (HM) compared to other models. These improvements indicate that our approach better balances performance on base and novel data, thus achieving stronger generalization compared to the prior prompt learning techniques.", + "bbox": [ + 75, + 304, + 468, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In particular, LLaMP excels in fine-grained datasets requiring detailed analysis. On FGVCAircraft, LLaMP surpasses PSRC by $4.57\\%$ on base accuracy and $1.75\\%$ on HM, highlighting its strong understanding of detailed aircraft features. Furthermore, on EuroSAT, LLaMP achieves improvements of $9.76\\%$ and $5.28\\%$ on novel accuracy and HM, respectively. We also observe similar performance gains on StanfordCars, where LLaMP outperforms by $3.29\\%$ on base accuracy and $1.31\\%$ on HM. The information embedded in LLM enables LLaMP to capture and utilize the rich semantic information necessary for distinguishing between closely related categories.", + "bbox": [ + 75, + 396, + 468, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-Shot Classification. LLaMP also achieves improvements across these classification datasets in few-shot classification tasks. As in Tab.2, with an average classification accuracy of $83.81\\%$ . Notably, on FGVCAircraft and StanfordCars, LLaMP shows a significant improvement over PSRC, further demonstrating that the knowledge from language models benefits the recognition of fine-grained object categories, which aligns with our observation on zero-shot base-to-novel generalization. Moreover, on DTD, where MaPLe and PSRC achieve around $72\\%$ accuracy, LLaMP achieves a higher accuracy of $74.17\\%$ , underscoring its ability to recognize textures.", + "bbox": [ + 75, + 577, + 468, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 770, + 230, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is the knowledge from LLM helping? In Tab. 3, we show that the knowledge from LLM benefits in both ways: Without training, performance of ordinary CLIP model can be improved by introducing noun phrases; The LLaMP framework shows further improvement after training.", + "bbox": [ + 75, + 792, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noun phrases are parsed from the LLM's responses the prompt of \"Describe [STH]\". We then use the tem", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/644211bcd4075aa0c799a9dc35d2eebcec7fa7de671298366b78b4a10d58830c.jpg", + "table_caption": [ + "Table 2. Few shot classification results with 16 shots. Numbers in the bracket indicate the average performance over $1/2/4/8/16$ shots." + ], + "table_footnote": [], + "table_body": "
MethodLLMBaseNovelHM
CLIP✓69.3474.2271.70
70.9574.9372.79
LLaMP✓82.2176.4479.22
85.1677.7181.27
", + "bbox": [ + 545, + 301, + 846, + 401 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/82d95e9394564cdf9cb81b6389a2030505e2d9017286443817d2d88f119de9fb.jpg", + "table_caption": [ + "Table 3. Ablation study on the LLM Knowledge." + ], + "table_footnote": [], + "table_body": "
LPQOKVFFN%BaseNovelHM
✓.0385.0077.2980.96
✓✓✓3385.2077.4581.14
✓✓✓8385.0577.7381.22
✓✓✓✓10085.2377.5681.21
✓✓1785.1677.7181.27
", + "bbox": [ + 504, + 434, + 888, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the Training Strategy. “%” indicates the ratio of parameters trained compared to fully tuning a layer.", + "bbox": [ + 500, + 554, + 890, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plate, \"A photo of [STH] with [NP]\" to generate the NP-augmented text embedding for CLIP. We take the average of all augmented embeddings for classification. In Tab. 3 we show that even ordinary CLIP can benefit from incorporating LLMs' knowledge.", + "bbox": [ + 496, + 598, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, the comparison between LLaMP and LLaMP without the LLM indicates that merely integrating LoRA [13] to the vision encoder is not beneficial. The \"LLaMP without LLM\" is essentially an ordinary prompting learning model plus LoRAs in the vision encoder. We show that the improved vision encoding capacity only benefits when the quality of text embeddings s are enhanced by incorporating LLMs' knowledge through LLaMP.", + "bbox": [ + 496, + 672, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoder Training strategy. We categorize trainable parameters of $\\mathcal{D}_N$ into four groups: learnable prompts (LP), Query and Output projections (QO), Key and Value projections (KV), and the feed-forward network (FFN). Tab. 4 indicates LLaMP can achieve desirable results by just learning the prompts of $\\mathcal{D}$ . One step further, adding QO into optimization achieves the best performance. Although other", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28459", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a7aa7d3ddc9a8573ebaf8270c0603fcd5a1b885ade38e508e4c38e0f1e87ea2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPriorsBaseNovelHM
LLaMPX84.9077.5981.08
Plain85.2677.5681.22
NP85.1677.7181.27
", + "bbox": [ + 120, + 88, + 426, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ccb74f6db239837352b7451632af8b0544c7984b8810ce38464e46b458c7c8d8.jpg", + "table_caption": [ + "Table 5. Ablation Study on Pre-generated Text Priors. $\\times$ refers to \"without textual priors\" and NP stands for noun phrases." + ], + "table_footnote": [], + "table_body": "
MethodBaseNovelHM
LLM Only81.7435.8249.81
LLaMP85.1677.7181.27
", + "bbox": [ + 138, + 214, + 406, + 276 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ed65280b1c1c6a8f28208dd6804fd00a6bae8d8dd44772482f62188d12a96b76.jpg", + "table_caption": [ + "Table 6. The CLIP text encoder helps adaptation." + ], + "table_footnote": [], + "table_body": "
SchemeBaseNovelHM
Prompt × 984.6777.2880.81
LoRA × 1284.8977.2780.90
Prompt ×6 + LoRA ×685.1677.7181.27
", + "bbox": [ + 96, + 311, + 450, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Study on Vision Tuning Scheme. Our hybrid design achieves the best performance.", + "bbox": [ + 76, + 405, + 467, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "setups introduce much more trainable parameters, they can not surpass the \"LP + QO\" strategy.", + "bbox": [ + 76, + 444, + 467, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Textual Priors. We study the effect of pregenerated textual priors on LLaMP. We compare three different approaches: without textual priors, using plain responses as the prior, and LLaMP, which takes parsed noun phrases. Tab. 5 shows that LLaMP can achieve over $81\\%$ on HM without pre-generated priors, while adding parsed noun phrases as textual priors further pushes the HM to $81.27\\%$ .", + "bbox": [ + 76, + 474, + 467, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP as the bridge. One may wonder if it is possible to replace CLIP text encoder with a large language model. Here, we study two setups: i) LLM as encoder, which treats the output of the language model, $\\tilde{h}_l$ as the text feature; ii) LLaMP, which treat $\\tilde{h}_l$ as part of the text input prompt. Tab. 6 reveals that relying solely on the LLM results in poor accuracy for novel categories. This supports our hypothesis that aligning LLMs with vision encoders generally requires a more extensive dataset. Furthermore, LLaMP's design significantly improves the novel accuracy by $40\\%$ .", + "bbox": [ + 75, + 580, + 467, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Vision Training Strategy. As ViT-16/B has 12 transformer layers, we compare different vision training strategies within LLaMP in Tab. 7. Apart from the default hybrid scheme, we evaluate setups including prompt learning at first 9 layers (P9), a similar setup to PSRC [19], and LoRA [13] in all 12 layers. The results suggest that the scheme leverages the strengths of both prompt learning and LoRA, addressing potential bottlenecks in the vision encoder and enhancing overall performance in LLaMP.", + "bbox": [ + 75, + 733, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of LLM Prompts. We vary the number of LLM prompts and study their effects on LLaMP. As in", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5eb53d774d0e54e537515c705435c6df8314a142113bfec24752075d117f18bc.jpg", + "image_caption": [ + "Figure 3. Effect of LLM Prompts on Harmonic Mean. 16 prompts achieve the most balanced performance." + ], + "image_footnote": [], + "bbox": [ + 514, + 95, + 879, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e434f605bf4506380301043f978de0813f9f16be30e16f768ac6c04911f40740.jpg", + "image_caption": [ + "Figure 4. Visualization of LLaMP Predictions by GradCAM [35] Fig. 3, using 16 prompts optimizes the LLM's capabilities, achieving the highest harmonic mean at $81.27\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 503, + 247, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations. In Fig. 4, we visualize the gradient heatmap of input images from FGVCAircraft and EuroSAT, through GradCAM [35]. The figure shows that, LLaMP can capture distinctive features that matches LLM's description.", + "bbox": [ + 498, + 518, + 890, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 593, + 617, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our study shows that the encyclopedic knowledge from LLMs is beneficial for low-shot image classification as extra class-specific information. To leverage such knowledge, we propose LLaMP, a framework that adapts LLMs as prompt learners for the CLIP model. Over two common low-shot scenarios: zero-shot generalization and few-shot learning, LLaMP demonstrates notable improvements compared with previous state-of-the-arts on a spectrum of datasets.", + "bbox": [ + 496, + 619, + 890, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While LLaMP reveals an effective way in leveraging LLMs' knowledge, both modalities, vision and language, only interact at the finest feature level. Given the broader LLM-aided knowledge from the language side, the performance can be potentially further improved by introducing language priors at earlier vision encoding stages.", + "bbox": [ + 496, + 739, + 890, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 845, + 650, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported, in part, by the Office of Naval Research under grant #N00014-21-1-2802.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28460", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, et al. Falcon-40b: an open large language model with state-of-the-art performance. Technical report, Technical report, Technology Innovation Institute, 2023. 2", + "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2", + "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461. Springer, 2014. 5, 6, 7", + "[4] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In ICCV, pages 122-131, 2021. 2", + "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 2", + "[6] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2", + "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5, 6, 7", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5, 6, 7", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 6", + "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR workshop, pages 178-178. IEEE, 2004. 5, 6, 7", + "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 6, 7", + "[12] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 5", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 6, 7, 8" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2", + "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4", + "[16] Michael Kampffmeyer, Yinbo Chen, Xiaodan Liang, Hao Wang, Yujia Zhang, and Eric P Xing. Rethinking knowledge graph propagation for zero-shot learning. In CVPR, pages 11487-11496, 2019. 2", + "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 2, 6", + "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 5, 6, 7", + "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 5, 6, 7, 8", + "[20] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th ICLR, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 2", + "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 5, 6, 7", + "[22] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. TPAMI, 36(3):453-465, 2013. 2", + "[23] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 2, 6", + "[24] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 6", + "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 5, 6, 7", + "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5, 6, 7", + "[27] OpenAI. Chatgpt, 2023. Available at https://openai.com/chatgpt.2", + "[28] OpenAI. Gpt-4 technical report, 2023. 1, 2, 5", + "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28461", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, pages 8026-8037, 2019. 5", + "[31] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5, 2023. 4", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7", + "[33] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 3", + "[34] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In CVPR, pages 49-58, 2016. 2", + "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626, 2017. 8", + "[36] Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D Manning, and Andrew Y Ng. Zero-shot learning through cross-modal transfer. arXiv preprint arXiv:1301.3666, 2013. 2", + "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7", + "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 4", + "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 4, 6", + "[40] Xiaolong Wang, Yufei Ye, and Abhinav Gupta. Zero-shot recognition via semantic embeddings and knowledge graphs. In CVPR, pages 6857-6866, 2018. 2", + "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 3", + "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3", + "[43] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online, 2020. Association for Computational Linguistics. 4", + "[44] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In CVPR, pages 5542-5551, 2018. 2", + "[45] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. IJCV, 119(1):3-22, 2016. 5, 6, 7", + "[46] Li Zhang, Tao Xiang, and Shaogang Gong. Learning a deep embedding model for zero-shot learning. In CVPR, pages 2021-2030, 2017. 2", + "[47] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2", + "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 5, 6, 7", + "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7", + "[50] Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR, pages 1004-1013, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 654 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28462", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_model.json b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d990e3e92263f48dd1bc417650e6d6bdb9fa44be --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.75, + 0.175 + ], + "angle": 0, + "content": "Large Language Models are Good Prompt Learners for Low-Shot Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.204, + 0.804, + 0.257 + ], + "angle": 0, + "content": "Zhaoheng Zheng Jingmin Wei Xuefeng Hu Haidong Zhu Ram Nevatia Viterbi School of Engineering University of Southern California" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.26, + 0.761, + 0.275 + ], + "angle": 0, + "content": "{zhaoheng.zheng, jingminw, xuefengh, haidongz, nevatia}@usc.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.31, + 0.314, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.645 + ], + "angle": 0, + "content": "Low-shot image classification, where training images are limited or inaccessible, has benefited from recent progress on pre-trained vision-language (VL) models with strong generalizability, e.g. CLIP. Prompt learning methods built with VL models generate text features from the class names that only have confined class-specific information. Large Language Models (LLMs), with their vast encyclopedic knowledge, emerge as the complement. Thus, in this paper, we discuss the integration of LLMs to enhance pre-trained VL models, specifically on low-shot classification. However, the domain gap between language and vision blocks the direct application of LLMs. Thus, we propose LLaMP, Large Language Models as Prompt learners, that produces adaptive prompts for the CLIP text encoder, establishing it as the connecting bridge. Experiments show that, compared with other state-of-the-art prompt learning methods, LLaMP yields better performance on both zero-shot generalization and few-shot image classification, over a spectrum of 11 datasets. Code will be made available at: https://github.com/zhaohengz/LLaMP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Low-shot image classification tasks, including few-shot and zero-shot variants, are to learn from a set of class names along with a limited or null set of images. Such capacities are crucial for the extension and generalization of vision systems. Vision-Language (VL) models trained on large-scale web data, such as CLIP [32] and ALIGN [14], provide a new paradigm due to their generalization capabilities that includes zero-shot classification, and have been used in recent work [17-19, 23, 24, 48, 49]. Due to the scarcity of images for training, methods built for both tasks rely heavily on merely category names as the source of class-specific knowledge, resulting in a shortage of distinguishable descriptions. Meanwhile, Large Language Models (LLMs)," + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.31, + 0.89, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.405, + 0.798, + 0.418 + ], + "angle": 0, + "content": "(a) Visual Descriptions from an LLM." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.422, + 0.888, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.554, + 0.824, + 0.567 + ], + "angle": 0, + "content": "(b) LLMs' Knowledge Boosts the Performance." + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.579, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Figure 1. Demonstration of LLaMP: (a) LLMs can provide visual descriptions for fine-grained object categories; (b) Zero-shot baseto-novel generalization benefits from the LLM knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.894, + 0.703 + ], + "angle": 0, + "content": "e.g. GPT-4 [28] and LLaMA [38, 39], have demonstrated their encyclopedic knowledge and thus can provide linguistic visual descriptions for objects. Here, we investigate how to leverage LLMs for low-shot image classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The emergence of prompt learning has provided an efficient way to adapt large pre-trained models. Previous work has explored various strategies to prompt vision-language (VL) models, including vision-conditioned text prompt learning [48], joint VL prompt learning [18] and self-regulated VL prompts[19]. On the text side, regardless of the learning strategy, learned prompt vectors are shared across all categories. The only difference among text inputs is the class name. In low-shot scenarios where visual data is limited, the extraction of class-specific knowledge from textual inputs becomes essential. However, the current paradigm, which relies on the CLIP text encoder to distinguish between class names, faces challenges, partic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28453" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "ularly with fine-grained target categories. For example, in FGVCAircraft [25], the class name \"Yak-40\", can barely provide any information for recognizing the object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.47, + 0.381 + ], + "angle": 0, + "content": "Large Language Models, trained with large text corpora, are good candidates to serve as the complement. As in Fig. 1a, being queried about \"Yak-40\", the LLM generates a sentence detailing the visual appearance of the Yak-40 that can be further parsed into noun phrases and integrated into text prompts, providing richer information, compared with the ordinary prompt. We also show in Fig. 1b that by simply incorporating noun phrases extracted from a LLM's response, the performance of the ordinary CLIP models is improved by more than \\(1\\%\\) without any training. Although recent prompt-learning based methods have shown notable improvements, it is non-trivial to apply them on textual visual descriptions generated by LLMs. Thus, instead of directly taking LLM generations as the textual input, we aim at producing class-specific representations by adapting LLMs to low-shot image classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.47, + 0.685 + ], + "angle": 0, + "content": "One challenge of the adaption is the domain gap between vision and language. When trained exclusively with textual corpora, the latent feature space of a LLM significantly diverges from that of its visual counterpart. Even worse, the data scarcity under the low-shot scenario make it virtually impossible to align two spaces through plain contrastive loss. We argue that, the CLIP text encoder, which is trained to project features from the language domain into the joint VL domain, can serve as the bridge. Thus, we propose the LLaMP framework, Large Language Models as Prompt learners, which leverages LLMs to learn informative prompts for CLIP models. In LLaMP, we treat LLMs as the prompt learner of the CLIP text encoder. More specifically, for each object category, LLaMP extracts corresponding knowledge from the LLM and yields class-specific prompt vectors, which are further combined with class-agnostic prompt embeddings (as in previous approaches), and encoded by the CLIP text encoder. We design an efficient tuning pipeline to avoid fully fine-tuning the language model while performing effective adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Following the protocol in [48, 49], we evaluate LLaMP with two typical scenarios: zero-shot base-to-novel generalization [49] and few-shot image classification. For each scenario, we run LLaMP with 11 datasets covering a spectrum of tasks. On average, LLaMP achieves a \\(1.3\\%\\) boost on the harmonic mean against the state-of-the-art PSRC [19], and \\(9.6\\%\\) over the ordinary CLIP [32], on base-to-novel generalization. We also observe an average improvement of \\(0.94\\%\\) on 16-shot image classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In summary, our approach makes use of Large Language Models to improve performance in low-shot image classification scenarios. The main contributions are: i) To the best of our knowledge, we are the first to investigate how to use the encyclopedic knowledge inherent in Large Lan" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "guage Models (LLMs) to enhance low-shot image classification; ii) We design a framework, LLaMP, to effectively adapt LLMs for image classification, without training the entire language model, and achieve state-of-the-art in both few-shot and zero-shot settings; iii) We conduct extensive analysis investigating the effectiveness of each components of LLaMP, and discuss the optimal setup for LLM-aided image classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.226, + 0.642, + 0.243 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.252, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Large Language Models (LLMs). Recent years have witnessed remarkable progress in scaling up the size and capabilities of LLMs. Zhang et al. [47] first introduced a suite of transformers pre-trained at scale, followed by PaLM [6]. ChatGPT/GPT-4 [27, 28] emerged as a milestone conversational model, demonstrated impressive conversational abilities as a generalist model. Vicuna [5] further advanced by learning from ChatGPT, while LLaMA [38] demonstrates that larger scale training yields stronger foundation models. The subsequent LLaMA-2 [39] and PaLM-2 [2] achieved further gains in scale, efficiency and reasoning. Most recently, Almazrouei et al. [1] released Falcon, a 400B model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Zero-Shot Learning (ZSL). ZSL stands in contrast to traditional fully-supervised paradigms. Instead of relying on direct visual training samples, it leverages side information that can be drawn from a multitude of non-visual domains, including attributes [22], word embeddings [36, 40], and descriptive texts [34]. Zhang et al. [46] designed an embedding model to bridge the gap between seen and unseen categories. Concurrently, studies like [4, 44, 50] have spotlighted that generative models can produce features for unseen categories. Moreover, Graph Convolution Networks (GCN) [20] has been explored in research such as [16, 40] for further generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Prompt Learning. With the progress in large-scale vision-language models, such as CLIP [32] and ALIGN [14], which reveal their capacity in zero-shot transferability, prompt learning has emerged as an efficient learning scheme, where learnable prompts are appended to the input to fine-tune models. For low-shot image classification, CoOp [49] and CoCoOp [48], which modeled context words as learnable vectors to automate prompt engineering, have shown significant improvements over regular CLIP. MaPLe [18] further employed a hierarchical multi-modal prompting strategy across transformer blocks for progressive feature modeling. Kan et al. [17] incorporated external knowledge by designing knowledge-aware prompts and adaptation head for better generalization. Lee et al. [23] used masked attention to prevent internal representation shift for better generalization. Khattak et al. [19] further improved prompt learning by guiding prompts to balance task-specific and task-agnostic knowledge via mutual agreement maximization and prompt ensemble." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28454" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.185, + 0.108 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.114, + 0.218, + 0.129 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.138, + 0.473, + 0.432 + ], + "angle": 0, + "content": "Similar to previous CLIP-based learning approaches, we consider the classification problem as an image-text matching problem. We denote the image encoder and the text encoder, in CLIP-like models, as \\(\\mathcal{F}\\) and \\(\\mathcal{G}\\), parameterized by \\(\\theta_{\\mathcal{F}}\\) and \\(\\theta_{\\mathcal{G}}\\), respectively. An input image \\(\\boldsymbol{x} \\in \\mathbb{R}^{C \\times H \\times W}\\) is split into \\(M\\) equal-sized patches which are converted into a sequence of embeddings \\(\\tilde{\\boldsymbol{x}} = \\{\\boldsymbol{e}_{cls}, \\boldsymbol{e}_1, \\boldsymbol{e}_2, \\dots, \\boldsymbol{e}_M\\}\\). The visual input sequence \\(\\tilde{\\boldsymbol{x}}\\) is encoded by the image encoder, producing the image feature \\(\\boldsymbol{f} = \\mathcal{F}(\\tilde{\\boldsymbol{x}})\\). On the text side, the text label \\(y\\) and the associated name is formatted as \"A photo of [STH]\" and tokenized into a sequence of tokens \\(\\tilde{\\boldsymbol{y}} = \\{\\boldsymbol{t}_{bos}, \\boldsymbol{t}_1, \\boldsymbol{t}_2, \\dots, \\boldsymbol{t}_L, \\boldsymbol{t}_{eos}\\}\\), where \\(L\\) is the length of input tokens. The input sequence is then encoded into \\(\\boldsymbol{g} = \\mathcal{G}(\\tilde{\\boldsymbol{y}})\\). For image classification, target class labels \\(\\{1, 2, \\dots, C\\}\\) are encoded into text features \\(g_i\\). The classification is done by picking the class that has the highest similarity with the vision feature: \\(\\hat{y} = \\mathrm{argmax}_i \\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}_i)\\), where \\(\\mathcal{C}\\) is the softmax cosine-similarity function \\(\\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}) = \\frac{\\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g} / \\tau)}{\\sum_{i=1}^{C} \\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g}_j / \\tau)}\\) with temperature \\(\\tau\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.47, + 0.597 + ], + "angle": 0, + "content": "Multimodal Prompting Learning. Given the size of the CLIP model, fine-tuning the entire model becomes infeasible. As both image and text encoders are built with standard transformer architecture, prompt learning, which tunes the model by combining trainable prompts with hidden states has been applied on the text encoder [48, 49], the image encoder [15, 41, 42], or both [18, 19, 33]. Similar to [19, 33], we build our method following the vision-language prompting paradigm, with deep prompting [15, 19], which not only insert prompts to the input layer, but to later encoder layers." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.473, + 0.84 + ], + "angle": 0, + "content": "More specifically, for each transformer layer that takes prompts, we define \\(V\\) learnable visual prompts \\(\\pmb{p}_v = \\{p_v^1, p_v^2, \\dots, p_v^V\\}\\) and \\(T\\) learnable language prompts \\(\\pmb{p}_t = \\{p_t^1, p_t^2, \\dots, p_t^T\\}\\). For the \\(i\\)-th vision encoder layer, visual prompts \\(\\pmb{p}_v^i\\) are appended to input embeddings: \\(\\tilde{\\pmb{x}}_p^i = \\{e_{cls}^i, e_1^i, e_2^i, \\dots, e_M^i, p_v^i\\}\\). The prompt-augmented vision feature, \\(\\pmb{f}_p = \\mathcal{F}(\\tilde{\\pmb{x}}_p)\\), is produced by jointly encoding prompts and the image. As the ViT [9] architecture in CLIP adopts the bi-directional attention mechanism, the placement of \\(\\pmb{p}_v\\) has no effect on \\(\\pmb{f}_p\\). On the language side, prompts are concatenated with the input of the \\(i\\)-th text encoder: \\(\\tilde{\\pmb{y}}_p^i = \\{\\pmb{t}_{bos}^i, \\pmb{p}_t^i, \\pmb{t}_1^i, \\pmb{t}_2^i, \\dots, \\pmb{t}_L^i, \\pmb{t}_{eos}^i\\}\\). \\(\\tilde{\\pmb{y}}_p\\) is further processed by the text encoder, resulting in the prompt-augmented language feature \\(\\pmb{g}_p = \\mathcal{G}(\\tilde{\\pmb{y}}_p)\\). More specifically, prompts to the first layer \\(\\pmb{p}_t^1\\) are initialized with the embeddings of \"A photo of a\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Low-Rank Adaptation [13] (LoRA). As a parameter-efficient tuning technique, LoRA is designed to adapt large transformer model without updating original model weights. The LoRA technique is, in particular, applied to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "linear projection layers. More specifically, for a linear layer with weight \\( W_0 \\in \\mathbb{R}^{d \\times k} \\), LoRA creates \\( \\Delta W \\) by learning two low rank matrices \\( B \\in \\mathbb{R}^{d \\times r} \\) and \\( A \\in \\mathbb{R}^{r \\times k} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.148, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} = \\left(W _ {0} + \\Delta W\\right) \\boldsymbol {x} = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.174, + 0.892, + 0.219 + ], + "angle": 0, + "content": "We adopt a hybrid tuning scheme on the vision encoder, which performs prompt learning on the first few layers and applies LoRA on the rest." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.228, + 0.836, + 0.244 + ], + "angle": 0, + "content": "3.2. Adaptive Prompt Learning with LLMs" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.251, + 0.892, + 0.313 + ], + "angle": 0, + "content": "The goal of prompt tuning is to find a set of optimal prompts \\(\\pmb{p} = \\{\\pmb{p}_v, \\pmb{p}_t\\}\\) which maximizes the log likelihood of \\(P(x, y | \\theta_{\\mathcal{F}}, \\theta_{\\mathcal{G}})\\) over target downstream distribution \\((x, y) \\sim (X, Y)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.322, + 0.892, + 0.36 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} = \\underset {\\boldsymbol {p}} {\\operatorname {a r g m a x}} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim (\\boldsymbol {X}, \\boldsymbol {Y})} \\log \\mathcal {C} \\left(\\mathcal {F} \\left(\\boldsymbol {x}; \\boldsymbol {p} _ {v}\\right), \\mathcal {G} \\left(\\boldsymbol {y}; \\boldsymbol {p} _ {t}\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.481 + ], + "angle": 0, + "content": "However, the \\( p \\) optimized through Eqn. 2 has two issues. First, \\( p \\) is shared for all categories for the downstream task, while the optimal prompt for each category might be different. Second, in low-shot scenarios, \\( p \\) are usually empirically estimated from a limited training-set \\( X^{train} \\) with limited categories \\( \\{1, 2, \\dots, C^{base}\\} \\), and therefore such \\( p \\) can often be over-fitted to the small training-set \\( X^{train} \\) and fail to generalize to novel categories outside \\( \\{1, 2, \\dots, C^{base}\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.482, + 0.892, + 0.603 + ], + "angle": 0, + "content": "To overcome these problems, we propose to learn a meta function on the language side \\( \\pmb{p}_t = \\Theta(y) \\) which can adaptively estimate the optimal prompt for each category. An intuitive way to estimate proper prompts \\( \\pmb{p} \\) for category name \\( y \\) is to take advantage of the knowledge of the pre-trained Large Language Models (LLM) \\( \\mathcal{D} \\) and extract discriminative descriptions of category \\( y \\). For example, given the input text \\( z: \\text{\"Describe {y}\"} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.613, + 0.892, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} = \\left\\{p _ {1}, p _ {2}, \\dots , p _ {k} \\right\\} = \\mathcal {D} (\\boldsymbol {z}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.64, + 0.851, + 0.655 + ], + "angle": 0, + "content": "while \\(p_i\\) being sequentially generated by \\(\\mathcal{D}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.664, + 0.892, + 0.689 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\mathcal {D} (\\boldsymbol {z}, t _ {1}, \\dots , t _ {i - 1}) = \\mathcal {D} ^ {(i)} (\\boldsymbol {z}) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.687, + 0.673, + 0.701 + ], + "angle": 0, + "content": "\\[\nt _ {i} = \\mathcal {M} (p _ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.789 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}^{(i)}\\) is the \\(i\\)-th forward iteration of \\(\\mathcal{D}\\), and \\(\\mathcal{M}\\) maps continuous hidden states into discrete language tokens. To accelerate the process and to obtain \\(p\\) in one pass, we approximate the above process with \\(K\\) learnable prompts \\(\\pmb{p}_l = \\{\\theta_1,\\dots,\\theta_K\\}\\) so that" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.799, + 0.892, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {t} = \\Theta (y) = \\mathcal {D} \\left(\\left\\{\\theta_ {1}, \\dots , \\theta_ {K} \\right\\} | z\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Discussion. While Large Language Models (LLMs) possess robust foundational knowledge within the linguistic domain, it is not feasible to directly substitute the text encoder of CLIP with an LLM. The reason lies in the inherent divergence between the LLM's latent space, which" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.09, + 0.891, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.315, + 0.894, + 0.371 + ], + "angle": 0, + "content": "Figure 2. An Overview of the LLaMP Framework: We first generate the knowledge cache by passing the query prompt through the LLM \\(\\mathcal{D}\\) and use the knowledge cache to encode \\(\\pmb{p}_l\\), resulting in the adaptive prompts \\(\\tilde{\\pmb{h}}_l^i = W\\pmb{h}_l^i + b^i\\) for the CLIP text encoder. \\(\\tilde{\\pmb{h}}_l\\) is combined with regular learnable prompts of \\(\\mathcal{G}\\) to generate the final text feature vector \\(\\pmb{g}_p\\). The image feature vector \\(\\pmb{f}_p\\) is obtained through a hybrid-tuning strategy combining prompt learning and low-rank adaptation (LoRA)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.471, + 0.501 + ], + "angle": 0, + "content": "is purely language-oriented, and the image-focused latent space of vision encoders. Attempting a direct alignment via contrastive learning would require an extensive dataset that is typically beyond the scope of low-shot learning. To bridge this gap, we introduce LLaMP—an adaptive prompt learning framework that leverages the LLM to craft class-specific prompt vectors, to reinforce the text encoder for low-shot image classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.304, + 0.523 + ], + "angle": 0, + "content": "3.3. The LLaMP Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.531, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Fig. 2 shows an overview of the LLaMP framework. For convenience, we denote the decoder-only LLM as \\(\\mathcal{D}\\). The input to the decoder \\(\\mathcal{D}\\) consists of two components: textual prompts \\(\\pmb{y}\\) in the form of sentences, tokenized as \\(\\tilde{\\pmb{y}}\\), and learnable prompts \\(\\pmb{p}_l\\). We append prompt embeddings to the end of the input sequence and obtain the last hidden states of \\(\\mathcal{D}\\) as the feature \\(\\pmb{h}_l\\):" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.646, + 0.47, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {h} _ {l} = \\mathcal {D} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {p} _ {l}) [ L + 1: L + K ], L = \\operatorname {L e n g t h} (\\tilde {\\boldsymbol {y}}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Hidden states of \\(\\mathcal{D}\\) are then mapped to the input space of the CLIP text encoder by the projection matrix \\(W\\in \\mathbb{R}^{d_1\\times d_2}\\), where \\(d_{1}\\) and \\(d_{2}\\) are respectively the hidden sizes of the LLM \\(\\mathcal{D}\\) and the CLIP text encoder \\(\\mathcal{G}\\). A set of prompt-specific biases \\(b\\in \\mathbb{R}^{K\\times d_2}\\) are added:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.756, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {h}} _ {l} = W \\boldsymbol {h} _ {l} + b \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.47, + 0.875 + ], + "angle": 0, + "content": "We combine \\(\\tilde{h}_l\\) from LLM with regular learnable prompts, as in previous approaches [19], to construct the input for CLIP text encoder. Similar to deep prompting [15, 19], we create layer-specific prompts through different \\(W\\) matrices and \\(b\\) vectors. For the \\(i\\)-th layer, we let \\(\\tilde{\\pmb{h}}_l^i = W^i\\pmb{h}_l + b^i\\) and the entire sequence is constructed as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.884, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {y}} _ {l} ^ {i} = \\left\\{\\boldsymbol {t} _ {b o s} ^ {i}, \\boldsymbol {p} _ {t} ^ {i}, \\boldsymbol {t} _ {1} ^ {i}, \\boldsymbol {t} _ {2} ^ {i}, \\dots , \\boldsymbol {t} _ {L} ^ {i}, \\tilde {\\boldsymbol {h}} _ {l} ^ {i}, \\boldsymbol {t} _ {e o s} ^ {i} \\right\\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.529 + ], + "angle": 0, + "content": "LLM Knowledge Cache. A Large Language Model (LLM), as implied by its name, typically comprises billions of parameters. For example, the most compact LLaMA [38, 39] model has 7B parameters. Thus, even performing prompt learning on a LLM become impractical. The memory consumption to store gradients for back propagation can go beyond the limit of mainstream GPUs. Instead, the causal attention mechanism inherent in decoder-only LLMs, where the embedding of an input token only depends on the preceding tokens, facilitates a feasible workaround." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.893, + 0.639 + ], + "angle": 0, + "content": "As previously mentioned, the prompt embeddings \\( \\pmb{p}_l \\) are appended to the end of text tokens \\( \\tilde{\\pmb{y}} \\). According to the causal attention mechanism, \\( \\tilde{\\pmb{y}} \\) is encoded independently of \\( \\pmb{p}_l \\). Thus, we design a two-stage process, where we create the LLM knowledge cache by passing \\( \\tilde{\\pmb{y}} \\) through \\( \\mathcal{D} \\) and leverage the cache to convert \\( \\pmb{p}_l \\) into class-specific embeddings for the CLIP text encoder \\( \\mathcal{G} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.822 + ], + "angle": 0, + "content": "To compute the attention of a token, the only dependency is the Key and Value vectors from the preceding tokens. Thus, we adopt the KV-cache [31, 43], a technique used in inference acceleration of LLMs, to create the knowledge cache. At the first stage, we pass text tokens \\(\\tilde{\\pmb{y}}\\) through the language model \\(\\mathcal{D}\\) and save the Keys and Values as the knowledge cache for the second stage. Once computed, the knowledge cache remains fixed throughout the entire training process and bears the information that is needed for further computation. Thus, in LLaMP, we leverage the knowledge cache obtained at the first stage to generate class-specific prompt embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "At the second stage, we create class-specific prompt embeddings from the pre-computed knowledge cache. As \\( \\pmb{p}_l \\) is not initialized in the natural language domain, it need not pass through the entire LLM; instead, we insert those prompts \\( \\pmb{p}_l \\) to the last layer of the LLM \\( \\mathcal{D}_N \\). It is achieved" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.423, + 0.108 + ], + "angle": 0, + "content": "by encoding them alongside the cache from \\(\\tilde{\\pmb{y}}\\), as in" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.119, + 0.47, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {H} _ {l} = \\mathcal {D} _ {N} \\left(\\boldsymbol {K} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {V} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {p} _ {l}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.146, + 0.47, + 0.236 + ], + "angle": 0, + "content": "where \\( K_{\\tilde{y}}, V_{\\tilde{y}} \\) represent the knowledge cache. This design enables LLaMP to efficiently learn informative prompt embeddings for the CLIP encoder \\( \\mathcal{G} \\). It accomplishes this by incurring modest training costs, compared with training the entire LLM. Simultaneously, it maintains the essential knowledge inherent in the LLM decoder \\( \\mathcal{D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.523 + ], + "angle": 0, + "content": "Training Targets of LLaMP. Although the training strategy in Eqn. 9 has reduced the number of learnable parameters, a full decoder inside a LLM still consists of an enormous number of parameters. For example, one layer in LLaMA-7B bears 200M parameters, making training of the entire layer costly. As the goal is to leverage the knowledge from LLM, altering a full layer can lead to the loss of knowledge. Thus, as shown in Fig. 2, a typical decoder layer has two major components: the self attention module, consisting Query, Key, Value and Output projection layers, and the Feed-Forward Network (FFN). LLaMP targets the Query and Output projection layer inside the self-attention module. By updating the Query layer, LLM prompts \\( p_l \\) are learned to distill pertinent information from the knowledge cache and the Output layer projects it to the latent space. We keep the Key and Value layers frozen to ensure the alignment between \\( p_l \\) and knowledge cache. We leave the FFN unchanged to preserve the knowledge. Further discussions regarding these choices are made in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Textual Priors from Pre-Generated Responses. We extend the initial prompt, \"In one sentence, describe the distinctive appearance of [STH]\", by incorporating the response generated by the language model into the input sequence. This approach enriches the base content: the generated text provides a clear and explicit description of the object's appearance, acting as a valuable informative prior for language model adaptation. However, it is common for responses from an LLM to include filler words like \"sure\" for sentence structure coherence. To refine the input, we parse the noun phrases from the LLM's response through spaCy [12], an NLP engine, and merge them with the initial prompt, forming a more focused and informative language prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Textual Augmentations. Following the insights of Khattak et al. [19], which highlight the performance benefits of diverse textual inputs, we aim to further augment the text inputs used in the CLIP text encoder. Our approach, building upon the methods in [19, 48], incorporates handcrafted templates and expands their diversity through a two-step process: i) We introduce noun phrases into the existing templates for CLIP, for example, transforming \"A photo of [STH]\" to \"A photo of [STH] with [NP]\", thereby enriching the descriptive content; ii) We create a variety of new prompt templates for the LLM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "similar to “In one sentence, describe the distinctive appearance of [STH]” through GPT-4 [28], to further diversify the text input." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.145, + 0.715, + 0.161 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.168, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Similar to PSRC [19], our objective function consists of three components: The main cross-entropy loss \\(\\mathcal{L}_{CE}\\), feature-level L1 regularization \\(\\mathcal{L}_{l1}\\), and soft distillation loss \\(\\mathcal{L}_{dist}\\). Given \\(\\mathcal{C}\\) training categories and \\(\\mathcal{N}\\) training samples, \\(\\mathcal{L}_{CE}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.252, + 0.892, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} = - \\frac {1}{\\mathcal {N}} \\sum_ {i} \\log \\frac {\\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} / \\tau\\right)}{\\sum \\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {j} / \\tau\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.299, + 0.892, + 0.332 + ], + "angle": 0, + "content": "The L1 regularization is computed between learned features \\( \\pmb{f}_p, \\pmb{g}_p \\) and pre-trained CLIP features \\( \\hat{\\pmb{f}}, \\hat{\\pmb{g}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.338, + 0.892, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l 1} = \\frac {1}{\\mathcal {N}} \\sum_ {i} \\lambda_ {v} \\left| \\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {f}} ^ {i} \\right| + \\frac {1}{C} \\sum_ {i} \\lambda_ {t} \\left| \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {g}} ^ {i} \\right|, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.381, + 0.892, + 0.427 + ], + "angle": 0, + "content": "where \\(\\lambda_v\\) and \\(\\lambda_t\\) are coefficients. The prediction of LLaMP is further bound by the KL-Divergence between predicted distributions of LLaMP and vanilla CLIP:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.436, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\lambda_ {\\text {d i s t}} D _ {K L} \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}}, \\hat {\\boldsymbol {f}} \\cdot \\hat {\\boldsymbol {g}}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.892, + 0.493 + ], + "angle": 0, + "content": "We sum all three losses up as the final objective function: \\(\\mathcal{L} = \\mathcal{L}_{CE} + \\mathcal{L}_{l1} + \\mathcal{L}_{dist}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.554 + ], + "angle": 0, + "content": "During training, we randomly sample one LLM template as the input of LLaMP for each batch. For inference, we compute the probability distribution predicted from each input template and average them." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.634, + 0.584 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.676, + 0.607 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Datasets. Similar to previous work [18, 19, 48], in our study, we evaluate LLaMP performance over a spectrum of classification tasks with 11 datasets, including ImageNet [8] and Caltech101 [10] for generic image classification, OxfordPets [29], StanfordCars [21], Flowers102 [26], Food101 [3], and FGVCAircraft [25] for fine-grained classification, SUN397 [45] for scene recognition, UCF101 [37] for action recognition, DTD [7] for texture classification, and EuroSAT [11] for satellite image recognition." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Scenarios & Metrics. We evaluate LLaMP on two typical low-shot scenarios: zero-shot base-to-novel generalization and few-shot image classification. In zero-shot base-to-novel generalization, the base classes are seen during training, while the novel classes are unseen. We measure models performance through accuracies of base and novel classes, and the harmonic mean of the two. For few-shot classification, we assess the accuracy with 16 shots per class." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We build LLaMP through the PyTorch [30] framework. All models are trained with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28457" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.867, + 0.284 + ], + "angle": 0, + "content": "
MethodAverageImageNet [8]Caltech101 [10]OxfordPets [29]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3474.2271.7072.4368.1470.2296.8494.0095.4091.1797.2694.12
CoOp [49]82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp [48]80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
KAPT* [17]78.4170.5274.2671.1065.2068.0297.1093.5395.2893.1396.5394.80
ProDA [24]81.5672.3076.6576.6670.5473.4797.7494.3696.0295.4397.7696.58
MaPLe [18]82.2875.1478.5575.4070.3272.7298.2793.2395.6895.4397.8396.62
RPO [23]81.1375.0077.7876.6071.5774.0097.9794.3796.0394.6397.5096.05
PSRC [19]84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
LLaMP85.1677.7181.2777.9971.2774.4898.4595.8597.1396.3197.7497.02
Δ w.r.t. PSRC+0.90+1.61+1.30+0.39+0.54+0.47+0.35+1.82+1.11+0.98+0.44+0.72
" + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.288, + 0.865, + 0.484 + ], + "angle": 0, + "content": "
MethodStanfordCars [21]Flowers102[26]Food101 [3]FGVCAircraft [25]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]63.3774.8968.6572.0877.8074.8390.1091.2290.6627.1936.2931.09
CoOp [49]78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp [48]70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
KAPT* [17]69.4766.2067.7995.0071.2081.4086.1387.0686.5929.6728.7329.19
ProDA [24]72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
MaPLe [18]74.7071.2072.9197.7068.6880.6690.3088.5789.4336.9034.1335.46
RPO [23]73.8775.5374.6994.1376.6784.5090.3390.8390.5837.3334.2035.70
PSRC [19]78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
LLaMP81.5674.5477.8997.8277.4086.4291.0591.9391.4947.3037.6141.90
Δ w.r.t. PSRC+3.29-0.43+1.31-0.25+0.90+0.47+0.38+0.40+0.39+4.57-0.26+1.75
" + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.488, + 0.865, + 0.684 + ], + "angle": 0, + "content": "
MethodSUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3675.3572.2353.2459.9056.3756.4864.0560.0370.5377.5073.85
CoOp [49]80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp [48]79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
KAPT* [17]79.4074.3376.7875.9758.3065.9784.8067.5775.2180.8367.1073.33
ProDA [24]80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
MaPLe [18]78.4776.9377.7980.6756.4866.4483.9066.0073.8885.2371.9778.04
RPO [23]80.6077.8079.1876.7062.1368.6186.6368.9776.7983.6775.4379.34
PSRC [19]82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
LLaMP83.4179.9081.6283.4964.4972.7791.9383.6687.6087.1380.6683.77
Δ w.r.t. PSRC+0.74+1.43+1.10+0.12+1.52+1.02-0.97+9.76+5.28+0.03+1.86+1.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.696, + 0.895, + 0.738 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on base-to-novel generalization. LLaMP shows strong generalization results over previous approaches on 11 image classification tasks. Absolute gains over PSRC are indicated in blue. *KAPT is trained with ViT-B/32 image encoder instead of ViT-B/16." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.746, + 0.473, + 0.899 + ], + "angle": 0, + "content": "2 NVIDIA A100 40GB GPUs. For LLaMP, we adopt LLaMA2-7B [39] as the language model \\(\\mathcal{D}\\), and ViT-B/16 [9] as the image encoder, following [18, 19, 48, 49]. On the text side, we set prompt learning depth to 9. To tune the vision encoder, we adopt the hybrid tuning scheme which performs deep prompt learning on the first 6 layers and LoRA on the rest. Similar to [13], LoRA is applied to the Query and Value projection layers inside attention modules. The number of \\(p_l\\) prompts, \\(K\\), is set to 16. We set a global learning rate of 2E-4 with a batch size of 8. The learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.747, + 0.892, + 0.779 + ], + "angle": 0, + "content": "rate of LoRA modules is set to 2E-5. \\(\\lambda_{t},\\lambda_{v}\\) and \\(\\lambda_{dist}\\) are set to 25, 10 and 2.5, respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.787, + 0.723, + 0.803 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Zero-Shot Base-to-Novel Generalization. LLaMP outperforms existing state-of-the-art prompt learning methods on most metrics of 11 classification datasets in the baseto-novel generalization benchmark. As shown in Tab. 1, compared to the latest model PSRC [19], LLaMP achieves average gains of \\(0.90\\%\\) in base accuracy, \\(1.61\\%\\) in novel" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28458" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.842, + 0.273 + ], + "angle": 0, + "content": "
16-Shot Classification
AverageImageNet [8]Collect [10]Pets [29]Cars [21]Flowers [26]Food [3]Aircraft [25]SUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
CLIP [32]78.79 (65.02)67.3195.4385.3480.4497.3782.9045.3673.2869.9687.2182.11
CoOp [49]79.89 (73.82)71.8795.5791.8783.0797.0784.2043.4074.6769.8784.9382.23
CoCoOp [48]74.90 (70.70)70.8395.1693.3471.5787.8487.2531.2172.1563.0473.3278.14
MaPLe [18]81.79 (75.58)72.3396.0092.8383.5797.0085.3348.4075.5371.3392.3385.03
PSRC [19]82.87 (77.90)73.1796.0793.6783.8397.6087.5050.8377.2372.7392.4386.47
LLaMP83.81 (78.50)73.4997.0894.2186.0798.0687.6256.0777.0274.1791.3186.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.877, + 0.296 + ], + "angle": 0, + "content": "Table 2. Few shot classification results with 16 shots. Numbers in the bracket indicate the average performance over \\( 1/2/4/8/16 \\) shots." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.305, + 0.469, + 0.397 + ], + "angle": 0, + "content": "accuracy, and \\(1.30\\%\\) in harmonic mean on average. Moreover, LLaMP consistently achieves higher harmonic means (HM) compared to other models. These improvements indicate that our approach better balances performance on base and novel data, thus achieving stronger generalization compared to the prior prompt learning techniques." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.469, + 0.577 + ], + "angle": 0, + "content": "In particular, LLaMP excels in fine-grained datasets requiring detailed analysis. On FGVCAircraft, LLaMP surpasses PSRC by \\(4.57\\%\\) on base accuracy and \\(1.75\\%\\) on HM, highlighting its strong understanding of detailed aircraft features. Furthermore, on EuroSAT, LLaMP achieves improvements of \\(9.76\\%\\) and \\(5.28\\%\\) on novel accuracy and HM, respectively. We also observe similar performance gains on StanfordCars, where LLaMP outperforms by \\(3.29\\%\\) on base accuracy and \\(1.31\\%\\) on HM. The information embedded in LLM enables LLaMP to capture and utilize the rich semantic information necessary for distinguishing between closely related categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.469, + 0.76 + ], + "angle": 0, + "content": "Few-Shot Classification. LLaMP also achieves improvements across these classification datasets in few-shot classification tasks. As in Tab.2, with an average classification accuracy of \\(83.81\\%\\). Notably, on FGVCAircraft and StanfordCars, LLaMP shows a significant improvement over PSRC, further demonstrating that the knowledge from language models benefits the recognition of fine-grained object categories, which aligns with our observation on zero-shot base-to-novel generalization. Moreover, on DTD, where MaPLe and PSRC achieve around \\(72\\%\\) accuracy, LLaMP achieves a higher accuracy of \\(74.17\\%\\), underscoring its ability to recognize textures." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.231, + 0.787 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Is the knowledge from LLM helping? In Tab. 3, we show that the knowledge from LLM benefits in both ways: Without training, performance of ordinary CLIP model can be improved by introducing noun phrases; The LLaMP framework shows further improvement after training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Noun phrases are parsed from the LLM's responses the prompt of \"Describe [STH]\". We then use the tem" + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.303, + 0.848, + 0.402 + ], + "angle": 0, + "content": "
MethodLLMBaseNovelHM
CLIP✓69.3474.2271.70
70.9574.9372.79
LLaMP✓82.2176.4479.22
85.1677.7181.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.412, + 0.842, + 0.426 + ], + "angle": 0, + "content": "Table 3. Ablation study on the LLM Knowledge." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.435, + 0.889, + 0.545 + ], + "angle": 0, + "content": "
LPQOKVFFN%BaseNovelHM
✓.0385.0077.2980.96
✓✓✓3385.2077.4581.14
✓✓✓8385.0577.7381.22
✓✓✓✓10085.2377.5681.21
✓✓1785.1677.7181.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.555, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Table 4. Ablation study on the Training Strategy. “%” indicates the ratio of parameters trained compared to fully tuning a layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.674 + ], + "angle": 0, + "content": "plate, \"A photo of [STH] with [NP]\" to generate the NP-augmented text embedding for CLIP. We take the average of all augmented embeddings for classification. In Tab. 3 we show that even ordinary CLIP can benefit from incorporating LLMs' knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Furthermore, the comparison between LLaMP and LLaMP without the LLM indicates that merely integrating LoRA [13] to the vision encoder is not beneficial. The \"LLaMP without LLM\" is essentially an ordinary prompting learning model plus LoRAs in the vision encoder. We show that the improved vision encoding capacity only benefits when the quality of text embeddings s are enhanced by incorporating LLMs' knowledge through LLaMP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Decoder Training strategy. We categorize trainable parameters of \\(\\mathcal{D}_N\\) into four groups: learnable prompts (LP), Query and Output projections (QO), Key and Value projections (KV), and the feed-forward network (FFN). Tab. 4 indicates LLaMP can achieve desirable results by just learning the prompts of \\(\\mathcal{D}\\). One step further, adding QO into optimization achieves the best performance. Although other" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28459" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.428, + 0.166 + ], + "angle": 0, + "content": "
MethodPriorsBaseNovelHM
LLaMPX84.9077.5981.08
Plain85.2677.5681.22
NP85.1677.7181.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.177, + 0.47, + 0.204 + ], + "angle": 0, + "content": "Table 5. Ablation Study on Pre-generated Text Priors. \\( \\times \\) refers to \"without textual priors\" and NP stands for noun phrases." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.215, + 0.408, + 0.277 + ], + "angle": 0, + "content": "
MethodBaseNovelHM
LLM Only81.7435.8249.81
LLaMP85.1677.7181.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.288, + 0.419, + 0.302 + ], + "angle": 0, + "content": "Table 6. The CLIP text encoder helps adaptation." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.312, + 0.451, + 0.395 + ], + "angle": 0, + "content": "
SchemeBaseNovelHM
Prompt × 984.6777.2880.81
LoRA × 1284.8977.2780.90
Prompt ×6 + LoRA ×685.1677.7181.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.406, + 0.468, + 0.434 + ], + "angle": 0, + "content": "Table 7. Study on Vision Tuning Scheme. Our hybrid design achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.468, + 0.474 + ], + "angle": 0, + "content": "setups introduce much more trainable parameters, they can not surpass the \"LP + QO\" strategy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.476, + 0.468, + 0.581 + ], + "angle": 0, + "content": "Effect of Textual Priors. We study the effect of pregenerated textual priors on LLaMP. We compare three different approaches: without textual priors, using plain responses as the prior, and LLaMP, which takes parsed noun phrases. Tab. 5 shows that LLaMP can achieve over \\(81\\%\\) on HM without pre-generated priors, while adding parsed noun phrases as textual priors further pushes the HM to \\(81.27\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.734 + ], + "angle": 0, + "content": "CLIP as the bridge. One may wonder if it is possible to replace CLIP text encoder with a large language model. Here, we study two setups: i) LLM as encoder, which treats the output of the language model, \\(\\tilde{h}_l\\) as the text feature; ii) LLaMP, which treat \\(\\tilde{h}_l\\) as part of the text input prompt. Tab. 6 reveals that relying solely on the LLM results in poor accuracy for novel categories. This supports our hypothesis that aligning LLMs with vision encoders generally requires a more extensive dataset. Furthermore, LLaMP's design significantly improves the novel accuracy by \\(40\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.468, + 0.869 + ], + "angle": 0, + "content": "Vision Training Strategy. As ViT-16/B has 12 transformer layers, we compare different vision training strategies within LLaMP in Tab. 7. Apart from the default hybrid scheme, we evaluate setups including prompt learning at first 9 layers (P9), a similar setup to PSRC [19], and LoRA [13] in all 12 layers. The results suggest that the scheme leverages the strengths of both prompt learning and LoRA, addressing potential bottlenecks in the vision encoder and enhancing overall performance in LLaMP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Number of LLM Prompts. We vary the number of LLM prompts and study their effects on LLaMP. As in" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.097, + 0.88, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.217, + 0.892, + 0.245 + ], + "angle": 0, + "content": "Figure 3. Effect of LLM Prompts on Harmonic Mean. 16 prompts achieve the most balanced performance." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.248, + 0.892, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.477, + 0.89, + 0.519 + ], + "angle": 0, + "content": "Figure 4. Visualization of LLaMP Predictions by GradCAM [35] Fig. 3, using 16 prompts optimizes the LLM's capabilities, achieving the highest harmonic mean at \\(81.27\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Visualizations. In Fig. 4, we visualize the gradient heatmap of input images from FGVCAircraft and EuroSAT, through GradCAM [35]. The figure shows that, LLaMP can capture distinctive features that matches LLM's description." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.594, + 0.618, + 0.609 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.74 + ], + "angle": 0, + "content": "Our study shows that the encyclopedic knowledge from LLMs is beneficial for low-shot image classification as extra class-specific information. To leverage such knowledge, we propose LLaMP, a framework that adapts LLMs as prompt learners for the CLIP model. Over two common low-shot scenarios: zero-shot generalization and few-shot learning, LLaMP demonstrates notable improvements compared with previous state-of-the-arts on a spectrum of datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.741, + 0.892, + 0.832 + ], + "angle": 0, + "content": "Limitations. While LLaMP reveals an effective way in leveraging LLMs' knowledge, both modalities, vision and language, only interact at the finest feature level. Given the broader LLM-aided knowledge from the language side, the performance can be potentially further improved by introducing language priors at earlier vision encoding stages." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.651, + 0.862 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This research was supported, in part, by the Office of Naval Research under grant #N00014-21-1-2802." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28460" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[1] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, et al. Falcon-40b: an open large language model with state-of-the-art performance. Technical report, Technical report, Technology Innovation Institute, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.47, + 0.3 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461. Springer, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In ICCV, pages 122-131, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.671 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR workshop, pages 178-178. IEEE, 2004. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[12] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.272 + ], + "angle": 0, + "content": "[16] Michael Kampffmeyer, Yinbo Chen, Xiaodan Liang, Hao Wang, Yujia Zhang, and Eric P Xing. Rethinking knowledge graph propagation for zero-shot learning. In CVPR, pages 11487-11496, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.328 + ], + "angle": 0, + "content": "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.893, + 0.384 + ], + "angle": 0, + "content": "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[20] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th ICLR, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.893, + 0.552 + ], + "angle": 0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.893, + 0.594 + ], + "angle": 0, + "content": "[22] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. TPAMI, 36(3):453-465, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.893, + 0.65 + ], + "angle": 0, + "content": "[23] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[24] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[27] OpenAI. Chatgpt, 2023. Available at https://openai.com/chatgpt.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.806, + 0.859 + ], + "angle": 0, + "content": "[28] OpenAI. Gpt-4 technical report, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In \\(CVPR\\), pages 3498-3505. IEEE, 2012. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28461" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, pages 8026-8037, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[31] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.316 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.318, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[33] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.432 + ], + "angle": 0, + "content": "[34] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In CVPR, pages 49-58, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[36] Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D Manning, and Andrew Y Ng. Zero-shot learning through cross-modal transfer. arXiv preprint arXiv:1301.3666, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.472, + 0.603 + ], + "angle": 0, + "content": "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.472, + 0.673 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.472, + 0.744 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[40] Xiaolong Wang, Yufei Ye, and Abhinav Gupta. Zero-shot recognition via semantic embeddings and knowledge graphs. In CVPR, pages 6857-6866, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.302 + ], + "angle": 0, + "content": "[43] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online, 2020. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[44] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In CVPR, pages 5542-5551, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[45] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. IJCV, 119(1):3-22, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[46] Li Zhang, Tao Xiang, and Shaogang Gong. Learning a deep embedding model for zero-shot learning. In CVPR, pages 2021-2030, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.445, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[47] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.557 + ], + "angle": 0, + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.558, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[50] Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR, pages 1004-1013, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28462" + } + ] +] \ No newline at end of file diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_origin.pdf b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b9daa10ce38ba258a06c539d50ac306d55a32b57 --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/9c680495-254d-41be-9282-c0c5d7c9a2e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5dce1767bc53d7c7e59b83fb4448a6d8b11d3b1e2d0c25835f1b46451224cb1 +size 1387970 diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/full.md b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c72dee2eef4a63ea069cefb03fb881d37ad8af6 --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/full.md @@ -0,0 +1,324 @@ +# Large Language Models are Good Prompt Learners for Low-Shot Image Classification + +Zhaoheng Zheng Jingmin Wei Xuefeng Hu Haidong Zhu Ram Nevatia Viterbi School of Engineering University of Southern California + +{zhaoheng.zheng, jingminw, xuefengh, haidongz, nevatia}@usc.edu + +# Abstract + +Low-shot image classification, where training images are limited or inaccessible, has benefited from recent progress on pre-trained vision-language (VL) models with strong generalizability, e.g. CLIP. Prompt learning methods built with VL models generate text features from the class names that only have confined class-specific information. Large Language Models (LLMs), with their vast encyclopedic knowledge, emerge as the complement. Thus, in this paper, we discuss the integration of LLMs to enhance pre-trained VL models, specifically on low-shot classification. However, the domain gap between language and vision blocks the direct application of LLMs. Thus, we propose LLaMP, Large Language Models as Prompt learners, that produces adaptive prompts for the CLIP text encoder, establishing it as the connecting bridge. Experiments show that, compared with other state-of-the-art prompt learning methods, LLaMP yields better performance on both zero-shot generalization and few-shot image classification, over a spectrum of 11 datasets. Code will be made available at: https://github.com/zhaohengz/LLaMP. + +# 1. Introduction + +Low-shot image classification tasks, including few-shot and zero-shot variants, are to learn from a set of class names along with a limited or null set of images. Such capacities are crucial for the extension and generalization of vision systems. Vision-Language (VL) models trained on large-scale web data, such as CLIP [32] and ALIGN [14], provide a new paradigm due to their generalization capabilities that includes zero-shot classification, and have been used in recent work [17-19, 23, 24, 48, 49]. Due to the scarcity of images for training, methods built for both tasks rely heavily on merely category names as the source of class-specific knowledge, resulting in a shortage of distinguishable descriptions. Meanwhile, Large Language Models (LLMs), + +![](images/13ba31720a95e4f52c8efd8e4be9287dc1acb43ea3910e2996b90dc8748222a6.jpg) +(a) Visual Descriptions from an LLM. + +![](images/44ef9588350c8f3b111c503448edbde37cb3f6386a248e4c5a54c4b4c7fc6ba8.jpg) +(b) LLMs' Knowledge Boosts the Performance. +Figure 1. Demonstration of LLaMP: (a) LLMs can provide visual descriptions for fine-grained object categories; (b) Zero-shot baseto-novel generalization benefits from the LLM knowledge. + +e.g. GPT-4 [28] and LLaMA [38, 39], have demonstrated their encyclopedic knowledge and thus can provide linguistic visual descriptions for objects. Here, we investigate how to leverage LLMs for low-shot image classification. + +The emergence of prompt learning has provided an efficient way to adapt large pre-trained models. Previous work has explored various strategies to prompt vision-language (VL) models, including vision-conditioned text prompt learning [48], joint VL prompt learning [18] and self-regulated VL prompts[19]. On the text side, regardless of the learning strategy, learned prompt vectors are shared across all categories. The only difference among text inputs is the class name. In low-shot scenarios where visual data is limited, the extraction of class-specific knowledge from textual inputs becomes essential. However, the current paradigm, which relies on the CLIP text encoder to distinguish between class names, faces challenges, partic + +ularly with fine-grained target categories. For example, in FGVCAircraft [25], the class name "Yak-40", can barely provide any information for recognizing the object. + +Large Language Models, trained with large text corpora, are good candidates to serve as the complement. As in Fig. 1a, being queried about "Yak-40", the LLM generates a sentence detailing the visual appearance of the Yak-40 that can be further parsed into noun phrases and integrated into text prompts, providing richer information, compared with the ordinary prompt. We also show in Fig. 1b that by simply incorporating noun phrases extracted from a LLM's response, the performance of the ordinary CLIP models is improved by more than $1\%$ without any training. Although recent prompt-learning based methods have shown notable improvements, it is non-trivial to apply them on textual visual descriptions generated by LLMs. Thus, instead of directly taking LLM generations as the textual input, we aim at producing class-specific representations by adapting LLMs to low-shot image classification. + +One challenge of the adaption is the domain gap between vision and language. When trained exclusively with textual corpora, the latent feature space of a LLM significantly diverges from that of its visual counterpart. Even worse, the data scarcity under the low-shot scenario make it virtually impossible to align two spaces through plain contrastive loss. We argue that, the CLIP text encoder, which is trained to project features from the language domain into the joint VL domain, can serve as the bridge. Thus, we propose the LLaMP framework, Large Language Models as Prompt learners, which leverages LLMs to learn informative prompts for CLIP models. In LLaMP, we treat LLMs as the prompt learner of the CLIP text encoder. More specifically, for each object category, LLaMP extracts corresponding knowledge from the LLM and yields class-specific prompt vectors, which are further combined with class-agnostic prompt embeddings (as in previous approaches), and encoded by the CLIP text encoder. We design an efficient tuning pipeline to avoid fully fine-tuning the language model while performing effective adaptation. + +Following the protocol in [48, 49], we evaluate LLaMP with two typical scenarios: zero-shot base-to-novel generalization [49] and few-shot image classification. For each scenario, we run LLaMP with 11 datasets covering a spectrum of tasks. On average, LLaMP achieves a $1.3\%$ boost on the harmonic mean against the state-of-the-art PSRC [19], and $9.6\%$ over the ordinary CLIP [32], on base-to-novel generalization. We also observe an average improvement of $0.94\%$ on 16-shot image classification. + +In summary, our approach makes use of Large Language Models to improve performance in low-shot image classification scenarios. The main contributions are: i) To the best of our knowledge, we are the first to investigate how to use the encyclopedic knowledge inherent in Large Lan + +guage Models (LLMs) to enhance low-shot image classification; ii) We design a framework, LLaMP, to effectively adapt LLMs for image classification, without training the entire language model, and achieve state-of-the-art in both few-shot and zero-shot settings; iii) We conduct extensive analysis investigating the effectiveness of each components of LLaMP, and discuss the optimal setup for LLM-aided image classification. + +# 2. Related Work + +Large Language Models (LLMs). Recent years have witnessed remarkable progress in scaling up the size and capabilities of LLMs. Zhang et al. [47] first introduced a suite of transformers pre-trained at scale, followed by PaLM [6]. ChatGPT/GPT-4 [27, 28] emerged as a milestone conversational model, demonstrated impressive conversational abilities as a generalist model. Vicuna [5] further advanced by learning from ChatGPT, while LLaMA [38] demonstrates that larger scale training yields stronger foundation models. The subsequent LLaMA-2 [39] and PaLM-2 [2] achieved further gains in scale, efficiency and reasoning. Most recently, Almazrouei et al. [1] released Falcon, a 400B model. + +Zero-Shot Learning (ZSL). ZSL stands in contrast to traditional fully-supervised paradigms. Instead of relying on direct visual training samples, it leverages side information that can be drawn from a multitude of non-visual domains, including attributes [22], word embeddings [36, 40], and descriptive texts [34]. Zhang et al. [46] designed an embedding model to bridge the gap between seen and unseen categories. Concurrently, studies like [4, 44, 50] have spotlighted that generative models can produce features for unseen categories. Moreover, Graph Convolution Networks (GCN) [20] has been explored in research such as [16, 40] for further generalization. + +Prompt Learning. With the progress in large-scale vision-language models, such as CLIP [32] and ALIGN [14], which reveal their capacity in zero-shot transferability, prompt learning has emerged as an efficient learning scheme, where learnable prompts are appended to the input to fine-tune models. For low-shot image classification, CoOp [49] and CoCoOp [48], which modeled context words as learnable vectors to automate prompt engineering, have shown significant improvements over regular CLIP. MaPLe [18] further employed a hierarchical multi-modal prompting strategy across transformer blocks for progressive feature modeling. Kan et al. [17] incorporated external knowledge by designing knowledge-aware prompts and adaptation head for better generalization. Lee et al. [23] used masked attention to prevent internal representation shift for better generalization. Khattak et al. [19] further improved prompt learning by guiding prompts to balance task-specific and task-agnostic knowledge via mutual agreement maximization and prompt ensemble. + +# 3. Approach + +# 3.1. Preliminaries + +Similar to previous CLIP-based learning approaches, we consider the classification problem as an image-text matching problem. We denote the image encoder and the text encoder, in CLIP-like models, as $\mathcal{F}$ and $\mathcal{G}$ , parameterized by $\theta_{\mathcal{F}}$ and $\theta_{\mathcal{G}}$ , respectively. An input image $\boldsymbol{x} \in \mathbb{R}^{C \times H \times W}$ is split into $M$ equal-sized patches which are converted into a sequence of embeddings $\tilde{\boldsymbol{x}} = \{\boldsymbol{e}_{cls}, \boldsymbol{e}_1, \boldsymbol{e}_2, \dots, \boldsymbol{e}_M\}$ . The visual input sequence $\tilde{\boldsymbol{x}}$ is encoded by the image encoder, producing the image feature $\boldsymbol{f} = \mathcal{F}(\tilde{\boldsymbol{x}})$ . On the text side, the text label $y$ and the associated name is formatted as "A photo of [STH]" and tokenized into a sequence of tokens $\tilde{\boldsymbol{y}} = \{\boldsymbol{t}_{bos}, \boldsymbol{t}_1, \boldsymbol{t}_2, \dots, \boldsymbol{t}_L, \boldsymbol{t}_{eos}\}$ , where $L$ is the length of input tokens. The input sequence is then encoded into $\boldsymbol{g} = \mathcal{G}(\tilde{\boldsymbol{y}})$ . For image classification, target class labels $\{1, 2, \dots, C\}$ are encoded into text features $g_i$ . The classification is done by picking the class that has the highest similarity with the vision feature: $\hat{y} = \mathrm{argmax}_i \mathcal{C}(\boldsymbol{f}, \boldsymbol{g}_i)$ , where $\mathcal{C}$ is the softmax cosine-similarity function $\mathcal{C}(\boldsymbol{f}, \boldsymbol{g}) = \frac{\exp(\boldsymbol{f} \cdot \boldsymbol{g} / \tau)}{\sum_{i=1}^{C} \exp(\boldsymbol{f} \cdot \boldsymbol{g}_j / \tau)}$ with temperature $\tau$ . + +Multimodal Prompting Learning. Given the size of the CLIP model, fine-tuning the entire model becomes infeasible. As both image and text encoders are built with standard transformer architecture, prompt learning, which tunes the model by combining trainable prompts with hidden states has been applied on the text encoder [48, 49], the image encoder [15, 41, 42], or both [18, 19, 33]. Similar to [19, 33], we build our method following the vision-language prompting paradigm, with deep prompting [15, 19], which not only insert prompts to the input layer, but to later encoder layers. + +More specifically, for each transformer layer that takes prompts, we define $V$ learnable visual prompts $\pmb{p}_v = \{p_v^1, p_v^2, \dots, p_v^V\}$ and $T$ learnable language prompts $\pmb{p}_t = \{p_t^1, p_t^2, \dots, p_t^T\}$ . For the $i$ -th vision encoder layer, visual prompts $\pmb{p}_v^i$ are appended to input embeddings: $\tilde{\pmb{x}}_p^i = \{e_{cls}^i, e_1^i, e_2^i, \dots, e_M^i, p_v^i\}$ . The prompt-augmented vision feature, $\pmb{f}_p = \mathcal{F}(\tilde{\pmb{x}}_p)$ , is produced by jointly encoding prompts and the image. As the ViT [9] architecture in CLIP adopts the bi-directional attention mechanism, the placement of $\pmb{p}_v$ has no effect on $\pmb{f}_p$ . On the language side, prompts are concatenated with the input of the $i$ -th text encoder: $\tilde{\pmb{y}}_p^i = \{\pmb{t}_{bos}^i, \pmb{p}_t^i, \pmb{t}_1^i, \pmb{t}_2^i, \dots, \pmb{t}_L^i, \pmb{t}_{eos}^i\}$ . $\tilde{\pmb{y}}_p$ is further processed by the text encoder, resulting in the prompt-augmented language feature $\pmb{g}_p = \mathcal{G}(\tilde{\pmb{y}}_p)$ . More specifically, prompts to the first layer $\pmb{p}_t^1$ are initialized with the embeddings of "A photo of a". + +Low-Rank Adaptation [13] (LoRA). As a parameter-efficient tuning technique, LoRA is designed to adapt large transformer model without updating original model weights. The LoRA technique is, in particular, applied to + +linear projection layers. More specifically, for a linear layer with weight $W_0 \in \mathbb{R}^{d \times k}$ , LoRA creates $\Delta W$ by learning two low rank matrices $B \in \mathbb{R}^{d \times r}$ and $A \in \mathbb{R}^{r \times k}$ : + +$$ +\boldsymbol {h} = \left(W _ {0} + \Delta W\right) \boldsymbol {x} = W _ {0} \boldsymbol {x} + B A \boldsymbol {x}. \tag {1} +$$ + +We adopt a hybrid tuning scheme on the vision encoder, which performs prompt learning on the first few layers and applies LoRA on the rest. + +# 3.2. Adaptive Prompt Learning with LLMs + +The goal of prompt tuning is to find a set of optimal prompts $\pmb{p} = \{\pmb{p}_v, \pmb{p}_t\}$ which maximizes the log likelihood of $P(x, y | \theta_{\mathcal{F}}, \theta_{\mathcal{G}})$ over target downstream distribution $(x, y) \sim (X, Y)$ : + +$$ +\boldsymbol {p} = \underset {\boldsymbol {p}} {\operatorname {a r g m a x}} \mathbb {E} _ {(\boldsymbol {x}, \boldsymbol {y}) \sim (\boldsymbol {X}, \boldsymbol {Y})} \log \mathcal {C} \left(\mathcal {F} \left(\boldsymbol {x}; \boldsymbol {p} _ {v}\right), \mathcal {G} \left(\boldsymbol {y}; \boldsymbol {p} _ {t}\right)\right) \tag {2} +$$ + +However, the $p$ optimized through Eqn. 2 has two issues. First, $p$ is shared for all categories for the downstream task, while the optimal prompt for each category might be different. Second, in low-shot scenarios, $p$ are usually empirically estimated from a limited training-set $X^{train}$ with limited categories $\{1, 2, \dots, C^{base}\}$ , and therefore such $p$ can often be over-fitted to the small training-set $X^{train}$ and fail to generalize to novel categories outside $\{1, 2, \dots, C^{base}\}$ . + +To overcome these problems, we propose to learn a meta function on the language side $\pmb{p}_t = \Theta(y)$ which can adaptively estimate the optimal prompt for each category. An intuitive way to estimate proper prompts $\pmb{p}$ for category name $y$ is to take advantage of the knowledge of the pre-trained Large Language Models (LLM) $\mathcal{D}$ and extract discriminative descriptions of category $y$ . For example, given the input text $z: \text{"Describe {y}"}$ , + +$$ +\boldsymbol {p} _ {t} = \left\{p _ {1}, p _ {2}, \dots , p _ {k} \right\} = \mathcal {D} (\boldsymbol {z}). \tag {3} +$$ + +while $p_i$ being sequentially generated by $\mathcal{D}$ such that + +$$ +p _ {i} = \mathcal {D} (\boldsymbol {z}, t _ {1}, \dots , t _ {i - 1}) = \mathcal {D} ^ {(i)} (\boldsymbol {z}) \tag {4} +$$ + +$$ +t _ {i} = \mathcal {M} (p _ {i}), +$$ + +where $\mathcal{D}^{(i)}$ is the $i$ -th forward iteration of $\mathcal{D}$ , and $\mathcal{M}$ maps continuous hidden states into discrete language tokens. To accelerate the process and to obtain $p$ in one pass, we approximate the above process with $K$ learnable prompts $\pmb{p}_l = \{\theta_1,\dots,\theta_K\}$ so that + +$$ +\boldsymbol {p} _ {t} = \Theta (y) = \mathcal {D} \left(\left\{\theta_ {1}, \dots , \theta_ {K} \right\} | z\right) \tag {5} +$$ + +Discussion. While Large Language Models (LLMs) possess robust foundational knowledge within the linguistic domain, it is not feasible to directly substitute the text encoder of CLIP with an LLM. The reason lies in the inherent divergence between the LLM's latent space, which + +![](images/283906f7dd2e1c395cde2299a97a11fd24d0a477ca2588799710fb4f0a87cefd.jpg) +Figure 2. An Overview of the LLaMP Framework: We first generate the knowledge cache by passing the query prompt through the LLM $\mathcal{D}$ and use the knowledge cache to encode $\pmb{p}_l$ , resulting in the adaptive prompts $\tilde{\pmb{h}}_l^i = W\pmb{h}_l^i + b^i$ for the CLIP text encoder. $\tilde{\pmb{h}}_l$ is combined with regular learnable prompts of $\mathcal{G}$ to generate the final text feature vector $\pmb{g}_p$ . The image feature vector $\pmb{f}_p$ is obtained through a hybrid-tuning strategy combining prompt learning and low-rank adaptation (LoRA). + +is purely language-oriented, and the image-focused latent space of vision encoders. Attempting a direct alignment via contrastive learning would require an extensive dataset that is typically beyond the scope of low-shot learning. To bridge this gap, we introduce LLaMP—an adaptive prompt learning framework that leverages the LLM to craft class-specific prompt vectors, to reinforce the text encoder for low-shot image classification. + +# 3.3. The LLaMP Framework + +Fig. 2 shows an overview of the LLaMP framework. For convenience, we denote the decoder-only LLM as $\mathcal{D}$ . The input to the decoder $\mathcal{D}$ consists of two components: textual prompts $\pmb{y}$ in the form of sentences, tokenized as $\tilde{\pmb{y}}$ , and learnable prompts $\pmb{p}_l$ . We append prompt embeddings to the end of the input sequence and obtain the last hidden states of $\mathcal{D}$ as the feature $\pmb{h}_l$ : + +$$ +\boldsymbol {h} _ {l} = \mathcal {D} (\tilde {\boldsymbol {y}}, \boldsymbol {p} _ {l}) [ L + 1: L + K ], L = \operatorname {L e n g t h} (\tilde {\boldsymbol {y}}). \tag {6} +$$ + +Hidden states of $\mathcal{D}$ are then mapped to the input space of the CLIP text encoder by the projection matrix $W\in \mathbb{R}^{d_1\times d_2}$ , where $d_{1}$ and $d_{2}$ are respectively the hidden sizes of the LLM $\mathcal{D}$ and the CLIP text encoder $\mathcal{G}$ . A set of prompt-specific biases $b\in \mathbb{R}^{K\times d_2}$ are added: + +$$ +\tilde {\boldsymbol {h}} _ {l} = W \boldsymbol {h} _ {l} + b \tag {7} +$$ + +We combine $\tilde{h}_l$ from LLM with regular learnable prompts, as in previous approaches [19], to construct the input for CLIP text encoder. Similar to deep prompting [15, 19], we create layer-specific prompts through different $W$ matrices and $b$ vectors. For the $i$ -th layer, we let $\tilde{\pmb{h}}_l^i = W^i\pmb{h}_l + b^i$ and the entire sequence is constructed as + +$$ +\tilde {\boldsymbol {y}} _ {l} ^ {i} = \left\{\boldsymbol {t} _ {b o s} ^ {i}, \boldsymbol {p} _ {t} ^ {i}, \boldsymbol {t} _ {1} ^ {i}, \boldsymbol {t} _ {2} ^ {i}, \dots , \boldsymbol {t} _ {L} ^ {i}, \tilde {\boldsymbol {h}} _ {l} ^ {i}, \boldsymbol {t} _ {e o s} ^ {i} \right\} \tag {8} +$$ + +LLM Knowledge Cache. A Large Language Model (LLM), as implied by its name, typically comprises billions of parameters. For example, the most compact LLaMA [38, 39] model has 7B parameters. Thus, even performing prompt learning on a LLM become impractical. The memory consumption to store gradients for back propagation can go beyond the limit of mainstream GPUs. Instead, the causal attention mechanism inherent in decoder-only LLMs, where the embedding of an input token only depends on the preceding tokens, facilitates a feasible workaround. + +As previously mentioned, the prompt embeddings $\pmb{p}_l$ are appended to the end of text tokens $\tilde{\pmb{y}}$ . According to the causal attention mechanism, $\tilde{\pmb{y}}$ is encoded independently of $\pmb{p}_l$ . Thus, we design a two-stage process, where we create the LLM knowledge cache by passing $\tilde{\pmb{y}}$ through $\mathcal{D}$ and leverage the cache to convert $\pmb{p}_l$ into class-specific embeddings for the CLIP text encoder $\mathcal{G}$ . + +To compute the attention of a token, the only dependency is the Key and Value vectors from the preceding tokens. Thus, we adopt the KV-cache [31, 43], a technique used in inference acceleration of LLMs, to create the knowledge cache. At the first stage, we pass text tokens $\tilde{\pmb{y}}$ through the language model $\mathcal{D}$ and save the Keys and Values as the knowledge cache for the second stage. Once computed, the knowledge cache remains fixed throughout the entire training process and bears the information that is needed for further computation. Thus, in LLaMP, we leverage the knowledge cache obtained at the first stage to generate class-specific prompt embeddings. + +At the second stage, we create class-specific prompt embeddings from the pre-computed knowledge cache. As $\pmb{p}_l$ is not initialized in the natural language domain, it need not pass through the entire LLM; instead, we insert those prompts $\pmb{p}_l$ to the last layer of the LLM $\mathcal{D}_N$ . It is achieved + +by encoding them alongside the cache from $\tilde{\pmb{y}}$ , as in + +$$ +\boldsymbol {H} _ {l} = \mathcal {D} _ {N} \left(\boldsymbol {K} _ {\tilde {\boldsymbol {y}}}, \boldsymbol {V} _ {\tilde {\boldsymbol {y}}}, \boldsymbol {p} _ {l}\right), \tag {9} +$$ + +where $K_{\tilde{y}}, V_{\tilde{y}}$ represent the knowledge cache. This design enables LLaMP to efficiently learn informative prompt embeddings for the CLIP encoder $\mathcal{G}$ . It accomplishes this by incurring modest training costs, compared with training the entire LLM. Simultaneously, it maintains the essential knowledge inherent in the LLM decoder $\mathcal{D}$ . + +Training Targets of LLaMP. Although the training strategy in Eqn. 9 has reduced the number of learnable parameters, a full decoder inside a LLM still consists of an enormous number of parameters. For example, one layer in LLaMA-7B bears 200M parameters, making training of the entire layer costly. As the goal is to leverage the knowledge from LLM, altering a full layer can lead to the loss of knowledge. Thus, as shown in Fig. 2, a typical decoder layer has two major components: the self attention module, consisting Query, Key, Value and Output projection layers, and the Feed-Forward Network (FFN). LLaMP targets the Query and Output projection layer inside the self-attention module. By updating the Query layer, LLM prompts $p_l$ are learned to distill pertinent information from the knowledge cache and the Output layer projects it to the latent space. We keep the Key and Value layers frozen to ensure the alignment between $p_l$ and knowledge cache. We leave the FFN unchanged to preserve the knowledge. Further discussions regarding these choices are made in Sec. 4.3. + +Textual Priors from Pre-Generated Responses. We extend the initial prompt, "In one sentence, describe the distinctive appearance of [STH]", by incorporating the response generated by the language model into the input sequence. This approach enriches the base content: the generated text provides a clear and explicit description of the object's appearance, acting as a valuable informative prior for language model adaptation. However, it is common for responses from an LLM to include filler words like "sure" for sentence structure coherence. To refine the input, we parse the noun phrases from the LLM's response through spaCy [12], an NLP engine, and merge them with the initial prompt, forming a more focused and informative language prior. + +Textual Augmentations. Following the insights of Khattak et al. [19], which highlight the performance benefits of diverse textual inputs, we aim to further augment the text inputs used in the CLIP text encoder. Our approach, building upon the methods in [19, 48], incorporates handcrafted templates and expands their diversity through a two-step process: i) We introduce noun phrases into the existing templates for CLIP, for example, transforming "A photo of [STH]" to "A photo of [STH] with [NP]", thereby enriching the descriptive content; ii) We create a variety of new prompt templates for the LLM + +similar to “In one sentence, describe the distinctive appearance of [STH]” through GPT-4 [28], to further diversify the text input. + +# 3.4. Training and Inference + +Similar to PSRC [19], our objective function consists of three components: The main cross-entropy loss $\mathcal{L}_{CE}$ , feature-level L1 regularization $\mathcal{L}_{l1}$ , and soft distillation loss $\mathcal{L}_{dist}$ . Given $\mathcal{C}$ training categories and $\mathcal{N}$ training samples, $\mathcal{L}_{CE}$ is defined as + +$$ +\mathcal {L} _ {C E} = - \frac {1}{\mathcal {N}} \sum_ {i} \log \frac {\exp \left(\boldsymbol {f} _ {\boldsymbol {p}} ^ {i} \cdot \boldsymbol {g} _ {\boldsymbol {p}} ^ {i} / \tau\right)}{\sum \exp \left(\boldsymbol {f} _ {\boldsymbol {p}} ^ {i} \cdot \boldsymbol {g} _ {\boldsymbol {p}} ^ {j} / \tau\right)}. \tag {10} +$$ + +The L1 regularization is computed between learned features $\pmb{f}_p, \pmb{g}_p$ and pre-trained CLIP features $\hat{\pmb{f}}, \hat{\pmb{g}}$ : + +$$ +\mathcal {L} _ {l 1} = \frac {1}{\mathcal {N}} \sum_ {i} \lambda_ {v} \left| \boldsymbol {f} _ {\boldsymbol {p}} ^ {i} - \hat {\boldsymbol {f}} ^ {i} \right| + \frac {1}{C} \sum_ {i} \lambda_ {t} \left| \boldsymbol {g} _ {\boldsymbol {p}} ^ {i} - \hat {\boldsymbol {g}} ^ {i} \right|, \tag {11} +$$ + +where $\lambda_v$ and $\lambda_t$ are coefficients. The prediction of LLaMP is further bound by the KL-Divergence between predicted distributions of LLaMP and vanilla CLIP: + +$$ +\mathcal {L} _ {\text {d i s t}} = \lambda_ {\text {d i s t}} D _ {K L} \left(\boldsymbol {f} _ {\boldsymbol {p}} \cdot \boldsymbol {g} _ {\boldsymbol {p}}, \hat {\boldsymbol {f}} \cdot \hat {\boldsymbol {g}}\right). \tag {12} +$$ + +We sum all three losses up as the final objective function: $\mathcal{L} = \mathcal{L}_{CE} + \mathcal{L}_{l1} + \mathcal{L}_{dist}$ . + +During training, we randomly sample one LLM template as the input of LLaMP for each batch. For inference, we compute the probability distribution predicted from each input template and average them. + +# 4. Experiments + +# 4.1. Experiment Setup + +Datasets. Similar to previous work [18, 19, 48], in our study, we evaluate LLaMP performance over a spectrum of classification tasks with 11 datasets, including ImageNet [8] and Caltech101 [10] for generic image classification, OxfordPets [29], StanfordCars [21], Flowers102 [26], Food101 [3], and FGVCAircraft [25] for fine-grained classification, SUN397 [45] for scene recognition, UCF101 [37] for action recognition, DTD [7] for texture classification, and EuroSAT [11] for satellite image recognition. + +Scenarios & Metrics. We evaluate LLaMP on two typical low-shot scenarios: zero-shot base-to-novel generalization and few-shot image classification. In zero-shot base-to-novel generalization, the base classes are seen during training, while the novel classes are unseen. We measure models performance through accuracies of base and novel classes, and the harmonic mean of the two. For few-shot classification, we assess the accuracy with 16 shots per class. + +Implementation Details. We build LLaMP through the PyTorch [30] framework. All models are trained with + +
MethodAverageImageNet [8]Caltech101 [10]OxfordPets [29]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3474.2271.7072.4368.1470.2296.8494.0095.4091.1797.2694.12
CoOp [49]82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp [48]80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
KAPT* [17]78.4170.5274.2671.1065.2068.0297.1093.5395.2893.1396.5394.80
ProDA [24]81.5672.3076.6576.6670.5473.4797.7494.3696.0295.4397.7696.58
MaPLe [18]82.2875.1478.5575.4070.3272.7298.2793.2395.6895.4397.8396.62
RPO [23]81.1375.0077.7876.6071.5774.0097.9794.3796.0394.6397.5096.05
PSRC [19]84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
LLaMP85.1677.7181.2777.9971.2774.4898.4595.8597.1396.3197.7497.02
Δ w.r.t. PSRC+0.90+1.61+1.30+0.39+0.54+0.47+0.35+1.82+1.11+0.98+0.44+0.72
+ +
MethodStanfordCars [21]Flowers102[26]Food101 [3]FGVCAircraft [25]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]63.3774.8968.6572.0877.8074.8390.1091.2290.6627.1936.2931.09
CoOp [49]78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp [48]70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
KAPT* [17]69.4766.2067.7995.0071.2081.4086.1387.0686.5929.6728.7329.19
ProDA [24]72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
MaPLe [18]74.7071.2072.9197.7068.6880.6690.3088.5789.4336.9034.1335.46
RPO [23]73.8775.5374.6994.1376.6784.5090.3390.8390.5837.3334.2035.70
PSRC [19]78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
LLaMP81.5674.5477.8997.8277.4086.4291.0591.9391.4947.3037.6141.90
Δ w.r.t. PSRC+3.29-0.43+1.31-0.25+0.90+0.47+0.38+0.40+0.39+4.57-0.26+1.75
+ +
MethodSUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3675.3572.2353.2459.9056.3756.4864.0560.0370.5377.5073.85
CoOp [49]80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp [48]79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
KAPT* [17]79.4074.3376.7875.9758.3065.9784.8067.5775.2180.8367.1073.33
ProDA [24]80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
MaPLe [18]78.4776.9377.7980.6756.4866.4483.9066.0073.8885.2371.9778.04
RPO [23]80.6077.8079.1876.7062.1368.6186.6368.9776.7983.6775.4379.34
PSRC [19]82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
LLaMP83.4179.9081.6283.4964.4972.7791.9383.6687.6087.1380.6683.77
Δ w.r.t. PSRC+0.74+1.43+1.10+0.12+1.52+1.02-0.97+9.76+5.28+0.03+1.86+1.03
+ +Table 1. Comparison with state-of-the-art methods on base-to-novel generalization. LLaMP shows strong generalization results over previous approaches on 11 image classification tasks. Absolute gains over PSRC are indicated in blue. *KAPT is trained with ViT-B/32 image encoder instead of ViT-B/16. + +2 NVIDIA A100 40GB GPUs. For LLaMP, we adopt LLaMA2-7B [39] as the language model $\mathcal{D}$ , and ViT-B/16 [9] as the image encoder, following [18, 19, 48, 49]. On the text side, we set prompt learning depth to 9. To tune the vision encoder, we adopt the hybrid tuning scheme which performs deep prompt learning on the first 6 layers and LoRA on the rest. Similar to [13], LoRA is applied to the Query and Value projection layers inside attention modules. The number of $p_l$ prompts, $K$ , is set to 16. We set a global learning rate of 2E-4 with a batch size of 8. The learning + +rate of LoRA modules is set to 2E-5. $\lambda_{t},\lambda_{v}$ and $\lambda_{dist}$ are set to 25, 10 and 2.5, respectively. + +# 4.2. Quantitative Evaluation + +Zero-Shot Base-to-Novel Generalization. LLaMP outperforms existing state-of-the-art prompt learning methods on most metrics of 11 classification datasets in the baseto-novel generalization benchmark. As shown in Tab. 1, compared to the latest model PSRC [19], LLaMP achieves average gains of $0.90\%$ in base accuracy, $1.61\%$ in novel + +
16-Shot Classification
AverageImageNet [8]Collect [10]Pets [29]Cars [21]Flowers [26]Food [3]Aircraft [25]SUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
CLIP [32]78.79 (65.02)67.3195.4385.3480.4497.3782.9045.3673.2869.9687.2182.11
CoOp [49]79.89 (73.82)71.8795.5791.8783.0797.0784.2043.4074.6769.8784.9382.23
CoCoOp [48]74.90 (70.70)70.8395.1693.3471.5787.8487.2531.2172.1563.0473.3278.14
MaPLe [18]81.79 (75.58)72.3396.0092.8383.5797.0085.3348.4075.5371.3392.3385.03
PSRC [19]82.87 (77.90)73.1796.0793.6783.8397.6087.5050.8377.2372.7392.4386.47
LLaMP83.81 (78.50)73.4997.0894.2186.0798.0687.6256.0777.0274.1791.3186.84
+ +accuracy, and $1.30\%$ in harmonic mean on average. Moreover, LLaMP consistently achieves higher harmonic means (HM) compared to other models. These improvements indicate that our approach better balances performance on base and novel data, thus achieving stronger generalization compared to the prior prompt learning techniques. + +In particular, LLaMP excels in fine-grained datasets requiring detailed analysis. On FGVCAircraft, LLaMP surpasses PSRC by $4.57\%$ on base accuracy and $1.75\%$ on HM, highlighting its strong understanding of detailed aircraft features. Furthermore, on EuroSAT, LLaMP achieves improvements of $9.76\%$ and $5.28\%$ on novel accuracy and HM, respectively. We also observe similar performance gains on StanfordCars, where LLaMP outperforms by $3.29\%$ on base accuracy and $1.31\%$ on HM. The information embedded in LLM enables LLaMP to capture and utilize the rich semantic information necessary for distinguishing between closely related categories. + +Few-Shot Classification. LLaMP also achieves improvements across these classification datasets in few-shot classification tasks. As in Tab.2, with an average classification accuracy of $83.81\%$ . Notably, on FGVCAircraft and StanfordCars, LLaMP shows a significant improvement over PSRC, further demonstrating that the knowledge from language models benefits the recognition of fine-grained object categories, which aligns with our observation on zero-shot base-to-novel generalization. Moreover, on DTD, where MaPLe and PSRC achieve around $72\%$ accuracy, LLaMP achieves a higher accuracy of $74.17\%$ , underscoring its ability to recognize textures. + +# 4.3. Ablation Study + +Is the knowledge from LLM helping? In Tab. 3, we show that the knowledge from LLM benefits in both ways: Without training, performance of ordinary CLIP model can be improved by introducing noun phrases; The LLaMP framework shows further improvement after training. + +Noun phrases are parsed from the LLM's responses the prompt of "Describe [STH]". We then use the tem + +Table 2. Few shot classification results with 16 shots. Numbers in the bracket indicate the average performance over $1/2/4/8/16$ shots. + +
MethodLLMBaseNovelHM
CLIP✓69.3474.2271.70
70.9574.9372.79
LLaMP✓82.2176.4479.22
85.1677.7181.27
+ +Table 3. Ablation study on the LLM Knowledge. + +
LPQOKVFFN%BaseNovelHM
✓.0385.0077.2980.96
✓✓✓3385.2077.4581.14
✓✓✓8385.0577.7381.22
✓✓✓✓10085.2377.5681.21
✓✓1785.1677.7181.27
+ +Table 4. Ablation study on the Training Strategy. “%” indicates the ratio of parameters trained compared to fully tuning a layer. + +plate, "A photo of [STH] with [NP]" to generate the NP-augmented text embedding for CLIP. We take the average of all augmented embeddings for classification. In Tab. 3 we show that even ordinary CLIP can benefit from incorporating LLMs' knowledge. + +Furthermore, the comparison between LLaMP and LLaMP without the LLM indicates that merely integrating LoRA [13] to the vision encoder is not beneficial. The "LLaMP without LLM" is essentially an ordinary prompting learning model plus LoRAs in the vision encoder. We show that the improved vision encoding capacity only benefits when the quality of text embeddings s are enhanced by incorporating LLMs' knowledge through LLaMP. + +Decoder Training strategy. We categorize trainable parameters of $\mathcal{D}_N$ into four groups: learnable prompts (LP), Query and Output projections (QO), Key and Value projections (KV), and the feed-forward network (FFN). Tab. 4 indicates LLaMP can achieve desirable results by just learning the prompts of $\mathcal{D}$ . One step further, adding QO into optimization achieves the best performance. Although other + +
MethodPriorsBaseNovelHM
LLaMPX84.9077.5981.08
Plain85.2677.5681.22
NP85.1677.7181.27
+ +Table 5. Ablation Study on Pre-generated Text Priors. $\times$ refers to "without textual priors" and NP stands for noun phrases. + +
MethodBaseNovelHM
LLM Only81.7435.8249.81
LLaMP85.1677.7181.27
+ +Table 6. The CLIP text encoder helps adaptation. + +
SchemeBaseNovelHM
Prompt × 984.6777.2880.81
LoRA × 1284.8977.2780.90
Prompt ×6 + LoRA ×685.1677.7181.27
+ +Table 7. Study on Vision Tuning Scheme. Our hybrid design achieves the best performance. + +setups introduce much more trainable parameters, they can not surpass the "LP + QO" strategy. + +Effect of Textual Priors. We study the effect of pregenerated textual priors on LLaMP. We compare three different approaches: without textual priors, using plain responses as the prior, and LLaMP, which takes parsed noun phrases. Tab. 5 shows that LLaMP can achieve over $81\%$ on HM without pre-generated priors, while adding parsed noun phrases as textual priors further pushes the HM to $81.27\%$ . + +CLIP as the bridge. One may wonder if it is possible to replace CLIP text encoder with a large language model. Here, we study two setups: i) LLM as encoder, which treats the output of the language model, $\tilde{h}_l$ as the text feature; ii) LLaMP, which treat $\tilde{h}_l$ as part of the text input prompt. Tab. 6 reveals that relying solely on the LLM results in poor accuracy for novel categories. This supports our hypothesis that aligning LLMs with vision encoders generally requires a more extensive dataset. Furthermore, LLaMP's design significantly improves the novel accuracy by $40\%$ . + +Vision Training Strategy. As ViT-16/B has 12 transformer layers, we compare different vision training strategies within LLaMP in Tab. 7. Apart from the default hybrid scheme, we evaluate setups including prompt learning at first 9 layers (P9), a similar setup to PSRC [19], and LoRA [13] in all 12 layers. The results suggest that the scheme leverages the strengths of both prompt learning and LoRA, addressing potential bottlenecks in the vision encoder and enhancing overall performance in LLaMP. + +Number of LLM Prompts. We vary the number of LLM prompts and study their effects on LLaMP. As in + +![](images/5eb53d774d0e54e537515c705435c6df8314a142113bfec24752075d117f18bc.jpg) +Figure 3. Effect of LLM Prompts on Harmonic Mean. 16 prompts achieve the most balanced performance. + +![](images/e434f605bf4506380301043f978de0813f9f16be30e16f768ac6c04911f40740.jpg) +Figure 4. Visualization of LLaMP Predictions by GradCAM [35] Fig. 3, using 16 prompts optimizes the LLM's capabilities, achieving the highest harmonic mean at $81.27\%$ . + +Visualizations. In Fig. 4, we visualize the gradient heatmap of input images from FGVCAircraft and EuroSAT, through GradCAM [35]. The figure shows that, LLaMP can capture distinctive features that matches LLM's description. + +# 5. Conclusion + +Our study shows that the encyclopedic knowledge from LLMs is beneficial for low-shot image classification as extra class-specific information. To leverage such knowledge, we propose LLaMP, a framework that adapts LLMs as prompt learners for the CLIP model. Over two common low-shot scenarios: zero-shot generalization and few-shot learning, LLaMP demonstrates notable improvements compared with previous state-of-the-arts on a spectrum of datasets. + +Limitations. While LLaMP reveals an effective way in leveraging LLMs' knowledge, both modalities, vision and language, only interact at the finest feature level. Given the broader LLM-aided knowledge from the language side, the performance can be potentially further improved by introducing language priors at earlier vision encoding stages. + +# Acknowledgment + +This research was supported, in part, by the Office of Naval Research under grant #N00014-21-1-2802. + +# References + +[1] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, et al. Falcon-40b: an open large language model with state-of-the-art performance. Technical report, Technical report, Technology Innovation Institute, 2023. 2 +[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2 +[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461. Springer, 2014. 5, 6, 7 +[4] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In ICCV, pages 122-131, 2021. 2 +[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 2 +[6] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2 +[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5, 6, 7 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5, 6, 7 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 6 +[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR workshop, pages 178-178. IEEE, 2004. 5, 6, 7 +[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 6, 7 +[12] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 5 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 6, 7, 8 + +[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2 +[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4 +[16] Michael Kampffmeyer, Yinbo Chen, Xiaodan Liang, Hao Wang, Yujia Zhang, and Eric P Xing. Rethinking knowledge graph propagation for zero-shot learning. In CVPR, pages 11487-11496, 2019. 2 +[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 2, 6 +[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 5, 6, 7 +[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[20] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th ICLR, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 2 +[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 5, 6, 7 +[22] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. TPAMI, 36(3):453-465, 2013. 2 +[23] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 2, 6 +[24] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 6 +[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 5, 6, 7 +[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5, 6, 7 +[27] OpenAI. Chatgpt, 2023. Available at https://openai.com/chatgpt.2 +[28] OpenAI. Gpt-4 technical report, 2023. 1, 2, 5 +[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In $CVPR$ , pages 3498-3505. IEEE, 2012. 5, 6, 7 + +[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, pages 8026-8037, 2019. 5 +[31] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5, 2023. 4 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7 +[33] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 3 +[34] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In CVPR, pages 49-58, 2016. 2 +[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626, 2017. 8 +[36] Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D Manning, and Andrew Y Ng. Zero-shot learning through cross-modal transfer. arXiv preprint arXiv:1301.3666, 2013. 2 +[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7 +[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 4 +[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 4, 6 +[40] Xiaolong Wang, Yufei Ye, and Abhinav Gupta. Zero-shot recognition via semantic embeddings and knowledge graphs. In CVPR, pages 6857-6866, 2018. 2 +[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 3 +[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer + +Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3 +[43] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online, 2020. Association for Computational Linguistics. 4 +[44] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In CVPR, pages 5542-5551, 2018. 2 +[45] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. IJCV, 119(1):3-22, 2016. 5, 6, 7 +[46] Li Zhang, Tao Xiang, and Shaogang Gong. Learning a deep embedding model for zero-shot learning. In CVPR, pages 2021-2030, 2017. 2 +[47] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2 +[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 5, 6, 7 +[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7 +[50] Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR, pages 1004-1013, 2018. 2 \ No newline at end of file diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/images.zip b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..457826cc1a0c75fe98fff4d24df70152dda7731a --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c30de2d07697e7aaa3ac8d4e45910cb816d0a172addd82186173b9904f4411f1 +size 736709 diff --git a/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/layout.json b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2c430e3e0b7e7db3e44dfbf6297c47014f30fff --- /dev/null +++ b/2024/Large Language Models are Good Prompt Learners for Low-Shot Image Classification/layout.json @@ -0,0 +1,9626 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 459, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 459, + 138 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 459, + 138 + ], + "type": "text", + "content": "Large Language Models are Good Prompt Learners for Low-Shot Image Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 161, + 492, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 161, + 492, + 203 + ], + "spans": [ + { + "bbox": [ + 101, + 161, + 492, + 203 + ], + "type": "text", + "content": "Zhaoheng Zheng Jingmin Wei Xuefeng Hu Haidong Zhu Ram Nevatia Viterbi School of Engineering University of Southern California" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 205, + 465, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 205, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 125, + 205, + 465, + 217 + ], + "type": "text", + "content": "{zhaoheng.zheng, jingminw, xuefengh, haidongz, nevatia}@usc.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 245, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 271, + 290, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 510 + ], + "type": "text", + "content": "Low-shot image classification, where training images are limited or inaccessible, has benefited from recent progress on pre-trained vision-language (VL) models with strong generalizability, e.g. CLIP. Prompt learning methods built with VL models generate text features from the class names that only have confined class-specific information. Large Language Models (LLMs), with their vast encyclopedic knowledge, emerge as the complement. Thus, in this paper, we discuss the integration of LLMs to enhance pre-trained VL models, specifically on low-shot classification. However, the domain gap between language and vision blocks the direct application of LLMs. Thus, we propose LLaMP, Large Language Models as Prompt learners, that produces adaptive prompts for the CLIP text encoder, establishing it as the connecting bridge. Experiments show that, compared with other state-of-the-art prompt learning methods, LLaMP yields better performance on both zero-shot generalization and few-shot image classification, over a spectrum of 11 datasets. Code will be made available at: https://github.com/zhaohengz/LLaMP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 128, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "Low-shot image classification tasks, including few-shot and zero-shot variants, are to learn from a set of class names along with a limited or null set of images. Such capacities are crucial for the extension and generalization of vision systems. Vision-Language (VL) models trained on large-scale web data, such as CLIP [32] and ALIGN [14], provide a new paradigm due to their generalization capabilities that includes zero-shot classification, and have been used in recent work [17-19, 23, 24, 48, 49]. Due to the scarcity of images for training, methods built for both tasks rely heavily on merely category names as the source of class-specific knowledge, resulting in a shortage of distinguishable descriptions. Meanwhile, Large Language Models (LLMs)," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 245, + 544, + 316 + ], + "blocks": [ + { + "bbox": [ + 310, + 245, + 544, + 316 + ], + "lines": [ + { + "bbox": [ + 310, + 245, + 544, + 316 + ], + "spans": [ + { + "bbox": [ + 310, + 245, + 544, + 316 + ], + "type": "image", + "image_path": "13ba31720a95e4f52c8efd8e4be9287dc1acb43ea3910e2996b90dc8748222a6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 320, + 488, + 331 + ], + "lines": [ + { + "bbox": [ + 364, + 320, + 488, + 331 + ], + "spans": [ + { + "bbox": [ + 364, + 320, + 488, + 331 + ], + "type": "text", + "content": "(a) Visual Descriptions from an LLM." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 334, + 543, + 435 + ], + "blocks": [ + { + "bbox": [ + 310, + 334, + 543, + 435 + ], + "lines": [ + { + "bbox": [ + 310, + 334, + 543, + 435 + ], + "spans": [ + { + "bbox": [ + 310, + 334, + 543, + 435 + ], + "type": "image", + "image_path": "44ef9588350c8f3b111c503448edbde37cb3f6386a248e4c5a54c4b4c7fc6ba8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 438, + 504, + 449 + ], + "lines": [ + { + "bbox": [ + 348, + 438, + 504, + 449 + ], + "spans": [ + { + "bbox": [ + 348, + 438, + 504, + 449 + ], + "type": "text", + "content": "(b) LLMs' Knowledge Boosts the Performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 458, + 546, + 492 + ], + "lines": [ + { + "bbox": [ + 305, + 458, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 458, + 546, + 492 + ], + "type": "text", + "content": "Figure 1. Demonstration of LLaMP: (a) LLMs can provide visual descriptions for fine-grained object categories; (b) Zero-shot baseto-novel generalization benefits from the LLM knowledge." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 507, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 547, + 556 + ], + "type": "text", + "content": "e.g. GPT-4 [28] and LLaMA [38, 39], have demonstrated their encyclopedic knowledge and thus can provide linguistic visual descriptions for objects. Here, we investigate how to leverage LLMs for low-shot image classification." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 715 + ], + "type": "text", + "content": "The emergence of prompt learning has provided an efficient way to adapt large pre-trained models. Previous work has explored various strategies to prompt vision-language (VL) models, including vision-conditioned text prompt learning [48], joint VL prompt learning [18] and self-regulated VL prompts[19]. On the text side, regardless of the learning strategy, learned prompt vectors are shared across all categories. The only difference among text inputs is the class name. In low-shot scenarios where visual data is limited, the extraction of class-specific knowledge from textual inputs becomes essential. However, the current paradigm, which relies on the CLIP text encoder to distinguish between class names, faces challenges, partic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28453" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "ularly with fine-grained target categories. For example, in FGVCAircraft [25], the class name \"Yak-40\", can barely provide any information for recognizing the object." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 287, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 301 + ], + "type": "text", + "content": "Large Language Models, trained with large text corpora, are good candidates to serve as the complement. As in Fig. 1a, being queried about \"Yak-40\", the LLM generates a sentence detailing the visual appearance of the Yak-40 that can be further parsed into noun phrases and integrated into text prompts, providing richer information, compared with the ordinary prompt. We also show in Fig. 1b that by simply incorporating noun phrases extracted from a LLM's response, the performance of the ordinary CLIP models is improved by more than " + }, + { + "bbox": [ + 46, + 110, + 287, + 301 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 110, + 287, + 301 + ], + "type": "text", + "content": " without any training. Although recent prompt-learning based methods have shown notable improvements, it is non-trivial to apply them on textual visual descriptions generated by LLMs. Thus, instead of directly taking LLM generations as the textual input, we aim at producing class-specific representations by adapting LLMs to low-shot image classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 303, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 287, + 542 + ], + "type": "text", + "content": "One challenge of the adaption is the domain gap between vision and language. When trained exclusively with textual corpora, the latent feature space of a LLM significantly diverges from that of its visual counterpart. Even worse, the data scarcity under the low-shot scenario make it virtually impossible to align two spaces through plain contrastive loss. We argue that, the CLIP text encoder, which is trained to project features from the language domain into the joint VL domain, can serve as the bridge. Thus, we propose the LLaMP framework, Large Language Models as Prompt learners, which leverages LLMs to learn informative prompts for CLIP models. In LLaMP, we treat LLMs as the prompt learner of the CLIP text encoder. More specifically, for each object category, LLaMP extracts corresponding knowledge from the LLM and yields class-specific prompt vectors, which are further combined with class-agnostic prompt embeddings (as in previous approaches), and encoded by the CLIP text encoder. We design an efficient tuning pipeline to avoid fully fine-tuning the language model while performing effective adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "Following the protocol in [48, 49], we evaluate LLaMP with two typical scenarios: zero-shot base-to-novel generalization [49] and few-shot image classification. For each scenario, we run LLaMP with 11 datasets covering a spectrum of tasks. On average, LLaMP achieves a " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " boost on the harmonic mean against the state-of-the-art PSRC [19], and " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "9.6\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " over the ordinary CLIP [32], on base-to-novel generalization. We also observe an average improvement of " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "0.94\\%" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " on 16-shot image classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "In summary, our approach makes use of Large Language Models to improve performance in low-shot image classification scenarios. The main contributions are: i) To the best of our knowledge, we are the first to investigate how to use the encyclopedic knowledge inherent in Large Lan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "guage Models (LLMs) to enhance low-shot image classification; ii) We design a framework, LLaMP, to effectively adapt LLMs for image classification, without training the entire language model, and achieve state-of-the-art in both few-shot and zero-shot settings; iii) We conduct extensive analysis investigating the effectiveness of each components of LLaMP, and discuss the optimal setup for LLM-aided image classification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "spans": [ + { + "bbox": [ + 306, + 178, + 392, + 192 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 199, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 199, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 545, + 342 + ], + "type": "text", + "content": "Large Language Models (LLMs). Recent years have witnessed remarkable progress in scaling up the size and capabilities of LLMs. Zhang et al. [47] first introduced a suite of transformers pre-trained at scale, followed by PaLM [6]. ChatGPT/GPT-4 [27, 28] emerged as a milestone conversational model, demonstrated impressive conversational abilities as a generalist model. Vicuna [5] further advanced by learning from ChatGPT, while LLaMA [38] demonstrates that larger scale training yields stronger foundation models. The subsequent LLaMA-2 [39] and PaLM-2 [2] achieved further gains in scale, efficiency and reasoning. Most recently, Almazrouei et al. [1] released Falcon, a 400B model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": "Zero-Shot Learning (ZSL). ZSL stands in contrast to traditional fully-supervised paradigms. Instead of relying on direct visual training samples, it leverages side information that can be drawn from a multitude of non-visual domains, including attributes [22], word embeddings [36, 40], and descriptive texts [34]. Zhang et al. [46] designed an embedding model to bridge the gap between seen and unseen categories. Concurrently, studies like [4, 44, 50] have spotlighted that generative models can produce features for unseen categories. Moreover, Graph Convolution Networks (GCN) [20] has been explored in research such as [16, 40] for further generalization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Prompt Learning. With the progress in large-scale vision-language models, such as CLIP [32] and ALIGN [14], which reveal their capacity in zero-shot transferability, prompt learning has emerged as an efficient learning scheme, where learnable prompts are appended to the input to fine-tune models. For low-shot image classification, CoOp [49] and CoCoOp [48], which modeled context words as learnable vectors to automate prompt engineering, have shown significant improvements over regular CLIP. MaPLe [18] further employed a hierarchical multi-modal prompting strategy across transformer blocks for progressive feature modeling. Kan et al. [17] incorporated external knowledge by designing knowledge-aware prompts and adaptation head for better generalization. Lee et al. [23] used masked attention to prevent internal representation shift for better generalization. Khattak et al. [19] further improved prompt learning by guiding prompts to balance task-specific and task-agnostic knowledge via mutual agreement maximization and prompt ensemble." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28454" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 113, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 113, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 113, + 85 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 90, + 133, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 90, + 133, + 102 + ], + "spans": [ + { + "bbox": [ + 47, + 90, + 133, + 102 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": "Similar to previous CLIP-based learning approaches, we consider the classification problem as an image-text matching problem. We denote the image encoder and the text encoder, in CLIP-like models, as " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{F}}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ", respectively. An input image " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " is split into " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " equal-sized patches which are converted into a sequence of embeddings " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}} = \\{\\boldsymbol{e}_{cls}, \\boldsymbol{e}_1, \\boldsymbol{e}_2, \\dots, \\boldsymbol{e}_M\\}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ". The visual input sequence " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{x}}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " is encoded by the image encoder, producing the image feature " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\boldsymbol{f} = \\mathcal{F}(\\tilde{\\boldsymbol{x}})" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ". On the text side, the text label " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " and the associated name is formatted as \"A photo of [STH]\" and tokenized into a sequence of tokens " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{y}} = \\{\\boldsymbol{t}_{bos}, \\boldsymbol{t}_1, \\boldsymbol{t}_2, \\dots, \\boldsymbol{t}_L, \\boldsymbol{t}_{eos}\\}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " is the length of input tokens. The input sequence is then encoded into " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\boldsymbol{g} = \\mathcal{G}(\\tilde{\\boldsymbol{y}})" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ". For image classification, target class labels " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\{1, 2, \\dots, C\\}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " are encoded into text features " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ". The classification is done by picking the class that has the highest similarity with the vision feature: " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\mathrm{argmax}_i \\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}_i)" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " is the softmax cosine-similarity function " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\boldsymbol{f}, \\boldsymbol{g}) = \\frac{\\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g} / \\tau)}{\\sum_{i=1}^{C} \\exp(\\boldsymbol{f} \\cdot \\boldsymbol{g}_j / \\tau)}" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": " with temperature " + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 109, + 289, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 342, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 472 + ], + "type": "text", + "content": "Multimodal Prompting Learning. Given the size of the CLIP model, fine-tuning the entire model becomes infeasible. As both image and text encoders are built with standard transformer architecture, prompt learning, which tunes the model by combining trainable prompts with hidden states has been applied on the text encoder [48, 49], the image encoder [15, 41, 42], or both [18, 19, 33]. Similar to [19, 33], we build our method following the vision-language prompting paradigm, with deep prompting [15, 19], which not only insert prompts to the input layer, but to later encoder layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": "More specifically, for each transformer layer that takes prompts, we define " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " learnable visual prompts " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{p}_v = \\{p_v^1, p_v^2, \\dots, p_v^V\\}" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " learnable language prompts " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t = \\{p_t^1, p_t^2, \\dots, p_t^T\\}" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": "-th vision encoder layer, visual prompts " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{p}_v^i" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " are appended to input embeddings: " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_p^i = \\{e_{cls}^i, e_1^i, e_2^i, \\dots, e_M^i, p_v^i\\}" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ". The prompt-augmented vision feature, " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p = \\mathcal{F}(\\tilde{\\pmb{x}}_p)" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ", is produced by jointly encoding prompts and the image. As the ViT [9] architecture in CLIP adopts the bi-directional attention mechanism, the placement of " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{p}_v" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " has no effect on " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ". On the language side, prompts are concatenated with the input of the " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": "-th text encoder: " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}_p^i = \\{\\pmb{t}_{bos}^i, \\pmb{p}_t^i, \\pmb{t}_1^i, \\pmb{t}_2^i, \\dots, \\pmb{t}_L^i, \\pmb{t}_{eos}^i\\}" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}_p" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " is further processed by the text encoder, resulting in the prompt-augmented language feature " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{g}_p = \\mathcal{G}(\\tilde{\\pmb{y}}_p)" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": ". More specifically, prompts to the first layer " + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t^1" + }, + { + "bbox": [ + 47, + 473, + 289, + 665 + ], + "type": "text", + "content": " are initialized with the embeddings of \"A photo of a\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Low-Rank Adaptation [13] (LoRA). As a parameter-efficient tuning technique, LoRA is designed to adapt large transformer model without updating original model weights. The LoRA technique is, in particular, applied to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "linear projection layers. More specifically, for a linear layer with weight " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times k}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ", LoRA creates " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " by learning two low rank matrices " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{d \\times r}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times k}" + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 348, + 117, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 117, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 348, + 117, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} = \\left(W _ {0} + \\Delta W\\right) \\boldsymbol {x} = W _ {0} \\boldsymbol {x} + B A \\boldsymbol {x}. \\tag {1}", + "image_path": "a541df1ba4f8aa31b1cf3f09dd088612ce1e3974c46acde4e46734b9effce86f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 305, + 137, + 545, + 173 + ], + "type": "text", + "content": "We adopt a hybrid tuning scheme on the vision encoder, which performs prompt learning on the first few layers and applies LoRA on the rest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 180, + 511, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 511, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 511, + 193 + ], + "type": "text", + "content": "3.2. Adaptive Prompt Learning with LLMs" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "text", + "content": "The goal of prompt tuning is to find a set of optimal prompts " + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\{\\pmb{p}_v, \\pmb{p}_t\\}" + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "text", + "content": " which maximizes the log likelihood of " + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "inline_equation", + "content": "P(x, y | \\theta_{\\mathcal{F}}, \\theta_{\\mathcal{G}})" + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "text", + "content": " over target downstream distribution " + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "inline_equation", + "content": "(x, y) \\sim (X, Y)" + }, + { + "bbox": [ + 305, + 198, + 545, + 247 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 255, + 545, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 255, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 545, + 285 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} = \\underset {\\boldsymbol {p}} {\\operatorname {a r g m a x}} \\mathbb {E} _ {(\\boldsymbol {x}, \\boldsymbol {y}) \\sim (\\boldsymbol {X}, \\boldsymbol {Y})} \\log \\mathcal {C} \\left(\\mathcal {F} \\left(\\boldsymbol {x}; \\boldsymbol {p} _ {v}\\right), \\mathcal {G} \\left(\\boldsymbol {y}; \\boldsymbol {p} _ {t}\\right)\\right) \\tag {2}", + "image_path": "52300915ac6a867fc937049351c1dffe5a7b6b0033a939c1f554a9778f773307.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": "However, the " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " optimized through Eqn. 2 has two issues. First, " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " is shared for all categories for the downstream task, while the optimal prompt for each category might be different. Second, in low-shot scenarios, " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " are usually empirically estimated from a limited training-set " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "X^{train}" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " with limited categories " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\{1, 2, \\dots, C^{base}\\}" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": ", and therefore such " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " can often be over-fitted to the small training-set " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "X^{train}" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": " and fail to generalize to novel categories outside " + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\{1, 2, \\dots, C^{base}\\}" + }, + { + "bbox": [ + 304, + 285, + 545, + 380 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": "To overcome these problems, we propose to learn a meta function on the language side " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{p}_t = \\Theta(y)" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": " which can adaptively estimate the optimal prompt for each category. An intuitive way to estimate proper prompts " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": " for category name " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": " is to take advantage of the knowledge of the pre-trained Large Language Models (LLM) " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": " and extract discriminative descriptions of category " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": ". For example, given the input text " + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "inline_equation", + "content": "z: \\text{\"Describe {y}\"}" + }, + { + "bbox": [ + 304, + 381, + 545, + 477 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 363, + 485, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 485, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 363, + 485, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} = \\left\\{p _ {1}, p _ {2}, \\dots , p _ {k} \\right\\} = \\mathcal {D} (\\boldsymbol {z}). \\tag {3}", + "image_path": "a8e1662c503a1e4ef26f20d58e0832961ce7484de7163002acf0321f3ef02c11.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "text", + "content": "while " + }, + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "text", + "content": " being sequentially generated by " + }, + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 305, + 506, + 520, + 518 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 356, + 525, + 545, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 525, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 356, + 525, + 545, + 545 + ], + "type": "interline_equation", + "content": "p _ {i} = \\mathcal {D} (\\boldsymbol {z}, t _ {1}, \\dots , t _ {i - 1}) = \\mathcal {D} ^ {(i)} (\\boldsymbol {z}) \\tag {4}", + "image_path": "bc4db135b9a308636c2836b9f20cb9674442c823f56be9992489709b6887c0f5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 544, + 411, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 544, + 411, + 555 + ], + "spans": [ + { + "bbox": [ + 359, + 544, + 411, + 555 + ], + "type": "interline_equation", + "content": "t _ {i} = \\mathcal {M} (p _ {i}),", + "image_path": "57ffffb86fc45278c951f439567070dc4874864ab9267e9c455c480c702c1884.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{(i)}" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": "-th forward iteration of " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": " maps continuous hidden states into discrete language tokens. To accelerate the process and to obtain " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": " in one pass, we approximate the above process with " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": " learnable prompts " + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l = \\{\\theta_1,\\dots,\\theta_K\\}" + }, + { + "bbox": [ + 304, + 563, + 545, + 624 + ], + "type": "text", + "content": " so that" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 358, + 632, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 632, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 358, + 632, + 545, + 646 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {t} = \\Theta (y) = \\mathcal {D} \\left(\\left\\{\\theta_ {1}, \\dots , \\theta_ {K} \\right\\} | z\\right) \\tag {5}", + "image_path": "7314ef7b5e37f39281ace17a0af9b5100e9d8ab82ae48f2bb74a820d065134d1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Discussion. While Large Language Models (LLMs) possess robust foundational knowledge within the linguistic domain, it is not feasible to directly substitute the text encoder of CLIP with an LLM. The reason lies in the inherent divergence between the LLM's latent space, which" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28455" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 545, + 240 + ], + "type": "image", + "image_path": "283906f7dd2e1c395cde2299a97a11fd24d0a477ca2588799710fb4f0a87cefd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": "Figure 2. An Overview of the LLaMP Framework: We first generate the knowledge cache by passing the query prompt through the LLM " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": " and use the knowledge cache to encode " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": ", resulting in the adaptive prompts " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}_l^i = W\\pmb{h}_l^i + b^i" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": " for the CLIP text encoder. " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}_l" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": " is combined with regular learnable prompts of " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": " to generate the final text feature vector " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{g}_p" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": ". The image feature vector " + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 46, + 249, + 547, + 293 + ], + "type": "text", + "content": " is obtained through a hybrid-tuning strategy combining prompt learning and low-rank adaptation (LoRA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 396 + ], + "type": "text", + "content": "is purely language-oriented, and the image-focused latent space of vision encoders. Attempting a direct alignment via contrastive learning would require an extensive dataset that is typically beyond the scope of low-shot learning. To bridge this gap, we introduce LLaMP—an adaptive prompt learning framework that leverages the LLM to craft class-specific prompt vectors, to reinforce the text encoder for low-shot image classification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 402, + 186, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 186, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 186, + 414 + ], + "type": "text", + "content": "3.3. The LLaMP Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": "Fig. 2 shows an overview of the LLaMP framework. For convenience, we denote the decoder-only LLM as " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": ". The input to the decoder " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": " consists of two components: textual prompts " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": " in the form of sentences, tokenized as " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": ", and learnable prompts " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": ". We append prompt embeddings to the end of the input sequence and obtain the last hidden states of " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": " as the feature " + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "inline_equation", + "content": "\\pmb{h}_l" + }, + { + "bbox": [ + 46, + 420, + 287, + 504 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 511, + 287, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 287, + 526 + ], + "type": "interline_equation", + "content": "\\boldsymbol {h} _ {l} = \\mathcal {D} (\\tilde {\\boldsymbol {y}}, \\boldsymbol {p} _ {l}) [ L + 1: L + K ], L = \\operatorname {L e n g t h} (\\tilde {\\boldsymbol {y}}). \\tag {6}", + "image_path": "6314563c62f56cf6260bbc81f4825b2b9f331aec776d5469c375a9f1fc66c61e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": "Hidden states of " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": " are then mapped to the input space of the CLIP text encoder by the projection matrix " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "W\\in \\mathbb{R}^{d_1\\times d_2}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "d_{1}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "d_{2}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": " are respectively the hidden sizes of the LLM " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": " and the CLIP text encoder " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": ". A set of prompt-specific biases " + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "inline_equation", + "content": "b\\in \\mathbb{R}^{K\\times d_2}" + }, + { + "bbox": [ + 46, + 532, + 287, + 592 + ], + "type": "text", + "content": " are added:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 598, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 598, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 135, + 598, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {h}} _ {l} = W \\boldsymbol {h} _ {l} + b \\tag {7}", + "image_path": "083e0dddc0f7f78279573f5d20e32a0b209a740f2316eae1425ad009e04efad2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": "We combine " + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\tilde{h}_l" + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": " from LLM with regular learnable prompts, as in previous approaches [19], to construct the input for CLIP text encoder. Similar to deep prompting [15, 19], we create layer-specific prompts through different " + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": " matrices and " + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": " vectors. For the " + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": "-th layer, we let " + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{h}}_l^i = W^i\\pmb{h}_l + b^i" + }, + { + "bbox": [ + 46, + 619, + 287, + 693 + ], + "type": "text", + "content": " and the entire sequence is constructed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 86, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {y}} _ {l} ^ {i} = \\left\\{\\boldsymbol {t} _ {b o s} ^ {i}, \\boldsymbol {p} _ {t} ^ {i}, \\boldsymbol {t} _ {1} ^ {i}, \\boldsymbol {t} _ {2} ^ {i}, \\dots , \\boldsymbol {t} _ {L} ^ {i}, \\tilde {\\boldsymbol {h}} _ {l} ^ {i}, \\boldsymbol {t} _ {e o s} ^ {i} \\right\\} \\tag {8}", + "image_path": "b16d7979d833639c49806d3e04e076ec89daaff8ad4b11a459b2b798de1326bc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 299, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 418 + ], + "type": "text", + "content": "LLM Knowledge Cache. A Large Language Model (LLM), as implied by its name, typically comprises billions of parameters. For example, the most compact LLaMA [38, 39] model has 7B parameters. Thus, even performing prompt learning on a LLM become impractical. The memory consumption to store gradients for back propagation can go beyond the limit of mainstream GPUs. Instead, the causal attention mechanism inherent in decoder-only LLMs, where the embedding of an input token only depends on the preceding tokens, facilitates a feasible workaround." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": "As previously mentioned, the prompt embeddings " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": " are appended to the end of text tokens " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": ". According to the causal attention mechanism, " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": " is encoded independently of " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": ". Thus, we design a two-stage process, where we create the LLM knowledge cache by passing " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": " and leverage the cache to convert " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": " into class-specific embeddings for the CLIP text encoder " + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 422, + 546, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": "To compute the attention of a token, the only dependency is the Key and Value vectors from the preceding tokens. Thus, we adopt the KV-cache [31, 43], a technique used in inference acceleration of LLMs, to create the knowledge cache. At the first stage, we pass text tokens " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " through the language model " + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 508, + 545, + 651 + ], + "type": "text", + "content": " and save the Keys and Values as the knowledge cache for the second stage. Once computed, the knowledge cache remains fixed throughout the entire training process and bears the information that is needed for further computation. Thus, in LLaMP, we leverage the knowledge cache obtained at the first stage to generate class-specific prompt embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "At the second stage, we create class-specific prompt embeddings from the pre-computed knowledge cache. As " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " is not initialized in the natural language domain, it need not pass through the entire LLM; instead, we insert those prompts " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{p}_l" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": " to the last layer of the LLM " + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": ". It is achieved" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28456" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "content": "by encoding them alongside the cache from " + }, + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{y}}" + }, + { + "bbox": [ + 47, + 72, + 258, + 85 + ], + "type": "text", + "content": ", as in" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 287, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 94, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 287, + 107 + ], + "type": "interline_equation", + "content": "\\boldsymbol {H} _ {l} = \\mathcal {D} _ {N} \\left(\\boldsymbol {K} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {V} _ {\\tilde {\\boldsymbol {y}}}, \\boldsymbol {p} _ {l}\\right), \\tag {9}", + "image_path": "41733c8dfc9faaa988dee2f0eee587662c59f481eca6001a2b9e221237470868.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "inline_equation", + "content": "K_{\\tilde{y}}, V_{\\tilde{y}}" + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "text", + "content": " represent the knowledge cache. This design enables LLaMP to efficiently learn informative prompt embeddings for the CLIP encoder " + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "text", + "content": ". It accomplishes this by incurring modest training costs, compared with training the entire LLM. Simultaneously, it maintains the essential knowledge inherent in the LLM decoder " + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 115, + 287, + 186 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "text", + "content": "Training Targets of LLaMP. Although the training strategy in Eqn. 9 has reduced the number of learnable parameters, a full decoder inside a LLM still consists of an enormous number of parameters. For example, one layer in LLaMA-7B bears 200M parameters, making training of the entire layer costly. As the goal is to leverage the knowledge from LLM, altering a full layer can lead to the loss of knowledge. Thus, as shown in Fig. 2, a typical decoder layer has two major components: the self attention module, consisting Query, Key, Value and Output projection layers, and the Feed-Forward Network (FFN). LLaMP targets the Query and Output projection layer inside the self-attention module. By updating the Query layer, LLM prompts " + }, + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "text", + "content": " are learned to distill pertinent information from the knowledge cache and the Output layer projects it to the latent space. We keep the Key and Value layers frozen to ensure the alignment between " + }, + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 187, + 288, + 414 + ], + "type": "text", + "content": " and knowledge cache. We leave the FFN unchanged to preserve the knowledge. Further discussions regarding these choices are made in Sec. 4.3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 582 + ], + "type": "text", + "content": "Textual Priors from Pre-Generated Responses. We extend the initial prompt, \"In one sentence, describe the distinctive appearance of [STH]\", by incorporating the response generated by the language model into the input sequence. This approach enriches the base content: the generated text provides a clear and explicit description of the object's appearance, acting as a valuable informative prior for language model adaptation. However, it is common for responses from an LLM to include filler words like \"sure\" for sentence structure coherence. To refine the input, we parse the noun phrases from the LLM's response through spaCy [12], an NLP engine, and merge them with the initial prompt, forming a more focused and informative language prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 714 + ], + "type": "text", + "content": "Textual Augmentations. Following the insights of Khattak et al. [19], which highlight the performance benefits of diverse textual inputs, we aim to further augment the text inputs used in the CLIP text encoder. Our approach, building upon the methods in [19, 48], incorporates handcrafted templates and expands their diversity through a two-step process: i) We introduce noun phrases into the existing templates for CLIP, for example, transforming \"A photo of [STH]\" to \"A photo of [STH] with [NP]\", thereby enriching the descriptive content; ii) We create a variety of new prompt templates for the LLM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "similar to “In one sentence, describe the distinctive appearance of [STH]” through GPT-4 [28], to further diversify the text input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 114, + 437, + 127 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 114, + 437, + 127 + ], + "spans": [ + { + "bbox": [ + 305, + 114, + 437, + 127 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": "Similar to PSRC [19], our objective function consists of three components: The main cross-entropy loss " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": ", feature-level L1 regularization " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{l1}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": ", and soft distillation loss " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dist}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": " training categories and " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": " training samples, " + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 304, + 133, + 547, + 194 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 340, + 199, + 545, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 199, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 340, + 199, + 545, + 230 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} = - \\frac {1}{\\mathcal {N}} \\sum_ {i} \\log \\frac {\\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} / \\tau\\right)}{\\sum \\exp \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {j} / \\tau\\right)}. \\tag {10}", + "image_path": "bab545d417c7ede27875cf335099373ada7eca8875aff511a86de14ca599d244.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "text", + "content": "The L1 regularization is computed between learned features " + }, + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p, \\pmb{g}_p" + }, + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "text", + "content": " and pre-trained CLIP features " + }, + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{f}}, \\hat{\\pmb{g}}" + }, + { + "bbox": [ + 305, + 236, + 545, + 262 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 267, + 545, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 267, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 318, + 267, + 545, + 295 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l 1} = \\frac {1}{\\mathcal {N}} \\sum_ {i} \\lambda_ {v} \\left| \\boldsymbol {f} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {f}} ^ {i} \\right| + \\frac {1}{C} \\sum_ {i} \\lambda_ {t} \\left| \\boldsymbol {g} _ {\\boldsymbol {p}} ^ {i} - \\hat {\\boldsymbol {g}} ^ {i} \\right|, \\tag {11}", + "image_path": "161693e81e6ead7904c7c1fb9c334ce707f0744904644ff3fe41c69bdf010453.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\lambda_v" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 304, + 301, + 545, + 338 + ], + "type": "text", + "content": " are coefficients. The prediction of LLaMP is further bound by the KL-Divergence between predicted distributions of LLaMP and vanilla CLIP:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 345, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 345, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 354, + 345, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\lambda_ {\\text {d i s t}} D _ {K L} \\left(\\boldsymbol {f} _ {\\boldsymbol {p}} \\cdot \\boldsymbol {g} _ {\\boldsymbol {p}}, \\hat {\\boldsymbol {f}} \\cdot \\hat {\\boldsymbol {g}}\\right). \\tag {12}", + "image_path": "9bf819afc2fbfc28a613d1bb03437a306ca44b2d8645ade3744e7a6290ccb556.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 366, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 390 + ], + "type": "text", + "content": "We sum all three losses up as the final objective function: " + }, + { + "bbox": [ + 305, + 366, + 545, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{CE} + \\mathcal{L}_{l1} + \\mathcal{L}_{dist}" + }, + { + "bbox": [ + 305, + 366, + 545, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 438 + ], + "type": "text", + "content": "During training, we randomly sample one LLM template as the input of LLaMP for each batch. For inference, we compute the probability distribution predicted from each input template and average them." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 448, + 388, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 388, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 388, + 462 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 468, + 413, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 413, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 413, + 480 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 594 + ], + "type": "text", + "content": "Datasets. Similar to previous work [18, 19, 48], in our study, we evaluate LLaMP performance over a spectrum of classification tasks with 11 datasets, including ImageNet [8] and Caltech101 [10] for generic image classification, OxfordPets [29], StanfordCars [21], Flowers102 [26], Food101 [3], and FGVCAircraft [25] for fine-grained classification, SUN397 [45] for scene recognition, UCF101 [37] for action recognition, DTD [7] for texture classification, and EuroSAT [11] for satellite image recognition." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Scenarios & Metrics. We evaluate LLaMP on two typical low-shot scenarios: zero-shot base-to-novel generalization and few-shot image classification. In zero-shot base-to-novel generalization, the base classes are seen during training, while the novel classes are unseen. We measure models performance through accuracies of base and novel classes, and the harmonic mean of the two. For few-shot classification, we assess the accuracy with 16 shots per class." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Implementation Details. We build LLaMP through the PyTorch [30] framework. All models are trained with" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28457" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 530, + 224 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 530, + 224 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 530, + 224 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 530, + 224 + ], + "type": "table", + "html": "
MethodAverageImageNet [8]Caltech101 [10]OxfordPets [29]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3474.2271.7072.4368.1470.2296.8494.0095.4091.1797.2694.12
CoOp [49]82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp [48]80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
KAPT* [17]78.4170.5274.2671.1065.2068.0297.1093.5395.2893.1396.5394.80
ProDA [24]81.5672.3076.6576.6670.5473.4797.7494.3696.0295.4397.7696.58
MaPLe [18]82.2875.1478.5575.4070.3272.7298.2793.2395.6895.4397.8396.62
RPO [23]81.1375.0077.7876.6071.5774.0097.9794.3796.0394.6397.5096.05
PSRC [19]84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
LLaMP85.1677.7181.2777.9971.2774.4898.4595.8597.1396.3197.7497.02
Δ w.r.t. PSRC+0.90+1.61+1.30+0.39+0.54+0.47+0.35+1.82+1.11+0.98+0.44+0.72
", + "image_path": "e31197af8adaa8d8b7eae93e47dfe3fd81a52d8a78cffd09ae0dc33bb1edb96a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 228, + 529, + 383 + ], + "blocks": [ + { + "bbox": [ + 65, + 228, + 529, + 383 + ], + "lines": [ + { + "bbox": [ + 65, + 228, + 529, + 383 + ], + "spans": [ + { + "bbox": [ + 65, + 228, + 529, + 383 + ], + "type": "table", + "html": "
MethodStanfordCars [21]Flowers102[26]Food101 [3]FGVCAircraft [25]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]63.3774.8968.6572.0877.8074.8390.1091.2290.6627.1936.2931.09
CoOp [49]78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp [48]70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
KAPT* [17]69.4766.2067.7995.0071.2081.4086.1387.0686.5929.6728.7329.19
ProDA [24]72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
MaPLe [18]74.7071.2072.9197.7068.6880.6690.3088.5789.4336.9034.1335.46
RPO [23]73.8775.5374.6994.1376.6784.5090.3390.8390.5837.3334.2035.70
PSRC [19]78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
LLaMP81.5674.5477.8997.8277.4086.4291.0591.9391.4947.3037.6141.90
Δ w.r.t. PSRC+3.29-0.43+1.31-0.25+0.90+0.47+0.38+0.40+0.39+4.57-0.26+1.75
", + "image_path": "72616f794bb299e77101e81471b3aa570be7d7de907b66acd6dd735a05110dc0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 386, + 529, + 541 + ], + "blocks": [ + { + "bbox": [ + 65, + 386, + 529, + 541 + ], + "lines": [ + { + "bbox": [ + 65, + 386, + 529, + 541 + ], + "spans": [ + { + "bbox": [ + 65, + 386, + 529, + 541 + ], + "type": "table", + "html": "
MethodSUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CLIP [32]69.3675.3572.2353.2459.9056.3756.4864.0560.0370.5377.5073.85
CoOp [49]80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp [48]79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
KAPT* [17]79.4074.3376.7875.9758.3065.9784.8067.5775.2180.8367.1073.33
ProDA [24]80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
MaPLe [18]78.4776.9377.7980.6756.4866.4483.9066.0073.8885.2371.9778.04
RPO [23]80.6077.8079.1876.7062.1368.6186.6368.9776.7983.6775.4379.34
PSRC [19]82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
LLaMP83.4179.9081.6283.4964.4972.7791.9383.6687.6087.1380.6683.77
Δ w.r.t. PSRC+0.74+1.43+1.10+0.12+1.52+1.02-0.97+9.76+5.28+0.03+1.86+1.03
", + "image_path": "49bb5878a17491d4a0b157d7e8880559049eb07afca1468b5b03940175ab0b6d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 551, + 547, + 584 + ], + "lines": [ + { + "bbox": [ + 46, + 551, + 547, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 547, + 584 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on base-to-novel generalization. LLaMP shows strong generalization results over previous approaches on 11 image classification tasks. Absolute gains over PSRC are indicated in blue. *KAPT is trained with ViT-B/32 image encoder instead of ViT-B/16." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "text", + "content": "2 NVIDIA A100 40GB GPUs. For LLaMP, we adopt LLaMA2-7B [39] as the language model " + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "text", + "content": ", and ViT-B/16 [9] as the image encoder, following [18, 19, 48, 49]. On the text side, we set prompt learning depth to 9. To tune the vision encoder, we adopt the hybrid tuning scheme which performs deep prompt learning on the first 6 layers and LoRA on the rest. Similar to [13], LoRA is applied to the Query and Value projection layers inside attention modules. The number of " + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "text", + "content": " prompts, " + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 590, + 289, + 712 + ], + "type": "text", + "content": ", is set to 16. We set a global learning rate of 2E-4 with a batch size of 8. The learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": "rate of LoRA modules is set to 2E-5. " + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{t},\\lambda_{v}" + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{dist}" + }, + { + "bbox": [ + 304, + 591, + 545, + 616 + ], + "type": "text", + "content": " are set to 25, 10 and 2.5, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 623, + 442, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 623, + 442, + 635 + ], + "spans": [ + { + "bbox": [ + 305, + 623, + 442, + 635 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Zero-Shot Base-to-Novel Generalization. LLaMP outperforms existing state-of-the-art prompt learning methods on most metrics of 11 classification datasets in the baseto-novel generalization benchmark. As shown in Tab. 1, compared to the latest model PSRC [19], LLaMP achieves average gains of " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "0.90\\%" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " in base accuracy, " + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.61\\%" + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": " in novel" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28458" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 515, + 216 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 515, + 216 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 515, + 216 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 515, + 216 + ], + "type": "table", + "html": "
16-Shot Classification
AverageImageNet [8]Collect [10]Pets [29]Cars [21]Flowers [26]Food [3]Aircraft [25]SUN397 [45]DTD [7]EuroSAT [11]UCF101 [37]
CLIP [32]78.79 (65.02)67.3195.4385.3480.4497.3782.9045.3673.2869.9687.2182.11
CoOp [49]79.89 (73.82)71.8795.5791.8783.0797.0784.2043.4074.6769.8784.9382.23
CoCoOp [48]74.90 (70.70)70.8395.1693.3471.5787.8487.2531.2172.1563.0473.3278.14
MaPLe [18]81.79 (75.58)72.3396.0092.8383.5797.0085.3348.4075.5371.3392.3385.03
PSRC [19]82.87 (77.90)73.1796.0793.6783.8397.6087.5050.8377.2372.7392.4386.47
LLaMP83.81 (78.50)73.4997.0894.2186.0798.0687.6256.0777.0274.1791.3186.84
", + "image_path": "908bf87a600e47c599f21fe166edd5ef55e1b62e24f4faa081e57d22fd4b317b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "text", + "content": "accuracy, and " + }, + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "inline_equation", + "content": "1.30\\%" + }, + { + "bbox": [ + 46, + 241, + 287, + 314 + ], + "type": "text", + "content": " in harmonic mean on average. Moreover, LLaMP consistently achieves higher harmonic means (HM) compared to other models. These improvements indicate that our approach better balances performance on base and novel data, thus achieving stronger generalization compared to the prior prompt learning techniques." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": "In particular, LLaMP excels in fine-grained datasets requiring detailed analysis. On FGVCAircraft, LLaMP surpasses PSRC by " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "4.57\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " on base accuracy and " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "1.75\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " on HM, highlighting its strong understanding of detailed aircraft features. Furthermore, on EuroSAT, LLaMP achieves improvements of " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "9.76\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "5.28\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " on novel accuracy and HM, respectively. We also observe similar performance gains on StanfordCars, where LLaMP outperforms by " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "3.29\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " on base accuracy and " + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "inline_equation", + "content": "1.31\\%" + }, + { + "bbox": [ + 46, + 314, + 287, + 456 + ], + "type": "text", + "content": " on HM. The information embedded in LLM enables LLaMP to capture and utilize the rich semantic information necessary for distinguishing between closely related categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": "Few-Shot Classification. LLaMP also achieves improvements across these classification datasets in few-shot classification tasks. As in Tab.2, with an average classification accuracy of " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "83.81\\%" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": ". Notably, on FGVCAircraft and StanfordCars, LLaMP shows a significant improvement over PSRC, further demonstrating that the knowledge from language models benefits the recognition of fine-grained object categories, which aligns with our observation on zero-shot base-to-novel generalization. Moreover, on DTD, where MaPLe and PSRC achieve around " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " accuracy, LLaMP achieves a higher accuracy of " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "74.17\\%" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": ", underscoring its ability to recognize textures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 141, + 623 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 689 + ], + "type": "text", + "content": "Is the knowledge from LLM helping? In Tab. 3, we show that the knowledge from LLM benefits in both ways: Without training, performance of ordinary CLIP model can be improved by introducing noun phrases; The LLaMP framework shows further improvement after training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Noun phrases are parsed from the LLM's responses the prompt of \"Describe [STH]\". We then use the tem" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 239, + 518, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 536, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 536, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 536, + 234 + ], + "type": "text", + "content": "Table 2. Few shot classification results with 16 shots. Numbers in the bracket indicate the average performance over " + }, + { + "bbox": [ + 55, + 223, + 536, + 234 + ], + "type": "inline_equation", + "content": "1/2/4/8/16" + }, + { + "bbox": [ + 55, + 223, + 536, + 234 + ], + "type": "text", + "content": " shots." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 239, + 518, + 318 + ], + "lines": [ + { + "bbox": [ + 334, + 239, + 518, + 318 + ], + "spans": [ + { + "bbox": [ + 334, + 239, + 518, + 318 + ], + "type": "table", + "html": "
MethodLLMBaseNovelHM
CLIP✓69.3474.2271.70
70.9574.9372.79
LLaMP✓82.2176.4479.22
85.1677.7181.27
", + "image_path": "644211bcd4075aa0c799a9dc35d2eebcec7fa7de671298366b78b4a10d58830c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 344, + 544, + 431 + ], + "blocks": [ + { + "bbox": [ + 336, + 326, + 515, + 337 + ], + "lines": [ + { + "bbox": [ + 336, + 326, + 515, + 337 + ], + "spans": [ + { + "bbox": [ + 336, + 326, + 515, + 337 + ], + "type": "text", + "content": "Table 3. Ablation study on the LLM Knowledge." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 344, + 544, + 431 + ], + "lines": [ + { + "bbox": [ + 309, + 344, + 544, + 431 + ], + "spans": [ + { + "bbox": [ + 309, + 344, + 544, + 431 + ], + "type": "table", + "html": "
LPQOKVFFN%BaseNovelHM
✓.0385.0077.2980.96
✓✓✓3385.2077.4581.14
✓✓✓8385.0577.7381.22
✓✓✓✓10085.2377.5681.21
✓✓1785.1677.7181.27
", + "image_path": "82d95e9394564cdf9cb81b6389a2030505e2d9017286443817d2d88f119de9fb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 439, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 306, + 439, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 545, + 462 + ], + "type": "text", + "content": "Table 4. Ablation study on the Training Strategy. “%” indicates the ratio of parameters trained compared to fully tuning a layer." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 533 + ], + "type": "text", + "content": "plate, \"A photo of [STH] with [NP]\" to generate the NP-augmented text embedding for CLIP. We take the average of all augmented embeddings for classification. In Tab. 3 we show that even ordinary CLIP can benefit from incorporating LLMs' knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 629 + ], + "type": "text", + "content": "Furthermore, the comparison between LLaMP and LLaMP without the LLM indicates that merely integrating LoRA [13] to the vision encoder is not beneficial. The \"LLaMP without LLM\" is essentially an ordinary prompting learning model plus LoRAs in the vision encoder. We show that the improved vision encoding capacity only benefits when the quality of text embeddings s are enhanced by incorporating LLMs' knowledge through LLaMP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Decoder Training strategy. We categorize trainable parameters of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_N" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " into four groups: learnable prompts (LP), Query and Output projections (QO), Key and Value projections (KV), and the feed-forward network (FFN). Tab. 4 indicates LLaMP can achieve desirable results by just learning the prompts of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". One step further, adding QO into optimization achieves the best performance. Although other" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28459" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 261, + 131 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 261, + 131 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 261, + 131 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 261, + 131 + ], + "type": "table", + "html": "
MethodPriorsBaseNovelHM
LLaMPX84.9077.5981.08
Plain85.2677.5681.22
NP85.1677.7181.27
", + "image_path": "a7aa7d3ddc9a8573ebaf8270c0603fcd5a1b885ade38e508e4c38e0f1e87ea2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 170, + 249, + 219 + ], + "blocks": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "type": "text", + "content": "Table 5. Ablation Study on Pre-generated Text Priors. " + }, + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 47, + 140, + 287, + 161 + ], + "type": "text", + "content": " refers to \"without textual priors\" and NP stands for noun phrases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 170, + 249, + 219 + ], + "lines": [ + { + "bbox": [ + 85, + 170, + 249, + 219 + ], + "spans": [ + { + "bbox": [ + 85, + 170, + 249, + 219 + ], + "type": "table", + "html": "
MethodBaseNovelHM
LLM Only81.7435.8249.81
LLaMP85.1677.7181.27
", + "image_path": "ccb74f6db239837352b7451632af8b0544c7984b8810ce38464e46b458c7c8d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 247, + 276, + 312 + ], + "blocks": [ + { + "bbox": [ + 77, + 228, + 256, + 239 + ], + "lines": [ + { + "bbox": [ + 77, + 228, + 256, + 239 + ], + "spans": [ + { + "bbox": [ + 77, + 228, + 256, + 239 + ], + "type": "text", + "content": "Table 6. The CLIP text encoder helps adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 247, + 276, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 247, + 276, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 247, + 276, + 312 + ], + "type": "table", + "html": "
SchemeBaseNovelHM
Prompt × 984.6777.2880.81
LoRA × 1284.8977.2780.90
Prompt ×6 + LoRA ×685.1677.7181.27
", + "image_path": "ed65280b1c1c6a8f28208dd6804fd00a6bae8d8dd44772482f62188d12a96b76.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 321, + 286, + 343 + ], + "lines": [ + { + "bbox": [ + 47, + 321, + 286, + 343 + ], + "spans": [ + { + "bbox": [ + 47, + 321, + 286, + 343 + ], + "type": "text", + "content": "Table 7. Study on Vision Tuning Scheme. Our hybrid design achieves the best performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 352, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 286, + 375 + ], + "type": "text", + "content": "setups introduce much more trainable parameters, they can not surpass the \"LP + QO\" strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "text", + "content": "Effect of Textual Priors. We study the effect of pregenerated textual priors on LLaMP. We compare three different approaches: without textual priors, using plain responses as the prior, and LLaMP, which takes parsed noun phrases. Tab. 5 shows that LLaMP can achieve over " + }, + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "inline_equation", + "content": "81\\%" + }, + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "text", + "content": " on HM without pre-generated priors, while adding parsed noun phrases as textual priors further pushes the HM to " + }, + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "inline_equation", + "content": "81.27\\%" + }, + { + "bbox": [ + 47, + 376, + 286, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "content": "CLIP as the bridge. One may wonder if it is possible to replace CLIP text encoder with a large language model. Here, we study two setups: i) LLM as encoder, which treats the output of the language model, " + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{h}_l" + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "content": " as the text feature; ii) LLaMP, which treat " + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "inline_equation", + "content": "\\tilde{h}_l" + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "content": " as part of the text input prompt. Tab. 6 reveals that relying solely on the LLM results in poor accuracy for novel categories. This supports our hypothesis that aligning LLMs with vision encoders generally requires a more extensive dataset. Furthermore, LLaMP's design significantly improves the novel accuracy by " + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 460, + 286, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 581, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 286, + 688 + ], + "type": "text", + "content": "Vision Training Strategy. As ViT-16/B has 12 transformer layers, we compare different vision training strategies within LLaMP in Tab. 7. Apart from the default hybrid scheme, we evaluate setups including prompt learning at first 9 layers (P9), a similar setup to PSRC [19], and LoRA [13] in all 12 layers. The results suggest that the scheme leverages the strengths of both prompt learning and LoRA, addressing potential bottlenecks in the vision encoder and enhancing overall performance in LLaMP." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Number of LLM Prompts. We vary the number of LLM prompts and study their effects on LLaMP. As in" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 76, + 538, + 167 + ], + "blocks": [ + { + "bbox": [ + 315, + 76, + 538, + 167 + ], + "lines": [ + { + "bbox": [ + 315, + 76, + 538, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 76, + 538, + 167 + ], + "type": "image", + "image_path": "5eb53d774d0e54e537515c705435c6df8314a142113bfec24752075d117f18bc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 171, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 545, + 194 + ], + "type": "text", + "content": "Figure 3. Effect of LLM Prompts on Harmonic Mean. 16 prompts achieve the most balanced performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 308, + 196, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 308, + 196, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 308, + 196, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 196, + 545, + 369 + ], + "type": "image", + "image_path": "e434f605bf4506380301043f978de0813f9f16be30e16f768ac6c04911f40740.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 377, + 544, + 411 + ], + "lines": [ + { + "bbox": [ + 305, + 377, + 544, + 411 + ], + "spans": [ + { + "bbox": [ + 305, + 377, + 544, + 411 + ], + "type": "text", + "content": "Figure 4. Visualization of LLaMP Predictions by GradCAM [35] Fig. 3, using 16 prompts optimizes the LLM's capabilities, achieving the highest harmonic mean at " + }, + { + "bbox": [ + 305, + 377, + 544, + 411 + ], + "type": "inline_equation", + "content": "81.27\\%" + }, + { + "bbox": [ + 305, + 377, + 544, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": "Visualizations. In Fig. 4, we visualize the gradient heatmap of input images from FGVCAircraft and EuroSAT, through GradCAM [35]. The figure shows that, LLaMP can capture distinctive features that matches LLM's description." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "spans": [ + { + "bbox": [ + 306, + 470, + 378, + 482 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 586 + ], + "type": "text", + "content": "Our study shows that the encyclopedic knowledge from LLMs is beneficial for low-shot image classification as extra class-specific information. To leverage such knowledge, we propose LLaMP, a framework that adapts LLMs as prompt learners for the CLIP model. Over two common low-shot scenarios: zero-shot generalization and few-shot learning, LLaMP demonstrates notable improvements compared with previous state-of-the-arts on a spectrum of datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 586, + 545, + 658 + ], + "type": "text", + "content": "Limitations. While LLaMP reveals an effective way in leveraging LLMs' knowledge, both modalities, vision and language, only interact at the finest feature level. Given the broader LLM-aided knowledge from the language side, the performance can be potentially further improved by introducing language priors at earlier vision encoding stages." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 670, + 398, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 398, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 398, + 682 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This research was supported, in part, by the Office of Naval Research under grant #N00014-21-1-2802." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28460" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 157 + ], + "type": "text", + "content": "[1] Ebetesam Almazrouei, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Merouane Debbah, Etienne Goffinet, Daniel Heslow, Julien Launay, Quentin Malartic, et al. Falcon-40b: an open large language model with state-of-the-art performance. Technical report, Technical report, Technology Innovation Institute, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 203 + ], + "type": "text", + "content": "[2] Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv preprint arXiv:2305.10403, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 237 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461. Springer, 2014. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 282 + ], + "type": "text", + "content": "[4] Shiming Chen, Wenjie Wang, Beihao Xia, Qinmu Peng, Xinge You, Feng Zheng, and Ling Shao. Free: Feature refinement for generalized zero-shot learning. In ICCV, pages 122-131, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 339 + ], + "type": "text", + "content": "[5] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 53, + 284, + 287, + 339 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 53, + 284, + 287, + 339 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 396 + ], + "type": "text", + "content": "[6] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 430 + ], + "type": "text", + "content": "[7] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 464 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255. IEEE, 2009. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 531 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 577 + ], + "type": "text", + "content": "[10] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPR workshop, pages 178-178. IEEE, 2004. 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 633 + ], + "type": "text", + "content": "[11] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[12] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[14] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[15] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 215 + ], + "type": "text", + "content": "[16] Michael Kampffmeyer, Yinbo Chen, Xiaodan Liang, Hao Wang, Yujia Zhang, and Eric P Xing. Rethinking knowledge graph propagation for zero-shot learning. In CVPR, pages 11487-11496, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 259 + ], + "type": "text", + "content": "[17] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 546, + 304 + ], + "type": "text", + "content": "[18] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 304, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 546, + 358 + ], + "type": "text", + "content": "[19] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 546, + 403 + ], + "type": "text", + "content": "[20] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In 5th ICLR, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 546, + 437 + ], + "type": "text", + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, pages 554–561, 2013. 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 546, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 546, + 470 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 546, + 470 + ], + "type": "text", + "content": "[22] Christoph H Lampert, Hannes Nickisch, and Stefan Harmeling. Attribute-based classification for zero-shot visual object categorization. TPAMI, 36(3):453-465, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 546, + 514 + ], + "type": "text", + "content": "[23] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 1, 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "text", + "content": "[24] Yuning Lu, Jianzhuang Liu, Yonggang Zhang, Yajing Liu, and Xinmei Tian. Prompt distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5206-5215, 2022. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 559, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 602 + ], + "type": "text", + "content": "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 647 + ], + "type": "text", + "content": "[26] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729. IEEE, 2008. 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 669 + ], + "type": "text", + "content": "[27] OpenAI. Chatgpt, 2023. Available at https://openai.com/chatgpt.2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 670, + 493, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 493, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 493, + 680 + ], + "type": "text", + "content": "[28] OpenAI. Gpt-4 technical report, 2023. 1, 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In " + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "inline_equation", + "content": "CVPR" + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": ", pages 3498-3505. IEEE, 2012. 5, 6, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28461" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, pages 8026-8037, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[31] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 250 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 251, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 251, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 251, + 288, + 308 + ], + "type": "text", + "content": "[33] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 342 + ], + "type": "text", + "content": "[34] Scott Reed, Zeynep Akata, Honglak Lee, and Bernt Schiele. Learning deep representations of fine-grained visual descriptions. In CVPR, pages 49-58, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 397 + ], + "type": "text", + "content": "[35] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626, 2017. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 442 + ], + "type": "text", + "content": "[36] Richard Socher, Milind Ganjoo, Hamsa Sridhar, Osbert Bastani, Christopher D Manning, and Andrew Y Ng. Zero-shot learning through cross-modal transfer. arXiv preprint arXiv:1301.3666, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 477 + ], + "type": "text", + "content": "[37] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "text", + "content": "[38] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 589 + ], + "type": "text", + "content": "[39] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 624 + ], + "type": "text", + "content": "[40] Xiaolong Wang, Yufei Ye, and Abhinav Gupta. Zero-shot recognition via semantic embeddings and knowledge graphs. In CVPR, pages 6857-6866, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 689 + ], + "type": "text", + "content": "[41] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaoqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In European Conference on Computer Vision, pages 631-648. Springer, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[42] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 518 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 116 + ], + "type": "text", + "content": "Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 239 + ], + "type": "text", + "content": "[43] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierrick Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 38-45, Online, 2020. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 272 + ], + "type": "text", + "content": "[44] Yongqin Xian, Tobias Lorenz, Bernt Schiele, and Zeynep Akata. Feature generating networks for zero-shot learning. In CVPR, pages 5542-5551, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 274, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 316 + ], + "type": "text", + "content": "[45] Jianxiong Xiao, Krista A Ehinger, James Hays, Antonio Torralba, and Aude Oliva. Sun database: Exploring a large collection of scene categories. IJCV, 119(1):3-22, 2016. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 350 + ], + "type": "text", + "content": "[46] Li Zhang, Tao Xiang, and Shaogang Gong. Learning a deep embedding model for zero-shot learning. In CVPR, pages 2021-2030, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 352, + 545, + 406 + ], + "type": "text", + "content": "[47] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 408, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 441 + ], + "type": "text", + "content": "[48] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 474 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 518 + ], + "type": "text", + "content": "[50] Yizhe Zhu, Mohamed Elhoseiny, Bingchen Liu, Xi Peng, and Ahmed Elgammal. A generative adversarial approach for zero-shot learning from noisy texts. In CVPR, pages 1004-1013, 2018. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28462" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_content_list.json b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..161517c13b27205df08461b84157093dc6bf4b67 --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_content_list.json @@ -0,0 +1,1741 @@ +[ + { + "type": "text", + "text": "Latency Correction for Event-guided Deblurring and Frame Interpolation", + "text_level": 1, + "bbox": [ + 109, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixin Yang $^{1,2,\\dagger}$ Jinxiu Liang $^{1,2}$ Bohan Yu $^{1,2}$ Yan Chen $^{3}$ Jimmy S. Ren $^{3}$ Boxin Shi $^{1,2,*}$", + "bbox": [ + 125, + 179, + 844, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "bbox": [ + 140, + 200, + 828, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{3}$ SenseTime Research $\\{\\text { yangyixin93,cssherryliang,ybh1998,shiboxin }\\} @$ pku.edu.cn yanchenace@gmail.com rensijie@sensetime.com", + "bbox": [ + 84, + 217, + 883, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras, with their high temporal resolution, dynamic range, and low power consumption, are particularly good at time-sensitive applications like deblurring and frame interpolation. However, their performance is hindered by latency variability, especially under low-light conditions and with fast-moving objects. This paper addresses the challenge of latency in event cameras — the temporal discrepancy between the actual occurrence of changes in the corresponding timestamp assigned by the sensor. Focusing on event-guided deblurring and frame interpolation tasks, we propose a latency correction method based on a parameterized latency model. To enable data-driven learning, we develop an event-based temporal fidelity to describe the sharpness of latent images reconstructed from events and the corresponding blurry images, and reformulate the event-based double integral model differentiable to latency. The proposed method is validated using synthetic and real-world datasets, demonstrating the benefits of latency correction for deblurring and interpolation across different lighting conditions.", + "bbox": [ + 75, + 318, + 473, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 652, + 209, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emerging as a groundbreaking visual sensing paradigm, event cameras [1, 12, 13, 20, 22] provide a novel perspective to capturing and interpreting dynamic scenes. They have advantages of high temporal resolution (the discrete measurement resolution of the event detection time, in the order of microseconds [8]), high dynamic range (e.g., 120 dB for DAVIS346), and low power consumption. They asynchronously record changes in logarithmic pixel intensity only when they exceed a predefined threshold. Their unique characteristics have led to impressive performance gains in various time-sensitive applications, including deblurring [5, 19, 24, 26, 32] and frame interpolation [19, 25, 30].", + "bbox": [ + 75, + 679, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/766f2bb6c7df79a6ba5f677d693446d68924ec73330ae2a0a68c5000f8c3626d.jpg", + "image_caption": [ + "Figure 1. Illustration of latency in an event camera, and its influence on an event-guided deblurring algorithm. (a) The aerial view of ideal vs. real events triggered by a black point moving upwards on a white background, where the real events are captured in low-light conditions (below 0.1 Lux). (b) Deblurring results using the Event-based Double Integral (EDI) [19] method, guided by events captured under different illumination conditions. (c) Improved deblurring results from EDI [19] using event with latency corrected by the proposed method." + ], + "image_footnote": [], + "bbox": [ + 501, + 285, + 712, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/053a46be857106f0a9c92d0f53f3edf55067bcd84a6ae06868c1adb8ee83bab7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 287, + 888, + 358 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95d2c6f0c172e3f4049a30243b704b656d2f75ce34a4e3032a01c9a4f5b3b5aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 362, + 888, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ideal event model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this, exhibiting a temporal discrepancy between the actual moment of intensity change and the corresponding event timestamp. As claimed by the inventors of the event-based sensors, in well-lit environments above 1k Lux, event cameras can achieve latency as low as $3\\mu \\mathrm{s}$ [22]. However, in dimmer conditions for a few Lux, the latency can exceed 1 ms [22]. The challenge in modeling this temporal discrepancy arises from the sensor pixels' non-linear and variable response to illumination, which is further complicated by manufacturing imperfection, inherent electronic noise, and the unpredictable nature of real-world scenes [13, 15, 16, 21]. Therefore, the profile of the temporal discrepancy is spatial-varying, inconsistent, and noisy. All these complexities make algorithms for handling them scarce.", + "bbox": [ + 496, + 566, + 892, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Without specifying the nature of the temporal discrepancy, we consider the collective temporal discrepancies in the timing of the camera's response to real-world changes, a phenomenon we term \"latency\", where the aforementioned sensor-level latency is one of the dominant components.", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is done during Yixin's internship at SenseTime.", + "bbox": [ + 93, + 875, + 410, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: shiboxin@pku.edu.cn", + "bbox": [ + 93, + 887, + 341, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24977", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the complexity of the latency formation model, latency presents a strong relation with illuminance. As illustrated in Figure 1 (a), compared with the ideal events generated instantaneously when the intensity changes, the real events suffer from latency in illuminance below 0.1 Lux. The latency presents a significant challenge for time-sensitive applications [19, 24, 25, 30], which assume ideal event triggering and struggle to restore sharp edges due to the severe latency, as demonstrated in Figure 1 (b) vs. (c).", + "bbox": [ + 75, + 90, + 468, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we focus on latency correction for event-guided imaging deblurring [19, 24, 28] and interpolation [19, 25, 31] in low-light and fast-moving scenes, where event cameras should play an important role. Based on the analysis and modeling of latency, we propose to predict latency in event cameras by conceptualizing latency as a variable function—potentially a constant for a specific scene or a polynomial function of the illuminance level, given a fixed event camera type—we endeavor to elucidate the relationship between latency and illuminance.", + "bbox": [ + 75, + 231, + 470, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose a latency correction method for captured events and blurry images in an event-image hybrid sensor (e.g., DAVIS346). Inspired by the latency properties related to illuminance observed through controlled experiments, we characterize the latency-intensity relationship using a parameterized curve. To estimate the optimal latency, we introduce a novel objective to assess the sharpness profile of a latent image reconstructed from the latency-corrected events and the blurry image, which we term event-based temporal fidelity. Building upon them, we collect data under different scene illuminance and advocate for a self-supervised approach to latency correction. The EDI model [18, 19] is reformulated to be differentiable for latency to enable gradient-based optimization, therefore enhancing the performance of deblurring and interpolation.", + "bbox": [ + 75, + 387, + 468, + 614 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The efficacy of the proposed method is validated using both synthetic and real data, demonstrating improvements in image quality for deblurring and interpolation under challenging lighting conditions where the latency in events becomes severe. The key contributions of our work include:", + "bbox": [ + 75, + 619, + 468, + 695 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- the first data-driven latency correction method grounded in a differentiable reformulation of the event integral with respect to latency;", + "- the event-based temporal fidelity for measuring the sharpness profile of a latent image reconstructed by latency-corrected events and the corresponding blurry image; and", + "- a parameterized latency-intensity curve for modeling the relationship of latency in events and the intensity of the corresponding blurry image." + ], + "bbox": [ + 76, + 699, + 467, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method provides an approach to understanding and mitigating the effects of latency, paving the way for more accurate and reliable event-based imaging in dynamic environments with challenging lighting conditions.", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 500, + 89, + 645, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We investigate the influence of latency on imaging algorithms in event-image hybrid systems, particularly focusing on deblurring and interpolation tasks due to their sensitivity to timing in events.", + "bbox": [ + 498, + 117, + 890, + 178 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Analysis of latency in event cameras. The triggering of events involves three subprocesses: photon reception, event trigger, and event readout. Notably, latency primarily arises during photon reception and event readout due to the electronic and physical design of these cameras [8]. Latency characteristics vary among different event camera models, ranging from minimal latency of $15\\mu s$ [13] to $4\\mu s$ [20] under 1k Lux conditions. Despite improvements in camera design, latency persists at nearly 1 ms under 1 Lux [1], posing challenges in low-light scenarios. Existing studies predominantly focus on sensor design aspects of event latency [1, 12, 13, 22] with a limited number examining latency under varying illuminance levels [6, 21]. In low-light conditions, latency induced by photoreceptor bandwidth becomes predominant [16], which is proportional to photocurrent (or illuminance) at low illuminance level, increasing monotonically with the illuminance level [2, 13, 15, 16, 21]. Consequently, the latency decreases as illuminance increases [7, 13, 21]. The photoreceptor circuit's low-pass characteristic [15, 16], has led to its modeling as a first- or second-order low-pass filter in various event simulators [7, 10]. While this provides a simplified relationship between latency and photocurrent, accurately determining photocurrent is often invalid for arbitrary imaging setups due to the complex imaging processes involving shuttering, apertures, etc. In order to save costs, most event cameras timestamp each event as they are being read out from the pixel array [8], which introduces readout latency that is significantly influenced by sensitivity and contrast [2, 8, 10, 15, 21]. The complexity of these factors contributing to latency in event cameras explains the scarcity of algorithms specifically designed to address this challenge for event-image hybrid systems.", + "bbox": [ + 498, + 184, + 890, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event guided imaging algorithms. Event cameras record changes during the exposure time of blurry images, which makes the problem of deblurring and interpolation more well-posed. Pan et al. [18, 19] introduced the EDI model for high-frame-rate video reconstruction from blurry videos using event data, applicable in both deblurring and interpolation. For deblurring, Jiang et al. [9] combined convolutional recurrent networks with directional event filtering, Shang et al. [23] utilized the nearest sharp frames and events for deblurring, Xu et al. [28] improved model generalization through self-supervised learning, Teng et al. [26] focused on event representations for image enhancement, and Sun et al. [24] developed an event-image fusion module. For joint deblurring and interpolation, Lin et al. [14] employed neu", + "bbox": [ + 498, + 689, + 892, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24978", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ral networks and dynamic filtering, while Wang et al. [27] and Zhang et al. [30, 31] integrated multiple techniques, including event-enhanced sparse learning and self-supervised frameworks. Sun et al. [25] proposed a bidirectional recurrent network for fusion tasks.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For these algorithms assuming ideal event triggering, the reconstruction of the latent image between or within exposure time heavily depends on the precision of timestamps in accompanying events. Latency presence can significantly impair algorithm performance. To the best of our knowledge, these has been no approach to address latency for event-guided deblurring and interpolation.", + "bbox": [ + 75, + 167, + 470, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 286, + 250, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we aim to tackle the latency in events for event-guided deblurring and frame interpolation. Section 3.1 delves into the deviation of events with latency from the ideal event triggering mechanism and the event-based double integral model with latency. Based on this model, we propose a learning-based latency correction method for event-guided deblurring and frame interpolation. Section 3.2 introduces the objective for optimization, the event-based temporal fidelity. To allow gradient-based optimization, Section 3.3 reformulates event-based double integral to be differentiable respective to latency, where the latency is further parameterized into a polynomial of intensity on a per-pixel basis. Training details are shown in Section 3.4.", + "bbox": [ + 75, + 309, + 470, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Analysis on the latency in events", + "text_level": 1, + "bbox": [ + 76, + 513, + 361, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latency in event triggering. In an event camera, the photons hitting the photodetector of a pixel are transduced into an electrical signal and then compared against a predefined threshold to determine whether the change in intensity is substantial enough to trigger an \"event\". If this threshold is surpassed, the event is recorded with a high-resolution timestamp. Mathematically, the pixels of an event camera independently respond to variations in the continuous brightness signal $I_{p}(t)$ at pixel position $\\boldsymbol{p} = (p_x,p_y)^\\top$ and time $t$ . An event $e = (\\boldsymbol{p},t,\\sigma)$ is initiated at pixel $\\boldsymbol{p}$ and time $t$ with polarity $\\sigma \\in \\{+1, - 1\\}$ indicating ON or OFF events when the logarithmic change of $I_{p}$ exceeds a dispatched threshold $\\theta$ since the last event triggered at the pixel $\\boldsymbol{p}$ and time $t_{\\mathrm{ref}}$ :", + "bbox": [ + 75, + 539, + 468, + 751 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\log I _ {\\boldsymbol {p}} (t) - \\log I _ {\\boldsymbol {p}} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\theta . \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 757, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In tasks involving image reconstruction where per-pixel intensity is needed, the logarithmic intensity changes for continuous time neighboring recorded timestamps are often approximated as an equality involving the threshold value $\\theta$ and polarity $\\sigma$ with the quantization errors ignored.", + "bbox": [ + 75, + 779, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ideal model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this ideal, ex", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hibiting a temporal latency $l$ between the actual change in intensity and the event's triggering. The presence of latency exacerbates the approximation error of the intensity changes (e.g., Eq. (2) in the following), particularly affecting time-sensitive tasks like deblurring and frame interpolation. In environments with controlled lighting and motion, empirical methods have been employed to study and characterize latency in event cameras, such as using a second-order low-pass filter [13, 20] with parameters that are influenced by illumination conditions [6, 7, 13]. In the unpredictable flux of real-world settings, latency becomes markedly more unpredictable and defies precise modeling. The complexity arises from the non-linear and non-uniform factors. For instance, pixel responses across the sensor array exhibit variability due to manufacturing differences, and the light received by each pixel fluctuates rapidly with movement in the scene or changes in illumination conditions. Moreover, the existence of noise exacerbates the difficulty.", + "bbox": [ + 496, + 90, + 890, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based double integral model with latency. An event-image hybrid sensor (e.g., DAVIS346) is able to capture pixel-wise aligned events and images, synergistically harnessing the strengths of both modalities. This integration involves vibrant color representation characteristic of frame-based sensors, coupled with the high temporal resolution and dynamic range offered by event cameras. The mathematical foundation for the interplay between events and images in such a system is encapsulated by the EDI model, which is originally proposed for applications in deblurring and frame interpolation [18, 19].", + "bbox": [ + 496, + 367, + 890, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proportional change for images at time $\\tau$ and time $t$ can be measured by the integral of the $N$ events $\\mathbb{E} = \\{e_k\\}_{k=1}^N$ triggered from $\\tau$ to $t$ as:", + "bbox": [ + 498, + 534, + 890, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\boldsymbol {p}} (t) = I _ {\\boldsymbol {p}} (\\tau) \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 584, + 890, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{\\pmb{p}}(s)$ is a function of continuous time:", + "bbox": [ + 500, + 623, + 807, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\boldsymbol {p}} (s) = \\sigma_ {k} \\delta_ {t _ {k}} (s), \\quad s \\in (t _ {k - 1}, t _ {k} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 643, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "whenever there is an event $e_k = (\\pmb{p}_k, t_k, \\sigma_k)$ triggered and $\\delta_{t_k}(s)$ is an impulse function at time $t_k$ with unit integral. A blurry image $B_p(\\tau)$ with exposure duration $t \\in [\\tau - \\Delta \\tau, \\tau + \\Delta \\tau]$ can be represented as the average of the latent image $I_p(t)$ over the exposure duration [18], which can be formulated as:", + "bbox": [ + 498, + 666, + 890, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB _ {\\boldsymbol {p}} (\\tau) = \\frac {1}{2 \\Delta \\tau} I _ {\\boldsymbol {p}} (\\tau) R _ {\\boldsymbol {p}} (\\tau), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 758, + 890, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R_{\\pmb{p}}(\\tau)$ is defined as:", + "bbox": [ + 500, + 791, + 679, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 811, + 890, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It indicates that the latent image can be reconstructed by:", + "bbox": [ + 500, + 851, + 874, + 866 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\boldsymbol {p}} (\\tau) = 2 \\Delta \\tau \\frac {B _ {\\boldsymbol {p}} (\\tau)}{R _ {\\boldsymbol {p}} (\\tau)}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 869, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24979", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a spatially-varying latency $l_{\\mathbf{p}}(\\tau)$ during exposure time $[\\tau - \\Delta \\tau, \\tau + \\Delta \\tau]$ into the EDI map $R_{\\mathbf{p}}(\\tau)$ in Eq. (5):", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau - l _ {\\boldsymbol {p}} (\\tau)} ^ {t - l _ {\\boldsymbol {p}} (\\tau)} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 141, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, the event as well as the exposure time for each latent image are discrete. Suppose the blurry image $B$ compromises $N_{I}$ latent images, and $N_{j}$ events at pixel $\\pmb{p}$ are triggered during exposure time $[\\tau - l_{\\pmb{p}}(\\tau), t - l_{\\pmb{p}}(\\tau)]$ , the discrete formulation of Eq. (7) is:", + "bbox": [ + 76, + 185, + 468, + 261 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\sum_ {j = 1} ^ {N _ {I}} \\exp \\left(\\theta \\sum_ {k = 1} ^ {N _ {j}} \\sigma_ {k}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 266, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent image reconstructed from blurry image $B$ and event image $E$ can then be obtained by:", + "bbox": [ + 76, + 320, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {I} _ {\\boldsymbol {p}} (\\tau) = N _ {I} \\frac {B _ {\\boldsymbol {p}} (\\tau)}{\\widehat {R} _ {\\boldsymbol {p}} (\\tau)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 354, + 468, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\pmb{p}$ and $\\tau$ will be omitted in the following for simplicity.", + "bbox": [ + 76, + 393, + 437, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Event-based temporal fidelity", + "text_level": 1, + "bbox": [ + 76, + 417, + 339, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider events $\\mathbb{E} = \\{e_k\\}_{k=1}^N$ captured in the exposure duration of the blurry image $B$ . To achieve the goal of latency correction, a no-reference-based objective should be defined since the ground truth is hard to collect. To this end, we propose event-based temporal fidelity (ETF) $\\mathcal{L}_{\\mathrm{ETF}}$ to quantify the sharpness profile of latent images $I$ reconstructed from events with latency $l_p$ corrected and their corresponding blurry image $B$ , as shown in Figure 2 (a).", + "bbox": [ + 75, + 439, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuitively, $\\mathcal{L}_{\\mathrm{ETF}}$ should have smaller values as the estimated latency approaches its optimal values. Sharp images have more high-frequency contents compared with blurry images, which can be assessed by measuring the magnitude of its derivatives [3]. As shown in Figure 1 (d), the latent image results of EDI obtained from events with large latency appear with less sharpness and thicker edges. This motivates us to optimize the latency $l_{p}$ by enforcing regularizations of the sharpness profile of the latent image $I$ reconstructed from the EDI model in Eq. (9). Specifically, we define the average edge thickness of image gradients $\\nabla I_p = (\\nabla_x I_p, \\nabla_y I_p)$ as the number of pixels with values greater than a parameter $\\lambda$ ,", + "bbox": [ + 75, + 561, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\lambda} (I) = \\sum_ {\\boldsymbol {p}} H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 762, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H(\\cdot)$ is the Heaviside function.", + "bbox": [ + 76, + 800, + 326, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Integrating the right-side of the equation over $\\lambda$ with a decreasing function $\\rho$ , we can obtain an $\\lambda$ -irrelevant representation of $s(I)$ :", + "bbox": [ + 76, + 815, + 468, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (I) = \\int_ {0} ^ {\\infty} \\sum_ {\\boldsymbol {p}} \\rho (\\lambda) H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right) d \\lambda . \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 864, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e3cc03fe458577520879a89881bd4c164b3269ba8f2faf73105bd9413a3e9d9b.jpg", + "image_caption": [ + "(a) The proposed latency correction method" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 885, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d900b7fb95b2c3fd92a4739c71395346be410a7c31597e7c167d002d0fa54fa3.jpg", + "image_caption": [ + "(b) $\\mathcal{L}_{\\mathrm{ETF}}$ vs. constant $l$", + "Figure 2. Illustration of the proposed latency correction method. (a) The pipeline of the proposed method, which takes synchronized blurry image $B$ and events as input, and output the latency $l$ by polynomial function $l(L)$ . (b) The value of the proposed event temporal fidelity (ETF) $\\mathcal{L}_{\\mathrm{ETF}}$ decreases as the estimated latency approaches ground-truth latency, which validates the efficacy of ETF as an objective. (c) Latency-differentiable reformulation of event integration, enabling gradient-based optimization." + ], + "image_footnote": [], + "bbox": [ + 522, + 224, + 669, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1292622a7a85402013c098c1fe8c26335eea58d5527c597d09e62b0f381c9b89.jpg", + "image_caption": [ + "(c) Differentiable event integral" + ], + "image_footnote": [], + "bbox": [ + 694, + 226, + 864, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the function $\\rho (\\lambda) = e^{-\\lambda}$ in our experiments1. To pay more attention to the smoothness area, we multiply the image gradients $\\nabla I_{p}$ with an amplifier $\\epsilon$ .", + "bbox": [ + 496, + 465, + 890, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noted that, for the robustness to scenes with possible thick edges on the image, $\\epsilon$ can not be set too small, which brings challenges for detecting high-frequency details. To compensate for this, we define another term about the contrast of the latent image $I$ . Motivated by the fact that sharper images tend to have large contrast, we evaluate the contrast as the average local variation with local window $\\mathcal{N}(\\pmb{p})$ centered at pixel $\\pmb{p}$ [17]:", + "bbox": [ + 496, + 512, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (I) = \\sum_ {\\boldsymbol {p}} \\frac {1}{\\left| \\mathcal {N} _ {\\boldsymbol {p}} \\right|} \\sum_ {\\boldsymbol {q} \\in \\mathcal {N} _ {\\boldsymbol {p}}} \\left(I _ {\\boldsymbol {q}} - \\mu_ {I} \\left(\\mathcal {N} _ {\\boldsymbol {p}}\\right)\\right) ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 637, + 890, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_I(\\mathcal{N}_p)$ is the mean of the pixel value inside the window $\\mathcal{N}_p$ . The reason we use a local variation rather than a global one is that the blurry effects caused by motion are not always globally consistent. Maximizing the local variation can emphasize the local sharpness and reduce the blur for each local window. In practice, we adopt local standard deviation $\\sqrt{h(I)}$ in our experiments.", + "bbox": [ + 496, + 680, + 890, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For consecutive frames in a video, we can assume that their overall brightness level are similar, and thus the latency values are similar. Therefore, we define an indicator of temporal consistency by assuming Gaussian-distributed latency across consecutive frames. Given the latency of the last adjacent frame is $l'$ , the indicator defined on the latency", + "bbox": [ + 496, + 786, + 890, + 877 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 Please see the supplementary material for more details about $s(I)$", + "bbox": [ + 517, + 886, + 874, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24980", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$l$ of the current frame is formulated as:", + "bbox": [ + 76, + 90, + 334, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng (l) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {\\left(l - l ^ {\\prime}\\right) ^ {2}}{2}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 107, + 468, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, $\\mathcal{L}_{\\mathrm{ETF}}$ to determine optimal latency is defined as:", + "bbox": [ + 76, + 145, + 450, + 160 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E T F}} (I, l) = \\frac {s (\\epsilon I)}{\\sqrt {h (I)} g (l)}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 165, + 468, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The efficacy of the proposed $\\mathcal{L}_{\\mathrm{ETF}}$ is validated in Figure 2 (b), which shows that its value decreases as the latency estimation error decreases.", + "bbox": [ + 76, + 204, + 468, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Learning-based latency correction", + "text_level": 1, + "bbox": [ + 76, + 260, + 375, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the objective defined by $\\mathcal{L}_{\\mathrm{ETF}}$ , we introduce a self-supervised approach founded on a parametrized form of latency, and a reformulated event-based double integral model that is differentiable with respect to latency.", + "bbox": [ + 76, + 284, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latency-illuminance curve. The physical factors that have influences on the latency of events cannot be easily measured for the arbitrary situation. Fortunately, for deblurring and frame interpolation algorithms using event-image hybrid sensors such as [18, 19], the intensity of the image can indirectly indicate the illuminance levels in certain regions. Inspired by the previous findings about the relationship between latency and illuminance levels [7, 16, 21], we establish a latency-intensity relationship using a parameterized curve, which is validated in controlled environments (see illustrations in Figure 3 and discussions in Section 4.1 for more details, which shows that the latency increases as the illuminance decreases). To elucidate this relationship, we model the latency-illuminance curve as a $K$ -th order polynomial $\\psi$ of the illuminance $L$ at pixel $p$ :", + "bbox": [ + 76, + 349, + 468, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi (L) = \\sum_ {i = 0} ^ {K} a _ {i} \\cdot f (L) ^ {i}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 580, + 468, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{a_i\\}_{i = 0}^K$ are the parameters, $f(\\cdot)$ is a fixed function to map the illuminance $L$ into monotone decrement for consideration of easier optimization. The curve in Eq. (15) can effectively be a guidance for regularizing the estimation of spatially varying latency. In experiments, the illuminance $L$ is unavailable. We adopt the blurry image $B$ to approximate illuminance $L$ to estimate the spatially-varying latency, resulting in $l_{p} = \\psi (\\frac{B_{p}}{2\\Delta\\tau})$ .", + "bbox": [ + 76, + 627, + 468, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latency-differentiable event integral. It is noted that the latency $l_{p}$ derived from Eq. (9) is an index-like parameter, which has no derivatives. To allow gradient-based optimization, we establish a differentiable relationship between latency in events and specific indicators as shown in Figure 2 (c). By reformulating the event integral to be latency-differentiable, we derive a gradient chain linking $\\{a_{i}\\}_{i = 0}^{K}$ to $\\mathcal{L}_{\\mathrm{ETF}}$ via the EDI model with latency.", + "bbox": [ + 76, + 753, + 468, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to Eq. (3) where events $e_k = (\\pmb{p}_k, t_k, \\sigma_k)$ are represented by an impulse function, we employ a piece-wise constant function:", + "bbox": [ + 500, + 90, + 890, + 133 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\boldsymbol {p}} ^ {(k)} (s) = \\frac {\\sigma_ {k}}{t _ {k} - t _ {k - 1}}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right], \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 137, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the changes $\\sigma_{k}$ is equally split into the several duration. Then integration of events in duration $(t_{k - 1},t_k]$ becomes:", + "bbox": [ + 498, + 170, + 890, + 213 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\int E _ {\\boldsymbol {p}} ^ {(k)} (s) d s = \\sigma_ {k}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right]. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 217, + 890, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For simplicity, we denote $E_{\\pmb{p}}^{(k)}(s)$ as $E_{\\pmb{p}}^{(k)}$ . Consider the $N$ events $\\{e_k\\}_{k=1}^N$ triggered at pixel $\\pmb{p}$ during exposure time $[\\tau - l_p, t - l_p]$ where $t_0 < \\tau - l_p < t_1$ and $t_{N-1} < t - l_p < t_n$ , we can rewrite the inner integral part in Eq. (7) into:", + "bbox": [ + 498, + 253, + 890, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int_ {\\tau - l _ {p}} ^ {t - l _ {p}} E _ {\\pmb {p}} (s) d s \\\\ = \\int_ {\\tau - l _ {p}} ^ {t _ {1}} E _ {\\boldsymbol {p}} ^ {(1)} d s + \\sum_ {k = 2} ^ {N - 1} \\int_ {t _ {k - 1}} ^ {t _ {k}} E _ {\\boldsymbol {p}} ^ {(k)} d s + \\int_ {t _ {N - 1}} ^ {t - l _ {\\boldsymbol {p}}} E _ {\\boldsymbol {p}} ^ {(N)} d s \\\\ = \\left(t _ {1} - l _ {\\boldsymbol {p}} - \\tau\\right) \\sigma_ {1} + \\sum_ {k = 2} ^ {N - 1} \\sigma_ {k} + \\left(t - l _ {\\boldsymbol {p}} - t _ {N - 1}\\right) \\sigma_ {n}, \\tag {18} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 321, + 890, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which is differentiable in the neighborhood of $l_{p}$ with fixed $N$ . It is noted that $N$ will change when $\\tau - l_{p} < t_{0}$ or $t - l_{p} \\geq t_{N}$ , in which discontinuous changes on event triggering time will occur. While this might lead to a local minimum solution in optimization for a specific sample, it can be alleviated when optimizing over different samples.", + "bbox": [ + 498, + 468, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatially-varying latency correction. With the differentiable form of event integral for parameters $\\{a_i\\}_{i=0}^n$ , we can now perform gradient descent in the dataset by defining a loss function $\\mathcal{L}_{\\mathrm{total}}$ . Firstly, $\\mathcal{L}_{\\mathrm{ETF}}$ defined in Eq. (14) is used as a term of the loss function. Heuristically, most of the latency $l_p$ should lie in the range from 0 to the exposure time of the last blurry image, which can be optimized by the reformulated EDI. To ensure the latency does not exceed the previous exposure time so that the derivative can be calculated, a clipping operation is applied on $l_p$ . A boundary loss is also used to limit the range of latency:", + "bbox": [ + 498, + 565, + 890, + 732 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {b o u n d}} = \\sum_ {l _ {p} > 0} l _ {p} + \\sum_ {l _ {p} < \\tau - \\tau^ {\\prime} - \\Delta \\tau} - l _ {p}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 736, + 890, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To deal with $l_{p}$ exceeds the range of $[0, \\tau - \\tau' - \\Delta \\tau]$ and disables derivative computation due to gradients with too large magnitudes, we additionally optimize an intermediate constant latency $l_{\\mathrm{const}}$ shared across the sensor array during the exposure time of the blurry image. It has a relatively small solution space and can be solved efficiently by searching. $l_{\\mathrm{const}}$ provides an effective estimation of the overall latency (whose effectiveness is validated in the ex", + "bbox": [ + 498, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Please see the supplementary material for more details about $f(\\cdot)$", + "bbox": [ + 94, + 886, + 450, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24981", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "periment results on synthetic data in Section 4.2). Thus, it is being adopted in the term $\\mathcal{L}_{\\mathrm{mse}}$ to minimize the distance between the current estimated latency and $l_{\\mathrm{const}}$ . $l^{\\prime}$ in Eq. (13) is replaced by $l_{\\mathrm{const}}$ to regularize using the frame's constant latency, rather than adjacent frames'.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "From the latency-illuminance curve estimated previously, we can observe that the curve should be monotonic. Thus we enforce that the function $l_{p}$ is monotonically decreasing as the intensity increases. To apply the monotonicity constraint, we introduce the following constraint on the derivative of a function $l_{p}$ :", + "bbox": [ + 75, + 167, + 467, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d e v}} = \\sum \\max \\left(f ^ {\\prime} \\left(\\frac {B}{2 \\Delta \\tau}\\right) \\sum_ {i = 0} ^ {K} i \\cdot a _ {i} \\cdot f \\left(\\frac {B}{2 \\Delta \\tau}\\right) ^ {i - 1}, 0\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 263, + 467, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B$ contains different values on each pixel to ensure monotonically decreasing on all the data samples.", + "bbox": [ + 76, + 309, + 467, + 339 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, the loss function to be minimized is defined as:", + "bbox": [ + 96, + 340, + 464, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {\\text {E T F}} + \\beta \\mathcal {L} _ {\\text {b o u n d}} + \\eta \\mathcal {L} _ {\\text {m s e}} + \\zeta \\mathcal {L} _ {\\text {d e v}}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 361, + 467, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha, \\beta, \\eta, \\zeta$ are hyper-parameters for balancing different terms. Empirically, we set $\\alpha = 1, \\beta = 1, \\eta = 20, \\zeta = 0.1$ in our experiments.", + "bbox": [ + 76, + 383, + 467, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Training details", + "text_level": 1, + "bbox": [ + 76, + 440, + 233, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset preparation. For latency correction training, we utilized a DAVIS346 event camera [1] to obtain blurry images and corresponding events. Our real-scene dataset comprises 9 videos, where 100 frames and their neighboring events from each video are selected for training.", + "bbox": [ + 75, + 468, + 467, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other implementation details. To simplify the implementation, we stack events into 500 bins to maintain high temporal resolution. Our model is implemented on the Pytorch framework and runs on an NVIDIA GeForce RTX 3090 GPU. We use ADAM [11] with default parameter setting for optimization. The batch size is set as 1. The learning rate was fixed to $10^{-5}$ in the earlier 2 epochs and reduced to 0 in the last 3 epochs by the linear decay strategy. Constant initialization to zero for the non-latency assumption is used for initialization.", + "bbox": [ + 75, + 547, + 467, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 714, + 207, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Latency estimation in controlled environments", + "text_level": 1, + "bbox": [ + 76, + 739, + 467, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments in controlled settings measure actual latency and illuminance. Our setup, shown in Figure 3 (a), includes a light source, an Arduino board, a power source, a photometer, and an event camera (DVS346Mono) with its lens removed to prevent light attenuation. The Arduino board was programmed to switch off the light at a known time $t_{\\mathrm{ideal}}$ , and the subsequent OFF events were assumed to be triggered by this change. The latency of the $i$ -th triggered event was calculated as the difference between $t_{\\mathrm{ideal}}$", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1c520d6d45ffee476ee0e94ca90bd7056a98031ed6c0d3ea1fcabcdc20711f5.jpg", + "image_caption": [ + "(a) Latency estimation system" + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 705, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76b93f404960be0820c8fe1142643c2266b878e7112be868a58bcea08795fc24.jpg", + "image_caption": [ + "(b) Event trigger rate vs. time" + ], + "image_footnote": [], + "bbox": [ + 707, + 89, + 888, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01f09bed8f57f760866a5b8665d4fdd04d551275aead5e5cd2aefabcdcd5cd48.jpg", + "image_caption": [ + "(c) Latency vs. Illuminance", + "Figure 3. Estimation of latency-illuminance curve. (a) A system designed to investigate the relationship between illuminance levels and latency, approximated using an $K$ -th order polynomial. (b) The comparison of event triggering rate under three varied lighting conditions. It is approximated by the first $1\\%$ OFF events in our experiments. (c) Estimated latency under controlled lighting, highlighting increased latency with diminishing light. (d) Results of the polynomial approximation." + ], + "image_footnote": [], + "bbox": [ + 506, + 205, + 705, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb6499a94a2293af0b518513eb87c9b9a4cdd73bdfd20b0f7e87c23d8bd49ed5.jpg", + "image_caption": [ + "(d) Latency curve from Eq. (2)" + ], + "image_footnote": [], + "bbox": [ + 707, + 205, + 888, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the time $t_{\\mathrm{actual}}^{(i)}$ . Illuminance is modulated by adjusting light voltage and measured alongside events by a photometer (with the lowest detectable value being 0.1 Lux). This process was repeated to mitigate noise effects (about 80 times in our experiments). The latency level of a given illuminance was calculated by the event triggering rate as shown in Fig. 3 (b), which is approximated by the median latency when the first $1\\%$ OFF events among all the samples are triggered for each process. By varying the illuminance, we empirically find that the latency increases as the illuminance decreases, as shown in Fig. 3 (c). We use the proposed parameterized curve to approximate the relationship between latency and illuminance, whose results are shown in Fig. 3 (d). It demonstrates an accurate curve fitting with $K = 5$ . Due to the difficulty in illuminance measuring, the estimated parameter cannot be directly applied in deblurring and interpolation for arbitrary scenes.", + "bbox": [ + 496, + 465, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on synthetic data", + "text_level": 1, + "bbox": [ + 500, + 734, + 754, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. As the proposed method serves as a preprocessing step for latency correction of events, we evaluate the proposed methods based on two downstream tasks, that is, deblurring, and interpolation. Specifically, we test our method for deblurring algorithms including EDI [19], EFNet [24], and GEM [31], and for interpolation algorithms including EDI [19] and REFID [25]. We directly use the released pretrain model for comparison.", + "bbox": [ + 496, + 761, + 890, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To qualitatively evaluate downstream tasks, we", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24982", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ffb1dcda35a65fdd0939c2e61c8a94c0395d411c7e41bb5df202ec89b259d61.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on synthetic data. We correct the latency of events by the constant latency correction (\"W/ corr. const.\") method and compared it with the ground truth events (GT input). All the methods are tested with the same input." + ], + "image_footnote": [], + "bbox": [ + 80, + 75, + 890, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1483ae6e1e03041fc736740e4272eeef6dc1f93c48314386e8895f0a4d4061f.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation on synthetic data. $\\uparrow (\\downarrow)$ means higher (lower) is better. The \"Diff\" shows the percentage improvement from \"W/o corr.\" to \"W/ corr. const.\" compared with \"GT input\"." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓
EDI [19]GT input28.4890.8390.489
W/o corr.26.0010.8100.500
DeblurringW/ corr. const.27.4710.8280.488
Diff. (%)+5.15+2.15-2.45
EFNet [24]GT input33.7610.9720.105
W/o corr.33.3300.9710.107
W/ corr. const.33.7780.9710.106
Diff. (%)+0.130-0.95
GEM [31]GT input30.6110.9630.183
W/o corr.27.7140.9340.198
W/ corr. const.30.3720.9580.185
Diff. (%)+8.68+2.49-7.10
EDI [19]GT input22.2930.8400.330
InterpolationW/o corr.22.1250.8340.331
W/ corr. const.22.2650.8380.331
Diff. (%)+0.63+0.470
REFID [25]GT input34.5220.9740.109
W/o corr.28.2440.9270.142
W/ corr. const.34.2170.9700.111
Diff. (%)+17.30+4.41-28.44
", + "bbox": [ + 78, + 483, + 468, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "synthesize events with constant latency $l_{\\tau}$ for each pair of events and its corresponding blurry image. In particular, we utilize the event provided in the HighREV dataset [25] as ideal events, and then we add a constant latency to the event-image pairs for different event sequences. We employ", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the training data in the HighREV dataset which contains 1615 frames from 19 videos for testing to provide more convincing results. The spatially-varying latency correction is highly related to the properties of event cameras and depends on different scene illuminances, which can not be faithfully synthetic by event simulators [7, 10]. Therefore, we only provide results of the constant latency estimated by the proposed methods for quantitative comparisons.", + "bbox": [ + 496, + 424, + 890, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. To show the corrected performance, all the tested methods are evaluated with four kinds of input, events without latency (terms \"GT input\"), events with latency which is not corrected by our method (terms \"W/o corr.\"), the events corrected by constant latency $l_{\\mathrm{const}}$ (terms \"W/ corr. const.). We adopt three commonly used metrics to measure the performance of deblurring and interpolation, peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and the perceptual error with learned perceptual image patch similarity (LPIPS) [29], respectively.", + "bbox": [ + 496, + 547, + 892, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. The quantitative evaluation results are shown in Table 1. We add the \"Diff.\" row to show the improvement of the proposed method. It shows that the proposed method can improve the performance of deblurring and interpolation by event latency correction. For deep learning methods like EFNet [24], they implicitly considered latency effects in training data generation, which makes them more robust to latency in events. However, their performance can also be improved with the latency correction. The qualitative evaluation results are shown in Figure 4, which show that the proposed method can explicitly reduce the artifacts brought by latency, and it performs more closely to the ground truth events (GT input). Specifically, the latency correction can", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24983", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/206ec58a7943040845c5eb6fe3849466592b9cb2143602b926df79a602de296e.jpg", + "image_caption": [ + "Figure 5. The qualitative comparison on real data. We correct the latency of events by constant latency correction (\"W/ corr. Const.\") and spatially-varying latency correction (\"W/ corr\"). All the methods are tested with the same input." + ], + "image_footnote": [], + "bbox": [ + 76, + 77, + 890, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reduce the light strips brought by previous edges, in which the events are triggered with latency. The non-learning interpolation algorithm EDI [19] is more strict with the time accuracy and the noise in events, which makes it still blurry even with the GT input.", + "bbox": [ + 75, + 411, + 470, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Evaluation on real data", + "text_level": 1, + "bbox": [ + 76, + 494, + 294, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In order to reveal the performance of the proposed method, we evaluate our method on real data captured by DAVIS346Mono. Because the ground truth sharp images are unavailable, we can only qualitatively evaluate the proposed method on real data. The results on real data $^2$ are shown in the left two columns of Figure 5. It shows that both the proposed latency correction methods can reduce the artifacts and blurry effects. Processed by the proposed latency correction methods, the number of inaccuracy events was explicitly reduced to help the deblurring and interpolation algorithms to produce natural and sharp edges.", + "bbox": [ + 75, + 517, + 468, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 690, + 227, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the performance of spatially-varying latency correction with constant latency correction on real data as shown in the right two columns in Figure 5. It shows that events corrected by the spatially-varying method have more natural views than constant latency correction, which can also provide sharper results3. Benefiting from the data captured under different illumination conditions, the spatially varying latency correction can also provide more accurate latency estimation for specific scenes. Along with the different illumination conditions, it can also address the different latency properties for different pixels.", + "bbox": [ + 75, + 713, + 468, + 878 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 409, + 619, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present the first method for latency correction in improving event-guided deblurring and interpolation tasks. Our approach introduces event-based temporal fidelity to evaluate the sharpness of latent images reconstructed from latency-corrected events and their corresponding blurry images. This objective guides the optimization of parameters in a parametrized latency model through gradient-based optimization, facilitated by a reformulated, latency-differentiable, event-based double integral model. The efficacy of our methods is demonstrated using both synthetic and real-world data sets.", + "bbox": [ + 496, + 435, + 890, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although with potential for broad applicability across different event cameras, the present implementation of the proposed method is specifically tailored for DAVIS346Mono cameras due to that it is the only event camera we have with on-chip synchronization and our current method requires perfect temporal synchronization between the image and events. Moreover, due to challenges in measuring illuminance and curve variability, the practicality of estimated parameters of the latency-illuminance curve varies across sensors. The curves for more sensors, while significant, has not been explored in this paper.", + "bbox": [ + 496, + 612, + 892, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 784, + 660, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803), Beijing Natural Science Foundation (Grant No. L233024), National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), and SenseTime Collaborative Research Grant.", + "bbox": [ + 496, + 799, + 890, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3More comparisons are provided in the supplementary material.", + "bbox": [ + 94, + 886, + 433, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24984", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 2014. 1, 2, 6", + "[2] Tobi Delbruck and Carver A Mead. Adaptive photoreceptor with wide dynamic range. In Proc. of IEEE International Symposium on Circuits and Systems-ISCAS, pages 339–342, 1994. 2", + "[3] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus Is All You Need: Loss Functions for Event-Based Vision. In Proc. of Computer Vision and Pattern Recognition, 2019. 4", + "[4] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26:1272-1282, 2004. 1", + "[5] Chen Haoyu, Teng Minggui, Shi Boxin, Wang Yizhou, and Huang Tiejun. Learning to Deblur and Generate High frame rate video with an event camera. arXiv preprint arXiv:2003.00847, 2020. 1", + "[6] Ondrej Holevský, Radoslav Skoviera, Václav Hlaváč, and Roman Vitek. Experimental Comparison between Event and Global Shutter Cameras. Sensors, 21(4), 2021. 2, 3", + "[7] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic DVS events. In Proc. of Computer Vision and Pattern Recognition Workshops, 2021. 2, 3, 5, 7", + "[8] invitation. Understanding the Performance of Neuromorphic Event-based Vision Sensors. https://inivation.com/wp-content/uploads/2020/05/White-Paper-May-2020.pdf, 2020.1, 2", + "[9] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning Event-Based Motion Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2020. 2", + "[10] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, 15:702765, 2021. 2, 7", + "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[12] Juan Antonio Lénero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 $\\mu$ s Latency Asynchronous Frame-Free Event-Driven Dynamic-Vision-Sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 1, 2", + "[13] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1, 2, 3", + "[14] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In Proc. of European Conference on Computer Vision, 2020. 2", + "[15] Brian McReynolds, Rui Graca, and Tobi Delbruck. Experimental methods to predict dynamic vision sensor event" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "camera performance. Optical Engineering, 61(7):074103-074103, 2022. 1, 2", + "[16] Brian J Mcreynolds, Rui Graca, Daniel O'Keefe, Rachel Oliver, Richard Balthazor, Nathaniel George, and Matthew McHarg. Modeling and decoding event-based sensor lightning response. In Proc. of Unconventional Imaging, Sensing, and Adaptive Optics 2023, 2023. 1, 2, 5", + "[17] Patrenahalli M. Narendra and Robert C. Fitch. Real-time Adaptive contrast enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, (6):655-661, 1981. 4", + "[18] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proc. of Computer Vision and Pattern Recognition, 2019. 2, 3, 5", + "[19] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2, 3, 5, 6, 7, 8", + "[20] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2, 3", + "[21] Jonah P Sengupta. Demystifying Event-Based Camera Latency: Sensor Speed Dependence on Pixel Biasing, Light, and Spatial Activity. In Proc. of SPIE Conference 13045: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXXV, 2023. 1, 2, 5", + "[22] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. A $128 \\times 128$ $1.5\\%$ Contrast Sensitivity $0.9\\%$ FPN $3\\mu s$ Latency $4\\mathrm{mW}$ Asynchronous Frame-Free Dynamic Vision Sensor Using Transimpedance Preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 2", + "[23] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proc. of International Conference on Computer Vision, 2021. 2", + "[24] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based Fusion for Motion Deblurring with Cross-modal Attention. In Proc. of European Conference on Computer Vision, 2022. 1, 2, 6, 7, 5", + "[25] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based Frame Interpolation with Ad-Hoc Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 6, 7, 5", + "[26] Minggui Teng, Chu Zhou, Hanyue Lou, and Boxin Shi. NEST: Neural event stack for event-based image enhancement. In Proc. of European Conference on Computer Vision, 2022. 1, 2", + "[27] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In Proc. of European Conference on Computer Vision, 2020. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24985", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion Deblurring with Real Events. In Proc. of International Conference on Computer Vision, pages 2583-2592, 2021. 2", + "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of Computer Vision and Pattern Recognition, 2018. 7", + "[30] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proc. of Computer Vision and Pattern Recognition, 2022. 1, 2, 3", + "[31] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In Proc. of International Conference on Computer Vision, 2023. 2, 3, 6, 7, 5", + "[32] Chu Zhou, Minggui Teng, Jin Han, Jinxiu Liang, Chao Xu, Gang Cao, and Boxin Shi. Deblurring Low-Light Images with Events. International Journal of Computer Vision, 131 (5):1284-1298, 2023. 1" + ], + "bbox": [ + 78, + 90, + 468, + 358 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24986", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_model.json b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..afb6c5afedbea3d29bf8a1a0dbda5077f924b478 --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Latency Correction for Event-guided Deblurring and Frame Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.845, + 0.2 + ], + "angle": 0, + "content": "Yixin Yang\\(^{1,2,\\dagger}\\) Jinxiu Liang\\(^{1,2}\\) Bohan Yu\\(^{1,2}\\) Yan Chen\\(^{3}\\) Jimmy S. Ren\\(^{3}\\) Boxin Shi\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.201, + 0.829, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\) National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.218, + 0.884, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University \\(^{3}\\) SenseTime Research \\(\\{\\text { yangyixin93,cssherryliang,ybh1998,shiboxin }\\} @\\) pku.edu.cn yanchenace@gmail.com rensijie@sensetime.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.622 + ], + "angle": 0, + "content": "Event cameras, with their high temporal resolution, dynamic range, and low power consumption, are particularly good at time-sensitive applications like deblurring and frame interpolation. However, their performance is hindered by latency variability, especially under low-light conditions and with fast-moving objects. This paper addresses the challenge of latency in event cameras — the temporal discrepancy between the actual occurrence of changes in the corresponding timestamp assigned by the sensor. Focusing on event-guided deblurring and frame interpolation tasks, we propose a latency correction method based on a parameterized latency model. To enable data-driven learning, we develop an event-based temporal fidelity to describe the sharpness of latent images reconstructed from events and the corresponding blurry images, and reformulate the event-based double integral model differentiable to latency. The proposed method is validated using synthetic and real-world datasets, demonstrating the benefits of latency correction for deblurring and interpolation across different lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.654, + 0.21, + 0.67 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Emerging as a groundbreaking visual sensing paradigm, event cameras [1, 12, 13, 20, 22] provide a novel perspective to capturing and interpreting dynamic scenes. They have advantages of high temporal resolution (the discrete measurement resolution of the event detection time, in the order of microseconds [8]), high dynamic range (e.g., 120 dB for DAVIS346), and low power consumption. They asynchronously record changes in logarithmic pixel intensity only when they exceed a predefined threshold. Their unique characteristics have led to impressive performance gains in various time-sensitive applications, including deblurring [5, 19, 24, 26, 32] and frame interpolation [19, 25, 30]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.286, + 0.714, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.288, + 0.89, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.363, + 0.889, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.432, + 0.893, + 0.557 + ], + "angle": 0, + "content": "Figure 1. Illustration of latency in an event camera, and its influence on an event-guided deblurring algorithm. (a) The aerial view of ideal vs. real events triggered by a black point moving upwards on a white background, where the real events are captured in low-light conditions (below 0.1 Lux). (b) Deblurring results using the Event-based Double Integral (EDI) [19] method, guided by events captured under different illumination conditions. (c) Improved deblurring results from EDI [19] using event with latency corrected by the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.893, + 0.824 + ], + "angle": 0, + "content": "The ideal event model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this, exhibiting a temporal discrepancy between the actual moment of intensity change and the corresponding event timestamp. As claimed by the inventors of the event-based sensors, in well-lit environments above 1k Lux, event cameras can achieve latency as low as \\(3\\mu \\mathrm{s}\\) [22]. However, in dimmer conditions for a few Lux, the latency can exceed 1 ms [22]. The challenge in modeling this temporal discrepancy arises from the sensor pixels' non-linear and variable response to illumination, which is further complicated by manufacturing imperfection, inherent electronic noise, and the unpredictable nature of real-world scenes [13, 15, 16, 21]. Therefore, the profile of the temporal discrepancy is spatial-varying, inconsistent, and noisy. All these complexities make algorithms for handling them scarce." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Without specifying the nature of the temporal discrepancy, we consider the collective temporal discrepancies in the timing of the camera's response to real-world changes, a phenomenon we term \"latency\", where the aforementioned sensor-level latency is one of the dominant components." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.411, + 0.888 + ], + "angle": 0, + "content": "This work is done during Yixin's internship at SenseTime." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.888, + 0.342, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author: shiboxin@pku.edu.cn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.411, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24977" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.228 + ], + "angle": 0, + "content": "Despite the complexity of the latency formation model, latency presents a strong relation with illuminance. As illustrated in Figure 1 (a), compared with the ideal events generated instantaneously when the intensity changes, the real events suffer from latency in illuminance below 0.1 Lux. The latency presents a significant challenge for time-sensitive applications [19, 24, 25, 30], which assume ideal event triggering and struggle to restore sharp edges due to the severe latency, as demonstrated in Figure 1 (b) vs. (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.471, + 0.385 + ], + "angle": 0, + "content": "In this paper, we focus on latency correction for event-guided imaging deblurring [19, 24, 28] and interpolation [19, 25, 31] in low-light and fast-moving scenes, where event cameras should play an important role. Based on the analysis and modeling of latency, we propose to predict latency in event cameras by conceptualizing latency as a variable function—potentially a constant for a specific scene or a polynomial function of the illuminance level, given a fixed event camera type—we endeavor to elucidate the relationship between latency and illuminance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.616 + ], + "angle": 0, + "content": "Specifically, we propose a latency correction method for captured events and blurry images in an event-image hybrid sensor (e.g., DAVIS346). Inspired by the latency properties related to illuminance observed through controlled experiments, we characterize the latency-intensity relationship using a parameterized curve. To estimate the optimal latency, we introduce a novel objective to assess the sharpness profile of a latent image reconstructed from the latency-corrected events and the blurry image, which we term event-based temporal fidelity. Building upon them, we collect data under different scene illuminance and advocate for a self-supervised approach to latency correction. The EDI model [18, 19] is reformulated to be differentiable for latency to enable gradient-based optimization, therefore enhancing the performance of deblurring and interpolation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.62, + 0.47, + 0.696 + ], + "angle": 0, + "content": "The efficacy of the proposed method is validated using both synthetic and real data, demonstrating improvements in image quality for deblurring and interpolation under challenging lighting conditions where the latency in events becomes severe. The key contributions of our work include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.7, + 0.468, + 0.745 + ], + "angle": 0, + "content": "- the first data-driven latency correction method grounded in a differentiable reformulation of the event integral with respect to latency;" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.746, + 0.468, + 0.79 + ], + "angle": 0, + "content": "- the event-based temporal fidelity for measuring the sharpness profile of a latent image reconstructed by latency-corrected events and the corresponding blurry image; and" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.792, + 0.468, + 0.835 + ], + "angle": 0, + "content": "- a parameterized latency-intensity curve for modeling the relationship of latency in events and the intensity of the corresponding blurry image." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.7, + 0.468, + 0.835 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The proposed method provides an approach to understanding and mitigating the effects of latency, paving the way for more accurate and reliable event-based imaging in dynamic environments with challenging lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.646, + 0.107 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.118, + 0.892, + 0.179 + ], + "angle": 0, + "content": "We investigate the influence of latency on imaging algorithms in event-image hybrid systems, particularly focusing on deblurring and interpolation tasks due to their sensitivity to timing in events." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.685 + ], + "angle": 0, + "content": "Analysis of latency in event cameras. The triggering of events involves three subprocesses: photon reception, event trigger, and event readout. Notably, latency primarily arises during photon reception and event readout due to the electronic and physical design of these cameras [8]. Latency characteristics vary among different event camera models, ranging from minimal latency of \\(15\\mu s\\) [13] to \\(4\\mu s\\) [20] under 1k Lux conditions. Despite improvements in camera design, latency persists at nearly 1 ms under 1 Lux [1], posing challenges in low-light scenarios. Existing studies predominantly focus on sensor design aspects of event latency [1, 12, 13, 22] with a limited number examining latency under varying illuminance levels [6, 21]. In low-light conditions, latency induced by photoreceptor bandwidth becomes predominant [16], which is proportional to photocurrent (or illuminance) at low illuminance level, increasing monotonically with the illuminance level [2, 13, 15, 16, 21]. Consequently, the latency decreases as illuminance increases [7, 13, 21]. The photoreceptor circuit's low-pass characteristic [15, 16], has led to its modeling as a first- or second-order low-pass filter in various event simulators [7, 10]. While this provides a simplified relationship between latency and photocurrent, accurately determining photocurrent is often invalid for arbitrary imaging setups due to the complex imaging processes involving shuttering, apertures, etc. In order to save costs, most event cameras timestamp each event as they are being read out from the pixel array [8], which introduces readout latency that is significantly influenced by sensitivity and contrast [2, 8, 10, 15, 21]. The complexity of these factors contributing to latency in event cameras explains the scarcity of algorithms specifically designed to address this challenge for event-image hybrid systems." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Event guided imaging algorithms. Event cameras record changes during the exposure time of blurry images, which makes the problem of deblurring and interpolation more well-posed. Pan et al. [18, 19] introduced the EDI model for high-frame-rate video reconstruction from blurry videos using event data, applicable in both deblurring and interpolation. For deblurring, Jiang et al. [9] combined convolutional recurrent networks with directional event filtering, Shang et al. [23] utilized the nearest sharp frames and events for deblurring, Xu et al. [28] improved model generalization through self-supervised learning, Teng et al. [26] focused on event representations for image enhancement, and Sun et al. [24] developed an event-image fusion module. For joint deblurring and interpolation, Lin et al. [14] employed neu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24978" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "ral networks and dynamic filtering, while Wang et al. [27] and Zhang et al. [30, 31] integrated multiple techniques, including event-enhanced sparse learning and self-supervised frameworks. Sun et al. [25] proposed a bidirectional recurrent network for fusion tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.471, + 0.274 + ], + "angle": 0, + "content": "For these algorithms assuming ideal event triggering, the reconstruction of the latent image between or within exposure time heavily depends on the precision of timestamps in accompanying events. Latency presence can significantly impair algorithm performance. To the best of our knowledge, these has been no approach to address latency for event-guided deblurring and interpolation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.251, + 0.304 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.507 + ], + "angle": 0, + "content": "In this paper, we aim to tackle the latency in events for event-guided deblurring and frame interpolation. Section 3.1 delves into the deviation of events with latency from the ideal event triggering mechanism and the event-based double integral model with latency. Based on this model, we propose a learning-based latency correction method for event-guided deblurring and frame interpolation. Section 3.2 introduces the objective for optimization, the event-based temporal fidelity. To allow gradient-based optimization, Section 3.3 reformulates event-based double integral to be differentiable respective to latency, where the latency is further parameterized into a polynomial of intensity on a per-pixel basis. Training details are shown in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.362, + 0.531 + ], + "angle": 0, + "content": "3.1. Analysis on the latency in events" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.752 + ], + "angle": 0, + "content": "Latency in event triggering. In an event camera, the photons hitting the photodetector of a pixel are transduced into an electrical signal and then compared against a predefined threshold to determine whether the change in intensity is substantial enough to trigger an \"event\". If this threshold is surpassed, the event is recorded with a high-resolution timestamp. Mathematically, the pixels of an event camera independently respond to variations in the continuous brightness signal \\( I_{p}(t) \\) at pixel position \\( \\boldsymbol{p} = (p_x,p_y)^\\top \\) and time \\( t \\). An event \\( e = (\\boldsymbol{p},t,\\sigma) \\) is initiated at pixel \\( \\boldsymbol{p} \\) and time \\( t \\) with polarity \\( \\sigma \\in \\{+1, - 1\\} \\) indicating ON or OFF events when the logarithmic change of \\( I_{p} \\) exceeds a dispatched threshold \\( \\theta \\) since the last event triggered at the pixel \\( \\boldsymbol{p} \\) and time \\( t_{\\mathrm{ref}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.758, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\log I _ {\\boldsymbol {p}} (t) - \\log I _ {\\boldsymbol {p}} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\theta . \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.855 + ], + "angle": 0, + "content": "In tasks involving image reconstruction where per-pixel intensity is needed, the logarithmic intensity changes for continuous time neighboring recorded timestamps are often approximated as an equality involving the threshold value \\(\\theta\\) and polarity \\(\\sigma\\) with the quantization errors ignored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The ideal model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this ideal, ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.364 + ], + "angle": 0, + "content": "hibiting a temporal latency \\( l \\) between the actual change in intensity and the event's triggering. The presence of latency exacerbates the approximation error of the intensity changes (e.g., Eq. (2) in the following), particularly affecting time-sensitive tasks like deblurring and frame interpolation. In environments with controlled lighting and motion, empirical methods have been employed to study and characterize latency in event cameras, such as using a second-order low-pass filter [13, 20] with parameters that are influenced by illumination conditions [6, 7, 13]. In the unpredictable flux of real-world settings, latency becomes markedly more unpredictable and defies precise modeling. The complexity arises from the non-linear and non-uniform factors. For instance, pixel responses across the sensor array exhibit variability due to manufacturing differences, and the light received by each pixel fluctuates rapidly with movement in the scene or changes in illumination conditions. Moreover, the existence of noise exacerbates the difficulty." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.892, + 0.534 + ], + "angle": 0, + "content": "Event-based double integral model with latency. An event-image hybrid sensor (e.g., DAVIS346) is able to capture pixel-wise aligned events and images, synergistically harnessing the strengths of both modalities. This integration involves vibrant color representation characteristic of frame-based sensors, coupled with the high temporal resolution and dynamic range offered by event cameras. The mathematical foundation for the interplay between events and images in such a system is encapsulated by the EDI model, which is originally proposed for applications in deblurring and frame interpolation [18, 19]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.535, + 0.892, + 0.58 + ], + "angle": 0, + "content": "The proportional change for images at time \\(\\tau\\) and time \\(t\\) can be measured by the integral of the \\(N\\) events \\(\\mathbb{E} = \\{e_k\\}_{k=1}^N\\) triggered from \\(\\tau\\) to \\(t\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.585, + 0.892, + 0.62 + ], + "angle": 0, + "content": "\\[\nI _ {\\boldsymbol {p}} (t) = I _ {\\boldsymbol {p}} (\\tau) \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.624, + 0.808, + 0.64 + ], + "angle": 0, + "content": "where \\(E_{\\pmb{p}}(s)\\) is a function of continuous time:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.644, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\nE _ {\\boldsymbol {p}} (s) = \\sigma_ {k} \\delta_ {t _ {k}} (s), \\quad s \\in (t _ {k - 1}, t _ {k} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.667, + 0.892, + 0.756 + ], + "angle": 0, + "content": "whenever there is an event \\( e_k = (\\pmb{p}_k, t_k, \\sigma_k) \\) triggered and \\( \\delta_{t_k}(s) \\) is an impulse function at time \\( t_k \\) with unit integral. A blurry image \\( B_p(\\tau) \\) with exposure duration \\( t \\in [\\tau - \\Delta \\tau, \\tau + \\Delta \\tau] \\) can be represented as the average of the latent image \\( I_p(t) \\) over the exposure duration [18], which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.759, + 0.892, + 0.788 + ], + "angle": 0, + "content": "\\[\nB _ {\\boldsymbol {p}} (\\tau) = \\frac {1}{2 \\Delta \\tau} I _ {\\boldsymbol {p}} (\\tau) R _ {\\boldsymbol {p}} (\\tau), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.792, + 0.681, + 0.807 + ], + "angle": 0, + "content": "where \\(R_{\\pmb{p}}(\\tau)\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.812, + 0.892, + 0.848 + ], + "angle": 0, + "content": "\\[\nR _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.852, + 0.875, + 0.867 + ], + "angle": 0, + "content": "It indicates that the latent image can be reconstructed by:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.871, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nI _ {\\boldsymbol {p}} (\\tau) = 2 \\Delta \\tau \\frac {B _ {\\boldsymbol {p}} (\\tau)}{R _ {\\boldsymbol {p}} (\\tau)}. \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24979" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "We introduce a spatially-varying latency \\( l_{\\mathbf{p}}(\\tau) \\) during exposure time \\( [\\tau - \\Delta \\tau, \\tau + \\Delta \\tau] \\) into the EDI map \\( R_{\\mathbf{p}}(\\tau) \\) in Eq. (5):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.142, + 0.47, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau - l _ {\\boldsymbol {p}} (\\tau)} ^ {t - l _ {\\boldsymbol {p}} (\\tau)} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.186, + 0.469, + 0.262 + ], + "angle": 0, + "content": "In practice, the event as well as the exposure time for each latent image are discrete. Suppose the blurry image \\(B\\) compromises \\(N_{I}\\) latent images, and \\(N_{j}\\) events at pixel \\(\\pmb{p}\\) are triggered during exposure time \\([\\tau - l_{\\pmb{p}}(\\tau), t - l_{\\pmb{p}}(\\tau)]\\), the discrete formulation of Eq. (7) is:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.267, + 0.469, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\sum_ {j = 1} ^ {N _ {I}} \\exp \\left(\\theta \\sum_ {k = 1} ^ {N _ {j}} \\sigma_ {k}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.469, + 0.352 + ], + "angle": 0, + "content": "The latent image reconstructed from blurry image \\(B\\) and event image \\(E\\) can then be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.356, + 0.469, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\widehat {I} _ {\\boldsymbol {p}} (\\tau) = N _ {I} \\frac {B _ {\\boldsymbol {p}} (\\tau)}{\\widehat {R} _ {\\boldsymbol {p}} (\\tau)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.395, + 0.438, + 0.411 + ], + "angle": 0, + "content": "\\(\\pmb{p}\\) and \\(\\tau\\) will be omitted in the following for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.418, + 0.34, + 0.435 + ], + "angle": 0, + "content": "3.2. Event-based temporal fidelity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.469, + 0.562 + ], + "angle": 0, + "content": "We consider events \\(\\mathbb{E} = \\{e_k\\}_{k=1}^N\\) captured in the exposure duration of the blurry image \\(B\\). To achieve the goal of latency correction, a no-reference-based objective should be defined since the ground truth is hard to collect. To this end, we propose event-based temporal fidelity (ETF) \\(\\mathcal{L}_{\\mathrm{ETF}}\\) to quantify the sharpness profile of latent images \\(I\\) reconstructed from events with latency \\(l_p\\) corrected and their corresponding blurry image \\(B\\), as shown in Figure 2 (a)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.759 + ], + "angle": 0, + "content": "Intuitively, \\(\\mathcal{L}_{\\mathrm{ETF}}\\) should have smaller values as the estimated latency approaches its optimal values. Sharp images have more high-frequency contents compared with blurry images, which can be assessed by measuring the magnitude of its derivatives [3]. As shown in Figure 1 (d), the latent image results of EDI obtained from events with large latency appear with less sharpness and thicker edges. This motivates us to optimize the latency \\(l_{p}\\) by enforcing regularizations of the sharpness profile of the latent image \\(I\\) reconstructed from the EDI model in Eq. (9). Specifically, we define the average edge thickness of image gradients \\(\\nabla I_p = (\\nabla_x I_p, \\nabla_y I_p)\\) as the number of pixels with values greater than a parameter \\(\\lambda\\)," + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.763, + 0.469, + 0.797 + ], + "angle": 0, + "content": "\\[\ns _ {\\lambda} (I) = \\sum_ {\\boldsymbol {p}} H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.801, + 0.328, + 0.815 + ], + "angle": 0, + "content": "where \\(H(\\cdot)\\) is the Heaviside function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.816, + 0.469, + 0.861 + ], + "angle": 0, + "content": "Integrating the right-side of the equation over \\(\\lambda\\) with a decreasing function \\(\\rho\\), we can obtain an \\(\\lambda\\)-irrelevant representation of \\(s(I)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.866, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\ns (I) = \\int_ {0} ^ {\\infty} \\sum_ {\\boldsymbol {p}} \\rho (\\lambda) H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right) d \\lambda . \\tag {11}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.887, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.208, + 0.821, + 0.22 + ], + "angle": 0, + "content": "(a) The proposed latency correction method" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.226, + 0.671, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.322, + 0.653, + 0.334 + ], + "angle": 0, + "content": "(b) \\(\\mathcal{L}_{\\mathrm{ETF}}\\) vs. constant \\(l\\)" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.227, + 0.865, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.321, + 0.866, + 0.333 + ], + "angle": 0, + "content": "(c) Differentiable event integral" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed latency correction method. (a) The pipeline of the proposed method, which takes synchronized blurry image \\( B \\) and events as input, and output the latency \\( l \\) by polynomial function \\( l(L) \\). (b) The value of the proposed event temporal fidelity (ETF) \\( \\mathcal{L}_{\\mathrm{ETF}} \\) decreases as the estimated latency approaches ground-truth latency, which validates the efficacy of ETF as an objective. (c) Latency-differentiable reformulation of event integration, enabling gradient-based optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.514 + ], + "angle": 0, + "content": "We adopt the function \\(\\rho (\\lambda) = e^{-\\lambda}\\) in our experiments1. To pay more attention to the smoothness area, we multiply the image gradients \\(\\nabla I_{p}\\) with an amplifier \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.513, + 0.892, + 0.634 + ], + "angle": 0, + "content": "It is noted that, for the robustness to scenes with possible thick edges on the image, \\(\\epsilon\\) can not be set too small, which brings challenges for detecting high-frequency details. To compensate for this, we define another term about the contrast of the latent image \\(I\\). Motivated by the fact that sharper images tend to have large contrast, we evaluate the contrast as the average local variation with local window \\(\\mathcal{N}(\\pmb{p})\\) centered at pixel \\(\\pmb{p}\\) [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.638, + 0.892, + 0.676 + ], + "angle": 0, + "content": "\\[\nh (I) = \\sum_ {\\boldsymbol {p}} \\frac {1}{\\left| \\mathcal {N} _ {\\boldsymbol {p}} \\right|} \\sum_ {\\boldsymbol {q} \\in \\mathcal {N} _ {\\boldsymbol {p}}} \\left(I _ {\\boldsymbol {q}} - \\mu_ {I} \\left(\\mathcal {N} _ {\\boldsymbol {p}}\\right)\\right) ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.681, + 0.892, + 0.787 + ], + "angle": 0, + "content": "where \\(\\mu_I(\\mathcal{N}_p)\\) is the mean of the pixel value inside the window \\(\\mathcal{N}_p\\). The reason we use a local variation rather than a global one is that the blurry effects caused by motion are not always globally consistent. Maximizing the local variation can emphasize the local sharpness and reduce the blur for each local window. In practice, we adopt local standard deviation \\(\\sqrt{h(I)}\\) in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.892, + 0.878 + ], + "angle": 0, + "content": "For consecutive frames in a video, we can assume that their overall brightness level are similar, and thus the latency values are similar. Therefore, we define an indicator of temporal consistency by assuming Gaussian-distributed latency across consecutive frames. Given the latency of the last adjacent frame is \\( l' \\), the indicator defined on the latency" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.875, + 0.901 + ], + "angle": 0, + "content": "1 Please see the supplementary material for more details about \\(s(I)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24980" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.336, + 0.105 + ], + "angle": 0, + "content": "\\(l\\) of the current frame is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.108, + 0.469, + 0.143 + ], + "angle": 0, + "content": "\\[\ng (l) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {\\left(l - l ^ {\\prime}\\right) ^ {2}}{2}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.146, + 0.452, + 0.161 + ], + "angle": 0, + "content": "Overall, \\(\\mathcal{L}_{\\mathrm{ETF}}\\) to determine optimal latency is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.166, + 0.469, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E T F}} (I, l) = \\frac {s (\\epsilon I)}{\\sqrt {h (I)} g (l)}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.205, + 0.469, + 0.25 + ], + "angle": 0, + "content": "The efficacy of the proposed \\(\\mathcal{L}_{\\mathrm{ETF}}\\) is validated in Figure 2 (b), which shows that its value decreases as the latency estimation error decreases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.261, + 0.377, + 0.277 + ], + "angle": 0, + "content": "3.3. Learning-based latency correction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.285, + 0.469, + 0.346 + ], + "angle": 0, + "content": "Based on the objective defined by \\(\\mathcal{L}_{\\mathrm{ETF}}\\), we introduce a self-supervised approach founded on a parametrized form of latency, and a reformulated event-based double integral model that is differentiable with respect to latency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.35, + 0.47, + 0.577 + ], + "angle": 0, + "content": "Latency-illuminance curve. The physical factors that have influences on the latency of events cannot be easily measured for the arbitrary situation. Fortunately, for deblurring and frame interpolation algorithms using event-image hybrid sensors such as [18, 19], the intensity of the image can indirectly indicate the illuminance levels in certain regions. Inspired by the previous findings about the relationship between latency and illuminance levels [7, 16, 21], we establish a latency-intensity relationship using a parameterized curve, which is validated in controlled environments (see illustrations in Figure 3 and discussions in Section 4.1 for more details, which shows that the latency increases as the illuminance decreases). To elucidate this relationship, we model the latency-illuminance curve as a \\(K\\)-th order polynomial \\(\\psi\\) of the illuminance \\(L\\) at pixel \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.582, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\psi (L) = \\sum_ {i = 0} ^ {K} a _ {i} \\cdot f (L) ^ {i}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.628, + 0.469, + 0.753 + ], + "angle": 0, + "content": "where \\(\\{a_i\\}_{i = 0}^K\\) are the parameters, \\(f(\\cdot)\\) is a fixed function to map the illuminance \\(L\\) into monotone decrement for consideration of easier optimization. The curve in Eq. (15) can effectively be a guidance for regularizing the estimation of spatially varying latency. In experiments, the illuminance \\(L\\) is unavailable. We adopt the blurry image \\(B\\) to approximate illuminance \\(L\\) to estimate the spatially-varying latency, resulting in \\(l_{p} = \\psi (\\frac{B_{p}}{2\\Delta\\tau})\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.755, + 0.469, + 0.876 + ], + "angle": 0, + "content": "Latency-differentiable event integral. It is noted that the latency \\( l_{p} \\) derived from Eq. (9) is an index-like parameter, which has no derivatives. To allow gradient-based optimization, we establish a differentiable relationship between latency in events and specific indicators as shown in Figure 2 (c). By reformulating the event integral to be latency-differentiable, we derive a gradient chain linking \\( \\{a_{i}\\}_{i = 0}^{K} \\) to \\( \\mathcal{L}_{\\mathrm{ETF}} \\) via the EDI model with latency." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.135 + ], + "angle": 0, + "content": "In contrast to Eq. (3) where events \\(e_k = (\\pmb{p}_k, t_k, \\sigma_k)\\) are represented by an impulse function, we employ a piece-wise constant function:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.138, + 0.892, + 0.167 + ], + "angle": 0, + "content": "\\[\nE _ {\\boldsymbol {p}} ^ {(k)} (s) = \\frac {\\sigma_ {k}}{t _ {k} - t _ {k - 1}}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right], \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.171, + 0.892, + 0.214 + ], + "angle": 0, + "content": "where the changes \\(\\sigma_{k}\\) is equally split into the several duration. Then integration of events in duration \\((t_{k - 1},t_k]\\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.218, + 0.892, + 0.249 + ], + "angle": 0, + "content": "\\[\n\\int E _ {\\boldsymbol {p}} ^ {(k)} (s) d s = \\sigma_ {k}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right]. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.255, + 0.892, + 0.318 + ], + "angle": 0, + "content": "For simplicity, we denote \\( E_{\\pmb{p}}^{(k)}(s) \\) as \\( E_{\\pmb{p}}^{(k)} \\). Consider the \\( N \\) events \\( \\{e_k\\}_{k=1}^N \\) triggered at pixel \\( \\pmb{p} \\) during exposure time \\( [\\tau - l_p, t - l_p] \\) where \\( t_0 < \\tau - l_p < t_1 \\) and \\( t_{N-1} < t - l_p < t_n \\), we can rewrite the inner integral part in Eq. (7) into:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.323, + 0.891, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int_ {\\tau - l _ {p}} ^ {t - l _ {p}} E _ {\\pmb {p}} (s) d s \\\\ = \\int_ {\\tau - l _ {p}} ^ {t _ {1}} E _ {\\boldsymbol {p}} ^ {(1)} d s + \\sum_ {k = 2} ^ {N - 1} \\int_ {t _ {k - 1}} ^ {t _ {k}} E _ {\\boldsymbol {p}} ^ {(k)} d s + \\int_ {t _ {N - 1}} ^ {t - l _ {\\boldsymbol {p}}} E _ {\\boldsymbol {p}} ^ {(N)} d s \\\\ = \\left(t _ {1} - l _ {\\boldsymbol {p}} - \\tau\\right) \\sigma_ {1} + \\sum_ {k = 2} ^ {N - 1} \\sigma_ {k} + \\left(t - l _ {\\boldsymbol {p}} - t _ {N - 1}\\right) \\sigma_ {n}, \\tag {18} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.469, + 0.892, + 0.56 + ], + "angle": 0, + "content": "which is differentiable in the neighborhood of \\( l_{p} \\) with fixed \\( N \\). It is noted that \\( N \\) will change when \\( \\tau - l_{p} < t_{0} \\) or \\( t - l_{p} \\geq t_{N} \\), in which discontinuous changes on event triggering time will occur. While this might lead to a local minimum solution in optimization for a specific sample, it can be alleviated when optimizing over different samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.566, + 0.892, + 0.733 + ], + "angle": 0, + "content": "Spatially-varying latency correction. With the differentiable form of event integral for parameters \\(\\{a_i\\}_{i=0}^n\\), we can now perform gradient descent in the dataset by defining a loss function \\(\\mathcal{L}_{\\mathrm{total}}\\). Firstly, \\(\\mathcal{L}_{\\mathrm{ETF}}\\) defined in Eq. (14) is used as a term of the loss function. Heuristically, most of the latency \\(l_p\\) should lie in the range from 0 to the exposure time of the last blurry image, which can be optimized by the reformulated EDI. To ensure the latency does not exceed the previous exposure time so that the derivative can be calculated, a clipping operation is applied on \\(l_p\\). A boundary loss is also used to limit the range of latency:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.737, + 0.892, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {b o u n d}} = \\sum_ {l _ {p} > 0} l _ {p} + \\sum_ {l _ {p} < \\tau - \\tau^ {\\prime} - \\Delta \\tau} - l _ {p}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To deal with \\(l_{p}\\) exceeds the range of \\([0, \\tau - \\tau' - \\Delta \\tau]\\) and disables derivative computation due to gradients with too large magnitudes, we additionally optimize an intermediate constant latency \\(l_{\\mathrm{const}}\\) shared across the sensor array during the exposure time of the blurry image. It has a relatively small solution space and can be solved efficiently by searching. \\(l_{\\mathrm{const}}\\) provides an effective estimation of the overall latency (whose effectiveness is validated in the ex" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.451, + 0.901 + ], + "angle": 0, + "content": "2Please see the supplementary material for more details about \\(f(\\cdot)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "periment results on synthetic data in Section 4.2). Thus, it is being adopted in the term \\(\\mathcal{L}_{\\mathrm{mse}}\\) to minimize the distance between the current estimated latency and \\(l_{\\mathrm{const}}\\). \\(l^{\\prime}\\) in Eq. (13) is replaced by \\(l_{\\mathrm{const}}\\) to regularize using the frame's constant latency, rather than adjacent frames'." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.26 + ], + "angle": 0, + "content": "From the latency-illuminance curve estimated previously, we can observe that the curve should be monotonic. Thus we enforce that the function \\( l_{p} \\) is monotonically decreasing as the intensity increases. To apply the monotonicity constraint, we introduce the following constraint on the derivative of a function \\( l_{p} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.265, + 0.468, + 0.305 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d e v}} = \\sum \\max \\left(f ^ {\\prime} \\left(\\frac {B}{2 \\Delta \\tau}\\right) \\sum_ {i = 0} ^ {K} i \\cdot a _ {i} \\cdot f \\left(\\frac {B}{2 \\Delta \\tau}\\right) ^ {i - 1}, 0\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.31, + 0.468, + 0.34 + ], + "angle": 0, + "content": "where \\( B \\) contains different values on each pixel to ensure monotonically decreasing on all the data samples." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.341, + 0.465, + 0.355 + ], + "angle": 0, + "content": "Overall, the loss function to be minimized is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.362, + 0.468, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {\\text {E T F}} + \\beta \\mathcal {L} _ {\\text {b o u n d}} + \\eta \\mathcal {L} _ {\\text {m s e}} + \\zeta \\mathcal {L} _ {\\text {d e v}}, \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.384, + 0.468, + 0.429 + ], + "angle": 0, + "content": "where \\(\\alpha, \\beta, \\eta, \\zeta\\) are hyper-parameters for balancing different terms. Empirically, we set \\(\\alpha = 1, \\beta = 1, \\eta = 20, \\zeta = 0.1\\) in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.441, + 0.235, + 0.457 + ], + "angle": 0, + "content": "3.4. Training details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.468, + 0.545 + ], + "angle": 0, + "content": "Dataset preparation. For latency correction training, we utilized a DAVIS346 event camera [1] to obtain blurry images and corresponding events. Our real-scene dataset comprises 9 videos, where 100 frames and their neighboring events from each video are selected for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.549, + 0.468, + 0.699 + ], + "angle": 0, + "content": "Other implementation details. To simplify the implementation, we stack events into 500 bins to maintain high temporal resolution. Our model is implemented on the Pytorch framework and runs on an NVIDIA GeForce RTX 3090 GPU. We use ADAM [11] with default parameter setting for optimization. The batch size is set as 1. The learning rate was fixed to \\(10^{-5}\\) in the earlier 2 epochs and reduced to 0 in the last 3 epochs by the linear decay strategy. Constant initialization to zero for the non-latency assumption is used for initialization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.715, + 0.209, + 0.732 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.468, + 0.757 + ], + "angle": 0, + "content": "4.1. Latency estimation in controlled environments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Experiments in controlled settings measure actual latency and illuminance. Our setup, shown in Figure 3 (a), includes a light source, an Arduino board, a power source, a photometer, and an event camera (DVS346Mono) with its lens removed to prevent light attenuation. The Arduino board was programmed to switch off the light at a known time \\( t_{\\mathrm{ideal}} \\), and the subsequent OFF events were assumed to be triggered by this change. The latency of the \\( i \\)-th triggered event was calculated as the difference between \\( t_{\\mathrm{ideal}} \\)" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.707, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.194, + 0.68, + 0.205 + ], + "angle": 0, + "content": "(a) Latency estimation system" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.09, + 0.89, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.194, + 0.878, + 0.205 + ], + "angle": 0, + "content": "(b) Event trigger rate vs. time" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.206, + 0.706, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.324, + 0.672, + 0.334 + ], + "angle": 0, + "content": "(c) Latency vs. Illuminance" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.207, + 0.89, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.324, + 0.878, + 0.334 + ], + "angle": 0, + "content": "(d) Latency curve from Eq. (2)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Figure 3. Estimation of latency-illuminance curve. (a) A system designed to investigate the relationship between illuminance levels and latency, approximated using an \\(K\\)-th order polynomial. (b) The comparison of event triggering rate under three varied lighting conditions. It is approximated by the first \\(1\\%\\) OFF events in our experiments. (c) Estimated latency under controlled lighting, highlighting increased latency with diminishing light. (d) Results of the polynomial approximation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.726 + ], + "angle": 0, + "content": "and the time \\( t_{\\mathrm{actual}}^{(i)} \\). Illuminance is modulated by adjusting light voltage and measured alongside events by a photometer (with the lowest detectable value being 0.1 Lux). This process was repeated to mitigate noise effects (about 80 times in our experiments). The latency level of a given illuminance was calculated by the event triggering rate as shown in Fig. 3 (b), which is approximated by the median latency when the first \\( 1\\% \\) OFF events among all the samples are triggered for each process. By varying the illuminance, we empirically find that the latency increases as the illuminance decreases, as shown in Fig. 3 (c). We use the proposed parameterized curve to approximate the relationship between latency and illuminance, whose results are shown in Fig. 3 (d). It demonstrates an accurate curve fitting with \\( K = 5 \\). Due to the difficulty in illuminance measuring, the estimated parameter cannot be directly applied in deblurring and interpolation for arbitrary scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.735, + 0.755, + 0.75 + ], + "angle": 0, + "content": "4.2. Evaluation on synthetic data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.882 + ], + "angle": 0, + "content": "Compared methods. As the proposed method serves as a preprocessing step for latency correction of events, we evaluate the proposed methods based on two downstream tasks, that is, deblurring, and interpolation. Specifically, we test our method for deblurring algorithms including EDI [19], EFNet [24], and GEM [31], and for interpolation algorithms including EDI [19] and REFID [25]. We directly use the released pretrain model for comparison." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Dataset. To qualitatively evaluate downstream tasks, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.076, + 0.891, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.893, + 0.418 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on synthetic data. We correct the latency of events by the constant latency correction (\"W/ corr. const.\") method and compared it with the ground truth events (GT input). All the methods are tested with the same input." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.478 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation on synthetic data. \\(\\uparrow (\\downarrow)\\) means higher (lower) is better. The \"Diff\" shows the percentage improvement from \"W/o corr.\" to \"W/ corr. const.\" compared with \"GT input\"." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.484, + 0.47, + 0.81 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓
EDI [19]GT input28.4890.8390.489
W/o corr.26.0010.8100.500
DeblurringW/ corr. const.27.4710.8280.488
Diff. (%)+5.15+2.15-2.45
EFNet [24]GT input33.7610.9720.105
W/o corr.33.3300.9710.107
W/ corr. const.33.7780.9710.106
Diff. (%)+0.130-0.95
GEM [31]GT input30.6110.9630.183
W/o corr.27.7140.9340.198
W/ corr. const.30.3720.9580.185
Diff. (%)+8.68+2.49-7.10
EDI [19]GT input22.2930.8400.330
InterpolationW/o corr.22.1250.8340.331
W/ corr. const.22.2650.8380.331
Diff. (%)+0.63+0.470
REFID [25]GT input34.5220.9740.109
W/o corr.28.2440.9270.142
W/ corr. const.34.2170.9700.111
Diff. (%)+17.30+4.41-28.44
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "synthesize events with constant latency \\( l_{\\tau} \\) for each pair of events and its corresponding blurry image. In particular, we utilize the event provided in the HighREV dataset [25] as ideal events, and then we add a constant latency to the event-image pairs for different event sequences. We employ" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.546 + ], + "angle": 0, + "content": "the training data in the HighREV dataset which contains 1615 frames from 19 videos for testing to provide more convincing results. The spatially-varying latency correction is highly related to the properties of event cameras and depends on different scene illuminances, which can not be faithfully synthetic by event simulators [7, 10]. Therefore, we only provide results of the constant latency estimated by the proposed methods for quantitative comparisons." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.893, + 0.701 + ], + "angle": 0, + "content": "Metrics. To show the corrected performance, all the tested methods are evaluated with four kinds of input, events without latency (terms \"GT input\"), events with latency which is not corrected by our method (terms \"W/o corr.\"), the events corrected by constant latency \\( l_{\\mathrm{const}} \\) (terms \"W/ corr. const.). We adopt three commonly used metrics to measure the performance of deblurring and interpolation, peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and the perceptual error with learned perceptual image patch similarity (LPIPS) [29], respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Results. The quantitative evaluation results are shown in Table 1. We add the \"Diff.\" row to show the improvement of the proposed method. It shows that the proposed method can improve the performance of deblurring and interpolation by event latency correction. For deep learning methods like EFNet [24], they implicitly considered latency effects in training data generation, which makes them more robust to latency in events. However, their performance can also be improved with the latency correction. The qualitative evaluation results are shown in Figure 4, which show that the proposed method can explicitly reduce the artifacts brought by latency, and it performs more closely to the ground truth events (GT input). Specifically, the latency correction can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.078, + 0.891, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.893, + 0.4 + ], + "angle": 0, + "content": "Figure 5. The qualitative comparison on real data. We correct the latency of events by constant latency correction (\"W/ corr. Const.\") and spatially-varying latency correction (\"W/ corr\"). All the methods are tested with the same input." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.412, + 0.471, + 0.489 + ], + "angle": 0, + "content": "reduce the light strips brought by previous edges, in which the events are triggered with latency. The non-learning interpolation algorithm EDI [19] is more strict with the time accuracy and the noise in events, which makes it still blurry even with the GT input." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.495, + 0.295, + 0.509 + ], + "angle": 0, + "content": "4.3. Evaluation on real data" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.47, + 0.685 + ], + "angle": 0, + "content": "In order to reveal the performance of the proposed method, we evaluate our method on real data captured by DAVIS346Mono. Because the ground truth sharp images are unavailable, we can only qualitatively evaluate the proposed method on real data. The results on real data\\(^2\\) are shown in the left two columns of Figure 5. It shows that both the proposed latency correction methods can reduce the artifacts and blurry effects. Processed by the proposed latency correction methods, the number of inaccuracy events was explicitly reduced to help the deblurring and interpolation algorithms to produce natural and sharp edges." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.691, + 0.228, + 0.708 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.88 + ], + "angle": 0, + "content": "We compare the performance of spatially-varying latency correction with constant latency correction on real data as shown in the right two columns in Figure 5. It shows that events corrected by the spatially-varying method have more natural views than constant latency correction, which can also provide sharper results3. Benefiting from the data captured under different illumination conditions, the spatially varying latency correction can also provide more accurate latency estimation for specific scenes. Along with the different illumination conditions, it can also address the different latency properties for different pixels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.62, + 0.426 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.602 + ], + "angle": 0, + "content": "In this paper, we present the first method for latency correction in improving event-guided deblurring and interpolation tasks. Our approach introduces event-based temporal fidelity to evaluate the sharpness of latent images reconstructed from latency-corrected events and their corresponding blurry images. This objective guides the optimization of parameters in a parametrized latency model through gradient-based optimization, facilitated by a reformulated, latency-differentiable, event-based double integral model. The efficacy of our methods is demonstrated using both synthetic and real-world data sets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.893, + 0.78 + ], + "angle": 0, + "content": "Limitations. Although with potential for broad applicability across different event cameras, the present implementation of the proposed method is specifically tailored for DAVIS346Mono cameras due to that it is the only event camera we have with on-chip synchronization and our current method requires perfect temporal synchronization between the image and events. Moreover, due to challenges in measuring illuminance and curve variability, the practicality of estimated parameters of the latency-illuminance curve varies across sensors. The curves for more sensors, while significant, has not been explored in this paper." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.661, + 0.801 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.892, + 0.89 + ], + "angle": 0, + "content": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803), Beijing Natural Science Foundation (Grant No. L233024), National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), and SenseTime Collaborative Research Grant." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.434, + 0.901 + ], + "angle": 0, + "content": "3More comparisons are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24984" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 2014. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.225 + ], + "angle": 0, + "content": "[2] Tobi Delbruck and Carver A Mead. Adaptive photoreceptor with wide dynamic range. In Proc. of IEEE International Symposium on Circuits and Systems-ISCAS, pages 339–342, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.227, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus Is All You Need: Loss Functions for Event-Based Vision. In Proc. of Computer Vision and Pattern Recognition, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26:1272-1282, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[5] Chen Haoyu, Teng Minggui, Shi Boxin, Wang Yizhou, and Huang Tiejun. Learning to Deblur and Generate High frame rate video with an event camera. arXiv preprint arXiv:2003.00847, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[6] Ondrej Holevský, Radoslav Skoviera, Václav Hlaváč, and Roman Vitek. Experimental Comparison between Event and Global Shutter Cameras. Sensors, 21(4), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[7] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic DVS events. In Proc. of Computer Vision and Pattern Recognition Workshops, 2021. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[8] invitation. Understanding the Performance of Neuromorphic Event-based Vision Sensors. https://inivation.com/wp-content/uploads/2020/05/White-Paper-May-2020.pdf, 2020.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.59 + ], + "angle": 0, + "content": "[9] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning Event-Based Motion Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.592, + 0.471, + 0.647 + ], + "angle": 0, + "content": "[10] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, 15:702765, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.691, + 0.471, + 0.759 + ], + "angle": 0, + "content": "[12] Juan Antonio Lénero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 \\(\\mu\\)s Latency Asynchronous Frame-Free Event-Driven Dynamic-Vision-Sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.761, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[13] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[14] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In Proc. of European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[15] Brian McReynolds, Rui Graca, and Tobi Delbruck. Experimental methods to predict dynamic vision sensor event" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "camera performance. Optical Engineering, 61(7):074103-074103, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[16] Brian J Mcreynolds, Rui Graca, Daniel O'Keefe, Rachel Oliver, Richard Balthazor, Nathaniel George, and Matthew McHarg. Modeling and decoding event-based sensor lightning response. In Proc. of Unconventional Imaging, Sensing, and Adaptive Optics 2023, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[17] Patrenahalli M. Narendra and Robert C. Fitch. Real-time Adaptive contrast enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, (6):655-661, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.253, + 0.895, + 0.308 + ], + "angle": 0, + "content": "[18] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proc. of Computer Vision and Pattern Recognition, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.311, + 0.892, + 0.38 + ], + "angle": 0, + "content": "[19] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[20] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[21] Jonah P Sengupta. Demystifying Event-Based Camera Latency: Sensor Speed Dependence on Pixel Biasing, Light, and Spatial Activity. In Proc. of SPIE Conference 13045: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXXV, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[22] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. A \\(128 \\times 128\\) \\(1.5\\%\\) Contrast Sensitivity \\(0.9\\%\\) FPN \\(3\\mu s\\) Latency \\(4\\mathrm{mW}\\) Asynchronous Frame-Free Dynamic Vision Sensor Using Transimpedance Preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.599, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[23] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proc. of International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[24] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based Fusion for Motion Deblurring with Cross-modal Attention. In Proc. of European Conference on Computer Vision, 2022. 1, 2, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[25] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based Frame Interpolation with Ad-Hoc Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.801, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[26] Minggui Teng, Chu Zhou, Hanyue Lou, and Boxin Shi. NEST: Neural event stack for event-based image enhancement. In Proc. of European Conference on Computer Vision, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[27] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In Proc. of European Conference on Computer Vision, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24985" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion Deblurring with Real Events. In Proc. of International Conference on Computer Vision, pages 2583-2592, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of Computer Vision and Pattern Recognition, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[30] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proc. of Computer Vision and Pattern Recognition, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.249, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[31] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In Proc. of International Conference on Computer Vision, 2023. 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.359 + ], + "angle": 0, + "content": "[32] Chu Zhou, Minggui Teng, Jin Han, Jinxiu Liang, Chao Xu, Gang Cao, and Boxin Shi. Deblurring Low-Light Images with Events. International Journal of Computer Vision, 131 (5):1284-1298, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24986" + } + ] +] \ No newline at end of file diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_origin.pdf b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6de7cde3cb3dd5cd0ae3cab89440bd1dca7a5ac5 --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/29c2331e-d68b-477e-86d2-3c185688337a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2fe345fe3209b5099500984710ad95034ef71b1ab89ae5c99bf5a2f237a1422 +size 5277771 diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/full.md b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e42eb4cfd6eb3f104b07c3783b07b8fe6ff9e19d --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/full.md @@ -0,0 +1,345 @@ +# Latency Correction for Event-guided Deblurring and Frame Interpolation + +Yixin Yang $^{1,2,\dagger}$ Jinxiu Liang $^{1,2}$ Bohan Yu $^{1,2}$ Yan Chen $^{3}$ Jimmy S. Ren $^{3}$ Boxin Shi $^{1,2,*}$ + +$^{1}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University + +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{3}$ SenseTime Research $\{\text { yangyixin93,cssherryliang,ybh1998,shiboxin }\} @$ pku.edu.cn yanchenace@gmail.com rensijie@sensetime.com + +# Abstract + +Event cameras, with their high temporal resolution, dynamic range, and low power consumption, are particularly good at time-sensitive applications like deblurring and frame interpolation. However, their performance is hindered by latency variability, especially under low-light conditions and with fast-moving objects. This paper addresses the challenge of latency in event cameras — the temporal discrepancy between the actual occurrence of changes in the corresponding timestamp assigned by the sensor. Focusing on event-guided deblurring and frame interpolation tasks, we propose a latency correction method based on a parameterized latency model. To enable data-driven learning, we develop an event-based temporal fidelity to describe the sharpness of latent images reconstructed from events and the corresponding blurry images, and reformulate the event-based double integral model differentiable to latency. The proposed method is validated using synthetic and real-world datasets, demonstrating the benefits of latency correction for deblurring and interpolation across different lighting conditions. + +# 1. Introduction + +Emerging as a groundbreaking visual sensing paradigm, event cameras [1, 12, 13, 20, 22] provide a novel perspective to capturing and interpreting dynamic scenes. They have advantages of high temporal resolution (the discrete measurement resolution of the event detection time, in the order of microseconds [8]), high dynamic range (e.g., 120 dB for DAVIS346), and low power consumption. They asynchronously record changes in logarithmic pixel intensity only when they exceed a predefined threshold. Their unique characteristics have led to impressive performance gains in various time-sensitive applications, including deblurring [5, 19, 24, 26, 32] and frame interpolation [19, 25, 30]. + +![](images/766f2bb6c7df79a6ba5f677d693446d68924ec73330ae2a0a68c5000f8c3626d.jpg) +Figure 1. Illustration of latency in an event camera, and its influence on an event-guided deblurring algorithm. (a) The aerial view of ideal vs. real events triggered by a black point moving upwards on a white background, where the real events are captured in low-light conditions (below 0.1 Lux). (b) Deblurring results using the Event-based Double Integral (EDI) [19] method, guided by events captured under different illumination conditions. (c) Improved deblurring results from EDI [19] using event with latency corrected by the proposed method. + +![](images/053a46be857106f0a9c92d0f53f3edf55067bcd84a6ae06868c1adb8ee83bab7.jpg) + +![](images/95d2c6f0c172e3f4049a30243b704b656d2f75ce34a4e3032a01c9a4f5b3b5aa.jpg) + +The ideal event model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this, exhibiting a temporal discrepancy between the actual moment of intensity change and the corresponding event timestamp. As claimed by the inventors of the event-based sensors, in well-lit environments above 1k Lux, event cameras can achieve latency as low as $3\mu \mathrm{s}$ [22]. However, in dimmer conditions for a few Lux, the latency can exceed 1 ms [22]. The challenge in modeling this temporal discrepancy arises from the sensor pixels' non-linear and variable response to illumination, which is further complicated by manufacturing imperfection, inherent electronic noise, and the unpredictable nature of real-world scenes [13, 15, 16, 21]. Therefore, the profile of the temporal discrepancy is spatial-varying, inconsistent, and noisy. All these complexities make algorithms for handling them scarce. + +Without specifying the nature of the temporal discrepancy, we consider the collective temporal discrepancies in the timing of the camera's response to real-world changes, a phenomenon we term "latency", where the aforementioned sensor-level latency is one of the dominant components. + +Despite the complexity of the latency formation model, latency presents a strong relation with illuminance. As illustrated in Figure 1 (a), compared with the ideal events generated instantaneously when the intensity changes, the real events suffer from latency in illuminance below 0.1 Lux. The latency presents a significant challenge for time-sensitive applications [19, 24, 25, 30], which assume ideal event triggering and struggle to restore sharp edges due to the severe latency, as demonstrated in Figure 1 (b) vs. (c). + +In this paper, we focus on latency correction for event-guided imaging deblurring [19, 24, 28] and interpolation [19, 25, 31] in low-light and fast-moving scenes, where event cameras should play an important role. Based on the analysis and modeling of latency, we propose to predict latency in event cameras by conceptualizing latency as a variable function—potentially a constant for a specific scene or a polynomial function of the illuminance level, given a fixed event camera type—we endeavor to elucidate the relationship between latency and illuminance. + +Specifically, we propose a latency correction method for captured events and blurry images in an event-image hybrid sensor (e.g., DAVIS346). Inspired by the latency properties related to illuminance observed through controlled experiments, we characterize the latency-intensity relationship using a parameterized curve. To estimate the optimal latency, we introduce a novel objective to assess the sharpness profile of a latent image reconstructed from the latency-corrected events and the blurry image, which we term event-based temporal fidelity. Building upon them, we collect data under different scene illuminance and advocate for a self-supervised approach to latency correction. The EDI model [18, 19] is reformulated to be differentiable for latency to enable gradient-based optimization, therefore enhancing the performance of deblurring and interpolation. + +The efficacy of the proposed method is validated using both synthetic and real data, demonstrating improvements in image quality for deblurring and interpolation under challenging lighting conditions where the latency in events becomes severe. The key contributions of our work include: + +- the first data-driven latency correction method grounded in a differentiable reformulation of the event integral with respect to latency; +- the event-based temporal fidelity for measuring the sharpness profile of a latent image reconstructed by latency-corrected events and the corresponding blurry image; and +- a parameterized latency-intensity curve for modeling the relationship of latency in events and the intensity of the corresponding blurry image. + +The proposed method provides an approach to understanding and mitigating the effects of latency, paving the way for more accurate and reliable event-based imaging in dynamic environments with challenging lighting conditions. + +# 2. Related works + +We investigate the influence of latency on imaging algorithms in event-image hybrid systems, particularly focusing on deblurring and interpolation tasks due to their sensitivity to timing in events. + +Analysis of latency in event cameras. The triggering of events involves three subprocesses: photon reception, event trigger, and event readout. Notably, latency primarily arises during photon reception and event readout due to the electronic and physical design of these cameras [8]. Latency characteristics vary among different event camera models, ranging from minimal latency of $15\mu s$ [13] to $4\mu s$ [20] under 1k Lux conditions. Despite improvements in camera design, latency persists at nearly 1 ms under 1 Lux [1], posing challenges in low-light scenarios. Existing studies predominantly focus on sensor design aspects of event latency [1, 12, 13, 22] with a limited number examining latency under varying illuminance levels [6, 21]. In low-light conditions, latency induced by photoreceptor bandwidth becomes predominant [16], which is proportional to photocurrent (or illuminance) at low illuminance level, increasing monotonically with the illuminance level [2, 13, 15, 16, 21]. Consequently, the latency decreases as illuminance increases [7, 13, 21]. The photoreceptor circuit's low-pass characteristic [15, 16], has led to its modeling as a first- or second-order low-pass filter in various event simulators [7, 10]. While this provides a simplified relationship between latency and photocurrent, accurately determining photocurrent is often invalid for arbitrary imaging setups due to the complex imaging processes involving shuttering, apertures, etc. In order to save costs, most event cameras timestamp each event as they are being read out from the pixel array [8], which introduces readout latency that is significantly influenced by sensitivity and contrast [2, 8, 10, 15, 21]. The complexity of these factors contributing to latency in event cameras explains the scarcity of algorithms specifically designed to address this challenge for event-image hybrid systems. + +Event guided imaging algorithms. Event cameras record changes during the exposure time of blurry images, which makes the problem of deblurring and interpolation more well-posed. Pan et al. [18, 19] introduced the EDI model for high-frame-rate video reconstruction from blurry videos using event data, applicable in both deblurring and interpolation. For deblurring, Jiang et al. [9] combined convolutional recurrent networks with directional event filtering, Shang et al. [23] utilized the nearest sharp frames and events for deblurring, Xu et al. [28] improved model generalization through self-supervised learning, Teng et al. [26] focused on event representations for image enhancement, and Sun et al. [24] developed an event-image fusion module. For joint deblurring and interpolation, Lin et al. [14] employed neu + +ral networks and dynamic filtering, while Wang et al. [27] and Zhang et al. [30, 31] integrated multiple techniques, including event-enhanced sparse learning and self-supervised frameworks. Sun et al. [25] proposed a bidirectional recurrent network for fusion tasks. + +For these algorithms assuming ideal event triggering, the reconstruction of the latent image between or within exposure time heavily depends on the precision of timestamps in accompanying events. Latency presence can significantly impair algorithm performance. To the best of our knowledge, these has been no approach to address latency for event-guided deblurring and interpolation. + +# 3. Proposed Method + +In this paper, we aim to tackle the latency in events for event-guided deblurring and frame interpolation. Section 3.1 delves into the deviation of events with latency from the ideal event triggering mechanism and the event-based double integral model with latency. Based on this model, we propose a learning-based latency correction method for event-guided deblurring and frame interpolation. Section 3.2 introduces the objective for optimization, the event-based temporal fidelity. To allow gradient-based optimization, Section 3.3 reformulates event-based double integral to be differentiable respective to latency, where the latency is further parameterized into a polynomial of intensity on a per-pixel basis. Training details are shown in Section 3.4. + +# 3.1. Analysis on the latency in events + +Latency in event triggering. In an event camera, the photons hitting the photodetector of a pixel are transduced into an electrical signal and then compared against a predefined threshold to determine whether the change in intensity is substantial enough to trigger an "event". If this threshold is surpassed, the event is recorded with a high-resolution timestamp. Mathematically, the pixels of an event camera independently respond to variations in the continuous brightness signal $I_{p}(t)$ at pixel position $\boldsymbol{p} = (p_x,p_y)^\top$ and time $t$ . An event $e = (\boldsymbol{p},t,\sigma)$ is initiated at pixel $\boldsymbol{p}$ and time $t$ with polarity $\sigma \in \{+1, - 1\}$ indicating ON or OFF events when the logarithmic change of $I_{p}$ exceeds a dispatched threshold $\theta$ since the last event triggered at the pixel $\boldsymbol{p}$ and time $t_{\mathrm{ref}}$ : + +$$ +\left\| \log I _ {\boldsymbol {p}} (t) - \log I _ {\boldsymbol {p}} \left(t _ {\text {r e f}}\right) \right\| \geq \theta . \tag {1} +$$ + +In tasks involving image reconstruction where per-pixel intensity is needed, the logarithmic intensity changes for continuous time neighboring recorded timestamps are often approximated as an equality involving the threshold value $\theta$ and polarity $\sigma$ with the quantization errors ignored. + +The ideal model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this ideal, ex + +hibiting a temporal latency $l$ between the actual change in intensity and the event's triggering. The presence of latency exacerbates the approximation error of the intensity changes (e.g., Eq. (2) in the following), particularly affecting time-sensitive tasks like deblurring and frame interpolation. In environments with controlled lighting and motion, empirical methods have been employed to study and characterize latency in event cameras, such as using a second-order low-pass filter [13, 20] with parameters that are influenced by illumination conditions [6, 7, 13]. In the unpredictable flux of real-world settings, latency becomes markedly more unpredictable and defies precise modeling. The complexity arises from the non-linear and non-uniform factors. For instance, pixel responses across the sensor array exhibit variability due to manufacturing differences, and the light received by each pixel fluctuates rapidly with movement in the scene or changes in illumination conditions. Moreover, the existence of noise exacerbates the difficulty. + +Event-based double integral model with latency. An event-image hybrid sensor (e.g., DAVIS346) is able to capture pixel-wise aligned events and images, synergistically harnessing the strengths of both modalities. This integration involves vibrant color representation characteristic of frame-based sensors, coupled with the high temporal resolution and dynamic range offered by event cameras. The mathematical foundation for the interplay between events and images in such a system is encapsulated by the EDI model, which is originally proposed for applications in deblurring and frame interpolation [18, 19]. + +The proportional change for images at time $\tau$ and time $t$ can be measured by the integral of the $N$ events $\mathbb{E} = \{e_k\}_{k=1}^N$ triggered from $\tau$ to $t$ as: + +$$ +I _ {\boldsymbol {p}} (t) = I _ {\boldsymbol {p}} (\tau) \exp \left(\theta \int_ {\tau} ^ {t} E _ {\boldsymbol {p}} (s) d s\right), \tag {2} +$$ + +where $E_{\pmb{p}}(s)$ is a function of continuous time: + +$$ +E _ {\boldsymbol {p}} (s) = \sigma_ {k} \delta_ {t _ {k}} (s), \quad s \in (t _ {k - 1}, t _ {k} ], \tag {3} +$$ + +whenever there is an event $e_k = (\pmb{p}_k, t_k, \sigma_k)$ triggered and $\delta_{t_k}(s)$ is an impulse function at time $t_k$ with unit integral. A blurry image $B_p(\tau)$ with exposure duration $t \in [\tau - \Delta \tau, \tau + \Delta \tau]$ can be represented as the average of the latent image $I_p(t)$ over the exposure duration [18], which can be formulated as: + +$$ +B _ {\boldsymbol {p}} (\tau) = \frac {1}{2 \Delta \tau} I _ {\boldsymbol {p}} (\tau) R _ {\boldsymbol {p}} (\tau), \tag {4} +$$ + +where $R_{\pmb{p}}(\tau)$ is defined as: + +$$ +R _ {\boldsymbol {p}} (\tau) = \int_ {\tau - \Delta \tau} ^ {\tau + \Delta \tau} \exp \left(\theta \int_ {\tau} ^ {t} E _ {\boldsymbol {p}} (s) d s\right) d t. \tag {5} +$$ + +It indicates that the latent image can be reconstructed by: + +$$ +I _ {\boldsymbol {p}} (\tau) = 2 \Delta \tau \frac {B _ {\boldsymbol {p}} (\tau)}{R _ {\boldsymbol {p}} (\tau)}. \tag {6} +$$ + +We introduce a spatially-varying latency $l_{\mathbf{p}}(\tau)$ during exposure time $[\tau - \Delta \tau, \tau + \Delta \tau]$ into the EDI map $R_{\mathbf{p}}(\tau)$ in Eq. (5): + +$$ +\widehat {R} _ {\boldsymbol {p}} (\tau) = \int_ {\tau - \Delta \tau} ^ {\tau + \Delta \tau} \exp \left(\theta \int_ {\tau - l _ {\boldsymbol {p}} (\tau)} ^ {t - l _ {\boldsymbol {p}} (\tau)} E _ {\boldsymbol {p}} (s) d s\right) d t. \tag {7} +$$ + +In practice, the event as well as the exposure time for each latent image are discrete. Suppose the blurry image $B$ compromises $N_{I}$ latent images, and $N_{j}$ events at pixel $\pmb{p}$ are triggered during exposure time $[\tau - l_{\pmb{p}}(\tau), t - l_{\pmb{p}}(\tau)]$ , the discrete formulation of Eq. (7) is: + +$$ +\widehat {R} _ {\boldsymbol {p}} (\tau) = \sum_ {j = 1} ^ {N _ {I}} \exp \left(\theta \sum_ {k = 1} ^ {N _ {j}} \sigma_ {k}\right). \tag {8} +$$ + +The latent image reconstructed from blurry image $B$ and event image $E$ can then be obtained by: + +$$ +\widehat {I} _ {\boldsymbol {p}} (\tau) = N _ {I} \frac {B _ {\boldsymbol {p}} (\tau)}{\widehat {R} _ {\boldsymbol {p}} (\tau)}. \tag {9} +$$ + +$\pmb{p}$ and $\tau$ will be omitted in the following for simplicity. + +# 3.2. Event-based temporal fidelity + +We consider events $\mathbb{E} = \{e_k\}_{k=1}^N$ captured in the exposure duration of the blurry image $B$ . To achieve the goal of latency correction, a no-reference-based objective should be defined since the ground truth is hard to collect. To this end, we propose event-based temporal fidelity (ETF) $\mathcal{L}_{\mathrm{ETF}}$ to quantify the sharpness profile of latent images $I$ reconstructed from events with latency $l_p$ corrected and their corresponding blurry image $B$ , as shown in Figure 2 (a). + +Intuitively, $\mathcal{L}_{\mathrm{ETF}}$ should have smaller values as the estimated latency approaches its optimal values. Sharp images have more high-frequency contents compared with blurry images, which can be assessed by measuring the magnitude of its derivatives [3]. As shown in Figure 1 (d), the latent image results of EDI obtained from events with large latency appear with less sharpness and thicker edges. This motivates us to optimize the latency $l_{p}$ by enforcing regularizations of the sharpness profile of the latent image $I$ reconstructed from the EDI model in Eq. (9). Specifically, we define the average edge thickness of image gradients $\nabla I_p = (\nabla_x I_p, \nabla_y I_p)$ as the number of pixels with values greater than a parameter $\lambda$ , + +$$ +s _ {\lambda} (I) = \sum_ {\boldsymbol {p}} H \left(\left\| \nabla I _ {\boldsymbol {p}} \right\| _ {2} - \lambda\right), \tag {10} +$$ + +where $H(\cdot)$ is the Heaviside function. + +Integrating the right-side of the equation over $\lambda$ with a decreasing function $\rho$ , we can obtain an $\lambda$ -irrelevant representation of $s(I)$ : + +$$ +s (I) = \int_ {0} ^ {\infty} \sum_ {\boldsymbol {p}} \rho (\lambda) H \left(\left\| \nabla I _ {\boldsymbol {p}} \right\| _ {2} - \lambda\right) d \lambda . \tag {11} +$$ + +![](images/e3cc03fe458577520879a89881bd4c164b3269ba8f2faf73105bd9413a3e9d9b.jpg) +(a) The proposed latency correction method + +![](images/d900b7fb95b2c3fd92a4739c71395346be410a7c31597e7c167d002d0fa54fa3.jpg) +(b) $\mathcal{L}_{\mathrm{ETF}}$ vs. constant $l$ +Figure 2. Illustration of the proposed latency correction method. (a) The pipeline of the proposed method, which takes synchronized blurry image $B$ and events as input, and output the latency $l$ by polynomial function $l(L)$ . (b) The value of the proposed event temporal fidelity (ETF) $\mathcal{L}_{\mathrm{ETF}}$ decreases as the estimated latency approaches ground-truth latency, which validates the efficacy of ETF as an objective. (c) Latency-differentiable reformulation of event integration, enabling gradient-based optimization. + +![](images/1292622a7a85402013c098c1fe8c26335eea58d5527c597d09e62b0f381c9b89.jpg) +(c) Differentiable event integral + +We adopt the function $\rho (\lambda) = e^{-\lambda}$ in our experiments1. To pay more attention to the smoothness area, we multiply the image gradients $\nabla I_{p}$ with an amplifier $\epsilon$ . + +It is noted that, for the robustness to scenes with possible thick edges on the image, $\epsilon$ can not be set too small, which brings challenges for detecting high-frequency details. To compensate for this, we define another term about the contrast of the latent image $I$ . Motivated by the fact that sharper images tend to have large contrast, we evaluate the contrast as the average local variation with local window $\mathcal{N}(\pmb{p})$ centered at pixel $\pmb{p}$ [17]: + +$$ +h (I) = \sum_ {\boldsymbol {p}} \frac {1}{\left| \mathcal {N} _ {\boldsymbol {p}} \right|} \sum_ {\boldsymbol {q} \in \mathcal {N} _ {\boldsymbol {p}}} \left(I _ {\boldsymbol {q}} - \mu_ {I} \left(\mathcal {N} _ {\boldsymbol {p}}\right)\right) ^ {2}, \tag {12} +$$ + +where $\mu_I(\mathcal{N}_p)$ is the mean of the pixel value inside the window $\mathcal{N}_p$ . The reason we use a local variation rather than a global one is that the blurry effects caused by motion are not always globally consistent. Maximizing the local variation can emphasize the local sharpness and reduce the blur for each local window. In practice, we adopt local standard deviation $\sqrt{h(I)}$ in our experiments. + +For consecutive frames in a video, we can assume that their overall brightness level are similar, and thus the latency values are similar. Therefore, we define an indicator of temporal consistency by assuming Gaussian-distributed latency across consecutive frames. Given the latency of the last adjacent frame is $l'$ , the indicator defined on the latency + +$l$ of the current frame is formulated as: + +$$ +g (l) = \frac {1}{2 \pi} \exp \left(- \frac {\left(l - l ^ {\prime}\right) ^ {2}}{2}\right). \tag {13} +$$ + +Overall, $\mathcal{L}_{\mathrm{ETF}}$ to determine optimal latency is defined as: + +$$ +\mathcal {L} _ {\mathrm {E T F}} (I, l) = \frac {s (\epsilon I)}{\sqrt {h (I)} g (l)}. \tag {14} +$$ + +The efficacy of the proposed $\mathcal{L}_{\mathrm{ETF}}$ is validated in Figure 2 (b), which shows that its value decreases as the latency estimation error decreases. + +# 3.3. Learning-based latency correction + +Based on the objective defined by $\mathcal{L}_{\mathrm{ETF}}$ , we introduce a self-supervised approach founded on a parametrized form of latency, and a reformulated event-based double integral model that is differentiable with respect to latency. + +Latency-illuminance curve. The physical factors that have influences on the latency of events cannot be easily measured for the arbitrary situation. Fortunately, for deblurring and frame interpolation algorithms using event-image hybrid sensors such as [18, 19], the intensity of the image can indirectly indicate the illuminance levels in certain regions. Inspired by the previous findings about the relationship between latency and illuminance levels [7, 16, 21], we establish a latency-intensity relationship using a parameterized curve, which is validated in controlled environments (see illustrations in Figure 3 and discussions in Section 4.1 for more details, which shows that the latency increases as the illuminance decreases). To elucidate this relationship, we model the latency-illuminance curve as a $K$ -th order polynomial $\psi$ of the illuminance $L$ at pixel $p$ : + +$$ +\psi (L) = \sum_ {i = 0} ^ {K} a _ {i} \cdot f (L) ^ {i}, \tag {15} +$$ + +where $\{a_i\}_{i = 0}^K$ are the parameters, $f(\cdot)$ is a fixed function to map the illuminance $L$ into monotone decrement for consideration of easier optimization. The curve in Eq. (15) can effectively be a guidance for regularizing the estimation of spatially varying latency. In experiments, the illuminance $L$ is unavailable. We adopt the blurry image $B$ to approximate illuminance $L$ to estimate the spatially-varying latency, resulting in $l_{p} = \psi (\frac{B_{p}}{2\Delta\tau})$ . + +Latency-differentiable event integral. It is noted that the latency $l_{p}$ derived from Eq. (9) is an index-like parameter, which has no derivatives. To allow gradient-based optimization, we establish a differentiable relationship between latency in events and specific indicators as shown in Figure 2 (c). By reformulating the event integral to be latency-differentiable, we derive a gradient chain linking $\{a_{i}\}_{i = 0}^{K}$ to $\mathcal{L}_{\mathrm{ETF}}$ via the EDI model with latency. + +In contrast to Eq. (3) where events $e_k = (\pmb{p}_k, t_k, \sigma_k)$ are represented by an impulse function, we employ a piece-wise constant function: + +$$ +E _ {\boldsymbol {p}} ^ {(k)} (s) = \frac {\sigma_ {k}}{t _ {k} - t _ {k - 1}}, \quad s \in \left(t _ {k - 1}, t _ {k} \right], \tag {16} +$$ + +where the changes $\sigma_{k}$ is equally split into the several duration. Then integration of events in duration $(t_{k - 1},t_k]$ becomes: + +$$ +\int E _ {\boldsymbol {p}} ^ {(k)} (s) d s = \sigma_ {k}, \quad s \in \left(t _ {k - 1}, t _ {k} \right]. \tag {17} +$$ + +For simplicity, we denote $E_{\pmb{p}}^{(k)}(s)$ as $E_{\pmb{p}}^{(k)}$ . Consider the $N$ events $\{e_k\}_{k=1}^N$ triggered at pixel $\pmb{p}$ during exposure time $[\tau - l_p, t - l_p]$ where $t_0 < \tau - l_p < t_1$ and $t_{N-1} < t - l_p < t_n$ , we can rewrite the inner integral part in Eq. (7) into: + +$$ +\begin{array}{l} \int_ {\tau - l _ {p}} ^ {t - l _ {p}} E _ {\pmb {p}} (s) d s \\ = \int_ {\tau - l _ {p}} ^ {t _ {1}} E _ {\boldsymbol {p}} ^ {(1)} d s + \sum_ {k = 2} ^ {N - 1} \int_ {t _ {k - 1}} ^ {t _ {k}} E _ {\boldsymbol {p}} ^ {(k)} d s + \int_ {t _ {N - 1}} ^ {t - l _ {\boldsymbol {p}}} E _ {\boldsymbol {p}} ^ {(N)} d s \\ = \left(t _ {1} - l _ {\boldsymbol {p}} - \tau\right) \sigma_ {1} + \sum_ {k = 2} ^ {N - 1} \sigma_ {k} + \left(t - l _ {\boldsymbol {p}} - t _ {N - 1}\right) \sigma_ {n}, \tag {18} \\ \end{array} +$$ + +which is differentiable in the neighborhood of $l_{p}$ with fixed $N$ . It is noted that $N$ will change when $\tau - l_{p} < t_{0}$ or $t - l_{p} \geq t_{N}$ , in which discontinuous changes on event triggering time will occur. While this might lead to a local minimum solution in optimization for a specific sample, it can be alleviated when optimizing over different samples. + +Spatially-varying latency correction. With the differentiable form of event integral for parameters $\{a_i\}_{i=0}^n$ , we can now perform gradient descent in the dataset by defining a loss function $\mathcal{L}_{\mathrm{total}}$ . Firstly, $\mathcal{L}_{\mathrm{ETF}}$ defined in Eq. (14) is used as a term of the loss function. Heuristically, most of the latency $l_p$ should lie in the range from 0 to the exposure time of the last blurry image, which can be optimized by the reformulated EDI. To ensure the latency does not exceed the previous exposure time so that the derivative can be calculated, a clipping operation is applied on $l_p$ . A boundary loss is also used to limit the range of latency: + +$$ +\mathcal {L} _ {\text {b o u n d}} = \sum_ {l _ {p} > 0} l _ {p} + \sum_ {l _ {p} < \tau - \tau^ {\prime} - \Delta \tau} - l _ {p}. \tag {19} +$$ + +To deal with $l_{p}$ exceeds the range of $[0, \tau - \tau' - \Delta \tau]$ and disables derivative computation due to gradients with too large magnitudes, we additionally optimize an intermediate constant latency $l_{\mathrm{const}}$ shared across the sensor array during the exposure time of the blurry image. It has a relatively small solution space and can be solved efficiently by searching. $l_{\mathrm{const}}$ provides an effective estimation of the overall latency (whose effectiveness is validated in the ex + +periment results on synthetic data in Section 4.2). Thus, it is being adopted in the term $\mathcal{L}_{\mathrm{mse}}$ to minimize the distance between the current estimated latency and $l_{\mathrm{const}}$ . $l^{\prime}$ in Eq. (13) is replaced by $l_{\mathrm{const}}$ to regularize using the frame's constant latency, rather than adjacent frames'. + +From the latency-illuminance curve estimated previously, we can observe that the curve should be monotonic. Thus we enforce that the function $l_{p}$ is monotonically decreasing as the intensity increases. To apply the monotonicity constraint, we introduce the following constraint on the derivative of a function $l_{p}$ : + +$$ +\mathcal {L} _ {\mathrm {d e v}} = \sum \max \left(f ^ {\prime} \left(\frac {B}{2 \Delta \tau}\right) \sum_ {i = 0} ^ {K} i \cdot a _ {i} \cdot f \left(\frac {B}{2 \Delta \tau}\right) ^ {i - 1}, 0\right), \tag {20} +$$ + +where $B$ contains different values on each pixel to ensure monotonically decreasing on all the data samples. + +Overall, the loss function to be minimized is defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \alpha \mathcal {L} _ {\text {E T F}} + \beta \mathcal {L} _ {\text {b o u n d}} + \eta \mathcal {L} _ {\text {m s e}} + \zeta \mathcal {L} _ {\text {d e v}}, \tag {21} +$$ + +where $\alpha, \beta, \eta, \zeta$ are hyper-parameters for balancing different terms. Empirically, we set $\alpha = 1, \beta = 1, \eta = 20, \zeta = 0.1$ in our experiments. + +# 3.4. Training details + +Dataset preparation. For latency correction training, we utilized a DAVIS346 event camera [1] to obtain blurry images and corresponding events. Our real-scene dataset comprises 9 videos, where 100 frames and their neighboring events from each video are selected for training. + +Other implementation details. To simplify the implementation, we stack events into 500 bins to maintain high temporal resolution. Our model is implemented on the Pytorch framework and runs on an NVIDIA GeForce RTX 3090 GPU. We use ADAM [11] with default parameter setting for optimization. The batch size is set as 1. The learning rate was fixed to $10^{-5}$ in the earlier 2 epochs and reduced to 0 in the last 3 epochs by the linear decay strategy. Constant initialization to zero for the non-latency assumption is used for initialization. + +# 4. Experiments + +# 4.1. Latency estimation in controlled environments + +Experiments in controlled settings measure actual latency and illuminance. Our setup, shown in Figure 3 (a), includes a light source, an Arduino board, a power source, a photometer, and an event camera (DVS346Mono) with its lens removed to prevent light attenuation. The Arduino board was programmed to switch off the light at a known time $t_{\mathrm{ideal}}$ , and the subsequent OFF events were assumed to be triggered by this change. The latency of the $i$ -th triggered event was calculated as the difference between $t_{\mathrm{ideal}}$ + +![](images/c1c520d6d45ffee476ee0e94ca90bd7056a98031ed6c0d3ea1fcabcdc20711f5.jpg) +(a) Latency estimation system + +![](images/76b93f404960be0820c8fe1142643c2266b878e7112be868a58bcea08795fc24.jpg) +(b) Event trigger rate vs. time + +![](images/01f09bed8f57f760866a5b8665d4fdd04d551275aead5e5cd2aefabcdcd5cd48.jpg) +(c) Latency vs. Illuminance +Figure 3. Estimation of latency-illuminance curve. (a) A system designed to investigate the relationship between illuminance levels and latency, approximated using an $K$ -th order polynomial. (b) The comparison of event triggering rate under three varied lighting conditions. It is approximated by the first $1\%$ OFF events in our experiments. (c) Estimated latency under controlled lighting, highlighting increased latency with diminishing light. (d) Results of the polynomial approximation. + +![](images/cb6499a94a2293af0b518513eb87c9b9a4cdd73bdfd20b0f7e87c23d8bd49ed5.jpg) +(d) Latency curve from Eq. (2) + +and the time $t_{\mathrm{actual}}^{(i)}$ . Illuminance is modulated by adjusting light voltage and measured alongside events by a photometer (with the lowest detectable value being 0.1 Lux). This process was repeated to mitigate noise effects (about 80 times in our experiments). The latency level of a given illuminance was calculated by the event triggering rate as shown in Fig. 3 (b), which is approximated by the median latency when the first $1\%$ OFF events among all the samples are triggered for each process. By varying the illuminance, we empirically find that the latency increases as the illuminance decreases, as shown in Fig. 3 (c). We use the proposed parameterized curve to approximate the relationship between latency and illuminance, whose results are shown in Fig. 3 (d). It demonstrates an accurate curve fitting with $K = 5$ . Due to the difficulty in illuminance measuring, the estimated parameter cannot be directly applied in deblurring and interpolation for arbitrary scenes. + +# 4.2. Evaluation on synthetic data + +Compared methods. As the proposed method serves as a preprocessing step for latency correction of events, we evaluate the proposed methods based on two downstream tasks, that is, deblurring, and interpolation. Specifically, we test our method for deblurring algorithms including EDI [19], EFNet [24], and GEM [31], and for interpolation algorithms including EDI [19] and REFID [25]. We directly use the released pretrain model for comparison. + +Dataset. To qualitatively evaluate downstream tasks, we + +![](images/9ffb1dcda35a65fdd0939c2e61c8a94c0395d411c7e41bb5df202ec89b259d61.jpg) +Figure 4. Qualitative comparisons on synthetic data. We correct the latency of events by the constant latency correction ("W/ corr. const.") method and compared it with the ground truth events (GT input). All the methods are tested with the same input. + +Table 1. Quantitative evaluation on synthetic data. $\uparrow (\downarrow)$ means higher (lower) is better. The "Diff" shows the percentage improvement from "W/o corr." to "W/ corr. const." compared with "GT input". + +
PSNR↑SSIM↑LPIPS↓
EDI [19]GT input28.4890.8390.489
W/o corr.26.0010.8100.500
DeblurringW/ corr. const.27.4710.8280.488
Diff. (%)+5.15+2.15-2.45
EFNet [24]GT input33.7610.9720.105
W/o corr.33.3300.9710.107
W/ corr. const.33.7780.9710.106
Diff. (%)+0.130-0.95
GEM [31]GT input30.6110.9630.183
W/o corr.27.7140.9340.198
W/ corr. const.30.3720.9580.185
Diff. (%)+8.68+2.49-7.10
EDI [19]GT input22.2930.8400.330
InterpolationW/o corr.22.1250.8340.331
W/ corr. const.22.2650.8380.331
Diff. (%)+0.63+0.470
REFID [25]GT input34.5220.9740.109
W/o corr.28.2440.9270.142
W/ corr. const.34.2170.9700.111
Diff. (%)+17.30+4.41-28.44
+ +synthesize events with constant latency $l_{\tau}$ for each pair of events and its corresponding blurry image. In particular, we utilize the event provided in the HighREV dataset [25] as ideal events, and then we add a constant latency to the event-image pairs for different event sequences. We employ + +the training data in the HighREV dataset which contains 1615 frames from 19 videos for testing to provide more convincing results. The spatially-varying latency correction is highly related to the properties of event cameras and depends on different scene illuminances, which can not be faithfully synthetic by event simulators [7, 10]. Therefore, we only provide results of the constant latency estimated by the proposed methods for quantitative comparisons. + +Metrics. To show the corrected performance, all the tested methods are evaluated with four kinds of input, events without latency (terms "GT input"), events with latency which is not corrected by our method (terms "W/o corr."), the events corrected by constant latency $l_{\mathrm{const}}$ (terms "W/ corr. const.). We adopt three commonly used metrics to measure the performance of deblurring and interpolation, peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and the perceptual error with learned perceptual image patch similarity (LPIPS) [29], respectively. + +Results. The quantitative evaluation results are shown in Table 1. We add the "Diff." row to show the improvement of the proposed method. It shows that the proposed method can improve the performance of deblurring and interpolation by event latency correction. For deep learning methods like EFNet [24], they implicitly considered latency effects in training data generation, which makes them more robust to latency in events. However, their performance can also be improved with the latency correction. The qualitative evaluation results are shown in Figure 4, which show that the proposed method can explicitly reduce the artifacts brought by latency, and it performs more closely to the ground truth events (GT input). Specifically, the latency correction can + +![](images/206ec58a7943040845c5eb6fe3849466592b9cb2143602b926df79a602de296e.jpg) +Figure 5. The qualitative comparison on real data. We correct the latency of events by constant latency correction ("W/ corr. Const.") and spatially-varying latency correction ("W/ corr"). All the methods are tested with the same input. + +reduce the light strips brought by previous edges, in which the events are triggered with latency. The non-learning interpolation algorithm EDI [19] is more strict with the time accuracy and the noise in events, which makes it still blurry even with the GT input. + +# 4.3. Evaluation on real data + +In order to reveal the performance of the proposed method, we evaluate our method on real data captured by DAVIS346Mono. Because the ground truth sharp images are unavailable, we can only qualitatively evaluate the proposed method on real data. The results on real data $^2$ are shown in the left two columns of Figure 5. It shows that both the proposed latency correction methods can reduce the artifacts and blurry effects. Processed by the proposed latency correction methods, the number of inaccuracy events was explicitly reduced to help the deblurring and interpolation algorithms to produce natural and sharp edges. + +# 4.4. Ablation study + +We compare the performance of spatially-varying latency correction with constant latency correction on real data as shown in the right two columns in Figure 5. It shows that events corrected by the spatially-varying method have more natural views than constant latency correction, which can also provide sharper results3. Benefiting from the data captured under different illumination conditions, the spatially varying latency correction can also provide more accurate latency estimation for specific scenes. Along with the different illumination conditions, it can also address the different latency properties for different pixels. + +# 5. Conclusion + +In this paper, we present the first method for latency correction in improving event-guided deblurring and interpolation tasks. Our approach introduces event-based temporal fidelity to evaluate the sharpness of latent images reconstructed from latency-corrected events and their corresponding blurry images. This objective guides the optimization of parameters in a parametrized latency model through gradient-based optimization, facilitated by a reformulated, latency-differentiable, event-based double integral model. The efficacy of our methods is demonstrated using both synthetic and real-world data sets. + +Limitations. Although with potential for broad applicability across different event cameras, the present implementation of the proposed method is specifically tailored for DAVIS346Mono cameras due to that it is the only event camera we have with on-chip synchronization and our current method requires perfect temporal synchronization between the image and events. Moreover, due to challenges in measuring illuminance and curve variability, the practicality of estimated parameters of the latency-illuminance curve varies across sensors. The curves for more sensors, while significant, has not been explored in this paper. + +# Acknowledgement + +This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803), Beijing Natural Science Foundation (Grant No. L233024), National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), and SenseTime Collaborative Research Grant. + +# References + +[1] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 2014. 1, 2, 6 +[2] Tobi Delbruck and Carver A Mead. Adaptive photoreceptor with wide dynamic range. In Proc. of IEEE International Symposium on Circuits and Systems-ISCAS, pages 339–342, 1994. 2 +[3] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus Is All You Need: Loss Functions for Event-Based Vision. In Proc. of Computer Vision and Pattern Recognition, 2019. 4 +[4] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26:1272-1282, 2004. 1 +[5] Chen Haoyu, Teng Minggui, Shi Boxin, Wang Yizhou, and Huang Tiejun. Learning to Deblur and Generate High frame rate video with an event camera. arXiv preprint arXiv:2003.00847, 2020. 1 +[6] Ondrej Holevský, Radoslav Skoviera, Václav Hlaváč, and Roman Vitek. Experimental Comparison between Event and Global Shutter Cameras. Sensors, 21(4), 2021. 2, 3 +[7] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic DVS events. In Proc. of Computer Vision and Pattern Recognition Workshops, 2021. 2, 3, 5, 7 +[8] invitation. Understanding the Performance of Neuromorphic Event-based Vision Sensors. https://inivation.com/wp-content/uploads/2020/05/White-Paper-May-2020.pdf, 2020.1, 2 +[9] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning Event-Based Motion Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2020. 2 +[10] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, 15:702765, 2021. 2, 7 +[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[12] Juan Antonio Lénero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 $\mu$ s Latency Asynchronous Frame-Free Event-Driven Dynamic-Vision-Sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 1, 2 +[13] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1, 2, 3 +[14] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In Proc. of European Conference on Computer Vision, 2020. 2 +[15] Brian McReynolds, Rui Graca, and Tobi Delbruck. Experimental methods to predict dynamic vision sensor event + +camera performance. Optical Engineering, 61(7):074103-074103, 2022. 1, 2 +[16] Brian J Mcreynolds, Rui Graca, Daniel O'Keefe, Rachel Oliver, Richard Balthazor, Nathaniel George, and Matthew McHarg. Modeling and decoding event-based sensor lightning response. In Proc. of Unconventional Imaging, Sensing, and Adaptive Optics 2023, 2023. 1, 2, 5 +[17] Patrenahalli M. Narendra and Robert C. Fitch. Real-time Adaptive contrast enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, (6):655-661, 1981. 4 +[18] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proc. of Computer Vision and Pattern Recognition, 2019. 2, 3, 5 +[19] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2, 3, 5, 6, 7, 8 +[20] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2, 3 +[21] Jonah P Sengupta. Demystifying Event-Based Camera Latency: Sensor Speed Dependence on Pixel Biasing, Light, and Spatial Activity. In Proc. of SPIE Conference 13045: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXXV, 2023. 1, 2, 5 +[22] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. A $128 \times 128$ $1.5\%$ Contrast Sensitivity $0.9\%$ FPN $3\mu s$ Latency $4\mathrm{mW}$ Asynchronous Frame-Free Dynamic Vision Sensor Using Transimpedance Preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 2 +[23] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proc. of International Conference on Computer Vision, 2021. 2 +[24] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based Fusion for Motion Deblurring with Cross-modal Attention. In Proc. of European Conference on Computer Vision, 2022. 1, 2, 6, 7, 5 +[25] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based Frame Interpolation with Ad-Hoc Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 6, 7, 5 +[26] Minggui Teng, Chu Zhou, Hanyue Lou, and Boxin Shi. NEST: Neural event stack for event-based image enhancement. In Proc. of European Conference on Computer Vision, 2022. 1, 2 +[27] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In Proc. of European Conference on Computer Vision, 2020. 3 + +[28] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion Deblurring with Real Events. In Proc. of International Conference on Computer Vision, pages 2583-2592, 2021. 2 +[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of Computer Vision and Pattern Recognition, 2018. 7 +[30] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proc. of Computer Vision and Pattern Recognition, 2022. 1, 2, 3 +[31] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In Proc. of International Conference on Computer Vision, 2023. 2, 3, 6, 7, 5 +[32] Chu Zhou, Minggui Teng, Jin Han, Jinxiu Liang, Chao Xu, Gang Cao, and Boxin Shi. Deblurring Low-Light Images with Events. International Journal of Computer Vision, 131 (5):1284-1298, 2023. 1 \ No newline at end of file diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/images.zip b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba126d790bba8194b665ebf1d95418e0d286b775 --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d6e160a7ffbae20448c27bd85476e9fb873ffa783f26894a8809adbf0983fb8 +size 632249 diff --git a/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/layout.json b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3507db942fff4814bd1936dc4c7d40db35a42d26 --- /dev/null +++ b/2024/Latency Correction for Event-guided Deblurring and Frame Interpolation/layout.json @@ -0,0 +1,9793 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 526, + 121 + ], + "type": "text", + "content": "Latency Correction for Event-guided Deblurring and Frame Interpolation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": "Yixin Yang" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": " Jinxiu Liang" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": " Bohan Yu" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": " Yan Chen" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": " Jimmy S. Ren" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 77, + 142, + 517, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 159, + 507, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 159, + 507, + 171 + ], + "spans": [ + { + "bbox": [ + 86, + 159, + 507, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 159, + 507, + 171 + ], + "type": "text", + "content": " National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "spans": [ + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University " + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "text", + "content": " SenseTime Research " + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "inline_equation", + "content": "\\{\\text { yangyixin93,cssherryliang,ybh1998,shiboxin }\\} @" + }, + { + "bbox": [ + 52, + 172, + 541, + 199 + ], + "type": "text", + "content": " pku.edu.cn yanchenace@gmail.com rensijie@sensetime.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 492 + ], + "type": "text", + "content": "Event cameras, with their high temporal resolution, dynamic range, and low power consumption, are particularly good at time-sensitive applications like deblurring and frame interpolation. However, their performance is hindered by latency variability, especially under low-light conditions and with fast-moving objects. This paper addresses the challenge of latency in event cameras — the temporal discrepancy between the actual occurrence of changes in the corresponding timestamp assigned by the sensor. Focusing on event-guided deblurring and frame interpolation tasks, we propose a latency correction method based on a parameterized latency model. To enable data-driven learning, we develop an event-based temporal fidelity to describe the sharpness of latent images reconstructed from events and the corresponding blurry images, and reformulate the event-based double integral model differentiable to latency. The proposed method is validated using synthetic and real-world datasets, demonstrating the benefits of latency correction for deblurring and interpolation across different lighting conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 128, + 530 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 538, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 682 + ], + "type": "text", + "content": "Emerging as a groundbreaking visual sensing paradigm, event cameras [1, 12, 13, 20, 22] provide a novel perspective to capturing and interpreting dynamic scenes. They have advantages of high temporal resolution (the discrete measurement resolution of the event detection time, in the order of microseconds [8]), high dynamic range (e.g., 120 dB for DAVIS346), and low power consumption. They asynchronously record changes in logarithmic pixel intensity only when they exceed a predefined threshold. Their unique characteristics have led to impressive performance gains in various time-sensitive applications, including deblurring [5, 19, 24, 26, 32] and frame interpolation [19, 25, 30]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 226, + 436, + 336 + ], + "blocks": [ + { + "bbox": [ + 307, + 226, + 436, + 336 + ], + "lines": [ + { + "bbox": [ + 307, + 226, + 436, + 336 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 436, + 336 + ], + "type": "image", + "image_path": "766f2bb6c7df79a6ba5f677d693446d68924ec73330ae2a0a68c5000f8c3626d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 342, + 546, + 441 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 546, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 546, + 441 + ], + "type": "text", + "content": "Figure 1. Illustration of latency in an event camera, and its influence on an event-guided deblurring algorithm. (a) The aerial view of ideal vs. real events triggered by a black point moving upwards on a white background, where the real events are captured in low-light conditions (below 0.1 Lux). (b) Deblurring results using the Event-based Double Integral (EDI) [19] method, guided by events captured under different illumination conditions. (c) Improved deblurring results from EDI [19] using event with latency corrected by the proposed method." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 440, + 228, + 544, + 284 + ], + "blocks": [ + { + "bbox": [ + 440, + 228, + 544, + 284 + ], + "lines": [ + { + "bbox": [ + 440, + 228, + 544, + 284 + ], + "spans": [ + { + "bbox": [ + 440, + 228, + 544, + 284 + ], + "type": "image", + "image_path": "053a46be857106f0a9c92d0f53f3edf55067bcd84a6ae06868c1adb8ee83bab7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 441, + 287, + 544, + 336 + ], + "blocks": [ + { + "bbox": [ + 441, + 287, + 544, + 336 + ], + "lines": [ + { + "bbox": [ + 441, + 287, + 544, + 336 + ], + "spans": [ + { + "bbox": [ + 441, + 287, + 544, + 336 + ], + "type": "image", + "image_path": "95d2c6f0c172e3f4049a30243b704b656d2f75ce34a4e3032a01c9a4f5b3b5aa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "text", + "content": "The ideal event model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this, exhibiting a temporal discrepancy between the actual moment of intensity change and the corresponding event timestamp. As claimed by the inventors of the event-based sensors, in well-lit environments above 1k Lux, event cameras can achieve latency as low as " + }, + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "inline_equation", + "content": "3\\mu \\mathrm{s}" + }, + { + "bbox": [ + 304, + 449, + 546, + 652 + ], + "type": "text", + "content": " [22]. However, in dimmer conditions for a few Lux, the latency can exceed 1 ms [22]. The challenge in modeling this temporal discrepancy arises from the sensor pixels' non-linear and variable response to illumination, which is further complicated by manufacturing imperfection, inherent electronic noise, and the unpredictable nature of real-world scenes [13, 15, 16, 21]. Therefore, the profile of the temporal discrepancy is spatial-varying, inconsistent, and noisy. All these complexities make algorithms for handling them scarce." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Without specifying the nature of the temporal discrepancy, we consider the collective temporal discrepancies in the timing of the camera's response to real-world changes, a phenomenon we term \"latency\", where the aforementioned sensor-level latency is one of the dominant components." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 251, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 251, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 251, + 703 + ], + "type": "text", + "content": "This work is done during Yixin's internship at SenseTime." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 703, + 209, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 209, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 209, + 712 + ], + "type": "text", + "content": "*Corresponding author: shiboxin@pku.edu.cn" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24977" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 180 + ], + "type": "text", + "content": "Despite the complexity of the latency formation model, latency presents a strong relation with illuminance. As illustrated in Figure 1 (a), compared with the ideal events generated instantaneously when the intensity changes, the real events suffer from latency in illuminance below 0.1 Lux. The latency presents a significant challenge for time-sensitive applications [19, 24, 25, 30], which assume ideal event triggering and struggle to restore sharp edges due to the severe latency, as demonstrated in Figure 1 (b) vs. (c)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 288, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 288, + 304 + ], + "type": "text", + "content": "In this paper, we focus on latency correction for event-guided imaging deblurring [19, 24, 28] and interpolation [19, 25, 31] in low-light and fast-moving scenes, where event cameras should play an important role. Based on the analysis and modeling of latency, we propose to predict latency in event cameras by conceptualizing latency as a variable function—potentially a constant for a specific scene or a polynomial function of the illuminance level, given a fixed event camera type—we endeavor to elucidate the relationship between latency and illuminance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 307, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 487 + ], + "type": "text", + "content": "Specifically, we propose a latency correction method for captured events and blurry images in an event-image hybrid sensor (e.g., DAVIS346). Inspired by the latency properties related to illuminance observed through controlled experiments, we characterize the latency-intensity relationship using a parameterized curve. To estimate the optimal latency, we introduce a novel objective to assess the sharpness profile of a latent image reconstructed from the latency-corrected events and the blurry image, which we term event-based temporal fidelity. Building upon them, we collect data under different scene illuminance and advocate for a self-supervised approach to latency correction. The EDI model [18, 19] is reformulated to be differentiable for latency to enable gradient-based optimization, therefore enhancing the performance of deblurring and interpolation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 551 + ], + "type": "text", + "content": "The efficacy of the proposed method is validated using both synthetic and real data, demonstrating improvements in image quality for deblurring and interpolation under challenging lighting conditions where the latency in events becomes severe. The key contributions of our work include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 554, + 286, + 661 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 554, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 286, + 590 + ], + "type": "text", + "content": "- the first data-driven latency correction method grounded in a differentiable reformulation of the event integral with respect to latency;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 590, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 286, + 625 + ], + "type": "text", + "content": "- the event-based temporal fidelity for measuring the sharpness profile of a latent image reconstructed by latency-corrected events and the corresponding blurry image; and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 627, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 286, + 661 + ], + "type": "text", + "content": "- a parameterized latency-intensity curve for modeling the relationship of latency in events and the intensity of the corresponding blurry image." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "The proposed method provides an approach to understanding and mitigating the effects of latency, paving the way for more accurate and reliable event-based imaging in dynamic environments with challenging lighting conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 71, + 395, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 395, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 395, + 84 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 93, + 545, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 93, + 545, + 141 + ], + "spans": [ + { + "bbox": [ + 305, + 93, + 545, + 141 + ], + "type": "text", + "content": "We investigate the influence of latency on imaging algorithms in event-image hybrid systems, particularly focusing on deblurring and interpolation tasks due to their sensitivity to timing in events." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "text", + "content": "Analysis of latency in event cameras. The triggering of events involves three subprocesses: photon reception, event trigger, and event readout. Notably, latency primarily arises during photon reception and event readout due to the electronic and physical design of these cameras [8]. Latency characteristics vary among different event camera models, ranging from minimal latency of " + }, + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "text", + "content": " [13] to " + }, + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "inline_equation", + "content": "4\\mu s" + }, + { + "bbox": [ + 305, + 146, + 545, + 542 + ], + "type": "text", + "content": " [20] under 1k Lux conditions. Despite improvements in camera design, latency persists at nearly 1 ms under 1 Lux [1], posing challenges in low-light scenarios. Existing studies predominantly focus on sensor design aspects of event latency [1, 12, 13, 22] with a limited number examining latency under varying illuminance levels [6, 21]. In low-light conditions, latency induced by photoreceptor bandwidth becomes predominant [16], which is proportional to photocurrent (or illuminance) at low illuminance level, increasing monotonically with the illuminance level [2, 13, 15, 16, 21]. Consequently, the latency decreases as illuminance increases [7, 13, 21]. The photoreceptor circuit's low-pass characteristic [15, 16], has led to its modeling as a first- or second-order low-pass filter in various event simulators [7, 10]. While this provides a simplified relationship between latency and photocurrent, accurately determining photocurrent is often invalid for arbitrary imaging setups due to the complex imaging processes involving shuttering, apertures, etc. In order to save costs, most event cameras timestamp each event as they are being read out from the pixel array [8], which introduces readout latency that is significantly influenced by sensitivity and contrast [2, 8, 10, 15, 21]. The complexity of these factors contributing to latency in event cameras explains the scarcity of algorithms specifically designed to address this challenge for event-image hybrid systems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 546, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 546, + 714 + ], + "type": "text", + "content": "Event guided imaging algorithms. Event cameras record changes during the exposure time of blurry images, which makes the problem of deblurring and interpolation more well-posed. Pan et al. [18, 19] introduced the EDI model for high-frame-rate video reconstruction from blurry videos using event data, applicable in both deblurring and interpolation. For deblurring, Jiang et al. [9] combined convolutional recurrent networks with directional event filtering, Shang et al. [23] utilized the nearest sharp frames and events for deblurring, Xu et al. [28] improved model generalization through self-supervised learning, Teng et al. [26] focused on event representations for image enhancement, and Sun et al. [24] developed an event-image fusion module. For joint deblurring and interpolation, Lin et al. [14] employed neu" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24978" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "ral networks and dynamic filtering, while Wang et al. [27] and Zhang et al. [30, 31] integrated multiple techniques, including event-enhanced sparse learning and self-supervised frameworks. Sun et al. [25] proposed a bidirectional recurrent network for fusion tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 288, + 217 + ], + "type": "text", + "content": "For these algorithms assuming ideal event triggering, the reconstruction of the latent image between or within exposure time heavily depends on the precision of timestamps in accompanying events. Latency presence can significantly impair algorithm performance. To the best of our knowledge, these has been no approach to address latency for event-guided deblurring and interpolation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 227, + 153, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 153, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 153, + 240 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 288, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 401 + ], + "type": "text", + "content": "In this paper, we aim to tackle the latency in events for event-guided deblurring and frame interpolation. Section 3.1 delves into the deviation of events with latency from the ideal event triggering mechanism and the event-based double integral model with latency. Based on this model, we propose a learning-based latency correction method for event-guided deblurring and frame interpolation. Section 3.2 introduces the objective for optimization, the event-based temporal fidelity. To allow gradient-based optimization, Section 3.3 reformulates event-based double integral to be differentiable respective to latency, where the latency is further parameterized into a polynomial of intensity on a per-pixel basis. Training details are shown in Section 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 221, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 221, + 420 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 221, + 420 + ], + "type": "text", + "content": "3.1. Analysis on the latency in events" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": "Latency in event triggering. In an event camera, the photons hitting the photodetector of a pixel are transduced into an electrical signal and then compared against a predefined threshold to determine whether the change in intensity is substantial enough to trigger an \"event\". If this threshold is surpassed, the event is recorded with a high-resolution timestamp. Mathematically, the pixels of an event camera independently respond to variations in the continuous brightness signal " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "I_{p}(t)" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " at pixel position " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p} = (p_x,p_y)^\\top" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": ". An event " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "e = (\\boldsymbol{p},t,\\sigma)" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " is initiated at pixel " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " with polarity " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\{+1, - 1\\}" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " indicating ON or OFF events when the logarithmic change of " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "I_{p}" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " exceeds a dispatched threshold " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " since the last event triggered at the pixel " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 427, + 287, + 595 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 600, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 600, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\left\\| \\log I _ {\\boldsymbol {p}} (t) - \\log I _ {\\boldsymbol {p}} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\theta . \\tag {1}", + "image_path": "b2352eccea7e58b14d11ecaf60080dccade6235af34121ca0f46c3a3590c9ed0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": "In tasks involving image reconstruction where per-pixel intensity is needed, the logarithmic intensity changes for continuous time neighboring recorded timestamps are often approximated as an equality involving the threshold value " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " and polarity " + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 617, + 287, + 677 + ], + "type": "text", + "content": " with the quantization errors ignored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The ideal model posits that event triggering is instantaneous upon the detection of intensity changes. However, real-world event timestamps deviate from this ideal, ex" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": "hibiting a temporal latency " + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 72, + 545, + 288 + ], + "type": "text", + "content": " between the actual change in intensity and the event's triggering. The presence of latency exacerbates the approximation error of the intensity changes (e.g., Eq. (2) in the following), particularly affecting time-sensitive tasks like deblurring and frame interpolation. In environments with controlled lighting and motion, empirical methods have been employed to study and characterize latency in event cameras, such as using a second-order low-pass filter [13, 20] with parameters that are influenced by illumination conditions [6, 7, 13]. In the unpredictable flux of real-world settings, latency becomes markedly more unpredictable and defies precise modeling. The complexity arises from the non-linear and non-uniform factors. For instance, pixel responses across the sensor array exhibit variability due to manufacturing differences, and the light received by each pixel fluctuates rapidly with movement in the scene or changes in illumination conditions. Moreover, the existence of noise exacerbates the difficulty." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 545, + 422 + ], + "type": "text", + "content": "Event-based double integral model with latency. An event-image hybrid sensor (e.g., DAVIS346) is able to capture pixel-wise aligned events and images, synergistically harnessing the strengths of both modalities. This integration involves vibrant color representation characteristic of frame-based sensors, coupled with the high temporal resolution and dynamic range offered by event cameras. The mathematical foundation for the interplay between events and images in such a system is encapsulated by the EDI model, which is originally proposed for applications in deblurring and frame interpolation [18, 19]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": "The proportional change for images at time " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " can be measured by the integral of the " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " events " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\mathbb{E} = \\{e_k\\}_{k=1}^N" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " triggered from " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 423, + 545, + 459 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 463, + 545, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 463, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 349, + 463, + 545, + 491 + ], + "type": "interline_equation", + "content": "I _ {\\boldsymbol {p}} (t) = I _ {\\boldsymbol {p}} (\\tau) \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right), \\tag {2}", + "image_path": "79d778d49e35fac7ed8d875f38e7c8ad1411dd78f72cfed7b84f0805ad8dd48a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 494, + 494, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 494, + 494, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 494, + 494, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 494, + 494, + 506 + ], + "type": "inline_equation", + "content": "E_{\\pmb{p}}(s)" + }, + { + "bbox": [ + 306, + 494, + 494, + 506 + ], + "type": "text", + "content": " is a function of continuous time:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 510, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 510, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 351, + 510, + 545, + 525 + ], + "type": "interline_equation", + "content": "E _ {\\boldsymbol {p}} (s) = \\sigma_ {k} \\delta_ {t _ {k}} (s), \\quad s \\in (t _ {k - 1}, t _ {k} ], \\tag {3}", + "image_path": "425a46e7bc5726c82e5654772199fb640b3d258a8c60618f5c37f74ef6dee4d9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": "whenever there is an event " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "e_k = (\\pmb{p}_k, t_k, \\sigma_k)" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " triggered and " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\delta_{t_k}(s)" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " is an impulse function at time " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " with unit integral. A blurry image " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "B_p(\\tau)" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " with exposure duration " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "t \\in [\\tau - \\Delta \\tau, \\tau + \\Delta \\tau]" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " can be represented as the average of the latent image " + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "inline_equation", + "content": "I_p(t)" + }, + { + "bbox": [ + 305, + 528, + 545, + 598 + ], + "type": "text", + "content": " over the exposure duration [18], which can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 368, + 601, + 545, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 601, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 368, + 601, + 545, + 624 + ], + "type": "interline_equation", + "content": "B _ {\\boldsymbol {p}} (\\tau) = \\frac {1}{2 \\Delta \\tau} I _ {\\boldsymbol {p}} (\\tau) R _ {\\boldsymbol {p}} (\\tau), \\tag {4}", + "image_path": "062681f21e7c8969e1f96bb7249a9eba5750e7157defb7202acfa07168b736e3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 627, + 416, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 416, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 416, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 627, + 416, + 639 + ], + "type": "inline_equation", + "content": "R_{\\pmb{p}}(\\tau)" + }, + { + "bbox": [ + 306, + 627, + 416, + 639 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 337, + 643, + 545, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 643, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 337, + 643, + 545, + 671 + ], + "type": "interline_equation", + "content": "R _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau} ^ {t} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {5}", + "image_path": "7bfbbad7bab8c938d847a429ed5e784bfbe3297e0ae02676ee1f83986ac4f197.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 674, + 535, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 674, + 535, + 686 + ], + "spans": [ + { + "bbox": [ + 306, + 674, + 535, + 686 + ], + "type": "text", + "content": "It indicates that the latent image can be reconstructed by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 381, + 689, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 689, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 381, + 689, + 545, + 717 + ], + "type": "interline_equation", + "content": "I _ {\\boldsymbol {p}} (\\tau) = 2 \\Delta \\tau \\frac {B _ {\\boldsymbol {p}} (\\tau)}{R _ {\\boldsymbol {p}} (\\tau)}. \\tag {6}", + "image_path": "791a9f5c1b19262230be4ca06e3d3cbb7a3cc90cdde8b184951cfe1d23f5647f.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24979" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "We introduce a spatially-varying latency " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "l_{\\mathbf{p}}(\\tau)" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " during exposure time " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "[\\tau - \\Delta \\tau, \\tau + \\Delta \\tau]" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " into the EDI map " + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{p}}(\\tau)" + }, + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": " in Eq. (5):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 112, + 287, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 112, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 112, + 287, + 144 + ], + "type": "interline_equation", + "content": "\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\int_ {\\tau - \\Delta \\tau} ^ {\\tau + \\Delta \\tau} \\exp \\left(\\theta \\int_ {\\tau - l _ {\\boldsymbol {p}} (\\tau)} ^ {t - l _ {\\boldsymbol {p}} (\\tau)} E _ {\\boldsymbol {p}} (s) d s\\right) d t. \\tag {7}", + "image_path": "8b66c5d8fd1f728a9969d09cbd3fb125b638cd634c244b31d2d7d4c24a31ea18.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": "In practice, the event as well as the exposure time for each latent image are discrete. Suppose the blurry image " + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": " compromises " + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "N_{I}" + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": " latent images, and " + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "N_{j}" + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": " events at pixel " + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": " are triggered during exposure time " + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "[\\tau - l_{\\pmb{p}}(\\tau), t - l_{\\pmb{p}}(\\tau)]" + }, + { + "bbox": [ + 47, + 147, + 287, + 207 + ], + "type": "text", + "content": ", the discrete formulation of Eq. (7) is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 211, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 211, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 102, + 211, + 287, + 251 + ], + "type": "interline_equation", + "content": "\\widehat {R} _ {\\boldsymbol {p}} (\\tau) = \\sum_ {j = 1} ^ {N _ {I}} \\exp \\left(\\theta \\sum_ {k = 1} ^ {N _ {j}} \\sigma_ {k}\\right). \\tag {8}", + "image_path": "ba033db2d899f359c31be1f2142001d044ee8c87c27ba3deed503a8d7dc0c703.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": "The latent image reconstructed from blurry image " + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": " and event image " + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": " can then be obtained by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 281, + 287, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 281, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 125, + 281, + 287, + 310 + ], + "type": "interline_equation", + "content": "\\widehat {I} _ {\\boldsymbol {p}} (\\tau) = N _ {I} \\frac {B _ {\\boldsymbol {p}} (\\tau)}{\\widehat {R} _ {\\boldsymbol {p}} (\\tau)}. \\tag {9}", + "image_path": "88789cfa3b7e0442eb6c71007128f4935e93e9fff03a3f1b529552fb124f26c0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 312, + 268, + 325 + ], + "type": "text", + "content": " will be omitted in the following for simplicity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 331, + 208, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 208, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 208, + 344 + ], + "type": "text", + "content": "3.2. Event-based temporal fidelity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": "We consider events " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbb{E} = \\{e_k\\}_{k=1}^N" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": " captured in the exposure duration of the blurry image " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": ". To achieve the goal of latency correction, a no-reference-based objective should be defined since the ground truth is hard to collect. To this end, we propose event-based temporal fidelity (ETF) " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": " to quantify the sharpness profile of latent images " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": " reconstructed from events with latency " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "l_p" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": " corrected and their corresponding blurry image " + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 348, + 287, + 445 + ], + "type": "text", + "content": ", as shown in Figure 2 (a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": " should have smaller values as the estimated latency approaches its optimal values. Sharp images have more high-frequency contents compared with blurry images, which can be assessed by measuring the magnitude of its derivatives [3]. As shown in Figure 1 (d), the latent image results of EDI obtained from events with large latency appear with less sharpness and thicker edges. This motivates us to optimize the latency " + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": " by enforcing regularizations of the sharpness profile of the latent image " + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": " reconstructed from the EDI model in Eq. (9). Specifically, we define the average edge thickness of image gradients " + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\nabla I_p = (\\nabla_x I_p, \\nabla_y I_p)" + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": " as the number of pixels with values greater than a parameter " + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 445, + 287, + 601 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 604, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 604, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 604, + 287, + 631 + ], + "type": "interline_equation", + "content": "s _ {\\lambda} (I) = \\sum_ {\\boldsymbol {p}} H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right), \\tag {10}", + "image_path": "ce8705027c555280807c38d4d012f71ad7548f9a08efc4cd10540e61d17c3cb7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 200, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 200, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 200, + 645 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 634, + 200, + 645 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 47, + 634, + 200, + 645 + ], + "type": "text", + "content": " is the Heaviside function." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "content": "Integrating the right-side of the equation over " + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "content": " with a decreasing function " + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "content": ", we can obtain an " + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "content": "-irrelevant representation of " + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "inline_equation", + "content": "s(I)" + }, + { + "bbox": [ + 47, + 646, + 287, + 681 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 80, + 685, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 685, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 685, + 287, + 715 + ], + "type": "interline_equation", + "content": "s (I) = \\int_ {0} ^ {\\infty} \\sum_ {\\boldsymbol {p}} \\rho (\\lambda) H \\left(\\left\\| \\nabla I _ {\\boldsymbol {p}} \\right\\| _ {2} - \\lambda\\right) d \\lambda . \\tag {11}", + "image_path": "8a2e92f3345becff443b3f984e6a6d299da969904328b4fc00e6789502c2151d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 542, + 163 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 542, + 163 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 542, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 542, + 163 + ], + "type": "image", + "image_path": "e3cc03fe458577520879a89881bd4c164b3269ba8f2faf73105bd9413a3e9d9b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 164, + 502, + 174 + ], + "lines": [ + { + "bbox": [ + 361, + 164, + 502, + 174 + ], + "spans": [ + { + "bbox": [ + 361, + 164, + 502, + 174 + ], + "type": "text", + "content": "(a) The proposed latency correction method" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 320, + 178, + 410, + 255 + ], + "blocks": [ + { + "bbox": [ + 320, + 178, + 410, + 255 + ], + "lines": [ + { + "bbox": [ + 320, + 178, + 410, + 255 + ], + "spans": [ + { + "bbox": [ + 320, + 178, + 410, + 255 + ], + "type": "image", + "image_path": "d900b7fb95b2c3fd92a4739c71395346be410a7c31597e7c167d002d0fa54fa3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "lines": [ + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "spans": [ + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "type": "text", + "content": " vs. constant " + }, + { + "bbox": [ + 328, + 255, + 399, + 264 + ], + "type": "inline_equation", + "content": "l" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed latency correction method. (a) The pipeline of the proposed method, which takes synchronized blurry image " + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "text", + "content": " and events as input, and output the latency " + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "text", + "content": " by polynomial function " + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "inline_equation", + "content": "l(L)" + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "text", + "content": ". (b) The value of the proposed event temporal fidelity (ETF) " + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 304, + 275, + 545, + 363 + ], + "type": "text", + "content": " decreases as the estimated latency approaches ground-truth latency, which validates the efficacy of ETF as an objective. (c) Latency-differentiable reformulation of event integration, enabling gradient-based optimization." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 425, + 179, + 529, + 253 + ], + "blocks": [ + { + "bbox": [ + 425, + 179, + 529, + 253 + ], + "lines": [ + { + "bbox": [ + 425, + 179, + 529, + 253 + ], + "spans": [ + { + "bbox": [ + 425, + 179, + 529, + 253 + ], + "type": "image", + "image_path": "1292622a7a85402013c098c1fe8c26335eea58d5527c597d09e62b0f381c9b89.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 254, + 529, + 263 + ], + "lines": [ + { + "bbox": [ + 430, + 254, + 529, + 263 + ], + "spans": [ + { + "bbox": [ + 430, + 254, + 529, + 263 + ], + "type": "text", + "content": "(c) Differentiable event integral" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "text", + "content": "We adopt the function " + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\rho (\\lambda) = e^{-\\lambda}" + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "text", + "content": " in our experiments1. To pay more attention to the smoothness area, we multiply the image gradients " + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\nabla I_{p}" + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "text", + "content": " with an amplifier " + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 369, + 545, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "content": "It is noted that, for the robustness to scenes with possible thick edges on the image, " + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "content": " can not be set too small, which brings challenges for detecting high-frequency details. To compensate for this, we define another term about the contrast of the latent image " + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "content": ". Motivated by the fact that sharper images tend to have large contrast, we evaluate the contrast as the average local variation with local window " + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\pmb{p})" + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "content": " centered at pixel " + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 406, + 545, + 502 + ], + "type": "text", + "content": " [17]:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 342, + 505, + 545, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 505, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 342, + 505, + 545, + 535 + ], + "type": "interline_equation", + "content": "h (I) = \\sum_ {\\boldsymbol {p}} \\frac {1}{\\left| \\mathcal {N} _ {\\boldsymbol {p}} \\right|} \\sum_ {\\boldsymbol {q} \\in \\mathcal {N} _ {\\boldsymbol {p}}} \\left(I _ {\\boldsymbol {q}} - \\mu_ {I} \\left(\\mathcal {N} _ {\\boldsymbol {p}}\\right)\\right) ^ {2}, \\tag {12}", + "image_path": "1eb22adc50d491c33be49353df23da621f7ea286b397374f2907e43b36d74a1b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "inline_equation", + "content": "\\mu_I(\\mathcal{N}_p)" + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "text", + "content": " is the mean of the pixel value inside the window " + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "text", + "content": ". The reason we use a local variation rather than a global one is that the blurry effects caused by motion are not always globally consistent. Maximizing the local variation can emphasize the local sharpness and reduce the blur for each local window. In practice, we adopt local standard deviation " + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "inline_equation", + "content": "\\sqrt{h(I)}" + }, + { + "bbox": [ + 304, + 539, + 545, + 623 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 623, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 545, + 695 + ], + "type": "text", + "content": "For consecutive frames in a video, we can assume that their overall brightness level are similar, and thus the latency values are similar. Therefore, we define an indicator of temporal consistency by assuming Gaussian-distributed latency across consecutive frames. Given the latency of the last adjacent frame is " + }, + { + "bbox": [ + 304, + 623, + 545, + 695 + ], + "type": "inline_equation", + "content": "l'" + }, + { + "bbox": [ + 304, + 623, + 545, + 695 + ], + "type": "text", + "content": ", the indicator defined on the latency" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 535, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 535, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 535, + 713 + ], + "type": "text", + "content": "1 Please see the supplementary material for more details about " + }, + { + "bbox": [ + 317, + 702, + 535, + 713 + ], + "type": "inline_equation", + "content": "s(I)" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24980" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 205, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 205, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 205, + 83 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 72, + 205, + 83 + ], + "type": "text", + "content": " of the current frame is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 85, + 287, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 85, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 104, + 85, + 287, + 113 + ], + "type": "interline_equation", + "content": "g (l) = \\frac {1}{2 \\pi} \\exp \\left(- \\frac {\\left(l - l ^ {\\prime}\\right) ^ {2}}{2}\\right). \\tag {13}", + "image_path": "87e4bcf3e0138e72de3807f29386911bde3282a14603b0ca4b25f38d7a121e25.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 276, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 276, + 127 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 276, + 127 + ], + "type": "text", + "content": "Overall, " + }, + { + "bbox": [ + 47, + 115, + 276, + 127 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 47, + 115, + 276, + 127 + ], + "type": "text", + "content": " to determine optimal latency is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 131, + 287, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 131, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 113, + 131, + 287, + 159 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E T F}} (I, l) = \\frac {s (\\epsilon I)}{\\sqrt {h (I)} g (l)}. \\tag {14}", + "image_path": "de695b50489d32bf69a2a8e062bb87b92cc17e4d35d2b61014cf26f7a97e9b79.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 162, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 162, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 287, + 198 + ], + "type": "text", + "content": "The efficacy of the proposed " + }, + { + "bbox": [ + 47, + 162, + 287, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 47, + 162, + 287, + 198 + ], + "type": "text", + "content": " is validated in Figure 2 (b), which shows that its value decreases as the latency estimation error decreases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 206, + 230, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 206, + 230, + 219 + ], + "spans": [ + { + "bbox": [ + 47, + 206, + 230, + 219 + ], + "type": "text", + "content": "3.3. Learning-based latency correction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 225, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 287, + 274 + ], + "type": "text", + "content": "Based on the objective defined by " + }, + { + "bbox": [ + 47, + 225, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 47, + 225, + 287, + 274 + ], + "type": "text", + "content": ", we introduce a self-supervised approach founded on a parametrized form of latency, and a reformulated event-based double integral model that is differentiable with respect to latency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "content": "Latency-illuminance curve. The physical factors that have influences on the latency of events cannot be easily measured for the arbitrary situation. Fortunately, for deblurring and frame interpolation algorithms using event-image hybrid sensors such as [18, 19], the intensity of the image can indirectly indicate the illuminance levels in certain regions. Inspired by the previous findings about the relationship between latency and illuminance levels [7, 16, 21], we establish a latency-intensity relationship using a parameterized curve, which is validated in controlled environments (see illustrations in Figure 3 and discussions in Section 4.1 for more details, which shows that the latency increases as the illuminance decreases). To elucidate this relationship, we model the latency-illuminance curve as a " + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "content": "-th order polynomial " + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "content": " of the illuminance " + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 277, + 287, + 456 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 460, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 460, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 118, + 460, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\psi (L) = \\sum_ {i = 0} ^ {K} a _ {i} \\cdot f (L) ^ {i}, \\tag {15}", + "image_path": "dca40357ebbc22a8564a719b7356a014297442a8fefac44ec00f87561dcbf5e2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "\\{a_i\\}_{i = 0}^K" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " are the parameters, " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " is a fixed function to map the illuminance " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " into monotone decrement for consideration of easier optimization. The curve in Eq. (15) can effectively be a guidance for regularizing the estimation of spatially varying latency. In experiments, the illuminance " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " is unavailable. We adopt the blurry image " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " to approximate illuminance " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": " to estimate the spatially-varying latency, resulting in " + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "inline_equation", + "content": "l_{p} = \\psi (\\frac{B_{p}}{2\\Delta\\tau})" + }, + { + "bbox": [ + 47, + 497, + 287, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "text", + "content": "Latency-differentiable event integral. It is noted that the latency " + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "text", + "content": " derived from Eq. (9) is an index-like parameter, which has no derivatives. To allow gradient-based optimization, we establish a differentiable relationship between latency in events and specific indicators as shown in Figure 2 (c). By reformulating the event integral to be latency-differentiable, we derive a gradient chain linking " + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\{a_{i}\\}_{i = 0}^{K}" + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 47, + 597, + 287, + 693 + ], + "type": "text", + "content": " via the EDI model with latency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "text", + "content": "In contrast to Eq. (3) where events " + }, + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "inline_equation", + "content": "e_k = (\\pmb{p}_k, t_k, \\sigma_k)" + }, + { + "bbox": [ + 306, + 72, + 545, + 106 + ], + "type": "text", + "content": " are represented by an impulse function, we employ a piece-wise constant function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 109, + 545, + 132 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 109, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 345, + 109, + 545, + 132 + ], + "type": "interline_equation", + "content": "E _ {\\boldsymbol {p}} ^ {(k)} (s) = \\frac {\\sigma_ {k}}{t _ {k} - t _ {k - 1}}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right], \\tag {16}", + "image_path": "68f5f58ade774dfc08caf42d2f417436a5a1eff46c4ae9cd78df4211d73e9794.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "text", + "content": "where the changes " + }, + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "text", + "content": " is equally split into the several duration. Then integration of events in duration " + }, + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "inline_equation", + "content": "(t_{k - 1},t_k]" + }, + { + "bbox": [ + 305, + 135, + 545, + 169 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 349, + 172, + 545, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 172, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 349, + 172, + 545, + 197 + ], + "type": "interline_equation", + "content": "\\int E _ {\\boldsymbol {p}} ^ {(k)} (s) d s = \\sigma_ {k}, \\quad s \\in \\left(t _ {k - 1}, t _ {k} \\right]. \\tag {17}", + "image_path": "28c511d244097cba7c47803da7ae84b64d6fb9d111aafd2c73dd8032592c154d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": "For simplicity, we denote " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "E_{\\pmb{p}}^{(k)}(s)" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "E_{\\pmb{p}}^{(k)}" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": ". Consider the " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " events " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\{e_k\\}_{k=1}^N" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " triggered at pixel " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " during exposure time " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "[\\tau - l_p, t - l_p]" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "t_0 < \\tau - l_p < t_1" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "inline_equation", + "content": "t_{N-1} < t - l_p < t_n" + }, + { + "bbox": [ + 305, + 201, + 545, + 251 + ], + "type": "text", + "content": ", we can rewrite the inner integral part in Eq. (7) into:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 255, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 545, + 366 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int_ {\\tau - l _ {p}} ^ {t - l _ {p}} E _ {\\pmb {p}} (s) d s \\\\ = \\int_ {\\tau - l _ {p}} ^ {t _ {1}} E _ {\\boldsymbol {p}} ^ {(1)} d s + \\sum_ {k = 2} ^ {N - 1} \\int_ {t _ {k - 1}} ^ {t _ {k}} E _ {\\boldsymbol {p}} ^ {(k)} d s + \\int_ {t _ {N - 1}} ^ {t - l _ {\\boldsymbol {p}}} E _ {\\boldsymbol {p}} ^ {(N)} d s \\\\ = \\left(t _ {1} - l _ {\\boldsymbol {p}} - \\tau\\right) \\sigma_ {1} + \\sum_ {k = 2} ^ {N - 1} \\sigma_ {k} + \\left(t - l _ {\\boldsymbol {p}} - t _ {N - 1}\\right) \\sigma_ {n}, \\tag {18} \\\\ \\end{array}", + "image_path": "378e864e8127e6162e013589fb4870f20802c214b2ba12fcf8a6286a871d6778.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": "which is differentiable in the neighborhood of " + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": " with fixed " + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": ". It is noted that " + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": " will change when " + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "inline_equation", + "content": "\\tau - l_{p} < t_{0}" + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "inline_equation", + "content": "t - l_{p} \\geq t_{N}" + }, + { + "bbox": [ + 305, + 371, + 545, + 443 + ], + "type": "text", + "content": ", in which discontinuous changes on event triggering time will occur. While this might lead to a local minimum solution in optimization for a specific sample, it can be alleviated when optimizing over different samples." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": "Spatially-varying latency correction. With the differentiable form of event integral for parameters " + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\{a_i\\}_{i=0}^n" + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": ", we can now perform gradient descent in the dataset by defining a loss function " + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": ". Firstly, " + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ETF}}" + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": " defined in Eq. (14) is used as a term of the loss function. Heuristically, most of the latency " + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "inline_equation", + "content": "l_p" + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": " should lie in the range from 0 to the exposure time of the last blurry image, which can be optimized by the reformulated EDI. To ensure the latency does not exceed the previous exposure time so that the derivative can be calculated, a clipping operation is applied on " + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "inline_equation", + "content": "l_p" + }, + { + "bbox": [ + 305, + 448, + 545, + 580 + ], + "type": "text", + "content": ". A boundary loss is also used to limit the range of latency:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 350, + 583, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 583, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 350, + 583, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {b o u n d}} = \\sum_ {l _ {p} > 0} l _ {p} + \\sum_ {l _ {p} < \\tau - \\tau^ {\\prime} - \\Delta \\tau} - l _ {p}. \\tag {19}", + "image_path": "96ca0b95fe598d39f4744a860bf8528bcb63bf392bfd40836f236b4a63396079.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": "To deal with " + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": " exceeds the range of " + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "[0, \\tau - \\tau' - \\Delta \\tau]" + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": " and disables derivative computation due to gradients with too large magnitudes, we additionally optimize an intermediate constant latency " + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{const}}" + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": " shared across the sensor array during the exposure time of the blurry image. It has a relatively small solution space and can be solved efficiently by searching. " + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{const}}" + }, + { + "bbox": [ + 305, + 617, + 545, + 713 + ], + "type": "text", + "content": " provides an effective estimation of the overall latency (whose effectiveness is validated in the ex" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "type": "text", + "content": "2Please see the supplementary material for more details about " + }, + { + "bbox": [ + 58, + 702, + 276, + 713 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24981" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "periment results on synthetic data in Section 4.2). Thus, it is being adopted in the term " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mse}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " to minimize the distance between the current estimated latency and " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{const}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "l^{\\prime}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " in Eq. (13) is replaced by " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{const}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " to regularize using the frame's constant latency, rather than adjacent frames'." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": "From the latency-illuminance curve estimated previously, we can observe that the curve should be monotonic. Thus we enforce that the function " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": " is monotonically decreasing as the intensity increases. To apply the monotonicity constraint, we introduce the following constraint on the derivative of a function " + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "inline_equation", + "content": "l_{p}" + }, + { + "bbox": [ + 46, + 133, + 286, + 205 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 52, + 209, + 286, + 241 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 209, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 52, + 209, + 286, + 241 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d e v}} = \\sum \\max \\left(f ^ {\\prime} \\left(\\frac {B}{2 \\Delta \\tau}\\right) \\sum_ {i = 0} ^ {K} i \\cdot a _ {i} \\cdot f \\left(\\frac {B}{2 \\Delta \\tau}\\right) ^ {i - 1}, 0\\right), \\tag {20}", + "image_path": "578842a324a5c3e8243d445b242943f626881ad92b429cdd50bd526b207d3aae.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 245, + 286, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 245, + 286, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 245, + 286, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 245, + 286, + 269 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 245, + 286, + 269 + ], + "type": "text", + "content": " contains different values on each pixel to ensure monotonically decreasing on all the data samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 270, + 284, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 270, + 284, + 281 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 284, + 281 + ], + "type": "text", + "content": "Overall, the loss function to be minimized is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 286, + 286, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 286, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 69, + 286, + 286, + 299 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\alpha \\mathcal {L} _ {\\text {E T F}} + \\beta \\mathcal {L} _ {\\text {b o u n d}} + \\eta \\mathcal {L} _ {\\text {m s e}} + \\zeta \\mathcal {L} _ {\\text {d e v}}, \\tag {21}", + "image_path": "cd30161e2252ab6aa82a5c645bbfd89e13d35bccb1bf6a3abfd5fecd13aa57a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\eta, \\zeta" + }, + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "text", + "content": " are hyper-parameters for balancing different terms. Empirically, we set " + }, + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "inline_equation", + "content": "\\alpha = 1, \\beta = 1, \\eta = 20, \\zeta = 0.1" + }, + { + "bbox": [ + 47, + 304, + 286, + 339 + ], + "type": "text", + "content": " in our experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 349, + 143, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 349, + 143, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 143, + 361 + ], + "type": "text", + "content": "3.4. Training details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 371, + 286, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 286, + 431 + ], + "type": "text", + "content": "Dataset preparation. For latency correction training, we utilized a DAVIS346 event camera [1] to obtain blurry images and corresponding events. Our real-scene dataset comprises 9 videos, where 100 frames and their neighboring events from each video are selected for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 434, + 286, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 286, + 553 + ], + "type": "text", + "content": "Other implementation details. To simplify the implementation, we stack events into 500 bins to maintain high temporal resolution. Our model is implemented on the Pytorch framework and runs on an NVIDIA GeForce RTX 3090 GPU. We use ADAM [11] with default parameter setting for optimization. The batch size is set as 1. The learning rate was fixed to " + }, + { + "bbox": [ + 46, + 434, + 286, + 553 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 46, + 434, + 286, + 553 + ], + "type": "text", + "content": " in the earlier 2 epochs and reduced to 0 in the last 3 epochs by the linear decay strategy. Constant initialization to zero for the non-latency assumption is used for initialization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 566, + 127, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 566, + 127, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 566, + 127, + 579 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 586, + 286, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 286, + 599 + ], + "type": "text", + "content": "4.1. Latency estimation in controlled environments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "Experiments in controlled settings measure actual latency and illuminance. Our setup, shown in Figure 3 (a), includes a light source, an Arduino board, a power source, a photometer, and an event camera (DVS346Mono) with its lens removed to prevent light attenuation. The Arduino board was programmed to switch off the light at a known time " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ideal}}" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": ", and the subsequent OFF events were assumed to be triggered by this change. The latency of the " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "-th triggered event was calculated as the difference between " + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{ideal}}" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 432, + 152 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 432, + 152 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 432, + 152 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 432, + 152 + ], + "type": "image", + "image_path": "c1c520d6d45ffee476ee0e94ca90bd7056a98031ed6c0d3ea1fcabcdc20711f5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 153, + 416, + 162 + ], + "lines": [ + { + "bbox": [ + 325, + 153, + 416, + 162 + ], + "spans": [ + { + "bbox": [ + 325, + 153, + 416, + 162 + ], + "type": "text", + "content": "(a) Latency estimation system" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 544, + 153 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 544, + 153 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 544, + 153 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 544, + 153 + ], + "type": "image", + "image_path": "76b93f404960be0820c8fe1142643c2266b878e7112be868a58bcea08795fc24.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 153, + 537, + 162 + ], + "lines": [ + { + "bbox": [ + 448, + 153, + 537, + 162 + ], + "spans": [ + { + "bbox": [ + 448, + 153, + 537, + 162 + ], + "type": "text", + "content": "(b) Event trigger rate vs. time" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 310, + 163, + 432, + 256 + ], + "blocks": [ + { + "bbox": [ + 310, + 163, + 432, + 256 + ], + "lines": [ + { + "bbox": [ + 310, + 163, + 432, + 256 + ], + "spans": [ + { + "bbox": [ + 310, + 163, + 432, + 256 + ], + "type": "image", + "image_path": "01f09bed8f57f760866a5b8665d4fdd04d551275aead5e5cd2aefabcdcd5cd48.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 256, + 411, + 264 + ], + "lines": [ + { + "bbox": [ + 329, + 256, + 411, + 264 + ], + "spans": [ + { + "bbox": [ + 329, + 256, + 411, + 264 + ], + "type": "text", + "content": "(c) Latency vs. Illuminance" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "type": "text", + "content": "Figure 3. Estimation of latency-illuminance curve. (a) A system designed to investigate the relationship between illuminance levels and latency, approximated using an " + }, + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "type": "text", + "content": "-th order polynomial. (b) The comparison of event triggering rate under three varied lighting conditions. It is approximated by the first " + }, + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 277, + 545, + 365 + ], + "type": "text", + "content": " OFF events in our experiments. (c) Estimated latency under controlled lighting, highlighting increased latency with diminishing light. (d) Results of the polynomial approximation." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 163, + 544, + 256 + ], + "blocks": [ + { + "bbox": [ + 433, + 163, + 544, + 256 + ], + "lines": [ + { + "bbox": [ + 433, + 163, + 544, + 256 + ], + "spans": [ + { + "bbox": [ + 433, + 163, + 544, + 256 + ], + "type": "image", + "image_path": "cb6499a94a2293af0b518513eb87c9b9a4cdd73bdfd20b0f7e87c23d8bd49ed5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 256, + 537, + 264 + ], + "lines": [ + { + "bbox": [ + 448, + 256, + 537, + 264 + ], + "spans": [ + { + "bbox": [ + 448, + 256, + 537, + 264 + ], + "type": "text", + "content": "(d) Latency curve from Eq. (2)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "content": "and the time " + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{actual}}^{(i)}" + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "content": ". Illuminance is modulated by adjusting light voltage and measured alongside events by a photometer (with the lowest detectable value being 0.1 Lux). This process was repeated to mitigate noise effects (about 80 times in our experiments). The latency level of a given illuminance was calculated by the event triggering rate as shown in Fig. 3 (b), which is approximated by the median latency when the first " + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "content": " OFF events among all the samples are triggered for each process. By varying the illuminance, we empirically find that the latency increases as the illuminance decreases, as shown in Fig. 3 (c). We use the proposed parameterized curve to approximate the relationship between latency and illuminance, whose results are shown in Fig. 3 (d). It demonstrates an accurate curve fitting with " + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 304, + 369, + 545, + 574 + ], + "type": "text", + "content": ". Due to the difficulty in illuminance measuring, the estimated parameter cannot be directly applied in deblurring and interpolation for arbitrary scenes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 582, + 462, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 582, + 462, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 462, + 594 + ], + "type": "text", + "content": "4.2. Evaluation on synthetic data" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 603, + 545, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 698 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 698 + ], + "type": "text", + "content": "Compared methods. As the proposed method serves as a preprocessing step for latency correction of events, we evaluate the proposed methods based on two downstream tasks, that is, deblurring, and interpolation. Specifically, we test our method for deblurring algorithms including EDI [19], EFNet [24], and GEM [31], and for interpolation algorithms including EDI [19] and REFID [25]. We directly use the released pretrain model for comparison." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Dataset. To qualitatively evaluate downstream tasks, we" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24982" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 60, + 545, + 297 + ], + "blocks": [ + { + "bbox": [ + 49, + 60, + 545, + 297 + ], + "lines": [ + { + "bbox": [ + 49, + 60, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 60, + 545, + 297 + ], + "type": "image", + "image_path": "9ffb1dcda35a65fdd0939c2e61c8a94c0395d411c7e41bb5df202ec89b259d61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 331 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on synthetic data. We correct the latency of events by the constant latency correction (\"W/ corr. const.\") method and compared it with the ground truth events (GT input). All the methods are tested with the same input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 383, + 287, + 641 + ], + "blocks": [ + { + "bbox": [ + 46, + 335, + 288, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 378 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation on synthetic data. " + }, + { + "bbox": [ + 46, + 335, + 288, + 378 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 46, + 335, + 288, + 378 + ], + "type": "text", + "content": " means higher (lower) is better. The \"Diff\" shows the percentage improvement from \"W/o corr.\" to \"W/ corr. const.\" compared with \"GT input\"." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 383, + 287, + 641 + ], + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 641 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓
EDI [19]GT input28.4890.8390.489
W/o corr.26.0010.8100.500
DeblurringW/ corr. const.27.4710.8280.488
Diff. (%)+5.15+2.15-2.45
EFNet [24]GT input33.7610.9720.105
W/o corr.33.3300.9710.107
W/ corr. const.33.7780.9710.106
Diff. (%)+0.130-0.95
GEM [31]GT input30.6110.9630.183
W/o corr.27.7140.9340.198
W/ corr. const.30.3720.9580.185
Diff. (%)+8.68+2.49-7.10
EDI [19]GT input22.2930.8400.330
InterpolationW/o corr.22.1250.8340.331
W/ corr. const.22.2650.8380.331
Diff. (%)+0.63+0.470
REFID [25]GT input34.5220.9740.109
W/o corr.28.2440.9270.142
W/ corr. const.34.2170.9700.111
Diff. (%)+17.30+4.41-28.44
", + "image_path": "b1483ae6e1e03041fc736740e4272eeef6dc1f93c48314386e8895f0a4d4061f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "synthesize events with constant latency " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "l_{\\tau}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " for each pair of events and its corresponding blurry image. In particular, we utilize the event provided in the HighREV dataset [25] as ideal events, and then we add a constant latency to the event-image pairs for different event sequences. We employ" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 432 + ], + "type": "text", + "content": "the training data in the HighREV dataset which contains 1615 frames from 19 videos for testing to provide more convincing results. The spatially-varying latency correction is highly related to the properties of event cameras and depends on different scene illuminances, which can not be faithfully synthetic by event simulators [7, 10]. Therefore, we only provide results of the constant latency estimated by the proposed methods for quantitative comparisons." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 434, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 546, + 555 + ], + "type": "text", + "content": "Metrics. To show the corrected performance, all the tested methods are evaluated with four kinds of input, events without latency (terms \"GT input\"), events with latency which is not corrected by our method (terms \"W/o corr.\"), the events corrected by constant latency " + }, + { + "bbox": [ + 304, + 434, + 546, + 555 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{const}}" + }, + { + "bbox": [ + 304, + 434, + 546, + 555 + ], + "type": "text", + "content": " (terms \"W/ corr. const.). We adopt three commonly used metrics to measure the performance of deblurring and interpolation, peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and the perceptual error with learned perceptual image patch similarity (LPIPS) [29], respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "Results. The quantitative evaluation results are shown in Table 1. We add the \"Diff.\" row to show the improvement of the proposed method. It shows that the proposed method can improve the performance of deblurring and interpolation by event latency correction. For deep learning methods like EFNet [24], they implicitly considered latency effects in training data generation, which makes them more robust to latency in events. However, their performance can also be improved with the latency correction. The qualitative evaluation results are shown in Figure 4, which show that the proposed method can explicitly reduce the artifacts brought by latency, and it performs more closely to the ground truth events (GT input). Specifically, the latency correction can" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24983" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 61, + 545, + 292 + ], + "blocks": [ + { + "bbox": [ + 47, + 61, + 545, + 292 + ], + "lines": [ + { + "bbox": [ + 47, + 61, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 47, + 61, + 545, + 292 + ], + "type": "image", + "image_path": "206ec58a7943040845c5eb6fe3849466592b9cb2143602b926df79a602de296e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 546, + 316 + ], + "type": "text", + "content": "Figure 5. The qualitative comparison on real data. We correct the latency of events by constant latency correction (\"W/ corr. Const.\") and spatially-varying latency correction (\"W/ corr\"). All the methods are tested with the same input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 288, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 326, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 288, + 387 + ], + "type": "text", + "content": "reduce the light strips brought by previous edges, in which the events are triggered with latency. The non-learning interpolation algorithm EDI [19] is more strict with the time accuracy and the noise in events, which makes it still blurry even with the GT input." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 392, + 180, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 180, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 180, + 403 + ], + "type": "text", + "content": "4.3. Evaluation on real data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "text", + "content": "In order to reveal the performance of the proposed method, we evaluate our method on real data captured by DAVIS346Mono. Because the ground truth sharp images are unavailable, we can only qualitatively evaluate the proposed method on real data. The results on real data" + }, + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 410, + 287, + 542 + ], + "type": "text", + "content": " are shown in the left two columns of Figure 5. It shows that both the proposed latency correction methods can reduce the artifacts and blurry effects. Processed by the proposed latency correction methods, the number of inaccuracy events was explicitly reduced to help the deblurring and interpolation algorithms to produce natural and sharp edges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 547, + 139, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 139, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 139, + 560 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 565, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 696 + ], + "type": "text", + "content": "We compare the performance of spatially-varying latency correction with constant latency correction on real data as shown in the right two columns in Figure 5. It shows that events corrected by the spatially-varying method have more natural views than constant latency correction, which can also provide sharper results3. Benefiting from the data captured under different illumination conditions, the spatially varying latency correction can also provide more accurate latency estimation for specific scenes. Along with the different illumination conditions, it can also address the different latency properties for different pixels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 324, + 379, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 379, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 379, + 337 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 345, + 545, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 476 + ], + "type": "text", + "content": "In this paper, we present the first method for latency correction in improving event-guided deblurring and interpolation tasks. Our approach introduces event-based temporal fidelity to evaluate the sharpness of latent images reconstructed from latency-corrected events and their corresponding blurry images. This objective guides the optimization of parameters in a parametrized latency model through gradient-based optimization, facilitated by a reformulated, latency-differentiable, event-based double integral model. The efficacy of our methods is demonstrated using both synthetic and real-world data sets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 485, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 546, + 617 + ], + "type": "text", + "content": "Limitations. Although with potential for broad applicability across different event cameras, the present implementation of the proposed method is specifically tailored for DAVIS346Mono cameras due to that it is the only event camera we have with on-chip synchronization and our current method requires perfect temporal synchronization between the image and events. Moreover, due to challenges in measuring illuminance and curve variability, the practicality of estimated parameters of the latency-illuminance curve varies across sensors. The curves for more sensors, while significant, has not been explored in this paper." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 621, + 404, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 404, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 404, + 634 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 633, + 545, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 545, + 704 + ], + "type": "text", + "content": "This work was supported by National Science and Technology Major Project (Grant No. 2021ZD0109803), Beijing Natural Science Foundation (Grant No. L233024), National Natural Science Foundation of China (Grant No. 62302019, 62088102, 62136001), and SenseTime Collaborative Research Grant." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 265, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 265, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 265, + 713 + ], + "type": "text", + "content": "3More comparisons are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24984" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 2014. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 178 + ], + "type": "text", + "content": "[2] Tobi Delbruck and Carver A Mead. Adaptive photoreceptor with wide dynamic range. In Proc. of IEEE International Symposium on Circuits and Systems-ISCAS, pages 339–342, 1994. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 179, + 288, + 223 + ], + "type": "text", + "content": "[3] Guillermo Gallego, Mathias Gehrig, and Davide Scaramuzza. Focus Is All You Need: Loss Functions for Event-Based Vision. In Proc. of Computer Vision and Pattern Recognition, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] Michael D. Grossberg and Shree K. Nayar. Modeling the space of camera response functions. IEEE Transactions on Pattern Analysis and Machine Intelligence, 26:1272-1282, 2004. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 312 + ], + "type": "text", + "content": "[5] Chen Haoyu, Teng Minggui, Shi Boxin, Wang Yizhou, and Huang Tiejun. Learning to Deblur and Generate High frame rate video with an event camera. arXiv preprint arXiv:2003.00847, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 346 + ], + "type": "text", + "content": "[6] Ondrej Holevský, Radoslav Skoviera, Václav Hlaváč, and Roman Vitek. Experimental Comparison between Event and Global Shutter Cameras. Sensors, 21(4), 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[7] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic DVS events. In Proc. of Computer Vision and Pattern Recognition Workshops, 2021. 2, 3, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[8] invitation. Understanding the Performance of Neuromorphic Event-based Vision Sensors. https://inivation.com/wp-content/uploads/2020/05/White-Paper-May-2020.pdf, 2020.1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 467 + ], + "type": "text", + "content": "[9] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning Event-Based Motion Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 512 + ], + "type": "text", + "content": "[10] Damien Joubert, Alexandre Marcireau, Nic Ralph, Andrew Jolley, André van Schaik, and Gregory Cohen. Event camera simulator improvements via characterized parameters. Frontiers in Neuroscience, 15:702765, 2021. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 545 + ], + "type": "text", + "content": "[11] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "text", + "content": "[12] Juan Antonio Lénero-Bardallo, Teresa Serrano-Gotarredona, and Bernabé Linares-Barranco. A 3.6 " + }, + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 48, + 547, + 288, + 601 + ], + "type": "text", + "content": "s Latency Asynchronous Frame-Free Event-Driven Dynamic-Vision-Sensor. IEEE Journal of Solid-State Circuits, 46(6):1443-1455, 2011. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": "[13] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 48, + 602, + 288, + 646 + ], + "type": "text", + "content": " dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor. IEEE Journal of Solid-State Circuits, 43(2):566-576, 2008. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 690 + ], + "type": "text", + "content": "[14] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In Proc. of European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[15] Brian McReynolds, Rui Graca, and Tobi Delbruck. Experimental methods to predict dynamic vision sensor event" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "camera performance. Optical Engineering, 61(7):074103-074103, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 152 + ], + "type": "text", + "content": "[16] Brian J Mcreynolds, Rui Graca, Daniel O'Keefe, Rachel Oliver, Richard Balthazor, Nathaniel George, and Matthew McHarg. Modeling and decoding event-based sensor lightning response. In Proc. of Unconventional Imaging, Sensing, and Adaptive Optics 2023, 2023. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 196 + ], + "type": "text", + "content": "[17] Patrenahalli M. Narendra and Robert C. Fitch. Real-time Adaptive contrast enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence, (6):655-661, 1981. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 200, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 200, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 200, + 547, + 243 + ], + "type": "text", + "content": "[18] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proc. of Computer Vision and Pattern Recognition, 2019. 2, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 246, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 246, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 308, + 246, + 545, + 300 + ], + "type": "text", + "content": "[19] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 303, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 357 + ], + "type": "text", + "content": "[20] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor With Lossless Pixel-Level Video Compression and Time-Domain CDS. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2011. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 415 + ], + "type": "text", + "content": "[21] Jonah P Sengupta. Demystifying Event-Based Camera Latency: Sensor Speed Dependence on Pixel Biasing, Light, and Spatial Activity. In Proc. of SPIE Conference 13045: Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XXXV, 2023. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "text", + "content": "[22] Teresa Serrano-Gotarredona and Bernabé Linares-Barranco. A " + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "text", + "content": " Contrast Sensitivity " + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "text", + "content": " FPN " + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "text", + "content": " Latency " + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "inline_equation", + "content": "4\\mathrm{mW}" + }, + { + "bbox": [ + 307, + 417, + 545, + 472 + ], + "type": "text", + "content": " Asynchronous Frame-Free Dynamic Vision Sensor Using Transimpedance Preamplifiers. IEEE Journal of Solid-State Circuits, 48(3):827-838, 2013. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 545, + 518 + ], + "type": "text", + "content": "[23] Wei Shang, Dongwei Ren, Dongqing Zou, Jimmy S Ren, Ping Luo, and Wangmeng Zuo. Bringing events into video deblurring with non-consecutively blurry frames. In Proc. of International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 545, + 575 + ], + "type": "text", + "content": "[24] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based Fusion for Motion Deblurring with Cross-modal Attention. In Proc. of European Conference on Computer Vision, 2022. 1, 2, 6, 7, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 577, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 545, + 632 + ], + "type": "text", + "content": "[25] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based Frame Interpolation with Ad-Hoc Deblurring. In Proc. of Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 634, + 545, + 677 + ], + "type": "text", + "content": "[26] Minggui Teng, Chu Zhou, Hanyue Lou, and Boxin Shi. NEST: Neural event stack for event-based image enhancement. In Proc. of European Conference on Computer Vision, 2022. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 714 + ], + "type": "text", + "content": "[27] Bishan Wang, Jingwei He, Lei Yu, Gui-Song Xia, and Wen Yang. Event enhanced high-quality image recovery. In Proc. of European Conference on Computer Vision, 2020. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24985" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 284 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Fang Xu, Lei Yu, Bishan Wang, Wen Yang, Gui-Song Xia, Xu Jia, Zhendong Qiao, and Jianzhuang Liu. Motion Deblurring with Real Events. In Proc. of International Conference on Computer Vision, pages 2583-2592, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 162 + ], + "type": "text", + "content": "[29] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of Computer Vision and Pattern Recognition, 2018. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 195 + ], + "type": "text", + "content": "[30] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In Proc. of Computer Vision and Pattern Recognition, 2022. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 197, + 286, + 239 + ], + "type": "text", + "content": "[31] Xiang Zhang, Lei Yu, Wen Yang, Jianzhuang Liu, and GuiSong Xia. Generalizing Event-Based Motion Deblurring in Real-World Scenarios. In Proc. of International Conference on Computer Vision, 2023. 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "type": "text", + "content": "[32] Chu Zhou, Minggui Teng, Jin Han, Jinxiu Liang, Chao Xu, Gang Cao, and Boxin Shi. Deblurring Low-Light Images with Events. International Journal of Computer Vision, 131 (5):1284-1298, 2023. 1" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24986" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_content_list.json b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5a34ff424cd14d614e33448c1071eb2c5bd2dec3 --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_content_list.json @@ -0,0 +1,1328 @@ +[ + { + "type": "text", + "text": "Latent Modulated Function for Computational Optimal Continuous Image Representation", + "text_level": 1, + "bbox": [ + 202, + 130, + 769, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongyao He1 Zhi Jin1,2,*", + "bbox": [ + 374, + 202, + 601, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University $^{2}$ Guangdong Provincial Key Laboratory of Fire Science and Technology", + "bbox": [ + 133, + 220, + 836, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hezy28@mail2.sysu.edu.cn, jinzh26@mail.sysu.edu.cn", + "bbox": [ + 258, + 258, + 705, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent work Local Implicit Image Function (LIIF) and subsequent Implicit Neural Representation (INR) based works have achieved remarkable success in Arbitrary-Scale Super-Resolution (ASSR) by using MLP to decode Low-Resolution (LR) features. However, these continuous image representations typically implement decoding in High-Resolution (HR) High-Dimensional (HD) space, leading to a quadratic increase in computational cost and seriously hindering the practical applications of ASSR. To tackle this problem, we propose a novel Latent Modulated Function (LMF), which decouples the HR-HD decoding process into shared latent decoding in LR-HD space and independent rendering in HR Low-Dimensional (LD) space, thereby realizing the first computational optimal paradigm of continuous image representation. Specifically, LMF utilizes an HD MLP in latent space to generate latent modulations of each LR feature vector. This enables a modulated LD MLP in render space to quickly adapt to any input feature vector and perform rendering at arbitrary resolution. Furthermore, we leverage the positive correlation between modulation intensity and input image complexity to design a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to adjust the decoding efficiency based on the rendering precision. Extensive experiments demonstrate that converting existing INR-based ASSR methods to LMF can reduce the computational cost by up to $99.9\\%$ , accelerating inference by up to $57\\times$ , and save up to $76\\%$ of parameters, while maintaining competitive performance. The code is available at https://github.com/HeZongyao/LMF.", + "bbox": [ + 76, + 340, + 473, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 806, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image data has long been stored and presented as discrete 2D arrays with fixed resolutions, while the visual world we perceive is continuous and precise. As a result, there is a", + "bbox": [ + 75, + 832, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d93c6cf7827c50f011fcc04ba860ebfc839404c1ed92ff3fe1706222a1c5cfc8.jpg", + "image_caption": [ + "Figure 1. Efficiency comparisons $(320 \\times 180$ input) for ASSR. LMF-based ASSR methods significantly reduce computational cost (MACs), runtime (circle sizes), and parameters (colors)." + ], + "image_footnote": [], + "bbox": [ + 504, + 306, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "growing demand for arbitrary scale image resizing in various real-world scenarios. For instance, users often require continuous resizing of images during capture, viewing, and editing. Recent developments in Single Image SuperResolution (SISR) [7, 9, 12, 16, 20, 21, 23-25, 36, 45-47] have showcased significant advancements in reconstructing High-Resolution (HR) images from Low-Resolution (LR) counterparts. However, traditional interpolation methods are still the preferred choice for handling real-world image resizing. This is because SISR methods typically learn upsampling functions for a particular scale factor, requiring training different models for different scales. In contrast, Arbitrary-Scale Super-Resolution (ASSR) utilizes a single model to handle arbitrary scales, thus presenting significant potential in practical image-resizing applications.", + "bbox": [ + 496, + 551, + 892, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building on the success of Implicit Neural Representation (INR), Local Implicit Image Function (LIIF) [11] is a recent ASSR approach that learns to continuously represent 2D images through an implicit function. LIIF encodes an image as latent codes and employs a decoding MLP to predict pixel values based on coordinates and their neighboring latent codes. The use of continuous coordinates enables LIIF and subsequent INR-based ASSR works", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26026", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76fb012507164f8c3e888194a2cf40616913ae4e357db06ad6b86223a62fc94b.jpg", + "image_caption": [ + "Figure 2. The framework of our LMF-based continuous image representation. Given a latent code generated by the encoder, an HD latent MLP generates the latent modulation to adjust the features of an LD render MLP, thereby achieving efficient arbitrary-resolution rendering." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 880, + 218 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[5, 8, 14, 15, 22, 34, 37-41] to represent images at any resolution, achieving robust performance in up to $\\times 30$ SR task. However, existing INR-based ASSR methods adhere to an intuitive ASSR paradigm that uses a complex decoding function in HR High-Dimensional (HD) space to independently predict the pixel value of each HR coordinate. As a result, their computational cost and runtime increase quadratically with the scale factor, contradicting the goal of real-world image-resizing applications (refer to Figure 1). Upon closer examination, the significant computational redundancy of HR-HD decoding can be divided into three aspects. Firstly, there is inherent similarity within the HR region corresponding to an LR latent code, suggesting the need for merging shared computations within this region before independently predicting each HR pixel value. Secondly, this paradigm ignores the varying complexity of different LR latent codes and applies the same computational process to predict all the HR pixel values. Lastly, this paradigm couples the HD MLP parameters and HR computations, while such a large number of parameters are not always needed for each prediction.", + "bbox": [ + 75, + 265, + 472, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle this problem, we propose a novel Latent Modulated Function (LMF) to realize a computational optimal paradigm of continuous image representation. By leveraging latent modulation, LMF successfully disentangles the decoding process from HR-HD space into latent (LR-HD) and render (HR-LD) spaces. As shown in Figure 2, LMF first utilizes a latent MLP to generate the latent modulation of a given latent code. When querying the coordinates near the latent code, the latent modulation is applied to adjust each hidden linear layer in the render MLP. This modulation enables LMF to minimize the dimension of the render MLP without sacrificing performance. Furthermore, we observe a positive correlation between the latent modulation intensity and the input image complexity. Building on this insight, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm. CMSR not only allows the computational cost of rendering to be proportional to the input image complexity rather than the output resolution, but also provides the flexibility to balance precision and efficiency during testing. Computational optimal and fast ASSR holds great promise for future research and downstream appli", + "bbox": [ + 75, + 584, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cations. Experiments demonstrate that transforming existing INR-based ASSR methods into LMF reduces computational costs by $90.4\\%$ to $99.9\\%$ , accelerates inference by $2.4 \\times$ to $56.9 \\times$ , and saves $45.1\\%$ to $76.0\\%$ of parameters, while maintaining competitive PSNR performance.", + "bbox": [ + 496, + 265, + 890, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 340, + 859, + 354 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Latent Modulated Function (LMF) that realizes computational optimal continuous image representation, which decouples the HR-HD decoding into decoding in latent (LR-HD) and render (HR-LD) spaces.", + "- Based on the correlation between modulation intensity and input image complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to effortlessly balance the rendering efficiency and precision at test time.", + "- Extensive experiments demonstrate that seamlessly converting existing INR-based ASSR methods to LMF can achieve substantial reductions in computational cost, runtime, and parameters, while preserving performance." + ], + "bbox": [ + 500, + 356, + 890, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 549, + 640, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Implicit Neural Representation", + "text_level": 1, + "bbox": [ + 500, + 574, + 774, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "INR has demonstrated remarkable success for continuous signal representation across various tasks and modalities, including signal reconstruction [10, 27, 31, 32], image and video compression [6, 13, 35], and image generation [33, 43]. The highlight of INR research is NeRF [28] and subsequent works [2, 3, 30, 42], which learn MLPs to infer the RGB and density of any given coordinates and view directions in a 3D scene. However, a major challenge faced by INR is the extensive training process required to fit a signal from scratch. To address this, MINER [31] and KiloNeRF [30] propose the use of multiple small MLPs to independently learn different segments of a signal. Another solution is hyper-base network design [10, 13, 27, 33], where a pre-trained hyper network generates INR for any input signal during testing. To represent complex signals, hyper networks often generate modulations for base networks.", + "bbox": [ + 496, + 597, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Single-Image Super-Resolution", + "text_level": 1, + "bbox": [ + 500, + 847, + 774, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SISR as a classic low-level vision task has been studied for decades. In recent years, this field witnessed a variety of", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26027", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CNN-based networks, such as SRCNN [12], LapSRN [20], SRResNet [21], EDSR [25], MemNet [36], RCAN [46], RDN [47], and SRFBN [23]. Lately, image processing transformers, such as IPT [7], SwinIR [24], and HAT [9], have been proposed, which surpass CNN-based methods in SR performance using a large dataset. Although SISR methods have achieved remarkable performance, most of them are designed for a specific scale.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ASSR methods are designed to utilize a single network for all scales, showcasing significant advancements in practicality and convenience over previous works. Hu et al. proposed MetaSR [17] to generate an upsampling layer for ASSR using a meta-network. However, MetaSR exhibits poor generalization for large-scale SR. Building upon the success of INR, a novel continuous image representation was proposed in LIIF [11], where each pixel value is inferred from an MLP using its coordinate and neighboring 2D features. Subsequently, a variety of INR-based ASSR methods have been proposed [5, 8, 14, 15, 22, 34, 37-41]. Notably, Lee et al. proposed LTE [22], which introduces a dominant-frequency estimator to continuously capture fine details. Cao et al. introduced CiaoSR [5], a continuous implicit attention-in-attention network that learns ensemble weights for nearby local features, achieving State-Of-The-Art (SOTA) performance. Despite the robust performance of INR-based ASSR methods for scales up to $\\times 30$ , they adhere to the paradigm of decoding in HR-HD space, leading to prohibitively high computational cost and runtime.", + "bbox": [ + 75, + 210, + 470, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 527, + 174, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary and Motivation", + "text_level": 1, + "bbox": [ + 76, + 551, + 328, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In continuous image representations, an image $I_{in}$ is represented as 2D feature maps $F$ in latent space $\\mathbb{R}^{H_{in} \\times W_{in} \\times D_F}$ . A latent code $z^* \\in \\mathbb{R}^{D_F}$ represents a feature vector derived from the feature maps, which is used to decode its nearest coordinates in the continuous spatial domain $X \\subset \\mathbb{R}^2$ . Since the encoder can be any feature extractor from existing SISR methods, the essence of continuous image representations is the decoding function, which arbitrarily upsamples LR-HD features to HR-LD pixel values. A simple and intuitive paradigm is to learn a decoding function in HR-HD space that independently decodes the pixel value of each HR coordinate. This vanilla decoding function $f_{\\theta}$ parameterized as an HD MLP $\\theta$ can be formulated as:", + "bbox": [ + 75, + 575, + 468, + 770 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta} \\left(z ^ {*}, x _ {i, j} ^ {*}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 782, + 468, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{i,j}^{*}\\in X^{*}$ refers to any nearest coordinate of $z^{*}$ , and $I(x_{i,j}^{*})$ represents the predicted pixel value at $x_{i,j}^{*}$ .", + "bbox": [ + 76, + 809, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, as mentioned earlier in the Introduction section, this vanilla paradigm suffers from three aspects of computational redundancy. A computational optimal decoding function should allow computations proportional to", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the input image complexity rather than the output resolution. With this in mind, we propose to decouple the HR-HD decoding process into shared decoding in latent space and independent rendering in render space, along with a rendering algorithm adaptive to the input image content.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Latent Modulated Function", + "text_level": 1, + "bbox": [ + 498, + 174, + 750, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To eliminate the computational redundancy in HR-HD decoding, we first propose a two-stage decoding process comprising a latent MLP $\\theta_{l}$ and a render MLP $\\theta_{r}$ . The latent MLP handles the shared decoding process within the HR region of a latent code $z^{*}$ , while the render MLP independently predicts the pixel value of each HR coordinate. The LR-HD output $f_{\\theta_l}(z^*)$ from the latent MLP is utilized to assist the render MLP in rendering. However, directly using this output $f_{\\theta_l}(z^*)$ as the input of the render MLP is akin to using the latent code itself, as demonstrated in Eq. 1, resulting in no efficiency improvement yet:", + "bbox": [ + 496, + 196, + 890, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta_ {r}} \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right), x _ {i, j} ^ {*}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 371, + 890, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, to render arbitrary pixels accurately, it is necessary for the render MLP to operate in HR space. Therefore, the most viable approach to minimize rendering computations is to reduce the dimension or depth of the render MLP. However, an HD render MLP is still needed in Eq. 2 to fully receive and process the output $f_{\\theta_l}(z^*)$ , which contradicts our intention of utilizing an LD render MLP. To overcome these limitations, modulations become a suitable solution to connect the latent and render space. Modulations have been used in INR [13, 27, 33] and are often associated with the hyper-base network design. A hyper network converts an input image to modulations, which are then utilized by a small base network to map coordinates to a continuous image. However, these modulations are shared by the input image and work at the image level, resulting in unsatisfactory performance and challenging training. Hence, we introduce a latent modulation $m^{*} \\in \\mathbb{R}^{D_M}$ , which is generated by the latent MLP using the latent code $z^{*}$ . The latent modulation consists of the FiLM layers [29], where each layer consists of a scale modulation and a shift modulation:", + "bbox": [ + 496, + 396, + 892, + 698 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm ^ {*} = \\left[ \\alpha_ {1} ^ {*}, \\beta_ {1} ^ {*}, \\dots , \\alpha_ {K} ^ {*}, \\beta_ {K} ^ {*} \\right] = S p l i t \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 707, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha^{*}$ represents the element-wise scale modulation, $\\beta^{*}$ represents the element-wise shift modulation, and $K$ represents the depth of hidden layers in the render MLP.", + "bbox": [ + 496, + 731, + 890, + 775 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each HR coordinate, the latent modulation $m^{*}$ of the latent code $z^{*}$ nearest to this coordinate is applied to the render MLP, and is applied to the hidden features $h^{*} \\in \\mathbb{R}^{D_{H}}$ after each hidden linear layer:", + "bbox": [ + 496, + 776, + 890, + 837 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k + 1} ^ {*} = f _ {\\theta_ {r} ^ {k + 1}} \\left(\\sigma \\left(\\left(1 + \\alpha_ {k} ^ {*}\\right) \\odot h _ {k} ^ {*} + \\beta_ {k} ^ {*}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 845, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h_k^*$ represents the hidden features after the $k$ -th hidden linear layer, and $\\sigma$ represents the activation function.", + "bbox": [ + 496, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26028", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ecc20c03c6c14fe0ce269b16d540715265ee23591ea5c0665c4a77bda9afa34d.jpg", + "image_caption": [ + "Figure 3. Visualization $(\\times 4$ SR) of the normalized mean of shift modulation (left), and the normalized residual between the LM-LIIF predicted image and the bilinear upsampled image (right)." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 264, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ca0ef0fa8030fd66f06b6d03f4752b998a50135c9dc33bf6601303abc65fd04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 87, + 436, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, the proposed latent modulation effectively solves the incompatibility between LR-HD latent codes and HR-LD rendering and significantly disentangles the computations from parameters. For a latent modulation with 192 dimensions, we can accurately render images using a 7-layer render MLP with only 16 dimensions. This LD render MLP only learns the general rendering rules of natural images, while the latent modulation optimizes its rendering process based on the characteristics of each input latent code. As a result, a tiny render MLP can quickly adapt to any input latent code and accomplish efficient rendering at arbitrary resolution. A remaining concern is that the latent codes in existing INR-based ASSR methods often possess high dimensions, resulting in unnecessary computations when directly input into the render MLP. Therefore, we compress the HD latent code to an LD one $z_{c}^{*}$ using the latent MLP, extending Eq. 3 to the first stage in Eq. 5. Building on the latent modulation, we propose a computational optimal paradigm of continuous image representation, called Latent Modulated Function (LMF):", + "bbox": [ + 76, + 262, + 472, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: $[m^{*},z_{c}^{*}] = S p l i t(f_{\\theta_{l}}(z^{*}))$ (5)", + "bbox": [ + 151, + 570, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: $I(x_{i,j}^{*}) = f_{\\theta_r,m^*}(z_c^*,x_{i,j}^*)$", + "bbox": [ + 153, + 590, + 392, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LMF is a universal paradigm for efficient ASSR and is compatible with existing INR-based ASSR methods. By effortlessly relocating the time-consuming operations and large MLPs to latent space, LMF allows a minimal render MLP to perform the same level of ASSR as the original ASSR method. In addition, LMF effectively decouples the computational expensive techniques of feature unfolding and local ensemble [11]. Specifically, LMF only utilizes feature unfolding for the latent MLP, where nearby $3 \\times 3$ feature vectors are concatenated as inputs. LMF performs local ensemble exclusively for the render MLP, which merges the predictions from the top-left, top-right, bottom-left, and bottom-right latent codes to achieve continuous predictions.", + "bbox": [ + 75, + 613, + 470, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Controllable Multi-Scale Rendering", + "text_level": 1, + "bbox": [ + 76, + 816, + 390, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The HR regions of two latent codes often have different signal complexity, allowing for the use of different decoding functions. However, the challenge lies in obtaining the signal complexity of each latent code. The modulation scheme", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da796c39b49cd2f67a7bb8191164a67b5b76657f5707106e84e1ab0cfbf5d767.jpg", + "image_caption": [ + "Figure 4. The positive correlation between the means of shift modulation and the minimum rendering scale factors in LM-LIIF." + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 851, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Eq. 4 suggests that the intensity of latent modulation is proportional to the signal complexity of its latent code. A lower intensity implies that the signal can be sufficiently rendered by the render MLP itself. In Figure 3, we display the normalized mean of shift modulation (left), and the normalized residual between the SR result of LMF-based LIIF and the bilinear upsampled result (right). It is evident that the means of modulation closely align with the signal complexity of the predicted image. Regions with higher means of modulation consistently correspond to more complex textures and edges. Hence, we employ the mean of shift modulation as the intensity of latent modulation.", + "bbox": [ + 496, + 263, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In LMF, we define the signal complexity of the latent code $z^{*}$ as the minimum scale factor $s^{*}$ needed to fully represent its signal. In simple terms, any rendering with a scale factor $s, s > s^{*}$ will yield the same result as rendering with $s^{*}$ followed by upsample interpolation (i.e. bilinear) with scale $\\frac{s}{s^{*}}$ . Therefore, we only need to render $z^{*}$ with scale $\\min(s^{*}, s)$ when upsampling an image with a scale $s$ . To obtain the minimum scale factor of each mean of shift modulation, in Figure 4, we sample 100 means of shift modulation uniformly from the range of [0, 1] and measure their minimum scale factors on the LR images in the training set. Figure 4 clearly shows the positive correlation between the means of shift modulation and the minimum scale factors. However, since they do not have a one-to-one correspondence, we create a Scale2Mod table by sampling common scale factors and recording their minimum and maximum means of shift modulation.", + "bbox": [ + 496, + 446, + 892, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the Scale2Mods table, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to decode latent codes with their minimum scale factors. Firstly, we initialize $s_{-1}$ as 1, $I_{s_{-1}}$ as the LR image. For each sampled scale $s_i \\in [s_0, s_1, \\ldots, s]$ , we first upsample the last result $I_{s_{i-1}}$ using bilinear interpolation with $\\frac{s_i}{s_{i-1}}$ to obtain $I_{s_i}$ . We then query the Scale2Mods table to obtain the range of modulation means $[m_{min}^{s_i}, m_{max}^{s_i}]$ for $s_i$ . For $s_i$ , the render MLP only needs to render the latent codes whose corresponding modulation means falling within this range. We repeat the above process iteratively until we obtain the final SR result or until all latent codes have been decoded. For precision controllable rendering, when creating", + "bbox": [ + 496, + 702, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26029", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3c873803e3295f6b3a555ac8ec4491e2af5922ce1b093550f94afe3951677341.jpg", + "table_caption": [ + "Table 1. Quantitative comparison (PSNR (dB)/MACs) for ASSR on DIV2K validation set. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodParamsIn-scaleOut-of-scale
×2×3×4×6×12×18
Bicubic-31.01/-28.22/-26.66/-24.82/-22.27/-21.00/-
EDSR-b [25]-34.55/-30.90/-28.94/----
LIIF [11]355K34.67/3.567T30.96/3.567T29.00/3.567T26.75/3.567T23.71/3.567T22.17/3.567T
LM-LIIF170K34.65/125.32G30.95/65.37G28.99/44.38G26.75/25.30G23.72/6.61G22.19/2.95G
LTE [22]506K34.72/2.228T31.02/2.123T29.04/2.086T26.81/2.059T23.78/2.044T22.23/2.026T
LM-LTE278K34.71/198.81G30.99/100.75G29.03/66.43G26.78/34.96G23.74/9.07G22.21/4.10G
CiaoSR-1.4M [5]1.429M34.91/11.460T31.15/11.443T29.23/11.437T26.95/11.433T23.88/11.430T22.32/11.348T
LM-CiaoSR753K34.81/1.071T31.10/501.36G29.12/293.01G26.86/135.66G23.80/42.60G22.26/23.90G
", + "bbox": [ + 102, + 114, + 867, + 263 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7dedcc786b6edd74f38c50afb5570bb26264bb1f5a63046762fc3299d1715439.jpg", + "table_caption": [ + "Table 2. Quantitative comparison (PSNR (dB)) for in-scale SR on test sets. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodSet5Set14BSD100Urban100DIV2K100
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [47]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6134.9431.2229.19
LIIF [11]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6834.9931.2629.27
LM-LIIF38.1834.7232.5034.0630.5528.8332.3229.2827.7532.8228.8126.6634.9631.2529.27
LTE [22]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8135.0431.3229.33
LM-LTE38.2334.7632.5334.1130.5628.8632.3729.3127.7833.0328.9626.8035.0331.3129.32
CiaoSR-1.4M [5]38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1135.1531.4229.45
LM-CiaoSR38.2434.8132.5934.1130.6228.9032.3929.3227.8033.1329.1326.9535.0831.3629.37
SwinIR [24]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0735.2631.5029.48
LIIF [11]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1535.1731.4629.46
LM-LIIF38.3134.8832.7834.3130.7929.0232.4329.3727.8533.4629.4327.2135.2131.4829.49
LTE [22]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2435.2431.5029.51
LM-LTE38.3234.8832.7734.2830.7929.0132.4629.3927.8733.5229.4427.2435.2431.5029.50
CiaoSR-3.1M [5]38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4235.2931.5529.59
LM-CiaoSR38.3734.9432.8734.3330.8329.0832.4829.4327.9033.6529.5827.3935.3031.5529.56
", + "bbox": [ + 81, + 301, + 888, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the Scale2Mods table, we set an MSE threshold $t$ and define the signal complexity of a latent code as the minimum scale factor needed to render its signal within this error $t$ . By effortlessly adjusting the MSE threshold during testing, CMSR is capable of rendering with varying precision and speed, while adapting naturally to the image content.", + "bbox": [ + 75, + 549, + 468, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 655, + 209, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 76, + 680, + 272, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. Following the setups in previous ASSR methods [5, 11, 22], we employ the DIV2K dataset [1] as the training set. In each training iteration, we uniformly sample a downsampling scale $s$ from a continuous range of $[ \\times 1, \\times 4 ]$ . HR patches with spatial size of $48s \\times 48s$ are cropped and bicubic downsampled to LR patches with spatial size of $48 \\times 48$ . To supervise the training process, we randomly sample $48 \\times 48$ pixels from each HR patch. For evaluation, we present results on the DIV2K validation set (DIV2K 100) along with four widely-used benchmark test sets: Set5 [4], Set14 [44], BSD100 [26], and Urban100 [18].", + "bbox": [ + 75, + 703, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We transform three previous ASSR methods, namely LIIF [11], LTE [22], and CiaoSR", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[5], into our proposed LMF. LM-LIIF employs a 2-layer latent MLP with 208 dimensions, and a 7-layer render MLP with 16 dimensions. As for LM-LTE, we employ a 2-layer latent MLP with 256 dimensions, a 9-layer render MLP with 16 dimensions, and a local texture estimator [22] with 128 dimensions. LM-CiaoSR employs four 2-layer latent MLPs with 256 dimensions to generate key, value, and query information, followed by a 9-layer render MLP with 16 dimensions. The detailed architectures of LMF-based methods are provided in the supplementary material.", + "bbox": [ + 496, + 549, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize EDSR-baseline (EDSR-b) [25], RDN [47], and SwinIR [24] without their upsampling modules as encoders. For both EDSR-b and RDN encoders, we conduct training for 1000 epochs with a batch size of 16. The learning rate is initialized as $1 \\times 10^{-4}$ and decayed by a factor of 0.5 every 200 epochs. As for the SwinIR encoder, we conduct training for 1000 epochs with a batch size of 32. The learning rate is initialized as $2 \\times 10^{-4}$ and decayed by a factor of 0.5 at [500, 800, 900, 950] epochs. For optimization, we utilize the L1 pixel loss as the loss function and employ the Adam [19] optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ . Our training details align with previous ASSR works [5, 22].", + "bbox": [ + 496, + 702, + 892, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Settings. In line with previous methods [5, 11,", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26030", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f6b54eab31ade454104942ef8bcac0b8fe27a5247b5fda5b360207394ae80f83.jpg", + "table_caption": [ + "Table 3. Quantitative comparison (PSNR (dB)) for out-of-scale SR on test sets. The best results are highlighted in bold. The first 6 rows use RDN as the encoder, and the last 6 rows use SwinIR as the encoder." + ], + "table_footnote": [], + "table_body": "
MethodSet5Set14BSD100Urban100DIV2K100
×6×8×12×6×8×12×6×8×12×6×8×12×6×12×18
LIIF [11]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1526.9923.8922.34
LM-LIIF29.1827.2124.7126.6625.1723.3125.9824.9323.5824.1422.7721.1526.9923.9122.36
LTE [22]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.0423.9522.40
LM-LTE29.2127.2224.8226.6925.1523.3226.0124.9523.6124.2622.8621.2427.0323.9522.39
CiaoSR-1.4M [5]29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.1624.0622.48
LM-CiaoSR29.2427.2624.8026.7225.1823.3026.0324.8123.6024.4022.9821.3327.0723.9822.42
LIIF [11]29.4627.3624.9926.8225.3423.3926.0725.0123.6524.5923.1421.4327.1524.0222.43
LM-LIIF29.4627.3624.9826.8225.3723.4326.0825.0323.6524.6123.1521.4927.1824.0622.49
LTE [22]29.5027.3525.0726.8625.4223.4426.0925.0323.6624.6223.1721.5027.2024.0922.50
LM-LTE29.5227.3925.0126.8325.3823.4126.1025.0323.6624.6323.1721.5027.1924.0822.50
CiaoSR-3.1M [5]29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6027.2824.1522.54
LM-CiaoSR29.5227.4225.0226.8425.4123.4526.1225.0623.6824.7423.2621.5827.2424.1222.53
", + "bbox": [ + 81, + 128, + 888, + 335 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9296bc04e468dc0dcb5daa8c97468a019a2398227d5119a73495c4413b046f31.jpg", + "table_caption": [ + "Table 4. Efficiency comparison (MACs/GPU runtime) for ASSR. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
DecoderParamsGPU mem.180p270p360p
720p1080p2k720p1080p2k720p1080p2k
LIIF [11]355K7.7GB1.160T/1.1148s2.609T/2.5230s4.638T/4.5232s1.160T/1.1195s2.609T/2.5347s4.638T/4.5544s1.160T/1.1272s2.609T/2.5456s4.638T/4.5589s
LM-LIIF170K5.3GB14.43G/0.0246s18.72G/0.0474s24.72G/0.0795s25.39G/0.0363s32.46G/0.0593s37.50G/0.0919s40.74G/0.0541s47.81G/0.0770s57.71G/0.1107s
LTE [22]506K8.4GB694.34G/0.9135s1.506T/2.0366s2.666T/3.7578s714.11G/0.9191s1.526T/2.0497s2.685T/3.7694s741.80G/0.9258s1.553T/2.0637s2.712T/3.7791s
LM-LTE278K4.7GB21.60G/0.0382s25.87G/0.0750s33.21G/0.1254s39.53G/0.0545s48.59G/0.0905s52.88G/0.1414s64.63G/0.0774s73.69G/0.1123s86.38G/0.1640s
CiaoSR-1.4M [5]1.429M8.8GB3.718T/1.8820s8.362T/4.1718s14.864T/7.3619s3.721T/1.9838s8.365T/4.2999s14.867T/7.5611s3.725T/2.1355s8.370T/4.5064s14.871T/7.8366s
LM-CiaoSR753K8.9GB95.23G/0.2637s99.57G/0.3849s106.68G/0.5546s204.09G/0.5263s214.32G/0.6476s217.83G/0.8160s356.46G/0.8844s366.69G/1.0013s381.02G/1.1778s
", + "bbox": [ + 81, + 373, + 888, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "22], we employ PSNR as the objective criterion to assess quality performance. To evaluate efficiency, we measure the computational cost using Multiply-Accumulate Operations (MACs) and quantify the model speed using the runtime on an RTX 2080 Ti GPU. Additionally, we provide the number of parameters and GPU memory usage. Notably, the MACs, runtime, GPU memory, and parameters are evaluated solely on the decoder unless explicitly stated.", + "bbox": [ + 75, + 553, + 470, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. ASSR Comparison", + "text_level": 1, + "bbox": [ + 76, + 691, + 259, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 1, we compare our LMF-based methods with reference ASSR methods, namely LIIF [11], LTE [22], and CiaoSR [5], on the DIV2K validation set using EDSR-b as the encoder. We include the results of bicubic interpolation and EDSR-b with its initial upsampling module for reference. The results demonstrate that LMF-based methods achieve a significant reduction in computational cost while delivering competitive PSNR performance compared to the original methods. Compared to LIIF, LM-LIIF achieves a reduction of $96.6\\%$ to $99.9\\%$ in computational cost across all scales. Similarly, LM-LTE and LM-CiaoSR provide $91.3\\%$ to $99.8\\%$ and $90.4\\%$ to", + "bbox": [ + 75, + 719, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$99.8\\%$ reduction in computational cost compared to LTE and CiaoSR, respectively. However, LM-CiaoSR exhibits slightly lower PSNR than CiaoSR due to relocating the expensive attention-in-attention modules from HR-HD space to latent space for the minimal computational cost.", + "bbox": [ + 496, + 553, + 890, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2 and 3, we present the comprehensive PSNR evaluation for ASSR using RDN and SwinIR as the encoders. Notably, CiaoSR utilizes a 1.429M decoder for EDSR-b and RDN, and a 3.065M decoder for SwinIR, while our LM-CiaoSR employs the same decoder for all encoders. The results demonstrate the competitive PSNR performance of our LMF-based methods, especially LM-LIIF and LM-LTE, compared to the reference methods on benchmark test sets. Notably, LMF-based methods exhibit superior PSNR performance gains on the SwinIR encoder. The results presented in Table 1, 2, and 3 collectively validate the universal applicability of our LMF across different encoders, INR-based ASSR methods, test sets, and scales.", + "bbox": [ + 496, + 636, + 892, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 4, we present comparisons of MACs, runtime, GPU memory, and parameters for ASSR across multiple input and output resolutions. We utilize bicubic downsampled images from the DIV2K validation set as inputs for", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26031", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c95688909c8a3a98fc36451b03ecff56aa19ef690da168c9230e63e1b69b47b6.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison for ASSR. All of the ASSR methods use SwinIR as the encoder." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 880, + 296 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8968c3d8a12b677a57c57b070a6e274b45a720e0f2af6dadb09c76f387002bc.jpg", + "table_caption": [ + "Table 5. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on latent modulations and latent codes, using EDSR-b as the encoder." + ], + "table_footnote": [], + "table_body": "
DecoderLatent codeLatent modulationLatent MLPRender MLPParamsIn-scaleOut-of-scale
×2×4×8×16
LIIF576××5 × 256355K34.67/3.567T29.00/3.567T25.40/3.567T22.61/3.567T
LIIF (tiny)576××9 × 1612K33.61/117.44G28.18/117.44G24.78/117.44G22.16/117.44G
LIIF (C2F)192×2 × 1927 × 16202K34.11/221.54G28.63/114.62G25.11/73.04G22.38/58.91G
LM-LIIF64Scale2 × 1927 × 16136K34.61/110.94G28.96/46.87G25.38/30.79G22.61/26.58G
LM-LIIF64Shift2 × 1927 × 16136K34.59/110.94G28.94/46.87G25.36/30.79G22.59/26.58G
LM-LIIF64Scale & shift2 × 1927 × 16155K34.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
LM-LIIF16Scale & shift2 × 2087 × 16170K34.65/125.32G28.99/44.38G25.40/24.10G22.62/18.90G
", + "bbox": [ + 80, + 361, + 888, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "each resolution, and all results exclude the encoders. As shown in the table, our LMF-based methods exhibit a significant reduction in computational cost, ranging from $90.7\\%$ to $99.5\\%$ , and a substantial reduction in runtime, ranging from $58.6\\%$ to $98.2\\%$ , compared to the reference methods. Compared to the reference methods, LMF-based methods also exhibit a reduction in parameters, ranging from $45.1\\%$ to $52.1\\%$ , which is attributed to the decoupling of computations from parameters, demonstrating the significance of our elegant design over merely stacking parameters.", + "bbox": [ + 75, + 507, + 468, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. In Figure 5, we present a qualitative comparison for ASSR, using SwinIR as the encoder. The visual results demonstrate that our LMF-based methods achieve competitive visual quality across all scales compared to the reference methods. Remarkably, LMF-based methods achieve continuous and precise results while significantly reducing computational cost and runtime. This confirms that LMF effectively addresses the challenge of upsampling LR-HD features to HR-LD pixel values, eliminating the need for an HR-HD decoding function. More visual results are provided in the supplementary materials.", + "bbox": [ + 75, + 659, + 468, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 843, + 228, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we investigate the effect of each component in LMF. Both PSNR and MACs results on the DIV2K vali", + "bbox": [ + 75, + 869, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dation set are provided for a comprehensive evaluation, and all of the models use EDSR-b as the encoder. Notably, the CMSR algorithm is not used in Table 5 and Table 6.", + "bbox": [ + 496, + 507, + 890, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Latent Modulation. In Table 5, we evaluate the impact of latent modulation. Firstly, the comparison between LIIF and LIIF (tiny) reveals a significant decline in PSNR performance when shrinking the decoder from a 5-layer MLP with 256 dimensions to a 9-layer MLP with 16 dimensions. This performance drop is expected due to the substantial reduction in network parameters. However, our LM-LIIF model, which utilizes a 7-layer render MLP with 16 dimensions, outperforms LIIF in terms of PSNR, while consuming computational cost even lower than LIIF (tiny). Furthermore, we introduce a coarse-to-fine variant of LIIF, where both stages employ the same structure as LM-LIIF. Unlike LM-LIIF, LIIF (C2F) fails to achieve competitive PSNR performance compared to LIIF, highlighting the critical role of latent modulation in reducing the dimensions of the render MLP. These results also validate that HR-HD computations are redundant for decoding LR-HD features.", + "bbox": [ + 496, + 553, + 892, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 5, we also compare the performance of different types of latent modulations, namely the scale modulation, the shift modulation, and the combination of scale and shift modulations. The results demonstrate that LM-LIIF achieves the highest PSNR performance when both scale and shift modulations are employed, and the computational", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26032", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fceee7d91c70857265418261404b246856933a56a9d4311a524a0eabba083df5.jpg", + "table_caption": [ + "Table 6. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on the latent MLP and the render MLP, using EDSR-b as the encoder, and LM-LIIF as the decoder." + ], + "table_footnote": [], + "table_body": "
Latent MLPRender MLPIn-scaleOut-of-scale
×2×4×8×16
2 × 2562 × 12834.66/236.91G28.97/131.21G25.37/104.57G22.60/97.18G
2 × 2563 × 6434.67/232.18G29.00/126.48G25.41/99.85G22.63/92.50G
2 × 2565 × 3234.68/197.37G29.00/91.67G25.41/65.11G22.63/58.04G
4 × 2569 × 1634.64/258.37G28.98/87.79G25.40/45.04G22.62/34.13G
2 × 2569 × 1634.66/171.85G28.99/66.16G25.40/39.65G22.62/32.79G
2 × 1927 × 1634.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
2 × 1285 × 1634.63/79.76G28.98/35.02G25.39/23.79G22.62/20.83G
", + "bbox": [ + 78, + 142, + 468, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cost remains nearly identical. Finally, we compare using the latent code (64 dimensions) from the encoder and the compressed latent code (16 dimensions) from the latent MLP as the input of the render MLP. The results indicate that using the compressed latent code can further reduce the computational cost for large scales. Based on these findings, we incorporate both scale and shift modulations as well as the compressed latent code in our final LMF models.", + "bbox": [ + 75, + 356, + 468, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Network Structure. We evaluated different depths and dimensions of the latent MLP and the render MLP in Table 6. Notably, the compressed latent code is not used to ensure better alignment of depths and dimensions between the two MLPs. For the latent MLP, we observe that decreasing the depth of the latent MLP from 4 to 2 yields nearly the same PSNR performance, which suggests that the latent MLP does not require a deep architecture or a large number of parameters. However, when reducing the dimension of the latent MLP from 192 to 128, the PSNR performance begins to decrease. This observation confirms that HD computations are crucial for a powerful latent MLP.", + "bbox": [ + 75, + 477, + 468, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 6, we investigate the depths and dimensions of the render MLP by fixing the dimensions of the latent MLP and the latent modulation at 256. The results demonstrate that we can effortlessly reduce the dimension of the render MLP from 128 to 16, resulting in an impressive up to $66.26\\%$ reduction in computational cost, without compromising performance. When the latent modulation possesses a fixed dimension, it is consistently viable to utilize a lower dimensional rendering process, thereby minimizing the computational burden. All things considered, our final LM-LIIF model employs a 2-layer latent MLP with 208 dimensions (16 dimensions for the compressed latent code) and a 7-layer render MLP with 16 dimensions, optimizing computational efficiency without sacrificing performance.", + "bbox": [ + 75, + 657, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Controllable Multi-Scale Rendering. In Table 7, we evaluate the MSE thresholds when creating the Scale2Mod ta", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c8b0b128273aa1bf5c9e77dc7bfbfdb0411d28bae37616628f3aabf74f3ef02c.jpg", + "table_caption": [ + "Table 7. Ablation study (PSNR(dB)/MACs of rendering/MACs of decoder on DIV2K validation set) on MSE thresholds of CMSR, using EDSR-b as the encoder, and LM-LIIF as the decoder." + ], + "table_footnote": [], + "table_body": "
MSE thresholdIn-scaleOut-of-scale
×2×4×8×16
×34.65/17.41G/125.32G28.99/17.41G/44.38G25.40/17.37G/24.10G22.62/17.23G/18.90G
2 × 10-534.65/16.30G/124.22G28.99/13.67G/40.62G25.40/12.41G/19.14G22.62/3.48G/5.15G
1 × 10-434.62/13.63G/121.55G28.98/11.71G/38.69G25.39/8.35G/15.08G22.62/2.50G/4.17G
5 × 10-434.42/10.35G/118.27G28.93/8.94G/35.92G25.38/6.08G/12.81G22.61/2.29G/3.96G
", + "bbox": [ + 504, + 142, + 888, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ble for CMSR. Compared to the case without CMSR, a Scale2Mod table created with an MSE threshold of $2 \\times 10^{-5}$ achieves the same PSNR performance while reducing computational cost by $20.58\\%$ and $72.75\\%$ for scale $\\times 8$ and $\\times 16$ . With lower precision requirements, i.e., MSE thresholds of $1 \\times 10^{-4}$ and $5 \\times 10^{-4}$ , CMSR effectively reduces the computational cost even further while maintaining maximum PSNR performance. This adaptability is advantageous for real-world applications as the precision can be adjusted during testing without the need for model retraining. However, CMSR exclusively reduces the computational cost of the render MLP, which limits the computational cost reduction for small scales. Hence, in the Experiments section, we only employ CMSR with the MSE threshold $2 \\times 10^{-5}$ for out-of-distribution scales. More details and results regarding CMSR are provided in the supplementary material.", + "bbox": [ + 496, + 273, + 890, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 536, + 617, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a Latent Modulated Function (LMF), which realizes a computational optimal paradigm of continuous image representation. LMF utilizes an MLP in latent space to generate the latent modulation of each latent code. This enables a modulated MLP in render space to perform effective rendering at any resolution. Moreover, based on the correlation between modulation intensity and signal complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to balance the rendering precision and efficiency during testing. Experiments demonstrate that converting INR-based ASSR methods to LMF leads to substantial reductions in computational cost, runtime, and parameters, while preserving performance. Better forms of latent modulations and more efficient decoder architectures will be explored in future work.", + "bbox": [ + 496, + 564, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 811, + 679, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant No. 62071500; Supported by Shenzhen Science and Technology Program under Grant No. JCYJ20230807111107015.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26033", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2", + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. British machine vision conference, 2012. 5", + "[5] Jiezhang Cao, Qin Wang, Yongqin Xian, Yawei Li, Bingbing Ni, Zhiming Pi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Ciaosr: Continuous implicit attention-inattention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 5, 6", + "[6] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. Nerv: Neural representations for videos. Advances in Neural Information Processing Systems, 34:21557-21568, 2021. 2", + "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 3", + "[8] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 2, 3", + "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 3", + "[10] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, pages 170-187. Springer, 2022. 2", + "[11] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 3, 4, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184-199, 2014. 1, 3", + "[13] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2, 3", + "[14] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2, 3", + "[15] Zongyao He and Zhi Jin. Dynamic implicit image function for efficient arbitrary-scale image representation. arXiv preprint arXiv:2306.12321, 2023. 2, 3", + "[16] Zongyao He, Zhi Jin, and Yao Zhao. Srdrl: A blind superresolution framework with degradation reconstruction loss. IEEE Transactions on Multimedia, 24:2877-2889, 2021. 1", + "[17] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1575-1584, 2019. 3", + "[18] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5", + "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 1, 3", + "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 1, 3", + "[22] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 2, 3, 5, 6", + "[23] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 1, 3", + "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3, 5" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26034", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 3, 5", + "[26] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 416-423, 2001. 5", + "[27] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated periodic activations for generalizable local functional representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, 2021. 2, 3", + "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[29] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3", + "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[31] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2", + "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2", + "[33] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 3", + "[34] Gaochao Song, Qian Sun, Luo Zhang, Ran Su, Jianfeng Shi, and Ying He. Ope-sr: Orthogonal position encoding for designing a parameter-free upsampling module in arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10009-10020, 2023. 2, 3", + "[35] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74–91. Springer, 2022. 2", + "[36] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4539-4547, 2017. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4801-4810, 2021. 2, 3", + "[38] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Hang Wang, Zhengyan Tong, and Yutian Liu. Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1786-1795, 2023.", + "[39] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021.", + "[40] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021.", + "[41] Jie-En Yao, Li-Yuan Tsao, Yi-Chen Lo, Roy Tseng, Chia-Che Chang, and Chun-Yi Lee. Local implicit normalizing flow for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1776–1785, 2023. 2, 3", + "[42] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2", + "[43] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2", + "[44] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5", + "[45] Huanrong Zhang, Jie Xiao, and Zhi Jin. Multi-scale image super-resolution via a single extendable deep network. IEEE Journal of Selected Topics in Signal Processing, 15(2):253-263, 2020. 1", + "[46] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 3", + "[47] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 3, 5" + ], + "bbox": [ + 501, + 90, + 892, + 782 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26035", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_model.json b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68673aa38f89cae74637e3595ce41d0d080b005b --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_model.json @@ -0,0 +1,1892 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.77, + 0.177 + ], + "angle": 0, + "content": "Latent Modulated Function for Computational Optimal Continuous Image Representation" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.203, + 0.602, + 0.221 + ], + "angle": 0, + "content": "Zongyao He1 Zhi Jin1,2,*" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.221, + 0.838, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University \\(^{2}\\)Guangdong Provincial Key Laboratory of Fire Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.26, + 0.706, + 0.275 + ], + "angle": 0, + "content": "hezy28@mail2.sysu.edu.cn, jinzh26@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.341, + 0.474, + 0.78 + ], + "angle": 0, + "content": "The recent work Local Implicit Image Function (LIIF) and subsequent Implicit Neural Representation (INR) based works have achieved remarkable success in Arbitrary-Scale Super-Resolution (ASSR) by using MLP to decode Low-Resolution (LR) features. However, these continuous image representations typically implement decoding in High-Resolution (HR) High-Dimensional (HD) space, leading to a quadratic increase in computational cost and seriously hindering the practical applications of ASSR. To tackle this problem, we propose a novel Latent Modulated Function (LMF), which decouples the HR-HD decoding process into shared latent decoding in LR-HD space and independent rendering in HR Low-Dimensional (LD) space, thereby realizing the first computational optimal paradigm of continuous image representation. Specifically, LMF utilizes an HD MLP in latent space to generate latent modulations of each LR feature vector. This enables a modulated LD MLP in render space to quickly adapt to any input feature vector and perform rendering at arbitrary resolution. Furthermore, we leverage the positive correlation between modulation intensity and input image complexity to design a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to adjust the decoding efficiency based on the rendering precision. Extensive experiments demonstrate that converting existing INR-based ASSR methods to LMF can reduce the computational cost by up to \\(99.9\\%\\), accelerating inference by up to \\(57\\times\\), and save up to \\(76\\%\\) of parameters, while maintaining competitive performance. The code is available at https://github.com/HeZongyao/LMF." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.808, + 0.21, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.833, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Image data has long been stored and presented as discrete 2D arrays with fixed resolutions, while the visual world we perceive is continuous and precise. As a result, there is a" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.307, + 0.891, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.501, + 0.894, + 0.545 + ], + "angle": 0, + "content": "Figure 1. Efficiency comparisons \\((320 \\times 180\\) input) for ASSR. LMF-based ASSR methods significantly reduce computational cost (MACs), runtime (circle sizes), and parameters (colors)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.893, + 0.78 + ], + "angle": 0, + "content": "growing demand for arbitrary scale image resizing in various real-world scenarios. For instance, users often require continuous resizing of images during capture, viewing, and editing. Recent developments in Single Image SuperResolution (SISR) [7, 9, 12, 16, 20, 21, 23-25, 36, 45-47] have showcased significant advancements in reconstructing High-Resolution (HR) images from Low-Resolution (LR) counterparts. However, traditional interpolation methods are still the preferred choice for handling real-world image resizing. This is because SISR methods typically learn upsampling functions for a particular scale factor, requiring training different models for different scales. In contrast, Arbitrary-Scale Super-Resolution (ASSR) utilizes a single model to handle arbitrary scales, thus presenting significant potential in practical image-resizing applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Building on the success of Implicit Neural Representation (INR), Local Implicit Image Function (LIIF) [11] is a recent ASSR approach that learns to continuously represent 2D images through an implicit function. LIIF encodes an image as latent codes and employs a decoding MLP to predict pixel values based on coordinates and their neighboring latent codes. The use of continuous coordinates enables LIIF and subsequent INR-based ASSR works" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26026" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.088, + 0.882, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.894, + 0.258 + ], + "angle": 0, + "content": "Figure 2. The framework of our LMF-based continuous image representation. Given a latent code generated by the encoder, an HD latent MLP generates the latent modulation to adjust the features of an LD render MLP, thereby achieving efficient arbitrary-resolution rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.266, + 0.473, + 0.584 + ], + "angle": 0, + "content": "[5, 8, 14, 15, 22, 34, 37-41] to represent images at any resolution, achieving robust performance in up to \\(\\times 30\\) SR task. However, existing INR-based ASSR methods adhere to an intuitive ASSR paradigm that uses a complex decoding function in HR High-Dimensional (HD) space to independently predict the pixel value of each HR coordinate. As a result, their computational cost and runtime increase quadratically with the scale factor, contradicting the goal of real-world image-resizing applications (refer to Figure 1). Upon closer examination, the significant computational redundancy of HR-HD decoding can be divided into three aspects. Firstly, there is inherent similarity within the HR region corresponding to an LR latent code, suggesting the need for merging shared computations within this region before independently predicting each HR pixel value. Secondly, this paradigm ignores the varying complexity of different LR latent codes and applies the same computational process to predict all the HR pixel values. Lastly, this paradigm couples the HD MLP parameters and HR computations, while such a large number of parameters are not always needed for each prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To tackle this problem, we propose a novel Latent Modulated Function (LMF) to realize a computational optimal paradigm of continuous image representation. By leveraging latent modulation, LMF successfully disentangles the decoding process from HR-HD space into latent (LR-HD) and render (HR-LD) spaces. As shown in Figure 2, LMF first utilizes a latent MLP to generate the latent modulation of a given latent code. When querying the coordinates near the latent code, the latent modulation is applied to adjust each hidden linear layer in the render MLP. This modulation enables LMF to minimize the dimension of the render MLP without sacrificing performance. Furthermore, we observe a positive correlation between the latent modulation intensity and the input image complexity. Building on this insight, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm. CMSR not only allows the computational cost of rendering to be proportional to the input image complexity rather than the output resolution, but also provides the flexibility to balance precision and efficiency during testing. Computational optimal and fast ASSR holds great promise for future research and downstream appli" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.342 + ], + "angle": 0, + "content": "cations. Experiments demonstrate that transforming existing INR-based ASSR methods into LMF reduces computational costs by \\(90.4\\%\\) to \\(99.9\\%\\), accelerates inference by \\(2.4 \\times\\) to \\(56.9 \\times\\), and saves \\(45.1\\%\\) to \\(76.0\\%\\) of parameters, while maintaining competitive PSNR performance." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.342, + 0.86, + 0.356 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.357, + 0.892, + 0.417 + ], + "angle": 0, + "content": "- We propose a Latent Modulated Function (LMF) that realizes computational optimal continuous image representation, which decouples the HR-HD decoding into decoding in latent (LR-HD) and render (HR-LD) spaces." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.418, + 0.892, + 0.478 + ], + "angle": 0, + "content": "- Based on the correlation between modulation intensity and input image complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to effortlessly balance the rendering efficiency and precision at test time." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.479, + 0.892, + 0.539 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that seamlessly converting existing INR-based ASSR methods to LMF can achieve substantial reductions in computational cost, runtime, and parameters, while preserving performance." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.357, + 0.892, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.642, + 0.566 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.575, + 0.776, + 0.591 + ], + "angle": 0, + "content": "2.1. Implicit Neural Representation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.84 + ], + "angle": 0, + "content": "INR has demonstrated remarkable success for continuous signal representation across various tasks and modalities, including signal reconstruction [10, 27, 31, 32], image and video compression [6, 13, 35], and image generation [33, 43]. The highlight of INR research is NeRF [28] and subsequent works [2, 3, 30, 42], which learn MLPs to infer the RGB and density of any given coordinates and view directions in a 3D scene. However, a major challenge faced by INR is the extensive training process required to fit a signal from scratch. To address this, MINER [31] and KiloNeRF [30] propose the use of multiple small MLPs to independently learn different segments of a signal. Another solution is hyper-base network design [10, 13, 27, 33], where a pre-trained hyper network generates INR for any input signal during testing. To represent complex signals, hyper networks often generate modulations for base networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.776, + 0.864 + ], + "angle": 0, + "content": "2.2. Single-Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "SISR as a classic low-level vision task has been studied for decades. In recent years, this field witnessed a variety of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26027" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.212 + ], + "angle": 0, + "content": "CNN-based networks, such as SRCNN [12], LapSRN [20], SRResNet [21], EDSR [25], MemNet [36], RCAN [46], RDN [47], and SRFBN [23]. Lately, image processing transformers, such as IPT [7], SwinIR [24], and HAT [9], have been proposed, which surpass CNN-based methods in SR performance using a large dataset. Although SISR methods have achieved remarkable performance, most of them are designed for a specific scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.515 + ], + "angle": 0, + "content": "ASSR methods are designed to utilize a single network for all scales, showcasing significant advancements in practicality and convenience over previous works. Hu et al. proposed MetaSR [17] to generate an upsampling layer for ASSR using a meta-network. However, MetaSR exhibits poor generalization for large-scale SR. Building upon the success of INR, a novel continuous image representation was proposed in LIIF [11], where each pixel value is inferred from an MLP using its coordinate and neighboring 2D features. Subsequently, a variety of INR-based ASSR methods have been proposed [5, 8, 14, 15, 22, 34, 37-41]. Notably, Lee et al. proposed LTE [22], which introduces a dominant-frequency estimator to continuously capture fine details. Cao et al. introduced CiaoSR [5], a continuous implicit attention-in-attention network that learns ensemble weights for nearby local features, achieving State-Of-The-Art (SOTA) performance. Despite the robust performance of INR-based ASSR methods for scales up to \\(\\times 30\\), they adhere to the paradigm of decoding in HR-HD space, leading to prohibitively high computational cost and runtime." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.176, + 0.543 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.329, + 0.568 + ], + "angle": 0, + "content": "3.1. Preliminary and Motivation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.469, + 0.771 + ], + "angle": 0, + "content": "In continuous image representations, an image \\(I_{in}\\) is represented as 2D feature maps \\(F\\) in latent space \\(\\mathbb{R}^{H_{in} \\times W_{in} \\times D_F}\\). A latent code \\(z^* \\in \\mathbb{R}^{D_F}\\) represents a feature vector derived from the feature maps, which is used to decode its nearest coordinates in the continuous spatial domain \\(X \\subset \\mathbb{R}^2\\). Since the encoder can be any feature extractor from existing SISR methods, the essence of continuous image representations is the decoding function, which arbitrarily upsamples LR-HD features to HR-LD pixel values. A simple and intuitive paradigm is to learn a decoding function in HR-HD space that independently decodes the pixel value of each HR coordinate. This vanilla decoding function \\(f_{\\theta}\\) parameterized as an HD MLP \\(\\theta\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.783, + 0.469, + 0.802 + ], + "angle": 0, + "content": "\\[\nI \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta} \\left(z ^ {*}, x _ {i, j} ^ {*}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.81, + 0.469, + 0.841 + ], + "angle": 0, + "content": "where \\(x_{i,j}^{*}\\in X^{*}\\) refers to any nearest coordinate of \\(z^{*}\\), and \\(I(x_{i,j}^{*})\\) represents the predicted pixel value at \\(x_{i,j}^{*}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "However, as mentioned earlier in the Introduction section, this vanilla paradigm suffers from three aspects of computational redundancy. A computational optimal decoding function should allow computations proportional to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "the input image complexity rather than the output resolution. With this in mind, we propose to decouple the HR-HD decoding process into shared decoding in latent space and independent rendering in render space, along with a rendering algorithm adaptive to the input image content." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.175, + 0.75, + 0.19 + ], + "angle": 0, + "content": "3.2. Latent Modulated Function" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.365 + ], + "angle": 0, + "content": "To eliminate the computational redundancy in HR-HD decoding, we first propose a two-stage decoding process comprising a latent MLP \\(\\theta_{l}\\) and a render MLP \\(\\theta_{r}\\). The latent MLP handles the shared decoding process within the HR region of a latent code \\(z^{*}\\), while the render MLP independently predicts the pixel value of each HR coordinate. The LR-HD output \\(f_{\\theta_l}(z^*)\\) from the latent MLP is utilized to assist the render MLP in rendering. However, directly using this output \\(f_{\\theta_l}(z^*)\\) as the input of the render MLP is akin to using the latent code itself, as demonstrated in Eq. 1, resulting in no efficiency improvement yet:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.372, + 0.892, + 0.391 + ], + "angle": 0, + "content": "\\[\nI \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta_ {r}} \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right), x _ {i, j} ^ {*}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.893, + 0.699 + ], + "angle": 0, + "content": "Furthermore, to render arbitrary pixels accurately, it is necessary for the render MLP to operate in HR space. Therefore, the most viable approach to minimize rendering computations is to reduce the dimension or depth of the render MLP. However, an HD render MLP is still needed in Eq. 2 to fully receive and process the output \\( f_{\\theta_l}(z^*) \\), which contradicts our intention of utilizing an LD render MLP. To overcome these limitations, modulations become a suitable solution to connect the latent and render space. Modulations have been used in INR [13, 27, 33] and are often associated with the hyper-base network design. A hyper network converts an input image to modulations, which are then utilized by a small base network to map coordinates to a continuous image. However, these modulations are shared by the input image and work at the image level, resulting in unsatisfactory performance and challenging training. Hence, we introduce a latent modulation \\( m^{*} \\in \\mathbb{R}^{D_M} \\), which is generated by the latent MLP using the latent code \\( z^{*} \\). The latent modulation consists of the FiLM layers [29], where each layer consists of a scale modulation and a shift modulation:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.708, + 0.891, + 0.726 + ], + "angle": 0, + "content": "\\[\nm ^ {*} = \\left[ \\alpha_ {1} ^ {*}, \\beta_ {1} ^ {*}, \\dots , \\alpha_ {K} ^ {*}, \\beta_ {K} ^ {*} \\right] = S p l i t \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.732, + 0.891, + 0.776 + ], + "angle": 0, + "content": "where \\(\\alpha^{*}\\) represents the element-wise scale modulation, \\(\\beta^{*}\\) represents the element-wise shift modulation, and \\(K\\) represents the depth of hidden layers in the render MLP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.777, + 0.892, + 0.838 + ], + "angle": 0, + "content": "For each HR coordinate, the latent modulation \\(m^{*}\\) of the latent code \\(z^{*}\\) nearest to this coordinate is applied to the render MLP, and is applied to the hidden features \\(h^{*} \\in \\mathbb{R}^{D_{H}}\\) after each hidden linear layer:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.846, + 0.891, + 0.865 + ], + "angle": 0, + "content": "\\[\nh _ {k + 1} ^ {*} = f _ {\\theta_ {r} ^ {k + 1}} \\left(\\sigma \\left(\\left(1 + \\alpha_ {k} ^ {*}\\right) \\odot h _ {k} ^ {*} + \\beta_ {k} ^ {*}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\( h_k^* \\) represents the hidden features after the \\( k \\)-th hidden linear layer, and \\( \\sigma \\) represents the activation function." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26028" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.266, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.088, + 0.437, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Figure 3. Visualization \\((\\times 4\\) SR) of the normalized mean of shift modulation (left), and the normalized residual between the LM-LIIF predicted image and the bilinear upsampled image (right)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.263, + 0.473, + 0.565 + ], + "angle": 0, + "content": "Importantly, the proposed latent modulation effectively solves the incompatibility between LR-HD latent codes and HR-LD rendering and significantly disentangles the computations from parameters. For a latent modulation with 192 dimensions, we can accurately render images using a 7-layer render MLP with only 16 dimensions. This LD render MLP only learns the general rendering rules of natural images, while the latent modulation optimizes its rendering process based on the characteristics of each input latent code. As a result, a tiny render MLP can quickly adapt to any input latent code and accomplish efficient rendering at arbitrary resolution. A remaining concern is that the latent codes in existing INR-based ASSR methods often possess high dimensions, resulting in unnecessary computations when directly input into the render MLP. Therefore, we compress the HD latent code to an LD one \\( z_{c}^{*} \\) using the latent MLP, extending Eq. 3 to the first stage in Eq. 5. Building on the latent modulation, we propose a computational optimal paradigm of continuous image representation, called Latent Modulated Function (LMF):" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.571, + 0.469, + 0.595 + ], + "angle": 0, + "content": "Stage 1: \\([m^{*},z_{c}^{*}] = S p l i t(f_{\\theta_{l}}(z^{*}))\\) (5)" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.591, + 0.393, + 0.609 + ], + "angle": 0, + "content": "Stage 2: \\( I(x_{i,j}^{*}) = f_{\\theta_r,m^*}(z_c^*,x_{i,j}^*) \\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.81 + ], + "angle": 0, + "content": "LMF is a universal paradigm for efficient ASSR and is compatible with existing INR-based ASSR methods. By effortlessly relocating the time-consuming operations and large MLPs to latent space, LMF allows a minimal render MLP to perform the same level of ASSR as the original ASSR method. In addition, LMF effectively decouples the computational expensive techniques of feature unfolding and local ensemble [11]. Specifically, LMF only utilizes feature unfolding for the latent MLP, where nearby \\(3 \\times 3\\) feature vectors are concatenated as inputs. LMF performs local ensemble exclusively for the render MLP, which merges the predictions from the top-left, top-right, bottom-left, and bottom-right latent codes to achieve continuous predictions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.391, + 0.833 + ], + "angle": 0, + "content": "3.3. Controllable Multi-Scale Rendering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The HR regions of two latent codes often have different signal complexity, allowing for the use of different decoding functions. However, the challenge lies in obtaining the signal complexity of each latent code. The modulation scheme" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.852, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.227, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Figure 4. The positive correlation between the means of shift modulation and the minimum rendering scale factors in LM-LIIF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.446 + ], + "angle": 0, + "content": "in Eq. 4 suggests that the intensity of latent modulation is proportional to the signal complexity of its latent code. A lower intensity implies that the signal can be sufficiently rendered by the render MLP itself. In Figure 3, we display the normalized mean of shift modulation (left), and the normalized residual between the SR result of LMF-based LIIF and the bilinear upsampled result (right). It is evident that the means of modulation closely align with the signal complexity of the predicted image. Regions with higher means of modulation consistently correspond to more complex textures and edges. Hence, we employ the mean of shift modulation as the intensity of latent modulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.893, + 0.702 + ], + "angle": 0, + "content": "In LMF, we define the signal complexity of the latent code \\( z^{*} \\) as the minimum scale factor \\( s^{*} \\) needed to fully represent its signal. In simple terms, any rendering with a scale factor \\( s, s > s^{*} \\) will yield the same result as rendering with \\( s^{*} \\) followed by upsample interpolation (i.e. bilinear) with scale \\( \\frac{s}{s^{*}} \\). Therefore, we only need to render \\( z^{*} \\) with scale \\( \\min(s^{*}, s) \\) when upsampling an image with a scale \\( s \\). To obtain the minimum scale factor of each mean of shift modulation, in Figure 4, we sample 100 means of shift modulation uniformly from the range of [0, 1] and measure their minimum scale factors on the LR images in the training set. Figure 4 clearly shows the positive correlation between the means of shift modulation and the minimum scale factors. However, since they do not have a one-to-one correspondence, we create a Scale2Mod table by sampling common scale factors and recording their minimum and maximum means of shift modulation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Based on the Scale2Mods table, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to decode latent codes with their minimum scale factors. Firstly, we initialize \\( s_{-1} \\) as 1, \\( I_{s_{-1}} \\) as the LR image. For each sampled scale \\( s_i \\in [s_0, s_1, \\ldots, s] \\), we first upsample the last result \\( I_{s_{i-1}} \\) using bilinear interpolation with \\( \\frac{s_i}{s_{i-1}} \\) to obtain \\( I_{s_i} \\). We then query the Scale2Mods table to obtain the range of modulation means \\( [m_{min}^{s_i}, m_{max}^{s_i}] \\) for \\( s_i \\). For \\( s_i \\), the render MLP only needs to render the latent codes whose corresponding modulation means falling within this range. We repeat the above process iteratively until we obtain the final SR result or until all latent codes have been decoded. For precision controllable rendering, when creating" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26029" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.089, + 0.872, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison (PSNR (dB)/MACs) for ASSR on DIV2K validation set. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.115, + 0.868, + 0.265 + ], + "angle": 0, + "content": "
MethodParamsIn-scaleOut-of-scale
×2×3×4×6×12×18
Bicubic-31.01/-28.22/-26.66/-24.82/-22.27/-21.00/-
EDSR-b [25]-34.55/-30.90/-28.94/----
LIIF [11]355K34.67/3.567T30.96/3.567T29.00/3.567T26.75/3.567T23.71/3.567T22.17/3.567T
LM-LIIF170K34.65/125.32G30.95/65.37G28.99/44.38G26.75/25.30G23.72/6.61G22.19/2.95G
LTE [22]506K34.72/2.228T31.02/2.123T29.04/2.086T26.81/2.059T23.78/2.044T22.23/2.026T
LM-LTE278K34.71/198.81G30.99/100.75G29.03/66.43G26.78/34.96G23.74/9.07G22.21/4.10G
CiaoSR-1.4M [5]1.429M34.91/11.460T31.15/11.443T29.23/11.437T26.95/11.433T23.88/11.430T22.32/11.348T
LM-CiaoSR753K34.81/1.071T31.10/501.36G29.12/293.01G26.86/135.66G23.80/42.60G22.26/23.90G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.277, + 0.825, + 0.292 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison (PSNR (dB)) for in-scale SR on test sets. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.303, + 0.89, + 0.538 + ], + "angle": 0, + "content": "
MethodSet5Set14BSD100Urban100DIV2K100
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [47]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6134.9431.2229.19
LIIF [11]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6834.9931.2629.27
LM-LIIF38.1834.7232.5034.0630.5528.8332.3229.2827.7532.8228.8126.6634.9631.2529.27
LTE [22]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8135.0431.3229.33
LM-LTE38.2334.7632.5334.1130.5628.8632.3729.3127.7833.0328.9626.8035.0331.3129.32
CiaoSR-1.4M [5]38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1135.1531.4229.45
LM-CiaoSR38.2434.8132.5934.1130.6228.9032.3929.3227.8033.1329.1326.9535.0831.3629.37
SwinIR [24]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0735.2631.5029.48
LIIF [11]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1535.1731.4629.46
LM-LIIF38.3134.8832.7834.3130.7929.0232.4329.3727.8533.4629.4327.2135.2131.4829.49
LTE [22]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2435.2431.5029.51
LM-LTE38.3234.8832.7734.2830.7929.0132.4629.3927.8733.5229.4427.2435.2431.5029.50
CiaoSR-3.1M [5]38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4235.2931.5529.59
LM-CiaoSR38.3734.9432.8734.3330.8329.0832.4829.4327.9033.6529.5827.3935.3031.5529.56
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.47, + 0.643 + ], + "angle": 0, + "content": "the Scale2Mods table, we set an MSE threshold \\( t \\) and define the signal complexity of a latent code as the minimum scale factor needed to render its signal within this error \\( t \\). By effortlessly adjusting the MSE threshold during testing, CMSR is capable of rendering with varying precision and speed, while adapting naturally to the image content." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.656, + 0.21, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.274, + 0.697 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Dataset. Following the setups in previous ASSR methods [5, 11, 22], we employ the DIV2K dataset [1] as the training set. In each training iteration, we uniformly sample a downsampling scale \\(s\\) from a continuous range of \\([ \\times 1, \\times 4 ]\\). HR patches with spatial size of \\(48s \\times 48s\\) are cropped and bicubic downsampled to LR patches with spatial size of \\(48 \\times 48\\). To supervise the training process, we randomly sample \\(48 \\times 48\\) pixels from each HR patch. For evaluation, we present results on the DIV2K validation set (DIV2K 100) along with four widely-used benchmark test sets: Set5 [4], Set14 [44], BSD100 [26], and Urban100 [18]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We transform three previous ASSR methods, namely LIIF [11], LTE [22], and CiaoSR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[5], into our proposed LMF. LM-LIIF employs a 2-layer latent MLP with 208 dimensions, and a 7-layer render MLP with 16 dimensions. As for LM-LTE, we employ a 2-layer latent MLP with 256 dimensions, a 9-layer render MLP with 16 dimensions, and a local texture estimator [22] with 128 dimensions. LM-CiaoSR employs four 2-layer latent MLPs with 256 dimensions to generate key, value, and query information, followed by a 9-layer render MLP with 16 dimensions. The detailed architectures of LMF-based methods are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.893, + 0.884 + ], + "angle": 0, + "content": "We utilize EDSR-baseline (EDSR-b) [25], RDN [47], and SwinIR [24] without their upsampling modules as encoders. For both EDSR-b and RDN encoders, we conduct training for 1000 epochs with a batch size of 16. The learning rate is initialized as \\(1 \\times 10^{-4}\\) and decayed by a factor of 0.5 every 200 epochs. As for the SwinIR encoder, we conduct training for 1000 epochs with a batch size of 32. The learning rate is initialized as \\(2 \\times 10^{-4}\\) and decayed by a factor of 0.5 at [500, 800, 900, 950] epochs. For optimization, we utilize the L1 pixel loss as the loss function and employ the Adam [19] optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\). Our training details align with previous ASSR works [5, 22]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Settings. In line with previous methods [5, 11," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26030" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison (PSNR (dB)) for out-of-scale SR on test sets. The best results are highlighted in bold. The first 6 rows use RDN as the encoder, and the last 6 rows use SwinIR as the encoder." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.129, + 0.89, + 0.336 + ], + "angle": 0, + "content": "
MethodSet5Set14BSD100Urban100DIV2K100
×6×8×12×6×8×12×6×8×12×6×8×12×6×12×18
LIIF [11]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1526.9923.8922.34
LM-LIIF29.1827.2124.7126.6625.1723.3125.9824.9323.5824.1422.7721.1526.9923.9122.36
LTE [22]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.0423.9522.40
LM-LTE29.2127.2224.8226.6925.1523.3226.0124.9523.6124.2622.8621.2427.0323.9522.39
CiaoSR-1.4M [5]29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.1624.0622.48
LM-CiaoSR29.2427.2624.8026.7225.1823.3026.0324.8123.6024.4022.9821.3327.0723.9822.42
LIIF [11]29.4627.3624.9926.8225.3423.3926.0725.0123.6524.5923.1421.4327.1524.0222.43
LM-LIIF29.4627.3624.9826.8225.3723.4326.0825.0323.6524.6123.1521.4927.1824.0622.49
LTE [22]29.5027.3525.0726.8625.4223.4426.0925.0323.6624.6223.1721.5027.2024.0922.50
LM-LTE29.5227.3925.0126.8325.3823.4126.1025.0323.6624.6323.1721.5027.1924.0822.50
CiaoSR-3.1M [5]29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6027.2824.1522.54
LM-CiaoSR29.5227.4225.0226.8425.4123.4526.1225.0623.6824.7423.2621.5827.2424.1222.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.171, + 0.348, + 0.798, + 0.364 + ], + "angle": 0, + "content": "Table 4. Efficiency comparison (MACs/GPU runtime) for ASSR. The best results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.374, + 0.89, + 0.543 + ], + "angle": 0, + "content": "
DecoderParamsGPU mem.180p270p360p
720p1080p2k720p1080p2k720p1080p2k
LIIF [11]355K7.7GB1.160T/1.1148s2.609T/2.5230s4.638T/4.5232s1.160T/1.1195s2.609T/2.5347s4.638T/4.5544s1.160T/1.1272s2.609T/2.5456s4.638T/4.5589s
LM-LIIF170K5.3GB14.43G/0.0246s18.72G/0.0474s24.72G/0.0795s25.39G/0.0363s32.46G/0.0593s37.50G/0.0919s40.74G/0.0541s47.81G/0.0770s57.71G/0.1107s
LTE [22]506K8.4GB694.34G/0.9135s1.506T/2.0366s2.666T/3.7578s714.11G/0.9191s1.526T/2.0497s2.685T/3.7694s741.80G/0.9258s1.553T/2.0637s2.712T/3.7791s
LM-LTE278K4.7GB21.60G/0.0382s25.87G/0.0750s33.21G/0.1254s39.53G/0.0545s48.59G/0.0905s52.88G/0.1414s64.63G/0.0774s73.69G/0.1123s86.38G/0.1640s
CiaoSR-1.4M [5]1.429M8.8GB3.718T/1.8820s8.362T/4.1718s14.864T/7.3619s3.721T/1.9838s8.365T/4.2999s14.867T/7.5611s3.725T/2.1355s8.370T/4.5064s14.871T/7.8366s
LM-CiaoSR753K8.9GB95.23G/0.2637s99.57G/0.3849s106.68G/0.5546s204.09G/0.5263s214.32G/0.6476s217.83G/0.8160s356.46G/0.8844s366.69G/1.0013s381.02G/1.1778s
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.676 + ], + "angle": 0, + "content": "22], we employ PSNR as the objective criterion to assess quality performance. To evaluate efficiency, we measure the computational cost using Multiply-Accumulate Operations (MACs) and quantify the model speed using the runtime on an RTX 2080 Ti GPU. Additionally, we provide the number of parameters and GPU memory usage. Notably, the MACs, runtime, GPU memory, and parameters are evaluated solely on the decoder unless explicitly stated." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.26, + 0.71 + ], + "angle": 0, + "content": "4.2. ASSR Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. In Table 1, we compare our LMF-based methods with reference ASSR methods, namely LIIF [11], LTE [22], and CiaoSR [5], on the DIV2K validation set using EDSR-b as the encoder. We include the results of bicubic interpolation and EDSR-b with its initial upsampling module for reference. The results demonstrate that LMF-based methods achieve a significant reduction in computational cost while delivering competitive PSNR performance compared to the original methods. Compared to LIIF, LM-LIIF achieves a reduction of \\(96.6\\%\\) to \\(99.9\\%\\) in computational cost across all scales. Similarly, LM-LTE and LM-CiaoSR provide \\(91.3\\%\\) to \\(99.8\\%\\) and \\(90.4\\%\\) to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.631 + ], + "angle": 0, + "content": "\\(99.8\\%\\) reduction in computational cost compared to LTE and CiaoSR, respectively. However, LM-CiaoSR exhibits slightly lower PSNR than CiaoSR due to relocating the expensive attention-in-attention modules from HR-HD space to latent space for the minimal computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.893, + 0.834 + ], + "angle": 0, + "content": "In Table 2 and 3, we present the comprehensive PSNR evaluation for ASSR using RDN and SwinIR as the encoders. Notably, CiaoSR utilizes a 1.429M decoder for EDSR-b and RDN, and a 3.065M decoder for SwinIR, while our LM-CiaoSR employs the same decoder for all encoders. The results demonstrate the competitive PSNR performance of our LMF-based methods, especially LM-LIIF and LM-LTE, compared to the reference methods on benchmark test sets. Notably, LMF-based methods exhibit superior PSNR performance gains on the SwinIR encoder. The results presented in Table 1, 2, and 3 collectively validate the universal applicability of our LMF across different encoders, INR-based ASSR methods, test sets, and scales." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In Table 4, we present comparisons of MACs, runtime, GPU memory, and parameters for ASSR across multiple input and output resolutions. We utilize bicubic downsampled images from the DIV2K validation set as inputs for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26031" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.882, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.308, + 0.774, + 0.322 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison for ASSR. All of the ASSR methods use SwinIR as the encoder." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.337, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Table 5. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on latent modulations and latent codes, using EDSR-b as the encoder." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.362, + 0.889, + 0.497 + ], + "angle": 0, + "content": "
DecoderLatent codeLatent modulationLatent MLPRender MLPParamsIn-scaleOut-of-scale
×2×4×8×16
LIIF576××5 × 256355K34.67/3.567T29.00/3.567T25.40/3.567T22.61/3.567T
LIIF (tiny)576××9 × 1612K33.61/117.44G28.18/117.44G24.78/117.44G22.16/117.44G
LIIF (C2F)192×2 × 1927 × 16202K34.11/221.54G28.63/114.62G25.11/73.04G22.38/58.91G
LM-LIIF64Scale2 × 1927 × 16136K34.61/110.94G28.96/46.87G25.38/30.79G22.61/26.58G
LM-LIIF64Shift2 × 1927 × 16136K34.59/110.94G28.94/46.87G25.36/30.79G22.59/26.58G
LM-LIIF64Scale & shift2 × 1927 × 16155K34.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
LM-LIIF16Scale & shift2 × 2087 × 16170K34.65/125.32G28.99/44.38G25.40/24.10G22.62/18.90G
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.658 + ], + "angle": 0, + "content": "each resolution, and all results exclude the encoders. As shown in the table, our LMF-based methods exhibit a significant reduction in computational cost, ranging from \\(90.7\\%\\) to \\(99.5\\%\\), and a substantial reduction in runtime, ranging from \\(58.6\\%\\) to \\(98.2\\%\\), compared to the reference methods. Compared to the reference methods, LMF-based methods also exhibit a reduction in parameters, ranging from \\(45.1\\%\\) to \\(52.1\\%\\), which is attributed to the decoupling of computations from parameters, demonstrating the significance of our elegant design over merely stacking parameters." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.827 + ], + "angle": 0, + "content": "Qualitative Results. In Figure 5, we present a qualitative comparison for ASSR, using SwinIR as the encoder. The visual results demonstrate that our LMF-based methods achieve competitive visual quality across all scales compared to the reference methods. Remarkably, LMF-based methods achieve continuous and precise results while significantly reducing computational cost and runtime. This confirms that LMF effectively addresses the challenge of upsampling LR-HD features to HR-LD pixel values, eliminating the need for an HR-HD decoding function. More visual results are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.844, + 0.229, + 0.862 + ], + "angle": 0, + "content": "5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.9 + ], + "angle": 0, + "content": "In this section, we investigate the effect of each component in LMF. Both PSNR and MACs results on the DIV2K vali" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.892, + 0.553 + ], + "angle": 0, + "content": "dation set are provided for a comprehensive evaluation, and all of the models use EDSR-b as the encoder. Notably, the CMSR algorithm is not used in Table 5 and Table 6." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.893, + 0.809 + ], + "angle": 0, + "content": "Latent Modulation. In Table 5, we evaluate the impact of latent modulation. Firstly, the comparison between LIIF and LIIF (tiny) reveals a significant decline in PSNR performance when shrinking the decoder from a 5-layer MLP with 256 dimensions to a 9-layer MLP with 16 dimensions. This performance drop is expected due to the substantial reduction in network parameters. However, our LM-LIIF model, which utilizes a 7-layer render MLP with 16 dimensions, outperforms LIIF in terms of PSNR, while consuming computational cost even lower than LIIF (tiny). Furthermore, we introduce a coarse-to-fine variant of LIIF, where both stages employ the same structure as LM-LIIF. Unlike LM-LIIF, LIIF (C2F) fails to achieve competitive PSNR performance compared to LIIF, highlighting the critical role of latent modulation in reducing the dimensions of the render MLP. These results also validate that HR-HD computations are redundant for decoding LR-HD features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "In Table 5, we also compare the performance of different types of latent modulations, namely the scale modulation, the shift modulation, and the combination of scale and shift modulations. The results demonstrate that LM-LIIF achieves the highest PSNR performance when both scale and shift modulations are employed, and the computational" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26032" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.131 + ], + "angle": 0, + "content": "Table 6. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on the latent MLP and the render MLP, using EDSR-b as the encoder, and LM-LIIF as the decoder." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.143, + 0.469, + 0.344 + ], + "angle": 0, + "content": "
Latent MLPRender MLPIn-scaleOut-of-scale
×2×4×8×16
2 × 2562 × 12834.66/236.91G28.97/131.21G25.37/104.57G22.60/97.18G
2 × 2563 × 6434.67/232.18G29.00/126.48G25.41/99.85G22.63/92.50G
2 × 2565 × 3234.68/197.37G29.00/91.67G25.41/65.11G22.63/58.04G
4 × 2569 × 1634.64/258.37G28.98/87.79G25.40/45.04G22.62/34.13G
2 × 2569 × 1634.66/171.85G28.99/66.16G25.40/39.65G22.62/32.79G
2 × 1927 × 1634.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
2 × 1285 × 1634.63/79.76G28.98/35.02G25.39/23.79G22.62/20.83G
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.477 + ], + "angle": 0, + "content": "cost remains nearly identical. Finally, we compare using the latent code (64 dimensions) from the encoder and the compressed latent code (16 dimensions) from the latent MLP as the input of the render MLP. The results indicate that using the compressed latent code can further reduce the computational cost for large scales. Based on these findings, we incorporate both scale and shift modulations as well as the compressed latent code in our final LMF models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Network Structure. We evaluated different depths and dimensions of the latent MLP and the render MLP in Table 6. Notably, the compressed latent code is not used to ensure better alignment of depths and dimensions between the two MLPs. For the latent MLP, we observe that decreasing the depth of the latent MLP from 4 to 2 yields nearly the same PSNR performance, which suggests that the latent MLP does not require a deep architecture or a large number of parameters. However, when reducing the dimension of the latent MLP from 192 to 128, the PSNR performance begins to decrease. This observation confirms that HD computations are crucial for a powerful latent MLP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.871 + ], + "angle": 0, + "content": "In Table 6, we investigate the depths and dimensions of the render MLP by fixing the dimensions of the latent MLP and the latent modulation at 256. The results demonstrate that we can effortlessly reduce the dimension of the render MLP from 128 to 16, resulting in an impressive up to \\(66.26\\%\\) reduction in computational cost, without compromising performance. When the latent modulation possesses a fixed dimension, it is consistently viable to utilize a lower dimensional rendering process, thereby minimizing the computational burden. All things considered, our final LM-LIIF model employs a 2-layer latent MLP with 208 dimensions (16 dimensions for the compressed latent code) and a 7-layer render MLP with 16 dimensions, optimizing computational efficiency without sacrificing performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Controllable Multi-Scale Rendering. In Table 7, we evaluate the MSE thresholds when creating the Scale2Mod ta" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 7. Ablation study (PSNR(dB)/MACs of rendering/MACs of decoder on DIV2K validation set) on MSE thresholds of CMSR, using EDSR-b as the encoder, and LM-LIIF as the decoder." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.143, + 0.89, + 0.258 + ], + "angle": 0, + "content": "
MSE thresholdIn-scaleOut-of-scale
×2×4×8×16
×34.65/17.41G/125.32G28.99/17.41G/44.38G25.40/17.37G/24.10G22.62/17.23G/18.90G
2 × 10-534.65/16.30G/124.22G28.99/13.67G/40.62G25.40/12.41G/19.14G22.62/3.48G/5.15G
1 × 10-434.62/13.63G/121.55G28.98/11.71G/38.69G25.39/8.35G/15.08G22.62/2.50G/4.17G
5 × 10-434.42/10.35G/118.27G28.93/8.94G/35.92G25.38/6.08G/12.81G22.61/2.29G/3.96G
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.516 + ], + "angle": 0, + "content": "ble for CMSR. Compared to the case without CMSR, a Scale2Mod table created with an MSE threshold of \\( 2 \\times 10^{-5} \\) achieves the same PSNR performance while reducing computational cost by \\( 20.58\\% \\) and \\( 72.75\\% \\) for scale \\( \\times 8 \\) and \\( \\times 16 \\). With lower precision requirements, i.e., MSE thresholds of \\( 1 \\times 10^{-4} \\) and \\( 5 \\times 10^{-4} \\), CMSR effectively reduces the computational cost even further while maintaining maximum PSNR performance. This adaptability is advantageous for real-world applications as the precision can be adjusted during testing without the need for model retraining. However, CMSR exclusively reduces the computational cost of the render MLP, which limits the computational cost reduction for small scales. Hence, in the Experiments section, we only employ CMSR with the MSE threshold \\( 2 \\times 10^{-5} \\) for out-of-distribution scales. More details and results regarding CMSR are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.537, + 0.619, + 0.554 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.791 + ], + "angle": 0, + "content": "In this paper, we propose a Latent Modulated Function (LMF), which realizes a computational optimal paradigm of continuous image representation. LMF utilizes an MLP in latent space to generate the latent modulation of each latent code. This enables a modulated MLP in render space to perform effective rendering at any resolution. Moreover, based on the correlation between modulation intensity and signal complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to balance the rendering precision and efficiency during testing. Experiments demonstrate that converting INR-based ASSR methods to LMF leads to substantial reductions in computational cost, runtime, and parameters, while preserving performance. Better forms of latent modulations and more efficient decoder architectures will be explored in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.813, + 0.681, + 0.83 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 62071500; Supported by Shenzhen Science and Technology Program under Grant No. JCYJ20230807111107015." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26033" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. British machine vision conference, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.47, + 0.485 + ], + "angle": 0, + "content": "[5] Jiezhang Cao, Qin Wang, Yongqin Xian, Yawei Li, Bingbing Ni, Zhiming Pi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Ciaosr: Continuous implicit attention-inattention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[6] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. Nerv: Neural representations for videos. Advances in Neural Information Processing Systems, 34:21557-21568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.614 + ], + "angle": 0, + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[8] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.47, + 0.771 + ], + "angle": 0, + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[10] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, pages 170-187. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[11] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 3, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[12] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184-199, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[13] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[14] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.894, + 0.329 + ], + "angle": 0, + "content": "[15] Zongyao He and Zhi Jin. Dynamic implicit image function for efficient arbitrary-scale image representation. arXiv preprint arXiv:2306.12321, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[16] Zongyao He, Zhi Jin, and Yao Zhao. Srdrl: A blind superresolution framework with degradation reconstruction loss. IEEE Transactions on Multimedia, 24:2877-2889, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[17] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1575-1584, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.894, + 0.762 + ], + "angle": 0, + "content": "[22] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26034" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[26] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 416-423, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[27] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated periodic activations for generalizable local functional representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.328, + 0.471, + 0.397 + ], + "angle": 0, + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.399, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[29] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[31] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.706 + ], + "angle": 0, + "content": "[33] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[34] Gaochao Song, Qian Sun, Luo Zhang, Ran Su, Jianfeng Shi, and Ying He. Ope-sr: Orthogonal position encoding for designing a parameter-free upsampling module in arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10009-10020, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[35] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74–91. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[36] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4539-4547, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4801-4810, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[38] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Hang Wang, Zhengyan Tong, and Yutian Liu. Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1786-1795, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.289 + ], + "angle": 0, + "content": "[39] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.893, + 0.346 + ], + "angle": 0, + "content": "[40] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.893, + 0.416 + ], + "angle": 0, + "content": "[41] Jie-En Yao, Li-Yuan Tsao, Yi-Chen Lo, Roy Tseng, Chia-Che Chang, and Chun-Yi Lee. Local implicit normalizing flow for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1776–1785, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.417, + 0.893, + 0.473 + ], + "angle": 0, + "content": "[42] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.893, + 0.53 + ], + "angle": 0, + "content": "[43] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.53, + 0.893, + 0.599 + ], + "angle": 0, + "content": "[44] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.893, + 0.656 + ], + "angle": 0, + "content": "[45] Huanrong Zhang, Jie Xiao, and Zhi Jin. Multi-scale image super-resolution via a single extendable deep network. IEEE Journal of Selected Topics in Signal Processing, 15(2):253-263, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.657, + 0.893, + 0.713 + ], + "angle": 0, + "content": "[46] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.714, + 0.893, + 0.783 + ], + "angle": 0, + "content": "[47] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26035" + } + ] +] \ No newline at end of file diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_origin.pdf b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16853cec4d7768415ee7102850a839a558e33307 --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/637c4edb-98b1-40b4-8db4-8d4aa7226fc9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4140d2fb213ba198a2762195381b2960a8ca310b22c198d7c1a398481294108e +size 1700977 diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/full.md b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28b98126bc0307c7af793b9f5b9f64d4987e46e8 --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/full.md @@ -0,0 +1,261 @@ +# Latent Modulated Function for Computational Optimal Continuous Image Representation + +Zongyao He1 Zhi Jin1,2,* + +$^{1}$ School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University $^{2}$ Guangdong Provincial Key Laboratory of Fire Science and Technology + +hezy28@mail2.sysu.edu.cn, jinzh26@mail.sysu.edu.cn + +# Abstract + +The recent work Local Implicit Image Function (LIIF) and subsequent Implicit Neural Representation (INR) based works have achieved remarkable success in Arbitrary-Scale Super-Resolution (ASSR) by using MLP to decode Low-Resolution (LR) features. However, these continuous image representations typically implement decoding in High-Resolution (HR) High-Dimensional (HD) space, leading to a quadratic increase in computational cost and seriously hindering the practical applications of ASSR. To tackle this problem, we propose a novel Latent Modulated Function (LMF), which decouples the HR-HD decoding process into shared latent decoding in LR-HD space and independent rendering in HR Low-Dimensional (LD) space, thereby realizing the first computational optimal paradigm of continuous image representation. Specifically, LMF utilizes an HD MLP in latent space to generate latent modulations of each LR feature vector. This enables a modulated LD MLP in render space to quickly adapt to any input feature vector and perform rendering at arbitrary resolution. Furthermore, we leverage the positive correlation between modulation intensity and input image complexity to design a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to adjust the decoding efficiency based on the rendering precision. Extensive experiments demonstrate that converting existing INR-based ASSR methods to LMF can reduce the computational cost by up to $99.9\%$ , accelerating inference by up to $57\times$ , and save up to $76\%$ of parameters, while maintaining competitive performance. The code is available at https://github.com/HeZongyao/LMF. + +# 1. Introduction + +Image data has long been stored and presented as discrete 2D arrays with fixed resolutions, while the visual world we perceive is continuous and precise. As a result, there is a + +![](images/d93c6cf7827c50f011fcc04ba860ebfc839404c1ed92ff3fe1706222a1c5cfc8.jpg) +Figure 1. Efficiency comparisons $(320 \times 180$ input) for ASSR. LMF-based ASSR methods significantly reduce computational cost (MACs), runtime (circle sizes), and parameters (colors). + +growing demand for arbitrary scale image resizing in various real-world scenarios. For instance, users often require continuous resizing of images during capture, viewing, and editing. Recent developments in Single Image SuperResolution (SISR) [7, 9, 12, 16, 20, 21, 23-25, 36, 45-47] have showcased significant advancements in reconstructing High-Resolution (HR) images from Low-Resolution (LR) counterparts. However, traditional interpolation methods are still the preferred choice for handling real-world image resizing. This is because SISR methods typically learn upsampling functions for a particular scale factor, requiring training different models for different scales. In contrast, Arbitrary-Scale Super-Resolution (ASSR) utilizes a single model to handle arbitrary scales, thus presenting significant potential in practical image-resizing applications. + +Building on the success of Implicit Neural Representation (INR), Local Implicit Image Function (LIIF) [11] is a recent ASSR approach that learns to continuously represent 2D images through an implicit function. LIIF encodes an image as latent codes and employs a decoding MLP to predict pixel values based on coordinates and their neighboring latent codes. The use of continuous coordinates enables LIIF and subsequent INR-based ASSR works + +![](images/76fb012507164f8c3e888194a2cf40616913ae4e357db06ad6b86223a62fc94b.jpg) +Figure 2. The framework of our LMF-based continuous image representation. Given a latent code generated by the encoder, an HD latent MLP generates the latent modulation to adjust the features of an LD render MLP, thereby achieving efficient arbitrary-resolution rendering. + +[5, 8, 14, 15, 22, 34, 37-41] to represent images at any resolution, achieving robust performance in up to $\times 30$ SR task. However, existing INR-based ASSR methods adhere to an intuitive ASSR paradigm that uses a complex decoding function in HR High-Dimensional (HD) space to independently predict the pixel value of each HR coordinate. As a result, their computational cost and runtime increase quadratically with the scale factor, contradicting the goal of real-world image-resizing applications (refer to Figure 1). Upon closer examination, the significant computational redundancy of HR-HD decoding can be divided into three aspects. Firstly, there is inherent similarity within the HR region corresponding to an LR latent code, suggesting the need for merging shared computations within this region before independently predicting each HR pixel value. Secondly, this paradigm ignores the varying complexity of different LR latent codes and applies the same computational process to predict all the HR pixel values. Lastly, this paradigm couples the HD MLP parameters and HR computations, while such a large number of parameters are not always needed for each prediction. + +To tackle this problem, we propose a novel Latent Modulated Function (LMF) to realize a computational optimal paradigm of continuous image representation. By leveraging latent modulation, LMF successfully disentangles the decoding process from HR-HD space into latent (LR-HD) and render (HR-LD) spaces. As shown in Figure 2, LMF first utilizes a latent MLP to generate the latent modulation of a given latent code. When querying the coordinates near the latent code, the latent modulation is applied to adjust each hidden linear layer in the render MLP. This modulation enables LMF to minimize the dimension of the render MLP without sacrificing performance. Furthermore, we observe a positive correlation between the latent modulation intensity and the input image complexity. Building on this insight, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm. CMSR not only allows the computational cost of rendering to be proportional to the input image complexity rather than the output resolution, but also provides the flexibility to balance precision and efficiency during testing. Computational optimal and fast ASSR holds great promise for future research and downstream appli + +cations. Experiments demonstrate that transforming existing INR-based ASSR methods into LMF reduces computational costs by $90.4\%$ to $99.9\%$ , accelerates inference by $2.4 \times$ to $56.9 \times$ , and saves $45.1\%$ to $76.0\%$ of parameters, while maintaining competitive PSNR performance. + +Our main contributions are summarized as follows: + +- We propose a Latent Modulated Function (LMF) that realizes computational optimal continuous image representation, which decouples the HR-HD decoding into decoding in latent (LR-HD) and render (HR-LD) spaces. +- Based on the correlation between modulation intensity and input image complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to effortlessly balance the rendering efficiency and precision at test time. +- Extensive experiments demonstrate that seamlessly converting existing INR-based ASSR methods to LMF can achieve substantial reductions in computational cost, runtime, and parameters, while preserving performance. + +# 2. Related Work + +# 2.1. Implicit Neural Representation + +INR has demonstrated remarkable success for continuous signal representation across various tasks and modalities, including signal reconstruction [10, 27, 31, 32], image and video compression [6, 13, 35], and image generation [33, 43]. The highlight of INR research is NeRF [28] and subsequent works [2, 3, 30, 42], which learn MLPs to infer the RGB and density of any given coordinates and view directions in a 3D scene. However, a major challenge faced by INR is the extensive training process required to fit a signal from scratch. To address this, MINER [31] and KiloNeRF [30] propose the use of multiple small MLPs to independently learn different segments of a signal. Another solution is hyper-base network design [10, 13, 27, 33], where a pre-trained hyper network generates INR for any input signal during testing. To represent complex signals, hyper networks often generate modulations for base networks. + +# 2.2. Single-Image Super-Resolution + +SISR as a classic low-level vision task has been studied for decades. In recent years, this field witnessed a variety of + +CNN-based networks, such as SRCNN [12], LapSRN [20], SRResNet [21], EDSR [25], MemNet [36], RCAN [46], RDN [47], and SRFBN [23]. Lately, image processing transformers, such as IPT [7], SwinIR [24], and HAT [9], have been proposed, which surpass CNN-based methods in SR performance using a large dataset. Although SISR methods have achieved remarkable performance, most of them are designed for a specific scale. + +ASSR methods are designed to utilize a single network for all scales, showcasing significant advancements in practicality and convenience over previous works. Hu et al. proposed MetaSR [17] to generate an upsampling layer for ASSR using a meta-network. However, MetaSR exhibits poor generalization for large-scale SR. Building upon the success of INR, a novel continuous image representation was proposed in LIIF [11], where each pixel value is inferred from an MLP using its coordinate and neighboring 2D features. Subsequently, a variety of INR-based ASSR methods have been proposed [5, 8, 14, 15, 22, 34, 37-41]. Notably, Lee et al. proposed LTE [22], which introduces a dominant-frequency estimator to continuously capture fine details. Cao et al. introduced CiaoSR [5], a continuous implicit attention-in-attention network that learns ensemble weights for nearby local features, achieving State-Of-The-Art (SOTA) performance. Despite the robust performance of INR-based ASSR methods for scales up to $\times 30$ , they adhere to the paradigm of decoding in HR-HD space, leading to prohibitively high computational cost and runtime. + +# 3. Methods + +# 3.1. Preliminary and Motivation + +In continuous image representations, an image $I_{in}$ is represented as 2D feature maps $F$ in latent space $\mathbb{R}^{H_{in} \times W_{in} \times D_F}$ . A latent code $z^* \in \mathbb{R}^{D_F}$ represents a feature vector derived from the feature maps, which is used to decode its nearest coordinates in the continuous spatial domain $X \subset \mathbb{R}^2$ . Since the encoder can be any feature extractor from existing SISR methods, the essence of continuous image representations is the decoding function, which arbitrarily upsamples LR-HD features to HR-LD pixel values. A simple and intuitive paradigm is to learn a decoding function in HR-HD space that independently decodes the pixel value of each HR coordinate. This vanilla decoding function $f_{\theta}$ parameterized as an HD MLP $\theta$ can be formulated as: + +$$ +I \left(x _ {i, j} ^ {*}\right) = f _ {\theta} \left(z ^ {*}, x _ {i, j} ^ {*}\right), \tag {1} +$$ + +where $x_{i,j}^{*}\in X^{*}$ refers to any nearest coordinate of $z^{*}$ , and $I(x_{i,j}^{*})$ represents the predicted pixel value at $x_{i,j}^{*}$ . + +However, as mentioned earlier in the Introduction section, this vanilla paradigm suffers from three aspects of computational redundancy. A computational optimal decoding function should allow computations proportional to + +the input image complexity rather than the output resolution. With this in mind, we propose to decouple the HR-HD decoding process into shared decoding in latent space and independent rendering in render space, along with a rendering algorithm adaptive to the input image content. + +# 3.2. Latent Modulated Function + +To eliminate the computational redundancy in HR-HD decoding, we first propose a two-stage decoding process comprising a latent MLP $\theta_{l}$ and a render MLP $\theta_{r}$ . The latent MLP handles the shared decoding process within the HR region of a latent code $z^{*}$ , while the render MLP independently predicts the pixel value of each HR coordinate. The LR-HD output $f_{\theta_l}(z^*)$ from the latent MLP is utilized to assist the render MLP in rendering. However, directly using this output $f_{\theta_l}(z^*)$ as the input of the render MLP is akin to using the latent code itself, as demonstrated in Eq. 1, resulting in no efficiency improvement yet: + +$$ +I \left(x _ {i, j} ^ {*}\right) = f _ {\theta_ {r}} \left(f _ {\theta_ {l}} \left(z ^ {*}\right), x _ {i, j} ^ {*}\right). \tag {2} +$$ + +Furthermore, to render arbitrary pixels accurately, it is necessary for the render MLP to operate in HR space. Therefore, the most viable approach to minimize rendering computations is to reduce the dimension or depth of the render MLP. However, an HD render MLP is still needed in Eq. 2 to fully receive and process the output $f_{\theta_l}(z^*)$ , which contradicts our intention of utilizing an LD render MLP. To overcome these limitations, modulations become a suitable solution to connect the latent and render space. Modulations have been used in INR [13, 27, 33] and are often associated with the hyper-base network design. A hyper network converts an input image to modulations, which are then utilized by a small base network to map coordinates to a continuous image. However, these modulations are shared by the input image and work at the image level, resulting in unsatisfactory performance and challenging training. Hence, we introduce a latent modulation $m^{*} \in \mathbb{R}^{D_M}$ , which is generated by the latent MLP using the latent code $z^{*}$ . The latent modulation consists of the FiLM layers [29], where each layer consists of a scale modulation and a shift modulation: + +$$ +m ^ {*} = \left[ \alpha_ {1} ^ {*}, \beta_ {1} ^ {*}, \dots , \alpha_ {K} ^ {*}, \beta_ {K} ^ {*} \right] = S p l i t \left(f _ {\theta_ {l}} \left(z ^ {*}\right)\right), \tag {3} +$$ + +where $\alpha^{*}$ represents the element-wise scale modulation, $\beta^{*}$ represents the element-wise shift modulation, and $K$ represents the depth of hidden layers in the render MLP. + +For each HR coordinate, the latent modulation $m^{*}$ of the latent code $z^{*}$ nearest to this coordinate is applied to the render MLP, and is applied to the hidden features $h^{*} \in \mathbb{R}^{D_{H}}$ after each hidden linear layer: + +$$ +h _ {k + 1} ^ {*} = f _ {\theta_ {r} ^ {k + 1}} \left(\sigma \left(\left(1 + \alpha_ {k} ^ {*}\right) \odot h _ {k} ^ {*} + \beta_ {k} ^ {*}\right)\right), \tag {4} +$$ + +where $h_k^*$ represents the hidden features after the $k$ -th hidden linear layer, and $\sigma$ represents the activation function. + +![](images/ecc20c03c6c14fe0ce269b16d540715265ee23591ea5c0665c4a77bda9afa34d.jpg) +Figure 3. Visualization $(\times 4$ SR) of the normalized mean of shift modulation (left), and the normalized residual between the LM-LIIF predicted image and the bilinear upsampled image (right). + +![](images/6ca0ef0fa8030fd66f06b6d03f4752b998a50135c9dc33bf6601303abc65fd04.jpg) + +Importantly, the proposed latent modulation effectively solves the incompatibility between LR-HD latent codes and HR-LD rendering and significantly disentangles the computations from parameters. For a latent modulation with 192 dimensions, we can accurately render images using a 7-layer render MLP with only 16 dimensions. This LD render MLP only learns the general rendering rules of natural images, while the latent modulation optimizes its rendering process based on the characteristics of each input latent code. As a result, a tiny render MLP can quickly adapt to any input latent code and accomplish efficient rendering at arbitrary resolution. A remaining concern is that the latent codes in existing INR-based ASSR methods often possess high dimensions, resulting in unnecessary computations when directly input into the render MLP. Therefore, we compress the HD latent code to an LD one $z_{c}^{*}$ using the latent MLP, extending Eq. 3 to the first stage in Eq. 5. Building on the latent modulation, we propose a computational optimal paradigm of continuous image representation, called Latent Modulated Function (LMF): + +Stage 1: $[m^{*},z_{c}^{*}] = S p l i t(f_{\theta_{l}}(z^{*}))$ (5) + +Stage 2: $I(x_{i,j}^{*}) = f_{\theta_r,m^*}(z_c^*,x_{i,j}^*)$ + +LMF is a universal paradigm for efficient ASSR and is compatible with existing INR-based ASSR methods. By effortlessly relocating the time-consuming operations and large MLPs to latent space, LMF allows a minimal render MLP to perform the same level of ASSR as the original ASSR method. In addition, LMF effectively decouples the computational expensive techniques of feature unfolding and local ensemble [11]. Specifically, LMF only utilizes feature unfolding for the latent MLP, where nearby $3 \times 3$ feature vectors are concatenated as inputs. LMF performs local ensemble exclusively for the render MLP, which merges the predictions from the top-left, top-right, bottom-left, and bottom-right latent codes to achieve continuous predictions. + +# 3.3. Controllable Multi-Scale Rendering + +The HR regions of two latent codes often have different signal complexity, allowing for the use of different decoding functions. However, the challenge lies in obtaining the signal complexity of each latent code. The modulation scheme + +![](images/da796c39b49cd2f67a7bb8191164a67b5b76657f5707106e84e1ab0cfbf5d767.jpg) +Figure 4. The positive correlation between the means of shift modulation and the minimum rendering scale factors in LM-LIIF. + +in Eq. 4 suggests that the intensity of latent modulation is proportional to the signal complexity of its latent code. A lower intensity implies that the signal can be sufficiently rendered by the render MLP itself. In Figure 3, we display the normalized mean of shift modulation (left), and the normalized residual between the SR result of LMF-based LIIF and the bilinear upsampled result (right). It is evident that the means of modulation closely align with the signal complexity of the predicted image. Regions with higher means of modulation consistently correspond to more complex textures and edges. Hence, we employ the mean of shift modulation as the intensity of latent modulation. + +In LMF, we define the signal complexity of the latent code $z^{*}$ as the minimum scale factor $s^{*}$ needed to fully represent its signal. In simple terms, any rendering with a scale factor $s, s > s^{*}$ will yield the same result as rendering with $s^{*}$ followed by upsample interpolation (i.e. bilinear) with scale $\frac{s}{s^{*}}$ . Therefore, we only need to render $z^{*}$ with scale $\min(s^{*}, s)$ when upsampling an image with a scale $s$ . To obtain the minimum scale factor of each mean of shift modulation, in Figure 4, we sample 100 means of shift modulation uniformly from the range of [0, 1] and measure their minimum scale factors on the LR images in the training set. Figure 4 clearly shows the positive correlation between the means of shift modulation and the minimum scale factors. However, since they do not have a one-to-one correspondence, we create a Scale2Mod table by sampling common scale factors and recording their minimum and maximum means of shift modulation. + +Based on the Scale2Mods table, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to decode latent codes with their minimum scale factors. Firstly, we initialize $s_{-1}$ as 1, $I_{s_{-1}}$ as the LR image. For each sampled scale $s_i \in [s_0, s_1, \ldots, s]$ , we first upsample the last result $I_{s_{i-1}}$ using bilinear interpolation with $\frac{s_i}{s_{i-1}}$ to obtain $I_{s_i}$ . We then query the Scale2Mods table to obtain the range of modulation means $[m_{min}^{s_i}, m_{max}^{s_i}]$ for $s_i$ . For $s_i$ , the render MLP only needs to render the latent codes whose corresponding modulation means falling within this range. We repeat the above process iteratively until we obtain the final SR result or until all latent codes have been decoded. For precision controllable rendering, when creating + +Table 1. Quantitative comparison (PSNR (dB)/MACs) for ASSR on DIV2K validation set. The best results are highlighted in bold. + +
MethodParamsIn-scaleOut-of-scale
×2×3×4×6×12×18
Bicubic-31.01/-28.22/-26.66/-24.82/-22.27/-21.00/-
EDSR-b [25]-34.55/-30.90/-28.94/----
LIIF [11]355K34.67/3.567T30.96/3.567T29.00/3.567T26.75/3.567T23.71/3.567T22.17/3.567T
LM-LIIF170K34.65/125.32G30.95/65.37G28.99/44.38G26.75/25.30G23.72/6.61G22.19/2.95G
LTE [22]506K34.72/2.228T31.02/2.123T29.04/2.086T26.81/2.059T23.78/2.044T22.23/2.026T
LM-LTE278K34.71/198.81G30.99/100.75G29.03/66.43G26.78/34.96G23.74/9.07G22.21/4.10G
CiaoSR-1.4M [5]1.429M34.91/11.460T31.15/11.443T29.23/11.437T26.95/11.433T23.88/11.430T22.32/11.348T
LM-CiaoSR753K34.81/1.071T31.10/501.36G29.12/293.01G26.86/135.66G23.80/42.60G22.26/23.90G
+ +Table 2. Quantitative comparison (PSNR (dB)) for in-scale SR on test sets. The best results are highlighted in bold. + +
MethodSet5Set14BSD100Urban100DIV2K100
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [47]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6134.9431.2229.19
LIIF [11]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6834.9931.2629.27
LM-LIIF38.1834.7232.5034.0630.5528.8332.3229.2827.7532.8228.8126.6634.9631.2529.27
LTE [22]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8135.0431.3229.33
LM-LTE38.2334.7632.5334.1130.5628.8632.3729.3127.7833.0328.9626.8035.0331.3129.32
CiaoSR-1.4M [5]38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1135.1531.4229.45
LM-CiaoSR38.2434.8132.5934.1130.6228.9032.3929.3227.8033.1329.1326.9535.0831.3629.37
SwinIR [24]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0735.2631.5029.48
LIIF [11]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1535.1731.4629.46
LM-LIIF38.3134.8832.7834.3130.7929.0232.4329.3727.8533.4629.4327.2135.2131.4829.49
LTE [22]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2435.2431.5029.51
LM-LTE38.3234.8832.7734.2830.7929.0132.4629.3927.8733.5229.4427.2435.2431.5029.50
CiaoSR-3.1M [5]38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4235.2931.5529.59
LM-CiaoSR38.3734.9432.8734.3330.8329.0832.4829.4327.9033.6529.5827.3935.3031.5529.56
+ +the Scale2Mods table, we set an MSE threshold $t$ and define the signal complexity of a latent code as the minimum scale factor needed to render its signal within this error $t$ . By effortlessly adjusting the MSE threshold during testing, CMSR is capable of rendering with varying precision and speed, while adapting naturally to the image content. + +# 4. Experiments + +# 4.1. Experimental Setups + +Dataset. Following the setups in previous ASSR methods [5, 11, 22], we employ the DIV2K dataset [1] as the training set. In each training iteration, we uniformly sample a downsampling scale $s$ from a continuous range of $[ \times 1, \times 4 ]$ . HR patches with spatial size of $48s \times 48s$ are cropped and bicubic downsampled to LR patches with spatial size of $48 \times 48$ . To supervise the training process, we randomly sample $48 \times 48$ pixels from each HR patch. For evaluation, we present results on the DIV2K validation set (DIV2K 100) along with four widely-used benchmark test sets: Set5 [4], Set14 [44], BSD100 [26], and Urban100 [18]. + +Implementation Details. We transform three previous ASSR methods, namely LIIF [11], LTE [22], and CiaoSR + +[5], into our proposed LMF. LM-LIIF employs a 2-layer latent MLP with 208 dimensions, and a 7-layer render MLP with 16 dimensions. As for LM-LTE, we employ a 2-layer latent MLP with 256 dimensions, a 9-layer render MLP with 16 dimensions, and a local texture estimator [22] with 128 dimensions. LM-CiaoSR employs four 2-layer latent MLPs with 256 dimensions to generate key, value, and query information, followed by a 9-layer render MLP with 16 dimensions. The detailed architectures of LMF-based methods are provided in the supplementary material. + +We utilize EDSR-baseline (EDSR-b) [25], RDN [47], and SwinIR [24] without their upsampling modules as encoders. For both EDSR-b and RDN encoders, we conduct training for 1000 epochs with a batch size of 16. The learning rate is initialized as $1 \times 10^{-4}$ and decayed by a factor of 0.5 every 200 epochs. As for the SwinIR encoder, we conduct training for 1000 epochs with a batch size of 32. The learning rate is initialized as $2 \times 10^{-4}$ and decayed by a factor of 0.5 at [500, 800, 900, 950] epochs. For optimization, we utilize the L1 pixel loss as the loss function and employ the Adam [19] optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ . Our training details align with previous ASSR works [5, 22]. + +Evaluation Settings. In line with previous methods [5, 11, + +Table 3. Quantitative comparison (PSNR (dB)) for out-of-scale SR on test sets. The best results are highlighted in bold. The first 6 rows use RDN as the encoder, and the last 6 rows use SwinIR as the encoder. + +
MethodSet5Set14BSD100Urban100DIV2K100
×6×8×12×6×8×12×6×8×12×6×8×12×6×12×18
LIIF [11]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1526.9923.8922.34
LM-LIIF29.1827.2124.7126.6625.1723.3125.9824.9323.5824.1422.7721.1526.9923.9122.36
LTE [22]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.0423.9522.40
LM-LTE29.2127.2224.8226.6925.1523.3226.0124.9523.6124.2622.8621.2427.0323.9522.39
CiaoSR-1.4M [5]29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.1624.0622.48
LM-CiaoSR29.2427.2624.8026.7225.1823.3026.0324.8123.6024.4022.9821.3327.0723.9822.42
LIIF [11]29.4627.3624.9926.8225.3423.3926.0725.0123.6524.5923.1421.4327.1524.0222.43
LM-LIIF29.4627.3624.9826.8225.3723.4326.0825.0323.6524.6123.1521.4927.1824.0622.49
LTE [22]29.5027.3525.0726.8625.4223.4426.0925.0323.6624.6223.1721.5027.2024.0922.50
LM-LTE29.5227.3925.0126.8325.3823.4126.1025.0323.6624.6323.1721.5027.1924.0822.50
CiaoSR-3.1M [5]29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6027.2824.1522.54
LM-CiaoSR29.5227.4225.0226.8425.4123.4526.1225.0623.6824.7423.2621.5827.2424.1222.53
+ +Table 4. Efficiency comparison (MACs/GPU runtime) for ASSR. The best results are highlighted in bold. + +
DecoderParamsGPU mem.180p270p360p
720p1080p2k720p1080p2k720p1080p2k
LIIF [11]355K7.7GB1.160T/1.1148s2.609T/2.5230s4.638T/4.5232s1.160T/1.1195s2.609T/2.5347s4.638T/4.5544s1.160T/1.1272s2.609T/2.5456s4.638T/4.5589s
LM-LIIF170K5.3GB14.43G/0.0246s18.72G/0.0474s24.72G/0.0795s25.39G/0.0363s32.46G/0.0593s37.50G/0.0919s40.74G/0.0541s47.81G/0.0770s57.71G/0.1107s
LTE [22]506K8.4GB694.34G/0.9135s1.506T/2.0366s2.666T/3.7578s714.11G/0.9191s1.526T/2.0497s2.685T/3.7694s741.80G/0.9258s1.553T/2.0637s2.712T/3.7791s
LM-LTE278K4.7GB21.60G/0.0382s25.87G/0.0750s33.21G/0.1254s39.53G/0.0545s48.59G/0.0905s52.88G/0.1414s64.63G/0.0774s73.69G/0.1123s86.38G/0.1640s
CiaoSR-1.4M [5]1.429M8.8GB3.718T/1.8820s8.362T/4.1718s14.864T/7.3619s3.721T/1.9838s8.365T/4.2999s14.867T/7.5611s3.725T/2.1355s8.370T/4.5064s14.871T/7.8366s
LM-CiaoSR753K8.9GB95.23G/0.2637s99.57G/0.3849s106.68G/0.5546s204.09G/0.5263s214.32G/0.6476s217.83G/0.8160s356.46G/0.8844s366.69G/1.0013s381.02G/1.1778s
+ +22], we employ PSNR as the objective criterion to assess quality performance. To evaluate efficiency, we measure the computational cost using Multiply-Accumulate Operations (MACs) and quantify the model speed using the runtime on an RTX 2080 Ti GPU. Additionally, we provide the number of parameters and GPU memory usage. Notably, the MACs, runtime, GPU memory, and parameters are evaluated solely on the decoder unless explicitly stated. + +# 4.2. ASSR Comparison + +Quantitative Results. In Table 1, we compare our LMF-based methods with reference ASSR methods, namely LIIF [11], LTE [22], and CiaoSR [5], on the DIV2K validation set using EDSR-b as the encoder. We include the results of bicubic interpolation and EDSR-b with its initial upsampling module for reference. The results demonstrate that LMF-based methods achieve a significant reduction in computational cost while delivering competitive PSNR performance compared to the original methods. Compared to LIIF, LM-LIIF achieves a reduction of $96.6\%$ to $99.9\%$ in computational cost across all scales. Similarly, LM-LTE and LM-CiaoSR provide $91.3\%$ to $99.8\%$ and $90.4\%$ to + +$99.8\%$ reduction in computational cost compared to LTE and CiaoSR, respectively. However, LM-CiaoSR exhibits slightly lower PSNR than CiaoSR due to relocating the expensive attention-in-attention modules from HR-HD space to latent space for the minimal computational cost. + +In Table 2 and 3, we present the comprehensive PSNR evaluation for ASSR using RDN and SwinIR as the encoders. Notably, CiaoSR utilizes a 1.429M decoder for EDSR-b and RDN, and a 3.065M decoder for SwinIR, while our LM-CiaoSR employs the same decoder for all encoders. The results demonstrate the competitive PSNR performance of our LMF-based methods, especially LM-LIIF and LM-LTE, compared to the reference methods on benchmark test sets. Notably, LMF-based methods exhibit superior PSNR performance gains on the SwinIR encoder. The results presented in Table 1, 2, and 3 collectively validate the universal applicability of our LMF across different encoders, INR-based ASSR methods, test sets, and scales. + +In Table 4, we present comparisons of MACs, runtime, GPU memory, and parameters for ASSR across multiple input and output resolutions. We utilize bicubic downsampled images from the DIV2K validation set as inputs for + +![](images/c95688909c8a3a98fc36451b03ecff56aa19ef690da168c9230e63e1b69b47b6.jpg) +Figure 5. Qualitative comparison for ASSR. All of the ASSR methods use SwinIR as the encoder. + +Table 5. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on latent modulations and latent codes, using EDSR-b as the encoder. + +
DecoderLatent codeLatent modulationLatent MLPRender MLPParamsIn-scaleOut-of-scale
×2×4×8×16
LIIF576××5 × 256355K34.67/3.567T29.00/3.567T25.40/3.567T22.61/3.567T
LIIF (tiny)576××9 × 1612K33.61/117.44G28.18/117.44G24.78/117.44G22.16/117.44G
LIIF (C2F)192×2 × 1927 × 16202K34.11/221.54G28.63/114.62G25.11/73.04G22.38/58.91G
LM-LIIF64Scale2 × 1927 × 16136K34.61/110.94G28.96/46.87G25.38/30.79G22.61/26.58G
LM-LIIF64Shift2 × 1927 × 16136K34.59/110.94G28.94/46.87G25.36/30.79G22.59/26.58G
LM-LIIF64Scale & shift2 × 1927 × 16155K34.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
LM-LIIF16Scale & shift2 × 2087 × 16170K34.65/125.32G28.99/44.38G25.40/24.10G22.62/18.90G
+ +each resolution, and all results exclude the encoders. As shown in the table, our LMF-based methods exhibit a significant reduction in computational cost, ranging from $90.7\%$ to $99.5\%$ , and a substantial reduction in runtime, ranging from $58.6\%$ to $98.2\%$ , compared to the reference methods. Compared to the reference methods, LMF-based methods also exhibit a reduction in parameters, ranging from $45.1\%$ to $52.1\%$ , which is attributed to the decoupling of computations from parameters, demonstrating the significance of our elegant design over merely stacking parameters. + +Qualitative Results. In Figure 5, we present a qualitative comparison for ASSR, using SwinIR as the encoder. The visual results demonstrate that our LMF-based methods achieve competitive visual quality across all scales compared to the reference methods. Remarkably, LMF-based methods achieve continuous and precise results while significantly reducing computational cost and runtime. This confirms that LMF effectively addresses the challenge of upsampling LR-HD features to HR-LD pixel values, eliminating the need for an HR-HD decoding function. More visual results are provided in the supplementary materials. + +# 5. Ablation Study + +In this section, we investigate the effect of each component in LMF. Both PSNR and MACs results on the DIV2K vali + +dation set are provided for a comprehensive evaluation, and all of the models use EDSR-b as the encoder. Notably, the CMSR algorithm is not used in Table 5 and Table 6. + +Latent Modulation. In Table 5, we evaluate the impact of latent modulation. Firstly, the comparison between LIIF and LIIF (tiny) reveals a significant decline in PSNR performance when shrinking the decoder from a 5-layer MLP with 256 dimensions to a 9-layer MLP with 16 dimensions. This performance drop is expected due to the substantial reduction in network parameters. However, our LM-LIIF model, which utilizes a 7-layer render MLP with 16 dimensions, outperforms LIIF in terms of PSNR, while consuming computational cost even lower than LIIF (tiny). Furthermore, we introduce a coarse-to-fine variant of LIIF, where both stages employ the same structure as LM-LIIF. Unlike LM-LIIF, LIIF (C2F) fails to achieve competitive PSNR performance compared to LIIF, highlighting the critical role of latent modulation in reducing the dimensions of the render MLP. These results also validate that HR-HD computations are redundant for decoding LR-HD features. + +In Table 5, we also compare the performance of different types of latent modulations, namely the scale modulation, the shift modulation, and the combination of scale and shift modulations. The results demonstrate that LM-LIIF achieves the highest PSNR performance when both scale and shift modulations are employed, and the computational + +Table 6. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on the latent MLP and the render MLP, using EDSR-b as the encoder, and LM-LIIF as the decoder. + +
Latent MLPRender MLPIn-scaleOut-of-scale
×2×4×8×16
2 × 2562 × 12834.66/236.91G28.97/131.21G25.37/104.57G22.60/97.18G
2 × 2563 × 6434.67/232.18G29.00/126.48G25.41/99.85G22.63/92.50G
2 × 2565 × 3234.68/197.37G29.00/91.67G25.41/65.11G22.63/58.04G
4 × 2569 × 1634.64/258.37G28.98/87.79G25.40/45.04G22.62/34.13G
2 × 2569 × 1634.66/171.85G28.99/66.16G25.40/39.65G22.62/32.79G
2 × 1927 × 1634.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
2 × 1285 × 1634.63/79.76G28.98/35.02G25.39/23.79G22.62/20.83G
+ +cost remains nearly identical. Finally, we compare using the latent code (64 dimensions) from the encoder and the compressed latent code (16 dimensions) from the latent MLP as the input of the render MLP. The results indicate that using the compressed latent code can further reduce the computational cost for large scales. Based on these findings, we incorporate both scale and shift modulations as well as the compressed latent code in our final LMF models. + +Network Structure. We evaluated different depths and dimensions of the latent MLP and the render MLP in Table 6. Notably, the compressed latent code is not used to ensure better alignment of depths and dimensions between the two MLPs. For the latent MLP, we observe that decreasing the depth of the latent MLP from 4 to 2 yields nearly the same PSNR performance, which suggests that the latent MLP does not require a deep architecture or a large number of parameters. However, when reducing the dimension of the latent MLP from 192 to 128, the PSNR performance begins to decrease. This observation confirms that HD computations are crucial for a powerful latent MLP. + +In Table 6, we investigate the depths and dimensions of the render MLP by fixing the dimensions of the latent MLP and the latent modulation at 256. The results demonstrate that we can effortlessly reduce the dimension of the render MLP from 128 to 16, resulting in an impressive up to $66.26\%$ reduction in computational cost, without compromising performance. When the latent modulation possesses a fixed dimension, it is consistently viable to utilize a lower dimensional rendering process, thereby minimizing the computational burden. All things considered, our final LM-LIIF model employs a 2-layer latent MLP with 208 dimensions (16 dimensions for the compressed latent code) and a 7-layer render MLP with 16 dimensions, optimizing computational efficiency without sacrificing performance. + +Controllable Multi-Scale Rendering. In Table 7, we evaluate the MSE thresholds when creating the Scale2Mod ta + +Table 7. Ablation study (PSNR(dB)/MACs of rendering/MACs of decoder on DIV2K validation set) on MSE thresholds of CMSR, using EDSR-b as the encoder, and LM-LIIF as the decoder. + +
MSE thresholdIn-scaleOut-of-scale
×2×4×8×16
×34.65/17.41G/125.32G28.99/17.41G/44.38G25.40/17.37G/24.10G22.62/17.23G/18.90G
2 × 10-534.65/16.30G/124.22G28.99/13.67G/40.62G25.40/12.41G/19.14G22.62/3.48G/5.15G
1 × 10-434.62/13.63G/121.55G28.98/11.71G/38.69G25.39/8.35G/15.08G22.62/2.50G/4.17G
5 × 10-434.42/10.35G/118.27G28.93/8.94G/35.92G25.38/6.08G/12.81G22.61/2.29G/3.96G
+ +ble for CMSR. Compared to the case without CMSR, a Scale2Mod table created with an MSE threshold of $2 \times 10^{-5}$ achieves the same PSNR performance while reducing computational cost by $20.58\%$ and $72.75\%$ for scale $\times 8$ and $\times 16$ . With lower precision requirements, i.e., MSE thresholds of $1 \times 10^{-4}$ and $5 \times 10^{-4}$ , CMSR effectively reduces the computational cost even further while maintaining maximum PSNR performance. This adaptability is advantageous for real-world applications as the precision can be adjusted during testing without the need for model retraining. However, CMSR exclusively reduces the computational cost of the render MLP, which limits the computational cost reduction for small scales. Hence, in the Experiments section, we only employ CMSR with the MSE threshold $2 \times 10^{-5}$ for out-of-distribution scales. More details and results regarding CMSR are provided in the supplementary material. + +# 6. Conclusion + +In this paper, we propose a Latent Modulated Function (LMF), which realizes a computational optimal paradigm of continuous image representation. LMF utilizes an MLP in latent space to generate the latent modulation of each latent code. This enables a modulated MLP in render space to perform effective rendering at any resolution. Moreover, based on the correlation between modulation intensity and signal complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to balance the rendering precision and efficiency during testing. Experiments demonstrate that converting INR-based ASSR methods to LMF leads to substantial reductions in computational cost, runtime, and parameters, while preserving performance. Better forms of latent modulations and more efficient decoder architectures will be explored in future work. + +# 7. Acknowledgement + +This work was supported by the National Natural Science Foundation of China under Grant No. 62071500; Supported by Shenzhen Science and Technology Program under Grant No. JCYJ20230807111107015. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2 +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. British machine vision conference, 2012. 5 +[5] Jiezhang Cao, Qin Wang, Yongqin Xian, Yawei Li, Bingbing Ni, Zhiming Pi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Ciaosr: Continuous implicit attention-inattention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 5, 6 +[6] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. Nerv: Neural representations for videos. Advances in Neural Information Processing Systems, 34:21557-21568, 2021. 2 +[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 3 +[8] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 2, 3 +[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 3 +[10] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, pages 170-187. Springer, 2022. 2 +[11] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 3, 4, 5, 6 + +[12] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184-199, 2014. 1, 3 +[13] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2, 3 +[14] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2, 3 +[15] Zongyao He and Zhi Jin. Dynamic implicit image function for efficient arbitrary-scale image representation. arXiv preprint arXiv:2306.12321, 2023. 2, 3 +[16] Zongyao He, Zhi Jin, and Yao Zhao. Srdrl: A blind superresolution framework with degradation reconstruction loss. IEEE Transactions on Multimedia, 24:2877-2889, 2021. 1 +[17] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1575-1584, 2019. 3 +[18] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5 +[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 1, 3 +[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 1, 3 +[22] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 2, 3, 5, 6 +[23] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 1, 3 +[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3, 5 + +[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 3, 5 +[26] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 416-423, 2001. 5 +[27] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated periodic activations for generalizable local functional representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, 2021. 2, 3 +[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[29] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3 +[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[31] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2 +[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2 +[33] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 3 +[34] Gaochao Song, Qian Sun, Luo Zhang, Ran Su, Jianfeng Shi, and Ying He. Ope-sr: Orthogonal position encoding for designing a parameter-free upsampling module in arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10009-10020, 2023. 2, 3 +[35] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74–91. Springer, 2022. 2 +[36] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4539-4547, 2017. 1, 3 + +[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4801-4810, 2021. 2, 3 +[38] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Hang Wang, Zhengyan Tong, and Yutian Liu. Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1786-1795, 2023. +[39] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. +[40] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021. +[41] Jie-En Yao, Li-Yuan Tsao, Yi-Chen Lo, Roy Tseng, Chia-Che Chang, and Chun-Yi Lee. Local implicit normalizing flow for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1776–1785, 2023. 2, 3 +[42] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2 +[43] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2 +[44] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5 +[45] Huanrong Zhang, Jie Xiao, and Zhi Jin. Multi-scale image super-resolution via a single extendable deep network. IEEE Journal of Selected Topics in Signal Processing, 15(2):253-263, 2020. 1 +[46] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 3 +[47] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 3, 5 \ No newline at end of file diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/images.zip b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..524a1d71966e43ecec9ab5cd9c60055acb772bb9 --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29a69f3c41ac74eafc26388b8c7388a5c940607b5b582c84fb93e44bdf8f09c8 +size 870759 diff --git a/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/layout.json b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b4b717b1bb3de49e0fcf627b4cab75d547742b9 --- /dev/null +++ b/2024/Latent Modulated Function for Computational Optimal Continuous Image Representation/layout.json @@ -0,0 +1,7688 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 471, + 140 + ], + "type": "text", + "content": "Latent Modulated Function for Computational Optimal Continuous Image Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 229, + 160, + 368, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 160, + 368, + 175 + ], + "spans": [ + { + "bbox": [ + 229, + 160, + 368, + 175 + ], + "type": "text", + "content": "Zongyao He1 Zhi Jin1,2,*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "spans": [ + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "type": "text", + "content": "School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University " + }, + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 175, + 512, + 204 + ], + "type": "text", + "content": "Guangdong Provincial Key Laboratory of Fire Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 158, + 205, + 432, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 205, + 432, + 217 + ], + "spans": [ + { + "bbox": [ + 158, + 205, + 432, + 217 + ], + "type": "text", + "content": "hezy28@mail2.sysu.edu.cn, jinzh26@mail.sysu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "text", + "content": "The recent work Local Implicit Image Function (LIIF) and subsequent Implicit Neural Representation (INR) based works have achieved remarkable success in Arbitrary-Scale Super-Resolution (ASSR) by using MLP to decode Low-Resolution (LR) features. However, these continuous image representations typically implement decoding in High-Resolution (HR) High-Dimensional (HD) space, leading to a quadratic increase in computational cost and seriously hindering the practical applications of ASSR. To tackle this problem, we propose a novel Latent Modulated Function (LMF), which decouples the HR-HD decoding process into shared latent decoding in LR-HD space and independent rendering in HR Low-Dimensional (LD) space, thereby realizing the first computational optimal paradigm of continuous image representation. Specifically, LMF utilizes an HD MLP in latent space to generate latent modulations of each LR feature vector. This enables a modulated LD MLP in render space to quickly adapt to any input feature vector and perform rendering at arbitrary resolution. Furthermore, we leverage the positive correlation between modulation intensity and input image complexity to design a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to adjust the decoding efficiency based on the rendering precision. Extensive experiments demonstrate that converting existing INR-based ASSR methods to LMF can reduce the computational cost by up to " + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "text", + "content": ", accelerating inference by up to " + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "inline_equation", + "content": "57\\times" + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "text", + "content": ", and save up to " + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 47, + 270, + 290, + 617 + ], + "type": "text", + "content": " of parameters, while maintaining competitive performance. The code is available at https://github.com/HeZongyao/LMF." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 659, + 288, + 696 + ], + "type": "text", + "content": "Image data has long been stored and presented as discrete 2D arrays with fixed resolutions, while the visual world we perceive is continuous and precise. As a result, there is a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 545, + 388 + ], + "type": "image", + "image_path": "d93c6cf7827c50f011fcc04ba860ebfc839404c1ed92ff3fe1706222a1c5cfc8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 547, + 431 + ], + "type": "text", + "content": "Figure 1. Efficiency comparisons " + }, + { + "bbox": [ + 305, + 396, + 547, + 431 + ], + "type": "inline_equation", + "content": "(320 \\times 180" + }, + { + "bbox": [ + 305, + 396, + 547, + 431 + ], + "type": "text", + "content": " input) for ASSR. LMF-based ASSR methods significantly reduce computational cost (MACs), runtime (circle sizes), and parameters (colors)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 617 + ], + "type": "text", + "content": "growing demand for arbitrary scale image resizing in various real-world scenarios. For instance, users often require continuous resizing of images during capture, viewing, and editing. Recent developments in Single Image SuperResolution (SISR) [7, 9, 12, 16, 20, 21, 23-25, 36, 45-47] have showcased significant advancements in reconstructing High-Resolution (HR) images from Low-Resolution (LR) counterparts. However, traditional interpolation methods are still the preferred choice for handling real-world image resizing. This is because SISR methods typically learn upsampling functions for a particular scale factor, requiring training different models for different scales. In contrast, Arbitrary-Scale Super-Resolution (ASSR) utilizes a single model to handle arbitrary scales, thus presenting significant potential in practical image-resizing applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Building on the success of Implicit Neural Representation (INR), Local Implicit Image Function (LIIF) [11] is a recent ASSR approach that learns to continuously represent 2D images through an implicit function. LIIF encodes an image as latent codes and employs a decoding MLP to predict pixel values based on coordinates and their neighboring latent codes. The use of continuous coordinates enables LIIF and subsequent INR-based ASSR works" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26026" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 539, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 539, + 173 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 539, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 539, + 173 + ], + "type": "image", + "image_path": "76fb012507164f8c3e888194a2cf40616913ae4e357db06ad6b86223a62fc94b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 547, + 204 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 547, + 204 + ], + "type": "text", + "content": "Figure 2. The framework of our LMF-based continuous image representation. Given a latent code generated by the encoder, an HD latent MLP generates the latent modulation to adjust the features of an LD render MLP, thereby achieving efficient arbitrary-resolution rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 210, + 289, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 210, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 289, + 462 + ], + "type": "text", + "content": "[5, 8, 14, 15, 22, 34, 37-41] to represent images at any resolution, achieving robust performance in up to " + }, + { + "bbox": [ + 46, + 210, + 289, + 462 + ], + "type": "inline_equation", + "content": "\\times 30" + }, + { + "bbox": [ + 46, + 210, + 289, + 462 + ], + "type": "text", + "content": " SR task. However, existing INR-based ASSR methods adhere to an intuitive ASSR paradigm that uses a complex decoding function in HR High-Dimensional (HD) space to independently predict the pixel value of each HR coordinate. As a result, their computational cost and runtime increase quadratically with the scale factor, contradicting the goal of real-world image-resizing applications (refer to Figure 1). Upon closer examination, the significant computational redundancy of HR-HD decoding can be divided into three aspects. Firstly, there is inherent similarity within the HR region corresponding to an LR latent code, suggesting the need for merging shared computations within this region before independently predicting each HR pixel value. Secondly, this paradigm ignores the varying complexity of different LR latent codes and applies the same computational process to predict all the HR pixel values. Lastly, this paradigm couples the HD MLP parameters and HR computations, while such a large number of parameters are not always needed for each prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "type": "text", + "content": "To tackle this problem, we propose a novel Latent Modulated Function (LMF) to realize a computational optimal paradigm of continuous image representation. By leveraging latent modulation, LMF successfully disentangles the decoding process from HR-HD space into latent (LR-HD) and render (HR-LD) spaces. As shown in Figure 2, LMF first utilizes a latent MLP to generate the latent modulation of a given latent code. When querying the coordinates near the latent code, the latent modulation is applied to adjust each hidden linear layer in the render MLP. This modulation enables LMF to minimize the dimension of the render MLP without sacrificing performance. Furthermore, we observe a positive correlation between the latent modulation intensity and the input image complexity. Building on this insight, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm. CMSR not only allows the computational cost of rendering to be proportional to the input image complexity rather than the output resolution, but also provides the flexibility to balance precision and efficiency during testing. Computational optimal and fast ASSR holds great promise for future research and downstream appli" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": "cations. Experiments demonstrate that transforming existing INR-based ASSR methods into LMF reduces computational costs by " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "90.4\\%" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": ", accelerates inference by " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "2.4 \\times" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "56.9 \\times" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": ", and saves " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "45.1\\%" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "inline_equation", + "content": "76.0\\%" + }, + { + "bbox": [ + 304, + 210, + 545, + 270 + ], + "type": "text", + "content": " of parameters, while maintaining competitive PSNR performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 270, + 526, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 270, + 526, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 526, + 281 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 282, + 545, + 426 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 282, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 282, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 282, + 545, + 330 + ], + "type": "text", + "content": "- We propose a Latent Modulated Function (LMF) that realizes computational optimal continuous image representation, which decouples the HR-HD decoding into decoding in latent (LR-HD) and render (HR-LD) spaces." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 331, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 331, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 545, + 378 + ], + "type": "text", + "content": "- Based on the correlation between modulation intensity and input image complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to effortlessly balance the rendering efficiency and precision at test time." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 379, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 545, + 426 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that seamlessly converting existing INR-based ASSR methods to LMF can achieve substantial reductions in computational cost, runtime, and parameters, while preserving performance." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 392, + 448 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 455, + 474, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 474, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 474, + 468 + ], + "type": "text", + "content": "2.1. Implicit Neural Representation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 665 + ], + "type": "text", + "content": "INR has demonstrated remarkable success for continuous signal representation across various tasks and modalities, including signal reconstruction [10, 27, 31, 32], image and video compression [6, 13, 35], and image generation [33, 43]. The highlight of INR research is NeRF [28] and subsequent works [2, 3, 30, 42], which learn MLPs to infer the RGB and density of any given coordinates and view directions in a 3D scene. However, a major challenge faced by INR is the extensive training process required to fit a signal from scratch. To address this, MINER [31] and KiloNeRF [30] propose the use of multiple small MLPs to independently learn different segments of a signal. Another solution is hyper-base network design [10, 13, 27, 33], where a pre-trained hyper network generates INR for any input signal during testing. To represent complex signals, hyper networks often generate modulations for base networks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 671, + 474, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 474, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 474, + 684 + ], + "type": "text", + "content": "2.2. Single-Image Super-Resolution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "SISR as a classic low-level vision task has been studied for decades. In recent years, this field witnessed a variety of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26027" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "CNN-based networks, such as SRCNN [12], LapSRN [20], SRResNet [21], EDSR [25], MemNet [36], RCAN [46], RDN [47], and SRFBN [23]. Lately, image processing transformers, such as IPT [7], SwinIR [24], and HAT [9], have been proposed, which surpass CNN-based methods in SR performance using a large dataset. Although SISR methods have achieved remarkable performance, most of them are designed for a specific scale." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "content": "ASSR methods are designed to utilize a single network for all scales, showcasing significant advancements in practicality and convenience over previous works. Hu et al. proposed MetaSR [17] to generate an upsampling layer for ASSR using a meta-network. However, MetaSR exhibits poor generalization for large-scale SR. Building upon the success of INR, a novel continuous image representation was proposed in LIIF [11], where each pixel value is inferred from an MLP using its coordinate and neighboring 2D features. Subsequently, a variety of INR-based ASSR methods have been proposed [5, 8, 14, 15, 22, 34, 37-41]. Notably, Lee et al. proposed LTE [22], which introduces a dominant-frequency estimator to continuously capture fine details. Cao et al. introduced CiaoSR [5], a continuous implicit attention-in-attention network that learns ensemble weights for nearby local features, achieving State-Of-The-Art (SOTA) performance. Despite the robust performance of INR-based ASSR methods for scales up to " + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "inline_equation", + "content": "\\times 30" + }, + { + "bbox": [ + 46, + 167, + 288, + 407 + ], + "type": "text", + "content": ", they adhere to the paradigm of decoding in HR-HD space, leading to prohibitively high computational cost and runtime." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 418, + 107, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 107, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 107, + 430 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 201, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 201, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 201, + 449 + ], + "type": "text", + "content": "3.1. Preliminary and Motivation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": "In continuous image representations, an image " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "I_{in}" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": " is represented as 2D feature maps " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": " in latent space " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H_{in} \\times W_{in} \\times D_F}" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": ". A latent code " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "z^* \\in \\mathbb{R}^{D_F}" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": " represents a feature vector derived from the feature maps, which is used to decode its nearest coordinates in the continuous spatial domain " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "X \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": ". Since the encoder can be any feature extractor from existing SISR methods, the essence of continuous image representations is the decoding function, which arbitrarily upsamples LR-HD features to HR-LD pixel values. A simple and intuitive paradigm is to learn a decoding function in HR-HD space that independently decodes the pixel value of each HR coordinate. This vanilla decoding function " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": " parameterized as an HD MLP " + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 456, + 287, + 610 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 620, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 620, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 120, + 620, + 287, + 635 + ], + "type": "interline_equation", + "content": "I \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta} \\left(z ^ {*}, x _ {i, j} ^ {*}\\right), \\tag {1}", + "image_path": "cf764cecf0f1993a29976bf0e84ec4ac91a67ce2154c1776b8b94ded0c2f705e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "inline_equation", + "content": "x_{i,j}^{*}\\in X^{*}" + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "content": " refers to any nearest coordinate of " + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "inline_equation", + "content": "I(x_{i,j}^{*})" + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "content": " represents the predicted pixel value at " + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "inline_equation", + "content": "x_{i,j}^{*}" + }, + { + "bbox": [ + 47, + 641, + 287, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "However, as mentioned earlier in the Introduction section, this vanilla paradigm suffers from three aspects of computational redundancy. A computational optimal decoding function should allow computations proportional to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "the input image complexity rather than the output resolution. With this in mind, we propose to decouple the HR-HD decoding process into shared decoding in latent space and independent rendering in render space, along with a rendering algorithm adaptive to the input image content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 138, + 459, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 138, + 459, + 150 + ], + "spans": [ + { + "bbox": [ + 305, + 138, + 459, + 150 + ], + "type": "text", + "content": "3.2. Latent Modulated Function" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": "To eliminate the computational redundancy in HR-HD decoding, we first propose a two-stage decoding process comprising a latent MLP " + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": " and a render MLP " + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\theta_{r}" + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": ". The latent MLP handles the shared decoding process within the HR region of a latent code " + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": ", while the render MLP independently predicts the pixel value of each HR coordinate. The LR-HD output " + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta_l}(z^*)" + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": " from the latent MLP is utilized to assist the render MLP in rendering. However, directly using this output " + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta_l}(z^*)" + }, + { + "bbox": [ + 304, + 156, + 545, + 289 + ], + "type": "text", + "content": " as the input of the render MLP is akin to using the latent code itself, as demonstrated in Eq. 1, resulting in no efficiency improvement yet:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 294, + 545, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 294, + 545, + 309 + ], + "spans": [ + { + "bbox": [ + 367, + 294, + 545, + 309 + ], + "type": "interline_equation", + "content": "I \\left(x _ {i, j} ^ {*}\\right) = f _ {\\theta_ {r}} \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right), x _ {i, j} ^ {*}\\right). \\tag {2}", + "image_path": "813ceda28ac34c33b2b8cc28d5ae54806ea49394258901706e2aeb2364116d09.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "text", + "content": "Furthermore, to render arbitrary pixels accurately, it is necessary for the render MLP to operate in HR space. Therefore, the most viable approach to minimize rendering computations is to reduce the dimension or depth of the render MLP. However, an HD render MLP is still needed in Eq. 2 to fully receive and process the output " + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "inline_equation", + "content": "f_{\\theta_l}(z^*)" + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "text", + "content": ", which contradicts our intention of utilizing an LD render MLP. To overcome these limitations, modulations become a suitable solution to connect the latent and render space. Modulations have been used in INR [13, 27, 33] and are often associated with the hyper-base network design. A hyper network converts an input image to modulations, which are then utilized by a small base network to map coordinates to a continuous image. However, these modulations are shared by the input image and work at the image level, resulting in unsatisfactory performance and challenging training. Hence, we introduce a latent modulation " + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "inline_equation", + "content": "m^{*} \\in \\mathbb{R}^{D_M}" + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "text", + "content": ", which is generated by the latent MLP using the latent code " + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 314, + 546, + 553 + ], + "type": "text", + "content": ". The latent modulation consists of the FiLM layers [29], where each layer consists of a scale modulation and a shift modulation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 560, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 560, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 330, + 560, + 545, + 574 + ], + "type": "interline_equation", + "content": "m ^ {*} = \\left[ \\alpha_ {1} ^ {*}, \\beta_ {1} ^ {*}, \\dots , \\alpha_ {K} ^ {*}, \\beta_ {K} ^ {*} \\right] = S p l i t \\left(f _ {\\theta_ {l}} \\left(z ^ {*}\\right)\\right), \\tag {3}", + "image_path": "f0b3edbd859795eb8e25903fc8344f24f5c178253f4c7f34794bb95f133652eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\alpha^{*}" + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "text", + "content": " represents the element-wise scale modulation, " + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "inline_equation", + "content": "\\beta^{*}" + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "text", + "content": " represents the element-wise shift modulation, and " + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 579, + 545, + 614 + ], + "type": "text", + "content": " represents the depth of hidden layers in the render MLP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": "For each HR coordinate, the latent modulation " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "m^{*}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " of the latent code " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " nearest to this coordinate is applied to the render MLP, and is applied to the hidden features " + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "inline_equation", + "content": "h^{*} \\in \\mathbb{R}^{D_{H}}" + }, + { + "bbox": [ + 304, + 615, + 545, + 663 + ], + "type": "text", + "content": " after each hidden linear layer:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 341, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 341, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "h _ {k + 1} ^ {*} = f _ {\\theta_ {r} ^ {k + 1}} \\left(\\sigma \\left(\\left(1 + \\alpha_ {k} ^ {*}\\right) \\odot h _ {k} ^ {*} + \\beta_ {k} ^ {*}\\right)\\right), \\tag {4}", + "image_path": "fc8245c850de96e74aa771d8cb838e47423be42024ab9af672fb30dd743564bb.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_k^*" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents the hidden features after the " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": "-th hidden linear layer, and " + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 689, + 545, + 713 + ], + "type": "text", + "content": " represents the activation function." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26028" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 162, + 160 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 162, + 160 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 162, + 160 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 162, + 160 + ], + "type": "image", + "image_path": "ecc20c03c6c14fe0ce269b16d540715265ee23591ea5c0665c4a77bda9afa34d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 168, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 202 + ], + "type": "text", + "content": "Figure 3. Visualization " + }, + { + "bbox": [ + 47, + 168, + 287, + 202 + ], + "type": "inline_equation", + "content": "(\\times 4" + }, + { + "bbox": [ + 47, + 168, + 287, + 202 + ], + "type": "text", + "content": " SR) of the normalized mean of shift modulation (left), and the normalized residual between the LM-LIIF predicted image and the bilinear upsampled image (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 69, + 267, + 161 + ], + "blocks": [ + { + "bbox": [ + 175, + 69, + 267, + 161 + ], + "lines": [ + { + "bbox": [ + 175, + 69, + 267, + 161 + ], + "spans": [ + { + "bbox": [ + 175, + 69, + 267, + 161 + ], + "type": "image", + "image_path": "6ca0ef0fa8030fd66f06b6d03f4752b998a50135c9dc33bf6601303abc65fd04.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 208, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 208, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 289, + 447 + ], + "type": "text", + "content": "Importantly, the proposed latent modulation effectively solves the incompatibility between LR-HD latent codes and HR-LD rendering and significantly disentangles the computations from parameters. For a latent modulation with 192 dimensions, we can accurately render images using a 7-layer render MLP with only 16 dimensions. This LD render MLP only learns the general rendering rules of natural images, while the latent modulation optimizes its rendering process based on the characteristics of each input latent code. As a result, a tiny render MLP can quickly adapt to any input latent code and accomplish efficient rendering at arbitrary resolution. A remaining concern is that the latent codes in existing INR-based ASSR methods often possess high dimensions, resulting in unnecessary computations when directly input into the render MLP. Therefore, we compress the HD latent code to an LD one " + }, + { + "bbox": [ + 47, + 208, + 289, + 447 + ], + "type": "inline_equation", + "content": "z_{c}^{*}" + }, + { + "bbox": [ + 47, + 208, + 289, + 447 + ], + "type": "text", + "content": " using the latent MLP, extending Eq. 3 to the first stage in Eq. 5. Building on the latent modulation, we propose a computational optimal paradigm of continuous image representation, called Latent Modulated Function (LMF):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 452, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 452, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 93, + 452, + 287, + 471 + ], + "type": "text", + "content": "Stage 1: " + }, + { + "bbox": [ + 93, + 452, + 287, + 471 + ], + "type": "inline_equation", + "content": "[m^{*},z_{c}^{*}] = S p l i t(f_{\\theta_{l}}(z^{*}))" + }, + { + "bbox": [ + 93, + 452, + 287, + 471 + ], + "type": "text", + "content": " (5)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 468, + 240, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 468, + 240, + 482 + ], + "spans": [ + { + "bbox": [ + 94, + 468, + 240, + 482 + ], + "type": "text", + "content": "Stage 2: " + }, + { + "bbox": [ + 94, + 468, + 240, + 482 + ], + "type": "inline_equation", + "content": "I(x_{i,j}^{*}) = f_{\\theta_r,m^*}(z_c^*,x_{i,j}^*)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 288, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 641 + ], + "type": "text", + "content": "LMF is a universal paradigm for efficient ASSR and is compatible with existing INR-based ASSR methods. By effortlessly relocating the time-consuming operations and large MLPs to latent space, LMF allows a minimal render MLP to perform the same level of ASSR as the original ASSR method. In addition, LMF effectively decouples the computational expensive techniques of feature unfolding and local ensemble [11]. Specifically, LMF only utilizes feature unfolding for the latent MLP, where nearby " + }, + { + "bbox": [ + 46, + 486, + 288, + 641 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 486, + 288, + 641 + ], + "type": "text", + "content": " feature vectors are concatenated as inputs. LMF performs local ensemble exclusively for the render MLP, which merges the predictions from the top-left, top-right, bottom-left, and bottom-right latent codes to achieve continuous predictions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 239, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 239, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 239, + 659 + ], + "type": "text", + "content": "3.3. Controllable Multi-Scale Rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "The HR regions of two latent codes often have different signal complexity, allowing for the use of different decoding functions. However, the challenge lies in obtaining the signal complexity of each latent code. The modulation scheme" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 521, + 170 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 521, + 170 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 521, + 170 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 521, + 170 + ], + "type": "image", + "image_path": "da796c39b49cd2f67a7bb8191164a67b5b76657f5707106e84e1ab0cfbf5d767.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 306, + 179, + 545, + 201 + ], + "type": "text", + "content": "Figure 4. The positive correlation between the means of shift modulation and the minimum rendering scale factors in LM-LIIF." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "type": "text", + "content": "in Eq. 4 suggests that the intensity of latent modulation is proportional to the signal complexity of its latent code. A lower intensity implies that the signal can be sufficiently rendered by the render MLP itself. In Figure 3, we display the normalized mean of shift modulation (left), and the normalized residual between the SR result of LMF-based LIIF and the bilinear upsampled result (right). It is evident that the means of modulation closely align with the signal complexity of the predicted image. Regions with higher means of modulation consistently correspond to more complex textures and edges. Hence, we employ the mean of shift modulation as the intensity of latent modulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": "In LMF, we define the signal complexity of the latent code " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " as the minimum scale factor " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "s^{*}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " needed to fully represent its signal. In simple terms, any rendering with a scale factor " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "s, s > s^{*}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " will yield the same result as rendering with " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "s^{*}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " followed by upsample interpolation (i.e. bilinear) with scale " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\frac{s}{s^{*}}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": ". Therefore, we only need to render " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " with scale " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\min(s^{*}, s)" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": " when upsampling an image with a scale " + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 354, + 546, + 555 + ], + "type": "text", + "content": ". To obtain the minimum scale factor of each mean of shift modulation, in Figure 4, we sample 100 means of shift modulation uniformly from the range of [0, 1] and measure their minimum scale factors on the LR images in the training set. Figure 4 clearly shows the positive correlation between the means of shift modulation and the minimum scale factors. However, since they do not have a one-to-one correspondence, we create a Scale2Mod table by sampling common scale factors and recording their minimum and maximum means of shift modulation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": "Based on the Scale2Mods table, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm to decode latent codes with their minimum scale factors. Firstly, we initialize " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "s_{-1}" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": " as 1, " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_{s_{-1}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": " as the LR image. For each sampled scale " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "s_i \\in [s_0, s_1, \\ldots, s]" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": ", we first upsample the last result " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_{s_{i-1}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": " using bilinear interpolation with " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\frac{s_i}{s_{i-1}}" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "I_{s_i}" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": ". We then query the Scale2Mods table to obtain the range of modulation means " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "[m_{min}^{s_i}, m_{max}^{s_i}]" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 304, + 556, + 547, + 713 + ], + "type": "text", + "content": ", the render MLP only needs to render the latent codes whose corresponding modulation means falling within this range. We repeat the above process iteratively until we obtain the final SR result or until all latent codes have been decoded. For precision controllable rendering, when creating" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26029" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 91, + 531, + 209 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 533, + 82 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 533, + 82 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 533, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparison (PSNR (dB)/MACs) for ASSR on DIV2K validation set. The best results are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 91, + 531, + 209 + ], + "lines": [ + { + "bbox": [ + 63, + 91, + 531, + 209 + ], + "spans": [ + { + "bbox": [ + 63, + 91, + 531, + 209 + ], + "type": "table", + "html": "
MethodParamsIn-scaleOut-of-scale
×2×3×4×6×12×18
Bicubic-31.01/-28.22/-26.66/-24.82/-22.27/-21.00/-
EDSR-b [25]-34.55/-30.90/-28.94/----
LIIF [11]355K34.67/3.567T30.96/3.567T29.00/3.567T26.75/3.567T23.71/3.567T22.17/3.567T
LM-LIIF170K34.65/125.32G30.95/65.37G28.99/44.38G26.75/25.30G23.72/6.61G22.19/2.95G
LTE [22]506K34.72/2.228T31.02/2.123T29.04/2.086T26.81/2.059T23.78/2.044T22.23/2.026T
LM-LTE278K34.71/198.81G30.99/100.75G29.03/66.43G26.78/34.96G23.74/9.07G22.21/4.10G
CiaoSR-1.4M [5]1.429M34.91/11.460T31.15/11.443T29.23/11.437T26.95/11.433T23.88/11.430T22.32/11.348T
LM-CiaoSR753K34.81/1.071T31.10/501.36G29.12/293.01G26.86/135.66G23.80/42.60G22.26/23.90G
", + "image_path": "3c873803e3295f6b3a555ac8ec4491e2af5922ce1b093550f94afe3951677341.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 239, + 544, + 426 + ], + "blocks": [ + { + "bbox": [ + 87, + 219, + 504, + 231 + ], + "lines": [ + { + "bbox": [ + 87, + 219, + 504, + 231 + ], + "spans": [ + { + "bbox": [ + 87, + 219, + 504, + 231 + ], + "type": "text", + "content": "Table 2. Quantitative comparison (PSNR (dB)) for in-scale SR on test sets. The best results are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 239, + 544, + 426 + ], + "lines": [ + { + "bbox": [ + 50, + 239, + 544, + 426 + ], + "spans": [ + { + "bbox": [ + 50, + 239, + 544, + 426 + ], + "type": "table", + "html": "
MethodSet5Set14BSD100Urban100DIV2K100
×2×3×4×2×3×4×2×3×4×2×3×4×2×3×4
RDN [47]38.2434.7132.4734.0130.5728.8132.3429.2627.7232.8928.8026.6134.9431.2229.19
LIIF [11]38.1734.6832.5033.9730.5328.8032.3229.2627.7432.8728.8226.6834.9931.2629.27
LM-LIIF38.1834.7232.5034.0630.5528.8332.3229.2827.7532.8228.8126.6634.9631.2529.27
LTE [22]38.2334.7232.6134.0930.5828.8832.3629.3027.7733.0428.9726.8135.0431.3229.33
LM-LTE38.2334.7632.5334.1130.5628.8632.3729.3127.7833.0328.9626.8035.0331.3129.32
CiaoSR-1.4M [5]38.2934.8532.6634.2230.6528.9332.4129.3427.8333.3029.1727.1135.1531.4229.45
LM-CiaoSR38.2434.8132.5934.1130.6228.9032.3929.3227.8033.1329.1326.9535.0831.3629.37
SwinIR [24]38.3534.8932.7234.1430.7728.9432.4429.3727.8333.4029.2927.0735.2631.5029.48
LIIF [11]38.2834.8732.7334.1430.7528.9832.3929.3427.8433.3629.3327.1535.1731.4629.46
LM-LIIF38.3134.8832.7834.3130.7929.0232.4329.3727.8533.4629.4327.2135.2131.4829.49
LTE [22]38.3334.8932.8134.2530.8029.0632.4429.3927.8633.5029.4127.2435.2431.5029.51
LM-LTE38.3234.8832.7734.2830.7929.0132.4629.3927.8733.5229.4427.2435.2431.5029.50
CiaoSR-3.1M [5]38.3834.9132.8434.3330.8229.0832.4729.4227.9033.6529.5227.4235.2931.5529.59
LM-CiaoSR38.3734.9432.8734.3330.8329.0832.4829.4327.9033.6529.5827.3935.3031.5529.56
", + "image_path": "7dedcc786b6edd74f38c50afb5570bb26264bb1f5a63046762fc3299d1715439.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "text", + "content": "the Scale2Mods table, we set an MSE threshold " + }, + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "text", + "content": " and define the signal complexity of a latent code as the minimum scale factor needed to render its signal within this error " + }, + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 435, + 287, + 509 + ], + "type": "text", + "content": ". By effortlessly adjusting the MSE threshold during testing, CMSR is capable of rendering with varying precision and speed, while adapting naturally to the image content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 519, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 128, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 539, + 167, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 167, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 167, + 552 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": "Dataset. Following the setups in previous ASSR methods [5, 11, 22], we employ the DIV2K dataset [1] as the training set. In each training iteration, we uniformly sample a downsampling scale " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " from a continuous range of " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "[ \\times 1, \\times 4 ]" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": ". HR patches with spatial size of " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "48s \\times 48s" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " are cropped and bicubic downsampled to LR patches with spatial size of " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": ". To supervise the training process, we randomly sample " + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 46, + 557, + 287, + 689 + ], + "type": "text", + "content": " pixels from each HR patch. For evaluation, we present results on the DIV2K validation set (DIV2K 100) along with four widely-used benchmark test sets: Set5 [4], Set14 [44], BSD100 [26], and Urban100 [18]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Implementation Details. We transform three previous ASSR methods, namely LIIF [11], LTE [22], and CiaoSR" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 545, + 555 + ], + "type": "text", + "content": "[5], into our proposed LMF. LM-LIIF employs a 2-layer latent MLP with 208 dimensions, and a 7-layer render MLP with 16 dimensions. As for LM-LTE, we employ a 2-layer latent MLP with 256 dimensions, a 9-layer render MLP with 16 dimensions, and a local texture estimator [22] with 128 dimensions. LM-CiaoSR employs four 2-layer latent MLPs with 256 dimensions to generate key, value, and query information, followed by a 9-layer render MLP with 16 dimensions. The detailed architectures of LMF-based methods are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": "We utilize EDSR-baseline (EDSR-b) [25], RDN [47], and SwinIR [24] without their upsampling modules as encoders. For both EDSR-b and RDN encoders, we conduct training for 1000 epochs with a batch size of 16. The learning rate is initialized as " + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": " and decayed by a factor of 0.5 every 200 epochs. As for the SwinIR encoder, we conduct training for 1000 epochs with a batch size of 32. The learning rate is initialized as " + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": " and decayed by a factor of 0.5 at [500, 800, 900, 950] epochs. For optimization, we utilize the L1 pixel loss as the loss function and employ the Adam [19] optimizer with " + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 556, + 546, + 700 + ], + "type": "text", + "content": ". Our training details align with previous ASSR works [5, 22]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Settings. In line with previous methods [5, 11," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26030" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 102, + 544, + 266 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 3. Quantitative comparison (PSNR (dB)) for out-of-scale SR on test sets. The best results are highlighted in bold. The first 6 rows use RDN as the encoder, and the last 6 rows use SwinIR as the encoder." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 102, + 544, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 266 + ], + "type": "table", + "html": "
MethodSet5Set14BSD100Urban100DIV2K100
×6×8×12×6×8×12×6×8×12×6×8×12×6×12×18
LIIF [11]29.1527.1424.8626.6425.1523.2425.9824.9123.5724.2022.7921.1526.9923.8922.34
LM-LIIF29.1827.2124.7126.6625.1723.3125.9824.9323.5824.1422.7721.1526.9923.9122.36
LTE [22]29.3227.2624.7926.7125.1623.3126.0124.9523.6024.2822.8821.2227.0423.9522.40
LM-LTE29.2127.2224.8226.6925.1523.3226.0124.9523.6124.2622.8621.2427.0323.9522.39
CiaoSR-1.4M [5]29.4627.3624.9226.7925.2823.3726.0725.0023.6424.5823.1321.4227.1624.0622.48
LM-CiaoSR29.2427.2624.8026.7225.1823.3026.0324.8123.6024.4022.9821.3327.0723.9822.42
LIIF [11]29.4627.3624.9926.8225.3423.3926.0725.0123.6524.5923.1421.4327.1524.0222.43
LM-LIIF29.4627.3624.9826.8225.3723.4326.0825.0323.6524.6123.1521.4927.1824.0622.49
LTE [22]29.5027.3525.0726.8625.4223.4426.0925.0323.6624.6223.1721.5027.2024.0922.50
LM-LTE29.5227.3925.0126.8325.3823.4126.1025.0323.6624.6323.1721.5027.1924.0822.50
CiaoSR-3.1M [5]29.6227.4524.9626.8825.4223.3826.1325.0723.6824.8423.3421.6027.2824.1522.54
LM-CiaoSR29.5227.4225.0226.8425.4123.4526.1225.0623.6824.7423.2621.5827.2424.1222.53
", + "image_path": "f6b54eab31ade454104942ef8bcac0b8fe27a5247b5fda5b360207394ae80f83.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 296, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 104, + 275, + 488, + 288 + ], + "lines": [ + { + "bbox": [ + 104, + 275, + 488, + 288 + ], + "spans": [ + { + "bbox": [ + 104, + 275, + 488, + 288 + ], + "type": "text", + "content": "Table 4. Efficiency comparison (MACs/GPU runtime) for ASSR. The best results are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 296, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 296, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 296, + 544, + 430 + ], + "type": "table", + "html": "
DecoderParamsGPU mem.180p270p360p
720p1080p2k720p1080p2k720p1080p2k
LIIF [11]355K7.7GB1.160T/1.1148s2.609T/2.5230s4.638T/4.5232s1.160T/1.1195s2.609T/2.5347s4.638T/4.5544s1.160T/1.1272s2.609T/2.5456s4.638T/4.5589s
LM-LIIF170K5.3GB14.43G/0.0246s18.72G/0.0474s24.72G/0.0795s25.39G/0.0363s32.46G/0.0593s37.50G/0.0919s40.74G/0.0541s47.81G/0.0770s57.71G/0.1107s
LTE [22]506K8.4GB694.34G/0.9135s1.506T/2.0366s2.666T/3.7578s714.11G/0.9191s1.526T/2.0497s2.685T/3.7694s741.80G/0.9258s1.553T/2.0637s2.712T/3.7791s
LM-LTE278K4.7GB21.60G/0.0382s25.87G/0.0750s33.21G/0.1254s39.53G/0.0545s48.59G/0.0905s52.88G/0.1414s64.63G/0.0774s73.69G/0.1123s86.38G/0.1640s
CiaoSR-1.4M [5]1.429M8.8GB3.718T/1.8820s8.362T/4.1718s14.864T/7.3619s3.721T/1.9838s8.365T/4.2999s14.867T/7.5611s3.725T/2.1355s8.370T/4.5064s14.871T/7.8366s
LM-CiaoSR753K8.9GB95.23G/0.2637s99.57G/0.3849s106.68G/0.5546s204.09G/0.5263s214.32G/0.6476s217.83G/0.8160s356.46G/0.8844s366.69G/1.0013s381.02G/1.1778s
", + "image_path": "9296bc04e468dc0dcb5daa8c97468a019a2398227d5119a73495c4413b046f31.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 438, + 288, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 535 + ], + "type": "text", + "content": "22], we employ PSNR as the objective criterion to assess quality performance. To evaluate efficiency, we measure the computational cost using Multiply-Accumulate Operations (MACs) and quantify the model speed using the runtime on an RTX 2080 Ti GPU. Additionally, we provide the number of parameters and GPU memory usage. Notably, the MACs, runtime, GPU memory, and parameters are evaluated solely on the decoder unless explicitly stated." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 159, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 159, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 159, + 562 + ], + "type": "text", + "content": "4.2. ASSR Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": "Quantitative Results. In Table 1, we compare our LMF-based methods with reference ASSR methods, namely LIIF [11], LTE [22], and CiaoSR [5], on the DIV2K validation set using EDSR-b as the encoder. We include the results of bicubic interpolation and EDSR-b with its initial upsampling module for reference. The results demonstrate that LMF-based methods achieve a significant reduction in computational cost while delivering competitive PSNR performance compared to the original methods. Compared to LIIF, LM-LIIF achieves a reduction of " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "96.6\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "99.9\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " in computational cost across all scales. Similarly, LM-LTE and LM-CiaoSR provide " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "91.3\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "inline_equation", + "content": "90.4\\%" + }, + { + "bbox": [ + 46, + 570, + 288, + 713 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 438, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 499 + ], + "type": "inline_equation", + "content": "99.8\\%" + }, + { + "bbox": [ + 304, + 438, + 545, + 499 + ], + "type": "text", + "content": " reduction in computational cost compared to LTE and CiaoSR, respectively. However, LM-CiaoSR exhibits slightly lower PSNR than CiaoSR due to relocating the expensive attention-in-attention modules from HR-HD space to latent space for the minimal computational cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 504, + 546, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 546, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 546, + 660 + ], + "type": "text", + "content": "In Table 2 and 3, we present the comprehensive PSNR evaluation for ASSR using RDN and SwinIR as the encoders. Notably, CiaoSR utilizes a 1.429M decoder for EDSR-b and RDN, and a 3.065M decoder for SwinIR, while our LM-CiaoSR employs the same decoder for all encoders. The results demonstrate the competitive PSNR performance of our LMF-based methods, especially LM-LIIF and LM-LTE, compared to the reference methods on benchmark test sets. Notably, LMF-based methods exhibit superior PSNR performance gains on the SwinIR encoder. The results presented in Table 1, 2, and 3 collectively validate the universal applicability of our LMF across different encoders, INR-based ASSR methods, test sets, and scales." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "In Table 4, we present comparisons of MACs, runtime, GPU memory, and parameters for ASSR across multiple input and output resolutions. We utilize bicubic downsampled images from the DIV2K validation set as inputs for" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26031" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 539, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 235 + ], + "type": "image", + "image_path": "c95688909c8a3a98fc36451b03ecff56aa19ef690da168c9230e63e1b69b47b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 243, + 473, + 255 + ], + "lines": [ + { + "bbox": [ + 118, + 243, + 473, + 255 + ], + "spans": [ + { + "bbox": [ + 118, + 243, + 473, + 255 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison for ASSR. All of the ASSR methods use SwinIR as the encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 286, + 544, + 393 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 545, + 277 + ], + "type": "text", + "content": "Table 5. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on latent modulations and latent codes, using EDSR-b as the encoder." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 286, + 544, + 393 + ], + "lines": [ + { + "bbox": [ + 49, + 286, + 544, + 393 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 544, + 393 + ], + "type": "table", + "html": "
DecoderLatent codeLatent modulationLatent MLPRender MLPParamsIn-scaleOut-of-scale
×2×4×8×16
LIIF576××5 × 256355K34.67/3.567T29.00/3.567T25.40/3.567T22.61/3.567T
LIIF (tiny)576××9 × 1612K33.61/117.44G28.18/117.44G24.78/117.44G22.16/117.44G
LIIF (C2F)192×2 × 1927 × 16202K34.11/221.54G28.63/114.62G25.11/73.04G22.38/58.91G
LM-LIIF64Scale2 × 1927 × 16136K34.61/110.94G28.96/46.87G25.38/30.79G22.61/26.58G
LM-LIIF64Shift2 × 1927 × 16136K34.59/110.94G28.94/46.87G25.36/30.79G22.59/26.58G
LM-LIIF64Scale & shift2 × 1927 × 16155K34.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
LM-LIIF16Scale & shift2 × 2087 × 16170K34.65/125.32G28.99/44.38G25.40/24.10G22.62/18.90G
", + "image_path": "f8968c3d8a12b677a57c57b070a6e274b45a720e0f2af6dadb09c76f387002bc.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": "each resolution, and all results exclude the encoders. As shown in the table, our LMF-based methods exhibit a significant reduction in computational cost, ranging from " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "90.7\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "99.5\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": ", and a substantial reduction in runtime, ranging from " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "58.6\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "98.2\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": ", compared to the reference methods. Compared to the reference methods, LMF-based methods also exhibit a reduction in parameters, ranging from " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "45.1\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "inline_equation", + "content": "52.1\\%" + }, + { + "bbox": [ + 46, + 402, + 287, + 521 + ], + "type": "text", + "content": ", which is attributed to the decoupling of computations from parameters, demonstrating the significance of our elegant design over merely stacking parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 654 + ], + "type": "text", + "content": "Qualitative Results. In Figure 5, we present a qualitative comparison for ASSR, using SwinIR as the encoder. The visual results demonstrate that our LMF-based methods achieve competitive visual quality across all scales compared to the reference methods. Remarkably, LMF-based methods achieve continuous and precise results while significantly reducing computational cost and runtime. This confirms that LMF effectively addresses the challenge of upsampling LR-HD features to HR-LD pixel values, eliminating the need for an HR-HD decoding function. More visual results are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 668, + 140, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 140, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 140, + 682 + ], + "type": "text", + "content": "5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 712 + ], + "type": "text", + "content": "In this section, we investigate the effect of each component in LMF. Both PSNR and MACs results on the DIV2K vali" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 545, + 437 + ], + "type": "text", + "content": "dation set are provided for a comprehensive evaluation, and all of the models use EDSR-b as the encoder. Notably, the CMSR algorithm is not used in Table 5 and Table 6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 546, + 640 + ], + "type": "text", + "content": "Latent Modulation. In Table 5, we evaluate the impact of latent modulation. Firstly, the comparison between LIIF and LIIF (tiny) reveals a significant decline in PSNR performance when shrinking the decoder from a 5-layer MLP with 256 dimensions to a 9-layer MLP with 16 dimensions. This performance drop is expected due to the substantial reduction in network parameters. However, our LM-LIIF model, which utilizes a 7-layer render MLP with 16 dimensions, outperforms LIIF in terms of PSNR, while consuming computational cost even lower than LIIF (tiny). Furthermore, we introduce a coarse-to-fine variant of LIIF, where both stages employ the same structure as LM-LIIF. Unlike LM-LIIF, LIIF (C2F) fails to achieve competitive PSNR performance compared to LIIF, highlighting the critical role of latent modulation in reducing the dimensions of the render MLP. These results also validate that HR-HD computations are redundant for decoding LR-HD features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "In Table 5, we also compare the performance of different types of latent modulations, namely the scale modulation, the shift modulation, and the combination of scale and shift modulations. The results demonstrate that LM-LIIF achieves the highest PSNR performance when both scale and shift modulations are employed, and the computational" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26032" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 113, + 287, + 272 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 103 + ], + "type": "text", + "content": "Table 6. Ablation study (PSNR(dB)/MACs on DIV2K validation set) on the latent MLP and the render MLP, using EDSR-b as the encoder, and LM-LIIF as the decoder." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 113, + 287, + 272 + ], + "lines": [ + { + "bbox": [ + 48, + 113, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 113, + 287, + 272 + ], + "type": "table", + "html": "
Latent MLPRender MLPIn-scaleOut-of-scale
×2×4×8×16
2 × 2562 × 12834.66/236.91G28.97/131.21G25.37/104.57G22.60/97.18G
2 × 2563 × 6434.67/232.18G29.00/126.48G25.41/99.85G22.63/92.50G
2 × 2565 × 3234.68/197.37G29.00/91.67G25.41/65.11G22.63/58.04G
4 × 2569 × 1634.64/258.37G28.98/87.79G25.40/45.04G22.62/34.13G
2 × 2569 × 1634.66/171.85G28.99/66.16G25.40/39.65G22.62/32.79G
2 × 1927 × 1634.66/123.10G29.00/49.91G25.40/31.55G22.63/26.77G
2 × 1285 × 1634.63/79.76G28.98/35.02G25.39/23.79G22.62/20.83G
", + "image_path": "fceee7d91c70857265418261404b246856933a56a9d4311a524a0eabba083df5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 377 + ], + "type": "text", + "content": "cost remains nearly identical. Finally, we compare using the latent code (64 dimensions) from the encoder and the compressed latent code (16 dimensions) from the latent MLP as the input of the render MLP. The results indicate that using the compressed latent code can further reduce the computational cost for large scales. Based on these findings, we incorporate both scale and shift modulations as well as the compressed latent code in our final LMF models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 378, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 287, + 521 + ], + "type": "text", + "content": "Network Structure. We evaluated different depths and dimensions of the latent MLP and the render MLP in Table 6. Notably, the compressed latent code is not used to ensure better alignment of depths and dimensions between the two MLPs. For the latent MLP, we observe that decreasing the depth of the latent MLP from 4 to 2 yields nearly the same PSNR performance, which suggests that the latent MLP does not require a deep architecture or a large number of parameters. However, when reducing the dimension of the latent MLP from 192 to 128, the PSNR performance begins to decrease. This observation confirms that HD computations are crucial for a powerful latent MLP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "content": "In Table 6, we investigate the depths and dimensions of the render MLP by fixing the dimensions of the latent MLP and the latent modulation at 256. The results demonstrate that we can effortlessly reduce the dimension of the render MLP from 128 to 16, resulting in an impressive up to " + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "inline_equation", + "content": "66.26\\%" + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "content": " reduction in computational cost, without compromising performance. When the latent modulation possesses a fixed dimension, it is consistently viable to utilize a lower dimensional rendering process, thereby minimizing the computational burden. All things considered, our final LM-LIIF model employs a 2-layer latent MLP with 208 dimensions (16 dimensions for the compressed latent code) and a 7-layer render MLP with 16 dimensions, optimizing computational efficiency without sacrificing performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Controllable Multi-Scale Rendering. In Table 7, we evaluate the MSE thresholds when creating the Scale2Mod ta" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 113, + 544, + 204 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 7. Ablation study (PSNR(dB)/MACs of rendering/MACs of decoder on DIV2K validation set) on MSE thresholds of CMSR, using EDSR-b as the encoder, and LM-LIIF as the decoder." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 113, + 544, + 204 + ], + "lines": [ + { + "bbox": [ + 309, + 113, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 309, + 113, + 544, + 204 + ], + "type": "table", + "html": "
MSE thresholdIn-scaleOut-of-scale
×2×4×8×16
×34.65/17.41G/125.32G28.99/17.41G/44.38G25.40/17.37G/24.10G22.62/17.23G/18.90G
2 × 10-534.65/16.30G/124.22G28.99/13.67G/40.62G25.40/12.41G/19.14G22.62/3.48G/5.15G
1 × 10-434.62/13.63G/121.55G28.98/11.71G/38.69G25.39/8.35G/15.08G22.62/2.50G/4.17G
5 × 10-434.42/10.35G/118.27G28.93/8.94G/35.92G25.38/6.08G/12.81G22.61/2.29G/3.96G
", + "image_path": "c8b0b128273aa1bf5c9e77dc7bfbfdb0411d28bae37616628f3aabf74f3ef02c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": "ble for CMSR. Compared to the case without CMSR, a Scale2Mod table created with an MSE threshold of " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " achieves the same PSNR performance while reducing computational cost by " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "20.58\\%" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "72.75\\%" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " for scale " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": ". With lower precision requirements, i.e., MSE thresholds of " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": ", CMSR effectively reduces the computational cost even further while maintaining maximum PSNR performance. This adaptability is advantageous for real-world applications as the precision can be adjusted during testing without the need for model retraining. However, CMSR exclusively reduces the computational cost of the render MLP, which limits the computational cost reduction for small scales. Hence, in the Experiments section, we only employ CMSR with the MSE threshold " + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 217, + 545, + 408 + ], + "type": "text", + "content": " for out-of-distribution scales. More details and results regarding CMSR are provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 425, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 425, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 425, + 378, + 438 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 626 + ], + "type": "text", + "content": "In this paper, we propose a Latent Modulated Function (LMF), which realizes a computational optimal paradigm of continuous image representation. LMF utilizes an MLP in latent space to generate the latent modulation of each latent code. This enables a modulated MLP in render space to perform effective rendering at any resolution. Moreover, based on the correlation between modulation intensity and signal complexity, we propose a Controllable Multi-Scale Rendering (CMSR) algorithm, offering the flexibility to balance the rendering precision and efficiency during testing. Experiments demonstrate that converting INR-based ASSR methods to LMF leads to substantial reductions in computational cost, runtime, and parameters, while preserving performance. Better forms of latent modulations and more efficient decoder architectures will be explored in future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 643, + 416, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 643, + 416, + 657 + ], + "spans": [ + { + "bbox": [ + 306, + 643, + 416, + 657 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No. 62071500; Supported by Shenzhen Science and Technology Program under Grant No. JCYJ20230807111107015." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26033" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 213 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 288, + 270 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 316 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. British machine vision conference, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 384 + ], + "type": "text", + "content": "[5] Jiezhang Cao, Qin Wang, Yongqin Xian, Yawei Li, Bingbing Ni, Zhiming Pi, Kai Zhang, Yulun Zhang, Radu Timofte, and Luc Van Gool. Ciaosr: Continuous implicit attention-inattention network for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 430 + ], + "type": "text", + "content": "[6] Hao Chen, Bo He, Hanyu Wang, Yixuan Ren, Ser Nam Lim, and Abhinav Shrivastava. Nerv: Neural representations for videos. Advances in Neural Information Processing Systems, 34:21557-21568, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 431, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 486 + ], + "type": "text", + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 287, + 553 + ], + "type": "text", + "content": "[8] Hao-Wei Chen, Yu-Syuan Xu, Min-Fong Hong, Yi-Min Tsai, Hsien-Kai Kuo, and Chun-Yi Lee. Cascaded local implicit transformer for arbitrary-scale super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18257-18267, 2023. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 610 + ], + "type": "text", + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22367-22377, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 655 + ], + "type": "text", + "content": "[10] Yinbo Chen and Xiaolong Wang. Transformers as metalearners for implicit neural representations. In European Conference on Computer Vision, pages 170-187. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[11] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8628-8638, 2021. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[12] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European Conference on Computer Vision, pages 184-199, 2014. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 160 + ], + "type": "text", + "content": "[13] Emilien Dupont, Hrushikesh Loya, Milad Alizadeh, Adam Golinski, Yee Whye Teh, and Arnaud Doucet. Coin++: Neural compression across modalities. arXiv preprint arXiv:2201.12904, 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 547, + 226 + ], + "type": "text", + "content": "[14] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 547, + 260 + ], + "type": "text", + "content": "[15] Zongyao He and Zhi Jin. Dynamic implicit image function for efficient arbitrary-scale image representation. arXiv preprint arXiv:2306.12321, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[16] Zongyao He, Zhi Jin, and Yao Zhao. Srdrl: A blind superresolution framework with degradation reconstruction loss. IEEE Transactions on Multimedia, 24:2877-2889, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 547, + 348 + ], + "type": "text", + "content": "[17] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1575-1584, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[18] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 426 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 427, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 481 + ], + "type": "text", + "content": "[20] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 624-632, 2017. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 482, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 557 + ], + "type": "text", + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4681-4690, 2017. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 603 + ], + "type": "text", + "content": "[22] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 657 + ], + "type": "text", + "content": "[23] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwanggil Jeon, and Wei Wu. Feedback network for image superresolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3, 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26034" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 1, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 288, + 194 + ], + "type": "text", + "content": "[26] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 416-423, 2001. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 194, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 259 + ], + "type": "text", + "content": "[27] Ishit Mehta, Michael Gharbi, Connelly Barnes, Eli Shechtman, Ravi Ramamoorthi, and Manmohan Chandraker. Modulated periodic activations for generalizable local functional representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14214-14223, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 288, + 314 + ], + "type": "text", + "content": "[28] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 360 + ], + "type": "text", + "content": "[29] Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 414 + ], + "type": "text", + "content": "[30] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 415, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 468 + ], + "type": "text", + "content": "[31] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In European Conference on Computer Vision, pages 318-333. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 288, + 514 + ], + "type": "text", + "content": "[32] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 559 + ], + "type": "text", + "content": "[33] Ivan Skorokhodov, Savva Ignatyev, and Mohamed Elhoseiny. Adversarial generation of continuous images. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 10753-10764, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "text", + "content": "[34] Gaochao Song, Qian Sun, Luo Zhang, Ran Su, Jianfeng Shi, and Ying He. Ope-sr: Orthogonal position encoding for designing a parameter-free upsampling module in arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10009-10020, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[35] Yannick Strümpler, Janis Postels, Ren Yang, Luc Van Gool, and Federico Tombari. Implicit neural representations for image compression. In European Conference on Computer Vision, pages 74–91. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[36] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4539-4547, 2017. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 620 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 128 + ], + "type": "text", + "content": "[37] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4801-4810, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 184 + ], + "type": "text", + "content": "[38] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Hang Wang, Zhengyan Tong, and Yutian Liu. Deep arbitrary-scale image super-resolution via scale-equivariance pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1786-1795, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 546, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 228 + ], + "type": "text", + "content": "[39] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 546, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 546, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 546, + 274 + ], + "type": "text", + "content": "[40] Jingyu Yang, Sheng Shen, Huanjing Yue, and Kun Li. Implicit transformer network for screen content image continuous super-resolution. Advances in Neural Information Processing Systems, 34:13304-13315, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 274, + 546, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 546, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 546, + 329 + ], + "type": "text", + "content": "[41] Jie-En Yao, Li-Yuan Tsao, Yi-Chen Lo, Roy Tseng, Chia-Che Chang, and Chun-Yi Lee. Local implicit normalizing flow for arbitrary-scale image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1776–1785, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 330, + 546, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 330, + 546, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 330, + 546, + 374 + ], + "type": "text", + "content": "[42] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 546, + 419 + ], + "type": "text", + "content": "[43] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 419, + 546, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 419, + 546, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 419, + 546, + 474 + ], + "type": "text", + "content": "[44] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference*, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 546, + 519 + ], + "type": "text", + "content": "[45] Huanrong Zhang, Jie Xiao, and Zhi Jin. Multi-scale image super-resolution via a single extendable deep network. IEEE Journal of Selected Topics in Signal Processing, 15(2):253-263, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 520, + 546, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 520, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 307, + 520, + 546, + 564 + ], + "type": "text", + "content": "[46] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In European Conference on Computer Vision, pages 286-301, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 565, + 546, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 565, + 546, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 565, + 546, + 620 + ], + "type": "text", + "content": "[47] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2472-2481, 2018. 1, 3, 5" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26035" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_content_list.json b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79ebdf887f0c9f7f00491ff65ddbe55f7fee9687 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_content_list.json @@ -0,0 +1,1598 @@ +[ + { + "type": "text", + "text": "Layout-Agnostic Scene Text Image Synthesis with Diffusion Models", + "text_level": 1, + "bbox": [ + 145, + 130, + 825, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qilong Zhangli $^{1,2}$ Jindong Jiang $^{1}$ Di Liu $^{1}$ Licheng Yu $^{2}$ Xiaoliang Dai $^{2}$ Ankit Ramchandani $^{2}$ Guan Pang $^{2}$ Dimitris N. Metaxas $^{1}$ Praveen Krishnan $^{2}$ $^{1}$ Rutgers University $^{2}$ Meta AI", + "bbox": [ + 166, + 179, + 802, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f55640a5b3b582496b41378a992dbf9d9fc654a370869b4e2dde4fdee6165b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 256, + 215, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04a97ac60502742b86f90d22d7e63f0bd14f6dd79f490ed21c38912a2fa5f490.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 256, + 346, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f815a24adc7ed4dcaa374f82ee5cd965c15299301bcbd0012777ccfe495c50a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 257, + 480, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c5778c4ebca9ce0dffbdd641959d0a550b13806a2c5397459da8c948aa80141.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 256, + 614, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4dc5618bcc17b750a0d57b496e4b835b32867de9b80b4bdc928d059eb8461bc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 257, + 746, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d202a5ca24575dbc24364e29e7ba09716b64a7d004a07cc8258d4ba62a5dd8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 257, + 883, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7075f232c8a55425353c54295037708fff3632fddc4318ca47586f93b97d6236.jpg", + "image_caption": [ + "Figure 1. Scene Text Generation. Qualitative samples of scene text images generated by our model are presented. These images contain visually appealing texts that are coherent with the background and are created without relying on any spatial information or predefined layouts as input, thereby enhancing the Text-to-Image (T2I) Diffusion Model's capability to generate text." + ], + "image_footnote": [], + "bbox": [ + 81, + 391, + 212, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f02803c29233d644775d35b7dc4adf040bdc8fd6960441839d769a7fa4c8277.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 391, + 346, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7a9488ce58b78fc30f20d95a2843eabed3388becb69910d4f330d67ca9e5749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 391, + 482, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d2a70b1d257505d341b536e3d09ebf0d88c749f2cc936092cc8f0f689d109c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 391, + 614, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3150e268735cacda835a1a633cc0aef4315e0b6af240b566591e543f457d8cf.jpg", + "image_caption": [ + "'Time Traveling Device'" + ], + "image_footnote": [], + "bbox": [ + 616, + 391, + 748, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f89071be178b7d838071078cee11e331c3ffa4b6fb0b814c3eaf3fa56eb682e0.jpg", + "image_caption": [ + "in the foreground, with the text", + "'contemplate the clouds' in", + "roundedcursive" + ], + "image_footnote": [], + "bbox": [ + 750, + 391, + 883, + 513 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 608, + 313, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While diffusion models have significantly advanced the quality of image generation, their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts, an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges, this paper introduces SceneTextGen, a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so, SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties, coupled with a character-level instance segmentation model and a word", + "bbox": [ + 73, + 643, + 473, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods.", + "bbox": [ + 496, + 609, + 893, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 736, + 632, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The text-to-image [50, 57] task has gained popularity with advancements in diffusion models [41, 45, 49], significantly enhancing the quality of image generation. However, seamlessly integrating clear and contextually appropriate text into images is a persistent challenge. Text plays a vital role in many domains, including media content creation and artistic design, yet current diffusion models often struggle to produce text that is lexically correct, in valid font style, and that naturally complements the overall image aesthetics.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7496", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt: A vintage postage stamp showing a painting of mountains and the text 'California'.", + "bbox": [ + 84, + 93, + 454, + 117 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8027ef464ee97173c5a5da9e6b544bb2bbb1cea7f23407d24e06571c4dd84193.jpg", + "image_caption": [ + "(a) Layout-Constrained Models" + ], + "image_footnote": [], + "bbox": [ + 86, + 119, + 263, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/46d07de1bcafdcab222e6d427c6fc0e5f5f2a2db2b40c0b6c23becb3c334ca15.jpg", + "image_caption": [ + "(b) Layout-Agnostic SceneTextGen", + "Figure 2. Models reliant on predefined text layouts for input exhibit limitations such as constrained font diversity and static text positioning during each inference, leading to a lack of variability in style and arrangement." + ], + "image_footnote": [], + "bbox": [ + 279, + 118, + 459, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional methods [54, 58, 62] of creating scene text images typically formulate this problem as scene text editing which only involves editing or adding text to an existing scene image. These methods have challenges handling complex backgrounds, font styles and lighting variations. While recent models [4, 8, 33, 45, 49] have made strides in addressing these limitations by enhancing text encoding strategies [33] or employing predefined text layouts [8], they still face significant constraints in generating visual text. These constraints become apparent when observing the limited diversity in font styles and the static positioning of text, as shown in Fig. 2. Such rigidity in layout and font selection hampers the capacity of generative models to produce text that is stylistically varied and contextually aligned with the image content.", + "bbox": [ + 75, + 371, + 468, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose SceneTextGen, a novel framework that capitalizes on the capabilities of latent diffusion models to infuse text into scene images with greater diversity and authenticity. Our approach is specifically engineered to transcend the limitations of predefined layouts, enabling more flexible text placement and an expansive assortment of text styles.", + "bbox": [ + 75, + 599, + 468, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The two primary contributions of this work are: (i) the integration of character-level encoder to capture the typographic properties of visual text and carefully injecting it into the cross attention layers of the diffusion model, and (ii) introducing a novel word spotting loss using a pre-trained OCR along with a character segmentation loss to make the network more faithful in generating visual text. The character level encoder naturally blends with the existing diffusion model architecture and therefore allows the network to learn the layout of visual text implicitly rather than constraining itself to some pre-defined form. Our comprehensive evaluations confirm that SceneTextGen surpasses contemporary methods, facilitating the generation of images with text that", + "bbox": [ + 75, + 704, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "is both aesthetically pleasing and rich in variety.", + "bbox": [ + 498, + 90, + 816, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 121, + 640, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text 2 Image Generation Models", + "text_level": 1, + "bbox": [ + 500, + 146, + 785, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the realm of text-to-image generation, diffusion models [11, 16, 27, 37, 41, 44, 49, 63, 64] represent a significant leap forward. Different from GAN-based models [10, 17, 67], diffusion models employ a stochastic process that iteratively adds noise to an image and learn to reverse this process to generate images from textual descriptions. Their capacity to produce high-quality, detailed visual content from text prompts has been well documented. The Latent Diffusion Model (LDM)[42] further enhances this approach by operating in a compressed image latent space, improving both efficiency and image quality. It also facilitates high-quality conditional generation through cross-attention text conditioning, leading to various downstream applications [12, 13, 15, 19, 55]. DALL-E [41], renowned for its novel approach of combining discrete VAEs with transformer language models, has shown remarkable ability in generating diverse and complex images from textual descriptions, showcasing the potential of transformer architectures in creative generative tasks. Deepfloyd [49], utilizing the robust T5 text encoders [40], not only enhances image quality but also facilitates nuanced understanding of complex prompts, thereby allowing for more accurate and context-aware visual representations. ControlNet [63], has demonstrated exceptional performance in conditioned image generation by providing the model with references such as skeleton, canny edge images, and segmentation maps. However, these models frequently encounter difficulties when it comes to the generation of text within images (see Fig. 4), a task requiring the text to be not only visually integrated but also contextually pertinent to the image content. This challenge stems from the complexity of modeling the fine-grained interplay between visual and textual elements, ensuring that the generated text is legible, aesthetically fitting, and semantically in sync with the image. Prior attempts to refine this aspect have led to improvements [8, 49, 63], yet the generation of contextually coherent text in images remains a largely unsolved problem, underscoring the need for more focused research.", + "bbox": [ + 496, + 170, + 890, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Scene Text Generation", + "text_level": 1, + "bbox": [ + 500, + 755, + 709, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The success of adversarial networks such as GANs [7, 10, 17, 56, 67] for image generation and style transfer [20, 21] gave rise to scene text generation methods which can generate text at the granularity of glyphs [1, 26] or individual words [23, 54, 58]. Many previous works focused on the particular task of scene text editing where the model learns a style from a reference image and renders the target content in that style. Methods such as SRNet [54], SWAPText [58]", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7497", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "decomposes the problem into: (1) learning the foreground text using style and content, (2) background in-painting network to remove existing text, and (3) a blending network to merge foreground and background. These methods often fail in learning the correct style and removing existing text from complex backgrounds. StyleBrush [23] proposes a self-supervised approach to disentangle content and style and generate word images in a one-shot manner. All these previous methods are limited to generating individual words, requires reference word style images and does not generalize to generate the entire scene text image.", + "bbox": [ + 75, + 90, + 470, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial fusion GAN (SF-GAN) [62] generates text images by superimposing a foreground content image, transformed to match the style and geometry of a background image. This approach is aimed more at pure text synthesis on image regions without text, whereas in this work we aim to generate both image and text in a manner that reflects its natural appearance in real-world contexts.", + "bbox": [ + 75, + 261, + 468, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the significant progress of diffusion models in text-to-image generation, recent methods in scene text generation adapt these models towards producing more legible visual text. DIFFSTE [18] enhances scene text editing with a dual encoder design in diffusion models, promoting text legibility and style control. It is adept at mapping text instructions to images, showcasing zero-shot capabilities for rendering text in novel font variations and interpreting informal natural language instructions. This method is however a scene text editing method which generates single keywords on specific regions defined by a mask, whereas we propose a scene text generation network. One of the closest work in this space is TextDiffuser [8] which address the problem of scene text generation by decomposing the problem into two-stages. In the first stage, a transformer model create the layout mask for keywords extracted from text prompts. This is then taken as condition data while formulating the diffusion model to generate scene text images. They also introduce the character segmentation loss which helps in generating legible text. ControlNet [63], though not originally designed for scene text image rendering, has been effectively adapted for this purpose. It utilizes canny edge maps as conditional inputs, sourced from printed text images generated by a layout model, to fine-tune the diffusion model's output. The use of pre-defined layout for visual text generation in [8, 63], seriously limits the diversity of text styles, fonts and even the layouts. We believe this is due to the inherent difficulty in predicting layouts independently without the general image guidance. In our work, we avoid the need of pre-defined layout and make the network implicitly learn layout along with image generation using our novel way of injecting character level features. We also introduce a word spotting loss which augments the character segmentation loss proposed in [8] to generate more legible visual texts.", + "bbox": [ + 75, + 369, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Scene Text Recognition", + "text_level": 1, + "bbox": [ + 498, + 90, + 714, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advances in computer vision have laid a foundation for sophisticated analytical techniques in various domains [2, 9, 14, 30-32, 36, 46, 53, 60, 61, 65]. Especially in scene text image recognition, most existing works split the process into two stages: a text detection [3, 28, 29, 34, 48, 66] module to detect words or characters from complex backgrounds, and a text recognition [2, 24, 47, 52] module which transcribes the text into unicode characters given a cropped word image. More recently, end-2-end methods [6, 25, 38, 43, 51] have become popular due to the benefits of joint training of detector and recognizer to share contextual information.", + "bbox": [ + 496, + 113, + 890, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the nexus of scene text generation and recognition, leveraging pre-trained scene text recognition or word spotting models as guidance during diffusion-based text-to-image synthesis has surfaced as an innovative strategy. For simplicity, in this paper we refer to a scene text recognition module as OCR (optical character recognition). The integration of OCR-derived losses enables the refinement of generative models to produce text that is not just visually coherent but also contextually accurate. This confluence of generative modeling prowess with OCR accuracy paves the way for novel research avenues to generate images with text that is both authentic to read and visually integrated.", + "bbox": [ + 496, + 295, + 892, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 489, + 633, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 500, + 513, + 622, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to facilitate layout-free text image generation with diverse layouts and styles. A straightforward approach would be to directly fine-tune an existing latent diffusion model with text images. However, our early investigation suggests that this strategy did not yield significant improvements compared to the original LDM. We hypothesize that this limitation is due to two primary factors. First, the language encoder of the LDM, primarily designed for semantic interpretation, fails to capture character-specific information adequately for text rendering. This encoder tends to provide more semantic than structural information about the text, necessitating a dedicated network for encoding the text and guiding the model on its visual representation. Second, the conventional denoising loss used in diffusion models seems insufficient for accurately rendering text in images, often leading to text regions resembling textual patterns without the distinct features of text strokes. To address these challenges, we propose integrating a character-level encoder and a hierarchical cross-attention mechanism to learn character-level context information. Additionally, we introduce two auxiliary losses at the word and character levels to emphasize the text presentation. The subsequent sections will detail each of these components and their contribution to enhancing our model's performance.", + "bbox": [ + 496, + 537, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7498", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f58c1f1343c61e708eb975832790db88f5c42350b11e93ff3b6ee2090278e56e.jpg", + "image_caption": [ + "Figure 3. Model Framework: SceneTextGen employs a character-level encoder to extract detailed character-specific features. During loss computation, the model leverages both word-level and character-level supervisions to guide the recovery of the image, in addition to the standard denoising loss. This dual-level supervision enhances the model's ability to accurately generate and refine text within scenes." + ], + "image_footnote": [], + "bbox": [ + 78, + 93, + 883, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 385, + 217, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed method is trained on a corpus of visual text images. Given an image $\\mathcal{I}$ with textual caption $c$ , we denote $w_{i}$ , where $i \\in [0, N]$ , as the visual text (words) present in this image. We also assume that for each word, we know its location, given in terms of the bounding box. For all practical purposes, we assume this can be pre-computed using a pre-trained OCR detection [28] and recognition network [5]. Please note, the OCR bounding box and transcriptions are typically noisy. Fig. 3 presents an overview of our proposed method. A latent diffusion model has three key components: a CLIP encoder for semantic embedding of text and images, an autoencoder for dimensionality reduction and feature extraction, and a UNet-based structure for effective image-text synthesis and manipulation. The core architecture of our method follows the latent diffusion model, which is adapted to incorporate the proposed character-level features and content-based loss functions.", + "bbox": [ + 75, + 409, + 473, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Character-Level Encoding", + "text_level": 1, + "bbox": [ + 76, + 679, + 316, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach begins by extracting and ordering the visual text words from the image based on its locations as given by the bounding box. We sort the word boxes following the conventional reading pattern, i.e., from left to right and top to bottom. This provides us the naive ordering of visual text which is then tokenized at the character level to capture its typographic information. Then, a character encoder is used to encode these tokenized characters into a high-dimensional feature space, allowing an accurate transcription of the text's spelling and appearance. The derived text features encapsulate the typographical details and are poised for the subsequent integration with the image's latent features, which have already been contextually primed", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by the initial cross-attention with the encoded caption features obtained from the CLIP [39] text encoder.", + "bbox": [ + 498, + 386, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Hierarchical Text Integration Process", + "text_level": 1, + "bbox": [ + 498, + 429, + 823, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have employed a sequential cross-attention mechanism to integrate character information into the U-Net architecture effectively. This method is inspired by the observation that the cross-attention of captioning features is pivotal in shaping the overall spatial structure of an image, a notion supported by prior studies [15]. Leveraging this concept, our model initially constructs the general spatial layout of the image content, guided by the caption, using the first cross-attention layer. It then focuses on the precise rendering of characters in a subsequent cross-attention phase. This approach establishes a hierarchical text integration process, facilitating the development of a preliminary visual scaffold that is both thematically and contextually coherent. It ensures that the textual elements are accurately positioned and seamlessly integrated into the overall image structure.", + "bbox": [ + 496, + 454, + 892, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. OCR-Guided Diffusion for Text Accuracy", + "text_level": 1, + "bbox": [ + 498, + 693, + 856, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure the textual accuracy of generated images, our model incorporates an OCR loss, utilizing the predictions from the end to end pre-trained GLASS model [43]. Upon each iteration of the diffusion process, the UNet predicts a denoising step, from which we derive an estimation of clean image $x_0$ . This estimated $x_0$ is then decoded through a Variational Autoencoder (VAE) [22] to reconstruct an image.", + "bbox": [ + 496, + 718, + 890, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The GLASS OCR model performs inference on this reconstructed image to produce word-level recognition results, which are represented as a tensor $P$ with shape $[N,L,K]$ , where $N$ is the number of words detected, $L$ is the maximum length of any word, and $K$ is the size of the", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7499", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "character set. The ground truth for these detections is represented as a tensor $G$ with shape $[N, L]$ , where each entry is the character index for the corresponding position, and non-character positions are marked with zero.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The OCR loss $(\\mathcal{L}_{\\mathrm{OCR}})$ is computed using a masked cross-entropy function, which is formulated as follows:", + "bbox": [ + 76, + 151, + 468, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O C R}} = - \\frac {1}{\\sum M} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i j} \\cdot \\log \\left(\\frac {\\exp \\left(P _ {i j} [ G _ {i j} ]\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(P _ {i j} [ k ]\\right)}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 200, + 467, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $M$ is a binary mask tensor that has the same shape as $G$ and indicates the valid character positions (i.e., where $G_{ij} \\neq 0$ ). $M_{ij}$ represents the binary value of the mask at the $i$ -th word and $j$ -th character position, $P_{ij}$ is the predicted probability distribution over the character set, and $G_{ij}$ is the ground truth character index at that position.", + "bbox": [ + 76, + 257, + 468, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating this OCR loss into the training regime, we guide the diffusion model to produce text that is not only visually coherent but also textually accurate, as recognized by the GLASS [43] OCR model, thereby enhancing the overall fidelity of the generated images.", + "bbox": [ + 76, + 347, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Refinement of Text Generation with Character-Level Constraints", + "text_level": 1, + "bbox": [ + 76, + 431, + 468, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the iterative refinement of our diffusion model, we observed an unintended consequence of the OCR loss; the model tended to generate images with repetitive words. This issue is potentially attributed to the blurry nature of images at higher noise levels during training, which could render the OCR loss counterproductive. The word-level OCR loss, while ensuring textual accuracy, imposes no explicit constraint on the quantity of text within the image, inadvertently encouraging the model to generate excessive text.", + "bbox": [ + 75, + 469, + 468, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we augmented our loss function with a character-level segmentation loss, which acts directly on the latent space rather than the recovered image. After obtaining the predicted latent features of a image $x_0$ from the UNet, we proceed in two directions: we decode $x_0$ using the VAE to compute the word-level OCR loss on the recovered image (as explained earlier), and we also apply $x_0$ to a pre-trained character-level segmentation model based on U-Net adapted from [8]. This model outputs a 96-dimensional feature map (corresponding to the length of the alphabet plus one for non-character pixels) with a spatial resolution of $64 \\times 64$ . The character-aware loss is then computed via cross-entropy between this feature map and a resized character-level segmentation mask $C$ .", + "bbox": [ + 75, + 604, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, the total loss function is a composite of the denoising loss, the word-level recognition loss, and the character-level segmentation loss, expressed as:", + "bbox": [ + 76, + 816, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s i n g}} + \\lambda_ {\\text {w o r d}} \\mathcal {L} _ {\\text {O C R - w o r d}} + \\lambda_ {\\text {c h a r}} \\mathcal {L} _ {\\text {O C R - c h a r}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{denoising}}$ is the denoising loss, $\\mathcal{L}_{\\mathrm{OCR - word}}$ is the word-level OCR loss, $\\mathcal{L}_{\\mathrm{OCR - char}}$ is the character-level segmentation loss, and $\\lambda_{\\mathrm{word}}$ and $\\lambda_{\\mathrm{char}}$ are weighting coefficients balancing the contribution of each term.", + "bbox": [ + 498, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The character-level segmentation loss is formulated as:", + "bbox": [ + 517, + 152, + 880, + 166 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {O C R - c h a r}} = - \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {9 6} y _ {h w c} \\log \\left(\\hat {y} _ {h w c}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 190, + 890, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ and $W$ are the height and width of the feature map, $y$ is the ground truth character segmentation mask, and $\\hat{y}$ is the predicted character probability map from the segmentation model.", + "bbox": [ + 498, + 242, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By integrating character-level information directly in the latent space, we impose a structured constraint on text generation, promoting both the accuracy and the appropriate quantity of text in the generated images.", + "bbox": [ + 498, + 303, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 377, + 630, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 401, + 715, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets For the training of our model, we utilized the publicly available MARIO dataset from [8], excluding MARIO-TMDB, and MARIO-OpenLibrary subsets as they are not publicly accessible. Upon removing any corrupted images, our final dataset comprised 7,249,449 image-caption pairs. In addition, to bolster our model's capability in generating a broader range of concepts (beyond text-centric images), we integrated 2,110,745 non-text images. These additional images, accompanied by text pairs, were selected based on a minimum predicted aesthetics score of 6.25, allowing for joint training to enhance overall performance.", + "bbox": [ + 498, + 425, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We conducted quantitative comparisons of our SceneTextGen method against several leading approaches, including LDM[42], ControlNet[63], TextDiffuser[8], GlyphControl[59], and DeepFloyd[49], utilizing the publicly available code and pre-trained models for fairness. Notably, DeepFloyd is distinguished by its dual superresolution modules, enabling it to produce high-resolution images at 1024x1024 pixels, in contrast to the 512x512 pixel images generated by the other models. For ControlNet comparisons, we employed Canny edge maps of printed text images created by the initial model stage of TextDiffuser as conditioning inputs. However, due to the unavailability of APIs, open-source code or checkpoints, we could not extend our comparative analysis to includeImagen[45], eDiff-i[4], or GlyphDraw[35].", + "bbox": [ + 498, + 625, + 890, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Criteria We assess text rendering quality using the MARIO-7M-Eval, MARIO-TMDB-Eval, and", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MARIO-OpenLibrary-Eval datasets. Our evaluation is twofold: firstly, through CLIPScore, which measures the cosine similarity between the image and text representations from CLIP; and secondly, via OCR Evaluation, which leverages existing OCR tools to detect and recognize text regions within the generated images. Metrics such as Average Precision, Average Recall, F1 Score, and Accuracy are employed to determine the presence of keywords in the generated images. During training, the input of the character encoder is the ground truth OCR texts. During the inference process, the character encoder receives as its input the text from captions provided by the user. These captions are enclosed in quotation marks as specified in the user's prompts, and the input of the CLIP text encoder is the full caption. For each generated and ground truth image pair, we utilize the easy-ocr library for OCR detection and recognition, followed by Hungarian Matching between the sets of texts, applying the Levenshtein distance (or edit distance) on the matched text pairs for the OCR evaluation.", + "bbox": [ + 75, + 90, + 472, + 380 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 T2I Visual Text Performance Evaluation" + ], + "code_body": "function METRICS(M) \nP ← PRECISION(M) \nR ← RECALL(M) \nF ← F1(P, R) \nA ← ACCURACY(M) \nreturn {P, R, F, A} \nend function \n1: GT ← OCR(Ground Truth Images) \n2: Gen ← OCR(Generated Images) \n3: Scores ← an empty list \n4: for each pair in ZIP(GT, Gen) do \n5: C ← COSTMAT(pair) \n6: M ← HUNGARIAN(C) \n7: Scores ← Scores ∪ {METRICS(M)} \n8: end for \n9: return AVERAGE(Scores)", + "bbox": [ + 81, + 410, + 472, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pseudo Code for OCR Performance Evaluation The methodology described in Algorithm 1 demonstrates the steps taken to assess the performance of the text-to-image conversion. In the initial step, the algorithm applies OCR to both the ground truth and the generated images. This process results in two sets of text outputs, which are then paired for comparison. The Hungarian algorithm is employed here to find the optimal matching between elements (words) of these two sets, minimizing the overall difference between the matched pairs. This is crucial for an objective and accurate comparison. For each matched pair, we calculate a cost matrix, which serves as the input for the Hungarian algorithm. The output of this step is a matching matrix $M$ which represents the best possible alignment between the text elements in the ground truth and the generated images. Sub", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sequentially, the algorithm computes key performance metrics for each pair: Precision, Recall, F1 Score, and Accuracy. Precision focuses on the accuracy of the replicated text, recall measures the completeness, F1 score provides a balance between precision and recall, and accuracy gives an overall effectiveness of the text replication. Finally, the algorithm averages these scores across all image pairs to provide an overall performance evaluation of the text-to-image conversion process.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 498, + 239, + 694, + 255 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experimental analysis in Tab. 1 provides a direct comparison of the OCR based recognition scores among various models as measured in terms of Precision, Recall, F1 scores, and Accuracy. Our results indicate that SceneTextGen consistently outperforms competing models in most metrics. Latent Diffusion Model, lacking a sophisticated mechanism for text comprehension, typically under-performs, leading to lower OCR scores. In contrast, DeepFloyd[49] incorporates a T5 encoder which aids in textual understanding, thereby enhancing the quality of the generated text. However, its performance is still limited due to an insufficient character-level understanding.", + "bbox": [ + 496, + 263, + 890, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ControlNet[63], TextDiffuser[8], and GlyphControl[59], which utilize predefined text layouts or spatial information, show mixed results. While the explicit introduction of text information allows ControlNet to achieve high OCR scores, the resulting text often appears artificial and lacks seamless integration within the images (see Fig. 4).", + "bbox": [ + 496, + 446, + 893, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-dataset Generalization Ability As demonstrated in Tab. 1, SceneTextGen-7M, despite being trained solely on the MARIO-7M dataset, exhibits strong generalization capabilities. It maintains robust OCR scores across evaluation sets from both the TMDB and OpenLibrary datasets, underscoring its adaptability and the efficacy of its training methodology.", + "bbox": [ + 496, + 560, + 893, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Measuring Font Style Diversity", + "text_level": 1, + "bbox": [ + 498, + 678, + 774, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To quantitatively assess the diversity of font styles generated by SceneTextGen, we utilized a pretrained VGG-based font recognition model trained on synthesized text images. This approach involved first extracting text image patches using a pretrained Optical Character Recognition (OCR) model. These patches were then processed through the font recognition model to retrieve features from the penultimate layer. By applying t-SNE for dimensionality reduction, we visualized the feature space to examine the proximity and diversity of text generated by different methods.", + "bbox": [ + 496, + 703, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As illustrated in Fig. 5, the rendered text images — which serve as a baseline — were created by printing text onto a white canvas using the layout generator from [8] with", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7501", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e4bf3f722347d1a5ddb8193a9789d1068b91d664880ae0e6b7adf7d083dae5ad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-defined Text LayoutMARIO-7MTMDBOpenLibrary
AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)
TextDiffuser-10M [8]✓0.61350.42890.46830.34250.44010.42430.39940.28890.56170.38160.42420.2916
GlyphControl-10M* [59]✓0.50750.41180.41400.28830.36350.40820.34570.23620.47340.37620.38360.2534
Latent Diffusion Model0.14820.16900.12960.07530.17170.25790.17030.09740.19110.28730.19230.1106
DeepFloyd [49]0.24670.27880.22060.13660.22840.37380.24490.15000.25550.39100.26350.1610
ControlNet [63]✓0.51020.44440.42380.29810.30750.46670.32840.21940.40500.42730.36900.2415
TextDiffuser-7M [8]✓0.47780.34470.36820.25120.31980.33620.29610.19300.42570.31460.33180.2108
SceneTextGen-7M(Ours)0.52740.44200.44240.30880.38130.47160.37900.26020.41360.45190.39450.2571
", + "bbox": [ + 78, + 88, + 893, + 222 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparative analysis of OCR based text recognition scores across different models. Note, 7M denotes models that were trained on the MARIO-7M dataset (7 million images with texts). In contrast, the TextDiffuser-10M category includes models trained on an expanded dataset collection that encompasses MARIO-7M, MARIO-TMDB, and MARIO-OpenLibrary. * denotes LAION-Glyph dataset.", + "bbox": [ + 75, + 233, + 893, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80028703f90f109493fb0f4762b77da7e6e6d8e7b018270029d0d3fd0288ba95.jpg", + "image_caption": [ + "Figure 4. Comparative visualization of generated images. We present a side-by-side comparison of images generated from the same text prompt across different existing methods (with results generated by [8] as a proxy. Human faces are blurred for ethical considerations.). Each row corresponds to a unique prompt, showcasing the visual quality, text clarity, and contextual coherence achieved by each method." + ], + "image_footnote": [], + "bbox": [ + 117, + 292, + 846, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the default 'Arial' font. This process establishes a reference for the feature distribution of 'Artificial Texts.' ControlNet [63], constrained by predefined text canny edges, exhibits a feature distribution closely aligned with the rendered text images. This proximity suggests its limitations in integrating text seamlessly within images, a finding corroborated by the qualitative results shown in Fig. 4. TextDiffuser [8], although performing well in most instances, still shares some feature space with ControlNet, indicating occasional production of artificial or unnatural texts. In contrast, SceneTextGen — operating independently of any predefined canny edges or layouts — demonstrates a distinct distribution with minimal overlap with the 'rendered text", + "bbox": [ + 75, + 702, + 472, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images,' signifying its robustness in generating naturalistic text within the context of scene images. Example visualizations are in Fig. 6. In addition, to better understand the text layout given by each model, we also show in Fig. 7 the overall distribution of visual text in the generated images and ground truth images.", + "bbox": [ + 496, + 702, + 893, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Weight of Each Loss", + "text_level": 1, + "bbox": [ + 498, + 801, + 694, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present two ablation studies to evaluate the impact of different loss function configurations on OCR performance. Table 3 explores the effects of using combinations of word-level and character-level losses. Table 4 examines the impact of varying the weights of these loss func", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7502", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2efd9dacf5575ecf1e01b44aeb9c3614741854d46f131a561664ac26547126a2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-defined \nText LayoutCLIP Score(↑)
TextDiffuser-10M [8]✓0.3436
GlyphControl-10M* [59]✓0.3450
Latent Diffusion Model0.3015
DeepFloyd [49]0.3267
ControlNet [63]✓0.3424
TextDiffuser-7M [8]✓0.3385
SceneTextGen-7M(Ours)0.3455
", + "bbox": [ + 114, + 88, + 431, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7831a1c1e096e1d49e1b4c36fc750f3461a00c47479e296deeec5f251a9ccdbe.jpg", + "image_caption": [ + "Figure 5. t-SNE representation of text region embeddings derived from the penultimate layer features of a font recognition model." + ], + "image_footnote": [], + "bbox": [ + 117, + 292, + 429, + 454 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ef7b853fb4ee53bc1682a40646e8d09b64a9425fd4ffdb9d884730acfa2b6a7.jpg", + "image_caption": [ + "Caption: A sign that says 'Hello World'", + "Figure 6. Diversity of font styles. SceneTextGen is able to generate visually appealing and diverse font styles and layouts for text, without any style or layout prompts." + ], + "image_footnote": [], + "bbox": [ + 78, + 507, + 467, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions. In both studies, we assess the Average Precision (AP) and Accuracy (AC) scores to understand how these configurations influence the model's ability to accurately recognize and generate text in images.", + "bbox": [ + 75, + 729, + 468, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 76, + 800, + 204, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While SceneTextGen demonstrates superior performance in scene text image generation, challenges remain. Limited scene complexity: Models struggle to accurately generate complex elements in conjunction with text, such as a car with a 'green' sticker specifically in the back win", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41803e143ea7eb1dc0edc55fd7980a2b6c70bcd13e44856cc39fad113dccd7d7.jpg", + "image_caption": [ + "TextDiffuser" + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 624, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e545863a265564e7b10efca04de74032c86f0a9af1f22c7da2d20da0d9abb54.jpg", + "image_caption": [ + "SceneTextGen", + "Figure 7. Distribution of visual text coordinates in the generated image w.r.t. the area of each box. As expected from real-life observations, the visual text in the ground truth images tends to be located at the center of each image." + ], + "image_footnote": [], + "bbox": [ + 624, + 90, + 741, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c80c44c344aa77d47da102c5fb105c22f02ab4e454559c6b660bdb6a81aefb6.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 741, + 90, + 859, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3236ad3b3229c97911e4f8e31574c9ef420632755eca427c7c3ce3119e4824dd.jpg", + "table_caption": [ + "Table 2. Comparison of CLIP scores reflecting the overall image quality generated by various models. SceneTextGen-7M achieves the highest CLIP score, indicating superior text-image alignment without relying on text layouts. * denotes LAION-Glyph dataset." + ], + "table_footnote": [], + "table_body": "
LwordLcharAPAC
✓0.50820.3109
✓0.50830.3035
✓✓0.52740.3088
", + "bbox": [ + 517, + 275, + 666, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/39bd0e56f406fca446442d98439d11d92adf5c8066099ff626bfeb16d7545f7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WlamClamAPAC
120.41070.2782
0.10.20.45870.2895
0.010.020.52740.3088
", + "bbox": [ + 728, + 275, + 872, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Effects of loss function combinations on the MARIO-7M-Eval dataset", + "bbox": [ + 500, + 339, + 683, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Impact of varying loss weights on the MARIO-7M-eval dataset", + "bbox": [ + 707, + 339, + 890, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3bc893c50e3f640c1674a55b5cc69176eaf23d13967ea125f84e43565c60239.jpg", + "image_caption": [ + "Figure 8. Failure cases in complex scene interpretation and processing of lengthy prompts for text." + ], + "image_footnote": [], + "bbox": [ + 506, + 397, + 880, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dow.(Fig. 8, Case 1). Long text handling: Text accuracy and coherence decrease with increasing length (Fig. 8, Case 2). These areas require further development.", + "bbox": [ + 500, + 612, + 890, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 671, + 617, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SceneTextGen, incorporating a character-level encoder and hierarchical text integration, offers advancements in scene text image generation. Despite improved text rendering, limitations arise in generating complex visuals and handling lengthy text. These challenges highlight the ongoing difficulty in reconciling textual accuracy with broader image synthesis. This work furthers our understanding of text-image generation, paving the way for future exploration.", + "bbox": [ + 496, + 696, + 890, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417.", + "bbox": [ + 496, + 835, + 892, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7503", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 2", + "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3", + "[3] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9365-9374, 2019. 3", + "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 5", + "[5] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 4", + "[6] Michal Bušta, Yash Patel, and Jiri Matas. E2e-plt-an unconstrained end-to-end method for multi-language scene text. In Computer Vision-ACCV 2018 Workshops: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers 14, pages 127–143. Springer, 2019. 3", + "[7] Qi Chang, Zhennan Yan, Mu Zhou, Di Liu, Khalid Sawalha, Meng Ye, Qilong Zhangli, Mikael Kanski, Subhi Al'Aref, Leon Axel, et al. Deeprecon: Joint 2d cardiac segmentation and 3d volume reconstruction via a structure-specific generative method. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 567-577. Springer, 2022. 2", + "[8] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 2, 3, 5, 6, 7, 8", + "[9] Yunhe Gao, Mu Zhou, Di Liu, Zhennan Yan, Shaoting Zhang, and Dimitris N Metaxas. A data-scalable transformer for medical image segmentation: architecture, model efficiency, and benchmark. arXiv preprint arXiv:2203.00131, 2022. 3", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2", + "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4291-4301, 2024. 2", + "[14] Xiaoxiao He, Chaowei Tan, Bo Liu, Liping Si, Weiwu Yao, Liang Zhao, Di Liu, Qilong Zhangli, Qi Chang, Kang Li, et al. Dealing with heterogeneous 3d mr knee images: A federated few-shot learning method with dual knowledge distillation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 3", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 2, 4", + "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2", + "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[18] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 3", + "[19] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Advances in Neural Information Processing Systems, pages 8563-8601, 2023. 2", + "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2", + "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4", + "[23] Praveen Krishnan, Rama Kovvuri, Guan Pang, Boris Vassilev, and Tal Hassner. Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3", + "[24] Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention modeling forOCR in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2231-2239, 2016. 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7504", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13094–13102, 2023. 3", + "[26] Wei Li, Yongxing He, Yanwei Qi, Zejian Li, and Yongchuan Tang. Fet-gan: Font and effect transfer via k-shot adaptive instance normalization. In Proceedings of the AAAI conference on artificial intelligence, pages 1717-1724, 2020. 2", + "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2", + "[28] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In Proceedings of the AAAI conference on artificial intelligence, pages 11474-11481, 2020. 3, 4", + "[29] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 3", + "[30] Di Liu, Yunhe Gao, Qilong Zhangli, Ligong Han, Xiaoxiao He, Zhaoyang Xia, Song Wen, Qi Chang, Zhennan Yan, Mu Zhou, et al. Transfusion: multi-view divergent fusion for medical image segmentation with transformers. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 485-495. Springer, 2022. 3", + "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023.", + "[32] Di Liu, Anastasis Stathopoulos, Qilong Zhangli, Yunhe Gao, and Dimitris Metaxas. Lepard: Learning explicit part discovery for 3d articulated shape reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[33] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022. 2", + "[34] Jianqi Ma, Weiyuan Shao, Hao Ye, Li Wang, Hong Wang, Yingbin Zheng, and Xiangyang Xue. Arbitrary-oriented scene text detection via rotation proposals. IEEE transactions on multimedia, 20(11):3111-3122, 2018. 3", + "[35] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coher-ently. arXiv preprint arXiv:2303.17870, 2023. 5", + "[36] Carlos Martin-Isla, Víctor M Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J Fulton, Tewodros Weldebirhan Arega, Kumaradevan Punithakumar, et al. Deep learning segmentation of the right ventricle in cardiac mri: The" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "m&ms challenge. IEEE Journal of Biomedical and Health Informatics, 2023. 3", + "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[38] Siyang Qin, Alessandro Bissacco, Michalis Raptis, Yasuhisa Fujii, and Ying Xiao. Towards unconstrained end-to-end text spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4704-4714, 2019. 3", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2", + "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2", + "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5", + "[43] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In European Conference on Computer Vision, pages 249-266. Springer, 2022. 3, 4, 5", + "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2", + "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2, 5", + "[46] Lohrasb Ross Sayadi, Usama S Hamdan, Qilong Zhangli, and Raj M Vyas. Harnessing the power of artificial intelligence to teach cleft lip surgery. Plastic and Reconstructive Surgery-Global Open, 10(7):e4451, 2022. 3", + "[47] Baoguang Shi, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4168-4176, 2016. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7505", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3", + "[49] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd-if. https://github.com/deep-floyd/IF, 2023. GitHub Repository. 1, 2, 5, 6, 7, 8", + "[50] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiaoyuan Jing, Fei Wu, and Bingkun Bao. Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 1", + "[51] Peng Wang, Hui Li, and Chunhua Shen. Towards end-to-end text spotting in natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7266-7281, 2021. 3", + "[52] Peng Wang, Cheng Da, and Cong Yao. Multi-granularity prediction for scene text recognition. In European Conference on Computer Vision, pages 339-355. Springer, 2022. 3", + "[53] Song Wen, Hao Wang, Di Liu, Qilong Zhangli, and Dimitris Metaxas. Second-order graph odes for multi-agent trajectory forecasting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5101-5110, 2024. 3", + "[54] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In Proceedings of the 27th ACM international conference on multimedia, pages 1500-1508, 2019. 2", + "[55] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotpdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2", + "[56] Zhaoyang Xia, Yuxiao Chen, Qilong Zhangli, Matt Huener-fauth, Carol Neidle, and Dimitris Metaxas. Sign language video anonymization. In Proceedings of the LREC2022 10th Workshop on the Representation and Processing of Sign Languages: Multilingual Sign Language Resources, Marseille, France, 25 June 2022, 2022. 2", + "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1", + "[58] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14700-14709, 2020. 2", + "[59] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7, 8", + "[60] Jiachi Ye, Haoyan Kang, Hao Wang, Chen Shen, Belal Jahannia, Elham Heidari, Nvid Asadizanjani, Mohammad-Ali Miri, Volker J Sorger, and Hamed Dalir. Demultiplexing oam" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "beams via fourier optical convolutional neural network. In Laser Beam Shaping XXIII, pages 16-33. SPIE, 2023. 3", + "[61] Jiachi Ye, Maria Solyanik, Zibo Hu, Hamed Dalir, Behrouz Movahhed Nouri, and Volker J Sorger. Free-space optical multiplexed orbital angular momentum beam identification system using fourier optical convolutional layer based on 4f system. In Complex Light and Optical Forces XVII, pages 70-80. SPIE, 2023. 3", + "[62] Fangneng Zhan, Hongyuan Zhu, and Shijian Lu. Spatial fusion gan for image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3653-3662, 2019. 2, 3", + "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 5, 6, 7, 8", + "[64] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2", + "[65] Qilong Zhangli, Jingru Yi, Di Liu, Xiaoxiao He, Zhaoyang Xia, Qi Chang, Ligong Han, Yunhe Gao, Song Wen, Haiming Tang, et al. Region proposal rectification towards robust instance segmentation of biological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2022. 3", + "[66] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 3", + "[67] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 612 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7506", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_model.json b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..82b0fd9460070529a943c26291a6b4f3344559d6 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.826, + 0.154 + ], + "angle": 0, + "content": "Layout-Agnostic Scene Text Image Synthesis with Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.18, + 0.803, + 0.235 + ], + "angle": 0, + "content": "Qilong Zhangli\\(^{1,2}\\) Jindong Jiang\\(^{1}\\) Di Liu\\(^{1}\\) Licheng Yu\\(^{2}\\) Xiaoliang Dai\\(^{2}\\) Ankit Ramchandani\\(^{2}\\) Guan Pang\\(^{2}\\) Dimitris N. Metaxas\\(^{1}\\) Praveen Krishnan\\(^{2}\\) \\(^{1}\\)Rutgers University \\(^{2}\\)Meta AI" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.257, + 0.216, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.257, + 0.348, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.258, + 0.482, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.257, + 0.616, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.258, + 0.748, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.258, + 0.884, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.392, + 0.214, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.392, + 0.348, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.392, + 0.483, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.392, + 0.616, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.392, + 0.749, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.392, + 0.884, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.515, + 0.735, + 0.523 + ], + "angle": 0, + "content": "'Time Traveling Device'" + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.515, + 0.882, + 0.523 + ], + "angle": 0, + "content": "in the foreground, with the text" + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.524, + 0.873, + 0.531 + ], + "angle": 0, + "content": "'contemplate the clouds' in" + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.533, + 0.851, + 0.542 + ], + "angle": 0, + "content": "roundedcursive" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.554, + 0.895, + 0.598 + ], + "angle": 0, + "content": "Figure 1. Scene Text Generation. Qualitative samples of scene text images generated by our model are presented. These images contain visually appealing texts that are coherent with the background and are created without relying on any spatial information or predefined layouts as input, thereby enhancing the Text-to-Image (T2I) Diffusion Model's capability to generate text." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.609, + 0.314, + 0.625 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.644, + 0.474, + 0.901 + ], + "angle": 0, + "content": "While diffusion models have significantly advanced the quality of image generation, their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts, an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges, this paper introduces SceneTextGen, a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so, SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties, coupled with a character-level instance segmentation model and a word" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.703 + ], + "angle": 0, + "content": "level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.737, + 0.633, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The text-to-image [50, 57] task has gained popularity with advancements in diffusion models [41, 45, 49], significantly enhancing the quality of image generation. However, seamlessly integrating clear and contextually appropriate text into images is a persistent challenge. Text plays a vital role in many domains, including media content creation and artistic design, yet current diffusion models often struggle to produce text that is lexically correct, in valid font style, and that naturally complements the overall image aesthetics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7496" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.086, + 0.094, + 0.455, + 0.118 + ], + "angle": 0, + "content": "Prompt: A vintage postage stamp showing a painting of mountains and the text 'California'." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.12, + 0.264, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.26, + 0.255, + 0.272 + ], + "angle": 0, + "content": "(a) Layout-Constrained Models" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.119, + 0.46, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.26, + 0.46, + 0.273 + ], + "angle": 0, + "content": "(b) Layout-Agnostic SceneTextGen" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Models reliant on predefined text layouts for input exhibit limitations such as constrained font diversity and static text positioning during each inference, leading to a lack of variability in style and arrangement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.599 + ], + "angle": 0, + "content": "Traditional methods [54, 58, 62] of creating scene text images typically formulate this problem as scene text editing which only involves editing or adding text to an existing scene image. These methods have challenges handling complex backgrounds, font styles and lighting variations. While recent models [4, 8, 33, 45, 49] have made strides in addressing these limitations by enhancing text encoding strategies [33] or employing predefined text layouts [8], they still face significant constraints in generating visual text. These constraints become apparent when observing the limited diversity in font styles and the static positioning of text, as shown in Fig. 2. Such rigidity in layout and font selection hampers the capacity of generative models to produce text that is stylistically varied and contextually aligned with the image content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.6, + 0.469, + 0.705 + ], + "angle": 0, + "content": "To address these issues, we propose SceneTextGen, a novel framework that capitalizes on the capabilities of latent diffusion models to infuse text into scene images with greater diversity and authenticity. Our approach is specifically engineered to transcend the limitations of predefined layouts, enabling more flexible text placement and an expansive assortment of text styles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The two primary contributions of this work are: (i) the integration of character-level encoder to capture the typographic properties of visual text and carefully injecting it into the cross attention layers of the diffusion model, and (ii) introducing a novel word spotting loss using a pre-trained OCR along with a character segmentation loss to make the network more faithful in generating visual text. The character level encoder naturally blends with the existing diffusion model architecture and therefore allows the network to learn the layout of visual text implicitly rather than constraining itself to some pre-defined form. Our comprehensive evaluations confirm that SceneTextGen surpasses contemporary methods, facilitating the generation of images with text that" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.818, + 0.108 + ], + "angle": 0, + "content": "is both aesthetically pleasing and rich in variety." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.122, + 0.642, + 0.137 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.147, + 0.787, + 0.164 + ], + "angle": 0, + "content": "2.1. Text 2 Image Generation Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.171, + 0.892, + 0.744 + ], + "angle": 0, + "content": "In the realm of text-to-image generation, diffusion models [11, 16, 27, 37, 41, 44, 49, 63, 64] represent a significant leap forward. Different from GAN-based models [10, 17, 67], diffusion models employ a stochastic process that iteratively adds noise to an image and learn to reverse this process to generate images from textual descriptions. Their capacity to produce high-quality, detailed visual content from text prompts has been well documented. The Latent Diffusion Model (LDM)[42] further enhances this approach by operating in a compressed image latent space, improving both efficiency and image quality. It also facilitates high-quality conditional generation through cross-attention text conditioning, leading to various downstream applications [12, 13, 15, 19, 55]. DALL-E [41], renowned for its novel approach of combining discrete VAEs with transformer language models, has shown remarkable ability in generating diverse and complex images from textual descriptions, showcasing the potential of transformer architectures in creative generative tasks. Deepfloyd [49], utilizing the robust T5 text encoders [40], not only enhances image quality but also facilitates nuanced understanding of complex prompts, thereby allowing for more accurate and context-aware visual representations. ControlNet [63], has demonstrated exceptional performance in conditioned image generation by providing the model with references such as skeleton, canny edge images, and segmentation maps. However, these models frequently encounter difficulties when it comes to the generation of text within images (see Fig. 4), a task requiring the text to be not only visually integrated but also contextually pertinent to the image content. This challenge stems from the complexity of modeling the fine-grained interplay between visual and textual elements, ensuring that the generated text is legible, aesthetically fitting, and semantically in sync with the image. Prior attempts to refine this aspect have led to improvements [8, 49, 63], yet the generation of contextually coherent text in images remains a largely unsolved problem, underscoring the need for more focused research." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.71, + 0.771 + ], + "angle": 0, + "content": "2.2. Scene Text Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The success of adversarial networks such as GANs [7, 10, 17, 56, 67] for image generation and style transfer [20, 21] gave rise to scene text generation methods which can generate text at the granularity of glyphs [1, 26] or individual words [23, 54, 58]. Many previous works focused on the particular task of scene text editing where the model learns a style from a reference image and renders the target content in that style. Methods such as SRNet [54], SWAPText [58]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7497" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.258 + ], + "angle": 0, + "content": "decomposes the problem into: (1) learning the foreground text using style and content, (2) background in-painting network to remove existing text, and (3) a blending network to merge foreground and background. These methods often fail in learning the correct style and removing existing text from complex backgrounds. StyleBrush [23] proposes a self-supervised approach to disentangle content and style and generate word images in a one-shot manner. All these previous methods are limited to generating individual words, requires reference word style images and does not generalize to generate the entire scene text image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.368 + ], + "angle": 0, + "content": "Spatial fusion GAN (SF-GAN) [62] generates text images by superimposing a foreground content image, transformed to match the style and geometry of a background image. This approach is aimed more at pure text synthesis on image regions without text, whereas in this work we aim to generate both image and text in a manner that reflects its natural appearance in real-world contexts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.471, + 0.901 + ], + "angle": 0, + "content": "With the significant progress of diffusion models in text-to-image generation, recent methods in scene text generation adapt these models towards producing more legible visual text. DIFFSTE [18] enhances scene text editing with a dual encoder design in diffusion models, promoting text legibility and style control. It is adept at mapping text instructions to images, showcasing zero-shot capabilities for rendering text in novel font variations and interpreting informal natural language instructions. This method is however a scene text editing method which generates single keywords on specific regions defined by a mask, whereas we propose a scene text generation network. One of the closest work in this space is TextDiffuser [8] which address the problem of scene text generation by decomposing the problem into two-stages. In the first stage, a transformer model create the layout mask for keywords extracted from text prompts. This is then taken as condition data while formulating the diffusion model to generate scene text images. They also introduce the character segmentation loss which helps in generating legible text. ControlNet [63], though not originally designed for scene text image rendering, has been effectively adapted for this purpose. It utilizes canny edge maps as conditional inputs, sourced from printed text images generated by a layout model, to fine-tune the diffusion model's output. The use of pre-defined layout for visual text generation in [8, 63], seriously limits the diversity of text styles, fonts and even the layouts. We believe this is due to the inherent difficulty in predicting layouts independently without the general image guidance. In our work, we avoid the need of pre-defined layout and make the network implicitly learn layout along with image generation using our novel way of injecting character level features. We also introduce a word spotting loss which augments the character segmentation loss proposed in [8] to generate more legible visual texts." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.715, + 0.108 + ], + "angle": 0, + "content": "2.3. Scene Text Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Advances in computer vision have laid a foundation for sophisticated analytical techniques in various domains [2, 9, 14, 30-32, 36, 46, 53, 60, 61, 65]. Especially in scene text image recognition, most existing works split the process into two stages: a text detection [3, 28, 29, 34, 48, 66] module to detect words or characters from complex backgrounds, and a text recognition [2, 24, 47, 52] module which transcribes the text into unicode characters given a cropped word image. More recently, end-2-end methods [6, 25, 38, 43, 51] have become popular due to the benefits of joint training of detector and recognizer to share contextual information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.893, + 0.477 + ], + "angle": 0, + "content": "In the nexus of scene text generation and recognition, leveraging pre-trained scene text recognition or word spotting models as guidance during diffusion-based text-to-image synthesis has surfaced as an innovative strategy. For simplicity, in this paper we refer to a scene text recognition module as OCR (optical character recognition). The integration of OCR-derived losses enables the refinement of generative models to produce text that is not just visually coherent but also contextually accurate. This confluence of generative modeling prowess with OCR accuracy paves the way for novel research avenues to generate images with text that is both authentic to read and visually integrated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.634, + 0.508 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.515, + 0.623, + 0.53 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our objective is to facilitate layout-free text image generation with diverse layouts and styles. A straightforward approach would be to directly fine-tune an existing latent diffusion model with text images. However, our early investigation suggests that this strategy did not yield significant improvements compared to the original LDM. We hypothesize that this limitation is due to two primary factors. First, the language encoder of the LDM, primarily designed for semantic interpretation, fails to capture character-specific information adequately for text rendering. This encoder tends to provide more semantic than structural information about the text, necessitating a dedicated network for encoding the text and guiding the model on its visual representation. Second, the conventional denoising loss used in diffusion models seems insufficient for accurately rendering text in images, often leading to text regions resembling textual patterns without the distinct features of text strokes. To address these challenges, we propose integrating a character-level encoder and a hierarchical cross-attention mechanism to learn character-level context information. Additionally, we introduce two auxiliary losses at the word and character levels to emphasize the text presentation. The subsequent sections will detail each of these components and their contribution to enhancing our model's performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.094, + 0.885, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.895, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Model Framework: SceneTextGen employs a character-level encoder to extract detailed character-specific features. During loss computation, the model leverages both word-level and character-level supervisions to guide the recovery of the image, in addition to the standard denoising loss. This dual-level supervision enhances the model's ability to accurately generate and refine text within scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.386, + 0.218, + 0.401 + ], + "angle": 0, + "content": "3.2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.474, + 0.669 + ], + "angle": 0, + "content": "The proposed method is trained on a corpus of visual text images. Given an image \\(\\mathcal{I}\\) with textual caption \\(c\\), we denote \\(w_{i}\\), where \\(i \\in [0, N]\\), as the visual text (words) present in this image. We also assume that for each word, we know its location, given in terms of the bounding box. For all practical purposes, we assume this can be pre-computed using a pre-trained OCR detection [28] and recognition network [5]. Please note, the OCR bounding box and transcriptions are typically noisy. Fig. 3 presents an overview of our proposed method. A latent diffusion model has three key components: a CLIP encoder for semantic embedding of text and images, an autoencoder for dimensionality reduction and feature extraction, and a UNet-based structure for effective image-text synthesis and manipulation. The core architecture of our method follows the latent diffusion model, which is adapted to incorporate the proposed character-level features and content-based loss functions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.318, + 0.696 + ], + "angle": 0, + "content": "3.3. Character-Level Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our approach begins by extracting and ordering the visual text words from the image based on its locations as given by the bounding box. We sort the word boxes following the conventional reading pattern, i.e., from left to right and top to bottom. This provides us the naive ordering of visual text which is then tokenized at the character level to capture its typographic information. Then, a character encoder is used to encode these tokenized characters into a high-dimensional feature space, allowing an accurate transcription of the text's spelling and appearance. The derived text features encapsulate the typographical details and are poised for the subsequent integration with the image's latent features, which have already been contextually primed" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.892, + 0.418 + ], + "angle": 0, + "content": "by the initial cross-attention with the encoded caption features obtained from the CLIP [39] text encoder." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.43, + 0.825, + 0.446 + ], + "angle": 0, + "content": "3.4. Hierarchical Text Integration Process" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.893, + 0.681 + ], + "angle": 0, + "content": "We have employed a sequential cross-attention mechanism to integrate character information into the U-Net architecture effectively. This method is inspired by the observation that the cross-attention of captioning features is pivotal in shaping the overall spatial structure of an image, a notion supported by prior studies [15]. Leveraging this concept, our model initially constructs the general spatial layout of the image content, guided by the caption, using the first cross-attention layer. It then focuses on the precise rendering of characters in a subsequent cross-attention phase. This approach establishes a hierarchical text integration process, facilitating the development of a preliminary visual scaffold that is both thematically and contextually coherent. It ensures that the textual elements are accurately positioned and seamlessly integrated into the overall image structure." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.694, + 0.857, + 0.71 + ], + "angle": 0, + "content": "3.5. OCR-Guided Diffusion for Text Accuracy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.824 + ], + "angle": 0, + "content": "To ensure the textual accuracy of generated images, our model incorporates an OCR loss, utilizing the predictions from the end to end pre-trained GLASS model [43]. Upon each iteration of the diffusion process, the UNet predicts a denoising step, from which we derive an estimation of clean image \\( x_0 \\). This estimated \\( x_0 \\) is then decoded through a Variational Autoencoder (VAE) [22] to reconstruct an image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The GLASS OCR model performs inference on this reconstructed image to produce word-level recognition results, which are represented as a tensor \\( P \\) with shape \\( [N,L,K] \\), where \\( N \\) is the number of words detected, \\( L \\) is the maximum length of any word, and \\( K \\) is the size of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "character set. The ground truth for these detections is represented as a tensor \\( G \\) with shape \\([N, L]\\), where each entry is the character index for the corresponding position, and non-character positions are marked with zero." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.152, + 0.47, + 0.182 + ], + "angle": 0, + "content": "The OCR loss \\((\\mathcal{L}_{\\mathrm{OCR}})\\) is computed using a masked cross-entropy function, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.201, + 0.468, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O C R}} = - \\frac {1}{\\sum M} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i j} \\cdot \\log \\left(\\frac {\\exp \\left(P _ {i j} [ G _ {i j} ]\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(P _ {i j} [ k ]\\right)}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.258, + 0.469, + 0.348 + ], + "angle": 0, + "content": "Here, \\( M \\) is a binary mask tensor that has the same shape as \\( G \\) and indicates the valid character positions (i.e., where \\( G_{ij} \\neq 0 \\)). \\( M_{ij} \\) represents the binary value of the mask at the \\( i \\)-th word and \\( j \\)-th character position, \\( P_{ij} \\) is the predicted probability distribution over the character set, and \\( G_{ij} \\) is the ground truth character index at that position." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.348, + 0.469, + 0.424 + ], + "angle": 0, + "content": "By integrating this OCR loss into the training regime, we guide the diffusion model to produce text that is not only visually coherent but also textually accurate, as recognized by the GLASS [43] OCR model, thereby enhancing the overall fidelity of the generated images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.432, + 0.469, + 0.461 + ], + "angle": 0, + "content": "3.6. Refinement of Text Generation with Character-Level Constraints" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.469, + 0.605 + ], + "angle": 0, + "content": "During the iterative refinement of our diffusion model, we observed an unintended consequence of the OCR loss; the model tended to generate images with repetitive words. This issue is potentially attributed to the blurry nature of images at higher noise levels during training, which could render the OCR loss counterproductive. The word-level OCR loss, while ensuring textual accuracy, imposes no explicit constraint on the quantity of text within the image, inadvertently encouraging the model to generate excessive text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.469, + 0.816 + ], + "angle": 0, + "content": "To address this, we augmented our loss function with a character-level segmentation loss, which acts directly on the latent space rather than the recovered image. After obtaining the predicted latent features of a image \\( x_0 \\) from the UNet, we proceed in two directions: we decode \\( x_0 \\) using the VAE to compute the word-level OCR loss on the recovered image (as explained earlier), and we also apply \\( x_0 \\) to a pre-trained character-level segmentation model based on U-Net adapted from [8]. This model outputs a 96-dimensional feature map (corresponding to the length of the alphabet plus one for non-character pixels) with a spatial resolution of \\( 64 \\times 64 \\). The character-aware loss is then computed via cross-entropy between this feature map and a resized character-level segmentation mask \\( C \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.817, + 0.469, + 0.862 + ], + "angle": 0, + "content": "Thus, the total loss function is a composite of the denoising loss, the word-level recognition loss, and the character-level segmentation loss, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s i n g}} + \\lambda_ {\\text {w o r d}} \\mathcal {L} _ {\\text {O C R - w o r d}} + \\lambda_ {\\text {c h a r}} \\mathcal {L} _ {\\text {O C R - c h a r}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{denoising}}\\) is the denoising loss, \\(\\mathcal{L}_{\\mathrm{OCR - word}}\\) is the word-level OCR loss, \\(\\mathcal{L}_{\\mathrm{OCR - char}}\\) is the character-level segmentation loss, and \\(\\lambda_{\\mathrm{word}}\\) and \\(\\lambda_{\\mathrm{char}}\\) are weighting coefficients balancing the contribution of each term." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.153, + 0.882, + 0.167 + ], + "angle": 0, + "content": "The character-level segmentation loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.191, + 0.892, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {O C R - c h a r}} = - \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {9 6} y _ {h w c} \\log \\left(\\hat {y} _ {h w c}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.302 + ], + "angle": 0, + "content": "where \\( H \\) and \\( W \\) are the height and width of the feature map, \\( y \\) is the ground truth character segmentation mask, and \\( \\hat{y} \\) is the predicted character probability map from the segmentation model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.365 + ], + "angle": 0, + "content": "By integrating character-level information directly in the latent space, we impose a structured constraint on text generation, promoting both the accuracy and the appropriate quantity of text in the generated images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.632, + 0.394 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.402, + 0.717, + 0.418 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.426, + 0.892, + 0.606 + ], + "angle": 0, + "content": "Datasets For the training of our model, we utilized the publicly available MARIO dataset from [8], excluding MARIO-TMDB, and MARIO-OpenLibrary subsets as they are not publicly accessible. Upon removing any corrupted images, our final dataset comprised 7,249,449 image-caption pairs. In addition, to bolster our model's capability in generating a broader range of concepts (beyond text-centric images), we integrated 2,110,745 non-text images. These additional images, accompanied by text pairs, were selected based on a minimum predicted aesthetics score of 6.25, allowing for joint training to enhance overall performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.626, + 0.892, + 0.852 + ], + "angle": 0, + "content": "Baselines We conducted quantitative comparisons of our SceneTextGen method against several leading approaches, including LDM[42], ControlNet[63], TextDiffuser[8], GlyphControl[59], and DeepFloyd[49], utilizing the publicly available code and pre-trained models for fairness. Notably, DeepFloyd is distinguished by its dual superresolution modules, enabling it to produce high-resolution images at 1024x1024 pixels, in contrast to the 512x512 pixel images generated by the other models. For ControlNet comparisons, we employed Canny edge maps of printed text images created by the initial model stage of TextDiffuser as conditioning inputs. However, due to the unavailability of APIs, open-source code or checkpoints, we could not extend our comparative analysis to includeImagen[45], eDiff-i[4], or GlyphDraw[35]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation Criteria We assess text rendering quality using the MARIO-7M-Eval, MARIO-TMDB-Eval, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7500" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.381 + ], + "angle": 0, + "content": "MARIO-OpenLibrary-Eval datasets. Our evaluation is twofold: firstly, through CLIPScore, which measures the cosine similarity between the image and text representations from CLIP; and secondly, via OCR Evaluation, which leverages existing OCR tools to detect and recognize text regions within the generated images. Metrics such as Average Precision, Average Recall, F1 Score, and Accuracy are employed to determine the presence of keywords in the generated images. During training, the input of the character encoder is the ground truth OCR texts. During the inference process, the character encoder receives as its input the text from captions provided by the user. These captions are enclosed in quotation marks as specified in the user's prompts, and the input of the CLIP text encoder is the full caption. For each generated and ground truth image pair, we utilize the easy-ocr library for OCR detection and recognition, followed by Hungarian Matching between the sets of texts, applying the Levenshtein distance (or edit distance) on the matched text pairs for the OCR evaluation." + }, + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.393, + 0.435, + 0.408 + ], + "angle": 0, + "content": "Algorithm 1 T2I Visual Text Performance Evaluation" + }, + { + "type": "algorithm", + "bbox": [ + 0.083, + 0.411, + 0.473, + 0.641 + ], + "angle": 0, + "content": "function METRICS(M) \nP ← PRECISION(M) \nR ← RECALL(M) \nF ← F1(P, R) \nA ← ACCURACY(M) \nreturn {P, R, F, A} \nend function \n1: GT ← OCR(Ground Truth Images) \n2: Gen ← OCR(Generated Images) \n3: Scores ← an empty list \n4: for each pair in ZIP(GT, Gen) do \n5: C ← COSTMAT(pair) \n6: M ← HUNGARIAN(C) \n7: Scores ← Scores ∪ {METRICS(M)} \n8: end for \n9: return AVERAGE(Scores)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Pseudo Code for OCR Performance Evaluation The methodology described in Algorithm 1 demonstrates the steps taken to assess the performance of the text-to-image conversion. In the initial step, the algorithm applies OCR to both the ground truth and the generated images. This process results in two sets of text outputs, which are then paired for comparison. The Hungarian algorithm is employed here to find the optimal matching between elements (words) of these two sets, minimizing the overall difference between the matched pairs. This is crucial for an objective and accurate comparison. For each matched pair, we calculate a cost matrix, which serves as the input for the Hungarian algorithm. The output of this step is a matching matrix \\( M \\) which represents the best possible alignment between the text elements in the ground truth and the generated images. Sub" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "sequentially, the algorithm computes key performance metrics for each pair: Precision, Recall, F1 Score, and Accuracy. Precision focuses on the accuracy of the replicated text, recall measures the completeness, F1 score provides a balance between precision and recall, and accuracy gives an overall effectiveness of the text replication. Finally, the algorithm averages these scores across all image pairs to provide an overall performance evaluation of the text-to-image conversion process." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.24, + 0.695, + 0.256 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.265, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Our experimental analysis in Tab. 1 provides a direct comparison of the OCR based recognition scores among various models as measured in terms of Precision, Recall, F1 scores, and Accuracy. Our results indicate that SceneTextGen consistently outperforms competing models in most metrics. Latent Diffusion Model, lacking a sophisticated mechanism for text comprehension, typically under-performs, leading to lower OCR scores. In contrast, DeepFloyd[49] incorporates a T5 encoder which aids in textual understanding, thereby enhancing the quality of the generated text. However, its performance is still limited due to an insufficient character-level understanding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.894, + 0.538 + ], + "angle": 0, + "content": "ControlNet[63], TextDiffuser[8], and GlyphControl[59], which utilize predefined text layouts or spatial information, show mixed results. While the explicit introduction of text information allows ControlNet to achieve high OCR scores, the resulting text often appears artificial and lacks seamless integration within the images (see Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.561, + 0.895, + 0.669 + ], + "angle": 0, + "content": "Cross-dataset Generalization Ability As demonstrated in Tab. 1, SceneTextGen-7M, despite being trained solely on the MARIO-7M dataset, exhibits strong generalization capabilities. It maintains robust OCR scores across evaluation sets from both the TMDB and OpenLibrary datasets, underscoring its adaptability and the efficacy of its training methodology." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.679, + 0.776, + 0.696 + ], + "angle": 0, + "content": "4.3. Measuring Font Style Diversity" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.854 + ], + "angle": 0, + "content": "To quantitatively assess the diversity of font styles generated by SceneTextGen, we utilized a pretrained VGG-based font recognition model trained on synthesized text images. This approach involved first extracting text image patches using a pretrained Optical Character Recognition (OCR) model. These patches were then processed through the font recognition model to retrieve features from the penultimate layer. By applying t-SNE for dimensionality reduction, we visualized the feature space to examine the proximity and diversity of text generated by different methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 5, the rendered text images — which serve as a baseline — were created by printing text onto a white canvas using the layout generator from [8] with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7501" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.223 + ], + "angle": 0, + "content": "
Pre-defined Text LayoutMARIO-7MTMDBOpenLibrary
AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)
TextDiffuser-10M [8]✓0.61350.42890.46830.34250.44010.42430.39940.28890.56170.38160.42420.2916
GlyphControl-10M* [59]✓0.50750.41180.41400.28830.36350.40820.34570.23620.47340.37620.38360.2534
Latent Diffusion Model0.14820.16900.12960.07530.17170.25790.17030.09740.19110.28730.19230.1106
DeepFloyd [49]0.24670.27880.22060.13660.22840.37380.24490.15000.25550.39100.26350.1610
ControlNet [63]✓0.51020.44440.42380.29810.30750.46670.32840.21940.40500.42730.36900.2415
TextDiffuser-7M [8]✓0.47780.34470.36820.25120.31980.33620.29610.19300.42570.31460.33180.2108
SceneTextGen-7M(Ours)0.52740.44200.44240.30880.38130.47160.37900.26020.41360.45190.39450.2571
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.234, + 0.895, + 0.276 + ], + "angle": 0, + "content": "Table 1. Comparative analysis of OCR based text recognition scores across different models. Note, 7M denotes models that were trained on the MARIO-7M dataset (7 million images with texts). In contrast, the TextDiffuser-10M category includes models trained on an expanded dataset collection that encompasses MARIO-7M, MARIO-TMDB, and MARIO-OpenLibrary. * denotes LAION-Glyph dataset." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.293, + 0.847, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.635, + 0.895, + 0.679 + ], + "angle": 0, + "content": "Figure 4. Comparative visualization of generated images. We present a side-by-side comparison of images generated from the same text prompt across different existing methods (with results generated by [8] as a proxy. Human faces are blurred for ethical considerations.). Each row corresponds to a unique prompt, showcasing the visual quality, text clarity, and contextual coherence achieved by each method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.473, + 0.9 + ], + "angle": 0, + "content": "the default 'Arial' font. This process establishes a reference for the feature distribution of 'Artificial Texts.' ControlNet [63], constrained by predefined text canny edges, exhibits a feature distribution closely aligned with the rendered text images. This proximity suggests its limitations in integrating text seamlessly within images, a finding corroborated by the qualitative results shown in Fig. 4. TextDiffuser [8], although performing well in most instances, still shares some feature space with ControlNet, indicating occasional production of artificial or unnatural texts. In contrast, SceneTextGen — operating independently of any predefined canny edges or layouts — demonstrates a distinct distribution with minimal overlap with the 'rendered text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.795 + ], + "angle": 0, + "content": "images,' signifying its robustness in generating naturalistic text within the context of scene images. Example visualizations are in Fig. 6. In addition, to better understand the text layout given by each model, we also show in Fig. 7 the overall distribution of visual text in the generated images and ground truth images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.802, + 0.696, + 0.818 + ], + "angle": 0, + "content": "4.4. Weight of Each Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we present two ablation studies to evaluate the impact of different loss function configurations on OCR performance. Table 3 explores the effects of using combinations of word-level and character-level losses. Table 4 examines the impact of varying the weights of these loss func" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7502" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.432, + 0.214 + ], + "angle": 0, + "content": "
Pre-defined \nText LayoutCLIP Score(↑)
TextDiffuser-10M [8]✓0.3436
GlyphControl-10M* [59]✓0.3450
Latent Diffusion Model0.3015
DeepFloyd [49]0.3267
ControlNet [63]✓0.3424
TextDiffuser-7M [8]✓0.3385
SceneTextGen-7M(Ours)0.3455
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.224, + 0.47, + 0.281 + ], + "angle": 0, + "content": "Table 2. Comparison of CLIP scores reflecting the overall image quality generated by various models. SceneTextGen-7M achieves the highest CLIP score, indicating superior text-image alignment without relying on text layouts. * denotes LAION-Glyph dataset." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.294, + 0.43, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.495 + ], + "angle": 0, + "content": "Figure 5. t-SNE representation of text region embeddings derived from the penultimate layer features of a font recognition model." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.508, + 0.468, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.631, + 0.373, + 0.644 + ], + "angle": 0, + "content": "Caption: A sign that says 'Hello World'" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.704 + ], + "angle": 0, + "content": "Figure 6. Diversity of font styles. SceneTextGen is able to generate visually appealing and diverse font styles and layouts for text, without any style or layout prompts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.791 + ], + "angle": 0, + "content": "tions. In both studies, we assess the Average Precision (AP) and Accuracy (AC) scores to understand how these configurations influence the model's ability to accurately recognize and generate text in images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.205, + 0.817 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "While SceneTextGen demonstrates superior performance in scene text image generation, challenges remain. Limited scene complexity: Models struggle to accurately generate complex elements in conjunction with text, such as a car with a 'green' sticker specifically in the back win" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.625, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.178, + 0.605, + 0.19 + ], + "angle": 0, + "content": "TextDiffuser" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.092, + 0.742, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.178, + 0.725, + 0.19 + ], + "angle": 0, + "content": "SceneTextGen" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.092, + 0.861, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.178, + 0.842, + 0.19 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.207, + 0.894, + 0.263 + ], + "angle": 0, + "content": "Figure 7. Distribution of visual text coordinates in the generated image w.r.t. the area of each box. As expected from real-life observations, the visual text in the ground truth images tends to be located at the center of each image." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.276, + 0.668, + 0.328 + ], + "angle": 0, + "content": "
LwordLcharAPAC
✓0.50820.3109
✓0.50830.3035
✓✓0.52740.3088
" + }, + { + "type": "table", + "bbox": [ + 0.729, + 0.276, + 0.874, + 0.328 + ], + "angle": 0, + "content": "
WlamClamAPAC
120.41070.2782
0.10.20.45870.2895
0.010.020.52740.3088
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.34, + 0.684, + 0.38 + ], + "angle": 0, + "content": "Table 3. Effects of loss function combinations on the MARIO-7M-Eval dataset" + }, + { + "type": "table_caption", + "bbox": [ + 0.709, + 0.34, + 0.892, + 0.38 + ], + "angle": 0, + "content": "Table 4. Impact of varying loss weights on the MARIO-7M-eval dataset" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.398, + 0.882, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.558, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Figure 8. Failure cases in complex scene interpretation and processing of lengthy prompts for text." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.613, + 0.892, + 0.659 + ], + "angle": 0, + "content": "dow.(Fig. 8, Case 1). Long text handling: Text accuracy and coherence decrease with increasing length (Fig. 8, Case 2). These areas require further development." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.619, + 0.687 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.818 + ], + "angle": 0, + "content": "SceneTextGen, incorporating a character-level encoder and hierarchical text integration, offers advancements in scene text image generation. Despite improved text rendering, limitations arise in generating complex visuals and handling lengthy text. These challenges highlight the ongoing difficulty in reconciling textual accuracy with broader image synthesis. This work furthers our understanding of text-image generation, paving the way for future exploration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.836, + 0.893, + 0.896 + ], + "angle": 0, + "content": "Acknowledgements This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7503" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[3] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9365-9374, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.464 + ], + "angle": 0, + "content": "[5] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[6] Michal Bušta, Yash Patel, and Jiri Matas. E2e-plt-an unconstrained end-to-end method for multi-language scene text. In Computer Vision-ACCV 2018 Workshops: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers 14, pages 127–143. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[7] Qi Chang, Zhennan Yan, Mu Zhou, Di Liu, Khalid Sawalha, Meng Ye, Qilong Zhangli, Mikael Kanski, Subhi Al'Aref, Leon Axel, et al. Deeprecon: Joint 2d cardiac segmentation and 3d volume reconstruction via a structure-specific generative method. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 567-577. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.703 + ], + "angle": 0, + "content": "[8] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[9] Yunhe Gao, Mu Zhou, Di Liu, Zhennan Yan, Shaoting Zhang, and Dimitris N Metaxas. A data-scalable transformer for medical image segmentation: architecture, model efficiency, and benchmark. arXiv preprint arXiv:2203.00131, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.162 + ], + "angle": 0, + "content": "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.895, + 0.272 + ], + "angle": 0, + "content": "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4291-4301, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[14] Xiaoxiao He, Chaowei Tan, Bo Liu, Liping Si, Weiwu Yao, Liang Zhao, Di Liu, Qilong Zhangli, Qi Chang, Kang Li, et al. Dealing with heterogeneous 3d mr knee images: A federated few-shot learning method with dual knowledge distillation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.424 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.895, + 0.482 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.893, + 0.55 + ], + "angle": 0, + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.552, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[18] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.893, + 0.65 + ], + "angle": 0, + "content": "[19] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Advances in Neural Information Processing Systems, pages 8563-8601, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.692 + ], + "angle": 0, + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.895, + 0.762 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[23] Praveen Krishnan, Rama Kovvuri, Guan Pang, Boris Vassilev, and Tal Hassner. Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[24] Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention modeling forOCR in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2231-2239, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7504" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13094–13102, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[26] Wei Li, Yongxing He, Yanwei Qi, Zejian Li, and Yongchuan Tang. Fet-gan: Font and effect transfer via k-shot adaptive instance normalization. In Proceedings of the AAAI conference on artificial intelligence, pages 1717-1724, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[28] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In Proceedings of the AAAI conference on artificial intelligence, pages 11474-11481, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[29] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[30] Di Liu, Yunhe Gao, Qilong Zhangli, Ligong Han, Xiaoxiao He, Zhaoyang Xia, Song Wen, Qi Chang, Zhennan Yan, Mu Zhou, et al. Transfusion: multi-view divergent fusion for medical image segmentation with transformers. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 485-495. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.594 + ], + "angle": 0, + "content": "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.65 + ], + "angle": 0, + "content": "[32] Di Liu, Anastasis Stathopoulos, Qilong Zhangli, Yunhe Gao, and Dimitris Metaxas. Lepard: Learning explicit part discovery for 3d articulated shape reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[33] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[34] Jianqi Ma, Weiyuan Shao, Hao Ye, Li Wang, Hong Wang, Yingbin Zheng, and Xiangyang Xue. Arbitrary-oriented scene text detection via rotation proposals. IEEE transactions on multimedia, 20(11):3111-3122, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[35] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coher-ently. arXiv preprint arXiv:2303.17870, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[36] Carlos Martin-Isla, Víctor M Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J Fulton, Tewodros Weldebirhan Arega, Kumaradevan Punithakumar, et al. Deep learning segmentation of the right ventricle in cardiac mri: The" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "m&ms challenge. IEEE Journal of Biomedical and Health Informatics, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[38] Siyang Qin, Alessandro Bissacco, Michalis Raptis, Yasuhisa Fujii, and Ying Xiao. Towards unconstrained end-to-end text spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4704-4714, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.474 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[43] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In European Conference on Computer Vision, pages 249-266. Springer, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.605, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[46] Lohrasb Ross Sayadi, Usama S Hamdan, Qilong Zhangli, and Raj M Vyas. Harnessing the power of artificial intelligence to teach cleft lip surgery. Plastic and Reconstructive Surgery-Global Open, 10(7):e4451, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[47] Baoguang Shi, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4168-4176, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[48] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[49] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd-if. https://github.com/deep-floyd/IF, 2023. GitHub Repository. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[50] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiaoyuan Jing, Fei Wu, and Bingkun Bao. Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[51] Peng Wang, Hui Li, and Chunhua Shen. Towards end-to-end text spotting in natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7266-7281, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.388 + ], + "angle": 0, + "content": "[52] Peng Wang, Cheng Da, and Cong Yao. Multi-granularity prediction for scene text recognition. In European Conference on Computer Vision, pages 339-355. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[53] Song Wen, Hao Wang, Di Liu, Qilong Zhangli, and Dimitris Metaxas. Second-order graph odes for multi-agent trajectory forecasting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5101-5110, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[54] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In Proceedings of the 27th ACM international conference on multimedia, pages 1500-1508, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[55] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotpdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[56] Zhaoyang Xia, Yuxiao Chen, Qilong Zhangli, Matt Huener-fauth, Carol Neidle, and Dimitris Metaxas. Sign language video anonymization. In Proceedings of the LREC2022 10th Workshop on the Representation and Processing of Sign Languages: Multilingual Sign Language Resources, Marseille, France, 25 June 2022, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[58] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14700-14709, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[59] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[60] Jiachi Ye, Haoyan Kang, Hao Wang, Chen Shen, Belal Jahannia, Elham Heidari, Nvid Asadizanjani, Mohammad-Ali Miri, Volker J Sorger, and Hamed Dalir. Demultiplexing oam" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "beams via fourier optical convolutional neural network. In Laser Beam Shaping XXIII, pages 16-33. SPIE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[61] Jiachi Ye, Maria Solyanik, Zibo Hu, Hamed Dalir, Behrouz Movahhed Nouri, and Volker J Sorger. Free-space optical multiplexed orbital angular momentum beam identification system using fourier optical convolutional layer based on 4f system. In Complex Light and Optical Forces XVII, pages 70-80. SPIE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[62] Fangneng Zhan, Hongyuan Zhu, and Shijian Lu. Spatial fusion gan for image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3653-3662, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[64] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.39, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[65] Qilong Zhangli, Jingru Yi, Di Liu, Xiaoxiao He, Zhaoyang Xia, Qi Chang, Ligong Han, Yunhe Gao, Song Wen, Haiming Tang, et al. Region proposal rectification towards robust instance segmentation of biological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[66] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[67] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7506" + } + ] +] \ No newline at end of file diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_origin.pdf b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a4642517953f2bbe9a78a14b54c50c61f1743d40 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/cffc0cc1-9751-4215-8102-3e240ce1394d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5036e66a7f6dcfd84fba842432bd8961526e28282568fb3fe64a817000f1a264 +size 4842096 diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/full.md b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..861ae5c6762b86b9794be3f65fde8d2a41dcb4d0 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/full.md @@ -0,0 +1,333 @@ +# Layout-Agnostic Scene Text Image Synthesis with Diffusion Models + +Qilong Zhangli $^{1,2}$ Jindong Jiang $^{1}$ Di Liu $^{1}$ Licheng Yu $^{2}$ Xiaoliang Dai $^{2}$ Ankit Ramchandani $^{2}$ Guan Pang $^{2}$ Dimitris N. Metaxas $^{1}$ Praveen Krishnan $^{2}$ $^{1}$ Rutgers University $^{2}$ Meta AI + +![](images/4f55640a5b3b582496b41378a992dbf9d9fc654a370869b4e2dde4fdee6165b7.jpg) + +![](images/04a97ac60502742b86f90d22d7e63f0bd14f6dd79f490ed21c38912a2fa5f490.jpg) + +![](images/7f815a24adc7ed4dcaa374f82ee5cd965c15299301bcbd0012777ccfe495c50a.jpg) + +![](images/1c5778c4ebca9ce0dffbdd641959d0a550b13806a2c5397459da8c948aa80141.jpg) + +![](images/4dc5618bcc17b750a0d57b496e4b835b32867de9b80b4bdc928d059eb8461bc1.jpg) + +![](images/5d202a5ca24575dbc24364e29e7ba09716b64a7d004a07cc8258d4ba62a5dd8f.jpg) + +![](images/7075f232c8a55425353c54295037708fff3632fddc4318ca47586f93b97d6236.jpg) +Figure 1. Scene Text Generation. Qualitative samples of scene text images generated by our model are presented. These images contain visually appealing texts that are coherent with the background and are created without relying on any spatial information or predefined layouts as input, thereby enhancing the Text-to-Image (T2I) Diffusion Model's capability to generate text. + +![](images/5f02803c29233d644775d35b7dc4adf040bdc8fd6960441839d769a7fa4c8277.jpg) + +![](images/a7a9488ce58b78fc30f20d95a2843eabed3388becb69910d4f330d67ca9e5749.jpg) + +![](images/0d2a70b1d257505d341b536e3d09ebf0d88c749f2cc936092cc8f0f689d109c8.jpg) + +![](images/e3150e268735cacda835a1a633cc0aef4315e0b6af240b566591e543f457d8cf.jpg) +'Time Traveling Device' + +![](images/f89071be178b7d838071078cee11e331c3ffa4b6fb0b814c3eaf3fa56eb682e0.jpg) +in the foreground, with the text +'contemplate the clouds' in +roundedcursive + +# Abstract + +While diffusion models have significantly advanced the quality of image generation, their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts, an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges, this paper introduces SceneTextGen, a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so, SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties, coupled with a character-level instance segmentation model and a word + +level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods. + +# 1. Introduction + +The text-to-image [50, 57] task has gained popularity with advancements in diffusion models [41, 45, 49], significantly enhancing the quality of image generation. However, seamlessly integrating clear and contextually appropriate text into images is a persistent challenge. Text plays a vital role in many domains, including media content creation and artistic design, yet current diffusion models often struggle to produce text that is lexically correct, in valid font style, and that naturally complements the overall image aesthetics. + +Prompt: A vintage postage stamp showing a painting of mountains and the text 'California'. + +![](images/8027ef464ee97173c5a5da9e6b544bb2bbb1cea7f23407d24e06571c4dd84193.jpg) +(a) Layout-Constrained Models + +![](images/46d07de1bcafdcab222e6d427c6fc0e5f5f2a2db2b40c0b6c23becb3c334ca15.jpg) +(b) Layout-Agnostic SceneTextGen +Figure 2. Models reliant on predefined text layouts for input exhibit limitations such as constrained font diversity and static text positioning during each inference, leading to a lack of variability in style and arrangement. + +Traditional methods [54, 58, 62] of creating scene text images typically formulate this problem as scene text editing which only involves editing or adding text to an existing scene image. These methods have challenges handling complex backgrounds, font styles and lighting variations. While recent models [4, 8, 33, 45, 49] have made strides in addressing these limitations by enhancing text encoding strategies [33] or employing predefined text layouts [8], they still face significant constraints in generating visual text. These constraints become apparent when observing the limited diversity in font styles and the static positioning of text, as shown in Fig. 2. Such rigidity in layout and font selection hampers the capacity of generative models to produce text that is stylistically varied and contextually aligned with the image content. + +To address these issues, we propose SceneTextGen, a novel framework that capitalizes on the capabilities of latent diffusion models to infuse text into scene images with greater diversity and authenticity. Our approach is specifically engineered to transcend the limitations of predefined layouts, enabling more flexible text placement and an expansive assortment of text styles. + +The two primary contributions of this work are: (i) the integration of character-level encoder to capture the typographic properties of visual text and carefully injecting it into the cross attention layers of the diffusion model, and (ii) introducing a novel word spotting loss using a pre-trained OCR along with a character segmentation loss to make the network more faithful in generating visual text. The character level encoder naturally blends with the existing diffusion model architecture and therefore allows the network to learn the layout of visual text implicitly rather than constraining itself to some pre-defined form. Our comprehensive evaluations confirm that SceneTextGen surpasses contemporary methods, facilitating the generation of images with text that + +is both aesthetically pleasing and rich in variety. + +# 2. Related Work + +# 2.1. Text 2 Image Generation Models + +In the realm of text-to-image generation, diffusion models [11, 16, 27, 37, 41, 44, 49, 63, 64] represent a significant leap forward. Different from GAN-based models [10, 17, 67], diffusion models employ a stochastic process that iteratively adds noise to an image and learn to reverse this process to generate images from textual descriptions. Their capacity to produce high-quality, detailed visual content from text prompts has been well documented. The Latent Diffusion Model (LDM)[42] further enhances this approach by operating in a compressed image latent space, improving both efficiency and image quality. It also facilitates high-quality conditional generation through cross-attention text conditioning, leading to various downstream applications [12, 13, 15, 19, 55]. DALL-E [41], renowned for its novel approach of combining discrete VAEs with transformer language models, has shown remarkable ability in generating diverse and complex images from textual descriptions, showcasing the potential of transformer architectures in creative generative tasks. Deepfloyd [49], utilizing the robust T5 text encoders [40], not only enhances image quality but also facilitates nuanced understanding of complex prompts, thereby allowing for more accurate and context-aware visual representations. ControlNet [63], has demonstrated exceptional performance in conditioned image generation by providing the model with references such as skeleton, canny edge images, and segmentation maps. However, these models frequently encounter difficulties when it comes to the generation of text within images (see Fig. 4), a task requiring the text to be not only visually integrated but also contextually pertinent to the image content. This challenge stems from the complexity of modeling the fine-grained interplay between visual and textual elements, ensuring that the generated text is legible, aesthetically fitting, and semantically in sync with the image. Prior attempts to refine this aspect have led to improvements [8, 49, 63], yet the generation of contextually coherent text in images remains a largely unsolved problem, underscoring the need for more focused research. + +# 2.2. Scene Text Generation + +The success of adversarial networks such as GANs [7, 10, 17, 56, 67] for image generation and style transfer [20, 21] gave rise to scene text generation methods which can generate text at the granularity of glyphs [1, 26] or individual words [23, 54, 58]. Many previous works focused on the particular task of scene text editing where the model learns a style from a reference image and renders the target content in that style. Methods such as SRNet [54], SWAPText [58] + +decomposes the problem into: (1) learning the foreground text using style and content, (2) background in-painting network to remove existing text, and (3) a blending network to merge foreground and background. These methods often fail in learning the correct style and removing existing text from complex backgrounds. StyleBrush [23] proposes a self-supervised approach to disentangle content and style and generate word images in a one-shot manner. All these previous methods are limited to generating individual words, requires reference word style images and does not generalize to generate the entire scene text image. + +Spatial fusion GAN (SF-GAN) [62] generates text images by superimposing a foreground content image, transformed to match the style and geometry of a background image. This approach is aimed more at pure text synthesis on image regions without text, whereas in this work we aim to generate both image and text in a manner that reflects its natural appearance in real-world contexts. + +With the significant progress of diffusion models in text-to-image generation, recent methods in scene text generation adapt these models towards producing more legible visual text. DIFFSTE [18] enhances scene text editing with a dual encoder design in diffusion models, promoting text legibility and style control. It is adept at mapping text instructions to images, showcasing zero-shot capabilities for rendering text in novel font variations and interpreting informal natural language instructions. This method is however a scene text editing method which generates single keywords on specific regions defined by a mask, whereas we propose a scene text generation network. One of the closest work in this space is TextDiffuser [8] which address the problem of scene text generation by decomposing the problem into two-stages. In the first stage, a transformer model create the layout mask for keywords extracted from text prompts. This is then taken as condition data while formulating the diffusion model to generate scene text images. They also introduce the character segmentation loss which helps in generating legible text. ControlNet [63], though not originally designed for scene text image rendering, has been effectively adapted for this purpose. It utilizes canny edge maps as conditional inputs, sourced from printed text images generated by a layout model, to fine-tune the diffusion model's output. The use of pre-defined layout for visual text generation in [8, 63], seriously limits the diversity of text styles, fonts and even the layouts. We believe this is due to the inherent difficulty in predicting layouts independently without the general image guidance. In our work, we avoid the need of pre-defined layout and make the network implicitly learn layout along with image generation using our novel way of injecting character level features. We also introduce a word spotting loss which augments the character segmentation loss proposed in [8] to generate more legible visual texts. + +# 2.3. Scene Text Recognition + +Advances in computer vision have laid a foundation for sophisticated analytical techniques in various domains [2, 9, 14, 30-32, 36, 46, 53, 60, 61, 65]. Especially in scene text image recognition, most existing works split the process into two stages: a text detection [3, 28, 29, 34, 48, 66] module to detect words or characters from complex backgrounds, and a text recognition [2, 24, 47, 52] module which transcribes the text into unicode characters given a cropped word image. More recently, end-2-end methods [6, 25, 38, 43, 51] have become popular due to the benefits of joint training of detector and recognizer to share contextual information. + +In the nexus of scene text generation and recognition, leveraging pre-trained scene text recognition or word spotting models as guidance during diffusion-based text-to-image synthesis has surfaced as an innovative strategy. For simplicity, in this paper we refer to a scene text recognition module as OCR (optical character recognition). The integration of OCR-derived losses enables the refinement of generative models to produce text that is not just visually coherent but also contextually accurate. This confluence of generative modeling prowess with OCR accuracy paves the way for novel research avenues to generate images with text that is both authentic to read and visually integrated. + +# 3. Methodology + +# 3.1. Motivation + +Our objective is to facilitate layout-free text image generation with diverse layouts and styles. A straightforward approach would be to directly fine-tune an existing latent diffusion model with text images. However, our early investigation suggests that this strategy did not yield significant improvements compared to the original LDM. We hypothesize that this limitation is due to two primary factors. First, the language encoder of the LDM, primarily designed for semantic interpretation, fails to capture character-specific information adequately for text rendering. This encoder tends to provide more semantic than structural information about the text, necessitating a dedicated network for encoding the text and guiding the model on its visual representation. Second, the conventional denoising loss used in diffusion models seems insufficient for accurately rendering text in images, often leading to text regions resembling textual patterns without the distinct features of text strokes. To address these challenges, we propose integrating a character-level encoder and a hierarchical cross-attention mechanism to learn character-level context information. Additionally, we introduce two auxiliary losses at the word and character levels to emphasize the text presentation. The subsequent sections will detail each of these components and their contribution to enhancing our model's performance. + +![](images/f58c1f1343c61e708eb975832790db88f5c42350b11e93ff3b6ee2090278e56e.jpg) +Figure 3. Model Framework: SceneTextGen employs a character-level encoder to extract detailed character-specific features. During loss computation, the model leverages both word-level and character-level supervisions to guide the recovery of the image, in addition to the standard denoising loss. This dual-level supervision enhances the model's ability to accurately generate and refine text within scenes. + +# 3.2. Preliminaries + +The proposed method is trained on a corpus of visual text images. Given an image $\mathcal{I}$ with textual caption $c$ , we denote $w_{i}$ , where $i \in [0, N]$ , as the visual text (words) present in this image. We also assume that for each word, we know its location, given in terms of the bounding box. For all practical purposes, we assume this can be pre-computed using a pre-trained OCR detection [28] and recognition network [5]. Please note, the OCR bounding box and transcriptions are typically noisy. Fig. 3 presents an overview of our proposed method. A latent diffusion model has three key components: a CLIP encoder for semantic embedding of text and images, an autoencoder for dimensionality reduction and feature extraction, and a UNet-based structure for effective image-text synthesis and manipulation. The core architecture of our method follows the latent diffusion model, which is adapted to incorporate the proposed character-level features and content-based loss functions. + +# 3.3. Character-Level Encoding + +Our approach begins by extracting and ordering the visual text words from the image based on its locations as given by the bounding box. We sort the word boxes following the conventional reading pattern, i.e., from left to right and top to bottom. This provides us the naive ordering of visual text which is then tokenized at the character level to capture its typographic information. Then, a character encoder is used to encode these tokenized characters into a high-dimensional feature space, allowing an accurate transcription of the text's spelling and appearance. The derived text features encapsulate the typographical details and are poised for the subsequent integration with the image's latent features, which have already been contextually primed + +by the initial cross-attention with the encoded caption features obtained from the CLIP [39] text encoder. + +# 3.4. Hierarchical Text Integration Process + +We have employed a sequential cross-attention mechanism to integrate character information into the U-Net architecture effectively. This method is inspired by the observation that the cross-attention of captioning features is pivotal in shaping the overall spatial structure of an image, a notion supported by prior studies [15]. Leveraging this concept, our model initially constructs the general spatial layout of the image content, guided by the caption, using the first cross-attention layer. It then focuses on the precise rendering of characters in a subsequent cross-attention phase. This approach establishes a hierarchical text integration process, facilitating the development of a preliminary visual scaffold that is both thematically and contextually coherent. It ensures that the textual elements are accurately positioned and seamlessly integrated into the overall image structure. + +# 3.5. OCR-Guided Diffusion for Text Accuracy + +To ensure the textual accuracy of generated images, our model incorporates an OCR loss, utilizing the predictions from the end to end pre-trained GLASS model [43]. Upon each iteration of the diffusion process, the UNet predicts a denoising step, from which we derive an estimation of clean image $x_0$ . This estimated $x_0$ is then decoded through a Variational Autoencoder (VAE) [22] to reconstruct an image. + +The GLASS OCR model performs inference on this reconstructed image to produce word-level recognition results, which are represented as a tensor $P$ with shape $[N,L,K]$ , where $N$ is the number of words detected, $L$ is the maximum length of any word, and $K$ is the size of the + +character set. The ground truth for these detections is represented as a tensor $G$ with shape $[N, L]$ , where each entry is the character index for the corresponding position, and non-character positions are marked with zero. + +The OCR loss $(\mathcal{L}_{\mathrm{OCR}})$ is computed using a masked cross-entropy function, which is formulated as follows: + +$$ +\mathcal {L} _ {\mathrm {O C R}} = - \frac {1}{\sum M} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {L} M _ {i j} \cdot \log \left(\frac {\exp \left(P _ {i j} [ G _ {i j} ]\right)}{\sum_ {k = 1} ^ {K} \exp \left(P _ {i j} [ k ]\right)}\right) \tag {1} +$$ + +Here, $M$ is a binary mask tensor that has the same shape as $G$ and indicates the valid character positions (i.e., where $G_{ij} \neq 0$ ). $M_{ij}$ represents the binary value of the mask at the $i$ -th word and $j$ -th character position, $P_{ij}$ is the predicted probability distribution over the character set, and $G_{ij}$ is the ground truth character index at that position. + +By integrating this OCR loss into the training regime, we guide the diffusion model to produce text that is not only visually coherent but also textually accurate, as recognized by the GLASS [43] OCR model, thereby enhancing the overall fidelity of the generated images. + +# 3.6. Refinement of Text Generation with Character-Level Constraints + +During the iterative refinement of our diffusion model, we observed an unintended consequence of the OCR loss; the model tended to generate images with repetitive words. This issue is potentially attributed to the blurry nature of images at higher noise levels during training, which could render the OCR loss counterproductive. The word-level OCR loss, while ensuring textual accuracy, imposes no explicit constraint on the quantity of text within the image, inadvertently encouraging the model to generate excessive text. + +To address this, we augmented our loss function with a character-level segmentation loss, which acts directly on the latent space rather than the recovered image. After obtaining the predicted latent features of a image $x_0$ from the UNet, we proceed in two directions: we decode $x_0$ using the VAE to compute the word-level OCR loss on the recovered image (as explained earlier), and we also apply $x_0$ to a pre-trained character-level segmentation model based on U-Net adapted from [8]. This model outputs a 96-dimensional feature map (corresponding to the length of the alphabet plus one for non-character pixels) with a spatial resolution of $64 \times 64$ . The character-aware loss is then computed via cross-entropy between this feature map and a resized character-level segmentation mask $C$ . + +Thus, the total loss function is a composite of the denoising loss, the word-level recognition loss, and the character-level segmentation loss, expressed as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {d e n o i s i n g}} + \lambda_ {\text {w o r d}} \mathcal {L} _ {\text {O C R - w o r d}} + \lambda_ {\text {c h a r}} \mathcal {L} _ {\text {O C R - c h a r}} \tag {2} +$$ + +where $\mathcal{L}_{\mathrm{denoising}}$ is the denoising loss, $\mathcal{L}_{\mathrm{OCR - word}}$ is the word-level OCR loss, $\mathcal{L}_{\mathrm{OCR - char}}$ is the character-level segmentation loss, and $\lambda_{\mathrm{word}}$ and $\lambda_{\mathrm{char}}$ are weighting coefficients balancing the contribution of each term. + +The character-level segmentation loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {O C R - c h a r}} = - \frac {1}{H W} \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \sum_ {c = 1} ^ {9 6} y _ {h w c} \log \left(\hat {y} _ {h w c}\right) \tag {3} +$$ + +where $H$ and $W$ are the height and width of the feature map, $y$ is the ground truth character segmentation mask, and $\hat{y}$ is the predicted character probability map from the segmentation model. + +By integrating character-level information directly in the latent space, we impose a structured constraint on text generation, promoting both the accuracy and the appropriate quantity of text in the generated images. + +# 4. Experiments + +# 4.1. Implementation Details + +Datasets For the training of our model, we utilized the publicly available MARIO dataset from [8], excluding MARIO-TMDB, and MARIO-OpenLibrary subsets as they are not publicly accessible. Upon removing any corrupted images, our final dataset comprised 7,249,449 image-caption pairs. In addition, to bolster our model's capability in generating a broader range of concepts (beyond text-centric images), we integrated 2,110,745 non-text images. These additional images, accompanied by text pairs, were selected based on a minimum predicted aesthetics score of 6.25, allowing for joint training to enhance overall performance. + +Baselines We conducted quantitative comparisons of our SceneTextGen method against several leading approaches, including LDM[42], ControlNet[63], TextDiffuser[8], GlyphControl[59], and DeepFloyd[49], utilizing the publicly available code and pre-trained models for fairness. Notably, DeepFloyd is distinguished by its dual superresolution modules, enabling it to produce high-resolution images at 1024x1024 pixels, in contrast to the 512x512 pixel images generated by the other models. For ControlNet comparisons, we employed Canny edge maps of printed text images created by the initial model stage of TextDiffuser as conditioning inputs. However, due to the unavailability of APIs, open-source code or checkpoints, we could not extend our comparative analysis to includeImagen[45], eDiff-i[4], or GlyphDraw[35]. + +Evaluation Criteria We assess text rendering quality using the MARIO-7M-Eval, MARIO-TMDB-Eval, and + +MARIO-OpenLibrary-Eval datasets. Our evaluation is twofold: firstly, through CLIPScore, which measures the cosine similarity between the image and text representations from CLIP; and secondly, via OCR Evaluation, which leverages existing OCR tools to detect and recognize text regions within the generated images. Metrics such as Average Precision, Average Recall, F1 Score, and Accuracy are employed to determine the presence of keywords in the generated images. During training, the input of the character encoder is the ground truth OCR texts. During the inference process, the character encoder receives as its input the text from captions provided by the user. These captions are enclosed in quotation marks as specified in the user's prompts, and the input of the CLIP text encoder is the full caption. For each generated and ground truth image pair, we utilize the easy-ocr library for OCR detection and recognition, followed by Hungarian Matching between the sets of texts, applying the Levenshtein distance (or edit distance) on the matched text pairs for the OCR evaluation. + +Algorithm 1 T2I Visual Text Performance Evaluation +function METRICS(M) +P ← PRECISION(M) +R ← RECALL(M) +F ← F1(P, R) +A ← ACCURACY(M) +return {P, R, F, A} +end function +1: GT ← OCR(Ground Truth Images) +2: Gen ← OCR(Generated Images) +3: Scores ← an empty list +4: for each pair in ZIP(GT, Gen) do +5: C ← COSTMAT(pair) +6: M ← HUNGARIAN(C) +7: Scores ← Scores ∪ {METRICS(M)} +8: end for +9: return AVERAGE(Scores) + +Pseudo Code for OCR Performance Evaluation The methodology described in Algorithm 1 demonstrates the steps taken to assess the performance of the text-to-image conversion. In the initial step, the algorithm applies OCR to both the ground truth and the generated images. This process results in two sets of text outputs, which are then paired for comparison. The Hungarian algorithm is employed here to find the optimal matching between elements (words) of these two sets, minimizing the overall difference between the matched pairs. This is crucial for an objective and accurate comparison. For each matched pair, we calculate a cost matrix, which serves as the input for the Hungarian algorithm. The output of this step is a matching matrix $M$ which represents the best possible alignment between the text elements in the ground truth and the generated images. Sub + +sequentially, the algorithm computes key performance metrics for each pair: Precision, Recall, F1 Score, and Accuracy. Precision focuses on the accuracy of the replicated text, recall measures the completeness, F1 score provides a balance between precision and recall, and accuracy gives an overall effectiveness of the text replication. Finally, the algorithm averages these scores across all image pairs to provide an overall performance evaluation of the text-to-image conversion process. + +# 4.2. Quantitative Results + +Our experimental analysis in Tab. 1 provides a direct comparison of the OCR based recognition scores among various models as measured in terms of Precision, Recall, F1 scores, and Accuracy. Our results indicate that SceneTextGen consistently outperforms competing models in most metrics. Latent Diffusion Model, lacking a sophisticated mechanism for text comprehension, typically under-performs, leading to lower OCR scores. In contrast, DeepFloyd[49] incorporates a T5 encoder which aids in textual understanding, thereby enhancing the quality of the generated text. However, its performance is still limited due to an insufficient character-level understanding. + +ControlNet[63], TextDiffuser[8], and GlyphControl[59], which utilize predefined text layouts or spatial information, show mixed results. While the explicit introduction of text information allows ControlNet to achieve high OCR scores, the resulting text often appears artificial and lacks seamless integration within the images (see Fig. 4). + +Cross-dataset Generalization Ability As demonstrated in Tab. 1, SceneTextGen-7M, despite being trained solely on the MARIO-7M dataset, exhibits strong generalization capabilities. It maintains robust OCR scores across evaluation sets from both the TMDB and OpenLibrary datasets, underscoring its adaptability and the efficacy of its training methodology. + +# 4.3. Measuring Font Style Diversity + +To quantitatively assess the diversity of font styles generated by SceneTextGen, we utilized a pretrained VGG-based font recognition model trained on synthesized text images. This approach involved first extracting text image patches using a pretrained Optical Character Recognition (OCR) model. These patches were then processed through the font recognition model to retrieve features from the penultimate layer. By applying t-SNE for dimensionality reduction, we visualized the feature space to examine the proximity and diversity of text generated by different methods. + +As illustrated in Fig. 5, the rendered text images — which serve as a baseline — were created by printing text onto a white canvas using the layout generator from [8] with + +
Pre-defined Text LayoutMARIO-7MTMDBOpenLibrary
AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)
TextDiffuser-10M [8]✓0.61350.42890.46830.34250.44010.42430.39940.28890.56170.38160.42420.2916
GlyphControl-10M* [59]✓0.50750.41180.41400.28830.36350.40820.34570.23620.47340.37620.38360.2534
Latent Diffusion Model0.14820.16900.12960.07530.17170.25790.17030.09740.19110.28730.19230.1106
DeepFloyd [49]0.24670.27880.22060.13660.22840.37380.24490.15000.25550.39100.26350.1610
ControlNet [63]✓0.51020.44440.42380.29810.30750.46670.32840.21940.40500.42730.36900.2415
TextDiffuser-7M [8]✓0.47780.34470.36820.25120.31980.33620.29610.19300.42570.31460.33180.2108
SceneTextGen-7M(Ours)0.52740.44200.44240.30880.38130.47160.37900.26020.41360.45190.39450.2571
+ +Table 1. Comparative analysis of OCR based text recognition scores across different models. Note, 7M denotes models that were trained on the MARIO-7M dataset (7 million images with texts). In contrast, the TextDiffuser-10M category includes models trained on an expanded dataset collection that encompasses MARIO-7M, MARIO-TMDB, and MARIO-OpenLibrary. * denotes LAION-Glyph dataset. + +![](images/80028703f90f109493fb0f4762b77da7e6e6d8e7b018270029d0d3fd0288ba95.jpg) +Figure 4. Comparative visualization of generated images. We present a side-by-side comparison of images generated from the same text prompt across different existing methods (with results generated by [8] as a proxy. Human faces are blurred for ethical considerations.). Each row corresponds to a unique prompt, showcasing the visual quality, text clarity, and contextual coherence achieved by each method. + +the default 'Arial' font. This process establishes a reference for the feature distribution of 'Artificial Texts.' ControlNet [63], constrained by predefined text canny edges, exhibits a feature distribution closely aligned with the rendered text images. This proximity suggests its limitations in integrating text seamlessly within images, a finding corroborated by the qualitative results shown in Fig. 4. TextDiffuser [8], although performing well in most instances, still shares some feature space with ControlNet, indicating occasional production of artificial or unnatural texts. In contrast, SceneTextGen — operating independently of any predefined canny edges or layouts — demonstrates a distinct distribution with minimal overlap with the 'rendered text + +images,' signifying its robustness in generating naturalistic text within the context of scene images. Example visualizations are in Fig. 6. In addition, to better understand the text layout given by each model, we also show in Fig. 7 the overall distribution of visual text in the generated images and ground truth images. + +# 4.4. Weight of Each Loss + +In this section, we present two ablation studies to evaluate the impact of different loss function configurations on OCR performance. Table 3 explores the effects of using combinations of word-level and character-level losses. Table 4 examines the impact of varying the weights of these loss func + +
Pre-defined +Text LayoutCLIP Score(↑)
TextDiffuser-10M [8]✓0.3436
GlyphControl-10M* [59]✓0.3450
Latent Diffusion Model0.3015
DeepFloyd [49]0.3267
ControlNet [63]✓0.3424
TextDiffuser-7M [8]✓0.3385
SceneTextGen-7M(Ours)0.3455
+ +![](images/7831a1c1e096e1d49e1b4c36fc750f3461a00c47479e296deeec5f251a9ccdbe.jpg) +Figure 5. t-SNE representation of text region embeddings derived from the penultimate layer features of a font recognition model. + +![](images/8ef7b853fb4ee53bc1682a40646e8d09b64a9425fd4ffdb9d884730acfa2b6a7.jpg) +Caption: A sign that says 'Hello World' +Figure 6. Diversity of font styles. SceneTextGen is able to generate visually appealing and diverse font styles and layouts for text, without any style or layout prompts. + +tions. In both studies, we assess the Average Precision (AP) and Accuracy (AC) scores to understand how these configurations influence the model's ability to accurately recognize and generate text in images. + +# 4.5. Limitations + +While SceneTextGen demonstrates superior performance in scene text image generation, challenges remain. Limited scene complexity: Models struggle to accurately generate complex elements in conjunction with text, such as a car with a 'green' sticker specifically in the back win + +![](images/41803e143ea7eb1dc0edc55fd7980a2b6c70bcd13e44856cc39fad113dccd7d7.jpg) +TextDiffuser + +![](images/5e545863a265564e7b10efca04de74032c86f0a9af1f22c7da2d20da0d9abb54.jpg) +SceneTextGen +Figure 7. Distribution of visual text coordinates in the generated image w.r.t. the area of each box. As expected from real-life observations, the visual text in the ground truth images tends to be located at the center of each image. + +![](images/0c80c44c344aa77d47da102c5fb105c22f02ab4e454559c6b660bdb6a81aefb6.jpg) +Ground Truth + +Table 2. Comparison of CLIP scores reflecting the overall image quality generated by various models. SceneTextGen-7M achieves the highest CLIP score, indicating superior text-image alignment without relying on text layouts. * denotes LAION-Glyph dataset. + +
LwordLcharAPAC
✓0.50820.3109
✓0.50830.3035
✓✓0.52740.3088
+ +
WlamClamAPAC
120.41070.2782
0.10.20.45870.2895
0.010.020.52740.3088
+ +Table 3. Effects of loss function combinations on the MARIO-7M-Eval dataset + +Table 4. Impact of varying loss weights on the MARIO-7M-eval dataset + +![](images/a3bc893c50e3f640c1674a55b5cc69176eaf23d13967ea125f84e43565c60239.jpg) +Figure 8. Failure cases in complex scene interpretation and processing of lengthy prompts for text. + +dow.(Fig. 8, Case 1). Long text handling: Text accuracy and coherence decrease with increasing length (Fig. 8, Case 2). These areas require further development. + +# 5. Conclusion + +SceneTextGen, incorporating a character-level encoder and hierarchical text integration, offers advancements in scene text image generation. Despite improved text rendering, limitations arise in generating complex visuals and handling lengthy text. These challenges highlight the ongoing difficulty in reconciling textual accuracy with broader image synthesis. This work furthers our understanding of text-image generation, paving the way for future exploration. + +Acknowledgements This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417. + +# References + +[1] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 2 +[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3 +[3] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9365-9374, 2019. 3 +[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 5 +[5] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 4 +[6] Michal Bušta, Yash Patel, and Jiri Matas. E2e-plt-an unconstrained end-to-end method for multi-language scene text. In Computer Vision-ACCV 2018 Workshops: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers 14, pages 127–143. Springer, 2019. 3 +[7] Qi Chang, Zhennan Yan, Mu Zhou, Di Liu, Khalid Sawalha, Meng Ye, Qilong Zhangli, Mikael Kanski, Subhi Al'Aref, Leon Axel, et al. Deeprecon: Joint 2d cardiac segmentation and 3d volume reconstruction via a structure-specific generative method. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 567-577. Springer, 2022. 2 +[8] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 2, 3, 5, 6, 7, 8 +[9] Yunhe Gao, Mu Zhou, Di Liu, Zhennan Yan, Shaoting Zhang, and Dimitris N Metaxas. A data-scalable transformer for medical image segmentation: architecture, model efficiency, and benchmark. arXiv preprint arXiv:2203.00131, 2022. 3 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2 + +[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2 +[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4291-4301, 2024. 2 +[14] Xiaoxiao He, Chaowei Tan, Bo Liu, Liping Si, Weiwu Yao, Liang Zhao, Di Liu, Qilong Zhangli, Qi Chang, Kang Li, et al. Dealing with heterogeneous 3d mr knee images: A federated few-shot learning method with dual knowledge distillation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 3 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 2, 4 +[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2 +[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[18] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 3 +[19] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Advances in Neural Information Processing Systems, pages 8563-8601, 2023. 2 +[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2 +[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4 +[23] Praveen Krishnan, Rama Kovvuri, Guan Pang, Boris Vassilev, and Tal Hassner. Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3 +[24] Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention modeling forOCR in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2231-2239, 2016. 3 + +[25] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13094–13102, 2023. 3 +[26] Wei Li, Yongxing He, Yanwei Qi, Zejian Li, and Yongchuan Tang. Fet-gan: Font and effect transfer via k-shot adaptive instance normalization. In Proceedings of the AAAI conference on artificial intelligence, pages 1717-1724, 2020. 2 +[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2 +[28] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In Proceedings of the AAAI conference on artificial intelligence, pages 11474-11481, 2020. 3, 4 +[29] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 3 +[30] Di Liu, Yunhe Gao, Qilong Zhangli, Ligong Han, Xiaoxiao He, Zhaoyang Xia, Song Wen, Qi Chang, Zhennan Yan, Mu Zhou, et al. Transfusion: multi-view divergent fusion for medical image segmentation with transformers. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 485-495. Springer, 2022. 3 +[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023. +[32] Di Liu, Anastasis Stathopoulos, Qilong Zhangli, Yunhe Gao, and Dimitris Metaxas. Lepard: Learning explicit part discovery for 3d articulated shape reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3 +[33] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022. 2 +[34] Jianqi Ma, Weiyuan Shao, Hao Ye, Li Wang, Hong Wang, Yingbin Zheng, and Xiangyang Xue. Arbitrary-oriented scene text detection via rotation proposals. IEEE transactions on multimedia, 20(11):3111-3122, 2018. 3 +[35] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coher-ently. arXiv preprint arXiv:2303.17870, 2023. 5 +[36] Carlos Martin-Isla, Víctor M Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J Fulton, Tewodros Weldebirhan Arega, Kumaradevan Punithakumar, et al. Deep learning segmentation of the right ventricle in cardiac mri: The + +m&ms challenge. IEEE Journal of Biomedical and Health Informatics, 2023. 3 +[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[38] Siyang Qin, Alessandro Bissacco, Michalis Raptis, Yasuhisa Fujii, and Ying Xiao. Towards unconstrained end-to-end text spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4704-4714, 2019. 3 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2 +[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2 +[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5 +[43] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In European Conference on Computer Vision, pages 249-266. Springer, 2022. 3, 4, 5 +[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2 +[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2, 5 +[46] Lohrasb Ross Sayadi, Usama S Hamdan, Qilong Zhangli, and Raj M Vyas. Harnessing the power of artificial intelligence to teach cleft lip surgery. Plastic and Reconstructive Surgery-Global Open, 10(7):e4451, 2022. 3 +[47] Baoguang Shi, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4168-4176, 2016. 3 + +[48] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3 +[49] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd-if. https://github.com/deep-floyd/IF, 2023. GitHub Repository. 1, 2, 5, 6, 7, 8 +[50] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiaoyuan Jing, Fei Wu, and Bingkun Bao. Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 1 +[51] Peng Wang, Hui Li, and Chunhua Shen. Towards end-to-end text spotting in natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7266-7281, 2021. 3 +[52] Peng Wang, Cheng Da, and Cong Yao. Multi-granularity prediction for scene text recognition. In European Conference on Computer Vision, pages 339-355. Springer, 2022. 3 +[53] Song Wen, Hao Wang, Di Liu, Qilong Zhangli, and Dimitris Metaxas. Second-order graph odes for multi-agent trajectory forecasting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5101-5110, 2024. 3 +[54] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In Proceedings of the 27th ACM international conference on multimedia, pages 1500-1508, 2019. 2 +[55] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotpdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2 +[56] Zhaoyang Xia, Yuxiao Chen, Qilong Zhangli, Matt Huener-fauth, Carol Neidle, and Dimitris Metaxas. Sign language video anonymization. In Proceedings of the LREC2022 10th Workshop on the Representation and Processing of Sign Languages: Multilingual Sign Language Resources, Marseille, France, 25 June 2022, 2022. 2 +[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1 +[58] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14700-14709, 2020. 2 +[59] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7, 8 +[60] Jiachi Ye, Haoyan Kang, Hao Wang, Chen Shen, Belal Jahannia, Elham Heidari, Nvid Asadizanjani, Mohammad-Ali Miri, Volker J Sorger, and Hamed Dalir. Demultiplexing oam + +beams via fourier optical convolutional neural network. In Laser Beam Shaping XXIII, pages 16-33. SPIE, 2023. 3 +[61] Jiachi Ye, Maria Solyanik, Zibo Hu, Hamed Dalir, Behrouz Movahhed Nouri, and Volker J Sorger. Free-space optical multiplexed orbital angular momentum beam identification system using fourier optical convolutional layer based on 4f system. In Complex Light and Optical Forces XVII, pages 70-80. SPIE, 2023. 3 +[62] Fangneng Zhan, Hongyuan Zhu, and Shijian Lu. Spatial fusion gan for image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3653-3662, 2019. 2, 3 +[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 5, 6, 7, 8 +[64] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2 +[65] Qilong Zhangli, Jingru Yi, Di Liu, Xiaoxiao He, Zhaoyang Xia, Qi Chang, Ligong Han, Yunhe Gao, Song Wen, Haiming Tang, et al. Region proposal rectification towards robust instance segmentation of biological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2022. 3 +[66] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 3 +[67] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/images.zip b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c84b39ccef1d7a8439f7d3493488dad093191c96 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f0e9bd6f3d0731570002de3a22540a7c53e2dc10e94378b8593806f2904b44b +size 800942 diff --git a/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/layout.json b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a4a61442cafb39e9d5389a9e8bb5f34cda58dde7 --- /dev/null +++ b/2024/Layout-Agnostic Scene Text Image Synthesis with Diffusion Models/layout.json @@ -0,0 +1,8423 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 505, + 121 + ], + "type": "text", + "content": "Layout-Agnostic Scene Text Image Synthesis with Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": "Qilong Zhangli" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Jindong Jiang" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Di Liu" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Licheng Yu" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Xiaoliang Dai" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Ankit Ramchandani" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Guan Pang" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Dimitris N. Metaxas" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": " Praveen Krishnan" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": "Rutgers University " + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 142, + 491, + 186 + ], + "type": "text", + "content": "Meta AI" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 203, + 132, + 306 + ], + "blocks": [ + { + "bbox": [ + 50, + 203, + 132, + 306 + ], + "lines": [ + { + "bbox": [ + 50, + 203, + 132, + 306 + ], + "spans": [ + { + "bbox": [ + 50, + 203, + 132, + 306 + ], + "type": "image", + "image_path": "4f55640a5b3b582496b41378a992dbf9d9fc654a370869b4e2dde4fdee6165b7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 203, + 212, + 308 + ], + "blocks": [ + { + "bbox": [ + 132, + 203, + 212, + 308 + ], + "lines": [ + { + "bbox": [ + 132, + 203, + 212, + 308 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 212, + 308 + ], + "type": "image", + "image_path": "04a97ac60502742b86f90d22d7e63f0bd14f6dd79f490ed21c38912a2fa5f490.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 204, + 294, + 300 + ], + "blocks": [ + { + "bbox": [ + 214, + 204, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 214, + 204, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 214, + 204, + 294, + 300 + ], + "type": "image", + "image_path": "7f815a24adc7ed4dcaa374f82ee5cd965c15299301bcbd0012777ccfe495c50a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 296, + 203, + 376, + 300 + ], + "blocks": [ + { + "bbox": [ + 296, + 203, + 376, + 300 + ], + "lines": [ + { + "bbox": [ + 296, + 203, + 376, + 300 + ], + "spans": [ + { + "bbox": [ + 296, + 203, + 376, + 300 + ], + "type": "image", + "image_path": "1c5778c4ebca9ce0dffbdd641959d0a550b13806a2c5397459da8c948aa80141.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 377, + 204, + 457, + 300 + ], + "blocks": [ + { + "bbox": [ + 377, + 204, + 457, + 300 + ], + "lines": [ + { + "bbox": [ + 377, + 204, + 457, + 300 + ], + "spans": [ + { + "bbox": [ + 377, + 204, + 457, + 300 + ], + "type": "image", + "image_path": "4dc5618bcc17b750a0d57b496e4b835b32867de9b80b4bdc928d059eb8461bc1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 459, + 204, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 459, + 204, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 459, + 204, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 459, + 204, + 541, + 308 + ], + "type": "image", + "image_path": "5d202a5ca24575dbc24364e29e7ba09716b64a7d004a07cc8258d4ba62a5dd8f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 310, + 130, + 407 + ], + "blocks": [ + { + "bbox": [ + 50, + 310, + 130, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 310, + 130, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 310, + 130, + 407 + ], + "type": "image", + "image_path": "7075f232c8a55425353c54295037708fff3632fddc4318ca47586f93b97d6236.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 438, + 547, + 473 + ], + "lines": [ + { + "bbox": [ + 46, + 438, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 547, + 473 + ], + "type": "text", + "content": "Figure 1. Scene Text Generation. Qualitative samples of scene text images generated by our model are presented. These images contain visually appealing texts that are coherent with the background and are created without relying on any spatial information or predefined layouts as input, thereby enhancing the Text-to-Image (T2I) Diffusion Model's capability to generate text." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 132, + 310, + 212, + 407 + ], + "blocks": [ + { + "bbox": [ + 132, + 310, + 212, + 407 + ], + "lines": [ + { + "bbox": [ + 132, + 310, + 212, + 407 + ], + "spans": [ + { + "bbox": [ + 132, + 310, + 212, + 407 + ], + "type": "image", + "image_path": "5f02803c29233d644775d35b7dc4adf040bdc8fd6960441839d769a7fa4c8277.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 214, + 310, + 295, + 406 + ], + "blocks": [ + { + "bbox": [ + 214, + 310, + 295, + 406 + ], + "lines": [ + { + "bbox": [ + 214, + 310, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 214, + 310, + 295, + 406 + ], + "type": "image", + "image_path": "a7a9488ce58b78fc30f20d95a2843eabed3388becb69910d4f330d67ca9e5749.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 296, + 310, + 376, + 406 + ], + "blocks": [ + { + "bbox": [ + 296, + 310, + 376, + 406 + ], + "lines": [ + { + "bbox": [ + 296, + 310, + 376, + 406 + ], + "spans": [ + { + "bbox": [ + 296, + 310, + 376, + 406 + ], + "type": "image", + "image_path": "0d2a70b1d257505d341b536e3d09ebf0d88c749f2cc936092cc8f0f689d109c8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 377, + 310, + 458, + 407 + ], + "blocks": [ + { + "bbox": [ + 377, + 310, + 458, + 407 + ], + "lines": [ + { + "bbox": [ + 377, + 310, + 458, + 407 + ], + "spans": [ + { + "bbox": [ + 377, + 310, + 458, + 407 + ], + "type": "image", + "image_path": "e3150e268735cacda835a1a633cc0aef4315e0b6af240b566591e543f457d8cf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 407, + 449, + 414 + ], + "lines": [ + { + "bbox": [ + 377, + 407, + 449, + 414 + ], + "spans": [ + { + "bbox": [ + 377, + 407, + 449, + 414 + ], + "type": "text", + "content": "'Time Traveling Device'" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 459, + 310, + 541, + 407 + ], + "blocks": [ + { + "bbox": [ + 459, + 310, + 541, + 407 + ], + "lines": [ + { + "bbox": [ + 459, + 310, + 541, + 407 + ], + "spans": [ + { + "bbox": [ + 459, + 310, + 541, + 407 + ], + "type": "image", + "image_path": "f89071be178b7d838071078cee11e331c3ffa4b6fb0b814c3eaf3fa56eb682e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 407, + 539, + 414 + ], + "lines": [ + { + "bbox": [ + 459, + 407, + 539, + 414 + ], + "spans": [ + { + "bbox": [ + 459, + 407, + 539, + 414 + ], + "type": "text", + "content": "in the foreground, with the text" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 415, + 534, + 420 + ], + "lines": [ + { + "bbox": [ + 463, + 415, + 534, + 420 + ], + "spans": [ + { + "bbox": [ + 463, + 415, + 534, + 420 + ], + "type": "text", + "content": "'contemplate the clouds' in" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 478, + 422, + 520, + 429 + ], + "lines": [ + { + "bbox": [ + 478, + 422, + 520, + 429 + ], + "spans": [ + { + "bbox": [ + 478, + 422, + 520, + 429 + ], + "type": "text", + "content": "roundedcursive" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 482, + 192, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 482, + 192, + 495 + ], + "spans": [ + { + "bbox": [ + 143, + 482, + 192, + 495 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 45, + 510, + 290, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 510, + 290, + 713 + ], + "spans": [ + { + "bbox": [ + 45, + 510, + 290, + 713 + ], + "type": "text", + "content": "While diffusion models have significantly advanced the quality of image generation, their capability to accurately and coherently render text within these images remains a substantial challenge. Conventional diffusion-based methods for scene text generation are typically limited by their reliance on an intermediate layout output. This dependency often results in a constrained diversity of text styles and fonts, an inherent limitation stemming from the deterministic nature of the layout generation phase. To address these challenges, this paper introduces SceneTextGen, a novel diffusion-based model specifically designed to circumvent the need for a predefined layout stage. By doing so, SceneTextGen facilitates a more natural and varied representation of text. The novelty of SceneTextGen lies in its integration of three key components: a character-level encoder for capturing detailed typographic properties, coupled with a character-level instance segmentation model and a word" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 483, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 556 + ], + "type": "text", + "content": "level spotting model to address the issues of unwanted text generation and minor character inaccuracies. We validate the performance of our method by demonstrating improved character recognition rates on generated images across different public visual text datasets in comparison to both standard diffusion based methods and text specific methods." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 583, + 387, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 583, + 387, + 596 + ], + "spans": [ + { + "bbox": [ + 306, + 583, + 387, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "The text-to-image [50, 57] task has gained popularity with advancements in diffusion models [41, 45, 49], significantly enhancing the quality of image generation. However, seamlessly integrating clear and contextually appropriate text into images is a persistent challenge. Text plays a vital role in many domains, including media content creation and artistic design, yet current diffusion models often struggle to produce text that is lexically correct, in valid font style, and that naturally complements the overall image aesthetics." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7496" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 52, + 74, + 278, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 74, + 278, + 93 + ], + "spans": [ + { + "bbox": [ + 52, + 74, + 278, + 93 + ], + "type": "text", + "content": "Prompt: A vintage postage stamp showing a painting of mountains and the text 'California'." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 95, + 161, + 203 + ], + "blocks": [ + { + "bbox": [ + 53, + 95, + 161, + 203 + ], + "lines": [ + { + "bbox": [ + 53, + 95, + 161, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 95, + 161, + 203 + ], + "type": "image", + "image_path": "8027ef464ee97173c5a5da9e6b544bb2bbb1cea7f23407d24e06571c4dd84193.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 205, + 156, + 215 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 156, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 156, + 215 + ], + "type": "text", + "content": "(a) Layout-Constrained Models" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 171, + 94, + 281, + 204 + ], + "blocks": [ + { + "bbox": [ + 171, + 94, + 281, + 204 + ], + "lines": [ + { + "bbox": [ + 171, + 94, + 281, + 204 + ], + "spans": [ + { + "bbox": [ + 171, + 94, + 281, + 204 + ], + "type": "image", + "image_path": "46d07de1bcafdcab222e6d427c6fc0e5f5f2a2db2b40c0b6c23becb3c334ca15.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 205, + 281, + 216 + ], + "lines": [ + { + "bbox": [ + 167, + 205, + 281, + 216 + ], + "spans": [ + { + "bbox": [ + 167, + 205, + 281, + 216 + ], + "type": "text", + "content": "(b) Layout-Agnostic SceneTextGen" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 229, + 288, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 274 + ], + "type": "text", + "content": "Figure 2. Models reliant on predefined text layouts for input exhibit limitations such as constrained font diversity and static text positioning during each inference, leading to a lack of variability in style and arrangement." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 474 + ], + "type": "text", + "content": "Traditional methods [54, 58, 62] of creating scene text images typically formulate this problem as scene text editing which only involves editing or adding text to an existing scene image. These methods have challenges handling complex backgrounds, font styles and lighting variations. While recent models [4, 8, 33, 45, 49] have made strides in addressing these limitations by enhancing text encoding strategies [33] or employing predefined text layouts [8], they still face significant constraints in generating visual text. These constraints become apparent when observing the limited diversity in font styles and the static positioning of text, as shown in Fig. 2. Such rigidity in layout and font selection hampers the capacity of generative models to produce text that is stylistically varied and contextually aligned with the image content." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 475, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 558 + ], + "type": "text", + "content": "To address these issues, we propose SceneTextGen, a novel framework that capitalizes on the capabilities of latent diffusion models to infuse text into scene images with greater diversity and authenticity. Our approach is specifically engineered to transcend the limitations of predefined layouts, enabling more flexible text placement and an expansive assortment of text styles." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 714 + ], + "type": "text", + "content": "The two primary contributions of this work are: (i) the integration of character-level encoder to capture the typographic properties of visual text and carefully injecting it into the cross attention layers of the diffusion model, and (ii) introducing a novel word spotting loss using a pre-trained OCR along with a character segmentation loss to make the network more faithful in generating visual text. The character level encoder naturally blends with the existing diffusion model architecture and therefore allows the network to learn the layout of visual text implicitly rather than constraining itself to some pre-defined form. Our comprehensive evaluations confirm that SceneTextGen surpasses contemporary methods, facilitating the generation of images with text that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 500, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 500, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 500, + 85 + ], + "type": "text", + "content": "is both aesthetically pleasing and rich in variety." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 96, + 392, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 96, + 392, + 108 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 392, + 108 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 116, + 481, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 116, + 481, + 129 + ], + "spans": [ + { + "bbox": [ + 306, + 116, + 481, + 129 + ], + "type": "text", + "content": "2.1. Text 2 Image Generation Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 135, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 135, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 135, + 545, + 589 + ], + "type": "text", + "content": "In the realm of text-to-image generation, diffusion models [11, 16, 27, 37, 41, 44, 49, 63, 64] represent a significant leap forward. Different from GAN-based models [10, 17, 67], diffusion models employ a stochastic process that iteratively adds noise to an image and learn to reverse this process to generate images from textual descriptions. Their capacity to produce high-quality, detailed visual content from text prompts has been well documented. The Latent Diffusion Model (LDM)[42] further enhances this approach by operating in a compressed image latent space, improving both efficiency and image quality. It also facilitates high-quality conditional generation through cross-attention text conditioning, leading to various downstream applications [12, 13, 15, 19, 55]. DALL-E [41], renowned for its novel approach of combining discrete VAEs with transformer language models, has shown remarkable ability in generating diverse and complex images from textual descriptions, showcasing the potential of transformer architectures in creative generative tasks. Deepfloyd [49], utilizing the robust T5 text encoders [40], not only enhances image quality but also facilitates nuanced understanding of complex prompts, thereby allowing for more accurate and context-aware visual representations. ControlNet [63], has demonstrated exceptional performance in conditioned image generation by providing the model with references such as skeleton, canny edge images, and segmentation maps. However, these models frequently encounter difficulties when it comes to the generation of text within images (see Fig. 4), a task requiring the text to be not only visually integrated but also contextually pertinent to the image content. This challenge stems from the complexity of modeling the fine-grained interplay between visual and textual elements, ensuring that the generated text is legible, aesthetically fitting, and semantically in sync with the image. Prior attempts to refine this aspect have led to improvements [8, 49, 63], yet the generation of contextually coherent text in images remains a largely unsolved problem, underscoring the need for more focused research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 598, + 434, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 434, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 434, + 610 + ], + "type": "text", + "content": "2.2. Scene Text Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "The success of adversarial networks such as GANs [7, 10, 17, 56, 67] for image generation and style transfer [20, 21] gave rise to scene text generation methods which can generate text at the granularity of glyphs [1, 26] or individual words [23, 54, 58]. Many previous works focused on the particular task of scene text editing where the model learns a style from a reference image and renders the target content in that style. Methods such as SRNet [54], SWAPText [58]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7497" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 204 + ], + "type": "text", + "content": "decomposes the problem into: (1) learning the foreground text using style and content, (2) background in-painting network to remove existing text, and (3) a blending network to merge foreground and background. These methods often fail in learning the correct style and removing existing text from complex backgrounds. StyleBrush [23] proposes a self-supervised approach to disentangle content and style and generate word images in a one-shot manner. All these previous methods are limited to generating individual words, requires reference word style images and does not generalize to generate the entire scene text image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 291 + ], + "type": "text", + "content": "Spatial fusion GAN (SF-GAN) [62] generates text images by superimposing a foreground content image, transformed to match the style and geometry of a background image. This approach is aimed more at pure text synthesis on image regions without text, whereas in this work we aim to generate both image and text in a manner that reflects its natural appearance in real-world contexts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 713 + ], + "type": "text", + "content": "With the significant progress of diffusion models in text-to-image generation, recent methods in scene text generation adapt these models towards producing more legible visual text. DIFFSTE [18] enhances scene text editing with a dual encoder design in diffusion models, promoting text legibility and style control. It is adept at mapping text instructions to images, showcasing zero-shot capabilities for rendering text in novel font variations and interpreting informal natural language instructions. This method is however a scene text editing method which generates single keywords on specific regions defined by a mask, whereas we propose a scene text generation network. One of the closest work in this space is TextDiffuser [8] which address the problem of scene text generation by decomposing the problem into two-stages. In the first stage, a transformer model create the layout mask for keywords extracted from text prompts. This is then taken as condition data while formulating the diffusion model to generate scene text images. They also introduce the character segmentation loss which helps in generating legible text. ControlNet [63], though not originally designed for scene text image rendering, has been effectively adapted for this purpose. It utilizes canny edge maps as conditional inputs, sourced from printed text images generated by a layout model, to fine-tune the diffusion model's output. The use of pre-defined layout for visual text generation in [8, 63], seriously limits the diversity of text styles, fonts and even the layouts. We believe this is due to the inherent difficulty in predicting layouts independently without the general image guidance. In our work, we avoid the need of pre-defined layout and make the network implicitly learn layout along with image generation using our novel way of injecting character level features. We also introduce a word spotting loss which augments the character segmentation loss proposed in [8] to generate more legible visual texts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 437, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 437, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 437, + 85 + ], + "type": "text", + "content": "2.3. Scene Text Recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 233 + ], + "type": "text", + "content": "Advances in computer vision have laid a foundation for sophisticated analytical techniques in various domains [2, 9, 14, 30-32, 36, 46, 53, 60, 61, 65]. Especially in scene text image recognition, most existing works split the process into two stages: a text detection [3, 28, 29, 34, 48, 66] module to detect words or characters from complex backgrounds, and a text recognition [2, 24, 47, 52] module which transcribes the text into unicode characters given a cropped word image. More recently, end-2-end methods [6, 25, 38, 43, 51] have become popular due to the benefits of joint training of detector and recognizer to share contextual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 234, + 546, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 546, + 377 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 546, + 377 + ], + "type": "text", + "content": "In the nexus of scene text generation and recognition, leveraging pre-trained scene text recognition or word spotting models as guidance during diffusion-based text-to-image synthesis has surfaced as an innovative strategy. For simplicity, in this paper we refer to a scene text recognition module as OCR (optical character recognition). The integration of OCR-derived losses enables the refinement of generative models to produce text that is not just visually coherent but also contextually accurate. This confluence of generative modeling prowess with OCR accuracy paves the way for novel research avenues to generate images with text that is both authentic to read and visually integrated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 388, + 388, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 388, + 402 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 388, + 402 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 407, + 381, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 407, + 381, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 407, + 381, + 419 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 713 + ], + "type": "text", + "content": "Our objective is to facilitate layout-free text image generation with diverse layouts and styles. A straightforward approach would be to directly fine-tune an existing latent diffusion model with text images. However, our early investigation suggests that this strategy did not yield significant improvements compared to the original LDM. We hypothesize that this limitation is due to two primary factors. First, the language encoder of the LDM, primarily designed for semantic interpretation, fails to capture character-specific information adequately for text rendering. This encoder tends to provide more semantic than structural information about the text, necessitating a dedicated network for encoding the text and guiding the model on its visual representation. Second, the conventional denoising loss used in diffusion models seems insufficient for accurately rendering text in images, often leading to text regions resembling textual patterns without the distinct features of text strokes. To address these challenges, we propose integrating a character-level encoder and a hierarchical cross-attention mechanism to learn character-level context information. Additionally, we introduce two auxiliary losses at the word and character levels to emphasize the text presentation. The subsequent sections will detail each of these components and their contribution to enhancing our model's performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7498" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 74, + 541, + 240 + ], + "blocks": [ + { + "bbox": [ + 48, + 74, + 541, + 240 + ], + "lines": [ + { + "bbox": [ + 48, + 74, + 541, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 74, + 541, + 240 + ], + "type": "image", + "image_path": "f58c1f1343c61e708eb975832790db88f5c42350b11e93ff3b6ee2090278e56e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 285 + ], + "type": "text", + "content": "Figure 3. Model Framework: SceneTextGen employs a character-level encoder to extract detailed character-specific features. During loss computation, the model leverages both word-level and character-level supervisions to guide the recovery of the image, in addition to the standard denoising loss. This dual-level supervision enhances the model's ability to accurately generate and refine text within scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 305, + 133, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 133, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 133, + 317 + ], + "type": "text", + "content": "3.2. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "content": "The proposed method is trained on a corpus of visual text images. Given an image " + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "content": " with textual caption " + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "inline_equation", + "content": "i \\in [0, N]" + }, + { + "bbox": [ + 46, + 324, + 290, + 529 + ], + "type": "text", + "content": ", as the visual text (words) present in this image. We also assume that for each word, we know its location, given in terms of the bounding box. For all practical purposes, we assume this can be pre-computed using a pre-trained OCR detection [28] and recognition network [5]. Please note, the OCR bounding box and transcriptions are typically noisy. Fig. 3 presents an overview of our proposed method. A latent diffusion model has three key components: a CLIP encoder for semantic embedding of text and images, an autoencoder for dimensionality reduction and feature extraction, and a UNet-based structure for effective image-text synthesis and manipulation. The core architecture of our method follows the latent diffusion model, which is adapted to incorporate the proposed character-level features and content-based loss functions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 538, + 194, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 194, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 194, + 551 + ], + "type": "text", + "content": "3.3. Character-Level Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Our approach begins by extracting and ordering the visual text words from the image based on its locations as given by the bounding box. We sort the word boxes following the conventional reading pattern, i.e., from left to right and top to bottom. This provides us the naive ordering of visual text which is then tokenized at the character level to capture its typographic information. Then, a character encoder is used to encode these tokenized characters into a high-dimensional feature space, allowing an accurate transcription of the text's spelling and appearance. The derived text features encapsulate the typographical details and are poised for the subsequent integration with the image's latent features, which have already been contextually primed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 331 + ], + "type": "text", + "content": "by the initial cross-attention with the encoded caption features obtained from the CLIP [39] text encoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 340, + 504, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 340, + 504, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 340, + 504, + 353 + ], + "type": "text", + "content": "3.4. Hierarchical Text Integration Process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 360, + 546, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 546, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 546, + 539 + ], + "type": "text", + "content": "We have employed a sequential cross-attention mechanism to integrate character information into the U-Net architecture effectively. This method is inspired by the observation that the cross-attention of captioning features is pivotal in shaping the overall spatial structure of an image, a notion supported by prior studies [15]. Leveraging this concept, our model initially constructs the general spatial layout of the image content, guided by the caption, using the first cross-attention layer. It then focuses on the precise rendering of characters in a subsequent cross-attention phase. This approach establishes a hierarchical text integration process, facilitating the development of a preliminary visual scaffold that is both thematically and contextually coherent. It ensures that the textual elements are accurately positioned and seamlessly integrated into the overall image structure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 549, + 524, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 549, + 524, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 549, + 524, + 562 + ], + "type": "text", + "content": "3.5. OCR-Guided Diffusion for Text Accuracy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": "To ensure the textual accuracy of generated images, our model incorporates an OCR loss, utilizing the predictions from the end to end pre-trained GLASS model [43]. Upon each iteration of the diffusion process, the UNet predicts a denoising step, from which we derive an estimation of clean image " + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": ". This estimated " + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 569, + 545, + 652 + ], + "type": "text", + "content": " is then decoded through a Variational Autoencoder (VAE) [22] to reconstruct an image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The GLASS OCR model performs inference on this reconstructed image to produce word-level recognition results, which are represented as a tensor " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "[N,L,K]" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is the number of words detected, " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is the maximum length of any word, and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is the size of the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7499" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "character set. The ground truth for these detections is represented as a tensor " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "inline_equation", + "content": "[N, L]" + }, + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": ", where each entry is the character index for the corresponding position, and non-character positions are marked with zero." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 120, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 120, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 47, + 120, + 287, + 144 + ], + "type": "text", + "content": "The OCR loss " + }, + { + "bbox": [ + 47, + 120, + 287, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{OCR}})" + }, + { + "bbox": [ + 47, + 120, + 287, + 144 + ], + "type": "text", + "content": " is computed using a masked cross-entropy function, which is formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 286, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 286, + 203 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O C R}} = - \\frac {1}{\\sum M} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {L} M _ {i j} \\cdot \\log \\left(\\frac {\\exp \\left(P _ {i j} [ G _ {i j} ]\\right)}{\\sum_ {k = 1} ^ {K} \\exp \\left(P _ {i j} [ k ]\\right)}\\right) \\tag {1}", + "image_path": "24853d6420e8bc680cdc51abea753dc34c7351a2d80e832bc4e794d8f68fe33f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": " is a binary mask tensor that has the same shape as " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": " and indicates the valid character positions (i.e., where " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "G_{ij} \\neq 0" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "M_{ij}" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": " represents the binary value of the mask at the " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": "-th word and " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": "-th character position, " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "P_{ij}" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": " is the predicted probability distribution over the character set, and " + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "inline_equation", + "content": "G_{ij}" + }, + { + "bbox": [ + 47, + 204, + 287, + 275 + ], + "type": "text", + "content": " is the ground truth character index at that position." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 275, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 275, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 287, + 335 + ], + "type": "text", + "content": "By integrating this OCR loss into the training regime, we guide the diffusion model to produce text that is not only visually coherent but also textually accurate, as recognized by the GLASS [43] OCR model, thereby enhancing the overall fidelity of the generated images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 365 + ], + "type": "text", + "content": "3.6. Refinement of Text Generation with Character-Level Constraints" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 372, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 287, + 479 + ], + "type": "text", + "content": "During the iterative refinement of our diffusion model, we observed an unintended consequence of the OCR loss; the model tended to generate images with repetitive words. This issue is potentially attributed to the blurry nature of images at higher noise levels during training, which could render the OCR loss counterproductive. The word-level OCR loss, while ensuring textual accuracy, imposes no explicit constraint on the quantity of text within the image, inadvertently encouraging the model to generate excessive text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": "To address this, we augmented our loss function with a character-level segmentation loss, which acts directly on the latent space rather than the recovered image. After obtaining the predicted latent features of a image " + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": " from the UNet, we proceed in two directions: we decode " + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": " using the VAE to compute the word-level OCR loss on the recovered image (as explained earlier), and we also apply " + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": " to a pre-trained character-level segmentation model based on U-Net adapted from [8]. This model outputs a 96-dimensional feature map (corresponding to the length of the alphabet plus one for non-character pixels) with a spatial resolution of " + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": ". The character-aware loss is then computed via cross-entropy between this feature map and a resized character-level segmentation mask " + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 479, + 287, + 646 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 287, + 682 + ], + "type": "text", + "content": "Thus, the total loss function is a composite of the denoising loss, the word-level recognition loss, and the character-level segmentation loss, expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 701, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {d e n o i s i n g}} + \\lambda_ {\\text {w o r d}} \\mathcal {L} _ {\\text {O C R - w o r d}} + \\lambda_ {\\text {c h a r}} \\mathcal {L} _ {\\text {O C R - c h a r}} \\tag {2}", + "image_path": "af42e74a68e36cba91e626f51558c31c4c63ed8c8952d7a3106e39e7120f4da8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{denoising}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " is the denoising loss, " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OCR - word}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " is the word-level OCR loss, " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{OCR - char}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " is the character-level segmentation loss, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{word}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{char}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 119 + ], + "type": "text", + "content": " are weighting coefficients balancing the contribution of each term." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 539, + 132 + ], + "type": "text", + "content": "The character-level segmentation loss is formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 151, + 545, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 151, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 321, + 151, + 545, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {O C R - c h a r}} = - \\frac {1}{H W} \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {9 6} y _ {h w c} \\log \\left(\\hat {y} _ {h w c}\\right) \\tag {3}", + "image_path": "d3eee7451c5bb9903102c0213d34c897257bb0e54fcba1c1109d80b2d18a980d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": " are the height and width of the feature map, " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": " is the ground truth character segmentation mask, and " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": " is the predicted character probability map from the segmentation model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 240, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 289 + ], + "type": "text", + "content": "By integrating character-level information directly in the latent space, we impose a structured constraint on text generation, promoting both the accuracy and the appropriate quantity of text in the generated images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 299, + 386, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 386, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 386, + 312 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 318, + 438, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 318, + 438, + 331 + ], + "spans": [ + { + "bbox": [ + 306, + 318, + 438, + 331 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 337, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 337, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 337, + 545, + 479 + ], + "type": "text", + "content": "Datasets For the training of our model, we utilized the publicly available MARIO dataset from [8], excluding MARIO-TMDB, and MARIO-OpenLibrary subsets as they are not publicly accessible. Upon removing any corrupted images, our final dataset comprised 7,249,449 image-caption pairs. In addition, to bolster our model's capability in generating a broader range of concepts (beyond text-centric images), we integrated 2,110,745 non-text images. These additional images, accompanied by text pairs, were selected based on a minimum predicted aesthetics score of 6.25, allowing for joint training to enhance overall performance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 495, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 495, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 305, + 495, + 545, + 674 + ], + "type": "text", + "content": "Baselines We conducted quantitative comparisons of our SceneTextGen method against several leading approaches, including LDM[42], ControlNet[63], TextDiffuser[8], GlyphControl[59], and DeepFloyd[49], utilizing the publicly available code and pre-trained models for fairness. Notably, DeepFloyd is distinguished by its dual superresolution modules, enabling it to produce high-resolution images at 1024x1024 pixels, in contrast to the 512x512 pixel images generated by the other models. For ControlNet comparisons, we employed Canny edge maps of printed text images created by the initial model stage of TextDiffuser as conditioning inputs. However, due to the unavailability of APIs, open-source code or checkpoints, we could not extend our comparative analysis to includeImagen[45], eDiff-i[4], or GlyphDraw[35]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Evaluation Criteria We assess text rendering quality using the MARIO-7M-Eval, MARIO-TMDB-Eval, and" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7500" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "content": "MARIO-OpenLibrary-Eval datasets. Our evaluation is twofold: firstly, through CLIPScore, which measures the cosine similarity between the image and text representations from CLIP; and secondly, via OCR Evaluation, which leverages existing OCR tools to detect and recognize text regions within the generated images. Metrics such as Average Precision, Average Recall, F1 Score, and Accuracy are employed to determine the presence of keywords in the generated images. During training, the input of the character encoder is the ground truth OCR texts. During the inference process, the character encoder receives as its input the text from captions provided by the user. These captions are enclosed in quotation marks as specified in the user's prompts, and the input of the CLIP text encoder is the full caption. For each generated and ground truth image pair, we utilize the easy-ocr library for OCR detection and recognition, followed by Hungarian Matching between the sets of texts, applying the Levenshtein distance (or edit distance) on the matched text pairs for the OCR evaluation." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 50, + 325, + 289, + 507 + ], + "blocks": [ + { + "bbox": [ + 48, + 311, + 266, + 323 + ], + "lines": [ + { + "bbox": [ + 48, + 311, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 266, + 323 + ], + "type": "text", + "content": "Algorithm 1 T2I Visual Text Performance Evaluation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 50, + 325, + 289, + 507 + ], + "lines": [ + { + "bbox": [ + 50, + 325, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 50, + 325, + 289, + 507 + ], + "type": "text", + "content": "function METRICS(M) \nP ← PRECISION(M) \nR ← RECALL(M) \nF ← F1(P, R) \nA ← ACCURACY(M) \nreturn {P, R, F, A} \nend function \n1: GT ← OCR(Ground Truth Images) \n2: Gen ← OCR(Generated Images) \n3: Scores ← an empty list \n4: for each pair in ZIP(GT, Gen) do \n5: C ← COSTMAT(pair) \n6: M ← HUNGARIAN(C) \n7: Scores ← Scores ∪ {METRICS(M)} \n8: end for \n9: return AVERAGE(Scores)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Pseudo Code for OCR Performance Evaluation The methodology described in Algorithm 1 demonstrates the steps taken to assess the performance of the text-to-image conversion. In the initial step, the algorithm applies OCR to both the ground truth and the generated images. This process results in two sets of text outputs, which are then paired for comparison. The Hungarian algorithm is employed here to find the optimal matching between elements (words) of these two sets, minimizing the overall difference between the matched pairs. This is crucial for an objective and accurate comparison. For each matched pair, we calculate a cost matrix, which serves as the input for the Hungarian algorithm. The output of this step is a matching matrix " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " which represents the best possible alignment between the text elements in the ground truth and the generated images. Sub" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "sequentially, the algorithm computes key performance metrics for each pair: Precision, Recall, F1 Score, and Accuracy. Precision focuses on the accuracy of the replicated text, recall measures the completeness, F1 score provides a balance between precision and recall, and accuracy gives an overall effectiveness of the text replication. Finally, the algorithm averages these scores across all image pairs to provide an overall performance evaluation of the text-to-image conversion process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 190, + 425, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 190, + 425, + 202 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 425, + 202 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 545, + 353 + ], + "type": "text", + "content": "Our experimental analysis in Tab. 1 provides a direct comparison of the OCR based recognition scores among various models as measured in terms of Precision, Recall, F1 scores, and Accuracy. Our results indicate that SceneTextGen consistently outperforms competing models in most metrics. Latent Diffusion Model, lacking a sophisticated mechanism for text comprehension, typically under-performs, leading to lower OCR scores. In contrast, DeepFloyd[49] incorporates a T5 encoder which aids in textual understanding, thereby enhancing the quality of the generated text. However, its performance is still limited due to an insufficient character-level understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 354, + 547, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 426 + ], + "type": "text", + "content": "ControlNet[63], TextDiffuser[8], and GlyphControl[59], which utilize predefined text layouts or spatial information, show mixed results. While the explicit introduction of text information allows ControlNet to achieve high OCR scores, the resulting text often appears artificial and lacks seamless integration within the images (see Fig. 4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 444, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 444, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 444, + 547, + 529 + ], + "type": "text", + "content": "Cross-dataset Generalization Ability As demonstrated in Tab. 1, SceneTextGen-7M, despite being trained solely on the MARIO-7M dataset, exhibits strong generalization capabilities. It maintains robust OCR scores across evaluation sets from both the TMDB and OpenLibrary datasets, underscoring its adaptability and the efficacy of its training methodology." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 537, + 474, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 537, + 474, + 551 + ], + "spans": [ + { + "bbox": [ + 305, + 537, + 474, + 551 + ], + "type": "text", + "content": "4.3. Measuring Font Style Diversity" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 676 + ], + "type": "text", + "content": "To quantitatively assess the diversity of font styles generated by SceneTextGen, we utilized a pretrained VGG-based font recognition model trained on synthesized text images. This approach involved first extracting text image patches using a pretrained Optical Character Recognition (OCR) model. These patches were then processed through the font recognition model to retrieve features from the penultimate layer. By applying t-SNE for dimensionality reduction, we visualized the feature space to examine the proximity and diversity of text generated by different methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 5, the rendered text images — which serve as a baseline — were created by printing text onto a white canvas using the layout generator from [8] with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7501" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 176 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 176 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 176 + ], + "type": "table", + "html": "
Pre-defined Text LayoutMARIO-7MTMDBOpenLibrary
AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)AP(↑)AR(↑)F1(↑)AC(↑)
TextDiffuser-10M [8]✓0.61350.42890.46830.34250.44010.42430.39940.28890.56170.38160.42420.2916
GlyphControl-10M* [59]✓0.50750.41180.41400.28830.36350.40820.34570.23620.47340.37620.38360.2534
Latent Diffusion Model0.14820.16900.12960.07530.17170.25790.17030.09740.19110.28730.19230.1106
DeepFloyd [49]0.24670.27880.22060.13660.22840.37380.24490.15000.25550.39100.26350.1610
ControlNet [63]✓0.51020.44440.42380.29810.30750.46670.32840.21940.40500.42730.36900.2415
TextDiffuser-7M [8]✓0.47780.34470.36820.25120.31980.33620.29610.19300.42570.31460.33180.2108
SceneTextGen-7M(Ours)0.52740.44200.44240.30880.38130.47160.37900.26020.41360.45190.39450.2571
", + "image_path": "e4bf3f722347d1a5ddb8193a9789d1068b91d664880ae0e6b7adf7d083dae5ad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 547, + 218 + ], + "type": "text", + "content": "Table 1. Comparative analysis of OCR based text recognition scores across different models. Note, 7M denotes models that were trained on the MARIO-7M dataset (7 million images with texts). In contrast, the TextDiffuser-10M category includes models trained on an expanded dataset collection that encompasses MARIO-7M, MARIO-TMDB, and MARIO-OpenLibrary. * denotes LAION-Glyph dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 72, + 232, + 518, + 494 + ], + "blocks": [ + { + "bbox": [ + 72, + 232, + 518, + 494 + ], + "lines": [ + { + "bbox": [ + 72, + 232, + 518, + 494 + ], + "spans": [ + { + "bbox": [ + 72, + 232, + 518, + 494 + ], + "type": "image", + "image_path": "80028703f90f109493fb0f4762b77da7e6e6d8e7b018270029d0d3fd0288ba95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 502, + 547, + 537 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 547, + 537 + ], + "type": "text", + "content": "Figure 4. Comparative visualization of generated images. We present a side-by-side comparison of images generated from the same text prompt across different existing methods (with results generated by [8] as a proxy. Human faces are blurred for ethical considerations.). Each row corresponds to a unique prompt, showcasing the visual quality, text clarity, and contextual coherence achieved by each method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 556, + 289, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 289, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 289, + 712 + ], + "type": "text", + "content": "the default 'Arial' font. This process establishes a reference for the feature distribution of 'Artificial Texts.' ControlNet [63], constrained by predefined text canny edges, exhibits a feature distribution closely aligned with the rendered text images. This proximity suggests its limitations in integrating text seamlessly within images, a finding corroborated by the qualitative results shown in Fig. 4. TextDiffuser [8], although performing well in most instances, still shares some feature space with ControlNet, indicating occasional production of artificial or unnatural texts. In contrast, SceneTextGen — operating independently of any predefined canny edges or layouts — demonstrates a distinct distribution with minimal overlap with the 'rendered text" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 556, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 629 + ], + "type": "text", + "content": "images,' signifying its robustness in generating naturalistic text within the context of scene images. Example visualizations are in Fig. 6. In addition, to better understand the text layout given by each model, we also show in Fig. 7 the overall distribution of visual text in the generated images and ground truth images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 635, + 425, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 635, + 425, + 647 + ], + "spans": [ + { + "bbox": [ + 305, + 635, + 425, + 647 + ], + "type": "text", + "content": "4.4. Weight of Each Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "In this section, we present two ablation studies to evaluate the impact of different loss function configurations on OCR performance. Table 3 explores the effects of using combinations of word-level and character-level losses. Table 4 examines the impact of varying the weights of these loss func" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7502" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 264, + 169 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 264, + 169 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 264, + 169 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 264, + 169 + ], + "type": "table", + "html": "
Pre-defined \nText LayoutCLIP Score(↑)
TextDiffuser-10M [8]✓0.3436
GlyphControl-10M* [59]✓0.3450
Latent Diffusion Model0.3015
DeepFloyd [49]0.3267
ControlNet [63]✓0.3424
TextDiffuser-7M [8]✓0.3385
SceneTextGen-7M(Ours)0.3455
", + "image_path": "2efd9dacf5575ecf1e01b44aeb9c3614741854d46f131a561664ac26547126a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 232, + 263, + 360 + ], + "blocks": [ + { + "bbox": [ + 72, + 232, + 263, + 360 + ], + "lines": [ + { + "bbox": [ + 72, + 232, + 263, + 360 + ], + "spans": [ + { + "bbox": [ + 72, + 232, + 263, + 360 + ], + "type": "image", + "image_path": "7831a1c1e096e1d49e1b4c36fc750f3461a00c47479e296deeec5f251a9ccdbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 392 + ], + "type": "text", + "content": "Figure 5. t-SNE representation of text region embeddings derived from the penultimate layer features of a font recognition model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 402, + 286, + 498 + ], + "blocks": [ + { + "bbox": [ + 48, + 402, + 286, + 498 + ], + "lines": [ + { + "bbox": [ + 48, + 402, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 402, + 286, + 498 + ], + "type": "image", + "image_path": "8ef7b853fb4ee53bc1682a40646e8d09b64a9425fd4ffdb9d884730acfa2b6a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 499, + 228, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 499, + 228, + 510 + ], + "spans": [ + { + "bbox": [ + 104, + 499, + 228, + 510 + ], + "type": "text", + "content": "Caption: A sign that says 'Hello World'" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 557 + ], + "type": "text", + "content": "Figure 6. Diversity of font styles. SceneTextGen is able to generate visually appealing and diverse font styles and layouts for text, without any style or layout prompts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "type": "text", + "content": "tions. In both studies, we assess the Average Precision (AP) and Accuracy (AC) scores to understand how these configurations influence the model's ability to accurately recognize and generate text in images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 634, + 125, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 125, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 125, + 647 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "While SceneTextGen demonstrates superior performance in scene text image generation, challenges remain. Limited scene complexity: Models struggle to accurately generate complex elements in conjunction with text, such as a car with a 'green' sticker specifically in the back win" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 382, + 137 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 382, + 137 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 382, + 137 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 382, + 137 + ], + "type": "image", + "image_path": "41803e143ea7eb1dc0edc55fd7980a2b6c70bcd13e44856cc39fad113dccd7d7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 140, + 370, + 150 + ], + "lines": [ + { + "bbox": [ + 331, + 140, + 370, + 150 + ], + "spans": [ + { + "bbox": [ + 331, + 140, + 370, + 150 + ], + "type": "text", + "content": "TextDiffuser" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 382, + 72, + 454, + 137 + ], + "blocks": [ + { + "bbox": [ + 382, + 72, + 454, + 137 + ], + "lines": [ + { + "bbox": [ + 382, + 72, + 454, + 137 + ], + "spans": [ + { + "bbox": [ + 382, + 72, + 454, + 137 + ], + "type": "image", + "image_path": "5e545863a265564e7b10efca04de74032c86f0a9af1f22c7da2d20da0d9abb54.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 140, + 443, + 150 + ], + "lines": [ + { + "bbox": [ + 399, + 140, + 443, + 150 + ], + "spans": [ + { + "bbox": [ + 399, + 140, + 443, + 150 + ], + "type": "text", + "content": "SceneTextGen" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 163, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 304, + 163, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 547, + 208 + ], + "type": "text", + "content": "Figure 7. Distribution of visual text coordinates in the generated image w.r.t. the area of each box. As expected from real-life observations, the visual text in the ground truth images tends to be located at the center of each image." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 454, + 72, + 526, + 137 + ], + "blocks": [ + { + "bbox": [ + 454, + 72, + 526, + 137 + ], + "lines": [ + { + "bbox": [ + 454, + 72, + 526, + 137 + ], + "spans": [ + { + "bbox": [ + 454, + 72, + 526, + 137 + ], + "type": "image", + "image_path": "0c80c44c344aa77d47da102c5fb105c22f02ab4e454559c6b660bdb6a81aefb6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 140, + 515, + 150 + ], + "lines": [ + { + "bbox": [ + 473, + 140, + 515, + 150 + ], + "spans": [ + { + "bbox": [ + 473, + 140, + 515, + 150 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 218, + 408, + 259 + ], + "blocks": [ + { + "bbox": [ + 46, + 177, + 287, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 177, + 287, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 177, + 287, + 222 + ], + "type": "text", + "content": "Table 2. Comparison of CLIP scores reflecting the overall image quality generated by various models. SceneTextGen-7M achieves the highest CLIP score, indicating superior text-image alignment without relying on text layouts. * denotes LAION-Glyph dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 218, + 408, + 259 + ], + "lines": [ + { + "bbox": [ + 317, + 218, + 408, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 408, + 259 + ], + "type": "table", + "html": "
LwordLcharAPAC
✓0.50820.3109
✓0.50830.3035
✓✓0.52740.3088
", + "image_path": "3236ad3b3229c97911e4f8e31574c9ef420632755eca427c7c3ce3119e4824dd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 446, + 218, + 534, + 259 + ], + "blocks": [ + { + "bbox": [ + 446, + 218, + 534, + 259 + ], + "lines": [ + { + "bbox": [ + 446, + 218, + 534, + 259 + ], + "spans": [ + { + "bbox": [ + 446, + 218, + 534, + 259 + ], + "type": "table", + "html": "
WlamClamAPAC
120.41070.2782
0.10.20.45870.2895
0.010.020.52740.3088
", + "image_path": "39bd0e56f406fca446442d98439d11d92adf5c8066099ff626bfeb16d7545f7d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 269, + 418, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 269, + 418, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 418, + 300 + ], + "type": "text", + "content": "Table 3. Effects of loss function combinations on the MARIO-7M-Eval dataset" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 433, + 269, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 433, + 269, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 433, + 269, + 545, + 300 + ], + "type": "text", + "content": "Table 4. Impact of varying loss weights on the MARIO-7M-eval dataset" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 315, + 539, + 430 + ], + "blocks": [ + { + "bbox": [ + 310, + 315, + 539, + 430 + ], + "lines": [ + { + "bbox": [ + 310, + 315, + 539, + 430 + ], + "spans": [ + { + "bbox": [ + 310, + 315, + 539, + 430 + ], + "type": "image", + "image_path": "a3bc893c50e3f640c1674a55b5cc69176eaf23d13967ea125f84e43565c60239.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 441, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 306, + 441, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 545, + 464 + ], + "type": "text", + "content": "Figure 8. Failure cases in complex scene interpretation and processing of lengthy prompts for text." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 485, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 545, + 521 + ], + "type": "text", + "content": "dow.(Fig. 8, Case 1). Long text handling: Text accuracy and coherence decrease with increasing length (Fig. 8, Case 2). These areas require further development." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 532, + 378, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 378, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 378, + 544 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 647 + ], + "type": "text", + "content": "SceneTextGen, incorporating a character-level encoder and hierarchical text integration, offers advancements in scene text image generation. Despite improved text rendering, limitations arise in generating complex visuals and handling lengthy text. These challenges highlight the ongoing difficulty in reconciling textual accuracy with broader image synthesis. This work furthers our understanding of text-image generation, paving the way for future exploration." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 662, + 546, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 662, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 304, + 662, + 546, + 709 + ], + "type": "text", + "content": "Acknowledgements This research project has been partially funded by research grants to Dimitris N. Metaxas through NSF: 2310966, 2235405, 2212301, 2003874, and FA9550-23-1-0417." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7503" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Samaneh Azadi, Matthew Fisher, Vladimir G Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-content gan for few-shot font style transfer. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7564-7573, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 213 + ], + "type": "text", + "content": "[2] Jeonghun Baek, Geewook Kim, Junyeop Lee, Sungrae Park, Dongyoon Han, Sangdoo Yun, Seong Joon Oh, and Hwalsuk Lee. What is wrong with scene text recognition model comparisons? dataset and model analysis. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4715-4723, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 267 + ], + "type": "text", + "content": "[3] Youngmin Baek, Bado Lee, Dongyoon Han, Sangdoo Yun, and Hwalsuk Lee. Character region awareness for text detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9365-9374, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[4] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[5] Darwin Bautista and Rowel Atienza. Scene text recognition with permuted autoregressive sequence models. In European Conference on Computer Vision, pages 178-196. Springer, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 434 + ], + "type": "text", + "content": "[6] Michal Bušta, Yash Patel, and Jiri Matas. E2e-plt-an unconstrained end-to-end method for multi-language scene text. In Computer Vision-ACCV 2018 Workshops: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers 14, pages 127–143. Springer, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 513 + ], + "type": "text", + "content": "[7] Qi Chang, Zhennan Yan, Mu Zhou, Di Liu, Khalid Sawalha, Meng Ye, Qilong Zhangli, Mikael Kanski, Subhi Al'Aref, Leon Axel, et al. Deeprecon: Joint 2d cardiac segmentation and 3d volume reconstruction via a structure-specific generative method. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 567-577. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 556 + ], + "type": "text", + "content": "[8] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint arXiv:2305.10855, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 287, + 612 + ], + "type": "text", + "content": "[9] Yunhe Gao, Mu Zhou, Di Liu, Zhennan Yan, Shaoting Zhang, and Dimitris N Metaxas. A data-scalable transformer for medical image segmentation: architecture, model efficiency, and benchmark. arXiv preprint arXiv:2203.00131, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[11] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10696-10706, 2022. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 128 + ], + "type": "text", + "content": "[12] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7323-7334, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 215 + ], + "type": "text", + "content": "[13] Ligong Han, Song Wen, Qi Chen, Zhixing Zhang, Kunpeng Song, Mengwei Ren, Ruijiang Gao, Anastasis Stathopoulos, Xiaoxiao He, Yuxiao Chen, Di Liu, Qilong Zhangli, Jindong Jiang, Zhaoyang Xia, Akash Srivastava, and Dimitris Metaxas. Proxedit: Improving tuning-free real image editing with proximal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4291-4301, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 217, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 546, + 282 + ], + "type": "text", + "content": "[14] Xiaoxiao He, Chaowei Tan, Bo Liu, Liping Si, Weiwu Yao, Liang Zhao, Di Liu, Qilong Zhangli, Qi Chang, Kang Li, et al. Dealing with heterogeneous 3d mr knee images: A federated few-shot learning method with dual knowledge distillation. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 283, + 546, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 335 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, 2022. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 337, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 547, + 381 + ], + "type": "text", + "content": "[16] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. The Journal of Machine Learning Research, 23(1):2249-2281, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 382, + 546, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 546, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 546, + 435 + ], + "type": "text", + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 546, + 480 + ], + "type": "text", + "content": "[18] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "type": "text", + "content": "[19] Jindong Jiang, Fei Deng, Gautam Singh, and Sungjin Ahn. Object-centric slot diffusion. In Advances in Neural Information Processing Systems, pages 8563-8601, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 548 + ], + "type": "text", + "content": "[20] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 547, + 603 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 603, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 624 + ], + "type": "text", + "content": "[22] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 625, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 546, + 669 + ], + "type": "text", + "content": "[23] Praveen Krishnan, Rama Kovvuri, Guan Pang, Boris Vassilev, and Tal Hassner. Textstylebrush: Transfer of text aesthetics from a single example. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[24] Chen-Yu Lee and Simon Osindero. Recursive recurrent nets with attention modeling forOCR in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2231-2239, 2016. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7504" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Minghao Li, Tengchao Lv, Jingye Chen, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, and Furu Wei. Trocr: Transformer-based optical character recognition with pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13094–13102, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 172 + ], + "type": "text", + "content": "[26] Wei Li, Yongxing He, Yanwei Qi, Zejian Li, and Yongchuan Tang. Fet-gan: Font and effect transfer via k-shot adaptive instance normalization. In Proceedings of the AAAI conference on artificial intelligence, pages 1717-1724, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 228 + ], + "type": "text", + "content": "[27] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 271 + ], + "type": "text", + "content": "[28] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In Proceedings of the AAAI conference on artificial intelligence, pages 11474-11481, 2020. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 327 + ], + "type": "text", + "content": "[29] Ron Litman, Oron Anschel, Shahar Tsiper, Roeit Litman, Shai Mazor, and R Manmatha. Scatter: selective context attentional scene text recognizer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11962-11972, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 403 + ], + "type": "text", + "content": "[30] Di Liu, Yunhe Gao, Qilong Zhangli, Ligong Han, Xiaoxiao He, Zhaoyang Xia, Song Wen, Qi Chang, Zhennan Yan, Mu Zhou, et al. Transfusion: multi-view divergent fusion for medical image segmentation with transformers. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 485-495. Springer, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 470 + ], + "type": "text", + "content": "[31] Di Liu, Xiang Yu, Meng Ye, Qilong Zhangli, Zhuowei Li, Zhixing Zhang, and Dimitris N Metaxas. Deformer: Integrating transformers with deformable models for 3d shape abstraction from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14236-14246, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[32] Di Liu, Anastasis Stathopoulos, Qilong Zhangli, Yunhe Gao, and Dimitris Metaxas. Lepard: Learning explicit part discovery for 3d articulated shape reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[33] Rosanne Liu, Dan Garrette, Chitwan Sahara, William Chan, Adam Roberts, Sharan Narang, Irina Blok, RJ Mical, Mohammad Norouzi, and Noah Constant. Character-aware models improve visual text rendering. arXiv preprint arXiv:2212.10562, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[34] Jianqi Ma, Weiyuan Shao, Hao Ye, Li Wang, Hong Wang, Yingbin Zheng, and Xiangyang Xue. Arbitrary-oriented scene text detection via rotation proposals. IEEE transactions on multimedia, 20(11):3111-3122, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[35] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coher-ently. arXiv preprint arXiv:2303.17870, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[36] Carlos Martin-Isla, Víctor M Campello, Cristian Izquierdo, Kaisar Kushibar, Carla Sendra-Balcells, Polyxeni Gkontra, Alireza Sojoudi, Mitchell J Fulton, Tewodros Weldebirhan Arega, Kumaradevan Punithakumar, et al. Deep learning segmentation of the right ventricle in cardiac mri: The" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "m&ms challenge. IEEE Journal of Biomedical and Health Informatics, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[37] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 196 + ], + "type": "text", + "content": "[38] Siyang Qin, Alessandro Bissacco, Michalis Raptis, Yasuhisa Fujii, and Ying Xiao. Towards unconstrained end-to-end text spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4704-4714, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 198, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 262 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[40] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485-5551, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 375 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 432 + ], + "type": "text", + "content": "[42] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 478 + ], + "type": "text", + "content": "[43] Roi Ronen, Shahar Tsiper, Oron Anschel, Inbal Lavi, Amir Markovitz, and R Manmatha. Glass: Global to local attention for scene-text spotting. In European Conference on Computer Vision, pages 249-266. Springer, 2022. 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "text", + "content": "[44] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 545, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 611 + ], + "type": "text", + "content": "[45] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems, 35:36479-36494, 2022. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[46] Lohrasb Ross Sayadi, Usama S Hamdan, Qilong Zhangli, and Raj M Vyas. Harnessing the power of artificial intelligence to teach cleft lip surgery. Plastic and Reconstructive Surgery-Global Open, 10(7):e4451, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[47] Baoguang Shi, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Robust scene text recognition with automatic rectification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4168-4176, 2016. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7505" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[48] Baoguang Shi, Mingkun Yang, Xinggang Wang, Pengyuan Lyu, Cong Yao, and Xiang Bai. Aster: An attentional scene text recognizer with flexible rectification. IEEE transactions on pattern analysis and machine intelligence, 41(9):2035-2048, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[49] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd-if. https://github.com/deep-floyd/IF, 2023. GitHub Repository. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[50] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiaoyuan Jing, Fei Wu, and Bingkun Bao. Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[51] Peng Wang, Hui Li, and Chunhua Shen. Towards end-to-end text spotting in natural scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7266-7281, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 307 + ], + "type": "text", + "content": "[52] Peng Wang, Cheng Da, and Cong Yao. Multi-granularity prediction for scene text recognition. In European Conference on Computer Vision, pages 339-355. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 287, + 363 + ], + "type": "text", + "content": "[53] Song Wen, Hao Wang, Di Liu, Qilong Zhangli, and Dimitris Metaxas. Second-order graph odes for multi-agent trajectory forecasting. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5101-5110, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 409 + ], + "type": "text", + "content": "[54] Liang Wu, Chengquan Zhang, Jiaming Liu, Junyu Han, Jingtuo Liu, Errui Ding, and Xiang Bai. Editing text in the wild. In Proceedings of the 27th ACM international conference on multimedia, pages 1500-1508, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 454 + ], + "type": "text", + "content": "[55] Ziyi Wu, Jingyu Hu, Wuyue Lu, Igor Gilitschenski, and Animesh Garg. Slotpdiffusion: Object-centric generative modeling with diffusion models. Advances in Neural Information Processing Systems, 36:50932-50958, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 521 + ], + "type": "text", + "content": "[56] Zhaoyang Xia, Yuxiao Chen, Qilong Zhangli, Matt Huener-fauth, Carol Neidle, and Dimitris Metaxas. Sign language video anonymization. In Proceedings of the LREC2022 10th Workshop on the Representation and Processing of Sign Languages: Multilingual Sign Language Resources, Marseille, France, 25 June 2022, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "text", + "content": "[57] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 634 + ], + "type": "text", + "content": "[58] Qiangpeng Yang, Jun Huang, and Wei Lin. Swaptext: Image based texts transfer in scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14700-14709, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 679 + ], + "type": "text", + "content": "[59] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[60] Jiachi Ye, Haoyan Kang, Hao Wang, Chen Shen, Belal Jahannia, Elham Heidari, Nvid Asadizanjani, Mohammad-Ali Miri, Volker J Sorger, and Hamed Dalir. Demultiplexing oam" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 485 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "beams via fourier optical convolutional neural network. In Laser Beam Shaping XXIII, pages 16-33. SPIE, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 162 + ], + "type": "text", + "content": "[61] Jiachi Ye, Maria Solyanik, Zibo Hu, Hamed Dalir, Behrouz Movahhed Nouri, and Volker J Sorger. Free-space optical multiplexed orbital angular momentum beam identification system using fourier optical convolutional layer based on 4f system. In Complex Light and Optical Forces XVII, pages 70-80. SPIE, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 206 + ], + "type": "text", + "content": "[62] Fangneng Zhan, Hongyuan Zhu, and Shijian Lu. Spatial fusion gan for image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3653-3662, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 251 + ], + "type": "text", + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 307 + ], + "type": "text", + "content": "[64] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 374 + ], + "type": "text", + "content": "[65] Qilong Zhangli, Jingru Yi, Di Liu, Xiaoxiao He, Zhaoyang Xia, Qi Chang, Ligong Han, Yunhe Gao, Song Wen, Haiming Tang, et al. Region proposal rectification towards robust instance segmentation of biological images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "text", + "content": "[66] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 485 + ], + "type": "text", + "content": "[67] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7506" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_content_list.json b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7600c06b650e24b4f765934dd25cfc940f0cc91 --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_content_list.json @@ -0,0 +1,1537 @@ +[ + { + "type": "text", + "text": "LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding", + "text_level": 1, + "bbox": [ + 83, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Min Liang, Jia-Wei Ma, Xiaobin Zhu*, Jingyan Qin, Xu-Cheng Yin \nUniversity of Science and Technology Beijing", + "bbox": [ + 218, + 180, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lm20200126, mjw20151001}@@hotmail.com, {zhuxiaobin, xuchengyin}@ustb.edu.cn, qinjingyanking@foxmail.com", + "bbox": [ + 150, + 219, + 815, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing scene text detectors generally focus on accurately detecting single-level (i.e., word-level, line-level, or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts, detecting multi-level texts while exploring their contextual information is critical. To this end, we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection, which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer, WordDecoder, LineDecoder, and ParaDecoder are proposed to be responsible for word-level text prediction, line-level text prediction, and paragraph-level text prediction, respectively. Meanwhile, WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships, respectively. In addition, we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection.", + "bbox": [ + 75, + 318, + 473, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 695, + 209, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reading and understanding texts in scene images and digital documents is important in various real-world applications, including visual recognition [40], scene understanding [4], and text-based VQA [1, 30]. Compared to words, phrases and sentences express more semantic messages, and paragraphs convey richer contextual semantic messages. For better reading and understanding texts in scenes, it is necessary to simultaneously detect multi-level texts (i.e., word-", + "bbox": [ + 75, + 720, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7530836704e0217b162f3083a33106d8f66ccbf3885507e0d4da6e265955ca93.jpg", + "image_caption": [ + "Figure 1. Top: LayoutFormer is introduced for hierarchical text detection, which simultaneously detects multi-level texts (i.e., word-level, line-level, and paragraph-level) and predicts their geometric layouts. Bottom: Pipeline comparisons of (a) single-level detector, (b) two-level detector, and (c) hierarchical text detector." + ], + "image_footnote": [], + "bbox": [ + 504, + 285, + 890, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "level for words, line-level for phrases and sentences, and paragraph-level for paragraphs) and predict their geometric layouts [16, 44], i.e., hierarchical text detection [25]. Hierarchical texts can deliver complete and meaningful text messages, which is beneficial for various downstream tasks.", + "bbox": [ + 496, + 638, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although existing scene text detectors have achieved promising performance, they [32, 47] often concentrate on detecting single-level text, which decodes the encoded features of an input image into their representation of independent text entities, e.g., words, lines, or paragraphs. As shown in Fig. 1 (a), single-level detectors only predict one level of text results and do not explore the contextual relationship among different levels of texts. This makes them unable to deliver more informative messages for facilitating scene text understanding. Recently, some methods have investigated two-level text detection. A schematic of the framework is shown in Fig. 1 (b). CUTE [43] proposes a", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. This work was supported by National Key Research and Development Program of China (2020AAA0109701), National Science Fund for Distinguished Young Scholars (62125601), and National Natural Science Foundation of China (62076024, 62172035).", + "bbox": [ + 75, + 851, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15665", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two-stage network, which first adopts a Transformer structure like DETR [3] to perform word/line bounding boxes and then directly learns the sequential relationships between detected text boxes based on visual features after ROI operation and positional information. Unified Detector [24] proposes a unified scene text detection and layout analysis network, which first uses a Transformer-based structure [35] to generate word/line mask maps and then models their affiliate relationships. These two methods are all supervised by an affinity matrix or index label, directly modelling relationships on the word-level/line-level detection results. They only rely on the former detection results to model contextual relationships, making the model's performance highly limited by the text detection results at the first stage. Moreover, this paradigm cannot be easily extended to the task of text detection at more than two levels.", + "bbox": [ + 75, + 90, + 472, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing scene text detectors generally adopt multi-scale feature maps that contain rich information for accurate prediction. However, methods based on the Transformer encoder-decoder structure often suffer from huge computational and memory costs on high-resolution feature maps. This is because the attention weight computation in the Transformer decoder is of linear computation w.r.t. pixel numbers. There have been several work to tackle the issues. Deformable DETR [51] introduces the multi-scale deformable attention module to select a small set of key locations. AdaMixer [10] adaptively samples features over the space and scales of an object. However, these methods are based on estimated offsets of instance coordinates, not applicable to segmentation methods. In image segmentation, Mask2Former [6] claims that local features can well update query features and proposes to use image features of foreground regions. TSP [31] adopts a RoIAlign operation to extract features of RoIs. However, these methods still suffer from complex operations. We think that it is enough to update query features with only foreground features of a few representative pixels.", + "bbox": [ + 75, + 335, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a hierarchical text detector named LayoutFormer, which can simultaneously produce hierarchical text detection boxes and their layout relationships. Unlike using inter-level relationships as supervision, we directly adopt multi-level text boxes as supervision, which can significantly reduce the dependence on the former text detection results and implicitly exploit multi-level textual information. In LayoutFormer, we propose three Transformer-based modules, i.e., WordDecoder, LineDecoder, and ParaDecoder, for detecting word-level texts, line-level texts, and paragraph-level texts, respectively. We first adopt text line as the detection unit, and LineDecoder updates and forms line query features. Because a line can be split into several words, WordDecoder learns the deviation of each word within a line based on the line query features and forms final word query features. ParaDecoder aggre", + "bbox": [ + 75, + 659, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gates line query features by learning line-paragraph relationships and forms final paragraph query features. In addition, we propose a Prior Location Sampler, which adaptively selects a few representative features that are highly relevant to scene texts. The sampled features are used for cross-attention operation to update query features. Prior Location Sampler can improve the model performance while reducing the training cost. Extensive experiments verify the superior performance on hierarchical text detection.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 517, + 233, + 821, + 247 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an innovative hierarchical text detector that simultaneously detects multi-level text instances and explores their contextual information for predicting the geometric layouts, finally performing hierarchical outputs.", + "- We propose WordDecoder, LineDecoder, and ParaDecoder in LayoutFormer to decode word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively.", + "- We propose a Prior Location Sampler to adaptively select a few representative foreground features for updating queries, which can significantly reduce the computational cost while improving the detection performance.", + "- Extensive experiments on multiple publicly available datasets verify the state-of-the-art performance of our LayoutFormer." + ], + "bbox": [ + 500, + 253, + 890, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 523, + 640, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Text Detection", + "text_level": 1, + "bbox": [ + 500, + 554, + 694, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning-based scene text detection methods have been widely investigated and roughly divided into bottom-up and top-down methods. The bottom-up methods generally detect local components [23, 29, 33] or text pixels [19, 37] for grouping text instances. TextSnake [23] describes a text instance as a sequence of ordered, overlapping disks. The pixel-based methods use some auxiliary information (e.g., similarity vectors in PAN [38], and threshold map in DB [19]) or post-processings (e.g., a progressive scale expansion algorithm in PSENet [37]) to generate text instances. The top-down methods [18, 50] directly predict bounding boxes of scene text instances. To detect arbitrary-shaped texts, some methods [39, 49, 52] localize the key points on the contours of scene texts. Recently, DETR [3] presents a Transformer-based architecture for object detection and achieved great success. Inspired by it, many researchers [24, 32, 47] apply the Transformer structure to scene text detection. Overall, existing scene text detection methods mainly address single-level text detection, ignoring the text detection of multiple levels and further predicting the hierarchical relationships.", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15666", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fa45a20519ffc544cc0c89557cd32075adb119f536671e24b81d9b556f7031c9.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed LayoutFormer, which mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders (WordDecoder, LineDecoder, and ParaDecoder) and three predictors. $M_l$ denotes predicted mask maps of text lines by $l$ -th Transformer decoder, and $F_l$ denotes the corresponding image features generated by $M_l$ . We first predict text lines, and then predict words and paragraphs. In testing, word-level results produce line-level results based on word-line relationships, and then paragraph-level results are produced based on line-paragraph relationships, finally obtaining hierarchical text detection results." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 867, + 342 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/44d73bab8c299b7d593fc88bf3c73480d904b7b9fc18fa672c11ed04f3b8f9e7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Detection UnitQueryMemory/GBPQ
word25612.153.09
line25610.657.48
paragraph2568.540.34
", + "bbox": [ + 102, + 448, + 444, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparisons of word, line, and paragraph as text detection unit on the HierText validation set, respectively. \"Query\" denotes the number of text queries. \"Memory\" denotes GPU Memory consumed in training.", + "bbox": [ + 75, + 525, + 470, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Layout Analysis", + "text_level": 1, + "bbox": [ + 76, + 606, + 238, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Document Layout analysis has made great progress, which can be mainly divided into object detection-based methods [17, 27, 34] and segmentation-based methods [2, 16, 41]. Inspired by object detection [15, 28] and semantic segmentation [5, 12, 22], these methods treat semantically coherent text blocks as a special kind of object. However, they fail to produce word or line-level detections and can only be used in companies with standalone text detectors, increasing the complexity of the pipeline. Some work [36] takes a hierarchical view and applies graph-based models on the finest granularity, i.e., individual words, to analyze the layout.", + "bbox": [ + 75, + 628, + 468, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, some work has investigated layout analysis in scene text detection. CUTE [43] detects contextual text blocks consisting of one or multiple ordered integral text units. Their model is a two-stage structure, which first detects integral text units and then models the relationship of integral text units by a designed relation module. Unified Detector [24] introduces a novel task of unified scene text", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "detection and layout analysis, which models the two closely related tasks with a unified model. It produces mask outputs for the text detection task and an affinity matrix for the layout analysis task. The above methods are highly dependent on text detection results at the first stage. Differently, we use multi-level text boxes as supervision to learn the relationship among different levels for layout analysis.", + "bbox": [ + 496, + 450, + 893, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 575, + 635, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 603, + 629, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "How to effectively model hierarchical text detection tasks? Is the top-down or bottom-up approach better? As we know, the text layouts in real scene images is highly complex, bringing significant challenges to hierarchical text detection. According to experiments and analyses, we found that paragraph-level text boxes are considerably irregular and always contain much background, which makes existing scene text detection approaches difficult to model text instances. For example, as shown in the second row of Fig. 3, a paragraph text box annotates the entire image. As listed in Tab. 1, the detection performance of paragraphs is much lower than that of words and lines. Thus, the top-down modeling approach is considered inappropriate. Both word and line are generally used as text detection units. However, text lines always contain more complete semantic information than words. The aspect ratios and sizes of words are considerably more variable than text lines, making detection more challenging. Meanwhile, if the word is used", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15667", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/78c6e6b056116af6e2270121115adc3b256ccb815d7f857326ba933bde754c8b.jpg", + "image_caption": [ + "Figure 3. Some examples of word-level (first column), line-level (second column), and paragraph-level (third column) text boxes." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 452, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as the detection unit, the detection error will accumulate bigger after two aggregation processes. Hence, we finally choose line-level as our detection unit, generate word-level texts by splitting them downward, and get paragraph-level texts by aggregating them upward.", + "bbox": [ + 75, + 311, + 470, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network Architecture", + "text_level": 1, + "bbox": [ + 76, + 397, + 282, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of our LayoutFormer is illustrated in Fig. 2. It mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders and three predictors, respectively, corresponding to word-level text detection, line-level text detection and paragraph-level text detection. Specifically, given an input image $I \\in R^{H \\times W \\times 3}$ , the feature encoder with a ResNet-50 [14] backbone and a Transformer encoder [6] extracts and enhances the features and then generates multi-scale feature maps, i.e., $C_4 \\in R^{\\frac{H}{32} \\times \\frac{W}{32} \\times C}$ , $C_3 \\in R^{\\frac{H}{16} \\times \\frac{W}{16} \\times C}$ , $C_2 \\in R^{\\frac{H}{8} \\times \\frac{W}{8} \\times C}$ and $C_1 \\in R^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}$ . The typical value we use is $C = 256$ . Then, Prior Location Sampler adaptively selects image features for the Transformer decoder. Afterwards, the LineDecoder, WordDecoder, and ParaDecoder update line query features $X_{line} \\in R^{N_{line} \\times C}$ , word query features $X_{word} \\in R^{N_{word} \\times C}$ , and paragraph query features $X_{para} \\in R^{N_{para} \\times C}$ , respectively. The three predictors convert $X_{line}$ , $X_{word}$ , and $X_{para}$ to line-level prediction $(P_{line}, M_{line})$ , word-level prediction $(P_{word}, M_{word})$ , and paragraph-level prediction $(P_{para}, M_{para})$ , where $P_* \\in R^{N_* \\times 2}$ is text/non-text probability values of $N_*$ instances and $M_* \\in R^{N_* \\times \\frac{H}{4} \\times \\frac{W}{4}}$ is mask maps of $N_*$ instances.", + "bbox": [ + 75, + 420, + 472, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Prior Location Sampler", + "text_level": 1, + "bbox": [ + 76, + 763, + 295, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask2Former [6] proposes masked attention to attend within the foreground region of the predicted mask for each query, which can be computed as", + "bbox": [ + 75, + 787, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {l} = \\operatorname {s o f t m a x} \\left(B _ {l - 1} + Q _ {l} K _ {l} ^ {T}\\right) V _ {l} + X _ {l - 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 844, + 468, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l$ is the layer index, $B_{l-1}$ is the binary mask prediction of $l-1$ -th layer, $X_{l}$ is query features at the $l$ -th layer, $Q_{l}$", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5666002bc55488412c0a002c68d73966bf59fafa8a2098c8fda747d502823a03.jpg", + "image_caption": [ + "Figure 4. Illustration of the Prior Location Sampler, which selects representative foreground features with Top- $K$ scores in predicted mask maps and then aggregates with weights." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 883, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is query features $X_{l-1}$ under linear transformation, $K_l$ and $V_l$ are the input image features under linear transformation. $X_0$ denotes input query features to the Transformer decoder. $B_0$ is the binary mask prediction obtained from $X_0$ . The input image features are the entire multi-scale feature maps from a feature encoder. Their computational complexity of Eq. (1) is $O(NH'W'C)$ , where $H'$ and $W'$ are the height and width of a specific feature map respectively, and $N$ is the number of queries. The cross-attention operation suffers from a linear complexity growth with the spatial size of feature maps. Moreover, the main accompanying problem is the high demand for computing resources.", + "bbox": [ + 496, + 310, + 892, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Thus, we propose the Prior Location Sampler to select representative foreground features with Top- $K$ scores in mask maps and obtain the image features $F_{l}\\in R^{K\\times N\\times C}$ for $l$ -th decoder layer. Our computational complexity of Eq. (1) is $O(NKC)$ . In our implementation, the value $K\\ll H^{\\prime}W^{\\prime}$ , and thus the complexity of our cross-attention operation could be significantly reduced. The computational resource consumption due to an increase in the number of feature maps or an increase in the scale of feature maps is negligible.", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the Prior Location Sampler is illustrated in Fig. 4. Specifically, $M_{l}$ denotes the predicted mask maps of $l$ -th decoder layer, and", + "bbox": [ + 496, + 643, + 890, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l} = \\left\\{M _ {l n} \\in R ^ {\\frac {H}{4} \\times \\frac {W}{4}} | n = 1, 2, \\dots , N \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 695, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each $M_{ln}$ , we use 0.5 for thresholding. We sort scores in $M_{ln}$ and select features with Top- $K$ scores [13] in multi-scale feature maps $\\{C_i|i = 1,2,3,4\\}$ respectively, where $f_{K}(\\cdot)$ is sampling operation. Then, we assign [20, 28] text ROIs of different scales to the pyramid levels for calculating weights [10] as Eq. (4), where $S_{ln}$ is the cumulative sum of the binarized $M_{ln}$ , i.e., the area of a candidate text, and 56 [20] is a scaling factor. The selected features for $M_{ln}$ are gathered into $F_{ln}\\in R^{K\\times C}$ :", + "bbox": [ + 496, + 719, + 893, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l n} = \\sum_ {i = 1} ^ {4} w _ {l i} f _ {K} \\left(C _ {i}, M _ {l n}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 863, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15668", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {l n} = \\left\\lfloor \\log_ {2} \\frac {S _ {l n}}{5 6} \\right\\rfloor , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 113, + 468, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {l} = \\text {s o f t m a x} \\left[ - \\frac {\\left(t _ {l n} - \\iota + 1\\right) ^ {2}}{2} \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 148, + 383, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final features $F_{l}\\in R^{K\\times N\\times C}$ are denoted as:", + "bbox": [ + 99, + 181, + 423, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l} = \\left\\{F _ {l n} \\mid n = 1, 2, \\dots , N \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 210, + 467, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Transformer Decoder", + "text_level": 1, + "bbox": [ + 76, + 233, + 279, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LineDecoder. We feed the selected image features $F_{l-1} \\in R^{K \\times N_{line} \\times C}$ and $N_{line}$ learnable line queries into LineDecoder, where $K = 32$ . LineDecoder uses the standard Transformer decoder structure, consisting of self-attention, cross-attention, and feedforward network. We also add the positional embeddings to queries and keys at every self-attention and cross-attention layer. We refine line query features layer-by-layer by setting $L_1 = 4$ . The updated line query features are adopted to predict $l$ -th mask maps, which are fed into Prior Location Sampler to generate image features for the next decoder layer.", + "bbox": [ + 76, + 255, + 467, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "WordDecoder. We frist hypothesize that there are $N_{per\\_line}$ words in each line and initialize $N_{per\\_line}$ learnable word queries. Thus, the number of words in an image can be formulated as the product of the number of lines in an image and words per line, as", + "bbox": [ + 76, + 422, + 467, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN _ {\\text {w o r d}} = N _ {\\text {l i n e}} \\times N _ {\\text {p e r l i n e}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 508, + 467, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "So, we can decouple the self-attention between words into self-attention between lines and self-attention between words per line. Based on line queries from LineDecoder, we only need to compute self-attention between $N_{per\\_line}$ , greatly reducing the computational cost. Then we obtain word query features $X_{word} \\in R^{N_{word} \\times C}$ by", + "bbox": [ + 76, + 532, + 467, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {w o r d}} = X _ {\\text {l i n e}} X _ {\\text {p e r - l i n e}} ^ {T}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 632, + 467, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{line} \\in R^{N_{line} \\times 1 \\times C}$ is line query features and $X_{per\\_line} \\in R^{N_{per\\_line} \\times 1 \\times C}$ is learnable word query features per line. Then cross-attention followed by feedforward network adopts image features $F_{l} \\in R^{K \\times N_{line} \\times C}$ with $K = 64$ to update $X_{word}$ . The image features are obtained by Prior Location Sampler, which feeds the predicted mask maps of the last decoder layer in LineDecoder. In our experiments, we set $L_{3} = 1$ . With our word decoupling mechanism, every $N_{per\\_line}$ queries predicts words within a text line for capturing word-line relationships.", + "bbox": [ + 76, + 657, + 467, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ParaDecoder. We adaptively learn line-paragraph relationships and aggregate line query features for paragraph-level texts to get paragraph query features. We initialize $N_{para}$ learnable paragraph queries. ParaDecoder also consists of self-attention, cross-attention, and feedforward networks. It uses $X_{line}$ line query features as inputs of cross attention", + "bbox": [ + 76, + 810, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to update paragraph queries. Positional embeddings are added to queries and keys at every self-attention and cross-attention layer. In our experiments, we adopt $X_{line}$ of last decoder layer in LineDecoder and set $L_{2} = 1$ . In this process, the learned attention weights $(R^{N_{para} \\times N_{line}})$ in cross-attention of ParaDecoder are line-paragraph relationships.", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optimization", + "text_level": 1, + "bbox": [ + 500, + 190, + 637, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. Instead of relation matrices, we use multi-level text boxes as supervision. The total loss is formulated as:", + "bbox": [ + 500, + 214, + 890, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {l i n e}} + L _ {\\text {w o r d}} + L _ {\\text {p a r a}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 256, + 890, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{line}$ , $L_{word}$ , and $L_{para}$ respectively denote losses of line-level, word-level, and paragraph-level, which can be formulated as:", + "bbox": [ + 498, + 282, + 890, + 327 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {*} = \\lambda_ {c l s} L _ {c l s} \\left(P _ {*}, P _ {*} ^ {\\prime}\\right) + \\lambda_ {m a s k} L _ {m a s k} \\left(M _ {*}, M _ {*} ^ {\\prime}\\right), \\\\ * \\in \\{w o r d, l i n e, p a r a \\}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 339, + 890, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{*}$ and $P_{*}^{\\prime}$ are the class predictions and their corresponding ground truth, $M_{*}$ and $M_{*}^{\\prime}$ are the mask predictions and their corresponding ground truth, $L_{cls}$ denotes binary cross-entropy loss, $L_{mask}$ is the sum of binary cross-entropy loss and dice loss. We set $\\lambda_{cls} = 2.0$ for predictions matched with ground truth and 0.1 for unmatched predictions. $\\lambda_{mask}$ is set to 5.0.", + "bbox": [ + 498, + 386, + 890, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. We present a simple inference procedure that converts class predictions and mask predictions to text boxes, which mainly consists of four steps: (1) the text queries whose class scores are over 0.3 will be selected; (2) the mask predictions of selected queries are binarized by 0.5 to get the binary maps; (3) the connected regions are obtained from the binary maps; (4) the text queries whose averaged foreground mask probability is below 0.9 will be filtered. For hierarchical text detection task, we frist generate the word-level detection boxes by the above inference procedure. Then, we use word-line relationships from WordDecoder to form a word-line geometric layouts and further use line-paragraph relationships from ParaDecoder to form a line-paragraph geometric layouts, finally performing hierarchical text detection outputs.", + "bbox": [ + 498, + 492, + 890, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 731, + 630, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 756, + 601, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HierText [24]. It is a hierarchical text detection dataset consisting of 8,281 training images, 1,724 validation images, and 1,634 testing images. It annotates images hierarchically, which first annotates word locations with polygons, then clusters words into lines and lines into paragraphs.", + "bbox": [ + 498, + 779, + 890, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MSRA-TD500 [45]. It is a line-level annotated arbitrary-oriented long text dataset. It consists of 300 training images and 200 testing images collected from natural scenes.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15669", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5fdb9239e714b82a272dc22fbdd47512491349e399b169b1ab5799d043d8fc1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTotal-TextHierText-Line
RPFMemory/GBRPFTPQMemory/GB
C2-C4masked attention84.6887.7986.2115.365.5682.7373.1577.5756.7418.1
C2-C4Top-K&concat84.6687.8086.2015.165.3482.5672.9577.3756.4418.3
C1-C4Top-K&concat84.4688.1586.2716.265.8882.9673.4477.5356.9420.3
C1-C4Top-K&weight85.0789.2687.1214.266.6584.1074.3777.2957.4817.0
", + "bbox": [ + 99, + 88, + 867, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CTW1500 [21]. It is an arbitrary-shaped scene text dataset that consists of 1,000 training images and 500 testing images. In this dataset, the annotations of text instances are line-level and labelled by a polygon with 14 key points.", + "bbox": [ + 75, + 253, + 467, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total-Text [7]. It is an arbitrary-shaped scene text dataset that contains 1,255 training and 300 testing images, which are annotated by the word-level polygon with 14 key points.", + "bbox": [ + 76, + 314, + 467, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 369, + 294, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiments, for HierText, we only use the training images of HierText to train the models by $60k$ steps for single-level text detection and $80k$ steps for multi-level text detection. In testing, we keep the aspect ratio of testing images and resize them to 1,120 height. For MSRA-TD500, CTW1500, and Total-Text, we adopt SynthText [11] to pretrain the models by $150k$ steps. Then, we finetune the models on the corresponding real-world datasets by $40k$ steps. In testing, we keep the aspect ratio of testing images and resize them to 800 height on three datasets.", + "bbox": [ + 75, + 393, + 467, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In training, we use AdamW [26] optimizer and the step learning rate schedule with an initial learning rate of 0.0001 and a weight decay of 0.05. A learning rate multiplier of 0.1 is applied to the ResNet-50 backbone. We decay the learning rate at 0.9 and 0.95 fractions of the total number of training steps by a factor of 10. We train our models with a batch size of 8. Data augmentation includes: (1) Randomly Flipping; (2) Randomly rotate them in range $(-10^{\\circ}$ to $10^{\\circ})$ ; (3) Randomly resize them in range (0.5 to 3.0); (4) Randomly cropping. Finally, we resize the images to $640\\times 640$", + "bbox": [ + 75, + 545, + 467, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 76, + 705, + 261, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For MSRA-TD500, CTW1500, and CTW1500 datasets, we follow the standard evaluation protocol Recall (R), Precision (P), and F-measure (F). For the HierText dataset, we follow [24] and adopt Panoptic Quality (PQ) as the main evaluation metric for the hierarchical text detection task.", + "bbox": [ + 75, + 729, + 467, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 816, + 240, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$K$ value in Prior Location Sampler. We compare the different Top- $K$ values in Prior Location Sampler. Results are listed in Tab. 3. Selecting too few foreground points may fail to capture and learn the features of an entire text. Select", + "bbox": [ + 75, + 839, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6092ec11d1023c36658a3c1e3f9c8762227c93482510ba54550d88385ae4bcf.jpg", + "table_caption": [ + "Table 2. Ablation comparison of Prior Location Sampler on the Total-Text and line-level of the HierText validation set. \"Memory\" denotes GPU Memory consumed in training. $C_1 - C_4$ denotes the multi-scale feature maps used in the Transformer decoder. The results show that \"Top-K&weight\" with $C_1 - C_4$ achieves a considerable improvement and consumes fewer training resources than other methods." + ], + "table_footnote": [], + "table_body": "
KTotal-TextHierText-Line
RPFRPFPQ
1684.3488.0186.1465.3481.1272.3855.54
3285.0789.2687.1266.6584.1074.3757.48
6484.7588.6386.6468.6784.4475.7558.86
", + "bbox": [ + 506, + 251, + 887, + 329 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation comparison of Top- $K$ value on Total-Text and line-level of the HierText validation set.", + "bbox": [ + 500, + 338, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78a97c23a2766b7a8dde7b68011fdf84741510fb585cb31ff5f4369067463a0a.jpg", + "image_caption": [ + "Figure 5. Visualization of Top- $K$ sampling points in Prior Location Sampler." + ], + "image_footnote": [], + "bbox": [ + 504, + 378, + 885, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing too many foreground points may introduce noise, which is detrimental to the performance of the model. For Total-Text, we set $K$ to 32, and the same for MSRA-TD500 and CTW1500. While text lines of HierText may have larger aspect ratios and scales, the performance of $K = 64$ is better than that of $K = 32$ . However, we finally set $K$ to 32 for a trade-off between performance and efficiency.", + "bbox": [ + 496, + 522, + 890, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation of Prior Location Sampler. We evaluate the effectiveness of the Prior Location Sampler on the Total-Text and HierText validation set. Results are listed in Tab. 2. \" $C_1 - C_4$ \" denotes that multi-scale feature maps $\\{C_i | i = 1, 2, 3, 4\\}$ are used in Transformer decoder. \"masked attention\" is followed from Mask2Former [6], which calculates the attention matrix by additionally adding a binary mask prediction. \"Top-K&concat\" applies Top-K strategy to multi-scale feature maps and then performs concatenation. \"Top-K&weight\" applies Top-K strategy to multiscale feature maps and then element-wise sums in weights. Inspired by Mask2Former, the first row of Tab. 2 feeds the entire feature map into the Transformer decoder. The experimental results show that \"Top-K&weight\" apparently achieves a significant improvement in harmonic accuracy and memory. Meanwhile, due to using a small number of foreground points, our method consumes fewer training resources (GPU Memory) than other methods.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15670", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c38f98e61ec1d65aaad8b4753c179bbf65f7ab11eaa363d364400db5fe2d0fce.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodWordLineParagraph
RPFTPQRPFTPQRPFTPQ
LayoutFormer57.0475.2564.8975.8849.2450.3976.3260.7076.1246.2145.8973.1456.4076.1942.97
", + "bbox": [ + 127, + 89, + 844, + 143 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/108a76d52945c395ad9f4948118e249a50bfcf43bd0b650a862bedc638e55c91.jpg", + "table_caption": [ + "Table 4. Experimental results of word-line-paragraph three-level text detection on the HierText testing set." + ], + "table_footnote": [], + "table_body": "
MethodNperLineNlineWordLine
PQPQ
In the below experiments, we always set L3 to 1.
Xline1012044.0236.35
Fl1012050.2948.60
In the below experiments, we use Fl as image features.
L3 = 11012050.2948.60
L3 = 41012051.0744.05
L3 = 1612046.2139.60
L3 = 1620050.1246.00
", + "bbox": [ + 89, + 191, + 459, + 338 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/098a04bdafce0b0722de2016f0d27027a9cbe3e2a731f48bf39db870b935c4bd.jpg", + "table_caption": [ + "Table 5. Ablation comparison of WordDecoder on the HierText validation set. $N_{per\\_line}$ and $N_{line}$ denote the number of word queries in each line and text line queries, respectively." + ], + "table_footnote": [], + "table_body": "
MethodNlineNparaLineParagraph
PQPQ
In the below experiments, we always set L2 to 1.
Xline12812858.3151.98
Fl12812857.2746.15
", + "bbox": [ + 91, + 405, + 455, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation of WordDecoder. We conduct ablation experiments by comparing different settings of image features, $L_{3}$ , $N_{per\\_line}$ and $N_{line}$ in WordDecoder. Results are listed in Tab. 5. Using $F_{l}$ as image features for updating word queries is better than $X_{line}$ . This is because to identify different text lines, $X_{line}$ extracts consistent features of each text line, whereas $X_{per\\_line}$ needs to learn differentiated features within a text line to identify different words within a line. In addition, we find that $L_{3}$ being set to 1 performs better than being set to 4. Thus, we adopt $L_{3} = 1$ by default. We also conduct comparative experiments with different $N_{per\\_line}$ and $N_{line}$ . We can conclude that the larger the query number, the higher the performance.", + "bbox": [ + 75, + 564, + 468, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation of ParaDecoder. We compare different image features for ParaDecoder. Results are listed in Tab. 6. Adopting $X_{line}$ as image features is better than $F_{l}$ . We think that after the learning of LineDecoder, $X_{line}$ is able to characterize text lines well.", + "bbox": [ + 75, + 761, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 76, + 845, + 447, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hierarchical Text Detection. In this section, we evaluate the performance on the HierText testing set for hierarchical", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/17613fa338d6ddb5a1e3a716e06eb6613ded099256db7bdeaee9b37041fdf23f.jpg", + "table_caption": [ + "Table 6. Ablation comparison of ParaDecoder on the HierText validation set. $N_{line}$ and $N_{para}$ denote the number of text line queries and paragraph queries, respectively." + ], + "table_footnote": [], + "table_body": "
MethodNwordNlineWordLine
PQPQ
LayoutFormer120012050.3548.62
", + "bbox": [ + 516, + 191, + 877, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Experimental results of word-line two-level text detection on the HierText testing set. $N_{word}$ and $N_{line}$ denote the number of word queries and text line queries, respectively.", + "bbox": [ + 498, + 255, + 890, + 299 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fa2316c2eec527dfe017818d4a7cdec7c32eb241944e717dded40c382d86115.jpg", + "image_caption": [ + "Figure 6. The hierarchical text detection examples on the HierText. The visualization images in each row are in turn the input image, word-level detection results, line-level detection results, and paragraph-level detection results." + ], + "image_footnote": [], + "bbox": [ + 506, + 311, + 887, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "text detection. To our knowledge, few work has conducted research on hierarchical text detection. Hence, we mainly compare with [24]. Results are listed in Tab. 7, Tab. 9, and Tab. 4. Meanwhile, we demonstrate hierarchical detection results on images in Fig. 6. Both qualitative and quantitative verify the advances of our method. And we can observe that our method can precisely detect hierarchical texts.", + "bbox": [ + 496, + 628, + 890, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For word-line detection, our results are listed in Tab. 7. Because there is no any prior research on this, we provide a baseline to facilitate comparisons with other methods. Benefiting from our decoupling operation on words, our $N_{word}$ can reach 1,200, much higher than Unified Detector [24].", + "bbox": [ + 496, + 734, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For line-paragraph detection, our LayoutFormer achieves the best text detection and layout analysis performance. Results are listed in Tab. 9. \"GCP API\", \"GCN-PP\", \"Mask-RCNN-Cluster\", \"Max-DeepLab-Cluster\" are two-stage approaches, while \"Unified Detector\" is an end-to-end unified approach. Our LayoutFormer with", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7849654e5281b3634de9f98c1443ff423ebadf01101ef5a786642da728b0bfde.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVenueLine DetectionWord Detection
MSRA-TD500CTW1500Total-Text
RPFRPFRPF
TextSnake [23]ECCV1873.983.278.367.985.375.674.582.778.4
MSR [42]IJCAI1976.787.481.777.883.880.773.085.278.6
LOMO [48]CVPR19---76.585.780.879.387.683.3
PAN [38]ICCV1983.884.484.181.286.483.781.089.385.0
ContourNet [39]CVPR20---84.183.783.983.986.985.4
DB [19]AAAI2079.291.584.980.286.983.482.587.184.7
TextFuseNet [46]IJCAI20---85.085.885.483.287.585.3
FCENet [53]CVPR21---83.487.685.582.589.385.8
BPNet [49]ICCV2180.6885.4082.9781.4587.8184.5184.6590.2787.37
PCR [8]CVPR2183.590.887.082.387.284.782.088.585.2
I3CL [9]IJCV22---84.587.485.983.789.286.3
FSG [32]CVPR2284.891.688.182.488.185.285.790.788.1
Unified Detector [24]CVPR2387.4488.0487.7087.4484.5685.9791.0684.9687.90
DPText-DETR [47]AAAI23---86.291.788.886.491.889.0
LayoutFormer-88.3291.9590.1084.2688.1686.1785.0789.2687.12
", + "bbox": [ + 117, + 88, + 854, + 334 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/293cbef44fd980e45ee96ee3aad49c54b18a0f590732c0a02676ecc1387fc212.jpg", + "table_caption": [ + "Table 8. Experimental results of single-level text detection. We set $N$ to 100 on three datasets. On MSRA-TD500 and CTW1500, our LayoutFormer achieves the state-of-the-art performance. On Total-Text, our LayoutFormer achieves competitive results." + ], + "table_footnote": [], + "table_body": "
MethodNlineNparaLineParagraph
PQPQ
GCP API [24]--56.1746.33
GCN-PP [24]384-62.2350.10
Mask-RCNN-Cluster [24]384-62.2351.67
Max-DeepLab-Cluster [24]384-62.2352.52
128-58.7651.48
Unified Detector [24]256--52.50
384-62.2353.60
LayoutFormer38420062.3753.76
", + "bbox": [ + 80, + 398, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Experimental results of line-paragraph two-level text detection on the HierText testing set. $N_{line}$ denotes the number of text line queries.", + "bbox": [ + 75, + 561, + 470, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$N_{line} = 384$ outperforms Unified Detector [24] with $N_{line} = 384$ by $0.14\\%$ and $0.16\\%$ in terms of PQ of text detection and layout analysis, respectively.", + "bbox": [ + 76, + 636, + 468, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For word-line-paragraph detection, there is no any prior research on three-level text detection and layout analysis. We only list our performance on Tab. 4, which can be used by subsequent methods for comparison. Our Layout-Former achieves $49.24\\%$ , $46.21\\%$ , and $42.97\\%$ in terms of PQ of word-level, line-level, and paragraph-level, respectively. The inference speed of our method is 3.0 FPS.", + "bbox": [ + 75, + 684, + 468, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Single-level Text Detection. In this section, we evaluate the performance of our model on the most widely used benchmarks for single-level scene text detection, i.e., MSRAD500 and CTW1500 for line-level, Total-Text for word-level. Results are listed in Tab. 8. The quantitative results demonstrate the advances of our model.", + "bbox": [ + 75, + 792, + 468, + 881 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For line detection, we achieve the state-of-the-art results", + "bbox": [ + 96, + 885, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on MSRA-TD500. Notably, our LayoutFormer achieves $88.32\\%$ , $91.95\\%$ , and $90.10\\%$ in terms of Recall, Precision, and F-measure, respectively, which significantly outperforms other methods with a great margin. For example, LayoutFormer outperforms PCR [8], FSG [32], and Unified Detector [24] by $3.10\\%$ , $2.00\\%$ , $2.40\\%$ in terms of F-measure, respectively. On CTW1500, LayoutFormer achieves $84.26\\%$ , $88.16\\%$ , and $86.17\\%$ in terms of Recall, Precision, and F-measure, respectively, achieving the promising performance.", + "bbox": [ + 496, + 400, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For word detection, we achieve competitive results on Total-Text. The performance of our LayoutFormer is slightly lower than FSG [32] and Unified Detector [24], in which the former utilizes more training epochs and the latter utilizes larger training datasets and more training epochs. DPText-DETR [47] pre-trains their models on more datasets and these datasets are specifically for improving the performance of arbitrary-shape texts.", + "bbox": [ + 496, + 553, + 890, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 691, + 619, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a hierarchical text detector for promoting scene text understanding, which simultaneously detects multi-level texts and predicts their geometric layouts. In LayoutFormer, we propose WordDecoder, LineDecoder, and ParaDecoder to detect word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively. Through this research, we hope to provide a baseline for hierarchical text detection and inspire other hierarchical tasks. In addition, we will continue to promote and improve the hierarchical architecture for scene text understanding.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15672", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Furkan Biten, Rubén Tito, Andrés Mafla, Lluis Gómez i Bigorda, Marçal Rusinol, C. V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4290-4300, 2019. 1", + "[2] Samuele Capobianco, Leonardo Scommegna, and Simone Marinai. Historical handwritten document segmentation by using a weighted loss. In ANNPR, pages 395-406, 2018. 3", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 2", + "[4] Praneeth Chakravarthula, Jim Aldon D'Souza, Ethan Tseng, Joe Bartusek, and Felix Heide. Seeing with sound: Long-range acoustic beamforming for multimodal scene understanding. In CVPR, pages 982-991, 2023. 1", + "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2018. 3", + "[6] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1280–1289, 2022. 2, 4, 6", + "[7] Chee-Kheng Ch'ng, Chee Seng Chan, and Cheng-Lin Liu. Total-text: toward orientation robustness in scene text detection. IJDAR, 23(1):31-52, 2020. 6", + "[8] Pengwen Dai, Sanyi Zhang, Hua Zhang, and Xiaochun Cao. Progressive contour regression for arbitrary-shape scene text detection. In CVPR, pages 7393-7402, 2021. 8", + "[9] Bo Du, Jian Ye, Jing Zhang, Juhua Liu, and Dacheng Tao. I3CL: intra- and inter-instance collaborative learning for arbitrary-shaped scene text detection. IJCV, 130(8):1961-1977, 2022. 8", + "[10] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5354-5363, 2022. 2, 4", + "[11] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6", + "[12] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. 2022. 3", + "[13] Junjie He, Pengyu Li, Yifeng Geng, and Xuansong Xie. Fastinst: A simple query-based model for real-time instance segmentation. In CVPR, pages 23663-23672, 2023. 4", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, pages 2980-2988, 2017. 3", + "[16] Joonho Lee, Hideaki Hayashi, Wataru Ohyama, and Seiichi Uchida. Page segmentation using a convolutional neural net" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "work with trainable co-occurrence features. In ICDAR, pages 1023-1028, 2019. 1, 3", + "[17] Kai Li, Curtis Wigington, Chris Tensmeyer, Handong Zhao, Nikolaos Barmpalios, Vlad I. Morariu, Varun Manjunatha, Tong Sun, and Yun Fu. Cross-domain document object detection: Benchmark suite and method. In CVPR, pages 12912-12921, 2020. 3", + "[18] Minghui Liao, Baoguang Shi, Xiang Bai, Xinggang Wang, and Wenyu Liu. Textboxes: A fast text detector with a single deep neural network. In AAAI, pages 4161-4167, 2017. 2", + "[19] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In AAAI, pages 11474-11481, 2020. 2, 8", + "[20] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, pages 936-944, 2017. 4", + "[21] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. PR, 90:337-345, 2019. 6", + "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3", + "[23] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 19-35, 2018. 2, 8", + "[24] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1039-1049, 2022. 2, 3, 5, 6, 7, 8", + "[25] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. ICDAR 2023 competition on hierarchical text detection and recognition. In ICDAR, pages 483-497, 2023. 1", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[27] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascadetabnet: An approach for end to end table detection and structure recognition from image-based documents. In CVPR, pages 2439-2447, 2020. 3", + "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 91-99, 2015. 3, 4", + "[29] Baoguang Shi, Xiang Bai, and Serge J. Belongie. Detecting oriented text in natural images by linking segments. In CVPR, pages 3482-3490, 2017. 2", + "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, pages 8317-8326, 2019. 1", + "[31] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3591-3600, 2021. 2", + "[32] Jingqun Tang, Wenqing Zhang, Hongye Liu, Mingkun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be bet-" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ter than all: Feature sampling and grouping for scene text detection. In CVPR, pages 4553-4562, 2022. 1, 2, 8", + "[33] Zhi Tian, Weilin Huang, Tong He, Pan He, and Yu Qiao. Detecting text in natural image with connectionist text proposal network. In ECCV, pages 56-72, 2016. 2", + "[34] Matheus Palhares Viana and Dario Augusto Borges Oliveira. Fast cnn-based document layout analysis. In ICCV, pages 1173-1180, 2017. 3", + "[35] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan L. Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 2", + "[36] Renshen Wang, Yasuhisa Fujii, and Ashok C. Popat. Postocr paragraph recognition by graph convolutional networks. In WACV, pages 2533-2542, 2022. 3", + "[37] Wenhai Wang, Enze Xie, Xiang Li, Wenbo Hou, Tong Lu, Gang Yu, and Shuai Shao. Shape robust text detection with progressive scale expansion network. In CVPR, pages 9336-9345, 2019. 2", + "[38] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In ICCV, pages 8439-8448, 2019. 2, 8", + "[39] Yuxin Wang, Hongtao Xie, Zheng-Jun Zha, Mengting Xing, Zilong Fu, and Yongdong Zhang. Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In CVPR, pages 11750-11759, 2020. 2, 8", + "[40] Guanghui Xu, Shuaicheng Niu, Mingkui Tan, Yucheng Luo, Qing Du, and Qi Wu. Towards accurate text-based image captioning with content diversity exploration. In CVPR, pages 12637-12646, 2021. 1", + "[41] Yue Xu, Fei Yin, Zhaoxiang Zhang, and Cheng-Lin Liu. Multi-task layout analysis for historical handwritten documents using fully convolutional networks. In IJCAI, pages 1057-1063, 2018. 3", + "[42] Chuhui Xue, Shijian Lu, and Wei Zhang. MSR: multi-scale shape regression for scene text detection. In IJCAI, pages 989–995, 2019. 8", + "[43] Chuhui Xue, Jiaxing Huang, Wenqing Zhang, Shijian Lu, Changhu Wang, and Song Bai. Contextual text block detection towards scene text understanding. In ECCV, pages 374-391, 2022. 1, 3", + "[44] Xiao Yang, Ersin Yumer, Paul Arente, Mike Kraley, Daniel Kifer, and C. Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In CVPR, pages 4342-4351, 2017. 1", + "[45] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090, 2012. 5", + "[46] Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. Textfusenet: Scene text detection with richer fused features. In IJCAI, pages 516-522, 2020. 8", + "[47] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dpext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, 2023. 1, 2, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Chengquan Zhang, Borong Liang, Zuming Huang, Mengyi En, Junyu Han, Errui Ding, and Xinghao Ding. Look more than once: An accurate detector for text of arbitrary shapes. In CVPR, pages 10552-10561, 2019. 8", + "[49] Shi-Xue Zhang, Xiaobin Zhu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Adaptive boundary proposal network for arbitrary shape text detection. In ICCV, pages 1285-1294, 2021. 2, 8", + "[50] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. In CVPR, pages 2642-2651, 2017. 2", + "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 2", + "[52] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 2", + "[53] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 8" + ], + "bbox": [ + 501, + 92, + 890, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15674", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_model.json b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40039f8affdc42999d57c42b44ed3685998f1bbe --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.084, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.181, + 0.751, + 0.218 + ], + "angle": 0, + "content": "Min Liang, Jia-Wei Ma, Xiaobin Zhu*, Jingyan Qin, Xu-Cheng Yin \nUniversity of Science and Technology Beijing" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.22, + 0.816, + 0.252 + ], + "angle": 0, + "content": "{lm20200126, mjw20151001}@@hotmail.com, {zhuxiaobin, xuchengyin}@ustb.edu.cn, qinjingyanking@foxmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.668 + ], + "angle": 0, + "content": "Existing scene text detectors generally focus on accurately detecting single-level (i.e., word-level, line-level, or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts, detecting multi-level texts while exploring their contextual information is critical. To this end, we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection, which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer, WordDecoder, LineDecoder, and ParaDecoder are proposed to be responsible for word-level text prediction, line-level text prediction, and paragraph-level text prediction, respectively. Meanwhile, WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships, respectively. In addition, we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.696, + 0.21, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Reading and understanding texts in scene images and digital documents is important in various real-world applications, including visual recognition [40], scene understanding [4], and text-based VQA [1, 30]. Compared to words, phrases and sentences express more semantic messages, and paragraphs convey richer contextual semantic messages. For better reading and understanding texts in scenes, it is necessary to simultaneously detect multi-level texts (i.e., word-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author. This work was supported by National Key Research and Development Program of China (2020AAA0109701), National Science Fund for Distinguished Young Scholars (62125601), and National Natural Science Foundation of China (62076024, 62172035)." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.286, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.535, + 0.894, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Top: LayoutFormer is introduced for hierarchical text detection, which simultaneously detects multi-level texts (i.e., word-level, line-level, and paragraph-level) and predicts their geometric layouts. Bottom: Pipeline comparisons of (a) single-level detector, (b) two-level detector, and (c) hierarchical text detector." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.894, + 0.716 + ], + "angle": 0, + "content": "level for words, line-level for phrases and sentences, and paragraph-level for paragraphs) and predict their geometric layouts [16, 44], i.e., hierarchical text detection [25]. Hierarchical texts can deliver complete and meaningful text messages, which is beneficial for various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Although existing scene text detectors have achieved promising performance, they [32, 47] often concentrate on detecting single-level text, which decodes the encoded features of an input image into their representation of independent text entities, e.g., words, lines, or paragraphs. As shown in Fig. 1 (a), single-level detectors only predict one level of text results and do not explore the contextual relationship among different levels of texts. This makes them unable to deliver more informative messages for facilitating scene text understanding. Recently, some methods have investigated two-level text detection. A schematic of the framework is shown in Fig. 1 (b). CUTE [43] proposes a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.332 + ], + "angle": 0, + "content": "two-stage network, which first adopts a Transformer structure like DETR [3] to perform word/line bounding boxes and then directly learns the sequential relationships between detected text boxes based on visual features after ROI operation and positional information. Unified Detector [24] proposes a unified scene text detection and layout analysis network, which first uses a Transformer-based structure [35] to generate word/line mask maps and then models their affiliate relationships. These two methods are all supervised by an affinity matrix or index label, directly modelling relationships on the word-level/line-level detection results. They only rely on the former detection results to model contextual relationships, making the model's performance highly limited by the text detection results at the first stage. Moreover, this paradigm cannot be easily extended to the task of text detection at more than two levels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Existing scene text detectors generally adopt multi-scale feature maps that contain rich information for accurate prediction. However, methods based on the Transformer encoder-decoder structure often suffer from huge computational and memory costs on high-resolution feature maps. This is because the attention weight computation in the Transformer decoder is of linear computation w.r.t. pixel numbers. There have been several work to tackle the issues. Deformable DETR [51] introduces the multi-scale deformable attention module to select a small set of key locations. AdaMixer [10] adaptively samples features over the space and scales of an object. However, these methods are based on estimated offsets of instance coordinates, not applicable to segmentation methods. In image segmentation, Mask2Former [6] claims that local features can well update query features and proposes to use image features of foreground regions. TSP [31] adopts a RoIAlign operation to extract features of RoIs. However, these methods still suffer from complex operations. We think that it is enough to update query features with only foreground features of a few representative pixels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a hierarchical text detector named LayoutFormer, which can simultaneously produce hierarchical text detection boxes and their layout relationships. Unlike using inter-level relationships as supervision, we directly adopt multi-level text boxes as supervision, which can significantly reduce the dependence on the former text detection results and implicitly exploit multi-level textual information. In LayoutFormer, we propose three Transformer-based modules, i.e., WordDecoder, LineDecoder, and ParaDecoder, for detecting word-level texts, line-level texts, and paragraph-level texts, respectively. We first adopt text line as the detection unit, and LineDecoder updates and forms line query features. Because a line can be split into several words, WordDecoder learns the deviation of each word within a line based on the line query features and forms final word query features. ParaDecoder aggre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "gates line query features by learning line-paragraph relationships and forms final paragraph query features. In addition, we propose a Prior Location Sampler, which adaptively selects a few representative features that are highly relevant to scene texts. The sampled features are used for cross-attention operation to update query features. Prior Location Sampler can improve the model performance while reducing the training cost. Extensive experiments verify the superior performance on hierarchical text detection." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.234, + 0.822, + 0.248 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.314 + ], + "angle": 0, + "content": "- We propose an innovative hierarchical text detector that simultaneously detects multi-level text instances and explores their contextual information for predicting the geometric layouts, finally performing hierarchical outputs." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.315, + 0.892, + 0.389 + ], + "angle": 0, + "content": "- We propose WordDecoder, LineDecoder, and ParaDecoder in LayoutFormer to decode word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.39, + 0.892, + 0.449 + ], + "angle": 0, + "content": "- We propose a Prior Location Sampler to adaptively select a few representative foreground features for updating queries, which can significantly reduce the computational cost while improving the detection performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.45, + 0.892, + 0.495 + ], + "angle": 0, + "content": "- Extensive experiments on multiple publicly available datasets verify the state-of-the-art performance of our LayoutFormer." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.642, + 0.54 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.696, + 0.57 + ], + "angle": 0, + "content": "2.1. Scene Text Detection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Deep learning-based scene text detection methods have been widely investigated and roughly divided into bottom-up and top-down methods. The bottom-up methods generally detect local components [23, 29, 33] or text pixels [19, 37] for grouping text instances. TextSnake [23] describes a text instance as a sequence of ordered, overlapping disks. The pixel-based methods use some auxiliary information (e.g., similarity vectors in PAN [38], and threshold map in DB [19]) or post-processings (e.g., a progressive scale expansion algorithm in PSENet [37]) to generate text instances. The top-down methods [18, 50] directly predict bounding boxes of scene text instances. To detect arbitrary-shaped texts, some methods [39, 49, 52] localize the key points on the contours of scene texts. Recently, DETR [3] presents a Transformer-based architecture for object detection and achieved great success. Inspired by it, many researchers [24, 32, 47] apply the Transformer structure to scene text detection. Overall, existing scene text detection methods mainly address single-level text detection, ignoring the text detection of multiple levels and further predicting the hierarchical relationships." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15666" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.868, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.357, + 0.895, + 0.429 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed LayoutFormer, which mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders (WordDecoder, LineDecoder, and ParaDecoder) and three predictors. \\(M_l\\) denotes predicted mask maps of text lines by \\(l\\)-th Transformer decoder, and \\(F_l\\) denotes the corresponding image features generated by \\(M_l\\). We first predict text lines, and then predict words and paragraphs. In testing, word-level results produce line-level results based on word-line relationships, and then paragraph-level results are produced based on line-paragraph relationships, finally obtaining hierarchical text detection results." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.449, + 0.445, + 0.516 + ], + "angle": 0, + "content": "
Detection UnitQueryMemory/GBPQ
word25612.153.09
line25610.657.48
paragraph2568.540.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.526, + 0.472, + 0.582 + ], + "angle": 0, + "content": "Table 1. Comparisons of word, line, and paragraph as text detection unit on the HierText validation set, respectively. \"Query\" denotes the number of text queries. \"Memory\" denotes GPU Memory consumed in training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.607, + 0.24, + 0.623 + ], + "angle": 0, + "content": "2.2. Layout Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Document Layout analysis has made great progress, which can be mainly divided into object detection-based methods [17, 27, 34] and segmentation-based methods [2, 16, 41]. Inspired by object detection [15, 28] and semantic segmentation [5, 12, 22], these methods treat semantically coherent text blocks as a special kind of object. However, they fail to produce word or line-level detections and can only be used in companies with standalone text detectors, increasing the complexity of the pipeline. Some work [36] takes a hierarchical view and applies graph-based models on the finest granularity, i.e., individual words, to analyze the layout." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Recently, some work has investigated layout analysis in scene text detection. CUTE [43] detects contextual text blocks consisting of one or multiple ordered integral text units. Their model is a two-stage structure, which first detects integral text units and then models the relationship of integral text units by a designed relation module. Unified Detector [24] introduces a novel task of unified scene text" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.558 + ], + "angle": 0, + "content": "detection and layout analysis, which models the two closely related tasks with a unified model. It produces mask outputs for the text detection task and an affinity matrix for the layout analysis task. The above methods are highly dependent on text detection results at the first stage. Differently, we use multi-level text boxes as supervision to learn the relationship among different levels for layout analysis." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.637, + 0.595 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.63, + 0.621 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "How to effectively model hierarchical text detection tasks? Is the top-down or bottom-up approach better? As we know, the text layouts in real scene images is highly complex, bringing significant challenges to hierarchical text detection. According to experiments and analyses, we found that paragraph-level text boxes are considerably irregular and always contain much background, which makes existing scene text detection approaches difficult to model text instances. For example, as shown in the second row of Fig. 3, a paragraph text box annotates the entire image. As listed in Tab. 1, the detection performance of paragraphs is much lower than that of words and lines. Thus, the top-down modeling approach is considered inappropriate. Both word and line are generally used as text detection units. However, text lines always contain more complete semantic information than words. The aspect ratios and sizes of words are considerably more variable than text lines, making detection more challenging. Meanwhile, if the word is used" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15667" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.454, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.26, + 0.47, + 0.289 + ], + "angle": 0, + "content": "Figure 3. Some examples of word-level (first column), line-level (second column), and paragraph-level (third column) text boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.388 + ], + "angle": 0, + "content": "as the detection unit, the detection error will accumulate bigger after two aggregation processes. Hence, we finally choose line-level as our detection unit, generate word-level texts by splitting them downward, and get paragraph-level texts by aggregating them upward." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.398, + 0.283, + 0.413 + ], + "angle": 0, + "content": "3.2. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.473, + 0.755 + ], + "angle": 0, + "content": "The architecture of our LayoutFormer is illustrated in Fig. 2. It mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders and three predictors, respectively, corresponding to word-level text detection, line-level text detection and paragraph-level text detection. Specifically, given an input image \\( I \\in R^{H \\times W \\times 3} \\), the feature encoder with a ResNet-50 [14] backbone and a Transformer encoder [6] extracts and enhances the features and then generates multi-scale feature maps, i.e., \\( C_4 \\in R^{\\frac{H}{32} \\times \\frac{W}{32} \\times C} \\), \\( C_3 \\in R^{\\frac{H}{16} \\times \\frac{W}{16} \\times C} \\), \\( C_2 \\in R^{\\frac{H}{8} \\times \\frac{W}{8} \\times C} \\) and \\( C_1 \\in R^{\\frac{H}{4} \\times \\frac{W}{4} \\times C} \\). The typical value we use is \\( C = 256 \\). Then, Prior Location Sampler adaptively selects image features for the Transformer decoder. Afterwards, the LineDecoder, WordDecoder, and ParaDecoder update line query features \\( X_{line} \\in R^{N_{line} \\times C} \\), word query features \\( X_{word} \\in R^{N_{word} \\times C} \\), and paragraph query features \\( X_{para} \\in R^{N_{para} \\times C} \\), respectively. The three predictors convert \\( X_{line} \\), \\( X_{word} \\), and \\( X_{para} \\) to line-level prediction \\( (P_{line}, M_{line}) \\), word-level prediction \\( (P_{word}, M_{word}) \\), and paragraph-level prediction \\( (P_{para}, M_{para}) \\), where \\( P_* \\in R^{N_* \\times 2} \\) is text/non-text probability values of \\( N_* \\) instances and \\( M_* \\in R^{N_* \\times \\frac{H}{4} \\times \\frac{W}{4}} \\) is mask maps of \\( N_* \\) instances." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.764, + 0.296, + 0.78 + ], + "angle": 0, + "content": "3.3. Prior Location Sampler" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.788, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Mask2Former [6] proposes masked attention to attend within the foreground region of the predicted mask for each query, which can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.845, + 0.469, + 0.862 + ], + "angle": 0, + "content": "\\[\nX _ {l} = \\operatorname {s o f t m a x} \\left(B _ {l - 1} + Q _ {l} K _ {l} ^ {T}\\right) V _ {l} + X _ {l - 1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(l\\) is the layer index, \\(B_{l-1}\\) is the binary mask prediction of \\(l-1\\)-th layer, \\(X_{l}\\) is query features at the \\(l\\)-th layer, \\(Q_{l}\\)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.885, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.25, + 0.893, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Illustration of the Prior Location Sampler, which selects representative foreground features with Top- \\( K \\) scores in predicted mask maps and then aggregates with weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.893, + 0.492 + ], + "angle": 0, + "content": "is query features \\( X_{l-1} \\) under linear transformation, \\( K_l \\) and \\( V_l \\) are the input image features under linear transformation. \\( X_0 \\) denotes input query features to the Transformer decoder. \\( B_0 \\) is the binary mask prediction obtained from \\( X_0 \\). The input image features are the entire multi-scale feature maps from a feature encoder. Their computational complexity of Eq. (1) is \\( O(NH'W'C) \\), where \\( H' \\) and \\( W' \\) are the height and width of a specific feature map respectively, and \\( N \\) is the number of queries. The cross-attention operation suffers from a linear complexity growth with the spatial size of feature maps. Moreover, the main accompanying problem is the high demand for computing resources." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Thus, we propose the Prior Location Sampler to select representative foreground features with Top- \\(K\\) scores in mask maps and obtain the image features \\(F_{l}\\in R^{K\\times N\\times C}\\) for \\(l\\)-th decoder layer. Our computational complexity of Eq. (1) is \\(O(NKC)\\). In our implementation, the value \\(K\\ll H^{\\prime}W^{\\prime}\\), and thus the complexity of our cross-attention operation could be significantly reduced. The computational resource consumption due to an increase in the number of feature maps or an increase in the scale of feature maps is negligible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The architecture of the Prior Location Sampler is illustrated in Fig. 4. Specifically, \\( M_{l} \\) denotes the predicted mask maps of \\( l \\)-th decoder layer, and" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.696, + 0.891, + 0.716 + ], + "angle": 0, + "content": "\\[\nM _ {l} = \\left\\{M _ {l n} \\in R ^ {\\frac {H}{4} \\times \\frac {W}{4}} | n = 1, 2, \\dots , N \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.894, + 0.856 + ], + "angle": 0, + "content": "For each \\(M_{ln}\\), we use 0.5 for thresholding. We sort scores in \\(M_{ln}\\) and select features with Top- \\(K\\) scores [13] in multi-scale feature maps \\(\\{C_i|i = 1,2,3,4\\}\\) respectively, where \\(f_{K}(\\cdot)\\) is sampling operation. Then, we assign [20, 28] text ROIs of different scales to the pyramid levels for calculating weights [10] as Eq. (4), where \\(S_{ln}\\) is the cumulative sum of the binarized \\(M_{ln}\\), i.e., the area of a candidate text, and 56 [20] is a scaling factor. The selected features for \\(M_{ln}\\) are gathered into \\(F_{ln}\\in R^{K\\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.864, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\nF _ {l n} = \\sum_ {i = 1} ^ {4} w _ {l i} f _ {K} \\left(C _ {i}, M _ {l n}\\right), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15668" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.195, + 0.114, + 0.469, + 0.151 + ], + "angle": 0, + "content": "\\[\nt _ {l n} = \\left\\lfloor \\log_ {2} \\frac {S _ {l n}}{5 6} \\right\\rfloor , \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.15, + 0.385, + 0.175 + ], + "angle": 0, + "content": "\\[\nw _ {l} = \\text {s o f t m a x} \\left[ - \\frac {\\left(t _ {l n} - \\iota + 1\\right) ^ {2}}{2} \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.183, + 0.424, + 0.199 + ], + "angle": 0, + "content": "The final features \\(F_{l}\\in R^{K\\times N\\times C}\\) are denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.211, + 0.468, + 0.226 + ], + "angle": 0, + "content": "\\[\nF _ {l} = \\left\\{F _ {l n} \\mid n = 1, 2, \\dots , N \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.234, + 0.28, + 0.248 + ], + "angle": 0, + "content": "3.4. Transformer Decoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.256, + 0.468, + 0.422 + ], + "angle": 0, + "content": "LineDecoder. We feed the selected image features \\( F_{l-1} \\in R^{K \\times N_{line} \\times C} \\) and \\( N_{line} \\) learnable line queries into LineDecoder, where \\( K = 32 \\). LineDecoder uses the standard Transformer decoder structure, consisting of self-attention, cross-attention, and feedforward network. We also add the positional embeddings to queries and keys at every self-attention and cross-attention layer. We refine line query features layer-by-layer by setting \\( L_1 = 4 \\). The updated line query features are adopted to predict \\( l \\)-th mask maps, which are fed into Prior Location Sampler to generate image features for the next decoder layer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.423, + 0.468, + 0.497 + ], + "angle": 0, + "content": "WordDecoder. We frist hypothesize that there are \\( N_{per\\_line} \\) words in each line and initialize \\( N_{per\\_line} \\) learnable word queries. Thus, the number of words in an image can be formulated as the product of the number of lines in an image and words per line, as" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.51, + 0.468, + 0.526 + ], + "angle": 0, + "content": "\\[\nN _ {\\text {w o r d}} = N _ {\\text {l i n e}} \\times N _ {\\text {p e r l i n e}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.468, + 0.623 + ], + "angle": 0, + "content": "So, we can decouple the self-attention between words into self-attention between lines and self-attention between words per line. Based on line queries from LineDecoder, we only need to compute self-attention between \\( N_{per\\_line} \\), greatly reducing the computational cost. Then we obtain word query features \\( X_{word} \\in R^{N_{word} \\times C} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.633, + 0.468, + 0.652 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {w o r d}} = X _ {\\text {l i n e}} X _ {\\text {p e r - l i n e}} ^ {T}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.468, + 0.81 + ], + "angle": 0, + "content": "where \\( X_{line} \\in R^{N_{line} \\times 1 \\times C} \\) is line query features and \\( X_{per\\_line} \\in R^{N_{per\\_line} \\times 1 \\times C} \\) is learnable word query features per line. Then cross-attention followed by feedforward network adopts image features \\( F_{l} \\in R^{K \\times N_{line} \\times C} \\) with \\( K = 64 \\) to update \\( X_{word} \\). The image features are obtained by Prior Location Sampler, which feeds the predicted mask maps of the last decoder layer in LineDecoder. In our experiments, we set \\( L_{3} = 1 \\). With our word decoupling mechanism, every \\( N_{per\\_line} \\) queries predicts words within a text line for capturing word-line relationships." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.468, + 0.9 + ], + "angle": 0, + "content": "ParaDecoder. We adaptively learn line-paragraph relationships and aggregate line query features for paragraph-level texts to get paragraph query features. We initialize \\( N_{para} \\) learnable paragraph queries. ParaDecoder also consists of self-attention, cross-attention, and feedforward networks. It uses \\( X_{line} \\) line query features as inputs of cross attention" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "to update paragraph queries. Positional embeddings are added to queries and keys at every self-attention and cross-attention layer. In our experiments, we adopt \\( X_{line} \\) of last decoder layer in LineDecoder and set \\( L_{2} = 1 \\). In this process, the learned attention weights \\( (R^{N_{para} \\times N_{line}}) \\) in cross-attention of ParaDecoder are line-paragraph relationships." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.191, + 0.638, + 0.207 + ], + "angle": 0, + "content": "3.5. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.215, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Training. Instead of relation matrices, we use multi-level text boxes as supervision. The total loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.257, + 0.891, + 0.273 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {l i n e}} + L _ {\\text {w o r d}} + L _ {\\text {p a r a}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.891, + 0.328 + ], + "angle": 0, + "content": "where \\( L_{line} \\), \\( L_{word} \\), and \\( L_{para} \\) respectively denote losses of line-level, word-level, and paragraph-level, which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.34, + 0.891, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {*} = \\lambda_ {c l s} L _ {c l s} \\left(P _ {*}, P _ {*} ^ {\\prime}\\right) + \\lambda_ {m a s k} L _ {m a s k} \\left(M _ {*}, M _ {*} ^ {\\prime}\\right), \\\\ * \\in \\{w o r d, l i n e, p a r a \\}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.387, + 0.891, + 0.492 + ], + "angle": 0, + "content": "where \\( P_{*} \\) and \\( P_{*}^{\\prime} \\) are the class predictions and their corresponding ground truth, \\( M_{*} \\) and \\( M_{*}^{\\prime} \\) are the mask predictions and their corresponding ground truth, \\( L_{cls} \\) denotes binary cross-entropy loss, \\( L_{mask} \\) is the sum of binary cross-entropy loss and dice loss. We set \\( \\lambda_{cls} = 2.0 \\) for predictions matched with ground truth and 0.1 for unmatched predictions. \\( \\lambda_{mask} \\) is set to 5.0." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.493, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Inference. We present a simple inference procedure that converts class predictions and mask predictions to text boxes, which mainly consists of four steps: (1) the text queries whose class scores are over 0.3 will be selected; (2) the mask predictions of selected queries are binarized by 0.5 to get the binary maps; (3) the connected regions are obtained from the binary maps; (4) the text queries whose averaged foreground mask probability is below 0.9 will be filtered. For hierarchical text detection task, we frist generate the word-level detection boxes by the above inference procedure. Then, we use word-line relationships from WordDecoder to form a word-line geometric layouts and further use line-paragraph relationships from ParaDecoder to form a line-paragraph geometric layouts, finally performing hierarchical text detection outputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.732, + 0.632, + 0.749 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.602, + 0.771 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.891, + 0.856 + ], + "angle": 0, + "content": "HierText [24]. It is a hierarchical text detection dataset consisting of 8,281 training images, 1,724 validation images, and 1,634 testing images. It annotates images hierarchically, which first annotates word locations with polygons, then clusters words into lines and lines into paragraphs." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.891, + 0.901 + ], + "angle": 0, + "content": "MSRA-TD500 [45]. It is a line-level annotated arbitrary-oriented long text dataset. It consists of 300 training images and 200 testing images collected from natural scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15669" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.868, + 0.18 + ], + "angle": 0, + "content": "
MethodTotal-TextHierText-Line
RPFMemory/GBRPFTPQMemory/GB
C2-C4masked attention84.6887.7986.2115.365.5682.7373.1577.5756.7418.1
C2-C4Top-K&concat84.6687.8086.2015.165.3482.5672.9577.3756.4418.3
C1-C4Top-K&concat84.4688.1586.2716.265.8882.9673.4477.5356.9420.3
C1-C4Top-K&weight85.0789.2687.1214.266.6584.1074.3777.2957.4817.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.189, + 0.892, + 0.232 + ], + "angle": 0, + "content": "Table 2. Ablation comparison of Prior Location Sampler on the Total-Text and line-level of the HierText validation set. \"Memory\" denotes GPU Memory consumed in training. \\(C_1 - C_4\\) denotes the multi-scale feature maps used in the Transformer decoder. The results show that \"Top-K&weight\" with \\(C_1 - C_4\\) achieves a considerable improvement and consumes fewer training resources than other methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.468, + 0.315 + ], + "angle": 0, + "content": "CTW1500 [21]. It is an arbitrary-shaped scene text dataset that consists of 1,000 training images and 500 testing images. In this dataset, the annotations of text instances are line-level and labelled by a polygon with 14 key points." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.315, + 0.468, + 0.36 + ], + "angle": 0, + "content": "Total-Text [7]. It is an arbitrary-shaped scene text dataset that contains 1,255 training and 300 testing images, which are annotated by the word-level polygon with 14 key points." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.371, + 0.295, + 0.387 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.468, + 0.544 + ], + "angle": 0, + "content": "In our experiments, for HierText, we only use the training images of HierText to train the models by \\(60k\\) steps for single-level text detection and \\(80k\\) steps for multi-level text detection. In testing, we keep the aspect ratio of testing images and resize them to 1,120 height. For MSRA-TD500, CTW1500, and Total-Text, we adopt SynthText [11] to pretrain the models by \\(150k\\) steps. Then, we finetune the models on the corresponding real-world datasets by \\(40k\\) steps. In testing, we keep the aspect ratio of testing images and resize them to 800 height on three datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.546, + 0.468, + 0.696 + ], + "angle": 0, + "content": "In training, we use AdamW [26] optimizer and the step learning rate schedule with an initial learning rate of 0.0001 and a weight decay of 0.05. A learning rate multiplier of 0.1 is applied to the ResNet-50 backbone. We decay the learning rate at 0.9 and 0.95 fractions of the total number of training steps by a factor of 10. We train our models with a batch size of 8. Data augmentation includes: (1) Randomly Flipping; (2) Randomly rotate them in range \\((-10^{\\circ}\\) to \\(10^{\\circ})\\); (3) Randomly resize them in range (0.5 to 3.0); (4) Randomly cropping. Finally, we resize the images to \\(640\\times 640\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.262, + 0.722 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.468, + 0.806 + ], + "angle": 0, + "content": "For MSRA-TD500, CTW1500, and CTW1500 datasets, we follow the standard evaluation protocol Recall (R), Precision (P), and F-measure (F). For the HierText dataset, we follow [24] and adopt Panoptic Quality (PQ) as the main evaluation metric for the hierarchical text detection task." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.241, + 0.832 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.9 + ], + "angle": 0, + "content": "\\(K\\) value in Prior Location Sampler. We compare the different Top- \\(K\\) values in Prior Location Sampler. Results are listed in Tab. 3. Selecting too few foreground points may fail to capture and learn the features of an entire text. Select" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.252, + 0.888, + 0.33 + ], + "angle": 0, + "content": "
KTotal-TextHierText-Line
RPFRPFPQ
1684.3488.0186.1465.3481.1272.3855.54
3285.0789.2687.1266.6584.1074.3757.48
6484.7588.6386.6468.6784.4475.7558.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.339, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Table 3. Ablation comparison of Top- \\( K \\) value on Total-Text and line-level of the HierText validation set." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.38, + 0.886, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.475, + 0.89, + 0.503 + ], + "angle": 0, + "content": "Figure 5. Visualization of Top- \\( K \\) sampling points in Prior Location Sampler." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.629 + ], + "angle": 0, + "content": "ing too many foreground points may introduce noise, which is detrimental to the performance of the model. For Total-Text, we set \\( K \\) to 32, and the same for MSRA-TD500 and CTW1500. While text lines of HierText may have larger aspect ratios and scales, the performance of \\( K = 64 \\) is better than that of \\( K = 32 \\). However, we finally set \\( K \\) to 32 for a trade-off between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Implementation of Prior Location Sampler. We evaluate the effectiveness of the Prior Location Sampler on the Total-Text and HierText validation set. Results are listed in Tab. 2. \" \\(C_1 - C_4\\) \" denotes that multi-scale feature maps \\(\\{C_i | i = 1, 2, 3, 4\\}\\) are used in Transformer decoder. \"masked attention\" is followed from Mask2Former [6], which calculates the attention matrix by additionally adding a binary mask prediction. \"Top-K&concat\" applies Top-K strategy to multi-scale feature maps and then performs concatenation. \"Top-K&weight\" applies Top-K strategy to multiscale feature maps and then element-wise sums in weights. Inspired by Mask2Former, the first row of Tab. 2 feeds the entire feature map into the Transformer decoder. The experimental results show that \"Top-K&weight\" apparently achieves a significant improvement in harmonic accuracy and memory. Meanwhile, due to using a small number of foreground points, our method consumes fewer training resources (GPU Memory) than other methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15670" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.128, + 0.09, + 0.845, + 0.144 + ], + "angle": 0, + "content": "
MethodWordLineParagraph
RPFTPQRPFTPQRPFTPQ
LayoutFormer57.0475.2564.8975.8849.2450.3976.3260.7076.1246.2145.8973.1456.4076.1942.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.17, + 0.153, + 0.798, + 0.168 + ], + "angle": 0, + "content": "Table 4. Experimental results of word-line-paragraph three-level text detection on the HierText testing set." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.192, + 0.46, + 0.339 + ], + "angle": 0, + "content": "
MethodNperLineNlineWordLine
PQPQ
In the below experiments, we always set L3 to 1.
Xline1012044.0236.35
Fl1012050.2948.60
In the below experiments, we use Fl as image features.
L3 = 11012050.2948.60
L3 = 41012051.0744.05
L3 = 1612046.2139.60
L3 = 1620050.1246.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.35, + 0.471, + 0.392 + ], + "angle": 0, + "content": "Table 5. Ablation comparison of WordDecoder on the HierText validation set. \\( N_{per\\_line} \\) and \\( N_{line} \\) denote the number of word queries in each line and text line queries, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.406, + 0.456, + 0.485 + ], + "angle": 0, + "content": "
MethodNlineNparaLineParagraph
PQPQ
In the below experiments, we always set L2 to 1.
Xline12812858.3151.98
Fl12812857.2746.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.539 + ], + "angle": 0, + "content": "Table 6. Ablation comparison of ParaDecoder on the HierText validation set. \\( N_{line} \\) and \\( N_{para} \\) denote the number of text line queries and paragraph queries, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.762 + ], + "angle": 0, + "content": "Implementation of WordDecoder. We conduct ablation experiments by comparing different settings of image features, \\( L_{3} \\), \\( N_{per\\_line} \\) and \\( N_{line} \\) in WordDecoder. Results are listed in Tab. 5. Using \\( F_{l} \\) as image features for updating word queries is better than \\( X_{line} \\). This is because to identify different text lines, \\( X_{line} \\) extracts consistent features of each text line, whereas \\( X_{per\\_line} \\) needs to learn differentiated features within a text line to identify different words within a line. In addition, we find that \\( L_{3} \\) being set to 1 performs better than being set to 4. Thus, we adopt \\( L_{3} = 1 \\) by default. We also conduct comparative experiments with different \\( N_{per\\_line} \\) and \\( N_{line} \\). We can conclude that the larger the query number, the higher the performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.762, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Implementation of ParaDecoder. We compare different image features for ParaDecoder. Results are listed in Tab. 6. Adopting \\( X_{line} \\) as image features is better than \\( F_{l} \\). We think that after the learning of LineDecoder, \\( X_{line} \\) is able to characterize text lines well." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.449, + 0.863 + ], + "angle": 0, + "content": "4.5. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Hierarchical Text Detection. In this section, we evaluate the performance on the HierText testing set for hierarchical" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.192, + 0.878, + 0.246 + ], + "angle": 0, + "content": "
MethodNwordNlineWordLine
PQPQ
LayoutFormer120012050.3548.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.3 + ], + "angle": 0, + "content": "Table 7. Experimental results of word-line two-level text detection on the HierText testing set. \\( N_{word} \\) and \\( N_{line} \\) denote the number of word queries and text line queries, respectively." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.313, + 0.888, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.547, + 0.892, + 0.603 + ], + "angle": 0, + "content": "Figure 6. The hierarchical text detection examples on the HierText. The visualization images in each row are in turn the input image, word-level detection results, line-level detection results, and paragraph-level detection results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.734 + ], + "angle": 0, + "content": "text detection. To our knowledge, few work has conducted research on hierarchical text detection. Hence, we mainly compare with [24]. Results are listed in Tab. 7, Tab. 9, and Tab. 4. Meanwhile, we demonstrate hierarchical detection results on images in Fig. 6. Both qualitative and quantitative verify the advances of our method. And we can observe that our method can precisely detect hierarchical texts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.81 + ], + "angle": 0, + "content": "For word-line detection, our results are listed in Tab. 7. Because there is no any prior research on this, we provide a baseline to facilitate comparisons with other methods. Benefiting from our decoupling operation on words, our \\( N_{word} \\) can reach 1,200, much higher than Unified Detector [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For line-paragraph detection, our LayoutFormer achieves the best text detection and layout analysis performance. Results are listed in Tab. 9. \"GCP API\", \"GCN-PP\", \"Mask-RCNN-Cluster\", \"Max-DeepLab-Cluster\" are two-stage approaches, while \"Unified Detector\" is an end-to-end unified approach. Our LayoutFormer with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15671" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.855, + 0.335 + ], + "angle": 0, + "content": "
MethodVenueLine DetectionWord Detection
MSRA-TD500CTW1500Total-Text
RPFRPFRPF
TextSnake [23]ECCV1873.983.278.367.985.375.674.582.778.4
MSR [42]IJCAI1976.787.481.777.883.880.773.085.278.6
LOMO [48]CVPR19---76.585.780.879.387.683.3
PAN [38]ICCV1983.884.484.181.286.483.781.089.385.0
ContourNet [39]CVPR20---84.183.783.983.986.985.4
DB [19]AAAI2079.291.584.980.286.983.482.587.184.7
TextFuseNet [46]IJCAI20---85.085.885.483.287.585.3
FCENet [53]CVPR21---83.487.685.582.589.385.8
BPNet [49]ICCV2180.6885.4082.9781.4587.8184.5184.6590.2787.37
PCR [8]CVPR2183.590.887.082.387.284.782.088.585.2
I3CL [9]IJCV22---84.587.485.983.789.286.3
FSG [32]CVPR2284.891.688.182.488.185.285.790.788.1
Unified Detector [24]CVPR2387.4488.0487.7087.4484.5685.9791.0684.9687.90
DPText-DETR [47]AAAI23---86.291.788.886.491.889.0
LayoutFormer-88.3291.9590.1084.2688.1686.1785.0789.2687.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.346, + 0.894, + 0.377 + ], + "angle": 0, + "content": "Table 8. Experimental results of single-level text detection. We set \\(N\\) to 100 on three datasets. On MSRA-TD500 and CTW1500, our LayoutFormer achieves the state-of-the-art performance. On Total-Text, our LayoutFormer achieves competitive results." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.399, + 0.469, + 0.552 + ], + "angle": 0, + "content": "
MethodNlineNparaLineParagraph
PQPQ
GCP API [24]--56.1746.33
GCN-PP [24]384-62.2350.10
Mask-RCNN-Cluster [24]384-62.2351.67
Max-DeepLab-Cluster [24]384-62.2352.52
128-58.7651.48
Unified Detector [24]256--52.50
384-62.2353.60
LayoutFormer38420062.3753.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.563, + 0.472, + 0.606 + ], + "angle": 0, + "content": "Table 9. Experimental results of line-paragraph two-level text detection on the HierText testing set. \\( N_{line} \\) denotes the number of text line queries." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.637, + 0.47, + 0.682 + ], + "angle": 0, + "content": "\\(N_{line} = 384\\) outperforms Unified Detector [24] with \\(N_{line} = 384\\) by \\(0.14\\%\\) and \\(0.16\\%\\) in terms of PQ of text detection and layout analysis, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.79 + ], + "angle": 0, + "content": "For word-line-paragraph detection, there is no any prior research on three-level text detection and layout analysis. We only list our performance on Tab. 4, which can be used by subsequent methods for comparison. Our Layout-Former achieves \\(49.24\\%\\), \\(46.21\\%\\), and \\(42.97\\%\\) in terms of PQ of word-level, line-level, and paragraph-level, respectively. The inference speed of our method is 3.0 FPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.469, + 0.882 + ], + "angle": 0, + "content": "Single-level Text Detection. In this section, we evaluate the performance of our model on the most widely used benchmarks for single-level scene text detection, i.e., MSRAD500 and CTW1500 for line-level, Total-Text for word-level. Results are listed in Tab. 8. The quantitative results demonstrate the advances of our model." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "For line detection, we achieve the state-of-the-art results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.553 + ], + "angle": 0, + "content": "on MSRA-TD500. Notably, our LayoutFormer achieves \\(88.32\\%\\), \\(91.95\\%\\), and \\(90.10\\%\\) in terms of Recall, Precision, and F-measure, respectively, which significantly outperforms other methods with a great margin. For example, LayoutFormer outperforms PCR [8], FSG [32], and Unified Detector [24] by \\(3.10\\%\\), \\(2.00\\%\\), \\(2.40\\%\\) in terms of F-measure, respectively. On CTW1500, LayoutFormer achieves \\(84.26\\%\\), \\(88.16\\%\\), and \\(86.17\\%\\) in terms of Recall, Precision, and F-measure, respectively, achieving the promising performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.676 + ], + "angle": 0, + "content": "For word detection, we achieve competitive results on Total-Text. The performance of our LayoutFormer is slightly lower than FSG [32] and Unified Detector [24], in which the former utilizes more training epochs and the latter utilizes larger training datasets and more training epochs. DPText-DETR [47] pre-trains their models on more datasets and these datasets are specifically for improving the performance of arbitrary-shape texts." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.693, + 0.62, + 0.709 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a hierarchical text detector for promoting scene text understanding, which simultaneously detects multi-level texts and predicts their geometric layouts. In LayoutFormer, we propose WordDecoder, LineDecoder, and ParaDecoder to detect word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively. Through this research, we hope to provide a baseline for hierarchical text detection and inspire other hierarchical tasks. In addition, we will continue to promote and improve the hierarchical architecture for scene text understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15672" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.17 + ], + "angle": 0, + "content": "[1] Ali Furkan Biten, Rubén Tito, Andrés Mafla, Lluis Gómez i Bigorda, Marçal Rusinol, C. V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4290-4300, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.469, + 0.214 + ], + "angle": 0, + "content": "[2] Samuele Capobianco, Leonardo Scommegna, and Simone Marinai. Historical handwritten document segmentation by using a weighted loss. In ANNPR, pages 395-406, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[4] Praneeth Chakravarthula, Jim Aldon D'Souza, Ethan Tseng, Joe Bartusek, and Felix Heide. Seeing with sound: Long-range acoustic beamforming for multimodal scene understanding. In CVPR, pages 982-991, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.469, + 0.455 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1280–1289, 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.468, + 0.498 + ], + "angle": 0, + "content": "[7] Chee-Kheng Ch'ng, Chee Seng Chan, and Cheng-Lin Liu. Total-text: toward orientation robustness in scene text detection. IJDAR, 23(1):31-52, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[8] Pengwen Dai, Sanyi Zhang, Hua Zhang, and Xiaochun Cao. Progressive contour regression for arbitrary-shape scene text detection. In CVPR, pages 7393-7402, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.544, + 0.469, + 0.598 + ], + "angle": 0, + "content": "[9] Bo Du, Jian Ye, Jing Zhang, Juhua Liu, and Dacheng Tao. I3CL: intra- and inter-instance collaborative learning for arbitrary-shaped scene text detection. IJCV, 130(8):1961-1977, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[10] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5354-5363, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.644, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[11] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.741 + ], + "angle": 0, + "content": "[12] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[13] Junjie He, Pengyu Li, Yifeng Geng, and Xuansong Xie. Fastinst: A simple query-based model for real-time instance segmentation. In CVPR, pages 23663-23672, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, pages 2980-2988, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[16] Joonho Lee, Hideaki Hayashi, Wataru Ohyama, and Seiichi Uchida. Page segmentation using a convolutional neural net" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "work with trainable co-occurrence features. In ICDAR, pages 1023-1028, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[17] Kai Li, Curtis Wigington, Chris Tensmeyer, Handong Zhao, Nikolaos Barmpalios, Vlad I. Morariu, Varun Manjunatha, Tong Sun, and Yun Fu. Cross-domain document object detection: Benchmark suite and method. In CVPR, pages 12912-12921, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[18] Minghui Liao, Baoguang Shi, Xiang Bai, Xinggang Wang, and Wenyu Liu. Textboxes: A fast text detector with a single deep neural network. In AAAI, pages 4161-4167, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.236, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[19] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In AAAI, pages 11474-11481, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, pages 936-944, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[21] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. PR, 90:337-345, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[23] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 19-35, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[24] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1039-1049, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[25] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. ICDAR 2023 competition on hierarchical text detection and recognition. In ICDAR, pages 483-497, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[27] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascadetabnet: An approach for end to end table detection and structure recognition from image-based documents. In CVPR, pages 2439-2447, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.688, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 91-99, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[29] Baoguang Shi, Xiang Bai, and Serge J. Belongie. Detecting oriented text in natural images by linking segments. In CVPR, pages 3482-3490, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, pages 8317-8326, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[31] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3591-3600, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Jingqun Tang, Wenqing Zhang, Hongye Liu, Mingkun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be bet-" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15673" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "ter than all: Feature sampling and grouping for scene text detection. In CVPR, pages 4553-4562, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.164 + ], + "angle": 0, + "content": "[33] Zhi Tian, Weilin Huang, Tong He, Pan He, and Yu Qiao. Detecting text in natural image with connectionist text proposal network. In ECCV, pages 56-72, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.167, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[34] Matheus Palhares Viana and Dario Augusto Borges Oliveira. Fast cnn-based document layout analysis. In ICCV, pages 1173-1180, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.21, + 0.469, + 0.264 + ], + "angle": 0, + "content": "[35] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan L. Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.268, + 0.469, + 0.309 + ], + "angle": 0, + "content": "[36] Renshen Wang, Yasuhisa Fujii, and Ashok C. Popat. Postocr paragraph recognition by graph convolutional networks. In WACV, pages 2533-2542, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[37] Wenhai Wang, Enze Xie, Xiang Li, Wenbo Hou, Tong Lu, Gang Yu, and Shuai Shao. Shape robust text detection with progressive scale expansion network. In CVPR, pages 9336-9345, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[38] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In ICCV, pages 8439-8448, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[39] Yuxin Wang, Hongtao Xie, Zheng-Jun Zha, Mengting Xing, Zilong Fu, and Yongdong Zhang. Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In CVPR, pages 11750-11759, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.54 + ], + "angle": 0, + "content": "[40] Guanghui Xu, Shuaicheng Niu, Mingkui Tan, Yucheng Luo, Qing Du, and Qi Wu. Towards accurate text-based image captioning with content diversity exploration. In CVPR, pages 12637-12646, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.542, + 0.469, + 0.597 + ], + "angle": 0, + "content": "[41] Yue Xu, Fei Yin, Zhaoxiang Zhang, and Cheng-Lin Liu. Multi-task layout analysis for historical handwritten documents using fully convolutional networks. In IJCAI, pages 1057-1063, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.469, + 0.64 + ], + "angle": 0, + "content": "[42] Chuhui Xue, Shijian Lu, and Wei Zhang. MSR: multi-scale shape regression for scene text detection. In IJCAI, pages 989–995, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[43] Chuhui Xue, Jiaxing Huang, Wenqing Zhang, Shijian Lu, Changhu Wang, and Song Bai. Contextual text block detection towards scene text understanding. In ECCV, pages 374-391, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.469, + 0.756 + ], + "angle": 0, + "content": "[44] Xiao Yang, Ersin Yumer, Paul Arente, Mike Kraley, Daniel Kifer, and C. Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In CVPR, pages 4342-4351, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[45] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[46] Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. Textfusenet: Scene text detection with richer fused features. In IJCAI, pages 516-522, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dpext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, 2023. 1, 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[48] Chengquan Zhang, Borong Liang, Zuming Huang, Mengyi En, Junyu Han, Errui Ding, and Xinghao Ding. Look more than once: An accurate detector for text of arbitrary shapes. In CVPR, pages 10552-10561, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[49] Shi-Xue Zhang, Xiaobin Zhu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Adaptive boundary proposal network for arbitrary shape text detection. In ICCV, pages 1285-1294, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[50] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. In CVPR, pages 2642-2651, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[52] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[53] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15674" + } + ] +] \ No newline at end of file diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_origin.pdf b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..438b714eebf6a30afd0c328fdcb967c58ed6f0b6 --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/09e4cb52-2d02-4e8e-9918-66498721ecee_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21904230b79dda0e28ea7988e8f2b97a188fa14739b47f6fb0ab64f555b0738b +size 6465719 diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/full.md b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..980b9ea1cba21cf806b033dea4985b8ed77c86fb --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/full.md @@ -0,0 +1,314 @@ +# LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding + +Min Liang, Jia-Wei Ma, Xiaobin Zhu*, Jingyan Qin, Xu-Cheng Yin +University of Science and Technology Beijing + +{lm20200126, mjw20151001}@@hotmail.com, {zhuxiaobin, xuchengyin}@ustb.edu.cn, qinjingyanking@foxmail.com + +# Abstract + +Existing scene text detectors generally focus on accurately detecting single-level (i.e., word-level, line-level, or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts, detecting multi-level texts while exploring their contextual information is critical. To this end, we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection, which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer, WordDecoder, LineDecoder, and ParaDecoder are proposed to be responsible for word-level text prediction, line-level text prediction, and paragraph-level text prediction, respectively. Meanwhile, WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships, respectively. In addition, we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection. + +# 1. Introduction + +Reading and understanding texts in scene images and digital documents is important in various real-world applications, including visual recognition [40], scene understanding [4], and text-based VQA [1, 30]. Compared to words, phrases and sentences express more semantic messages, and paragraphs convey richer contextual semantic messages. For better reading and understanding texts in scenes, it is necessary to simultaneously detect multi-level texts (i.e., word- + +![](images/7530836704e0217b162f3083a33106d8f66ccbf3885507e0d4da6e265955ca93.jpg) +Figure 1. Top: LayoutFormer is introduced for hierarchical text detection, which simultaneously detects multi-level texts (i.e., word-level, line-level, and paragraph-level) and predicts their geometric layouts. Bottom: Pipeline comparisons of (a) single-level detector, (b) two-level detector, and (c) hierarchical text detector. + +level for words, line-level for phrases and sentences, and paragraph-level for paragraphs) and predict their geometric layouts [16, 44], i.e., hierarchical text detection [25]. Hierarchical texts can deliver complete and meaningful text messages, which is beneficial for various downstream tasks. + +Although existing scene text detectors have achieved promising performance, they [32, 47] often concentrate on detecting single-level text, which decodes the encoded features of an input image into their representation of independent text entities, e.g., words, lines, or paragraphs. As shown in Fig. 1 (a), single-level detectors only predict one level of text results and do not explore the contextual relationship among different levels of texts. This makes them unable to deliver more informative messages for facilitating scene text understanding. Recently, some methods have investigated two-level text detection. A schematic of the framework is shown in Fig. 1 (b). CUTE [43] proposes a + +two-stage network, which first adopts a Transformer structure like DETR [3] to perform word/line bounding boxes and then directly learns the sequential relationships between detected text boxes based on visual features after ROI operation and positional information. Unified Detector [24] proposes a unified scene text detection and layout analysis network, which first uses a Transformer-based structure [35] to generate word/line mask maps and then models their affiliate relationships. These two methods are all supervised by an affinity matrix or index label, directly modelling relationships on the word-level/line-level detection results. They only rely on the former detection results to model contextual relationships, making the model's performance highly limited by the text detection results at the first stage. Moreover, this paradigm cannot be easily extended to the task of text detection at more than two levels. + +Existing scene text detectors generally adopt multi-scale feature maps that contain rich information for accurate prediction. However, methods based on the Transformer encoder-decoder structure often suffer from huge computational and memory costs on high-resolution feature maps. This is because the attention weight computation in the Transformer decoder is of linear computation w.r.t. pixel numbers. There have been several work to tackle the issues. Deformable DETR [51] introduces the multi-scale deformable attention module to select a small set of key locations. AdaMixer [10] adaptively samples features over the space and scales of an object. However, these methods are based on estimated offsets of instance coordinates, not applicable to segmentation methods. In image segmentation, Mask2Former [6] claims that local features can well update query features and proposes to use image features of foreground regions. TSP [31] adopts a RoIAlign operation to extract features of RoIs. However, these methods still suffer from complex operations. We think that it is enough to update query features with only foreground features of a few representative pixels. + +In this paper, we propose a hierarchical text detector named LayoutFormer, which can simultaneously produce hierarchical text detection boxes and their layout relationships. Unlike using inter-level relationships as supervision, we directly adopt multi-level text boxes as supervision, which can significantly reduce the dependence on the former text detection results and implicitly exploit multi-level textual information. In LayoutFormer, we propose three Transformer-based modules, i.e., WordDecoder, LineDecoder, and ParaDecoder, for detecting word-level texts, line-level texts, and paragraph-level texts, respectively. We first adopt text line as the detection unit, and LineDecoder updates and forms line query features. Because a line can be split into several words, WordDecoder learns the deviation of each word within a line based on the line query features and forms final word query features. ParaDecoder aggre + +gates line query features by learning line-paragraph relationships and forms final paragraph query features. In addition, we propose a Prior Location Sampler, which adaptively selects a few representative features that are highly relevant to scene texts. The sampled features are used for cross-attention operation to update query features. Prior Location Sampler can improve the model performance while reducing the training cost. Extensive experiments verify the superior performance on hierarchical text detection. + +In summary, our contributions are as follows: + +- We propose an innovative hierarchical text detector that simultaneously detects multi-level text instances and explores their contextual information for predicting the geometric layouts, finally performing hierarchical outputs. +- We propose WordDecoder, LineDecoder, and ParaDecoder in LayoutFormer to decode word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively. +- We propose a Prior Location Sampler to adaptively select a few representative foreground features for updating queries, which can significantly reduce the computational cost while improving the detection performance. +- Extensive experiments on multiple publicly available datasets verify the state-of-the-art performance of our LayoutFormer. + +# 2. Related Work + +# 2.1. Scene Text Detection + +Deep learning-based scene text detection methods have been widely investigated and roughly divided into bottom-up and top-down methods. The bottom-up methods generally detect local components [23, 29, 33] or text pixels [19, 37] for grouping text instances. TextSnake [23] describes a text instance as a sequence of ordered, overlapping disks. The pixel-based methods use some auxiliary information (e.g., similarity vectors in PAN [38], and threshold map in DB [19]) or post-processings (e.g., a progressive scale expansion algorithm in PSENet [37]) to generate text instances. The top-down methods [18, 50] directly predict bounding boxes of scene text instances. To detect arbitrary-shaped texts, some methods [39, 49, 52] localize the key points on the contours of scene texts. Recently, DETR [3] presents a Transformer-based architecture for object detection and achieved great success. Inspired by it, many researchers [24, 32, 47] apply the Transformer structure to scene text detection. Overall, existing scene text detection methods mainly address single-level text detection, ignoring the text detection of multiple levels and further predicting the hierarchical relationships. + +![](images/fa45a20519ffc544cc0c89557cd32075adb119f536671e24b81d9b556f7031c9.jpg) +Figure 2. Illustration of the proposed LayoutFormer, which mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders (WordDecoder, LineDecoder, and ParaDecoder) and three predictors. $M_l$ denotes predicted mask maps of text lines by $l$ -th Transformer decoder, and $F_l$ denotes the corresponding image features generated by $M_l$ . We first predict text lines, and then predict words and paragraphs. In testing, word-level results produce line-level results based on word-line relationships, and then paragraph-level results are produced based on line-paragraph relationships, finally obtaining hierarchical text detection results. + +
Detection UnitQueryMemory/GBPQ
word25612.153.09
line25610.657.48
paragraph2568.540.34
+ +Table 1. Comparisons of word, line, and paragraph as text detection unit on the HierText validation set, respectively. "Query" denotes the number of text queries. "Memory" denotes GPU Memory consumed in training. + +# 2.2. Layout Analysis + +Document Layout analysis has made great progress, which can be mainly divided into object detection-based methods [17, 27, 34] and segmentation-based methods [2, 16, 41]. Inspired by object detection [15, 28] and semantic segmentation [5, 12, 22], these methods treat semantically coherent text blocks as a special kind of object. However, they fail to produce word or line-level detections and can only be used in companies with standalone text detectors, increasing the complexity of the pipeline. Some work [36] takes a hierarchical view and applies graph-based models on the finest granularity, i.e., individual words, to analyze the layout. + +Recently, some work has investigated layout analysis in scene text detection. CUTE [43] detects contextual text blocks consisting of one or multiple ordered integral text units. Their model is a two-stage structure, which first detects integral text units and then models the relationship of integral text units by a designed relation module. Unified Detector [24] introduces a novel task of unified scene text + +detection and layout analysis, which models the two closely related tasks with a unified model. It produces mask outputs for the text detection task and an affinity matrix for the layout analysis task. The above methods are highly dependent on text detection results at the first stage. Differently, we use multi-level text boxes as supervision to learn the relationship among different levels for layout analysis. + +# 3. Methodology + +# 3.1. Preliminary + +How to effectively model hierarchical text detection tasks? Is the top-down or bottom-up approach better? As we know, the text layouts in real scene images is highly complex, bringing significant challenges to hierarchical text detection. According to experiments and analyses, we found that paragraph-level text boxes are considerably irregular and always contain much background, which makes existing scene text detection approaches difficult to model text instances. For example, as shown in the second row of Fig. 3, a paragraph text box annotates the entire image. As listed in Tab. 1, the detection performance of paragraphs is much lower than that of words and lines. Thus, the top-down modeling approach is considered inappropriate. Both word and line are generally used as text detection units. However, text lines always contain more complete semantic information than words. The aspect ratios and sizes of words are considerably more variable than text lines, making detection more challenging. Meanwhile, if the word is used + +![](images/78c6e6b056116af6e2270121115adc3b256ccb815d7f857326ba933bde754c8b.jpg) +Figure 3. Some examples of word-level (first column), line-level (second column), and paragraph-level (third column) text boxes. + +as the detection unit, the detection error will accumulate bigger after two aggregation processes. Hence, we finally choose line-level as our detection unit, generate word-level texts by splitting them downward, and get paragraph-level texts by aggregating them upward. + +# 3.2. Network Architecture + +The architecture of our LayoutFormer is illustrated in Fig. 2. It mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders and three predictors, respectively, corresponding to word-level text detection, line-level text detection and paragraph-level text detection. Specifically, given an input image $I \in R^{H \times W \times 3}$ , the feature encoder with a ResNet-50 [14] backbone and a Transformer encoder [6] extracts and enhances the features and then generates multi-scale feature maps, i.e., $C_4 \in R^{\frac{H}{32} \times \frac{W}{32} \times C}$ , $C_3 \in R^{\frac{H}{16} \times \frac{W}{16} \times C}$ , $C_2 \in R^{\frac{H}{8} \times \frac{W}{8} \times C}$ and $C_1 \in R^{\frac{H}{4} \times \frac{W}{4} \times C}$ . The typical value we use is $C = 256$ . Then, Prior Location Sampler adaptively selects image features for the Transformer decoder. Afterwards, the LineDecoder, WordDecoder, and ParaDecoder update line query features $X_{line} \in R^{N_{line} \times C}$ , word query features $X_{word} \in R^{N_{word} \times C}$ , and paragraph query features $X_{para} \in R^{N_{para} \times C}$ , respectively. The three predictors convert $X_{line}$ , $X_{word}$ , and $X_{para}$ to line-level prediction $(P_{line}, M_{line})$ , word-level prediction $(P_{word}, M_{word})$ , and paragraph-level prediction $(P_{para}, M_{para})$ , where $P_* \in R^{N_* \times 2}$ is text/non-text probability values of $N_*$ instances and $M_* \in R^{N_* \times \frac{H}{4} \times \frac{W}{4}}$ is mask maps of $N_*$ instances. + +# 3.3. Prior Location Sampler + +Mask2Former [6] proposes masked attention to attend within the foreground region of the predicted mask for each query, which can be computed as + +$$ +X _ {l} = \operatorname {s o f t m a x} \left(B _ {l - 1} + Q _ {l} K _ {l} ^ {T}\right) V _ {l} + X _ {l - 1}, \tag {1} +$$ + +where $l$ is the layer index, $B_{l-1}$ is the binary mask prediction of $l-1$ -th layer, $X_{l}$ is query features at the $l$ -th layer, $Q_{l}$ + +![](images/5666002bc55488412c0a002c68d73966bf59fafa8a2098c8fda747d502823a03.jpg) +Figure 4. Illustration of the Prior Location Sampler, which selects representative foreground features with Top- $K$ scores in predicted mask maps and then aggregates with weights. + +is query features $X_{l-1}$ under linear transformation, $K_l$ and $V_l$ are the input image features under linear transformation. $X_0$ denotes input query features to the Transformer decoder. $B_0$ is the binary mask prediction obtained from $X_0$ . The input image features are the entire multi-scale feature maps from a feature encoder. Their computational complexity of Eq. (1) is $O(NH'W'C)$ , where $H'$ and $W'$ are the height and width of a specific feature map respectively, and $N$ is the number of queries. The cross-attention operation suffers from a linear complexity growth with the spatial size of feature maps. Moreover, the main accompanying problem is the high demand for computing resources. + +Thus, we propose the Prior Location Sampler to select representative foreground features with Top- $K$ scores in mask maps and obtain the image features $F_{l}\in R^{K\times N\times C}$ for $l$ -th decoder layer. Our computational complexity of Eq. (1) is $O(NKC)$ . In our implementation, the value $K\ll H^{\prime}W^{\prime}$ , and thus the complexity of our cross-attention operation could be significantly reduced. The computational resource consumption due to an increase in the number of feature maps or an increase in the scale of feature maps is negligible. + +The architecture of the Prior Location Sampler is illustrated in Fig. 4. Specifically, $M_{l}$ denotes the predicted mask maps of $l$ -th decoder layer, and + +$$ +M _ {l} = \left\{M _ {l n} \in R ^ {\frac {H}{4} \times \frac {W}{4}} | n = 1, 2, \dots , N \right\}. \tag {2} +$$ + +For each $M_{ln}$ , we use 0.5 for thresholding. We sort scores in $M_{ln}$ and select features with Top- $K$ scores [13] in multi-scale feature maps $\{C_i|i = 1,2,3,4\}$ respectively, where $f_{K}(\cdot)$ is sampling operation. Then, we assign [20, 28] text ROIs of different scales to the pyramid levels for calculating weights [10] as Eq. (4), where $S_{ln}$ is the cumulative sum of the binarized $M_{ln}$ , i.e., the area of a candidate text, and 56 [20] is a scaling factor. The selected features for $M_{ln}$ are gathered into $F_{ln}\in R^{K\times C}$ : + +$$ +F _ {l n} = \sum_ {i = 1} ^ {4} w _ {l i} f _ {K} \left(C _ {i}, M _ {l n}\right), \tag {3} +$$ + +$$ +t _ {l n} = \left\lfloor \log_ {2} \frac {S _ {l n}}{5 6} \right\rfloor , \tag {4} +$$ + +$$ +w _ {l} = \text {s o f t m a x} \left[ - \frac {\left(t _ {l n} - \iota + 1\right) ^ {2}}{2} \right]. +$$ + +The final features $F_{l}\in R^{K\times N\times C}$ are denoted as: + +$$ +F _ {l} = \left\{F _ {l n} \mid n = 1, 2, \dots , N \right\}. \tag {5} +$$ + +# 3.4. Transformer Decoder + +LineDecoder. We feed the selected image features $F_{l-1} \in R^{K \times N_{line} \times C}$ and $N_{line}$ learnable line queries into LineDecoder, where $K = 32$ . LineDecoder uses the standard Transformer decoder structure, consisting of self-attention, cross-attention, and feedforward network. We also add the positional embeddings to queries and keys at every self-attention and cross-attention layer. We refine line query features layer-by-layer by setting $L_1 = 4$ . The updated line query features are adopted to predict $l$ -th mask maps, which are fed into Prior Location Sampler to generate image features for the next decoder layer. + +WordDecoder. We frist hypothesize that there are $N_{per\_line}$ words in each line and initialize $N_{per\_line}$ learnable word queries. Thus, the number of words in an image can be formulated as the product of the number of lines in an image and words per line, as + +$$ +N _ {\text {w o r d}} = N _ {\text {l i n e}} \times N _ {\text {p e r l i n e}}. \tag {6} +$$ + +So, we can decouple the self-attention between words into self-attention between lines and self-attention between words per line. Based on line queries from LineDecoder, we only need to compute self-attention between $N_{per\_line}$ , greatly reducing the computational cost. Then we obtain word query features $X_{word} \in R^{N_{word} \times C}$ by + +$$ +X _ {\text {w o r d}} = X _ {\text {l i n e}} X _ {\text {p e r - l i n e}} ^ {T}, \tag {7} +$$ + +where $X_{line} \in R^{N_{line} \times 1 \times C}$ is line query features and $X_{per\_line} \in R^{N_{per\_line} \times 1 \times C}$ is learnable word query features per line. Then cross-attention followed by feedforward network adopts image features $F_{l} \in R^{K \times N_{line} \times C}$ with $K = 64$ to update $X_{word}$ . The image features are obtained by Prior Location Sampler, which feeds the predicted mask maps of the last decoder layer in LineDecoder. In our experiments, we set $L_{3} = 1$ . With our word decoupling mechanism, every $N_{per\_line}$ queries predicts words within a text line for capturing word-line relationships. + +ParaDecoder. We adaptively learn line-paragraph relationships and aggregate line query features for paragraph-level texts to get paragraph query features. We initialize $N_{para}$ learnable paragraph queries. ParaDecoder also consists of self-attention, cross-attention, and feedforward networks. It uses $X_{line}$ line query features as inputs of cross attention + +to update paragraph queries. Positional embeddings are added to queries and keys at every self-attention and cross-attention layer. In our experiments, we adopt $X_{line}$ of last decoder layer in LineDecoder and set $L_{2} = 1$ . In this process, the learned attention weights $(R^{N_{para} \times N_{line}})$ in cross-attention of ParaDecoder are line-paragraph relationships. + +# 3.5. Optimization + +Training. Instead of relation matrices, we use multi-level text boxes as supervision. The total loss is formulated as: + +$$ +L = L _ {\text {l i n e}} + L _ {\text {w o r d}} + L _ {\text {p a r a}}, \tag {8} +$$ + +where $L_{line}$ , $L_{word}$ , and $L_{para}$ respectively denote losses of line-level, word-level, and paragraph-level, which can be formulated as: + +$$ +\begin{array}{l} L _ {*} = \lambda_ {c l s} L _ {c l s} \left(P _ {*}, P _ {*} ^ {\prime}\right) + \lambda_ {m a s k} L _ {m a s k} \left(M _ {*}, M _ {*} ^ {\prime}\right), \\ * \in \{w o r d, l i n e, p a r a \}, \tag {9} \\ \end{array} +$$ + +where $P_{*}$ and $P_{*}^{\prime}$ are the class predictions and their corresponding ground truth, $M_{*}$ and $M_{*}^{\prime}$ are the mask predictions and their corresponding ground truth, $L_{cls}$ denotes binary cross-entropy loss, $L_{mask}$ is the sum of binary cross-entropy loss and dice loss. We set $\lambda_{cls} = 2.0$ for predictions matched with ground truth and 0.1 for unmatched predictions. $\lambda_{mask}$ is set to 5.0. + +Inference. We present a simple inference procedure that converts class predictions and mask predictions to text boxes, which mainly consists of four steps: (1) the text queries whose class scores are over 0.3 will be selected; (2) the mask predictions of selected queries are binarized by 0.5 to get the binary maps; (3) the connected regions are obtained from the binary maps; (4) the text queries whose averaged foreground mask probability is below 0.9 will be filtered. For hierarchical text detection task, we frist generate the word-level detection boxes by the above inference procedure. Then, we use word-line relationships from WordDecoder to form a word-line geometric layouts and further use line-paragraph relationships from ParaDecoder to form a line-paragraph geometric layouts, finally performing hierarchical text detection outputs. + +# 4. Experiments + +# 4.1. Datasets + +HierText [24]. It is a hierarchical text detection dataset consisting of 8,281 training images, 1,724 validation images, and 1,634 testing images. It annotates images hierarchically, which first annotates word locations with polygons, then clusters words into lines and lines into paragraphs. + +MSRA-TD500 [45]. It is a line-level annotated arbitrary-oriented long text dataset. It consists of 300 training images and 200 testing images collected from natural scenes. + +
MethodTotal-TextHierText-Line
RPFMemory/GBRPFTPQMemory/GB
C2-C4masked attention84.6887.7986.2115.365.5682.7373.1577.5756.7418.1
C2-C4Top-K&concat84.6687.8086.2015.165.3482.5672.9577.3756.4418.3
C1-C4Top-K&concat84.4688.1586.2716.265.8882.9673.4477.5356.9420.3
C1-C4Top-K&weight85.0789.2687.1214.266.6584.1074.3777.2957.4817.0
+ +CTW1500 [21]. It is an arbitrary-shaped scene text dataset that consists of 1,000 training images and 500 testing images. In this dataset, the annotations of text instances are line-level and labelled by a polygon with 14 key points. + +Total-Text [7]. It is an arbitrary-shaped scene text dataset that contains 1,255 training and 300 testing images, which are annotated by the word-level polygon with 14 key points. + +# 4.2. Implementation Details + +In our experiments, for HierText, we only use the training images of HierText to train the models by $60k$ steps for single-level text detection and $80k$ steps for multi-level text detection. In testing, we keep the aspect ratio of testing images and resize them to 1,120 height. For MSRA-TD500, CTW1500, and Total-Text, we adopt SynthText [11] to pretrain the models by $150k$ steps. Then, we finetune the models on the corresponding real-world datasets by $40k$ steps. In testing, we keep the aspect ratio of testing images and resize them to 800 height on three datasets. + +In training, we use AdamW [26] optimizer and the step learning rate schedule with an initial learning rate of 0.0001 and a weight decay of 0.05. A learning rate multiplier of 0.1 is applied to the ResNet-50 backbone. We decay the learning rate at 0.9 and 0.95 fractions of the total number of training steps by a factor of 10. We train our models with a batch size of 8. Data augmentation includes: (1) Randomly Flipping; (2) Randomly rotate them in range $(-10^{\circ}$ to $10^{\circ})$ ; (3) Randomly resize them in range (0.5 to 3.0); (4) Randomly cropping. Finally, we resize the images to $640\times 640$ + +# 4.3. Evaluation Metrics + +For MSRA-TD500, CTW1500, and CTW1500 datasets, we follow the standard evaluation protocol Recall (R), Precision (P), and F-measure (F). For the HierText dataset, we follow [24] and adopt Panoptic Quality (PQ) as the main evaluation metric for the hierarchical text detection task. + +# 4.4. Ablation Studies + +$K$ value in Prior Location Sampler. We compare the different Top- $K$ values in Prior Location Sampler. Results are listed in Tab. 3. Selecting too few foreground points may fail to capture and learn the features of an entire text. Select + +Table 2. Ablation comparison of Prior Location Sampler on the Total-Text and line-level of the HierText validation set. "Memory" denotes GPU Memory consumed in training. $C_1 - C_4$ denotes the multi-scale feature maps used in the Transformer decoder. The results show that "Top-K&weight" with $C_1 - C_4$ achieves a considerable improvement and consumes fewer training resources than other methods. + +
KTotal-TextHierText-Line
RPFRPFPQ
1684.3488.0186.1465.3481.1272.3855.54
3285.0789.2687.1266.6584.1074.3757.48
6484.7588.6386.6468.6784.4475.7558.86
+ +Table 3. Ablation comparison of Top- $K$ value on Total-Text and line-level of the HierText validation set. + +![](images/78a97c23a2766b7a8dde7b68011fdf84741510fb585cb31ff5f4369067463a0a.jpg) +Figure 5. Visualization of Top- $K$ sampling points in Prior Location Sampler. + +ing too many foreground points may introduce noise, which is detrimental to the performance of the model. For Total-Text, we set $K$ to 32, and the same for MSRA-TD500 and CTW1500. While text lines of HierText may have larger aspect ratios and scales, the performance of $K = 64$ is better than that of $K = 32$ . However, we finally set $K$ to 32 for a trade-off between performance and efficiency. + +Implementation of Prior Location Sampler. We evaluate the effectiveness of the Prior Location Sampler on the Total-Text and HierText validation set. Results are listed in Tab. 2. " $C_1 - C_4$ " denotes that multi-scale feature maps $\{C_i | i = 1, 2, 3, 4\}$ are used in Transformer decoder. "masked attention" is followed from Mask2Former [6], which calculates the attention matrix by additionally adding a binary mask prediction. "Top-K&concat" applies Top-K strategy to multi-scale feature maps and then performs concatenation. "Top-K&weight" applies Top-K strategy to multiscale feature maps and then element-wise sums in weights. Inspired by Mask2Former, the first row of Tab. 2 feeds the entire feature map into the Transformer decoder. The experimental results show that "Top-K&weight" apparently achieves a significant improvement in harmonic accuracy and memory. Meanwhile, due to using a small number of foreground points, our method consumes fewer training resources (GPU Memory) than other methods. + +
MethodWordLineParagraph
RPFTPQRPFTPQRPFTPQ
LayoutFormer57.0475.2564.8975.8849.2450.3976.3260.7076.1246.2145.8973.1456.4076.1942.97
+ +Table 4. Experimental results of word-line-paragraph three-level text detection on the HierText testing set. + +
MethodNperLineNlineWordLine
PQPQ
In the below experiments, we always set L3 to 1.
Xline1012044.0236.35
Fl1012050.2948.60
In the below experiments, we use Fl as image features.
L3 = 11012050.2948.60
L3 = 41012051.0744.05
L3 = 1612046.2139.60
L3 = 1620050.1246.00
+ +Table 5. Ablation comparison of WordDecoder on the HierText validation set. $N_{per\_line}$ and $N_{line}$ denote the number of word queries in each line and text line queries, respectively. + +
MethodNlineNparaLineParagraph
PQPQ
In the below experiments, we always set L2 to 1.
Xline12812858.3151.98
Fl12812857.2746.15
+ +Implementation of WordDecoder. We conduct ablation experiments by comparing different settings of image features, $L_{3}$ , $N_{per\_line}$ and $N_{line}$ in WordDecoder. Results are listed in Tab. 5. Using $F_{l}$ as image features for updating word queries is better than $X_{line}$ . This is because to identify different text lines, $X_{line}$ extracts consistent features of each text line, whereas $X_{per\_line}$ needs to learn differentiated features within a text line to identify different words within a line. In addition, we find that $L_{3}$ being set to 1 performs better than being set to 4. Thus, we adopt $L_{3} = 1$ by default. We also conduct comparative experiments with different $N_{per\_line}$ and $N_{line}$ . We can conclude that the larger the query number, the higher the performance. + +Implementation of ParaDecoder. We compare different image features for ParaDecoder. Results are listed in Tab. 6. Adopting $X_{line}$ as image features is better than $F_{l}$ . We think that after the learning of LineDecoder, $X_{line}$ is able to characterize text lines well. + +# 4.5. Comparisons with State-of-the-art Methods + +Hierarchical Text Detection. In this section, we evaluate the performance on the HierText testing set for hierarchical + +Table 6. Ablation comparison of ParaDecoder on the HierText validation set. $N_{line}$ and $N_{para}$ denote the number of text line queries and paragraph queries, respectively. + +
MethodNwordNlineWordLine
PQPQ
LayoutFormer120012050.3548.62
+ +Table 7. Experimental results of word-line two-level text detection on the HierText testing set. $N_{word}$ and $N_{line}$ denote the number of word queries and text line queries, respectively. + +![](images/1fa2316c2eec527dfe017818d4a7cdec7c32eb241944e717dded40c382d86115.jpg) +Figure 6. The hierarchical text detection examples on the HierText. The visualization images in each row are in turn the input image, word-level detection results, line-level detection results, and paragraph-level detection results. + +text detection. To our knowledge, few work has conducted research on hierarchical text detection. Hence, we mainly compare with [24]. Results are listed in Tab. 7, Tab. 9, and Tab. 4. Meanwhile, we demonstrate hierarchical detection results on images in Fig. 6. Both qualitative and quantitative verify the advances of our method. And we can observe that our method can precisely detect hierarchical texts. + +For word-line detection, our results are listed in Tab. 7. Because there is no any prior research on this, we provide a baseline to facilitate comparisons with other methods. Benefiting from our decoupling operation on words, our $N_{word}$ can reach 1,200, much higher than Unified Detector [24]. + +For line-paragraph detection, our LayoutFormer achieves the best text detection and layout analysis performance. Results are listed in Tab. 9. "GCP API", "GCN-PP", "Mask-RCNN-Cluster", "Max-DeepLab-Cluster" are two-stage approaches, while "Unified Detector" is an end-to-end unified approach. Our LayoutFormer with + +
MethodVenueLine DetectionWord Detection
MSRA-TD500CTW1500Total-Text
RPFRPFRPF
TextSnake [23]ECCV1873.983.278.367.985.375.674.582.778.4
MSR [42]IJCAI1976.787.481.777.883.880.773.085.278.6
LOMO [48]CVPR19---76.585.780.879.387.683.3
PAN [38]ICCV1983.884.484.181.286.483.781.089.385.0
ContourNet [39]CVPR20---84.183.783.983.986.985.4
DB [19]AAAI2079.291.584.980.286.983.482.587.184.7
TextFuseNet [46]IJCAI20---85.085.885.483.287.585.3
FCENet [53]CVPR21---83.487.685.582.589.385.8
BPNet [49]ICCV2180.6885.4082.9781.4587.8184.5184.6590.2787.37
PCR [8]CVPR2183.590.887.082.387.284.782.088.585.2
I3CL [9]IJCV22---84.587.485.983.789.286.3
FSG [32]CVPR2284.891.688.182.488.185.285.790.788.1
Unified Detector [24]CVPR2387.4488.0487.7087.4484.5685.9791.0684.9687.90
DPText-DETR [47]AAAI23---86.291.788.886.491.889.0
LayoutFormer-88.3291.9590.1084.2688.1686.1785.0789.2687.12
+ +Table 8. Experimental results of single-level text detection. We set $N$ to 100 on three datasets. On MSRA-TD500 and CTW1500, our LayoutFormer achieves the state-of-the-art performance. On Total-Text, our LayoutFormer achieves competitive results. + +
MethodNlineNparaLineParagraph
PQPQ
GCP API [24]--56.1746.33
GCN-PP [24]384-62.2350.10
Mask-RCNN-Cluster [24]384-62.2351.67
Max-DeepLab-Cluster [24]384-62.2352.52
128-58.7651.48
Unified Detector [24]256--52.50
384-62.2353.60
LayoutFormer38420062.3753.76
+ +Table 9. Experimental results of line-paragraph two-level text detection on the HierText testing set. $N_{line}$ denotes the number of text line queries. + +$N_{line} = 384$ outperforms Unified Detector [24] with $N_{line} = 384$ by $0.14\%$ and $0.16\%$ in terms of PQ of text detection and layout analysis, respectively. + +For word-line-paragraph detection, there is no any prior research on three-level text detection and layout analysis. We only list our performance on Tab. 4, which can be used by subsequent methods for comparison. Our Layout-Former achieves $49.24\%$ , $46.21\%$ , and $42.97\%$ in terms of PQ of word-level, line-level, and paragraph-level, respectively. The inference speed of our method is 3.0 FPS. + +Single-level Text Detection. In this section, we evaluate the performance of our model on the most widely used benchmarks for single-level scene text detection, i.e., MSRAD500 and CTW1500 for line-level, Total-Text for word-level. Results are listed in Tab. 8. The quantitative results demonstrate the advances of our model. + +For line detection, we achieve the state-of-the-art results + +on MSRA-TD500. Notably, our LayoutFormer achieves $88.32\%$ , $91.95\%$ , and $90.10\%$ in terms of Recall, Precision, and F-measure, respectively, which significantly outperforms other methods with a great margin. For example, LayoutFormer outperforms PCR [8], FSG [32], and Unified Detector [24] by $3.10\%$ , $2.00\%$ , $2.40\%$ in terms of F-measure, respectively. On CTW1500, LayoutFormer achieves $84.26\%$ , $88.16\%$ , and $86.17\%$ in terms of Recall, Precision, and F-measure, respectively, achieving the promising performance. + +For word detection, we achieve competitive results on Total-Text. The performance of our LayoutFormer is slightly lower than FSG [32] and Unified Detector [24], in which the former utilizes more training epochs and the latter utilizes larger training datasets and more training epochs. DPText-DETR [47] pre-trains their models on more datasets and these datasets are specifically for improving the performance of arbitrary-shape texts. + +# 5. Conclusion + +In this work, we propose a hierarchical text detector for promoting scene text understanding, which simultaneously detects multi-level texts and predicts their geometric layouts. In LayoutFormer, we propose WordDecoder, LineDecoder, and ParaDecoder to detect word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively. Through this research, we hope to provide a baseline for hierarchical text detection and inspire other hierarchical tasks. In addition, we will continue to promote and improve the hierarchical architecture for scene text understanding. + +# References + +[1] Ali Furkan Biten, Rubén Tito, Andrés Mafla, Lluis Gómez i Bigorda, Marçal Rusinol, C. V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4290-4300, 2019. 1 +[2] Samuele Capobianco, Leonardo Scommegna, and Simone Marinai. Historical handwritten document segmentation by using a weighted loss. In ANNPR, pages 395-406, 2018. 3 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 2 +[4] Praneeth Chakravarthula, Jim Aldon D'Souza, Ethan Tseng, Joe Bartusek, and Felix Heide. Seeing with sound: Long-range acoustic beamforming for multimodal scene understanding. In CVPR, pages 982-991, 2023. 1 +[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2018. 3 +[6] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1280–1289, 2022. 2, 4, 6 +[7] Chee-Kheng Ch'ng, Chee Seng Chan, and Cheng-Lin Liu. Total-text: toward orientation robustness in scene text detection. IJDAR, 23(1):31-52, 2020. 6 +[8] Pengwen Dai, Sanyi Zhang, Hua Zhang, and Xiaochun Cao. Progressive contour regression for arbitrary-shape scene text detection. In CVPR, pages 7393-7402, 2021. 8 +[9] Bo Du, Jian Ye, Jing Zhang, Juhua Liu, and Dacheng Tao. I3CL: intra- and inter-instance collaborative learning for arbitrary-shaped scene text detection. IJCV, 130(8):1961-1977, 2022. 8 +[10] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5354-5363, 2022. 2, 4 +[11] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6 +[12] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. 2022. 3 +[13] Junjie He, Pengyu Li, Yifeng Geng, and Xuansong Xie. Fastinst: A simple query-based model for real-time instance segmentation. In CVPR, pages 23663-23672, 2023. 4 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, pages 2980-2988, 2017. 3 +[16] Joonho Lee, Hideaki Hayashi, Wataru Ohyama, and Seiichi Uchida. Page segmentation using a convolutional neural net + +work with trainable co-occurrence features. In ICDAR, pages 1023-1028, 2019. 1, 3 +[17] Kai Li, Curtis Wigington, Chris Tensmeyer, Handong Zhao, Nikolaos Barmpalios, Vlad I. Morariu, Varun Manjunatha, Tong Sun, and Yun Fu. Cross-domain document object detection: Benchmark suite and method. In CVPR, pages 12912-12921, 2020. 3 +[18] Minghui Liao, Baoguang Shi, Xiang Bai, Xinggang Wang, and Wenyu Liu. Textboxes: A fast text detector with a single deep neural network. In AAAI, pages 4161-4167, 2017. 2 +[19] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In AAAI, pages 11474-11481, 2020. 2, 8 +[20] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, pages 936-944, 2017. 4 +[21] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. PR, 90:337-345, 2019. 6 +[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3 +[23] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 19-35, 2018. 2, 8 +[24] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1039-1049, 2022. 2, 3, 5, 6, 7, 8 +[25] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. ICDAR 2023 competition on hierarchical text detection and recognition. In ICDAR, pages 483-497, 2023. 1 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 +[27] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascadetabnet: An approach for end to end table detection and structure recognition from image-based documents. In CVPR, pages 2439-2447, 2020. 3 +[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 91-99, 2015. 3, 4 +[29] Baoguang Shi, Xiang Bai, and Serge J. Belongie. Detecting oriented text in natural images by linking segments. In CVPR, pages 3482-3490, 2017. 2 +[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, pages 8317-8326, 2019. 1 +[31] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3591-3600, 2021. 2 +[32] Jingqun Tang, Wenqing Zhang, Hongye Liu, Mingkun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be bet- + +ter than all: Feature sampling and grouping for scene text detection. In CVPR, pages 4553-4562, 2022. 1, 2, 8 +[33] Zhi Tian, Weilin Huang, Tong He, Pan He, and Yu Qiao. Detecting text in natural image with connectionist text proposal network. In ECCV, pages 56-72, 2016. 2 +[34] Matheus Palhares Viana and Dario Augusto Borges Oliveira. Fast cnn-based document layout analysis. In ICCV, pages 1173-1180, 2017. 3 +[35] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan L. Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 2 +[36] Renshen Wang, Yasuhisa Fujii, and Ashok C. Popat. Postocr paragraph recognition by graph convolutional networks. In WACV, pages 2533-2542, 2022. 3 +[37] Wenhai Wang, Enze Xie, Xiang Li, Wenbo Hou, Tong Lu, Gang Yu, and Shuai Shao. Shape robust text detection with progressive scale expansion network. In CVPR, pages 9336-9345, 2019. 2 +[38] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In ICCV, pages 8439-8448, 2019. 2, 8 +[39] Yuxin Wang, Hongtao Xie, Zheng-Jun Zha, Mengting Xing, Zilong Fu, and Yongdong Zhang. Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In CVPR, pages 11750-11759, 2020. 2, 8 +[40] Guanghui Xu, Shuaicheng Niu, Mingkui Tan, Yucheng Luo, Qing Du, and Qi Wu. Towards accurate text-based image captioning with content diversity exploration. In CVPR, pages 12637-12646, 2021. 1 +[41] Yue Xu, Fei Yin, Zhaoxiang Zhang, and Cheng-Lin Liu. Multi-task layout analysis for historical handwritten documents using fully convolutional networks. In IJCAI, pages 1057-1063, 2018. 3 +[42] Chuhui Xue, Shijian Lu, and Wei Zhang. MSR: multi-scale shape regression for scene text detection. In IJCAI, pages 989–995, 2019. 8 +[43] Chuhui Xue, Jiaxing Huang, Wenqing Zhang, Shijian Lu, Changhu Wang, and Song Bai. Contextual text block detection towards scene text understanding. In ECCV, pages 374-391, 2022. 1, 3 +[44] Xiao Yang, Ersin Yumer, Paul Arente, Mike Kraley, Daniel Kifer, and C. Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In CVPR, pages 4342-4351, 2017. 1 +[45] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090, 2012. 5 +[46] Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. Textfusenet: Scene text detection with richer fused features. In IJCAI, pages 516-522, 2020. 8 +[47] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dpext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, 2023. 1, 2, 8 + +[48] Chengquan Zhang, Borong Liang, Zuming Huang, Mengyi En, Junyu Han, Errui Ding, and Xinghao Ding. Look more than once: An accurate detector for text of arbitrary shapes. In CVPR, pages 10552-10561, 2019. 8 +[49] Shi-Xue Zhang, Xiaobin Zhu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Adaptive boundary proposal network for arbitrary shape text detection. In ICCV, pages 1285-1294, 2021. 2, 8 +[50] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. In CVPR, pages 2642-2651, 2017. 2 +[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 2 +[52] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 2 +[53] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 8 \ No newline at end of file diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/images.zip b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0286aa399cde2c682dbc9897b4e7ce2f6a0fee8f --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61396100e8db9d26561439006d5cb554030a4f7e55a9bea473a782c16e250421 +size 726464 diff --git a/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/layout.json b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52ab1cb809e7e265541433867dc5690c9ea99fb2 --- /dev/null +++ b/2024/LayoutFormer_ Hierarchical Text Detection Towards Scene Text Understanding/layout.json @@ -0,0 +1,9555 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 51, + 103, + 543, + 121 + ], + "type": "text", + "content": "LayoutFormer: Hierarchical Text Detection Towards Scene Text Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 143, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 143, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 134, + 143, + 459, + 172 + ], + "type": "text", + "content": "Min Liang, Jia-Wei Ma, Xiaobin Zhu*, Jingyan Qin, Xu-Cheng Yin \nUniversity of Science and Technology Beijing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 174, + 499, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 174, + 499, + 199 + ], + "spans": [ + { + "bbox": [ + 92, + 174, + 499, + 199 + ], + "type": "text", + "content": "{lm20200126, mjw20151001}@@hotmail.com, {zhuxiaobin, xuchengyin}@ustb.edu.cn, qinjingyanking@foxmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 529 + ], + "type": "text", + "content": "Existing scene text detectors generally focus on accurately detecting single-level (i.e., word-level, line-level, or paragraph-level) text entities without exploring the relationships among different levels of text entities. To comprehensively understand scene texts, detecting multi-level texts while exploring their contextual information is critical. To this end, we propose a unified framework (dubbed LayoutFormer) for hierarchical text detection, which simultaneously conducts multi-level text detection and predicts the geometric layouts for promoting scene text understanding. In LayoutFormer, WordDecoder, LineDecoder, and ParaDecoder are proposed to be responsible for word-level text prediction, line-level text prediction, and paragraph-level text prediction, respectively. Meanwhile, WordDecoder and ParaDecoder adaptively learn word-line and line-paragraph relationships, respectively. In addition, we propose a Prior Location Sampler to be used on multi-scale features to adaptively select a few representative foreground features for updating text queries. It can improve hierarchical detection performance while significantly reducing the computational cost. Comprehensive experiments verify that our method achieves state-of-the-art performance on single-level and hierarchical text detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 128, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 571, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 668 + ], + "type": "text", + "content": "Reading and understanding texts in scene images and digital documents is important in various real-world applications, including visual recognition [40], scene understanding [4], and text-based VQA [1, 30]. Compared to words, phrases and sentences express more semantic messages, and paragraphs convey richer contextual semantic messages. For better reading and understanding texts in scenes, it is necessary to simultaneously detect multi-level texts (i.e., word-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 309, + 226, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 309, + 226, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 309, + 226, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 309, + 226, + 545, + 415 + ], + "type": "image", + "image_path": "7530836704e0217b162f3083a33106d8f66ccbf3885507e0d4da6e265955ca93.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 423, + 547, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 423, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 423, + 547, + 478 + ], + "type": "text", + "content": "Figure 1. Top: LayoutFormer is introduced for hierarchical text detection, which simultaneously detects multi-level texts (i.e., word-level, line-level, and paragraph-level) and predicts their geometric layouts. Bottom: Pipeline comparisons of (a) single-level detector, (b) two-level detector, and (c) hierarchical text detector." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 567 + ], + "type": "text", + "content": "level for words, line-level for phrases and sentences, and paragraph-level for paragraphs) and predict their geometric layouts [16, 44], i.e., hierarchical text detection [25]. Hierarchical texts can deliver complete and meaningful text messages, which is beneficial for various downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Although existing scene text detectors have achieved promising performance, they [32, 47] often concentrate on detecting single-level text, which decodes the encoded features of an input image into their representation of independent text entities, e.g., words, lines, or paragraphs. As shown in Fig. 1 (a), single-level detectors only predict one level of text results and do not explore the contextual relationship among different levels of texts. This makes them unable to deliver more informative messages for facilitating scene text understanding. Recently, some methods have investigated two-level text detection. A schematic of the framework is shown in Fig. 1 (b). CUTE [43] proposes a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 712 + ], + "type": "text", + "content": "*Corresponding Author. This work was supported by National Key Research and Development Program of China (2020AAA0109701), National Science Fund for Distinguished Young Scholars (62125601), and National Natural Science Foundation of China (62076024, 62172035)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15665" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 262 + ], + "type": "text", + "content": "two-stage network, which first adopts a Transformer structure like DETR [3] to perform word/line bounding boxes and then directly learns the sequential relationships between detected text boxes based on visual features after ROI operation and positional information. Unified Detector [24] proposes a unified scene text detection and layout analysis network, which first uses a Transformer-based structure [35] to generate word/line mask maps and then models their affiliate relationships. These two methods are all supervised by an affinity matrix or index label, directly modelling relationships on the word-level/line-level detection results. They only rely on the former detection results to model contextual relationships, making the model's performance highly limited by the text detection results at the first stage. Moreover, this paradigm cannot be easily extended to the task of text detection at more than two levels." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 518 + ], + "type": "text", + "content": "Existing scene text detectors generally adopt multi-scale feature maps that contain rich information for accurate prediction. However, methods based on the Transformer encoder-decoder structure often suffer from huge computational and memory costs on high-resolution feature maps. This is because the attention weight computation in the Transformer decoder is of linear computation w.r.t. pixel numbers. There have been several work to tackle the issues. Deformable DETR [51] introduces the multi-scale deformable attention module to select a small set of key locations. AdaMixer [10] adaptively samples features over the space and scales of an object. However, these methods are based on estimated offsets of instance coordinates, not applicable to segmentation methods. In image segmentation, Mask2Former [6] claims that local features can well update query features and proposes to use image features of foreground regions. TSP [31] adopts a RoIAlign operation to extract features of RoIs. However, these methods still suffer from complex operations. We think that it is enough to update query features with only foreground features of a few representative pixels." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we propose a hierarchical text detector named LayoutFormer, which can simultaneously produce hierarchical text detection boxes and their layout relationships. Unlike using inter-level relationships as supervision, we directly adopt multi-level text boxes as supervision, which can significantly reduce the dependence on the former text detection results and implicitly exploit multi-level textual information. In LayoutFormer, we propose three Transformer-based modules, i.e., WordDecoder, LineDecoder, and ParaDecoder, for detecting word-level texts, line-level texts, and paragraph-level texts, respectively. We first adopt text line as the detection unit, and LineDecoder updates and forms line query features. Because a line can be split into several words, WordDecoder learns the deviation of each word within a line based on the line query features and forms final word query features. ParaDecoder aggre" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "gates line query features by learning line-paragraph relationships and forms final paragraph query features. In addition, we propose a Prior Location Sampler, which adaptively selects a few representative features that are highly relevant to scene texts. The sampled features are used for cross-attention operation to update query features. Prior Location Sampler can improve the model performance while reducing the training cost. Extensive experiments verify the superior performance on hierarchical text detection." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 185, + 503, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 503, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 503, + 196 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 201, + 545, + 392 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 201, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 545, + 248 + ], + "type": "text", + "content": "- We propose an innovative hierarchical text detector that simultaneously detects multi-level text instances and explores their contextual information for predicting the geometric layouts, finally performing hierarchical outputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 545, + 308 + ], + "type": "text", + "content": "- We propose WordDecoder, LineDecoder, and ParaDecoder in LayoutFormer to decode word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 308, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 545, + 355 + ], + "type": "text", + "content": "- We propose a Prior Location Sampler to adaptively select a few representative foreground features for updating queries, which can significantly reduce the computational cost while improving the detection performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 356, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 356, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 306, + 356, + 545, + 392 + ], + "type": "text", + "content": "- Extensive experiments on multiple publicly available datasets verify the state-of-the-art performance of our LayoutFormer." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 392, + 427 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 425, + 451 + ], + "type": "text", + "content": "2.1. Scene Text Detection" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "Deep learning-based scene text detection methods have been widely investigated and roughly divided into bottom-up and top-down methods. The bottom-up methods generally detect local components [23, 29, 33] or text pixels [19, 37] for grouping text instances. TextSnake [23] describes a text instance as a sequence of ordered, overlapping disks. The pixel-based methods use some auxiliary information (e.g., similarity vectors in PAN [38], and threshold map in DB [19]) or post-processings (e.g., a progressive scale expansion algorithm in PSENet [37]) to generate text instances. The top-down methods [18, 50] directly predict bounding boxes of scene text instances. To detect arbitrary-shaped texts, some methods [39, 49, 52] localize the key points on the contours of scene texts. Recently, DETR [3] presents a Transformer-based architecture for object detection and achieved great success. Inspired by it, many researchers [24, 32, 47] apply the Transformer structure to scene text detection. Overall, existing scene text detection methods mainly address single-level text detection, ignoring the text detection of multiple levels and further predicting the hierarchical relationships." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15666" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 531, + 271 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 531, + 271 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 531, + 271 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 531, + 271 + ], + "type": "image", + "image_path": "fa45a20519ffc544cc0c89557cd32075adb119f536671e24b81d9b556f7031c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "lines": [ + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed LayoutFormer, which mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders (WordDecoder, LineDecoder, and ParaDecoder) and three predictors. " + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "inline_equation", + "content": "M_l" + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "text", + "content": " denotes predicted mask maps of text lines by " + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "text", + "content": "-th Transformer decoder, and " + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "inline_equation", + "content": "F_l" + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "text", + "content": " denotes the corresponding image features generated by " + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "inline_equation", + "content": "M_l" + }, + { + "bbox": [ + 46, + 282, + 547, + 339 + ], + "type": "text", + "content": ". We first predict text lines, and then predict words and paragraphs. In testing, word-level results produce line-level results based on word-line relationships, and then paragraph-level results are produced based on line-paragraph relationships, finally obtaining hierarchical text detection results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 355, + 272, + 408 + ], + "blocks": [ + { + "bbox": [ + 63, + 355, + 272, + 408 + ], + "lines": [ + { + "bbox": [ + 63, + 355, + 272, + 408 + ], + "spans": [ + { + "bbox": [ + 63, + 355, + 272, + 408 + ], + "type": "table", + "html": "
Detection UnitQueryMemory/GBPQ
word25612.153.09
line25610.657.48
paragraph2568.540.34
", + "image_path": "44d73bab8c299b7d593fc88bf3c73480d904b7b9fc18fa672c11ed04f3b8f9e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 288, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 416, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 288, + 460 + ], + "type": "text", + "content": "Table 1. Comparisons of word, line, and paragraph as text detection unit on the HierText validation set, respectively. \"Query\" denotes the number of text queries. \"Memory\" denotes GPU Memory consumed in training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 480, + 146, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 146, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 146, + 493 + ], + "type": "text", + "content": "2.2. Layout Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": "Document Layout analysis has made great progress, which can be mainly divided into object detection-based methods [17, 27, 34] and segmentation-based methods [2, 16, 41]. Inspired by object detection [15, 28] and semantic segmentation [5, 12, 22], these methods treat semantically coherent text blocks as a special kind of object. However, they fail to produce word or line-level detections and can only be used in companies with standalone text detectors, increasing the complexity of the pipeline. Some work [36] takes a hierarchical view and applies graph-based models on the finest granularity, i.e., individual words, to analyze the layout." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Recently, some work has investigated layout analysis in scene text detection. CUTE [43] detects contextual text blocks consisting of one or multiple ordered integral text units. Their model is a two-stage structure, which first detects integral text units and then models the relationship of integral text units by a designed relation module. Unified Detector [24] introduces a novel task of unified scene text" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 357, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 441 + ], + "type": "text", + "content": "detection and layout analysis, which models the two closely related tasks with a unified model. It produces mask outputs for the text detection task and an affinity matrix for the layout analysis task. The above methods are highly dependent on text detection results at the first stage. Differently, we use multi-level text boxes as supervision to learn the relationship among different levels for layout analysis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 456, + 389, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 389, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 389, + 471 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 478, + 385, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 385, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 385, + 491 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "How to effectively model hierarchical text detection tasks? Is the top-down or bottom-up approach better? As we know, the text layouts in real scene images is highly complex, bringing significant challenges to hierarchical text detection. According to experiments and analyses, we found that paragraph-level text boxes are considerably irregular and always contain much background, which makes existing scene text detection approaches difficult to model text instances. For example, as shown in the second row of Fig. 3, a paragraph text box annotates the entire image. As listed in Tab. 1, the detection performance of paragraphs is much lower than that of words and lines. Thus, the top-down modeling approach is considered inappropriate. Both word and line are generally used as text detection units. However, text lines always contain more complete semantic information than words. The aspect ratios and sizes of words are considerably more variable than text lines, making detection more challenging. Meanwhile, if the word is used" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15667" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 277, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 277, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 277, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 277, + 198 + ], + "type": "image", + "image_path": "78c6e6b056116af6e2270121115adc3b256ccb815d7f857326ba933bde754c8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 287, + 228 + ], + "type": "text", + "content": "Figure 3. Some examples of word-level (first column), line-level (second column), and paragraph-level (third column) text boxes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "type": "text", + "content": "as the detection unit, the detection error will accumulate bigger after two aggregation processes. Hence, we finally choose line-level as our detection unit, generate word-level texts by splitting them downward, and get paragraph-level texts by aggregating them upward." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 315, + 173, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 173, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 173, + 327 + ], + "type": "text", + "content": "3.2. Network Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": "The architecture of our LayoutFormer is illustrated in Fig. 2. It mainly consists of a feature encoder, a Prior Location Sampler, three Transformer decoders and three predictors, respectively, corresponding to word-level text detection, line-level text detection and paragraph-level text detection. Specifically, given an input image " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "I \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", the feature encoder with a ResNet-50 [14] backbone and a Transformer encoder [6] extracts and enhances the features and then generates multi-scale feature maps, i.e., " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "C_4 \\in R^{\\frac{H}{32} \\times \\frac{W}{32} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "C_3 \\in R^{\\frac{H}{16} \\times \\frac{W}{16} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "C_2 \\in R^{\\frac{H}{8} \\times \\frac{W}{8} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "C_1 \\in R^{\\frac{H}{4} \\times \\frac{W}{4} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ". The typical value we use is " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "C = 256" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ". Then, Prior Location Sampler adaptively selects image features for the Transformer decoder. Afterwards, the LineDecoder, WordDecoder, and ParaDecoder update line query features " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{line} \\in R^{N_{line} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", word query features " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{word} \\in R^{N_{word} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", and paragraph query features " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{para} \\in R^{N_{para} \\times C}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", respectively. The three predictors convert " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{word}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "X_{para}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " to line-level prediction " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "(P_{line}, M_{line})" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", word-level prediction " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "(P_{word}, M_{word})" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", and paragraph-level prediction " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "(P_{para}, M_{para})" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "P_* \\in R^{N_* \\times 2}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " is text/non-text probability values of " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "N_*" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " instances and " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "M_* \\in R^{N_* \\times \\frac{H}{4} \\times \\frac{W}{4}}" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " is mask maps of " + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "inline_equation", + "content": "N_*" + }, + { + "bbox": [ + 46, + 333, + 289, + 597 + ], + "type": "text", + "content": " instances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 605, + 181, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 181, + 617 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 181, + 617 + ], + "type": "text", + "content": "3.3. Prior Location Sampler" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 624, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 624, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 624, + 287, + 659 + ], + "type": "text", + "content": "Mask2Former [6] proposes masked attention to attend within the foreground region of the predicted mask for each query, which can be computed as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 669, + 287, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 669, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 77, + 669, + 287, + 682 + ], + "type": "interline_equation", + "content": "X _ {l} = \\operatorname {s o f t m a x} \\left(B _ {l - 1} + Q _ {l} K _ {l} ^ {T}\\right) V _ {l} + X _ {l - 1}, \\tag {1}", + "image_path": "c8fd7faeef921ca71bc7f47c60a011b5e318e5e1f4f0dd841654fe762d92ad88.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the layer index, " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "B_{l-1}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is the binary mask prediction of " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "l-1" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is query features at the " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "-th layer, " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "Q_{l}" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 541, + 190 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 541, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 541, + 190 + ], + "type": "image", + "image_path": "5666002bc55488412c0a002c68d73966bf59fafa8a2098c8fda747d502823a03.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 198, + 546, + 231 + ], + "lines": [ + { + "bbox": [ + 305, + 198, + 546, + 231 + ], + "spans": [ + { + "bbox": [ + 305, + 198, + 546, + 231 + ], + "type": "text", + "content": "Figure 4. Illustration of the Prior Location Sampler, which selects representative foreground features with Top- " + }, + { + "bbox": [ + 305, + 198, + 546, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 198, + 546, + 231 + ], + "type": "text", + "content": " scores in predicted mask maps and then aggregates with weights." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": "is query features " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "X_{l-1}" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " under linear transformation, " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "K_l" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "V_l" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " are the input image features under linear transformation. " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " denotes input query features to the Transformer decoder. " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "B_0" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " is the binary mask prediction obtained from " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "X_0" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": ". The input image features are the entire multi-scale feature maps from a feature encoder. Their computational complexity of Eq. (1) is " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "O(NH'W'C)" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "H'" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "W'" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " are the height and width of a specific feature map respectively, and " + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 246, + 546, + 389 + ], + "type": "text", + "content": " is the number of queries. The cross-attention operation suffers from a linear complexity growth with the spatial size of feature maps. Moreover, the main accompanying problem is the high demand for computing resources." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "Thus, we propose the Prior Location Sampler to select representative foreground features with Top- " + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": " scores in mask maps and obtain the image features " + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "inline_equation", + "content": "F_{l}\\in R^{K\\times N\\times C}" + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "-th decoder layer. Our computational complexity of Eq. (1) is " + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "inline_equation", + "content": "O(NKC)" + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": ". In our implementation, the value " + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "inline_equation", + "content": "K\\ll H^{\\prime}W^{\\prime}" + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": ", and thus the complexity of our cross-attention operation could be significantly reduced. The computational resource consumption due to an increase in the number of feature maps or an increase in the scale of feature maps is negligible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "content": "The architecture of the Prior Location Sampler is illustrated in Fig. 4. Specifically, " + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "inline_equation", + "content": "M_{l}" + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "content": " denotes the predicted mask maps of " + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 510, + 545, + 545 + ], + "type": "text", + "content": "-th decoder layer, and" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 551, + 545, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 551, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 342, + 551, + 545, + 567 + ], + "type": "interline_equation", + "content": "M _ {l} = \\left\\{M _ {l n} \\in R ^ {\\frac {H}{4} \\times \\frac {W}{4}} | n = 1, 2, \\dots , N \\right\\}. \\tag {2}", + "image_path": "86c97053cf9c1f2e251ed87a02c60eae447021915ac6fd0a6ec4ee7ba04491d2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": "For each " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "M_{ln}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ", we use 0.5 for thresholding. We sort scores in " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "M_{ln}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " and select features with Top- " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " scores [13] in multi-scale feature maps " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\{C_i|i = 1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "f_{K}(\\cdot)" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is sampling operation. Then, we assign [20, 28] text ROIs of different scales to the pyramid levels for calculating weights [10] as Eq. (4), where " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "S_{ln}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " is the cumulative sum of the binarized " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "M_{ln}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ", i.e., the area of a candidate text, and 56 [20] is a scaling factor. The selected features for " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "M_{ln}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": " are gathered into " + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "inline_equation", + "content": "F_{ln}\\in R^{K\\times C}" + }, + { + "bbox": [ + 304, + 570, + 547, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 684, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 684, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 367, + 684, + 545, + 715 + ], + "type": "interline_equation", + "content": "F _ {l n} = \\sum_ {i = 1} ^ {4} w _ {l i} f _ {K} \\left(C _ {i}, M _ {l n}\\right), \\tag {3}", + "image_path": "085b3e61bedcc781c2292dbdb8945d170cb633d76e59a6a683228a5c2f75aea9.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15668" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 119, + 90, + 287, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 90, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 119, + 90, + 287, + 119 + ], + "type": "interline_equation", + "content": "t _ {l n} = \\left\\lfloor \\log_ {2} \\frac {S _ {l n}}{5 6} \\right\\rfloor , \\tag {4}", + "image_path": "10cdd21beca2785af58dca7b1fe8013c9b4d272fdb1718c93af13323c97faaee.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 118, + 235, + 138 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 118, + 235, + 138 + ], + "spans": [ + { + "bbox": [ + 99, + 118, + 235, + 138 + ], + "type": "interline_equation", + "content": "w _ {l} = \\text {s o f t m a x} \\left[ - \\frac {\\left(t _ {l n} - \\iota + 1\\right) ^ {2}}{2} \\right].", + "image_path": "b473d29c70398928e68a061fe028c4d111254754b3320be937bd1c09eb376f3a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 144, + 259, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 144, + 259, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 144, + 259, + 157 + ], + "type": "text", + "content": "The final features " + }, + { + "bbox": [ + 61, + 144, + 259, + 157 + ], + "type": "inline_equation", + "content": "F_{l}\\in R^{K\\times N\\times C}" + }, + { + "bbox": [ + 61, + 144, + 259, + 157 + ], + "type": "text", + "content": " are denoted as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 167, + 286, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 167, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 111, + 167, + 286, + 178 + ], + "type": "interline_equation", + "content": "F _ {l} = \\left\\{F _ {l n} \\mid n = 1, 2, \\dots , N \\right\\}. \\tag {5}", + "image_path": "b4cb849f2ddafed1c4e7145a3e8b8efb8b13fdf02a28394cbd9f92328a1c17b9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 185, + 171, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 185, + 171, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 185, + 171, + 196 + ], + "type": "text", + "content": "3.4. Transformer Decoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": "LineDecoder. We feed the selected image features " + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "inline_equation", + "content": "F_{l-1} \\in R^{K \\times N_{line} \\times C}" + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": " learnable line queries into LineDecoder, where " + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "inline_equation", + "content": "K = 32" + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": ". LineDecoder uses the standard Transformer decoder structure, consisting of self-attention, cross-attention, and feedforward network. We also add the positional embeddings to queries and keys at every self-attention and cross-attention layer. We refine line query features layer-by-layer by setting " + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "inline_equation", + "content": "L_1 = 4" + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": ". The updated line query features are adopted to predict " + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 202, + 286, + 334 + ], + "type": "text", + "content": "-th mask maps, which are fed into Prior Location Sampler to generate image features for the next decoder layer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "text", + "content": "WordDecoder. We frist hypothesize that there are " + }, + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "text", + "content": " words in each line and initialize " + }, + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 47, + 335, + 286, + 393 + ], + "type": "text", + "content": " learnable word queries. Thus, the number of words in an image can be formulated as the product of the number of lines in an image and words per line, as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 108, + 403, + 286, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 403, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 108, + 403, + 286, + 416 + ], + "type": "interline_equation", + "content": "N _ {\\text {w o r d}} = N _ {\\text {l i n e}} \\times N _ {\\text {p e r l i n e}}. \\tag {6}", + "image_path": "e512abf07950c8f9f2457df98d2b82e4f6a704fbcd49cdd0d4a701c00d769884.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "text", + "content": "So, we can decouple the self-attention between words into self-attention between lines and self-attention between words per line. Based on line queries from LineDecoder, we only need to compute self-attention between " + }, + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "text", + "content": ", greatly reducing the computational cost. Then we obtain word query features " + }, + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "inline_equation", + "content": "X_{word} \\in R^{N_{word} \\times C}" + }, + { + "bbox": [ + 47, + 422, + 286, + 493 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 501, + 286, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 501, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 114, + 501, + 286, + 516 + ], + "type": "interline_equation", + "content": "X _ {\\text {w o r d}} = X _ {\\text {l i n e}} X _ {\\text {p e r - l i n e}} ^ {T}, \\tag {7}", + "image_path": "9a4d821c34ed09bfe93e47f94914a637a8e730e1ac6b0dc68b37bed64629962d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "X_{line} \\in R^{N_{line} \\times 1 \\times C}" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": " is line query features and " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "X_{per\\_line} \\in R^{N_{per\\_line} \\times 1 \\times C}" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": " is learnable word query features per line. Then cross-attention followed by feedforward network adopts image features " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "F_{l} \\in R^{K \\times N_{line} \\times C}" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "K = 64" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": " to update " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "X_{word}" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": ". The image features are obtained by Prior Location Sampler, which feeds the predicted mask maps of the last decoder layer in LineDecoder. In our experiments, we set " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "L_{3} = 1" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": ". With our word decoupling mechanism, every " + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 47, + 521, + 286, + 641 + ], + "type": "text", + "content": " queries predicts words within a text line for capturing word-line relationships." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "text", + "content": "ParaDecoder. We adaptively learn line-paragraph relationships and aggregate line query features for paragraph-level texts to get paragraph query features. We initialize " + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "N_{para}" + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "text", + "content": " learnable paragraph queries. ParaDecoder also consists of self-attention, cross-attention, and feedforward networks. It uses " + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 47, + 642, + 286, + 712 + ], + "type": "text", + "content": " line query features as inputs of cross attention" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "to update paragraph queries. Positional embeddings are added to queries and keys at every self-attention and cross-attention layer. In our experiments, we adopt " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " of last decoder layer in LineDecoder and set " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "L_{2} = 1" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": ". In this process, the learned attention weights " + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "(R^{N_{para} \\times N_{line}})" + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": " in cross-attention of ParaDecoder are line-paragraph relationships." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 151, + 390, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 151, + 390, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 151, + 390, + 163 + ], + "type": "text", + "content": "3.5. Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 545, + 193 + ], + "type": "text", + "content": "Training. Instead of relation matrices, we use multi-level text boxes as supervision. The total loss is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 203, + 545, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 203, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 365, + 203, + 545, + 216 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {l i n e}} + L _ {\\text {w o r d}} + L _ {\\text {p a r a}}, \\tag {8}", + "image_path": "a084e258ca6835c7a6d982c1ae65d8a085bacad4668f3b1e89dd74ea247a8cef.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "inline_equation", + "content": "L_{line}" + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "inline_equation", + "content": "L_{word}" + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "inline_equation", + "content": "L_{para}" + }, + { + "bbox": [ + 305, + 224, + 545, + 259 + ], + "type": "text", + "content": " respectively denote losses of line-level, word-level, and paragraph-level, which can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 269, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 269, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 324, + 269, + 545, + 297 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {*} = \\lambda_ {c l s} L _ {c l s} \\left(P _ {*}, P _ {*} ^ {\\prime}\\right) + \\lambda_ {m a s k} L _ {m a s k} \\left(M _ {*}, M _ {*} ^ {\\prime}\\right), \\\\ * \\in \\{w o r d, l i n e, p a r a \\}, \\tag {9} \\\\ \\end{array}", + "image_path": "b29f3a0aa68f04eff0826c890433276f4cb3718231fea3e43e1583158427069f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "P_{*}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "P_{*}^{\\prime}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " are the class predictions and their corresponding ground truth, " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "M_{*}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "M_{*}^{\\prime}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " are the mask predictions and their corresponding ground truth, " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " denotes binary cross-entropy loss, " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " is the sum of binary cross-entropy loss and dice loss. We set " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\lambda_{cls} = 2.0" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " for predictions matched with ground truth and 0.1 for unmatched predictions. " + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\lambda_{mask}" + }, + { + "bbox": [ + 305, + 306, + 545, + 389 + ], + "type": "text", + "content": " is set to 5.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 390, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 390, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 390, + 545, + 568 + ], + "type": "text", + "content": "Inference. We present a simple inference procedure that converts class predictions and mask predictions to text boxes, which mainly consists of four steps: (1) the text queries whose class scores are over 0.3 will be selected; (2) the mask predictions of selected queries are binarized by 0.5 to get the binary maps; (3) the connected regions are obtained from the binary maps; (4) the text queries whose averaged foreground mask probability is below 0.9 will be filtered. For hierarchical text detection task, we frist generate the word-level detection boxes by the above inference procedure. Then, we use word-line relationships from WordDecoder to form a word-line geometric layouts and further use line-paragraph relationships from ParaDecoder to form a line-paragraph geometric layouts, finally performing hierarchical text detection outputs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 579, + 386, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 386, + 593 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 386, + 593 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 599, + 368, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 368, + 610 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 368, + 610 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 545, + 677 + ], + "type": "text", + "content": "HierText [24]. It is a hierarchical text detection dataset consisting of 8,281 training images, 1,724 validation images, and 1,634 testing images. It annotates images hierarchically, which first annotates word locations with polygons, then clusters words into lines and lines into paragraphs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "MSRA-TD500 [45]. It is a line-level annotated arbitrary-oriented long text dataset. It consists of 300 training images and 200 testing images collected from natural scenes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15669" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 531, + 142 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 531, + 142 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 531, + 142 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 531, + 142 + ], + "type": "table", + "html": "
MethodTotal-TextHierText-Line
RPFMemory/GBRPFTPQMemory/GB
C2-C4masked attention84.6887.7986.2115.365.5682.7373.1577.5756.7418.1
C2-C4Top-K&concat84.6687.8086.2015.165.3482.5672.9577.3756.4418.3
C1-C4Top-K&concat84.4688.1586.2716.265.8882.9673.4477.5356.9420.3
C1-C4Top-K&weight85.0789.2687.1214.266.6584.1074.3777.2957.4817.0
", + "image_path": "5fdb9239e714b82a272dc22fbdd47512491349e399b169b1ab5799d043d8fc1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 286, + 249 + ], + "type": "text", + "content": "CTW1500 [21]. It is an arbitrary-shaped scene text dataset that consists of 1,000 training images and 500 testing images. In this dataset, the annotations of text instances are line-level and labelled by a polygon with 14 key points." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 249, + 286, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 286, + 285 + ], + "type": "text", + "content": "Total-Text [7]. It is an arbitrary-shaped scene text dataset that contains 1,255 training and 300 testing images, which are annotated by the word-level polygon with 14 key points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 293, + 180, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 180, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 180, + 306 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "content": "In our experiments, for HierText, we only use the training images of HierText to train the models by " + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "inline_equation", + "content": "60k" + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "content": " steps for single-level text detection and " + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "inline_equation", + "content": "80k" + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "content": " steps for multi-level text detection. In testing, we keep the aspect ratio of testing images and resize them to 1,120 height. For MSRA-TD500, CTW1500, and Total-Text, we adopt SynthText [11] to pretrain the models by " + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "inline_equation", + "content": "150k" + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "content": " steps. Then, we finetune the models on the corresponding real-world datasets by " + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "inline_equation", + "content": "40k" + }, + { + "bbox": [ + 46, + 312, + 286, + 430 + ], + "type": "text", + "content": " steps. In testing, we keep the aspect ratio of testing images and resize them to 800 height on three datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "text", + "content": "In training, we use AdamW [26] optimizer and the step learning rate schedule with an initial learning rate of 0.0001 and a weight decay of 0.05. A learning rate multiplier of 0.1 is applied to the ResNet-50 backbone. We decay the learning rate at 0.9 and 0.95 fractions of the total number of training steps by a factor of 10. We train our models with a batch size of 8. Data augmentation includes: (1) Randomly Flipping; (2) Randomly rotate them in range " + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "(-10^{\\circ}" + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "10^{\\circ})" + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "text", + "content": "; (3) Randomly resize them in range (0.5 to 3.0); (4) Randomly cropping. Finally, we resize the images to " + }, + { + "bbox": [ + 46, + 432, + 286, + 551 + ], + "type": "inline_equation", + "content": "640\\times 640" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 559, + 160, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 160, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 160, + 571 + ], + "type": "text", + "content": "4.3. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 286, + 638 + ], + "type": "text", + "content": "For MSRA-TD500, CTW1500, and CTW1500 datasets, we follow the standard evaluation protocol Recall (R), Precision (P), and F-measure (F). For the HierText dataset, we follow [24] and adopt Panoptic Quality (PQ) as the main evaluation metric for the hierarchical text detection task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 147, + 658 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": " value in Prior Location Sampler. We compare the different Top- " + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 665, + 286, + 712 + ], + "type": "text", + "content": " values in Prior Location Sampler. Results are listed in Tab. 3. Selecting too few foreground points may fail to capture and learn the features of an entire text. Select" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 310, + 199, + 543, + 261 + ], + "blocks": [ + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "type": "text", + "content": "Table 2. Ablation comparison of Prior Location Sampler on the Total-Text and line-level of the HierText validation set. \"Memory\" denotes GPU Memory consumed in training. " + }, + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "type": "inline_equation", + "content": "C_1 - C_4" + }, + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "type": "text", + "content": " denotes the multi-scale feature maps used in the Transformer decoder. The results show that \"Top-K&weight\" with " + }, + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "type": "inline_equation", + "content": "C_1 - C_4" + }, + { + "bbox": [ + 46, + 149, + 545, + 183 + ], + "type": "text", + "content": " achieves a considerable improvement and consumes fewer training resources than other methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 199, + 543, + 261 + ], + "lines": [ + { + "bbox": [ + 310, + 199, + 543, + 261 + ], + "spans": [ + { + "bbox": [ + 310, + 199, + 543, + 261 + ], + "type": "table", + "html": "
KTotal-TextHierText-Line
RPFRPFPQ
1684.3488.0186.1465.3481.1272.3855.54
3285.0789.2687.1266.6584.1074.3757.48
6484.7588.6386.6468.6784.4475.7558.86
", + "image_path": "e6092ec11d1023c36658a3c1e3f9c8762227c93482510ba54550d88385ae4bcf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 268, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 306, + 268, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 306, + 268, + 545, + 289 + ], + "type": "text", + "content": "Table 3. Ablation comparison of Top- " + }, + { + "bbox": [ + 306, + 268, + 545, + 289 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 268, + 545, + 289 + ], + "type": "text", + "content": " value on Total-Text and line-level of the HierText validation set." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 309, + 300, + 542, + 369 + ], + "blocks": [ + { + "bbox": [ + 309, + 300, + 542, + 369 + ], + "lines": [ + { + "bbox": [ + 309, + 300, + 542, + 369 + ], + "spans": [ + { + "bbox": [ + 309, + 300, + 542, + 369 + ], + "type": "image", + "image_path": "78a97c23a2766b7a8dde7b68011fdf84741510fb585cb31ff5f4369067463a0a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 376, + 544, + 398 + ], + "lines": [ + { + "bbox": [ + 305, + 376, + 544, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 544, + 398 + ], + "type": "text", + "content": "Figure 5. Visualization of Top- " + }, + { + "bbox": [ + 305, + 376, + 544, + 398 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 376, + 544, + 398 + ], + "type": "text", + "content": " sampling points in Prior Location Sampler." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "content": "ing too many foreground points may introduce noise, which is detrimental to the performance of the model. For Total-Text, we set " + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "content": " to 32, and the same for MSRA-TD500 and CTW1500. While text lines of HierText may have larger aspect ratios and scales, the performance of " + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "inline_equation", + "content": "K = 64" + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "content": " is better than that of " + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "inline_equation", + "content": "K = 32" + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "content": ". However, we finally set " + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 414, + 545, + 498 + ], + "type": "text", + "content": " to 32 for a trade-off between performance and efficiency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "Implementation of Prior Location Sampler. We evaluate the effectiveness of the Prior Location Sampler on the Total-Text and HierText validation set. Results are listed in Tab. 2. \" " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_1 - C_4" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " \" denotes that multi-scale feature maps " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{C_i | i = 1, 2, 3, 4\\}" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " are used in Transformer decoder. \"masked attention\" is followed from Mask2Former [6], which calculates the attention matrix by additionally adding a binary mask prediction. \"Top-K&concat\" applies Top-K strategy to multi-scale feature maps and then performs concatenation. \"Top-K&weight\" applies Top-K strategy to multiscale feature maps and then element-wise sums in weights. Inspired by Mask2Former, the first row of Tab. 2 feeds the entire feature map into the Transformer decoder. The experimental results show that \"Top-K&weight\" apparently achieves a significant improvement in harmonic accuracy and memory. Meanwhile, due to using a small number of foreground points, our method consumes fewer training resources (GPU Memory) than other methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15670" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 71, + 517, + 114 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 517, + 114 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 517, + 114 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 517, + 114 + ], + "type": "table", + "html": "
MethodWordLineParagraph
RPFTPQRPFTPQRPFTPQ
LayoutFormer57.0475.2564.8975.8849.2450.3976.3260.7076.1246.2145.8973.1456.4076.1942.97
", + "image_path": "c38f98e61ec1d65aaad8b4753c179bbf65f7ab11eaa363d364400db5fe2d0fce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 152, + 281, + 268 + ], + "blocks": [ + { + "bbox": [ + 104, + 121, + 488, + 133 + ], + "lines": [ + { + "bbox": [ + 104, + 121, + 488, + 133 + ], + "spans": [ + { + "bbox": [ + 104, + 121, + 488, + 133 + ], + "type": "text", + "content": "Table 4. Experimental results of word-line-paragraph three-level text detection on the HierText testing set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 152, + 281, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 281, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 281, + 268 + ], + "type": "table", + "html": "
MethodNperLineNlineWordLine
PQPQ
In the below experiments, we always set L3 to 1.
Xline1012044.0236.35
Fl1012050.2948.60
In the below experiments, we use Fl as image features.
L3 = 11012050.2948.60
L3 = 41012051.0744.05
L3 = 1612046.2139.60
L3 = 1620050.1246.00
", + "image_path": "108a76d52945c395ad9f4948118e249a50bfcf43bd0b650a862bedc638e55c91.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 321, + 279, + 384 + ], + "blocks": [ + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "type": "text", + "content": "Table 5. Ablation comparison of WordDecoder on the HierText validation set. " + }, + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 46, + 277, + 288, + 310 + ], + "type": "text", + "content": " denote the number of word queries in each line and text line queries, respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 321, + 279, + 384 + ], + "lines": [ + { + "bbox": [ + 56, + 321, + 279, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 279, + 384 + ], + "type": "table", + "html": "
MethodNlineNparaLineParagraph
PQPQ
In the below experiments, we always set L2 to 1.
Xline12812858.3151.98
Fl12812857.2746.15
", + "image_path": "098a04bdafce0b0722de2016f0d27027a9cbe3e2a731f48bf39db870b935c4bd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": "Implementation of WordDecoder. We conduct ablation experiments by comparing different settings of image features, " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " in WordDecoder. Results are listed in Tab. 5. Using " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " as image features for updating word queries is better than " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": ". This is because to identify different text lines, " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " extracts consistent features of each text line, whereas " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "X_{per\\_line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " needs to learn differentiated features within a text line to identify different words within a line. In addition, we find that " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "L_{3}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " being set to 1 performs better than being set to 4. Thus, we adopt " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "L_{3} = 1" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " by default. We also conduct comparative experiments with different " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "N_{per\\_line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 46, + 447, + 287, + 603 + ], + "type": "text", + "content": ". We can conclude that the larger the query number, the higher the performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": "Implementation of ParaDecoder. We compare different image features for ParaDecoder. Results are listed in Tab. 6. Adopting " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " as image features is better than " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": ". We think that after the learning of LineDecoder, " + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "inline_equation", + "content": "X_{line}" + }, + { + "bbox": [ + 46, + 603, + 287, + 662 + ], + "type": "text", + "content": " is able to characterize text lines well." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 274, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 274, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 274, + 683 + ], + "type": "text", + "content": "4.5. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Hierarchical Text Detection. In this section, we evaluate the performance on the HierText testing set for hierarchical" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 152, + 537, + 194 + ], + "blocks": [ + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "type": "text", + "content": "Table 6. Ablation comparison of ParaDecoder on the HierText validation set. " + }, + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "type": "inline_equation", + "content": "N_{para}" + }, + { + "bbox": [ + 46, + 392, + 287, + 426 + ], + "type": "text", + "content": " denote the number of text line queries and paragraph queries, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 152, + 537, + 194 + ], + "lines": [ + { + "bbox": [ + 316, + 152, + 537, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 537, + 194 + ], + "type": "table", + "html": "
MethodNwordNlineWordLine
PQPQ
LayoutFormer120012050.3548.62
", + "image_path": "17613fa338d6ddb5a1e3a716e06eb6613ded099256db7bdeaee9b37041fdf23f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": "Table 7. Experimental results of word-line two-level text detection on the HierText testing set. " + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "inline_equation", + "content": "N_{word}" + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 305, + 202, + 545, + 237 + ], + "type": "text", + "content": " denote the number of word queries and text line queries, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 247, + 543, + 425 + ], + "blocks": [ + { + "bbox": [ + 310, + 247, + 543, + 425 + ], + "lines": [ + { + "bbox": [ + 310, + 247, + 543, + 425 + ], + "spans": [ + { + "bbox": [ + 310, + 247, + 543, + 425 + ], + "type": "image", + "image_path": "1fa2316c2eec527dfe017818d4a7cdec7c32eb241944e717dded40c382d86115.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 433, + 545, + 477 + ], + "lines": [ + { + "bbox": [ + 305, + 433, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 305, + 433, + 545, + 477 + ], + "type": "text", + "content": "Figure 6. The hierarchical text detection examples on the HierText. The visualization images in each row are in turn the input image, word-level detection results, line-level detection results, and paragraph-level detection results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 581 + ], + "type": "text", + "content": "text detection. To our knowledge, few work has conducted research on hierarchical text detection. Hence, we mainly compare with [24]. Results are listed in Tab. 7, Tab. 9, and Tab. 4. Meanwhile, we demonstrate hierarchical detection results on images in Fig. 6. Both qualitative and quantitative verify the advances of our method. And we can observe that our method can precisely detect hierarchical texts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": "For word-line detection, our results are listed in Tab. 7. Because there is no any prior research on this, we provide a baseline to facilitate comparisons with other methods. Benefiting from our decoupling operation on words, our " + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "inline_equation", + "content": "N_{word}" + }, + { + "bbox": [ + 304, + 582, + 545, + 641 + ], + "type": "text", + "content": " can reach 1,200, much higher than Unified Detector [24]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "For line-paragraph detection, our LayoutFormer achieves the best text detection and layout analysis performance. Results are listed in Tab. 9. \"GCP API\", \"GCN-PP\", \"Mask-RCNN-Cluster\", \"Max-DeepLab-Cluster\" are two-stage approaches, while \"Unified Detector\" is an end-to-end unified approach. Our LayoutFormer with" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15671" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 523, + 265 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 523, + 265 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 523, + 265 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 523, + 265 + ], + "type": "table", + "html": "
MethodVenueLine DetectionWord Detection
MSRA-TD500CTW1500Total-Text
RPFRPFRPF
TextSnake [23]ECCV1873.983.278.367.985.375.674.582.778.4
MSR [42]IJCAI1976.787.481.777.883.880.773.085.278.6
LOMO [48]CVPR19---76.585.780.879.387.683.3
PAN [38]ICCV1983.884.484.181.286.483.781.089.385.0
ContourNet [39]CVPR20---84.183.783.983.986.985.4
DB [19]AAAI2079.291.584.980.286.983.482.587.184.7
TextFuseNet [46]IJCAI20---85.085.885.483.287.585.3
FCENet [53]CVPR21---83.487.685.582.589.385.8
BPNet [49]ICCV2180.6885.4082.9781.4587.8184.5184.6590.2787.37
PCR [8]CVPR2183.590.887.082.387.284.782.088.585.2
I3CL [9]IJCV22---84.587.485.983.789.286.3
FSG [32]CVPR2284.891.688.182.488.185.285.790.788.1
Unified Detector [24]CVPR2387.4488.0487.7087.4484.5685.9791.0684.9687.90
DPText-DETR [47]AAAI23---86.291.788.886.491.889.0
LayoutFormer-88.3291.9590.1084.2688.1686.1785.0789.2687.12
", + "image_path": "7849654e5281b3634de9f98c1443ff423ebadf01101ef5a786642da728b0bfde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 316, + 287, + 437 + ], + "blocks": [ + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "type": "text", + "content": "Table 8. Experimental results of single-level text detection. We set " + }, + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 274, + 547, + 298 + ], + "type": "text", + "content": " to 100 on three datasets. On MSRA-TD500 and CTW1500, our LayoutFormer achieves the state-of-the-art performance. On Total-Text, our LayoutFormer achieves competitive results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 316, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 49, + 316, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 287, + 437 + ], + "type": "table", + "html": "
MethodNlineNparaLineParagraph
PQPQ
GCP API [24]--56.1746.33
GCN-PP [24]384-62.2350.10
Mask-RCNN-Cluster [24]384-62.2351.67
Max-DeepLab-Cluster [24]384-62.2352.52
128-58.7651.48
Unified Detector [24]256--52.50
384-62.2353.60
LayoutFormer38420062.3753.76
", + "image_path": "293cbef44fd980e45ee96ee3aad49c54b18a0f590732c0a02676ecc1387fc212.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "text", + "content": "Table 9. Experimental results of line-paragraph two-level text detection on the HierText testing set. " + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "inline_equation", + "content": "N_{line}" + }, + { + "bbox": [ + 46, + 445, + 288, + 479 + ], + "type": "text", + "content": " denotes the number of text line queries." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "N_{line} = 384" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " outperforms Unified Detector [24] with " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "N_{line} = 384" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "0.14\\%" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "inline_equation", + "content": "0.16\\%" + }, + { + "bbox": [ + 47, + 504, + 287, + 540 + ], + "type": "text", + "content": " in terms of PQ of text detection and layout analysis, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": "For word-line-paragraph detection, there is no any prior research on three-level text detection and layout analysis. We only list our performance on Tab. 4, which can be used by subsequent methods for comparison. Our Layout-Former achieves " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "49.24\\%" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "46.21\\%" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "42.97\\%" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " in terms of PQ of word-level, line-level, and paragraph-level, respectively. The inference speed of our method is 3.0 FPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 628, + 287, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 287, + 698 + ], + "type": "text", + "content": "Single-level Text Detection. In this section, we evaluate the performance of our model on the most widely used benchmarks for single-level scene text detection, i.e., MSRAD500 and CTW1500 for line-level, Total-Text for word-level. Results are listed in Tab. 8. The quantitative results demonstrate the advances of our model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 712 + ], + "type": "text", + "content": "For line detection, we achieve the state-of-the-art results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": "on MSRA-TD500. Notably, our LayoutFormer achieves " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "88.32\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "91.95\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "90.10\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": " in terms of Recall, Precision, and F-measure, respectively, which significantly outperforms other methods with a great margin. For example, LayoutFormer outperforms PCR [8], FSG [32], and Unified Detector [24] by " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "3.10\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "2.00\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "2.40\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": " in terms of F-measure, respectively. On CTW1500, LayoutFormer achieves " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "84.26\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "88.16\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "inline_equation", + "content": "86.17\\%" + }, + { + "bbox": [ + 304, + 317, + 545, + 437 + ], + "type": "text", + "content": " in terms of Recall, Precision, and F-measure, respectively, achieving the promising performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 535 + ], + "type": "text", + "content": "For word detection, we achieve competitive results on Total-Text. The performance of our LayoutFormer is slightly lower than FSG [32] and Unified Detector [24], in which the former utilizes more training epochs and the latter utilizes larger training datasets and more training epochs. DPText-DETR [47] pre-trains their models on more datasets and these datasets are specifically for improving the performance of arbitrary-shape texts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "spans": [ + { + "bbox": [ + 306, + 548, + 379, + 561 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose a hierarchical text detector for promoting scene text understanding, which simultaneously detects multi-level texts and predicts their geometric layouts. In LayoutFormer, we propose WordDecoder, LineDecoder, and ParaDecoder to detect word-level text, line-level text, and paragraph-level text, respectively. WordDecoder and ParaDecoder adaptively learn word-line relationships and line-paragraph relationships, respectively. Through this research, we hope to provide a baseline for hierarchical text detection and inspire other hierarchical tasks. In addition, we will continue to promote and improve the hierarchical architecture for scene text understanding." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15672" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Ali Furkan Biten, Rubén Tito, Andrés Mafla, Lluis Gómez i Bigorda, Marçal Rusinol, C. V. Jawahar, Ernest Valveny, and Dimosthenis Karatzas. Scene text visual question answering. In ICCV, pages 4290-4300, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 169 + ], + "type": "text", + "content": "[2] Samuele Capobianco, Leonardo Scommegna, and Simone Marinai. Historical handwritten document segmentation by using a weighted loss. In ANNPR, pages 395-406, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 287, + 213 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 287, + 258 + ], + "type": "text", + "content": "[4] Praneeth Chakravarthula, Jim Aldon D'Souza, Ethan Tseng, Joe Bartusek, and Felix Heide. Seeing with sound: Long-range acoustic beamforming for multimodal scene understanding. In CVPR, pages 982-991, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 287, + 315 + ], + "type": "text", + "content": "[5] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell., 40(4):834-848, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 360 + ], + "type": "text", + "content": "[6] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1280–1289, 2022. 2, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 362, + 286, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 286, + 394 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 286, + 394 + ], + "type": "text", + "content": "[7] Chee-Kheng Ch'ng, Chee Seng Chan, and Cheng-Lin Liu. Total-text: toward orientation robustness in scene text detection. IJDAR, 23(1):31-52, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 287, + 429 + ], + "type": "text", + "content": "[8] Pengwen Dai, Sanyi Zhang, Hua Zhang, and Xiaochun Cao. Progressive contour regression for arbitrary-shape scene text detection. In CVPR, pages 7393-7402, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 430, + 287, + 473 + ], + "type": "text", + "content": "[9] Bo Du, Jian Ye, Jing Zhang, Juhua Liu, and Dacheng Tao. I3CL: intra- and inter-instance collaborative learning for arbitrary-shaped scene text detection. IJCV, 130(8):1961-1977, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 475, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 287, + 508 + ], + "type": "text", + "content": "[10] Ziteng Gao, Limin Wang, Bing Han, and Sheng Guo. Adamixer: A fast-converging query-based object detector. In CVPR, pages 5354-5363, 2022. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 542 + ], + "type": "text", + "content": "[11] Ankush Gupta, Andrea Vedaldi, and Andrew Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 586 + ], + "type": "text", + "content": "[12] Fei He, Haoyang Zhang, Naiyu Gao, Jian Jia, Yanhu Shan, Xin Zhao, and Kaiqi Huang. Inspro: Propagating instance query and proposal for online video instance segmentation. 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[13] Junjie He, Pengyu Li, Yifeng Geng, and Xuansong Xie. Fastinst: A simple query-based model for real-time instance segmentation. In CVPR, pages 23663-23672, 2023. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In ICCV, pages 2980-2988, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[16] Joonho Lee, Hideaki Hayashi, Wataru Ohyama, and Seiichi Uchida. Page segmentation using a convolutional neural net" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "work with trainable co-occurrence features. In ICDAR, pages 1023-1028, 2019. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 149 + ], + "type": "text", + "content": "[17] Kai Li, Curtis Wigington, Chris Tensmeyer, Handong Zhao, Nikolaos Barmpalios, Vlad I. Morariu, Varun Manjunatha, Tong Sun, and Yun Fu. Cross-domain document object detection: Benchmark suite and method. In CVPR, pages 12912-12921, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 185 + ], + "type": "text", + "content": "[18] Minghui Liao, Baoguang Shi, Xiang Bai, Xinggang Wang, and Wenyu Liu. Textboxes: A fast text detector with a single deep neural network. In AAAI, pages 4161-4167, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 218 + ], + "type": "text", + "content": "[19] Minghui Liao, Zhaoyi Wan, Cong Yao, Kai Chen, and Xiang Bai. Real-time scene text detection with differentiable binarization. In AAAI, pages 11474-11481, 2020. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 220, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 261 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Piotr Dólar, Ross B. Girshick, Kaiming He, Bharath Hariharan, and Serge J. Belongie. Feature pyramid networks for object detection. In CVPR, pages 936-944, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 297 + ], + "type": "text", + "content": "[21] Yuliang Liu, Lianwen Jin, Shuai tao Zhang, Canjie Luo, and Sheng Zhang. Curved scene text detection via transverse and longitudinal sequence connection. PR, 90:337-345, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 331 + ], + "type": "text", + "content": "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, pages 3431-3440, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 376 + ], + "type": "text", + "content": "[23] Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He, Wenhao Wu, and Cong Yao. Textsnake: A flexible representation for detecting text of arbitrary shapes. In ECCV, pages 19-35, 2018. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "text", + "content": "[24] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. Towards end-to-end unified scene text detection and layout analysis. In CVPR, pages 1039-1049, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 465 + ], + "type": "text", + "content": "[25] Shangbang Long, Siyang Qin, Dmitry Panteleev, Alessandro Bissacco, Yasuhisa Fujii, and Michalis Raptis. ICDAR 2023 competition on hierarchical text detection and recognition. In ICDAR, pages 483-497, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 487 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 489, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 542 + ], + "type": "text", + "content": "[27] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascadetabnet: An approach for end to end table detection and structure recognition from image-based documents. In CVPR, pages 2439-2447, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 578 + ], + "type": "text", + "content": "[28] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. pages 91-99, 2015. 3, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "text", + "content": "[29] Baoguang Shi, Xiang Bai, and Serge J. Belongie. Detecting oriented text in natural images by linking segments. In CVPR, pages 3482-3490, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 655 + ], + "type": "text", + "content": "[30] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards VQA models that can read. In CVPR, pages 8317-8326, 2019. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 689 + ], + "type": "text", + "content": "[31] Zhiqing Sun, Shengcao Cao, Yiming Yang, and Kris Kitani. Rethinking transformer-based set prediction for object detection. In ICCV, pages 3591-3600, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[32] Jingqun Tang, Wenqing Zhang, Hongye Liu, Mingkun Yang, Bo Jiang, Guanglong Hu, and Xiang Bai. Few could be bet-" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15673" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "ter than all: Feature sampling and grouping for scene text detection. In CVPR, pages 4553-4562, 2022. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[33] Zhi Tian, Weilin Huang, Tong He, Pan He, and Yu Qiao. Detecting text in natural image with connectionist text proposal network. In ECCV, pages 56-72, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 132, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 132, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 132, + 287, + 163 + ], + "type": "text", + "content": "[34] Matheus Palhares Viana and Dario Augusto Borges Oliveira. Fast cnn-based document layout analysis. In ICCV, pages 1173-1180, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 166, + 287, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 166, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 48, + 166, + 287, + 209 + ], + "type": "text", + "content": "[35] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan L. Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In CVPR, pages 5463-5474, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 212, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 212, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 287, + 244 + ], + "type": "text", + "content": "[36] Renshen Wang, Yasuhisa Fujii, and Ashok C. Popat. Postocr paragraph recognition by graph convolutional networks. In WACV, pages 2533-2542, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 289 + ], + "type": "text", + "content": "[37] Wenhai Wang, Enze Xie, Xiang Li, Wenbo Hou, Tong Lu, Gang Yu, and Shuai Shao. Shape robust text detection with progressive scale expansion network. In CVPR, pages 9336-9345, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 292, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 287, + 336 + ], + "type": "text", + "content": "[38] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network. In ICCV, pages 8439-8448, 2019. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 380 + ], + "type": "text", + "content": "[39] Yuxin Wang, Hongtao Xie, Zheng-Jun Zha, Mengting Xing, Zilong Fu, and Yongdong Zhang. Contournet: Taking a further step toward accurate arbitrary-shaped scene text detection. In CVPR, pages 11750-11759, 2020. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 383, + 287, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 427 + ], + "type": "text", + "content": "[40] Guanghui Xu, Shuaicheng Niu, Mingkui Tan, Yucheng Luo, Qing Du, and Qi Wu. Towards accurate text-based image captioning with content diversity exploration. In CVPR, pages 12637-12646, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 429, + 287, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 429, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 429, + 287, + 472 + ], + "type": "text", + "content": "[41] Yue Xu, Fei Yin, Zhaoxiang Zhang, and Cheng-Lin Liu. Multi-task layout analysis for historical handwritten documents using fully convolutional networks. In IJCAI, pages 1057-1063, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 287, + 506 + ], + "type": "text", + "content": "[42] Chuhui Xue, Shijian Lu, and Wei Zhang. MSR: multi-scale shape regression for scene text detection. In IJCAI, pages 989–995, 2019. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 552 + ], + "type": "text", + "content": "[43] Chuhui Xue, Jiaxing Huang, Wenqing Zhang, Shijian Lu, Changhu Wang, and Song Bai. Contextual text block detection towards scene text understanding. In ECCV, pages 374-391, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 554, + 287, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 287, + 598 + ], + "type": "text", + "content": "[44] Xiao Yang, Ersin Yumer, Paul Arente, Mike Kraley, Daniel Kifer, and C. Lee Giles. Learning to extract semantic structure from documents using multimodal fully convolutional neural networks. In CVPR, pages 4342-4351, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 287, + 632 + ], + "type": "text", + "content": "[45] Cong Yao, Xiang Bai, Wenyu Liu, Yi Ma, and Zhuowen Tu. Detecting texts of arbitrary orientations in natural images. In CVPR, pages 1083-1090, 2012. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[46] Jian Ye, Zhe Chen, Juhua Liu, and Bo Du. Textfusenet: Scene text detection with richer fused features. In IJCAI, pages 516-522, 2020. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[47] Maoyuan Ye, Jing Zhang, Shanshan Zhao, Juhua Liu, Bo Du, and Dacheng Tao. Dpext-detr: Towards better scene text detection with dynamic points in transformer. In AAAI, 2023. 1, 2, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 329 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[48] Chengquan Zhang, Borong Liang, Zuming Huang, Mengyi En, Junyu Han, Errui Ding, and Xinghao Ding. Look more than once: An accurate detector for text of arbitrary shapes. In CVPR, pages 10552-10561, 2019. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[49] Shi-Xue Zhang, Xiaobin Zhu, Chun Yang, Hongfa Wang, and Xu-Cheng Yin. Adaptive boundary proposal network for arbitrary shape text detection. In ICCV, pages 1285-1294, 2021. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[50] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. EAST: an efficient and accurate scene text detector. In CVPR, pages 2642-2651, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "text", + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ICLR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[52] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 329 + ], + "type": "text", + "content": "[53] Yiqin Zhu, Jianyong Chen, Lingyu Liang, Zhanghui Kuang, Lianwen Jin, and Wayne Zhang. Fourier contour embedding for arbitrary-shaped text detection. In CVPR, pages 3123-3131, 2021. 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15674" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_content_list.json b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d00ecfff5671658380908ca64b13d26ee5428985 --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_content_list.json @@ -0,0 +1,1662 @@ +[ + { + "type": "text", + "text": "LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding", + "text_level": 1, + "bbox": [ + 112, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuwei Luo $^{1*}$ , Yufan Shen $^{12*}$ , Zhaoqing Zhu $^{1*}$ , Qi Zheng $^{1}$ , Zhi Yu $^{2}$ , Cong Yao $^{1}$ , Alibaba Group, Zhejiang University", + "bbox": [ + 166, + 202, + 797, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{luochuwei,zzhaoqing.z,zhengqisjtu,yaocong2010}@gmail.com", + "bbox": [ + 232, + 241, + 736, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{syficy,yuzhirenzhe}@zju.edu.cn", + "bbox": [ + 346, + 258, + 620, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, leveraging large language models (LLMs) or multimodal large language models (MLLMs) for document understanding has been proven very promising. However, previous works that employ LLMs/MLLMs for document understanding have not fully explored and utilized the document layout information, which is vital for precise document understanding. In this paper, we propose LayoutLLM, an LLM/MLLM based method for document understanding. The core of LayoutLLM is a layout instruction tuning strategy, which is specially designed to enhance the comprehension and utilization of document layouts. The proposed layout instruction tuning strategy consists of two components: Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. To capture the characteristics of document layout in Layout-aware Pre-training, three groups of pretraining tasks, corresponding to document-level, region-level and segment-level information, are introduced. Furthermore, a novel module called layout chain-of-thought (LayoutCoT) is devised to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers. LayoutCoT is effective for boosting the performance of document understanding. Meanwhile, it brings a certain degree of interpretability, which could facilitate manual inspection and correction. Experiments on standard benchmarks show that the proposed LayoutLLM significantly outperforms existing methods that adopt open-source 7B LLMs/MLLMs for document understanding.", + "bbox": [ + 76, + 340, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 758, + 209, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Document AI [7], including its document understanding tasks such as document VQA [33, 47] and document visual information extraction [18, 19, 37], is currently a hot topic in both academia and industry. In recent years, document pre-trained models [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 54, 55, 59] have achieved excellent performance in doc", + "bbox": [ + 76, + 784, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6023a6bc2d8f1d903a3ed2d3637c8de18a88b928f5b5c37b27166e39f321bec8.jpg", + "image_caption": [ + "(1)LLMs" + ], + "image_footnote": [], + "bbox": [ + 504, + 308, + 617, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9950b96af1868905b0333d6c074d8b39ed0765fb45e4606715ef7ffe55314596.jpg", + "image_caption": [ + "(II)MLMs" + ], + "image_footnote": [], + "bbox": [ + 627, + 309, + 746, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/930f4118ac7aca512809d2f20097a1a27dfa7de897d6c4277681788fb22f12bd.jpg", + "image_caption": [ + "(III) LayoutLLM (Ours)" + ], + "image_footnote": [], + "bbox": [ + 756, + 309, + 870, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e0a4723496339970e0d87b8bfbf5942039012f5a0ffced855ac97b6c87e36d0.jpg", + "image_caption": [ + "(b) Training paradigms" + ], + "image_footnote": [], + "bbox": [ + 506, + 419, + 651, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a748fae26f0c2f517931304e7719f8fe618cbcf011533e653ad0141c8e1870d6.jpg", + "image_caption": [ + "(a) Model architectures" + ], + "image_footnote": [], + "bbox": [ + 653, + 420, + 771, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d03aba8ef4cf57671481d86956f799803e8d1c835f732fd97f88171320e3dc7.jpg", + "image_caption": [ + "Figure 1. LLMs/MLLMs for document understanding. The LayoutLLM is an LLM/MLLM based method that integrates a document pre-trained model as encoder. It is trained by the newly proposed layout instruction tuning strategy which consists of Layout-aware Pre-training and Layout-aware Supervised Fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 772, + 420, + 867, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ument AI downstream tasks. However, due to the necessity for fine-tuning on corresponding downstream task data, it is challenging to directly adapt such pre-trained models for zero-shot document understanding. In this paper, zero-shot refers to not using training sets of downstream tasks.", + "bbox": [ + 496, + 580, + 890, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large language models (LLMs) such as ChatGPT [35] and LLaMA [49, 50], or multimodal large language models (MLLMs) like GPT-4V [1, 36, 56], have shown remarkable zero-shot capabilities across various applications. For Document AI, as shown in Fig. 1 (a), (I) directly prompting LLMs with document text [15, 39] and (II) training document-based MLLMs [3, 57, 60] have also achieved promising results under the zero-shot setting [3, 39, 57, 60].", + "bbox": [ + 496, + 657, + 890, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It is widely accepted that document layout information is vital for document understanding [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. However, it is difficult to convey document layout information by directly feeding text to LLMs. As Fig. 1(a)(I) shows, representing documents as either flattened plain text or layout text such as text with coordinates [15, 39, 44, 64] is often used for LLMs.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15630", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Flattened plain text completely excludes any layout information of the document [54]. Additionally, as Tab. 1 shows, using layout text that represents both textual and layout information as inputs for LLMs does not guarantee LLMs can effectively comprehend this formatted text.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, existing works that employ MLLMs for document understanding also have not fully explored the document layout information. Document-based MLLMs integrate visual models [11, 42] with LLM [48-50, 61] for document understanding. As Fig. 1(b) shows, they are typically based on pre-training and supervised fine-tuning (SFT) on document datasets. In the pre-training stage, tasks such as image captioning [29, 57, 60] or generating all text in a document as flattened plain text [9, 27, 60] are commonly applied. Both these image captions [43, 46] and plain text only provide a brief representation and fail to capture the layout information of the document. So it is difficult for the model to learn document layout in the existing pre-training stage. In the SFT stage, document-related VQA or information extraction data [3, 57] is often used. The answers are directly provided during SFT, lacking explicit learning about document layout. In summary, current approaches using plain or layout text to prompt LLMs and training document-based MLLMs have not effectively captured layout information, limiting their zero-shot document understanding capability. Therefore, for better document understanding with the power of LLMs, it is necessary to investigate how to effectively incorporate layout information into LLMs.", + "bbox": [ + 75, + 169, + 472, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose LayoutLLM, an LLM/MLLM based method for document understanding, in which a layout instruction tuning strategy is designed to enhance the comprehension of document layouts. Different from the existing MLLMs that use a general visual pre-trained model [11, 42] as the encoder, we integrate document pretrained models [2, 12, 16, 17, 23, 25, 32, 52, 54, 55] as the encoder in order to better leverage the model's foundational understanding capability for documents. The proposed layout instruction tuning consists of two stages: layout-aware pre-training and layout-aware supervised fine-tuning (SFT). Due to the complex nature of documents in real-world scenarios, encompassing rich textual content and diverse layout structures, achieving a thorough understanding involves not only capturing the document's fundamental content at global but also delving into local details. In the layout-aware pre-training stage, to ensure the model learns not only the global information of documents but also detailed information at different hierarchical levels, three groups of different level pre-training tasks are proposed: document-level, region-level, and segment-level. All the proposed pretraining tasks are unified in the format of instruction tuning.", + "bbox": [ + 75, + 518, + 472, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, in the layout-aware SFT stage, to enhance the model's comprehension and utilization of layout information for question answering, a novel strategy called", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LayoutCoT is proposed, motivated by the chain-of-thought (CoT) [21, 53] ability in LLMs. Unlike existing methods that are directly supervised by the answer to the document understanding question, LayoutCoT consists of three successive steps: Question Analysis, Relevant Area Concentration, and Answer Formation. Through these steps, the model gains a deeper understanding of the questions, becomes capable of focusing the relevant areas instead of searching answers in the entire document and can leverage the specific characteristics of identified areas (such as tables, paragraphs, etc.) to accurately infer the answers. It not only brings a certain degree of interpretability, but also provides a feasible way for manual intervention or correction of model results. Extensive zero-shot experiments on five widely-used document understanding benchmarks demonstrate the effectiveness of the proposed LayoutLLM.", + "bbox": [ + 496, + 90, + 893, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 517, + 333, + 821, + 345 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) To better learn document layouts from global to local in layout-aware pre-training, three groups of different level pre-training tasks, which are all implemented through instruction tuning, are proposed.", + "2) A novel LayoutCoT strategy is proposed to achieve layout-aware supervised fine-tuning. It enables Layout-LLM to focus on the relevant document area and leverage the region's features to generate accurate answers, exhibiting a certain degree of interpretability.", + "3) Experimental results on zero-shot document understanding tasks show that the proposed LayoutLLM significantly outperforms existing methods that adopt LLMs/MLLMs for document understanding, demonstrating the great potential of document layout modeling." + ], + "bbox": [ + 491, + 349, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 573, + 650, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-trained models for document understanding. Document pre-trained models have demonstrated the effectiveness of layout information in document understanding [2, 5, 8, 10, 12, 13, 16, 17, 20, 22, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. As a pioneer, LayoutLM [54] is the first to encode spatial coordinates of text for layout representation learning in pre-training documents. The following works [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 55, 59] then joint text, layout and images in document pre-training by combining visual models as document image encoders with the text and layout transformers, and various works [5, 10, 20, 22] start to explore pre-training end-to-end models for document understanding. These studies have achieved significant advancements in document understanding by exploring various model architectures [2, 8, 10, 12, 13, 20, 25, 26, 38, 41, 52, 59] and attention mechanisms [16, 17, 55] for modeling layout information. These methods also have proposed layout-related pre-training tasks that have been demonstrated to be highly effective in document understanding tasks. For instance,", + "bbox": [ + 496, + 598, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15631", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tasks like masked vision-language modeling [13, 17, 55], where the model is required to generate the original text corresponding to the randomly masked text in the document; position masking [31, 51], involving the randomly position masking and subsequent recovery of position information in the document; geometric pre-training [26, 32], focusing on learning direction, distance, etc.; and layout-aware generation tasks [5, 22], aiming to make the model generate structured text with layout information. However, due to the necessity of fine-tuning with annotated data for downstream tasks, these efforts face challenges in extending to zero-shot document understanding.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs/MLLMs for document understanding. Recently, LLMs like ChatGPT [35] and MLLMs like GPT-4 [36, 56] have demonstrated remarkable zero-shot performance across a wide range of NLP/CV tasks. Leveraging LLMs/MLLMs for zero-shot document understanding has also shown promising progress [3, 30, 39, 45, 57, 60]. Perot et al. [39] explore the use of LLMs for document visual information extraction, emphasizing the importance of the document layout. LLaVAR [60] which extends LLaVA [28, 29] to the document domain is pre-trained by generating plain text in the document image. During SFT, it is trained by document-related instructions which are generated by GPT-4. Expanding on mPLUG-Owl [58], mPLUG-DocOwl [57] is trained using publicly available datasets for document understanding. It includes tasks like document-level image captioning, direct information extraction, and direct document VQA. Moreover, Qwen-VL [3] proposes a general MLLM that performs well on document understanding tasks, utilizing document-level pre-training and direct VQA for SFT. Though existing LLMs/MLLMs have shown promising results in document understanding, their limited focus on document layout in pre-training and SFT has hindered their ability to achieve higher accuracy in zero-shot document understanding and better interpretability.", + "bbox": [ + 75, + 272, + 472, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LayoutLLM", + "text_level": 1, + "bbox": [ + 76, + 647, + 207, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LayoutLLM is an LLM/MLLM based method that incorporates document pre-trained models for document understanding. To enhance the document layout comprehension in LayoutLLM, a novel layout instruction tuning strategy is proposed, which consists of two stages: layout-aware pretraining and layout-aware supervised fine-tuning (SFT).", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 76, + 771, + 264, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of LayoutLLM is shown in Fig. 2. In LayoutLLM, given an input document image and its corresponding text and layout information, the document pre-trained model encoder is required to obtain the multimodal document features. Then, these features are encoded by multimodal projectors, and together with the instruction embeddings, fed into the LLM to generate the final results.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38e33a72c3a2e8f530d4ef31b412f7eb52b03c6a13c3fc7a3cc5dc0dff274d67.jpg", + "image_caption": [ + "Figure 2. Overall architecture of LayoutLLM." + ], + "image_footnote": [], + "bbox": [ + 553, + 89, + 841, + 231 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Document pre-trained model encoder. To leverage the foundational document comprehension capability of document pre-trained models, in this work, we utilize LayoutLMv3 [17], a widely-used document pre-training model, as our basic document encoder. The document image, text, and layout are initially inputted into the document pretrained model (DocPTM). They are then encoded by the DocPTM to obtain the corresponding features as follows:", + "bbox": [ + 498, + 255, + 890, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {V}, F _ {T} = \\operatorname {D o c P T M} (V, T, B o x) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 391, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $V$ represents the document image, $T = T_{0:n - 1}$ and $Box = Box_{0:n - 1}$ indicate the text sequences in the document and their corresponding bounding-box coordinates respectively. After being encoded by the DocPTM, the visual features of the document $F_{V} = F_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{0}}$ and the text layout features $F_{T} = F_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{0}}$ are acquired. $m$ signifies the number of visual features and $n$ represents the number of tokens contained in the document. $d_0$ denotes the dimension of DocPTM feature space.", + "bbox": [ + 498, + 411, + 890, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal projectors. To project multi-modality features from DocPTM into the LLM's embedding space, inspired by the simple yet effective projector design in LLaVA [28, 29], two different Multilayer Perceptrons (MLPs) are used as visual projector and text & layout projector respectively. Formally, the projected features can be obtained by:", + "bbox": [ + 498, + 547, + 890, + 637 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH _ {V} = P _ {V} \\left(F _ {V}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 643, + 890, + 660 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH _ {T} = P _ {T} \\left(F _ {T}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 664, + 890, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H_{V} = H_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{1}}$ is the feature encoded by the visual projector, $H_{T} = H_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{1}}$ is the feature encoded by the text & layout projector, and $d_{1}$ is the dimensional of the LLM embedding features.", + "bbox": [ + 496, + 688, + 890, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large language model. Finally, the $H_V$ , $H_T$ and the embedding of the question's instruction text, $I_q = I_{q_0:l_q - 1}$ , are inputted together into the LLM, generating the target answer $I_a = I_{a_0:l_a - 1}$ . $l_q$ and $l_a$ represent the length of the question's instruction text and the answer text, respectively.", + "bbox": [ + 498, + 750, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Layout Instruction Tuning", + "text_level": 1, + "bbox": [ + 500, + 832, + 741, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LayoutLLM model is trained using the layout instruction tuning, which consists of two stages: layout-aware pretraining and layout-aware SFT.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15632", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13b8337221be70f40d2a4e8849fcd0cd0072db60f067a6e54a1f07b7d6790542.jpg", + "image_caption": [ + "Figure 3. Overview of the Layout Instruction Tuning. (a) Document-level, region-level, and segment-level pre-training tasks, unified in instruction tuning format, are introduced. (b) A novel module called LayoutCoT is designed to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers through three intermediate steps." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 883, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Layout-aware Pre-training", + "text_level": 1, + "bbox": [ + 76, + 383, + 318, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of pre-training the LayoutLLM is to enhance the model's comprehensive understanding of documents at different levels through layout learning, rather than only focusing on global document-level understanding like existing MLLM methods [57, 60]. To this end, during the pre-training stage, three different level pre-training strategies are simultaneously applied to the LayoutLLM, namely document-level, region-level, and segment-level.", + "bbox": [ + 75, + 412, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Document-level To enable the model to possess fundamental global document understanding, pre-training tasks, namely Document Dense Description (DDD) and Text and Layout Reconstruction (TLR), are proposed. As Fig. 3 (a) shows, like the image caption task, the DDD task requires the model to learn to describe the input document. Moreover, in the DDD task, the descriptions for document images are more detailed. For instance, in the document image caption data used for the LLaVAR [60] pre-training, captions contain an average of 36.27 words, while in the proposed dataset of the DDD task, the description contains an average of 373.25 words. Through the DDD task, the model can obtain basic document-level information, such as document type and detailed content. The TLR task aims to reconstruct the complete text and layout information of a document and output it in the format “<\\{box\\}, \\{text\\} >”. The TLR task aligns the text and layout embeddings output from DocPTM with the LayoutLLM's LLM space. Consequently, it enables the LLM in LayoutLLM to comprehend the text and layout information contained in the documents.", + "bbox": [ + 75, + 536, + 468, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region-level The information contained in specific regions of a document, such as the titles, figures, tables, is essential for document understanding [4, 6, 8, 24, 62]. These regions serve as important characteristics that differentiate", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a document from plain text in natural language. For the LayoutLLM to achieve the basic region-level understanding, two pre-training tasks, namely Document Layout Analysis (DLA) and Table Understanding (TU), are utilized. The DLA task is achieved in two ways as shown in Fig. 3. One involves locating the layout region based on the layout type, the other involves identifying the type of a given area. Furthermore, the table region differs from other regions in that it requires additional focus on 2D layout understanding. The TU task enables the model to understand the basic row and column information in the table region of a document. As shown in Fig. 3, the TU task includes instruction tuning for the number of rows and columns, logical coordinates, and the content within rows and columns.", + "bbox": [ + 496, + 383, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Segment-level Early works on document pre-trained models [13, 17, 26, 31, 32, 51, 55] have demonstrated the effectiveness of segment-level document pre-training tasks to document layout understanding ability, such as masked vision language modeling (MVLM) [13, 17, 55], position masking [31, 51], and geometric pre-training [26, 32]. Inspired by these works, to make LayoutLLM have segment-level layout understanding, these tasks are transformed into instruction formats for pre-training as Fig. 3(a) shows. For the MVLM instructions, random masking of the text input to LayoutLLM is performed, and the model is instruction tuned by asking the masked words and answering them. For mask position instruction, the layout information (coordinates) to a specific text line, when input to LayoutLLM, is randomly set to 0. The instruction is constructed by asking about the text line with zeroed coordinates and requesting the model to respond with the original coordinates with text content. For geometric layout instruction, text lines are randomly selected, and an instruction is constructed by asking questions about the direction and distance between them.", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15633", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Layout-aware Supervised Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 90, + 395, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the SFT stage of existing document-based MLLMs, models are directly supervised by the answer to the document understanding instructions. Consequently, these methods lack explicit learning of document layout which is crucial for document understanding. Considering this limitation, and inspired by previous works related to chain-of-though (CoT) [21, 53], which have shown that inferencing with intermediate steps can greatly enhance performance. A novel module called LayoutCoT is proposed, which incorporates the layout information into every intermediate step of CoT explicitly. Meanwhile, by introducing the layout-aware intermediate steps, the answer process gains a certain degree of interpretability for LayoutLLM and also provides interactive correction possibilities based on LayoutCoT.", + "bbox": [ + 75, + 114, + 468, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LayoutCoT Details. As shown in Fig. 3(b), the LayoutCoT involves the following three intermediate steps:", + "bbox": [ + 76, + 325, + 470, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1: Question Analysis. To effectively address a document understanding problem, analyzing the key characteristics of the question is very important. Identifying the question type, such as table understanding or entity extraction from paragraphs, and assessing whether the question is a straightforward extraction query or a more complex reasoning problem, can help guide the direction for the subsequent inference process. Therefore, to give basic guidance to the subsequent steps, the question analysis step is designed, encompassing an analysis of the question type from a layout perspective and a detailed understanding of the question itself. Benefiting from the layout understanding ability by layout-aware pre-training, this step can extract the types and key information mentioned in the question, which are related to the specific characteristics of the document.", + "bbox": [ + 75, + 357, + 468, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2: Relevant Area Concentration. For most document understanding tasks, the entire document contains a large amount of irrelevant information that may confuse the model [5]. This step aims to focus on the relevant area and generate its location information, which is used to assist the model to accurately infer the answer. Benefiting from the layout information conveyed by step 1 and the positioning capabilities learned from the region&segment-level pretraining, the model can accurately generate the location of the relevant area. For example in Fig. 3(b), according to the question type \"table\" in step 1, the relevant \"table\" can be located. By guiding the model to focus on the relevant area, this step largely narrows the search scope, increasing the possibility of giving the right answer. Meanwhile, the location information provides a way for visual inspection and interactive correction (see Sec. 4.7 for details).", + "bbox": [ + 75, + 584, + 468, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 3: Answer Formation. Finally, the last step, the answer formation, provides explanations based on the layout characteristics of the relevant areas located in step 2 and key points analyzed in step 1 to get the final answer. For example in Fig. 3(b), for a \"table\" type question, this step in-", + "bbox": [ + 75, + 825, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 CONSTRUCT(D): LayoutCoT Construction.", + "text_level": 1, + "bbox": [ + 504, + 87, + 888, + 102 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Definition: $\\mathcal{H}$ : Document HTMLs; $\\mathcal{I}$ : Document Images; $\\mathcal{T}$ : MRC Texts; $\\mathcal{R}$ : Document Language Representation; QA: QA pairs; $T_{c}$ : Text CoT; $\\mathcal{L}_{c}$ : LayoutCoT;", + "Input: Document Dataset $\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\}$", + "$(\\mathcal{D}_H = \\{\\mathcal{H}\\}, \\mathcal{D}_I = \\{\\mathcal{I}\\}, \\mathcal{D}_M = \\{\\mathcal{T}, \\mathcal{Q}\\mathcal{A}\\})$ ;", + "Output: Constructed Dataset $\\mathcal{D}_c$", + "1: Procedure CONSTRUCT $(\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\})$", + "2: 1) $\\mathcal{R} = getDocRep(\\mathcal{I})$ if $\\mathcal{D} \\subseteq \\mathcal{D}_I$ else $\\mathcal{H}$ if $\\mathcal{D} \\subseteq \\mathcal{D}_H$ else pass;", + "3: 2) $\\mathcal{Q}\\mathcal{A},\\mathcal{T}_c = getQACoT(\\mathcal{D})$ if $\\mathcal{D}\\subseteq \\mathcal{D}_M$ else getQACoTGPT(R);", + "4: 3) $\\mathcal{L}_c = getLayoutCoT(\\mathcal{T}_c)$", + "5: 4) if $\\mathcal{D} \\subseteq \\{\\mathcal{D}_H, \\mathcal{D}_M\\}$ : $\\mathcal{I} = Htm12Img(\\mathcal{D}_H \\cap \\mathcal{H} : \\mathcal{T})$ ;", + "6: return $\\mathcal{D}_c\\gets \\{\\mathcal{I},\\mathcal{Q}\\mathcal{A},\\mathcal{L}_c\\}$" + ], + "bbox": [ + 508, + 104, + 888, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "volves analyzing the row and column in the relevant table in step 2, and referencing the answer step-by-step. For a \"key-value\" question, analyzing the keywords in concentrated areas can help get the final answer. Analyzing answers in different ways based on the features of different layout regions not only improves the document understanding performance but also brings a certain level of interpretability.", + "bbox": [ + 496, + 275, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LayoutCoT Construction. Given the need for both text and image annotations in constructing LayoutCoT, manual labeling can be difficult. Algorithm 1 proposes a manual-labeling-free method, generating LayoutCoT data using public datasets with GPT (GPT-3.5 Turbo). It involves representing document text and layout in a format understandable by GPT. GPT is then utilized to generate document-content-based QA and corresponding text CoT. Finally, use rules for transforming the text CoT to LayoutCoT.", + "bbox": [ + 496, + 382, + 890, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Three types of publicly available document datasets are focused on: HTML documents $(\\mathcal{D}_H)$ , image documents $(\\mathcal{D}_I)$ , and text documents $(\\mathcal{D}_M)$ for machine reading comprehension (MRC). The construction process is as follows:", + "bbox": [ + 496, + 518, + 890, + 580 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Document Representation: To fully leverage the capabilities of GPT, it is crucial to ensure that the document content fed to GPT contains accurate layout information. Since HTML is the formatted language that can represent documents accurately, $\\mathcal{D}_H$ is represented using the original HTML. By transforming HTML to PDF and using the PDF parser, the text and bounding-boxes are obtained. For $\\mathcal{D}_I$ , the layout-aware text [15] is used. The text and bounding-boxes in $\\mathcal{D}_I$ are from the original dataset annotations.", + "2) QA&Text CoT Generation: The language representation $\\mathcal{R}$ for the document is employed for prompting GPT to generate QA pairs $\\mathcal{QA}$ with text CoTs $\\mathcal{T}_c$ . In addition, the $\\mathcal{D}_M$ includes the QA pairs and reasoning process, thereby directly reusing the $\\mathcal{QA}$ and manually organizing $\\mathcal{T}_c$ . The generated $\\mathcal{T}_c$ includes the step 1 (question analysis) and step 3 (answer formation) for LayoutCoT, and locates all relevant sentences in the document for $\\mathcal{QA}$ .", + "3) LayoutCoT Generation: The step 1 & 3 in $\\mathcal{T}_c$ are used as the step 1 & 3 in $\\mathcal{L}_c$ . To construct the step 2 (relevant area concentration) of $\\mathcal{L}_c$ , the union bounding-box of all located relevant sentences in $\\mathcal{T}_c$ are taken as the relevant area." + ], + "bbox": [ + 496, + 580, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15634", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4) Document Images Generation: For $\\mathcal{D}_H$ and $\\mathcal{D}_M$ , the HTMLs and MRC text are converted to images. Overall, the document images $\\mathcal{I}$ , generated QA pairs $\\mathcal{QA}$ and LayoutCoTs $\\mathcal{L}_c$ constitute the final LayoutCoT dataset $\\mathcal{D}_c$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 169, + 209, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset Collection", + "text_level": 1, + "bbox": [ + 76, + 195, + 253, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layout-aware pre-training data of LayoutLLM is from publicly available document understanding datasets. It does not incorporate any data from the training, validation, and test sets of downstream benchmarks. Region-level pretraining tasks, most document-level and segment-level tasks are self-supervised. Thus, only document images and images converted from PDFs in the original datasets, along with the corresponding OCR or text layout results from PDF parsing, are needed. For these tasks, data is randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65]. Particularly, data for document dense description is from inputting the document text content into GPT-3.5 Turbo, prompting it to generate an average of 373.25 words document dense descriptions. For the region-level tasks, specifically the document layout analysis task, publicly available document layout analysis datasets are utilized, including PubLayNet [62], DocLayNet [40], and Docbank [24]. Data for another region-level task, table understanding, is sourced from PubTabNet [63] with its table annotations. All data is transformed into the instruction format illustrated in Fig. 3(a). In total, 5.7 million instructions are constructed, with a ratio of 1:4:4 for document-level, region-level, and segment-level tasks, respectively. For detailed instruction templates and dataset descriptions, please refer to the supplementary material.", + "bbox": [ + 75, + 219, + 472, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layout-aware SFT data of LayoutLLM is generated by GPT (GPT-3.5 Turbo) and converted from existing textual Machine Reading Comprehension (MRC) datasets, as discussed in Sec. 3.2.2. To generate high-quality document-based textual QA and textual CoT, it is essential to make GPT comprehend the document layout. So, the document is represented using both layout text [15] and HTML. Similar to the pre-training data, the $D_I$ in Algorithm 1 is also randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65] for building layout text. The $D_H$ in Algorithm 1 is from GPT's free generation. The $D_M$ in Algorithm 1 is randomly sampled from the FeTaQA [34] which is a wikipedia question answering dataset. A total of 300K instructions are constructed, with a ratio of 5:4.5:0.5 for $D_I$ , $D_H$ , and $D_M$ , respectively. For detailed prompt templates of document-based text QA and text CoT generation using GPT, prompts for HTML generation using GPT, and dataset description, please refer to the supplementary material.", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training Setup", + "text_level": 1, + "bbox": [ + 498, + 90, + 653, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The encoder weight of LayoutLLM is initialized from the LayoutLMv3-large [17] which is a widely-used document pre-trained model. And the LLM backbone is initialized from Vicuna-7B-v1.5 [61]. Other parameters are randomly initialized. During pre-training, the LLM is frozen, and the parameters of the two projectors and document pre-trained model encoder are updated. During SFT, both LLM and two projectors are fine-tuned while keeping the document pretrained model encoder frozen. For detailed training setup, please refer to the supplementary material.", + "bbox": [ + 496, + 112, + 893, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation Setup", + "text_level": 1, + "bbox": [ + 498, + 272, + 669, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The zero-shot ability is highly expected in real-world document understanding scenarios [7, 30, 45]. Therefore, zero-shot evaluations on widely-used document understanding benchmarks including document visual question answering (Document VQA) and visual information extraction (VIE) are conducted. Only the test sets are utilized in all benchmarks and only the official provided image, text, and layout information are used. The Document VQA datasets comprise the DocVQA[33] test set, consisting of 5,188 questions, and the VisualMRC[47] test set containing 6,708 questions. Following the evaluation metric settings of the original datasets, the ANLS [33] is utilized for evaluating DocVQA, and Rouge-L is used for evaluating VisualMRC. For the VIE task, FUNSD [19], CORD [37], and SROIE [18] are used. The test set of FUNSD comprises 50 form images, each annotated with entity-level headers, questions, answers, and others, along with entity linking annotations. CORD's test set consists of 100 receipt images, annotated with 30 entity types, such as the tax amount, total price, etc. SROIE's test set includes 347 receipt images, annotated with 4 entity types: company, date, address, and total. To prompt LLMs/MLLMs for zero-shot VIE, annotations in VIE datasets are transformed into question answering format (QA for VIE). For key-value annotations with linking in the FUNSD, the format is {Q: What is the \"key\" in the document? A: \"value)}. For entity annotations in CORD and SROIE, directly asking for the entity in the document like {Q: What is the address in the document? A: \"the address annotation\"} is utilized. Following DocVQA, the QA for VIE task is evaluated by ANLS.", + "bbox": [ + 496, + 295, + 893, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Main Results", + "text_level": 1, + "bbox": [ + 498, + 756, + 638, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1, the zero-shot document understanding performance of LayoutLLM and existing open-source LLMs and MLLMs is evaluated. Generally, the existing LLMs are better than MLLMs for zero-shot document VQA and VIE. For example in the results on DocVQA, most LLMs can achieve a performance of around $60\\%$ or higher, while most MLLMs can only attain around $10\\%$ , except for mPLUG-DocOWL and Qwen-VL that trained with the", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15635", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e82125f67fba9f385d2e11fec999c61386d341ac2f72e66fc462411316d44554.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsDocument VQAQA for VIE
DocVQAVisualMRCFUNSDCORDSROIE
Fine-tuned PTMLayoutLMv3 [17]83.37*-92.08*‡97.46*‡-
LLMLlama2-7B [50]61.3429.7340.784.3915.86
Llama2-7B-chat [50]64.9952.8448.2047.7068.97
Plain TextVicuna-7B [61]61.3953.6349.7944.6767.49
Vicuna-1.5-7B [61]66.9952.1348.0651.4068.20
LLMLlama2-7B [50]37.3233.8251.4028.0434.96
Llama2-7B-chat [50]56.5549.2658.3450.9351.15
Layout Text (Text + Box) [15]Vicuna-7B [61]37.2152.5542.7346.5945.43
Vicuna-1.5-7B [61]56.8147.2259.6356.1366.20
MLLMLLaVAR-7B [60]11.6†36.371.7113.552.38
LLaVA-1.5-7B [28]13.3435.231.9318.063.83
mPLUG-DocOWL-7B [57]62.2*†----
Qwen-VL-7B [3]65.1*†42.5247.0930.0058.59
LayoutLLM-7B△ (Ours)74.2555.7378.6562.2170.97
LayoutLLM-7B★ (Ours)74.2755.7679.9863.1072.12
", + "bbox": [ + 96, + 78, + 870, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training set. One possible reason is that it's difficult for these MLLMs to obtain accurate textual information from document images. Additionally, for LLMs, using Plain Text and Layout Text respectively as the document representation are further discussed, where the Layout Text introduces layout information by adding text coordinates in the format: {text:\"text\", box:[x1,y1,x2,y2]} [15]. Compared to the Plain Text, the Layout Text variant doesn't show stable performance improvements, noticeable in certain tasks, for example in Vicuna-1.5, an improvement in VIE (FUNSD $48.06\\%$ to $59.63\\%$ ) but a decline in DocVQA $(66.99\\%$ to $56.81\\%)$ . LLMs may lack the ability to learn this formatted layout text, and directly adding layout information (e.g., coordinates) to the text will also highly increase the token length, making the answer inference more challenging.", + "bbox": [ + 75, + 383, + 472, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to the prior SOTA model, LayoutLMv3, which is fine-tuned using the training set of downstream tasks, LayoutLLM demonstrates competitive performance on the DocVQA benchmark. Compared with these LLMs and MLLMs, LayoutLLM achieves consistent and significant improvements over them on all evaluation benchmarks. Notably, LayoutLLM which employs zero-shot performance, outperforms mPLUG-DocOWL and Qwen-VL by around $10\\%$ on the DocVQA dataset, both of which are trained with this dataset. This demonstrates that LayoutLLM can learn more robust and discriminative representations for document understanding. Furthermore, experiments of the different initialization of the LLM backbone have all achieved optimal results across all benchmarks, substantiating that LayoutLLM can adapt to various LLMs. In summary, our method explores a more effective way to utilize layout information for document understanding, which significantly improves the performance of zero-shot document understanding.", + "bbox": [ + 75, + 612, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5737a2582d79468e1080f86076b20ef23e3ac7d1528229e840545e6057d4619.jpg", + "table_caption": [ + "Table 1. Zero-shot document understanding results on open-source LLMs and MLLMs. * signifies training set use; unmarked results are zero-shot. Results marked with ${}^{ \\ddagger }$ are F1 scores for VIE. Results marked with $\\dagger$ are from the original paper and others are re-implemented by us. $\\bigtriangleup$ marks LayoutLLM's LLM backbone as initialized with Llama2-7B-chat,and $\\star$ with Vicuna-1.5-7B." + ], + "table_footnote": [], + "table_body": "
#Layout-aware \nPre-trainingLayout-aware \nSFTDocVQAFUNSD
070.8270.96
1✓72.3174.02
2✓✓74.2779.98
", + "bbox": [ + 503, + 376, + 888, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on the DocVQA and FUNSD test sets.", + "bbox": [ + 506, + 446, + 880, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 465, + 650, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To better verify the effectiveness of the layout-aware pretraining and layout-aware SFT in the layout instruction tuning, an ablation study is conducted (see Tab. 2).", + "bbox": [ + 496, + 489, + 890, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Initial baseline. The #0 baseline disables both layout-aware pre-training and layout-aware SFT. It only adopts SFT (same SFT data but without LayoutCoT steps) on the LayoutLLM. Even without any alignment pre-training and LayoutCoT steps guidance, the baseline outperforms previous SOTAs, achieving $70.82\\%$ on DocVQA and $70.96\\%$ on FUNSD. This indicates the document understanding ability of DocPTM benefits document understanding with LLMs. Effect of Layout-aware pre-training. In #1, the layout-aware pre-training and the same SFT with #0 is conducted. Compared to #0, benefiting from the basic document understanding capability learned through the layout-aware pretraining, #1 shows an increase of $1.49\\%$ on DocVQA and $3.06\\%$ on FUNSD. It can be observed that the basic document understanding ability learned from the layout-aware pre-training significantly enhances the performance of the basic key-value extraction tasks in the FUNSD dataset. Compared to the FUNSD VIE task, it also shows that the DocVQA is a more complex task. Effect of Layout-aware SFT. Compared with #1, #2 further incorporates the layout-aware SFT strategy, resulting in a performance gain of $1.96\\%$ on DocVQA and even $5.96\\%$ on FUNSD. This indicates the LayoutCoT in layout-aware SFT can help the LayoutLLM to handle complex document tasks and it is effic", + "bbox": [ + 496, + 537, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15636", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c085931ced6d4bf149854cf8647c6ef4f5d296c89fbdf02912e98b6fdafc124.jpg", + "image_caption": [ + "Question: What is the Income Tax Withheld for the month of July 1954?" + ], + "image_footnote": [], + "bbox": [ + 138, + 77, + 259, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce4b7d257b1d72b26e1b07fd7edbcb9f66bf7af5fda50855d64dcf90090b89b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 78, + 408, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b336250946d452584eaa12662d0d52a3dd13d5966a47a782dda79845f69f9589.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 160, + 256, + 188 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/097a6af6bd5cb4f86bd9ad43ac2415512d63639b7e8a071980726536d774504e.jpg", + "image_caption": [ + "Question: What is the Total for Exit 2004?" + ], + "image_footnote": [], + "bbox": [ + 140, + 191, + 258, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90fd5a4718cf92ef8920d9ede940eea486e3a53c844d7738b74cdbf619e76b94.jpg", + "image_caption": [ + "Figure 4. Qualitative results on DocVQA. Green boxes are the areas concentrated in the step 2 of LayoutCoT." + ], + "image_footnote": [], + "bbox": [ + 138, + 268, + 256, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72b69f819c37d11880391fc1b4164a638f805e8c984d634efc4c474e23d6d3f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 258, + 292, + 269, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b107afefe8faf6e0623f542bbfff496f92c4926ed48bfd822b1f78a84927f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 188, + 408, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/683193296cabd548f6372c9f5491b9a70051cf08d79d1ecaaf247a5ad245bb74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 290, + 356, + 301 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tive for boosting the performance on both document VQA and VIE. Moreover, it also promotes a certain degree of interpretability. Overall, the ablation study demonstrates the importance of layout-aware pre-training and layout-aware SFT for zero-shot document understanding.", + "bbox": [ + 75, + 337, + 468, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 76, + 422, + 261, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Two examples are shown in Fig. 4. Through combined with layout-aware pre-training and layout-aware SFT, Layout-LLM can accurately focus on the relevant areas, utilize the layout information to assist in problem-solving and provide interpretability. For example, in Fig. 4(a) question about key-value extraction in up-down layout, different from the left-right variant, relies more on document layout to infer the right answer. Since the keywords \"Income Tax\" in the question often co-occur with numerical data, Vicuna-1.5 and Qwen-VL find numerical answers relying more on the semantics than the layout, resulting in incorrect responses. In contrast, benefiting from the layout-aware pre-training, our model can effectively leverage layout information to give accurate answers. In addition, the model using LayoutCoT can further provide the location and the reasoning process, showing a certain degree of interpretability. But in certain situations, only combined with layout pre-training, our model might fail to give accurate answers. As shown in Fig. 4(b), without LayoutCoT, our model identifies \"Exit 2003\" as the relevant column and generates a wrong answer. However, with the help of LayoutCoT, LayoutLLM can correctly identify the question type as \"Table\", locate the relevant table area, and finally infer the right answer from the corresponding \"Exit 2004\" column.", + "bbox": [ + 75, + 445, + 472, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Interactive Correction with LayoutCoT", + "text_level": 1, + "bbox": [ + 76, + 816, + 416, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since LayoutCoT runs in a step-by-step fashion and produces intermediate results in the inference stage, it can facilitate interactive inspection and correction, when processing a document. As shown in Fig. 5(a), there are two areas", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8641010ee6173b2ae131e22ffec8f9ff88ea4049185be5ac26e93b051f28a303.jpg", + "image_caption": [ + "Question: In the table given on the left bottom of", + "Step 1 (Question Analysis): The question is about key-value extraction. And the question asks for the 'vendor'.", + "Step 2 (Relevant Area Concentration): The answer can be found", + "in the area: [28, 123, 317, 148].", + "Step 3 (Answer Formation): The number entered for 'vendor' in the text is:", + "Step 1 (Question Analysis):", + "Step 2 (Relevant Area Concentration): The answer can be found in the text: [46,703,467,830].", + "Step 3 (Answer Formation): The number entered for 'vendor' in", + "the table is 013684", + "Final answer: 013684." + ], + "image_footnote": [], + "bbox": [ + 547, + 84, + 674, + 157 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 549, + 178, + 674, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question: What is the heading of second paragraph?", + "bbox": [ + 549, + 255, + 674, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/671c37c21e3f401282dd5824acdf3d36cac071b217411f43cd15690c94a9bf77.jpg", + "image_caption": [ + "Table 1 Table 2" + ], + "image_footnote": [], + "bbox": [ + 754, + 121, + 764, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fe3bb4ef182a5215bc43bae56bf3f48b87db9051de41450fc76c96cf3f7f0ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 121, + 776, + 126 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01e428f05a981fa67b4f30d0c169b7b5a6b2b29167a7010ff79388567c85968f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 128, + 764, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d9ad74267586c88775852aa6428a6c2d49a534a5a18b72175f0b83117d0c48b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 128, + 776, + 132 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a39ea0d7bcfd074e9a0c8e0da7fe9e03ffd767ac2a6e6a69682e06d31f397a3.jpg", + "image_caption": [ + "the page, what is it?", + "GT Answer: 013684", + "GT Answer: PRELIMINARY STATEMENT", + "GT Answer: 013684", + "Step 1 (Question Analysis): The question is about Factoid. And the question asks for the heading of the second paragraph... Step 2 (Relevant Area Concentration): The answer can be found in the area: 134, 280, 654, 3731.", + "Step 3 (Answer Formation): The heading of the second", + "paragraph is \\*TAKEDA'S RESPONSE AND OBJECTIONS TO", + "PRACTITIONER'S SUPPLEMENTAL INJURY CLAIMS. ADDITIONAL ACTOS PERSONAL INJURY CLAIMS*.", + "Final answer: TAXEDA'S RESPONSE AND OBJECTIONS TO", + "□", + "Step 1 (Question Analysis):", + "Step 2 (Relevant Area Concentration): The answer can be found", + "in the area: [117, 566, 678, 681].", + "Step 3 (Answer Formation): The heading of the second paragraph is \"PRESUMMARY STATEMENT\".", + "Final answer: PRELIMINARY STATEMENT", + "tion with LayoutCoT. Green represents and answers, while Red represents the list viewed in digital version." + ], + "image_footnote": [], + "bbox": [ + 549, + 176, + 674, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Question: What is the heading of second paragraph?", + "bbox": [ + 549, + 255, + 674, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GT Answer: PRELIMINARY STATEMENT", + "bbox": [ + 549, + 263, + 637, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55eab25f39acc79a21497930688dbd25324e6abce888c915d9c89a42c6238ec4.jpg", + "image_caption": [ + "PLAINTIPS SUPPLEMENTAL INTERROGATORY 2 REGRADING ADDITIONAL ACTS PERONEAL INJURY CLAIMS" + ], + "image_footnote": [], + "bbox": [ + 790, + 233, + 797, + 237 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in the image that are relevant to the keyword \"vendor\" of the question. LayoutCoT focused on a wrong area containing \"vendor\", as it missed \"left bottom\" in the question, the answer was incorrect. However, after the right area is given manually, it can finally give the correct answer. Similarly, in Fig. 5(b) the question asks about \"the heading of the second paragraph\". However, the term \"paragraph\" does not have a universal definition, and in this case, the sentences in the area below the main heading were considered as a paragraph, causing the model to predict \"TAKEDA's...CLAIMS\", which was incorrect according to the GT. Once the right \"second paragraph\" region is fed to the model, the answer can be successfully revised. This unique ability of LayoutCoT could be very valuable in high-stake scenarios (e.g., a bank transaction), where the standards are extremely high, and manual checking and correction (i.e., human-in-the-loop) are required.", + "bbox": [ + 496, + 332, + 892, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 606, + 622, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through LayoutCoT, LayoutLLM demonstrates the capability of interactive correction, but in real-world applications, this is not enough. The ability to refuse false-positive outputs and generate hints (e.g. \"The answer is not mentioned in the document.\") is crucial. However, it is currently absent in LayoutLLM. In addition, despite achieving notable improvements through layout-aware pre-training, LayoutLLM struggles in precisely understanding region-level relationships, as evidenced in Fig. 5(a). We will study how to endow LayoutLLM with such abilities.", + "bbox": [ + 496, + 631, + 890, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 799, + 617, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose LayoutLLM for document understanding, in which a layout instruction tuning strategy comprising layout-aware pre-training and layout-aware SFT is designed to improve the comprehension of document layouts. Extensive experiments confirm the effectiveness of LayoutLLM.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15637", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] GPT-4V(ision) system card. 2023. 1", + "[2] Srikar Appalaraju, Bhavan Jasani, and Bhargava Urala Kota. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 4171-4186, 2021. 1, 2", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3, 7", + "[4] Galal M Binmakhashen and Sabri A Mahmoud. Document layout analysis: a comprehensive survey. ACM Computing Surveys (CSUR), 52(6):1-36, 2019. 4", + "[5] Haoyu Cao, Changcun Bao, Chaohu Liu, Huang Chen, Kun Yin, Hao Liu, Yinsong Liu, Deqiang Jiang, and Xing Sun. Attention where it matters: Rethinking visual document understanding with selective region concentration. In ICCV, pages 19517-19527, 2023. 2, 3, 5", + "[6] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 4", + "[7] Lei Cui, Yiheng Xu, Tengchao Lv, and Furu Wei. Document ai: Benchmarks, models and applications. arXiv preprint arXiv:2111.08609, 2021. 1, 6", + "[8] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 1, 2, 4", + "[9] Cheng Da, Peng Wang, and Cong Yao. Multi-granularity prediction with learnable fusion for scene text recognition. arXiv preprint arXiv:2307.13244, 2023. 2", + "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In ECCV, pages 280–296. Springer, 2022. 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[12] Jieuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Nikolaos Barmpalios, Rajiv Jain, Ani Nenkova, and Tong Sun. Unified pretraining framework for document understanding. In NeurIPS, 2021. 1, 2", + "[13] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xlayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In CVPR, pages 4583-4592, 2022. 1, 2, 3, 4", + "[14] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In ICDAR, pages 991-995. IEEE, 2015. 6" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. ICCV, 2023. 1, 5, 6, 7", + "[16] Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents. In AAAI, 2022. 1, 2", + "[17] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM Multimedia, 2022. 1, 2, 3, 4, 6, 7", + "[18] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 competition on scanned receipt OCR and information extraction. In ICDAR. IEEE, 2019. 1, 6", + "[19] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents, 2019. 1, 6", + "[20] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In ECCV, pages 498- 517.Springer,2022.2", + "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. NeurIPS, 35:22199-22213, 2022. 2, 5", + "[22] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In ICML, pages 18893–18912. PMLR, 2023. 2, 3", + "[23] Chenliang Li, Bin Bi, and Ming Yan. StructuralLM: Structural pre-training for form understanding. In ACL, 2021. 1, 2", + "[24] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 4, 6", + "[25] Peizhao Li, Jiuxiang Gu, and Jason Kuen. SelfDoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 1, 2", + "[26] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. Structext: Structured text understanding with multimodal transformers. In ACM Multimedia, pages 1912-1920, 2021. 1, 2, 3, 4", + "[27] Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE transactions on pattern analysis and machine intelligence, 45(1):919-931, 2022. 2", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 7", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15638", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Yang Liu, Biao Yang, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 3, 6", + "[31] Chuwei Luo, Guozhi Tang, Qi Zheng, Cong Yao, Lianwen Jin, Chenliang Li, Yang Xue, and Luo Si. Bi-vldoc: Bidirectional vision-language modeling for visually-rich document understanding. arXiv preprint arXiv:2206.13155, 2022. 3, 4", + "[32] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. Geolayoutm: Geometric pre-training for visual information extraction. In CVPR, pages 7092-7101, 2023. 1, 2, 3, 4", + "[33] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 1, 6", + "[34] Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Hailey Schoelkopf, Riley Kong, Xiangru Tang, et al. Fetaqa: Freeform table question answering. TACL, 10:35-49, 2022. 6", + "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1, 3", + "[36] R OpenAI. Gpt-4 technical report. arXiv, pages 2303-08774, 2023. 1, 3", + "[37] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. {CORD}: A consolidated receipt dataset for post-{ocr} parsing. In Workshop on Document Intelligence at NeurIPS 2019, 2019. 1, 6", + "[38] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. In EMNLP findings, 2022. 1, 2", + "[39] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Jiaqi Mu, Hao Zhang, and Nan Hua. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 1, 3", + "[40] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A large human-annotated dataset for document layout segmentation. In SIGKDD. ACM, 2022. 6", + "[41] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Płka. Going full-tilt boogie on document understanding with text-image layout transformer. In ICDAR, pages 732-747. Springer, 2021. 1, 2", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2", + "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for train" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing next generation image-text models. NeurIPS, 35:25278-25294, 2022. 2", + "[44] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1", + "[45] Yongxin Shi, Dezhi Peng, Wenhui Liao, Zening Lin, Xinhong Chen, Chongyu Liu, Yuyi Zhang, and Lianwen Jin. ExploringOCR capabilities of GPT-4V(ision): A quantitative and in-depth evaluation. arXiv preprint arXiv:2310.16809, 2023. 3, 6", + "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758. Springer, 2020. 2", + "[47] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 1, 6", + "[48] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.2", + "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[50] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 1, 2, 7", + "[51] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance text-Layout interaction in multi-modal pre-training for document understanding. In ACL, 2023. 3, 4", + "[52] Jiapeng Wang, Lianwen Jin, and Kai Ding. Lilt: A simple yet effective language-independent layout transformer for structured document understanding. In ACL, 2022. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15639", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 2, 5", + "[54] Yiheng Xu, Minghao Li, Lei Cui, and Shaohan Huang. LayoutLM: Pre-training of text and layout for document image understanding. In KDD, pages 1192-1200, 2020. 1, 2", + "[55] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 1, 2, 3, 4", + "[56] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with GPT-4V(ision). arXiv preprint arXiv:2309.17421, 9, 2023. 1, 3", + "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 1, 2, 3, 4, 7", + "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3", + "[59] Yuechen Yu, Yulin Li, Chengquan Zhang, Xiaogiang Zhang, Zengyuan Guo, Xiameng Qin, Kun Yao, Junyu Han, Errui Ding, and Jingdong Wang. Structextv2: Masked visual-textual prediction for document image pre-training. In ICLR, 2023. 1, 2", + "[60] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 3, 4, 7", + "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2, 6, 7", + "[62] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 4, 6", + "[63] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 6", + "[64] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 1", + "[65] Stepan Simsa, Milan Sulc, Michal Uričář, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyás Skalický, Jiří Matas, Antoine Doucet, Mickaël Courstaty, and Dimosthenis Karatzas. Docile benchmark for document information localization and extraction, 2023. 6" + ], + "bbox": [ + 78, + 90, + 467, + 893 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15640", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_model.json b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5de6b468669d0428cb5b0c2b99d7c8198f6a62e2 --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_model.json @@ -0,0 +1,2796 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.203, + 0.798, + 0.24 + ], + "angle": 0, + "content": "Chuwei Luo\\(^{1*}\\), Yufan Shen\\(^{12*}\\), Zhaoqing Zhu\\(^{1*}\\), Qi Zheng\\(^{1}\\), Zhi Yu\\(^{2}\\), Cong Yao\\(^{1}\\), Alibaba Group, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.242, + 0.737, + 0.257 + ], + "angle": 0, + "content": "{luochuwei,zzhaoqing.z,zhengqisjtu,yaocong2010}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.26, + 0.621, + 0.275 + ], + "angle": 0, + "content": "{syficy,yuzhirenzhe}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.474, + 0.751 + ], + "angle": 0, + "content": "Recently, leveraging large language models (LLMs) or multimodal large language models (MLLMs) for document understanding has been proven very promising. However, previous works that employ LLMs/MLLMs for document understanding have not fully explored and utilized the document layout information, which is vital for precise document understanding. In this paper, we propose LayoutLLM, an LLM/MLLM based method for document understanding. The core of LayoutLLM is a layout instruction tuning strategy, which is specially designed to enhance the comprehension and utilization of document layouts. The proposed layout instruction tuning strategy consists of two components: Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. To capture the characteristics of document layout in Layout-aware Pre-training, three groups of pretraining tasks, corresponding to document-level, region-level and segment-level information, are introduced. Furthermore, a novel module called layout chain-of-thought (LayoutCoT) is devised to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers. LayoutCoT is effective for boosting the performance of document understanding. Meanwhile, it brings a certain degree of interpretability, which could facilitate manual inspection and correction. Experiments on standard benchmarks show that the proposed LayoutLLM significantly outperforms existing methods that adopt open-source 7B LLMs/MLLMs for document understanding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.759, + 0.21, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.785, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Document AI [7], including its document understanding tasks such as document VQA [33, 47] and document visual information extraction [18, 19, 37], is currently a hot topic in both academia and industry. In recent years, document pre-trained models [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 54, 55, 59] have achieved excellent performance in doc" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.309, + 0.619, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.388, + 0.576, + 0.396 + ], + "angle": 0, + "content": "(1)LLMs" + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.31, + 0.747, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.388, + 0.704, + 0.396 + ], + "angle": 0, + "content": "(II)MLMs" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.31, + 0.871, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.388, + 0.853, + 0.396 + ], + "angle": 0, + "content": "(III) LayoutLLM (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.402, + 0.737, + 0.412 + ], + "angle": 0, + "content": "(a) Model architectures" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.42, + 0.653, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.421, + 0.772, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.421, + 0.869, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.479, + 0.737, + 0.489 + ], + "angle": 0, + "content": "(b) Training paradigms" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.504, + 0.892, + 0.574 + ], + "angle": 0, + "content": "Figure 1. LLMs/MLLMs for document understanding. The LayoutLLM is an LLM/MLLM based method that integrates a document pre-trained model as encoder. It is trained by the newly proposed layout instruction tuning strategy which consists of Layout-aware Pre-training and Layout-aware Supervised Fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.582, + 0.892, + 0.657 + ], + "angle": 0, + "content": "ument AI downstream tasks. However, due to the necessity for fine-tuning on corresponding downstream task data, it is challenging to directly adapt such pre-trained models for zero-shot document understanding. In this paper, zero-shot refers to not using training sets of downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Recently, large language models (LLMs) such as ChatGPT [35] and LLaMA [49, 50], or multimodal large language models (MLLMs) like GPT-4V [1, 36, 56], have shown remarkable zero-shot capabilities across various applications. For Document AI, as shown in Fig. 1 (a), (I) directly prompting LLMs with document text [15, 39] and (II) training document-based MLLMs [3, 57, 60] have also achieved promising results under the zero-shot setting [3, 39, 57, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It is widely accepted that document layout information is vital for document understanding [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. However, it is difficult to convey document layout information by directly feeding text to LLMs. As Fig. 1(a)(I) shows, representing documents as either flattened plain text or layout text such as text with coordinates [15, 39, 44, 64] is often used for LLMs." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15630" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "Flattened plain text completely excludes any layout information of the document [54]. Additionally, as Tab. 1 shows, using layout text that represents both textual and layout information as inputs for LLMs does not guarantee LLMs can effectively comprehend this formatted text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.473, + 0.518 + ], + "angle": 0, + "content": "Moreover, existing works that employ MLLMs for document understanding also have not fully explored the document layout information. Document-based MLLMs integrate visual models [11, 42] with LLM [48-50, 61] for document understanding. As Fig. 1(b) shows, they are typically based on pre-training and supervised fine-tuning (SFT) on document datasets. In the pre-training stage, tasks such as image captioning [29, 57, 60] or generating all text in a document as flattened plain text [9, 27, 60] are commonly applied. Both these image captions [43, 46] and plain text only provide a brief representation and fail to capture the layout information of the document. So it is difficult for the model to learn document layout in the existing pre-training stage. In the SFT stage, document-related VQA or information extraction data [3, 57] is often used. The answers are directly provided during SFT, lacking explicit learning about document layout. In summary, current approaches using plain or layout text to prompt LLMs and training document-based MLLMs have not effectively captured layout information, limiting their zero-shot document understanding capability. Therefore, for better document understanding with the power of LLMs, it is necessary to investigate how to effectively incorporate layout information into LLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.473, + 0.854 + ], + "angle": 0, + "content": "To this end, we propose LayoutLLM, an LLM/MLLM based method for document understanding, in which a layout instruction tuning strategy is designed to enhance the comprehension of document layouts. Different from the existing MLLMs that use a general visual pre-trained model [11, 42] as the encoder, we integrate document pretrained models [2, 12, 16, 17, 23, 25, 32, 52, 54, 55] as the encoder in order to better leverage the model's foundational understanding capability for documents. The proposed layout instruction tuning consists of two stages: layout-aware pre-training and layout-aware supervised fine-tuning (SFT). Due to the complex nature of documents in real-world scenarios, encompassing rich textual content and diverse layout structures, achieving a thorough understanding involves not only capturing the document's fundamental content at global but also delving into local details. In the layout-aware pre-training stage, to ensure the model learns not only the global information of documents but also detailed information at different hierarchical levels, three groups of different level pre-training tasks are proposed: document-level, region-level, and segment-level. All the proposed pretraining tasks are unified in the format of instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Furthermore, in the layout-aware SFT stage, to enhance the model's comprehension and utilization of layout information for question answering, a novel strategy called" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.333 + ], + "angle": 0, + "content": "LayoutCoT is proposed, motivated by the chain-of-thought (CoT) [21, 53] ability in LLMs. Unlike existing methods that are directly supervised by the answer to the document understanding question, LayoutCoT consists of three successive steps: Question Analysis, Relevant Area Concentration, and Answer Formation. Through these steps, the model gains a deeper understanding of the questions, becomes capable of focusing the relevant areas instead of searching answers in the entire document and can leverage the specific characteristics of identified areas (such as tables, paragraphs, etc.) to accurately infer the answers. It not only brings a certain degree of interpretability, but also provides a feasible way for manual intervention or correction of model results. Extensive zero-shot experiments on five widely-used document understanding benchmarks demonstrate the effectiveness of the proposed LayoutLLM." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.334, + 0.822, + 0.347 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.494, + 0.35, + 0.892, + 0.41 + ], + "angle": 0, + "content": "1) To better learn document layouts from global to local in layout-aware pre-training, three groups of different level pre-training tasks, which are all implemented through instruction tuning, are proposed." + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.411, + 0.892, + 0.486 + ], + "angle": 0, + "content": "2) A novel LayoutCoT strategy is proposed to achieve layout-aware supervised fine-tuning. It enables Layout-LLM to focus on the relevant document area and leverage the region's features to generate accurate answers, exhibiting a certain degree of interpretability." + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.486, + 0.892, + 0.562 + ], + "angle": 0, + "content": "3) Experimental results on zero-shot document understanding tasks show that the proposed LayoutLLM significantly outperforms existing methods that adopt LLMs/MLLMs for document understanding, demonstrating the great potential of document layout modeling." + }, + { + "type": "list", + "bbox": [ + 0.493, + 0.35, + 0.892, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.651, + 0.589 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Pre-trained models for document understanding. Document pre-trained models have demonstrated the effectiveness of layout information in document understanding [2, 5, 8, 10, 12, 13, 16, 17, 20, 22, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. As a pioneer, LayoutLM [54] is the first to encode spatial coordinates of text for layout representation learning in pre-training documents. The following works [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 55, 59] then joint text, layout and images in document pre-training by combining visual models as document image encoders with the text and layout transformers, and various works [5, 10, 20, 22] start to explore pre-training end-to-end models for document understanding. These studies have achieved significant advancements in document understanding by exploring various model architectures [2, 8, 10, 12, 13, 20, 25, 26, 38, 41, 52, 59] and attention mechanisms [16, 17, 55] for modeling layout information. These methods also have proposed layout-related pre-training tasks that have been demonstrated to be highly effective in document understanding tasks. For instance," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15631" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "tasks like masked vision-language modeling [13, 17, 55], where the model is required to generate the original text corresponding to the randomly masked text in the document; position masking [31, 51], involving the randomly position masking and subsequent recovery of position information in the document; geometric pre-training [26, 32], focusing on learning direction, distance, etc.; and layout-aware generation tasks [5, 22], aiming to make the model generate structured text with layout information. However, due to the necessity of fine-tuning with annotated data for downstream tasks, these efforts face challenges in extending to zero-shot document understanding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.637 + ], + "angle": 0, + "content": "LLMs/MLLMs for document understanding. Recently, LLMs like ChatGPT [35] and MLLMs like GPT-4 [36, 56] have demonstrated remarkable zero-shot performance across a wide range of NLP/CV tasks. Leveraging LLMs/MLLMs for zero-shot document understanding has also shown promising progress [3, 30, 39, 45, 57, 60]. Perot et al. [39] explore the use of LLMs for document visual information extraction, emphasizing the importance of the document layout. LLaVAR [60] which extends LLaVA [28, 29] to the document domain is pre-trained by generating plain text in the document image. During SFT, it is trained by document-related instructions which are generated by GPT-4. Expanding on mPLUG-Owl [58], mPLUG-DocOwl [57] is trained using publicly available datasets for document understanding. It includes tasks like document-level image captioning, direct information extraction, and direct document VQA. Moreover, Qwen-VL [3] proposes a general MLLM that performs well on document understanding tasks, utilizing document-level pre-training and direct VQA for SFT. Though existing LLMs/MLLMs have shown promising results in document understanding, their limited focus on document layout in pre-training and SFT has hindered their ability to achieve higher accuracy in zero-shot document understanding and better interpretability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.648, + 0.208, + 0.664 + ], + "angle": 0, + "content": "3. LayoutLLM" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.765 + ], + "angle": 0, + "content": "LayoutLLM is an LLM/MLLM based method that incorporates document pre-trained models for document understanding. To enhance the document layout comprehension in LayoutLLM, a novel layout instruction tuning strategy is proposed, which consists of two stages: layout-aware pretraining and layout-aware supervised fine-tuning (SFT)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.265, + 0.787 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The overall architecture of LayoutLLM is shown in Fig. 2. In LayoutLLM, given an input document image and its corresponding text and layout information, the document pre-trained model encoder is required to obtain the multimodal document features. Then, these features are encoded by multimodal projectors, and together with the instruction embeddings, fed into the LLM to generate the final results." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.09, + 0.843, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.237, + 0.833, + 0.251 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of LayoutLLM." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.256, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Document pre-trained model encoder. To leverage the foundational document comprehension capability of document pre-trained models, in this work, we utilize LayoutLMv3 [17], a widely-used document pre-training model, as our basic document encoder. The document image, text, and layout are initially inputted into the document pretrained model (DocPTM). They are then encoded by the DocPTM to obtain the corresponding features as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.392, + 0.892, + 0.408 + ], + "angle": 0, + "content": "\\[\nF _ {V}, F _ {T} = \\operatorname {D o c P T M} (V, T, B o x) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.892, + 0.547 + ], + "angle": 0, + "content": "where \\(V\\) represents the document image, \\(T = T_{0:n - 1}\\) and \\(Box = Box_{0:n - 1}\\) indicate the text sequences in the document and their corresponding bounding-box coordinates respectively. After being encoded by the DocPTM, the visual features of the document \\(F_{V} = F_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{0}}\\) and the text layout features \\(F_{T} = F_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{0}}\\) are acquired. \\(m\\) signifies the number of visual features and \\(n\\) represents the number of tokens contained in the document. \\(d_0\\) denotes the dimension of DocPTM feature space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.548, + 0.892, + 0.638 + ], + "angle": 0, + "content": "Multimodal projectors. To project multi-modality features from DocPTM into the LLM's embedding space, inspired by the simple yet effective projector design in LLaVA [28, 29], two different Multilayer Perceptrons (MLPs) are used as visual projector and text & layout projector respectively. Formally, the projected features can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.645, + 0.891, + 0.661 + ], + "angle": 0, + "content": "\\[\nH _ {V} = P _ {V} \\left(F _ {V}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.665, + 0.891, + 0.681 + ], + "angle": 0, + "content": "\\[\nH _ {T} = P _ {T} \\left(F _ {T}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.891, + 0.749 + ], + "angle": 0, + "content": "where \\(H_{V} = H_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{1}}\\) is the feature encoded by the visual projector, \\(H_{T} = H_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{1}}\\) is the feature encoded by the text & layout projector, and \\(d_{1}\\) is the dimensional of the LLM embedding features." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Large language model. Finally, the \\( H_V \\), \\( H_T \\) and the embedding of the question's instruction text, \\( I_q = I_{q_0:l_q - 1} \\), are inputted together into the LLM, generating the target answer \\( I_a = I_{a_0:l_a - 1} \\). \\( l_q \\) and \\( l_a \\) represent the length of the question's instruction text and the answer text, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.742, + 0.849 + ], + "angle": 0, + "content": "3.2. Layout Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The LayoutLLM model is trained using the layout instruction tuning, which consists of two stages: layout-aware pretraining and layout-aware SFT." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15632" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.885, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Figure 3. Overview of the Layout Instruction Tuning. (a) Document-level, region-level, and segment-level pre-training tasks, unified in instruction tuning format, are introduced. (b) A novel module called LayoutCoT is designed to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers through three intermediate steps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.385, + 0.319, + 0.4 + ], + "angle": 0, + "content": "3.2.1 Layout-aware Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.469, + 0.534 + ], + "angle": 0, + "content": "The goal of pre-training the LayoutLLM is to enhance the model's comprehensive understanding of documents at different levels through layout learning, rather than only focusing on global document-level understanding like existing MLLM methods [57, 60]. To this end, during the pre-training stage, three different level pre-training strategies are simultaneously applied to the LayoutLLM, namely document-level, region-level, and segment-level." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Document-level To enable the model to possess fundamental global document understanding, pre-training tasks, namely Document Dense Description (DDD) and Text and Layout Reconstruction (TLR), are proposed. As Fig. 3 (a) shows, like the image caption task, the DDD task requires the model to learn to describe the input document. Moreover, in the DDD task, the descriptions for document images are more detailed. For instance, in the document image caption data used for the LLaVAR [60] pre-training, captions contain an average of 36.27 words, while in the proposed dataset of the DDD task, the description contains an average of 373.25 words. Through the DDD task, the model can obtain basic document-level information, such as document type and detailed content. The TLR task aims to reconstruct the complete text and layout information of a document and output it in the format “<\\{box\\}, \\{text\\} >”. The TLR task aligns the text and layout embeddings output from DocPTM with the LayoutLLM's LLM space. Consequently, it enables the LLM in LayoutLLM to comprehend the text and layout information contained in the documents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Region-level The information contained in specific regions of a document, such as the titles, figures, tables, is essential for document understanding [4, 6, 8, 24, 62]. These regions serve as important characteristics that differentiate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.595 + ], + "angle": 0, + "content": "a document from plain text in natural language. For the LayoutLLM to achieve the basic region-level understanding, two pre-training tasks, namely Document Layout Analysis (DLA) and Table Understanding (TU), are utilized. The DLA task is achieved in two ways as shown in Fig. 3. One involves locating the layout region based on the layout type, the other involves identifying the type of a given area. Furthermore, the table region differs from other regions in that it requires additional focus on 2D layout understanding. The TU task enables the model to understand the basic row and column information in the table region of a document. As shown in Fig. 3, the TU task includes instruction tuning for the number of rows and columns, logical coordinates, and the content within rows and columns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Segment-level Early works on document pre-trained models [13, 17, 26, 31, 32, 51, 55] have demonstrated the effectiveness of segment-level document pre-training tasks to document layout understanding ability, such as masked vision language modeling (MVLM) [13, 17, 55], position masking [31, 51], and geometric pre-training [26, 32]. Inspired by these works, to make LayoutLLM have segment-level layout understanding, these tasks are transformed into instruction formats for pre-training as Fig. 3(a) shows. For the MVLM instructions, random masking of the text input to LayoutLLM is performed, and the model is instruction tuned by asking the masked words and answering them. For mask position instruction, the layout information (coordinates) to a specific text line, when input to LayoutLLM, is randomly set to 0. The instruction is constructed by asking about the text line with zeroed coordinates and requesting the model to respond with the original coordinates with text content. For geometric layout instruction, text lines are randomly selected, and an instruction is constructed by asking questions about the direction and distance between them." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15633" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.396, + 0.107 + ], + "angle": 0, + "content": "3.2.2 Layout-aware Supervised Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.326 + ], + "angle": 0, + "content": "In the SFT stage of existing document-based MLLMs, models are directly supervised by the answer to the document understanding instructions. Consequently, these methods lack explicit learning of document layout which is crucial for document understanding. Considering this limitation, and inspired by previous works related to chain-of-though (CoT) [21, 53], which have shown that inferencing with intermediate steps can greatly enhance performance. A novel module called LayoutCoT is proposed, which incorporates the layout information into every intermediate step of CoT explicitly. Meanwhile, by introducing the layout-aware intermediate steps, the answer process gains a certain degree of interpretability for LayoutLLM and also provides interactive correction possibilities based on LayoutCoT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.327, + 0.471, + 0.357 + ], + "angle": 0, + "content": "LayoutCoT Details. As shown in Fig. 3(b), the LayoutCoT involves the following three intermediate steps:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.583 + ], + "angle": 0, + "content": "Step 1: Question Analysis. To effectively address a document understanding problem, analyzing the key characteristics of the question is very important. Identifying the question type, such as table understanding or entity extraction from paragraphs, and assessing whether the question is a straightforward extraction query or a more complex reasoning problem, can help guide the direction for the subsequent inference process. Therefore, to give basic guidance to the subsequent steps, the question analysis step is designed, encompassing an analysis of the question type from a layout perspective and a detailed understanding of the question itself. Benefiting from the layout understanding ability by layout-aware pre-training, this step can extract the types and key information mentioned in the question, which are related to the specific characteristics of the document." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.824 + ], + "angle": 0, + "content": "Step 2: Relevant Area Concentration. For most document understanding tasks, the entire document contains a large amount of irrelevant information that may confuse the model [5]. This step aims to focus on the relevant area and generate its location information, which is used to assist the model to accurately infer the answer. Benefiting from the layout information conveyed by step 1 and the positioning capabilities learned from the region&segment-level pretraining, the model can accurately generate the location of the relevant area. For example in Fig. 3(b), according to the question type \"table\" in step 1, the relevant \"table\" can be located. By guiding the model to focus on the relevant area, this step largely narrows the search scope, increasing the possibility of giving the right answer. Meanwhile, the location information provides a way for visual inspection and interactive correction (see Sec. 4.7 for details)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Step 3: Answer Formation. Finally, the last step, the answer formation, provides explanations based on the layout characteristics of the relevant areas located in step 2 and key points analyzed in step 1 to get the final answer. For example in Fig. 3(b), for a \"table\" type question, this step in-" + }, + { + "type": "title", + "bbox": [ + 0.505, + 0.088, + 0.889, + 0.103 + ], + "angle": 0, + "content": "Algorithm 1 CONSTRUCT(D): LayoutCoT Construction." + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.106, + 0.89, + 0.142 + ], + "angle": 0, + "content": "Definition: \\(\\mathcal{H}\\) : Document HTMLs; \\(\\mathcal{I}\\) : Document Images; \\(\\mathcal{T}\\) : MRC Texts; \\(\\mathcal{R}\\) : Document Language Representation; QA: QA pairs; \\(T_{c}\\) : Text CoT; \\(\\mathcal{L}_{c}\\) : LayoutCoT;" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.143, + 0.771, + 0.156 + ], + "angle": 0, + "content": "Input: Document Dataset \\(\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\}\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.528, + 0.156, + 0.753, + 0.168 + ], + "angle": 0, + "content": "\\((\\mathcal{D}_H = \\{\\mathcal{H}\\}, \\mathcal{D}_I = \\{\\mathcal{I}\\}, \\mathcal{D}_M = \\{\\mathcal{T}, \\mathcal{Q}\\mathcal{A}\\})\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.168, + 0.687, + 0.179 + ], + "angle": 0, + "content": "Output: Constructed Dataset \\(\\mathcal{D}_c\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.183, + 0.796, + 0.196 + ], + "angle": 0, + "content": "1: Procedure CONSTRUCT \\((\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\})\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.196, + 0.88, + 0.208 + ], + "angle": 0, + "content": "2: 1) \\(\\mathcal{R} = getDocRep(\\mathcal{I})\\) if \\(\\mathcal{D} \\subseteq \\mathcal{D}_I\\) else \\(\\mathcal{H}\\) if \\(\\mathcal{D} \\subseteq \\mathcal{D}_H\\) else pass;" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.208, + 0.89, + 0.22 + ], + "angle": 0, + "content": "3: 2) \\(\\mathcal{Q}\\mathcal{A},\\mathcal{T}_c = getQACoT(\\mathcal{D})\\) if \\(\\mathcal{D}\\subseteq \\mathcal{D}_M\\) else getQACoTGPT(R);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.22, + 0.688, + 0.232 + ], + "angle": 0, + "content": "4: 3) \\(\\mathcal{L}_c = getLayoutCoT(\\mathcal{T}_c)\\)" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.232, + 0.831, + 0.244 + ], + "angle": 0, + "content": "5: 4) if \\(\\mathcal{D} \\subseteq \\{\\mathcal{D}_H, \\mathcal{D}_M\\}\\): \\(\\mathcal{I} = Htm12Img(\\mathcal{D}_H \\cap \\mathcal{H} : \\mathcal{T})\\);" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.244, + 0.678, + 0.256 + ], + "angle": 0, + "content": "6: return \\(\\mathcal{D}_c\\gets \\{\\mathcal{I},\\mathcal{Q}\\mathcal{A},\\mathcal{L}_c\\}\\)" + }, + { + "type": "list", + "bbox": [ + 0.509, + 0.106, + 0.89, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.382 + ], + "angle": 0, + "content": "volves analyzing the row and column in the relevant table in step 2, and referencing the answer step-by-step. For a \"key-value\" question, analyzing the keywords in concentrated areas can help get the final answer. Analyzing answers in different ways based on the features of different layout regions not only improves the document understanding performance but also brings a certain level of interpretability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.518 + ], + "angle": 0, + "content": "LayoutCoT Construction. Given the need for both text and image annotations in constructing LayoutCoT, manual labeling can be difficult. Algorithm 1 proposes a manual-labeling-free method, generating LayoutCoT data using public datasets with GPT (GPT-3.5 Turbo). It involves representing document text and layout in a format understandable by GPT. GPT is then utilized to generate document-content-based QA and corresponding text CoT. Finally, use rules for transforming the text CoT to LayoutCoT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Three types of publicly available document datasets are focused on: HTML documents \\((\\mathcal{D}_H)\\), image documents \\((\\mathcal{D}_I)\\), and text documents \\((\\mathcal{D}_M)\\) for machine reading comprehension (MRC). The construction process is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.717 + ], + "angle": 0, + "content": "1) Document Representation: To fully leverage the capabilities of GPT, it is crucial to ensure that the document content fed to GPT contains accurate layout information. Since HTML is the formatted language that can represent documents accurately, \\( \\mathcal{D}_H \\) is represented using the original HTML. By transforming HTML to PDF and using the PDF parser, the text and bounding-boxes are obtained. For \\( \\mathcal{D}_I \\), the layout-aware text [15] is used. The text and bounding-boxes in \\( \\mathcal{D}_I \\) are from the original dataset annotations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.838 + ], + "angle": 0, + "content": "2) QA&Text CoT Generation: The language representation \\(\\mathcal{R}\\) for the document is employed for prompting GPT to generate QA pairs \\(\\mathcal{QA}\\) with text CoTs \\(\\mathcal{T}_c\\). In addition, the \\(\\mathcal{D}_M\\) includes the QA pairs and reasoning process, thereby directly reusing the \\(\\mathcal{QA}\\) and manually organizing \\(\\mathcal{T}_c\\). The generated \\(\\mathcal{T}_c\\) includes the step 1 (question analysis) and step 3 (answer formation) for LayoutCoT, and locates all relevant sentences in the document for \\(\\mathcal{QA}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "3) LayoutCoT Generation: The step 1 & 3 in \\(\\mathcal{T}_c\\) are used as the step 1 & 3 in \\(\\mathcal{L}_c\\). To construct the step 2 (relevant area concentration) of \\(\\mathcal{L}_c\\), the union bounding-box of all located relevant sentences in \\(\\mathcal{T}_c\\) are taken as the relevant area." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15634" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "4) Document Images Generation: For \\(\\mathcal{D}_H\\) and \\(\\mathcal{D}_M\\), the HTMLs and MRC text are converted to images. Overall, the document images \\(\\mathcal{I}\\), generated QA pairs \\(\\mathcal{QA}\\) and LayoutCoTs \\(\\mathcal{L}_c\\) constitute the final LayoutCoT dataset \\(\\mathcal{D}_c\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.17, + 0.21, + 0.187 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.196, + 0.254, + 0.211 + ], + "angle": 0, + "content": "4.1. Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.473, + 0.614 + ], + "angle": 0, + "content": "Layout-aware pre-training data of LayoutLLM is from publicly available document understanding datasets. It does not incorporate any data from the training, validation, and test sets of downstream benchmarks. Region-level pretraining tasks, most document-level and segment-level tasks are self-supervised. Thus, only document images and images converted from PDFs in the original datasets, along with the corresponding OCR or text layout results from PDF parsing, are needed. For these tasks, data is randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65]. Particularly, data for document dense description is from inputting the document text content into GPT-3.5 Turbo, prompting it to generate an average of 373.25 words document dense descriptions. For the region-level tasks, specifically the document layout analysis task, publicly available document layout analysis datasets are utilized, including PubLayNet [62], DocLayNet [40], and Docbank [24]. Data for another region-level task, table understanding, is sourced from PubTabNet [63] with its table annotations. All data is transformed into the instruction format illustrated in Fig. 3(a). In total, 5.7 million instructions are constructed, with a ratio of 1:4:4 for document-level, region-level, and segment-level tasks, respectively. For detailed instruction templates and dataset descriptions, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Layout-aware SFT data of LayoutLLM is generated by GPT (GPT-3.5 Turbo) and converted from existing textual Machine Reading Comprehension (MRC) datasets, as discussed in Sec. 3.2.2. To generate high-quality document-based textual QA and textual CoT, it is essential to make GPT comprehend the document layout. So, the document is represented using both layout text [15] and HTML. Similar to the pre-training data, the \\(D_I\\) in Algorithm 1 is also randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65] for building layout text. The \\(D_H\\) in Algorithm 1 is from GPT's free generation. The \\(D_M\\) in Algorithm 1 is randomly sampled from the FeTaQA [34] which is a wikipedia question answering dataset. A total of 300K instructions are constructed, with a ratio of 5:4.5:0.5 for \\(D_I\\), \\(D_H\\), and \\(D_M\\), respectively. For detailed prompt templates of document-based text QA and text CoT generation using GPT, prompts for HTML generation using GPT, and dataset description, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.655, + 0.108 + ], + "angle": 0, + "content": "4.2. Training Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.895, + 0.268 + ], + "angle": 0, + "content": "The encoder weight of LayoutLLM is initialized from the LayoutLMv3-large [17] which is a widely-used document pre-trained model. And the LLM backbone is initialized from Vicuna-7B-v1.5 [61]. Other parameters are randomly initialized. During pre-training, the LLM is frozen, and the parameters of the two projectors and document pre-trained model encoder are updated. During SFT, both LLM and two projectors are fine-tuned while keeping the document pretrained model encoder frozen. For detailed training setup, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.273, + 0.671, + 0.29 + ], + "angle": 0, + "content": "4.3. Evaluation Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.895, + 0.751 + ], + "angle": 0, + "content": "The zero-shot ability is highly expected in real-world document understanding scenarios [7, 30, 45]. Therefore, zero-shot evaluations on widely-used document understanding benchmarks including document visual question answering (Document VQA) and visual information extraction (VIE) are conducted. Only the test sets are utilized in all benchmarks and only the official provided image, text, and layout information are used. The Document VQA datasets comprise the DocVQA[33] test set, consisting of 5,188 questions, and the VisualMRC[47] test set containing 6,708 questions. Following the evaluation metric settings of the original datasets, the ANLS [33] is utilized for evaluating DocVQA, and Rouge-L is used for evaluating VisualMRC. For the VIE task, FUNSD [19], CORD [37], and SROIE [18] are used. The test set of FUNSD comprises 50 form images, each annotated with entity-level headers, questions, answers, and others, along with entity linking annotations. CORD's test set consists of 100 receipt images, annotated with 30 entity types, such as the tax amount, total price, etc. SROIE's test set includes 347 receipt images, annotated with 4 entity types: company, date, address, and total. To prompt LLMs/MLLMs for zero-shot VIE, annotations in VIE datasets are transformed into question answering format (QA for VIE). For key-value annotations with linking in the FUNSD, the format is {Q: What is the \"key\" in the document? A: \"value)}. For entity annotations in CORD and SROIE, directly asking for the entity in the document like {Q: What is the address in the document? A: \"the address annotation\"} is utilized. Following DocVQA, the QA for VIE task is evaluated by ANLS." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.757, + 0.64, + 0.773 + ], + "angle": 0, + "content": "4.4. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 1, the zero-shot document understanding performance of LayoutLLM and existing open-source LLMs and MLLMs is evaluated. Generally, the existing LLMs are better than MLLMs for zero-shot document VQA and VIE. For example in the results on DocVQA, most LLMs can achieve a performance of around \\(60\\%\\) or higher, while most MLLMs can only attain around \\(10\\%\\), except for mPLUG-DocOWL and Qwen-VL that trained with the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15635" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.079, + 0.872, + 0.326 + ], + "angle": 0, + "content": "
ModelsDocument VQAQA for VIE
DocVQAVisualMRCFUNSDCORDSROIE
Fine-tuned PTMLayoutLMv3 [17]83.37*-92.08*‡97.46*‡-
LLMLlama2-7B [50]61.3429.7340.784.3915.86
Llama2-7B-chat [50]64.9952.8448.2047.7068.97
Plain TextVicuna-7B [61]61.3953.6349.7944.6767.49
Vicuna-1.5-7B [61]66.9952.1348.0651.4068.20
LLMLlama2-7B [50]37.3233.8251.4028.0434.96
Llama2-7B-chat [50]56.5549.2658.3450.9351.15
Layout Text (Text + Box) [15]Vicuna-7B [61]37.2152.5542.7346.5945.43
Vicuna-1.5-7B [61]56.8147.2259.6356.1366.20
MLLMLLaVAR-7B [60]11.6†36.371.7113.552.38
LLaVA-1.5-7B [28]13.3435.231.9318.063.83
mPLUG-DocOWL-7B [57]62.2*†----
Qwen-VL-7B [3]65.1*†42.5247.0930.0058.59
LayoutLLM-7B△ (Ours)74.2555.7378.6562.2170.97
LayoutLLM-7B★ (Ours)74.2755.7679.9863.1072.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Table 1. Zero-shot document understanding results on open-source LLMs and MLLMs. * signifies training set use; unmarked results are zero-shot. Results marked with \\( {}^{ \\ddagger } \\) are F1 scores for VIE. Results marked with \\( \\dagger \\) are from the original paper and others are re-implemented by us. \\( \\bigtriangleup \\) marks LayoutLLM's LLM backbone as initialized with Llama2-7B-chat,and \\( \\star \\) with Vicuna-1.5-7B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.612 + ], + "angle": 0, + "content": "training set. One possible reason is that it's difficult for these MLLMs to obtain accurate textual information from document images. Additionally, for LLMs, using Plain Text and Layout Text respectively as the document representation are further discussed, where the Layout Text introduces layout information by adding text coordinates in the format: {text:\"text\", box:[x1,y1,x2,y2]} [15]. Compared to the Plain Text, the Layout Text variant doesn't show stable performance improvements, noticeable in certain tasks, for example in Vicuna-1.5, an improvement in VIE (FUNSD \\(48.06\\%\\) to \\(59.63\\%\\)) but a decline in DocVQA \\((66.99\\%\\) to \\(56.81\\%)\\). LLMs may lack the ability to learn this formatted layout text, and directly adding layout information (e.g., coordinates) to the text will also highly increase the token length, making the answer inference more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Compared to the prior SOTA model, LayoutLMv3, which is fine-tuned using the training set of downstream tasks, LayoutLLM demonstrates competitive performance on the DocVQA benchmark. Compared with these LLMs and MLLMs, LayoutLLM achieves consistent and significant improvements over them on all evaluation benchmarks. Notably, LayoutLLM which employs zero-shot performance, outperforms mPLUG-DocOWL and Qwen-VL by around \\(10\\%\\) on the DocVQA dataset, both of which are trained with this dataset. This demonstrates that LayoutLLM can learn more robust and discriminative representations for document understanding. Furthermore, experiments of the different initialization of the LLM backbone have all achieved optimal results across all benchmarks, substantiating that LayoutLLM can adapt to various LLMs. In summary, our method explores a more effective way to utilize layout information for document understanding, which significantly improves the performance of zero-shot document understanding." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.377, + 0.89, + 0.445 + ], + "angle": 0, + "content": "
#Layout-aware \nPre-trainingLayout-aware \nSFTDocVQAFUNSD
070.8270.96
1✓72.3174.02
2✓✓74.2779.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.508, + 0.448, + 0.882, + 0.462 + ], + "angle": 0, + "content": "Table 2. Ablation study on the DocVQA and FUNSD test sets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.651, + 0.483 + ], + "angle": 0, + "content": "4.5. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.892, + 0.536 + ], + "angle": 0, + "content": "To better verify the effectiveness of the layout-aware pretraining and layout-aware SFT in the layout instruction tuning, an ablation study is conducted (see Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Initial baseline. The #0 baseline disables both layout-aware pre-training and layout-aware SFT. It only adopts SFT (same SFT data but without LayoutCoT steps) on the LayoutLLM. Even without any alignment pre-training and LayoutCoT steps guidance, the baseline outperforms previous SOTAs, achieving \\(70.82\\%\\) on DocVQA and \\(70.96\\%\\) on FUNSD. This indicates the document understanding ability of DocPTM benefits document understanding with LLMs. Effect of Layout-aware pre-training. In #1, the layout-aware pre-training and the same SFT with #0 is conducted. Compared to #0, benefiting from the basic document understanding capability learned through the layout-aware pretraining, #1 shows an increase of \\(1.49\\%\\) on DocVQA and \\(3.06\\%\\) on FUNSD. It can be observed that the basic document understanding ability learned from the layout-aware pre-training significantly enhances the performance of the basic key-value extraction tasks in the FUNSD dataset. Compared to the FUNSD VIE task, it also shows that the DocVQA is a more complex task. Effect of Layout-aware SFT. Compared with #1, #2 further incorporates the layout-aware SFT strategy, resulting in a performance gain of \\(1.96\\%\\) on DocVQA and even \\(5.96\\%\\) on FUNSD. This indicates the LayoutCoT in layout-aware SFT can help the LayoutLLM to handle complex document tasks and it is effic" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.078, + 0.26, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.148, + 0.259, + 0.158 + ], + "angle": 0, + "content": "Question: What is the Income Tax Withheld for the month of July 1954?" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.079, + 0.409, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.161, + 0.258, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.192, + 0.259, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.261, + 0.247, + 0.267 + ], + "angle": 0, + "content": "Question: What is the Total for Exit 2004?" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.27, + 0.258, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.293, + 0.271, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.189, + 0.409, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.291, + 0.357, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on DocVQA. Green boxes are the areas concentrated in the step 2 of LayoutCoT." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.469, + 0.413 + ], + "angle": 0, + "content": "tive for boosting the performance on both document VQA and VIE. Moreover, it also promotes a certain degree of interpretability. Overall, the ablation study demonstrates the importance of layout-aware pre-training and layout-aware SFT for zero-shot document understanding." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.423, + 0.262, + 0.438 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.473, + 0.809 + ], + "angle": 0, + "content": "Two examples are shown in Fig. 4. Through combined with layout-aware pre-training and layout-aware SFT, Layout-LLM can accurately focus on the relevant areas, utilize the layout information to assist in problem-solving and provide interpretability. For example, in Fig. 4(a) question about key-value extraction in up-down layout, different from the left-right variant, relies more on document layout to infer the right answer. Since the keywords \"Income Tax\" in the question often co-occur with numerical data, Vicuna-1.5 and Qwen-VL find numerical answers relying more on the semantics than the layout, resulting in incorrect responses. In contrast, benefiting from the layout-aware pre-training, our model can effectively leverage layout information to give accurate answers. In addition, the model using LayoutCoT can further provide the location and the reasoning process, showing a certain degree of interpretability. But in certain situations, only combined with layout pre-training, our model might fail to give accurate answers. As shown in Fig. 4(b), without LayoutCoT, our model identifies \"Exit 2003\" as the relevant column and generates a wrong answer. However, with the help of LayoutCoT, LayoutLLM can correctly identify the question type as \"Table\", locate the relevant table area, and finally infer the right answer from the corresponding \"Exit 2004\" column." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.418, + 0.833 + ], + "angle": 0, + "content": "4.7. Interactive Correction with LayoutCoT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Since LayoutCoT runs in a step-by-step fashion and produces intermediate results in the inference stage, it can facilitate interactive inspection and correction, when processing a document. As shown in Fig. 5(a), there are two areas" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.085, + 0.675, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.159, + 0.67, + 0.165 + ], + "angle": 0, + "content": "Question: In the table given on the left bottom of" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.166, + 0.598, + 0.172 + ], + "angle": 0, + "content": "the page, what is it?" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.173, + 0.597, + 0.179 + ], + "angle": 0, + "content": "GT Answer: 013684" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.179, + 0.676, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.256, + 0.675, + 0.263 + ], + "angle": 0, + "content": "Question: What is the heading of second paragraph?" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.264, + 0.638, + 0.271 + ], + "angle": 0, + "content": "GT Answer: PRELIMINARY STATEMENT" + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.086, + 0.824, + 0.094 + ], + "angle": 0, + "content": "Step 1 (Question Analysis): The question is about key-value extraction. And the question asks for the 'vendor'." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.095, + 0.831, + 0.101 + ], + "angle": 0, + "content": "Step 2 (Relevant Area Concentration): The answer can be found" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.102, + 0.761, + 0.107 + ], + "angle": 0, + "content": "in the area: [28, 123, 317, 148]." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.108, + 0.829, + 0.114 + ], + "angle": 0, + "content": "Step 3 (Answer Formation): The number entered for 'vendor' in the text is:" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.115, + 0.773, + 0.121 + ], + "angle": 0, + "content": "Table 1 Table 2" + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.122, + 0.765, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.122, + 0.777, + 0.127 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.129, + 0.765, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.129, + 0.777, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.137, + 0.746, + 0.142 + ], + "angle": 0, + "content": "Step 1 (Question Analysis):" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.143, + 0.831, + 0.15 + ], + "angle": 0, + "content": "Step 2 (Relevant Area Concentration): The answer can be found in the text: [46,703,467,830]." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.151, + 0.829, + 0.157 + ], + "angle": 0, + "content": "Step 3 (Answer Formation): The number entered for 'vendor' in" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.157, + 0.732, + 0.162 + ], + "angle": 0, + "content": "the table is 013684" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.163, + 0.736, + 0.169 + ], + "angle": 0, + "content": "Final answer: 013684." + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.17, + 0.597, + 0.176 + ], + "angle": 0, + "content": "GT Answer: 013684" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.177, + 0.676, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.256, + 0.675, + 0.263 + ], + "angle": 0, + "content": "Question: What is the heading of second paragraph?" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.264, + 0.638, + 0.27 + ], + "angle": 0, + "content": "GT Answer: PRELIMINARY STATEMENT" + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.179, + 0.829, + 0.199 + ], + "angle": 0, + "content": "Step 1 (Question Analysis): The question is about Factoid. And the question asks for the heading of the second paragraph... Step 2 (Relevant Area Concentration): The answer can be found in the area: 134, 280, 654, 3731." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.201, + 0.812, + 0.206 + ], + "angle": 0, + "content": "Step 3 (Answer Formation): The heading of the second" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.207, + 0.82, + 0.212 + ], + "angle": 0, + "content": "paragraph is \\*TAKEDA'S RESPONSE AND OBJECTIONS TO" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.213, + 0.819, + 0.219 + ], + "angle": 0, + "content": "PRACTITIONER'S SUPPLEMENTAL INJURY CLAIMS. ADDITIONAL ACTOS PERSONAL INJURY CLAIMS*." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.22, + 0.824, + 0.226 + ], + "angle": 0, + "content": "Final answer: TAXEDA'S RESPONSE AND OBJECTIONS TO" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.227, + 0.835, + 0.233 + ], + "angle": 0, + "content": "PLAINTIPS SUPPLEMENTAL INTERROGATORY 2 REGRADING ADDITIONAL ACTS PERONEAL INJURY CLAIMS" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.234, + 0.765, + 0.239 + ], + "angle": 0, + "content": "□" + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.234, + 0.799, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.242, + 0.745, + 0.247 + ], + "angle": 0, + "content": "Step 1 (Question Analysis):" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.248, + 0.83, + 0.254 + ], + "angle": 0, + "content": "Step 2 (Relevant Area Concentration): The answer can be found" + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.255, + 0.765, + 0.26 + ], + "angle": 0, + "content": "in the area: [117, 566, 678, 681]." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.261, + 0.81, + 0.266 + ], + "angle": 0, + "content": "Step 3 (Answer Formation): The heading of the second paragraph is \"PRESUMMARY STATEMENT\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.682, + 0.267, + 0.779, + 0.274 + ], + "angle": 0, + "content": "Final answer: PRELIMINARY STATEMENT" + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.278, + 0.892, + 0.319 + ], + "angle": 0, + "content": "tion with LayoutCoT. Green represents and answers, while Red represents the list viewed in digital version." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.893, + 0.591 + ], + "angle": 0, + "content": "in the image that are relevant to the keyword \"vendor\" of the question. LayoutCoT focused on a wrong area containing \"vendor\", as it missed \"left bottom\" in the question, the answer was incorrect. However, after the right area is given manually, it can finally give the correct answer. Similarly, in Fig. 5(b) the question asks about \"the heading of the second paragraph\". However, the term \"paragraph\" does not have a universal definition, and in this case, the sentences in the area below the main heading were considered as a paragraph, causing the model to predict \"TAKEDA's...CLAIMS\", which was incorrect according to the GT. Once the right \"second paragraph\" region is fed to the model, the answer can be successfully revised. This unique ability of LayoutCoT could be very valuable in high-stake scenarios (e.g., a bank transaction), where the standards are extremely high, and manual checking and correction (i.e., human-in-the-loop) are required." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.623, + 0.621 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Through LayoutCoT, LayoutLLM demonstrates the capability of interactive correction, but in real-world applications, this is not enough. The ability to refuse false-positive outputs and generate hints (e.g. \"The answer is not mentioned in the document.\") is crucial. However, it is currently absent in LayoutLLM. In addition, despite achieving notable improvements through layout-aware pre-training, LayoutLLM struggles in precisely understanding region-level relationships, as evidenced in Fig. 5(a). We will study how to endow LayoutLLM with such abilities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.619, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We propose LayoutLLM for document understanding, in which a layout instruction tuning strategy comprising layout-aware pre-training and layout-aware SFT is designed to improve the comprehension of document layouts. Extensive experiments confirm the effectiveness of LayoutLLM." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15637" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.33, + 0.128 + ], + "angle": 0, + "content": "[1] GPT-4V(ision) system card. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.131, + 0.469, + 0.172 + ], + "angle": 0, + "content": "[2] Srikar Appalaraju, Bhavan Jasani, and Bhargava Urala Kota. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 4171-4186, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.175, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[4] Galal M Binmakhashen and Sabri A Mahmoud. Document layout analysis: a comprehensive survey. ACM Computing Surveys (CSUR), 52(6):1-36, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.289, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[5] Haoyu Cao, Changcun Bao, Chaohu Liu, Huang Chen, Kun Yin, Hao Liu, Yinsong Liu, Deqiang Jiang, and Xing Sun. Attention where it matters: Rethinking visual document understanding with selective region concentration. In ICCV, pages 19517-19527, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[6] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[7] Lei Cui, Yiheng Xu, Tengchao Lv, and Furu Wei. Document ai: Benchmarks, models and applications. arXiv preprint arXiv:2111.08609, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[8] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[9] Cheng Da, Peng Wang, and Cong Yao. Multi-granularity prediction with learnable fusion for scene text recognition. arXiv preprint arXiv:2307.13244, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In ECCV, pages 280–296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.633, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.771 + ], + "angle": 0, + "content": "[12] Jieuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Nikolaos Barmpalios, Rajiv Jain, Ani Nenkova, and Tong Sun. Unified pretraining framework for document understanding. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[13] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xlayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In CVPR, pages 4583-4592, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[14] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In ICDAR, pages 991-995. IEEE, 2015. 6" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. ICCV, 2023. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[16] Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents. In AAAI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[17] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM Multimedia, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[18] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 competition on scanned receipt OCR and information extraction. In ICDAR. IEEE, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[19] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[20] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In ECCV, pages 498- 517.Springer,2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. NeurIPS, 35:22199-22213, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[22] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In ICML, pages 18893–18912. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[23] Chenliang Li, Bin Bi, and Ming Yan. StructuralLM: Structural pre-training for form understanding. In ACL, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[24] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[25] Peizhao Li, Jiuxiang Gu, and Jason Kuen. SelfDoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[26] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. Structext: Structured text understanding with multimodal transformers. In ACM Multimedia, pages 1912-1920, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[27] Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE transactions on pattern analysis and machine intelligence, 45(1):919-931, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15638" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Yang Liu, Biao Yang, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[31] Chuwei Luo, Guozhi Tang, Qi Zheng, Cong Yao, Lianwen Jin, Chenliang Li, Yang Xue, and Luo Si. Bi-vldoc: Bidirectional vision-language modeling for visually-rich document understanding. arXiv preprint arXiv:2206.13155, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.262 + ], + "angle": 0, + "content": "[32] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. Geolayoutm: Geometric pre-training for visual information extraction. In CVPR, pages 7092-7101, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.304 + ], + "angle": 0, + "content": "[33] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.47, + 0.361 + ], + "angle": 0, + "content": "[34] Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Hailey Schoelkopf, Riley Kong, Xiangru Tang, et al. Fetaqa: Freeform table question answering. TACL, 10:35-49, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.363, + 0.469, + 0.39 + ], + "angle": 0, + "content": "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.469, + 0.419 + ], + "angle": 0, + "content": "[36] R OpenAI. Gpt-4 technical report. arXiv, pages 2303-08774, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.422, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[37] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. {CORD}: A consolidated receipt dataset for post-{ocr} parsing. In Workshop on Document Intelligence at NeurIPS 2019, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.492, + 0.469, + 0.575 + ], + "angle": 0, + "content": "[38] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. In EMNLP findings, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.646 + ], + "angle": 0, + "content": "[39] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Jiaqi Mu, Hao Zhang, and Nan Hua. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.648, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[40] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A large human-annotated dataset for document layout segmentation. In SIGKDD. ACM, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[41] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Płka. Going full-tilt boogie on document understanding with text-image layout transformer. In ICDAR, pages 732-747. Springer, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for train" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ing next generation image-text models. NeurIPS, 35:25278-25294, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[44] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[45] Yongxin Shi, Dezhi Peng, Wenhui Liao, Zening Lin, Xinhong Chen, Chongyu Liu, Yuyi Zhang, and Lianwen Jin. ExploringOCR capabilities of GPT-4V(ision): A quantitative and in-depth evaluation. arXiv preprint arXiv:2310.16809, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[47] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.367, + 0.892, + 0.435 + ], + "angle": 0, + "content": "[48] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.438, + 0.892, + 0.521 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.524, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[51] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance text-Layout interaction in multi-modal pre-training for document understanding. In ACL, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[52] Jiapeng Wang, Lianwen Jin, and Kai Ding. Lilt: A simple yet effective language-independent layout transformer for structured document understanding. In ACL, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15639" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.147 + ], + "angle": 0, + "content": "[53] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[54] Yiheng Xu, Minghao Li, Lei Cui, and Shaohan Huang. LayoutLM: Pre-training of text and layout for document image understanding. In KDD, pages 1192-1200, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[55] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[56] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with GPT-4V(ision). arXiv preprint arXiv:2309.17421, 9, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.32, + 0.468, + 0.387 + ], + "angle": 0, + "content": "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.39, + 0.468, + 0.471 + ], + "angle": 0, + "content": "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.474, + 0.468, + 0.541 + ], + "angle": 0, + "content": "[59] Yuechen Yu, Yulin Li, Chengquan Zhang, Xiaogiang Zhang, Zengyuan Guo, Xiameng Qin, Kun Yao, Junyu Han, Errui Ding, and Jingdong Wang. Structextv2: Masked visual-textual prediction for document image pre-training. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.545, + 0.468, + 0.599 + ], + "angle": 0, + "content": "[60] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.602, + 0.468, + 0.668 + ], + "angle": 0, + "content": "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.468, + 0.712 + ], + "angle": 0, + "content": "[62] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.715, + 0.468, + 0.755 + ], + "angle": 0, + "content": "[63] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.758, + 0.468, + 0.825 + ], + "angle": 0, + "content": "[64] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.828, + 0.468, + 0.895 + ], + "angle": 0, + "content": "[65] Stepan Simsa, Milan Sulc, Michal Uričář, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyás Skalický, Jiří Matas, Antoine Doucet, Mickaël Courstaty, and Dimosthenis Karatzas. Docile benchmark for document information localization and extraction, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15640" + } + ] +] \ No newline at end of file diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_origin.pdf b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c4565a44861b41e1591f3c4eaf9e7eae5b5e55a --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/a3bb284c-623e-4013-87c6-088a22a81e99_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47313f9d7de86483165404db6507e728c2f506cad2005711a1514f8c2dceb276 +size 2563587 diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/full.md b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28ca03d29ae0a902dca5fb5f549697a9316fd8e1 --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/full.md @@ -0,0 +1,352 @@ +# LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding + +Chuwei Luo $^{1*}$ , Yufan Shen $^{12*}$ , Zhaoqing Zhu $^{1*}$ , Qi Zheng $^{1}$ , Zhi Yu $^{2}$ , Cong Yao $^{1}$ , Alibaba Group, Zhejiang University + +{luochuwei,zzhaoqing.z,zhengqisjtu,yaocong2010}@gmail.com + +{syficy,yuzhirenzhe}@zju.edu.cn + +# Abstract + +Recently, leveraging large language models (LLMs) or multimodal large language models (MLLMs) for document understanding has been proven very promising. However, previous works that employ LLMs/MLLMs for document understanding have not fully explored and utilized the document layout information, which is vital for precise document understanding. In this paper, we propose LayoutLLM, an LLM/MLLM based method for document understanding. The core of LayoutLLM is a layout instruction tuning strategy, which is specially designed to enhance the comprehension and utilization of document layouts. The proposed layout instruction tuning strategy consists of two components: Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. To capture the characteristics of document layout in Layout-aware Pre-training, three groups of pretraining tasks, corresponding to document-level, region-level and segment-level information, are introduced. Furthermore, a novel module called layout chain-of-thought (LayoutCoT) is devised to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers. LayoutCoT is effective for boosting the performance of document understanding. Meanwhile, it brings a certain degree of interpretability, which could facilitate manual inspection and correction. Experiments on standard benchmarks show that the proposed LayoutLLM significantly outperforms existing methods that adopt open-source 7B LLMs/MLLMs for document understanding. + +# 1. Introduction + +Document AI [7], including its document understanding tasks such as document VQA [33, 47] and document visual information extraction [18, 19, 37], is currently a hot topic in both academia and industry. In recent years, document pre-trained models [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 54, 55, 59] have achieved excellent performance in doc + +![](images/6023a6bc2d8f1d903a3ed2d3637c8de18a88b928f5b5c37b27166e39f321bec8.jpg) +(1)LLMs + +![](images/9950b96af1868905b0333d6c074d8b39ed0765fb45e4606715ef7ffe55314596.jpg) +(II)MLMs + +![](images/930f4118ac7aca512809d2f20097a1a27dfa7de897d6c4277681788fb22f12bd.jpg) +(III) LayoutLLM (Ours) + +![](images/6e0a4723496339970e0d87b8bfbf5942039012f5a0ffced855ac97b6c87e36d0.jpg) +(b) Training paradigms + +![](images/a748fae26f0c2f517931304e7719f8fe618cbcf011533e653ad0141c8e1870d6.jpg) +(a) Model architectures + +![](images/6d03aba8ef4cf57671481d86956f799803e8d1c835f732fd97f88171320e3dc7.jpg) +Figure 1. LLMs/MLLMs for document understanding. The LayoutLLM is an LLM/MLLM based method that integrates a document pre-trained model as encoder. It is trained by the newly proposed layout instruction tuning strategy which consists of Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. + +ument AI downstream tasks. However, due to the necessity for fine-tuning on corresponding downstream task data, it is challenging to directly adapt such pre-trained models for zero-shot document understanding. In this paper, zero-shot refers to not using training sets of downstream tasks. + +Recently, large language models (LLMs) such as ChatGPT [35] and LLaMA [49, 50], or multimodal large language models (MLLMs) like GPT-4V [1, 36, 56], have shown remarkable zero-shot capabilities across various applications. For Document AI, as shown in Fig. 1 (a), (I) directly prompting LLMs with document text [15, 39] and (II) training document-based MLLMs [3, 57, 60] have also achieved promising results under the zero-shot setting [3, 39, 57, 60]. + +It is widely accepted that document layout information is vital for document understanding [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. However, it is difficult to convey document layout information by directly feeding text to LLMs. As Fig. 1(a)(I) shows, representing documents as either flattened plain text or layout text such as text with coordinates [15, 39, 44, 64] is often used for LLMs. + +Flattened plain text completely excludes any layout information of the document [54]. Additionally, as Tab. 1 shows, using layout text that represents both textual and layout information as inputs for LLMs does not guarantee LLMs can effectively comprehend this formatted text. + +Moreover, existing works that employ MLLMs for document understanding also have not fully explored the document layout information. Document-based MLLMs integrate visual models [11, 42] with LLM [48-50, 61] for document understanding. As Fig. 1(b) shows, they are typically based on pre-training and supervised fine-tuning (SFT) on document datasets. In the pre-training stage, tasks such as image captioning [29, 57, 60] or generating all text in a document as flattened plain text [9, 27, 60] are commonly applied. Both these image captions [43, 46] and plain text only provide a brief representation and fail to capture the layout information of the document. So it is difficult for the model to learn document layout in the existing pre-training stage. In the SFT stage, document-related VQA or information extraction data [3, 57] is often used. The answers are directly provided during SFT, lacking explicit learning about document layout. In summary, current approaches using plain or layout text to prompt LLMs and training document-based MLLMs have not effectively captured layout information, limiting their zero-shot document understanding capability. Therefore, for better document understanding with the power of LLMs, it is necessary to investigate how to effectively incorporate layout information into LLMs. + +To this end, we propose LayoutLLM, an LLM/MLLM based method for document understanding, in which a layout instruction tuning strategy is designed to enhance the comprehension of document layouts. Different from the existing MLLMs that use a general visual pre-trained model [11, 42] as the encoder, we integrate document pretrained models [2, 12, 16, 17, 23, 25, 32, 52, 54, 55] as the encoder in order to better leverage the model's foundational understanding capability for documents. The proposed layout instruction tuning consists of two stages: layout-aware pre-training and layout-aware supervised fine-tuning (SFT). Due to the complex nature of documents in real-world scenarios, encompassing rich textual content and diverse layout structures, achieving a thorough understanding involves not only capturing the document's fundamental content at global but also delving into local details. In the layout-aware pre-training stage, to ensure the model learns not only the global information of documents but also detailed information at different hierarchical levels, three groups of different level pre-training tasks are proposed: document-level, region-level, and segment-level. All the proposed pretraining tasks are unified in the format of instruction tuning. + +Furthermore, in the layout-aware SFT stage, to enhance the model's comprehension and utilization of layout information for question answering, a novel strategy called + +LayoutCoT is proposed, motivated by the chain-of-thought (CoT) [21, 53] ability in LLMs. Unlike existing methods that are directly supervised by the answer to the document understanding question, LayoutCoT consists of three successive steps: Question Analysis, Relevant Area Concentration, and Answer Formation. Through these steps, the model gains a deeper understanding of the questions, becomes capable of focusing the relevant areas instead of searching answers in the entire document and can leverage the specific characteristics of identified areas (such as tables, paragraphs, etc.) to accurately infer the answers. It not only brings a certain degree of interpretability, but also provides a feasible way for manual intervention or correction of model results. Extensive zero-shot experiments on five widely-used document understanding benchmarks demonstrate the effectiveness of the proposed LayoutLLM. + +Our contributions are summarized as follows: + +1) To better learn document layouts from global to local in layout-aware pre-training, three groups of different level pre-training tasks, which are all implemented through instruction tuning, are proposed. +2) A novel LayoutCoT strategy is proposed to achieve layout-aware supervised fine-tuning. It enables Layout-LLM to focus on the relevant document area and leverage the region's features to generate accurate answers, exhibiting a certain degree of interpretability. +3) Experimental results on zero-shot document understanding tasks show that the proposed LayoutLLM significantly outperforms existing methods that adopt LLMs/MLLMs for document understanding, demonstrating the great potential of document layout modeling. + +# 2. Related Works + +Pre-trained models for document understanding. Document pre-trained models have demonstrated the effectiveness of layout information in document understanding [2, 5, 8, 10, 12, 13, 16, 17, 20, 22, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. As a pioneer, LayoutLM [54] is the first to encode spatial coordinates of text for layout representation learning in pre-training documents. The following works [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 55, 59] then joint text, layout and images in document pre-training by combining visual models as document image encoders with the text and layout transformers, and various works [5, 10, 20, 22] start to explore pre-training end-to-end models for document understanding. These studies have achieved significant advancements in document understanding by exploring various model architectures [2, 8, 10, 12, 13, 20, 25, 26, 38, 41, 52, 59] and attention mechanisms [16, 17, 55] for modeling layout information. These methods also have proposed layout-related pre-training tasks that have been demonstrated to be highly effective in document understanding tasks. For instance, + +tasks like masked vision-language modeling [13, 17, 55], where the model is required to generate the original text corresponding to the randomly masked text in the document; position masking [31, 51], involving the randomly position masking and subsequent recovery of position information in the document; geometric pre-training [26, 32], focusing on learning direction, distance, etc.; and layout-aware generation tasks [5, 22], aiming to make the model generate structured text with layout information. However, due to the necessity of fine-tuning with annotated data for downstream tasks, these efforts face challenges in extending to zero-shot document understanding. + +LLMs/MLLMs for document understanding. Recently, LLMs like ChatGPT [35] and MLLMs like GPT-4 [36, 56] have demonstrated remarkable zero-shot performance across a wide range of NLP/CV tasks. Leveraging LLMs/MLLMs for zero-shot document understanding has also shown promising progress [3, 30, 39, 45, 57, 60]. Perot et al. [39] explore the use of LLMs for document visual information extraction, emphasizing the importance of the document layout. LLaVAR [60] which extends LLaVA [28, 29] to the document domain is pre-trained by generating plain text in the document image. During SFT, it is trained by document-related instructions which are generated by GPT-4. Expanding on mPLUG-Owl [58], mPLUG-DocOwl [57] is trained using publicly available datasets for document understanding. It includes tasks like document-level image captioning, direct information extraction, and direct document VQA. Moreover, Qwen-VL [3] proposes a general MLLM that performs well on document understanding tasks, utilizing document-level pre-training and direct VQA for SFT. Though existing LLMs/MLLMs have shown promising results in document understanding, their limited focus on document layout in pre-training and SFT has hindered their ability to achieve higher accuracy in zero-shot document understanding and better interpretability. + +# 3. LayoutLLM + +LayoutLLM is an LLM/MLLM based method that incorporates document pre-trained models for document understanding. To enhance the document layout comprehension in LayoutLLM, a novel layout instruction tuning strategy is proposed, which consists of two stages: layout-aware pretraining and layout-aware supervised fine-tuning (SFT). + +# 3.1. Model Architecture + +The overall architecture of LayoutLLM is shown in Fig. 2. In LayoutLLM, given an input document image and its corresponding text and layout information, the document pre-trained model encoder is required to obtain the multimodal document features. Then, these features are encoded by multimodal projectors, and together with the instruction embeddings, fed into the LLM to generate the final results. + +![](images/38e33a72c3a2e8f530d4ef31b412f7eb52b03c6a13c3fc7a3cc5dc0dff274d67.jpg) +Figure 2. Overall architecture of LayoutLLM. + +Document pre-trained model encoder. To leverage the foundational document comprehension capability of document pre-trained models, in this work, we utilize LayoutLMv3 [17], a widely-used document pre-training model, as our basic document encoder. The document image, text, and layout are initially inputted into the document pretrained model (DocPTM). They are then encoded by the DocPTM to obtain the corresponding features as follows: + +$$ +F _ {V}, F _ {T} = \operatorname {D o c P T M} (V, T, B o x) \tag {1} +$$ + +where $V$ represents the document image, $T = T_{0:n - 1}$ and $Box = Box_{0:n - 1}$ indicate the text sequences in the document and their corresponding bounding-box coordinates respectively. After being encoded by the DocPTM, the visual features of the document $F_{V} = F_{V_{0}:V_{m - 1}}\in \mathbb{R}^{d_{0}}$ and the text layout features $F_{T} = F_{T_{0}:T_{n - 1}}\in \mathbb{R}^{d_{0}}$ are acquired. $m$ signifies the number of visual features and $n$ represents the number of tokens contained in the document. $d_0$ denotes the dimension of DocPTM feature space. + +Multimodal projectors. To project multi-modality features from DocPTM into the LLM's embedding space, inspired by the simple yet effective projector design in LLaVA [28, 29], two different Multilayer Perceptrons (MLPs) are used as visual projector and text & layout projector respectively. Formally, the projected features can be obtained by: + +$$ +H _ {V} = P _ {V} \left(F _ {V}\right) \tag {2} +$$ + +$$ +H _ {T} = P _ {T} \left(F _ {T}\right) \tag {3} +$$ + +where $H_{V} = H_{V_{0}:V_{m - 1}}\in \mathbb{R}^{d_{1}}$ is the feature encoded by the visual projector, $H_{T} = H_{T_{0}:T_{n - 1}}\in \mathbb{R}^{d_{1}}$ is the feature encoded by the text & layout projector, and $d_{1}$ is the dimensional of the LLM embedding features. + +Large language model. Finally, the $H_V$ , $H_T$ and the embedding of the question's instruction text, $I_q = I_{q_0:l_q - 1}$ , are inputted together into the LLM, generating the target answer $I_a = I_{a_0:l_a - 1}$ . $l_q$ and $l_a$ represent the length of the question's instruction text and the answer text, respectively. + +# 3.2. Layout Instruction Tuning + +The LayoutLLM model is trained using the layout instruction tuning, which consists of two stages: layout-aware pretraining and layout-aware SFT. + +![](images/13b8337221be70f40d2a4e8849fcd0cd0072db60f067a6e54a1f07b7d6790542.jpg) +Figure 3. Overview of the Layout Instruction Tuning. (a) Document-level, region-level, and segment-level pre-training tasks, unified in instruction tuning format, are introduced. (b) A novel module called LayoutCoT is designed to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers through three intermediate steps. + +# 3.2.1 Layout-aware Pre-training + +The goal of pre-training the LayoutLLM is to enhance the model's comprehensive understanding of documents at different levels through layout learning, rather than only focusing on global document-level understanding like existing MLLM methods [57, 60]. To this end, during the pre-training stage, three different level pre-training strategies are simultaneously applied to the LayoutLLM, namely document-level, region-level, and segment-level. + +Document-level To enable the model to possess fundamental global document understanding, pre-training tasks, namely Document Dense Description (DDD) and Text and Layout Reconstruction (TLR), are proposed. As Fig. 3 (a) shows, like the image caption task, the DDD task requires the model to learn to describe the input document. Moreover, in the DDD task, the descriptions for document images are more detailed. For instance, in the document image caption data used for the LLaVAR [60] pre-training, captions contain an average of 36.27 words, while in the proposed dataset of the DDD task, the description contains an average of 373.25 words. Through the DDD task, the model can obtain basic document-level information, such as document type and detailed content. The TLR task aims to reconstruct the complete text and layout information of a document and output it in the format “<\{box\}, \{text\} >”. The TLR task aligns the text and layout embeddings output from DocPTM with the LayoutLLM's LLM space. Consequently, it enables the LLM in LayoutLLM to comprehend the text and layout information contained in the documents. + +Region-level The information contained in specific regions of a document, such as the titles, figures, tables, is essential for document understanding [4, 6, 8, 24, 62]. These regions serve as important characteristics that differentiate + +a document from plain text in natural language. For the LayoutLLM to achieve the basic region-level understanding, two pre-training tasks, namely Document Layout Analysis (DLA) and Table Understanding (TU), are utilized. The DLA task is achieved in two ways as shown in Fig. 3. One involves locating the layout region based on the layout type, the other involves identifying the type of a given area. Furthermore, the table region differs from other regions in that it requires additional focus on 2D layout understanding. The TU task enables the model to understand the basic row and column information in the table region of a document. As shown in Fig. 3, the TU task includes instruction tuning for the number of rows and columns, logical coordinates, and the content within rows and columns. + +Segment-level Early works on document pre-trained models [13, 17, 26, 31, 32, 51, 55] have demonstrated the effectiveness of segment-level document pre-training tasks to document layout understanding ability, such as masked vision language modeling (MVLM) [13, 17, 55], position masking [31, 51], and geometric pre-training [26, 32]. Inspired by these works, to make LayoutLLM have segment-level layout understanding, these tasks are transformed into instruction formats for pre-training as Fig. 3(a) shows. For the MVLM instructions, random masking of the text input to LayoutLLM is performed, and the model is instruction tuned by asking the masked words and answering them. For mask position instruction, the layout information (coordinates) to a specific text line, when input to LayoutLLM, is randomly set to 0. The instruction is constructed by asking about the text line with zeroed coordinates and requesting the model to respond with the original coordinates with text content. For geometric layout instruction, text lines are randomly selected, and an instruction is constructed by asking questions about the direction and distance between them. + +# 3.2.2 Layout-aware Supervised Fine-tuning + +In the SFT stage of existing document-based MLLMs, models are directly supervised by the answer to the document understanding instructions. Consequently, these methods lack explicit learning of document layout which is crucial for document understanding. Considering this limitation, and inspired by previous works related to chain-of-though (CoT) [21, 53], which have shown that inferencing with intermediate steps can greatly enhance performance. A novel module called LayoutCoT is proposed, which incorporates the layout information into every intermediate step of CoT explicitly. Meanwhile, by introducing the layout-aware intermediate steps, the answer process gains a certain degree of interpretability for LayoutLLM and also provides interactive correction possibilities based on LayoutCoT. + +LayoutCoT Details. As shown in Fig. 3(b), the LayoutCoT involves the following three intermediate steps: + +Step 1: Question Analysis. To effectively address a document understanding problem, analyzing the key characteristics of the question is very important. Identifying the question type, such as table understanding or entity extraction from paragraphs, and assessing whether the question is a straightforward extraction query or a more complex reasoning problem, can help guide the direction for the subsequent inference process. Therefore, to give basic guidance to the subsequent steps, the question analysis step is designed, encompassing an analysis of the question type from a layout perspective and a detailed understanding of the question itself. Benefiting from the layout understanding ability by layout-aware pre-training, this step can extract the types and key information mentioned in the question, which are related to the specific characteristics of the document. + +Step 2: Relevant Area Concentration. For most document understanding tasks, the entire document contains a large amount of irrelevant information that may confuse the model [5]. This step aims to focus on the relevant area and generate its location information, which is used to assist the model to accurately infer the answer. Benefiting from the layout information conveyed by step 1 and the positioning capabilities learned from the region&segment-level pretraining, the model can accurately generate the location of the relevant area. For example in Fig. 3(b), according to the question type "table" in step 1, the relevant "table" can be located. By guiding the model to focus on the relevant area, this step largely narrows the search scope, increasing the possibility of giving the right answer. Meanwhile, the location information provides a way for visual inspection and interactive correction (see Sec. 4.7 for details). + +Step 3: Answer Formation. Finally, the last step, the answer formation, provides explanations based on the layout characteristics of the relevant areas located in step 2 and key points analyzed in step 1 to get the final answer. For example in Fig. 3(b), for a "table" type question, this step in- + +# Algorithm 1 CONSTRUCT(D): LayoutCoT Construction. + +Definition: $\mathcal{H}$ : Document HTMLs; $\mathcal{I}$ : Document Images; $\mathcal{T}$ : MRC Texts; $\mathcal{R}$ : Document Language Representation; QA: QA pairs; $T_{c}$ : Text CoT; $\mathcal{L}_{c}$ : LayoutCoT; +Input: Document Dataset $\mathcal{D} = \{\mathcal{D}_H,\mathcal{D}_I,\mathcal{D}_M\}$ +$(\mathcal{D}_H = \{\mathcal{H}\}, \mathcal{D}_I = \{\mathcal{I}\}, \mathcal{D}_M = \{\mathcal{T}, \mathcal{Q}\mathcal{A}\})$ ; +Output: Constructed Dataset $\mathcal{D}_c$ +1: Procedure CONSTRUCT $(\mathcal{D} = \{\mathcal{D}_H,\mathcal{D}_I,\mathcal{D}_M\})$ +2: 1) $\mathcal{R} = getDocRep(\mathcal{I})$ if $\mathcal{D} \subseteq \mathcal{D}_I$ else $\mathcal{H}$ if $\mathcal{D} \subseteq \mathcal{D}_H$ else pass; +3: 2) $\mathcal{Q}\mathcal{A},\mathcal{T}_c = getQACoT(\mathcal{D})$ if $\mathcal{D}\subseteq \mathcal{D}_M$ else getQACoTGPT(R); +4: 3) $\mathcal{L}_c = getLayoutCoT(\mathcal{T}_c)$ +5: 4) if $\mathcal{D} \subseteq \{\mathcal{D}_H, \mathcal{D}_M\}$ : $\mathcal{I} = Htm12Img(\mathcal{D}_H \cap \mathcal{H} : \mathcal{T})$ ; +6: return $\mathcal{D}_c\gets \{\mathcal{I},\mathcal{Q}\mathcal{A},\mathcal{L}_c\}$ + +volves analyzing the row and column in the relevant table in step 2, and referencing the answer step-by-step. For a "key-value" question, analyzing the keywords in concentrated areas can help get the final answer. Analyzing answers in different ways based on the features of different layout regions not only improves the document understanding performance but also brings a certain level of interpretability. + +LayoutCoT Construction. Given the need for both text and image annotations in constructing LayoutCoT, manual labeling can be difficult. Algorithm 1 proposes a manual-labeling-free method, generating LayoutCoT data using public datasets with GPT (GPT-3.5 Turbo). It involves representing document text and layout in a format understandable by GPT. GPT is then utilized to generate document-content-based QA and corresponding text CoT. Finally, use rules for transforming the text CoT to LayoutCoT. + +Three types of publicly available document datasets are focused on: HTML documents $(\mathcal{D}_H)$ , image documents $(\mathcal{D}_I)$ , and text documents $(\mathcal{D}_M)$ for machine reading comprehension (MRC). The construction process is as follows: + +1) Document Representation: To fully leverage the capabilities of GPT, it is crucial to ensure that the document content fed to GPT contains accurate layout information. Since HTML is the formatted language that can represent documents accurately, $\mathcal{D}_H$ is represented using the original HTML. By transforming HTML to PDF and using the PDF parser, the text and bounding-boxes are obtained. For $\mathcal{D}_I$ , the layout-aware text [15] is used. The text and bounding-boxes in $\mathcal{D}_I$ are from the original dataset annotations. +2) QA&Text CoT Generation: The language representation $\mathcal{R}$ for the document is employed for prompting GPT to generate QA pairs $\mathcal{QA}$ with text CoTs $\mathcal{T}_c$ . In addition, the $\mathcal{D}_M$ includes the QA pairs and reasoning process, thereby directly reusing the $\mathcal{QA}$ and manually organizing $\mathcal{T}_c$ . The generated $\mathcal{T}_c$ includes the step 1 (question analysis) and step 3 (answer formation) for LayoutCoT, and locates all relevant sentences in the document for $\mathcal{QA}$ . +3) LayoutCoT Generation: The step 1 & 3 in $\mathcal{T}_c$ are used as the step 1 & 3 in $\mathcal{L}_c$ . To construct the step 2 (relevant area concentration) of $\mathcal{L}_c$ , the union bounding-box of all located relevant sentences in $\mathcal{T}_c$ are taken as the relevant area. + +4) Document Images Generation: For $\mathcal{D}_H$ and $\mathcal{D}_M$ , the HTMLs and MRC text are converted to images. Overall, the document images $\mathcal{I}$ , generated QA pairs $\mathcal{QA}$ and LayoutCoTs $\mathcal{L}_c$ constitute the final LayoutCoT dataset $\mathcal{D}_c$ . + +# 4. Experiments + +# 4.1. Dataset Collection + +Layout-aware pre-training data of LayoutLLM is from publicly available document understanding datasets. It does not incorporate any data from the training, validation, and test sets of downstream benchmarks. Region-level pretraining tasks, most document-level and segment-level tasks are self-supervised. Thus, only document images and images converted from PDFs in the original datasets, along with the corresponding OCR or text layout results from PDF parsing, are needed. For these tasks, data is randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65]. Particularly, data for document dense description is from inputting the document text content into GPT-3.5 Turbo, prompting it to generate an average of 373.25 words document dense descriptions. For the region-level tasks, specifically the document layout analysis task, publicly available document layout analysis datasets are utilized, including PubLayNet [62], DocLayNet [40], and Docbank [24]. Data for another region-level task, table understanding, is sourced from PubTabNet [63] with its table annotations. All data is transformed into the instruction format illustrated in Fig. 3(a). In total, 5.7 million instructions are constructed, with a ratio of 1:4:4 for document-level, region-level, and segment-level tasks, respectively. For detailed instruction templates and dataset descriptions, please refer to the supplementary material. + +Layout-aware SFT data of LayoutLLM is generated by GPT (GPT-3.5 Turbo) and converted from existing textual Machine Reading Comprehension (MRC) datasets, as discussed in Sec. 3.2.2. To generate high-quality document-based textual QA and textual CoT, it is essential to make GPT comprehend the document layout. So, the document is represented using both layout text [15] and HTML. Similar to the pre-training data, the $D_I$ in Algorithm 1 is also randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65] for building layout text. The $D_H$ in Algorithm 1 is from GPT's free generation. The $D_M$ in Algorithm 1 is randomly sampled from the FeTaQA [34] which is a wikipedia question answering dataset. A total of 300K instructions are constructed, with a ratio of 5:4.5:0.5 for $D_I$ , $D_H$ , and $D_M$ , respectively. For detailed prompt templates of document-based text QA and text CoT generation using GPT, prompts for HTML generation using GPT, and dataset description, please refer to the supplementary material. + +# 4.2. Training Setup + +The encoder weight of LayoutLLM is initialized from the LayoutLMv3-large [17] which is a widely-used document pre-trained model. And the LLM backbone is initialized from Vicuna-7B-v1.5 [61]. Other parameters are randomly initialized. During pre-training, the LLM is frozen, and the parameters of the two projectors and document pre-trained model encoder are updated. During SFT, both LLM and two projectors are fine-tuned while keeping the document pretrained model encoder frozen. For detailed training setup, please refer to the supplementary material. + +# 4.3. Evaluation Setup + +The zero-shot ability is highly expected in real-world document understanding scenarios [7, 30, 45]. Therefore, zero-shot evaluations on widely-used document understanding benchmarks including document visual question answering (Document VQA) and visual information extraction (VIE) are conducted. Only the test sets are utilized in all benchmarks and only the official provided image, text, and layout information are used. The Document VQA datasets comprise the DocVQA[33] test set, consisting of 5,188 questions, and the VisualMRC[47] test set containing 6,708 questions. Following the evaluation metric settings of the original datasets, the ANLS [33] is utilized for evaluating DocVQA, and Rouge-L is used for evaluating VisualMRC. For the VIE task, FUNSD [19], CORD [37], and SROIE [18] are used. The test set of FUNSD comprises 50 form images, each annotated with entity-level headers, questions, answers, and others, along with entity linking annotations. CORD's test set consists of 100 receipt images, annotated with 30 entity types, such as the tax amount, total price, etc. SROIE's test set includes 347 receipt images, annotated with 4 entity types: company, date, address, and total. To prompt LLMs/MLLMs for zero-shot VIE, annotations in VIE datasets are transformed into question answering format (QA for VIE). For key-value annotations with linking in the FUNSD, the format is {Q: What is the "key" in the document? A: "value)}. For entity annotations in CORD and SROIE, directly asking for the entity in the document like {Q: What is the address in the document? A: "the address annotation"} is utilized. Following DocVQA, the QA for VIE task is evaluated by ANLS. + +# 4.4. Main Results + +As shown in Tab. 1, the zero-shot document understanding performance of LayoutLLM and existing open-source LLMs and MLLMs is evaluated. Generally, the existing LLMs are better than MLLMs for zero-shot document VQA and VIE. For example in the results on DocVQA, most LLMs can achieve a performance of around $60\%$ or higher, while most MLLMs can only attain around $10\%$ , except for mPLUG-DocOWL and Qwen-VL that trained with the + +
ModelsDocument VQAQA for VIE
DocVQAVisualMRCFUNSDCORDSROIE
Fine-tuned PTMLayoutLMv3 [17]83.37*-92.08*‡97.46*‡-
LLMLlama2-7B [50]61.3429.7340.784.3915.86
Llama2-7B-chat [50]64.9952.8448.2047.7068.97
Plain TextVicuna-7B [61]61.3953.6349.7944.6767.49
Vicuna-1.5-7B [61]66.9952.1348.0651.4068.20
LLMLlama2-7B [50]37.3233.8251.4028.0434.96
Llama2-7B-chat [50]56.5549.2658.3450.9351.15
Layout Text (Text + Box) [15]Vicuna-7B [61]37.2152.5542.7346.5945.43
Vicuna-1.5-7B [61]56.8147.2259.6356.1366.20
MLLMLLaVAR-7B [60]11.6†36.371.7113.552.38
LLaVA-1.5-7B [28]13.3435.231.9318.063.83
mPLUG-DocOWL-7B [57]62.2*†----
Qwen-VL-7B [3]65.1*†42.5247.0930.0058.59
LayoutLLM-7B△ (Ours)74.2555.7378.6562.2170.97
LayoutLLM-7B★ (Ours)74.2755.7679.9863.1072.12
+ +training set. One possible reason is that it's difficult for these MLLMs to obtain accurate textual information from document images. Additionally, for LLMs, using Plain Text and Layout Text respectively as the document representation are further discussed, where the Layout Text introduces layout information by adding text coordinates in the format: {text:"text", box:[x1,y1,x2,y2]} [15]. Compared to the Plain Text, the Layout Text variant doesn't show stable performance improvements, noticeable in certain tasks, for example in Vicuna-1.5, an improvement in VIE (FUNSD $48.06\%$ to $59.63\%$ ) but a decline in DocVQA $(66.99\%$ to $56.81\%)$ . LLMs may lack the ability to learn this formatted layout text, and directly adding layout information (e.g., coordinates) to the text will also highly increase the token length, making the answer inference more challenging. + +Compared to the prior SOTA model, LayoutLMv3, which is fine-tuned using the training set of downstream tasks, LayoutLLM demonstrates competitive performance on the DocVQA benchmark. Compared with these LLMs and MLLMs, LayoutLLM achieves consistent and significant improvements over them on all evaluation benchmarks. Notably, LayoutLLM which employs zero-shot performance, outperforms mPLUG-DocOWL and Qwen-VL by around $10\%$ on the DocVQA dataset, both of which are trained with this dataset. This demonstrates that LayoutLLM can learn more robust and discriminative representations for document understanding. Furthermore, experiments of the different initialization of the LLM backbone have all achieved optimal results across all benchmarks, substantiating that LayoutLLM can adapt to various LLMs. In summary, our method explores a more effective way to utilize layout information for document understanding, which significantly improves the performance of zero-shot document understanding. + +Table 1. Zero-shot document understanding results on open-source LLMs and MLLMs. * signifies training set use; unmarked results are zero-shot. Results marked with ${}^{ \ddagger }$ are F1 scores for VIE. Results marked with $\dagger$ are from the original paper and others are re-implemented by us. $\bigtriangleup$ marks LayoutLLM's LLM backbone as initialized with Llama2-7B-chat,and $\star$ with Vicuna-1.5-7B. + +
#Layout-aware +Pre-trainingLayout-aware +SFTDocVQAFUNSD
070.8270.96
1✓72.3174.02
2✓✓74.2779.98
+ +Table 2. Ablation study on the DocVQA and FUNSD test sets. + +# 4.5. Ablation study + +To better verify the effectiveness of the layout-aware pretraining and layout-aware SFT in the layout instruction tuning, an ablation study is conducted (see Tab. 2). + +Initial baseline. The #0 baseline disables both layout-aware pre-training and layout-aware SFT. It only adopts SFT (same SFT data but without LayoutCoT steps) on the LayoutLLM. Even without any alignment pre-training and LayoutCoT steps guidance, the baseline outperforms previous SOTAs, achieving $70.82\%$ on DocVQA and $70.96\%$ on FUNSD. This indicates the document understanding ability of DocPTM benefits document understanding with LLMs. Effect of Layout-aware pre-training. In #1, the layout-aware pre-training and the same SFT with #0 is conducted. Compared to #0, benefiting from the basic document understanding capability learned through the layout-aware pretraining, #1 shows an increase of $1.49\%$ on DocVQA and $3.06\%$ on FUNSD. It can be observed that the basic document understanding ability learned from the layout-aware pre-training significantly enhances the performance of the basic key-value extraction tasks in the FUNSD dataset. Compared to the FUNSD VIE task, it also shows that the DocVQA is a more complex task. Effect of Layout-aware SFT. Compared with #1, #2 further incorporates the layout-aware SFT strategy, resulting in a performance gain of $1.96\%$ on DocVQA and even $5.96\%$ on FUNSD. This indicates the LayoutCoT in layout-aware SFT can help the LayoutLLM to handle complex document tasks and it is effic + +![](images/2c085931ced6d4bf149854cf8647c6ef4f5d296c89fbdf02912e98b6fdafc124.jpg) +Question: What is the Income Tax Withheld for the month of July 1954? + +![](images/ce4b7d257b1d72b26e1b07fd7edbcb9f66bf7af5fda50855d64dcf90090b89b7.jpg) + +![](images/b336250946d452584eaa12662d0d52a3dd13d5966a47a782dda79845f69f9589.jpg) + +![](images/097a6af6bd5cb4f86bd9ad43ac2415512d63639b7e8a071980726536d774504e.jpg) +Question: What is the Total for Exit 2004? + +![](images/90fd5a4718cf92ef8920d9ede940eea486e3a53c844d7738b74cdbf619e76b94.jpg) +Figure 4. Qualitative results on DocVQA. Green boxes are the areas concentrated in the step 2 of LayoutCoT. + +![](images/72b69f819c37d11880391fc1b4164a638f805e8c984d634efc4c474e23d6d3f4.jpg) + +![](images/3b107afefe8faf6e0623f542bbfff496f92c4926ed48bfd822b1f78a84927f63.jpg) + +![](images/683193296cabd548f6372c9f5491b9a70051cf08d79d1ecaaf247a5ad245bb74.jpg) + +tive for boosting the performance on both document VQA and VIE. Moreover, it also promotes a certain degree of interpretability. Overall, the ablation study demonstrates the importance of layout-aware pre-training and layout-aware SFT for zero-shot document understanding. + +# 4.6. Qualitative Results + +Two examples are shown in Fig. 4. Through combined with layout-aware pre-training and layout-aware SFT, Layout-LLM can accurately focus on the relevant areas, utilize the layout information to assist in problem-solving and provide interpretability. For example, in Fig. 4(a) question about key-value extraction in up-down layout, different from the left-right variant, relies more on document layout to infer the right answer. Since the keywords "Income Tax" in the question often co-occur with numerical data, Vicuna-1.5 and Qwen-VL find numerical answers relying more on the semantics than the layout, resulting in incorrect responses. In contrast, benefiting from the layout-aware pre-training, our model can effectively leverage layout information to give accurate answers. In addition, the model using LayoutCoT can further provide the location and the reasoning process, showing a certain degree of interpretability. But in certain situations, only combined with layout pre-training, our model might fail to give accurate answers. As shown in Fig. 4(b), without LayoutCoT, our model identifies "Exit 2003" as the relevant column and generates a wrong answer. However, with the help of LayoutCoT, LayoutLLM can correctly identify the question type as "Table", locate the relevant table area, and finally infer the right answer from the corresponding "Exit 2004" column. + +# 4.7. Interactive Correction with LayoutCoT + +Since LayoutCoT runs in a step-by-step fashion and produces intermediate results in the inference stage, it can facilitate interactive inspection and correction, when processing a document. As shown in Fig. 5(a), there are two areas + +![](images/8641010ee6173b2ae131e22ffec8f9ff88ea4049185be5ac26e93b051f28a303.jpg) +Question: In the table given on the left bottom of +Step 1 (Question Analysis): The question is about key-value extraction. And the question asks for the 'vendor'. +Step 2 (Relevant Area Concentration): The answer can be found +in the area: [28, 123, 317, 148]. +Step 3 (Answer Formation): The number entered for 'vendor' in the text is: +Step 1 (Question Analysis): +Step 2 (Relevant Area Concentration): The answer can be found in the text: [46,703,467,830]. +Step 3 (Answer Formation): The number entered for 'vendor' in +the table is 013684 +Final answer: 013684. + +Question: What is the heading of second paragraph? + +![](images/671c37c21e3f401282dd5824acdf3d36cac071b217411f43cd15690c94a9bf77.jpg) +Table 1 Table 2 + +![](images/fe3bb4ef182a5215bc43bae56bf3f48b87db9051de41450fc76c96cf3f7f0ba0.jpg) + +![](images/01e428f05a981fa67b4f30d0c169b7b5a6b2b29167a7010ff79388567c85968f.jpg) + +![](images/7d9ad74267586c88775852aa6428a6c2d49a534a5a18b72175f0b83117d0c48b.jpg) + +![](images/6a39ea0d7bcfd074e9a0c8e0da7fe9e03ffd767ac2a6e6a69682e06d31f397a3.jpg) +the page, what is it? +GT Answer: 013684 +GT Answer: PRELIMINARY STATEMENT +GT Answer: 013684 +Step 1 (Question Analysis): The question is about Factoid. And the question asks for the heading of the second paragraph... Step 2 (Relevant Area Concentration): The answer can be found in the area: 134, 280, 654, 3731. +Step 3 (Answer Formation): The heading of the second +paragraph is \*TAKEDA'S RESPONSE AND OBJECTIONS TO +PRACTITIONER'S SUPPLEMENTAL INJURY CLAIMS. ADDITIONAL ACTOS PERSONAL INJURY CLAIMS*. +Final answer: TAXEDA'S RESPONSE AND OBJECTIONS TO +□ +Step 1 (Question Analysis): +Step 2 (Relevant Area Concentration): The answer can be found +in the area: [117, 566, 678, 681]. +Step 3 (Answer Formation): The heading of the second paragraph is "PRESUMMARY STATEMENT". +Final answer: PRELIMINARY STATEMENT +tion with LayoutCoT. Green represents and answers, while Red represents the list viewed in digital version. + +Question: What is the heading of second paragraph? + +GT Answer: PRELIMINARY STATEMENT + +![](images/55eab25f39acc79a21497930688dbd25324e6abce888c915d9c89a42c6238ec4.jpg) +PLAINTIPS SUPPLEMENTAL INTERROGATORY 2 REGRADING ADDITIONAL ACTS PERONEAL INJURY CLAIMS + +in the image that are relevant to the keyword "vendor" of the question. LayoutCoT focused on a wrong area containing "vendor", as it missed "left bottom" in the question, the answer was incorrect. However, after the right area is given manually, it can finally give the correct answer. Similarly, in Fig. 5(b) the question asks about "the heading of the second paragraph". However, the term "paragraph" does not have a universal definition, and in this case, the sentences in the area below the main heading were considered as a paragraph, causing the model to predict "TAKEDA's...CLAIMS", which was incorrect according to the GT. Once the right "second paragraph" region is fed to the model, the answer can be successfully revised. This unique ability of LayoutCoT could be very valuable in high-stake scenarios (e.g., a bank transaction), where the standards are extremely high, and manual checking and correction (i.e., human-in-the-loop) are required. + +# 5. Limitations + +Through LayoutCoT, LayoutLLM demonstrates the capability of interactive correction, but in real-world applications, this is not enough. The ability to refuse false-positive outputs and generate hints (e.g. "The answer is not mentioned in the document.") is crucial. However, it is currently absent in LayoutLLM. In addition, despite achieving notable improvements through layout-aware pre-training, LayoutLLM struggles in precisely understanding region-level relationships, as evidenced in Fig. 5(a). We will study how to endow LayoutLLM with such abilities. + +# 6. Conclusion + +We propose LayoutLLM for document understanding, in which a layout instruction tuning strategy comprising layout-aware pre-training and layout-aware SFT is designed to improve the comprehension of document layouts. Extensive experiments confirm the effectiveness of LayoutLLM. + +# References + +[1] GPT-4V(ision) system card. 2023. 1 +[2] Srikar Appalaraju, Bhavan Jasani, and Bhargava Urala Kota. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 4171-4186, 2021. 1, 2 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3, 7 +[4] Galal M Binmakhashen and Sabri A Mahmoud. Document layout analysis: a comprehensive survey. ACM Computing Surveys (CSUR), 52(6):1-36, 2019. 4 +[5] Haoyu Cao, Changcun Bao, Chaohu Liu, Huang Chen, Kun Yin, Hao Liu, Yinsong Liu, Deqiang Jiang, and Xing Sun. Attention where it matters: Rethinking visual document understanding with selective region concentration. In ICCV, pages 19517-19527, 2023. 2, 3, 5 +[6] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 4 +[7] Lei Cui, Yiheng Xu, Tengchao Lv, and Furu Wei. Document ai: Benchmarks, models and applications. arXiv preprint arXiv:2111.08609, 2021. 1, 6 +[8] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 1, 2, 4 +[9] Cheng Da, Peng Wang, and Cong Yao. Multi-granularity prediction with learnable fusion for scene text recognition. arXiv preprint arXiv:2307.13244, 2023. 2 +[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In ECCV, pages 280–296. Springer, 2022. 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[12] Jieuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Nikolaos Barmpalios, Rajiv Jain, Ani Nenkova, and Tong Sun. Unified pretraining framework for document understanding. In NeurIPS, 2021. 1, 2 +[13] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xlayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In CVPR, pages 4583-4592, 2022. 1, 2, 3, 4 +[14] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In ICDAR, pages 991-995. IEEE, 2015. 6 + +[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. ICCV, 2023. 1, 5, 6, 7 +[16] Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents. In AAAI, 2022. 1, 2 +[17] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM Multimedia, 2022. 1, 2, 3, 4, 6, 7 +[18] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 competition on scanned receipt OCR and information extraction. In ICDAR. IEEE, 2019. 1, 6 +[19] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents, 2019. 1, 6 +[20] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In ECCV, pages 498- 517.Springer,2022.2 +[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. NeurIPS, 35:22199-22213, 2022. 2, 5 +[22] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In ICML, pages 18893–18912. PMLR, 2023. 2, 3 +[23] Chenliang Li, Bin Bi, and Ming Yan. StructuralLM: Structural pre-training for form understanding. In ACL, 2021. 1, 2 +[24] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 4, 6 +[25] Peizhao Li, Jiuxiang Gu, and Jason Kuen. SelfDoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 1, 2 +[26] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. Structext: Structured text understanding with multimodal transformers. In ACM Multimedia, pages 1912-1920, 2021. 1, 2, 3, 4 +[27] Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE transactions on pattern analysis and machine intelligence, 45(1):919-931, 2022. 2 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 7 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3 + +[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Yang Liu, Biao Yang, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 3, 6 +[31] Chuwei Luo, Guozhi Tang, Qi Zheng, Cong Yao, Lianwen Jin, Chenliang Li, Yang Xue, and Luo Si. Bi-vldoc: Bidirectional vision-language modeling for visually-rich document understanding. arXiv preprint arXiv:2206.13155, 2022. 3, 4 +[32] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. Geolayoutm: Geometric pre-training for visual information extraction. In CVPR, pages 7092-7101, 2023. 1, 2, 3, 4 +[33] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 1, 6 +[34] Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Hailey Schoelkopf, Riley Kong, Xiangru Tang, et al. Fetaqa: Freeform table question answering. TACL, 10:35-49, 2022. 6 +[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1, 3 +[36] R OpenAI. Gpt-4 technical report. arXiv, pages 2303-08774, 2023. 1, 3 +[37] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. {CORD}: A consolidated receipt dataset for post-{ocr} parsing. In Workshop on Document Intelligence at NeurIPS 2019, 2019. 1, 6 +[38] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. In EMNLP findings, 2022. 1, 2 +[39] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Jiaqi Mu, Hao Zhang, and Nan Hua. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 1, 3 +[40] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A large human-annotated dataset for document layout segmentation. In SIGKDD. ACM, 2022. 6 +[41] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Płka. Going full-tilt boogie on document understanding with text-image layout transformer. In ICDAR, pages 732-747. Springer, 2021. 1, 2 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2 +[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for train + +ing next generation image-text models. NeurIPS, 35:25278-25294, 2022. 2 +[44] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1 +[45] Yongxin Shi, Dezhi Peng, Wenhui Liao, Zening Lin, Xinhong Chen, Chongyu Liu, Yuyi Zhang, and Lianwen Jin. ExploringOCR capabilities of GPT-4V(ision): A quantitative and in-depth evaluation. arXiv preprint arXiv:2310.16809, 2023. 3, 6 +[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758. Springer, 2020. 2 +[47] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 1, 6 +[48] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.2 +[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[50] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 1, 2, 7 +[51] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance text-Layout interaction in multi-modal pre-training for document understanding. In ACL, 2023. 3, 4 +[52] Jiapeng Wang, Lianwen Jin, and Kai Ding. Lilt: A simple yet effective language-independent layout transformer for structured document understanding. In ACL, 2022. 1, 2 + +[53] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 2, 5 +[54] Yiheng Xu, Minghao Li, Lei Cui, and Shaohan Huang. LayoutLM: Pre-training of text and layout for document image understanding. In KDD, pages 1192-1200, 2020. 1, 2 +[55] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 1, 2, 3, 4 +[56] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with GPT-4V(ision). arXiv preprint arXiv:2309.17421, 9, 2023. 1, 3 +[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 1, 2, 3, 4, 7 +[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3 +[59] Yuechen Yu, Yulin Li, Chengquan Zhang, Xiaogiang Zhang, Zengyuan Guo, Xiameng Qin, Kun Yao, Junyu Han, Errui Ding, and Jingdong Wang. Structextv2: Masked visual-textual prediction for document image pre-training. In ICLR, 2023. 1, 2 +[60] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 3, 4, 7 +[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2, 6, 7 +[62] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 4, 6 +[63] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 6 +[64] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 1 +[65] Stepan Simsa, Milan Sulc, Michal Uričář, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyás Skalický, Jiří Matas, Antoine Doucet, Mickaël Courstaty, and Dimosthenis Karatzas. Docile benchmark for document information localization and extraction, 2023. 6 \ No newline at end of file diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/images.zip b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..18b04a1794e169e36b87e317e29ea130b6877305 --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c95e23ab115742425dbd47aee34eebf377511d8515e8944163f0d53b078a1ced +size 410696 diff --git a/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/layout.json b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59a8e1d2bf81754a4afd002b9e4b4ebf1010810f --- /dev/null +++ b/2024/LayoutLLM_ Layout Instruction Tuning with Large Language Models for Document Understanding/layout.json @@ -0,0 +1,10336 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 523, + 140 + ], + "type": "text", + "content": "LayoutLLM: Layout Instruction Tuning with Large Language Models for Document Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": "Chuwei Luo" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Yufan Shen" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Zhaoqing Zhu" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Qi Zheng" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Zhi Yu" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Cong Yao" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Alibaba Group, Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 191, + 451, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 191, + 451, + 203 + ], + "spans": [ + { + "bbox": [ + 142, + 191, + 451, + 203 + ], + "type": "text", + "content": "{luochuwei,zzhaoqing.z,zhengqisjtu,yaocong2010}@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 205, + 380, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 205, + 380, + 217 + ], + "spans": [ + { + "bbox": [ + 212, + 205, + 380, + 217 + ], + "type": "text", + "content": "{syficy,yuzhirenzhe}@zju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 270, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 290, + 594 + ], + "type": "text", + "content": "Recently, leveraging large language models (LLMs) or multimodal large language models (MLLMs) for document understanding has been proven very promising. However, previous works that employ LLMs/MLLMs for document understanding have not fully explored and utilized the document layout information, which is vital for precise document understanding. In this paper, we propose LayoutLLM, an LLM/MLLM based method for document understanding. The core of LayoutLLM is a layout instruction tuning strategy, which is specially designed to enhance the comprehension and utilization of document layouts. The proposed layout instruction tuning strategy consists of two components: Layout-aware Pre-training and Layout-aware Supervised Fine-tuning. To capture the characteristics of document layout in Layout-aware Pre-training, three groups of pretraining tasks, corresponding to document-level, region-level and segment-level information, are introduced. Furthermore, a novel module called layout chain-of-thought (LayoutCoT) is devised to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers. LayoutCoT is effective for boosting the performance of document understanding. Meanwhile, it brings a certain degree of interpretability, which could facilitate manual inspection and correction. Experiments on standard benchmarks show that the proposed LayoutLLM significantly outperforms existing methods that adopt open-source 7B LLMs/MLLMs for document understanding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 128, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 621, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 287, + 693 + ], + "type": "text", + "content": "Document AI [7], including its document understanding tasks such as document VQA [33, 47] and document visual information extraction [18, 19, 37], is currently a hot topic in both academia and industry. In recent years, document pre-trained models [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 54, 55, 59] have achieved excellent performance in doc" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 244, + 378, + 304 + ], + "blocks": [ + { + "bbox": [ + 309, + 244, + 378, + 304 + ], + "lines": [ + { + "bbox": [ + 309, + 244, + 378, + 304 + ], + "spans": [ + { + "bbox": [ + 309, + 244, + 378, + 304 + ], + "type": "image", + "image_path": "6023a6bc2d8f1d903a3ed2d3637c8de18a88b928f5b5c37b27166e39f321bec8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 307, + 352, + 313 + ], + "lines": [ + { + "bbox": [ + 332, + 307, + 352, + 313 + ], + "spans": [ + { + "bbox": [ + 332, + 307, + 352, + 313 + ], + "type": "text", + "content": "(1)LLMs" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 384, + 245, + 457, + 304 + ], + "blocks": [ + { + "bbox": [ + 384, + 245, + 457, + 304 + ], + "lines": [ + { + "bbox": [ + 384, + 245, + 457, + 304 + ], + "spans": [ + { + "bbox": [ + 384, + 245, + 457, + 304 + ], + "type": "image", + "image_path": "9950b96af1868905b0333d6c074d8b39ed0765fb45e4606715ef7ffe55314596.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 307, + 430, + 313 + ], + "lines": [ + { + "bbox": [ + 408, + 307, + 430, + 313 + ], + "spans": [ + { + "bbox": [ + 408, + 307, + 430, + 313 + ], + "type": "text", + "content": "(II)MLMs" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 463, + 245, + 533, + 304 + ], + "blocks": [ + { + "bbox": [ + 463, + 245, + 533, + 304 + ], + "lines": [ + { + "bbox": [ + 463, + 245, + 533, + 304 + ], + "spans": [ + { + "bbox": [ + 463, + 245, + 533, + 304 + ], + "type": "image", + "image_path": "930f4118ac7aca512809d2f20097a1a27dfa7de897d6c4277681788fb22f12bd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 307, + 522, + 313 + ], + "lines": [ + { + "bbox": [ + 476, + 307, + 522, + 313 + ], + "spans": [ + { + "bbox": [ + 476, + 307, + 522, + 313 + ], + "type": "text", + "content": "(III) LayoutLLM (Ours)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 310, + 332, + 399, + 376 + ], + "blocks": [ + { + "bbox": [ + 310, + 332, + 399, + 376 + ], + "lines": [ + { + "bbox": [ + 310, + 332, + 399, + 376 + ], + "spans": [ + { + "bbox": [ + 310, + 332, + 399, + 376 + ], + "type": "image", + "image_path": "6e0a4723496339970e0d87b8bfbf5942039012f5a0ffced855ac97b6c87e36d0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 379, + 451, + 387 + ], + "lines": [ + { + "bbox": [ + 389, + 379, + 451, + 387 + ], + "spans": [ + { + "bbox": [ + 389, + 379, + 451, + 387 + ], + "type": "text", + "content": "(b) Training paradigms" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 400, + 333, + 472, + 375 + ], + "blocks": [ + { + "bbox": [ + 390, + 318, + 451, + 326 + ], + "lines": [ + { + "bbox": [ + 390, + 318, + 451, + 326 + ], + "spans": [ + { + "bbox": [ + 390, + 318, + 451, + 326 + ], + "type": "text", + "content": "(a) Model architectures" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 400, + 333, + 472, + 375 + ], + "lines": [ + { + "bbox": [ + 400, + 333, + 472, + 375 + ], + "spans": [ + { + "bbox": [ + 400, + 333, + 472, + 375 + ], + "type": "image", + "image_path": "a748fae26f0c2f517931304e7719f8fe618cbcf011533e653ad0141c8e1870d6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 473, + 333, + 531, + 375 + ], + "blocks": [ + { + "bbox": [ + 473, + 333, + 531, + 375 + ], + "lines": [ + { + "bbox": [ + 473, + 333, + 531, + 375 + ], + "spans": [ + { + "bbox": [ + 473, + 333, + 531, + 375 + ], + "type": "image", + "image_path": "6d03aba8ef4cf57671481d86956f799803e8d1c835f732fd97f88171320e3dc7.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 399, + 545, + 454 + ], + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 454 + ], + "type": "text", + "content": "Figure 1. LLMs/MLLMs for document understanding. The LayoutLLM is an LLM/MLLM based method that integrates a document pre-trained model as encoder. It is trained by the newly proposed layout instruction tuning strategy which consists of Layout-aware Pre-training and Layout-aware Supervised Fine-tuning." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 520 + ], + "type": "text", + "content": "ument AI downstream tasks. However, due to the necessity for fine-tuning on corresponding downstream task data, it is challenging to directly adapt such pre-trained models for zero-shot document understanding. In this paper, zero-shot refers to not using training sets of downstream tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": "Recently, large language models (LLMs) such as ChatGPT [35] and LLaMA [49, 50], or multimodal large language models (MLLMs) like GPT-4V [1, 36, 56], have shown remarkable zero-shot capabilities across various applications. For Document AI, as shown in Fig. 1 (a), (I) directly prompting LLMs with document text [15, 39] and (II) training document-based MLLMs [3, 57, 60] have also achieved promising results under the zero-shot setting [3, 39, 57, 60]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "It is widely accepted that document layout information is vital for document understanding [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. However, it is difficult to convey document layout information by directly feeding text to LLMs. As Fig. 1(a)(I) shows, representing documents as either flattened plain text or layout text such as text with coordinates [15, 39, 44, 64] is often used for LLMs." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15630" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "Flattened plain text completely excludes any layout information of the document [54]. Additionally, as Tab. 1 shows, using layout text that represents both textual and layout information as inputs for LLMs does not guarantee LLMs can effectively comprehend this formatted text." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 289, + 410 + ], + "type": "text", + "content": "Moreover, existing works that employ MLLMs for document understanding also have not fully explored the document layout information. Document-based MLLMs integrate visual models [11, 42] with LLM [48-50, 61] for document understanding. As Fig. 1(b) shows, they are typically based on pre-training and supervised fine-tuning (SFT) on document datasets. In the pre-training stage, tasks such as image captioning [29, 57, 60] or generating all text in a document as flattened plain text [9, 27, 60] are commonly applied. Both these image captions [43, 46] and plain text only provide a brief representation and fail to capture the layout information of the document. So it is difficult for the model to learn document layout in the existing pre-training stage. In the SFT stage, document-related VQA or information extraction data [3, 57] is often used. The answers are directly provided during SFT, lacking explicit learning about document layout. In summary, current approaches using plain or layout text to prompt LLMs and training document-based MLLMs have not effectively captured layout information, limiting their zero-shot document understanding capability. Therefore, for better document understanding with the power of LLMs, it is necessary to investigate how to effectively incorporate layout information into LLMs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 289, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 676 + ], + "type": "text", + "content": "To this end, we propose LayoutLLM, an LLM/MLLM based method for document understanding, in which a layout instruction tuning strategy is designed to enhance the comprehension of document layouts. Different from the existing MLLMs that use a general visual pre-trained model [11, 42] as the encoder, we integrate document pretrained models [2, 12, 16, 17, 23, 25, 32, 52, 54, 55] as the encoder in order to better leverage the model's foundational understanding capability for documents. The proposed layout instruction tuning consists of two stages: layout-aware pre-training and layout-aware supervised fine-tuning (SFT). Due to the complex nature of documents in real-world scenarios, encompassing rich textual content and diverse layout structures, achieving a thorough understanding involves not only capturing the document's fundamental content at global but also delving into local details. In the layout-aware pre-training stage, to ensure the model learns not only the global information of documents but also detailed information at different hierarchical levels, three groups of different level pre-training tasks are proposed: document-level, region-level, and segment-level. All the proposed pretraining tasks are unified in the format of instruction tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Furthermore, in the layout-aware SFT stage, to enhance the model's comprehension and utilization of layout information for question answering, a novel strategy called" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 263 + ], + "type": "text", + "content": "LayoutCoT is proposed, motivated by the chain-of-thought (CoT) [21, 53] ability in LLMs. Unlike existing methods that are directly supervised by the answer to the document understanding question, LayoutCoT consists of three successive steps: Question Analysis, Relevant Area Concentration, and Answer Formation. Through these steps, the model gains a deeper understanding of the questions, becomes capable of focusing the relevant areas instead of searching answers in the entire document and can leverage the specific characteristics of identified areas (such as tables, paragraphs, etc.) to accurately infer the answers. It not only brings a certain degree of interpretability, but also provides a feasible way for manual intervention or correction of model results. Extensive zero-shot experiments on five widely-used document understanding benchmarks demonstrate the effectiveness of the proposed LayoutLLM." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 264, + 503, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 503, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 503, + 274 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 301, + 277, + 545, + 445 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 302, + 277, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 277, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 302, + 277, + 545, + 324 + ], + "type": "text", + "content": "1) To better learn document layouts from global to local in layout-aware pre-training, three groups of different level pre-training tasks, which are all implemented through instruction tuning, are proposed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 325, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 325, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 301, + 325, + 545, + 384 + ], + "type": "text", + "content": "2) A novel LayoutCoT strategy is proposed to achieve layout-aware supervised fine-tuning. It enables Layout-LLM to focus on the relevant document area and leverage the region's features to generate accurate answers, exhibiting a certain degree of interpretability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 301, + 384, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 384, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 301, + 384, + 545, + 445 + ], + "type": "text", + "content": "3) Experimental results on zero-shot document understanding tasks show that the proposed LayoutLLM significantly outperforms existing methods that adopt LLMs/MLLMs for document understanding, demonstrating the great potential of document layout modeling." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 454, + 398, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 398, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 398, + 466 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 713 + ], + "type": "text", + "content": "Pre-trained models for document understanding. Document pre-trained models have demonstrated the effectiveness of layout information in document understanding [2, 5, 8, 10, 12, 13, 16, 17, 20, 22, 23, 25, 26, 32, 38, 41, 52, 54, 55, 59]. As a pioneer, LayoutLM [54] is the first to encode spatial coordinates of text for layout representation learning in pre-training documents. The following works [2, 8, 12, 13, 16, 17, 23, 25, 26, 32, 38, 52, 55, 59] then joint text, layout and images in document pre-training by combining visual models as document image encoders with the text and layout transformers, and various works [5, 10, 20, 22] start to explore pre-training end-to-end models for document understanding. These studies have achieved significant advancements in document understanding by exploring various model architectures [2, 8, 10, 12, 13, 20, 25, 26, 38, 41, 52, 59] and attention mechanisms [16, 17, 55] for modeling layout information. These methods also have proposed layout-related pre-training tasks that have been demonstrated to be highly effective in document understanding tasks. For instance," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15631" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "tasks like masked vision-language modeling [13, 17, 55], where the model is required to generate the original text corresponding to the randomly masked text in the document; position masking [31, 51], involving the randomly position masking and subsequent recovery of position information in the document; geometric pre-training [26, 32], focusing on learning direction, distance, etc.; and layout-aware generation tasks [5, 22], aiming to make the model generate structured text with layout information. However, due to the necessity of fine-tuning with annotated data for downstream tasks, these efforts face challenges in extending to zero-shot document understanding." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 504 + ], + "type": "text", + "content": "LLMs/MLLMs for document understanding. Recently, LLMs like ChatGPT [35] and MLLMs like GPT-4 [36, 56] have demonstrated remarkable zero-shot performance across a wide range of NLP/CV tasks. Leveraging LLMs/MLLMs for zero-shot document understanding has also shown promising progress [3, 30, 39, 45, 57, 60]. Perot et al. [39] explore the use of LLMs for document visual information extraction, emphasizing the importance of the document layout. LLaVAR [60] which extends LLaVA [28, 29] to the document domain is pre-trained by generating plain text in the document image. During SFT, it is trained by document-related instructions which are generated by GPT-4. Expanding on mPLUG-Owl [58], mPLUG-DocOwl [57] is trained using publicly available datasets for document understanding. It includes tasks like document-level image captioning, direct information extraction, and direct document VQA. Moreover, Qwen-VL [3] proposes a general MLLM that performs well on document understanding tasks, utilizing document-level pre-training and direct VQA for SFT. Though existing LLMs/MLLMs have shown promising results in document understanding, their limited focus on document layout in pre-training and SFT has hindered their ability to achieve higher accuracy in zero-shot document understanding and better interpretability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 513, + 127, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 513, + 127, + 525 + ], + "spans": [ + { + "bbox": [ + 47, + 513, + 127, + 525 + ], + "type": "text", + "content": "3. LayoutLLM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "LayoutLLM is an LLM/MLLM based method that incorporates document pre-trained models for document understanding. To enhance the document layout comprehension in LayoutLLM, a novel layout instruction tuning strategy is proposed, which consists of two stages: layout-aware pretraining and layout-aware supervised fine-tuning (SFT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 162, + 623 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "The overall architecture of LayoutLLM is shown in Fig. 2. In LayoutLLM, given an input document image and its corresponding text and layout information, the document pre-trained model encoder is required to obtain the multimodal document features. Then, these features are encoded by multimodal projectors, and together with the instruction embeddings, fed into the LLM to generate the final results." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 339, + 71, + 515, + 183 + ], + "blocks": [ + { + "bbox": [ + 339, + 71, + 515, + 183 + ], + "lines": [ + { + "bbox": [ + 339, + 71, + 515, + 183 + ], + "spans": [ + { + "bbox": [ + 339, + 71, + 515, + 183 + ], + "type": "image", + "image_path": "38e33a72c3a2e8f530d4ef31b412f7eb52b03c6a13c3fc7a3cc5dc0dff274d67.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 187, + 509, + 198 + ], + "lines": [ + { + "bbox": [ + 341, + 187, + 509, + 198 + ], + "spans": [ + { + "bbox": [ + 341, + 187, + 509, + 198 + ], + "type": "text", + "content": "Figure 2. Overall architecture of LayoutLLM." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 202, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 202, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 545, + 299 + ], + "type": "text", + "content": "Document pre-trained model encoder. To leverage the foundational document comprehension capability of document pre-trained models, in this work, we utilize LayoutLMv3 [17], a widely-used document pre-training model, as our basic document encoder. The document image, text, and layout are initially inputted into the document pretrained model (DocPTM). They are then encoded by the DocPTM to obtain the corresponding features as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 310, + 545, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 310, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 351, + 310, + 545, + 323 + ], + "type": "interline_equation", + "content": "F _ {V}, F _ {T} = \\operatorname {D o c P T M} (V, T, B o x) \\tag {1}", + "image_path": "a18fc2b1011c1279f8853b7f534e7e39f6666e47a6c52f26effd6f76154562c3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " represents the document image, " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "T = T_{0:n - 1}" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "Box = Box_{0:n - 1}" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " indicate the text sequences in the document and their corresponding bounding-box coordinates respectively. After being encoded by the DocPTM, the visual features of the document " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "F_{V} = F_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{0}}" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " and the text layout features " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "F_{T} = F_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{0}}" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " are acquired. " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " signifies the number of visual features and " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " represents the number of tokens contained in the document. " + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "inline_equation", + "content": "d_0" + }, + { + "bbox": [ + 305, + 326, + 545, + 433 + ], + "type": "text", + "content": " denotes the dimension of DocPTM feature space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 434, + 545, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 505 + ], + "type": "text", + "content": "Multimodal projectors. To project multi-modality features from DocPTM into the LLM's embedding space, inspired by the simple yet effective projector design in LLaVA [28, 29], two different Multilayer Perceptrons (MLPs) are used as visual projector and text & layout projector respectively. Formally, the projected features can be obtained by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 510, + 545, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 510, + 545, + 523 + ], + "spans": [ + { + "bbox": [ + 392, + 510, + 545, + 523 + ], + "type": "interline_equation", + "content": "H _ {V} = P _ {V} \\left(F _ {V}\\right) \\tag {2}", + "image_path": "a7be258369c8f2bda4b120afad1e857045d7bf551e5e56ec7efb6073a9decd43.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 393, + 526, + 545, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 526, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 393, + 526, + 545, + 539 + ], + "type": "interline_equation", + "content": "H _ {T} = P _ {T} \\left(F _ {T}\\right) \\tag {3}", + "image_path": "c68f59d661fee8184e218bf38290b9b54e6101e3760d76e9cc3f57fb44585f38.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "inline_equation", + "content": "H_{V} = H_{V_{0}:V_{m - 1}}\\in \\mathbb{R}^{d_{1}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "text", + "content": " is the feature encoded by the visual projector, " + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "inline_equation", + "content": "H_{T} = H_{T_{0}:T_{n - 1}}\\in \\mathbb{R}^{d_{1}}" + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "text", + "content": " is the feature encoded by the text & layout projector, and " + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "inline_equation", + "content": "d_{1}" + }, + { + "bbox": [ + 304, + 545, + 545, + 593 + ], + "type": "text", + "content": " is the dimensional of the LLM embedding features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": "Large language model. Finally, the " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "H_V" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "H_T" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": " and the embedding of the question's instruction text, " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "I_q = I_{q_0:l_q - 1}" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": ", are inputted together into the LLM, generating the target answer " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "I_a = I_{a_0:l_a - 1}" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "l_q" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "inline_equation", + "content": "l_a" + }, + { + "bbox": [ + 305, + 594, + 545, + 654 + ], + "type": "text", + "content": " represent the length of the question's instruction text and the answer text, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 659, + 454, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 454, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 454, + 672 + ], + "type": "text", + "content": "3.2. Layout Instruction Tuning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "The LayoutLLM model is trained using the layout instruction tuning, which consists of two stages: layout-aware pretraining and layout-aware SFT." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15632" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 541, + 247 + ], + "type": "image", + "image_path": "13b8337221be70f40d2a4e8849fcd0cd0072db60f067a6e54a1f07b7d6790542.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 546, + 289 + ], + "type": "text", + "content": "Figure 3. Overview of the Layout Instruction Tuning. (a) Document-level, region-level, and segment-level pre-training tasks, unified in instruction tuning format, are introduced. (b) A novel module called LayoutCoT is designed to enable LayoutLLM to focus on regions relevant to the question and generate accurate answers through three intermediate steps." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 304, + 195, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 304, + 195, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 304, + 195, + 316 + ], + "type": "text", + "content": "3.2.1 Layout-aware Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 422 + ], + "type": "text", + "content": "The goal of pre-training the LayoutLLM is to enhance the model's comprehensive understanding of documents at different levels through layout learning, rather than only focusing on global document-level understanding like existing MLLM methods [57, 60]. To this end, during the pre-training stage, three different level pre-training strategies are simultaneously applied to the LayoutLLM, namely document-level, region-level, and segment-level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 425, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 664 + ], + "type": "text", + "content": "Document-level To enable the model to possess fundamental global document understanding, pre-training tasks, namely Document Dense Description (DDD) and Text and Layout Reconstruction (TLR), are proposed. As Fig. 3 (a) shows, like the image caption task, the DDD task requires the model to learn to describe the input document. Moreover, in the DDD task, the descriptions for document images are more detailed. For instance, in the document image caption data used for the LLaVAR [60] pre-training, captions contain an average of 36.27 words, while in the proposed dataset of the DDD task, the description contains an average of 373.25 words. Through the DDD task, the model can obtain basic document-level information, such as document type and detailed content. The TLR task aims to reconstruct the complete text and layout information of a document and output it in the format “<\\{box\\}, \\{text\\} >”. The TLR task aligns the text and layout embeddings output from DocPTM with the LayoutLLM's LLM space. Consequently, it enables the LLM in LayoutLLM to comprehend the text and layout information contained in the documents." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "Region-level The information contained in specific regions of a document, such as the titles, figures, tables, is essential for document understanding [4, 6, 8, 24, 62]. These regions serve as important characteristics that differentiate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 304, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 471 + ], + "type": "text", + "content": "a document from plain text in natural language. For the LayoutLLM to achieve the basic region-level understanding, two pre-training tasks, namely Document Layout Analysis (DLA) and Table Understanding (TU), are utilized. The DLA task is achieved in two ways as shown in Fig. 3. One involves locating the layout region based on the layout type, the other involves identifying the type of a given area. Furthermore, the table region differs from other regions in that it requires additional focus on 2D layout understanding. The TU task enables the model to understand the basic row and column information in the table region of a document. As shown in Fig. 3, the TU task includes instruction tuning for the number of rows and columns, logical coordinates, and the content within rows and columns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "Segment-level Early works on document pre-trained models [13, 17, 26, 31, 32, 51, 55] have demonstrated the effectiveness of segment-level document pre-training tasks to document layout understanding ability, such as masked vision language modeling (MVLM) [13, 17, 55], position masking [31, 51], and geometric pre-training [26, 32]. Inspired by these works, to make LayoutLLM have segment-level layout understanding, these tasks are transformed into instruction formats for pre-training as Fig. 3(a) shows. For the MVLM instructions, random masking of the text input to LayoutLLM is performed, and the model is instruction tuned by asking the masked words and answering them. For mask position instruction, the layout information (coordinates) to a specific text line, when input to LayoutLLM, is randomly set to 0. The instruction is constructed by asking about the text line with zeroed coordinates and requesting the model to respond with the original coordinates with text content. For geometric layout instruction, text lines are randomly selected, and an instruction is constructed by asking questions about the direction and distance between them." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15633" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 242, + 84 + ], + "type": "text", + "content": "3.2.2 Layout-aware Supervised Fine-tuning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 258 + ], + "type": "text", + "content": "In the SFT stage of existing document-based MLLMs, models are directly supervised by the answer to the document understanding instructions. Consequently, these methods lack explicit learning of document layout which is crucial for document understanding. Considering this limitation, and inspired by previous works related to chain-of-though (CoT) [21, 53], which have shown that inferencing with intermediate steps can greatly enhance performance. A novel module called LayoutCoT is proposed, which incorporates the layout information into every intermediate step of CoT explicitly. Meanwhile, by introducing the layout-aware intermediate steps, the answer process gains a certain degree of interpretability for LayoutLLM and also provides interactive correction possibilities based on LayoutCoT." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 288, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 288, + 282 + ], + "type": "text", + "content": "LayoutCoT Details. As shown in Fig. 3(b), the LayoutCoT involves the following three intermediate steps:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 461 + ], + "type": "text", + "content": "Step 1: Question Analysis. To effectively address a document understanding problem, analyzing the key characteristics of the question is very important. Identifying the question type, such as table understanding or entity extraction from paragraphs, and assessing whether the question is a straightforward extraction query or a more complex reasoning problem, can help guide the direction for the subsequent inference process. Therefore, to give basic guidance to the subsequent steps, the question analysis step is designed, encompassing an analysis of the question type from a layout perspective and a detailed understanding of the question itself. Benefiting from the layout understanding ability by layout-aware pre-training, this step can extract the types and key information mentioned in the question, which are related to the specific characteristics of the document." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 652 + ], + "type": "text", + "content": "Step 2: Relevant Area Concentration. For most document understanding tasks, the entire document contains a large amount of irrelevant information that may confuse the model [5]. This step aims to focus on the relevant area and generate its location information, which is used to assist the model to accurately infer the answer. Benefiting from the layout information conveyed by step 1 and the positioning capabilities learned from the region&segment-level pretraining, the model can accurately generate the location of the relevant area. For example in Fig. 3(b), according to the question type \"table\" in step 1, the relevant \"table\" can be located. By guiding the model to focus on the relevant area, this step largely narrows the search scope, increasing the possibility of giving the right answer. Meanwhile, the location information provides a way for visual inspection and interactive correction (see Sec. 4.7 for details)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 714 + ], + "type": "text", + "content": "Step 3: Answer Formation. Finally, the last step, the answer formation, provides explanations based on the layout characteristics of the relevant areas located in step 2 and key points analyzed in step 1 to get the final answer. For example in Fig. 3(b), for a \"table\" type question, this step in-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 309, + 69, + 544, + 81 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 69, + 544, + 81 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 544, + 81 + ], + "type": "text", + "content": "Algorithm 1 CONSTRUCT(D): LayoutCoT Construction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 311, + 83, + 544, + 202 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "spans": [ + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": "Definition: " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : Document HTMLs; " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : Document Images; " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : MRC Texts; " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : Document Language Representation; QA: QA pairs; " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : Text CoT; " + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{c}" + }, + { + "bbox": [ + 311, + 83, + 544, + 112 + ], + "type": "text", + "content": " : LayoutCoT;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 113, + 471, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 113, + 471, + 123 + ], + "spans": [ + { + "bbox": [ + 312, + 113, + 471, + 123 + ], + "type": "text", + "content": "Input: Document Dataset " + }, + { + "bbox": [ + 312, + 113, + 471, + 123 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 123, + 460, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 123, + 460, + 133 + ], + "spans": [ + { + "bbox": [ + 323, + 123, + 460, + 133 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_H = \\{\\mathcal{H}\\}, \\mathcal{D}_I = \\{\\mathcal{I}\\}, \\mathcal{D}_M = \\{\\mathcal{T}, \\mathcal{Q}\\mathcal{A}\\})" + }, + { + "bbox": [ + 323, + 123, + 460, + 133 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 133, + 420, + 141 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 133, + 420, + 141 + ], + "spans": [ + { + "bbox": [ + 312, + 133, + 420, + 141 + ], + "type": "text", + "content": "Output: Constructed Dataset " + }, + { + "bbox": [ + 312, + 133, + 420, + 141 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 144, + 487, + 155 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 144, + 487, + 155 + ], + "spans": [ + { + "bbox": [ + 312, + 144, + 487, + 155 + ], + "type": "text", + "content": "1: Procedure CONSTRUCT " + }, + { + "bbox": [ + 312, + 144, + 487, + 155 + ], + "type": "inline_equation", + "content": "(\\mathcal{D} = \\{\\mathcal{D}_H,\\mathcal{D}_I,\\mathcal{D}_M\\})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "spans": [ + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "text", + "content": "2: 1) " + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = getDocRep(\\mathcal{I})" + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\subseteq \\mathcal{D}_I" + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "text", + "content": " else " + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\subseteq \\mathcal{D}_H" + }, + { + "bbox": [ + 312, + 155, + 538, + 164 + ], + "type": "text", + "content": " else pass;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "text", + "content": "3: 2) " + }, + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}\\mathcal{A},\\mathcal{T}_c = getQACoT(\\mathcal{D})" + }, + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\subseteq \\mathcal{D}_M" + }, + { + "bbox": [ + 312, + 164, + 544, + 174 + ], + "type": "text", + "content": " else getQACoTGPT(R);" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 174, + 421, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 174, + 421, + 183 + ], + "spans": [ + { + "bbox": [ + 312, + 174, + 421, + 183 + ], + "type": "text", + "content": "4: 3) " + }, + { + "bbox": [ + 312, + 174, + 421, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c = getLayoutCoT(\\mathcal{T}_c)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "spans": [ + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "text", + "content": "5: 4) if " + }, + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\subseteq \\{\\mathcal{D}_H, \\mathcal{D}_M\\}" + }, + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = Htm12Img(\\mathcal{D}_H \\cap \\mathcal{H} : \\mathcal{T})" + }, + { + "bbox": [ + 312, + 183, + 508, + 193 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 193, + 414, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 193, + 414, + 202 + ], + "spans": [ + { + "bbox": [ + 312, + 193, + 414, + 202 + ], + "type": "text", + "content": "6: return " + }, + { + "bbox": [ + 312, + 193, + 414, + 202 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c\\gets \\{\\mathcal{I},\\mathcal{Q}\\mathcal{A},\\mathcal{L}_c\\}" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 304, + 218, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 302 + ], + "type": "text", + "content": "volves analyzing the row and column in the relevant table in step 2, and referencing the answer step-by-step. For a \"key-value\" question, analyzing the keywords in concentrated areas can help get the final answer. Analyzing answers in different ways based on the features of different layout regions not only improves the document understanding performance but also brings a certain level of interpretability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 303, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 410 + ], + "type": "text", + "content": "LayoutCoT Construction. Given the need for both text and image annotations in constructing LayoutCoT, manual labeling can be difficult. Algorithm 1 proposes a manual-labeling-free method, generating LayoutCoT data using public datasets with GPT (GPT-3.5 Turbo). It involves representing document text and layout in a format understandable by GPT. GPT is then utilized to generate document-content-based QA and corresponding text CoT. Finally, use rules for transforming the text CoT to LayoutCoT." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "text", + "content": "Three types of publicly available document datasets are focused on: HTML documents " + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_H)" + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "text", + "content": ", image documents " + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_I)" + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "text", + "content": ", and text documents " + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_M)" + }, + { + "bbox": [ + 304, + 411, + 545, + 460 + ], + "type": "text", + "content": " for machine reading comprehension (MRC). The construction process is as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 460, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": "1) Document Representation: To fully leverage the capabilities of GPT, it is crucial to ensure that the document content fed to GPT contains accurate layout information. Since HTML is the formatted language that can represent documents accurately, " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_H" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " is represented using the original HTML. By transforming HTML to PDF and using the PDF parser, the text and bounding-boxes are obtained. For " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_I" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": ", the layout-aware text [15] is used. The text and bounding-boxes in " + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_I" + }, + { + "bbox": [ + 304, + 460, + 545, + 567 + ], + "type": "text", + "content": " are from the original dataset annotations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": "2) QA&Text CoT Generation: The language representation " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": " for the document is employed for prompting GPT to generate QA pairs " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{QA}" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": " with text CoTs " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": ". In addition, the " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_M" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": " includes the QA pairs and reasoning process, thereby directly reusing the " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{QA}" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": " and manually organizing " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": ". The generated " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": " includes the step 1 (question analysis) and step 3 (answer formation) for LayoutCoT, and locates all relevant sentences in the document for " + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{QA}" + }, + { + "bbox": [ + 304, + 569, + 545, + 663 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "3) LayoutCoT Generation: The step 1 & 3 in " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " are used as the step 1 & 3 in " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": ". To construct the step 2 (relevant area concentration) of " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": ", the union bounding-box of all located relevant sentences in " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " are taken as the relevant area." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15634" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "4) Document Images Generation: For " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_H" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_M" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", the HTMLs and MRC text are converted to images. Overall, the document images " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": ", generated QA pairs " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{QA}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and LayoutCoTs " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " constitute the final LayoutCoT dataset " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 134, + 128, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 128, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 128, + 148 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 155, + 155, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 155, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 155, + 167 + ], + "type": "text", + "content": "4.1. Dataset Collection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 174, + 289, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 289, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 289, + 486 + ], + "type": "text", + "content": "Layout-aware pre-training data of LayoutLLM is from publicly available document understanding datasets. It does not incorporate any data from the training, validation, and test sets of downstream benchmarks. Region-level pretraining tasks, most document-level and segment-level tasks are self-supervised. Thus, only document images and images converted from PDFs in the original datasets, along with the corresponding OCR or text layout results from PDF parsing, are needed. For these tasks, data is randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65]. Particularly, data for document dense description is from inputting the document text content into GPT-3.5 Turbo, prompting it to generate an average of 373.25 words document dense descriptions. For the region-level tasks, specifically the document layout analysis task, publicly available document layout analysis datasets are utilized, including PubLayNet [62], DocLayNet [40], and Docbank [24]. Data for another region-level task, table understanding, is sourced from PubTabNet [63] with its table annotations. All data is transformed into the instruction format illustrated in Fig. 3(a). In total, 5.7 million instructions are constructed, with a ratio of 1:4:4 for document-level, region-level, and segment-level tasks, respectively. For detailed instruction templates and dataset descriptions, please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "Layout-aware SFT data of LayoutLLM is generated by GPT (GPT-3.5 Turbo) and converted from existing textual Machine Reading Comprehension (MRC) datasets, as discussed in Sec. 3.2.2. To generate high-quality document-based textual QA and textual CoT, it is essential to make GPT comprehend the document layout. So, the document is represented using both layout text [15] and HTML. Similar to the pre-training data, the " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_I" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " in Algorithm 1 is also randomly sampled from PubLayNet [62], DocLayNet [40], Docbank [24], RVL-CDIP [14], and DocILE [65] for building layout text. The " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " in Algorithm 1 is from GPT's free generation. The " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_M" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " in Algorithm 1 is randomly sampled from the FeTaQA [34] which is a wikipedia question answering dataset. A total of 300K instructions are constructed, with a ratio of 5:4.5:0.5 for " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_I" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "D_M" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": ", respectively. For detailed prompt templates of document-based text QA and text CoT generation using GPT, prompts for HTML generation using GPT, and dataset description, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 400, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 400, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 400, + 85 + ], + "type": "text", + "content": "4.2. Training Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 89, + 547, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 547, + 212 + ], + "type": "text", + "content": "The encoder weight of LayoutLLM is initialized from the LayoutLMv3-large [17] which is a widely-used document pre-trained model. And the LLM backbone is initialized from Vicuna-7B-v1.5 [61]. Other parameters are randomly initialized. During pre-training, the LLM is frozen, and the parameters of the two projectors and document pre-trained model encoder are updated. During SFT, both LLM and two projectors are fine-tuned while keeping the document pretrained model encoder frozen. For detailed training setup, please refer to the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 216, + 410, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 216, + 410, + 229 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 410, + 229 + ], + "type": "text", + "content": "4.3. Evaluation Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 234, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 547, + 594 + ], + "type": "text", + "content": "The zero-shot ability is highly expected in real-world document understanding scenarios [7, 30, 45]. Therefore, zero-shot evaluations on widely-used document understanding benchmarks including document visual question answering (Document VQA) and visual information extraction (VIE) are conducted. Only the test sets are utilized in all benchmarks and only the official provided image, text, and layout information are used. The Document VQA datasets comprise the DocVQA[33] test set, consisting of 5,188 questions, and the VisualMRC[47] test set containing 6,708 questions. Following the evaluation metric settings of the original datasets, the ANLS [33] is utilized for evaluating DocVQA, and Rouge-L is used for evaluating VisualMRC. For the VIE task, FUNSD [19], CORD [37], and SROIE [18] are used. The test set of FUNSD comprises 50 form images, each annotated with entity-level headers, questions, answers, and others, along with entity linking annotations. CORD's test set consists of 100 receipt images, annotated with 30 entity types, such as the tax amount, total price, etc. SROIE's test set includes 347 receipt images, annotated with 4 entity types: company, date, address, and total. To prompt LLMs/MLLMs for zero-shot VIE, annotations in VIE datasets are transformed into question answering format (QA for VIE). For key-value annotations with linking in the FUNSD, the format is {Q: What is the \"key\" in the document? A: \"value)}. For entity annotations in CORD and SROIE, directly asking for the entity in the document like {Q: What is the address in the document? A: \"the address annotation\"} is utilized. Following DocVQA, the QA for VIE task is evaluated by ANLS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 599, + 391, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 391, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 391, + 612 + ], + "type": "text", + "content": "4.4. Main Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "As shown in Tab. 1, the zero-shot document understanding performance of LayoutLLM and existing open-source LLMs and MLLMs is evaluated. Generally, the existing LLMs are better than MLLMs for zero-shot document VQA and VIE. For example in the results on DocVQA, most LLMs can achieve a performance of around " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": " or higher, while most MLLMs can only attain around " + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": ", except for mPLUG-DocOWL and Qwen-VL that trained with the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15635" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 62, + 533, + 258 + ], + "blocks": [ + { + "bbox": [ + 59, + 62, + 533, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 62, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 62, + 533, + 258 + ], + "type": "table", + "html": "
ModelsDocument VQAQA for VIE
DocVQAVisualMRCFUNSDCORDSROIE
Fine-tuned PTMLayoutLMv3 [17]83.37*-92.08*‡97.46*‡-
LLMLlama2-7B [50]61.3429.7340.784.3915.86
Llama2-7B-chat [50]64.9952.8448.2047.7068.97
Plain TextVicuna-7B [61]61.3953.6349.7944.6767.49
Vicuna-1.5-7B [61]66.9952.1348.0651.4068.20
LLMLlama2-7B [50]37.3233.8251.4028.0434.96
Llama2-7B-chat [50]56.5549.2658.3450.9351.15
Layout Text (Text + Box) [15]Vicuna-7B [61]37.2152.5542.7346.5945.43
Vicuna-1.5-7B [61]56.8147.2259.6356.1366.20
MLLMLLaVAR-7B [60]11.6†36.371.7113.552.38
LLaVA-1.5-7B [28]13.3435.231.9318.063.83
mPLUG-DocOWL-7B [57]62.2*†----
Qwen-VL-7B [3]65.1*†42.5247.0930.0058.59
LayoutLLM-7B△ (Ours)74.2555.7378.6562.2170.97
LayoutLLM-7B★ (Ours)74.2755.7679.9863.1072.12
", + "image_path": "e82125f67fba9f385d2e11fec999c61386d341ac2f72e66fc462411316d44554.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "content": "training set. One possible reason is that it's difficult for these MLLMs to obtain accurate textual information from document images. Additionally, for LLMs, using Plain Text and Layout Text respectively as the document representation are further discussed, where the Layout Text introduces layout information by adding text coordinates in the format: {text:\"text\", box:[x1,y1,x2,y2]} [15]. Compared to the Plain Text, the Layout Text variant doesn't show stable performance improvements, noticeable in certain tasks, for example in Vicuna-1.5, an improvement in VIE (FUNSD " + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "inline_equation", + "content": "48.06\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "inline_equation", + "content": "59.63\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "content": ") but a decline in DocVQA " + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "inline_equation", + "content": "(66.99\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "inline_equation", + "content": "56.81\\%)" + }, + { + "bbox": [ + 46, + 304, + 289, + 484 + ], + "type": "text", + "content": ". LLMs may lack the ability to learn this formatted layout text, and directly adding layout information (e.g., coordinates) to the text will also highly increase the token length, making the answer inference more challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 485, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 289, + 715 + ], + "type": "text", + "content": "Compared to the prior SOTA model, LayoutLMv3, which is fine-tuned using the training set of downstream tasks, LayoutLLM demonstrates competitive performance on the DocVQA benchmark. Compared with these LLMs and MLLMs, LayoutLLM achieves consistent and significant improvements over them on all evaluation benchmarks. Notably, LayoutLLM which employs zero-shot performance, outperforms mPLUG-DocOWL and Qwen-VL by around " + }, + { + "bbox": [ + 46, + 485, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 485, + 289, + 715 + ], + "type": "text", + "content": " on the DocVQA dataset, both of which are trained with this dataset. This demonstrates that LayoutLLM can learn more robust and discriminative representations for document understanding. Furthermore, experiments of the different initialization of the LLM backbone have all achieved optimal results across all benchmarks, substantiating that LayoutLLM can adapt to various LLMs. In summary, our method explores a more effective way to utilize layout information for document understanding, which significantly improves the performance of zero-shot document understanding." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 308, + 298, + 544, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": "Table 1. Zero-shot document understanding results on open-source LLMs and MLLMs. * signifies training set use; unmarked results are zero-shot. Results marked with " + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "inline_equation", + "content": "{}^{ \\ddagger }" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": " are F1 scores for VIE. Results marked with " + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": " are from the original paper and others are re-implemented by us. " + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\bigtriangleup" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": " marks LayoutLLM's LLM backbone as initialized with Llama2-7B-chat,and " + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": " with Vicuna-1.5-7B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 298, + 544, + 352 + ], + "lines": [ + { + "bbox": [ + 308, + 298, + 544, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 544, + 352 + ], + "type": "table", + "html": "
#Layout-aware \nPre-trainingLayout-aware \nSFTDocVQAFUNSD
070.8270.96
1✓72.3174.02
2✓✓74.2779.98
", + "image_path": "e5737a2582d79468e1080f86076b20ef23e3ac7d1528229e840545e6057d4619.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 310, + 354, + 539, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 354, + 539, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 354, + 539, + 365 + ], + "type": "text", + "content": "Table 2. Ablation study on the DocVQA and FUNSD test sets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 369, + 398, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 398, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 398, + 382 + ], + "type": "text", + "content": "4.5. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 424 + ], + "type": "text", + "content": "To better verify the effectiveness of the layout-aware pretraining and layout-aware SFT in the layout instruction tuning, an ablation study is conducted (see Tab. 2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": "Initial baseline. The #0 baseline disables both layout-aware pre-training and layout-aware SFT. It only adopts SFT (same SFT data but without LayoutCoT steps) on the LayoutLLM. Even without any alignment pre-training and LayoutCoT steps guidance, the baseline outperforms previous SOTAs, achieving " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "70.82\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on DocVQA and " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "70.96\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on FUNSD. This indicates the document understanding ability of DocPTM benefits document understanding with LLMs. Effect of Layout-aware pre-training. In #1, the layout-aware pre-training and the same SFT with #0 is conducted. Compared to #0, benefiting from the basic document understanding capability learned through the layout-aware pretraining, #1 shows an increase of " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "1.49\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on DocVQA and " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "3.06\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on FUNSD. It can be observed that the basic document understanding ability learned from the layout-aware pre-training significantly enhances the performance of the basic key-value extraction tasks in the FUNSD dataset. Compared to the FUNSD VIE task, it also shows that the DocVQA is a more complex task. Effect of Layout-aware SFT. Compared with #1, #2 further incorporates the layout-aware SFT strategy, resulting in a performance gain of " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on DocVQA and even " + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "inline_equation", + "content": "5.96\\%" + }, + { + "bbox": [ + 304, + 426, + 546, + 714 + ], + "type": "text", + "content": " on FUNSD. This indicates the LayoutCoT in layout-aware SFT can help the LayoutLLM to handle complex document tasks and it is effic" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15636" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 61, + 159, + 116 + ], + "blocks": [ + { + "bbox": [ + 85, + 61, + 159, + 116 + ], + "lines": [ + { + "bbox": [ + 85, + 61, + 159, + 116 + ], + "spans": [ + { + "bbox": [ + 85, + 61, + 159, + 116 + ], + "type": "image", + "image_path": "2c085931ced6d4bf149854cf8647c6ef4f5d296c89fbdf02912e98b6fdafc124.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 117, + 158, + 125 + ], + "lines": [ + { + "bbox": [ + 85, + 117, + 158, + 125 + ], + "spans": [ + { + "bbox": [ + 85, + 117, + 158, + 125 + ], + "type": "text", + "content": "Question: What is the Income Tax Withheld for the month of July 1954?" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 164, + 62, + 250, + 137 + ], + "blocks": [ + { + "bbox": [ + 164, + 62, + 250, + 137 + ], + "lines": [ + { + "bbox": [ + 164, + 62, + 250, + 137 + ], + "spans": [ + { + "bbox": [ + 164, + 62, + 250, + 137 + ], + "type": "image", + "image_path": "ce4b7d257b1d72b26e1b07fd7edbcb9f66bf7af5fda50855d64dcf90090b89b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 85, + 127, + 157, + 149 + ], + "blocks": [ + { + "bbox": [ + 85, + 127, + 157, + 149 + ], + "lines": [ + { + "bbox": [ + 85, + 127, + 157, + 149 + ], + "spans": [ + { + "bbox": [ + 85, + 127, + 157, + 149 + ], + "type": "image", + "image_path": "b336250946d452584eaa12662d0d52a3dd13d5966a47a782dda79845f69f9589.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 86, + 152, + 158, + 204 + ], + "blocks": [ + { + "bbox": [ + 86, + 152, + 158, + 204 + ], + "lines": [ + { + "bbox": [ + 86, + 152, + 158, + 204 + ], + "spans": [ + { + "bbox": [ + 86, + 152, + 158, + 204 + ], + "type": "image", + "image_path": "097a6af6bd5cb4f86bd9ad43ac2415512d63639b7e8a071980726536d774504e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 206, + 151, + 211 + ], + "lines": [ + { + "bbox": [ + 85, + 206, + 151, + 211 + ], + "spans": [ + { + "bbox": [ + 85, + 206, + 151, + 211 + ], + "type": "text", + "content": "Question: What is the Total for Exit 2004?" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 85, + 213, + 157, + 237 + ], + "blocks": [ + { + "bbox": [ + 85, + 213, + 157, + 237 + ], + "lines": [ + { + "bbox": [ + 85, + 213, + 157, + 237 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 157, + 237 + ], + "type": "image", + "image_path": "90fd5a4718cf92ef8920d9ede940eea486e3a53c844d7738b74cdbf619e76b94.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 287, + 262 + ], + "type": "text", + "content": "Figure 4. Qualitative results on DocVQA. Green boxes are the areas concentrated in the step 2 of LayoutCoT." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 158, + 232, + 165, + 237 + ], + "blocks": [ + { + "bbox": [ + 158, + 232, + 165, + 237 + ], + "lines": [ + { + "bbox": [ + 158, + 232, + 165, + 237 + ], + "spans": [ + { + "bbox": [ + 158, + 232, + 165, + 237 + ], + "type": "image", + "image_path": "72b69f819c37d11880391fc1b4164a638f805e8c984d634efc4c474e23d6d3f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 166, + 149, + 250, + 228 + ], + "blocks": [ + { + "bbox": [ + 166, + 149, + 250, + 228 + ], + "lines": [ + { + "bbox": [ + 166, + 149, + 250, + 228 + ], + "spans": [ + { + "bbox": [ + 166, + 149, + 250, + 228 + ], + "type": "image", + "image_path": "3b107afefe8faf6e0623f542bbfff496f92c4926ed48bfd822b1f78a84927f63.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 195, + 230, + 218, + 239 + ], + "blocks": [ + { + "bbox": [ + 195, + 230, + 218, + 239 + ], + "lines": [ + { + "bbox": [ + 195, + 230, + 218, + 239 + ], + "spans": [ + { + "bbox": [ + 195, + 230, + 218, + 239 + ], + "type": "image", + "image_path": "683193296cabd548f6372c9f5491b9a70051cf08d79d1ecaaf247a5ad245bb74.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 267, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 287, + 327 + ], + "type": "text", + "content": "tive for boosting the performance on both document VQA and VIE. Moreover, it also promotes a certain degree of interpretability. Overall, the ablation study demonstrates the importance of layout-aware pre-training and layout-aware SFT for zero-shot document understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 335, + 160, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 160, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 160, + 346 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 353, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 289, + 640 + ], + "type": "text", + "content": "Two examples are shown in Fig. 4. Through combined with layout-aware pre-training and layout-aware SFT, Layout-LLM can accurately focus on the relevant areas, utilize the layout information to assist in problem-solving and provide interpretability. For example, in Fig. 4(a) question about key-value extraction in up-down layout, different from the left-right variant, relies more on document layout to infer the right answer. Since the keywords \"Income Tax\" in the question often co-occur with numerical data, Vicuna-1.5 and Qwen-VL find numerical answers relying more on the semantics than the layout, resulting in incorrect responses. In contrast, benefiting from the layout-aware pre-training, our model can effectively leverage layout information to give accurate answers. In addition, the model using LayoutCoT can further provide the location and the reasoning process, showing a certain degree of interpretability. But in certain situations, only combined with layout pre-training, our model might fail to give accurate answers. As shown in Fig. 4(b), without LayoutCoT, our model identifies \"Exit 2003\" as the relevant column and generates a wrong answer. However, with the help of LayoutCoT, LayoutLLM can correctly identify the question type as \"Table\", locate the relevant table area, and finally infer the right answer from the corresponding \"Exit 2004\" column." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 647, + 255, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 255, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 255, + 659 + ], + "type": "text", + "content": "4.7. Interactive Correction with LayoutCoT" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "Since LayoutCoT runs in a step-by-step fashion and produces intermediate results in the inference stage, it can facilitate interactive inspection and correction, when processing a document. As shown in Fig. 5(a), there are two areas" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 335, + 67, + 413, + 125 + ], + "blocks": [ + { + "bbox": [ + 335, + 67, + 413, + 125 + ], + "lines": [ + { + "bbox": [ + 335, + 67, + 413, + 125 + ], + "spans": [ + { + "bbox": [ + 335, + 67, + 413, + 125 + ], + "type": "image", + "image_path": "8641010ee6173b2ae131e22ffec8f9ff88ea4049185be5ac26e93b051f28a303.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 125, + 410, + 130 + ], + "lines": [ + { + "bbox": [ + 336, + 125, + 410, + 130 + ], + "spans": [ + { + "bbox": [ + 336, + 125, + 410, + 130 + ], + "type": "text", + "content": "Question: In the table given on the left bottom of" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 416, + 68, + 504, + 74 + ], + "lines": [ + { + "bbox": [ + 416, + 68, + 504, + 74 + ], + "spans": [ + { + "bbox": [ + 416, + 68, + 504, + 74 + ], + "type": "text", + "content": "Step 1 (Question Analysis): The question is about key-value extraction. And the question asks for the 'vendor'." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 75, + 508, + 79 + ], + "lines": [ + { + "bbox": [ + 417, + 75, + 508, + 79 + ], + "spans": [ + { + "bbox": [ + 417, + 75, + 508, + 79 + ], + "type": "text", + "content": "Step 2 (Relevant Area Concentration): The answer can be found" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 80, + 465, + 84 + ], + "lines": [ + { + "bbox": [ + 417, + 80, + 465, + 84 + ], + "spans": [ + { + "bbox": [ + 417, + 80, + 465, + 84 + ], + "type": "text", + "content": "in the area: [28, 123, 317, 148]." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 85, + 507, + 90 + ], + "lines": [ + { + "bbox": [ + 417, + 85, + 507, + 90 + ], + "spans": [ + { + "bbox": [ + 417, + 85, + 507, + 90 + ], + "type": "text", + "content": "Step 3 (Answer Formation): The number entered for 'vendor' in the text is:" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 108, + 456, + 112 + ], + "lines": [ + { + "bbox": [ + 417, + 108, + 456, + 112 + ], + "spans": [ + { + "bbox": [ + 417, + 108, + 456, + 112 + ], + "type": "text", + "content": "Step 1 (Question Analysis):" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 113, + 508, + 118 + ], + "lines": [ + { + "bbox": [ + 417, + 113, + 508, + 118 + ], + "spans": [ + { + "bbox": [ + 417, + 113, + 508, + 118 + ], + "type": "text", + "content": "Step 2 (Relevant Area Concentration): The answer can be found in the text: [46,703,467,830]." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 119, + 507, + 124 + ], + "lines": [ + { + "bbox": [ + 417, + 119, + 507, + 124 + ], + "spans": [ + { + "bbox": [ + 417, + 119, + 507, + 124 + ], + "type": "text", + "content": "Step 3 (Answer Formation): The number entered for 'vendor' in" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 124, + 447, + 128 + ], + "lines": [ + { + "bbox": [ + 417, + 124, + 447, + 128 + ], + "spans": [ + { + "bbox": [ + 417, + 124, + 447, + 128 + ], + "type": "text", + "content": "the table is 013684" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 129, + 450, + 133 + ], + "lines": [ + { + "bbox": [ + 417, + 129, + 450, + 133 + ], + "spans": [ + { + "bbox": [ + 417, + 129, + 450, + 133 + ], + "type": "text", + "content": "Final answer: 013684." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 141, + 413, + 201 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 141, + 413, + 201 + ], + "spans": [ + { + "bbox": [ + 336, + 141, + 413, + 201 + ], + "type": "image", + "image_path": "86624bcb5655974bb4b5b0957c487e29a2b2a7391f60c6301b6fe543a1cd22cc.jpg" + } + ] + } + ], + "index": 20, + "type": "text" + }, + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "spans": [ + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "type": "text", + "content": "Question: What is the heading of second paragraph?" + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 462, + 96, + 468, + 100 + ], + "blocks": [ + { + "bbox": [ + 417, + 91, + 473, + 95 + ], + "lines": [ + { + "bbox": [ + 417, + 91, + 473, + 95 + ], + "spans": [ + { + "bbox": [ + 417, + 91, + 473, + 95 + ], + "type": "text", + "content": "Table 1 Table 2" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 96, + 468, + 100 + ], + "lines": [ + { + "bbox": [ + 462, + 96, + 468, + 100 + ], + "spans": [ + { + "bbox": [ + 462, + 96, + 468, + 100 + ], + "type": "image", + "image_path": "671c37c21e3f401282dd5824acdf3d36cac071b217411f43cd15690c94a9bf77.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 469, + 96, + 475, + 100 + ], + "blocks": [ + { + "bbox": [ + 469, + 96, + 475, + 100 + ], + "lines": [ + { + "bbox": [ + 469, + 96, + 475, + 100 + ], + "spans": [ + { + "bbox": [ + 469, + 96, + 475, + 100 + ], + "type": "image", + "image_path": "fe3bb4ef182a5215bc43bae56bf3f48b87db9051de41450fc76c96cf3f7f0ba0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 462, + 102, + 468, + 105 + ], + "blocks": [ + { + "bbox": [ + 462, + 102, + 468, + 105 + ], + "lines": [ + { + "bbox": [ + 462, + 102, + 468, + 105 + ], + "spans": [ + { + "bbox": [ + 462, + 102, + 468, + 105 + ], + "type": "image", + "image_path": "01e428f05a981fa67b4f30d0c169b7b5a6b2b29167a7010ff79388567c85968f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 469, + 102, + 475, + 105 + ], + "blocks": [ + { + "bbox": [ + 469, + 102, + 475, + 105 + ], + "lines": [ + { + "bbox": [ + 469, + 102, + 475, + 105 + ], + "spans": [ + { + "bbox": [ + 469, + 102, + 475, + 105 + ], + "type": "image", + "image_path": "7d9ad74267586c88775852aa6428a6c2d49a534a5a18b72175f0b83117d0c48b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 336, + 140, + 413, + 201 + ], + "blocks": [ + { + "bbox": [ + 336, + 131, + 365, + 136 + ], + "lines": [ + { + "bbox": [ + 336, + 131, + 365, + 136 + ], + "spans": [ + { + "bbox": [ + 336, + 131, + 365, + 136 + ], + "type": "text", + "content": "the page, what is it?" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 137, + 365, + 141 + ], + "lines": [ + { + "bbox": [ + 336, + 137, + 365, + 141 + ], + "spans": [ + { + "bbox": [ + 336, + 137, + 365, + 141 + ], + "type": "text", + "content": "GT Answer: 013684" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 209, + 390, + 214 + ], + "lines": [ + { + "bbox": [ + 336, + 209, + 390, + 214 + ], + "spans": [ + { + "bbox": [ + 336, + 209, + 390, + 214 + ], + "type": "text", + "content": "GT Answer: PRELIMINARY STATEMENT" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 335, + 134, + 365, + 139 + ], + "lines": [ + { + "bbox": [ + 335, + 134, + 365, + 139 + ], + "spans": [ + { + "bbox": [ + 335, + 134, + 365, + 139 + ], + "type": "text", + "content": "GT Answer: 013684" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 336, + 140, + 413, + 201 + ], + "lines": [ + { + "bbox": [ + 336, + 140, + 413, + 201 + ], + "spans": [ + { + "bbox": [ + 336, + 140, + 413, + 201 + ], + "type": "image", + "image_path": "6a39ea0d7bcfd074e9a0c8e0da7fe9e03ffd767ac2a6e6a69682e06d31f397a3.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 141, + 507, + 157 + ], + "lines": [ + { + "bbox": [ + 407, + 141, + 507, + 157 + ], + "spans": [ + { + "bbox": [ + 407, + 141, + 507, + 157 + ], + "type": "text", + "content": "Step 1 (Question Analysis): The question is about Factoid. And the question asks for the heading of the second paragraph... Step 2 (Relevant Area Concentration): The answer can be found in the area: 134, 280, 654, 3731." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 159, + 496, + 163 + ], + "lines": [ + { + "bbox": [ + 417, + 159, + 496, + 163 + ], + "spans": [ + { + "bbox": [ + 417, + 159, + 496, + 163 + ], + "type": "text", + "content": "Step 3 (Answer Formation): The heading of the second" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 163, + 501, + 167 + ], + "lines": [ + { + "bbox": [ + 417, + 163, + 501, + 167 + ], + "spans": [ + { + "bbox": [ + 417, + 163, + 501, + 167 + ], + "type": "text", + "content": "paragraph is \\*TAKEDA'S RESPONSE AND OBJECTIONS TO" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 168, + 501, + 173 + ], + "lines": [ + { + "bbox": [ + 417, + 168, + 501, + 173 + ], + "spans": [ + { + "bbox": [ + 417, + 168, + 501, + 173 + ], + "type": "text", + "content": "PRACTITIONER'S SUPPLEMENTAL INJURY CLAIMS. ADDITIONAL ACTOS PERSONAL INJURY CLAIMS*." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 174, + 504, + 178 + ], + "lines": [ + { + "bbox": [ + 417, + 174, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 417, + 174, + 504, + 178 + ], + "type": "text", + "content": "Final answer: TAXEDA'S RESPONSE AND OBJECTIONS TO" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 185, + 468, + 189 + ], + "lines": [ + { + "bbox": [ + 417, + 185, + 468, + 189 + ], + "spans": [ + { + "bbox": [ + 417, + 185, + 468, + 189 + ], + "type": "text", + "content": "□" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 191, + 455, + 195 + ], + "lines": [ + { + "bbox": [ + 417, + 191, + 455, + 195 + ], + "spans": [ + { + "bbox": [ + 417, + 191, + 455, + 195 + ], + "type": "text", + "content": "Step 1 (Question Analysis):" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 196, + 507, + 201 + ], + "lines": [ + { + "bbox": [ + 417, + 196, + 507, + 201 + ], + "spans": [ + { + "bbox": [ + 417, + 196, + 507, + 201 + ], + "type": "text", + "content": "Step 2 (Relevant Area Concentration): The answer can be found" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 201, + 468, + 205 + ], + "lines": [ + { + "bbox": [ + 417, + 201, + 468, + 205 + ], + "spans": [ + { + "bbox": [ + 417, + 201, + 468, + 205 + ], + "type": "text", + "content": "in the area: [117, 566, 678, 681]." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 206, + 495, + 210 + ], + "lines": [ + { + "bbox": [ + 417, + 206, + 495, + 210 + ], + "spans": [ + { + "bbox": [ + 417, + 206, + 495, + 210 + ], + "type": "text", + "content": "Step 3 (Answer Formation): The heading of the second paragraph is \"PRESUMMARY STATEMENT\"." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 417, + 211, + 476, + 217 + ], + "lines": [ + { + "bbox": [ + 417, + 211, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 417, + 211, + 476, + 217 + ], + "type": "text", + "content": "Final answer: PRELIMINARY STATEMENT" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 220, + 545, + 252 + ], + "lines": [ + { + "bbox": [ + 406, + 220, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 406, + 220, + 545, + 252 + ], + "type": "text", + "content": "tion with LayoutCoT. Green represents and answers, while Red represents the list viewed in digital version." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "spans": [ + { + "bbox": [ + 336, + 202, + 413, + 208 + ], + "type": "text", + "content": "Question: What is the heading of second paragraph?" + } + ] + } + ], + "index": 39, + "type": "text" + }, + { + "bbox": [ + 336, + 209, + 390, + 213 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 209, + 390, + 213 + ], + "spans": [ + { + "bbox": [ + 336, + 209, + 390, + 213 + ], + "type": "text", + "content": "GT Answer: PRELIMINARY STATEMENT" + } + ] + } + ], + "index": 40, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 484, + 185, + 488, + 188 + ], + "blocks": [ + { + "bbox": [ + 417, + 179, + 511, + 184 + ], + "lines": [ + { + "bbox": [ + 417, + 179, + 511, + 184 + ], + "spans": [ + { + "bbox": [ + 417, + 179, + 511, + 184 + ], + "type": "text", + "content": "PLAINTIPS SUPPLEMENTAL INTERROGATORY 2 REGRADING ADDITIONAL ACTS PERONEAL INJURY CLAIMS" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 185, + 488, + 188 + ], + "lines": [ + { + "bbox": [ + 484, + 185, + 488, + 188 + ], + "spans": [ + { + "bbox": [ + 484, + 185, + 488, + 188 + ], + "type": "image", + "image_path": "55eab25f39acc79a21497930688dbd25324e6abce888c915d9c89a42c6238ec4.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 304, + 263, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 546, + 468 + ], + "type": "text", + "content": "in the image that are relevant to the keyword \"vendor\" of the question. LayoutCoT focused on a wrong area containing \"vendor\", as it missed \"left bottom\" in the question, the answer was incorrect. However, after the right area is given manually, it can finally give the correct answer. Similarly, in Fig. 5(b) the question asks about \"the heading of the second paragraph\". However, the term \"paragraph\" does not have a universal definition, and in this case, the sentences in the area below the main heading were considered as a paragraph, causing the model to predict \"TAKEDA's...CLAIMS\", which was incorrect according to the GT. Once the right \"second paragraph\" region is fed to the model, the answer can be successfully revised. This unique ability of LayoutCoT could be very valuable in high-stake scenarios (e.g., a bank transaction), where the standards are extremely high, and manual checking and correction (i.e., human-in-the-loop) are required." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 306, + 480, + 381, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 381, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 381, + 491 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 620 + ], + "type": "text", + "content": "Through LayoutCoT, LayoutLLM demonstrates the capability of interactive correction, but in real-world applications, this is not enough. The ability to refuse false-positive outputs and generate hints (e.g. \"The answer is not mentioned in the document.\") is crucial. However, it is currently absent in LayoutLLM. In addition, despite achieving notable improvements through layout-aware pre-training, LayoutLLM struggles in precisely understanding region-level relationships, as evidenced in Fig. 5(a). We will study how to endow LayoutLLM with such abilities." + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 378, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We propose LayoutLLM for document understanding, in which a layout instruction tuning strategy comprising layout-aware pre-training and layout-aware SFT is designed to improve the comprehension of document layouts. Extensive experiments confirm the effectiveness of LayoutLLM." + } + ] + } + ], + "index": 59 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15637" + } + ] + } + ], + "index": 60 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 201, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 201, + 101 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 201, + 101 + ], + "type": "text", + "content": "[1] GPT-4V(ision) system card. 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 287, + 136 + ], + "type": "text", + "content": "[2] Srikar Appalaraju, Bhavan Jasani, and Bhargava Urala Kota. DocFormer: End-to-end transformer for document understanding. In ICCV, pages 4171-4186, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 138, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 138, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 138, + 287, + 192 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 226 + ], + "type": "text", + "content": "[4] Galal M Binmakhashen and Sabri A Mahmoud. Document layout analysis: a comprehensive survey. ACM Computing Surveys (CSUR), 52(6):1-36, 2019. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 287, + 282 + ], + "type": "text", + "content": "[5] Haoyu Cao, Changcun Bao, Chaohu Liu, Huang Chen, Kun Yin, Hao Liu, Yinsong Liu, Deqiang Jiang, and Xing Sun. Attention where it matters: Rethinking visual document understanding with selective region concentration. In ICCV, pages 19517-19527, 2023. 2, 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 287, + 349 + ], + "type": "text", + "content": "[6] Hiuyi Cheng, Peirong Zhang, Sihang Wu, Jiaxin Zhang, Qiyuan Zhu, Zecheng Xie, Jing Li, Kai Ding, and Lianwen Jin. M6doc: A large-scale multi-format, multi-type, multilayout, multi-language, multi-annotation category dataset for modern document layout analysis. In CVPR, pages 15138–15147, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 384 + ], + "type": "text", + "content": "[7] Lei Cui, Yiheng Xu, Tengchao Lv, and Furu Wei. Document ai: Benchmarks, models and applications. arXiv preprint arXiv:2111.08609, 2021. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 419 + ], + "type": "text", + "content": "[8] Cheng Da, Chuwei Luo, Qi Zheng, and Cong Yao. Vision grid transformer for document layout analysis. In ICCV, pages 19462-19472, 2023. 1, 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 421, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 286, + 453 + ], + "type": "text", + "content": "[9] Cheng Da, Peng Wang, and Cong Yao. Multi-granularity prediction with learnable fusion for scene text recognition. arXiv preprint arXiv:2307.13244, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[10] Brian Davis, Bryan Morse, Brian Price, Chris Tensmeyer, Curtis Wigington, and Vlad Morariu. End-to-end document recognition and understanding with dessurt. In ECCV, pages 280–296. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 565 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 610 + ], + "type": "text", + "content": "[12] Jieuxiang Gu, Jason Kuen, Vlad I Morariu, Handong Zhao, Nikolaos Barmpalios, Rajiv Jain, Ani Nenkova, and Tong Sun. Unified pretraining framework for document understanding. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 666 + ], + "type": "text", + "content": "[13] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xlayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In CVPR, pages 4583-4592, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In ICDAR, pages 991-995. IEEE, 2015. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Jiabang He, Lei Wang, Yi Hu, Ning Liu, Hui Liu, Xing Xu, and Heng Tao Shen. Icl-d3ie: In-context learning with diverse demonstrations updating for document information extraction. ICCV, 2023. 1, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[16] Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, and Sungrae Park. Bros: A pre-trained language model focusing on text and layout for better key information extraction from documents. In AAAI, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 204 + ], + "type": "text", + "content": "[17] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In ACM Multimedia, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 248 + ], + "type": "text", + "content": "[18] Zheng Huang, Kai Chen, Jianhua He, Xiang Bai, Dimosthenis Karatzas, Shijian Lu, and C. V. Jawahar. ICDAR2019 competition on scanned receipt OCR and information extraction. In ICDAR. IEEE, 2019. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 281 + ], + "type": "text", + "content": "[19] Guillaume Jaume, Hazim Kemal Ekenel, and Jean-Philippe Thiran. Funsd: A dataset for form understanding in noisy scanned documents, 2019. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[20] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In ECCV, pages 498- 517.Springer,2022.2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[21] Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. NeurIPS, 35:22199-22213, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 435 + ], + "type": "text", + "content": "[22] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In ICML, pages 18893–18912. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 468 + ], + "type": "text", + "content": "[23] Chenliang Li, Bin Bi, and Ming Yan. StructuralLM: Structural pre-training for form understanding. In ACL, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[24] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. arXiv preprint arXiv:2006.01038, 2020. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "text", + "content": "[25] Peizhao Li, Jiuxiang Gu, and Jason Kuen. SelfDoc: Self-supervised document representation learning. In CVPR, pages 5652-5660, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 601 + ], + "type": "text", + "content": "[26] Yulin Li, Yuxi Qian, Yuechen Yu, Xiameng Qin, Chengquan Zhang, Yan Liu, Kun Yao, Junyu Han, Jingtuo Liu, and Errui Ding. Structext: Structured text understanding with multimodal transformers. In ACM Multimedia, pages 1912-1920, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 656 + ], + "type": "text", + "content": "[27] Minghui Liao, Zhisheng Zou, Zhaoyi Wan, Cong Yao, and Xiang Bai. Real-time scene text detection with differentiable binarization and adaptive scale fusion. IEEE transactions on pattern analysis and machine intelligence, 45(1):919-931, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 690 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15638" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Yuliang Liu, Zhang Li, Hongliang Li, Wenwen Yu, Yang Liu, Biao Yang, Mingxin Huang, Dezhi Peng, Mingyu Liu, Mingrui Chen, Chunyuan Li, Xucheng Yin, Cheng lin Liu, Lianwen Jin, and Xiang Bai. On the hidden mystery ofOCR in large multimodal models, 2023. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[31] Chuwei Luo, Guozhi Tang, Qi Zheng, Cong Yao, Lianwen Jin, Chenliang Li, Yang Xue, and Luo Si. Bi-vldoc: Bidirectional vision-language modeling for visually-rich document understanding. arXiv preprint arXiv:2206.13155, 2022. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 207 + ], + "type": "text", + "content": "[32] Chuwei Luo, Changxu Cheng, Qi Zheng, and Cong Yao. Geolayoutm: Geometric pre-training for visual information extraction. In CVPR, pages 7092-7101, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 240 + ], + "type": "text", + "content": "[33] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 285 + ], + "type": "text", + "content": "[34] Linyong Nan, Chiachun Hsieh, Ziming Mao, Xi Victoria Lin, Neha Verma, Rui Zhang, Wojciech Krysciński, Hailey Schoelkopf, Riley Kong, Xiangru Tang, et al. Fetaqa: Freeform table question answering. TACL, 10:35-49, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 287, + 287, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 287, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 287, + 287, + 308 + ], + "type": "text", + "content": "[35] OpenAI. Introducing chatgpt. https://openai.com/blog/chatgpt, 2022.1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 311, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 287, + 331 + ], + "type": "text", + "content": "[36] R OpenAI. Gpt-4 technical report. arXiv, pages 2303-08774, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 334, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 334, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 334, + 287, + 387 + ], + "type": "text", + "content": "[37] Seunghyun Park, Seung Shin, Bado Lee, Junyeop Lee, Jaeheung Surh, Minjoon Seo, and Hwalsuk Lee. {CORD}: A consolidated receipt dataset for post-{ocr} parsing. In Workshop on Document Intelligence at NeurIPS 2019, 2019. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 389, + 287, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 389, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 48, + 389, + 287, + 455 + ], + "type": "text", + "content": "[38] Qiming Peng, Yinxu Pan, Wenjin Wang, Bin Luo, Zhenyu Zhang, Zhengjie Huang, Teng Hu, Weichong Yin, Yongfeng Chen, Yin Zhang, Shikun Feng, Yu Sun, Hao Tian, Hua Wu, and Haifeng Wang. Ernie-Layout: Layout knowledge enhanced pre-training for visually-rich document understanding. In EMNLP findings, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 511 + ], + "type": "text", + "content": "[39] Vincent Perot, Kai Kang, Florian Luisier, Guolong Su, Xiaoyu Sun, Ramya Sree Boppana, Zilong Wang, Jiaqi Mu, Hao Zhang, and Nan Hua. Lmdx: Language model-based document information extraction and localization. arXiv preprint arXiv:2309.10952, 2023. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 513, + 287, + 555 + ], + "type": "text", + "content": "[40] Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. DocLayNet: A large human-annotated dataset for document layout segmentation. In SIGKDD. ACM, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 612 + ], + "type": "text", + "content": "[41] Rafal Powalski, Łukasz Borchmann, Dawid Jurkiewicz, Tomasz Dwojak, Michal Pietruszka, and Gabriela Płka. Going full-tilt boogie on document understanding with text-image layout transformer. In ICDAR, pages 732-747. Springer, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[43] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for train" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "ing next generation image-text models. NeurIPS, 35:25278-25294, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 151 + ], + "type": "text", + "content": "[44] Baoguang Shi, Xiang Bai, and Cong Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE transactions on pattern analysis and machine intelligence, 39(11):2298-2304, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 545, + 207 + ], + "type": "text", + "content": "[45] Yongxin Shi, Dezhi Peng, Wenhui Liao, Zening Lin, Xinhong Chen, Chongyu Liu, Yuyi Zhang, and Lianwen Jin. ExploringOCR capabilities of GPT-4V(ision): A quantitative and in-depth evaluation. arXiv preprint arXiv:2310.16809, 2023. 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 253 + ], + "type": "text", + "content": "[46] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In ECCV, pages 742-758. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 255, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 288 + ], + "type": "text", + "content": "[47] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 290, + 545, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 290, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 308, + 290, + 545, + 344 + ], + "type": "text", + "content": "[48] Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/stanford_alpaca, 2023.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 346, + 545, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 346, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 308, + 346, + 545, + 412 + ], + "type": "text", + "content": "[49] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 415, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 643 + ], + "type": "text", + "content": "[50] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Viktor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models, 2023. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[51] Yi Tu, Ya Guo, Huan Chen, and Jinyang Tang. LayoutMask: Enhance text-Layout interaction in multi-modal pre-training for document understanding. In ACL, 2023. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 712 + ], + "type": "text", + "content": "[52] Jiapeng Wang, Lianwen Jin, and Kai Ding. Lilt: A simple yet effective language-independent layout transformer for structured document understanding. In ACL, 2022. 1, 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15639" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 708 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 116 + ], + "type": "text", + "content": "[53] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 286, + 150 + ], + "type": "text", + "content": "[54] Yiheng Xu, Minghao Li, Lei Cui, and Shaohan Huang. LayoutLM: Pre-training of text and layout for document image understanding. In KDD, pages 1192-1200, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "type": "text", + "content": "[55] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. Layoutmv2: Multi-modal pre-training for visually-rich document understanding. In ACL, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 251 + ], + "type": "text", + "content": "[56] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with GPT-4V(ision). arXiv preprint arXiv:2309.17421, 9, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 253, + 286, + 306 + ], + "type": "text", + "content": "[57] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Yuhao Dan, Chenlin Zhao, Guohai Xu, Chenliang Li, Junfeng Tian, et al. mplug-docowl: Modularized multimodal large language model for document understanding. arXiv preprint arXiv:2307.02499, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 308, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 286, + 373 + ], + "type": "text", + "content": "[58] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng Tian, Qian Qi, Ji Zhang, and Fei Huang. mplug-owl: Modularization empowers large language models with multimodality, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 375, + 286, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 375, + 286, + 428 + ], + "spans": [ + { + "bbox": [ + 49, + 375, + 286, + 428 + ], + "type": "text", + "content": "[59] Yuechen Yu, Yulin Li, Chengquan Zhang, Xiaogiang Zhang, Zengyuan Guo, Xiameng Qin, Kun Yao, Junyu Han, Errui Ding, and Jingdong Wang. Structextv2: Masked visual-textual prediction for document image pre-training. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 431, + 286, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 431, + 286, + 474 + ], + "spans": [ + { + "bbox": [ + 49, + 431, + 286, + 474 + ], + "type": "text", + "content": "[60] Yanzhe Zhang, Ruiyi Zhang, Jiuxiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 476, + 286, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 476, + 286, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 476, + 286, + 529 + ], + "type": "text", + "content": "[61] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 286, + 563 + ], + "type": "text", + "content": "[62] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: largest dataset ever for document layout analysis. In ICDAR, pages 1015-1022. IEEE, 2019. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 566, + 286, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 566, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 49, + 566, + 286, + 597 + ], + "type": "text", + "content": "[63] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In ECCV, pages 564-580. Springer, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 600, + 286, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 600, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 49, + 600, + 286, + 653 + ], + "type": "text", + "content": "[64] Xinyu Zhou, Cong Yao, He Wen, Yuzhi Wang, Shuchang Zhou, Weiran He, and Jiajun Liang. East: an efficient and accurate scene text detector. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 5551-5560, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 655, + 286, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 655, + 286, + 708 + ], + "spans": [ + { + "bbox": [ + 49, + 655, + 286, + 708 + ], + "type": "text", + "content": "[65] Stepan Simsa, Milan Sulc, Michal Uričář, Yash Patel, Ahmed Hamdi, Matej Kocián, Matyás Skalický, Jiří Matas, Antoine Doucet, Mickaël Courstaty, and Dimosthenis Karatzas. Docile benchmark for document information localization and extraction, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15640" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_content_list.json b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..73be9ce52e972383da1e3d1ec34155dc0a77038c --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_content_list.json @@ -0,0 +1,1645 @@ +[ + { + "type": "text", + "text": "LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example", + "text_level": 1, + "bbox": [ + 78, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Soyeon Yoon* Kwan Yun* Kwanggyoon Seo Sihun Cha Jung Eun Yoo Junyong Noh KAIST, Visual Media Lab", + "bbox": [ + 107, + 203, + 854, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{thoyeony, yunandy, skg1023, chacorp, jey920, junyongnoh}@kaist.ac.kr", + "bbox": [ + 174, + 244, + 784, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38c9f02eb51e5361cf0e14ca87fb34a950c377f46f76ae34436680573ddeed2b.jpg", + "image_caption": [ + "(a) Diverse deformation targets and output of LeGO" + ], + "image_footnote": [], + "bbox": [ + 124, + 297, + 488, + 512 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e22b3963195b072c0d67965648471e991e66d2458316c9b8f154107a849ad188.jpg", + "image_caption": [ + "(b) Facial animation of stylized avatar", + "Figure 1. (a) The proposed method demonstrates robustness to unseen face identities and topologies and effectively generates stylized output faces with desired topologies. (b) Our stylized avatars can be animated using 3DMM blend shapes." + ], + "image_footnote": [], + "bbox": [ + 511, + 315, + 898, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 575, + 313, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D face stylization have made significant strides in few to zero-shot settings. However, the degree of stylization achieved by existing methods is often not sufficient for practical applications because they are mostly based on statistical 3D Morphable Models (3DMM) with limited variations. To this end, we propose a method that can produce a highly stylized 3D face model with desired topology. Our methods train a surface deformation network with 3DMM and translate its domain to the target style using a paired exemplar. The network achieves stylization of the 3D face mesh by mimicking the style of the target using a differentiable renderer and directional CLIP losses. Additionally, during the inference process, we utilize a Mesh Agnostic Encoder (MAGE) that takes deformation target, a mesh of diverse topologies as input to the stylization process and encodes its shape into our latent space. The resulting stylized face model can be animated by commonly used 3DMM blend shapes. A set of quantitative and qualitative evaluations demonstrate that our method can produce", + "bbox": [ + 75, + 613, + 472, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "highly stylized face meshes according to a given style and output them in a desired topology. We also demonstrate example applications of our method including image-based stylized avatar generation, linear interpolation of geometric styles, and facial animation of stylized avatars.", + "bbox": [ + 496, + 578, + 893, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 669, + 632, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crafting animatable stylized 3D avatars that encapsulate both personal identity and character style requires extensive efforts from skilled artists. When creating animated films, the artists design stylized 3D avatars whose facial appearance matches the theme of the entire content while putting careful effort into preserving the idiosyncrasy of the actors. Similarly, on social media, artists create numerous stylized presets so that the combinations of these presets can represent diverse identities.", + "bbox": [ + 496, + 694, + 890, + 829 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To reduce the burden of manual crafting effort, generating stylized 3D faces has been a prominent area of research. Recent attempts include 3D-aware generative ad", + "bbox": [ + 496, + 830, + 890, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work", + "bbox": [ + 517, + 886, + 769, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4505", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "versarial networks (GANs) and denoising diffusion models (DMs) [1, 12, 15, 18, 46], generating a stylized texture for the 3D morphable model (3DMM) [2, 47], text-based geometry deformation [7, 22, 24-26], and surface deformation methods [13, 43]. These methods have successfully demonstrated the possibility of producing high-quality and diverse stylized 3D faces, although each method has distinct challenges.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify three key elements for stylized avatar creation so that the results can be practically useful.", + "bbox": [ + 76, + 212, + 468, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Avatar creation in a desired topology that is compatible with conventional CG pipelines.", + "2. Extending stylization capabilities beyond 3DMM.", + "3. Generating stylized avatars that are animatable using blend shapes." + ], + "bbox": [ + 76, + 242, + 468, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Where each component meets the standards and demands of the dynamic entertainment landscape 1) allowing creators to reuse existing animation rigs and texture maps across different models.; 2) achieving greater diversity and flexibility in expressing unique and non-conventional characters.; 3) enabling animators to achieve coherent and natural movements across various facial features and expressions. While recent methods have made meaningful strides in one or two aspects of these key elements, there has not been a method that satisfies all three elements, as summarized in Figure 2.", + "bbox": [ + 75, + 316, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7b4ce791e7fd7bdb263df39ef1a0b9397188fbcf2ac5bc4e333a4b1a4226b2ad.jpg", + "image_caption": [ + "Figure 2. Comparison of different stylized 3D face generation methods and their limitations in meeting key elements. 3D-aware methods cannot generate 3D face in desired topologies. 3DMM-based methods have a limited stylization capability. Text-based deformation models are not directly animatable. The proposed method meets the goal of all three components." + ], + "image_footnote": [], + "bbox": [ + 58, + 482, + 467, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose a novel method that can generate stylized 3D face meshes. This is achieved by translating the domain of a pre-trained surface deformation network based on one of the most widely used 3DMM model, FLAME [19], to a target style domain. We achieve this goal by first training the surface deformation network with the FLAME decoder to leverage its linear shape space combined with global expression space. During fine-tuning, we employ a directional CLIP-based domain adaptation method [16, 35, 48], widely used in 2D domain, to retain the face identity while reflecting the desired style.", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, to seamlessly integrate this 2D-based training method into the stylized 3D face generation task, we propose a hierarchical rendering scheme that captures local and global facial features, ensuring effective training and identity preservation. In the inference stage, we introduce a Mesh Agnostic Encoder (MAGE) to enable mesh agnostic stylization for an input, which we call a deformation target that has various mesh topologies. MAGE is composed of pre-trained encoders from Neural Face Rigging (NFR) [34] and latent mapping networks, which establish correspondences between shapes by encoding mesh representations into a topology-invariant latent space. This enhances the versatility and applicability of our approach in the context of 3D face stylization. As shown in Figure 1, our method can generate stylized 3D face models with varied mesh topologies that are equipped with the animation capability of 3DMM while ensuring consistency across diverse deformation target mesh representations.", + "bbox": [ + 496, + 90, + 893, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 378, + 640, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Stylized 3D face generation", + "text_level": 1, + "bbox": [ + 500, + 402, + 746, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With recent advances in the GANs and DMs that utilize neural fields for 3D-aware face generation [3, 10], the generation of faces in diverse styles has gained popularity [1, 12, 15, 18]. By leveraging 2D priors from generative models, which capture various patterns and variations observed in the extensive 2D training data, these methods can generate 3D faces with various styles. However, despite their success in producing consistent multi-view images through neural rendering, creating stylized faces in a desired topology is challenging, limiting their suitability for using existing graphics tools across different faces. On the other hand, 3DMM-based personalized 3D face generation methods with text-guidance [2, 47] have shown success in producing high-fidelity stylized textures for 3D face models. However, these methods confine the shape of the generated face models within the 3DMM shape space, constraining possibilities for geometric exaggerations or abstraction beyond the training data.", + "bbox": [ + 496, + 426, + 890, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Learning-Based 3D Shape Network", + "text_level": 1, + "bbox": [ + 500, + 710, + 807, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, learning implicit functions for 3D shapes has demonstrated remarkable performance in representing complex geometric structures [8, 23, 27, 30, 31]. In particular, DIF-Net [5] adopted MLPs for learning a standard signed distance function (SDF) and a volumetric deformation function, leading to comprehensive mapping between the produced SDFs. Most related to our work, DD3C [13] utilized template deformation for 3D caricature auto-decoder. They found that modeling each shape as a deformation of a fixed temaplate surface effective compared to absolute position. We advance a step further by training a surface deformation", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4506", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network on 3DMM and transfer its domain into stylization.", + "bbox": [ + 76, + 90, + 467, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another line of research aims to transfer shape deformation [39]. Recently, attempts have been made to transfer deformation utilizing examples [40, 43, 45] or text [7, 24, 25]. These methods can deform a mesh into a desired style and identity using a given example. However these deformation methods may not be as effective when it comes to coherent control of animation.", + "bbox": [ + 75, + 107, + 467, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Few-Shot Domain Adaptation", + "text_level": 1, + "bbox": [ + 76, + 224, + 341, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain adaptation refers to the process of utilizing a neural network that has been initially trained on a large dataset from a source domain, and subsequently fine-tuning it on a smaller dataset from a target domain. Few-shot domain adaptation is widely researched in the 2D domain, especially with generative priors [4, 12, 20, 28, 29, 41, 42, 44]. Attempts have been made in one-shot domain adaptation to utilize the semantic capabilities of vision-language networks like CLIP [35]. These networks can provide direction for distinguishing between identity and style [6, 16, 48]. We adopt this two-way directional guidance from CLIP and combine it with a differentiable renderer to effectively stylize the face mesh while preserving the original identity using a paired exemplar.", + "bbox": [ + 75, + 250, + 467, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Mesh Agnostic Deformation", + "text_level": 1, + "bbox": [ + 76, + 473, + 328, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, mesh agnostic networks [9, 11, 32, 33, 38] demonstrated a great potential for learning 3D information such as dense correspondences between different 3D shape representations without relying on consistent topology or vertex ordering. These networks operate on meshes and encode shapes into a topology-invariant latent space. Specifically, NFR [34] proposed an autoencoder framework for facial expression retargeting across different mesh topologies. The framework utilizes separate expression and identity encoders, both functioning in a topology-agnostic manner.", + "bbox": [ + 75, + 497, + 467, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We integrate components of this approach into our method by mapping their embeddings into our latent space. This network translates the output of the pre-trained encoder from NFR to the latent space of our surface deformation network. This enables any geometric shape of the deformation target to be encoded and fed into our surface deformation network.", + "bbox": [ + 75, + 648, + 467, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 771, + 166, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our research focuses on training and fine-tuning a surface deformation network to generate stylized 3D faces with diverse shapes and expressions. We start with the source face deformation network $D_{S}$ that deforms the template face to a face with different identities and expressions. Thereafter, we fine-tune it into a target style face deformation network $D_{T}$ using a paired exemplar. This process is outlined below.", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We first train $D_{S}$ using FLAME in a self-supervised manner, enabling the creation of versatile head meshes with different shapes and expressions.", + "2. For fine-tuning, we assemble a paired exemplar that consists of an identity exemplar mesh $M_S$ and a style exemplar mesh $M_T$ , sharing identity.", + "3. Both $M_S$ and $M_T$ jointly serve as example guidance to fine-tune $D_T$ .", + "4. At inference, using MAGE and $D_{T}$ , the deformation target of diverse topologies can be translated into a stylized face." + ], + "bbox": [ + 500, + 90, + 890, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A detailed description of these processes will be provided in the following subsections.", + "bbox": [ + 500, + 258, + 890, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Deformation Network as Parametric Model", + "text_level": 1, + "bbox": [ + 500, + 297, + 867, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$D_{S}$ is a surface deformation network that deforms a template face with the given latent vectors. We train the network to generate diverse human faces with FLAME, in a self-supervising manner. FLAME can manipulate both the global head shape and local expressions through its shape parameters $\\vec{\\beta} \\in \\mathbb{R}^{300}$ , and expression parameters $\\vec{\\psi} \\in \\mathbb{R}^{100}$ , which are the components of the FLAME parameters $\\Phi$ . This empowers $D_{S}$ to generate diverse geometric face shapes and expressions.", + "bbox": [ + 498, + 321, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate face meshes using these shape and expression components of $\\Phi$ as input, we employ mapping networks, $\\mathcal{M}_{shape}$ and $\\mathcal{M}_{exp}$ , which consist of MultiLayer Perceptrons (MLP). Each mapping network separately transforms $\\vec{\\beta}$ and $\\vec{\\psi}$ into latent vectors, $z_{s}$ and $z_{e}$ , respectively.", + "bbox": [ + 498, + 462, + 890, + 554 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {s} = \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}), \\quad z _ {e} = \\mathcal {M} _ {\\text {e x p}} (\\vec {\\psi}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 555, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These latent vectors are then fed into $D_{S}$ , enabling the creation of diverse face meshes imbued with a wide range of expressions. The mapping networks $\\mathcal{M}_{shape}$ and $\\mathcal{M}_{exp}$ are trained jointly with $D_{S}$ during the training process. The following losses are employed:", + "bbox": [ + 498, + 580, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} (\\vec {\\beta}, \\vec {\\psi}) = \\left\\| \\operatorname {F L A M E} (\\vec {\\beta}, \\vec {\\psi}) - D _ {S} \\left(\\left[ \\begin{array}{l} \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}) \\\\ \\mathcal {M} _ {\\exp} (\\vec {\\psi}) \\end{array} \\right]\\right) \\right\\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 667, + 890, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Surface-Intensive Mesh Sampling To enable $D_{S}$ to represent various geometries, we introduce a surface-intensive mesh sampling (SIMS) strategy during training. SIMS is performed by randomly selecting points from the surfaces of the face mesh, $FLAME(\\vec{\\beta},\\vec{\\psi})$ , during training $D_{S}$ . Specifically, we sample approximately 4 times more points from the surface than the number of vertices of the face mesh. Experiment results are reported in Sec. 4.2.", + "bbox": [ + 498, + 722, + 890, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LeGO", + "text_level": 1, + "bbox": [ + 500, + 854, + 584, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The fine-tuning process utilizes a paired exemplar, $M_S$ and $M_T$ , to guide the adaptation of $D_T$ for the generation of", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4507", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c1ce92d59794f03772abbfd385e751411f605deb0c55f32b5e6912dd137cd0b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 84, + 890, + 203 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c9823072a971478dd9d876b739437c5e189760037433427a52ae61e8cd35bfe3.jpg", + "image_caption": [ + "Figure 3. Overview of our method: The upper box illustrates the inference stage, where our method takes diverse deformation targets and generates stylized outputs. In the lower-left box, the training process of Mesh Agnostic Encoder (MAGE) is depicted. In the lower-right box, the fine-tuning process of $D_{T}$ is illustrated." + ], + "image_footnote": [], + "bbox": [ + 80, + 204, + 483, + 425 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aecb9ef82435c93ea841f65e8d8dc807b605c1e57accf11bd2365acd633e6765.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 204, + 890, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "stylized 3D faces. An identity exemplar mesh, $M_S$ , is generated from the FLAME decoder using $\\Phi_{ref}$ , and a style exemplar mesh, $M_T$ , is manually crafted from $M_S$ . During each iteration of fine-tuning, $\\Phi_{sample}$ is sampled randomly to generate face mesh $M_S^{samp}$ from $D_S$ and corresponding stylized 3D face $M_T^{samp}$ from $D_T$ . In the fine-tuning process, quantities of $M_S$ , $M_T$ , $D_S$ , and $D_T$ are all singular, representing one pair of meshes and one pair of networks.", + "bbox": [ + 75, + 484, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a hierarchical rendering scheme designed to preserve semantically important features, such as the shape and facial components, in the face mesh. With a differentiable renderer, this approach captures significant facial features from both local and global perspectives, enhancing stylization fidelity and identity preservation. To further enhance this process, we incorporate 2D-based losses, including CLIP reconstruction loss $L_{\\mathrm{CLIP}}$ , CLIP indomain loss $L_{in}$ , CLIP across-domain loss $L_{across}$ , and 3D-based losses such as vertex reconstruction loss $L_{vert}$ and style loss $L_{style}$ . Details of each loss are elaborated in Section 3.3.", + "bbox": [ + 75, + 607, + 470, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Functions", + "text_level": 1, + "bbox": [ + 76, + 800, + 228, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vertex Reconstruction Loss The vertex reconstruction loss $L_{vert}$ guides $D_T$ in learning to deform a style exemplar mesh $M_T$ from $[z_s^{ref}, z_e^{ref}]$ , which is mapped from $\\Phi_{ref}$ . The loss utilizes Mean Squared Error (MSE) to ensure that the vertices of the predicted mesh $M_T^*$ , generated by $D_T$ ,", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "closely match $M_T$ . The loss can be written as follows:", + "bbox": [ + 498, + 484, + 859, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {v e r t} = \\left\\| M _ {T} - M _ {T} ^ {*} \\right\\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 512, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP Reconstruction Loss The CLIP reconstruction loss $L_{\\mathrm{CLIP}}$ serves to maintain semantic consistency between the deformed mesh $M_T^*$ and $M_T$ in the CLIP space. This is crucial because even minor displacements in 3D space can result in surface irregularities or undesirable shading variations. The CLIP reconstruction loss can be written as follows:", + "bbox": [ + 496, + 545, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {C L I P}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {*}\\right)\\right) \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 656, + 890, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$R_{l,v}$ represents the hierarchical rendering of a differentiable renderer from level $l$ and view direction $v$ . Each view is anchored at a predefined position on a face mesh, including the front and both sides of the face. Each level corresponds to the distance from the face to the camera. At the highest level, where the face is captured in close-up, additional images are rendered from significant facial features such as the nose, eyes, and lips. $E_C$ encodes rendered images into CLIP embeddings.", + "bbox": [ + 496, + 699, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP Directional Loss Inspired from one-shot stylization methods in the 2D domain [48], we incorporate the CLIP in-domain loss $L_{in}$ and the CLIP across-domain loss", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4508", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{across}$ . $L_{in}$ ensures that the direction of two distinct faces in the source domain remains consistent in two distinct stylized faces of the target style domain. In contrast, $L_{across}$ ensures that the direction of a face from the source domain and its corresponding stylized face in the target domain is preserved across different faces and their corresponding stylized faces. This can be shown in the lower right of Figure 3. Therefore our losses for CLIP in-domain and across-domain can be written as follows:", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {i n}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\tag {5} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {s a m p}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) \\right\\| _ {2} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 238, + 468, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {a c r o s s}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\right. \\tag {6} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right)\\right) \\| _ {2} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 301, + 468, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_S^{samp}$ and $M_T^{samp}$ refer to $D_S[z_s^{samp};z_e^{samp}]$ and $D_T[z_s^{samp};z_e^{samp}]$ , respectively, where $[z_s^{samp};z_e^{samp}]$ is the latent vectors mapped from $\\Phi_{sample}$ .", + "bbox": [ + 76, + 358, + 468, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Style Loss We introduce a novel style loss $L_{style}$ to capture the style of $M_T$ by utilizing surface normals. Style loss compares surface normal, which are strongly correlated with the semantic information of the mesh [5]. A straightforward approach would involve comparing $M_T$ with $M_T^{samp}$ . However, this will constrain $M_T^{samp}$ to have same expression as $M_T$ , which could degrade the animatability. To address this, we construct a pseudo pair for normal calculations. This is done by generating a face mesh through the concatenation of $z_e^{ref}$ from $M_T$ and $z_s^{samp}$ . This approach is utilized to compute the style loss, $L_{style}$ without degrading the animatability while enhancing style adherence. In short, $L_{style}$ ensures alignment of the surface normals from $D_T([z_s^{samp}; z_e^{ref}])$ with $M_T$ . The formulation of the style loss can be written as follows:", + "bbox": [ + 75, + 422, + 472, + 650 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s t y l e}} = \\sum_ {f \\in S} \\left(1 - \\frac {n _ {f} \\cdot n _ {f} ^ {\\prime}}{\\left| n _ {f} \\right| \\left| n _ {f} ^ {\\prime} \\right|}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 660, + 468, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_f$ refers to the surface normal from $M_T$ while $n'_f$ refers to the corresponding normal from $D_T([z_s^{samp}; z_e^{ref}])$ .", + "bbox": [ + 76, + 705, + 467, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final objective function can be written as follows:", + "bbox": [ + 96, + 736, + 452, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {v e r t} L _ {v e r t} + \\lambda_ {\\mathrm {C L I P}} L _ {\\mathrm {C L I P}} + \\lambda_ {i n} L _ {i n} \\tag {8} \\\\ + \\lambda_ {a c r o s s} L _ {a c r o s s} + \\lambda_ {s t y l e} L _ {s t y l e} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 761, + 468, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{vert}$ , $\\lambda_{\\mathrm{CLIP}}$ , $\\lambda_{in}$ , $\\lambda_{across}$ , and $\\lambda_{style}$ are the weight for each loss term.", + "bbox": [ + 76, + 806, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mesh Agnostic Encoder", + "text_level": 1, + "bbox": [ + 76, + 847, + 297, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NFR [34] employs DiffusionNet [38] for encoding and decoding face meshes, facilitating face retargeting across", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different topologies. Given NFR's proficiency in extracting identity and expression details from faces with varying topologies, we extend its encoders to create a MAGE, as depicted in Figure 3. MAGE consists of ID2ID and exp2exp, both being MLPs that receive embeddings from each ID encoder and expression encoder pre-trained in NFR. Finally, given intermediate vectors from both ID2ID and exp2exp, the latent mapper outputs the latent vectors for $D_{S}$ .", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MAGE is trained in a self-supervised manner by randomly sampling $\\beta$ and $\\psi$ and passing them through the mapping networks of $D_{S}$ to obtain $[z_{s};z_{e}]$ . Training involves comparing $[\\hat{z}_s;\\hat{z}_e]$ , predicted from MAGE, with their corresponding ground truth latent vectors $[z_s;z_e]$ . The objective function of the encoder uses a MSE loss as follows:", + "bbox": [ + 496, + 212, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {e n c}} = \\left\\| \\left[ z _ {s}; z _ {e} \\right] - \\left[ \\hat {z} _ {s}; \\hat {z} _ {e} \\right] \\right\\| _ {2} ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 316, + 890, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[ \\hat{z}_s; \\hat{z}_e ]$ is MAGE $[D_S([z_s;z_e])]$ . With this encoder, we can project a face mesh with diverse topologies into the latent space of $D_S$ .", + "bbox": [ + 496, + 342, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 500, + 398, + 609, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training, $D_{T}$ is capable of generating a stylized 3D face mesh with a desired topology from a deformation target mesh that has an arbitrary topology. This is accomplished by first projecting the deformation target into a latent vector, using MAGE. This latent vector is then fed as input to $D_{T}$ , allowing it to deform the template face of the desired topology into a stylized 3D face mesh. The resulting stylized face preserves both the identity of the deformation target and the desired style while providing animation control.", + "bbox": [ + 496, + 422, + 890, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 574, + 633, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 595, + 709, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method based on three key elements for stylized 3D head avatar creation: Avatar creation in a desired topology, stylization beyond 3DMM, and animation capability with blend shapes. Also, implementation details are provided in supplementary material.", + "bbox": [ + 496, + 614, + 890, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Avatar in a Desired Topology To evaluate the capability of our method in producing a mesh with a desired topology from the deformation target of arbitrary topologies, we first apply a re-meshing technique that includes Loop subdivision [21] and mesh simplification [37] to meshes from FLAME. We also use original meshes from the CoMA [36] and ICT-FaceKit [17] directly as additional diverse meshes. To visually confirm that our method can align different mesh topologies with a desired template face, we enhance the resulting meshes with a checkerboard texture. The visual results, as shown in Figure 4, demonstrate that our method consistently produces stylized 3D faces with the desired topology, regardless of variations in the topology of the deformation target.", + "bbox": [ + 496, + 704, + 890, + 915 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4509", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1b383397d4ace2329250b51d7f5ed7551c817793476eef4fffb332c323b633f.jpg", + "image_caption": [ + "Figure 4. Demonstration of stylized 3D faces with desired topology, regardless of deformation target variations." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 467, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Stylization Capability We evaluate our method across various styles to assess its stylization capability. As shown in Figure 5, our method can generate a broad spectrum of styles encompassing human-like and non-human-like faces while preserving the original identities of the given face mesh. Our method also demonstrates the capability to achieve stylization across various geometries, including face masks and point clouds.", + "bbox": [ + 75, + 446, + 470, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b8b9308a7bd4554173aa44cc65e74da2372f3ffc683567f71653c0c3528b4e97.jpg", + "image_caption": [ + "Figure 5. Stylization results across diverse styles and identities. Our approach generates varied styles while preserving the deformation target identity and generalizing to diverse geometric representations like masks and point clouds." + ], + "image_footnote": [], + "bbox": [ + 81, + 583, + 467, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Animation Capability Using parameters from 3DMM, our method can animate facial expressions in the resulting stylized 3D faces. As illustrated in Figure 6, our method can generate 3D stylized faces with various expressions. This allows applications such as video-driven stylized talking heads, which are elaborated in supplementary material.", + "bbox": [ + 498, + 90, + 890, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3693ca614ca6b7633101a8f131125b1f858b3d45cfac1212c7561c87f0b5ea4b.jpg", + "image_caption": [ + "Figure 6. Visualization of dynamic expressions in stylized 3D faces." + ], + "image_footnote": [], + "bbox": [ + 504, + 196, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 428, + 720, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a quantitative evaluation, we measured the average CLIP style preservation (CLIP-SP) and CLIP identity preservation (CLIP-IP) scores. These metrics reflect the trade-off between preserving style and identity, making their averages crucial for validating stylization. For both metrics, we calculated the cosine similarity of CLIP embeddings from rendered meshes. Specifically, for CLIP-SP, we compared the embeddings of the generated stylized face mesh and style exemplar mesh. For CLIP-IP, we compared the embeddings of the generated stylized face mesh and deformation target mesh. For dataset, we sampled 8 different face mesh from FLAME without expression and manually crafted stylized mesh corresponding to each identities. Also, for the quantitative evaluation, we randomly sampled another 10 different identities without expression to be used as deformation targets.", + "bbox": [ + 496, + 452, + 890, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with Baselines on Mesh Stylization We compared our stylized face generation results with those produced by baselines that can deform a mesh into different styles: Deformation Transfer [39], TextDeformer [7], and X-mesh [22]. In case of Deformation Transfer [39], we identified the correspondences between the identity exemplar mesh and the deformation target using 68 facial landmarks along with 9 additional points on the forehead in order to apply learned deformation. TextDeformer and X-mesh employ a text-guided deformation that operates on an input mesh. For both methods, the input face is deformed based on the same descriptive text to generate stylized outputs. All these baselines were evaluated on three different", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4510", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af9a6089dd28008237c4a66e5b817fdfb96024e6cbcefece5f002fd961c9335e.jpg", + "table_caption": [ + "Table 1. Quantitative results are presented for comparison with baselines and ablations, featuring averages of CLIP-SP and CLIP-IP. The highest scores are denoted in bold, while the second highest scores are underlined." + ], + "table_footnote": [], + "table_body": "
Mesh TypeFLAMELoopSimplifiedOverall-Average
MethodsCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageAverage
Ours0.98670.93470.96070.98330.93500.95920.95590.89870.92730.9491
Ours w/o Lstyle0.98090.93740.95910.97680.93390.95530.95420.89870.92650.9470
Ours w direct Lstyle0.98650.93480.96070.98330.93560.95950.95500.89730.92620.9488
w/o hierarchical rendering0.98460.93490.95970.98240.93340.95790.95690.89850.92770.9484
Text deformer0.86410.83630.85020.86300.84250.85280.85340.81730.83530.8461
X-mesh0.86260.86020.86140.88170.89500.88840.86350.84480.85410.8680
Deformation transfer0.97680.93760.95720.95980.95760.95870.92540.92880.92710.9477
", + "bbox": [ + 78, + 123, + 893, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mesh topologies: FLAME and FLAME with Loop subdivision and mesh simplification.", + "bbox": [ + 75, + 263, + 468, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present qualitative comparison results on stylized 3D face generation in Figure 7. Deformation Transfer is capable of generating faces that follow the style; however, it exhibits artifacts, including anomalies and concavities on the surface due to the calculation of displacement and its direct transfer. In contrast, text-based methods fall short in following styles. Our method, on the other hand, can generate stylized 3D faces that adhere to both the input identity and style example without any artifacts.", + "bbox": [ + 75, + 294, + 468, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In quantitative evaluation, shown in Table 1, our method obtained the highest average scores computed from CLIP-SP and CLIP-IP. This clearly demonstrates that our approach achieves better performance in mesh stylization while preserving input identity compared to deformation baselines. More details and examples are provided in the supplementary material.", + "bbox": [ + 75, + 429, + 470, + 536 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49ae9faeb5d6783b98f9bcdac24cfd62f117a32575ccc4a4e3b01af0baa00a24.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison on stylization. Our method adheres to style and identity without artifacts unlike Deformation Transfer. Text-based approaches fail to match styles from text." + ], + "image_footnote": [], + "bbox": [ + 84, + 547, + 467, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Surface-Intensive Mesh Sampling To test the proposed sampling method SIMS, we conducted a study on the reconstruction task comparing three different sampling variants for training $D_{S}$ . SIMS is surface-intensive sampling, using approximately four times more points sampled from the mesh surface compared to mesh vertices. Hybrid sampling, originally proposed in DD3C [13], sampled points both ver", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/efde3f5507e568cc21ea0a3cfe05dab14c8fa831e7d54e59b26bdaaae6fdf9e5.jpg", + "table_caption": [ + "Table 2. Reconstruction comparison with mesh sampling methods" + ], + "table_footnote": [], + "table_body": "
Reconstruction loss↓ (all in e-5)
Methodsoriginalsimplifiedloop-1loop-2average
SIMS(Ours)1.7901.6351.6121.5911.657
Hybrid2.3002.0192.1382.1182.144
Vertex1.9555.8663.2582.9593.509
", + "bbox": [ + 514, + 282, + 874, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tices and faces in similar ratio ( $\\sim 1.1$ times). Vertex-only used just the vertices of the mesh.", + "bbox": [ + 498, + 375, + 890, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the experiment, we calculated the reconstruction loss on four different experiment settings: (1) original FLAME, (2) a simplified FLAME mesh with 1/4 the vertex count, (3) FLAME with one loop subdivision, and (4) FLAME with two loop subdivisions. The reconstruction loss was compared to the ground truth position using mean squared error. As shown in Table 2, SIMS achieved the lowest error by a large margin on all experiments. Hybrid sampling generally performed better than vertex-only sampling.", + "bbox": [ + 498, + 405, + 890, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study To perform an ablation study, we trained our model $D_{T}$ using different settings: 1) \"Ours w/o $L_{\\text{style}}$ ,\" where the style loss was removed; 2) \"Ours w direct $L_{\\text{style}}$ ,\" where the style loss was directly compared between $D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}])$ and $M_{T}$ ; 3) \"Ours w/o hierarchical rendering,\" where a single canonical view was used to calculate the $L_{\\text{in}}$ and $L_{\\text{across}}$ . The results are displayed in Figure 8 and Table 1. The results produced by \"Ours\" and \"Ours w direct $L_{\\text{style}}$ \" do not show any artifacts unlike the results produced by w/o $L_{\\text{style}}$ and w/o hierarchical rendering, which exhibit surface artifacts. However, \"Ours w. direct $L_{\\text{style}}$ \" forces $D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}])$ and $M_{T}$ , which have different expressions, to have the same normals; thus, it discards the expression. As a result, while \"Ours w. direct $L_{\\text{style}}$ \" does stylize effectively, it cannot animate. Conversely, \"Ours\" not only stylizes well but is also animatable. Additional ablation study results regarding animation are presented in the supplementary material.", + "bbox": [ + 498, + 550, + 892, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study We conducted a user study with 34 participants to evaluate different stylized face generation methods on human perception. A total of eight different styles were used. Each participant was presented with 20 questions and asked to choose the result that best followed the style while", + "bbox": [ + 498, + 824, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4511", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b565d288cf8580c30a2dd2e2db458e9c19dce4818526603c989ad5ae44b0b181.jpg", + "image_caption": [ + "Figure 8. Visualization on ablation study." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 468, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "preserving the identity. As shown in Table 3, our method received higher scores compared to the baseline methods. Ours outperformed the text-based stylization methods by a large margin and scored higher than the method requiring correspondence specification.", + "bbox": [ + 75, + 231, + 468, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9c2b9c43ee7b2fe2e8da00510f4c4c4a3c372ff3d721bf51b441ba9582629742.jpg", + "table_caption": [ + "Table 3. Results from the user perceptual study given style exemplar and the deformation target. The percentage represents the selected frequency." + ], + "table_footnote": [], + "table_body": "
MethodUser Score
Ours60.65%
Deformation transfer38.17%
Text deformer0.44%
X-mesh0.74%
", + "bbox": [ + 166, + 367, + 375, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Applications", + "text_level": 1, + "bbox": [ + 76, + 436, + 207, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.1. Style Interpolation", + "text_level": 1, + "bbox": [ + 76, + 460, + 256, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LeGO enables linear style interpolation as illustrated in Figure 9, through weight blending. Diverse styles are seamlessly blended to create new stylized meshes by linearly interpolating weights from $D_{T1}$ and $D_{T2}$ using following equation:", + "bbox": [ + 75, + 484, + 468, + 559 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nW _ {\\text {n e w}} = \\alpha W _ {D _ {T 1}} + (1 - \\alpha) W _ {D _ {T 2}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 564, + 468, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, $W_{D_{T1}}$ and $W_{D_{T2}}$ denote the network weights for $D_{T1}$ and $D_{T2}$ , respectively, while $\\alpha$ represents the blending weight controlling the interpolation.", + "bbox": [ + 76, + 593, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7570eedb8d42d8e498594d4d2147b9431b72da8842d2276191bfbaf4f28cceb1.jpg", + "image_caption": [ + "Figure 9. Linear interpolation of Style A and Style B" + ], + "image_footnote": [], + "bbox": [ + 81, + 652, + 468, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Image-based 3D Stylized Avatar Generation", + "text_level": 1, + "bbox": [ + 76, + 794, + 449, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A stylized 3D face can be generated from a single portrait by first using methods that reconstruct 3D faces from 2D portraits. Among these methods, we employed MICA [49] to reconstruct a 3D face from an image. This reconstructed shape was then fed into LeGO to create stylized faces. The results are visualized in Figure 10.", + "bbox": [ + 75, + 818, + 468, + 907 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c72e2a57b95ad2397a6b5ca4b46376fe154b673f43951e9e564fb428f873ee51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 89, + 879, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3792c6cdf2a385cafff6dfafd3a563b0080dd89454b4e53acd5db4397d7e2a8.jpg", + "image_caption": [ + "Figure 10. Visualization of generating stylized 3D faces from 2D portraits [14]." + ], + "image_footnote": [], + "bbox": [ + 511, + 138, + 635, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7b2331c88f4bc2663c3387d46399f258445701aa4f19682747b966667d8f119.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 152, + 870, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitation and Conclusion", + "text_level": 1, + "bbox": [ + 498, + 353, + 746, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a novel approach for generating stylized 3D face meshes, considering three key elements. First, we proposed a surface deformation network that can generate a face in the desired topology using SIMS. Second, by the domain adaptation with hierarchical rendering, we achieved superior stylization capability. Lastly, Using 3DMM prior, we can generate a stylized face equipped with the animation capability.", + "bbox": [ + 496, + 378, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we proposed MAGE for practical usage, which can take diverse mesh topologies as input, and a novel style loss that adheres to the style effectively while preserving animation ability. Comprehensive experimental results demonstrate that our method is capable of generating a stylized mesh with consistent topology given deformation target meshes exhibiting significant topological variation.", + "bbox": [ + 496, + 500, + 890, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method shows promising results and enables significant advancements in practical avatar creation, it also has challenges to address. At inference, achieving a stylized output with the same topology as the input requires a two-stage process involving template replacement with the mean face mesh. Addressing these efficiency and practicality challenges is crucial for further enhancing our approach.", + "bbox": [ + 496, + 607, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: $90\\%$ ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: $10\\%$ ).", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4512", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. arXiv preprint arXiv:2301.02700, 2023. 2", + "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. 2", + "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2", + "[4] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3", + "[5] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 2, 5", + "[6] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3", + "[7] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 2, 3, 6", + "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2", + "[9] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3", + "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2", + "[11] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 3", + "[12] Wonjoon Jin, Nuri Ryu, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dr. 3d: Adapting 3d gans to artistic drawings. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3", + "[13] Yucheol Jung, Wonjong Jang, Soongjin Kim, Jiaolong Yang, Xin Tong, and Seungyong Lee. Deep deformable 3d caricatures with learned shape control. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 7", + "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 8", + "[15] Gwanghyun Kim and Se Young Chun. Datid-3d: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14203–14213, 2023. 2", + "[16] Seongtae Kim, Kyoungkook Kang, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dynagan: Dynamic few-shot adaptation of gans to multiple domains. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3", + "[17] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 5", + "[18] Shaoxu Li. Instruct-video2 avatar: Video-to- avatar generation with instructions. arXiv preprint arXiv:2306.02903, 2023.2", + "[19] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2", + "[20] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3", + "[21] Charles Loop. Smooth subdivision surfaces based on triangles. 1987. 5", + "[22] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 2, 6", + "[23] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3", + "[25] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3", + "[26] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2", + "[27] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019. 2" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4513", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3", + "[29] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 3", + "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2", + "[31] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2", + "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3", + "[34] Dafei Qin, Jun Saito, Noam Aigerman, Thibault Groueix, and Taku Komura. Neural face rigging for animating and retargeting facial meshes in the wild. arXiv preprint arXiv:2305.08296, 2023. 2, 3, 5", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[36] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European conference on computer vision (ECCV), pages 704-720, 2018. 5", + "[37] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 5", + "[38] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics (TOG), 41(3): 1-16, 2022. 3, 5", + "[39] Robert W Sumner and Jovan Popovic. Deformation transfer for triangle meshes. ACM Transactions on graphics (TOG), 23(3):399-405, 2004. 3, 6", + "[40] Minhyuk Sung, Zhenyu Jiang, Panos Achlioptas, Niloy J Mitra, and Leonidas J Guibas. Deformsyncnet: Deformation transfer via synchronized shape deformation spaces. arXiv preprint arXiv:2009.01456, 2020. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3", + "[42] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11204-11213, 2022. 3", + "[43] Xirui Yan, Zhenbo Yu, Bingbing Ni, and Hang Wang. Cross-species 3d face morphing via alignment-aware controller. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3018-3026, 2022. 2, 3", + "[44] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3", + "[45] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 3", + "[46] Chi Zhang, Yiwen Chen, Yijun Fu, Zhenglin Zhou, Gang Yu, Billzb Wang, Bin Fu, Tao Chen, Guosheng Lin, and Chunhua Shen. StyleAvatar3d: Leveraging image-text diffusion models for high-fidelity 3d avatar generation. arXiv preprint arXiv:2305.19012, 2023. 2", + "[47] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2", + "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2022. 2, 3, 4", + "[49] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 250-269. Springer, 2022. 8" + ], + "bbox": [ + 501, + 92, + 893, + 724 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4514", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_model.json b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c65f4a8fb898a734f09f4478f6d2007ac612ae0 --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_model.json @@ -0,0 +1,2288 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.204, + 0.856, + 0.242 + ], + "angle": 0, + "content": "Soyeon Yoon* Kwan Yun* Kwanggyoon Seo Sihun Cha Jung Eun Yoo Junyong Noh KAIST, Visual Media Lab" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.245, + 0.785, + 0.261 + ], + "angle": 0, + "content": "{thoyeony, yunandy, skg1023, chacorp, jey920, junyongnoh}@kaist.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.125, + 0.298, + 0.49, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.519, + 0.443, + 0.533 + ], + "angle": 0, + "content": "(a) Diverse deformation targets and output of LeGO" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.316, + 0.9, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.519, + 0.786, + 0.532 + ], + "angle": 0, + "content": "(b) Facial animation of stylized avatar" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.537, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1. (a) The proposed method demonstrates robustness to unseen face identities and topologies and effectively generates stylized output faces with desired topologies. (b) Our stylized avatars can be animated using 3DMM blend shapes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.577, + 0.314, + 0.593 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Recent advances in 3D face stylization have made significant strides in few to zero-shot settings. However, the degree of stylization achieved by existing methods is often not sufficient for practical applications because they are mostly based on statistical 3D Morphable Models (3DMM) with limited variations. To this end, we propose a method that can produce a highly stylized 3D face model with desired topology. Our methods train a surface deformation network with 3DMM and translate its domain to the target style using a paired exemplar. The network achieves stylization of the 3D face mesh by mimicking the style of the target using a differentiable renderer and directional CLIP losses. Additionally, during the inference process, we utilize a Mesh Agnostic Encoder (MAGE) that takes deformation target, a mesh of diverse topologies as input to the stylization process and encodes its shape into our latent space. The resulting stylized face model can be animated by commonly used 3DMM blend shapes. A set of quantitative and qualitative evaluations demonstrate that our method can produce" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.895, + 0.656 + ], + "angle": 0, + "content": "highly stylized face meshes according to a given style and output them in a desired topology. We also demonstrate example applications of our method including image-based stylized avatar generation, linear interpolation of geometric styles, and facial animation of stylized avatars." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.67, + 0.633, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.83 + ], + "angle": 0, + "content": "Crafting animatable stylized 3D avatars that encapsulate both personal identity and character style requires extensive efforts from skilled artists. When creating animated films, the artists design stylized 3D avatars whose facial appearance matches the theme of the entire content while putting careful effort into preserving the idiosyncrasy of the actors. Similarly, on social media, artists create numerous stylized presets so that the combinations of these presets can represent diverse identities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.832, + 0.892, + 0.879 + ], + "angle": 0, + "content": "To reduce the burden of manual crafting effort, generating stylized 3D faces has been a prominent area of research. Recent attempts include 3D-aware generative ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.771, + 0.901 + ], + "angle": 0, + "content": "*These authors contributed equally to this work" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4505" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "versarial networks (GANs) and denoising diffusion models (DMs) [1, 12, 15, 18, 46], generating a stylized texture for the 3D morphable model (3DMM) [2, 47], text-based geometry deformation [7, 22, 24-26], and surface deformation methods [13, 43]. These methods have successfully demonstrated the possibility of producing high-quality and diverse stylized 3D faces, although each method has distinct challenges." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.243 + ], + "angle": 0, + "content": "We identify three key elements for stylized avatar creation so that the results can be practically useful." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.243, + 0.468, + 0.272 + ], + "angle": 0, + "content": "1. Avatar creation in a desired topology that is compatible with conventional CG pipelines." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.273, + 0.429, + 0.287 + ], + "angle": 0, + "content": "2. Extending stylization capabilities beyond 3DMM." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.288, + 0.469, + 0.317 + ], + "angle": 0, + "content": "3. Generating stylized avatars that are animatable using blend shapes." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.243, + 0.469, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Where each component meets the standards and demands of the dynamic entertainment landscape 1) allowing creators to reuse existing animation rigs and texture maps across different models.; 2) achieving greater diversity and flexibility in expressing unique and non-conventional characters.; 3) enabling animators to achieve coherent and natural movements across various facial features and expressions. While recent methods have made meaningful strides in one or two aspects of these key elements, there has not been a method that satisfies all three elements, as summarized in Figure 2." + }, + { + "type": "image", + "bbox": [ + 0.059, + 0.483, + 0.468, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.636, + 0.47, + 0.721 + ], + "angle": 0, + "content": "Figure 2. Comparison of different stylized 3D face generation methods and their limitations in meeting key elements. 3D-aware methods cannot generate 3D face in desired topologies. 3DMM-based methods have a limited stylization capability. Text-based deformation models are not directly animatable. The proposed method meets the goal of all three components." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "To address this, we propose a novel method that can generate stylized 3D face meshes. This is achieved by translating the domain of a pre-trained surface deformation network based on one of the most widely used 3DMM model, FLAME [19], to a target style domain. We achieve this goal by first training the surface deformation network with the FLAME decoder to leverage its linear shape space combined with global expression space. During fine-tuning, we employ a directional CLIP-based domain adaptation method [16, 35, 48], widely used in 2D domain, to retain the face identity while reflecting the desired style." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.365 + ], + "angle": 0, + "content": "In addition, to seamlessly integrate this 2D-based training method into the stylized 3D face generation task, we propose a hierarchical rendering scheme that captures local and global facial features, ensuring effective training and identity preservation. In the inference stage, we introduce a Mesh Agnostic Encoder (MAGE) to enable mesh agnostic stylization for an input, which we call a deformation target that has various mesh topologies. MAGE is composed of pre-trained encoders from Neural Face Rigging (NFR) [34] and latent mapping networks, which establish correspondences between shapes by encoding mesh representations into a topology-invariant latent space. This enhances the versatility and applicability of our approach in the context of 3D face stylization. As shown in Figure 1, our method can generate stylized 3D face models with varied mesh topologies that are equipped with the animation capability of 3DMM while ensuring consistency across diverse deformation target mesh representations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.379, + 0.642, + 0.395 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.404, + 0.747, + 0.421 + ], + "angle": 0, + "content": "2.1. Stylized 3D face generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.701 + ], + "angle": 0, + "content": "With recent advances in the GANs and DMs that utilize neural fields for 3D-aware face generation [3, 10], the generation of faces in diverse styles has gained popularity [1, 12, 15, 18]. By leveraging 2D priors from generative models, which capture various patterns and variations observed in the extensive 2D training data, these methods can generate 3D faces with various styles. However, despite their success in producing consistent multi-view images through neural rendering, creating stylized faces in a desired topology is challenging, limiting their suitability for using existing graphics tools across different faces. On the other hand, 3DMM-based personalized 3D face generation methods with text-guidance [2, 47] have shown success in producing high-fidelity stylized textures for 3D face models. However, these methods confine the shape of the generated face models within the 3DMM shape space, constraining possibilities for geometric exaggerations or abstraction beyond the training data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.808, + 0.728 + ], + "angle": 0, + "content": "2.2. Learning-Based 3D Shape Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, learning implicit functions for 3D shapes has demonstrated remarkable performance in representing complex geometric structures [8, 23, 27, 30, 31]. In particular, DIF-Net [5] adopted MLPs for learning a standard signed distance function (SDF) and a volumetric deformation function, leading to comprehensive mapping between the produced SDFs. Most related to our work, DD3C [13] utilized template deformation for 3D caricature auto-decoder. They found that modeling each shape as a deformation of a fixed temaplate surface effective compared to absolute position. We advance a step further by training a surface deformation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4506" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.106 + ], + "angle": 0, + "content": "network on 3DMM and transfer its domain into stylization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.212 + ], + "angle": 0, + "content": "Another line of research aims to transfer shape deformation [39]. Recently, attempts have been made to transfer deformation utilizing examples [40, 43, 45] or text [7, 24, 25]. These methods can deform a mesh into a desired style and identity using a given example. However these deformation methods may not be as effective when it comes to coherent control of animation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.226, + 0.342, + 0.241 + ], + "angle": 0, + "content": "2.3. Few-Shot Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.25, + 0.468, + 0.461 + ], + "angle": 0, + "content": "Domain adaptation refers to the process of utilizing a neural network that has been initially trained on a large dataset from a source domain, and subsequently fine-tuning it on a smaller dataset from a target domain. Few-shot domain adaptation is widely researched in the 2D domain, especially with generative priors [4, 12, 20, 28, 29, 41, 42, 44]. Attempts have been made in one-shot domain adaptation to utilize the semantic capabilities of vision-language networks like CLIP [35]. These networks can provide direction for distinguishing between identity and style [6, 16, 48]. We adopt this two-way directional guidance from CLIP and combine it with a differentiable renderer to effectively stylize the face mesh while preserving the original identity using a paired exemplar." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.474, + 0.329, + 0.489 + ], + "angle": 0, + "content": "2.4. Mesh Agnostic Deformation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.468, + 0.649 + ], + "angle": 0, + "content": "Recently, mesh agnostic networks [9, 11, 32, 33, 38] demonstrated a great potential for learning 3D information such as dense correspondences between different 3D shape representations without relying on consistent topology or vertex ordering. These networks operate on meshes and encode shapes into a topology-invariant latent space. Specifically, NFR [34] proposed an autoencoder framework for facial expression retargeting across different mesh topologies. The framework utilizes separate expression and identity encoders, both functioning in a topology-agnostic manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.468, + 0.755 + ], + "angle": 0, + "content": "We integrate components of this approach into our method by mapping their embeddings into our latent space. This network translates the output of the pre-trained encoder from NFR to the latent space of our surface deformation network. This enables any geometric shape of the deformation target to be encoded and fed into our surface deformation network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.168, + 0.787 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our research focuses on training and fine-tuning a surface deformation network to generate stylized 3D faces with diverse shapes and expressions. We start with the source face deformation network \\( D_{S} \\) that deforms the template face to a face with different identities and expressions. Thereafter, we fine-tune it into a target style face deformation network \\( D_{T} \\) using a paired exemplar. This process is outlined below." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.136 + ], + "angle": 0, + "content": "1. We first train \\( D_{S} \\) using FLAME in a self-supervised manner, enabling the creation of versatile head meshes with different shapes and expressions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.137, + 0.891, + 0.182 + ], + "angle": 0, + "content": "2. For fine-tuning, we assemble a paired exemplar that consists of an identity exemplar mesh \\( M_S \\) and a style exemplar mesh \\( M_T \\), sharing identity." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.182, + 0.891, + 0.211 + ], + "angle": 0, + "content": "3. Both \\( M_S \\) and \\( M_T \\) jointly serve as example guidance to fine-tune \\( D_T \\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.212, + 0.891, + 0.256 + ], + "angle": 0, + "content": "4. At inference, using MAGE and \\( D_{T} \\), the deformation target of diverse topologies can be translated into a stylized face." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.891, + 0.29 + ], + "angle": 0, + "content": "A detailed description of these processes will be provided in the following subsections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.299, + 0.869, + 0.314 + ], + "angle": 0, + "content": "3.1. Deformation Network as Parametric Model" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.322, + 0.891, + 0.462 + ], + "angle": 0, + "content": "\\(D_{S}\\) is a surface deformation network that deforms a template face with the given latent vectors. We train the network to generate diverse human faces with FLAME, in a self-supervising manner. FLAME can manipulate both the global head shape and local expressions through its shape parameters \\(\\vec{\\beta} \\in \\mathbb{R}^{300}\\), and expression parameters \\(\\vec{\\psi} \\in \\mathbb{R}^{100}\\), which are the components of the FLAME parameters \\(\\Phi\\). This empowers \\(D_{S}\\) to generate diverse geometric face shapes and expressions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.463, + 0.891, + 0.555 + ], + "angle": 0, + "content": "To generate face meshes using these shape and expression components of \\(\\Phi\\) as input, we employ mapping networks, \\(\\mathcal{M}_{shape}\\) and \\(\\mathcal{M}_{exp}\\), which consist of MultiLayer Perceptrons (MLP). Each mapping network separately transforms \\(\\vec{\\beta}\\) and \\(\\vec{\\psi}\\) into latent vectors, \\(z_{s}\\) and \\(z_{e}\\), respectively." + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.556, + 0.891, + 0.577 + ], + "angle": 0, + "content": "\\[\nz _ {s} = \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}), \\quad z _ {e} = \\mathcal {M} _ {\\text {e x p}} (\\vec {\\psi}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.582, + 0.891, + 0.658 + ], + "angle": 0, + "content": "These latent vectors are then fed into \\( D_{S} \\), enabling the creation of diverse face meshes imbued with a wide range of expressions. The mapping networks \\( \\mathcal{M}_{shape} \\) and \\( \\mathcal{M}_{exp} \\) are trained jointly with \\( D_{S} \\) during the training process. The following losses are employed:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.668, + 0.891, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} (\\vec {\\beta}, \\vec {\\psi}) = \\left\\| \\operatorname {F L A M E} (\\vec {\\beta}, \\vec {\\psi}) - D _ {S} \\left(\\left[ \\begin{array}{l} \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}) \\\\ \\mathcal {M} _ {\\exp} (\\vec {\\psi}) \\end{array} \\right]\\right) \\right\\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.723, + 0.891, + 0.845 + ], + "angle": 0, + "content": "Surface-Intensive Mesh Sampling To enable \\(D_{S}\\) to represent various geometries, we introduce a surface-intensive mesh sampling (SIMS) strategy during training. SIMS is performed by randomly selecting points from the surfaces of the face mesh, \\(FLAME(\\vec{\\beta},\\vec{\\psi})\\), during training \\(D_{S}\\). Specifically, we sample approximately 4 times more points from the surface than the number of vertices of the face mesh. Experiment results are reported in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.855, + 0.585, + 0.868 + ], + "angle": 0, + "content": "3.2. LeGO" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "The fine-tuning process utilizes a paired exemplar, \\( M_S \\) and \\( M_T \\), to guide the adaptation of \\( D_T \\) for the generation of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4507" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.085, + 0.891, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.205, + 0.484, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.205, + 0.891, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.428, + 0.893, + 0.471 + ], + "angle": 0, + "content": "Figure 3. Overview of our method: The upper box illustrates the inference stage, where our method takes diverse deformation targets and generates stylized outputs. In the lower-left box, the training process of Mesh Agnostic Encoder (MAGE) is depicted. In the lower-right box, the fine-tuning process of \\( D_{T} \\) is illustrated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.607 + ], + "angle": 0, + "content": "stylized 3D faces. An identity exemplar mesh, \\( M_S \\), is generated from the FLAME decoder using \\( \\Phi_{ref} \\), and a style exemplar mesh, \\( M_T \\), is manually crafted from \\( M_S \\). During each iteration of fine-tuning, \\( \\Phi_{sample} \\) is sampled randomly to generate face mesh \\( M_S^{samp} \\) from \\( D_S \\) and corresponding stylized 3D face \\( M_T^{samp} \\) from \\( D_T \\). In the fine-tuning process, quantities of \\( M_S \\), \\( M_T \\), \\( D_S \\), and \\( D_T \\) are all singular, representing one pair of meshes and one pair of networks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.471, + 0.789 + ], + "angle": 0, + "content": "We introduce a hierarchical rendering scheme designed to preserve semantically important features, such as the shape and facial components, in the face mesh. With a differentiable renderer, this approach captures significant facial features from both local and global perspectives, enhancing stylization fidelity and identity preservation. To further enhance this process, we incorporate 2D-based losses, including CLIP reconstruction loss \\( L_{\\mathrm{CLIP}} \\), CLIP indomain loss \\( L_{in} \\), CLIP across-domain loss \\( L_{across} \\), and 3D-based losses such as vertex reconstruction loss \\( L_{vert} \\) and style loss \\( L_{style} \\). Details of each loss are elaborated in Section 3.3." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.23, + 0.816 + ], + "angle": 0, + "content": "3.3. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Vertex Reconstruction Loss The vertex reconstruction loss \\( L_{vert} \\) guides \\( D_T \\) in learning to deform a style exemplar mesh \\( M_T \\) from \\( [z_s^{ref}, z_e^{ref}] \\), which is mapped from \\( \\Phi_{ref} \\). The loss utilizes Mean Squared Error (MSE) to ensure that the vertices of the predicted mesh \\( M_T^* \\), generated by \\( D_T \\)," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.486, + 0.86, + 0.501 + ], + "angle": 0, + "content": "closely match \\(M_T\\). The loss can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.513, + 0.891, + 0.53 + ], + "angle": 0, + "content": "\\[\nL _ {v e r t} = \\left\\| M _ {T} - M _ {T} ^ {*} \\right\\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.651 + ], + "angle": 0, + "content": "CLIP Reconstruction Loss The CLIP reconstruction loss \\( L_{\\mathrm{CLIP}} \\) serves to maintain semantic consistency between the deformed mesh \\( M_T^* \\) and \\( M_T \\) in the CLIP space. This is crucial because even minor displacements in 3D space can result in surface irregularities or undesirable shading variations. The CLIP reconstruction loss can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.657, + 0.891, + 0.699 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {C L I P}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {*}\\right)\\right) \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.837 + ], + "angle": 0, + "content": "\\(R_{l,v}\\) represents the hierarchical rendering of a differentiable renderer from level \\(l\\) and view direction \\(v\\). Each view is anchored at a predefined position on a face mesh, including the front and both sides of the face. Each level corresponds to the distance from the face to the camera. At the highest level, where the face is captured in close-up, additional images are rendered from significant facial features such as the nose, eyes, and lips. \\(E_C\\) encodes rendered images into CLIP embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "CLIP Directional Loss Inspired from one-shot stylization methods in the 2D domain [48], we incorporate the CLIP in-domain loss \\( L_{in} \\) and the CLIP across-domain loss" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4508" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "\\(L_{across}\\). \\(L_{in}\\) ensures that the direction of two distinct faces in the source domain remains consistent in two distinct stylized faces of the target style domain. In contrast, \\(L_{across}\\) ensures that the direction of a face from the source domain and its corresponding stylized face in the target domain is preserved across different faces and their corresponding stylized faces. This can be shown in the lower right of Figure 3. Therefore our losses for CLIP in-domain and across-domain can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.239, + 0.47, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {i n}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\tag {5} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {s a m p}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) \\right\\| _ {2} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.303, + 0.47, + 0.35 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {a c r o s s}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\right. \\tag {6} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right)\\right) \\| _ {2} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.359, + 0.47, + 0.407 + ], + "angle": 0, + "content": "where \\( M_S^{samp} \\) and \\( M_T^{samp} \\) refer to \\( D_S[z_s^{samp};z_e^{samp}] \\) and \\( D_T[z_s^{samp};z_e^{samp}] \\), respectively, where \\( [z_s^{samp};z_e^{samp}] \\) is the latent vectors mapped from \\( \\Phi_{sample} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.473, + 0.651 + ], + "angle": 0, + "content": "Style Loss We introduce a novel style loss \\( L_{style} \\) to capture the style of \\( M_T \\) by utilizing surface normals. Style loss compares surface normal, which are strongly correlated with the semantic information of the mesh [5]. A straightforward approach would involve comparing \\( M_T \\) with \\( M_T^{samp} \\). However, this will constrain \\( M_T^{samp} \\) to have same expression as \\( M_T \\), which could degrade the animatability. To address this, we construct a pseudo pair for normal calculations. This is done by generating a face mesh through the concatenation of \\( z_e^{ref} \\) from \\( M_T \\) and \\( z_s^{samp} \\). This approach is utilized to compute the style loss, \\( L_{style} \\) without degrading the animatability while enhancing style adherence. In short, \\( L_{style} \\) ensures alignment of the surface normals from \\( D_T([z_s^{samp}; z_e^{ref}]) \\) with \\( M_T \\). The formulation of the style loss can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.661, + 0.47, + 0.7 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s t y l e}} = \\sum_ {f \\in S} \\left(1 - \\frac {n _ {f} \\cdot n _ {f} ^ {\\prime}}{\\left| n _ {f} \\right| \\left| n _ {f} ^ {\\prime} \\right|}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.706, + 0.468, + 0.737 + ], + "angle": 0, + "content": "where \\(n_f\\) refers to the surface normal from \\(M_T\\) while \\(n'_f\\) refers to the corresponding normal from \\(D_T([z_s^{samp}; z_e^{ref}])\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.737, + 0.453, + 0.751 + ], + "angle": 0, + "content": "The final objective function can be written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.762, + 0.47, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {t o t a l} = \\lambda_ {v e r t} L _ {v e r t} + \\lambda_ {\\mathrm {C L I P}} L _ {\\mathrm {C L I P}} + \\lambda_ {i n} L _ {i n} \\tag {8} \\\\ + \\lambda_ {a c r o s s} L _ {a c r o s s} + \\lambda_ {s t y l e} L _ {s t y l e} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.808, + 0.47, + 0.838 + ], + "angle": 0, + "content": "where \\(\\lambda_{vert}\\), \\(\\lambda_{\\mathrm{CLIP}}\\), \\(\\lambda_{in}\\), \\(\\lambda_{across}\\), and \\(\\lambda_{style}\\) are the weight for each loss term." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.298, + 0.864 + ], + "angle": 0, + "content": "3.4. Mesh Agnostic Encoder" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "NFR [34] employs DiffusionNet [38] for encoding and decoding face meshes, facilitating face retargeting across" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "different topologies. Given NFR's proficiency in extracting identity and expression details from faces with varying topologies, we extend its encoders to create a MAGE, as depicted in Figure 3. MAGE consists of ID2ID and exp2exp, both being MLPs that receive embeddings from each ID encoder and expression encoder pre-trained in NFR. Finally, given intermediate vectors from both ID2ID and exp2exp, the latent mapper outputs the latent vectors for \\( D_{S} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.303 + ], + "angle": 0, + "content": "MAGE is trained in a self-supervised manner by randomly sampling \\(\\beta\\) and \\(\\psi\\) and passing them through the mapping networks of \\(D_{S}\\) to obtain \\([z_{s};z_{e}]\\). Training involves comparing \\([\\hat{z}_s;\\hat{z}_e]\\), predicted from MAGE, with their corresponding ground truth latent vectors \\([z_s;z_e]\\). The objective function of the encoder uses a MSE loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.318, + 0.892, + 0.337 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {e n c}} = \\left\\| \\left[ z _ {s}; z _ {e} \\right] - \\left[ \\hat {z} _ {s}; \\hat {z} _ {e} \\right] \\right\\| _ {2} ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.389 + ], + "angle": 0, + "content": "where \\([ \\hat{z}_s; \\hat{z}_e ]\\) is MAGE \\([D_S([z_s;z_e])]\\). With this encoder, we can project a face mesh with diverse topologies into the latent space of \\( D_S \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.4, + 0.61, + 0.415 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.56 + ], + "angle": 0, + "content": "After training, \\( D_{T} \\) is capable of generating a stylized 3D face mesh with a desired topology from a deformation target mesh that has an arbitrary topology. This is accomplished by first projecting the deformation target into a latent vector, using MAGE. This latent vector is then fed as input to \\( D_{T} \\), allowing it to deform the template face of the desired topology into a stylized 3D face mesh. The resulting stylized face preserves both the identity of the deformation target and the desired style while providing animation control." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.575, + 0.634, + 0.591 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.71, + 0.611 + ], + "angle": 0, + "content": "4.1. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.892, + 0.692 + ], + "angle": 0, + "content": "We evaluate our method based on three key elements for stylized 3D head avatar creation: Avatar creation in a desired topology, stylization beyond 3DMM, and animation capability with blend shapes. Also, implementation details are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.916 + ], + "angle": 0, + "content": "Avatar in a Desired Topology To evaluate the capability of our method in producing a mesh with a desired topology from the deformation target of arbitrary topologies, we first apply a re-meshing technique that includes Loop subdivision [21] and mesh simplification [37] to meshes from FLAME. We also use original meshes from the CoMA [36] and ICT-FaceKit [17] directly as additional diverse meshes. To visually confirm that our method can align different mesh topologies with a desired template face, we enhance the resulting meshes with a checkerboard texture. The visual results, as shown in Figure 4, demonstrate that our method consistently produces stylized 3D faces with the desired topology, regardless of variations in the topology of the deformation target." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4509" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.468, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Figure 4. Demonstration of stylized 3D faces with desired topology, regardless of deformation target variations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.471, + 0.568 + ], + "angle": 0, + "content": "Stylization Capability We evaluate our method across various styles to assess its stylization capability. As shown in Figure 5, our method can generate a broad spectrum of styles encompassing human-like and non-human-like faces while preserving the original identities of the given face mesh. Our method also demonstrates the capability to achieve stylization across various geometries, including face masks and point clouds." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.584, + 0.468, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.847, + 0.47, + 0.904 + ], + "angle": 0, + "content": "Figure 5. Stylization results across diverse styles and identities. Our approach generates varied styles while preserving the deformation target identity and generalizing to diverse geometric representations like masks and point clouds." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.184 + ], + "angle": 0, + "content": "Animation Capability Using parameters from 3DMM, our method can animate facial expressions in the resulting stylized 3D faces. As illustrated in Figure 6, our method can generate 3D stylized faces with various expressions. This allows applications such as video-driven stylized talking heads, which are elaborated in supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.197, + 0.891, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Figure 6. Visualization of dynamic expressions in stylized 3D faces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.429, + 0.722, + 0.445 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.453, + 0.892, + 0.694 + ], + "angle": 0, + "content": "For a quantitative evaluation, we measured the average CLIP style preservation (CLIP-SP) and CLIP identity preservation (CLIP-IP) scores. These metrics reflect the trade-off between preserving style and identity, making their averages crucial for validating stylization. For both metrics, we calculated the cosine similarity of CLIP embeddings from rendered meshes. Specifically, for CLIP-SP, we compared the embeddings of the generated stylized face mesh and style exemplar mesh. For CLIP-IP, we compared the embeddings of the generated stylized face mesh and deformation target mesh. For dataset, we sampled 8 different face mesh from FLAME without expression and manually crafted stylized mesh corresponding to each identities. Also, for the quantitative evaluation, we randomly sampled another 10 different identities without expression to be used as deformation targets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Comparison with Baselines on Mesh Stylization We compared our stylized face generation results with those produced by baselines that can deform a mesh into different styles: Deformation Transfer [39], TextDeformer [7], and X-mesh [22]. In case of Deformation Transfer [39], we identified the correspondences between the identity exemplar mesh and the deformation target using 68 facial landmarks along with 9 additional points on the forehead in order to apply learned deformation. TextDeformer and X-mesh employ a text-guided deformation that operates on an input mesh. For both methods, the input face is deformed based on the same descriptive text to generate stylized outputs. All these baselines were evaluated on three different" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4510" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results are presented for comparison with baselines and ablations, featuring averages of CLIP-SP and CLIP-IP. The highest scores are denoted in bold, while the second highest scores are underlined." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.124, + 0.895, + 0.248 + ], + "angle": 0, + "content": "
Mesh TypeFLAMELoopSimplifiedOverall-Average
MethodsCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageAverage
Ours0.98670.93470.96070.98330.93500.95920.95590.89870.92730.9491
Ours w/o Lstyle0.98090.93740.95910.97680.93390.95530.95420.89870.92650.9470
Ours w direct Lstyle0.98650.93480.96070.98330.93560.95950.95500.89730.92620.9488
w/o hierarchical rendering0.98460.93490.95970.98240.93340.95790.95690.89850.92770.9484
Text deformer0.86410.83630.85020.86300.84250.85280.85340.81730.83530.8461
X-mesh0.86260.86020.86140.88170.89500.88840.86350.84480.85410.8680
Deformation transfer0.97680.93760.95720.95980.95760.95870.92540.92880.92710.9477
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.264, + 0.47, + 0.294 + ], + "angle": 0, + "content": "mesh topologies: FLAME and FLAME with Loop subdivision and mesh simplification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.43 + ], + "angle": 0, + "content": "We present qualitative comparison results on stylized 3D face generation in Figure 7. Deformation Transfer is capable of generating faces that follow the style; however, it exhibits artifacts, including anomalies and concavities on the surface due to the calculation of displacement and its direct transfer. In contrast, text-based methods fall short in following styles. Our method, on the other hand, can generate stylized 3D faces that adhere to both the input identity and style example without any artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.537 + ], + "angle": 0, + "content": "In quantitative evaluation, shown in Table 1, our method obtained the highest average scores computed from CLIP-SP and CLIP-IP. This clearly demonstrates that our approach achieves better performance in mesh stylization while preserving input identity compared to deformation baselines. More details and examples are provided in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.548, + 0.468, + 0.731 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.743, + 0.47, + 0.786 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison on stylization. Our method adheres to style and identity without artifacts unlike Deformation Transfer. Text-based approaches fail to match styles from text." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Surface-Intensive Mesh Sampling To test the proposed sampling method SIMS, we conducted a study on the reconstruction task comparing three different sampling variants for training \\(D_{S}\\). SIMS is surface-intensive sampling, using approximately four times more points sampled from the mesh surface compared to mesh vertices. Hybrid sampling, originally proposed in DD3C [13], sampled points both ver" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.262, + 0.892, + 0.276 + ], + "angle": 0, + "content": "Table 2. Reconstruction comparison with mesh sampling methods" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.283, + 0.875, + 0.35 + ], + "angle": 0, + "content": "
Reconstruction loss↓ (all in e-5)
Methodsoriginalsimplifiedloop-1loop-2average
SIMS(Ours)1.7901.6351.6121.5911.657
Hybrid2.3002.0192.1382.1182.144
Vertex1.9555.8663.2582.9593.509
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.891, + 0.404 + ], + "angle": 0, + "content": "tices and faces in similar ratio (\\(\\sim 1.1\\) times). Vertex-only used just the vertices of the mesh." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.406, + 0.892, + 0.543 + ], + "angle": 0, + "content": "For the experiment, we calculated the reconstruction loss on four different experiment settings: (1) original FLAME, (2) a simplified FLAME mesh with 1/4 the vertex count, (3) FLAME with one loop subdivision, and (4) FLAME with two loop subdivisions. The reconstruction loss was compared to the ground truth position using mean squared error. As shown in Table 2, SIMS achieved the lowest error by a large margin on all experiments. Hybrid sampling generally performed better than vertex-only sampling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.893, + 0.824 + ], + "angle": 0, + "content": "Ablation Study To perform an ablation study, we trained our model \\( D_{T} \\) using different settings: 1) \"Ours w/o \\( L_{\\text{style}} \\),\" where the style loss was removed; 2) \"Ours w direct \\( L_{\\text{style}} \\),\" where the style loss was directly compared between \\( D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}]) \\) and \\( M_{T} \\); 3) \"Ours w/o hierarchical rendering,\" where a single canonical view was used to calculate the \\( L_{\\text{in}} \\) and \\( L_{\\text{across}} \\). The results are displayed in Figure 8 and Table 1. The results produced by \"Ours\" and \"Ours w direct \\( L_{\\text{style}} \\)\" do not show any artifacts unlike the results produced by w/o \\( L_{\\text{style}} \\) and w/o hierarchical rendering, which exhibit surface artifacts. However, \"Ours w. direct \\( L_{\\text{style}} \\)\" forces \\( D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}]) \\) and \\( M_{T} \\), which have different expressions, to have the same normals; thus, it discards the expression. As a result, while \"Ours w. direct \\( L_{\\text{style}} \\)\" does stylize effectively, it cannot animate. Conversely, \"Ours\" not only stylizes well but is also animatable. Additional ablation study results regarding animation are presented in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.894, + 0.901 + ], + "angle": 0, + "content": "User Study We conducted a user study with 34 participants to evaluate different stylized face generation methods on human perception. A total of eight different styles were used. Each participant was presented with 20 questions and asked to choose the result that best followed the style while" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4511" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.09, + 0.47, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.197, + 0.396, + 0.211 + ], + "angle": 0, + "content": "Figure 8. Visualization on ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.307 + ], + "angle": 0, + "content": "preserving the identity. As shown in Table 3, our method received higher scores compared to the baseline methods. Ours outperformed the text-based stylization methods by a large margin and scored higher than the method requiring correspondence specification." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.321, + 0.469, + 0.363 + ], + "angle": 0, + "content": "Table 3. Results from the user perceptual study given style exemplar and the deformation target. The percentage represents the selected frequency." + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.368, + 0.377, + 0.426 + ], + "angle": 0, + "content": "
MethodUser Score
Ours60.65%
Deformation transfer38.17%
Text deformer0.44%
X-mesh0.74%
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.208, + 0.453 + ], + "angle": 0, + "content": "5. Applications" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.462, + 0.258, + 0.478 + ], + "angle": 0, + "content": "5.1. Style Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.469, + 0.56 + ], + "angle": 0, + "content": "LeGO enables linear style interpolation as illustrated in Figure 9, through weight blending. Diverse styles are seamlessly blended to create new stylized meshes by linearly interpolating weights from \\(D_{T1}\\) and \\(D_{T2}\\) using following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.565, + 0.469, + 0.582 + ], + "angle": 0, + "content": "\\[\nW _ {\\text {n e w}} = \\alpha W _ {D _ {T 1}} + (1 - \\alpha) W _ {D _ {T 2}} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.594, + 0.47, + 0.64 + ], + "angle": 0, + "content": "Here, \\( W_{D_{T1}} \\) and \\( W_{D_{T2}} \\) denote the network weights for \\( D_{T1} \\) and \\( D_{T2} \\), respectively, while \\( \\alpha \\) represents the blending weight controlling the interpolation." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.653, + 0.47, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.762, + 0.431, + 0.776 + ], + "angle": 0, + "content": "Figure 9. Linear interpolation of Style A and Style B" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.795, + 0.45, + 0.811 + ], + "angle": 0, + "content": "5.2. Image-based 3D Stylized Avatar Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.819, + 0.469, + 0.909 + ], + "angle": 0, + "content": "A stylized 3D face can be generated from a single portrait by first using methods that reconstruct 3D faces from 2D portraits. Among these methods, we employed MICA [49] to reconstruct a 3D face from an image. This reconstructed shape was then fed into LeGO to create stylized faces. The results are visualized in Figure 10." + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.09, + 0.88, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.14, + 0.636, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.153, + 0.871, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.306, + 0.892, + 0.334 + ], + "angle": 0, + "content": "Figure 10. Visualization of generating stylized 3D faces from 2D portraits [14]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.354, + 0.748, + 0.37 + ], + "angle": 0, + "content": "6. Limitation and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.892, + 0.501 + ], + "angle": 0, + "content": "We presented a novel approach for generating stylized 3D face meshes, considering three key elements. First, we proposed a surface deformation network that can generate a face in the desired topology using SIMS. Second, by the domain adaptation with hierarchical rendering, we achieved superior stylization capability. Lastly, Using 3DMM prior, we can generate a stylized face equipped with the animation capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.892, + 0.607 + ], + "angle": 0, + "content": "In addition, we proposed MAGE for practical usage, which can take diverse mesh topologies as input, and a novel style loss that adheres to the style effectively while preserving animation ability. Comprehensive experimental results demonstrate that our method is capable of generating a stylized mesh with consistent topology given deformation target meshes exhibiting significant topological variation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.713 + ], + "angle": 0, + "content": "While our method shows promising results and enables significant advancements in practical avatar creation, it also has challenges to address. At inference, achieving a stylized output with the same topology as the input requires a two-stage process involving template replacement with the mean face mesh. Addressing these efficiency and practicality challenges is crucial for further enhancing our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: \\(90\\%\\) ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: \\(10\\%\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4512" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. arXiv preprint arXiv:2301.02700, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[5] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.478 + ], + "angle": 0, + "content": "[6] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[7] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.578 + ], + "angle": 0, + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.579, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[9] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.47, + 0.703 + ], + "angle": 0, + "content": "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[11] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[12] Wonjoon Jin, Nuri Ryu, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dr. 3d: Adapting 3d gans to artistic drawings. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[13] Yucheol Jung, Wonjong Jang, Soongjin Kim, Jiaolong Yang, Xin Tong, and Seungyong Lee. Deep deformable 3d caricatures with learned shape control. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.191 + ], + "angle": 0, + "content": "[15] Gwanghyun Kim and Se Young Chun. Datid-3d: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14203–14213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[16] Seongtae Kim, Kyoungkook Kang, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dynagan: Dynamic few-shot adaptation of gans to multiple domains. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[17] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[18] Shaoxu Li. Instruct-video2 avatar: Video-to- avatar generation with instructions. arXiv preprint arXiv:2306.02903, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[19] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[20] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.491 + ], + "angle": 0, + "content": "[21] Charles Loop. Smooth subdivision surfaces based on triangles. 1987. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.492, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[22] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[23] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[25] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[26] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4513" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[29] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.47, + 0.36 + ], + "angle": 0, + "content": "[31] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.43 + ], + "angle": 0, + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[34] Dafei Qin, Jun Saito, Noam Aigerman, Thibault Groueix, and Taku Komura. Neural face rigging for animating and retargeting facial meshes in the wild. arXiv preprint arXiv:2305.08296, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[36] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European conference on computer vision (ECCV), pages 704-720, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[37] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[38] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics (TOG), 41(3): 1-16, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[39] Robert W Sumner and Jovan Popovic. Deformation transfer for triangle meshes. ACM Transactions on graphics (TOG), 23(3):399-405, 2004. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[40] Minhyuk Sung, Zhenyu Jiang, Panos Achlioptas, Niloy J Mitra, and Leonidas J Guibas. Deformsyncnet: Deformation transfer via synchronized shape deformation spaces. arXiv preprint arXiv:2009.01456, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[41] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.245 + ], + "angle": 0, + "content": "[42] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11204-11213, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[43] Xirui Yan, Zhenbo Yu, Bingbing Ni, and Hang Wang. Cross-species 3d face morphing via alignment-aware controller. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3018-3026, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.372 + ], + "angle": 0, + "content": "[44] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.892, + 0.443 + ], + "angle": 0, + "content": "[45] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.513 + ], + "angle": 0, + "content": "[46] Chi Zhang, Yiwen Chen, Yijun Fu, Zhenglin Zhou, Gang Yu, Billzb Wang, Bin Fu, Tao Chen, Guosheng Lin, and Chunhua Shen. StyleAvatar3d: Leveraging image-text diffusion models for high-fidelity 3d avatar generation. arXiv preprint arXiv:2305.19012, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[47] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.586, + 0.892, + 0.653 + ], + "angle": 0, + "content": "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.656, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[49] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 250-269. Springer, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4514" + } + ] +] \ No newline at end of file diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_origin.pdf b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ef7ab55d23c6f97bbfdb254e0253dae12cb3ccfa --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/6836350f-601e-42fa-b6e0-49050e114ece_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145476b548dd139ff57b3e60b196edf590929af399162155f5efe1c23fb605a7 +size 7323808 diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/full.md b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae493310bf3bfcf0dcfd9c80b2cbc7dd29b38ffb --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/full.md @@ -0,0 +1,330 @@ +# LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example + +Soyeon Yoon* Kwan Yun* Kwanggyoon Seo Sihun Cha Jung Eun Yoo Junyong Noh KAIST, Visual Media Lab + +{thoyeony, yunandy, skg1023, chacorp, jey920, junyongnoh}@kaist.ac.kr + +![](images/38c9f02eb51e5361cf0e14ca87fb34a950c377f46f76ae34436680573ddeed2b.jpg) +(a) Diverse deformation targets and output of LeGO + +![](images/e22b3963195b072c0d67965648471e991e66d2458316c9b8f154107a849ad188.jpg) +(b) Facial animation of stylized avatar +Figure 1. (a) The proposed method demonstrates robustness to unseen face identities and topologies and effectively generates stylized output faces with desired topologies. (b) Our stylized avatars can be animated using 3DMM blend shapes. + +# Abstract + +Recent advances in 3D face stylization have made significant strides in few to zero-shot settings. However, the degree of stylization achieved by existing methods is often not sufficient for practical applications because they are mostly based on statistical 3D Morphable Models (3DMM) with limited variations. To this end, we propose a method that can produce a highly stylized 3D face model with desired topology. Our methods train a surface deformation network with 3DMM and translate its domain to the target style using a paired exemplar. The network achieves stylization of the 3D face mesh by mimicking the style of the target using a differentiable renderer and directional CLIP losses. Additionally, during the inference process, we utilize a Mesh Agnostic Encoder (MAGE) that takes deformation target, a mesh of diverse topologies as input to the stylization process and encodes its shape into our latent space. The resulting stylized face model can be animated by commonly used 3DMM blend shapes. A set of quantitative and qualitative evaluations demonstrate that our method can produce + +highly stylized face meshes according to a given style and output them in a desired topology. We also demonstrate example applications of our method including image-based stylized avatar generation, linear interpolation of geometric styles, and facial animation of stylized avatars. + +# 1. Introduction + +Crafting animatable stylized 3D avatars that encapsulate both personal identity and character style requires extensive efforts from skilled artists. When creating animated films, the artists design stylized 3D avatars whose facial appearance matches the theme of the entire content while putting careful effort into preserving the idiosyncrasy of the actors. Similarly, on social media, artists create numerous stylized presets so that the combinations of these presets can represent diverse identities. + +To reduce the burden of manual crafting effort, generating stylized 3D faces has been a prominent area of research. Recent attempts include 3D-aware generative ad + +versarial networks (GANs) and denoising diffusion models (DMs) [1, 12, 15, 18, 46], generating a stylized texture for the 3D morphable model (3DMM) [2, 47], text-based geometry deformation [7, 22, 24-26], and surface deformation methods [13, 43]. These methods have successfully demonstrated the possibility of producing high-quality and diverse stylized 3D faces, although each method has distinct challenges. + +We identify three key elements for stylized avatar creation so that the results can be practically useful. + +1. Avatar creation in a desired topology that is compatible with conventional CG pipelines. +2. Extending stylization capabilities beyond 3DMM. +3. Generating stylized avatars that are animatable using blend shapes. + +Where each component meets the standards and demands of the dynamic entertainment landscape 1) allowing creators to reuse existing animation rigs and texture maps across different models.; 2) achieving greater diversity and flexibility in expressing unique and non-conventional characters.; 3) enabling animators to achieve coherent and natural movements across various facial features and expressions. While recent methods have made meaningful strides in one or two aspects of these key elements, there has not been a method that satisfies all three elements, as summarized in Figure 2. + +![](images/7b4ce791e7fd7bdb263df39ef1a0b9397188fbcf2ac5bc4e333a4b1a4226b2ad.jpg) +Figure 2. Comparison of different stylized 3D face generation methods and their limitations in meeting key elements. 3D-aware methods cannot generate 3D face in desired topologies. 3DMM-based methods have a limited stylization capability. Text-based deformation models are not directly animatable. The proposed method meets the goal of all three components. + +To address this, we propose a novel method that can generate stylized 3D face meshes. This is achieved by translating the domain of a pre-trained surface deformation network based on one of the most widely used 3DMM model, FLAME [19], to a target style domain. We achieve this goal by first training the surface deformation network with the FLAME decoder to leverage its linear shape space combined with global expression space. During fine-tuning, we employ a directional CLIP-based domain adaptation method [16, 35, 48], widely used in 2D domain, to retain the face identity while reflecting the desired style. + +In addition, to seamlessly integrate this 2D-based training method into the stylized 3D face generation task, we propose a hierarchical rendering scheme that captures local and global facial features, ensuring effective training and identity preservation. In the inference stage, we introduce a Mesh Agnostic Encoder (MAGE) to enable mesh agnostic stylization for an input, which we call a deformation target that has various mesh topologies. MAGE is composed of pre-trained encoders from Neural Face Rigging (NFR) [34] and latent mapping networks, which establish correspondences between shapes by encoding mesh representations into a topology-invariant latent space. This enhances the versatility and applicability of our approach in the context of 3D face stylization. As shown in Figure 1, our method can generate stylized 3D face models with varied mesh topologies that are equipped with the animation capability of 3DMM while ensuring consistency across diverse deformation target mesh representations. + +# 2. Related Work + +# 2.1. Stylized 3D face generation + +With recent advances in the GANs and DMs that utilize neural fields for 3D-aware face generation [3, 10], the generation of faces in diverse styles has gained popularity [1, 12, 15, 18]. By leveraging 2D priors from generative models, which capture various patterns and variations observed in the extensive 2D training data, these methods can generate 3D faces with various styles. However, despite their success in producing consistent multi-view images through neural rendering, creating stylized faces in a desired topology is challenging, limiting their suitability for using existing graphics tools across different faces. On the other hand, 3DMM-based personalized 3D face generation methods with text-guidance [2, 47] have shown success in producing high-fidelity stylized textures for 3D face models. However, these methods confine the shape of the generated face models within the 3DMM shape space, constraining possibilities for geometric exaggerations or abstraction beyond the training data. + +# 2.2. Learning-Based 3D Shape Network + +Recently, learning implicit functions for 3D shapes has demonstrated remarkable performance in representing complex geometric structures [8, 23, 27, 30, 31]. In particular, DIF-Net [5] adopted MLPs for learning a standard signed distance function (SDF) and a volumetric deformation function, leading to comprehensive mapping between the produced SDFs. Most related to our work, DD3C [13] utilized template deformation for 3D caricature auto-decoder. They found that modeling each shape as a deformation of a fixed temaplate surface effective compared to absolute position. We advance a step further by training a surface deformation + +network on 3DMM and transfer its domain into stylization. + +Another line of research aims to transfer shape deformation [39]. Recently, attempts have been made to transfer deformation utilizing examples [40, 43, 45] or text [7, 24, 25]. These methods can deform a mesh into a desired style and identity using a given example. However these deformation methods may not be as effective when it comes to coherent control of animation. + +# 2.3. Few-Shot Domain Adaptation + +Domain adaptation refers to the process of utilizing a neural network that has been initially trained on a large dataset from a source domain, and subsequently fine-tuning it on a smaller dataset from a target domain. Few-shot domain adaptation is widely researched in the 2D domain, especially with generative priors [4, 12, 20, 28, 29, 41, 42, 44]. Attempts have been made in one-shot domain adaptation to utilize the semantic capabilities of vision-language networks like CLIP [35]. These networks can provide direction for distinguishing between identity and style [6, 16, 48]. We adopt this two-way directional guidance from CLIP and combine it with a differentiable renderer to effectively stylize the face mesh while preserving the original identity using a paired exemplar. + +# 2.4. Mesh Agnostic Deformation + +Recently, mesh agnostic networks [9, 11, 32, 33, 38] demonstrated a great potential for learning 3D information such as dense correspondences between different 3D shape representations without relying on consistent topology or vertex ordering. These networks operate on meshes and encode shapes into a topology-invariant latent space. Specifically, NFR [34] proposed an autoencoder framework for facial expression retargeting across different mesh topologies. The framework utilizes separate expression and identity encoders, both functioning in a topology-agnostic manner. + +We integrate components of this approach into our method by mapping their embeddings into our latent space. This network translates the output of the pre-trained encoder from NFR to the latent space of our surface deformation network. This enables any geometric shape of the deformation target to be encoded and fed into our surface deformation network. + +# 3. Method + +Our research focuses on training and fine-tuning a surface deformation network to generate stylized 3D faces with diverse shapes and expressions. We start with the source face deformation network $D_{S}$ that deforms the template face to a face with different identities and expressions. Thereafter, we fine-tune it into a target style face deformation network $D_{T}$ using a paired exemplar. This process is outlined below. + +1. We first train $D_{S}$ using FLAME in a self-supervised manner, enabling the creation of versatile head meshes with different shapes and expressions. +2. For fine-tuning, we assemble a paired exemplar that consists of an identity exemplar mesh $M_S$ and a style exemplar mesh $M_T$ , sharing identity. +3. Both $M_S$ and $M_T$ jointly serve as example guidance to fine-tune $D_T$ . +4. At inference, using MAGE and $D_{T}$ , the deformation target of diverse topologies can be translated into a stylized face. + +A detailed description of these processes will be provided in the following subsections. + +# 3.1. Deformation Network as Parametric Model + +$D_{S}$ is a surface deformation network that deforms a template face with the given latent vectors. We train the network to generate diverse human faces with FLAME, in a self-supervising manner. FLAME can manipulate both the global head shape and local expressions through its shape parameters $\vec{\beta} \in \mathbb{R}^{300}$ , and expression parameters $\vec{\psi} \in \mathbb{R}^{100}$ , which are the components of the FLAME parameters $\Phi$ . This empowers $D_{S}$ to generate diverse geometric face shapes and expressions. + +To generate face meshes using these shape and expression components of $\Phi$ as input, we employ mapping networks, $\mathcal{M}_{shape}$ and $\mathcal{M}_{exp}$ , which consist of MultiLayer Perceptrons (MLP). Each mapping network separately transforms $\vec{\beta}$ and $\vec{\psi}$ into latent vectors, $z_{s}$ and $z_{e}$ , respectively. + +$$ +z _ {s} = \mathcal {M} _ {\text {s h a p e}} (\vec {\beta}), \quad z _ {e} = \mathcal {M} _ {\text {e x p}} (\vec {\psi}) \tag {1} +$$ + +These latent vectors are then fed into $D_{S}$ , enabling the creation of diverse face meshes imbued with a wide range of expressions. The mapping networks $\mathcal{M}_{shape}$ and $\mathcal{M}_{exp}$ are trained jointly with $D_{S}$ during the training process. The following losses are employed: + +$$ +\mathrm {L} (\vec {\beta}, \vec {\psi}) = \left\| \operatorname {F L A M E} (\vec {\beta}, \vec {\psi}) - D _ {S} \left(\left[ \begin{array}{l} \mathcal {M} _ {\text {s h a p e}} (\vec {\beta}) \\ \mathcal {M} _ {\exp} (\vec {\psi}) \end{array} \right]\right) \right\| _ {2} ^ {2} \tag {2} +$$ + +Surface-Intensive Mesh Sampling To enable $D_{S}$ to represent various geometries, we introduce a surface-intensive mesh sampling (SIMS) strategy during training. SIMS is performed by randomly selecting points from the surfaces of the face mesh, $FLAME(\vec{\beta},\vec{\psi})$ , during training $D_{S}$ . Specifically, we sample approximately 4 times more points from the surface than the number of vertices of the face mesh. Experiment results are reported in Sec. 4.2. + +# 3.2. LeGO + +The fine-tuning process utilizes a paired exemplar, $M_S$ and $M_T$ , to guide the adaptation of $D_T$ for the generation of + +![](images/c1ce92d59794f03772abbfd385e751411f605deb0c55f32b5e6912dd137cd0b2.jpg) + +![](images/c9823072a971478dd9d876b739437c5e189760037433427a52ae61e8cd35bfe3.jpg) +Figure 3. Overview of our method: The upper box illustrates the inference stage, where our method takes diverse deformation targets and generates stylized outputs. In the lower-left box, the training process of Mesh Agnostic Encoder (MAGE) is depicted. In the lower-right box, the fine-tuning process of $D_{T}$ is illustrated. + +![](images/aecb9ef82435c93ea841f65e8d8dc807b605c1e57accf11bd2365acd633e6765.jpg) + +stylized 3D faces. An identity exemplar mesh, $M_S$ , is generated from the FLAME decoder using $\Phi_{ref}$ , and a style exemplar mesh, $M_T$ , is manually crafted from $M_S$ . During each iteration of fine-tuning, $\Phi_{sample}$ is sampled randomly to generate face mesh $M_S^{samp}$ from $D_S$ and corresponding stylized 3D face $M_T^{samp}$ from $D_T$ . In the fine-tuning process, quantities of $M_S$ , $M_T$ , $D_S$ , and $D_T$ are all singular, representing one pair of meshes and one pair of networks. + +We introduce a hierarchical rendering scheme designed to preserve semantically important features, such as the shape and facial components, in the face mesh. With a differentiable renderer, this approach captures significant facial features from both local and global perspectives, enhancing stylization fidelity and identity preservation. To further enhance this process, we incorporate 2D-based losses, including CLIP reconstruction loss $L_{\mathrm{CLIP}}$ , CLIP indomain loss $L_{in}$ , CLIP across-domain loss $L_{across}$ , and 3D-based losses such as vertex reconstruction loss $L_{vert}$ and style loss $L_{style}$ . Details of each loss are elaborated in Section 3.3. + +# 3.3. Loss Functions + +Vertex Reconstruction Loss The vertex reconstruction loss $L_{vert}$ guides $D_T$ in learning to deform a style exemplar mesh $M_T$ from $[z_s^{ref}, z_e^{ref}]$ , which is mapped from $\Phi_{ref}$ . The loss utilizes Mean Squared Error (MSE) to ensure that the vertices of the predicted mesh $M_T^*$ , generated by $D_T$ , + +closely match $M_T$ . The loss can be written as follows: + +$$ +L _ {v e r t} = \left\| M _ {T} - M _ {T} ^ {*} \right\| _ {2} ^ {2} \tag {3} +$$ + +CLIP Reconstruction Loss The CLIP reconstruction loss $L_{\mathrm{CLIP}}$ serves to maintain semantic consistency between the deformed mesh $M_T^*$ and $M_T$ in the CLIP space. This is crucial because even minor displacements in 3D space can result in surface irregularities or undesirable shading variations. The CLIP reconstruction loss can be written as follows: + +$$ +L _ {\text {C L I P}} = \sum_ {l \in L} \sum_ {v \in V _ {l}} \| E _ {C} \left(R _ {l, v} \left(M _ {T}\right)\right) - E _ {C} \left(R _ {l, v} \left(M _ {T} ^ {*}\right)\right) \| _ {2} ^ {2} \tag {4} +$$ + +$R_{l,v}$ represents the hierarchical rendering of a differentiable renderer from level $l$ and view direction $v$ . Each view is anchored at a predefined position on a face mesh, including the front and both sides of the face. Each level corresponds to the distance from the face to the camera. At the highest level, where the face is captured in close-up, additional images are rendered from significant facial features such as the nose, eyes, and lips. $E_C$ encodes rendered images into CLIP embeddings. + +CLIP Directional Loss Inspired from one-shot stylization methods in the 2D domain [48], we incorporate the CLIP in-domain loss $L_{in}$ and the CLIP across-domain loss + +$L_{across}$ . $L_{in}$ ensures that the direction of two distinct faces in the source domain remains consistent in two distinct stylized faces of the target style domain. In contrast, $L_{across}$ ensures that the direction of a face from the source domain and its corresponding stylized face in the target domain is preserved across different faces and their corresponding stylized faces. This can be shown in the lower right of Figure 3. Therefore our losses for CLIP in-domain and across-domain can be written as follows: + +$$ +\begin{array}{l} L _ {\text {i n}} = \sum_ {l \in L} \sum_ {v \in V _ {l}} \| \left(E _ {C} \left(R _ {l, v} \left(M _ {S} ^ {\text {s a m p}}\right)\right) - E _ {C} \left(R _ {l, v} \left(M _ {S}\right)\right)\right) \tag {5} \\ - \left(E _ {C} \left(R _ {l, v} \left(M _ {T} ^ {s a m p}\right)\right) - E _ {C} \left(R _ {l, v} \left(M _ {T}\right)\right) \right\| _ {2} ^ {2} \\ \end{array} +$$ + +$$ +\begin{array}{l} L _ {\text {a c r o s s}} = \sum_ {l \in L} \sum_ {v \in V _ {l}} \| \left(E _ {C} \left(R _ {l, v} \left(M _ {T}\right)\right) - \left(E _ {C} \left(R _ {l, v} \left(M _ {S}\right)\right)\right) \right. \tag {6} \\ - \left(E _ {C} \left(R _ {l, v} \left(M _ {T} ^ {\text {s a m p}}\right)\right) - E _ {C} \left(R _ {l, v} \left(M _ {S} ^ {\text {s a m p}}\right)\right)\right) \| _ {2} ^ {2} \\ \end{array} +$$ + +where $M_S^{samp}$ and $M_T^{samp}$ refer to $D_S[z_s^{samp};z_e^{samp}]$ and $D_T[z_s^{samp};z_e^{samp}]$ , respectively, where $[z_s^{samp};z_e^{samp}]$ is the latent vectors mapped from $\Phi_{sample}$ . + +Style Loss We introduce a novel style loss $L_{style}$ to capture the style of $M_T$ by utilizing surface normals. Style loss compares surface normal, which are strongly correlated with the semantic information of the mesh [5]. A straightforward approach would involve comparing $M_T$ with $M_T^{samp}$ . However, this will constrain $M_T^{samp}$ to have same expression as $M_T$ , which could degrade the animatability. To address this, we construct a pseudo pair for normal calculations. This is done by generating a face mesh through the concatenation of $z_e^{ref}$ from $M_T$ and $z_s^{samp}$ . This approach is utilized to compute the style loss, $L_{style}$ without degrading the animatability while enhancing style adherence. In short, $L_{style}$ ensures alignment of the surface normals from $D_T([z_s^{samp}; z_e^{ref}])$ with $M_T$ . The formulation of the style loss can be written as follows: + +$$ +L _ {\text {s t y l e}} = \sum_ {f \in S} \left(1 - \frac {n _ {f} \cdot n _ {f} ^ {\prime}}{\left| n _ {f} \right| \left| n _ {f} ^ {\prime} \right|}\right) \tag {7} +$$ + +where $n_f$ refers to the surface normal from $M_T$ while $n'_f$ refers to the corresponding normal from $D_T([z_s^{samp}; z_e^{ref}])$ . + +The final objective function can be written as follows: + +$$ +\begin{array}{l} L _ {t o t a l} = \lambda_ {v e r t} L _ {v e r t} + \lambda_ {\mathrm {C L I P}} L _ {\mathrm {C L I P}} + \lambda_ {i n} L _ {i n} \tag {8} \\ + \lambda_ {a c r o s s} L _ {a c r o s s} + \lambda_ {s t y l e} L _ {s t y l e} \\ \end{array} +$$ + +where $\lambda_{vert}$ , $\lambda_{\mathrm{CLIP}}$ , $\lambda_{in}$ , $\lambda_{across}$ , and $\lambda_{style}$ are the weight for each loss term. + +# 3.4. Mesh Agnostic Encoder + +NFR [34] employs DiffusionNet [38] for encoding and decoding face meshes, facilitating face retargeting across + +different topologies. Given NFR's proficiency in extracting identity and expression details from faces with varying topologies, we extend its encoders to create a MAGE, as depicted in Figure 3. MAGE consists of ID2ID and exp2exp, both being MLPs that receive embeddings from each ID encoder and expression encoder pre-trained in NFR. Finally, given intermediate vectors from both ID2ID and exp2exp, the latent mapper outputs the latent vectors for $D_{S}$ . + +MAGE is trained in a self-supervised manner by randomly sampling $\beta$ and $\psi$ and passing them through the mapping networks of $D_{S}$ to obtain $[z_{s};z_{e}]$ . Training involves comparing $[\hat{z}_s;\hat{z}_e]$ , predicted from MAGE, with their corresponding ground truth latent vectors $[z_s;z_e]$ . The objective function of the encoder uses a MSE loss as follows: + +$$ +L _ {\mathrm {e n c}} = \left\| \left[ z _ {s}; z _ {e} \right] - \left[ \hat {z} _ {s}; \hat {z} _ {e} \right] \right\| _ {2} ^ {2} \tag {9} +$$ + +where $[ \hat{z}_s; \hat{z}_e ]$ is MAGE $[D_S([z_s;z_e])]$ . With this encoder, we can project a face mesh with diverse topologies into the latent space of $D_S$ . + +# 3.5. Inference + +After training, $D_{T}$ is capable of generating a stylized 3D face mesh with a desired topology from a deformation target mesh that has an arbitrary topology. This is accomplished by first projecting the deformation target into a latent vector, using MAGE. This latent vector is then fed as input to $D_{T}$ , allowing it to deform the template face of the desired topology into a stylized 3D face mesh. The resulting stylized face preserves both the identity of the deformation target and the desired style while providing animation control. + +# 4. Experiments + +# 4.1. Qualitative Evaluation + +We evaluate our method based on three key elements for stylized 3D head avatar creation: Avatar creation in a desired topology, stylization beyond 3DMM, and animation capability with blend shapes. Also, implementation details are provided in supplementary material. + +Avatar in a Desired Topology To evaluate the capability of our method in producing a mesh with a desired topology from the deformation target of arbitrary topologies, we first apply a re-meshing technique that includes Loop subdivision [21] and mesh simplification [37] to meshes from FLAME. We also use original meshes from the CoMA [36] and ICT-FaceKit [17] directly as additional diverse meshes. To visually confirm that our method can align different mesh topologies with a desired template face, we enhance the resulting meshes with a checkerboard texture. The visual results, as shown in Figure 4, demonstrate that our method consistently produces stylized 3D faces with the desired topology, regardless of variations in the topology of the deformation target. + +![](images/c1b383397d4ace2329250b51d7f5ed7551c817793476eef4fffb332c323b633f.jpg) +Figure 4. Demonstration of stylized 3D faces with desired topology, regardless of deformation target variations. + +Stylization Capability We evaluate our method across various styles to assess its stylization capability. As shown in Figure 5, our method can generate a broad spectrum of styles encompassing human-like and non-human-like faces while preserving the original identities of the given face mesh. Our method also demonstrates the capability to achieve stylization across various geometries, including face masks and point clouds. + +![](images/b8b9308a7bd4554173aa44cc65e74da2372f3ffc683567f71653c0c3528b4e97.jpg) +Figure 5. Stylization results across diverse styles and identities. Our approach generates varied styles while preserving the deformation target identity and generalizing to diverse geometric representations like masks and point clouds. + +Animation Capability Using parameters from 3DMM, our method can animate facial expressions in the resulting stylized 3D faces. As illustrated in Figure 6, our method can generate 3D stylized faces with various expressions. This allows applications such as video-driven stylized talking heads, which are elaborated in supplementary material. + +![](images/3693ca614ca6b7633101a8f131125b1f858b3d45cfac1212c7561c87f0b5ea4b.jpg) +Figure 6. Visualization of dynamic expressions in stylized 3D faces. + +# 4.2. Quantitative Evaluation + +For a quantitative evaluation, we measured the average CLIP style preservation (CLIP-SP) and CLIP identity preservation (CLIP-IP) scores. These metrics reflect the trade-off between preserving style and identity, making their averages crucial for validating stylization. For both metrics, we calculated the cosine similarity of CLIP embeddings from rendered meshes. Specifically, for CLIP-SP, we compared the embeddings of the generated stylized face mesh and style exemplar mesh. For CLIP-IP, we compared the embeddings of the generated stylized face mesh and deformation target mesh. For dataset, we sampled 8 different face mesh from FLAME without expression and manually crafted stylized mesh corresponding to each identities. Also, for the quantitative evaluation, we randomly sampled another 10 different identities without expression to be used as deformation targets. + +Comparison with Baselines on Mesh Stylization We compared our stylized face generation results with those produced by baselines that can deform a mesh into different styles: Deformation Transfer [39], TextDeformer [7], and X-mesh [22]. In case of Deformation Transfer [39], we identified the correspondences between the identity exemplar mesh and the deformation target using 68 facial landmarks along with 9 additional points on the forehead in order to apply learned deformation. TextDeformer and X-mesh employ a text-guided deformation that operates on an input mesh. For both methods, the input face is deformed based on the same descriptive text to generate stylized outputs. All these baselines were evaluated on three different + +Table 1. Quantitative results are presented for comparison with baselines and ablations, featuring averages of CLIP-SP and CLIP-IP. The highest scores are denoted in bold, while the second highest scores are underlined. + +
Mesh TypeFLAMELoopSimplifiedOverall-Average
MethodsCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageAverage
Ours0.98670.93470.96070.98330.93500.95920.95590.89870.92730.9491
Ours w/o Lstyle0.98090.93740.95910.97680.93390.95530.95420.89870.92650.9470
Ours w direct Lstyle0.98650.93480.96070.98330.93560.95950.95500.89730.92620.9488
w/o hierarchical rendering0.98460.93490.95970.98240.93340.95790.95690.89850.92770.9484
Text deformer0.86410.83630.85020.86300.84250.85280.85340.81730.83530.8461
X-mesh0.86260.86020.86140.88170.89500.88840.86350.84480.85410.8680
Deformation transfer0.97680.93760.95720.95980.95760.95870.92540.92880.92710.9477
+ +mesh topologies: FLAME and FLAME with Loop subdivision and mesh simplification. + +We present qualitative comparison results on stylized 3D face generation in Figure 7. Deformation Transfer is capable of generating faces that follow the style; however, it exhibits artifacts, including anomalies and concavities on the surface due to the calculation of displacement and its direct transfer. In contrast, text-based methods fall short in following styles. Our method, on the other hand, can generate stylized 3D faces that adhere to both the input identity and style example without any artifacts. + +In quantitative evaluation, shown in Table 1, our method obtained the highest average scores computed from CLIP-SP and CLIP-IP. This clearly demonstrates that our approach achieves better performance in mesh stylization while preserving input identity compared to deformation baselines. More details and examples are provided in the supplementary material. + +![](images/49ae9faeb5d6783b98f9bcdac24cfd62f117a32575ccc4a4e3b01af0baa00a24.jpg) +Figure 7. Qualitative comparison on stylization. Our method adheres to style and identity without artifacts unlike Deformation Transfer. Text-based approaches fail to match styles from text. + +Surface-Intensive Mesh Sampling To test the proposed sampling method SIMS, we conducted a study on the reconstruction task comparing three different sampling variants for training $D_{S}$ . SIMS is surface-intensive sampling, using approximately four times more points sampled from the mesh surface compared to mesh vertices. Hybrid sampling, originally proposed in DD3C [13], sampled points both ver + +Table 2. Reconstruction comparison with mesh sampling methods + +
Reconstruction loss↓ (all in e-5)
Methodsoriginalsimplifiedloop-1loop-2average
SIMS(Ours)1.7901.6351.6121.5911.657
Hybrid2.3002.0192.1382.1182.144
Vertex1.9555.8663.2582.9593.509
+ +tices and faces in similar ratio ( $\sim 1.1$ times). Vertex-only used just the vertices of the mesh. + +For the experiment, we calculated the reconstruction loss on four different experiment settings: (1) original FLAME, (2) a simplified FLAME mesh with 1/4 the vertex count, (3) FLAME with one loop subdivision, and (4) FLAME with two loop subdivisions. The reconstruction loss was compared to the ground truth position using mean squared error. As shown in Table 2, SIMS achieved the lowest error by a large margin on all experiments. Hybrid sampling generally performed better than vertex-only sampling. + +Ablation Study To perform an ablation study, we trained our model $D_{T}$ using different settings: 1) "Ours w/o $L_{\text{style}}$ ," where the style loss was removed; 2) "Ours w direct $L_{\text{style}}$ ," where the style loss was directly compared between $D_{T}([z_{s}^{\text{samp}}; z_{e}^{\text{samp}}])$ and $M_{T}$ ; 3) "Ours w/o hierarchical rendering," where a single canonical view was used to calculate the $L_{\text{in}}$ and $L_{\text{across}}$ . The results are displayed in Figure 8 and Table 1. The results produced by "Ours" and "Ours w direct $L_{\text{style}}$ " do not show any artifacts unlike the results produced by w/o $L_{\text{style}}$ and w/o hierarchical rendering, which exhibit surface artifacts. However, "Ours w. direct $L_{\text{style}}$ " forces $D_{T}([z_{s}^{\text{samp}}; z_{e}^{\text{samp}}])$ and $M_{T}$ , which have different expressions, to have the same normals; thus, it discards the expression. As a result, while "Ours w. direct $L_{\text{style}}$ " does stylize effectively, it cannot animate. Conversely, "Ours" not only stylizes well but is also animatable. Additional ablation study results regarding animation are presented in the supplementary material. + +User Study We conducted a user study with 34 participants to evaluate different stylized face generation methods on human perception. A total of eight different styles were used. Each participant was presented with 20 questions and asked to choose the result that best followed the style while + +![](images/b565d288cf8580c30a2dd2e2db458e9c19dce4818526603c989ad5ae44b0b181.jpg) +Figure 8. Visualization on ablation study. + +preserving the identity. As shown in Table 3, our method received higher scores compared to the baseline methods. Ours outperformed the text-based stylization methods by a large margin and scored higher than the method requiring correspondence specification. + +Table 3. Results from the user perceptual study given style exemplar and the deformation target. The percentage represents the selected frequency. + +
MethodUser Score
Ours60.65%
Deformation transfer38.17%
Text deformer0.44%
X-mesh0.74%
+ +# 5. Applications + +# 5.1. Style Interpolation + +LeGO enables linear style interpolation as illustrated in Figure 9, through weight blending. Diverse styles are seamlessly blended to create new stylized meshes by linearly interpolating weights from $D_{T1}$ and $D_{T2}$ using following equation: + +$$ +W _ {\text {n e w}} = \alpha W _ {D _ {T 1}} + (1 - \alpha) W _ {D _ {T 2}} \tag {10} +$$ + +Here, $W_{D_{T1}}$ and $W_{D_{T2}}$ denote the network weights for $D_{T1}$ and $D_{T2}$ , respectively, while $\alpha$ represents the blending weight controlling the interpolation. + +![](images/7570eedb8d42d8e498594d4d2147b9431b72da8842d2276191bfbaf4f28cceb1.jpg) +Figure 9. Linear interpolation of Style A and Style B + +# 5.2. Image-based 3D Stylized Avatar Generation + +A stylized 3D face can be generated from a single portrait by first using methods that reconstruct 3D faces from 2D portraits. Among these methods, we employed MICA [49] to reconstruct a 3D face from an image. This reconstructed shape was then fed into LeGO to create stylized faces. The results are visualized in Figure 10. + +![](images/c72e2a57b95ad2397a6b5ca4b46376fe154b673f43951e9e564fb428f873ee51.jpg) + +![](images/f3792c6cdf2a385cafff6dfafd3a563b0080dd89454b4e53acd5db4397d7e2a8.jpg) +Figure 10. Visualization of generating stylized 3D faces from 2D portraits [14]. + +![](images/b7b2331c88f4bc2663c3387d46399f258445701aa4f19682747b966667d8f119.jpg) + +# 6. Limitation and Conclusion + +We presented a novel approach for generating stylized 3D face meshes, considering three key elements. First, we proposed a surface deformation network that can generate a face in the desired topology using SIMS. Second, by the domain adaptation with hierarchical rendering, we achieved superior stylization capability. Lastly, Using 3DMM prior, we can generate a stylized face equipped with the animation capability. + +In addition, we proposed MAGE for practical usage, which can take diverse mesh topologies as input, and a novel style loss that adheres to the style effectively while preserving animation ability. Comprehensive experimental results demonstrate that our method is capable of generating a stylized mesh with consistent topology given deformation target meshes exhibiting significant topological variation. + +While our method shows promising results and enables significant advancements in practical avatar creation, it also has challenges to address. At inference, achieving a stylized output with the same topology as the input requires a two-stage process involving template replacement with the mean face mesh. Addressing these efficiency and practicality challenges is crucial for further enhancing our approach. + +Acknowledgement This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: $90\%$ ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: $10\%$ ). + +# References + +[1] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. arXiv preprint arXiv:2301.02700, 2023. 2 +[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. 2 +[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2 +[4] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3 +[5] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 2, 5 +[6] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3 +[7] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 2, 3, 6 +[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2 +[9] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3 +[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2 +[11] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 3 +[12] Wonjoon Jin, Nuri Ryu, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dr. 3d: Adapting 3d gans to artistic drawings. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3 +[13] Yucheol Jung, Wonjong Jang, Soongjin Kim, Jiaolong Yang, Xin Tong, and Seungyong Lee. Deep deformable 3d caricatures with learned shape control. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 7 +[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. + +In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 8 +[15] Gwanghyun Kim and Se Young Chun. Datid-3d: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14203–14213, 2023. 2 +[16] Seongtae Kim, Kyoungkook Kang, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dynagan: Dynamic few-shot adaptation of gans to multiple domains. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3 +[17] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 5 +[18] Shaoxu Li. Instruct-video2 avatar: Video-to- avatar generation with instructions. arXiv preprint arXiv:2306.02903, 2023.2 +[19] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2 +[20] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3 +[21] Charles Loop. Smooth subdivision surfaces based on triangles. 1987. 5 +[22] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 2, 6 +[23] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3 +[25] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3 +[26] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2 +[27] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019. 2 + +[28] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3 +[29] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 3 +[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2 +[31] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2 +[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3 +[34] Dafei Qin, Jun Saito, Noam Aigerman, Thibault Groueix, and Taku Komura. Neural face rigging for animating and retargeting facial meshes in the wild. arXiv preprint arXiv:2305.08296, 2023. 2, 3, 5 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 +[36] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European conference on computer vision (ECCV), pages 704-720, 2018. 5 +[37] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 5 +[38] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics (TOG), 41(3): 1-16, 2022. 3, 5 +[39] Robert W Sumner and Jovan Popovic. Deformation transfer for triangle meshes. ACM Transactions on graphics (TOG), 23(3):399-405, 2004. 3, 6 +[40] Minhyuk Sung, Zhenyu Jiang, Panos Achlioptas, Niloy J Mitra, and Leonidas J Guibas. Deformsyncnet: Deformation transfer via synchronized shape deformation spaces. arXiv preprint arXiv:2009.01456, 2020. 3 + +[41] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3 +[42] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11204-11213, 2022. 3 +[43] Xirui Yan, Zhenbo Yu, Bingbing Ni, and Hang Wang. Cross-species 3d face morphing via alignment-aware controller. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3018-3026, 2022. 2, 3 +[44] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3 +[45] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 3 +[46] Chi Zhang, Yiwen Chen, Yijun Fu, Zhenglin Zhou, Gang Yu, Billzb Wang, Bin Fu, Tao Chen, Guosheng Lin, and Chunhua Shen. StyleAvatar3d: Leveraging image-text diffusion models for high-fidelity 3d avatar generation. arXiv preprint arXiv:2305.19012, 2023. 2 +[47] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2 +[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2022. 2, 3, 4 +[49] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 250-269. Springer, 2022. 8 \ No newline at end of file diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/images.zip b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9d4daaed056ca48984aa05603967954a1a91d599 --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:722f17cf923f8a738e2d274276c824e86b6bef721cd8bfdd5b44f698c67707d6 +size 631767 diff --git a/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/layout.json b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fa44a46ea6fcf772db54c498f39989b8c95a3455 --- /dev/null +++ b/2024/LeGO_ Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example/layout.json @@ -0,0 +1,9659 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 102, + 547, + 140 + ], + "type": "text", + "content": "LeGO: Leveraging a Surface Deformation Network for Animatable Stylized Face Generation with One Example" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 161, + 523, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 161, + 523, + 191 + ], + "spans": [ + { + "bbox": [ + 66, + 161, + 523, + 191 + ], + "type": "text", + "content": "Soyeon Yoon* Kwan Yun* Kwanggyoon Seo Sihun Cha Jung Eun Yoo Junyong Noh KAIST, Visual Media Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 194, + 480, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 194, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 107, + 194, + 480, + 206 + ], + "type": "text", + "content": "{thoyeony, yunandy, skg1023, chacorp, jey920, junyongnoh}@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 76, + 236, + 299, + 406 + ], + "blocks": [ + { + "bbox": [ + 76, + 236, + 299, + 406 + ], + "lines": [ + { + "bbox": [ + 76, + 236, + 299, + 406 + ], + "spans": [ + { + "bbox": [ + 76, + 236, + 299, + 406 + ], + "type": "image", + "image_path": "38c9f02eb51e5361cf0e14ca87fb34a950c377f46f76ae34436680573ddeed2b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 411, + 271, + 422 + ], + "lines": [ + { + "bbox": [ + 92, + 411, + 271, + 422 + ], + "spans": [ + { + "bbox": [ + 92, + 411, + 271, + 422 + ], + "type": "text", + "content": "(a) Diverse deformation targets and output of LeGO" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 313, + 250, + 550, + 401 + ], + "blocks": [ + { + "bbox": [ + 313, + 250, + 550, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 250, + 550, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 550, + 401 + ], + "type": "image", + "image_path": "e22b3963195b072c0d67965648471e991e66d2458316c9b8f154107a849ad188.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 411, + 481, + 421 + ], + "lines": [ + { + "bbox": [ + 349, + 411, + 481, + 421 + ], + "spans": [ + { + "bbox": [ + 349, + 411, + 481, + 421 + ], + "type": "text", + "content": "(b) Facial animation of stylized avatar" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 547, + 448 + ], + "type": "text", + "content": "Figure 1. (a) The proposed method demonstrates robustness to unseen face identities and topologies and effectively generates stylized output faces with desired topologies. (b) Our stylized avatars can be animated using 3DMM blend shapes." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "spans": [ + { + "bbox": [ + 143, + 456, + 192, + 469 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 714 + ], + "type": "text", + "content": "Recent advances in 3D face stylization have made significant strides in few to zero-shot settings. However, the degree of stylization achieved by existing methods is often not sufficient for practical applications because they are mostly based on statistical 3D Morphable Models (3DMM) with limited variations. To this end, we propose a method that can produce a highly stylized 3D face model with desired topology. Our methods train a surface deformation network with 3DMM and translate its domain to the target style using a paired exemplar. The network achieves stylization of the 3D face mesh by mimicking the style of the target using a differentiable renderer and directional CLIP losses. Additionally, during the inference process, we utilize a Mesh Agnostic Encoder (MAGE) that takes deformation target, a mesh of diverse topologies as input to the stylization process and encodes its shape into our latent space. The resulting stylized face model can be animated by commonly used 3DMM blend shapes. A set of quantitative and qualitative evaluations demonstrate that our method can produce" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 458, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 547, + 519 + ], + "type": "text", + "content": "highly stylized face meshes according to a given style and output them in a desired topology. We also demonstrate example applications of our method including image-based stylized avatar generation, linear interpolation of geometric styles, and facial animation of stylized avatars." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 530, + 387, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 387, + 542 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 387, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 550, + 545, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 657 + ], + "type": "text", + "content": "Crafting animatable stylized 3D avatars that encapsulate both personal identity and character style requires extensive efforts from skilled artists. When creating animated films, the artists design stylized 3D avatars whose facial appearance matches the theme of the entire content while putting careful effort into preserving the idiosyncrasy of the actors. Similarly, on social media, artists create numerous stylized presets so that the combinations of these presets can represent diverse identities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 658, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 696 + ], + "type": "text", + "content": "To reduce the burden of manual crafting effort, generating stylized 3D faces has been a prominent area of research. Recent attempts include 3D-aware generative ad" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 702, + 471, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 471, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 471, + 713 + ], + "type": "text", + "content": "*These authors contributed equally to this work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4505" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "versarial networks (GANs) and denoising diffusion models (DMs) [1, 12, 15, 18, 46], generating a stylized texture for the 3D morphable model (3DMM) [2, 47], text-based geometry deformation [7, 22, 24-26], and surface deformation methods [13, 43]. These methods have successfully demonstrated the possibility of producing high-quality and diverse stylized 3D faces, although each method has distinct challenges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 287, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 192 + ], + "type": "text", + "content": "We identify three key elements for stylized avatar creation so that the results can be practically useful." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 192, + 287, + 251 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 192, + 286, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 192, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 47, + 192, + 286, + 215 + ], + "type": "text", + "content": "1. Avatar creation in a desired topology that is compatible with conventional CG pipelines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 216, + 262, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 216, + 262, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 216, + 262, + 227 + ], + "type": "text", + "content": "2. Extending stylization capabilities beyond 3DMM." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "type": "text", + "content": "3. Generating stylized avatars that are animatable using blend shapes." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 251, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 371 + ], + "type": "text", + "content": "Where each component meets the standards and demands of the dynamic entertainment landscape 1) allowing creators to reuse existing animation rigs and texture maps across different models.; 2) achieving greater diversity and flexibility in expressing unique and non-conventional characters.; 3) enabling animators to achieve coherent and natural movements across various facial features and expressions. While recent methods have made meaningful strides in one or two aspects of these key elements, there has not been a method that satisfies all three elements, as summarized in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 36, + 382, + 286, + 495 + ], + "blocks": [ + { + "bbox": [ + 36, + 382, + 286, + 495 + ], + "lines": [ + { + "bbox": [ + 36, + 382, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 36, + 382, + 286, + 495 + ], + "type": "image", + "image_path": "7b4ce791e7fd7bdb263df39ef1a0b9397188fbcf2ac5bc4e333a4b1a4226b2ad.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 503, + 287, + 571 + ], + "lines": [ + { + "bbox": [ + 46, + 503, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 287, + 571 + ], + "type": "text", + "content": "Figure 2. Comparison of different stylized 3D face generation methods and their limitations in meeting key elements. 3D-aware methods cannot generate 3D face in desired topologies. 3DMM-based methods have a limited stylization capability. Text-based deformation models are not directly animatable. The proposed method meets the goal of all three components." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "To address this, we propose a novel method that can generate stylized 3D face meshes. This is achieved by translating the domain of a pre-trained surface deformation network based on one of the most widely used 3DMM model, FLAME [19], to a target style domain. We achieve this goal by first training the surface deformation network with the FLAME decoder to leverage its linear shape space combined with global expression space. During fine-tuning, we employ a directional CLIP-based domain adaptation method [16, 35, 48], widely used in 2D domain, to retain the face identity while reflecting the desired style." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 289 + ], + "type": "text", + "content": "In addition, to seamlessly integrate this 2D-based training method into the stylized 3D face generation task, we propose a hierarchical rendering scheme that captures local and global facial features, ensuring effective training and identity preservation. In the inference stage, we introduce a Mesh Agnostic Encoder (MAGE) to enable mesh agnostic stylization for an input, which we call a deformation target that has various mesh topologies. MAGE is composed of pre-trained encoders from Neural Face Rigging (NFR) [34] and latent mapping networks, which establish correspondences between shapes by encoding mesh representations into a topology-invariant latent space. This enhances the versatility and applicability of our approach in the context of 3D face stylization. As shown in Figure 1, our method can generate stylized 3D face models with varied mesh topologies that are equipped with the animation capability of 3DMM while ensuring consistency across diverse deformation target mesh representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 300, + 392, + 312 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 319, + 457, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 319, + 457, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 457, + 333 + ], + "type": "text", + "content": "2.1. Stylized 3D face generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 338, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 555 + ], + "type": "text", + "content": "With recent advances in the GANs and DMs that utilize neural fields for 3D-aware face generation [3, 10], the generation of faces in diverse styles has gained popularity [1, 12, 15, 18]. By leveraging 2D priors from generative models, which capture various patterns and variations observed in the extensive 2D training data, these methods can generate 3D faces with various styles. However, despite their success in producing consistent multi-view images through neural rendering, creating stylized faces in a desired topology is challenging, limiting their suitability for using existing graphics tools across different faces. On the other hand, 3DMM-based personalized 3D face generation methods with text-guidance [2, 47] have shown success in producing high-fidelity stylized textures for 3D face models. However, these methods confine the shape of the generated face models within the 3DMM shape space, constraining possibilities for geometric exaggerations or abstraction beyond the training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 563, + 494, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 494, + 576 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 494, + 576 + ], + "type": "text", + "content": "2.2. Learning-Based 3D Shape Network" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Recently, learning implicit functions for 3D shapes has demonstrated remarkable performance in representing complex geometric structures [8, 23, 27, 30, 31]. In particular, DIF-Net [5] adopted MLPs for learning a standard signed distance function (SDF) and a volumetric deformation function, leading to comprehensive mapping between the produced SDFs. Most related to our work, DD3C [13] utilized template deformation for 3D caricature auto-decoder. They found that modeling each shape as a deformation of a fixed temaplate surface effective compared to absolute position. We advance a step further by training a surface deformation" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4506" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 83 + ], + "type": "text", + "content": "network on 3DMM and transfer its domain into stylization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 167 + ], + "type": "text", + "content": "Another line of research aims to transfer shape deformation [39]. Recently, attempts have been made to transfer deformation utilizing examples [40, 43, 45] or text [7, 24, 25]. These methods can deform a mesh into a desired style and identity using a given example. However these deformation methods may not be as effective when it comes to coherent control of animation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 178, + 209, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 178, + 209, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 178, + 209, + 190 + ], + "type": "text", + "content": "2.3. Few-Shot Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 198, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 286, + 365 + ], + "type": "text", + "content": "Domain adaptation refers to the process of utilizing a neural network that has been initially trained on a large dataset from a source domain, and subsequently fine-tuning it on a smaller dataset from a target domain. Few-shot domain adaptation is widely researched in the 2D domain, especially with generative priors [4, 12, 20, 28, 29, 41, 42, 44]. Attempts have been made in one-shot domain adaptation to utilize the semantic capabilities of vision-language networks like CLIP [35]. These networks can provide direction for distinguishing between identity and style [6, 16, 48]. We adopt this two-way directional guidance from CLIP and combine it with a differentiable renderer to effectively stylize the face mesh while preserving the original identity using a paired exemplar." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 375, + 201, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 375, + 201, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 375, + 201, + 387 + ], + "type": "text", + "content": "2.4. Mesh Agnostic Deformation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 394, + 286, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 286, + 514 + ], + "type": "text", + "content": "Recently, mesh agnostic networks [9, 11, 32, 33, 38] demonstrated a great potential for learning 3D information such as dense correspondences between different 3D shape representations without relying on consistent topology or vertex ordering. These networks operate on meshes and encode shapes into a topology-invariant latent space. Specifically, NFR [34] proposed an autoencoder framework for facial expression retargeting across different mesh topologies. The framework utilizes separate expression and identity encoders, both functioning in a topology-agnostic manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 514, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 286, + 597 + ], + "type": "text", + "content": "We integrate components of this approach into our method by mapping their embeddings into our latent space. This network translates the output of the pre-trained encoder from NFR to the latent space of our surface deformation network. This enables any geometric shape of the deformation target to be encoded and fed into our surface deformation network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 611, + 102, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 102, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 102, + 623 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "Our research focuses on training and fine-tuning a surface deformation network to generate stylized 3D faces with diverse shapes and expressions. We start with the source face deformation network " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " that deforms the template face to a face with different identities and expressions. Thereafter, we fine-tune it into a target style face deformation network " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " using a paired exemplar. This process is outlined below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 545, + 202 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": "1. We first train " + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": " using FLAME in a self-supervised manner, enabling the creation of versatile head meshes with different shapes and expressions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "content": "2. For fine-tuning, we assemble a paired exemplar that consists of an identity exemplar mesh " + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "content": " and a style exemplar mesh " + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 306, + 108, + 545, + 144 + ], + "type": "text", + "content": ", sharing identity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": "3. Both " + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": " jointly serve as example guidance to fine-tune " + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 167, + 545, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 545, + 202 + ], + "type": "text", + "content": "4. At inference, using MAGE and " + }, + { + "bbox": [ + 306, + 167, + 545, + 202 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 306, + 167, + 545, + 202 + ], + "type": "text", + "content": ", the deformation target of diverse topologies can be translated into a stylized face." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 205, + 545, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 545, + 229 + ], + "type": "text", + "content": "A detailed description of these processes will be provided in the following subsections." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 236, + 531, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 236, + 531, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 236, + 531, + 248 + ], + "type": "text", + "content": "3.1. Deformation Network as Parametric Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "content": " is a surface deformation network that deforms a template face with the given latent vectors. We train the network to generate diverse human faces with FLAME, in a self-supervising manner. FLAME can manipulate both the global head shape and local expressions through its shape parameters " + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\vec{\\beta} \\in \\mathbb{R}^{300}" + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "content": ", and expression parameters " + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\vec{\\psi} \\in \\mathbb{R}^{100}" + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "content": ", which are the components of the FLAME parameters " + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "content": ". This empowers " + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 255, + 545, + 365 + ], + "type": "text", + "content": " to generate diverse geometric face shapes and expressions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": "To generate face meshes using these shape and expression components of " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " as input, we employ mapping networks, " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{shape}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{exp}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": ", which consist of MultiLayer Perceptrons (MLP). Each mapping network separately transforms " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\vec{\\beta}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "\\vec{\\psi}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " into latent vectors, " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 305, + 366, + 545, + 439 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 348, + 440, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 440, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 348, + 440, + 545, + 456 + ], + "type": "interline_equation", + "content": "z _ {s} = \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}), \\quad z _ {e} = \\mathcal {M} _ {\\text {e x p}} (\\vec {\\psi}) \\tag {1}", + "image_path": "f0ff4886ca0fca99d98475c02e0c47ddd0348f4c294a43f2c7b4d475e94edcfd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "content": "These latent vectors are then fed into " + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "content": ", enabling the creation of diverse face meshes imbued with a wide range of expressions. The mapping networks " + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{shape}" + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{exp}" + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "content": " are trained jointly with " + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 460, + 545, + 521 + ], + "type": "text", + "content": " during the training process. The following losses are employed:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 529, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\mathrm {L} (\\vec {\\beta}, \\vec {\\psi}) = \\left\\| \\operatorname {F L A M E} (\\vec {\\beta}, \\vec {\\psi}) - D _ {S} \\left(\\left[ \\begin{array}{l} \\mathcal {M} _ {\\text {s h a p e}} (\\vec {\\beta}) \\\\ \\mathcal {M} _ {\\exp} (\\vec {\\psi}) \\end{array} \\right]\\right) \\right\\| _ {2} ^ {2} \\tag {2}", + "image_path": "167d71d68c66cdf1d1294458c04a63fc5ec2c5cc9e42a189e3ac57c432ee5cf3.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "text", + "content": "Surface-Intensive Mesh Sampling To enable " + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "text", + "content": " to represent various geometries, we introduce a surface-intensive mesh sampling (SIMS) strategy during training. SIMS is performed by randomly selecting points from the surfaces of the face mesh, " + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "FLAME(\\vec{\\beta},\\vec{\\psi})" + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "text", + "content": ", during training " + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 305, + 572, + 545, + 669 + ], + "type": "text", + "content": ". Specifically, we sample approximately 4 times more points from the surface than the number of vertices of the face mesh. Experiment results are reported in Sec. 4.2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 677, + 358, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 358, + 687 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 358, + 687 + ], + "type": "text", + "content": "3.2. LeGO" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "The fine-tuning process utilizes a paired exemplar, " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": ", to guide the adaptation of " + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": " for the generation of" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4507" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 67, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 49, + 67, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 49, + 67, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 67, + 545, + 161 + ], + "type": "image", + "image_path": "c1ce92d59794f03772abbfd385e751411f605deb0c55f32b5e6912dd137cd0b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 162, + 296, + 337 + ], + "blocks": [ + { + "bbox": [ + 49, + 162, + 296, + 337 + ], + "lines": [ + { + "bbox": [ + 49, + 162, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 296, + 337 + ], + "type": "image", + "image_path": "c9823072a971478dd9d876b739437c5e189760037433427a52ae61e8cd35bfe3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 338, + 546, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 338, + 546, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 546, + 373 + ], + "type": "text", + "content": "Figure 3. Overview of our method: The upper box illustrates the inference stage, where our method takes diverse deformation targets and generates stylized outputs. In the lower-left box, the training process of Mesh Agnostic Encoder (MAGE) is depicted. In the lower-right box, the fine-tuning process of " + }, + { + "bbox": [ + 46, + 338, + 546, + 373 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 46, + 338, + 546, + 373 + ], + "type": "text", + "content": " is illustrated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 162, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 296, + 162, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 296, + 162, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 296, + 162, + 545, + 337 + ], + "type": "image", + "image_path": "aecb9ef82435c93ea841f65e8d8dc807b605c1e57accf11bd2365acd633e6765.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": "stylized 3D faces. An identity exemplar mesh, " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", is generated from the FLAME decoder using " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\Phi_{ref}" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", and a style exemplar mesh, " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", is manually crafted from " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ". During each iteration of fine-tuning, " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "\\Phi_{sample}" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": " is sampled randomly to generate face mesh " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_S^{samp}" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "D_S" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": " and corresponding stylized 3D face " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_T^{samp}" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ". In the fine-tuning process, quantities of " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_S" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "D_S" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 384, + 287, + 480 + ], + "type": "text", + "content": " are all singular, representing one pair of meshes and one pair of networks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": "We introduce a hierarchical rendering scheme designed to preserve semantically important features, such as the shape and facial components, in the face mesh. With a differentiable renderer, this approach captures significant facial features from both local and global perspectives, enhancing stylization fidelity and identity preservation. To further enhance this process, we incorporate 2D-based losses, including CLIP reconstruction loss " + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": ", CLIP indomain loss " + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": ", CLIP across-domain loss " + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "inline_equation", + "content": "L_{across}" + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": ", and 3D-based losses such as vertex reconstruction loss " + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "inline_equation", + "content": "L_{vert}" + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": " and style loss " + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "inline_equation", + "content": "L_{style}" + }, + { + "bbox": [ + 46, + 481, + 288, + 624 + ], + "type": "text", + "content": ". Details of each loss are elaborated in Section 3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 634, + 140, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 140, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 140, + 646 + ], + "type": "text", + "content": "3.3. Loss Functions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Vertex Reconstruction Loss The vertex reconstruction loss " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "L_{vert}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " guides " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " in learning to deform a style exemplar mesh " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "[z_s^{ref}, z_e^{ref}]" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", which is mapped from " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\Phi_{ref}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". The loss utilizes Mean Squared Error (MSE) to ensure that the vertices of the predicted mesh " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "M_T^*" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ", generated by " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "D_T" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 384, + 526, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 384, + 526, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 526, + 396 + ], + "type": "text", + "content": "closely match " + }, + { + "bbox": [ + 305, + 384, + 526, + 396 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 305, + 384, + 526, + 396 + ], + "type": "text", + "content": ". The loss can be written as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 375, + 406, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 406, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 375, + 406, + 545, + 419 + ], + "type": "interline_equation", + "content": "L _ {v e r t} = \\left\\| M _ {T} - M _ {T} ^ {*} \\right\\| _ {2} ^ {2} \\tag {3}", + "image_path": "350b373e91875a8d42bab570a06272011865572216f3a9ff2ea1832ad820a424.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": "CLIP Reconstruction Loss The CLIP reconstruction loss " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " serves to maintain semantic consistency between the deformed mesh " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "M_T^*" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " in the CLIP space. This is crucial because even minor displacements in 3D space can result in surface irregularities or undesirable shading variations. The CLIP reconstruction loss can be written as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 520, + 545, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 545, + 553 + ], + "type": "interline_equation", + "content": "L _ {\\text {C L I P}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {*}\\right)\\right) \\| _ {2} ^ {2} \\tag {4}", + "image_path": "9fdf3e7880b8ba63896eb2c6aaf69d6bcd6b5d63c5962e248016e0d07800e9ee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "inline_equation", + "content": "R_{l,v}" + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "text", + "content": " represents the hierarchical rendering of a differentiable renderer from level " + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "text", + "content": " and view direction " + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "text", + "content": ". Each view is anchored at a predefined position on a face mesh, including the front and both sides of the face. Each level corresponds to the distance from the face to the camera. At the highest level, where the face is captured in close-up, additional images are rendered from significant facial features such as the nose, eyes, and lips. " + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "inline_equation", + "content": "E_C" + }, + { + "bbox": [ + 304, + 554, + 545, + 662 + ], + "type": "text", + "content": " encodes rendered images into CLIP embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "CLIP Directional Loss Inspired from one-shot stylization methods in the 2D domain [48], we incorporate the CLIP in-domain loss " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and the CLIP across-domain loss" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4508" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "inline_equation", + "content": "L_{across}" + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "inline_equation", + "content": "L_{in}" + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": " ensures that the direction of two distinct faces in the source domain remains consistent in two distinct stylized faces of the target style domain. In contrast, " + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "inline_equation", + "content": "L_{across}" + }, + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": " ensures that the direction of a face from the source domain and its corresponding stylized face in the target domain is preserved across different faces and their corresponding stylized faces. This can be shown in the lower right of Figure 3. Therefore our losses for CLIP in-domain and across-domain can be written as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 189, + 287, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 189, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 287, + 228 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {i n}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\tag {5} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {s a m p}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) \\right\\| _ {2} ^ {2} \\\\ \\end{array}", + "image_path": "416c14ec8b332864d2492a8794b7aa26ce8012882c5d71adf4f8c00d643642c1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 239, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 287, + 277 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {a c r o s s}} = \\sum_ {l \\in L} \\sum_ {v \\in V _ {l}} \\| \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T}\\right)\\right) - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {S}\\right)\\right)\\right) \\right. \\tag {6} \\\\ - \\left(E _ {C} \\left(R _ {l, v} \\left(M _ {T} ^ {\\text {s a m p}}\\right)\\right) - E _ {C} \\left(R _ {l, v} \\left(M _ {S} ^ {\\text {s a m p}}\\right)\\right)\\right) \\| _ {2} ^ {2} \\\\ \\end{array}", + "image_path": "36255630f38d0d30dea454cab41bedca12df611211d8dc7f016652783af19578.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "M_S^{samp}" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "M_T^{samp}" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": " refer to " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "D_S[z_s^{samp};z_e^{samp}]" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "D_T[z_s^{samp};z_e^{samp}]" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "[z_s^{samp};z_e^{samp}]" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": " is the latent vectors mapped from " + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "inline_equation", + "content": "\\Phi_{sample}" + }, + { + "bbox": [ + 47, + 284, + 287, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": "Style Loss We introduce a novel style loss " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "L_{style}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " to capture the style of " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " by utilizing surface normals. Style loss compares surface normal, which are strongly correlated with the semantic information of the mesh [5]. A straightforward approach would involve comparing " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T^{samp}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": ". However, this will constrain " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T^{samp}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " to have same expression as " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": ", which could degrade the animatability. To address this, we construct a pseudo pair for normal calculations. This is done by generating a face mesh through the concatenation of " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "z_e^{ref}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "z_s^{samp}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": ". This approach is utilized to compute the style loss, " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "L_{style}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " without degrading the animatability while enhancing style adherence. In short, " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "L_{style}" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " ensures alignment of the surface normals from " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "D_T([z_s^{samp}; z_e^{ref}])" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 46, + 335, + 289, + 515 + ], + "type": "text", + "content": ". The formulation of the style loss can be written as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 523, + 287, + 554 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 523, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 104, + 523, + 287, + 554 + ], + "type": "interline_equation", + "content": "L _ {\\text {s t y l e}} = \\sum_ {f \\in S} \\left(1 - \\frac {n _ {f} \\cdot n _ {f} ^ {\\prime}}{\\left| n _ {f} \\right| \\left| n _ {f} ^ {\\prime} \\right|}\\right) \\tag {7}", + "image_path": "a05e1804b377928c5451024fc5f8855eb2753ddb1ae68e4991f8f89b1e0510b3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "inline_equation", + "content": "n_f" + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "content": " refers to the surface normal from " + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "inline_equation", + "content": "M_T" + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "inline_equation", + "content": "n'_f" + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "content": " refers to the corresponding normal from " + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "inline_equation", + "content": "D_T([z_s^{samp}; z_e^{ref}])" + }, + { + "bbox": [ + 47, + 559, + 286, + 583 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 583, + 277, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 583, + 277, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 583, + 277, + 594 + ], + "type": "text", + "content": "The final objective function can be written as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 73, + 603, + 287, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 603, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 73, + 603, + 287, + 632 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {t o t a l} = \\lambda_ {v e r t} L _ {v e r t} + \\lambda_ {\\mathrm {C L I P}} L _ {\\mathrm {C L I P}} + \\lambda_ {i n} L _ {i n} \\tag {8} \\\\ + \\lambda_ {a c r o s s} L _ {a c r o s s} + \\lambda_ {s t y l e} L _ {s t y l e} \\\\ \\end{array}", + "image_path": "e2e22fb1a982a3ca4514e67b80f5b18bed1fc4ed713baac314db672e35d3c697.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{vert}" + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{in}" + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{across}" + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\lambda_{style}" + }, + { + "bbox": [ + 47, + 639, + 287, + 663 + ], + "type": "text", + "content": " are the weight for each loss term." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 182, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 182, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 182, + 684 + ], + "type": "text", + "content": "3.4. Mesh Agnostic Encoder" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "NFR [34] employs DiffusionNet [38] for encoding and decoding face meshes, facilitating face retargeting across" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "different topologies. Given NFR's proficiency in extracting identity and expression details from faces with varying topologies, we extend its encoders to create a MAGE, as depicted in Figure 3. MAGE consists of ID2ID and exp2exp, both being MLPs that receive embeddings from each ID encoder and expression encoder pre-trained in NFR. Finally, given intermediate vectors from both ID2ID and exp2exp, the latent mapper outputs the latent vectors for " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": "MAGE is trained in a self-supervised manner by randomly sampling " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": " and passing them through the mapping networks of " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "[z_{s};z_{e}]" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": ". Training involves comparing " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "[\\hat{z}_s;\\hat{z}_e]" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": ", predicted from MAGE, with their corresponding ground truth latent vectors " + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "inline_equation", + "content": "[z_s;z_e]" + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": ". The objective function of the encoder uses a MSE loss as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 251, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 251, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 367, + 251, + 545, + 266 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {e n c}} = \\left\\| \\left[ z _ {s}; z _ {e} \\right] - \\left[ \\hat {z} _ {s}; \\hat {z} _ {e} \\right] \\right\\| _ {2} ^ {2} \\tag {9}", + "image_path": "bdc90b8787fd818635b97effdafb6f21bfdacd4ec74f5518874e7067f17b3773.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "inline_equation", + "content": "[ \\hat{z}_s; \\hat{z}_e ]" + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": " is MAGE " + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "inline_equation", + "content": "[D_S([z_s;z_e])]" + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": ". With this encoder, we can project a face mesh with diverse topologies into the latent space of " + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "inline_equation", + "content": "D_S" + }, + { + "bbox": [ + 304, + 271, + 545, + 308 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 316, + 373, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 373, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 373, + 328 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "content": "After training, " + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "content": " is capable of generating a stylized 3D face mesh with a desired topology from a deformation target mesh that has an arbitrary topology. This is accomplished by first projecting the deformation target into a latent vector, using MAGE. This latent vector is then fed as input to " + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 304, + 335, + 545, + 443 + ], + "type": "text", + "content": ", allowing it to deform the template face of the desired topology into a stylized 3D face mesh. The resulting stylized face preserves both the identity of the deformation target and the desired style while providing animation control." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 455, + 388, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 455, + 388, + 468 + ], + "spans": [ + { + "bbox": [ + 305, + 455, + 388, + 468 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 472, + 434, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 434, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 434, + 483 + ], + "type": "text", + "content": "4.1. Qualitative Evaluation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 545, + 548 + ], + "type": "text", + "content": "We evaluate our method based on three key elements for stylized 3D head avatar creation: Avatar creation in a desired topology, stylization beyond 3DMM, and animation capability with blend shapes. Also, implementation details are provided in supplementary material." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 545, + 725 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 725 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 725 + ], + "type": "text", + "content": "Avatar in a Desired Topology To evaluate the capability of our method in producing a mesh with a desired topology from the deformation target of arbitrary topologies, we first apply a re-meshing technique that includes Loop subdivision [21] and mesh simplification [37] to meshes from FLAME. We also use original meshes from the CoMA [36] and ICT-FaceKit [17] directly as additional diverse meshes. To visually confirm that our method can align different mesh topologies with a desired template face, we enhance the resulting meshes with a checkerboard texture. The visual results, as shown in Figure 4, demonstrate that our method consistently produces stylized 3D faces with the desired topology, regardless of variations in the topology of the deformation target." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4509" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 286, + 304 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 286, + 304 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 286, + 304 + ], + "type": "image", + "image_path": "c1b383397d4ace2329250b51d7f5ed7551c817793476eef4fffb332c323b633f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 312, + 288, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 335 + ], + "type": "text", + "content": "Figure 4. Demonstration of stylized 3D faces with desired topology, regardless of deformation target variations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 354, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 449 + ], + "type": "text", + "content": "Stylization Capability We evaluate our method across various styles to assess its stylization capability. As shown in Figure 5, our method can generate a broad spectrum of styles encompassing human-like and non-human-like faces while preserving the original identities of the given face mesh. Our method also demonstrates the capability to achieve stylization across various geometries, including face masks and point clouds." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 462, + 286, + 666 + ], + "blocks": [ + { + "bbox": [ + 50, + 462, + 286, + 666 + ], + "lines": [ + { + "bbox": [ + 50, + 462, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 50, + 462, + 286, + 666 + ], + "type": "image", + "image_path": "b8b9308a7bd4554173aa44cc65e74da2372f3ffc683567f71653c0c3528b4e97.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "lines": [ + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 287, + 715 + ], + "type": "text", + "content": "Figure 5. Stylization results across diverse styles and identities. Our approach generates varied styles while preserving the deformation target identity and generalizing to diverse geometric representations like masks and point clouds." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 145 + ], + "type": "text", + "content": "Animation Capability Using parameters from 3DMM, our method can animate facial expressions in the resulting stylized 3D faces. As illustrated in Figure 6, our method can generate 3D stylized faces with various expressions. This allows applications such as video-driven stylized talking heads, which are elaborated in supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 156, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 309, + 156, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 309, + 156, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 309, + 156, + 545, + 289 + ], + "type": "image", + "image_path": "3693ca614ca6b7633101a8f131125b1f858b3d45cfac1212c7561c87f0b5ea4b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 302, + 545, + 324 + ], + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 324 + ], + "type": "text", + "content": "Figure 6. Visualization of dynamic expressions in stylized 3D faces." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 339, + 441, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 339, + 441, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 339, + 441, + 352 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 358, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 358, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 304, + 358, + 545, + 549 + ], + "type": "text", + "content": "For a quantitative evaluation, we measured the average CLIP style preservation (CLIP-SP) and CLIP identity preservation (CLIP-IP) scores. These metrics reflect the trade-off between preserving style and identity, making their averages crucial for validating stylization. For both metrics, we calculated the cosine similarity of CLIP embeddings from rendered meshes. Specifically, for CLIP-SP, we compared the embeddings of the generated stylized face mesh and style exemplar mesh. For CLIP-IP, we compared the embeddings of the generated stylized face mesh and deformation target mesh. For dataset, we sampled 8 different face mesh from FLAME without expression and manually crafted stylized mesh corresponding to each identities. Also, for the quantitative evaluation, we randomly sampled another 10 different identities without expression to be used as deformation targets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Comparison with Baselines on Mesh Stylization We compared our stylized face generation results with those produced by baselines that can deform a mesh into different styles: Deformation Transfer [39], TextDeformer [7], and X-mesh [22]. In case of Deformation Transfer [39], we identified the correspondences between the identity exemplar mesh and the deformation target using 68 facial landmarks along with 9 additional points on the forehead in order to apply learned deformation. TextDeformer and X-mesh employ a text-guided deformation that operates on an input mesh. For both methods, the input face is deformed based on the same descriptive text to generate stylized outputs. All these baselines were evaluated on three different" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4510" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 98, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results are presented for comparison with baselines and ablations, featuring averages of CLIP-SP and CLIP-IP. The highest scores are denoted in bold, while the second highest scores are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 98, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 48, + 98, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 98, + 547, + 196 + ], + "type": "table", + "html": "
Mesh TypeFLAMELoopSimplifiedOverall-Average
MethodsCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageCLIP-SPCLIP-IPAverageAverage
Ours0.98670.93470.96070.98330.93500.95920.95590.89870.92730.9491
Ours w/o Lstyle0.98090.93740.95910.97680.93390.95530.95420.89870.92650.9470
Ours w direct Lstyle0.98650.93480.96070.98330.93560.95950.95500.89730.92620.9488
w/o hierarchical rendering0.98460.93490.95970.98240.93340.95790.95690.89850.92770.9484
Text deformer0.86410.83630.85020.86300.84250.85280.85340.81730.83530.8461
X-mesh0.86260.86020.86140.88170.89500.88840.86350.84480.85410.8680
Deformation transfer0.97680.93760.95720.95980.95760.95870.92540.92880.92710.9477
", + "image_path": "af9a6089dd28008237c4a66e5b817fdfb96024e6cbcefece5f002fd961c9335e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 287, + 232 + ], + "type": "text", + "content": "mesh topologies: FLAME and FLAME with Loop subdivision and mesh simplification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 233, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 340 + ], + "type": "text", + "content": "We present qualitative comparison results on stylized 3D face generation in Figure 7. Deformation Transfer is capable of generating faces that follow the style; however, it exhibits artifacts, including anomalies and concavities on the surface due to the calculation of displacement and its direct transfer. In contrast, text-based methods fall short in following styles. Our method, on the other hand, can generate stylized 3D faces that adhere to both the input identity and style example without any artifacts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 340, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 425 + ], + "type": "text", + "content": "In quantitative evaluation, shown in Table 1, our method obtained the highest average scores computed from CLIP-SP and CLIP-IP. This clearly demonstrates that our approach achieves better performance in mesh stylization while preserving input identity compared to deformation baselines. More details and examples are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 434, + 286, + 578 + ], + "blocks": [ + { + "bbox": [ + 52, + 434, + 286, + 578 + ], + "lines": [ + { + "bbox": [ + 52, + 434, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 52, + 434, + 286, + 578 + ], + "type": "image", + "image_path": "49ae9faeb5d6783b98f9bcdac24cfd62f117a32575ccc4a4e3b01af0baa00a24.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 588, + 287, + 622 + ], + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 622 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison on stylization. Our method adheres to style and identity without artifacts unlike Deformation Transfer. Text-based approaches fail to match styles from text." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Surface-Intensive Mesh Sampling To test the proposed sampling method SIMS, we conducted a study on the reconstruction task comparing three different sampling variants for training " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ". SIMS is surface-intensive sampling, using approximately four times more points sampled from the mesh surface compared to mesh vertices. Hybrid sampling, originally proposed in DD3C [13], sampled points both ver" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 224, + 535, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 207, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 306, + 207, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 306, + 207, + 545, + 218 + ], + "type": "text", + "content": "Table 2. Reconstruction comparison with mesh sampling methods" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 224, + 535, + 277 + ], + "lines": [ + { + "bbox": [ + 315, + 224, + 535, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 535, + 277 + ], + "type": "table", + "html": "
Reconstruction loss↓ (all in e-5)
Methodsoriginalsimplifiedloop-1loop-2average
SIMS(Ours)1.7901.6351.6121.5911.657
Hybrid2.3002.0192.1382.1182.144
Vertex1.9555.8663.2582.9593.509
", + "image_path": "efde3f5507e568cc21ea0a3cfe05dab14c8fa831e7d54e59b26bdaaae6fdf9e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "type": "text", + "content": "tices and faces in similar ratio (" + }, + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\sim 1.1" + }, + { + "bbox": [ + 305, + 297, + 545, + 319 + ], + "type": "text", + "content": " times). Vertex-only used just the vertices of the mesh." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 321, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 321, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 321, + 545, + 430 + ], + "type": "text", + "content": "For the experiment, we calculated the reconstruction loss on four different experiment settings: (1) original FLAME, (2) a simplified FLAME mesh with 1/4 the vertex count, (3) FLAME with one loop subdivision, and (4) FLAME with two loop subdivisions. The reconstruction loss was compared to the ground truth position using mean squared error. As shown in Table 2, SIMS achieved the lowest error by a large margin on all experiments. Hybrid sampling generally performed better than vertex-only sampling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": "Ablation Study To perform an ablation study, we trained our model " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": " using different settings: 1) \"Ours w/o " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": ",\" where the style loss was removed; 2) \"Ours w direct " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": ",\" where the style loss was directly compared between " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}])" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "M_{T}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": "; 3) \"Ours w/o hierarchical rendering,\" where a single canonical view was used to calculate the " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{in}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{across}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": ". The results are displayed in Figure 8 and Table 1. The results produced by \"Ours\" and \"Ours w direct " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": "\" do not show any artifacts unlike the results produced by w/o " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": " and w/o hierarchical rendering, which exhibit surface artifacts. However, \"Ours w. direct " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": "\" forces " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "D_{T}([z_{s}^{\\text{samp}}; z_{e}^{\\text{samp}}])" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "M_{T}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": ", which have different expressions, to have the same normals; thus, it discards the expression. As a result, while \"Ours w. direct " + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "inline_equation", + "content": "L_{\\text{style}}" + }, + { + "bbox": [ + 305, + 436, + 546, + 652 + ], + "type": "text", + "content": "\" does stylize effectively, it cannot animate. Conversely, \"Ours\" not only stylizes well but is also animatable. Additional ablation study results regarding animation are presented in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 547, + 713 + ], + "type": "text", + "content": "User Study We conducted a user study with 34 participants to evaluate different stylized face generation methods on human perception. A total of eight different styles were used. Each participant was presented with 20 questions and asked to choose the result that best followed the style while" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4511" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 287, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 287, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 287, + 148 + ], + "type": "image", + "image_path": "b565d288cf8580c30a2dd2e2db458e9c19dce4818526603c989ad5ae44b0b181.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 156, + 242, + 167 + ], + "lines": [ + { + "bbox": [ + 91, + 156, + 242, + 167 + ], + "spans": [ + { + "bbox": [ + 91, + 156, + 242, + 167 + ], + "type": "text", + "content": "Figure 8. Visualization on ablation study." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 243 + ], + "type": "text", + "content": "preserving the identity. As shown in Table 3, our method received higher scores compared to the baseline methods. Ours outperformed the text-based stylization methods by a large margin and scored higher than the method requiring correspondence specification." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 102, + 291, + 230, + 337 + ], + "blocks": [ + { + "bbox": [ + 46, + 254, + 287, + 287 + ], + "lines": [ + { + "bbox": [ + 46, + 254, + 287, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 254, + 287, + 287 + ], + "type": "text", + "content": "Table 3. Results from the user perceptual study given style exemplar and the deformation target. The percentage represents the selected frequency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 291, + 230, + 337 + ], + "lines": [ + { + "bbox": [ + 102, + 291, + 230, + 337 + ], + "spans": [ + { + "bbox": [ + 102, + 291, + 230, + 337 + ], + "type": "table", + "html": "
MethodUser Score
Ours60.65%
Deformation transfer38.17%
Text deformer0.44%
X-mesh0.74%
", + "image_path": "9c2b9c43ee7b2fe2e8da00510f4c4c4a3c372ff3d721bf51b441ba9582629742.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 346, + 127, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 127, + 358 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 127, + 358 + ], + "type": "text", + "content": "5. Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 365, + 157, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 157, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 157, + 378 + ], + "type": "text", + "content": "5.1. Style Interpolation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "content": "LeGO enables linear style interpolation as illustrated in Figure 9, through weight blending. Diverse styles are seamlessly blended to create new stylized meshes by linearly interpolating weights from " + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "inline_equation", + "content": "D_{T1}" + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "inline_equation", + "content": "D_{T2}" + }, + { + "bbox": [ + 46, + 384, + 287, + 443 + ], + "type": "text", + "content": " using following equation:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 447, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 447, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 97, + 447, + 287, + 460 + ], + "type": "interline_equation", + "content": "W _ {\\text {n e w}} = \\alpha W _ {D _ {T 1}} + (1 - \\alpha) W _ {D _ {T 2}} \\tag {10}", + "image_path": "12f44a8b0614145060c8fc8bda88e5d8553919964c71dde7e006dd81d8631bec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "inline_equation", + "content": "W_{D_{T1}}" + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "inline_equation", + "content": "W_{D_{T2}}" + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": " denote the network weights for " + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "inline_equation", + "content": "D_{T1}" + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "inline_equation", + "content": "D_{T2}" + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": ", respectively, while " + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 470, + 287, + 506 + ], + "type": "text", + "content": " represents the blending weight controlling the interpolation." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 517, + 287, + 594 + ], + "blocks": [ + { + "bbox": [ + 50, + 517, + 287, + 594 + ], + "lines": [ + { + "bbox": [ + 50, + 517, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 50, + 517, + 287, + 594 + ], + "type": "image", + "image_path": "7570eedb8d42d8e498594d4d2147b9431b72da8842d2276191bfbaf4f28cceb1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 603, + 263, + 614 + ], + "lines": [ + { + "bbox": [ + 70, + 603, + 263, + 614 + ], + "spans": [ + { + "bbox": [ + 70, + 603, + 263, + 614 + ], + "type": "text", + "content": "Figure 9. Linear interpolation of Style A and Style B" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 629, + 275, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 275, + 642 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 275, + 642 + ], + "type": "text", + "content": "5.2. Image-based 3D Stylized Avatar Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 648, + 287, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 648, + 287, + 719 + ], + "spans": [ + { + "bbox": [ + 46, + 648, + 287, + 719 + ], + "type": "text", + "content": "A stylized 3D face can be generated from a single portrait by first using methods that reconstruct 3D faces from 2D portraits. Among these methods, we employed MICA [49] to reconstruct a 3D face from an image. This reconstructed shape was then fed into LeGO to create stylized faces. The results are visualized in Figure 10." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 392, + 71, + 538, + 121 + ], + "blocks": [ + { + "bbox": [ + 392, + 71, + 538, + 121 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 538, + 121 + ], + "type": "image", + "image_path": "c72e2a57b95ad2397a6b5ca4b46376fe154b673f43951e9e564fb428f873ee51.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 313, + 110, + 389, + 232 + ], + "blocks": [ + { + "bbox": [ + 313, + 110, + 389, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 110, + 389, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 389, + 232 + ], + "type": "image", + "image_path": "f3792c6cdf2a385cafff6dfafd3a563b0080dd89454b4e53acd5db4397d7e2a8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 242, + 545, + 264 + ], + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 264 + ], + "type": "text", + "content": "Figure 10. Visualization of generating stylized 3D faces from 2D portraits [14]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 121, + 533, + 232 + ], + "blocks": [ + { + "bbox": [ + 395, + 121, + 533, + 232 + ], + "lines": [ + { + "bbox": [ + 395, + 121, + 533, + 232 + ], + "spans": [ + { + "bbox": [ + 395, + 121, + 533, + 232 + ], + "type": "image", + "image_path": "b7b2331c88f4bc2663c3387d46399f258445701aa4f19682747b966667d8f119.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 280, + 457, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 457, + 293 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 457, + 293 + ], + "type": "text", + "content": "6. Limitation and Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 396 + ], + "type": "text", + "content": "We presented a novel approach for generating stylized 3D face meshes, considering three key elements. First, we proposed a surface deformation network that can generate a face in the desired topology using SIMS. Second, by the domain adaptation with hierarchical rendering, we achieved superior stylization capability. Lastly, Using 3DMM prior, we can generate a stylized face equipped with the animation capability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 396, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 480 + ], + "type": "text", + "content": "In addition, we proposed MAGE for practical usage, which can take diverse mesh topologies as input, and a novel style loss that adheres to the style effectively while preserving animation ability. Comprehensive experimental results demonstrate that our method is capable of generating a stylized mesh with consistent topology given deformation target meshes exhibiting significant topological variation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 481, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 564 + ], + "type": "text", + "content": "While our method shows promising results and enables significant advancements in practical avatar creation, it also has challenges to address. At inference, achieving a stylized output with the same topology as the input requires a two-stage process involving template replacement with the mean face mesh. Addressing these efficiency and practicality challenges is crucial for further enhancing our approach." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2023 (Project Name: Development of Universal Fashion Creation Platform Technology for Avatar Personality Expression, Project Number: RS-2023-00228331, Contribution Rate: " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " ) and Korea Institute for Advancement of Technology(KIAT) grant funded by the Korea Government(MOTIE) (Project Name: The Competency Development Program for Industry Specialist, Project Number: P0012746, Contribution Rate: " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4512" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Rameen Abdal, Hsin-Ying Lee, Peihao Zhu, Menglei Chai, Aliaksandr Siarohin, Peter Wonka, and Sergey Tulyakov. 3davatargan: Bridging domains for personalized editable avatars. arXiv preprint arXiv:2301.02700, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "text", + "content": "[2] Shivangi Aneja, Justus Thies, Angela Dai, and Matthias Nießner. Clipface: Text-guided editing of textured 3d morphable models. arXiv preprint arXiv:2212.01406, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 235 + ], + "type": "text", + "content": "[3] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16123-16133, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 280 + ], + "type": "text", + "content": "[4] Min Jin Chong and David Forsyth. Jojogan: One shot face stylization. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 128-152. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 334 + ], + "type": "text", + "content": "[5] Yu Deng, Jiaolong Yang, and Xin Tong. Deformed implicit field: Modeling 3d shapes with learned dense correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10286-10296, 2021. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 378 + ], + "type": "text", + "content": "[6] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 423 + ], + "type": "text", + "content": "[7] William Gao, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. arXiv preprint arXiv:2304.13348, 2023. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 457 + ], + "type": "text", + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 458, + 288, + 513 + ], + "type": "text", + "content": "[9] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3d surface generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 216-224, 2018. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 556 + ], + "type": "text", + "content": "[10] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 601 + ], + "type": "text", + "content": "[11] Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. MeshCNN: a network with an edge. ACM Transactions on Graphics (ToG), 38(4):1-12, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 646 + ], + "type": "text", + "content": "[12] Wonjoon Jin, Nuri Ryu, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dr. 3d: Adapting 3d gans to artistic drawings. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[13] Yucheol Jung, Wonjong Jang, Soongjin Kim, Jiaolong Yang, Xin Tong, and Seungyong Lee. Deep deformable 3d caricatures with learned shape control. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 151 + ], + "type": "text", + "content": "[15] Gwanghyun Kim and Se Young Chun. Datid-3d: Diversity-preserved domain adaptation using text-to-image diffusion for 3d generative model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14203–14213, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 547, + 196 + ], + "type": "text", + "content": "[16] Seongtae Kim, Kyoungkook Kang, Geonung Kim, Seung-Hwan Baek, and Sunghyun Cho. Dynagan: Dynamic few-shot adaptation of gans to multiple domains. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 547, + 262 + ], + "type": "text", + "content": "[17] Ruilong Li, Karl Bladin, Yajie Zhao, Chinmay Chinara, Owen Ingraham, Pengda Xiang, Xinglei Ren, Pratusha Prasad, Bipin Kishore, Jun Xing, et al. Learning formation of physically-based face attributes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3410-3419, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "text", + "content": "[18] Shaoxu Li. Instruct-video2 avatar: Video-to- avatar generation with instructions. arXiv preprint arXiv:2306.02903, 2023.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 331 + ], + "type": "text", + "content": "[19] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[20] Yijun Li, Richard Zhang, Jingwan Lu, and Eli Shechtman. Few-shot image generation with elastic weight consolidation. arXiv preprint arXiv:2012.02780, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 388 + ], + "type": "text", + "content": "[21] Charles Loop. Smooth subdivision surfaces based on triangles. 1987. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 389, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 389, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 389, + 545, + 455 + ], + "type": "text", + "content": "[22] Yiwei Ma, Xiaqing Zhang, Xiaoshuai Sun, Jiayi Ji, Haowei Wang, Guannan Jiang, Weilin Zhuang, and Rongrong Ji. X-mesh: Towards fast and accurate text-driven 3d stylization via dynamic textual guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2749-2760, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 511 + ], + "type": "text", + "content": "[23] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[24] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 612 + ], + "type": "text", + "content": "[25] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 Conference Papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[26] Thu Nguyen-Phuoc, Gabriel Schwartz, Yuting Ye, Stephen Lombardi, and Lei Xiao. Alteredavatar: Stylizing dynamic 3d avatars with fast style adaptation. arXiv preprint arXiv:2305.19245, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[27] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Occupancy flow: 4d reconstruction by learning particle dynamics. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5379-5389, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4513" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] Atsuhiro Noguchi and Tatsuya Harada. Image generation from small datasets via batch statistics adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2750-2758, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 173 + ], + "type": "text", + "content": "[29] Utkarsh Ojha, Yijun Li, Jingwan Lu, Alexei A Efros, Yong Jae Lee, Eli Shechtman, and Richard Zhang. Few-shot image generation via cross-domain correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 285 + ], + "type": "text", + "content": "[31] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 523-540. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 340 + ], + "type": "text", + "content": "[32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 386 + ], + "type": "text", + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 430 + ], + "type": "text", + "content": "[34] Dafei Qin, Jun Saito, Noam Aigerman, Thibault Groueix, and Taku Komura. Neural face rigging for animating and retargeting facial meshes in the wild. arXiv preprint arXiv:2305.08296, 2023. 2, 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 433, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 498 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[36] Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European conference on computer vision (ECCV), pages 704-720, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[37] William J Schroeder, Jonathan A Zarge, and William E Lorensen. Decimation of triangle meshes. In Proceedings of the 19th annual conference on Computer graphics and interactive techniques, pages 65-70, 1992. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[38] Nicholas Sharp, Souhaib Attaiki, Keenan Crane, and Maks Ovsjanikov. Diffusionnet: Discretization agnostic learning on surfaces. ACM Transactions on Graphics (TOG), 41(3): 1-16, 2022. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 667 + ], + "type": "text", + "content": "[39] Robert W Sumner and Jovan Popovic. Deformation transfer for triangle meshes. ACM Transactions on graphics (TOG), 23(3):399-405, 2004. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[40] Minhyuk Sung, Zhenyu Jiang, Panos Achlioptas, Niloy J Mitra, and Leonidas J Guibas. Deformsyncnet: Deformation transfer via synchronized shape deformation spaces. arXiv preprint arXiv:2009.01456, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 574 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 139 + ], + "type": "text", + "content": "[41] Yaxing Wang, Abel Gonzalez-Garcia, David Berga, Luis Herranz, Fahad Shahbaz Khan, and Joost van de Weijer. Minegan: effective knowledge transfer from gans to target domains with few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9332-9341, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 194 + ], + "type": "text", + "content": "[42] Jiayu Xiao, Liang Li, Chaofei Wang, Zheng-Jun Zha, and Qingming Huang. Few shot generative model adaption via relaxed spatial structural alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11204-11213, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 239 + ], + "type": "text", + "content": "[43] Xirui Yan, Zhenbo Yu, Bingbing Ni, and Hang Wang. Cross-species 3d face morphing via alignment-aware controller. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3018-3026, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 240, + 545, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 294 + ], + "type": "text", + "content": "[44] Shuai Yang, Liming Jiang, Ziwei Liu, and Chen Change Loy. Pastiche master: Exemplar-based high-resolution portrait style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7693-7702, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 350 + ], + "type": "text", + "content": "[45] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3d deformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 75-83, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 353, + 545, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 406 + ], + "type": "text", + "content": "[46] Chi Zhang, Yiwen Chen, Yijun Fu, Zhenglin Zhou, Gang Yu, Billzb Wang, Bin Fu, Tao Chen, Guosheng Lin, and Chunhua Shen. StyleAvatar3d: Leveraging image-text diffusion models for high-fidelity 3d avatar generation. arXiv preprint arXiv:2305.19012, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 545, + 461 + ], + "type": "text", + "content": "[47] Longwen Zhang, Qiwei Qiu, Hongyang Lin, Qixuan Zhang, Cheng Shi, Wei Yang, Ye Shi, Sibei Yang, Lan Xu, and Jingyi Yu. Dreamface: Progressive generation of animatable 3d faces under text guidance. arXiv preprint arXiv:2304.03117, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 464, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 545, + 517 + ], + "type": "text", + "content": "[48] Peihao Zhu, Rameen Abdal, John Femiani, and Peter Wonka. Mind the gap: Domain gap control for single shot domain adaptation for generative adversarial networks. In International Conference on Learning Representations, 2022. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 519, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 519, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 519, + 545, + 574 + ], + "type": "text", + "content": "[49] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Towards metrical reconstruction of human faces. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XIII, pages 250-269. Springer, 2022. 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4514" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_content_list.json b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e4d4a5c023e9d6ed2225bc9d3315055ce30045a3 --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_content_list.json @@ -0,0 +1,1527 @@ +[ + { + "type": "text", + "text": "Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning", + "text_level": 1, + "bbox": [ + 99, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joshua C. Zhao, Ahaan Dabholkar, Atul Sharma, Saurabh Bagchi Purdue University", + "bbox": [ + 225, + 203, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhao1207,adabholk,sharm438,sbagchi}@purdue.edu", + "bbox": [ + 274, + 241, + 691, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Federated learning is a decentralized learning paradigm introduced to preserve privacy of client data. Despite this, prior work has shown that an attacker at the server can still reconstruct the private training data using only the client updates. These attacks are known as data reconstruction attacks and fall into two major categories: gradient inversion (GI) and linear layer leakage attacks (LLL). However, despite demonstrating the effectiveness of these attacks in breaching privacy, prior work has not investigated the usefulness of the reconstructed data for downstream tasks. In this work, we explore data reconstruction attacks through the lens of training and improving models with leaked data. We demonstrate the effectiveness of both GI and LLL attacks in maliciously training models using the leaked data more accurately than a benign federated learning strategy. Counter-intuitively, this bump in training quality can occur despite limited reconstruction quality or a small total number of leaked images. Finally, we show the limitations of these attacks for downstream training, individually for GI attacks and for LLL attacks.", + "bbox": [ + 75, + 323, + 473, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 638, + 209, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With growing concerns of data privacy, federated learning (FL) [17] has gained traction as a potential privacy-preserving method for training machine learning models. Compared to centralized learning where training is done on data localized at a central server, FL takes a decentralized approach where participating clients train a model on their local data and send their updates to the server. A typical training round involves a server sending a model to the clients, the clients training the model using their local data, and finally having the clients send their updates to the server for aggregation. However, the privacy-preserving property of FL only holds if the updates cannot be used to extract sensitive information about the local training data.", + "bbox": [ + 75, + 657, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite only sending updates, prior works have shown the ability of attackers to gain information about the private training data through membership inference attacks [3,", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3bea72727bea4569b30fade7e1446724c05af839d2023a8d61b20a1dc5f5d26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 290, + 867, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4837e62ab0979054111d2495c3728b81b524954bf875568d00e077bde60f6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 377, + 867, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2e2fe93a86039f39c7cc7966222d872c01e7c854198f4127cd54630fdaf9b1dc.jpg", + "image_caption": [ + "Figure 1. Training using leaked data." + ], + "image_footnote": [], + "bbox": [ + 521, + 468, + 864, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "19, 22], property-inference attacks [15, 18], or GAN-based methods [9, 23]. Within these privacy attacks, data reconstruction has stood out as the most powerful, allowing attackers to directly recover the training data of clients [20, 24]. Within data reconstruction attacks, gradient inversion (optimization-based attacks) [6, 26, 32] and linear layer leakage [1, 5, 30] attacks stand out as the most common.", + "bbox": [ + 498, + 597, + 890, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While prior works have discussed the power and the limitations of these attacks in the context of an attacker simply breaching privacy [10, 29], to the best of our knowledge, none has discussed the effectiveness of using the leaked data for downstream model training. Figure 1 shows this process of training using leaked data. With the onset of deep learning, data has become a valuable commodity. This growth in AI also means that the value of data no longer lies just in breaching privacy of the raw data, but in the ability of the data to be used downstream for creating powerful models. Therefore, one measure of the success of a data reconstruction attack should be whether the leaked data is useful for machine learning training, an aspect omitted in prior work.", + "bbox": [ + 496, + 704, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In order to discuss training on leaked data, we first introduce the attacks. Gradient inversion (GI) attacks typically reconstruct data through an iterative process where the distance between a gradient computed from a dummy image and the ground truth gradient is minimized by an optimizer. Label reconstruction, the process of recovering the batch labels from the client gradient, also plays a critical role in affecting the final reconstruction quality. These labels are typically leaked prior to optimization from the gradients of the final layer of the network [7, 16, 26, 28] and are matched with reconstructed images during the optimization. These attacks have shown success with individual updates of smaller batch sizes, up to 48 on ImageNet [26] or 100 on CIFAR [6]. However, this only considers breaching privacy, as a batch size of 48 on ImageNet or 100 on CIFAR result in low quality reconstructions and/or a very small number of identifiable images. Even for a modest batch size of 8 or 16 on CIFAR-10, some batch reconstructions already fail. Given the importance of data quality for training models, a sufficient condition for success is no longer if any images are identifiable, but rather the overall usefulness of the leaked data for the downstream training task. Here, even reconstructions with low similarity scores to the ground truth may still positively contribute to training.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to GI, linear layer leakage (LLL) does not suffer from problems with reconstruction quality. Using modification of a fully-connected (linear, FC) layer, an attacker can directly recover a proportion of the images of a client batch [1, 5, 30]. Reconstructed images are done by solving a simple linear equation and have little computation overhead. However, these attacks typically require modification of the model architecture and can add a large overhead in terms of model size [29]. While data quality is not an issue, when using the leaked data for training models, labeling the data becomes a problem. Unlike the optimization attacks, even if labels are known prior to reconstruction, the LLL process does not match them to the leaked images. As a result, after reconstruction there would be a set of leaked images and labels, where the images are not matched with their corresponding label. Furthermore, because only a proportion of images in the given batch are leaked for LLL, even with a full set of labels, some labels will not have a leaked image to go with. This can lead to a tedious process of manually labeling leaked images.", + "bbox": [ + 76, + 457, + 472, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Viewing data reconstruction attacks through the lens of training models brings up many questions. How do models trained with leaked images compare to centralized or federated learning? Does the reconstruction quality impact how well models perform when trained on leaked data? How does the lack of label matching during reconstruction affect linear layer leakage? In this work we address these largely unanswered questions and highlight the problems that arise in this new setting. Our contributions go as follows:", + "bbox": [ + 76, + 763, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate the effectiveness of training models using leaked data compared to federated learning and the centralized baselines. Using the data from linear layer leakage and optimization, models trained on leaked data can achieve $20.40\\%$ and $17.58\\%$ higher accuracy on CIFAR-10 compared to federated learning.", + "- We show that similarity metrics such as PSNR do not show if reconstructed images are useful for training, as even some of the worst images improve models. Training on CIFAR-10 with bad reconstructions generated by Inverting Gradients (batch size 16) with PSNR $< 14$ results in a $58.29\\%$ test accuracy. Removing these images and training only on images $>14$ PSNR results in only a minor accuracy decrease from $76.83\\%$ to $75.48\\%$ .", + "- We quantify the effects and highlight the crucial limitations of current data reconstruction attacks when viewed through the lens of downstream model training. While gradient inversion can breach privacy for a batch size of 100 on CIFAR-10, this is impractical for training as even increasing the batch size from 4 to 16 results in a performance drop from $90.34\\%$ to $76.83\\%$ ." + ], + "bbox": [ + 500, + 90, + 893, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 416, + 638, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gradient inversion (GI) attacks operate under an honest-but-curious threat model where a server only knows the update and model. The method involves feeding a dummy image into a model and computing the subsequent gradient. The dummy image starts out initialized as random noise, and an optimizer iteratively minimizes the distance between the computed gradient and the ground truth client gradient.", + "bbox": [ + 496, + 435, + 890, + 541 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {*} = \\arg \\min _ {x} \\left\\| \\nabla L (x, y, \\theta) - \\nabla W \\right\\| _ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 547, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The intuition behind these attacks is that a similar computed dummy gradient $\\nabla L(x,y,\\theta)$ and ground truth gradient $\\nabla W$ will result from similar images. Thus, as the gradients become closer, the final reconstructed image $x^{*}$ should become close to the ground truth training image. The label $y$ is assumed to be known prior to optimization. For a batch reconstruction with multiple labels, the labels are matched with the reconstructed images during optimization (as the loss and gradient are computed with these image-label pairs). Regularizers have also been introduced to assist the optimization process [6, 26]. However, these can result in reconstructed image artifacts.", + "bbox": [ + 496, + 571, + 890, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Linear layer leakage (LLL) attacks are built on the observation that a fully-connected (FC) layer leaks the input to the layer through the gradients [4, 21].", + "bbox": [ + 496, + 753, + 893, + 800 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {i} = \\frac {\\delta L}{\\delta W ^ {i}} / \\frac {\\delta L}{\\delta B ^ {i}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 804, + 890, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, $\\frac{\\delta L}{\\delta W^i}$ is the weight gradient and $\\frac{\\delta L}{\\delta B^i}$ is the bias gradient of a neuron. Neuron $i$ is activated by an image and $x^i$ is the respective reconstructed image. However, this assumes that only a single image activated the neuron. If multiple", + "bbox": [ + 496, + 838, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12248", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images activate the same neuron, reconstructed image $x^{i}$ becomes a combination of the images.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate this, the methods of trap weights [1] and Robbing the Fed [5] were proposed. Trap weights initialize the weights of the FC layer randomly as half positive and half negative, where the negative parameters are sampled from a slightly larger magnitude. This method has difficulty with scaling [30] compared to Robbing the Fed, which proposed a method where the FC layer weights are used to measure a property of the images (e.g., average pixel brightness). The biases of the layer were then set to fit the distribution of the dataset and a ReLU activation was used to threshold the activation. Images were reconstructed as", + "bbox": [ + 75, + 121, + 467, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {i} = \\left(\\frac {\\delta L}{\\delta W ^ {i}} - \\frac {\\delta L}{\\delta W ^ {i + 1}}\\right) / \\left(\\frac {\\delta L}{\\delta B ^ {i}} - \\frac {\\delta L}{\\delta B ^ {i + 1}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 291, + 467, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $i + 1$ indicates the neuron with a bias being the next highest cutoff. This method requires basic knowledge about the input data distribution, but achieves a higher leakage rate compared to trap weights and is scalable to aggregation.", + "bbox": [ + 75, + 325, + 467, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the case of FedAvg, the sparse variant of Robbing the Fed utilizes a two-sided activation function (such as Hard tanh) and weight/bias scaling to maintain the same leakage as in FedSGD. However, the sparse variant is not scalable and leads to precision problems when attacking secure aggregation [2] or larger batches. The LOKI attack addresses the scalability problems of both methods through the introduction of a convolutional layer that splits the scaling between number of clients and batch size in secure aggregation [30]. They further introduce a convolutional scaling factor (CSF) which achieves a higher leakage rate in FedAvg without increasing the FC layer size.", + "bbox": [ + 75, + 386, + 467, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While there have been many works exploring data reconstruction, prior work has not evaluated these attacks in the context of downstream tasks. The works on GI [6, 26, 32] and LLL [1, 5, 30] attacks only discuss the reconstruction quality of the methods in terms of standard image metrics such as PSNR (peak signal-to-noise ratio), SSIM (structural similarity index measure), or LPIPS (learned perceptual image patch similarity). However, these metrics only measure how similar the reconstruction are to the ground truth and only give a vague sense of how useful the data is for training. Another work discussed the limitations of GI attacks [10] measured by image similarity. In [29], LLL attacks were measured in terms of the resource overhead added by the attacks. However, these works also did not discuss the usefulness of reconstructed data post-leakage.", + "bbox": [ + 75, + 568, + 467, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to bridge the gap and explore the important facet of training machine learning models on leaked data from reconstruction attacks. Viewing leaked data beyond privacy breach brings to light key weaknesses of current attacks in generating useful data. While LLL reconstructs high quality images, the other important half of data, namely having matching labels, is missing. GI attacks can", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "breach privacy for larger batch sizes, but the quality of the leaked data would be nearly useless for training models. In the age of machine learning and artificial intelligence, the value of data has increased tremendously. Therefore, when looking at leaked data, it is important to measure success in terms of usefulness to model training.", + "bbox": [ + 498, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Training on leaked data", + "text_level": 1, + "bbox": [ + 498, + 188, + 722, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data reconstruction attacks have largely focused on the quality of the recovered image in gradient inversion (GI) or the efficiency of linear layer leakage (LLL). However, label reconstruction is important for training on leaked data. Within the pipeline of GI attacks, labels are typically recovered prior to reconstruction, as they assist in the optimization problem [10]. Recent works have shown that label restoration with large batches and duplicate labels is possible [7, 16]. More accurate labels and the added flexibility of duplicate class images directly improves prior GI attacks such as Inverting Gradients [6] and GradInversion [26].", + "bbox": [ + 496, + 205, + 890, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Interestingly, one strength of LLL has been in the ability to reconstruct images without knowledge of the data labels. While the class labels can affect the distribution of the client dataset (e.g., average image brightness can be different between classes) [30], the labels themselves do not help in the reconstruction. Knowing labels helps guide GI attacks (as gradients are computed from data and label pairs), but for LLL, images activating the neurons are reconstructed using only the weight and bias gradients of the FC layer as with Equation 3. Labels are necessary to train models using leaked data, but even with a set of recovered labels, the leaked images would still remain unmatched. Unlike GI which matches during optimization, images leaked by LLL would need to be manually matched. For this work, we experiment with cases when all labels are known and matched and alternatively, when only a small portion of labels are known and matched. Nonetheless, the label leakage and matching issue stands as an open problem for LLL attacks.", + "bbox": [ + 496, + 372, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identifying which images are reconstructed correctly is also a challenging problem for LLL. For FedSGD, bins without activation can be identified, as after the subtraction process the reconstruction will be zero. However, if multiple images activate the same bin, the reconstruction is non-zero. By checking for non-zero values, we cannot directly identify whether the reconstruction is a single image or a combination of multiple. Manually filtering this can become a tedious process in practice. This is somewhat mitigated in the FedAvg case by LOKI. Across separate FedAvg local mini-batches, multiple images will not activate the same neuron, resulting in muchFewew overlapping reconstructions. Therefore the only images that can cause multiple activations are within the same mini-batch [30]. In this work, we manually filter out multiple image reconstructions in order to test the quality of the data. However, especially for LLL in FedSGD, the problem of how to automate", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12249", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/434c327510eddc12822a5fa5b8bd324c52e96d1482f5ee27d7c7c68f737cbf6d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MNIST
Batch sizeModelTime per batch (s)Time for entire dataset (days)
8DNN54.794.76
1662.342.71
3264.321.40
CIFAR-10
Batch sizeModelTime per batch (s)Time for entire dataset (days)
4ResNet-18422.8161.17
8437.5431.65
16435.3315.75
", + "bbox": [ + 130, + 88, + 416, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "removal or how to use these multiple image reconstructions for training still remains an important item of future work.", + "bbox": [ + 75, + 285, + 468, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 319, + 209, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of our work is to discuss the effectiveness of leaked data when used for the downstream task of image classification. We use centralized training on the entire benign dataset and federated learning (with only updates and no data shared by the clients) as the baselines for comparison. Typically, these serve as the upper bound and the lower bound respectively for the model accuracy. For federated learning, we use two settings of 10 and 50 clients. We use a non-IID bias of 0.5 and FedAvg with 3 local iterations. Additional training results for the IID setting and FedSGD are included in the supplementary.", + "bbox": [ + 75, + 338, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use Inverting Gradients [6] as the gradient inversion (GI) attack. Another candidate is GradInversion, which is a SOTA GI attack (especially for ImageNet-sized images). However, we discarded this as the author code for this attack is not available and the third-party online implementations did not achieve comparable results to the original paper. For linear layer leakage (LLL), we use the LOKI attack [30]. We use LOKI as it achieves SOTA FedAvg leakage rate and equivalent SOTA leakage rate in FedSGD compared to Robbing the Fed. Model performance for other LLL methods [1, 5] can be found in the supplementary.", + "bbox": [ + 75, + 503, + 468, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use MNIST [13], CIFAR-10 [11], and the Tiny ImageNet [12] datasets. For the MNIST dataset, we train using a simple DNN with 2 convolutional layers and 2 FC layers. For CIFAR-10 and Tiny Imagenet, we use a ResNet-18 [8]. We evaluate on the validation set for Tiny ImageNet because the test set is unlabeled. The plots only show a single run for each setting, but the (numerical) testing performance is reported as an average over 5 runs. We run all experiments on NVIDIA A100 80GB GPUs.", + "bbox": [ + 75, + 670, + 468, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Gradient inversion computation time", + "text_level": 1, + "bbox": [ + 76, + 809, + 397, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to LLL attacks, the computational overhead of GI attacks when reconstructing data, while done offline, is still non-negligible. In order to fully evaluate the leaked data, we ran the Inverting Gradients attack on the entire dataset of MNIST with batch sizes of 8, 16, and 32. We", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/df84ec2c4b1f6ed3f71b29b828d27c248bd6340b02c88e24a9fdb0e90c73aaea.jpg", + "table_caption": [ + "Table 1. Computation time for Inverting Gradients running on a NVIDIA A100 80GB GPU. Small batches have more total batches and take more time for the whole dataset. Computation time is significantly higher for CIFAR-10 given the more complex model." + ], + "table_footnote": [], + "table_body": "
CIFAR-10MNIST
Batch sizePSNRSSIMBatch sizePSNRSSIM
427.880.9067823.160.6996
820.770.72151619.150.6506
1615.940.39783215.780.4537
", + "bbox": [ + 529, + 88, + 864, + 143 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2eedcc4b4e7b2f4f8c223f6ec355c0b36aab3d269ebcd33088e21c5aa8d02082.jpg", + "table_caption": [ + "(a) Gradient Inversion PSNR/SSIM" + ], + "table_footnote": [], + "table_body": "
CIFAR-10MNISTTiny ImageNet
FC factorLeaked imagesPercentLeaked imagesPercentLeaked imagesPercent
84378887.585254887.588580485.80
43946478.934596676.617514975.15
22988259.763579559.665801258.01
11824236.482181536.363539335.39
", + "bbox": [ + 511, + 162, + 880, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Linear layer attack leakage rate", + "bbox": [ + 602, + 242, + 789, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. (a) Average PSNR↑/SSIM↑ scores of gradient inversion (Inverting Gradients) reconstructions for various batch sizes and (b) number of leaked images/leakage rate of linear layer leakage attack (LOKI) for several FC layer sizes.", + "bbox": [ + 498, + 257, + 890, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "also ran over CIFAR-10 with batch sizes of 4, 8, and 16. We run 10,000 iterations of optimization for each batch. We evaluate the Inverting Gradients attack in the best case scenario with an untrained model where all labels are correctly recovered prior to optimization.", + "bbox": [ + 496, + 319, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows the amount of time taken by Inverting Gradients run on a NVIDIA A100 80GB GPU for MNIST and CIFAR-10 on a DNN and ResNet-18 respectively. There is some variance in the average amount of time based on batch size, but when attacking the entire dataset, smaller batch sizes always take longer time as the number of total batches is much larger. CIFAR-10 takes significantly longer to reconstruct each batch given the more complex model of ResNet-18 compared to a DNN. For a batch size of 4, CIFAR-10 takes 422.81 seconds per batch. Iterating across the entire dataset on a single GPU takes 61.17 days.", + "bbox": [ + 496, + 395, + 892, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Leakage statistics", + "text_level": 1, + "bbox": [ + 500, + 566, + 671, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For GI we use Inverting Gradients [6], and for LLL we use LOKI [30]. For Inverting Grads, we use a batch size of 4, 8, and 16 on CIFAR-10 and 8, 16 and 32 on MNIST. For LOKI, we choose the FC layer size based on the batch size. We vary this with factors of 1, 2, 4 and 8 on CIFAR-10, MNIST, and Tiny ImageNet. We use a client batch size of 64 for all datasets, so an FC size factor of 2 would mean $64 \\cdot 2 = 128$ units in the FC layer.", + "bbox": [ + 496, + 582, + 890, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2a gives the average PSNR and SSIM metrics (where a higher score indicates greater similarity with the ground truth) for the reconstructions across the entire dataset of CIFAR-10 and MNIST from Inverting Gradients. Increasing the batch size results in a lower PSNR and SSIM score for both CIFAR-10 and MNIST.", + "bbox": [ + 496, + 703, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2b reports the number of leaked images for LOKI. For an FC factor of 4, LOKI leaks $78.93\\%$ , $76.61\\%$ , and $75.15\\%$ of images on the CIFAR-10, MNIST, and Tiny ImageNet datasets. Increasing the FC layer size increases the model size overhead, but results in a higher leakage rate as images are more likely to activate different neurons. Decreasing the size creates a smaller model size overhead but", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12250", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "results in lower leakage rate.", + "bbox": [ + 76, + 90, + 269, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Training on leaked data from scratch", + "text_level": 1, + "bbox": [ + 76, + 111, + 397, + 126 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start with a discussion on the models that can be trained from the leaked data of both GI and LLL attacks. The attacks are applied under the FedSGD setting. We also start with assuming that all labels are known prior to training — this is a best case assumption for the downstream training. For CIFAR-10, we train the centralized models using SGD with momentum 0.9 and weight decay 0.0001. We use a batch size of 128 and train for 200 epochs starting with an initial learning rate of 0.1 and a scheduler that decreases on plateau. For MNIST, we use SGD without momentum or weight decay and train for 40 epochs with a LR of 0.01. For Tiny ImageNet we use Adam with learning rate 0.001 and train for 50 epochs with a batch size of 1024. For federated learning, we use 400 training rounds for CIFAR-10 and MNIST and 500 rounds for Tiny ImageNet. Client batch size of 32, 128, and 256 are used for MNIST, CIFAR-10, and Tiny ImageNet. The test accuracy is downsampled to fit on the same axis with centralized training.", + "bbox": [ + 76, + 128, + 468, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3a and Figure 2a show the results for CIFAR-10 for centralized training, federated learning (FedAvg), and training with leaked data. Centralized training achieves a $94.38\\%$ accuracy and federated learning achieves a $72.76\\%$ and $68.71\\%$ peak accuracy with 10 and 50 clients respectively. Centralized and FL indeed provide the upper bound and the lower bound of the accuracy here. Across all attack settings, GI and LLL achieve higher model accuracy compared to federated learning. The reconstruction quality of GI is adversely affected by larger batch sizes and in turn also impacts the final model performance. With a batch size of 4, 8, and 16, models trained on the leaked data from GI achieves $90.34\\%$ , $86.16\\%$ , and $76.83\\%$ accuracy. With even larger batch sizes, it is likely that the model performance will drop below federated learning. The performance of models using LLL data is more stable. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the accuracy only drops from $93.16\\%$ to $88.86\\%$ . This is spite of a large drop in the number of leaked images from 43788 ( $87.58\\%$ ) to 18242 ( $36.48\\%$ ) as shown in Table 2b. This is a subtle and consequential result — the downstream training task can still be accomplished despite a big drop in the amount of data leakage. All prior work had stopped at the stage of evaluating the proportion of leakage and therefore had missed out on this insight.", + "bbox": [ + 76, + 401, + 468, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3b and Figure 2b show the results for MNIST. LLL performs better than the FedAvg baseline in all cases. With an FC size factor of 8, LOKI achieves a $98.82\\%$ test accuracy (nearly the same regardless of the FC size factor), only $0.07\\%$ lower than centralized at $98.89\\%$ . GI performs slightly worse than federated learning, achieving $95.96\\%$ , $95.50\\%$ , and $94.29\\%$ at batch sizes 8, 16, and 32 compared to $96.17\\%$ and $96.18\\%$ with 10 and 50 clients in FedAvg.", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We believe this is because the noise in the reconstructions hampers model performance with GI. This affects MNIST more severely than for CIFAR because the MNIST images are more sparse and sensitive to noise.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The top-1 validation accuracy for Tiny ImageNet are given in Table 3c and Figure 2c. The total number of leaked images hurts LLL more here than in the other two datasets. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the validation accuracy drops from $46.70\\%$ to $35.20\\%$ , an $11.5\\%$ drop in performance. Federated learning achieves $37.00\\%$ and $35.06\\%$ accuracy with 10 and 50 clients respectively. FL with 10 clients has slightly better accuracy than LLL with an FC size factor of 1. However, any LLL setting with a larger FC layer achieves higher accuracy than both settings in federated learning.", + "bbox": [ + 496, + 152, + 892, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. FedAvg with leaked images from LOKI", + "text_level": 1, + "bbox": [ + 498, + 320, + 836, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We explore the impact of leaking images using LLL in FedAvg when training models. We use 8 local iterations of mini-batch size 8, a local dataset size of 64, and a learning rate of 1e-4. Following the LOKI attack [30], $\\mathrm{CSF} = 500$ is used to achieve a higher leakage rate and reconstruction quality. Similar to before, we apply the attack across batches covering the entire CIFAR-10 training dataset.", + "bbox": [ + 496, + 337, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 4 shows the leakage rate of LOKI in FedAvg compared to FedSGD. Using $CSF = 500$ , the leakage rate of LOKI in FedAvg is substantially higher than the leakage rates in the FedSGD setting. Using an FC layer of half the size, the leakage rate in FedAvg is comparable to that of FedSGD. For example, LOKI FedAvg factor 1 achieves a leakage rate of $58.39\\%$ compared to LOKI FedSGD factor 2 which achieves a leakage rate of $59.76\\%$ .", + "bbox": [ + 496, + 443, + 892, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 5 shows the testing accuracy of LOKI FedAvg for the different FC layer sizes. The test accuracies are very comparable to the FedSGD settings. However, with a smaller FC layer size factor, the effect of having a larger total number of leaked images becomes visible. With FC factor 1, LOKI FedAvg achieves $91.11\\%$ accuracy while LOKI FedSGD achieves $88.86\\%$ .", + "bbox": [ + 496, + 564, + 890, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Training with semi-supervised learning", + "text_level": 1, + "bbox": [ + 498, + 672, + 836, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to train machine learning models in a supervised fashion, images leaked through LLL need to be labeled. Previously, we assumed that all labels were known and matched to the images prior to training, but this is impractical. Within a LLL batch reconstruction, the order of images is not preserved. Even if labels were recovered like for GI attacks, this would result in a set of unmatched ground truth labels and images. Manual hand labeling is tedious and this inability to match labels is a weakness of LLL attacks only when it comes to downstream tasks.", + "bbox": [ + 496, + 689, + 890, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, we assume that only a small portion of images are labeled. This setting is similar to semi-supervised learning (SSL) [25, 31] and we use CoMatch [14], a semi-supervised algorithm, to train the mod", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/18955b8583f2939a32deb4eedb4008f0527f0d3b4d995ac7826ec7606e01c2d0.jpg", + "image_caption": [ + "(a) CIFAR-10", + "Table 3. (a) CIFAR-10 and (b) MNIST test accuracy, (c) Tiny ImageNet top-1 validation accuracy trained from scratch with leaked data. CIFAR-10 and Tiny ImageNet are trained with a ResNet-18 and MNIST is trained with a DNN. For the two attacks (gradient inversion and linear layer leakage), training happens with leaked data. Second column indicates the number of clients in federated learning (FedAvg), the batch size for the gradient inversion attack, and the fully-connected layer size factor for linear layer leakage. Best accuracy is used for FedAvg and final accuracy for all other settings." + ], + "image_footnote": [], + "bbox": [ + 83, + 90, + 344, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b9c9ceb9aecf6de5784cf9f95866c9b17ecbb84294e6d4e903c40347399fe9cd.jpg", + "image_caption": [ + "(b) MNIST", + "Figure 2. Training models on (a) CIFAR-10, (b) MNIST, and (c) Tiny ImageNet with leaked data compared to centralized and federated learning training. Both linear layer leakage and gradient inversion achieve higher accuracy than the federated learning (FedAvg) baseline for CIFAR-10 in all cases. For MNIST, LLL (LOKI) nearly reaches centralized accuracy while GI performs slightly worse than FL. Top-1 validation accuracy used when training models on Tiny ImageNet. Here, LLL performs better with a FC layer size factor of 2 or higher." + ], + "image_footnote": [], + "bbox": [ + 354, + 90, + 612, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f73af5f6922e58ce7f7088b8c495d2a771907580156a0131deb309d8d33c90d1.jpg", + "image_caption": [ + "(c) Tiny ImageNet" + ], + "image_footnote": [], + "bbox": [ + 625, + 90, + 885, + 222 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/98f10f62088a826fbe7c4f02f84413b24574b64dd8c32cd0a78d982710907761.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-94.38
FedAvg1072.76
5068.71
Gradient inversion (Inverting grads [6])490.34
886.16
1676.83
893.16
Linear layer leakage (LOKI [30])492.94
291.90
188.86
", + "bbox": [ + 83, + 306, + 344, + 443 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e5676606207bc897a7eae3fb4dc140f48e4e5d4152044c53af12f04ec1ec28ad.jpg", + "table_caption": [ + "(a) CIFAR-10" + ], + "table_footnote": [], + "table_body": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-98.89
FedAvg1096.17
5096.18
Gradient inversion (Inverting grads [6])895.96
1695.50
3294.29
898.82
Linear layer leakage (LOKI [30])498.70
298.72
198.46
", + "bbox": [ + 354, + 306, + 616, + 443 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/48c4f9c896068650377fb65a05876315147f82c67b146795f4d1c2d783c5565b.jpg", + "table_caption": [ + "(b) MNIST" + ], + "table_footnote": [], + "table_body": "
Num. clients / FC factorValidation accuracy
Centralized-48.55
FedAvg1037.00
5035.06
Linear layer leakage (LOKI [30])846.70
445.09
241.15
135.20
", + "bbox": [ + 622, + 325, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab5cbcfdf3a15ac69ed120d590e7b488b560e1efa3521dfc676df42bba8208ab.jpg", + "table_caption": [ + "(c) Tiny ImageNet" + ], + "table_footnote": [], + "table_body": "
FC size factorLOKI FedSGDLOKI FedAvg
887.58 (43788)92.19 (46097)
478.93 (39464)85.98 (42989)
259.76 (29882)75.29 (37645)
136.48 (18242)58.39 (29196)
", + "bbox": [ + 160, + 537, + 387, + 601 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/31d3e565cdc0a4a62a8098b7373f195d5e07984e95977dc833603b9337c09a4a.jpg", + "table_caption": [ + "Table 4. Leakage rate % (leaked images) of LOKI in FedSGD and FedAvg on CIFAR-10 for several FC layer sizes. FedAvg with half the FC layer size has comparable leakage rate to FedSGD." + ], + "table_footnote": [], + "table_body": "
FC factorFedSGD accuracyFedAvg accuracy
Linear layer leakage (LOKI [30])893.1693.31
492.9492.88
291.9092.35
188.8691.11
", + "bbox": [ + 124, + 654, + 424, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. CIFAR-10 test accuracy with LOKI FedAvg.", + "bbox": [ + 109, + 723, + 433, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "els. We use a WideResNet-28-2 [27] and train for 450 epochs on CIFAR-10. Figure 3a fixes the number of known labels to be 40 and shows the test accuracy curve with different fully connected layer sizes. We note that when there is only a small proportion of leaked data, this leads to a pronounced decrease in performance in SSL. LOKI with an FC factor of 1 (with $36.48\\%$ leakage rate) illustrates this, as it only achieves $78.31\\%$ max accuracy compared to the CIFAR-10 40 label baseline of $92.18\\%$ . With a factor of 2, the accuracy improves and jumps to $88.91\\%$ as the leak", + "bbox": [ + 75, + 750, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age rate also substantially increases to $59.76\\%$ . With factors 4 and 8 (78.93% and 87.58% leakage rate), the accuracy is very close to the baseline at 91.18% and 91.84% respectively. Figure 3b shows the test accuracy plot of LOKI size factor 4 with the number of known labels being 20, 40, and 250. With 20 and 250 labels, LOKI factor 4 achieves 86.69% and 91.65% peak accuracy. We note that there is a greater variability in accuracy with a smaller number of labels, a trend consistent with other SSL methods [14, 25, 31].", + "bbox": [ + 496, + 539, + 890, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.6. Starting training from federated models", + "text_level": 1, + "bbox": [ + 498, + 685, + 843, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the previous experiments worked with leaked data from all batches in the dataset, such large amounts of leaked data may not always be available. For example, LLL attacks may affect model performance due to manipulation of the model parameters and/or architecture. In later rounds when model performance is high, this may be easier to detect. Similarly, GI attacks such as Inverting Gradients [6] have better reconstruction quality with an untrained network. In both cases, it is possible that the attack is only applied during a few initial rounds of training and only a small portion of data is leaked. Training from scratch with a small portion of leaked data does not achieve good model performance alone, so instead, we can initialize from the fully trained FL", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96a88cd287b7aa7e619fa2b9b30c5f1578bac4dc4720164d481efb25845c8989.jpg", + "image_caption": [ + "(a) Varying leakage rate" + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 415, + 231 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a19abd6be1601b4fe053ebee474da233da9f51943d5494b2499860b9f36c49f.jpg", + "image_caption": [ + "(b) Varying number of labels" + ], + "image_footnote": [], + "bbox": [ + 130, + 255, + 411, + 396 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d48916e9d49d4a26cd1b8c08844cf60c755bd2b6a184520b246e91eda313b904.jpg", + "image_caption": [ + "Figure 3. Semi-supervised learning using CoMatch on CIFAR-10 with a WideResNet with (a) a varying FC size and leakage rate (LR) for LOKI and the known labels fixed at 40 and (b) a fixed number of leaked images and 20, 40, and 250 known labels.", + "Figure 4. FL model trained with 50 clients as start point for training with leaked LLL data. Highlighted area indicates improvement above the FL model and training on the leaked data alone." + ], + "image_footnote": [], + "bbox": [ + 133, + 484, + 415, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model and then train it centrally using the leaked data.", + "bbox": [ + 76, + 686, + 434, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use smaller portions of the leaked data from LOKI FC size factor of 8 for this experiment. With only $5\\%$ (2500) of leaked images in CIFAR-10, a model trained from scratch can only achieve $61.15\\%$ test accuracy. With $2\\%$ (1000 images) of data, the model performance drops to $39.63\\%$ . Both cases are below an FL model with 50 clients that achieves a $64.08\\%$ final test accuracy. However, by using the FL model as the initial starting point, even a small amount of leaked data allows for an increase in performance. We take the model that has been trained using FL and then additionally train that model centrally. Using only $2\\%$ and $5\\%$ of the leaked data, the models are able to achieve $68.25\\%$ and $74.25\\%$ final testing ac", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebc175d5d9b7228c9a88d58b34e345a73275e717ec964b4421324e0e6d22a0e3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR% ims keptTest accuracy
>2015.8870.04
>1827.5971.52
>1640.7873.73
>1460.2275.48
>1284.4276.16
", + "bbox": [ + 514, + 88, + 679, + 162 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4efa67a88f32285a2d9bc47d4f7e064bcaa736d93b4282a5385eaad980ec3bcf.jpg", + "table_caption": [ + "(a) PSNR above threshold" + ], + "table_footnote": [], + "table_body": "
PSNR% ims keptTest accuracy
< 2084.1273.03
< 1872.4167.09
< 1659.2263.71
< 1439.7858.29
< 1215.5845.05
", + "bbox": [ + 707, + 88, + 874, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) PSNR below threshold", + "bbox": [ + 720, + 172, + 861, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Training models using CIFAR-10 leaked data from inverting gradients batch size 16. Only the reconstructions with a PSNR (a) above and (b) below the threshold are used in training.", + "bbox": [ + 500, + 186, + 890, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "curacy, a $28.62\\%$ and $13.10\\%$ improvement from starting from scratch. Figure 4 shows the improvement in test accuracy achieved by using the trained FL model as initialization instead of starting from scratch. The improvement is especially noticeable with small percentages of leaked data.", + "bbox": [ + 496, + 239, + 890, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Quality of data reconstruction", + "text_level": 1, + "bbox": [ + 500, + 319, + 767, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quality of reconstruction is an important concern for GI attacks. As the batch size continues to increase, the reconstruction quality decreases. Table 2a also shows this relationship. From the previous experiments, we also see a negative relationship between the increasing batch size and the usefulness of the leaked data in downstream model training.", + "bbox": [ + 496, + 335, + 890, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we explore whether poor quality reconstructions are still useful for training models. Using the CIFAR-10 leaked dataset created using Inverting Gradients on a batch size of 16, we first sort the reconstructions by their PSNR value. Our first set of experiments removes the worst images from training. Table 6a shows the percent of images above the PSNR threshold and the final test accuracy when trained on that data. Compared to the baseline accuracy of $76.83\\%$ achieved by using the entire leaked dataset, even removing the worst images with a $\\mathrm{PSNR} < 12$ from the training set results in a small but non-zero performance loss. Then we take a different look at this question by now training on the worst images. Table 6b shows the test accuracy of models trained on all images below a PSNR threshold. Even when training on the lowest quality images with a PSNR $< 12$ , the model achieves a $45.05\\%$ final testing accuracy, significantly worse than the baseline, but much above zero.", + "bbox": [ + 496, + 426, + 892, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the previous results, we see that even leaked data with a low reconstruction quality is still useful to the model training process. This message is a new insight — all prior works when calculating the leak rate do not count poorly reconstructed images. We additionally include results using SSIM in the supplementary section. Visual examples of poor quality reconstructions used to train the models are also included in the supplement.", + "bbox": [ + 496, + 684, + 890, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.8. Observations on reconstruction quality trends We additionally note a few trends manually observed when going through the GI batch reconstructions of CIFAR-10.", + "text_level": 1, + "bbox": [ + 498, + 808, + 888, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While some reconstructions have very low quality, only images of the same label will be swapped within the batch during label matching. This is a desirable property for train", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12253", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/037c55b704c969df021352cee8be54e08a1c7271bf868a3b7ca118b575e4e55a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 90, + 444, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cdd38955952bead9e00475752c68023aec498d6130665d635920cb69892334be.jpg", + "image_caption": [ + "Figure 5. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Ground truth images are on top and reconstructions are on the bottom. All labels in the batch are different and reconstructed images are high quality.", + "Figure 6. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Duplicate labels exist between images 0, 4, 7 and images 2, 3. Reconstruction 3 contains parts of the content from both images 2 and 3. Reconstructions 0, 4, and 7 appear almost as noise. However, the other images still have reasonable quality." + ], + "image_footnote": [], + "bbox": [ + 102, + 233, + 444, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing on leaked data. As long as the set of labels is leaked correctly, the optimization process will match images to a correct label. So even though the image quality is poor (and can be entirely unrecognizable), since the label is correctly matched, such an image still does help in the training (this buttresses the message from Tables 6a and 6b). Naturally, as the PSNR goes down, the image helps less and less.", + "bbox": [ + 75, + 392, + 467, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also note that when all labels are different within a batch, the reconstruction quality of the entire batch of images is almost always good. An example on CIFAR-10 is shown in Figure 5. Usually when the reconstructions are poor, it occurs between groups of images in the same batch that belong to the same class. Poor reconstructions can appear as almost entirely noise, but occasionally images can have visible pieces of each image from the same class in the batch stitched together. The example batch in Figure 6 depicts these properties of poorly reconstructed images. As a side note, we observe that a batch may have a mix of well reconstructed and poorly reconstructed images. We also note that the reverse mapping of the same label images to poor reconstructions is not true — having duplicate labels does not always imply that the reconstructions will be bad.", + "bbox": [ + 75, + 500, + 467, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 731, + 191, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a main goal of our work, we have explored the effect of leaked data through today's leading data reconstruction attacks on downstream model training. Our main message is that it is important to also consider the use of data beyond reconstruction quality and image similarity. It is important to consider how far these leaked samples help in a downstream training task. From the experiments, we have seen that current reconstruction attacks are powerful, but still lack in several critical areas. For gradient inversion attacks, the reconstruction quality does pose an issue for", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training models. Another impediment to practical use of this attack type is the time and computational cost.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Linear layer leakage attacks do not suffer from reconstruction quality issues and are much lighter weight computationally, but still suffer from other challenges. While LOKI has increased the efficiency of the FC layer leakage in FedAvg, there is still room for improvement (especially in FedSGD). As shown in [29], the model size overhead added by these attacks can be very large, especially if secure aggregation is used. Furthermore, the lack of labels poses a large challenge in training models on the data. With a small dataset like CIFAR-10, hand labeling 40 images is not too challenging. However, larger datasets will require more total labels. For example, SimMatch with $1\\%$ of labels on ImageNet can achieve $67.2\\%$ top-1 accuracy [31]. However, even $1\\%$ of the total images in ImageNet-1k is 12,812 images. Even if the images were leaked, hand labeling would be extremely hard. Furthermore, we see from Figure 3a that having less total data and fewer labels results in an even steeper decrease in performance.", + "bbox": [ + 496, + 121, + 890, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While not discussed previously, there are many other aspects that affect data reconstruction attacks when considering training models. For example, how does the non-IID aspect of client data affect reconstruction? Other relevant settings of federated learning such as asynchronous federated learning, client selection, and differential privacy can also affect the final leaked dataset. Centralized training also allows for more complex and larger models without the communication or computational restrictions of clients in federated learning. These topics can be explored in future work.", + "bbox": [ + 496, + 393, + 890, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 550, + 617, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We examine data reconstruction attacks through the lens of training models on the leaked data. While highlighting how the weaknesses of gradient inversion in terms of reconstruction quality affect downstream model training, we also show that even poorly reconstructed images are useful for training. We also discuss how the label matching problem for linear layer leakage can be mitigated through semisupervised learning. Under ideal conditions, we demonstrate that leaked data from both gradient inversion and linear layer leakage attacks are able to train powerful models comparable to even a centralized baseline. On CIFAR-10, gradient inversion and linear layer leakage attacks achieve $90.34\\%$ and $93.16\\%$ testing accuracy respectively, $17.58\\%$ and $20.40\\%$ higher than federated learning and only $4.04\\%$ and $1.22\\%$ lower compared to centralized training.", + "bbox": [ + 496, + 568, + 890, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This material is based in part upon work supported by Adobe Research, the Army Research Office under Contract number W911NF-2020-221, and the National Science Foundation under Grant Number CCF-1919197. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors.", + "bbox": [ + 496, + 796, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. 8th IEEE European Symposium on Security and Privacy (IEEE Euro S&P), 2023. 1, 2, 3, 4", + "[2] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 3", + "[3] Christopher A Choquette-Choo, Florian Tramer, Nicholas Carlini, and Nicolas Papernot. Label-only membership inference attacks. In International conference on machine learning, pages 1964-1974. PMLR, 2021. 1", + "[4] Lixin Fan, Kam Woh Ng, Ce Ju, Tianyu Zhang, Chang Liu, Chee Seng Chan, and Qiang Yang. Rethinking privacy preserving deep learning: How to evaluate and thwart privacy attacks. In Federated Learning, pages 32-50. Springer, 2020. 2", + "[5] Liam H Fowl, Jonas Geiping, Wojciech Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. In International Conference on Learning Representations, 2022. 1, 2, 3, 4", + "[6] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33:16937-16947, 2020. 1, 2, 3, 4, 6", + "[7] Jiahui Geng, Yongli Mou, Feifei Li, Qing Li, Oya Bayan, Stefan Decker, and Chunming Rong. Towards general deep leakage in federated learning. arXiv preprint arXiv:2110.09074, 2021. 2, 3", + "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[9] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603-618, 2017. 1", + "[10] Yangsibo Huang, Samyak Gupta, Zhao Song, Kai Li, and Sanjeev Arora. Evaluating gradient inversion attacks and defenses in federated learning. Advances in Neural Information Processing Systems, 34:7232-7241, 2021. 1, 3", + "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 4", + "[12] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 4", + "[13] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 4", + "[14] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 5, 6", + "[15] Xinjian Luo, Yuncheng Wu, Xiaokui Xiao, and Beng Chin Ooi. Feature inference attack on model predictions in vertical federated learning. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), pages 181-192. IEEE, 2021. 1", + "[16] Kailang Ma, Yu Sun, Jian Cui, Dawei Li, Zhenyu Guan, and Jianwei Liu. Instance-wise batch label restoration via gradients in federated learning. In The Eleventh International Conference on Learning Representations, 2022. 2, 3", + "[17] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1", + "[18] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), pages 691-706. IEEE, 2019. 1", + "[19] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In 2019 IEEE symposium on security and privacy (SP), pages 739-753. IEEE, 2019. 1", + "[20] Dario Pasquini, Danilo Francati, and Giuseppe Ateniese. Eluding secure aggregation in federated learning via model inconsistency. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 2429-2443, 2022. 1", + "[21] Le Trieu Phong, Yoshinori Aono, Takuya Hayashi, Lihua Wang, and Shiho Moriai. Privacy-preserving deep learning: Revisited and enhanced. In International Conference on Applications and Techniques in Information Security, pages 100-110. Springer, 2017. 2", + "[22] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1", + "[23] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE Conference on Computer Communications, pages 2512-2520. IEEE, 2019. 1", + "[24] Yuxin Wen, Jonas Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. International Conference on Machine Learning, 2022. 1", + "[25] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 5, 6", + "[26] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337-16346, 2021. 1, 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12255", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6", + "[28] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. arXiv preprint arXiv:2001.02610, 2020. 2", + "[29] Joshua C Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. The resource problem of using linear layer leakage attack in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2023. 1, 2, 3, 8", + "[30] Joshua Christian Zhao, Atul Sharma, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. Loki: Large-scale data reconstruction attack against federated learning through model manipulation. In 2024 IEEE Symposium on Security and Privacy (SP), pages 30-30. IEEE Computer Society, 2023. 1, 2, 3, 4, 5, 6", + "[31] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 5, 6, 8", + "[32] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 3" + ], + "bbox": [ + 78, + 90, + 467, + 443 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12256", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_model.json b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16edfae42cac8fc725c425328bda9be1fe97d3af --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.204, + 0.744, + 0.24 + ], + "angle": 0, + "content": "Joshua C. Zhao, Ahaan Dabholkar, Atul Sharma, Saurabh Bagchi Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.242, + 0.692, + 0.256 + ], + "angle": 0, + "content": "{zhao1207,adabholk,sharm438,sbagchi}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.324, + 0.474, + 0.628 + ], + "angle": 0, + "content": "Federated learning is a decentralized learning paradigm introduced to preserve privacy of client data. Despite this, prior work has shown that an attacker at the server can still reconstruct the private training data using only the client updates. These attacks are known as data reconstruction attacks and fall into two major categories: gradient inversion (GI) and linear layer leakage attacks (LLL). However, despite demonstrating the effectiveness of these attacks in breaching privacy, prior work has not investigated the usefulness of the reconstructed data for downstream tasks. In this work, we explore data reconstruction attacks through the lens of training and improving models with leaked data. We demonstrate the effectiveness of both GI and LLL attacks in maliciously training models using the leaked data more accurately than a benign federated learning strategy. Counter-intuitively, this bump in training quality can occur despite limited reconstruction quality or a small total number of leaked images. Finally, we show the limitations of these attacks for downstream training, individually for GI attacks and for LLL attacks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.64, + 0.21, + 0.655 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.856 + ], + "angle": 0, + "content": "With growing concerns of data privacy, federated learning (FL) [17] has gained traction as a potential privacy-preserving method for training machine learning models. Compared to centralized learning where training is done on data localized at a central server, FL takes a decentralized approach where participating clients train a model on their local data and send their updates to the server. A typical training round involves a server sending a model to the clients, the clients training the model using their local data, and finally having the clients send their updates to the server for aggregation. However, the privacy-preserving property of FL only holds if the updates cannot be used to extract sensitive information about the local training data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Despite only sending updates, prior works have shown the ability of attackers to gain information about the private training data through membership inference attacks [3," + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.291, + 0.868, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.378, + 0.868, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.469, + 0.865, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.581, + 0.808, + 0.595 + ], + "angle": 0, + "content": "Figure 1. Training using leaked data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.598, + 0.892, + 0.704 + ], + "angle": 0, + "content": "19, 22], property-inference attacks [15, 18], or GAN-based methods [9, 23]. Within these privacy attacks, data reconstruction has stood out as the most powerful, allowing attackers to directly recover the training data of clients [20, 24]. Within data reconstruction attacks, gradient inversion (optimization-based attacks) [6, 26, 32] and linear layer leakage [1, 5, 30] attacks stand out as the most common." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.902 + ], + "angle": 0, + "content": "While prior works have discussed the power and the limitations of these attacks in the context of an attacker simply breaching privacy [10, 29], to the best of our knowledge, none has discussed the effectiveness of using the leaked data for downstream model training. Figure 1 shows this process of training using leaked data. With the onset of deep learning, data has become a valuable commodity. This growth in AI also means that the value of data no longer lies just in breaching privacy of the raw data, but in the ability of the data to be used downstream for creating powerful models. Therefore, one measure of the success of a data reconstruction attack should be whether the leaked data is useful for machine learning training, an aspect omitted in prior work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12247" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "In order to discuss training on leaked data, we first introduce the attacks. Gradient inversion (GI) attacks typically reconstruct data through an iterative process where the distance between a gradient computed from a dummy image and the ground truth gradient is minimized by an optimizer. Label reconstruction, the process of recovering the batch labels from the client gradient, also plays a critical role in affecting the final reconstruction quality. These labels are typically leaked prior to optimization from the gradients of the final layer of the network [7, 16, 26, 28] and are matched with reconstructed images during the optimization. These attacks have shown success with individual updates of smaller batch sizes, up to 48 on ImageNet [26] or 100 on CIFAR [6]. However, this only considers breaching privacy, as a batch size of 48 on ImageNet or 100 on CIFAR result in low quality reconstructions and/or a very small number of identifiable images. Even for a modest batch size of 8 or 16 on CIFAR-10, some batch reconstructions already fail. Given the importance of data quality for training models, a sufficient condition for success is no longer if any images are identifiable, but rather the overall usefulness of the leaked data for the downstream training task. Here, even reconstructions with low similarity scores to the ground truth may still positively contribute to training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.473, + 0.761 + ], + "angle": 0, + "content": "Compared to GI, linear layer leakage (LLL) does not suffer from problems with reconstruction quality. Using modification of a fully-connected (linear, FC) layer, an attacker can directly recover a proportion of the images of a client batch [1, 5, 30]. Reconstructed images are done by solving a simple linear equation and have little computation overhead. However, these attacks typically require modification of the model architecture and can add a large overhead in terms of model size [29]. While data quality is not an issue, when using the leaked data for training models, labeling the data becomes a problem. Unlike the optimization attacks, even if labels are known prior to reconstruction, the LLL process does not match them to the leaked images. As a result, after reconstruction there would be a set of leaked images and labels, where the images are not matched with their corresponding label. Furthermore, because only a proportion of images in the given batch are leaked for LLL, even with a full set of labels, some labels will not have a leaked image to go with. This can lead to a tedious process of manually labeling leaked images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Viewing data reconstruction attacks through the lens of training models brings up many questions. How do models trained with leaked images compare to centralized or federated learning? Does the reconstruction quality impact how well models perform when trained on leaked data? How does the lack of label matching during reconstruction affect linear layer leakage? In this work we address these largely unanswered questions and highlight the problems that arise in this new setting. Our contributions go as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "- We demonstrate the effectiveness of training models using leaked data compared to federated learning and the centralized baselines. Using the data from linear layer leakage and optimization, models trained on leaked data can achieve \\(20.40\\%\\) and \\(17.58\\%\\) higher accuracy on CIFAR-10 compared to federated learning." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.183, + 0.895, + 0.302 + ], + "angle": 0, + "content": "- We show that similarity metrics such as PSNR do not show if reconstructed images are useful for training, as even some of the worst images improve models. Training on CIFAR-10 with bad reconstructions generated by Inverting Gradients (batch size 16) with PSNR \\(< 14\\) results in a \\(58.29\\%\\) test accuracy. Removing these images and training only on images \\(>14\\) PSNR results in only a minor accuracy decrease from \\(76.83\\%\\) to \\(75.48\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.303, + 0.894, + 0.409 + ], + "angle": 0, + "content": "- We quantify the effects and highlight the crucial limitations of current data reconstruction attacks when viewed through the lens of downstream model training. While gradient inversion can breach privacy for a batch size of 100 on CIFAR-10, this is impractical for training as even increasing the batch size from 4 to 16 results in a performance drop from \\(90.34\\%\\) to \\(76.83\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.64, + 0.433 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Gradient inversion (GI) attacks operate under an honest-but-curious threat model where a server only knows the update and model. The method involves feeding a dummy image into a model and computing the subsequent gradient. The dummy image starts out initialized as random noise, and an optimizer iteratively minimizes the distance between the computed gradient and the ground truth client gradient." + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.548, + 0.892, + 0.568 + ], + "angle": 0, + "content": "\\[\nx ^ {*} = \\arg \\min _ {x} \\left\\| \\nabla L (x, y, \\theta) - \\nabla W \\right\\| _ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.892, + 0.755 + ], + "angle": 0, + "content": "The intuition behind these attacks is that a similar computed dummy gradient \\(\\nabla L(x,y,\\theta)\\) and ground truth gradient \\(\\nabla W\\) will result from similar images. Thus, as the gradients become closer, the final reconstructed image \\(x^{*}\\) should become close to the ground truth training image. The label \\(y\\) is assumed to be known prior to optimization. For a batch reconstruction with multiple labels, the labels are matched with the reconstructed images during optimization (as the loss and gradient are computed with these image-label pairs). Regularizers have also been introduced to assist the optimization process [6, 26]. However, these can result in reconstructed image artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.894, + 0.801 + ], + "angle": 0, + "content": "Linear layer leakage (LLL) attacks are built on the observation that a fully-connected (FC) layer leaks the input to the layer through the gradients [4, 21]." + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.805, + 0.892, + 0.836 + ], + "angle": 0, + "content": "\\[\nx ^ {i} = \\frac {\\delta L}{\\delta W ^ {i}} / \\frac {\\delta L}{\\delta B ^ {i}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here, \\(\\frac{\\delta L}{\\delta W^i}\\) is the weight gradient and \\(\\frac{\\delta L}{\\delta B^i}\\) is the bias gradient of a neuron. Neuron \\(i\\) is activated by an image and \\(x^i\\) is the respective reconstructed image. However, this assumes that only a single image activated the neuron. If multiple" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.468, + 0.121 + ], + "angle": 0, + "content": "images activate the same neuron, reconstructed image \\( x^{i} \\) becomes a combination of the images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.288 + ], + "angle": 0, + "content": "To mitigate this, the methods of trap weights [1] and Robbing the Fed [5] were proposed. Trap weights initialize the weights of the FC layer randomly as half positive and half negative, where the negative parameters are sampled from a slightly larger magnitude. This method has difficulty with scaling [30] compared to Robbing the Fed, which proposed a method where the FC layer weights are used to measure a property of the images (e.g., average pixel brightness). The biases of the layer were then set to fit the distribution of the dataset and a ReLU activation was used to threshold the activation. Images were reconstructed as" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.292, + 0.468, + 0.323 + ], + "angle": 0, + "content": "\\[\nx ^ {i} = \\left(\\frac {\\delta L}{\\delta W ^ {i}} - \\frac {\\delta L}{\\delta W ^ {i + 1}}\\right) / \\left(\\frac {\\delta L}{\\delta B ^ {i}} - \\frac {\\delta L}{\\delta B ^ {i + 1}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.468, + 0.386 + ], + "angle": 0, + "content": "where \\( i + 1 \\) indicates the neuron with a bias being the next highest cutoff. This method requires basic knowledge about the input data distribution, but achieves a higher leakage rate compared to trap weights and is scalable to aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.468, + 0.568 + ], + "angle": 0, + "content": "In the case of FedAvg, the sparse variant of Robbing the Fed utilizes a two-sided activation function (such as Hard tanh) and weight/bias scaling to maintain the same leakage as in FedSGD. However, the sparse variant is not scalable and leads to precision problems when attacking secure aggregation [2] or larger batches. The LOKI attack addresses the scalability problems of both methods through the introduction of a convolutional layer that splits the scaling between number of clients and batch size in secure aggregation [30]. They further introduce a convolutional scaling factor (CSF) which achieves a higher leakage rate in FedAvg without increasing the FC layer size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.468, + 0.794 + ], + "angle": 0, + "content": "While there have been many works exploring data reconstruction, prior work has not evaluated these attacks in the context of downstream tasks. The works on GI [6, 26, 32] and LLL [1, 5, 30] attacks only discuss the reconstruction quality of the methods in terms of standard image metrics such as PSNR (peak signal-to-noise ratio), SSIM (structural similarity index measure), or LPIPS (learned perceptual image patch similarity). However, these metrics only measure how similar the reconstruction are to the ground truth and only give a vague sense of how useful the data is for training. Another work discussed the limitations of GI attacks [10] measured by image similarity. In [29], LLL attacks were measured in terms of the resource overhead added by the attacks. However, these works also did not discuss the usefulness of reconstructed data post-leakage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "This work aims to bridge the gap and explore the important facet of training machine learning models on leaked data from reconstruction attacks. Viewing leaked data beyond privacy breach brings to light key weaknesses of current attacks in generating useful data. While LLL reconstructs high quality images, the other important half of data, namely having matching labels, is missing. GI attacks can" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "breach privacy for larger batch sizes, but the quality of the leaked data would be nearly useless for training models. In the age of machine learning and artificial intelligence, the value of data has increased tremendously. Therefore, when looking at leaked data, it is important to measure success in terms of usefulness to model training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.189, + 0.723, + 0.205 + ], + "angle": 0, + "content": "3. Training on leaked data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.207, + 0.892, + 0.373 + ], + "angle": 0, + "content": "Data reconstruction attacks have largely focused on the quality of the recovered image in gradient inversion (GI) or the efficiency of linear layer leakage (LLL). However, label reconstruction is important for training on leaked data. Within the pipeline of GI attacks, labels are typically recovered prior to reconstruction, as they assist in the optimization problem [10]. Recent works have shown that label restoration with large batches and duplicate labels is possible [7, 16]. More accurate labels and the added flexibility of duplicate class images directly improves prior GI attacks such as Inverting Gradients [6] and GradInversion [26]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.373, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Interestingly, one strength of LLL has been in the ability to reconstruct images without knowledge of the data labels. While the class labels can affect the distribution of the client dataset (e.g., average image brightness can be different between classes) [30], the labels themselves do not help in the reconstruction. Knowing labels helps guide GI attacks (as gradients are computed from data and label pairs), but for LLL, images activating the neurons are reconstructed using only the weight and bias gradients of the FC layer as with Equation 3. Labels are necessary to train models using leaked data, but even with a set of recovered labels, the leaked images would still remain unmatched. Unlike GI which matches during optimization, images leaked by LLL would need to be manually matched. For this work, we experiment with cases when all labels are known and matched and alternatively, when only a small portion of labels are known and matched. Nonetheless, the label leakage and matching issue stands as an open problem for LLL attacks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Identifying which images are reconstructed correctly is also a challenging problem for LLL. For FedSGD, bins without activation can be identified, as after the subtraction process the reconstruction will be zero. However, if multiple images activate the same bin, the reconstruction is non-zero. By checking for non-zero values, we cannot directly identify whether the reconstruction is a single image or a combination of multiple. Manually filtering this can become a tedious process in practice. This is somewhat mitigated in the FedAvg case by LOKI. Across separate FedAvg local mini-batches, multiple images will not activate the same neuron, resulting in muchFewew overlapping reconstructions. Therefore the only images that can cause multiple activations are within the same mini-batch [30]. In this work, we manually filter out multiple image reconstructions in order to test the quality of the data. However, especially for LLL in FedSGD, the problem of how to automate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12249" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.418, + 0.215 + ], + "angle": 0, + "content": "
MNIST
Batch sizeModelTime per batch (s)Time for entire dataset (days)
8DNN54.794.76
1662.342.71
3264.321.40
CIFAR-10
Batch sizeModelTime per batch (s)Time for entire dataset (days)
4ResNet-18422.8161.17
8437.5431.65
16435.3315.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.278 + ], + "angle": 0, + "content": "Table 1. Computation time for Inverting Gradients running on a NVIDIA A100 80GB GPU. Small batches have more total batches and take more time for the whole dataset. Computation time is significantly higher for CIFAR-10 given the more complex model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.469, + 0.317 + ], + "angle": 0, + "content": "removal or how to use these multiple image reconstructions for training still remains an important item of future work." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.32, + 0.21, + 0.337 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.47, + 0.504 + ], + "angle": 0, + "content": "The goal of our work is to discuss the effectiveness of leaked data when used for the downstream task of image classification. We use centralized training on the entire benign dataset and federated learning (with only updates and no data shared by the clients) as the baselines for comparison. Typically, these serve as the upper bound and the lower bound respectively for the model accuracy. For federated learning, we use two settings of 10 and 50 clients. We use a non-IID bias of 0.5 and FedAvg with 3 local iterations. Additional training results for the IID setting and FedSGD are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.47, + 0.67 + ], + "angle": 0, + "content": "We use Inverting Gradients [6] as the gradient inversion (GI) attack. Another candidate is GradInversion, which is a SOTA GI attack (especially for ImageNet-sized images). However, we discarded this as the author code for this attack is not available and the third-party online implementations did not achieve comparable results to the original paper. For linear layer leakage (LLL), we use the LOKI attack [30]. We use LOKI as it achieves SOTA FedAvg leakage rate and equivalent SOTA leakage rate in FedSGD compared to Robbing the Fed. Model performance for other LLL methods [1, 5] can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.469, + 0.806 + ], + "angle": 0, + "content": "We use MNIST [13], CIFAR-10 [11], and the Tiny ImageNet [12] datasets. For the MNIST dataset, we train using a simple DNN with 2 convolutional layers and 2 FC layers. For CIFAR-10 and Tiny Imagenet, we use a ResNet-18 [8]. We evaluate on the validation set for Tiny ImageNet because the test set is unlabeled. The plots only show a single run for each setting, but the (numerical) testing performance is reported as an average over 5 runs. We run all experiments on NVIDIA A100 80GB GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.81, + 0.398, + 0.825 + ], + "angle": 0, + "content": "4.1. Gradient inversion computation time" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.9 + ], + "angle": 0, + "content": "Compared to LLL attacks, the computational overhead of GI attacks when reconstructing data, while done offline, is still non-negligible. In order to fully evaluate the leaked data, we ran the Inverting Gradients attack on the entire dataset of MNIST with batch sizes of 8, 16, and 32. We" + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.089, + 0.865, + 0.144 + ], + "angle": 0, + "content": "
CIFAR-10MNIST
Batch sizePSNRSSIMBatch sizePSNRSSIM
427.880.9067823.160.6996
820.770.72151619.150.6506
1615.940.39783215.780.4537
" + }, + { + "type": "table_caption", + "bbox": [ + 0.602, + 0.148, + 0.792, + 0.159 + ], + "angle": 0, + "content": "(a) Gradient Inversion PSNR/SSIM" + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.163, + 0.882, + 0.239 + ], + "angle": 0, + "content": "
CIFAR-10MNISTTiny ImageNet
FC factorLeaked imagesPercentLeaked imagesPercentLeaked imagesPercent
84378887.585254887.588580485.80
43946478.934596676.617514975.15
22988259.763579559.665801258.01
11824236.482181536.363539335.39
" + }, + { + "type": "table_caption", + "bbox": [ + 0.603, + 0.243, + 0.79, + 0.256 + ], + "angle": 0, + "content": "(b) Linear layer attack leakage rate" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.258, + 0.892, + 0.314 + ], + "angle": 0, + "content": "Table 2. (a) Average PSNR↑/SSIM↑ scores of gradient inversion (Inverting Gradients) reconstructions for various batch sizes and (b) number of leaked images/leakage rate of linear layer leakage attack (LOKI) for several FC layer sizes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.395 + ], + "angle": 0, + "content": "also ran over CIFAR-10 with batch sizes of 4, 8, and 16. We run 10,000 iterations of optimization for each batch. We evaluate the Inverting Gradients attack in the best case scenario with an untrained model where all labels are correctly recovered prior to optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.562 + ], + "angle": 0, + "content": "Table 1 shows the amount of time taken by Inverting Gradients run on a NVIDIA A100 80GB GPU for MNIST and CIFAR-10 on a DNN and ResNet-18 respectively. There is some variance in the average amount of time based on batch size, but when attacking the entire dataset, smaller batch sizes always take longer time as the number of total batches is much larger. CIFAR-10 takes significantly longer to reconstruct each batch given the more complex model of ResNet-18 compared to a DNN. For a batch size of 4, CIFAR-10 takes 422.81 seconds per batch. Iterating across the entire dataset on a single GPU takes 61.17 days." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.672, + 0.582 + ], + "angle": 0, + "content": "4.2. Leakage statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.703 + ], + "angle": 0, + "content": "For GI we use Inverting Gradients [6], and for LLL we use LOKI [30]. For Inverting Grads, we use a batch size of 4, 8, and 16 on CIFAR-10 and 8, 16 and 32 on MNIST. For LOKI, we choose the FC layer size based on the batch size. We vary this with factors of 1, 2, 4 and 8 on CIFAR-10, MNIST, and Tiny ImageNet. We use a client batch size of 64 for all datasets, so an FC size factor of 2 would mean \\(64 \\cdot 2 = 128\\) units in the FC layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Table 2a gives the average PSNR and SSIM metrics (where a higher score indicates greater similarity with the ground truth) for the reconstructions across the entire dataset of CIFAR-10 and MNIST from Inverting Gradients. Increasing the batch size results in a lower PSNR and SSIM score for both CIFAR-10 and MNIST." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 2b reports the number of leaked images for LOKI. For an FC factor of 4, LOKI leaks \\(78.93\\%\\), \\(76.61\\%\\), and \\(75.15\\%\\) of images on the CIFAR-10, MNIST, and Tiny ImageNet datasets. Increasing the FC layer size increases the model size overhead, but results in a higher leakage rate as images are more likely to activate different neurons. Decreasing the size creates a smaller model size overhead but" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12250" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.27, + 0.106 + ], + "angle": 0, + "content": "results in lower leakage rate." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.112, + 0.398, + 0.127 + ], + "angle": 0, + "content": "4.3. Training on leaked data from scratch" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.129, + 0.47, + 0.4 + ], + "angle": 0, + "content": "We start with a discussion on the models that can be trained from the leaked data of both GI and LLL attacks. The attacks are applied under the FedSGD setting. We also start with assuming that all labels are known prior to training — this is a best case assumption for the downstream training. For CIFAR-10, we train the centralized models using SGD with momentum 0.9 and weight decay 0.0001. We use a batch size of 128 and train for 200 epochs starting with an initial learning rate of 0.1 and a scheduler that decreases on plateau. For MNIST, we use SGD without momentum or weight decay and train for 40 epochs with a LR of 0.01. For Tiny ImageNet we use Adam with learning rate 0.001 and train for 50 epochs with a batch size of 1024. For federated learning, we use 400 training rounds for CIFAR-10 and MNIST and 500 rounds for Tiny ImageNet. Client batch size of 32, 128, and 256 are used for MNIST, CIFAR-10, and Tiny ImageNet. The test accuracy is downsampled to fit on the same axis with centralized training." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.47, + 0.779 + ], + "angle": 0, + "content": "Table 3a and Figure 2a show the results for CIFAR-10 for centralized training, federated learning (FedAvg), and training with leaked data. Centralized training achieves a \\(94.38\\%\\) accuracy and federated learning achieves a \\(72.76\\%\\) and \\(68.71\\%\\) peak accuracy with 10 and 50 clients respectively. Centralized and FL indeed provide the upper bound and the lower bound of the accuracy here. Across all attack settings, GI and LLL achieve higher model accuracy compared to federated learning. The reconstruction quality of GI is adversely affected by larger batch sizes and in turn also impacts the final model performance. With a batch size of 4, 8, and 16, models trained on the leaked data from GI achieves \\(90.34\\%\\), \\(86.16\\%\\), and \\(76.83\\%\\) accuracy. With even larger batch sizes, it is likely that the model performance will drop below federated learning. The performance of models using LLL data is more stable. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the accuracy only drops from \\(93.16\\%\\) to \\(88.86\\%\\). This is spite of a large drop in the number of leaked images from 43788 (\\(87.58\\%\\)) to 18242 (\\(36.48\\%\\)) as shown in Table 2b. This is a subtle and consequential result — the downstream training task can still be accomplished despite a big drop in the amount of data leakage. All prior work had stopped at the stage of evaluating the proportion of leakage and therefore had missed out on this insight." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Table 3b and Figure 2b show the results for MNIST. LLL performs better than the FedAvg baseline in all cases. With an FC size factor of 8, LOKI achieves a \\(98.82\\%\\) test accuracy (nearly the same regardless of the FC size factor), only \\(0.07\\%\\) lower than centralized at \\(98.89\\%\\). GI performs slightly worse than federated learning, achieving \\(95.96\\%\\), \\(95.50\\%\\), and \\(94.29\\%\\) at batch sizes 8, 16, and 32 compared to \\(96.17\\%\\) and \\(96.18\\%\\) with 10 and 50 clients in FedAvg." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "We believe this is because the noise in the reconstructions hampers model performance with GI. This affects MNIST more severely than for CIFAR because the MNIST images are more sparse and sensitive to noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.318 + ], + "angle": 0, + "content": "The top-1 validation accuracy for Tiny ImageNet are given in Table 3c and Figure 2c. The total number of leaked images hurts LLL more here than in the other two datasets. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the validation accuracy drops from \\(46.70\\%\\) to \\(35.20\\%\\), an \\(11.5\\%\\) drop in performance. Federated learning achieves \\(37.00\\%\\) and \\(35.06\\%\\) accuracy with 10 and 50 clients respectively. FL with 10 clients has slightly better accuracy than LLL with an FC size factor of 1. However, any LLL setting with a larger FC layer achieves higher accuracy than both settings in federated learning." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.321, + 0.837, + 0.337 + ], + "angle": 0, + "content": "4.4. FedAvg with leaked images from LOKI" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.443 + ], + "angle": 0, + "content": "We explore the impact of leaking images using LLL in FedAvg when training models. We use 8 local iterations of mini-batch size 8, a local dataset size of 64, and a learning rate of 1e-4. Following the LOKI attack [30], \\(\\mathrm{CSF} = 500\\) is used to achieve a higher leakage rate and reconstruction quality. Similar to before, we apply the attack across batches covering the entire CIFAR-10 training dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.893, + 0.564 + ], + "angle": 0, + "content": "Table 4 shows the leakage rate of LOKI in FedAvg compared to FedSGD. Using \\( CSF = 500 \\), the leakage rate of LOKI in FedAvg is substantially higher than the leakage rates in the FedSGD setting. Using an FC layer of half the size, the leakage rate in FedAvg is comparable to that of FedSGD. For example, LOKI FedAvg factor 1 achieves a leakage rate of \\( 58.39\\% \\) compared to LOKI FedSGD factor 2 which achieves a leakage rate of \\( 59.76\\% \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.669 + ], + "angle": 0, + "content": "Table 5 shows the testing accuracy of LOKI FedAvg for the different FC layer sizes. The test accuracies are very comparable to the FedSGD settings. However, with a smaller FC layer size factor, the effect of having a larger total number of leaked images becomes visible. With FC factor 1, LOKI FedAvg achieves \\(91.11\\%\\) accuracy while LOKI FedSGD achieves \\(88.86\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.673, + 0.837, + 0.689 + ], + "angle": 0, + "content": "4.5. Training with semi-supervised learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In order to train machine learning models in a supervised fashion, images leaked through LLL need to be labeled. Previously, we assumed that all labels were known and matched to the images prior to training, but this is impractical. Within a LLL batch reconstruction, the order of images is not preserved. Even if labels were recovered like for GI attacks, this would result in a set of unmatched ground truth labels and images. Manual hand labeling is tedious and this inability to match labels is a weakness of LLL attacks only when it comes to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this experiment, we assume that only a small portion of images are labeled. This setting is similar to semi-supervised learning (SSL) [25, 31] and we use CoMatch [14], a semi-supervised algorithm, to train the mod" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.345, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.229, + 0.253, + 0.242 + ], + "angle": 0, + "content": "(a) CIFAR-10" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.092, + 0.614, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.229, + 0.517, + 0.242 + ], + "angle": 0, + "content": "(b) MNIST" + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.092, + 0.887, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.229, + 0.807, + 0.242 + ], + "angle": 0, + "content": "(c) Tiny ImageNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.248, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Training models on (a) CIFAR-10, (b) MNIST, and (c) Tiny ImageNet with leaked data compared to centralized and federated learning training. Both linear layer leakage and gradient inversion achieve higher accuracy than the federated learning (FedAvg) baseline for CIFAR-10 in all cases. For MNIST, LLL (LOKI) nearly reaches centralized accuracy while GI performs slightly worse than FL. Top-1 validation accuracy used when training models on Tiny ImageNet. Here, LLL performs better with a FC layer size factor of 2 or higher." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.307, + 0.346, + 0.444 + ], + "angle": 0, + "content": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-94.38
FedAvg1072.76
5068.71
Gradient inversion (Inverting grads [6])490.34
886.16
1676.83
893.16
Linear layer leakage (LOKI [30])492.94
291.90
188.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.447, + 0.253, + 0.458 + ], + "angle": 0, + "content": "(a) CIFAR-10" + }, + { + "type": "table", + "bbox": [ + 0.356, + 0.307, + 0.617, + 0.444 + ], + "angle": 0, + "content": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-98.89
FedAvg1096.17
5096.18
Gradient inversion (Inverting grads [6])895.96
1695.50
3294.29
898.82
Linear layer leakage (LOKI [30])498.70
298.72
198.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.455, + 0.447, + 0.517, + 0.458 + ], + "angle": 0, + "content": "(b) MNIST" + }, + { + "type": "table", + "bbox": [ + 0.623, + 0.327, + 0.892, + 0.423 + ], + "angle": 0, + "content": "
Num. clients / FC factorValidation accuracy
Centralized-48.55
FedAvg1037.00
5035.06
Linear layer leakage (LOKI [30])846.70
445.09
241.15
135.20
" + }, + { + "type": "table_caption", + "bbox": [ + 0.707, + 0.426, + 0.807, + 0.439 + ], + "angle": 0, + "content": "(c) Tiny ImageNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.459, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Table 3. (a) CIFAR-10 and (b) MNIST test accuracy, (c) Tiny ImageNet top-1 validation accuracy trained from scratch with leaked data. CIFAR-10 and Tiny ImageNet are trained with a ResNet-18 and MNIST is trained with a DNN. For the two attacks (gradient inversion and linear layer leakage), training happens with leaked data. Second column indicates the number of clients in federated learning (FedAvg), the batch size for the gradient inversion attack, and the fully-connected layer size factor for linear layer leakage. Best accuracy is used for FedAvg and final accuracy for all other settings." + }, + { + "type": "table", + "bbox": [ + 0.161, + 0.538, + 0.388, + 0.602 + ], + "angle": 0, + "content": "
FC size factorLOKI FedSGDLOKI FedAvg
887.58 (43788)92.19 (46097)
478.93 (39464)85.98 (42989)
259.76 (29882)75.29 (37645)
136.48 (18242)58.39 (29196)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.607, + 0.471, + 0.649 + ], + "angle": 0, + "content": "Table 4. Leakage rate % (leaked images) of LOKI in FedSGD and FedAvg on CIFAR-10 for several FC layer sizes. FedAvg with half the FC layer size has comparable leakage rate to FedSGD." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.655, + 0.425, + 0.72 + ], + "angle": 0, + "content": "
FC factorFedSGD accuracyFedAvg accuracy
Linear layer leakage (LOKI [30])893.1693.31
492.9492.88
291.9092.35
188.8691.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.724, + 0.434, + 0.739 + ], + "angle": 0, + "content": "Table 5. CIFAR-10 test accuracy with LOKI FedAvg." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.901 + ], + "angle": 0, + "content": "els. We use a WideResNet-28-2 [27] and train for 450 epochs on CIFAR-10. Figure 3a fixes the number of known labels to be 40 and shows the test accuracy curve with different fully connected layer sizes. We note that when there is only a small proportion of leaked data, this leads to a pronounced decrease in performance in SSL. LOKI with an FC factor of 1 (with \\(36.48\\%\\) leakage rate) illustrates this, as it only achieves \\(78.31\\%\\) max accuracy compared to the CIFAR-10 40 label baseline of \\(92.18\\%\\). With a factor of 2, the accuracy improves and jumps to \\(88.91\\%\\) as the leak" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.678 + ], + "angle": 0, + "content": "age rate also substantially increases to \\(59.76\\%\\). With factors 4 and 8 (78.93% and 87.58% leakage rate), the accuracy is very close to the baseline at 91.18% and 91.84% respectively. Figure 3b shows the test accuracy plot of LOKI size factor 4 with the number of known labels being 20, 40, and 250. With 20 and 250 labels, LOKI factor 4 achieves 86.69% and 91.65% peak accuracy. We note that there is a greater variability in accuracy with a smaller number of labels, a trend consistent with other SSL methods [14, 25, 31]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.686, + 0.844, + 0.702 + ], + "angle": 0, + "content": "4.6. Starting training from federated models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While the previous experiments worked with leaked data from all batches in the dataset, such large amounts of leaked data may not always be available. For example, LLL attacks may affect model performance due to manipulation of the model parameters and/or architecture. In later rounds when model performance is high, this may be easier to detect. Similarly, GI attacks such as Inverting Gradients [6] have better reconstruction quality with an untrained network. In both cases, it is possible that the attack is only applied during a few initial rounds of training and only a small portion of data is leaked. Training from scratch with a small portion of leaked data does not achieve good model performance alone, so instead, we can initialize from the fully trained FL" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12252" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.416, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.21, + 0.239, + 0.338, + 0.251 + ], + "angle": 0, + "content": "(a) Varying leakage rate" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.256, + 0.413, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.405, + 0.349, + 0.417 + ], + "angle": 0, + "content": "(b) Varying number of labels" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Semi-supervised learning using CoMatch on CIFAR-10 with a WideResNet with (a) a varying FC size and leakage rate (LR) for LOKI and the known labels fixed at 40 and (b) a fixed number of leaked images and 20, 40, and 250 known labels." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.485, + 0.416, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Figure 4. FL model trained with 50 clients as start point for training with leaked LLL data. Highlighted area indicates improvement above the FL model and training on the leaked data alone." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.687, + 0.436, + 0.702 + ], + "angle": 0, + "content": "model and then train it centrally using the leaked data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We use smaller portions of the leaked data from LOKI FC size factor of 8 for this experiment. With only \\(5\\%\\) (2500) of leaked images in CIFAR-10, a model trained from scratch can only achieve \\(61.15\\%\\) test accuracy. With \\(2\\%\\) (1000 images) of data, the model performance drops to \\(39.63\\%\\). Both cases are below an FL model with 50 clients that achieves a \\(64.08\\%\\) final test accuracy. However, by using the FL model as the initial starting point, even a small amount of leaked data allows for an increase in performance. We take the model that has been trained using FL and then additionally train that model centrally. Using only \\(2\\%\\) and \\(5\\%\\) of the leaked data, the models are able to achieve \\(68.25\\%\\) and \\(74.25\\%\\) final testing ac" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.681, + 0.164 + ], + "angle": 0, + "content": "
PSNR% ims keptTest accuracy
>2015.8870.04
>1827.5971.52
>1640.7873.73
>1460.2275.48
>1284.4276.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.174, + 0.668, + 0.186 + ], + "angle": 0, + "content": "(a) PSNR above threshold" + }, + { + "type": "table", + "bbox": [ + 0.708, + 0.089, + 0.875, + 0.164 + ], + "angle": 0, + "content": "
PSNR% ims keptTest accuracy
< 2084.1273.03
< 1872.4167.09
< 1659.2263.71
< 1439.7858.29
< 1215.5845.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.721, + 0.174, + 0.862, + 0.186 + ], + "angle": 0, + "content": "(b) PSNR below threshold" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.188, + 0.892, + 0.23 + ], + "angle": 0, + "content": "Table 6. Training models using CIFAR-10 leaked data from inverting gradients batch size 16. Only the reconstructions with a PSNR (a) above and (b) below the threshold are used in training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.24, + 0.892, + 0.316 + ], + "angle": 0, + "content": "curacy, a \\(28.62\\%\\) and \\(13.10\\%\\) improvement from starting from scratch. Figure 4 shows the improvement in test accuracy achieved by using the trained FL model as initialization instead of starting from scratch. The improvement is especially noticeable with small percentages of leaked data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.768, + 0.335 + ], + "angle": 0, + "content": "4.7. Quality of data reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.336, + 0.892, + 0.426 + ], + "angle": 0, + "content": "Quality of reconstruction is an important concern for GI attacks. As the batch size continues to increase, the reconstruction quality decreases. Table 2a also shows this relationship. From the previous experiments, we also see a negative relationship between the increasing batch size and the usefulness of the leaked data in downstream model training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.893, + 0.683 + ], + "angle": 0, + "content": "Here, we explore whether poor quality reconstructions are still useful for training models. Using the CIFAR-10 leaked dataset created using Inverting Gradients on a batch size of 16, we first sort the reconstructions by their PSNR value. Our first set of experiments removes the worst images from training. Table 6a shows the percent of images above the PSNR threshold and the final test accuracy when trained on that data. Compared to the baseline accuracy of \\(76.83\\%\\) achieved by using the entire leaked dataset, even removing the worst images with a \\(\\mathrm{PSNR} < 12\\) from the training set results in a small but non-zero performance loss. Then we take a different look at this question by now training on the worst images. Table 6b shows the test accuracy of models trained on all images below a PSNR threshold. Even when training on the lowest quality images with a PSNR \\(< 12\\), the model achieves a \\(45.05\\%\\) final testing accuracy, significantly worse than the baseline, but much above zero." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.892, + 0.805 + ], + "angle": 0, + "content": "From the previous results, we see that even leaked data with a low reconstruction quality is still useful to the model training process. This message is a new insight — all prior works when calculating the leak rate do not count poorly reconstructed images. We additionally include results using SSIM in the supplementary section. Visual examples of poor quality reconstructions used to train the models are also included in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.809, + 0.89, + 0.855 + ], + "angle": 0, + "content": "4.8. Observations on reconstruction quality trends We additionally note a few trends manually observed when going through the GI batch reconstructions of CIFAR-10." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "While some reconstructions have very low quality, only images of the same label will be swapped within the batch during label matching. This is a desirable property for train" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.092, + 0.445, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.178, + 0.47, + 0.233 + ], + "angle": 0, + "content": "Figure 5. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Ground truth images are on top and reconstructions are on the bottom. All labels in the batch are different and reconstructed images are high quality." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.234, + 0.445, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.319, + 0.47, + 0.388 + ], + "angle": 0, + "content": "Figure 6. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Duplicate labels exist between images 0, 4, 7 and images 2, 3. Reconstruction 3 contains parts of the content from both images 2 and 3. Reconstructions 0, 4, and 7 appear almost as noise. However, the other images still have reasonable quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.468, + 0.499 + ], + "angle": 0, + "content": "ing on leaked data. As long as the set of labels is leaked correctly, the optimization process will match images to a correct label. So even though the image quality is poor (and can be entirely unrecognizable), since the label is correctly matched, such an image still does help in the training (this buttresses the message from Tables 6a and 6b). Naturally, as the PSNR goes down, the image helps less and less." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.468, + 0.725 + ], + "angle": 0, + "content": "We also note that when all labels are different within a batch, the reconstruction quality of the entire batch of images is almost always good. An example on CIFAR-10 is shown in Figure 5. Usually when the reconstructions are poor, it occurs between groups of images in the same batch that belong to the same class. Poor reconstructions can appear as almost entirely noise, but occasionally images can have visible pieces of each image from the same class in the batch stitched together. The example batch in Figure 6 depicts these properties of poorly reconstructed images. As a side note, we observe that a batch may have a mix of well reconstructed and poorly reconstructed images. We also note that the reverse mapping of the same label images to poor reconstructions is not true — having duplicate labels does not always imply that the reconstructions will be bad." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.732, + 0.192, + 0.747 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As a main goal of our work, we have explored the effect of leaked data through today's leading data reconstruction attacks on downstream model training. Our main message is that it is important to also consider the use of data beyond reconstruction quality and image similarity. It is important to consider how far these leaked samples help in a downstream training task. From the experiments, we have seen that current reconstruction attacks are powerful, but still lack in several critical areas. For gradient inversion attacks, the reconstruction quality does pose an issue for" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "training models. Another impediment to practical use of this attack type is the time and computational cost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.393 + ], + "angle": 0, + "content": "Linear layer leakage attacks do not suffer from reconstruction quality issues and are much lighter weight computationally, but still suffer from other challenges. While LOKI has increased the efficiency of the FC layer leakage in FedAvg, there is still room for improvement (especially in FedSGD). As shown in [29], the model size overhead added by these attacks can be very large, especially if secure aggregation is used. Furthermore, the lack of labels poses a large challenge in training models on the data. With a small dataset like CIFAR-10, hand labeling 40 images is not too challenging. However, larger datasets will require more total labels. For example, SimMatch with \\(1\\%\\) of labels on ImageNet can achieve \\(67.2\\%\\) top-1 accuracy [31]. However, even \\(1\\%\\) of the total images in ImageNet-1k is 12,812 images. Even if the images were leaked, hand labeling would be extremely hard. Furthermore, we see from Figure 3a that having less total data and fewer labels results in an even steeper decrease in performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.545 + ], + "angle": 0, + "content": "While not discussed previously, there are many other aspects that affect data reconstruction attacks when considering training models. For example, how does the non-IID aspect of client data affect reconstruction? Other relevant settings of federated learning such as asynchronous federated learning, client selection, and differential privacy can also affect the final leaked dataset. Centralized training also allows for more complex and larger models without the communication or computational restrictions of clients in federated learning. These topics can be explored in future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.551, + 0.619, + 0.566 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We examine data reconstruction attacks through the lens of training models on the leaked data. While highlighting how the weaknesses of gradient inversion in terms of reconstruction quality affect downstream model training, we also show that even poorly reconstructed images are useful for training. We also discuss how the label matching problem for linear layer leakage can be mitigated through semisupervised learning. Under ideal conditions, we demonstrate that leaked data from both gradient inversion and linear layer leakage attacks are able to train powerful models comparable to even a centralized baseline. On CIFAR-10, gradient inversion and linear layer leakage attacks achieve \\(90.34\\%\\) and \\(93.16\\%\\) testing accuracy respectively, \\(17.58\\%\\) and \\(20.40\\%\\) higher than federated learning and only \\(4.04\\%\\) and \\(1.22\\%\\) lower compared to centralized training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This material is based in part upon work supported by Adobe Research, the Army Research Office under Contract number W911NF-2020-221, and the National Science Foundation under Grant Number CCF-1919197. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12254" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. 8th IEEE European Symposium on Security and Privacy (IEEE Euro S&P), 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.268 + ], + "angle": 0, + "content": "[2] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Christopher A Choquette-Choo, Florian Tramer, Nicholas Carlini, and Nicolas Papernot. Label-only membership inference attacks. In International conference on machine learning, pages 1964-1974. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.325, + 0.469, + 0.393 + ], + "angle": 0, + "content": "[4] Lixin Fan, Kam Woh Ng, Ce Ju, Tianyu Zhang, Chang Liu, Chee Seng Chan, and Qiang Yang. Rethinking privacy preserving deep learning: How to evaluate and thwart privacy attacks. In Federated Learning, pages 32-50. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[5] Liam H Fowl, Jonas Geiping, Wojciech Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. In International Conference on Learning Representations, 2022. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.533 + ], + "angle": 0, + "content": "[6] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33:16937-16947, 2020. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.535, + 0.469, + 0.589 + ], + "angle": 0, + "content": "[7] Jiahui Geng, Yongli Mou, Feifei Li, Qing Li, Oya Bayan, Stefan Decker, and Chunming Rong. Towards general deep leakage in federated learning. arXiv preprint arXiv:2110.09074, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[9] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603-618, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[10] Yangsibo Huang, Samyak Gupta, Zhao Song, Kai Li, and Sanjeev Arora. Evaluating gradient inversion attacks and defenses in federated learning. Advances in Neural Information Processing Systems, 34:7232-7241, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[12] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[13] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[15] Xinjian Luo, Yuncheng Wu, Xiaokui Xiao, and Beng Chin Ooi. Feature inference attack on model predictions in vertical federated learning. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), pages 181-192. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[16] Kailang Ma, Yu Sun, Jian Cui, Dawei Li, Zhenyu Guan, and Jianwei Liu. Instance-wise batch label restoration via gradients in federated learning. In The Eleventh International Conference on Learning Representations, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[17] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[18] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), pages 691-706. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In 2019 IEEE symposium on security and privacy (SP), pages 739-753. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[20] Dario Pasquini, Danilo Francati, and Giuseppe Ateniese. Eluding secure aggregation in federated learning via model inconsistency. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 2429-2443, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Le Trieu Phong, Yoshinori Aono, Takuya Hayashi, Lihua Wang, and Shiho Moriai. Privacy-preserving deep learning: Revisited and enhanced. In International Conference on Applications and Techniques in Information Security, pages 100-110. Springer, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[22] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE Conference on Computer Communications, pages 2512-2520. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[24] Yuxin Wen, Jonas Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. International Conference on Machine Learning, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337-16346, 2021. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12255" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[27] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.468, + 0.162 + ], + "angle": 0, + "content": "[28] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. arXiv preprint arXiv:2001.02610, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.468, + 0.247 + ], + "angle": 0, + "content": "[29] Joshua C Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. The resource problem of using linear layer leakage attack in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2023. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[30] Joshua Christian Zhao, Atul Sharma, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. Loki: Large-scale data reconstruction attack against federated learning through model manipulation. In 2024 IEEE Symposium on Security and Privacy (SP), pages 30-30. IEEE Computer Society, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.468, + 0.402 + ], + "angle": 0, + "content": "[31] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.468, + 0.444 + ], + "angle": 0, + "content": "[32] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12256" + } + ] +] \ No newline at end of file diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_origin.pdf b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..960bfae670b9bfba3501c84f6040e9429f3823ce --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/c7bd675a-7448-4fdd-958a-3d9af33c562c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70f2de443c85f22ed16f7efd56ecd077bf66fd386cf3dd55c1bc741686d97191 +size 997711 diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/full.md b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..45ba4683d77d94dfdd1030c8e8a553f0be1e2f8b --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/full.md @@ -0,0 +1,289 @@ +# Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning + +Joshua C. Zhao, Ahaan Dabholkar, Atul Sharma, Saurabh Bagchi Purdue University + +{zhao1207,adabholk,sharm438,sbagchi}@purdue.edu + +# Abstract + +Federated learning is a decentralized learning paradigm introduced to preserve privacy of client data. Despite this, prior work has shown that an attacker at the server can still reconstruct the private training data using only the client updates. These attacks are known as data reconstruction attacks and fall into two major categories: gradient inversion (GI) and linear layer leakage attacks (LLL). However, despite demonstrating the effectiveness of these attacks in breaching privacy, prior work has not investigated the usefulness of the reconstructed data for downstream tasks. In this work, we explore data reconstruction attacks through the lens of training and improving models with leaked data. We demonstrate the effectiveness of both GI and LLL attacks in maliciously training models using the leaked data more accurately than a benign federated learning strategy. Counter-intuitively, this bump in training quality can occur despite limited reconstruction quality or a small total number of leaked images. Finally, we show the limitations of these attacks for downstream training, individually for GI attacks and for LLL attacks. + +# 1. Introduction + +With growing concerns of data privacy, federated learning (FL) [17] has gained traction as a potential privacy-preserving method for training machine learning models. Compared to centralized learning where training is done on data localized at a central server, FL takes a decentralized approach where participating clients train a model on their local data and send their updates to the server. A typical training round involves a server sending a model to the clients, the clients training the model using their local data, and finally having the clients send their updates to the server for aggregation. However, the privacy-preserving property of FL only holds if the updates cannot be used to extract sensitive information about the local training data. + +Despite only sending updates, prior works have shown the ability of attackers to gain information about the private training data through membership inference attacks [3, + +![](images/b3bea72727bea4569b30fade7e1446724c05af839d2023a8d61b20a1dc5f5d26.jpg) + +![](images/c4837e62ab0979054111d2495c3728b81b524954bf875568d00e077bde60f6fd.jpg) + +![](images/2e2fe93a86039f39c7cc7966222d872c01e7c854198f4127cd54630fdaf9b1dc.jpg) +Figure 1. Training using leaked data. + +19, 22], property-inference attacks [15, 18], or GAN-based methods [9, 23]. Within these privacy attacks, data reconstruction has stood out as the most powerful, allowing attackers to directly recover the training data of clients [20, 24]. Within data reconstruction attacks, gradient inversion (optimization-based attacks) [6, 26, 32] and linear layer leakage [1, 5, 30] attacks stand out as the most common. + +While prior works have discussed the power and the limitations of these attacks in the context of an attacker simply breaching privacy [10, 29], to the best of our knowledge, none has discussed the effectiveness of using the leaked data for downstream model training. Figure 1 shows this process of training using leaked data. With the onset of deep learning, data has become a valuable commodity. This growth in AI also means that the value of data no longer lies just in breaching privacy of the raw data, but in the ability of the data to be used downstream for creating powerful models. Therefore, one measure of the success of a data reconstruction attack should be whether the leaked data is useful for machine learning training, an aspect omitted in prior work. + +In order to discuss training on leaked data, we first introduce the attacks. Gradient inversion (GI) attacks typically reconstruct data through an iterative process where the distance between a gradient computed from a dummy image and the ground truth gradient is minimized by an optimizer. Label reconstruction, the process of recovering the batch labels from the client gradient, also plays a critical role in affecting the final reconstruction quality. These labels are typically leaked prior to optimization from the gradients of the final layer of the network [7, 16, 26, 28] and are matched with reconstructed images during the optimization. These attacks have shown success with individual updates of smaller batch sizes, up to 48 on ImageNet [26] or 100 on CIFAR [6]. However, this only considers breaching privacy, as a batch size of 48 on ImageNet or 100 on CIFAR result in low quality reconstructions and/or a very small number of identifiable images. Even for a modest batch size of 8 or 16 on CIFAR-10, some batch reconstructions already fail. Given the importance of data quality for training models, a sufficient condition for success is no longer if any images are identifiable, but rather the overall usefulness of the leaked data for the downstream training task. Here, even reconstructions with low similarity scores to the ground truth may still positively contribute to training. + +Compared to GI, linear layer leakage (LLL) does not suffer from problems with reconstruction quality. Using modification of a fully-connected (linear, FC) layer, an attacker can directly recover a proportion of the images of a client batch [1, 5, 30]. Reconstructed images are done by solving a simple linear equation and have little computation overhead. However, these attacks typically require modification of the model architecture and can add a large overhead in terms of model size [29]. While data quality is not an issue, when using the leaked data for training models, labeling the data becomes a problem. Unlike the optimization attacks, even if labels are known prior to reconstruction, the LLL process does not match them to the leaked images. As a result, after reconstruction there would be a set of leaked images and labels, where the images are not matched with their corresponding label. Furthermore, because only a proportion of images in the given batch are leaked for LLL, even with a full set of labels, some labels will not have a leaked image to go with. This can lead to a tedious process of manually labeling leaked images. + +Viewing data reconstruction attacks through the lens of training models brings up many questions. How do models trained with leaked images compare to centralized or federated learning? Does the reconstruction quality impact how well models perform when trained on leaked data? How does the lack of label matching during reconstruction affect linear layer leakage? In this work we address these largely unanswered questions and highlight the problems that arise in this new setting. Our contributions go as follows: + +- We demonstrate the effectiveness of training models using leaked data compared to federated learning and the centralized baselines. Using the data from linear layer leakage and optimization, models trained on leaked data can achieve $20.40\%$ and $17.58\%$ higher accuracy on CIFAR-10 compared to federated learning. +- We show that similarity metrics such as PSNR do not show if reconstructed images are useful for training, as even some of the worst images improve models. Training on CIFAR-10 with bad reconstructions generated by Inverting Gradients (batch size 16) with PSNR $< 14$ results in a $58.29\%$ test accuracy. Removing these images and training only on images $>14$ PSNR results in only a minor accuracy decrease from $76.83\%$ to $75.48\%$ . +- We quantify the effects and highlight the crucial limitations of current data reconstruction attacks when viewed through the lens of downstream model training. While gradient inversion can breach privacy for a batch size of 100 on CIFAR-10, this is impractical for training as even increasing the batch size from 4 to 16 results in a performance drop from $90.34\%$ to $76.83\%$ . + +# 2. Related work + +Gradient inversion (GI) attacks operate under an honest-but-curious threat model where a server only knows the update and model. The method involves feeding a dummy image into a model and computing the subsequent gradient. The dummy image starts out initialized as random noise, and an optimizer iteratively minimizes the distance between the computed gradient and the ground truth client gradient. + +$$ +x ^ {*} = \arg \min _ {x} \left\| \nabla L (x, y, \theta) - \nabla W \right\| _ {2} \tag {1} +$$ + +The intuition behind these attacks is that a similar computed dummy gradient $\nabla L(x,y,\theta)$ and ground truth gradient $\nabla W$ will result from similar images. Thus, as the gradients become closer, the final reconstructed image $x^{*}$ should become close to the ground truth training image. The label $y$ is assumed to be known prior to optimization. For a batch reconstruction with multiple labels, the labels are matched with the reconstructed images during optimization (as the loss and gradient are computed with these image-label pairs). Regularizers have also been introduced to assist the optimization process [6, 26]. However, these can result in reconstructed image artifacts. + +Linear layer leakage (LLL) attacks are built on the observation that a fully-connected (FC) layer leaks the input to the layer through the gradients [4, 21]. + +$$ +x ^ {i} = \frac {\delta L}{\delta W ^ {i}} / \frac {\delta L}{\delta B ^ {i}} \tag {2} +$$ + +Here, $\frac{\delta L}{\delta W^i}$ is the weight gradient and $\frac{\delta L}{\delta B^i}$ is the bias gradient of a neuron. Neuron $i$ is activated by an image and $x^i$ is the respective reconstructed image. However, this assumes that only a single image activated the neuron. If multiple + +images activate the same neuron, reconstructed image $x^{i}$ becomes a combination of the images. + +To mitigate this, the methods of trap weights [1] and Robbing the Fed [5] were proposed. Trap weights initialize the weights of the FC layer randomly as half positive and half negative, where the negative parameters are sampled from a slightly larger magnitude. This method has difficulty with scaling [30] compared to Robbing the Fed, which proposed a method where the FC layer weights are used to measure a property of the images (e.g., average pixel brightness). The biases of the layer were then set to fit the distribution of the dataset and a ReLU activation was used to threshold the activation. Images were reconstructed as + +$$ +x ^ {i} = \left(\frac {\delta L}{\delta W ^ {i}} - \frac {\delta L}{\delta W ^ {i + 1}}\right) / \left(\frac {\delta L}{\delta B ^ {i}} - \frac {\delta L}{\delta B ^ {i + 1}}\right) \tag {3} +$$ + +where $i + 1$ indicates the neuron with a bias being the next highest cutoff. This method requires basic knowledge about the input data distribution, but achieves a higher leakage rate compared to trap weights and is scalable to aggregation. + +In the case of FedAvg, the sparse variant of Robbing the Fed utilizes a two-sided activation function (such as Hard tanh) and weight/bias scaling to maintain the same leakage as in FedSGD. However, the sparse variant is not scalable and leads to precision problems when attacking secure aggregation [2] or larger batches. The LOKI attack addresses the scalability problems of both methods through the introduction of a convolutional layer that splits the scaling between number of clients and batch size in secure aggregation [30]. They further introduce a convolutional scaling factor (CSF) which achieves a higher leakage rate in FedAvg without increasing the FC layer size. + +While there have been many works exploring data reconstruction, prior work has not evaluated these attacks in the context of downstream tasks. The works on GI [6, 26, 32] and LLL [1, 5, 30] attacks only discuss the reconstruction quality of the methods in terms of standard image metrics such as PSNR (peak signal-to-noise ratio), SSIM (structural similarity index measure), or LPIPS (learned perceptual image patch similarity). However, these metrics only measure how similar the reconstruction are to the ground truth and only give a vague sense of how useful the data is for training. Another work discussed the limitations of GI attacks [10] measured by image similarity. In [29], LLL attacks were measured in terms of the resource overhead added by the attacks. However, these works also did not discuss the usefulness of reconstructed data post-leakage. + +This work aims to bridge the gap and explore the important facet of training machine learning models on leaked data from reconstruction attacks. Viewing leaked data beyond privacy breach brings to light key weaknesses of current attacks in generating useful data. While LLL reconstructs high quality images, the other important half of data, namely having matching labels, is missing. GI attacks can + +breach privacy for larger batch sizes, but the quality of the leaked data would be nearly useless for training models. In the age of machine learning and artificial intelligence, the value of data has increased tremendously. Therefore, when looking at leaked data, it is important to measure success in terms of usefulness to model training. + +# 3. Training on leaked data + +Data reconstruction attacks have largely focused on the quality of the recovered image in gradient inversion (GI) or the efficiency of linear layer leakage (LLL). However, label reconstruction is important for training on leaked data. Within the pipeline of GI attacks, labels are typically recovered prior to reconstruction, as they assist in the optimization problem [10]. Recent works have shown that label restoration with large batches and duplicate labels is possible [7, 16]. More accurate labels and the added flexibility of duplicate class images directly improves prior GI attacks such as Inverting Gradients [6] and GradInversion [26]. + +Interestingly, one strength of LLL has been in the ability to reconstruct images without knowledge of the data labels. While the class labels can affect the distribution of the client dataset (e.g., average image brightness can be different between classes) [30], the labels themselves do not help in the reconstruction. Knowing labels helps guide GI attacks (as gradients are computed from data and label pairs), but for LLL, images activating the neurons are reconstructed using only the weight and bias gradients of the FC layer as with Equation 3. Labels are necessary to train models using leaked data, but even with a set of recovered labels, the leaked images would still remain unmatched. Unlike GI which matches during optimization, images leaked by LLL would need to be manually matched. For this work, we experiment with cases when all labels are known and matched and alternatively, when only a small portion of labels are known and matched. Nonetheless, the label leakage and matching issue stands as an open problem for LLL attacks. + +Identifying which images are reconstructed correctly is also a challenging problem for LLL. For FedSGD, bins without activation can be identified, as after the subtraction process the reconstruction will be zero. However, if multiple images activate the same bin, the reconstruction is non-zero. By checking for non-zero values, we cannot directly identify whether the reconstruction is a single image or a combination of multiple. Manually filtering this can become a tedious process in practice. This is somewhat mitigated in the FedAvg case by LOKI. Across separate FedAvg local mini-batches, multiple images will not activate the same neuron, resulting in muchFewew overlapping reconstructions. Therefore the only images that can cause multiple activations are within the same mini-batch [30]. In this work, we manually filter out multiple image reconstructions in order to test the quality of the data. However, especially for LLL in FedSGD, the problem of how to automate + +
MNIST
Batch sizeModelTime per batch (s)Time for entire dataset (days)
8DNN54.794.76
1662.342.71
3264.321.40
CIFAR-10
Batch sizeModelTime per batch (s)Time for entire dataset (days)
4ResNet-18422.8161.17
8437.5431.65
16435.3315.75
+ +removal or how to use these multiple image reconstructions for training still remains an important item of future work. + +# 4. Experiments + +The goal of our work is to discuss the effectiveness of leaked data when used for the downstream task of image classification. We use centralized training on the entire benign dataset and federated learning (with only updates and no data shared by the clients) as the baselines for comparison. Typically, these serve as the upper bound and the lower bound respectively for the model accuracy. For federated learning, we use two settings of 10 and 50 clients. We use a non-IID bias of 0.5 and FedAvg with 3 local iterations. Additional training results for the IID setting and FedSGD are included in the supplementary. + +We use Inverting Gradients [6] as the gradient inversion (GI) attack. Another candidate is GradInversion, which is a SOTA GI attack (especially for ImageNet-sized images). However, we discarded this as the author code for this attack is not available and the third-party online implementations did not achieve comparable results to the original paper. For linear layer leakage (LLL), we use the LOKI attack [30]. We use LOKI as it achieves SOTA FedAvg leakage rate and equivalent SOTA leakage rate in FedSGD compared to Robbing the Fed. Model performance for other LLL methods [1, 5] can be found in the supplementary. + +We use MNIST [13], CIFAR-10 [11], and the Tiny ImageNet [12] datasets. For the MNIST dataset, we train using a simple DNN with 2 convolutional layers and 2 FC layers. For CIFAR-10 and Tiny Imagenet, we use a ResNet-18 [8]. We evaluate on the validation set for Tiny ImageNet because the test set is unlabeled. The plots only show a single run for each setting, but the (numerical) testing performance is reported as an average over 5 runs. We run all experiments on NVIDIA A100 80GB GPUs. + +# 4.1. Gradient inversion computation time + +Compared to LLL attacks, the computational overhead of GI attacks when reconstructing data, while done offline, is still non-negligible. In order to fully evaluate the leaked data, we ran the Inverting Gradients attack on the entire dataset of MNIST with batch sizes of 8, 16, and 32. We + +Table 1. Computation time for Inverting Gradients running on a NVIDIA A100 80GB GPU. Small batches have more total batches and take more time for the whole dataset. Computation time is significantly higher for CIFAR-10 given the more complex model. + +
CIFAR-10MNIST
Batch sizePSNRSSIMBatch sizePSNRSSIM
427.880.9067823.160.6996
820.770.72151619.150.6506
1615.940.39783215.780.4537
+ +(a) Gradient Inversion PSNR/SSIM + +
CIFAR-10MNISTTiny ImageNet
FC factorLeaked imagesPercentLeaked imagesPercentLeaked imagesPercent
84378887.585254887.588580485.80
43946478.934596676.617514975.15
22988259.763579559.665801258.01
11824236.482181536.363539335.39
+ +(b) Linear layer attack leakage rate + +Table 2. (a) Average PSNR↑/SSIM↑ scores of gradient inversion (Inverting Gradients) reconstructions for various batch sizes and (b) number of leaked images/leakage rate of linear layer leakage attack (LOKI) for several FC layer sizes. + +also ran over CIFAR-10 with batch sizes of 4, 8, and 16. We run 10,000 iterations of optimization for each batch. We evaluate the Inverting Gradients attack in the best case scenario with an untrained model where all labels are correctly recovered prior to optimization. + +Table 1 shows the amount of time taken by Inverting Gradients run on a NVIDIA A100 80GB GPU for MNIST and CIFAR-10 on a DNN and ResNet-18 respectively. There is some variance in the average amount of time based on batch size, but when attacking the entire dataset, smaller batch sizes always take longer time as the number of total batches is much larger. CIFAR-10 takes significantly longer to reconstruct each batch given the more complex model of ResNet-18 compared to a DNN. For a batch size of 4, CIFAR-10 takes 422.81 seconds per batch. Iterating across the entire dataset on a single GPU takes 61.17 days. + +# 4.2. Leakage statistics + +For GI we use Inverting Gradients [6], and for LLL we use LOKI [30]. For Inverting Grads, we use a batch size of 4, 8, and 16 on CIFAR-10 and 8, 16 and 32 on MNIST. For LOKI, we choose the FC layer size based on the batch size. We vary this with factors of 1, 2, 4 and 8 on CIFAR-10, MNIST, and Tiny ImageNet. We use a client batch size of 64 for all datasets, so an FC size factor of 2 would mean $64 \cdot 2 = 128$ units in the FC layer. + +Table 2a gives the average PSNR and SSIM metrics (where a higher score indicates greater similarity with the ground truth) for the reconstructions across the entire dataset of CIFAR-10 and MNIST from Inverting Gradients. Increasing the batch size results in a lower PSNR and SSIM score for both CIFAR-10 and MNIST. + +Table 2b reports the number of leaked images for LOKI. For an FC factor of 4, LOKI leaks $78.93\%$ , $76.61\%$ , and $75.15\%$ of images on the CIFAR-10, MNIST, and Tiny ImageNet datasets. Increasing the FC layer size increases the model size overhead, but results in a higher leakage rate as images are more likely to activate different neurons. Decreasing the size creates a smaller model size overhead but + +results in lower leakage rate. + +# 4.3. Training on leaked data from scratch + +We start with a discussion on the models that can be trained from the leaked data of both GI and LLL attacks. The attacks are applied under the FedSGD setting. We also start with assuming that all labels are known prior to training — this is a best case assumption for the downstream training. For CIFAR-10, we train the centralized models using SGD with momentum 0.9 and weight decay 0.0001. We use a batch size of 128 and train for 200 epochs starting with an initial learning rate of 0.1 and a scheduler that decreases on plateau. For MNIST, we use SGD without momentum or weight decay and train for 40 epochs with a LR of 0.01. For Tiny ImageNet we use Adam with learning rate 0.001 and train for 50 epochs with a batch size of 1024. For federated learning, we use 400 training rounds for CIFAR-10 and MNIST and 500 rounds for Tiny ImageNet. Client batch size of 32, 128, and 256 are used for MNIST, CIFAR-10, and Tiny ImageNet. The test accuracy is downsampled to fit on the same axis with centralized training. + +Table 3a and Figure 2a show the results for CIFAR-10 for centralized training, federated learning (FedAvg), and training with leaked data. Centralized training achieves a $94.38\%$ accuracy and federated learning achieves a $72.76\%$ and $68.71\%$ peak accuracy with 10 and 50 clients respectively. Centralized and FL indeed provide the upper bound and the lower bound of the accuracy here. Across all attack settings, GI and LLL achieve higher model accuracy compared to federated learning. The reconstruction quality of GI is adversely affected by larger batch sizes and in turn also impacts the final model performance. With a batch size of 4, 8, and 16, models trained on the leaked data from GI achieves $90.34\%$ , $86.16\%$ , and $76.83\%$ accuracy. With even larger batch sizes, it is likely that the model performance will drop below federated learning. The performance of models using LLL data is more stable. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the accuracy only drops from $93.16\%$ to $88.86\%$ . This is spite of a large drop in the number of leaked images from 43788 ( $87.58\%$ ) to 18242 ( $36.48\%$ ) as shown in Table 2b. This is a subtle and consequential result — the downstream training task can still be accomplished despite a big drop in the amount of data leakage. All prior work had stopped at the stage of evaluating the proportion of leakage and therefore had missed out on this insight. + +Table 3b and Figure 2b show the results for MNIST. LLL performs better than the FedAvg baseline in all cases. With an FC size factor of 8, LOKI achieves a $98.82\%$ test accuracy (nearly the same regardless of the FC size factor), only $0.07\%$ lower than centralized at $98.89\%$ . GI performs slightly worse than federated learning, achieving $95.96\%$ , $95.50\%$ , and $94.29\%$ at batch sizes 8, 16, and 32 compared to $96.17\%$ and $96.18\%$ with 10 and 50 clients in FedAvg. + +We believe this is because the noise in the reconstructions hampers model performance with GI. This affects MNIST more severely than for CIFAR because the MNIST images are more sparse and sensitive to noise. + +The top-1 validation accuracy for Tiny ImageNet are given in Table 3c and Figure 2c. The total number of leaked images hurts LLL more here than in the other two datasets. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the validation accuracy drops from $46.70\%$ to $35.20\%$ , an $11.5\%$ drop in performance. Federated learning achieves $37.00\%$ and $35.06\%$ accuracy with 10 and 50 clients respectively. FL with 10 clients has slightly better accuracy than LLL with an FC size factor of 1. However, any LLL setting with a larger FC layer achieves higher accuracy than both settings in federated learning. + +# 4.4. FedAvg with leaked images from LOKI + +We explore the impact of leaking images using LLL in FedAvg when training models. We use 8 local iterations of mini-batch size 8, a local dataset size of 64, and a learning rate of 1e-4. Following the LOKI attack [30], $\mathrm{CSF} = 500$ is used to achieve a higher leakage rate and reconstruction quality. Similar to before, we apply the attack across batches covering the entire CIFAR-10 training dataset. + +Table 4 shows the leakage rate of LOKI in FedAvg compared to FedSGD. Using $CSF = 500$ , the leakage rate of LOKI in FedAvg is substantially higher than the leakage rates in the FedSGD setting. Using an FC layer of half the size, the leakage rate in FedAvg is comparable to that of FedSGD. For example, LOKI FedAvg factor 1 achieves a leakage rate of $58.39\%$ compared to LOKI FedSGD factor 2 which achieves a leakage rate of $59.76\%$ . + +Table 5 shows the testing accuracy of LOKI FedAvg for the different FC layer sizes. The test accuracies are very comparable to the FedSGD settings. However, with a smaller FC layer size factor, the effect of having a larger total number of leaked images becomes visible. With FC factor 1, LOKI FedAvg achieves $91.11\%$ accuracy while LOKI FedSGD achieves $88.86\%$ . + +# 4.5. Training with semi-supervised learning + +In order to train machine learning models in a supervised fashion, images leaked through LLL need to be labeled. Previously, we assumed that all labels were known and matched to the images prior to training, but this is impractical. Within a LLL batch reconstruction, the order of images is not preserved. Even if labels were recovered like for GI attacks, this would result in a set of unmatched ground truth labels and images. Manual hand labeling is tedious and this inability to match labels is a weakness of LLL attacks only when it comes to downstream tasks. + +In this experiment, we assume that only a small portion of images are labeled. This setting is similar to semi-supervised learning (SSL) [25, 31] and we use CoMatch [14], a semi-supervised algorithm, to train the mod + +![](images/18955b8583f2939a32deb4eedb4008f0527f0d3b4d995ac7826ec7606e01c2d0.jpg) +(a) CIFAR-10 +Table 3. (a) CIFAR-10 and (b) MNIST test accuracy, (c) Tiny ImageNet top-1 validation accuracy trained from scratch with leaked data. CIFAR-10 and Tiny ImageNet are trained with a ResNet-18 and MNIST is trained with a DNN. For the two attacks (gradient inversion and linear layer leakage), training happens with leaked data. Second column indicates the number of clients in federated learning (FedAvg), the batch size for the gradient inversion attack, and the fully-connected layer size factor for linear layer leakage. Best accuracy is used for FedAvg and final accuracy for all other settings. + +![](images/b9c9ceb9aecf6de5784cf9f95866c9b17ecbb84294e6d4e903c40347399fe9cd.jpg) +(b) MNIST +Figure 2. Training models on (a) CIFAR-10, (b) MNIST, and (c) Tiny ImageNet with leaked data compared to centralized and federated learning training. Both linear layer leakage and gradient inversion achieve higher accuracy than the federated learning (FedAvg) baseline for CIFAR-10 in all cases. For MNIST, LLL (LOKI) nearly reaches centralized accuracy while GI performs slightly worse than FL. Top-1 validation accuracy used when training models on Tiny ImageNet. Here, LLL performs better with a FC layer size factor of 2 or higher. + +![](images/f73af5f6922e58ce7f7088b8c495d2a771907580156a0131deb309d8d33c90d1.jpg) +(c) Tiny ImageNet + +
Num. clients / Batch size / FC factorTest accuracy
Centralized-94.38
FedAvg1072.76
5068.71
Gradient inversion (Inverting grads [6])490.34
886.16
1676.83
893.16
Linear layer leakage (LOKI [30])492.94
291.90
188.86
+ +(a) CIFAR-10 + +
Num. clients / Batch size / FC factorTest accuracy
Centralized-98.89
FedAvg1096.17
5096.18
Gradient inversion (Inverting grads [6])895.96
1695.50
3294.29
898.82
Linear layer leakage (LOKI [30])498.70
298.72
198.46
+ +(b) MNIST + +
Num. clients / FC factorValidation accuracy
Centralized-48.55
FedAvg1037.00
5035.06
Linear layer leakage (LOKI [30])846.70
445.09
241.15
135.20
+ +(c) Tiny ImageNet + +
FC size factorLOKI FedSGDLOKI FedAvg
887.58 (43788)92.19 (46097)
478.93 (39464)85.98 (42989)
259.76 (29882)75.29 (37645)
136.48 (18242)58.39 (29196)
+ +Table 4. Leakage rate % (leaked images) of LOKI in FedSGD and FedAvg on CIFAR-10 for several FC layer sizes. FedAvg with half the FC layer size has comparable leakage rate to FedSGD. + +
FC factorFedSGD accuracyFedAvg accuracy
Linear layer leakage (LOKI [30])893.1693.31
492.9492.88
291.9092.35
188.8691.11
+ +Table 5. CIFAR-10 test accuracy with LOKI FedAvg. + +els. We use a WideResNet-28-2 [27] and train for 450 epochs on CIFAR-10. Figure 3a fixes the number of known labels to be 40 and shows the test accuracy curve with different fully connected layer sizes. We note that when there is only a small proportion of leaked data, this leads to a pronounced decrease in performance in SSL. LOKI with an FC factor of 1 (with $36.48\%$ leakage rate) illustrates this, as it only achieves $78.31\%$ max accuracy compared to the CIFAR-10 40 label baseline of $92.18\%$ . With a factor of 2, the accuracy improves and jumps to $88.91\%$ as the leak + +age rate also substantially increases to $59.76\%$ . With factors 4 and 8 (78.93% and 87.58% leakage rate), the accuracy is very close to the baseline at 91.18% and 91.84% respectively. Figure 3b shows the test accuracy plot of LOKI size factor 4 with the number of known labels being 20, 40, and 250. With 20 and 250 labels, LOKI factor 4 achieves 86.69% and 91.65% peak accuracy. We note that there is a greater variability in accuracy with a smaller number of labels, a trend consistent with other SSL methods [14, 25, 31]. + +# 4.6. Starting training from federated models + +While the previous experiments worked with leaked data from all batches in the dataset, such large amounts of leaked data may not always be available. For example, LLL attacks may affect model performance due to manipulation of the model parameters and/or architecture. In later rounds when model performance is high, this may be easier to detect. Similarly, GI attacks such as Inverting Gradients [6] have better reconstruction quality with an untrained network. In both cases, it is possible that the attack is only applied during a few initial rounds of training and only a small portion of data is leaked. Training from scratch with a small portion of leaked data does not achieve good model performance alone, so instead, we can initialize from the fully trained FL + +![](images/96a88cd287b7aa7e619fa2b9b30c5f1578bac4dc4720164d481efb25845c8989.jpg) +(a) Varying leakage rate + +![](images/4a19abd6be1601b4fe053ebee474da233da9f51943d5494b2499860b9f36c49f.jpg) +(b) Varying number of labels + +![](images/d48916e9d49d4a26cd1b8c08844cf60c755bd2b6a184520b246e91eda313b904.jpg) +Figure 3. Semi-supervised learning using CoMatch on CIFAR-10 with a WideResNet with (a) a varying FC size and leakage rate (LR) for LOKI and the known labels fixed at 40 and (b) a fixed number of leaked images and 20, 40, and 250 known labels. +Figure 4. FL model trained with 50 clients as start point for training with leaked LLL data. Highlighted area indicates improvement above the FL model and training on the leaked data alone. + +model and then train it centrally using the leaked data. + +We use smaller portions of the leaked data from LOKI FC size factor of 8 for this experiment. With only $5\%$ (2500) of leaked images in CIFAR-10, a model trained from scratch can only achieve $61.15\%$ test accuracy. With $2\%$ (1000 images) of data, the model performance drops to $39.63\%$ . Both cases are below an FL model with 50 clients that achieves a $64.08\%$ final test accuracy. However, by using the FL model as the initial starting point, even a small amount of leaked data allows for an increase in performance. We take the model that has been trained using FL and then additionally train that model centrally. Using only $2\%$ and $5\%$ of the leaked data, the models are able to achieve $68.25\%$ and $74.25\%$ final testing ac + +
PSNR% ims keptTest accuracy
>2015.8870.04
>1827.5971.52
>1640.7873.73
>1460.2275.48
>1284.4276.16
+ +(a) PSNR above threshold + +
PSNR% ims keptTest accuracy
< 2084.1273.03
< 1872.4167.09
< 1659.2263.71
< 1439.7858.29
< 1215.5845.05
+ +(b) PSNR below threshold + +Table 6. Training models using CIFAR-10 leaked data from inverting gradients batch size 16. Only the reconstructions with a PSNR (a) above and (b) below the threshold are used in training. + +curacy, a $28.62\%$ and $13.10\%$ improvement from starting from scratch. Figure 4 shows the improvement in test accuracy achieved by using the trained FL model as initialization instead of starting from scratch. The improvement is especially noticeable with small percentages of leaked data. + +# 4.7. Quality of data reconstruction + +Quality of reconstruction is an important concern for GI attacks. As the batch size continues to increase, the reconstruction quality decreases. Table 2a also shows this relationship. From the previous experiments, we also see a negative relationship between the increasing batch size and the usefulness of the leaked data in downstream model training. + +Here, we explore whether poor quality reconstructions are still useful for training models. Using the CIFAR-10 leaked dataset created using Inverting Gradients on a batch size of 16, we first sort the reconstructions by their PSNR value. Our first set of experiments removes the worst images from training. Table 6a shows the percent of images above the PSNR threshold and the final test accuracy when trained on that data. Compared to the baseline accuracy of $76.83\%$ achieved by using the entire leaked dataset, even removing the worst images with a $\mathrm{PSNR} < 12$ from the training set results in a small but non-zero performance loss. Then we take a different look at this question by now training on the worst images. Table 6b shows the test accuracy of models trained on all images below a PSNR threshold. Even when training on the lowest quality images with a PSNR $< 12$ , the model achieves a $45.05\%$ final testing accuracy, significantly worse than the baseline, but much above zero. + +From the previous results, we see that even leaked data with a low reconstruction quality is still useful to the model training process. This message is a new insight — all prior works when calculating the leak rate do not count poorly reconstructed images. We additionally include results using SSIM in the supplementary section. Visual examples of poor quality reconstructions used to train the models are also included in the supplement. + +# 4.8. Observations on reconstruction quality trends We additionally note a few trends manually observed when going through the GI batch reconstructions of CIFAR-10. + +While some reconstructions have very low quality, only images of the same label will be swapped within the batch during label matching. This is a desirable property for train + +![](images/037c55b704c969df021352cee8be54e08a1c7271bf868a3b7ca118b575e4e55a.jpg) + +![](images/cdd38955952bead9e00475752c68023aec498d6130665d635920cb69892334be.jpg) +Figure 5. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Ground truth images are on top and reconstructions are on the bottom. All labels in the batch are different and reconstructed images are high quality. +Figure 6. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Duplicate labels exist between images 0, 4, 7 and images 2, 3. Reconstruction 3 contains parts of the content from both images 2 and 3. Reconstructions 0, 4, and 7 appear almost as noise. However, the other images still have reasonable quality. + +ing on leaked data. As long as the set of labels is leaked correctly, the optimization process will match images to a correct label. So even though the image quality is poor (and can be entirely unrecognizable), since the label is correctly matched, such an image still does help in the training (this buttresses the message from Tables 6a and 6b). Naturally, as the PSNR goes down, the image helps less and less. + +We also note that when all labels are different within a batch, the reconstruction quality of the entire batch of images is almost always good. An example on CIFAR-10 is shown in Figure 5. Usually when the reconstructions are poor, it occurs between groups of images in the same batch that belong to the same class. Poor reconstructions can appear as almost entirely noise, but occasionally images can have visible pieces of each image from the same class in the batch stitched together. The example batch in Figure 6 depicts these properties of poorly reconstructed images. As a side note, we observe that a batch may have a mix of well reconstructed and poorly reconstructed images. We also note that the reverse mapping of the same label images to poor reconstructions is not true — having duplicate labels does not always imply that the reconstructions will be bad. + +# 5. Discussion + +As a main goal of our work, we have explored the effect of leaked data through today's leading data reconstruction attacks on downstream model training. Our main message is that it is important to also consider the use of data beyond reconstruction quality and image similarity. It is important to consider how far these leaked samples help in a downstream training task. From the experiments, we have seen that current reconstruction attacks are powerful, but still lack in several critical areas. For gradient inversion attacks, the reconstruction quality does pose an issue for + +training models. Another impediment to practical use of this attack type is the time and computational cost. + +Linear layer leakage attacks do not suffer from reconstruction quality issues and are much lighter weight computationally, but still suffer from other challenges. While LOKI has increased the efficiency of the FC layer leakage in FedAvg, there is still room for improvement (especially in FedSGD). As shown in [29], the model size overhead added by these attacks can be very large, especially if secure aggregation is used. Furthermore, the lack of labels poses a large challenge in training models on the data. With a small dataset like CIFAR-10, hand labeling 40 images is not too challenging. However, larger datasets will require more total labels. For example, SimMatch with $1\%$ of labels on ImageNet can achieve $67.2\%$ top-1 accuracy [31]. However, even $1\%$ of the total images in ImageNet-1k is 12,812 images. Even if the images were leaked, hand labeling would be extremely hard. Furthermore, we see from Figure 3a that having less total data and fewer labels results in an even steeper decrease in performance. + +While not discussed previously, there are many other aspects that affect data reconstruction attacks when considering training models. For example, how does the non-IID aspect of client data affect reconstruction? Other relevant settings of federated learning such as asynchronous federated learning, client selection, and differential privacy can also affect the final leaked dataset. Centralized training also allows for more complex and larger models without the communication or computational restrictions of clients in federated learning. These topics can be explored in future work. + +# 6. Conclusion + +We examine data reconstruction attacks through the lens of training models on the leaked data. While highlighting how the weaknesses of gradient inversion in terms of reconstruction quality affect downstream model training, we also show that even poorly reconstructed images are useful for training. We also discuss how the label matching problem for linear layer leakage can be mitigated through semisupervised learning. Under ideal conditions, we demonstrate that leaked data from both gradient inversion and linear layer leakage attacks are able to train powerful models comparable to even a centralized baseline. On CIFAR-10, gradient inversion and linear layer leakage attacks achieve $90.34\%$ and $93.16\%$ testing accuracy respectively, $17.58\%$ and $20.40\%$ higher than federated learning and only $4.04\%$ and $1.22\%$ lower compared to centralized training. + +Acknowledgements. This material is based in part upon work supported by Adobe Research, the Army Research Office under Contract number W911NF-2020-221, and the National Science Foundation under Grant Number CCF-1919197. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors. + +# References + +[1] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. 8th IEEE European Symposium on Security and Privacy (IEEE Euro S&P), 2023. 1, 2, 3, 4 +[2] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 3 +[3] Christopher A Choquette-Choo, Florian Tramer, Nicholas Carlini, and Nicolas Papernot. Label-only membership inference attacks. In International conference on machine learning, pages 1964-1974. PMLR, 2021. 1 +[4] Lixin Fan, Kam Woh Ng, Ce Ju, Tianyu Zhang, Chang Liu, Chee Seng Chan, and Qiang Yang. Rethinking privacy preserving deep learning: How to evaluate and thwart privacy attacks. In Federated Learning, pages 32-50. Springer, 2020. 2 +[5] Liam H Fowl, Jonas Geiping, Wojciech Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. In International Conference on Learning Representations, 2022. 1, 2, 3, 4 +[6] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33:16937-16947, 2020. 1, 2, 3, 4, 6 +[7] Jiahui Geng, Yongli Mou, Feifei Li, Qing Li, Oya Bayan, Stefan Decker, and Chunming Rong. Towards general deep leakage in federated learning. arXiv preprint arXiv:2110.09074, 2021. 2, 3 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[9] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603-618, 2017. 1 +[10] Yangsibo Huang, Samyak Gupta, Zhao Song, Kai Li, and Sanjeev Arora. Evaluating gradient inversion attacks and defenses in federated learning. Advances in Neural Information Processing Systems, 34:7232-7241, 2021. 1, 3 +[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 4 +[12] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 4 +[13] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 4 +[14] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza + +tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 5, 6 +[15] Xinjian Luo, Yuncheng Wu, Xiaokui Xiao, and Beng Chin Ooi. Feature inference attack on model predictions in vertical federated learning. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), pages 181-192. IEEE, 2021. 1 +[16] Kailang Ma, Yu Sun, Jian Cui, Dawei Li, Zhenyu Guan, and Jianwei Liu. Instance-wise batch label restoration via gradients in federated learning. In The Eleventh International Conference on Learning Representations, 2022. 2, 3 +[17] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1 +[18] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), pages 691-706. IEEE, 2019. 1 +[19] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In 2019 IEEE symposium on security and privacy (SP), pages 739-753. IEEE, 2019. 1 +[20] Dario Pasquini, Danilo Francati, and Giuseppe Ateniese. Eluding secure aggregation in federated learning via model inconsistency. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 2429-2443, 2022. 1 +[21] Le Trieu Phong, Yoshinori Aono, Takuya Hayashi, Lihua Wang, and Shiho Moriai. Privacy-preserving deep learning: Revisited and enhanced. In International Conference on Applications and Techniques in Information Security, pages 100-110. Springer, 2017. 2 +[22] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1 +[23] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE Conference on Computer Communications, pages 2512-2520. IEEE, 2019. 1 +[24] Yuxin Wen, Jonas Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. International Conference on Machine Learning, 2022. 1 +[25] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 5, 6 +[26] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337-16346, 2021. 1, 2, 3 + +[27] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6 +[28] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. arXiv preprint arXiv:2001.02610, 2020. 2 +[29] Joshua C Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. The resource problem of using linear layer leakage attack in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2023. 1, 2, 3, 8 +[30] Joshua Christian Zhao, Atul Sharma, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. Loki: Large-scale data reconstruction attack against federated learning through model manipulation. In 2024 IEEE Symposium on Security and Privacy (SP), pages 30-30. IEEE Computer Society, 2023. 1, 2, 3, 4, 5, 6 +[31] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 5, 6, 8 +[32] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 3 \ No newline at end of file diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/images.zip b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..64e0697e16e47175fdceaf8f88da79b2205bed82 --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04e08f5839974f9c446d78c31c77738591a0fc8c6809a9360aa72e8e7e550bd6 +size 410449 diff --git a/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/layout.json b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03dd848816f32a6b2ab6a45d77544c69951346b1 --- /dev/null +++ b/2024/Leak and Learn_ An Attacker's Cookbook to Train Using Leaked Data from Federated Learning/layout.json @@ -0,0 +1,7865 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 533, + 140 + ], + "type": "text", + "content": "Leak and Learn: An Attacker's Cookbook to Train Using Leaked Data from Federated Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 138, + 161, + 455, + 190 + ], + "type": "text", + "content": "Joshua C. Zhao, Ahaan Dabholkar, Atul Sharma, Saurabh Bagchi Purdue University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "type": "text", + "content": "{zhao1207,adabholk,sharm438,sbagchi}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 256, + 290, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 256, + 290, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 290, + 497 + ], + "type": "text", + "content": "Federated learning is a decentralized learning paradigm introduced to preserve privacy of client data. Despite this, prior work has shown that an attacker at the server can still reconstruct the private training data using only the client updates. These attacks are known as data reconstruction attacks and fall into two major categories: gradient inversion (GI) and linear layer leakage attacks (LLL). However, despite demonstrating the effectiveness of these attacks in breaching privacy, prior work has not investigated the usefulness of the reconstructed data for downstream tasks. In this work, we explore data reconstruction attacks through the lens of training and improving models with leaked data. We demonstrate the effectiveness of both GI and LLL attacks in maliciously training models using the leaked data more accurately than a benign federated learning strategy. Counter-intuitively, this bump in training quality can occur despite limited reconstruction quality or a small total number of leaked images. Finally, we show the limitations of these attacks for downstream training, individually for GI attacks and for LLL attacks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 128, + 518 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 677 + ], + "type": "text", + "content": "With growing concerns of data privacy, federated learning (FL) [17] has gained traction as a potential privacy-preserving method for training machine learning models. Compared to centralized learning where training is done on data localized at a central server, FL takes a decentralized approach where participating clients train a model on their local data and send their updates to the server. A typical training round involves a server sending a model to the clients, the clients training the model using their local data, and finally having the clients send their updates to the server for aggregation. However, the privacy-preserving property of FL only holds if the updates cannot be used to extract sensitive information about the local training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Despite only sending updates, prior works have shown the ability of attackers to gain information about the private training data through membership inference attacks [3," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 230, + 531, + 293 + ], + "blocks": [ + { + "bbox": [ + 318, + 230, + 531, + 293 + ], + "lines": [ + { + "bbox": [ + 318, + 230, + 531, + 293 + ], + "spans": [ + { + "bbox": [ + 318, + 230, + 531, + 293 + ], + "type": "image", + "image_path": "b3bea72727bea4569b30fade7e1446724c05af839d2023a8d61b20a1dc5f5d26.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 299, + 531, + 369 + ], + "blocks": [ + { + "bbox": [ + 319, + 299, + 531, + 369 + ], + "lines": [ + { + "bbox": [ + 319, + 299, + 531, + 369 + ], + "spans": [ + { + "bbox": [ + 319, + 299, + 531, + 369 + ], + "type": "image", + "image_path": "c4837e62ab0979054111d2495c3728b81b524954bf875568d00e077bde60f6fd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 371, + 529, + 456 + ], + "blocks": [ + { + "bbox": [ + 319, + 371, + 529, + 456 + ], + "lines": [ + { + "bbox": [ + 319, + 371, + 529, + 456 + ], + "spans": [ + { + "bbox": [ + 319, + 371, + 529, + 456 + ], + "type": "image", + "image_path": "2e2fe93a86039f39c7cc7966222d872c01e7c854198f4127cd54630fdaf9b1dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 460, + 494, + 471 + ], + "lines": [ + { + "bbox": [ + 356, + 460, + 494, + 471 + ], + "spans": [ + { + "bbox": [ + 356, + 460, + 494, + 471 + ], + "type": "text", + "content": "Figure 1. Training using leaked data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 473, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 473, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 473, + 545, + 557 + ], + "type": "text", + "content": "19, 22], property-inference attacks [15, 18], or GAN-based methods [9, 23]. Within these privacy attacks, data reconstruction has stood out as the most powerful, allowing attackers to directly recover the training data of clients [20, 24]. Within data reconstruction attacks, gradient inversion (optimization-based attacks) [6, 26, 32] and linear layer leakage [1, 5, 30] attacks stand out as the most common." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": "While prior works have discussed the power and the limitations of these attacks in the context of an attacker simply breaching privacy [10, 29], to the best of our knowledge, none has discussed the effectiveness of using the leaked data for downstream model training. Figure 1 shows this process of training using leaked data. With the onset of deep learning, data has become a valuable commodity. This growth in AI also means that the value of data no longer lies just in breaching privacy of the raw data, but in the ability of the data to be used downstream for creating powerful models. Therefore, one measure of the success of a data reconstruction attack should be whether the leaked data is useful for machine learning training, an aspect omitted in prior work." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12247" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "In order to discuss training on leaked data, we first introduce the attacks. Gradient inversion (GI) attacks typically reconstruct data through an iterative process where the distance between a gradient computed from a dummy image and the ground truth gradient is minimized by an optimizer. Label reconstruction, the process of recovering the batch labels from the client gradient, also plays a critical role in affecting the final reconstruction quality. These labels are typically leaked prior to optimization from the gradients of the final layer of the network [7, 16, 26, 28] and are matched with reconstructed images during the optimization. These attacks have shown success with individual updates of smaller batch sizes, up to 48 on ImageNet [26] or 100 on CIFAR [6]. However, this only considers breaching privacy, as a batch size of 48 on ImageNet or 100 on CIFAR result in low quality reconstructions and/or a very small number of identifiable images. Even for a modest batch size of 8 or 16 on CIFAR-10, some batch reconstructions already fail. Given the importance of data quality for training models, a sufficient condition for success is no longer if any images are identifiable, but rather the overall usefulness of the leaked data for the downstream training task. Here, even reconstructions with low similarity scores to the ground truth may still positively contribute to training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 289, + 602 + ], + "type": "text", + "content": "Compared to GI, linear layer leakage (LLL) does not suffer from problems with reconstruction quality. Using modification of a fully-connected (linear, FC) layer, an attacker can directly recover a proportion of the images of a client batch [1, 5, 30]. Reconstructed images are done by solving a simple linear equation and have little computation overhead. However, these attacks typically require modification of the model architecture and can add a large overhead in terms of model size [29]. While data quality is not an issue, when using the leaked data for training models, labeling the data becomes a problem. Unlike the optimization attacks, even if labels are known prior to reconstruction, the LLL process does not match them to the leaked images. As a result, after reconstruction there would be a set of leaked images and labels, where the images are not matched with their corresponding label. Furthermore, because only a proportion of images in the given batch are leaked for LLL, even with a full set of labels, some labels will not have a leaked image to go with. This can lead to a tedious process of manually labeling leaked images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 288, + 715 + ], + "type": "text", + "content": "Viewing data reconstruction attacks through the lens of training models brings up many questions. How do models trained with leaked images compare to centralized or federated learning? Does the reconstruction quality impact how well models perform when trained on leaked data? How does the lack of label matching during reconstruction affect linear layer leakage? In this work we address these largely unanswered questions and highlight the problems that arise in this new setting. Our contributions go as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 72, + 547, + 323 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": "- We demonstrate the effectiveness of training models using leaked data compared to federated learning and the centralized baselines. Using the data from linear layer leakage and optimization, models trained on leaked data can achieve " + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "20.40\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "17.58\\%" + }, + { + "bbox": [ + 306, + 72, + 545, + 144 + ], + "type": "text", + "content": " higher accuracy on CIFAR-10 compared to federated learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": "- We show that similarity metrics such as PSNR do not show if reconstructed images are useful for training, as even some of the worst images improve models. Training on CIFAR-10 with bad reconstructions generated by Inverting Gradients (batch size 16) with PSNR " + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "inline_equation", + "content": "< 14" + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": " results in a " + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "inline_equation", + "content": "58.29\\%" + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": " test accuracy. Removing these images and training only on images " + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "inline_equation", + "content": ">14" + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": " PSNR results in only a minor accuracy decrease from " + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "inline_equation", + "content": "76.83\\%" + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "inline_equation", + "content": "75.48\\%" + }, + { + "bbox": [ + 306, + 144, + 547, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "text", + "content": "- We quantify the effects and highlight the crucial limitations of current data reconstruction attacks when viewed through the lens of downstream model training. While gradient inversion can breach privacy for a batch size of 100 on CIFAR-10, this is impractical for training as even increasing the batch size from 4 to 16 results in a performance drop from " + }, + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "inline_equation", + "content": "90.34\\%" + }, + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "inline_equation", + "content": "76.83\\%" + }, + { + "bbox": [ + 306, + 239, + 547, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 330, + 391, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 391, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 391, + 342 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 345, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 429 + ], + "type": "text", + "content": "Gradient inversion (GI) attacks operate under an honest-but-curious threat model where a server only knows the update and model. The method involves feeding a dummy image into a model and computing the subsequent gradient. The dummy image starts out initialized as random noise, and an optimizer iteratively minimizes the distance between the computed gradient and the ground truth client gradient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 434, + 545, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 434, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 347, + 434, + 545, + 449 + ], + "type": "interline_equation", + "content": "x ^ {*} = \\arg \\min _ {x} \\left\\| \\nabla L (x, y, \\theta) - \\nabla W \\right\\| _ {2} \\tag {1}", + "image_path": "f4812b08f2c9f8eb8b8865169126bfeb5bc25e9c9a77493ab3cf621f10d31a2d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "content": "The intuition behind these attacks is that a similar computed dummy gradient " + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\nabla L(x,y,\\theta)" + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "content": " and ground truth gradient " + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "inline_equation", + "content": "\\nabla W" + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "content": " will result from similar images. Thus, as the gradients become closer, the final reconstructed image " + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "inline_equation", + "content": "x^{*}" + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "content": " should become close to the ground truth training image. The label " + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 453, + 545, + 597 + ], + "type": "text", + "content": " is assumed to be known prior to optimization. For a batch reconstruction with multiple labels, the labels are matched with the reconstructed images during optimization (as the loss and gradient are computed with these image-label pairs). Regularizers have also been introduced to assist the optimization process [6, 26]. However, these can result in reconstructed image artifacts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 597, + 547, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 547, + 634 + ], + "type": "text", + "content": "Linear layer leakage (LLL) attacks are built on the observation that a fully-connected (FC) layer leaks the input to the layer through the gradients [4, 21]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 390, + 637, + 545, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 637, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 390, + 637, + 545, + 662 + ], + "type": "interline_equation", + "content": "x ^ {i} = \\frac {\\delta L}{\\delta W ^ {i}} / \\frac {\\delta L}{\\delta B ^ {i}} \\tag {2}", + "image_path": "b8b58da5107e1b9d85661ef81486d22ebd3dcd98fa0ec530c659d4e6d500ff83.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\delta L}{\\delta W^i}" + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "content": " is the weight gradient and " + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\delta L}{\\delta B^i}" + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "content": " is the bias gradient of a neuron. Neuron " + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "content": " is activated by an image and " + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 304, + 664, + 547, + 715 + ], + "type": "text", + "content": " is the respective reconstructed image. However, this assumes that only a single image activated the neuron. If multiple" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12248" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "images activate the same neuron, reconstructed image " + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": " becomes a combination of the images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 228 + ], + "type": "text", + "content": "To mitigate this, the methods of trap weights [1] and Robbing the Fed [5] were proposed. Trap weights initialize the weights of the FC layer randomly as half positive and half negative, where the negative parameters are sampled from a slightly larger magnitude. This method has difficulty with scaling [30] compared to Robbing the Fed, which proposed a method where the FC layer weights are used to measure a property of the images (e.g., average pixel brightness). The biases of the layer were then set to fit the distribution of the dataset and a ReLU activation was used to threshold the activation. Images were reconstructed as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 231, + 286, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 231, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 81, + 231, + 286, + 255 + ], + "type": "interline_equation", + "content": "x ^ {i} = \\left(\\frac {\\delta L}{\\delta W ^ {i}} - \\frac {\\delta L}{\\delta W ^ {i + 1}}\\right) / \\left(\\frac {\\delta L}{\\delta B ^ {i}} - \\frac {\\delta L}{\\delta B ^ {i + 1}}\\right) \\tag {3}", + "image_path": "245b1c6ba9e178783185e0aed26183d93dd80566e76dd88f203ff8fff6de45df.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 286, + 305 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 258, + 286, + 305 + ], + "type": "inline_equation", + "content": "i + 1" + }, + { + "bbox": [ + 46, + 258, + 286, + 305 + ], + "type": "text", + "content": " indicates the neuron with a bias being the next highest cutoff. This method requires basic knowledge about the input data distribution, but achieves a higher leakage rate compared to trap weights and is scalable to aggregation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 306, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 286, + 449 + ], + "type": "text", + "content": "In the case of FedAvg, the sparse variant of Robbing the Fed utilizes a two-sided activation function (such as Hard tanh) and weight/bias scaling to maintain the same leakage as in FedSGD. However, the sparse variant is not scalable and leads to precision problems when attacking secure aggregation [2] or larger batches. The LOKI attack addresses the scalability problems of both methods through the introduction of a convolutional layer that splits the scaling between number of clients and batch size in secure aggregation [30]. They further introduce a convolutional scaling factor (CSF) which achieves a higher leakage rate in FedAvg without increasing the FC layer size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 286, + 628 + ], + "type": "text", + "content": "While there have been many works exploring data reconstruction, prior work has not evaluated these attacks in the context of downstream tasks. The works on GI [6, 26, 32] and LLL [1, 5, 30] attacks only discuss the reconstruction quality of the methods in terms of standard image metrics such as PSNR (peak signal-to-noise ratio), SSIM (structural similarity index measure), or LPIPS (learned perceptual image patch similarity). However, these metrics only measure how similar the reconstruction are to the ground truth and only give a vague sense of how useful the data is for training. Another work discussed the limitations of GI attacks [10] measured by image similarity. In [29], LLL attacks were measured in terms of the resource overhead added by the attacks. However, these works also did not discuss the usefulness of reconstructed data post-leakage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "This work aims to bridge the gap and explore the important facet of training machine learning models on leaked data from reconstruction attacks. Viewing leaked data beyond privacy breach brings to light key weaknesses of current attacks in generating useful data. While LLL reconstructs high quality images, the other important half of data, namely having matching labels, is missing. GI attacks can" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 144 + ], + "type": "text", + "content": "breach privacy for larger batch sizes, but the quality of the leaked data would be nearly useless for training models. In the age of machine learning and artificial intelligence, the value of data has increased tremendously. Therefore, when looking at leaked data, it is important to measure success in terms of usefulness to model training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 149, + 442, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 149, + 442, + 162 + ], + "spans": [ + { + "bbox": [ + 305, + 149, + 442, + 162 + ], + "type": "text", + "content": "3. Training on leaked data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 163, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 163, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 163, + 545, + 295 + ], + "type": "text", + "content": "Data reconstruction attacks have largely focused on the quality of the recovered image in gradient inversion (GI) or the efficiency of linear layer leakage (LLL). However, label reconstruction is important for training on leaked data. Within the pipeline of GI attacks, labels are typically recovered prior to reconstruction, as they assist in the optimization problem [10]. Recent works have shown that label restoration with large batches and duplicate labels is possible [7, 16]. More accurate labels and the added flexibility of duplicate class images directly improves prior GI attacks such as Inverting Gradients [6] and GradInversion [26]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 295, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 295, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 295, + 545, + 510 + ], + "type": "text", + "content": "Interestingly, one strength of LLL has been in the ability to reconstruct images without knowledge of the data labels. While the class labels can affect the distribution of the client dataset (e.g., average image brightness can be different between classes) [30], the labels themselves do not help in the reconstruction. Knowing labels helps guide GI attacks (as gradients are computed from data and label pairs), but for LLL, images activating the neurons are reconstructed using only the weight and bias gradients of the FC layer as with Equation 3. Labels are necessary to train models using leaked data, but even with a set of recovered labels, the leaked images would still remain unmatched. Unlike GI which matches during optimization, images leaked by LLL would need to be manually matched. For this work, we experiment with cases when all labels are known and matched and alternatively, when only a small portion of labels are known and matched. Nonetheless, the label leakage and matching issue stands as an open problem for LLL attacks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Identifying which images are reconstructed correctly is also a challenging problem for LLL. For FedSGD, bins without activation can be identified, as after the subtraction process the reconstruction will be zero. However, if multiple images activate the same bin, the reconstruction is non-zero. By checking for non-zero values, we cannot directly identify whether the reconstruction is a single image or a combination of multiple. Manually filtering this can become a tedious process in practice. This is somewhat mitigated in the FedAvg case by LOKI. Across separate FedAvg local mini-batches, multiple images will not activate the same neuron, resulting in muchFewew overlapping reconstructions. Therefore the only images that can cause multiple activations are within the same mini-batch [30]. In this work, we manually filter out multiple image reconstructions in order to test the quality of the data. However, especially for LLL in FedSGD, the problem of how to automate" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12249" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 255, + 170 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 255, + 170 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 255, + 170 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 255, + 170 + ], + "type": "table", + "html": "
MNIST
Batch sizeModelTime per batch (s)Time for entire dataset (days)
8DNN54.794.76
1662.342.71
3264.321.40
CIFAR-10
Batch sizeModelTime per batch (s)Time for entire dataset (days)
4ResNet-18422.8161.17
8437.5431.65
16435.3315.75
", + "image_path": "434c327510eddc12822a5fa5b8bd324c52e96d1482f5ee27d7c7c68f737cbf6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 251 + ], + "type": "text", + "content": "removal or how to use these multiple image reconstructions for training still remains an important item of future work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 253, + 128, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 128, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 128, + 266 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 287, + 399 + ], + "type": "text", + "content": "The goal of our work is to discuss the effectiveness of leaked data when used for the downstream task of image classification. We use centralized training on the entire benign dataset and federated learning (with only updates and no data shared by the clients) as the baselines for comparison. Typically, these serve as the upper bound and the lower bound respectively for the model accuracy. For federated learning, we use two settings of 10 and 50 clients. We use a non-IID bias of 0.5 and FedAvg with 3 local iterations. Additional training results for the IID setting and FedSGD are included in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 399, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 530 + ], + "type": "text", + "content": "We use Inverting Gradients [6] as the gradient inversion (GI) attack. Another candidate is GradInversion, which is a SOTA GI attack (especially for ImageNet-sized images). However, we discarded this as the author code for this attack is not available and the third-party online implementations did not achieve comparable results to the original paper. For linear layer leakage (LLL), we use the LOKI attack [30]. We use LOKI as it achieves SOTA FedAvg leakage rate and equivalent SOTA leakage rate in FedSGD compared to Robbing the Fed. Model performance for other LLL methods [1, 5] can be found in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 531, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 638 + ], + "type": "text", + "content": "We use MNIST [13], CIFAR-10 [11], and the Tiny ImageNet [12] datasets. For the MNIST dataset, we train using a simple DNN with 2 convolutional layers and 2 FC layers. For CIFAR-10 and Tiny Imagenet, we use a ResNet-18 [8]. We evaluate on the validation set for Tiny ImageNet because the test set is unlabeled. The plots only show a single run for each setting, but the (numerical) testing performance is reported as an average over 5 runs. We run all experiments on NVIDIA A100 80GB GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 641, + 243, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 641, + 243, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 641, + 243, + 653 + ], + "type": "text", + "content": "4.1. Gradient inversion computation time" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "Compared to LLL attacks, the computational overhead of GI attacks when reconstructing data, while done offline, is still non-negligible. In order to fully evaluate the leaked data, we ran the Inverting Gradients attack on the entire dataset of MNIST with batch sizes of 8, 16, and 32. We" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 324, + 70, + 529, + 114 + ], + "blocks": [ + { + "bbox": [ + 46, + 175, + 287, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 220 + ], + "type": "text", + "content": "Table 1. Computation time for Inverting Gradients running on a NVIDIA A100 80GB GPU. Small batches have more total batches and take more time for the whole dataset. Computation time is significantly higher for CIFAR-10 given the more complex model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 70, + 529, + 114 + ], + "lines": [ + { + "bbox": [ + 324, + 70, + 529, + 114 + ], + "spans": [ + { + "bbox": [ + 324, + 70, + 529, + 114 + ], + "type": "table", + "html": "
CIFAR-10MNIST
Batch sizePSNRSSIMBatch sizePSNRSSIM
427.880.9067823.160.6996
820.770.72151619.150.6506
1615.940.39783215.780.4537
", + "image_path": "df84ec2c4b1f6ed3f71b29b828d27c248bd6340b02c88e24a9fdb0e90c73aaea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 129, + 539, + 189 + ], + "blocks": [ + { + "bbox": [ + 368, + 117, + 484, + 125 + ], + "lines": [ + { + "bbox": [ + 368, + 117, + 484, + 125 + ], + "spans": [ + { + "bbox": [ + 368, + 117, + 484, + 125 + ], + "type": "text", + "content": "(a) Gradient Inversion PSNR/SSIM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 129, + 539, + 189 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 539, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 539, + 189 + ], + "type": "table", + "html": "
CIFAR-10MNISTTiny ImageNet
FC factorLeaked imagesPercentLeaked imagesPercentLeaked imagesPercent
84378887.585254887.588580485.80
43946478.934596676.617514975.15
22988259.763579559.665801258.01
11824236.482181536.363539335.39
", + "image_path": "2eedcc4b4e7b2f4f8c223f6ec355c0b36aab3d269ebcd33088e21c5aa8d02082.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 192, + 483, + 202 + ], + "lines": [ + { + "bbox": [ + 369, + 192, + 483, + 202 + ], + "spans": [ + { + "bbox": [ + 369, + 192, + 483, + 202 + ], + "type": "text", + "content": "(b) Linear layer attack leakage rate" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 204, + 545, + 248 + ], + "lines": [ + { + "bbox": [ + 305, + 204, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 305, + 204, + 545, + 248 + ], + "type": "text", + "content": "Table 2. (a) Average PSNR↑/SSIM↑ scores of gradient inversion (Inverting Gradients) reconstructions for various batch sizes and (b) number of leaked images/leakage rate of linear layer leakage attack (LOKI) for several FC layer sizes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 312 + ], + "type": "text", + "content": "also ran over CIFAR-10 with batch sizes of 4, 8, and 16. We run 10,000 iterations of optimization for each batch. We evaluate the Inverting Gradients attack in the best case scenario with an untrained model where all labels are correctly recovered prior to optimization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 445 + ], + "type": "text", + "content": "Table 1 shows the amount of time taken by Inverting Gradients run on a NVIDIA A100 80GB GPU for MNIST and CIFAR-10 on a DNN and ResNet-18 respectively. There is some variance in the average amount of time based on batch size, but when attacking the entire dataset, smaller batch sizes always take longer time as the number of total batches is much larger. CIFAR-10 takes significantly longer to reconstruct each batch given the more complex model of ResNet-18 compared to a DNN. For a batch size of 4, CIFAR-10 takes 422.81 seconds per batch. Iterating across the entire dataset on a single GPU takes 61.17 days." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 449, + 411, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 411, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 411, + 460 + ], + "type": "text", + "content": "4.2. Leakage statistics" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "content": "For GI we use Inverting Gradients [6], and for LLL we use LOKI [30]. For Inverting Grads, we use a batch size of 4, 8, and 16 on CIFAR-10 and 8, 16 and 32 on MNIST. For LOKI, we choose the FC layer size based on the batch size. We vary this with factors of 1, 2, 4 and 8 on CIFAR-10, MNIST, and Tiny ImageNet. We use a client batch size of 64 for all datasets, so an FC size factor of 2 would mean " + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "inline_equation", + "content": "64 \\cdot 2 = 128" + }, + { + "bbox": [ + 304, + 461, + 545, + 556 + ], + "type": "text", + "content": " units in the FC layer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 628 + ], + "type": "text", + "content": "Table 2a gives the average PSNR and SSIM metrics (where a higher score indicates greater similarity with the ground truth) for the reconstructions across the entire dataset of CIFAR-10 and MNIST from Inverting Gradients. Increasing the batch size results in a lower PSNR and SSIM score for both CIFAR-10 and MNIST." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Table 2b reports the number of leaked images for LOKI. For an FC factor of 4, LOKI leaks " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "78.93\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "76.61\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "75.15\\%" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " of images on the CIFAR-10, MNIST, and Tiny ImageNet datasets. Increasing the FC layer size increases the model size overhead, but results in a higher leakage rate as images are more likely to activate different neurons. Decreasing the size creates a smaller model size overhead but" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12250" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 165, + 83 + ], + "type": "text", + "content": "results in lower leakage rate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 88, + 243, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 88, + 243, + 100 + ], + "spans": [ + { + "bbox": [ + 47, + 88, + 243, + 100 + ], + "type": "text", + "content": "4.3. Training on leaked data from scratch" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 102, + 287, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 102, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 287, + 316 + ], + "type": "text", + "content": "We start with a discussion on the models that can be trained from the leaked data of both GI and LLL attacks. The attacks are applied under the FedSGD setting. We also start with assuming that all labels are known prior to training — this is a best case assumption for the downstream training. For CIFAR-10, we train the centralized models using SGD with momentum 0.9 and weight decay 0.0001. We use a batch size of 128 and train for 200 epochs starting with an initial learning rate of 0.1 and a scheduler that decreases on plateau. For MNIST, we use SGD without momentum or weight decay and train for 40 epochs with a LR of 0.01. For Tiny ImageNet we use Adam with learning rate 0.001 and train for 50 epochs with a batch size of 1024. For federated learning, we use 400 training rounds for CIFAR-10 and MNIST and 500 rounds for Tiny ImageNet. Client batch size of 32, 128, and 256 are used for MNIST, CIFAR-10, and Tiny ImageNet. The test accuracy is downsampled to fit on the same axis with centralized training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": "Table 3a and Figure 2a show the results for CIFAR-10 for centralized training, federated learning (FedAvg), and training with leaked data. Centralized training achieves a " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "94.38\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": " accuracy and federated learning achieves a " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "72.76\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "68.71\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": " peak accuracy with 10 and 50 clients respectively. Centralized and FL indeed provide the upper bound and the lower bound of the accuracy here. Across all attack settings, GI and LLL achieve higher model accuracy compared to federated learning. The reconstruction quality of GI is adversely affected by larger batch sizes and in turn also impacts the final model performance. With a batch size of 4, 8, and 16, models trained on the leaked data from GI achieves " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "90.34\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "86.16\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "76.83\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": " accuracy. With even larger batch sizes, it is likely that the model performance will drop below federated learning. The performance of models using LLL data is more stable. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the accuracy only drops from " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "93.16\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "88.86\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": ". This is spite of a large drop in the number of leaked images from 43788 (" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "87.58\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": ") to 18242 (" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "inline_equation", + "content": "36.48\\%" + }, + { + "bbox": [ + 47, + 318, + 287, + 616 + ], + "type": "text", + "content": ") as shown in Table 2b. This is a subtle and consequential result — the downstream training task can still be accomplished despite a big drop in the amount of data leakage. All prior work had stopped at the stage of evaluating the proportion of leakage and therefore had missed out on this insight." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Table 3b and Figure 2b show the results for MNIST. LLL performs better than the FedAvg baseline in all cases. With an FC size factor of 8, LOKI achieves a " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "98.82\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " test accuracy (nearly the same regardless of the FC size factor), only " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "0.07\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " lower than centralized at " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "98.89\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ". GI performs slightly worse than federated learning, achieving " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "95.96\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "95.50\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "94.29\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " at batch sizes 8, 16, and 32 compared to " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "96.17\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "96.18\\%" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " with 10 and 50 clients in FedAvg." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "We believe this is because the noise in the reconstructions hampers model performance with GI. This affects MNIST more severely than for CIFAR because the MNIST images are more sparse and sensitive to noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": "The top-1 validation accuracy for Tiny ImageNet are given in Table 3c and Figure 2c. The total number of leaked images hurts LLL more here than in the other two datasets. Between an FC layer size of 512 (factor 8) and 64 (factor 1), the validation accuracy drops from " + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "inline_equation", + "content": "46.70\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "inline_equation", + "content": "35.20\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": ", an " + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "inline_equation", + "content": "11.5\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": " drop in performance. Federated learning achieves " + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "inline_equation", + "content": "37.00\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "inline_equation", + "content": "35.06\\%" + }, + { + "bbox": [ + 304, + 121, + 546, + 251 + ], + "type": "text", + "content": " accuracy with 10 and 50 clients respectively. FL with 10 clients has slightly better accuracy than LLL with an FC size factor of 1. However, any LLL setting with a larger FC layer achieves higher accuracy than both settings in federated learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 254, + 512, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 512, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 512, + 266 + ], + "type": "text", + "content": "4.4. FedAvg with leaked images from LOKI" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 267, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 350 + ], + "type": "text", + "content": "We explore the impact of leaking images using LLL in FedAvg when training models. We use 8 local iterations of mini-batch size 8, a local dataset size of 64, and a learning rate of 1e-4. Following the LOKI attack [30], " + }, + { + "bbox": [ + 304, + 267, + 545, + 350 + ], + "type": "inline_equation", + "content": "\\mathrm{CSF} = 500" + }, + { + "bbox": [ + 304, + 267, + 545, + 350 + ], + "type": "text", + "content": " is used to achieve a higher leakage rate and reconstruction quality. Similar to before, we apply the attack across batches covering the entire CIFAR-10 training dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "text", + "content": "Table 4 shows the leakage rate of LOKI in FedAvg compared to FedSGD. Using " + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "inline_equation", + "content": "CSF = 500" + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "text", + "content": ", the leakage rate of LOKI in FedAvg is substantially higher than the leakage rates in the FedSGD setting. Using an FC layer of half the size, the leakage rate in FedAvg is comparable to that of FedSGD. For example, LOKI FedAvg factor 1 achieves a leakage rate of " + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "inline_equation", + "content": "58.39\\%" + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "text", + "content": " compared to LOKI FedSGD factor 2 which achieves a leakage rate of " + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "inline_equation", + "content": "59.76\\%" + }, + { + "bbox": [ + 304, + 351, + 546, + 446 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "text", + "content": "Table 5 shows the testing accuracy of LOKI FedAvg for the different FC layer sizes. The test accuracies are very comparable to the FedSGD settings. However, with a smaller FC layer size factor, the effect of having a larger total number of leaked images becomes visible. With FC factor 1, LOKI FedAvg achieves " + }, + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "inline_equation", + "content": "91.11\\%" + }, + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "text", + "content": " accuracy while LOKI FedSGD achieves " + }, + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "inline_equation", + "content": "88.86\\%" + }, + { + "bbox": [ + 304, + 447, + 545, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 533, + 512, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 512, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 512, + 545 + ], + "type": "text", + "content": "4.5. Training with semi-supervised learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 664 + ], + "type": "text", + "content": "In order to train machine learning models in a supervised fashion, images leaked through LLL need to be labeled. Previously, we assumed that all labels were known and matched to the images prior to training, but this is impractical. Within a LLL batch reconstruction, the order of images is not preserved. Even if labels were recovered like for GI attacks, this would result in a set of unmatched ground truth labels and images. Manual hand labeling is tedious and this inability to match labels is a weakness of LLL attacks only when it comes to downstream tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "In this experiment, we assume that only a small portion of images are labeled. This setting is similar to semi-supervised learning (SSL) [25, 31] and we use CoMatch [14], a semi-supervised algorithm, to train the mod" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12251" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 211, + 177 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 211, + 177 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 211, + 177 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 211, + 177 + ], + "type": "image", + "image_path": "18955b8583f2939a32deb4eedb4008f0527f0d3b4d995ac7826ec7606e01c2d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 108, + 181, + 154, + 191 + ], + "lines": [ + { + "bbox": [ + 108, + 181, + 154, + 191 + ], + "spans": [ + { + "bbox": [ + 108, + 181, + 154, + 191 + ], + "type": "text", + "content": "(a) CIFAR-10" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 363, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 46, + 363, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 547, + 418 + ], + "type": "text", + "content": "Table 3. (a) CIFAR-10 and (b) MNIST test accuracy, (c) Tiny ImageNet top-1 validation accuracy trained from scratch with leaked data. CIFAR-10 and Tiny ImageNet are trained with a ResNet-18 and MNIST is trained with a DNN. For the two attacks (gradient inversion and linear layer leakage), training happens with leaked data. Second column indicates the number of clients in federated learning (FedAvg), the batch size for the gradient inversion attack, and the fully-connected layer size factor for linear layer leakage. Best accuracy is used for FedAvg and final accuracy for all other settings." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 217, + 72, + 375, + 176 + ], + "blocks": [ + { + "bbox": [ + 217, + 72, + 375, + 176 + ], + "lines": [ + { + "bbox": [ + 217, + 72, + 375, + 176 + ], + "spans": [ + { + "bbox": [ + 217, + 72, + 375, + 176 + ], + "type": "image", + "image_path": "b9c9ceb9aecf6de5784cf9f95866c9b17ecbb84294e6d4e903c40347399fe9cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 181, + 316, + 191 + ], + "lines": [ + { + "bbox": [ + 278, + 181, + 316, + 191 + ], + "spans": [ + { + "bbox": [ + 278, + 181, + 316, + 191 + ], + "type": "text", + "content": "(b) MNIST" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 196, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 196, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 547, + 240 + ], + "type": "text", + "content": "Figure 2. Training models on (a) CIFAR-10, (b) MNIST, and (c) Tiny ImageNet with leaked data compared to centralized and federated learning training. Both linear layer leakage and gradient inversion achieve higher accuracy than the federated learning (FedAvg) baseline for CIFAR-10 in all cases. For MNIST, LLL (LOKI) nearly reaches centralized accuracy while GI performs slightly worse than FL. Top-1 validation accuracy used when training models on Tiny ImageNet. Here, LLL performs better with a FC layer size factor of 2 or higher." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 383, + 72, + 542, + 176 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 542, + 176 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 542, + 176 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 542, + 176 + ], + "type": "image", + "image_path": "f73af5f6922e58ce7f7088b8c495d2a771907580156a0131deb309d8d33c90d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 181, + 493, + 191 + ], + "lines": [ + { + "bbox": [ + 432, + 181, + 493, + 191 + ], + "spans": [ + { + "bbox": [ + 432, + 181, + 493, + 191 + ], + "type": "text", + "content": "(c) Tiny ImageNet" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 51, + 243, + 211, + 351 + ], + "blocks": [ + { + "bbox": [ + 51, + 243, + 211, + 351 + ], + "lines": [ + { + "bbox": [ + 51, + 243, + 211, + 351 + ], + "spans": [ + { + "bbox": [ + 51, + 243, + 211, + 351 + ], + "type": "table", + "html": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-94.38
FedAvg1072.76
5068.71
Gradient inversion (Inverting grads [6])490.34
886.16
1676.83
893.16
Linear layer leakage (LOKI [30])492.94
291.90
188.86
", + "image_path": "98f10f62088a826fbe7c4f02f84413b24574b64dd8c32cd0a78d982710907761.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 217, + 243, + 377, + 351 + ], + "blocks": [ + { + "bbox": [ + 108, + 354, + 154, + 362 + ], + "lines": [ + { + "bbox": [ + 108, + 354, + 154, + 362 + ], + "spans": [ + { + "bbox": [ + 108, + 354, + 154, + 362 + ], + "type": "text", + "content": "(a) CIFAR-10" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 217, + 243, + 377, + 351 + ], + "lines": [ + { + "bbox": [ + 217, + 243, + 377, + 351 + ], + "spans": [ + { + "bbox": [ + 217, + 243, + 377, + 351 + ], + "type": "table", + "html": "
Num. clients / Batch size / FC factorTest accuracy
Centralized-98.89
FedAvg1096.17
5096.18
Gradient inversion (Inverting grads [6])895.96
1695.50
3294.29
898.82
Linear layer leakage (LOKI [30])498.70
298.72
198.46
", + "image_path": "e5676606207bc897a7eae3fb4dc140f48e4e5d4152044c53af12f04ec1ec28ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 381, + 258, + 545, + 335 + ], + "blocks": [ + { + "bbox": [ + 278, + 354, + 316, + 362 + ], + "lines": [ + { + "bbox": [ + 278, + 354, + 316, + 362 + ], + "spans": [ + { + "bbox": [ + 278, + 354, + 316, + 362 + ], + "type": "text", + "content": "(b) MNIST" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 381, + 258, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 381, + 258, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 381, + 258, + 545, + 335 + ], + "type": "table", + "html": "
Num. clients / FC factorValidation accuracy
Centralized-48.55
FedAvg1037.00
5035.06
Linear layer leakage (LOKI [30])846.70
445.09
241.15
135.20
", + "image_path": "48c4f9c896068650377fb65a05876315147f82c67b146795f4d1c2d783c5565b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 98, + 426, + 237, + 476 + ], + "blocks": [ + { + "bbox": [ + 432, + 337, + 493, + 347 + ], + "lines": [ + { + "bbox": [ + 432, + 337, + 493, + 347 + ], + "spans": [ + { + "bbox": [ + 432, + 337, + 493, + 347 + ], + "type": "text", + "content": "(c) Tiny ImageNet" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 98, + 426, + 237, + 476 + ], + "lines": [ + { + "bbox": [ + 98, + 426, + 237, + 476 + ], + "spans": [ + { + "bbox": [ + 98, + 426, + 237, + 476 + ], + "type": "table", + "html": "
FC size factorLOKI FedSGDLOKI FedAvg
887.58 (43788)92.19 (46097)
478.93 (39464)85.98 (42989)
259.76 (29882)75.29 (37645)
136.48 (18242)58.39 (29196)
", + "image_path": "ab5cbcfdf3a15ac69ed120d590e7b488b560e1efa3521dfc676df42bba8208ab.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 76, + 518, + 260, + 570 + ], + "blocks": [ + { + "bbox": [ + 46, + 480, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 46, + 480, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 480, + 288, + 514 + ], + "type": "text", + "content": "Table 4. Leakage rate % (leaked images) of LOKI in FedSGD and FedAvg on CIFAR-10 for several FC layer sizes. FedAvg with half the FC layer size has comparable leakage rate to FedSGD." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 518, + 260, + 570 + ], + "lines": [ + { + "bbox": [ + 76, + 518, + 260, + 570 + ], + "spans": [ + { + "bbox": [ + 76, + 518, + 260, + 570 + ], + "type": "table", + "html": "
FC factorFedSGD accuracyFedAvg accuracy
Linear layer leakage (LOKI [30])893.1693.31
492.9492.88
291.9092.35
188.8691.11
", + "image_path": "31d3e565cdc0a4a62a8098b7373f195d5e07984e95977dc833603b9337c09a4a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 573, + 265, + 585 + ], + "lines": [ + { + "bbox": [ + 67, + 573, + 265, + 585 + ], + "spans": [ + { + "bbox": [ + 67, + 573, + 265, + 585 + ], + "type": "text", + "content": "Table 5. CIFAR-10 test accuracy with LOKI FedAvg." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": "els. We use a WideResNet-28-2 [27] and train for 450 epochs on CIFAR-10. Figure 3a fixes the number of known labels to be 40 and shows the test accuracy curve with different fully connected layer sizes. We note that when there is only a small proportion of leaked data, this leads to a pronounced decrease in performance in SSL. LOKI with an FC factor of 1 (with " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "36.48\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": " leakage rate) illustrates this, as it only achieves " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "78.31\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": " max accuracy compared to the CIFAR-10 40 label baseline of " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "92.18\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": ". With a factor of 2, the accuracy improves and jumps to " + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "inline_equation", + "content": "88.91\\%" + }, + { + "bbox": [ + 46, + 594, + 289, + 713 + ], + "type": "text", + "content": " as the leak" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 427, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 536 + ], + "type": "text", + "content": "age rate also substantially increases to " + }, + { + "bbox": [ + 304, + 427, + 545, + 536 + ], + "type": "inline_equation", + "content": "59.76\\%" + }, + { + "bbox": [ + 304, + 427, + 545, + 536 + ], + "type": "text", + "content": ". With factors 4 and 8 (78.93% and 87.58% leakage rate), the accuracy is very close to the baseline at 91.18% and 91.84% respectively. Figure 3b shows the test accuracy plot of LOKI size factor 4 with the number of known labels being 20, 40, and 250. With 20 and 250 labels, LOKI factor 4 achieves 86.69% and 91.65% peak accuracy. We note that there is a greater variability in accuracy with a smaller number of labels, a trend consistent with other SSL methods [14, 25, 31]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 543, + 516, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 543, + 516, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 543, + 516, + 555 + ], + "type": "text", + "content": "4.6. Starting training from federated models" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "While the previous experiments worked with leaked data from all batches in the dataset, such large amounts of leaked data may not always be available. For example, LLL attacks may affect model performance due to manipulation of the model parameters and/or architecture. In later rounds when model performance is high, this may be easier to detect. Similarly, GI attacks such as Inverting Gradients [6] have better reconstruction quality with an untrained network. In both cases, it is possible that the attack is only applied during a few initial rounds of training and only a small portion of data is leaked. Training from scratch with a small portion of leaked data does not achieve good model performance alone, so instead, we can initialize from the fully trained FL" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12252" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 254, + 183 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 254, + 183 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 254, + 183 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 254, + 183 + ], + "type": "image", + "image_path": "96a88cd287b7aa7e619fa2b9b30c5f1578bac4dc4720164d481efb25845c8989.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 128, + 189, + 206, + 198 + ], + "lines": [ + { + "bbox": [ + 128, + 189, + 206, + 198 + ], + "spans": [ + { + "bbox": [ + 128, + 189, + 206, + 198 + ], + "type": "text", + "content": "(a) Varying leakage rate" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 80, + 202, + 252, + 314 + ], + "blocks": [ + { + "bbox": [ + 80, + 202, + 252, + 314 + ], + "lines": [ + { + "bbox": [ + 80, + 202, + 252, + 314 + ], + "spans": [ + { + "bbox": [ + 80, + 202, + 252, + 314 + ], + "type": "image", + "image_path": "4a19abd6be1601b4fe053ebee474da233da9f51943d5494b2499860b9f36c49f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 320, + 213, + 330 + ], + "lines": [ + { + "bbox": [ + 119, + 320, + 213, + 330 + ], + "spans": [ + { + "bbox": [ + 119, + 320, + 213, + 330 + ], + "type": "text", + "content": "(b) Varying number of labels" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 82, + 384, + 254, + 495 + ], + "blocks": [ + { + "bbox": [ + 47, + 334, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 378 + ], + "type": "text", + "content": "Figure 3. Semi-supervised learning using CoMatch on CIFAR-10 with a WideResNet with (a) a varying FC size and leakage rate (LR) for LOKI and the known labels fixed at 40 and (b) a fixed number of leaked images and 20, 40, and 250 known labels." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 82, + 384, + 254, + 495 + ], + "lines": [ + { + "bbox": [ + 82, + 384, + 254, + 495 + ], + "spans": [ + { + "bbox": [ + 82, + 384, + 254, + 495 + ], + "type": "image", + "image_path": "d48916e9d49d4a26cd1b8c08844cf60c755bd2b6a184520b246e91eda313b904.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 502, + 287, + 535 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 535 + ], + "type": "text", + "content": "Figure 4. FL model trained with 50 clients as start point for training with leaked LLL data. Highlighted area indicates improvement above the FL model and training on the leaked data alone." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 544, + 266, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 266, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 266, + 555 + ], + "type": "text", + "content": "model and then train it centrally using the leaked data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "We use smaller portions of the leaked data from LOKI FC size factor of 8 for this experiment. With only " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " (2500) of leaked images in CIFAR-10, a model trained from scratch can only achieve " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "61.15\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " test accuracy. With " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " (1000 images) of data, the model performance drops to " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "39.63\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": ". Both cases are below an FL model with 50 clients that achieves a " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "64.08\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " final test accuracy. However, by using the FL model as the initial starting point, even a small amount of leaked data allows for an increase in performance. We take the model that has been trained using FL and then additionally train that model centrally. Using only " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " of the leaked data, the models are able to achieve " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "68.25\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "inline_equation", + "content": "74.25\\%" + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": " final testing ac" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 416, + 129 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 416, + 129 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 416, + 129 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 416, + 129 + ], + "type": "table", + "html": "
PSNR% ims keptTest accuracy
>2015.8870.04
>1827.5971.52
>1640.7873.73
>1460.2275.48
>1284.4276.16
", + "image_path": "ebc175d5d9b7228c9a88d58b34e345a73275e717ec964b4421324e0e6d22a0e3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 433, + 70, + 535, + 129 + ], + "blocks": [ + { + "bbox": [ + 323, + 137, + 408, + 147 + ], + "lines": [ + { + "bbox": [ + 323, + 137, + 408, + 147 + ], + "spans": [ + { + "bbox": [ + 323, + 137, + 408, + 147 + ], + "type": "text", + "content": "(a) PSNR above threshold" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 433, + 70, + 535, + 129 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 535, + 129 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 535, + 129 + ], + "type": "table", + "html": "
PSNR% ims keptTest accuracy
< 2084.1273.03
< 1872.4167.09
< 1659.2263.71
< 1439.7858.29
< 1215.5845.05
", + "image_path": "4efa67a88f32285a2d9bc47d4f7e064bcaa736d93b4282a5385eaad980ec3bcf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 441, + 137, + 527, + 147 + ], + "lines": [ + { + "bbox": [ + 441, + 137, + 527, + 147 + ], + "spans": [ + { + "bbox": [ + 441, + 137, + 527, + 147 + ], + "type": "text", + "content": "(b) PSNR below threshold" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 148, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 306, + 148, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 545, + 182 + ], + "type": "text", + "content": "Table 6. Training models using CIFAR-10 leaked data from inverting gradients batch size 16. Only the reconstructions with a PSNR (a) above and (b) below the threshold are used in training." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "content": "curacy, a " + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "28.62\\%" + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "inline_equation", + "content": "13.10\\%" + }, + { + "bbox": [ + 304, + 190, + 545, + 250 + ], + "type": "text", + "content": " improvement from starting from scratch. Figure 4 shows the improvement in test accuracy achieved by using the trained FL model as initialization instead of starting from scratch. The improvement is especially noticeable with small percentages of leaked data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 253, + 470, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 470, + 265 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 470, + 265 + ], + "type": "text", + "content": "4.7. Quality of data reconstruction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 266, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 337 + ], + "type": "text", + "content": "Quality of reconstruction is an important concern for GI attacks. As the batch size continues to increase, the reconstruction quality decreases. Table 2a also shows this relationship. From the previous experiments, we also see a negative relationship between the increasing batch size and the usefulness of the leaked data in downstream model training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "content": "Here, we explore whether poor quality reconstructions are still useful for training models. Using the CIFAR-10 leaked dataset created using Inverting Gradients on a batch size of 16, we first sort the reconstructions by their PSNR value. Our first set of experiments removes the worst images from training. Table 6a shows the percent of images above the PSNR threshold and the final test accuracy when trained on that data. Compared to the baseline accuracy of " + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "inline_equation", + "content": "76.83\\%" + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "content": " achieved by using the entire leaked dataset, even removing the worst images with a " + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "inline_equation", + "content": "\\mathrm{PSNR} < 12" + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "content": " from the training set results in a small but non-zero performance loss. Then we take a different look at this question by now training on the worst images. Table 6b shows the test accuracy of models trained on all images below a PSNR threshold. Even when training on the lowest quality images with a PSNR " + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "inline_equation", + "content": "< 12" + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "content": ", the model achieves a " + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "inline_equation", + "content": "45.05\\%" + }, + { + "bbox": [ + 304, + 338, + 546, + 540 + ], + "type": "text", + "content": " final testing accuracy, significantly worse than the baseline, but much above zero." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 542, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 542, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 545, + 637 + ], + "type": "text", + "content": "From the previous results, we see that even leaked data with a low reconstruction quality is still useful to the model training process. This message is a new insight — all prior works when calculating the leak rate do not count poorly reconstructed images. We additionally include results using SSIM in the supplementary section. Visual examples of poor quality reconstructions used to train the models are also included in the supplement." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 640, + 544, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 640, + 544, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 640, + 544, + 677 + ], + "type": "text", + "content": "4.8. Observations on reconstruction quality trends We additionally note a few trends manually observed when going through the GI batch reconstructions of CIFAR-10." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "While some reconstructions have very low quality, only images of the same label will be swapped within the batch during label matching. This is a desirable property for train" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12253" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 272, + 133 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 272, + 133 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 272, + 133 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 272, + 133 + ], + "type": "image", + "image_path": "037c55b704c969df021352cee8be54e08a1c7271bf868a3b7ca118b575e4e55a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 185, + 272, + 244 + ], + "blocks": [ + { + "bbox": [ + 46, + 140, + 287, + 184 + ], + "lines": [ + { + "bbox": [ + 46, + 140, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 140, + 287, + 184 + ], + "type": "text", + "content": "Figure 5. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Ground truth images are on top and reconstructions are on the bottom. All labels in the batch are different and reconstructed images are high quality." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 185, + 272, + 244 + ], + "lines": [ + { + "bbox": [ + 63, + 185, + 272, + 244 + ], + "spans": [ + { + "bbox": [ + 63, + 185, + 272, + 244 + ], + "type": "image", + "image_path": "cdd38955952bead9e00475752c68023aec498d6130665d635920cb69892334be.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 252, + 287, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 252, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 287, + 307 + ], + "type": "text", + "content": "Figure 6. Reconstructions on CIFAR-10 from Inverting Gradients batch size 8. Duplicate labels exist between images 0, 4, 7 and images 2, 3. Reconstruction 3 contains parts of the content from both images 2 and 3. Reconstructions 0, 4, and 7 appear almost as noise. However, the other images still have reasonable quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 311, + 286, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 286, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 286, + 395 + ], + "type": "text", + "content": "ing on leaked data. As long as the set of labels is leaked correctly, the optimization process will match images to a correct label. So even though the image quality is poor (and can be entirely unrecognizable), since the label is correctly matched, such an image still does help in the training (this buttresses the message from Tables 6a and 6b). Naturally, as the PSNR goes down, the image helps less and less." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 396, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 286, + 574 + ], + "type": "text", + "content": "We also note that when all labels are different within a batch, the reconstruction quality of the entire batch of images is almost always good. An example on CIFAR-10 is shown in Figure 5. Usually when the reconstructions are poor, it occurs between groups of images in the same batch that belong to the same class. Poor reconstructions can appear as almost entirely noise, but occasionally images can have visible pieces of each image from the same class in the batch stitched together. The example batch in Figure 6 depicts these properties of poorly reconstructed images. As a side note, we observe that a batch may have a mix of well reconstructed and poorly reconstructed images. We also note that the reverse mapping of the same label images to poor reconstructions is not true — having duplicate labels does not always imply that the reconstructions will be bad." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 579, + 117, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 117, + 591 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 117, + 591 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "As a main goal of our work, we have explored the effect of leaked data through today's leading data reconstruction attacks on downstream model training. Our main message is that it is important to also consider the use of data beyond reconstruction quality and image similarity. It is important to consider how far these leaked samples help in a downstream training task. From the experiments, we have seen that current reconstruction attacks are powerful, but still lack in several critical areas. For gradient inversion attacks, the reconstruction quality does pose an issue for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "training models. Another impediment to practical use of this attack type is the time and computational cost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "text", + "content": "Linear layer leakage attacks do not suffer from reconstruction quality issues and are much lighter weight computationally, but still suffer from other challenges. While LOKI has increased the efficiency of the FC layer leakage in FedAvg, there is still room for improvement (especially in FedSGD). As shown in [29], the model size overhead added by these attacks can be very large, especially if secure aggregation is used. Furthermore, the lack of labels poses a large challenge in training models on the data. With a small dataset like CIFAR-10, hand labeling 40 images is not too challenging. However, larger datasets will require more total labels. For example, SimMatch with " + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "text", + "content": " of labels on ImageNet can achieve " + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "inline_equation", + "content": "67.2\\%" + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "text", + "content": " top-1 accuracy [31]. However, even " + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 96, + 545, + 311 + ], + "type": "text", + "content": " of the total images in ImageNet-1k is 12,812 images. Even if the images were leaked, hand labeling would be extremely hard. Furthermore, we see from Figure 3a that having less total data and fewer labels results in an even steeper decrease in performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 312, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 431 + ], + "type": "text", + "content": "While not discussed previously, there are many other aspects that affect data reconstruction attacks when considering training models. For example, how does the non-IID aspect of client data affect reconstruction? Other relevant settings of federated learning such as asynchronous federated learning, client selection, and differential privacy can also affect the final leaked dataset. Centralized training also allows for more complex and larger models without the communication or computational restrictions of clients in federated learning. These topics can be explored in future work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 436, + 378, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 436, + 378, + 448 + ], + "spans": [ + { + "bbox": [ + 306, + 436, + 378, + 448 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": "We examine data reconstruction attacks through the lens of training models on the leaked data. While highlighting how the weaknesses of gradient inversion in terms of reconstruction quality affect downstream model training, we also show that even poorly reconstructed images are useful for training. We also discuss how the label matching problem for linear layer leakage can be mitigated through semisupervised learning. Under ideal conditions, we demonstrate that leaked data from both gradient inversion and linear layer leakage attacks are able to train powerful models comparable to even a centralized baseline. On CIFAR-10, gradient inversion and linear layer leakage attacks achieve " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "90.34\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "93.16\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " testing accuracy respectively, " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "17.58\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "20.40\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " higher than federated learning and only " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "4.04\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 304, + 450, + 545, + 629 + ], + "type": "text", + "content": " lower compared to centralized training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This material is based in part upon work supported by Adobe Research, the Army Research Office under Contract number W911NF-2020-221, and the National Science Foundation under Grant Number CCF-1919197. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the sponsors." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12254" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Franziska Boenisch, Adam Dziedzic, Roei Schuster, Ali Shahin Shamsabadi, Ilia Shumailov, and Nicolas Papernot. When the curious abandon honesty: Federated learning is not private. 8th IEEE European Symposium on Security and Privacy (IEEE Euro S&P), 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 212 + ], + "type": "text", + "content": "[2] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth. Practical secure aggregation for privacy-preserving machine learning. In proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, pages 1175-1191, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 256 + ], + "type": "text", + "content": "[3] Christopher A Choquette-Choo, Florian Tramer, Nicholas Carlini, and Nicolas Papernot. Label-only membership inference attacks. In International conference on machine learning, pages 1964-1974. PMLR, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 257, + 287, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 257, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 287, + 311 + ], + "type": "text", + "content": "[4] Lixin Fan, Kam Woh Ng, Ce Ju, Tianyu Zhang, Chang Liu, Chee Seng Chan, and Qiang Yang. Rethinking privacy preserving deep learning: How to evaluate and thwart privacy attacks. In Federated Learning, pages 32-50. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 367 + ], + "type": "text", + "content": "[5] Liam H Fowl, Jonas Geiping, Wojciech Czaja, Micah Goldblum, and Tom Goldstein. Robbing the fed: Directly obtaining private data in federated learning with modified models. In International Conference on Learning Representations, 2022. 1, 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 422 + ], + "type": "text", + "content": "[6] Jonas Geiping, Hartmut Bauermeister, Hannah Dröge, and Michael Moeller. Inverting gradients-how easy is it to break privacy in federated learning? Advances in Neural Information Processing Systems, 33:16937-16947, 2020. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 423, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 287, + 466 + ], + "type": "text", + "content": "[7] Jiahui Geng, Yongli Mou, Feifei Li, Qing Li, Oya Bayan, Stefan Decker, and Chunming Rong. Towards general deep leakage in federated learning. arXiv preprint arXiv:2110.09074, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 512 + ], + "type": "text", + "content": "[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 567 + ], + "type": "text", + "content": "[9] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deep models under the gan: information leakage from collaborative deep learning. In Proceedings of the 2017 ACM SIGSAC conference on computer and communications security, pages 603-618, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[10] Yangsibo Huang, Samyak Gupta, Zhao Song, Kai Li, and Sanjeev Arora. Evaluating gradient inversion attacks and defenses in federated learning. Advances in Neural Information Processing Systems, 34:7232-7241, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[11] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Toronto, 2009. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 667 + ], + "type": "text", + "content": "[12] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 689 + ], + "type": "text", + "content": "[13] Yann LeCun. The mnist database of handwritten digits. http://yann.lecun.com/exdb/mnist/, 1998. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Junnan Li, Caiming Xiong, and Steven CH Hoi. Comatch: Semi-supervised learning with contrastive graph regulariza" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9475-9484, 2021. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 148 + ], + "type": "text", + "content": "[15] Xinjian Luo, Yuncheng Wu, Xiaokui Xiao, and Beng Chin Ooi. Feature inference attack on model predictions in vertical federated learning. In 2021 IEEE 37th International Conference on Data Engineering (ICDE), pages 181-192. IEEE, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 194 + ], + "type": "text", + "content": "[16] Kailang Ma, Yu Sun, Jian Cui, Dawei Li, Zhenyu Guan, and Jianwei Liu. Instance-wise batch label restoration via gradients in federated learning. In The Eleventh International Conference on Learning Representations, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "type": "text", + "content": "[17] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 293 + ], + "type": "text", + "content": "[18] Luca Melis, Congzheng Song, Emiliano De Cristofaro, and Vitaly Shmatikov. Exploiting unintended feature leakage in collaborative learning. In 2019 IEEE symposium on security and privacy (SP), pages 691-706. IEEE, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 348 + ], + "type": "text", + "content": "[19] Milad Nasr, Reza Shokri, and Amir Houmansadr. Comprehensive privacy analysis of deep learning: Passive and active white-box inference attacks against centralized and federated learning. In 2019 IEEE symposium on security and privacy (SP), pages 739-753. IEEE, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[20] Dario Pasquini, Danilo Francati, and Giuseppe Ateniese. Eluding secure aggregation in federated learning via model inconsistency. In Proceedings of the 2022 ACM SIGSAC Conference on Computer and Communications Security, pages 2429-2443, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[21] Le Trieu Phong, Yoshinori Aono, Takuya Hayashi, Lihua Wang, and Shiho Moriai. Privacy-preserving deep learning: Revisited and enhanced. In International Conference on Applications and Techniques in Information Security, pages 100-110. Springer, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[22] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 558 + ], + "type": "text", + "content": "[23] Zhibo Wang, Mengkai Song, Zhifei Zhang, Yang Song, Qian Wang, and Hairong Qi. Beyond inferring class representatives: User-level privacy leakage from federated learning. In IEEE INFOCOM 2019-IEEE Conference on Computer Communications, pages 2512-2520. IEEE, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[24] Yuxin Wen, Jonas Geiping, Liam Fowl, Micah Goldblum, and Tom Goldstein. Fishing for user data in large-batch federated learning via gradient magnification. International Conference on Machine Learning, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[25] Fan Yang, Kai Wu, Shuyi Zhang, Guannan Jiang, Yong Liu, Feng Zheng, Wei Zhang, Chengjie Wang, and Long Zeng. Class-aware contrastive semi-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14421-14430, 2022. 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[26] Hongxu Yin, Arun Mallya, Arash Vahdat, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. See through gradients: Image batch recovery via gradinversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16337-16346, 2021. 1, 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12255" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 351 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[27] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 286, + 128 + ], + "type": "text", + "content": "[28] Bo Zhao, Konda Reddy Mopuri, and Hakan Bilen. idlg: Improved deep leakage from gradients. arXiv preprint arXiv:2001.02610, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 286, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 286, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 286, + 195 + ], + "type": "text", + "content": "[29] Joshua C Zhao, Ahmed Roushdy Elkordy, Atul Sharma, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. The resource problem of using linear layer leakage attack in federated learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2023. 1, 2, 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 286, + 262 + ], + "type": "text", + "content": "[30] Joshua Christian Zhao, Atul Sharma, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Salman Avestimehr, and Saurabh Bagchi. Loki: Large-scale data reconstruction attack against federated learning through model manipulation. In 2024 IEEE Symposium on Security and Privacy (SP), pages 30-30. IEEE Computer Society, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 286, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 286, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 286, + 318 + ], + "type": "text", + "content": "[31] Mingkai Zheng, Shan You, Lang Huang, Fei Wang, Chen Qian, and Chang Xu. Simmatch: Semi-supervised learning with similarity matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14471-14481, 2022. 5, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 286, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 286, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 286, + 351 + ], + "type": "text", + "content": "[32] Ligeng Zhu, Zhijian Liu, and Song Han. Deep leakage from gradients. Advances in neural information processing systems, 32, 2019. 1, 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12256" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_content_list.json b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ccb93658569470f6daa4129c0e7a3d69e72a8a4 --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_content_list.json @@ -0,0 +1,1850 @@ +[ + { + "type": "text", + "text": "Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering", + "text_level": 1, + "bbox": [ + 81, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suyuan Liu1, Ke Liang1, Zhibin Dong1, Siwei Wang2, Xihong Yang1, Sihang Zhou1, En Zhu1, Xinwang Liu1* \n1National University of Defense Technology, Changsha, China \n2Intelligent Game and Decision Lab, Beijing, China", + "bbox": [ + 84, + 200, + 950, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{suyuanliu,xinwangliu,enzhu}@nudt.edu.cn", + "bbox": [ + 336, + 258, + 691, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, anchor-based methods have achieved promising progress in multi-view clustering. The performances of these methods are significantly affected by the quality of the anchors. However, the anchors generated by previous works solely rely on single-view information, ignoring the correlation among different views. In particular, we observe that similar patterns are more likely to exist between similar views so such correlation information can be leveraged to enhance the quality of the anchors, which is also omitted. To this end, we propose a novel plug-and-play anchor enhancement strategy through view correlation for multi-view clustering. Specifically, we construct a view graph based on aligned initial anchor graphs to explore inter-view correlations. By learning from view correlation, we enhance the anchors of the current view using the relationships between anchors and samples on neighboring views, thereby narrowing the spatial distribution of anchors on similar views. Experimental results on seven datasets demonstrate the superiority of our proposed method over other existing methods. Furthermore, extensive comparative experiments validate the effectiveness of the proposed anchor enhancement module when applied to various anchor-based methods.", + "bbox": [ + 75, + 342, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 722, + 209, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale scenarios have always been a hot concern with the data explosion in our real world. Among these scenarios, large-scale multi-view clustering (MVC) has also drawn increasing attention in recent years[2, 4, 8, 18, 35]. In general, large-scale MVC methods can be categorized into two types, i.e., matrix factorization-based methods and anchor-based methods. The former methods [11, 24, 41, 50] aim to generate a low-dimensional representation for further clustering by decomposing the original data, thereby avoiding the computational costs of similarity matrix construc", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2696bbd21b59aa35ef1a797cc514518d9ff939b0579d1af45fb497bea0883d18.jpg", + "image_caption": [ + "Figure 1. An example illustration of clustering results affected by initial anchors. The first two methods keep the anchors fixed, while the latter two methods update the anchors through iteration-optimization. Their clustering performances are influenced to varying degrees by different anchor generation approaches." + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 885, + 462 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion. However, their efficiency usually decreases along with increasing dimension, thus constraining their scalability[10, 36]. Inspired by prototype learning [16, 48], anchors are introduced into MVC to reduce the scope for clustering from $n$ to $m$ , which denotes the number of samples and anchors, respectively. Specifically, anchor-based MVC methods [13, 27, 45] construct an anchor graph $(n \\times m)$ rather than a similarity matrix $(n \\times n)$ . It will alleviate the aforementioned computation and scalability problems.", + "bbox": [ + 496, + 563, + 893, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anchor-based MVC methods typically consist of three procedures, including (1) anchor generation, (2) anchor graph construction, and (3) clustering. Early attempts usually generate anchors at first, which are fixed during constructing anchor graphs[31, 32, 47]. For example, SFMC [19] selects anchors by iterative sampling based on feature similarity. Besides, LMVSC [14] regards cluster centers generated by $k$ -means of each view as the anchors. However, these methods are sensitive to the initial anchors (See Fig. 1), where the anchor selection is primitive and sometimes results in compromised performance.", + "bbox": [ + 496, + 700, + 893, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the above issues, researchers make efforts on anchor optimization to guarantee better quality anchors[17,", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26151", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "26, 38]. For example, SMVSC [33] proposed a unified framework that combines anchor generation and anchor-graph construction, achieving iterative updates to optimize the anchors. Building upon this, UDBGL [7] introduced the Laplacian rank constraint to ensure that the anchor graph possesses a discrete cluster structure. The above iteration-optimization strategies actually aim to get a better combination of the anchor and its corresponding anchor graph instead of enhancing the quality of the anchors themselves. Meanwhile, to the best of our knowledge, almost all existing anchor optimization models belong to this type. Although they can get better and more stable clustering performance, they still suffer from the initial state for optimization[6, 15], which leaves us a great challenge as well as the opportunity, i.e., what is the new strategy to alleviate such problem? In this paper, we provide a potential solution to it.", + "bbox": [ + 76, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observe that most anchor generation methods only rely on a single view[3, 12, 30]. Multiple works have proven that information of different views for the same sample will benefit the expressive ability of the sample [23, 29, 34, 49]. Similarly, different views for the same anchors should also be used to generate better anchors. Specifically, utilizing information across views and ensuring the consistency of anchor graphs on each view will benefit the anchor generation procedure. However, such a characteristic is not used for anchor enhancement in previous anchor-based MVC methods, and our strategy is developed based on it.", + "bbox": [ + 76, + 334, + 468, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose an Anchor Enhancement strategy with View Correlation (AEVC) for multi-view clustering (MVC). The framework of AEVC is shown in Fig. 2. The proposed approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which enhance the quality of anchors. Additionally, we improve the clustering performance by revising the existing anchor graph construction method based on view correlation. Specifically, we first align the initial anchors and construct a view-graph based on the aligned anchor graph to find the neighbors of each view. Then, new anchors for the current view are generated with the anchor graph of neighboring views, thereby enhancing the initial anchors. Finally, the inter-view correlations are combined to learn a consistent anchor graph between views with the enhanced anchors. AEVC optimizes anchors by absorbing information from similar views, improving initial anchor quality, and reducing distribution bias caused by the independent anchors generation in each view. The main contributions of our work can be summarized as follows:", + "bbox": [ + 76, + 502, + 472, + 819 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose AEVC, a plug-and-play anchor enhancement strategy for MVC based on view correlation, to address the issue of poor initial anchor quality affecting the performance of MVC.", + "- By learning from the view correlation, AEVC incorpo" + ], + "bbox": [ + 76, + 825, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/89574ce77c2e9aa7ee1accf6245c120eb95e5f018a0f390a605f7205ac8c126c.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed AEVC. Specifically, in the view-graph learning module, AEVC primarily performs anchor alignment and learns the view-graph. In the anchor enhancement module, AEVC leverages the neighboring relationships from the view-graph to enhance the anchors. Finally, AEVC executes anchor graph construction based on the enhanced anchors." + ], + "image_footnote": [], + "bbox": [ + 496, + 89, + 895, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rates the anchor distribution information from neighboring views into the process of anchor enhancement. Theoretical and experimental validations demonstrate that the anchors enhanced by AEVC exhibit a tighter inter-view relationship.", + "bbox": [ + 511, + 398, + 890, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We compared our proposed approach to other anchor-based multi-view methods and demonstrated its superiority. Extensive experiments verify the effectiveness and generalization of the proposed AEVC strategy when integrated with typical anchor-based MVC methods.", + "bbox": [ + 500, + 474, + 890, + 550 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 565, + 640, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anchor-based Multi-view Clustering", + "text_level": 1, + "bbox": [ + 500, + 590, + 816, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges posed by traditional multiview clustering methods in handling large-scale datasets, Anchor-based MVC has been proposed in recent years[37, 45, 55]. The core idea of anchor-based MVC is to generate a set of representative anchors from the entire sample pool, where the number of anchors is typically much smaller than the total number of samples[25, 54]. In contrast to conventional methods that require establishing relationships among all samples, anchor-based approaches only involve computing similarities between anchors and samples. This significantly reduces the complexity associated with constructing the relationship graph. Additionally, adopting a reduced-dimensional graph further mitigates computational burdens, facilitating subsequent multi-view fusion and clustering with reduced computational overhead.", + "bbox": [ + 496, + 613, + 890, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional anchor-based MVC methods generate anchors by sampling[30, 31, 46]. Subsequent anchor graph construction and clustering are performed based on the initially determined anchors. For instance, Kang et al.[14]", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26152", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and Yang et al. [46] advocate applying $k$ -means to the original samples, selecting cluster centers as anchors to enhance their representativeness. Xia et al. [44] introduces a variance-based de-correlation anchor selection strategy, leveraging the internal data structure to augment the coverage capability of anchors. The performance of such methods is highly influenced by the quality of the initially generated anchors.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Anchor Optimization Strategy", + "text_level": 1, + "bbox": [ + 76, + 223, + 341, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the above challenges, extensive research efforts have been devoted to anchor optimization in order to achieve superior anchor quality. Iteration-optimization methods propose to update anchors and anchor graphs, utilizing the anchor graph obtained from the previous iteration to update the anchors[51, 53, 56]. Sun et al. [33] initially propose learning anchors and constructing anchor graphs through iterative updates within the same framework. Liu et al. [26] further introduce Laplacian rank constraints, enabling the anchor graph to generate labels without directly needing post-processing. However, the results of iterative update algorithms are significantly influenced by the initial values of variables[6, 15]. Instead of directly enhancing the quality of the anchors themselves, the above methods primarily focus on obtaining an improved combination of the anchor and its corresponding anchor graph.", + "bbox": [ + 75, + 248, + 472, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. View Correlation Learning", + "text_level": 1, + "bbox": [ + 76, + 501, + 321, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-view and multi-modal scenarios are widely seen nowadays [1, 20]. Among them, investigating the relationships between multiple objects is a crucial topic [5, 21, 22], especially for multi-view learning, which holds for anchor-based multi-view clustering[7, 44]. Some existing methods delve into the consistency of views by aligning anchors[42, 43, 52]. For instance, Li et al. [19] concatenates raw data from all views to select anchors, ensuring correspondence among anchors across views. Wang et al. [39] takes a step further by proposing an anchor alignment framework based on the features and structural information of anchors. Another category of methods employs early fusion to avoid conflicts between views[9, 28]. For instance, Wang et al. [40] use a projection matrix to learn shared anchors and anchor graphs in the latent space. However, the first category of methods solely utilizes alignment information between views, while the second category underutilizes complementary information among views. In the next section, we will introduce a novel approach based on a view graph, leveraging inter-view interactions to enhance anchors.", + "bbox": [ + 75, + 525, + 472, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 166, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed Anchor Enhancement strategy with View Correlation (AEVC) for", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multi-view clustering. We begin by sequentially describing the three modules of AEVC. Subsequently, a complexity analysis of the proposed algorithm is provided. Lastly, a theoretical analysis is conducted to evaluate the effectiveness of AEVC.", + "bbox": [ + 496, + 90, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 179, + 609, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing anchor-based MVC methods can be divided into three steps: (1) anchor generation, (2) anchor graph construction, (3) clustering. The clustering performance of existing anchor-based MVC methods heavily relies on the quality of initially generated anchors. Although iteration-optimization strategies alleviate this issue by exploring combinations of anchors and anchor graphs, they are still impacted by the optimization of the initial state. To address this problem, we propose an anchor enhancement strategy based on view correlation for multi-view clustering. Specifically, different from existing methods, our approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which helps to enhance the quality of anchors. Furthermore, we revise the existing anchor graph construction method based on view correlation, further improving the clustering performance.", + "bbox": [ + 496, + 203, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. View-graph Learning", + "text_level": 1, + "bbox": [ + 500, + 473, + 702, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given multi-view data $\\{\\mathbf{X}_i\\}_{i=1}^v$ , where $\\mathbf{X}_i \\in \\mathbb{R}^{d_i \\times n}$ is composed of $n$ samples with $d_i$ dimension. Denoting $\\mathbf{A}_i^{(0)} \\in \\mathbb{R}^{d_i \\times m}$ as the generated anchor matrix in the $i$ -th view, and $\\mathbf{Z}_i^{(0)} \\in \\mathbb{R}^{m \\times n}$ represents the corresponding anchor graph, where $m$ is the number of anchors. Due to the varying dimensions and inconsistent distributions of data across different views, it is not feasible to directly assess the distance between $\\mathbf{X}_i$ and $\\mathbf{X}_j$ . However, it is easy to calculate the similarity between anchor graphs from different views due to the same size of anchor graphs.", + "bbox": [ + 496, + 496, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nevertheless, the independently generated anchors on each view introduce the Anchor-Unaligned Problem [39]. This means that directly computing the similarity between anchor graphs of different views would be inaccurate due to the misalignment of rows corresponding to anchors. Thus, an anchor alignment operation must be performed before view-graph learning.", + "bbox": [ + 496, + 654, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing anchor alignment strategies typically directly align the anchors on the first view, which carries the risk of being influenced by noise view. Here, we first concatenate the multi-view data column-wise to obtain $\\hat{\\mathbf{X}} = [\\mathbf{X}_1; \\ldots; \\mathbf{X}_v] \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times n}$ . Then, we generate the anchors $\\hat{\\mathbf{A}} \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times m}$ and the corresponding anchor graph $\\hat{\\mathbf{Z}} \\in \\mathbb{R}^{m \\times n}$ . According to [39], we can obtain the permutation matrix $\\mathbf{P}_i$ in the $i$ -th view by minimizing the distance between $\\hat{\\mathbf{Z}}$ and the permuted $\\mathbf{Z}_i^{(0)}$ , the formula is", + "bbox": [ + 496, + 761, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26153", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as follows,", + "bbox": [ + 76, + 90, + 151, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {P} _ {i}} \\left\\| \\hat {\\mathbf {Z}} - \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 114, + 468, + 147 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {s . t .} \\mathbf {P} _ {i} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} ^ {\\top} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} \\in \\{0, 1 \\} ^ {m \\times m}.\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 148, + 413, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the anchor matrix and anchor graph can be aligned with the permutation matrix as follows:", + "bbox": [ + 76, + 176, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\mathbf {A} _ {i} ^ {(0)} \\mathbf {P} _ {i} ^ {\\top}, \\\\ \\mathbf {Z} _ {i} = \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)}. \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 217, + 468, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After performing the alignment operation, we calculate the view graph $\\mathbf{S} \\in \\mathbb{R}^{v \\times v}$ as follows,", + "bbox": [ + 76, + 277, + 468, + 308 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\mathbf {s} _ {i j} = \\frac {1}{\\| \\mathbf {Z} _ {i} - \\mathbf {Z} _ {j} \\| _ {\\mathbf {F}} ^ {2}}, & i \\neq j, \\\\ \\mathbf {s} _ {i j} = 0, & i = j,. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 318, + 468, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{s}_{ij}$ denotes the similarity between the $i$ -th and the $j$ -th view.", + "bbox": [ + 76, + 378, + 468, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Anchor Enhancement Strategy", + "text_level": 1, + "bbox": [ + 76, + 417, + 349, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Most existing methods independently generate anchors on each view, ignoring the correlation between views. In fact, anchors on similar views should have similar distributions, which means that the corresponding anchor graphs should be close to each other. Based on this assumption, we propose to enhance the anchors of the current view with information on neighboring views. One intuitive idea is to directly update the anchors based on the anchors of neighboring views. However, the dimensions of anchors on different views are not the same, making direct transfer infeasible. Instead of utilizing feature consistency, we focus on structural consistency. Specifically, we reconstruct new anchors by leveraging the anchor graphs from neighboring views and the data from the current view, thereby enhancing the initial anchors. The formulation of the anchor enhancement on the $i$ -th view is as follows:", + "bbox": [ + 76, + 441, + 468, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {A}} _ {i} = \\frac {1}{\\mu} \\left(\\mathbf {A} _ {i} + \\gamma \\sum_ {\\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right)} \\hat {s} _ {i j} \\mathbf {X} _ {i} \\mathbf {Z} _ {j} ^ {\\top}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 691, + 468, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu = 1 + \\gamma \\sum_{\\mathbf{X}_j \\in \\Omega_\\zeta(\\mathbf{X}_i)} \\hat{s}_{ij}$ ensures that the enhanced anchors are a convex combination of the initial anchors and the reconstructed anchors. $\\Omega_\\zeta(\\mathbf{X}_i)$ denotes the $\\zeta$ -neighbors of the $i$ -th view, which is defined as:", + "bbox": [ + 76, + 741, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right) = \\arg \\max _ {\\mathbf {X} _ {j}} \\sum_ {j = 1} | s _ {i j} |. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 811, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\gamma$ is the enhancement rate, which is used to adjust the influence of the initial anchors on the enhanced version. $\\hat{\\mathbf{S}}$ represents the normalized view-graph based on the following", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "settings, which controls the influence of different neighbors during anchor enhancement:", + "bbox": [ + 500, + 90, + 890, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {s} _ {i j} \\left\\{ \\begin{array}{l l} s _ {i j}, & \\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right), \\\\ 0, & \\mathbf {X} _ {j} \\notin \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right). \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 131, + 890, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By enhancing the anchors with Eq. (4), the quality of the initial anchors is improved. Moreover, the enhancement further achieves a better alignment between anchors on similar views. In the subsequent sections, we will analyze the aforementioned conclusions from both theoretical and experimental perspectives. Furthermore, our proposed anchor enhancement module is plug-and-play and we will showcase its performance improvement in the experiments by integrating different anchor-based multi-view clustering methods.", + "bbox": [ + 498, + 184, + 890, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Anchor Graph Construction", + "text_level": 1, + "bbox": [ + 500, + 344, + 754, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining high-quality anchors through our proposed anchor enhancement module, it is crucial to construct a reliable anchor graph that ensures view consistency. Since the enhanced anchors are aligned, we can directly generate a common anchor graph. However, the contributions of different views in constructing the anchor graph are not equal. To mitigate the influence of view discrepancies on building a consistent representation, we assign higher weights to views that have stronger associations with other views. Specifically, the contribution weight $\\omega_{i} = \\sum_{j=1}^{v} s_{ij}$ for the $i$ -th view is computed as the sum of its similarities with all other views. Furthermore, to avoid the impact of noise during anchor graph construction, we introduce the learned anchor graph $\\hat{\\mathbf{Z}}$ from the concatenated views as guidance. Ultimately, the anchor graph construction is formulated as follows:", + "bbox": [ + 498, + 367, + 890, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {Z}} \\omega_ {i} \\left\\| \\mathbf {X} _ {i} - \\tilde {\\mathbf {A}} _ {i} \\mathbf {Z} \\right\\| _ {\\mathbf {F}} ^ {2} + \\lambda \\left\\| \\mathbf {Z} - \\hat {\\mathbf {Z}} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 614, + 890, + 647 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {s . t .} \\mathbf {Z} \\geq 0, \\mathbf {Z} ^ {\\top} \\mathbf {1} = \\mathbf {1}. \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 647, + 766, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the independence between each column of $\\mathbf{Z}$ , let $z_{j}$ denote the $j$ -th column of $\\mathbf{Z}$ . The aforementioned problem can be transformed into $n$ Quadratic Programming(QP) problems, formulated as follows:", + "bbox": [ + 500, + 674, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\frac {1}{2} \\mathbf {z} _ {j} ^ {\\top} \\mathbf {H} \\mathbf {z} _ {j} + f \\mathbf {z} _ {j}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 743, + 890, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {s . t .} \\mathbf {z} _ {j} \\geq 0, \\mathbf {z} _ {j} ^ {\\top} \\mathbf {1} = 1, \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 765, + 767, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the matrix $\\mathbf{H} = 2\\left(\\sum_{i=1}^{v}\\omega_{i}\\tilde{\\mathbf{A}}_{i}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\mathbf{I}\\right)$ and the vector $f = -2\\left(\\sum_{i=1}^{v}\\omega_{i}\\mathbf{X}_{i[:,j]}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\hat{z}_{j}\\right)$ . The optimal $\\mathbf{Z}$ can be obtained by solving the above QP problem. Finally, the results can be generated by performing $k$ -means on the left singular vector of $\\mathbf{Z}$ [14]. The entire procedure of AEVC is shown in Algorithm 1.", + "bbox": [ + 498, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26154", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: The proposed AEVC", + "text_level": 1, + "bbox": [ + 86, + 94, + 323, + 109 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Multi-view datasets $\\{\\mathbf{X}_i\\}_{i = 1}^v$ , the number of anchors $m$ , the parameter $\\gamma$ $\\lambda$ and the number of clusters $k$", + "bbox": [ + 96, + 112, + 442, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Clustering results.", + "bbox": [ + 98, + 157, + 279, + 172 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 Initialize $\\{\\mathbf{A}_i^{(0)}\\}_{i = 1}^v$ $\\hat{\\mathbf{A}}$ $\\{\\mathbf{Z}_i^{(0)}\\}_{i = 1}^v$ and $\\hat{\\mathbf{Z}}$ // View-graph Learning", + "2 for $i = 1\\to v$ do", + "3 Calculate permutation matrix $\\mathbf{P}_i$ with Eq. (1); 4 Perform alignment with Eq. (2);", + "5 Learn view-graph S as defined in Eq. (3); // Anchor Enhancement", + "6 for $i = 1\\to v$ do", + "7 Find $\\zeta$ -neighbors $\\Omega_{\\zeta}(\\mathbf{X}_i)$ as defined in Eq. (5);", + "8 Calculate normalized view-graph $\\hat{\\mathbf{S}}$ with Eq. (6);", + "9 for $i = 1\\rightarrow v$ do", + "10 Enhance $\\mathbf{A}_i$ with Eq. (4);" + ], + "bbox": [ + 78, + 176, + 433, + 375 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "// Anchor Graph Construction", + "11 Obtain $\\mathbf{Z}$ by solving Eq. (7);", + "12 Perform $k$ -means on the left singular vector of $\\mathbf{Z}$ ." + ], + "bbox": [ + 78, + 378, + 421, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 459, + 194, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computational complexity. The computational complexity of the proposed AEVC mainly consists of three modules. The complexity of the view-graph learning module is $\\mathcal{O}(m^3 v + nmv)$ . Specifically, the computation of $v$ permutation matrices $\\mathbf{P}_i$ has a complexity of $\\mathcal{O}(m^3 v)$ , and the computation of the view graph has a complexity of $\\mathcal{O}(nmv)$ . The time complexity of the anchor enhancement module is $\\mathcal{O}(nmdv)$ , where $d = \\sum_{i=1}^{v} d_v$ , primarily arising from matrix multiplication during enhancement. The complexity of the anchor graph construction module is $\\mathcal{O}(nm^3 v + nmdv)$ , where solving all QP problems requires a time complexity of $\\mathcal{O}(nm^3 v)$ , and the involved matrix multiplication requires $\\mathcal{O}(nm^3 v + nmdv)$ . Therefore, the overall computational complexity of AEVC is $\\mathcal{O}(nm^3 v + m^3 v + nmdv + nmv)$ , which exhibits linear complexity with respect to the number of samples.", + "bbox": [ + 75, + 482, + 468, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Space complexity. During the execution of the AEVC algorithm, the main variables that need to be stored include matrices $\\mathbf{X}_i$ , $\\mathbf{A}_i$ , $\\mathbf{Z}_i$ , $\\hat{\\mathbf{A}}$ , $\\hat{\\mathbf{Z}}$ , $\\mathbf{P}_i$ , $\\mathbf{S}$ , and $\\hat{\\mathbf{S}}$ . The overall space complexity of AEVC is $\\mathcal{O}(nd + ndv + nmdv + v^2)$ , which is also linear with respect to the number of samples and can be extended to large-scale scenarios.", + "bbox": [ + 75, + 724, + 468, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two Property of enhanced anchors. We have discovered that the anchors enhanced by the proposed AEVC exhibit stronger inter-view consistency. The theoretical analysis is provided below.", + "bbox": [ + 75, + 816, + 468, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assumption 1. Given a two-view data $\\{\\mathbf{X}_1, \\mathbf{X}_2\\}$ , denoting", + "bbox": [ + 76, + 885, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\{\\mathbf{A}_1,\\mathbf{A}_2\\}$ are the generated initial anchors, and $\\{\\mathbf{Z}_1,\\mathbf{Z}_2\\}$ are the corresponding anchor graphs under the assumption that $\\mathbf{X}_1 = \\mathbf{A}_1\\mathbf{Z}_1$ and $\\mathbf{X}_2 = \\mathbf{A}_2\\mathbf{Z}_2$", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the convenience of derivation, we simplify the anchor enhancement strategy in Eq. (4) with $\\alpha \\in (0,1)$ as follows:", + "bbox": [ + 496, + 143, + 890, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {A}} _ {1} = \\alpha \\mathbf {A} _ {1} + (1 - \\alpha) \\mathbf {X} _ {1} \\mathbf {Z} _ {2} ^ {\\top}, \\\\ \\tilde {\\mathbf {A}} _ {2} = \\alpha \\mathbf {A} _ {2} + (1 - \\alpha) \\mathbf {X} _ {2} \\mathbf {Z} _ {1} ^ {\\top}. \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 195, + 890, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The corresponding anchor graphs $\\{\\tilde{\\mathbf{Z}}_1,\\tilde{\\mathbf{Z}}_2\\}$ are constructed under the same assumption that $\\mathbf{X}_1 = \\tilde{\\mathbf{A}}_1\\tilde{\\mathbf{Z}}_1$ and $\\mathbf{X}_2 = \\tilde{\\mathbf{A}}_2\\tilde{\\mathbf{Z}}_2$ . Then we can derive that", + "bbox": [ + 498, + 246, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {Z}} _ {1} = \\alpha \\mathbf {Z} _ {1} + (1 - \\alpha) \\mathbf {Z} _ {2}, \\\\ \\tilde {\\mathbf {Z}} _ {2} = \\alpha \\mathbf {Z} _ {2} + (1 - \\alpha) \\mathbf {Z} _ {1}. \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 300, + 890, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can deduce the following properties:", + "bbox": [ + 519, + 349, + 787, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Property 1. The disparity among the anchor graphs on each view is reduced after anchor enhancement. Mathematically,", + "bbox": [ + 498, + 373, + 890, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 425, + 890, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. According to Eq. (10), we can derive that", + "bbox": [ + 500, + 460, + 823, + 476 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} = \\| (2 \\alpha - 1) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = \\left| 2 \\alpha - 1 \\right| \\cdot \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}} \\tag {12} \\\\ < \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 483, + 890, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This completes the proof.", + "bbox": [ + 500, + 556, + 669, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Property 1 implies that anchor enhancement leads to a more consistent distribution of anchors in similar views, thereby establishing a tighter alignment.", + "bbox": [ + 496, + 578, + 890, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Property 2. Assuming there exists a value $\\beta \\in (0,1)$ such that $\\mathbf{Z} = \\beta \\mathbf{Z}_1 + (1 - \\beta)\\mathbf{Z}_2$ is the optimal consistent anchor graph among all views, the enhanced anchor graph is closer to the optimal version after enhancement than before for all $\\alpha \\in (2\\beta -1,1)\\cap (0,1)$ . Mathematically,", + "bbox": [ + 496, + 632, + 890, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\| _ {\\mathbf {F}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 715, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proof. According to Eq. (10), we can derive that", + "bbox": [ + 500, + 751, + 823, + 766 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} = \\| (\\alpha - \\beta) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = | \\alpha - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\\\ < | 1 - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\tag {14} \\\\ = \\left\\| (1 - \\beta) \\left(\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}\\right) \\right\\| _ {\\mathbf {F}} \\\\ = \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 773, + 890, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This completes the proof.", + "bbox": [ + 500, + 885, + 669, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1f847bc1d8bb9a2294dd727278e238a03f0f0c5d8e113a34d4780731614362e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 874, + 886, + 890, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26155", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Property 2 implies that anchor enhancement helps reduce the gap between the anchor graph on each view and the optimal consistent anchor graph. It is easy to derive the multiview extended versions of the above properties.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 169, + 200, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 76, + 195, + 253, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Seven multi-view datasets are employed in our experiments: Dermatology1 contains diagnostic information for six types of skin diseases. ForestTypes2 consists of satellite images of different forest types. BDGP3 comprises both images and textual descriptions of Drosophila embryos. Reuters4 is a news text dataset. Hdigit5 consists of a set of handwritten digital images. VGGFace2 is a large-scale face recognition dataset. CIFAR100 is a subset of a large-scale tiny image dataset. Detail description is shown in Table 1.", + "bbox": [ + 75, + 220, + 468, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare the proposed AEVC with existing state-of-the-art anchor-based MVC methods: Large-scale multi-view subspace clustering in linear time (LMVSC)[14]; Scalable Multi-view Subspace Clustering with Unified Anchors (SMVSC)[33]; Fast multi-view clustering via nonnegative and orthogonal factorization (FMCNOF)[46]; Fast Parameter-Free Multi-View Subspace Clustering With Consensus Anchor Guidance (FPMVSCAG)[40]; Efficient Multi-View Clustering via Unified and Discrete Bipartite Graph Learning (UDBGL)[7]; Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity (FastMICE)[10]; Flexible and Diverse Anchor Graph Fusion for Scalable Multi-View Clustering (FDAGF)[55]; Align then Fusion: Generalized Large-scale Multi-view Clustering with Anchor Matching Correspondences (FMVACC)[39].", + "bbox": [ + 75, + 373, + 470, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For the above-compared algorithm, we have tuned their parameters to align with the relevant literature. We conducted a search for three hyperparameters in the proposed AEVC method using predefined lists. Specifically, the anchor number $m$ is chosen from the set $[k, 2k, 5k]$ , where $k$ represents the cluster number. The enhancement rate $\\gamma$ is selected from $[0.1, 1, 10, 100]$ , and the balancing hyperparameter $\\lambda$ is chosen from the range $[0.0001, 0.01, 1, 100]$ . To ensure a fair comparison, algorithms relying on $k$ -means for result acquisition underwent 50 independent runs, and the results were averaged for accuracy assessment. Our experimental evaluation incorporated", + "bbox": [ + 75, + 616, + 468, + 797 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f68f89be3ec5f40d29aa7a028cc64c3f366f26a111d3e20ed401db6d744f2236.jpg", + "table_caption": [ + "Table 1. Benchmark datasets description." + ], + "table_footnote": [], + "table_body": "
Datasetnvkdp(p=1,...,v)
Dermatology3582612/22
ForestTypes523349/9/9
BDGP2500351000/500/250
Reuters7200564819/4810/4892/4858/4777
Hdigit10000210784/256
VGGFace23628741003944/576/512/640
CIFAR100500004100944/576/512/640
", + "bbox": [ + 504, + 104, + 887, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "four distinct metrics: clustering Accuracy (ACC), Normalized Mutual Information (NMI), Purity, and F-score.", + "bbox": [ + 498, + 236, + 890, + 267 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Superiority of AEVC", + "text_level": 1, + "bbox": [ + 500, + 277, + 699, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our proposed AEVC with eight state-of-the-art anchor-based multi-view clustering methods on seven datasets, as illustrated in Table 2. From the results, we have the following observations:", + "bbox": [ + 498, + 301, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The experimental results reveal that our proposed algorithm consistently outperforms other anchor-based MVC methods across all datasets, confirming the superiority of AEVC. In terms of ACC, AEVC surpasses the second-best algorithm by $6.9\\%$ , $4.1\\%$ , $1.9\\%$ , $9.9\\%$ , $1.3\\%$ , $15.8\\%$ , and $9.5\\%$ on different datasets.", + "- Compared to previous methods based on iteratively optimized anchors (SMVSC, FPMVS-CAG, UDBGL, FDAGF, FMVACC), AEVC exhibits superior clustering performance, emphasizing the effectiveness of our anchor enhancement approach over iteration-optimization methods." + ], + "bbox": [ + 500, + 364, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Effectiveness of Enhanced Anchors", + "text_level": 1, + "bbox": [ + 500, + 559, + 805, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We carried out an ablation study to demonstrate the effectiveness of the proposed anchor enhancement strategy. In our approach, the anchor enhancement strategy typically comprises two steps: anchor alignment in the view-graph learning module and anchor enhancement module. We denote the method of removing the entire anchor enhancement strategy as a baseline and compare it with those where anchor alignment and enhancement are sequentially incorporated. From the experimental results shown in Table 3, it is observed that anchor alignment leads to an improvement in clustering performance, as misalignment of anchors can result in incorrect multi-view fusion[39]. Furthermore, with the application of our anchor enhancement strategy, there is a further improvement in clustering performance, providing strong evidence for the effectiveness of the anchor enhancement strategy.", + "bbox": [ + 496, + 583, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of the enhanced anchors, we incorporate randomly generated anchors from the original data at different proportions and showcase the corresponding clustering performance in Fig. 3(a). For instance, when the proportion is $10\\%$ , it means that $90\\%$", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 http://archive.ics.uci.edu/ml/datasets/Dermatology", + "bbox": [ + 99, + 811, + 374, + 825 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2http://archive.ics.uci.edu/dataset/333/", + "bbox": [ + 102, + 825, + 305, + 837 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3 https://www.fruitfly.org/", + "bbox": [ + 102, + 837, + 236, + 849 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "4http://www.researchAtt.com/%7Elewis/reuters21578.html", + "bbox": [ + 102, + 849, + 413, + 862 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "5https://cs.nyu.edu/%7Erowei/data.html", + "bbox": [ + 102, + 862, + 318, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{6}$ https://www.robots.ox.ac.uk/vgg/data/vgg_face2/", + "bbox": [ + 102, + 875, + 366, + 887 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "7https://www.cs.toronto.edu/%7Ekriz/cifar.html", + "bbox": [ + 102, + 887, + 354, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26156", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/60cbc4a9c4db7d2ca490269d25aff3b7884c8d4323f2b89359a292180cd8d7f7.jpg", + "table_caption": [ + "Table 2. Clustering performance comparison of eight anchor-based MVC methods on seven datasets. The best is marked in bold and underlined, the second best is marked in bold." + ], + "table_footnote": [], + "table_body": "
DatasetsLMVSCSMVSCFMCNOFFPMVS-CAGUDBGLFastMICEFDAGFFMVACCProposed
ACC
Dermatology79.02±0.0778.64±0.0562.01±0.0082.96±0.0784.08±0.0087.15±0.0084.41±7.9084.39±4.3493.20±4.50
ForestTypes77.68±0.0472.82±0.0145.51±0.0072.22±0.0577.06±0.0077.06±0.0076.40±7.1379.29±0.2382.53±0.09
BDGP49.52±0.0237.22±0.0231.08±0.0032.62±0.0139.28±0.0049.00±0.0047.30±3.1659.51±4.0060.65±0.62
Reuters27.19±0.0124.41±0.0119.25±0.0017.33±0.0022.32±0.0026.00±0.0025.04±0.5027.22±0.2429.91±0.02
Hdigit86.87±0.0865.99±0.0332.84±0.0064.22±0.0531.39±0.0085.13±0.0077.69±7.8386.55±4.0888.01±2.97
VGGFace6.09±0.007.21±0.003.47±0.006.02±0.005.39±0.005.26±0.006.50±0.336.68±0.168.35±0.07
CIFAR1008.92±0.008.48±0.004.40±0.007.46±0.007.83±0.009.81±0.007.47±0.137.04±0.1210.74±0.08
NMI
Dermatology70.17±0.0466.62±0.0354.24±0.0071.90±0.0587.41±0.0082.73±0.0079.70±4.3476.27±2.1988.73±3.01
ForestTypes53.56±0.0246.56±0.0116.99±0.0048.20±0.0452.04±0.0056.53±0.0054.80±5.3555.33±0.1559.63±0.09
BDGP25.85±0.029.85±0.0110.29±0.0010.02±0.0114.90±0.0026.03±0.0021.90±1.9535.55±4.5330.92±0.65
Reuters5.67±0.015.45±0.010.64±0.000.13±0.001.79±0.007.23±0.003.36±0.485.05±0.087.07±0.01
Hdigit89.33±0.0356.16±0.0228.57±0.0055.80±0.0319.04±0.0088.67±0.0070.00±2.6076.96±3.0978.09±1.59
VGGFace11.92±0.0014.13±0.005.81±0.0012.33±0.0011.00±0.0010.79±0.0012.32±0.3312.97±0.1514.33±0.11
CIFAR10015.49±0.0014.83±0.008.49±0.0013.87±0.0014.83±0.0016.82±0.0013.93±0.1511.45±0.1616.94±0.14
Purity
Dermatology80.97±0.0480.35±0.0462.85±0.0083.55±0.0784.64±0.0087.15±0.0087.86±4.5185.38±2.0494.03±2.35
ForestTypes77.99±0.0272.82±0.0145.51±0.0072.73±0.0477.06±0.0078.93±0.0077.61±4.7479.29±0.2382.53±0.09
BDGP49.64±0.0237.80±0.0133.32±0.0034.82±0.0140.16±0.0049.44±0.0047.36±3.0859.73±3.6760.65±0.62
Reuters28.22±0.0025.12±0.0119.78±0.0017.89±0.0023.31±0.0028.07±0.0025.45±0.5927.50±0.1730.84±0.02
Hdigit88.93±0.0666.28±0.0332.85±0.0064.34±0.0531.94±0.0088.43±0.0079.82±5.6686.59±4.0188.22±2.25
VGGFace7.02±0.007.73±0.003.50±0.006.33±0.005.77±0.006.11±0.007.17±0.337.47±0.179.43±0.08
CIFAR10010.41±0.009.09±0.004.49±0.007.84±0.008.69±0.0011.11±0.008.28±0.137.89±0.1412.29±0.13
", + "bbox": [ + 80, + 127, + 890, + 439 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0dfa20e04212551ecbf6ce8a876bb060e93e9c353502c2bc9d9cec32be6ebc6b.jpg", + "table_caption": [ + "Table 3. Ablation study on anchor enhancement module." + ], + "table_footnote": [], + "table_body": "
MethodsDermatologyForestTypesBDGP
Baseline90.87±3.6070.76±0.0554.38±0.10
+Align91.82±3.7875.51±0.1757.97±0.09
+Align+Enh.93.20±4.5082.53±0.0960.65±0.62
", + "bbox": [ + 81, + 487, + 470, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the reconstructed anchors and $10\\%$ of random anchors are included during the enhancement process. From the results, as the proportion of randomly generated anchors increases, the clustering performance on both datasets continuously deteriorates, thus validating the effectiveness of our enhanced anchors.", + "bbox": [ + 75, + 575, + 468, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the two properties of enhanced anchors, we calculate the values on both sides of the equations in the properties on two datasets. As shown in Fig. 3(b), the difference between the anchor graphs in the two views before and after anchor enhancement decreases $\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\tilde{\\mathbf{Z}}_2\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}_2\\right\\|_{\\mathbf{F}}\\right)$ , validating Property 1. Additionally, the difference between the anchor graph and the optimal consistent anchor graph reduces before and after anchor enhancement $\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}}\\right)$ , validating Property 2.", + "bbox": [ + 75, + 667, + 470, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Effectiveness of Anchor Graph Construction", + "text_level": 1, + "bbox": [ + 76, + 830, + 457, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the effectiveness of the revised anchor graph construction module, we contrast our proposed approach with two versions without the view weight and", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5f79e0224c7f545cc3586d8c56a5a1308f5e5294ab93cc71c4c9eea90663ab9.jpg", + "image_caption": [ + "(a) Validation of enhancement strategy" + ], + "image_footnote": [], + "bbox": [ + 501, + 464, + 696, + 575 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78325abf0eb9f3e1177bbc8ac387b91002c10a59a6c81a0f8e630be974867eab.jpg", + "image_caption": [ + "(b) Validation of Property 1 & 2", + "Figure 3. Validation of the proposed AEVC's effectiveness: (a) As the proportion of randomly generated anchors increases, the corresponding clustering performance decreases monotonically. (b) After anchor enhancement, the differences in anchor graphs between different views are reduced, and the enhanced anchor graph is closer to the optimal consensus anchor graph." + ], + "image_footnote": [], + "bbox": [ + 696, + 465, + 890, + 577 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8b99aa18815b8bc2050d73a6307198e24195b051b1d6b94096ccf29d4f118387.jpg", + "table_caption": [ + "Table 4. Ablation study on revised anchor graph construction module. Other results are provided in supplementary materials." + ], + "table_footnote": [], + "table_body": "
MethodsDermatologyForestTypesBDGP
w/o. view weight91.60±6.2681.45±0.0060.50±0.46
w/o. regularized term89.98±3.4079.75±0.1060.64±0.62
Proposed93.20±4.5082.53±0.0960.65±0.62
", + "bbox": [ + 504, + 738, + 901, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "regularized term respectively, as illustrated in Table 4. After incorporating weights learned from the views during fusion, there is a noticeable improvement in the clustering performance of the method. This highlights the advantage of assessing the importance of views in multi-view clustering", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26157", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0dc928d91d001417a42e14d27619922a2b7c318f088e2d662a5b0dd16e581f6b.jpg", + "table_caption": [ + "Table 5. Performance comparison between various anchor-based MVC w./w.o. our AEVC strategy." + ], + "table_footnote": [], + "table_body": "
MethodsDermatologyForestTypesBDGPReutersHdigitVGGFaceCIFAR100
Fixed Anchors Methods
FMCNOF62.01±0.0045.51±0.0031.08±0.0019.25±0.0032.84±0.003.47±0.004.40±0.00
FMCNOF-AE73.18±0.0069.60±0.0037.64±0.0022.79±0.0034.31±0.003.21±0.005.05±0.00
LMVSC79.02±0.0777.68±0.0449.52±0.0227.19±0.0186.87±0.086.09±0.008.92±0.00
LMVSC-AE88.23±7.8080.90±5.4952.13±0.3428.84±0.6785.74±5.726.83±0.168.81±0.16
Updated Anchors Methods
UDBGL84.08±0.0077.06±0.0039.28±0.0022.32±0.0031.39±0.005.39±0.007.83±0.00
UDBGL-AE81.01±0.0077.82±0.0046.92±0.0028.53±0.0044.60±0.005.65±0.008.57±0.00
FMVACC84.39±4.3479.29±0.2359.51±4.0027.22±0.2486.55±4.086.68±0.167.04±0.12
FMVACC-AE88.41±3.1381.08±0.2359.36±3.0333.12±2.4988.42±2.937.13±0.177.28±0.15
", + "bbox": [ + 122, + 113, + 854, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "through the proposed strategy. After removing the regularization term, there was a decline in performance across different datasets, which further confirms the effectiveness of constructing anchor graphs guided by $\\hat{\\mathbf{Z}}$ .", + "bbox": [ + 75, + 301, + 468, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Transferability Analysis", + "text_level": 1, + "bbox": [ + 76, + 373, + 299, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the transferability of AEVC, Table 5 presents a comparison of clustering accuracy between various types of anchor-based multi-view clustering methods with and without our anchor enhancement module. In this experiment, we selected two methods from each of the two categories employing different anchor strategies. Specifically, FMCNOF and LMVSC represent methods with fixed anchors, while UDBGL and FMVACC belong to methods with iteratively optimized anchors. As observed in the table, the clustering performance of the four anchor-based multi-view clustering methods experiences significant improvement across the majority of datasets after incorporating our proposed module. For instance, on the Reuters dataset, the clustering performance is notably enhanced by $18.4\\%$ , $6.1\\%$ , $27.8\\%$ , and $21.7\\%$ for FMCNOF, LMVSC, UDBGL, and FMVACC, respectively. This substantial improvement provides strong evidence for the effectiveness of AEVC when transferred to other anchor-based methods.", + "bbox": [ + 75, + 397, + 472, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Sensitivity Analysis", + "text_level": 1, + "bbox": [ + 76, + 680, + 264, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method involves three hyperparameters: the number of anchors $m$ , the anchor enhancement rate $\\gamma$ , and the hyperparameter for balancing the regularization term $\\lambda$ . Due to space constraints, we primarily analyze the sensitivity of the latter two parameters here, while the analysis regarding $m$ will be presented in the supplementary materials. To demonstrate the impact of $\\gamma$ and $\\lambda$ , we fix $m$ and conduct grid searches within specified ranges. Results presented in Fig. 4 indicate that our method is significantly influenced by the anchor enhancement rate, as the extreme rate can disrupt the balance between newly generated and original anchors. Specifically, on the Dermatology dataset, setting $\\gamma$ to 100 yields the highest clustering accuracy.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be366ee45049e67f42935c740ed0e7663931453db5efe43949f6d36133f3a076.jpg", + "image_caption": [ + "(a) Dermatology" + ], + "image_footnote": [], + "bbox": [ + 519, + 309, + 684, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8aa03ee39764fb58b9cdb1186f1d15d6918641c6ec84e8550c89b2fe8e1b7f0a.jpg", + "image_caption": [ + "(b) ForestTypes", + "Figure 4. Sensitivity analysis of $\\lambda$ and $\\gamma$ on two benchmark datasets. Others are displayed in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 718, + 309, + 883, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 515, + 617, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an Anchor Enhancement strategy with View Correlation for multi-view clustering termed AEVC. Different from existing anchor-based MVC methods, AEVC introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module. It can effectively enhance the quality of anchors, which is proven both experimentally and theoretically. Besides, we propose a revised anchor graph construction module based on the anchor enhancement strategy, which further improves the clustering performances. Experimental results show the promising capacity of AEVC, especially when integrating with other anchor-based methods. In the future, we will explore adaptive learning of enhancement rate.", + "bbox": [ + 496, + 547, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 797, + 669, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (No. 62325604), and the National Natural Science Foundation of China (No. 62276271 and No. 62306324).", + "bbox": [ + 496, + 830, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26158", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3", + "[2] Jie Chen, Hua Mao, Dezhong Peng, Changqing Zhang, and Xi Peng. Multiview clustering by consensus spectral rotation fusion. IEEE Transactions on Image Processing, 2023. 1", + "[3] Yongyong Chen, Xiaojia Zhao, Zheng Zhang, Youfa Liu, Jingyong Su, and Yicong Zhou. Tensor learning meets dynamic anchor learning: From complete to incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Siwei Wang, Xinzhong Zhu, Xinwang Liu, and En Zhu. Iterative deep structural graph contrast clustering for multiview raw data. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2023. 1", + "[5] Zhibin Dong, Siwei Wang, Jiaqi Jin, Xinwang Liu, and En Zhu. Cross-view topology based consistent and complementary information for deep multi-view clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19440-19451, 2023. 3", + "[6] Jürgen Dorn, Mario Girsch, Günther Skele, and Wolfgang Slany. Comparison of iterative improvement techniques for schedule optimization. European Journal of Operational Research, 94(2):349-361, 1996. 2, 3", + "[7] Si-Guo Fang, Dong Huang, Xiao-Sha Cai, Chang-Dong Wang, Chaobo He, and Yong Tang. Efficient multi-view clustering via unified and discrete bipartite graph learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2, 3, 6", + "[8] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2023. 1", + "[9] Wenjue He, Zheng Zhang, Yongyong Chen, and Jie Wen. Structured anchor-inferred graph learning for universal incomplete multi-view clustering. World Wide Web, 26(1): 375-399, 2023. 3", + "[10] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 2023. 1, 6", + "[11] Yanyong Huang, Zongxin Shen, Yuxin Cai, Xiwen Yi, Dongjie Wang, Fengmao Lv, and Tianrui Li. Imufs: Complementary and consensus learning-based incomplete multiview unsupervised feature selection. IEEE Transactions on Knowledge and Data Engineering, 2023. 1", + "[12] Jintian Ji and Songhe Feng. Anchor structure regularization induced multi-view subspace clustering via enhanced tensor rank minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19343-19352, 2023. 2", + "[13] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11600-11609, 2023. 1", + "[14] Zhao Kang, Wangtao Zhou, Zhitong Zhao, Junming Shao, Meng Han, and Zenglin Xu. Large-scale multi-view subspace clustering in linear time. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4412-4419, 2020. 1, 2, 4, 6", + "[15] Borhan Kazimipour, Xiaodong Li, and A Kai Qin. Initialization methods for large scale global optimization. In 2013 IEEE congress on evolutionary computation, pages 2750-2757. IEEE, 2013. 2, 3", + "[16] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 1", + "[17] Lusi Li and Haibo He. Bipartite graph based multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2020. 1", + "[18] Ruihuang Li, Changqing Zhang, Qinghua Hu, Pengfei Zhu, and Zheng Wang. Flexible multi-view representation learning for subspace clustering. In *IJCAI*, pages 2916–2922, 2019. 1", + "[19] Xuelong Li, Han Zhang, Rong Wang, and Feiping Nie. Multi-view clustering: A scalable and parameter-free bipartite graph fusion method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 3", + "[20] Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, and Fuchun Sun. Reasoning over different types of knowledge graphs: Static, temporal and multi-modal. arXiv preprint arXiv:2212.05767, 2022. 3", + "[21] Ke Liang, Yue Liu, Sihang Zhou, Wenxuan Tu, Yi Wen, Xihong Yang, Xiangjun Dong, and Xinwang Liu. Knowledge graph contrastive learning based on relation-symmetrical structure. IEEE Transactions on Knowledge and Data Engineering, pages 1-12, 2023. 3", + "[22] Ke Liang, Lingyuan Meng, Sihang Zhou, Siwei Wang, Wenxuan Tu, Yue Liu, Meng Liu, and Xinwang Liu. Message intercommunication for inductive relation reasoning. arXiv preprint arXiv:2305.14074, 2023. 3", + "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11174-11183, 2021. 2", + "[24] Jialu Liu, Chi Wang, Jing Gao, and Jiawei Han. Multiview clustering via joint nonnegative matrix factorization. In Proceedings of the 2013 SIAM International Conference on Data Mining, pages 252-260. SIAM, 2013. 1", + "[25] Suyuan Liu, Xinwang Liu, Siwei Wang, Xin Niu, and En Zhu. Fast incomplete multi-view clustering with view-independent anchors. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[26] Suyuan Liu, Siwei Wang, Pei Zhang, Kai Xu, Xinwang Liu, Changwang Zhang, and Feng Gao. Efficient one-pass multiview subspace clustering with consensus anchors. In Pro-" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26159", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the AAAI Conference on Artificial Intelligence, pages 7576-7584, 2022. 2, 3", + "[27] Suyuan Liu, Qing Liao, Siwei Wang, Xinwang Liu, and En Zhu. Robust and consistent anchor graph learning for multiview clustering. IEEE Transactions on Knowledge and Data Engineering, 2024. 1", + "[28] Xun Lu and Songhe Feng. Structure diversity-induced anchor graph fusion for multi-view clustering. ACM Transactions on Knowledge Discovery from Data, 17(2):1-18, 2023. 3", + "[29] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In International conference on machine learning, pages 5092-5101. PMLR, 2019. 2", + "[30] Qianyao Qiang, Bin Zhang, Fei Wang, and Feiping Nie. Fast multi-view discrete clustering with anchor graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9360-9367, 2021. 2", + "[31] Shaojun Shi, Feiping Nie, Rong Wang, and Xuelong Li. Fast multi-view clustering via prototype graph. IEEE Transactions on Knowledge and Data Engineering, 35(1):443-455, 2021. 1, 2", + "[32] Xiaochuang Shu, Xiangdong Zhang, Quanxue Gao, Ming Yang, Rong Wang, and Xinbo Gao. Self-weighted anchor graph learning for multi-view clustering. IEEE Transactions on Multimedia, 2022. 1", + "[33] Mengjing Sun, Pei Zhang, Siwei Wang, Sihang Zhou, Wenxuan Tu, Xinwang Liu, En Zhu, and Changjian Wang. Scalable multi-view subspace clustering with unified anchors. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3528-3536, 2021. 2, 3, 6", + "[34] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 2", + "[35] Xinhang Wan, Jiyuan Liu, Weixuan Liang, Xinwang Liu, Yi Wen, and En Zhu. Continual multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, page 3676-3684, New York, NY, USA, 2022. Association for Computing Machinery. 1", + "[36] Xinhang Wan, Xinwang Liu, Jiyuan Liu, Siwei Wang, Yi Wen, Weixuan Liang, En Zhu, Zhe Liu, and Lu Zhou. Autoweighted multi-view clustering for large-scale data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10078-10086, 2023. 1", + "[37] Jun Wang, Chang Tang, Zhiguo Wan, Wei Zhang, Kun Sun, and Albert Y Zomaya. Efficient and effective one-step multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2", + "[38] Siwei Wang, Xinwang Liu, Li Liu, Wenxuan Tu, Xinzhong Zhu, Jiyuan Liu, Sihang Zhou, and En Zhu. Highly-efficient incomplete large-scale multi-view clustering with consensus bipartite graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9776-9785, 2022. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 3, 6", + "[40] Siwei Wang, Xinwang Liu, Xinzhong Zhu, Pei Zhang, Yi Zhang, Feng Gao, and En Zhu. Fast parameter-free multiview subspace clustering with consensus anchor guidance. IEEE Trans. Image Process., 31:556-568, 2022. 3, 6", + "[41] Yang Wang, Lin Wu, Xuemin Lin, and Junbin Gao. Multiview spectral clustering via structured low-rank matrix factorization. IEEE transactions on neural networks and learning systems, 29(10):4833-4843, 2018. 1", + "[42] Yi Wen, Siwei Wang, Ke Liang, Weixuan Liang, Xinhang Wan, Xinwang Liu, Suyuan Liu, Jiyuan Liu, and En Zhu. Scalable incomplete multi-view clustering with structure alignment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3031-3040, 2023. 3", + "[43] Yi Wen, Siwei Wang, Qing Liao, Weixuan Liang, Ke Liang, Xinhang Wan, and Xinwang Liu. Unpaired multi-view graph clustering with cross-view structure matching. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[44] Wei Xia, Quanxue Gao, Qianqian Wang, Xinbo Gao, Chris Ding, and Dacheng Tao. Tensorized bipartite graph learning for multi-view clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3", + "[45] Weiqing Yan, Jindong Xu, Jinglei Liu, Guanghui Yue, and Chang Tang. Bipartite graph-based discriminative feature learning for multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3403-3411, 2022. 1, 2", + "[46] Ben Yang, Xuetao Zhang, Feiping Nie, Fei Wang, Weizhong Yu, and Rong Wang. Fast multi-view clustering via nonnegative and orthogonal factorization. IEEE Transactions on Image Processing, 30:2575-2586, 2020. 2, 3, 6", + "[47] Ben Yang, Xuetao Zhang, Zhiping Lin, Feiping Nie, Badong Chen, and Fei Wang. Efficient and robust multiview clustering with anchor graph regularization. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6200-6213, 2022. 1", + "[48] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 1", + "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2", + "[50] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In Proceedings of the 31st ACM International Conference on Multimedia, pages 337-346, 2023. 1", + "[51] Jiali You, Zhenwen Ren, Xiaojian You, Haoran Li, and Yuancheng Yao. Priori anchor labels supervised scalable multi-view bipartite graph clustering. In Proceedings of the" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26160", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AAAI Conference on Artificial Intelligence, pages 10972-10979, 2023. 3", + "[52] Shengju Yu, Siwei Wang, Yi Wen, Ziming Wang, Zhigang Luo, En Zhu, and Xinwang Liu. How to construct corresponding anchors for incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3", + "[53] Chao Zhang, Huaxiong Li, Wei Lv, Zizheng Huang, Yang Gao, and Chunlin Chen. Enhanced tensor low-rank and sparse representation recovery for incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11174-11182, 2023. 3", + "[54] Nan Zhang, Xiaoqin Zhang, and Shiliang Sun. Efficient multiview representation learning with correntropy and anchor graph. IEEE Transactions on Knowledge & Data Engineering, (01):1-14, 2023. 2", + "[55] Pei Zhang, Siwei Wang, Liang Li, Changwang Zhang, Xinwang Liu, En Zhu, Zhe Liu, Lu Zhou, and Lei Luo. Let the data choose: flexible and diverse anchor graph fusion for scalable multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11262-11269, 2023. 2, 6", + "[56] Xiaojia Zhao, Qiangqiang Shen, Yongyong Chen, Yongsheng Liang, Junxin Chen, and Yicong Zhou. Self-completed bipartite graph learning for fast incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + ], + "bbox": [ + 78, + 90, + 468, + 470 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26161", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_model.json b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bd562aec5c1ffbd71563ba1a7fe976ba4dc08bf5 --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.082, + 0.13, + 0.892, + 0.178 + ], + "angle": 0, + "content": "Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.202, + 0.951, + 0.257 + ], + "angle": 0, + "content": "Suyuan Liu1, Ke Liang1, Zhibin Dong1, Siwei Wang2, Xihong Yang1, Sihang Zhou1, En Zhu1, Xinwang Liu1* \n1National University of Defense Technology, Changsha, China \n2Intelligent Game and Decision Lab, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.337, + 0.26, + 0.692, + 0.275 + ], + "angle": 0, + "content": "{suyuanliu,xinwangliu,enzhu}@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.474, + 0.691 + ], + "angle": 0, + "content": "In recent years, anchor-based methods have achieved promising progress in multi-view clustering. The performances of these methods are significantly affected by the quality of the anchors. However, the anchors generated by previous works solely rely on single-view information, ignoring the correlation among different views. In particular, we observe that similar patterns are more likely to exist between similar views so such correlation information can be leveraged to enhance the quality of the anchors, which is also omitted. To this end, we propose a novel plug-and-play anchor enhancement strategy through view correlation for multi-view clustering. Specifically, we construct a view graph based on aligned initial anchor graphs to explore inter-view correlations. By learning from view correlation, we enhance the anchors of the current view using the relationships between anchors and samples on neighboring views, thereby narrowing the spatial distribution of anchors on similar views. Experimental results on seven datasets demonstrate the superiority of our proposed method over other existing methods. Furthermore, extensive comparative experiments validate the effectiveness of the proposed anchor enhancement module when applied to various anchor-based methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.723, + 0.21, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Large-scale scenarios have always been a hot concern with the data explosion in our real world. Among these scenarios, large-scale multi-view clustering (MVC) has also drawn increasing attention in recent years[2, 4, 8, 18, 35]. In general, large-scale MVC methods can be categorized into two types, i.e., matrix factorization-based methods and anchor-based methods. The former methods [11, 24, 41, 50] aim to generate a low-dimensional representation for further clustering by decomposing the original data, thereby avoiding the computational costs of similarity matrix construc" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.312, + 0.887, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.474, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Figure 1. An example illustration of clustering results affected by initial anchors. The first two methods keep the anchors fixed, while the latter two methods update the anchors through iteration-optimization. Their clustering performances are influenced to varying degrees by different anchor generation approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.701 + ], + "angle": 0, + "content": "tion. However, their efficiency usually decreases along with increasing dimension, thus constraining their scalability[10, 36]. Inspired by prototype learning [16, 48], anchors are introduced into MVC to reduce the scope for clustering from \\( n \\) to \\( m \\), which denotes the number of samples and anchors, respectively. Specifically, anchor-based MVC methods [13, 27, 45] construct an anchor graph \\( (n \\times m) \\) rather than a similarity matrix \\( (n \\times n) \\). It will alleviate the aforementioned computation and scalability problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.895, + 0.869 + ], + "angle": 0, + "content": "Anchor-based MVC methods typically consist of three procedures, including (1) anchor generation, (2) anchor graph construction, and (3) clustering. Early attempts usually generate anchors at first, which are fixed during constructing anchor graphs[31, 32, 47]. For example, SFMC [19] selects anchors by iterative sampling based on feature similarity. Besides, LMVSC [14] regards cluster centers generated by \\( k \\)-means of each view as the anchors. However, these methods are sensitive to the initial anchors (See Fig. 1), where the anchor selection is primitive and sometimes results in compromised performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "To address the above issues, researchers make efforts on anchor optimization to guarantee better quality anchors[17," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "26151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.333 + ], + "angle": 0, + "content": "26, 38]. For example, SMVSC [33] proposed a unified framework that combines anchor generation and anchor-graph construction, achieving iterative updates to optimize the anchors. Building upon this, UDBGL [7] introduced the Laplacian rank constraint to ensure that the anchor graph possesses a discrete cluster structure. The above iteration-optimization strategies actually aim to get a better combination of the anchor and its corresponding anchor graph instead of enhancing the quality of the anchors themselves. Meanwhile, to the best of our knowledge, almost all existing anchor optimization models belong to this type. Although they can get better and more stable clustering performance, they still suffer from the initial state for optimization[6, 15], which leaves us a great challenge as well as the opportunity, i.e., what is the new strategy to alleviate such problem? In this paper, we provide a potential solution to it." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.47, + 0.501 + ], + "angle": 0, + "content": "We observe that most anchor generation methods only rely on a single view[3, 12, 30]. Multiple works have proven that information of different views for the same sample will benefit the expressive ability of the sample [23, 29, 34, 49]. Similarly, different views for the same anchors should also be used to generate better anchors. Specifically, utilizing information across views and ensuring the consistency of anchor graphs on each view will benefit the anchor generation procedure. However, such a characteristic is not used for anchor enhancement in previous anchor-based MVC methods, and our strategy is developed based on it." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.503, + 0.473, + 0.82 + ], + "angle": 0, + "content": "In this work, we propose an Anchor Enhancement strategy with View Correlation (AEVC) for multi-view clustering (MVC). The framework of AEVC is shown in Fig. 2. The proposed approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which enhance the quality of anchors. Additionally, we improve the clustering performance by revising the existing anchor graph construction method based on view correlation. Specifically, we first align the initial anchors and construct a view-graph based on the aligned anchor graph to find the neighbors of each view. Then, new anchors for the current view are generated with the anchor graph of neighboring views, thereby enhancing the initial anchors. Finally, the inter-view correlations are combined to learn a consistent anchor graph between views with the enhanced anchors. AEVC optimizes anchors by absorbing information from similar views, improving initial anchor quality, and reducing distribution bias caused by the independent anchors generation in each view. The main contributions of our work can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.884 + ], + "angle": 0, + "content": "- We propose AEVC, a plug-and-play anchor enhancement strategy for MVC based on view correlation, to address the issue of poor initial anchor quality affecting the performance of MVC." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "- By learning from the view correlation, AEVC incorpo" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.09, + 0.896, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.301, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed AEVC. Specifically, in the view-graph learning module, AEVC primarily performs anchor alignment and learns the view-graph. In the anchor enhancement module, AEVC leverages the neighboring relationships from the view-graph to enhance the anchors. Finally, AEVC executes anchor graph construction based on the enhanced anchors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.892, + 0.476 + ], + "angle": 0, + "content": "rates the anchor distribution information from neighboring views into the process of anchor enhancement. Theoretical and experimental validations demonstrate that the anchors enhanced by AEVC exhibit a tighter inter-view relationship." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.551 + ], + "angle": 0, + "content": "- We compared our proposed approach to other anchor-based multi-view methods and demonstrated its superiority. Extensive experiments verify the effectiveness and generalization of the proposed AEVC strategy when integrated with typical anchor-based MVC methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.566, + 0.642, + 0.582 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.591, + 0.818, + 0.607 + ], + "angle": 0, + "content": "2.1. Anchor-based Multi-view Clustering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To address the challenges posed by traditional multiview clustering methods in handling large-scale datasets, Anchor-based MVC has been proposed in recent years[37, 45, 55]. The core idea of anchor-based MVC is to generate a set of representative anchors from the entire sample pool, where the number of anchors is typically much smaller than the total number of samples[25, 54]. In contrast to conventional methods that require establishing relationships among all samples, anchor-based approaches only involve computing similarities between anchors and samples. This significantly reduces the complexity associated with constructing the relationship graph. Additionally, adopting a reduced-dimensional graph further mitigates computational burdens, facilitating subsequent multi-view fusion and clustering with reduced computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Traditional anchor-based MVC methods generate anchors by sampling[30, 31, 46]. Subsequent anchor graph construction and clustering are performed based on the initially determined anchors. For instance, Kang et al.[14]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26152" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "and Yang et al. [46] advocate applying \\( k \\)-means to the original samples, selecting cluster centers as anchors to enhance their representativeness. Xia et al. [44] introduces a variance-based de-correlation anchor selection strategy, leveraging the internal data structure to augment the coverage capability of anchors. The performance of such methods is highly influenced by the quality of the initially generated anchors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.224, + 0.342, + 0.24 + ], + "angle": 0, + "content": "2.2. Anchor Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.249, + 0.473, + 0.491 + ], + "angle": 0, + "content": "To tackle the above challenges, extensive research efforts have been devoted to anchor optimization in order to achieve superior anchor quality. Iteration-optimization methods propose to update anchors and anchor graphs, utilizing the anchor graph obtained from the previous iteration to update the anchors[51, 53, 56]. Sun et al. [33] initially propose learning anchors and constructing anchor graphs through iterative updates within the same framework. Liu et al. [26] further introduce Laplacian rank constraints, enabling the anchor graph to generate labels without directly needing post-processing. However, the results of iterative update algorithms are significantly influenced by the initial values of variables[6, 15]. Instead of directly enhancing the quality of the anchors themselves, the above methods primarily focus on obtaining an improved combination of the anchor and its corresponding anchor graph." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.322, + 0.518 + ], + "angle": 0, + "content": "2.3. View Correlation Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.473, + 0.83 + ], + "angle": 0, + "content": "Multi-view and multi-modal scenarios are widely seen nowadays [1, 20]. Among them, investigating the relationships between multiple objects is a crucial topic [5, 21, 22], especially for multi-view learning, which holds for anchor-based multi-view clustering[7, 44]. Some existing methods delve into the consistency of views by aligning anchors[42, 43, 52]. For instance, Li et al. [19] concatenates raw data from all views to select anchors, ensuring correspondence among anchors across views. Wang et al. [39] takes a step further by proposing an anchor alignment framework based on the features and structural information of anchors. Another category of methods employs early fusion to avoid conflicts between views[9, 28]. For instance, Wang et al. [40] use a projection matrix to learn shared anchors and anchor graphs in the latent space. However, the first category of methods solely utilizes alignment information between views, while the second category underutilizes complementary information among views. In the next section, we will introduce a novel approach based on a view graph, leveraging inter-view interactions to enhance anchors." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.168, + 0.859 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce our proposed Anchor Enhancement strategy with View Correlation (AEVC) for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.166 + ], + "angle": 0, + "content": "multi-view clustering. We begin by sequentially describing the three modules of AEVC. Subsequently, a complexity analysis of the proposed algorithm is provided. Lastly, a theoretical analysis is conducted to evaluate the effectiveness of AEVC." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.18, + 0.61, + 0.194 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.204, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Existing anchor-based MVC methods can be divided into three steps: (1) anchor generation, (2) anchor graph construction, (3) clustering. The clustering performance of existing anchor-based MVC methods heavily relies on the quality of initially generated anchors. Although iteration-optimization strategies alleviate this issue by exploring combinations of anchors and anchor graphs, they are still impacted by the optimization of the initial state. To address this problem, we propose an anchor enhancement strategy based on view correlation for multi-view clustering. Specifically, different from existing methods, our approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which helps to enhance the quality of anchors. Furthermore, we revise the existing anchor graph construction method based on view correlation, further improving the clustering performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.474, + 0.704, + 0.49 + ], + "angle": 0, + "content": "3.2. View-graph Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Given multi-view data \\(\\{\\mathbf{X}_i\\}_{i=1}^v\\), where \\(\\mathbf{X}_i \\in \\mathbb{R}^{d_i \\times n}\\) is composed of \\(n\\) samples with \\(d_i\\) dimension. Denoting \\(\\mathbf{A}_i^{(0)} \\in \\mathbb{R}^{d_i \\times m}\\) as the generated anchor matrix in the \\(i\\)-th view, and \\(\\mathbf{Z}_i^{(0)} \\in \\mathbb{R}^{m \\times n}\\) represents the corresponding anchor graph, where \\(m\\) is the number of anchors. Due to the varying dimensions and inconsistent distributions of data across different views, it is not feasible to directly assess the distance between \\(\\mathbf{X}_i\\) and \\(\\mathbf{X}_j\\). However, it is easy to calculate the similarity between anchor graphs from different views due to the same size of anchor graphs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Nevertheless, the independently generated anchors on each view introduce the Anchor-Unaligned Problem [39]. This means that directly computing the similarity between anchor graphs of different views would be inaccurate due to the misalignment of rows corresponding to anchors. Thus, an anchor alignment operation must be performed before view-graph learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Existing anchor alignment strategies typically directly align the anchors on the first view, which carries the risk of being influenced by noise view. Here, we first concatenate the multi-view data column-wise to obtain \\(\\hat{\\mathbf{X}} = [\\mathbf{X}_1; \\ldots; \\mathbf{X}_v] \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times n}\\). Then, we generate the anchors \\(\\hat{\\mathbf{A}} \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times m}\\) and the corresponding anchor graph \\(\\hat{\\mathbf{Z}} \\in \\mathbb{R}^{m \\times n}\\). According to [39], we can obtain the permutation matrix \\(\\mathbf{P}_i\\) in the \\(i\\)-th view by minimizing the distance between \\(\\hat{\\mathbf{Z}}\\) and the permuted \\(\\mathbf{Z}_i^{(0)}\\), the formula is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.152, + 0.105 + ], + "angle": 0, + "content": "as follows," + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.116, + 0.469, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {P} _ {i}} \\left\\| \\hat {\\mathbf {Z}} - \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.149, + 0.414, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\mathrm {s . t .} \\mathbf {P} _ {i} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} ^ {\\top} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} \\in \\{0, 1 \\} ^ {m \\times m}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.177, + 0.469, + 0.207 + ], + "angle": 0, + "content": "Then the anchor matrix and anchor graph can be aligned with the permutation matrix as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.218, + 0.469, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\mathbf {A} _ {i} ^ {(0)} \\mathbf {P} _ {i} ^ {\\top}, \\\\ \\mathbf {Z} _ {i} = \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)}. \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.279, + 0.469, + 0.309 + ], + "angle": 0, + "content": "After performing the alignment operation, we calculate the view graph \\(\\mathbf{S} \\in \\mathbb{R}^{v \\times v}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.319, + 0.469, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\mathbf {s} _ {i j} = \\frac {1}{\\| \\mathbf {Z} _ {i} - \\mathbf {Z} _ {j} \\| _ {\\mathbf {F}} ^ {2}}, & i \\neq j, \\\\ \\mathbf {s} _ {i j} = 0, & i = j,. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.379, + 0.469, + 0.409 + ], + "angle": 0, + "content": "where \\(\\mathbf{s}_{ij}\\) denotes the similarity between the \\(i\\)-th and the \\(j\\)-th view." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.419, + 0.351, + 0.435 + ], + "angle": 0, + "content": "3.3. Anchor Enhancement Strategy" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.681 + ], + "angle": 0, + "content": "Most existing methods independently generate anchors on each view, ignoring the correlation between views. In fact, anchors on similar views should have similar distributions, which means that the corresponding anchor graphs should be close to each other. Based on this assumption, we propose to enhance the anchors of the current view with information on neighboring views. One intuitive idea is to directly update the anchors based on the anchors of neighboring views. However, the dimensions of anchors on different views are not the same, making direct transfer infeasible. Instead of utilizing feature consistency, we focus on structural consistency. Specifically, we reconstruct new anchors by leveraging the anchor graphs from neighboring views and the data from the current view, thereby enhancing the initial anchors. The formulation of the anchor enhancement on the \\(i\\)-th view is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.692, + 0.469, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {A}} _ {i} = \\frac {1}{\\mu} \\left(\\mathbf {A} _ {i} + \\gamma \\sum_ {\\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right)} \\hat {s} _ {i j} \\mathbf {X} _ {i} \\mathbf {Z} _ {j} ^ {\\top}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.742, + 0.469, + 0.802 + ], + "angle": 0, + "content": "where \\(\\mu = 1 + \\gamma \\sum_{\\mathbf{X}_j \\in \\Omega_\\zeta(\\mathbf{X}_i)} \\hat{s}_{ij}\\) ensures that the enhanced anchors are a convex combination of the initial anchors and the reconstructed anchors. \\(\\Omega_\\zeta(\\mathbf{X}_i)\\) denotes the \\(\\zeta\\)-neighbors of the \\(i\\)-th view, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.812, + 0.469, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right) = \\arg \\max _ {\\mathbf {X} _ {j}} \\sum_ {j = 1} | s _ {i j} |. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(\\gamma\\) is the enhancement rate, which is used to adjust the influence of the initial anchors on the enhanced version. \\(\\hat{\\mathbf{S}}\\) represents the normalized view-graph based on the following" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "settings, which controls the influence of different neighbors during anchor enhancement:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.132, + 0.891, + 0.176 + ], + "angle": 0, + "content": "\\[\n\\hat {s} _ {i j} \\left\\{ \\begin{array}{l l} s _ {i j}, & \\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right), \\\\ 0, & \\mathbf {X} _ {j} \\notin \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right). \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.185, + 0.892, + 0.334 + ], + "angle": 0, + "content": "By enhancing the anchors with Eq. (4), the quality of the initial anchors is improved. Moreover, the enhancement further achieves a better alignment between anchors on similar views. In the subsequent sections, we will analyze the aforementioned conclusions from both theoretical and experimental perspectives. Furthermore, our proposed anchor enhancement module is plug-and-play and we will showcase its performance improvement in the experiments by integrating different anchor-based multi-view clustering methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.345, + 0.755, + 0.361 + ], + "angle": 0, + "content": "3.4. Anchor Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.608 + ], + "angle": 0, + "content": "After obtaining high-quality anchors through our proposed anchor enhancement module, it is crucial to construct a reliable anchor graph that ensures view consistency. Since the enhanced anchors are aligned, we can directly generate a common anchor graph. However, the contributions of different views in constructing the anchor graph are not equal. To mitigate the influence of view discrepancies on building a consistent representation, we assign higher weights to views that have stronger associations with other views. Specifically, the contribution weight \\(\\omega_{i} = \\sum_{j=1}^{v} s_{ij}\\) for the \\(i\\)-th view is computed as the sum of its similarities with all other views. Furthermore, to avoid the impact of noise during anchor graph construction, we introduce the learned anchor graph \\(\\hat{\\mathbf{Z}}\\) from the concatenated views as guidance. Ultimately, the anchor graph construction is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.616, + 0.891, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {Z}} \\omega_ {i} \\left\\| \\mathbf {X} _ {i} - \\tilde {\\mathbf {A}} _ {i} \\mathbf {Z} \\right\\| _ {\\mathbf {F}} ^ {2} + \\lambda \\left\\| \\mathbf {Z} - \\hat {\\mathbf {Z}} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.648, + 0.767, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {s . t .} \\mathbf {Z} \\geq 0, \\mathbf {Z} ^ {\\top} \\mathbf {1} = \\mathbf {1}. \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.675, + 0.891, + 0.735 + ], + "angle": 0, + "content": "Due to the independence between each column of \\(\\mathbf{Z}\\), let \\(z_{j}\\) denote the \\(j\\)-th column of \\(\\mathbf{Z}\\). The aforementioned problem can be transformed into \\(n\\) Quadratic Programming(QP) problems, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.744, + 0.891, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\min \\frac {1}{2} \\mathbf {z} _ {j} ^ {\\top} \\mathbf {H} \\mathbf {z} _ {j} + f \\mathbf {z} _ {j}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.766, + 0.769, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {s . t .} \\mathbf {z} _ {j} \\geq 0, \\mathbf {z} _ {j} ^ {\\top} \\mathbf {1} = 1, \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where the matrix \\(\\mathbf{H} = 2\\left(\\sum_{i=1}^{v}\\omega_{i}\\tilde{\\mathbf{A}}_{i}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\mathbf{I}\\right)\\) and the vector \\(f = -2\\left(\\sum_{i=1}^{v}\\omega_{i}\\mathbf{X}_{i[:,j]}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\hat{z}_{j}\\right)\\). The optimal \\(\\mathbf{Z}\\) can be obtained by solving the above QP problem. Finally, the results can be generated by performing \\(k\\)-means on the left singular vector of \\(\\mathbf{Z}\\) [14]. The entire procedure of AEVC is shown in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.088, + 0.095, + 0.325, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: The proposed AEVC" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.113, + 0.443, + 0.158 + ], + "angle": 0, + "content": "Input: Multi-view datasets \\(\\{\\mathbf{X}_i\\}_{i = 1}^v\\) , the number of anchors \\(m\\) , the parameter \\(\\gamma\\) \\(\\lambda\\) and the number of clusters \\(k\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.159, + 0.281, + 0.174 + ], + "angle": 0, + "content": "Output: Clustering results." + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.178, + 0.38, + 0.212 + ], + "angle": 0, + "content": "1 Initialize \\(\\{\\mathbf{A}_i^{(0)}\\}_{i = 1}^v\\) \\(\\hat{\\mathbf{A}}\\) \\(\\{\\mathbf{Z}_i^{(0)}\\}_{i = 1}^v\\) and \\(\\hat{\\mathbf{Z}}\\) // View-graph Learning" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.213, + 0.218, + 0.226 + ], + "angle": 0, + "content": "2 for \\(i = 1\\to v\\) do" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.227, + 0.428, + 0.259 + ], + "angle": 0, + "content": "3 Calculate permutation matrix \\(\\mathbf{P}_i\\) with Eq. (1); 4 Perform alignment with Eq. (2);" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.262, + 0.374, + 0.291 + ], + "angle": 0, + "content": "5 Learn view-graph S as defined in Eq. (3); // Anchor Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.293, + 0.218, + 0.306 + ], + "angle": 0, + "content": "6 for \\(i = 1\\to v\\) do" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.308, + 0.434, + 0.325 + ], + "angle": 0, + "content": "7 Find \\(\\zeta\\)-neighbors \\(\\Omega_{\\zeta}(\\mathbf{X}_i)\\) as defined in Eq. (5);" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.328, + 0.421, + 0.344 + ], + "angle": 0, + "content": "8 Calculate normalized view-graph \\(\\hat{\\mathbf{S}}\\) with Eq. (6);" + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.345, + 0.218, + 0.357 + ], + "angle": 0, + "content": "9 for \\(i = 1\\rightarrow v\\) do" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.359, + 0.295, + 0.375 + ], + "angle": 0, + "content": "10 Enhance \\(\\mathbf{A}_i\\) with Eq. (4);" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.178, + 0.434, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.38, + 0.375, + 0.393 + ], + "angle": 0, + "content": "// Anchor Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.395, + 0.289, + 0.409 + ], + "angle": 0, + "content": "11 Obtain \\(\\mathbf{Z}\\) by solving Eq. (7);" + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.41, + 0.423, + 0.424 + ], + "angle": 0, + "content": "12 Perform \\(k\\)-means on the left singular vector of \\(\\mathbf{Z}\\)." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.38, + 0.423, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.46, + 0.195, + 0.473 + ], + "angle": 0, + "content": "3.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.725 + ], + "angle": 0, + "content": "Computational complexity. The computational complexity of the proposed AEVC mainly consists of three modules. The complexity of the view-graph learning module is \\(\\mathcal{O}(m^3 v + nmv)\\). Specifically, the computation of \\(v\\) permutation matrices \\(\\mathbf{P}_i\\) has a complexity of \\(\\mathcal{O}(m^3 v)\\), and the computation of the view graph has a complexity of \\(\\mathcal{O}(nmv)\\). The time complexity of the anchor enhancement module is \\(\\mathcal{O}(nmdv)\\), where \\(d = \\sum_{i=1}^{v} d_v\\), primarily arising from matrix multiplication during enhancement. The complexity of the anchor graph construction module is \\(\\mathcal{O}(nm^3 v + nmdv)\\), where solving all QP problems requires a time complexity of \\(\\mathcal{O}(nm^3 v)\\), and the involved matrix multiplication requires \\(\\mathcal{O}(nm^3 v + nmdv)\\). Therefore, the overall computational complexity of AEVC is \\(\\mathcal{O}(nm^3 v + m^3 v + nmdv + nmv)\\), which exhibits linear complexity with respect to the number of samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.469, + 0.816 + ], + "angle": 0, + "content": "Space complexity. During the execution of the AEVC algorithm, the main variables that need to be stored include matrices \\(\\mathbf{X}_i\\), \\(\\mathbf{A}_i\\), \\(\\mathbf{Z}_i\\), \\(\\hat{\\mathbf{A}}\\), \\(\\hat{\\mathbf{Z}}\\), \\(\\mathbf{P}_i\\), \\(\\mathbf{S}\\), and \\(\\hat{\\mathbf{S}}\\). The overall space complexity of AEVC is \\(\\mathcal{O}(nd + ndv + nmdv + v^2)\\), which is also linear with respect to the number of samples and can be extended to large-scale scenarios." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.817, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Two Property of enhanced anchors. We have discovered that the anchors enhanced by the proposed AEVC exhibit stronger inter-view consistency. The theoretical analysis is provided below." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Assumption 1. Given a two-view data \\(\\{\\mathbf{X}_1, \\mathbf{X}_2\\}\\), denoting" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.138 + ], + "angle": 0, + "content": "\\(\\{\\mathbf{A}_1,\\mathbf{A}_2\\}\\) are the generated initial anchors, and \\(\\{\\mathbf{Z}_1,\\mathbf{Z}_2\\}\\) are the corresponding anchor graphs under the assumption that \\(\\mathbf{X}_1 = \\mathbf{A}_1\\mathbf{Z}_1\\) and \\(\\mathbf{X}_2 = \\mathbf{A}_2\\mathbf{Z}_2\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.145, + 0.892, + 0.189 + ], + "angle": 0, + "content": "For the convenience of derivation, we simplify the anchor enhancement strategy in Eq. (4) with \\(\\alpha \\in (0,1)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.196, + 0.892, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {A}} _ {1} = \\alpha \\mathbf {A} _ {1} + (1 - \\alpha) \\mathbf {X} _ {1} \\mathbf {Z} _ {2} ^ {\\top}, \\\\ \\tilde {\\mathbf {A}} _ {2} = \\alpha \\mathbf {A} _ {2} + (1 - \\alpha) \\mathbf {X} _ {2} \\mathbf {Z} _ {1} ^ {\\top}. \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.247, + 0.892, + 0.293 + ], + "angle": 0, + "content": "The corresponding anchor graphs \\(\\{\\tilde{\\mathbf{Z}}_1,\\tilde{\\mathbf{Z}}_2\\}\\) are constructed under the same assumption that \\(\\mathbf{X}_1 = \\tilde{\\mathbf{A}}_1\\tilde{\\mathbf{Z}}_1\\) and \\(\\mathbf{X}_2 = \\tilde{\\mathbf{A}}_2\\tilde{\\mathbf{Z}}_2\\). Then we can derive that" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.301, + 0.891, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {Z}} _ {1} = \\alpha \\mathbf {Z} _ {1} + (1 - \\alpha) \\mathbf {Z} _ {2}, \\\\ \\tilde {\\mathbf {Z}} _ {2} = \\alpha \\mathbf {Z} _ {2} + (1 - \\alpha) \\mathbf {Z} _ {1}. \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.35, + 0.788, + 0.366 + ], + "angle": 0, + "content": "We can deduce the following properties:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.374, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Property 1. The disparity among the anchor graphs on each view is reduced after anchor enhancement. Mathematically," + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.426, + 0.892, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.461, + 0.825, + 0.477 + ], + "angle": 0, + "content": "Proof. According to Eq. (10), we can derive that" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.484, + 0.891, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} = \\| (2 \\alpha - 1) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = \\left| 2 \\alpha - 1 \\right| \\cdot \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}} \\tag {12} \\\\ < \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.558, + 0.671, + 0.573 + ], + "angle": 0, + "content": "This completes the proof." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.579, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Property 1 implies that anchor enhancement leads to a more consistent distribution of anchors in similar views, thereby establishing a tighter alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Property 2. Assuming there exists a value \\(\\beta \\in (0,1)\\) such that \\(\\mathbf{Z} = \\beta \\mathbf{Z}_1 + (1 - \\beta)\\mathbf{Z}_2\\) is the optimal consistent anchor graph among all views, the enhanced anchor graph is closer to the optimal version after enhancement than before for all \\(\\alpha \\in (2\\beta -1,1)\\cap (0,1)\\). Mathematically," + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.716, + 0.891, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\| _ {\\mathbf {F}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.752, + 0.825, + 0.767 + ], + "angle": 0, + "content": "Proof. According to Eq. (10), we can derive that" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.775, + 0.891, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} = \\| (\\alpha - \\beta) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = | \\alpha - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\\\ < | 1 - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\tag {14} \\\\ = \\left\\| (1 - \\beta) \\left(\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}\\right) \\right\\| _ {\\mathbf {F}} \\\\ = \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.671, + 0.901 + ], + "angle": 0, + "content": "This completes the proof." + }, + { + "type": "image", + "bbox": [ + 0.875, + 0.887, + 0.892, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26155" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "Property 2 implies that anchor enhancement helps reduce the gap between the anchor graph on each view and the optimal consistent anchor graph. It is easy to derive the multiview extended versions of the above properties." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.17, + 0.202, + 0.187 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.254, + 0.213 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.221, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Datasets. Seven multi-view datasets are employed in our experiments: Dermatology1 contains diagnostic information for six types of skin diseases. ForestTypes2 consists of satellite images of different forest types. BDGP3 comprises both images and textual descriptions of Drosophila embryos. Reuters4 is a news text dataset. Hdigit5 consists of a set of handwritten digital images. VGGFace2 is a large-scale face recognition dataset. CIFAR100 is a subset of a large-scale tiny image dataset. Detail description is shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.471, + 0.616 + ], + "angle": 0, + "content": "Compared methods. We compare the proposed AEVC with existing state-of-the-art anchor-based MVC methods: Large-scale multi-view subspace clustering in linear time (LMVSC)[14]; Scalable Multi-view Subspace Clustering with Unified Anchors (SMVSC)[33]; Fast multi-view clustering via nonnegative and orthogonal factorization (FMCNOF)[46]; Fast Parameter-Free Multi-View Subspace Clustering With Consensus Anchor Guidance (FPMVSCAG)[40]; Efficient Multi-View Clustering via Unified and Discrete Bipartite Graph Learning (UDBGL)[7]; Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity (FastMICE)[10]; Flexible and Diverse Anchor Graph Fusion for Scalable Multi-View Clustering (FDAGF)[55]; Align then Fusion: Generalized Large-scale Multi-view Clustering with Anchor Matching Correspondences (FMVACC)[39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.799 + ], + "angle": 0, + "content": "Implementation details. For the above-compared algorithm, we have tuned their parameters to align with the relevant literature. We conducted a search for three hyperparameters in the proposed AEVC method using predefined lists. Specifically, the anchor number \\( m \\) is chosen from the set \\( [k, 2k, 5k] \\), where \\( k \\) represents the cluster number. The enhancement rate \\( \\gamma \\) is selected from \\( [0.1, 1, 10, 100] \\), and the balancing hyperparameter \\( \\lambda \\) is chosen from the range \\( [0.0001, 0.01, 1, 100] \\). To ensure a fair comparison, algorithms relying on \\( k \\)-means for result acquisition underwent 50 independent runs, and the results were averaged for accuracy assessment. Our experimental evaluation incorporated" + }, + { + "type": "table_caption", + "bbox": [ + 0.573, + 0.09, + 0.819, + 0.103 + ], + "angle": 0, + "content": "Table 1. Benchmark datasets description." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.105, + 0.888, + 0.224 + ], + "angle": 0, + "content": "
Datasetnvkdp(p=1,...,v)
Dermatology3582612/22
ForestTypes523349/9/9
BDGP2500351000/500/250
Reuters7200564819/4810/4892/4858/4777
Hdigit10000210784/256
VGGFace23628741003944/576/512/640
CIFAR100500004100944/576/512/640
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.268 + ], + "angle": 0, + "content": "four distinct metrics: clustering Accuracy (ACC), Normalized Mutual Information (NMI), Purity, and F-score." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.278, + 0.7, + 0.294 + ], + "angle": 0, + "content": "4.2. Superiority of AEVC" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.302, + 0.892, + 0.361 + ], + "angle": 0, + "content": "We compare our proposed AEVC with eight state-of-the-art anchor-based multi-view clustering methods on seven datasets, as illustrated in Table 2. From the results, we have the following observations:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.366, + 0.892, + 0.455 + ], + "angle": 0, + "content": "- The experimental results reveal that our proposed algorithm consistently outperforms other anchor-based MVC methods across all datasets, confirming the superiority of AEVC. In terms of ACC, AEVC surpasses the second-best algorithm by \\(6.9\\%\\), \\(4.1\\%\\), \\(1.9\\%\\), \\(9.9\\%\\), \\(1.3\\%\\), \\(15.8\\%\\), and \\(9.5\\%\\) on different datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.456, + 0.892, + 0.545 + ], + "angle": 0, + "content": "- Compared to previous methods based on iteratively optimized anchors (SMVSC, FPMVS-CAG, UDBGL, FDAGF, FMVACC), AEVC exhibits superior clustering performance, emphasizing the effectiveness of our anchor enhancement approach over iteration-optimization methods." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.366, + 0.892, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.807, + 0.575 + ], + "angle": 0, + "content": "4.3. Effectiveness of Enhanced Anchors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.825 + ], + "angle": 0, + "content": "We carried out an ablation study to demonstrate the effectiveness of the proposed anchor enhancement strategy. In our approach, the anchor enhancement strategy typically comprises two steps: anchor alignment in the view-graph learning module and anchor enhancement module. We denote the method of removing the entire anchor enhancement strategy as a baseline and compare it with those where anchor alignment and enhancement are sequentially incorporated. From the experimental results shown in Table 3, it is observed that anchor alignment leads to an improvement in clustering performance, as misalignment of anchors can result in incorrect multi-view fusion[39]. Furthermore, with the application of our anchor enhancement strategy, there is a further improvement in clustering performance, providing strong evidence for the effectiveness of the anchor enhancement strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of the enhanced anchors, we incorporate randomly generated anchors from the original data at different proportions and showcase the corresponding clustering performance in Fig. 3(a). For instance, when the proportion is \\(10\\%\\), it means that \\(90\\%\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.813, + 0.375, + 0.827 + ], + "angle": 0, + "content": "1 http://archive.ics.uci.edu/ml/datasets/Dermatology" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.827, + 0.307, + 0.838 + ], + "angle": 0, + "content": "2http://archive.ics.uci.edu/dataset/333/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.838, + 0.238, + 0.851 + ], + "angle": 0, + "content": "3 https://www.fruitfly.org/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.851, + 0.414, + 0.863 + ], + "angle": 0, + "content": "4http://www.researchAtt.com/%7Elewis/reuters21578.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.863, + 0.319, + 0.876 + ], + "angle": 0, + "content": "5https://cs.nyu.edu/%7Erowei/data.html" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.876, + 0.367, + 0.888 + ], + "angle": 0, + "content": "\\(^{6}\\)https://www.robots.ox.ac.uk/vgg/data/vgg_face2/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.888, + 0.355, + 0.9 + ], + "angle": 0, + "content": "7https://www.cs.toronto.edu/%7Ekriz/cifar.html" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.813, + 0.414, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26156" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 2. Clustering performance comparison of eight anchor-based MVC methods on seven datasets. The best is marked in bold and underlined, the second best is marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.128, + 0.892, + 0.44 + ], + "angle": 0, + "content": "
DatasetsLMVSCSMVSCFMCNOFFPMVS-CAGUDBGLFastMICEFDAGFFMVACCProposed
ACC
Dermatology79.02±0.0778.64±0.0562.01±0.0082.96±0.0784.08±0.0087.15±0.0084.41±7.9084.39±4.3493.20±4.50
ForestTypes77.68±0.0472.82±0.0145.51±0.0072.22±0.0577.06±0.0077.06±0.0076.40±7.1379.29±0.2382.53±0.09
BDGP49.52±0.0237.22±0.0231.08±0.0032.62±0.0139.28±0.0049.00±0.0047.30±3.1659.51±4.0060.65±0.62
Reuters27.19±0.0124.41±0.0119.25±0.0017.33±0.0022.32±0.0026.00±0.0025.04±0.5027.22±0.2429.91±0.02
Hdigit86.87±0.0865.99±0.0332.84±0.0064.22±0.0531.39±0.0085.13±0.0077.69±7.8386.55±4.0888.01±2.97
VGGFace6.09±0.007.21±0.003.47±0.006.02±0.005.39±0.005.26±0.006.50±0.336.68±0.168.35±0.07
CIFAR1008.92±0.008.48±0.004.40±0.007.46±0.007.83±0.009.81±0.007.47±0.137.04±0.1210.74±0.08
NMI
Dermatology70.17±0.0466.62±0.0354.24±0.0071.90±0.0587.41±0.0082.73±0.0079.70±4.3476.27±2.1988.73±3.01
ForestTypes53.56±0.0246.56±0.0116.99±0.0048.20±0.0452.04±0.0056.53±0.0054.80±5.3555.33±0.1559.63±0.09
BDGP25.85±0.029.85±0.0110.29±0.0010.02±0.0114.90±0.0026.03±0.0021.90±1.9535.55±4.5330.92±0.65
Reuters5.67±0.015.45±0.010.64±0.000.13±0.001.79±0.007.23±0.003.36±0.485.05±0.087.07±0.01
Hdigit89.33±0.0356.16±0.0228.57±0.0055.80±0.0319.04±0.0088.67±0.0070.00±2.6076.96±3.0978.09±1.59
VGGFace11.92±0.0014.13±0.005.81±0.0012.33±0.0011.00±0.0010.79±0.0012.32±0.3312.97±0.1514.33±0.11
CIFAR10015.49±0.0014.83±0.008.49±0.0013.87±0.0014.83±0.0016.82±0.0013.93±0.1511.45±0.1616.94±0.14
Purity
Dermatology80.97±0.0480.35±0.0462.85±0.0083.55±0.0784.64±0.0087.15±0.0087.86±4.5185.38±2.0494.03±2.35
ForestTypes77.99±0.0272.82±0.0145.51±0.0072.73±0.0477.06±0.0078.93±0.0077.61±4.7479.29±0.2382.53±0.09
BDGP49.64±0.0237.80±0.0133.32±0.0034.82±0.0140.16±0.0049.44±0.0047.36±3.0859.73±3.6760.65±0.62
Reuters28.22±0.0025.12±0.0119.78±0.0017.89±0.0023.31±0.0028.07±0.0025.45±0.5927.50±0.1730.84±0.02
Hdigit88.93±0.0666.28±0.0332.85±0.0064.34±0.0531.94±0.0088.43±0.0079.82±5.6686.59±4.0188.22±2.25
VGGFace7.02±0.007.73±0.003.50±0.006.33±0.005.77±0.006.11±0.007.17±0.337.47±0.179.43±0.08
CIFAR10010.41±0.009.09±0.004.49±0.007.84±0.008.69±0.0011.11±0.008.28±0.137.89±0.1412.29±0.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.463, + 0.443, + 0.478 + ], + "angle": 0, + "content": "Table 3. Ablation study on anchor enhancement module." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.488, + 0.472, + 0.551 + ], + "angle": 0, + "content": "
MethodsDermatologyForestTypesBDGP
Baseline90.87±3.6070.76±0.0554.38±0.10
+Align91.82±3.7875.51±0.1757.97±0.09
+Align+Enh.93.20±4.5082.53±0.0960.65±0.62
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.667 + ], + "angle": 0, + "content": "of the reconstructed anchors and \\(10\\%\\) of random anchors are included during the enhancement process. From the results, as the proportion of randomly generated anchors increases, the clustering performance on both datasets continuously deteriorates, thus validating the effectiveness of our enhanced anchors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.471, + 0.822 + ], + "angle": 0, + "content": "To verify the two properties of enhanced anchors, we calculate the values on both sides of the equations in the properties on two datasets. As shown in Fig. 3(b), the difference between the anchor graphs in the two views before and after anchor enhancement decreases \\(\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\tilde{\\mathbf{Z}}_2\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}_2\\right\\|_{\\mathbf{F}}\\right)\\), validating Property 1. Additionally, the difference between the anchor graph and the optimal consistent anchor graph reduces before and after anchor enhancement \\(\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}}\\right)\\), validating Property 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.458, + 0.849 + ], + "angle": 0, + "content": "4.4. Effectiveness of Anchor Graph Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To further validate the effectiveness of the revised anchor graph construction module, we contrast our proposed approach with two versions without the view weight and" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.465, + 0.697, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.581, + 0.692, + 0.593 + ], + "angle": 0, + "content": "(a) Validation of enhancement strategy" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.466, + 0.892, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.581, + 0.871, + 0.593 + ], + "angle": 0, + "content": "(b) Validation of Property 1 & 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.604, + 0.895, + 0.689 + ], + "angle": 0, + "content": "Figure 3. Validation of the proposed AEVC's effectiveness: (a) As the proportion of randomly generated anchors increases, the corresponding clustering performance decreases monotonically. (b) After anchor enhancement, the differences in anchor graphs between different views are reduced, and the enhanced anchor graph is closer to the optimal consensus anchor graph." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.7, + 0.892, + 0.731 + ], + "angle": 0, + "content": "Table 4. Ablation study on revised anchor graph construction module. Other results are provided in supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.739, + 0.902, + 0.796 + ], + "angle": 0, + "content": "
MethodsDermatologyForestTypesBDGP
w/o. view weight91.60±6.2681.45±0.0060.50±0.46
w/o. regularized term89.98±3.4079.75±0.1060.64±0.62
Proposed93.20±4.5082.53±0.0960.65±0.62
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "regularized term respectively, as illustrated in Table 4. After incorporating weights learned from the views during fusion, there is a noticeable improvement in the clustering performance of the method. This highlights the advantage of assessing the importance of views in multi-view clustering" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26157" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.191, + 0.09, + 0.777, + 0.104 + ], + "angle": 0, + "content": "Table 5. Performance comparison between various anchor-based MVC w./w.o. our AEVC strategy." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.114, + 0.855, + 0.278 + ], + "angle": 0, + "content": "
MethodsDermatologyForestTypesBDGPReutersHdigitVGGFaceCIFAR100
Fixed Anchors Methods
FMCNOF62.01±0.0045.51±0.0031.08±0.0019.25±0.0032.84±0.003.47±0.004.40±0.00
FMCNOF-AE73.18±0.0069.60±0.0037.64±0.0022.79±0.0034.31±0.003.21±0.005.05±0.00
LMVSC79.02±0.0777.68±0.0449.52±0.0227.19±0.0186.87±0.086.09±0.008.92±0.00
LMVSC-AE88.23±7.8080.90±5.4952.13±0.3428.84±0.6785.74±5.726.83±0.168.81±0.16
Updated Anchors Methods
UDBGL84.08±0.0077.06±0.0039.28±0.0022.32±0.0031.39±0.005.39±0.007.83±0.00
UDBGL-AE81.01±0.0077.82±0.0046.92±0.0028.53±0.0044.60±0.005.65±0.008.57±0.00
FMVACC84.39±4.3479.29±0.2359.51±4.0027.22±0.2486.55±4.086.68±0.167.04±0.12
FMVACC-AE88.41±3.1381.08±0.2359.36±3.0333.12±2.4988.42±2.937.13±0.177.28±0.15
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.47, + 0.362 + ], + "angle": 0, + "content": "through the proposed strategy. After removing the regularization term, there was a decline in performance across different datasets, which further confirms the effectiveness of constructing anchor graphs guided by \\(\\hat{\\mathbf{Z}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.374, + 0.3, + 0.39 + ], + "angle": 0, + "content": "4.5. Transferability Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.473, + 0.671 + ], + "angle": 0, + "content": "To validate the transferability of AEVC, Table 5 presents a comparison of clustering accuracy between various types of anchor-based multi-view clustering methods with and without our anchor enhancement module. In this experiment, we selected two methods from each of the two categories employing different anchor strategies. Specifically, FMCNOF and LMVSC represent methods with fixed anchors, while UDBGL and FMVACC belong to methods with iteratively optimized anchors. As observed in the table, the clustering performance of the four anchor-based multi-view clustering methods experiences significant improvement across the majority of datasets after incorporating our proposed module. For instance, on the Reuters dataset, the clustering performance is notably enhanced by \\(18.4\\%\\), \\(6.1\\%\\), \\(27.8\\%\\), and \\(21.7\\%\\) for FMCNOF, LMVSC, UDBGL, and FMVACC, respectively. This substantial improvement provides strong evidence for the effectiveness of AEVC when transferred to other anchor-based methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.265, + 0.697 + ], + "angle": 0, + "content": "4.6. Sensitivity Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Our proposed method involves three hyperparameters: the number of anchors \\( m \\), the anchor enhancement rate \\( \\gamma \\), and the hyperparameter for balancing the regularization term \\( \\lambda \\). Due to space constraints, we primarily analyze the sensitivity of the latter two parameters here, while the analysis regarding \\( m \\) will be presented in the supplementary materials. To demonstrate the impact of \\( \\gamma \\) and \\( \\lambda \\), we fix \\( m \\) and conduct grid searches within specified ranges. Results presented in Fig. 4 indicate that our method is significantly influenced by the anchor enhancement rate, as the extreme rate can disrupt the balance between newly generated and original anchors. Specifically, on the Dermatology dataset, setting \\( \\gamma \\) to 100 yields the highest clustering accuracy." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.31, + 0.685, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.423, + 0.645, + 0.435 + ], + "angle": 0, + "content": "(a) Dermatology" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.31, + 0.884, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.423, + 0.846, + 0.435 + ], + "angle": 0, + "content": "(b) ForestTypes" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.446, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Figure 4. Sensitivity analysis of \\(\\lambda\\) and \\(\\gamma\\) on two benchmark datasets. Others are displayed in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.516, + 0.619, + 0.531 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.761 + ], + "angle": 0, + "content": "In this paper, we propose an Anchor Enhancement strategy with View Correlation for multi-view clustering termed AEVC. Different from existing anchor-based MVC methods, AEVC introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module. It can effectively enhance the quality of anchors, which is proven both experimentally and theoretically. Besides, we propose a revised anchor graph construction module based on the anchor enhancement strategy, which further improves the clustering performances. Experimental results show the promising capacity of AEVC, especially when integrating with other anchor-based methods. In the future, we will explore adaptive learning of enhancement rate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.799, + 0.671, + 0.816 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.831, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (No. 62325604), and the National Natural Science Foundation of China (No. 62276271 and No. 62306324)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26158" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Jie Chen, Hua Mao, Dezhong Peng, Changqing Zhang, and Xi Peng. Multiview clustering by consensus spectral rotation fusion. IEEE Transactions on Image Processing, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.282 + ], + "angle": 0, + "content": "[3] Yongyong Chen, Xiaojia Zhao, Zheng Zhang, Youfa Liu, Jingyong Su, and Yicong Zhou. Tensor learning meets dynamic anchor learning: From complete to incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.354 + ], + "angle": 0, + "content": "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Siwei Wang, Xinzhong Zhu, Xinwang Liu, and En Zhu. Iterative deep structural graph contrast clustering for multiview raw data. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[5] Zhibin Dong, Siwei Wang, Jiaqi Jin, Xinwang Liu, and En Zhu. Cross-view topology based consistent and complementary information for deep multi-view clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19440-19451, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[6] Jürgen Dorn, Mario Girsch, Günther Skele, and Wolfgang Slany. Comparison of iterative improvement techniques for schedule optimization. European Journal of Operational Research, 94(2):349-361, 1996. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.549 + ], + "angle": 0, + "content": "[7] Si-Guo Fang, Dong Huang, Xiao-Sha Cai, Chang-Dong Wang, Chaobo He, and Yong Tang. Efficient multi-view clustering via unified and discrete bipartite graph learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[8] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Wenjue He, Zheng Zhang, Yongyong Chen, and Jie Wen. Structured anchor-inferred graph learning for universal incomplete multi-view clustering. World Wide Web, 26(1): 375-399, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[11] Yanyong Huang, Zongxin Shen, Yuxin Cai, Xiwen Yi, Dongjie Wang, Fengmao Lv, and Tianrui Li. Imufs: Complementary and consensus learning-based incomplete multiview unsupervised feature selection. IEEE Transactions on Knowledge and Data Engineering, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[12] Jintian Ji and Songhe Feng. Anchor structure regularization induced multi-view subspace clustering via enhanced tensor rank minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19343-19352, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[13] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11600-11609, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[14] Zhao Kang, Wangtao Zhou, Zhitong Zhao, Junming Shao, Meng Han, and Zenglin Xu. Large-scale multi-view subspace clustering in linear time. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4412-4419, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[15] Borhan Kazimipour, Xiaodong Li, and A Kai Qin. Initialization methods for large scale global optimization. In 2013 IEEE congress on evolutionary computation, pages 2750-2757. IEEE, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.319 + ], + "angle": 0, + "content": "[16] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.894, + 0.362 + ], + "angle": 0, + "content": "[17] Lusi Li and Haibo He. Bipartite graph based multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.894, + 0.417 + ], + "angle": 0, + "content": "[18] Ruihuang Li, Changqing Zhang, Qinghua Hu, Pengfei Zhu, and Zheng Wang. Flexible multi-view representation learning for subspace clustering. In *IJCAI*, pages 2916–2922, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.894, + 0.476 + ], + "angle": 0, + "content": "[19] Xuelong Li, Han Zhang, Rong Wang, and Feiping Nie. Multi-view clustering: A scalable and parameter-free bipartite graph fusion method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.894, + 0.545 + ], + "angle": 0, + "content": "[20] Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, and Fuchun Sun. Reasoning over different types of knowledge graphs: Static, temporal and multi-modal. arXiv preprint arXiv:2212.05767, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.617 + ], + "angle": 0, + "content": "[21] Ke Liang, Yue Liu, Sihang Zhou, Wenxuan Tu, Yi Wen, Xihong Yang, Xiangjun Dong, and Xinwang Liu. Knowledge graph contrastive learning based on relation-symmetrical structure. IEEE Transactions on Knowledge and Data Engineering, pages 1-12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.673 + ], + "angle": 0, + "content": "[22] Ke Liang, Lingyuan Meng, Sihang Zhou, Siwei Wang, Wenxuan Tu, Yue Liu, Meng Liu, and Xinwang Liu. Message intercommunication for inductive relation reasoning. arXiv preprint arXiv:2305.14074, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.894, + 0.744 + ], + "angle": 0, + "content": "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11174-11183, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[24] Jialu Liu, Chi Wang, Jing Gao, and Jiawei Han. Multiview clustering via joint nonnegative matrix factorization. In Proceedings of the 2013 SIAM International Conference on Data Mining, pages 252-260. SIAM, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.894, + 0.857 + ], + "angle": 0, + "content": "[25] Suyuan Liu, Xinwang Liu, Siwei Wang, Xin Niu, and En Zhu. Fast incomplete multi-view clustering with view-independent anchors. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[26] Suyuan Liu, Siwei Wang, Pei Zhang, Kai Xu, Xinwang Liu, Changwang Zhang, and Feng Gao. Efficient one-pass multiview subspace clustering with consensus anchors. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 7576-7584, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.123, + 0.469, + 0.178 + ], + "angle": 0, + "content": "[27] Suyuan Liu, Qing Liao, Siwei Wang, Xinwang Liu, and En Zhu. Robust and consistent anchor graph learning for multiview clustering. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.469, + 0.234 + ], + "angle": 0, + "content": "[28] Xun Lu and Songhe Feng. Structure diversity-induced anchor graph fusion for multi-view clustering. ACM Transactions on Knowledge Discovery from Data, 17(2):1-18, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.239, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[29] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In International conference on machine learning, pages 5092-5101. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.469, + 0.353 + ], + "angle": 0, + "content": "[30] Qianyao Qiang, Bin Zhang, Fei Wang, and Feiping Nie. Fast multi-view discrete clustering with anchor graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9360-9367, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.355, + 0.469, + 0.41 + ], + "angle": 0, + "content": "[31] Shaojun Shi, Feiping Nie, Rong Wang, and Xuelong Li. Fast multi-view clustering via prototype graph. IEEE Transactions on Knowledge and Data Engineering, 35(1):443-455, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[32] Xiaochuang Shu, Xiangdong Zhang, Quanxue Gao, Ming Yang, Rong Wang, and Xinbo Gao. Self-weighted anchor graph learning for multi-view clustering. IEEE Transactions on Multimedia, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[33] Mengjing Sun, Pei Zhang, Siwei Wang, Sihang Zhou, Wenxuan Tu, Xinwang Liu, En Zhu, and Changjian Wang. Scalable multi-view subspace clustering with unified anchors. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3528-3536, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.544, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[34] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.685 + ], + "angle": 0, + "content": "[35] Xinhang Wan, Jiyuan Liu, Weixuan Liang, Xinwang Liu, Yi Wen, and En Zhu. Continual multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, page 3676-3684, New York, NY, USA, 2022. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[36] Xinhang Wan, Xinwang Liu, Jiyuan Liu, Siwei Wang, Yi Wen, Weixuan Liang, En Zhu, Zhe Liu, and Lu Zhou. Autoweighted multi-view clustering for large-scale data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10078-10086, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[37] Jun Wang, Chang Tang, Zhiguo Wan, Wei Zhang, Kun Sun, and Albert Y Zomaya. Efficient and effective one-step multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[38] Siwei Wang, Xinwang Liu, Li Liu, Wenxuan Tu, Xinzhong Zhu, Jiyuan Liu, Sihang Zhou, and En Zhu. Highly-efficient incomplete large-scale multi-view clustering with consensus bipartite graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9776-9785, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[39] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[40] Siwei Wang, Xinwang Liu, Xinzhong Zhu, Pei Zhang, Yi Zhang, Feng Gao, and En Zhu. Fast parameter-free multiview subspace clustering with consensus anchor guidance. IEEE Trans. Image Process., 31:556-568, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[41] Yang Wang, Lin Wu, Xuemin Lin, and Junbin Gao. Multiview spectral clustering via structured low-rank matrix factorization. IEEE transactions on neural networks and learning systems, 29(10):4833-4843, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[42] Yi Wen, Siwei Wang, Ke Liang, Weixuan Liang, Xinhang Wan, Xinwang Liu, Suyuan Liu, Jiyuan Liu, and En Zhu. Scalable incomplete multi-view clustering with structure alignment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3031-3040, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[43] Yi Wen, Siwei Wang, Qing Liao, Weixuan Liang, Ke Liang, Xinhang Wan, and Xinwang Liu. Unpaired multi-view graph clustering with cross-view structure matching. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[44] Wei Xia, Quanxue Gao, Qianqian Wang, Xinbo Gao, Chris Ding, and Dacheng Tao. Tensorized bipartite graph learning for multi-view clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[45] Weiqing Yan, Jindong Xu, Jinglei Liu, Guanghui Yue, and Chang Tang. Bipartite graph-based discriminative feature learning for multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3403-3411, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[46] Ben Yang, Xuetao Zhang, Feiping Nie, Fei Wang, Weizhong Yu, and Rong Wang. Fast multi-view clustering via nonnegative and orthogonal factorization. IEEE Transactions on Image Processing, 30:2575-2586, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.657 + ], + "angle": 0, + "content": "[47] Ben Yang, Xuetao Zhang, Zhiping Lin, Feiping Nie, Badong Chen, and Fei Wang. Efficient and robust multiview clustering with anchor graph regularization. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6200-6213, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[48] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[50] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In Proceedings of the 31st ACM International Conference on Multimedia, pages 337-346, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[51] Jiali You, Zhenwen Ren, Xiaojian You, Haoran Li, and Yuancheng Yao. Priori anchor labels supervised scalable multi-view bipartite graph clustering. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.119 + ], + "angle": 0, + "content": "AAAI Conference on Artificial Intelligence, pages 10972-10979, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.189 + ], + "angle": 0, + "content": "[52] Shengju Yu, Siwei Wang, Yi Wen, Ziming Wang, Zhigang Luo, En Zhu, and Xinwang Liu. How to construct corresponding anchors for incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[53] Chao Zhang, Huaxiong Li, Wei Lv, Zizheng Huang, Yang Gao, and Chunlin Chen. Enhanced tensor low-rank and sparse representation recovery for incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11174-11182, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[54] Nan Zhang, Xiaoqin Zhang, and Shiliang Sun. Efficient multiview representation learning with correntropy and anchor graph. IEEE Transactions on Knowledge & Data Engineering, (01):1-14, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.401 + ], + "angle": 0, + "content": "[55] Pei Zhang, Siwei Wang, Liang Li, Changwang Zhang, Xinwang Liu, En Zhu, Zhe Liu, Lu Zhou, and Lei Luo. Let the data choose: flexible and diverse anchor graph fusion for scalable multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11262-11269, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.403, + 0.469, + 0.472 + ], + "angle": 0, + "content": "[56] Xiaojia Zhao, Qiangqiang Shen, Yongyong Chen, Yongsheng Liang, Junxin Chen, and Yicong Zhou. Self-completed bipartite graph learning for fast incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26161" + } + ] +] \ No newline at end of file diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_origin.pdf b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..49aa8d10c55429375e6cc38192002e73f3e9d79a --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/4eed6c41-c8e8-44a9-84e6-f4256d6b5e7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d9e21b5fca97961885a48d37db411437e8a9334c1954bfe6be4270e42328096 +size 856328 diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/full.md b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3834c50eaa48d48ecbd45b7d8b07f3701452f321 --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/full.md @@ -0,0 +1,381 @@ +# Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering + +Suyuan Liu1, Ke Liang1, Zhibin Dong1, Siwei Wang2, Xihong Yang1, Sihang Zhou1, En Zhu1, Xinwang Liu1* +1National University of Defense Technology, Changsha, China +2Intelligent Game and Decision Lab, Beijing, China + +{suyuanliu,xinwangliu,enzhu}@nudt.edu.cn + +# Abstract + +In recent years, anchor-based methods have achieved promising progress in multi-view clustering. The performances of these methods are significantly affected by the quality of the anchors. However, the anchors generated by previous works solely rely on single-view information, ignoring the correlation among different views. In particular, we observe that similar patterns are more likely to exist between similar views so such correlation information can be leveraged to enhance the quality of the anchors, which is also omitted. To this end, we propose a novel plug-and-play anchor enhancement strategy through view correlation for multi-view clustering. Specifically, we construct a view graph based on aligned initial anchor graphs to explore inter-view correlations. By learning from view correlation, we enhance the anchors of the current view using the relationships between anchors and samples on neighboring views, thereby narrowing the spatial distribution of anchors on similar views. Experimental results on seven datasets demonstrate the superiority of our proposed method over other existing methods. Furthermore, extensive comparative experiments validate the effectiveness of the proposed anchor enhancement module when applied to various anchor-based methods. + +# 1. Introduction + +Large-scale scenarios have always been a hot concern with the data explosion in our real world. Among these scenarios, large-scale multi-view clustering (MVC) has also drawn increasing attention in recent years[2, 4, 8, 18, 35]. In general, large-scale MVC methods can be categorized into two types, i.e., matrix factorization-based methods and anchor-based methods. The former methods [11, 24, 41, 50] aim to generate a low-dimensional representation for further clustering by decomposing the original data, thereby avoiding the computational costs of similarity matrix construc + +![](images/2696bbd21b59aa35ef1a797cc514518d9ff939b0579d1af45fb497bea0883d18.jpg) +Figure 1. An example illustration of clustering results affected by initial anchors. The first two methods keep the anchors fixed, while the latter two methods update the anchors through iteration-optimization. Their clustering performances are influenced to varying degrees by different anchor generation approaches. + +tion. However, their efficiency usually decreases along with increasing dimension, thus constraining their scalability[10, 36]. Inspired by prototype learning [16, 48], anchors are introduced into MVC to reduce the scope for clustering from $n$ to $m$ , which denotes the number of samples and anchors, respectively. Specifically, anchor-based MVC methods [13, 27, 45] construct an anchor graph $(n \times m)$ rather than a similarity matrix $(n \times n)$ . It will alleviate the aforementioned computation and scalability problems. + +Anchor-based MVC methods typically consist of three procedures, including (1) anchor generation, (2) anchor graph construction, and (3) clustering. Early attempts usually generate anchors at first, which are fixed during constructing anchor graphs[31, 32, 47]. For example, SFMC [19] selects anchors by iterative sampling based on feature similarity. Besides, LMVSC [14] regards cluster centers generated by $k$ -means of each view as the anchors. However, these methods are sensitive to the initial anchors (See Fig. 1), where the anchor selection is primitive and sometimes results in compromised performance. + +To address the above issues, researchers make efforts on anchor optimization to guarantee better quality anchors[17, + +26, 38]. For example, SMVSC [33] proposed a unified framework that combines anchor generation and anchor-graph construction, achieving iterative updates to optimize the anchors. Building upon this, UDBGL [7] introduced the Laplacian rank constraint to ensure that the anchor graph possesses a discrete cluster structure. The above iteration-optimization strategies actually aim to get a better combination of the anchor and its corresponding anchor graph instead of enhancing the quality of the anchors themselves. Meanwhile, to the best of our knowledge, almost all existing anchor optimization models belong to this type. Although they can get better and more stable clustering performance, they still suffer from the initial state for optimization[6, 15], which leaves us a great challenge as well as the opportunity, i.e., what is the new strategy to alleviate such problem? In this paper, we provide a potential solution to it. + +We observe that most anchor generation methods only rely on a single view[3, 12, 30]. Multiple works have proven that information of different views for the same sample will benefit the expressive ability of the sample [23, 29, 34, 49]. Similarly, different views for the same anchors should also be used to generate better anchors. Specifically, utilizing information across views and ensuring the consistency of anchor graphs on each view will benefit the anchor generation procedure. However, such a characteristic is not used for anchor enhancement in previous anchor-based MVC methods, and our strategy is developed based on it. + +In this work, we propose an Anchor Enhancement strategy with View Correlation (AEVC) for multi-view clustering (MVC). The framework of AEVC is shown in Fig. 2. The proposed approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which enhance the quality of anchors. Additionally, we improve the clustering performance by revising the existing anchor graph construction method based on view correlation. Specifically, we first align the initial anchors and construct a view-graph based on the aligned anchor graph to find the neighbors of each view. Then, new anchors for the current view are generated with the anchor graph of neighboring views, thereby enhancing the initial anchors. Finally, the inter-view correlations are combined to learn a consistent anchor graph between views with the enhanced anchors. AEVC optimizes anchors by absorbing information from similar views, improving initial anchor quality, and reducing distribution bias caused by the independent anchors generation in each view. The main contributions of our work can be summarized as follows: + +- We propose AEVC, a plug-and-play anchor enhancement strategy for MVC based on view correlation, to address the issue of poor initial anchor quality affecting the performance of MVC. +- By learning from the view correlation, AEVC incorpo + +![](images/89574ce77c2e9aa7ee1accf6245c120eb95e5f018a0f390a605f7205ac8c126c.jpg) +Figure 2. The framework of our proposed AEVC. Specifically, in the view-graph learning module, AEVC primarily performs anchor alignment and learns the view-graph. In the anchor enhancement module, AEVC leverages the neighboring relationships from the view-graph to enhance the anchors. Finally, AEVC executes anchor graph construction based on the enhanced anchors. + +rates the anchor distribution information from neighboring views into the process of anchor enhancement. Theoretical and experimental validations demonstrate that the anchors enhanced by AEVC exhibit a tighter inter-view relationship. + +- We compared our proposed approach to other anchor-based multi-view methods and demonstrated its superiority. Extensive experiments verify the effectiveness and generalization of the proposed AEVC strategy when integrated with typical anchor-based MVC methods. + +# 2. Related Work + +# 2.1. Anchor-based Multi-view Clustering + +To address the challenges posed by traditional multiview clustering methods in handling large-scale datasets, Anchor-based MVC has been proposed in recent years[37, 45, 55]. The core idea of anchor-based MVC is to generate a set of representative anchors from the entire sample pool, where the number of anchors is typically much smaller than the total number of samples[25, 54]. In contrast to conventional methods that require establishing relationships among all samples, anchor-based approaches only involve computing similarities between anchors and samples. This significantly reduces the complexity associated with constructing the relationship graph. Additionally, adopting a reduced-dimensional graph further mitigates computational burdens, facilitating subsequent multi-view fusion and clustering with reduced computational overhead. + +Traditional anchor-based MVC methods generate anchors by sampling[30, 31, 46]. Subsequent anchor graph construction and clustering are performed based on the initially determined anchors. For instance, Kang et al.[14] + +and Yang et al. [46] advocate applying $k$ -means to the original samples, selecting cluster centers as anchors to enhance their representativeness. Xia et al. [44] introduces a variance-based de-correlation anchor selection strategy, leveraging the internal data structure to augment the coverage capability of anchors. The performance of such methods is highly influenced by the quality of the initially generated anchors. + +# 2.2. Anchor Optimization Strategy + +To tackle the above challenges, extensive research efforts have been devoted to anchor optimization in order to achieve superior anchor quality. Iteration-optimization methods propose to update anchors and anchor graphs, utilizing the anchor graph obtained from the previous iteration to update the anchors[51, 53, 56]. Sun et al. [33] initially propose learning anchors and constructing anchor graphs through iterative updates within the same framework. Liu et al. [26] further introduce Laplacian rank constraints, enabling the anchor graph to generate labels without directly needing post-processing. However, the results of iterative update algorithms are significantly influenced by the initial values of variables[6, 15]. Instead of directly enhancing the quality of the anchors themselves, the above methods primarily focus on obtaining an improved combination of the anchor and its corresponding anchor graph. + +# 2.3. View Correlation Learning + +Multi-view and multi-modal scenarios are widely seen nowadays [1, 20]. Among them, investigating the relationships between multiple objects is a crucial topic [5, 21, 22], especially for multi-view learning, which holds for anchor-based multi-view clustering[7, 44]. Some existing methods delve into the consistency of views by aligning anchors[42, 43, 52]. For instance, Li et al. [19] concatenates raw data from all views to select anchors, ensuring correspondence among anchors across views. Wang et al. [39] takes a step further by proposing an anchor alignment framework based on the features and structural information of anchors. Another category of methods employs early fusion to avoid conflicts between views[9, 28]. For instance, Wang et al. [40] use a projection matrix to learn shared anchors and anchor graphs in the latent space. However, the first category of methods solely utilizes alignment information between views, while the second category underutilizes complementary information among views. In the next section, we will introduce a novel approach based on a view graph, leveraging inter-view interactions to enhance anchors. + +# 3. Method + +In this section, we introduce our proposed Anchor Enhancement strategy with View Correlation (AEVC) for + +multi-view clustering. We begin by sequentially describing the three modules of AEVC. Subsequently, a complexity analysis of the proposed algorithm is provided. Lastly, a theoretical analysis is conducted to evaluate the effectiveness of AEVC. + +# 3.1. Overview + +Existing anchor-based MVC methods can be divided into three steps: (1) anchor generation, (2) anchor graph construction, (3) clustering. The clustering performance of existing anchor-based MVC methods heavily relies on the quality of initially generated anchors. Although iteration-optimization strategies alleviate this issue by exploring combinations of anchors and anchor graphs, they are still impacted by the optimization of the initial state. To address this problem, we propose an anchor enhancement strategy based on view correlation for multi-view clustering. Specifically, different from existing methods, our approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which helps to enhance the quality of anchors. Furthermore, we revise the existing anchor graph construction method based on view correlation, further improving the clustering performance. + +# 3.2. View-graph Learning + +Given multi-view data $\{\mathbf{X}_i\}_{i=1}^v$ , where $\mathbf{X}_i \in \mathbb{R}^{d_i \times n}$ is composed of $n$ samples with $d_i$ dimension. Denoting $\mathbf{A}_i^{(0)} \in \mathbb{R}^{d_i \times m}$ as the generated anchor matrix in the $i$ -th view, and $\mathbf{Z}_i^{(0)} \in \mathbb{R}^{m \times n}$ represents the corresponding anchor graph, where $m$ is the number of anchors. Due to the varying dimensions and inconsistent distributions of data across different views, it is not feasible to directly assess the distance between $\mathbf{X}_i$ and $\mathbf{X}_j$ . However, it is easy to calculate the similarity between anchor graphs from different views due to the same size of anchor graphs. + +Nevertheless, the independently generated anchors on each view introduce the Anchor-Unaligned Problem [39]. This means that directly computing the similarity between anchor graphs of different views would be inaccurate due to the misalignment of rows corresponding to anchors. Thus, an anchor alignment operation must be performed before view-graph learning. + +Existing anchor alignment strategies typically directly align the anchors on the first view, which carries the risk of being influenced by noise view. Here, we first concatenate the multi-view data column-wise to obtain $\hat{\mathbf{X}} = [\mathbf{X}_1; \ldots; \mathbf{X}_v] \in \mathbb{R}^{\sum_{i=1}^{v} d_i \times n}$ . Then, we generate the anchors $\hat{\mathbf{A}} \in \mathbb{R}^{\sum_{i=1}^{v} d_i \times m}$ and the corresponding anchor graph $\hat{\mathbf{Z}} \in \mathbb{R}^{m \times n}$ . According to [39], we can obtain the permutation matrix $\mathbf{P}_i$ in the $i$ -th view by minimizing the distance between $\hat{\mathbf{Z}}$ and the permuted $\mathbf{Z}_i^{(0)}$ , the formula is + +as follows, + +$$ +\min _ {\mathbf {P} _ {i}} \left\| \hat {\mathbf {Z}} - \mathbf {P} _ {i} \mathbf {Z} _ {i} ^ {(0)} \right\| _ {\mathbf {F}} ^ {2}, \tag {1} +$$ + +$$ +\mathrm {s . t .} \mathbf {P} _ {i} \mathbf {1} = \mathbf {1}, \mathbf {P} _ {i} ^ {\top} \mathbf {1} = \mathbf {1}, \mathbf {P} _ {i} \in \{0, 1 \} ^ {m \times m}. +$$ + +Then the anchor matrix and anchor graph can be aligned with the permutation matrix as follows: + +$$ +\left\{ \begin{array}{l} \mathbf {A} _ {i} = \mathbf {A} _ {i} ^ {(0)} \mathbf {P} _ {i} ^ {\top}, \\ \mathbf {Z} _ {i} = \mathbf {P} _ {i} \mathbf {Z} _ {i} ^ {(0)}. \end{array} \right. \tag {2} +$$ + +After performing the alignment operation, we calculate the view graph $\mathbf{S} \in \mathbb{R}^{v \times v}$ as follows, + +$$ +\left\{ \begin{array}{l l} \mathbf {s} _ {i j} = \frac {1}{\| \mathbf {Z} _ {i} - \mathbf {Z} _ {j} \| _ {\mathbf {F}} ^ {2}}, & i \neq j, \\ \mathbf {s} _ {i j} = 0, & i = j,. \end{array} \right. \tag {3} +$$ + +where $\mathbf{s}_{ij}$ denotes the similarity between the $i$ -th and the $j$ -th view. + +# 3.3. Anchor Enhancement Strategy + +Most existing methods independently generate anchors on each view, ignoring the correlation between views. In fact, anchors on similar views should have similar distributions, which means that the corresponding anchor graphs should be close to each other. Based on this assumption, we propose to enhance the anchors of the current view with information on neighboring views. One intuitive idea is to directly update the anchors based on the anchors of neighboring views. However, the dimensions of anchors on different views are not the same, making direct transfer infeasible. Instead of utilizing feature consistency, we focus on structural consistency. Specifically, we reconstruct new anchors by leveraging the anchor graphs from neighboring views and the data from the current view, thereby enhancing the initial anchors. The formulation of the anchor enhancement on the $i$ -th view is as follows: + +$$ +\tilde {\mathbf {A}} _ {i} = \frac {1}{\mu} \left(\mathbf {A} _ {i} + \gamma \sum_ {\mathbf {X} _ {j} \in \Omega_ {\zeta} \left(\mathbf {X} _ {i}\right)} \hat {s} _ {i j} \mathbf {X} _ {i} \mathbf {Z} _ {j} ^ {\top}\right). \tag {4} +$$ + +where $\mu = 1 + \gamma \sum_{\mathbf{X}_j \in \Omega_\zeta(\mathbf{X}_i)} \hat{s}_{ij}$ ensures that the enhanced anchors are a convex combination of the initial anchors and the reconstructed anchors. $\Omega_\zeta(\mathbf{X}_i)$ denotes the $\zeta$ -neighbors of the $i$ -th view, which is defined as: + +$$ +\Omega_ {\zeta} \left(\mathbf {X} _ {i}\right) = \arg \max _ {\mathbf {X} _ {j}} \sum_ {j = 1} | s _ {i j} |. \tag {5} +$$ + +$\gamma$ is the enhancement rate, which is used to adjust the influence of the initial anchors on the enhanced version. $\hat{\mathbf{S}}$ represents the normalized view-graph based on the following + +settings, which controls the influence of different neighbors during anchor enhancement: + +$$ +\hat {s} _ {i j} \left\{ \begin{array}{l l} s _ {i j}, & \mathbf {X} _ {j} \in \Omega_ {\zeta} \left(\mathbf {X} _ {i}\right), \\ 0, & \mathbf {X} _ {j} \notin \Omega_ {\zeta} \left(\mathbf {X} _ {i}\right). \end{array} \right. \tag {6} +$$ + +By enhancing the anchors with Eq. (4), the quality of the initial anchors is improved. Moreover, the enhancement further achieves a better alignment between anchors on similar views. In the subsequent sections, we will analyze the aforementioned conclusions from both theoretical and experimental perspectives. Furthermore, our proposed anchor enhancement module is plug-and-play and we will showcase its performance improvement in the experiments by integrating different anchor-based multi-view clustering methods. + +# 3.4. Anchor Graph Construction + +After obtaining high-quality anchors through our proposed anchor enhancement module, it is crucial to construct a reliable anchor graph that ensures view consistency. Since the enhanced anchors are aligned, we can directly generate a common anchor graph. However, the contributions of different views in constructing the anchor graph are not equal. To mitigate the influence of view discrepancies on building a consistent representation, we assign higher weights to views that have stronger associations with other views. Specifically, the contribution weight $\omega_{i} = \sum_{j=1}^{v} s_{ij}$ for the $i$ -th view is computed as the sum of its similarities with all other views. Furthermore, to avoid the impact of noise during anchor graph construction, we introduce the learned anchor graph $\hat{\mathbf{Z}}$ from the concatenated views as guidance. Ultimately, the anchor graph construction is formulated as follows: + +$$ +\min _ {\mathbf {Z}} \omega_ {i} \left\| \mathbf {X} _ {i} - \tilde {\mathbf {A}} _ {i} \mathbf {Z} \right\| _ {\mathbf {F}} ^ {2} + \lambda \left\| \mathbf {Z} - \hat {\mathbf {Z}} \right\| _ {\mathbf {F}} ^ {2}, \tag {7} +$$ + +$$ +\begin{array}{l} \text {s . t .} \mathbf {Z} \geq 0, \mathbf {Z} ^ {\top} \mathbf {1} = \mathbf {1}. \end{array} +$$ + +Due to the independence between each column of $\mathbf{Z}$ , let $z_{j}$ denote the $j$ -th column of $\mathbf{Z}$ . The aforementioned problem can be transformed into $n$ Quadratic Programming(QP) problems, formulated as follows: + +$$ +\min \frac {1}{2} \mathbf {z} _ {j} ^ {\top} \mathbf {H} \mathbf {z} _ {j} + f \mathbf {z} _ {j}, \tag {8} +$$ + +$$ +\begin{array}{l} \text {s . t .} \mathbf {z} _ {j} \geq 0, \mathbf {z} _ {j} ^ {\top} \mathbf {1} = 1, \end{array} +$$ + +where the matrix $\mathbf{H} = 2\left(\sum_{i=1}^{v}\omega_{i}\tilde{\mathbf{A}}_{i}^{\top}\tilde{\mathbf{A}}_{i} + \lambda \mathbf{I}\right)$ and the vector $f = -2\left(\sum_{i=1}^{v}\omega_{i}\mathbf{X}_{i[:,j]}^{\top}\tilde{\mathbf{A}}_{i} + \lambda \hat{z}_{j}\right)$ . The optimal $\mathbf{Z}$ can be obtained by solving the above QP problem. Finally, the results can be generated by performing $k$ -means on the left singular vector of $\mathbf{Z}$ [14]. The entire procedure of AEVC is shown in Algorithm 1. + +# Algorithm 1: The proposed AEVC + +Input: Multi-view datasets $\{\mathbf{X}_i\}_{i = 1}^v$ , the number of anchors $m$ , the parameter $\gamma$ $\lambda$ and the number of clusters $k$ + +Output: Clustering results. + +1 Initialize $\{\mathbf{A}_i^{(0)}\}_{i = 1}^v$ $\hat{\mathbf{A}}$ $\{\mathbf{Z}_i^{(0)}\}_{i = 1}^v$ and $\hat{\mathbf{Z}}$ // View-graph Learning +2 for $i = 1\to v$ do +3 Calculate permutation matrix $\mathbf{P}_i$ with Eq. (1); 4 Perform alignment with Eq. (2); +5 Learn view-graph S as defined in Eq. (3); // Anchor Enhancement +6 for $i = 1\to v$ do +7 Find $\zeta$ -neighbors $\Omega_{\zeta}(\mathbf{X}_i)$ as defined in Eq. (5); +8 Calculate normalized view-graph $\hat{\mathbf{S}}$ with Eq. (6); +9 for $i = 1\rightarrow v$ do +10 Enhance $\mathbf{A}_i$ with Eq. (4); + +// Anchor Graph Construction +11 Obtain $\mathbf{Z}$ by solving Eq. (7); +12 Perform $k$ -means on the left singular vector of $\mathbf{Z}$ . + +# 3.5. Discussion + +Computational complexity. The computational complexity of the proposed AEVC mainly consists of three modules. The complexity of the view-graph learning module is $\mathcal{O}(m^3 v + nmv)$ . Specifically, the computation of $v$ permutation matrices $\mathbf{P}_i$ has a complexity of $\mathcal{O}(m^3 v)$ , and the computation of the view graph has a complexity of $\mathcal{O}(nmv)$ . The time complexity of the anchor enhancement module is $\mathcal{O}(nmdv)$ , where $d = \sum_{i=1}^{v} d_v$ , primarily arising from matrix multiplication during enhancement. The complexity of the anchor graph construction module is $\mathcal{O}(nm^3 v + nmdv)$ , where solving all QP problems requires a time complexity of $\mathcal{O}(nm^3 v)$ , and the involved matrix multiplication requires $\mathcal{O}(nm^3 v + nmdv)$ . Therefore, the overall computational complexity of AEVC is $\mathcal{O}(nm^3 v + m^3 v + nmdv + nmv)$ , which exhibits linear complexity with respect to the number of samples. + +Space complexity. During the execution of the AEVC algorithm, the main variables that need to be stored include matrices $\mathbf{X}_i$ , $\mathbf{A}_i$ , $\mathbf{Z}_i$ , $\hat{\mathbf{A}}$ , $\hat{\mathbf{Z}}$ , $\mathbf{P}_i$ , $\mathbf{S}$ , and $\hat{\mathbf{S}}$ . The overall space complexity of AEVC is $\mathcal{O}(nd + ndv + nmdv + v^2)$ , which is also linear with respect to the number of samples and can be extended to large-scale scenarios. + +Two Property of enhanced anchors. We have discovered that the anchors enhanced by the proposed AEVC exhibit stronger inter-view consistency. The theoretical analysis is provided below. + +Assumption 1. Given a two-view data $\{\mathbf{X}_1, \mathbf{X}_2\}$ , denoting + +$\{\mathbf{A}_1,\mathbf{A}_2\}$ are the generated initial anchors, and $\{\mathbf{Z}_1,\mathbf{Z}_2\}$ are the corresponding anchor graphs under the assumption that $\mathbf{X}_1 = \mathbf{A}_1\mathbf{Z}_1$ and $\mathbf{X}_2 = \mathbf{A}_2\mathbf{Z}_2$ + +For the convenience of derivation, we simplify the anchor enhancement strategy in Eq. (4) with $\alpha \in (0,1)$ as follows: + +$$ +\left\{ \begin{array}{l} \tilde {\mathbf {A}} _ {1} = \alpha \mathbf {A} _ {1} + (1 - \alpha) \mathbf {X} _ {1} \mathbf {Z} _ {2} ^ {\top}, \\ \tilde {\mathbf {A}} _ {2} = \alpha \mathbf {A} _ {2} + (1 - \alpha) \mathbf {X} _ {2} \mathbf {Z} _ {1} ^ {\top}. \end{array} \right. \tag {9} +$$ + +The corresponding anchor graphs $\{\tilde{\mathbf{Z}}_1,\tilde{\mathbf{Z}}_2\}$ are constructed under the same assumption that $\mathbf{X}_1 = \tilde{\mathbf{A}}_1\tilde{\mathbf{Z}}_1$ and $\mathbf{X}_2 = \tilde{\mathbf{A}}_2\tilde{\mathbf{Z}}_2$ . Then we can derive that + +$$ +\left\{ \begin{array}{l} \tilde {\mathbf {Z}} _ {1} = \alpha \mathbf {Z} _ {1} + (1 - \alpha) \mathbf {Z} _ {2}, \\ \tilde {\mathbf {Z}} _ {2} = \alpha \mathbf {Z} _ {2} + (1 - \alpha) \mathbf {Z} _ {1}. \end{array} \right. \tag {10} +$$ + +We can deduce the following properties: + +Property 1. The disparity among the anchor graphs on each view is reduced after anchor enhancement. Mathematically, + +$$ +\left\| \tilde {\mathbf {Z}} _ {1} - \tilde {\mathbf {Z}} _ {2} \right\| _ {\mathbf {F}} < \| \mathbf {Z} _ {1} - \mathbf {Z} _ {2} \| _ {\mathbf {F}}. \tag {11} +$$ + +Proof. According to Eq. (10), we can derive that + +$$ +\begin{array}{l} \left\| \tilde {\mathbf {Z}} _ {1} - \tilde {\mathbf {Z}} _ {2} \right\| _ {\mathbf {F}} = \| (2 \alpha - 1) (\mathbf {Z} _ {1} - \mathbf {Z} _ {2}) \| _ {\mathbf {F}} \\ = \left| 2 \alpha - 1 \right| \cdot \left\| \mathbf {Z} _ {1} - \mathbf {Z} _ {2} \right\| _ {\mathbf {F}} \tag {12} \\ < \left\| \mathbf {Z} _ {1} - \mathbf {Z} _ {2} \right\| _ {\mathbf {F}}. \\ \end{array} +$$ + +This completes the proof. + +Property 1 implies that anchor enhancement leads to a more consistent distribution of anchors in similar views, thereby establishing a tighter alignment. + +Property 2. Assuming there exists a value $\beta \in (0,1)$ such that $\mathbf{Z} = \beta \mathbf{Z}_1 + (1 - \beta)\mathbf{Z}_2$ is the optimal consistent anchor graph among all views, the enhanced anchor graph is closer to the optimal version after enhancement than before for all $\alpha \in (2\beta -1,1)\cap (0,1)$ . Mathematically, + +$$ +\left\| \tilde {\mathbf {Z}} _ {1} - \mathbf {Z} \right\| _ {\mathbf {F}} < \| \mathbf {Z} _ {1} - \mathbf {Z} \| _ {\mathbf {F}}. \tag {13} +$$ + +Proof. According to Eq. (10), we can derive that + +$$ +\begin{array}{l} \left\| \tilde {\mathbf {Z}} _ {1} - \mathbf {Z} \right\| _ {\mathbf {F}} = \| (\alpha - \beta) (\mathbf {Z} _ {1} - \mathbf {Z} _ {2}) \| _ {\mathbf {F}} \\ = | \alpha - \beta | \cdot \| \mathbf {Z} _ {1} - \mathbf {Z} _ {2} \| _ {\mathbf {F}} \\ < | 1 - \beta | \cdot \| \mathbf {Z} _ {1} - \mathbf {Z} _ {2} \| _ {\mathbf {F}} \tag {14} \\ = \left\| (1 - \beta) \left(\mathbf {Z} _ {1} - \mathbf {Z} _ {2}\right) \right\| _ {\mathbf {F}} \\ = \left\| \mathbf {Z} _ {1} - \mathbf {Z} \right\| _ {\mathbf {F}}. \\ \end{array} +$$ + +This completes the proof. + +![](images/e1f847bc1d8bb9a2294dd727278e238a03f0f0c5d8e113a34d4780731614362e.jpg) + +Property 2 implies that anchor enhancement helps reduce the gap between the anchor graph on each view and the optimal consistent anchor graph. It is easy to derive the multiview extended versions of the above properties. + +# 4. Experiment + +# 4.1. Experiment Setup + +Datasets. Seven multi-view datasets are employed in our experiments: Dermatology1 contains diagnostic information for six types of skin diseases. ForestTypes2 consists of satellite images of different forest types. BDGP3 comprises both images and textual descriptions of Drosophila embryos. Reuters4 is a news text dataset. Hdigit5 consists of a set of handwritten digital images. VGGFace2 is a large-scale face recognition dataset. CIFAR100 is a subset of a large-scale tiny image dataset. Detail description is shown in Table 1. + +Compared methods. We compare the proposed AEVC with existing state-of-the-art anchor-based MVC methods: Large-scale multi-view subspace clustering in linear time (LMVSC)[14]; Scalable Multi-view Subspace Clustering with Unified Anchors (SMVSC)[33]; Fast multi-view clustering via nonnegative and orthogonal factorization (FMCNOF)[46]; Fast Parameter-Free Multi-View Subspace Clustering With Consensus Anchor Guidance (FPMVSCAG)[40]; Efficient Multi-View Clustering via Unified and Discrete Bipartite Graph Learning (UDBGL)[7]; Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity (FastMICE)[10]; Flexible and Diverse Anchor Graph Fusion for Scalable Multi-View Clustering (FDAGF)[55]; Align then Fusion: Generalized Large-scale Multi-view Clustering with Anchor Matching Correspondences (FMVACC)[39]. + +Implementation details. For the above-compared algorithm, we have tuned their parameters to align with the relevant literature. We conducted a search for three hyperparameters in the proposed AEVC method using predefined lists. Specifically, the anchor number $m$ is chosen from the set $[k, 2k, 5k]$ , where $k$ represents the cluster number. The enhancement rate $\gamma$ is selected from $[0.1, 1, 10, 100]$ , and the balancing hyperparameter $\lambda$ is chosen from the range $[0.0001, 0.01, 1, 100]$ . To ensure a fair comparison, algorithms relying on $k$ -means for result acquisition underwent 50 independent runs, and the results were averaged for accuracy assessment. Our experimental evaluation incorporated + +Table 1. Benchmark datasets description. + +
Datasetnvkdp(p=1,...,v)
Dermatology3582612/22
ForestTypes523349/9/9
BDGP2500351000/500/250
Reuters7200564819/4810/4892/4858/4777
Hdigit10000210784/256
VGGFace23628741003944/576/512/640
CIFAR100500004100944/576/512/640
+ +four distinct metrics: clustering Accuracy (ACC), Normalized Mutual Information (NMI), Purity, and F-score. + +# 4.2. Superiority of AEVC + +We compare our proposed AEVC with eight state-of-the-art anchor-based multi-view clustering methods on seven datasets, as illustrated in Table 2. From the results, we have the following observations: + +- The experimental results reveal that our proposed algorithm consistently outperforms other anchor-based MVC methods across all datasets, confirming the superiority of AEVC. In terms of ACC, AEVC surpasses the second-best algorithm by $6.9\%$ , $4.1\%$ , $1.9\%$ , $9.9\%$ , $1.3\%$ , $15.8\%$ , and $9.5\%$ on different datasets. +- Compared to previous methods based on iteratively optimized anchors (SMVSC, FPMVS-CAG, UDBGL, FDAGF, FMVACC), AEVC exhibits superior clustering performance, emphasizing the effectiveness of our anchor enhancement approach over iteration-optimization methods. + +# 4.3. Effectiveness of Enhanced Anchors + +We carried out an ablation study to demonstrate the effectiveness of the proposed anchor enhancement strategy. In our approach, the anchor enhancement strategy typically comprises two steps: anchor alignment in the view-graph learning module and anchor enhancement module. We denote the method of removing the entire anchor enhancement strategy as a baseline and compare it with those where anchor alignment and enhancement are sequentially incorporated. From the experimental results shown in Table 3, it is observed that anchor alignment leads to an improvement in clustering performance, as misalignment of anchors can result in incorrect multi-view fusion[39]. Furthermore, with the application of our anchor enhancement strategy, there is a further improvement in clustering performance, providing strong evidence for the effectiveness of the anchor enhancement strategy. + +To further demonstrate the effectiveness of the enhanced anchors, we incorporate randomly generated anchors from the original data at different proportions and showcase the corresponding clustering performance in Fig. 3(a). For instance, when the proportion is $10\%$ , it means that $90\%$ + +Table 2. Clustering performance comparison of eight anchor-based MVC methods on seven datasets. The best is marked in bold and underlined, the second best is marked in bold. + +
DatasetsLMVSCSMVSCFMCNOFFPMVS-CAGUDBGLFastMICEFDAGFFMVACCProposed
ACC
Dermatology79.02±0.0778.64±0.0562.01±0.0082.96±0.0784.08±0.0087.15±0.0084.41±7.9084.39±4.3493.20±4.50
ForestTypes77.68±0.0472.82±0.0145.51±0.0072.22±0.0577.06±0.0077.06±0.0076.40±7.1379.29±0.2382.53±0.09
BDGP49.52±0.0237.22±0.0231.08±0.0032.62±0.0139.28±0.0049.00±0.0047.30±3.1659.51±4.0060.65±0.62
Reuters27.19±0.0124.41±0.0119.25±0.0017.33±0.0022.32±0.0026.00±0.0025.04±0.5027.22±0.2429.91±0.02
Hdigit86.87±0.0865.99±0.0332.84±0.0064.22±0.0531.39±0.0085.13±0.0077.69±7.8386.55±4.0888.01±2.97
VGGFace6.09±0.007.21±0.003.47±0.006.02±0.005.39±0.005.26±0.006.50±0.336.68±0.168.35±0.07
CIFAR1008.92±0.008.48±0.004.40±0.007.46±0.007.83±0.009.81±0.007.47±0.137.04±0.1210.74±0.08
NMI
Dermatology70.17±0.0466.62±0.0354.24±0.0071.90±0.0587.41±0.0082.73±0.0079.70±4.3476.27±2.1988.73±3.01
ForestTypes53.56±0.0246.56±0.0116.99±0.0048.20±0.0452.04±0.0056.53±0.0054.80±5.3555.33±0.1559.63±0.09
BDGP25.85±0.029.85±0.0110.29±0.0010.02±0.0114.90±0.0026.03±0.0021.90±1.9535.55±4.5330.92±0.65
Reuters5.67±0.015.45±0.010.64±0.000.13±0.001.79±0.007.23±0.003.36±0.485.05±0.087.07±0.01
Hdigit89.33±0.0356.16±0.0228.57±0.0055.80±0.0319.04±0.0088.67±0.0070.00±2.6076.96±3.0978.09±1.59
VGGFace11.92±0.0014.13±0.005.81±0.0012.33±0.0011.00±0.0010.79±0.0012.32±0.3312.97±0.1514.33±0.11
CIFAR10015.49±0.0014.83±0.008.49±0.0013.87±0.0014.83±0.0016.82±0.0013.93±0.1511.45±0.1616.94±0.14
Purity
Dermatology80.97±0.0480.35±0.0462.85±0.0083.55±0.0784.64±0.0087.15±0.0087.86±4.5185.38±2.0494.03±2.35
ForestTypes77.99±0.0272.82±0.0145.51±0.0072.73±0.0477.06±0.0078.93±0.0077.61±4.7479.29±0.2382.53±0.09
BDGP49.64±0.0237.80±0.0133.32±0.0034.82±0.0140.16±0.0049.44±0.0047.36±3.0859.73±3.6760.65±0.62
Reuters28.22±0.0025.12±0.0119.78±0.0017.89±0.0023.31±0.0028.07±0.0025.45±0.5927.50±0.1730.84±0.02
Hdigit88.93±0.0666.28±0.0332.85±0.0064.34±0.0531.94±0.0088.43±0.0079.82±5.6686.59±4.0188.22±2.25
VGGFace7.02±0.007.73±0.003.50±0.006.33±0.005.77±0.006.11±0.007.17±0.337.47±0.179.43±0.08
CIFAR10010.41±0.009.09±0.004.49±0.007.84±0.008.69±0.0011.11±0.008.28±0.137.89±0.1412.29±0.13
+ +Table 3. Ablation study on anchor enhancement module. + +
MethodsDermatologyForestTypesBDGP
Baseline90.87±3.6070.76±0.0554.38±0.10
+Align91.82±3.7875.51±0.1757.97±0.09
+Align+Enh.93.20±4.5082.53±0.0960.65±0.62
+ +of the reconstructed anchors and $10\%$ of random anchors are included during the enhancement process. From the results, as the proportion of randomly generated anchors increases, the clustering performance on both datasets continuously deteriorates, thus validating the effectiveness of our enhanced anchors. + +To verify the two properties of enhanced anchors, we calculate the values on both sides of the equations in the properties on two datasets. As shown in Fig. 3(b), the difference between the anchor graphs in the two views before and after anchor enhancement decreases $\left(\left\| \tilde{\mathbf{Z}}_1 - \tilde{\mathbf{Z}}_2\right\|_{\mathbf{F}} < \left\| \mathbf{Z}_1 - \mathbf{Z}_2\right\|_{\mathbf{F}}\right)$ , validating Property 1. Additionally, the difference between the anchor graph and the optimal consistent anchor graph reduces before and after anchor enhancement $\left(\left\| \tilde{\mathbf{Z}}_1 - \mathbf{Z}\right\|_{\mathbf{F}} < \left\| \mathbf{Z}_1 - \mathbf{Z}\right\|_{\mathbf{F}}\right)$ , validating Property 2. + +# 4.4. Effectiveness of Anchor Graph Construction + +To further validate the effectiveness of the revised anchor graph construction module, we contrast our proposed approach with two versions without the view weight and + +![](images/f5f79e0224c7f545cc3586d8c56a5a1308f5e5294ab93cc71c4c9eea90663ab9.jpg) +(a) Validation of enhancement strategy + +![](images/78325abf0eb9f3e1177bbc8ac387b91002c10a59a6c81a0f8e630be974867eab.jpg) +(b) Validation of Property 1 & 2 +Figure 3. Validation of the proposed AEVC's effectiveness: (a) As the proportion of randomly generated anchors increases, the corresponding clustering performance decreases monotonically. (b) After anchor enhancement, the differences in anchor graphs between different views are reduced, and the enhanced anchor graph is closer to the optimal consensus anchor graph. + +Table 4. Ablation study on revised anchor graph construction module. Other results are provided in supplementary materials. + +
MethodsDermatologyForestTypesBDGP
w/o. view weight91.60±6.2681.45±0.0060.50±0.46
w/o. regularized term89.98±3.4079.75±0.1060.64±0.62
Proposed93.20±4.5082.53±0.0960.65±0.62
+ +regularized term respectively, as illustrated in Table 4. After incorporating weights learned from the views during fusion, there is a noticeable improvement in the clustering performance of the method. This highlights the advantage of assessing the importance of views in multi-view clustering + +Table 5. Performance comparison between various anchor-based MVC w./w.o. our AEVC strategy. + +
MethodsDermatologyForestTypesBDGPReutersHdigitVGGFaceCIFAR100
Fixed Anchors Methods
FMCNOF62.01±0.0045.51±0.0031.08±0.0019.25±0.0032.84±0.003.47±0.004.40±0.00
FMCNOF-AE73.18±0.0069.60±0.0037.64±0.0022.79±0.0034.31±0.003.21±0.005.05±0.00
LMVSC79.02±0.0777.68±0.0449.52±0.0227.19±0.0186.87±0.086.09±0.008.92±0.00
LMVSC-AE88.23±7.8080.90±5.4952.13±0.3428.84±0.6785.74±5.726.83±0.168.81±0.16
Updated Anchors Methods
UDBGL84.08±0.0077.06±0.0039.28±0.0022.32±0.0031.39±0.005.39±0.007.83±0.00
UDBGL-AE81.01±0.0077.82±0.0046.92±0.0028.53±0.0044.60±0.005.65±0.008.57±0.00
FMVACC84.39±4.3479.29±0.2359.51±4.0027.22±0.2486.55±4.086.68±0.167.04±0.12
FMVACC-AE88.41±3.1381.08±0.2359.36±3.0333.12±2.4988.42±2.937.13±0.177.28±0.15
+ +through the proposed strategy. After removing the regularization term, there was a decline in performance across different datasets, which further confirms the effectiveness of constructing anchor graphs guided by $\hat{\mathbf{Z}}$ . + +# 4.5. Transferability Analysis + +To validate the transferability of AEVC, Table 5 presents a comparison of clustering accuracy between various types of anchor-based multi-view clustering methods with and without our anchor enhancement module. In this experiment, we selected two methods from each of the two categories employing different anchor strategies. Specifically, FMCNOF and LMVSC represent methods with fixed anchors, while UDBGL and FMVACC belong to methods with iteratively optimized anchors. As observed in the table, the clustering performance of the four anchor-based multi-view clustering methods experiences significant improvement across the majority of datasets after incorporating our proposed module. For instance, on the Reuters dataset, the clustering performance is notably enhanced by $18.4\%$ , $6.1\%$ , $27.8\%$ , and $21.7\%$ for FMCNOF, LMVSC, UDBGL, and FMVACC, respectively. This substantial improvement provides strong evidence for the effectiveness of AEVC when transferred to other anchor-based methods. + +# 4.6. Sensitivity Analysis + +Our proposed method involves three hyperparameters: the number of anchors $m$ , the anchor enhancement rate $\gamma$ , and the hyperparameter for balancing the regularization term $\lambda$ . Due to space constraints, we primarily analyze the sensitivity of the latter two parameters here, while the analysis regarding $m$ will be presented in the supplementary materials. To demonstrate the impact of $\gamma$ and $\lambda$ , we fix $m$ and conduct grid searches within specified ranges. Results presented in Fig. 4 indicate that our method is significantly influenced by the anchor enhancement rate, as the extreme rate can disrupt the balance between newly generated and original anchors. Specifically, on the Dermatology dataset, setting $\gamma$ to 100 yields the highest clustering accuracy. + +![](images/be366ee45049e67f42935c740ed0e7663931453db5efe43949f6d36133f3a076.jpg) +(a) Dermatology + +![](images/8aa03ee39764fb58b9cdb1186f1d15d6918641c6ec84e8550c89b2fe8e1b7f0a.jpg) +(b) ForestTypes +Figure 4. Sensitivity analysis of $\lambda$ and $\gamma$ on two benchmark datasets. Others are displayed in supplementary materials. + +# 5. Conclusion + +In this paper, we propose an Anchor Enhancement strategy with View Correlation for multi-view clustering termed AEVC. Different from existing anchor-based MVC methods, AEVC introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module. It can effectively enhance the quality of anchors, which is proven both experimentally and theoretically. Besides, we propose a revised anchor graph construction module based on the anchor enhancement strategy, which further improves the clustering performances. Experimental results show the promising capacity of AEVC, especially when integrating with other anchor-based methods. In the future, we will explore adaptive learning of enhancement rate. + +# 6. Acknowledgment + +This work was supported by the National Science Fund for Distinguished Young Scholars of China (No. 62325604), and the National Natural Science Foundation of China (No. 62276271 and No. 62306324). + +# References + +[1] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3 +[2] Jie Chen, Hua Mao, Dezhong Peng, Changqing Zhang, and Xi Peng. Multiview clustering by consensus spectral rotation fusion. IEEE Transactions on Image Processing, 2023. 1 +[3] Yongyong Chen, Xiaojia Zhao, Zheng Zhang, Youfa Liu, Jingyong Su, and Yicong Zhou. Tensor learning meets dynamic anchor learning: From complete to incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Siwei Wang, Xinzhong Zhu, Xinwang Liu, and En Zhu. Iterative deep structural graph contrast clustering for multiview raw data. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2023. 1 +[5] Zhibin Dong, Siwei Wang, Jiaqi Jin, Xinwang Liu, and En Zhu. Cross-view topology based consistent and complementary information for deep multi-view clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19440-19451, 2023. 3 +[6] Jürgen Dorn, Mario Girsch, Günther Skele, and Wolfgang Slany. Comparison of iterative improvement techniques for schedule optimization. European Journal of Operational Research, 94(2):349-361, 1996. 2, 3 +[7] Si-Guo Fang, Dong Huang, Xiao-Sha Cai, Chang-Dong Wang, Chaobo He, and Yong Tang. Efficient multi-view clustering via unified and discrete bipartite graph learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2, 3, 6 +[8] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2023. 1 +[9] Wenjue He, Zheng Zhang, Yongyong Chen, and Jie Wen. Structured anchor-inferred graph learning for universal incomplete multi-view clustering. World Wide Web, 26(1): 375-399, 2023. 3 +[10] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 2023. 1, 6 +[11] Yanyong Huang, Zongxin Shen, Yuxin Cai, Xiwen Yi, Dongjie Wang, Fengmao Lv, and Tianrui Li. Imufs: Complementary and consensus learning-based incomplete multiview unsupervised feature selection. IEEE Transactions on Knowledge and Data Engineering, 2023. 1 +[12] Jintian Ji and Songhe Feng. Anchor structure regularization induced multi-view subspace clustering via enhanced tensor rank minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19343-19352, 2023. 2 +[13] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11600-11609, 2023. 1 +[14] Zhao Kang, Wangtao Zhou, Zhitong Zhao, Junming Shao, Meng Han, and Zenglin Xu. Large-scale multi-view subspace clustering in linear time. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4412-4419, 2020. 1, 2, 4, 6 +[15] Borhan Kazimipour, Xiaodong Li, and A Kai Qin. Initialization methods for large scale global optimization. In 2013 IEEE congress on evolutionary computation, pages 2750-2757. IEEE, 2013. 2, 3 +[16] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 1 +[17] Lusi Li and Haibo He. Bipartite graph based multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2020. 1 +[18] Ruihuang Li, Changqing Zhang, Qinghua Hu, Pengfei Zhu, and Zheng Wang. Flexible multi-view representation learning for subspace clustering. In *IJCAI*, pages 2916–2922, 2019. 1 +[19] Xuelong Li, Han Zhang, Rong Wang, and Feiping Nie. Multi-view clustering: A scalable and parameter-free bipartite graph fusion method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 3 +[20] Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, and Fuchun Sun. Reasoning over different types of knowledge graphs: Static, temporal and multi-modal. arXiv preprint arXiv:2212.05767, 2022. 3 +[21] Ke Liang, Yue Liu, Sihang Zhou, Wenxuan Tu, Yi Wen, Xihong Yang, Xiangjun Dong, and Xinwang Liu. Knowledge graph contrastive learning based on relation-symmetrical structure. IEEE Transactions on Knowledge and Data Engineering, pages 1-12, 2023. 3 +[22] Ke Liang, Lingyuan Meng, Sihang Zhou, Siwei Wang, Wenxuan Tu, Yue Liu, Meng Liu, and Xinwang Liu. Message intercommunication for inductive relation reasoning. arXiv preprint arXiv:2305.14074, 2023. 3 +[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11174-11183, 2021. 2 +[24] Jialu Liu, Chi Wang, Jing Gao, and Jiawei Han. Multiview clustering via joint nonnegative matrix factorization. In Proceedings of the 2013 SIAM International Conference on Data Mining, pages 252-260. SIAM, 2013. 1 +[25] Suyuan Liu, Xinwang Liu, Siwei Wang, Xin Niu, and En Zhu. Fast incomplete multi-view clustering with view-independent anchors. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[26] Suyuan Liu, Siwei Wang, Pei Zhang, Kai Xu, Xinwang Liu, Changwang Zhang, and Feng Gao. Efficient one-pass multiview subspace clustering with consensus anchors. In Pro- + +ceedings of the AAAI Conference on Artificial Intelligence, pages 7576-7584, 2022. 2, 3 +[27] Suyuan Liu, Qing Liao, Siwei Wang, Xinwang Liu, and En Zhu. Robust and consistent anchor graph learning for multiview clustering. IEEE Transactions on Knowledge and Data Engineering, 2024. 1 +[28] Xun Lu and Songhe Feng. Structure diversity-induced anchor graph fusion for multi-view clustering. ACM Transactions on Knowledge Discovery from Data, 17(2):1-18, 2023. 3 +[29] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In International conference on machine learning, pages 5092-5101. PMLR, 2019. 2 +[30] Qianyao Qiang, Bin Zhang, Fei Wang, and Feiping Nie. Fast multi-view discrete clustering with anchor graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9360-9367, 2021. 2 +[31] Shaojun Shi, Feiping Nie, Rong Wang, and Xuelong Li. Fast multi-view clustering via prototype graph. IEEE Transactions on Knowledge and Data Engineering, 35(1):443-455, 2021. 1, 2 +[32] Xiaochuang Shu, Xiangdong Zhang, Quanxue Gao, Ming Yang, Rong Wang, and Xinbo Gao. Self-weighted anchor graph learning for multi-view clustering. IEEE Transactions on Multimedia, 2022. 1 +[33] Mengjing Sun, Pei Zhang, Siwei Wang, Sihang Zhou, Wenxuan Tu, Xinwang Liu, En Zhu, and Changjian Wang. Scalable multi-view subspace clustering with unified anchors. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3528-3536, 2021. 2, 3, 6 +[34] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 2 +[35] Xinhang Wan, Jiyuan Liu, Weixuan Liang, Xinwang Liu, Yi Wen, and En Zhu. Continual multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, page 3676-3684, New York, NY, USA, 2022. Association for Computing Machinery. 1 +[36] Xinhang Wan, Xinwang Liu, Jiyuan Liu, Siwei Wang, Yi Wen, Weixuan Liang, En Zhu, Zhe Liu, and Lu Zhou. Autoweighted multi-view clustering for large-scale data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10078-10086, 2023. 1 +[37] Jun Wang, Chang Tang, Zhiguo Wan, Wei Zhang, Kun Sun, and Albert Y Zomaya. Efficient and effective one-step multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2 +[38] Siwei Wang, Xinwang Liu, Li Liu, Wenxuan Tu, Xinzhong Zhu, Jiyuan Liu, Sihang Zhou, and En Zhu. Highly-efficient incomplete large-scale multi-view clustering with consensus bipartite graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9776-9785, 2022. 2 + +[39] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 3, 6 +[40] Siwei Wang, Xinwang Liu, Xinzhong Zhu, Pei Zhang, Yi Zhang, Feng Gao, and En Zhu. Fast parameter-free multiview subspace clustering with consensus anchor guidance. IEEE Trans. Image Process., 31:556-568, 2022. 3, 6 +[41] Yang Wang, Lin Wu, Xuemin Lin, and Junbin Gao. Multiview spectral clustering via structured low-rank matrix factorization. IEEE transactions on neural networks and learning systems, 29(10):4833-4843, 2018. 1 +[42] Yi Wen, Siwei Wang, Ke Liang, Weixuan Liang, Xinhang Wan, Xinwang Liu, Suyuan Liu, Jiyuan Liu, and En Zhu. Scalable incomplete multi-view clustering with structure alignment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3031-3040, 2023. 3 +[43] Yi Wen, Siwei Wang, Qing Liao, Weixuan Liang, Ke Liang, Xinhang Wan, and Xinwang Liu. Unpaired multi-view graph clustering with cross-view structure matching. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[44] Wei Xia, Quanxue Gao, Qianqian Wang, Xinbo Gao, Chris Ding, and Dacheng Tao. Tensorized bipartite graph learning for multi-view clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3 +[45] Weiqing Yan, Jindong Xu, Jinglei Liu, Guanghui Yue, and Chang Tang. Bipartite graph-based discriminative feature learning for multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3403-3411, 2022. 1, 2 +[46] Ben Yang, Xuetao Zhang, Feiping Nie, Fei Wang, Weizhong Yu, and Rong Wang. Fast multi-view clustering via nonnegative and orthogonal factorization. IEEE Transactions on Image Processing, 30:2575-2586, 2020. 2, 3, 6 +[47] Ben Yang, Xuetao Zhang, Zhiping Lin, Feiping Nie, Badong Chen, and Fei Wang. Efficient and robust multiview clustering with anchor graph regularization. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6200-6213, 2022. 1 +[48] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 1 +[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2 +[50] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In Proceedings of the 31st ACM International Conference on Multimedia, pages 337-346, 2023. 1 +[51] Jiali You, Zhenwen Ren, Xiaojian You, Haoran Li, and Yuancheng Yao. Priori anchor labels supervised scalable multi-view bipartite graph clustering. In Proceedings of the + +AAAI Conference on Artificial Intelligence, pages 10972-10979, 2023. 3 +[52] Shengju Yu, Siwei Wang, Yi Wen, Ziming Wang, Zhigang Luo, En Zhu, and Xinwang Liu. How to construct corresponding anchors for incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3 +[53] Chao Zhang, Huaxiong Li, Wei Lv, Zizheng Huang, Yang Gao, and Chunlin Chen. Enhanced tensor low-rank and sparse representation recovery for incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11174-11182, 2023. 3 +[54] Nan Zhang, Xiaoqin Zhang, and Shiliang Sun. Efficient multiview representation learning with correntropy and anchor graph. IEEE Transactions on Knowledge & Data Engineering, (01):1-14, 2023. 2 +[55] Pei Zhang, Siwei Wang, Liang Li, Changwang Zhang, Xinwang Liu, En Zhu, Zhe Liu, Lu Zhou, and Lei Luo. Let the data choose: flexible and diverse anchor graph fusion for scalable multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11262-11269, 2023. 2, 6 +[56] Xiaojia Zhao, Qiangqiang Shen, Yongyong Chen, Yongsheng Liang, Junxin Chen, and Yicong Zhou. Self-completed bipartite graph learning for fast incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3 \ No newline at end of file diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/images.zip b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dddb7853c5891a2dbc98248a1d2ca0653f2952e0 --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d734b4ef680ab01df02d275aed9554f26dd02d131cd438a1dadf52da2d4a528 +size 621467 diff --git a/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/layout.json b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f35ed49ec963db72300e4072f739051a3fb3c1d8 --- /dev/null +++ b/2024/Learn from View Correlation_ An Anchor Enhancement Strategy for Multi-view Clustering/layout.json @@ -0,0 +1,10619 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 50, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 545, + 140 + ], + "type": "text", + "content": "Learn from View Correlation: An Anchor Enhancement Strategy for Multi-view Clustering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 159, + 582, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 159, + 582, + 203 + ], + "spans": [ + { + "bbox": [ + 52, + 159, + 582, + 203 + ], + "type": "text", + "content": "Suyuan Liu1, Ke Liang1, Zhibin Dong1, Siwei Wang2, Xihong Yang1, Sihang Zhou1, En Zhu1, Xinwang Liu1* \n1National University of Defense Technology, Changsha, China \n2Intelligent Game and Decision Lab, Beijing, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 206, + 205, + 423, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 205, + 423, + 217 + ], + "spans": [ + { + "bbox": [ + 206, + 205, + 423, + 217 + ], + "type": "text", + "content": "{suyuanliu,xinwangliu,enzhu}@nudt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 271, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 290, + 547 + ], + "type": "text", + "content": "In recent years, anchor-based methods have achieved promising progress in multi-view clustering. The performances of these methods are significantly affected by the quality of the anchors. However, the anchors generated by previous works solely rely on single-view information, ignoring the correlation among different views. In particular, we observe that similar patterns are more likely to exist between similar views so such correlation information can be leveraged to enhance the quality of the anchors, which is also omitted. To this end, we propose a novel plug-and-play anchor enhancement strategy through view correlation for multi-view clustering. Specifically, we construct a view graph based on aligned initial anchor graphs to explore inter-view correlations. By learning from view correlation, we enhance the anchors of the current view using the relationships between anchors and samples on neighboring views, thereby narrowing the spatial distribution of anchors on similar views. Experimental results on seven datasets demonstrate the superiority of our proposed method over other existing methods. Furthermore, extensive comparative experiments validate the effectiveness of the proposed anchor enhancement module when applied to various anchor-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "Large-scale scenarios have always been a hot concern with the data explosion in our real world. Among these scenarios, large-scale multi-view clustering (MVC) has also drawn increasing attention in recent years[2, 4, 8, 18, 35]. In general, large-scale MVC methods can be categorized into two types, i.e., matrix factorization-based methods and anchor-based methods. The former methods [11, 24, 41, 50] aim to generate a low-dimensional representation for further clustering by decomposing the original data, thereby avoiding the computational costs of similarity matrix construc" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 542, + 366 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 542, + 366 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 542, + 366 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 542, + 366 + ], + "type": "image", + "image_path": "2696bbd21b59aa35ef1a797cc514518d9ff939b0579d1af45fb497bea0883d18.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 375, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 431 + ], + "type": "text", + "content": "Figure 1. An example illustration of clustering results affected by initial anchors. The first two methods keep the anchors fixed, while the latter two methods update the anchors through iteration-optimization. Their clustering performances are influenced to varying degrees by different anchor generation approaches." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "content": "tion. However, their efficiency usually decreases along with increasing dimension, thus constraining their scalability[10, 36]. Inspired by prototype learning [16, 48], anchors are introduced into MVC to reduce the scope for clustering from " + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "content": ", which denotes the number of samples and anchors, respectively. Specifically, anchor-based MVC methods [13, 27, 45] construct an anchor graph " + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "inline_equation", + "content": "(n \\times m)" + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "content": " rather than a similarity matrix " + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "inline_equation", + "content": "(n \\times n)" + }, + { + "bbox": [ + 304, + 446, + 547, + 555 + ], + "type": "text", + "content": ". It will alleviate the aforementioned computation and scalability problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 555, + 547, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 547, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 547, + 688 + ], + "type": "text", + "content": "Anchor-based MVC methods typically consist of three procedures, including (1) anchor generation, (2) anchor graph construction, and (3) clustering. Early attempts usually generate anchors at first, which are fixed during constructing anchor graphs[31, 32, 47]. For example, SFMC [19] selects anchors by iterative sampling based on feature similarity. Besides, LMVSC [14] regards cluster centers generated by " + }, + { + "bbox": [ + 304, + 555, + 547, + 688 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 555, + 547, + 688 + ], + "type": "text", + "content": "-means of each view as the anchors. However, these methods are sensitive to the initial anchors (See Fig. 1), where the anchor selection is primitive and sometimes results in compromised performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "To address the above issues, researchers make efforts on anchor optimization to guarantee better quality anchors[17," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "26151" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 263 + ], + "type": "text", + "content": "26, 38]. For example, SMVSC [33] proposed a unified framework that combines anchor generation and anchor-graph construction, achieving iterative updates to optimize the anchors. Building upon this, UDBGL [7] introduced the Laplacian rank constraint to ensure that the anchor graph possesses a discrete cluster structure. The above iteration-optimization strategies actually aim to get a better combination of the anchor and its corresponding anchor graph instead of enhancing the quality of the anchors themselves. Meanwhile, to the best of our knowledge, almost all existing anchor optimization models belong to this type. Although they can get better and more stable clustering performance, they still suffer from the initial state for optimization[6, 15], which leaves us a great challenge as well as the opportunity, i.e., what is the new strategy to alleviate such problem? In this paper, we provide a potential solution to it." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 265, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 287, + 396 + ], + "type": "text", + "content": "We observe that most anchor generation methods only rely on a single view[3, 12, 30]. Multiple works have proven that information of different views for the same sample will benefit the expressive ability of the sample [23, 29, 34, 49]. Similarly, different views for the same anchors should also be used to generate better anchors. Specifically, utilizing information across views and ensuring the consistency of anchor graphs on each view will benefit the anchor generation procedure. However, such a characteristic is not used for anchor enhancement in previous anchor-based MVC methods, and our strategy is developed based on it." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 398, + 289, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 289, + 649 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 289, + 649 + ], + "type": "text", + "content": "In this work, we propose an Anchor Enhancement strategy with View Correlation (AEVC) for multi-view clustering (MVC). The framework of AEVC is shown in Fig. 2. The proposed approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which enhance the quality of anchors. Additionally, we improve the clustering performance by revising the existing anchor graph construction method based on view correlation. Specifically, we first align the initial anchors and construct a view-graph based on the aligned anchor graph to find the neighbors of each view. Then, new anchors for the current view are generated with the anchor graph of neighboring views, thereby enhancing the initial anchors. Finally, the inter-view correlations are combined to learn a consistent anchor graph between views with the enhanced anchors. AEVC optimizes anchors by absorbing information from similar views, improving initial anchor quality, and reducing distribution bias caused by the independent anchors generation in each view. The main contributions of our work can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 654, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 700 + ], + "type": "text", + "content": "- We propose AEVC, a plug-and-play anchor enhancement strategy for MVC based on view correlation, to address the issue of poor initial anchor quality affecting the performance of MVC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "- By learning from the view correlation, AEVC incorpo" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 304, + 71, + 548, + 230 + ], + "blocks": [ + { + "bbox": [ + 304, + 71, + 548, + 230 + ], + "lines": [ + { + "bbox": [ + 304, + 71, + 548, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 71, + 548, + 230 + ], + "type": "image", + "image_path": "89574ce77c2e9aa7ee1accf6245c120eb95e5f018a0f390a605f7205ac8c126c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 238, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 304, + 238, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 547, + 304 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed AEVC. Specifically, in the view-graph learning module, AEVC primarily performs anchor alignment and learns the view-graph. In the anchor enhancement module, AEVC leverages the neighboring relationships from the view-graph to enhance the anchors. Finally, AEVC executes anchor graph construction based on the enhanced anchors." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 545, + 376 + ], + "type": "text", + "content": "rates the anchor distribution information from neighboring views into the process of anchor enhancement. Theoretical and experimental validations demonstrate that the anchors enhanced by AEVC exhibit a tighter inter-view relationship." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 376, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 436 + ], + "type": "text", + "content": "- We compared our proposed approach to other anchor-based multi-view methods and demonstrated its superiority. Extensive experiments verify the effectiveness and generalization of the proposed AEVC strategy when integrated with typical anchor-based MVC methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 448, + 392, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 448, + 392, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 448, + 392, + 460 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 468, + 500, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 500, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 500, + 480 + ], + "type": "text", + "content": "2.1. Anchor-based Multi-view Clustering" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 665 + ], + "type": "text", + "content": "To address the challenges posed by traditional multiview clustering methods in handling large-scale datasets, Anchor-based MVC has been proposed in recent years[37, 45, 55]. The core idea of anchor-based MVC is to generate a set of representative anchors from the entire sample pool, where the number of anchors is typically much smaller than the total number of samples[25, 54]. In contrast to conventional methods that require establishing relationships among all samples, anchor-based approaches only involve computing similarities between anchors and samples. This significantly reduces the complexity associated with constructing the relationship graph. Additionally, adopting a reduced-dimensional graph further mitigates computational burdens, facilitating subsequent multi-view fusion and clustering with reduced computational overhead." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Traditional anchor-based MVC methods generate anchors by sampling[30, 31, 46]. Subsequent anchor graph construction and clustering are performed based on the initially determined anchors. For instance, Kang et al.[14]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26152" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "and Yang et al. [46] advocate applying " + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "-means to the original samples, selecting cluster centers as anchors to enhance their representativeness. Xia et al. [44] introduces a variance-based de-correlation anchor selection strategy, leveraging the internal data structure to augment the coverage capability of anchors. The performance of such methods is highly influenced by the quality of the initially generated anchors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 177, + 209, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 209, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 209, + 190 + ], + "type": "text", + "content": "2.2. Anchor Optimization Strategy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 197, + 289, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 197, + 289, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 289, + 388 + ], + "type": "text", + "content": "To tackle the above challenges, extensive research efforts have been devoted to anchor optimization in order to achieve superior anchor quality. Iteration-optimization methods propose to update anchors and anchor graphs, utilizing the anchor graph obtained from the previous iteration to update the anchors[51, 53, 56]. Sun et al. [33] initially propose learning anchors and constructing anchor graphs through iterative updates within the same framework. Liu et al. [26] further introduce Laplacian rank constraints, enabling the anchor graph to generate labels without directly needing post-processing. However, the results of iterative update algorithms are significantly influenced by the initial values of variables[6, 15]. Instead of directly enhancing the quality of the anchors themselves, the above methods primarily focus on obtaining an improved combination of the anchor and its corresponding anchor graph." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 397, + 197, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 197, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 197, + 410 + ], + "type": "text", + "content": "2.3. View Correlation Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 416, + 289, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 289, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 289, + 657 + ], + "type": "text", + "content": "Multi-view and multi-modal scenarios are widely seen nowadays [1, 20]. Among them, investigating the relationships between multiple objects is a crucial topic [5, 21, 22], especially for multi-view learning, which holds for anchor-based multi-view clustering[7, 44]. Some existing methods delve into the consistency of views by aligning anchors[42, 43, 52]. For instance, Li et al. [19] concatenates raw data from all views to select anchors, ensuring correspondence among anchors across views. Wang et al. [39] takes a step further by proposing an anchor alignment framework based on the features and structural information of anchors. Another category of methods employs early fusion to avoid conflicts between views[9, 28]. For instance, Wang et al. [40] use a projection matrix to learn shared anchors and anchor graphs in the latent space. However, the first category of methods solely utilizes alignment information between views, while the second category underutilizes complementary information among views. In the next section, we will introduce a novel approach based on a view graph, leveraging inter-view interactions to enhance anchors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 669, + 102, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 102, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 102, + 680 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "In this section, we introduce our proposed Anchor Enhancement strategy with View Correlation (AEVC) for" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 131 + ], + "type": "text", + "content": "multi-view clustering. We begin by sequentially describing the three modules of AEVC. Subsequently, a complexity analysis of the proposed algorithm is provided. Lastly, a theoretical analysis is conducted to evaluate the effectiveness of AEVC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 142, + 373, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 142, + 373, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 142, + 373, + 153 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 161, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 161, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 161, + 545, + 365 + ], + "type": "text", + "content": "Existing anchor-based MVC methods can be divided into three steps: (1) anchor generation, (2) anchor graph construction, (3) clustering. The clustering performance of existing anchor-based MVC methods heavily relies on the quality of initially generated anchors. Although iteration-optimization strategies alleviate this issue by exploring combinations of anchors and anchor graphs, they are still impacted by the optimization of the initial state. To address this problem, we propose an anchor enhancement strategy based on view correlation for multi-view clustering. Specifically, different from existing methods, our approach introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module, which helps to enhance the quality of anchors. Furthermore, we revise the existing anchor graph construction method based on view correlation, further improving the clustering performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 375, + 430, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 375, + 430, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 375, + 430, + 388 + ], + "type": "text", + "content": "3.2. View-graph Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": "Given multi-view data " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_i\\}_{i=1}^v" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i \\in \\mathbb{R}^{d_i \\times n}" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " is composed of " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " samples with " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " dimension. Denoting " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^{(0)} \\in \\mathbb{R}^{d_i \\times m}" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " as the generated anchor matrix in the " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": "-th view, and " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i^{(0)} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " represents the corresponding anchor graph, where " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " is the number of anchors. Due to the varying dimensions and inconsistent distributions of data across different views, it is not feasible to directly assess the distance between " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_j" + }, + { + "bbox": [ + 304, + 393, + 545, + 517 + ], + "type": "text", + "content": ". However, it is easy to calculate the similarity between anchor graphs from different views due to the same size of anchor graphs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 518, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 601 + ], + "type": "text", + "content": "Nevertheless, the independently generated anchors on each view introduce the Anchor-Unaligned Problem [39]. This means that directly computing the similarity between anchor graphs of different views would be inaccurate due to the misalignment of rows corresponding to anchors. Thus, an anchor alignment operation must be performed before view-graph learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": "Existing anchor alignment strategies typically directly align the anchors on the first view, which carries the risk of being influenced by noise view. Here, we first concatenate the multi-view data column-wise to obtain " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = [\\mathbf{X}_1; \\ldots; \\mathbf{X}_v] \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times n}" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": ". Then, we generate the anchors " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}} \\in \\mathbb{R}^{\\sum_{i=1}^{v} d_i \\times m}" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": " and the corresponding anchor graph " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}} \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": ". According to [39], we can obtain the permutation matrix " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": "-th view by minimizing the distance between " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": " and the permuted " + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i^{(0)}" + }, + { + "bbox": [ + 304, + 603, + 545, + 713 + ], + "type": "text", + "content": ", the formula is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26153" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 93, + 83 + ], + "type": "text", + "content": "as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 91, + 287, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 91, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 91, + 287, + 117 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {P} _ {i}} \\left\\| \\hat {\\mathbf {Z}} - \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {1}", + "image_path": "83c6663580548736fe07e757357eec2254c69ac511c6c75c0cadedb55195504b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 118, + 253, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 118, + 253, + 131 + ], + "spans": [ + { + "bbox": [ + 81, + 118, + 253, + 131 + ], + "type": "interline_equation", + "content": "\\mathrm {s . t .} \\mathbf {P} _ {i} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} ^ {\\top} \\mathbf {1} = \\mathbf {1}, \\mathbf {P} _ {i} \\in \\{0, 1 \\} ^ {m \\times m}.", + "image_path": "570c2f2e7c9d2ee161ecdc642d0c806c8a261fff6ac42da2fa8ac6998059c5f7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 163 + ], + "type": "text", + "content": "Then the anchor matrix and anchor graph can be aligned with the permutation matrix as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 172, + 287, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 172, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 125, + 172, + 287, + 212 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\mathbf {A} _ {i} = \\mathbf {A} _ {i} ^ {(0)} \\mathbf {P} _ {i} ^ {\\top}, \\\\ \\mathbf {Z} _ {i} = \\mathbf {P} _ {i} \\mathbf {Z} _ {i} ^ {(0)}. \\end{array} \\right. \\tag {2}", + "image_path": "2629b68de503392a039b7d6c50e7705d127047b2deb6b9cd6d8495724db67fde.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 220, + 287, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 220, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 287, + 244 + ], + "type": "text", + "content": "After performing the alignment operation, we calculate the view graph " + }, + { + "bbox": [ + 47, + 220, + 287, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{v \\times v}" + }, + { + "bbox": [ + 47, + 220, + 287, + 244 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 252, + 287, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 252, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 287, + 291 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\mathbf {s} _ {i j} = \\frac {1}{\\| \\mathbf {Z} _ {i} - \\mathbf {Z} _ {j} \\| _ {\\mathbf {F}} ^ {2}}, & i \\neq j, \\\\ \\mathbf {s} _ {i j} = 0, & i = j,. \\end{array} \\right. \\tag {3}", + "image_path": "c981eb3c7900140143157868b8d534377e8a2fe3da9ea1039dca8a8446748ba0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{ij}" + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "text", + "content": " denotes the similarity between the " + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "text", + "content": "-th and the " + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 47, + 300, + 287, + 323 + ], + "type": "text", + "content": "-th view." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 331, + 214, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 214, + 344 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 214, + 344 + ], + "type": "text", + "content": "3.3. Anchor Enhancement Strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 350, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 539 + ], + "type": "text", + "content": "Most existing methods independently generate anchors on each view, ignoring the correlation between views. In fact, anchors on similar views should have similar distributions, which means that the corresponding anchor graphs should be close to each other. Based on this assumption, we propose to enhance the anchors of the current view with information on neighboring views. One intuitive idea is to directly update the anchors based on the anchors of neighboring views. However, the dimensions of anchors on different views are not the same, making direct transfer infeasible. Instead of utilizing feature consistency, we focus on structural consistency. Specifically, we reconstruct new anchors by leveraging the anchor graphs from neighboring views and the data from the current view, thereby enhancing the initial anchors. The formulation of the anchor enhancement on the " + }, + { + "bbox": [ + 47, + 350, + 287, + 539 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 350, + 287, + 539 + ], + "type": "text", + "content": "-th view is as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 87, + 548, + 287, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 548, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 87, + 548, + 287, + 578 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {A}} _ {i} = \\frac {1}{\\mu} \\left(\\mathbf {A} _ {i} + \\gamma \\sum_ {\\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right)} \\hat {s} _ {i j} \\mathbf {X} _ {i} \\mathbf {Z} _ {j} ^ {\\top}\\right). \\tag {4}", + "image_path": "2ab4ec9b80d941e2432043e6c91e4ad31dc2931cdc29c6dc1b31dcd78c836e8c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mu = 1 + \\gamma \\sum_{\\mathbf{X}_j \\in \\Omega_\\zeta(\\mathbf{X}_i)} \\hat{s}_{ij}" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": " ensures that the enhanced anchors are a convex combination of the initial anchors and the reconstructed anchors. " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\Omega_\\zeta(\\mathbf{X}_i)" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": "-neighbors of the " + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 587, + 287, + 635 + ], + "type": "text", + "content": "-th view, which is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 643, + 287, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 643, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 287, + 668 + ], + "type": "interline_equation", + "content": "\\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right) = \\arg \\max _ {\\mathbf {X} _ {j}} \\sum_ {j = 1} | s _ {i j} |. \\tag {5}", + "image_path": "601c917a1ad2a6030e2e63de157a294b09c093a41125bbe60ca1c3f6f2bbf9ef.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is the enhancement rate, which is used to adjust the influence of the initial anchors on the enhanced version. " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " represents the normalized view-graph based on the following" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 95 + ], + "type": "text", + "content": "settings, which controls the influence of different neighbors during anchor enhancement:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 104, + 545, + 139 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 104, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 364, + 104, + 545, + 139 + ], + "type": "interline_equation", + "content": "\\hat {s} _ {i j} \\left\\{ \\begin{array}{l l} s _ {i j}, & \\mathbf {X} _ {j} \\in \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right), \\\\ 0, & \\mathbf {X} _ {j} \\notin \\Omega_ {\\zeta} \\left(\\mathbf {X} _ {i}\\right). \\end{array} \\right. \\tag {6}", + "image_path": "bd39f33145d19b5dc6aa403c91c7638e4d5d1048de41d6b12ae02c84e37ab84d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 146, + 545, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 146, + 545, + 264 + ], + "spans": [ + { + "bbox": [ + 305, + 146, + 545, + 264 + ], + "type": "text", + "content": "By enhancing the anchors with Eq. (4), the quality of the initial anchors is improved. Moreover, the enhancement further achieves a better alignment between anchors on similar views. In the subsequent sections, we will analyze the aforementioned conclusions from both theoretical and experimental perspectives. Furthermore, our proposed anchor enhancement module is plug-and-play and we will showcase its performance improvement in the experiments by integrating different anchor-based multi-view clustering methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 273, + 462, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 273, + 462, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 462, + 285 + ], + "type": "text", + "content": "3.4. Anchor Graph Construction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "text", + "content": "After obtaining high-quality anchors through our proposed anchor enhancement module, it is crucial to construct a reliable anchor graph that ensures view consistency. Since the enhanced anchors are aligned, we can directly generate a common anchor graph. However, the contributions of different views in constructing the anchor graph are not equal. To mitigate the influence of view discrepancies on building a consistent representation, we assign higher weights to views that have stronger associations with other views. Specifically, the contribution weight " + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\omega_{i} = \\sum_{j=1}^{v} s_{ij}" + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "text", + "content": "-th view is computed as the sum of its similarities with all other views. Furthermore, to avoid the impact of noise during anchor graph construction, we introduce the learned anchor graph " + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 305, + 291, + 545, + 481 + ], + "type": "text", + "content": " from the concatenated views as guidance. Ultimately, the anchor graph construction is formulated as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 342, + 487, + 545, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 487, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 342, + 487, + 545, + 513 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {Z}} \\omega_ {i} \\left\\| \\mathbf {X} _ {i} - \\tilde {\\mathbf {A}} _ {i} \\mathbf {Z} \\right\\| _ {\\mathbf {F}} ^ {2} + \\lambda \\left\\| \\mathbf {Z} - \\hat {\\mathbf {Z}} \\right\\| _ {\\mathbf {F}} ^ {2}, \\tag {7}", + "image_path": "a9be63b5fcc917e7eec7cd2bbb919ddb1870c7e3c65a86527228af6170cd60ab.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 382, + 513, + 469, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 513, + 469, + 525 + ], + "spans": [ + { + "bbox": [ + 382, + 513, + 469, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {s . t .} \\mathbf {Z} \\geq 0, \\mathbf {Z} ^ {\\top} \\mathbf {1} = \\mathbf {1}. \\end{array}", + "image_path": "66601548a40a249ef934dbf0d429a625f89f7168297ba93396697b2794b2824f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": "Due to the independence between each column of " + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": ". The aforementioned problem can be transformed into " + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 306, + 534, + 545, + 582 + ], + "type": "text", + "content": " Quadratic Programming(QP) problems, formulated as follows:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 381, + 589, + 545, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 589, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 381, + 589, + 545, + 610 + ], + "type": "interline_equation", + "content": "\\min \\frac {1}{2} \\mathbf {z} _ {j} ^ {\\top} \\mathbf {H} \\mathbf {z} _ {j} + f \\mathbf {z} _ {j}, \\tag {8}", + "image_path": "8e3de09cc49e69f97ab154d5649d3b9df139d38f661de526d55df13b0852a6fb.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 383, + 606, + 470, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 606, + 470, + 620 + ], + "spans": [ + { + "bbox": [ + 383, + 606, + 470, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {s . t .} \\mathbf {z} _ {j} \\geq 0, \\mathbf {z} _ {j} ^ {\\top} \\mathbf {1} = 1, \\end{array}", + "image_path": "2584ae7a5768b4f2e7bbd209cb66254806cdcd03e5b848b9d4b28c9f32952c2f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "where the matrix " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H} = 2\\left(\\sum_{i=1}^{v}\\omega_{i}\\tilde{\\mathbf{A}}_{i}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\mathbf{I}\\right)" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " and the vector " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "f = -2\\left(\\sum_{i=1}^{v}\\omega_{i}\\mathbf{X}_{i[:,j]}^{\\top}\\tilde{\\mathbf{A}}_{i} + \\lambda \\hat{z}_{j}\\right)" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": ". The optimal " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " can be obtained by solving the above QP problem. Finally, the results can be generated by performing " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": "-means on the left singular vector of " + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 305, + 629, + 545, + 713 + ], + "type": "text", + "content": " [14]. The entire procedure of AEVC is shown in Algorithm 1." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26154" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 75, + 198, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 75, + 198, + 87 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 198, + 87 + ], + "type": "text", + "content": "Algorithm 1: The proposed AEVC" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "text", + "content": "Input: Multi-view datasets " + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_i\\}_{i = 1}^v" + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "text", + "content": " , the number of anchors " + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "text", + "content": " , the parameter " + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "text", + "content": " and the number of clusters " + }, + { + "bbox": [ + 59, + 89, + 271, + 125 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 60, + 125, + 171, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 125, + 171, + 137 + ], + "spans": [ + { + "bbox": [ + 60, + 125, + 171, + 137 + ], + "type": "text", + "content": "Output: Clustering results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 140, + 265, + 297 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "spans": [ + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "text", + "content": "1 Initialize " + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_i^{(0)}\\}_{i = 1}^v" + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_i^{(0)}\\}_{i = 1}^v" + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 52, + 140, + 232, + 167 + ], + "type": "text", + "content": " // View-graph Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 168, + 133, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 168, + 133, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 168, + 133, + 178 + ], + "type": "text", + "content": "2 for " + }, + { + "bbox": [ + 51, + 168, + 133, + 178 + ], + "type": "inline_equation", + "content": "i = 1\\to v" + }, + { + "bbox": [ + 51, + 168, + 133, + 178 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 179, + 261, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 179, + 261, + 205 + ], + "spans": [ + { + "bbox": [ + 52, + 179, + 261, + 205 + ], + "type": "text", + "content": "3 Calculate permutation matrix " + }, + { + "bbox": [ + 52, + 179, + 261, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 52, + 179, + 261, + 205 + ], + "type": "text", + "content": " with Eq. (1); 4 Perform alignment with Eq. (2);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 207, + 228, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 207, + 228, + 230 + ], + "spans": [ + { + "bbox": [ + 52, + 207, + 228, + 230 + ], + "type": "text", + "content": "5 Learn view-graph S as defined in Eq. (3); // Anchor Enhancement" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 232, + 133, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 232, + 133, + 242 + ], + "spans": [ + { + "bbox": [ + 52, + 232, + 133, + 242 + ], + "type": "text", + "content": "6 for " + }, + { + "bbox": [ + 52, + 232, + 133, + 242 + ], + "type": "inline_equation", + "content": "i = 1\\to v" + }, + { + "bbox": [ + 52, + 232, + 133, + 242 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "spans": [ + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "text", + "content": "7 Find " + }, + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "text", + "content": "-neighbors " + }, + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "inline_equation", + "content": "\\Omega_{\\zeta}(\\mathbf{X}_i)" + }, + { + "bbox": [ + 51, + 243, + 265, + 257 + ], + "type": "text", + "content": " as defined in Eq. (5);" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 51, + 259, + 257, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 259, + 257, + 272 + ], + "spans": [ + { + "bbox": [ + 51, + 259, + 257, + 272 + ], + "type": "text", + "content": "8 Calculate normalized view-graph " + }, + { + "bbox": [ + 51, + 259, + 257, + 272 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 51, + 259, + 257, + 272 + ], + "type": "text", + "content": " with Eq. (6);" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 51, + 273, + 133, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 273, + 133, + 282 + ], + "spans": [ + { + "bbox": [ + 51, + 273, + 133, + 282 + ], + "type": "text", + "content": "9 for " + }, + { + "bbox": [ + 51, + 273, + 133, + 282 + ], + "type": "inline_equation", + "content": "i = 1\\rightarrow v" + }, + { + "bbox": [ + 51, + 273, + 133, + 282 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 284, + 180, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 180, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 180, + 297 + ], + "type": "text", + "content": "10 Enhance " + }, + { + "bbox": [ + 48, + 284, + 180, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 48, + 284, + 180, + 297 + ], + "type": "text", + "content": " with Eq. (4);" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 300, + 258, + 335 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 60, + 300, + 229, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 300, + 229, + 311 + ], + "spans": [ + { + "bbox": [ + 60, + 300, + 229, + 311 + ], + "type": "text", + "content": "// Anchor Graph Construction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 312, + 176, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 312, + 176, + 323 + ], + "spans": [ + { + "bbox": [ + 49, + 312, + 176, + 323 + ], + "type": "text", + "content": "11 Obtain " + }, + { + "bbox": [ + 49, + 312, + 176, + 323 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 49, + 312, + 176, + 323 + ], + "type": "text", + "content": " by solving Eq. (7);" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "text", + "content": "12 Perform " + }, + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "text", + "content": "-means on the left singular vector of " + }, + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 48, + 324, + 258, + 335 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 364, + 119, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 364, + 119, + 374 + ], + "spans": [ + { + "bbox": [ + 47, + 364, + 119, + 374 + ], + "type": "text", + "content": "3.5. Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": "Computational complexity. The computational complexity of the proposed AEVC mainly consists of three modules. The complexity of the view-graph learning module is " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(m^3 v + nmv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ". Specifically, the computation of " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": " permutation matrices " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": " has a complexity of " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(m^3 v)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", and the computation of the view graph has a complexity of " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nmv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ". The time complexity of the anchor enhancement module is " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nmdv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "d = \\sum_{i=1}^{v} d_v" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", primarily arising from matrix multiplication during enhancement. The complexity of the anchor graph construction module is " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nm^3 v + nmdv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", where solving all QP problems requires a time complexity of " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nm^3 v)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", and the involved matrix multiplication requires " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nm^3 v + nmdv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ". Therefore, the overall computational complexity of AEVC is " + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nm^3 v + m^3 v + nmdv + nmv)" + }, + { + "bbox": [ + 46, + 382, + 287, + 574 + ], + "type": "text", + "content": ", which exhibits linear complexity with respect to the number of samples." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": "Space complexity. During the execution of the AEVC algorithm, the main variables that need to be stored include matrices " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{A}}" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{S}}" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ". The overall space complexity of AEVC is " + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(nd + ndv + nmdv + v^2)" + }, + { + "bbox": [ + 46, + 574, + 287, + 646 + ], + "type": "text", + "content": ", which is also linear with respect to the number of samples and can be extended to large-scale scenarios." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 647, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 694 + ], + "type": "text", + "content": "Two Property of enhanced anchors. We have discovered that the anchors enhanced by the proposed AEVC exhibit stronger inter-view consistency. The theoretical analysis is provided below." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": "Assumption 1. Given a two-view data " + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_1, \\mathbf{X}_2\\}" + }, + { + "bbox": [ + 47, + 701, + 287, + 714 + ], + "type": "text", + "content": ", denoting" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{A}_1,\\mathbf{A}_2\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " are the generated initial anchors, and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_1,\\mathbf{Z}_2\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " are the corresponding anchor graphs under the assumption that " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1 = \\mathbf{A}_1\\mathbf{Z}_1" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2 = \\mathbf{A}_2\\mathbf{Z}_2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 114, + 545, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 545, + 149 + ], + "type": "text", + "content": "For the convenience of derivation, we simplify the anchor enhancement strategy in Eq. (4) with " + }, + { + "bbox": [ + 304, + 114, + 545, + 149 + ], + "type": "inline_equation", + "content": "\\alpha \\in (0,1)" + }, + { + "bbox": [ + 304, + 114, + 545, + 149 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 356, + 155, + 545, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 155, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 356, + 155, + 545, + 188 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {A}} _ {1} = \\alpha \\mathbf {A} _ {1} + (1 - \\alpha) \\mathbf {X} _ {1} \\mathbf {Z} _ {2} ^ {\\top}, \\\\ \\tilde {\\mathbf {A}} _ {2} = \\alpha \\mathbf {A} _ {2} + (1 - \\alpha) \\mathbf {X} _ {2} \\mathbf {Z} _ {1} ^ {\\top}. \\end{array} \\right. \\tag {9}", + "image_path": "8c7234f9dc0cfef2d5f7f98d0cab9a8041eaa02b3e8137b743ddfbc11898fa20.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "content": "The corresponding anchor graphs " + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\mathbf{Z}}_1,\\tilde{\\mathbf{Z}}_2\\}" + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "content": " are constructed under the same assumption that " + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1 = \\tilde{\\mathbf{A}}_1\\tilde{\\mathbf{Z}}_1" + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2 = \\tilde{\\mathbf{A}}_2\\tilde{\\mathbf{Z}}_2" + }, + { + "bbox": [ + 305, + 195, + 545, + 232 + ], + "type": "text", + "content": ". Then we can derive that" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 365, + 238, + 545, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 238, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 365, + 238, + 545, + 270 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\tilde {\\mathbf {Z}} _ {1} = \\alpha \\mathbf {Z} _ {1} + (1 - \\alpha) \\mathbf {Z} _ {2}, \\\\ \\tilde {\\mathbf {Z}} _ {2} = \\alpha \\mathbf {Z} _ {2} + (1 - \\alpha) \\mathbf {Z} _ {1}. \\end{array} \\right. \\tag {10}", + "image_path": "11def0077fdf395ceb922b7a3c0cf92b84cc9241e37bc3dd30fa61738f1e3031.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 277, + 482, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 277, + 482, + 289 + ], + "spans": [ + { + "bbox": [ + 318, + 277, + 482, + 289 + ], + "type": "text", + "content": "We can deduce the following properties:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 296, + 545, + 331 + ], + "type": "text", + "content": "Property 1. The disparity among the anchor graphs on each view is reduced after anchor enhancement. Mathematically," + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 364, + 337, + 545, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 337, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 364, + 337, + 545, + 360 + ], + "type": "interline_equation", + "content": "\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}}. \\tag {11}", + "image_path": "2447a95f2156521046bb86b74a6022f771f7ca994c5ae3610b11db7de517100c.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 365, + 504, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 504, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 504, + 377 + ], + "type": "text", + "content": "Proof. According to Eq. (10), we can derive that" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 344, + 383, + 545, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 383, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 344, + 383, + 545, + 435 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\tilde {\\mathbf {Z}} _ {2} \\right\\| _ {\\mathbf {F}} = \\| (2 \\alpha - 1) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = \\left| 2 \\alpha - 1 \\right| \\cdot \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}} \\tag {12} \\\\ < \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}", + "image_path": "c852db67fbd130a048310d1344906f496e69c4df0dbae2fc06014a9c7d3c096b.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 441, + 410, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 410, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 410, + 453 + ], + "type": "text", + "content": "This completes the proof." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 458, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 458, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 458, + 545, + 495 + ], + "type": "text", + "content": "Property 1 implies that anchor enhancement leads to a more consistent distribution of anchors in similar views, thereby establishing a tighter alignment." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": "Property 2. Assuming there exists a value " + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\beta \\in (0,1)" + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\beta \\mathbf{Z}_1 + (1 - \\beta)\\mathbf{Z}_2" + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": " is the optimal consistent anchor graph among all views, the enhanced anchor graph is closer to the optimal version after enhancement than before for all " + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "inline_equation", + "content": "\\alpha \\in (2\\beta -1,1)\\cap (0,1)" + }, + { + "bbox": [ + 304, + 501, + 545, + 561 + ], + "type": "text", + "content": ". Mathematically," + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 368, + 567, + 545, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 567, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 368, + 567, + 545, + 589 + ], + "type": "interline_equation", + "content": "\\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} < \\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\| _ {\\mathbf {F}}. \\tag {13}", + "image_path": "134356b6c86b33d6253e18e0396d95d84c4e86ece10a61bb7511185db296d912.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 595, + 504, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 595, + 504, + 607 + ], + "spans": [ + { + "bbox": [ + 306, + 595, + 504, + 607 + ], + "type": "text", + "content": "Proof. According to Eq. (10), we can derive that" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 348, + 613, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 613, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 348, + 613, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\| \\tilde {\\mathbf {Z}} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}} = \\| (\\alpha - \\beta) (\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}) \\| _ {\\mathbf {F}} \\\\ = | \\alpha - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\\\ < | 1 - \\beta | \\cdot \\| \\mathbf {Z} _ {1} - \\mathbf {Z} _ {2} \\| _ {\\mathbf {F}} \\tag {14} \\\\ = \\left\\| (1 - \\beta) \\left(\\mathbf {Z} _ {1} - \\mathbf {Z} _ {2}\\right) \\right\\| _ {\\mathbf {F}} \\\\ = \\left\\| \\mathbf {Z} _ {1} - \\mathbf {Z} \\right\\| _ {\\mathbf {F}}. \\\\ \\end{array}", + "image_path": "493c17bbdddb401c4cabffc78c3edb7bf018ad6e5e601a9bb5c91a1c43a6e17e.jpg" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 701, + 410, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 410, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 410, + 713 + ], + "type": "text", + "content": "This completes the proof." + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 535, + 702, + 545, + 711 + ], + "blocks": [ + { + "bbox": [ + 535, + 702, + 545, + 711 + ], + "lines": [ + { + "bbox": [ + 535, + 702, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 535, + 702, + 545, + 711 + ], + "type": "image", + "image_path": "e1f847bc1d8bb9a2294dd727278e238a03f0f0c5d8e113a34d4780731614362e.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26155" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "Property 2 implies that anchor enhancement helps reduce the gap between the anchor graph on each view and the optimal consistent anchor graph. It is easy to derive the multiview extended versions of the above properties." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 134, + 123, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 134, + 123, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 134, + 123, + 148 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 155, + 155, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 155, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 155, + 168 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 175, + 287, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 293 + ], + "type": "text", + "content": "Datasets. Seven multi-view datasets are employed in our experiments: Dermatology1 contains diagnostic information for six types of skin diseases. ForestTypes2 consists of satellite images of different forest types. BDGP3 comprises both images and textual descriptions of Drosophila embryos. Reuters4 is a news text dataset. Hdigit5 consists of a set of handwritten digital images. VGGFace2 is a large-scale face recognition dataset. CIFAR100 is a subset of a large-scale tiny image dataset. Detail description is shown in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 296, + 288, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 288, + 487 + ], + "type": "text", + "content": "Compared methods. We compare the proposed AEVC with existing state-of-the-art anchor-based MVC methods: Large-scale multi-view subspace clustering in linear time (LMVSC)[14]; Scalable Multi-view Subspace Clustering with Unified Anchors (SMVSC)[33]; Fast multi-view clustering via nonnegative and orthogonal factorization (FMCNOF)[46]; Fast Parameter-Free Multi-View Subspace Clustering With Consensus Anchor Guidance (FPMVSCAG)[40]; Efficient Multi-View Clustering via Unified and Discrete Bipartite Graph Learning (UDBGL)[7]; Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity (FastMICE)[10]; Flexible and Diverse Anchor Graph Fusion for Scalable Multi-View Clustering (FDAGF)[55]; Align then Fusion: Generalized Large-scale Multi-view Clustering with Anchor Matching Correspondences (FMVACC)[39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": "Implementation details. For the above-compared algorithm, we have tuned their parameters to align with the relevant literature. We conducted a search for three hyperparameters in the proposed AEVC method using predefined lists. Specifically, the anchor number " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": " is chosen from the set " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "[k, 2k, 5k]" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": " represents the cluster number. The enhancement rate " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": " is selected from " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "[0.1, 1, 10, 100]" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": ", and the balancing hyperparameter " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": " is chosen from the range " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "[0.0001, 0.01, 1, 100]" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": ". To ensure a fair comparison, algorithms relying on " + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 488, + 287, + 632 + ], + "type": "text", + "content": "-means for result acquisition underwent 50 independent runs, and the results were averaged for accuracy assessment. Our experimental evaluation incorporated" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 83, + 543, + 177 + ], + "blocks": [ + { + "bbox": [ + 350, + 71, + 501, + 81 + ], + "lines": [ + { + "bbox": [ + 350, + 71, + 501, + 81 + ], + "spans": [ + { + "bbox": [ + 350, + 71, + 501, + 81 + ], + "type": "text", + "content": "Table 1. Benchmark datasets description." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 83, + 543, + 177 + ], + "lines": [ + { + "bbox": [ + 309, + 83, + 543, + 177 + ], + "spans": [ + { + "bbox": [ + 309, + 83, + 543, + 177 + ], + "type": "table", + "html": "
Datasetnvkdp(p=1,...,v)
Dermatology3582612/22
ForestTypes523349/9/9
BDGP2500351000/500/250
Reuters7200564819/4810/4892/4858/4777
Hdigit10000210784/256
VGGFace23628741003944/576/512/640
CIFAR100500004100944/576/512/640
", + "image_path": "f68f89be3ec5f40d29aa7a028cc64c3f366f26a111d3e20ed401db6d744f2236.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "type": "text", + "content": "four distinct metrics: clustering Accuracy (ACC), Normalized Mutual Information (NMI), Purity, and F-score." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 220, + 428, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 428, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 428, + 232 + ], + "type": "text", + "content": "4.2. Superiority of AEVC" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 239, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 285 + ], + "type": "text", + "content": "We compare our proposed AEVC with eight state-of-the-art anchor-based multi-view clustering methods on seven datasets, as illustrated in Table 2. From the results, we have the following observations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 289, + 545, + 431 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": "- The experimental results reveal that our proposed algorithm consistently outperforms other anchor-based MVC methods across all datasets, confirming the superiority of AEVC. In terms of ACC, AEVC surpasses the second-best algorithm by " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "9.9\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 306, + 289, + 545, + 360 + ], + "type": "text", + "content": " on different datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 361, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 545, + 431 + ], + "type": "text", + "content": "- Compared to previous methods based on iteratively optimized anchors (SMVSC, FPMVS-CAG, UDBGL, FDAGF, FMVACC), AEVC exhibits superior clustering performance, emphasizing the effectiveness of our anchor enhancement approach over iteration-optimization methods." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 443, + 493, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 493, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 493, + 455 + ], + "type": "text", + "content": "4.3. Effectiveness of Enhanced Anchors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 653 + ], + "type": "text", + "content": "We carried out an ablation study to demonstrate the effectiveness of the proposed anchor enhancement strategy. In our approach, the anchor enhancement strategy typically comprises two steps: anchor alignment in the view-graph learning module and anchor enhancement module. We denote the method of removing the entire anchor enhancement strategy as a baseline and compare it with those where anchor alignment and enhancement are sequentially incorporated. From the experimental results shown in Table 3, it is observed that anchor alignment leads to an improvement in clustering performance, as misalignment of anchors can result in incorrect multi-view fusion[39]. Furthermore, with the application of our anchor enhancement strategy, there is a further improvement in clustering performance, providing strong evidence for the effectiveness of the anchor enhancement strategy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of the enhanced anchors, we incorporate randomly generated anchors from the original data at different proportions and showcase the corresponding clustering performance in Fig. 3(a). For instance, when the proportion is " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": ", it means that " + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "inline_equation", + "content": "90\\%" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 643, + 229, + 654 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 643, + 229, + 654 + ], + "spans": [ + { + "bbox": [ + 61, + 643, + 229, + 654 + ], + "type": "text", + "content": "1 http://archive.ics.uci.edu/ml/datasets/Dermatology" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 654, + 187, + 663 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 654, + 187, + 663 + ], + "spans": [ + { + "bbox": [ + 63, + 654, + 187, + 663 + ], + "type": "text", + "content": "2http://archive.ics.uci.edu/dataset/333/" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 63, + 663, + 145, + 673 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 663, + 145, + 673 + ], + "spans": [ + { + "bbox": [ + 63, + 663, + 145, + 673 + ], + "type": "text", + "content": "3 https://www.fruitfly.org/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 63, + 673, + 253, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 673, + 253, + 683 + ], + "spans": [ + { + "bbox": [ + 63, + 673, + 253, + 683 + ], + "type": "text", + "content": "4http://www.researchAtt.com/%7Elewis/reuters21578.html" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 63, + 683, + 195, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 683, + 195, + 693 + ], + "spans": [ + { + "bbox": [ + 63, + 683, + 195, + 693 + ], + "type": "text", + "content": "5https://cs.nyu.edu/%7Erowei/data.html" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 63, + 693, + 224, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 693, + 224, + 703 + ], + "spans": [ + { + "bbox": [ + 63, + 693, + 224, + 703 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 63, + 693, + 224, + 703 + ], + "type": "text", + "content": "https://www.robots.ox.ac.uk/vgg/data/vgg_face2/" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 63, + 703, + 217, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 217, + 712 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 217, + 712 + ], + "type": "text", + "content": "7https://www.cs.toronto.edu/%7Ekriz/cifar.html" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26156" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 101, + 545, + 348 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 2. Clustering performance comparison of eight anchor-based MVC methods on seven datasets. The best is marked in bold and underlined, the second best is marked in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 101, + 545, + 348 + ], + "lines": [ + { + "bbox": [ + 49, + 101, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 101, + 545, + 348 + ], + "type": "table", + "html": "
DatasetsLMVSCSMVSCFMCNOFFPMVS-CAGUDBGLFastMICEFDAGFFMVACCProposed
ACC
Dermatology79.02±0.0778.64±0.0562.01±0.0082.96±0.0784.08±0.0087.15±0.0084.41±7.9084.39±4.3493.20±4.50
ForestTypes77.68±0.0472.82±0.0145.51±0.0072.22±0.0577.06±0.0077.06±0.0076.40±7.1379.29±0.2382.53±0.09
BDGP49.52±0.0237.22±0.0231.08±0.0032.62±0.0139.28±0.0049.00±0.0047.30±3.1659.51±4.0060.65±0.62
Reuters27.19±0.0124.41±0.0119.25±0.0017.33±0.0022.32±0.0026.00±0.0025.04±0.5027.22±0.2429.91±0.02
Hdigit86.87±0.0865.99±0.0332.84±0.0064.22±0.0531.39±0.0085.13±0.0077.69±7.8386.55±4.0888.01±2.97
VGGFace6.09±0.007.21±0.003.47±0.006.02±0.005.39±0.005.26±0.006.50±0.336.68±0.168.35±0.07
CIFAR1008.92±0.008.48±0.004.40±0.007.46±0.007.83±0.009.81±0.007.47±0.137.04±0.1210.74±0.08
NMI
Dermatology70.17±0.0466.62±0.0354.24±0.0071.90±0.0587.41±0.0082.73±0.0079.70±4.3476.27±2.1988.73±3.01
ForestTypes53.56±0.0246.56±0.0116.99±0.0048.20±0.0452.04±0.0056.53±0.0054.80±5.3555.33±0.1559.63±0.09
BDGP25.85±0.029.85±0.0110.29±0.0010.02±0.0114.90±0.0026.03±0.0021.90±1.9535.55±4.5330.92±0.65
Reuters5.67±0.015.45±0.010.64±0.000.13±0.001.79±0.007.23±0.003.36±0.485.05±0.087.07±0.01
Hdigit89.33±0.0356.16±0.0228.57±0.0055.80±0.0319.04±0.0088.67±0.0070.00±2.6076.96±3.0978.09±1.59
VGGFace11.92±0.0014.13±0.005.81±0.0012.33±0.0011.00±0.0010.79±0.0012.32±0.3312.97±0.1514.33±0.11
CIFAR10015.49±0.0014.83±0.008.49±0.0013.87±0.0014.83±0.0016.82±0.0013.93±0.1511.45±0.1616.94±0.14
Purity
Dermatology80.97±0.0480.35±0.0462.85±0.0083.55±0.0784.64±0.0087.15±0.0087.86±4.5185.38±2.0494.03±2.35
ForestTypes77.99±0.0272.82±0.0145.51±0.0072.73±0.0477.06±0.0078.93±0.0077.61±4.7479.29±0.2382.53±0.09
BDGP49.64±0.0237.80±0.0133.32±0.0034.82±0.0140.16±0.0049.44±0.0047.36±3.0859.73±3.6760.65±0.62
Reuters28.22±0.0025.12±0.0119.78±0.0017.89±0.0023.31±0.0028.07±0.0025.45±0.5927.50±0.1730.84±0.02
Hdigit88.93±0.0666.28±0.0332.85±0.0064.34±0.0531.94±0.0088.43±0.0079.82±5.6686.59±4.0188.22±2.25
VGGFace7.02±0.007.73±0.003.50±0.006.33±0.005.77±0.006.11±0.007.17±0.337.47±0.179.43±0.08
CIFAR10010.41±0.009.09±0.004.49±0.007.84±0.008.69±0.0011.11±0.008.28±0.137.89±0.1412.29±0.13
", + "image_path": "60cbc4a9c4db7d2ca490269d25aff3b7884c8d4323f2b89359a292180cd8d7f7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 386, + 288, + 436 + ], + "blocks": [ + { + "bbox": [ + 63, + 366, + 271, + 378 + ], + "lines": [ + { + "bbox": [ + 63, + 366, + 271, + 378 + ], + "spans": [ + { + "bbox": [ + 63, + 366, + 271, + 378 + ], + "type": "text", + "content": "Table 3. Ablation study on anchor enhancement module." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 386, + 288, + 436 + ], + "lines": [ + { + "bbox": [ + 50, + 386, + 288, + 436 + ], + "spans": [ + { + "bbox": [ + 50, + 386, + 288, + 436 + ], + "type": "table", + "html": "
MethodsDermatologyForestTypesBDGP
Baseline90.87±3.6070.76±0.0554.38±0.10
+Align91.82±3.7875.51±0.1757.97±0.09
+Align+Enh.93.20±4.5082.53±0.0960.65±0.62
", + "image_path": "0dfa20e04212551ecbf6ce8a876bb060e93e9c353502c2bc9d9cec32be6ebc6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 528 + ], + "type": "text", + "content": "of the reconstructed anchors and " + }, + { + "bbox": [ + 46, + 456, + 287, + 528 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 456, + 287, + 528 + ], + "type": "text", + "content": " of random anchors are included during the enhancement process. From the results, as the proportion of randomly generated anchors increases, the clustering performance on both datasets continuously deteriorates, thus validating the effectiveness of our enhanced anchors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "text", + "content": "To verify the two properties of enhanced anchors, we calculate the values on both sides of the equations in the properties on two datasets. As shown in Fig. 3(b), the difference between the anchor graphs in the two views before and after anchor enhancement decreases " + }, + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\tilde{\\mathbf{Z}}_2\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}_2\\right\\|_{\\mathbf{F}}\\right)" + }, + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "text", + "content": ", validating Property 1. Additionally, the difference between the anchor graph and the optimal consistent anchor graph reduces before and after anchor enhancement " + }, + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "inline_equation", + "content": "\\left(\\left\\| \\tilde{\\mathbf{Z}}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}} < \\left\\| \\mathbf{Z}_1 - \\mathbf{Z}\\right\\|_{\\mathbf{F}}\\right)" + }, + { + "bbox": [ + 46, + 529, + 288, + 651 + ], + "type": "text", + "content": ", validating Property 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 658, + 280, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 280, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 280, + 672 + ], + "type": "text", + "content": "4.4. Effectiveness of Anchor Graph Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "To further validate the effectiveness of the revised anchor graph construction module, we contrast our proposed approach with two versions without the view weight and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 368, + 426, + 456 + ], + "blocks": [ + { + "bbox": [ + 307, + 368, + 426, + 456 + ], + "lines": [ + { + "bbox": [ + 307, + 368, + 426, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 368, + 426, + 456 + ], + "type": "image", + "image_path": "f5f79e0224c7f545cc3586d8c56a5a1308f5e5294ab93cc71c4c9eea90663ab9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 460, + 423, + 469 + ], + "lines": [ + { + "bbox": [ + 312, + 460, + 423, + 469 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 423, + 469 + ], + "type": "text", + "content": "(a) Validation of enhancement strategy" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 369, + 545, + 457 + ], + "blocks": [ + { + "bbox": [ + 426, + 369, + 545, + 457 + ], + "lines": [ + { + "bbox": [ + 426, + 369, + 545, + 457 + ], + "spans": [ + { + "bbox": [ + 426, + 369, + 545, + 457 + ], + "type": "image", + "image_path": "78325abf0eb9f3e1177bbc8ac387b91002c10a59a6c81a0f8e630be974867eab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 460, + 533, + 469 + ], + "lines": [ + { + "bbox": [ + 440, + 460, + 533, + 469 + ], + "spans": [ + { + "bbox": [ + 440, + 460, + 533, + 469 + ], + "type": "text", + "content": "(b) Validation of Property 1 & 2" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 478, + 547, + 545 + ], + "lines": [ + { + "bbox": [ + 304, + 478, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 547, + 545 + ], + "type": "text", + "content": "Figure 3. Validation of the proposed AEVC's effectiveness: (a) As the proportion of randomly generated anchors increases, the corresponding clustering performance decreases monotonically. (b) After anchor enhancement, the differences in anchor graphs between different views are reduced, and the enhanced anchor graph is closer to the optimal consensus anchor graph." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 585, + 552, + 630 + ], + "blocks": [ + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "lines": [ + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 305, + 554, + 545, + 578 + ], + "type": "text", + "content": "Table 4. Ablation study on revised anchor graph construction module. Other results are provided in supplementary materials." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 585, + 552, + 630 + ], + "lines": [ + { + "bbox": [ + 309, + 585, + 552, + 630 + ], + "spans": [ + { + "bbox": [ + 309, + 585, + 552, + 630 + ], + "type": "table", + "html": "
MethodsDermatologyForestTypesBDGP
w/o. view weight91.60±6.2681.45±0.0060.50±0.46
w/o. regularized term89.98±3.4079.75±0.1060.64±0.62
Proposed93.20±4.5082.53±0.0960.65±0.62
", + "image_path": "8b99aa18815b8bc2050d73a6307198e24195b051b1d6b94096ccf29d4f118387.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "regularized term respectively, as illustrated in Table 4. After incorporating weights learned from the views during fusion, there is a noticeable improvement in the clustering performance of the method. This highlights the advantage of assessing the importance of views in multi-view clustering" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26157" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 90, + 523, + 220 + ], + "blocks": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "lines": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "spans": [ + { + "bbox": [ + 116, + 71, + 475, + 82 + ], + "type": "text", + "content": "Table 5. Performance comparison between various anchor-based MVC w./w.o. our AEVC strategy." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 90, + 523, + 220 + ], + "lines": [ + { + "bbox": [ + 75, + 90, + 523, + 220 + ], + "spans": [ + { + "bbox": [ + 75, + 90, + 523, + 220 + ], + "type": "table", + "html": "
MethodsDermatologyForestTypesBDGPReutersHdigitVGGFaceCIFAR100
Fixed Anchors Methods
FMCNOF62.01±0.0045.51±0.0031.08±0.0019.25±0.0032.84±0.003.47±0.004.40±0.00
FMCNOF-AE73.18±0.0069.60±0.0037.64±0.0022.79±0.0034.31±0.003.21±0.005.05±0.00
LMVSC79.02±0.0777.68±0.0449.52±0.0227.19±0.0186.87±0.086.09±0.008.92±0.00
LMVSC-AE88.23±7.8080.90±5.4952.13±0.3428.84±0.6785.74±5.726.83±0.168.81±0.16
Updated Anchors Methods
UDBGL84.08±0.0077.06±0.0039.28±0.0022.32±0.0031.39±0.005.39±0.007.83±0.00
UDBGL-AE81.01±0.0077.82±0.0046.92±0.0028.53±0.0044.60±0.005.65±0.008.57±0.00
FMVACC84.39±4.3479.29±0.2359.51±4.0027.22±0.2486.55±4.086.68±0.167.04±0.12
FMVACC-AE88.41±3.1381.08±0.2359.36±3.0333.12±2.4988.42±2.937.13±0.177.28±0.15
", + "image_path": "0dc928d91d001417a42e14d27619922a2b7c318f088e2d662a5b0dd16e581f6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "content": "through the proposed strategy. After removing the regularization term, there was a decline in performance across different datasets, which further confirms the effectiveness of constructing anchor graphs guided by " + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 46, + 239, + 287, + 286 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 296, + 183, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 296, + 183, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 183, + 308 + ], + "type": "text", + "content": "4.5. Transferability Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "content": "To validate the transferability of AEVC, Table 5 presents a comparison of clustering accuracy between various types of anchor-based multi-view clustering methods with and without our anchor enhancement module. In this experiment, we selected two methods from each of the two categories employing different anchor strategies. Specifically, FMCNOF and LMVSC represent methods with fixed anchors, while UDBGL and FMVACC belong to methods with iteratively optimized anchors. As observed in the table, the clustering performance of the four anchor-based multi-view clustering methods experiences significant improvement across the majority of datasets after incorporating our proposed module. For instance, on the Reuters dataset, the clustering performance is notably enhanced by " + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "inline_equation", + "content": "27.8\\%" + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "inline_equation", + "content": "21.7\\%" + }, + { + "bbox": [ + 46, + 315, + 289, + 531 + ], + "type": "text", + "content": " for FMCNOF, LMVSC, UDBGL, and FMVACC, respectively. This substantial improvement provides strong evidence for the effectiveness of AEVC when transferred to other anchor-based methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 162, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 162, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 162, + 552 + ], + "type": "text", + "content": "4.6. Sensitivity Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Our proposed method involves three hyperparameters: the number of anchors " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", the anchor enhancement rate " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", and the hyperparameter for balancing the regularization term " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ". Due to space constraints, we primarily analyze the sensitivity of the latter two parameters here, while the analysis regarding " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " will be presented in the supplementary materials. To demonstrate the impact of " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", we fix " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and conduct grid searches within specified ranges. Results presented in Fig. 4 indicate that our method is significantly influenced by the anchor enhancement rate, as the extreme rate can disrupt the balance between newly generated and original anchors. Specifically, on the Dermatology dataset, setting " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " to 100 yields the highest clustering accuracy." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 245, + 419, + 327 + ], + "blocks": [ + { + "bbox": [ + 318, + 245, + 419, + 327 + ], + "lines": [ + { + "bbox": [ + 318, + 245, + 419, + 327 + ], + "spans": [ + { + "bbox": [ + 318, + 245, + 419, + 327 + ], + "type": "image", + "image_path": "be366ee45049e67f42935c740ed0e7663931453db5efe43949f6d36133f3a076.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 335, + 394, + 344 + ], + "lines": [ + { + "bbox": [ + 345, + 335, + 394, + 344 + ], + "spans": [ + { + "bbox": [ + 345, + 335, + 394, + 344 + ], + "type": "text", + "content": "(a) Dermatology" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 245, + 541, + 327 + ], + "blocks": [ + { + "bbox": [ + 440, + 245, + 541, + 327 + ], + "lines": [ + { + "bbox": [ + 440, + 245, + 541, + 327 + ], + "spans": [ + { + "bbox": [ + 440, + 245, + 541, + 327 + ], + "type": "image", + "image_path": "8aa03ee39764fb58b9cdb1186f1d15d6918641c6ec84e8550c89b2fe8e1b7f0a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 335, + 517, + 344 + ], + "lines": [ + { + "bbox": [ + 470, + 335, + 517, + 344 + ], + "spans": [ + { + "bbox": [ + 470, + 335, + 517, + 344 + ], + "type": "text", + "content": "(b) ForestTypes" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": "Figure 4. Sensitivity analysis of " + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 353, + 545, + 375 + ], + "type": "text", + "content": " on two benchmark datasets. Others are displayed in supplementary materials." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 408, + 378, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 408, + 378, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 378, + 420 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 602 + ], + "type": "text", + "content": "In this paper, we propose an Anchor Enhancement strategy with View Correlation for multi-view clustering termed AEVC. Different from existing anchor-based MVC methods, AEVC introduces a novel plug-and-play anchor enhancement strategy consisting of two modules: the view-graph learning module and the anchor enhancement module. It can effectively enhance the quality of anchors, which is proven both experimentally and theoretically. Besides, we propose a revised anchor graph construction module based on the anchor enhancement strategy, which further improves the clustering performances. Experimental results show the promising capacity of AEVC, especially when integrating with other anchor-based methods. In the future, we will explore adaptive learning of enhancement rate." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 632, + 410, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 632, + 410, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 632, + 410, + 646 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 658, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 703 + ], + "type": "text", + "content": "This work was supported by the National Science Fund for Distinguished Young Scholars of China (No. 62325604), and the National Natural Science Foundation of China (No. 62276271 and No. 62306324)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26158" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE transactions on pattern analysis and machine intelligence, 41(2):423-443, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "text", + "content": "[2] Jie Chen, Hua Mao, Dezhong Peng, Changqing Zhang, and Xi Peng. Multiview clustering by consensus spectral rotation fusion. IEEE Transactions on Image Processing, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 223 + ], + "type": "text", + "content": "[3] Yongyong Chen, Xiaojia Zhao, Zheng Zhang, Youfa Liu, Jingyong Su, and Yicong Zhou. Tensor learning meets dynamic anchor learning: From complete to incomplete multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 280 + ], + "type": "text", + "content": "[4] Zhibin Dong, Jiaqi Jin, Yuyang Xiao, Siwei Wang, Xinzhong Zhu, Xinwang Liu, and En Zhu. Iterative deep structural graph contrast clustering for multiview raw data. IEEE Transactions on Neural Networks and Learning Systems, pages 1-13, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 335 + ], + "type": "text", + "content": "[5] Zhibin Dong, Siwei Wang, Jiaqi Jin, Xinwang Liu, and En Zhu. Cross-view topology based consistent and complementary information for deep multi-view clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 19440-19451, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 379 + ], + "type": "text", + "content": "[6] Jürgen Dorn, Mario Girsch, Günther Skele, and Wolfgang Slany. Comparison of iterative improvement techniques for schedule optimization. European Journal of Operational Research, 94(2):349-361, 1996. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "text", + "content": "[7] Si-Guo Fang, Dong Huang, Xiao-Sha Cai, Chang-Dong Wang, Chaobo He, and Yong Tang. Efficient multi-view clustering via unified and discrete bipartite graph learning. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 288, + 479 + ], + "type": "text", + "content": "[8] Uno Fang, Man Li, Jianxin Li, Longxiang Gao, Tao Jia, and Yanchun Zhang. A comprehensive survey on multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 523 + ], + "type": "text", + "content": "[9] Wenjue He, Zheng Zhang, Yongyong Chen, and Jie Wen. Structured anchor-inferred graph learning for universal incomplete multi-view clustering. World Wide Web, 26(1): 375-399, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Fast multi-view clustering via ensembles: Towards scalability, superiority, and simplicity. IEEE Transactions on Knowledge and Data Engineering, 2023. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 624 + ], + "type": "text", + "content": "[11] Yanyong Huang, Zongxin Shen, Yuxin Cai, Xiwen Yi, Dongjie Wang, Fengmao Lv, and Tianrui Li. Imufs: Complementary and consensus learning-based incomplete multiview unsupervised feature selection. IEEE Transactions on Knowledge and Data Engineering, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 678 + ], + "type": "text", + "content": "[12] Jintian Ji and Songhe Feng. Anchor structure regularization induced multi-view subspace clustering via enhanced tensor rank minimization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19343-19352, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[13] Jiaqi Jin, Siwei Wang, Zhibin Dong, Xinwang Liu, and En Zhu. Deep incomplete multi-view clustering with cross-view partial sample and prototype alignment. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11600-11609, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 150 + ], + "type": "text", + "content": "[14] Zhao Kang, Wangtao Zhou, Zhitong Zhao, Junming Shao, Meng Han, and Zenglin Xu. Large-scale multi-view subspace clustering in linear time. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4412-4419, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[15] Borhan Kazimipour, Xiaodong Li, and A Kai Qin. Initialization methods for large scale global optimization. In 2013 IEEE congress on evolutionary computation, pages 2750-2757. IEEE, 2013. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 547, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 252 + ], + "type": "text", + "content": "[16] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334-8343, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 253, + 547, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 547, + 286 + ], + "type": "text", + "content": "[17] Lusi Li and Haibo He. Bipartite graph based multi-view clustering. IEEE Transactions on Knowledge and Data Engineering, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 287, + 547, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 547, + 330 + ], + "type": "text", + "content": "[18] Ruihuang Li, Changqing Zhang, Qinghua Hu, Pengfei Zhu, and Zheng Wang. Flexible multi-view representation learning for subspace clustering. In *IJCAI*, pages 2916–2922, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 331, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 376 + ], + "type": "text", + "content": "[19] Xuelong Li, Han Zhang, Rong Wang, and Feiping Nie. Multi-view clustering: A scalable and parameter-free bipartite graph fusion method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 377, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 547, + 431 + ], + "type": "text", + "content": "[20] Ke Liang, Lingyuan Meng, Meng Liu, Yue Liu, Wenxuan Tu, Siwei Wang, Sihang Zhou, Xinwang Liu, and Fuchun Sun. Reasoning over different types of knowledge graphs: Static, temporal and multi-modal. arXiv preprint arXiv:2212.05767, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 547, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 488 + ], + "type": "text", + "content": "[21] Ke Liang, Yue Liu, Sihang Zhou, Wenxuan Tu, Yi Wen, Xihong Yang, Xiangjun Dong, and Xinwang Liu. Knowledge graph contrastive learning based on relation-symmetrical structure. IEEE Transactions on Knowledge and Data Engineering, pages 1-12, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 533 + ], + "type": "text", + "content": "[22] Ke Liang, Lingyuan Meng, Sihang Zhou, Siwei Wang, Wenxuan Tu, Yue Liu, Meng Liu, and Xinwang Liu. Message intercommunication for inductive relation reasoning. arXiv preprint arXiv:2305.14074, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 547, + 589 + ], + "type": "text", + "content": "[23] Yijie Lin, Yuanbiao Gou, Zitao Liu, Boyun Li, Jiancheng Lv, and Xi Peng. Completer: Incomplete multi-view clustering via contrastive prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11174-11183, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 634 + ], + "type": "text", + "content": "[24] Jialu Liu, Chi Wang, Jing Gao, and Jiawei Han. Multiview clustering via joint nonnegative matrix factorization. In Proceedings of the 2013 SIAM International Conference on Data Mining, pages 252-260. SIAM, 2013. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 678 + ], + "type": "text", + "content": "[25] Suyuan Liu, Xinwang Liu, Siwei Wang, Xin Niu, and En Zhu. Fast incomplete multi-view clustering with view-independent anchors. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[26] Suyuan Liu, Siwei Wang, Pei Zhang, Kai Xu, Xinwang Liu, Changwang Zhang, and Feng Gao. Efficient one-pass multiview subspace clustering with consensus anchors. In Pro-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26159" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "ceedings of the AAAI Conference on Artificial Intelligence, pages 7576-7584, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 287, + 140 + ], + "type": "text", + "content": "[27] Suyuan Liu, Qing Liao, Siwei Wang, Xinwang Liu, and En Zhu. Robust and consistent anchor graph learning for multiview clustering. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 185 + ], + "type": "text", + "content": "[28] Xun Lu and Songhe Feng. Structure diversity-induced anchor graph fusion for multi-view clustering. ACM Transactions on Knowledge Discovery from Data, 17(2):1-18, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 189, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 189, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 189, + 287, + 233 + ], + "type": "text", + "content": "[29] Xi Peng, Zhenyu Huang, Jiancheng Lv, Hongyuan Zhu, and Joey Tianyi Zhou. Comic: Multi-view clustering without parameter selection. In International conference on machine learning, pages 5092-5101. PMLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 235, + 287, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 287, + 279 + ], + "type": "text", + "content": "[30] Qianyao Qiang, Bin Zhang, Fei Wang, and Feiping Nie. Fast multi-view discrete clustering with anchor graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9360-9367, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 281, + 287, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 281, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 281, + 287, + 324 + ], + "type": "text", + "content": "[31] Shaojun Shi, Feiping Nie, Rong Wang, and Xuelong Li. Fast multi-view clustering via prototype graph. IEEE Transactions on Knowledge and Data Engineering, 35(1):443-455, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[32] Xiaochuang Shu, Xiangdong Zhang, Quanxue Gao, Ming Yang, Rong Wang, and Xinbo Gao. Self-weighted anchor graph learning for multi-view clustering. IEEE Transactions on Multimedia, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 373, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 429 + ], + "type": "text", + "content": "[33] Mengjing Sun, Pei Zhang, Siwei Wang, Sihang Zhou, Wenxuan Tu, Xinwang Liu, En Zhu, and Changjian Wang. Scalable multi-view subspace clustering with unified anchors. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3528-3536, 2021. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "type": "text", + "content": "[34] Huayi Tang and Yong Liu. Deep safe multi-view clustering: Reducing the risk of clustering performance degradation caused by view increase. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 202-211, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 542 + ], + "type": "text", + "content": "[35] Xinhang Wan, Jiyuan Liu, Weixuan Liang, Xinwang Liu, Yi Wen, and En Zhu. Continual multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, page 3676-3684, New York, NY, USA, 2022. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 599 + ], + "type": "text", + "content": "[36] Xinhang Wan, Xinwang Liu, Jiyuan Liu, Siwei Wang, Yi Wen, Weixuan Liang, En Zhu, Zhe Liu, and Lu Zhou. Autoweighted multi-view clustering for large-scale data. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10078-10086, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[37] Jun Wang, Chang Tang, Zhiguo Wan, Wei Zhang, Kun Sun, and Albert Y Zomaya. Efficient and effective one-step multiview clustering. IEEE Transactions on Neural Networks and Learning Systems, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[38] Siwei Wang, Xinwang Liu, Li Liu, Wenxuan Tu, Xinzhong Zhu, Jiyuan Liu, Sihang Zhou, and En Zhu. Highly-efficient incomplete large-scale multi-view clustering with consensus bipartite graph. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9776-9785, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[39] Siwei Wang, Xinwang Liu, Suyuan Liu, Jiaqi Jin, Wenxuan Tu, Xinzhong Zhu, and En Zhu. Align then fusion: Generalized large-scale multi-view clustering with anchor matching correspondences. Advances in Neural Information Processing Systems, 35:5882-5895, 2022. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[40] Siwei Wang, Xinwang Liu, Xinzhong Zhu, Pei Zhang, Yi Zhang, Feng Gao, and En Zhu. Fast parameter-free multiview subspace clustering with consensus anchor guidance. IEEE Trans. Image Process., 31:556-568, 2022. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[41] Yang Wang, Lin Wu, Xuemin Lin, and Junbin Gao. Multiview spectral clustering via structured low-rank matrix factorization. IEEE transactions on neural networks and learning systems, 29(10):4833-4843, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[42] Yi Wen, Siwei Wang, Ke Liang, Weixuan Liang, Xinhang Wan, Xinwang Liu, Suyuan Liu, Jiyuan Liu, and En Zhu. Scalable incomplete multi-view clustering with structure alignment. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3031-3040, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[43] Yi Wen, Siwei Wang, Qing Liao, Weixuan Liang, Ke Liang, Xinhang Wan, and Xinwang Liu. Unpaired multi-view graph clustering with cross-view structure matching. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 364 + ], + "type": "text", + "content": "[44] Wei Xia, Quanxue Gao, Qianqian Wang, Xinbo Gao, Chris Ding, and Dacheng Tao. Tensorized bipartite graph learning for multi-view clustering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 420 + ], + "type": "text", + "content": "[45] Weiqing Yan, Jindong Xu, Jinglei Liu, Guanghui Yue, and Chang Tang. Bipartite graph-based discriminative feature learning for multi-view clustering. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3403-3411, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[46] Ben Yang, Xuetao Zhang, Feiping Nie, Fei Wang, Weizhong Yu, and Rong Wang. Fast multi-view clustering via nonnegative and orthogonal factorization. IEEE Transactions on Image Processing, 30:2575-2586, 2020. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 520 + ], + "type": "text", + "content": "[47] Ben Yang, Xuetao Zhang, Zhiping Lin, Feiping Nie, Badong Chen, and Fei Wang. Efficient and robust multiview clustering with anchor graph regularization. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):6200-6213, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 577 + ], + "type": "text", + "content": "[48] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[49] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[50] Xihong Yang, Jin Jiaqi, Siwei Wang, Ke Liang, Yue Liu, Yi Wen, Suyuan Liu, Sihang Zhou, Xinwang Liu, and En Zhu. Dealmvc: Dual contrastive calibration for multi-view clustering. In Proceedings of the 31st ACM International Conference on Multimedia, pages 337-346, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[51] Jiali You, Zhenwen Ren, Xiaojian You, Haoran Li, and Yuancheng Yao. Priori anchor labels supervised scalable multi-view bipartite graph clustering. In Proceedings of the" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26160" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 373 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "AAAI Conference on Artificial Intelligence, pages 10972-10979, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[52] Shengju Yu, Siwei Wang, Yi Wen, Ziming Wang, Zhigang Luo, En Zhu, and Xinwang Liu. How to construct corresponding anchors for incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 207 + ], + "type": "text", + "content": "[53] Chao Zhang, Huaxiong Li, Wei Lv, Zizheng Huang, Yang Gao, and Chunlin Chen. Enhanced tensor low-rank and sparse representation recovery for incomplete multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11174-11182, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[54] Nan Zhang, Xiaoqin Zhang, and Shiliang Sun. Efficient multiview representation learning with correntropy and anchor graph. IEEE Transactions on Knowledge & Data Engineering, (01):1-14, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 317 + ], + "type": "text", + "content": "[55] Pei Zhang, Siwei Wang, Liang Li, Changwang Zhang, Xinwang Liu, En Zhu, Zhe Liu, Lu Zhou, and Lei Luo. Let the data choose: flexible and diverse anchor graph fusion for scalable multi-view clustering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11262-11269, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[56] Xiaojia Zhao, Qiangqiang Shen, Yongyong Chen, Yongsheng Liang, Junxin Chen, and Yicong Zhou. Self-completed bipartite graph learning for fast incomplete multiview clustering. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26161" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_content_list.json b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b887299da77ff34b57976488427fc6f8ecce644f --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation", + "text_level": 1, + "bbox": [ + 104, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyun Wang \nBeihang University", + "bbox": [ + 184, + 181, + 351, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "19231136@buaa.edu.cn", + "bbox": [ + 184, + 219, + 361, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guoliang Kang * \nBeihang University, Zhongguancun Laboratory", + "bbox": [ + 413, + 181, + 787, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kgl.prml@gmail.com", + "bbox": [ + 516, + 219, + 679, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works utilize CLIP to perform the challenging unsupervised semantic segmentation task where only images without annotations are available. However, we observe that when adopting CLIP to such a pixel-level understanding task, unexpected bias occurs. Previous works don't explicitly model such bias, which largely constrains the segmentation performance. In this paper, we propose to explicitly model and rectify the bias existing in CLIP to facilitate the unsupervised semantic segmentation. Specifically, we design a learnable \"Reference\" prompt to encode class-preference bias and project the positional embedding of vision transformer to represent space-preference bias. Via a simple element-wise subtraction, we rectify the logits of CLIP classifier. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then a contrastive loss between masked visual feature and the text features of different classes is imposed to facilitate the effective bias modeling. To further improve the segmentation, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture via minimizing our designed mask-guided, feature-guided and text-guided loss terms. Extensive experiments on standard benchmarks demonstrate that our method performs favorably against previous state-of-the-arts. The implementation is available at https://github.com/dogehhh/ReLUCP.", + "bbox": [ + 76, + 301, + 473, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 710, + 209, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation aims to attach a semantic label to each pixel of an image. Since the rising of deep learning [27, 28, 45, 54], semantic segmentation has been widely adopted in real-world applications, e.g., autonomous driving, medical image segmentation, etc. Conventional approaches [6, 9, 10, 37, 38, 53, 60] for semantic segmentation have achieved remarkable performance. However, the superior performance of those methods relies heavily on large amounts of fully annotated masks. Collecting such high", + "bbox": [ + 75, + 737, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality pixel-level annotations can be both time-consuming and expensive, e.g., some annotations for specialized tasks require massive expert knowledge, some are even inaccessible due to privacy reasons, etc. Therefore, it is necessary to explore unsupervised semantic segmentation where only images without annotations are available.", + "bbox": [ + 496, + 268, + 893, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised semantic segmentation (USS) has been studied for years. Many non-language-guided USS methods have been proposed, e.g., clustering-based methods [11, 25, 29, 34, 42], contrastive-learning-based methods [21, 51], boundary-based methods [23], etc. Despite promising progress achieved, there still exhibits a large performance gap between USS and the supervised segmentation methods. Besides, these methods typically obtain class-agnostic masks and have to depend on additional processing (e.g., Hungarian matching) to assign semantic labels to the masks, rendering them less practical in real scenarios.", + "bbox": [ + 496, + 364, + 893, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, large-scale visual-language pre-trained models, e.g., CLIP [45], demonstrate superior zero-shot classification performance by comparing the degree of alignment between image feature and text features of different categories. A few CLIP-based USS approaches [22, 46, 49, 62] emerge and show remarkable performance improvement compared with the non-language-guided USS methods. These models require no access to any types of annotations, and directly assign a label to each pixel, benefiting from the aligned vision and text embedding space of CLIP. However, good alignment between image-level visual feature and textural feature doesn't necessarily mean good alignment between pixel-level visual feature and textural feature. Thus, for CLIP, unexpected bias may inevitably appear. Previous works didn't explicitly model such bias, which may largely constrain their segmentation performance.", + "bbox": [ + 496, + 534, + 895, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We observe two kinds of bias existing in CLIP. From one aspect, as shown in Fig. 1(a), CLIP exhibits space-preference bias. CLIP performs apparently better for segmenting central objects than the ones distributed near the image boundary. It can be reflected by the fact that mIoU decreases as the distance between the centroids of object and the image increases. From the other aspect, as shown in Fig. 1(b), there exists class-preference bias between similar", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4102", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2575f8c34e09cc48afa5a280e80590666f01f27d3f00618934a9264dac09ada8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 86, + 89, + 433, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7eef6c1530938f32fb3538d7ea98bbfac462f22e9913064392a5ad1dcc130a33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 90, + 534, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c53632a59c17113761d19526488bf5af2b07fa34e63fe3c259912d58e6918ab7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 90, + 633, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6bde63d342e6c7cf94b87467dbbd8374044b1eaacbbb4601078c8f00e39d141a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 727, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/351504f4a6cc247c3417f7c3100894262dfd94b013516e9f264df8f30cc2cb39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 90, + 828, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f7b4dce1ae444e3cf1495e5c3bbba378cd07204a38b923cd6b115b3f243963bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 441, + 160, + 534, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa747899bbf73067470b5d0090df55432d29e6021d1647db492039cdd43cee5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 160, + 633, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/93451363de034c4fb6fb3c28886744a2de50e6592136a88ead8875e46dfe7a3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 160, + 728, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3b4e2e2eb6f5c95b077e839363f3360f83be3f79bc4a9f0e6dc8863c2b59a8c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 160, + 828, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7cb6556268707b46e596e279189ea6b1bf6f5bef8224dfc978eba5bbe6d8e409.jpg", + "image_caption": [ + "Figure 1. (a) Space-preference bias. (Left): The relationship between distance $(x$ -axis) and mIoU $(y$ -axis) is drawn on PASCAL VOC [18]. The distance means the spatial distance between the centroids of the object and the image and mIoU is computed based on predictions and ground truth. The curve shows that CLIP [45] (green) is apparently better at segmentation for central objects than boundary ones, but our method (blue) effectively mitigates this bias. More details about how we draw this figure has been shown in our supplementary material. (Right): Visualizations also show our improvement on space-preference bias qualitatively. (b) Class-preference bias. (Left): We randomly select 6 classes from PASCAL Context [40] and draw the confusion matrix of CLIP and our model. It shows that beside the ground truth, CLIP also prefers to assign an incorrect but relevant label to a pixel in quite a few cases, while our results show apparent improvement. (Right): The visualizations are consistent with what we observed in confusion matrix. For example, for a \"cow\", CLIP tends to classify it as \"dog\" incorrectly." + ], + "image_footnote": [], + "bbox": [ + 86, + 224, + 888, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "categories in CLIP. For example, according to visualization (right), when ground truth is \"cow\", CLIP tends to incorrectly classify it as \"dog\". We also show such trend between randomly selected classes by confusion matrix (left). Elements on diagonal line represent right classification, while others are false. We observe a wide range of class-preference bias introduced by CLIP.", + "bbox": [ + 75, + 484, + 468, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose to explicitly model and rectify the bias of CLIP to facilitate the USS. Specifically, we design two kinds of text inputs for each class, which are named as \"Reference\" and \"Query\" respectively. The Query is manually designed while the Reference contains learnable prompts. We adopt the text features of Query and Reference as classifiers to generate the Query and Reference logits respectively for each pixel of image. The Query logits represent the segmentation ability of original CLIP, while the Reference logits are expected to reflect the bias of CLIP preferring a specific class. Additionally, we project the positional embedding of CLIP's vision transformer to generate positional logits for each image. We expect the positional logits to represent space-preference bias of CLIP. Then, we remove the class-preference and the space-preference bias from original CLIP via a logit-subtraction mechanism, i.e., subtracting the Reference logits and the positional logits from the Query logits. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then the contrastive loss between masked visual", + "bbox": [ + 75, + 598, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "feature and the text features of different categories is imposed to facilitate the effective bias modeling and rectification. To further improve the segmentation performance, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture, via mask-guided distillation, feature-guided distillation and text-guided learning.", + "bbox": [ + 496, + 484, + 890, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on standard semantic segmentation benchmarks, including PASCAL VOC [18], PASCAL Context [40] and ADE20K [61]. Experiment results demonstrate that our method performs favourably against previous state-of-the-arts. Notably, on PASCAL VOC, our method outperforms CLIP S4 [22] by $3.4\\%$ . Extensive ablation studies verify the effectiveness of each design in our framework.", + "bbox": [ + 496, + 579, + 890, + 699 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows.", + "bbox": [ + 517, + 702, + 821, + 715 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We observe that when applying CLIP to pixel-level understanding tasks, unexpected bias including space-preference bias and class-preference bias, occurs. Such bias may largely constrain the segmentation performance of CLIP-based segmentation models.", + "- We propose to explicitly model the class-preference and space-preference bias of CLIP via learnable Reference text inputs and projection of positional embedding. Through a simple logit-subtraction mechanism and the contrastive loss built on masked features, we effectively rectify the bias of CLIP.", + "- We conduct extensive experiments on segmentation" + ], + "bbox": [ + 517, + 719, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4103", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "benchmarks under the unsupervised setting. Experiment results show superior performance of our method to previous state-of-the-arts.", + "bbox": [ + 109, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 150, + 218, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained vision-language models. Pre-trained vision-language models (VLMs) [8, 14, 32, 33, 35] have developed rapidly with the help of large-scale image-text pairs available on the Internet. Recently, CLIP [45], ALIGN [26] and Slip [41] have made great progress on learning visual and textual representations jointly by using contrastive learning. With the image-level alignment with text, pre-trained VLMs have strong ability for zero-shot classification task and can be transferred to various downstream tasks, such as object detection [17, 54] and semantic segmentation [62].", + "bbox": [ + 76, + 176, + 468, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unsupervised semantic segmentation. While conventional approaches of semantic segmentation [37, 38, 53, 60] rely on pixel-level annotations and weakly-supervised methods [1, 16, 44, 56] still ask for image-level labels, unsupervised semantic segmentation (USS) methods [11, 20, 42, 50, 51] explores to train a segmentation model without any annotations. Models like [2, 7] adopt generative model [12] to separate foreground with background or generate corresponding masks. SegSort [23], HSG[29] and ACSeg [34] use clustering strategy, while IIC [25] uses mutual information maximization to perform unsupervised learning. MaskContrast [51] introduces contrastive learning into USS. Others like DSM [39] and LNE [13] exploit features extracted from self-supervised model, and combine it with spectral graph theory. However, the methods mentioned above either fail to segment images with multi-category objects or show a large performance gap with the supervised methods. Besides, they can only obtain class-agnostic masks and have to depend on additional strategy, such as Hungarian-matching algorithm [30], to match corresponding category with masks. Recently, pre-trained vision-language models are adopted in USS. MaskCLIP [62] modifies the image encoder of CLIP to generate patch-level features and directly performs segmentation with text features as classifiers. CLIP-py [46] performs contrastive learning between visual features from self-supervised ViT [4] and text features from CLIP. ReCo [49] performs image retrieval with CLIP and extracts class-wise embedding as classifier with co-segmentation. CLIP-S4 [22] learns pixel embeddings with pixel-segment contrastive learning and aligns such embeddings with CLIP in terms of embedding and semantic consistency. These methods can directly assign a category to each pixel, whose setting is named as language-guided unsupervised semantic segmentation. However, directly applying CLIP may bring prior bias, including space-preference bias and class-preference bias. As we know, there is no previous method trying to solve these bias and we manage to explicitly model the bias, and rectify them by", + "bbox": [ + 76, + 327, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "element-wise subtraction.", + "bbox": [ + 498, + 92, + 671, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Language-guided semantic segmentation. Recently, many works are exploring semantic segmentation guided by language under different settings. Zero-shot works [3, 31, 43, 55] split classes into seen and unseen set. During the training period, only masks of seen classes are provided. For inference, models are tested on both seen and unseen classes, but the test data is still in the same distribution with the training data. Open-vocabulary works [5, 19, 36, 47, 57-59] are trained in one scenario with extra annotations including class-agnostic masks or image captions, but are used for predicating segmentation masks of novel classes in other scenarios. From the technical view, our method also falls into the category of language-guided semantic segmentation. However, we consider the unsupervised setting. In this setting, we have access to images without any annotations during training. The training and inference images are sampled from the same distributions and the same set of categories. Such a setting is different to the typical zero-shot or open-vocabulary setting.", + "bbox": [ + 496, + 106, + 890, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 407, + 591, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Background In this work, we aim to rectify the bias of CLIP for unsupervised semantic segmentation. In USS, we only have access to images without any types of annotations to train the segmentation model. For training and inference, the same set of categories are considered and the data distributions are assumed to be the same.", + "bbox": [ + 496, + 431, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview The general framework of our method is illustrated in Fig. 2. We aim to rectify the bias of CLIP including the class-preference bias and the space-preference bias, to facilitate unsupervised semantic segmentation. From a high level, class-preference bias reflects the shift of CLIP predictions towards specific classes, while space-preference bias reflects the shift of CLIP predictions towards specific spatial locations. Both biases will be finally reflected in the predicted logit maps. A reasonable way to rectify the bias is to subtract its logit maps from the normal logit maps predicted via original CLIP.", + "bbox": [ + 496, + 523, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To realize our goal, we firstly forward the image $I \\in \\mathbb{R}^{3 \\times H \\times W}$ through the image encoder of CLIP to obtain the patch-level image features $Z$ . We design two kinds of text inputs for each class. One is the manually designed Query text input $Q$ and the other one is the Reference text input $R$ which consists of a learnable prompt and the name of class. For each class, passing $Q$ and $R$ through the text encoder of CLIP, we obtain two text embeddings $W_{q}$ and $W_{r}$ , which serve as the weights of query segmentation head and reference segmentation head respectively. Taking the same visual feature $Z$ as input, the query and reference segmentation heads output a query logit map $M_{q}$ (Sec. 3.1) and Reference logit map $M_{r}$ (Sec. 3.2) respectively. Meanwhile, positional embedding $p$ is sent into a", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4104", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learnable convolutional network to generate positional logit map $M_{p}$ (Sec. 3.3). Then the bias logit map $M_{b}$ can be obtained by adding Reference logit map $M_{r}$ and positional logit map $M_{p}$ . We then subtract $M_{b}$ from $M_{q}$ to generate the rectified logits $M$ . Based on the rectified logits, we generate masks for different classes. Then a contrastive loss is imposed between the masked visual features and the text features of different categories to encourage the bias modeling and rectification (Sec. 3.4). In order to enhance the segmentation performance, we distill the knowledge of rectified CLIP to the advanced segmentation architecture with specifically designed mask-guided, feature-guided and text-guided loss terms (Sec. 3.5). In both the rectification and distillation stages, we keep CLIP frozen.", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Baseline: Directly Segment with CLIP", + "text_level": 1, + "bbox": [ + 76, + 311, + 410, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following MaskCLIP [62], we adapt pre-trained CLIP [45] (ViT-B/16) to the semantic segmentation task. We remove the query and key embedding layers of last attention but reformulate the value embedding layer and the last linear projection layer into two respective $1 \\times 1$ convolutional layers. Therefore, the image encoder can not only generate local features for dense predictions, but also keep the visual-language association in CLIP by freezing its pre-trained weights. We forward image $I$ through the image encoder and obtain patch-level features $Z \\in \\mathbb{R}^{n \\times D}$ ( $n$ is the number of patches and $D$ is the dimension of features in CLIP).", + "bbox": [ + 75, + 333, + 468, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each text in Query $Q = \\{Q_{1}, Q_{2}, \\dots, Q_{C}\\}$ ( $C$ is the number of classes) is an ensemble of several manually designed prompts, e.g., \"a good/large/bad photo of a [CLS]\" where [CLS] denotes the a specific class name. Passing $Q$ through the text encoder, we obtain its text embeddings $W_{q}$ . We treat text embeddings $W_{q}$ as the weight of segmentation head to perform $1 \\times 1$ convolution. By sending features $Z$ to the segmentation head, we get a Query logit map $M_{q} \\in \\mathbb{R}^{n \\times C}$ . Then the segmentation mask can be predicted by arg max operation on $M_{q}$ .", + "bbox": [ + 75, + 500, + 472, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learn Class-Preference Bias", + "text_level": 1, + "bbox": [ + 76, + 659, + 333, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to explicitly model the class-preference bias brought by pre-trained CLIP for specific datasets, we design a Reference $R_{i}$ , $i = \\{1,2,\\dots ,C\\}$ , as additional text input for each class. Inspired by CoOp [63], $R_{i}$ consists of a learnable prompt which is shared across all the classes for efficiency, and a class name $[CLS]$ , which can be formed as", + "bbox": [ + 75, + 683, + 470, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {i} = \\left[ v _ {1} \\right] \\left[ v _ {2} \\right] \\dots \\left[ v _ {l} \\right] \\dots \\left[ v _ {L} \\right] [ C L S ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 787, + 468, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each $[v_l](l \\in \\{1, \\dots, L\\})$ is a vector with dimension $D$ and serves as a word embedding. The $L$ is a constant representing the number of word embeddings to learn. Totally, there are 77 word embeddings for a text of CLIP. As two word embeddings are used for class name and two for indicating start and end of a text, we set $L$ to 73.", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Passing the Reference $R$ through text encoder, we obtain reference text embedding $W_{r}\\in \\mathbb{R}^{C\\times D}$ . The text embedding $W_{r}$ is then directly utilized as the weights of segmentation head to perform $1\\times 1$ convolution on visual feature $Z$ and output a Reference logit map $M_{r}\\in \\mathbb{R}^{n\\times C}$ .", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we obtain different Reference logit maps for different classes, we expect the Reference logit map to encode the class-preference bias to facilitate the following bias rectification process. It is worth noting that we make Reference $R$ learnable but keep Query $Q$ fixed. It is because when we make them both learnable, Query may also capture some class-preference bias and technically we cannot guarantee which text embedding should encode the class-preference bias, resulting in an implicit bias modeling. Thus, in our framework, we choose not to make the $Q$ learnable to just encourage $R$ encode the bias. Such a design cooperates with the following logit subtraction mechanism to make the bias modeling and rectification more meaningful and effective.", + "bbox": [ + 496, + 167, + 892, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Learn Space-Preference Bias", + "text_level": 1, + "bbox": [ + 498, + 377, + 758, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In ViT [15], positional embeddings (PE) are important for encoding spatial information to features. Thus, we assume the space-preference bias should depend on the positional embeddings (PE) and choose to learn a projection of PE to represent the space-preference bias.", + "bbox": [ + 496, + 401, + 890, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design a 3-layer $3 \\times 3$ convolutional network, and each convolutional layer is followed by a batch normalization. Specifically, we project PE $p$ by the designed convolutional network to obtain positional logits $M_p \\in \\mathbb{R}^{n \\times 1}$ . During the training process, the projection network is optimized to model the space-preference bias in positional logits $M_p$ . Since PE is shared across all categories, the learned space-preference bias is also shared across all the categories.", + "bbox": [ + 496, + 479, + 892, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Rectify Bias with Contrastive Learning Loss", + "text_level": 1, + "bbox": [ + 498, + 612, + 875, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By keeping Query prompt $Q$ fixed, the Query logits $M_{q}$ represents the natural prediction ability of CLIP model, which may contain class-preference bias and space-preference bias. With learnable Reference prompt and projection of PE, we explicitly encode the class-preference bias and space-preference bias into the Reference logit map $M_{r}$ and positional logit map $M_{p}$ respectively. We then add $M_{r}$ and $M_{p}$ together to depict the final bias $M_{b}$ , i.e.,", + "bbox": [ + 496, + 637, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {b} = M _ {r} + M _ {p} ^ {*}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 772, + 890, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we simply expand $M_p$ to $M_p^* \\in \\mathbb{R}^{n \\times C}$ which means we repeat each $M_p$ for $C$ times.", + "bbox": [ + 496, + 804, + 890, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we perform a simple element-wise subtraction between $M_q$ and $M_b$ to generate the rectified logit map $M$ ,", + "bbox": [ + 498, + 835, + 890, + 867 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM = M _ {q} - M _ {b}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4105", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9632a2b4208aba1e3dbd642e4b06ea94125ba29c48b529ef06c6b493f60b1344.jpg", + "image_caption": [ + "(a) Rectification Stage" + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 491, + 255 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dacd7504c4a63c683af85660184ef7567e39f64c721250c4e26b64ff8471921b.jpg", + "image_caption": [ + "(b) Distillation", + "Figure 2. Method overview. We propose a new framework for language-guided unsupervised semantic segmentation. (a) Rectification Stage: At this stage, we aim to rectify the class-preference and space-preference bias of CLIP. (b) Distillation: We distill knowledge from the rectified CLIP to the advanced segmentation architecture with the mask-guided, feature-guided and text-guided loss terms." + ], + "image_footnote": [], + "bbox": [ + 495, + 90, + 890, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From a high-level, this operation can be interpreted as subtracting the \"bias\" from the predictions with text features of Query as segmentation head.", + "bbox": [ + 75, + 334, + 470, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the arg max operation is not differentiable, we utilize a Gumbel-Softmax trick [24] to generate the candidate masks for each class with $\\tau_{1}$ as the temperature, i.e.,", + "bbox": [ + 75, + 378, + 468, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = \\text {G u m b e l - S o f t m a x} (M, \\tau_ {1}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 431, + 468, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We apply $\\hat{M}$ to the feature $Z$ to get the masked features $Z_{g}$ , which are expected to encode the features of objects with different classes. We then design a contrastive loss to supervise the learning via aligning masked features with text features of different classes. As the masked features represent the objects of interest without context, we utilize $W_{q}$ generated by text input $Q$ for alignment.", + "bbox": [ + 76, + 455, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, as it is usually impossible for an image to contain all the classes of interest, we need to infer what classes exist in the current image. Our strategy is as follows. Firstly, we choose to extract global visual feature from image encoder of CLIP. Then we calculate the similarity scores between the text input of a specific class and the global visual feature of the image. We only choose the classes whose similarity scores are higher than a threshold $t$ as the potential classes exist in the current image.", + "bbox": [ + 75, + 561, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we may select $K$ pseudo labels $\\{c_1, c_2, \\dots, c_K\\}$ for each image. We perform a global average pooling of $Z_g$ and compute the similarities between pooled visual features of pseudo labels and text features of all the categories. The contrastive loss can be defined as", + "bbox": [ + 75, + 696, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\log \\frac {\\exp \\left\\{S _ {c _ {k} , c _ {k}} / \\tau \\right\\}}{\\sum_ {j = 1} ^ {C} \\exp \\left\\{S _ {c _ {k} , j} / \\tau \\right\\}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 777, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S_{c_k,j}$ denotes the similarity between visual feature of pseudo label $c_k$ and text feature of $j$ -th ( $j \\in \\{1,2,\\dots,C\\}$ ) category and the $\\tau$ is a constant.", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A better modeling of bias yields more accurate estimations of object masks. Then the masked features of objects", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "are more aligned with corresponding text features. As a result, the contrastive loss (Eq. (5)) will be lower. In contrast, a worse modeling of bias results in higher contrastive loss. Thus, minimizing Eq. (5) will drive the model to update towards making more accurate mask predictions, i.e., rectifying the bias of CLIP when adapted to the downstream USS task.", + "bbox": [ + 496, + 334, + 890, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Distillation for Enhanced Results", + "text_level": 1, + "bbox": [ + 498, + 446, + 790, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As CLIP is not specifically designed for segmentation tasks, we choose to distill the knowledge from the rectified CLIP (teacher) to the advanced segmentation architecture (student) to enhance the feature representations and finally improve the segmentation performance. In our paper, we choose DeepLab V2 [6] as the student network which directly inherits and fixes the Query and Reference segmentation heads, and utilizes logit subtraction mechanism to generate the final segmentation masks. In our design, the rectified CLIP works as a teacher to guide the feature learning of DeepLab V2 with three designed loss terms.", + "bbox": [ + 496, + 469, + 890, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask-guided loss. We directly exploit segmentation masks generated from the bias rectification stage as pseudo labels. We compute a cross-entropy loss between the pseudo labels $\\tilde{M}$ and predictions $M^D$ from DeepLab V2, i.e.,", + "bbox": [ + 496, + 636, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{H W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\log P _ {i j} (\\tilde {M} _ {i j}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 709, + 890, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H$ and $W$ denote the height and width of an image. The $P_{ij}(\\tilde{M}_{ij}) = \\mathrm{Softmax}(M_{ij}^D)_{\\tilde{M}_{ij}}$ , which means the predicted probability with respect to the pseudo label $\\tilde{M}_{ij}$ .", + "bbox": [ + 496, + 753, + 890, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-guided loss. After obtaining features $Z$ from visual backbone of rectified CLIP and $\\hat{Z}$ from student visual backbone, we resize two features to the same shape of original image by bilinear interpolation. We then perform a L1 loss between two features", + "bbox": [ + 496, + 804, + 890, + 877 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t}} = \\left| \\left| Z - \\hat {Z} \\right| \\right| _ {1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 883, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4106", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, visual features from student network is aligned with visual and textual features of rectified CLIP, providing an important basis for the following text-guided loss.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-guided loss. We adopt the same strategy as the Rectification stage (Sec. 3.4) to compute contrastive loss as our text-guided loss $\\mathcal{L}_{text}$ . The only difference is that we use feature $\\hat{Z}$ and masks generated by the student network to obtain the corresponding masked features.", + "bbox": [ + 75, + 136, + 467, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Total distillation loss. Total distillation loss $\\mathcal{L}_{distill}$ is calculated as follows, where $\\alpha$ and $\\beta$ are both constants and set to 0.5. Effect of each loss term is studied in Sec. 4.3.", + "bbox": [ + 75, + 210, + 467, + 257 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {f e a t}} + \\beta \\mathcal {L} _ {\\text {t e x t}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 267, + 467, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 292, + 209, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 318, + 156, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on three standard benchmarks for semantic segmentation, including PASCAL VOC 2012 [18], PASCAL Context [40] and ADE20K [61]. PASCAL VOC 2012 (1,464/1,449 train/validation) contains 20 object classes, while PASCAL Context (4,998/5,105 train/validation) is an extension of PASCAL VOC 2010 and we consider 59 most common classes in our experiments. ADE20K (20,210/2,000 train/validation) is a segmentation dataset with various scenes and 150 most common categories are considered.", + "bbox": [ + 75, + 340, + 467, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. For the image encoder of CLIP, we adopt ViT-B/16 as visual backbone. For the text encoder of CLIP, we adopt Transformer [52]. During the whole training period, we keep both of the encoders frozen. We use conventional data augmentations including random cropping and random flipping. Relevant hyper-parameters for each dataset, including number of rectification epochs, number of distillation iterations and parameters of data augmentation are shown in our supplementary material. For both stages, we use a SGD [48] optimizer with a learning rate of 0.01 and a weight decay of 0.0005. We adopt the poly strategy with the power of 0.9 for learning rate. In our experiment, we report the mean intersection over union (mIoU) as evaluation metric.", + "bbox": [ + 75, + 492, + 467, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare with previous USS methods to verify the superiority of our method, including MaskCLIP $(+)$ [62], CLIP-S4 [22], ReCo [49], CLIPpy [46], etc. From the technical view, although our method falls into the category of text-guided segmentation, we choose not to compare with text-guided methods which are designed for zero-shot or open-vocabulary setting (e.g., GroupViT [57], TCL [5] and ViewCo [47]). Two main differences exist between the zero-shot/open-vocabulary setting and the unsupervised setting: 1) in zero-shot/open-vocabulary, the category sets for training and inference are non-overlapped, but in USS the category sets are shared; 2) text-guided zero-shot/open-vocabulary methods usually rely on large-scale", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/330846b29a901e9a6ac4c44150a7ebf593351f207d20c5eff4dd4267138b9970.jpg", + "table_caption": [ + "Table 1. Comparison with non-language-guided unsupervised semantic segmentation methods on PASCAL VOC." + ], + "table_footnote": [], + "table_body": "
ModelLCmIoU
IIC [25]✓9.8
SegSort [23]✓11.7
Deep Spectral Methods [39]✗37.2
HSG [29]✗41.9
ACSeg [34]✗47.1
MaskContrast [51]✓49.6
Ours (rectification)✗58.5
Ours (distill)✗75.4
Ours (distill)✓76.1
", + "bbox": [ + 539, + 128, + 846, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "image-text pairs or class-agnostic masks to supervise the training, but in USS we only utilize the images without any types of annotations to train the model.", + "bbox": [ + 498, + 306, + 890, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA methods", + "text_level": 1, + "bbox": [ + 498, + 362, + 790, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with both previous non-language-guided USS methods (Table 1) and CLIP-based USS methods (Table 2). In all tables, we use \"Ours (rectification)\" to denote the segmentation results after the rectification stage and \"Ours (distill)\" to denote the segmentation results after distillation. Unless otherwise stated, results from previous methods are directly cited from the original papers.", + "bbox": [ + 496, + 386, + 890, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, our model shows remarkably better segmentation results compared with the conventional non-language-guided USS methods. In order to further evaluate the strength of features extracted from our distillation model, we also report our results of linear classification (LC). Following MaskContrast [51], we fix our model to generate pixel embeddings and train a linear classifier to generate semantic segmentation masks. The results of LC also prove that our method extracts strong features and achieves significant gains.", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 2, we also make a comparison with typical CLIP-based methods, including MaskCLIP, CLIPpy, ReCo and CLIP-S4. These methods all consider the same CLIP-based USS setting with ours. The results show our method performs favorably against previous CLIP-based methods. For example, after the rectification stage, our method outperforms MaskCLIP by $9.0\\%$ , $4.1\\%$ and $1.6\\%$ respectively on the three datasets, while after distillation, our method outperforms $\\mathrm{MaskCLIP + }$ by $5.4\\%$ , $2.7\\%$ and $2.1\\%$ . Our method also shows better results than CLIPpy by $20.8\\%$ and $0.8\\%$ on PASCAL VOC and ADE20K and outperforms CLIP-S4 by $3.4\\%$ and $0.2\\%$ on PASCAL VOC and PASCAL Context. Since ReCo employs a \"context elimination\" (CE) trick which may introduce prior knowledge, we also report the results of ReCo by removing this trick (ReCo w/o CE in Table 2). Our method outperforms ReCo obviously, e.g., on PASCAL VOC, ours (rectification) outper", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4107", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e96df8ce99ceafd4509e47896b340fe78b6c6bf3086aa53e1b8cc9925bd09e6.jpg", + "table_caption": [ + "Table 2. Comparison with CLIP-based unsupervised semantic segmentation methods on various benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodVOCContextADE
CLIPpy [46]54.6/13.5
MaskCLIP [62]49.521.79.5
MaskCLIP+ [62]70.031.112.2
ReCo [49]55.226.2/
ReCo (w/o CE)54.823.1/
CLIP-S4 [22]72.033.6/
Ours (rectification)58.525.811.1
Ours (distill)75.433.814.3
", + "bbox": [ + 107, + 128, + 433, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8421e69c7116fabbec9a9b29faa933b2d97cbb2b9c7ee073c7c05060246ece9.jpg", + "table_caption": [ + "Table 3. Ablation on whether Query should be learnable. Results show that fixing Query performs favorably against making Query learnable." + ], + "table_footnote": [], + "table_body": "
QueryReferenceVOCContextADE
✓✓56.723.79.4
X✓56.724.410.5
", + "bbox": [ + 106, + 349, + 433, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "forms ReCo and ReCo (w/o CE) by $3.3\\%$ and $3.7\\%$ respectively.", + "bbox": [ + 76, + 415, + 467, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 76, + 454, + 227, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Should Query be learnable? We study on whether the prompt of Query should be learnable. As shown in Table 3, we conduct experiments with learnable and fixed Query respectively. As we aim to show the effect of fixing Query, we don't utilize the projected positional logits to calculate the numbers shown in the table. The numbers show that fixing Query performs favorably against making Query learnable. This verifies that we should only learn Reference but keep Query fixed as we discuss in Sec. 3.2.", + "bbox": [ + 75, + 477, + 467, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different bias modeling. As shown in Table 4, compared with the baseline without any bias rectification (numbers with only \"Query\"), utilizing learable Reference to model the class-preference bias brings remarkable performance improvement, e.g., $7.2\\%$ mIoU on PASCAL VOC, $2.7\\%$ mIoU on PASCAL Context and $1.0\\%$ mIoU on ADE20K. While introducing projection of positional embedding to model the space-preference bias, the numbers are further improved, i.e., around $1\\%$ better than modeling the class-preference bias only. Those results verify the effectiveness of bias modeling by introducing learnable Reference and projection of positional embedding.", + "bbox": [ + 75, + 613, + 467, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of element-wise subtraction on bias rectification. In order to validate the effect of our element-wise subtraction, we conduct experiments in Table 5. We compare our subtraction mechanism with an alternative solution, i.e., instead of subtracting logits encoding bias (i.e., Reference logits plus positional logits), we add all the logits. Comparisons shown in the table verify the effectiveness of our", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b884a09fa8ff0739dcad062827fb478849ddf78862d602d14bb096d8982ec2f9.jpg", + "table_caption": [ + "Table 4. Effect of bias modeling. Results show effect of both bias modeling quantitatively and each contributes to better results." + ], + "table_footnote": [], + "table_body": "
QueryReferencePEVOCContextADE
✓49.521.79.5
✓✓56.724.410.5
✓✓✓58.525.811.1
", + "bbox": [ + 511, + 128, + 877, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b47a2eefadab7b46db1076d055193b988896bf2ff38855dde8b069a8edb78450.jpg", + "table_caption": [ + "Table 5. Effectiveness of element-wise subtraction. Results show that the element-wise subtraction removes bias from CLIP." + ], + "table_footnote": [], + "table_body": "
QueryRef+PEsubaddVOCContext
✓XXX49.521.7
✓✓X✓57.824.9
✓✓✓X58.525.8
", + "bbox": [ + 519, + 258, + 864, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "subtraction way. As we aim to remove the bias from the original CLIP, we speculate that the subtraction operation may work as a strong prior which regularizes the training to facilitate meaningful and effective bias modeling. The effect of bias rectification by element-wise subtraction is further visualized in Fig. 3.", + "bbox": [ + 496, + 340, + 890, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of Bias Modeling. In Fig. 3 (b), we illustrate how we explicitly model and rectify the class-preference bias. As shown in the first column of Fig. 3 (b), the original CLIP (see \"Query\" mask) tends to misclassify part of a \"person\" (see the ground-truth mask denoted as \"GT\") into a \"boat\". Such a mistake is reflected in the comparison between the logit heatmaps for the \"person\" channel and the \"boat\" channel: in the area misclassified as \"boat\", the logits for \"boat\" are relatively higher than those for \"person\". In contrast, for Reference logit map, the person-channel logits are quite low, while the boat-channel logits are generally very high, especially for the misclassified area. Consequently, via the proposed logit subtraction operation, we obtain the rectified logits (see the last column), where the boat channel is largely suppressed. Finally, we obtain a much better mask (see \"Ours\" in the first column).", + "bbox": [ + 496, + 431, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 3 (c), we aim to show that the projection of positional embedding (PE) effectively models the space-preference bias. By comparing the results with and without rectifying logits projected by PE (see the dashed boxes of the last two columns), we find that rectification with PE largely improves the segmentation performance in the boundary areas. The results illustrate that the projection of PE does encode the space-preference bias and rectifying such bias may largely improve the segmentation results.", + "bbox": [ + 496, + 672, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different distillation loss terms. We conduct experiments on PASCAL VOC to validate the effect of each loss term of our distillation framework, including the mask-guided, feature-guided and text-guided loss terms. From Table 7, all the loss terms contribute to better performance.", + "bbox": [ + 496, + 809, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of segmentation head for distillation. As illustrated", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4108", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2085a7ff846c08481c69541cf64fe3d979f818cb9e131f98ef34b9ad308fd423.jpg", + "image_caption": [ + "Figure 3. (a) Qualitative Results: We visualize segmentation results on PASCAL Context. From the visualization, we observe that our model outperforms MaskCLIP $(+)$ obviously by rectifying both class-preference bias and space-preference bias. (b) Class-preference bias: In order to explain how Reference explicitly models the class-preference bias, we show the heatmap of Reference logits. (c) Space-preference bias: The segmentation within dashed boxes shows the effectiveness of PE projection on rectifying space-preference bias." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 495, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d62c94cdbc5ddb8cba0e249a0cf1658006ca0fc919781fbf04b60085c31c61c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 890, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9ba8efd3af1bde7e8b5ad712b680b3dc3dd68f6e260144384a30904ba71d52ba.jpg", + "table_caption": [ + "Table 6. Effect of segmentation head for distillation. We perform distillation with both original classification head (Ori. Head) and structure from our rectification stage" + ], + "table_footnote": [], + "table_body": "
Ori. HeadQueryReferencePE Proj.mIoU
✓XXX73.8
X✓XX74.1
X✓✓X75.4
X✓✓✓75.0
", + "bbox": [ + 86, + 463, + 452, + 553 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f198ffd3fc4cefebbea7192ef3ce98c4301eb3a51a2e8aa2449045a7109364b3.jpg", + "table_caption": [ + "Table 7. Ablation results on distillation loss. According to the results, each loss item contributes to better distillation result." + ], + "table_footnote": [], + "table_body": "
MaskFeatureTextmIoU
✓73.0
✓✓74.1
✓✓✓75.4
", + "bbox": [ + 151, + 593, + 387, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig 2 (b), instead of directly using original classification head at the distillation stage, the student network (i.e., DeepLab V2 [6]) inherits and fixes the Query and Reference segmentation heads from our rectification stage, and utilizes logit subtraction mechanism to generate the final segmentation masks. In Table 6, we conduct experiments with original head (\"Ori. Head\") of DeepLab V2 and other types of segmentation heads which are directly inherited from our rectification stage on PASCAL VOC. From the results, we observe that it is better to inherit the components from our rectification stage than adopting the original classification head. However, introducing projection of PE in segmentation head cannot bring further improvement.", + "bbox": [ + 75, + 688, + 468, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. We visualize our segmentation re", + "bbox": [ + 76, + 885, + 467, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults in Fig. 3 (a). It can be observed that there exists apparent space-preference bias and class-preference bias in the segmentation results of original CLIP (MaskCLIP) and these bias still cannot be removed even after distillation (MaskCLIP+). However, our model outperforms MaskCLIP(+) obviously by rectifying the bias of CLIP for unsupervised semantic segmentation.", + "bbox": [ + 496, + 412, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 556, + 617, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new framework for language-guided unsupervised semantic segmentation. We observe bias, including space-preference bias and class-preference bias, exists in CLIP when directly applying CLIP to segmentation task. We propose using additional Reference to learn class-preference bias and projecting positional embedding to represent space-preference bias, and then manage to rectify them by a simple element-wise logit subtraction mechanism. For further improving the segmentation performance, we distill the knowledge from rectified CLIP to advanced segmentation backbone with specifically designed losses. Extensive experiments demonstrate that our method achieve superior segmentation performance compared to previous state-of-the-arts. We hope our work may inspire future research to investigate how to better adapt CLIP to complex visual understanding tasks.", + "bbox": [ + 496, + 590, + 890, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This research project is supported by National Natural Science Foundation (NNSF) of China under Grant 92370114 and National Key R&D Program of China under Grant 2022YFB3104900.", + "bbox": [ + 496, + 839, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4109", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 3", + "[2] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[5] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 6, 1", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2018. 1, 5, 8", + "[7] Mickaël Chen, Thierry Artières, and Ludovic Denoyer. Unsupervised object segmentation by redrawing. Advances in neural information processing systems, 32, 2019. 3", + "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European Conference on Computer Vision, pages 104-120, 2020. 3", + "[9] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 1", + "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1", + "[11] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 3", + "[12] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 3", + "[13] Zhijie Deng and Yucen Luo. Learning neural eigenfunctions for unsupervised semantic segmentation. arXiv preprint arXiv:2304.02841, 2023. 3" + ], + "bbox": [ + 76, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3", + "[15] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4", + "[16] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 3", + "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3", + "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 6", + "[19] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 540-557. Springer, 2022. 3", + "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. 3", + "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1", + "[22] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2023. 1, 2, 3, 6, 7", + "[23] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7334-7344, 2019. 1, 3, 6", + "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings, 2017. 5", + "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 1, 3, 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4110", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3", + "[27] Guoliang Kang, Lu Jiang, Yunchao Wei, Yi Yang, and Alexander Hauptmann. Contrastive adaptation network for single-and multi-source domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 44(4): 1793-1804, 2020. 1", + "[28] Guoliang Kang, Yunchao Wei, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Pixel-level cycle association: A new perspective for domain adaptive semantic segmentation. Advances in neural information processing systems, 33: 3569-3580, 2020. 1", + "[29] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571–2581, 2022. 1, 3, 6", + "[30] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3", + "[31] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 3", + "[32] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI conference on artificial intelligence, pages 11336-11344, 2020. 3", + "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3", + "[34] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 1, 3, 6", + "[35] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 3", + "[36] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, et al. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1, 3", + "[38] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3", + "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3, 6", + "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 2, 6", + "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European Conference on Computer Vision, pages 529-544. Springer, 2022. 3", + "[42] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 142-158. Springer, 2020. 1, 3", + "[43] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3", + "[44] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1713-1721, 2015. 3", + "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4", + "[46] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. arXiv preprint arXiv:2210.09996, 2022. 1, 3, 6, 7", + "[47] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 3, 6", + "[48] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4111", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In Advances in Neural Information Processing Systems. 1, 3, 6, 7", + "[50] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 3", + "[51] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 3, 6", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6", + "[53] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 1, 3", + "[54] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. arXiv preprint arXiv:2303.11749, 2023. 1, 3", + "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3", + "[56] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. C2am: contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989–998, 2022. 3", + "[57] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 3, 6, 1", + "[58] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023.", + "[59] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 3", + "[60] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1, 3" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] B. Zhou, Z. Hang, Francesco Xavier Puig Fernandez, S. Fidler, and A. Torralba. Scene parsing through ade20k dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6", + "[62] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 1, 3, 4, 6, 7", + "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + ], + "bbox": [ + 501, + 90, + 892, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4112", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_model.json b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44648a956d0f861cf1bb4e712f5ade65fdf4c07f --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.182, + 0.352, + 0.217 + ], + "angle": 0, + "content": "Jingyun Wang \nBeihang University" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.22, + 0.362, + 0.233 + ], + "angle": 0, + "content": "19231136@buaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.414, + 0.182, + 0.788, + 0.217 + ], + "angle": 0, + "content": "Guoliang Kang * \nBeihang University, Zhongguancun Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.22, + 0.681, + 0.234 + ], + "angle": 0, + "content": "kgl.prml@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.681 + ], + "angle": 0, + "content": "Recent works utilize CLIP to perform the challenging unsupervised semantic segmentation task where only images without annotations are available. However, we observe that when adopting CLIP to such a pixel-level understanding task, unexpected bias occurs. Previous works don't explicitly model such bias, which largely constrains the segmentation performance. In this paper, we propose to explicitly model and rectify the bias existing in CLIP to facilitate the unsupervised semantic segmentation. Specifically, we design a learnable \"Reference\" prompt to encode class-preference bias and project the positional embedding of vision transformer to represent space-preference bias. Via a simple element-wise subtraction, we rectify the logits of CLIP classifier. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then a contrastive loss between masked visual feature and the text features of different classes is imposed to facilitate the effective bias modeling. To further improve the segmentation, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture via minimizing our designed mask-guided, feature-guided and text-guided loss terms. Extensive experiments on standard benchmarks demonstrate that our method performs favorably against previous state-of-the-arts. The implementation is available at https://github.com/dogehhh/ReLUCP." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.21, + 0.727 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.47, + 0.875 + ], + "angle": 0, + "content": "Semantic segmentation aims to attach a semantic label to each pixel of an image. Since the rising of deep learning [27, 28, 45, 54], semantic segmentation has been widely adopted in real-world applications, e.g., autonomous driving, medical image segmentation, etc. Conventional approaches [6, 9, 10, 37, 38, 53, 60] for semantic segmentation have achieved remarkable performance. However, the superior performance of those methods relies heavily on large amounts of fully annotated masks. Collecting such high" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.362 + ], + "angle": 0, + "content": "quality pixel-level annotations can be both time-consuming and expensive, e.g., some annotations for specialized tasks require massive expert knowledge, some are even inaccessible due to privacy reasons, etc. Therefore, it is necessary to explore unsupervised semantic segmentation where only images without annotations are available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.895, + 0.532 + ], + "angle": 0, + "content": "Unsupervised semantic segmentation (USS) has been studied for years. Many non-language-guided USS methods have been proposed, e.g., clustering-based methods [11, 25, 29, 34, 42], contrastive-learning-based methods [21, 51], boundary-based methods [23], etc. Despite promising progress achieved, there still exhibits a large performance gap between USS and the supervised segmentation methods. Besides, these methods typically obtain class-agnostic masks and have to depend on additional processing (e.g., Hungarian matching) to assign semantic labels to the masks, rendering them less practical in real scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.896, + 0.777 + ], + "angle": 0, + "content": "Recently, large-scale visual-language pre-trained models, e.g., CLIP [45], demonstrate superior zero-shot classification performance by comparing the degree of alignment between image feature and text features of different categories. A few CLIP-based USS approaches [22, 46, 49, 62] emerge and show remarkable performance improvement compared with the non-language-guided USS methods. These models require no access to any types of annotations, and directly assign a label to each pixel, benefiting from the aligned vision and text embedding space of CLIP. However, good alignment between image-level visual feature and textural feature doesn't necessarily mean good alignment between pixel-level visual feature and textural feature. Thus, for CLIP, unexpected bias may inevitably appear. Previous works didn't explicitly model such bias, which may largely constrain their segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We observe two kinds of bias existing in CLIP. From one aspect, as shown in Fig. 1(a), CLIP exhibits space-preference bias. CLIP performs apparently better for segmenting central objects than the ones distributed near the image boundary. It can be reflected by the fact that mIoU decreases as the distance between the centroids of object and the image increases. From the other aspect, as shown in Fig. 1(b), there exists class-preference bias between similar" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4102" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.09, + 0.434, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.091, + 0.535, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.092, + 0.634, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.092, + 0.728, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.092, + 0.829, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.161, + 0.535, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.161, + 0.635, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.161, + 0.729, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.161, + 0.829, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.226, + 0.889, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.892, + 0.482 + ], + "angle": 0, + "content": "Figure 1. (a) Space-preference bias. (Left): The relationship between distance \\((x\\)-axis) and mIoU \\((y\\)-axis) is drawn on PASCAL VOC [18]. The distance means the spatial distance between the centroids of the object and the image and mIoU is computed based on predictions and ground truth. The curve shows that CLIP [45] (green) is apparently better at segmentation for central objects than boundary ones, but our method (blue) effectively mitigates this bias. More details about how we draw this figure has been shown in our supplementary material. (Right): Visualizations also show our improvement on space-preference bias qualitatively. (b) Class-preference bias. (Left): We randomly select 6 classes from PASCAL Context [40] and draw the confusion matrix of CLIP and our model. It shows that beside the ground truth, CLIP also prefers to assign an incorrect but relevant label to a pixel in quite a few cases, while our results show apparent improvement. (Right): The visualizations are consistent with what we observed in confusion matrix. For example, for a \"cow\", CLIP tends to classify it as \"dog\" incorrectly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.469, + 0.593 + ], + "angle": 0, + "content": "categories in CLIP. For example, according to visualization (right), when ground truth is \"cow\", CLIP tends to incorrectly classify it as \"dog\". We also show such trend between randomly selected classes by confusion matrix (left). Elements on diagonal line represent right classification, while others are false. We observe a wide range of class-preference bias introduced by CLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose to explicitly model and rectify the bias of CLIP to facilitate the USS. Specifically, we design two kinds of text inputs for each class, which are named as \"Reference\" and \"Query\" respectively. The Query is manually designed while the Reference contains learnable prompts. We adopt the text features of Query and Reference as classifiers to generate the Query and Reference logits respectively for each pixel of image. The Query logits represent the segmentation ability of original CLIP, while the Reference logits are expected to reflect the bias of CLIP preferring a specific class. Additionally, we project the positional embedding of CLIP's vision transformer to generate positional logits for each image. We expect the positional logits to represent space-preference bias of CLIP. Then, we remove the class-preference and the space-preference bias from original CLIP via a logit-subtraction mechanism, i.e., subtracting the Reference logits and the positional logits from the Query logits. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then the contrastive loss between masked visual" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.577 + ], + "angle": 0, + "content": "feature and the text features of different categories is imposed to facilitate the effective bias modeling and rectification. To further improve the segmentation performance, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture, via mask-guided distillation, feature-guided distillation and text-guided learning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.7 + ], + "angle": 0, + "content": "We conduct extensive experiments on standard semantic segmentation benchmarks, including PASCAL VOC [18], PASCAL Context [40] and ADE20K [61]. Experiment results demonstrate that our method performs favourably against previous state-of-the-arts. Notably, on PASCAL VOC, our method outperforms CLIP S4 [22] by \\(3.4\\%\\). Extensive ablation studies verify the effectiveness of each design in our framework." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.703, + 0.822, + 0.716 + ], + "angle": 0, + "content": "Our contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.72, + 0.892, + 0.795 + ], + "angle": 0, + "content": "- We observe that when applying CLIP to pixel-level understanding tasks, unexpected bias including space-preference bias and class-preference bias, occurs. Such bias may largely constrain the segmentation performance of CLIP-based segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.796, + 0.892, + 0.886 + ], + "angle": 0, + "content": "- We propose to explicitly model the class-preference and space-preference bias of CLIP via learnable Reference text inputs and projection of positional embedding. Through a simple logit-subtraction mechanism and the contrastive loss built on masked features, we effectively rectify the bias of CLIP." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.887, + 0.892, + 0.902 + ], + "angle": 0, + "content": "- We conduct extensive experiments on segmentation" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.72, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "benchmarks under the unsupervised setting. Experiment results show superior performance of our method to previous state-of-the-arts." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.151, + 0.22, + 0.167 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.177, + 0.47, + 0.327 + ], + "angle": 0, + "content": "Pre-trained vision-language models. Pre-trained vision-language models (VLMs) [8, 14, 32, 33, 35] have developed rapidly with the help of large-scale image-text pairs available on the Internet. Recently, CLIP [45], ALIGN [26] and Slip [41] have made great progress on learning visual and textual representations jointly by using contrastive learning. With the image-level alignment with text, pre-trained VLMs have strong ability for zero-shot classification task and can be transferred to various downstream tasks, such as object detection [17, 54] and semantic segmentation [62]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Unsupervised semantic segmentation. While conventional approaches of semantic segmentation [37, 38, 53, 60] rely on pixel-level annotations and weakly-supervised methods [1, 16, 44, 56] still ask for image-level labels, unsupervised semantic segmentation (USS) methods [11, 20, 42, 50, 51] explores to train a segmentation model without any annotations. Models like [2, 7] adopt generative model [12] to separate foreground with background or generate corresponding masks. SegSort [23], HSG[29] and ACSeg [34] use clustering strategy, while IIC [25] uses mutual information maximization to perform unsupervised learning. MaskContrast [51] introduces contrastive learning into USS. Others like DSM [39] and LNE [13] exploit features extracted from self-supervised model, and combine it with spectral graph theory. However, the methods mentioned above either fail to segment images with multi-category objects or show a large performance gap with the supervised methods. Besides, they can only obtain class-agnostic masks and have to depend on additional strategy, such as Hungarian-matching algorithm [30], to match corresponding category with masks. Recently, pre-trained vision-language models are adopted in USS. MaskCLIP [62] modifies the image encoder of CLIP to generate patch-level features and directly performs segmentation with text features as classifiers. CLIP-py [46] performs contrastive learning between visual features from self-supervised ViT [4] and text features from CLIP. ReCo [49] performs image retrieval with CLIP and extracts class-wise embedding as classifier with co-segmentation. CLIP-S4 [22] learns pixel embeddings with pixel-segment contrastive learning and aligns such embeddings with CLIP in terms of embedding and semantic consistency. These methods can directly assign a category to each pixel, whose setting is named as language-guided unsupervised semantic segmentation. However, directly applying CLIP may bring prior bias, including space-preference bias and class-preference bias. As we know, there is no previous method trying to solve these bias and we manage to explicitly model the bias, and rectify them by" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.093, + 0.672, + 0.106 + ], + "angle": 0, + "content": "element-wise subtraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Language-guided semantic segmentation. Recently, many works are exploring semantic segmentation guided by language under different settings. Zero-shot works [3, 31, 43, 55] split classes into seen and unseen set. During the training period, only masks of seen classes are provided. For inference, models are tested on both seen and unseen classes, but the test data is still in the same distribution with the training data. Open-vocabulary works [5, 19, 36, 47, 57-59] are trained in one scenario with extra annotations including class-agnostic masks or image captions, but are used for predicating segmentation masks of novel classes in other scenarios. From the technical view, our method also falls into the category of language-guided semantic segmentation. However, we consider the unsupervised setting. In this setting, we have access to images without any annotations during training. The training and inference images are sampled from the same distributions and the same set of categories. Such a setting is different to the typical zero-shot or open-vocabulary setting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.593, + 0.423 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Background In this work, we aim to rectify the bias of CLIP for unsupervised semantic segmentation. In USS, we only have access to images without any types of annotations to train the segmentation model. For training and inference, the same set of categories are considered and the data distributions are assumed to be the same." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Overview The general framework of our method is illustrated in Fig. 2. We aim to rectify the bias of CLIP including the class-preference bias and the space-preference bias, to facilitate unsupervised semantic segmentation. From a high level, class-preference bias reflects the shift of CLIP predictions towards specific classes, while space-preference bias reflects the shift of CLIP predictions towards specific spatial locations. Both biases will be finally reflected in the predicted logit maps. A reasonable way to rectify the bias is to subtract its logit maps from the normal logit maps predicted via original CLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To realize our goal, we firstly forward the image \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\) through the image encoder of CLIP to obtain the patch-level image features \\(Z\\). We design two kinds of text inputs for each class. One is the manually designed Query text input \\(Q\\) and the other one is the Reference text input \\(R\\) which consists of a learnable prompt and the name of class. For each class, passing \\(Q\\) and \\(R\\) through the text encoder of CLIP, we obtain two text embeddings \\(W_{q}\\) and \\(W_{r}\\), which serve as the weights of query segmentation head and reference segmentation head respectively. Taking the same visual feature \\(Z\\) as input, the query and reference segmentation heads output a query logit map \\(M_{q}\\) (Sec. 3.1) and Reference logit map \\(M_{r}\\) (Sec. 3.2) respectively. Meanwhile, positional embedding \\(p\\) is sent into a" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "learnable convolutional network to generate positional logit map \\( M_{p} \\) (Sec. 3.3). Then the bias logit map \\( M_{b} \\) can be obtained by adding Reference logit map \\( M_{r} \\) and positional logit map \\( M_{p} \\). We then subtract \\( M_{b} \\) from \\( M_{q} \\) to generate the rectified logits \\( M \\). Based on the rectified logits, we generate masks for different classes. Then a contrastive loss is imposed between the masked visual features and the text features of different categories to encourage the bias modeling and rectification (Sec. 3.4). In order to enhance the segmentation performance, we distill the knowledge of rectified CLIP to the advanced segmentation architecture with specifically designed mask-guided, feature-guided and text-guided loss terms (Sec. 3.5). In both the rectification and distillation stages, we keep CLIP frozen." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.312, + 0.411, + 0.328 + ], + "angle": 0, + "content": "3.1. Baseline: Directly Segment with CLIP" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.47, + 0.501 + ], + "angle": 0, + "content": "Following MaskCLIP [62], we adapt pre-trained CLIP [45] (ViT-B/16) to the semantic segmentation task. We remove the query and key embedding layers of last attention but reformulate the value embedding layer and the last linear projection layer into two respective \\( 1 \\times 1 \\) convolutional layers. Therefore, the image encoder can not only generate local features for dense predictions, but also keep the visual-language association in CLIP by freezing its pre-trained weights. We forward image \\( I \\) through the image encoder and obtain patch-level features \\( Z \\in \\mathbb{R}^{n \\times D} \\) (\\( n \\) is the number of patches and \\( D \\) is the dimension of features in CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Each text in Query \\( Q = \\{Q_{1}, Q_{2}, \\dots, Q_{C}\\} \\) (\\( C \\) is the number of classes) is an ensemble of several manually designed prompts, e.g., \"a good/large/bad photo of a [CLS]\" where [CLS] denotes the a specific class name. Passing \\( Q \\) through the text encoder, we obtain its text embeddings \\( W_{q} \\). We treat text embeddings \\( W_{q} \\) as the weight of segmentation head to perform \\( 1 \\times 1 \\) convolution. By sending features \\( Z \\) to the segmentation head, we get a Query logit map \\( M_{q} \\in \\mathbb{R}^{n \\times C} \\). Then the segmentation mask can be predicted by arg max operation on \\( M_{q} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.66, + 0.334, + 0.675 + ], + "angle": 0, + "content": "3.2. Learn Class-Preference Bias" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.471, + 0.777 + ], + "angle": 0, + "content": "In order to explicitly model the class-preference bias brought by pre-trained CLIP for specific datasets, we design a Reference \\( R_{i} \\), \\( i = \\{1,2,\\dots ,C\\} \\), as additional text input for each class. Inspired by CoOp [63], \\( R_{i} \\) consists of a learnable prompt which is shared across all the classes for efficiency, and a class name \\( [CLS] \\), which can be formed as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.788, + 0.47, + 0.806 + ], + "angle": 0, + "content": "\\[\nR _ {i} = \\left[ v _ {1} \\right] \\left[ v _ {2} \\right] \\dots \\left[ v _ {l} \\right] \\dots \\left[ v _ {L} \\right] [ C L S ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "where each \\([v_l](l \\in \\{1, \\dots, L\\})\\) is a vector with dimension \\(D\\) and serves as a word embedding. The \\(L\\) is a constant representing the number of word embeddings to learn. Totally, there are 77 word embeddings for a text of CLIP. As two word embeddings are used for class name and two for indicating start and end of a text, we set \\(L\\) to 73." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Passing the Reference \\(R\\) through text encoder, we obtain reference text embedding \\(W_{r}\\in \\mathbb{R}^{C\\times D}\\). The text embedding \\(W_{r}\\) is then directly utilized as the weights of segmentation head to perform \\(1\\times 1\\) convolution on visual feature \\(Z\\) and output a Reference logit map \\(M_{r}\\in \\mathbb{R}^{n\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.169, + 0.893, + 0.366 + ], + "angle": 0, + "content": "As we obtain different Reference logit maps for different classes, we expect the Reference logit map to encode the class-preference bias to facilitate the following bias rectification process. It is worth noting that we make Reference \\( R \\) learnable but keep Query \\( Q \\) fixed. It is because when we make them both learnable, Query may also capture some class-preference bias and technically we cannot guarantee which text embedding should encode the class-preference bias, resulting in an implicit bias modeling. Thus, in our framework, we choose not to make the \\( Q \\) learnable to just encourage \\( R \\) encode the bias. Such a design cooperates with the following logit subtraction mechanism to make the bias modeling and rectification more meaningful and effective." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.378, + 0.759, + 0.394 + ], + "angle": 0, + "content": "3.3. Learn Space-Preference Bias" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.478 + ], + "angle": 0, + "content": "In ViT [15], positional embeddings (PE) are important for encoding spatial information to features. Thus, we assume the space-preference bias should depend on the positional embeddings (PE) and choose to learn a projection of PE to represent the space-preference bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.48, + 0.893, + 0.601 + ], + "angle": 0, + "content": "We design a 3-layer \\(3 \\times 3\\) convolutional network, and each convolutional layer is followed by a batch normalization. Specifically, we project PE \\(p\\) by the designed convolutional network to obtain positional logits \\(M_p \\in \\mathbb{R}^{n \\times 1}\\). During the training process, the projection network is optimized to model the space-preference bias in positional logits \\(M_p\\). Since PE is shared across all categories, the learned space-preference bias is also shared across all the categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.613, + 0.877, + 0.63 + ], + "angle": 0, + "content": "3.4. Rectify Bias with Contrastive Learning Loss" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.638, + 0.892, + 0.76 + ], + "angle": 0, + "content": "By keeping Query prompt \\( Q \\) fixed, the Query logits \\( M_{q} \\) represents the natural prediction ability of CLIP model, which may contain class-preference bias and space-preference bias. With learnable Reference prompt and projection of PE, we explicitly encode the class-preference bias and space-preference bias into the Reference logit map \\( M_{r} \\) and positional logit map \\( M_{p} \\) respectively. We then add \\( M_{r} \\) and \\( M_{p} \\) together to depict the final bias \\( M_{b} \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.773, + 0.891, + 0.792 + ], + "angle": 0, + "content": "\\[\nM _ {b} = M _ {r} + M _ {p} ^ {*}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.891, + 0.836 + ], + "angle": 0, + "content": "where we simply expand \\(M_p\\) to \\(M_p^* \\in \\mathbb{R}^{n \\times C}\\) which means we repeat each \\(M_p\\) for \\(C\\) times." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.837, + 0.892, + 0.868 + ], + "angle": 0, + "content": "Then we perform a simple element-wise subtraction between \\( M_q \\) and \\( M_b \\) to generate the rectified logit map \\( M \\)," + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\nM = M _ {q} - M _ {b}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4105" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.493, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.261, + 0.323, + 0.272 + ], + "angle": 0, + "content": "(a) Rectification Stage" + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.092, + 0.891, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.261, + 0.736, + 0.272 + ], + "angle": 0, + "content": "(b) Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.894, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Method overview. We propose a new framework for language-guided unsupervised semantic segmentation. (a) Rectification Stage: At this stage, we aim to rectify the class-preference and space-preference bias of CLIP. (b) Distillation: We distill knowledge from the rectified CLIP to the advanced segmentation architecture with the mask-guided, feature-guided and text-guided loss terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.471, + 0.379 + ], + "angle": 0, + "content": "From a high-level, this operation can be interpreted as subtracting the \"bias\" from the predictions with text features of Query as segmentation head." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.426 + ], + "angle": 0, + "content": "Since the arg max operation is not differentiable, we utilize a Gumbel-Softmax trick [24] to generate the candidate masks for each class with \\(\\tau_{1}\\) as the temperature, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.433, + 0.469, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = \\text {G u m b e l - S o f t m a x} (M, \\tau_ {1}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.456, + 0.47, + 0.563 + ], + "angle": 0, + "content": "We apply \\(\\hat{M}\\) to the feature \\(Z\\) to get the masked features \\(Z_{g}\\), which are expected to encode the features of objects with different classes. We then design a contrastive loss to supervise the learning via aligning masked features with text features of different classes. As the masked features represent the objects of interest without context, we utilize \\(W_{q}\\) generated by text input \\(Q\\) for alignment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.563, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Additionally, as it is usually impossible for an image to contain all the classes of interest, we need to infer what classes exist in the current image. Our strategy is as follows. Firstly, we choose to extract global visual feature from image encoder of CLIP. Then we calculate the similarity scores between the text input of a specific class and the global visual feature of the image. We only choose the classes whose similarity scores are higher than a threshold \\( t \\) as the potential classes exist in the current image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.698, + 0.469, + 0.773 + ], + "angle": 0, + "content": "Then we may select \\(K\\) pseudo labels \\(\\{c_1, c_2, \\dots, c_K\\}\\) for each image. We perform a global average pooling of \\(Z_g\\) and compute the similarities between pooled visual features of pseudo labels and text features of all the categories. The contrastive loss can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.778, + 0.469, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\log \\frac {\\exp \\left\\{S _ {c _ {k} , c _ {k}} / \\tau \\right\\}}{\\sum_ {j = 1} ^ {C} \\exp \\left\\{S _ {c _ {k} , j} / \\tau \\right\\}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.87 + ], + "angle": 0, + "content": "where \\( S_{c_k,j} \\) denotes the similarity between visual feature of pseudo label \\( c_k \\) and text feature of \\( j \\)-th (\\( j \\in \\{1,2,\\dots,C\\} \\)) category and the \\( \\tau \\) is a constant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "A better modeling of bias yields more accurate estimations of object masks. Then the masked features of objects" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.335, + 0.892, + 0.44 + ], + "angle": 0, + "content": "are more aligned with corresponding text features. As a result, the contrastive loss (Eq. (5)) will be lower. In contrast, a worse modeling of bias results in higher contrastive loss. Thus, minimizing Eq. (5) will drive the model to update towards making more accurate mask predictions, i.e., rectifying the bias of CLIP when adapted to the downstream USS task." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.448, + 0.792, + 0.462 + ], + "angle": 0, + "content": "3.5. Distillation for Enhanced Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.637 + ], + "angle": 0, + "content": "As CLIP is not specifically designed for segmentation tasks, we choose to distill the knowledge from the rectified CLIP (teacher) to the advanced segmentation architecture (student) to enhance the feature representations and finally improve the segmentation performance. In our paper, we choose DeepLab V2 [6] as the student network which directly inherits and fixes the Query and Reference segmentation heads, and utilizes logit subtraction mechanism to generate the final segmentation masks. In our design, the rectified CLIP works as a teacher to guide the feature learning of DeepLab V2 with three designed loss terms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Mask-guided loss. We directly exploit segmentation masks generated from the bias rectification stage as pseudo labels. We compute a cross-entropy loss between the pseudo labels \\(\\tilde{M}\\) and predictions \\(M^D\\) from DeepLab V2, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.71, + 0.892, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{H W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\log P _ {i j} (\\tilde {M} _ {i j}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.754, + 0.892, + 0.804 + ], + "angle": 0, + "content": "where \\(H\\) and \\(W\\) denote the height and width of an image. The \\(P_{ij}(\\tilde{M}_{ij}) = \\mathrm{Softmax}(M_{ij}^D)_{\\tilde{M}_{ij}}\\), which means the predicted probability with respect to the pseudo label \\(\\tilde{M}_{ij}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.805, + 0.892, + 0.878 + ], + "angle": 0, + "content": "Feature-guided loss. After obtaining features \\( Z \\) from visual backbone of rectified CLIP and \\( \\hat{Z} \\) from student visual backbone, we resize two features to the same shape of original image by bilinear interpolation. We then perform a L1 loss between two features" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.884, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t}} = \\left| \\left| Z - \\hat {Z} \\right| \\right| _ {1} \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4106" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "Therefore, visual features from student network is aligned with visual and textual features of rectified CLIP, providing an important basis for the following text-guided loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.212 + ], + "angle": 0, + "content": "Text-guided loss. We adopt the same strategy as the Rectification stage (Sec. 3.4) to compute contrastive loss as our text-guided loss \\(\\mathcal{L}_{text}\\). The only difference is that we use feature \\(\\hat{Z}\\) and masks generated by the student network to obtain the corresponding masked features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.468, + 0.258 + ], + "angle": 0, + "content": "Total distillation loss. Total distillation loss \\(\\mathcal{L}_{distill}\\) is calculated as follows, where \\(\\alpha\\) and \\(\\beta\\) are both constants and set to 0.5. Effect of each loss term is studied in Sec. 4.3." + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.268, + 0.468, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {f e a t}} + \\beta \\mathcal {L} _ {\\text {t e x t}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.294, + 0.21, + 0.312 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.319, + 0.158, + 0.335 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.468, + 0.492 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on three standard benchmarks for semantic segmentation, including PASCAL VOC 2012 [18], PASCAL Context [40] and ADE20K [61]. PASCAL VOC 2012 (1,464/1,449 train/validation) contains 20 object classes, while PASCAL Context (4,998/5,105 train/validation) is an extension of PASCAL VOC 2010 and we consider 59 most common classes in our experiments. ADE20K (20,210/2,000 train/validation) is a segmentation dataset with various scenes and 150 most common categories are considered." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Implementation details. For the image encoder of CLIP, we adopt ViT-B/16 as visual backbone. For the text encoder of CLIP, we adopt Transformer [52]. During the whole training period, we keep both of the encoders frozen. We use conventional data augmentations including random cropping and random flipping. Relevant hyper-parameters for each dataset, including number of rectification epochs, number of distillation iterations and parameters of data augmentation are shown in our supplementary material. For both stages, we use a SGD [48] optimizer with a learning rate of 0.01 and a weight decay of 0.0005. We adopt the poly strategy with the power of 0.9 for learning rate. In our experiment, we report the mean intersection over union (mIoU) as evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare with previous USS methods to verify the superiority of our method, including MaskCLIP \\((+)\\) [62], CLIP-S4 [22], ReCo [49], CLIPpy [46], etc. From the technical view, although our method falls into the category of text-guided segmentation, we choose not to compare with text-guided methods which are designed for zero-shot or open-vocabulary setting (e.g., GroupViT [57], TCL [5] and ViewCo [47]). Two main differences exist between the zero-shot/open-vocabulary setting and the unsupervised setting: 1) in zero-shot/open-vocabulary, the category sets for training and inference are non-overlapped, but in USS the category sets are shared; 2) text-guided zero-shot/open-vocabulary methods usually rely on large-scale" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison with non-language-guided unsupervised semantic segmentation methods on PASCAL VOC." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.129, + 0.847, + 0.303 + ], + "angle": 0, + "content": "
ModelLCmIoU
IIC [25]✓9.8
SegSort [23]✓11.7
Deep Spectral Methods [39]✗37.2
HSG [29]✗41.9
ACSeg [34]✗47.1
MaskContrast [51]✓49.6
Ours (rectification)✗58.5
Ours (distill)✗75.4
Ours (distill)✓76.1
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.308, + 0.892, + 0.354 + ], + "angle": 0, + "content": "image-text pairs or class-agnostic masks to supervise the training, but in USS we only utilize the images without any types of annotations to train the model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.363, + 0.791, + 0.38 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.493 + ], + "angle": 0, + "content": "We compare our method with both previous non-language-guided USS methods (Table 1) and CLIP-based USS methods (Table 2). In all tables, we use \"Ours (rectification)\" to denote the segmentation results after the rectification stage and \"Ours (distill)\" to denote the segmentation results after distillation. Unless otherwise stated, results from previous methods are directly cited from the original papers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "As shown in Table 1, our model shows remarkably better segmentation results compared with the conventional non-language-guided USS methods. In order to further evaluate the strength of features extracted from our distillation model, we also report our results of linear classification (LC). Following MaskContrast [51], we fix our model to generate pixel embeddings and train a linear classifier to generate semantic segmentation masks. The results of LC also prove that our method extracts strong features and achieves significant gains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Table 2, we also make a comparison with typical CLIP-based methods, including MaskCLIP, CLIPpy, ReCo and CLIP-S4. These methods all consider the same CLIP-based USS setting with ours. The results show our method performs favorably against previous CLIP-based methods. For example, after the rectification stage, our method outperforms MaskCLIP by \\(9.0\\%\\), \\(4.1\\%\\) and \\(1.6\\%\\) respectively on the three datasets, while after distillation, our method outperforms \\(\\mathrm{MaskCLIP + }\\) by \\(5.4\\%\\), \\(2.7\\%\\) and \\(2.1\\%\\). Our method also shows better results than CLIPpy by \\(20.8\\%\\) and \\(0.8\\%\\) on PASCAL VOC and ADE20K and outperforms CLIP-S4 by \\(3.4\\%\\) and \\(0.2\\%\\) on PASCAL VOC and PASCAL Context. Since ReCo employs a \"context elimination\" (CE) trick which may introduce prior knowledge, we also report the results of ReCo by removing this trick (ReCo w/o CE in Table 2). Our method outperforms ReCo obviously, e.g., on PASCAL VOC, ours (rectification) outper" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4107" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison with CLIP-based unsupervised semantic segmentation methods on various benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.129, + 0.434, + 0.289 + ], + "angle": 0, + "content": "
MethodVOCContextADE
CLIPpy [46]54.6/13.5
MaskCLIP [62]49.521.79.5
MaskCLIP+ [62]70.031.112.2
ReCo [49]55.226.2/
ReCo (w/o CE)54.823.1/
CLIP-S4 [22]72.033.6/
Ours (rectification)58.525.811.1
Ours (distill)75.433.814.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.297, + 0.47, + 0.34 + ], + "angle": 0, + "content": "Table 3. Ablation on whether Query should be learnable. Results show that fixing Query performs favorably against making Query learnable." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.35, + 0.434, + 0.412 + ], + "angle": 0, + "content": "
QueryReferenceVOCContextADE
✓✓56.723.79.4
X✓56.724.410.5
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.416, + 0.468, + 0.446 + ], + "angle": 0, + "content": "forms ReCo and ReCo (w/o CE) by \\(3.3\\%\\) and \\(3.7\\%\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.455, + 0.228, + 0.471 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.612 + ], + "angle": 0, + "content": "Should Query be learnable? We study on whether the prompt of Query should be learnable. As shown in Table 3, we conduct experiments with learnable and fixed Query respectively. As we aim to show the effect of fixing Query, we don't utilize the projected positional logits to calculate the numbers shown in the table. The numbers show that fixing Query performs favorably against making Query learnable. This verifies that we should only learn Reference but keep Query fixed as we discuss in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Effect of different bias modeling. As shown in Table 4, compared with the baseline without any bias rectification (numbers with only \"Query\"), utilizing learable Reference to model the class-preference bias brings remarkable performance improvement, e.g., \\(7.2\\%\\) mIoU on PASCAL VOC, \\(2.7\\%\\) mIoU on PASCAL Context and \\(1.0\\%\\) mIoU on ADE20K. While introducing projection of positional embedding to model the space-preference bias, the numbers are further improved, i.e., around \\(1\\%\\) better than modeling the class-preference bias only. Those results verify the effectiveness of bias modeling by introducing learnable Reference and projection of positional embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of element-wise subtraction on bias rectification. In order to validate the effect of our element-wise subtraction, we conduct experiments in Table 5. We compare our subtraction mechanism with an alternative solution, i.e., instead of subtracting logits encoding bias (i.e., Reference logits plus positional logits), we add all the logits. Comparisons shown in the table verify the effectiveness of our" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 4. Effect of bias modeling. Results show effect of both bias modeling quantitatively and each contributes to better results." + }, + { + "type": "table", + "bbox": [ + 0.512, + 0.129, + 0.878, + 0.206 + ], + "angle": 0, + "content": "
QueryReferencePEVOCContextADE
✓49.521.79.5
✓✓56.724.410.5
✓✓✓58.525.811.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.22, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 5. Effectiveness of element-wise subtraction. Results show that the element-wise subtraction removes bias from CLIP." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.26, + 0.866, + 0.336 + ], + "angle": 0, + "content": "
QueryRef+PEsubaddVOCContext
✓XXX49.521.7
✓✓X✓57.824.9
✓✓✓X58.525.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.431 + ], + "angle": 0, + "content": "subtraction way. As we aim to remove the bias from the original CLIP, we speculate that the subtraction operation may work as a strong prior which regularizes the training to facilitate meaningful and effective bias modeling. The effect of bias rectification by element-wise subtraction is further visualized in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Visualization of Bias Modeling. In Fig. 3 (b), we illustrate how we explicitly model and rectify the class-preference bias. As shown in the first column of Fig. 3 (b), the original CLIP (see \"Query\" mask) tends to misclassify part of a \"person\" (see the ground-truth mask denoted as \"GT\") into a \"boat\". Such a mistake is reflected in the comparison between the logit heatmaps for the \"person\" channel and the \"boat\" channel: in the area misclassified as \"boat\", the logits for \"boat\" are relatively higher than those for \"person\". In contrast, for Reference logit map, the person-channel logits are quite low, while the boat-channel logits are generally very high, especially for the misclassified area. Consequently, via the proposed logit subtraction operation, we obtain the rectified logits (see the last column), where the boat channel is largely suppressed. Finally, we obtain a much better mask (see \"Ours\" in the first column)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.809 + ], + "angle": 0, + "content": "In Fig. 3 (c), we aim to show that the projection of positional embedding (PE) effectively models the space-preference bias. By comparing the results with and without rectifying logits projected by PE (see the dashed boxes of the last two columns), we find that rectification with PE largely improves the segmentation performance in the boundary areas. The results illustrate that the projection of PE does encode the space-preference bias and rectifying such bias may largely improve the segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Effect of different distillation loss terms. We conduct experiments on PASCAL VOC to validate the effect of each loss term of our distillation framework, including the mask-guided, feature-guided and text-guided loss terms. From Table 7, all the loss terms contribute to better performance." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Effect of segmentation head for distillation. As illustrated" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.496, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.337, + 0.894, + 0.395 + ], + "angle": 0, + "content": "Figure 3. (a) Qualitative Results: We visualize segmentation results on PASCAL Context. From the visualization, we observe that our model outperforms MaskCLIP \\((+)\\) obviously by rectifying both class-preference bias and space-preference bias. (b) Class-preference bias: In order to explain how Reference explicitly models the class-preference bias, we show the heatmap of Reference logits. (c) Space-preference bias: The segmentation within dashed boxes shows the effectiveness of PE projection on rectifying space-preference bias." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.41, + 0.471, + 0.455 + ], + "angle": 0, + "content": "Table 6. Effect of segmentation head for distillation. We perform distillation with both original classification head (Ori. Head) and structure from our rectification stage" + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.464, + 0.453, + 0.554 + ], + "angle": 0, + "content": "
Ori. HeadQueryReferencePE Proj.mIoU
✓XXX73.8
X✓XX74.1
X✓✓X75.4
X✓✓✓75.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.555, + 0.468, + 0.583 + ], + "angle": 0, + "content": "Table 7. Ablation results on distillation loss. According to the results, each loss item contributes to better distillation result." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.594, + 0.388, + 0.671 + ], + "angle": 0, + "content": "
MaskFeatureTextmIoU
✓73.0
✓✓74.1
✓✓✓75.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.886 + ], + "angle": 0, + "content": "in Fig 2 (b), instead of directly using original classification head at the distillation stage, the student network (i.e., DeepLab V2 [6]) inherits and fixes the Query and Reference segmentation heads from our rectification stage, and utilizes logit subtraction mechanism to generate the final segmentation masks. In Table 6, we conduct experiments with original head (\"Ori. Head\") of DeepLab V2 and other types of segmentation heads which are directly inherited from our rectification stage on PASCAL VOC. From the results, we observe that it is better to inherit the components from our rectification stage than adopting the original classification head. However, introducing projection of PE in segmentation head cannot bring further improvement." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. We visualize our segmentation re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.892, + 0.52 + ], + "angle": 0, + "content": "sults in Fig. 3 (a). It can be observed that there exists apparent space-preference bias and class-preference bias in the segmentation results of original CLIP (MaskCLIP) and these bias still cannot be removed even after distillation (MaskCLIP+). However, our model outperforms MaskCLIP(+) obviously by rectifying the bias of CLIP for unsupervised semantic segmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.619, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.591, + 0.892, + 0.833 + ], + "angle": 0, + "content": "In this paper, we propose a new framework for language-guided unsupervised semantic segmentation. We observe bias, including space-preference bias and class-preference bias, exists in CLIP when directly applying CLIP to segmentation task. We propose using additional Reference to learn class-preference bias and projecting positional embedding to represent space-preference bias, and then manage to rectify them by a simple element-wise logit subtraction mechanism. For further improving the segmentation performance, we distill the knowledge from rectified CLIP to advanced segmentation backbone with specifically designed losses. Extensive experiments demonstrate that our method achieve superior segmentation performance compared to previous state-of-the-arts. We hope our work may inspire future research to investigate how to better adapt CLIP to complex visual understanding tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement This research project is supported by National Natural Science Foundation (NNSF) of China under Grant 92370114 and National Key R&D Program of China under Grant 2022YFB3104900." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.472, + 0.415 + ], + "angle": 0, + "content": "[5] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.471, + 0.497 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2018. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Mickaël Chen, Thierry Artières, and Ludovic Denoyer. Unsupervised object segmentation by redrawing. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.601 + ], + "angle": 0, + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European Conference on Computer Vision, pages 104-120, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.66, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.731, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[11] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.802, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[12] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[13] Zhijie Deng and Yucen Luo. Learning neural eigenfunctions for unsupervised semantic segmentation. arXiv preprint arXiv:2304.02841, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[14] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[15] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[16] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.893, + 0.483 + ], + "angle": 0, + "content": "[19] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 540-557. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.893, + 0.538 + ], + "angle": 0, + "content": "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.607 + ], + "angle": 0, + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.893, + 0.677 + ], + "angle": 0, + "content": "[22] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.759 + ], + "angle": 0, + "content": "[23] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7334-7344, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[26] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[27] Guoliang Kang, Lu Jiang, Yunchao Wei, Yi Yang, and Alexander Hauptmann. Contrastive adaptation network for single-and multi-source domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 44(4): 1793-1804, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.471, + 0.317 + ], + "angle": 0, + "content": "[28] Guoliang Kang, Yunchao Wei, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Pixel-level cycle association: A new perspective for domain adaptive semantic segmentation. Advances in neural information processing systems, 33: 3569-3580, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[29] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571–2581, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.445 + ], + "angle": 0, + "content": "[30] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[31] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.587 + ], + "angle": 0, + "content": "[32] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI conference on artificial intelligence, pages 11336-11344, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.47, + 0.66 + ], + "angle": 0, + "content": "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[34] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.816 + ], + "angle": 0, + "content": "[35] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[36] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[37] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, et al. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[38] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European Conference on Computer Vision, pages 529-544. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[42] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 142-158. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[43] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[44] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1713-1721, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[46] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. arXiv preprint arXiv:2210.09996, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[47] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4111" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[49] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In Advances in Neural Information Processing Systems. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[50] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.288 + ], + "angle": 0, + "content": "[51] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[53] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.471, + 0.469 + ], + "angle": 0, + "content": "[54] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. arXiv preprint arXiv:2303.11749, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.471, + 0.539 + ], + "angle": 0, + "content": "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.621 + ], + "angle": 0, + "content": "[56] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. C2am: contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989–998, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.471, + 0.704 + ], + "angle": 0, + "content": "[57] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 3, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[58] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[59] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[60] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[61] B. Zhou, Z. Hang, Francesco Xavier Puig Fernandez, S. Fidler, and A. Torralba. Scene parsing through ade20k dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.218 + ], + "angle": 0, + "content": "[62] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.893, + 0.275 + ], + "angle": 0, + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4112" + } + ] +] \ No newline at end of file diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_origin.pdf b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ebf57a4549ac56a096a0a2e7a7f2c91bcee4cb07 --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/4af240bb-fb73-4b13-b278-60a5b44fd376_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b53af70d8efc9910b7d6f865243916f5c07781283f9b5ae9562b320916b6b1c4 +size 2261890 diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/full.md b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..16a25a4a95873cf8f3d37eb7b2693b6c3c2f097b --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/full.md @@ -0,0 +1,341 @@ +# Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation + +Jingyun Wang +Beihang University + +19231136@buaa.edu.cn + +Guoliang Kang * +Beihang University, Zhongguancun Laboratory + +kgl.prml@gmail.com + +# Abstract + +Recent works utilize CLIP to perform the challenging unsupervised semantic segmentation task where only images without annotations are available. However, we observe that when adopting CLIP to such a pixel-level understanding task, unexpected bias occurs. Previous works don't explicitly model such bias, which largely constrains the segmentation performance. In this paper, we propose to explicitly model and rectify the bias existing in CLIP to facilitate the unsupervised semantic segmentation. Specifically, we design a learnable "Reference" prompt to encode class-preference bias and project the positional embedding of vision transformer to represent space-preference bias. Via a simple element-wise subtraction, we rectify the logits of CLIP classifier. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then a contrastive loss between masked visual feature and the text features of different classes is imposed to facilitate the effective bias modeling. To further improve the segmentation, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture via minimizing our designed mask-guided, feature-guided and text-guided loss terms. Extensive experiments on standard benchmarks demonstrate that our method performs favorably against previous state-of-the-arts. The implementation is available at https://github.com/dogehhh/ReLUCP. + +# 1. Introduction + +Semantic segmentation aims to attach a semantic label to each pixel of an image. Since the rising of deep learning [27, 28, 45, 54], semantic segmentation has been widely adopted in real-world applications, e.g., autonomous driving, medical image segmentation, etc. Conventional approaches [6, 9, 10, 37, 38, 53, 60] for semantic segmentation have achieved remarkable performance. However, the superior performance of those methods relies heavily on large amounts of fully annotated masks. Collecting such high + +quality pixel-level annotations can be both time-consuming and expensive, e.g., some annotations for specialized tasks require massive expert knowledge, some are even inaccessible due to privacy reasons, etc. Therefore, it is necessary to explore unsupervised semantic segmentation where only images without annotations are available. + +Unsupervised semantic segmentation (USS) has been studied for years. Many non-language-guided USS methods have been proposed, e.g., clustering-based methods [11, 25, 29, 34, 42], contrastive-learning-based methods [21, 51], boundary-based methods [23], etc. Despite promising progress achieved, there still exhibits a large performance gap between USS and the supervised segmentation methods. Besides, these methods typically obtain class-agnostic masks and have to depend on additional processing (e.g., Hungarian matching) to assign semantic labels to the masks, rendering them less practical in real scenarios. + +Recently, large-scale visual-language pre-trained models, e.g., CLIP [45], demonstrate superior zero-shot classification performance by comparing the degree of alignment between image feature and text features of different categories. A few CLIP-based USS approaches [22, 46, 49, 62] emerge and show remarkable performance improvement compared with the non-language-guided USS methods. These models require no access to any types of annotations, and directly assign a label to each pixel, benefiting from the aligned vision and text embedding space of CLIP. However, good alignment between image-level visual feature and textural feature doesn't necessarily mean good alignment between pixel-level visual feature and textural feature. Thus, for CLIP, unexpected bias may inevitably appear. Previous works didn't explicitly model such bias, which may largely constrain their segmentation performance. + +We observe two kinds of bias existing in CLIP. From one aspect, as shown in Fig. 1(a), CLIP exhibits space-preference bias. CLIP performs apparently better for segmenting central objects than the ones distributed near the image boundary. It can be reflected by the fact that mIoU decreases as the distance between the centroids of object and the image increases. From the other aspect, as shown in Fig. 1(b), there exists class-preference bias between similar + +![](images/2575f8c34e09cc48afa5a280e80590666f01f27d3f00618934a9264dac09ada8.jpg) + +![](images/7eef6c1530938f32fb3538d7ea98bbfac462f22e9913064392a5ad1dcc130a33.jpg) + +![](images/c53632a59c17113761d19526488bf5af2b07fa34e63fe3c259912d58e6918ab7.jpg) + +![](images/6bde63d342e6c7cf94b87467dbbd8374044b1eaacbbb4601078c8f00e39d141a.jpg) + +![](images/351504f4a6cc247c3417f7c3100894262dfd94b013516e9f264df8f30cc2cb39.jpg) + +![](images/f7b4dce1ae444e3cf1495e5c3bbba378cd07204a38b923cd6b115b3f243963bc.jpg) + +![](images/aa747899bbf73067470b5d0090df55432d29e6021d1647db492039cdd43cee5b.jpg) + +![](images/93451363de034c4fb6fb3c28886744a2de50e6592136a88ead8875e46dfe7a3f.jpg) + +![](images/3b4e2e2eb6f5c95b077e839363f3360f83be3f79bc4a9f0e6dc8863c2b59a8c3.jpg) + +![](images/7cb6556268707b46e596e279189ea6b1bf6f5bef8224dfc978eba5bbe6d8e409.jpg) +Figure 1. (a) Space-preference bias. (Left): The relationship between distance $(x$ -axis) and mIoU $(y$ -axis) is drawn on PASCAL VOC [18]. The distance means the spatial distance between the centroids of the object and the image and mIoU is computed based on predictions and ground truth. The curve shows that CLIP [45] (green) is apparently better at segmentation for central objects than boundary ones, but our method (blue) effectively mitigates this bias. More details about how we draw this figure has been shown in our supplementary material. (Right): Visualizations also show our improvement on space-preference bias qualitatively. (b) Class-preference bias. (Left): We randomly select 6 classes from PASCAL Context [40] and draw the confusion matrix of CLIP and our model. It shows that beside the ground truth, CLIP also prefers to assign an incorrect but relevant label to a pixel in quite a few cases, while our results show apparent improvement. (Right): The visualizations are consistent with what we observed in confusion matrix. For example, for a "cow", CLIP tends to classify it as "dog" incorrectly. + +categories in CLIP. For example, according to visualization (right), when ground truth is "cow", CLIP tends to incorrectly classify it as "dog". We also show such trend between randomly selected classes by confusion matrix (left). Elements on diagonal line represent right classification, while others are false. We observe a wide range of class-preference bias introduced by CLIP. + +In this paper, we propose to explicitly model and rectify the bias of CLIP to facilitate the USS. Specifically, we design two kinds of text inputs for each class, which are named as "Reference" and "Query" respectively. The Query is manually designed while the Reference contains learnable prompts. We adopt the text features of Query and Reference as classifiers to generate the Query and Reference logits respectively for each pixel of image. The Query logits represent the segmentation ability of original CLIP, while the Reference logits are expected to reflect the bias of CLIP preferring a specific class. Additionally, we project the positional embedding of CLIP's vision transformer to generate positional logits for each image. We expect the positional logits to represent space-preference bias of CLIP. Then, we remove the class-preference and the space-preference bias from original CLIP via a logit-subtraction mechanism, i.e., subtracting the Reference logits and the positional logits from the Query logits. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then the contrastive loss between masked visual + +feature and the text features of different categories is imposed to facilitate the effective bias modeling and rectification. To further improve the segmentation performance, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture, via mask-guided distillation, feature-guided distillation and text-guided learning. + +We conduct extensive experiments on standard semantic segmentation benchmarks, including PASCAL VOC [18], PASCAL Context [40] and ADE20K [61]. Experiment results demonstrate that our method performs favourably against previous state-of-the-arts. Notably, on PASCAL VOC, our method outperforms CLIP S4 [22] by $3.4\%$ . Extensive ablation studies verify the effectiveness of each design in our framework. + +Our contributions are summarized as follows. + +- We observe that when applying CLIP to pixel-level understanding tasks, unexpected bias including space-preference bias and class-preference bias, occurs. Such bias may largely constrain the segmentation performance of CLIP-based segmentation models. +- We propose to explicitly model the class-preference and space-preference bias of CLIP via learnable Reference text inputs and projection of positional embedding. Through a simple logit-subtraction mechanism and the contrastive loss built on masked features, we effectively rectify the bias of CLIP. +- We conduct extensive experiments on segmentation + +benchmarks under the unsupervised setting. Experiment results show superior performance of our method to previous state-of-the-arts. + +# 2. Related Work + +Pre-trained vision-language models. Pre-trained vision-language models (VLMs) [8, 14, 32, 33, 35] have developed rapidly with the help of large-scale image-text pairs available on the Internet. Recently, CLIP [45], ALIGN [26] and Slip [41] have made great progress on learning visual and textual representations jointly by using contrastive learning. With the image-level alignment with text, pre-trained VLMs have strong ability for zero-shot classification task and can be transferred to various downstream tasks, such as object detection [17, 54] and semantic segmentation [62]. + +Unsupervised semantic segmentation. While conventional approaches of semantic segmentation [37, 38, 53, 60] rely on pixel-level annotations and weakly-supervised methods [1, 16, 44, 56] still ask for image-level labels, unsupervised semantic segmentation (USS) methods [11, 20, 42, 50, 51] explores to train a segmentation model without any annotations. Models like [2, 7] adopt generative model [12] to separate foreground with background or generate corresponding masks. SegSort [23], HSG[29] and ACSeg [34] use clustering strategy, while IIC [25] uses mutual information maximization to perform unsupervised learning. MaskContrast [51] introduces contrastive learning into USS. Others like DSM [39] and LNE [13] exploit features extracted from self-supervised model, and combine it with spectral graph theory. However, the methods mentioned above either fail to segment images with multi-category objects or show a large performance gap with the supervised methods. Besides, they can only obtain class-agnostic masks and have to depend on additional strategy, such as Hungarian-matching algorithm [30], to match corresponding category with masks. Recently, pre-trained vision-language models are adopted in USS. MaskCLIP [62] modifies the image encoder of CLIP to generate patch-level features and directly performs segmentation with text features as classifiers. CLIP-py [46] performs contrastive learning between visual features from self-supervised ViT [4] and text features from CLIP. ReCo [49] performs image retrieval with CLIP and extracts class-wise embedding as classifier with co-segmentation. CLIP-S4 [22] learns pixel embeddings with pixel-segment contrastive learning and aligns such embeddings with CLIP in terms of embedding and semantic consistency. These methods can directly assign a category to each pixel, whose setting is named as language-guided unsupervised semantic segmentation. However, directly applying CLIP may bring prior bias, including space-preference bias and class-preference bias. As we know, there is no previous method trying to solve these bias and we manage to explicitly model the bias, and rectify them by + +element-wise subtraction. + +Language-guided semantic segmentation. Recently, many works are exploring semantic segmentation guided by language under different settings. Zero-shot works [3, 31, 43, 55] split classes into seen and unseen set. During the training period, only masks of seen classes are provided. For inference, models are tested on both seen and unseen classes, but the test data is still in the same distribution with the training data. Open-vocabulary works [5, 19, 36, 47, 57-59] are trained in one scenario with extra annotations including class-agnostic masks or image captions, but are used for predicating segmentation masks of novel classes in other scenarios. From the technical view, our method also falls into the category of language-guided semantic segmentation. However, we consider the unsupervised setting. In this setting, we have access to images without any annotations during training. The training and inference images are sampled from the same distributions and the same set of categories. Such a setting is different to the typical zero-shot or open-vocabulary setting. + +# 3. Method + +Background In this work, we aim to rectify the bias of CLIP for unsupervised semantic segmentation. In USS, we only have access to images without any types of annotations to train the segmentation model. For training and inference, the same set of categories are considered and the data distributions are assumed to be the same. + +Overview The general framework of our method is illustrated in Fig. 2. We aim to rectify the bias of CLIP including the class-preference bias and the space-preference bias, to facilitate unsupervised semantic segmentation. From a high level, class-preference bias reflects the shift of CLIP predictions towards specific classes, while space-preference bias reflects the shift of CLIP predictions towards specific spatial locations. Both biases will be finally reflected in the predicted logit maps. A reasonable way to rectify the bias is to subtract its logit maps from the normal logit maps predicted via original CLIP. + +To realize our goal, we firstly forward the image $I \in \mathbb{R}^{3 \times H \times W}$ through the image encoder of CLIP to obtain the patch-level image features $Z$ . We design two kinds of text inputs for each class. One is the manually designed Query text input $Q$ and the other one is the Reference text input $R$ which consists of a learnable prompt and the name of class. For each class, passing $Q$ and $R$ through the text encoder of CLIP, we obtain two text embeddings $W_{q}$ and $W_{r}$ , which serve as the weights of query segmentation head and reference segmentation head respectively. Taking the same visual feature $Z$ as input, the query and reference segmentation heads output a query logit map $M_{q}$ (Sec. 3.1) and Reference logit map $M_{r}$ (Sec. 3.2) respectively. Meanwhile, positional embedding $p$ is sent into a + +learnable convolutional network to generate positional logit map $M_{p}$ (Sec. 3.3). Then the bias logit map $M_{b}$ can be obtained by adding Reference logit map $M_{r}$ and positional logit map $M_{p}$ . We then subtract $M_{b}$ from $M_{q}$ to generate the rectified logits $M$ . Based on the rectified logits, we generate masks for different classes. Then a contrastive loss is imposed between the masked visual features and the text features of different categories to encourage the bias modeling and rectification (Sec. 3.4). In order to enhance the segmentation performance, we distill the knowledge of rectified CLIP to the advanced segmentation architecture with specifically designed mask-guided, feature-guided and text-guided loss terms (Sec. 3.5). In both the rectification and distillation stages, we keep CLIP frozen. + +# 3.1. Baseline: Directly Segment with CLIP + +Following MaskCLIP [62], we adapt pre-trained CLIP [45] (ViT-B/16) to the semantic segmentation task. We remove the query and key embedding layers of last attention but reformulate the value embedding layer and the last linear projection layer into two respective $1 \times 1$ convolutional layers. Therefore, the image encoder can not only generate local features for dense predictions, but also keep the visual-language association in CLIP by freezing its pre-trained weights. We forward image $I$ through the image encoder and obtain patch-level features $Z \in \mathbb{R}^{n \times D}$ ( $n$ is the number of patches and $D$ is the dimension of features in CLIP). + +Each text in Query $Q = \{Q_{1}, Q_{2}, \dots, Q_{C}\}$ ( $C$ is the number of classes) is an ensemble of several manually designed prompts, e.g., "a good/large/bad photo of a [CLS]" where [CLS] denotes the a specific class name. Passing $Q$ through the text encoder, we obtain its text embeddings $W_{q}$ . We treat text embeddings $W_{q}$ as the weight of segmentation head to perform $1 \times 1$ convolution. By sending features $Z$ to the segmentation head, we get a Query logit map $M_{q} \in \mathbb{R}^{n \times C}$ . Then the segmentation mask can be predicted by arg max operation on $M_{q}$ . + +# 3.2. Learn Class-Preference Bias + +In order to explicitly model the class-preference bias brought by pre-trained CLIP for specific datasets, we design a Reference $R_{i}$ , $i = \{1,2,\dots ,C\}$ , as additional text input for each class. Inspired by CoOp [63], $R_{i}$ consists of a learnable prompt which is shared across all the classes for efficiency, and a class name $[CLS]$ , which can be formed as + +$$ +R _ {i} = \left[ v _ {1} \right] \left[ v _ {2} \right] \dots \left[ v _ {l} \right] \dots \left[ v _ {L} \right] [ C L S ], \tag {1} +$$ + +where each $[v_l](l \in \{1, \dots, L\})$ is a vector with dimension $D$ and serves as a word embedding. The $L$ is a constant representing the number of word embeddings to learn. Totally, there are 77 word embeddings for a text of CLIP. As two word embeddings are used for class name and two for indicating start and end of a text, we set $L$ to 73. + +Passing the Reference $R$ through text encoder, we obtain reference text embedding $W_{r}\in \mathbb{R}^{C\times D}$ . The text embedding $W_{r}$ is then directly utilized as the weights of segmentation head to perform $1\times 1$ convolution on visual feature $Z$ and output a Reference logit map $M_{r}\in \mathbb{R}^{n\times C}$ . + +As we obtain different Reference logit maps for different classes, we expect the Reference logit map to encode the class-preference bias to facilitate the following bias rectification process. It is worth noting that we make Reference $R$ learnable but keep Query $Q$ fixed. It is because when we make them both learnable, Query may also capture some class-preference bias and technically we cannot guarantee which text embedding should encode the class-preference bias, resulting in an implicit bias modeling. Thus, in our framework, we choose not to make the $Q$ learnable to just encourage $R$ encode the bias. Such a design cooperates with the following logit subtraction mechanism to make the bias modeling and rectification more meaningful and effective. + +# 3.3. Learn Space-Preference Bias + +In ViT [15], positional embeddings (PE) are important for encoding spatial information to features. Thus, we assume the space-preference bias should depend on the positional embeddings (PE) and choose to learn a projection of PE to represent the space-preference bias. + +We design a 3-layer $3 \times 3$ convolutional network, and each convolutional layer is followed by a batch normalization. Specifically, we project PE $p$ by the designed convolutional network to obtain positional logits $M_p \in \mathbb{R}^{n \times 1}$ . During the training process, the projection network is optimized to model the space-preference bias in positional logits $M_p$ . Since PE is shared across all categories, the learned space-preference bias is also shared across all the categories. + +# 3.4. Rectify Bias with Contrastive Learning Loss + +By keeping Query prompt $Q$ fixed, the Query logits $M_{q}$ represents the natural prediction ability of CLIP model, which may contain class-preference bias and space-preference bias. With learnable Reference prompt and projection of PE, we explicitly encode the class-preference bias and space-preference bias into the Reference logit map $M_{r}$ and positional logit map $M_{p}$ respectively. We then add $M_{r}$ and $M_{p}$ together to depict the final bias $M_{b}$ , i.e., + +$$ +M _ {b} = M _ {r} + M _ {p} ^ {*}, \tag {2} +$$ + +where we simply expand $M_p$ to $M_p^* \in \mathbb{R}^{n \times C}$ which means we repeat each $M_p$ for $C$ times. + +Then we perform a simple element-wise subtraction between $M_q$ and $M_b$ to generate the rectified logit map $M$ , + +$$ +M = M _ {q} - M _ {b}. \tag {3} +$$ + +![](images/9632a2b4208aba1e3dbd642e4b06ea94125ba29c48b529ef06c6b493f60b1344.jpg) +(a) Rectification Stage + +![](images/dacd7504c4a63c683af85660184ef7567e39f64c721250c4e26b64ff8471921b.jpg) +(b) Distillation +Figure 2. Method overview. We propose a new framework for language-guided unsupervised semantic segmentation. (a) Rectification Stage: At this stage, we aim to rectify the class-preference and space-preference bias of CLIP. (b) Distillation: We distill knowledge from the rectified CLIP to the advanced segmentation architecture with the mask-guided, feature-guided and text-guided loss terms. + +From a high-level, this operation can be interpreted as subtracting the "bias" from the predictions with text features of Query as segmentation head. + +Since the arg max operation is not differentiable, we utilize a Gumbel-Softmax trick [24] to generate the candidate masks for each class with $\tau_{1}$ as the temperature, i.e., + +$$ +\hat {M} = \text {G u m b e l - S o f t m a x} (M, \tau_ {1}). \tag {4} +$$ + +We apply $\hat{M}$ to the feature $Z$ to get the masked features $Z_{g}$ , which are expected to encode the features of objects with different classes. We then design a contrastive loss to supervise the learning via aligning masked features with text features of different classes. As the masked features represent the objects of interest without context, we utilize $W_{q}$ generated by text input $Q$ for alignment. + +Additionally, as it is usually impossible for an image to contain all the classes of interest, we need to infer what classes exist in the current image. Our strategy is as follows. Firstly, we choose to extract global visual feature from image encoder of CLIP. Then we calculate the similarity scores between the text input of a specific class and the global visual feature of the image. We only choose the classes whose similarity scores are higher than a threshold $t$ as the potential classes exist in the current image. + +Then we may select $K$ pseudo labels $\{c_1, c_2, \dots, c_K\}$ for each image. We perform a global average pooling of $Z_g$ and compute the similarities between pooled visual features of pseudo labels and text features of all the categories. The contrastive loss can be defined as + +$$ +\mathcal {L} = - \frac {1}{K} \sum_ {k = 1} ^ {K} \log \frac {\exp \left\{S _ {c _ {k} , c _ {k}} / \tau \right\}}{\sum_ {j = 1} ^ {C} \exp \left\{S _ {c _ {k} , j} / \tau \right\}} \tag {5} +$$ + +where $S_{c_k,j}$ denotes the similarity between visual feature of pseudo label $c_k$ and text feature of $j$ -th ( $j \in \{1,2,\dots,C\}$ ) category and the $\tau$ is a constant. + +A better modeling of bias yields more accurate estimations of object masks. Then the masked features of objects + +are more aligned with corresponding text features. As a result, the contrastive loss (Eq. (5)) will be lower. In contrast, a worse modeling of bias results in higher contrastive loss. Thus, minimizing Eq. (5) will drive the model to update towards making more accurate mask predictions, i.e., rectifying the bias of CLIP when adapted to the downstream USS task. + +# 3.5. Distillation for Enhanced Results + +As CLIP is not specifically designed for segmentation tasks, we choose to distill the knowledge from the rectified CLIP (teacher) to the advanced segmentation architecture (student) to enhance the feature representations and finally improve the segmentation performance. In our paper, we choose DeepLab V2 [6] as the student network which directly inherits and fixes the Query and Reference segmentation heads, and utilizes logit subtraction mechanism to generate the final segmentation masks. In our design, the rectified CLIP works as a teacher to guide the feature learning of DeepLab V2 with three designed loss terms. + +Mask-guided loss. We directly exploit segmentation masks generated from the bias rectification stage as pseudo labels. We compute a cross-entropy loss between the pseudo labels $\tilde{M}$ and predictions $M^D$ from DeepLab V2, i.e., + +$$ +\mathcal {L} _ {\text {m a s k}} = - \frac {1}{H W} \sum_ {i} ^ {H} \sum_ {j} ^ {W} \log P _ {i j} (\tilde {M} _ {i j}) \tag {6} +$$ + +where $H$ and $W$ denote the height and width of an image. The $P_{ij}(\tilde{M}_{ij}) = \mathrm{Softmax}(M_{ij}^D)_{\tilde{M}_{ij}}$ , which means the predicted probability with respect to the pseudo label $\tilde{M}_{ij}$ . + +Feature-guided loss. After obtaining features $Z$ from visual backbone of rectified CLIP and $\hat{Z}$ from student visual backbone, we resize two features to the same shape of original image by bilinear interpolation. We then perform a L1 loss between two features + +$$ +\mathcal {L} _ {\text {f e a t}} = \left| \left| Z - \hat {Z} \right| \right| _ {1} \tag {7} +$$ + +Therefore, visual features from student network is aligned with visual and textual features of rectified CLIP, providing an important basis for the following text-guided loss. + +Text-guided loss. We adopt the same strategy as the Rectification stage (Sec. 3.4) to compute contrastive loss as our text-guided loss $\mathcal{L}_{text}$ . The only difference is that we use feature $\hat{Z}$ and masks generated by the student network to obtain the corresponding masked features. + +Total distillation loss. Total distillation loss $\mathcal{L}_{distill}$ is calculated as follows, where $\alpha$ and $\beta$ are both constants and set to 0.5. Effect of each loss term is studied in Sec. 4.3. + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = \mathcal {L} _ {\text {m a s k}} + \alpha \mathcal {L} _ {\text {f e a t}} + \beta \mathcal {L} _ {\text {t e x t}}. \tag {8} +$$ + +# 4. Experiments + +# 4.1. Setup + +Datasets. We conduct experiments on three standard benchmarks for semantic segmentation, including PASCAL VOC 2012 [18], PASCAL Context [40] and ADE20K [61]. PASCAL VOC 2012 (1,464/1,449 train/validation) contains 20 object classes, while PASCAL Context (4,998/5,105 train/validation) is an extension of PASCAL VOC 2010 and we consider 59 most common classes in our experiments. ADE20K (20,210/2,000 train/validation) is a segmentation dataset with various scenes and 150 most common categories are considered. + +Implementation details. For the image encoder of CLIP, we adopt ViT-B/16 as visual backbone. For the text encoder of CLIP, we adopt Transformer [52]. During the whole training period, we keep both of the encoders frozen. We use conventional data augmentations including random cropping and random flipping. Relevant hyper-parameters for each dataset, including number of rectification epochs, number of distillation iterations and parameters of data augmentation are shown in our supplementary material. For both stages, we use a SGD [48] optimizer with a learning rate of 0.01 and a weight decay of 0.0005. We adopt the poly strategy with the power of 0.9 for learning rate. In our experiment, we report the mean intersection over union (mIoU) as evaluation metric. + +Baselines. We compare with previous USS methods to verify the superiority of our method, including MaskCLIP $(+)$ [62], CLIP-S4 [22], ReCo [49], CLIPpy [46], etc. From the technical view, although our method falls into the category of text-guided segmentation, we choose not to compare with text-guided methods which are designed for zero-shot or open-vocabulary setting (e.g., GroupViT [57], TCL [5] and ViewCo [47]). Two main differences exist between the zero-shot/open-vocabulary setting and the unsupervised setting: 1) in zero-shot/open-vocabulary, the category sets for training and inference are non-overlapped, but in USS the category sets are shared; 2) text-guided zero-shot/open-vocabulary methods usually rely on large-scale + +Table 1. Comparison with non-language-guided unsupervised semantic segmentation methods on PASCAL VOC. + +
ModelLCmIoU
IIC [25]✓9.8
SegSort [23]✓11.7
Deep Spectral Methods [39]✗37.2
HSG [29]✗41.9
ACSeg [34]✗47.1
MaskContrast [51]✓49.6
Ours (rectification)✗58.5
Ours (distill)✗75.4
Ours (distill)✓76.1
+ +image-text pairs or class-agnostic masks to supervise the training, but in USS we only utilize the images without any types of annotations to train the model. + +# 4.2. Comparison with SOTA methods + +We compare our method with both previous non-language-guided USS methods (Table 1) and CLIP-based USS methods (Table 2). In all tables, we use "Ours (rectification)" to denote the segmentation results after the rectification stage and "Ours (distill)" to denote the segmentation results after distillation. Unless otherwise stated, results from previous methods are directly cited from the original papers. + +As shown in Table 1, our model shows remarkably better segmentation results compared with the conventional non-language-guided USS methods. In order to further evaluate the strength of features extracted from our distillation model, we also report our results of linear classification (LC). Following MaskContrast [51], we fix our model to generate pixel embeddings and train a linear classifier to generate semantic segmentation masks. The results of LC also prove that our method extracts strong features and achieves significant gains. + +In Table 2, we also make a comparison with typical CLIP-based methods, including MaskCLIP, CLIPpy, ReCo and CLIP-S4. These methods all consider the same CLIP-based USS setting with ours. The results show our method performs favorably against previous CLIP-based methods. For example, after the rectification stage, our method outperforms MaskCLIP by $9.0\%$ , $4.1\%$ and $1.6\%$ respectively on the three datasets, while after distillation, our method outperforms $\mathrm{MaskCLIP + }$ by $5.4\%$ , $2.7\%$ and $2.1\%$ . Our method also shows better results than CLIPpy by $20.8\%$ and $0.8\%$ on PASCAL VOC and ADE20K and outperforms CLIP-S4 by $3.4\%$ and $0.2\%$ on PASCAL VOC and PASCAL Context. Since ReCo employs a "context elimination" (CE) trick which may introduce prior knowledge, we also report the results of ReCo by removing this trick (ReCo w/o CE in Table 2). Our method outperforms ReCo obviously, e.g., on PASCAL VOC, ours (rectification) outper + +Table 2. Comparison with CLIP-based unsupervised semantic segmentation methods on various benchmarks. + +
MethodVOCContextADE
CLIPpy [46]54.6/13.5
MaskCLIP [62]49.521.79.5
MaskCLIP+ [62]70.031.112.2
ReCo [49]55.226.2/
ReCo (w/o CE)54.823.1/
CLIP-S4 [22]72.033.6/
Ours (rectification)58.525.811.1
Ours (distill)75.433.814.3
+ +Table 3. Ablation on whether Query should be learnable. Results show that fixing Query performs favorably against making Query learnable. + +
QueryReferenceVOCContextADE
✓✓56.723.79.4
X✓56.724.410.5
+ +forms ReCo and ReCo (w/o CE) by $3.3\%$ and $3.7\%$ respectively. + +# 4.3. Ablation study + +Should Query be learnable? We study on whether the prompt of Query should be learnable. As shown in Table 3, we conduct experiments with learnable and fixed Query respectively. As we aim to show the effect of fixing Query, we don't utilize the projected positional logits to calculate the numbers shown in the table. The numbers show that fixing Query performs favorably against making Query learnable. This verifies that we should only learn Reference but keep Query fixed as we discuss in Sec. 3.2. + +Effect of different bias modeling. As shown in Table 4, compared with the baseline without any bias rectification (numbers with only "Query"), utilizing learable Reference to model the class-preference bias brings remarkable performance improvement, e.g., $7.2\%$ mIoU on PASCAL VOC, $2.7\%$ mIoU on PASCAL Context and $1.0\%$ mIoU on ADE20K. While introducing projection of positional embedding to model the space-preference bias, the numbers are further improved, i.e., around $1\%$ better than modeling the class-preference bias only. Those results verify the effectiveness of bias modeling by introducing learnable Reference and projection of positional embedding. + +Effect of element-wise subtraction on bias rectification. In order to validate the effect of our element-wise subtraction, we conduct experiments in Table 5. We compare our subtraction mechanism with an alternative solution, i.e., instead of subtracting logits encoding bias (i.e., Reference logits plus positional logits), we add all the logits. Comparisons shown in the table verify the effectiveness of our + +Table 4. Effect of bias modeling. Results show effect of both bias modeling quantitatively and each contributes to better results. + +
QueryReferencePEVOCContextADE
✓49.521.79.5
✓✓56.724.410.5
✓✓✓58.525.811.1
+ +Table 5. Effectiveness of element-wise subtraction. Results show that the element-wise subtraction removes bias from CLIP. + +
QueryRef+PEsubaddVOCContext
✓XXX49.521.7
✓✓X✓57.824.9
✓✓✓X58.525.8
+ +subtraction way. As we aim to remove the bias from the original CLIP, we speculate that the subtraction operation may work as a strong prior which regularizes the training to facilitate meaningful and effective bias modeling. The effect of bias rectification by element-wise subtraction is further visualized in Fig. 3. + +Visualization of Bias Modeling. In Fig. 3 (b), we illustrate how we explicitly model and rectify the class-preference bias. As shown in the first column of Fig. 3 (b), the original CLIP (see "Query" mask) tends to misclassify part of a "person" (see the ground-truth mask denoted as "GT") into a "boat". Such a mistake is reflected in the comparison between the logit heatmaps for the "person" channel and the "boat" channel: in the area misclassified as "boat", the logits for "boat" are relatively higher than those for "person". In contrast, for Reference logit map, the person-channel logits are quite low, while the boat-channel logits are generally very high, especially for the misclassified area. Consequently, via the proposed logit subtraction operation, we obtain the rectified logits (see the last column), where the boat channel is largely suppressed. Finally, we obtain a much better mask (see "Ours" in the first column). + +In Fig. 3 (c), we aim to show that the projection of positional embedding (PE) effectively models the space-preference bias. By comparing the results with and without rectifying logits projected by PE (see the dashed boxes of the last two columns), we find that rectification with PE largely improves the segmentation performance in the boundary areas. The results illustrate that the projection of PE does encode the space-preference bias and rectifying such bias may largely improve the segmentation results. + +Effect of different distillation loss terms. We conduct experiments on PASCAL VOC to validate the effect of each loss term of our distillation framework, including the mask-guided, feature-guided and text-guided loss terms. From Table 7, all the loss terms contribute to better performance. + +Effect of segmentation head for distillation. As illustrated + +![](images/2085a7ff846c08481c69541cf64fe3d979f818cb9e131f98ef34b9ad308fd423.jpg) +Figure 3. (a) Qualitative Results: We visualize segmentation results on PASCAL Context. From the visualization, we observe that our model outperforms MaskCLIP $(+)$ obviously by rectifying both class-preference bias and space-preference bias. (b) Class-preference bias: In order to explain how Reference explicitly models the class-preference bias, we show the heatmap of Reference logits. (c) Space-preference bias: The segmentation within dashed boxes shows the effectiveness of PE projection on rectifying space-preference bias. + +![](images/d62c94cdbc5ddb8cba0e249a0cf1658006ca0fc919781fbf04b60085c31c61c7.jpg) + +Table 6. Effect of segmentation head for distillation. We perform distillation with both original classification head (Ori. Head) and structure from our rectification stage + +
Ori. HeadQueryReferencePE Proj.mIoU
✓XXX73.8
X✓XX74.1
X✓✓X75.4
X✓✓✓75.0
+ +Table 7. Ablation results on distillation loss. According to the results, each loss item contributes to better distillation result. + +
MaskFeatureTextmIoU
✓73.0
✓✓74.1
✓✓✓75.4
+ +in Fig 2 (b), instead of directly using original classification head at the distillation stage, the student network (i.e., DeepLab V2 [6]) inherits and fixes the Query and Reference segmentation heads from our rectification stage, and utilizes logit subtraction mechanism to generate the final segmentation masks. In Table 6, we conduct experiments with original head ("Ori. Head") of DeepLab V2 and other types of segmentation heads which are directly inherited from our rectification stage on PASCAL VOC. From the results, we observe that it is better to inherit the components from our rectification stage than adopting the original classification head. However, introducing projection of PE in segmentation head cannot bring further improvement. + +Qualitative Results. We visualize our segmentation re + +sults in Fig. 3 (a). It can be observed that there exists apparent space-preference bias and class-preference bias in the segmentation results of original CLIP (MaskCLIP) and these bias still cannot be removed even after distillation (MaskCLIP+). However, our model outperforms MaskCLIP(+) obviously by rectifying the bias of CLIP for unsupervised semantic segmentation. + +# 5. Conclusion + +In this paper, we propose a new framework for language-guided unsupervised semantic segmentation. We observe bias, including space-preference bias and class-preference bias, exists in CLIP when directly applying CLIP to segmentation task. We propose using additional Reference to learn class-preference bias and projecting positional embedding to represent space-preference bias, and then manage to rectify them by a simple element-wise logit subtraction mechanism. For further improving the segmentation performance, we distill the knowledge from rectified CLIP to advanced segmentation backbone with specifically designed losses. Extensive experiments demonstrate that our method achieve superior segmentation performance compared to previous state-of-the-arts. We hope our work may inspire future research to investigate how to better adapt CLIP to complex visual understanding tasks. + +Acknowledgement This research project is supported by National Natural Science Foundation (NNSF) of China under Grant 92370114 and National Key R&D Program of China under Grant 2022YFB3104900. + +# References + +[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 3 +[2] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 3 +[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[5] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 6, 1 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2018. 1, 5, 8 +[7] Mickaël Chen, Thierry Artières, and Ludovic Denoyer. Unsupervised object segmentation by redrawing. Advances in neural information processing systems, 32, 2019. 3 +[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European Conference on Computer Vision, pages 104-120, 2020. 3 +[9] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 1 +[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1 +[11] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 3 +[12] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 3 +[13] Zhijie Deng and Yucen Luo. Learning neural eigenfunctions for unsupervised semantic segmentation. arXiv preprint arXiv:2304.02841, 2023. 3 + +[14] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3 +[15] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4 +[16] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 3 +[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3 +[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 6 +[19] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 540-557. Springer, 2022. 3 +[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. 3 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1 +[22] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2023. 1, 2, 3, 6, 7 +[23] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7334-7344, 2019. 1, 3, 6 +[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings, 2017. 5 +[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 1, 3, 6 + +[26] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3 +[27] Guoliang Kang, Lu Jiang, Yunchao Wei, Yi Yang, and Alexander Hauptmann. Contrastive adaptation network for single-and multi-source domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 44(4): 1793-1804, 2020. 1 +[28] Guoliang Kang, Yunchao Wei, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Pixel-level cycle association: A new perspective for domain adaptive semantic segmentation. Advances in neural information processing systems, 33: 3569-3580, 2020. 1 +[29] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571–2581, 2022. 1, 3, 6 +[30] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3 +[31] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 3 +[32] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI conference on artificial intelligence, pages 11336-11344, 2020. 3 +[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3 +[34] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 1, 3, 6 +[35] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 3 +[36] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 3 + +[37] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, et al. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1, 3 +[38] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3 +[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3, 6 +[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 2, 6 +[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European Conference on Computer Vision, pages 529-544. Springer, 2022. 3 +[42] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 142-158. Springer, 2020. 1, 3 +[43] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3 +[44] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1713-1721, 2015. 3 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4 +[46] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. arXiv preprint arXiv:2210.09996, 2022. 1, 3, 6, 7 +[47] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 3, 6 +[48] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 6 + +[49] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In Advances in Neural Information Processing Systems. 1, 3, 6, 7 +[50] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 3 +[51] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 3, 6 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6 +[53] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 1, 3 +[54] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. arXiv preprint arXiv:2303.11749, 2023. 1, 3 +[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3 +[56] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. C2am: contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989–998, 2022. 3 +[57] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 3, 6, 1 +[58] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023. +[59] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 3 +[60] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1, 3 + +[61] B. Zhou, Z. Hang, Francesco Xavier Puig Fernandez, S. Fidler, and A. Torralba. Scene parsing through ade20k dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6 +[62] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 1, 3, 4, 6, 7 +[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4 \ No newline at end of file diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/images.zip b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e8a07bc8edb5fb34f74de04cebf8dfa4457df4b2 --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b05550f5f783549a5f3c6cd08f6aae3eda3151c0367376539fa32907e65e8308 +size 504293 diff --git a/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/layout.json b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e18ea006a2ee29fae90a0d06d4b2e8a296ea0a04 --- /dev/null +++ b/2024/Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation/layout.json @@ -0,0 +1,9736 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 531, + 121 + ], + "type": "text", + "content": "Learn to Rectify the Bias of CLIP for Unsupervised Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 144, + 215, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 144, + 215, + 171 + ], + "spans": [ + { + "bbox": [ + 113, + 144, + 215, + 171 + ], + "type": "text", + "content": "Jingyun Wang \nBeihang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 174, + 221, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 174, + 221, + 184 + ], + "spans": [ + { + "bbox": [ + 113, + 174, + 221, + 184 + ], + "type": "text", + "content": "19231136@buaa.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 253, + 144, + 482, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 144, + 482, + 171 + ], + "spans": [ + { + "bbox": [ + 253, + 144, + 482, + 171 + ], + "type": "text", + "content": "Guoliang Kang * \nBeihang University, Zhongguancun Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 174, + 416, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 416, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 416, + 185 + ], + "type": "text", + "content": "kgl.prml@gmail.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 539 + ], + "type": "text", + "content": "Recent works utilize CLIP to perform the challenging unsupervised semantic segmentation task where only images without annotations are available. However, we observe that when adopting CLIP to such a pixel-level understanding task, unexpected bias occurs. Previous works don't explicitly model such bias, which largely constrains the segmentation performance. In this paper, we propose to explicitly model and rectify the bias existing in CLIP to facilitate the unsupervised semantic segmentation. Specifically, we design a learnable \"Reference\" prompt to encode class-preference bias and project the positional embedding of vision transformer to represent space-preference bias. Via a simple element-wise subtraction, we rectify the logits of CLIP classifier. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then a contrastive loss between masked visual feature and the text features of different classes is imposed to facilitate the effective bias modeling. To further improve the segmentation, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture via minimizing our designed mask-guided, feature-guided and text-guided loss terms. Extensive experiments on standard benchmarks demonstrate that our method performs favorably against previous state-of-the-arts. The implementation is available at https://github.com/dogehhh/ReLUCP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 128, + 575 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 287, + 693 + ], + "type": "text", + "content": "Semantic segmentation aims to attach a semantic label to each pixel of an image. Since the rising of deep learning [27, 28, 45, 54], semantic segmentation has been widely adopted in real-world applications, e.g., autonomous driving, medical image segmentation, etc. Conventional approaches [6, 9, 10, 37, 38, 53, 60] for semantic segmentation have achieved remarkable performance. However, the superior performance of those methods relies heavily on large amounts of fully annotated masks. Collecting such high" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 286 + ], + "type": "text", + "content": "quality pixel-level annotations can be both time-consuming and expensive, e.g., some annotations for specialized tasks require massive expert knowledge, some are even inaccessible due to privacy reasons, etc. Therefore, it is necessary to explore unsupervised semantic segmentation where only images without annotations are available." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "type": "text", + "content": "Unsupervised semantic segmentation (USS) has been studied for years. Many non-language-guided USS methods have been proposed, e.g., clustering-based methods [11, 25, 29, 34, 42], contrastive-learning-based methods [21, 51], boundary-based methods [23], etc. Despite promising progress achieved, there still exhibits a large performance gap between USS and the supervised segmentation methods. Besides, these methods typically obtain class-agnostic masks and have to depend on additional processing (e.g., Hungarian matching) to assign semantic labels to the masks, rendering them less practical in real scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 548, + 615 + ], + "type": "text", + "content": "Recently, large-scale visual-language pre-trained models, e.g., CLIP [45], demonstrate superior zero-shot classification performance by comparing the degree of alignment between image feature and text features of different categories. A few CLIP-based USS approaches [22, 46, 49, 62] emerge and show remarkable performance improvement compared with the non-language-guided USS methods. These models require no access to any types of annotations, and directly assign a label to each pixel, benefiting from the aligned vision and text embedding space of CLIP. However, good alignment between image-level visual feature and textural feature doesn't necessarily mean good alignment between pixel-level visual feature and textural feature. Thus, for CLIP, unexpected bias may inevitably appear. Previous works didn't explicitly model such bias, which may largely constrain their segmentation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "We observe two kinds of bias existing in CLIP. From one aspect, as shown in Fig. 1(a), CLIP exhibits space-preference bias. CLIP performs apparently better for segmenting central objects than the ones distributed near the image boundary. It can be reflected by the fact that mIoU decreases as the distance between the centroids of object and the image increases. From the other aspect, as shown in Fig. 1(b), there exists class-preference bias between similar" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4102" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 71, + 265, + 178 + ], + "blocks": [ + { + "bbox": [ + 53, + 71, + 265, + 178 + ], + "lines": [ + { + "bbox": [ + 53, + 71, + 265, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 71, + 265, + 178 + ], + "type": "image", + "image_path": "2575f8c34e09cc48afa5a280e80590666f01f27d3f00618934a9264dac09ada8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 269, + 72, + 327, + 125 + ], + "blocks": [ + { + "bbox": [ + 269, + 72, + 327, + 125 + ], + "lines": [ + { + "bbox": [ + 269, + 72, + 327, + 125 + ], + "spans": [ + { + "bbox": [ + 269, + 72, + 327, + 125 + ], + "type": "image", + "image_path": "7eef6c1530938f32fb3538d7ea98bbfac462f22e9913064392a5ad1dcc130a33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 330, + 72, + 388, + 125 + ], + "blocks": [ + { + "bbox": [ + 330, + 72, + 388, + 125 + ], + "lines": [ + { + "bbox": [ + 330, + 72, + 388, + 125 + ], + "spans": [ + { + "bbox": [ + 330, + 72, + 388, + 125 + ], + "type": "image", + "image_path": "c53632a59c17113761d19526488bf5af2b07fa34e63fe3c259912d58e6918ab7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 445, + 125 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 445, + 125 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 445, + 125 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 445, + 125 + ], + "type": "image", + "image_path": "6bde63d342e6c7cf94b87467dbbd8374044b1eaacbbb4601078c8f00e39d141a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 449, + 72, + 507, + 125 + ], + "blocks": [ + { + "bbox": [ + 449, + 72, + 507, + 125 + ], + "lines": [ + { + "bbox": [ + 449, + 72, + 507, + 125 + ], + "spans": [ + { + "bbox": [ + 449, + 72, + 507, + 125 + ], + "type": "image", + "image_path": "351504f4a6cc247c3417f7c3100894262dfd94b013516e9f264df8f30cc2cb39.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 270, + 127, + 327, + 171 + ], + "blocks": [ + { + "bbox": [ + 270, + 127, + 327, + 171 + ], + "lines": [ + { + "bbox": [ + 270, + 127, + 327, + 171 + ], + "spans": [ + { + "bbox": [ + 270, + 127, + 327, + 171 + ], + "type": "image", + "image_path": "f7b4dce1ae444e3cf1495e5c3bbba378cd07204a38b923cd6b115b3f243963bc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 330, + 127, + 388, + 171 + ], + "blocks": [ + { + "bbox": [ + 330, + 127, + 388, + 171 + ], + "lines": [ + { + "bbox": [ + 330, + 127, + 388, + 171 + ], + "spans": [ + { + "bbox": [ + 330, + 127, + 388, + 171 + ], + "type": "image", + "image_path": "aa747899bbf73067470b5d0090df55432d29e6021d1647db492039cdd43cee5b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 127, + 446, + 171 + ], + "blocks": [ + { + "bbox": [ + 391, + 127, + 446, + 171 + ], + "lines": [ + { + "bbox": [ + 391, + 127, + 446, + 171 + ], + "spans": [ + { + "bbox": [ + 391, + 127, + 446, + 171 + ], + "type": "image", + "image_path": "93451363de034c4fb6fb3c28886744a2de50e6592136a88ead8875e46dfe7a3f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 449, + 127, + 507, + 171 + ], + "blocks": [ + { + "bbox": [ + 449, + 127, + 507, + 171 + ], + "lines": [ + { + "bbox": [ + 449, + 127, + 507, + 171 + ], + "spans": [ + { + "bbox": [ + 449, + 127, + 507, + 171 + ], + "type": "image", + "image_path": "3b4e2e2eb6f5c95b077e839363f3360f83be3f79bc4a9f0e6dc8863c2b59a8c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 178, + 544, + 274 + ], + "blocks": [ + { + "bbox": [ + 53, + 178, + 544, + 274 + ], + "lines": [ + { + "bbox": [ + 53, + 178, + 544, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 544, + 274 + ], + "type": "image", + "image_path": "7cb6556268707b46e596e279189ea6b1bf6f5bef8224dfc978eba5bbe6d8e409.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "type": "text", + "content": "Figure 1. (a) Space-preference bias. (Left): The relationship between distance " + }, + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "type": "inline_equation", + "content": "(x" + }, + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "type": "text", + "content": "-axis) and mIoU " + }, + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "type": "inline_equation", + "content": "(y" + }, + { + "bbox": [ + 46, + 281, + 545, + 381 + ], + "type": "text", + "content": "-axis) is drawn on PASCAL VOC [18]. The distance means the spatial distance between the centroids of the object and the image and mIoU is computed based on predictions and ground truth. The curve shows that CLIP [45] (green) is apparently better at segmentation for central objects than boundary ones, but our method (blue) effectively mitigates this bias. More details about how we draw this figure has been shown in our supplementary material. (Right): Visualizations also show our improvement on space-preference bias qualitatively. (b) Class-preference bias. (Left): We randomly select 6 classes from PASCAL Context [40] and draw the confusion matrix of CLIP and our model. It shows that beside the ground truth, CLIP also prefers to assign an incorrect but relevant label to a pixel in quite a few cases, while our results show apparent improvement. (Right): The visualizations are consistent with what we observed in confusion matrix. For example, for a \"cow\", CLIP tends to classify it as \"dog\" incorrectly." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 384, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 469 + ], + "type": "text", + "content": "categories in CLIP. For example, according to visualization (right), when ground truth is \"cow\", CLIP tends to incorrectly classify it as \"dog\". We also show such trend between randomly selected classes by confusion matrix (left). Elements on diagonal line represent right classification, while others are false. We observe a wide range of class-preference bias introduced by CLIP." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we propose to explicitly model and rectify the bias of CLIP to facilitate the USS. Specifically, we design two kinds of text inputs for each class, which are named as \"Reference\" and \"Query\" respectively. The Query is manually designed while the Reference contains learnable prompts. We adopt the text features of Query and Reference as classifiers to generate the Query and Reference logits respectively for each pixel of image. The Query logits represent the segmentation ability of original CLIP, while the Reference logits are expected to reflect the bias of CLIP preferring a specific class. Additionally, we project the positional embedding of CLIP's vision transformer to generate positional logits for each image. We expect the positional logits to represent space-preference bias of CLIP. Then, we remove the class-preference and the space-preference bias from original CLIP via a logit-subtraction mechanism, i.e., subtracting the Reference logits and the positional logits from the Query logits. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then the contrastive loss between masked visual" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 456 + ], + "type": "text", + "content": "feature and the text features of different categories is imposed to facilitate the effective bias modeling and rectification. To further improve the segmentation performance, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture, via mask-guided distillation, feature-guided distillation and text-guided learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "content": "We conduct extensive experiments on standard semantic segmentation benchmarks, including PASCAL VOC [18], PASCAL Context [40] and ADE20K [61]. Experiment results demonstrate that our method performs favourably against previous state-of-the-arts. Notably, on PASCAL VOC, our method outperforms CLIP S4 [22] by " + }, + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 459, + 545, + 554 + ], + "type": "text", + "content": ". Extensive ablation studies verify the effectiveness of each design in our framework." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 556, + 503, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 503, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 503, + 567 + ], + "type": "text", + "content": "Our contributions are summarized as follows." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 570, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 317, + 570, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 545, + 629 + ], + "type": "text", + "content": "- We observe that when applying CLIP to pixel-level understanding tasks, unexpected bias including space-preference bias and class-preference bias, occurs. Such bias may largely constrain the segmentation performance of CLIP-based segmentation models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 630, + 545, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 630, + 545, + 701 + ], + "spans": [ + { + "bbox": [ + 317, + 630, + 545, + 701 + ], + "type": "text", + "content": "- We propose to explicitly model the class-preference and space-preference bias of CLIP via learnable Reference text inputs and projection of positional embedding. Through a simple logit-subtraction mechanism and the contrastive loss built on masked features, we effectively rectify the bias of CLIP." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 545, + 714 + ], + "type": "text", + "content": "- We conduct extensive experiments on segmentation" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4103" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 109 + ], + "type": "text", + "content": "benchmarks under the unsupervised setting. Experiment results show superior performance of our method to previous state-of-the-arts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 134, + 132 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 258 + ], + "type": "text", + "content": "Pre-trained vision-language models. Pre-trained vision-language models (VLMs) [8, 14, 32, 33, 35] have developed rapidly with the help of large-scale image-text pairs available on the Internet. Recently, CLIP [45], ALIGN [26] and Slip [41] have made great progress on learning visual and textual representations jointly by using contrastive learning. With the image-level alignment with text, pre-trained VLMs have strong ability for zero-shot classification task and can be transferred to various downstream tasks, such as object detection [17, 54] and semantic segmentation [62]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 289, + 715 + ], + "type": "text", + "content": "Unsupervised semantic segmentation. While conventional approaches of semantic segmentation [37, 38, 53, 60] rely on pixel-level annotations and weakly-supervised methods [1, 16, 44, 56] still ask for image-level labels, unsupervised semantic segmentation (USS) methods [11, 20, 42, 50, 51] explores to train a segmentation model without any annotations. Models like [2, 7] adopt generative model [12] to separate foreground with background or generate corresponding masks. SegSort [23], HSG[29] and ACSeg [34] use clustering strategy, while IIC [25] uses mutual information maximization to perform unsupervised learning. MaskContrast [51] introduces contrastive learning into USS. Others like DSM [39] and LNE [13] exploit features extracted from self-supervised model, and combine it with spectral graph theory. However, the methods mentioned above either fail to segment images with multi-category objects or show a large performance gap with the supervised methods. Besides, they can only obtain class-agnostic masks and have to depend on additional strategy, such as Hungarian-matching algorithm [30], to match corresponding category with masks. Recently, pre-trained vision-language models are adopted in USS. MaskCLIP [62] modifies the image encoder of CLIP to generate patch-level features and directly performs segmentation with text features as classifiers. CLIP-py [46] performs contrastive learning between visual features from self-supervised ViT [4] and text features from CLIP. ReCo [49] performs image retrieval with CLIP and extracts class-wise embedding as classifier with co-segmentation. CLIP-S4 [22] learns pixel embeddings with pixel-segment contrastive learning and aligns such embeddings with CLIP in terms of embedding and semantic consistency. These methods can directly assign a category to each pixel, whose setting is named as language-guided unsupervised semantic segmentation. However, directly applying CLIP may bring prior bias, including space-preference bias and class-preference bias. As we know, there is no previous method trying to solve these bias and we manage to explicitly model the bias, and rectify them by" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 73, + 411, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 73, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 73, + 411, + 83 + ], + "type": "text", + "content": "element-wise subtraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 84, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 312 + ], + "type": "text", + "content": "Language-guided semantic segmentation. Recently, many works are exploring semantic segmentation guided by language under different settings. Zero-shot works [3, 31, 43, 55] split classes into seen and unseen set. During the training period, only masks of seen classes are provided. For inference, models are tested on both seen and unseen classes, but the test data is still in the same distribution with the training data. Open-vocabulary works [5, 19, 36, 47, 57-59] are trained in one scenario with extra annotations including class-agnostic masks or image captions, but are used for predicating segmentation masks of novel classes in other scenarios. From the technical view, our method also falls into the category of language-guided semantic segmentation. However, we consider the unsupervised setting. In this setting, we have access to images without any annotations during training. The training and inference images are sampled from the same distributions and the same set of categories. Such a setting is different to the typical zero-shot or open-vocabulary setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 323, + 362, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 362, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 362, + 335 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": "Background In this work, we aim to rectify the bias of CLIP for unsupervised semantic segmentation. In USS, we only have access to images without any types of annotations to train the segmentation model. For training and inference, the same set of categories are considered and the data distributions are assumed to be the same." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 545 + ], + "type": "text", + "content": "Overview The general framework of our method is illustrated in Fig. 2. We aim to rectify the bias of CLIP including the class-preference bias and the space-preference bias, to facilitate unsupervised semantic segmentation. From a high level, class-preference bias reflects the shift of CLIP predictions towards specific classes, while space-preference bias reflects the shift of CLIP predictions towards specific spatial locations. Both biases will be finally reflected in the predicted logit maps. A reasonable way to rectify the bias is to subtract its logit maps from the normal logit maps predicted via original CLIP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "To realize our goal, we firstly forward the image " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " through the image encoder of CLIP to obtain the patch-level image features " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ". We design two kinds of text inputs for each class. One is the manually designed Query text input " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " and the other one is the Reference text input " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " which consists of a learnable prompt and the name of class. For each class, passing " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " through the text encoder of CLIP, we obtain two text embeddings " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "W_{r}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ", which serve as the weights of query segmentation head and reference segmentation head respectively. Taking the same visual feature " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " as input, the query and reference segmentation heads output a query logit map " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " (Sec. 3.1) and Reference logit map " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " (Sec. 3.2) respectively. Meanwhile, positional embedding " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " is sent into a" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4104" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "learnable convolutional network to generate positional logit map " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " (Sec. 3.3). Then the bias logit map " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " can be obtained by adding Reference logit map " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " and positional logit map " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ". We then subtract " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": " to generate the rectified logits " + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": ". Based on the rectified logits, we generate masks for different classes. Then a contrastive loss is imposed between the masked visual features and the text features of different categories to encourage the bias modeling and rectification (Sec. 3.4). In order to enhance the segmentation performance, we distill the knowledge of rectified CLIP to the advanced segmentation architecture with specifically designed mask-guided, feature-guided and text-guided loss terms (Sec. 3.5). In both the rectification and distillation stages, we keep CLIP frozen." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 247, + 251, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 251, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 251, + 259 + ], + "type": "text", + "content": "3.1. Baseline: Directly Segment with CLIP" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": "Following MaskCLIP [62], we adapt pre-trained CLIP [45] (ViT-B/16) to the semantic segmentation task. We remove the query and key embedding layers of last attention but reformulate the value embedding layer and the last linear projection layer into two respective " + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": " convolutional layers. Therefore, the image encoder can not only generate local features for dense predictions, but also keep the visual-language association in CLIP by freezing its pre-trained weights. We forward image " + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": " through the image encoder and obtain patch-level features " + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{n \\times D}" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": " is the number of patches and " + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 264, + 287, + 396 + ], + "type": "text", + "content": " is the dimension of features in CLIP)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": "Each text in Query " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "Q = \\{Q_{1}, Q_{2}, \\dots, Q_{C}\\}" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " is the number of classes) is an ensemble of several manually designed prompts, e.g., \"a good/large/bad photo of a [CLS]\" where [CLS] denotes the a specific class name. Passing " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " through the text encoder, we obtain its text embeddings " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": ". We treat text embeddings " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " as the weight of segmentation head to perform " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " convolution. By sending features " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": " to the segmentation head, we get a Query logit map " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "M_{q} \\in \\mathbb{R}^{n \\times C}" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": ". Then the segmentation mask can be predicted by arg max operation on " + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 46, + 396, + 289, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 522, + 204, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 204, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 204, + 534 + ], + "type": "text", + "content": "3.2. Learn Class-Preference Bias" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "content": "In order to explicitly model the class-preference bias brought by pre-trained CLIP for specific datasets, we design a Reference " + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "inline_equation", + "content": "i = \\{1,2,\\dots ,C\\}" + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "content": ", as additional text input for each class. Inspired by CoOp [63], " + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "content": " consists of a learnable prompt which is shared across all the classes for efficiency, and a class name " + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "inline_equation", + "content": "[CLS]" + }, + { + "bbox": [ + 46, + 541, + 288, + 615 + ], + "type": "text", + "content": ", which can be formed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 624, + 287, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 624, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 100, + 624, + 287, + 638 + ], + "type": "interline_equation", + "content": "R _ {i} = \\left[ v _ {1} \\right] \\left[ v _ {2} \\right] \\dots \\left[ v _ {l} \\right] \\dots \\left[ v _ {L} \\right] [ C L S ], \\tag {1}", + "image_path": "d15535a2d0365d4f1015b5421390dc3b6f53d335f461a6a4996c5896f3af6103.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "where each " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "[v_l](l \\in \\{1, \\dots, L\\})" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is a vector with dimension " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and serves as a word embedding. The " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is a constant representing the number of word embeddings to learn. Totally, there are 77 word embeddings for a text of CLIP. As two word embeddings are used for class name and two for indicating start and end of a text, we set " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " to 73." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Passing the Reference " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " through text encoder, we obtain reference text embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "W_{r}\\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". The text embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "W_{r}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is then directly utilized as the weights of segmentation head to perform " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " convolution on visual feature " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " and output a Reference logit map " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "M_{r}\\in \\mathbb{R}^{n\\times C}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "content": "As we obtain different Reference logit maps for different classes, we expect the Reference logit map to encode the class-preference bias to facilitate the following bias rectification process. It is worth noting that we make Reference " + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "content": " learnable but keep Query " + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "content": " fixed. It is because when we make them both learnable, Query may also capture some class-preference bias and technically we cannot guarantee which text embedding should encode the class-preference bias, resulting in an implicit bias modeling. Thus, in our framework, we choose not to make the " + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "content": " learnable to just encourage " + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 133, + 546, + 289 + ], + "type": "text", + "content": " encode the bias. Such a design cooperates with the following logit subtraction mechanism to make the bias modeling and rectification more meaningful and effective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 299, + 464, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 299, + 464, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 464, + 312 + ], + "type": "text", + "content": "3.3. Learn Space-Preference Bias" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 378 + ], + "type": "text", + "content": "In ViT [15], positional embeddings (PE) are important for encoding spatial information to features. Thus, we assume the space-preference bias should depend on the positional embeddings (PE) and choose to learn a projection of PE to represent the space-preference bias." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "content": "We design a 3-layer " + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "content": " convolutional network, and each convolutional layer is followed by a batch normalization. Specifically, we project PE " + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "content": " by the designed convolutional network to obtain positional logits " + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "inline_equation", + "content": "M_p \\in \\mathbb{R}^{n \\times 1}" + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "content": ". During the training process, the projection network is optimized to model the space-preference bias in positional logits " + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "inline_equation", + "content": "M_p" + }, + { + "bbox": [ + 304, + 380, + 546, + 475 + ], + "type": "text", + "content": ". Since PE is shared across all categories, the learned space-preference bias is also shared across all the categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 485, + 536, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 485, + 536, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 485, + 536, + 498 + ], + "type": "text", + "content": "3.4. Rectify Bias with Contrastive Learning Loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": "By keeping Query prompt " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " fixed, the Query logits " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{q}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " represents the natural prediction ability of CLIP model, which may contain class-preference bias and space-preference bias. With learnable Reference prompt and projection of PE, we explicitly encode the class-preference bias and space-preference bias into the Reference logit map " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " and positional logit map " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " respectively. We then add " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{r}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{p}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": " together to depict the final bias " + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 304, + 505, + 545, + 601 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 612, + 545, + 627 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 612, + 545, + 627 + ], + "spans": [ + { + "bbox": [ + 389, + 612, + 545, + 627 + ], + "type": "interline_equation", + "content": "M _ {b} = M _ {r} + M _ {p} ^ {*}, \\tag {2}", + "image_path": "142ba827c4c094bf0f52ed7adcd4ee3a23c888436df5c0cba0c5887075d29f31.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "content": "where we simply expand " + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "inline_equation", + "content": "M_p" + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "inline_equation", + "content": "M_p^* \\in \\mathbb{R}^{n \\times C}" + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "content": " which means we repeat each " + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "inline_equation", + "content": "M_p" + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 637, + 545, + 662 + ], + "type": "text", + "content": " times." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "content": "Then we perform a simple element-wise subtraction between " + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "inline_equation", + "content": "M_q" + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "inline_equation", + "content": "M_b" + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "content": " to generate the rectified logit map " + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 305, + 662, + 545, + 687 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 391, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 391, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "M = M _ {q} - M _ {b}. \\tag {3}", + "image_path": "5785d6df26bbffe595fc20716f2578cff3703b2eb9928666d0617b6fb45022ac.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4105" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 301, + 202 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 301, + 202 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 301, + 202 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 301, + 202 + ], + "type": "image", + "image_path": "9632a2b4208aba1e3dbd642e4b06ea94125ba29c48b529ef06c6b493f60b1344.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 206, + 197, + 215 + ], + "lines": [ + { + "bbox": [ + 136, + 206, + 197, + 215 + ], + "spans": [ + { + "bbox": [ + 136, + 206, + 197, + 215 + ], + "type": "text", + "content": "(a) Rectification Stage" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 303, + 72, + 545, + 202 + ], + "blocks": [ + { + "bbox": [ + 303, + 72, + 545, + 202 + ], + "lines": [ + { + "bbox": [ + 303, + 72, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 303, + 72, + 545, + 202 + ], + "type": "image", + "image_path": "dacd7504c4a63c683af85660184ef7567e39f64c721250c4e26b64ff8471921b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 206, + 450, + 215 + ], + "lines": [ + { + "bbox": [ + 409, + 206, + 450, + 215 + ], + "spans": [ + { + "bbox": [ + 409, + 206, + 450, + 215 + ], + "type": "text", + "content": "(b) Distillation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 224, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 547, + 258 + ], + "type": "text", + "content": "Figure 2. Method overview. We propose a new framework for language-guided unsupervised semantic segmentation. (a) Rectification Stage: At this stage, we aim to rectify the class-preference and space-preference bias of CLIP. (b) Distillation: We distill knowledge from the rectified CLIP to the advanced segmentation architecture with the mask-guided, feature-guided and text-guided loss terms." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 265, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 288, + 300 + ], + "type": "text", + "content": "From a high-level, this operation can be interpreted as subtracting the \"bias\" from the predictions with text features of Query as segmentation head." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "text", + "content": "Since the arg max operation is not differentiable, we utilize a Gumbel-Softmax trick [24] to generate the candidate masks for each class with " + }, + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 46, + 300, + 287, + 337 + ], + "type": "text", + "content": " as the temperature, i.e.," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 342, + 287, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 342, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 100, + 342, + 287, + 356 + ], + "type": "interline_equation", + "content": "\\hat {M} = \\text {G u m b e l - S o f t m a x} (M, \\tau_ {1}). \\tag {4}", + "image_path": "720cf027092c5053ce036adb0e57668ff71e4856f1d4ba45a00b0f9be135fe3f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": "We apply " + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": " to the feature " + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": " to get the masked features " + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "Z_{g}" + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": ", which are expected to encode the features of objects with different classes. We then design a contrastive loss to supervise the learning via aligning masked features with text features of different classes. As the masked features represent the objects of interest without context, we utilize " + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": " generated by text input " + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 47, + 361, + 287, + 445 + ], + "type": "text", + "content": " for alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 445, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 552 + ], + "type": "text", + "content": "Additionally, as it is usually impossible for an image to contain all the classes of interest, we need to infer what classes exist in the current image. Our strategy is as follows. Firstly, we choose to extract global visual feature from image encoder of CLIP. Then we calculate the similarity scores between the text input of a specific class and the global visual feature of the image. We only choose the classes whose similarity scores are higher than a threshold " + }, + { + "bbox": [ + 46, + 445, + 287, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 445, + 287, + 552 + ], + "type": "text", + "content": " as the potential classes exist in the current image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "content": "Then we may select " + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "content": " pseudo labels " + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "\\{c_1, c_2, \\dots, c_K\\}" + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "content": " for each image. We perform a global average pooling of " + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "inline_equation", + "content": "Z_g" + }, + { + "bbox": [ + 46, + 552, + 287, + 612 + ], + "type": "text", + "content": " and compute the similarities between pooled visual features of pseudo labels and text features of all the categories. The contrastive loss can be defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 616, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 616, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 86, + 616, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\frac {1}{K} \\sum_ {k = 1} ^ {K} \\log \\frac {\\exp \\left\\{S _ {c _ {k} , c _ {k}} / \\tau \\right\\}}{\\sum_ {j = 1} ^ {C} \\exp \\left\\{S _ {c _ {k} , j} / \\tau \\right\\}} \\tag {5}", + "image_path": "0726a17baf8aa8df3d76033a845103f02f54b85bbe092e79b9c858ad53d731de.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "S_{c_k,j}" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " denotes the similarity between visual feature of pseudo label " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " and text feature of " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "-th (" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "j \\in \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": ") category and the " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": " is a constant." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "A better modeling of bias yields more accurate estimations of object masks. Then the masked features of objects" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 265, + 545, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 265, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 304, + 265, + 545, + 348 + ], + "type": "text", + "content": "are more aligned with corresponding text features. As a result, the contrastive loss (Eq. (5)) will be lower. In contrast, a worse modeling of bias results in higher contrastive loss. Thus, minimizing Eq. (5) will drive the model to update towards making more accurate mask predictions, i.e., rectifying the bias of CLIP when adapted to the downstream USS task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 354, + 484, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 484, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 484, + 365 + ], + "type": "text", + "content": "3.5. Distillation for Enhanced Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 372, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 504 + ], + "type": "text", + "content": "As CLIP is not specifically designed for segmentation tasks, we choose to distill the knowledge from the rectified CLIP (teacher) to the advanced segmentation architecture (student) to enhance the feature representations and finally improve the segmentation performance. In our paper, we choose DeepLab V2 [6] as the student network which directly inherits and fixes the Query and Reference segmentation heads, and utilizes logit subtraction mechanism to generate the final segmentation masks. In our design, the rectified CLIP works as a teacher to guide the feature learning of DeepLab V2 with three designed loss terms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "text", + "content": "Mask-guided loss. We directly exploit segmentation masks generated from the bias rectification stage as pseudo labels. We compute a cross-entropy loss between the pseudo labels " + }, + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\tilde{M}" + }, + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "text", + "content": " and predictions " + }, + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "inline_equation", + "content": "M^D" + }, + { + "bbox": [ + 304, + 504, + 545, + 563 + ], + "type": "text", + "content": " from DeepLab V2, i.e.," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 345, + 562, + 545, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 562, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 345, + 562, + 545, + 595 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = - \\frac {1}{H W} \\sum_ {i} ^ {H} \\sum_ {j} ^ {W} \\log P _ {i j} (\\tilde {M} _ {i j}) \\tag {6}", + "image_path": "6026a5ddcb2d25ef59e5dea9935a444f041359f34c0859fae8e82d29e81d2673.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "content": " denote the height and width of an image. The " + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "inline_equation", + "content": "P_{ij}(\\tilde{M}_{ij}) = \\mathrm{Softmax}(M_{ij}^D)_{\\tilde{M}_{ij}}" + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "content": ", which means the predicted probability with respect to the pseudo label " + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "inline_equation", + "content": "\\tilde{M}_{ij}" + }, + { + "bbox": [ + 304, + 597, + 545, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "text", + "content": "Feature-guided loss. After obtaining features " + }, + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "text", + "content": " from visual backbone of rectified CLIP and " + }, + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 304, + 637, + 545, + 695 + ], + "type": "text", + "content": " from student visual backbone, we resize two features to the same shape of original image by bilinear interpolation. We then perform a L1 loss between two features" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 384, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t}} = \\left| \\left| Z - \\hat {Z} \\right| \\right| _ {1} \\tag {7}", + "image_path": "f9b22f68cd566fdd22a53c3bc45854ef3f30f6ce5d685be9dcfd29e2b4000d7c.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4106" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "Therefore, visual features from student network is aligned with visual and textual features of rectified CLIP, providing an important basis for the following text-guided loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "text", + "content": "Text-guided loss. We adopt the same strategy as the Rectification stage (Sec. 3.4) to compute contrastive loss as our text-guided loss " + }, + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{text}" + }, + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "text", + "content": ". The only difference is that we use feature " + }, + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 46, + 108, + 286, + 167 + ], + "type": "text", + "content": " and masks generated by the student network to obtain the corresponding masked features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": "Total distillation loss. Total distillation loss " + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{distill}" + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": " is calculated as follows, where " + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 167, + 286, + 204 + ], + "type": "text", + "content": " are both constants and set to 0.5. Effect of each loss term is studied in Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 212, + 286, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 212, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 88, + 212, + 286, + 224 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} = \\mathcal {L} _ {\\text {m a s k}} + \\alpha \\mathcal {L} _ {\\text {f e a t}} + \\beta \\mathcal {L} _ {\\text {t e x t}}. \\tag {8}", + "image_path": "664b2fcf58133c6592bcebef1b5e88120a004d74c823f9ea8d78985e417628d3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 232, + 128, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 128, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 128, + 247 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 252, + 96, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 252, + 96, + 265 + ], + "spans": [ + { + "bbox": [ + 47, + 252, + 96, + 265 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 270, + 286, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 286, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 286, + 389 + ], + "type": "text", + "content": "Datasets. We conduct experiments on three standard benchmarks for semantic segmentation, including PASCAL VOC 2012 [18], PASCAL Context [40] and ADE20K [61]. PASCAL VOC 2012 (1,464/1,449 train/validation) contains 20 object classes, while PASCAL Context (4,998/5,105 train/validation) is an extension of PASCAL VOC 2010 and we consider 59 most common classes in our experiments. ADE20K (20,210/2,000 train/validation) is a segmentation dataset with various scenes and 150 most common categories are considered." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 390, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 557 + ], + "type": "text", + "content": "Implementation details. For the image encoder of CLIP, we adopt ViT-B/16 as visual backbone. For the text encoder of CLIP, we adopt Transformer [52]. During the whole training period, we keep both of the encoders frozen. We use conventional data augmentations including random cropping and random flipping. Relevant hyper-parameters for each dataset, including number of rectification epochs, number of distillation iterations and parameters of data augmentation are shown in our supplementary material. For both stages, we use a SGD [48] optimizer with a learning rate of 0.01 and a weight decay of 0.0005. We adopt the poly strategy with the power of 0.9 for learning rate. In our experiment, we report the mean intersection over union (mIoU) as evaluation metric." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "Baselines. We compare with previous USS methods to verify the superiority of our method, including MaskCLIP " + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "inline_equation", + "content": "(+)" + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": " [62], CLIP-S4 [22], ReCo [49], CLIPpy [46], etc. From the technical view, although our method falls into the category of text-guided segmentation, we choose not to compare with text-guided methods which are designed for zero-shot or open-vocabulary setting (e.g., GroupViT [57], TCL [5] and ViewCo [47]). Two main differences exist between the zero-shot/open-vocabulary setting and the unsupervised setting: 1) in zero-shot/open-vocabulary, the category sets for training and inference are non-overlapped, but in USS the category sets are shared; 2) text-guided zero-shot/open-vocabulary methods usually rely on large-scale" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 330, + 102, + 518, + 239 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Comparison with non-language-guided unsupervised semantic segmentation methods on PASCAL VOC." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 102, + 518, + 239 + ], + "lines": [ + { + "bbox": [ + 330, + 102, + 518, + 239 + ], + "spans": [ + { + "bbox": [ + 330, + 102, + 518, + 239 + ], + "type": "table", + "html": "
ModelLCmIoU
IIC [25]✓9.8
SegSort [23]✓11.7
Deep Spectral Methods [39]✗37.2
HSG [29]✗41.9
ACSeg [34]✗47.1
MaskContrast [51]✓49.6
Ours (rectification)✗58.5
Ours (distill)✗75.4
Ours (distill)✓76.1
", + "image_path": "330846b29a901e9a6ac4c44150a7ebf593351f207d20c5eff4dd4267138b9970.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 280 + ], + "type": "text", + "content": "image-text pairs or class-agnostic masks to supervise the training, but in USS we only utilize the images without any types of annotations to train the model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 287, + 484, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 287, + 484, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 287, + 484, + 300 + ], + "type": "text", + "content": "4.2. Comparison with SOTA methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 390 + ], + "type": "text", + "content": "We compare our method with both previous non-language-guided USS methods (Table 1) and CLIP-based USS methods (Table 2). In all tables, we use \"Ours (rectification)\" to denote the segmentation results after the rectification stage and \"Ours (distill)\" to denote the segmentation results after distillation. Unless otherwise stated, results from previous methods are directly cited from the original papers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "As shown in Table 1, our model shows remarkably better segmentation results compared with the conventional non-language-guided USS methods. In order to further evaluate the strength of features extracted from our distillation model, we also report our results of linear classification (LC). Following MaskContrast [51], we fix our model to generate pixel embeddings and train a linear classifier to generate semantic segmentation masks. The results of LC also prove that our method extracts strong features and achieves significant gains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In Table 2, we also make a comparison with typical CLIP-based methods, including MaskCLIP, CLIPpy, ReCo and CLIP-S4. These methods all consider the same CLIP-based USS setting with ours. The results show our method performs favorably against previous CLIP-based methods. For example, after the rectification stage, our method outperforms MaskCLIP by " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "9.0\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " respectively on the three datasets, while after distillation, our method outperforms " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MaskCLIP + }" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ". Our method also shows better results than CLIPpy by " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "20.8\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " on PASCAL VOC and ADE20K and outperforms CLIP-S4 by " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " on PASCAL VOC and PASCAL Context. Since ReCo employs a \"context elimination\" (CE) trick which may introduce prior knowledge, we also report the results of ReCo by removing this trick (ReCo w/o CE in Table 2). Our method outperforms ReCo obviously, e.g., on PASCAL VOC, ours (rectification) outper" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4107" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 102, + 265, + 228 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 94 + ], + "type": "text", + "content": "Table 2. Comparison with CLIP-based unsupervised semantic segmentation methods on various benchmarks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 102, + 265, + 228 + ], + "lines": [ + { + "bbox": [ + 66, + 102, + 265, + 228 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 265, + 228 + ], + "type": "table", + "html": "
MethodVOCContextADE
CLIPpy [46]54.6/13.5
MaskCLIP [62]49.521.79.5
MaskCLIP+ [62]70.031.112.2
ReCo [49]55.226.2/
ReCo (w/o CE)54.823.1/
CLIP-S4 [22]72.033.6/
Ours (rectification)58.525.811.1
Ours (distill)75.433.814.3
", + "image_path": "4e96df8ce99ceafd4509e47896b340fe78b6c6bf3086aa53e1b8cc9925bd09e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 277, + 265, + 326 + ], + "blocks": [ + { + "bbox": [ + 47, + 235, + 287, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 235, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 287, + 269 + ], + "type": "text", + "content": "Table 3. Ablation on whether Query should be learnable. Results show that fixing Query performs favorably against making Query learnable." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 277, + 265, + 326 + ], + "lines": [ + { + "bbox": [ + 65, + 277, + 265, + 326 + ], + "spans": [ + { + "bbox": [ + 65, + 277, + 265, + 326 + ], + "type": "table", + "html": "
QueryReferenceVOCContextADE
✓✓56.723.79.4
X✓56.724.410.5
", + "image_path": "b8421e69c7116fabbec9a9b29faa933b2d97cbb2b9c7ee073c7c05060246ece9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "text", + "content": "forms ReCo and ReCo (w/o CE) by " + }, + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 47, + 329, + 286, + 353 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 360, + 139, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 139, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 139, + 373 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 378, + 286, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 484 + ], + "type": "text", + "content": "Should Query be learnable? We study on whether the prompt of Query should be learnable. As shown in Table 3, we conduct experiments with learnable and fixed Query respectively. As we aim to show the effect of fixing Query, we don't utilize the projected positional logits to calculate the numbers shown in the table. The numbers show that fixing Query performs favorably against making Query learnable. This verifies that we should only learn Reference but keep Query fixed as we discuss in Sec. 3.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "content": "Effect of different bias modeling. As shown in Table 4, compared with the baseline without any bias rectification (numbers with only \"Query\"), utilizing learable Reference to model the class-preference bias brings remarkable performance improvement, e.g., " + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "content": " mIoU on PASCAL VOC, " + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "content": " mIoU on PASCAL Context and " + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "content": " mIoU on ADE20K. While introducing projection of positional embedding to model the space-preference bias, the numbers are further improved, i.e., around " + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 628 + ], + "type": "text", + "content": " better than modeling the class-preference bias only. Those results verify the effectiveness of bias modeling by introducing learnable Reference and projection of positional embedding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Effect of element-wise subtraction on bias rectification. In order to validate the effect of our element-wise subtraction, we conduct experiments in Table 5. We compare our subtraction mechanism with an alternative solution, i.e., instead of subtracting logits encoding bias (i.e., Reference logits plus positional logits), we add all the logits. Comparisons shown in the table verify the effectiveness of our" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 102, + 537, + 163 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 4. Effect of bias modeling. Results show effect of both bias modeling quantitatively and each contributes to better results." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 102, + 537, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 102, + 537, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 537, + 163 + ], + "type": "table", + "html": "
QueryReferencePEVOCContextADE
✓49.521.79.5
✓✓56.724.410.5
✓✓✓58.525.811.1
", + "image_path": "b884a09fa8ff0739dcad062827fb478849ddf78862d602d14bb096d8982ec2f9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 205, + 529, + 266 + ], + "blocks": [ + { + "bbox": [ + 306, + 174, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 306, + 174, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 174, + 545, + 196 + ], + "type": "text", + "content": "Table 5. Effectiveness of element-wise subtraction. Results show that the element-wise subtraction removes bias from CLIP." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 205, + 529, + 266 + ], + "lines": [ + { + "bbox": [ + 318, + 205, + 529, + 266 + ], + "spans": [ + { + "bbox": [ + 318, + 205, + 529, + 266 + ], + "type": "table", + "html": "
QueryRef+PEsubaddVOCContext
✓XXX49.521.7
✓✓X✓57.824.9
✓✓✓X58.525.8
", + "image_path": "b47a2eefadab7b46db1076d055193b988896bf2ff38855dde8b069a8edb78450.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 341 + ], + "type": "text", + "content": "subtraction way. As we aim to remove the bias from the original CLIP, we speculate that the subtraction operation may work as a strong prior which regularizes the training to facilitate meaningful and effective bias modeling. The effect of bias rectification by element-wise subtraction is further visualized in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 533 + ], + "type": "text", + "content": "Visualization of Bias Modeling. In Fig. 3 (b), we illustrate how we explicitly model and rectify the class-preference bias. As shown in the first column of Fig. 3 (b), the original CLIP (see \"Query\" mask) tends to misclassify part of a \"person\" (see the ground-truth mask denoted as \"GT\") into a \"boat\". Such a mistake is reflected in the comparison between the logit heatmaps for the \"person\" channel and the \"boat\" channel: in the area misclassified as \"boat\", the logits for \"boat\" are relatively higher than those for \"person\". In contrast, for Reference logit map, the person-channel logits are quite low, while the boat-channel logits are generally very high, especially for the misclassified area. Consequently, via the proposed logit subtraction operation, we obtain the rectified logits (see the last column), where the boat channel is largely suppressed. Finally, we obtain a much better mask (see \"Ours\" in the first column)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 640 + ], + "type": "text", + "content": "In Fig. 3 (c), we aim to show that the projection of positional embedding (PE) effectively models the space-preference bias. By comparing the results with and without rectifying logits projected by PE (see the dashed boxes of the last two columns), we find that rectification with PE largely improves the segmentation performance in the boundary areas. The results illustrate that the projection of PE does encode the space-preference bias and rectifying such bias may largely improve the segmentation results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "content": "Effect of different distillation loss terms. We conduct experiments on PASCAL VOC to validate the effect of each loss term of our distillation framework, including the mask-guided, feature-guided and text-guided loss terms. From Table 7, all the loss terms contribute to better performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "Effect of segmentation head for distillation. As illustrated" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4108" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 303, + 259 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 303, + 259 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 303, + 259 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 303, + 259 + ], + "type": "image", + "image_path": "2085a7ff846c08481c69541cf64fe3d979f818cb9e131f98ef34b9ad308fd423.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": "Figure 3. (a) Qualitative Results: We visualize segmentation results on PASCAL Context. From the visualization, we observe that our model outperforms MaskCLIP " + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "inline_equation", + "content": "(+)" + }, + { + "bbox": [ + 46, + 266, + 547, + 312 + ], + "type": "text", + "content": " obviously by rectifying both class-preference bias and space-preference bias. (b) Class-preference bias: In order to explain how Reference explicitly models the class-preference bias, we show the heatmap of Reference logits. (c) Space-preference bias: The segmentation within dashed boxes shows the effectiveness of PE projection on rectifying space-preference bias." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 258 + ], + "type": "image", + "image_path": "d62c94cdbc5ddb8cba0e249a0cf1658006ca0fc919781fbf04b60085c31c61c7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 367, + 277, + 438 + ], + "blocks": [ + { + "bbox": [ + 47, + 324, + 288, + 360 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 288, + 360 + ], + "type": "text", + "content": "Table 6. Effect of segmentation head for distillation. We perform distillation with both original classification head (Ori. Head) and structure from our rectification stage" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 367, + 277, + 438 + ], + "lines": [ + { + "bbox": [ + 53, + 367, + 277, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 367, + 277, + 438 + ], + "type": "table", + "html": "
Ori. HeadQueryReferencePE Proj.mIoU
✓XXX73.8
X✓XX74.1
X✓✓X75.4
X✓✓✓75.0
", + "image_path": "9ba8efd3af1bde7e8b5ad712b680b3dc3dd68f6e260144384a30904ba71d52ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 93, + 470, + 237, + 531 + ], + "blocks": [ + { + "bbox": [ + 47, + 439, + 286, + 461 + ], + "lines": [ + { + "bbox": [ + 47, + 439, + 286, + 461 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 286, + 461 + ], + "type": "text", + "content": "Table 7. Ablation results on distillation loss. According to the results, each loss item contributes to better distillation result." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 470, + 237, + 531 + ], + "lines": [ + { + "bbox": [ + 93, + 470, + 237, + 531 + ], + "spans": [ + { + "bbox": [ + 93, + 470, + 237, + 531 + ], + "type": "table", + "html": "
MaskFeatureTextmIoU
✓73.0
✓✓74.1
✓✓✓75.4
", + "image_path": "f198ffd3fc4cefebbea7192ef3ce98c4301eb3a51a2e8aa2449045a7109364b3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 701 + ], + "type": "text", + "content": "in Fig 2 (b), instead of directly using original classification head at the distillation stage, the student network (i.e., DeepLab V2 [6]) inherits and fixes the Query and Reference segmentation heads from our rectification stage, and utilizes logit subtraction mechanism to generate the final segmentation masks. In Table 6, we conduct experiments with original head (\"Ori. Head\") of DeepLab V2 and other types of segmentation heads which are directly inherited from our rectification stage on PASCAL VOC. From the results, we observe that it is better to inherit the components from our rectification stage than adopting the original classification head. However, introducing projection of PE in segmentation head cannot bring further improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 286, + 714 + ], + "type": "text", + "content": "Qualitative Results. We visualize our segmentation re" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 327, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 411 + ], + "type": "text", + "content": "sults in Fig. 3 (a). It can be observed that there exists apparent space-preference bias and class-preference bias in the segmentation results of original CLIP (MaskCLIP) and these bias still cannot be removed even after distillation (MaskCLIP+). However, our model outperforms MaskCLIP(+) obviously by rectifying the bias of CLIP for unsupervised semantic segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 378, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 468, + 545, + 659 + ], + "type": "text", + "content": "In this paper, we propose a new framework for language-guided unsupervised semantic segmentation. We observe bias, including space-preference bias and class-preference bias, exists in CLIP when directly applying CLIP to segmentation task. We propose using additional Reference to learn class-preference bias and projecting positional embedding to represent space-preference bias, and then manage to rectify them by a simple element-wise logit subtraction mechanism. For further improving the segmentation performance, we distill the knowledge from rectified CLIP to advanced segmentation backbone with specifically designed losses. Extensive experiments demonstrate that our method achieve superior segmentation performance compared to previous state-of-the-arts. We hope our work may inspire future research to investigate how to better adapt CLIP to complex visual understanding tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 712 + ], + "type": "text", + "content": "Acknowledgement This research project is supported by National Natural Science Foundation (NNSF) of China under Grant 92370114 and National Key R&D Program of China under Grant 2022YFB3104900." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4109" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 180 + ], + "type": "text", + "content": "[2] Adam Bielski and Paolo Favaro. Emergence of object segmentation in perturbed generative models. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 215 + ], + "type": "text", + "content": "[3] Maxime Bucher, Tuan-Hung Vu, Matthieu Cord, and Patrick Pérez. Zero-shot semantic segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "text", + "content": "[5] Junbum Cha, Jonghwan Mun, and Byungseok Roh. Learning to generate text-grounded mask for open-world semantic segmentation from only image-text pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11165-11174, 2023. 3, 6, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 393 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2018. 1, 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 430 + ], + "type": "text", + "content": "[7] Mickaël Chen, Thierry Artières, and Ludovic Denoyer. Unsupervised object segmentation by redrawing. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 475 + ], + "type": "text", + "content": "[8] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In European Conference on Computer Vision, pages 104-120, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "text", + "content": "[9] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 522, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 288, + 577 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 288, + 633 + ], + "type": "text", + "content": "[11] Jang Hyun Cho, Utkarsh Mall, Kavita Bala, and Bharath Hariharan. Picie: Unsupervised semantic segmentation using invariance and equivariance in clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16794-16804, 2021. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 288, + 678 + ], + "type": "text", + "content": "[12] Antonia Creswell, Tom White, Vincent Dumoulin, Kai Arulkumaran, Biswa Sengupta, and Anil A Bharath. Generative adversarial networks: An overview. IEEE Signal Processing Magazine, 35(1):53-65, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 713 + ], + "type": "text", + "content": "[13] Zhijie Deng and Yucen Luo. Learning neural eigenfunctions for unsupervised semantic segmentation. arXiv preprint arXiv:2304.02841, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[14] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[15] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, and S. Gelly. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 226 + ], + "type": "text", + "content": "[16] Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to-prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 282 + ], + "type": "text", + "content": "[17] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 327 + ], + "type": "text", + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 546, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 382 + ], + "type": "text", + "content": "[19] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level labels. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXVI, pages 540-557. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 546, + 426 + ], + "type": "text", + "content": "[20] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T. Freeman. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 480 + ], + "type": "text", + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 482, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 546, + 536 + ], + "type": "text", + "content": "[22] Wenbin He, Suphanut Jamonnak, Liang Gou, and Liu Ren. Clip-s4: Language-guided self-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11207-11216, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 537, + 546, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 601 + ], + "type": "text", + "content": "[23] Jyh-Jing Hwang, Stella X Yu, Jianbo Shi, Maxwell D Collins, Tien-Ju Yang, Xiao Zhang, and Liang-Chieh Chen. Segsort: Segmentation by discriminative sorting of segments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7334-7344, 2019. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "type": "text", + "content": "[24] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "text", + "content": "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9865-9874, 2019. 1, 3, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4110" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[26] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[27] Guoliang Kang, Lu Jiang, Yunchao Wei, Yi Yang, and Alexander Hauptmann. Contrastive adaptation network for single-and multi-source domain adaptation. IEEE transactions on pattern analysis and machine intelligence, 44(4): 1793-1804, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 251 + ], + "type": "text", + "content": "[28] Guoliang Kang, Yunchao Wei, Yi Yang, Yueting Zhuang, and Alexander Hauptmann. Pixel-level cycle association: A new perspective for domain adaptive semantic segmentation. Advances in neural information processing systems, 33: 3569-3580, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 253, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 319 + ], + "type": "text", + "content": "[29] Tsung-Wei Ke, Jyh-Jing Hwang, Yunhui Guo, Xudong Wang, and Stella X Yu. Unsupervised hierarchical semantic segmentation with multiview cosegmentation and clustering transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2571–2581, 2022. 1, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 320, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 352 + ], + "type": "text", + "content": "[30] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[31] Boyi Li, Kilian Q. Weinberger, Serge J. Belongie, Vladlen Koltun, and René Ranftl. Language-driven semantic segmentation. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 464 + ], + "type": "text", + "content": "[32] Gen Li, Nan Duan, Yuejian Fang, Ming Gong, and Daxin Jiang. Unicoder-vl: A universal encoder for vision and language by cross-modal pre-training. In Proceedings of the AAAI conference on artificial intelligence, pages 11336-11344, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 522 + ], + "type": "text", + "content": "[33] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 589 + ], + "type": "text", + "content": "[34] Kehan Li, Zhennan Wang, Zesen Cheng, Runyi Yu, Yian Zhao, Guoli Song, Chang Liu, Li Yuan, and Jie Chen. Acseg: Adaptive conceptualization for unsupervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2023. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 590, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 646 + ], + "type": "text", + "content": "[35] Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu. Hero: Hierarchical encoder for video+ language omni-representation pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2046-2065, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[36] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7061-7070, 2023. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[37] Chen Liang-Chieh, George Papandreou, Iasonas Kokkinos, Kevin Murphy, et al. Semantic image segmentation with deep convolutional nets and fully connected crfs. In International Conference on Learning Representations, 2015. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[38] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[39] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 242, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 307 + ], + "type": "text", + "content": "[40] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 352 + ], + "type": "text", + "content": "[41] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European Conference on Computer Vision, pages 529-544. Springer, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[42] Yassine Ouali, Céline Hudelot, and Myriam Tami. Autoregressive unsupervised image segmentation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII 16, pages 142-158. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 464 + ], + "type": "text", + "content": "[43] Giuseppe Pastore, Fabio Cermelli, Yongqin Xian, Massimiliano Mancini, Zeynep Akata, and Barbara Caputo. A closer look at self-training for zero-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2693-2702, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "text", + "content": "[44] Pedro O Pinheiro and Ronan Collobert. From image-level to pixel-level labeling with convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1713-1721, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 577 + ], + "type": "text", + "content": "[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "type": "text", + "content": "[46] Kanchana Ranasinghe, Brandon McKinzie, Sachin Ravi, Yinfei Yang, Alexander Toshev, and Jonathon Shlens. Perceptual grouping in vision-language models. arXiv preprint arXiv:2210.09996, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[47] Pengzhen Ren, Changlin Li, Hang Xu, Yi Zhu, Guangrun Wang, Jianzhuang Liu, Xiaojun Chang, and Xiaodan Liang. Viewco: Discovering text-supervised segmentation masks via multi-view semantic consistency. arXiv preprint arXiv:2302.10307, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[48] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4111" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[49] Gyungin Shin, Weidi Xie, and Samuel Albanie. Reco: Retrieve and co-segment for zero-shot transfer. In Advances in Neural Information Processing Systems. 1, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "text", + "content": "[50] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part X, pages 268-285. Springer, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 228 + ], + "type": "text", + "content": "[51] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10052-10062, 2021. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 288, + 327 + ], + "type": "text", + "content": "[53] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11534-11542, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 371 + ], + "type": "text", + "content": "[54] Zhenyu Wang, Yali Li, Xi Chen, Ser-Nam Lim, Antonio Torralba, Hengshuang Zhao, and Shengjin Wang. Detecting everything in the open world: Towards universal object detection. arXiv preprint arXiv:2303.11749, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 288, + 426 + ], + "type": "text", + "content": "[55] Yongqin Xian, Subhabrata Choudhury, Yang He, Bernt Schiele, and Zeynep Akata. Semantic projection network for zero-and few-label semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8256-8265, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 288, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 491 + ], + "type": "text", + "content": "[56] Jinheng Xie, Jianfeng Xiang, Junliang Chen, Xianxu Hou, Xiaodong Zhao, and Linlin Shen. C2am: contrastive learning of class-agnostic activation map for weakly supervised object localization and semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989–998, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 288, + 557 + ], + "type": "text", + "content": "[57] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18134-18144, 2022. 3, 6, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 613 + ], + "type": "text", + "content": "[58] Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Yi Wang, Yu Qiao, and Weidi Xie. Learning open-vocabulary semantic segmentation models from natural language supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2935-2944, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "type": "text", + "content": "[59] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xiang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 714 + ], + "type": "text", + "content": "[60] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 1, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 217 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[61] B. Zhou, Z. Hang, Francesco Xavier Puig Fernandez, S. Fidler, and A. Torralba. Scene parsing through ade20k dataset. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 172 + ], + "type": "text", + "content": "[62] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVIII, pages 696-712. Springer, 2022. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 172, + 546, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 546, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 546, + 217 + ], + "type": "text", + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4112" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_content_list.json b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fa2bbe2cbd3bd0f97ccb5aae5434942c0beaa15a --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_content_list.json @@ -0,0 +1,1800 @@ +[ + { + "type": "text", + "text": "Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans", + "text_level": 1, + "bbox": [ + 142, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Romain Loiseau1,2", + "bbox": [ + 98, + 180, + 251, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "romain.loiseau@enpc.fr", + "bbox": [ + 89, + 204, + 264, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elliot Vincent1,3", + "bbox": [ + 321, + 180, + 450, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "elliot.vincent@enpc.fr", + "bbox": [ + 300, + 204, + 473, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mathieu Aubry", + "bbox": [ + 527, + 180, + 656, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mathieu.aubry@enpc.fr", + "bbox": [ + 509, + 205, + 676, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Loic Landrieu $^{1,2}$", + "bbox": [ + 728, + 180, + 864, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "loic. landrieu@enpc.fr", + "bbox": [ + 714, + 205, + 880, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France", + "bbox": [ + 238, + 226, + 728, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France", + "bbox": [ + 274, + 244, + 694, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 INRIA Paris, France", + "bbox": [ + 393, + 265, + 571, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 318, + 313, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose an unsupervised method for parsing large 3D scans of real-world scenes with easily-interpretable shapes. This work aims to provide a practical tool for analyzing 3D scenes in the context of aerial surveying and mapping, without the need for user annotations. Our approach is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical 3D shapes. The resulting reconstruction is visually interpretable and can be used to perform unsupervised instance and low-shot semantic segmentation of complex scenes. We demonstrate the usefulness of our model on a novel dataset of seven large aerial Li-DAR scans from diverse real-world scenarios. Our approach outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. Our code and dataset are available at https://romainloiseau.fr/learnable-earth-parser/.", + "bbox": [ + 75, + 349, + 473, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 648, + 209, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern aerial 3D scanning technologies open up unprecedented opportunities for environmental monitoring and economic intelligence. However, their practical use remains challenging due to the complexity of real-world scenes, the diversity of usage scenarios, and the difficulty of annotation. Therefore, our aim is to develop an approach that could help perform diverse tasks—from counting trees in a forest or identifying the various components of a factory to measuring the surface of greenhouses or monitor urban growth—all without human supervision.", + "bbox": [ + 75, + 674, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To do so, we address two important limitations of existing 3D deep learning methods. First, they are often primarily designed, trained, and tested on synthetic [7, 46, 59, 72] or highly curated data [2, 30, 39], which fail to capture the endless variability of the real world. Moreover, they often", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8501e7e11e85de94600e8c49e4cab8f1f2be7c828e79f17b6869bedaeebe3cf2.jpg", + "image_caption": [ + "Figure 1. Learnable Earth Parser. Our unsupervised method takes large aerial 3D scans as input and model them with a small set of learned 3D prototypes. Our approach is trained without annotation and produce legible decompositions of complex scenes, which can be used for semantic and instance segmentation." + ], + "image_footnote": [], + "bbox": [ + 501, + 320, + 901, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "assume that annotations are available for tasks of interest. Second, even unsupervised approaches [1, 77] often rely on learning abstract feature representations, making them difficult to interpret [76]. Although some work has attempted to decompose 3D shapes into meaningful components without supervision [11, 43, 53, 67], they were all designed on simple synthetic shapes and none generalizes to real data.", + "bbox": [ + 496, + 763, + 893, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, we present the Learnable Earth Parser, an unsupervised deep learning method", + "bbox": [ + 498, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27874", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8ff0cd769002fc64a16aecb1caf1d09aa06d61cd036dba3a5c47ff5ee233c19.jpg", + "image_caption": [ + "Figure 2. Method Overview. Our model approximates an input point cloud $\\mathbf{X}$ with $S$ slot models. Each slot maps $\\mathbf{X}$ to an affine 3D deformation $\\mathcal{T}_s(\\mathbf{X})$ , a slot activation probability $\\alpha_{s}$ , and the joint probabilities $\\beta_s^1,\\dots,\\beta_s^K$ of the slot being activated and choosing one of the $K$ learnable prototype point clouds $\\mathbf{P}^1,\\dots,\\mathbf{P}^K$ . The output $\\mathcal{M}_s(\\mathbf{X})$ of an activated slot $s$ is obtained by applying the transformation $\\mathcal{T}_s(\\mathbf{X})$ to its most likely prototype. Non-activated slots do not contribute to the output." + ], + "image_footnote": [], + "bbox": [ + 119, + 85, + 851, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "designed to decompose large-scale 3D point clouds into interpretable parts. Our model learns a small set of 3D prototypical shapes that are selected, positioned, rotated, and resized to reconstruct an input point cloud. We introduce a novel probabilistic formulation that enables the design of a reconstruction loss for learning jointly the 3D prototypes, but also to select and position them.", + "bbox": [ + 75, + 314, + 470, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our approach, we created a new open-access dataset consisting of 7 aerial LiDAR scans, covering $7.7\\mathrm{km}^2$ and containing 98 million 3D points with annotations in diverse urban and natural environments. Our results demonstrate that the Learnable Earth Parser learns decompositions superior to traditional and deep learning baselines, leading to convincing performance for semantic and instance segmentation, as shown in Figure 1. We believe that our contributions provide researchers and practitioners with new tools and resources to tackle the challenges of real-world 3D data.", + "bbox": [ + 75, + 420, + 468, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 599, + 215, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed unsupervised method uses point cloud reconstruction as a proxy to learn to decompose large aerial point clouds and is evaluated on a novel and diverse dataset of 3D scans. In the following, we briefly present related works for primitive-based point cloud decomposition, automatic decomposition of LiDAR data, and an overview of existing aerial LiDAR datasets.", + "bbox": [ + 75, + 625, + 468, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Primitive-based point cloud decomposition. Modeling shapes as a set of primitives such as generalized cylinders [5] or superquadrics [3] has a rich history in vision and graphics [26]. Classical applications include reverse engineering [4], shape completion [61, 66], and shape editing [14]. A variety of methods have been developed to find primitives in unstructured 3D scenes, including seed growing techniques [33, 34], genetic algorithms [8], approaches [17, 38, 56, 60] based on RANSAC [13], and probabilistic methods [40, 71].", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For this problem, like for many others in computer vision, deep learning has become the dominant paradigm. However, supervised methods [36, 80] are limited by the availability of annotated datasets. Following the seminal work of Tulsiani et al. [67], unsupervised approaches that simply rely on a reconstruction loss to learn primitive decomposition are the most common and the most closely related to our work [53, 55]. An important challenge for these approaches is to model a variable number of primitives. This has been addressed using recurrent networks [35, 62, 80], capsule networks [78], reinforcement learning strategies [67] or, most similar to our approach, computing the Chamfer distance based on a probabilistic model [53]. Our work is also related to approaches that learn prototypical shapes instead of being restricted to a predefined family of parametric primitives [11, 43].", + "bbox": [ + 496, + 314, + 890, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, most of these methods are designed, trained and evaluated on well-curated synthetic object datasets, such as ModelNet [72], ShapeNet [7], or D-FAUST [6], and are typically designed to handle single objects from known categories. In contrast, our approach can handle complex scenes composed of many objects with significant variety.", + "bbox": [ + 496, + 556, + 890, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Decomposition of LiDAR scans. Automatically decomposing large LiDAR scans poses unique challenges due to their size and diversity [73]. Some previous approaches use simple shape primitives, such as lines [10, 21], planes [16, 19, 51], or volumes [37], but these may not be flexible enough to capture the complexity of real-world scans. Other approaches are designed for specific object classes, like trees [47] or buildings [24, 32], but they are limited in their ability to represent a wide range of shapes. In contrast, our Learnable Earth Parser overcomes these limitations by learning ad-hoc prototypes for each new scene, ensuring both expressivity and adaptability. LiDAR data are often treated as digital elevation models, i.e. images with pixel elevations [20, 23, 42]. Thus, our work is related to image-based primitive prediction [22, 29, 54, 57] and unsupervised", + "bbox": [ + 496, + 672, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-object image segmentation [41, 49, 64, 75]. However, 3D point clouds have higher precision and can better represent multi-layered structures such as forest areas.", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aerial LiDAR datasets. The increased availability of aerial LiDAR technology has led to the multiplication of open datasets [15, 52, 74, 79] of varying sizes from 1 to 10 $\\mathrm{km}^2$ [63, 68]. However, these scans are limited to dense urban environments and do not capture the challenge of modeling diverse terrains. Some specialized datasets focus on forested areas [27, 70]. Our proposed dataset is of similar scale, spanning $7.7\\mathrm{km}^2$ , but covers a variety of urban, natural, and rural scenes, making it more representative of the diversity of possible usage scenarios.", + "bbox": [ + 75, + 156, + 470, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 321, + 168, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to learn to break down a point cloud into simpler and more easily understandable components. To achieve this, we propose an analysis-by-synthesis approach where we train a highly-constrained model $\\mathcal{M}$ to approximate a point cloud as a combination of learned 3D shapes.", + "bbox": [ + 75, + 347, + 468, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Probabilistic Scene Reconstruction Model", + "text_level": 1, + "bbox": [ + 76, + 431, + 433, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our model first selects up to $S$ shapes from $K$ learnable 3D shapes, then positions and deforms them to best approximate an input point cloud $\\mathbf{X}$ . We propose a probabilistic formulation of the selection process, which can be seen as an extension of the model of Paschalidou et al. [53] to multiple free-form shapes instead of a single parametric family.", + "bbox": [ + 75, + 455, + 468, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learnable shape prototypes. Following Loiseau et al. [43], we define $K$ point clouds $\\mathbf{P}^1,\\dots ,\\mathbf{P}^K$ that we refer to as prototypes. Each prototype is meant to represent a single instance of a recurring 3D structure in the considered scene. The points' coordinates are free parameters of the model and learned directly.", + "bbox": [ + 75, + 580, + 468, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene reconstruction model. Our full model $\\mathcal{M}$ is the combination of $S$ reconstruction models $\\mathcal{M}_s$ , which we refer to as slots in analogy to the Slot Attention approach [41]. Each slot contributes to the final reconstruction only if it is activated. Slot activation is determined by a binary variable $a_s$ : $\\mathcal{M}_s$ is activated if and only if $a_s = 1$ . The output of $\\mathcal{M}(\\mathbf{X})$ is the combination of the reconstructions from all activated slots:", + "bbox": [ + 75, + 691, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\mathbf {X}) = \\bigcup_ {\\substack {s = 1 \\dots S \\\\ a _ {s} = 1}} \\mathcal {M} _ {s} (\\mathbf {X}). \\tag{1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 824, + 468, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each slot model outputs a point cloud which is the deformation of one learnable prototype chosen from $\\mathbf{P}^1, \\dots, \\mathbf{P}^K$ .", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea9845e46b2a4f8fbafe1137999f9fb3b950160164bc24602c779f64c39471cd.jpg", + "image_caption": [ + "Figure 3. Earth Parser Dataset. Our dataset contains 7 scenes representing various urban and natural environments acquired by aerial LiDAR. The illustration of the power plant and the greenhouses display the complete scenes, while other ones display a subset of each scene (between 25 and $50\\%$ of the total area)." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 890, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We associate to each slot $s$ a network $\\mathcal{T}_s$ which maps $\\mathbf{X}$ to an affine transformation in 3D space $\\mathcal{T}_s(\\mathbf{X})$ . The output $\\mathcal{M}_s(\\mathbf{X})$ of slot $s$ is determined by a variable $b_{s} \\in \\{1, \\dots, K\\}$ . If $b_{s} = k$ , then the output of $\\mathcal{M}_s(\\mathbf{X})$ is $\\mathbf{Y}_s^k$ , the result of applying the transformation $\\mathcal{T}_s(\\mathbf{X})$ to the prototype $\\mathbf{P}^k$ :", + "bbox": [ + 498, + 375, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {s} ^ {k} = \\mathcal {T} _ {s} (\\mathbf {X}) [ \\mathbf {P} ^ {k} ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 474, + 890, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Please note that $\\mathbf{Y}_s^k$ is a function of $\\mathbf{X}$ . However, to keep our notations simple, we omit this dependence.", + "bbox": [ + 498, + 506, + 890, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Probabilistic modeling. We make our reconstruction model probabilistic by modeling $a$ and $b$ as random variables following (multi-)Bernoulli distributions. We call $\\alpha_{s}$ the probability that slot $s$ is activated and $\\beta_{s}^{k}$ the probability that it is activated and selects the prototype $k$ :", + "bbox": [ + 498, + 554, + 890, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(a _ {s} = 1\\right) = \\alpha_ {s}, \\quad p \\left(a _ {s} = 1, b _ {s} = k\\right) = \\beta_ {s} ^ {k}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 641, + 890, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each slot $s$ , we predict the vector $(1 - \\alpha_{s},\\beta_{s}^{1},\\dots \\beta_{s}^{K})$ with a neural network taking the point cloud $\\mathbf{X}$ as input and finishing with a softmax layer. Again, we don't write the dependency of the $\\alpha_{s}$ and $\\beta_{s}$ on $\\mathbf{X}$ explicitly to simplify the notations. The complete model $\\mathcal{M}(\\mathbf{X})$ and the slots models $\\mathcal{M}_s(\\mathbf{X})$ can now be seen as random variables, producing different potential reconstructions with probabilities given by $\\alpha$ and $\\beta$ . During inference, we consider only slots with $\\alpha_{s} > 0.5$ and select the prototype with highest $\\beta_s^k$ . However, during training, we compute all reconstructions $\\mathbf{Y}_s^k$ .", + "bbox": [ + 498, + 670, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Training Losses", + "text_level": 1, + "bbox": [ + 500, + 832, + 658, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a large 3D scene, we train our model by sampling square patches $\\mathbf{X}$ from the scene. For each batch of patches, we minimize a loss composed of a reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27876", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and several regularization terms $\\mathcal{L}_{\\mathrm{reg}}$ implementing different priors:", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {M}) = \\mathbb {E} _ {\\mathbf {X}} \\left[ \\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) \\right] + \\mathcal {L} _ {\\text {r e g}} (\\mathcal {M}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 128, + 468, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction loss. We define the reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ as the sum of two losses:", + "bbox": [ + 76, + 162, + 468, + 191 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) = \\mathcal {L} _ {\\text {a c c}} (\\mathcal {M}, \\mathbf {X}) + \\mathcal {L} _ {\\text {c o v}} (\\mathcal {M}, \\mathbf {X}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 199, + 468, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{acc}}$ encourages likely reconstructions of $\\mathcal{M}(\\mathbf{X})$ to accurately approximate $\\mathbf{X}$ , and $\\mathcal{L}_{\\mathrm{cov}}$ ensures coverage, i.e., that each point of $\\mathbf{X}$ is well-reconstructed by at least one activated model. We define each term using the asymmetric Chamfer distance $d$ between two point clouds $\\mathbf{X}$ and $\\mathbf{Y}$ :", + "bbox": [ + 76, + 223, + 468, + 299 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd (\\mathbf {X}, \\mathbf {Y}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\min _ {y \\in \\mathbf {Y}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 316, + 468, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $|.|$ the number of points in a point cloud. We write $d(x, \\mathbf{Y})$ the distance between the point $x$ and its closest point in $\\mathbf{Y}$ .", + "bbox": [ + 76, + 356, + 468, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define $\\mathcal{L}_{\\mathrm{acc}}$ as the average over all slots $s$ of the expected distance between $\\mathcal{M}_s(\\mathbf{X})$ and $\\mathbf{X}$ :", + "bbox": [ + 76, + 401, + 468, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a c c}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {a _ {s}, b _ {s}} [ d (\\mathcal {M} _ {s} (\\mathbf {X}), \\mathbf {X}) ] (7) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(\\mathbf {Y} _ {s} ^ {k}, \\mathbf {X}\\right). (8) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 439, + 468, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conversely, we define $\\mathcal{L}_{\\mathrm{cov}}$ as the average over all points $x$ of $\\mathbf{X}$ of the expected distance between $x$ and its closest point in the reconstruction:", + "bbox": [ + 76, + 520, + 468, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\mathbb {E} _ {a, b} \\left[ \\min _ {s \\mid a _ {s} = 1} d (x, \\mathcal {M} _ {s} (\\mathbf {X})) \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 570, + 468, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the ideas of Paschalidou et al. [53], we first define $\\Delta (x,s)$ as the expected distance between $x$ and $\\mathcal{M}_s(\\mathbf{X})$ conditionally to the slot $s$ being activated:", + "bbox": [ + 76, + 612, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta (x, s) = \\mathbb {E} _ {b _ {s} | a _ {s} = 1} [ d (x, \\mathcal {M} _ {s} (\\mathbf {X})) ] (10) \\\\ = \\frac {1}{\\alpha_ {s}} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(x, \\mathbf {Y} _ {s} ^ {k}\\right). (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 664, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we compute for each point $x$ a permutation $\\sigma_x$ of $[1, S]$ such that $\\Delta(x, \\sigma_x(s))$ is non-decreasing, i.e.:", + "bbox": [ + 76, + 727, + 468, + 757 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta (x, \\sigma_ {x} (1)) \\leq \\dots \\leq \\Delta (x, \\sigma_ {x} (S)). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 763, + 468, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If $s$ is the closest activated slot to $x$ , then all the slots closer to $x$ must be deactivated. This observation leads us to rewrite $\\mathcal{L}_{\\mathrm{cov}}$ as follows:", + "bbox": [ + 76, + 787, + 468, + 833 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\sum_ {s = 1} ^ {S} \\Delta (x, s) \\alpha_ {s} \\prod_ {r < \\sigma_ {x} (s)} \\left(1 - \\alpha_ {\\sigma_ {x} ^ {- 1} (r)}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 839, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\sigma^{-1}$ the inverse permutation of $\\sigma$ .", + "bbox": [ + 76, + 883, + 331, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularization losses. As is often necessary with fully unsupervised reconstruction methods [43, 48, 49], we define several regularization losses implementing priors on the model output to prevent degenerate local minima:", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- To encourage the slot activation to be sparse, we use the following loss penalizing slot activation:", + "bbox": [ + 500, + 152, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a c t}} (\\mathcal {M}) = \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 191, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- To avoid slots and prototypes that are never used, we use the following losses, which we compute batch-wise:", + "bbox": [ + 500, + 234, + 890, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s l o t}} (\\mathcal {M}) = - \\sum_ {s = 1} ^ {S} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {s} ]}{\\sum_ {t = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {t} ]}, \\epsilon_ {S}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 273, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r o t o}} (\\mathcal {M}) = - \\sum_ {k = 1} ^ {K} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} \\left[ \\sum_ {s = 1} ^ {S} \\beta_ {s} ^ {k} \\right]}{\\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]}, \\epsilon_ {K}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 313, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\epsilon_S$ and $\\epsilon_K$ hyperparameters setting the smallest acceptable relative use frequency for a slot or prototype.", + "bbox": [ + 500, + 362, + 888, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The full regularization loss is a weighted sum of the three losses described above:", + "bbox": [ + 500, + 392, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {a c t}} \\mathcal {L} _ {\\text {a c t}} + \\lambda_ {\\text {s l o t}} \\mathcal {L} _ {\\text {s l o t}} + \\lambda_ {\\text {p r o t o}} \\mathcal {L} _ {\\text {p r o t o}}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 431, + 890, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use $\\lambda_{\\mathrm{act}} = 10^{-4}$ , $\\lambda_{\\mathrm{slot}} = \\lambda_{\\mathrm{proto}} = 0.1$ and $\\epsilon_S = \\epsilon_K = 0.1$ in all our experiments on the Earth Parser Dataset.", + "bbox": [ + 500, + 457, + 890, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training and Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 497, + 820, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model configuration. We process the input point cloud $\\mathbf{X}$ using a similar architecture as in [53]. We first voxelize it with a $64\\times 64\\times 64$ grid and map it to a vector using a sequence of 6 3D sparse convolutions [9] and 6 strided convolutions. The resulting representation is then transformed using one linear layer for each slot. These features are decoded by simple 2-layer MLPs: one generates the distribution parameters $\\alpha_{s}$ and $\\beta_s^k$ , and the other ones the parameters of the 3D transformations $\\mathcal{T}_s(X)$ . The transformations include an anisotropic scaling, a $y$ -axis tilt of $\\pm \\pi /10$ , a rotation around the $z$ -axis, and a translation, in this specific order. We use $S = 64$ slots and $K = 6$ prototypes as default parameters. See the supplementary material for details.", + "bbox": [ + 498, + 520, + 890, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The intensity of the return signal of each point is available in the LiDAR scans. We associate each prototype with a single learnable intensity parameter and perform the Chamfer distance (Equation 6) in 4 dimensions: spatial coordinates normalized to $[0,1]^3$ and intensity to $[0,0.1]$ .", + "bbox": [ + 498, + 715, + 890, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Curriculum learning. The network predicts simultaneously the slot's probability distributions and their deformations. This results in many concurrent degrees of freedom and can make the training process unstable. Therefore, following Monnier et al. [49] and Loiseau et al. [43], we implement a multi-stage curriculum learning strategy. We first", + "bbox": [ + 498, + 809, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27877", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1922eaf39f98874f35a5cbc8f06007a23c6a9c6d444bf126e65343b5b5ddd33b.jpg", + "image_caption": [ + "Figure 4. Reconstruction Quality. We show two partial scenes with their RGB and intensity values, as well as their reconstruction by our method and competing models. We use the prototypes' intensity to color the points or pixels. As SuperQuadrics does not model the intensity, we use a random colour for each quadric." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 903, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "initialize the prototypes as point clouds uniformly sampled from a random cuboid and gradually unfreeze the model parameters in the following order: (i) translation, rotation, tilt, slot activation, and choice of prototype; (ii) intensities of the prototypes, when available; (iii) scales of the prototypes; (iv) positions of the prototypes' 3D points; (v) anisotropic scalings of the prototypes. As shown in Section 4, each step of this curriculum scheme improves the performances.", + "bbox": [ + 75, + 340, + 472, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prototypes selection. We automatically select the number of prototypes for a complete scene using a simple greedy algorithm. We measure the increase of reconstruction loss when preventing the model from selecting each prototype individually. We remove the prototype with the lowest increase if it is lower than $5\\%$ , and iterate.", + "bbox": [ + 75, + 479, + 468, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our model is trained separately for each scene by randomly sampling square patches. During training, the patches are subsampled to a maximum $10^{5}$ points. Each stage of the curriculum is trained until convergence. We use the ADAM optimizer [28] with a learning rate of $10^{-4}$ and default parameters. See the supplementary material for preprocessing, training and evaluation details.", + "bbox": [ + 75, + 588, + 468, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 707, + 163, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we assess the ability of our method to parse complex 3D aerial data. We give in Section 4.1 an overview of our proposed dataset of aerial LiDAR scans. In Section 4.2, we then discuss our evaluation metrics and baselines. Finally, we present a quantitative (Section 4.3) and qualitative (Section 4.4) analysis of our results.", + "bbox": [ + 75, + 732, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Earth Parser Dataset", + "text_level": 1, + "bbox": [ + 76, + 832, + 277, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a new dataset to train and evaluate parsing methods on large, uncurated aerial LiDAR scans. We use data from the French Mapping Agency associated to the", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LiDAR-HD project [45]. Each scan is composed of several airborne LiDAR acquisitions taken at different angles, leading to a minimum resolution of 20 points/ $\\mathrm{m}^2$ . The points are associated with their laser reflectance (intensity), and colorized based on asynchronous aerial photography.", + "bbox": [ + 496, + 340, + 890, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We selected 7 scenes, covering over $7.7\\mathrm{km}^2$ and a total of 98 million 3D points, with diverse content and complexity, such as dense habitations, forests, or complex industrial facilities. We associate most 3D points with a coarse semantic label, such as ground, building, or vegetation. The characteristics of the scenes are detailed in Table 2 and each is visualized in Figure 3.", + "bbox": [ + 496, + 417, + 892, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics and Baselines", + "text_level": 1, + "bbox": [ + 498, + 535, + 794, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We quantitatively evaluate the performances for reconstruction and semantic segmentation of our model and several unsupervised scene decomposition approaches.", + "bbox": [ + 496, + 559, + 890, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. As our goal is to summarize a point cloud using few prototypes, the quality of the reconstruction is critical. We measure it with the symmetric Chamfer distance (\"Cham.\" in the Tables) between the input and the output point clouds of our model.", + "bbox": [ + 496, + 627, + 890, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By associating a class to each prototype's points, we can propagate labels from the reconstruction to the input cloud and perform semantic segmentation. We evaluate the quality of this segmentation with the class-averaged Intersection-over-Union (mIoU) metric.", + "bbox": [ + 496, + 704, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a practical scenario, an operator can manually annotate the points of the 3D prototypes, allowing for the segmentation of the entirety of $\\mathbf{X}$ with minimal effort. To perform automatic evaluation, we follow the standard practice for evaluating clustering and unsupervised segmentation methods [25, 31, 48]: we assign to each prototype's point the most frequent class of its closest point of the input after the reconstruction.", + "bbox": [ + 496, + 780, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27878", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bd640d8a62a91437b3b7e4af311a2c6cac59a0860c115a72b77a61a33453f4bb.jpg", + "table_caption": [ + "Table 1. Results on the Earth Parser Dataset. We report the quality of the reconstruction (Cham.) and semantic segmentation (mIoU) on each of the scenes of our Earth Parser Dataset. While our method does not always provide the most faithful reconstructions, it leads to the most accurate point classification." + ], + "table_footnote": [], + "table_body": "
Rec.SemanticCrop FieldsForestGreenhousesMarinaPower PlantUrbanWindturbines
Cham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoU
k-means (i,z) [44]X✓—93.8—71.5—39.3—41.4—42.8—56.5—87.6
SuperQuadrics [53]3DX0.86—1.04—0.60—0.93—0.58—0.40—13.5—
DTI-Sprites [49]2.5D+i✓6.1083.214.5940.25.3642.06.1641.45.3629.02.9947.336.1925.9
AtlasNet v2 [11]3D+i✓1.0743.11.5871.40.5649.10.7342.10.4541.60.6348.89.4748.1
Ours3D+i✓0.7296.90.8883.70.4091.30.8278.70.4452.20.2983.26.6593.4
", + "bbox": [ + 84, + 141, + 890, + 253 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/147271780cc9be1a8be836706a8c18b124a941d29f7e964d6a2d110e5e2f72b6.jpg", + "table_caption": [ + "Table 2. Earth Parser Dataset. Our proposed dataset is composed of 7 diverse scenes acquired by aerial LiDAR." + ], + "table_footnote": [], + "table_body": "
NameSurface in km2# points ×106annotation ratio in %num. of classes
Crop Fields1.119.777.42
Forest1.146.797.82
Greenhouses0.11.395.63
Marina0.10.592.72
Power Plant0.28.678.44
Urban1.115.795.93
Wind turbines4.25.699.83
Total7.798.391.6—
", + "bbox": [ + 76, + 303, + 470, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We adapt several unsupervised approaches for scene reconstruction and/or semantic segmentation tasks to provide comparisons for our approach:", + "bbox": [ + 75, + 492, + 468, + 537 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- k-means. We cluster the points of the input with the k-means algorithm [44] using as many clusters as we use prototypes. We obtain the best results by using a combination of the point's intensity and elevation as features for clustering. We then assign to each centroid its most frequent class, and propagate this label to the entire cluster, leading to a semantic segmentation. This method does not reconstruct the input, but gives us a simple and surprisingly strong baseline for semantic segmentation.", + "- SuperQuadrics revisited. We use the method of Paschalidou et al. [53] to learn to approximate scenes with an adaptive number of superquadrics [3]. It provides a baseline for reconstruction and a qualitative comparison for instance segmentation, shown in supplementary material.", + "- DTI-Sprites. We use the point cloud to construct a digital elevation model, i.e. a 2.5D image of resolution $32 \\times 32$ where each pixel has an elevation and intensity value. We adapt the unsupervised image decomposition approach of Monnier et al. [49] to break down this image into a set of 2.5D sprites. We evaluate the reconstruction and segmentation by sampling 25 3D point per pixel, transferring the pixel's label to the points, and interpolating their elevations.", + "- AtlasNet v2. This extension [11] of AtlasNet [18] uses a fixed number of learnable prototype point clouds to recon" + ], + "bbox": [ + 75, + 537, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/991f28f9223c412d48668a4e142f8063feb07bd41c2ba30e2b1606f64bf24886.jpg", + "table_caption": [ + "Table 3. Ablation Study. We evaluate the effect of our prototype selection post-processing, our model's degrees of freedom, and our different regularization losses." + ], + "table_footnote": [], + "table_body": "
UrbanMarina
Cham.mIoUCham.mIoU
expressivityLearnable Earth Parser0.2983.20.8278.7
→w/o post-processing0.2883.70.9678.3
→w/o aniso-scale0.3382.41.0467.2
→w/o prototypes0.3668.31.0742.8
→w/o scales0.5558.91.3340.8
→w/o intensities0.5558.71.0940.8
losses→w/o Lact0.1754.10.8056.9
→w/o Lslot0.2577.80.8143.7
→w/o Lproto0.2857.20.9740.7
", + "bbox": [ + 501, + 316, + 890, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "struct its input. It can be evaluated for both reconstruction and semantic segmentation in a way similar to ours. We extend it to handle intensity in a manner akin to our approach, which improves its segmentation results.", + "bbox": [ + 496, + 547, + 890, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to our method, we train all baselines except k-means by sampling square patches in each scene. Figure 4 shows the output of the reconstruction methods.", + "bbox": [ + 496, + 608, + 890, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 500, + 662, + 692, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of our approach with the proposed baselines on the Earth Parser Dataset, as well as two publicly available datasets.", + "bbox": [ + 496, + 686, + 890, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Earth Parser Dataset. We provide quantitative reconstruction and semantic segmentation results in Table 1, and illustrations in Figure 4. Despite being highly constrained, our model yields the best reconstruction in 6 out of 7 scenes. Moreover, we significantly outperform the other methods for semantic segmentation across all scenes.", + "bbox": [ + 496, + 750, + 890, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite its simplicity, the k-means baseline provides strong semantic segmentation performance, beating the other baselines in 5 of 7 annotated scenes. DTI-Sprites [49] has lower reconstruction and segmentation quality, which", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27879", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/461aef78b71a06e44745f68ff5e9e1a1d0936d1761b282cb4a6b6f0626686597.jpg", + "table_caption": [ + "Table 4. Synthetic Shapes. We train our method on all planes from ShapeNet-Part [58], with random rotations around $z$ -axis. We show the reconstruction of an input plane and the prototypes learned on the dataset." + ], + "table_footnote": [], + "table_body": "
AtlasNet v2 [11]SuperQuadrics [53]Ours
Recons.
Protos.---
Cham.1.462.911.34
mIoU34.5-68.6
", + "bbox": [ + 76, + 155, + 470, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9eb4d67c3081644fac27d3682de68df885118fbb50e80eccfb81eab181c3104d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cham.mIoU
k-means (i,z) [44]—52.7
Superquadrics [53]0.33—
DTI-Sprites [49]2.2553.6
AtlasNet v2 [11]0.3154.9
Ours0.2982.7
", + "bbox": [ + 76, + 316, + 282, + 425 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0e64fd1741ce88e0d62b422b28e7456390c76b85e6646b1463fc89c09278a09.jpg", + "image_caption": [ + "Figure 5. Results on DALES [68]. We report quantitative and qualitative results for one tile from DALES." + ], + "image_footnote": [], + "bbox": [ + 294, + 318, + 473, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is expected as it models a 3D point cloud in 2.5D. AtlasNet v2 [11] provides good reconstructions but segmentation fails for scenes such as Crop Fields or Urban due its inability to adapt its prototype usage to the input. On the contrary, SuperQuadrics [53] can adjust the number of superquadric it uses and, to some degree, their shape. However, this method uses a single parametric family for all prototypes and fails to reconstruct complex real-world scenes such as Power Plant. Thanks to its probabilistic slot selection, our method can handle inputs with a varying number of objects using only a small set of learned prototypical shapes.", + "bbox": [ + 75, + 479, + 468, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study. We evaluate the impact of various components of our model and report the results in Table 3. First, we observe that the prototype selection post-processing has limited impact on the quality of the prediction and reconstructions, but allows us to adapt and significantly decrease the number of prototypes used for each scene. Second, we evaluate the impact of reducing the expressivity of our model. We successively remove: (i) the anisotropic scaling of $\\mathcal{T}$ , and the possibility of learning the prototype's (ii) points position, (iii) scale, and (iv) intensity. As expected, each degree if liberty removed decreases the quality of the 3D reconstruction and segmentation.", + "bbox": [ + 75, + 672, + 467, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the impact of the different regularization introduced in Section 3.2. The losses related to slot activation have a marginal effect on reconstruction quality but", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "significantly affect the performance of semantic segmentation. This suggests that using slots sparingly but equally is important for prototypes to specialize for specific objects, but is not necessary to the expressivity of the reconstruction model. The prototype activation loss prevents unused and possibly degenerate prototypes and therefore improves both segmentation and reconstruction.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the impact of reducing the number of prototypes or slots. When we set $K = 3$ instead of 6, the average semantic segmentation mIoU drops by 26.3 points, while reconstruction is minimally affected, showing only a $2.3\\%$ increase in Chamfer distance. Conversely, setting $S = 32$ instead of 64 leads to a significant $+47.3\\%$ increase of Chamfer distance, with a modest drop 0.4 of semantic segmentation mIoU. These findings highlight the importance of maintaining sufficient diversity in the prototypes to enable specialization for different object types, whereas the number of slots strongly influences the expressiveness of the reconstruction model.", + "bbox": [ + 496, + 196, + 892, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ShapeNet. We evaluate our model on 2690 planes from ShapeNet-Part [58], whose points are annotated as wing, engine, tail, or body. We randomly rotate the shapes around the $z$ -axis during training and evaluation. We report in Table 4 the performances and reconstructions for our approach, AtlasNet v2 [11], and SuperQuadrics revisited [53]. Our method handle rotations better than AtlasNet v2, and manages to successfully locate the tail of the planes. While SuperQuadrics' reconstructions make sense qualitatively, they are worse in terms of accuracy than ours, and do not enable semantic segmentation.", + "bbox": [ + 496, + 396, + 892, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DALES. We also train and evaluate our model on DALES [68], a dataset of aerial LiDAR scans of urban area, with a restricted class set: ground, vegetation, and buildings. Our model significantly outperforms competing reconstruction-based approaches, and yields similar performance than for the Urban tile of the Earth Parser Dataset; see Figure 5.", + "bbox": [ + 496, + 580, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 500, + 696, + 683, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instance segmentation. We can perform instance segmentation simply by considering each slot as a different instance. This is particularly interesting for parsing natural woodlands, a key endeavor for forest management [50] and biomass estimation [12]. While this task has a long history of handcrafted approaches [69], current deep learning approaches are mostly limited to artificial or low-density forests [65]. As shown qualitatively in Figure 6, our Learnable Earth Parser can learn without any supervision to separate individual trees in dense forests, or boats in a marina. We evaluate quantitatively the performance of our model in terms of instance segmentation, we counted manually 88", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27880", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/edd8816a7c438417c0fb5ddd49665ba2b19559f637662b1f52c3c0dc3af2ee50.jpg", + "image_caption": [ + "Figure 6. Instance Segmentation. We can identify the reconstruction of each slot as a separate instance, allowing us to perform instance segmentation of complex data such as dense natural forests or a marina. For this visualization, we considered the points associated to \"trees\" or \"boat hull\" prototypes and color each of their instance randomly." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 468, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "boat instances in 10 distinct portions of the Marina scene. We then computed the Mean Relative Error (MRE) of the prediction given by the number of boat-like prototypes (prototype #3 in Figure 8) in these zones. Our method yields a MRE of $7,4\\%$ , 2.6 times lower than DTI-Sprites [49], demonstrating its strong performances.", + "bbox": [ + 75, + 431, + 468, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Semantic segmentation. After annotating the prototypes, our model can perform convincing semantic segmentation of complex scenes, as shown in Figure 7.", + "bbox": [ + 75, + 542, + 468, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interpretable prototypes. In Figure 8, we show prototypes learned on our Earth Parser Dataset with colors showing the associated semantic label for each point. These prototypes give at a glance insights on the content of these real-world scenes. Our model is able to learn a wide variety of shapes, such as boats' masts, wind turbines, or greenhouses. We also observe that the prototypes are typically associated with a single object type. Empirically, the average class distribution of the nearest neighbors of each prototypes' point exhibits the same normalized entropy (0.22) as a Bernoulli distribution with probability 0.964. The highly constrained nature of the deformations $\\mathcal{T}$ prevents slots from repurposing the same prototypes for different objects, making the learned prototypes easy to identify and manually annotate.", + "bbox": [ + 75, + 604, + 470, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 829, + 194, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a novel unsupervised method for parsing complex real-world aerial scans into simple parts using a small set of learned prototypical shapes. We demonstrate", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d0f9e92889933ddcee42309ba4271befd917345318a3674e19da68a0fed51349.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 893, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a19f73eddc8fa3dec4ac0e97498ea4647b08baffbed6866f693077d4a985ad0.jpg", + "image_caption": [ + "Figure 7. Semantic Segmentation. Our model can perform semantic segmentation of large real-world scene based on annotated prototypes. Black points in the ground truth are unannotated.", + "Figure 8. Learned Prototypes. We display three of the prototypes chosen during the selection process for various scenes." + ], + "image_footnote": [], + "bbox": [ + 501, + 369, + 890, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the quality and interpretability of our results on a novel dataset of aerial LiDAR scans. To the best of our knowledge, we are the first to demonstrate the possibility of performing deep unsupervised 3D shape analysis on such a challenging real-world dataset. We believe that our results open new perspectives for computer-assisted environment monitoring and economic intelligence.", + "bbox": [ + 496, + 556, + 893, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported in part by ANR project READY3D ANR-19-CE23-0007, ANR under the France 2030 program under the reference ANR-23-PEIA-0008, and was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012096R2 made by GENCI. The work of MA was partly supported by the European Research Council (ERC project DISCOVER, number 101076028). The scenes of Earth Parser Dataset were acquired and annotated by the LiDAR-HD project [45]. We thank Zenodo for hosting the dataset. We thank Zeynep Sonat Baltaci, Emile Blettery, Nicolas Dufour, Antoine Guedon, Helen Mair Rawsthorne, Tom Monnier, Damien Robert, Mathis Petrovich and Yannis Siglidis for inspiring discussions and valuable feedback.", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antonio Alliegro, Davide Boscaini, and Tatiana Tommasi. Joint supervised and self-supervised learning for 3D real world challenges. In ICPR, 2021. 1", + "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1", + "[3] Alan H Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1981. 2, 6", + "[4] Pál Benkő, Ralph R Martin, and Tamás Várady. Algorithms for reverse engineering boundary representation models. Computer-Aided Design, 2001. 2", + "[5] Thomas Binford. Visual perception by computer. In Proc. IEEE Conf. on Systems and Control, 1975. 2", + "[6] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 2", + "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2", + "[8] YH Chen and CY Liu. Quadric surface extraction using genetic algorithms. Computer-Aided Design, 1999. 2", + "[9] Spconv Contributors. SPConv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4", + "[10] Yang Cui, Qingquan Li, Bisheng Yang, Wen Xiao, Chi Chen, and Zhen Dong. Automatic 3-d reconstruction of indoor environment with mobile laser scanning point clouds. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 2", + "[11] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. NeurIPS, 2019. 1, 2, 5, 6, 7", + "[12] Antonio Ferraz, Sassan Saatchi, Clément Mallet, Stéphane Jacquemoud, Gil Gonçalves, Carlos Alberto Silva, Paula Soares, Margarida Tomé, and Luisa Pereira. Airborne Li-DAR estimation of aboveground forest biomass in the absence of field inventory. Remote Sensing, 2016. 7", + "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 2", + "[14] Ran Gal, Olga Sorkine, Niloy J Mitra, and Daniel Cohen-Or. iWIRES: An analyze-and-edit approach to shape manipulation. In ACM SIGGRAPH. 2009. 2", + "[15] Weixiao Gao, Liangliang Nan, Bas Boom, and Hugo Ledoux. SUM: A benchmark dataset of semantic urban meshes. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 3", + "[16] Xuming Ge, Bo Wu, Yuan Li, and Han Hu. A multiprimitive-based hierarchical optimal approach for semantic labeling of ALS point clouds. Remote Sensing, 2019. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Paulo FU Gotardo, Olga Regina Pereira Bellon, and Luciano Silva. Range image segmentation by surface extraction using an improved robust estimator. In CVPR, 2003. 2", + "[18] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 6", + "[19] Stéphane Guinard, Loic Landrieu, Laurent Caraffa, and Bruno Vallet. Piecewise-planar approximation of large 3d data as graph-structured optimization. ISPRS Annals of Photogrammetry, Remote Sensing & Spatial Information Sciences, 2019. 2", + "[20] Florent Guiotte, Minh-Tan Pham, Romain Dambreville, Thomas Corpetti, and Sébastien Lefèvre. Semantic segmentation of LiDAR points clouds: Rasterization beyond digital elevation models. IEEE Geoscience and Remote Sensing Letters, 2020. 2", + "[21] Bo Guo, Qingquan Li, Xianfeng Huang, and Chisheng Wang. An improved method for power-line reconstruction from point cloud data. Remote sensing, 2016. 2", + "[22] Abhinav Gupta, Alexei Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In ECCV, 2010. 2", + "[23] Norbert Haala and Claus Brenner. Extraction of buildings and trees in urban environments. ISPRS journal of photogrammetry and remote sensing, 1999. 2", + "[24] Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3D: large-scale building reconstruction from airborne LiDAR point clouds. Remote Sensing, 2022. 2", + "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. CVPR, 2019. 5", + "[26] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3D data. In Computer Graphics Forum. Wiley Online Library, 2019. 2", + "[27] Ekaterina Kalinicheva, Loic Landrieu, Clément Mallet, and Nesrine Chehata. Multi-layer modeling of dense vegetation from aerial lidar scans. In CVPR Workshops, 2022. 3", + "[28] Diederik P Kingma and Jimmy Ba. ADAM: A method for stochastic optimization. *ICLR*, 2014. 5", + "[29] Florian Kluger, Hanno Ackermann, Eric Brachmann, Michael Ying Yang, and Bodo Rosenhahn. Cuboids revisited: Learning robust 3D shape fitting to single RGB images. In CVPR, 2021. 2", + "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 1", + "[31] Adam Kosiorek, Sara Sabour, Yee Whye Teh, and Geoffrey E Hinton. Stacked capsule autoencoders. NeurIPS, 2019. 5", + "[32] Florent Lafarge, Xavier Descombes, Josiane Zerubia, and Marc Pierrot-Deseilligny. Automatic 3D buildup reconstruction from DEMs: An application to PLEIADES simulations. In ISPRS, 2006. 2", + "[33] Aleš Leonardis, Alok Gupta, and Ruzena Bajcsy. Segmentation of range images as the search for geometric parametric models. International Journal of Computer Vision, 1995. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27882", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Ales Leonardis, Ales Jaklic, and Franc Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1997. 2", + "[35] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics, 2017. 2", + "[36] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3D point clouds. In CVPR, 2019. 2", + "[37] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In ECCV, 2016. 2", + "[38] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. In ACM SIGGRAPH. 2011. 2", + "[39] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1", + "[40] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In CVPR, 2022. 2", + "[41] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 3", + "[42] Suresh K Lodha, Edward J Kreps, David P Helmbold, and Darren Fitzpatrick. Aerial LiDAR data classification using support vector machines. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06). IEEE, 2006. 2", + "[43] Romain Loiseau, Tom Monnier, Mathieu Aubry, and Loic Landrieu. Representing shape collections with alignment-aware linear models. In 3DV, 2021. 1, 2, 3, 4", + "[44] J MacQueen. Classification and analysis of multivariate observations. In 5th Berkeley Symp. Math. Statist. Probability, 1967. 6, 7", + "[45] IGN (French mapping agency). LiDAR-HD: A 3D mapping of France's soil and subsoil, 2021. 5, 8", + "[46] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J.Davison. SceneNet RGB-D: Can 5M synthetic images beat generic imagenet pre-training on indoor segmentation? ICCV, 2017. 1", + "[47] Jie Mei, Liqiang Zhang, Shihao Wu, Zhen Wang, and Liang Zhang. 3d tree modeling from incomplete point clouds via optimization and 11-mst. International Journal of Geographical Information Science, 2017. 2", + "[48] Tom Monnier, Thibault Groueix, and Mathieu Aubry. Deep Transformation-Invariant Clustering. NeurIPS, 2020. 4, 5", + "[49] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In ICCV, 2021. 3, 4, 5, 6, 7, 8", + "[50] Felix Morsdorf, Erich Meier, Benjamin Kötz, Klaus I Itten, Matthias Dobbertin, and Britta Allgower. LIDAR-based geometric reconstruction of boreal type forest stands at single" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tree level for forest and wildland fire management. Remote sensing of environment, 2004. 7", + "[51] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In ICCV, 2017. 2", + "[52] Joachim Niemeyer, Franz Rottensteiner, and Uwe Soergel. Contextual classification of LiDAR data and building object detection in urban areas. ISPRS journal of photogrammetry and remote sensing, 2014. 3", + "[53] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, 2019. 1, 2, 3, 4, 5, 6, 7", + "[54] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In CVPR, 2020. 2", + "[55] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 2", + "[56] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. Monteboxfinder: Detecting and filtering primitives to fit a noisy point cloud. In ECCV. Springer, 2022. 2", + "[57] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2", + "[58] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR Workshop, 2015. 7", + "[59] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR 2015 Workshop on Functionality, Physics, Intentionality and Causality, 2015. 1", + "[60] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. In Computer Graphics Forum, 2007. 2", + "[61] Ruwen Schnabel, Patrick Degener, and Reinhard Klein. Completion and reconstruction with primitive shapes. In Computer Graphics Forum. Wiley Online Library, 2009. 2", + "[62] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 2", + "[63] Nina M Singer and Vijayan K Asari. DALES objects: A large scale benchmark dataset for instance segmentation in aerial LiDAR. IEEE Access, 2021. 3", + "[64] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 3", + "[65] Chenxin Sun, Chengwei Huang, Huaiqing Zhang, Bangqian Chen, Feng An, Liwen Wang, and Ting Yun. Individual tree crown segmentation and crown width extraction from a heightmap derived from aerial laser scanning data using a deep learning framework. Frontiers in plant science, 2022. 7", + "[66] Minhyuk Sung, Vladimir G Kim, Roland Angst, and Leonidas Guibas. Data-driven structural priors for shape completion. ACM Transactions on Graphics, 2015. 2", + "[67] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27883", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Nina Varney, Vijayan K Asari, and Quinn Graehling. DALES: A large-scale aerial LiDAR data set for semantic segmentation. In CVPR Workshop, 2020. 3, 7", + "[69] Jari Vauhkonen, Liviu Ene, Sandeep Gupta, Johannes Heinzel, Johan Holmgren, Juho Pitkänen, Svein Solberg, Yunsheng Wang, Holger Weinacker, K Marius Hauglin, et al. Comparative testing of single-tree detection algorithms under different types of forest. Forestry, 2012. 7", + "[70] Ben G Weinstein, Sergio Marconi, Stephanie Bohlman, Alina Zare, and Ethan White. Individual tree-crown detection in RGB imagery using semi-supervised deep learning neural networks. Remote Sensing, 2019. 3", + "[71] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In ECCV, 2022. 2", + "[72] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 2", + "[73] Shaobo Xia, Dong Chen, Ruisheng Wang, Jonathan Li, and Xinchang Zhang. Geometric primitives in LiDAR point clouds: A review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020. 2", + "[74] Zhen Ye, Yusheng Xu, Rong Huang, Xiaohua Tong, Xin Li, Xiangfeng Liu, Kuifeng Luan, Ludwig Hoegner, and Uwe Stilla. LASDU: A large-scale aerial LiDAR dataset for semantic labeling in dense urban areas. ISPRS International Journal of Geo-Information, 2020. 3", + "[75] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised Discovery of Object Radiance Fields. *ICLR*, 2022. 3", + "[76] Yu Zhang, Peter Tino, Aleks Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1", + "[77] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In ICCV, 2021. 1", + "[78] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3D point capsule networks. In CVPR, 2019. 2", + "[79] SM Zolanvari, Susana Ruano, Aakanksha Rana, Alan Cummins, Rogerio Eduardo da Silva, Morteza Rahbar, and Aljosa Smolic. DublinCity: annotated LiDAR point cloud and its applications. BMVC, 2019. 3", + "[80] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In ICCV, 2017. 2" + ], + "bbox": [ + 78, + 90, + 467, + 742 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27884", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_model.json b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a9922440e5573116693306ed4d31730863774f9a --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_model.json @@ -0,0 +1,2708 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.181, + 0.253, + 0.198 + ], + "angle": 0, + "content": "Romain Loiseau1,2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.205, + 0.266, + 0.218 + ], + "angle": 0, + "content": "romain.loiseau@enpc.fr" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.181, + 0.452, + 0.198 + ], + "angle": 0, + "content": "Elliot Vincent1,3" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.205, + 0.475, + 0.218 + ], + "angle": 0, + "content": "elliot.vincent@enpc.fr" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.181, + 0.658, + 0.199 + ], + "angle": 0, + "content": "Mathieu Aubry" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.206, + 0.678, + 0.218 + ], + "angle": 0, + "content": "mathieu.aubry@enpc.fr" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.181, + 0.865, + 0.198 + ], + "angle": 0, + "content": "Loic Landrieu\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.206, + 0.882, + 0.218 + ], + "angle": 0, + "content": "loic. landrieu@enpc.fr" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.227, + 0.73, + 0.245 + ], + "angle": 0, + "content": "1 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.246, + 0.696, + 0.264 + ], + "angle": 0, + "content": "2 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.266, + 0.572, + 0.283 + ], + "angle": 0, + "content": "3 INRIA Paris, France" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.319, + 0.314, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.474, + 0.624 + ], + "angle": 0, + "content": "We propose an unsupervised method for parsing large 3D scans of real-world scenes with easily-interpretable shapes. This work aims to provide a practical tool for analyzing 3D scenes in the context of aerial surveying and mapping, without the need for user annotations. Our approach is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical 3D shapes. The resulting reconstruction is visually interpretable and can be used to perform unsupervised instance and low-shot semantic segmentation of complex scenes. We demonstrate the usefulness of our model on a novel dataset of seven large aerial Li-DAR scans from diverse real-world scenarios. Our approach outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. Our code and dataset are available at https://romainloiseau.fr/learnable-earth-parser/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.649, + 0.21, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Modern aerial 3D scanning technologies open up unprecedented opportunities for environmental monitoring and economic intelligence. However, their practical use remains challenging due to the complexity of real-world scenes, the diversity of usage scenarios, and the difficulty of annotation. Therefore, our aim is to develop an approach that could help perform diverse tasks—from counting trees in a forest or identifying the various components of a factory to measuring the surface of greenhouses or monitor urban growth—all without human supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To do so, we address two important limitations of existing 3D deep learning methods. First, they are often primarily designed, trained, and tested on synthetic [7, 46, 59, 72] or highly curated data [2, 30, 39], which fail to capture the endless variability of the real world. Moreover, they often" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.321, + 0.902, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.679, + 0.895, + 0.75 + ], + "angle": 0, + "content": "Figure 1. Learnable Earth Parser. Our unsupervised method takes large aerial 3D scans as input and model them with a small set of learned 3D prototypes. Our approach is trained without annotation and produce legible decompositions of complex scenes, which can be used for semantic and instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.87 + ], + "angle": 0, + "content": "assume that annotations are available for tasks of interest. Second, even unsupervised approaches [1, 77] often rely on learning abstract feature representations, making them difficult to interpret [76]. Although some work has attempted to decompose 3D shapes into meaningful components without supervision [11, 43, 53, 67], they were all designed on simple synthetic shapes and none generalizes to real data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "To overcome these limitations, we present the Learnable Earth Parser, an unsupervised deep learning method" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27874" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.087, + 0.852, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.245, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Method Overview. Our model approximates an input point cloud \\(\\mathbf{X}\\) with \\(S\\) slot models. Each slot maps \\(\\mathbf{X}\\) to an affine 3D deformation \\(\\mathcal{T}_s(\\mathbf{X})\\), a slot activation probability \\(\\alpha_{s}\\), and the joint probabilities \\(\\beta_s^1,\\dots,\\beta_s^K\\) of the slot being activated and choosing one of the \\(K\\) learnable prototype point clouds \\(\\mathbf{P}^1,\\dots,\\mathbf{P}^K\\). The output \\(\\mathcal{M}_s(\\mathbf{X})\\) of an activated slot \\(s\\) is obtained by applying the transformation \\(\\mathcal{T}_s(\\mathbf{X})\\) to its most likely prototype. Non-activated slots do not contribute to the output." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.471, + 0.42 + ], + "angle": 0, + "content": "designed to decompose large-scale 3D point clouds into interpretable parts. Our model learns a small set of 3D prototypical shapes that are selected, positioned, rotated, and resized to reconstruct an input point cloud. We introduce a novel probabilistic formulation that enables the design of a reconstruction loss for learning jointly the 3D prototypes, but also to select and position them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.588 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our approach, we created a new open-access dataset consisting of 7 aerial LiDAR scans, covering \\(7.7\\mathrm{km}^2\\) and containing 98 million 3D points with annotations in diverse urban and natural environments. Our results demonstrate that the Learnable Earth Parser learns decompositions superior to traditional and deep learning baselines, leading to convincing performance for semantic and instance segmentation, as shown in Figure 1. We believe that our contributions provide researchers and practitioners with new tools and resources to tackle the challenges of real-world 3D data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.216, + 0.617 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.732 + ], + "angle": 0, + "content": "Our proposed unsupervised method uses point cloud reconstruction as a proxy to learn to decompose large aerial point clouds and is evaluated on a novel and diverse dataset of 3D scans. In the following, we briefly present related works for primitive-based point cloud decomposition, automatic decomposition of LiDAR data, and an overview of existing aerial LiDAR datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Primitive-based point cloud decomposition. Modeling shapes as a set of primitives such as generalized cylinders [5] or superquadrics [3] has a rich history in vision and graphics [26]. Classical applications include reverse engineering [4], shape completion [61, 66], and shape editing [14]. A variety of methods have been developed to find primitives in unstructured 3D scenes, including seed growing techniques [33, 34], genetic algorithms [8], approaches [17, 38, 56, 60] based on RANSAC [13], and probabilistic methods [40, 71]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.557 + ], + "angle": 0, + "content": "For this problem, like for many others in computer vision, deep learning has become the dominant paradigm. However, supervised methods [36, 80] are limited by the availability of annotated datasets. Following the seminal work of Tulsiani et al. [67], unsupervised approaches that simply rely on a reconstruction loss to learn primitive decomposition are the most common and the most closely related to our work [53, 55]. An important challenge for these approaches is to model a variable number of primitives. This has been addressed using recurrent networks [35, 62, 80], capsule networks [78], reinforcement learning strategies [67] or, most similar to our approach, computing the Chamfer distance based on a probabilistic model [53]. Our work is also related to approaches that learn prototypical shapes instead of being restricted to a predefined family of parametric primitives [11, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.892, + 0.65 + ], + "angle": 0, + "content": "However, most of these methods are designed, trained and evaluated on well-curated synthetic object datasets, such as ModelNet [72], ShapeNet [7], or D-FAUST [6], and are typically designed to handle single objects from known categories. In contrast, our approach can handle complex scenes composed of many objects with significant variety." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Decomposition of LiDAR scans. Automatically decomposing large LiDAR scans poses unique challenges due to their size and diversity [73]. Some previous approaches use simple shape primitives, such as lines [10, 21], planes [16, 19, 51], or volumes [37], but these may not be flexible enough to capture the complexity of real-world scans. Other approaches are designed for specific object classes, like trees [47] or buildings [24, 32], but they are limited in their ability to represent a wide range of shapes. In contrast, our Learnable Earth Parser overcomes these limitations by learning ad-hoc prototypes for each new scene, ensuring both expressivity and adaptability. LiDAR data are often treated as digital elevation models, i.e. images with pixel elevations [20, 23, 42]. Thus, our work is related to image-based primitive prediction [22, 29, 54, 57] and unsupervised" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27875" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "multi-object image segmentation [41, 49, 64, 75]. However, 3D point clouds have higher precision and can better represent multi-layered structures such as forest areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.157, + 0.471, + 0.309 + ], + "angle": 0, + "content": "Aerial LiDAR datasets. The increased availability of aerial LiDAR technology has led to the multiplication of open datasets [15, 52, 74, 79] of varying sizes from 1 to 10 \\(\\mathrm{km}^2\\) [63, 68]. However, these scans are limited to dense urban environments and do not capture the challenge of modeling diverse terrains. Some specialized datasets focus on forested areas [27, 70]. Our proposed dataset is of similar scale, spanning \\(7.7\\mathrm{km}^2\\), but covers a variety of urban, natural, and rural scenes, making it more representative of the diversity of possible usage scenarios." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.322, + 0.169, + 0.337 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Our goal is to learn to break down a point cloud into simpler and more easily understandable components. To achieve this, we propose an analysis-by-synthesis approach where we train a highly-constrained model \\(\\mathcal{M}\\) to approximate a point cloud as a combination of learned 3D shapes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.434, + 0.449 + ], + "angle": 0, + "content": "3.1. Probabilistic Scene Reconstruction Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.563 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our model first selects up to \\(S\\) shapes from \\(K\\) learnable 3D shapes, then positions and deforms them to best approximate an input point cloud \\(\\mathbf{X}\\). We propose a probabilistic formulation of the selection process, which can be seen as an extension of the model of Paschalidou et al. [53] to multiple free-form shapes instead of a single parametric family." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Learnable shape prototypes. Following Loiseau et al. [43], we define \\(K\\) point clouds \\(\\mathbf{P}^1,\\dots ,\\mathbf{P}^K\\) that we refer to as prototypes. Each prototype is meant to represent a single instance of a recurring 3D structure in the considered scene. The points' coordinates are free parameters of the model and learned directly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Scene reconstruction model. Our full model \\(\\mathcal{M}\\) is the combination of \\(S\\) reconstruction models \\(\\mathcal{M}_s\\), which we refer to as slots in analogy to the Slot Attention approach [41]. Each slot contributes to the final reconstruction only if it is activated. Slot activation is determined by a binary variable \\(a_s\\): \\(\\mathcal{M}_s\\) is activated if and only if \\(a_s = 1\\). The output of \\(\\mathcal{M}(\\mathbf{X})\\) is the combination of the reconstructions from all activated slots:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.825, + 0.469, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\mathbf {X}) = \\bigcup_ {\\substack {s = 1 \\dots S \\\\ a _ {s} = 1}} \\mathcal {M} _ {s} (\\mathbf {X}). \\tag{1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Each slot model outputs a point cloud which is the deformation of one learnable prototype chosen from \\(\\mathbf{P}^1, \\dots, \\mathbf{P}^K\\)." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.891, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.292, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Earth Parser Dataset. Our dataset contains 7 scenes representing various urban and natural environments acquired by aerial LiDAR. The illustration of the power plant and the greenhouses display the complete scenes, while other ones display a subset of each scene (between 25 and \\(50\\%\\) of the total area)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.375, + 0.892, + 0.466 + ], + "angle": 0, + "content": "We associate to each slot \\(s\\) a network \\(\\mathcal{T}_s\\) which maps \\(\\mathbf{X}\\) to an affine transformation in 3D space \\(\\mathcal{T}_s(\\mathbf{X})\\). The output \\(\\mathcal{M}_s(\\mathbf{X})\\) of slot \\(s\\) is determined by a variable \\(b_{s} \\in \\{1, \\dots, K\\}\\). If \\(b_{s} = k\\), then the output of \\(\\mathcal{M}_s(\\mathbf{X})\\) is \\(\\mathbf{Y}_s^k\\), the result of applying the transformation \\(\\mathcal{T}_s(\\mathbf{X})\\) to the prototype \\(\\mathbf{P}^k\\):" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.476, + 0.891, + 0.495 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {s} ^ {k} = \\mathcal {T} _ {s} (\\mathbf {X}) [ \\mathbf {P} ^ {k} ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.507, + 0.892, + 0.537 + ], + "angle": 0, + "content": "Please note that \\(\\mathbf{Y}_s^k\\) is a function of \\(\\mathbf{X}\\). However, to keep our notations simple, we omit this dependence." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.632 + ], + "angle": 0, + "content": "Probabilistic modeling. We make our reconstruction model probabilistic by modeling \\( a \\) and \\( b \\) as random variables following (multi-)Bernoulli distributions. We call \\( \\alpha_{s} \\) the probability that slot \\( s \\) is activated and \\( \\beta_{s}^{k} \\) the probability that it is activated and selects the prototype \\( k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.642, + 0.891, + 0.66 + ], + "angle": 0, + "content": "\\[\np \\left(a _ {s} = 1\\right) = \\alpha_ {s}, \\quad p \\left(a _ {s} = 1, b _ {s} = k\\right) = \\beta_ {s} ^ {k}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.671, + 0.892, + 0.825 + ], + "angle": 0, + "content": "For each slot \\(s\\), we predict the vector \\((1 - \\alpha_{s},\\beta_{s}^{1},\\dots \\beta_{s}^{K})\\) with a neural network taking the point cloud \\(\\mathbf{X}\\) as input and finishing with a softmax layer. Again, we don't write the dependency of the \\(\\alpha_{s}\\) and \\(\\beta_{s}\\) on \\(\\mathbf{X}\\) explicitly to simplify the notations. The complete model \\(\\mathcal{M}(\\mathbf{X})\\) and the slots models \\(\\mathcal{M}_s(\\mathbf{X})\\) can now be seen as random variables, producing different potential reconstructions with probabilities given by \\(\\alpha\\) and \\(\\beta\\). During inference, we consider only slots with \\(\\alpha_{s} > 0.5\\) and select the prototype with highest \\(\\beta_s^k\\). However, during training, we compute all reconstructions \\(\\mathbf{Y}_s^k\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.659, + 0.849 + ], + "angle": 0, + "content": "3.2. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given a large 3D scene, we train our model by sampling square patches \\(\\mathbf{X}\\) from the scene. For each batch of patches, we minimize a loss composed of a reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27876" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "and several regularization terms \\(\\mathcal{L}_{\\mathrm{reg}}\\) implementing different priors:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.129, + 0.469, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {M}) = \\mathbb {E} _ {\\mathbf {X}} \\left[ \\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) \\right] + \\mathcal {L} _ {\\text {r e g}} (\\mathcal {M}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.163, + 0.47, + 0.192 + ], + "angle": 0, + "content": "Reconstruction loss. We define the reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) as the sum of two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.2, + 0.469, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) = \\mathcal {L} _ {\\text {a c c}} (\\mathcal {M}, \\mathbf {X}) + \\mathcal {L} _ {\\text {c o v}} (\\mathcal {M}, \\mathbf {X}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.469, + 0.3 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{acc}}\\) encourages likely reconstructions of \\(\\mathcal{M}(\\mathbf{X})\\) to accurately approximate \\(\\mathbf{X}\\), and \\(\\mathcal{L}_{\\mathrm{cov}}\\) ensures coverage, i.e., that each point of \\(\\mathbf{X}\\) is well-reconstructed by at least one activated model. We define each term using the asymmetric Chamfer distance \\(d\\) between two point clouds \\(\\mathbf{X}\\) and \\(\\mathbf{Y}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.317, + 0.469, + 0.351 + ], + "angle": 0, + "content": "\\[\nd (\\mathbf {X}, \\mathbf {Y}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\min _ {y \\in \\mathbf {Y}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.469, + 0.401 + ], + "angle": 0, + "content": "with \\( |.| \\) the number of points in a point cloud. We write \\( d(x, \\mathbf{Y}) \\) the distance between the point \\( x \\) and its closest point in \\( \\mathbf{Y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.402, + 0.469, + 0.433 + ], + "angle": 0, + "content": "We define \\(\\mathcal{L}_{\\mathrm{acc}}\\) as the average over all slots \\(s\\) of the expected distance between \\(\\mathcal{M}_s(\\mathbf{X})\\) and \\(\\mathbf{X}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.44, + 0.469, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a c c}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {a _ {s}, b _ {s}} [ d (\\mathcal {M} _ {s} (\\mathbf {X}), \\mathbf {X}) ] (7) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(\\mathbf {Y} _ {s} ^ {k}, \\mathbf {X}\\right). (8) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.521, + 0.469, + 0.565 + ], + "angle": 0, + "content": "Conversely, we define \\(\\mathcal{L}_{\\mathrm{cov}}\\) as the average over all points \\(x\\) of \\(\\mathbf{X}\\) of the expected distance between \\(x\\) and its closest point in the reconstruction:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.571, + 0.469, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\mathbb {E} _ {a, b} \\left[ \\min _ {s \\mid a _ {s} = 1} d (x, \\mathcal {M} _ {s} (\\mathbf {X})) \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.613, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Following the ideas of Paschalidou et al. [53], we first define \\(\\Delta (x,s)\\) as the expected distance between \\(x\\) and \\(\\mathcal{M}_s(\\mathbf{X})\\) conditionally to the slot \\(s\\) being activated:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.665, + 0.469, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta (x, s) = \\mathbb {E} _ {b _ {s} | a _ {s} = 1} [ d (x, \\mathcal {M} _ {s} (\\mathbf {X})) ] (10) \\\\ = \\frac {1}{\\alpha_ {s}} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(x, \\mathbf {Y} _ {s} ^ {k}\\right). (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.728, + 0.47, + 0.758 + ], + "angle": 0, + "content": "Next, we compute for each point \\( x \\) a permutation \\( \\sigma_x \\) of \\( [1, S] \\) such that \\( \\Delta(x, \\sigma_x(s)) \\) is non-decreasing, i.e.:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.765, + 0.469, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\Delta (x, \\sigma_ {x} (1)) \\leq \\dots \\leq \\Delta (x, \\sigma_ {x} (S)). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.788, + 0.469, + 0.834 + ], + "angle": 0, + "content": "If \\( s \\) is the closest activated slot to \\( x \\), then all the slots closer to \\( x \\) must be deactivated. This observation leads us to rewrite \\( \\mathcal{L}_{\\mathrm{cov}} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.84, + 0.469, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\sum_ {s = 1} ^ {S} \\Delta (x, s) \\alpha_ {s} \\prod_ {r < \\sigma_ {x} (s)} \\left(1 - \\alpha_ {\\sigma_ {x} ^ {- 1} (r)}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.885, + 0.333, + 0.901 + ], + "angle": 0, + "content": "with \\(\\sigma^{-1}\\) the inverse permutation of \\(\\sigma\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Regularization losses. As is often necessary with fully unsupervised reconstruction methods [43, 48, 49], we define several regularization losses implementing priors on the model output to prevent degenerate local minima:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.153, + 0.892, + 0.183 + ], + "angle": 0, + "content": "- To encourage the slot activation to be sparse, we use the following loss penalizing slot activation:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.192, + 0.892, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a c t}} (\\mathcal {M}) = \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.236, + 0.892, + 0.266 + ], + "angle": 0, + "content": "- To avoid slots and prototypes that are never used, we use the following losses, which we compute batch-wise:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.274, + 0.892, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s l o t}} (\\mathcal {M}) = - \\sum_ {s = 1} ^ {S} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {s} ]}{\\sum_ {t = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {t} ]}, \\epsilon_ {S}\\right), \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.314, + 0.892, + 0.355 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r o t o}} (\\mathcal {M}) = - \\sum_ {k = 1} ^ {K} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} \\left[ \\sum_ {s = 1} ^ {S} \\beta_ {s} ^ {k} \\right]}{\\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]}, \\epsilon_ {K}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.363, + 0.89, + 0.393 + ], + "angle": 0, + "content": "with \\(\\epsilon_S\\) and \\(\\epsilon_K\\) hyperparameters setting the smallest acceptable relative use frequency for a slot or prototype." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.393, + 0.891, + 0.422 + ], + "angle": 0, + "content": "The full regularization loss is a weighted sum of the three losses described above:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.433, + 0.892, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {a c t}} \\mathcal {L} _ {\\text {a c t}} + \\lambda_ {\\text {s l o t}} \\mathcal {L} _ {\\text {s l o t}} + \\lambda_ {\\text {p r o t o}} \\mathcal {L} _ {\\text {p r o t o}}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.458, + 0.892, + 0.489 + ], + "angle": 0, + "content": "where we use \\(\\lambda_{\\mathrm{act}} = 10^{-4}\\), \\(\\lambda_{\\mathrm{slot}} = \\lambda_{\\mathrm{proto}} = 0.1\\) and \\(\\epsilon_S = \\epsilon_K = 0.1\\) in all our experiments on the Earth Parser Dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.821, + 0.514 + ], + "angle": 0, + "content": "3.3. Training and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.521, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Model configuration. We process the input point cloud \\(\\mathbf{X}\\) using a similar architecture as in [53]. We first voxelize it with a \\(64\\times 64\\times 64\\) grid and map it to a vector using a sequence of 6 3D sparse convolutions [9] and 6 strided convolutions. The resulting representation is then transformed using one linear layer for each slot. These features are decoded by simple 2-layer MLPs: one generates the distribution parameters \\(\\alpha_{s}\\) and \\(\\beta_s^k\\), and the other ones the parameters of the 3D transformations \\(\\mathcal{T}_s(X)\\). The transformations include an anisotropic scaling, a \\(y\\)-axis tilt of \\(\\pm \\pi /10\\), a rotation around the \\(z\\)-axis, and a translation, in this specific order. We use \\(S = 64\\) slots and \\(K = 6\\) prototypes as default parameters. See the supplementary material for details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.717, + 0.892, + 0.793 + ], + "angle": 0, + "content": "The intensity of the return signal of each point is available in the LiDAR scans. We associate each prototype with a single learnable intensity parameter and perform the Chamfer distance (Equation 6) in 4 dimensions: spatial coordinates normalized to \\([0,1]^3\\) and intensity to \\([0,0.1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Curriculum learning. The network predicts simultaneously the slot's probability distributions and their deformations. This results in many concurrent degrees of freedom and can make the training process unstable. Therefore, following Monnier et al. [49] and Loiseau et al. [43], we implement a multi-stage curriculum learning strategy. We first" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.904, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.895, + 0.329 + ], + "angle": 0, + "content": "Figure 4. Reconstruction Quality. We show two partial scenes with their RGB and intensity values, as well as their reconstruction by our method and competing models. We use the prototypes' intensity to color the points or pixels. As SuperQuadrics does not model the intensity, we use a random colour for each quadric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.473, + 0.465 + ], + "angle": 0, + "content": "initialize the prototypes as point clouds uniformly sampled from a random cuboid and gradually unfreeze the model parameters in the following order: (i) translation, rotation, tilt, slot activation, and choice of prototype; (ii) intensities of the prototypes, when available; (iii) scales of the prototypes; (iv) positions of the prototypes' 3D points; (v) anisotropic scalings of the prototypes. As shown in Section 4, each step of this curriculum scheme improves the performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Prototypes selection. We automatically select the number of prototypes for a complete scene using a simple greedy algorithm. We measure the increase of reconstruction loss when preventing the model from selecting each prototype individually. We remove the prototype with the lowest increase if it is lower than \\(5\\%\\), and iterate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Implementation details. Our model is trained separately for each scene by randomly sampling square patches. During training, the patches are subsampled to a maximum \\(10^{5}\\) points. Each stage of the curriculum is trained until convergence. We use the ADAM optimizer [28] with a learning rate of \\(10^{-4}\\) and default parameters. See the supplementary material for preprocessing, training and evaluation details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.708, + 0.165, + 0.723 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.47, + 0.825 + ], + "angle": 0, + "content": "In this section, we assess the ability of our method to parse complex 3D aerial data. We give in Section 4.1 an overview of our proposed dataset of aerial LiDAR scans. In Section 4.2, we then discuss our evaluation metrics and baselines. Finally, we present a quantitative (Section 4.3) and qualitative (Section 4.4) analysis of our results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.279, + 0.848 + ], + "angle": 0, + "content": "4.1. Earth Parser Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We introduce a new dataset to train and evaluate parsing methods on large, uncurated aerial LiDAR scans. We use data from the French Mapping Agency associated to the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.417 + ], + "angle": 0, + "content": "LiDAR-HD project [45]. Each scan is composed of several airborne LiDAR acquisitions taken at different angles, leading to a minimum resolution of 20 points/\\(\\mathrm{m}^2\\). The points are associated with their laser reflectance (intensity), and colorized based on asynchronous aerial photography." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.525 + ], + "angle": 0, + "content": "We selected 7 scenes, covering over \\(7.7\\mathrm{km}^2\\) and a total of 98 million 3D points, with diverse content and complexity, such as dense habitations, forests, or complex industrial facilities. We associate most 3D points with a coarse semantic label, such as ground, building, or vegetation. The characteristics of the scenes are detailed in Table 2 and each is visualized in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.536, + 0.795, + 0.551 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.56, + 0.892, + 0.606 + ], + "angle": 0, + "content": "We quantitatively evaluate the performances for reconstruction and semantic segmentation of our model and several unsupervised scene decomposition approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.703 + ], + "angle": 0, + "content": "Evaluation metrics. As our goal is to summarize a point cloud using few prototypes, the quality of the reconstruction is critical. We measure it with the symmetric Chamfer distance (\"Cham.\" in the Tables) between the input and the output point clouds of our model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.78 + ], + "angle": 0, + "content": "By associating a class to each prototype's points, we can propagate labels from the reconstruction to the input cloud and perform semantic segmentation. We evaluate the quality of this segmentation with the class-averaged Intersection-over-Union (mIoU) metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In a practical scenario, an operator can manually annotate the points of the 3D prototypes, allowing for the segmentation of the entirety of \\(\\mathbf{X}\\) with minimal effort. To perform automatic evaluation, we follow the standard practice for evaluating clustering and unsupervised segmentation methods [25, 31, 48]: we assign to each prototype's point the most frequent class of its closest point of the input after the reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27878" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Results on the Earth Parser Dataset. We report the quality of the reconstruction (Cham.) and semantic segmentation (mIoU) on each of the scenes of our Earth Parser Dataset. While our method does not always provide the most faithful reconstructions, it leads to the most accurate point classification." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.142, + 0.891, + 0.254 + ], + "angle": 0, + "content": "
Rec.SemanticCrop FieldsForestGreenhousesMarinaPower PlantUrbanWindturbines
Cham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoU
k-means (i,z) [44]X✓—93.8—71.5—39.3—41.4—42.8—56.5—87.6
SuperQuadrics [53]3DX0.86—1.04—0.60—0.93—0.58—0.40—13.5—
DTI-Sprites [49]2.5D+i✓6.1083.214.5940.25.3642.06.1641.45.3629.02.9947.336.1925.9
AtlasNet v2 [11]3D+i✓1.0743.11.5871.40.5649.10.7342.10.4541.60.6348.89.4748.1
Ours3D+i✓0.7296.90.8883.70.4091.30.8278.70.4452.20.2983.26.6593.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.265, + 0.471, + 0.294 + ], + "angle": 0, + "content": "Table 2. Earth Parser Dataset. Our proposed dataset is composed of 7 diverse scenes acquired by aerial LiDAR." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.304, + 0.472, + 0.478 + ], + "angle": 0, + "content": "
NameSurface in km2# points ×106annotation ratio in %num. of classes
Crop Fields1.119.777.42
Forest1.146.797.82
Greenhouses0.11.395.63
Marina0.10.592.72
Power Plant0.28.678.44
Urban1.115.795.93
Wind turbines4.25.699.83
Total7.798.391.6—
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Baselines. We adapt several unsupervised approaches for scene reconstruction and/or semantic segmentation tasks to provide comparisons for our approach:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.674 + ], + "angle": 0, + "content": "- k-means. We cluster the points of the input with the k-means algorithm [44] using as many clusters as we use prototypes. We obtain the best results by using a combination of the point's intensity and elevation as features for clustering. We then assign to each centroid its most frequent class, and propagate this label to the entire cluster, leading to a semantic segmentation. This method does not reconstruct the input, but gives us a simple and surprisingly strong baseline for semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.75 + ], + "angle": 0, + "content": "- SuperQuadrics revisited. We use the method of Paschalidou et al. [53] to learn to approximate scenes with an adaptive number of superquadrics [3]. It provides a baseline for reconstruction and a qualitative comparison for instance segmentation, shown in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.871 + ], + "angle": 0, + "content": "- DTI-Sprites. We use the point cloud to construct a digital elevation model, i.e. a 2.5D image of resolution \\(32 \\times 32\\) where each pixel has an elevation and intensity value. We adapt the unsupervised image decomposition approach of Monnier et al. [49] to break down this image into a set of 2.5D sprites. We evaluate the reconstruction and segmentation by sampling 25 3D point per pixel, transferring the pixel's label to the points, and interpolating their elevations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- AtlasNet v2. This extension [11] of AtlasNet [18] uses a fixed number of learnable prototype point clouds to recon" + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.539, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.265, + 0.895, + 0.307 + ], + "angle": 0, + "content": "Table 3. Ablation Study. We evaluate the effect of our prototype selection post-processing, our model's degrees of freedom, and our different regularization losses." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.318, + 0.892, + 0.525 + ], + "angle": 0, + "content": "
UrbanMarina
Cham.mIoUCham.mIoU
expressivityLearnable Earth Parser0.2983.20.8278.7
→w/o post-processing0.2883.70.9678.3
→w/o aniso-scale0.3382.41.0467.2
→w/o prototypes0.3668.31.0742.8
→w/o scales0.5558.91.3340.8
→w/o intensities0.5558.71.0940.8
losses→w/o Lact0.1754.10.8056.9
→w/o Lslot0.2577.80.8143.7
→w/o Lproto0.2857.20.9740.7
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.609 + ], + "angle": 0, + "content": "struct its input. It can be evaluated for both reconstruction and semantic segmentation in a way similar to ours. We extend it to handle intensity in a manner akin to our approach, which improves its segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.892, + 0.655 + ], + "angle": 0, + "content": "Similar to our method, we train all baselines except k-means by sampling square patches in each scene. Figure 4 shows the output of the reconstruction methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.694, + 0.68 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.733 + ], + "angle": 0, + "content": "We compare the performance of our approach with the proposed baselines on the Earth Parser Dataset, as well as two publicly available datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Earth Parser Dataset. We provide quantitative reconstruction and semantic segmentation results in Table 1, and illustrations in Figure 4. Despite being highly constrained, our model yields the best reconstruction in 6 out of 7 scenes. Moreover, we significantly outperform the other methods for semantic segmentation across all scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Despite its simplicity, the k-means baseline provides strong semantic segmentation performance, beating the other baselines in 5 of 7 annotated scenes. DTI-Sprites [49] has lower reconstruction and segmentation quality, which" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27879" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.145 + ], + "angle": 0, + "content": "Table 4. Synthetic Shapes. We train our method on all planes from ShapeNet-Part [58], with random rotations around \\( z \\)-axis. We show the reconstruction of an input plane and the prototypes learned on the dataset." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.156, + 0.472, + 0.307 + ], + "angle": 0, + "content": "
AtlasNet v2 [11]SuperQuadrics [53]Ours
Recons.
Protos.---
Cham.1.462.911.34
mIoU34.5-68.6
" + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.318, + 0.284, + 0.426 + ], + "angle": 0, + "content": "
Cham.mIoU
k-means (i,z) [44]—52.7
Superquadrics [53]0.33—
DTI-Sprites [49]2.2553.6
AtlasNet v2 [11]0.3154.9
Ours0.2982.7
" + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.319, + 0.474, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.465 + ], + "angle": 0, + "content": "Figure 5. Results on DALES [68]. We report quantitative and qualitative results for one tile from DALES." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.648 + ], + "angle": 0, + "content": "is expected as it models a 3D point cloud in 2.5D. AtlasNet v2 [11] provides good reconstructions but segmentation fails for scenes such as Crop Fields or Urban due its inability to adapt its prototype usage to the input. On the contrary, SuperQuadrics [53] can adjust the number of superquadric it uses and, to some degree, their shape. However, this method uses a single parametric family for all prototypes and fails to reconstruct complex real-world scenes such as Power Plant. Thanks to its probabilistic slot selection, our method can handle inputs with a varying number of objects using only a small set of learned prototypical shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.468, + 0.853 + ], + "angle": 0, + "content": "Ablation study. We evaluate the impact of various components of our model and report the results in Table 3. First, we observe that the prototype selection post-processing has limited impact on the quality of the prediction and reconstructions, but allows us to adapt and significantly decrease the number of prototypes used for each scene. Second, we evaluate the impact of reducing the expressivity of our model. We successively remove: (i) the anisotropic scaling of \\(\\mathcal{T}\\), and the possibility of learning the prototype's (ii) points position, (iii) scale, and (iv) intensity. As expected, each degree if liberty removed decreases the quality of the 3D reconstruction and segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We study the impact of the different regularization introduced in Section 3.2. The losses related to slot activation have a marginal effect on reconstruction quality but" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.197 + ], + "angle": 0, + "content": "significantly affect the performance of semantic segmentation. This suggests that using slots sparingly but equally is important for prototypes to specialize for specific objects, but is not necessary to the expressivity of the reconstruction model. The prototype activation loss prevents unused and possibly degenerate prototypes and therefore improves both segmentation and reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.893, + 0.378 + ], + "angle": 0, + "content": "We evaluate the impact of reducing the number of prototypes or slots. When we set \\( K = 3 \\) instead of 6, the average semantic segmentation mIoU drops by 26.3 points, while reconstruction is minimally affected, showing only a \\( 2.3\\% \\) increase in Chamfer distance. Conversely, setting \\( S = 32 \\) instead of 64 leads to a significant \\( +47.3\\% \\) increase of Chamfer distance, with a modest drop 0.4 of semantic segmentation mIoU. These findings highlight the importance of maintaining sufficient diversity in the prototypes to enable specialization for different object types, whereas the number of slots strongly influences the expressiveness of the reconstruction model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.893, + 0.563 + ], + "angle": 0, + "content": "ShapeNet. We evaluate our model on 2690 planes from ShapeNet-Part [58], whose points are annotated as wing, engine, tail, or body. We randomly rotate the shapes around the \\(z\\)-axis during training and evaluation. We report in Table 4 the performances and reconstructions for our approach, AtlasNet v2 [11], and SuperQuadrics revisited [53]. Our method handle rotations better than AtlasNet v2, and manages to successfully locate the tail of the planes. While SuperQuadrics' reconstructions make sense qualitatively, they are worse in terms of accuracy than ours, and do not enable semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.687 + ], + "angle": 0, + "content": "DALES. We also train and evaluate our model on DALES [68], a dataset of aerial LiDAR scans of urban area, with a restricted class set: ground, vegetation, and buildings. Our model significantly outperforms competing reconstruction-based approaches, and yields similar performance than for the Urban tile of the Earth Parser Dataset; see Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.684, + 0.713 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Instance segmentation. We can perform instance segmentation simply by considering each slot as a different instance. This is particularly interesting for parsing natural woodlands, a key endeavor for forest management [50] and biomass estimation [12]. While this task has a long history of handcrafted approaches [69], current deep learning approaches are mostly limited to artificial or low-density forests [65]. As shown qualitatively in Figure 6, our Learnable Earth Parser can learn without any supervision to separate individual trees in dense forests, or boats in a marina. We evaluate quantitatively the performance of our model in terms of instance segmentation, we counted manually 88" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27880" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.47, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.421 + ], + "angle": 0, + "content": "Figure 6. Instance Segmentation. We can identify the reconstruction of each slot as a separate instance, allowing us to perform instance segmentation of complex data such as dense natural forests or a marina. For this visualization, we considered the points associated to \"trees\" or \"boat hull\" prototypes and color each of their instance randomly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.526 + ], + "angle": 0, + "content": "boat instances in 10 distinct portions of the Marina scene. We then computed the Mean Relative Error (MRE) of the prediction given by the number of boat-like prototypes (prototype #3 in Figure 8) in these zones. Our method yields a MRE of \\(7,4\\%\\), 2.6 times lower than DTI-Sprites [49], demonstrating its strong performances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.588 + ], + "angle": 0, + "content": "Semantic segmentation. After annotating the prototypes, our model can perform convincing semantic segmentation of complex scenes, as shown in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.471, + 0.818 + ], + "angle": 0, + "content": "Interpretable prototypes. In Figure 8, we show prototypes learned on our Earth Parser Dataset with colors showing the associated semantic label for each point. These prototypes give at a glance insights on the content of these real-world scenes. Our model is able to learn a wide variety of shapes, such as boats' masts, wind turbines, or greenhouses. We also observe that the prototypes are typically associated with a single object type. Empirically, the average class distribution of the nearest neighbors of each prototypes' point exhibits the same normalized entropy (0.22) as a Bernoulli distribution with probability 0.964. The highly constrained nature of the deformations \\(\\mathcal{T}\\) prevents slots from repurposing the same prototypes for different objects, making the learned prototypes easy to identify and manually annotate." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.196, + 0.847 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We introduced a novel unsupervised method for parsing complex real-world aerial scans into simple parts using a small set of learned prototypical shapes. We demonstrate" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.895, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.326, + 0.895, + 0.368 + ], + "angle": 0, + "content": "Figure 7. Semantic Segmentation. Our model can perform semantic segmentation of large real-world scene based on annotated prototypes. Black points in the ground truth are unannotated." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.371, + 0.892, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.513, + 0.892, + 0.543 + ], + "angle": 0, + "content": "Figure 8. Learned Prototypes. We display three of the prototypes chosen during the selection process for various scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.894, + 0.665 + ], + "angle": 0, + "content": "the quality and interpretability of our results on a novel dataset of aerial LiDAR scans. To the best of our knowledge, we are the first to demonstrate the possibility of performing deep unsupervised 3D shape analysis on such a challenging real-world dataset. We believe that our results open new perspectives for computer-assisted environment monitoring and economic intelligence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This work was supported in part by ANR project READY3D ANR-19-CE23-0007, ANR under the France 2030 program under the reference ANR-23-PEIA-0008, and was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012096R2 made by GENCI. The work of MA was partly supported by the European Research Council (ERC project DISCOVER, number 101076028). The scenes of Earth Parser Dataset were acquired and annotated by the LiDAR-HD project [45]. We thank Zenodo for hosting the dataset. We thank Zeynep Sonat Baltaci, Emile Blettery, Nicolas Dufour, Antoine Guedon, Helen Mair Rawsthorne, Tom Monnier, Damien Robert, Mathis Petrovich and Yannis Siglidis for inspiring discussions and valuable feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27881" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.156 + ], + "angle": 0, + "content": "[1] Antonio Alliegro, Davide Boscaini, and Tatiana Tommasi. Joint supervised and self-supervised learning for 3D real world challenges. In ICPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.2 + ], + "angle": 0, + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.202, + 0.468, + 0.242 + ], + "angle": 0, + "content": "[3] Alan H Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1981. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.245, + 0.468, + 0.286 + ], + "angle": 0, + "content": "[4] Pál Benkő, Ralph R Martin, and Tamás Várady. Algorithms for reverse engineering boundary representation models. Computer-Aided Design, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.288, + 0.467, + 0.315 + ], + "angle": 0, + "content": "[5] Thomas Binford. Visual perception by computer. In Proc. IEEE Conf. on Systems and Control, 1975. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.318, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[6] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.361, + 0.468, + 0.429 + ], + "angle": 0, + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.432, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[8] YH Chen and CY Liu. Quadric surface extraction using genetic algorithms. Computer-Aided Design, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.461, + 0.468, + 0.5 + ], + "angle": 0, + "content": "[9] Spconv Contributors. SPConv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[10] Yang Cui, Qingquan Li, Bisheng Yang, Wen Xiao, Chi Chen, and Zhen Dong. Automatic 3-d reconstruction of indoor environment with mobile laser scanning point clouds. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.629 + ], + "angle": 0, + "content": "[11] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. NeurIPS, 2019. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[12] Antonio Ferraz, Sassan Saatchi, Clément Mallet, Stéphane Jacquemoud, Gil Gonçalves, Carlos Alberto Silva, Paula Soares, Margarida Tomé, and Luisa Pereira. Airborne Li-DAR estimation of aboveground forest biomass in the absence of field inventory. Remote Sensing, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.468, + 0.8 + ], + "angle": 0, + "content": "[14] Ran Gal, Olga Sorkine, Niloy J Mitra, and Daniel Cohen-Or. iWIRES: An analyze-and-edit approach to shape manipulation. In ACM SIGGRAPH. 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[15] Weixiao Gao, Liangliang Nan, Bas Boom, and Hugo Ledoux. SUM: A benchmark dataset of semantic urban meshes. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[16] Xuming Ge, Bo Wu, Yuan Li, and Han Hu. A multiprimitive-based hierarchical optimal approach for semantic labeling of ALS point clouds. Remote Sensing, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[17] Paulo FU Gotardo, Olga Regina Pereira Bellon, and Luciano Silva. Range image segmentation by surface extraction using an improved robust estimator. In CVPR, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[18] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[19] Stéphane Guinard, Loic Landrieu, Laurent Caraffa, and Bruno Vallet. Piecewise-planar approximation of large 3d data as graph-structured optimization. ISPRS Annals of Photogrammetry, Remote Sensing & Spatial Information Sciences, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[20] Florent Guiotte, Minh-Tan Pham, Romain Dambreville, Thomas Corpetti, and Sébastien Lefèvre. Semantic segmentation of LiDAR points clouds: Rasterization beyond digital elevation models. IEEE Geoscience and Remote Sensing Letters, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[21] Bo Guo, Qingquan Li, Xianfeng Huang, and Chisheng Wang. An improved method for power-line reconstruction from point cloud data. Remote sensing, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[22] Abhinav Gupta, Alexei Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In ECCV, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[23] Norbert Haala and Claus Brenner. Extraction of buildings and trees in urban environments. ISPRS journal of photogrammetry and remote sensing, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.495 + ], + "angle": 0, + "content": "[24] Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3D: large-scale building reconstruction from airborne LiDAR point clouds. Remote Sensing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. CVPR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[26] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3D data. In Computer Graphics Forum. Wiley Online Library, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.596, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[27] Ekaterina Kalinicheva, Loic Landrieu, Clément Mallet, and Nesrine Chehata. Multi-layer modeling of dense vegetation from aerial lidar scans. In CVPR Workshops, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[28] Diederik P Kingma and Jimmy Ba. ADAM: A method for stochastic optimization. *ICLR*, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[29] Florian Kluger, Hanno Ackermann, Eric Brachmann, Michael Ying Yang, and Bodo Rosenhahn. Cuboids revisited: Learning robust 3D shape fitting to single RGB images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[31] Adam Kosiorek, Sara Sabour, Yee Whye Teh, and Geoffrey E Hinton. Stacked capsule autoencoders. NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[32] Florent Lafarge, Xavier Descombes, Josiane Zerubia, and Marc Pierrot-Deseilligny. Automatic 3D buildup reconstruction from DEMs: An application to PLEIADES simulations. In ISPRS, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[33] Aleš Leonardis, Alok Gupta, and Ruzena Bajcsy. Segmentation of range images as the search for geometric parametric models. International Journal of Computer Vision, 1995. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27882" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[34] Ales Leonardis, Ales Jaklic, and Franc Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[35] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[36] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3D point clouds. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[37] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.342 + ], + "angle": 0, + "content": "[38] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. In ACM SIGGRAPH. 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[39] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.469, + 0.441 + ], + "angle": 0, + "content": "[40] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[41] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[42] Suresh K Lodha, Edward J Kreps, David P Helmbold, and Darren Fitzpatrick. Aerial LiDAR data classification using support vector machines. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06). IEEE, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[43] Romain Loiseau, Tom Monnier, Mathieu Aubry, and Loic Landrieu. Representing shape collections with alignment-aware linear models. In 3DV, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[44] J MacQueen. Classification and analysis of multivariate observations. In 5th Berkeley Symp. Math. Statist. Probability, 1967. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.678 + ], + "angle": 0, + "content": "[45] IGN (French mapping agency). LiDAR-HD: A 3D mapping of France's soil and subsoil, 2021. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[46] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J.Davison. SceneNet RGB-D: Can 5M synthetic images beat generic imagenet pre-training on indoor segmentation? ICCV, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[47] Jie Mei, Liqiang Zhang, Shihao Wu, Zhen Wang, and Liang Zhang. 3d tree modeling from incomplete point clouds via optimization and 11-mst. International Journal of Geographical Information Science, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[48] Tom Monnier, Thibault Groueix, and Mathieu Aubry. Deep Transformation-Invariant Clustering. NeurIPS, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[49] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In ICCV, 2021. 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[50] Felix Morsdorf, Erich Meier, Benjamin Kötz, Klaus I Itten, Matthias Dobbertin, and Britta Allgower. LIDAR-based geometric reconstruction of boreal type forest stands at single" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tree level for forest and wildland fire management. Remote sensing of environment, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[51] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[52] Joachim Niemeyer, Franz Rottensteiner, and Uwe Soergel. Contextual classification of LiDAR data and building object detection in urban areas. ISPRS journal of photogrammetry and remote sensing, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[53] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, 2019. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.246, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[54] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[55] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[56] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. Monteboxfinder: Detecting and filtering primitives to fit a noisy point cloud. In ECCV. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[57] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.429, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[58] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR Workshop, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[59] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR 2015 Workshop on Functionality, Physics, Intentionality and Causality, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[60] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. In Computer Graphics Forum, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[61] Ruwen Schnabel, Patrick Degener, and Reinhard Klein. Completion and reconstruction with primitive shapes. In Computer Graphics Forum. Wiley Online Library, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[62] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[63] Nina M Singer and Vijayan K Asari. DALES objects: A large scale benchmark dataset for instance segmentation in aerial LiDAR. IEEE Access, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[64] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[65] Chenxin Sun, Chengwei Huang, Huaiqing Zhang, Bangqian Chen, Feng An, Liwen Wang, and Ting Yun. Individual tree crown segmentation and crown width extraction from a heightmap derived from aerial laser scanning data using a deep learning framework. Frontiers in plant science, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[66] Minhyuk Sung, Vladimir G Kim, Roland Angst, and Leonidas Guibas. Data-driven structural priors for shape completion. ACM Transactions on Graphics, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[67] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27883" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[68] Nina Varney, Vijayan K Asari, and Quinn Graehling. DALES: A large-scale aerial LiDAR data set for semantic segmentation. In CVPR Workshop, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.205 + ], + "angle": 0, + "content": "[69] Jari Vauhkonen, Liviu Ene, Sandeep Gupta, Johannes Heinzel, Johan Holmgren, Juho Pitkänen, Svein Solberg, Yunsheng Wang, Holger Weinacker, K Marius Hauglin, et al. Comparative testing of single-tree detection algorithms under different types of forest. Forestry, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[70] Ben G Weinstein, Sergio Marconi, Stephanie Bohlman, Alina Zare, and Ethan White. Individual tree-crown detection in RGB imagery using semi-supervised deep learning neural networks. Remote Sensing, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.263, + 0.468, + 0.303 + ], + "angle": 0, + "content": "[71] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.305, + 0.468, + 0.359 + ], + "angle": 0, + "content": "[72] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[73] Shaobo Xia, Dong Chen, Ruisheng Wang, Jonathan Li, and Xinchang Zhang. Geometric primitives in LiDAR point clouds: A review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.419, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[74] Zhen Ye, Yusheng Xu, Rong Huang, Xiaohua Tong, Xin Li, Xiangfeng Liu, Kuifeng Luan, Ludwig Hoegner, and Uwe Stilla. LASDU: A large-scale aerial LiDAR dataset for semantic labeling in dense urban areas. ISPRS International Journal of Geo-Information, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.529 + ], + "angle": 0, + "content": "[75] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised Discovery of Object Radiance Fields. *ICLR*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[76] Yu Zhang, Peter Tino, Aleks Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.468, + 0.615 + ], + "angle": 0, + "content": "[77] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.618, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[78] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3D point capsule networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[79] SM Zolanvari, Susana Ruano, Aakanksha Rana, Alan Cummins, Rogerio Eduardo da Silva, Morteza Rahbar, and Aljosa Smolic. DublinCity: annotated LiDAR point cloud and its applications. BMVC, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.704, + 0.468, + 0.743 + ], + "angle": 0, + "content": "[80] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In ICCV, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27884" + } + ] +] \ No newline at end of file diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_origin.pdf b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e3cd9ab24b2bc711a9cfdac9df897a4096c26be2 --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/5317043d-970e-43ef-b2f4-dc56567d2d02_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1905ffc47f2b08a672fc0ff7f188d025b7b482722c9e6e2b94ae3ba259c6221 +size 7678742 diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/full.md b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/full.md new file mode 100644 index 0000000000000000000000000000000000000000..231b6fa148d8a8739d7d2b328d333c811b4e84b2 --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/full.md @@ -0,0 +1,392 @@ +# Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans + +Romain Loiseau1,2 + +romain.loiseau@enpc.fr + +Elliot Vincent1,3 + +elliot.vincent@enpc.fr + +Mathieu Aubry + +mathieu.aubry@enpc.fr + +Loic Landrieu $^{1,2}$ + +loic. landrieu@enpc.fr + +1 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France + +2 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France + +3 INRIA Paris, France + +# Abstract + +We propose an unsupervised method for parsing large 3D scans of real-world scenes with easily-interpretable shapes. This work aims to provide a practical tool for analyzing 3D scenes in the context of aerial surveying and mapping, without the need for user annotations. Our approach is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical 3D shapes. The resulting reconstruction is visually interpretable and can be used to perform unsupervised instance and low-shot semantic segmentation of complex scenes. We demonstrate the usefulness of our model on a novel dataset of seven large aerial Li-DAR scans from diverse real-world scenarios. Our approach outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. Our code and dataset are available at https://romainloiseau.fr/learnable-earth-parser/. + +# 1. Introduction + +Modern aerial 3D scanning technologies open up unprecedented opportunities for environmental monitoring and economic intelligence. However, their practical use remains challenging due to the complexity of real-world scenes, the diversity of usage scenarios, and the difficulty of annotation. Therefore, our aim is to develop an approach that could help perform diverse tasks—from counting trees in a forest or identifying the various components of a factory to measuring the surface of greenhouses or monitor urban growth—all without human supervision. + +To do so, we address two important limitations of existing 3D deep learning methods. First, they are often primarily designed, trained, and tested on synthetic [7, 46, 59, 72] or highly curated data [2, 30, 39], which fail to capture the endless variability of the real world. Moreover, they often + +![](images/8501e7e11e85de94600e8c49e4cab8f1f2be7c828e79f17b6869bedaeebe3cf2.jpg) +Figure 1. Learnable Earth Parser. Our unsupervised method takes large aerial 3D scans as input and model them with a small set of learned 3D prototypes. Our approach is trained without annotation and produce legible decompositions of complex scenes, which can be used for semantic and instance segmentation. + +assume that annotations are available for tasks of interest. Second, even unsupervised approaches [1, 77] often rely on learning abstract feature representations, making them difficult to interpret [76]. Although some work has attempted to decompose 3D shapes into meaningful components without supervision [11, 43, 53, 67], they were all designed on simple synthetic shapes and none generalizes to real data. + +To overcome these limitations, we present the Learnable Earth Parser, an unsupervised deep learning method + +![](images/b8ff0cd769002fc64a16aecb1caf1d09aa06d61cd036dba3a5c47ff5ee233c19.jpg) +Figure 2. Method Overview. Our model approximates an input point cloud $\mathbf{X}$ with $S$ slot models. Each slot maps $\mathbf{X}$ to an affine 3D deformation $\mathcal{T}_s(\mathbf{X})$ , a slot activation probability $\alpha_{s}$ , and the joint probabilities $\beta_s^1,\dots,\beta_s^K$ of the slot being activated and choosing one of the $K$ learnable prototype point clouds $\mathbf{P}^1,\dots,\mathbf{P}^K$ . The output $\mathcal{M}_s(\mathbf{X})$ of an activated slot $s$ is obtained by applying the transformation $\mathcal{T}_s(\mathbf{X})$ to its most likely prototype. Non-activated slots do not contribute to the output. + +designed to decompose large-scale 3D point clouds into interpretable parts. Our model learns a small set of 3D prototypical shapes that are selected, positioned, rotated, and resized to reconstruct an input point cloud. We introduce a novel probabilistic formulation that enables the design of a reconstruction loss for learning jointly the 3D prototypes, but also to select and position them. + +To evaluate the effectiveness of our approach, we created a new open-access dataset consisting of 7 aerial LiDAR scans, covering $7.7\mathrm{km}^2$ and containing 98 million 3D points with annotations in diverse urban and natural environments. Our results demonstrate that the Learnable Earth Parser learns decompositions superior to traditional and deep learning baselines, leading to convincing performance for semantic and instance segmentation, as shown in Figure 1. We believe that our contributions provide researchers and practitioners with new tools and resources to tackle the challenges of real-world 3D data. + +# 2. Related work + +Our proposed unsupervised method uses point cloud reconstruction as a proxy to learn to decompose large aerial point clouds and is evaluated on a novel and diverse dataset of 3D scans. In the following, we briefly present related works for primitive-based point cloud decomposition, automatic decomposition of LiDAR data, and an overview of existing aerial LiDAR datasets. + +Primitive-based point cloud decomposition. Modeling shapes as a set of primitives such as generalized cylinders [5] or superquadrics [3] has a rich history in vision and graphics [26]. Classical applications include reverse engineering [4], shape completion [61, 66], and shape editing [14]. A variety of methods have been developed to find primitives in unstructured 3D scenes, including seed growing techniques [33, 34], genetic algorithms [8], approaches [17, 38, 56, 60] based on RANSAC [13], and probabilistic methods [40, 71]. + +For this problem, like for many others in computer vision, deep learning has become the dominant paradigm. However, supervised methods [36, 80] are limited by the availability of annotated datasets. Following the seminal work of Tulsiani et al. [67], unsupervised approaches that simply rely on a reconstruction loss to learn primitive decomposition are the most common and the most closely related to our work [53, 55]. An important challenge for these approaches is to model a variable number of primitives. This has been addressed using recurrent networks [35, 62, 80], capsule networks [78], reinforcement learning strategies [67] or, most similar to our approach, computing the Chamfer distance based on a probabilistic model [53]. Our work is also related to approaches that learn prototypical shapes instead of being restricted to a predefined family of parametric primitives [11, 43]. + +However, most of these methods are designed, trained and evaluated on well-curated synthetic object datasets, such as ModelNet [72], ShapeNet [7], or D-FAUST [6], and are typically designed to handle single objects from known categories. In contrast, our approach can handle complex scenes composed of many objects with significant variety. + +Decomposition of LiDAR scans. Automatically decomposing large LiDAR scans poses unique challenges due to their size and diversity [73]. Some previous approaches use simple shape primitives, such as lines [10, 21], planes [16, 19, 51], or volumes [37], but these may not be flexible enough to capture the complexity of real-world scans. Other approaches are designed for specific object classes, like trees [47] or buildings [24, 32], but they are limited in their ability to represent a wide range of shapes. In contrast, our Learnable Earth Parser overcomes these limitations by learning ad-hoc prototypes for each new scene, ensuring both expressivity and adaptability. LiDAR data are often treated as digital elevation models, i.e. images with pixel elevations [20, 23, 42]. Thus, our work is related to image-based primitive prediction [22, 29, 54, 57] and unsupervised + +multi-object image segmentation [41, 49, 64, 75]. However, 3D point clouds have higher precision and can better represent multi-layered structures such as forest areas. + +Aerial LiDAR datasets. The increased availability of aerial LiDAR technology has led to the multiplication of open datasets [15, 52, 74, 79] of varying sizes from 1 to 10 $\mathrm{km}^2$ [63, 68]. However, these scans are limited to dense urban environments and do not capture the challenge of modeling diverse terrains. Some specialized datasets focus on forested areas [27, 70]. Our proposed dataset is of similar scale, spanning $7.7\mathrm{km}^2$ , but covers a variety of urban, natural, and rural scenes, making it more representative of the diversity of possible usage scenarios. + +# 3. Method + +Our goal is to learn to break down a point cloud into simpler and more easily understandable components. To achieve this, we propose an analysis-by-synthesis approach where we train a highly-constrained model $\mathcal{M}$ to approximate a point cloud as a combination of learned 3D shapes. + +# 3.1. Probabilistic Scene Reconstruction Model + +As illustrated in Figure 2, our model first selects up to $S$ shapes from $K$ learnable 3D shapes, then positions and deforms them to best approximate an input point cloud $\mathbf{X}$ . We propose a probabilistic formulation of the selection process, which can be seen as an extension of the model of Paschalidou et al. [53] to multiple free-form shapes instead of a single parametric family. + +Learnable shape prototypes. Following Loiseau et al. [43], we define $K$ point clouds $\mathbf{P}^1,\dots ,\mathbf{P}^K$ that we refer to as prototypes. Each prototype is meant to represent a single instance of a recurring 3D structure in the considered scene. The points' coordinates are free parameters of the model and learned directly. + +Scene reconstruction model. Our full model $\mathcal{M}$ is the combination of $S$ reconstruction models $\mathcal{M}_s$ , which we refer to as slots in analogy to the Slot Attention approach [41]. Each slot contributes to the final reconstruction only if it is activated. Slot activation is determined by a binary variable $a_s$ : $\mathcal{M}_s$ is activated if and only if $a_s = 1$ . The output of $\mathcal{M}(\mathbf{X})$ is the combination of the reconstructions from all activated slots: + +$$ +\mathcal {M} (\mathbf {X}) = \bigcup_ {\substack {s = 1 \dots S \\ a _ {s} = 1}} \mathcal {M} _ {s} (\mathbf {X}). \tag{1} +$$ + +Each slot model outputs a point cloud which is the deformation of one learnable prototype chosen from $\mathbf{P}^1, \dots, \mathbf{P}^K$ . + +![](images/ea9845e46b2a4f8fbafe1137999f9fb3b950160164bc24602c779f64c39471cd.jpg) +Figure 3. Earth Parser Dataset. Our dataset contains 7 scenes representing various urban and natural environments acquired by aerial LiDAR. The illustration of the power plant and the greenhouses display the complete scenes, while other ones display a subset of each scene (between 25 and $50\%$ of the total area). + +We associate to each slot $s$ a network $\mathcal{T}_s$ which maps $\mathbf{X}$ to an affine transformation in 3D space $\mathcal{T}_s(\mathbf{X})$ . The output $\mathcal{M}_s(\mathbf{X})$ of slot $s$ is determined by a variable $b_{s} \in \{1, \dots, K\}$ . If $b_{s} = k$ , then the output of $\mathcal{M}_s(\mathbf{X})$ is $\mathbf{Y}_s^k$ , the result of applying the transformation $\mathcal{T}_s(\mathbf{X})$ to the prototype $\mathbf{P}^k$ : + +$$ +\mathbf {Y} _ {s} ^ {k} = \mathcal {T} _ {s} (\mathbf {X}) [ \mathbf {P} ^ {k} ]. \tag {2} +$$ + +Please note that $\mathbf{Y}_s^k$ is a function of $\mathbf{X}$ . However, to keep our notations simple, we omit this dependence. + +Probabilistic modeling. We make our reconstruction model probabilistic by modeling $a$ and $b$ as random variables following (multi-)Bernoulli distributions. We call $\alpha_{s}$ the probability that slot $s$ is activated and $\beta_{s}^{k}$ the probability that it is activated and selects the prototype $k$ : + +$$ +p \left(a _ {s} = 1\right) = \alpha_ {s}, \quad p \left(a _ {s} = 1, b _ {s} = k\right) = \beta_ {s} ^ {k}. \tag {3} +$$ + +For each slot $s$ , we predict the vector $(1 - \alpha_{s},\beta_{s}^{1},\dots \beta_{s}^{K})$ with a neural network taking the point cloud $\mathbf{X}$ as input and finishing with a softmax layer. Again, we don't write the dependency of the $\alpha_{s}$ and $\beta_{s}$ on $\mathbf{X}$ explicitly to simplify the notations. The complete model $\mathcal{M}(\mathbf{X})$ and the slots models $\mathcal{M}_s(\mathbf{X})$ can now be seen as random variables, producing different potential reconstructions with probabilities given by $\alpha$ and $\beta$ . During inference, we consider only slots with $\alpha_{s} > 0.5$ and select the prototype with highest $\beta_s^k$ . However, during training, we compute all reconstructions $\mathbf{Y}_s^k$ . + +# 3.2. Training Losses + +Given a large 3D scene, we train our model by sampling square patches $\mathbf{X}$ from the scene. For each batch of patches, we minimize a loss composed of a reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ + +and several regularization terms $\mathcal{L}_{\mathrm{reg}}$ implementing different priors: + +$$ +\mathcal {L} (\mathcal {M}) = \mathbb {E} _ {\mathbf {X}} \left[ \mathcal {L} _ {\text {r e c}} (\mathcal {M}, \mathbf {X}) \right] + \mathcal {L} _ {\text {r e g}} (\mathcal {M}). \tag {4} +$$ + +Reconstruction loss. We define the reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ as the sum of two losses: + +$$ +\mathcal {L} _ {\text {r e c}} (\mathcal {M}, \mathbf {X}) = \mathcal {L} _ {\text {a c c}} (\mathcal {M}, \mathbf {X}) + \mathcal {L} _ {\text {c o v}} (\mathcal {M}, \mathbf {X}). \tag {5} +$$ + +$\mathcal{L}_{\mathrm{acc}}$ encourages likely reconstructions of $\mathcal{M}(\mathbf{X})$ to accurately approximate $\mathbf{X}$ , and $\mathcal{L}_{\mathrm{cov}}$ ensures coverage, i.e., that each point of $\mathbf{X}$ is well-reconstructed by at least one activated model. We define each term using the asymmetric Chamfer distance $d$ between two point clouds $\mathbf{X}$ and $\mathbf{Y}$ : + +$$ +d (\mathbf {X}, \mathbf {Y}) = \frac {1}{| \mathbf {X} |} \sum_ {x \in \mathbf {X}} \min _ {y \in \mathbf {Y}} \| x - y \| _ {2} ^ {2}, \tag {6} +$$ + +with $|.|$ the number of points in a point cloud. We write $d(x, \mathbf{Y})$ the distance between the point $x$ and its closest point in $\mathbf{Y}$ . + +We define $\mathcal{L}_{\mathrm{acc}}$ as the average over all slots $s$ of the expected distance between $\mathcal{M}_s(\mathbf{X})$ and $\mathbf{X}$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {a c c}} (\mathcal {M}, \mathbf {X}) = \frac {1}{S} \sum_ {s = 1} ^ {S} \mathbb {E} _ {a _ {s}, b _ {s}} [ d (\mathcal {M} _ {s} (\mathbf {X}), \mathbf {X}) ] (7) \\ = \frac {1}{S} \sum_ {s = 1} ^ {S} \sum_ {k = 1} ^ {K} \beta_ {s} ^ {k} d \left(\mathbf {Y} _ {s} ^ {k}, \mathbf {X}\right). (8) \\ \end{array} +$$ + +Conversely, we define $\mathcal{L}_{\mathrm{cov}}$ as the average over all points $x$ of $\mathbf{X}$ of the expected distance between $x$ and its closest point in the reconstruction: + +$$ +\mathcal {L} _ {\mathrm {c o v}} (\mathcal {M}, \mathbf {X}) = \frac {1}{| \mathbf {X} |} \sum_ {x \in \mathbf {X}} \mathbb {E} _ {a, b} \left[ \min _ {s \mid a _ {s} = 1} d (x, \mathcal {M} _ {s} (\mathbf {X})) \right]. \tag {9} +$$ + +Following the ideas of Paschalidou et al. [53], we first define $\Delta (x,s)$ as the expected distance between $x$ and $\mathcal{M}_s(\mathbf{X})$ conditionally to the slot $s$ being activated: + +$$ +\begin{array}{l} \Delta (x, s) = \mathbb {E} _ {b _ {s} | a _ {s} = 1} [ d (x, \mathcal {M} _ {s} (\mathbf {X})) ] (10) \\ = \frac {1}{\alpha_ {s}} \sum_ {k = 1} ^ {K} \beta_ {s} ^ {k} d \left(x, \mathbf {Y} _ {s} ^ {k}\right). (11) \\ \end{array} +$$ + +Next, we compute for each point $x$ a permutation $\sigma_x$ of $[1, S]$ such that $\Delta(x, \sigma_x(s))$ is non-decreasing, i.e.: + +$$ +\Delta (x, \sigma_ {x} (1)) \leq \dots \leq \Delta (x, \sigma_ {x} (S)). \tag {12} +$$ + +If $s$ is the closest activated slot to $x$ , then all the slots closer to $x$ must be deactivated. This observation leads us to rewrite $\mathcal{L}_{\mathrm{cov}}$ as follows: + +$$ +\mathcal {L} _ {\mathrm {c o v}} (\mathcal {M}, \mathbf {X}) = \frac {1}{| \mathbf {X} |} \sum_ {x \in \mathbf {X}} \sum_ {s = 1} ^ {S} \Delta (x, s) \alpha_ {s} \prod_ {r < \sigma_ {x} (s)} \left(1 - \alpha_ {\sigma_ {x} ^ {- 1} (r)}\right), \tag {13} +$$ + +with $\sigma^{-1}$ the inverse permutation of $\sigma$ . + +Regularization losses. As is often necessary with fully unsupervised reconstruction methods [43, 48, 49], we define several regularization losses implementing priors on the model output to prevent degenerate local minima: + +- To encourage the slot activation to be sparse, we use the following loss penalizing slot activation: + +$$ +\mathcal {L} _ {\mathrm {a c t}} (\mathcal {M}) = \sum_ {s = 1} ^ {S} \mathbb {E} _ {\mathbf {X}} \left[ \alpha_ {s} \right]. \tag {14} +$$ + +- To avoid slots and prototypes that are never used, we use the following losses, which we compute batch-wise: + +$$ +\mathcal {L} _ {\text {s l o t}} (\mathcal {M}) = - \sum_ {s = 1} ^ {S} \min \left(\frac {\mathbb {E} _ {\mathbf {X}} [ \alpha_ {s} ]}{\sum_ {t = 1} ^ {S} \mathbb {E} _ {\mathbf {X}} [ \alpha_ {t} ]}, \epsilon_ {S}\right), \tag {15} +$$ + +$$ +\mathcal {L} _ {\text {p r o t o}} (\mathcal {M}) = - \sum_ {k = 1} ^ {K} \min \left(\frac {\mathbb {E} _ {\mathbf {X}} \left[ \sum_ {s = 1} ^ {S} \beta_ {s} ^ {k} \right]}{\sum_ {s = 1} ^ {S} \mathbb {E} _ {\mathbf {X}} \left[ \alpha_ {s} \right]}, \epsilon_ {K}\right), \tag {16} +$$ + +with $\epsilon_S$ and $\epsilon_K$ hyperparameters setting the smallest acceptable relative use frequency for a slot or prototype. + +The full regularization loss is a weighted sum of the three losses described above: + +$$ +\mathcal {L} _ {\text {r e g}} = \lambda_ {\text {a c t}} \mathcal {L} _ {\text {a c t}} + \lambda_ {\text {s l o t}} \mathcal {L} _ {\text {s l o t}} + \lambda_ {\text {p r o t o}} \mathcal {L} _ {\text {p r o t o}}, \tag {17} +$$ + +where we use $\lambda_{\mathrm{act}} = 10^{-4}$ , $\lambda_{\mathrm{slot}} = \lambda_{\mathrm{proto}} = 0.1$ and $\epsilon_S = \epsilon_K = 0.1$ in all our experiments on the Earth Parser Dataset. + +# 3.3. Training and Implementation Details + +Model configuration. We process the input point cloud $\mathbf{X}$ using a similar architecture as in [53]. We first voxelize it with a $64\times 64\times 64$ grid and map it to a vector using a sequence of 6 3D sparse convolutions [9] and 6 strided convolutions. The resulting representation is then transformed using one linear layer for each slot. These features are decoded by simple 2-layer MLPs: one generates the distribution parameters $\alpha_{s}$ and $\beta_s^k$ , and the other ones the parameters of the 3D transformations $\mathcal{T}_s(X)$ . The transformations include an anisotropic scaling, a $y$ -axis tilt of $\pm \pi /10$ , a rotation around the $z$ -axis, and a translation, in this specific order. We use $S = 64$ slots and $K = 6$ prototypes as default parameters. See the supplementary material for details. + +The intensity of the return signal of each point is available in the LiDAR scans. We associate each prototype with a single learnable intensity parameter and perform the Chamfer distance (Equation 6) in 4 dimensions: spatial coordinates normalized to $[0,1]^3$ and intensity to $[0,0.1]$ . + +Curriculum learning. The network predicts simultaneously the slot's probability distributions and their deformations. This results in many concurrent degrees of freedom and can make the training process unstable. Therefore, following Monnier et al. [49] and Loiseau et al. [43], we implement a multi-stage curriculum learning strategy. We first + +![](images/1922eaf39f98874f35a5cbc8f06007a23c6a9c6d444bf126e65343b5b5ddd33b.jpg) +Figure 4. Reconstruction Quality. We show two partial scenes with their RGB and intensity values, as well as their reconstruction by our method and competing models. We use the prototypes' intensity to color the points or pixels. As SuperQuadrics does not model the intensity, we use a random colour for each quadric. + +initialize the prototypes as point clouds uniformly sampled from a random cuboid and gradually unfreeze the model parameters in the following order: (i) translation, rotation, tilt, slot activation, and choice of prototype; (ii) intensities of the prototypes, when available; (iii) scales of the prototypes; (iv) positions of the prototypes' 3D points; (v) anisotropic scalings of the prototypes. As shown in Section 4, each step of this curriculum scheme improves the performances. + +Prototypes selection. We automatically select the number of prototypes for a complete scene using a simple greedy algorithm. We measure the increase of reconstruction loss when preventing the model from selecting each prototype individually. We remove the prototype with the lowest increase if it is lower than $5\%$ , and iterate. + +Implementation details. Our model is trained separately for each scene by randomly sampling square patches. During training, the patches are subsampled to a maximum $10^{5}$ points. Each stage of the curriculum is trained until convergence. We use the ADAM optimizer [28] with a learning rate of $10^{-4}$ and default parameters. See the supplementary material for preprocessing, training and evaluation details. + +# 4. Results + +In this section, we assess the ability of our method to parse complex 3D aerial data. We give in Section 4.1 an overview of our proposed dataset of aerial LiDAR scans. In Section 4.2, we then discuss our evaluation metrics and baselines. Finally, we present a quantitative (Section 4.3) and qualitative (Section 4.4) analysis of our results. + +# 4.1. Earth Parser Dataset + +We introduce a new dataset to train and evaluate parsing methods on large, uncurated aerial LiDAR scans. We use data from the French Mapping Agency associated to the + +LiDAR-HD project [45]. Each scan is composed of several airborne LiDAR acquisitions taken at different angles, leading to a minimum resolution of 20 points/ $\mathrm{m}^2$ . The points are associated with their laser reflectance (intensity), and colorized based on asynchronous aerial photography. + +We selected 7 scenes, covering over $7.7\mathrm{km}^2$ and a total of 98 million 3D points, with diverse content and complexity, such as dense habitations, forests, or complex industrial facilities. We associate most 3D points with a coarse semantic label, such as ground, building, or vegetation. The characteristics of the scenes are detailed in Table 2 and each is visualized in Figure 3. + +# 4.2. Evaluation Metrics and Baselines + +We quantitatively evaluate the performances for reconstruction and semantic segmentation of our model and several unsupervised scene decomposition approaches. + +Evaluation metrics. As our goal is to summarize a point cloud using few prototypes, the quality of the reconstruction is critical. We measure it with the symmetric Chamfer distance ("Cham." in the Tables) between the input and the output point clouds of our model. + +By associating a class to each prototype's points, we can propagate labels from the reconstruction to the input cloud and perform semantic segmentation. We evaluate the quality of this segmentation with the class-averaged Intersection-over-Union (mIoU) metric. + +In a practical scenario, an operator can manually annotate the points of the 3D prototypes, allowing for the segmentation of the entirety of $\mathbf{X}$ with minimal effort. To perform automatic evaluation, we follow the standard practice for evaluating clustering and unsupervised segmentation methods [25, 31, 48]: we assign to each prototype's point the most frequent class of its closest point of the input after the reconstruction. + +Table 1. Results on the Earth Parser Dataset. We report the quality of the reconstruction (Cham.) and semantic segmentation (mIoU) on each of the scenes of our Earth Parser Dataset. While our method does not always provide the most faithful reconstructions, it leads to the most accurate point classification. + +
Rec.SemanticCrop FieldsForestGreenhousesMarinaPower PlantUrbanWindturbines
Cham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoU
k-means (i,z) [44]X✓—93.8—71.5—39.3—41.4—42.8—56.5—87.6
SuperQuadrics [53]3DX0.86—1.04—0.60—0.93—0.58—0.40—13.5—
DTI-Sprites [49]2.5D+i✓6.1083.214.5940.25.3642.06.1641.45.3629.02.9947.336.1925.9
AtlasNet v2 [11]3D+i✓1.0743.11.5871.40.5649.10.7342.10.4541.60.6348.89.4748.1
Ours3D+i✓0.7296.90.8883.70.4091.30.8278.70.4452.20.2983.26.6593.4
+ +Table 2. Earth Parser Dataset. Our proposed dataset is composed of 7 diverse scenes acquired by aerial LiDAR. + +
NameSurface in km2# points ×106annotation ratio in %num. of classes
Crop Fields1.119.777.42
Forest1.146.797.82
Greenhouses0.11.395.63
Marina0.10.592.72
Power Plant0.28.678.44
Urban1.115.795.93
Wind turbines4.25.699.83
Total7.798.391.6—
+ +Baselines. We adapt several unsupervised approaches for scene reconstruction and/or semantic segmentation tasks to provide comparisons for our approach: + +- k-means. We cluster the points of the input with the k-means algorithm [44] using as many clusters as we use prototypes. We obtain the best results by using a combination of the point's intensity and elevation as features for clustering. We then assign to each centroid its most frequent class, and propagate this label to the entire cluster, leading to a semantic segmentation. This method does not reconstruct the input, but gives us a simple and surprisingly strong baseline for semantic segmentation. +- SuperQuadrics revisited. We use the method of Paschalidou et al. [53] to learn to approximate scenes with an adaptive number of superquadrics [3]. It provides a baseline for reconstruction and a qualitative comparison for instance segmentation, shown in supplementary material. +- DTI-Sprites. We use the point cloud to construct a digital elevation model, i.e. a 2.5D image of resolution $32 \times 32$ where each pixel has an elevation and intensity value. We adapt the unsupervised image decomposition approach of Monnier et al. [49] to break down this image into a set of 2.5D sprites. We evaluate the reconstruction and segmentation by sampling 25 3D point per pixel, transferring the pixel's label to the points, and interpolating their elevations. +- AtlasNet v2. This extension [11] of AtlasNet [18] uses a fixed number of learnable prototype point clouds to recon + +Table 3. Ablation Study. We evaluate the effect of our prototype selection post-processing, our model's degrees of freedom, and our different regularization losses. + +
UrbanMarina
Cham.mIoUCham.mIoU
expressivityLearnable Earth Parser0.2983.20.8278.7
→w/o post-processing0.2883.70.9678.3
→w/o aniso-scale0.3382.41.0467.2
→w/o prototypes0.3668.31.0742.8
→w/o scales0.5558.91.3340.8
→w/o intensities0.5558.71.0940.8
losses→w/o Lact0.1754.10.8056.9
→w/o Lslot0.2577.80.8143.7
→w/o Lproto0.2857.20.9740.7
+ +struct its input. It can be evaluated for both reconstruction and semantic segmentation in a way similar to ours. We extend it to handle intensity in a manner akin to our approach, which improves its segmentation results. + +Similar to our method, we train all baselines except k-means by sampling square patches in each scene. Figure 4 shows the output of the reconstruction methods. + +# 4.3.Quantitative Results + +We compare the performance of our approach with the proposed baselines on the Earth Parser Dataset, as well as two publicly available datasets. + +Earth Parser Dataset. We provide quantitative reconstruction and semantic segmentation results in Table 1, and illustrations in Figure 4. Despite being highly constrained, our model yields the best reconstruction in 6 out of 7 scenes. Moreover, we significantly outperform the other methods for semantic segmentation across all scenes. + +Despite its simplicity, the k-means baseline provides strong semantic segmentation performance, beating the other baselines in 5 of 7 annotated scenes. DTI-Sprites [49] has lower reconstruction and segmentation quality, which + +Table 4. Synthetic Shapes. We train our method on all planes from ShapeNet-Part [58], with random rotations around $z$ -axis. We show the reconstruction of an input plane and the prototypes learned on the dataset. + +
AtlasNet v2 [11]SuperQuadrics [53]Ours
Recons.
Protos.---
Cham.1.462.911.34
mIoU34.5-68.6
+ +
Cham.mIoU
k-means (i,z) [44]—52.7
Superquadrics [53]0.33—
DTI-Sprites [49]2.2553.6
AtlasNet v2 [11]0.3154.9
Ours0.2982.7
+ +![](images/c0e64fd1741ce88e0d62b422b28e7456390c76b85e6646b1463fc89c09278a09.jpg) +Figure 5. Results on DALES [68]. We report quantitative and qualitative results for one tile from DALES. + +is expected as it models a 3D point cloud in 2.5D. AtlasNet v2 [11] provides good reconstructions but segmentation fails for scenes such as Crop Fields or Urban due its inability to adapt its prototype usage to the input. On the contrary, SuperQuadrics [53] can adjust the number of superquadric it uses and, to some degree, their shape. However, this method uses a single parametric family for all prototypes and fails to reconstruct complex real-world scenes such as Power Plant. Thanks to its probabilistic slot selection, our method can handle inputs with a varying number of objects using only a small set of learned prototypical shapes. + +Ablation study. We evaluate the impact of various components of our model and report the results in Table 3. First, we observe that the prototype selection post-processing has limited impact on the quality of the prediction and reconstructions, but allows us to adapt and significantly decrease the number of prototypes used for each scene. Second, we evaluate the impact of reducing the expressivity of our model. We successively remove: (i) the anisotropic scaling of $\mathcal{T}$ , and the possibility of learning the prototype's (ii) points position, (iii) scale, and (iv) intensity. As expected, each degree if liberty removed decreases the quality of the 3D reconstruction and segmentation. + +We study the impact of the different regularization introduced in Section 3.2. The losses related to slot activation have a marginal effect on reconstruction quality but + +significantly affect the performance of semantic segmentation. This suggests that using slots sparingly but equally is important for prototypes to specialize for specific objects, but is not necessary to the expressivity of the reconstruction model. The prototype activation loss prevents unused and possibly degenerate prototypes and therefore improves both segmentation and reconstruction. + +We evaluate the impact of reducing the number of prototypes or slots. When we set $K = 3$ instead of 6, the average semantic segmentation mIoU drops by 26.3 points, while reconstruction is minimally affected, showing only a $2.3\%$ increase in Chamfer distance. Conversely, setting $S = 32$ instead of 64 leads to a significant $+47.3\%$ increase of Chamfer distance, with a modest drop 0.4 of semantic segmentation mIoU. These findings highlight the importance of maintaining sufficient diversity in the prototypes to enable specialization for different object types, whereas the number of slots strongly influences the expressiveness of the reconstruction model. + +ShapeNet. We evaluate our model on 2690 planes from ShapeNet-Part [58], whose points are annotated as wing, engine, tail, or body. We randomly rotate the shapes around the $z$ -axis during training and evaluation. We report in Table 4 the performances and reconstructions for our approach, AtlasNet v2 [11], and SuperQuadrics revisited [53]. Our method handle rotations better than AtlasNet v2, and manages to successfully locate the tail of the planes. While SuperQuadrics' reconstructions make sense qualitatively, they are worse in terms of accuracy than ours, and do not enable semantic segmentation. + +DALES. We also train and evaluate our model on DALES [68], a dataset of aerial LiDAR scans of urban area, with a restricted class set: ground, vegetation, and buildings. Our model significantly outperforms competing reconstruction-based approaches, and yields similar performance than for the Urban tile of the Earth Parser Dataset; see Figure 5. + +# 4.4. Qualitative Results + +Instance segmentation. We can perform instance segmentation simply by considering each slot as a different instance. This is particularly interesting for parsing natural woodlands, a key endeavor for forest management [50] and biomass estimation [12]. While this task has a long history of handcrafted approaches [69], current deep learning approaches are mostly limited to artificial or low-density forests [65]. As shown qualitatively in Figure 6, our Learnable Earth Parser can learn without any supervision to separate individual trees in dense forests, or boats in a marina. We evaluate quantitatively the performance of our model in terms of instance segmentation, we counted manually 88 + +![](images/edd8816a7c438417c0fb5ddd49665ba2b19559f637662b1f52c3c0dc3af2ee50.jpg) +Figure 6. Instance Segmentation. We can identify the reconstruction of each slot as a separate instance, allowing us to perform instance segmentation of complex data such as dense natural forests or a marina. For this visualization, we considered the points associated to "trees" or "boat hull" prototypes and color each of their instance randomly. + +boat instances in 10 distinct portions of the Marina scene. We then computed the Mean Relative Error (MRE) of the prediction given by the number of boat-like prototypes (prototype #3 in Figure 8) in these zones. Our method yields a MRE of $7,4\%$ , 2.6 times lower than DTI-Sprites [49], demonstrating its strong performances. + +Semantic segmentation. After annotating the prototypes, our model can perform convincing semantic segmentation of complex scenes, as shown in Figure 7. + +Interpretable prototypes. In Figure 8, we show prototypes learned on our Earth Parser Dataset with colors showing the associated semantic label for each point. These prototypes give at a glance insights on the content of these real-world scenes. Our model is able to learn a wide variety of shapes, such as boats' masts, wind turbines, or greenhouses. We also observe that the prototypes are typically associated with a single object type. Empirically, the average class distribution of the nearest neighbors of each prototypes' point exhibits the same normalized entropy (0.22) as a Bernoulli distribution with probability 0.964. The highly constrained nature of the deformations $\mathcal{T}$ prevents slots from repurposing the same prototypes for different objects, making the learned prototypes easy to identify and manually annotate. + +# 5. Conclusion + +We introduced a novel unsupervised method for parsing complex real-world aerial scans into simple parts using a small set of learned prototypical shapes. We demonstrate + +![](images/d0f9e92889933ddcee42309ba4271befd917345318a3674e19da68a0fed51349.jpg) + +![](images/7a19f73eddc8fa3dec4ac0e97498ea4647b08baffbed6866f693077d4a985ad0.jpg) +Figure 7. Semantic Segmentation. Our model can perform semantic segmentation of large real-world scene based on annotated prototypes. Black points in the ground truth are unannotated. +Figure 8. Learned Prototypes. We display three of the prototypes chosen during the selection process for various scenes. + +the quality and interpretability of our results on a novel dataset of aerial LiDAR scans. To the best of our knowledge, we are the first to demonstrate the possibility of performing deep unsupervised 3D shape analysis on such a challenging real-world dataset. We believe that our results open new perspectives for computer-assisted environment monitoring and economic intelligence. + +Acknowledgements This work was supported in part by ANR project READY3D ANR-19-CE23-0007, ANR under the France 2030 program under the reference ANR-23-PEIA-0008, and was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012096R2 made by GENCI. The work of MA was partly supported by the European Research Council (ERC project DISCOVER, number 101076028). The scenes of Earth Parser Dataset were acquired and annotated by the LiDAR-HD project [45]. We thank Zenodo for hosting the dataset. We thank Zeynep Sonat Baltaci, Emile Blettery, Nicolas Dufour, Antoine Guedon, Helen Mair Rawsthorne, Tom Monnier, Damien Robert, Mathis Petrovich and Yannis Siglidis for inspiring discussions and valuable feedback. + +# References + +[1] Antonio Alliegro, Davide Boscaini, and Tatiana Tommasi. Joint supervised and self-supervised learning for 3D real world challenges. In ICPR, 2021. 1 +[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1 +[3] Alan H Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1981. 2, 6 +[4] Pál Benkő, Ralph R Martin, and Tamás Várady. Algorithms for reverse engineering boundary representation models. Computer-Aided Design, 2001. 2 +[5] Thomas Binford. Visual perception by computer. In Proc. IEEE Conf. on Systems and Control, 1975. 2 +[6] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 2 +[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2 +[8] YH Chen and CY Liu. Quadric surface extraction using genetic algorithms. Computer-Aided Design, 1999. 2 +[9] Spconv Contributors. SPConv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4 +[10] Yang Cui, Qingquan Li, Bisheng Yang, Wen Xiao, Chi Chen, and Zhen Dong. Automatic 3-d reconstruction of indoor environment with mobile laser scanning point clouds. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 2 +[11] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. NeurIPS, 2019. 1, 2, 5, 6, 7 +[12] Antonio Ferraz, Sassan Saatchi, Clément Mallet, Stéphane Jacquemoud, Gil Gonçalves, Carlos Alberto Silva, Paula Soares, Margarida Tomé, and Luisa Pereira. Airborne Li-DAR estimation of aboveground forest biomass in the absence of field inventory. Remote Sensing, 2016. 7 +[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 2 +[14] Ran Gal, Olga Sorkine, Niloy J Mitra, and Daniel Cohen-Or. iWIRES: An analyze-and-edit approach to shape manipulation. In ACM SIGGRAPH. 2009. 2 +[15] Weixiao Gao, Liangliang Nan, Bas Boom, and Hugo Ledoux. SUM: A benchmark dataset of semantic urban meshes. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 3 +[16] Xuming Ge, Bo Wu, Yuan Li, and Han Hu. A multiprimitive-based hierarchical optimal approach for semantic labeling of ALS point clouds. Remote Sensing, 2019. 2 + +[17] Paulo FU Gotardo, Olga Regina Pereira Bellon, and Luciano Silva. Range image segmentation by surface extraction using an improved robust estimator. In CVPR, 2003. 2 +[18] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 6 +[19] Stéphane Guinard, Loic Landrieu, Laurent Caraffa, and Bruno Vallet. Piecewise-planar approximation of large 3d data as graph-structured optimization. ISPRS Annals of Photogrammetry, Remote Sensing & Spatial Information Sciences, 2019. 2 +[20] Florent Guiotte, Minh-Tan Pham, Romain Dambreville, Thomas Corpetti, and Sébastien Lefèvre. Semantic segmentation of LiDAR points clouds: Rasterization beyond digital elevation models. IEEE Geoscience and Remote Sensing Letters, 2020. 2 +[21] Bo Guo, Qingquan Li, Xianfeng Huang, and Chisheng Wang. An improved method for power-line reconstruction from point cloud data. Remote sensing, 2016. 2 +[22] Abhinav Gupta, Alexei Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In ECCV, 2010. 2 +[23] Norbert Haala and Claus Brenner. Extraction of buildings and trees in urban environments. ISPRS journal of photogrammetry and remote sensing, 1999. 2 +[24] Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3D: large-scale building reconstruction from airborne LiDAR point clouds. Remote Sensing, 2022. 2 +[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. CVPR, 2019. 5 +[26] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3D data. In Computer Graphics Forum. Wiley Online Library, 2019. 2 +[27] Ekaterina Kalinicheva, Loic Landrieu, Clément Mallet, and Nesrine Chehata. Multi-layer modeling of dense vegetation from aerial lidar scans. In CVPR Workshops, 2022. 3 +[28] Diederik P Kingma and Jimmy Ba. ADAM: A method for stochastic optimization. *ICLR*, 2014. 5 +[29] Florian Kluger, Hanno Ackermann, Eric Brachmann, Michael Ying Yang, and Bodo Rosenhahn. Cuboids revisited: Learning robust 3D shape fitting to single RGB images. In CVPR, 2021. 2 +[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 1 +[31] Adam Kosiorek, Sara Sabour, Yee Whye Teh, and Geoffrey E Hinton. Stacked capsule autoencoders. NeurIPS, 2019. 5 +[32] Florent Lafarge, Xavier Descombes, Josiane Zerubia, and Marc Pierrot-Deseilligny. Automatic 3D buildup reconstruction from DEMs: An application to PLEIADES simulations. In ISPRS, 2006. 2 +[33] Aleš Leonardis, Alok Gupta, and Ruzena Bajcsy. Segmentation of range images as the search for geometric parametric models. International Journal of Computer Vision, 1995. 2 + +[34] Ales Leonardis, Ales Jaklic, and Franc Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1997. 2 +[35] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics, 2017. 2 +[36] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3D point clouds. In CVPR, 2019. 2 +[37] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In ECCV, 2016. 2 +[38] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. In ACM SIGGRAPH. 2011. 2 +[39] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 +[40] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In CVPR, 2022. 2 +[41] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 3 +[42] Suresh K Lodha, Edward J Kreps, David P Helmbold, and Darren Fitzpatrick. Aerial LiDAR data classification using support vector machines. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06). IEEE, 2006. 2 +[43] Romain Loiseau, Tom Monnier, Mathieu Aubry, and Loic Landrieu. Representing shape collections with alignment-aware linear models. In 3DV, 2021. 1, 2, 3, 4 +[44] J MacQueen. Classification and analysis of multivariate observations. In 5th Berkeley Symp. Math. Statist. Probability, 1967. 6, 7 +[45] IGN (French mapping agency). LiDAR-HD: A 3D mapping of France's soil and subsoil, 2021. 5, 8 +[46] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J.Davison. SceneNet RGB-D: Can 5M synthetic images beat generic imagenet pre-training on indoor segmentation? ICCV, 2017. 1 +[47] Jie Mei, Liqiang Zhang, Shihao Wu, Zhen Wang, and Liang Zhang. 3d tree modeling from incomplete point clouds via optimization and 11-mst. International Journal of Geographical Information Science, 2017. 2 +[48] Tom Monnier, Thibault Groueix, and Mathieu Aubry. Deep Transformation-Invariant Clustering. NeurIPS, 2020. 4, 5 +[49] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In ICCV, 2021. 3, 4, 5, 6, 7, 8 +[50] Felix Morsdorf, Erich Meier, Benjamin Kötz, Klaus I Itten, Matthias Dobbertin, and Britta Allgower. LIDAR-based geometric reconstruction of boreal type forest stands at single + +tree level for forest and wildland fire management. Remote sensing of environment, 2004. 7 +[51] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In ICCV, 2017. 2 +[52] Joachim Niemeyer, Franz Rottensteiner, and Uwe Soergel. Contextual classification of LiDAR data and building object detection in urban areas. ISPRS journal of photogrammetry and remote sensing, 2014. 3 +[53] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, 2019. 1, 2, 3, 4, 5, 6, 7 +[54] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In CVPR, 2020. 2 +[55] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 2 +[56] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. Monteboxfinder: Detecting and filtering primitives to fit a noisy point cloud. In ECCV. Springer, 2022. 2 +[57] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2 +[58] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR Workshop, 2015. 7 +[59] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR 2015 Workshop on Functionality, Physics, Intentionality and Causality, 2015. 1 +[60] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. In Computer Graphics Forum, 2007. 2 +[61] Ruwen Schnabel, Patrick Degener, and Reinhard Klein. Completion and reconstruction with primitive shapes. In Computer Graphics Forum. Wiley Online Library, 2009. 2 +[62] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 2 +[63] Nina M Singer and Vijayan K Asari. DALES objects: A large scale benchmark dataset for instance segmentation in aerial LiDAR. IEEE Access, 2021. 3 +[64] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 3 +[65] Chenxin Sun, Chengwei Huang, Huaiqing Zhang, Bangqian Chen, Feng An, Liwen Wang, and Ting Yun. Individual tree crown segmentation and crown width extraction from a heightmap derived from aerial laser scanning data using a deep learning framework. Frontiers in plant science, 2022. 7 +[66] Minhyuk Sung, Vladimir G Kim, Roland Angst, and Leonidas Guibas. Data-driven structural priors for shape completion. ACM Transactions on Graphics, 2015. 2 +[67] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2 + +[68] Nina Varney, Vijayan K Asari, and Quinn Graehling. DALES: A large-scale aerial LiDAR data set for semantic segmentation. In CVPR Workshop, 2020. 3, 7 +[69] Jari Vauhkonen, Liviu Ene, Sandeep Gupta, Johannes Heinzel, Johan Holmgren, Juho Pitkänen, Svein Solberg, Yunsheng Wang, Holger Weinacker, K Marius Hauglin, et al. Comparative testing of single-tree detection algorithms under different types of forest. Forestry, 2012. 7 +[70] Ben G Weinstein, Sergio Marconi, Stephanie Bohlman, Alina Zare, and Ethan White. Individual tree-crown detection in RGB imagery using semi-supervised deep learning neural networks. Remote Sensing, 2019. 3 +[71] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In ECCV, 2022. 2 +[72] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 2 +[73] Shaobo Xia, Dong Chen, Ruisheng Wang, Jonathan Li, and Xinchang Zhang. Geometric primitives in LiDAR point clouds: A review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020. 2 +[74] Zhen Ye, Yusheng Xu, Rong Huang, Xiaohua Tong, Xin Li, Xiangfeng Liu, Kuifeng Luan, Ludwig Hoegner, and Uwe Stilla. LASDU: A large-scale aerial LiDAR dataset for semantic labeling in dense urban areas. ISPRS International Journal of Geo-Information, 2020. 3 +[75] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised Discovery of Object Radiance Fields. *ICLR*, 2022. 3 +[76] Yu Zhang, Peter Tino, Aleks Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1 +[77] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In ICCV, 2021. 1 +[78] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3D point capsule networks. In CVPR, 2019. 2 +[79] SM Zolanvari, Susana Ruano, Aakanksha Rana, Alan Cummins, Rogerio Eduardo da Silva, Morteza Rahbar, and Aljosa Smolic. DublinCity: annotated LiDAR point cloud and its applications. BMVC, 2019. 3 +[80] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In ICCV, 2017. 2 \ No newline at end of file diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/images.zip b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..03d372d61539b7c262bb824f33472f8e15a206dc --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:246a45c15bd5a828f2982bef2b7dd556bae92ea8b9cc931a6f932f8eae204464 +size 724992 diff --git a/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/layout.json b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ded176fda6274792c0f1363a03ff0c18a2ca82e2 --- /dev/null +++ b/2024/Learnable Earth Parser_ Discovering 3D Prototypes in Aerial Scans/layout.json @@ -0,0 +1,10891 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "type": "text", + "content": "Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 143, + 154, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 143, + 154, + 156 + ], + "spans": [ + { + "bbox": [ + 60, + 143, + 154, + 156 + ], + "type": "text", + "content": "Romain Loiseau1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 162, + 162, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 162, + 162, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 162, + 172 + ], + "type": "text", + "content": "romain.loiseau@enpc.fr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 143, + 276, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 143, + 276, + 156 + ], + "spans": [ + { + "bbox": [ + 197, + 143, + 276, + 156 + ], + "type": "text", + "content": "Elliot Vincent1,3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 162, + 290, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 162, + 290, + 172 + ], + "spans": [ + { + "bbox": [ + 184, + 162, + 290, + 172 + ], + "type": "text", + "content": "elliot.vincent@enpc.fr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 143, + 402, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 143, + 402, + 157 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 402, + 157 + ], + "type": "text", + "content": "Mathieu Aubry" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 163, + 414, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 163, + 414, + 172 + ], + "spans": [ + { + "bbox": [ + 312, + 163, + 414, + 172 + ], + "type": "text", + "content": "mathieu.aubry@enpc.fr" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 446, + 143, + 529, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 143, + 529, + 156 + ], + "spans": [ + { + "bbox": [ + 446, + 143, + 529, + 156 + ], + "type": "text", + "content": "Loic Landrieu" + }, + { + "bbox": [ + 446, + 143, + 529, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 437, + 163, + 539, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 163, + 539, + 172 + ], + "spans": [ + { + "bbox": [ + 437, + 163, + 539, + 172 + ], + "type": "text", + "content": "loic. landrieu@enpc.fr" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 146, + 179, + 446, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 179, + 446, + 194 + ], + "spans": [ + { + "bbox": [ + 146, + 179, + 446, + 194 + ], + "type": "text", + "content": "1 LIGM, Ecole des Ponts, Univ Gustave Eiffel, CNRS, France" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 168, + 194, + 425, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 194, + 425, + 209 + ], + "spans": [ + { + "bbox": [ + 168, + 194, + 425, + 209 + ], + "type": "text", + "content": "2 Univ Gustave Eiffel, IGN, ENSG, LASTIG, France" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 241, + 210, + 350, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 210, + 350, + 224 + ], + "spans": [ + { + "bbox": [ + 241, + 210, + 350, + 224 + ], + "type": "text", + "content": "3 INRIA Paris, France" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "spans": [ + { + "bbox": [ + 143, + 252, + 192, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 277, + 290, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 290, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 290, + 494 + ], + "type": "text", + "content": "We propose an unsupervised method for parsing large 3D scans of real-world scenes with easily-interpretable shapes. This work aims to provide a practical tool for analyzing 3D scenes in the context of aerial surveying and mapping, without the need for user annotations. Our approach is based on a probabilistic reconstruction model that decomposes an input 3D point cloud into a small set of learned prototypical 3D shapes. The resulting reconstruction is visually interpretable and can be used to perform unsupervised instance and low-shot semantic segmentation of complex scenes. We demonstrate the usefulness of our model on a novel dataset of seven large aerial Li-DAR scans from diverse real-world scenarios. Our approach outperforms state-of-the-art unsupervised methods in terms of decomposition accuracy while remaining visually interpretable. Our code and dataset are available at https://romainloiseau.fr/learnable-earth-parser/." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 128, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 653 + ], + "type": "text", + "content": "Modern aerial 3D scanning technologies open up unprecedented opportunities for environmental monitoring and economic intelligence. However, their practical use remains challenging due to the complexity of real-world scenes, the diversity of usage scenarios, and the difficulty of annotation. Therefore, our aim is to develop an approach that could help perform diverse tasks—from counting trees in a forest or identifying the various components of a factory to measuring the surface of greenhouses or monitor urban growth—all without human supervision." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "To do so, we address two important limitations of existing 3D deep learning methods. First, they are often primarily designed, trained, and tested on synthetic [7, 46, 59, 72] or highly curated data [2, 30, 39], which fail to capture the endless variability of the real world. Moreover, they often" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 254, + 552, + 534 + ], + "blocks": [ + { + "bbox": [ + 307, + 254, + 552, + 534 + ], + "lines": [ + { + "bbox": [ + 307, + 254, + 552, + 534 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 552, + 534 + ], + "type": "image", + "image_path": "8501e7e11e85de94600e8c49e4cab8f1f2be7c828e79f17b6869bedaeebe3cf2.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 537, + 547, + 594 + ], + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 594 + ], + "type": "text", + "content": "Figure 1. Learnable Earth Parser. Our unsupervised method takes large aerial 3D scans as input and model them with a small set of learned 3D prototypes. Our approach is trained without annotation and produce legible decompositions of complex scenes, which can be used for semantic and instance segmentation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 689 + ], + "type": "text", + "content": "assume that annotations are available for tasks of interest. Second, even unsupervised approaches [1, 77] often rely on learning abstract feature representations, making them difficult to interpret [76]. Although some work has attempted to decompose 3D shapes into meaningful components without supervision [11, 43, 53, 67], they were all designed on simple synthetic shapes and none generalizes to real data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 714 + ], + "type": "text", + "content": "To overcome these limitations, we present the Learnable Earth Parser, an unsupervised deep learning method" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27874" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 68, + 521, + 192 + ], + "blocks": [ + { + "bbox": [ + 73, + 68, + 521, + 192 + ], + "lines": [ + { + "bbox": [ + 73, + 68, + 521, + 192 + ], + "spans": [ + { + "bbox": [ + 73, + 68, + 521, + 192 + ], + "type": "image", + "image_path": "b8ff0cd769002fc64a16aecb1caf1d09aa06d61cd036dba3a5c47ff5ee233c19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": "Figure 2. Method Overview. Our model approximates an input point cloud " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " slot models. Each slot maps " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " to an affine 3D deformation " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": ", a slot activation probability " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": ", and the joint probabilities " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\beta_s^1,\\dots,\\beta_s^K" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " of the slot being activated and choosing one of the " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " learnable prototype point clouds " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^1,\\dots,\\mathbf{P}^K" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " of an activated slot " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " is obtained by applying the transformation " + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 194, + 547, + 240 + ], + "type": "text", + "content": " to its most likely prototype. Non-activated slots do not contribute to the output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 249, + 288, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 288, + 332 + ], + "type": "text", + "content": "designed to decompose large-scale 3D point clouds into interpretable parts. Our model learns a small set of 3D prototypical shapes that are selected, positioned, rotated, and resized to reconstruct an input point cloud. We introduce a novel probabilistic formulation that enables the design of a reconstruction loss for learning jointly the 3D prototypes, but also to select and position them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 333, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 465 + ], + "type": "text", + "content": "To evaluate the effectiveness of our approach, we created a new open-access dataset consisting of 7 aerial LiDAR scans, covering " + }, + { + "bbox": [ + 46, + 333, + 287, + 465 + ], + "type": "inline_equation", + "content": "7.7\\mathrm{km}^2" + }, + { + "bbox": [ + 46, + 333, + 287, + 465 + ], + "type": "text", + "content": " and containing 98 million 3D points with annotations in diverse urban and natural environments. Our results demonstrate that the Learnable Earth Parser learns decompositions superior to traditional and deep learning baselines, leading to convincing performance for semantic and instance segmentation, as shown in Figure 1. We believe that our contributions provide researchers and practitioners with new tools and resources to tackle the challenges of real-world 3D data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 475, + 132, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 132, + 488 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 132, + 488 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 579 + ], + "type": "text", + "content": "Our proposed unsupervised method uses point cloud reconstruction as a proxy to learn to decompose large aerial point clouds and is evaluated on a novel and diverse dataset of 3D scans. In the following, we briefly present related works for primitive-based point cloud decomposition, automatic decomposition of LiDAR data, and an overview of existing aerial LiDAR datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Primitive-based point cloud decomposition. Modeling shapes as a set of primitives such as generalized cylinders [5] or superquadrics [3] has a rich history in vision and graphics [26]. Classical applications include reverse engineering [4], shape completion [61, 66], and shape editing [14]. A variety of methods have been developed to find primitives in unstructured 3D scenes, including seed growing techniques [33, 34], genetic algorithms [8], approaches [17, 38, 56, 60] based on RANSAC [13], and probabilistic methods [40, 71]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 249, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 441 + ], + "type": "text", + "content": "For this problem, like for many others in computer vision, deep learning has become the dominant paradigm. However, supervised methods [36, 80] are limited by the availability of annotated datasets. Following the seminal work of Tulsiani et al. [67], unsupervised approaches that simply rely on a reconstruction loss to learn primitive decomposition are the most common and the most closely related to our work [53, 55]. An important challenge for these approaches is to model a variable number of primitives. This has been addressed using recurrent networks [35, 62, 80], capsule networks [78], reinforcement learning strategies [67] or, most similar to our approach, computing the Chamfer distance based on a probabilistic model [53]. Our work is also related to approaches that learn prototypical shapes instead of being restricted to a predefined family of parametric primitives [11, 43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 514 + ], + "type": "text", + "content": "However, most of these methods are designed, trained and evaluated on well-curated synthetic object datasets, such as ModelNet [72], ShapeNet [7], or D-FAUST [6], and are typically designed to handle single objects from known categories. In contrast, our approach can handle complex scenes composed of many objects with significant variety." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 713 + ], + "type": "text", + "content": "Decomposition of LiDAR scans. Automatically decomposing large LiDAR scans poses unique challenges due to their size and diversity [73]. Some previous approaches use simple shape primitives, such as lines [10, 21], planes [16, 19, 51], or volumes [37], but these may not be flexible enough to capture the complexity of real-world scans. Other approaches are designed for specific object classes, like trees [47] or buildings [24, 32], but they are limited in their ability to represent a wide range of shapes. In contrast, our Learnable Earth Parser overcomes these limitations by learning ad-hoc prototypes for each new scene, ensuring both expressivity and adaptability. LiDAR data are often treated as digital elevation models, i.e. images with pixel elevations [20, 23, 42]. Thus, our work is related to image-based primitive prediction [22, 29, 54, 57] and unsupervised" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27875" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "multi-object image segmentation [41, 49, 64, 75]. However, 3D point clouds have higher precision and can better represent multi-layered structures such as forest areas." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "spans": [ + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "text", + "content": "Aerial LiDAR datasets. The increased availability of aerial LiDAR technology has led to the multiplication of open datasets [15, 52, 74, 79] of varying sizes from 1 to 10 " + }, + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "inline_equation", + "content": "\\mathrm{km}^2" + }, + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "text", + "content": " [63, 68]. However, these scans are limited to dense urban environments and do not capture the challenge of modeling diverse terrains. Some specialized datasets focus on forested areas [27, 70]. Our proposed dataset is of similar scale, spanning " + }, + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "inline_equation", + "content": "7.7\\mathrm{km}^2" + }, + { + "bbox": [ + 46, + 124, + 288, + 244 + ], + "type": "text", + "content": ", but covers a variety of urban, natural, and rural scenes, making it more representative of the diversity of possible usage scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 255, + 103, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 255, + 103, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 103, + 266 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "content": "Our goal is to learn to break down a point cloud into simpler and more easily understandable components. To achieve this, we propose an analysis-by-synthesis approach where we train a highly-constrained model " + }, + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 275, + 287, + 335 + ], + "type": "text", + "content": " to approximate a point cloud as a combination of learned 3D shapes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 342, + 265, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 265, + 355 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 265, + 355 + ], + "type": "text", + "content": "3.1. Probabilistic Scene Reconstruction Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": "As illustrated in Figure 2, our model first selects up to " + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": " shapes from " + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": " learnable 3D shapes, then positions and deforms them to best approximate an input point cloud " + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 361, + 287, + 445 + ], + "type": "text", + "content": ". We propose a probabilistic formulation of the selection process, which can be seen as an extension of the model of Paschalidou et al. [53] to multiple free-form shapes instead of a single parametric family." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": "Learnable shape prototypes. Following Loiseau et al. [43], we define " + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": " point clouds " + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^1,\\dots ,\\mathbf{P}^K" + }, + { + "bbox": [ + 46, + 460, + 287, + 533 + ], + "type": "text", + "content": " that we refer to as prototypes. Each prototype is meant to represent a single instance of a recurring 3D structure in the considered scene. The points' coordinates are free parameters of the model and learned directly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": "Scene reconstruction model. Our full model " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": " is the combination of " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": " reconstruction models " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": ", which we refer to as slots in analogy to the Slot Attention approach [41]. Each slot contributes to the final reconstruction only if it is activated. Slot activation is determined by a binary variable " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "a_s" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": " is activated if and only if " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "a_s = 1" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": ". The output of " + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{X})" + }, + { + "bbox": [ + 46, + 548, + 287, + 643 + ], + "type": "text", + "content": " is the combination of the reconstructions from all activated slots:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 653, + 287, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 653, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 113, + 653, + 287, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\mathbf {X}) = \\bigcup_ {\\substack {s = 1 \\dots S \\\\ a _ {s} = 1}} \\mathcal {M} _ {s} (\\mathbf {X}). \\tag{1}", + "image_path": "f362f72a0cfe63ee5652d4d28b395761598fa3f754172c24c828af6f00a1f6c2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Each slot model outputs a point cloud which is the deformation of one learnable prototype chosen from " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^1, \\dots, \\mathbf{P}^K" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 545, + 223 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 223 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 223 + ], + "type": "image", + "image_path": "ea9845e46b2a4f8fbafe1137999f9fb3b950160164bc24602c779f64c39471cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 231, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 546, + 285 + ], + "type": "text", + "content": "Figure 3. Earth Parser Dataset. Our dataset contains 7 scenes representing various urban and natural environments acquired by aerial LiDAR. The illustration of the power plant and the greenhouses display the complete scenes, while other ones display a subset of each scene (between 25 and " + }, + { + "bbox": [ + 305, + 231, + 546, + 285 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 231, + 546, + 285 + ], + "type": "text", + "content": " of the total area)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": "We associate to each slot " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " a network " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " which maps " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " to an affine transformation in 3D space " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " of slot " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " is determined by a variable " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "b_{s} \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "b_{s} = k" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": ", then the output of " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_s^k" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": ", the result of applying the transformation " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": " to the prototype " + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^k" + }, + { + "bbox": [ + 305, + 297, + 545, + 369 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 376, + 545, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 376, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 384, + 376, + 545, + 392 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {s} ^ {k} = \\mathcal {T} _ {s} (\\mathbf {X}) [ \\mathbf {P} ^ {k} ]. \\tag {2}", + "image_path": "d61999902d367389bd9d0210e02e0e760b36cc77b5c3211cc295db61ca5017eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "content": "Please note that " + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_s^k" + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "content": " is a function of " + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 401, + 545, + 425 + ], + "type": "text", + "content": ". However, to keep our notations simple, we omit this dependence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": "Probabilistic modeling. We make our reconstruction model probabilistic by modeling " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": " as random variables following (multi-)Bernoulli distributions. We call " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": " the probability that slot " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": " is activated and " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\beta_{s}^{k}" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": " the probability that it is activated and selects the prototype " + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 439, + 545, + 500 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 508, + 545, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 508, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 336, + 508, + 545, + 522 + ], + "type": "interline_equation", + "content": "p \\left(a _ {s} = 1\\right) = \\alpha_ {s}, \\quad p \\left(a _ {s} = 1, b _ {s} = k\\right) = \\beta_ {s} ^ {k}. \\tag {3}", + "image_path": "7dc292edeaeea66f4b69ef92387a9c94aa427ea3d370168c6e3f49737a932394.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": "For each slot " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": ", we predict the vector " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "(1 - \\alpha_{s},\\beta_{s}^{1},\\dots \\beta_{s}^{K})" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " with a neural network taking the point cloud " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " as input and finishing with a softmax layer. Again, we don't write the dependency of the " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\beta_{s}" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " explicitly to simplify the notations. The complete model " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " and the slots models " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " can now be seen as random variables, producing different potential reconstructions with probabilities given by " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": ". During inference, we consider only slots with " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\alpha_{s} > 0.5" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": " and select the prototype with highest " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\beta_s^k" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": ". However, during training, we compute all reconstructions " + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_s^k" + }, + { + "bbox": [ + 305, + 531, + 545, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 659, + 403, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 403, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 403, + 672 + ], + "type": "text", + "content": "3.2. Training Losses" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "Given a large 3D scene, we train our model by sampling square patches " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " from the scene. For each batch of patches, we minimize a loss composed of a reconstruction loss " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27876" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "and several regularization terms " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": " implementing different priors:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 102, + 287, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 287, + 117 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {M}) = \\mathbb {E} _ {\\mathbf {X}} \\left[ \\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) \\right] + \\mathcal {L} _ {\\text {r e g}} (\\mathcal {M}). \\tag {4}", + "image_path": "57f515d237853ab252931f30fa9e1913168894ea950d6bd6530fee395aa92934.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "type": "text", + "content": "Reconstruction loss. We define the reconstruction loss " + }, + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 47, + 129, + 287, + 152 + ], + "type": "text", + "content": " as the sum of two losses:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 158, + 287, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 158, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 158, + 287, + 171 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} (\\mathcal {M}, \\mathbf {X}) = \\mathcal {L} _ {\\text {a c c}} (\\mathcal {M}, \\mathbf {X}) + \\mathcal {L} _ {\\text {c o v}} (\\mathcal {M}, \\mathbf {X}). \\tag {5}", + "image_path": "31b6fc4bdb6256b75c89deb1f4877123e41828429c0efd1da7ae3df649f0ee59.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " encourages likely reconstructions of " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{X})" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " to accurately approximate " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cov}}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " ensures coverage, i.e., that each point of " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " is well-reconstructed by at least one activated model. We define each term using the asymmetric Chamfer distance " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " between two point clouds " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 47, + 177, + 287, + 237 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 251, + 287, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 251, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 97, + 251, + 287, + 277 + ], + "type": "interline_equation", + "content": "d (\\mathbf {X}, \\mathbf {Y}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\min _ {y \\in \\mathbf {Y}} \\| x - y \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "8f889701f0872e699597c6382cb9a9449e736b2a6720c1de6ee4dfca9d4ff3c7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "inline_equation", + "content": "|.|" + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "content": " the number of points in a point cloud. We write " + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "inline_equation", + "content": "d(x, \\mathbf{Y})" + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "content": " the distance between the point " + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "content": " and its closest point in " + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 47, + 282, + 287, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "content": " as the average over all slots " + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "content": " of the expected distance between " + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 318, + 287, + 342 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 348, + 287, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 348, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 74, + 348, + 287, + 407 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a c c}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {a _ {s}, b _ {s}} [ d (\\mathcal {M} _ {s} (\\mathbf {X}), \\mathbf {X}) ] (7) \\\\ = \\frac {1}{S} \\sum_ {s = 1} ^ {S} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(\\mathbf {Y} _ {s} ^ {k}, \\mathbf {X}\\right). (8) \\\\ \\end{array}", + "image_path": "6751ecbde883eb8b93d36a454a834422729abccb3533373e7b3aa1c3880c8d20.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "content": "Conversely, we define " + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cov}}" + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "content": " as the average over all points " + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "content": " of the expected distance between " + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 412, + 287, + 447 + ], + "type": "text", + "content": " and its closest point in the reconstruction:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 452, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 452, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 52, + 452, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\mathbb {E} _ {a, b} \\left[ \\min _ {s \\mid a _ {s} = 1} d (x, \\mathcal {M} _ {s} (\\mathbf {X})) \\right]. \\tag {9}", + "image_path": "dfe4755e3d6dcb6a697538e3e3c53b64cc2c73a6742dab99971b2bcebcfbd150.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "content": "Following the ideas of Paschalidou et al. [53], we first define " + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\Delta (x,s)" + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "content": " as the expected distance between " + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_s(\\mathbf{X})" + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "content": " conditionally to the slot " + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 485, + 287, + 521 + ], + "type": "text", + "content": " being activated:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 526, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 526, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 93, + 526, + 287, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta (x, s) = \\mathbb {E} _ {b _ {s} | a _ {s} = 1} [ d (x, \\mathcal {M} _ {s} (\\mathbf {X})) ] (10) \\\\ = \\frac {1}{\\alpha_ {s}} \\sum_ {k = 1} ^ {K} \\beta_ {s} ^ {k} d \\left(x, \\mathbf {Y} _ {s} ^ {k}\\right). (11) \\\\ \\end{array}", + "image_path": "1cdbf0ad4abd62034baf5a775dcfb858ebcdb48bc7119213bd841ba6c4880173.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "content": "Next, we compute for each point " + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "content": " a permutation " + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "inline_equation", + "content": "\\sigma_x" + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "inline_equation", + "content": "[1, S]" + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "inline_equation", + "content": "\\Delta(x, \\sigma_x(s))" + }, + { + "bbox": [ + 47, + 576, + 287, + 600 + ], + "type": "text", + "content": " is non-decreasing, i.e.:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 605, + 287, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 605, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 93, + 605, + 287, + 619 + ], + "type": "interline_equation", + "content": "\\Delta (x, \\sigma_ {x} (1)) \\leq \\dots \\leq \\Delta (x, \\sigma_ {x} (S)). \\tag {12}", + "image_path": "13ea92f3617c8ce13b178923da95383dd4db2580fc10c733fecd9812f66a9985.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "content": " is the closest activated slot to " + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "content": ", then all the slots closer to " + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "content": " must be deactivated. This observation leads us to rewrite " + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cov}}" + }, + { + "bbox": [ + 47, + 624, + 287, + 660 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 665, + 287, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 287, + 695 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o v}} (\\mathcal {M}, \\mathbf {X}) = \\frac {1}{| \\mathbf {X} |} \\sum_ {x \\in \\mathbf {X}} \\sum_ {s = 1} ^ {S} \\Delta (x, s) \\alpha_ {s} \\prod_ {r < \\sigma_ {x} (s)} \\left(1 - \\alpha_ {\\sigma_ {x} ^ {- 1} (r)}\\right), \\tag {13}", + "image_path": "e91e84a1049c2d5dcf448a2d817bdf727dad6ebcaeec46592f293635821ef0a0.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "inline_equation", + "content": "\\sigma^{-1}" + }, + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "text", + "content": " the inverse permutation of " + }, + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 700, + 203, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "Regularization losses. As is often necessary with fully unsupervised reconstruction methods [43, 48, 49], we define several regularization losses implementing priors on the model output to prevent degenerate local minima:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 545, + 144 + ], + "type": "text", + "content": "- To encourage the slot activation to be sparse, we use the following loss penalizing slot activation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 375, + 152, + 545, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 152, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 375, + 152, + 545, + 179 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a c t}} (\\mathcal {M}) = \\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]. \\tag {14}", + "image_path": "c3c0cc75d67e2ba862fa0b20308e6b6bc5fd3b426431ee180010334e9a64fa44.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 186, + 545, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 545, + 210 + ], + "type": "text", + "content": "- To avoid slots and prototypes that are never used, we use the following losses, which we compute batch-wise:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 327, + 217, + 545, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 217, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 327, + 217, + 545, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s l o t}} (\\mathcal {M}) = - \\sum_ {s = 1} ^ {S} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {s} ]}{\\sum_ {t = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} [ \\alpha_ {t} ]}, \\epsilon_ {S}\\right), \\tag {15}", + "image_path": "6501825b0113a86059ee762ffbe7cf5397a86a08c920fcfc9d03acd052da9493.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 323, + 248, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 248, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 323, + 248, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r o t o}} (\\mathcal {M}) = - \\sum_ {k = 1} ^ {K} \\min \\left(\\frac {\\mathbb {E} _ {\\mathbf {X}} \\left[ \\sum_ {s = 1} ^ {S} \\beta_ {s} ^ {k} \\right]}{\\sum_ {s = 1} ^ {S} \\mathbb {E} _ {\\mathbf {X}} \\left[ \\alpha_ {s} \\right]}, \\epsilon_ {K}\\right), \\tag {16}", + "image_path": "a5b73ac4ab61d6e31a588791493e207180fb5ad9a7cb37f7d778d0335d2a80e0.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "inline_equation", + "content": "\\epsilon_S" + }, + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "inline_equation", + "content": "\\epsilon_K" + }, + { + "bbox": [ + 306, + 287, + 544, + 311 + ], + "type": "text", + "content": " hyperparameters setting the smallest acceptable relative use frequency for a slot or prototype." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 311, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 545, + 334 + ], + "type": "text", + "content": "The full regularization loss is a weighted sum of the three losses described above:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 345, + 342, + 545, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 342, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 345, + 342, + 545, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\lambda_ {\\text {a c t}} \\mathcal {L} _ {\\text {a c t}} + \\lambda_ {\\text {s l o t}} \\mathcal {L} _ {\\text {s l o t}} + \\lambda_ {\\text {p r o t o}} \\mathcal {L} _ {\\text {p r o t o}}, \\tag {17}", + "image_path": "5949f9f72ca4aa101f044016a512d8b6e4ccaa003e7addb3d52cdd4b97d42fca.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "text", + "content": "where we use " + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{act}} = 10^{-4}" + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{slot}} = \\lambda_{\\mathrm{proto}} = 0.1" + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\epsilon_S = \\epsilon_K = 0.1" + }, + { + "bbox": [ + 306, + 362, + 545, + 387 + ], + "type": "text", + "content": " in all our experiments on the Earth Parser Dataset." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 394, + 502, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 502, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 502, + 407 + ], + "type": "text", + "content": "3.3. Training and Implementation Details" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": "Model configuration. We process the input point cloud " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": " using a similar architecture as in [53]. We first voxelize it with a " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "64\\times 64\\times 64" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": " grid and map it to a vector using a sequence of 6 3D sparse convolutions [9] and 6 strided convolutions. The resulting representation is then transformed using one linear layer for each slot. These features are decoded by simple 2-layer MLPs: one generates the distribution parameters " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\alpha_{s}" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\beta_s^k" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": ", and the other ones the parameters of the 3D transformations " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s(X)" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": ". The transformations include an anisotropic scaling, a " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": "-axis tilt of " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\pm \\pi /10" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": ", a rotation around the " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": "-axis, and a translation, in this specific order. We use " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "S = 64" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": " slots and " + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "inline_equation", + "content": "K = 6" + }, + { + "bbox": [ + 305, + 412, + 545, + 567 + ], + "type": "text", + "content": " prototypes as default parameters. See the supplementary material for details." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "text", + "content": "The intensity of the return signal of each point is available in the LiDAR scans. We associate each prototype with a single learnable intensity parameter and perform the Chamfer distance (Equation 6) in 4 dimensions: spatial coordinates normalized to " + }, + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "[0,1]^3" + }, + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "text", + "content": " and intensity to " + }, + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "inline_equation", + "content": "[0,0.1]" + }, + { + "bbox": [ + 305, + 567, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "Curriculum learning. The network predicts simultaneously the slot's probability distributions and their deformations. This results in many concurrent degrees of freedom and can make the training process unstable. Therefore, following Monnier et al. [49] and Loiseau et al. [43], we implement a multi-stage curriculum learning strategy. We first" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27877" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 553, + 222 + ], + "type": "image", + "image_path": "1922eaf39f98874f35a5cbc8f06007a23c6a9c6d444bf126e65343b5b5ddd33b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 547, + 260 + ], + "type": "text", + "content": "Figure 4. Reconstruction Quality. We show two partial scenes with their RGB and intensity values, as well as their reconstruction by our method and competing models. We use the prototypes' intensity to color the points or pixels. As SuperQuadrics does not model the intensity, we use a random colour for each quadric." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 289, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 368 + ], + "type": "text", + "content": "initialize the prototypes as point clouds uniformly sampled from a random cuboid and gradually unfreeze the model parameters in the following order: (i) translation, rotation, tilt, slot activation, and choice of prototype; (ii) intensities of the prototypes, when available; (iii) scales of the prototypes; (iv) positions of the prototypes' 3D points; (v) anisotropic scalings of the prototypes. As shown in Section 4, each step of this curriculum scheme improves the performances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": "Prototypes selection. We automatically select the number of prototypes for a complete scene using a simple greedy algorithm. We measure the increase of reconstruction loss when preventing the model from selecting each prototype individually. We remove the prototype with the lowest increase if it is lower than " + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 380, + 287, + 452 + ], + "type": "text", + "content": ", and iterate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "text", + "content": "Implementation details. Our model is trained separately for each scene by randomly sampling square patches. During training, the patches are subsampled to a maximum " + }, + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "text", + "content": " points. Each stage of the curriculum is trained until convergence. We use the ADAM optimizer [28] with a learning rate of " + }, + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 466, + 287, + 551 + ], + "type": "text", + "content": " and default parameters. See the supplementary material for preprocessing, training and evaluation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 560, + 100, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 100, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 100, + 572 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 580, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 653 + ], + "type": "text", + "content": "In this section, we assess the ability of our method to parse complex 3D aerial data. We give in Section 4.1 an overview of our proposed dataset of aerial LiDAR scans. In Section 4.2, we then discuss our evaluation metrics and baselines. Finally, we present a quantitative (Section 4.3) and qualitative (Section 4.4) analysis of our results." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 659, + 170, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 170, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 170, + 671 + ], + "type": "text", + "content": "4.1. Earth Parser Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We introduce a new dataset to train and evaluate parsing methods on large, uncurated aerial LiDAR scans. We use data from the French Mapping Agency associated to the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "text", + "content": "LiDAR-HD project [45]. Each scan is composed of several airborne LiDAR acquisitions taken at different angles, leading to a minimum resolution of 20 points/" + }, + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathrm{m}^2" + }, + { + "bbox": [ + 304, + 270, + 545, + 330 + ], + "type": "text", + "content": ". The points are associated with their laser reflectance (intensity), and colorized based on asynchronous aerial photography." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 331, + 546, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 415 + ], + "type": "text", + "content": "We selected 7 scenes, covering over " + }, + { + "bbox": [ + 304, + 331, + 546, + 415 + ], + "type": "inline_equation", + "content": "7.7\\mathrm{km}^2" + }, + { + "bbox": [ + 304, + 331, + 546, + 415 + ], + "type": "text", + "content": " and a total of 98 million 3D points, with diverse content and complexity, such as dense habitations, forests, or complex industrial facilities. We associate most 3D points with a coarse semantic label, such as ground, building, or vegetation. The characteristics of the scenes are detailed in Table 2 and each is visualized in Figure 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 424, + 486, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 424, + 486, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 424, + 486, + 436 + ], + "type": "text", + "content": "4.2. Evaluation Metrics and Baselines" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 443, + 545, + 479 + ], + "type": "text", + "content": "We quantitatively evaluate the performances for reconstruction and semantic segmentation of our model and several unsupervised scene decomposition approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 556 + ], + "type": "text", + "content": "Evaluation metrics. As our goal is to summarize a point cloud using few prototypes, the quality of the reconstruction is critical. We measure it with the symmetric Chamfer distance (\"Cham.\" in the Tables) between the input and the output point clouds of our model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 617 + ], + "type": "text", + "content": "By associating a class to each prototype's points, we can propagate labels from the reconstruction to the input cloud and perform semantic segmentation. We evaluate the quality of this segmentation with the class-averaged Intersection-over-Union (mIoU) metric." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": "In a practical scenario, an operator can manually annotate the points of the 3D prototypes, allowing for the segmentation of the entirety of " + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 618, + 546, + 713 + ], + "type": "text", + "content": " with minimal effort. To perform automatic evaluation, we follow the standard practice for evaluating clustering and unsupervised segmentation methods [25, 31, 48]: we assign to each prototype's point the most frequent class of its closest point of the input after the reconstruction." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27878" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 112, + 545, + 201 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Results on the Earth Parser Dataset. We report the quality of the reconstruction (Cham.) and semantic segmentation (mIoU) on each of the scenes of our Earth Parser Dataset. While our method does not always provide the most faithful reconstructions, it leads to the most accurate point classification." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 112, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 52, + 112, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 52, + 112, + 545, + 201 + ], + "type": "table", + "html": "
Rec.SemanticCrop FieldsForestGreenhousesMarinaPower PlantUrbanWindturbines
Cham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoUCham.mIoU
k-means (i,z) [44]X✓—93.8—71.5—39.3—41.4—42.8—56.5—87.6
SuperQuadrics [53]3DX0.86—1.04—0.60—0.93—0.58—0.40—13.5—
DTI-Sprites [49]2.5D+i✓6.1083.214.5940.25.3642.06.1641.45.3629.02.9947.336.1925.9
AtlasNet v2 [11]3D+i✓1.0743.11.5871.40.5649.10.7342.10.4541.60.6348.89.4748.1
Ours3D+i✓0.7296.90.8883.70.4091.30.8278.70.4452.20.2983.26.6593.4
", + "image_path": "bd640d8a62a91437b3b7e4af311a2c6cac59a0860c115a72b77a61a33453f4bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 240, + 288, + 378 + ], + "blocks": [ + { + "bbox": [ + 47, + 209, + 288, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 209, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 288, + 232 + ], + "type": "text", + "content": "Table 2. Earth Parser Dataset. Our proposed dataset is composed of 7 diverse scenes acquired by aerial LiDAR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 240, + 288, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 240, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 288, + 378 + ], + "type": "table", + "html": "
NameSurface in km2# points ×106annotation ratio in %num. of classes
Crop Fields1.119.777.42
Forest1.146.797.82
Greenhouses0.11.395.63
Marina0.10.592.72
Power Plant0.28.678.44
Urban1.115.795.93
Wind turbines4.25.699.83
Total7.798.391.6—
", + "image_path": "147271780cc9be1a8be836706a8c18b124a941d29f7e964d6a2d110e5e2f72b6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 426 + ], + "type": "text", + "content": "Baselines. We adapt several unsupervised approaches for scene reconstruction and/or semantic segmentation tasks to provide comparisons for our approach:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 533 + ], + "type": "text", + "content": "- k-means. We cluster the points of the input with the k-means algorithm [44] using as many clusters as we use prototypes. We obtain the best results by using a combination of the point's intensity and elevation as features for clustering. We then assign to each centroid its most frequent class, and propagate this label to the entire cluster, leading to a semantic segmentation. This method does not reconstruct the input, but gives us a simple and surprisingly strong baseline for semantic segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 594 + ], + "type": "text", + "content": "- SuperQuadrics revisited. We use the method of Paschalidou et al. [53] to learn to approximate scenes with an adaptive number of superquadrics [3]. It provides a baseline for reconstruction and a qualitative comparison for instance segmentation, shown in supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "content": "- DTI-Sprites. We use the point cloud to construct a digital elevation model, i.e. a 2.5D image of resolution " + }, + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 594, + 288, + 689 + ], + "type": "text", + "content": " where each pixel has an elevation and intensity value. We adapt the unsupervised image decomposition approach of Monnier et al. [49] to break down this image into a set of 2.5D sprites. We evaluate the reconstruction and segmentation by sampling 25 3D point per pixel, transferring the pixel's label to the points, and interpolating their elevations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "- AtlasNet v2. This extension [11] of AtlasNet [18] uses a fixed number of learnable prototype point clouds to recon" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "type": "table", + "bbox": [ + 307, + 251, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 305, + 209, + 547, + 243 + ], + "lines": [ + { + "bbox": [ + 305, + 209, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 209, + 547, + 243 + ], + "type": "text", + "content": "Table 3. Ablation Study. We evaluate the effect of our prototype selection post-processing, our model's degrees of freedom, and our different regularization losses." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 251, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 307, + 251, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 545, + 415 + ], + "type": "table", + "html": "
UrbanMarina
Cham.mIoUCham.mIoU
expressivityLearnable Earth Parser0.2983.20.8278.7
→w/o post-processing0.2883.70.9678.3
→w/o aniso-scale0.3382.41.0467.2
→w/o prototypes0.3668.31.0742.8
→w/o scales0.5558.91.3340.8
→w/o intensities0.5558.71.0940.8
losses→w/o Lact0.1754.10.8056.9
→w/o Lslot0.2577.80.8143.7
→w/o Lproto0.2857.20.9740.7
", + "image_path": "991f28f9223c412d48668a4e142f8063feb07bd41c2ba30e2b1606f64bf24886.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 434, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 482 + ], + "type": "text", + "content": "struct its input. It can be evaluated for both reconstruction and semantic segmentation in a way similar to ours. We extend it to handle intensity in a manner akin to our approach, which improves its segmentation results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 545, + 518 + ], + "type": "text", + "content": "Similar to our method, we train all baselines except k-means by sampling square patches in each scene. Figure 4 shows the output of the reconstruction methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 525, + 424, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 424, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 424, + 538 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 544, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 580 + ], + "type": "text", + "content": "We compare the performance of our approach with the proposed baselines on the Earth Parser Dataset, as well as two publicly available datasets." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 665 + ], + "type": "text", + "content": "Earth Parser Dataset. We provide quantitative reconstruction and semantic segmentation results in Table 1, and illustrations in Figure 4. Despite being highly constrained, our model yields the best reconstruction in 6 out of 7 scenes. Moreover, we significantly outperform the other methods for semantic segmentation across all scenes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Despite its simplicity, the k-means baseline provides strong semantic segmentation performance, beating the other baselines in 5 of 7 annotated scenes. DTI-Sprites [49] has lower reconstruction and segmentation quality, which" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27879" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 123, + 288, + 243 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "type": "text", + "content": "Table 4. Synthetic Shapes. We train our method on all planes from ShapeNet-Part [58], with random rotations around " + }, + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 70, + 288, + 114 + ], + "type": "text", + "content": "-axis. We show the reconstruction of an input plane and the prototypes learned on the dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 123, + 288, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 123, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 123, + 288, + 243 + ], + "type": "table", + "html": "
AtlasNet v2 [11]SuperQuadrics [53]Ours
Recons.
Protos.---
Cham.1.462.911.34
mIoU34.5-68.6
", + "image_path": "461aef78b71a06e44745f68ff5e9e1a1d0936d1761b282cb4a6b6f0626686597.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 251, + 173, + 337 + ], + "blocks": [ + { + "bbox": [ + 47, + 251, + 173, + 337 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 173, + 337 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 173, + 337 + ], + "type": "table", + "html": "
Cham.mIoU
k-means (i,z) [44]—52.7
Superquadrics [53]0.33—
DTI-Sprites [49]2.2553.6
AtlasNet v2 [11]0.3154.9
Ours0.2982.7
", + "image_path": "9eb4d67c3081644fac27d3682de68df885118fbb50e80eccfb81eab181c3104d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 180, + 252, + 290, + 330 + ], + "blocks": [ + { + "bbox": [ + 180, + 252, + 290, + 330 + ], + "lines": [ + { + "bbox": [ + 180, + 252, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 180, + 252, + 290, + 330 + ], + "type": "image", + "image_path": "c0e64fd1741ce88e0d62b422b28e7456390c76b85e6646b1463fc89c09278a09.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 287, + 368 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 368 + ], + "type": "text", + "content": "Figure 5. Results on DALES [68]. We report quantitative and qualitative results for one tile from DALES." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 513 + ], + "type": "text", + "content": "is expected as it models a 3D point cloud in 2.5D. AtlasNet v2 [11] provides good reconstructions but segmentation fails for scenes such as Crop Fields or Urban due its inability to adapt its prototype usage to the input. On the contrary, SuperQuadrics [53] can adjust the number of superquadric it uses and, to some degree, their shape. However, this method uses a single parametric family for all prototypes and fails to reconstruct complex real-world scenes such as Power Plant. Thanks to its probabilistic slot selection, our method can handle inputs with a varying number of objects using only a small set of learned prototypical shapes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 675 + ], + "type": "text", + "content": "Ablation study. We evaluate the impact of various components of our model and report the results in Table 3. First, we observe that the prototype selection post-processing has limited impact on the quality of the prediction and reconstructions, but allows us to adapt and significantly decrease the number of prototypes used for each scene. Second, we evaluate the impact of reducing the expressivity of our model. We successively remove: (i) the anisotropic scaling of " + }, + { + "bbox": [ + 46, + 533, + 286, + 675 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 533, + 286, + 675 + ], + "type": "text", + "content": ", and the possibility of learning the prototype's (ii) points position, (iii) scale, and (iv) intensity. As expected, each degree if liberty removed decreases the quality of the 3D reconstruction and segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "We study the impact of the different regularization introduced in Section 3.2. The losses related to slot activation have a marginal effect on reconstruction quality but" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "significantly affect the performance of semantic segmentation. This suggests that using slots sparingly but equally is important for prototypes to specialize for specific objects, but is not necessary to the expressivity of the reconstruction model. The prototype activation loss prevents unused and possibly degenerate prototypes and therefore improves both segmentation and reconstruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "content": "We evaluate the impact of reducing the number of prototypes or slots. When we set " + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "content": " instead of 6, the average semantic segmentation mIoU drops by 26.3 points, while reconstruction is minimally affected, showing only a " + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "content": " increase in Chamfer distance. Conversely, setting " + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "inline_equation", + "content": "S = 32" + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "content": " instead of 64 leads to a significant " + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "inline_equation", + "content": "+47.3\\%" + }, + { + "bbox": [ + 304, + 156, + 546, + 299 + ], + "type": "text", + "content": " increase of Chamfer distance, with a modest drop 0.4 of semantic segmentation mIoU. These findings highlight the importance of maintaining sufficient diversity in the prototypes to enable specialization for different object types, whereas the number of slots strongly influences the expressiveness of the reconstruction model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "text", + "content": "ShapeNet. We evaluate our model on 2690 planes from ShapeNet-Part [58], whose points are annotated as wing, engine, tail, or body. We randomly rotate the shapes around the " + }, + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 314, + 546, + 445 + ], + "type": "text", + "content": "-axis during training and evaluation. We report in Table 4 the performances and reconstructions for our approach, AtlasNet v2 [11], and SuperQuadrics revisited [53]. Our method handle rotations better than AtlasNet v2, and manages to successfully locate the tail of the planes. While SuperQuadrics' reconstructions make sense qualitatively, they are worse in terms of accuracy than ours, and do not enable semantic segmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 544 + ], + "type": "text", + "content": "DALES. We also train and evaluate our model on DALES [68], a dataset of aerial LiDAR scans of urban area, with a restricted class set: ground, vegetation, and buildings. Our model significantly outperforms competing reconstruction-based approaches, and yields similar performance than for the Urban tile of the Earth Parser Dataset; see Figure 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 552, + 418, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 418, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 418, + 564 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Instance segmentation. We can perform instance segmentation simply by considering each slot as a different instance. This is particularly interesting for parsing natural woodlands, a key endeavor for forest management [50] and biomass estimation [12]. While this task has a long history of handcrafted approaches [69], current deep learning approaches are mostly limited to artificial or low-density forests [65]. As shown qualitatively in Figure 6, our Learnable Earth Parser can learn without any supervision to separate individual trees in dense forests, or boats in a marina. We evaluate quantitatively the performance of our model in terms of instance segmentation, we counted manually 88" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27880" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 287, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 287, + 255 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 287, + 255 + ], + "type": "image", + "image_path": "edd8816a7c438417c0fb5ddd49665ba2b19559f637662b1f52c3c0dc3af2ee50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 333 + ], + "type": "text", + "content": "Figure 6. Instance Segmentation. We can identify the reconstruction of each slot as a separate instance, allowing us to perform instance segmentation of complex data such as dense natural forests or a marina. For this visualization, we considered the points associated to \"trees\" or \"boat hull\" prototypes and color each of their instance randomly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 342, + 287, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 416 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 416 + ], + "type": "text", + "content": "boat instances in 10 distinct portions of the Marina scene. We then computed the Mean Relative Error (MRE) of the prediction given by the number of boat-like prototypes (prototype #3 in Figure 8) in these zones. Our method yields a MRE of " + }, + { + "bbox": [ + 46, + 342, + 287, + 416 + ], + "type": "inline_equation", + "content": "7,4\\%" + }, + { + "bbox": [ + 46, + 342, + 287, + 416 + ], + "type": "text", + "content": ", 2.6 times lower than DTI-Sprites [49], demonstrating its strong performances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 465 + ], + "type": "text", + "content": "Semantic segmentation. After annotating the prototypes, our model can perform convincing semantic segmentation of complex scenes, as shown in Figure 7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 647 + ], + "type": "text", + "content": "Interpretable prototypes. In Figure 8, we show prototypes learned on our Earth Parser Dataset with colors showing the associated semantic label for each point. These prototypes give at a glance insights on the content of these real-world scenes. Our model is able to learn a wide variety of shapes, such as boats' masts, wind turbines, or greenhouses. We also observe that the prototypes are typically associated with a single object type. Empirically, the average class distribution of the nearest neighbors of each prototypes' point exhibits the same normalized entropy (0.22) as a Bernoulli distribution with probability 0.964. The highly constrained nature of the deformations " + }, + { + "bbox": [ + 46, + 479, + 288, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 479, + 288, + 647 + ], + "type": "text", + "content": " prevents slots from repurposing the same prototypes for different objects, making the learned prototypes easy to identify and manually annotate." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 119, + 670 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We introduced a novel unsupervised method for parsing complex real-world aerial scans into simple parts using a small set of learned prototypical shapes. We demonstrate" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 547, + 247 + ], + "type": "image", + "image_path": "d0f9e92889933ddcee42309ba4271befd917345318a3674e19da68a0fed51349.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 293, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 305, + 258, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 305, + 258, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 258, + 547, + 291 + ], + "type": "text", + "content": "Figure 7. Semantic Segmentation. Our model can perform semantic segmentation of large real-world scene based on annotated prototypes. Black points in the ground truth are unannotated." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 293, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 307, + 293, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 545, + 397 + ], + "type": "image", + "image_path": "7a19f73eddc8fa3dec4ac0e97498ea4647b08baffbed6866f693077d4a985ad0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 406, + 545, + 430 + ], + "type": "text", + "content": "Figure 8. Learned Prototypes. We display three of the prototypes chosen during the selection process for various scenes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 441, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 526 + ], + "type": "text", + "content": "the quality and interpretability of our results on a novel dataset of aerial LiDAR scans. To the best of our knowledge, we are the first to demonstrate the possibility of performing deep unsupervised 3D shape analysis on such a challenging real-world dataset. We believe that our results open new perspectives for computer-assisted environment monitoring and economic intelligence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements This work was supported in part by ANR project READY3D ANR-19-CE23-0007, ANR under the France 2030 program under the reference ANR-23-PEIA-0008, and was granted access to the HPC resources of IDRIS under the allocation 2022-AD011012096R2 made by GENCI. The work of MA was partly supported by the European Research Council (ERC project DISCOVER, number 101076028). The scenes of Earth Parser Dataset were acquired and annotated by the LiDAR-HD project [45]. We thank Zenodo for hosting the dataset. We thank Zeynep Sonat Baltaci, Emile Blettery, Nicolas Dufour, Antoine Guedon, Helen Mair Rawsthorne, Tom Monnier, Damien Robert, Mathis Petrovich and Yannis Siglidis for inspiring discussions and valuable feedback." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27881" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 123 + ], + "type": "text", + "content": "[1] Antonio Alliegro, Davide Boscaini, and Tatiana Tommasi. Joint supervised and self-supervised learning for 3D real world challenges. In ICPR, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 158 + ], + "type": "text", + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, 2016. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 286, + 191 + ], + "type": "text", + "content": "[3] Alan H Barr. Superquadrics and angle-preserving transformations. IEEE Computer graphics and Applications, 1981. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 286, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 286, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 286, + 226 + ], + "type": "text", + "content": "[4] Pál Benkő, Ralph R Martin, and Tamás Várady. Algorithms for reverse engineering boundary representation models. Computer-Aided Design, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 228, + 285, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 285, + 249 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 285, + 249 + ], + "type": "text", + "content": "[5] Thomas Binford. Visual perception by computer. In Proc. IEEE Conf. on Systems and Control, 1975. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 286, + 282 + ], + "type": "text", + "content": "[6] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 285, + 286, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 286, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 286, + 339 + ], + "type": "text", + "content": "[7] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 342, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 286, + 363 + ], + "type": "text", + "content": "[8] YH Chen and CY Liu. Quadric surface extraction using genetic algorithms. Computer-Aided Design, 1999. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 365, + 286, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 365, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 286, + 396 + ], + "type": "text", + "content": "[9] Spconv Contributors. SPConv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 453 + ], + "type": "text", + "content": "[10] Yang Cui, Qingquan Li, Bisheng Yang, Wen Xiao, Chi Chen, and Zhen Dong. Automatic 3-d reconstruction of indoor environment with mobile laser scanning point clouds. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 498 + ], + "type": "text", + "content": "[11] Theo Deprelle, Thibault Groueix, Matthew Fisher, Vladimir Kim, Bryan Russell, and Mathieu Aubry. Learning elementary structures for 3D shape generation and matching. NeurIPS, 2019. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 286, + 554 + ], + "type": "text", + "content": "[12] Antonio Ferraz, Sassan Saatchi, Clément Mallet, Stéphane Jacquemoud, Gil Gonçalves, Carlos Alberto Silva, Paula Soares, Margarida Tomé, and Luisa Pereira. Airborne Li-DAR estimation of aboveground forest biomass in the absence of field inventory. Remote Sensing, 2016. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 556, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 286, + 599 + ], + "type": "text", + "content": "[13] Martin A Fischler and Robert C Bolles. Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 1981. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 633 + ], + "type": "text", + "content": "[14] Ran Gal, Olga Sorkine, Niloy J Mitra, and Daniel Cohen-Or. iWIRES: An analyze-and-edit approach to shape manipulation. In ACM SIGGRAPH. 2009. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 678 + ], + "type": "text", + "content": "[15] Weixiao Gao, Liangliang Nan, Bas Boom, and Hugo Ledoux. SUM: A benchmark dataset of semantic urban meshes. ISPRS Journal of Photogrammetry and Remote Sensing, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[16] Xuming Ge, Bo Wu, Yuan Li, and Han Hu. A multiprimitive-based hierarchical optimal approach for semantic labeling of ALS point clouds. Remote Sensing, 2019. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Paulo FU Gotardo, Olga Regina Pereira Bellon, and Luciano Silva. Range image segmentation by surface extraction using an improved robust estimator. In CVPR, 2003. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 107, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 148 + ], + "type": "text", + "content": "[18] Thibault Groueix, Matthew Fisher, Vladimir G Kim, Bryan C Russell, and Mathieu Aubry. A papier-mâché approach to learning 3D surface generation. In CVPR, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 204 + ], + "type": "text", + "content": "[19] Stéphane Guinard, Loic Landrieu, Laurent Caraffa, and Bruno Vallet. Piecewise-planar approximation of large 3d data as graph-structured optimization. ISPRS Annals of Photogrammetry, Remote Sensing & Spatial Information Sciences, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 258 + ], + "type": "text", + "content": "[20] Florent Guiotte, Minh-Tan Pham, Romain Dambreville, Thomas Corpetti, and Sébastien Lefèvre. Semantic segmentation of LiDAR points clouds: Rasterization beyond digital elevation models. IEEE Geoscience and Remote Sensing Letters, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 293 + ], + "type": "text", + "content": "[21] Bo Guo, Qingquan Li, Xianfeng Huang, and Chisheng Wang. An improved method for power-line reconstruction from point cloud data. Remote sensing, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 326 + ], + "type": "text", + "content": "[22] Abhinav Gupta, Alexei Efros, and Martial Hebert. Blocks world revisited: Image understanding using qualitative geometry and mechanics. In ECCV, 2010. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "text", + "content": "[23] Norbert Haala and Claus Brenner. Extraction of buildings and trees in urban environments. ISPRS journal of photogrammetry and remote sensing, 1999. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 392 + ], + "type": "text", + "content": "[24] Jin Huang, Jantien Stoter, Ravi Peters, and Liangliang Nan. City3D: large-scale building reconstruction from airborne LiDAR point clouds. Remote Sensing, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 425 + ], + "type": "text", + "content": "[25] Xu Ji, Joao F Henriques, and Andrea Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. CVPR, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 470 + ], + "type": "text", + "content": "[26] Adrien Kaiser, Jose Alonso Ybanez Zepeda, and Tamy Boubekeur. A survey of simple geometric primitives detection methods for captured 3D data. In Computer Graphics Forum. Wiley Online Library, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 503 + ], + "type": "text", + "content": "[27] Ekaterina Kalinicheva, Loic Landrieu, Clément Mallet, and Nesrine Chehata. Multi-layer modeling of dense vegetation from aerial lidar scans. In CVPR Workshops, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 505, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 525 + ], + "type": "text", + "content": "[28] Diederik P Kingma and Jimmy Ba. ADAM: A method for stochastic optimization. *ICLR*, 2014. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 568 + ], + "type": "text", + "content": "[29] Florian Kluger, Hanno Ackermann, Eric Brachmann, Michael Ying Yang, and Bodo Rosenhahn. Cuboids revisited: Learning robust 3D shape fitting to single RGB images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 571, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 602 + ], + "type": "text", + "content": "[30] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 634 + ], + "type": "text", + "content": "[31] Adam Kosiorek, Sara Sabour, Yee Whye Teh, and Geoffrey E Hinton. Stacked capsule autoencoders. NeurIPS, 2019. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 637, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 678 + ], + "type": "text", + "content": "[32] Florent Lafarge, Xavier Descombes, Josiane Zerubia, and Marc Pierrot-Deseilligny. Automatic 3D buildup reconstruction from DEMs: An application to PLEIADES simulations. In ISPRS, 2006. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[33] Aleš Leonardis, Alok Gupta, and Ruzena Bajcsy. Segmentation of range images as the search for geometric parametric models. International Journal of Computer Vision, 1995. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27882" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[34] Ales Leonardis, Ales Jaklic, and Franc Solina. Superquadrics for segmenting and modeling range data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1997. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[35] Jun Li, Kai Xu, Siddhartha Chaudhuri, Ersin Yumer, Hao Zhang, and Leonidas Guibas. Grass: Generative recursive autoencoders for shape structures. ACM Transactions on Graphics, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "text", + "content": "[36] Lingxiao Li, Minhyuk Sung, Anastasia Dubrovina, Li Yi, and Leonidas J Guibas. Supervised fitting of geometric primitives to 3D point clouds. In CVPR, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 226 + ], + "type": "text", + "content": "[37] Minglei Li, Peter Wonka, and Liangliang Nan. Manhattanworld urban reconstruction from point clouds. In ECCV, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "text", + "content": "[38] Yangyan Li, Xiaokun Wu, Yiorgos Chrysathou, Andrei Sharf, Daniel Cohen-Or, and Niloy J Mitra. Globfit: Consistently fitting primitives by discovering global relations. In ACM SIGGRAPH. 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "type": "text", + "content": "[39] Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 349 + ], + "type": "text", + "content": "[40] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Robust and accurate superquadric recovery: A probabilistic approach. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 392 + ], + "type": "text", + "content": "[41] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 448 + ], + "type": "text", + "content": "[42] Suresh K Lodha, Edward J Kreps, David P Helmbold, and Darren Fitzpatrick. Aerial LiDAR data classification using support vector machines. In Third International Symposium on 3D Data Processing, Visualization, and Transmission (3DPVT'06). IEEE, 2006. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 480 + ], + "type": "text", + "content": "[43] Romain Loiseau, Tom Monnier, Mathieu Aubry, and Loic Landrieu. Representing shape collections with alignment-aware linear models. In 3DV, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 287, + 514 + ], + "type": "text", + "content": "[44] J MacQueen. Classification and analysis of multivariate observations. In 5th Berkeley Symp. Math. Statist. Probability, 1967. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 536 + ], + "type": "text", + "content": "[45] IGN (French mapping agency). LiDAR-HD: A 3D mapping of France's soil and subsoil, 2021. 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[46] John McCormac, Ankur Handa, Stefan Leutenegger, and Andrew J.Davison. SceneNet RGB-D: Can 5M synthetic images beat generic imagenet pre-training on indoor segmentation? ICCV, 2017. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 624 + ], + "type": "text", + "content": "[47] Jie Mei, Liqiang Zhang, Shihao Wu, Zhen Wang, and Liang Zhang. 3d tree modeling from incomplete point clouds via optimization and 11-mst. International Journal of Geographical Information Science, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 647 + ], + "type": "text", + "content": "[48] Tom Monnier, Thibault Groueix, and Mathieu Aubry. Deep Transformation-Invariant Clustering. NeurIPS, 2020. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[49] Tom Monnier, Elliot Vincent, Jean Ponce, and Mathieu Aubry. Unsupervised layered image decomposition into object prototypes. In ICCV, 2021. 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[50] Felix Morsdorf, Erich Meier, Benjamin Kötz, Klaus I Itten, Matthias Dobbertin, and Britta Allgower. LIDAR-based geometric reconstruction of boreal type forest stands at single" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "tree level for forest and wildland fire management. Remote sensing of environment, 2004. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 117 + ], + "type": "text", + "content": "[51] Liangliang Nan and Peter Wonka. Polyfit: Polygonal surface reconstruction from point clouds. In ICCV, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[52] Joachim Niemeyer, Franz Rottensteiner, and Uwe Soergel. Contextual classification of LiDAR data and building object detection in urban areas. ISPRS journal of photogrammetry and remote sensing, 2014. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[53] Despoina Paschalidou, Ali Osman Ulusoy, and Andreas Geiger. Superquadrics revisited: Learning 3D shape parsing beyond cuboids. In CVPR, 2019. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 194, + 545, + 228 + ], + "type": "text", + "content": "[54] Despoina Paschalidou, Luc Van Gool, and Andreas Geiger. Learning unsupervised hierarchical part decomposition of 3D objects from a single RGB image. In CVPR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[55] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural parts: Learning expressive 3D shape abstractions with invertible neural networks. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 304 + ], + "type": "text", + "content": "[56] Michael Ramamonjisoa, Sinisa Stekovic, and Vincent Lepetit. Monteboxfinder: Detecting and filtering primitives to fit a noisy point cloud. In ECCV. Springer, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "type": "text", + "content": "[57] Lawrence G. Roberts. Machine perception of three-dimensional solids. PhD thesis, Massachusetts Institute of Technology, 1963. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 339, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 339, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 339, + 545, + 371 + ], + "type": "text", + "content": "[58] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR Workshop, 2015. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[59] Manolis Savva, Angel X. Chang, and Pat Hanrahan. Semantically-Enriched 3D Models for Common-sense Knowledge. CVPR 2015 Workshop on Functionality, Physics, Intentionality and Causality, 2015. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "text", + "content": "[60] Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient ransac for point-cloud shape detection. In Computer Graphics Forum, 2007. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 481 + ], + "type": "text", + "content": "[61] Ruwen Schnabel, Patrick Degener, and Reinhard Klein. Completion and reconstruction with primitive shapes. In Computer Graphics Forum. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 514 + ], + "type": "text", + "content": "[62] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In CVPR, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 548 + ], + "type": "text", + "content": "[63] Nina M Singer and Vijayan K Asari. DALES objects: A large scale benchmark dataset for instance segmentation in aerial LiDAR. IEEE Access, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[64] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 645 + ], + "type": "text", + "content": "[65] Chenxin Sun, Chengwei Huang, Huaiqing Zhang, Bangqian Chen, Feng An, Liwen Wang, and Ting Yun. Individual tree crown segmentation and crown width extraction from a heightmap derived from aerial laser scanning data using a deep learning framework. Frontiers in plant science, 2022. 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[66] Minhyuk Sung, Vladimir G Kim, Roland Angst, and Leonidas Guibas. Data-driven structural priors for shape completion. ACM Transactions on Graphics, 2015. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[67] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning shape abstractions by assembling volumetric primitives. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27883" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 588 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[68] Nina Varney, Vijayan K Asari, and Quinn Graehling. DALES: A large-scale aerial LiDAR data set for semantic segmentation. In CVPR Workshop, 2020. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 162 + ], + "type": "text", + "content": "[69] Jari Vauhkonen, Liviu Ene, Sandeep Gupta, Johannes Heinzel, Johan Holmgren, Juho Pitkänen, Svein Solberg, Yunsheng Wang, Holger Weinacker, K Marius Hauglin, et al. Comparative testing of single-tree detection algorithms under different types of forest. Forestry, 2012. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 286, + 206 + ], + "type": "text", + "content": "[70] Ben G Weinstein, Sergio Marconi, Stephanie Bohlman, Alina Zare, and Ethan White. Individual tree-crown detection in RGB imagery using semi-supervised deep learning neural networks. Remote Sensing, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 208, + 286, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 208, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 208, + 286, + 239 + ], + "type": "text", + "content": "[71] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S Chirikjian. Primitive-based shape abstraction via nonparametric bayesian inference. In ECCV, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 241, + 286, + 284 + ], + "type": "text", + "content": "[72] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3D ShapeNets: A deep representation for volumetric shapes. In CVPR, 2015. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 286, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 286, + 330 + ], + "type": "text", + "content": "[73] Shaobo Xia, Dong Chen, Ruisheng Wang, Jonathan Li, and Xinchang Zhang. Geometric primitives in LiDAR point clouds: A review. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 331, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 331, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 286, + 385 + ], + "type": "text", + "content": "[74] Zhen Ye, Yusheng Xu, Rong Huang, Xiaohua Tong, Xin Li, Xiangfeng Liu, Kuifeng Luan, Ludwig Hoegner, and Uwe Stilla. LASDU: A large-scale aerial LiDAR dataset for semantic labeling in dense urban areas. ISPRS International Journal of Geo-Information, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 387, + 286, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 418 + ], + "type": "text", + "content": "[75] Hong-Xing Yu, Leonidas J. Guibas, and Jiajun Wu. Unsupervised Discovery of Object Radiance Fields. *ICLR*, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 286, + 453 + ], + "type": "text", + "content": "[76] Yu Zhang, Peter Tino, Aleks Leonardis, and Ke Tang. A survey on neural network interpretability. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 286, + 487 + ], + "type": "text", + "content": "[77] Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In ICCV, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 489, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 489, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 489, + 286, + 510 + ], + "type": "text", + "content": "[78] Yongheng Zhao, Tolga Birdal, Haowen Deng, and Federico Tombari. 3D point capsule networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 512, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 286, + 555 + ], + "type": "text", + "content": "[79] SM Zolanvari, Susana Ruano, Aakanksha Rana, Alan Cummins, Rogerio Eduardo da Silva, Morteza Rahbar, and Aljosa Smolic. DublinCity: annotated LiDAR point cloud and its applications. BMVC, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 557, + 286, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 557, + 286, + 588 + ], + "spans": [ + { + "bbox": [ + 49, + 557, + 286, + 588 + ], + "type": "text", + "content": "[80] Chuhang Zou, Ersin Yumer, Jimei Yang, Duygu Ceylan, and Derek Hoiem. 3D-PRNN: Generating shape primitives with recurrent neural networks. In ICCV, 2017. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27884" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_content_list.json b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa967d46558d34cee187d8364364c9fbf17a17df --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_content_list.json @@ -0,0 +1,1466 @@ +[ + { + "type": "text", + "text": "Learned Lossless Image Compression based on Bit Plane Slicing", + "text_level": 1, + "bbox": [ + 161, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Zhang1 Huairui Wang1 Zhenzhong Chen1* Shan Liu2 \n1Wuhan University 2Tencent Media Lab \n{zhe_zhang, zzchen}@whu.edu.cn", + "bbox": [ + 232, + 179, + 736, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autoregressive Initial Bits (ArIB), a framework that combines subimage autoregression and latent variable models, has shown its advantages in lossless image compression. However, in current methods, the image splitting makes the information of latent variables being uniformly distributed in each subimage, and causes inadequate use of latent variables in addition to posterior collapse. To tackle these issues, we introduce Bit Plane Slicing (BPS), splitting images in the bit plane dimension with the considerations on different importance for latent variables. Thus, BPS provides a more effective representation by arranging subimages with decreasing importance for latent variables. To solve the problem of the increased number of dimensions caused by BPS, we further propose a dimension-tailored autoregressive model that tailors autoregression methods for each dimension based on their characteristics, efficiently capturing the dependencies in plane, space, and color dimensions. As shown in the extensive experimental results, our method demonstrates the superior compression performance with comparable inference speed, when compared to the state-of-the-art normalizing-flow-based methods. The code is at https://github.com/ZZ022/ArIB-BPS.", + "bbox": [ + 76, + 301, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 666, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lossless image compression is widely applied in many domains where the preservation of high-quality images is demanded such as photography, scientific exploration, and remote sensing. The significance of these compression techniques lies in their ability to reduce image file sizes while preserving the original content, accomplished through the exploitation of inherent image correlations.", + "bbox": [ + 75, + 691, + 468, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fundamentally, Shannon's source coding theorem [30] sets a lower bound for coding length based on the image's entropy, which is determined by its inherent distribution. Consequently, many deep generative models", + "bbox": [ + 75, + 800, + 468, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[4], which have exhibited remarkable efficacy in distribution modeling, have found widespread application within the realm of lossless image compression. These encompass a variety of methodologies, including autoregressive methods (ARMs) [29, 37], variational autoencoders (VAEs) [19, 22, 34, 35], normalizing flows (NFs) [12, 13, 41, 42], and diffusion models [14, 17]. These diverse approaches involve trade-offs between compression performance and inference speed. Furthermore, compression performance encompasses dataset compression, which often approaches the theoretical limits optimized by neural networks, and single-image compression, which is affected by the initial bits required for bits back coding [34]. Bits back coding is utilized in VAEs, continuous NFs, and variational diffusion models. In summary, per pixel ARMs and diffusion models demonstrate impressive compression performance but necessitate impractical inference time [22, 28]. Among the remaining methods, continuous NFs offer the best dataset compression but exhibit weaker performance in single-image compression.", + "bbox": [ + 500, + 268, + 893, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A potential option for achieving a better trade-off is the combined use of subimage autoregression and VAEs, a kind of latent variable model. This approach has shown competitive performance in the field of lossless image compression, particularly in its single-image compression performance when using the Autoregressive Initial Bits (ArIB) framework [28]. ArIB excludes latent variables for some subimages and uses their bits as initial bits, significantly reducing the overhead associated with initial bits for single-image compression.", + "bbox": [ + 496, + 577, + 893, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current methods [28] split the image in the space dimension. Pixels with the same coordinate in different subimages are adjacent in the original image. This splitting uniformly distributes the information of latent variables in each subimage. Thus, each subimage carries similar importance for latent variables. However, latent variables are inadequately used due to their exclusion for some subimages. In addition, there's a growing risk of posterior collapse [21] along the subimage sequence due to the growing volume of the autoregressive prior. Posterior collapse, a phenomenon where non-useful latent variables are learned, adversely im", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author. This work was supported by the National Natural Science Foundation of China (Grant No. 62036005) and Tencent.", + "bbox": [ + 75, + 875, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27579", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pacts compression performance. Given the importance of all subimages for latent variables, current methods suffer from posterior collapse [28].", + "bbox": [ + 75, + 90, + 470, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose the utilization of bit plane slicing (BPS). BPS splits images along the bit plane dimension, with a reduction in the volume of global data modality from the most significant plane (MSP) to the least significant plane (LSP) (see Figure 2). Given that latent variables are utilized for learning global data modalities [22, 28], the importance for latent variables decreases from MSP to LSP (see Figure 3). We arrange subimages with decreasing importance for latent variables. Hence, a more adequate utilization of latent variables is achieved by excluding subimages with minor importance. Moreover, we enhance autoregression by introducing the autoregression from MSP to LSP. This autoregression has a mitigated risk of posterior collapse since subimages important for latent variables are subject to a low risk of posterior collapse.", + "bbox": [ + 75, + 140, + 468, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The introduction of BPS increases the number of dimensions of subimages. When combining autoregression in the plane dimension with that in the existing space and color dimensions, an efficient method is required. To solve the problem, we propose a dimension-tailored autoregressive model. Autoregressive methods are tailored based on the unique characteristics of each dimension: a non-neural distribution model for the simplest color space, a simple neural network for the space dimension, and a much more complex network for the challenging plane dimension. This model demonstrates a substantial improvement in performance, balanced with a suitable increase in complexity, as evidenced by our experimental results.", + "bbox": [ + 75, + 371, + 468, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consequently, our method outperforms the current ArIB methods in compression performance. Moreover, compared to the state-of-the-art NF-based methods, our approach demonstrates superior compression performance with comparable inference time.", + "bbox": [ + 75, + 571, + 468, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions could be summarized as follows:", + "bbox": [ + 94, + 651, + 434, + 665 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce bit plane slicing (BPS) for ArIB, since it splits the subimages with different importance for latent variables. By arranging subimages with decreasing importance for latent variables, BPS enables adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse.", + "- We propose a dimension-tailored autoregressive model. This model tailors autoregression methods for each dimension based on their different characteristics, efficiently modeling dependencies in an increased number of dimensions caused by BPS.", + "- With the merit of the BPS and dimension-tailored autoregressive model, our method surpasses the state-of-the-art normalizing-flow-based methods in compression performance with comparable inference time." + ], + "bbox": [ + 76, + 671, + 468, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive Models compress images by predicting the probability of the current component based on previously decoded ones. Per-pixel ARMs [29, 37] decode images pixel by pixel, yielding impressive compression performance. However, as image resolution increases, the number of required network evaluations grows linearly, leading to impractically long decoding times. To mitigate this issue, subimage ARMs divide an image into a fixed number of subimages, significantly reducing inference time in comparison to per-pixel ARMs. Various image-splitting methods have been proposed, including spatial splitting [9], channelwise splitting [24], unevenly channel-wise splitting [10], lossy and residual decomposition [23], autoregressive subpixel convolution [28], and low-high frequency decomposition [26]. In the neural compression community, autoregression in the bit plane dimension has been adopted for order-agnostic autoregressive diffusion models [14], progressive compression [15], and high bit-depth medical image compression [38]. These methods employ BPS for different goals. Our method leverages autoregression in the bit plane dimension for ArIB to enhance autoregression in a mitigated risk of posterior collapse.", + "bbox": [ + 496, + 114, + 890, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent variable models utilize extracted latent variables from input images to aid in compression. L3C [22] employs a neural network to extract latent variables as auxiliary information. VAE-based methods [19, 34, 35] use posterior sampling for obtaining latent variables to improve performance. LBB [12] defines a posterior distribution for latent variables, enabling discretization for continuous NFs, which is also applied in iFlow [41]. However, posterior sampling requires bits back coding [34], which introduces significant overhead for single-image compression.", + "bbox": [ + 496, + 448, + 890, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Combined methods integrate latent variable models and ARMs. They have witnessed great success in the neural lossy compression community [25] and have been explored in lossless compression recently [16, 28]. This combination enables the efficient capture of both local and global data modalities. SHVC [28] further introduces ArIB based on the combination to reduce the overhead by initial bits.", + "bbox": [ + 496, + 599, + 890, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 719, + 589, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce our Autoregressive Initial Bits with Bit Plane Slicing (ArIB-BPS) method. To provide a clear overview of our method, Figure 1 illustrates the key components and workflow of ArIB-BPS.", + "bbox": [ + 496, + 744, + 890, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Background", + "text_level": 1, + "bbox": [ + 500, + 816, + 630, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subimage Autoregressive Models split the image $x$ into a fixed number of subimages $\\{x_1, x_2, \\ldots, x_n\\}$ . The probability of $x$ is factorized into the product of conditional distributions as follows", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27580", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7d03d7ce93ed414d6a85d9c89617f41b9d69e3b349da280f12e71f1600478ff8.jpg", + "image_caption": [ + "Figure 1. Overview of our method: the left subfigure illustrates the encoding process. Here, $s$ denotes the split index. The depth of the orange indicates the complexity of the networks in the dimension-tailored autoregression. In the encoding process, the insignificant planes $x_{s+1:s}$ are first encoded conditioned on the significant planes $x_{1:s}$ . Then, the posterior distribution $q(z|x_{1:s})$ is obtained from $x_{1:s}$ . The latent variables $\\hat{z}$ are obtained by decoding from the previously encoded bitstream using $q(z|x_{1:s})$ . Next, $x_{1:s}$ is encoded, conditioned on $\\hat{z}$ . Finally, $\\hat{z}$ is encoded with prior distribution $p(z)$ . The decoding order is the reverse of the encoding process." + ], + "image_footnote": [], + "bbox": [ + 148, + 90, + 424, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/188c710655e20d7238d1f936dcfb9019b8583fd5245c6c067194229d733d008e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 436, + 87, + 830, + 305 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (x) = p \\left(x _ {1}\\right) \\prod_ {i = 2} ^ {n} p \\left(x _ {i} \\mid x _ {1: i - 1}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 428, + 468, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent variable models extract latent variables $z$ from the input image $x$ to assist compression. For VAEs and continuous NFs, the approach to obtain $z$ is sampling for a posterior distribution $q(z|x)$ . In this case, bits back coding [34] is required. It encodes the image via the following steps:", + "bbox": [ + 75, + 474, + 468, + 549 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Step 1: Decode $z$ with $q(z|x)$ from initial bits.", + "Step 2: Encode $x$ with $p(x|z)$ .", + "Step 3: Encode $z$ with $p(z)$ ." + ], + "bbox": [ + 94, + 550, + 410, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $p(\\cdot)$ represents the prior distribution. The asymmetric numeral system (ANS) [8], an entropy coder with a reversed order between encoding and decoding, is required.", + "bbox": [ + 75, + 595, + 468, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LBB [12] introduces local bits back, using $q(z|x)$ to discretize continuous NFs for lossless compression. Moreover, VAEs use neural networks to obtain $q(z|x)$ . When sequentially encoding a set of images $\\{x^{(1)}, x^{(2)}, \\ldots, x^{(N)}\\}$ , bits from previously encoded images could be used as initial bits for the current image. For VAEs, the average code length for encoding the image set is", + "bbox": [ + 75, + 642, + 468, + 748 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\bar {C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(- \\log p \\left(x ^ {(i)} \\mid z ^ {(i)}\\right) - \\log p \\left(z ^ {(i)}\\right) \\right. \\tag {2} \\\\ \\left. + \\log q \\left(z ^ {(i)} \\mid x ^ {(i)}\\right)\\right) - \\frac {1}{N} \\log q \\left(z ^ {(1)} \\mid x ^ {(1)}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 771, + 468, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first term, which is the Evidence Lower Bound (ELBO), can be optimized by neural networks. The second term can be amortized when compressing a dataset.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, when compressing a single image, the overhead $-\\log q(z|x)$ is not negligible, which limits the usage of these methods for single-image compression. Nonetheless, single-image compression is required for image compression applications, and it enables parallel compression when compressing a dataset.", + "bbox": [ + 496, + 416, + 890, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When combined with subimage autoregression, this overhead could be reduced via autoregressive initial bits (ArIB) [28]. ArIB partitions the subimage sequence into two subsets: $x_{1:s}$ and $x_{s+1:n}$ . Latent variables are employed exclusively for $x_{1:s}$ , while the bits used to compress $x_{s+1:n}$ serve as initial bits. If the number of bits provided is larger than the number of bits required, the overhead for bits back coding is eliminated.", + "bbox": [ + 496, + 507, + 892, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Bit Plane Slicing for Autoregressive Initial Bits", + "text_level": 1, + "bbox": [ + 498, + 635, + 890, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The current method, SHVC-ArIB [28], splits the image in the space dimension. Its splitting ensures pixels with the same coordinates are adjacent in the original image. Thus, each subimage is equally important for latent variables. This splitting causes inadequate use of latent variables, as latent variables are excluded for some subimages that are important for latent variables. For SHVC-ArIB, this incurs a performance cost compared to utilizing latent variables for all subimages. Moreover, the performance of SHVC-ArIB is constrained by posterior collapse [21]. Posterior collapse refers to the collapse of the posterior to the prior, indicating that non-useful information is extracted by latent variables. A common cause of posterior collapse is highly capable decoders [21]. Specifically, a strong autoregressive prior enhances the decoder's capability, increasing the likelihood of posterior collapse. Consequently, subimages positioned at", + "bbox": [ + 496, + 657, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/54b2e46c782b8aeb60659c71d05e33ff7addb2d37bbdfc740b3b26f31ac3e712.jpg", + "image_caption": [ + "Figure 2. Visualization for each bit plane. The planes are arranged from MSP to LSP, from left-top to right-bottom. The volume of global data modality decreases from MSP to LSP." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 439, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the end of the autoregressive sequence are in a high risk of posterior collapse due to their strong autoregressive priors. In SHVC-ArIB, enhancing spatial autoregression leads to slight improvements or even a decline in performance because of posterior collapse. In this paper, we introduce bit plane slicing (BPS) to address these issues.", + "bbox": [ + 75, + 297, + 468, + 387 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bit Plane Slicing is a well-established technique in digital image processing. For a $d$ -bit image, it partitions the entire image into $d$ bit planes. Here, each pixel intensity is represented as $x$ , and the $l^{th}$ bit is represented as $x^l$ . The value of $x^l$ can be determined using the following equation:", + "bbox": [ + 75, + 388, + 468, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx ^ {l} = \\left\\lfloor x / 2 ^ {d - l} \\right\\rfloor \\mod 2 \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 478, + 468, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this equation, $\\lfloor x\\rfloor$ denotes the largest integer less than or equal to $x$ . The original pixel value $x$ can be reconstructed using the equation:", + "bbox": [ + 75, + 505, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx = \\sum_ {l = 1} ^ {d} 2 ^ {d - l} \\times x ^ {l} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 564, + 468, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we focus specifically on 8-bit images. We define the significant planes as those with smaller $l$ values, and the insignificant planes as those with larger $l$ values. The volume of global data modality decreases from the most significant plane (MSP) to the least significant plane (LSP), as illustrated in Figure 2. Considering latent variables are employed to learn global data modalities [22, 28], we hypothesize that the importance for latent variables similarly decreases from MSP to LSP. To test this hypothesis, we employ the latent variable model from [34] for each bit plane. Subsequently, we evaluate the Bit Per Dimension (BPD) savings and calculate the bit rate percentage attributed to latent variables within the entire subimage. The results, illustrated in Figure 3, display a consistent descending trend from MSP to LSP, with values becoming negligible in the final few insignificant planes. This observation corroborates our initial hypothesis.", + "bbox": [ + 75, + 613, + 468, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this finding, we employ BPS for ArIB. We partition the image into two segments using a slice index rang-", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing in $\\{1, \\ldots, 8\\}$ , denoted as $s$ . The first segment comprises the significant planes, represented as $x^{1:s}$ , while the second segment includes the insignificant planes, denoted as $x^{s+1:8}$ . Latent variables $z$ are exclusively used for the significant planes, while ARMs are employed for both segments. This approach enables a more adequate use of latent variables since $x^{1:s}$ is more important for them. Moreover, we introduce plane-by-plane autoregression from MSP to LSP to enhance autoregression. As the volume of autoregressive prior increases along the subimage sequence, the risk of posterior collapse intensifies sequentially. In this autoregression, subimages that are more important for latent variables are placed in positions where they are less likely to encounter posterior collapse. In comparison, SHVC-ArIB split the image into subimages with similar importance, meaning that subimages in positions where there's a high risk of posterior collapse are also important for latent variables. Thus, the overall risk of posterior collapse is mitigated by our method.", + "bbox": [ + 496, + 90, + 890, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall bit rate of the image $x$ can be expressed as follows, where $R_{sig}$ and $R_{ins}$ denote the bit rates associated with significant and insignificant planes, respectively:", + "bbox": [ + 496, + 378, + 890, + 424 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R _ {x} = R _ {s i g} + R _ {i n s} \\\\ = - \\log p \\left(x ^ {1: s} \\mid z\\right) - \\log p \\left(x ^ {s + 1: 8} \\mid x ^ {1: s}\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 436, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further propose some modifications to suit our ArIB with BPS method.", + "bbox": [ + 500, + 481, + 890, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Latent Variables are commonly utilized in VAEs to improve performance. These variables are structured in a hierarchy comprising $L$ layers, denoted as $z = \\{z_{1},\\dots ,z_{L}\\}$ . The prior probability distribution of $z$ is factorized as follows:", + "bbox": [ + 500, + 512, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(z _ {1: L}\\right) = p \\left(z _ {L}\\right) \\prod_ {i = 1} ^ {L - 1} p \\left(z _ {i} \\mid z _ {i + 1: L}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 601, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two methods for obtaining the posterior $q(z|x^{1:s})$ : top-down, which proceeds from layer $L$ to layer 1, and bottom-up, which goes from layer 1 to layer $L$ . The former has been shown to exhibit superior compression performance [32, 35], while the latter can leverage the bit-swap technique [19] to reduce the number of initial bits required. Given that the bit rate increases from MSP to LSP [43], insignificant planes can contribute a larger number of initial bits. Consequently, we opt for the top-down inference models, a choice that differentiates our approach from SHVC-ArIB. In this context, the posterior distribution can be expressed as follows:", + "bbox": [ + 496, + 650, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {1: L} \\mid x ^ {1: s}\\right) = q \\left(z _ {L} \\mid x ^ {1: s}\\right) \\prod_ {i = 1} ^ {L - 1} q \\left(z _ {i} \\mid z _ {i + 1: L}, x ^ {1: s}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 856, + 890, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27582", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a9ad64b720dd402783f67b9793b19c96f95a77337ebbb9728f28cf3b5755bb5.jpg", + "image_caption": [ + "Figure 3. BPD savings achieved using latent variables and the bit rate percentage attributed to latent variables for each bit plane. The results are averaged over the CIFAR10, ImageNet32, and ImageNet64 [6] datasets. The importance for latent variables decreases from MSP to LSP." + ], + "image_footnote": [], + "bbox": [ + 101, + 85, + 446, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discretization also plays a pivotal role: a higher precision tends to enhance compression performance but also requires more initial bits. Given the variation in available initial bits across different images, we propose an image-adaptive discretization strategy.", + "bbox": [ + 75, + 455, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our method, we utilize equal mass discretization [35] for the top-down model. This ensures that all latent variables maintain equal probability mass for $p(z)$ . We define the discretization precision $k$ , which discretizes $z$ within $2^{k}$ possible values. It is important to note that $q(z|x)$ closely approximates $p(z)$ , as the Kullback-Leibler divergence between them is minimized during optimization. As a result, each element of $z$ necessitates roughly $-\\log (1 / (2^k)) = k$ initial bits. Furthermore, the size of the latent variables, denoted as $|z|$ , can be quantified. The bit rate $R_{ins}$ can be determined by measuring the length of the encoded bitstream. Following these calculations, the precision $k$ is derived using the equation:", + "bbox": [ + 75, + 532, + 470, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nk = \\left\\lfloor \\frac {R _ {i n s}}{| z |} \\right\\rfloor \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 739, + 468, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For training VAEs, the bit rate of latent variables is calculated with the condition that they are discretized to high precision. In practical scenarios, most images exhibit a sufficiently large precision value $k$ , ensuring negligible impact on compression performance when employing discretization. Nevertheless, low precision is required in minor cases, where the discretization hurts the compression performance. To tackle this, we propose a discretized", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sampling technique for sampling latent variables for finetuning. More specifically, both the prior and posterior distributions are modeled using a univariate logistic distribution, characterized by mean parameters $\\mu_p$ and $\\mu_q$ , respectively. We sample $\\epsilon$ from a standard logistic distribution. The calculation of a sampled latent variable $\\hat{z}$ and its corresponding bit rate can be represented as follows:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} t _ {l} = \\left\\lceil 2 ^ {k} \\cdot \\operatorname {s i g m o i d} \\left(\\epsilon + \\mu_ {q} - \\mu_ {p}\\right)\\right\\rceil - 1, \\\\ \\hat {z} = \\operatorname {l o g i t} \\left(\\frac {t _ {l} + 0 . 5}{2 ^ {k}}\\right) + \\mu_ {p}, \\\\ R _ {\\hat {z}} = \\log (\\operatorname {s i g m o i d} (\\operatorname {l o g i t} \\left(\\frac {t _ {l} + 1}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q}) \\tag {9} \\\\ - \\operatorname {s i g m o i d} (\\log \\mathrm {i t} \\left(\\frac {t _ {l}}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q})) + k, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 219, + 890, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lceil x\\rceil$ signifies the smallest integer greater than or equal to $x$ . During back-propagation, the Straight Through Estimator [3] is utilized to manage this operation.", + "bbox": [ + 496, + 354, + 890, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dimension-Tailored Autoregressive Model", + "text_level": 1, + "bbox": [ + 500, + 409, + 857, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BPS results in an increased number of dimensions for autoregression, specifically, the bit plane dimension. In addition to the plane dimension, we consider autoregression in space and color dimensions, which have demonstrated their efficacy in previous studies [26, 40]. Using a uniform approach to model all dimensions is employed in previous works [14]. However, this method has certain limitations. First, these dimensions present uneven modeling challenges. Employing a one-size-fits-all network to address all dimensions directly may either compromise modeling accuracy in the more complex dimensions or lead to inefficient use of computational resources in simpler dimensions. Second, strong autoregression in the space and color dimensions can potentially lead to posterior collapse. To address these issues, we propose a dimension-tailored autoregressive model that tailors the approach for each dimension based on its specific characteristics. An overview is shown in the right subfigures of Figure 1.", + "bbox": [ + 496, + 431, + 890, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our autoregressive model operates in three stages. In the first stage, we focus on the plane dimension, which presents a significant challenge due to the extensive range of possible values that the pixels in its preceding planes, denoted as $x^{1:i - 1}$ , can take on, amounting to $2^{i - 1}$ possibilities. To effectively leverage the autoregressive prior from $x^{1:i - 1}$ and apply it to $x^i$ , we employ a sophisticated plane context model, $\\Phi_p$ . To capture the global data modality embedded in the preceding planes, we adopt a U-Net architecture [27]. Specifically, we use two models that employ the architecture of diffusion models [7]: one for the significant planes and the other for the insignificant planes. Time embedding is used to differentiate planes.", + "bbox": [ + 496, + 704, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second stage, we address the space dimension. Compared to the plane dimension, this dimension is less complex, as each bit in the current bit plane can only take one of two possible values. We adopt a 4-stage checkerboard model [9, 26], which involves dividing the bit plane $x^{i}$ into four sub-planes, denoted as $x_{1}^{i}, x_{2}^{i}, x_{3}^{i}, x_{4}^{i}$ , as illustrated in Figure 1. For each sub-plane $x_{j}^{i}$ , we utilize a simple network, $\\Phi_{s}$ , to extract the spatial autoregressive prior $F_{s}$ from $x_{1:j-1}^{i}$ . Then, a parameter network $\\Theta$ is used to fuse information from both the plane and space dimensions. To effectively integrate information, we employ a relatively complex network for $\\Theta$ . Although the complexity of $\\Theta$ does contribute to the overall computational demand of our model, the complexity remains significantly lower than that of employing the architecture of $\\Phi_{p}$ for both dimensions, resulting in a more efficient modeling process. More specifically, $\\Phi_{s}$ is a single ResBlock [11], and $\\Theta$ combines an ECANet [39] with several wider ResBlocks.", + "bbox": [ + 76, + 90, + 472, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The third stage is for the color dimension. As identified in PixelCNN++ [29], the correlations among the color channels of a pixel are typically straightforward, and deep networks are often unnecessary for their modeling. Therefore, we adopt a non-network approach for the color dimension. Specifically, we model the distribution of bits in each color channel using a Bernoulli distribution, where the probability parameter $p = P(x = 1)$ is determined autoregressively based on the bits with the same spatial position in previously decoded channels. In our method, $\\Theta$ predict six parameters $\\theta = \\{f_r, f_g, f_b, \\alpha, \\beta, \\gamma\\}$ , and model the probability for each channel as follows:", + "bbox": [ + 76, + 362, + 472, + 542 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {r} = \\operatorname {s i g m o i d} \\left(f _ {r}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 547, + 276, + 564 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {g} = \\operatorname {s i g m o i d} \\left(f _ {g} + \\alpha \\cdot x _ {r}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 566, + 468, + 583 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {b} = \\operatorname {s i g m o i d} \\left(f _ {b} + \\beta \\cdot x _ {r} + \\gamma \\cdot x _ {g}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 585, + 392, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The 1st and 2nd stages of our dimension-tailored autoregressive model can be expressed as follows:", + "bbox": [ + 76, + 607, + 468, + 637 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {p ^ {i}} = \\Phi_ {p} (x ^ {1: i - 1})\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 642, + 282, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {s _ {j} ^ {i}} = \\Phi_ {s} \\left(x _ {1: j - 1} ^ {i}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 664, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}, F _ {z}), & i \\leq s \\\\ \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}), & i > s \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 686, + 383, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, we initially derive $F_{p}$ for the entire bit plane. Subsequently, for each of the four subimages in the bit plane, we compute $F_{s}$ . These are then concatenated with $F_{p}$ and $F_{z}$ (only for significant planes) and fed into $\\Theta$ to determine the six parameters $\\theta$ . Finally, we calculate the probability using non-network color autoregression as outlined in Equation 10.", + "bbox": [ + 76, + 734, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss function of our method is expressed as:", + "bbox": [ + 76, + 840, + 377, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} l = R _ {z} + R _ {x} \\tag {12} \\\\ = - \\log p (z) + \\log q (z | x ^ {1: s}) + R _ {x} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 866, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Architecture and Training Details", + "text_level": 1, + "bbox": [ + 500, + 114, + 795, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In accordance with prior research, our experimental setup entails training three models using three benchmark datasets: CIFAR10, ImageNet32, and ImageNet64 [6]. For all models, we set the split index, $s$ , to 4, based on Figure 3, and utilize 3 layers of hierarchical latent variables. For latent variables, we apply a factor 2 downsampling for each hierarchy of latent variables in latent encoders. The channel number for latent variables is set to 3 across all models. For $\\Phi_p$ , the number of hierarchical levels in the U-Net is set to 4 across all models.", + "bbox": [ + 496, + 138, + 890, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We utilize the Adam optimizer [18] with an initial learning rate of 0.0001, and we apply step decay as the learning rate schedule. For fine-tuning with discretized sampling, we employ Equation 8 with the previously trained model to obtain the precision $k$ for each image in the training set. All experiments are conducted on a GeForce GTX 1080 Ti.", + "bbox": [ + 496, + 290, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compression Performance", + "text_level": 1, + "bbox": [ + 500, + 392, + 740, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of our method across four distinct datasets: CIFAR10, ImageNet32, ImageNet64, and DIV2K [1], using bits per dimension (BPD) as our primary metric. The first three datasets are of low resolution, while the last one, DIV2K, is a high-resolution dataset. Following most previous works, we exclude per-pixel autoregressive methods [29, 37] and diffusion models [14, 17] from our evaluation due to their impractical time complexity.", + "bbox": [ + 496, + 416, + 890, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method evaluates performance in two distinct compression settings: dataset compression and single-image compression. In dataset compression, we sequentially compress 5,000 images with a fixed discretization precision of 10. For single-image compression, each image's $k$ value is calculated using Equation 8 and clipped between 3 and 10. If the calculated $k$ is less than 3, additional $(3 - k) \\times |z|$ ( $|z|$ represents the size of $z$ ) bits are added to ensure successful encoding. These possible bits, along with the $k$ value and image bits, are written to a bitstream. The single image compression performance is determined by measuring the bitstream size. Following [28, 41, 42], we evaluate the performance on the DIV2K dataset using the model trained on ImageNet64 by cropping images into $64 \\times 64$ patches.", + "bbox": [ + 496, + 537, + 890, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against advanced approaches from traditional methods, learned methods without bits back coding, and learned methods with bits back coding. We present the detailed results of our experiments in Table 1. Our method demonstrates superior compression performance, particularly in single-image compression. In dataset compression, our method excels in three out of four datasets, achieving the second-best performance on the remaining dataset. In single-image compression, our method outperforms other methods in all datasets. We note", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27584", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cb3a9fab59f2e27cf96619c81feb39be1f99b3b9341136662b78674da68c9a07.jpg", + "table_caption": [ + "Table 1. Compression performance in BPD for four datasets, where lower values indicate better performance. The best performance is highlighted in bold. *: ImageNet32, **: ImageNet64, +: Flickr2k [33], -: OpenImage [20]. 1st Group: traditional methods, 2nd Group: learned methods without bits back coding, 3rd Group: learned methods with bits back coding. Some values of SHVC and SHVC-ArIB are not available, and we list their available lower bound here. △: These values are obtained by adding dataset performance and initial bits." + ], + "table_footnote": [], + "table_body": "
Dataset CompressionSingle-Image Compression
MethodCIFAR10ImageNet32ImageNet64DIV2KCIFAR10ImageNet32ImageNet64DIV2K
PNG [5]5.896.395.744.235.896.395.744.23
FLIF [31]4.194.524.542.914.194.524.542.91
JPEG-XL [2]5.746.395.892.795.746.395.892.79
L3C [22]-4.764.423.09--4.764.423.09-
IDF [13]3.344.183.90-3.344.183.90-
IDF++ [36]3.264.123.81-3.264.123.81-
LC-FDNet [26]3.77+4.40+4.28+2.71+3.77+4.40+4.28+2.71+
Bit-Swap [19]3.824.50--6.536.97--
Hilloc [35]3.56*4.203.90*-----
SHVC [28]3.163.983.682.57**>3.96>4.78>4.48-
SHVC-ArIB [28]>3.16>3.98>3.68->3.16>3.98>3.68-
iVPF [42]3.204.033.752.68**9.20Δ---
LBB [12]3.123.883.70-42.98Δ49.84Δ41.70Δ-
iFlow [41]3.123.883.702.57**37.40Δ38.27Δ38.12Δ-
ArIB-BPS (ours)3.063.913.632.55**3.073.923.642.59**
", + "bbox": [ + 120, + 155, + 854, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dfad600da2958c9936e67386406f6693ec84146be756c0c495aaf4d5c78ae788.jpg", + "table_caption": [ + "Table 2. Comparison with advanced and available flow-based methods. The batchsize is 64 following LBB and iFlow. We train the small model to align with the inference speed of iFlow on ImageNet64." + ], + "table_footnote": [], + "table_body": "
Compression Performance (BPD) ↓Inference Time (ms/sample) ↓
datasetMethodtheoreticaldatasetsingleencodedecode
CIFAR10LBB [12]3.1163.11842.97864.9464.94
iFlow [41]3.1163.11837.39817.3847.56
ArIB-BPS3.0483.0573.07014.9314.93
ImageNet32LBB [12]3.8713.87549.835194.14194.14
iFlow [41]3.8713.87338.27374.84119.30
ArIB-BPS3.9043.9113.91852.9652.96
ImageNet64LBB [12]3.7013.70341.70395.5795.57
iFlow [41]3.7013.70338.12337.4958.08
ArIB-BPS (Small)3.6773.6853.69343.6543.65
ArIB-BPS3.6243.6333.641107.80107.80
", + "bbox": [ + 173, + 462, + 797, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that precise values for the single-image compression performance of SHVC, as well as the compression performance of SHVC-ArIB, are not available. However, the lower bound could be obtained. Moreover, it could be inferred that our method surpasses SHVC-ArIB by an average of more than 0.10 BPD on low-resolution datasets in both compression settings. Additional details are provided in the Appendix.", + "bbox": [ + 75, + 678, + 470, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of our approach, we conducted a detailed comparison between our method and continuous normalizing-flow-based methods, which exhibit state-of-the-art performance in dataset compression. As Table 2 illustrates, ArIB-BPS outperforms iFlow in both single-image and dataset compression performance for CIFAR10 and ImageNet64, with comparable inference time.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For ImageNet32, although ArIB-BPS lags slightly behind in dataset compression performance, it excels in other metrics, particularly in single-image compression. These results imply that our method achieves a better trade-off between compression performance and inference speed.", + "bbox": [ + 496, + 678, + 893, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 498, + 768, + 666, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct three ablation studies on the CIFAR10 dataset to show the efficiency of our dimension-tailored autoregressive model and the effectiveness of discretized sampling. We note that results in Table 3 and 4 are evaluated using models without using discretized sampling for fine-tuning. Therefore, the results of ArIB-BPS are slightly different from those in Table 2.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27585", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34109f5638c6e4f64eec5b20fa8484c6f4dd8b878de149f6c32bf993b67f1b06.jpg", + "table_caption": [ + "Table 3. Ablation for the effectiveness of autoregression on plane, space, and color dimensions. Inference time is identical for encoding and decoding. The batchsize is 64." + ], + "table_footnote": [], + "table_body": "
planespacecolorTheoretical BPD ↓Inference Time (ms/sample) ↓
✓✓✓3.05214.93
✓✓3.45814.92
✓✓3.47010.36
✓✓3.5225.83
", + "bbox": [ + 106, + 142, + 439, + 232 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9e26ebc08a0e06234e02448b6c53ee47df608b76896b6408650c225372958997.jpg", + "table_caption": [ + "Table 4. Compression performance and inference time of various usages of the architecture of $\\Phi_p$ . The batch size is 64. P: plane. S: space. C: color. Sig: significant planes. Ins: insignificant planes. We also train two small models to further validate our design." + ], + "table_footnote": [], + "table_body": "
Theoretical BPD ↓Inference Time (ms/sample) ↓
SigInsTotalSigInsTotal
P (Ours)1.0412.0113.0526.308.6414.93
PS1.0521.9933.04514.4624.9739.42
PSC1.0431.9713.01543.8374.97118.80
PS (Small)1.0552.0703.1259.0811.4420.51
PSC (Small)1.0962.2183.3149.6413.9623.59
", + "bbox": [ + 81, + 316, + 468, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1 Dimension-Tailored Autoregressive Model", + "text_level": 1, + "bbox": [ + 76, + 464, + 423, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, we conduct experiments to assess the impact of autoregression in each dimension. To do this, we train and evaluate three additional models, each with autoregression removed in either the color, space, or plane dimension. In the model that does not incorporate plane autoregression, we employ the color autoregressive model from PixelCNN++ [29] to ensure compatibility. Additionally, we enhance $\\Phi_s$ by employing the architecture of $\\Phi_p$ , accommodating the increased complexity in the spatial dimension, which now has 256 possible values per pixel instead of just 2. In our experiments, this configuration results in improved performance. Table 3 clearly demonstrates that incorporating autoregression in each dimension consistently leads to significant improvements in compression performance. We note that our autoregression for the color dimension incurs only a negligible overhead in inference time.", + "bbox": [ + 75, + 491, + 468, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we conduct an evaluation to assess the efficiency of our dimension-tailored design. Specifically, we train two additional models: one using the architecture of $\\Phi_p$ for both the plane and space dimensions and the other using the architecture of $\\Phi_p$ for all three dimensions. Subsequently, we evaluate the impact of these modifications on significant and insignificant planes separately. As depicted in Table 4, the dimension-tailored autoregressive model achieves superior performance in significant planes with lower complexity. This superiority is attributed to reduced hindrance to the latent variables, facilitated by sim", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ed6e446a7050d348666012dc92dd0c28350ffa96cecb3c79bc4d103d508fdaf.jpg", + "image_caption": [ + "Figure 4. Average BPD savings through fine-tuning using discretized sampling across various discretization precisions." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 875, + 256 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pler spatial autoregression and color-wise autoregression, as reflected in higher bit rates for latent variables. In the case of insignificant planes, although strengthened autoregression can lead to performance gains, it results in significantly longer inference time. To further validate our method's efficiency, we train two small models with reduced inference time. Our method demonstrates better compression performance. Considering the trade-off between inference time and compression performance, our design is more efficient.", + "bbox": [ + 496, + 320, + 890, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2 Discretized Sampling", + "text_level": 1, + "bbox": [ + 500, + 474, + 702, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assess the performance improvement after fine-tuning across various $k$ values, as illustrated in Figure 4. The method improves performance across all precision values, particularly at lower precision levels. While low-precision cases form a minor segment, their optimization boosts performance in edge cases.", + "bbox": [ + 496, + 498, + 890, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 603, + 617, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed ArIB-BPS, a new approach for lossless image compression. BPS enables us to achieve adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse by arranging subimages with decreasing importance for latent variables. In addition, we present a dimension-tailored autoregressive model that takes into account the different characteristics of each dimension. It efficiently captures dependencies in space, color, and plane dimensions. Experimental results demonstrate that our method achieves superior compression performance compared to the state-of-the-art normalizing-flow-based methods with comparable inference time.", + "bbox": [ + 496, + 628, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The limitation of our work lies in the complexity of higher bit-depth. The complexity of our method would increase as the bit-depth increases due to the interdependency among bit planes. To extend our method for high-bit-depth images, efficient strategies such as handling multiple planes in one step should be explored in our future work.", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27586", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6", + "[2] Jyrki Alakuijala, Ruud Van Asseldonk, Sami Boukortt, Martin Bruse, Iulia-Maria Comsa, Moritz Firsching, Thomas Fischbacher, Evgenii Kliuchnikov, Sebastian Gomez, Robert Obryk, et al. Jpeg xl next-generation image compression architecture and coding tools. In Applications of Digital Image Processing XLII, pages 112-124. SPIE, 2019. 7", + "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5", + "[4] Sam Bond-Taylor, Adam Leach, Yang Long, and Chris George Willcocks. Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1", + "[5] Thomas Boutell. PNG (portable network graphics) specification version 1.0. Technical report, 1997. 7", + "[6] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017. 5, 6", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 5", + "[8] Jarek Duda. Asymmetric numeral systems. arXiv preprint arXiv:0902.0271, 2009. 3", + "[9] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 6", + "[10] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 6", + "[12] Jonathan Ho, Evan Lohn, and Pieter Abbeel. Compression with flows via local bits-back coding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 3, 7", + "[13] Emiel Hoogeboom, Jorn Peters, Rianne Van Den Berg, and Max Welling. Integer discrete flows and lossless compression. Advances in Neural Information Processing Systems, 32, 2019. 1, 7", + "[14] Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. In International Conference on Learning Representations, 2022. 1, 2, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Seungmin Jeon, Kwang Pyo Choi, Youngo Park, and Chang-Su Kim. Context-based trit-plane coding for progressive image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14348-14357, 2023. 2", + "[16] Ning Kang, Shanzhao Qiu, Shifeng Zhang, Zhenguo Li, and Shu-Tao Xia. Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739-3748, 2022. 2", + "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in Neural Information Processing Systems, 34:21696-21707, 2021. 1, 6", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Friso Kingma, Pieter Abbeel, and Jonathan Ho. Bit-swap: Recursive bits-back coding for lossless compression with hierarchical latent variables. In International Conference on Machine Learning, pages 3408-3417, 2019. 1, 2, 4, 7", + "[20] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 7", + "[21] James Lucas, George Tucker, Roger B Grosse, and Mohammad Norouzi. Don't blame the elbo! a linear vae perspective on posterior collapse. Advances in Neural Information Processing Systems, 32, 2019. 1, 3", + "[22] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Practical full resolution learned lossless image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10629-10638, 2019. 1, 2, 4, 7", + "[23] Fabian Mentzer, Luc Van Gool, and Michael Tschannen. Learning better lossless compression using lossy compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6638-6647, 2020. 2", + "[24] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing, pages 3339-3343. IEEE, 2020. 2", + "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in Neural Information Processing Systems, 31, 2018. 2", + "[26] Hochang Rhee, Yeong Il Jang, Seyun Kim, and Nam Ik Cho. Lc-fdnet: Learned lossless image compression with frequency decomposition network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6033-6042, 2022. 2, 5, 6, 7", + "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 5", + "[28] Tom Ryder, Chen Zhang, Ning Kang, and Shifeng Zhang. Split hierarchical variational compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 386-395, 2022. 1, 2, 3, 4, 6, 7", + "[29] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P. Kingma. PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations, 2017. 1, 2, 6, 8", + "[30] Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379-423, 1948. 1", + "[31] Jon Sneyers and Pieter Wuille. Flif: Free lossless image format based on maniac compression. In IEEE International Conference on Image Processing, pages 66-70. IEEE, 2016.", + "[32] Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. Ladder variational autoencoders. Advances in Neural Information Processing Systems, 29, 2016. 4", + "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 114-125, 2017. 7", + "[34] James Townsend, Thomas Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. In International Conference on Learning Representations, 2018. 1, 2, 3, 4", + "[35] James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: lossless image compression with hierarchical latent variable models. In International Conference on Learning Representations, 2019. 1, 2, 4, 5, 7", + "[36] Rianne van den Berg, Alexey A Gritsenko, Mostafa Dehghani, Casper Kaae Sønderby, and Tim Salimans. Idf++: Analyzing and improving integer discrete flows for lossless compression. In International Conference on Learning Representations, 2020. 7", + "[37] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, 2016. 1, 2, 6", + "[38] Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu. Learning lossless compression for high bit-depth medical imaging. In 2023 IEEE International Conference on Multimedia and Expo, pages 2549-2554. IEEE, 2023. 2", + "[39] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 6", + "[40] Honglei Zhang, Francesco Cricri, Hamed R Tavakoli, Nannan Zou, Emre Aksu, and Miska M Hannuksela. Lossless" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image compression using a multi-scale progressive statistical model. In Proceedings of the Asian Conference on Computer Vision, 2020. 5", + "[41] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 1, 2, 6, 7", + "[42] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 1, 6, 7", + "[43] Zhizheng Zhang, Zhibo Chen, Jianxin Lin, and Weiping Li. Learned scalable image compression with bidirectional context disentanglement network. In IEEE International Conference on Multimedia and Expo, pages 1438-1443. IEEE, 2019. 4" + ], + "bbox": [ + 501, + 92, + 890, + 343 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27588", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_model.json b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3c5346ab9bdc0d4de8da09fd46453c5f615a5c48 --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Learned Lossless Image Compression based on Bit Plane Slicing" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.18, + 0.737, + 0.234 + ], + "angle": 0, + "content": "Zhe Zhang1 Huairui Wang1 Zhenzhong Chen1* Shan Liu2 \n1Wuhan University 2Tencent Media Lab \n{zhe_zhang, zzchen}@whu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Autoregressive Initial Bits (ArIB), a framework that combines subimage autoregression and latent variable models, has shown its advantages in lossless image compression. However, in current methods, the image splitting makes the information of latent variables being uniformly distributed in each subimage, and causes inadequate use of latent variables in addition to posterior collapse. To tackle these issues, we introduce Bit Plane Slicing (BPS), splitting images in the bit plane dimension with the considerations on different importance for latent variables. Thus, BPS provides a more effective representation by arranging subimages with decreasing importance for latent variables. To solve the problem of the increased number of dimensions caused by BPS, we further propose a dimension-tailored autoregressive model that tailors autoregression methods for each dimension based on their characteristics, efficiently capturing the dependencies in plane, space, and color dimensions. As shown in the extensive experimental results, our method demonstrates the superior compression performance with comparable inference speed, when compared to the state-of-the-art normalizing-flow-based methods. The code is at https://github.com/ZZ022/ArIB-BPS." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.667, + 0.21, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.8 + ], + "angle": 0, + "content": "Lossless image compression is widely applied in many domains where the preservation of high-quality images is demanded such as photography, scientific exploration, and remote sensing. The significance of these compression techniques lies in their ability to reduce image file sizes while preserving the original content, accomplished through the exploitation of inherent image correlations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.47, + 0.862 + ], + "angle": 0, + "content": "Fundamentally, Shannon's source coding theorem [30] sets a lower bound for coding length based on the image's entropy, which is determined by its inherent distribution. Consequently, many deep generative models" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.895, + 0.573 + ], + "angle": 0, + "content": "[4], which have exhibited remarkable efficacy in distribution modeling, have found widespread application within the realm of lossless image compression. These encompass a variety of methodologies, including autoregressive methods (ARMs) [29, 37], variational autoencoders (VAEs) [19, 22, 34, 35], normalizing flows (NFs) [12, 13, 41, 42], and diffusion models [14, 17]. These diverse approaches involve trade-offs between compression performance and inference speed. Furthermore, compression performance encompasses dataset compression, which often approaches the theoretical limits optimized by neural networks, and single-image compression, which is affected by the initial bits required for bits back coding [34]. Bits back coding is utilized in VAEs, continuous NFs, and variational diffusion models. In summary, per pixel ARMs and diffusion models demonstrate impressive compression performance but necessitate impractical inference time [22, 28]. Among the remaining methods, continuous NFs offer the best dataset compression but exhibit weaker performance in single-image compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.732 + ], + "angle": 0, + "content": "A potential option for achieving a better trade-off is the combined use of subimage autoregression and VAEs, a kind of latent variable model. This approach has shown competitive performance in the field of lossless image compression, particularly in its single-image compression performance when using the Autoregressive Initial Bits (ArIB) framework [28]. ArIB excludes latent variables for some subimages and uses their bits as initial bits, significantly reducing the overhead associated with initial bits for single-image compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Current methods [28] split the image in the space dimension. Pixels with the same coordinate in different subimages are adjacent in the original image. This splitting uniformly distributes the information of latent variables in each subimage. Thus, each subimage carries similar importance for latent variables. However, latent variables are inadequately used due to their exclusion for some subimages. In addition, there's a growing risk of posterior collapse [21] along the subimage sequence due to the growing volume of the autoregressive prior. Posterior collapse, a phenomenon where non-useful latent variables are learned, adversely im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.47, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author. This work was supported by the National Natural Science Foundation of China (Grant No. 62036005) and Tencent." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "pacts compression performance. Given the importance of all subimages for latent variables, current methods suffer from posterior collapse [28]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.141, + 0.47, + 0.368 + ], + "angle": 0, + "content": "To address these issues, we propose the utilization of bit plane slicing (BPS). BPS splits images along the bit plane dimension, with a reduction in the volume of global data modality from the most significant plane (MSP) to the least significant plane (LSP) (see Figure 2). Given that latent variables are utilized for learning global data modalities [22, 28], the importance for latent variables decreases from MSP to LSP (see Figure 3). We arrange subimages with decreasing importance for latent variables. Hence, a more adequate utilization of latent variables is achieved by excluding subimages with minor importance. Moreover, we enhance autoregression by introducing the autoregression from MSP to LSP. This autoregression has a mitigated risk of posterior collapse since subimages important for latent variables are subject to a low risk of posterior collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.568 + ], + "angle": 0, + "content": "The introduction of BPS increases the number of dimensions of subimages. When combining autoregression in the plane dimension with that in the existing space and color dimensions, an efficient method is required. To solve the problem, we propose a dimension-tailored autoregressive model. Autoregressive methods are tailored based on the unique characteristics of each dimension: a non-neural distribution model for the simplest color space, a simple neural network for the space dimension, and a much more complex network for the challenging plane dimension. This model demonstrates a substantial improvement in performance, balanced with a suitable increase in complexity, as evidenced by our experimental results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.469, + 0.648 + ], + "angle": 0, + "content": "Consequently, our method outperforms the current ArIB methods in compression performance. Moreover, compared to the state-of-the-art NF-based methods, our approach demonstrates superior compression performance with comparable inference time." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.652, + 0.435, + 0.666 + ], + "angle": 0, + "content": "Our contributions could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.672, + 0.469, + 0.762 + ], + "angle": 0, + "content": "- We introduce bit plane slicing (BPS) for ArIB, since it splits the subimages with different importance for latent variables. By arranging subimages with decreasing importance for latent variables, BPS enables adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.763, + 0.469, + 0.837 + ], + "angle": 0, + "content": "- We propose a dimension-tailored autoregressive model. This model tailors autoregression methods for each dimension based on their different characteristics, efficiently modeling dependencies in an increased number of dimensions caused by BPS." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.838, + 0.469, + 0.898 + ], + "angle": 0, + "content": "- With the merit of the BPS and dimension-tailored autoregressive model, our method surpasses the state-of-the-art normalizing-flow-based methods in compression performance with comparable inference time." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.672, + 0.469, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Autoregressive Models compress images by predicting the probability of the current component based on previously decoded ones. Per-pixel ARMs [29, 37] decode images pixel by pixel, yielding impressive compression performance. However, as image resolution increases, the number of required network evaluations grows linearly, leading to impractically long decoding times. To mitigate this issue, subimage ARMs divide an image into a fixed number of subimages, significantly reducing inference time in comparison to per-pixel ARMs. Various image-splitting methods have been proposed, including spatial splitting [9], channelwise splitting [24], unevenly channel-wise splitting [10], lossy and residual decomposition [23], autoregressive subpixel convolution [28], and low-high frequency decomposition [26]. In the neural compression community, autoregression in the bit plane dimension has been adopted for order-agnostic autoregressive diffusion models [14], progressive compression [15], and high bit-depth medical image compression [38]. These methods employ BPS for different goals. Our method leverages autoregression in the bit plane dimension for ArIB to enhance autoregression in a mitigated risk of posterior collapse." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.449, + 0.892, + 0.599 + ], + "angle": 0, + "content": "Latent variable models utilize extracted latent variables from input images to aid in compression. L3C [22] employs a neural network to extract latent variables as auxiliary information. VAE-based methods [19, 34, 35] use posterior sampling for obtaining latent variables to improve performance. LBB [12] defines a posterior distribution for latent variables, enabling discretization for continuous NFs, which is also applied in iFlow [41]. However, posterior sampling requires bits back coding [34], which introduces significant overhead for single-image compression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Combined methods integrate latent variable models and ARMs. They have witnessed great success in the neural lossy compression community [25] and have been explored in lossless compression recently [16, 28]. This combination enables the efficient capture of both local and global data modalities. SHVC [28] further introduces ArIB based on the combination to reduce the overhead by initial bits." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.591, + 0.735 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.806 + ], + "angle": 0, + "content": "In this section, we introduce our Autoregressive Initial Bits with Bit Plane Slicing (ArIB-BPS) method. To provide a clear overview of our method, Figure 1 illustrates the key components and workflow of ArIB-BPS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.631, + 0.832 + ], + "angle": 0, + "content": "3.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Subimage Autoregressive Models split the image \\( x \\) into a fixed number of subimages \\( \\{x_1, x_2, \\ldots, x_n\\} \\). The probability of \\( x \\) is factorized into the product of conditional distributions as follows" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27580" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.091, + 0.426, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.088, + 0.831, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.895, + 0.392 + ], + "angle": 0, + "content": "Figure 1. Overview of our method: the left subfigure illustrates the encoding process. Here, \\( s \\) denotes the split index. The depth of the orange indicates the complexity of the networks in the dimension-tailored autoregression. In the encoding process, the insignificant planes \\( x_{s+1:s} \\) are first encoded conditioned on the significant planes \\( x_{1:s} \\). Then, the posterior distribution \\( q(z|x_{1:s}) \\) is obtained from \\( x_{1:s} \\). The latent variables \\( \\hat{z} \\) are obtained by decoding from the previously encoded bitstream using \\( q(z|x_{1:s}) \\). Next, \\( x_{1:s} \\) is encoded, conditioned on \\( \\hat{z} \\). Finally, \\( \\hat{z} \\) is encoded with prior distribution \\( p(z) \\). The decoding order is the reverse of the encoding process." + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.429, + 0.47, + 0.468 + ], + "angle": 0, + "content": "\\[\np (x) = p \\left(x _ {1}\\right) \\prod_ {i = 2} ^ {n} p \\left(x _ {i} \\mid x _ {1: i - 1}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Latent variable models extract latent variables \\( z \\) from the input image \\( x \\) to assist compression. For VAEs and continuous NFs, the approach to obtain \\( z \\) is sampling for a posterior distribution \\( q(z|x) \\). In this case, bits back coding [34] is required. It encodes the image via the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.551, + 0.411, + 0.566 + ], + "angle": 0, + "content": "Step 1: Decode \\(z\\) with \\(q(z|x)\\) from initial bits." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.566, + 0.304, + 0.581 + ], + "angle": 0, + "content": "Step 2: Encode \\( x \\) with \\( p(x|z) \\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.581, + 0.29, + 0.596 + ], + "angle": 0, + "content": "Step 3: Encode \\(z\\) with \\(p(z)\\)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.551, + 0.411, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Here, \\( p(\\cdot) \\) represents the prior distribution. The asymmetric numeral system (ANS) [8], an entropy coder with a reversed order between encoding and decoding, is required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.749 + ], + "angle": 0, + "content": "LBB [12] introduces local bits back, using \\( q(z|x) \\) to discretize continuous NFs for lossless compression. Moreover, VAEs use neural networks to obtain \\( q(z|x) \\). When sequentially encoding a set of images \\( \\{x^{(1)}, x^{(2)}, \\ldots, x^{(N)}\\} \\), bits from previously encoded images could be used as initial bits for the current image. For VAEs, the average code length for encoding the image set is" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.772, + 0.469, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\bar {C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(- \\log p \\left(x ^ {(i)} \\mid z ^ {(i)}\\right) - \\log p \\left(z ^ {(i)}\\right) \\right. \\tag {2} \\\\ \\left. + \\log q \\left(z ^ {(i)} \\mid x ^ {(i)}\\right)\\right) - \\frac {1}{N} \\log q \\left(z ^ {(1)} \\mid x ^ {(1)}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The first term, which is the Evidence Lower Bound (ELBO), can be optimized by neural networks. The second term can be amortized when compressing a dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.507 + ], + "angle": 0, + "content": "However, when compressing a single image, the overhead \\(-\\log q(z|x)\\) is not negligible, which limits the usage of these methods for single-image compression. Nonetheless, single-image compression is required for image compression applications, and it enables parallel compression when compressing a dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.893, + 0.628 + ], + "angle": 0, + "content": "When combined with subimage autoregression, this overhead could be reduced via autoregressive initial bits (ArIB) [28]. ArIB partitions the subimage sequence into two subsets: \\( x_{1:s} \\) and \\( x_{s+1:n} \\). Latent variables are employed exclusively for \\( x_{1:s} \\), while the bits used to compress \\( x_{s+1:n} \\) serve as initial bits. If the number of bits provided is larger than the number of bits required, the overhead for bits back coding is eliminated." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.636, + 0.892, + 0.652 + ], + "angle": 0, + "content": "3.2. Bit Plane Slicing for Autoregressive Initial Bits" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The current method, SHVC-ArIB [28], splits the image in the space dimension. Its splitting ensures pixels with the same coordinates are adjacent in the original image. Thus, each subimage is equally important for latent variables. This splitting causes inadequate use of latent variables, as latent variables are excluded for some subimages that are important for latent variables. For SHVC-ArIB, this incurs a performance cost compared to utilizing latent variables for all subimages. Moreover, the performance of SHVC-ArIB is constrained by posterior collapse [21]. Posterior collapse refers to the collapse of the posterior to the prior, indicating that non-useful information is extracted by latent variables. A common cause of posterior collapse is highly capable decoders [21]. Specifically, a strong autoregressive prior enhances the decoder's capability, increasing the likelihood of posterior collapse. Consequently, subimages positioned at" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.09, + 0.44, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Visualization for each bit plane. The planes are arranged from MSP to LSP, from left-top to right-bottom. The volume of global data modality decreases from MSP to LSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.388 + ], + "angle": 0, + "content": "the end of the autoregressive sequence are in a high risk of posterior collapse due to their strong autoregressive priors. In SHVC-ArIB, enhancing spatial autoregression leads to slight improvements or even a decline in performance because of posterior collapse. In this paper, we introduce bit plane slicing (BPS) to address these issues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.469, + 0.465 + ], + "angle": 0, + "content": "Bit Plane Slicing is a well-established technique in digital image processing. For a \\(d\\)-bit image, it partitions the entire image into \\(d\\) bit planes. Here, each pixel intensity is represented as \\(x\\), and the \\(l^{th}\\) bit is represented as \\(x^l\\). The value of \\(x^l\\) can be determined using the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.479, + 0.47, + 0.499 + ], + "angle": 0, + "content": "\\[\nx ^ {l} = \\left\\lfloor x / 2 ^ {d - l} \\right\\rfloor \\mod 2 \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.552 + ], + "angle": 0, + "content": "In this equation, \\(\\lfloor x\\rfloor\\) denotes the largest integer less than or equal to \\(x\\). The original pixel value \\(x\\) can be reconstructed using the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.565, + 0.47, + 0.606 + ], + "angle": 0, + "content": "\\[\nx = \\sum_ {l = 1} ^ {d} 2 ^ {d - l} \\times x ^ {l} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.869 + ], + "angle": 0, + "content": "In this paper, we focus specifically on 8-bit images. We define the significant planes as those with smaller \\( l \\) values, and the insignificant planes as those with larger \\( l \\) values. The volume of global data modality decreases from the most significant plane (MSP) to the least significant plane (LSP), as illustrated in Figure 2. Considering latent variables are employed to learn global data modalities [22, 28], we hypothesize that the importance for latent variables similarly decreases from MSP to LSP. To test this hypothesis, we employ the latent variable model from [34] for each bit plane. Subsequently, we evaluate the Bit Per Dimension (BPD) savings and calculate the bit rate percentage attributed to latent variables within the entire subimage. The results, illustrated in Figure 3, display a consistent descending trend from MSP to LSP, with values becoming negligible in the final few insignificant planes. This observation corroborates our initial hypothesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Based on this finding, we employ BPS for ArIB. We partition the image into two segments using a slice index rang-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.379 + ], + "angle": 0, + "content": "ing in \\(\\{1, \\ldots, 8\\}\\), denoted as \\(s\\). The first segment comprises the significant planes, represented as \\(x^{1:s}\\), while the second segment includes the insignificant planes, denoted as \\(x^{s+1:8}\\). Latent variables \\(z\\) are exclusively used for the significant planes, while ARMs are employed for both segments. This approach enables a more adequate use of latent variables since \\(x^{1:s}\\) is more important for them. Moreover, we introduce plane-by-plane autoregression from MSP to LSP to enhance autoregression. As the volume of autoregressive prior increases along the subimage sequence, the risk of posterior collapse intensifies sequentially. In this autoregression, subimages that are more important for latent variables are placed in positions where they are less likely to encounter posterior collapse. In comparison, SHVC-ArIB split the image into subimages with similar importance, meaning that subimages in positions where there's a high risk of posterior collapse are also important for latent variables. Thus, the overall risk of posterior collapse is mitigated by our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.425 + ], + "angle": 0, + "content": "The overall bit rate of the image \\( x \\) can be expressed as follows, where \\( R_{sig} \\) and \\( R_{ins} \\) denote the bit rates associated with significant and insignificant planes, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.438, + 0.891, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R _ {x} = R _ {s i g} + R _ {i n s} \\\\ = - \\log p \\left(x ^ {1: s} \\mid z\\right) - \\log p \\left(x ^ {s + 1: 8} \\mid x ^ {1: s}\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.482, + 0.891, + 0.511 + ], + "angle": 0, + "content": "We further propose some modifications to suit our ArIB with BPS method." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.513, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Hierarchical Latent Variables are commonly utilized in VAEs to improve performance. These variables are structured in a hierarchy comprising \\( L \\) layers, denoted as \\( z = \\{z_{1},\\dots ,z_{L}\\} \\). The prior probability distribution of \\( z \\) is factorized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.602, + 0.892, + 0.643 + ], + "angle": 0, + "content": "\\[\np \\left(z _ {1: L}\\right) = p \\left(z _ {L}\\right) \\prod_ {i = 1} ^ {L - 1} p \\left(z _ {i} \\mid z _ {i + 1: L}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.892, + 0.832 + ], + "angle": 0, + "content": "There are two methods for obtaining the posterior \\( q(z|x^{1:s}) \\): top-down, which proceeds from layer \\( L \\) to layer 1, and bottom-up, which goes from layer 1 to layer \\( L \\). The former has been shown to exhibit superior compression performance [32, 35], while the latter can leverage the bit-swap technique [19] to reduce the number of initial bits required. Given that the bit rate increases from MSP to LSP [43], insignificant planes can contribute a larger number of initial bits. Consequently, we opt for the top-down inference models, a choice that differentiates our approach from SHVC-ArIB. In this context, the posterior distribution can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.857, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {1: L} \\mid x ^ {1: s}\\right) = q \\left(z _ {L} \\mid x ^ {1: s}\\right) \\prod_ {i = 1} ^ {L - 1} q \\left(z _ {i} \\mid z _ {i + 1: L}, x ^ {1: s}\\right) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27582" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.087, + 0.447, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.36, + 0.473, + 0.43 + ], + "angle": 0, + "content": "Figure 3. BPD savings achieved using latent variables and the bit rate percentage attributed to latent variables for each bit plane. The results are averaged over the CIFAR10, ImageNet32, and ImageNet64 [6] datasets. The importance for latent variables decreases from MSP to LSP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.457, + 0.47, + 0.534 + ], + "angle": 0, + "content": "Discretization also plays a pivotal role: a higher precision tends to enhance compression performance but also requires more initial bits. Given the variation in available initial bits across different images, we propose an image-adaptive discretization strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.471, + 0.73 + ], + "angle": 0, + "content": "In our method, we utilize equal mass discretization [35] for the top-down model. This ensures that all latent variables maintain equal probability mass for \\( p(z) \\). We define the discretization precision \\( k \\), which discretizes \\( z \\) within \\( 2^{k} \\) possible values. It is important to note that \\( q(z|x) \\) closely approximates \\( p(z) \\), as the Kullback-Leibler divergence between them is minimized during optimization. As a result, each element of \\( z \\) necessitates roughly \\( -\\log (1 / (2^k)) = k \\) initial bits. Furthermore, the size of the latent variables, denoted as \\( |z| \\), can be quantified. The bit rate \\( R_{ins} \\) can be determined by measuring the length of the encoded bitstream. Following these calculations, the precision \\( k \\) is derived using the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.741, + 0.47, + 0.775 + ], + "angle": 0, + "content": "\\[\nk = \\left\\lfloor \\frac {R _ {i n s}}{| z |} \\right\\rfloor \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For training VAEs, the bit rate of latent variables is calculated with the condition that they are discretized to high precision. In practical scenarios, most images exhibit a sufficiently large precision value \\( k \\), ensuring negligible impact on compression performance when employing discretization. Nevertheless, low precision is required in minor cases, where the discretization hurts the compression performance. To tackle this, we propose a discretized" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "sampling technique for sampling latent variables for finetuning. More specifically, both the prior and posterior distributions are modeled using a univariate logistic distribution, characterized by mean parameters \\(\\mu_p\\) and \\(\\mu_q\\), respectively. We sample \\(\\epsilon\\) from a standard logistic distribution. The calculation of a sampled latent variable \\(\\hat{z}\\) and its corresponding bit rate can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.22, + 0.892, + 0.345 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} t _ {l} = \\left\\lceil 2 ^ {k} \\cdot \\operatorname {s i g m o i d} \\left(\\epsilon + \\mu_ {q} - \\mu_ {p}\\right)\\right\\rceil - 1, \\\\ \\hat {z} = \\operatorname {l o g i t} \\left(\\frac {t _ {l} + 0 . 5}{2 ^ {k}}\\right) + \\mu_ {p}, \\\\ R _ {\\hat {z}} = \\log (\\operatorname {s i g m o i d} (\\operatorname {l o g i t} \\left(\\frac {t _ {l} + 1}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q}) \\tag {9} \\\\ - \\operatorname {s i g m o i d} (\\log \\mathrm {i t} \\left(\\frac {t _ {l}}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q})) + k, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.402 + ], + "angle": 0, + "content": "where \\(\\lceil x\\rceil\\) signifies the smallest integer greater than or equal to \\(x\\). During back-propagation, the Straight Through Estimator [3] is utilized to manage this operation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.859, + 0.426 + ], + "angle": 0, + "content": "3.3. Dimension-Tailored Autoregressive Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.704 + ], + "angle": 0, + "content": "BPS results in an increased number of dimensions for autoregression, specifically, the bit plane dimension. In addition to the plane dimension, we consider autoregression in space and color dimensions, which have demonstrated their efficacy in previous studies [26, 40]. Using a uniform approach to model all dimensions is employed in previous works [14]. However, this method has certain limitations. First, these dimensions present uneven modeling challenges. Employing a one-size-fits-all network to address all dimensions directly may either compromise modeling accuracy in the more complex dimensions or lead to inefficient use of computational resources in simpler dimensions. Second, strong autoregression in the space and color dimensions can potentially lead to posterior collapse. To address these issues, we propose a dimension-tailored autoregressive model that tailors the approach for each dimension based on its specific characteristics. An overview is shown in the right subfigures of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our autoregressive model operates in three stages. In the first stage, we focus on the plane dimension, which presents a significant challenge due to the extensive range of possible values that the pixels in its preceding planes, denoted as \\( x^{1:i - 1} \\), can take on, amounting to \\( 2^{i - 1} \\) possibilities. To effectively leverage the autoregressive prior from \\( x^{1:i - 1} \\) and apply it to \\( x^i \\), we employ a sophisticated plane context model, \\( \\Phi_p \\). To capture the global data modality embedded in the preceding planes, we adopt a U-Net architecture [27]. Specifically, we use two models that employ the architecture of diffusion models [7]: one for the significant planes and the other for the insignificant planes. Time embedding is used to differentiate planes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.362 + ], + "angle": 0, + "content": "In the second stage, we address the space dimension. Compared to the plane dimension, this dimension is less complex, as each bit in the current bit plane can only take one of two possible values. We adopt a 4-stage checkerboard model [9, 26], which involves dividing the bit plane \\( x^{i} \\) into four sub-planes, denoted as \\( x_{1}^{i}, x_{2}^{i}, x_{3}^{i}, x_{4}^{i} \\), as illustrated in Figure 1. For each sub-plane \\( x_{j}^{i} \\), we utilize a simple network, \\( \\Phi_{s} \\), to extract the spatial autoregressive prior \\( F_{s} \\) from \\( x_{1:j-1}^{i} \\). Then, a parameter network \\( \\Theta \\) is used to fuse information from both the plane and space dimensions. To effectively integrate information, we employ a relatively complex network for \\( \\Theta \\). Although the complexity of \\( \\Theta \\) does contribute to the overall computational demand of our model, the complexity remains significantly lower than that of employing the architecture of \\( \\Phi_{p} \\) for both dimensions, resulting in a more efficient modeling process. More specifically, \\( \\Phi_{s} \\) is a single ResBlock [11], and \\( \\Theta \\) combines an ECANet [39] with several wider ResBlocks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.363, + 0.473, + 0.543 + ], + "angle": 0, + "content": "The third stage is for the color dimension. As identified in PixelCNN++ [29], the correlations among the color channels of a pixel are typically straightforward, and deep networks are often unnecessary for their modeling. Therefore, we adopt a non-network approach for the color dimension. Specifically, we model the distribution of bits in each color channel using a Bernoulli distribution, where the probability parameter \\( p = P(x = 1) \\) is determined autoregressively based on the bits with the same spatial position in previously decoded channels. In our method, \\( \\Theta \\) predict six parameters \\( \\theta = \\{f_r, f_g, f_b, \\alpha, \\beta, \\gamma\\} \\), and model the probability for each channel as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.548, + 0.277, + 0.565 + ], + "angle": 0, + "content": "\\[\np _ {r} = \\operatorname {s i g m o i d} \\left(f _ {r}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.567, + 0.469, + 0.584 + ], + "angle": 0, + "content": "\\[\np _ {g} = \\operatorname {s i g m o i d} \\left(f _ {g} + \\alpha \\cdot x _ {r}\\right) \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.586, + 0.393, + 0.603 + ], + "angle": 0, + "content": "\\[\np _ {b} = \\operatorname {s i g m o i d} \\left(f _ {b} + \\beta \\cdot x _ {r} + \\gamma \\cdot x _ {g}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.608, + 0.47, + 0.638 + ], + "angle": 0, + "content": "The 1st and 2nd stages of our dimension-tailored autoregressive model can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.643, + 0.284, + 0.663 + ], + "angle": 0, + "content": "\\[\nF _ {p ^ {i}} = \\Phi_ {p} (x ^ {1: i - 1})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.665, + 0.469, + 0.692 + ], + "angle": 0, + "content": "\\[\nF _ {s _ {j} ^ {i}} = \\Phi_ {s} \\left(x _ {1: j - 1} ^ {i}\\right) \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.688, + 0.384, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\theta_ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}, F _ {z}), & i \\leq s \\\\ \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}), & i > s \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.735, + 0.47, + 0.84 + ], + "angle": 0, + "content": "In summary, we initially derive \\( F_{p} \\) for the entire bit plane. Subsequently, for each of the four subimages in the bit plane, we compute \\( F_{s} \\). These are then concatenated with \\( F_{p} \\) and \\( F_{z} \\) (only for significant planes) and fed into \\( \\Theta \\) to determine the six parameters \\( \\theta \\). Finally, we calculate the probability using non-network color autoregression as outlined in Equation 10." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.378, + 0.855 + ], + "angle": 0, + "content": "Loss function of our method is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.867, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} l = R _ {z} + R _ {x} \\tag {12} \\\\ = - \\log p (z) + \\log q (z | x ^ {1: s}) + R _ {x} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.116, + 0.796, + 0.132 + ], + "angle": 0, + "content": "4.1. Architecture and Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.139, + 0.892, + 0.289 + ], + "angle": 0, + "content": "In accordance with prior research, our experimental setup entails training three models using three benchmark datasets: CIFAR10, ImageNet32, and ImageNet64 [6]. For all models, we set the split index, \\(s\\), to 4, based on Figure 3, and utilize 3 layers of hierarchical latent variables. For latent variables, we apply a factor 2 downsampling for each hierarchy of latent variables in latent encoders. The channel number for latent variables is set to 3 across all models. For \\(\\Phi_p\\), the number of hierarchical levels in the U-Net is set to 4 across all models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.381 + ], + "angle": 0, + "content": "We utilize the Adam optimizer [18] with an initial learning rate of 0.0001, and we apply step decay as the learning rate schedule. For fine-tuning with discretized sampling, we employ Equation 8 with the previously trained model to obtain the precision \\( k \\) for each image in the training set. All experiments are conducted on a GeForce GTX 1080 Ti." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.393, + 0.741, + 0.409 + ], + "angle": 0, + "content": "4.2. Compression Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.537 + ], + "angle": 0, + "content": "We evaluate the performance of our method across four distinct datasets: CIFAR10, ImageNet32, ImageNet64, and DIV2K [1], using bits per dimension (BPD) as our primary metric. The first three datasets are of low resolution, while the last one, DIV2K, is a high-resolution dataset. Following most previous works, we exclude per-pixel autoregressive methods [29, 37] and diffusion models [14, 17] from our evaluation due to their impractical time complexity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Our method evaluates performance in two distinct compression settings: dataset compression and single-image compression. In dataset compression, we sequentially compress 5,000 images with a fixed discretization precision of 10. For single-image compression, each image's \\( k \\) value is calculated using Equation 8 and clipped between 3 and 10. If the calculated \\( k \\) is less than 3, additional \\( (3 - k) \\times |z| \\) (\\( |z| \\) represents the size of \\( z \\)) bits are added to ensure successful encoding. These possible bits, along with the \\( k \\) value and image bits, are written to a bitstream. The single image compression performance is determined by measuring the bitstream size. Following [28, 41, 42], we evaluate the performance on the DIV2K dataset using the model trained on ImageNet64 by cropping images into \\( 64 \\times 64 \\) patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare our method against advanced approaches from traditional methods, learned methods without bits back coding, and learned methods with bits back coding. We present the detailed results of our experiments in Table 1. Our method demonstrates superior compression performance, particularly in single-image compression. In dataset compression, our method excels in three out of four datasets, achieving the second-best performance on the remaining dataset. In single-image compression, our method outperforms other methods in all datasets. We note" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27584" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.149 + ], + "angle": 0, + "content": "Table 1. Compression performance in BPD for four datasets, where lower values indicate better performance. The best performance is highlighted in bold. *: ImageNet32, **: ImageNet64, +: Flickr2k [33], -: OpenImage [20]. 1st Group: traditional methods, 2nd Group: learned methods without bits back coding, 3rd Group: learned methods with bits back coding. Some values of SHVC and SHVC-ArIB are not available, and we list their available lower bound here. △: These values are obtained by adding dataset performance and initial bits." + }, + { + "type": "table", + "bbox": [ + 0.121, + 0.156, + 0.855, + 0.413 + ], + "angle": 0, + "content": "
Dataset CompressionSingle-Image Compression
MethodCIFAR10ImageNet32ImageNet64DIV2KCIFAR10ImageNet32ImageNet64DIV2K
PNG [5]5.896.395.744.235.896.395.744.23
FLIF [31]4.194.524.542.914.194.524.542.91
JPEG-XL [2]5.746.395.892.795.746.395.892.79
L3C [22]-4.764.423.09--4.764.423.09-
IDF [13]3.344.183.90-3.344.183.90-
IDF++ [36]3.264.123.81-3.264.123.81-
LC-FDNet [26]3.77+4.40+4.28+2.71+3.77+4.40+4.28+2.71+
Bit-Swap [19]3.824.50--6.536.97--
Hilloc [35]3.56*4.203.90*-----
SHVC [28]3.163.983.682.57**>3.96>4.78>4.48-
SHVC-ArIB [28]>3.16>3.98>3.68->3.16>3.98>3.68-
iVPF [42]3.204.033.752.68**9.20Δ---
LBB [12]3.123.883.70-42.98Δ49.84Δ41.70Δ-
iFlow [41]3.123.883.702.57**37.40Δ38.27Δ38.12Δ-
ArIB-BPS (ours)3.063.913.632.55**3.073.923.642.59**
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.424, + 0.893, + 0.455 + ], + "angle": 0, + "content": "Table 2. Comparison with advanced and available flow-based methods. The batchsize is 64 following LBB and iFlow. We train the small model to align with the inference speed of iFlow on ImageNet64." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.463, + 0.799, + 0.655 + ], + "angle": 0, + "content": "
Compression Performance (BPD) ↓Inference Time (ms/sample) ↓
datasetMethodtheoreticaldatasetsingleencodedecode
CIFAR10LBB [12]3.1163.11842.97864.9464.94
iFlow [41]3.1163.11837.39817.3847.56
ArIB-BPS3.0483.0573.07014.9314.93
ImageNet32LBB [12]3.8713.87549.835194.14194.14
iFlow [41]3.8713.87338.27374.84119.30
ArIB-BPS3.9043.9113.91852.9652.96
ImageNet64LBB [12]3.7013.70341.70395.5795.57
iFlow [41]3.7013.70338.12337.4958.08
ArIB-BPS (Small)3.6773.6853.69343.6543.65
ArIB-BPS3.6243.6333.641107.80107.80
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.472, + 0.786 + ], + "angle": 0, + "content": "that precise values for the single-image compression performance of SHVC, as well as the compression performance of SHVC-ArIB, are not available. However, the lower bound could be obtained. Moreover, it could be inferred that our method surpasses SHVC-ArIB by an average of more than 0.10 BPD on low-resolution datasets in both compression settings. Additional details are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of our approach, we conducted a detailed comparison between our method and continuous normalizing-flow-based methods, which exhibit state-of-the-art performance in dataset compression. As Table 2 illustrates, ArIB-BPS outperforms iFlow in both single-image and dataset compression performance for CIFAR10 and ImageNet64, with comparable inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.895, + 0.756 + ], + "angle": 0, + "content": "For ImageNet32, although ArIB-BPS lags slightly behind in dataset compression performance, it excels in other metrics, particularly in single-image compression. These results imply that our method achieves a better trade-off between compression performance and inference speed." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.77, + 0.667, + 0.785 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We conduct three ablation studies on the CIFAR10 dataset to show the efficiency of our dimension-tailored autoregressive model and the effectiveness of discretized sampling. We note that results in Table 3 and 4 are evaluated using models without using discretized sampling for fine-tuning. Therefore, the results of ArIB-BPS are slightly different from those in Table 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27585" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.132 + ], + "angle": 0, + "content": "Table 3. Ablation for the effectiveness of autoregression on plane, space, and color dimensions. Inference time is identical for encoding and decoding. The batchsize is 64." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.143, + 0.44, + 0.233 + ], + "angle": 0, + "content": "
planespacecolorTheoretical BPD ↓Inference Time (ms/sample) ↓
✓✓✓3.05214.93
✓✓3.45814.92
✓✓3.47010.36
✓✓3.5225.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.249, + 0.472, + 0.306 + ], + "angle": 0, + "content": "Table 4. Compression performance and inference time of various usages of the architecture of \\( \\Phi_p \\). The batch size is 64. P: plane. S: space. C: color. Sig: significant planes. Ins: insignificant planes. We also train two small models to further validate our design." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.317, + 0.47, + 0.436 + ], + "angle": 0, + "content": "
Theoretical BPD ↓Inference Time (ms/sample) ↓
SigInsTotalSigInsTotal
P (Ours)1.0412.0113.0526.308.6414.93
PS1.0521.9933.04514.4624.9739.42
PSC1.0431.9713.01543.8374.97118.80
PS (Small)1.0552.0703.1259.0811.4420.51
PSC (Small)1.0962.2183.3149.6413.9623.59
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.465, + 0.424, + 0.48 + ], + "angle": 0, + "content": "4.3.1 Dimension-Tailored Autoregressive Model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Firstly, we conduct experiments to assess the impact of autoregression in each dimension. To do this, we train and evaluate three additional models, each with autoregression removed in either the color, space, or plane dimension. In the model that does not incorporate plane autoregression, we employ the color autoregressive model from PixelCNN++ [29] to ensure compatibility. Additionally, we enhance \\(\\Phi_s\\) by employing the architecture of \\(\\Phi_p\\), accommodating the increased complexity in the spatial dimension, which now has 256 possible values per pixel instead of just 2. In our experiments, this configuration results in improved performance. Table 3 clearly demonstrates that incorporating autoregression in each dimension consistently leads to significant improvements in compression performance. We note that our autoregression for the color dimension incurs only a negligible overhead in inference time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we conduct an evaluation to assess the efficiency of our dimension-tailored design. Specifically, we train two additional models: one using the architecture of \\(\\Phi_p\\) for both the plane and space dimensions and the other using the architecture of \\(\\Phi_p\\) for all three dimensions. Subsequently, we evaluate the impact of these modifications on significant and insignificant planes separately. As depicted in Table 4, the dimension-tailored autoregressive model achieves superior performance in significant planes with lower complexity. This superiority is attributed to reduced hindrance to the latent variables, facilitated by sim" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.877, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.267, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Figure 4. Average BPD savings through fine-tuning using discretized sampling across various discretization precisions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.892, + 0.458 + ], + "angle": 0, + "content": "pler spatial autoregression and color-wise autoregression, as reflected in higher bit rates for latent variables. In the case of insignificant planes, although strengthened autoregression can lead to performance gains, it results in significantly longer inference time. To further validate our method's efficiency, we train two small models with reduced inference time. Our method demonstrates better compression performance. Considering the trade-off between inference time and compression performance, our design is more efficient." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.476, + 0.704, + 0.492 + ], + "angle": 0, + "content": "4.3.2 Discretized Sampling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.499, + 0.892, + 0.591 + ], + "angle": 0, + "content": "We assess the performance improvement after fine-tuning across various \\( k \\) values, as illustrated in Figure 4. The method improves performance across all precision values, particularly at lower precision levels. While low-precision cases form a minor segment, their optimization boosts performance in edge cases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.619, + 0.619 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We have proposed ArIB-BPS, a new approach for lossless image compression. BPS enables us to achieve adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse by arranging subimages with decreasing importance for latent variables. In addition, we present a dimension-tailored autoregressive model that takes into account the different characteristics of each dimension. It efficiently captures dependencies in space, color, and plane dimensions. Experimental results demonstrate that our method achieves superior compression performance compared to the state-of-the-art normalizing-flow-based methods with comparable inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The limitation of our work lies in the complexity of higher bit-depth. The complexity of our method would increase as the bit-depth increases due to the interdependency among bit planes. To extend our method for high-bit-depth images, efficient strategies such as handling multiple planes in one step should be explored in our future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27586" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.257 + ], + "angle": 0, + "content": "[2] Jyrki Alakuijala, Ruud Van Asseldonk, Sami Boukortt, Martin Bruse, Iulia-Maria Comsa, Moritz Firsching, Thomas Fischbacher, Evgenii Kliuchnikov, Sebastian Gomez, Robert Obryk, et al. Jpeg xl next-generation image compression architecture and coding tools. In Applications of Digital Image Processing XLII, pages 112-124. SPIE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.471, + 0.312 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.471, + 0.384 + ], + "angle": 0, + "content": "[4] Sam Bond-Taylor, Adam Leach, Yang Long, and Chris George Willcocks. Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.385, + 0.471, + 0.413 + ], + "angle": 0, + "content": "[5] Thomas Boutell. PNG (portable network graphics) specification version 1.0. Technical report, 1997. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.471, + 0.457 + ], + "angle": 0, + "content": "[6] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.471, + 0.5 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[8] Jarek Duda. Asymmetric numeral systems. arXiv preprint arXiv:0902.0271, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[9] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[10] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.47, + 0.786 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Evan Lohn, and Pieter Abbeel. Compression with flows via local bits-back coding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Emiel Hoogeboom, Jorn Peters, Rianne Van Den Berg, and Max Welling. Integer discrete flows and lossless compression. Advances in Neural Information Processing Systems, 32, 2019. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. In International Conference on Learning Representations, 2022. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[15] Seungmin Jeon, Kwang Pyo Choi, Youngo Park, and Chang-Su Kim. Context-based trit-plane coding for progressive image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14348-14357, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[16] Ning Kang, Shanzhao Qiu, Shifeng Zhang, Zhenguo Li, and Shu-Tao Xia. Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739-3748, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in Neural Information Processing Systems, 34:21696-21707, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[19] Friso Kingma, Pieter Abbeel, and Jonathan Ho. Bit-swap: Recursive bits-back coding for lossless compression with hierarchical latent variables. In International Conference on Machine Learning, pages 3408-3417, 2019. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[20] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] James Lucas, George Tucker, Roger B Grosse, and Mohammad Norouzi. Don't blame the elbo! a linear vae perspective on posterior collapse. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.616 + ], + "angle": 0, + "content": "[22] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Practical full resolution learned lossless image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10629-10638, 2019. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.685 + ], + "angle": 0, + "content": "[23] Fabian Mentzer, Luc Van Gool, and Michael Tschannen. Learning better lossless compression using lossy compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6638-6647, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[24] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing, pages 3339-3343. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[26] Hochang Rhee, Yeong Il Jang, Seyun Kim, and Nam Ik Cho. Lc-fdnet: Learned lossless image compression with frequency decomposition network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6033-6042, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "tation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.191 + ], + "angle": 0, + "content": "[28] Tom Ryder, Chen Zhang, Ning Kang, and Shifeng Zhang. Split hierarchical variational compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 386-395, 2022. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[29] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P. Kingma. PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations, 2017. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.471, + 0.303 + ], + "angle": 0, + "content": "[30] Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379-423, 1948. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.471, + 0.358 + ], + "angle": 0, + "content": "[31] Jon Sneyers and Pieter Wuille. Flif: Free lossless image format based on maniac compression. In IEEE International Conference on Image Processing, pages 66-70. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[32] Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. Ladder variational autoencoders. Advances in Neural Information Processing Systems, 29, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 114-125, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.491, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[34] James Townsend, Thomas Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. In International Conference on Learning Representations, 2018. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[35] James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: lossless image compression with hierarchical latent variable models. In International Conference on Learning Representations, 2019. 1, 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[36] Rianne van den Berg, Alexey A Gritsenko, Mostafa Dehghani, Casper Kaae Sønderby, and Tim Salimans. Idf++: Analyzing and improving integer discrete flows for lossless compression. In International Conference on Learning Representations, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[37] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[38] Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu. Learning lossless compression for high bit-depth medical imaging. In 2023 IEEE International Conference on Multimedia and Expo, pages 2549-2554. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[39] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[40] Honglei Zhang, Francesco Cricri, Hamed R Tavakoli, Nannan Zou, Emre Aksu, and Miska M Hannuksela. Lossless" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "image compression using a multi-scale progressive statistical model. In Proceedings of the Asian Conference on Computer Vision, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[42] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[43] Zhizheng Zhang, Zhibo Chen, Jianxin Lin, and Weiping Li. Learned scalable image compression with bidirectional context disentanglement network. In IEEE International Conference on Multimedia and Expo, pages 1438-1443. IEEE, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27588" + } + ] +] \ No newline at end of file diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_origin.pdf b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2dc0cb7658608b6352285b1d2953e5e154a7a8b0 --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/493cad4c-eae1-44ac-813f-9e0c42d64707_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea254b806eac2c1b6e22698dbcfb244cecf3a1ba2112db86a8ee1b048144cb50 +size 1718106 diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/full.md b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34806eb221b9ec640d6a252721e834737dbb0239 --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/full.md @@ -0,0 +1,306 @@ +# Learned Lossless Image Compression based on Bit Plane Slicing + +Zhe Zhang1 Huairui Wang1 Zhenzhong Chen1* Shan Liu2 +1Wuhan University 2Tencent Media Lab +{zhe_zhang, zzchen}@whu.edu.cn + +# Abstract + +Autoregressive Initial Bits (ArIB), a framework that combines subimage autoregression and latent variable models, has shown its advantages in lossless image compression. However, in current methods, the image splitting makes the information of latent variables being uniformly distributed in each subimage, and causes inadequate use of latent variables in addition to posterior collapse. To tackle these issues, we introduce Bit Plane Slicing (BPS), splitting images in the bit plane dimension with the considerations on different importance for latent variables. Thus, BPS provides a more effective representation by arranging subimages with decreasing importance for latent variables. To solve the problem of the increased number of dimensions caused by BPS, we further propose a dimension-tailored autoregressive model that tailors autoregression methods for each dimension based on their characteristics, efficiently capturing the dependencies in plane, space, and color dimensions. As shown in the extensive experimental results, our method demonstrates the superior compression performance with comparable inference speed, when compared to the state-of-the-art normalizing-flow-based methods. The code is at https://github.com/ZZ022/ArIB-BPS. + +# 1. Introduction + +Lossless image compression is widely applied in many domains where the preservation of high-quality images is demanded such as photography, scientific exploration, and remote sensing. The significance of these compression techniques lies in their ability to reduce image file sizes while preserving the original content, accomplished through the exploitation of inherent image correlations. + +Fundamentally, Shannon's source coding theorem [30] sets a lower bound for coding length based on the image's entropy, which is determined by its inherent distribution. Consequently, many deep generative models + +[4], which have exhibited remarkable efficacy in distribution modeling, have found widespread application within the realm of lossless image compression. These encompass a variety of methodologies, including autoregressive methods (ARMs) [29, 37], variational autoencoders (VAEs) [19, 22, 34, 35], normalizing flows (NFs) [12, 13, 41, 42], and diffusion models [14, 17]. These diverse approaches involve trade-offs between compression performance and inference speed. Furthermore, compression performance encompasses dataset compression, which often approaches the theoretical limits optimized by neural networks, and single-image compression, which is affected by the initial bits required for bits back coding [34]. Bits back coding is utilized in VAEs, continuous NFs, and variational diffusion models. In summary, per pixel ARMs and diffusion models demonstrate impressive compression performance but necessitate impractical inference time [22, 28]. Among the remaining methods, continuous NFs offer the best dataset compression but exhibit weaker performance in single-image compression. + +A potential option for achieving a better trade-off is the combined use of subimage autoregression and VAEs, a kind of latent variable model. This approach has shown competitive performance in the field of lossless image compression, particularly in its single-image compression performance when using the Autoregressive Initial Bits (ArIB) framework [28]. ArIB excludes latent variables for some subimages and uses their bits as initial bits, significantly reducing the overhead associated with initial bits for single-image compression. + +Current methods [28] split the image in the space dimension. Pixels with the same coordinate in different subimages are adjacent in the original image. This splitting uniformly distributes the information of latent variables in each subimage. Thus, each subimage carries similar importance for latent variables. However, latent variables are inadequately used due to their exclusion for some subimages. In addition, there's a growing risk of posterior collapse [21] along the subimage sequence due to the growing volume of the autoregressive prior. Posterior collapse, a phenomenon where non-useful latent variables are learned, adversely im + +pacts compression performance. Given the importance of all subimages for latent variables, current methods suffer from posterior collapse [28]. + +To address these issues, we propose the utilization of bit plane slicing (BPS). BPS splits images along the bit plane dimension, with a reduction in the volume of global data modality from the most significant plane (MSP) to the least significant plane (LSP) (see Figure 2). Given that latent variables are utilized for learning global data modalities [22, 28], the importance for latent variables decreases from MSP to LSP (see Figure 3). We arrange subimages with decreasing importance for latent variables. Hence, a more adequate utilization of latent variables is achieved by excluding subimages with minor importance. Moreover, we enhance autoregression by introducing the autoregression from MSP to LSP. This autoregression has a mitigated risk of posterior collapse since subimages important for latent variables are subject to a low risk of posterior collapse. + +The introduction of BPS increases the number of dimensions of subimages. When combining autoregression in the plane dimension with that in the existing space and color dimensions, an efficient method is required. To solve the problem, we propose a dimension-tailored autoregressive model. Autoregressive methods are tailored based on the unique characteristics of each dimension: a non-neural distribution model for the simplest color space, a simple neural network for the space dimension, and a much more complex network for the challenging plane dimension. This model demonstrates a substantial improvement in performance, balanced with a suitable increase in complexity, as evidenced by our experimental results. + +Consequently, our method outperforms the current ArIB methods in compression performance. Moreover, compared to the state-of-the-art NF-based methods, our approach demonstrates superior compression performance with comparable inference time. + +Our contributions could be summarized as follows: + +- We introduce bit plane slicing (BPS) for ArIB, since it splits the subimages with different importance for latent variables. By arranging subimages with decreasing importance for latent variables, BPS enables adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse. +- We propose a dimension-tailored autoregressive model. This model tailors autoregression methods for each dimension based on their different characteristics, efficiently modeling dependencies in an increased number of dimensions caused by BPS. +- With the merit of the BPS and dimension-tailored autoregressive model, our method surpasses the state-of-the-art normalizing-flow-based methods in compression performance with comparable inference time. + +# 2. Related Work + +Autoregressive Models compress images by predicting the probability of the current component based on previously decoded ones. Per-pixel ARMs [29, 37] decode images pixel by pixel, yielding impressive compression performance. However, as image resolution increases, the number of required network evaluations grows linearly, leading to impractically long decoding times. To mitigate this issue, subimage ARMs divide an image into a fixed number of subimages, significantly reducing inference time in comparison to per-pixel ARMs. Various image-splitting methods have been proposed, including spatial splitting [9], channelwise splitting [24], unevenly channel-wise splitting [10], lossy and residual decomposition [23], autoregressive subpixel convolution [28], and low-high frequency decomposition [26]. In the neural compression community, autoregression in the bit plane dimension has been adopted for order-agnostic autoregressive diffusion models [14], progressive compression [15], and high bit-depth medical image compression [38]. These methods employ BPS for different goals. Our method leverages autoregression in the bit plane dimension for ArIB to enhance autoregression in a mitigated risk of posterior collapse. + +Latent variable models utilize extracted latent variables from input images to aid in compression. L3C [22] employs a neural network to extract latent variables as auxiliary information. VAE-based methods [19, 34, 35] use posterior sampling for obtaining latent variables to improve performance. LBB [12] defines a posterior distribution for latent variables, enabling discretization for continuous NFs, which is also applied in iFlow [41]. However, posterior sampling requires bits back coding [34], which introduces significant overhead for single-image compression. + +Combined methods integrate latent variable models and ARMs. They have witnessed great success in the neural lossy compression community [25] and have been explored in lossless compression recently [16, 28]. This combination enables the efficient capture of both local and global data modalities. SHVC [28] further introduces ArIB based on the combination to reduce the overhead by initial bits. + +# 3. Method + +In this section, we introduce our Autoregressive Initial Bits with Bit Plane Slicing (ArIB-BPS) method. To provide a clear overview of our method, Figure 1 illustrates the key components and workflow of ArIB-BPS. + +# 3.1. Background + +Subimage Autoregressive Models split the image $x$ into a fixed number of subimages $\{x_1, x_2, \ldots, x_n\}$ . The probability of $x$ is factorized into the product of conditional distributions as follows + +![](images/7d03d7ce93ed414d6a85d9c89617f41b9d69e3b349da280f12e71f1600478ff8.jpg) +Figure 1. Overview of our method: the left subfigure illustrates the encoding process. Here, $s$ denotes the split index. The depth of the orange indicates the complexity of the networks in the dimension-tailored autoregression. In the encoding process, the insignificant planes $x_{s+1:s}$ are first encoded conditioned on the significant planes $x_{1:s}$ . Then, the posterior distribution $q(z|x_{1:s})$ is obtained from $x_{1:s}$ . The latent variables $\hat{z}$ are obtained by decoding from the previously encoded bitstream using $q(z|x_{1:s})$ . Next, $x_{1:s}$ is encoded, conditioned on $\hat{z}$ . Finally, $\hat{z}$ is encoded with prior distribution $p(z)$ . The decoding order is the reverse of the encoding process. + +![](images/188c710655e20d7238d1f936dcfb9019b8583fd5245c6c067194229d733d008e.jpg) + +$$ +p (x) = p \left(x _ {1}\right) \prod_ {i = 2} ^ {n} p \left(x _ {i} \mid x _ {1: i - 1}\right) \tag {1} +$$ + +Latent variable models extract latent variables $z$ from the input image $x$ to assist compression. For VAEs and continuous NFs, the approach to obtain $z$ is sampling for a posterior distribution $q(z|x)$ . In this case, bits back coding [34] is required. It encodes the image via the following steps: + +Step 1: Decode $z$ with $q(z|x)$ from initial bits. +Step 2: Encode $x$ with $p(x|z)$ . +Step 3: Encode $z$ with $p(z)$ . + +Here, $p(\cdot)$ represents the prior distribution. The asymmetric numeral system (ANS) [8], an entropy coder with a reversed order between encoding and decoding, is required. + +LBB [12] introduces local bits back, using $q(z|x)$ to discretize continuous NFs for lossless compression. Moreover, VAEs use neural networks to obtain $q(z|x)$ . When sequentially encoding a set of images $\{x^{(1)}, x^{(2)}, \ldots, x^{(N)}\}$ , bits from previously encoded images could be used as initial bits for the current image. For VAEs, the average code length for encoding the image set is + +$$ +\begin{array}{l} \bar {C} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(- \log p \left(x ^ {(i)} \mid z ^ {(i)}\right) - \log p \left(z ^ {(i)}\right) \right. \tag {2} \\ \left. + \log q \left(z ^ {(i)} \mid x ^ {(i)}\right)\right) - \frac {1}{N} \log q \left(z ^ {(1)} \mid x ^ {(1)}\right) \\ \end{array} +$$ + +The first term, which is the Evidence Lower Bound (ELBO), can be optimized by neural networks. The second term can be amortized when compressing a dataset. + +However, when compressing a single image, the overhead $-\log q(z|x)$ is not negligible, which limits the usage of these methods for single-image compression. Nonetheless, single-image compression is required for image compression applications, and it enables parallel compression when compressing a dataset. + +When combined with subimage autoregression, this overhead could be reduced via autoregressive initial bits (ArIB) [28]. ArIB partitions the subimage sequence into two subsets: $x_{1:s}$ and $x_{s+1:n}$ . Latent variables are employed exclusively for $x_{1:s}$ , while the bits used to compress $x_{s+1:n}$ serve as initial bits. If the number of bits provided is larger than the number of bits required, the overhead for bits back coding is eliminated. + +# 3.2. Bit Plane Slicing for Autoregressive Initial Bits + +The current method, SHVC-ArIB [28], splits the image in the space dimension. Its splitting ensures pixels with the same coordinates are adjacent in the original image. Thus, each subimage is equally important for latent variables. This splitting causes inadequate use of latent variables, as latent variables are excluded for some subimages that are important for latent variables. For SHVC-ArIB, this incurs a performance cost compared to utilizing latent variables for all subimages. Moreover, the performance of SHVC-ArIB is constrained by posterior collapse [21]. Posterior collapse refers to the collapse of the posterior to the prior, indicating that non-useful information is extracted by latent variables. A common cause of posterior collapse is highly capable decoders [21]. Specifically, a strong autoregressive prior enhances the decoder's capability, increasing the likelihood of posterior collapse. Consequently, subimages positioned at + +![](images/54b2e46c782b8aeb60659c71d05e33ff7addb2d37bbdfc740b3b26f31ac3e712.jpg) +Figure 2. Visualization for each bit plane. The planes are arranged from MSP to LSP, from left-top to right-bottom. The volume of global data modality decreases from MSP to LSP. + +the end of the autoregressive sequence are in a high risk of posterior collapse due to their strong autoregressive priors. In SHVC-ArIB, enhancing spatial autoregression leads to slight improvements or even a decline in performance because of posterior collapse. In this paper, we introduce bit plane slicing (BPS) to address these issues. + +Bit Plane Slicing is a well-established technique in digital image processing. For a $d$ -bit image, it partitions the entire image into $d$ bit planes. Here, each pixel intensity is represented as $x$ , and the $l^{th}$ bit is represented as $x^l$ . The value of $x^l$ can be determined using the following equation: + +$$ +x ^ {l} = \left\lfloor x / 2 ^ {d - l} \right\rfloor \mod 2 \tag {3} +$$ + +In this equation, $\lfloor x\rfloor$ denotes the largest integer less than or equal to $x$ . The original pixel value $x$ can be reconstructed using the equation: + +$$ +x = \sum_ {l = 1} ^ {d} 2 ^ {d - l} \times x ^ {l} \tag {4} +$$ + +In this paper, we focus specifically on 8-bit images. We define the significant planes as those with smaller $l$ values, and the insignificant planes as those with larger $l$ values. The volume of global data modality decreases from the most significant plane (MSP) to the least significant plane (LSP), as illustrated in Figure 2. Considering latent variables are employed to learn global data modalities [22, 28], we hypothesize that the importance for latent variables similarly decreases from MSP to LSP. To test this hypothesis, we employ the latent variable model from [34] for each bit plane. Subsequently, we evaluate the Bit Per Dimension (BPD) savings and calculate the bit rate percentage attributed to latent variables within the entire subimage. The results, illustrated in Figure 3, display a consistent descending trend from MSP to LSP, with values becoming negligible in the final few insignificant planes. This observation corroborates our initial hypothesis. + +Based on this finding, we employ BPS for ArIB. We partition the image into two segments using a slice index rang- + +ing in $\{1, \ldots, 8\}$ , denoted as $s$ . The first segment comprises the significant planes, represented as $x^{1:s}$ , while the second segment includes the insignificant planes, denoted as $x^{s+1:8}$ . Latent variables $z$ are exclusively used for the significant planes, while ARMs are employed for both segments. This approach enables a more adequate use of latent variables since $x^{1:s}$ is more important for them. Moreover, we introduce plane-by-plane autoregression from MSP to LSP to enhance autoregression. As the volume of autoregressive prior increases along the subimage sequence, the risk of posterior collapse intensifies sequentially. In this autoregression, subimages that are more important for latent variables are placed in positions where they are less likely to encounter posterior collapse. In comparison, SHVC-ArIB split the image into subimages with similar importance, meaning that subimages in positions where there's a high risk of posterior collapse are also important for latent variables. Thus, the overall risk of posterior collapse is mitigated by our method. + +The overall bit rate of the image $x$ can be expressed as follows, where $R_{sig}$ and $R_{ins}$ denote the bit rates associated with significant and insignificant planes, respectively: + +$$ +\begin{array}{l} R _ {x} = R _ {s i g} + R _ {i n s} \\ = - \log p \left(x ^ {1: s} \mid z\right) - \log p \left(x ^ {s + 1: 8} \mid x ^ {1: s}\right) \tag {5} \\ \end{array} +$$ + +We further propose some modifications to suit our ArIB with BPS method. + +Hierarchical Latent Variables are commonly utilized in VAEs to improve performance. These variables are structured in a hierarchy comprising $L$ layers, denoted as $z = \{z_{1},\dots ,z_{L}\}$ . The prior probability distribution of $z$ is factorized as follows: + +$$ +p \left(z _ {1: L}\right) = p \left(z _ {L}\right) \prod_ {i = 1} ^ {L - 1} p \left(z _ {i} \mid z _ {i + 1: L}\right) \tag {6} +$$ + +There are two methods for obtaining the posterior $q(z|x^{1:s})$ : top-down, which proceeds from layer $L$ to layer 1, and bottom-up, which goes from layer 1 to layer $L$ . The former has been shown to exhibit superior compression performance [32, 35], while the latter can leverage the bit-swap technique [19] to reduce the number of initial bits required. Given that the bit rate increases from MSP to LSP [43], insignificant planes can contribute a larger number of initial bits. Consequently, we opt for the top-down inference models, a choice that differentiates our approach from SHVC-ArIB. In this context, the posterior distribution can be expressed as follows: + +$$ +q \left(z _ {1: L} \mid x ^ {1: s}\right) = q \left(z _ {L} \mid x ^ {1: s}\right) \prod_ {i = 1} ^ {L - 1} q \left(z _ {i} \mid z _ {i + 1: L}, x ^ {1: s}\right) \tag {7} +$$ + +![](images/0a9ad64b720dd402783f67b9793b19c96f95a77337ebbb9728f28cf3b5755bb5.jpg) +Figure 3. BPD savings achieved using latent variables and the bit rate percentage attributed to latent variables for each bit plane. The results are averaged over the CIFAR10, ImageNet32, and ImageNet64 [6] datasets. The importance for latent variables decreases from MSP to LSP. + +Discretization also plays a pivotal role: a higher precision tends to enhance compression performance but also requires more initial bits. Given the variation in available initial bits across different images, we propose an image-adaptive discretization strategy. + +In our method, we utilize equal mass discretization [35] for the top-down model. This ensures that all latent variables maintain equal probability mass for $p(z)$ . We define the discretization precision $k$ , which discretizes $z$ within $2^{k}$ possible values. It is important to note that $q(z|x)$ closely approximates $p(z)$ , as the Kullback-Leibler divergence between them is minimized during optimization. As a result, each element of $z$ necessitates roughly $-\log (1 / (2^k)) = k$ initial bits. Furthermore, the size of the latent variables, denoted as $|z|$ , can be quantified. The bit rate $R_{ins}$ can be determined by measuring the length of the encoded bitstream. Following these calculations, the precision $k$ is derived using the equation: + +$$ +k = \left\lfloor \frac {R _ {i n s}}{| z |} \right\rfloor \tag {8} +$$ + +For training VAEs, the bit rate of latent variables is calculated with the condition that they are discretized to high precision. In practical scenarios, most images exhibit a sufficiently large precision value $k$ , ensuring negligible impact on compression performance when employing discretization. Nevertheless, low precision is required in minor cases, where the discretization hurts the compression performance. To tackle this, we propose a discretized + +sampling technique for sampling latent variables for finetuning. More specifically, both the prior and posterior distributions are modeled using a univariate logistic distribution, characterized by mean parameters $\mu_p$ and $\mu_q$ , respectively. We sample $\epsilon$ from a standard logistic distribution. The calculation of a sampled latent variable $\hat{z}$ and its corresponding bit rate can be represented as follows: + +$$ +\begin{array}{l} t _ {l} = \left\lceil 2 ^ {k} \cdot \operatorname {s i g m o i d} \left(\epsilon + \mu_ {q} - \mu_ {p}\right)\right\rceil - 1, \\ \hat {z} = \operatorname {l o g i t} \left(\frac {t _ {l} + 0 . 5}{2 ^ {k}}\right) + \mu_ {p}, \\ R _ {\hat {z}} = \log (\operatorname {s i g m o i d} (\operatorname {l o g i t} \left(\frac {t _ {l} + 1}{2 ^ {k}}\right) + \mu_ {p} - \mu_ {q}) \tag {9} \\ - \operatorname {s i g m o i d} (\log \mathrm {i t} \left(\frac {t _ {l}}{2 ^ {k}}\right) + \mu_ {p} - \mu_ {q})) + k, \\ \end{array} +$$ + +where $\lceil x\rceil$ signifies the smallest integer greater than or equal to $x$ . During back-propagation, the Straight Through Estimator [3] is utilized to manage this operation. + +# 3.3. Dimension-Tailored Autoregressive Model + +BPS results in an increased number of dimensions for autoregression, specifically, the bit plane dimension. In addition to the plane dimension, we consider autoregression in space and color dimensions, which have demonstrated their efficacy in previous studies [26, 40]. Using a uniform approach to model all dimensions is employed in previous works [14]. However, this method has certain limitations. First, these dimensions present uneven modeling challenges. Employing a one-size-fits-all network to address all dimensions directly may either compromise modeling accuracy in the more complex dimensions or lead to inefficient use of computational resources in simpler dimensions. Second, strong autoregression in the space and color dimensions can potentially lead to posterior collapse. To address these issues, we propose a dimension-tailored autoregressive model that tailors the approach for each dimension based on its specific characteristics. An overview is shown in the right subfigures of Figure 1. + +Our autoregressive model operates in three stages. In the first stage, we focus on the plane dimension, which presents a significant challenge due to the extensive range of possible values that the pixels in its preceding planes, denoted as $x^{1:i - 1}$ , can take on, amounting to $2^{i - 1}$ possibilities. To effectively leverage the autoregressive prior from $x^{1:i - 1}$ and apply it to $x^i$ , we employ a sophisticated plane context model, $\Phi_p$ . To capture the global data modality embedded in the preceding planes, we adopt a U-Net architecture [27]. Specifically, we use two models that employ the architecture of diffusion models [7]: one for the significant planes and the other for the insignificant planes. Time embedding is used to differentiate planes. + +In the second stage, we address the space dimension. Compared to the plane dimension, this dimension is less complex, as each bit in the current bit plane can only take one of two possible values. We adopt a 4-stage checkerboard model [9, 26], which involves dividing the bit plane $x^{i}$ into four sub-planes, denoted as $x_{1}^{i}, x_{2}^{i}, x_{3}^{i}, x_{4}^{i}$ , as illustrated in Figure 1. For each sub-plane $x_{j}^{i}$ , we utilize a simple network, $\Phi_{s}$ , to extract the spatial autoregressive prior $F_{s}$ from $x_{1:j-1}^{i}$ . Then, a parameter network $\Theta$ is used to fuse information from both the plane and space dimensions. To effectively integrate information, we employ a relatively complex network for $\Theta$ . Although the complexity of $\Theta$ does contribute to the overall computational demand of our model, the complexity remains significantly lower than that of employing the architecture of $\Phi_{p}$ for both dimensions, resulting in a more efficient modeling process. More specifically, $\Phi_{s}$ is a single ResBlock [11], and $\Theta$ combines an ECANet [39] with several wider ResBlocks. + +The third stage is for the color dimension. As identified in PixelCNN++ [29], the correlations among the color channels of a pixel are typically straightforward, and deep networks are often unnecessary for their modeling. Therefore, we adopt a non-network approach for the color dimension. Specifically, we model the distribution of bits in each color channel using a Bernoulli distribution, where the probability parameter $p = P(x = 1)$ is determined autoregressively based on the bits with the same spatial position in previously decoded channels. In our method, $\Theta$ predict six parameters $\theta = \{f_r, f_g, f_b, \alpha, \beta, \gamma\}$ , and model the probability for each channel as follows: + +$$ +p _ {r} = \operatorname {s i g m o i d} \left(f _ {r}\right) +$$ + +$$ +p _ {g} = \operatorname {s i g m o i d} \left(f _ {g} + \alpha \cdot x _ {r}\right) \tag {10} +$$ + +$$ +p _ {b} = \operatorname {s i g m o i d} \left(f _ {b} + \beta \cdot x _ {r} + \gamma \cdot x _ {g}\right) +$$ + +The 1st and 2nd stages of our dimension-tailored autoregressive model can be expressed as follows: + +$$ +F _ {p ^ {i}} = \Phi_ {p} (x ^ {1: i - 1}) +$$ + +$$ +F _ {s _ {j} ^ {i}} = \Phi_ {s} \left(x _ {1: j - 1} ^ {i}\right) \tag {11} +$$ + +$$ +\theta_ {j} ^ {i} = \left\{ \begin{array}{l l} \Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}, F _ {z}), & i \leq s \\ \Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}), & i > s \end{array} \right. +$$ + +In summary, we initially derive $F_{p}$ for the entire bit plane. Subsequently, for each of the four subimages in the bit plane, we compute $F_{s}$ . These are then concatenated with $F_{p}$ and $F_{z}$ (only for significant planes) and fed into $\Theta$ to determine the six parameters $\theta$ . Finally, we calculate the probability using non-network color autoregression as outlined in Equation 10. + +Loss function of our method is expressed as: + +$$ +\begin{array}{l} l = R _ {z} + R _ {x} \tag {12} \\ = - \log p (z) + \log q (z | x ^ {1: s}) + R _ {x} \\ \end{array} +$$ + +# 4. Experiments + +# 4.1. Architecture and Training Details + +In accordance with prior research, our experimental setup entails training three models using three benchmark datasets: CIFAR10, ImageNet32, and ImageNet64 [6]. For all models, we set the split index, $s$ , to 4, based on Figure 3, and utilize 3 layers of hierarchical latent variables. For latent variables, we apply a factor 2 downsampling for each hierarchy of latent variables in latent encoders. The channel number for latent variables is set to 3 across all models. For $\Phi_p$ , the number of hierarchical levels in the U-Net is set to 4 across all models. + +We utilize the Adam optimizer [18] with an initial learning rate of 0.0001, and we apply step decay as the learning rate schedule. For fine-tuning with discretized sampling, we employ Equation 8 with the previously trained model to obtain the precision $k$ for each image in the training set. All experiments are conducted on a GeForce GTX 1080 Ti. + +# 4.2. Compression Performance + +We evaluate the performance of our method across four distinct datasets: CIFAR10, ImageNet32, ImageNet64, and DIV2K [1], using bits per dimension (BPD) as our primary metric. The first three datasets are of low resolution, while the last one, DIV2K, is a high-resolution dataset. Following most previous works, we exclude per-pixel autoregressive methods [29, 37] and diffusion models [14, 17] from our evaluation due to their impractical time complexity. + +Our method evaluates performance in two distinct compression settings: dataset compression and single-image compression. In dataset compression, we sequentially compress 5,000 images with a fixed discretization precision of 10. For single-image compression, each image's $k$ value is calculated using Equation 8 and clipped between 3 and 10. If the calculated $k$ is less than 3, additional $(3 - k) \times |z|$ ( $|z|$ represents the size of $z$ ) bits are added to ensure successful encoding. These possible bits, along with the $k$ value and image bits, are written to a bitstream. The single image compression performance is determined by measuring the bitstream size. Following [28, 41, 42], we evaluate the performance on the DIV2K dataset using the model trained on ImageNet64 by cropping images into $64 \times 64$ patches. + +We compare our method against advanced approaches from traditional methods, learned methods without bits back coding, and learned methods with bits back coding. We present the detailed results of our experiments in Table 1. Our method demonstrates superior compression performance, particularly in single-image compression. In dataset compression, our method excels in three out of four datasets, achieving the second-best performance on the remaining dataset. In single-image compression, our method outperforms other methods in all datasets. We note + +Table 1. Compression performance in BPD for four datasets, where lower values indicate better performance. The best performance is highlighted in bold. *: ImageNet32, **: ImageNet64, +: Flickr2k [33], -: OpenImage [20]. 1st Group: traditional methods, 2nd Group: learned methods without bits back coding, 3rd Group: learned methods with bits back coding. Some values of SHVC and SHVC-ArIB are not available, and we list their available lower bound here. △: These values are obtained by adding dataset performance and initial bits. + +
Dataset CompressionSingle-Image Compression
MethodCIFAR10ImageNet32ImageNet64DIV2KCIFAR10ImageNet32ImageNet64DIV2K
PNG [5]5.896.395.744.235.896.395.744.23
FLIF [31]4.194.524.542.914.194.524.542.91
JPEG-XL [2]5.746.395.892.795.746.395.892.79
L3C [22]-4.764.423.09--4.764.423.09-
IDF [13]3.344.183.90-3.344.183.90-
IDF++ [36]3.264.123.81-3.264.123.81-
LC-FDNet [26]3.77+4.40+4.28+2.71+3.77+4.40+4.28+2.71+
Bit-Swap [19]3.824.50--6.536.97--
Hilloc [35]3.56*4.203.90*-----
SHVC [28]3.163.983.682.57**>3.96>4.78>4.48-
SHVC-ArIB [28]>3.16>3.98>3.68->3.16>3.98>3.68-
iVPF [42]3.204.033.752.68**9.20Δ---
LBB [12]3.123.883.70-42.98Δ49.84Δ41.70Δ-
iFlow [41]3.123.883.702.57**37.40Δ38.27Δ38.12Δ-
ArIB-BPS (ours)3.063.913.632.55**3.073.923.642.59**
+ +Table 2. Comparison with advanced and available flow-based methods. The batchsize is 64 following LBB and iFlow. We train the small model to align with the inference speed of iFlow on ImageNet64. + +
Compression Performance (BPD) ↓Inference Time (ms/sample) ↓
datasetMethodtheoreticaldatasetsingleencodedecode
CIFAR10LBB [12]3.1163.11842.97864.9464.94
iFlow [41]3.1163.11837.39817.3847.56
ArIB-BPS3.0483.0573.07014.9314.93
ImageNet32LBB [12]3.8713.87549.835194.14194.14
iFlow [41]3.8713.87338.27374.84119.30
ArIB-BPS3.9043.9113.91852.9652.96
ImageNet64LBB [12]3.7013.70341.70395.5795.57
iFlow [41]3.7013.70338.12337.4958.08
ArIB-BPS (Small)3.6773.6853.69343.6543.65
ArIB-BPS3.6243.6333.641107.80107.80
+ +that precise values for the single-image compression performance of SHVC, as well as the compression performance of SHVC-ArIB, are not available. However, the lower bound could be obtained. Moreover, it could be inferred that our method surpasses SHVC-ArIB by an average of more than 0.10 BPD on low-resolution datasets in both compression settings. Additional details are provided in the Appendix. + +To further demonstrate the effectiveness of our approach, we conducted a detailed comparison between our method and continuous normalizing-flow-based methods, which exhibit state-of-the-art performance in dataset compression. As Table 2 illustrates, ArIB-BPS outperforms iFlow in both single-image and dataset compression performance for CIFAR10 and ImageNet64, with comparable inference time. + +For ImageNet32, although ArIB-BPS lags slightly behind in dataset compression performance, it excels in other metrics, particularly in single-image compression. These results imply that our method achieves a better trade-off between compression performance and inference speed. + +# 4.3. Ablation Studies + +We conduct three ablation studies on the CIFAR10 dataset to show the efficiency of our dimension-tailored autoregressive model and the effectiveness of discretized sampling. We note that results in Table 3 and 4 are evaluated using models without using discretized sampling for fine-tuning. Therefore, the results of ArIB-BPS are slightly different from those in Table 2. + +Table 3. Ablation for the effectiveness of autoregression on plane, space, and color dimensions. Inference time is identical for encoding and decoding. The batchsize is 64. + +
planespacecolorTheoretical BPD ↓Inference Time (ms/sample) ↓
✓✓✓3.05214.93
✓✓3.45814.92
✓✓3.47010.36
✓✓3.5225.83
+ +Table 4. Compression performance and inference time of various usages of the architecture of $\Phi_p$ . The batch size is 64. P: plane. S: space. C: color. Sig: significant planes. Ins: insignificant planes. We also train two small models to further validate our design. + +
Theoretical BPD ↓Inference Time (ms/sample) ↓
SigInsTotalSigInsTotal
P (Ours)1.0412.0113.0526.308.6414.93
PS1.0521.9933.04514.4624.9739.42
PSC1.0431.9713.01543.8374.97118.80
PS (Small)1.0552.0703.1259.0811.4420.51
PSC (Small)1.0962.2183.3149.6413.9623.59
+ +# 4.3.1 Dimension-Tailored Autoregressive Model + +Firstly, we conduct experiments to assess the impact of autoregression in each dimension. To do this, we train and evaluate three additional models, each with autoregression removed in either the color, space, or plane dimension. In the model that does not incorporate plane autoregression, we employ the color autoregressive model from PixelCNN++ [29] to ensure compatibility. Additionally, we enhance $\Phi_s$ by employing the architecture of $\Phi_p$ , accommodating the increased complexity in the spatial dimension, which now has 256 possible values per pixel instead of just 2. In our experiments, this configuration results in improved performance. Table 3 clearly demonstrates that incorporating autoregression in each dimension consistently leads to significant improvements in compression performance. We note that our autoregression for the color dimension incurs only a negligible overhead in inference time. + +Furthermore, we conduct an evaluation to assess the efficiency of our dimension-tailored design. Specifically, we train two additional models: one using the architecture of $\Phi_p$ for both the plane and space dimensions and the other using the architecture of $\Phi_p$ for all three dimensions. Subsequently, we evaluate the impact of these modifications on significant and insignificant planes separately. As depicted in Table 4, the dimension-tailored autoregressive model achieves superior performance in significant planes with lower complexity. This superiority is attributed to reduced hindrance to the latent variables, facilitated by sim + +![](images/2ed6e446a7050d348666012dc92dd0c28350ffa96cecb3c79bc4d103d508fdaf.jpg) +Figure 4. Average BPD savings through fine-tuning using discretized sampling across various discretization precisions. + +pler spatial autoregression and color-wise autoregression, as reflected in higher bit rates for latent variables. In the case of insignificant planes, although strengthened autoregression can lead to performance gains, it results in significantly longer inference time. To further validate our method's efficiency, we train two small models with reduced inference time. Our method demonstrates better compression performance. Considering the trade-off between inference time and compression performance, our design is more efficient. + +# 4.3.2 Discretized Sampling + +We assess the performance improvement after fine-tuning across various $k$ values, as illustrated in Figure 4. The method improves performance across all precision values, particularly at lower precision levels. While low-precision cases form a minor segment, their optimization boosts performance in edge cases. + +# 5. Conclusion + +We have proposed ArIB-BPS, a new approach for lossless image compression. BPS enables us to achieve adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse by arranging subimages with decreasing importance for latent variables. In addition, we present a dimension-tailored autoregressive model that takes into account the different characteristics of each dimension. It efficiently captures dependencies in space, color, and plane dimensions. Experimental results demonstrate that our method achieves superior compression performance compared to the state-of-the-art normalizing-flow-based methods with comparable inference time. + +The limitation of our work lies in the complexity of higher bit-depth. The complexity of our method would increase as the bit-depth increases due to the interdependency among bit planes. To extend our method for high-bit-depth images, efficient strategies such as handling multiple planes in one step should be explored in our future work. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6 +[2] Jyrki Alakuijala, Ruud Van Asseldonk, Sami Boukortt, Martin Bruse, Iulia-Maria Comsa, Moritz Firsching, Thomas Fischbacher, Evgenii Kliuchnikov, Sebastian Gomez, Robert Obryk, et al. Jpeg xl next-generation image compression architecture and coding tools. In Applications of Digital Image Processing XLII, pages 112-124. SPIE, 2019. 7 +[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5 +[4] Sam Bond-Taylor, Adam Leach, Yang Long, and Chris George Willcocks. Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1 +[5] Thomas Boutell. PNG (portable network graphics) specification version 1.0. Technical report, 1997. 7 +[6] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017. 5, 6 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 5 +[8] Jarek Duda. Asymmetric numeral systems. arXiv preprint arXiv:0902.0271, 2009. 3 +[9] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 6 +[10] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 6 +[12] Jonathan Ho, Evan Lohn, and Pieter Abbeel. Compression with flows via local bits-back coding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 3, 7 +[13] Emiel Hoogeboom, Jorn Peters, Rianne Van Den Berg, and Max Welling. Integer discrete flows and lossless compression. Advances in Neural Information Processing Systems, 32, 2019. 1, 7 +[14] Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. In International Conference on Learning Representations, 2022. 1, 2, 5, 6 + +[15] Seungmin Jeon, Kwang Pyo Choi, Youngo Park, and Chang-Su Kim. Context-based trit-plane coding for progressive image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14348-14357, 2023. 2 +[16] Ning Kang, Shanzhao Qiu, Shifeng Zhang, Zhenguo Li, and Shu-Tao Xia. Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739-3748, 2022. 2 +[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in Neural Information Processing Systems, 34:21696-21707, 2021. 1, 6 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Friso Kingma, Pieter Abbeel, and Jonathan Ho. Bit-swap: Recursive bits-back coding for lossless compression with hierarchical latent variables. In International Conference on Machine Learning, pages 3408-3417, 2019. 1, 2, 4, 7 +[20] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 7 +[21] James Lucas, George Tucker, Roger B Grosse, and Mohammad Norouzi. Don't blame the elbo! a linear vae perspective on posterior collapse. Advances in Neural Information Processing Systems, 32, 2019. 1, 3 +[22] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Practical full resolution learned lossless image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10629-10638, 2019. 1, 2, 4, 7 +[23] Fabian Mentzer, Luc Van Gool, and Michael Tschannen. Learning better lossless compression using lossy compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6638-6647, 2020. 2 +[24] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing, pages 3339-3343. IEEE, 2020. 2 +[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in Neural Information Processing Systems, 31, 2018. 2 +[26] Hochang Rhee, Yeong Il Jang, Seyun Kim, and Nam Ik Cho. Lc-fdnet: Learned lossless image compression with frequency decomposition network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6033-6042, 2022. 2, 5, 6, 7 +[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen + +tation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 5 +[28] Tom Ryder, Chen Zhang, Ning Kang, and Shifeng Zhang. Split hierarchical variational compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 386-395, 2022. 1, 2, 3, 4, 6, 7 +[29] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P. Kingma. PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations, 2017. 1, 2, 6, 8 +[30] Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379-423, 1948. 1 +[31] Jon Sneyers and Pieter Wuille. Flif: Free lossless image format based on maniac compression. In IEEE International Conference on Image Processing, pages 66-70. IEEE, 2016. +[32] Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. Ladder variational autoencoders. Advances in Neural Information Processing Systems, 29, 2016. 4 +[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 114-125, 2017. 7 +[34] James Townsend, Thomas Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. In International Conference on Learning Representations, 2018. 1, 2, 3, 4 +[35] James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: lossless image compression with hierarchical latent variable models. In International Conference on Learning Representations, 2019. 1, 2, 4, 5, 7 +[36] Rianne van den Berg, Alexey A Gritsenko, Mostafa Dehghani, Casper Kaae Sønderby, and Tim Salimans. Idf++: Analyzing and improving integer discrete flows for lossless compression. In International Conference on Learning Representations, 2020. 7 +[37] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, 2016. 1, 2, 6 +[38] Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu. Learning lossless compression for high bit-depth medical imaging. In 2023 IEEE International Conference on Multimedia and Expo, pages 2549-2554. IEEE, 2023. 2 +[39] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 6 +[40] Honglei Zhang, Francesco Cricri, Hamed R Tavakoli, Nannan Zou, Emre Aksu, and Miska M Hannuksela. Lossless + +image compression using a multi-scale progressive statistical model. In Proceedings of the Asian Conference on Computer Vision, 2020. 5 +[41] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 1, 2, 6, 7 +[42] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 1, 6, 7 +[43] Zhizheng Zhang, Zhibo Chen, Jianxin Lin, and Weiping Li. Learned scalable image compression with bidirectional context disentanglement network. In IEEE International Conference on Multimedia and Expo, pages 1438-1443. IEEE, 2019. 4 \ No newline at end of file diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/images.zip b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5cda900a725137c7fc278280d0c50a52442d413a --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a67a6721c1bf08ef65f380fa177e0b5a5352a810bb139c79f102025ac99651dc +size 459670 diff --git a/2024/Learned Lossless Image Compression based on Bit Plane Slicing/layout.json b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dca57bde7b747acb6839493844fdaafabdec8e45 --- /dev/null +++ b/2024/Learned Lossless Image Compression based on Bit Plane Slicing/layout.json @@ -0,0 +1,8450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 496, + 121 + ], + "type": "text", + "content": "Learned Lossless Image Compression based on Bit Plane Slicing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 142, + 451, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 142, + 451, + 185 + ], + "spans": [ + { + "bbox": [ + 142, + 142, + 451, + 185 + ], + "type": "text", + "content": "Zhe Zhang1 Huairui Wang1 Zhenzhong Chen1* Shan Liu2 \n1Wuhan University 2Tencent Media Lab \n{zhe_zhang, zzchen}@whu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 239, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 502 + ], + "type": "text", + "content": "Autoregressive Initial Bits (ArIB), a framework that combines subimage autoregression and latent variable models, has shown its advantages in lossless image compression. However, in current methods, the image splitting makes the information of latent variables being uniformly distributed in each subimage, and causes inadequate use of latent variables in addition to posterior collapse. To tackle these issues, we introduce Bit Plane Slicing (BPS), splitting images in the bit plane dimension with the considerations on different importance for latent variables. Thus, BPS provides a more effective representation by arranging subimages with decreasing importance for latent variables. To solve the problem of the increased number of dimensions caused by BPS, we further propose a dimension-tailored autoregressive model that tailors autoregression methods for each dimension based on their characteristics, efficiently capturing the dependencies in plane, space, and color dimensions. As shown in the extensive experimental results, our method demonstrates the superior compression performance with comparable inference speed, when compared to the state-of-the-art normalizing-flow-based methods. The code is at https://github.com/ZZ022/ArIB-BPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 633 + ], + "type": "text", + "content": "Lossless image compression is widely applied in many domains where the preservation of high-quality images is demanded such as photography, scientific exploration, and remote sensing. The significance of these compression techniques lies in their ability to reduce image file sizes while preserving the original content, accomplished through the exploitation of inherent image correlations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 634, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 287, + 682 + ], + "type": "text", + "content": "Fundamentally, Shannon's source coding theorem [30] sets a lower bound for coding length based on the image's entropy, which is determined by its inherent distribution. Consequently, many deep generative models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 547, + 453 + ], + "type": "text", + "content": "[4], which have exhibited remarkable efficacy in distribution modeling, have found widespread application within the realm of lossless image compression. These encompass a variety of methodologies, including autoregressive methods (ARMs) [29, 37], variational autoencoders (VAEs) [19, 22, 34, 35], normalizing flows (NFs) [12, 13, 41, 42], and diffusion models [14, 17]. These diverse approaches involve trade-offs between compression performance and inference speed. Furthermore, compression performance encompasses dataset compression, which often approaches the theoretical limits optimized by neural networks, and single-image compression, which is affected by the initial bits required for bits back coding [34]. Bits back coding is utilized in VAEs, continuous NFs, and variational diffusion models. In summary, per pixel ARMs and diffusion models demonstrate impressive compression performance but necessitate impractical inference time [22, 28]. Among the remaining methods, continuous NFs offer the best dataset compression but exhibit weaker performance in single-image compression." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 457, + 547, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 579 + ], + "type": "text", + "content": "A potential option for achieving a better trade-off is the combined use of subimage autoregression and VAEs, a kind of latent variable model. This approach has shown competitive performance in the field of lossless image compression, particularly in its single-image compression performance when using the Autoregressive Initial Bits (ArIB) framework [28]. ArIB excludes latent variables for some subimages and uses their bits as initial bits, significantly reducing the overhead associated with initial bits for single-image compression." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Current methods [28] split the image in the space dimension. Pixels with the same coordinate in different subimages are adjacent in the original image. This splitting uniformly distributes the information of latent variables in each subimage. Thus, each subimage carries similar importance for latent variables. However, latent variables are inadequately used due to their exclusion for some subimages. In addition, there's a growing risk of posterior collapse [21] along the subimage sequence due to the growing volume of the autoregressive prior. Posterior collapse, a phenomenon where non-useful latent variables are learned, adversely im" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 713 + ], + "type": "text", + "content": "*Corresponding author. This work was supported by the National Natural Science Foundation of China (Grant No. 62036005) and Tencent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27579" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 108 + ], + "type": "text", + "content": "pacts compression performance. Given the importance of all subimages for latent variables, current methods suffer from posterior collapse [28]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 111, + 287, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 111, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 111, + 287, + 291 + ], + "type": "text", + "content": "To address these issues, we propose the utilization of bit plane slicing (BPS). BPS splits images along the bit plane dimension, with a reduction in the volume of global data modality from the most significant plane (MSP) to the least significant plane (LSP) (see Figure 2). Given that latent variables are utilized for learning global data modalities [22, 28], the importance for latent variables decreases from MSP to LSP (see Figure 3). We arrange subimages with decreasing importance for latent variables. Hence, a more adequate utilization of latent variables is achieved by excluding subimages with minor importance. Moreover, we enhance autoregression by introducing the autoregression from MSP to LSP. This autoregression has a mitigated risk of posterior collapse since subimages important for latent variables are subject to a low risk of posterior collapse." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 294, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 449 + ], + "type": "text", + "content": "The introduction of BPS increases the number of dimensions of subimages. When combining autoregression in the plane dimension with that in the existing space and color dimensions, an efficient method is required. To solve the problem, we propose a dimension-tailored autoregressive model. Autoregressive methods are tailored based on the unique characteristics of each dimension: a non-neural distribution model for the simplest color space, a simple neural network for the space dimension, and a much more complex network for the challenging plane dimension. This model demonstrates a substantial improvement in performance, balanced with a suitable increase in complexity, as evidenced by our experimental results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 513 + ], + "type": "text", + "content": "Consequently, our method outperforms the current ArIB methods in compression performance. Moreover, compared to the state-of-the-art NF-based methods, our approach demonstrates superior compression performance with comparable inference time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 516, + 266, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 516, + 266, + 527 + ], + "spans": [ + { + "bbox": [ + 58, + 516, + 266, + 527 + ], + "type": "text", + "content": "Our contributions could be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 532, + 287, + 711 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 532, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 287, + 603 + ], + "type": "text", + "content": "- We introduce bit plane slicing (BPS) for ArIB, since it splits the subimages with different importance for latent variables. By arranging subimages with decreasing importance for latent variables, BPS enables adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 604, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 287, + 662 + ], + "type": "text", + "content": "- We propose a dimension-tailored autoregressive model. This model tailors autoregression methods for each dimension based on their different characteristics, efficiently modeling dependencies in an increased number of dimensions caused by BPS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 663, + 287, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 663, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 47, + 663, + 287, + 711 + ], + "type": "text", + "content": "- With the merit of the BPS and dimension-tailored autoregressive model, our method surpasses the state-of-the-art normalizing-flow-based methods in compression performance with comparable inference time." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 354 + ], + "type": "text", + "content": "Autoregressive Models compress images by predicting the probability of the current component based on previously decoded ones. Per-pixel ARMs [29, 37] decode images pixel by pixel, yielding impressive compression performance. However, as image resolution increases, the number of required network evaluations grows linearly, leading to impractically long decoding times. To mitigate this issue, subimage ARMs divide an image into a fixed number of subimages, significantly reducing inference time in comparison to per-pixel ARMs. Various image-splitting methods have been proposed, including spatial splitting [9], channelwise splitting [24], unevenly channel-wise splitting [10], lossy and residual decomposition [23], autoregressive subpixel convolution [28], and low-high frequency decomposition [26]. In the neural compression community, autoregression in the bit plane dimension has been adopted for order-agnostic autoregressive diffusion models [14], progressive compression [15], and high bit-depth medical image compression [38]. These methods employ BPS for different goals. Our method leverages autoregression in the bit plane dimension for ArIB to enhance autoregression in a mitigated risk of posterior collapse." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 355, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 355, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 355, + 545, + 474 + ], + "type": "text", + "content": "Latent variable models utilize extracted latent variables from input images to aid in compression. L3C [22] employs a neural network to extract latent variables as auxiliary information. VAE-based methods [19, 34, 35] use posterior sampling for obtaining latent variables to improve performance. LBB [12] defines a posterior distribution for latent variables, enabling discretization for continuous NFs, which is also applied in iFlow [41]. However, posterior sampling requires bits back coding [34], which introduces significant overhead for single-image compression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 475, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 545, + 558 + ], + "type": "text", + "content": "Combined methods integrate latent variable models and ARMs. They have witnessed great success in the neural lossy compression community [25] and have been explored in lossless compression recently [16, 28]. This combination enables the efficient capture of both local and global data modalities. SHVC [28] further introduces ArIB based on the combination to reduce the overhead by initial bits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 570, + 361, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 361, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 361, + 582 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 638 + ], + "type": "text", + "content": "In this section, we introduce our Autoregressive Initial Bits with Bit Plane Slicing (ArIB-BPS) method. To provide a clear overview of our method, Figure 1 illustrates the key components and workflow of ArIB-BPS." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 647, + 386, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 386, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 386, + 658 + ], + "type": "text", + "content": "3.1. Background" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "Subimage Autoregressive Models split the image " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " into a fixed number of subimages " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\{x_1, x_2, \\ldots, x_n\\}" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": ". The probability of " + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": " is factorized into the product of conditional distributions as follows" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27580" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 260, + 240 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 260, + 240 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 260, + 240 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 260, + 240 + ], + "type": "image", + "image_path": "7d03d7ce93ed414d6a85d9c89617f41b9d69e3b349da280f12e71f1600478ff8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": "Figure 1. Overview of our method: the left subfigure illustrates the encoding process. Here, " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " denotes the split index. The depth of the orange indicates the complexity of the networks in the dimension-tailored autoregression. In the encoding process, the insignificant planes " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "x_{s+1:s}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " are first encoded conditioned on the significant planes " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "x_{1:s}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": ". Then, the posterior distribution " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "q(z|x_{1:s})" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "x_{1:s}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": ". The latent variables " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " are obtained by decoding from the previously encoded bitstream using " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "q(z|x_{1:s})" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": ". Next, " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "x_{1:s}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " is encoded, conditioned on " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": " is encoded with prior distribution " + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 46, + 253, + 547, + 310 + ], + "type": "text", + "content": ". The decoding order is the reverse of the encoding process." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 69, + 508, + 242 + ], + "blocks": [ + { + "bbox": [ + 267, + 69, + 508, + 242 + ], + "lines": [ + { + "bbox": [ + 267, + 69, + 508, + 242 + ], + "spans": [ + { + "bbox": [ + 267, + 69, + 508, + 242 + ], + "type": "image", + "image_path": "188c710655e20d7238d1f936dcfb9019b8583fd5245c6c067194229d733d008e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 339, + 287, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 339, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 104, + 339, + 287, + 370 + ], + "type": "interline_equation", + "content": "p (x) = p \\left(x _ {1}\\right) \\prod_ {i = 2} ^ {n} p \\left(x _ {i} \\mid x _ {1: i - 1}\\right) \\tag {1}", + "image_path": "2cd478b7dfbb3517b2311d9620d4ae4513cfe7990fc2a0039e9c5cda3b62ec69.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": "Latent variable models extract latent variables " + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": " from the input image " + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": " to assist compression. For VAEs and continuous NFs, the approach to obtain " + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": " is sampling for a posterior distribution " + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "inline_equation", + "content": "q(z|x)" + }, + { + "bbox": [ + 46, + 376, + 287, + 435 + ], + "type": "text", + "content": ". In this case, bits back coding [34] is required. It encodes the image via the following steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 436, + 251, + 472 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "spans": [ + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "text", + "content": "Step 1: Decode " + }, + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "inline_equation", + "content": "q(z|x)" + }, + { + "bbox": [ + 58, + 436, + 251, + 448 + ], + "type": "text", + "content": " from initial bits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "spans": [ + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "text", + "content": "Step 2: Encode " + }, + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "inline_equation", + "content": "p(x|z)" + }, + { + "bbox": [ + 58, + 448, + 186, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "spans": [ + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "text", + "content": "Step 3: Encode " + }, + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 58, + 460, + 177, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "inline_equation", + "content": "p(\\cdot)" + }, + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "text", + "content": " represents the prior distribution. The asymmetric numeral system (ANS) [8], an entropy coder with a reversed order between encoding and decoding, is required." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "content": "LBB [12] introduces local bits back, using " + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "inline_equation", + "content": "q(z|x)" + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "content": " to discretize continuous NFs for lossless compression. Moreover, VAEs use neural networks to obtain " + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "inline_equation", + "content": "q(z|x)" + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "content": ". When sequentially encoding a set of images " + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\{x^{(1)}, x^{(2)}, \\ldots, x^{(N)}\\}" + }, + { + "bbox": [ + 46, + 509, + 287, + 593 + ], + "type": "text", + "content": ", bits from previously encoded images could be used as initial bits for the current image. For VAEs, the average code length for encoding the image set is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 73, + 611, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 611, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 73, + 611, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\bar {C} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(- \\log p \\left(x ^ {(i)} \\mid z ^ {(i)}\\right) - \\log p \\left(z ^ {(i)}\\right) \\right. \\tag {2} \\\\ \\left. + \\log q \\left(z ^ {(i)} \\mid x ^ {(i)}\\right)\\right) - \\frac {1}{N} \\log q \\left(z ^ {(1)} \\mid x ^ {(1)}\\right) \\\\ \\end{array}", + "image_path": "82a5e896051c8d1b5037d325464704b64dfb1a0388bb5f97c35a8deb2faaca14.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "The first term, which is the Evidence Lower Bound (ELBO), can be optimized by neural networks. The second term can be amortized when compressing a dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 330, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 401 + ], + "type": "text", + "content": "However, when compressing a single image, the overhead " + }, + { + "bbox": [ + 304, + 330, + 545, + 401 + ], + "type": "inline_equation", + "content": "-\\log q(z|x)" + }, + { + "bbox": [ + 304, + 330, + 545, + 401 + ], + "type": "text", + "content": " is not negligible, which limits the usage of these methods for single-image compression. Nonetheless, single-image compression is required for image compression applications, and it enables parallel compression when compressing a dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "content": "When combined with subimage autoregression, this overhead could be reduced via autoregressive initial bits (ArIB) [28]. ArIB partitions the subimage sequence into two subsets: " + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "inline_equation", + "content": "x_{1:s}" + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "inline_equation", + "content": "x_{s+1:n}" + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "content": ". Latent variables are employed exclusively for " + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "inline_equation", + "content": "x_{1:s}" + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "content": ", while the bits used to compress " + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "inline_equation", + "content": "x_{s+1:n}" + }, + { + "bbox": [ + 304, + 402, + 546, + 497 + ], + "type": "text", + "content": " serve as initial bits. If the number of bits provided is larger than the number of bits required, the overhead for bits back coding is eliminated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 503, + 545, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 503, + 545, + 516 + ], + "spans": [ + { + "bbox": [ + 305, + 503, + 545, + 516 + ], + "type": "text", + "content": "3.2. Bit Plane Slicing for Autoregressive Initial Bits" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 714 + ], + "type": "text", + "content": "The current method, SHVC-ArIB [28], splits the image in the space dimension. Its splitting ensures pixels with the same coordinates are adjacent in the original image. Thus, each subimage is equally important for latent variables. This splitting causes inadequate use of latent variables, as latent variables are excluded for some subimages that are important for latent variables. For SHVC-ArIB, this incurs a performance cost compared to utilizing latent variables for all subimages. Moreover, the performance of SHVC-ArIB is constrained by posterior collapse [21]. Posterior collapse refers to the collapse of the posterior to the prior, indicating that non-useful information is extracted by latent variables. A common cause of posterior collapse is highly capable decoders [21]. Specifically, a strong autoregressive prior enhances the decoder's capability, increasing the likelihood of posterior collapse. Consequently, subimages positioned at" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27581" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 269, + 172 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 269, + 172 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 269, + 172 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 269, + 172 + ], + "type": "image", + "image_path": "54b2e46c782b8aeb60659c71d05e33ff7addb2d37bbdfc740b3b26f31ac3e712.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 180, + 289, + 214 + ], + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 214 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 214 + ], + "type": "text", + "content": "Figure 2. Visualization for each bit plane. The planes are arranged from MSP to LSP, from left-top to right-bottom. The volume of global data modality decreases from MSP to LSP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 307 + ], + "type": "text", + "content": "the end of the autoregressive sequence are in a high risk of posterior collapse due to their strong autoregressive priors. In SHVC-ArIB, enhancing spatial autoregression leads to slight improvements or even a decline in performance because of posterior collapse. In this paper, we introduce bit plane slicing (BPS) to address these issues." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": "Bit Plane Slicing is a well-established technique in digital image processing. For a " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": "-bit image, it partitions the entire image into " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": " bit planes. Here, each pixel intensity is represented as " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": ", and the " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "l^{th}" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": " bit is represented as " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "x^l" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": ". The value of " + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "inline_equation", + "content": "x^l" + }, + { + "bbox": [ + 46, + 308, + 287, + 368 + ], + "type": "text", + "content": " can be determined using the following equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 379, + 287, + 395 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 379, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 116, + 379, + 287, + 395 + ], + "type": "interline_equation", + "content": "x ^ {l} = \\left\\lfloor x / 2 ^ {d - l} \\right\\rfloor \\mod 2 \\tag {3}", + "image_path": "2cfca02f1cb6c669cf4bce4fa2f0dbf35029d82985492ed016a2b191bad1a858.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\lfloor x\\rfloor" + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "text", + "content": " denotes the largest integer less than or equal to " + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "text", + "content": ". The original pixel value " + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 400, + 287, + 437 + ], + "type": "text", + "content": " can be reconstructed using the equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 447, + 287, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 127, + 447, + 287, + 479 + ], + "type": "interline_equation", + "content": "x = \\sum_ {l = 1} ^ {d} 2 ^ {d - l} \\times x ^ {l} \\tag {4}", + "image_path": "5566dcfc86500ab947ac4bce5437bceaadbd99cd4939cba021ff5ff0c3bac815.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": "In this paper, we focus specifically on 8-bit images. We define the significant planes as those with smaller " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": " values, and the insignificant planes as those with larger " + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 486, + 287, + 688 + ], + "type": "text", + "content": " values. The volume of global data modality decreases from the most significant plane (MSP) to the least significant plane (LSP), as illustrated in Figure 2. Considering latent variables are employed to learn global data modalities [22, 28], we hypothesize that the importance for latent variables similarly decreases from MSP to LSP. To test this hypothesis, we employ the latent variable model from [34] for each bit plane. Subsequently, we evaluate the Bit Per Dimension (BPD) savings and calculate the bit rate percentage attributed to latent variables within the entire subimage. The results, illustrated in Figure 3, display a consistent descending trend from MSP to LSP, with values becoming negligible in the final few insignificant planes. This observation corroborates our initial hypothesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Based on this finding, we employ BPS for ArIB. We partition the image into two segments using a slice index rang-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": "ing in " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\{1, \\ldots, 8\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ". The first segment comprises the significant planes, represented as " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "x^{1:s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ", while the second segment includes the insignificant planes, denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "x^{s+1:8}" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": ". Latent variables " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " are exclusively used for the significant planes, while ARMs are employed for both segments. This approach enables a more adequate use of latent variables since " + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "inline_equation", + "content": "x^{1:s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 300 + ], + "type": "text", + "content": " is more important for them. Moreover, we introduce plane-by-plane autoregression from MSP to LSP to enhance autoregression. As the volume of autoregressive prior increases along the subimage sequence, the risk of posterior collapse intensifies sequentially. In this autoregression, subimages that are more important for latent variables are placed in positions where they are less likely to encounter posterior collapse. In comparison, SHVC-ArIB split the image into subimages with similar importance, meaning that subimages in positions where there's a high risk of posterior collapse are also important for latent variables. Thus, the overall risk of posterior collapse is mitigated by our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "text", + "content": "The overall bit rate of the image " + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "text", + "content": " can be expressed as follows, where " + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "inline_equation", + "content": "R_{sig}" + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "inline_equation", + "content": "R_{ins}" + }, + { + "bbox": [ + 304, + 300, + 545, + 336 + ], + "type": "text", + "content": " denote the bit rates associated with significant and insignificant planes, respectively:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 338, + 346, + 545, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 346, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 338, + 346, + 545, + 376 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R _ {x} = R _ {s i g} + R _ {i n s} \\\\ = - \\log p \\left(x ^ {1: s} \\mid z\\right) - \\log p \\left(x ^ {s + 1: 8} \\mid x ^ {1: s}\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "b3133b30d82a6047b90eb63bb0e2824d1a0cb3827aff4feccff8709add9c8d91.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 381, + 545, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 381, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 381, + 545, + 404 + ], + "type": "text", + "content": "We further propose some modifications to suit our ArIB with BPS method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "text", + "content": "Hierarchical Latent Variables are commonly utilized in VAEs to improve performance. These variables are structured in a hierarchy comprising " + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "text", + "content": " layers, denoted as " + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "inline_equation", + "content": "z = \\{z_{1},\\dots ,z_{L}\\}" + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "text", + "content": ". The prior probability distribution of " + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 406, + 545, + 464 + ], + "type": "text", + "content": " is factorized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 356, + 476, + 545, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 476, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 356, + 476, + 545, + 509 + ], + "type": "interline_equation", + "content": "p \\left(z _ {1: L}\\right) = p \\left(z _ {L}\\right) \\prod_ {i = 1} ^ {L - 1} p \\left(z _ {i} \\mid z _ {i + 1: L}\\right) \\tag {6}", + "image_path": "fb0411305683bd02547f55e336d6c73928a721fc1c99d34964852e5135d320f9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": "There are two methods for obtaining the posterior " + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "inline_equation", + "content": "q(z|x^{1:s})" + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": ": top-down, which proceeds from layer " + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": " to layer 1, and bottom-up, which goes from layer 1 to layer " + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 515, + 545, + 658 + ], + "type": "text", + "content": ". The former has been shown to exhibit superior compression performance [32, 35], while the latter can leverage the bit-swap technique [19] to reduce the number of initial bits required. Given that the bit rate increases from MSP to LSP [43], insignificant planes can contribute a larger number of initial bits. Consequently, we opt for the top-down inference models, a choice that differentiates our approach from SHVC-ArIB. In this context, the posterior distribution can be expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 678, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 678, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 321, + 678, + 545, + 711 + ], + "type": "interline_equation", + "content": "q \\left(z _ {1: L} \\mid x ^ {1: s}\\right) = q \\left(z _ {L} \\mid x ^ {1: s}\\right) \\prod_ {i = 1} ^ {L - 1} q \\left(z _ {i} \\mid z _ {i + 1: L}, x ^ {1: s}\\right) \\tag {7}", + "image_path": "404835a623f423b7981dc0e85c9643be3eab366eb490c92b80c940daed03a896.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27582" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 68, + 273, + 277 + ], + "blocks": [ + { + "bbox": [ + 62, + 68, + 273, + 277 + ], + "lines": [ + { + "bbox": [ + 62, + 68, + 273, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 68, + 273, + 277 + ], + "type": "image", + "image_path": "0a9ad64b720dd402783f67b9793b19c96f95a77337ebbb9728f28cf3b5755bb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 285, + 289, + 340 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 289, + 340 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 289, + 340 + ], + "type": "text", + "content": "Figure 3. BPD savings achieved using latent variables and the bit rate percentage attributed to latent variables for each bit plane. The results are averaged over the CIFAR10, ImageNet32, and ImageNet64 [6] datasets. The importance for latent variables decreases from MSP to LSP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 422 + ], + "type": "text", + "content": "Discretization also plays a pivotal role: a higher precision tends to enhance compression performance but also requires more initial bits. Given the variation in available initial bits across different images, we propose an image-adaptive discretization strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": "In our method, we utilize equal mass discretization [35] for the top-down model. This ensures that all latent variables maintain equal probability mass for " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": ". We define the discretization precision " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": ", which discretizes " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "2^{k}" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " possible values. It is important to note that " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "q(z|x)" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " closely approximates " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": ", as the Kullback-Leibler divergence between them is minimized during optimization. As a result, each element of " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " necessitates roughly " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "-\\log (1 / (2^k)) = k" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " initial bits. Furthermore, the size of the latent variables, denoted as " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "|z|" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": ", can be quantified. The bit rate " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "R_{ins}" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " can be determined by measuring the length of the encoded bitstream. Following these calculations, the precision " + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 422, + 288, + 578 + ], + "type": "text", + "content": " is derived using the equation:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 586, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 586, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 140, + 586, + 287, + 613 + ], + "type": "interline_equation", + "content": "k = \\left\\lfloor \\frac {R _ {i n s}}{| z |} \\right\\rfloor \\tag {8}", + "image_path": "b59dd75c4e4c0a319bc46726c3c676918c34b1170fe676df5eaa4581c8d9d846.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "For training VAEs, the bit rate of latent variables is calculated with the condition that they are discretized to high precision. In practical scenarios, most images exhibit a sufficiently large precision value " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ", ensuring negligible impact on compression performance when employing discretization. Nevertheless, low precision is required in minor cases, where the discretization hurts the compression performance. To tackle this, we propose a discretized" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "sampling technique for sampling latent variables for finetuning. More specifically, both the prior and posterior distributions are modeled using a univariate logistic distribution, characterized by mean parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mu_p" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mu_q" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", respectively. We sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " from a standard logistic distribution. The calculation of a sampled latent variable " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and its corresponding bit rate can be represented as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 174, + 545, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 174, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 323, + 174, + 545, + 273 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} t _ {l} = \\left\\lceil 2 ^ {k} \\cdot \\operatorname {s i g m o i d} \\left(\\epsilon + \\mu_ {q} - \\mu_ {p}\\right)\\right\\rceil - 1, \\\\ \\hat {z} = \\operatorname {l o g i t} \\left(\\frac {t _ {l} + 0 . 5}{2 ^ {k}}\\right) + \\mu_ {p}, \\\\ R _ {\\hat {z}} = \\log (\\operatorname {s i g m o i d} (\\operatorname {l o g i t} \\left(\\frac {t _ {l} + 1}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q}) \\tag {9} \\\\ - \\operatorname {s i g m o i d} (\\log \\mathrm {i t} \\left(\\frac {t _ {l}}{2 ^ {k}}\\right) + \\mu_ {p} - \\mu_ {q})) + k, \\\\ \\end{array}", + "image_path": "a59234ef91402882c3409a813e4a571cc04a58cb39f0e987a29fc6a8a1ff72b1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "inline_equation", + "content": "\\lceil x\\rceil" + }, + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "text", + "content": " signifies the smallest integer greater than or equal to " + }, + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 281, + 545, + 318 + ], + "type": "text", + "content": ". During back-propagation, the Straight Through Estimator [3] is utilized to manage this operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 324, + 525, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 525, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 525, + 337 + ], + "type": "text", + "content": "3.3. Dimension-Tailored Autoregressive Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 557 + ], + "type": "text", + "content": "BPS results in an increased number of dimensions for autoregression, specifically, the bit plane dimension. In addition to the plane dimension, we consider autoregression in space and color dimensions, which have demonstrated their efficacy in previous studies [26, 40]. Using a uniform approach to model all dimensions is employed in previous works [14]. However, this method has certain limitations. First, these dimensions present uneven modeling challenges. Employing a one-size-fits-all network to address all dimensions directly may either compromise modeling accuracy in the more complex dimensions or lead to inefficient use of computational resources in simpler dimensions. Second, strong autoregression in the space and color dimensions can potentially lead to posterior collapse. To address these issues, we propose a dimension-tailored autoregressive model that tailors the approach for each dimension based on its specific characteristics. An overview is shown in the right subfigures of Figure 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": "Our autoregressive model operates in three stages. In the first stage, we focus on the plane dimension, which presents a significant challenge due to the extensive range of possible values that the pixels in its preceding planes, denoted as " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "x^{1:i - 1}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", can take on, amounting to " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "2^{i - 1}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " possibilities. To effectively leverage the autoregressive prior from " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "x^{1:i - 1}" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": " and apply it to " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "x^i" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ", we employ a sophisticated plane context model, " + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 304, + 558, + 545, + 714 + ], + "type": "text", + "content": ". To capture the global data modality embedded in the preceding planes, we adopt a U-Net architecture [27]. Specifically, we use two models that employ the architecture of diffusion models [7]: one for the significant planes and the other for the insignificant planes. Time embedding is used to differentiate planes." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27583" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": "In the second stage, we address the space dimension. Compared to the plane dimension, this dimension is less complex, as each bit in the current bit plane can only take one of two possible values. We adopt a 4-stage checkerboard model [9, 26], which involves dividing the bit plane " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "x^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " into four sub-planes, denoted as " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "x_{1}^{i}, x_{2}^{i}, x_{3}^{i}, x_{4}^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": ", as illustrated in Figure 1. For each sub-plane " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "x_{j}^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": ", we utilize a simple network, " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Phi_{s}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": ", to extract the spatial autoregressive prior " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "x_{1:j-1}^{i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": ". Then, a parameter network " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " is used to fuse information from both the plane and space dimensions. To effectively integrate information, we employ a relatively complex network for " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": ". Although the complexity of " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " does contribute to the overall computational demand of our model, the complexity remains significantly lower than that of employing the architecture of " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Phi_{p}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " for both dimensions, resulting in a more efficient modeling process. More specifically, " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Phi_{s}" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " is a single ResBlock [11], and " + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 72, + 289, + 286 + ], + "type": "text", + "content": " combines an ECANet [39] with several wider ResBlocks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "text", + "content": "The third stage is for the color dimension. As identified in PixelCNN++ [29], the correlations among the color channels of a pixel are typically straightforward, and deep networks are often unnecessary for their modeling. Therefore, we adopt a non-network approach for the color dimension. Specifically, we model the distribution of bits in each color channel using a Bernoulli distribution, where the probability parameter " + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "inline_equation", + "content": "p = P(x = 1)" + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "text", + "content": " is determined autoregressively based on the bits with the same spatial position in previously decoded channels. In our method, " + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "text", + "content": " predict six parameters " + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "inline_equation", + "content": "\\theta = \\{f_r, f_g, f_b, \\alpha, \\beta, \\gamma\\}" + }, + { + "bbox": [ + 47, + 287, + 289, + 430 + ], + "type": "text", + "content": ", and model the probability for each channel as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 434, + 169, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 434, + 169, + 447 + ], + "spans": [ + { + "bbox": [ + 93, + 434, + 169, + 447 + ], + "type": "interline_equation", + "content": "p _ {r} = \\operatorname {s i g m o i d} \\left(f _ {r}\\right)", + "image_path": "1cd7a4b0eaed77c768143f00d4ad6a928204d8d8f52d347265653e282c22bd5f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 449, + 287, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 449, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 94, + 449, + 287, + 462 + ], + "type": "interline_equation", + "content": "p _ {g} = \\operatorname {s i g m o i d} \\left(f _ {g} + \\alpha \\cdot x _ {r}\\right) \\tag {10}", + "image_path": "c37f144f098bb75b42fae8fcc20d02c827f6aae3d20ecec38df13532249ea695.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 464, + 240, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 464, + 240, + 477 + ], + "spans": [ + { + "bbox": [ + 94, + 464, + 240, + 477 + ], + "type": "interline_equation", + "content": "p _ {b} = \\operatorname {s i g m o i d} \\left(f _ {b} + \\beta \\cdot x _ {r} + \\gamma \\cdot x _ {g}\\right)", + "image_path": "231d7858e2c873b6afd61cb278f2159d61c42cec294e77ee74029244f5b27ab0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 481, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 287, + 505 + ], + "type": "text", + "content": "The 1st and 2nd stages of our dimension-tailored autoregressive model can be expressed as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 509, + 173, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 509, + 173, + 525 + ], + "spans": [ + { + "bbox": [ + 97, + 509, + 173, + 525 + ], + "type": "interline_equation", + "content": "F _ {p ^ {i}} = \\Phi_ {p} (x ^ {1: i - 1})", + "image_path": "c9b114c3a2599e43ff397d67a869da9e0dc54d1978cee2109539f73f77091e53.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 99, + 526, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 526, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 99, + 526, + 287, + 548 + ], + "type": "interline_equation", + "content": "F _ {s _ {j} ^ {i}} = \\Phi_ {s} \\left(x _ {1: j - 1} ^ {i}\\right) \\tag {11}", + "image_path": "52d7371b8a360f5f1563b0eab4763b1c302dda3603dfffab159e86d04071192f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 544, + 235, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 544, + 235, + 578 + ], + "spans": [ + { + "bbox": [ + 104, + 544, + 235, + 578 + ], + "type": "interline_equation", + "content": "\\theta_ {j} ^ {i} = \\left\\{ \\begin{array}{l l} \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}, F _ {z}), & i \\leq s \\\\ \\Theta (F _ {p ^ {i}}, F _ {s _ {j} ^ {i}}), & i > s \\end{array} \\right.", + "image_path": "65bb1659c18a8201164885bc0fd865ce3f05706b71ec51b8fe783b0caddd0f38.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": "In summary, we initially derive " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": " for the entire bit plane. Subsequently, for each of the four subimages in the bit plane, we compute " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "F_{s}" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": ". These are then concatenated with " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "F_{z}" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": " (only for significant planes) and fed into " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": " to determine the six parameters " + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 582, + 287, + 665 + ], + "type": "text", + "content": ". Finally, we calculate the probability using non-network color autoregression as outlined in Equation 10." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 666, + 231, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 231, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 231, + 677 + ], + "type": "text", + "content": "Loss function of our method is expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 686, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 686, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 91, + 686, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} l = R _ {z} + R _ {x} \\tag {12} \\\\ = - \\log p (z) + \\log q (z | x ^ {1: s}) + R _ {x} \\\\ \\end{array}", + "image_path": "faa0ba02118c58973d4f4682d9bad014af48cdc4a97ec5ad6ca4cf04816ce785.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 91, + 487, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 487, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 487, + 104 + ], + "type": "text", + "content": "4.1. Architecture and Training Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "content": "In accordance with prior research, our experimental setup entails training three models using three benchmark datasets: CIFAR10, ImageNet32, and ImageNet64 [6]. For all models, we set the split index, " + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "content": ", to 4, based on Figure 3, and utilize 3 layers of hierarchical latent variables. For latent variables, we apply a factor 2 downsampling for each hierarchy of latent variables in latent encoders. The channel number for latent variables is set to 3 across all models. For " + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "content": ", the number of hierarchical levels in the U-Net is set to 4 across all models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 230, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 301 + ], + "type": "text", + "content": "We utilize the Adam optimizer [18] with an initial learning rate of 0.0001, and we apply step decay as the learning rate schedule. For fine-tuning with discretized sampling, we employ Equation 8 with the previously trained model to obtain the precision " + }, + { + "bbox": [ + 304, + 230, + 545, + 301 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 230, + 545, + 301 + ], + "type": "text", + "content": " for each image in the training set. All experiments are conducted on a GeForce GTX 1080 Ti." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 311, + 453, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 311, + 453, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 311, + 453, + 323 + ], + "type": "text", + "content": "4.2. Compression Performance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 330, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 425 + ], + "type": "text", + "content": "We evaluate the performance of our method across four distinct datasets: CIFAR10, ImageNet32, ImageNet64, and DIV2K [1], using bits per dimension (BPD) as our primary metric. The first three datasets are of low resolution, while the last one, DIV2K, is a high-resolution dataset. Following most previous works, we exclude per-pixel autoregressive methods [29, 37] and diffusion models [14, 17] from our evaluation due to their impractical time complexity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": "Our method evaluates performance in two distinct compression settings: dataset compression and single-image compression. In dataset compression, we sequentially compress 5,000 images with a fixed discretization precision of 10. For single-image compression, each image's " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " value is calculated using Equation 8 and clipped between 3 and 10. If the calculated " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " is less than 3, additional " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "(3 - k) \\times |z|" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "|z|" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " represents the size of " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": ") bits are added to ensure successful encoding. These possible bits, along with the " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " value and image bits, are written to a bitstream. The single image compression performance is determined by measuring the bitstream size. Following [28, 41, 42], we evaluate the performance on the DIV2K dataset using the model trained on ImageNet64 by cropping images into " + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 426, + 545, + 594 + ], + "type": "text", + "content": " patches." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We compare our method against advanced approaches from traditional methods, learned methods without bits back coding, and learned methods with bits back coding. We present the detailed results of our experiments in Table 1. Our method demonstrates superior compression performance, particularly in single-image compression. In dataset compression, our method excels in three out of four datasets, achieving the second-best performance on the remaining dataset. In single-image compression, our method outperforms other methods in all datasets. We note" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27584" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 123, + 523, + 327 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 118 + ], + "type": "text", + "content": "Table 1. Compression performance in BPD for four datasets, where lower values indicate better performance. The best performance is highlighted in bold. *: ImageNet32, **: ImageNet64, +: Flickr2k [33], -: OpenImage [20]. 1st Group: traditional methods, 2nd Group: learned methods without bits back coding, 3rd Group: learned methods with bits back coding. Some values of SHVC and SHVC-ArIB are not available, and we list their available lower bound here. △: These values are obtained by adding dataset performance and initial bits." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 123, + 523, + 327 + ], + "lines": [ + { + "bbox": [ + 74, + 123, + 523, + 327 + ], + "spans": [ + { + "bbox": [ + 74, + 123, + 523, + 327 + ], + "type": "table", + "html": "
Dataset CompressionSingle-Image Compression
MethodCIFAR10ImageNet32ImageNet64DIV2KCIFAR10ImageNet32ImageNet64DIV2K
PNG [5]5.896.395.744.235.896.395.744.23
FLIF [31]4.194.524.542.914.194.524.542.91
JPEG-XL [2]5.746.395.892.795.746.395.892.79
L3C [22]-4.764.423.09--4.764.423.09-
IDF [13]3.344.183.90-3.344.183.90-
IDF++ [36]3.264.123.81-3.264.123.81-
LC-FDNet [26]3.77+4.40+4.28+2.71+3.77+4.40+4.28+2.71+
Bit-Swap [19]3.824.50--6.536.97--
Hilloc [35]3.56*4.203.90*-----
SHVC [28]3.163.983.682.57**>3.96>4.78>4.48-
SHVC-ArIB [28]>3.16>3.98>3.68->3.16>3.98>3.68-
iVPF [42]3.204.033.752.68**9.20Δ---
LBB [12]3.123.883.70-42.98Δ49.84Δ41.70Δ-
iFlow [41]3.123.883.702.57**37.40Δ38.27Δ38.12Δ-
ArIB-BPS (ours)3.063.913.632.55**3.073.923.642.59**
", + "image_path": "cb3a9fab59f2e27cf96619c81feb39be1f99b3b9341136662b78674da68c9a07.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 366, + 488, + 518 + ], + "blocks": [ + { + "bbox": [ + 46, + 335, + 546, + 360 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 546, + 360 + ], + "type": "text", + "content": "Table 2. Comparison with advanced and available flow-based methods. The batchsize is 64 following LBB and iFlow. We train the small model to align with the inference speed of iFlow on ImageNet64." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 366, + 488, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 488, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 488, + 518 + ], + "type": "table", + "html": "
Compression Performance (BPD) ↓Inference Time (ms/sample) ↓
datasetMethodtheoreticaldatasetsingleencodedecode
CIFAR10LBB [12]3.1163.11842.97864.9464.94
iFlow [41]3.1163.11837.39817.3847.56
ArIB-BPS3.0483.0573.07014.9314.93
ImageNet32LBB [12]3.8713.87549.835194.14194.14
iFlow [41]3.8713.87338.27374.84119.30
ArIB-BPS3.9043.9113.91852.9652.96
ImageNet64LBB [12]3.7013.70341.70395.5795.57
iFlow [41]3.7013.70338.12337.4958.08
ArIB-BPS (Small)3.6773.6853.69343.6543.65
ArIB-BPS3.6243.6333.641107.80107.80
", + "image_path": "dfad600da2958c9936e67386406f6693ec84146be756c0c495aaf4d5c78ae788.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 537, + 288, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 288, + 622 + ], + "type": "text", + "content": "that precise values for the single-image compression performance of SHVC, as well as the compression performance of SHVC-ArIB, are not available. However, the lower bound could be obtained. Moreover, it could be inferred that our method surpasses SHVC-ArIB by an average of more than 0.10 BPD on low-resolution datasets in both compression settings. Additional details are provided in the Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of our approach, we conducted a detailed comparison between our method and continuous normalizing-flow-based methods, which exhibit state-of-the-art performance in dataset compression. As Table 2 illustrates, ArIB-BPS outperforms iFlow in both single-image and dataset compression performance for CIFAR10 and ImageNet64, with comparable inference time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 598 + ], + "type": "text", + "content": "For ImageNet32, although ArIB-BPS lags slightly behind in dataset compression performance, it excels in other metrics, particularly in single-image compression. These results imply that our method achieves a better trade-off between compression performance and inference speed." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 609, + 408, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 408, + 621 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 408, + 621 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "We conduct three ablation studies on the CIFAR10 dataset to show the efficiency of our dimension-tailored autoregressive model and the effectiveness of discretized sampling. We note that results in Table 3 and 4 are evaluated using models without using discretized sampling for fine-tuning. Therefore, the results of ArIB-BPS are slightly different from those in Table 2." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27585" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 113, + 269, + 184 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 104 + ], + "type": "text", + "content": "Table 3. Ablation for the effectiveness of autoregression on plane, space, and color dimensions. Inference time is identical for encoding and decoding. The batchsize is 64." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 113, + 269, + 184 + ], + "lines": [ + { + "bbox": [ + 65, + 113, + 269, + 184 + ], + "spans": [ + { + "bbox": [ + 65, + 113, + 269, + 184 + ], + "type": "table", + "html": "
planespacecolorTheoretical BPD ↓Inference Time (ms/sample) ↓
✓✓✓3.05214.93
✓✓3.45814.92
✓✓3.47010.36
✓✓3.5225.83
", + "image_path": "34109f5638c6e4f64eec5b20fa8484c6f4dd8b878de149f6c32bf993b67f1b06.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 251, + 287, + 345 + ], + "blocks": [ + { + "bbox": [ + 47, + 197, + 288, + 242 + ], + "lines": [ + { + "bbox": [ + 47, + 197, + 288, + 242 + ], + "spans": [ + { + "bbox": [ + 47, + 197, + 288, + 242 + ], + "type": "text", + "content": "Table 4. Compression performance and inference time of various usages of the architecture of " + }, + { + "bbox": [ + 47, + 197, + 288, + 242 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 47, + 197, + 288, + 242 + ], + "type": "text", + "content": ". The batch size is 64. P: plane. S: space. C: color. Sig: significant planes. Ins: insignificant planes. We also train two small models to further validate our design." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 251, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 50, + 251, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 50, + 251, + 287, + 345 + ], + "type": "table", + "html": "
Theoretical BPD ↓Inference Time (ms/sample) ↓
SigInsTotalSigInsTotal
P (Ours)1.0412.0113.0526.308.6414.93
PS1.0521.9933.04514.4624.9739.42
PSC1.0431.9713.01543.8374.97118.80
PS (Small)1.0552.0703.1259.0811.4420.51
PSC (Small)1.0962.2183.3149.6413.9623.59
", + "image_path": "9e26ebc08a0e06234e02448b6c53ee47df608b76896b6408650c225372958997.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 368, + 259, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 368, + 259, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 259, + 380 + ], + "type": "text", + "content": "4.3.1 Dimension-Tailored Autoregressive Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "text", + "content": "Firstly, we conduct experiments to assess the impact of autoregression in each dimension. To do this, we train and evaluate three additional models, each with autoregression removed in either the color, space, or plane dimension. In the model that does not incorporate plane autoregression, we employ the color autoregressive model from PixelCNN++ [29] to ensure compatibility. Additionally, we enhance " + }, + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Phi_s" + }, + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "text", + "content": " by employing the architecture of " + }, + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 46, + 389, + 287, + 581 + ], + "type": "text", + "content": ", accommodating the increased complexity in the spatial dimension, which now has 256 possible values per pixel instead of just 2. In our experiments, this configuration results in improved performance. Table 3 clearly demonstrates that incorporating autoregression in each dimension consistently leads to significant improvements in compression performance. We note that our autoregression for the color dimension incurs only a negligible overhead in inference time." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Furthermore, we conduct an evaluation to assess the efficiency of our dimension-tailored design. Specifically, we train two additional models: one using the architecture of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " for both the plane and space dimensions and the other using the architecture of " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " for all three dimensions. Subsequently, we evaluate the impact of these modifications on significant and insignificant planes separately. As depicted in Table 4, the dimension-tailored autoregressive model achieves superior performance in significant planes with lower complexity. This superiority is attributed to reduced hindrance to the latent variables, facilitated by sim" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 536, + 203 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 536, + 203 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 536, + 203 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 536, + 203 + ], + "type": "image", + "image_path": "2ed6e446a7050d348666012dc92dd0c28350ffa96cecb3c79bc4d103d508fdaf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 211, + 545, + 234 + ], + "type": "text", + "content": "Figure 4. Average BPD savings through fine-tuning using discretized sampling across various discretization precisions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 254, + 545, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 545, + 362 + ], + "type": "text", + "content": "pler spatial autoregression and color-wise autoregression, as reflected in higher bit rates for latent variables. In the case of insignificant planes, although strengthened autoregression can lead to performance gains, it results in significantly longer inference time. To further validate our method's efficiency, we train two small models with reduced inference time. Our method demonstrates better compression performance. Considering the trade-off between inference time and compression performance, our design is more efficient." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 376, + 430, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 430, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 430, + 389 + ], + "type": "text", + "content": "4.3.2 Discretized Sampling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 395, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 395, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 395, + 545, + 468 + ], + "type": "text", + "content": "We assess the performance improvement after fine-tuning across various " + }, + { + "bbox": [ + 304, + 395, + 545, + 468 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 395, + 545, + 468 + ], + "type": "text", + "content": " values, as illustrated in Figure 4. The method improves performance across all precision values, particularly at lower precision levels. While low-precision cases form a minor segment, their optimization boosts performance in edge cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 378, + 490 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": "We have proposed ArIB-BPS, a new approach for lossless image compression. BPS enables us to achieve adequate use of latent variables as well as enhanced autoregression in a mitigated risk of posterior collapse by arranging subimages with decreasing importance for latent variables. In addition, we present a dimension-tailored autoregressive model that takes into account the different characteristics of each dimension. It efficiently captures dependencies in space, color, and plane dimensions. Experimental results demonstrate that our method achieves superior compression performance compared to the state-of-the-art normalizing-flow-based methods with comparable inference time." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "The limitation of our work lies in the complexity of higher bit-depth. The complexity of our method would increase as the bit-depth increases due to the interdependency among bit planes. To extend our method for high-bit-depth images, efficient strategies such as handling multiple planes in one step should be explored in our future work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27586" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 203 + ], + "type": "text", + "content": "[2] Jyrki Alakuijala, Ruud Van Asseldonk, Sami Boukortt, Martin Bruse, Iulia-Maria Comsa, Moritz Firsching, Thomas Fischbacher, Evgenii Kliuchnikov, Sebastian Gomez, Robert Obryk, et al. Jpeg xl next-generation image compression architecture and coding tools. In Applications of Digital Image Processing XLII, pages 112-124. SPIE, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 247 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 288, + 304 + ], + "type": "text", + "content": "[4] Sam Bond-Taylor, Adam Leach, Yang Long, and Chris George Willcocks. Deep generative modelling: A comparative review of vaes, gans, normalizing flows, energy-based and autoregressive models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 304, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 304, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 304, + 288, + 327 + ], + "type": "text", + "content": "[5] Thomas Boutell. PNG (portable network graphics) specification version 1.0. Technical report, 1997. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 328, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 288, + 361 + ], + "type": "text", + "content": "[6] Patryk Chrabaszcz, Ilya Loshchilov, and Frank Hutter. A downsampled variant of imagenet as an alternative to the cifar datasets. arXiv preprint arXiv:1707.08819, 2017. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 396 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34:8780-8794, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 288, + 418 + ], + "type": "text", + "content": "[8] Jarek Duda. Asymmetric numeral systems. arXiv preprint arXiv:0902.0271, 2009. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 421, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 288, + 475 + ], + "type": "text", + "content": "[9] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 542 + ], + "type": "text", + "content": "[10] Dailan He, Ziming Yang, Weikun Peng, Rui Ma, Hongwei Qin, and Yan Wang. Elic: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5718-5727, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[12] Jonathan Ho, Evan Lohn, and Pieter Abbeel. Compression with flows via local bits-back coding. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[13] Emiel Hoogeboom, Jorn Peters, Rianne Van Den Berg, and Max Welling. Integer discrete flows and lossless compression. Advances in Neural Information Processing Systems, 32, 2019. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[14] Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, Rianne van den Berg, and Tim Salimans. Autoregressive diffusion models. In International Conference on Learning Representations, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[15] Seungmin Jeon, Kwang Pyo Choi, Youngo Park, and Chang-Su Kim. Context-based trit-plane coding for progressive image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14348-14357, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 184 + ], + "type": "text", + "content": "[16] Ning Kang, Shanzhao Qiu, Shifeng Zhang, Zhenguo Li, and Shu-Tao Xia. Pilc: Practical image lossless compression with an end-to-endgpu oriented neural framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3739-3748, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 217 + ], + "type": "text", + "content": "[17] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in Neural Information Processing Systems, 34:21696-21707, 2021. 1, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 297 + ], + "type": "text", + "content": "[19] Friso Kingma, Pieter Abbeel, and Jonathan Ho. Bit-swap: Recursive bits-back coding for lossless compression with hierarchical latent variables. In International Conference on Machine Learning, pages 3408-3417, 2019. 1, 2, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 298, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 386 + ], + "type": "text", + "content": "[20] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[21] James Lucas, George Tucker, Roger B Grosse, and Mohammad Norouzi. Don't blame the elbo! a linear vae perspective on posterior collapse. Advances in Neural Information Processing Systems, 32, 2019. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "text", + "content": "[22] Fabian Mentzer, Eirikur Agustsson, Michael Tschannen, Radu Timofte, and Luc Van Gool. Practical full resolution learned lossless image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10629-10638, 2019. 1, 2, 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 542 + ], + "type": "text", + "content": "[23] Fabian Mentzer, Luc Van Gool, and Michael Tschannen. Learning better lossless compression using lossy compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6638-6647, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 588 + ], + "type": "text", + "content": "[24] David Minnen and Saurabh Singh. Channel-wise autoregressive entropy models for learned image compression. In 2020 IEEE International Conference on Image Processing, pages 3339-3343. IEEE, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 633 + ], + "type": "text", + "content": "[25] David Minnen, Johannes Balle, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[26] Hochang Rhee, Yeong Il Jang, Seyun Kim, and Nam Ik Cho. Lc-fdnet: Learned lossless image compression with frequency decomposition network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6033-6042, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmen" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27587" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "tation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 234-241. Springer, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 151 + ], + "type": "text", + "content": "[28] Tom Ryder, Chen Zhang, Ning Kang, and Shifeng Zhang. Split hierarchical variational compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 386-395, 2022. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 206 + ], + "type": "text", + "content": "[29] Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P. Kingma. PixelCNN++: Improving the pixelCNN with discretized logistic mixture likelihood and other modifications. In International Conference on Learning Representations, 2017. 1, 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 288, + 239 + ], + "type": "text", + "content": "[30] Claude Elwood Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27(3):379-423, 1948. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 288, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 288, + 283 + ], + "type": "text", + "content": "[31] Jon Sneyers and Pieter Wuille. Flif: Free lossless image format based on maniac compression. In IEEE International Conference on Image Processing, pages 66-70. IEEE, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 288, + 330 + ], + "type": "text", + "content": "[32] Casper Kaae Sønderby, Tapani Raiko, Lars Maaløe, Søren Kaae Sønderby, and Ole Winther. Ladder variational autoencoders. Advances in Neural Information Processing Systems, 29, 2016. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 387 + ], + "type": "text", + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 114-125, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 431 + ], + "type": "text", + "content": "[34] James Townsend, Thomas Bird, and David Barber. Practical lossless compression with latent variables using bits back coding. In International Conference on Learning Representations, 2018. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 477 + ], + "type": "text", + "content": "[35] James Townsend, Thomas Bird, Julius Kunze, and David Barber. Hilloc: lossless image compression with hierarchical latent variable models. In International Conference on Learning Representations, 2019. 1, 2, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 288, + 533 + ], + "type": "text", + "content": "[36] Rianne van den Berg, Alexey A Gritsenko, Mostafa Dehghani, Casper Kaae Sønderby, and Tim Salimans. Idf++: Analyzing and improving integer discrete flows for lossless compression. In International Conference on Learning Representations, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 578 + ], + "type": "text", + "content": "[37] Aäron van den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In Proceedings of The 33rd International Conference on Machine Learning, pages 1747-1756, 2016. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[38] Kai Wang, Yuanchao Bai, Deming Zhai, Daxin Li, Junjun Jiang, and Xianming Liu. Learning lossless compression for high bit-depth medical imaging. In 2023 IEEE International Conference on Multimedia and Expo, pages 2549-2554. IEEE, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[39] Qilong Wang, Banggu Wu, Pengfei Zhu, Peihua Li, Wangmeng Zuo, and Qinghua Hu. Eca-net: Efficient channel attention for deep convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11534-11542, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[40] Honglei Zhang, Francesco Cricri, Hamed R Tavakoli, Nannan Zou, Emre Aksu, and Miska M Hannuksela. Lossless" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 272 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "image compression using a multi-scale progressive statistical model. In Proceedings of the Asian Conference on Computer Vision, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[41] Shifeng Zhang, Ning Kang, Tom Ryder, and Zhenguo Li. iflow: Numerically invertible flows for efficient lossless compression via a uniform coder. Advances in Neural Information Processing Systems, 34:5822-5833, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[42] Shifeng Zhang, Chen Zhang, Ning Kang, and Zhenguo Li. ivpf: Numerical invertible volume preserving flow for efficient lossless compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 620-629, 2021. 1, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 272 + ], + "type": "text", + "content": "[43] Zhizheng Zhang, Zhibo Chen, Jianxin Lin, and Weiping Li. Learned scalable image compression with bidirectional context disentanglement network. In IEEE International Conference on Multimedia and Expo, pages 1438-1443. IEEE, 2019. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27588" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_content_list.json b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..03f3b7c250de6b57822c3597f841d09277064699 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_content_list.json @@ -0,0 +1,1640 @@ +[ + { + "type": "text", + "text": "Learned representation-guided diffusion models for large-image generation", + "text_level": 1, + "bbox": [ + 106, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexandros Graikos* Srikar Yellapragada* Minh-Quan Le Saarthak Kapse Prateek Prasanna Joel Saltz Dimitris Samaras Stony Brook University", + "bbox": [ + 220, + 180, + 767, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To synthesize high-fidelity samples, diffusion models typically require auxiliary data to guide the generation process. However, it is impractical to procure the painstaking patch-level annotation effort required in specialized domains like histopathology and satellite imagery; it is often performed by domain experts and involves hundreds of millions of patches. Modern-day self-supervised learning (SSL) representations encode rich semantic and visual information. In this paper, we posit that such representations are expressive enough to act as proxies to fine-grained human labels. We introduce a novel approach that trains diffusion models conditioned on embeddings from SSL. Our diffusion models successfully project these features back to high-quality histopathology and remote sensing images. In addition, we construct larger images by assembling spatially consistent patches inferred from SSL embeddings, preserving long-range dependencies. Augmenting real data by generating variations of real images improves downstream classifier accuracy for patch-level and larger, image-scale classification tasks. Our models are effective even on datasets not encountered during training, demonstrating their robustness and generalizability. Generating images from learned embeddings is agnostic to the source of the embeddings. The SSL embeddings used to generate a large image can either be extracted from a reference image, or sampled from an auxiliary model conditioned on any related modality (e.g. class labels, text, genomic data). As proof of concept, we introduce the text-to-large image synthesis paradigm where we successfully synthesize large pathology and satellite images out of text descriptions.", + "bbox": [ + 75, + 300, + 473, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 781, + 209, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models produce high-quality and diverse samples across a spectrum of generative tasks [8, 24]. This leap forward has been enabled by the simultaneous curation of large-scale multi-modal datasets [40] and the development", + "bbox": [ + 75, + 806, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/26a83e1423cf49b6cdaa97b008ab8b1a43ea216117eed332430fd6ae51c85da4.jpg", + "image_caption": [ + "Figure 1. We propose using SSL features to condition diffusion models. This allows us to construct large images by assembling consistent patches inferred from a spatial arrangement of SSL embeddings. The generated image retains the semantics of the embeddings used as a condition, maintaining the forested and open areas from the reference. Best viewed zoomed-in." + ], + "image_footnote": [], + "bbox": [ + 501, + 266, + 893, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of efficient conditioning mechanisms [35, 37]. The key to unlocking the models' capabilities is to integrate auxiliary information during training and inference [8, 15, 28].", + "bbox": [ + 496, + 534, + 892, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale human-annotated datasets are mostly limited to image-caption pairs, collected from easily accessible online repositories and labeled by non-expert annotators. However, in domains such as digital histopathology and remote sensing, where gigapixel scale images provide vast amounts of unlabeled data, annotation proves challenging. Moreover, the process requires expert knowledge and is more difficult at a finer scale, i.e., captioning large crops of the gigapixel image is simpler than captioning smaller patches. Based on our estimates (see supplemental), annotating the entire TCGA-BRCA dataset with captions would take $\\approx 40,000$ hours of pathologist's time. Replicating the impressive results of diffusion models in these domains has been limited by the scarcity of fine-grained per-image conditioning, vital for high-quality image synthesis [29].", + "bbox": [ + 496, + 582, + 893, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern-day self-supervised learning (SSL) representations [5] encode rich semantic and visual information. Features from trained self-supervised models serve as compact image representations and are widely used to perform discriminative downstream tasks successfully [6, 11, 45], proving that these compressed representations indeed encode", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. Correspondence to agraikos@cs.stonybrook.edu", + "bbox": [ + 93, + 875, + 462, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code is available at this link", + "bbox": [ + 98, + 887, + 251, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8532", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "useful semantic information about the images. We hypothesize that such SSL representations are already expressive enough to act as proxies to fine-grained human labels. If this is true, these representations should be able to condition the training of effective diffusion models in these domains. In this novel approach, we utilize self-supervised feature extractors as image annotators; these features provide necessary per-image conditioning signals at the highest resolutions, required for diffusion model training.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments show that conditioning with expressive self-supervised features leads to precise control over the image content. The SSL features are adept at identifying complex patterns and structures in the images, while the diffusion model learns to translate them into visual components accurately (Fig. 3). This motivates us to perform large-image synthesis by locally controlling the appearance using SSL conditioning and dictating the global structure through the spatial arrangement of the conditions.", + "bbox": [ + 75, + 229, + 472, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach synthesizes large images in a patch-based manner, using a single image diffusion model at the highest resolution. We represent a large image as a grid of SSL embeddings, where each serves as a representation of a large image neighborhood. The whole image is then synthesized by generating consistent patches that capture both the local properties, as given by the local patch conditioning, and the spatial arrangement of the conditioning features. If we change this spatial arrangement, we are, in fact, editing how semantic elements are arranged globally in the large image. This strategy enables the controllable generation of images of virtually any size without significantly increased computation compared to the base patch-level model.", + "bbox": [ + 75, + 367, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate a large image, our approach requires the patch diffusion model and the spatially-arranged conditioning. We can start with a reference large image as source and extract SSL embeddings from non-overlapping segments, enabling our method to synthesize a variation of the original image (Fig. 1).", + "bbox": [ + 75, + 564, + 472, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utilizing SSL embeddings as conditions allows us to have the necessary control over image generation, at the expense of an explainable and easy-to-use conditioning mechanism. Nevertheless, we argue that since generating images from learned embeddings is agnostic to the embedding source, there are simple ways to combine control over generated images with explainability. We propose training auxiliary models to transform higher-level conditioning signals, such as text captions, to the learned patch representations. To demonstrate this versatility, we introduce text-to-large-image synthesis by training an auxiliary model to sample a spatial arrangement of embeddings from a text description.", + "bbox": [ + 75, + 657, + 472, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We train patch-level diffusion models using self-supervised features as conditioning on digital histopathology (TCGA [4]) and satellite image (NAIP [43]) datasets. We perform extensive evaluations and demonstrate the ad", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vantages of SSL conditioning and our large-image generation framework on synthesis and classification tasks. Our model achieves exceptional patch-level and large-image quality, the ability to improve classifiers through data augmentation even when synthesizing out-of-distribution data, and effective fusion of diffusion and SSL features for downstream applications. Finally, we are the first to perform text-to-large image synthesis, which should be of significant community interest as vision-language models (VLMs) for pathology and satellite images gain traction.", + "bbox": [ + 496, + 90, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 517, + 242, + 820, + 256 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a novel method to train diffusion models with self-supervised learning features as conditioning and generate high-quality images in the histopathology and satellite image domains.", + "- We present a framework for large-image synthesis, based on self-supervised guided diffusion, that maintains contextual integrity and image realism over large areas.", + "- We demonstrate the applicability of our model in various classification tasks and showcase its unique ability to augment out-of-distribution datasets.", + "- We introduce text-to-large image generation for digital histopathology and satellite images, highlighting the versatility of our approach." + ], + "bbox": [ + 500, + 257, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 467, + 637, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models: Introduced for image generation by Ho et al. [16], diffusion models have evolved considerably. These enhancements include class conditioning [28], architectural improvements and gradient-based guidance [8], and classifier-free guidance [15]. Latent Diffusion Models (LDMs) [37] proposed a two-step training process with a Variational Autoencoder (VAE) compressing input images into a lower-dimensional latent space and a diffusion model trained in this latent space. Denoising Diffusion Implicit Models (DDIM) [42] accelerate the sampling process by $10 - 50 \\times$ . Self-guided diffusion models [17] also utilize self-supervised learning by quantizing SSL embeddings. In contrast to our approach, their quantization discards useful information from the SSL embeddings.", + "bbox": [ + 496, + 492, + 893, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training generative models directly at the gigapixel resolution is infeasible. An alternative is to generate the images in a coarse-to-fine manner hierarchically. This has already been applied to natural images, where chaining multiple diffusion models generates images up to $1024 \\times 1024$ resolution [32, 39]. However, it is still limited, as it substantially increases the parameter count and is inherently constrained by the final target resolution.", + "bbox": [ + 496, + 704, + 893, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the context of digital histopathology, works have been limited to training unconditional [26] or class-conditioned [27, 47] diffusion models. For text conditioning, pathology text reports were used to provide context on the whole-slide scale [48]. Similar approaches have been", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8533", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/65afc4a7a334f2e01696fa705fe49eb3c0cc1ba8c0d1d3d78050bf41e73fd57b.jpg", + "image_caption": [ + "Figure 2. (a) We train diffusion models on patches $I$ (e.g. the one in the green box) taken from a large image conditioned on SSL embeddings. (b) We present our large image generation framework in 4 steps: (i) We extract a set of spatially arranged embeddings from a reference image or sample them from an auxiliary model. (ii) For every location $(i,j)$ , we compute a conditioning vector $\\lambda_{i,j}$ by interpolating the spatial grid of embeddings. (iii) At every diffusion step, we denoise the patch $F(i,j)$ using the conditioning $\\lambda_{i,j}$ . (iv) The next step is computed by averaging the denoising updates of all patches that overlap at $(i,j)$ ." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 854, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "applied to satellite data [10, 41]. Apart from high-quality images without manual annotation, our SSL conditioning is also necessary for large-image synthesis, as all previous conditioning methods would not be sensitive to the intricate differences between neighboring patches. In recent work, DiffInfinite [1] explored large-image generation using segmentation masks as conditioning. We argue this is still suboptimal as it requires accurate, human-annotated masks for training and a mask-generating model during inference.", + "bbox": [ + 75, + 479, + 468, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning Self-supervised learning (SSL) refers to both discriminative [2, 13] and distillation [5] approaches that aim to learn representations of the data without supervision. In this work, we are mainly interested in self-supervised learning for histopathology. Notable recent developments include the Hierarchical Image Pyramid Transformer (HIPT) [6], which utilizes the inherent hierarchical structure of Whole Slide Images (WSIs), CTransPath [45], a hybrid CNN and multi-scale Swin Transformer model, and iBOT [11], a masked image modeling method. These models excel as patch-level feature extractors for WSIs by leveraging the unique structure of the large-image data and capture the important semantic information that we require for conditioning a generative model.", + "bbox": [ + 75, + 617, + 470, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 168, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose training a diffusion model on patches $I$ drawn from large histopathology and satellite images, using self", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "supervised embeddings $y$ as conditioning. Furthermore, we present a patch-based approach that utilizes the SSL-conditioned diffusion model to synthesize arbitrarily large images. An overview of our method is presented in Fig. 2.", + "bbox": [ + 498, + 479, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Learned representation-guided diffusion", + "text_level": 1, + "bbox": [ + 498, + 551, + 846, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pre-trained self-supervised feature extractor, we employ LDMs [37] to learn the distribution over the large-image patches $p(\\mathcal{I})$ . LDMs are comprised of three components: an image-compressing Variational Autoencoder (VAE) that transforms the input images to latent representations, a U-Net denoiser to learn a denoising diffusion process that transforms Gaussian noise to latents, and a conditioning mechanism that controls the diffusion process. The conditioning is performed in our setting with a single vector $y$ , obtained from the self-supervised model, integrated via a cross-attention mechanism. We train the LDM using pairs of image patches $I$ and the corresponding extracted self-supervised embeddings $y$ .", + "bbox": [ + 496, + 577, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Large image generation", + "text_level": 1, + "bbox": [ + 500, + 786, + 718, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to synthesize high-quality large images, that not only capture global structure but also maintain spatial consistency. As shown in Fig. 2, we replicate the semantics in each patch with SSL-guided LDM. At the same time, we preserve the global arrangement of these semantics as defined by the grid of patches in the reference image. Future", + "bbox": [ + 496, + 810, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8534", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work can explore alternative approaches to ensure spatial alignment, including using topological constraints or integrating low-resolution information.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the MultiDiffusion [3] methodology to generate large images using only a patch-based diffusion model. We can represent the diffusion model as a learned mapping from images and conditions to images:", + "bbox": [ + 75, + 136, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi : \\mathcal {I} \\times \\mathcal {Y} \\rightarrow \\mathcal {I}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 207, + 468, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{I} = \\mathbb{R}^{H\\times W\\times C}$ are the large-image patches and $\\mathcal{V} = \\mathbb{R}^d$ are the per-patch conditions. To generate a patch, we initialize $I_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and sequentially transform to a \"clean\" image $I_0$ following the trained LDM:", + "bbox": [ + 76, + 232, + 468, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {t - 1} = \\Phi \\left(I _ {t} \\mid y\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 305, + 468, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for $t = T, T - 1, \\ldots, 1$ . We assume that a large image can be formed as a spatial grid of $P \\times P$ patches of size $H \\times W$ . We then define the large-image diffusion model as", + "bbox": [ + 75, + 330, + 468, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi : \\mathcal {J} \\times \\mathcal {Z} \\rightarrow \\mathcal {J} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 387, + 468, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{J} = \\mathbb{R}^{PH\\times PW\\times C}$ are the large images and $\\mathcal{Z} = \\underbrace{\\mathcal{Y}\\times\\mathcal{Y}\\cdots\\times\\mathcal{Y}}_{P^2}$ are the conditioning vectors of all patches.", + "bbox": [ + 76, + 412, + 468, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The process $\\Psi(J_t \\mid z) = J_{t-1}$ , with $J_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ , can be approximated by first defining mappings between the two different image and conditioning spaces:", + "bbox": [ + 75, + 459, + 468, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i, j} ^ {\\text {p r o j}}: \\mathcal {J} \\rightarrow I, \\quad \\lambda_ {i, j}: \\mathcal {Z} \\rightarrow \\mathcal {Y} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 513, + 468, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, for the case of large-image generation, we set $F_{i,j}^{\\mathrm{proj}}$ to be a crop (projection) of the large image, centered at $i,j$ , and $\\lambda_{i,j}$ the conditioning $y_{i,j} \\in \\mathbb{R}^d$ at $i,j$ . Since we only have conditioning vectors at the centers of the patches, we can use spatial interpolation algorithms to implement $\\lambda$ . This assumes that the interpolant $y_{i,j}$ is a valid conditioning vector for the diffusion process, which we validate experimentally. Then, $\\Psi$ is defined as", + "bbox": [ + 75, + 544, + 468, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(J _ {t} \\mid z\\right) = \\underset {J \\in \\mathcal {J}} {\\arg \\min } \\sum_ {i, j} \\| F _ {i, j} ^ {\\operatorname {p r o j}} (J) - \\Phi \\left(F _ {i, j} ^ {\\operatorname {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right) \\| ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 676, + 478, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be solved in closed-form by setting each pixel $i,j$ of $J$ to the average of all the patch-diffusion updates", + "bbox": [ + 76, + 722, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(J _ {t} \\mid z\\right) = \\sum_ {i, j} \\frac {F _ {i , j} ^ {\\text {u n p r o j}} (\\mathbf {1})}{\\sum_ {k , l} F _ {k , l} ^ {\\text {u n p r o j}} (\\mathbf {1})} \\otimes F _ {i, j} ^ {\\text {u n p r o j}} \\left(\\Phi \\left(F _ {i, j} ^ {\\text {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 762, + 491, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ denotes a patch image where all values are set to 1 and $F_{i,j}^{\\mathrm{unproj}}$ is the inverse mapping of pixels from the crop centered at $i, j$ back to the large image.", + "bbox": [ + 76, + 825, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are able to generate images larger than the ones produced by the patch diffusion model. At the same time, we", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "control what each patch looks like, which is crucial in maintaining the semantic integrity of the larger image. The self-supervised conditions can capture the variations between neighboring patches necessary for producing realistic results. Generating large images with coarser conditioning, such as global text prompts [48], would lead to uniform texture regions (see supplementary).", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Controllable large-image synthesis", + "text_level": 1, + "bbox": [ + 498, + 207, + 802, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the grid of self-supervised embeddings $z$ cannot be manipulated in a human-interpretable manner, we argue that it is simple to assert more control over the generated images. As illustrated in Fig. 2, this control can be attained by training auxiliary models $p(z \\mid c)$ that translate higher-level conditioning signals, such as text captions $c$ , to learned patch representations $z$ .", + "bbox": [ + 496, + 231, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since there is no available dataset of paired large images and captions, we resort to pre-trained multi-modal models, such as Quilt [18], CLIP [34] and BLIP [22], to provide the text conditioning. We construct training sets by extracting self-supervised embeddings from training-set large images and pairing them with multi-modal image embeddings or generated captions. During inference, we first sample SSL embeddings from the learned distribution $p(z \\mid c)$ , then utilize our patch diffusion models to synthesize a large image.", + "bbox": [ + 496, + 337, + 890, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Image Generation Experiments", + "text_level": 1, + "bbox": [ + 498, + 488, + 787, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 498, + 513, + 601, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train diffusion models on digital histopathology images from The Genome Cancer Atlas (TCGA) [4] and satellite imagery from the National Agriculture Imagery Program (NAIP) [43]. Specifically, we used the TCGA-BRCA (Breast Invasive Carcinoma Collection), TCGA-CRC (COAD + READ Colorectal Carcinoma) datasets, and the Chesapeake Land Cover dataset [36].", + "bbox": [ + 496, + 537, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the TCGA-BRCA and TCGA-CRC datasets, we use images at $20\\times$ magnification, and developed diffusion models conditioned on embeddings from HIPT [6] and iBOT [11], which were pre-trained on PanCancer TCGA. In the case of the HIPT model, we specifically used its patch-level ViT. Additionally, for the TCGA-BRCA dataset, we train a model at $5\\times$ magnification using embeddings from CTransPath [44] for additional evaluations.", + "bbox": [ + 496, + 643, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Chesapeake Land Cover dataset contains 732 NAIP tiles, each measuring a $6\\mathrm{km} \\times 7.5\\mathrm{km}$ area at $1\\mathrm{m}$ resolution. We extract $256 \\times 256$ non-overlapping pixel patches, resulting in 667,000 patches total. Given the absence of publicly available self-supervised learning models tailored to the NAIP data, we train a Vision Transformer (ViT-B/16) [9] using the DINO framework [5]. We then use the learned DINO embeddings to train the diffusion model on pairs of image patches and self-supervised embeddings.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8535", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba1d5d338f16fefd0c738ebe359bf2b54dc18718b0713ebc4ff6e088aa0e890b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 169, + 88, + 413, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5dc80e3652ccf483806329ff59547965b5b0632cdbea960a4c020f9bd7b016b7.jpg", + "image_caption": [ + "Figure 3. (Top) Patches $(256 \\times 256)$ from our models, and the corresponding reference real patches used to generate them. The SSL-guided LDM replicates the semantics of the reference patch. (Bottom) Large images $(1024 \\times 1024)$ from our models, and the corresponding reference real images used to generate them. We preserve the global arrangement of the semantics defined in the reference image." + ], + "image_footnote": [], + "bbox": [ + 117, + 200, + 292, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5154abae3d87d938c1ae2bbe62ee395860662e64a888e34f6d1555fbc93acac6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 200, + 478, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b4278c644829390cb668ca1b1eb2b59480392fab49ee3a600aed545a2c220ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 88, + 785, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dfa2631a7bf199267834b76d3a52a4e627ecaa20416568fa3f427a8257833f23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 200, + 666, + 347 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/924f86354ace31619d0d18db2deadf84eb67ccb2449dd71ff9984e7e014b9f86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 200, + 851, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For patch-level augmentation, we employ the NCT-CRC dataset [19], which has 100,000 Colorectal cancer (CRC) patches from 86 patients. Each $224 \\times 224$ pixel patch at $20 \\times$ magnification is annotated with one of nine distinct tissue class labels.", + "bbox": [ + 75, + 426, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For large-image augmentation, apart from TCGA-BRCA, we also use the BACH dataset [33]. Introduced in the ICIAR 2018 Grand Challenge, the dataset contains 400 H&E-stained Breast Cancer images of size $2048 \\times 1536$ pixels, evenly distributed across four categories: normal, benign, in-situ carcinoma, and invasive carcinoma.", + "bbox": [ + 75, + 502, + 468, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 602, + 290, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For all our experiments, we train the LDM on $256 \\times 256$ pixel patches, following PathLDM [48], which fine-tunes an ImageNet-trained [38] U-Net denoiser and uses a $4 \\times$ downsampling VQ-VAE, instead of the default LDM configuration. These modifications were deemed necessary for applying LDMs on large-image domains.", + "bbox": [ + 75, + 625, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our models on 6 NVIDIA RTX 8000 GPUs, with a batch size of 100 per GPU, utilizing code and pretrained checkpoints from LDM [37]. We set the learning rate at $10^{-4}$ with a warmup of 10,000 steps. We apply DDIM with 50 steps and a guidance scale of 1.75 for both patch sampling and large-image generation. When generating large images we apply the patch diffusion model with a stride $>1$ depending on the desired target quality.", + "bbox": [ + 75, + 715, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Image quality results", + "text_level": 1, + "bbox": [ + 76, + 847, + 274, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Fig. 3, we present synthetic patches and large images from our TCGA-BRCA and NAIP models, along with the", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding references from which the self-supervised embeddings were extracted. We evaluate our method's per-patch and large-image generation quality by computing FID scores [14] using the Clean-FID implementation [31]. We generate 10,000 patches $(256 \\times 256)$ and 3,000 large images $(1024 \\times 1024)$ from diffusion models trained on the TCGA-BRCA, TCGA-CRC, and NAIP datasets. Since our generative model requires a self-supervised conditioning vector $y$ (or multiple vectors $z$ ) for each synthetic image (or large image), we randomly sample embeddings from reference images in the training set to generate images for evaluation.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For patches, we measure FID against the real images (\"Vanilla FID\"). For large images, we follow the evaluation strategy of MultiDiffusion [3] and use FID to compare the distribution of $256 \\times 256$ crops from synthesized large images to that of real image patches of the same size (\"Crop FID\"). We also measure FID directly between the large images and ground truth data using CLIP [34] (\"CLIP FID\").", + "bbox": [ + 496, + 594, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the similarity between the reference and generated large images, we resize from $1024 \\times 1024$ to $256 \\times 256$ and use an SSL model to extract embeddings. We compute the cosine similarity between the paired embeddings (\"Embedding Similarity\").", + "bbox": [ + 496, + 700, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch-level quality: Our models achieve low patch FID scores across all datasets (Tab. 1). For TCGA-BRCA patches, our \"Vanilla FID\" of 6.98 is comparable to the current state-of-the-art [48] (7.64 at $10\\times$ ). We attain similar, low FID scores for the smaller CRC and NAIP datasets.", + "bbox": [ + 496, + 777, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic large image quality: We present the \"Crop FID\", \"CLIP FID\" and \"Embedding Similarity\" results for large images in Table 1. The \"Crop FID\" of the large im", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8536", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3775aee523e5c5cd2079966ec7022269248d014276949f037b76967ade51b5ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Training imagesPatch levelLarge image level
Vanilla FIDCrop FIDCLIP FIDEmb similarity
BRCA 20x15M6.9815.517.430.924
CRC 20x8M6.788.87.340.938
NAIP667k11.543.766.86-
BRCA 5x976k9.74-6.64-
", + "bbox": [ + 81, + 88, + 464, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. FID scores for our generated patch and large images. Our patch-level BRCA model is on par with SoTA [48] (7.64 at $10\\times$ ). \"CLIP FID\" and \"Embedding Similarity\" demonstrate our large images' realism and contextual accuracy.", + "bbox": [ + 76, + 203, + 468, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages is comparable to the \"Vanilla FID\" on the BRCA and CRC datasets, showing that patches from synthesized large images have similar semantic content to the ground truth patches. We attribute the worse \"Crop FID\" for the NAIP model to the limited number of samples available for both SSL and diffusion model training. The \"Crop FID\" is consistently higher than the patch-level FID, which is expected as the large-image generation framework only approximates the distribution of the large images and does not have access to the true conditioning at every location.", + "bbox": [ + 75, + 286, + 467, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our low \"CLIP FID\" scores indicate that the generated large images are similar to real images when resized to $224 \\times 224$ pixels. This indicates that our SSL-guided large-image generation successfully retains the larger-scale semantic arrangements of real data. For NAIP, our model is not as good at synthesizing high-frequency details, which explains the large discrepancy between \"Crop FID\" and \"CLIP FID\". Additionally, when comparing the \"CLIP FID\" of large images synthesized by the $20 \\times$ BRCA model and resized to $5 \\times$ , to images from a model trained directly on $5 \\times$ data, we see minimal difference (7.43 vs. 6.64).", + "bbox": [ + 75, + 438, + 467, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the contextual similarity between synthetic and reference large images for the BRCA and CRC data. We compute cosine similarity between large images using CTransPath embeddings. Our BRCA and CRC models demonstrate high \"Embedding Similarity\" scores of 0.924 and 0.938, respectively, reflecting our framework's effectiveness in preserving the integrity of context and key features on the large-image scale.", + "bbox": [ + 75, + 604, + 467, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Image Augmentation Experiments", + "text_level": 1, + "bbox": [ + 76, + 739, + 387, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 3, apart from visual fidelity, the synthetic images preserve the intricate characteristics of the reference images, both on the patch and large-image scale. These characteristics include the nuanced textural elements, histological staining, and cell structure. This correspondence between real and synthetic images demonstrates the detailed and varied information captured by the self-supervised embeddings used as conditioning. In conjunction with the high \"Embedding Similarity\", it justifies using variations of im", + "bbox": [ + 75, + 763, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages generated from SSL embeddings for patch and large-image level data augmentation.", + "bbox": [ + 498, + 90, + 888, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Having a powerful generative model enables us to perform data augmentation for tasks where we can control the augmented image label using conditioning. In our setting, our diffusion models are not trained with class labels; instead, we synthesize a novel image using the conditioning from a reference patch or region.", + "bbox": [ + 496, + 123, + 890, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assume that i) the self-supervised embedding used as conditioning contains information about the target label and ii) the diffusion-generated variations of an image do not alter this target label information. We validate both assumptions experimentally, by augmenting training sets on patch and large-image level tissue classification tasks, including a Multiple Instance Learning (MIL) task.", + "bbox": [ + 496, + 214, + 890, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Large-image augmentation on TCGA-BRCA: We examine two histopathology slide-level binary classification tasks on TCGA-BRCA: BRCA Subtyping (Invasive Ductal Carcinoma (IDC) vs Invasive Lobular Carcinoma (ILC)) and HRD prediction. We utilize a minimal dataset, just $10\\%$ of real WSI data (100 WSIs), to train MIL algorithms. We generate an equal set of synthetic images for 100 additional WSIs using training set images as reference.", + "bbox": [ + 496, + 321, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ 10-fold cross-validation to divide our dataset into training and testing segments. Within each fold, two multiple instance learning (MIL) models, CLAM-SB [25] and DSMIL [21], are trained on two sets: one with real data and another combining real and synthetic data. To train the MIL models, we extract features using the CTransPath ViT [44]. The results, detailed in Table 2, indicate that models trained on the augmented datasets consistently surpass their real-only counterparts, regardless of the MIL algorithm used. This demonstrates the value of the synthetic images generated by our method, confirming their efficacy as comparable to real images for training purposes.", + "bbox": [ + 496, + 444, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Large-image augmentation on BACH: We double the training set of BACH [33] by adding as many synthetic large-images, produced by the TCGA-BRCA diffusion model. From each $2048 \\times 1536$ pixel training set image, we extract a $8 \\times 6$ SSL embedding grid to generate a variation with the same label. For classification, we employ a ConvNeXt V2huge [46] model pre-trained on ImageNet. We train a 2-layer MLP classifier on top of the penultimate layer features and evaluate it on the official test set.", + "bbox": [ + 496, + 627, + 890, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results, presented in Table 3a, reveal a notable improvement in classifier performance, from $78\\%$ to $83\\%$ . This improvement again confirms the high quality of our synthetic data while also highlighting the versatility of our approach. Despite being trained on $256 \\times 256$ patches from TCGA-BRCA, the model generalizes to produce realistic large images from a completely different dataset. We attribute this generalization capability to the expressiveness of the SSL features and the potency of the diffusion model", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8537", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52b65db27a4cd5190299070928a7ac0b56d07026e276650dfe9c4a057b6df7dc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TCGA-BRCA SubtypingTCGA-BRCA HRD
Method1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real
CLAM-SB0.7250.8120.8860.8980.910.6030.6440.6490.7650.787
DSMIL0.6090.6590.8380.8560.9050.5170.5540.5630.6390.669
", + "bbox": [ + 81, + 88, + 893, + 170 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to accurately portray them in images. While our model does not reach the current SoTA accuracy of $87\\%$ , achieved by an ensemble approach [7], the simplicity and ability to integrate with other models make for a noteworthy contribution.", + "bbox": [ + 76, + 236, + 468, + 297 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8aecc5424c9a5dd6201534e124557ae021263ebd9285cddf007ce70c906a4fb.jpg", + "table_caption": [ + "Table 2. The inclusion of synthetic data consistently enhances AUC across various MIL architectures and BRCA tasks. The dataset contains 1000 real images, so “ $10\\%+$ synthetic” indicates training with 100 real and 100 synthetic WSIs, with the remainder used for testing." + ], + "table_footnote": [], + "table_body": "
Training DataTest Acc
Real78 %
Synthetic70 %
Real + synthetic83 %
SoTA [7]87 %
", + "bbox": [ + 81, + 311, + 267, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c2ddf479da98bbf4d59810380072eb4f2f913b67ab973e625f1a8cf6b8e4920d.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
Training DataVal Acc
Real93.8 %
Synthetic90.19 %
Real + synthetic96.27 %
SoTA [20]96.26 %
", + "bbox": [ + 274, + 311, + 460, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b)", + "bbox": [ + 357, + 386, + 375, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Our data augmentations provide notable improvements for the BACH (a) and CRC-VAL-HE (b) datasets. Notably, the diffusion training data does not overlap with the data of the augmented datasets.", + "bbox": [ + 75, + 402, + 468, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Patch-level image augmentation: We further investigate our out-of-distribution generalization capabilities by augmenting the NCT-CRC dataset [19]. We leverage our diffusion model trained on the TCGA-CRC data, which does not overlap with NCT-CRC, conditioned on embeddings from an iBOT ViT [11]. We generate an augmented dataset of equal size to the original by using the SSL embedding of every patch in the NCT-CRC training set to synthesize a corresponding image of the same label.", + "bbox": [ + 75, + 478, + 468, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We train an ImageNet pre-trained ResNet-50 [12] network on three splits: real images only, synthetic images only, and a combination of both. Evaluation on the CRC-VAL-HE-7K test set demonstrates a significant performance increase when synthetic data is introduced, with classifier accuracy rising from $93.8\\%$ to $96.27\\%$ as presented in Table 3b. We match the current SoTA [20], which used an ensemble of deep models, with a model agnostic approach. As in the previous experiment, the diffusion model, now trained on the significantly smaller TCGA-CRC data, can also effectively synthesize images from a completely different dataset by only controlling the self-supervised conditioning.", + "bbox": [ + 75, + 614, + 470, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Text-to-large image synthesis", + "text_level": 1, + "bbox": [ + 76, + 828, + 344, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For previous tasks, our large image generation approach synthesizes variations of an existing set of images from the pre-computed self-supervised embeddings. In Sec. 3.3 we", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "discussed how to control large image generation with auxiliary signals from any domain (class labels, text captions, etc.), by training models $p(z \\mid c)$ that can be combined with the embedding-conditioned image generation. We demonstrate controllable image synthesis with text-to-large image generation experiments on histopathology and satellite data. We measure the similarity between synthetic images and the text prompts used in generating them using vision-language models (VLMs).", + "bbox": [ + 496, + 236, + 890, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-to-large histopathology images: We utilize the CRC and BRCA diffusion models to generate $1024 \\times 1024$ pixel images from text prompts. We first construct training sets by pairing $4 \\times 4$ SSL embedding grids $z$ from large BRCA and CRC images, with their corresponding Quilt [18] image embeddings $c$ . We then train an auxiliary diffusion model to sample $p(z \\mid c)$ . During inference, we use a text embedding $c'$ as a proxy for the image embedding, to sample $z$ and synthesize a large image. To bridge the gap between the image and text Quilt embeddings [23, 30] we perturb the image embeddings when training the auxiliary diffusion model with Gaussian noise of variance $\\sigma^2 = 0.1$ .", + "bbox": [ + 496, + 377, + 892, + 558 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the text-to-large image pipeline, we generate images from a pre-defined set of classes described in natural language; non-malignant benign tissue, malignant in-situ carcinoma, malignant invasive carcinoma, normal breast tissue for BRCA and colon adenocarcinoma, benign colonic tissue for CRC. We construct the confusion matrix of zero-shot classifiers on the synthesized data. We used two different VLMs as zero-shot classifiers, Quilt and BiomedCLIP [49]. The results presented in Fig. 4 demonstrate our ability to synthesize images consistent with the text prompts. The capabilities of the VLM used, limit our synthetic image generation. The lower performance of BRCA vs. CRC is consistent with the results reported in Quilt [18]. Furthermore, we asked an expert pathologist to classify 100 synthetic CRC images as benign or adenocarcinoma images. Their evaluation showed an $89.9\\%$ agreement rate with the labels used for image generation, indicating consistency in our text-to-large image pipeline. We show examples of synthesized images in the supplementary.", + "bbox": [ + 496, + 563, + 893, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-to-large satellite images: To synthesize novel satellite images we first create a training set of $30\\mathrm{k}1024\\times 1024$ pixel large NAIP images, and pair them with captions", + "bbox": [ + 500, + 854, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8538", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from a BLIP model [22]. We train a diffusion model to sample the $4 \\times 4$ SSL embeddings $z$ from the captions $c$ . To evaluate, we create a separate set of 1000 NAIP image-caption pairs and measure the CLIP similarity between generated images and the given captions. We achieve a CLIP similarity score of 0.22, showing that we can effectively learn the mapping from text to large images with this hierarchical approach. Although our CLIP similarity is slightly worse than the scores reported for text-to-image Stable Diffusion models ( $>0.24$ ) [32], we expect this drop in performance as we trained with machine-generated captions. Training and generated image-caption pairs are provided in supplementary.", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5b98b4f0e2b4e5466e0373f7f4eb758b9b9e7588b8e30d20f1ca7935afddbb7.jpg", + "image_caption": [ + "Figure 4. Confusion matrix of zero-shot classification for novel TCGA-CRC and TCGA-BRCA synthetic images." + ], + "image_footnote": [], + "bbox": [ + 76, + 282, + 472, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Combining self-supervised embeddings with diffusion", + "text_level": 1, + "bbox": [ + 76, + 585, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate our patch-generating models by posing the question does the diffusion model learn more about the data than the self-supervised learning model? We hypothesize that by combining the pre-trained self-supervised embedding with the denoising task we can improve the learned representations of the data, leading to better performance in downstream tasks, which are performed with features from self-supervised learning. To validate this hypothesis, we utilize the trained diffusion model as a feature extractor and apply a Multiple-instance learning (MIL) approach for the slide-level classification task of subtyping Breast Cancer. For each patch in a whole-slide image (WSI) we first extract the self-supervised embeddings, which are then used as conditioning to obtain features from the denoiser's U-Net bottleneck layer at a fixed timestep $t = 50$ .", + "bbox": [ + 75, + 628, + 468, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 4, we evaluate the effectiveness of our novel fusion of generative diffusion and self-supervised embeddings. We compare the performance of MIL algorithms [21,", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[25] using features derived from our approach at $20 \\times$ and $5 \\times$ magnification (LDMHIPT/CTransPath) against using only the self-supervised conditioning (HIPT/ CTransPath). We used a 10-fold cross-validation strategy consistent with the data splits from HIPT [6], training the MIL algorithms on both the full dataset (100%) and a reduced subset (25%).", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results indicate that integrating self-supervised features into the diffusion model as conditioning leads to learning better representations and improves whole-slide classification. By fusing the generative knowledge from the diffusion process with the discriminative capabilities of the self-supervised embeddings, we construct a successful model for both discriminative and generative tasks.", + "bbox": [ + 496, + 183, + 892, + 289 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e7ec8e574ec5b36b08b4b5870882f50319d1044862e5c8920fb6b71c547c2d6a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MagFeatures25% training100% training
CLAM-SBDSMILCLAM-SBDSMIL
20×HIPT0.7880.7840.8610.839
LDMHIPT0.8420.7950.9080.894
5×CTransPath0.9000.8960.9190.910
LDMCTransPath0.9130.9050.9230.936
", + "bbox": [ + 504, + 305, + 888, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. 10-fold cross-validation AUC for BRCA Histological subtyping. $\\mathrm{LDM}_{\\mathrm{HIPT}}$ denotes diffusion features conditioned on HIPT embeddings. The fusion of SSL and diffusion features outperforms the SSL features by themselves.", + "bbox": [ + 496, + 402, + 890, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 496, + 617, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a novel approach to training diffusion models in large-image domains, such as digital histopathology and remote sensing. We overcome the need for fine-grained annotation by introducing self-supervised representation guided diffusion models, achieving remarkable image synthesis results on the patch level. Our approach also enables us to synthesize high-quality large images, where we have the ability to dictate the global structure by controlling the spatial arrangement of the conditions. We evaluated the usefulness of our synthetic images on a number of patch and large image-level tasks, as well as introduced a text-to-large image generation framework. Naively augmenting whole-slide images is a time-consuming process. We leave to future work the exploration of adaptive augmentation strategies that choose which image parts to augment. We believe these results illustrate the great potential for this technology to lead to bespoke foundational models for specialized domains, comparable to existing models for natural images.", + "bbox": [ + 496, + 523, + 890, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This research was partially supported by NCI awards 5U24CA215109, 1R21CA258493-01A1, UH3CA225021, NSF grants IIS-2123920, IIS-2212046 and Stony Brook Profund 2022 seed funding. We thank Rajarsi Gupta for his valuable feedback.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8539", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marco Aversa, Gabriel Nobis, Miriam Hagele, Kai Standvoss, Mihaela Chirica, Roderick Murray-Smith, Ahmed Alaa, Lukas Ruff, Daniela Ivanova, Wojciech Samek, et al. Diffinfinite: Large mask-image synthesis via parallel random patch diffusion in histopathology. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3", + "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1737-1752. PMLR, 2023. 4, 5", + "[4] JN Cancer Genome Atlas Research Network et al. The cancer genome atlas pan-cancer analysis project. Nat. Genet, 45 (10):1113-1120, 2013. 2, 4", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 3, 4", + "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 1, 3, 4, 8", + "[7] Sai Saketh Chennamsetty, Mohammed Safwan, and Varghese Alex. Classification of breast cancer histology image using ensemble of pre-trained neural networks. In Image Analysis and Recognition: 15th International Conference, ICIAR 2018, Póvoa de Varzim, Portugal, June 27-29, 2018, Proceedings 15, pages 804-811. Springer, 2018. 7", + "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[10] Miguel Espinosa and Elliot J Crowley. Generate your own scotland: Satellite image generation conditioned on maps. arXiv preprint arXiv:2308.16648, 2023. 3", + "[11] Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathology with masked image modeling. medRxiv, pages 2023-07, 2023. 1, 3, 4, 7", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[17] Vincent Tao Hu, David W Zhang, Yuki M Asano, Gertjan J Burghouts, and Cees GM Snoek. Self-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18413-18422, 2023. 2", + "[18] Wisdom Oluchi Ikezogwo, Mehmet Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Stefan Chan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. *Quilt-1m: One million image-text pairs for histopathology.* arXiv preprint arXiv:2306.11207, 2023. 4, 7", + "[19] Jakob Nikolas Kather, Niels Halama, and Alexander Marx. 100,000 histological images of human colorectal cancer and healthy tissue, 2018. 5, 7", + "[20] Anurodh Kumar, Amit Vishwakarma, and Varun Bajaj. Crccn-net: Automated framework for classification of colorectal tissue using histopathological images. Biomedical Signal Processing and Control, 79:104172, 2023. 7", + "[21] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 6, 8", + "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4, 8", + "[23] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 7", + "[24] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 1", + "[25] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8540", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Puria Azadi Moghadam, Sanne Van Dalen, Karina C Martin, Jochen Lennerz, Stephen Yip, Hossein Farahani, and Ali Bashashati. A morphology focused diffusion probabilistic model for synthesis of histopathology images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2000-2009, 2023. 2", + "[27] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, et al. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13 (1):12098, 2023. 2", + "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2", + "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1", + "[30] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 7", + "[31] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 8", + "[33] António Polónia, Catarina Eloy, and Paulo Aguiar. BACH Dataset : Grand Challenge on Breast Cancer Histology images, 2020. 5, 6", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5", + "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[36] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 4", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5", + "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2", + "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1", + "[41] Ahmad Sebaq and Mohamed ElHelw. Rsdiff: Remote sensing image generation from text using diffusion model. arXiv preprint arXiv:2309.02455, 2023. 3", + "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2", + "[43] USGS. National agriculture imagery program (NAIP), 2023. https://www.usgs.gov/centers/eros/science/usgs-eros-archive-aerial-photography - national - agriculture - imagery-program-naip.2,4", + "[44] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Junzhou Huang, Wei Yang, and Xiao Han. Transpath: Transformer-based self-supervised learning for histopathological image classification. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VIII 24, pages 186-195. Springer, 2021. 4, 6", + "[45] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis, 81:102559, 2022. 1, 3", + "[46] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 6", + "[47] Xuan Xu, Saarthak Kapse, Rajarsi Gupta, and Prateek Prasanna. Vit-dae: Transformer-driven diffusion autoencoder for histopathology image analysis. arXiv preprint arXiv:2304.01053, 2023. 2", + "[48] Srikar Yellapragada, Alexandros Graikos, Prateek Prasanna, Tahsin Kurc, Joel Saltz, and Dimitris Samaras. PathLDM: Text conditioned latent diffusion model for histopathology. arXiv preprint arXiv:2309.00748, 2023. 2, 4, 5, 6", + "[49] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8541", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Valluri, Cliff Wong, et al. Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915, 2023. 7", + "bbox": [ + 109, + 90, + 468, + 133 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8542", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_model.json b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d40c0c2d34d735a337dce79e4d02fde142de967 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Learned representation-guided diffusion models for large-image generation" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.181, + 0.768, + 0.234 + ], + "angle": 0, + "content": "Alexandros Graikos* Srikar Yellapragada* Minh-Quan Le Saarthak Kapse Prateek Prasanna Joel Saltz Dimitris Samaras Stony Brook University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.755 + ], + "angle": 0, + "content": "To synthesize high-fidelity samples, diffusion models typically require auxiliary data to guide the generation process. However, it is impractical to procure the painstaking patch-level annotation effort required in specialized domains like histopathology and satellite imagery; it is often performed by domain experts and involves hundreds of millions of patches. Modern-day self-supervised learning (SSL) representations encode rich semantic and visual information. In this paper, we posit that such representations are expressive enough to act as proxies to fine-grained human labels. We introduce a novel approach that trains diffusion models conditioned on embeddings from SSL. Our diffusion models successfully project these features back to high-quality histopathology and remote sensing images. In addition, we construct larger images by assembling spatially consistent patches inferred from SSL embeddings, preserving long-range dependencies. Augmenting real data by generating variations of real images improves downstream classifier accuracy for patch-level and larger, image-scale classification tasks. Our models are effective even on datasets not encountered during training, demonstrating their robustness and generalizability. Generating images from learned embeddings is agnostic to the source of the embeddings. The SSL embeddings used to generate a large image can either be extracted from a reference image, or sampled from an auxiliary model conditioned on any related modality (e.g. class labels, text, genomic data). As proof of concept, we introduce the text-to-large image synthesis paradigm where we successfully synthesize large pathology and satellite images out of text descriptions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.782, + 0.21, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.807, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Diffusion models produce high-quality and diverse samples across a spectrum of generative tasks [8, 24]. This leap forward has been enabled by the simultaneous curation of large-scale multi-modal datasets [40] and the development" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.267, + 0.895, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.421, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 1. We propose using SSL features to condition diffusion models. This allows us to construct large images by assembling consistent patches inferred from a spatial arrangement of SSL embeddings. The generated image retains the semantics of the embeddings used as a condition, maintaining the forested and open areas from the reference. Best viewed zoomed-in." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.581 + ], + "angle": 0, + "content": "of efficient conditioning mechanisms [35, 37]. The key to unlocking the models' capabilities is to integrate auxiliary information during training and inference [8, 15, 28]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.809 + ], + "angle": 0, + "content": "Large-scale human-annotated datasets are mostly limited to image-caption pairs, collected from easily accessible online repositories and labeled by non-expert annotators. However, in domains such as digital histopathology and remote sensing, where gigapixel scale images provide vast amounts of unlabeled data, annotation proves challenging. Moreover, the process requires expert knowledge and is more difficult at a finer scale, i.e., captioning large crops of the gigapixel image is simpler than captioning smaller patches. Based on our estimates (see supplemental), annotating the entire TCGA-BRCA dataset with captions would take \\(\\approx 40,000\\) hours of pathologist's time. Replicating the impressive results of diffusion models in these domains has been limited by the scarcity of fine-grained per-image conditioning, vital for high-quality image synthesis [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Modern-day self-supervised learning (SSL) representations [5] encode rich semantic and visual information. Features from trained self-supervised models serve as compact image representations and are widely used to perform discriminative downstream tasks successfully [6, 11, 45], proving that these compressed representations indeed encode" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.463, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution. Correspondence to agraikos@cs.stonybrook.edu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.253, + 0.899 + ], + "angle": 0, + "content": "\\(^{1}\\)Code is available at this link" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.463, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "useful semantic information about the images. We hypothesize that such SSL representations are already expressive enough to act as proxies to fine-grained human labels. If this is true, these representations should be able to condition the training of effective diffusion models in these domains. In this novel approach, we utilize self-supervised feature extractors as image annotators; these features provide necessary per-image conditioning signals at the highest resolutions, required for diffusion model training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.473, + 0.365 + ], + "angle": 0, + "content": "Our experiments show that conditioning with expressive self-supervised features leads to precise control over the image content. The SSL features are adept at identifying complex patterns and structures in the images, while the diffusion model learns to translate them into visual components accurately (Fig. 3). This motivates us to perform large-image synthesis by locally controlling the appearance using SSL conditioning and dictating the global structure through the spatial arrangement of the conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.563 + ], + "angle": 0, + "content": "Our approach synthesizes large images in a patch-based manner, using a single image diffusion model at the highest resolution. We represent a large image as a grid of SSL embeddings, where each serves as a representation of a large image neighborhood. The whole image is then synthesized by generating consistent patches that capture both the local properties, as given by the local patch conditioning, and the spatial arrangement of the conditioning features. If we change this spatial arrangement, we are, in fact, editing how semantic elements are arranged globally in the large image. This strategy enables the controllable generation of images of virtually any size without significantly increased computation compared to the base patch-level model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.473, + 0.656 + ], + "angle": 0, + "content": "To generate a large image, our approach requires the patch diffusion model and the spatially-arranged conditioning. We can start with a reference large image as source and extract SSL embeddings from non-overlapping segments, enabling our method to synthesize a variation of the original image (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.473, + 0.839 + ], + "angle": 0, + "content": "Utilizing SSL embeddings as conditions allows us to have the necessary control over image generation, at the expense of an explainable and easy-to-use conditioning mechanism. Nevertheless, we argue that since generating images from learned embeddings is agnostic to the embedding source, there are simple ways to combine control over generated images with explainability. We propose training auxiliary models to transform higher-level conditioning signals, such as text captions, to the learned patch representations. To demonstrate this versatility, we introduce text-to-large-image synthesis by training an auxiliary model to sample a spatial arrangement of embeddings from a text description." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "We train patch-level diffusion models using self-supervised features as conditioning on digital histopathology (TCGA [4]) and satellite image (NAIP [43]) datasets. We perform extensive evaluations and demonstrate the ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.243 + ], + "angle": 0, + "content": "vantages of SSL conditioning and our large-image generation framework on synthesis and classification tasks. Our model achieves exceptional patch-level and large-image quality, the ability to improve classifiers through data augmentation even when synthesizing out-of-distribution data, and effective fusion of diffusion and SSL features for downstream applications. Finally, we are the first to perform text-to-large image synthesis, which should be of significant community interest as vision-language models (VLMs) for pathology and satellite images gain traction." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.243, + 0.821, + 0.257 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.258, + 0.892, + 0.317 + ], + "angle": 0, + "content": "- We develop a novel method to train diffusion models with self-supervised learning features as conditioning and generate high-quality images in the histopathology and satellite image domains." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.318, + 0.892, + 0.363 + ], + "angle": 0, + "content": "- We present a framework for large-image synthesis, based on self-supervised guided diffusion, that maintains contextual integrity and image realism over large areas." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.364, + 0.892, + 0.407 + ], + "angle": 0, + "content": "- We demonstrate the applicability of our model in various classification tasks and showcase its unique ability to augment out-of-distribution datasets." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.409, + 0.892, + 0.455 + ], + "angle": 0, + "content": "- We introduce text-to-large image generation for digital histopathology and satellite images, highlighting the versatility of our approach." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.258, + 0.892, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.468, + 0.638, + 0.484 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.704 + ], + "angle": 0, + "content": "Diffusion models: Introduced for image generation by Ho et al. [16], diffusion models have evolved considerably. These enhancements include class conditioning [28], architectural improvements and gradient-based guidance [8], and classifier-free guidance [15]. Latent Diffusion Models (LDMs) [37] proposed a two-step training process with a Variational Autoencoder (VAE) compressing input images into a lower-dimensional latent space and a diffusion model trained in this latent space. Denoising Diffusion Implicit Models (DDIM) [42] accelerate the sampling process by \\(10 - 50 \\times\\). Self-guided diffusion models [17] also utilize self-supervised learning by quantizing SSL embeddings. In contrast to our approach, their quantization discards useful information from the SSL embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Training generative models directly at the gigapixel resolution is infeasible. An alternative is to generate the images in a coarse-to-fine manner hierarchically. This has already been applied to natural images, where chaining multiple diffusion models generates images up to \\(1024 \\times 1024\\) resolution [32, 39]. However, it is still limited, as it substantially increases the parameter count and is inherently constrained by the final target resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In the context of digital histopathology, works have been limited to training unconditional [26] or class-conditioned [27, 47] diffusion models. For text conditioning, pathology text reports were used to provide context on the whole-slide scale [48]. Similar approaches have been" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.856, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.893, + 0.456 + ], + "angle": 0, + "content": "Figure 2. (a) We train diffusion models on patches \\( I \\) (e.g. the one in the green box) taken from a large image conditioned on SSL embeddings. (b) We present our large image generation framework in 4 steps: (i) We extract a set of spatially arranged embeddings from a reference image or sample them from an auxiliary model. (ii) For every location \\( (i,j) \\), we compute a conditioning vector \\( \\lambda_{i,j} \\) by interpolating the spatial grid of embeddings. (iii) At every diffusion step, we denoise the patch \\( F(i,j) \\) using the conditioning \\( \\lambda_{i,j} \\). (iv) The next step is computed by averaging the denoising updates of all patches that overlap at \\( (i,j) \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.47, + 0.617 + ], + "angle": 0, + "content": "applied to satellite data [10, 41]. Apart from high-quality images without manual annotation, our SSL conditioning is also necessary for large-image synthesis, as all previous conditioning methods would not be sensitive to the intricate differences between neighboring patches. In recent work, DiffInfinite [1] explored large-image generation using segmentation masks as conditioning. We argue this is still suboptimal as it requires accurate, human-annotated masks for training and a mask-generating model during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Self-Supervised Learning Self-supervised learning (SSL) refers to both discriminative [2, 13] and distillation [5] approaches that aim to learn representations of the data without supervision. In this work, we are mainly interested in self-supervised learning for histopathology. Notable recent developments include the Hierarchical Image Pyramid Transformer (HIPT) [6], which utilizes the inherent hierarchical structure of Whole Slide Images (WSIs), CTransPath [45], a hybrid CNN and multi-scale Swin Transformer model, and iBOT [11], a masked image modeling method. These models excel as patch-level feature extractors for WSIs by leveraging the unique structure of the large-image data and capture the important semantic information that we require for conditioning a generative model." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.169, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We propose training a diffusion model on patches \\(I\\) drawn from large histopathology and satellite images, using self" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.542 + ], + "angle": 0, + "content": "supervised embeddings \\( y \\) as conditioning. Furthermore, we present a patch-based approach that utilizes the SSL-conditioned diffusion model to synthesize arbitrarily large images. An overview of our method is presented in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.553, + 0.847, + 0.57 + ], + "angle": 0, + "content": "3.1. Learned representation-guided diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Given a pre-trained self-supervised feature extractor, we employ LDMs [37] to learn the distribution over the large-image patches \\( p(\\mathcal{I}) \\). LDMs are comprised of three components: an image-compressing Variational Autoencoder (VAE) that transforms the input images to latent representations, a U-Net denoiser to learn a denoising diffusion process that transforms Gaussian noise to latents, and a conditioning mechanism that controls the diffusion process. The conditioning is performed in our setting with a single vector \\( y \\), obtained from the self-supervised model, integrated via a cross-attention mechanism. We train the LDM using pairs of image patches \\( I \\) and the corresponding extracted self-supervised embeddings \\( y \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.719, + 0.803 + ], + "angle": 0, + "content": "3.2. Large image generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our goal is to synthesize high-quality large images, that not only capture global structure but also maintain spatial consistency. As shown in Fig. 2, we replicate the semantics in each patch with SSL-guided LDM. At the same time, we preserve the global arrangement of these semantics as defined by the grid of patches in the reference image. Future" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8534" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "work can explore alternative approaches to ensure spatial alignment, including using topological constraints or integrating low-resolution information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.197 + ], + "angle": 0, + "content": "We follow the MultiDiffusion [3] methodology to generate large images using only a patch-based diffusion model. We can represent the diffusion model as a learned mapping from images and conditions to images:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.208, + 0.469, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\Phi : \\mathcal {I} \\times \\mathcal {Y} \\rightarrow \\mathcal {I}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.233, + 0.469, + 0.295 + ], + "angle": 0, + "content": "where \\(\\mathcal{I} = \\mathbb{R}^{H\\times W\\times C}\\) are the large-image patches and \\(\\mathcal{V} = \\mathbb{R}^d\\) are the per-patch conditions. To generate a patch, we initialize \\(I_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and sequentially transform to a \"clean\" image \\(I_0\\) following the trained LDM:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.306, + 0.469, + 0.322 + ], + "angle": 0, + "content": "\\[\nI _ {t - 1} = \\Phi \\left(I _ {t} \\mid y\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.469, + 0.378 + ], + "angle": 0, + "content": "for \\( t = T, T - 1, \\ldots, 1 \\). We assume that a large image can be formed as a spatial grid of \\( P \\times P \\) patches of size \\( H \\times W \\). We then define the large-image diffusion model as" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.388, + 0.469, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\Psi : \\mathcal {J} \\times \\mathcal {Z} \\rightarrow \\mathcal {J} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.413, + 0.469, + 0.459 + ], + "angle": 0, + "content": "where \\(\\mathcal{J} = \\mathbb{R}^{PH\\times PW\\times C}\\) are the large images and \\(\\mathcal{Z} = \\underbrace{\\mathcal{Y}\\times\\mathcal{Y}\\cdots\\times\\mathcal{Y}}_{P^2}\\) are the conditioning vectors of all patches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.469, + 0.506 + ], + "angle": 0, + "content": "The process \\(\\Psi(J_t \\mid z) = J_{t-1}\\), with \\(J_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\), can be approximated by first defining mappings between the two different image and conditioning spaces:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.515, + 0.469, + 0.534 + ], + "angle": 0, + "content": "\\[\nF _ {i, j} ^ {\\text {p r o j}}: \\mathcal {J} \\rightarrow I, \\quad \\lambda_ {i, j}: \\mathcal {Z} \\rightarrow \\mathcal {Y} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.667 + ], + "angle": 0, + "content": "where, for the case of large-image generation, we set \\(F_{i,j}^{\\mathrm{proj}}\\) to be a crop (projection) of the large image, centered at \\(i,j\\), and \\(\\lambda_{i,j}\\) the conditioning \\(y_{i,j} \\in \\mathbb{R}^d\\) at \\(i,j\\). Since we only have conditioning vectors at the centers of the patches, we can use spatial interpolation algorithms to implement \\(\\lambda\\). This assumes that the interpolant \\(y_{i,j}\\) is a valid conditioning vector for the diffusion process, which we validate experimentally. Then, \\(\\Psi\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.677, + 0.479, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(J _ {t} \\mid z\\right) = \\underset {J \\in \\mathcal {J}} {\\arg \\min } \\sum_ {i, j} \\| F _ {i, j} ^ {\\operatorname {p r o j}} (J) - \\Phi \\left(F _ {i, j} ^ {\\operatorname {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right) \\| ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.723, + 0.469, + 0.754 + ], + "angle": 0, + "content": "which can be solved in closed-form by setting each pixel \\(i,j\\) of \\(J\\) to the average of all the patch-diffusion updates" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.763, + 0.492, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(J _ {t} \\mid z\\right) = \\sum_ {i, j} \\frac {F _ {i , j} ^ {\\text {u n p r o j}} (\\mathbf {1})}{\\sum_ {k , l} F _ {k , l} ^ {\\text {u n p r o j}} (\\mathbf {1})} \\otimes F _ {i, j} ^ {\\text {u n p r o j}} \\left(\\Phi \\left(F _ {i, j} ^ {\\text {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) denotes a patch image where all values are set to 1 and \\(F_{i,j}^{\\mathrm{unproj}}\\) is the inverse mapping of pixels from the crop centered at \\(i, j\\) back to the large image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We are able to generate images larger than the ones produced by the patch diffusion model. At the same time, we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "control what each patch looks like, which is crucial in maintaining the semantic integrity of the larger image. The self-supervised conditions can capture the variations between neighboring patches necessary for producing realistic results. Generating large images with coarser conditioning, such as global text prompts [48], would lead to uniform texture regions (see supplementary)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.208, + 0.803, + 0.224 + ], + "angle": 0, + "content": "3.3. Controllable large-image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.232, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Although the grid of self-supervised embeddings \\( z \\) cannot be manipulated in a human-interpretable manner, we argue that it is simple to assert more control over the generated images. As illustrated in Fig. 2, this control can be attained by training auxiliary models \\( p(z \\mid c) \\) that translate higher-level conditioning signals, such as text captions \\( c \\), to learned patch representations \\( z \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Since there is no available dataset of paired large images and captions, we resort to pre-trained multi-modal models, such as Quilt [18], CLIP [34] and BLIP [22], to provide the text conditioning. We construct training sets by extracting self-supervised embeddings from training-set large images and pairing them with multi-modal image embeddings or generated captions. During inference, we first sample SSL embeddings from the learned distribution \\( p(z \\mid c) \\), then utilize our patch diffusion models to synthesize a large image." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.489, + 0.788, + 0.506 + ], + "angle": 0, + "content": "4. Image Generation Experiments" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.514, + 0.602, + 0.528 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.643 + ], + "angle": 0, + "content": "We train diffusion models on digital histopathology images from The Genome Cancer Atlas (TCGA) [4] and satellite imagery from the National Agriculture Imagery Program (NAIP) [43]. Specifically, we used the TCGA-BRCA (Breast Invasive Carcinoma Collection), TCGA-CRC (COAD + READ Colorectal Carcinoma) datasets, and the Chesapeake Land Cover dataset [36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.764 + ], + "angle": 0, + "content": "For the TCGA-BRCA and TCGA-CRC datasets, we use images at \\(20\\times\\) magnification, and developed diffusion models conditioned on embeddings from HIPT [6] and iBOT [11], which were pre-trained on PanCancer TCGA. In the case of the HIPT model, we specifically used its patch-level ViT. Additionally, for the TCGA-BRCA dataset, we train a model at \\(5\\times\\) magnification using embeddings from CTransPath [44] for additional evaluations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The Chesapeake Land Cover dataset contains 732 NAIP tiles, each measuring a \\(6\\mathrm{km} \\times 7.5\\mathrm{km}\\) area at \\(1\\mathrm{m}\\) resolution. We extract \\(256 \\times 256\\) non-overlapping pixel patches, resulting in 667,000 patches total. Given the absence of publicly available self-supervised learning models tailored to the NAIP data, we train a Vision Transformer (ViT-B/16) [9] using the DINO framework [5]. We then use the learned DINO embeddings to train the diffusion model on pairs of image patches and self-supervised embeddings." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.17, + 0.089, + 0.415, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.201, + 0.294, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.202, + 0.48, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.787, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.202, + 0.667, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.202, + 0.852, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.359, + 0.893, + 0.403 + ], + "angle": 0, + "content": "Figure 3. (Top) Patches \\((256 \\times 256)\\) from our models, and the corresponding reference real patches used to generate them. The SSL-guided LDM replicates the semantics of the reference patch. (Bottom) Large images \\((1024 \\times 1024)\\) from our models, and the corresponding reference real images used to generate them. We preserve the global arrangement of the semantics defined in the reference image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.501 + ], + "angle": 0, + "content": "For patch-level augmentation, we employ the NCT-CRC dataset [19], which has 100,000 Colorectal cancer (CRC) patches from 86 patients. Each \\(224 \\times 224\\) pixel patch at \\(20 \\times\\) magnification is annotated with one of nine distinct tissue class labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.503, + 0.47, + 0.593 + ], + "angle": 0, + "content": "For large-image augmentation, apart from TCGA-BRCA, we also use the BACH dataset [33]. Introduced in the ICIAR 2018 Grand Challenge, the dataset contains 400 H&E-stained Breast Cancer images of size \\(2048 \\times 1536\\) pixels, evenly distributed across four categories: normal, benign, in-situ carcinoma, and invasive carcinoma." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.291, + 0.619 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.469, + 0.716 + ], + "angle": 0, + "content": "For all our experiments, we train the LDM on \\(256 \\times 256\\) pixel patches, following PathLDM [48], which fine-tunes an ImageNet-trained [38] U-Net denoiser and uses a \\(4 \\times\\) downsampling VQ-VAE, instead of the default LDM configuration. These modifications were deemed necessary for applying LDMs on large-image domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.717, + 0.469, + 0.839 + ], + "angle": 0, + "content": "We train our models on 6 NVIDIA RTX 8000 GPUs, with a batch size of 100 per GPU, utilizing code and pretrained checkpoints from LDM [37]. We set the learning rate at \\(10^{-4}\\) with a warmup of 10,000 steps. We apply DDIM with 50 steps and a guidance scale of 1.75 for both patch sampling and large-image generation. When generating large images we apply the patch diffusion model with a stride \\(>1\\) depending on the desired target quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.276, + 0.864 + ], + "angle": 0, + "content": "4.3. Image quality results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In Fig. 3, we present synthetic patches and large images from our TCGA-BRCA and NAIP models, along with the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.593 + ], + "angle": 0, + "content": "corresponding references from which the self-supervised embeddings were extracted. We evaluate our method's per-patch and large-image generation quality by computing FID scores [14] using the Clean-FID implementation [31]. We generate 10,000 patches \\((256 \\times 256)\\) and 3,000 large images \\((1024 \\times 1024)\\) from diffusion models trained on the TCGA-BRCA, TCGA-CRC, and NAIP datasets. Since our generative model requires a self-supervised conditioning vector \\(y\\) (or multiple vectors \\(z\\)) for each synthetic image (or large image), we randomly sample embeddings from reference images in the training set to generate images for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.701 + ], + "angle": 0, + "content": "For patches, we measure FID against the real images (\"Vanilla FID\"). For large images, we follow the evaluation strategy of MultiDiffusion [3] and use FID to compare the distribution of \\(256 \\times 256\\) crops from synthesized large images to that of real image patches of the same size (\"Crop FID\"). We also measure FID directly between the large images and ground truth data using CLIP [34] (\"CLIP FID\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.778 + ], + "angle": 0, + "content": "To evaluate the similarity between the reference and generated large images, we resize from \\(1024 \\times 1024\\) to \\(256 \\times 256\\) and use an SSL model to extract embeddings. We compute the cosine similarity between the paired embeddings (\"Embedding Similarity\")." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Patch-level quality: Our models achieve low patch FID scores across all datasets (Tab. 1). For TCGA-BRCA patches, our \"Vanilla FID\" of 6.98 is comparable to the current state-of-the-art [48] (7.64 at \\(10\\times\\)). We attain similar, low FID scores for the smaller CRC and NAIP datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Synthetic large image quality: We present the \"Crop FID\", \"CLIP FID\" and \"Embedding Similarity\" results for large images in Table 1. The \"Crop FID\" of the large im" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8536" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.083, + 0.089, + 0.465, + 0.192 + ], + "angle": 0, + "content": "
Dataset# Training imagesPatch levelLarge image level
Vanilla FIDCrop FIDCLIP FIDEmb similarity
BRCA 20x15M6.9815.517.430.924
CRC 20x8M6.788.87.340.938
NAIP667k11.543.766.86-
BRCA 5x976k9.74-6.64-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.204, + 0.47, + 0.261 + ], + "angle": 0, + "content": "Table 1. FID scores for our generated patch and large images. Our patch-level BRCA model is on par with SoTA [48] (7.64 at \\(10\\times\\)). \"CLIP FID\" and \"Embedding Similarity\" demonstrate our large images' realism and contextual accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.287, + 0.468, + 0.437 + ], + "angle": 0, + "content": "ages is comparable to the \"Vanilla FID\" on the BRCA and CRC datasets, showing that patches from synthesized large images have similar semantic content to the ground truth patches. We attribute the worse \"Crop FID\" for the NAIP model to the limited number of samples available for both SSL and diffusion model training. The \"Crop FID\" is consistently higher than the patch-level FID, which is expected as the large-image generation framework only approximates the distribution of the large images and does not have access to the true conditioning at every location." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.468, + 0.603 + ], + "angle": 0, + "content": "Our low \"CLIP FID\" scores indicate that the generated large images are similar to real images when resized to \\(224 \\times 224\\) pixels. This indicates that our SSL-guided large-image generation successfully retains the larger-scale semantic arrangements of real data. For NAIP, our model is not as good at synthesizing high-frequency details, which explains the large discrepancy between \"Crop FID\" and \"CLIP FID\". Additionally, when comparing the \"CLIP FID\" of large images synthesized by the \\(20 \\times\\) BRCA model and resized to \\(5 \\times\\), to images from a model trained directly on \\(5 \\times\\) data, we see minimal difference (7.43 vs. 6.64)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.468, + 0.725 + ], + "angle": 0, + "content": "We evaluate the contextual similarity between synthetic and reference large images for the BRCA and CRC data. We compute cosine similarity between large images using CTransPath embeddings. Our BRCA and CRC models demonstrate high \"Embedding Similarity\" scores of 0.924 and 0.938, respectively, reflecting our framework's effectiveness in preserving the integrity of context and key features on the large-image scale." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.388, + 0.758 + ], + "angle": 0, + "content": "5. Image Augmentation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.468, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, apart from visual fidelity, the synthetic images preserve the intricate characteristics of the reference images, both on the patch and large-image scale. These characteristics include the nuanced textural elements, histological staining, and cell structure. This correspondence between real and synthetic images demonstrates the detailed and varied information captured by the self-supervised embeddings used as conditioning. In conjunction with the high \"Embedding Similarity\", it justifies using variations of im" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.121 + ], + "angle": 0, + "content": "ages generated from SSL embeddings for patch and large-image level data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.892, + 0.214 + ], + "angle": 0, + "content": "Having a powerful generative model enables us to perform data augmentation for tasks where we can control the augmented image label using conditioning. In our setting, our diffusion models are not trained with class labels; instead, we synthesize a novel image using the conditioning from a reference patch or region." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.892, + 0.32 + ], + "angle": 0, + "content": "We assume that i) the self-supervised embedding used as conditioning contains information about the target label and ii) the diffusion-generated variations of an image do not alter this target label information. We validate both assumptions experimentally, by augmenting training sets on patch and large-image level tissue classification tasks, including a Multiple Instance Learning (MIL) task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.443 + ], + "angle": 0, + "content": "Large-image augmentation on TCGA-BRCA: We examine two histopathology slide-level binary classification tasks on TCGA-BRCA: BRCA Subtyping (Invasive Ductal Carcinoma (IDC) vs Invasive Lobular Carcinoma (ILC)) and HRD prediction. We utilize a minimal dataset, just \\(10\\%\\) of real WSI data (100 WSIs), to train MIL algorithms. We generate an equal set of synthetic images for 100 additional WSIs using training set images as reference." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.626 + ], + "angle": 0, + "content": "We employ 10-fold cross-validation to divide our dataset into training and testing segments. Within each fold, two multiple instance learning (MIL) models, CLAM-SB [25] and DSMIL [21], are trained on two sets: one with real data and another combining real and synthetic data. To train the MIL models, we extract features using the CTransPath ViT [44]. The results, detailed in Table 2, indicate that models trained on the augmented datasets consistently surpass their real-only counterparts, regardless of the MIL algorithm used. This demonstrates the value of the synthetic images generated by our method, confirming their efficacy as comparable to real images for training purposes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.763 + ], + "angle": 0, + "content": "Large-image augmentation on BACH: We double the training set of BACH [33] by adding as many synthetic large-images, produced by the TCGA-BRCA diffusion model. From each \\(2048 \\times 1536\\) pixel training set image, we extract a \\(8 \\times 6\\) SSL embedding grid to generate a variation with the same label. For classification, we employ a ConvNeXt V2huge [46] model pre-trained on ImageNet. We train a 2-layer MLP classifier on top of the penultimate layer features and evaluate it on the official test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results, presented in Table 3a, reveal a notable improvement in classifier performance, from \\(78\\%\\) to \\(83\\%\\). This improvement again confirms the high quality of our synthetic data while also highlighting the versatility of our approach. Despite being trained on \\(256 \\times 256\\) patches from TCGA-BRCA, the model generalizes to produce realistic large images from a completely different dataset. We attribute this generalization capability to the expressiveness of the SSL features and the potency of the diffusion model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8537" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.894, + 0.171 + ], + "angle": 0, + "content": "
TCGA-BRCA SubtypingTCGA-BRCA HRD
Method1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real
CLAM-SB0.7250.8120.8860.8980.910.6030.6440.6490.7650.787
DSMIL0.6090.6590.8380.8560.9050.5170.5540.5630.6390.669
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.182, + 0.893, + 0.211 + ], + "angle": 0, + "content": "Table 2. The inclusion of synthetic data consistently enhances AUC across various MIL architectures and BRCA tasks. The dataset contains 1000 real images, so “\\(10\\%+\\) synthetic” indicates training with 100 real and 100 synthetic WSIs, with the remainder used for testing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.237, + 0.47, + 0.299 + ], + "angle": 0, + "content": "to accurately portray them in images. While our model does not reach the current SoTA accuracy of \\(87\\%\\), achieved by an ensemble approach [7], the simplicity and ability to integrate with other models make for a noteworthy contribution." + }, + { + "type": "table", + "bbox": [ + 0.083, + 0.313, + 0.268, + 0.385 + ], + "angle": 0, + "content": "
Training DataTest Acc
Real78 %
Synthetic70 %
Real + synthetic83 %
SoTA [7]87 %
" + }, + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.388, + 0.182, + 0.399 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.276, + 0.313, + 0.462, + 0.385 + ], + "angle": 0, + "content": "
Training DataVal Acc
Real93.8 %
Synthetic90.19 %
Real + synthetic96.27 %
SoTA [20]96.26 %
" + }, + { + "type": "table_caption", + "bbox": [ + 0.358, + 0.387, + 0.376, + 0.398 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.404, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Table 3. Our data augmentations provide notable improvements for the BACH (a) and CRC-VAL-HE (b) datasets. Notably, the diffusion training data does not overlap with the data of the augmented datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Patch-level image augmentation: We further investigate our out-of-distribution generalization capabilities by augmenting the NCT-CRC dataset [19]. We leverage our diffusion model trained on the TCGA-CRC data, which does not overlap with NCT-CRC, conditioned on embeddings from an iBOT ViT [11]. We generate an augmented dataset of equal size to the original by using the SSL embedding of every patch in the NCT-CRC training set to synthesize a corresponding image of the same label." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.471, + 0.812 + ], + "angle": 0, + "content": "We train an ImageNet pre-trained ResNet-50 [12] network on three splits: real images only, synthetic images only, and a combination of both. Evaluation on the CRC-VAL-HE-7K test set demonstrates a significant performance increase when synthetic data is introduced, with classifier accuracy rising from \\(93.8\\%\\) to \\(96.27\\%\\) as presented in Table 3b. We match the current SoTA [20], which used an ensemble of deep models, with a model agnostic approach. As in the previous experiment, the diffusion model, now trained on the significantly smaller TCGA-CRC data, can also effectively synthesize images from a completely different dataset by only controlling the self-supervised conditioning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.345, + 0.846 + ], + "angle": 0, + "content": "6. Text-to-large image synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For previous tasks, our large image generation approach synthesizes variations of an existing set of images from the pre-computed self-supervised embeddings. In Sec. 3.3 we" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.237, + 0.892, + 0.373 + ], + "angle": 0, + "content": "discussed how to control large image generation with auxiliary signals from any domain (class labels, text captions, etc.), by training models \\( p(z \\mid c) \\) that can be combined with the embedding-conditioned image generation. We demonstrate controllable image synthesis with text-to-large image generation experiments on histopathology and satellite data. We measure the similarity between synthetic images and the text prompts used in generating them using vision-language models (VLMs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.893, + 0.559 + ], + "angle": 0, + "content": "Text-to-large histopathology images: We utilize the CRC and BRCA diffusion models to generate \\(1024 \\times 1024\\) pixel images from text prompts. We first construct training sets by pairing \\(4 \\times 4\\) SSL embedding grids \\(z\\) from large BRCA and CRC images, with their corresponding Quilt [18] image embeddings \\(c\\). We then train an auxiliary diffusion model to sample \\(p(z \\mid c)\\). During inference, we use a text embedding \\(c'\\) as a proxy for the image embedding, to sample \\(z\\) and synthesize a large image. To bridge the gap between the image and text Quilt embeddings [23, 30] we perturb the image embeddings when training the auxiliary diffusion model with Gaussian noise of variance \\(\\sigma^2 = 0.1\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.895, + 0.852 + ], + "angle": 0, + "content": "To evaluate the text-to-large image pipeline, we generate images from a pre-defined set of classes described in natural language; non-malignant benign tissue, malignant in-situ carcinoma, malignant invasive carcinoma, normal breast tissue for BRCA and colon adenocarcinoma, benign colonic tissue for CRC. We construct the confusion matrix of zero-shot classifiers on the synthesized data. We used two different VLMs as zero-shot classifiers, Quilt and BiomedCLIP [49]. The results presented in Fig. 4 demonstrate our ability to synthesize images consistent with the text prompts. The capabilities of the VLM used, limit our synthetic image generation. The lower performance of BRCA vs. CRC is consistent with the results reported in Quilt [18]. Furthermore, we asked an expert pathologist to classify 100 synthetic CRC images as benign or adenocarcinoma images. Their evaluation showed an \\(89.9\\%\\) agreement rate with the labels used for image generation, indicating consistency in our text-to-large image pipeline. We show examples of synthesized images in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Text-to-large satellite images: To synthesize novel satellite images we first create a training set of \\(30\\mathrm{k}1024\\times 1024\\) pixel large NAIP images, and pair them with captions" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.274 + ], + "angle": 0, + "content": "from a BLIP model [22]. We train a diffusion model to sample the \\(4 \\times 4\\) SSL embeddings \\(z\\) from the captions \\(c\\). To evaluate, we create a separate set of 1000 NAIP image-caption pairs and measure the CLIP similarity between generated images and the given captions. We achieve a CLIP similarity score of 0.22, showing that we can effectively learn the mapping from text to large images with this hierarchical approach. Although our CLIP similarity is slightly worse than the scores reported for text-to-image Stable Diffusion models (\\(>0.24\\)) [32], we expect this drop in performance as we trained with machine-generated captions. Training and generated image-caption pairs are provided in supplementary." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.283, + 0.473, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.528, + 0.471, + 0.556 + ], + "angle": 0, + "content": "Figure 4. Confusion matrix of zero-shot classification for novel TCGA-CRC and TCGA-BRCA synthetic images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.586, + 0.47, + 0.62 + ], + "angle": 0, + "content": "7. Combining self-supervised embeddings with diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.856 + ], + "angle": 0, + "content": "We further evaluate our patch-generating models by posing the question does the diffusion model learn more about the data than the self-supervised learning model? We hypothesize that by combining the pre-trained self-supervised embedding with the denoising task we can improve the learned representations of the data, leading to better performance in downstream tasks, which are performed with features from self-supervised learning. To validate this hypothesis, we utilize the trained diffusion model as a feature extractor and apply a Multiple-instance learning (MIL) approach for the slide-level classification task of subtyping Breast Cancer. For each patch in a whole-slide image (WSI) we first extract the self-supervised embeddings, which are then used as conditioning to obtain features from the denoiser's U-Net bottleneck layer at a fixed timestep \\( t = 50 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In Tab. 4, we evaluate the effectiveness of our novel fusion of generative diffusion and self-supervised embeddings. We compare the performance of MIL algorithms [21," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "[25] using features derived from our approach at \\(20 \\times\\) and \\(5 \\times\\) magnification (LDMHIPT/CTransPath) against using only the self-supervised conditioning (HIPT/ CTransPath). We used a 10-fold cross-validation strategy consistent with the data splits from HIPT [6], training the MIL algorithms on both the full dataset (100%) and a reduced subset (25%)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.893, + 0.29 + ], + "angle": 0, + "content": "The results indicate that integrating self-supervised features into the diffusion model as conditioning leads to learning better representations and improves whole-slide classification. By fusing the generative knowledge from the diffusion process with the discriminative capabilities of the self-supervised embeddings, we construct a successful model for both discriminative and generative tasks." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.306, + 0.89, + 0.392 + ], + "angle": 0, + "content": "
MagFeatures25% training100% training
CLAM-SBDSMILCLAM-SBDSMIL
20×HIPT0.7880.7840.8610.839
LDMHIPT0.8420.7950.9080.894
5×CTransPath0.9000.8960.9190.910
LDMCTransPath0.9130.9050.9230.936
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Table 4. 10-fold cross-validation AUC for BRCA Histological subtyping. \\(\\mathrm{LDM}_{\\mathrm{HIPT}}\\) denotes diffusion features conditioned on HIPT embeddings. The fusion of SSL and diffusion features outperforms the SSL features by themselves." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.619, + 0.513 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.797 + ], + "angle": 0, + "content": "We presented a novel approach to training diffusion models in large-image domains, such as digital histopathology and remote sensing. We overcome the need for fine-grained annotation by introducing self-supervised representation guided diffusion models, achieving remarkable image synthesis results on the patch level. Our approach also enables us to synthesize high-quality large images, where we have the ability to dictate the global structure by controlling the spatial arrangement of the conditions. We evaluated the usefulness of our synthetic images on a number of patch and large image-level tasks, as well as introduced a text-to-large image generation framework. Naively augmenting whole-slide images is a time-consuming process. We leave to future work the exploration of adaptive augmentation strategies that choose which image parts to augment. We believe these results illustrate the great potential for this technology to lead to bespoke foundational models for specialized domains, comparable to existing models for natural images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements This research was partially supported by NCI awards 5U24CA215109, 1R21CA258493-01A1, UH3CA225021, NSF grants IIS-2123920, IIS-2212046 and Stony Brook Profund 2022 seed funding. We thank Rajarsi Gupta for his valuable feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8539" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.211 + ], + "angle": 0, + "content": "[1] Marco Aversa, Gabriel Nobis, Miriam Hagele, Kai Standvoss, Mihaela Chirica, Roderick Murray-Smith, Ahmed Alaa, Lukas Ruff, Daniela Ivanova, Wojciech Samek, et al. Diffinfinite: Large mask-image synthesis via parallel random patch diffusion in histopathology. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.254 + ], + "angle": 0, + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.47, + 0.324 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1737-1752. PMLR, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[4] JN Cancer Genome Atlas Research Network et al. The cancer genome atlas pan-cancer analysis project. Nat. Genet, 45 (10):1113-1120, 2013. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.437 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[7] Sai Saketh Chennamsetty, Mohammed Safwan, and Varghese Alex. Classification of breast cancer histology image using ensemble of pre-trained neural networks. In Image Analysis and Recognition: 15th International Conference, ICIAR 2018, Póvoa de Varzim, Portugal, June 27-29, 2018, Proceedings 15, pages 804-811. Springer, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.649, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[10] Miguel Espinosa and Elliot J Crowley. Generate your own scotland: Satellite image generation conditioned on maps. arXiv preprint arXiv:2308.16648, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[11] Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathology with masked image modeling. medRxiv, pages 2023-07, 2023. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[17] Vincent Tao Hu, David W Zhang, Yuki M Asano, Gertjan J Burghouts, and Cees GM Snoek. Self-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18413-18422, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] Wisdom Oluchi Ikezogwo, Mehmet Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Stefan Chan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. *Quilt-1m: One million image-text pairs for histopathology.* arXiv preprint arXiv:2306.11207, 2023. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Jakob Nikolas Kather, Niels Halama, and Alexander Marx. 100,000 histological images of human colorectal cancer and healthy tissue, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Anurodh Kumar, Amit Vishwakarma, and Varun Bajaj. Crccn-net: Automated framework for classification of colorectal tissue using histopathological images. Biomedical Signal Processing and Control, 79:104172, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.563, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[21] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[23] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8540" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.176 + ], + "angle": 0, + "content": "[26] Puria Azadi Moghadam, Sanne Van Dalen, Karina C Martin, Jochen Lennerz, Stephen Yip, Hossein Farahani, and Ali Bashashati. A morphology focused diffusion probabilistic model for synthesis of histopathology images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2000-2009, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[27] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, et al. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13 (1):12098, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.328 + ], + "angle": 0, + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[30] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.472, + 0.496 + ], + "angle": 0, + "content": "[31] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.609 + ], + "angle": 0, + "content": "[33] António Polónia, Catarina Eloy, and Paulo Aguiar. BACH Dataset : Grand Challenge on Breast Cancer Histology images, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.61, + 0.472, + 0.691 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.472, + 0.748 + ], + "angle": 0, + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[36] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.331 + ], + "angle": 0, + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[41] Ahmad Sebaq and Mohamed ElHelw. Rsdiff: Remote sensing image generation from text using diffusion model. arXiv preprint arXiv:2309.02455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.893, + 0.418 + ], + "angle": 0, + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.489 + ], + "angle": 0, + "content": "[43] USGS. National agriculture imagery program (NAIP), 2023. https://www.usgs.gov/centers/eros/science/usgs-eros-archive-aerial-photography - national - agriculture - imagery-program-naip.2,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.893, + 0.602 + ], + "angle": 0, + "content": "[44] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Junzhou Huang, Wei Yang, and Xiao Han. Transpath: Transformer-based self-supervised learning for histopathological image classification. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VIII 24, pages 186-195. Springer, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.672 + ], + "angle": 0, + "content": "[45] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis, 81:102559, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.893, + 0.757 + ], + "angle": 0, + "content": "[46] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.758, + 0.893, + 0.814 + ], + "angle": 0, + "content": "[47] Xuan Xu, Saarthak Kapse, Rajarsi Gupta, and Prateek Prasanna. Vit-dae: Transformer-driven diffusion autoencoder for histopathology image analysis. arXiv preprint arXiv:2304.01053, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[48] Srikar Yellapragada, Alexandros Graikos, Prateek Prasanna, Tahsin Kurc, Joel Saltz, and Dimitris Samaras. PathLDM: Text conditioned latent diffusion model for histopathology. arXiv preprint arXiv:2309.00748, 2023. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[49] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8541" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "Valluri, Cliff Wong, et al. Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915, 2023. 7" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8542" + } + ] +] \ No newline at end of file diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_origin.pdf b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..979fb12bd77666f75d3c45c7c656331c6a5e8c73 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/5695d81c-8814-42a9-899f-2c8baf3ccbb3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80303989b92d5e802d4afa076bb19b4d30f9a191d5c41d5b80be7084e18bfc87 +size 2116073 diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/full.md b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e8af0333f9bf33b752f74b21a01675b7551bb59 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/full.md @@ -0,0 +1,311 @@ +# Learned representation-guided diffusion models for large-image generation + +Alexandros Graikos* Srikar Yellapragada* Minh-Quan Le Saarthak Kapse Prateek Prasanna Joel Saltz Dimitris Samaras Stony Brook University + +# Abstract + +To synthesize high-fidelity samples, diffusion models typically require auxiliary data to guide the generation process. However, it is impractical to procure the painstaking patch-level annotation effort required in specialized domains like histopathology and satellite imagery; it is often performed by domain experts and involves hundreds of millions of patches. Modern-day self-supervised learning (SSL) representations encode rich semantic and visual information. In this paper, we posit that such representations are expressive enough to act as proxies to fine-grained human labels. We introduce a novel approach that trains diffusion models conditioned on embeddings from SSL. Our diffusion models successfully project these features back to high-quality histopathology and remote sensing images. In addition, we construct larger images by assembling spatially consistent patches inferred from SSL embeddings, preserving long-range dependencies. Augmenting real data by generating variations of real images improves downstream classifier accuracy for patch-level and larger, image-scale classification tasks. Our models are effective even on datasets not encountered during training, demonstrating their robustness and generalizability. Generating images from learned embeddings is agnostic to the source of the embeddings. The SSL embeddings used to generate a large image can either be extracted from a reference image, or sampled from an auxiliary model conditioned on any related modality (e.g. class labels, text, genomic data). As proof of concept, we introduce the text-to-large image synthesis paradigm where we successfully synthesize large pathology and satellite images out of text descriptions. + +# 1. Introduction + +Diffusion models produce high-quality and diverse samples across a spectrum of generative tasks [8, 24]. This leap forward has been enabled by the simultaneous curation of large-scale multi-modal datasets [40] and the development + +![](images/26a83e1423cf49b6cdaa97b008ab8b1a43ea216117eed332430fd6ae51c85da4.jpg) +Figure 1. We propose using SSL features to condition diffusion models. This allows us to construct large images by assembling consistent patches inferred from a spatial arrangement of SSL embeddings. The generated image retains the semantics of the embeddings used as a condition, maintaining the forested and open areas from the reference. Best viewed zoomed-in. + +of efficient conditioning mechanisms [35, 37]. The key to unlocking the models' capabilities is to integrate auxiliary information during training and inference [8, 15, 28]. + +Large-scale human-annotated datasets are mostly limited to image-caption pairs, collected from easily accessible online repositories and labeled by non-expert annotators. However, in domains such as digital histopathology and remote sensing, where gigapixel scale images provide vast amounts of unlabeled data, annotation proves challenging. Moreover, the process requires expert knowledge and is more difficult at a finer scale, i.e., captioning large crops of the gigapixel image is simpler than captioning smaller patches. Based on our estimates (see supplemental), annotating the entire TCGA-BRCA dataset with captions would take $\approx 40,000$ hours of pathologist's time. Replicating the impressive results of diffusion models in these domains has been limited by the scarcity of fine-grained per-image conditioning, vital for high-quality image synthesis [29]. + +Modern-day self-supervised learning (SSL) representations [5] encode rich semantic and visual information. Features from trained self-supervised models serve as compact image representations and are widely used to perform discriminative downstream tasks successfully [6, 11, 45], proving that these compressed representations indeed encode + +useful semantic information about the images. We hypothesize that such SSL representations are already expressive enough to act as proxies to fine-grained human labels. If this is true, these representations should be able to condition the training of effective diffusion models in these domains. In this novel approach, we utilize self-supervised feature extractors as image annotators; these features provide necessary per-image conditioning signals at the highest resolutions, required for diffusion model training. + +Our experiments show that conditioning with expressive self-supervised features leads to precise control over the image content. The SSL features are adept at identifying complex patterns and structures in the images, while the diffusion model learns to translate them into visual components accurately (Fig. 3). This motivates us to perform large-image synthesis by locally controlling the appearance using SSL conditioning and dictating the global structure through the spatial arrangement of the conditions. + +Our approach synthesizes large images in a patch-based manner, using a single image diffusion model at the highest resolution. We represent a large image as a grid of SSL embeddings, where each serves as a representation of a large image neighborhood. The whole image is then synthesized by generating consistent patches that capture both the local properties, as given by the local patch conditioning, and the spatial arrangement of the conditioning features. If we change this spatial arrangement, we are, in fact, editing how semantic elements are arranged globally in the large image. This strategy enables the controllable generation of images of virtually any size without significantly increased computation compared to the base patch-level model. + +To generate a large image, our approach requires the patch diffusion model and the spatially-arranged conditioning. We can start with a reference large image as source and extract SSL embeddings from non-overlapping segments, enabling our method to synthesize a variation of the original image (Fig. 1). + +Utilizing SSL embeddings as conditions allows us to have the necessary control over image generation, at the expense of an explainable and easy-to-use conditioning mechanism. Nevertheless, we argue that since generating images from learned embeddings is agnostic to the embedding source, there are simple ways to combine control over generated images with explainability. We propose training auxiliary models to transform higher-level conditioning signals, such as text captions, to the learned patch representations. To demonstrate this versatility, we introduce text-to-large-image synthesis by training an auxiliary model to sample a spatial arrangement of embeddings from a text description. + +We train patch-level diffusion models using self-supervised features as conditioning on digital histopathology (TCGA [4]) and satellite image (NAIP [43]) datasets. We perform extensive evaluations and demonstrate the ad + +vantages of SSL conditioning and our large-image generation framework on synthesis and classification tasks. Our model achieves exceptional patch-level and large-image quality, the ability to improve classifiers through data augmentation even when synthesizing out-of-distribution data, and effective fusion of diffusion and SSL features for downstream applications. Finally, we are the first to perform text-to-large image synthesis, which should be of significant community interest as vision-language models (VLMs) for pathology and satellite images gain traction. + +In summary, our contributions are as follows: + +- We develop a novel method to train diffusion models with self-supervised learning features as conditioning and generate high-quality images in the histopathology and satellite image domains. +- We present a framework for large-image synthesis, based on self-supervised guided diffusion, that maintains contextual integrity and image realism over large areas. +- We demonstrate the applicability of our model in various classification tasks and showcase its unique ability to augment out-of-distribution datasets. +- We introduce text-to-large image generation for digital histopathology and satellite images, highlighting the versatility of our approach. + +# 2. Related work + +Diffusion models: Introduced for image generation by Ho et al. [16], diffusion models have evolved considerably. These enhancements include class conditioning [28], architectural improvements and gradient-based guidance [8], and classifier-free guidance [15]. Latent Diffusion Models (LDMs) [37] proposed a two-step training process with a Variational Autoencoder (VAE) compressing input images into a lower-dimensional latent space and a diffusion model trained in this latent space. Denoising Diffusion Implicit Models (DDIM) [42] accelerate the sampling process by $10 - 50 \times$ . Self-guided diffusion models [17] also utilize self-supervised learning by quantizing SSL embeddings. In contrast to our approach, their quantization discards useful information from the SSL embeddings. + +Training generative models directly at the gigapixel resolution is infeasible. An alternative is to generate the images in a coarse-to-fine manner hierarchically. This has already been applied to natural images, where chaining multiple diffusion models generates images up to $1024 \times 1024$ resolution [32, 39]. However, it is still limited, as it substantially increases the parameter count and is inherently constrained by the final target resolution. + +In the context of digital histopathology, works have been limited to training unconditional [26] or class-conditioned [27, 47] diffusion models. For text conditioning, pathology text reports were used to provide context on the whole-slide scale [48]. Similar approaches have been + +![](images/65afc4a7a334f2e01696fa705fe49eb3c0cc1ba8c0d1d3d78050bf41e73fd57b.jpg) +Figure 2. (a) We train diffusion models on patches $I$ (e.g. the one in the green box) taken from a large image conditioned on SSL embeddings. (b) We present our large image generation framework in 4 steps: (i) We extract a set of spatially arranged embeddings from a reference image or sample them from an auxiliary model. (ii) For every location $(i,j)$ , we compute a conditioning vector $\lambda_{i,j}$ by interpolating the spatial grid of embeddings. (iii) At every diffusion step, we denoise the patch $F(i,j)$ using the conditioning $\lambda_{i,j}$ . (iv) The next step is computed by averaging the denoising updates of all patches that overlap at $(i,j)$ . + +applied to satellite data [10, 41]. Apart from high-quality images without manual annotation, our SSL conditioning is also necessary for large-image synthesis, as all previous conditioning methods would not be sensitive to the intricate differences between neighboring patches. In recent work, DiffInfinite [1] explored large-image generation using segmentation masks as conditioning. We argue this is still suboptimal as it requires accurate, human-annotated masks for training and a mask-generating model during inference. + +Self-Supervised Learning Self-supervised learning (SSL) refers to both discriminative [2, 13] and distillation [5] approaches that aim to learn representations of the data without supervision. In this work, we are mainly interested in self-supervised learning for histopathology. Notable recent developments include the Hierarchical Image Pyramid Transformer (HIPT) [6], which utilizes the inherent hierarchical structure of Whole Slide Images (WSIs), CTransPath [45], a hybrid CNN and multi-scale Swin Transformer model, and iBOT [11], a masked image modeling method. These models excel as patch-level feature extractors for WSIs by leveraging the unique structure of the large-image data and capture the important semantic information that we require for conditioning a generative model. + +# 3. Method + +We propose training a diffusion model on patches $I$ drawn from large histopathology and satellite images, using self + +supervised embeddings $y$ as conditioning. Furthermore, we present a patch-based approach that utilizes the SSL-conditioned diffusion model to synthesize arbitrarily large images. An overview of our method is presented in Fig. 2. + +# 3.1. Learned representation-guided diffusion + +Given a pre-trained self-supervised feature extractor, we employ LDMs [37] to learn the distribution over the large-image patches $p(\mathcal{I})$ . LDMs are comprised of three components: an image-compressing Variational Autoencoder (VAE) that transforms the input images to latent representations, a U-Net denoiser to learn a denoising diffusion process that transforms Gaussian noise to latents, and a conditioning mechanism that controls the diffusion process. The conditioning is performed in our setting with a single vector $y$ , obtained from the self-supervised model, integrated via a cross-attention mechanism. We train the LDM using pairs of image patches $I$ and the corresponding extracted self-supervised embeddings $y$ . + +# 3.2. Large image generation + +Our goal is to synthesize high-quality large images, that not only capture global structure but also maintain spatial consistency. As shown in Fig. 2, we replicate the semantics in each patch with SSL-guided LDM. At the same time, we preserve the global arrangement of these semantics as defined by the grid of patches in the reference image. Future + +work can explore alternative approaches to ensure spatial alignment, including using topological constraints or integrating low-resolution information. + +We follow the MultiDiffusion [3] methodology to generate large images using only a patch-based diffusion model. We can represent the diffusion model as a learned mapping from images and conditions to images: + +$$ +\Phi : \mathcal {I} \times \mathcal {Y} \rightarrow \mathcal {I}, \tag {1} +$$ + +where $\mathcal{I} = \mathbb{R}^{H\times W\times C}$ are the large-image patches and $\mathcal{V} = \mathbb{R}^d$ are the per-patch conditions. To generate a patch, we initialize $I_{T}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and sequentially transform to a "clean" image $I_0$ following the trained LDM: + +$$ +I _ {t - 1} = \Phi \left(I _ {t} \mid y\right) \tag {2} +$$ + +for $t = T, T - 1, \ldots, 1$ . We assume that a large image can be formed as a spatial grid of $P \times P$ patches of size $H \times W$ . We then define the large-image diffusion model as + +$$ +\Psi : \mathcal {J} \times \mathcal {Z} \rightarrow \mathcal {J} \tag {3} +$$ + +where $\mathcal{J} = \mathbb{R}^{PH\times PW\times C}$ are the large images and $\mathcal{Z} = \underbrace{\mathcal{Y}\times\mathcal{Y}\cdots\times\mathcal{Y}}_{P^2}$ are the conditioning vectors of all patches. + +The process $\Psi(J_t \mid z) = J_{t-1}$ , with $J_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , can be approximated by first defining mappings between the two different image and conditioning spaces: + +$$ +F _ {i, j} ^ {\text {p r o j}}: \mathcal {J} \rightarrow I, \quad \lambda_ {i, j}: \mathcal {Z} \rightarrow \mathcal {Y} \tag {4} +$$ + +where, for the case of large-image generation, we set $F_{i,j}^{\mathrm{proj}}$ to be a crop (projection) of the large image, centered at $i,j$ , and $\lambda_{i,j}$ the conditioning $y_{i,j} \in \mathbb{R}^d$ at $i,j$ . Since we only have conditioning vectors at the centers of the patches, we can use spatial interpolation algorithms to implement $\lambda$ . This assumes that the interpolant $y_{i,j}$ is a valid conditioning vector for the diffusion process, which we validate experimentally. Then, $\Psi$ is defined as + +$$ +\Psi \left(J _ {t} \mid z\right) = \underset {J \in \mathcal {J}} {\arg \min } \sum_ {i, j} \| F _ {i, j} ^ {\operatorname {p r o j}} (J) - \Phi \left(F _ {i, j} ^ {\operatorname {p r o j}} \left(J _ {t}\right) \mid \lambda_ {i, j} (z)\right) \| ^ {2} \tag {5} +$$ + +which can be solved in closed-form by setting each pixel $i,j$ of $J$ to the average of all the patch-diffusion updates + +$$ +\Psi \left(J _ {t} \mid z\right) = \sum_ {i, j} \frac {F _ {i , j} ^ {\text {u n p r o j}} (\mathbf {1})}{\sum_ {k , l} F _ {k , l} ^ {\text {u n p r o j}} (\mathbf {1})} \otimes F _ {i, j} ^ {\text {u n p r o j}} \left(\Phi \left(F _ {i, j} ^ {\text {p r o j}} \left(J _ {t}\right) \mid \lambda_ {i, j} (z)\right)\right) \tag {6} +$$ + +where $\mathbf{1}$ denotes a patch image where all values are set to 1 and $F_{i,j}^{\mathrm{unproj}}$ is the inverse mapping of pixels from the crop centered at $i, j$ back to the large image. + +We are able to generate images larger than the ones produced by the patch diffusion model. At the same time, we + +control what each patch looks like, which is crucial in maintaining the semantic integrity of the larger image. The self-supervised conditions can capture the variations between neighboring patches necessary for producing realistic results. Generating large images with coarser conditioning, such as global text prompts [48], would lead to uniform texture regions (see supplementary). + +# 3.3. Controllable large-image synthesis + +Although the grid of self-supervised embeddings $z$ cannot be manipulated in a human-interpretable manner, we argue that it is simple to assert more control over the generated images. As illustrated in Fig. 2, this control can be attained by training auxiliary models $p(z \mid c)$ that translate higher-level conditioning signals, such as text captions $c$ , to learned patch representations $z$ . + +Since there is no available dataset of paired large images and captions, we resort to pre-trained multi-modal models, such as Quilt [18], CLIP [34] and BLIP [22], to provide the text conditioning. We construct training sets by extracting self-supervised embeddings from training-set large images and pairing them with multi-modal image embeddings or generated captions. During inference, we first sample SSL embeddings from the learned distribution $p(z \mid c)$ , then utilize our patch diffusion models to synthesize a large image. + +# 4. Image Generation Experiments + +# 4.1. Datasets + +We train diffusion models on digital histopathology images from The Genome Cancer Atlas (TCGA) [4] and satellite imagery from the National Agriculture Imagery Program (NAIP) [43]. Specifically, we used the TCGA-BRCA (Breast Invasive Carcinoma Collection), TCGA-CRC (COAD + READ Colorectal Carcinoma) datasets, and the Chesapeake Land Cover dataset [36]. + +For the TCGA-BRCA and TCGA-CRC datasets, we use images at $20\times$ magnification, and developed diffusion models conditioned on embeddings from HIPT [6] and iBOT [11], which were pre-trained on PanCancer TCGA. In the case of the HIPT model, we specifically used its patch-level ViT. Additionally, for the TCGA-BRCA dataset, we train a model at $5\times$ magnification using embeddings from CTransPath [44] for additional evaluations. + +The Chesapeake Land Cover dataset contains 732 NAIP tiles, each measuring a $6\mathrm{km} \times 7.5\mathrm{km}$ area at $1\mathrm{m}$ resolution. We extract $256 \times 256$ non-overlapping pixel patches, resulting in 667,000 patches total. Given the absence of publicly available self-supervised learning models tailored to the NAIP data, we train a Vision Transformer (ViT-B/16) [9] using the DINO framework [5]. We then use the learned DINO embeddings to train the diffusion model on pairs of image patches and self-supervised embeddings. + +![](images/ba1d5d338f16fefd0c738ebe359bf2b54dc18718b0713ebc4ff6e088aa0e890b.jpg) + +![](images/5dc80e3652ccf483806329ff59547965b5b0632cdbea960a4c020f9bd7b016b7.jpg) +Figure 3. (Top) Patches $(256 \times 256)$ from our models, and the corresponding reference real patches used to generate them. The SSL-guided LDM replicates the semantics of the reference patch. (Bottom) Large images $(1024 \times 1024)$ from our models, and the corresponding reference real images used to generate them. We preserve the global arrangement of the semantics defined in the reference image. + +![](images/5154abae3d87d938c1ae2bbe62ee395860662e64a888e34f6d1555fbc93acac6.jpg) + +![](images/4b4278c644829390cb668ca1b1eb2b59480392fab49ee3a600aed545a2c220ac.jpg) + +![](images/dfa2631a7bf199267834b76d3a52a4e627ecaa20416568fa3f427a8257833f23.jpg) + +![](images/924f86354ace31619d0d18db2deadf84eb67ccb2449dd71ff9984e7e014b9f86.jpg) + +For patch-level augmentation, we employ the NCT-CRC dataset [19], which has 100,000 Colorectal cancer (CRC) patches from 86 patients. Each $224 \times 224$ pixel patch at $20 \times$ magnification is annotated with one of nine distinct tissue class labels. + +For large-image augmentation, apart from TCGA-BRCA, we also use the BACH dataset [33]. Introduced in the ICIAR 2018 Grand Challenge, the dataset contains 400 H&E-stained Breast Cancer images of size $2048 \times 1536$ pixels, evenly distributed across four categories: normal, benign, in-situ carcinoma, and invasive carcinoma. + +# 4.2. Implementation details + +For all our experiments, we train the LDM on $256 \times 256$ pixel patches, following PathLDM [48], which fine-tunes an ImageNet-trained [38] U-Net denoiser and uses a $4 \times$ downsampling VQ-VAE, instead of the default LDM configuration. These modifications were deemed necessary for applying LDMs on large-image domains. + +We train our models on 6 NVIDIA RTX 8000 GPUs, with a batch size of 100 per GPU, utilizing code and pretrained checkpoints from LDM [37]. We set the learning rate at $10^{-4}$ with a warmup of 10,000 steps. We apply DDIM with 50 steps and a guidance scale of 1.75 for both patch sampling and large-image generation. When generating large images we apply the patch diffusion model with a stride $>1$ depending on the desired target quality. + +# 4.3. Image quality results + +In Fig. 3, we present synthetic patches and large images from our TCGA-BRCA and NAIP models, along with the + +corresponding references from which the self-supervised embeddings were extracted. We evaluate our method's per-patch and large-image generation quality by computing FID scores [14] using the Clean-FID implementation [31]. We generate 10,000 patches $(256 \times 256)$ and 3,000 large images $(1024 \times 1024)$ from diffusion models trained on the TCGA-BRCA, TCGA-CRC, and NAIP datasets. Since our generative model requires a self-supervised conditioning vector $y$ (or multiple vectors $z$ ) for each synthetic image (or large image), we randomly sample embeddings from reference images in the training set to generate images for evaluation. + +For patches, we measure FID against the real images ("Vanilla FID"). For large images, we follow the evaluation strategy of MultiDiffusion [3] and use FID to compare the distribution of $256 \times 256$ crops from synthesized large images to that of real image patches of the same size ("Crop FID"). We also measure FID directly between the large images and ground truth data using CLIP [34] ("CLIP FID"). + +To evaluate the similarity between the reference and generated large images, we resize from $1024 \times 1024$ to $256 \times 256$ and use an SSL model to extract embeddings. We compute the cosine similarity between the paired embeddings ("Embedding Similarity"). + +Patch-level quality: Our models achieve low patch FID scores across all datasets (Tab. 1). For TCGA-BRCA patches, our "Vanilla FID" of 6.98 is comparable to the current state-of-the-art [48] (7.64 at $10\times$ ). We attain similar, low FID scores for the smaller CRC and NAIP datasets. + +Synthetic large image quality: We present the "Crop FID", "CLIP FID" and "Embedding Similarity" results for large images in Table 1. The "Crop FID" of the large im + +
Dataset# Training imagesPatch levelLarge image level
Vanilla FIDCrop FIDCLIP FIDEmb similarity
BRCA 20x15M6.9815.517.430.924
CRC 20x8M6.788.87.340.938
NAIP667k11.543.766.86-
BRCA 5x976k9.74-6.64-
+ +Table 1. FID scores for our generated patch and large images. Our patch-level BRCA model is on par with SoTA [48] (7.64 at $10\times$ ). "CLIP FID" and "Embedding Similarity" demonstrate our large images' realism and contextual accuracy. + +ages is comparable to the "Vanilla FID" on the BRCA and CRC datasets, showing that patches from synthesized large images have similar semantic content to the ground truth patches. We attribute the worse "Crop FID" for the NAIP model to the limited number of samples available for both SSL and diffusion model training. The "Crop FID" is consistently higher than the patch-level FID, which is expected as the large-image generation framework only approximates the distribution of the large images and does not have access to the true conditioning at every location. + +Our low "CLIP FID" scores indicate that the generated large images are similar to real images when resized to $224 \times 224$ pixels. This indicates that our SSL-guided large-image generation successfully retains the larger-scale semantic arrangements of real data. For NAIP, our model is not as good at synthesizing high-frequency details, which explains the large discrepancy between "Crop FID" and "CLIP FID". Additionally, when comparing the "CLIP FID" of large images synthesized by the $20 \times$ BRCA model and resized to $5 \times$ , to images from a model trained directly on $5 \times$ data, we see minimal difference (7.43 vs. 6.64). + +We evaluate the contextual similarity between synthetic and reference large images for the BRCA and CRC data. We compute cosine similarity between large images using CTransPath embeddings. Our BRCA and CRC models demonstrate high "Embedding Similarity" scores of 0.924 and 0.938, respectively, reflecting our framework's effectiveness in preserving the integrity of context and key features on the large-image scale. + +# 5. Image Augmentation Experiments + +As shown in Fig. 3, apart from visual fidelity, the synthetic images preserve the intricate characteristics of the reference images, both on the patch and large-image scale. These characteristics include the nuanced textural elements, histological staining, and cell structure. This correspondence between real and synthetic images demonstrates the detailed and varied information captured by the self-supervised embeddings used as conditioning. In conjunction with the high "Embedding Similarity", it justifies using variations of im + +ages generated from SSL embeddings for patch and large-image level data augmentation. + +Having a powerful generative model enables us to perform data augmentation for tasks where we can control the augmented image label using conditioning. In our setting, our diffusion models are not trained with class labels; instead, we synthesize a novel image using the conditioning from a reference patch or region. + +We assume that i) the self-supervised embedding used as conditioning contains information about the target label and ii) the diffusion-generated variations of an image do not alter this target label information. We validate both assumptions experimentally, by augmenting training sets on patch and large-image level tissue classification tasks, including a Multiple Instance Learning (MIL) task. + +Large-image augmentation on TCGA-BRCA: We examine two histopathology slide-level binary classification tasks on TCGA-BRCA: BRCA Subtyping (Invasive Ductal Carcinoma (IDC) vs Invasive Lobular Carcinoma (ILC)) and HRD prediction. We utilize a minimal dataset, just $10\%$ of real WSI data (100 WSIs), to train MIL algorithms. We generate an equal set of synthetic images for 100 additional WSIs using training set images as reference. + +We employ 10-fold cross-validation to divide our dataset into training and testing segments. Within each fold, two multiple instance learning (MIL) models, CLAM-SB [25] and DSMIL [21], are trained on two sets: one with real data and another combining real and synthetic data. To train the MIL models, we extract features using the CTransPath ViT [44]. The results, detailed in Table 2, indicate that models trained on the augmented datasets consistently surpass their real-only counterparts, regardless of the MIL algorithm used. This demonstrates the value of the synthetic images generated by our method, confirming their efficacy as comparable to real images for training purposes. + +Large-image augmentation on BACH: We double the training set of BACH [33] by adding as many synthetic large-images, produced by the TCGA-BRCA diffusion model. From each $2048 \times 1536$ pixel training set image, we extract a $8 \times 6$ SSL embedding grid to generate a variation with the same label. For classification, we employ a ConvNeXt V2huge [46] model pre-trained on ImageNet. We train a 2-layer MLP classifier on top of the penultimate layer features and evaluate it on the official test set. + +The results, presented in Table 3a, reveal a notable improvement in classifier performance, from $78\%$ to $83\%$ . This improvement again confirms the high quality of our synthetic data while also highlighting the versatility of our approach. Despite being trained on $256 \times 256$ patches from TCGA-BRCA, the model generalizes to produce realistic large images from a completely different dataset. We attribute this generalization capability to the expressiveness of the SSL features and the potency of the diffusion model + +
TCGA-BRCA SubtypingTCGA-BRCA HRD
Method1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real
CLAM-SB0.7250.8120.8860.8980.910.6030.6440.6490.7650.787
DSMIL0.6090.6590.8380.8560.9050.5170.5540.5630.6390.669
+ +to accurately portray them in images. While our model does not reach the current SoTA accuracy of $87\%$ , achieved by an ensemble approach [7], the simplicity and ability to integrate with other models make for a noteworthy contribution. + +Table 2. The inclusion of synthetic data consistently enhances AUC across various MIL architectures and BRCA tasks. The dataset contains 1000 real images, so “ $10\%+$ synthetic” indicates training with 100 real and 100 synthetic WSIs, with the remainder used for testing. + +
Training DataTest Acc
Real78 %
Synthetic70 %
Real + synthetic83 %
SoTA [7]87 %
+ +(a) + +
Training DataVal Acc
Real93.8 %
Synthetic90.19 %
Real + synthetic96.27 %
SoTA [20]96.26 %
+ +(b) + +Table 3. Our data augmentations provide notable improvements for the BACH (a) and CRC-VAL-HE (b) datasets. Notably, the diffusion training data does not overlap with the data of the augmented datasets. + +Patch-level image augmentation: We further investigate our out-of-distribution generalization capabilities by augmenting the NCT-CRC dataset [19]. We leverage our diffusion model trained on the TCGA-CRC data, which does not overlap with NCT-CRC, conditioned on embeddings from an iBOT ViT [11]. We generate an augmented dataset of equal size to the original by using the SSL embedding of every patch in the NCT-CRC training set to synthesize a corresponding image of the same label. + +We train an ImageNet pre-trained ResNet-50 [12] network on three splits: real images only, synthetic images only, and a combination of both. Evaluation on the CRC-VAL-HE-7K test set demonstrates a significant performance increase when synthetic data is introduced, with classifier accuracy rising from $93.8\%$ to $96.27\%$ as presented in Table 3b. We match the current SoTA [20], which used an ensemble of deep models, with a model agnostic approach. As in the previous experiment, the diffusion model, now trained on the significantly smaller TCGA-CRC data, can also effectively synthesize images from a completely different dataset by only controlling the self-supervised conditioning. + +# 6. Text-to-large image synthesis + +For previous tasks, our large image generation approach synthesizes variations of an existing set of images from the pre-computed self-supervised embeddings. In Sec. 3.3 we + +discussed how to control large image generation with auxiliary signals from any domain (class labels, text captions, etc.), by training models $p(z \mid c)$ that can be combined with the embedding-conditioned image generation. We demonstrate controllable image synthesis with text-to-large image generation experiments on histopathology and satellite data. We measure the similarity between synthetic images and the text prompts used in generating them using vision-language models (VLMs). + +Text-to-large histopathology images: We utilize the CRC and BRCA diffusion models to generate $1024 \times 1024$ pixel images from text prompts. We first construct training sets by pairing $4 \times 4$ SSL embedding grids $z$ from large BRCA and CRC images, with their corresponding Quilt [18] image embeddings $c$ . We then train an auxiliary diffusion model to sample $p(z \mid c)$ . During inference, we use a text embedding $c'$ as a proxy for the image embedding, to sample $z$ and synthesize a large image. To bridge the gap between the image and text Quilt embeddings [23, 30] we perturb the image embeddings when training the auxiliary diffusion model with Gaussian noise of variance $\sigma^2 = 0.1$ . + +To evaluate the text-to-large image pipeline, we generate images from a pre-defined set of classes described in natural language; non-malignant benign tissue, malignant in-situ carcinoma, malignant invasive carcinoma, normal breast tissue for BRCA and colon adenocarcinoma, benign colonic tissue for CRC. We construct the confusion matrix of zero-shot classifiers on the synthesized data. We used two different VLMs as zero-shot classifiers, Quilt and BiomedCLIP [49]. The results presented in Fig. 4 demonstrate our ability to synthesize images consistent with the text prompts. The capabilities of the VLM used, limit our synthetic image generation. The lower performance of BRCA vs. CRC is consistent with the results reported in Quilt [18]. Furthermore, we asked an expert pathologist to classify 100 synthetic CRC images as benign or adenocarcinoma images. Their evaluation showed an $89.9\%$ agreement rate with the labels used for image generation, indicating consistency in our text-to-large image pipeline. We show examples of synthesized images in the supplementary. + +Text-to-large satellite images: To synthesize novel satellite images we first create a training set of $30\mathrm{k}1024\times 1024$ pixel large NAIP images, and pair them with captions + +from a BLIP model [22]. We train a diffusion model to sample the $4 \times 4$ SSL embeddings $z$ from the captions $c$ . To evaluate, we create a separate set of 1000 NAIP image-caption pairs and measure the CLIP similarity between generated images and the given captions. We achieve a CLIP similarity score of 0.22, showing that we can effectively learn the mapping from text to large images with this hierarchical approach. Although our CLIP similarity is slightly worse than the scores reported for text-to-image Stable Diffusion models ( $>0.24$ ) [32], we expect this drop in performance as we trained with machine-generated captions. Training and generated image-caption pairs are provided in supplementary. + +![](images/a5b98b4f0e2b4e5466e0373f7f4eb758b9b9e7588b8e30d20f1ca7935afddbb7.jpg) +Figure 4. Confusion matrix of zero-shot classification for novel TCGA-CRC and TCGA-BRCA synthetic images. + +# 7. Combining self-supervised embeddings with diffusion + +We further evaluate our patch-generating models by posing the question does the diffusion model learn more about the data than the self-supervised learning model? We hypothesize that by combining the pre-trained self-supervised embedding with the denoising task we can improve the learned representations of the data, leading to better performance in downstream tasks, which are performed with features from self-supervised learning. To validate this hypothesis, we utilize the trained diffusion model as a feature extractor and apply a Multiple-instance learning (MIL) approach for the slide-level classification task of subtyping Breast Cancer. For each patch in a whole-slide image (WSI) we first extract the self-supervised embeddings, which are then used as conditioning to obtain features from the denoiser's U-Net bottleneck layer at a fixed timestep $t = 50$ . + +In Tab. 4, we evaluate the effectiveness of our novel fusion of generative diffusion and self-supervised embeddings. We compare the performance of MIL algorithms [21, + +[25] using features derived from our approach at $20 \times$ and $5 \times$ magnification (LDMHIPT/CTransPath) against using only the self-supervised conditioning (HIPT/ CTransPath). We used a 10-fold cross-validation strategy consistent with the data splits from HIPT [6], training the MIL algorithms on both the full dataset (100%) and a reduced subset (25%). + +The results indicate that integrating self-supervised features into the diffusion model as conditioning leads to learning better representations and improves whole-slide classification. By fusing the generative knowledge from the diffusion process with the discriminative capabilities of the self-supervised embeddings, we construct a successful model for both discriminative and generative tasks. + +
MagFeatures25% training100% training
CLAM-SBDSMILCLAM-SBDSMIL
20×HIPT0.7880.7840.8610.839
LDMHIPT0.8420.7950.9080.894
5×CTransPath0.9000.8960.9190.910
LDMCTransPath0.9130.9050.9230.936
+ +Table 4. 10-fold cross-validation AUC for BRCA Histological subtyping. $\mathrm{LDM}_{\mathrm{HIPT}}$ denotes diffusion features conditioned on HIPT embeddings. The fusion of SSL and diffusion features outperforms the SSL features by themselves. + +# 8. Conclusion + +We presented a novel approach to training diffusion models in large-image domains, such as digital histopathology and remote sensing. We overcome the need for fine-grained annotation by introducing self-supervised representation guided diffusion models, achieving remarkable image synthesis results on the patch level. Our approach also enables us to synthesize high-quality large images, where we have the ability to dictate the global structure by controlling the spatial arrangement of the conditions. We evaluated the usefulness of our synthetic images on a number of patch and large image-level tasks, as well as introduced a text-to-large image generation framework. Naively augmenting whole-slide images is a time-consuming process. We leave to future work the exploration of adaptive augmentation strategies that choose which image parts to augment. We believe these results illustrate the great potential for this technology to lead to bespoke foundational models for specialized domains, comparable to existing models for natural images. + +Acknowledgements This research was partially supported by NCI awards 5U24CA215109, 1R21CA258493-01A1, UH3CA225021, NSF grants IIS-2123920, IIS-2212046 and Stony Brook Profund 2022 seed funding. We thank Rajarsi Gupta for his valuable feedback. + +# References + +[1] Marco Aversa, Gabriel Nobis, Miriam Hagele, Kai Standvoss, Mihaela Chirica, Roderick Murray-Smith, Ahmed Alaa, Lukas Ruff, Daniela Ivanova, Wojciech Samek, et al. Diffinfinite: Large mask-image synthesis via parallel random patch diffusion in histopathology. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1737-1752. PMLR, 2023. 4, 5 +[4] JN Cancer Genome Atlas Research Network et al. The cancer genome atlas pan-cancer analysis project. Nat. Genet, 45 (10):1113-1120, 2013. 2, 4 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 3, 4 +[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 1, 3, 4, 8 +[7] Sai Saketh Chennamsetty, Mohammed Safwan, and Varghese Alex. Classification of breast cancer histology image using ensemble of pre-trained neural networks. In Image Analysis and Recognition: 15th International Conference, ICIAR 2018, Póvoa de Varzim, Portugal, June 27-29, 2018, Proceedings 15, pages 804-811. Springer, 2018. 7 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[10] Miguel Espinosa and Elliot J Crowley. Generate your own scotland: Satellite image generation conditioned on maps. arXiv preprint arXiv:2308.16648, 2023. 3 +[11] Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathology with masked image modeling. medRxiv, pages 2023-07, 2023. 1, 3, 4, 7 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7 + +[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[17] Vincent Tao Hu, David W Zhang, Yuki M Asano, Gertjan J Burghouts, and Cees GM Snoek. Self-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18413-18422, 2023. 2 +[18] Wisdom Oluchi Ikezogwo, Mehmet Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Stefan Chan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. *Quilt-1m: One million image-text pairs for histopathology.* arXiv preprint arXiv:2306.11207, 2023. 4, 7 +[19] Jakob Nikolas Kather, Niels Halama, and Alexander Marx. 100,000 histological images of human colorectal cancer and healthy tissue, 2018. 5, 7 +[20] Anurodh Kumar, Amit Vishwakarma, and Varun Bajaj. Crccn-net: Automated framework for classification of colorectal tissue using histopathological images. Biomedical Signal Processing and Control, 79:104172, 2023. 7 +[21] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 6, 8 +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4, 8 +[23] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 7 +[24] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 1 +[25] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 6, 8 + +[26] Puria Azadi Moghadam, Sanne Van Dalen, Karina C Martin, Jochen Lennerz, Stephen Yip, Hossein Farahani, and Ali Bashashati. A morphology focused diffusion probabilistic model for synthesis of histopathology images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2000-2009, 2023. 2 +[27] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, et al. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13 (1):12098, 2023. 2 +[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2 +[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1 +[30] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 7 +[31] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 8 +[33] António Polónia, Catarina Eloy, and Paulo Aguiar. BACH Dataset : Grand Challenge on Breast Cancer Histology images, 2020. 5, 6 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5 +[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[36] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 4 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5 + +[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5 +[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2 +[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1 +[41] Ahmad Sebaq and Mohamed ElHelw. Rsdiff: Remote sensing image generation from text using diffusion model. arXiv preprint arXiv:2309.02455, 2023. 3 +[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2 +[43] USGS. National agriculture imagery program (NAIP), 2023. https://www.usgs.gov/centers/eros/science/usgs-eros-archive-aerial-photography - national - agriculture - imagery-program-naip.2,4 +[44] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Junzhou Huang, Wei Yang, and Xiao Han. Transpath: Transformer-based self-supervised learning for histopathological image classification. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VIII 24, pages 186-195. Springer, 2021. 4, 6 +[45] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis, 81:102559, 2022. 1, 3 +[46] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 6 +[47] Xuan Xu, Saarthak Kapse, Rajarsi Gupta, and Prateek Prasanna. Vit-dae: Transformer-driven diffusion autoencoder for histopathology image analysis. arXiv preprint arXiv:2304.01053, 2023. 2 +[48] Srikar Yellapragada, Alexandros Graikos, Prateek Prasanna, Tahsin Kurc, Joel Saltz, and Dimitris Samaras. PathLDM: Text conditioned latent diffusion model for histopathology. arXiv preprint arXiv:2309.00748, 2023. 2, 4, 5, 6 +[49] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen + +Valluri, Cliff Wong, et al. Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915, 2023. 7 \ No newline at end of file diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/images.zip b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a46e64b8a016064b238f38017f63b0e0886dc14 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:275a512fde6941b43cab63a14956b4e30753209fdc2c9c8ea1351f910d150d5a +size 484826 diff --git a/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/layout.json b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f45f9308aece6690178336a72d2745a3b7e5f77 --- /dev/null +++ b/2024/Learned Representation-Guided Diffusion Models for Large-Image Generation/layout.json @@ -0,0 +1,8689 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 529, + 121 + ], + "type": "text", + "content": "Learned representation-guided diffusion models for large-image generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 143, + 470, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 143, + 470, + 185 + ], + "spans": [ + { + "bbox": [ + 135, + 143, + 470, + 185 + ], + "type": "text", + "content": "Alexandros Graikos* Srikar Yellapragada* Minh-Quan Le Saarthak Kapse Prateek Prasanna Joel Saltz Dimitris Samaras Stony Brook University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 597 + ], + "type": "text", + "content": "To synthesize high-fidelity samples, diffusion models typically require auxiliary data to guide the generation process. However, it is impractical to procure the painstaking patch-level annotation effort required in specialized domains like histopathology and satellite imagery; it is often performed by domain experts and involves hundreds of millions of patches. Modern-day self-supervised learning (SSL) representations encode rich semantic and visual information. In this paper, we posit that such representations are expressive enough to act as proxies to fine-grained human labels. We introduce a novel approach that trains diffusion models conditioned on embeddings from SSL. Our diffusion models successfully project these features back to high-quality histopathology and remote sensing images. In addition, we construct larger images by assembling spatially consistent patches inferred from SSL embeddings, preserving long-range dependencies. Augmenting real data by generating variations of real images improves downstream classifier accuracy for patch-level and larger, image-scale classification tasks. Our models are effective even on datasets not encountered during training, demonstrating their robustness and generalizability. Generating images from learned embeddings is agnostic to the source of the embeddings. The SSL embeddings used to generate a large image can either be extracted from a reference image, or sampled from an auxiliary model conditioned on any related modality (e.g. class labels, text, genomic data). As proof of concept, we introduce the text-to-large image synthesis paradigm where we successfully synthesize large pathology and satellite images out of text descriptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 619, + 128, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 287, + 687 + ], + "type": "text", + "content": "Diffusion models produce high-quality and diverse samples across a spectrum of generative tasks [8, 24]. This leap forward has been enabled by the simultaneous curation of large-scale multi-modal datasets [40] and the development" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 211, + 547, + 326 + ], + "blocks": [ + { + "bbox": [ + 307, + 211, + 547, + 326 + ], + "lines": [ + { + "bbox": [ + 307, + 211, + 547, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 547, + 326 + ], + "type": "image", + "image_path": "26a83e1423cf49b6cdaa97b008ab8b1a43ea216117eed332430fd6ae51c85da4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 333, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 333, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 547, + 399 + ], + "type": "text", + "content": "Figure 1. We propose using SSL features to condition diffusion models. This allows us to construct large images by assembling consistent patches inferred from a spatial arrangement of SSL embeddings. The generated image retains the semantics of the embeddings used as a condition, maintaining the forested and open areas from the reference. Best viewed zoomed-in." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 423, + 546, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 460 + ], + "type": "text", + "content": "of efficient conditioning mechanisms [35, 37]. The key to unlocking the models' capabilities is to integrate auxiliary information during training and inference [8, 15, 28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 640 + ], + "type": "text", + "content": "Large-scale human-annotated datasets are mostly limited to image-caption pairs, collected from easily accessible online repositories and labeled by non-expert annotators. However, in domains such as digital histopathology and remote sensing, where gigapixel scale images provide vast amounts of unlabeled data, annotation proves challenging. Moreover, the process requires expert knowledge and is more difficult at a finer scale, i.e., captioning large crops of the gigapixel image is simpler than captioning smaller patches. Based on our estimates (see supplemental), annotating the entire TCGA-BRCA dataset with captions would take " + }, + { + "bbox": [ + 304, + 461, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\approx 40,000" + }, + { + "bbox": [ + 304, + 461, + 547, + 640 + ], + "type": "text", + "content": " hours of pathologist's time. Replicating the impressive results of diffusion models in these domains has been limited by the scarcity of fine-grained per-image conditioning, vital for high-quality image synthesis [29]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Modern-day self-supervised learning (SSL) representations [5] encode rich semantic and visual information. Features from trained self-supervised models serve as compact image representations and are widely used to perform discriminative downstream tasks successfully [6, 11, 45], proving that these compressed representations indeed encode" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 283, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 283, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 283, + 703 + ], + "type": "text", + "content": "*Equal contribution. Correspondence to agraikos@cs.stonybrook.edu" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 60, + 703, + 154, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 154, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 154, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 60, + 703, + 154, + 712 + ], + "type": "text", + "content": "Code is available at this link" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8532" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "useful semantic information about the images. We hypothesize that such SSL representations are already expressive enough to act as proxies to fine-grained human labels. If this is true, these representations should be able to condition the training of effective diffusion models in these domains. In this novel approach, we utilize self-supervised feature extractors as image annotators; these features provide necessary per-image conditioning signals at the highest resolutions, required for diffusion model training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 182, + 289, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 289 + ], + "type": "text", + "content": "Our experiments show that conditioning with expressive self-supervised features leads to precise control over the image content. The SSL features are adept at identifying complex patterns and structures in the images, while the diffusion model learns to translate them into visual components accurately (Fig. 3). This motivates us to perform large-image synthesis by locally controlling the appearance using SSL conditioning and dictating the global structure through the spatial arrangement of the conditions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 291, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 445 + ], + "type": "text", + "content": "Our approach synthesizes large images in a patch-based manner, using a single image diffusion model at the highest resolution. We represent a large image as a grid of SSL embeddings, where each serves as a representation of a large image neighborhood. The whole image is then synthesized by generating consistent patches that capture both the local properties, as given by the local patch conditioning, and the spatial arrangement of the conditioning features. If we change this spatial arrangement, we are, in fact, editing how semantic elements are arranged globally in the large image. This strategy enables the controllable generation of images of virtually any size without significantly increased computation compared to the base patch-level model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 447, + 289, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 289, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 289, + 519 + ], + "type": "text", + "content": "To generate a large image, our approach requires the patch diffusion model and the spatially-arranged conditioning. We can start with a reference large image as source and extract SSL embeddings from non-overlapping segments, enabling our method to synthesize a variation of the original image (Fig. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 521, + 289, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 664 + ], + "type": "text", + "content": "Utilizing SSL embeddings as conditions allows us to have the necessary control over image generation, at the expense of an explainable and easy-to-use conditioning mechanism. Nevertheless, we argue that since generating images from learned embeddings is agnostic to the embedding source, there are simple ways to combine control over generated images with explainability. We propose training auxiliary models to transform higher-level conditioning signals, such as text captions, to the learned patch representations. To demonstrate this versatility, we introduce text-to-large-image synthesis by training an auxiliary model to sample a spatial arrangement of embeddings from a text description." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "We train patch-level diffusion models using self-supervised features as conditioning on digital histopathology (TCGA [4]) and satellite image (NAIP [43]) datasets. We perform extensive evaluations and demonstrate the ad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "content": "vantages of SSL conditioning and our large-image generation framework on synthesis and classification tasks. Our model achieves exceptional patch-level and large-image quality, the ability to improve classifiers through data augmentation even when synthesizing out-of-distribution data, and effective fusion of diffusion and SSL features for downstream applications. Finally, we are the first to perform text-to-large image synthesis, which should be of significant community interest as vision-language models (VLMs) for pathology and satellite images gain traction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 192, + 502, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 192, + 502, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 502, + 203 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 204, + 545, + 360 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 306, + 204, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 545, + 251 + ], + "type": "text", + "content": "- We develop a novel method to train diffusion models with self-supervised learning features as conditioning and generate high-quality images in the histopathology and satellite image domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 287 + ], + "type": "text", + "content": "- We present a framework for large-image synthesis, based on self-supervised guided diffusion, that maintains contextual integrity and image realism over large areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 288, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 288, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 306, + 288, + 545, + 322 + ], + "type": "text", + "content": "- We demonstrate the applicability of our model in various classification tasks and showcase its unique ability to augment out-of-distribution datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 323, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 545, + 360 + ], + "type": "text", + "content": "- We introduce text-to-large image generation for digital histopathology and satellite images, highlighting the versatility of our approach." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 390, + 383 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 390, + 547, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 557 + ], + "type": "text", + "content": "Diffusion models: Introduced for image generation by Ho et al. [16], diffusion models have evolved considerably. These enhancements include class conditioning [28], architectural improvements and gradient-based guidance [8], and classifier-free guidance [15]. Latent Diffusion Models (LDMs) [37] proposed a two-step training process with a Variational Autoencoder (VAE) compressing input images into a lower-dimensional latent space and a diffusion model trained in this latent space. Denoising Diffusion Implicit Models (DDIM) [42] accelerate the sampling process by " + }, + { + "bbox": [ + 304, + 390, + 547, + 557 + ], + "type": "inline_equation", + "content": "10 - 50 \\times" + }, + { + "bbox": [ + 304, + 390, + 547, + 557 + ], + "type": "text", + "content": ". Self-guided diffusion models [17] also utilize self-supervised learning by quantizing SSL embeddings. In contrast to our approach, their quantization discards useful information from the SSL embeddings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 558, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 653 + ], + "type": "text", + "content": "Training generative models directly at the gigapixel resolution is infeasible. An alternative is to generate the images in a coarse-to-fine manner hierarchically. This has already been applied to natural images, where chaining multiple diffusion models generates images up to " + }, + { + "bbox": [ + 304, + 558, + 547, + 653 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 558, + 547, + 653 + ], + "type": "text", + "content": " resolution [32, 39]. However, it is still limited, as it substantially increases the parameter count and is inherently constrained by the final target resolution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "In the context of digital histopathology, works have been limited to training unconditional [26] or class-conditioned [27, 47] diffusion models. For text conditioning, pathology text reports were used to provide context on the whole-slide scale [48]. Similar approaches have been" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8533" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 523, + 296 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 523, + 296 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 523, + 296 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 523, + 296 + ], + "type": "image", + "image_path": "65afc4a7a334f2e01696fa705fe49eb3c0cc1ba8c0d1d3d78050bf41e73fd57b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": "Figure 2. (a) We train diffusion models on patches " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": " (e.g. the one in the green box) taken from a large image conditioned on SSL embeddings. (b) We present our large image generation framework in 4 steps: (i) We extract a set of spatially arranged embeddings from a reference image or sample them from an auxiliary model. (ii) For every location " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": ", we compute a conditioning vector " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": " by interpolating the spatial grid of embeddings. (iii) At every diffusion step, we denoise the patch " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "F(i,j)" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": " using the conditioning " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": ". (iv) The next step is computed by averaging the denoising updates of all patches that overlap at " + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 304, + 546, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 488 + ], + "type": "text", + "content": "applied to satellite data [10, 41]. Apart from high-quality images without manual annotation, our SSL conditioning is also necessary for large-image synthesis, as all previous conditioning methods would not be sensitive to the intricate differences between neighboring patches. In recent work, DiffInfinite [1] explored large-image generation using segmentation masks as conditioning. We argue this is still suboptimal as it requires accurate, human-annotated masks for training and a mask-generating model during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 489, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 657 + ], + "type": "text", + "content": "Self-Supervised Learning Self-supervised learning (SSL) refers to both discriminative [2, 13] and distillation [5] approaches that aim to learn representations of the data without supervision. In this work, we are mainly interested in self-supervised learning for histopathology. Notable recent developments include the Hierarchical Image Pyramid Transformer (HIPT) [6], which utilizes the inherent hierarchical structure of Whole Slide Images (WSIs), CTransPath [45], a hybrid CNN and multi-scale Swin Transformer model, and iBOT [11], a masked image modeling method. These models excel as patch-level feature extractors for WSIs by leveraging the unique structure of the large-image data and capture the important semantic information that we require for conditioning a generative model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "We propose training a diffusion model on patches " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " drawn from large histopathology and satellite images, using self" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "content": "supervised embeddings " + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "content": " as conditioning. Furthermore, we present a patch-based approach that utilizes the SSL-conditioned diffusion model to synthesize arbitrarily large images. An overview of our method is presented in Fig. 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 437, + 518, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 518, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 518, + 451 + ], + "type": "text", + "content": "3.1. Learned representation-guided diffusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": "Given a pre-trained self-supervised feature extractor, we employ LDMs [37] to learn the distribution over the large-image patches " + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "inline_equation", + "content": "p(\\mathcal{I})" + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": ". LDMs are comprised of three components: an image-compressing Variational Autoencoder (VAE) that transforms the input images to latent representations, a U-Net denoiser to learn a denoising diffusion process that transforms Gaussian noise to latents, and a conditioning mechanism that controls the diffusion process. The conditioning is performed in our setting with a single vector " + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": ", obtained from the self-supervised model, integrated via a cross-attention mechanism. We train the LDM using pairs of image patches " + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": " and the corresponding extracted self-supervised embeddings " + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 457, + 545, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 623, + 440, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 440, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 440, + 635 + ], + "type": "text", + "content": "3.2. Large image generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 714 + ], + "type": "text", + "content": "Our goal is to synthesize high-quality large images, that not only capture global structure but also maintain spatial consistency. As shown in Fig. 2, we replicate the semantics in each patch with SSL-guided LDM. At the same time, we preserve the global arrangement of these semantics as defined by the grid of patches in the reference image. Future" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8534" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "work can explore alternative approaches to ensure spatial alignment, including using topological constraints or integrating low-resolution information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 156 + ], + "type": "text", + "content": "We follow the MultiDiffusion [3] methodology to generate large images using only a patch-based diffusion model. We can represent the diffusion model as a learned mapping from images and conditions to images:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 164, + 287, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 164, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 132, + 164, + 287, + 176 + ], + "type": "interline_equation", + "content": "\\Phi : \\mathcal {I} \\times \\mathcal {Y} \\rightarrow \\mathcal {I}, \\tag {1}", + "image_path": "8d51b60db095659ca05542cfc858c1481209c870f740479e1ebe31feb8b2f216.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "content": " are the large-image patches and " + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{V} = \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "content": " are the per-patch conditions. To generate a patch, we initialize " + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "inline_equation", + "content": "I_{T}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "content": " and sequentially transform to a \"clean\" image " + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "inline_equation", + "content": "I_0" + }, + { + "bbox": [ + 47, + 184, + 287, + 233 + ], + "type": "text", + "content": " following the trained LDM:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 242, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 242, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 131, + 242, + 287, + 255 + ], + "type": "interline_equation", + "content": "I _ {t - 1} = \\Phi \\left(I _ {t} \\mid y\\right) \\tag {2}", + "image_path": "158a4cf6885bc387dda50c31433f862d81b385d329f538d059fb20e1592c4c46.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "inline_equation", + "content": "t = T, T - 1, \\ldots, 1" + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "text", + "content": ". We assume that a large image can be formed as a spatial grid of " + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "text", + "content": " patches of size " + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 262, + 287, + 299 + ], + "type": "text", + "content": ". We then define the large-image diffusion model as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 307, + 287, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 307, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 131, + 307, + 287, + 319 + ], + "type": "interline_equation", + "content": "\\Psi : \\mathcal {J} \\times \\mathcal {Z} \\rightarrow \\mathcal {J} \\tag {3}", + "image_path": "488d84949dfc084e2d303b2a4c09569b14cc121f4c5a2a96d22b7a1584c31759.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = \\mathbb{R}^{PH\\times PW\\times C}" + }, + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "text", + "content": " are the large images and " + }, + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{Z} = \\underbrace{\\mathcal{Y}\\times\\mathcal{Y}\\cdots\\times\\mathcal{Y}}_{P^2}" + }, + { + "bbox": [ + 47, + 327, + 287, + 363 + ], + "type": "text", + "content": " are the conditioning vectors of all patches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "content": "The process " + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "inline_equation", + "content": "\\Psi(J_t \\mid z) = J_{t-1}" + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "inline_equation", + "content": "J_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 46, + 364, + 287, + 400 + ], + "type": "text", + "content": ", can be approximated by first defining mappings between the two different image and conditioning spaces:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 407, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 407, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 102, + 407, + 287, + 422 + ], + "type": "interline_equation", + "content": "F _ {i, j} ^ {\\text {p r o j}}: \\mathcal {J} \\rightarrow I, \\quad \\lambda_ {i, j}: \\mathcal {Z} \\rightarrow \\mathcal {Y} \\tag {4}", + "image_path": "b5c33a4b19408767bcdbb60ff1368a2fe5ad7b6edd1ebe6f9d75ab8d42fff917.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": "where, for the case of large-image generation, we set " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "F_{i,j}^{\\mathrm{proj}}" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": " to be a crop (projection) of the large image, centered at " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": " the conditioning " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "y_{i,j} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": ". Since we only have conditioning vectors at the centers of the patches, we can use spatial interpolation algorithms to implement " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": ". This assumes that the interpolant " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "y_{i,j}" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": " is a valid conditioning vector for the diffusion process, which we validate experimentally. Then, " + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 431, + 287, + 528 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 536, + 293, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 293, + 571 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 293, + 571 + ], + "type": "interline_equation", + "content": "\\Psi \\left(J _ {t} \\mid z\\right) = \\underset {J \\in \\mathcal {J}} {\\arg \\min } \\sum_ {i, j} \\| F _ {i, j} ^ {\\operatorname {p r o j}} (J) - \\Phi \\left(F _ {i, j} ^ {\\operatorname {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right) \\| ^ {2} \\tag {5}", + "image_path": "082365dbf7a45750b885f7cf29f195827a9645ab5baf4c423d024abf2aa54e92.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "text", + "content": "which can be solved in closed-form by setting each pixel " + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 47, + 572, + 287, + 597 + ], + "type": "text", + "content": " to the average of all the patch-diffusion updates" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 604, + 301, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 604, + 301, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 604, + 301, + 653 + ], + "type": "interline_equation", + "content": "\\Psi \\left(J _ {t} \\mid z\\right) = \\sum_ {i, j} \\frac {F _ {i , j} ^ {\\text {u n p r o j}} (\\mathbf {1})}{\\sum_ {k , l} F _ {k , l} ^ {\\text {u n p r o j}} (\\mathbf {1})} \\otimes F _ {i, j} ^ {\\text {u n p r o j}} \\left(\\Phi \\left(F _ {i, j} ^ {\\text {p r o j}} \\left(J _ {t}\\right) \\mid \\lambda_ {i, j} (z)\\right)\\right) \\tag {6}", + "image_path": "c2c6070b4b831dd509cb249822ff8fbd32a3c0254d3e92d72587765cde4e276d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "content": " denotes a patch image where all values are set to 1 and " + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "F_{i,j}^{\\mathrm{unproj}}" + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "content": " is the inverse mapping of pixels from the crop centered at " + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "content": " back to the large image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We are able to generate images larger than the ones produced by the patch diffusion model. At the same time, we" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "control what each patch looks like, which is crucial in maintaining the semantic integrity of the larger image. The self-supervised conditions can capture the variations between neighboring patches necessary for producing realistic results. Generating large images with coarser conditioning, such as global text prompts [48], would lead to uniform texture regions (see supplementary)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 164, + 491, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 491, + 177 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 491, + 177 + ], + "type": "text", + "content": "3.3. Controllable large-image synthesis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "content": "Although the grid of self-supervised embeddings " + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "content": " cannot be manipulated in a human-interpretable manner, we argue that it is simple to assert more control over the generated images. As illustrated in Fig. 2, this control can be attained by training auxiliary models " + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "p(z \\mid c)" + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "content": " that translate higher-level conditioning signals, such as text captions " + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "content": ", to learned patch representations " + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 183, + 545, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "text", + "content": "Since there is no available dataset of paired large images and captions, we resort to pre-trained multi-modal models, such as Quilt [18], CLIP [34] and BLIP [22], to provide the text conditioning. We construct training sets by extracting self-supervised embeddings from training-set large images and pairing them with multi-modal image embeddings or generated captions. During inference, we first sample SSL embeddings from the learned distribution " + }, + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "inline_equation", + "content": "p(z \\mid c)" + }, + { + "bbox": [ + 304, + 267, + 545, + 375 + ], + "type": "text", + "content": ", then utilize our patch diffusion models to synthesize a large image." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 387, + 482, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 387, + 482, + 400 + ], + "spans": [ + { + "bbox": [ + 305, + 387, + 482, + 400 + ], + "type": "text", + "content": "4. Image Generation Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 407, + 368, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 407, + 368, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 407, + 368, + 418 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 426, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 509 + ], + "type": "text", + "content": "We train diffusion models on digital histopathology images from The Genome Cancer Atlas (TCGA) [4] and satellite imagery from the National Agriculture Imagery Program (NAIP) [43]. Specifically, we used the TCGA-BRCA (Breast Invasive Carcinoma Collection), TCGA-CRC (COAD + READ Colorectal Carcinoma) datasets, and the Chesapeake Land Cover dataset [36]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": "For the TCGA-BRCA and TCGA-CRC datasets, we use images at " + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": " magnification, and developed diffusion models conditioned on embeddings from HIPT [6] and iBOT [11], which were pre-trained on PanCancer TCGA. In the case of the HIPT model, we specifically used its patch-level ViT. Additionally, for the TCGA-BRCA dataset, we train a model at " + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 304, + 510, + 545, + 605 + ], + "type": "text", + "content": " magnification using embeddings from CTransPath [44] for additional evaluations." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The Chesapeake Land Cover dataset contains 732 NAIP tiles, each measuring a " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "6\\mathrm{km} \\times 7.5\\mathrm{km}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " area at " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " resolution. We extract " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " non-overlapping pixel patches, resulting in 667,000 patches total. Given the absence of publicly available self-supervised learning models tailored to the NAIP data, we train a Vision Transformer (ViT-B/16) [9] using the DINO framework [5]. We then use the learned DINO embeddings to train the diffusion model on pairs of image patches and self-supervised embeddings." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8535" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 253, + 148 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 253, + 148 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 253, + 148 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 253, + 148 + ], + "type": "image", + "image_path": "ba1d5d338f16fefd0c738ebe359bf2b54dc18718b0713ebc4ff6e088aa0e890b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 159, + 179, + 275 + ], + "blocks": [ + { + "bbox": [ + 72, + 159, + 179, + 275 + ], + "lines": [ + { + "bbox": [ + 72, + 159, + 179, + 275 + ], + "spans": [ + { + "bbox": [ + 72, + 159, + 179, + 275 + ], + "type": "image", + "image_path": "5dc80e3652ccf483806329ff59547965b5b0632cdbea960a4c020f9bd7b016b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "text", + "content": "Figure 3. (Top) Patches " + }, + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "text", + "content": " from our models, and the corresponding reference real patches used to generate them. The SSL-guided LDM replicates the semantics of the reference patch. (Bottom) Large images " + }, + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "inline_equation", + "content": "(1024 \\times 1024)" + }, + { + "bbox": [ + 46, + 284, + 546, + 319 + ], + "type": "text", + "content": " from our models, and the corresponding reference real images used to generate them. We preserve the global arrangement of the semantics defined in the reference image." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 187, + 159, + 293, + 275 + ], + "blocks": [ + { + "bbox": [ + 187, + 159, + 293, + 275 + ], + "lines": [ + { + "bbox": [ + 187, + 159, + 293, + 275 + ], + "spans": [ + { + "bbox": [ + 187, + 159, + 293, + 275 + ], + "type": "image", + "image_path": "5154abae3d87d938c1ae2bbe62ee395860662e64a888e34f6d1555fbc93acac6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 481, + 148 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 481, + 148 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 481, + 148 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 481, + 148 + ], + "type": "image", + "image_path": "4b4278c644829390cb668ca1b1eb2b59480392fab49ee3a600aed545a2c220ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 301, + 159, + 408, + 275 + ], + "blocks": [ + { + "bbox": [ + 301, + 159, + 408, + 275 + ], + "lines": [ + { + "bbox": [ + 301, + 159, + 408, + 275 + ], + "spans": [ + { + "bbox": [ + 301, + 159, + 408, + 275 + ], + "type": "image", + "image_path": "dfa2631a7bf199267834b76d3a52a4e627ecaa20416568fa3f427a8257833f23.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 414, + 159, + 521, + 275 + ], + "blocks": [ + { + "bbox": [ + 414, + 159, + 521, + 275 + ], + "lines": [ + { + "bbox": [ + 414, + 159, + 521, + 275 + ], + "spans": [ + { + "bbox": [ + 414, + 159, + 521, + 275 + ], + "type": "image", + "image_path": "924f86354ace31619d0d18db2deadf84eb67ccb2449dd71ff9984e7e014b9f86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "text", + "content": "For patch-level augmentation, we employ the NCT-CRC dataset [19], which has 100,000 Colorectal cancer (CRC) patches from 86 patients. Each " + }, + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "text", + "content": " pixel patch at " + }, + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 46, + 338, + 287, + 396 + ], + "type": "text", + "content": " magnification is annotated with one of nine distinct tissue class labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 398, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 398, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 398, + 287, + 469 + ], + "type": "text", + "content": "For large-image augmentation, apart from TCGA-BRCA, we also use the BACH dataset [33]. Introduced in the ICIAR 2018 Grand Challenge, the dataset contains 400 H&E-stained Breast Cancer images of size " + }, + { + "bbox": [ + 46, + 398, + 287, + 469 + ], + "type": "inline_equation", + "content": "2048 \\times 1536" + }, + { + "bbox": [ + 46, + 398, + 287, + 469 + ], + "type": "text", + "content": " pixels, evenly distributed across four categories: normal, benign, in-situ carcinoma, and invasive carcinoma." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 477, + 178, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 178, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 178, + 490 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": "For all our experiments, we train the LDM on " + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": " pixel patches, following PathLDM [48], which fine-tunes an ImageNet-trained [38] U-Net denoiser and uses a " + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 495, + 287, + 567 + ], + "type": "text", + "content": " downsampling VQ-VAE, instead of the default LDM configuration. These modifications were deemed necessary for applying LDMs on large-image domains." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "content": "We train our models on 6 NVIDIA RTX 8000 GPUs, with a batch size of 100 per GPU, utilizing code and pretrained checkpoints from LDM [37]. We set the learning rate at " + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "content": " with a warmup of 10,000 steps. We apply DDIM with 50 steps and a guidance scale of 1.75 for both patch sampling and large-image generation. When generating large images we apply the patch diffusion model with a stride " + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "inline_equation", + "content": ">1" + }, + { + "bbox": [ + 46, + 567, + 287, + 664 + ], + "type": "text", + "content": " depending on the desired target quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 671, + 168, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 168, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 168, + 684 + ], + "type": "text", + "content": "4.3. Image quality results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "In Fig. 3, we present synthetic patches and large images from our TCGA-BRCA and NAIP models, along with the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "corresponding references from which the self-supervised embeddings were extracted. We evaluate our method's per-patch and large-image generation quality by computing FID scores [14] using the Clean-FID implementation [31]. We generate 10,000 patches " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " and 3,000 large images " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "(1024 \\times 1024)" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " from diffusion models trained on the TCGA-BRCA, TCGA-CRC, and NAIP datasets. Since our generative model requires a self-supervised conditioning vector " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": " (or multiple vectors " + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": ") for each synthetic image (or large image), we randomly sample embeddings from reference images in the training set to generate images for evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "text", + "content": "For patches, we measure FID against the real images (\"Vanilla FID\"). For large images, we follow the evaluation strategy of MultiDiffusion [3] and use FID to compare the distribution of " + }, + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 471, + 545, + 555 + ], + "type": "text", + "content": " crops from synthesized large images to that of real image patches of the same size (\"Crop FID\"). We also measure FID directly between the large images and ground truth data using CLIP [34] (\"CLIP FID\")." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": "To evaluate the similarity between the reference and generated large images, we resize from " + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": " and use an SSL model to extract embeddings. We compute the cosine similarity between the paired embeddings (\"Embedding Similarity\")." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": "Patch-level quality: Our models achieve low patch FID scores across all datasets (Tab. 1). For TCGA-BRCA patches, our \"Vanilla FID\" of 6.98 is comparable to the current state-of-the-art [48] (7.64 at " + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": "). We attain similar, low FID scores for the smaller CRC and NAIP datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Synthetic large image quality: We present the \"Crop FID\", \"CLIP FID\" and \"Embedding Similarity\" results for large images in Table 1. The \"Crop FID\" of the large im" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8536" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 284, + 152 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 284, + 152 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 284, + 152 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 284, + 152 + ], + "type": "table", + "html": "
Dataset# Training imagesPatch levelLarge image level
Vanilla FIDCrop FIDCLIP FIDEmb similarity
BRCA 20x15M6.9815.517.430.924
CRC 20x8M6.788.87.340.938
NAIP667k11.543.766.86-
BRCA 5x976k9.74-6.64-
", + "image_path": "3775aee523e5c5cd2079966ec7022269248d014276949f037b76967ade51b5ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 161, + 287, + 206 + ], + "lines": [ + { + "bbox": [ + 47, + 161, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 47, + 161, + 287, + 206 + ], + "type": "text", + "content": "Table 1. FID scores for our generated patch and large images. Our patch-level BRCA model is on par with SoTA [48] (7.64 at " + }, + { + "bbox": [ + 47, + 161, + 287, + 206 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 47, + 161, + 287, + 206 + ], + "type": "text", + "content": "). \"CLIP FID\" and \"Embedding Similarity\" demonstrate our large images' realism and contextual accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 227, + 286, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 227, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 286, + 346 + ], + "type": "text", + "content": "ages is comparable to the \"Vanilla FID\" on the BRCA and CRC datasets, showing that patches from synthesized large images have similar semantic content to the ground truth patches. We attribute the worse \"Crop FID\" for the NAIP model to the limited number of samples available for both SSL and diffusion model training. The \"Crop FID\" is consistently higher than the patch-level FID, which is expected as the large-image generation framework only approximates the distribution of the large images and does not have access to the true conditioning at every location." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": "Our low \"CLIP FID\" scores indicate that the generated large images are similar to real images when resized to " + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": " pixels. This indicates that our SSL-guided large-image generation successfully retains the larger-scale semantic arrangements of real data. For NAIP, our model is not as good at synthesizing high-frequency details, which explains the large discrepancy between \"Crop FID\" and \"CLIP FID\". Additionally, when comparing the \"CLIP FID\" of large images synthesized by the " + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": " BRCA model and resized to " + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": ", to images from a model trained directly on " + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 46, + 347, + 286, + 477 + ], + "type": "text", + "content": " data, we see minimal difference (7.43 vs. 6.64)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 286, + 574 + ], + "type": "text", + "content": "We evaluate the contextual similarity between synthetic and reference large images for the BRCA and CRC data. We compute cosine similarity between large images using CTransPath embeddings. Our BRCA and CRC models demonstrate high \"Embedding Similarity\" scores of 0.924 and 0.938, respectively, reflecting our framework's effectiveness in preserving the integrity of context and key features on the large-image scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 237, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 237, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 237, + 600 + ], + "type": "text", + "content": "5. Image Augmentation Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 286, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, apart from visual fidelity, the synthetic images preserve the intricate characteristics of the reference images, both on the patch and large-image scale. These characteristics include the nuanced textural elements, histological staining, and cell structure. This correspondence between real and synthetic images demonstrates the detailed and varied information captured by the self-supervised embeddings used as conditioning. In conjunction with the high \"Embedding Similarity\", it justifies using variations of im" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 95 + ], + "type": "text", + "content": "ages generated from SSL embeddings for patch and large-image level data augmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 98, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 545, + 169 + ], + "type": "text", + "content": "Having a powerful generative model enables us to perform data augmentation for tasks where we can control the augmented image label using conditioning. In our setting, our diffusion models are not trained with class labels; instead, we synthesize a novel image using the conditioning from a reference patch or region." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 545, + 253 + ], + "type": "text", + "content": "We assume that i) the self-supervised embedding used as conditioning contains information about the target label and ii) the diffusion-generated variations of an image do not alter this target label information. We validate both assumptions experimentally, by augmenting training sets on patch and large-image level tissue classification tasks, including a Multiple Instance Learning (MIL) task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": "Large-image augmentation on TCGA-BRCA: We examine two histopathology slide-level binary classification tasks on TCGA-BRCA: BRCA Subtyping (Invasive Ductal Carcinoma (IDC) vs Invasive Lobular Carcinoma (ILC)) and HRD prediction. We utilize a minimal dataset, just " + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 255, + 545, + 350 + ], + "type": "text", + "content": " of real WSI data (100 WSIs), to train MIL algorithms. We generate an equal set of synthetic images for 100 additional WSIs using training set images as reference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 495 + ], + "type": "text", + "content": "We employ 10-fold cross-validation to divide our dataset into training and testing segments. Within each fold, two multiple instance learning (MIL) models, CLAM-SB [25] and DSMIL [21], are trained on two sets: one with real data and another combining real and synthetic data. To train the MIL models, we extract features using the CTransPath ViT [44]. The results, detailed in Table 2, indicate that models trained on the augmented datasets consistently surpass their real-only counterparts, regardless of the MIL algorithm used. This demonstrates the value of the synthetic images generated by our method, confirming their efficacy as comparable to real images for training purposes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "text", + "content": "Large-image augmentation on BACH: We double the training set of BACH [33] by adding as many synthetic large-images, produced by the TCGA-BRCA diffusion model. From each " + }, + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "inline_equation", + "content": "2048 \\times 1536" + }, + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "text", + "content": " pixel training set image, we extract a " + }, + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "inline_equation", + "content": "8 \\times 6" + }, + { + "bbox": [ + 304, + 497, + 545, + 604 + ], + "type": "text", + "content": " SSL embedding grid to generate a variation with the same label. For classification, we employ a ConvNeXt V2huge [46] model pre-trained on ImageNet. We train a 2-layer MLP classifier on top of the penultimate layer features and evaluate it on the official test set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The results, presented in Table 3a, reveal a notable improvement in classifier performance, from " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". This improvement again confirms the high quality of our synthetic data while also highlighting the versatility of our approach. Despite being trained on " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " patches from TCGA-BRCA, the model generalizes to produce realistic large images from a completely different dataset. We attribute this generalization capability to the expressiveness of the SSL features and the potency of the diffusion model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8537" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 547, + 135 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 135 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 135 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 135 + ], + "type": "table", + "html": "
TCGA-BRCA SubtypingTCGA-BRCA HRD
Method1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real1% Real1% Real + synthetic10% Real10% Real + synthetic20% Real
CLAM-SB0.7250.8120.8860.8980.910.6030.6440.6490.7650.787
DSMIL0.6090.6590.8380.8560.9050.5170.5540.5630.6390.669
", + "image_path": "52b65db27a4cd5190299070928a7ac0b56d07026e276650dfe9c4a057b6df7dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 187, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 187, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 187, + 287, + 236 + ], + "type": "text", + "content": "to accurately portray them in images. While our model does not reach the current SoTA accuracy of " + }, + { + "bbox": [ + 47, + 187, + 287, + 236 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 47, + 187, + 287, + 236 + ], + "type": "text", + "content": ", achieved by an ensemble approach [7], the simplicity and ability to integrate with other models make for a noteworthy contribution." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 247, + 164, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 546, + 167 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 546, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 546, + 167 + ], + "type": "text", + "content": "Table 2. The inclusion of synthetic data consistently enhances AUC across various MIL architectures and BRCA tasks. The dataset contains 1000 real images, so “" + }, + { + "bbox": [ + 47, + 144, + 546, + 167 + ], + "type": "inline_equation", + "content": "10\\%+" + }, + { + "bbox": [ + 47, + 144, + 546, + 167 + ], + "type": "text", + "content": " synthetic” indicates training with 100 real and 100 synthetic WSIs, with the remainder used for testing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 247, + 164, + 304 + ], + "lines": [ + { + "bbox": [ + 50, + 247, + 164, + 304 + ], + "spans": [ + { + "bbox": [ + 50, + 247, + 164, + 304 + ], + "type": "table", + "html": "
Training DataTest Acc
Real78 %
Synthetic70 %
Real + synthetic83 %
SoTA [7]87 %
", + "image_path": "a8aecc5424c9a5dd6201534e124557ae021263ebd9285cddf007ce70c906a4fb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 168, + 247, + 282, + 304 + ], + "blocks": [ + { + "bbox": [ + 100, + 307, + 111, + 316 + ], + "lines": [ + { + "bbox": [ + 100, + 307, + 111, + 316 + ], + "spans": [ + { + "bbox": [ + 100, + 307, + 111, + 316 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 168, + 247, + 282, + 304 + ], + "lines": [ + { + "bbox": [ + 168, + 247, + 282, + 304 + ], + "spans": [ + { + "bbox": [ + 168, + 247, + 282, + 304 + ], + "type": "table", + "html": "
Training DataVal Acc
Real93.8 %
Synthetic90.19 %
Real + synthetic96.27 %
SoTA [20]96.26 %
", + "image_path": "c2ddf479da98bbf4d59810380072eb4f2f913b67ab973e625f1a8cf6b8e4920d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 219, + 306, + 230, + 315 + ], + "lines": [ + { + "bbox": [ + 219, + 306, + 230, + 315 + ], + "spans": [ + { + "bbox": [ + 219, + 306, + 230, + 315 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "type": "text", + "content": "Table 3. Our data augmentations provide notable improvements for the BACH (a) and CRC-VAL-HE (b) datasets. Notably, the diffusion training data does not overlap with the data of the augmented datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 486 + ], + "type": "text", + "content": "Patch-level image augmentation: We further investigate our out-of-distribution generalization capabilities by augmenting the NCT-CRC dataset [19]. We leverage our diffusion model trained on the TCGA-CRC data, which does not overlap with NCT-CRC, conditioned on embeddings from an iBOT ViT [11]. We generate an augmented dataset of equal size to the original by using the SSL embedding of every patch in the NCT-CRC training set to synthesize a corresponding image of the same label." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "text", + "content": "We train an ImageNet pre-trained ResNet-50 [12] network on three splits: real images only, synthetic images only, and a combination of both. Evaluation on the CRC-VAL-HE-7K test set demonstrates a significant performance increase when synthetic data is introduced, with classifier accuracy rising from " + }, + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "inline_equation", + "content": "93.8\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "inline_equation", + "content": "96.27\\%" + }, + { + "bbox": [ + 46, + 487, + 288, + 643 + ], + "type": "text", + "content": " as presented in Table 3b. We match the current SoTA [20], which used an ensemble of deep models, with a model agnostic approach. As in the previous experiment, the diffusion model, now trained on the significantly smaller TCGA-CRC data, can also effectively synthesize images from a completely different dataset by only controlling the self-supervised conditioning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 656, + 211, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 211, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 211, + 670 + ], + "type": "text", + "content": "6. Text-to-large image synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "For previous tasks, our large image generation approach synthesizes variations of an existing set of images from the pre-computed self-supervised embeddings. In Sec. 3.3 we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "content": "discussed how to control large image generation with auxiliary signals from any domain (class labels, text captions, etc.), by training models " + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "inline_equation", + "content": "p(z \\mid c)" + }, + { + "bbox": [ + 304, + 187, + 545, + 295 + ], + "type": "text", + "content": " that can be combined with the embedding-conditioned image generation. We demonstrate controllable image synthesis with text-to-large image generation experiments on histopathology and satellite data. We measure the similarity between synthetic images and the text prompts used in generating them using vision-language models (VLMs)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": "Text-to-large histopathology images: We utilize the CRC and BRCA diffusion models to generate " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": " pixel images from text prompts. We first construct training sets by pairing " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": " SSL embedding grids " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": " from large BRCA and CRC images, with their corresponding Quilt [18] image embeddings " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": ". We then train an auxiliary diffusion model to sample " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "p(z \\mid c)" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": ". During inference, we use a text embedding " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": " as a proxy for the image embedding, to sample " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": " and synthesize a large image. To bridge the gap between the image and text Quilt embeddings [23, 30] we perturb the image embeddings when training the auxiliary diffusion model with Gaussian noise of variance " + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "inline_equation", + "content": "\\sigma^2 = 0.1" + }, + { + "bbox": [ + 304, + 299, + 546, + 442 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 446, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 547, + 674 + ], + "type": "text", + "content": "To evaluate the text-to-large image pipeline, we generate images from a pre-defined set of classes described in natural language; non-malignant benign tissue, malignant in-situ carcinoma, malignant invasive carcinoma, normal breast tissue for BRCA and colon adenocarcinoma, benign colonic tissue for CRC. We construct the confusion matrix of zero-shot classifiers on the synthesized data. We used two different VLMs as zero-shot classifiers, Quilt and BiomedCLIP [49]. The results presented in Fig. 4 demonstrate our ability to synthesize images consistent with the text prompts. The capabilities of the VLM used, limit our synthetic image generation. The lower performance of BRCA vs. CRC is consistent with the results reported in Quilt [18]. Furthermore, we asked an expert pathologist to classify 100 synthetic CRC images as benign or adenocarcinoma images. Their evaluation showed an " + }, + { + "bbox": [ + 304, + 446, + 547, + 674 + ], + "type": "inline_equation", + "content": "89.9\\%" + }, + { + "bbox": [ + 304, + 446, + 547, + 674 + ], + "type": "text", + "content": " agreement rate with the labels used for image generation, indicating consistency in our text-to-large image pipeline. We show examples of synthesized images in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 677, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 715 + ], + "type": "text", + "content": "Text-to-large satellite images: To synthesize novel satellite images we first create a training set of " + }, + { + "bbox": [ + 306, + 677, + 545, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}1024\\times 1024" + }, + { + "bbox": [ + 306, + 677, + 545, + 715 + ], + "type": "text", + "content": " pixel large NAIP images, and pair them with captions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8538" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": "from a BLIP model [22]. We train a diffusion model to sample the " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " SSL embeddings " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " from the captions " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ". To evaluate, we create a separate set of 1000 NAIP image-caption pairs and measure the CLIP similarity between generated images and the given captions. We achieve a CLIP similarity score of 0.22, showing that we can effectively learn the mapping from text to large images with this hierarchical approach. Although our CLIP similarity is slightly worse than the scores reported for text-to-image Stable Diffusion models (" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": ">0.24" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ") [32], we expect this drop in performance as we trained with machine-generated captions. Training and generated image-caption pairs are provided in supplementary." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 224, + 289, + 411 + ], + "blocks": [ + { + "bbox": [ + 47, + 224, + 289, + 411 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 289, + 411 + ], + "type": "image", + "image_path": "a5b98b4f0e2b4e5466e0373f7f4eb758b9b9e7588b8e30d20f1ca7935afddbb7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 288, + 440 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 288, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 288, + 440 + ], + "type": "text", + "content": "Figure 4. Confusion matrix of zero-shot classification for novel TCGA-CRC and TCGA-BRCA synthetic images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 464, + 287, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 464, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 464, + 287, + 491 + ], + "type": "text", + "content": "7. Combining self-supervised embeddings with diffusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "text", + "content": "We further evaluate our patch-generating models by posing the question does the diffusion model learn more about the data than the self-supervised learning model? We hypothesize that by combining the pre-trained self-supervised embedding with the denoising task we can improve the learned representations of the data, leading to better performance in downstream tasks, which are performed with features from self-supervised learning. To validate this hypothesis, we utilize the trained diffusion model as a feature extractor and apply a Multiple-instance learning (MIL) approach for the slide-level classification task of subtyping Breast Cancer. For each patch in a whole-slide image (WSI) we first extract the self-supervised embeddings, which are then used as conditioning to obtain features from the denoiser's U-Net bottleneck layer at a fixed timestep " + }, + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "inline_equation", + "content": "t = 50" + }, + { + "bbox": [ + 46, + 498, + 287, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "In Tab. 4, we evaluate the effectiveness of our novel fusion of generative diffusion and self-supervised embeddings. We compare the performance of MIL algorithms [21," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "[25] using features derived from our approach at " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " magnification (LDMHIPT/CTransPath) against using only the self-supervised conditioning (HIPT/ CTransPath). We used a 10-fold cross-validation strategy consistent with the data splits from HIPT [6], training the MIL algorithms on both the full dataset (100%) and a reduced subset (25%)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 145, + 546, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 546, + 229 + ], + "type": "text", + "content": "The results indicate that integrating self-supervised features into the diffusion model as conditioning leads to learning better representations and improves whole-slide classification. By fusing the generative knowledge from the diffusion process with the discriminative capabilities of the self-supervised embeddings, we construct a successful model for both discriminative and generative tasks." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 242, + 544, + 310 + ], + "blocks": [ + { + "bbox": [ + 309, + 242, + 544, + 310 + ], + "lines": [ + { + "bbox": [ + 309, + 242, + 544, + 310 + ], + "spans": [ + { + "bbox": [ + 309, + 242, + 544, + 310 + ], + "type": "table", + "html": "
MagFeatures25% training100% training
CLAM-SBDSMILCLAM-SBDSMIL
20×HIPT0.7880.7840.8610.839
LDMHIPT0.8420.7950.9080.894
5×CTransPath0.9000.8960.9190.910
LDMCTransPath0.9130.9050.9230.936
", + "image_path": "e7ec8e574ec5b36b08b4b5870882f50319d1044862e5c8920fb6b71c547c2d6a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 319, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 363 + ], + "type": "text", + "content": "Table 4. 10-fold cross-validation AUC for BRCA Histological subtyping. " + }, + { + "bbox": [ + 304, + 319, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\mathrm{LDM}_{\\mathrm{HIPT}}" + }, + { + "bbox": [ + 304, + 319, + 545, + 363 + ], + "type": "text", + "content": " denotes diffusion features conditioned on HIPT embeddings. The fusion of SSL and diffusion features outperforms the SSL features by themselves." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 378, + 406 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 415, + 545, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 631 + ], + "type": "text", + "content": "We presented a novel approach to training diffusion models in large-image domains, such as digital histopathology and remote sensing. We overcome the need for fine-grained annotation by introducing self-supervised representation guided diffusion models, achieving remarkable image synthesis results on the patch level. Our approach also enables us to synthesize high-quality large images, where we have the ability to dictate the global structure by controlling the spatial arrangement of the conditions. We evaluated the usefulness of our synthetic images on a number of patch and large image-level tasks, as well as introduced a text-to-large image generation framework. Naively augmenting whole-slide images is a time-consuming process. We leave to future work the exploration of adaptive augmentation strategies that choose which image parts to augment. We believe these results illustrate the great potential for this technology to lead to bespoke foundational models for specialized domains, comparable to existing models for natural images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements This research was partially supported by NCI awards 5U24CA215109, 1R21CA258493-01A1, UH3CA225021, NSF grants IIS-2123920, IIS-2212046 and Stony Brook Profund 2022 seed funding. We thank Rajarsi Gupta for his valuable feedback." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8539" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 167 + ], + "type": "text", + "content": "[1] Marco Aversa, Gabriel Nobis, Miriam Hagele, Kai Standvoss, Mihaela Chirica, Roderick Murray-Smith, Ahmed Alaa, Lukas Ruff, Daniela Ivanova, Wojciech Samek, et al. Diffinfinite: Large mask-image synthesis via parallel random patch diffusion in histopathology. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 201 + ], + "type": "text", + "content": "[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 287, + 256 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 1737-1752. PMLR, 2023. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 290 + ], + "type": "text", + "content": "[4] JN Cancer Genome Atlas Research Network et al. The cancer genome atlas pan-cancer analysis project. Nat. Genet, 45 (10):1113-1120, 2013. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 346 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 288, + 412 + ], + "type": "text", + "content": "[6] Richard J Chen, Chengkuan Chen, Yicong Li, Tiffany Y Chen, Andrew D Trister, Rahul G Krishnan, and Faisal Mahmood. Scaling vision transformers to gigapixel images via hierarchical self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16144-16155, 2022. 1, 3, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 479 + ], + "type": "text", + "content": "[7] Sai Saketh Chennamsetty, Mohammed Safwan, and Varghese Alex. Classification of breast cancer histology image using ensemble of pre-trained neural networks. In Image Analysis and Recognition: 15th International Conference, ICIAR 2018, Póvoa de Varzim, Portugal, June 27-29, 2018, Proceedings 15, pages 804-811. Springer, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 513 + ], + "type": "text", + "content": "[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 287, + 578 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 613 + ], + "type": "text", + "content": "[10] Miguel Espinosa and Elliot J Crowley. Generate your own scotland: Satellite image generation conditioned on maps. arXiv preprint arXiv:2308.16648, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[11] Alexandre Filiot, Ridouane Ghermi, Antoine Olivier, Paul Jacob, Lucas Fidon, Alice Mac Kain, Charlie Saillard, and Jean-Baptiste Schiratti. Scaling self-supervised learning for histopathology with masked image modeling. medRxiv, pages 2023-07, 2023. 1, 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 207 + ], + "type": "text", + "content": "[15] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 241 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 296 + ], + "type": "text", + "content": "[17] Vincent Tao Hu, David W Zhang, Yuki M Asano, Gertjan J Burghouts, and Cees GM Snoek. Self-guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18413-18422, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 363 + ], + "type": "text", + "content": "[18] Wisdom Oluchi Ikezogwo, Mehmet Saygin Seyfioglu, Fatemeh Ghezloo, Dylan Stefan Chan Geva, Fatwir Sheikh Mohammed, Pavan Kumar Anand, Ranjay Krishna, and Linda Shapiro. *Quilt-1m: One million image-text pairs for histopathology.* arXiv preprint arXiv:2306.11207, 2023. 4, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[19] Jakob Nikolas Kather, Niels Halama, and Alexander Marx. 100,000 histological images of human colorectal cancer and healthy tissue, 2018. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 443 + ], + "type": "text", + "content": "[20] Anurodh Kumar, Amit Vishwakarma, and Varun Bajaj. Crccn-net: Automated framework for classification of colorectal tissue using histopathological images. Biomedical Signal Processing and Control, 79:104172, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 499 + ], + "type": "text", + "content": "[21] Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2021. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 4, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 611 + ], + "type": "text", + "content": "[23] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 657 + ], + "type": "text", + "content": "[24] Haohe Liu, Zehua Chen, Yi Yuan, Xinhao Mei, Xubo Liu, Danilo Mandic, Wenwu Wang, and Mark D Plumbley. Audioldm: Text-to-audio generation with latent diffusion models. arXiv preprint arXiv:2301.12503, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Ming Y Lu, Drew FK Williamson, Tiffany Y Chen, Richard J Chen, Matteo Barbieri, and Faisal Mahmood. Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering, 5(6):555-570, 2021. 6, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8540" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 139 + ], + "type": "text", + "content": "[26] Puria Azadi Moghadam, Sanne Van Dalen, Karina C Martin, Jochen Lennerz, Stephen Yip, Hossein Farahani, and Ali Bashashati. A morphology focused diffusion probabilistic model for synthesis of histopathology images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2000-2009, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 215 + ], + "type": "text", + "content": "[27] Gustav Müller-Franzes, Jan Moritz Niehues, Firas Khader, Soroosh Tayebi Arasteh, Christoph Haarburger, Christiane Kuhl, Tianci Wang, Tianyu Han, Teresa Nolte, Sven Nebelung, et al. A multimodal comparison of latent denoising diffusion probabilistic models and generative adversarial networks for medical image synthesis. Scientific Reports, 13 (1):12098, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 259 + ], + "type": "text", + "content": "[28] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, pages 8162-8171. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 327 + ], + "type": "text", + "content": "[29] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, pages 16784-16804. PMLR, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 288, + 360 + ], + "type": "text", + "content": "[30] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. arXiv preprint arXiv:2211.00575, 2022. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 392 + ], + "type": "text", + "content": "[31] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In CVPR, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 288, + 448 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 449, + 288, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 482 + ], + "type": "text", + "content": "[33] António Polónia, Catarina Eloy, and Paulo Aguiar. BACH Dataset : Grand Challenge on Breast Cancer Histology images, 2020. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 483, + 288, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 288, + 547 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 592 + ], + "type": "text", + "content": "[35] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 657 + ], + "type": "text", + "content": "[36] Caleb Robinson, Le Hou, Kolya Malkin, Rachel Soobitsky, Jacob Czawlytko, Bistra Dilkina, and Nebojsa Jojic. Large scale high-resolution land cover mapping with multi-resolution data. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12726-12735, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 714 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 139 + ], + "type": "text", + "content": "[38] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 546, + 194 + ], + "type": "text", + "content": "[39] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 262 + ], + "type": "text", + "content": "[40] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 263, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 546, + 297 + ], + "type": "text", + "content": "[41] Ahmad Sebaq and Mohamed ElHelw. Rsdiff: Remote sensing image generation from text using diffusion model. arXiv preprint arXiv:2309.02455, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 546, + 331 + ], + "type": "text", + "content": "[42] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 332, + 546, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 387 + ], + "type": "text", + "content": "[43] USGS. National agriculture imagery program (NAIP), 2023. https://www.usgs.gov/centers/eros/science/usgs-eros-archive-aerial-photography - national - agriculture - imagery-program-naip.2,4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 546, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 476 + ], + "type": "text", + "content": "[44] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Junzhou Huang, Wei Yang, and Xiao Han. Transpath: Transformer-based self-supervised learning for histopathological image classification. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part VIII 24, pages 186-195. Springer, 2021. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 478, + 546, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 532 + ], + "type": "text", + "content": "[45] Xiyue Wang, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Transformer-based unsupervised contrastive learning for histopathological image classification. Medical image analysis, 81:102559, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 533, + 546, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 546, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 546, + 599 + ], + "type": "text", + "content": "[46] Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. Convnext v2: Co-designing and scaling convnets with masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16133-16142, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 600, + 546, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 600, + 546, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 600, + 546, + 644 + ], + "type": "text", + "content": "[47] Xuan Xu, Saarthak Kapse, Rajarsi Gupta, and Prateek Prasanna. Vit-dae: Transformer-driven diffusion autoencoder for histopathology image analysis. arXiv preprint arXiv:2304.01053, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "text", + "content": "[48] Srikar Yellapragada, Alexandros Graikos, Prateek Prasanna, Tahsin Kurc, Joel Saltz, and Dimitris Samaras. PathLDM: Text conditioned latent diffusion model for histopathology. arXiv preprint arXiv:2309.00748, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 714 + ], + "type": "text", + "content": "[49] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8541" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Valluri, Cliff Wong, et al. Large-scale domain-specific pretraining for biomedical vision-language processing. arXiv preprint arXiv:2303.00915, 2023. 7" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "8542" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_content_list.json b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..939d39c0f63ac86af082a3310873ed44d895c911 --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_content_list.json @@ -0,0 +1,1562 @@ +[ + { + "type": "text", + "text": "Learned Scanpaths Aid Blind Panoramic Video Quality Assessment", + "text_level": 1, + "bbox": [ + 143, + 130, + 826, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kanglong Fan $^{1}$ , Wen Wen $^{1}$ , Mu Li $^{2*}$ , Yifan Peng $^{3}$ , and Kede Ma $^{1}$ $^{1}$ City University of Hong Kong, $^{2}$ Harbin Institute of Technology, Shenzhen \n $^{3}$ The University of Hong Kong \n{kanglofan2-c, wwen29-c}@my.cityu.edu.hk, limuhit@gmail.com \nevanpeng@hku.hk, kede.ma@cityu.edu.hk \nhttps://github.com/kalofan/AutoScanpathQA", + "bbox": [ + 145, + 179, + 825, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 321, + 313, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoramic videos have the advantage of providing an immersive and interactive viewing experience. Nevertheless, their spherical nature gives rise to various and uncertain user viewing behaviors, which poses significant challenges for panoramic video quality assessment (PVQA). In this work, we propose an end-to-end optimized, blind PVQA method with explicit modeling of user viewing patterns through visual scanpaths. Our method consists of two modules: a scanpath generator and a quality assessor. The scanpath generator is initially trained to predict future scanpaths by minimizing their expected code length and then jointly optimized with the quality assessor for quality prediction. Our blind PVQA method enables direct quality assessment of panoramic images by treating them as videos composed of identical frames. Experiments on three public panoramic image and video quality datasets, encompassing both synthetic and authentic distortions, validate the superiority of our blind PVQA model over existing methods.", + "bbox": [ + 75, + 353, + 473, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 655, + 209, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The rapid advancement of multimedia technologies has marked the beginning of a new era characterized by the proliferation of panoramic videos [20]. Such type of digital data offers an immersive and interactive viewing experience that is transforming the way we consume multimedia. Therefore, assessing and ensuring the visual quality of panoramic videos is increasingly important, as it shapes the users' viewing experience and the triumph of any product or service based on panoramic videos [40]. Unlike their planar counterparts, panoramic videos provide a $360^{\\circ}$ broad view with a spherical data structure, which poses significant computational challenges for panoramic video quality assessment (PVQA). Moreover, the diverse and uncertain user", + "bbox": [ + 75, + 680, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fba2825827a098313c5e55c4139537b6fedece13fe40771d2eabae889e529071.jpg", + "image_caption": [ + "Figure 1. Analogy between human subjects and our end-to-end optimized method for panoramic video quality assessment." + ], + "image_footnote": [], + "bbox": [ + 504, + 321, + 895, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "viewing behaviors (in the form of visual scanpaths) induced by the spherical structure further complicate the quality prediction process. Addressing these challenges requires novel PVQA models that take into account both the spherical data structure as well as user viewing patterns.", + "bbox": [ + 496, + 547, + 893, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the quality assessment of panoramic images and videos, three approaches are commonly employed: sphereto-plane projection onto a 2D plane, rectilinear projection onto multiple viewpoints, and direct processing using spherical operations. According to the Theorema Egregium by Gauss, all sphere-to-plane map projections [32, 35, 41] are impeded by non-uniform sampling and geometric distortions, which may bias subsequent planar quality prediction. While spherical operators give a better account for the panoramic data structure, they are generally computationally prohibitive and, more importantly, may not faithfully reflect user viewing patterns [4, 39, 40]. To overcome these computational difficulties, several methods seek to sample and process rectilinear viewpoints [7, 14, 27, 29, 36, 37]. Of particular interest are scanpath-based methods, which sample, along visual scanpaths [23, 24], sequences of rectilinear viewpoints at discrete time instances. This sampling process turns panoramic images and videos into moving-", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2599", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "camera videos, amenable to planar VQA.", + "bbox": [ + 76, + 90, + 354, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By closely imitating how humans perceive visual distortions in virtual environments (see Figure 1), scanpath-based methods [27, 29, 36] have demonstrated remarkable efficacy in the quality of panoramic images. Nonetheless, some methods [27] rely on human visual scanpaths for assessment, which are cumbersome and time-consuming to obtain, thus limiting their applications in fully-automated situations. Some other methods [29, 36] design and refine the scanpath generator separately from the quality predictor, which is bound to be suboptimal. Moreover, while all methods prove effective with panoramic images, their adaptability for use with panoramic videos remains unclear.", + "bbox": [ + 75, + 106, + 467, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we further pursue the scanpath-based methods for end-to-end optimized blind PVQA. Our method consists of two modules: a scanpath generator and a quality assessor. Our scanpath generator is probabilistic, which takes historical scanpaths as input and is pre-trained to predict future scanpaths by minimizing their expected code length [16]. The scanpath generator and the quality assessor are then jointly optimized to explain human perceptual scores of panoramic videos. To enable end-to-end optimization, we employ the reparameterization trick [12, 13] to allow differentiable scanpath sampling and adopt subgradients to handle discontinuities of the interpolation kernel [11] for viewpoint sequence generation. Our blind PVQA method not only eliminates the need for human scanpaths, but also supplies a lightweight and differentiable scanpath generator that can work with any planar VQA model. Furthermore, our method is \"backward compatible,\" in the sense that it handles panoramic images with no modification. We test the proposed blind PVQA models on three public panoramic image and video quality datasets [6, 30, 35], covering both synthetic and authentic distortions. Under both in-dataset and cross-dataset settings, our models consistently exhibit better quality prediction performance.", + "bbox": [ + 75, + 287, + 468, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 662, + 218, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We review two highly relevant topics, scanpath generation and quality assessment of panoramic images and videos.", + "bbox": [ + 76, + 686, + 467, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scanpath Generation", + "text_level": 1, + "bbox": [ + 76, + 726, + 276, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typical inputs to a panoramic scanpath generator include the saliency map, optical flow map, and historical scanpath. To improve saliency detection, Nguyen et al. [21] compiled a panoramic video saliency dataset, while Xu et al. [38] focused on relative viewpoint displacement prediction. Apart from the historical scanpath, Li et al. [15] incorporated \"future\" scanpaths from other users to facilitate cross-user transfer learning. Through an in-depth root-cause analysis, Rondon et al. [25] discovered that visual features have a minimal impact on the prediction of short-term scanpaths", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(e.g., $\\leq 2$ seconds). Motivated by their findings, Chao et al. [3] trained a Transformer [33] to predict future scanpaths based solely on historical scanpaths.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above-mentioned methods [3, 15, 21, 25, 38] treat scanpath generation as a deterministic prediction task, neglecting the inherent scanpath diversity and uncertainty. As a departure, Li et al. [16] formulated scanpath generation as a density estimation problem, which can be implemented by expected code length minimization. In our work, we adopt Li's approach [16] to learn multi-user viewing patterns and generate human-like scanpaths.", + "bbox": [ + 496, + 136, + 890, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Quality Assessment", + "text_level": 1, + "bbox": [ + 500, + 266, + 686, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current PVQA models are primarily derived from planar image and video quality methods, which are applied to three types of data representations: the projected 2D plane, spherical surface, and projected rectilinear viewpoint.", + "bbox": [ + 496, + 287, + 890, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Planar domain methods [32, 35, 41] aim to rectify geometric distortions and mitigate uneven sampling that results from the sphere-to-plane projection. These include the latitude-adaptive weighting [32], Craster parabolic projection [41], and pseudocylindrical representation [35]. Spherical domain methods, such as S-PSNR [40] and S-SSIM [4], compute and pool local quality measurements over the sphere. Yang et al. [39] trained a non-local spherical neural network [5, 34] to extract spatiotemporal information from panoramic videos. Viewport domain methods prioritize the extraction of visually informative viewpoints for quality analysis. Li et al. [14] introduced a two-step approach that involves viewpoint proposal and quality assessment. Xu et al. [37] built a graph over the extracted viewpoints, and Fu et al. [7] constructed hypergraphs to represent the semantic interactions between viewpoints. One limitation of current viewpoint proposal methods is that they do not accurately reflect the human viewing experience.", + "bbox": [ + 496, + 349, + 890, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sui et al. [27] pioneered scanpath-based methods for PVQA, under the category of viewpoint domain methods. To eliminate the dependency on human scanpaths, Sui et al. [29] adopted a deep Markov model [28] to generate scanpaths. Meanwhile, Wu et al. [36] handcrafted a simple scanpath generator based on the entropy feature and equator bias. These methods are tailored for panoramic images and are not end-to-end optimized. In contrast, we aim ambitiously for an end-to-end optimized quality assessment method for panoramic videos, with the added benefit of being backward compatible with panoramic images.", + "bbox": [ + 496, + 619, + 890, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 800, + 671, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our method consists of two modules: a scanpath generator and a quality assessor. Given a panoramic video, we first specify a starting point $(\\phi_0,\\theta_0)$ , a viewing duration $S$ , and $N$ initial paths. The scanpath generator autoregressively samples $N$ scanpaths based on the", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2600", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aaaae09a7792ea19c8d4fd9e2b33dbc66e49f5210b3ee9f93ee2e2e8dde563fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 90, + 885, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8051603bb40278b82cac6cec6539f854f752cc1b70e08b8d374624c7bbd87c3e.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed blind PVQA method, consisting of a scanpath generator and a quality assessor. The basic component of the scanpath generator is the scanpath generation unit (SGU), which utilizes the historical and causal relative scanpaths to produce the GMM parameters for differentiable sampling of the current viewpoint. By assembling $W$ SGUs, we create a scanpath generation block (SGB), which autoregressively predicts a future scanpath of $W$ viewpoints. We further stack $M$ SGBs to generate a long-term scanpath of $M \\times W + H$ viewpoints, where $H$ is the length of the initial path. By adjusting the number of initial paths (denoted by $N$ ), we can sample $N$ scanpaths, along which we produce $N$ viewpoint sequences as input to the quality assessor." + ], + "image_footnote": [], + "bbox": [ + 83, + 228, + 316, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/023c04e5c6b15a1c7a6f0723741c334cbd08ac997346fd0ca773e1003c44d22c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 229, + 887, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "initial and already generated path segments. Along these scanpaths, we apply a differentiable viewpoint generation technique to extract $N$ viewpoint sequences from the input panoramic video. Each viewpoint sequence (as a planar video) is fed to the quality assessor, whose predicted score is subsequently aggregated into an overall quality estimate of the panoramic video.", + "bbox": [ + 75, + 458, + 468, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Scanpath Generator", + "text_level": 1, + "bbox": [ + 76, + 574, + 269, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Probabilistic Scanpath Modeling. To capture the uncertainty and diversity of human scanpaths, we formulate panoramic scanpath generation as a density estimation problem:", + "bbox": [ + 75, + 597, + 468, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max p (\\boldsymbol {r} | \\boldsymbol {s}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 671, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s = \\{(\\phi_0,\\theta_0),\\dots ,(\\phi_t,\\theta_t),\\dots ,(\\phi_{T - 1},\\theta_{T - 1})\\}$ is the historical scanpath as the condition and $\\pmb {r} = \\{(\\phi_T,\\theta_T),\\ldots ,(\\phi_{T + W - 1},\\theta_{T + W - 1})\\}$ is the future scanpath to be predicted. Herein, $W$ is the prediction horizon, and $(\\phi_t,\\theta_t)$ is the $t$ -th viewpoint in the Euler coordinate system. Mathematically, $p(\\boldsymbol {r}|\\boldsymbol {s})$ can be decomposed as", + "bbox": [ + 76, + 699, + 467, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (\\boldsymbol {r} \\mid \\boldsymbol {s}) = \\prod_ {t = 0} ^ {W - 1} p \\left(\\phi_ {T + t}, \\theta_ {T + t} \\mid \\boldsymbol {s}, \\boldsymbol {c} _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 801, + 468, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{c}_t = \\{(\\phi_T, \\theta_T), \\dots, (\\phi_{T + t - 1}, \\theta_{T + t - 1})\\}$ is referred to as the causal path that includes all estimated viewpoints before $(\\phi_{T + t}, \\theta_{T + t})$ , and $c_0 = \\emptyset$ . The", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "chain rule suggests estimating the conditional probability $p\\left(\\phi_{T + t}, \\theta_{T + t} \\mid s, c_t\\right)$ autoregressively. We further make the Markovian assumption: prediction of the current viewpoint is conditionally independent of viewpoints that are temporally further distant, given the most recent $H$ viewpoints. This leads to a truncated historical path context $s_T = \\{(\\phi_{T - H}, \\theta_{T - H}), \\ldots, (\\phi_{T - 1}, \\theta_{T - 1})\\}$ .", + "bbox": [ + 496, + 458, + 890, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We parameterize the probability $p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)$ , where $\\boldsymbol{r}_t = (\\phi_{T + t}, \\theta_{T + t})$ , by a Gaussian mixture model (GMM) with $K$ components:", + "bbox": [ + 496, + 571, + 890, + 623 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {r} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathcal {N} _ {i} \\left(\\boldsymbol {r} _ {t}; \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 636, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ is the $i$ -th mixture weight, $\\pmb{\\mu}_{i}$ and $\\pmb{\\Sigma}_{i}$ represent the mean vector and the covariance matrix of the $i$ -th Gaussian component, respectively. This parametrization can be straightforwardly done by training a density estimation network for parameter estimation. As illustrated in Figure 2, this network is inside the scanpath generation unit (SGU) and is composed of two subnetworks to process the historical path context $s_{T}$ and the causal path context $c_{t}$ , which we denote by H-Net and C-Net, respectively. The concatenated features are fed to three prediction heads to estimate the weight vector, the mean vectors, and the covariance matrices of the GMM, respectively. We find empirically that incorporating the historical video frames as the visual context significantly increases computational demands with", + "bbox": [ + 496, + 689, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2601", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "only slight improvements in performance. Therefore, to keep the scanpath generator lightweight, we choose to omit the visual context. The detailed specifications of the density estimation network can be found in the supplementary material.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Estimating continuous probability density is generally difficult and may lead to overfitting. In particular, maximum likelihood estimation of the GMM parameters through direct optimization of Eq. (3) is challenging, due to the presence of singularities [2]. To circumvent this, we compute the probability mass $P\\left(\\bar{\\boldsymbol{r}}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)$ by discretizing and integrating the density $p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)$ :", + "bbox": [ + 75, + 167, + 467, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\int_ {\\Omega} p \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) d \\Omega . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 295, + 467, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\bar{\\boldsymbol{r}}_t$ represents the quantized value of $\\boldsymbol{r}_t$ by a uniform quantizer with a step size of $\\Delta$ :", + "bbox": [ + 76, + 340, + 467, + 371 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\xi} = Q (\\xi) = \\Delta \\left\\lfloor \\frac {\\xi}{\\Delta} + \\frac {1}{2} \\right\\rfloor , \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 377, + 467, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lfloor \\cdot \\rfloor$ denotes the floor function. $\\Omega = [\\bar{\\phi}_{T + t} - 1 / 2\\Delta, \\bar{\\phi}_{T + t} + 1 / 2\\Delta] \\times [\\bar{\\theta}_{T + t} - 1 / 2\\Delta, \\bar{\\theta}_{T + t} + 1 / 2\\Delta]$ is the integration interval. As pointed out in [16], the incorporation of quantization establishes the equivalence between scanpath generation and lossy scanpath compression.", + "bbox": [ + 76, + 416, + 467, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, the absolute Euler coordinate system is not user-centric, meaning that it is not centered at the user's current viewpoint, relative to historical and future viewpoints [16]. This may complicate the probabilistic modeling of scanpaths and the end-to-end optimization of blind PVQA. To address this, we convert the Euler coordinates to the relative $uv$ coordinates:", + "bbox": [ + 75, + 492, + 467, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {s}} _ {T - t} = \\boldsymbol {\\Psi} _ {T - t} (\\boldsymbol {s} _ {T}), \\text {f o r} t \\in \\{1, \\dots , H \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 606, + 467, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Psi_{T - t}(\\cdot)$ denotes the mapping of $s_T$ to the viewpoint centered at the reference viewpoint $(\\phi_{T - t},\\theta_{T - t})$ . By choosing each viewpoint in $s_T$ as the reference, we create $H$ relative scanpaths out of $s_T$ (see Figure 3), which serve as input to the H-Net. Meanwhile, we map the causal scanpath context $c_{t}$ and the viewpoint to be predicted $(\\phi_{T + t},\\theta_{T + t})$ to the last historical viewpoint centered at $(\\phi_{T - 1},\\theta_{T - 1})$ . We stack $W$ SGUs to form a scanpath generation block (SGB), which takes $s_T$ as input and predicts the future scanpath $r$ . Furthermore, we stack $M$ SGBs to form the scanpath generator, which is capable of predicting a very long-term scanpath of length $M\\times W + H$ (including the initial length $H$ ). The parameters of different SGUs are shared to enable variable-length scanpath generation by varying $M$ .", + "bbox": [ + 75, + 628, + 467, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable Scanpath Sampling. To enable end-to-end optimization of the proposed blind PVQA method, we propose a two-step differentiable sampling method to draw", + "bbox": [ + 75, + 854, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f180e081f6730f63d7e3aa992ba4370010d3b799f50acdea1a0bdf5df94e3a34.jpg", + "image_caption": [ + "Figure 3. Visualization of a relative scanpath projected from the sphere to the viewpoint." + ], + "image_footnote": [], + "bbox": [ + 540, + 84, + 851, + 207 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "viewpoints from the estimated GMM via the reparameterization trick [12, 13]. The first step is to select a Gaussian component from which to sample the viewpoint, according to the categorical distribution:", + "bbox": [ + 498, + 260, + 890, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {e} = \\text {o n e} _ {\\text {h o t}} \\left(\\underset {i \\in \\{1, \\dots K \\}} {\\arg \\max } \\left(\\log \\left(\\alpha_ {i}\\right) + g _ {i}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 329, + 890, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{i}$ is a sample drawn from the Gumbel(0,1) distribution, and $e$ is a one-hot vector. Eq. (7) is known as the Gumbel-Max trick [8], which is non-differentiable. We relax the arg max operator with a softmax function [12]:", + "bbox": [ + 498, + 380, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {e}} = \\operatorname {s o f t m a x} \\left(\\left(\\log (\\boldsymbol {\\alpha}) + \\boldsymbol {g}\\right) / \\tau\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 450, + 890, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ represents the temperature coefficient, $\\alpha = [\\alpha_{1},\\dots,\\alpha_{K}]$ , and $\\pmb {g} = [g_1,\\dots,g_K]$ . As $\\tau$ approaches zero, $\\hat{e}$ converges to $e$ . In the forward pass, $\\arg \\max$ is used directly, while in the backward pass, it is replaced by the softmax function. The second step involves sampling a viewpoint from the selected Gaussian component. Assuming the $i$ -th Gaussian component is selected, the linear reparameterization trick [13] suggests", + "bbox": [ + 498, + 478, + 890, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {L} _ {i} \\epsilon , \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 609, + 890, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\boldsymbol{r}}_t$ is the relative $uv$ coordinates of the $t$ -th future viewpoint. $\\boldsymbol{\\Sigma}_i = \\boldsymbol{L}_i\\boldsymbol{L}_i^{\\mathrm{T}}$ is the Cholesky decomposition, and $\\epsilon$ is a sample drawn from the $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . Differentiation of the Cholesky decomposition is complicated and sometimes numerically unstable [26]. Thus we assume the independence between the $uv$ coordinates, leading to the simplified reparameterization formula:", + "bbox": [ + 498, + 636, + 890, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 753, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes the element-wise product, and $\\sigma_{i}$ denotes the standard deviations of the $i$ -th Gaussian component. Through the two-step reparameterization, our sampling strategy ensures effective back-propagation. As suggested in [16], we additionally implement a proportional-integral-derivative (PID) controller [1] to further improve the smoothness of the sampled scanpaths (see the details in the supplementary material).", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2602", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differentiable Viewport Sequence Generation. Inspired by [11], we first parameterize the Euler sampling grid in terms of the relative $uv$ coordinates via the inverse transformation $\\Psi^{-1}$ (see Eq. (6)). Subsequently, the Euler coordinate $(\\phi, \\theta)$ is mapped to the discrete sampling position $(m, n)$ in the ERP domain:", + "bbox": [ + 76, + 90, + 472, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm = (0. 5 - \\phi / \\pi) H _ {e} - 0. 5, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 191, + 468, + 208 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nn = (\\theta / 2 \\pi + 0. 5) W _ {e} - 0. 5, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 210, + 468, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{e}$ and $W_{e}$ are the height and width of the video frame in ERP. Once the mapping between $(u, v)$ and $(m, n)$ is established, we construct a flow field that is the same size as the viewpoint. Within this field, each element records the corresponding pixel position $(m, n)$ . Given an ERP image and the flow field, we apply bilinear interpolation [11] to compute pixel values in the viewpoint and leverage its sub-gradients for back-propagation. This process yields $N$ viewpoint sequences, corresponding to $N$ initial paths.", + "bbox": [ + 75, + 236, + 470, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Quality Assessor", + "text_level": 1, + "bbox": [ + 76, + 380, + 243, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our probabilistic scanpath generator can work with any planar VQA model, whether it is differentiable or not. To enable end-to-end optimization of the scanpath generator and the quality assessor, and to make a fair comparison with existing blind PVQA models, we reuse three differentiable quality assessors from ScanpathVQA [35], GSR-S / GSR-X [29], and Assessor360 [36]. Specifically, the quality assessor of ScanpathVQA is a lightweight ResNet-18 network [9], with the classification head replaced by a quality estimator (a multilayer perceptron). The quality assessors of GSR-S / GSR-X are adapted from Video Swin-T [18] / X-Clip-B/32 [22]. The quality assessor of Assessor360 is modified from Swin-B [17] with the addition of a temporal analysis module. We feed each of the $N$ viewpoint sequences to the quality assessor to compute $N$ quality scores. The overall quality estimate is then computed by a simple average:", + "bbox": [ + 75, + 402, + 472, + 660 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\hat {q} _ {i}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 669, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization Strategy", + "text_level": 1, + "bbox": [ + 76, + 719, + 285, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We explore a three-stage training procedure for our blind PVQA model. In the first stage, we pre-train the density estimation network on the VRVQW dataset [35] by minimizing the expected code length of the generated scanpaths (also equivalent to minimizing the negative log-likelihood):", + "bbox": [ + 75, + 742, + 470, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {c o d e}} = - \\frac {1}{B W} \\sum_ {i = 1} ^ {B} \\sum_ {t = 0} ^ {W - 1} \\log_ {2} \\left(P \\left(\\bar {\\boldsymbol {r}} _ {t} ^ {(i)} \\mid \\boldsymbol {s} ^ {(i)}, \\boldsymbol {c} _ {t} ^ {(i)}\\right)\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 839, + 468, + 896 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ denotes the mini-batch size. During this stage of training, we use human scanpaths to fill in the causal path context, which can be efficiently implemented by a causal masking mechanism. In the second stage, we fix the parameters of the pre-trained scanpath generator and warm up the quality assessor by optimizing the Pearson linear correlation coefficient (PLCC) between human perceptual scores and model predictions. In the third stage, we end-to-end finetune the entire method. We find that the proposed three-stage optimization strategy accelerates convergence compared to the naive end-to-end optimization.", + "bbox": [ + 496, + 90, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 268, + 633, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first delineate the experimental setups, and then compare our method with current blind PVQA models under both in-dataset and cross-dataset settings. We further validate our scanpath generator in terms of explaining human perceptual scores and replicating human viewing patterns. Lastly, we conduct a series of ablation experiments to probe the impact of several key designs.", + "bbox": [ + 496, + 295, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 498, + 409, + 697, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We employ three panoramic image and video datasets: VRVQW [35], CVIQD [30], and OIQA [6]. The VRVQW dataset includes 502 panoramic videos that have a wide spectrum of authentic distortions. Each video is viewed under four unique viewing conditions to simulate the different quality of experience during the initial viewing. The CVIQD dataset comprises a total of 528 compressed panoramic images by JPEG, AVC, and HEVC, from 16 reference images. The OIQA dataset includes 320 panoramic images that have been altered from 16 reference images by JPEG compression, JPEG2000 compression, Gaussian blurring, and Gaussian noise contamination.", + "bbox": [ + 496, + 431, + 890, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For the scanpath generator, we set the length of the provided initial path $H$ and the predicted future path $W$ in the SGB to be identical and equal to 5. The number of Gaussian components $K$ in Eq. (3) is set to 3. The quantization step size $\\Delta$ in Eq. (5) is set to 0.2. The temperature coefficient $\\tau$ in Eq. (8) is set to 1. The number of stacked SGBs $M$ is set to 6 and 14 for the viewing duration of 7 and 15 seconds, respectively. For the quality assessor, the number of scanpaths $N$ in Eq. (13) is set to 20. The input viewpoint size $H_{v} \\times W_{v}$ is $224 \\times 224$ , corresponding to a field of view of $90^{\\circ} \\times 90^{\\circ}$ . The length of the viewpoint sequence is set to 7 regardless of the duration and frame rate of the original panoramic video. We split VRVQW randomly into the training, validation, and test sets according to the ratio of $6:2:2$ for 5 times, and report the mean results. Similarly, we split CVIQD and OIQA using a different ratio of $7:1:2$ for 5 times. The detailed configuration of our three-stage optimization strategy is detailed in the supplementary material.", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2603", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f152c502592d28a36bd95dc4e5e5e1c22f786d2d9462338de29530b6958c58a6.jpg", + "table_caption": [ + "Table 1. In-dataset comparison of blind PVQA methods on three panoramic image and video quality datasets. SRCC: Spearman's rank correlation coefficient. PLCC: Pearson linear correlation coefficient. The same evaluation metrics are applied in Tables 2, 3, 5, 6 and Figure 6. The best results on each dataset are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSRCCPLCC
VRVQW [35]NIQE [19]0.4010.365
MC360IQA [31]0.6690.671
Wen24 [35]0.7560.763
ScanpathVQA [35]0.7790.781
Assessor360 [36]0.4060.415
Ours (ScanpathVQA)0.8010.809
Ours (GSR-S)0.8040.807
Ours (GSR-X)0.8150.819
Ours (Assessor360)0.8220.823
CVIQD [30]NIQE [19]0.8470.878
MC360IQA [31]0.9170.939
Wen24 [35]0.9190.932
GSR-S [29]0.9050.937
GSR-X [29]0.9440.962
Assessor360 [36]0.9550.969
Ours (ScanpathVQA)0.9120.936
Ours (GSR-S)0.9300.958
Ours (GSR-X)0.9560.974
Ours (Assessor360)0.9720.983
OIQA [6]NIQE [19]0.7020.657
MC360IQA [31]0.9000.906
Wen24 [35]0.9050.907
GSR-S [29]0.9020.915
GSR-X [29]0.9450.954
Assessor360 [36]0.9460.953
Ours (ScanpathVQA)0.9150.922
Ours (GSR-S)0.9270.936
Ours (GSR-X)0.9560.967
Ours (Assessor360)0.9600.971
", + "bbox": [ + 106, + 183, + 444, + 589 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9283e8d4e44e662c2805909824640f32acffbd6d5f5424f4ba78878bf1d43971.jpg", + "table_caption": [ + "Table 2. Cross-dataset comparison of blind PVQA methods on the CVIQD [30] and OIQA [6] datasets. The arrow points from the training set to the test set." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSRCCPLCC
OIQAMC360IQA0.7980.842
GSR-X0.7620.841
Assessor3600.8590.893
CVIQDOurs (ScanpathVQA)0.7330.747
Ours (Assessor360)0.8720.904
CVIQDMC360IQA0.2880.349
GSR-X0.6950.718
Assessor3600.3380.467
OIQAOurs (ScanpathVQA)0.6360.658
Ours (Assessor360)0.7030.715
", + "bbox": [ + 117, + 646, + 426, + 813 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 829, + 217, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our blind PVQA method with seven existing models, including NIQE [19], MC360IQA [31], Wen24 [35], ScanpathVQA [35], GSR-S [29], GSR-X [29],", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/026be187e8de684a0dc88f86de2e14b35d342fdbe43c92a999249945ebb6818d.jpg", + "table_caption": [ + "Table 3. Comparison of different scanpath generators for explaining human perceptual scores." + ], + "table_footnote": [], + "table_body": "
DatasetMethodSRCCPLCC
VRVQWHuman scanpath0.7860.790
Random sampling0.0750.104
Heuristic sampling [36]0.4310.443
Xu18 [38]0.7120.717
TRACK [25]0.7450.749
Li23 [16]0.7900.794
Ours0.8050.814
CVIQDRandom sampling0.6320.640
Heuristic sampling [36]0.8680.870
ScanDMM [28]0.8560.864
Li23 [16]0.8140.827
Ours0.9280.940
OIQARandom sampling0.5140.536
Heuristic sampling [36]0.8610.872
ScanDMM [28]0.8650.877
Li23 [16]0.7930.799
Ours0.9140.917
", + "bbox": [ + 531, + 121, + 857, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and Assessor360 [36]. For image quality models such as MC360IQA [31] and Assessor360 [36], we retrain them on the VRVQW dataset. In the case of MC360IQA, we assign the video-level quality score to each key frame and use their temporally averaged score for testing. Assessor360's scanpath generator is adapted to videos by adjusting the semantic context associated with each key frame.", + "bbox": [ + 496, + 378, + 890, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-dataset Results. Table 1 shows the Spearman's rank correlation coefficient (SRCC) and $\\mathrm{PLCC}^1$ results under the in-dataset setting. It is evident that our learned scanpaths enhance the performance of all quality assessors compared to other scanpath-based methods. When integrated with the quality assessor from Assessor360 (i.e., a modified SwinB with a temporal analysis module), our proposed method achieves the best results on all three datasets. Furthermore, our scanpath generator can boost a simpler quality assessor (e.g., from ScanpathVQA [35] with approximately 11 million parameters) to reach performance levels similar to those of a more sophisticated quality assessor coupled with a weaker scanpath generator (e.g., GSR-S [29] with approximately 112 million parameters). Methods that overlook human viewing patterns, like NIQE [19] and MC360IQA [31], fail to accurately model the human perception of panoramic image and video quality, especially on VRVQW. Additionally, we find that assessing the quality of panoramic videos with authentic distortions tends to be more challenging than for panoramic images with synthetic distortions. This is anticipated because authentic distortions in panoramic videos typically present as a complex blend of various artifacts, localized in space and time.", + "bbox": [ + 496, + 484, + 892, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-dataset Results. Table 2 shows the SRCC and PLCC results under the cross-dataset settings on CVIQD [30] and", + "bbox": [ + 498, + 832, + 893, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1As standard practice, we apply a monotonic logistic function to compensate for the nonlinearity in model predictions before computing PLCC.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2604", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b751be8dd20d7ae0490c8cc01ac6bca8d388dcd7a2da5bdc7ee8643680c4e67e.jpg", + "table_caption": [ + "Table 4. Comparison of different scanpath generators for replicating human viewing patterns using the minimum orthodromic distance (minOD) and maximum temporal correlation (maxTC)." + ], + "table_footnote": [], + "table_body": "
MethodminOD ↓maxTC ↑
Heuristic sampling [36]1.3250.401
Xu18 [38]1.1850.637
TRACK [25]1.0670.699
Li23 [16]0.5420.796
Ours (w/o end-to-end optimization)0.5560.781
Ours0.5360.805
", + "bbox": [ + 91, + 133, + 454, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72c6a5319bac4183d4a9c5b93a002c2f646bc3aafcfbcdd34a916aaa792d3311.jpg", + "image_caption": [ + "Figure 4. Comparison of different scanpath predictors in terms of maxTC with different prediction horizons." + ], + "image_footnote": [], + "bbox": [ + 135, + 258, + 411, + 421 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OIQA [6]. Generally, models trained on the OIQA dataset show better generalizability. This is likely because OIQA encompasses a broader range of distortion types, compared to CVIQD, which only includes the compression artifacts. Assessor360 shows a noticeable performance drop when tested on OIQA, potentially indicative of overfitting. Our scanpath generator is capable of restoring Assessor360's performance, which provides a strong indication of its effectiveness through end-to-end optimization.", + "bbox": [ + 75, + 469, + 472, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Scanpath Generator Validation", + "text_level": 1, + "bbox": [ + 76, + 618, + 354, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Explaining Human Perceptual Scores. We conduct an apple-to/apple comparison of different scanpath generators in terms of explaining human perceptual scores by fixing the quality assessor to that used in ScanpathVQA. These include random sampling, heuristic sampling [36], Xu18 [38], TRACK [25], ScanDMM [28], Li23 [16], and our method. Table 3 shows the SRCC and PLCC results on the VRVQW, CVIQD and OIQA datasets. It is noteworthy that our scanpath generator outperforms all competing methods across all three datasets, even surpassing the human-level performance on VRVQW. The performance of random and heuristic sampling decreases sharply on VRVQW, due to the presence of spatiotemporally localized authentic distortions.", + "bbox": [ + 75, + 642, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Replicating Human Viewing Patterns. We also test different scanpath generators in terms of replicating human viewing patterns on the VRVQW dataset by comparing the predicted scanpaths to those of humans. We use two set-to-set", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5fc3e3debab2a7957e0a7b3088cf52690a9fb15f4afbc1aa1924bd9d8c941d2a.jpg", + "image_caption": [ + "Figure 5. Comparison of saliency maps generated from scanpaths by our method and those by humans." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 887, + 220 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3f7c5a1af1bb8b8efbe1299b950feaf24a54c69fa49b2df2f595455e37ccc029.jpg", + "table_caption": [ + "Table 5. Impact of optimization strategies on blind PVQA. Training protocol: 1) Two-stage w/ fixed scanpaths, 2) Two-stage w/ varied scanpaths, and 3) Three-stage w/ end-to-end optimization." + ], + "table_footnote": [], + "table_body": "
ProtocolVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
10.7690.7720.7100.7800.6880.742
20.7810.7850.8300.8590.7980.815
30.8050.8140.9280.9400.9140.917
", + "bbox": [ + 501, + 314, + 901, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "evaluation metrics: the minimum orthodromic distance (i.e., minOD) and maximum temporal correlation (i.e., maxTC), as suggested in [16]. Given a set of human scanpaths, $S = \\{s^{(i)}\\}_{i=1}^{|S|}$ , the minimum orthodromic distance between $S$ and the set of predicted scanpaths $\\hat{S} = \\{\\hat{s}^{(i)}\\}_{i=1}^{|S|}$ can be computed by", + "bbox": [ + 498, + 404, + 890, + 502 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\mathrm {O D} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\min _ {\\boldsymbol {s} \\in \\mathcal {S}, \\boldsymbol {s} \\in \\hat {\\mathcal {S}}} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 512, + 890, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the orthodromic distance $\\mathrm{OD}(\\cdot ,\\cdot)$ is defined as", + "bbox": [ + 500, + 554, + 854, + 570 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\operatorname {a r c c o s} \\left(\\cos \\left(\\phi_ {t}\\right) \\cos \\left(\\hat {\\phi} _ {t}\\right) \\cos \\left(\\theta_ {t} - \\hat {\\theta} _ {t}\\right) \\right. \\\\ + \\sin \\left(\\phi_ {t}\\right) \\sin \\left(\\hat {\\phi} _ {t}\\right)). \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 582, + 890, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The maximum temporal correlation between $\\mathcal{S}$ and $\\hat{\\mathcal{S}}$ is defined as", + "bbox": [ + 498, + 680, + 890, + 710 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\mathrm {T C} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\max _ {\\boldsymbol {s} \\in \\mathcal {S}, \\hat {\\boldsymbol {s}} \\in \\hat {\\mathcal {S}}} \\mathrm {T C} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 720, + 890, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where the temporal correlation is computed by", + "bbox": [ + 500, + 762, + 808, + 777 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {T C} \\left(\\boldsymbol {s} ^ {(i)}, \\boldsymbol {s} ^ {(j)}\\right) = \\\\ \\frac {1}{2} \\left(\\operatorname {P L C C} \\left(\\phi^ {(i)}, \\phi^ {(j)}\\right) + \\operatorname {P L C C} \\left(\\boldsymbol {\\theta} ^ {(i)}, \\boldsymbol {\\theta} ^ {(j)}\\right)\\right). \\end{array} \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 787, + 890, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents the minOD and maxTC results, from which we find that our end-to-end optimized scanpath generator delivers the best results, surpassing its independently", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6113d7a9c53f552e9e3567a67957980a972a1f1de90693679c2332c493cf6afc.jpg", + "table_caption": [ + "Table 6. Impact of visual context on blind PVQA. #Parameters added to the scanpath generator are also shown." + ], + "table_footnote": [], + "table_body": "
Method#ParametersVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
Ours1M0.8050.8140.9280.9400.9140.917
Ours w/ visual context27M0.8160.8230.9370.9560.9250.934
", + "bbox": [ + 202, + 107, + 763, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "optimized counterpart by a clear margin. The heuristic sampling [36] that depends on the simplified entropy features and equator bias, is inadequate for capturing human viewing patterns, especially for long-term prediction horizons (see Figure 4). Due to the deterministic nature, Xu18 [38] and TRACK [25] fail to accommodate the diversity and uncertainty inherent in human scanpaths, resulting in subpar performance. Incorporating historical video frames as the visual context, Li23 [16] shows performance on par with our method, reinforcing our assertion that visual context informs less about future viewpoints. Figure 5 demonstrates a comparison of the saliency maps derived from scanpaths by our method and those by humans, offering further proof of the close alignment of our scanpath generator and human viewing behaviors.", + "bbox": [ + 75, + 189, + 470, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 424, + 241, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Optimization Strategies. We explore three different optimization strategies: 1) a two-stage approach where the pre-trained scanpath generator produces a fixed set of scanpaths for the training of the quality assessor, 2) a similar two-stage approach but supplying a varied set of scanpaths in each epoch of training, and 3) the default three-stage approach that enables end-to-end optimization. From the results in Table 5, we find that the two-stage approach benefits from \"data augmentation\" with varied scanpaths in each epoch. Our three-stage end-to-end optimization strategy further boosts the accuracy of quality prediction by jointly finetuning both the scanpath generator and quality assessor.", + "bbox": [ + 75, + 446, + 468, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Visual Context. To assess the impact of visual context on blind PVQA, we enhance our scanpath generator with a video analysis network [16], implemented by a variant of ResNet-50 for frame-level feature extraction and aggregation. We subsequently integrate it with the ScanpathVQA quality assessor for blind PVQA, with the results shown in Table 6. We find that the visual context has a negligible effect on blind PVQA, and thus we exclude it in the generation of scanpaths.", + "bbox": [ + 75, + 643, + 468, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the Number and Length of Viewport Sequences. We investigate the effects of varying the number $N$ and length $L$ of viewpoint sequences on blind PVQA. We test $N$ values from $\\{5,10,15,20,50\\}$ and $L$ values from $\\{4,7,15\\}$ . Figure 6 shows the SRCC results on VRVQW, using the ScanpathVQA quality assessor for prediction. It is clear that $N = 20$ viewpoint sequences are sufficient for reliable quality assessment, with performance remaining stable", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3abb9f809bf7a66d4ab2271e69ac1348ec74016d2cabd6e9deb12476f2bcad2.jpg", + "image_caption": [ + "Figure 6. Impact of the number $N$ and length $L$ of viewpoint sequences on blind PVQA." + ], + "image_footnote": [], + "bbox": [ + 555, + 188, + 834, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with the increase in $N$ . For sequence length, $L = 7$ appears to be a wise choice. Further increasing $L$ does not noticeably affect the performance, but would lead to a considerable rise in computational demand. Conversely, a shorter viewpoint sequence results in a noticeable drop in performance due to the loss of information from excessive temporal downsampling. It is important to note that these findings are specific to the ScanpathVQA quality assessor and may differ from other assessors.", + "bbox": [ + 496, + 406, + 890, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 559, + 619, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced an end-to-end optimized blind PVQA method, consisting of a scanpath generator and a quality assessor. The proposed scanpath generator is differentiable and can be integrated with any planar VQA model, whose effectiveness has been thoroughly validated in supporting blind PVQA and in modeling human viewing patterns. Additionally, we have also devised a three-stage optimization strategy to facilitate training convergence, which aligns with current large-scale optimization practices that involve self-supervised pre-training followed by supervised finetuning, including initial warmup phases [10].", + "bbox": [ + 496, + 585, + 890, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 768, + 679, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (62071407 and 62102339), the Shenzhen Science and Technology Program (RCBS20221008093121052), the Research Grants Council of Hong Kong (ECS 2121382, ECS 27212822, and GRF 17201822), and the CCF-Tencent Rhino Bird Fund (9239061).", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2606", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kiam Heong Ang, Gregory Chong, and Yun Li. PID control system analysis, design, and technology. IEEE Transactions on Control Systems Technology, 13(4):559-576, 2005. 4", + "[2] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 4", + "[3] Fang-Yi Chao, Cagri Ozcinar, and Aljosa Smolic. Transformer-based long-term viewpoint prediction in $360^{\\circ}$ video: Scanpath is all you need. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2021. 2", + "[4] Sijia Chen, Yingxue Zhang, Yiming Li, Zhenzhong Chen, and Zhou Wang. Spherical structural similarity index for objective omnidirectional video quality assessment. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2018. 1, 2", + "[5] Taco S. Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, pages 1-15, 2018. 2", + "[6] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Yucheng Zhu, Yi Fang, and Xiaokang Yang. Perceptual quality assessment of omnidirectional images. In IEEE International Symposium on Circuits and Systems, pages 1-5, 2018. 2, 5, 6, 7", + "[7] Jun Fu, Chen Hou, Wei Zhou, Jiahua Xu, and Zhibo Chen. Adaptive hypergraph convolutional network for no-reference 360-degree image quality assessment. In ACM International Conference on Multimedia, pages 961-969, 2022. 1, 2", + "[8] Emil J. Gumbel. Statistical Theory of Extreme Values and Some Practical Applications. US Government Printing Office, 1948. 4", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5", + "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8", + "[11] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017-2025, 2015. 2, 5", + "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In International Conference on Learning Representations, pages 1–12, 2017. 2, 4", + "[13] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, pages 1–14, 2014. 2, 4", + "[14] Chen Li, Mai Xu, Lai Jiang, Shanyi Zhang, and Xiaoming Tao. Viewport proposal CNN for $360^{\\circ}$ video quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10169-10178, 2019. 1, 2", + "[15] Chenge Li, Weixi Zhang, Yong Liu, and Yao Wang. Very long term field of view prediction for 360-degree video streaming. In IEEE Conference on Multimedia Information Processing and Retrieval, pages 297-302, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Mu Li, Kanglong Fan, and Kede Ma. Scanpath prediction in panoramic videos via expected code length minimization. arXiv preprint arXiv:2305.02536, 2023. 2, 4, 6, 7, 8", + "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 5", + "[18] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 5", + "[19] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 6", + "[20] King-To Ng, Shing-Chow Chan, and Heung-Yeung Shum. Data compression and transmission aspects of panoramic videos. IEEE Transactions on Circuits and Systems for Video Technology, 15(1):82-95, 2005. 1", + "[21] Anh Nguyen, Zhisheng Yan, and Klara Nahrstedt. Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In ACM International Conference on Multimedia, pages 1190-1198, 2018. 2", + "[22] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18, 2022. 5", + "[23] David Noton and Lawrence Stark. Scanpaths in saccadic eye movements while viewing and recognizing patterns. Vision Research, 11(9):929-942, 1971. 1", + "[24] David Noton and Lawrence Stark. Scanpaths in eye movements during pattern perception. Science, 171(3968):308-311, 1971. 1", + "[25] Miguel Fabian Romero Rondon, Lucile Sassatelli, Ramón Aparicio-Pardo, and Frédéric Precioso. TRACK: A new method from a re-examination of deep architectures for head motion prediction in $360^{\\circ}$ videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5681-5699, 2022. 2, 6, 7, 8", + "[26] Stephen P. Smith. Differentiation of the Cholesky algorithm. Journal of Computational and Graphical Statistics, 4(2):134-147, 1995. 4", + "[27] Xiangjie Sui, Kede Ma, Yiru Yao, and Yuming Fang. Perceptual quality assessment of omnidirectional images as moving camera videos. IEEE Transactions on Visualization and Computer Graphics, 28(8):3022-3034, 2021. 1, 2", + "[28] Xiangjie Sui, Yuming Fang, Hanwei Zhu, Shiqi Wang, and Zhou Wang. ScanDMM: A deep markov model of scanpath prediction for $360^{\\circ}$ images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6989-6999, 2023. 2, 6, 7", + "[29] Xiangjie Sui, Hanwei Zhu, Xuelin Liu, Yuming Fang, Shiqi Wang, and Zhou Wang. Perceptual quality assessment of $360^{\\circ}$ images based on generative scanpath representation. arXiv preprint arXiv:2309.03472, 2023. 1, 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Wei Sun, Ke Gu, Siwei Ma, Wenhan Zhu, Ning Liu, and Guangtao Zhai. A large-scale compressed 360-degree spherical image database: From subjective quality evaluation to objective model comparison. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2018. 2, 5, 6", + "[31] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Huiyu Duan, and Siwei Ma. MC360IQA: A multi-channel CNN for blind 360-degree image quality assessment. IEEE Journal of Selected Topics in Signal Processing, 14(1):64-77, 2020. 6", + "[32] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters, 24(9):1408-1412, 2017. 1, 2", + "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000-6010, 2017. 2", + "[34] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 2", + "[35] Wen Wen, Mu Li, Yiru Yao, Xiangjie Sui, Yabin Zhang, Long Lan, Yuming Fang, and Kede Ma. Perceptual quality assessment of virtual reality videos in the wild. IEEE Transactions on Circuits and Systems for Video Technology, to appear, 2024. 1, 2, 5, 6", + "[36] Tianhe Wu, Shuwei Shi, Haoming Cai, Mingdeng Cao, Jing Xiao, Yinqiang Zheng, and Yujiu Yang. Assessor360: Multisession network for blind omnidirectional image quality assessment. In Advances in Neural Information Processing Systems, pages 1-22, 2023. 1, 2, 5, 6, 7, 8", + "[37] Jiahua Xu, Wei Zhou, and Zhibo Chen. Blind omnidirectional image quality assessment with viewpoint oriented graph convolutional networks. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1724-1737, 2020. 1, 2", + "[38] Yanyu Xu, Yanbing Dong, Junru Wu, Zhengzhong Sun, Zhiru Shi, Jingyi Yu, and Shenghua Gao. Gaze prediction in dynamic $360^{\\circ}$ immersive videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5333-5342, 2018. 2, 6, 7, 8", + "[39] Jiachen Yang, Tianlin Liu, Bin Jiang, Wen Lu, and Qinggang Meng. Panoramic video quality assessment based on non-local spherical CNN. IEEE Transactions on Multimedia, 23: 797-809, 2021. 1, 2", + "[40] Matt Yu, Haricharan Lakshman, and Bernd Girod. A framework to evaluate omnidirectional video coding schemes. In IEEE International Symposium on Mixed and Augmented Reality, pages 31-36, 2015. 1, 2", + "[41] Vladyslav Zakharchenko, Kwang Pyo Choi, and Jeong Hoon Park. Quality metric for spherical panoramic video. In SPIE Optics and Photonics for Information Processing X, pages 57-65, 2016. 1, 2" + ], + "bbox": [ + 78, + 90, + 470, + 838 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2608", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_model.json b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c8b76e322dc521bbafca27689d4a228fad908521 --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.131, + 0.828, + 0.152 + ], + "angle": 0, + "content": "Learned Scanpaths Aid Blind Panoramic Video Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.18, + 0.826, + 0.288 + ], + "angle": 0, + "content": "Kanglong Fan\\(^{1}\\), Wen Wen\\(^{1}\\), Mu Li\\(^{2*}\\), Yifan Peng\\(^{3}\\), and Kede Ma\\(^{1}\\) \n\\(^{1}\\) City University of Hong Kong, \\(^{2}\\) Harbin Institute of Technology, Shenzhen \n\\(^{3}\\) The University of Hong Kong \n{kanglofan2-c, wwen29-c}@my.cityu.edu.hk, limuhit@gmail.com \nevanpeng@hku.hk, kede.ma@cityu.edu.hk \nhttps://github.com/kalofan/AutoScanpathQA" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.322, + 0.314, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.354, + 0.474, + 0.626 + ], + "angle": 0, + "content": "Panoramic videos have the advantage of providing an immersive and interactive viewing experience. Nevertheless, their spherical nature gives rise to various and uncertain user viewing behaviors, which poses significant challenges for panoramic video quality assessment (PVQA). In this work, we propose an end-to-end optimized, blind PVQA method with explicit modeling of user viewing patterns through visual scanpaths. Our method consists of two modules: a scanpath generator and a quality assessor. The scanpath generator is initially trained to predict future scanpaths by minimizing their expected code length and then jointly optimized with the quality assessor for quality prediction. Our blind PVQA method enables direct quality assessment of panoramic images by treating them as videos composed of identical frames. Experiments on three public panoramic image and video quality datasets, encompassing both synthetic and authentic distortions, validate the superiority of our blind PVQA model over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.656, + 0.21, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.877 + ], + "angle": 0, + "content": "The rapid advancement of multimedia technologies has marked the beginning of a new era characterized by the proliferation of panoramic videos [20]. Such type of digital data offers an immersive and interactive viewing experience that is transforming the way we consume multimedia. Therefore, assessing and ensuring the visual quality of panoramic videos is increasingly important, as it shapes the users' viewing experience and the triumph of any product or service based on panoramic videos [40]. Unlike their planar counterparts, panoramic videos provide a \\(360^{\\circ}\\) broad view with a spherical data structure, which poses significant computational challenges for panoramic video quality assessment (PVQA). Moreover, the diverse and uncertain user" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.322, + 0.897, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.499, + 0.895, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Analogy between human subjects and our end-to-end optimized method for panoramic video quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.894, + 0.626 + ], + "angle": 0, + "content": "viewing behaviors (in the form of visual scanpaths) induced by the spherical structure further complicate the quality prediction process. Addressing these challenges requires novel PVQA models that take into account both the spherical data structure as well as user viewing patterns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In the quality assessment of panoramic images and videos, three approaches are commonly employed: sphereto-plane projection onto a 2D plane, rectilinear projection onto multiple viewpoints, and direct processing using spherical operations. According to the Theorema Egregium by Gauss, all sphere-to-plane map projections [32, 35, 41] are impeded by non-uniform sampling and geometric distortions, which may bias subsequent planar quality prediction. While spherical operators give a better account for the panoramic data structure, they are generally computationally prohibitive and, more importantly, may not faithfully reflect user viewing patterns [4, 39, 40]. To overcome these computational difficulties, several methods seek to sample and process rectilinear viewpoints [7, 14, 27, 29, 36, 37]. Of particular interest are scanpath-based methods, which sample, along visual scanpaths [23, 24], sequences of rectilinear viewpoints at discrete time instances. This sampling process turns panoramic images and videos into moving-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2599" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.356, + 0.106 + ], + "angle": 0, + "content": "camera videos, amenable to planar VQA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.287 + ], + "angle": 0, + "content": "By closely imitating how humans perceive visual distortions in virtual environments (see Figure 1), scanpath-based methods [27, 29, 36] have demonstrated remarkable efficacy in the quality of panoramic images. Nonetheless, some methods [27] rely on human visual scanpaths for assessment, which are cumbersome and time-consuming to obtain, thus limiting their applications in fully-automated situations. Some other methods [29, 36] design and refine the scanpath generator separately from the quality predictor, which is bound to be suboptimal. Moreover, while all methods prove effective with panoramic images, their adaptability for use with panoramic videos remains unclear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.47, + 0.651 + ], + "angle": 0, + "content": "In this work, we further pursue the scanpath-based methods for end-to-end optimized blind PVQA. Our method consists of two modules: a scanpath generator and a quality assessor. Our scanpath generator is probabilistic, which takes historical scanpaths as input and is pre-trained to predict future scanpaths by minimizing their expected code length [16]. The scanpath generator and the quality assessor are then jointly optimized to explain human perceptual scores of panoramic videos. To enable end-to-end optimization, we employ the reparameterization trick [12, 13] to allow differentiable scanpath sampling and adopt subgradients to handle discontinuities of the interpolation kernel [11] for viewpoint sequence generation. Our blind PVQA method not only eliminates the need for human scanpaths, but also supplies a lightweight and differentiable scanpath generator that can work with any planar VQA model. Furthermore, our method is \"backward compatible,\" in the sense that it handles panoramic images with no modification. We test the proposed blind PVQA models on three public panoramic image and video quality datasets [6, 30, 35], covering both synthetic and authentic distortions. Under both in-dataset and cross-dataset settings, our models consistently exhibit better quality prediction performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.663, + 0.22, + 0.679 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.688, + 0.468, + 0.718 + ], + "angle": 0, + "content": "We review two highly relevant topics, scanpath generation and quality assessment of panoramic images and videos." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.277, + 0.743 + ], + "angle": 0, + "content": "2.1. Scanpath Generation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Typical inputs to a panoramic scanpath generator include the saliency map, optical flow map, and historical scanpath. To improve saliency detection, Nguyen et al. [21] compiled a panoramic video saliency dataset, while Xu et al. [38] focused on relative viewpoint displacement prediction. Apart from the historical scanpath, Li et al. [15] incorporated \"future\" scanpaths from other users to facilitate cross-user transfer learning. Through an in-depth root-cause analysis, Rondon et al. [25] discovered that visual features have a minimal impact on the prediction of short-term scanpaths" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "(e.g., \\(\\leq 2\\) seconds). Motivated by their findings, Chao et al. [3] trained a Transformer [33] to predict future scanpaths based solely on historical scanpaths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.258 + ], + "angle": 0, + "content": "The above-mentioned methods [3, 15, 21, 25, 38] treat scanpath generation as a deterministic prediction task, neglecting the inherent scanpath diversity and uncertainty. As a departure, Li et al. [16] formulated scanpath generation as a density estimation problem, which can be implemented by expected code length minimization. In our work, we adopt Li's approach [16] to learn multi-user viewing patterns and generate human-like scanpaths." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.267, + 0.687, + 0.282 + ], + "angle": 0, + "content": "2.2. Quality Assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.35 + ], + "angle": 0, + "content": "Current PVQA models are primarily derived from planar image and video quality methods, which are applied to three types of data representations: the projected 2D plane, spherical surface, and projected rectilinear viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.621 + ], + "angle": 0, + "content": "Planar domain methods [32, 35, 41] aim to rectify geometric distortions and mitigate uneven sampling that results from the sphere-to-plane projection. These include the latitude-adaptive weighting [32], Craster parabolic projection [41], and pseudocylindrical representation [35]. Spherical domain methods, such as S-PSNR [40] and S-SSIM [4], compute and pool local quality measurements over the sphere. Yang et al. [39] trained a non-local spherical neural network [5, 34] to extract spatiotemporal information from panoramic videos. Viewport domain methods prioritize the extraction of visually informative viewpoints for quality analysis. Li et al. [14] introduced a two-step approach that involves viewpoint proposal and quality assessment. Xu et al. [37] built a graph over the extracted viewpoints, and Fu et al. [7] constructed hypergraphs to represent the semantic interactions between viewpoints. One limitation of current viewpoint proposal methods is that they do not accurately reflect the human viewing experience." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Sui et al. [27] pioneered scanpath-based methods for PVQA, under the category of viewpoint domain methods. To eliminate the dependency on human scanpaths, Sui et al. [29] adopted a deep Markov model [28] to generate scanpaths. Meanwhile, Wu et al. [36] handcrafted a simple scanpath generator based on the entropy feature and equator bias. These methods are tailored for panoramic images and are not end-to-end optimized. In contrast, we aim ambitiously for an end-to-end optimized quality assessment method for panoramic videos, with the added benefit of being backward compatible with panoramic images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.673, + 0.817 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our method consists of two modules: a scanpath generator and a quality assessor. Given a panoramic video, we first specify a starting point \\((\\phi_0,\\theta_0)\\), a viewing duration \\(S\\), and \\(N\\) initial paths. The scanpath generator autoregressively samples \\(N\\) scanpaths based on the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.092, + 0.887, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.229, + 0.318, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.23, + 0.888, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.362, + 0.892, + 0.447 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed blind PVQA method, consisting of a scanpath generator and a quality assessor. The basic component of the scanpath generator is the scanpath generation unit (SGU), which utilizes the historical and causal relative scanpaths to produce the GMM parameters for differentiable sampling of the current viewpoint. By assembling \\( W \\) SGUs, we create a scanpath generation block (SGB), which autoregressively predicts a future scanpath of \\( W \\) viewpoints. We further stack \\( M \\) SGBs to generate a long-term scanpath of \\( M \\times W + H \\) viewpoints, where \\( H \\) is the length of the initial path. By adjusting the number of initial paths (denoted by \\( N \\)), we can sample \\( N \\) scanpaths, along which we produce \\( N \\) viewpoint sequences as input to the quality assessor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.469, + 0.565 + ], + "angle": 0, + "content": "initial and already generated path segments. Along these scanpaths, we apply a differentiable viewpoint generation technique to extract \\( N \\) viewpoint sequences from the input panoramic video. Each viewpoint sequence (as a planar video) is fed to the quality assessor, whose predicted score is subsequently aggregated into an overall quality estimate of the panoramic video." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.271, + 0.59 + ], + "angle": 0, + "content": "3.1. Scanpath Generator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.469, + 0.659 + ], + "angle": 0, + "content": "Probabilistic Scanpath Modeling. To capture the uncertainty and diversity of human scanpaths, we formulate panoramic scanpath generation as a density estimation problem:" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.672, + 0.469, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\max p (\\boldsymbol {r} | \\boldsymbol {s}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.7, + 0.468, + 0.79 + ], + "angle": 0, + "content": "where \\(s = \\{(\\phi_0,\\theta_0),\\dots ,(\\phi_t,\\theta_t),\\dots ,(\\phi_{T - 1},\\theta_{T - 1})\\}\\) is the historical scanpath as the condition and \\(\\pmb {r} = \\{(\\phi_T,\\theta_T),\\ldots ,(\\phi_{T + W - 1},\\theta_{T + W - 1})\\}\\) is the future scanpath to be predicted. Herein, \\(W\\) is the prediction horizon, and \\((\\phi_t,\\theta_t)\\) is the \\(t\\) -th viewpoint in the Euler coordinate system. Mathematically, \\(p(\\boldsymbol {r}|\\boldsymbol {s})\\) can be decomposed as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.803, + 0.469, + 0.843 + ], + "angle": 0, + "content": "\\[\np (\\boldsymbol {r} \\mid \\boldsymbol {s}) = \\prod_ {t = 0} ^ {W - 1} p \\left(\\phi_ {T + t}, \\theta_ {T + t} \\mid \\boldsymbol {s}, \\boldsymbol {c} _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( \\pmb{c}_t = \\{(\\phi_T, \\theta_T), \\dots, (\\phi_{T + t - 1}, \\theta_{T + t - 1})\\} \\) is referred to as the causal path that includes all estimated viewpoints before \\( (\\phi_{T + t}, \\theta_{T + t}) \\), and \\( c_0 = \\emptyset \\). The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.572 + ], + "angle": 0, + "content": "chain rule suggests estimating the conditional probability \\( p\\left(\\phi_{T + t}, \\theta_{T + t} \\mid s, c_t\\right) \\) autoregressively. We further make the Markovian assumption: prediction of the current viewpoint is conditionally independent of viewpoints that are temporally further distant, given the most recent \\( H \\) viewpoints. This leads to a truncated historical path context \\( s_T = \\{(\\phi_{T - H}, \\theta_{T - H}), \\ldots, (\\phi_{T - 1}, \\theta_{T - 1})\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.624 + ], + "angle": 0, + "content": "We parameterize the probability \\( p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right) \\), where \\( \\boldsymbol{r}_t = (\\phi_{T + t}, \\theta_{T + t}) \\), by a Gaussian mixture model (GMM) with \\( K \\) components:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {r} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathcal {N} _ {i} \\left(\\boldsymbol {r} _ {t}; \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) is the \\(i\\)-th mixture weight, \\(\\pmb{\\mu}_{i}\\) and \\(\\pmb{\\Sigma}_{i}\\) represent the mean vector and the covariance matrix of the \\(i\\)-th Gaussian component, respectively. This parametrization can be straightforwardly done by training a density estimation network for parameter estimation. As illustrated in Figure 2, this network is inside the scanpath generation unit (SGU) and is composed of two subnetworks to process the historical path context \\(s_{T}\\) and the causal path context \\(c_{t}\\), which we denote by H-Net and C-Net, respectively. The concatenated features are fed to three prediction heads to estimate the weight vector, the mean vectors, and the covariance matrices of the GMM, respectively. We find empirically that incorporating the historical video frames as the visual context significantly increases computational demands with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2601" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "only slight improvements in performance. Therefore, to keep the scanpath generator lightweight, we choose to omit the visual context. The detailed specifications of the density estimation network can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.29 + ], + "angle": 0, + "content": "Estimating continuous probability density is generally difficult and may lead to overfitting. In particular, maximum likelihood estimation of the GMM parameters through direct optimization of Eq. (3) is challenging, due to the presence of singularities [2]. To circumvent this, we compute the probability mass \\( P\\left(\\bar{\\boldsymbol{r}}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right) \\) by discretizing and integrating the density \\( p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.296, + 0.468, + 0.336 + ], + "angle": 0, + "content": "\\[\nP \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\int_ {\\Omega} p \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) d \\Omega . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.342, + 0.468, + 0.372 + ], + "angle": 0, + "content": "\\(\\bar{\\boldsymbol{r}}_t\\) represents the quantized value of \\(\\boldsymbol{r}_t\\) by a uniform quantizer with a step size of \\(\\Delta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.378, + 0.468, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\bar {\\xi} = Q (\\xi) = \\Delta \\left\\lfloor \\frac {\\xi}{\\Delta} + \\frac {1}{2} \\right\\rfloor , \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.417, + 0.468, + 0.493 + ], + "angle": 0, + "content": "where \\(\\lfloor \\cdot \\rfloor\\) denotes the floor function. \\(\\Omega = [\\bar{\\phi}_{T + t} - 1 / 2\\Delta, \\bar{\\phi}_{T + t} + 1 / 2\\Delta] \\times [\\bar{\\theta}_{T + t} - 1 / 2\\Delta, \\bar{\\theta}_{T + t} + 1 / 2\\Delta]\\) is the integration interval. As pointed out in [16], the incorporation of quantization establishes the equivalence between scanpath generation and lossy scanpath compression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.468, + 0.598 + ], + "angle": 0, + "content": "Furthermore, the absolute Euler coordinate system is not user-centric, meaning that it is not centered at the user's current viewpoint, relative to historical and future viewpoints [16]. This may complicate the probabilistic modeling of scanpaths and the end-to-end optimization of blind PVQA. To address this, we convert the Euler coordinates to the relative \\( uv \\) coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.607, + 0.468, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {s}} _ {T - t} = \\boldsymbol {\\Psi} _ {T - t} (\\boldsymbol {s} _ {T}), \\text {f o r} t \\in \\{1, \\dots , H \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.468, + 0.854 + ], + "angle": 0, + "content": "where \\(\\Psi_{T - t}(\\cdot)\\) denotes the mapping of \\(s_T\\) to the viewpoint centered at the reference viewpoint \\((\\phi_{T - t},\\theta_{T - t})\\). By choosing each viewpoint in \\(s_T\\) as the reference, we create \\(H\\) relative scanpaths out of \\(s_T\\) (see Figure 3), which serve as input to the H-Net. Meanwhile, we map the causal scanpath context \\(c_{t}\\) and the viewpoint to be predicted \\((\\phi_{T + t},\\theta_{T + t})\\) to the last historical viewpoint centered at \\((\\phi_{T - 1},\\theta_{T - 1})\\). We stack \\(W\\) SGUs to form a scanpath generation block (SGB), which takes \\(s_T\\) as input and predicts the future scanpath \\(r\\). Furthermore, we stack \\(M\\) SGBs to form the scanpath generator, which is capable of predicting a very long-term scanpath of length \\(M\\times W + H\\) (including the initial length \\(H\\)). The parameters of different SGUs are shared to enable variable-length scanpath generation by varying \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Differentiable Scanpath Sampling. To enable end-to-end optimization of the proposed blind PVQA method, we propose a two-step differentiable sampling method to draw" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.085, + 0.852, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.249 + ], + "angle": 0, + "content": "Figure 3. Visualization of a relative scanpath projected from the sphere to the viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.261, + 0.892, + 0.321 + ], + "angle": 0, + "content": "viewpoints from the estimated GMM via the reparameterization trick [12, 13]. The first step is to select a Gaussian component from which to sample the viewpoint, according to the categorical distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.33, + 0.892, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {e} = \\text {o n e} _ {\\text {h o t}} \\left(\\underset {i \\in \\{1, \\dots K \\}} {\\arg \\max } \\left(\\log \\left(\\alpha_ {i}\\right) + g _ {i}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.381, + 0.891, + 0.442 + ], + "angle": 0, + "content": "where \\( g_{i} \\) is a sample drawn from the Gumbel(0,1) distribution, and \\( e \\) is a one-hot vector. Eq. (7) is known as the Gumbel-Max trick [8], which is non-differentiable. We relax the arg max operator with a softmax function [12]:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.452, + 0.891, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {e}} = \\operatorname {s o f t m a x} \\left(\\left(\\log (\\boldsymbol {\\alpha}) + \\boldsymbol {g}\\right) / \\tau\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.479, + 0.892, + 0.6 + ], + "angle": 0, + "content": "where \\(\\tau\\) represents the temperature coefficient, \\(\\alpha = [\\alpha_{1},\\dots,\\alpha_{K}]\\), and \\(\\pmb {g} = [g_1,\\dots,g_K]\\). As \\(\\tau\\) approaches zero, \\(\\hat{e}\\) converges to \\(e\\). In the forward pass, \\(\\arg \\max\\) is used directly, while in the backward pass, it is replaced by the softmax function. The second step involves sampling a viewpoint from the selected Gaussian component. Assuming the \\(i\\)-th Gaussian component is selected, the linear reparameterization trick [13] suggests" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.611, + 0.891, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {L} _ {i} \\epsilon , \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.637, + 0.892, + 0.743 + ], + "angle": 0, + "content": "where \\(\\tilde{\\boldsymbol{r}}_t\\) is the relative \\(uv\\) coordinates of the \\(t\\)-th future viewpoint. \\(\\boldsymbol{\\Sigma}_i = \\boldsymbol{L}_i\\boldsymbol{L}_i^{\\mathrm{T}}\\) is the Cholesky decomposition, and \\(\\epsilon\\) is a sample drawn from the \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). Differentiation of the Cholesky decomposition is complicated and sometimes numerically unstable [26]. Thus we assume the independence between the \\(uv\\) coordinates, leading to the simplified reparameterization formula:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.755, + 0.891, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes the element-wise product, and \\(\\sigma_{i}\\) denotes the standard deviations of the \\(i\\)-th Gaussian component. Through the two-step reparameterization, our sampling strategy ensures effective back-propagation. As suggested in [16], we additionally implement a proportional-integral-derivative (PID) controller [1] to further improve the smoothness of the sampled scanpaths (see the details in the supplementary material)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Differentiable Viewport Sequence Generation. Inspired by [11], we first parameterize the Euler sampling grid in terms of the relative \\( uv \\) coordinates via the inverse transformation \\( \\Psi^{-1} \\) (see Eq. (6)). Subsequently, the Euler coordinate \\( (\\phi, \\theta) \\) is mapped to the discrete sampling position \\( (m, n) \\) in the ERP domain:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.192, + 0.47, + 0.209 + ], + "angle": 0, + "content": "\\[\nm = (0. 5 - \\phi / \\pi) H _ {e} - 0. 5, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.211, + 0.469, + 0.227 + ], + "angle": 0, + "content": "\\[\nn = (\\theta / 2 \\pi + 0. 5) W _ {e} - 0. 5, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.471, + 0.375 + ], + "angle": 0, + "content": "where \\( H_{e} \\) and \\( W_{e} \\) are the height and width of the video frame in ERP. Once the mapping between \\( (u, v) \\) and \\( (m, n) \\) is established, we construct a flow field that is the same size as the viewpoint. Within this field, each element records the corresponding pixel position \\( (m, n) \\). Given an ERP image and the flow field, we apply bilinear interpolation [11] to compute pixel values in the viewpoint and leverage its sub-gradients for back-propagation. This process yields \\( N \\) viewpoint sequences, corresponding to \\( N \\) initial paths." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.381, + 0.245, + 0.398 + ], + "angle": 0, + "content": "3.2. Quality Assessor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.473, + 0.661 + ], + "angle": 0, + "content": "Our probabilistic scanpath generator can work with any planar VQA model, whether it is differentiable or not. To enable end-to-end optimization of the scanpath generator and the quality assessor, and to make a fair comparison with existing blind PVQA models, we reuse three differentiable quality assessors from ScanpathVQA [35], GSR-S / GSR-X [29], and Assessor360 [36]. Specifically, the quality assessor of ScanpathVQA is a lightweight ResNet-18 network [9], with the classification head replaced by a quality estimator (a multilayer perceptron). The quality assessors of GSR-S / GSR-X are adapted from Video Swin-T [18] / X-Clip-B/32 [22]. The quality assessor of Assessor360 is modified from Swin-B [17] with the addition of a temporal analysis module. We feed each of the \\(N\\) viewpoint sequences to the quality assessor to compute \\(N\\) quality scores. The overall quality estimate is then computed by a simple average:" + }, + { + "type": "equation", + "bbox": [ + 0.222, + 0.67, + 0.47, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\hat {q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\hat {q} _ {i}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.287, + 0.737 + ], + "angle": 0, + "content": "3.3. Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.471, + 0.82 + ], + "angle": 0, + "content": "We explore a three-stage training procedure for our blind PVQA model. In the first stage, we pre-train the density estimation network on the VRVQW dataset [35] by minimizing the expected code length of the generated scanpaths (also equivalent to minimizing the negative log-likelihood):" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.84, + 0.47, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {c o d e}} = - \\frac {1}{B W} \\sum_ {i = 1} ^ {B} \\sum_ {t = 0} ^ {W - 1} \\log_ {2} \\left(P \\left(\\bar {\\boldsymbol {r}} _ {t} ^ {(i)} \\mid \\boldsymbol {s} ^ {(i)}, \\boldsymbol {c} _ {t} ^ {(i)}\\right)\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.259 + ], + "angle": 0, + "content": "where \\( B \\) denotes the mini-batch size. During this stage of training, we use human scanpaths to fill in the causal path context, which can be efficiently implemented by a causal masking mechanism. In the second stage, we fix the parameters of the pre-trained scanpath generator and warm up the quality assessor by optimizing the Pearson linear correlation coefficient (PLCC) between human perceptual scores and model predictions. In the third stage, we end-to-end finetune the entire method. We find that the proposed three-stage optimization strategy accelerates convergence compared to the naive end-to-end optimization." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.27, + 0.634, + 0.288 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.403 + ], + "angle": 0, + "content": "In this section, we first delineate the experimental setups, and then compare our method with current blind PVQA models under both in-dataset and cross-dataset settings. We further validate our scanpath generator in terms of explaining human perceptual scores and replicating human viewing patterns. Lastly, we conduct a series of ablation experiments to probe the impact of several key designs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.41, + 0.698, + 0.427 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Datasets. We employ three panoramic image and video datasets: VRVQW [35], CVIQD [30], and OIQA [6]. The VRVQW dataset includes 502 panoramic videos that have a wide spectrum of authentic distortions. Each video is viewed under four unique viewing conditions to simulate the different quality of experience during the initial viewing. The CVIQD dataset comprises a total of 528 compressed panoramic images by JPEG, AVC, and HEVC, from 16 reference images. The OIQA dataset includes 320 panoramic images that have been altered from 16 reference images by JPEG compression, JPEG2000 compression, Gaussian blurring, and Gaussian noise contamination." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. For the scanpath generator, we set the length of the provided initial path \\( H \\) and the predicted future path \\( W \\) in the SGB to be identical and equal to 5. The number of Gaussian components \\( K \\) in Eq. (3) is set to 3. The quantization step size \\( \\Delta \\) in Eq. (5) is set to 0.2. The temperature coefficient \\( \\tau \\) in Eq. (8) is set to 1. The number of stacked SGBs \\( M \\) is set to 6 and 14 for the viewing duration of 7 and 15 seconds, respectively. For the quality assessor, the number of scanpaths \\( N \\) in Eq. (13) is set to 20. The input viewpoint size \\( H_{v} \\times W_{v} \\) is \\( 224 \\times 224 \\), corresponding to a field of view of \\( 90^{\\circ} \\times 90^{\\circ} \\). The length of the viewpoint sequence is set to 7 regardless of the duration and frame rate of the original panoramic video. We split VRVQW randomly into the training, validation, and test sets according to the ratio of \\( 6:2:2 \\) for 5 times, and report the mean results. Similarly, we split CVIQD and OIQA using a different ratio of \\( 7:1:2 \\) for 5 times. The detailed configuration of our three-stage optimization strategy is detailed in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2603" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.175 + ], + "angle": 0, + "content": "Table 1. In-dataset comparison of blind PVQA methods on three panoramic image and video quality datasets. SRCC: Spearman's rank correlation coefficient. PLCC: Pearson linear correlation coefficient. The same evaluation metrics are applied in Tables 2, 3, 5, 6 and Figure 6. The best results on each dataset are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.184, + 0.446, + 0.59 + ], + "angle": 0, + "content": "
DatasetMethodSRCCPLCC
VRVQW [35]NIQE [19]0.4010.365
MC360IQA [31]0.6690.671
Wen24 [35]0.7560.763
ScanpathVQA [35]0.7790.781
Assessor360 [36]0.4060.415
Ours (ScanpathVQA)0.8010.809
Ours (GSR-S)0.8040.807
Ours (GSR-X)0.8150.819
Ours (Assessor360)0.8220.823
CVIQD [30]NIQE [19]0.8470.878
MC360IQA [31]0.9170.939
Wen24 [35]0.9190.932
GSR-S [29]0.9050.937
GSR-X [29]0.9440.962
Assessor360 [36]0.9550.969
Ours (ScanpathVQA)0.9120.936
Ours (GSR-S)0.9300.958
Ours (GSR-X)0.9560.974
Ours (Assessor360)0.9720.983
OIQA [6]NIQE [19]0.7020.657
MC360IQA [31]0.9000.906
Wen24 [35]0.9050.907
GSR-S [29]0.9020.915
GSR-X [29]0.9450.954
Assessor360 [36]0.9460.953
Ours (ScanpathVQA)0.9150.922
Ours (GSR-S)0.9270.936
Ours (GSR-X)0.9560.967
Ours (Assessor360)0.9600.971
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.601, + 0.472, + 0.645 + ], + "angle": 0, + "content": "Table 2. Cross-dataset comparison of blind PVQA methods on the CVIQD [30] and OIQA [6] datasets. The arrow points from the training set to the test set." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.647, + 0.428, + 0.814 + ], + "angle": 0, + "content": "
DatasetMethodSRCCPLCC
OIQAMC360IQA0.7980.842
GSR-X0.7620.841
Assessor3600.8590.893
CVIQDOurs (ScanpathVQA)0.7330.747
Ours (Assessor360)0.8720.904
CVIQDMC360IQA0.2880.349
GSR-X0.6950.718
Assessor3600.3380.467
OIQAOurs (ScanpathVQA)0.6360.658
Ours (Assessor360)0.7030.715
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.218, + 0.845 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We compare our blind PVQA method with seven existing models, including NIQE [19], MC360IQA [31], Wen24 [35], ScanpathVQA [35], GSR-S [29], GSR-X [29]," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 3. Comparison of different scanpath generators for explaining human perceptual scores." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.122, + 0.859, + 0.366 + ], + "angle": 0, + "content": "
DatasetMethodSRCCPLCC
VRVQWHuman scanpath0.7860.790
Random sampling0.0750.104
Heuristic sampling [36]0.4310.443
Xu18 [38]0.7120.717
TRACK [25]0.7450.749
Li23 [16]0.7900.794
Ours0.8050.814
CVIQDRandom sampling0.6320.640
Heuristic sampling [36]0.8680.870
ScanDMM [28]0.8560.864
Li23 [16]0.8140.827
Ours0.9280.940
OIQARandom sampling0.5140.536
Heuristic sampling [36]0.8610.872
ScanDMM [28]0.8650.877
Li23 [16]0.7930.799
Ours0.9140.917
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.485 + ], + "angle": 0, + "content": "and Assessor360 [36]. For image quality models such as MC360IQA [31] and Assessor360 [36], we retrain them on the VRVQW dataset. In the case of MC360IQA, we assign the video-level quality score to each key frame and use their temporally averaged score for testing. Assessor360's scanpath generator is adapted to videos by adjusting the semantic context associated with each key frame." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.832 + ], + "angle": 0, + "content": "In-dataset Results. Table 1 shows the Spearman's rank correlation coefficient (SRCC) and \\(\\mathrm{PLCC}^1\\) results under the in-dataset setting. It is evident that our learned scanpaths enhance the performance of all quality assessors compared to other scanpath-based methods. When integrated with the quality assessor from Assessor360 (i.e., a modified SwinB with a temporal analysis module), our proposed method achieves the best results on all three datasets. Furthermore, our scanpath generator can boost a simpler quality assessor (e.g., from ScanpathVQA [35] with approximately 11 million parameters) to reach performance levels similar to those of a more sophisticated quality assessor coupled with a weaker scanpath generator (e.g., GSR-S [29] with approximately 112 million parameters). Methods that overlook human viewing patterns, like NIQE [19] and MC360IQA [31], fail to accurately model the human perception of panoramic image and video quality, especially on VRVQW. Additionally, we find that assessing the quality of panoramic videos with authentic distortions tends to be more challenging than for panoramic images with synthetic distortions. This is anticipated because authentic distortions in panoramic videos typically present as a complex blend of various artifacts, localized in space and time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.833, + 0.894, + 0.864 + ], + "angle": 0, + "content": "Cross-dataset Results. Table 2 shows the SRCC and PLCC results under the cross-dataset settings on CVIQD [30] and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1As standard practice, we apply a monotonic logistic function to compensate for the nonlinearity in model predictions before computing PLCC." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2604" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.473, + 0.132 + ], + "angle": 0, + "content": "Table 4. Comparison of different scanpath generators for replicating human viewing patterns using the minimum orthodromic distance (minOD) and maximum temporal correlation (maxTC)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.135, + 0.455, + 0.243 + ], + "angle": 0, + "content": "
MethodminOD ↓maxTC ↑
Heuristic sampling [36]1.3250.401
Xu18 [38]1.1850.637
TRACK [25]1.0670.699
Li23 [16]0.5420.796
Ours (w/o end-to-end optimization)0.5560.781
Ours0.5360.805
" + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.259, + 0.412, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.428, + 0.473, + 0.456 + ], + "angle": 0, + "content": "Figure 4. Comparison of different scanpath predictors in terms of maxTC with different prediction horizons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.609 + ], + "angle": 0, + "content": "OIQA [6]. Generally, models trained on the OIQA dataset show better generalizability. This is likely because OIQA encompasses a broader range of distortion types, compared to CVIQD, which only includes the compression artifacts. Assessor360 shows a noticeable performance drop when tested on OIQA, potentially indicative of overfitting. Our scanpath generator is capable of restoring Assessor360's performance, which provides a strong indication of its effectiveness through end-to-end optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.355, + 0.635 + ], + "angle": 0, + "content": "4.3. Scanpath Generator Validation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Explaining Human Perceptual Scores. We conduct an apple-to/apple comparison of different scanpath generators in terms of explaining human perceptual scores by fixing the quality assessor to that used in ScanpathVQA. These include random sampling, heuristic sampling [36], Xu18 [38], TRACK [25], ScanDMM [28], Li23 [16], and our method. Table 3 shows the SRCC and PLCC results on the VRVQW, CVIQD and OIQA datasets. It is noteworthy that our scanpath generator outperforms all competing methods across all three datasets, even surpassing the human-level performance on VRVQW. The performance of random and heuristic sampling decreases sharply on VRVQW, due to the presence of spatiotemporally localized authentic distortions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Replicating Human Viewing Patterns. We also test different scanpath generators in terms of replicating human viewing patterns on the VRVQW dataset by comparing the predicted scanpaths to those of humans. We use two set-to-set" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.888, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.225, + 0.892, + 0.254 + ], + "angle": 0, + "content": "Figure 5. Comparison of saliency maps generated from scanpaths by our method and those by humans." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.269, + 0.894, + 0.311 + ], + "angle": 0, + "content": "Table 5. Impact of optimization strategies on blind PVQA. Training protocol: 1) Two-stage w/ fixed scanpaths, 2) Two-stage w/ varied scanpaths, and 3) Three-stage w/ end-to-end optimization." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.315, + 0.903, + 0.391 + ], + "angle": 0, + "content": "
ProtocolVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
10.7690.7720.7100.7800.6880.742
20.7810.7850.8300.8590.7980.815
30.8050.8140.9280.9400.9140.917
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.405, + 0.892, + 0.503 + ], + "angle": 0, + "content": "evaluation metrics: the minimum orthodromic distance (i.e., minOD) and maximum temporal correlation (i.e., maxTC), as suggested in [16]. Given a set of human scanpaths, \\(S = \\{s^{(i)}\\}_{i=1}^{|S|}\\), the minimum orthodromic distance between \\(S\\) and the set of predicted scanpaths \\(\\hat{S} = \\{\\hat{s}^{(i)}\\}_{i=1}^{|S|}\\) can be computed by" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.513, + 0.892, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\min \\mathrm {O D} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\min _ {\\boldsymbol {s} \\in \\mathcal {S}, \\boldsymbol {s} \\in \\hat {\\mathcal {S}}} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.555, + 0.855, + 0.571 + ], + "angle": 0, + "content": "where the orthodromic distance \\(\\mathrm{OD}(\\cdot ,\\cdot)\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.583, + 0.892, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\operatorname {a r c c o s} \\left(\\cos \\left(\\phi_ {t}\\right) \\cos \\left(\\hat {\\phi} _ {t}\\right) \\cos \\left(\\theta_ {t} - \\hat {\\theta} _ {t}\\right) \\right. \\\\ + \\sin \\left(\\phi_ {t}\\right) \\sin \\left(\\hat {\\phi} _ {t}\\right)). \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.681, + 0.892, + 0.711 + ], + "angle": 0, + "content": "The maximum temporal correlation between \\( \\mathcal{S} \\) and \\( \\hat{\\mathcal{S}} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.721, + 0.892, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\max \\mathrm {T C} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\max _ {\\boldsymbol {s} \\in \\mathcal {S}, \\hat {\\boldsymbol {s}} \\in \\hat {\\mathcal {S}}} \\mathrm {T C} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.763, + 0.81, + 0.779 + ], + "angle": 0, + "content": "where the temporal correlation is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.788, + 0.892, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {T C} \\left(\\boldsymbol {s} ^ {(i)}, \\boldsymbol {s} ^ {(j)}\\right) = \\\\ \\frac {1}{2} \\left(\\operatorname {P L C C} \\left(\\phi^ {(i)}, \\phi^ {(j)}\\right) + \\operatorname {P L C C} \\left(\\boldsymbol {\\theta} ^ {(i)}, \\boldsymbol {\\theta} ^ {(j)}\\right)\\right). \\end{array} \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Table 4 presents the minOD and maxTC results, from which we find that our end-to-end optimized scanpath generator delivers the best results, surpassing its independently" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2605" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.156, + 0.089, + 0.813, + 0.104 + ], + "angle": 0, + "content": "Table 6. Impact of visual context on blind PVQA. #Parameters added to the scanpath generator are also shown." + }, + { + "type": "table", + "bbox": [ + 0.204, + 0.108, + 0.764, + 0.172 + ], + "angle": 0, + "content": "
Method#ParametersVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
Ours1M0.8050.8140.9280.9400.9140.917
Ours w/ visual context27M0.8160.8230.9370.9560.9250.934
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.19, + 0.471, + 0.417 + ], + "angle": 0, + "content": "optimized counterpart by a clear margin. The heuristic sampling [36] that depends on the simplified entropy features and equator bias, is inadequate for capturing human viewing patterns, especially for long-term prediction horizons (see Figure 4). Due to the deterministic nature, Xu18 [38] and TRACK [25] fail to accommodate the diversity and uncertainty inherent in human scanpaths, resulting in subpar performance. Incorporating historical video frames as the visual context, Li23 [16] shows performance on par with our method, reinforcing our assertion that visual context informs less about future viewpoints. Figure 5 demonstrates a comparison of the saliency maps derived from scanpaths by our method and those by humans, offering further proof of the close alignment of our scanpath generator and human viewing behaviors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.242, + 0.44 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.643 + ], + "angle": 0, + "content": "Impact of Optimization Strategies. We explore three different optimization strategies: 1) a two-stage approach where the pre-trained scanpath generator produces a fixed set of scanpaths for the training of the quality assessor, 2) a similar two-stage approach but supplying a varied set of scanpaths in each epoch of training, and 3) the default three-stage approach that enables end-to-end optimization. From the results in Table 5, we find that the two-stage approach benefits from \"data augmentation\" with varied scanpaths in each epoch. Our three-stage end-to-end optimization strategy further boosts the accuracy of quality prediction by jointly finetuning both the scanpath generator and quality assessor." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Impact of Visual Context. To assess the impact of visual context on blind PVQA, we enhance our scanpath generator with a video analysis network [16], implemented by a variant of ResNet-50 for frame-level feature extraction and aggregation. We subsequently integrate it with the ScanpathVQA quality assessor for blind PVQA, with the results shown in Table 6. We find that the visual context has a negligible effect on blind PVQA, and thus we exclude it in the generation of scanpaths." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Impact of the Number and Length of Viewport Sequences. We investigate the effects of varying the number \\(N\\) and length \\(L\\) of viewpoint sequences on blind PVQA. We test \\(N\\) values from \\(\\{5,10,15,20,50\\}\\) and \\(L\\) values from \\(\\{4,7,15\\}\\). Figure 6 shows the SRCC results on VRVQW, using the ScanpathVQA quality assessor for prediction. It is clear that \\(N = 20\\) viewpoint sequences are sufficient for reliable quality assessment, with performance remaining stable" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.189, + 0.835, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.392 + ], + "angle": 0, + "content": "Figure 6. Impact of the number \\(N\\) and length \\(L\\) of viewpoint sequences on blind PVQA." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.892, + 0.542 + ], + "angle": 0, + "content": "with the increase in \\( N \\). For sequence length, \\( L = 7 \\) appears to be a wise choice. Further increasing \\( L \\) does not noticeably affect the performance, but would lead to a considerable rise in computational demand. Conversely, a shorter viewpoint sequence results in a noticeable drop in performance due to the loss of information from excessive temporal downsampling. It is important to note that these findings are specific to the ScanpathVQA quality assessor and may differ from other assessors." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.62, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.753 + ], + "angle": 0, + "content": "We have introduced an end-to-end optimized blind PVQA method, consisting of a scanpath generator and a quality assessor. The proposed scanpath generator is differentiable and can be integrated with any planar VQA model, whose effectiveness has been thoroughly validated in supporting blind PVQA and in modeling human viewing patterns. Additionally, we have also devised a three-stage optimization strategy to facilitate training convergence, which aligns with current large-scale optimization practices that involve self-supervised pre-training followed by supervised finetuning, including initial warmup phases [10]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.681, + 0.787 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (62071407 and 62102339), the Shenzhen Science and Technology Program (RCBS20221008093121052), the Research Grants Council of Hong Kong (ECS 2121382, ECS 27212822, and GRF 17201822), and the CCF-Tencent Rhino Bird Fund (9239061)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2606" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Kiam Heong Ang, Gregory Chong, and Yun Li. PID control system analysis, design, and technology. IEEE Transactions on Control Systems Technology, 13(4):559-576, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.471, + 0.185 + ], + "angle": 0, + "content": "[2] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.471, + 0.254 + ], + "angle": 0, + "content": "[3] Fang-Yi Chao, Cagri Ozcinar, and Aljosa Smolic. Transformer-based long-term viewpoint prediction in \\(360^{\\circ}\\) video: Scanpath is all you need. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[4] Sijia Chen, Yingxue Zhang, Yiming Li, Zhenzhong Chen, and Zhou Wang. Spherical structural similarity index for objective omnidirectional video quality assessment. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.368 + ], + "angle": 0, + "content": "[5] Taco S. Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, pages 1-15, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.369, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Yucheng Zhu, Yi Fang, and Xiaokang Yang. Perceptual quality assessment of omnidirectional images. In IEEE International Symposium on Circuits and Systems, pages 1-5, 2018. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[7] Jun Fu, Chen Hou, Wei Zhou, Jiahua Xu, and Zhibo Chen. Adaptive hypergraph convolutional network for no-reference 360-degree image quality assessment. In ACM International Conference on Multimedia, pages 961-969, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.521 + ], + "angle": 0, + "content": "[8] Emil J. Gumbel. Statistical Theory of Extreme Values and Some Practical Applications. US Government Printing Office, 1948. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.577 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.471, + 0.634 + ], + "angle": 0, + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.471, + 0.689 + ], + "angle": 0, + "content": "[11] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017-2025, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.471, + 0.745 + ], + "angle": 0, + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In International Conference on Learning Representations, pages 1–12, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[13] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, pages 1–14, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[14] Chen Li, Mai Xu, Lai Jiang, Shanyi Zhang, and Xiaoming Tao. Viewport proposal CNN for \\(360^{\\circ}\\) video quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10169-10178, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] Chenge Li, Weixi Zhang, Yong Liu, and Yao Wang. Very long term field of view prediction for 360-degree video streaming. In IEEE Conference on Multimedia Information Processing and Retrieval, pages 297-302, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[16] Mu Li, Kanglong Fan, and Kede Ma. Scanpath prediction in panoramic videos via expected code length minimization. arXiv preprint arXiv:2305.02536, 2023. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[18] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[19] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[20] King-To Ng, Shing-Chow Chan, and Heung-Yeung Shum. Data compression and transmission aspects of panoramic videos. IEEE Transactions on Circuits and Systems for Video Technology, 15(1):82-95, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[21] Anh Nguyen, Zhisheng Yan, and Klara Nahrstedt. Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In ACM International Conference on Multimedia, pages 1190-1198, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[22] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[23] David Noton and Lawrence Stark. Scanpaths in saccadic eye movements while viewing and recognizing patterns. Vision Research, 11(9):929-942, 1971. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[24] David Noton and Lawrence Stark. Scanpaths in eye movements during pattern perception. Science, 171(3968):308-311, 1971. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[25] Miguel Fabian Romero Rondon, Lucile Sassatelli, Ramón Aparicio-Pardo, and Frédéric Precioso. TRACK: A new method from a re-examination of deep architectures for head motion prediction in \\(360^{\\circ}\\) videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5681-5699, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[26] Stephen P. Smith. Differentiation of the Cholesky algorithm. Journal of Computational and Graphical Statistics, 4(2):134-147, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[27] Xiangjie Sui, Kede Ma, Yiru Yao, and Yuming Fang. Perceptual quality assessment of omnidirectional images as moving camera videos. IEEE Transactions on Visualization and Computer Graphics, 28(8):3022-3034, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Xiangjie Sui, Yuming Fang, Hanwei Zhu, Shiqi Wang, and Zhou Wang. ScanDMM: A deep markov model of scanpath prediction for \\(360^{\\circ}\\) images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6989-6999, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Xiangjie Sui, Hanwei Zhu, Xuelin Liu, Yuming Fang, Shiqi Wang, and Zhou Wang. Perceptual quality assessment of \\(360^{\\circ}\\) images based on generative scanpath representation. arXiv preprint arXiv:2309.03472, 2023. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2607" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[30] Wei Sun, Ke Gu, Siwei Ma, Wenhan Zhu, Ning Liu, and Guangtao Zhai. A large-scale compressed 360-degree spherical image database: From subjective quality evaluation to objective model comparison. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2018. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[31] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Huiyu Duan, and Siwei Ma. MC360IQA: A multi-channel CNN for blind 360-degree image quality assessment. IEEE Journal of Selected Topics in Signal Processing, 14(1):64-77, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[32] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters, 24(9):1408-1412, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000-6010, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.387 + ], + "angle": 0, + "content": "[34] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[35] Wen Wen, Mu Li, Yiru Yao, Xiangjie Sui, Yabin Zhang, Long Lan, Yuming Fang, and Kede Ma. Perceptual quality assessment of virtual reality videos in the wild. IEEE Transactions on Circuits and Systems for Video Technology, to appear, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.471, + 0.53 + ], + "angle": 0, + "content": "[36] Tianhe Wu, Shuwei Shi, Haoming Cai, Mingdeng Cao, Jing Xiao, Yinqiang Zheng, and Yujiu Yang. Assessor360: Multisession network for blind omnidirectional image quality assessment. In Advances in Neural Information Processing Systems, pages 1-22, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.531, + 0.471, + 0.599 + ], + "angle": 0, + "content": "[37] Jiahua Xu, Wei Zhou, and Zhibo Chen. Blind omnidirectional image quality assessment with viewpoint oriented graph convolutional networks. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1724-1737, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.471, + 0.669 + ], + "angle": 0, + "content": "[38] Yanyu Xu, Yanbing Dong, Junru Wu, Zhengzhong Sun, Zhiru Shi, Jingyi Yu, and Shenghua Gao. Gaze prediction in dynamic \\(360^{\\circ}\\) immersive videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5333-5342, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.471, + 0.726 + ], + "angle": 0, + "content": "[39] Jiachen Yang, Tianlin Liu, Bin Jiang, Wen Lu, and Qinggang Meng. Panoramic video quality assessment based on non-local spherical CNN. IEEE Transactions on Multimedia, 23: 797-809, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.728, + 0.471, + 0.783 + ], + "angle": 0, + "content": "[40] Matt Yu, Haricharan Lakshman, and Bernd Girod. A framework to evaluate omnidirectional video coding schemes. In IEEE International Symposium on Mixed and Augmented Reality, pages 31-36, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.471, + 0.839 + ], + "angle": 0, + "content": "[41] Vladyslav Zakharchenko, Kwang Pyo Choi, and Jeong Hoon Park. Quality metric for spherical panoramic video. In SPIE Optics and Photonics for Information Processing X, pages 57-65, 2016. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2608" + } + ] +] \ No newline at end of file diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_origin.pdf b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f54ddda99d0992e3895f24e2ac2ddc2664f3dfee --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/4097298f-874f-411d-9a63-b8f9db525955_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e163929af402d4670709c82929f91bfa037731d7ea4c8166c24cc28bc67085b +size 1027957 diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/full.md b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad679ad8132320eeea313892349600f085eae21b --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/full.md @@ -0,0 +1,325 @@ +# Learned Scanpaths Aid Blind Panoramic Video Quality Assessment + +Kanglong Fan $^{1}$ , Wen Wen $^{1}$ , Mu Li $^{2*}$ , Yifan Peng $^{3}$ , and Kede Ma $^{1}$ $^{1}$ City University of Hong Kong, $^{2}$ Harbin Institute of Technology, Shenzhen + $^{3}$ The University of Hong Kong +{kanglofan2-c, wwen29-c}@my.cityu.edu.hk, limuhit@gmail.com +evanpeng@hku.hk, kede.ma@cityu.edu.hk +https://github.com/kalofan/AutoScanpathQA + +# Abstract + +Panoramic videos have the advantage of providing an immersive and interactive viewing experience. Nevertheless, their spherical nature gives rise to various and uncertain user viewing behaviors, which poses significant challenges for panoramic video quality assessment (PVQA). In this work, we propose an end-to-end optimized, blind PVQA method with explicit modeling of user viewing patterns through visual scanpaths. Our method consists of two modules: a scanpath generator and a quality assessor. The scanpath generator is initially trained to predict future scanpaths by minimizing their expected code length and then jointly optimized with the quality assessor for quality prediction. Our blind PVQA method enables direct quality assessment of panoramic images by treating them as videos composed of identical frames. Experiments on three public panoramic image and video quality datasets, encompassing both synthetic and authentic distortions, validate the superiority of our blind PVQA model over existing methods. + +# 1. Introduction + +The rapid advancement of multimedia technologies has marked the beginning of a new era characterized by the proliferation of panoramic videos [20]. Such type of digital data offers an immersive and interactive viewing experience that is transforming the way we consume multimedia. Therefore, assessing and ensuring the visual quality of panoramic videos is increasingly important, as it shapes the users' viewing experience and the triumph of any product or service based on panoramic videos [40]. Unlike their planar counterparts, panoramic videos provide a $360^{\circ}$ broad view with a spherical data structure, which poses significant computational challenges for panoramic video quality assessment (PVQA). Moreover, the diverse and uncertain user + +![](images/fba2825827a098313c5e55c4139537b6fedece13fe40771d2eabae889e529071.jpg) +Figure 1. Analogy between human subjects and our end-to-end optimized method for panoramic video quality assessment. + +viewing behaviors (in the form of visual scanpaths) induced by the spherical structure further complicate the quality prediction process. Addressing these challenges requires novel PVQA models that take into account both the spherical data structure as well as user viewing patterns. + +In the quality assessment of panoramic images and videos, three approaches are commonly employed: sphereto-plane projection onto a 2D plane, rectilinear projection onto multiple viewpoints, and direct processing using spherical operations. According to the Theorema Egregium by Gauss, all sphere-to-plane map projections [32, 35, 41] are impeded by non-uniform sampling and geometric distortions, which may bias subsequent planar quality prediction. While spherical operators give a better account for the panoramic data structure, they are generally computationally prohibitive and, more importantly, may not faithfully reflect user viewing patterns [4, 39, 40]. To overcome these computational difficulties, several methods seek to sample and process rectilinear viewpoints [7, 14, 27, 29, 36, 37]. Of particular interest are scanpath-based methods, which sample, along visual scanpaths [23, 24], sequences of rectilinear viewpoints at discrete time instances. This sampling process turns panoramic images and videos into moving- + +camera videos, amenable to planar VQA. + +By closely imitating how humans perceive visual distortions in virtual environments (see Figure 1), scanpath-based methods [27, 29, 36] have demonstrated remarkable efficacy in the quality of panoramic images. Nonetheless, some methods [27] rely on human visual scanpaths for assessment, which are cumbersome and time-consuming to obtain, thus limiting their applications in fully-automated situations. Some other methods [29, 36] design and refine the scanpath generator separately from the quality predictor, which is bound to be suboptimal. Moreover, while all methods prove effective with panoramic images, their adaptability for use with panoramic videos remains unclear. + +In this work, we further pursue the scanpath-based methods for end-to-end optimized blind PVQA. Our method consists of two modules: a scanpath generator and a quality assessor. Our scanpath generator is probabilistic, which takes historical scanpaths as input and is pre-trained to predict future scanpaths by minimizing their expected code length [16]. The scanpath generator and the quality assessor are then jointly optimized to explain human perceptual scores of panoramic videos. To enable end-to-end optimization, we employ the reparameterization trick [12, 13] to allow differentiable scanpath sampling and adopt subgradients to handle discontinuities of the interpolation kernel [11] for viewpoint sequence generation. Our blind PVQA method not only eliminates the need for human scanpaths, but also supplies a lightweight and differentiable scanpath generator that can work with any planar VQA model. Furthermore, our method is "backward compatible," in the sense that it handles panoramic images with no modification. We test the proposed blind PVQA models on three public panoramic image and video quality datasets [6, 30, 35], covering both synthetic and authentic distortions. Under both in-dataset and cross-dataset settings, our models consistently exhibit better quality prediction performance. + +# 2. Related Work + +We review two highly relevant topics, scanpath generation and quality assessment of panoramic images and videos. + +# 2.1. Scanpath Generation + +Typical inputs to a panoramic scanpath generator include the saliency map, optical flow map, and historical scanpath. To improve saliency detection, Nguyen et al. [21] compiled a panoramic video saliency dataset, while Xu et al. [38] focused on relative viewpoint displacement prediction. Apart from the historical scanpath, Li et al. [15] incorporated "future" scanpaths from other users to facilitate cross-user transfer learning. Through an in-depth root-cause analysis, Rondon et al. [25] discovered that visual features have a minimal impact on the prediction of short-term scanpaths + +(e.g., $\leq 2$ seconds). Motivated by their findings, Chao et al. [3] trained a Transformer [33] to predict future scanpaths based solely on historical scanpaths. + +The above-mentioned methods [3, 15, 21, 25, 38] treat scanpath generation as a deterministic prediction task, neglecting the inherent scanpath diversity and uncertainty. As a departure, Li et al. [16] formulated scanpath generation as a density estimation problem, which can be implemented by expected code length minimization. In our work, we adopt Li's approach [16] to learn multi-user viewing patterns and generate human-like scanpaths. + +# 2.2. Quality Assessment + +Current PVQA models are primarily derived from planar image and video quality methods, which are applied to three types of data representations: the projected 2D plane, spherical surface, and projected rectilinear viewpoint. + +Planar domain methods [32, 35, 41] aim to rectify geometric distortions and mitigate uneven sampling that results from the sphere-to-plane projection. These include the latitude-adaptive weighting [32], Craster parabolic projection [41], and pseudocylindrical representation [35]. Spherical domain methods, such as S-PSNR [40] and S-SSIM [4], compute and pool local quality measurements over the sphere. Yang et al. [39] trained a non-local spherical neural network [5, 34] to extract spatiotemporal information from panoramic videos. Viewport domain methods prioritize the extraction of visually informative viewpoints for quality analysis. Li et al. [14] introduced a two-step approach that involves viewpoint proposal and quality assessment. Xu et al. [37] built a graph over the extracted viewpoints, and Fu et al. [7] constructed hypergraphs to represent the semantic interactions between viewpoints. One limitation of current viewpoint proposal methods is that they do not accurately reflect the human viewing experience. + +Sui et al. [27] pioneered scanpath-based methods for PVQA, under the category of viewpoint domain methods. To eliminate the dependency on human scanpaths, Sui et al. [29] adopted a deep Markov model [28] to generate scanpaths. Meanwhile, Wu et al. [36] handcrafted a simple scanpath generator based on the entropy feature and equator bias. These methods are tailored for panoramic images and are not end-to-end optimized. In contrast, we aim ambitiously for an end-to-end optimized quality assessment method for panoramic videos, with the added benefit of being backward compatible with panoramic images. + +# 3. Proposed Method + +As illustrated in Figure 2, our method consists of two modules: a scanpath generator and a quality assessor. Given a panoramic video, we first specify a starting point $(\phi_0,\theta_0)$ , a viewing duration $S$ , and $N$ initial paths. The scanpath generator autoregressively samples $N$ scanpaths based on the + +![](images/aaaae09a7792ea19c8d4fd9e2b33dbc66e49f5210b3ee9f93ee2e2e8dde563fa.jpg) + +![](images/8051603bb40278b82cac6cec6539f854f752cc1b70e08b8d374624c7bbd87c3e.jpg) +Figure 2. Overview of the proposed blind PVQA method, consisting of a scanpath generator and a quality assessor. The basic component of the scanpath generator is the scanpath generation unit (SGU), which utilizes the historical and causal relative scanpaths to produce the GMM parameters for differentiable sampling of the current viewpoint. By assembling $W$ SGUs, we create a scanpath generation block (SGB), which autoregressively predicts a future scanpath of $W$ viewpoints. We further stack $M$ SGBs to generate a long-term scanpath of $M \times W + H$ viewpoints, where $H$ is the length of the initial path. By adjusting the number of initial paths (denoted by $N$ ), we can sample $N$ scanpaths, along which we produce $N$ viewpoint sequences as input to the quality assessor. + +![](images/023c04e5c6b15a1c7a6f0723741c334cbd08ac997346fd0ca773e1003c44d22c.jpg) + +initial and already generated path segments. Along these scanpaths, we apply a differentiable viewpoint generation technique to extract $N$ viewpoint sequences from the input panoramic video. Each viewpoint sequence (as a planar video) is fed to the quality assessor, whose predicted score is subsequently aggregated into an overall quality estimate of the panoramic video. + +# 3.1. Scanpath Generator + +Probabilistic Scanpath Modeling. To capture the uncertainty and diversity of human scanpaths, we formulate panoramic scanpath generation as a density estimation problem: + +$$ +\max p (\boldsymbol {r} | \boldsymbol {s}), \tag {1} +$$ + +where $s = \{(\phi_0,\theta_0),\dots ,(\phi_t,\theta_t),\dots ,(\phi_{T - 1},\theta_{T - 1})\}$ is the historical scanpath as the condition and $\pmb {r} = \{(\phi_T,\theta_T),\ldots ,(\phi_{T + W - 1},\theta_{T + W - 1})\}$ is the future scanpath to be predicted. Herein, $W$ is the prediction horizon, and $(\phi_t,\theta_t)$ is the $t$ -th viewpoint in the Euler coordinate system. Mathematically, $p(\boldsymbol {r}|\boldsymbol {s})$ can be decomposed as + +$$ +p (\boldsymbol {r} \mid \boldsymbol {s}) = \prod_ {t = 0} ^ {W - 1} p \left(\phi_ {T + t}, \theta_ {T + t} \mid \boldsymbol {s}, \boldsymbol {c} _ {t}\right), \tag {2} +$$ + +where $\pmb{c}_t = \{(\phi_T, \theta_T), \dots, (\phi_{T + t - 1}, \theta_{T + t - 1})\}$ is referred to as the causal path that includes all estimated viewpoints before $(\phi_{T + t}, \theta_{T + t})$ , and $c_0 = \emptyset$ . The + +chain rule suggests estimating the conditional probability $p\left(\phi_{T + t}, \theta_{T + t} \mid s, c_t\right)$ autoregressively. We further make the Markovian assumption: prediction of the current viewpoint is conditionally independent of viewpoints that are temporally further distant, given the most recent $H$ viewpoints. This leads to a truncated historical path context $s_T = \{(\phi_{T - H}, \theta_{T - H}), \ldots, (\phi_{T - 1}, \theta_{T - 1})\}$ . + +We parameterize the probability $p\left(\boldsymbol{r}_t \mid \boldsymbol{s}_T, \boldsymbol{c}_t\right)$ , where $\boldsymbol{r}_t = (\phi_{T + t}, \theta_{T + t})$ , by a Gaussian mixture model (GMM) with $K$ components: + +$$ +p \left(\boldsymbol {r} _ {t} \mid \boldsymbol {s} _ {T}, \boldsymbol {c} _ {t}\right) = \sum_ {i = 1} ^ {K} \alpha_ {i} \mathcal {N} _ {i} \left(\boldsymbol {r} _ {t}; \boldsymbol {\mu} _ {i}, \boldsymbol {\Sigma} _ {i}\right), \tag {3} +$$ + +where $\alpha_{i}$ is the $i$ -th mixture weight, $\pmb{\mu}_{i}$ and $\pmb{\Sigma}_{i}$ represent the mean vector and the covariance matrix of the $i$ -th Gaussian component, respectively. This parametrization can be straightforwardly done by training a density estimation network for parameter estimation. As illustrated in Figure 2, this network is inside the scanpath generation unit (SGU) and is composed of two subnetworks to process the historical path context $s_{T}$ and the causal path context $c_{t}$ , which we denote by H-Net and C-Net, respectively. The concatenated features are fed to three prediction heads to estimate the weight vector, the mean vectors, and the covariance matrices of the GMM, respectively. We find empirically that incorporating the historical video frames as the visual context significantly increases computational demands with + +only slight improvements in performance. Therefore, to keep the scanpath generator lightweight, we choose to omit the visual context. The detailed specifications of the density estimation network can be found in the supplementary material. + +Estimating continuous probability density is generally difficult and may lead to overfitting. In particular, maximum likelihood estimation of the GMM parameters through direct optimization of Eq. (3) is challenging, due to the presence of singularities [2]. To circumvent this, we compute the probability mass $P\left(\bar{\boldsymbol{r}}_t \mid \boldsymbol{s}_T, \boldsymbol{c}_t\right)$ by discretizing and integrating the density $p\left(\boldsymbol{r}_t \mid \boldsymbol{s}_T, \boldsymbol{c}_t\right)$ : + +$$ +P \left(\bar {\boldsymbol {r}} _ {t} \mid \boldsymbol {s} _ {T}, \boldsymbol {c} _ {t}\right) = \int_ {\Omega} p \left(\bar {\boldsymbol {r}} _ {t} \mid \boldsymbol {s} _ {T}, \boldsymbol {c} _ {t}\right) d \Omega . \tag {4} +$$ + +$\bar{\boldsymbol{r}}_t$ represents the quantized value of $\boldsymbol{r}_t$ by a uniform quantizer with a step size of $\Delta$ : + +$$ +\bar {\xi} = Q (\xi) = \Delta \left\lfloor \frac {\xi}{\Delta} + \frac {1}{2} \right\rfloor , \tag {5} +$$ + +where $\lfloor \cdot \rfloor$ denotes the floor function. $\Omega = [\bar{\phi}_{T + t} - 1 / 2\Delta, \bar{\phi}_{T + t} + 1 / 2\Delta] \times [\bar{\theta}_{T + t} - 1 / 2\Delta, \bar{\theta}_{T + t} + 1 / 2\Delta]$ is the integration interval. As pointed out in [16], the incorporation of quantization establishes the equivalence between scanpath generation and lossy scanpath compression. + +Furthermore, the absolute Euler coordinate system is not user-centric, meaning that it is not centered at the user's current viewpoint, relative to historical and future viewpoints [16]. This may complicate the probabilistic modeling of scanpaths and the end-to-end optimization of blind PVQA. To address this, we convert the Euler coordinates to the relative $uv$ coordinates: + +$$ +\tilde {\boldsymbol {s}} _ {T - t} = \boldsymbol {\Psi} _ {T - t} (\boldsymbol {s} _ {T}), \text {f o r} t \in \{1, \dots , H \}, \tag {6} +$$ + +where $\Psi_{T - t}(\cdot)$ denotes the mapping of $s_T$ to the viewpoint centered at the reference viewpoint $(\phi_{T - t},\theta_{T - t})$ . By choosing each viewpoint in $s_T$ as the reference, we create $H$ relative scanpaths out of $s_T$ (see Figure 3), which serve as input to the H-Net. Meanwhile, we map the causal scanpath context $c_{t}$ and the viewpoint to be predicted $(\phi_{T + t},\theta_{T + t})$ to the last historical viewpoint centered at $(\phi_{T - 1},\theta_{T - 1})$ . We stack $W$ SGUs to form a scanpath generation block (SGB), which takes $s_T$ as input and predicts the future scanpath $r$ . Furthermore, we stack $M$ SGBs to form the scanpath generator, which is capable of predicting a very long-term scanpath of length $M\times W + H$ (including the initial length $H$ ). The parameters of different SGUs are shared to enable variable-length scanpath generation by varying $M$ . + +Differentiable Scanpath Sampling. To enable end-to-end optimization of the proposed blind PVQA method, we propose a two-step differentiable sampling method to draw + +![](images/f180e081f6730f63d7e3aa992ba4370010d3b799f50acdea1a0bdf5df94e3a34.jpg) +Figure 3. Visualization of a relative scanpath projected from the sphere to the viewpoint. + +viewpoints from the estimated GMM via the reparameterization trick [12, 13]. The first step is to select a Gaussian component from which to sample the viewpoint, according to the categorical distribution: + +$$ +\boldsymbol {e} = \text {o n e} _ {\text {h o t}} \left(\underset {i \in \{1, \dots K \}} {\arg \max } \left(\log \left(\alpha_ {i}\right) + g _ {i}\right)\right), \tag {7} +$$ + +where $g_{i}$ is a sample drawn from the Gumbel(0,1) distribution, and $e$ is a one-hot vector. Eq. (7) is known as the Gumbel-Max trick [8], which is non-differentiable. We relax the arg max operator with a softmax function [12]: + +$$ +\hat {\boldsymbol {e}} = \operatorname {s o f t m a x} \left(\left(\log (\boldsymbol {\alpha}) + \boldsymbol {g}\right) / \tau\right), \tag {8} +$$ + +where $\tau$ represents the temperature coefficient, $\alpha = [\alpha_{1},\dots,\alpha_{K}]$ , and $\pmb {g} = [g_1,\dots,g_K]$ . As $\tau$ approaches zero, $\hat{e}$ converges to $e$ . In the forward pass, $\arg \max$ is used directly, while in the backward pass, it is replaced by the softmax function. The second step involves sampling a viewpoint from the selected Gaussian component. Assuming the $i$ -th Gaussian component is selected, the linear reparameterization trick [13] suggests + +$$ +\tilde {\boldsymbol {r}} _ {t} = \boldsymbol {u} _ {i} + \boldsymbol {L} _ {i} \epsilon , \tag {9} +$$ + +where $\tilde{\boldsymbol{r}}_t$ is the relative $uv$ coordinates of the $t$ -th future viewpoint. $\boldsymbol{\Sigma}_i = \boldsymbol{L}_i\boldsymbol{L}_i^{\mathrm{T}}$ is the Cholesky decomposition, and $\epsilon$ is a sample drawn from the $\mathcal{N}(\mathbf{0},\mathbf{I})$ . Differentiation of the Cholesky decomposition is complicated and sometimes numerically unstable [26]. Thus we assume the independence between the $uv$ coordinates, leading to the simplified reparameterization formula: + +$$ +\tilde {\boldsymbol {r}} _ {t} = \boldsymbol {u} _ {i} + \boldsymbol {\sigma} _ {i} \odot \boldsymbol {\epsilon}, \tag {10} +$$ + +where $\odot$ denotes the element-wise product, and $\sigma_{i}$ denotes the standard deviations of the $i$ -th Gaussian component. Through the two-step reparameterization, our sampling strategy ensures effective back-propagation. As suggested in [16], we additionally implement a proportional-integral-derivative (PID) controller [1] to further improve the smoothness of the sampled scanpaths (see the details in the supplementary material). + +Differentiable Viewport Sequence Generation. Inspired by [11], we first parameterize the Euler sampling grid in terms of the relative $uv$ coordinates via the inverse transformation $\Psi^{-1}$ (see Eq. (6)). Subsequently, the Euler coordinate $(\phi, \theta)$ is mapped to the discrete sampling position $(m, n)$ in the ERP domain: + +$$ +m = (0. 5 - \phi / \pi) H _ {e} - 0. 5, \tag {11} +$$ + +$$ +n = (\theta / 2 \pi + 0. 5) W _ {e} - 0. 5, \tag {12} +$$ + +where $H_{e}$ and $W_{e}$ are the height and width of the video frame in ERP. Once the mapping between $(u, v)$ and $(m, n)$ is established, we construct a flow field that is the same size as the viewpoint. Within this field, each element records the corresponding pixel position $(m, n)$ . Given an ERP image and the flow field, we apply bilinear interpolation [11] to compute pixel values in the viewpoint and leverage its sub-gradients for back-propagation. This process yields $N$ viewpoint sequences, corresponding to $N$ initial paths. + +# 3.2. Quality Assessor + +Our probabilistic scanpath generator can work with any planar VQA model, whether it is differentiable or not. To enable end-to-end optimization of the scanpath generator and the quality assessor, and to make a fair comparison with existing blind PVQA models, we reuse three differentiable quality assessors from ScanpathVQA [35], GSR-S / GSR-X [29], and Assessor360 [36]. Specifically, the quality assessor of ScanpathVQA is a lightweight ResNet-18 network [9], with the classification head replaced by a quality estimator (a multilayer perceptron). The quality assessors of GSR-S / GSR-X are adapted from Video Swin-T [18] / X-Clip-B/32 [22]. The quality assessor of Assessor360 is modified from Swin-B [17] with the addition of a temporal analysis module. We feed each of the $N$ viewpoint sequences to the quality assessor to compute $N$ quality scores. The overall quality estimate is then computed by a simple average: + +$$ +\hat {q} = \frac {1}{N} \sum_ {i = 1} ^ {N} \hat {q} _ {i}. \tag {13} +$$ + +# 3.3. Optimization Strategy + +We explore a three-stage training procedure for our blind PVQA model. In the first stage, we pre-train the density estimation network on the VRVQW dataset [35] by minimizing the expected code length of the generated scanpaths (also equivalent to minimizing the negative log-likelihood): + +$$ +\ell_ {\text {c o d e}} = - \frac {1}{B W} \sum_ {i = 1} ^ {B} \sum_ {t = 0} ^ {W - 1} \log_ {2} \left(P \left(\bar {\boldsymbol {r}} _ {t} ^ {(i)} \mid \boldsymbol {s} ^ {(i)}, \boldsymbol {c} _ {t} ^ {(i)}\right)\right), \tag {14} +$$ + +where $B$ denotes the mini-batch size. During this stage of training, we use human scanpaths to fill in the causal path context, which can be efficiently implemented by a causal masking mechanism. In the second stage, we fix the parameters of the pre-trained scanpath generator and warm up the quality assessor by optimizing the Pearson linear correlation coefficient (PLCC) between human perceptual scores and model predictions. In the third stage, we end-to-end finetune the entire method. We find that the proposed three-stage optimization strategy accelerates convergence compared to the naive end-to-end optimization. + +# 4. Experiments + +In this section, we first delineate the experimental setups, and then compare our method with current blind PVQA models under both in-dataset and cross-dataset settings. We further validate our scanpath generator in terms of explaining human perceptual scores and replicating human viewing patterns. Lastly, we conduct a series of ablation experiments to probe the impact of several key designs. + +# 4.1. Experimental Setups + +Datasets. We employ three panoramic image and video datasets: VRVQW [35], CVIQD [30], and OIQA [6]. The VRVQW dataset includes 502 panoramic videos that have a wide spectrum of authentic distortions. Each video is viewed under four unique viewing conditions to simulate the different quality of experience during the initial viewing. The CVIQD dataset comprises a total of 528 compressed panoramic images by JPEG, AVC, and HEVC, from 16 reference images. The OIQA dataset includes 320 panoramic images that have been altered from 16 reference images by JPEG compression, JPEG2000 compression, Gaussian blurring, and Gaussian noise contamination. + +Implementation Details. For the scanpath generator, we set the length of the provided initial path $H$ and the predicted future path $W$ in the SGB to be identical and equal to 5. The number of Gaussian components $K$ in Eq. (3) is set to 3. The quantization step size $\Delta$ in Eq. (5) is set to 0.2. The temperature coefficient $\tau$ in Eq. (8) is set to 1. The number of stacked SGBs $M$ is set to 6 and 14 for the viewing duration of 7 and 15 seconds, respectively. For the quality assessor, the number of scanpaths $N$ in Eq. (13) is set to 20. The input viewpoint size $H_{v} \times W_{v}$ is $224 \times 224$ , corresponding to a field of view of $90^{\circ} \times 90^{\circ}$ . The length of the viewpoint sequence is set to 7 regardless of the duration and frame rate of the original panoramic video. We split VRVQW randomly into the training, validation, and test sets according to the ratio of $6:2:2$ for 5 times, and report the mean results. Similarly, we split CVIQD and OIQA using a different ratio of $7:1:2$ for 5 times. The detailed configuration of our three-stage optimization strategy is detailed in the supplementary material. + +Table 1. In-dataset comparison of blind PVQA methods on three panoramic image and video quality datasets. SRCC: Spearman's rank correlation coefficient. PLCC: Pearson linear correlation coefficient. The same evaluation metrics are applied in Tables 2, 3, 5, 6 and Figure 6. The best results on each dataset are highlighted in bold. + +
DatasetMethodSRCCPLCC
VRVQW [35]NIQE [19]0.4010.365
MC360IQA [31]0.6690.671
Wen24 [35]0.7560.763
ScanpathVQA [35]0.7790.781
Assessor360 [36]0.4060.415
Ours (ScanpathVQA)0.8010.809
Ours (GSR-S)0.8040.807
Ours (GSR-X)0.8150.819
Ours (Assessor360)0.8220.823
CVIQD [30]NIQE [19]0.8470.878
MC360IQA [31]0.9170.939
Wen24 [35]0.9190.932
GSR-S [29]0.9050.937
GSR-X [29]0.9440.962
Assessor360 [36]0.9550.969
Ours (ScanpathVQA)0.9120.936
Ours (GSR-S)0.9300.958
Ours (GSR-X)0.9560.974
Ours (Assessor360)0.9720.983
OIQA [6]NIQE [19]0.7020.657
MC360IQA [31]0.9000.906
Wen24 [35]0.9050.907
GSR-S [29]0.9020.915
GSR-X [29]0.9450.954
Assessor360 [36]0.9460.953
Ours (ScanpathVQA)0.9150.922
Ours (GSR-S)0.9270.936
Ours (GSR-X)0.9560.967
Ours (Assessor360)0.9600.971
+ +Table 2. Cross-dataset comparison of blind PVQA methods on the CVIQD [30] and OIQA [6] datasets. The arrow points from the training set to the test set. + +
DatasetMethodSRCCPLCC
OIQAMC360IQA0.7980.842
GSR-X0.7620.841
Assessor3600.8590.893
CVIQDOurs (ScanpathVQA)0.7330.747
Ours (Assessor360)0.8720.904
CVIQDMC360IQA0.2880.349
GSR-X0.6950.718
Assessor3600.3380.467
OIQAOurs (ScanpathVQA)0.6360.658
Ours (Assessor360)0.7030.715
+ +# 4.2. Main Results + +We compare our blind PVQA method with seven existing models, including NIQE [19], MC360IQA [31], Wen24 [35], ScanpathVQA [35], GSR-S [29], GSR-X [29], + +Table 3. Comparison of different scanpath generators for explaining human perceptual scores. + +
DatasetMethodSRCCPLCC
VRVQWHuman scanpath0.7860.790
Random sampling0.0750.104
Heuristic sampling [36]0.4310.443
Xu18 [38]0.7120.717
TRACK [25]0.7450.749
Li23 [16]0.7900.794
Ours0.8050.814
CVIQDRandom sampling0.6320.640
Heuristic sampling [36]0.8680.870
ScanDMM [28]0.8560.864
Li23 [16]0.8140.827
Ours0.9280.940
OIQARandom sampling0.5140.536
Heuristic sampling [36]0.8610.872
ScanDMM [28]0.8650.877
Li23 [16]0.7930.799
Ours0.9140.917
+ +and Assessor360 [36]. For image quality models such as MC360IQA [31] and Assessor360 [36], we retrain them on the VRVQW dataset. In the case of MC360IQA, we assign the video-level quality score to each key frame and use their temporally averaged score for testing. Assessor360's scanpath generator is adapted to videos by adjusting the semantic context associated with each key frame. + +In-dataset Results. Table 1 shows the Spearman's rank correlation coefficient (SRCC) and $\mathrm{PLCC}^1$ results under the in-dataset setting. It is evident that our learned scanpaths enhance the performance of all quality assessors compared to other scanpath-based methods. When integrated with the quality assessor from Assessor360 (i.e., a modified SwinB with a temporal analysis module), our proposed method achieves the best results on all three datasets. Furthermore, our scanpath generator can boost a simpler quality assessor (e.g., from ScanpathVQA [35] with approximately 11 million parameters) to reach performance levels similar to those of a more sophisticated quality assessor coupled with a weaker scanpath generator (e.g., GSR-S [29] with approximately 112 million parameters). Methods that overlook human viewing patterns, like NIQE [19] and MC360IQA [31], fail to accurately model the human perception of panoramic image and video quality, especially on VRVQW. Additionally, we find that assessing the quality of panoramic videos with authentic distortions tends to be more challenging than for panoramic images with synthetic distortions. This is anticipated because authentic distortions in panoramic videos typically present as a complex blend of various artifacts, localized in space and time. + +Cross-dataset Results. Table 2 shows the SRCC and PLCC results under the cross-dataset settings on CVIQD [30] and + +Table 4. Comparison of different scanpath generators for replicating human viewing patterns using the minimum orthodromic distance (minOD) and maximum temporal correlation (maxTC). + +
MethodminOD ↓maxTC ↑
Heuristic sampling [36]1.3250.401
Xu18 [38]1.1850.637
TRACK [25]1.0670.699
Li23 [16]0.5420.796
Ours (w/o end-to-end optimization)0.5560.781
Ours0.5360.805
+ +![](images/72c6a5319bac4183d4a9c5b93a002c2f646bc3aafcfbcdd34a916aaa792d3311.jpg) +Figure 4. Comparison of different scanpath predictors in terms of maxTC with different prediction horizons. + +OIQA [6]. Generally, models trained on the OIQA dataset show better generalizability. This is likely because OIQA encompasses a broader range of distortion types, compared to CVIQD, which only includes the compression artifacts. Assessor360 shows a noticeable performance drop when tested on OIQA, potentially indicative of overfitting. Our scanpath generator is capable of restoring Assessor360's performance, which provides a strong indication of its effectiveness through end-to-end optimization. + +# 4.3. Scanpath Generator Validation + +Explaining Human Perceptual Scores. We conduct an apple-to/apple comparison of different scanpath generators in terms of explaining human perceptual scores by fixing the quality assessor to that used in ScanpathVQA. These include random sampling, heuristic sampling [36], Xu18 [38], TRACK [25], ScanDMM [28], Li23 [16], and our method. Table 3 shows the SRCC and PLCC results on the VRVQW, CVIQD and OIQA datasets. It is noteworthy that our scanpath generator outperforms all competing methods across all three datasets, even surpassing the human-level performance on VRVQW. The performance of random and heuristic sampling decreases sharply on VRVQW, due to the presence of spatiotemporally localized authentic distortions. + +Replicating Human Viewing Patterns. We also test different scanpath generators in terms of replicating human viewing patterns on the VRVQW dataset by comparing the predicted scanpaths to those of humans. We use two set-to-set + +![](images/5fc3e3debab2a7957e0a7b3088cf52690a9fb15f4afbc1aa1924bd9d8c941d2a.jpg) +Figure 5. Comparison of saliency maps generated from scanpaths by our method and those by humans. + +Table 5. Impact of optimization strategies on blind PVQA. Training protocol: 1) Two-stage w/ fixed scanpaths, 2) Two-stage w/ varied scanpaths, and 3) Three-stage w/ end-to-end optimization. + +
ProtocolVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
10.7690.7720.7100.7800.6880.742
20.7810.7850.8300.8590.7980.815
30.8050.8140.9280.9400.9140.917
+ +evaluation metrics: the minimum orthodromic distance (i.e., minOD) and maximum temporal correlation (i.e., maxTC), as suggested in [16]. Given a set of human scanpaths, $S = \{s^{(i)}\}_{i=1}^{|S|}$ , the minimum orthodromic distance between $S$ and the set of predicted scanpaths $\hat{S} = \{\hat{s}^{(i)}\}_{i=1}^{|S|}$ can be computed by + +$$ +\min \mathrm {O D} (\mathcal {S}, \hat {\mathcal {S}}) = \min _ {\boldsymbol {s} \in \mathcal {S}, \boldsymbol {s} \in \hat {\mathcal {S}}} \mathrm {O D} (\boldsymbol {s}, \hat {\boldsymbol {s}}), \tag {15} +$$ + +where the orthodromic distance $\mathrm{OD}(\cdot ,\cdot)$ is defined as + +$$ +\begin{array}{l} \mathrm {O D} (\boldsymbol {s}, \hat {\boldsymbol {s}}) = \frac {1}{T} \sum_ {t = 0} ^ {T - 1} \operatorname {a r c c o s} \left(\cos \left(\phi_ {t}\right) \cos \left(\hat {\phi} _ {t}\right) \cos \left(\theta_ {t} - \hat {\theta} _ {t}\right) \right. \\ + \sin \left(\phi_ {t}\right) \sin \left(\hat {\phi} _ {t}\right)). \tag {16} \\ \end{array} +$$ + +The maximum temporal correlation between $\mathcal{S}$ and $\hat{\mathcal{S}}$ is defined as + +$$ +\max \mathrm {T C} (\mathcal {S}, \hat {\mathcal {S}}) = \max _ {\boldsymbol {s} \in \mathcal {S}, \hat {\boldsymbol {s}} \in \hat {\mathcal {S}}} \mathrm {T C} (\boldsymbol {s}, \hat {\boldsymbol {s}}), \tag {17} +$$ + +where the temporal correlation is computed by + +$$ +\begin{array}{l} \operatorname {T C} \left(\boldsymbol {s} ^ {(i)}, \boldsymbol {s} ^ {(j)}\right) = \\ \frac {1}{2} \left(\operatorname {P L C C} \left(\phi^ {(i)}, \phi^ {(j)}\right) + \operatorname {P L C C} \left(\boldsymbol {\theta} ^ {(i)}, \boldsymbol {\theta} ^ {(j)}\right)\right). \end{array} \tag {18} +$$ + +Table 4 presents the minOD and maxTC results, from which we find that our end-to-end optimized scanpath generator delivers the best results, surpassing its independently + +Table 6. Impact of visual context on blind PVQA. #Parameters added to the scanpath generator are also shown. + +
Method#ParametersVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
Ours1M0.8050.8140.9280.9400.9140.917
Ours w/ visual context27M0.8160.8230.9370.9560.9250.934
+ +optimized counterpart by a clear margin. The heuristic sampling [36] that depends on the simplified entropy features and equator bias, is inadequate for capturing human viewing patterns, especially for long-term prediction horizons (see Figure 4). Due to the deterministic nature, Xu18 [38] and TRACK [25] fail to accommodate the diversity and uncertainty inherent in human scanpaths, resulting in subpar performance. Incorporating historical video frames as the visual context, Li23 [16] shows performance on par with our method, reinforcing our assertion that visual context informs less about future viewpoints. Figure 5 demonstrates a comparison of the saliency maps derived from scanpaths by our method and those by humans, offering further proof of the close alignment of our scanpath generator and human viewing behaviors. + +# 4.4. Ablation Studies + +Impact of Optimization Strategies. We explore three different optimization strategies: 1) a two-stage approach where the pre-trained scanpath generator produces a fixed set of scanpaths for the training of the quality assessor, 2) a similar two-stage approach but supplying a varied set of scanpaths in each epoch of training, and 3) the default three-stage approach that enables end-to-end optimization. From the results in Table 5, we find that the two-stage approach benefits from "data augmentation" with varied scanpaths in each epoch. Our three-stage end-to-end optimization strategy further boosts the accuracy of quality prediction by jointly finetuning both the scanpath generator and quality assessor. + +Impact of Visual Context. To assess the impact of visual context on blind PVQA, we enhance our scanpath generator with a video analysis network [16], implemented by a variant of ResNet-50 for frame-level feature extraction and aggregation. We subsequently integrate it with the ScanpathVQA quality assessor for blind PVQA, with the results shown in Table 6. We find that the visual context has a negligible effect on blind PVQA, and thus we exclude it in the generation of scanpaths. + +Impact of the Number and Length of Viewport Sequences. We investigate the effects of varying the number $N$ and length $L$ of viewpoint sequences on blind PVQA. We test $N$ values from $\{5,10,15,20,50\}$ and $L$ values from $\{4,7,15\}$ . Figure 6 shows the SRCC results on VRVQW, using the ScanpathVQA quality assessor for prediction. It is clear that $N = 20$ viewpoint sequences are sufficient for reliable quality assessment, with performance remaining stable + +![](images/b3abb9f809bf7a66d4ab2271e69ac1348ec74016d2cabd6e9deb12476f2bcad2.jpg) +Figure 6. Impact of the number $N$ and length $L$ of viewpoint sequences on blind PVQA. + +with the increase in $N$ . For sequence length, $L = 7$ appears to be a wise choice. Further increasing $L$ does not noticeably affect the performance, but would lead to a considerable rise in computational demand. Conversely, a shorter viewpoint sequence results in a noticeable drop in performance due to the loss of information from excessive temporal downsampling. It is important to note that these findings are specific to the ScanpathVQA quality assessor and may differ from other assessors. + +# 5. Conclusion + +We have introduced an end-to-end optimized blind PVQA method, consisting of a scanpath generator and a quality assessor. The proposed scanpath generator is differentiable and can be integrated with any planar VQA model, whose effectiveness has been thoroughly validated in supporting blind PVQA and in modeling human viewing patterns. Additionally, we have also devised a three-stage optimization strategy to facilitate training convergence, which aligns with current large-scale optimization practices that involve self-supervised pre-training followed by supervised finetuning, including initial warmup phases [10]. + +# 6. Acknowledgement + +This work was supported in part by the National Natural Science Foundation of China (62071407 and 62102339), the Shenzhen Science and Technology Program (RCBS20221008093121052), the Research Grants Council of Hong Kong (ECS 2121382, ECS 27212822, and GRF 17201822), and the CCF-Tencent Rhino Bird Fund (9239061). + +# References + +[1] Kiam Heong Ang, Gregory Chong, and Yun Li. PID control system analysis, design, and technology. IEEE Transactions on Control Systems Technology, 13(4):559-576, 2005. 4 +[2] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 4 +[3] Fang-Yi Chao, Cagri Ozcinar, and Aljosa Smolic. Transformer-based long-term viewpoint prediction in $360^{\circ}$ video: Scanpath is all you need. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2021. 2 +[4] Sijia Chen, Yingxue Zhang, Yiming Li, Zhenzhong Chen, and Zhou Wang. Spherical structural similarity index for objective omnidirectional video quality assessment. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2018. 1, 2 +[5] Taco S. Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, pages 1-15, 2018. 2 +[6] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Yucheng Zhu, Yi Fang, and Xiaokang Yang. Perceptual quality assessment of omnidirectional images. In IEEE International Symposium on Circuits and Systems, pages 1-5, 2018. 2, 5, 6, 7 +[7] Jun Fu, Chen Hou, Wei Zhou, Jiahua Xu, and Zhibo Chen. Adaptive hypergraph convolutional network for no-reference 360-degree image quality assessment. In ACM International Conference on Multimedia, pages 961-969, 2022. 1, 2 +[8] Emil J. Gumbel. Statistical Theory of Extreme Values and Some Practical Applications. US Government Printing Office, 1948. 4 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5 +[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8 +[11] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017-2025, 2015. 2, 5 +[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In International Conference on Learning Representations, pages 1–12, 2017. 2, 4 +[13] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, pages 1–14, 2014. 2, 4 +[14] Chen Li, Mai Xu, Lai Jiang, Shanyi Zhang, and Xiaoming Tao. Viewport proposal CNN for $360^{\circ}$ video quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10169-10178, 2019. 1, 2 +[15] Chenge Li, Weixi Zhang, Yong Liu, and Yao Wang. Very long term field of view prediction for 360-degree video streaming. In IEEE Conference on Multimedia Information Processing and Retrieval, pages 297-302, 2019. 2 + +[16] Mu Li, Kanglong Fan, and Kede Ma. Scanpath prediction in panoramic videos via expected code length minimization. arXiv preprint arXiv:2305.02536, 2023. 2, 4, 6, 7, 8 +[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 5 +[18] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 5 +[19] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 6 +[20] King-To Ng, Shing-Chow Chan, and Heung-Yeung Shum. Data compression and transmission aspects of panoramic videos. IEEE Transactions on Circuits and Systems for Video Technology, 15(1):82-95, 2005. 1 +[21] Anh Nguyen, Zhisheng Yan, and Klara Nahrstedt. Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In ACM International Conference on Multimedia, pages 1190-1198, 2018. 2 +[22] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18, 2022. 5 +[23] David Noton and Lawrence Stark. Scanpaths in saccadic eye movements while viewing and recognizing patterns. Vision Research, 11(9):929-942, 1971. 1 +[24] David Noton and Lawrence Stark. Scanpaths in eye movements during pattern perception. Science, 171(3968):308-311, 1971. 1 +[25] Miguel Fabian Romero Rondon, Lucile Sassatelli, Ramón Aparicio-Pardo, and Frédéric Precioso. TRACK: A new method from a re-examination of deep architectures for head motion prediction in $360^{\circ}$ videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5681-5699, 2022. 2, 6, 7, 8 +[26] Stephen P. Smith. Differentiation of the Cholesky algorithm. Journal of Computational and Graphical Statistics, 4(2):134-147, 1995. 4 +[27] Xiangjie Sui, Kede Ma, Yiru Yao, and Yuming Fang. Perceptual quality assessment of omnidirectional images as moving camera videos. IEEE Transactions on Visualization and Computer Graphics, 28(8):3022-3034, 2021. 1, 2 +[28] Xiangjie Sui, Yuming Fang, Hanwei Zhu, Shiqi Wang, and Zhou Wang. ScanDMM: A deep markov model of scanpath prediction for $360^{\circ}$ images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6989-6999, 2023. 2, 6, 7 +[29] Xiangjie Sui, Hanwei Zhu, Xuelin Liu, Yuming Fang, Shiqi Wang, and Zhou Wang. Perceptual quality assessment of $360^{\circ}$ images based on generative scanpath representation. arXiv preprint arXiv:2309.03472, 2023. 1, 2, 5, 6 + +[30] Wei Sun, Ke Gu, Siwei Ma, Wenhan Zhu, Ning Liu, and Guangtao Zhai. A large-scale compressed 360-degree spherical image database: From subjective quality evaluation to objective model comparison. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2018. 2, 5, 6 +[31] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Huiyu Duan, and Siwei Ma. MC360IQA: A multi-channel CNN for blind 360-degree image quality assessment. IEEE Journal of Selected Topics in Signal Processing, 14(1):64-77, 2020. 6 +[32] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters, 24(9):1408-1412, 2017. 1, 2 +[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000-6010, 2017. 2 +[34] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 2 +[35] Wen Wen, Mu Li, Yiru Yao, Xiangjie Sui, Yabin Zhang, Long Lan, Yuming Fang, and Kede Ma. Perceptual quality assessment of virtual reality videos in the wild. IEEE Transactions on Circuits and Systems for Video Technology, to appear, 2024. 1, 2, 5, 6 +[36] Tianhe Wu, Shuwei Shi, Haoming Cai, Mingdeng Cao, Jing Xiao, Yinqiang Zheng, and Yujiu Yang. Assessor360: Multisession network for blind omnidirectional image quality assessment. In Advances in Neural Information Processing Systems, pages 1-22, 2023. 1, 2, 5, 6, 7, 8 +[37] Jiahua Xu, Wei Zhou, and Zhibo Chen. Blind omnidirectional image quality assessment with viewpoint oriented graph convolutional networks. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1724-1737, 2020. 1, 2 +[38] Yanyu Xu, Yanbing Dong, Junru Wu, Zhengzhong Sun, Zhiru Shi, Jingyi Yu, and Shenghua Gao. Gaze prediction in dynamic $360^{\circ}$ immersive videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5333-5342, 2018. 2, 6, 7, 8 +[39] Jiachen Yang, Tianlin Liu, Bin Jiang, Wen Lu, and Qinggang Meng. Panoramic video quality assessment based on non-local spherical CNN. IEEE Transactions on Multimedia, 23: 797-809, 2021. 1, 2 +[40] Matt Yu, Haricharan Lakshman, and Bernd Girod. A framework to evaluate omnidirectional video coding schemes. In IEEE International Symposium on Mixed and Augmented Reality, pages 31-36, 2015. 1, 2 +[41] Vladyslav Zakharchenko, Kwang Pyo Choi, and Jeong Hoon Park. Quality metric for spherical panoramic video. In SPIE Optics and Photonics for Information Processing X, pages 57-65, 2016. 1, 2 \ No newline at end of file diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/images.zip b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5dc85030eadb447b40dea5052eded66a76e1d6ee --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b75e21b264e4579d6d8fbb3a67709329f8db1a9af6feb72f3456c0ed33726025 +size 512030 diff --git a/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/layout.json b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e7fbb20e126047948ea4757001ad10611ce8f71 --- /dev/null +++ b/2024/Learned Scanpaths Aid Blind Panoramic Video Quality Assessment/layout.json @@ -0,0 +1,9037 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 506, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 506, + 120 + ], + "type": "text", + "content": "Learned Scanpaths Aid Blind Panoramic Video Quality Assessment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": "Kanglong Fan" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": ", Wen Wen" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": ", Mu Li" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": ", Yifan Peng" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": ", and Kede Ma" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": " City University of Hong Kong, " + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": " Harbin Institute of Technology, Shenzhen \n" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 89, + 142, + 505, + 228 + ], + "type": "text", + "content": " The University of Hong Kong \n{kanglofan2-c, wwen29-c}@my.cityu.edu.hk, limuhit@gmail.com \nevanpeng@hku.hk, kede.ma@cityu.edu.hk \nhttps://github.com/kalofan/AutoScanpathQA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "spans": [ + { + "bbox": [ + 143, + 255, + 192, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 280, + 290, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 280, + 290, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 290, + 495 + ], + "type": "text", + "content": "Panoramic videos have the advantage of providing an immersive and interactive viewing experience. Nevertheless, their spherical nature gives rise to various and uncertain user viewing behaviors, which poses significant challenges for panoramic video quality assessment (PVQA). In this work, we propose an end-to-end optimized, blind PVQA method with explicit modeling of user viewing patterns through visual scanpaths. Our method consists of two modules: a scanpath generator and a quality assessor. The scanpath generator is initially trained to predict future scanpaths by minimizing their expected code length and then jointly optimized with the quality assessor for quality prediction. Our blind PVQA method enables direct quality assessment of panoramic images by treating them as videos composed of identical frames. Experiments on three public panoramic image and video quality datasets, encompassing both synthetic and authentic distortions, validate the superiority of our blind PVQA model over existing methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 519, + 128, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 128, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 128, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "text", + "content": "The rapid advancement of multimedia technologies has marked the beginning of a new era characterized by the proliferation of panoramic videos [20]. Such type of digital data offers an immersive and interactive viewing experience that is transforming the way we consume multimedia. Therefore, assessing and ensuring the visual quality of panoramic videos is increasingly important, as it shapes the users' viewing experience and the triumph of any product or service based on panoramic videos [40]. Unlike their planar counterparts, panoramic videos provide a " + }, + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 539, + 287, + 694 + ], + "type": "text", + "content": " broad view with a spherical data structure, which poses significant computational challenges for panoramic video quality assessment (PVQA). Moreover, the diverse and uncertain user" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 255, + 548, + 384 + ], + "blocks": [ + { + "bbox": [ + 309, + 255, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 309, + 255, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 309, + 255, + 548, + 384 + ], + "type": "image", + "image_path": "fba2825827a098313c5e55c4139537b6fedece13fe40771d2eabae889e529071.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 395, + 547, + 418 + ], + "lines": [ + { + "bbox": [ + 305, + 395, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 547, + 418 + ], + "type": "text", + "content": "Figure 1. Analogy between human subjects and our end-to-end optimized method for panoramic video quality assessment." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 495 + ], + "type": "text", + "content": "viewing behaviors (in the form of visual scanpaths) induced by the spherical structure further complicate the quality prediction process. Addressing these challenges requires novel PVQA models that take into account both the spherical data structure as well as user viewing patterns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "In the quality assessment of panoramic images and videos, three approaches are commonly employed: sphereto-plane projection onto a 2D plane, rectilinear projection onto multiple viewpoints, and direct processing using spherical operations. According to the Theorema Egregium by Gauss, all sphere-to-plane map projections [32, 35, 41] are impeded by non-uniform sampling and geometric distortions, which may bias subsequent planar quality prediction. While spherical operators give a better account for the panoramic data structure, they are generally computationally prohibitive and, more importantly, may not faithfully reflect user viewing patterns [4, 39, 40]. To overcome these computational difficulties, several methods seek to sample and process rectilinear viewpoints [7, 14, 27, 29, 36, 37]. Of particular interest are scanpath-based methods, which sample, along visual scanpaths [23, 24], sequences of rectilinear viewpoints at discrete time instances. This sampling process turns panoramic images and videos into moving-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2599" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 217, + 83 + ], + "type": "text", + "content": "camera videos, amenable to planar VQA." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 227 + ], + "type": "text", + "content": "By closely imitating how humans perceive visual distortions in virtual environments (see Figure 1), scanpath-based methods [27, 29, 36] have demonstrated remarkable efficacy in the quality of panoramic images. Nonetheless, some methods [27] rely on human visual scanpaths for assessment, which are cumbersome and time-consuming to obtain, thus limiting their applications in fully-automated situations. Some other methods [29, 36] design and refine the scanpath generator separately from the quality predictor, which is bound to be suboptimal. Moreover, while all methods prove effective with panoramic images, their adaptability for use with panoramic videos remains unclear." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 228, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 515 + ], + "type": "text", + "content": "In this work, we further pursue the scanpath-based methods for end-to-end optimized blind PVQA. Our method consists of two modules: a scanpath generator and a quality assessor. Our scanpath generator is probabilistic, which takes historical scanpaths as input and is pre-trained to predict future scanpaths by minimizing their expected code length [16]. The scanpath generator and the quality assessor are then jointly optimized to explain human perceptual scores of panoramic videos. To enable end-to-end optimization, we employ the reparameterization trick [12, 13] to allow differentiable scanpath sampling and adopt subgradients to handle discontinuities of the interpolation kernel [11] for viewpoint sequence generation. Our blind PVQA method not only eliminates the need for human scanpaths, but also supplies a lightweight and differentiable scanpath generator that can work with any planar VQA model. Furthermore, our method is \"backward compatible,\" in the sense that it handles panoramic images with no modification. We test the proposed blind PVQA models on three public panoramic image and video quality datasets [6, 30, 35], covering both synthetic and authentic distortions. Under both in-dataset and cross-dataset settings, our models consistently exhibit better quality prediction performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 525, + 134, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 134, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 134, + 537 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 286, + 568 + ], + "type": "text", + "content": "We review two highly relevant topics, scanpath generation and quality assessment of panoramic images and videos." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 575, + 169, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 169, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 169, + 588 + ], + "type": "text", + "content": "2.1. Scanpath Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Typical inputs to a panoramic scanpath generator include the saliency map, optical flow map, and historical scanpath. To improve saliency detection, Nguyen et al. [21] compiled a panoramic video saliency dataset, while Xu et al. [38] focused on relative viewpoint displacement prediction. Apart from the historical scanpath, Li et al. [15] incorporated \"future\" scanpaths from other users to facilitate cross-user transfer learning. Through an in-depth root-cause analysis, Rondon et al. [25] discovered that visual features have a minimal impact on the prediction of short-term scanpaths" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "(e.g., " + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "inline_equation", + "content": "\\leq 2" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": " seconds). Motivated by their findings, Chao et al. [3] trained a Transformer [33] to predict future scanpaths based solely on historical scanpaths." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 204 + ], + "type": "text", + "content": "The above-mentioned methods [3, 15, 21, 25, 38] treat scanpath generation as a deterministic prediction task, neglecting the inherent scanpath diversity and uncertainty. As a departure, Li et al. [16] formulated scanpath generation as a density estimation problem, which can be implemented by expected code length minimization. In our work, we adopt Li's approach [16] to learn multi-user viewing patterns and generate human-like scanpaths." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 211, + 420, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 420, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 420, + 223 + ], + "type": "text", + "content": "2.2. Quality Assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 277 + ], + "type": "text", + "content": "Current PVQA models are primarily derived from planar image and video quality methods, which are applied to three types of data representations: the projected 2D plane, spherical surface, and projected rectilinear viewpoint." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 277, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 491 + ], + "type": "text", + "content": "Planar domain methods [32, 35, 41] aim to rectify geometric distortions and mitigate uneven sampling that results from the sphere-to-plane projection. These include the latitude-adaptive weighting [32], Craster parabolic projection [41], and pseudocylindrical representation [35]. Spherical domain methods, such as S-PSNR [40] and S-SSIM [4], compute and pool local quality measurements over the sphere. Yang et al. [39] trained a non-local spherical neural network [5, 34] to extract spatiotemporal information from panoramic videos. Viewport domain methods prioritize the extraction of visually informative viewpoints for quality analysis. Li et al. [14] introduced a two-step approach that involves viewpoint proposal and quality assessment. Xu et al. [37] built a graph over the extracted viewpoints, and Fu et al. [7] constructed hypergraphs to represent the semantic interactions between viewpoints. One limitation of current viewpoint proposal methods is that they do not accurately reflect the human viewing experience." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 624 + ], + "type": "text", + "content": "Sui et al. [27] pioneered scanpath-based methods for PVQA, under the category of viewpoint domain methods. To eliminate the dependency on human scanpaths, Sui et al. [29] adopted a deep Markov model [28] to generate scanpaths. Meanwhile, Wu et al. [36] handcrafted a simple scanpath generator based on the entropy feature and equator bias. These methods are tailored for panoramic images and are not end-to-end optimized. In contrast, we aim ambitiously for an end-to-end optimized quality assessment method for panoramic videos, with the added benefit of being backward compatible with panoramic images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 411, + 647 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "As illustrated in Figure 2, our method consists of two modules: a scanpath generator and a quality assessor. Given a panoramic video, we first specify a starting point " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "(\\phi_0,\\theta_0)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", a viewing duration " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " initial paths. The scanpath generator autoregressively samples " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " scanpaths based on the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2600" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 72, + 542, + 178 + ], + "blocks": [ + { + "bbox": [ + 51, + 72, + 542, + 178 + ], + "lines": [ + { + "bbox": [ + 51, + 72, + 542, + 178 + ], + "spans": [ + { + "bbox": [ + 51, + 72, + 542, + 178 + ], + "type": "image", + "image_path": "aaaae09a7792ea19c8d4fd9e2b33dbc66e49f5210b3ee9f93ee2e2e8dde563fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 181, + 194, + 274 + ], + "blocks": [ + { + "bbox": [ + 51, + 181, + 194, + 274 + ], + "lines": [ + { + "bbox": [ + 51, + 181, + 194, + 274 + ], + "spans": [ + { + "bbox": [ + 51, + 181, + 194, + 274 + ], + "type": "image", + "image_path": "8051603bb40278b82cac6cec6539f854f752cc1b70e08b8d374624c7bbd87c3e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "lines": [ + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed blind PVQA method, consisting of a scanpath generator and a quality assessor. The basic component of the scanpath generator is the scanpath generation unit (SGU), which utilizes the historical and causal relative scanpaths to produce the GMM parameters for differentiable sampling of the current viewpoint. By assembling " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " SGUs, we create a scanpath generation block (SGB), which autoregressively predicts a future scanpath of " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " viewpoints. We further stack " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " SGBs to generate a long-term scanpath of " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "M \\times W + H" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " viewpoints, where " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " is the length of the initial path. By adjusting the number of initial paths (denoted by " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": "), we can sample " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " scanpaths, along which we produce " + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 286, + 545, + 354 + ], + "type": "text", + "content": " viewpoint sequences as input to the quality assessor." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 197, + 182, + 543, + 274 + ], + "blocks": [ + { + "bbox": [ + 197, + 182, + 543, + 274 + ], + "lines": [ + { + "bbox": [ + 197, + 182, + 543, + 274 + ], + "spans": [ + { + "bbox": [ + 197, + 182, + 543, + 274 + ], + "type": "image", + "image_path": "023c04e5c6b15a1c7a6f0723741c334cbd08ac997346fd0ca773e1003c44d22c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 287, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 447 + ], + "type": "text", + "content": "initial and already generated path segments. Along these scanpaths, we apply a differentiable viewpoint generation technique to extract " + }, + { + "bbox": [ + 46, + 363, + 287, + 447 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 363, + 287, + 447 + ], + "type": "text", + "content": " viewpoint sequences from the input panoramic video. Each viewpoint sequence (as a planar video) is fed to the quality assessor, whose predicted score is subsequently aggregated into an overall quality estimate of the panoramic video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 455, + 165, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 165, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 165, + 467 + ], + "type": "text", + "content": "3.1. Scanpath Generator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 521 + ], + "type": "text", + "content": "Probabilistic Scanpath Modeling. To capture the uncertainty and diversity of human scanpaths, we formulate panoramic scanpath generation as a density estimation problem:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 140, + 532, + 287, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 532, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 140, + 532, + 287, + 544 + ], + "type": "interline_equation", + "content": "\\max p (\\boldsymbol {r} | \\boldsymbol {s}), \\tag {1}", + "image_path": "bc1d77572ed2955d3deae705b761ad7a5ecc06870bde27f0607d51a8e642bd99.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "s = \\{(\\phi_0,\\theta_0),\\dots ,(\\phi_t,\\theta_t),\\dots ,(\\phi_{T - 1},\\theta_{T - 1})\\}" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " is the historical scanpath as the condition and " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "\\pmb {r} = \\{(\\phi_T,\\theta_T),\\ldots ,(\\phi_{T + W - 1},\\theta_{T + W - 1})\\}" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " is the future scanpath to be predicted. Herein, " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " is the prediction horizon, and " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "(\\phi_t,\\theta_t)" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " -th viewpoint in the Euler coordinate system. Mathematically, " + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol {r}|\\boldsymbol {s})" + }, + { + "bbox": [ + 47, + 554, + 286, + 625 + ], + "type": "text", + "content": " can be decomposed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 635, + 287, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 635, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 89, + 635, + 287, + 667 + ], + "type": "interline_equation", + "content": "p (\\boldsymbol {r} \\mid \\boldsymbol {s}) = \\prod_ {t = 0} ^ {W - 1} p \\left(\\phi_ {T + t}, \\theta_ {T + t} \\mid \\boldsymbol {s}, \\boldsymbol {c} _ {t}\\right), \\tag {2}", + "image_path": "4af34865d275389a6326ae9a4cb49c84312056d52540132f4c3fdc042855866b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{c}_t = \\{(\\phi_T, \\theta_T), \\dots, (\\phi_{T + t - 1}, \\theta_{T + t - 1})\\}" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": " is referred to as the causal path that includes all estimated viewpoints before " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\phi_{T + t}, \\theta_{T + t})" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "c_0 = \\emptyset" + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "text", + "content": "chain rule suggests estimating the conditional probability " + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "inline_equation", + "content": "p\\left(\\phi_{T + t}, \\theta_{T + t} \\mid s, c_t\\right)" + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "text", + "content": " autoregressively. We further make the Markovian assumption: prediction of the current viewpoint is conditionally independent of viewpoints that are temporally further distant, given the most recent " + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "text", + "content": " viewpoints. This leads to a truncated historical path context " + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "inline_equation", + "content": "s_T = \\{(\\phi_{T - H}, \\theta_{T - H}), \\ldots, (\\phi_{T - 1}, \\theta_{T - 1})\\}" + }, + { + "bbox": [ + 304, + 363, + 545, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "text", + "content": "We parameterize the probability " + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "inline_equation", + "content": "p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)" + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_t = (\\phi_{T + t}, \\theta_{T + t})" + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "text", + "content": ", by a Gaussian mixture model (GMM) with " + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 453, + 545, + 494 + ], + "type": "text", + "content": " components:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 344, + 504, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 344, + 504, + 545, + 536 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {r} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\sum_ {i = 1} ^ {K} \\alpha_ {i} \\mathcal {N} _ {i} \\left(\\boldsymbol {r} _ {t}; \\boldsymbol {\\mu} _ {i}, \\boldsymbol {\\Sigma} _ {i}\\right), \\tag {3}", + "image_path": "49a348ab25605d26f1fd0923da22611396e497df3f8bf0837f13d0d517cb3eab.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "-th mixture weight, " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_{i}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_{i}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " represent the mean vector and the covariance matrix of the " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": "-th Gaussian component, respectively. This parametrization can be straightforwardly done by training a density estimation network for parameter estimation. As illustrated in Figure 2, this network is inside the scanpath generation unit (SGU) and is composed of two subnetworks to process the historical path context " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "s_{T}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": " and the causal path context " + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 304, + 546, + 546, + 713 + ], + "type": "text", + "content": ", which we denote by H-Net and C-Net, respectively. The concatenated features are fed to three prediction heads to estimate the weight vector, the mean vectors, and the covariance matrices of the GMM, respectively. We find empirically that incorporating the historical video frames as the visual context significantly increases computational demands with" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2601" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "only slight improvements in performance. Therefore, to keep the scanpath generator lightweight, we choose to omit the visual context. The detailed specifications of the density estimation network can be found in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "text", + "content": "Estimating continuous probability density is generally difficult and may lead to overfitting. In particular, maximum likelihood estimation of the GMM parameters through direct optimization of Eq. (3) is challenging, due to the presence of singularities [2]. To circumvent this, we compute the probability mass " + }, + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "inline_equation", + "content": "P\\left(\\bar{\\boldsymbol{r}}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)" + }, + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "text", + "content": " by discretizing and integrating the density " + }, + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "inline_equation", + "content": "p\\left(\\boldsymbol{r}_t \\mid \\boldsymbol{s}_T, \\boldsymbol{c}_t\\right)" + }, + { + "bbox": [ + 46, + 133, + 286, + 229 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 92, + 234, + 286, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 234, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 92, + 234, + 286, + 266 + ], + "type": "interline_equation", + "content": "P \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) = \\int_ {\\Omega} p \\left(\\bar {\\boldsymbol {r}} _ {t} \\mid \\boldsymbol {s} _ {T}, \\boldsymbol {c} _ {t}\\right) d \\Omega . \\tag {4}", + "image_path": "9b388ac0f960127f287294d728e0ede06ac0f108cea76364a5003a14f6dd8bac.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{r}}_t" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": " represents the quantized value of " + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_t" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": " by a uniform quantizer with a step size of " + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 47, + 270, + 286, + 294 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 299, + 286, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 299, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 110, + 299, + 286, + 326 + ], + "type": "interline_equation", + "content": "\\bar {\\xi} = Q (\\xi) = \\Delta \\left\\lfloor \\frac {\\xi}{\\Delta} + \\frac {1}{2} \\right\\rfloor , \\tag {5}", + "image_path": "a038086374b228bc1fd60ed838231fef9845497cfdeec3392ee5877e457e66b2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "text", + "content": " denotes the floor function. " + }, + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "inline_equation", + "content": "\\Omega = [\\bar{\\phi}_{T + t} - 1 / 2\\Delta, \\bar{\\phi}_{T + t} + 1 / 2\\Delta] \\times [\\bar{\\theta}_{T + t} - 1 / 2\\Delta, \\bar{\\theta}_{T + t} + 1 / 2\\Delta]" + }, + { + "bbox": [ + 47, + 330, + 286, + 390 + ], + "type": "text", + "content": " is the integration interval. As pointed out in [16], the incorporation of quantization establishes the equivalence between scanpath generation and lossy scanpath compression." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": "Furthermore, the absolute Euler coordinate system is not user-centric, meaning that it is not centered at the user's current viewpoint, relative to historical and future viewpoints [16]. This may complicate the probabilistic modeling of scanpaths and the end-to-end optimization of blind PVQA. To address this, we convert the Euler coordinates to the relative " + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 46, + 390, + 286, + 473 + ], + "type": "text", + "content": " coordinates:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 84, + 480, + 286, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 480, + 286, + 493 + ], + "spans": [ + { + "bbox": [ + 84, + 480, + 286, + 493 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {s}} _ {T - t} = \\boldsymbol {\\Psi} _ {T - t} (\\boldsymbol {s} _ {T}), \\text {f o r} t \\in \\{1, \\dots , H \\}, \\tag {6}", + "image_path": "f241b770c9da5610cef84e4daa3ee28cad25a00e8f2a9d1b892f93b219b4306d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\Psi_{T - t}(\\cdot)" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " denotes the mapping of " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "s_T" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " to the viewpoint centered at the reference viewpoint " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "(\\phi_{T - t},\\theta_{T - t})" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": ". By choosing each viewpoint in " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "s_T" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " as the reference, we create " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " relative scanpaths out of " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "s_T" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " (see Figure 3), which serve as input to the H-Net. Meanwhile, we map the causal scanpath context " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " and the viewpoint to be predicted " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "(\\phi_{T + t},\\theta_{T + t})" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " to the last historical viewpoint centered at " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "(\\phi_{T - 1},\\theta_{T - 1})" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": ". We stack " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " SGUs to form a scanpath generation block (SGB), which takes " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "s_T" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " as input and predicts the future scanpath " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": ". Furthermore, we stack " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " SGBs to form the scanpath generator, which is capable of predicting a very long-term scanpath of length " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "M\\times W + H" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": " (including the initial length " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": "). The parameters of different SGUs are shared to enable variable-length scanpath generation by varying " + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 498, + 286, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 286, + 713 + ], + "type": "text", + "content": "Differentiable Scanpath Sampling. To enable end-to-end optimization of the proposed blind PVQA method, we propose a two-step differentiable sampling method to draw" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 67, + 521, + 164 + ], + "blocks": [ + { + "bbox": [ + 331, + 67, + 521, + 164 + ], + "lines": [ + { + "bbox": [ + 331, + 67, + 521, + 164 + ], + "spans": [ + { + "bbox": [ + 331, + 67, + 521, + 164 + ], + "type": "image", + "image_path": "f180e081f6730f63d7e3aa992ba4370010d3b799f50acdea1a0bdf5df94e3a34.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 175, + 545, + 197 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 197 + ], + "type": "text", + "content": "Figure 3. Visualization of a relative scanpath projected from the sphere to the viewpoint." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 206, + 545, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 545, + 254 + ], + "type": "text", + "content": "viewpoints from the estimated GMM via the reparameterization trick [12, 13]. The first step is to select a Gaussian component from which to sample the viewpoint, according to the categorical distribution:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 261, + 545, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 261, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 339, + 261, + 545, + 294 + ], + "type": "interline_equation", + "content": "\\boldsymbol {e} = \\text {o n e} _ {\\text {h o t}} \\left(\\underset {i \\in \\{1, \\dots K \\}} {\\arg \\max } \\left(\\log \\left(\\alpha_ {i}\\right) + g _ {i}\\right)\\right), \\tag {7}", + "image_path": "7730670a99ec06b652c7768621a40ef17695e7dccb372d18ec32baa4ef9c5c01.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "text", + "content": " is a sample drawn from the Gumbel(0,1) distribution, and " + }, + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 305, + 301, + 545, + 350 + ], + "type": "text", + "content": " is a one-hot vector. Eq. (7) is known as the Gumbel-Max trick [8], which is non-differentiable. We relax the arg max operator with a softmax function [12]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 357, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 357, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 359, + 357, + 545, + 371 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {e}} = \\operatorname {s o f t m a x} \\left(\\left(\\log (\\boldsymbol {\\alpha}) + \\boldsymbol {g}\\right) / \\tau\\right), \\tag {8}", + "image_path": "558d57be445c688f7a473e8a1fe51a7f89cd49a1713265cae5574c33a331ab2f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": " represents the temperature coefficient, " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\alpha = [\\alpha_{1},\\dots,\\alpha_{K}]" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\pmb {g} = [g_1,\\dots,g_K]" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": " approaches zero, " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\hat{e}" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": " converges to " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": ". In the forward pass, " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "\\arg \\max" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": " is used directly, while in the backward pass, it is replaced by the softmax function. The second step involves sampling a viewpoint from the selected Gaussian component. Assuming the " + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 379, + 545, + 475 + ], + "type": "text", + "content": "-th Gaussian component is selected, the linear reparameterization trick [13] suggests" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 483, + 545, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 483, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 392, + 483, + 545, + 496 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {L} _ {i} \\epsilon , \\tag {9}", + "image_path": "2587aa02ad0da540a09f78107a6a836bf382f1ce2b6908ff567a6239f1a62775.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{r}}_t" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": " is the relative " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": " coordinates of the " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": "-th future viewpoint. " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}_i = \\boldsymbol{L}_i\\boldsymbol{L}_i^{\\mathrm{T}}" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": " is the Cholesky decomposition, and " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": " is a sample drawn from the " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": ". Differentiation of the Cholesky decomposition is complicated and sometimes numerically unstable [26]. Thus we assume the independence between the " + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 305, + 504, + 545, + 588 + ], + "type": "text", + "content": " coordinates, leading to the simplified reparameterization formula:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 597, + 545, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 597, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 386, + 597, + 545, + 609 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {r}} _ {t} = \\boldsymbol {u} _ {i} + \\boldsymbol {\\sigma} _ {i} \\odot \\boldsymbol {\\epsilon}, \\tag {10}", + "image_path": "d7cc2644d77c9404f9cd43128014c8acd975d3f7fab57e09cca6c452578ee1a6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " denotes the element-wise product, and " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " denotes the standard deviations of the " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "-th Gaussian component. Through the two-step reparameterization, our sampling strategy ensures effective back-propagation. As suggested in [16], we additionally implement a proportional-integral-derivative (PID) controller [1] to further improve the smoothness of the sampled scanpaths (see the details in the supplementary material)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2602" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": "Differentiable Viewport Sequence Generation. Inspired by [11], we first parameterize the Euler sampling grid in terms of the relative " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " coordinates via the inverse transformation " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "\\Psi^{-1}" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " (see Eq. (6)). Subsequently, the Euler coordinate " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "(\\phi, \\theta)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " is mapped to the discrete sampling position " + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "(m, n)" + }, + { + "bbox": [ + 47, + 72, + 289, + 145 + ], + "type": "text", + "content": " in the ERP domain:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 152, + 287, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 152, + 287, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 152, + 287, + 165 + ], + "type": "interline_equation", + "content": "m = (0. 5 - \\phi / \\pi) H _ {e} - 0. 5, \\tag {11}", + "image_path": "869df7e224ce0ce68ef682495cc7e67cd639110d72809b4e62afd0a31e3907d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 167, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 167, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 110, + 167, + 287, + 179 + ], + "type": "interline_equation", + "content": "n = (\\theta / 2 \\pi + 0. 5) W _ {e} - 0. 5, \\tag {12}", + "image_path": "5c7ea217a2f8c0026988cea4224a118d6aeff4046aa00dba65aa562dd4eb9e56.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "H_{e}" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "W_{e}" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " are the height and width of the video frame in ERP. Once the mapping between " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "(m, n)" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " is established, we construct a flow field that is the same size as the viewpoint. Within this field, each element records the corresponding pixel position " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "(m, n)" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": ". Given an ERP image and the flow field, we apply bilinear interpolation [11] to compute pixel values in the viewpoint and leverage its sub-gradients for back-propagation. This process yields " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " viewpoint sequences, corresponding to " + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 187, + 288, + 297 + ], + "type": "text", + "content": " initial paths." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 301, + 149, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 301, + 149, + 315 + ], + "spans": [ + { + "bbox": [ + 47, + 301, + 149, + 315 + ], + "type": "text", + "content": "3.2. Quality Assessor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "text", + "content": "Our probabilistic scanpath generator can work with any planar VQA model, whether it is differentiable or not. To enable end-to-end optimization of the scanpath generator and the quality assessor, and to make a fair comparison with existing blind PVQA models, we reuse three differentiable quality assessors from ScanpathVQA [35], GSR-S / GSR-X [29], and Assessor360 [36]. Specifically, the quality assessor of ScanpathVQA is a lightweight ResNet-18 network [9], with the classification head replaced by a quality estimator (a multilayer perceptron). The quality assessors of GSR-S / GSR-X are adapted from Video Swin-T [18] / X-Clip-B/32 [22]. The quality assessor of Assessor360 is modified from Swin-B [17] with the addition of a temporal analysis module. We feed each of the " + }, + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "text", + "content": " viewpoint sequences to the quality assessor to compute " + }, + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 319, + 289, + 523 + ], + "type": "text", + "content": " quality scores. The overall quality estimate is then computed by a simple average:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 135, + 530, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 530, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 135, + 530, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\hat {q} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\hat {q} _ {i}. \\tag {13}", + "image_path": "f71ffefc6c542b0bee3e70b8c8889b58652c3cea649229e0416ac55a94631207.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 570, + 175, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 175, + 583 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 175, + 583 + ], + "type": "text", + "content": "3.3. Optimization Strategy" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 588, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 288, + 649 + ], + "type": "text", + "content": "We explore a three-stage training procedure for our blind PVQA model. In the first stage, we pre-train the density estimation network on the VRVQW dataset [35] by minimizing the expected code length of the generated scanpaths (also equivalent to minimizing the negative log-likelihood):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 665, + 287, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 287, + 710 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {c o d e}} = - \\frac {1}{B W} \\sum_ {i = 1} ^ {B} \\sum_ {t = 0} ^ {W - 1} \\log_ {2} \\left(P \\left(\\bar {\\boldsymbol {r}} _ {t} ^ {(i)} \\mid \\boldsymbol {s} ^ {(i)}, \\boldsymbol {c} _ {t} ^ {(i)}\\right)\\right), \\tag {14}", + "image_path": "7bc7bf6958a8ae81c0c8ea5b6a0ae2813af4b99696ad661f1d9656ffdef02991.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 304, + 72, + 547, + 205 + ], + "type": "text", + "content": " denotes the mini-batch size. During this stage of training, we use human scanpaths to fill in the causal path context, which can be efficiently implemented by a causal masking mechanism. In the second stage, we fix the parameters of the pre-trained scanpath generator and warm up the quality assessor by optimizing the Pearson linear correlation coefficient (PLCC) between human perceptual scores and model predictions. In the third stage, we end-to-end finetune the entire method. We find that the proposed three-stage optimization strategy accelerates convergence compared to the naive end-to-end optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 213, + 388, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 388, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 388, + 228 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 319 + ], + "type": "text", + "content": "In this section, we first delineate the experimental setups, and then compare our method with current blind PVQA models under both in-dataset and cross-dataset settings. We further validate our scanpath generator in terms of explaining human perceptual scores and replicating human viewing patterns. Lastly, we conduct a series of ablation experiments to probe the impact of several key designs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 324, + 427, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 427, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 427, + 338 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 486 + ], + "type": "text", + "content": "Datasets. We employ three panoramic image and video datasets: VRVQW [35], CVIQD [30], and OIQA [6]. The VRVQW dataset includes 502 panoramic videos that have a wide spectrum of authentic distortions. Each video is viewed under four unique viewing conditions to simulate the different quality of experience during the initial viewing. The CVIQD dataset comprises a total of 528 compressed panoramic images by JPEG, AVC, and HEVC, from 16 reference images. The OIQA dataset includes 320 panoramic images that have been altered from 16 reference images by JPEG compression, JPEG2000 compression, Gaussian blurring, and Gaussian noise contamination." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Implementation Details. For the scanpath generator, we set the length of the provided initial path " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " and the predicted future path " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " in the SGB to be identical and equal to 5. The number of Gaussian components " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (3) is set to 3. The quantization step size " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (5) is set to 0.2. The temperature coefficient " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (8) is set to 1. The number of stacked SGBs " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is set to 6 and 14 for the viewing duration of 7 and 15 seconds, respectively. For the quality assessor, the number of scanpaths " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " in Eq. (13) is set to 20. The input viewpoint size " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "H_{v} \\times W_{v}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": ", corresponding to a field of view of " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "90^{\\circ} \\times 90^{\\circ}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": ". The length of the viewpoint sequence is set to 7 regardless of the duration and frame rate of the original panoramic video. We split VRVQW randomly into the training, validation, and test sets according to the ratio of " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "6:2:2" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " for 5 times, and report the mean results. Similarly, we split CVIQD and OIQA using a different ratio of " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "7:1:2" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " for 5 times. The detailed configuration of our three-stage optimization strategy is detailed in the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2603" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 145, + 272, + 467 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 138 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 138 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 138 + ], + "type": "text", + "content": "Table 1. In-dataset comparison of blind PVQA methods on three panoramic image and video quality datasets. SRCC: Spearman's rank correlation coefficient. PLCC: Pearson linear correlation coefficient. The same evaluation metrics are applied in Tables 2, 3, 5, 6 and Figure 6. The best results on each dataset are highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 145, + 272, + 467 + ], + "lines": [ + { + "bbox": [ + 65, + 145, + 272, + 467 + ], + "spans": [ + { + "bbox": [ + 65, + 145, + 272, + 467 + ], + "type": "table", + "html": "
DatasetMethodSRCCPLCC
VRVQW [35]NIQE [19]0.4010.365
MC360IQA [31]0.6690.671
Wen24 [35]0.7560.763
ScanpathVQA [35]0.7790.781
Assessor360 [36]0.4060.415
Ours (ScanpathVQA)0.8010.809
Ours (GSR-S)0.8040.807
Ours (GSR-X)0.8150.819
Ours (Assessor360)0.8220.823
CVIQD [30]NIQE [19]0.8470.878
MC360IQA [31]0.9170.939
Wen24 [35]0.9190.932
GSR-S [29]0.9050.937
GSR-X [29]0.9440.962
Assessor360 [36]0.9550.969
Ours (ScanpathVQA)0.9120.936
Ours (GSR-S)0.9300.958
Ours (GSR-X)0.9560.974
Ours (Assessor360)0.9720.983
OIQA [6]NIQE [19]0.7020.657
MC360IQA [31]0.9000.906
Wen24 [35]0.9050.907
GSR-S [29]0.9020.915
GSR-X [29]0.9450.954
Assessor360 [36]0.9460.953
Ours (ScanpathVQA)0.9150.922
Ours (GSR-S)0.9270.936
Ours (GSR-X)0.9560.967
Ours (Assessor360)0.9600.971
", + "image_path": "f152c502592d28a36bd95dc4e5e5e1c22f786d2d9462338de29530b6958c58a6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 72, + 512, + 261, + 644 + ], + "blocks": [ + { + "bbox": [ + 46, + 475, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 510 + ], + "type": "text", + "content": "Table 2. Cross-dataset comparison of blind PVQA methods on the CVIQD [30] and OIQA [6] datasets. The arrow points from the training set to the test set." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 512, + 261, + 644 + ], + "lines": [ + { + "bbox": [ + 72, + 512, + 261, + 644 + ], + "spans": [ + { + "bbox": [ + 72, + 512, + 261, + 644 + ], + "type": "table", + "html": "
DatasetMethodSRCCPLCC
OIQAMC360IQA0.7980.842
GSR-X0.7620.841
Assessor3600.8590.893
CVIQDOurs (ScanpathVQA)0.7330.747
Ours (Assessor360)0.8720.904
CVIQDMC360IQA0.2880.349
GSR-X0.6950.718
Assessor3600.3380.467
OIQAOurs (ScanpathVQA)0.6360.658
Ours (Assessor360)0.7030.715
", + "image_path": "9283e8d4e44e662c2805909824640f32acffbd6d5f5424f4ba78878bf1d43971.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 133, + 669 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We compare our blind PVQA method with seven existing models, including NIQE [19], MC360IQA [31], Wen24 [35], ScanpathVQA [35], GSR-S [29], GSR-X [29]," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 325, + 96, + 525, + 289 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 3. Comparison of different scanpath generators for explaining human perceptual scores." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 96, + 525, + 289 + ], + "lines": [ + { + "bbox": [ + 325, + 96, + 525, + 289 + ], + "spans": [ + { + "bbox": [ + 325, + 96, + 525, + 289 + ], + "type": "table", + "html": "
DatasetMethodSRCCPLCC
VRVQWHuman scanpath0.7860.790
Random sampling0.0750.104
Heuristic sampling [36]0.4310.443
Xu18 [38]0.7120.717
TRACK [25]0.7450.749
Li23 [16]0.7900.794
Ours0.8050.814
CVIQDRandom sampling0.6320.640
Heuristic sampling [36]0.8680.870
ScanDMM [28]0.8560.864
Li23 [16]0.8140.827
Ours0.9280.940
OIQARandom sampling0.5140.536
Heuristic sampling [36]0.8610.872
ScanDMM [28]0.8650.877
Li23 [16]0.7930.799
Ours0.9140.917
", + "image_path": "026be187e8de684a0dc88f86de2e14b35d342fdbe43c92a999249945ebb6818d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 384 + ], + "type": "text", + "content": "and Assessor360 [36]. For image quality models such as MC360IQA [31] and Assessor360 [36], we retrain them on the VRVQW dataset. In the case of MC360IQA, we assign the video-level quality score to each key frame and use their temporally averaged score for testing. Assessor360's scanpath generator is adapted to videos by adjusting the semantic context associated with each key frame." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 384, + 546, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 658 + ], + "type": "text", + "content": "In-dataset Results. Table 1 shows the Spearman's rank correlation coefficient (SRCC) and " + }, + { + "bbox": [ + 304, + 384, + 546, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{PLCC}^1" + }, + { + "bbox": [ + 304, + 384, + 546, + 658 + ], + "type": "text", + "content": " results under the in-dataset setting. It is evident that our learned scanpaths enhance the performance of all quality assessors compared to other scanpath-based methods. When integrated with the quality assessor from Assessor360 (i.e., a modified SwinB with a temporal analysis module), our proposed method achieves the best results on all three datasets. Furthermore, our scanpath generator can boost a simpler quality assessor (e.g., from ScanpathVQA [35] with approximately 11 million parameters) to reach performance levels similar to those of a more sophisticated quality assessor coupled with a weaker scanpath generator (e.g., GSR-S [29] with approximately 112 million parameters). Methods that overlook human viewing patterns, like NIQE [19] and MC360IQA [31], fail to accurately model the human perception of panoramic image and video quality, especially on VRVQW. Additionally, we find that assessing the quality of panoramic videos with authentic distortions tends to be more challenging than for panoramic images with synthetic distortions. This is anticipated because authentic distortions in panoramic videos typically present as a complex blend of various artifacts, localized in space and time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 659, + 547, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 547, + 684 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 547, + 684 + ], + "type": "text", + "content": "Cross-dataset Results. Table 2 shows the SRCC and PLCC results under the cross-dataset settings on CVIQD [30] and" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "1As standard practice, we apply a monotonic logistic function to compensate for the nonlinearity in model predictions before computing PLCC." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2604" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 278, + 192 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 104 + ], + "type": "text", + "content": "Table 4. Comparison of different scanpath generators for replicating human viewing patterns using the minimum orthodromic distance (minOD) and maximum temporal correlation (maxTC)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 278, + 192 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 278, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 278, + 192 + ], + "type": "table", + "html": "
MethodminOD ↓maxTC ↑
Heuristic sampling [36]1.3250.401
Xu18 [38]1.1850.637
TRACK [25]1.0670.699
Li23 [16]0.5420.796
Ours (w/o end-to-end optimization)0.5560.781
Ours0.5360.805
", + "image_path": "b751be8dd20d7ae0490c8cc01ac6bca8d388dcd7a2da5bdc7ee8643680c4e67e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 83, + 205, + 252, + 334 + ], + "blocks": [ + { + "bbox": [ + 83, + 205, + 252, + 334 + ], + "lines": [ + { + "bbox": [ + 83, + 205, + 252, + 334 + ], + "spans": [ + { + "bbox": [ + 83, + 205, + 252, + 334 + ], + "type": "image", + "image_path": "72c6a5319bac4183d4a9c5b93a002c2f646bc3aafcfbcdd34a916aaa792d3311.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 338, + 289, + 361 + ], + "lines": [ + { + "bbox": [ + 47, + 338, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 289, + 361 + ], + "type": "text", + "content": "Figure 4. Comparison of different scanpath predictors in terms of maxTC with different prediction horizons." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 372, + 289, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 482 + ], + "type": "text", + "content": "OIQA [6]. Generally, models trained on the OIQA dataset show better generalizability. This is likely because OIQA encompasses a broader range of distortion types, compared to CVIQD, which only includes the compression artifacts. Assessor360 shows a noticeable performance drop when tested on OIQA, potentially indicative of overfitting. Our scanpath generator is capable of restoring Assessor360's performance, which provides a strong indication of its effectiveness through end-to-end optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 490, + 217, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 217, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 217, + 502 + ], + "type": "text", + "content": "4.3. Scanpath Generator Validation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 664 + ], + "type": "text", + "content": "Explaining Human Perceptual Scores. We conduct an apple-to/apple comparison of different scanpath generators in terms of explaining human perceptual scores by fixing the quality assessor to that used in ScanpathVQA. These include random sampling, heuristic sampling [36], Xu18 [38], TRACK [25], ScanDMM [28], Li23 [16], and our method. Table 3 shows the SRCC and PLCC results on the VRVQW, CVIQD and OIQA datasets. It is noteworthy that our scanpath generator outperforms all competing methods across all three datasets, even surpassing the human-level performance on VRVQW. The performance of random and heuristic sampling decreases sharply on VRVQW, due to the presence of spatiotemporally localized authentic distortions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "Replicating Human Viewing Patterns. We also test different scanpath generators in terms of replicating human viewing patterns on the VRVQW dataset by comparing the predicted scanpaths to those of humans. We use two set-to-set" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 543, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 543, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 543, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 543, + 175 + ], + "type": "image", + "image_path": "5fc3e3debab2a7957e0a7b3088cf52690a9fb15f4afbc1aa1924bd9d8c941d2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "lines": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 305, + 178, + 545, + 201 + ], + "type": "text", + "content": "Figure 5. Comparison of saliency maps generated from scanpaths by our method and those by humans." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 249, + 552, + 309 + ], + "blocks": [ + { + "bbox": [ + 305, + 213, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 547, + 246 + ], + "type": "text", + "content": "Table 5. Impact of optimization strategies on blind PVQA. Training protocol: 1) Two-stage w/ fixed scanpaths, 2) Two-stage w/ varied scanpaths, and 3) Three-stage w/ end-to-end optimization." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 249, + 552, + 309 + ], + "lines": [ + { + "bbox": [ + 307, + 249, + 552, + 309 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 552, + 309 + ], + "type": "table", + "html": "
ProtocolVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
10.7690.7720.7100.7800.6880.742
20.7810.7850.8300.8590.7980.815
30.8050.8140.9280.9400.9140.917
", + "image_path": "3f7c5a1af1bb8b8efbe1299b950feaf24a54c69fa49b2df2f595455e37ccc029.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "text", + "content": "evaluation metrics: the minimum orthodromic distance (i.e., minOD) and maximum temporal correlation (i.e., maxTC), as suggested in [16]. Given a set of human scanpaths, " + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "inline_equation", + "content": "S = \\{s^{(i)}\\}_{i=1}^{|S|}" + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "text", + "content": ", the minimum orthodromic distance between " + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "text", + "content": " and the set of predicted scanpaths " + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\{\\hat{s}^{(i)}\\}_{i=1}^{|S|}" + }, + { + "bbox": [ + 305, + 320, + 545, + 398 + ], + "type": "text", + "content": " can be computed by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 346, + 406, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 406, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 346, + 406, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\min \\mathrm {O D} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\min _ {\\boldsymbol {s} \\in \\mathcal {S}, \\boldsymbol {s} \\in \\hat {\\mathcal {S}}} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {15}", + "image_path": "3779488e8b31701769dbec53979f63ec1d9099d0efec0ef6852ac8b7e80ce69d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 439, + 523, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 523, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 523, + 452 + ], + "type": "text", + "content": "where the orthodromic distance " + }, + { + "bbox": [ + 306, + 439, + 523, + 452 + ], + "type": "inline_equation", + "content": "\\mathrm{OD}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 306, + 439, + 523, + 452 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 461, + 545, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 545, + 528 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {O D} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}) = \\frac {1}{T} \\sum_ {t = 0} ^ {T - 1} \\operatorname {a r c c o s} \\left(\\cos \\left(\\phi_ {t}\\right) \\cos \\left(\\hat {\\phi} _ {t}\\right) \\cos \\left(\\theta_ {t} - \\hat {\\theta} _ {t}\\right) \\right. \\\\ + \\sin \\left(\\phi_ {t}\\right) \\sin \\left(\\hat {\\phi} _ {t}\\right)). \\tag {16} \\\\ \\end{array}", + "image_path": "aea70d04dc1dce833acb597a2e8d313ef292dca245498ab2e592a2701a12a3b8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "text", + "content": "The maximum temporal correlation between " + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{S}}" + }, + { + "bbox": [ + 305, + 539, + 545, + 563 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 571, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 571, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 348, + 571, + 545, + 594 + ], + "type": "interline_equation", + "content": "\\max \\mathrm {T C} (\\mathcal {S}, \\hat {\\mathcal {S}}) = \\max _ {\\boldsymbol {s} \\in \\mathcal {S}, \\hat {\\boldsymbol {s}} \\in \\hat {\\mathcal {S}}} \\mathrm {T C} (\\boldsymbol {s}, \\hat {\\boldsymbol {s}}), \\tag {17}", + "image_path": "7bbb5b604e0448d6ec779474ba1f21af344aaf1ff1bcc9a8a5eb7c85ff4e46a6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 604, + 495, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 495, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 495, + 616 + ], + "type": "text", + "content": "where the temporal correlation is computed by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 624, + 545, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 624, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 318, + 624, + 545, + 671 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {T C} \\left(\\boldsymbol {s} ^ {(i)}, \\boldsymbol {s} ^ {(j)}\\right) = \\\\ \\frac {1}{2} \\left(\\operatorname {P L C C} \\left(\\phi^ {(i)}, \\phi^ {(j)}\\right) + \\operatorname {P L C C} \\left(\\boldsymbol {\\theta} ^ {(i)}, \\boldsymbol {\\theta} ^ {(j)}\\right)\\right). \\end{array} \\tag {18}", + "image_path": "e9620f8a610e423b8f9d3f9eb630d26f2f8c87274e55a79651d2485f1c6fb5b9.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Table 4 presents the minOD and maxTC results, from which we find that our end-to-end optimized scanpath generator delivers the best results, surpassing its independently" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2605" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 124, + 85, + 467, + 136 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 497, + 82 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 497, + 82 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 497, + 82 + ], + "type": "text", + "content": "Table 6. Impact of visual context on blind PVQA. #Parameters added to the scanpath generator are also shown." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 124, + 85, + 467, + 136 + ], + "lines": [ + { + "bbox": [ + 124, + 85, + 467, + 136 + ], + "spans": [ + { + "bbox": [ + 124, + 85, + 467, + 136 + ], + "type": "table", + "html": "
Method#ParametersVRVQWCVIQDOIQA
SRCCPLCCSRCCPLCCSRCCPLCC
Ours1M0.8050.8140.9280.9400.9140.917
Ours w/ visual context27M0.8160.8230.9370.9560.9250.934
", + "image_path": "6113d7a9c53f552e9e3567a67957980a972a1f1de90693679c2332c493cf6afc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 150, + 288, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 150, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 150, + 288, + 330 + ], + "type": "text", + "content": "optimized counterpart by a clear margin. The heuristic sampling [36] that depends on the simplified entropy features and equator bias, is inadequate for capturing human viewing patterns, especially for long-term prediction horizons (see Figure 4). Due to the deterministic nature, Xu18 [38] and TRACK [25] fail to accommodate the diversity and uncertainty inherent in human scanpaths, resulting in subpar performance. Incorporating historical video frames as the visual context, Li23 [16] shows performance on par with our method, reinforcing our assertion that visual context informs less about future viewpoints. Figure 5 demonstrates a comparison of the saliency maps derived from scanpaths by our method and those by humans, offering further proof of the close alignment of our scanpath generator and human viewing behaviors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 336, + 148, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 148, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 148, + 348 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 509 + ], + "type": "text", + "content": "Impact of Optimization Strategies. We explore three different optimization strategies: 1) a two-stage approach where the pre-trained scanpath generator produces a fixed set of scanpaths for the training of the quality assessor, 2) a similar two-stage approach but supplying a varied set of scanpaths in each epoch of training, and 3) the default three-stage approach that enables end-to-end optimization. From the results in Table 5, we find that the two-stage approach benefits from \"data augmentation\" with varied scanpaths in each epoch. Our three-stage end-to-end optimization strategy further boosts the accuracy of quality prediction by jointly finetuning both the scanpath generator and quality assessor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 617 + ], + "type": "text", + "content": "Impact of Visual Context. To assess the impact of visual context on blind PVQA, we enhance our scanpath generator with a video analysis network [16], implemented by a variant of ResNet-50 for frame-level feature extraction and aggregation. We subsequently integrate it with the ScanpathVQA quality assessor for blind PVQA, with the results shown in Table 6. We find that the visual context has a negligible effect on blind PVQA, and thus we exclude it in the generation of scanpaths." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Impact of the Number and Length of Viewport Sequences. We investigate the effects of varying the number " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and length " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " of viewpoint sequences on blind PVQA. We test " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " values from " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{5,10,15,20,50\\}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " values from " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{4,7,15\\}" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": ". Figure 6 shows the SRCC results on VRVQW, using the ScanpathVQA quality assessor for prediction. It is clear that " + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": " viewpoint sequences are sufficient for reliable quality assessment, with performance remaining stable" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 340, + 149, + 511, + 281 + ], + "blocks": [ + { + "bbox": [ + 340, + 149, + 511, + 281 + ], + "lines": [ + { + "bbox": [ + 340, + 149, + 511, + 281 + ], + "spans": [ + { + "bbox": [ + 340, + 149, + 511, + 281 + ], + "type": "image", + "image_path": "b3abb9f809bf7a66d4ab2271e69ac1348ec74016d2cabd6e9deb12476f2bcad2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "text", + "content": "Figure 6. Impact of the number " + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "text", + "content": " and length " + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 288, + 545, + 310 + ], + "type": "text", + "content": " of viewpoint sequences on blind PVQA." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "text", + "content": "with the increase in " + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "text", + "content": ". For sequence length, " + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "inline_equation", + "content": "L = 7" + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "text", + "content": " appears to be a wise choice. Further increasing " + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 322, + 545, + 429 + ], + "type": "text", + "content": " does not noticeably affect the performance, but would lead to a considerable rise in computational demand. Conversely, a shorter viewpoint sequence results in a noticeable drop in performance due to the loss of information from excessive temporal downsampling. It is important to note that these findings are specific to the ScanpathVQA quality assessor and may differ from other assessors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 379, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 596 + ], + "type": "text", + "content": "We have introduced an end-to-end optimized blind PVQA method, consisting of a scanpath generator and a quality assessor. The proposed scanpath generator is differentiable and can be integrated with any planar VQA model, whose effectiveness has been thoroughly validated in supporting blind PVQA and in modeling human viewing patterns. Additionally, we have also devised a three-stage optimization strategy to facilitate training convergence, which aligns with current large-scale optimization practices that involve self-supervised pre-training followed by supervised finetuning, including initial warmup phases [10]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 416, + 623 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (62071407 and 62102339), the Shenzhen Science and Technology Program (RCBS20221008093121052), the Research Grants Council of Hong Kong (ECS 2121382, ECS 27212822, and GRF 17201822), and the CCF-Tencent Rhino Bird Fund (9239061)." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2606" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Kiam Heong Ang, Gregory Chong, and Yun Li. PID control system analysis, design, and technology. IEEE Transactions on Control Systems Technology, 13(4):559-576, 2005. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 146 + ], + "type": "text", + "content": "[2] Christopher M. Bishop. Pattern Recognition and Machine Learning. Springer, 2006. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "text", + "content": "[3] Fang-Yi Chao, Cagri Ozcinar, and Aljosa Smolic. Transformer-based long-term viewpoint prediction in " + }, + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 53, + 148, + 288, + 201 + ], + "type": "text", + "content": " video: Scanpath is all you need. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[4] Sijia Chen, Yingxue Zhang, Yiming Li, Zhenzhong Chen, and Zhou Wang. Spherical structural similarity index for objective omnidirectional video quality assessment. In IEEE International Conference on Multimedia and Expo, pages 1-6, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "text", + "content": "[5] Taco S. Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, pages 1-15, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 292, + 288, + 335 + ], + "type": "text", + "content": "[6] Huiyu Duan, Guangtao Zhai, Xiongkuo Min, Yucheng Zhu, Yi Fang, and Xiaokang Yang. Perceptual quality assessment of omnidirectional images. In IEEE International Symposium on Circuits and Systems, pages 1-5, 2018. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 379 + ], + "type": "text", + "content": "[7] Jun Fu, Chen Hou, Wei Zhou, Jiahua Xu, and Zhibo Chen. Adaptive hypergraph convolutional network for no-reference 360-degree image quality assessment. In ACM International Conference on Multimedia, pages 961-969, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 412 + ], + "type": "text", + "content": "[8] Emil J. Gumbel. Statistical Theory of Extreme Values and Some Practical Applications. US Government Printing Office, 1948. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 456 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 288, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 288, + 502 + ], + "type": "text", + "content": "[10] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In IEEE Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 288, + 545 + ], + "type": "text", + "content": "[11] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems, pages 2017-2025, 2015. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 288, + 590 + ], + "type": "text", + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with Gumbel-Softmax. In International Conference on Learning Representations, pages 1–12, 2017. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 624 + ], + "type": "text", + "content": "[13] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In International Conference on Learning Representations, pages 1–14, 2014. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[14] Chen Li, Mai Xu, Lai Jiang, Shanyi Zhang, and Xiaoming Tao. Viewport proposal CNN for " + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": " video quality assessment. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10169-10178, 2019. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 713 + ], + "type": "text", + "content": "[15] Chenge Li, Weixi Zhang, Yong Liu, and Yao Wang. Very long term field of view prediction for 360-degree video streaming. In IEEE Conference on Multimedia Information Processing and Retrieval, pages 297-302, 2019. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[16] Mu Li, Kanglong Fan, and Kede Ma. Scanpath prediction in panoramic videos via expected code length minimization. arXiv preprint arXiv:2305.02536, 2023. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In IEEE International Conference on Computer Vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 207 + ], + "type": "text", + "content": "[18] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video Swin Transformer. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 208, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 241 + ], + "type": "text", + "content": "[19] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[20] King-To Ng, Shing-Chow Chan, and Heung-Yeung Shum. Data compression and transmission aspects of panoramic videos. IEEE Transactions on Circuits and Systems for Video Technology, 15(1):82-95, 2005. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 342 + ], + "type": "text", + "content": "[21] Anh Nguyen, Zhisheng Yan, and Klara Nahrstedt. Your attention is unique: Detecting 360-degree video saliency in head-mounted display for head movement prediction. In ACM International Conference on Multimedia, pages 1190-1198, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[22] Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. Expanding language-image pretrained models for general video recognition. In European Conference on Computer Vision, pages 1-18, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[23] David Noton and Lawrence Stark. Scanpaths in saccadic eye movements while viewing and recognizing patterns. Vision Research, 11(9):929-942, 1971. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 465 + ], + "type": "text", + "content": "[24] David Noton and Lawrence Stark. Scanpaths in eye movements during pattern perception. Science, 171(3968):308-311, 1971. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 467, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 533 + ], + "type": "text", + "content": "[25] Miguel Fabian Romero Rondon, Lucile Sassatelli, Ramón Aparicio-Pardo, and Frédéric Precioso. TRACK: A new method from a re-examination of deep architectures for head motion prediction in " + }, + { + "bbox": [ + 307, + 467, + 545, + 533 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 467, + 545, + 533 + ], + "type": "text", + "content": " videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5681-5699, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[26] Stephen P. Smith. Differentiation of the Cholesky algorithm. Journal of Computational and Graphical Statistics, 4(2):134-147, 1995. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 612 + ], + "type": "text", + "content": "[27] Xiangjie Sui, Kede Ma, Yiru Yao, and Yuming Fang. Perceptual quality assessment of omnidirectional images as moving camera videos. IEEE Transactions on Visualization and Computer Graphics, 28(8):3022-3034, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[28] Xiangjie Sui, Yuming Fang, Hanwei Zhu, Shiqi Wang, and Zhou Wang. ScanDMM: A deep markov model of scanpath prediction for " + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": " images. In IEEE Conference on Computer Vision and Pattern Recognition, pages 6989-6999, 2023. 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Xiangjie Sui, Hanwei Zhu, Xuelin Liu, Yuming Fang, Shiqi Wang, and Zhou Wang. Perceptual quality assessment of " + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": " images based on generative scanpath representation. arXiv preprint arXiv:2309.03472, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2607" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[30] Wei Sun, Ke Gu, Siwei Ma, Wenhan Zhu, Ning Liu, and Guangtao Zhai. A large-scale compressed 360-degree spherical image database: From subjective quality evaluation to objective model comparison. In IEEE International Workshop on Multimedia Signal Processing, pages 1-6, 2018. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 183 + ], + "type": "text", + "content": "[31] Wei Sun, Xiongkuo Min, Guangtao Zhai, Ke Gu, Huiyu Duan, and Siwei Ma. MC360IQA: A multi-channel CNN for blind 360-degree image quality assessment. IEEE Journal of Selected Topics in Signal Processing, 14(1):64-77, 2020. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 217 + ], + "type": "text", + "content": "[32] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE Signal Processing Letters, 24(9):1408-1412, 2017. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 288, + 262 + ], + "type": "text", + "content": "[33] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 6000-6010, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 306 + ], + "type": "text", + "content": "[34] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 7794-7803, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "text", + "content": "[35] Wen Wen, Mu Li, Yiru Yao, Xiangjie Sui, Yabin Zhang, Long Lan, Yuming Fang, and Kede Ma. Perceptual quality assessment of virtual reality videos in the wild. IEEE Transactions on Circuits and Systems for Video Technology, to appear, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 419 + ], + "type": "text", + "content": "[36] Tianhe Wu, Shuwei Shi, Haoming Cai, Mingdeng Cao, Jing Xiao, Yinqiang Zheng, and Yujiu Yang. Assessor360: Multisession network for blind omnidirectional image quality assessment. In Advances in Neural Information Processing Systems, pages 1-22, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 48, + 420, + 288, + 474 + ], + "type": "text", + "content": "[37] Jiahua Xu, Wei Zhou, and Zhibo Chen. Blind omnidirectional image quality assessment with viewpoint oriented graph convolutional networks. IEEE Transactions on Circuits and Systems for Video Technology, 31(5):1724-1737, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 476, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 288, + 529 + ], + "type": "text", + "content": "[38] Yanyu Xu, Yanbing Dong, Junru Wu, Zhengzhong Sun, Zhiru Shi, Jingyi Yu, and Shenghua Gao. Gaze prediction in dynamic " + }, + { + "bbox": [ + 48, + 476, + 288, + 529 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 476, + 288, + 529 + ], + "type": "text", + "content": " immersive videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5333-5342, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 574 + ], + "type": "text", + "content": "[39] Jiachen Yang, Tianlin Liu, Bin Jiang, Wen Lu, and Qinggang Meng. Panoramic video quality assessment based on non-local spherical CNN. IEEE Transactions on Multimedia, 23: 797-809, 2021. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 576, + 288, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 576, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 576, + 288, + 620 + ], + "type": "text", + "content": "[40] Matt Yu, Haricharan Lakshman, and Bernd Girod. A framework to evaluate omnidirectional video coding schemes. In IEEE International Symposium on Mixed and Augmented Reality, pages 31-36, 2015. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 622, + 288, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 288, + 664 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 288, + 664 + ], + "type": "text", + "content": "[41] Vladyslav Zakharchenko, Kwang Pyo Choi, and Jeong Hoon Park. Quality metric for spherical panoramic video. In SPIE Optics and Photonics for Information Processing X, pages 57-65, 2016. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2608" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_content_list.json b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b41dfcfc9d50d1f573540f2808d5abaf704cacf2 --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_content_list.json @@ -0,0 +1,1625 @@ +[ + { + "type": "text", + "text": "Learned Trajectory Embedding for Subspace Clustering", + "text_level": 1, + "bbox": [ + 199, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaroslava Lochman", + "bbox": [ + 212, + 179, + 380, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Carl Olsson1,2", + "bbox": [ + 436, + 179, + 552, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Christopher Zach1", + "bbox": [ + 607, + 179, + 754, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Chalmers University of Technology", + "bbox": [ + 235, + 204, + 529, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Lund University", + "bbox": [ + 588, + 205, + 728, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 258, + 313, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Clustering multiple motions from observed point trajectories is a fundamental task in understanding dynamic scenes. Most motion models require multiple tracks to estimate their parameters, hence identifying clusters when multiple motions are observed is a very challenging task. This is even aggravated for high-dimensional motion models. The starting point of our work is that this high-dimensionality of motion model can actually be leveraged to our advantage as sufficiently long trajectories identify the underlying motion uniquely in practice. Consequently, we propose to learn a mapping from trajectories to embedding vectors that represent the generating motion. The obtained trajectory embeddings are useful for clustering multiple observed motions, but are also trained to contain sufficient information to recover the parameters of the underlying motion by utilizing a geometric loss. We therefore are able to use only weak supervision from given motion segmentation to train this mapping. The entire algorithm consisting of trajectory embedding, clustering and motion parameter estimation is highly efficient. We conduct experiments on the Hopkins155, Hopkins12, and KT3DMoSeg datasets and show state-of-the-art performance of our proposed method for trajectory-based motion segmentation on full sequences and its competitiveness on the occluded sequences. Project page: https://ylochman.github.io/trajectory-embedding.", + "bbox": [ + 75, + 290, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-model fitting is a classical problem in computer vision, and a typical example is the task of estimating individual motions from point trajectories observed in images containing multiple moving (and potentially deforming) objects. A trajectory is a sequence of tracked point coordinates extracted in different frames and is therefore also a point in a high-dimensional space. The commonly used union-of-subspace models [61] assume that point trajectories originating from the same object/motion are samples from the", + "bbox": [ + 75, + 713, + 468, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "same (unknown) low dimensional subspace. The inference problem thus becomes that of simultaneously determining the parameters of underlying subspaces, the trajectory assignments to subspaces, and in practice also denoising of the data. Specifically, given an observation matrix $\\mathbb{M}_{2F\\times P}$ of point trajectories", + "bbox": [ + 500, + 258, + 893, + 349 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {M} _ {2 F \\times P} = \\left( \\begin{array}{c c c} | & & | \\\\ \\mathbf {x} _ {1} & \\dots & \\mathbf {x} _ {P} \\\\ | & & | \\end{array} \\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 361, + 890, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $F$ is the number of frames and $P$ is the number of points, the goal is to find a $P \\times P$ permutation matrix $\\mathsf{P}_{\\pi}$ that orders the trajectories into sub-matrices $\\{\\mathbf{X}_i\\}_{i=1}^c$ corresponding to different motions. Each motion is represented by a $2F \\times k_i$ basis matrix $\\mathsf{B}_i$ whose columns span the subspace and therefore $\\mathbf{X}_i = \\mathsf{B}_i\\mathsf{C}_i^T$ , where $\\mathsf{C}_i$ is a $k_i \\times P_i$ shape matrix. Hence", + "bbox": [ + 496, + 420, + 893, + 526 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M} _ {2 F \\times P} \\mathrm {P} _ {\\pi} \\approx \\left[ \\begin{array}{l l l} \\mathrm {B} _ {1} \\mathrm {C} _ {1} ^ {\\top} & \\dots & \\mathrm {B} _ {c} \\mathrm {C} _ {c} ^ {\\top} \\end{array} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 537, + 890, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\sum_{i=1}^{c} P_i = P$ , and ranks $k_i$ are assumed to be small. In this work we assume that the ranks $\\{k_1, \\ldots, k_c\\}$ are known which is a common scenario under multiple rigid motions [49]. While estimating a single motion model $\\mathsf{B}_i$ from a cluster of trajectories belonging to the same motion is relatively easy, the combination of cluster assignment and model estimation turns inference into a complicated chicken and egg problem. Further complications arise due to occlusion and partial observations which make it unlikely that we can observe all the full trajectories in all cases.", + "bbox": [ + 496, + 566, + 893, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work we seek to learn a feature representation that allows us to solve the clustering problem separately from the motion estimation. For this purpose we train a simple feed forward network (with the overall structure depicted in Fig. 1) taking a trajectory as input and returning a corresponding feature embedding. Using a metric learning loss we ensure that the features are close if and only if corresponding trajectories belong to the same motions. Our approach allows us to input trajectories of varying lengths and different starting and ending times, making it possible to cluster trajectories using their features without completing any missing data. In summary our main contributions are:", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1This work was supported by the Wallenberg Artificial Intelligence, Autonomous Systems and Software Program (WASP), funded by Knut and Alice Wallenberg Foundation.", + "bbox": [ + 76, + 862, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19092", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73d4e2bd8bd17b8fabda15038d42e7098775e3a7ad53220c6d3bf6c8b5816d99.jpg", + "image_caption": [ + "Figure 1. Schematic overview of the proposed approach. Input trajectories are passed through the feature extractor to obtain $\\{\\mathbf{f}_i\\}$ . The subspace estimator generates the time-dependent basis using coordinate-MLP, infers basis coefficients from $\\mathbf{f}$ , and outputs subspace B, where each sub-matrix $\\mathsf{B}_t \\in \\mathbb{R}^{2 \\times 4}$ is computed via linear combination of basis functions. $\\mathcal{L}_{\\mathrm{InfoNCE}}$ is the contrastive loss that uses ground truth cluster assignments, $\\mathcal{L}_{\\mathrm{Residual}}$ computes the trajectory-to-subspace geometric error, and $\\mathcal{L}_{\\mathrm{FeatDiff}}$ computes the feature difference between input and reconstructed trajectory. For motion segmentation, the trajectories are clustered in a feature space using $\\{\\mathbf{f}_i\\}$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 99, + 877, + 233 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a method to map point trajectories to discriminative feature representations, which can be seen as compact descriptions of the underlying generating motion.", + "- We show that clustering of trajectories can be performed reliably solely using those features, achieving state-of-the-art performance without the need for joint motion estimation and cluster assignments.", + "- The proposed feature extraction network is trained to be invariant with respect to random occlusions and pixel noise by utilizing a novel generative model that implicitly performs per-trajectory model estimation.", + "- We present and evaluate different ways to use the proposed network, such as trajectory clustering, trajectory completion and subspace estimation." + ], + "bbox": [ + 76, + 344, + 467, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Related Work", + "text_level": 1, + "bbox": [ + 76, + 570, + 220, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within geometric vision the traditionally dominant model-fitting approaches are robust statistical methods based on RANSAC [15]. In case of low dimensional models the approach provides a simple way of effectively exploring the search space. Iterative RANSAC [48], which sequentially finds models and removes their supporting inliers, is a simple option that handles settings with multiple models. It has been observed that the greedy nature of the approach makes it inefficient when a large number of models is present [6, 12, 21, 62]. In contrast, MultiRANSAC [62] simultaneously handles multiple models using iteratively fused consensus sets, and CONSAC [25] increases the efficiency of sampling by learning a sampling strategy based on previous detections. Energy minimization is used in [6, 12, 21] to select models among random proposals that jointly explain the data well. In [7] the hard point-to-model assignment is dropped in favour of progressively determining dominant models which enables clustering in consensus space.", + "bbox": [ + 75, + 595, + 468, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an alternative to consensus some works use preference [31, 32, 46]. The J-linkage approach [46] uses affini-", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ties between data and a set of models, followed by agglomerative clustering to assign the data to models. The T-linkage method [32] is based on a continuous relaxation of the binary assignment problem. In [31] the preference representation is used in a robust PCA formulation [10] while [33] uses a set cover approach.", + "bbox": [ + 496, + 344, + 890, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low dimensional geometric models have also been used to group longer point tracks, e.g. in [56] geometric two-view relations are used to create pairwise affinities for a spectral clustering stage. Segmentations computed from images pairs are combined through an averaging approach in [3, 4]. We note that for these local methods motion-specific properties should be observable across all image pairs in order to provide a largely coherent set of clusters.", + "bbox": [ + 496, + 441, + 890, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Subspace models are commonly occurring in geometric computer vision tasks. Early works like [38, 47] showed that structure from motion under affine camera models is essentially a subspace fitting problem. Further, shape deformations are modeled using a low dimensional subspace in [9]. The generalization to multiple independently moving objects leads to segmentation into multiple subspaces [8, 11, 17]. Approaches for handling degenerate and articulated cases were given in [45, 57, 58]. The general problem can be seen as an instance of subspace clustering which assumes that high dimensional data lies in a union of low-dimensional subspaces. The corresponding group of methods aim to simultaneously cluster the data points and estimate the underlying subspaces [1, 26], and may also perform matrix completion [14, 27, 50]. The generalized PCA approach (GPCA) [53-55] treats the union-of-subspaces as a zero-level-set of a higher degree polynomial which is fit to the data. A popular approach to the problem is sparse subspace clustering [5, 13], which has been shown to perform well on the benchmark provided in [49]. These methods are based on the notion of self-expressiveness, i.e. all data in a $d$ -dimensional subspace are linear combinations of $d$ points", + "bbox": [ + 496, + 569, + 892, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from that subspace. This enables an efficient approach that is provably correct in the case of disjoint subspaces (see Section 2). The LRR approach [29, 30] uses a dictionary basis and enforces low rank patterns among the coefficients.", + "bbox": [ + 75, + 90, + 467, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, several deep learning methods have been proposed for solving the problem of unsupervised subspace clustering. The first work by Ji et al. [23] used deep autoencoder and introduced a self-expressive layer. Later, a closed form solution of the self-expressive layer was derived [43] which improved its efficiency. Alternatively, it was shown [51] that using over-complete representations increases robustness to noise and improves overall performance. Zhang et al. [59] directly learns a self-expressive data representation. The approach is highly scalable since it circumvents the need for a quadratic number of coefficients of the original self-expressive model.", + "bbox": [ + 75, + 152, + 467, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While general subspace learning models typically do not use temporal information, this can be of great importance when analysing and predicting motion. For the reconstruction of deforming objects, Akhtar et al [2] noted the duality between trajectory space and shape space. The use of pre-defined trajectory-bases such as a DCT basis (motivated by the smoothness assumption of time-dependent trajectories [20, 37]) leads to a reduction in the number of unknowns [2]. Finally, a probabilistic framework for subspace clustering with temporal information is presented in [19].", + "bbox": [ + 75, + 332, + 467, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Overview of the Approach", + "text_level": 1, + "bbox": [ + 76, + 497, + 321, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we present our approach and the underlying architecture of our network. Our goal is to learn function $f_{\\theta}$ that takes an observed point trajectory as input and yields a (relatively low-dimensional) feature embedding that is appropriate for trajectory grouping. For this purpose we adopt a small PointNet-style network [39]. One of the advantages of this type of network, apart from its simplicity, is the built-in invariance with respect to the permutation of the points in the input tensor. In order to obtain a suitable feature representation, we specify a number of desirable properties that we try to infuse through the architecture and training:", + "bbox": [ + 75, + 523, + 467, + 688 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The features should be adequate for grouping/clustering of trajectories. For this reason we want a latent space with small within-cluster-distances and large between-cluster-distances. This is achieved using an InfoNCE loss.", + "- Inspired by traditional motion segmentation formulations, we aim for features to encode meaningful \"geometric information\" about the underlying motion subspaces. For this purpose, we design a decoding module that estimates a basis B for the underlying motion subspace that the trajectory belongs to. This prevents input memorization and greatly improves generalization to unseen data.", + "- Despite input tracks being of varying lengths we opt for a feature embedding of a fixed size. This simplifies treatment of problem instances with missing data and trajec" + ], + "bbox": [ + 76, + 689, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tories starting and ending at different times. The decoding module therefore takes an extra variable, specifying at which time instance the rows of the basis matrix B should be sampled. In this way the model allows us to sub-sample and predict motion information beyond the trajectory samples.", + "bbox": [ + 511, + 90, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Finally, feature embeddings and intermediate predictions are preferred to behave smoothly with respect to added input noise. To achieve this we construct a \"noise-free\" version of the observed trajectory from the predicted basis, which we require to be close to the original trajectory through a residual loss. Additionally, this reconstructed trajectory is subsequently fed into the network to obtain a corresponding feature vector, which is also required to be close to the original feature embedding.", + "bbox": [ + 500, + 181, + 890, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Because of the properties above we obtain a trajectory-feature from which we can extract the full motion subspace. Thus we essentially learn a function that maps a single trajectory to a subspace basis. For a general collection of subspaces, where trajectories can belong to multiple of them, this may not be possible (using a purely geometric approach). A simple example is line fitting to image points, where points in the intersection of two models can belong to any of them. In such cases geometry alone is not enough to resolve ambiguous assignments but other cues such as spatial smoothness of assignments [21] have to be used. However, in high-dimensional settings with low-dimensional subspaces the situation is somewhat different. In these cases a typical assumption is that of disjoint subspaces [13], that is, two subspaces only intersect at the origin, ensuring that there is a function that takes (non-zero) trajectories to their corresponding subspaces. We remark that while this assumption is sufficient for our purposes, it is not necessary. Since our method learns a data-driven mapping from point tracks to subspaces, it is able to focus on other trajectory properties than geometric ones. Therefore it can potentially handle the case of intersecting subspaces as long as the trajectory assignment is unambiguous.", + "bbox": [ + 496, + 318, + 890, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The disjoint subspace assumption We conclude this section by illustrating that the disjoint subspace assumption is valid in practical settings by analysing the data in the Hopkins 155 dataset. Figure 2 shows that re-using subspaces to explain trajectories in the other clusters generally leads to higher errors.", + "bbox": [ + 496, + 672, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, two linear subspaces $S_{i}$ and $S_{j}$ are disjoint if $\\dim(S_{i} \\oplus S_{j}) = \\dim(S_{i}) + \\dim(S_{j})$ . Hence if $\\mathsf{B}_{i}$ and $\\mathsf{B}_{j}$ contain bases for the two subspaces, the rank of the concatenated matrix, $\\eta_{ij} = \\mathrm{rank}((\\mathsf{B}_i \\mathsf{B}_j))$ , fulfills $\\eta_{ij} < \\mathrm{rank}(\\mathsf{B}_i) + \\mathrm{rank}(\\mathsf{B}_j)$ if the subspaces intersect. We extracted subsequences of varying length from the Hopkins 155 dataset and used the ground-truth clusters to generate the corresponding subspaces — for all pairs $\\{\\mathsf{B}_i, \\mathsf{B}_j\\}$ , where $i \\neq j$ and $\\mathrm{shape}(\\mathsf{B}_i) = \\mathrm{shape}(\\mathsf{B}_j)$ , the rank $\\eta_{ij}$ of the concatenated", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19094", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f2cf8a5e290db9d7728999b588d62c2b9f439ed21dcba9f5234c0a98852573cb.jpg", + "image_caption": [ + "Figure 2. RMS normalized error of the trajectory clusters with respect to the subspaces in Hopkins155 dataset (subsequences of length $F = 40, 50,$ and 60). An $(i,j)$ cell corresponds to the trajectory cluster of the $i^{\\text{th}}$ motion and the subspace of $j^{\\text{th}}$ motion." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 467, + 200 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matrix was found to be $8^{2}$ . This suggests that up to numerics the subspaces do not intersect. In the following Section we describe the specific architecture and the employed losses in our approach in more detail.", + "bbox": [ + 75, + 292, + 470, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Architecture and Training Losses", + "text_level": 1, + "bbox": [ + 76, + 369, + 382, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The schematic overview of the proposed architecture and training losses is shown in Fig. 1. Using the hypothesis that most observed trajectories come from a unique motion model, we aim to learn a function $f_{\\theta}$ mapping a point trajectory directly to a representation of the corresponding motion model. We interpret the output $f_{\\theta}(\\mathbf{x})$ as an abstract description of the generating motion, from which a concrete representation to measure geometric consistency can be extracted. We choose linear (low-dimensional) subspaces as the concrete realizations of motion models. Specifically, we learn mappings $f_{\\theta}$ and $g_{\\phi}$ using training data such that", + "bbox": [ + 75, + 393, + 468, + 561 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $f_{\\theta}$ maps a trajectory of varying length to a fixed-length feature embedding, and", + "2. $g_{\\phi}$ converts this feature vector for a given query time into the corresponding rows of the subspace representation." + ], + "bbox": [ + 76, + 561, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Feature Extraction", + "text_level": 1, + "bbox": [ + 76, + 632, + 259, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To construct a discriminative function $f_{\\theta}(\\cdot)$ , that is able to infer the motion from an input point trajectory, we first leverage the ability of the neural networks to compress high-dimensional complex data space into a lower-dimensional latent space with task specific invariance properties, from which it is easier to infer the output. A feature extractor for the point trajectories $f_{\\theta}: \\mathbf{x} \\mapsto \\mathbf{f} \\in \\mathbb{R}^{d}$ is therefore trained to produce invariant feature representations using ground truth clusters. In particular, an invariance with respect to the measurement noise and the variations within the motion clusters is of our interest. We therefore use a contrastive loss, namely a variant of the InfoNCE loss [36], to conduct metric learning, leading to a spread of feature embeddings", + "bbox": [ + 75, + 655, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in order to improve feature-based motion clustering,", + "bbox": [ + 498, + 90, + 846, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {I n f o N C E}} = - \\frac {1}{| \\mathcal {D} |} \\sum_ {(i, j, l, k) \\in \\mathcal {D}} \\log \\left(\\frac {p _ {i j}}{p _ {i j} + p _ {l k}}\\right), \\\\ p _ {i j} = \\exp \\left(- \\frac {\\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| _ {2} ^ {2}}{T}\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 114, + 890, + 193 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i = f_{\\theta}(\\mathbf{x}_i)$ and $\\mathcal{D}$ is the set of quadruples $(i,j,l,k)$ where $(i,k)$ belong to the same cluster and $(l,k)$ are random pairs, and $T$ is the temperature (hyper-)parameter, which we fix to 1. $\\mathcal{L}_{\\mathrm{InfoNCE}}$ is an instance of a supervised contrastive loss and is tightly connected to the triplet loss used in metric learning [24].", + "bbox": [ + 496, + 202, + 890, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned earlier, the proposed architecture resembles PointNet [39] with several modifications. The employed encoder consists of 1D temporal convolutional layers applied to each trajectory independently, with x and y values representing two input channels. A global max-pooling in a temporal domain is then applied on the time-dependent representations, followed by a small MLP and an $L_{2}$ normalization. The features are therefore lying on the $d - 1$ -dimensional unit sphere. Note that we do not apply pooling across the points (global context layer) due to the disjoint subspace assumption. For the experiments, we set the dimensionality of the latent space to $d = 128$ .", + "bbox": [ + 496, + 294, + 892, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Subspace Estimation", + "text_level": 1, + "bbox": [ + 500, + 483, + 696, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent vector $f_{\\theta}(\\mathbf{x})$ ideally identifies a motion model, but for geometric consistency this feature embedding requires conversion into a suitable mathematical motion model. We assume that observed point trajectories are generated by low-rank subspace models, and therefore the feature embedding is mapped by a subsequent network $g_{\\phi}$ to the corresponding matrix representation of the associated subspace. Specifically, the mapping $g_{\\phi}: \\mathcal{F} \\times \\mathbb{R} \\rightarrow \\mathbb{R}^{2 \\times r}$ produces rows $\\mathsf{B}_t$ for the given motion model $\\mathbf{f} = f_{\\theta}(\\mathbf{x})$ and the time index $t$ . Ideally the input trajectory $\\mathbf{x}$ lies in the predicted subspace; if the trajectory $\\mathbf{x}$ is consisting of (2D) points observed at times $t_1, \\ldots, t_F$ , then the output subspace is", + "bbox": [ + 496, + 506, + 892, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {B} = \\left( \\begin{array}{c} g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {1}\\right) \\\\ \\vdots \\\\ g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {F}\\right) \\end{array} \\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 713, + 890, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the (squared) residual is given in closed-form as", + "bbox": [ + 500, + 779, + 846, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {c}} \\| \\mathbf {x} - \\mathrm {B c} \\| _ {2} ^ {2} = \\| \\mathbf {x} - \\mathrm {B B} ^ {\\dagger} \\mathbf {x} \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 805, + 890, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{c}$ is the coefficient vector and $\\mathsf{B}^{\\dagger} = (\\mathsf{B}^{\\top}\\mathsf{B})^{-1}\\mathsf{B}^{\\top}$ . If the network identified the subspace correctly, then the residual is in the order of the observation noise level (and therefore small).", + "bbox": [ + 496, + 838, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2The matrix rank is computed as the number of singular values greater than $\\delta \\sigma_{1}$ , where $\\sigma_{1}$ is the largest singular value, $\\delta = 2F\\varepsilon$ , and $\\varepsilon \\approx 10^{-7}$ is the numerical zero for torch.float32.", + "bbox": [ + 75, + 862, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19095", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subspace basis functions The network $g_{\\phi}$ can be a small but generic MLP. Alternatively, we may leverage existing domain knowledge and choose the structure of $g_{\\phi}$ accordingly. In particular, we assume that all motions are linear combination of a finite, but continuous time trajectory basis $\\{h^j\\}$ . The basis functions $h^j$ can be fully fixed (such as a cosine basis with pre-determined frequencies and phase shifts [2]), parametric (such as the cosine basis with trainable frequencies and phase shifts [44]) or entirely non-parametric (i.e. represented by a generic MLP). Our choice of basis functions is a damped version of a cosine basis and is given as follows,", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\psi} ^ {j} (t) = e ^ {- \\left(\\alpha_ {j} (t - \\mu_ {j})\\right) ^ {2}} \\cos (\\beta_ {j} t + \\gamma_ {j}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 282, + 468, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi = (\\mu_j, \\alpha_j, \\beta_j, \\gamma_j)_j$ are trainable parameters. The mapping $g_{\\phi}$ can therefore be written as a composition", + "bbox": [ + 76, + 314, + 468, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\phi} (\\mathbf {f}, t) = \\omega_ {\\zeta} (\\mathbf {f}) \\mathbf {h} _ {\\psi} (t), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 354, + 468, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_{\\psi}(t) = (h_{\\psi}^{j}(t))_{j}$ , and the trainable function $\\omega_{\\zeta}$ extracts the basis coefficients from $\\mathbf{f}$ . If we have $N$ basis functions (i.e. $\\dim (\\mathbf{h}_{\\psi}(t)) = N$ ), then $\\omega_{\\zeta}(\\mathbf{f})$ is a $2r\\times N$ matrix. The total set of trainable parameters in the subspace estimation network is $\\phi = (\\zeta ,\\psi)$ . Finally, the network prediction can be compactly written as", + "bbox": [ + 76, + 385, + 468, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathsf {B} = \\mathrm {H} _ {\\psi} (\\mathbf {t}) \\Omega_ {\\zeta} \\left(f _ {\\theta} (\\mathbf {x})\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 488, + 468, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega_{\\zeta}(\\mathbf{f}) = \\mathrm{reshape}_{2N\\times r}\\left(\\omega_{\\zeta}(\\mathbf{f})\\right)$ and", + "bbox": [ + 76, + 515, + 354, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H} _ {\\psi} (\\mathbf {t}) = \\left( \\begin{array}{l l} \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} \\\\ \\vdots & \\\\ \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} \\end{array} \\right) \\in \\mathbb {R} ^ {2 F \\times 2 N}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 542, + 468, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set the number of subspace basis functions to $N = 64$ in our experiments. Since the predicted subspace matrix B augments the feature representation (see Sec. 3.3 below), we enforce a unique subspace representation by fixing the top block of B to the identity matrix (by multiplying with the respective inverse).", + "bbox": [ + 76, + 641, + 468, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Time dependence and time invariance The network represented by $g_{\\phi}$ can be interpreted as taking a \"feature\" token $\\mathbf{f}$ and a time coordinate $t$ as input, and is therefore somewhat analogous to the coordinate-MLP employed for implicit neural primitives such as [35, 44] or time series modeling [16]. The continuous time basis $\\{h^j\\}$ can be seen as domain-specific equivalent to positional coding, but its design is based on the expected motion patterns and does not necessarily meet the requirements of generic positional coding [40, 44, 60].", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The formulation above is intentionally not time-invariant: the same trajectory observed at different starting times may predict different motion models. This is usually a desired behavior for motion segmentation, e.g. two cars traveling the same path at different time form different motion segments. As with CNNs applied to images, low-level trajectory features are extracted by convolutional layers, and time-dependence is incorporated via final fully connected and pooling layers.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training", + "text_level": 1, + "bbox": [ + 500, + 234, + 604, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total training loss is the weighted sum of a contrastive loss $\\mathcal{L}_{\\mathrm{InfoNCE}}$ applied on $\\{\\mathbf{v}_i\\}$ , where $\\mathbf{v}_i = (\\mathrm{flatten}(\\mathsf{B}_i)^\\top \\mathbf{f}_i^\\top)^\\top$ , the reconstruction loss $\\mathcal{L}_{\\mathrm{Residual}}$", + "bbox": [ + 498, + 258, + 890, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {R e s i d u a l}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| \\mathbf {x} _ {j} - \\tilde {\\mathbf {x}} _ {j} \\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 313, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the feature difference loss", + "bbox": [ + 500, + 364, + 700, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {F e a t D i f f}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| f _ {\\theta} (\\mathbf {x} _ {j}) - f _ {\\theta} (\\tilde {\\mathbf {x}} _ {j}) \\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 385, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}_j = \\mathsf{B}(\\mathbf{x}_j)\\mathsf{B}(\\mathbf{x}_j)^\\dagger\\mathbf{x}_j$ is the projection of $\\mathbf{x}_j$ onto $\\mathsf{B}(\\mathbf{x}_j)$ . In short, $\\mathcal{L}_{\\mathrm{InfoNCE}}$ forces both an approximate invariance of the feature extractor $f_{\\theta}$ with respect to the variations within clusters and smoothness of the coordinate-MLP $g_{\\phi}$ , $\\mathcal{L}_{\\mathrm{Residual}}$ ensures geometric consistency of the subspaces with respect to input trajectories, and $\\mathcal{L}_{\\mathrm{FeatDiff}}$ forces an approximate invariance of $f_{\\theta}$ with respect to the pixel noise and can also be seen as enforcing smoothness of $f_{\\theta}$ . We train the network in two stages. First, we pre-train $f_{\\theta}$ with $\\mathcal{L}_{\\mathrm{InfoNCE}}$ until convergence. Then we add $g_{\\phi}$ and train it together with $f_{\\theta}$ using the total loss. To enforce the approximate invariance with respect to random detection failures, we synthesize missing points during training by applying dropout to the input sequences.", + "bbox": [ + 496, + 438, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Trajectory Completion", + "text_level": 1, + "bbox": [ + 500, + 657, + 712, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $B_{\\theta, \\phi}(\\mathbf{x}, \\mathbf{t})$ provide the subspace for an input trajectory $\\mathbf{x} \\in \\mathbb{R}^{2F}$ and time vector $\\mathbf{t} = (t_1, \\dots, t_F)^\\top$ . Assume now that $\\mathbf{x}$ is only partially observed, and let $\\mathbf{w} \\in \\{0, 1\\}^{2F}$ be the visibility mask. Further, let $\\bar{\\mathbf{w}} := \\mathbf{1} - \\mathbf{w}$ . The goal of the trajectory completion is to estimate missing values $\\bar{\\mathbf{x}}$ that minimize the following objective", + "bbox": [ + 496, + 681, + 890, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\bar {\\mathbf {x}}) := \\left\\| (I - B (\\bar {\\mathbf {x}}) B (\\bar {\\mathbf {x}}) ^ {\\dagger}) (w \\odot x + \\bar {w} \\odot \\bar {x}) \\right\\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 780, + 890, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf{B}(\\bar{\\mathbf{x}}) = B_{\\theta ,\\phi}(\\mathbf{w}\\odot \\mathbf{x} + \\bar{\\mathbf{w}}\\odot \\bar{\\mathbf{x}},\\mathbf{t})$ . Back-propagation for gradient-based determination of $\\bar{\\mathbf{x}}$ is in principle possible but computationally expensive. Alternatively, we rely on the approximate invariance of the feature extractor $f_{\\theta}$ with respect to the length of the input sequence and extrapolate the output subspace in time using $g_{\\phi}(\\mathbf{f},t)$ . More", + "bbox": [ + 496, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19096", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/bb7d54b216d6c9e029521371a1eec308e4ebbd7149caed979f55a1d1261b7013.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method2 motionsHopkins155AllHopkins12KT3DMoSeg
MeanMedianTimeMeanMedianTimeMeanMedianTimeMeanMedianMeanMedian
RANSAC [48]5.561.18175ms22.9422.03258ms9.763.21194ms----
GPCA [54]4.590.38324ms28.6628.26738ms10.342.54417ms--34.6033.95
MSL [45]4.140.0011h 4m8.231.761d 23h5.030.0019h 11m----
LSA [57]3.450.597.58s9.732.3315.96s4.940.909.47s--38.3038.58
ALC5 [42]3.030.00-6.261.02-3.760.265m 15s3.810.1724.3119.04
ALCsp [42]2.400.43-6.690.67-3.370.496m 11s1.281.07--
LRR [30]4.100.22-9.890.56-5.410.531.1s--33.6736.01
SSC [13]0.820.00-2.450.20-2.450.20920ms--33.8833.54
RSIM [22]0.780.00-1.770.28-1.010.00176ms0.680.70--
MultiConsensus [7]------4.40-40ms----
Ours0.630.07ms0.600.010ms0.620.09ms5.122.045.850.80
", + "bbox": [ + 76, + 88, + 929, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Motion segmentation results on the Hopkins155 [49], Hopkins12 [41] and KT3DMoSeg [56] datasets. The evaluation metrics is classification error $(\\%)$ . ' -' indicates that the results were not reported / no code available.", + "bbox": [ + 75, + 273, + 892, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "specifically, the subspace of $\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}$ can be estimated from $\\mathbf{x}_{\\mathrm{vis}}$ -a fully visible sub-trajectory of the trajectory $\\mathbf{x}$ as long as the feature representation $\\mathbf{f}$ can be inferred (i.e. $\\mathbf{x}_{\\mathrm{vis}}$ is sufficiently long). If we fix this estimate $\\mathbf{B}$ and let $\\mathsf{W} := \\mathrm{diag}(\\mathbf{w})$ , $\\bar{\\mathsf{W}} := \\mathrm{diag}(\\bar{\\mathsf{w}})$ , the new objective can be stated as", + "bbox": [ + 75, + 325, + 470, + 416 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {J} (\\bar {\\mathbf {x}}) := \\left\\| \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\mathbf {x} + \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\bar {\\mathbf {x}} \\right\\| ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 428, + 468, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the minimizer $\\bar{\\mathbf{x}}^*$ of $\\tilde{J}$ is given by", + "bbox": [ + 76, + 465, + 331, + 483 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathbf {x}} ^ {*} = - \\left(\\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {\\mathbf {w}}\\right) ^ {\\dagger} \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\mathbf {W x} =: \\mathrm {A} (\\mathrm {B}) \\mathbf {x}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 497, + 468, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is now possible to refine the estimated subspace using $\\bar{\\mathbf{x}}^*$ by feeding $\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}^*$ into the network. Overall, we apply the following iterative procedure for trajectory completion,", + "bbox": [ + 75, + 532, + 468, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{r l} \\mathsf {B} _ {0} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {x} _ {\\text {v i s}}, \\mathbf {t}) \\\\ \\bar {\\mathbf {x}} _ {i + 1} & \\leftarrow \\mathsf {A} (\\mathsf {B} _ {i}) \\mathbf {x} \\\\ \\mathsf {B} _ {i + 1} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {w} \\odot \\mathbf {x} + \\bar {\\mathbf {w}} \\odot \\bar {\\mathbf {x}} _ {i + 1}, \\mathbf {t}). \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 607, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70641f339b1ce61533b8b379eb5789df914a5deffdaedfa664b3a0dbec7aec9a.jpg", + "image_caption": [ + "Figure 3. Histograms of the classification errors for the groups with (left) 2 and (right) 3 motions in Hopkins155 dataset." + ], + "image_footnote": [], + "bbox": [ + 81, + 690, + 272, + 838 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2482bb9b3f2595898ed145e93ff942b73970873bd6d44eb021a4242821af9ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 688, + 464, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 324, + 632, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed network was trained on the two motion segmentation datasets: Hopkins155 [49] and KT3DMoSeg [56], yielding two different sets of weights. In our experiments, we used NVIDIA GeForce RTX 3080, CUDA 11.5, PyTorch 1.11.", + "bbox": [ + 496, + 349, + 890, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain groups of trajectories, it is sufficient to perform a partial forward pass through the (trained) feature extractor $f_{\\theta}$ and cluster in the feature space. We use hierarchical clustering as it provides a meaningful framework for selecting the number of clusters, i.e. motions. The subspaces can then be fit to each cluster. If the goal is to estimate the subspace for a single input trajectory, a full forward pass through $f_{\\theta}$ and $g_{\\phi}$ should be performed. For motion segmentation with missing data, the largest fully visible blocks of trajectories (but ignoring single occluded points) can be used to estimate the initial subspaces B, followed by trajectory completion as outlined in (15). The completed trajectories can then be used to cluster in the feature space.", + "bbox": [ + 496, + 425, + 892, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Motion Segmentation", + "text_level": 1, + "bbox": [ + 500, + 631, + 702, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of the proposed network in terms of motion segmentation on Hopkins155 [49], Hopkins12 [41], and KT3DMoSeg [56] datasets. We first compare the proposed method to the state-of-the-art algorithms that similarly assume rank-4 of B: sequential RANSAC [48], GPCA [54], MSL [45], LSA [57] (with the global subspace dimensionality $D = 4n$ for $n$ clusters), $\\mathrm{ALC}_5$ and $\\mathrm{ALC}_{\\mathrm{sp}}$ [42], LRR [30], SSC [13], RSIM [22], and MultiConsensus [7]. We then also compare our method to the other approaches making different assumptions on the underlying motion model [3, 4, 28, 56]. The metrics used to evaluate clustering performance is the classification error (\\%) which is standard for this benchmark [49].", + "bbox": [ + 496, + 654, + 893, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Hopkins155 dataset Hopkins155 [49] is the commonly used benchmark for trajectory-based motion", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19097", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb6f11bcbfd11722946152a0079c5a998a3ad2a92021a9f05d2a0994b55613a1.jpg", + "image_caption": [ + "Figure 4. The approximate invariance measured as an angle between features extracted from the reference trajectory and its perturbed version using (left) Gaussian pixel noise, (middle) perturbation within subspace, and (right) uniform occlusions. The gradually increasing level of corruption is shown on the x-axis." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 349, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34c5c2d4b6876d165fda2418c5ef64935724c2015b305639df503d5738f3facc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 89, + 612, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d56a9502e2b503c35590e5bd252a89b5368f3790416b0f4458253b4996d2fc7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 90, + 875, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "segmentation. The dataset consists of 155 sequences with 63-556 point trajectories generated by two or three motions in the scene. Each trajectory is fully visible, and the length varies from 15 to 100 frames. The averaged results are shown in the left block of Table 1, and the histograms of classification errors are shown in Fig. 3. Our network is very fast and produces the most accurate segmentations on fully visible trajectories.", + "bbox": [ + 75, + 305, + 470, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Hopkins12 dataset Hopkins12 [41] is a commonly used dataset for evaluating subspace clustering with missing data. It consists of 12 corrupted sequences from Hopkins155 dataset, where the corruption was done manually by labeling entries of trajectories as missing. The corruption rate of the observation matrix M varies from $1\\%$ to $23\\%$ . For $\\mathrm{ALC}_5$ and $\\mathrm{ALC_{sp}}$ [42], we report $l^1$ -based entry completion results. The results are shown in the middle block of Table 1. The proposed network shows promising results. Note that the maximum corruption rate of the individual trajectories in the Hopkins12 dataset is $93\\%$ and, unlike all the state-of-the-art approaches, the proposed trajectory completion method is not global, i.e. it works with individual trajectories. Therefore, if a provided individual sub-trajectory does not contain sufficient relevant information for motion estimation (which becomes more likely as the trajectory gets shorter), incorporating information from all tracks becomes necessary to disambiguate the true motion. Utilizing such global trajectory information is left for future research.", + "bbox": [ + 75, + 436, + 468, + 736 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on KT3DMoSeg dataset KT3DMoSeg [56] is an adapted version of the KITTI dataset [18, 34] containing 22 sequences with 158-1018 point trajectories and annotated segmentations. It is more challenging than Hopkins155 in several aspects. There is a dominant motion corresponding to static \"background\" scene which makes cluster distribution unbalanced. The trajectories are also highly occluded: the corruption rate of M goes up to $54\\%$ , and for the individual trajectories — up to $75\\%$ . Lastly, the sequences have shorter length of 10 to 20 frames. The results for this dataset", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are shown in the right block of Table 1 indicating that the proposed approach outperforms previous methods.", + "bbox": [ + 498, + 305, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods with different assumptions We also compared our approach to the other methods making different assumptions. Two-perspective-view (TPV) [28] assumes perspective cameras and uses pairwise epipolar relations. Spectral clustering methods proposed by [56]: kernel addition (KerAdd), co-regularization (CoReg), and subset constrained (Subset) spectral clustering also look at pairwise relations, but use fundamental matrices, affine matrices or homographies to compute the pairwise affinities. SYNCH [3] and MODE [4] similarly fit multiple fundamental matrices to each pair of images, but perform averaging. They assume that only pairwise matches are given. These methods are not directly comparable to ours, however the proposed method proves to be competitive as evidenced in Table 2.", + "bbox": [ + 496, + 342, + 892, + 553 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e19c972bab8f85a6480b434abcc640ccf5e1fccdd2b115b0260635417ba4161.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodHopkins155 MeanHopkins12 Med.KT3DMoSeg MeanMed.
TPV [28]2.34----
KerAdd [56]0.360.110.08.311.02
CoReg [56]0.460.060.07.920.75
Subset [56]0.310.060.08.080.71
SYNCH-tracks [3]3.675.460.57--
SYNCH [3]1.193.190.28--
MODE [4]1.374.330.38--
Our method0.625.122.045.850.80
", + "bbox": [ + 500, + 559, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison to the other methods making different assumptions. TPV [28] has no code publicly available.", + "bbox": [ + 498, + 696, + 890, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Approximate Invariances of $f_{\\theta}$", + "text_level": 1, + "bbox": [ + 500, + 741, + 769, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We experimentally validate the approximate invariance of the trained feature extractor $f_{\\theta}$ with respect to different types of synthetic data corruption: (1) Gaussian pixel noise with the standard deviation $\\sigma$ gradually increasing up to 10 px, (2) perturbations within the subspace, where the trajectories are first corrupted by Gaussian pixel noise with $\\sigma$ up to 300 px and then projected back to the underlying subspace, (3) synthesized uniform occlusions with missing data levels going up to 50%, and (4) tracking failures, where", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19098", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/54c344a69caf4d564a94fdce18966557b2ac3337c72a1fab150f3cfa15f8dda4.jpg", + "image_caption": [ + "Figure 5. The approximate invariance measured as an angle between features extracted from the reference trajectory and its shortened version for (left) shorter sequences with $F = 15 \\ldots 24$ , (middle) mid-length sequences with $F = 25 \\ldots 31$ , and (right) longer sequences with $F = 35 \\ldots 61$ . The gradually increasing percentage of sequence length reduction is shown on the x-axis." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 349, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d999bcf6c79aa629d8c035d99a2a535b35ba2f87890f0d2c8410bdd95b458931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 90, + 612, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8264cabf08a73bbc65ae118e87608d0a284f3ea6844eeb1988898deeed37f7b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 90, + 874, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the rearmost points are gradually excluded from the trajectories. All types of corruption are applied on the training sequences of Hopkins155 dataset. The invariances are measured using the angle between the features extracted from original and corrupted trajectories. The smaller the angle is, the more invariant the feature is to the certain type of data corruption.", + "bbox": [ + 75, + 306, + 470, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for the pixel noise, perturbations within the subspace, and uniform occlusions are shown in Fig. 4. Note that the features are $L_{2}$ -normalized in the $d$ -dimensional space for $d = 128$ . In such a high-dimensional space the two randomly picked vectors are likely to be almost orthogonal [52]. The network proves to be approximately invariant hence robust to these types of corruption, and especially to the pixel noise. Since 8-10 px is quite large (for the images of sizes $320 \\times 240$ , $640 \\times 480$ , or $720 \\times 480$ ), this result also shows robustness of $f_{\\theta}$ to occasional outlying points in the trajectory. The results for the synthesized tracking failures are shown in Fig. 5, where a single curve corresponds to a certain initial trajectory length $F$ . The angles are slightly higher for shorter sequences, which is expected since they contain less information. Note that during training we do not force the network to produce the same feature representations for the trajectories of the same point starting at the same time but having different length.", + "bbox": [ + 75, + 412, + 473, + 686 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2cbfb2bf4ea068025b17893b6eea77e0fbdf682ea68c279e949e97f095278ed0.jpg", + "image_caption": [ + "Figure 6. Mean pixel error of recovering from random occlusions for the corruption rates of (left) $10 - 40\\%$ and (right) $50 - 80\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 84, + 705, + 464, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Trajectory Completion", + "text_level": 1, + "bbox": [ + 500, + 306, + 712, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate the completion accuracy of the proposed method on the trajectories from Hopkins155 dataset with synthesized uniform occlusions. The performance is measured using the mean pixel error between the ground truth (i.e., detected/tracked) values of the synthetically occluded points and the predicted values. The error as a function of iterations for different corruption rates is shown in Fig. 6, where we split the results into a minor and a major corruption group. From the results we conclude that it is sufficient to run the algorithm (15) for 3 to 5 iterations for minor corruptions which takes 1s on average, and 5 to 20 iterations for major corruptions which takes 3s on average.", + "bbox": [ + 496, + 333, + 893, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 537, + 617, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we propose to learn a mapping from point trajectories to feature vectors in an embedding space, which fully identify the generating motion. We illustrate that these feature vectors can be directly used for subspace clustering, thereby eliminating the need for simultaneous estimation of assignments and subspace models. The proposed training loss ensures not only the utility of the obtained embedding for clustering, but also forces the feature vector to encode the actual motion parameters, which are retrieved by a dedicated coordinate-MLP. Comparisons to baselines on Hopkins155 and KT3DMoSeg highlight the efficiency and accuracy of our method, providing top clustering results with running time below 10ms.", + "bbox": [ + 496, + 565, + 890, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work we focus on the linear subspace framework as sole motion model to explain observed trajectories. One direction for future work is how to incorporate additional and/or different prior assumptions on the motion model into our method. Further, the linear subspace assumption is strongly tied to affine cameras, which only approximate the more realistic pinhole camera model. Therefore extending this work to more general camera models is a challenging direction for future research.", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19099", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopoulos, and Prabhakar Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. In Proceedings of the 1998 ACM SIGMOD international conference on Management of data, pages 94-105, 1998. 2", + "[2] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 3, 5", + "[3] Federica Arrigoni and Tomas Pajdla. Motion segmentation via synchronization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2, 6, 7", + "[4] Federica Arrigoni and Tomas Pajdla. Robust motion segmentation from pairwise matches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 671-681, 2019. 2, 6, 7", + "[5] Liang Bai and Jiye Liang. Sparse subspace clustering with entropy-norm. In International conference on machine learning, pages 561-568. PMLR, 2020. 2", + "[6] Daniel Barath and Jiri Matas. Progressive-x: Efficient, anytime, multi-model fitting algorithm. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3780-3788, 2019. 2", + "[7] Daniel Barath, Denys Rozumnyi, Ivan Eichhardt, Lerente Hajder, and Jiri Matas. Finding geometric models by clustering in the consensus space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5424, 2023. 2, 6", + "[8] T.E. Boult and L. Gottesfeld Brown. Factorization-based segmentation of motions. In Proceedings of the IEEE Workshop on Visual Motion, pages 179-186, 1991. 2", + "[9] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2", + "[10] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? J. ACM, 58(3), 2011. 2", + "[11] J. Costeira and T. Kanade. A multi-body factorization method for motion analysis. In Proceedings of IEEE International Conference on Computer Vision, pages 1071-1076, 1995. 2", + "[12] A. Delong, A. Osokin, H. Isack, and Y. Boykov. Fast approximate energy minimization with label costs. International Journal of Computer Vision, 96(1):1-27, 2012. 2", + "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE transactions on pattern analysis and machine intelligence, 35, 2013. 2, 3, 6", + "[14] Jicong Fan and Madeleine Udell. Online high rank matrix completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8690-8698, 2019. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2", + "[16] Elizabeth Fons, Alejandro Sztrajman, Yousef El-Laham, Alexandros Iosifidis, and Svitlana Vyetrenko. Hypertime: Implicit neural representation for time series. arXiv preprint arXiv:2208.05836, 2022. 5", + "[17] C.W. Gear. Feature grouping in moving objects. In Proceedings of 1994 IEEE Workshop on Motion of Non-rigid and Articulated Objects, pages 214-219, 1994. 2", + "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7", + "[19] Behnam Gholami and Vladimir Pavlovic. Probabilistic temporal subspace clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3066-3075, 2017. 3", + "[20] Paulo FU Gotardo and Aleix M Martinez. Computing smooth time trajectories for camera and deformable shape in structure from motion with occlusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(10):2051-2065, 2011. 3", + "[21] Hossam Isack and Yuri Boykov. Energy-based geometric multi-model fitting. International journal of computer vision, 97(2):123-147, 2012. 2, 3", + "[22] Pan Ji, Mathieu Salzmann, and Hongdong Li. Shape interaction matrix revisited and robustified: Efficient subspace clustering with corrupted and incomplete data. In Proceedings of the IEEE International Conference on computer vision, pages 4687-4695, 2015. 6", + "[23] Pan Ji, Tong Zhang, Hongdong Li, Mathieu Salzmann, and Ian Reid. Deep subspace clustering networks. Advances in neural information processing systems, 30, 2017. 3", + "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 4", + "[25] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 2", + "[26] Hans-Peter Kriegel, Peer Kroger, and Arthur Zimek. Clustering high-dimensional data: A survey on subspace clustering, pattern-based clustering, and correlation clustering. Acm transactions on knowledge discovery from data (tkdd), 3(1): 1-58, 2009. 2", + "[27] Connor Lane, Ron Boger, Chong You, Manolis Tsakiris, Benjamin Haeffele, and Rene Vidal. Classifying and comparing approaches to subspace clustering with missing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2", + "[28] Zhuwen Li, Jiaming Guo, Loong-Fah Cheong, and Steven Zhiying Zhou. Perspective motion segmentation via" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19100", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "collaborative clustering. In Proceedings of the IEEE international conference on computer vision, pages 1369-1376, 2013. 6, 7", + "[29] Guangcan Liu and Shuicheng Yan. Latent low-rank representation for subspace segmentation and feature extraction. In 2011 International Conference on Computer Vision, pages 1615-1622, 2011. 3", + "[30] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 3, 6", + "[31] Luca Magri and Fusiello Andrea. Robust multiple model fitting with preference analysis and low-rank approximation. In Proceedings of the British Machine Vision Conference 2015, pages 20-1, 2015. 2", + "[32] Luca Magri and Andrea Fusiello. T-linkage: A continuous relaxation of j-linkage for multi-model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3954-3961, 2014. 2", + "[33] Luca Magri and Andrea Fusiello. Multiple model fitting as a set coverage problem. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3318-3326, 2016. 2", + "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 7", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5", + "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[37] Hyun Soo Park, Takaaki Shiratori, Iain Matthews, and Yaser Sheikh. 3d reconstruction of a moving point from a series of 2d projections. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III 11, pages 158-171. Springer, 2010. 3", + "[38] C.J. Poelman and T. Kanade. A parapspective factorization method for shape and motion recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(3):206-218, 1997. 2", + "[39] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 4", + "[40] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 5", + "[41] Shankar Rao, Roberto Tron, Rene Vidal, and Yi Ma. Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(10):1832-1845, 2009. 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Shankar R Rao, Roberto Tron, René Vidal, and Yi Ma. Motion segmentation via robust subspace separation in the presence of outlying, incomplete, or corrupted trajectories. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 6, 7", + "[43] Junghoon Seo, Jamyoung Koo, and Taegyun Jeon. Deep closed-form subspace clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3", + "[44] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 5", + "[45] Yasuyuki Sugaya and Kenichi Kanatani. Geometric structure of degeneracy for multi-body motion segmentation. In International Workshop on Statistical Methods in Video Processing, pages 13-25. Springer, 2004. 2, 6", + "[46] Roberto Toldo and Andrea Fusiello. Robust multiple structures estimation with j-linkage. In Computer Vision – ECCV 2008, pages 537–547. Springer Berlin Heidelberg, 2008. 2", + "[47] Carlo Tomasi and Takeo Kanade. Shape and motion from image streams: a factorization method. International journal of computer vision, 9(2):137-154, 1992. 2", + "[48] Philip HS Torr. Geometric motion segmentation and model selection. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 356(1740):1321-1340, 1998. 2, 6", + "[49] Roberto Tron and René Vidal. A benchmark for the comparison of 3-d motion segmentation algorithms. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1, 2, 6", + "[50] Manolis Tsakiris and Rene Vidal. Theoretical analysis of sparse subspace clustering with missing entries. In International Conference on Machine Learning, pages 4975-4984. PMLR, 2018. 2", + "[51] Jeya Maria Jose Valanarasu and Vishal M Patel. Overcomplete deep subspace clustering networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 746-755, 2021. 3", + "[52] Roman Vershynin. High-dimensional probability: An introduction with applications in data science. Cambridge university press, 2018. 8", + "[53] René Vidal and Richard Hartley. Motion segmentation with missing data using powerfactorization and gpca. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II–II. IEEE, 2004. 2", + "[54] Rene Vidal, Yi Ma, and Shankar Sastry. Generalized principal component analysis (gpca). IEEE transactions on pattern analysis and machine intelligence, 27(12):1945-1959, 2005. 6", + "[55] René Vidal, Roberto Tron, and Richard Hartley. Multiframe motion segmentation with missing data using powerfactorization and gpca. International Journal of Computer Vision, 79:85-105, 2008. 2", + "[56] Xun Xu, Loong Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models." + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19101", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2859-2867, 2018. 2, 6, 7", + "[57] Jingyu Yan and Marc Pollefeys. A general framework for motion segmentation: Independent, articulated, rigid, non-rigid, degenerate and non-degenerate. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9, pages 94-106. Springer, 2006. 2, 6", + "[58] Lihi Zelnik-Manor and Michal Irani. Temporal factorization vs. spatial factorization. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8, pages 434-445. Springer, 2004. 2", + "[59] Shangzhi Zhang, Chong You, René Vidal, and Chun-Guang Li. Learning a self-expressive network for subspace clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12393-12403, 2021. 3", + "[60] Jianqiao Zheng, Sameera Ramasinghe, and Simon Lucey. Rethinking positional encoding. arXiv preprint arXiv:2107.02561, 2021. 5", + "[61] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1", + "[62] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE International Conference on Image Processing 2005, pages III-153. IEEE, 2005. 2" + ], + "bbox": [ + 78, + 90, + 468, + 513 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19102", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_model.json b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cde05795c5802218d2e7e1007cbdd3c1186307dc --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.2, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "Learned Trajectory Embedding for Subspace Clustering" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.18, + 0.381, + 0.198 + ], + "angle": 0, + "content": "Yaroslava Lochman" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.18, + 0.553, + 0.198 + ], + "angle": 0, + "content": "Carl Olsson1,2" + }, + { + "type": "text", + "bbox": [ + 0.608, + 0.18, + 0.756, + 0.2 + ], + "angle": 0, + "content": "Christopher Zach1" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.205, + 0.53, + 0.224 + ], + "angle": 0, + "content": "1Chalmers University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.589, + 0.206, + 0.729, + 0.224 + ], + "angle": 0, + "content": "\\(^{2}\\)Lund University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.259, + 0.314, + 0.275 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.474, + 0.671 + ], + "angle": 0, + "content": "1 Clustering multiple motions from observed point trajectories is a fundamental task in understanding dynamic scenes. Most motion models require multiple tracks to estimate their parameters, hence identifying clusters when multiple motions are observed is a very challenging task. This is even aggravated for high-dimensional motion models. The starting point of our work is that this high-dimensionality of motion model can actually be leveraged to our advantage as sufficiently long trajectories identify the underlying motion uniquely in practice. Consequently, we propose to learn a mapping from trajectories to embedding vectors that represent the generating motion. The obtained trajectory embeddings are useful for clustering multiple observed motions, but are also trained to contain sufficient information to recover the parameters of the underlying motion by utilizing a geometric loss. We therefore are able to use only weak supervision from given motion segmentation to train this mapping. The entire algorithm consisting of trajectory embedding, clustering and motion parameter estimation is highly efficient. We conduct experiments on the Hopkins155, Hopkins12, and KT3DMoSeg datasets and show state-of-the-art performance of our proposed method for trajectory-based motion segmentation on full sequences and its competitiveness on the occluded sequences. Project page: https://ylochman.github.io/trajectory-embedding." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.704 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.851 + ], + "angle": 0, + "content": "Multi-model fitting is a classical problem in computer vision, and a typical example is the task of estimating individual motions from point trajectories observed in images containing multiple moving (and potentially deforming) objects. A trajectory is a sequence of tracked point coordinates extracted in different frames and is therefore also a point in a high-dimensional space. The commonly used union-of-subspace models [61] assume that point trajectories originating from the same object/motion are samples from the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.26, + 0.895, + 0.351 + ], + "angle": 0, + "content": "same (unknown) low dimensional subspace. The inference problem thus becomes that of simultaneously determining the parameters of underlying subspaces, the trajectory assignments to subspaces, and in practice also denoising of the data. Specifically, given an observation matrix \\(\\mathbb{M}_{2F\\times P}\\) of point trajectories" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.362, + 0.892, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathbb {M} _ {2 F \\times P} = \\left( \\begin{array}{c c c} | & & | \\\\ \\mathbf {x} _ {1} & \\dots & \\mathbf {x} _ {P} \\\\ | & & | \\end{array} \\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.895, + 0.527 + ], + "angle": 0, + "content": "where \\(F\\) is the number of frames and \\(P\\) is the number of points, the goal is to find a \\(P \\times P\\) permutation matrix \\(\\mathsf{P}_{\\pi}\\) that orders the trajectories into sub-matrices \\(\\{\\mathbf{X}_i\\}_{i=1}^c\\) corresponding to different motions. Each motion is represented by a \\(2F \\times k_i\\) basis matrix \\(\\mathsf{B}_i\\) whose columns span the subspace and therefore \\(\\mathbf{X}_i = \\mathsf{B}_i\\mathsf{C}_i^T\\), where \\(\\mathsf{C}_i\\) is a \\(k_i \\times P_i\\) shape matrix. Hence" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.538, + 0.892, + 0.558 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M} _ {2 F \\times P} \\mathrm {P} _ {\\pi} \\approx \\left[ \\begin{array}{l l l} \\mathrm {B} _ {1} \\mathrm {C} _ {1} ^ {\\top} & \\dots & \\mathrm {B} _ {c} \\mathrm {C} _ {c} ^ {\\top} \\end{array} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.895, + 0.719 + ], + "angle": 0, + "content": "where \\(\\sum_{i=1}^{c} P_i = P\\), and ranks \\(k_i\\) are assumed to be small. In this work we assume that the ranks \\(\\{k_1, \\ldots, k_c\\}\\) are known which is a common scenario under multiple rigid motions [49]. While estimating a single motion model \\(\\mathsf{B}_i\\) from a cluster of trajectories belonging to the same motion is relatively easy, the combination of cluster assignment and model estimation turns inference into a complicated chicken and egg problem. Further complications arise due to occlusion and partial observations which make it unlikely that we can observe all the full trajectories in all cases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "In this work we seek to learn a feature representation that allows us to solve the clustering problem separately from the motion estimation. For this purpose we train a simple feed forward network (with the overall structure depicted in Fig. 1) taking a trajectory as input and returning a corresponding feature embedding. Using a metric learning loss we ensure that the features are close if and only if corresponding trajectories belong to the same motions. Our approach allows us to input trajectories of varying lengths and different starting and ending times, making it possible to cluster trajectories using their features without completing any missing data. In summary our main contributions are:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.863, + 0.47, + 0.901 + ], + "angle": 0, + "content": "1This work was supported by the Wallenberg Artificial Intelligence, Autonomous Systems and Software Program (WASP), funded by Knut and Alice Wallenberg Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19092" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.101, + 0.878, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.327 + ], + "angle": 0, + "content": "Figure 1. Schematic overview of the proposed approach. Input trajectories are passed through the feature extractor to obtain \\(\\{\\mathbf{f}_i\\}\\). The subspace estimator generates the time-dependent basis using coordinate-MLP, infers basis coefficients from \\(\\mathbf{f}\\), and outputs subspace B, where each sub-matrix \\(\\mathsf{B}_t \\in \\mathbb{R}^{2 \\times 4}\\) is computed via linear combination of basis functions. \\(\\mathcal{L}_{\\mathrm{InfoNCE}}\\) is the contrastive loss that uses ground truth cluster assignments, \\(\\mathcal{L}_{\\mathrm{Residual}}\\) computes the trajectory-to-subspace geometric error, and \\(\\mathcal{L}_{\\mathrm{FeatDiff}}\\) computes the feature difference between input and reconstructed trajectory. For motion segmentation, the trajectories are clustered in a feature space using \\(\\{\\mathbf{f}_i\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.345, + 0.468, + 0.39 + ], + "angle": 0, + "content": "- We present a method to map point trajectories to discriminative feature representations, which can be seen as compact descriptions of the underlying generating motion." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.391, + 0.468, + 0.45 + ], + "angle": 0, + "content": "- We show that clustering of trajectories can be performed reliably solely using those features, achieving state-of-the-art performance without the need for joint motion estimation and cluster assignments." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.451, + 0.468, + 0.51 + ], + "angle": 0, + "content": "- The proposed feature extraction network is trained to be invariant with respect to random occlusions and pixel noise by utilizing a novel generative model that implicitly performs per-trajectory model estimation." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.511, + 0.468, + 0.556 + ], + "angle": 0, + "content": "- We present and evaluate different ways to use the proposed network, such as trajectory clustering, trajectory completion and subspace estimation." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.345, + 0.468, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.571, + 0.221, + 0.587 + ], + "angle": 0, + "content": "1.1. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.869 + ], + "angle": 0, + "content": "Within geometric vision the traditionally dominant model-fitting approaches are robust statistical methods based on RANSAC [15]. In case of low dimensional models the approach provides a simple way of effectively exploring the search space. Iterative RANSAC [48], which sequentially finds models and removes their supporting inliers, is a simple option that handles settings with multiple models. It has been observed that the greedy nature of the approach makes it inefficient when a large number of models is present [6, 12, 21, 62]. In contrast, MultiRANSAC [62] simultaneously handles multiple models using iteratively fused consensus sets, and CONSAC [25] increases the efficiency of sampling by learning a sampling strategy based on previous detections. Energy minimization is used in [6, 12, 21] to select models among random proposals that jointly explain the data well. In [7] the hard point-to-model assignment is dropped in favour of progressively determining dominant models which enables clustering in consensus space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As an alternative to consensus some works use preference [31, 32, 46]. The J-linkage approach [46] uses affini-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.436 + ], + "angle": 0, + "content": "ties between data and a set of models, followed by agglomerative clustering to assign the data to models. The T-linkage method [32] is based on a continuous relaxation of the binary assignment problem. In [31] the preference representation is used in a robust PCA formulation [10] while [33] uses a set cover approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Low dimensional geometric models have also been used to group longer point tracks, e.g. in [56] geometric two-view relations are used to create pairwise affinities for a spectral clustering stage. Segmentations computed from images pairs are combined through an averaging approach in [3, 4]. We note that for these local methods motion-specific properties should be observable across all image pairs in order to provide a largely coherent set of clusters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Subspace models are commonly occurring in geometric computer vision tasks. Early works like [38, 47] showed that structure from motion under affine camera models is essentially a subspace fitting problem. Further, shape deformations are modeled using a low dimensional subspace in [9]. The generalization to multiple independently moving objects leads to segmentation into multiple subspaces [8, 11, 17]. Approaches for handling degenerate and articulated cases were given in [45, 57, 58]. The general problem can be seen as an instance of subspace clustering which assumes that high dimensional data lies in a union of low-dimensional subspaces. The corresponding group of methods aim to simultaneously cluster the data points and estimate the underlying subspaces [1, 26], and may also perform matrix completion [14, 27, 50]. The generalized PCA approach (GPCA) [53-55] treats the union-of-subspaces as a zero-level-set of a higher degree polynomial which is fit to the data. A popular approach to the problem is sparse subspace clustering [5, 13], which has been shown to perform well on the benchmark provided in [49]. These methods are based on the notion of self-expressiveness, i.e. all data in a \\(d\\)-dimensional subspace are linear combinations of \\(d\\) points" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19093" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.152 + ], + "angle": 0, + "content": "from that subspace. This enables an efficient approach that is provably correct in the case of disjoint subspaces (see Section 2). The LRR approach [29, 30] uses a dictionary basis and enforces low rank patterns among the coefficients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.468, + 0.332 + ], + "angle": 0, + "content": "More recently, several deep learning methods have been proposed for solving the problem of unsupervised subspace clustering. The first work by Ji et al. [23] used deep autoencoder and introduced a self-expressive layer. Later, a closed form solution of the self-expressive layer was derived [43] which improved its efficiency. Alternatively, it was shown [51] that using over-complete representations increases robustness to noise and improves overall performance. Zhang et al. [59] directly learns a self-expressive data representation. The approach is highly scalable since it circumvents the need for a quadratic number of coefficients of the original self-expressive model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.468, + 0.485 + ], + "angle": 0, + "content": "While general subspace learning models typically do not use temporal information, this can be of great importance when analysing and predicting motion. For the reconstruction of deforming objects, Akhtar et al [2] noted the duality between trajectory space and shape space. The use of pre-defined trajectory-bases such as a DCT basis (motivated by the smoothness assumption of time-dependent trajectories [20, 37]) leads to a reduction in the number of unknowns [2]. Finally, a probabilistic framework for subspace clustering with temporal information is presented in [19]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.322, + 0.515 + ], + "angle": 0, + "content": "2. Overview of the Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.468, + 0.689 + ], + "angle": 0, + "content": "In this section we present our approach and the underlying architecture of our network. Our goal is to learn function \\( f_{\\theta} \\) that takes an observed point trajectory as input and yields a (relatively low-dimensional) feature embedding that is appropriate for trajectory grouping. For this purpose we adopt a small PointNet-style network [39]. One of the advantages of this type of network, apart from its simplicity, is the built-in invariance with respect to the permutation of the points in the input tensor. In order to obtain a suitable feature representation, we specify a number of desirable properties that we try to infuse through the architecture and training:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.749 + ], + "angle": 0, + "content": "- The features should be adequate for grouping/clustering of trajectories. For this reason we want a latent space with small within-cluster-distances and large between-cluster-distances. This is achieved using an InfoNCE loss." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.75, + 0.468, + 0.855 + ], + "angle": 0, + "content": "- Inspired by traditional motion segmentation formulations, we aim for features to encode meaningful \"geometric information\" about the underlying motion subspaces. For this purpose, we design a decoding module that estimates a basis B for the underlying motion subspace that the trajectory belongs to. This prevents input memorization and greatly improves generalization to unseen data." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- Despite input tracks being of varying lengths we opt for a feature embedding of a fixed size. This simplifies treatment of problem instances with missing data and trajec" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.891, + 0.182 + ], + "angle": 0, + "content": "tories starting and ending at different times. The decoding module therefore takes an extra variable, specifying at which time instance the rows of the basis matrix B should be sampled. In this way the model allows us to sub-sample and predict motion information beyond the trajectory samples." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.183, + 0.892, + 0.318 + ], + "angle": 0, + "content": "- Finally, feature embeddings and intermediate predictions are preferred to behave smoothly with respect to added input noise. To achieve this we construct a \"noise-free\" version of the observed trajectory from the predicted basis, which we require to be close to the original trajectory through a residual loss. Additionally, this reconstructed trajectory is subsequently fed into the network to obtain a corresponding feature vector, which is also required to be close to the original feature embedding." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.892, + 0.665 + ], + "angle": 0, + "content": "Because of the properties above we obtain a trajectory-feature from which we can extract the full motion subspace. Thus we essentially learn a function that maps a single trajectory to a subspace basis. For a general collection of subspaces, where trajectories can belong to multiple of them, this may not be possible (using a purely geometric approach). A simple example is line fitting to image points, where points in the intersection of two models can belong to any of them. In such cases geometry alone is not enough to resolve ambiguous assignments but other cues such as spatial smoothness of assignments [21] have to be used. However, in high-dimensional settings with low-dimensional subspaces the situation is somewhat different. In these cases a typical assumption is that of disjoint subspaces [13], that is, two subspaces only intersect at the origin, ensuring that there is a function that takes (non-zero) trajectories to their corresponding subspaces. We remark that while this assumption is sufficient for our purposes, it is not necessary. Since our method learns a data-driven mapping from point tracks to subspaces, it is able to focus on other trajectory properties than geometric ones. Therefore it can potentially handle the case of intersecting subspaces as long as the trajectory assignment is unambiguous." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.891, + 0.764 + ], + "angle": 0, + "content": "The disjoint subspace assumption We conclude this section by illustrating that the disjoint subspace assumption is valid in practical settings by analysing the data in the Hopkins 155 dataset. Figure 2 shows that re-using subspaces to explain trajectories in the other clusters generally leads to higher errors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Formally, two linear subspaces \\(S_{i}\\) and \\(S_{j}\\) are disjoint if \\(\\dim(S_{i} \\oplus S_{j}) = \\dim(S_{i}) + \\dim(S_{j})\\). Hence if \\(\\mathsf{B}_{i}\\) and \\(\\mathsf{B}_{j}\\) contain bases for the two subspaces, the rank of the concatenated matrix, \\(\\eta_{ij} = \\mathrm{rank}((\\mathsf{B}_i \\mathsf{B}_j))\\), fulfills \\(\\eta_{ij} < \\mathrm{rank}(\\mathsf{B}_i) + \\mathrm{rank}(\\mathsf{B}_j)\\) if the subspaces intersect. We extracted subsequences of varying length from the Hopkins 155 dataset and used the ground-truth clusters to generate the corresponding subspaces — for all pairs \\(\\{\\mathsf{B}_i, \\mathsf{B}_j\\}\\), where \\(i \\neq j\\) and \\(\\mathrm{shape}(\\mathsf{B}_i) = \\mathrm{shape}(\\mathsf{B}_j)\\), the rank \\(\\eta_{ij}\\) of the concatenated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19094" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.468, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.211, + 0.471, + 0.268 + ], + "angle": 0, + "content": "Figure 2. RMS normalized error of the trajectory clusters with respect to the subspaces in Hopkins155 dataset (subsequences of length \\( F = 40, 50, \\) and 60). An \\( (i,j) \\) cell corresponds to the trajectory cluster of the \\( i^{\\text{th}} \\) motion and the subspace of \\( j^{\\text{th}} \\) motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.294, + 0.471, + 0.356 + ], + "angle": 0, + "content": "matrix was found to be \\( 8^{2} \\). This suggests that up to numerics the subspaces do not intersect. In the following Section we describe the specific architecture and the employed losses in our approach in more detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.37, + 0.383, + 0.387 + ], + "angle": 0, + "content": "3. Architecture and Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.562 + ], + "angle": 0, + "content": "The schematic overview of the proposed architecture and training losses is shown in Fig. 1. Using the hypothesis that most observed trajectories come from a unique motion model, we aim to learn a function \\( f_{\\theta} \\) mapping a point trajectory directly to a representation of the corresponding motion model. We interpret the output \\( f_{\\theta}(\\mathbf{x}) \\) as an abstract description of the generating motion, from which a concrete representation to measure geometric consistency can be extracted. We choose linear (low-dimensional) subspaces as the concrete realizations of motion models. Specifically, we learn mappings \\( f_{\\theta} \\) and \\( g_{\\phi} \\) using training data such that" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.562, + 0.469, + 0.592 + ], + "angle": 0, + "content": "1. \\( f_{\\theta} \\) maps a trajectory of varying length to a fixed-length feature embedding, and" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.592, + 0.469, + 0.623 + ], + "angle": 0, + "content": "2. \\( g_{\\phi} \\) converts this feature vector for a given query time into the corresponding rows of the subspace representation." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.562, + 0.469, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.633, + 0.26, + 0.647 + ], + "angle": 0, + "content": "3.1. Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.854 + ], + "angle": 0, + "content": "To construct a discriminative function \\( f_{\\theta}(\\cdot) \\), that is able to infer the motion from an input point trajectory, we first leverage the ability of the neural networks to compress high-dimensional complex data space into a lower-dimensional latent space with task specific invariance properties, from which it is easier to infer the output. A feature extractor for the point trajectories \\( f_{\\theta}: \\mathbf{x} \\mapsto \\mathbf{f} \\in \\mathbb{R}^{d} \\) is therefore trained to produce invariant feature representations using ground truth clusters. In particular, an invariance with respect to the measurement noise and the variations within the motion clusters is of our interest. We therefore use a contrastive loss, namely a variant of the InfoNCE loss [36], to conduct metric learning, leading to a spread of feature embeddings" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.901 + ], + "angle": 0, + "content": "2The matrix rank is computed as the number of singular values greater than \\(\\delta \\sigma_{1}\\), where \\(\\sigma_{1}\\) is the largest singular value, \\(\\delta = 2F\\varepsilon\\), and \\(\\varepsilon \\approx 10^{-7}\\) is the numerical zero for torch.float32." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.847, + 0.108 + ], + "angle": 0, + "content": "in order to improve feature-based motion clustering," + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.116, + 0.892, + 0.194 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {I n f o N C E}} = - \\frac {1}{| \\mathcal {D} |} \\sum_ {(i, j, l, k) \\in \\mathcal {D}} \\log \\left(\\frac {p _ {i j}}{p _ {i j} + p _ {l k}}\\right), \\\\ p _ {i j} = \\exp \\left(- \\frac {\\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| _ {2} ^ {2}}{T}\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.203, + 0.892, + 0.295 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i = f_{\\theta}(\\mathbf{x}_i)\\) and \\(\\mathcal{D}\\) is the set of quadruples \\((i,j,l,k)\\) where \\((i,k)\\) belong to the same cluster and \\((l,k)\\) are random pairs, and \\(T\\) is the temperature (hyper-)parameter, which we fix to 1. \\(\\mathcal{L}_{\\mathrm{InfoNCE}}\\) is an instance of a supervised contrastive loss and is tightly connected to the triplet loss used in metric learning [24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.476 + ], + "angle": 0, + "content": "As mentioned earlier, the proposed architecture resembles PointNet [39] with several modifications. The employed encoder consists of 1D temporal convolutional layers applied to each trajectory independently, with x and y values representing two input channels. A global max-pooling in a temporal domain is then applied on the time-dependent representations, followed by a small MLP and an \\( L_{2} \\) normalization. The features are therefore lying on the \\( d - 1 \\)-dimensional unit sphere. Note that we do not apply pooling across the points (global context layer) due to the disjoint subspace assumption. For the experiments, we set the dimensionality of the latent space to \\( d = 128 \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.697, + 0.501 + ], + "angle": 0, + "content": "3.2. Subspace Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.705 + ], + "angle": 0, + "content": "The latent vector \\( f_{\\theta}(\\mathbf{x}) \\) ideally identifies a motion model, but for geometric consistency this feature embedding requires conversion into a suitable mathematical motion model. We assume that observed point trajectories are generated by low-rank subspace models, and therefore the feature embedding is mapped by a subsequent network \\( g_{\\phi} \\) to the corresponding matrix representation of the associated subspace. Specifically, the mapping \\( g_{\\phi}: \\mathcal{F} \\times \\mathbb{R} \\rightarrow \\mathbb{R}^{2 \\times r} \\) produces rows \\( \\mathsf{B}_t \\) for the given motion model \\( \\mathbf{f} = f_{\\theta}(\\mathbf{x}) \\) and the time index \\( t \\). Ideally the input trajectory \\( \\mathbf{x} \\) lies in the predicted subspace; if the trajectory \\( \\mathbf{x} \\) is consisting of (2D) points observed at times \\( t_1, \\ldots, t_F \\), then the output subspace is" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.714, + 0.892, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathsf {B} = \\left( \\begin{array}{c} g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {1}\\right) \\\\ \\vdots \\\\ g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {F}\\right) \\end{array} \\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.78, + 0.847, + 0.796 + ], + "angle": 0, + "content": "and the (squared) residual is given in closed-form as" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.806, + 0.892, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {c}} \\| \\mathbf {x} - \\mathrm {B c} \\| _ {2} ^ {2} = \\| \\mathbf {x} - \\mathrm {B B} ^ {\\dagger} \\mathbf {x} \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.839, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{c}\\) is the coefficient vector and \\(\\mathsf{B}^{\\dagger} = (\\mathsf{B}^{\\top}\\mathsf{B})^{-1}\\mathsf{B}^{\\top}\\). If the network identified the subspace correctly, then the residual is in the order of the observation noise level (and therefore small)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19095" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "Subspace basis functions The network \\( g_{\\phi} \\) can be a small but generic MLP. Alternatively, we may leverage existing domain knowledge and choose the structure of \\( g_{\\phi} \\) accordingly. In particular, we assume that all motions are linear combination of a finite, but continuous time trajectory basis \\( \\{h^j\\} \\). The basis functions \\( h^j \\) can be fully fixed (such as a cosine basis with pre-determined frequencies and phase shifts [2]), parametric (such as the cosine basis with trainable frequencies and phase shifts [44]) or entirely non-parametric (i.e. represented by a generic MLP). Our choice of basis functions is a damped version of a cosine basis and is given as follows," + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.283, + 0.47, + 0.306 + ], + "angle": 0, + "content": "\\[\nh _ {\\psi} ^ {j} (t) = e ^ {- \\left(\\alpha_ {j} (t - \\mu_ {j})\\right) ^ {2}} \\cos (\\beta_ {j} t + \\gamma_ {j}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.315, + 0.47, + 0.346 + ], + "angle": 0, + "content": "where \\(\\psi = (\\mu_j, \\alpha_j, \\beta_j, \\gamma_j)_j\\) are trainable parameters. The mapping \\(g_{\\phi}\\) can therefore be written as a composition" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.356, + 0.469, + 0.373 + ], + "angle": 0, + "content": "\\[\ng _ {\\phi} (\\mathbf {f}, t) = \\omega_ {\\zeta} (\\mathbf {f}) \\mathbf {h} _ {\\psi} (t), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.386, + 0.47, + 0.478 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_{\\psi}(t) = (h_{\\psi}^{j}(t))_{j}\\), and the trainable function \\(\\omega_{\\zeta}\\) extracts the basis coefficients from \\(\\mathbf{f}\\). If we have \\(N\\) basis functions (i.e. \\(\\dim (\\mathbf{h}_{\\psi}(t)) = N\\)), then \\(\\omega_{\\zeta}(\\mathbf{f})\\) is a \\(2r\\times N\\) matrix. The total set of trainable parameters in the subspace estimation network is \\(\\phi = (\\zeta ,\\psi)\\). Finally, the network prediction can be compactly written as" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.489, + 0.469, + 0.506 + ], + "angle": 0, + "content": "\\[\n\\mathsf {B} = \\mathrm {H} _ {\\psi} (\\mathbf {t}) \\Omega_ {\\zeta} \\left(f _ {\\theta} (\\mathbf {x})\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.516, + 0.356, + 0.534 + ], + "angle": 0, + "content": "where \\(\\Omega_{\\zeta}(\\mathbf{f}) = \\mathrm{reshape}_{2N\\times r}\\left(\\omega_{\\zeta}(\\mathbf{f})\\right)\\) and" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.543, + 0.469, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H} _ {\\psi} (\\mathbf {t}) = \\left( \\begin{array}{l l} \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} \\\\ \\vdots & \\\\ \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} \\end{array} \\right) \\in \\mathbb {R} ^ {2 F \\times 2 N}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.47, + 0.732 + ], + "angle": 0, + "content": "We set the number of subspace basis functions to \\( N = 64 \\) in our experiments. Since the predicted subspace matrix B augments the feature representation (see Sec. 3.3 below), we enforce a unique subspace representation by fixing the top block of B to the identity matrix (by multiplying with the respective inverse)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Time dependence and time invariance The network represented by \\( g_{\\phi} \\) can be interpreted as taking a \"feature\" token \\( \\mathbf{f} \\) and a time coordinate \\( t \\) as input, and is therefore somewhat analogous to the coordinate-MLP employed for implicit neural primitives such as [35, 44] or time series modeling [16]. The continuous time basis \\( \\{h^j\\} \\) can be seen as domain-specific equivalent to positional coding, but its design is based on the expected motion patterns and does not necessarily meet the requirements of generic positional coding [40, 44, 60]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "The formulation above is intentionally not time-invariant: the same trajectory observed at different starting times may predict different motion models. This is usually a desired behavior for motion segmentation, e.g. two cars traveling the same path at different time form different motion segments. As with CNNs applied to images, low-level trajectory features are extracted by convolutional layers, and time-dependence is incorporated via final fully connected and pooling layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.236, + 0.605, + 0.252 + ], + "angle": 0, + "content": "3.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.259, + 0.892, + 0.307 + ], + "angle": 0, + "content": "The total training loss is the weighted sum of a contrastive loss \\(\\mathcal{L}_{\\mathrm{InfoNCE}}\\) applied on \\(\\{\\mathbf{v}_i\\}\\), where \\(\\mathbf{v}_i = (\\mathrm{flatten}(\\mathsf{B}_i)^\\top \\mathbf{f}_i^\\top)^\\top\\), the reconstruction loss \\(\\mathcal{L}_{\\mathrm{Residual}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.314, + 0.892, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {R e s i d u a l}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| \\mathbf {x} _ {j} - \\tilde {\\mathbf {x}} _ {j} \\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.365, + 0.702, + 0.379 + ], + "angle": 0, + "content": "and the feature difference loss" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.386, + 0.892, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {F e a t D i f f}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| f _ {\\theta} (\\mathbf {x} _ {j}) - f _ {\\theta} (\\tilde {\\mathbf {x}} _ {j}) \\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.651 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}_j = \\mathsf{B}(\\mathbf{x}_j)\\mathsf{B}(\\mathbf{x}_j)^\\dagger\\mathbf{x}_j\\) is the projection of \\(\\mathbf{x}_j\\) onto \\(\\mathsf{B}(\\mathbf{x}_j)\\). In short, \\(\\mathcal{L}_{\\mathrm{InfoNCE}}\\) forces both an approximate invariance of the feature extractor \\(f_{\\theta}\\) with respect to the variations within clusters and smoothness of the coordinate-MLP \\(g_{\\phi}\\), \\(\\mathcal{L}_{\\mathrm{Residual}}\\) ensures geometric consistency of the subspaces with respect to input trajectories, and \\(\\mathcal{L}_{\\mathrm{FeatDiff}}\\) forces an approximate invariance of \\(f_{\\theta}\\) with respect to the pixel noise and can also be seen as enforcing smoothness of \\(f_{\\theta}\\). We train the network in two stages. First, we pre-train \\(f_{\\theta}\\) with \\(\\mathcal{L}_{\\mathrm{InfoNCE}}\\) until convergence. Then we add \\(g_{\\phi}\\) and train it together with \\(f_{\\theta}\\) using the total loss. To enforce the approximate invariance with respect to random detection failures, we synthesize missing points during training by applying dropout to the input sequences." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.714, + 0.675 + ], + "angle": 0, + "content": "3.4. Trajectory Completion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Let \\(B_{\\theta, \\phi}(\\mathbf{x}, \\mathbf{t})\\) provide the subspace for an input trajectory \\(\\mathbf{x} \\in \\mathbb{R}^{2F}\\) and time vector \\(\\mathbf{t} = (t_1, \\dots, t_F)^\\top\\). Assume now that \\(\\mathbf{x}\\) is only partially observed, and let \\(\\mathbf{w} \\in \\{0, 1\\}^{2F}\\) be the visibility mask. Further, let \\(\\bar{\\mathbf{w}} := \\mathbf{1} - \\mathbf{w}\\). The goal of the trajectory completion is to estimate missing values \\(\\bar{\\mathbf{x}}\\) that minimize the following objective" + }, + { + "type": "equation", + "bbox": [ + 0.519, + 0.781, + 0.891, + 0.803 + ], + "angle": 0, + "content": "\\[\nJ (\\bar {\\mathbf {x}}) := \\left\\| (I - B (\\bar {\\mathbf {x}}) B (\\bar {\\mathbf {x}}) ^ {\\dagger}) (w \\odot x + \\bar {w} \\odot \\bar {x}) \\right\\| ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathsf{B}(\\bar{\\mathbf{x}}) = B_{\\theta ,\\phi}(\\mathbf{w}\\odot \\mathbf{x} + \\bar{\\mathbf{w}}\\odot \\bar{\\mathbf{x}},\\mathbf{t})\\) . Back-propagation for gradient-based determination of \\(\\bar{\\mathbf{x}}\\) is in principle possible but computationally expensive. Alternatively, we rely on the approximate invariance of the feature extractor \\(f_{\\theta}\\) with respect to the length of the input sequence and extrapolate the output subspace in time using \\(g_{\\phi}(\\mathbf{f},t)\\) . More" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19096" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.93, + 0.263 + ], + "angle": 0, + "content": "
Method2 motionsHopkins155AllHopkins12KT3DMoSeg
MeanMedianTimeMeanMedianTimeMeanMedianTimeMeanMedianMeanMedian
RANSAC [48]5.561.18175ms22.9422.03258ms9.763.21194ms----
GPCA [54]4.590.38324ms28.6628.26738ms10.342.54417ms--34.6033.95
MSL [45]4.140.0011h 4m8.231.761d 23h5.030.0019h 11m----
LSA [57]3.450.597.58s9.732.3315.96s4.940.909.47s--38.3038.58
ALC5 [42]3.030.00-6.261.02-3.760.265m 15s3.810.1724.3119.04
ALCsp [42]2.400.43-6.690.67-3.370.496m 11s1.281.07--
LRR [30]4.100.22-9.890.56-5.410.531.1s--33.6736.01
SSC [13]0.820.00-2.450.20-2.450.20920ms--33.8833.54
RSIM [22]0.780.00-1.770.28-1.010.00176ms0.680.70--
MultiConsensus [7]------4.40-40ms----
Ours0.630.07ms0.600.010ms0.620.09ms5.122.045.850.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.274, + 0.893, + 0.304 + ], + "angle": 0, + "content": "Table 1. Motion segmentation results on the Hopkins155 [49], Hopkins12 [41] and KT3DMoSeg [56] datasets. The evaluation metrics is classification error \\((\\%)\\). ' -' indicates that the results were not reported / no code available." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.471, + 0.417 + ], + "angle": 0, + "content": "specifically, the subspace of \\(\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}\\) can be estimated from \\(\\mathbf{x}_{\\mathrm{vis}}\\)-a fully visible sub-trajectory of the trajectory \\(\\mathbf{x}\\) as long as the feature representation \\(\\mathbf{f}\\) can be inferred (i.e. \\(\\mathbf{x}_{\\mathrm{vis}}\\) is sufficiently long). If we fix this estimate \\(\\mathbf{B}\\) and let \\(\\mathsf{W} := \\mathrm{diag}(\\mathbf{w})\\), \\(\\bar{\\mathsf{W}} := \\mathrm{diag}(\\bar{\\mathsf{w}})\\), the new objective can be stated as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.429, + 0.47, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\tilde {J} (\\bar {\\mathbf {x}}) := \\left\\| \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\mathbf {x} + \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\bar {\\mathbf {x}} \\right\\| ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.466, + 0.332, + 0.484 + ], + "angle": 0, + "content": "and the minimizer \\(\\bar{\\mathbf{x}}^*\\) of \\(\\tilde{J}\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.498, + 0.47, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathbf {x}} ^ {*} = - \\left(\\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {\\mathbf {w}}\\right) ^ {\\dagger} \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\mathbf {W x} =: \\mathrm {A} (\\mathrm {B}) \\mathbf {x}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.595 + ], + "angle": 0, + "content": "It is now possible to refine the estimated subspace using \\(\\bar{\\mathbf{x}}^*\\) by feeding \\(\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}^*\\) into the network. Overall, we apply the following iterative procedure for trajectory completion," + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.608, + 0.47, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{r l} \\mathsf {B} _ {0} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {x} _ {\\text {v i s}}, \\mathbf {t}) \\\\ \\bar {\\mathbf {x}} _ {i + 1} & \\leftarrow \\mathsf {A} (\\mathsf {B} _ {i}) \\mathbf {x} \\\\ \\mathsf {B} _ {i + 1} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {w} \\odot \\mathbf {x} + \\bar {\\mathbf {w}} \\odot \\bar {\\mathbf {x}} _ {i + 1}, \\mathbf {t}). \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.691, + 0.273, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.689, + 0.465, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.852, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Figure 3. Histograms of the classification errors for the groups with (left) 2 and (right) 3 motions in Hopkins155 dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.325, + 0.633, + 0.343 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.351, + 0.892, + 0.426 + ], + "angle": 0, + "content": "The proposed network was trained on the two motion segmentation datasets: Hopkins155 [49] and KT3DMoSeg [56], yielding two different sets of weights. In our experiments, we used NVIDIA GeForce RTX 3080, CUDA 11.5, PyTorch 1.11." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.426, + 0.893, + 0.624 + ], + "angle": 0, + "content": "To obtain groups of trajectories, it is sufficient to perform a partial forward pass through the (trained) feature extractor \\( f_{\\theta} \\) and cluster in the feature space. We use hierarchical clustering as it provides a meaningful framework for selecting the number of clusters, i.e. motions. The subspaces can then be fit to each cluster. If the goal is to estimate the subspace for a single input trajectory, a full forward pass through \\( f_{\\theta} \\) and \\( g_{\\phi} \\) should be performed. For motion segmentation with missing data, the largest fully visible blocks of trajectories (but ignoring single occluded points) can be used to estimate the initial subspaces B, followed by trajectory completion as outlined in (15). The completed trajectories can then be used to cluster in the feature space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.704, + 0.648 + ], + "angle": 0, + "content": "4.1. Motion Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.894, + 0.853 + ], + "angle": 0, + "content": "We evaluate the performance of the proposed network in terms of motion segmentation on Hopkins155 [49], Hopkins12 [41], and KT3DMoSeg [56] datasets. We first compare the proposed method to the state-of-the-art algorithms that similarly assume rank-4 of B: sequential RANSAC [48], GPCA [54], MSL [45], LSA [57] (with the global subspace dimensionality \\( D = 4n \\) for \\( n \\) clusters), \\( \\mathrm{ALC}_5 \\) and \\( \\mathrm{ALC}_{\\mathrm{sp}} \\) [42], LRR [30], SSC [13], RSIM [22], and MultiConsensus [7]. We then also compare our method to the other approaches making different assumptions on the underlying motion model [3, 4, 28, 56]. The metrics used to evaluate clustering performance is the classification error (\\%) which is standard for this benchmark [49]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Results on Hopkins155 dataset Hopkins155 [49] is the commonly used benchmark for trajectory-based motion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19097" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.35, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.09, + 0.613, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.091, + 0.877, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 4. The approximate invariance measured as an angle between features extracted from the reference trajectory and its perturbed version using (left) Gaussian pixel noise, (middle) perturbation within subspace, and (right) uniform occlusions. The gradually increasing level of corruption is shown on the x-axis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.306, + 0.471, + 0.426 + ], + "angle": 0, + "content": "segmentation. The dataset consists of 155 sequences with 63-556 point trajectories generated by two or three motions in the scene. Each trajectory is fully visible, and the length varies from 15 to 100 frames. The averaged results are shown in the left block of Table 1, and the histograms of classification errors are shown in Fig. 3. Our network is very fast and produces the most accurate segmentations on fully visible trajectories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Results on Hopkins12 dataset Hopkins12 [41] is a commonly used dataset for evaluating subspace clustering with missing data. It consists of 12 corrupted sequences from Hopkins155 dataset, where the corruption was done manually by labeling entries of trajectories as missing. The corruption rate of the observation matrix M varies from \\(1\\%\\) to \\(23\\%\\). For \\(\\mathrm{ALC}_5\\) and \\(\\mathrm{ALC_{sp}}\\) [42], we report \\(l^1\\)-based entry completion results. The results are shown in the middle block of Table 1. The proposed network shows promising results. Note that the maximum corruption rate of the individual trajectories in the Hopkins12 dataset is \\(93\\%\\) and, unlike all the state-of-the-art approaches, the proposed trajectory completion method is not global, i.e. it works with individual trajectories. Therefore, if a provided individual sub-trajectory does not contain sufficient relevant information for motion estimation (which becomes more likely as the trajectory gets shorter), incorporating information from all tracks becomes necessary to disambiguate the true motion. Utilizing such global trajectory information is left for future research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results on KT3DMoSeg dataset KT3DMoSeg [56] is an adapted version of the KITTI dataset [18, 34] containing 22 sequences with 158-1018 point trajectories and annotated segmentations. It is more challenging than Hopkins155 in several aspects. There is a dominant motion corresponding to static \"background\" scene which makes cluster distribution unbalanced. The trajectories are also highly occluded: the corruption rate of M goes up to \\(54\\%\\), and for the individual trajectories — up to \\(75\\%\\). Lastly, the sequences have shorter length of 10 to 20 frames. The results for this dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.306, + 0.891, + 0.336 + ], + "angle": 0, + "content": "are shown in the right block of Table 1 indicating that the proposed approach outperforms previous methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.893, + 0.554 + ], + "angle": 0, + "content": "Methods with different assumptions We also compared our approach to the other methods making different assumptions. Two-perspective-view (TPV) [28] assumes perspective cameras and uses pairwise epipolar relations. Spectral clustering methods proposed by [56]: kernel addition (KerAdd), co-regularization (CoReg), and subset constrained (Subset) spectral clustering also look at pairwise relations, but use fundamental matrices, affine matrices or homographies to compute the pairwise affinities. SYNCH [3] and MODE [4] similarly fit multiple fundamental matrices to each pair of images, but perform averaging. They assume that only pairwise matches are given. These methods are not directly comparable to ours, however the proposed method proves to be competitive as evidenced in Table 2." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.56, + 0.892, + 0.687 + ], + "angle": 0, + "content": "
MethodHopkins155 MeanHopkins12 Med.KT3DMoSeg MeanMed.
TPV [28]2.34----
KerAdd [56]0.360.110.08.311.02
CoReg [56]0.460.060.07.920.75
Subset [56]0.310.060.08.080.71
SYNCH-tracks [3]3.675.460.57--
SYNCH [3]1.193.190.28--
MODE [4]1.374.330.38--
Our method0.625.122.045.850.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.697, + 0.891, + 0.724 + ], + "angle": 0, + "content": "Table 2. Comparison to the other methods making different assumptions. TPV [28] has no code publicly available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.77, + 0.758 + ], + "angle": 0, + "content": "4.2. Approximate Invariances of \\( f_{\\theta} \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We experimentally validate the approximate invariance of the trained feature extractor \\( f_{\\theta} \\) with respect to different types of synthetic data corruption: (1) Gaussian pixel noise with the standard deviation \\( \\sigma \\) gradually increasing up to 10 px, (2) perturbations within the subspace, where the trajectories are first corrupted by Gaussian pixel noise with \\( \\sigma \\) up to 300 px and then projected back to the underlying subspace, (3) synthesized uniform occlusions with missing data levels going up to 50%, and (4) tracking failures, where" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19098" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.35, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.091, + 0.613, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.092, + 0.875, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.895, + 0.283 + ], + "angle": 0, + "content": "Figure 5. The approximate invariance measured as an angle between features extracted from the reference trajectory and its shortened version for (left) shorter sequences with \\( F = 15 \\ldots 24 \\), (middle) mid-length sequences with \\( F = 25 \\ldots 31 \\), and (right) longer sequences with \\( F = 35 \\ldots 61 \\). The gradually increasing percentage of sequence length reduction is shown on the x-axis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.471, + 0.413 + ], + "angle": 0, + "content": "the rearmost points are gradually excluded from the trajectories. All types of corruption are applied on the training sequences of Hopkins155 dataset. The invariances are measured using the angle between the features extracted from original and corrupted trajectories. The smaller the angle is, the more invariant the feature is to the certain type of data corruption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.474, + 0.687 + ], + "angle": 0, + "content": "Results for the pixel noise, perturbations within the subspace, and uniform occlusions are shown in Fig. 4. Note that the features are \\( L_{2} \\)-normalized in the \\( d \\)-dimensional space for \\( d = 128 \\). In such a high-dimensional space the two randomly picked vectors are likely to be almost orthogonal [52]. The network proves to be approximately invariant hence robust to these types of corruption, and especially to the pixel noise. Since 8-10 px is quite large (for the images of sizes \\( 320 \\times 240 \\), \\( 640 \\times 480 \\), or \\( 720 \\times 480 \\)), this result also shows robustness of \\( f_{\\theta} \\) to occasional outlying points in the trajectory. The results for the synthesized tracking failures are shown in Fig. 5, where a single curve corresponds to a certain initial trajectory length \\( F \\). The angles are slightly higher for shorter sequences, which is expected since they contain less information. Note that during training we do not force the network to produce the same feature representations for the trajectories of the same point starting at the same time but having different length." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.706, + 0.465, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.869, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Mean pixel error of recovering from random occlusions for the corruption rates of (left) \\(10 - 40\\%\\) and (right) \\(50 - 80\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.713, + 0.324 + ], + "angle": 0, + "content": "4.3. Trajectory Completion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.516 + ], + "angle": 0, + "content": "We further evaluate the completion accuracy of the proposed method on the trajectories from Hopkins155 dataset with synthesized uniform occlusions. The performance is measured using the mean pixel error between the ground truth (i.e., detected/tracked) values of the synthetically occluded points and the predicted values. The error as a function of iterations for different corruption rates is shown in Fig. 6, where we split the results into a minor and a major corruption group. From the results we conclude that it is sufficient to run the algorithm (15) for 3 to 5 iterations for minor corruptions which takes 1s on average, and 5 to 20 iterations for major corruptions which takes 3s on average." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.538, + 0.619, + 0.554 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.892, + 0.763 + ], + "angle": 0, + "content": "In this paper we propose to learn a mapping from point trajectories to feature vectors in an embedding space, which fully identify the generating motion. We illustrate that these feature vectors can be directly used for subspace clustering, thereby eliminating the need for simultaneous estimation of assignments and subspace models. The proposed training loss ensures not only the utility of the obtained embedding for clustering, but also forces the feature vector to encode the actual motion parameters, which are retrieved by a dedicated coordinate-MLP. Comparisons to baselines on Hopkins155 and KT3DMoSeg highlight the efficiency and accuracy of our method, providing top clustering results with running time below 10ms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this work we focus on the linear subspace framework as sole motion model to explain observed trajectories. One direction for future work is how to incorporate additional and/or different prior assumptions on the motion model into our method. Further, the linear subspace assumption is strongly tied to affine cameras, which only approximate the more realistic pinhole camera model. Therefore extending this work to more general camera models is a challenging direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19099" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopoulos, and Prabhakar Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. In Proceedings of the 1998 ACM SIGMOD international conference on Management of data, pages 94-105, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.47, + 0.299 + ], + "angle": 0, + "content": "[3] Federica Arrigoni and Tomas Pajdla. Motion segmentation via synchronization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.301, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[4] Federica Arrigoni and Tomas Pajdla. Robust motion segmentation from pairwise matches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 671-681, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.47, + 0.4 + ], + "angle": 0, + "content": "[5] Liang Bai and Jiye Liang. Sparse subspace clustering with entropy-norm. In International conference on machine learning, pages 561-568. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[6] Daniel Barath and Jiri Matas. Progressive-x: Efficient, anytime, multi-model fitting algorithm. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3780-3788, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[7] Daniel Barath, Denys Rozumnyi, Ivan Eichhardt, Lerente Hajder, and Jiri Matas. Finding geometric models by clustering in the consensus space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5424, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.572 + ], + "angle": 0, + "content": "[8] T.E. Boult and L. Gottesfeld Brown. Factorization-based segmentation of motions. In Proceedings of the IEEE Workshop on Visual Motion, pages 179-186, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[9] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.684 + ], + "angle": 0, + "content": "[10] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? J. ACM, 58(3), 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[11] J. Costeira and T. Kanade. A multi-body factorization method for motion analysis. In Proceedings of IEEE International Conference on Computer Vision, pages 1071-1076, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[12] A. Delong, A. Osokin, H. Isack, and Y. Boykov. Fast approximate energy minimization with label costs. International Journal of Computer Vision, 96(1):1-27, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE transactions on pattern analysis and machine intelligence, 35, 2013. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[14] Jicong Fan and Madeleine Udell. Online high rank matrix completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8690-8698, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[16] Elizabeth Fons, Alejandro Sztrajman, Yousef El-Laham, Alexandros Iosifidis, and Svitlana Vyetrenko. Hypertime: Implicit neural representation for time series. arXiv preprint arXiv:2208.05836, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[17] C.W. Gear. Feature grouping in moving objects. In Proceedings of 1994 IEEE Workshop on Motion of Non-rigid and Articulated Objects, pages 214-219, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[19] Behnam Gholami and Vladimir Pavlovic. Probabilistic temporal subspace clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3066-3075, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[20] Paulo FU Gotardo and Aleix M Martinez. Computing smooth time trajectories for camera and deformable shape in structure from motion with occlusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(10):2051-2065, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[21] Hossam Isack and Yuri Boykov. Energy-based geometric multi-model fitting. International journal of computer vision, 97(2):123-147, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[22] Pan Ji, Mathieu Salzmann, and Hongdong Li. Shape interaction matrix revisited and robustified: Efficient subspace clustering with corrupted and incomplete data. In Proceedings of the IEEE International Conference on computer vision, pages 4687-4695, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[23] Pan Ji, Tong Zhang, Hongdong Li, Mathieu Salzmann, and Ian Reid. Deep subspace clustering networks. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.733 + ], + "angle": 0, + "content": "[25] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[26] Hans-Peter Kriegel, Peer Kroger, and Arthur Zimek. Clustering high-dimensional data: A survey on subspace clustering, pattern-based clustering, and correlation clustering. Acm transactions on knowledge discovery from data (tkdd), 3(1): 1-58, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[27] Connor Lane, Ron Boger, Chong You, Manolis Tsakiris, Benjamin Haeffele, and Rene Vidal. Classifying and comparing approaches to subspace clustering with missing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Zhuwen Li, Jiaming Guo, Loong-Fah Cheong, and Steven Zhiying Zhou. Perspective motion segmentation via" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19100" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "collaborative clustering. In Proceedings of the IEEE international conference on computer vision, pages 1369-1376, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[29] Guangcan Liu and Shuicheng Yan. Latent low-rank representation for subspace segmentation and feature extraction. In 2011 International Conference on Computer Vision, pages 1615-1622, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[30] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.469, + 0.302 + ], + "angle": 0, + "content": "[31] Luca Magri and Fusiello Andrea. Robust multiple model fitting with preference analysis and low-rank approximation. In Proceedings of the British Machine Vision Conference 2015, pages 20-1, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[32] Luca Magri and Andrea Fusiello. T-linkage: A continuous relaxation of j-linkage for multi-model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3954-3961, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.359, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[33] Luca Magri and Andrea Fusiello. Multiple model fitting as a set coverage problem. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3318-3326, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.58 + ], + "angle": 0, + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.469, + 0.664 + ], + "angle": 0, + "content": "[37] Hyun Soo Park, Takaaki Shiratori, Iain Matthews, and Yaser Sheikh. 3d reconstruction of a moving point from a series of 2d projections. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III 11, pages 158-171. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[38] C.J. Poelman and T. Kanade. A parapspective factorization method for shape and motion recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(3):206-218, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[39] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[40] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Shankar Rao, Roberto Tron, Rene Vidal, and Yi Ma. Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(10):1832-1845, 2009. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[42] Shankar R Rao, Roberto Tron, René Vidal, and Yi Ma. Motion segmentation via robust subspace separation in the presence of outlying, incomplete, or corrupted trajectories. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[43] Junghoon Seo, Jamyoung Koo, and Taegyun Jeon. Deep closed-form subspace clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[44] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[45] Yasuyuki Sugaya and Kenichi Kanatani. Geometric structure of degeneracy for multi-body motion segmentation. In International Workshop on Statistical Methods in Video Processing, pages 13-25. Springer, 2004. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[46] Roberto Toldo and Andrea Fusiello. Robust multiple structures estimation with j-linkage. In Computer Vision – ECCV 2008, pages 537–547. Springer Berlin Heidelberg, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.418 + ], + "angle": 0, + "content": "[47] Carlo Tomasi and Takeo Kanade. Shape and motion from image streams: a factorization method. International journal of computer vision, 9(2):137-154, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[48] Philip HS Torr. Geometric motion segmentation and model selection. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 356(1740):1321-1340, 1998. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[49] Roberto Tron and René Vidal. A benchmark for the comparison of 3-d motion segmentation algorithms. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[50] Manolis Tsakiris and Rene Vidal. Theoretical analysis of sparse subspace clustering with missing entries. In International Conference on Machine Learning, pages 4975-4984. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.589, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[51] Jeya Maria Jose Valanarasu and Vishal M Patel. Overcomplete deep subspace clustering networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 746-755, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[52] Roman Vershynin. High-dimensional probability: An introduction with applications in data science. Cambridge university press, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[53] René Vidal and Richard Hartley. Motion segmentation with missing data using powerfactorization and gpca. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II–II. IEEE, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[54] Rene Vidal, Yi Ma, and Shankar Sastry. Generalized principal component analysis (gpca). IEEE transactions on pattern analysis and machine intelligence, 27(12):1945-1959, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[55] René Vidal, Roberto Tron, and Richard Hartley. Multiframe motion segmentation with missing data using powerfactorization and gpca. International Journal of Computer Vision, 79:85-105, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[56] Xun Xu, Loong Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19101" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2859-2867, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[57] Jingyu Yan and Marc Pollefeys. A general framework for motion segmentation: Independent, articulated, rigid, non-rigid, degenerate and non-degenerate. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9, pages 94-106. Springer, 2006. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[58] Lihi Zelnik-Manor and Michal Irani. Temporal factorization vs. spatial factorization. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8, pages 434-445. Springer, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.343 + ], + "angle": 0, + "content": "[59] Shangzhi Zhang, Chong You, René Vidal, and Chun-Guang Li. Learning a self-expressive network for subspace clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12393-12403, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.346, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[60] Jianqiao Zheng, Sameera Ramasinghe, and Simon Lucey. Rethinking positional encoding. arXiv preprint arXiv:2107.02561, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[61] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[62] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE International Conference on Image Processing 2005, pages III-153. IEEE, 2005. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19102" + } + ] +] \ No newline at end of file diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_origin.pdf b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..afacdea45eeecbf00cb79cb4e0c6a0351eeb6568 --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/37fd4122-aae1-41d4-895f-bd90782ed866_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eadd1b88822ef9266361ef478db2422ba59d275be2647a73f804b4d5e7131b6a +size 4514556 diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/full.md b/2024/Learned Trajectory Embedding for Subspace Clustering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e62bb541602aaa057b72004a4f140919b665990e --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/full.md @@ -0,0 +1,336 @@ +# Learned Trajectory Embedding for Subspace Clustering + +Yaroslava Lochman + +Carl Olsson1,2 + +Christopher Zach1 + +1Chalmers University of Technology + +$^{2}$ Lund University + +# Abstract + +1 Clustering multiple motions from observed point trajectories is a fundamental task in understanding dynamic scenes. Most motion models require multiple tracks to estimate their parameters, hence identifying clusters when multiple motions are observed is a very challenging task. This is even aggravated for high-dimensional motion models. The starting point of our work is that this high-dimensionality of motion model can actually be leveraged to our advantage as sufficiently long trajectories identify the underlying motion uniquely in practice. Consequently, we propose to learn a mapping from trajectories to embedding vectors that represent the generating motion. The obtained trajectory embeddings are useful for clustering multiple observed motions, but are also trained to contain sufficient information to recover the parameters of the underlying motion by utilizing a geometric loss. We therefore are able to use only weak supervision from given motion segmentation to train this mapping. The entire algorithm consisting of trajectory embedding, clustering and motion parameter estimation is highly efficient. We conduct experiments on the Hopkins155, Hopkins12, and KT3DMoSeg datasets and show state-of-the-art performance of our proposed method for trajectory-based motion segmentation on full sequences and its competitiveness on the occluded sequences. Project page: https://ylochman.github.io/trajectory-embedding. + +# 1. Introduction + +Multi-model fitting is a classical problem in computer vision, and a typical example is the task of estimating individual motions from point trajectories observed in images containing multiple moving (and potentially deforming) objects. A trajectory is a sequence of tracked point coordinates extracted in different frames and is therefore also a point in a high-dimensional space. The commonly used union-of-subspace models [61] assume that point trajectories originating from the same object/motion are samples from the + +same (unknown) low dimensional subspace. The inference problem thus becomes that of simultaneously determining the parameters of underlying subspaces, the trajectory assignments to subspaces, and in practice also denoising of the data. Specifically, given an observation matrix $\mathbb{M}_{2F\times P}$ of point trajectories + +$$ +\mathbb {M} _ {2 F \times P} = \left( \begin{array}{c c c} | & & | \\ \mathbf {x} _ {1} & \dots & \mathbf {x} _ {P} \\ | & & | \end{array} \right), \tag {1} +$$ + +where $F$ is the number of frames and $P$ is the number of points, the goal is to find a $P \times P$ permutation matrix $\mathsf{P}_{\pi}$ that orders the trajectories into sub-matrices $\{\mathbf{X}_i\}_{i=1}^c$ corresponding to different motions. Each motion is represented by a $2F \times k_i$ basis matrix $\mathsf{B}_i$ whose columns span the subspace and therefore $\mathbf{X}_i = \mathsf{B}_i\mathsf{C}_i^T$ , where $\mathsf{C}_i$ is a $k_i \times P_i$ shape matrix. Hence + +$$ +\mathrm {M} _ {2 F \times P} \mathrm {P} _ {\pi} \approx \left[ \begin{array}{l l l} \mathrm {B} _ {1} \mathrm {C} _ {1} ^ {\top} & \dots & \mathrm {B} _ {c} \mathrm {C} _ {c} ^ {\top} \end{array} \right], \tag {2} +$$ + +where $\sum_{i=1}^{c} P_i = P$ , and ranks $k_i$ are assumed to be small. In this work we assume that the ranks $\{k_1, \ldots, k_c\}$ are known which is a common scenario under multiple rigid motions [49]. While estimating a single motion model $\mathsf{B}_i$ from a cluster of trajectories belonging to the same motion is relatively easy, the combination of cluster assignment and model estimation turns inference into a complicated chicken and egg problem. Further complications arise due to occlusion and partial observations which make it unlikely that we can observe all the full trajectories in all cases. + +In this work we seek to learn a feature representation that allows us to solve the clustering problem separately from the motion estimation. For this purpose we train a simple feed forward network (with the overall structure depicted in Fig. 1) taking a trajectory as input and returning a corresponding feature embedding. Using a metric learning loss we ensure that the features are close if and only if corresponding trajectories belong to the same motions. Our approach allows us to input trajectories of varying lengths and different starting and ending times, making it possible to cluster trajectories using their features without completing any missing data. In summary our main contributions are: + +![](images/73d4e2bd8bd17b8fabda15038d42e7098775e3a7ad53220c6d3bf6c8b5816d99.jpg) +Figure 1. Schematic overview of the proposed approach. Input trajectories are passed through the feature extractor to obtain $\{\mathbf{f}_i\}$ . The subspace estimator generates the time-dependent basis using coordinate-MLP, infers basis coefficients from $\mathbf{f}$ , and outputs subspace B, where each sub-matrix $\mathsf{B}_t \in \mathbb{R}^{2 \times 4}$ is computed via linear combination of basis functions. $\mathcal{L}_{\mathrm{InfoNCE}}$ is the contrastive loss that uses ground truth cluster assignments, $\mathcal{L}_{\mathrm{Residual}}$ computes the trajectory-to-subspace geometric error, and $\mathcal{L}_{\mathrm{FeatDiff}}$ computes the feature difference between input and reconstructed trajectory. For motion segmentation, the trajectories are clustered in a feature space using $\{\mathbf{f}_i\}$ . + +- We present a method to map point trajectories to discriminative feature representations, which can be seen as compact descriptions of the underlying generating motion. +- We show that clustering of trajectories can be performed reliably solely using those features, achieving state-of-the-art performance without the need for joint motion estimation and cluster assignments. +- The proposed feature extraction network is trained to be invariant with respect to random occlusions and pixel noise by utilizing a novel generative model that implicitly performs per-trajectory model estimation. +- We present and evaluate different ways to use the proposed network, such as trajectory clustering, trajectory completion and subspace estimation. + +# 1.1. Related Work + +Within geometric vision the traditionally dominant model-fitting approaches are robust statistical methods based on RANSAC [15]. In case of low dimensional models the approach provides a simple way of effectively exploring the search space. Iterative RANSAC [48], which sequentially finds models and removes their supporting inliers, is a simple option that handles settings with multiple models. It has been observed that the greedy nature of the approach makes it inefficient when a large number of models is present [6, 12, 21, 62]. In contrast, MultiRANSAC [62] simultaneously handles multiple models using iteratively fused consensus sets, and CONSAC [25] increases the efficiency of sampling by learning a sampling strategy based on previous detections. Energy minimization is used in [6, 12, 21] to select models among random proposals that jointly explain the data well. In [7] the hard point-to-model assignment is dropped in favour of progressively determining dominant models which enables clustering in consensus space. + +As an alternative to consensus some works use preference [31, 32, 46]. The J-linkage approach [46] uses affini- + +ties between data and a set of models, followed by agglomerative clustering to assign the data to models. The T-linkage method [32] is based on a continuous relaxation of the binary assignment problem. In [31] the preference representation is used in a robust PCA formulation [10] while [33] uses a set cover approach. + +Low dimensional geometric models have also been used to group longer point tracks, e.g. in [56] geometric two-view relations are used to create pairwise affinities for a spectral clustering stage. Segmentations computed from images pairs are combined through an averaging approach in [3, 4]. We note that for these local methods motion-specific properties should be observable across all image pairs in order to provide a largely coherent set of clusters. + +Subspace models are commonly occurring in geometric computer vision tasks. Early works like [38, 47] showed that structure from motion under affine camera models is essentially a subspace fitting problem. Further, shape deformations are modeled using a low dimensional subspace in [9]. The generalization to multiple independently moving objects leads to segmentation into multiple subspaces [8, 11, 17]. Approaches for handling degenerate and articulated cases were given in [45, 57, 58]. The general problem can be seen as an instance of subspace clustering which assumes that high dimensional data lies in a union of low-dimensional subspaces. The corresponding group of methods aim to simultaneously cluster the data points and estimate the underlying subspaces [1, 26], and may also perform matrix completion [14, 27, 50]. The generalized PCA approach (GPCA) [53-55] treats the union-of-subspaces as a zero-level-set of a higher degree polynomial which is fit to the data. A popular approach to the problem is sparse subspace clustering [5, 13], which has been shown to perform well on the benchmark provided in [49]. These methods are based on the notion of self-expressiveness, i.e. all data in a $d$ -dimensional subspace are linear combinations of $d$ points + +from that subspace. This enables an efficient approach that is provably correct in the case of disjoint subspaces (see Section 2). The LRR approach [29, 30] uses a dictionary basis and enforces low rank patterns among the coefficients. + +More recently, several deep learning methods have been proposed for solving the problem of unsupervised subspace clustering. The first work by Ji et al. [23] used deep autoencoder and introduced a self-expressive layer. Later, a closed form solution of the self-expressive layer was derived [43] which improved its efficiency. Alternatively, it was shown [51] that using over-complete representations increases robustness to noise and improves overall performance. Zhang et al. [59] directly learns a self-expressive data representation. The approach is highly scalable since it circumvents the need for a quadratic number of coefficients of the original self-expressive model. + +While general subspace learning models typically do not use temporal information, this can be of great importance when analysing and predicting motion. For the reconstruction of deforming objects, Akhtar et al [2] noted the duality between trajectory space and shape space. The use of pre-defined trajectory-bases such as a DCT basis (motivated by the smoothness assumption of time-dependent trajectories [20, 37]) leads to a reduction in the number of unknowns [2]. Finally, a probabilistic framework for subspace clustering with temporal information is presented in [19]. + +# 2. Overview of the Approach + +In this section we present our approach and the underlying architecture of our network. Our goal is to learn function $f_{\theta}$ that takes an observed point trajectory as input and yields a (relatively low-dimensional) feature embedding that is appropriate for trajectory grouping. For this purpose we adopt a small PointNet-style network [39]. One of the advantages of this type of network, apart from its simplicity, is the built-in invariance with respect to the permutation of the points in the input tensor. In order to obtain a suitable feature representation, we specify a number of desirable properties that we try to infuse through the architecture and training: + +- The features should be adequate for grouping/clustering of trajectories. For this reason we want a latent space with small within-cluster-distances and large between-cluster-distances. This is achieved using an InfoNCE loss. +- Inspired by traditional motion segmentation formulations, we aim for features to encode meaningful "geometric information" about the underlying motion subspaces. For this purpose, we design a decoding module that estimates a basis B for the underlying motion subspace that the trajectory belongs to. This prevents input memorization and greatly improves generalization to unseen data. +- Despite input tracks being of varying lengths we opt for a feature embedding of a fixed size. This simplifies treatment of problem instances with missing data and trajec + +tories starting and ending at different times. The decoding module therefore takes an extra variable, specifying at which time instance the rows of the basis matrix B should be sampled. In this way the model allows us to sub-sample and predict motion information beyond the trajectory samples. + +- Finally, feature embeddings and intermediate predictions are preferred to behave smoothly with respect to added input noise. To achieve this we construct a "noise-free" version of the observed trajectory from the predicted basis, which we require to be close to the original trajectory through a residual loss. Additionally, this reconstructed trajectory is subsequently fed into the network to obtain a corresponding feature vector, which is also required to be close to the original feature embedding. + +Because of the properties above we obtain a trajectory-feature from which we can extract the full motion subspace. Thus we essentially learn a function that maps a single trajectory to a subspace basis. For a general collection of subspaces, where trajectories can belong to multiple of them, this may not be possible (using a purely geometric approach). A simple example is line fitting to image points, where points in the intersection of two models can belong to any of them. In such cases geometry alone is not enough to resolve ambiguous assignments but other cues such as spatial smoothness of assignments [21] have to be used. However, in high-dimensional settings with low-dimensional subspaces the situation is somewhat different. In these cases a typical assumption is that of disjoint subspaces [13], that is, two subspaces only intersect at the origin, ensuring that there is a function that takes (non-zero) trajectories to their corresponding subspaces. We remark that while this assumption is sufficient for our purposes, it is not necessary. Since our method learns a data-driven mapping from point tracks to subspaces, it is able to focus on other trajectory properties than geometric ones. Therefore it can potentially handle the case of intersecting subspaces as long as the trajectory assignment is unambiguous. + +The disjoint subspace assumption We conclude this section by illustrating that the disjoint subspace assumption is valid in practical settings by analysing the data in the Hopkins 155 dataset. Figure 2 shows that re-using subspaces to explain trajectories in the other clusters generally leads to higher errors. + +Formally, two linear subspaces $S_{i}$ and $S_{j}$ are disjoint if $\dim(S_{i} \oplus S_{j}) = \dim(S_{i}) + \dim(S_{j})$ . Hence if $\mathsf{B}_{i}$ and $\mathsf{B}_{j}$ contain bases for the two subspaces, the rank of the concatenated matrix, $\eta_{ij} = \mathrm{rank}((\mathsf{B}_i \mathsf{B}_j))$ , fulfills $\eta_{ij} < \mathrm{rank}(\mathsf{B}_i) + \mathrm{rank}(\mathsf{B}_j)$ if the subspaces intersect. We extracted subsequences of varying length from the Hopkins 155 dataset and used the ground-truth clusters to generate the corresponding subspaces — for all pairs $\{\mathsf{B}_i, \mathsf{B}_j\}$ , where $i \neq j$ and $\mathrm{shape}(\mathsf{B}_i) = \mathrm{shape}(\mathsf{B}_j)$ , the rank $\eta_{ij}$ of the concatenated + +![](images/f2cf8a5e290db9d7728999b588d62c2b9f439ed21dcba9f5234c0a98852573cb.jpg) +Figure 2. RMS normalized error of the trajectory clusters with respect to the subspaces in Hopkins155 dataset (subsequences of length $F = 40, 50,$ and 60). An $(i,j)$ cell corresponds to the trajectory cluster of the $i^{\text{th}}$ motion and the subspace of $j^{\text{th}}$ motion. + +matrix was found to be $8^{2}$ . This suggests that up to numerics the subspaces do not intersect. In the following Section we describe the specific architecture and the employed losses in our approach in more detail. + +# 3. Architecture and Training Losses + +The schematic overview of the proposed architecture and training losses is shown in Fig. 1. Using the hypothesis that most observed trajectories come from a unique motion model, we aim to learn a function $f_{\theta}$ mapping a point trajectory directly to a representation of the corresponding motion model. We interpret the output $f_{\theta}(\mathbf{x})$ as an abstract description of the generating motion, from which a concrete representation to measure geometric consistency can be extracted. We choose linear (low-dimensional) subspaces as the concrete realizations of motion models. Specifically, we learn mappings $f_{\theta}$ and $g_{\phi}$ using training data such that + +1. $f_{\theta}$ maps a trajectory of varying length to a fixed-length feature embedding, and +2. $g_{\phi}$ converts this feature vector for a given query time into the corresponding rows of the subspace representation. + +# 3.1. Feature Extraction + +To construct a discriminative function $f_{\theta}(\cdot)$ , that is able to infer the motion from an input point trajectory, we first leverage the ability of the neural networks to compress high-dimensional complex data space into a lower-dimensional latent space with task specific invariance properties, from which it is easier to infer the output. A feature extractor for the point trajectories $f_{\theta}: \mathbf{x} \mapsto \mathbf{f} \in \mathbb{R}^{d}$ is therefore trained to produce invariant feature representations using ground truth clusters. In particular, an invariance with respect to the measurement noise and the variations within the motion clusters is of our interest. We therefore use a contrastive loss, namely a variant of the InfoNCE loss [36], to conduct metric learning, leading to a spread of feature embeddings + +in order to improve feature-based motion clustering, + +$$ +\begin{array}{l} \mathcal {L} _ {\text {I n f o N C E}} = - \frac {1}{| \mathcal {D} |} \sum_ {(i, j, l, k) \in \mathcal {D}} \log \left(\frac {p _ {i j}}{p _ {i j} + p _ {l k}}\right), \\ p _ {i j} = \exp \left(- \frac {\| \mathbf {f} _ {i} - \mathbf {f} _ {j} \| _ {2} ^ {2}}{T}\right), \tag {3} \\ \end{array} +$$ + +where $\mathbf{f}_i = f_{\theta}(\mathbf{x}_i)$ and $\mathcal{D}$ is the set of quadruples $(i,j,l,k)$ where $(i,k)$ belong to the same cluster and $(l,k)$ are random pairs, and $T$ is the temperature (hyper-)parameter, which we fix to 1. $\mathcal{L}_{\mathrm{InfoNCE}}$ is an instance of a supervised contrastive loss and is tightly connected to the triplet loss used in metric learning [24]. + +As mentioned earlier, the proposed architecture resembles PointNet [39] with several modifications. The employed encoder consists of 1D temporal convolutional layers applied to each trajectory independently, with x and y values representing two input channels. A global max-pooling in a temporal domain is then applied on the time-dependent representations, followed by a small MLP and an $L_{2}$ normalization. The features are therefore lying on the $d - 1$ -dimensional unit sphere. Note that we do not apply pooling across the points (global context layer) due to the disjoint subspace assumption. For the experiments, we set the dimensionality of the latent space to $d = 128$ . + +# 3.2. Subspace Estimation + +The latent vector $f_{\theta}(\mathbf{x})$ ideally identifies a motion model, but for geometric consistency this feature embedding requires conversion into a suitable mathematical motion model. We assume that observed point trajectories are generated by low-rank subspace models, and therefore the feature embedding is mapped by a subsequent network $g_{\phi}$ to the corresponding matrix representation of the associated subspace. Specifically, the mapping $g_{\phi}: \mathcal{F} \times \mathbb{R} \rightarrow \mathbb{R}^{2 \times r}$ produces rows $\mathsf{B}_t$ for the given motion model $\mathbf{f} = f_{\theta}(\mathbf{x})$ and the time index $t$ . Ideally the input trajectory $\mathbf{x}$ lies in the predicted subspace; if the trajectory $\mathbf{x}$ is consisting of (2D) points observed at times $t_1, \ldots, t_F$ , then the output subspace is + +$$ +\mathsf {B} = \left( \begin{array}{c} g _ {\phi} \left(f _ {\theta} (\mathbf {x}), t _ {1}\right) \\ \vdots \\ g _ {\phi} \left(f _ {\theta} (\mathbf {x}), t _ {F}\right) \end{array} \right), \tag {4} +$$ + +and the (squared) residual is given in closed-form as + +$$ +\min _ {\mathbf {c}} \| \mathbf {x} - \mathrm {B c} \| _ {2} ^ {2} = \| \mathbf {x} - \mathrm {B B} ^ {\dagger} \mathbf {x} \| _ {2} ^ {2}, \tag {5} +$$ + +where $\mathbf{c}$ is the coefficient vector and $\mathsf{B}^{\dagger} = (\mathsf{B}^{\top}\mathsf{B})^{-1}\mathsf{B}^{\top}$ . If the network identified the subspace correctly, then the residual is in the order of the observation noise level (and therefore small). + +Subspace basis functions The network $g_{\phi}$ can be a small but generic MLP. Alternatively, we may leverage existing domain knowledge and choose the structure of $g_{\phi}$ accordingly. In particular, we assume that all motions are linear combination of a finite, but continuous time trajectory basis $\{h^j\}$ . The basis functions $h^j$ can be fully fixed (such as a cosine basis with pre-determined frequencies and phase shifts [2]), parametric (such as the cosine basis with trainable frequencies and phase shifts [44]) or entirely non-parametric (i.e. represented by a generic MLP). Our choice of basis functions is a damped version of a cosine basis and is given as follows, + +$$ +h _ {\psi} ^ {j} (t) = e ^ {- \left(\alpha_ {j} (t - \mu_ {j})\right) ^ {2}} \cos (\beta_ {j} t + \gamma_ {j}), \tag {6} +$$ + +where $\psi = (\mu_j, \alpha_j, \beta_j, \gamma_j)_j$ are trainable parameters. The mapping $g_{\phi}$ can therefore be written as a composition + +$$ +g _ {\phi} (\mathbf {f}, t) = \omega_ {\zeta} (\mathbf {f}) \mathbf {h} _ {\psi} (t), \tag {7} +$$ + +where $\mathbf{h}_{\psi}(t) = (h_{\psi}^{j}(t))_{j}$ , and the trainable function $\omega_{\zeta}$ extracts the basis coefficients from $\mathbf{f}$ . If we have $N$ basis functions (i.e. $\dim (\mathbf{h}_{\psi}(t)) = N$ ), then $\omega_{\zeta}(\mathbf{f})$ is a $2r\times N$ matrix. The total set of trainable parameters in the subspace estimation network is $\phi = (\zeta ,\psi)$ . Finally, the network prediction can be compactly written as + +$$ +\mathsf {B} = \mathrm {H} _ {\psi} (\mathbf {t}) \Omega_ {\zeta} \left(f _ {\theta} (\mathbf {x})\right), \tag {8} +$$ + +where $\Omega_{\zeta}(\mathbf{f}) = \mathrm{reshape}_{2N\times r}\left(\omega_{\zeta}(\mathbf{f})\right)$ and + +$$ +\mathrm {H} _ {\psi} (\mathbf {t}) = \left( \begin{array}{l l} \mathbf {h} _ {\psi} (t _ {1}) ^ {\top} & \\ & \mathbf {h} _ {\psi} (t _ {1}) ^ {\top} \\ \vdots & \\ \mathbf {h} _ {\psi} (t _ {F}) ^ {\top} & \\ & \mathbf {h} _ {\psi} (t _ {F}) ^ {\top} \end{array} \right) \in \mathbb {R} ^ {2 F \times 2 N}. \tag {9} +$$ + +We set the number of subspace basis functions to $N = 64$ in our experiments. Since the predicted subspace matrix B augments the feature representation (see Sec. 3.3 below), we enforce a unique subspace representation by fixing the top block of B to the identity matrix (by multiplying with the respective inverse). + +Time dependence and time invariance The network represented by $g_{\phi}$ can be interpreted as taking a "feature" token $\mathbf{f}$ and a time coordinate $t$ as input, and is therefore somewhat analogous to the coordinate-MLP employed for implicit neural primitives such as [35, 44] or time series modeling [16]. The continuous time basis $\{h^j\}$ can be seen as domain-specific equivalent to positional coding, but its design is based on the expected motion patterns and does not necessarily meet the requirements of generic positional coding [40, 44, 60]. + +The formulation above is intentionally not time-invariant: the same trajectory observed at different starting times may predict different motion models. This is usually a desired behavior for motion segmentation, e.g. two cars traveling the same path at different time form different motion segments. As with CNNs applied to images, low-level trajectory features are extracted by convolutional layers, and time-dependence is incorporated via final fully connected and pooling layers. + +# 3.3. Training + +The total training loss is the weighted sum of a contrastive loss $\mathcal{L}_{\mathrm{InfoNCE}}$ applied on $\{\mathbf{v}_i\}$ , where $\mathbf{v}_i = (\mathrm{flatten}(\mathsf{B}_i)^\top \mathbf{f}_i^\top)^\top$ , the reconstruction loss $\mathcal{L}_{\mathrm{Residual}}$ + +$$ +\mathcal {L} _ {\text {R e s i d u a l}} = \frac {1}{P} \sum_ {j = 1} ^ {P} \| \mathbf {x} _ {j} - \tilde {\mathbf {x}} _ {j} \| _ {2} ^ {2}, \tag {10} +$$ + +and the feature difference loss + +$$ +\mathcal {L} _ {\text {F e a t D i f f}} = \frac {1}{P} \sum_ {j = 1} ^ {P} \| f _ {\theta} (\mathbf {x} _ {j}) - f _ {\theta} (\tilde {\mathbf {x}} _ {j}) \| _ {2} ^ {2}, \tag {11} +$$ + +where $\tilde{\mathbf{x}}_j = \mathsf{B}(\mathbf{x}_j)\mathsf{B}(\mathbf{x}_j)^\dagger\mathbf{x}_j$ is the projection of $\mathbf{x}_j$ onto $\mathsf{B}(\mathbf{x}_j)$ . In short, $\mathcal{L}_{\mathrm{InfoNCE}}$ forces both an approximate invariance of the feature extractor $f_{\theta}$ with respect to the variations within clusters and smoothness of the coordinate-MLP $g_{\phi}$ , $\mathcal{L}_{\mathrm{Residual}}$ ensures geometric consistency of the subspaces with respect to input trajectories, and $\mathcal{L}_{\mathrm{FeatDiff}}$ forces an approximate invariance of $f_{\theta}$ with respect to the pixel noise and can also be seen as enforcing smoothness of $f_{\theta}$ . We train the network in two stages. First, we pre-train $f_{\theta}$ with $\mathcal{L}_{\mathrm{InfoNCE}}$ until convergence. Then we add $g_{\phi}$ and train it together with $f_{\theta}$ using the total loss. To enforce the approximate invariance with respect to random detection failures, we synthesize missing points during training by applying dropout to the input sequences. + +# 3.4. Trajectory Completion + +Let $B_{\theta, \phi}(\mathbf{x}, \mathbf{t})$ provide the subspace for an input trajectory $\mathbf{x} \in \mathbb{R}^{2F}$ and time vector $\mathbf{t} = (t_1, \dots, t_F)^\top$ . Assume now that $\mathbf{x}$ is only partially observed, and let $\mathbf{w} \in \{0, 1\}^{2F}$ be the visibility mask. Further, let $\bar{\mathbf{w}} := \mathbf{1} - \mathbf{w}$ . The goal of the trajectory completion is to estimate missing values $\bar{\mathbf{x}}$ that minimize the following objective + +$$ +J (\bar {\mathbf {x}}) := \left\| (I - B (\bar {\mathbf {x}}) B (\bar {\mathbf {x}}) ^ {\dagger}) (w \odot x + \bar {w} \odot \bar {x}) \right\| ^ {2}, \tag {12} +$$ + +where $\mathsf{B}(\bar{\mathbf{x}}) = B_{\theta ,\phi}(\mathbf{w}\odot \mathbf{x} + \bar{\mathbf{w}}\odot \bar{\mathbf{x}},\mathbf{t})$ . Back-propagation for gradient-based determination of $\bar{\mathbf{x}}$ is in principle possible but computationally expensive. Alternatively, we rely on the approximate invariance of the feature extractor $f_{\theta}$ with respect to the length of the input sequence and extrapolate the output subspace in time using $g_{\phi}(\mathbf{f},t)$ . More + +
Method2 motionsHopkins155AllHopkins12KT3DMoSeg
MeanMedianTimeMeanMedianTimeMeanMedianTimeMeanMedianMeanMedian
RANSAC [48]5.561.18175ms22.9422.03258ms9.763.21194ms----
GPCA [54]4.590.38324ms28.6628.26738ms10.342.54417ms--34.6033.95
MSL [45]4.140.0011h 4m8.231.761d 23h5.030.0019h 11m----
LSA [57]3.450.597.58s9.732.3315.96s4.940.909.47s--38.3038.58
ALC5 [42]3.030.00-6.261.02-3.760.265m 15s3.810.1724.3119.04
ALCsp [42]2.400.43-6.690.67-3.370.496m 11s1.281.07--
LRR [30]4.100.22-9.890.56-5.410.531.1s--33.6736.01
SSC [13]0.820.00-2.450.20-2.450.20920ms--33.8833.54
RSIM [22]0.780.00-1.770.28-1.010.00176ms0.680.70--
MultiConsensus [7]------4.40-40ms----
Ours0.630.07ms0.600.010ms0.620.09ms5.122.045.850.80
+ +Table 1. Motion segmentation results on the Hopkins155 [49], Hopkins12 [41] and KT3DMoSeg [56] datasets. The evaluation metrics is classification error $(\%)$ . ' -' indicates that the results were not reported / no code available. + +specifically, the subspace of $\mathbf{w} \odot \mathbf{x} + \bar{\mathbf{w}} \odot \bar{\mathbf{x}}$ can be estimated from $\mathbf{x}_{\mathrm{vis}}$ -a fully visible sub-trajectory of the trajectory $\mathbf{x}$ as long as the feature representation $\mathbf{f}$ can be inferred (i.e. $\mathbf{x}_{\mathrm{vis}}$ is sufficiently long). If we fix this estimate $\mathbf{B}$ and let $\mathsf{W} := \mathrm{diag}(\mathbf{w})$ , $\bar{\mathsf{W}} := \mathrm{diag}(\bar{\mathsf{w}})$ , the new objective can be stated as + +$$ +\tilde {J} (\bar {\mathbf {x}}) := \left\| \left(\mathrm {I} - \mathrm {B B} ^ {\dagger}\right) \bar {W} \mathbf {x} + \left(\mathrm {I} - \mathrm {B B} ^ {\dagger}\right) \bar {W} \bar {\mathbf {x}} \right\| ^ {2}, \tag {13} +$$ + +and the minimizer $\bar{\mathbf{x}}^*$ of $\tilde{J}$ is given by + +$$ +\bar {\mathbf {x}} ^ {*} = - \left(\left(\mathrm {I} - \mathrm {B B} ^ {\dagger}\right) \bar {\mathbf {w}}\right) ^ {\dagger} \left(\mathrm {I} - \mathrm {B B} ^ {\dagger}\right) \mathbf {W x} =: \mathrm {A} (\mathrm {B}) \mathbf {x}. \tag {14} +$$ + +It is now possible to refine the estimated subspace using $\bar{\mathbf{x}}^*$ by feeding $\mathbf{w} \odot \mathbf{x} + \bar{\mathbf{w}} \odot \bar{\mathbf{x}}^*$ into the network. Overall, we apply the following iterative procedure for trajectory completion, + +$$ +\left\{ \begin{array}{r l} \mathsf {B} _ {0} & \leftarrow B _ {\theta , \phi} (\mathbf {x} _ {\text {v i s}}, \mathbf {t}) \\ \bar {\mathbf {x}} _ {i + 1} & \leftarrow \mathsf {A} (\mathsf {B} _ {i}) \mathbf {x} \\ \mathsf {B} _ {i + 1} & \leftarrow B _ {\theta , \phi} (\mathbf {w} \odot \mathbf {x} + \bar {\mathbf {w}} \odot \bar {\mathbf {x}} _ {i + 1}, \mathbf {t}). \end{array} \right. \tag {15} +$$ + +![](images/70641f339b1ce61533b8b379eb5789df914a5deffdaedfa664b3a0dbec7aec9a.jpg) +Figure 3. Histograms of the classification errors for the groups with (left) 2 and (right) 3 motions in Hopkins155 dataset. + +![](images/e2482bb9b3f2595898ed145e93ff942b73970873bd6d44eb021a4242821af9ac.jpg) + +# 4. Experiments + +The proposed network was trained on the two motion segmentation datasets: Hopkins155 [49] and KT3DMoSeg [56], yielding two different sets of weights. In our experiments, we used NVIDIA GeForce RTX 3080, CUDA 11.5, PyTorch 1.11. + +To obtain groups of trajectories, it is sufficient to perform a partial forward pass through the (trained) feature extractor $f_{\theta}$ and cluster in the feature space. We use hierarchical clustering as it provides a meaningful framework for selecting the number of clusters, i.e. motions. The subspaces can then be fit to each cluster. If the goal is to estimate the subspace for a single input trajectory, a full forward pass through $f_{\theta}$ and $g_{\phi}$ should be performed. For motion segmentation with missing data, the largest fully visible blocks of trajectories (but ignoring single occluded points) can be used to estimate the initial subspaces B, followed by trajectory completion as outlined in (15). The completed trajectories can then be used to cluster in the feature space. + +# 4.1. Motion Segmentation + +We evaluate the performance of the proposed network in terms of motion segmentation on Hopkins155 [49], Hopkins12 [41], and KT3DMoSeg [56] datasets. We first compare the proposed method to the state-of-the-art algorithms that similarly assume rank-4 of B: sequential RANSAC [48], GPCA [54], MSL [45], LSA [57] (with the global subspace dimensionality $D = 4n$ for $n$ clusters), $\mathrm{ALC}_5$ and $\mathrm{ALC}_{\mathrm{sp}}$ [42], LRR [30], SSC [13], RSIM [22], and MultiConsensus [7]. We then also compare our method to the other approaches making different assumptions on the underlying motion model [3, 4, 28, 56]. The metrics used to evaluate clustering performance is the classification error (\%) which is standard for this benchmark [49]. + +Results on Hopkins155 dataset Hopkins155 [49] is the commonly used benchmark for trajectory-based motion + +![](images/eb6f11bcbfd11722946152a0079c5a998a3ad2a92021a9f05d2a0994b55613a1.jpg) +Figure 4. The approximate invariance measured as an angle between features extracted from the reference trajectory and its perturbed version using (left) Gaussian pixel noise, (middle) perturbation within subspace, and (right) uniform occlusions. The gradually increasing level of corruption is shown on the x-axis. + +![](images/34c5c2d4b6876d165fda2418c5ef64935724c2015b305639df503d5738f3facc.jpg) + +![](images/d56a9502e2b503c35590e5bd252a89b5368f3790416b0f4458253b4996d2fc7a.jpg) + +segmentation. The dataset consists of 155 sequences with 63-556 point trajectories generated by two or three motions in the scene. Each trajectory is fully visible, and the length varies from 15 to 100 frames. The averaged results are shown in the left block of Table 1, and the histograms of classification errors are shown in Fig. 3. Our network is very fast and produces the most accurate segmentations on fully visible trajectories. + +Results on Hopkins12 dataset Hopkins12 [41] is a commonly used dataset for evaluating subspace clustering with missing data. It consists of 12 corrupted sequences from Hopkins155 dataset, where the corruption was done manually by labeling entries of trajectories as missing. The corruption rate of the observation matrix M varies from $1\%$ to $23\%$ . For $\mathrm{ALC}_5$ and $\mathrm{ALC_{sp}}$ [42], we report $l^1$ -based entry completion results. The results are shown in the middle block of Table 1. The proposed network shows promising results. Note that the maximum corruption rate of the individual trajectories in the Hopkins12 dataset is $93\%$ and, unlike all the state-of-the-art approaches, the proposed trajectory completion method is not global, i.e. it works with individual trajectories. Therefore, if a provided individual sub-trajectory does not contain sufficient relevant information for motion estimation (which becomes more likely as the trajectory gets shorter), incorporating information from all tracks becomes necessary to disambiguate the true motion. Utilizing such global trajectory information is left for future research. + +Results on KT3DMoSeg dataset KT3DMoSeg [56] is an adapted version of the KITTI dataset [18, 34] containing 22 sequences with 158-1018 point trajectories and annotated segmentations. It is more challenging than Hopkins155 in several aspects. There is a dominant motion corresponding to static "background" scene which makes cluster distribution unbalanced. The trajectories are also highly occluded: the corruption rate of M goes up to $54\%$ , and for the individual trajectories — up to $75\%$ . Lastly, the sequences have shorter length of 10 to 20 frames. The results for this dataset + +are shown in the right block of Table 1 indicating that the proposed approach outperforms previous methods. + +Methods with different assumptions We also compared our approach to the other methods making different assumptions. Two-perspective-view (TPV) [28] assumes perspective cameras and uses pairwise epipolar relations. Spectral clustering methods proposed by [56]: kernel addition (KerAdd), co-regularization (CoReg), and subset constrained (Subset) spectral clustering also look at pairwise relations, but use fundamental matrices, affine matrices or homographies to compute the pairwise affinities. SYNCH [3] and MODE [4] similarly fit multiple fundamental matrices to each pair of images, but perform averaging. They assume that only pairwise matches are given. These methods are not directly comparable to ours, however the proposed method proves to be competitive as evidenced in Table 2. + +
MethodHopkins155 MeanHopkins12 Med.KT3DMoSeg MeanMed.
TPV [28]2.34----
KerAdd [56]0.360.110.08.311.02
CoReg [56]0.460.060.07.920.75
Subset [56]0.310.060.08.080.71
SYNCH-tracks [3]3.675.460.57--
SYNCH [3]1.193.190.28--
MODE [4]1.374.330.38--
Our method0.625.122.045.850.80
+ +Table 2. Comparison to the other methods making different assumptions. TPV [28] has no code publicly available. + +# 4.2. Approximate Invariances of $f_{\theta}$ + +We experimentally validate the approximate invariance of the trained feature extractor $f_{\theta}$ with respect to different types of synthetic data corruption: (1) Gaussian pixel noise with the standard deviation $\sigma$ gradually increasing up to 10 px, (2) perturbations within the subspace, where the trajectories are first corrupted by Gaussian pixel noise with $\sigma$ up to 300 px and then projected back to the underlying subspace, (3) synthesized uniform occlusions with missing data levels going up to 50%, and (4) tracking failures, where + +![](images/54c344a69caf4d564a94fdce18966557b2ac3337c72a1fab150f3cfa15f8dda4.jpg) +Figure 5. The approximate invariance measured as an angle between features extracted from the reference trajectory and its shortened version for (left) shorter sequences with $F = 15 \ldots 24$ , (middle) mid-length sequences with $F = 25 \ldots 31$ , and (right) longer sequences with $F = 35 \ldots 61$ . The gradually increasing percentage of sequence length reduction is shown on the x-axis. + +![](images/d999bcf6c79aa629d8c035d99a2a535b35ba2f87890f0d2c8410bdd95b458931.jpg) + +![](images/8264cabf08a73bbc65ae118e87608d0a284f3ea6844eeb1988898deeed37f7b8.jpg) + +the rearmost points are gradually excluded from the trajectories. All types of corruption are applied on the training sequences of Hopkins155 dataset. The invariances are measured using the angle between the features extracted from original and corrupted trajectories. The smaller the angle is, the more invariant the feature is to the certain type of data corruption. + +Results for the pixel noise, perturbations within the subspace, and uniform occlusions are shown in Fig. 4. Note that the features are $L_{2}$ -normalized in the $d$ -dimensional space for $d = 128$ . In such a high-dimensional space the two randomly picked vectors are likely to be almost orthogonal [52]. The network proves to be approximately invariant hence robust to these types of corruption, and especially to the pixel noise. Since 8-10 px is quite large (for the images of sizes $320 \times 240$ , $640 \times 480$ , or $720 \times 480$ ), this result also shows robustness of $f_{\theta}$ to occasional outlying points in the trajectory. The results for the synthesized tracking failures are shown in Fig. 5, where a single curve corresponds to a certain initial trajectory length $F$ . The angles are slightly higher for shorter sequences, which is expected since they contain less information. Note that during training we do not force the network to produce the same feature representations for the trajectories of the same point starting at the same time but having different length. + +![](images/2cbfb2bf4ea068025b17893b6eea77e0fbdf682ea68c279e949e97f095278ed0.jpg) +Figure 6. Mean pixel error of recovering from random occlusions for the corruption rates of (left) $10 - 40\%$ and (right) $50 - 80\%$ . + +# 4.3. Trajectory Completion + +We further evaluate the completion accuracy of the proposed method on the trajectories from Hopkins155 dataset with synthesized uniform occlusions. The performance is measured using the mean pixel error between the ground truth (i.e., detected/tracked) values of the synthetically occluded points and the predicted values. The error as a function of iterations for different corruption rates is shown in Fig. 6, where we split the results into a minor and a major corruption group. From the results we conclude that it is sufficient to run the algorithm (15) for 3 to 5 iterations for minor corruptions which takes 1s on average, and 5 to 20 iterations for major corruptions which takes 3s on average. + +# 5. Conclusion + +In this paper we propose to learn a mapping from point trajectories to feature vectors in an embedding space, which fully identify the generating motion. We illustrate that these feature vectors can be directly used for subspace clustering, thereby eliminating the need for simultaneous estimation of assignments and subspace models. The proposed training loss ensures not only the utility of the obtained embedding for clustering, but also forces the feature vector to encode the actual motion parameters, which are retrieved by a dedicated coordinate-MLP. Comparisons to baselines on Hopkins155 and KT3DMoSeg highlight the efficiency and accuracy of our method, providing top clustering results with running time below 10ms. + +In this work we focus on the linear subspace framework as sole motion model to explain observed trajectories. One direction for future work is how to incorporate additional and/or different prior assumptions on the motion model into our method. Further, the linear subspace assumption is strongly tied to affine cameras, which only approximate the more realistic pinhole camera model. Therefore extending this work to more general camera models is a challenging direction for future research. + +# References + +[1] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopoulos, and Prabhakar Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. In Proceedings of the 1998 ACM SIGMOD international conference on Management of data, pages 94-105, 1998. 2 +[2] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 3, 5 +[3] Federica Arrigoni and Tomas Pajdla. Motion segmentation via synchronization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2, 6, 7 +[4] Federica Arrigoni and Tomas Pajdla. Robust motion segmentation from pairwise matches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 671-681, 2019. 2, 6, 7 +[5] Liang Bai and Jiye Liang. Sparse subspace clustering with entropy-norm. In International conference on machine learning, pages 561-568. PMLR, 2020. 2 +[6] Daniel Barath and Jiri Matas. Progressive-x: Efficient, anytime, multi-model fitting algorithm. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3780-3788, 2019. 2 +[7] Daniel Barath, Denys Rozumnyi, Ivan Eichhardt, Lerente Hajder, and Jiri Matas. Finding geometric models by clustering in the consensus space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5424, 2023. 2, 6 +[8] T.E. Boult and L. Gottesfeld Brown. Factorization-based segmentation of motions. In Proceedings of the IEEE Workshop on Visual Motion, pages 179-186, 1991. 2 +[9] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2 +[10] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? J. ACM, 58(3), 2011. 2 +[11] J. Costeira and T. Kanade. A multi-body factorization method for motion analysis. In Proceedings of IEEE International Conference on Computer Vision, pages 1071-1076, 1995. 2 +[12] A. Delong, A. Osokin, H. Isack, and Y. Boykov. Fast approximate energy minimization with label costs. International Journal of Computer Vision, 96(1):1-27, 2012. 2 +[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE transactions on pattern analysis and machine intelligence, 35, 2013. 2, 3, 6 +[14] Jicong Fan and Madeleine Udell. Online high rank matrix completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8690-8698, 2019. 2 + +[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2 +[16] Elizabeth Fons, Alejandro Sztrajman, Yousef El-Laham, Alexandros Iosifidis, and Svitlana Vyetrenko. Hypertime: Implicit neural representation for time series. arXiv preprint arXiv:2208.05836, 2022. 5 +[17] C.W. Gear. Feature grouping in moving objects. In Proceedings of 1994 IEEE Workshop on Motion of Non-rigid and Articulated Objects, pages 214-219, 1994. 2 +[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7 +[19] Behnam Gholami and Vladimir Pavlovic. Probabilistic temporal subspace clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3066-3075, 2017. 3 +[20] Paulo FU Gotardo and Aleix M Martinez. Computing smooth time trajectories for camera and deformable shape in structure from motion with occlusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(10):2051-2065, 2011. 3 +[21] Hossam Isack and Yuri Boykov. Energy-based geometric multi-model fitting. International journal of computer vision, 97(2):123-147, 2012. 2, 3 +[22] Pan Ji, Mathieu Salzmann, and Hongdong Li. Shape interaction matrix revisited and robustified: Efficient subspace clustering with corrupted and incomplete data. In Proceedings of the IEEE International Conference on computer vision, pages 4687-4695, 2015. 6 +[23] Pan Ji, Tong Zhang, Hongdong Li, Mathieu Salzmann, and Ian Reid. Deep subspace clustering networks. Advances in neural information processing systems, 30, 2017. 3 +[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 4 +[25] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 2 +[26] Hans-Peter Kriegel, Peer Kroger, and Arthur Zimek. Clustering high-dimensional data: A survey on subspace clustering, pattern-based clustering, and correlation clustering. Acm transactions on knowledge discovery from data (tkdd), 3(1): 1-58, 2009. 2 +[27] Connor Lane, Ron Boger, Chong You, Manolis Tsakiris, Benjamin Haeffele, and Rene Vidal. Classifying and comparing approaches to subspace clustering with missing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2 +[28] Zhuwen Li, Jiaming Guo, Loong-Fah Cheong, and Steven Zhiying Zhou. Perspective motion segmentation via + +collaborative clustering. In Proceedings of the IEEE international conference on computer vision, pages 1369-1376, 2013. 6, 7 +[29] Guangcan Liu and Shuicheng Yan. Latent low-rank representation for subspace segmentation and feature extraction. In 2011 International Conference on Computer Vision, pages 1615-1622, 2011. 3 +[30] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 3, 6 +[31] Luca Magri and Fusiello Andrea. Robust multiple model fitting with preference analysis and low-rank approximation. In Proceedings of the British Machine Vision Conference 2015, pages 20-1, 2015. 2 +[32] Luca Magri and Andrea Fusiello. T-linkage: A continuous relaxation of j-linkage for multi-model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3954-3961, 2014. 2 +[33] Luca Magri and Andrea Fusiello. Multiple model fitting as a set coverage problem. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3318-3326, 2016. 2 +[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 7 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5 +[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[37] Hyun Soo Park, Takaaki Shiratori, Iain Matthews, and Yaser Sheikh. 3d reconstruction of a moving point from a series of 2d projections. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III 11, pages 158-171. Springer, 2010. 3 +[38] C.J. Poelman and T. Kanade. A parapspective factorization method for shape and motion recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(3):206-218, 1997. 2 +[39] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 4 +[40] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 5 +[41] Shankar Rao, Roberto Tron, Rene Vidal, and Yi Ma. Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(10):1832-1845, 2009. 6, 7 + +[42] Shankar R Rao, Roberto Tron, René Vidal, and Yi Ma. Motion segmentation via robust subspace separation in the presence of outlying, incomplete, or corrupted trajectories. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 6, 7 +[43] Junghoon Seo, Jamyoung Koo, and Taegyun Jeon. Deep closed-form subspace clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3 +[44] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 5 +[45] Yasuyuki Sugaya and Kenichi Kanatani. Geometric structure of degeneracy for multi-body motion segmentation. In International Workshop on Statistical Methods in Video Processing, pages 13-25. Springer, 2004. 2, 6 +[46] Roberto Toldo and Andrea Fusiello. Robust multiple structures estimation with j-linkage. In Computer Vision – ECCV 2008, pages 537–547. Springer Berlin Heidelberg, 2008. 2 +[47] Carlo Tomasi and Takeo Kanade. Shape and motion from image streams: a factorization method. International journal of computer vision, 9(2):137-154, 1992. 2 +[48] Philip HS Torr. Geometric motion segmentation and model selection. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 356(1740):1321-1340, 1998. 2, 6 +[49] Roberto Tron and René Vidal. A benchmark for the comparison of 3-d motion segmentation algorithms. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1, 2, 6 +[50] Manolis Tsakiris and Rene Vidal. Theoretical analysis of sparse subspace clustering with missing entries. In International Conference on Machine Learning, pages 4975-4984. PMLR, 2018. 2 +[51] Jeya Maria Jose Valanarasu and Vishal M Patel. Overcomplete deep subspace clustering networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 746-755, 2021. 3 +[52] Roman Vershynin. High-dimensional probability: An introduction with applications in data science. Cambridge university press, 2018. 8 +[53] René Vidal and Richard Hartley. Motion segmentation with missing data using powerfactorization and gpca. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II–II. IEEE, 2004. 2 +[54] Rene Vidal, Yi Ma, and Shankar Sastry. Generalized principal component analysis (gpca). IEEE transactions on pattern analysis and machine intelligence, 27(12):1945-1959, 2005. 6 +[55] René Vidal, Roberto Tron, and Richard Hartley. Multiframe motion segmentation with missing data using powerfactorization and gpca. International Journal of Computer Vision, 79:85-105, 2008. 2 +[56] Xun Xu, Loong Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models. + +In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2859-2867, 2018. 2, 6, 7 +[57] Jingyu Yan and Marc Pollefeys. A general framework for motion segmentation: Independent, articulated, rigid, non-rigid, degenerate and non-degenerate. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9, pages 94-106. Springer, 2006. 2, 6 +[58] Lihi Zelnik-Manor and Michal Irani. Temporal factorization vs. spatial factorization. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8, pages 434-445. Springer, 2004. 2 +[59] Shangzhi Zhang, Chong You, René Vidal, and Chun-Guang Li. Learning a self-expressive network for subspace clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12393-12403, 2021. 3 +[60] Jianqiao Zheng, Sameera Ramasinghe, and Simon Lucey. Rethinking positional encoding. arXiv preprint arXiv:2107.02561, 2021. 5 +[61] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1 +[62] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE International Conference on Image Processing 2005, pages III-153. IEEE, 2005. 2 \ No newline at end of file diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/images.zip b/2024/Learned Trajectory Embedding for Subspace Clustering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..465f838ca90d939f411b342d7090f7fa42a31489 --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79bf5ab94adfc364a791ea253c67d3fbc20e363b09205cc491b70520650fa93e +size 394714 diff --git a/2024/Learned Trajectory Embedding for Subspace Clustering/layout.json b/2024/Learned Trajectory Embedding for Subspace Clustering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..395b90a2c3017bb5e77b3d53666a5cf4c42cb024 --- /dev/null +++ b/2024/Learned Trajectory Embedding for Subspace Clustering/layout.json @@ -0,0 +1,10494 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 122, + 103, + 473, + 121 + ], + "type": "text", + "content": "Learned Trajectory Embedding for Subspace Clustering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 142, + 233, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 142, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 130, + 142, + 233, + 156 + ], + "type": "text", + "content": "Yaroslava Lochman" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 267, + 142, + 338, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 142, + 338, + 156 + ], + "spans": [ + { + "bbox": [ + 267, + 142, + 338, + 156 + ], + "type": "text", + "content": "Carl Olsson1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 372, + 142, + 462, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 142, + 462, + 158 + ], + "spans": [ + { + "bbox": [ + 372, + 142, + 462, + 158 + ], + "type": "text", + "content": "Christopher Zach1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 144, + 162, + 324, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 162, + 324, + 177 + ], + "spans": [ + { + "bbox": [ + 144, + 162, + 324, + 177 + ], + "type": "text", + "content": "1Chalmers University of Technology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 360, + 163, + 446, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 163, + 446, + 177 + ], + "spans": [ + { + "bbox": [ + 360, + 163, + 446, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 360, + 163, + 446, + 177 + ], + "type": "text", + "content": "Lund University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 192, + 217 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 230, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 290, + 531 + ], + "type": "text", + "content": "1 Clustering multiple motions from observed point trajectories is a fundamental task in understanding dynamic scenes. Most motion models require multiple tracks to estimate their parameters, hence identifying clusters when multiple motions are observed is a very challenging task. This is even aggravated for high-dimensional motion models. The starting point of our work is that this high-dimensionality of motion model can actually be leveraged to our advantage as sufficiently long trajectories identify the underlying motion uniquely in practice. Consequently, we propose to learn a mapping from trajectories to embedding vectors that represent the generating motion. The obtained trajectory embeddings are useful for clustering multiple observed motions, but are also trained to contain sufficient information to recover the parameters of the underlying motion by utilizing a geometric loss. We therefore are able to use only weak supervision from given motion segmentation to train this mapping. The entire algorithm consisting of trajectory embedding, clustering and motion parameter estimation is highly efficient. We conduct experiments on the Hopkins155, Hopkins12, and KT3DMoSeg datasets and show state-of-the-art performance of our proposed method for trajectory-based motion segmentation on full sequences and its competitiveness on the occluded sequences. Project page: https://ylochman.github.io/trajectory-embedding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 557 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 673 + ], + "type": "text", + "content": "Multi-model fitting is a classical problem in computer vision, and a typical example is the task of estimating individual motions from point trajectories observed in images containing multiple moving (and potentially deforming) objects. A trajectory is a sequence of tracked point coordinates extracted in different frames and is therefore also a point in a high-dimensional space. The commonly used union-of-subspace models [61] assume that point trajectories originating from the same object/motion are samples from the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 205, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 547, + 277 + ], + "type": "text", + "content": "same (unknown) low dimensional subspace. The inference problem thus becomes that of simultaneously determining the parameters of underlying subspaces, the trajectory assignments to subspaces, and in practice also denoising of the data. Specifically, given an observation matrix " + }, + { + "bbox": [ + 306, + 205, + 547, + 277 + ], + "type": "inline_equation", + "content": "\\mathbb{M}_{2F\\times P}" + }, + { + "bbox": [ + 306, + 205, + 547, + 277 + ], + "type": "text", + "content": " of point trajectories" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 364, + 286, + 545, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 286, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 364, + 286, + 545, + 326 + ], + "type": "interline_equation", + "content": "\\mathbb {M} _ {2 F \\times P} = \\left( \\begin{array}{c c c} | & & | \\\\ \\mathbf {x} _ {1} & \\dots & \\mathbf {x} _ {P} \\\\ | & & | \\end{array} \\right), \\tag {1}", + "image_path": "3ce58526565a057d0095b7e343891dd5523ab9daecfca4fed264678fd2d3d704.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " is the number of frames and " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " is the number of points, the goal is to find a " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " permutation matrix " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathsf{P}_{\\pi}" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " that orders the trajectories into sub-matrices " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{X}_i\\}_{i=1}^c" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " corresponding to different motions. Each motion is represented by a " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "2F \\times k_i" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " basis matrix " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_i" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " whose columns span the subspace and therefore " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i = \\mathsf{B}_i\\mathsf{C}_i^T" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathsf{C}_i" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "inline_equation", + "content": "k_i \\times P_i" + }, + { + "bbox": [ + 304, + 333, + 547, + 417 + ], + "type": "text", + "content": " shape matrix. Hence" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 355, + 426, + 545, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 426, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 355, + 426, + 545, + 441 + ], + "type": "interline_equation", + "content": "\\mathrm {M} _ {2 F \\times P} \\mathrm {P} _ {\\pi} \\approx \\left[ \\begin{array}{l l l} \\mathrm {B} _ {1} \\mathrm {C} _ {1} ^ {\\top} & \\dots & \\mathrm {B} _ {c} \\mathrm {C} _ {c} ^ {\\top} \\end{array} \\right], \\tag {2}", + "image_path": "73017ba4f76834aeaa45d6ee25782527a639d3d7bfb8752617f5bb3b08369412.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{c} P_i = P" + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": ", and ranks " + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "inline_equation", + "content": "k_i" + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": " are assumed to be small. In this work we assume that the ranks " + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\{k_1, \\ldots, k_c\\}" + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": " are known which is a common scenario under multiple rigid motions [49]. While estimating a single motion model " + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_i" + }, + { + "bbox": [ + 304, + 449, + 547, + 569 + ], + "type": "text", + "content": " from a cluster of trajectories belonging to the same motion is relatively easy, the combination of cluster assignment and model estimation turns inference into a complicated chicken and egg problem. Further complications arise due to occlusion and partial observations which make it unlikely that we can observe all the full trajectories in all cases." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "In this work we seek to learn a feature representation that allows us to solve the clustering problem separately from the motion estimation. For this purpose we train a simple feed forward network (with the overall structure depicted in Fig. 1) taking a trajectory as input and returning a corresponding feature embedding. Using a metric learning loss we ensure that the features are close if and only if corresponding trajectories belong to the same motions. Our approach allows us to input trajectories of varying lengths and different starting and ending times, making it possible to cluster trajectories using their features without completing any missing data. In summary our main contributions are:" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 683, + 287, + 713 + ], + "type": "text", + "content": "1This work was supported by the Wallenberg Artificial Intelligence, Autonomous Systems and Software Program (WASP), funded by Knut and Alice Wallenberg Foundation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19092" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 79, + 537, + 185 + ], + "blocks": [ + { + "bbox": [ + 60, + 79, + 537, + 185 + ], + "lines": [ + { + "bbox": [ + 60, + 79, + 537, + 185 + ], + "spans": [ + { + "bbox": [ + 60, + 79, + 537, + 185 + ], + "type": "image", + "image_path": "73d4e2bd8bd17b8fabda15038d42e7098775e3a7ad53220c6d3bf6c8b5816d99.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": "Figure 1. Schematic overview of the proposed approach. Input trajectories are passed through the feature extractor to obtain " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": ". The subspace estimator generates the time-dependent basis using coordinate-MLP, infers basis coefficients from " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": ", and outputs subspace B, where each sub-matrix " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_t \\in \\mathbb{R}^{2 \\times 4}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": " is computed via linear combination of basis functions. " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": " is the contrastive loss that uses ground truth cluster assignments, " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Residual}}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": " computes the trajectory-to-subspace geometric error, and " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FeatDiff}}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": " computes the feature difference between input and reconstructed trajectory. For motion segmentation, the trajectories are clustered in a feature space using " + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{f}_i\\}" + }, + { + "bbox": [ + 46, + 201, + 547, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 273, + 286, + 440 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 47, + 273, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 273, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 273, + 286, + 308 + ], + "type": "text", + "content": "- We present a method to map point trajectories to discriminative feature representations, which can be seen as compact descriptions of the underlying generating motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 309, + 286, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 286, + 356 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 286, + 356 + ], + "type": "text", + "content": "- We show that clustering of trajectories can be performed reliably solely using those features, achieving state-of-the-art performance without the need for joint motion estimation and cluster assignments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 357, + 286, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 286, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 286, + 403 + ], + "type": "text", + "content": "- The proposed feature extraction network is trained to be invariant with respect to random occlusions and pixel noise by utilizing a novel generative model that implicitly performs per-trajectory model estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 404, + 286, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 286, + 440 + ], + "type": "text", + "content": "- We present and evaluate different ways to use the proposed network, such as trajectory clustering, trajectory completion and subspace estimation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 452, + 135, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 135, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 135, + 464 + ], + "type": "text", + "content": "1.1. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 688 + ], + "type": "text", + "content": "Within geometric vision the traditionally dominant model-fitting approaches are robust statistical methods based on RANSAC [15]. In case of low dimensional models the approach provides a simple way of effectively exploring the search space. Iterative RANSAC [48], which sequentially finds models and removes their supporting inliers, is a simple option that handles settings with multiple models. It has been observed that the greedy nature of the approach makes it inefficient when a large number of models is present [6, 12, 21, 62]. In contrast, MultiRANSAC [62] simultaneously handles multiple models using iteratively fused consensus sets, and CONSAC [25] increases the efficiency of sampling by learning a sampling strategy based on previous detections. Energy minimization is used in [6, 12, 21] to select models among random proposals that jointly explain the data well. In [7] the hard point-to-model assignment is dropped in favour of progressively determining dominant models which enables clustering in consensus space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "As an alternative to consensus some works use preference [31, 32, 46]. The J-linkage approach [46] uses affini-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 273, + 545, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 345 + ], + "type": "text", + "content": "ties between data and a set of models, followed by agglomerative clustering to assign the data to models. The T-linkage method [32] is based on a continuous relaxation of the binary assignment problem. In [31] the preference representation is used in a robust PCA formulation [10] while [33] uses a set cover approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 350, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 445 + ], + "type": "text", + "content": "Low dimensional geometric models have also been used to group longer point tracks, e.g. in [56] geometric two-view relations are used to create pairwise affinities for a spectral clustering stage. Segmentations computed from images pairs are combined through an averaging approach in [3, 4]. We note that for these local methods motion-specific properties should be observable across all image pairs in order to provide a largely coherent set of clusters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "content": "Subspace models are commonly occurring in geometric computer vision tasks. Early works like [38, 47] showed that structure from motion under affine camera models is essentially a subspace fitting problem. Further, shape deformations are modeled using a low dimensional subspace in [9]. The generalization to multiple independently moving objects leads to segmentation into multiple subspaces [8, 11, 17]. Approaches for handling degenerate and articulated cases were given in [45, 57, 58]. The general problem can be seen as an instance of subspace clustering which assumes that high dimensional data lies in a union of low-dimensional subspaces. The corresponding group of methods aim to simultaneously cluster the data points and estimate the underlying subspaces [1, 26], and may also perform matrix completion [14, 27, 50]. The generalized PCA approach (GPCA) [53-55] treats the union-of-subspaces as a zero-level-set of a higher degree polynomial which is fit to the data. A popular approach to the problem is sparse subspace clustering [5, 13], which has been shown to perform well on the benchmark provided in [49]. These methods are based on the notion of self-expressiveness, i.e. all data in a " + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "content": "-dimensional subspace are linear combinations of " + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 451, + 546, + 713 + ], + "type": "text", + "content": " points" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19093" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 120 + ], + "type": "text", + "content": "from that subspace. This enables an efficient approach that is provably correct in the case of disjoint subspaces (see Section 2). The LRR approach [29, 30] uses a dictionary basis and enforces low rank patterns among the coefficients." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 286, + 262 + ], + "type": "text", + "content": "More recently, several deep learning methods have been proposed for solving the problem of unsupervised subspace clustering. The first work by Ji et al. [23] used deep autoencoder and introduced a self-expressive layer. Later, a closed form solution of the self-expressive layer was derived [43] which improved its efficiency. Alternatively, it was shown [51] that using over-complete representations increases robustness to noise and improves overall performance. Zhang et al. [59] directly learns a self-expressive data representation. The approach is highly scalable since it circumvents the need for a quadratic number of coefficients of the original self-expressive model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 286, + 384 + ], + "type": "text", + "content": "While general subspace learning models typically do not use temporal information, this can be of great importance when analysing and predicting motion. For the reconstruction of deforming objects, Akhtar et al [2] noted the duality between trajectory space and shape space. The use of pre-defined trajectory-bases such as a DCT basis (motivated by the smoothness assumption of time-dependent trajectories [20, 37]) leads to a reduction in the number of unknowns [2]. Finally, a probabilistic framework for subspace clustering with temporal information is presented in [19]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 197, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 197, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 197, + 407 + ], + "type": "text", + "content": "2. Overview of the Approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 286, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 286, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 286, + 545 + ], + "type": "text", + "content": "In this section we present our approach and the underlying architecture of our network. Our goal is to learn function " + }, + { + "bbox": [ + 46, + 415, + 286, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 415, + 286, + 545 + ], + "type": "text", + "content": " that takes an observed point trajectory as input and yields a (relatively low-dimensional) feature embedding that is appropriate for trajectory grouping. For this purpose we adopt a small PointNet-style network [39]. One of the advantages of this type of network, apart from its simplicity, is the built-in invariance with respect to the permutation of the points in the input tensor. In order to obtain a suitable feature representation, we specify a number of desirable properties that we try to infuse through the architecture and training:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 546, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 546, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 593 + ], + "type": "text", + "content": "- The features should be adequate for grouping/clustering of trajectories. For this reason we want a latent space with small within-cluster-distances and large between-cluster-distances. This is achieved using an InfoNCE loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 594, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 594, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 594, + 286, + 677 + ], + "type": "text", + "content": "- Inspired by traditional motion segmentation formulations, we aim for features to encode meaningful \"geometric information\" about the underlying motion subspaces. For this purpose, we design a decoding module that estimates a basis B for the underlying motion subspace that the trajectory belongs to. This prevents input memorization and greatly improves generalization to unseen data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "- Despite input tracks being of varying lengths we opt for a feature embedding of a fixed size. This simplifies treatment of problem instances with missing data and trajec" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 545, + 144 + ], + "type": "text", + "content": "tories starting and ending at different times. The decoding module therefore takes an extra variable, specifying at which time instance the rows of the basis matrix B should be sampled. In this way the model allows us to sub-sample and predict motion information beyond the trajectory samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 144, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 251 + ], + "type": "text", + "content": "- Finally, feature embeddings and intermediate predictions are preferred to behave smoothly with respect to added input noise. To achieve this we construct a \"noise-free\" version of the observed trajectory from the predicted basis, which we require to be close to the original trajectory through a residual loss. Additionally, this reconstructed trajectory is subsequently fed into the network to obtain a corresponding feature vector, which is also required to be close to the original feature embedding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 252, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 545, + 526 + ], + "type": "text", + "content": "Because of the properties above we obtain a trajectory-feature from which we can extract the full motion subspace. Thus we essentially learn a function that maps a single trajectory to a subspace basis. For a general collection of subspaces, where trajectories can belong to multiple of them, this may not be possible (using a purely geometric approach). A simple example is line fitting to image points, where points in the intersection of two models can belong to any of them. In such cases geometry alone is not enough to resolve ambiguous assignments but other cues such as spatial smoothness of assignments [21] have to be used. However, in high-dimensional settings with low-dimensional subspaces the situation is somewhat different. In these cases a typical assumption is that of disjoint subspaces [13], that is, two subspaces only intersect at the origin, ensuring that there is a function that takes (non-zero) trajectories to their corresponding subspaces. We remark that while this assumption is sufficient for our purposes, it is not necessary. Since our method learns a data-driven mapping from point tracks to subspaces, it is able to focus on other trajectory properties than geometric ones. Therefore it can potentially handle the case of intersecting subspaces as long as the trajectory assignment is unambiguous." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "content": "The disjoint subspace assumption We conclude this section by illustrating that the disjoint subspace assumption is valid in practical settings by analysing the data in the Hopkins 155 dataset. Figure 2 shows that re-using subspaces to explain trajectories in the other clusters generally leads to higher errors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Formally, two linear subspaces " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "S_{j}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " are disjoint if " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\dim(S_{i} \\oplus S_{j}) = \\dim(S_{i}) + \\dim(S_{j})" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". Hence if " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_{i}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_{j}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " contain bases for the two subspaces, the rank of the concatenated matrix, " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta_{ij} = \\mathrm{rank}((\\mathsf{B}_i \\mathsf{B}_j))" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", fulfills " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta_{ij} < \\mathrm{rank}(\\mathsf{B}_i) + \\mathrm{rank}(\\mathsf{B}_j)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " if the subspaces intersect. We extracted subsequences of varying length from the Hopkins 155 dataset and used the ground-truth clusters to generate the corresponding subspaces — for all pairs " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathsf{B}_i, \\mathsf{B}_j\\}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{shape}(\\mathsf{B}_i) = \\mathrm{shape}(\\mathsf{B}_j)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", the rank " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\eta_{ij}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " of the concatenated" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19094" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 286, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 286, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 286, + 159 + ], + "type": "image", + "image_path": "f2cf8a5e290db9d7728999b588d62c2b9f439ed21dcba9f5234c0a98852573cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": "Figure 2. RMS normalized error of the trajectory clusters with respect to the subspaces in Hopkins155 dataset (subsequences of length " + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "inline_equation", + "content": "F = 40, 50," + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": " and 60). An " + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": " cell corresponds to the trajectory cluster of the " + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": " motion and the subspace of " + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "inline_equation", + "content": "j^{\\text{th}}" + }, + { + "bbox": [ + 46, + 167, + 288, + 212 + ], + "type": "text", + "content": " motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 288, + 281 + ], + "type": "text", + "content": "matrix was found to be " + }, + { + "bbox": [ + 46, + 232, + 288, + 281 + ], + "type": "inline_equation", + "content": "8^{2}" + }, + { + "bbox": [ + 46, + 232, + 288, + 281 + ], + "type": "text", + "content": ". This suggests that up to numerics the subspaces do not intersect. In the following Section we describe the specific architecture and the employed losses in our approach in more detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 293, + 234, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 234, + 306 + ], + "type": "text", + "content": "3. Architecture and Training Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "content": "The schematic overview of the proposed architecture and training losses is shown in Fig. 1. Using the hypothesis that most observed trajectories come from a unique motion model, we aim to learn a function " + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "content": " mapping a point trajectory directly to a representation of the corresponding motion model. We interpret the output " + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "content": " as an abstract description of the generating motion, from which a concrete representation to measure geometric consistency can be extracted. We choose linear (low-dimensional) subspaces as the concrete realizations of motion models. Specifically, we learn mappings " + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 312, + 287, + 445 + ], + "type": "text", + "content": " using training data such that" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 445, + 287, + 493 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 47, + 445, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 287, + 468 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 47, + 445, + 287, + 468 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 47, + 445, + 287, + 468 + ], + "type": "text", + "content": " maps a trajectory of varying length to a fixed-length feature embedding, and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 468, + 287, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 493 + ], + "type": "text", + "content": "2. " + }, + { + "bbox": [ + 47, + 468, + 287, + 493 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 47, + 468, + 287, + 493 + ], + "type": "text", + "content": " converts this feature vector for a given query time into the corresponding rows of the subspace representation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 501, + 159, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 501, + 159, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 501, + 159, + 512 + ], + "type": "text", + "content": "3.1. Feature Extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "text", + "content": "To construct a discriminative function " + }, + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "text", + "content": ", that is able to infer the motion from an input point trajectory, we first leverage the ability of the neural networks to compress high-dimensional complex data space into a lower-dimensional latent space with task specific invariance properties, from which it is easier to infer the output. A feature extractor for the point trajectories " + }, + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "inline_equation", + "content": "f_{\\theta}: \\mathbf{x} \\mapsto \\mathbf{f} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 46, + 519, + 287, + 676 + ], + "type": "text", + "content": " is therefore trained to produce invariant feature representations using ground truth clusters. In particular, an invariance with respect to the measurement noise and the variations within the motion clusters is of our interest. We therefore use a contrastive loss, namely a variant of the InfoNCE loss [36], to conduct metric learning, leading to a spread of feature embeddings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 518, + 85 + ], + "type": "text", + "content": "in order to improve feature-based motion clustering," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 91, + 545, + 153 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 91, + 545, + 153 + ], + "spans": [ + { + "bbox": [ + 331, + 91, + 545, + 153 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {I n f o N C E}} = - \\frac {1}{| \\mathcal {D} |} \\sum_ {(i, j, l, k) \\in \\mathcal {D}} \\log \\left(\\frac {p _ {i j}}{p _ {i j} + p _ {l k}}\\right), \\\\ p _ {i j} = \\exp \\left(- \\frac {\\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| _ {2} ^ {2}}{T}\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "6f97db610c8a7d5e8c309af47fdbe68f2917ca2c3ed5d27556fab74dd7081385.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i = f_{\\theta}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " is the set of quadruples " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "(i,j,l,k)" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "(i,k)" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " belong to the same cluster and " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "(l,k)" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " are random pairs, and " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " is the temperature (hyper-)parameter, which we fix to 1. " + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 304, + 160, + 545, + 233 + ], + "type": "text", + "content": " is an instance of a supervised contrastive loss and is tightly connected to the triplet loss used in metric learning [24]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "text", + "content": "As mentioned earlier, the proposed architecture resembles PointNet [39] with several modifications. The employed encoder consists of 1D temporal convolutional layers applied to each trajectory independently, with x and y values representing two input channels. A global max-pooling in a temporal domain is then applied on the time-dependent representations, followed by a small MLP and an " + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "text", + "content": " normalization. The features are therefore lying on the " + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "inline_equation", + "content": "d - 1" + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "text", + "content": "-dimensional unit sphere. Note that we do not apply pooling across the points (global context layer) due to the disjoint subspace assumption. For the experiments, we set the dimensionality of the latent space to " + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "inline_equation", + "content": "d = 128" + }, + { + "bbox": [ + 304, + 233, + 546, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 383, + 426, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 426, + 396 + ], + "type": "text", + "content": "3.2. Subspace Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": "The latent vector " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " ideally identifies a motion model, but for geometric consistency this feature embedding requires conversion into a suitable mathematical motion model. We assume that observed point trajectories are generated by low-rank subspace models, and therefore the feature embedding is mapped by a subsequent network " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " to the corresponding matrix representation of the associated subspace. Specifically, the mapping " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "g_{\\phi}: \\mathcal{F} \\times \\mathbb{R} \\rightarrow \\mathbb{R}^{2 \\times r}" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " produces rows " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "\\mathsf{B}_t" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " for the given motion model " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{f} = f_{\\theta}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " and the time index " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": ". Ideally the input trajectory " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " lies in the predicted subspace; if the trajectory " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": " is consisting of (2D) points observed at times " + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "inline_equation", + "content": "t_1, \\ldots, t_F" + }, + { + "bbox": [ + 304, + 401, + 546, + 558 + ], + "type": "text", + "content": ", then the output subspace is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 565, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 565, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 375, + 565, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\mathsf {B} = \\left( \\begin{array}{c} g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {1}\\right) \\\\ \\vdots \\\\ g _ {\\phi} \\left(f _ {\\theta} (\\mathbf {x}), t _ {F}\\right) \\end{array} \\right), \\tag {4}", + "image_path": "8c6ff1757c77ff583f250d4684455e45cc0dc9c2d0dee1888361d110e0057412.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 617, + 518, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 617, + 518, + 630 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 518, + 630 + ], + "type": "text", + "content": "and the (squared) residual is given in closed-form as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 359, + 638, + 545, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 638, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 359, + 638, + 545, + 657 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {c}} \\| \\mathbf {x} - \\mathrm {B c} \\| _ {2} ^ {2} = \\| \\mathbf {x} - \\mathrm {B B} ^ {\\dagger} \\mathbf {x} \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "41b41667ef535adab3015edab6aec15f99c6a813584fff0c3b148722d5cfbc4b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "text", + "content": " is the coefficient vector and " + }, + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{B}^{\\dagger} = (\\mathsf{B}^{\\top}\\mathsf{B})^{-1}\\mathsf{B}^{\\top}" + }, + { + "bbox": [ + 304, + 664, + 546, + 713 + ], + "type": "text", + "content": ". If the network identified the subspace correctly, then the residual is in the order of the observation noise level (and therefore small)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": "2The matrix rank is computed as the number of singular values greater than " + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\delta \\sigma_{1}" + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{1}" + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": " is the largest singular value, " + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 2F\\varepsilon" + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\varepsilon \\approx 10^{-7}" + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": " is the numerical zero for torch.float32." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19095" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "Subspace basis functions The network " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " can be a small but generic MLP. Alternatively, we may leverage existing domain knowledge and choose the structure of " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " accordingly. In particular, we assume that all motions are linear combination of a finite, but continuous time trajectory basis " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\{h^j\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": ". The basis functions " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "h^j" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " can be fully fixed (such as a cosine basis with pre-determined frequencies and phase shifts [2]), parametric (such as the cosine basis with trainable frequencies and phase shifts [44]) or entirely non-parametric (i.e. represented by a generic MLP). Our choice of basis functions is a damped version of a cosine basis and is given as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 224, + 287, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 224, + 287, + 242 + ], + "spans": [ + { + "bbox": [ + 89, + 224, + 287, + 242 + ], + "type": "interline_equation", + "content": "h _ {\\psi} ^ {j} (t) = e ^ {- \\left(\\alpha_ {j} (t - \\mu_ {j})\\right) ^ {2}} \\cos (\\beta_ {j} t + \\gamma_ {j}), \\tag {6}", + "image_path": "317c213dd8f8463371c3423667de7506efbb9aba616b536de3c4e069689d6853.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "inline_equation", + "content": "\\psi = (\\mu_j, \\alpha_j, \\beta_j, \\gamma_j)_j" + }, + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "text", + "content": " are trainable parameters. The mapping " + }, + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 47, + 249, + 287, + 274 + ], + "type": "text", + "content": " can therefore be written as a composition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 281, + 287, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 281, + 287, + 295 + ], + "spans": [ + { + "bbox": [ + 118, + 281, + 287, + 295 + ], + "type": "interline_equation", + "content": "g _ {\\phi} (\\mathbf {f}, t) = \\omega_ {\\zeta} (\\mathbf {f}) \\mathbf {h} _ {\\psi} (t), \\tag {7}", + "image_path": "a448f1f5e87a90b4345487036a5a620e1f4b28c81d9313d064de6925a5808282.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\psi}(t) = (h_{\\psi}^{j}(t))_{j}" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": ", and the trainable function " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\omega_{\\zeta}" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": " extracts the basis coefficients from " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": ". If we have " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": " basis functions (i.e. " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\dim (\\mathbf{h}_{\\psi}(t)) = N" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": "), then " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\omega_{\\zeta}(\\mathbf{f})" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "2r\\times N" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": " matrix. The total set of trainable parameters in the subspace estimation network is " + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "inline_equation", + "content": "\\phi = (\\zeta ,\\psi)" + }, + { + "bbox": [ + 47, + 305, + 287, + 378 + ], + "type": "text", + "content": ". Finally, the network prediction can be compactly written as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 387, + 287, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 387, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 121, + 387, + 287, + 400 + ], + "type": "interline_equation", + "content": "\\mathsf {B} = \\mathrm {H} _ {\\psi} (\\mathbf {t}) \\Omega_ {\\zeta} \\left(f _ {\\theta} (\\mathbf {x})\\right), \\tag {8}", + "image_path": "c1caecf757da27cf51af942fe84e17da655431819a523f5cc10fa083fceb2bc2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 408, + 217, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 408, + 217, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 408, + 217, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 408, + 217, + 422 + ], + "type": "inline_equation", + "content": "\\Omega_{\\zeta}(\\mathbf{f}) = \\mathrm{reshape}_{2N\\times r}\\left(\\omega_{\\zeta}(\\mathbf{f})\\right)" + }, + { + "bbox": [ + 47, + 408, + 217, + 422 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 430, + 287, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 430, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 58, + 430, + 287, + 500 + ], + "type": "interline_equation", + "content": "\\mathrm {H} _ {\\psi} (\\mathbf {t}) = \\left( \\begin{array}{l l} \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {1}) ^ {\\top} \\\\ \\vdots & \\\\ \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} & \\\\ & \\mathbf {h} _ {\\psi} (t _ {F}) ^ {\\top} \\end{array} \\right) \\in \\mathbb {R} ^ {2 F \\times 2 N}. \\tag {9}", + "image_path": "346d9736c14d6dab492fa3a67133a8248cb4ee09c41a7a54d067feacc81beae4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 508, + 287, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 579 + ], + "type": "text", + "content": "We set the number of subspace basis functions to " + }, + { + "bbox": [ + 47, + 508, + 287, + 579 + ], + "type": "inline_equation", + "content": "N = 64" + }, + { + "bbox": [ + 47, + 508, + 287, + 579 + ], + "type": "text", + "content": " in our experiments. Since the predicted subspace matrix B augments the feature representation (see Sec. 3.3 below), we enforce a unique subspace representation by fixing the top block of B to the identity matrix (by multiplying with the respective inverse)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Time dependence and time invariance The network represented by " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " can be interpreted as taking a \"feature\" token " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " and a time coordinate " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " as input, and is therefore somewhat analogous to the coordinate-MLP employed for implicit neural primitives such as [35, 44] or time series modeling [16]. The continuous time basis " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\{h^j\\}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " can be seen as domain-specific equivalent to positional coding, but its design is based on the expected motion patterns and does not necessarily meet the requirements of generic positional coding [40, 44, 60]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "The formulation above is intentionally not time-invariant: the same trajectory observed at different starting times may predict different motion models. This is usually a desired behavior for motion segmentation, e.g. two cars traveling the same path at different time form different motion segments. As with CNNs applied to images, low-level trajectory features are extracted by convolutional layers, and time-dependence is incorporated via final fully connected and pooling layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 186, + 370, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 186, + 370, + 199 + ], + "spans": [ + { + "bbox": [ + 306, + 186, + 370, + 199 + ], + "type": "text", + "content": "3.3. Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "text", + "content": "The total training loss is the weighted sum of a contrastive loss " + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "text", + "content": " applied on " + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{v}_i\\}" + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i = (\\mathrm{flatten}(\\mathsf{B}_i)^\\top \\mathbf{f}_i^\\top)^\\top" + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "text", + "content": ", the reconstruction loss " + }, + { + "bbox": [ + 305, + 205, + 545, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Residual}}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 248, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 248, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 362, + 248, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {R e s i d u a l}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| \\mathbf {x} _ {j} - \\tilde {\\mathbf {x}} _ {j} \\| _ {2} ^ {2}, \\tag {10}", + "image_path": "feacbbfdeb2ae53d48e727ccdbae9e7b83cf4848999117f1e48f41d9be63e03e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 289, + 429, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 289, + 429, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 289, + 429, + 300 + ], + "type": "text", + "content": "and the feature difference loss" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 305, + 545, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 305, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 346, + 305, + 545, + 339 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {F e a t D i f f}} = \\frac {1}{P} \\sum_ {j = 1} ^ {P} \\| f _ {\\theta} (\\mathbf {x} _ {j}) - f _ {\\theta} (\\tilde {\\mathbf {x}} _ {j}) \\| _ {2} ^ {2}, \\tag {11}", + "image_path": "6f7d939fdbbbbc268229cd9c8681e68791694aabc6a58939bd943d8d78e2f578.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_j = \\mathsf{B}(\\mathbf{x}_j)\\mathsf{B}(\\mathbf{x}_j)^\\dagger\\mathbf{x}_j" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " is the projection of " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathsf{B}(\\mathbf{x}_j)" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": ". In short, " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " forces both an approximate invariance of the feature extractor " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " with respect to the variations within clusters and smoothness of the coordinate-MLP " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Residual}}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " ensures geometric consistency of the subspaces with respect to input trajectories, and " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FeatDiff}}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " forces an approximate invariance of " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " with respect to the pixel noise and can also be seen as enforcing smoothness of " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": ". We train the network in two stages. First, we pre-train " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{InfoNCE}}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " until convergence. Then we add " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " and train it together with " + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 347, + 545, + 515 + ], + "type": "text", + "content": " using the total loss. To enforce the approximate invariance with respect to random detection failures, we synthesize missing points during training by applying dropout to the input sequences." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 521, + 436, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 436, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 436, + 534 + ], + "type": "text", + "content": "3.4. Trajectory Completion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "B_{\\theta, \\phi}(\\mathbf{x}, \\mathbf{t})" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": " provide the subspace for an input trajectory " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{2F}" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": " and time vector " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{t} = (t_1, \\dots, t_F)^\\top" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": ". Assume now that " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": " is only partially observed, and let " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\in \\{0, 1\\}^{2F}" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": " be the visibility mask. Further, let " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{w}} := \\mathbf{1} - \\mathbf{w}" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": ". The goal of the trajectory completion is to estimate missing values " + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 540, + 545, + 612 + ], + "type": "text", + "content": " that minimize the following objective" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 618, + 545, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 618, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 618, + 545, + 635 + ], + "type": "interline_equation", + "content": "J (\\bar {\\mathbf {x}}) := \\left\\| (I - B (\\bar {\\mathbf {x}}) B (\\bar {\\mathbf {x}}) ^ {\\dagger}) (w \\odot x + \\bar {w} \\odot \\bar {x}) \\right\\| ^ {2}, \\tag {12}", + "image_path": "d0f7727be70343e28e10f3d26181373f103deced4140b636ec7e33aa5733629b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathsf{B}(\\bar{\\mathbf{x}}) = B_{\\theta ,\\phi}(\\mathbf{w}\\odot \\mathbf{x} + \\bar{\\mathbf{w}}\\odot \\bar{\\mathbf{x}},\\mathbf{t})" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " . Back-propagation for gradient-based determination of " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " is in principle possible but computationally expensive. Alternatively, we rely on the approximate invariance of the feature extractor " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " with respect to the length of the input sequence and extrapolate the output subspace in time using " + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "g_{\\phi}(\\mathbf{f},t)" + }, + { + "bbox": [ + 304, + 641, + 545, + 714 + ], + "type": "text", + "content": " . More" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19096" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 569, + 208 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 569, + 208 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 569, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 569, + 208 + ], + "type": "table", + "html": "
Method2 motionsHopkins155AllHopkins12KT3DMoSeg
MeanMedianTimeMeanMedianTimeMeanMedianTimeMeanMedianMeanMedian
RANSAC [48]5.561.18175ms22.9422.03258ms9.763.21194ms----
GPCA [54]4.590.38324ms28.6628.26738ms10.342.54417ms--34.6033.95
MSL [45]4.140.0011h 4m8.231.761d 23h5.030.0019h 11m----
LSA [57]3.450.597.58s9.732.3315.96s4.940.909.47s--38.3038.58
ALC5 [42]3.030.00-6.261.02-3.760.265m 15s3.810.1724.3119.04
ALCsp [42]2.400.43-6.690.67-3.370.496m 11s1.281.07--
LRR [30]4.100.22-9.890.56-5.410.531.1s--33.6736.01
SSC [13]0.820.00-2.450.20-2.450.20920ms--33.8833.54
RSIM [22]0.780.00-1.770.28-1.010.00176ms0.680.70--
MultiConsensus [7]------4.40-40ms----
Ours0.630.07ms0.600.010ms0.620.09ms5.122.045.850.80
", + "image_path": "bb7d54b216d6c9e029521371a1eec308e4ebbd7149caed979f55a1d1261b7013.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": "Table 1. Motion segmentation results on the Hopkins155 [49], Hopkins12 [41] and KT3DMoSeg [56] datasets. The evaluation metrics is classification error " + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 46, + 217, + 546, + 240 + ], + "type": "text", + "content": ". ' -' indicates that the results were not reported / no code available." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": "specifically, the subspace of " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": " can be estimated from " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": "-a fully visible sub-trajectory of the trajectory " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": " as long as the feature representation " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": " can be inferred (i.e. " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": " is sufficiently long). If we fix this estimate " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": " and let " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\mathsf{W} := \\mathrm{diag}(\\mathbf{w})" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "inline_equation", + "content": "\\bar{\\mathsf{W}} := \\mathrm{diag}(\\bar{\\mathsf{w}})" + }, + { + "bbox": [ + 46, + 258, + 288, + 330 + ], + "type": "text", + "content": ", the new objective can be stated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 339, + 287, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 339, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 69, + 339, + 287, + 358 + ], + "type": "interline_equation", + "content": "\\tilde {J} (\\bar {\\mathbf {x}}) := \\left\\| \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\mathbf {x} + \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {W} \\bar {\\mathbf {x}} \\right\\| ^ {2}, \\tag {13}", + "image_path": "73e4694f7558123edd6ef38344d767d8f19553a879c1b61461f4845da5b60cf5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "text", + "content": "and the minimizer " + }, + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{J}" + }, + { + "bbox": [ + 47, + 369, + 203, + 383 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 394, + 287, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 394, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 59, + 394, + 287, + 410 + ], + "type": "interline_equation", + "content": "\\bar {\\mathbf {x}} ^ {*} = - \\left(\\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\bar {\\mathbf {w}}\\right) ^ {\\dagger} \\left(\\mathrm {I} - \\mathrm {B B} ^ {\\dagger}\\right) \\mathbf {W x} =: \\mathrm {A} (\\mathrm {B}) \\mathbf {x}. \\tag {14}", + "image_path": "a9028b9627dc216c50c3519caed14eaaa08aa273669cdbd365c01f2e60ed1ddd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": "It is now possible to refine the estimated subspace using " + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": " by feeding " + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{w} \\odot \\mathbf{x} + \\bar{\\mathbf{w}} \\odot \\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 46, + 422, + 287, + 471 + ], + "type": "text", + "content": " into the network. Overall, we apply the following iterative procedure for trajectory completion," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 481, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 481, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 81, + 481, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{r l} \\mathsf {B} _ {0} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {x} _ {\\text {v i s}}, \\mathbf {t}) \\\\ \\bar {\\mathbf {x}} _ {i + 1} & \\leftarrow \\mathsf {A} (\\mathsf {B} _ {i}) \\mathbf {x} \\\\ \\mathsf {B} _ {i + 1} & \\leftarrow B _ {\\theta , \\phi} (\\mathbf {w} \\odot \\mathbf {x} + \\bar {\\mathbf {w}} \\odot \\bar {\\mathbf {x}} _ {i + 1}, \\mathbf {t}). \\end{array} \\right. \\tag {15}", + "image_path": "5de590f741db68b6d978fff2dabaf70063d7dcdc0dfff8eedb9c5d43da2fe37b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 50, + 547, + 167, + 664 + ], + "blocks": [ + { + "bbox": [ + 50, + 547, + 167, + 664 + ], + "lines": [ + { + "bbox": [ + 50, + 547, + 167, + 664 + ], + "spans": [ + { + "bbox": [ + 50, + 547, + 167, + 664 + ], + "type": "image", + "image_path": "70641f339b1ce61533b8b379eb5789df914a5deffdaedfa664b3a0dbec7aec9a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 674, + 287, + 696 + ], + "lines": [ + { + "bbox": [ + 46, + 674, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 674, + 287, + 696 + ], + "type": "text", + "content": "Figure 3. Histograms of the classification errors for the groups with (left) 2 and (right) 3 motions in Hopkins155 dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 169, + 545, + 284, + 664 + ], + "blocks": [ + { + "bbox": [ + 169, + 545, + 284, + 664 + ], + "lines": [ + { + "bbox": [ + 169, + 545, + 284, + 664 + ], + "spans": [ + { + "bbox": [ + 169, + 545, + 284, + 664 + ], + "type": "image", + "image_path": "e2482bb9b3f2595898ed145e93ff942b73970873bd6d44eb021a4242821af9ac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 257, + 387, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 257, + 387, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 257, + 387, + 271 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 337 + ], + "type": "text", + "content": "The proposed network was trained on the two motion segmentation datasets: Hopkins155 [49] and KT3DMoSeg [56], yielding two different sets of weights. In our experiments, we used NVIDIA GeForce RTX 3080, CUDA 11.5, PyTorch 1.11." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "text", + "content": "To obtain groups of trajectories, it is sufficient to perform a partial forward pass through the (trained) feature extractor " + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "text", + "content": " and cluster in the feature space. We use hierarchical clustering as it provides a meaningful framework for selecting the number of clusters, i.e. motions. The subspaces can then be fit to each cluster. If the goal is to estimate the subspace for a single input trajectory, a full forward pass through " + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 304, + 337, + 546, + 494 + ], + "type": "text", + "content": " should be performed. For motion segmentation with missing data, the largest fully visible blocks of trajectories (but ignoring single occluded points) can be used to estimate the initial subspaces B, followed by trajectory completion as outlined in (15). The completed trajectories can then be used to cluster in the feature space." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 500, + 430, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 430, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 430, + 513 + ], + "type": "text", + "content": "4.1. Motion Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "content": "We evaluate the performance of the proposed network in terms of motion segmentation on Hopkins155 [49], Hopkins12 [41], and KT3DMoSeg [56] datasets. We first compare the proposed method to the state-of-the-art algorithms that similarly assume rank-4 of B: sequential RANSAC [48], GPCA [54], MSL [45], LSA [57] (with the global subspace dimensionality " + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "inline_equation", + "content": "D = 4n" + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "content": " clusters), " + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{ALC}_5" + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{ALC}_{\\mathrm{sp}}" + }, + { + "bbox": [ + 304, + 518, + 547, + 675 + ], + "type": "text", + "content": " [42], LRR [30], SSC [13], RSIM [22], and MultiConsensus [7]. We then also compare our method to the other approaches making different assumptions on the underlying motion model [3, 4, 28, 56]. The metrics used to evaluate clustering performance is the classification error (\\%) which is standard for this benchmark [49]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "Results on Hopkins155 dataset Hopkins155 [49] is the commonly used benchmark for trajectory-based motion" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19097" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 214, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 214, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 214, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 214, + 178 + ], + "type": "image", + "image_path": "eb6f11bcbfd11722946152a0079c5a998a3ad2a92021a9f05d2a0994b55613a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 222 + ], + "type": "text", + "content": "Figure 4. The approximate invariance measured as an angle between features extracted from the reference trajectory and its perturbed version using (left) Gaussian pixel noise, (middle) perturbation within subspace, and (right) uniform occlusions. The gradually increasing level of corruption is shown on the x-axis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 71, + 375, + 178 + ], + "blocks": [ + { + "bbox": [ + 219, + 71, + 375, + 178 + ], + "lines": [ + { + "bbox": [ + 219, + 71, + 375, + 178 + ], + "spans": [ + { + "bbox": [ + 219, + 71, + 375, + 178 + ], + "type": "image", + "image_path": "34c5c2d4b6876d165fda2418c5ef64935724c2015b305639df503d5738f3facc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 536, + 179 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 536, + 179 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 536, + 179 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 536, + 179 + ], + "type": "image", + "image_path": "d56a9502e2b503c35590e5bd252a89b5368f3790416b0f4458253b4996d2fc7a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 242, + 288, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 242, + 288, + 337 + ], + "spans": [ + { + "bbox": [ + 46, + 242, + 288, + 337 + ], + "type": "text", + "content": "segmentation. The dataset consists of 155 sequences with 63-556 point trajectories generated by two or three motions in the scene. Each trajectory is fully visible, and the length varies from 15 to 100 frames. The averaged results are shown in the left block of Table 1, and the histograms of classification errors are shown in Fig. 3. Our network is very fast and produces the most accurate segmentations on fully visible trajectories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": "Results on Hopkins12 dataset Hopkins12 [41] is a commonly used dataset for evaluating subspace clustering with missing data. It consists of 12 corrupted sequences from Hopkins155 dataset, where the corruption was done manually by labeling entries of trajectories as missing. The corruption rate of the observation matrix M varies from " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "23\\%" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathrm{ALC}_5" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\mathrm{ALC_{sp}}" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": " [42], we report " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "l^1" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": "-based entry completion results. The results are shown in the middle block of Table 1. The proposed network shows promising results. Note that the maximum corruption rate of the individual trajectories in the Hopkins12 dataset is " + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 46, + 346, + 287, + 583 + ], + "type": "text", + "content": " and, unlike all the state-of-the-art approaches, the proposed trajectory completion method is not global, i.e. it works with individual trajectories. Therefore, if a provided individual sub-trajectory does not contain sufficient relevant information for motion estimation (which becomes more likely as the trajectory gets shorter), incorporating information from all tracks becomes necessary to disambiguate the true motion. Utilizing such global trajectory information is left for future research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Results on KT3DMoSeg dataset KT3DMoSeg [56] is an adapted version of the KITTI dataset [18, 34] containing 22 sequences with 158-1018 point trajectories and annotated segmentations. It is more challenging than Hopkins155 in several aspects. There is a dominant motion corresponding to static \"background\" scene which makes cluster distribution unbalanced. The trajectories are also highly occluded: the corruption rate of M goes up to " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ", and for the individual trajectories — up to " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ". Lastly, the sequences have shorter length of 10 to 20 frames. The results for this dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 242, + 545, + 266 + ], + "type": "text", + "content": "are shown in the right block of Table 1 indicating that the proposed approach outperforms previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 271, + 546, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 546, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 546, + 438 + ], + "type": "text", + "content": "Methods with different assumptions We also compared our approach to the other methods making different assumptions. Two-perspective-view (TPV) [28] assumes perspective cameras and uses pairwise epipolar relations. Spectral clustering methods proposed by [56]: kernel addition (KerAdd), co-regularization (CoReg), and subset constrained (Subset) spectral clustering also look at pairwise relations, but use fundamental matrices, affine matrices or homographies to compute the pairwise affinities. SYNCH [3] and MODE [4] similarly fit multiple fundamental matrices to each pair of images, but perform averaging. They assume that only pairwise matches are given. These methods are not directly comparable to ours, however the proposed method proves to be competitive as evidenced in Table 2." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 306, + 443, + 545, + 544 + ], + "blocks": [ + { + "bbox": [ + 306, + 443, + 545, + 544 + ], + "lines": [ + { + "bbox": [ + 306, + 443, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 545, + 544 + ], + "type": "table", + "html": "
MethodHopkins155 MeanHopkins12 Med.KT3DMoSeg MeanMed.
TPV [28]2.34----
KerAdd [56]0.360.110.08.311.02
CoReg [56]0.460.060.07.920.75
Subset [56]0.310.060.08.080.71
SYNCH-tracks [3]3.675.460.57--
SYNCH [3]1.193.190.28--
MODE [4]1.374.330.38--
Our method0.625.122.045.850.80
", + "image_path": "2e19c972bab8f85a6480b434abcc640ccf5e1fccdd2b115b0260635417ba4161.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 552, + 545, + 573 + ], + "lines": [ + { + "bbox": [ + 305, + 552, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 305, + 552, + 545, + 573 + ], + "type": "text", + "content": "Table 2. Comparison to the other methods making different assumptions. TPV [28] has no code publicly available." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 587, + 471, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 471, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 471, + 600 + ], + "type": "text", + "content": "4.2. Approximate Invariances of " + }, + { + "bbox": [ + 306, + 587, + 471, + 600 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "We experimentally validate the approximate invariance of the trained feature extractor " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " with respect to different types of synthetic data corruption: (1) Gaussian pixel noise with the standard deviation " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " gradually increasing up to 10 px, (2) perturbations within the subspace, where the trajectories are first corrupted by Gaussian pixel noise with " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " up to 300 px and then projected back to the underlying subspace, (3) synthesized uniform occlusions with missing data levels going up to 50%, and (4) tracking failures, where" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19098" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 214, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 214, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 214, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 214, + 178 + ], + "type": "image", + "image_path": "54c344a69caf4d564a94fdce18966557b2ac3337c72a1fab150f3cfa15f8dda4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "text", + "content": "Figure 5. The approximate invariance measured as an angle between features extracted from the reference trajectory and its shortened version for (left) shorter sequences with " + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "inline_equation", + "content": "F = 15 \\ldots 24" + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "text", + "content": ", (middle) mid-length sequences with " + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "inline_equation", + "content": "F = 25 \\ldots 31" + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "text", + "content": ", and (right) longer sequences with " + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "inline_equation", + "content": "F = 35 \\ldots 61" + }, + { + "bbox": [ + 46, + 189, + 547, + 224 + ], + "type": "text", + "content": ". The gradually increasing percentage of sequence length reduction is shown on the x-axis." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 375, + 178 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 375, + 178 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 375, + 178 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 375, + 178 + ], + "type": "image", + "image_path": "d999bcf6c79aa629d8c035d99a2a535b35ba2f87890f0d2c8410bdd95b458931.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 379, + 72, + 535, + 178 + ], + "blocks": [ + { + "bbox": [ + 379, + 72, + 535, + 178 + ], + "lines": [ + { + "bbox": [ + 379, + 72, + 535, + 178 + ], + "spans": [ + { + "bbox": [ + 379, + 72, + 535, + 178 + ], + "type": "image", + "image_path": "8264cabf08a73bbc65ae118e87608d0a284f3ea6844eeb1988898deeed37f7b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 327 + ], + "type": "text", + "content": "the rearmost points are gradually excluded from the trajectories. All types of corruption are applied on the training sequences of Hopkins155 dataset. The invariances are measured using the angle between the features extracted from original and corrupted trajectories. The smaller the angle is, the more invariant the feature is to the certain type of data corruption." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": "Results for the pixel noise, perturbations within the subspace, and uniform occlusions are shown in Fig. 4. Note that the features are " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": "-normalized in the " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": "-dimensional space for " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "d = 128" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": ". In such a high-dimensional space the two randomly picked vectors are likely to be almost orthogonal [52]. The network proves to be approximately invariant hence robust to these types of corruption, and especially to the pixel noise. Since 8-10 px is quite large (for the images of sizes " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "320 \\times 240" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": "), this result also shows robustness of " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": " to occasional outlying points in the trajectory. The results for the synthesized tracking failures are shown in Fig. 5, where a single curve corresponds to a certain initial trajectory length " + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 327, + 290, + 544 + ], + "type": "text", + "content": ". The angles are slightly higher for shorter sequences, which is expected since they contain less information. Note that during training we do not force the network to produce the same feature representations for the trajectories of the same point starting at the same time but having different length." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 52, + 559, + 284, + 681 + ], + "blocks": [ + { + "bbox": [ + 52, + 559, + 284, + 681 + ], + "lines": [ + { + "bbox": [ + 52, + 559, + 284, + 681 + ], + "spans": [ + { + "bbox": [ + 52, + 559, + 284, + 681 + ], + "type": "image", + "image_path": "2cbfb2bf4ea068025b17893b6eea77e0fbdf682ea68c279e949e97f095278ed0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "type": "text", + "content": "Figure 6. Mean pixel error of recovering from random occlusions for the corruption rates of (left) " + }, + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "type": "inline_equation", + "content": "10 - 40\\%" + }, + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "type": "text", + "content": " and (right) " + }, + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "type": "inline_equation", + "content": "50 - 80\\%" + }, + { + "bbox": [ + 46, + 688, + 287, + 710 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 243, + 436, + 256 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 436, + 256 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 436, + 256 + ], + "type": "text", + "content": "4.3. Trajectory Completion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 264, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 408 + ], + "type": "text", + "content": "We further evaluate the completion accuracy of the proposed method on the trajectories from Hopkins155 dataset with synthesized uniform occlusions. The performance is measured using the mean pixel error between the ground truth (i.e., detected/tracked) values of the synthetically occluded points and the predicted values. The error as a function of iterations for different corruption rates is shown in Fig. 6, where we split the results into a minor and a major corruption group. From the results we conclude that it is sufficient to run the algorithm (15) for 3 to 5 iterations for minor corruptions which takes 1s on average, and 5 to 20 iterations for major corruptions which takes 3s on average." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "spans": [ + { + "bbox": [ + 306, + 426, + 378, + 438 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 545, + 604 + ], + "type": "text", + "content": "In this paper we propose to learn a mapping from point trajectories to feature vectors in an embedding space, which fully identify the generating motion. We illustrate that these feature vectors can be directly used for subspace clustering, thereby eliminating the need for simultaneous estimation of assignments and subspace models. The proposed training loss ensures not only the utility of the obtained embedding for clustering, but also forces the feature vector to encode the actual motion parameters, which are retrieved by a dedicated coordinate-MLP. Comparisons to baselines on Hopkins155 and KT3DMoSeg highlight the efficiency and accuracy of our method, providing top clustering results with running time below 10ms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "In this work we focus on the linear subspace framework as sole motion model to explain observed trajectories. One direction for future work is how to incorporate additional and/or different prior assumptions on the motion model into our method. Further, the linear subspace assumption is strongly tied to affine cameras, which only approximate the more realistic pinhole camera model. Therefore extending this work to more general camera models is a challenging direction for future research." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19099" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopoulos, and Prabhakar Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. In Proceedings of the 1998 ACM SIGMOD international conference on Management of data, pages 94-105, 1998. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Ijaz Akhter, Yaser Sheikh, Sohaib Khan, and Takeo Kanade. Trajectory space: A dual representation for nonrigid structure from motion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(7):1442-1456, 2010. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 236 + ], + "type": "text", + "content": "[3] Federica Arrigoni and Tomas Pajdla. Motion segmentation via synchronization. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 238, + 287, + 282 + ], + "type": "text", + "content": "[4] Federica Arrigoni and Tomas Pajdla. Robust motion segmentation from pairwise matches. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 671-681, 2019. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 287, + 316 + ], + "type": "text", + "content": "[5] Liang Bai and Jiye Liang. Sparse subspace clustering with entropy-norm. In International conference on machine learning, pages 561-568. PMLR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[6] Daniel Barath and Jiri Matas. Progressive-x: Efficient, anytime, multi-model fitting algorithm. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3780-3788, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 418 + ], + "type": "text", + "content": "[7] Daniel Barath, Denys Rozumnyi, Ivan Eichhardt, Lerente Hajder, and Jiri Matas. Finding geometric models by clustering in the consensus space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5414-5424, 2023. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 453 + ], + "type": "text", + "content": "[8] T.E. Boult and L. Gottesfeld Brown. Factorization-based segmentation of motions. In Proceedings of the IEEE Workshop on Visual Motion, pages 179-186, 1991. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 508 + ], + "type": "text", + "content": "[9] Christoph Bregler, Aaron Hertzmann, and Henning Biermann. Recovering non-rigid 3d shape from image streams. In Proceedings IEEE Conference on Computer Vision and Pattern Recognition. CVPR 2000 (Cat. No. PR00662), pages 690-696. IEEE, 2000. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 510, + 287, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 541 + ], + "type": "text", + "content": "[10] Emmanuel J. Candès, Xiaodong Li, Yi Ma, and John Wright. Robust principal component analysis? J. ACM, 58(3), 2011. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 587 + ], + "type": "text", + "content": "[11] J. Costeira and T. Kanade. A multi-body factorization method for motion analysis. In Proceedings of IEEE International Conference on Computer Vision, pages 1071-1076, 1995. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 623 + ], + "type": "text", + "content": "[12] A. Delong, A. Osokin, H. Isack, and Y. Boykov. Fast approximate energy minimization with label costs. International Journal of Computer Vision, 96(1):1-27, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[13] Ehsan Elhamifar and René Vidal. Sparse subspace clustering: Algorithm, theory, and applications. IEEE transactions on pattern analysis and machine intelligence, 35, 2013. 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[14] Jicong Fan and Madeleine Udell. Online high rank matrix completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8690-8698, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 160 + ], + "type": "text", + "content": "[16] Elizabeth Fons, Alejandro Sztrajman, Yousef El-Laham, Alexandros Iosifidis, and Svitlana Vyetrenko. Hypertime: Implicit neural representation for time series. arXiv preprint arXiv:2208.05836, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[17] C.W. Gear. Feature grouping in moving objects. In Proceedings of 1994 IEEE Workshop on Motion of Non-rigid and Articulated Objects, pages 214-219, 1994. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "text", + "content": "[19] Behnam Gholami and Vladimir Pavlovic. Probabilistic temporal subspace clustering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3066-3075, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 337 + ], + "type": "text", + "content": "[20] Paulo FU Gotardo and Aleix M Martinez. Computing smooth time trajectories for camera and deformable shape in structure from motion with occlusion. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(10):2051-2065, 2011. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 370 + ], + "type": "text", + "content": "[21] Hossam Isack and Yuri Boykov. Energy-based geometric multi-model fitting. International journal of computer vision, 97(2):123-147, 2012. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 426 + ], + "type": "text", + "content": "[22] Pan Ji, Mathieu Salzmann, and Hongdong Li. Shape interaction matrix revisited and robustified: Efficient subspace clustering with corrupted and incomplete data. In Proceedings of the IEEE International Conference on computer vision, pages 4687-4695, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 459 + ], + "type": "text", + "content": "[23] Pan Ji, Tong Zhang, Hongdong Li, Mathieu Salzmann, and Ian Reid. Deep subspace clustering networks. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 513 + ], + "type": "text", + "content": "[24] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 515, + 545, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 580 + ], + "type": "text", + "content": "[25] Florian Kluger, Eric Brachmann, Hanno Ackermann, Carsten Rother, Michael Ying Yang, and Bodo Rosenhahn. Consac: Robust multi-model fitting by conditional sample consensus. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4634-4643, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 635 + ], + "type": "text", + "content": "[26] Hans-Peter Kriegel, Peer Kroger, and Arthur Zimek. Clustering high-dimensional data: A survey on subspace clustering, pattern-based clustering, and correlation clustering. Acm transactions on knowledge discovery from data (tkdd), 3(1): 1-58, 2009. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 691 + ], + "type": "text", + "content": "[27] Connor Lane, Ron Boger, Chong You, Manolis Tsakiris, Benjamin Haeffele, and Rene Vidal. Classifying and comparing approaches to subspace clustering with missing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0-0, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[28] Zhuwen Li, Jiaming Guo, Loong-Fah Cheong, and Steven Zhiying Zhou. Perspective motion segmentation via" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19100" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "collaborative clustering. In Proceedings of the IEEE international conference on computer vision, pages 1369-1376, 2013. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[29] Guangcan Liu and Shuicheng Yan. Latent low-rank representation for subspace segmentation and feature extraction. In 2011 International Conference on Computer Vision, pages 1615-1622, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[30] Guangcan Liu, Zhouchen Lin, Shuicheng Yan, Ju Sun, Yong Yu, and Yi Ma. Robust recovery of subspace structures by low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(1):171-184, 2012. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 239 + ], + "type": "text", + "content": "[31] Luca Magri and Fusiello Andrea. Robust multiple model fitting with preference analysis and low-rank approximation. In Proceedings of the British Machine Vision Conference 2015, pages 20-1, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 282 + ], + "type": "text", + "content": "[32] Luca Magri and Andrea Fusiello. T-linkage: A continuous relaxation of j-linkage for multi-model fitting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3954-3961, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 284, + 287, + 326 + ], + "type": "text", + "content": "[33] Luca Magri and Andrea Fusiello. Multiple model fitting as a set coverage problem. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3318-3326, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[34] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3061-3070, 2015. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 424 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 459 + ], + "type": "text", + "content": "[36] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 525 + ], + "type": "text", + "content": "[37] Hyun Soo Park, Takaaki Shiratori, Iain Matthews, and Yaser Sheikh. 3d reconstruction of a moving point from a series of 2d projections. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part III 11, pages 158-171. Springer, 2010. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 568 + ], + "type": "text", + "content": "[38] C.J. Poelman and T. Kanade. A parapspective factorization method for shape and motion recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(3):206-218, 1997. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 624 + ], + "type": "text", + "content": "[39] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[40] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mlps. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[41] Shankar Rao, Roberto Tron, Rene Vidal, and Yi Ma. Motion segmentation in the presence of outlying, incomplete, or corrupted trajectories. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32(10):1832-1845, 2009. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 128 + ], + "type": "text", + "content": "[42] Shankar R Rao, Roberto Tron, René Vidal, and Yi Ma. Motion segmentation via robust subspace separation in the presence of outlying, incomplete, or corrupted trajectories. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[43] Junghoon Seo, Jamyoung Koo, and Taegyun Jeon. Deep closed-form subspace clustering. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[44] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in neural information processing systems, 33:7462-7473, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[45] Yasuyuki Sugaya and Kenichi Kanatani. Geometric structure of degeneracy for multi-body motion segmentation. In International Workshop on Statistical Methods in Video Processing, pages 13-25. Springer, 2004. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[46] Roberto Toldo and Andrea Fusiello. Robust multiple structures estimation with j-linkage. In Computer Vision – ECCV 2008, pages 537–547. Springer Berlin Heidelberg, 2008. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 331 + ], + "type": "text", + "content": "[47] Carlo Tomasi and Takeo Kanade. Shape and motion from image streams: a factorization method. International journal of computer vision, 9(2):137-154, 1992. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 374 + ], + "type": "text", + "content": "[48] Philip HS Torr. Geometric motion segmentation and model selection. Philosophical Transactions of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 356(1740):1321-1340, 1998. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 421 + ], + "type": "text", + "content": "[49] Roberto Tron and René Vidal. A benchmark for the comparison of 3-d motion segmentation algorithms. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[50] Manolis Tsakiris and Rene Vidal. Theoretical analysis of sparse subspace clustering with missing entries. In International Conference on Machine Learning, pages 4975-4984. PMLR, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 466, + 545, + 510 + ], + "type": "text", + "content": "[51] Jeya Maria Jose Valanarasu and Vishal M Patel. Overcomplete deep subspace clustering networks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 746-755, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[52] Roman Vershynin. High-dimensional probability: An introduction with applications in data science. Cambridge university press, 2018. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 600 + ], + "type": "text", + "content": "[53] René Vidal and Richard Hartley. Motion segmentation with missing data using powerfactorization and gpca. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II–II. IEEE, 2004. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 644 + ], + "type": "text", + "content": "[54] Rene Vidal, Yi Ma, and Shankar Sastry. Generalized principal component analysis (gpca). IEEE transactions on pattern analysis and machine intelligence, 27(12):1945-1959, 2005. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[55] René Vidal, Roberto Tron, and Richard Hartley. Multiframe motion segmentation with missing data using powerfactorization and gpca. International Journal of Computer Vision, 79:85-105, 2008. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[56] Xun Xu, Loong Fah Cheong, and Zhuwen Li. Motion segmentation by exploiting complementary geometric models." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19101" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2859-2867, 2018. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 162 + ], + "type": "text", + "content": "[57] Jingyu Yan and Marc Pollefeys. A general framework for motion segmentation: Independent, articulated, rigid, non-rigid, degenerate and non-degenerate. In Computer Vision-ECCV 2006: 9th European Conference on Computer Vision, Graz, Austria, May 7-13, 2006, Proceedings, Part IV 9, pages 94-106. Springer, 2006. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[58] Lihi Zelnik-Manor and Michal Irani. Temporal factorization vs. spatial factorization. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision, Prague, Czech Republic, May 11-14, 2004. Proceedings, Part II 8, pages 434-445. Springer, 2004. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 271 + ], + "type": "text", + "content": "[59] Shangzhi Zhang, Chong You, René Vidal, and Chun-Guang Li. Learning a self-expressive network for subspace clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12393-12403, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 306 + ], + "type": "text", + "content": "[60] Jianqiao Zheng, Sameera Ramasinghe, and Simon Lucey. Rethinking positional encoding. arXiv preprint arXiv:2107.02561, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[61] Yingying Zhu, Dong Huang, Fernando De La Torre, and Simon Lucey. Complex non-rigid motion 3d reconstruction by union of subspaces. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 287, + 407 + ], + "type": "text", + "content": "[62] Marco Zuliani, Charles S Kenney, and BS Manjunath. The multiransac algorithm and its application to detect planar homographies. In IEEE International Conference on Image Processing 2005, pages III-153. IEEE, 2005. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19102" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file